Mercurial > projects > ldc
annotate druntime/src/compiler/dmd/arrayfloat.d @ 759:d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
author | Tomas Lindquist Olsen <tomas.l.olsen@gmail.com> |
---|---|
date | Tue, 11 Nov 2008 01:52:37 +0100 |
parents | |
children |
rev | line source |
---|---|
759
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1 /*************************** |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2 * D programming language http://www.digitalmars.com/d/ |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
3 * Runtime support for float array operations. |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
4 * Based on code originally written by Burton Radons. |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
5 * Placed in public domain. |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
6 */ |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
7 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
8 module rt.arrayfloat; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
9 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
10 private import util.cpuid; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
11 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
12 version (Unittest) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
13 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
14 /* This is so unit tests will test every CPU variant |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
15 */ |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
16 int cpuid; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
17 const int CPUID_MAX = 5; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
18 bool mmx() { return cpuid == 1 && util.cpuid.mmx(); } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
19 bool sse() { return cpuid == 2 && util.cpuid.sse(); } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
20 bool sse2() { return cpuid == 3 && util.cpuid.sse2(); } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
21 bool amd3dnow() { return cpuid == 4 && util.cpuid.amd3dnow(); } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
22 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
23 else |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
24 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
25 alias util.cpuid.mmx mmx; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
26 alias util.cpuid.sse sse; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
27 alias util.cpuid.sse2 sse2; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
28 alias util.cpuid.amd3dnow amd3dnow; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
29 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
30 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
31 //version = log; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
32 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
33 bool disjoint(T)(T[] a, T[] b) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
34 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
35 return (a.ptr + a.length <= b.ptr || b.ptr + b.length <= a.ptr); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
36 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
37 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
38 alias float T; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
39 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
40 extern (C): |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
41 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
42 /* ======================================================================== */ |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
43 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
44 /*********************** |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
45 * Computes: |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
46 * a[] = b[] + c[] |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
47 */ |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
48 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
49 T[] _arraySliceSliceAddSliceAssign_f(T[] a, T[] c, T[] b) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
50 in |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
51 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
52 assert(a.length == b.length && b.length == c.length); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
53 assert(disjoint(a, b)); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
54 assert(disjoint(a, c)); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
55 assert(disjoint(b, c)); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
56 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
57 body |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
58 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
59 //printf("_arraySliceSliceAddSliceAssign_f()\n"); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
60 auto aptr = a.ptr; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
61 auto aend = aptr + a.length; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
62 auto bptr = b.ptr; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
63 auto cptr = c.ptr; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
64 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
65 version (D_InlineAsm_X86) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
66 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
67 // SSE version is 834% faster |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
68 if (sse() && b.length >= 16) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
69 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
70 version (log) printf("\tsse unaligned\n"); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
71 auto n = aptr + (b.length & ~15); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
72 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
73 // Unaligned case |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
74 asm |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
75 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
76 mov EAX, bptr; // left operand |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
77 mov ECX, cptr; // right operand |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
78 mov ESI, aptr; // destination operand |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
79 mov EDI, n; // end comparison |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
80 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
81 align 8; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
82 startsseloopb: |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
83 movups XMM0, [EAX]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
84 movups XMM1, [EAX+16]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
85 movups XMM2, [EAX+32]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
86 movups XMM3, [EAX+48]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
87 add EAX, 64; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
88 movups XMM4, [ECX]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
89 movups XMM5, [ECX+16]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
90 movups XMM6, [ECX+32]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
91 movups XMM7, [ECX+48]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
92 add ESI, 64; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
93 addps XMM0, XMM4; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
94 addps XMM1, XMM5; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
95 addps XMM2, XMM6; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
96 addps XMM3, XMM7; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
97 add ECX, 64; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
98 movups [ESI+ 0-64], XMM0; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
99 movups [ESI+16-64], XMM1; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
100 movups [ESI+32-64], XMM2; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
101 movups [ESI+48-64], XMM3; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
102 cmp ESI, EDI; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
103 jb startsseloopb; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
104 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
105 mov aptr, ESI; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
106 mov bptr, EAX; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
107 mov cptr, ECX; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
108 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
109 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
110 else |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
111 // 3DNow! version is only 13% faster |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
112 if (amd3dnow() && b.length >= 8) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
113 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
114 version (log) printf("\tamd3dnow\n"); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
115 auto n = aptr + (b.length & ~7); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
116 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
117 asm |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
118 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
119 mov ESI, aptr; // destination operand |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
120 mov EDI, n; // end comparison |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
121 mov EAX, bptr; // left operand |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
122 mov ECX, cptr; // right operand |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
123 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
124 align 4; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
125 start3dnow: |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
126 movq MM0, [EAX]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
127 movq MM1, [EAX+8]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
128 movq MM2, [EAX+16]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
129 movq MM3, [EAX+24]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
130 pfadd MM0, [ECX]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
131 pfadd MM1, [ECX+8]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
132 pfadd MM2, [ECX+16]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
133 pfadd MM3, [ECX+24]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
134 movq [ESI], MM0; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
135 movq [ESI+8], MM1; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
136 movq [ESI+16], MM2; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
137 movq [ESI+24], MM3; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
138 add ECX, 32; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
139 add ESI, 32; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
140 add EAX, 32; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
141 cmp ESI, EDI; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
142 jb start3dnow; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
143 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
144 emms; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
145 mov aptr, ESI; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
146 mov bptr, EAX; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
147 mov cptr, ECX; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
148 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
149 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
150 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
151 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
152 // Handle remainder |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
153 version (log) if (aptr < aend) printf("\tbase\n"); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
154 while (aptr < aend) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
155 *aptr++ = *bptr++ + *cptr++; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
156 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
157 return a; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
158 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
159 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
160 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
161 unittest |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
162 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
163 printf("_arraySliceSliceAddSliceAssign_f unittest\n"); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
164 for (cpuid = 0; cpuid < CPUID_MAX; cpuid++) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
165 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
166 version (log) printf(" cpuid %d\n", cpuid); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
167 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
168 for (int j = 0; j < 2; j++) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
169 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
170 const int dim = 67; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
171 T[] a = new T[dim + j]; // aligned on 16 byte boundary |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
172 a = a[j .. dim + j]; // misalign for second iteration |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
173 T[] b = new T[dim + j]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
174 b = b[j .. dim + j]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
175 T[] c = new T[dim + j]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
176 c = c[j .. dim + j]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
177 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
178 for (int i = 0; i < dim; i++) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
179 { a[i] = cast(T)i; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
180 b[i] = cast(T)(i + 7); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
181 c[i] = cast(T)(i * 2); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
182 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
183 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
184 c[] = a[] + b[]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
185 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
186 for (int i = 0; i < dim; i++) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
187 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
188 if (c[i] != cast(T)(a[i] + b[i])) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
189 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
190 printf("[%d]: %g != %g + %g\n", i, c[i], a[i], b[i]); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
191 assert(0); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
192 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
193 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
194 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
195 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
196 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
197 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
198 /* ======================================================================== */ |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
199 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
200 /*********************** |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
201 * Computes: |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
202 * a[] = b[] - c[] |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
203 */ |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
204 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
205 T[] _arraySliceSliceMinSliceAssign_f(T[] a, T[] c, T[] b) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
206 in |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
207 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
208 assert(a.length == b.length && b.length == c.length); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
209 assert(disjoint(a, b)); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
210 assert(disjoint(a, c)); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
211 assert(disjoint(b, c)); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
212 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
213 body |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
214 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
215 auto aptr = a.ptr; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
216 auto aend = aptr + a.length; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
217 auto bptr = b.ptr; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
218 auto cptr = c.ptr; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
219 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
220 version (D_InlineAsm_X86) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
221 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
222 // SSE version is 834% faster |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
223 if (sse() && b.length >= 16) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
224 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
225 auto n = aptr + (b.length & ~15); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
226 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
227 // Unaligned case |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
228 asm |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
229 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
230 mov EAX, bptr; // left operand |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
231 mov ECX, cptr; // right operand |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
232 mov ESI, aptr; // destination operand |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
233 mov EDI, n; // end comparison |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
234 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
235 align 8; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
236 startsseloopb: |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
237 movups XMM0, [EAX]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
238 movups XMM1, [EAX+16]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
239 movups XMM2, [EAX+32]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
240 movups XMM3, [EAX+48]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
241 add EAX, 64; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
242 movups XMM4, [ECX]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
243 movups XMM5, [ECX+16]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
244 movups XMM6, [ECX+32]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
245 movups XMM7, [ECX+48]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
246 add ESI, 64; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
247 subps XMM0, XMM4; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
248 subps XMM1, XMM5; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
249 subps XMM2, XMM6; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
250 subps XMM3, XMM7; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
251 add ECX, 64; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
252 movups [ESI+ 0-64], XMM0; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
253 movups [ESI+16-64], XMM1; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
254 movups [ESI+32-64], XMM2; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
255 movups [ESI+48-64], XMM3; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
256 cmp ESI, EDI; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
257 jb startsseloopb; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
258 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
259 mov aptr, ESI; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
260 mov bptr, EAX; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
261 mov cptr, ECX; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
262 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
263 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
264 else |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
265 // 3DNow! version is only 13% faster |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
266 if (amd3dnow() && b.length >= 8) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
267 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
268 auto n = aptr + (b.length & ~7); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
269 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
270 asm |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
271 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
272 mov ESI, aptr; // destination operand |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
273 mov EDI, n; // end comparison |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
274 mov EAX, bptr; // left operand |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
275 mov ECX, cptr; // right operand |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
276 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
277 align 4; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
278 start3dnow: |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
279 movq MM0, [EAX]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
280 movq MM1, [EAX+8]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
281 movq MM2, [EAX+16]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
282 movq MM3, [EAX+24]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
283 pfsub MM0, [ECX]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
284 pfsub MM1, [ECX+8]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
285 pfsub MM2, [ECX+16]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
286 pfsub MM3, [ECX+24]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
287 movq [ESI], MM0; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
288 movq [ESI+8], MM1; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
289 movq [ESI+16], MM2; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
290 movq [ESI+24], MM3; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
291 add ECX, 32; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
292 add ESI, 32; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
293 add EAX, 32; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
294 cmp ESI, EDI; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
295 jb start3dnow; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
296 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
297 emms; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
298 mov aptr, ESI; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
299 mov bptr, EAX; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
300 mov cptr, ECX; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
301 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
302 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
303 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
304 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
305 // Handle remainder |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
306 while (aptr < aend) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
307 *aptr++ = *bptr++ - *cptr++; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
308 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
309 return a; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
310 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
311 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
312 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
313 unittest |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
314 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
315 printf("_arraySliceSliceMinSliceAssign_f unittest\n"); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
316 for (cpuid = 0; cpuid < CPUID_MAX; cpuid++) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
317 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
318 version (log) printf(" cpuid %d\n", cpuid); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
319 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
320 for (int j = 0; j < 2; j++) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
321 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
322 const int dim = 67; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
323 T[] a = new T[dim + j]; // aligned on 16 byte boundary |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
324 a = a[j .. dim + j]; // misalign for second iteration |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
325 T[] b = new T[dim + j]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
326 b = b[j .. dim + j]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
327 T[] c = new T[dim + j]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
328 c = c[j .. dim + j]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
329 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
330 for (int i = 0; i < dim; i++) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
331 { a[i] = cast(T)i; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
332 b[i] = cast(T)(i + 7); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
333 c[i] = cast(T)(i * 2); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
334 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
335 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
336 c[] = a[] - b[]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
337 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
338 for (int i = 0; i < dim; i++) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
339 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
340 if (c[i] != cast(T)(a[i] - b[i])) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
341 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
342 printf("[%d]: %g != %gd - %g\n", i, c[i], a[i], b[i]); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
343 assert(0); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
344 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
345 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
346 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
347 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
348 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
349 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
350 /* ======================================================================== */ |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
351 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
352 /*********************** |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
353 * Computes: |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
354 * a[] = b[] + value |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
355 */ |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
356 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
357 T[] _arraySliceExpAddSliceAssign_f(T[] a, T value, T[] b) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
358 in |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
359 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
360 assert(a.length == b.length); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
361 assert(disjoint(a, b)); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
362 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
363 body |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
364 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
365 //printf("_arraySliceExpAddSliceAssign_f()\n"); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
366 auto aptr = a.ptr; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
367 auto aend = aptr + a.length; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
368 auto bptr = b.ptr; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
369 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
370 version (D_InlineAsm_X86) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
371 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
372 // SSE version is 665% faster |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
373 if (sse() && a.length >= 16) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
374 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
375 auto n = aptr + (a.length & ~15); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
376 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
377 // Unaligned case |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
378 asm |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
379 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
380 mov EAX, bptr; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
381 mov ESI, aptr; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
382 mov EDI, n; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
383 movss XMM4, value; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
384 shufps XMM4, XMM4, 0; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
385 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
386 align 8; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
387 startsseloop: |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
388 add ESI, 64; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
389 movups XMM0, [EAX]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
390 movups XMM1, [EAX+16]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
391 movups XMM2, [EAX+32]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
392 movups XMM3, [EAX+48]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
393 add EAX, 64; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
394 addps XMM0, XMM4; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
395 addps XMM1, XMM4; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
396 addps XMM2, XMM4; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
397 addps XMM3, XMM4; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
398 movups [ESI+ 0-64], XMM0; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
399 movups [ESI+16-64], XMM1; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
400 movups [ESI+32-64], XMM2; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
401 movups [ESI+48-64], XMM3; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
402 cmp ESI, EDI; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
403 jb startsseloop; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
404 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
405 mov aptr, ESI; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
406 mov bptr, EAX; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
407 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
408 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
409 else |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
410 // 3DNow! version is 69% faster |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
411 if (amd3dnow() && a.length >= 8) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
412 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
413 auto n = aptr + (a.length & ~7); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
414 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
415 ulong w = *cast(uint *) &value; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
416 ulong v = w | (w << 32L); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
417 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
418 asm |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
419 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
420 mov ESI, aptr; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
421 mov EDI, n; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
422 mov EAX, bptr; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
423 movq MM4, qword ptr [v]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
424 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
425 align 8; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
426 start3dnow: |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
427 movq MM0, [EAX]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
428 movq MM1, [EAX+8]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
429 movq MM2, [EAX+16]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
430 movq MM3, [EAX+24]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
431 pfadd MM0, MM4; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
432 pfadd MM1, MM4; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
433 pfadd MM2, MM4; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
434 pfadd MM3, MM4; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
435 movq [ESI], MM0; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
436 movq [ESI+8], MM1; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
437 movq [ESI+16], MM2; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
438 movq [ESI+24], MM3; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
439 add ESI, 32; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
440 add EAX, 32; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
441 cmp ESI, EDI; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
442 jb start3dnow; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
443 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
444 emms; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
445 mov aptr, ESI; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
446 mov bptr, EAX; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
447 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
448 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
449 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
450 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
451 while (aptr < aend) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
452 *aptr++ = *bptr++ + value; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
453 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
454 return a; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
455 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
456 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
457 unittest |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
458 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
459 printf("_arraySliceExpAddSliceAssign_f unittest\n"); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
460 for (cpuid = 0; cpuid < CPUID_MAX; cpuid++) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
461 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
462 version (log) printf(" cpuid %d\n", cpuid); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
463 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
464 for (int j = 0; j < 2; j++) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
465 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
466 const int dim = 67; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
467 T[] a = new T[dim + j]; // aligned on 16 byte boundary |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
468 a = a[j .. dim + j]; // misalign for second iteration |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
469 T[] b = new T[dim + j]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
470 b = b[j .. dim + j]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
471 T[] c = new T[dim + j]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
472 c = c[j .. dim + j]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
473 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
474 for (int i = 0; i < dim; i++) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
475 { a[i] = cast(T)i; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
476 b[i] = cast(T)(i + 7); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
477 c[i] = cast(T)(i * 2); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
478 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
479 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
480 c[] = a[] + 6; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
481 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
482 for (int i = 0; i < dim; i++) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
483 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
484 if (c[i] != cast(T)(a[i] + 6)) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
485 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
486 printf("[%d]: %g != %g + 6\n", i, c[i], a[i]); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
487 assert(0); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
488 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
489 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
490 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
491 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
492 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
493 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
494 /* ======================================================================== */ |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
495 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
496 /*********************** |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
497 * Computes: |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
498 * a[] += value |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
499 */ |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
500 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
501 T[] _arrayExpSliceAddass_f(T[] a, T value) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
502 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
503 //printf("_arrayExpSliceAddass_f(a.length = %d, value = %Lg)\n", a.length, cast(real)value); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
504 auto aptr = a.ptr; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
505 auto aend = aptr + a.length; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
506 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
507 version (D_InlineAsm_X86) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
508 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
509 // SSE version is 302% faster |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
510 if (sse() && a.length >= 16) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
511 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
512 // align pointer |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
513 auto n = cast(T*)((cast(uint)aptr + 15) & ~15); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
514 while (aptr < n) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
515 *aptr++ += value; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
516 n = cast(T*)((cast(uint)aend) & ~15); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
517 if (aptr < n) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
518 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
519 // Aligned case |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
520 asm |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
521 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
522 mov ESI, aptr; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
523 mov EDI, n; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
524 movss XMM4, value; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
525 shufps XMM4, XMM4, 0; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
526 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
527 align 8; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
528 startsseloopa: |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
529 movaps XMM0, [ESI]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
530 movaps XMM1, [ESI+16]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
531 movaps XMM2, [ESI+32]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
532 movaps XMM3, [ESI+48]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
533 add ESI, 64; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
534 addps XMM0, XMM4; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
535 addps XMM1, XMM4; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
536 addps XMM2, XMM4; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
537 addps XMM3, XMM4; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
538 movaps [ESI+ 0-64], XMM0; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
539 movaps [ESI+16-64], XMM1; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
540 movaps [ESI+32-64], XMM2; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
541 movaps [ESI+48-64], XMM3; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
542 cmp ESI, EDI; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
543 jb startsseloopa; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
544 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
545 mov aptr, ESI; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
546 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
547 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
548 else |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
549 // 3DNow! version is 63% faster |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
550 if (amd3dnow() && a.length >= 8) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
551 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
552 auto n = aptr + (a.length & ~7); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
553 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
554 ulong w = *cast(uint *) &value; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
555 ulong v = w | (w << 32L); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
556 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
557 asm |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
558 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
559 mov ESI, dword ptr [aptr]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
560 mov EDI, dword ptr [n]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
561 movq MM4, qword ptr [v]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
562 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
563 align 8; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
564 start3dnow: |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
565 movq MM0, [ESI]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
566 movq MM1, [ESI+8]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
567 movq MM2, [ESI+16]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
568 movq MM3, [ESI+24]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
569 pfadd MM0, MM4; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
570 pfadd MM1, MM4; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
571 pfadd MM2, MM4; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
572 pfadd MM3, MM4; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
573 movq [ESI], MM0; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
574 movq [ESI+8], MM1; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
575 movq [ESI+16], MM2; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
576 movq [ESI+24], MM3; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
577 add ESI, 32; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
578 cmp ESI, EDI; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
579 jb start3dnow; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
580 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
581 emms; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
582 mov dword ptr [aptr], ESI; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
583 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
584 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
585 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
586 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
587 while (aptr < aend) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
588 *aptr++ += value; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
589 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
590 return a; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
591 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
592 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
593 unittest |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
594 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
595 printf("_arrayExpSliceAddass_f unittest\n"); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
596 for (cpuid = 0; cpuid < CPUID_MAX; cpuid++) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
597 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
598 version (log) printf(" cpuid %d\n", cpuid); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
599 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
600 for (int j = 0; j < 2; j++) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
601 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
602 const int dim = 67; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
603 T[] a = new T[dim + j]; // aligned on 16 byte boundary |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
604 a = a[j .. dim + j]; // misalign for second iteration |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
605 T[] b = new T[dim + j]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
606 b = b[j .. dim + j]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
607 T[] c = new T[dim + j]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
608 c = c[j .. dim + j]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
609 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
610 for (int i = 0; i < dim; i++) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
611 { a[i] = cast(T)i; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
612 b[i] = cast(T)(i + 7); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
613 c[i] = cast(T)(i * 2); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
614 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
615 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
616 a[] = c[]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
617 c[] += 6; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
618 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
619 for (int i = 0; i < dim; i++) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
620 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
621 if (c[i] != cast(T)(a[i] + 6)) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
622 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
623 printf("[%d]: %g != %g + 6\n", i, c[i], a[i]); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
624 assert(0); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
625 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
626 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
627 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
628 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
629 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
630 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
631 /* ======================================================================== */ |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
632 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
633 /*********************** |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
634 * Computes: |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
635 * a[] += b[] |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
636 */ |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
637 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
638 T[] _arraySliceSliceAddass_f(T[] a, T[] b) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
639 in |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
640 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
641 assert (a.length == b.length); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
642 assert (disjoint(a, b)); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
643 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
644 body |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
645 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
646 //printf("_arraySliceSliceAddass_f()\n"); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
647 auto aptr = a.ptr; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
648 auto aend = aptr + a.length; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
649 auto bptr = b.ptr; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
650 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
651 version (D_InlineAsm_X86) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
652 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
653 // SSE version is 468% faster |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
654 if (sse() && a.length >= 16) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
655 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
656 auto n = aptr + (a.length & ~15); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
657 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
658 // Unaligned case |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
659 asm |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
660 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
661 mov ECX, bptr; // right operand |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
662 mov ESI, aptr; // destination operand |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
663 mov EDI, n; // end comparison |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
664 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
665 align 8; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
666 startsseloopb: |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
667 movups XMM0, [ESI]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
668 movups XMM1, [ESI+16]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
669 movups XMM2, [ESI+32]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
670 movups XMM3, [ESI+48]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
671 add ESI, 64; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
672 movups XMM4, [ECX]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
673 movups XMM5, [ECX+16]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
674 movups XMM6, [ECX+32]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
675 movups XMM7, [ECX+48]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
676 add ECX, 64; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
677 addps XMM0, XMM4; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
678 addps XMM1, XMM5; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
679 addps XMM2, XMM6; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
680 addps XMM3, XMM7; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
681 movups [ESI+ 0-64], XMM0; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
682 movups [ESI+16-64], XMM1; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
683 movups [ESI+32-64], XMM2; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
684 movups [ESI+48-64], XMM3; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
685 cmp ESI, EDI; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
686 jb startsseloopb; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
687 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
688 mov aptr, ESI; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
689 mov bptr, ECX; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
690 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
691 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
692 else |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
693 // 3DNow! version is 57% faster |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
694 if (amd3dnow() && a.length >= 8) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
695 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
696 auto n = aptr + (a.length & ~7); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
697 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
698 asm |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
699 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
700 mov ESI, dword ptr [aptr]; // destination operand |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
701 mov EDI, dword ptr [n]; // end comparison |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
702 mov ECX, dword ptr [bptr]; // right operand |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
703 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
704 align 4; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
705 start3dnow: |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
706 movq MM0, [ESI]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
707 movq MM1, [ESI+8]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
708 movq MM2, [ESI+16]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
709 movq MM3, [ESI+24]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
710 pfadd MM0, [ECX]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
711 pfadd MM1, [ECX+8]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
712 pfadd MM2, [ECX+16]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
713 pfadd MM3, [ECX+24]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
714 movq [ESI], MM0; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
715 movq [ESI+8], MM1; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
716 movq [ESI+16], MM2; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
717 movq [ESI+24], MM3; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
718 add ESI, 32; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
719 add ECX, 32; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
720 cmp ESI, EDI; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
721 jb start3dnow; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
722 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
723 emms; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
724 mov dword ptr [aptr], ESI; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
725 mov dword ptr [bptr], ECX; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
726 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
727 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
728 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
729 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
730 while (aptr < aend) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
731 *aptr++ += *bptr++; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
732 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
733 return a; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
734 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
735 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
736 unittest |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
737 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
738 printf("_arraySliceSliceAddass_f unittest\n"); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
739 for (cpuid = 0; cpuid < CPUID_MAX; cpuid++) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
740 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
741 version (log) printf(" cpuid %d\n", cpuid); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
742 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
743 for (int j = 0; j < 2; j++) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
744 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
745 const int dim = 67; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
746 T[] a = new T[dim + j]; // aligned on 16 byte boundary |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
747 a = a[j .. dim + j]; // misalign for second iteration |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
748 T[] b = new T[dim + j]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
749 b = b[j .. dim + j]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
750 T[] c = new T[dim + j]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
751 c = c[j .. dim + j]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
752 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
753 for (int i = 0; i < dim; i++) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
754 { a[i] = cast(T)i; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
755 b[i] = cast(T)(i + 7); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
756 c[i] = cast(T)(i * 2); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
757 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
758 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
759 a[] = c[]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
760 c[] += b[]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
761 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
762 for (int i = 0; i < dim; i++) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
763 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
764 if (c[i] != cast(T)(a[i] + b[i])) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
765 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
766 printf("[%d]: %g != %g + %g\n", i, c[i], a[i], b[i]); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
767 assert(0); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
768 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
769 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
770 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
771 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
772 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
773 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
774 /* ======================================================================== */ |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
775 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
776 /*********************** |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
777 * Computes: |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
778 * a[] = b[] - value |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
779 */ |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
780 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
781 T[] _arraySliceExpMinSliceAssign_f(T[] a, T value, T[] b) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
782 in |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
783 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
784 assert (a.length == b.length); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
785 assert (disjoint(a, b)); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
786 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
787 body |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
788 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
789 //printf("_arraySliceExpMinSliceAssign_f()\n"); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
790 auto aptr = a.ptr; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
791 auto aend = aptr + a.length; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
792 auto bptr = b.ptr; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
793 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
794 version (D_InlineAsm_X86) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
795 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
796 // SSE version is 622% faster |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
797 if (sse() && a.length >= 16) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
798 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
799 auto n = aptr + (a.length & ~15); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
800 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
801 // Unaligned case |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
802 asm |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
803 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
804 mov EAX, bptr; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
805 mov ESI, aptr; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
806 mov EDI, n; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
807 movss XMM4, value; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
808 shufps XMM4, XMM4, 0; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
809 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
810 align 8; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
811 startsseloop: |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
812 add ESI, 64; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
813 movups XMM0, [EAX]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
814 movups XMM1, [EAX+16]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
815 movups XMM2, [EAX+32]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
816 movups XMM3, [EAX+48]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
817 add EAX, 64; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
818 subps XMM0, XMM4; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
819 subps XMM1, XMM4; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
820 subps XMM2, XMM4; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
821 subps XMM3, XMM4; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
822 movups [ESI+ 0-64], XMM0; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
823 movups [ESI+16-64], XMM1; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
824 movups [ESI+32-64], XMM2; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
825 movups [ESI+48-64], XMM3; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
826 cmp ESI, EDI; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
827 jb startsseloop; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
828 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
829 mov aptr, ESI; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
830 mov bptr, EAX; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
831 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
832 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
833 else |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
834 // 3DNow! version is 67% faster |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
835 if (amd3dnow() && a.length >= 8) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
836 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
837 auto n = aptr + (a.length & ~7); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
838 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
839 T[2] w; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
840 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
841 w[0] = w[1] = value; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
842 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
843 asm |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
844 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
845 mov ESI, dword ptr [aptr]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
846 mov EDI, dword ptr [n]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
847 mov EAX, dword ptr [bptr]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
848 movq MM4, qword ptr [w]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
849 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
850 align 8; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
851 start3dnow: |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
852 movq MM0, [EAX]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
853 movq MM1, [EAX+8]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
854 movq MM2, [EAX+16]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
855 movq MM3, [EAX+24]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
856 pfsub MM0, MM4; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
857 pfsub MM1, MM4; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
858 pfsub MM2, MM4; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
859 pfsub MM3, MM4; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
860 movq [ESI], MM0; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
861 movq [ESI+8], MM1; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
862 movq [ESI+16], MM2; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
863 movq [ESI+24], MM3; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
864 add ESI, 32; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
865 add EAX, 32; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
866 cmp ESI, EDI; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
867 jb start3dnow; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
868 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
869 emms; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
870 mov dword ptr [aptr], ESI; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
871 mov dword ptr [bptr], EAX; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
872 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
873 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
874 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
875 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
876 while (aptr < aend) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
877 *aptr++ = *bptr++ - value; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
878 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
879 return a; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
880 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
881 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
882 unittest |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
883 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
884 printf("_arraySliceExpMinSliceAssign_f unittest\n"); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
885 for (cpuid = 0; cpuid < CPUID_MAX; cpuid++) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
886 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
887 version (log) printf(" cpuid %d\n", cpuid); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
888 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
889 for (int j = 0; j < 2; j++) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
890 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
891 const int dim = 67; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
892 T[] a = new T[dim + j]; // aligned on 16 byte boundary |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
893 a = a[j .. dim + j]; // misalign for second iteration |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
894 T[] b = new T[dim + j]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
895 b = b[j .. dim + j]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
896 T[] c = new T[dim + j]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
897 c = c[j .. dim + j]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
898 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
899 for (int i = 0; i < dim; i++) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
900 { a[i] = cast(T)i; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
901 b[i] = cast(T)(i + 7); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
902 c[i] = cast(T)(i * 2); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
903 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
904 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
905 c[] = a[] - 6; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
906 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
907 for (int i = 0; i < dim; i++) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
908 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
909 if (c[i] != cast(T)(a[i] - 6)) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
910 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
911 printf("[%d]: %g != %g - 6\n", i, c[i], a[i]); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
912 assert(0); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
913 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
914 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
915 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
916 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
917 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
918 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
919 /* ======================================================================== */ |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
920 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
921 /*********************** |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
922 * Computes: |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
923 * a[] = value - b[] |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
924 */ |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
925 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
926 T[] _arrayExpSliceMinSliceAssign_f(T[] a, T[] b, T value) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
927 in |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
928 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
929 assert (a.length == b.length); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
930 assert (disjoint(a, b)); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
931 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
932 body |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
933 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
934 //printf("_arrayExpSliceMinSliceAssign_f()\n"); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
935 auto aptr = a.ptr; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
936 auto aend = aptr + a.length; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
937 auto bptr = b.ptr; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
938 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
939 version (D_InlineAsm_X86) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
940 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
941 // SSE version is 690% faster |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
942 if (sse() && a.length >= 16) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
943 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
944 auto n = aptr + (a.length & ~15); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
945 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
946 // Unaligned case |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
947 asm |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
948 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
949 mov EAX, bptr; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
950 mov ESI, aptr; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
951 mov EDI, n; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
952 movss XMM4, value; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
953 shufps XMM4, XMM4, 0; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
954 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
955 align 8; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
956 startsseloop: |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
957 add ESI, 64; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
958 movaps XMM5, XMM4; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
959 movaps XMM6, XMM4; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
960 movups XMM0, [EAX]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
961 movups XMM1, [EAX+16]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
962 movups XMM2, [EAX+32]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
963 movups XMM3, [EAX+48]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
964 add EAX, 64; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
965 subps XMM5, XMM0; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
966 subps XMM6, XMM1; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
967 movups [ESI+ 0-64], XMM5; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
968 movups [ESI+16-64], XMM6; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
969 movaps XMM5, XMM4; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
970 movaps XMM6, XMM4; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
971 subps XMM5, XMM2; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
972 subps XMM6, XMM3; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
973 movups [ESI+32-64], XMM5; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
974 movups [ESI+48-64], XMM6; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
975 cmp ESI, EDI; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
976 jb startsseloop; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
977 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
978 mov aptr, ESI; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
979 mov bptr, EAX; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
980 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
981 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
982 else |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
983 // 3DNow! version is 67% faster |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
984 if (amd3dnow() && a.length >= 8) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
985 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
986 auto n = aptr + (a.length & ~7); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
987 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
988 ulong w = *cast(uint *) &value; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
989 ulong v = w | (w << 32L); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
990 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
991 asm |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
992 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
993 mov ESI, aptr; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
994 mov EDI, n; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
995 mov EAX, bptr; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
996 movq MM4, qword ptr [v]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
997 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
998 align 8; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
999 start3dnow: |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1000 movq MM0, [EAX]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1001 movq MM1, [EAX+8]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1002 movq MM2, [EAX+16]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1003 movq MM3, [EAX+24]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1004 pfsubr MM0, MM4; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1005 pfsubr MM1, MM4; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1006 pfsubr MM2, MM4; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1007 pfsubr MM3, MM4; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1008 movq [ESI], MM0; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1009 movq [ESI+8], MM1; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1010 movq [ESI+16], MM2; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1011 movq [ESI+24], MM3; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1012 add ESI, 32; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1013 add EAX, 32; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1014 cmp ESI, EDI; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1015 jb start3dnow; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1016 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1017 emms; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1018 mov aptr, ESI; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1019 mov bptr, EAX; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1020 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1021 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1022 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1023 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1024 while (aptr < aend) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1025 *aptr++ = value - *bptr++; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1026 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1027 return a; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1028 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1029 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1030 unittest |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1031 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1032 printf("_arrayExpSliceMinSliceAssign_f unittest\n"); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1033 for (cpuid = 0; cpuid < CPUID_MAX; cpuid++) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1034 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1035 version (log) printf(" cpuid %d\n", cpuid); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1036 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1037 for (int j = 0; j < 2; j++) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1038 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1039 const int dim = 67; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1040 T[] a = new T[dim + j]; // aligned on 16 byte boundary |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1041 a = a[j .. dim + j]; // misalign for second iteration |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1042 T[] b = new T[dim + j]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1043 b = b[j .. dim + j]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1044 T[] c = new T[dim + j]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1045 c = c[j .. dim + j]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1046 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1047 for (int i = 0; i < dim; i++) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1048 { a[i] = cast(T)i; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1049 b[i] = cast(T)(i + 7); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1050 c[i] = cast(T)(i * 2); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1051 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1052 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1053 c[] = 6 - a[]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1054 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1055 for (int i = 0; i < dim; i++) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1056 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1057 if (c[i] != cast(T)(6 - a[i])) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1058 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1059 printf("[%d]: %g != 6 - %g\n", i, c[i], a[i]); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1060 assert(0); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1061 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1062 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1063 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1064 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1065 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1066 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1067 /* ======================================================================== */ |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1068 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1069 /*********************** |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1070 * Computes: |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1071 * a[] -= value |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1072 */ |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1073 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1074 T[] _arrayExpSliceMinass_f(T[] a, T value) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1075 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1076 //printf("_arrayExpSliceMinass_f(a.length = %d, value = %Lg)\n", a.length, cast(real)value); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1077 auto aptr = a.ptr; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1078 auto aend = aptr + a.length; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1079 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1080 version (D_InlineAsm_X86) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1081 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1082 // SSE version is 304% faster |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1083 if (sse() && a.length >= 16) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1084 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1085 // align pointer |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1086 auto n = cast(T*)((cast(uint)aptr + 15) & ~15); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1087 while (aptr < n) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1088 *aptr++ -= value; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1089 n = cast(T*)((cast(uint)aend) & ~15); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1090 if (aptr < n) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1091 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1092 // Aligned case |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1093 asm |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1094 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1095 mov ESI, aptr; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1096 mov EDI, n; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1097 movss XMM4, value; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1098 shufps XMM4, XMM4, 0; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1099 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1100 align 8; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1101 startsseloopa: |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1102 movaps XMM0, [ESI]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1103 movaps XMM1, [ESI+16]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1104 movaps XMM2, [ESI+32]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1105 movaps XMM3, [ESI+48]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1106 add ESI, 64; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1107 subps XMM0, XMM4; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1108 subps XMM1, XMM4; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1109 subps XMM2, XMM4; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1110 subps XMM3, XMM4; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1111 movaps [ESI+ 0-64], XMM0; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1112 movaps [ESI+16-64], XMM1; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1113 movaps [ESI+32-64], XMM2; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1114 movaps [ESI+48-64], XMM3; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1115 cmp ESI, EDI; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1116 jb startsseloopa; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1117 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1118 mov aptr, ESI; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1119 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1120 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1121 else |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1122 // 3DNow! version is 63% faster |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1123 if (amd3dnow() && a.length >= 8) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1124 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1125 auto n = aptr + (a.length & ~7); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1126 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1127 ulong w = *cast(uint *) &value; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1128 ulong v = w | (w << 32L); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1129 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1130 asm |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1131 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1132 mov ESI, dword ptr [aptr]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1133 mov EDI, dword ptr [n]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1134 movq MM4, qword ptr [v]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1135 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1136 align 8; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1137 start: |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1138 movq MM0, [ESI]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1139 movq MM1, [ESI+8]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1140 movq MM2, [ESI+16]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1141 movq MM3, [ESI+24]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1142 pfsub MM0, MM4; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1143 pfsub MM1, MM4; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1144 pfsub MM2, MM4; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1145 pfsub MM3, MM4; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1146 movq [ESI], MM0; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1147 movq [ESI+8], MM1; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1148 movq [ESI+16], MM2; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1149 movq [ESI+24], MM3; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1150 add ESI, 32; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1151 cmp ESI, EDI; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1152 jb start; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1153 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1154 emms; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1155 mov dword ptr [aptr], ESI; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1156 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1157 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1158 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1159 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1160 while (aptr < aend) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1161 *aptr++ -= value; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1162 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1163 return a; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1164 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1165 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1166 unittest |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1167 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1168 printf("_arrayExpSliceminass_f unittest\n"); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1169 for (cpuid = 0; cpuid < CPUID_MAX; cpuid++) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1170 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1171 version (log) printf(" cpuid %d\n", cpuid); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1172 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1173 for (int j = 0; j < 2; j++) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1174 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1175 const int dim = 67; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1176 T[] a = new T[dim + j]; // aligned on 16 byte boundary |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1177 a = a[j .. dim + j]; // misalign for second iteration |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1178 T[] b = new T[dim + j]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1179 b = b[j .. dim + j]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1180 T[] c = new T[dim + j]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1181 c = c[j .. dim + j]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1182 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1183 for (int i = 0; i < dim; i++) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1184 { a[i] = cast(T)i; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1185 b[i] = cast(T)(i + 7); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1186 c[i] = cast(T)(i * 2); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1187 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1188 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1189 a[] = c[]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1190 c[] -= 6; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1191 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1192 for (int i = 0; i < dim; i++) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1193 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1194 if (c[i] != cast(T)(a[i] - 6)) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1195 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1196 printf("[%d]: %g != %g - 6\n", i, c[i], a[i]); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1197 assert(0); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1198 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1199 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1200 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1201 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1202 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1203 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1204 /* ======================================================================== */ |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1205 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1206 /*********************** |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1207 * Computes: |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1208 * a[] -= b[] |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1209 */ |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1210 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1211 T[] _arraySliceSliceMinass_f(T[] a, T[] b) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1212 in |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1213 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1214 assert (a.length == b.length); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1215 assert (disjoint(a, b)); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1216 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1217 body |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1218 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1219 //printf("_arraySliceSliceMinass_f()\n"); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1220 auto aptr = a.ptr; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1221 auto aend = aptr + a.length; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1222 auto bptr = b.ptr; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1223 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1224 version (D_InlineAsm_X86) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1225 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1226 // SSE version is 468% faster |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1227 if (sse() && a.length >= 16) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1228 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1229 auto n = aptr + (a.length & ~15); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1230 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1231 // Unaligned case |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1232 asm |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1233 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1234 mov ECX, bptr; // right operand |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1235 mov ESI, aptr; // destination operand |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1236 mov EDI, n; // end comparison |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1237 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1238 align 8; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1239 startsseloopb: |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1240 movups XMM0, [ESI]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1241 movups XMM1, [ESI+16]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1242 movups XMM2, [ESI+32]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1243 movups XMM3, [ESI+48]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1244 add ESI, 64; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1245 movups XMM4, [ECX]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1246 movups XMM5, [ECX+16]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1247 movups XMM6, [ECX+32]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1248 movups XMM7, [ECX+48]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1249 add ECX, 64; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1250 subps XMM0, XMM4; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1251 subps XMM1, XMM5; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1252 subps XMM2, XMM6; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1253 subps XMM3, XMM7; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1254 movups [ESI+ 0-64], XMM0; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1255 movups [ESI+16-64], XMM1; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1256 movups [ESI+32-64], XMM2; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1257 movups [ESI+48-64], XMM3; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1258 cmp ESI, EDI; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1259 jb startsseloopb; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1260 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1261 mov aptr, ESI; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1262 mov bptr, ECX; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1263 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1264 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1265 else |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1266 // 3DNow! version is 57% faster |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1267 if (amd3dnow() && a.length >= 8) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1268 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1269 auto n = aptr + (a.length & ~7); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1270 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1271 asm |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1272 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1273 mov ESI, dword ptr [aptr]; // destination operand |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1274 mov EDI, dword ptr [n]; // end comparison |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1275 mov ECX, dword ptr [bptr]; // right operand |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1276 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1277 align 4; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1278 start: |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1279 movq MM0, [ESI]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1280 movq MM1, [ESI+8]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1281 movq MM2, [ESI+16]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1282 movq MM3, [ESI+24]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1283 pfsub MM0, [ECX]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1284 pfsub MM1, [ECX+8]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1285 pfsub MM2, [ECX+16]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1286 pfsub MM3, [ECX+24]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1287 movq [ESI], MM0; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1288 movq [ESI+8], MM1; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1289 movq [ESI+16], MM2; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1290 movq [ESI+24], MM3; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1291 add ESI, 32; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1292 add ECX, 32; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1293 cmp ESI, EDI; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1294 jb start; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1295 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1296 emms; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1297 mov aptr, ESI; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1298 mov bptr, ECX; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1299 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1300 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1301 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1302 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1303 while (aptr < aend) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1304 *aptr++ -= *bptr++; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1305 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1306 return a; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1307 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1308 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1309 unittest |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1310 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1311 printf("_arrayExpSliceMinass_f unittest\n"); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1312 for (cpuid = 0; cpuid < CPUID_MAX; cpuid++) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1313 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1314 version (log) printf(" cpuid %d\n", cpuid); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1315 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1316 for (int j = 0; j < 2; j++) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1317 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1318 const int dim = 67; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1319 T[] a = new T[dim + j]; // aligned on 16 byte boundary |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1320 a = a[j .. dim + j]; // misalign for second iteration |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1321 T[] b = new T[dim + j]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1322 b = b[j .. dim + j]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1323 T[] c = new T[dim + j]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1324 c = c[j .. dim + j]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1325 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1326 for (int i = 0; i < dim; i++) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1327 { a[i] = cast(T)i; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1328 b[i] = cast(T)(i + 7); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1329 c[i] = cast(T)(i * 2); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1330 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1331 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1332 a[] = c[]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1333 c[] -= 6; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1334 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1335 for (int i = 0; i < dim; i++) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1336 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1337 if (c[i] != cast(T)(a[i] - 6)) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1338 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1339 printf("[%d]: %g != %g - 6\n", i, c[i], a[i]); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1340 assert(0); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1341 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1342 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1343 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1344 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1345 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1346 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1347 /* ======================================================================== */ |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1348 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1349 /*********************** |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1350 * Computes: |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1351 * a[] = b[] * value |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1352 */ |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1353 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1354 T[] _arraySliceExpMulSliceAssign_f(T[] a, T value, T[] b) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1355 in |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1356 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1357 assert(a.length == b.length); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1358 assert(disjoint(a, b)); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1359 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1360 body |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1361 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1362 //printf("_arraySliceExpMulSliceAssign_f()\n"); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1363 auto aptr = a.ptr; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1364 auto aend = aptr + a.length; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1365 auto bptr = b.ptr; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1366 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1367 version (D_InlineAsm_X86) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1368 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1369 // SSE version is 607% faster |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1370 if (sse() && a.length >= 16) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1371 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1372 auto n = aptr + (a.length & ~15); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1373 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1374 // Unaligned case |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1375 asm |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1376 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1377 mov EAX, bptr; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1378 mov ESI, aptr; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1379 mov EDI, n; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1380 movss XMM4, value; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1381 shufps XMM4, XMM4, 0; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1382 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1383 align 8; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1384 startsseloop: |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1385 add ESI, 64; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1386 movups XMM0, [EAX]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1387 movups XMM1, [EAX+16]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1388 movups XMM2, [EAX+32]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1389 movups XMM3, [EAX+48]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1390 add EAX, 64; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1391 mulps XMM0, XMM4; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1392 mulps XMM1, XMM4; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1393 mulps XMM2, XMM4; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1394 mulps XMM3, XMM4; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1395 movups [ESI+ 0-64], XMM0; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1396 movups [ESI+16-64], XMM1; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1397 movups [ESI+32-64], XMM2; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1398 movups [ESI+48-64], XMM3; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1399 cmp ESI, EDI; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1400 jb startsseloop; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1401 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1402 mov aptr, ESI; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1403 mov bptr, EAX; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1404 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1405 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1406 else |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1407 // 3DNow! version is 69% faster |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1408 if (amd3dnow() && a.length >= 8) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1409 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1410 auto n = aptr + (a.length & ~7); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1411 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1412 ulong w = *cast(uint *) &value; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1413 ulong v = w | (w << 32L); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1414 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1415 asm |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1416 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1417 mov ESI, dword ptr [aptr]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1418 mov EDI, dword ptr [n]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1419 mov EAX, dword ptr [bptr]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1420 movq MM4, qword ptr [v]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1421 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1422 align 8; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1423 start: |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1424 movq MM0, [EAX]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1425 movq MM1, [EAX+8]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1426 movq MM2, [EAX+16]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1427 movq MM3, [EAX+24]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1428 pfmul MM0, MM4; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1429 pfmul MM1, MM4; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1430 pfmul MM2, MM4; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1431 pfmul MM3, MM4; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1432 movq [ESI], MM0; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1433 movq [ESI+8], MM1; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1434 movq [ESI+16], MM2; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1435 movq [ESI+24], MM3; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1436 add ESI, 32; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1437 add EAX, 32; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1438 cmp ESI, EDI; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1439 jb start; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1440 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1441 emms; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1442 mov aptr, ESI; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1443 mov bptr, EAX; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1444 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1445 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1446 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1447 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1448 while (aptr < aend) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1449 *aptr++ = *bptr++ * value; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1450 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1451 return a; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1452 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1453 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1454 unittest |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1455 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1456 printf("_arraySliceExpMulSliceAssign_f unittest\n"); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1457 for (cpuid = 0; cpuid < CPUID_MAX; cpuid++) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1458 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1459 version (log) printf(" cpuid %d\n", cpuid); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1460 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1461 for (int j = 0; j < 2; j++) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1462 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1463 const int dim = 67; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1464 T[] a = new T[dim + j]; // aligned on 16 byte boundary |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1465 a = a[j .. dim + j]; // misalign for second iteration |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1466 T[] b = new T[dim + j]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1467 b = b[j .. dim + j]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1468 T[] c = new T[dim + j]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1469 c = c[j .. dim + j]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1470 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1471 for (int i = 0; i < dim; i++) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1472 { a[i] = cast(T)i; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1473 b[i] = cast(T)(i + 7); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1474 c[i] = cast(T)(i * 2); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1475 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1476 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1477 c[] = a[] * 6; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1478 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1479 for (int i = 0; i < dim; i++) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1480 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1481 if (c[i] != cast(T)(a[i] * 6)) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1482 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1483 printf("[%d]: %g != %g * 6\n", i, c[i], a[i]); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1484 assert(0); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1485 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1486 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1487 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1488 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1489 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1490 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1491 /* ======================================================================== */ |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1492 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1493 /*********************** |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1494 * Computes: |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1495 * a[] = b[] * c[] |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1496 */ |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1497 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1498 T[] _arraySliceSliceMulSliceAssign_f(T[] a, T[] c, T[] b) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1499 in |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1500 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1501 assert(a.length == b.length && b.length == c.length); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1502 assert(disjoint(a, b)); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1503 assert(disjoint(a, c)); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1504 assert(disjoint(b, c)); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1505 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1506 body |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1507 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1508 //printf("_arraySliceSliceMulSliceAssign_f()\n"); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1509 auto aptr = a.ptr; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1510 auto aend = aptr + a.length; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1511 auto bptr = b.ptr; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1512 auto cptr = c.ptr; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1513 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1514 version (D_InlineAsm_X86) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1515 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1516 // SSE version is 833% faster |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1517 if (sse() && a.length >= 16) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1518 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1519 auto n = aptr + (a.length & ~15); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1520 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1521 // Unaligned case |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1522 asm |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1523 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1524 mov EAX, bptr; // left operand |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1525 mov ECX, cptr; // right operand |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1526 mov ESI, aptr; // destination operand |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1527 mov EDI, n; // end comparison |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1528 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1529 align 8; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1530 startsseloopb: |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1531 movups XMM0, [EAX]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1532 movups XMM1, [EAX+16]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1533 movups XMM2, [EAX+32]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1534 movups XMM3, [EAX+48]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1535 add ESI, 64; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1536 movups XMM4, [ECX]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1537 movups XMM5, [ECX+16]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1538 movups XMM6, [ECX+32]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1539 movups XMM7, [ECX+48]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1540 add EAX, 64; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1541 mulps XMM0, XMM4; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1542 mulps XMM1, XMM5; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1543 mulps XMM2, XMM6; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1544 mulps XMM3, XMM7; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1545 add ECX, 64; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1546 movups [ESI+ 0-64], XMM0; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1547 movups [ESI+16-64], XMM1; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1548 movups [ESI+32-64], XMM2; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1549 movups [ESI+48-64], XMM3; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1550 cmp ESI, EDI; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1551 jb startsseloopb; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1552 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1553 mov aptr, ESI; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1554 mov bptr, EAX; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1555 mov cptr, ECX; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1556 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1557 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1558 else |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1559 // 3DNow! version is only 13% faster |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1560 if (amd3dnow() && a.length >= 8) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1561 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1562 auto n = aptr + (a.length & ~7); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1563 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1564 asm |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1565 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1566 mov ESI, dword ptr [aptr]; // destination operand |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1567 mov EDI, dword ptr [n]; // end comparison |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1568 mov EAX, dword ptr [bptr]; // left operand |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1569 mov ECX, dword ptr [cptr]; // right operand |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1570 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1571 align 4; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1572 start: |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1573 movq MM0, [EAX]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1574 movq MM1, [EAX+8]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1575 movq MM2, [EAX+16]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1576 movq MM3, [EAX+24]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1577 pfmul MM0, [ECX]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1578 pfmul MM1, [ECX+8]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1579 pfmul MM2, [ECX+16]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1580 pfmul MM3, [ECX+24]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1581 movq [ESI], MM0; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1582 movq [ESI+8], MM1; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1583 movq [ESI+16], MM2; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1584 movq [ESI+24], MM3; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1585 add ECX, 32; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1586 add ESI, 32; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1587 add EAX, 32; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1588 cmp ESI, EDI; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1589 jb start; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1590 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1591 emms; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1592 mov aptr, ESI; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1593 mov bptr, EAX; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1594 mov cptr, ECX; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1595 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1596 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1597 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1598 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1599 while (aptr < aend) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1600 *aptr++ = *bptr++ * *cptr++; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1601 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1602 return a; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1603 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1604 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1605 unittest |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1606 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1607 printf("_arraySliceSliceMulSliceAssign_f unittest\n"); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1608 for (cpuid = 0; cpuid < CPUID_MAX; cpuid++) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1609 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1610 version (log) printf(" cpuid %d\n", cpuid); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1611 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1612 for (int j = 0; j < 2; j++) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1613 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1614 const int dim = 67; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1615 T[] a = new T[dim + j]; // aligned on 16 byte boundary |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1616 a = a[j .. dim + j]; // misalign for second iteration |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1617 T[] b = new T[dim + j]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1618 b = b[j .. dim + j]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1619 T[] c = new T[dim + j]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1620 c = c[j .. dim + j]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1621 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1622 for (int i = 0; i < dim; i++) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1623 { a[i] = cast(T)i; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1624 b[i] = cast(T)(i + 7); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1625 c[i] = cast(T)(i * 2); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1626 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1627 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1628 c[] = a[] * b[]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1629 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1630 for (int i = 0; i < dim; i++) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1631 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1632 if (c[i] != cast(T)(a[i] * b[i])) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1633 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1634 printf("[%d]: %g != %g * %g\n", i, c[i], a[i], b[i]); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1635 assert(0); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1636 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1637 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1638 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1639 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1640 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1641 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1642 /* ======================================================================== */ |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1643 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1644 /*********************** |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1645 * Computes: |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1646 * a[] *= value |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1647 */ |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1648 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1649 T[] _arrayExpSliceMulass_f(T[] a, T value) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1650 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1651 //printf("_arrayExpSliceMulass_f(a.length = %d, value = %Lg)\n", a.length, cast(real)value); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1652 auto aptr = a.ptr; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1653 auto aend = aptr + a.length; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1654 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1655 version (D_InlineAsm_X86) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1656 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1657 // SSE version is 303% faster |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1658 if (sse() && a.length >= 16) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1659 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1660 // align pointer |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1661 auto n = cast(T*)((cast(uint)aptr + 15) & ~15); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1662 while (aptr < n) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1663 *aptr++ *= value; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1664 n = cast(T*)((cast(uint)aend) & ~15); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1665 if (aptr < n) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1666 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1667 // Aligned case |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1668 asm |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1669 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1670 mov ESI, aptr; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1671 mov EDI, n; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1672 movss XMM4, value; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1673 shufps XMM4, XMM4, 0; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1674 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1675 align 8; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1676 startsseloopa: |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1677 movaps XMM0, [ESI]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1678 movaps XMM1, [ESI+16]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1679 movaps XMM2, [ESI+32]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1680 movaps XMM3, [ESI+48]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1681 add ESI, 64; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1682 mulps XMM0, XMM4; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1683 mulps XMM1, XMM4; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1684 mulps XMM2, XMM4; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1685 mulps XMM3, XMM4; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1686 movaps [ESI+ 0-64], XMM0; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1687 movaps [ESI+16-64], XMM1; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1688 movaps [ESI+32-64], XMM2; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1689 movaps [ESI+48-64], XMM3; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1690 cmp ESI, EDI; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1691 jb startsseloopa; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1692 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1693 mov aptr, ESI; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1694 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1695 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1696 else |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1697 // 3DNow! version is 63% faster |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1698 if (amd3dnow() && a.length >= 8) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1699 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1700 auto n = aptr + (a.length & ~7); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1701 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1702 ulong w = *cast(uint *) &value; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1703 ulong v = w | (w << 32L); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1704 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1705 asm |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1706 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1707 mov ESI, dword ptr [aptr]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1708 mov EDI, dword ptr [n]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1709 movq MM4, qword ptr [v]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1710 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1711 align 8; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1712 start: |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1713 movq MM0, [ESI]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1714 movq MM1, [ESI+8]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1715 movq MM2, [ESI+16]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1716 movq MM3, [ESI+24]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1717 pfmul MM0, MM4; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1718 pfmul MM1, MM4; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1719 pfmul MM2, MM4; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1720 pfmul MM3, MM4; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1721 movq [ESI], MM0; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1722 movq [ESI+8], MM1; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1723 movq [ESI+16], MM2; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1724 movq [ESI+24], MM3; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1725 add ESI, 32; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1726 cmp ESI, EDI; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1727 jb start; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1728 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1729 emms; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1730 mov dword ptr [aptr], ESI; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1731 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1732 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1733 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1734 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1735 while (aptr < aend) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1736 *aptr++ *= value; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1737 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1738 return a; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1739 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1740 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1741 unittest |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1742 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1743 printf("_arrayExpSliceMulass_f unittest\n"); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1744 for (cpuid = 0; cpuid < CPUID_MAX; cpuid++) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1745 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1746 version (log) printf(" cpuid %d\n", cpuid); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1747 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1748 for (int j = 0; j < 2; j++) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1749 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1750 const int dim = 67; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1751 T[] a = new T[dim + j]; // aligned on 16 byte boundary |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1752 a = a[j .. dim + j]; // misalign for second iteration |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1753 T[] b = new T[dim + j]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1754 b = b[j .. dim + j]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1755 T[] c = new T[dim + j]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1756 c = c[j .. dim + j]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1757 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1758 for (int i = 0; i < dim; i++) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1759 { a[i] = cast(T)i; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1760 b[i] = cast(T)(i + 7); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1761 c[i] = cast(T)(i * 2); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1762 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1763 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1764 a[] = c[]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1765 c[] *= 6; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1766 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1767 for (int i = 0; i < dim; i++) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1768 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1769 if (c[i] != cast(T)(a[i] * 6)) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1770 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1771 printf("[%d]: %g != %g * 6\n", i, c[i], a[i]); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1772 assert(0); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1773 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1774 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1775 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1776 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1777 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1778 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1779 /* ======================================================================== */ |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1780 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1781 /*********************** |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1782 * Computes: |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1783 * a[] *= b[] |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1784 */ |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1785 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1786 T[] _arraySliceSliceMulass_f(T[] a, T[] b) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1787 in |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1788 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1789 assert (a.length == b.length); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1790 assert (disjoint(a, b)); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1791 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1792 body |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1793 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1794 //printf("_arraySliceSliceMulass_f()\n"); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1795 auto aptr = a.ptr; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1796 auto aend = aptr + a.length; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1797 auto bptr = b.ptr; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1798 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1799 version (D_InlineAsm_X86) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1800 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1801 // SSE version is 525% faster |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1802 if (sse() && a.length >= 16) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1803 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1804 auto n = aptr + (a.length & ~15); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1805 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1806 // Unaligned case |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1807 asm |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1808 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1809 mov ECX, bptr; // right operand |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1810 mov ESI, aptr; // destination operand |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1811 mov EDI, n; // end comparison |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1812 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1813 align 8; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1814 startsseloopb: |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1815 movups XMM0, [ESI]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1816 movups XMM1, [ESI+16]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1817 movups XMM2, [ESI+32]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1818 movups XMM3, [ESI+48]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1819 add ESI, 64; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1820 movups XMM4, [ECX]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1821 movups XMM5, [ECX+16]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1822 movups XMM6, [ECX+32]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1823 movups XMM7, [ECX+48]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1824 add ECX, 64; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1825 mulps XMM0, XMM4; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1826 mulps XMM1, XMM5; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1827 mulps XMM2, XMM6; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1828 mulps XMM3, XMM7; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1829 movups [ESI+ 0-64], XMM0; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1830 movups [ESI+16-64], XMM1; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1831 movups [ESI+32-64], XMM2; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1832 movups [ESI+48-64], XMM3; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1833 cmp ESI, EDI; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1834 jb startsseloopb; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1835 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1836 mov aptr, ESI; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1837 mov bptr, ECX; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1838 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1839 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1840 else |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1841 // 3DNow! version is 57% faster |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1842 if (amd3dnow() && a.length >= 8) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1843 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1844 auto n = aptr + (a.length & ~7); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1845 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1846 asm |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1847 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1848 mov ESI, dword ptr [aptr]; // destination operand |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1849 mov EDI, dword ptr [n]; // end comparison |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1850 mov ECX, dword ptr [bptr]; // right operand |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1851 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1852 align 4; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1853 start: |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1854 movq MM0, [ESI]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1855 movq MM1, [ESI+8]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1856 movq MM2, [ESI+16]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1857 movq MM3, [ESI+24]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1858 pfmul MM0, [ECX]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1859 pfmul MM1, [ECX+8]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1860 pfmul MM2, [ECX+16]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1861 pfmul MM3, [ECX+24]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1862 movq [ESI], MM0; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1863 movq [ESI+8], MM1; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1864 movq [ESI+16], MM2; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1865 movq [ESI+24], MM3; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1866 add ESI, 32; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1867 add ECX, 32; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1868 cmp ESI, EDI; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1869 jb start; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1870 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1871 emms; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1872 mov dword ptr [aptr], ESI; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1873 mov dword ptr [bptr], ECX; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1874 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1875 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1876 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1877 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1878 while (aptr < aend) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1879 *aptr++ *= *bptr++; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1880 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1881 return a; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1882 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1883 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1884 unittest |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1885 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1886 printf("_arrayExpSliceMulass_f unittest\n"); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1887 for (cpuid = 0; cpuid < CPUID_MAX; cpuid++) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1888 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1889 version (log) printf(" cpuid %d\n", cpuid); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1890 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1891 for (int j = 0; j < 2; j++) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1892 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1893 const int dim = 67; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1894 T[] a = new T[dim + j]; // aligned on 16 byte boundary |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1895 a = a[j .. dim + j]; // misalign for second iteration |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1896 T[] b = new T[dim + j]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1897 b = b[j .. dim + j]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1898 T[] c = new T[dim + j]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1899 c = c[j .. dim + j]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1900 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1901 for (int i = 0; i < dim; i++) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1902 { a[i] = cast(T)i; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1903 b[i] = cast(T)(i + 7); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1904 c[i] = cast(T)(i * 2); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1905 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1906 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1907 a[] = c[]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1908 c[] *= 6; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1909 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1910 for (int i = 0; i < dim; i++) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1911 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1912 if (c[i] != cast(T)(a[i] * 6)) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1913 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1914 printf("[%d]: %g != %g * 6\n", i, c[i], a[i]); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1915 assert(0); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1916 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1917 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1918 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1919 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1920 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1921 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1922 /* ======================================================================== */ |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1923 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1924 /*********************** |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1925 * Computes: |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1926 * a[] = b[] / value |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1927 */ |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1928 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1929 T[] _arraySliceExpDivSliceAssign_f(T[] a, T value, T[] b) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1930 in |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1931 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1932 assert(a.length == b.length); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1933 assert(disjoint(a, b)); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1934 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1935 body |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1936 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1937 //printf("_arraySliceExpDivSliceAssign_f()\n"); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1938 auto aptr = a.ptr; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1939 auto aend = aptr + a.length; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1940 auto bptr = b.ptr; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1941 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1942 /* Multiplying by the reciprocal is faster, but does |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1943 * not produce as accurate an answer. |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1944 */ |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1945 T recip = cast(T)1 / value; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1946 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1947 version (D_InlineAsm_X86) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1948 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1949 // SSE version is 587% faster |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1950 if (sse() && a.length >= 16) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1951 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1952 auto n = aptr + (a.length & ~15); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1953 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1954 // Unaligned case |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1955 asm |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1956 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1957 mov EAX, bptr; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1958 mov ESI, aptr; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1959 mov EDI, n; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1960 movss XMM4, recip; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1961 //movss XMM4, value |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1962 //rcpss XMM4, XMM4 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1963 shufps XMM4, XMM4, 0; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1964 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1965 align 8; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1966 startsseloop: |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1967 add ESI, 64; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1968 movups XMM0, [EAX]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1969 movups XMM1, [EAX+16]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1970 movups XMM2, [EAX+32]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1971 movups XMM3, [EAX+48]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1972 add EAX, 64; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1973 mulps XMM0, XMM4; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1974 mulps XMM1, XMM4; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1975 mulps XMM2, XMM4; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1976 mulps XMM3, XMM4; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1977 //divps XMM0, XMM4; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1978 //divps XMM1, XMM4; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1979 //divps XMM2, XMM4; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1980 //divps XMM3, XMM4; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1981 movups [ESI+ 0-64], XMM0; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1982 movups [ESI+16-64], XMM1; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1983 movups [ESI+32-64], XMM2; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1984 movups [ESI+48-64], XMM3; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1985 cmp ESI, EDI; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1986 jb startsseloop; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1987 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1988 mov aptr, ESI; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1989 mov bptr, EAX; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1990 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1991 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1992 else |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1993 // 3DNow! version is 72% faster |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1994 if (amd3dnow() && a.length >= 8) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1995 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1996 auto n = aptr + (a.length & ~7); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1997 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1998 T[2] w = void; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1999 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2000 w[0] = recip; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2001 w[1] = recip; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2002 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2003 asm |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2004 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2005 mov ESI, dword ptr [aptr]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2006 mov EDI, dword ptr [n]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2007 mov EAX, dword ptr [bptr]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2008 movq MM4, qword ptr [w]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2009 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2010 align 8; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2011 start: |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2012 movq MM0, [EAX]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2013 movq MM1, [EAX+8]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2014 movq MM2, [EAX+16]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2015 movq MM3, [EAX+24]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2016 pfmul MM0, MM4; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2017 pfmul MM1, MM4; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2018 pfmul MM2, MM4; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2019 pfmul MM3, MM4; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2020 movq [ESI], MM0; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2021 movq [ESI+8], MM1; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2022 movq [ESI+16], MM2; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2023 movq [ESI+24], MM3; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2024 add ESI, 32; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2025 add EAX, 32; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2026 cmp ESI, EDI; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2027 jb start; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2028 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2029 emms; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2030 mov dword ptr [aptr], ESI; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2031 mov dword ptr [bptr], EAX; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2032 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2033 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2034 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2035 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2036 while (aptr < aend) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2037 *aptr++ = *bptr++ * recip; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2038 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2039 return a; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2040 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2041 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2042 unittest |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2043 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2044 printf("_arraySliceExpDivSliceAssign_f unittest\n"); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2045 for (cpuid = 0; cpuid < CPUID_MAX; cpuid++) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2046 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2047 version (log) printf(" cpuid %d\n", cpuid); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2048 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2049 for (int j = 0; j < 2; j++) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2050 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2051 const int dim = 67; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2052 T[] a = new T[dim + j]; // aligned on 16 byte boundary |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2053 a = a[j .. dim + j]; // misalign for second iteration |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2054 T[] b = new T[dim + j]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2055 b = b[j .. dim + j]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2056 T[] c = new T[dim + j]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2057 c = c[j .. dim + j]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2058 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2059 for (int i = 0; i < dim; i++) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2060 { a[i] = cast(T)i; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2061 b[i] = cast(T)(i + 7); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2062 c[i] = cast(T)(i * 2); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2063 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2064 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2065 c[] = a[] / 8; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2066 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2067 for (int i = 0; i < dim; i++) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2068 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2069 if (c[i] != cast(T)(a[i] / 8)) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2070 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2071 printf("[%d]: %g != %g / 8\n", i, c[i], a[i]); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2072 assert(0); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2073 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2074 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2075 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2076 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2077 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2078 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2079 /* ======================================================================== */ |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2080 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2081 /*********************** |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2082 * Computes: |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2083 * a[] /= value |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2084 */ |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2085 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2086 T[] _arrayExpSliceDivass_f(T[] a, T value) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2087 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2088 //printf("_arrayExpSliceDivass_f(a.length = %d, value = %Lg)\n", a.length, cast(real)value); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2089 auto aptr = a.ptr; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2090 auto aend = aptr + a.length; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2091 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2092 /* Multiplying by the reciprocal is faster, but does |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2093 * not produce as accurate an answer. |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2094 */ |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2095 T recip = cast(T)1 / value; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2096 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2097 version (D_InlineAsm_X86) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2098 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2099 // SSE version is 245% faster |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2100 if (sse() && a.length >= 16) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2101 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2102 // align pointer |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2103 auto n = cast(T*)((cast(uint)aptr + 15) & ~15); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2104 while (aptr < n) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2105 *aptr++ *= recip; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2106 n = cast(T*)((cast(uint)aend) & ~15); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2107 if (aptr < n) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2108 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2109 // Aligned case |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2110 asm |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2111 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2112 mov ESI, aptr; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2113 mov EDI, n; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2114 movss XMM4, recip; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2115 //movss XMM4, value |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2116 //rcpss XMM4, XMM4 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2117 shufps XMM4, XMM4, 0; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2118 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2119 align 8; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2120 startsseloopa: |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2121 movaps XMM0, [ESI]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2122 movaps XMM1, [ESI+16]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2123 movaps XMM2, [ESI+32]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2124 movaps XMM3, [ESI+48]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2125 add ESI, 64; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2126 mulps XMM0, XMM4; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2127 mulps XMM1, XMM4; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2128 mulps XMM2, XMM4; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2129 mulps XMM3, XMM4; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2130 //divps XMM0, XMM4; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2131 //divps XMM1, XMM4; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2132 //divps XMM2, XMM4; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2133 //divps XMM3, XMM4; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2134 movaps [ESI+ 0-64], XMM0; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2135 movaps [ESI+16-64], XMM1; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2136 movaps [ESI+32-64], XMM2; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2137 movaps [ESI+48-64], XMM3; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2138 cmp ESI, EDI; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2139 jb startsseloopa; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2140 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2141 mov aptr, ESI; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2142 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2143 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2144 else |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2145 // 3DNow! version is 57% faster |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2146 if (amd3dnow() && a.length >= 8) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2147 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2148 auto n = aptr + (a.length & ~7); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2149 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2150 T[2] w = void; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2151 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2152 w[0] = w[1] = recip; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2153 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2154 asm |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2155 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2156 mov ESI, dword ptr [aptr]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2157 mov EDI, dword ptr [n]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2158 movq MM4, qword ptr [w]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2159 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2160 align 8; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2161 start: |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2162 movq MM0, [ESI]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2163 movq MM1, [ESI+8]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2164 movq MM2, [ESI+16]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2165 movq MM3, [ESI+24]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2166 pfmul MM0, MM4; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2167 pfmul MM1, MM4; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2168 pfmul MM2, MM4; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2169 pfmul MM3, MM4; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2170 movq [ESI], MM0; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2171 movq [ESI+8], MM1; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2172 movq [ESI+16], MM2; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2173 movq [ESI+24], MM3; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2174 add ESI, 32; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2175 cmp ESI, EDI; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2176 jb start; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2177 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2178 emms; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2179 mov dword ptr [aptr], ESI; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2180 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2181 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2182 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2183 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2184 while (aptr < aend) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2185 *aptr++ *= recip; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2186 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2187 return a; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2188 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2189 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2190 unittest |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2191 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2192 printf("_arrayExpSliceDivass_f unittest\n"); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2193 for (cpuid = 0; cpuid < CPUID_MAX; cpuid++) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2194 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2195 version (log) printf(" cpuid %d\n", cpuid); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2196 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2197 for (int j = 0; j < 2; j++) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2198 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2199 const int dim = 67; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2200 T[] a = new T[dim + j]; // aligned on 16 byte boundary |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2201 a = a[j .. dim + j]; // misalign for second iteration |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2202 T[] b = new T[dim + j]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2203 b = b[j .. dim + j]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2204 T[] c = new T[dim + j]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2205 c = c[j .. dim + j]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2206 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2207 for (int i = 0; i < dim; i++) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2208 { a[i] = cast(T)i; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2209 b[i] = cast(T)(i + 7); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2210 c[i] = cast(T)(i * 2); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2211 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2212 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2213 a[] = c[]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2214 c[] /= 8; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2215 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2216 for (int i = 0; i < dim; i++) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2217 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2218 if (c[i] != cast(T)(a[i] / 8)) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2219 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2220 printf("[%d]: %g != %g / 8\n", i, c[i], a[i]); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2221 assert(0); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2222 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2223 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2224 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2225 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2226 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2227 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2228 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2229 /* ======================================================================== */ |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2230 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2231 /*********************** |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2232 * Computes: |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2233 * a[] -= b[] * value |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2234 */ |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2235 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2236 T[] _arraySliceExpMulSliceMinass_f(T[] a, T value, T[] b) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2237 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2238 return _arraySliceExpMulSliceAddass_f(a, -value, b); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2239 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2240 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2241 /*********************** |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2242 * Computes: |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2243 * a[] += b[] * value |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2244 */ |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2245 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2246 T[] _arraySliceExpMulSliceAddass_f(T[] a, T value, T[] b) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2247 in |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2248 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2249 assert(a.length == b.length); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2250 assert(disjoint(a, b)); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2251 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2252 body |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2253 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2254 auto aptr = a.ptr; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2255 auto aend = aptr + a.length; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2256 auto bptr = b.ptr; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2257 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2258 // Handle remainder |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2259 while (aptr < aend) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2260 *aptr++ += *bptr++ * value; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2261 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2262 return a; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2263 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2264 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2265 unittest |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2266 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2267 printf("_arraySliceExpMulSliceAddass_f unittest\n"); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2268 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2269 cpuid = 1; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2270 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2271 version (log) printf(" cpuid %d\n", cpuid); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2272 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2273 for (int j = 0; j < 1; j++) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2274 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2275 const int dim = 67; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2276 T[] a = new T[dim + j]; // aligned on 16 byte boundary |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2277 a = a[j .. dim + j]; // misalign for second iteration |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2278 T[] b = new T[dim + j]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2279 b = b[j .. dim + j]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2280 T[] c = new T[dim + j]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2281 c = c[j .. dim + j]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2282 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2283 for (int i = 0; i < dim; i++) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2284 { a[i] = cast(T)i; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2285 b[i] = cast(T)(i + 7); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2286 c[i] = cast(T)(i * 2); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2287 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2288 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2289 b[] = c[]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2290 c[] += a[] * 6; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2291 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2292 for (int i = 0; i < dim; i++) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2293 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2294 //printf("[%d]: %g ?= %g + %g * 6\n", i, c[i], b[i], a[i]); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2295 if (c[i] != cast(T)(b[i] + a[i] * 6)) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2296 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2297 printf("[%d]: %g ?= %g + %g * 6\n", i, c[i], b[i], a[i]); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2298 assert(0); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2299 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2300 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2301 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2302 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2303 } |