Mercurial > projects > ldc
annotate druntime/src/compiler/dmd/arrayshort.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 byte 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 /* Contains SSE2 and MMX versions of certain operations for wchar, short, |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
9 * and ushort ('u', 's' and 't' suffixes). |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
10 */ |
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 module rt.arrayshort; |
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 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
|
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 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
|
17 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
18 /* 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
|
19 */ |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
20 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
|
21 const int CPUID_MAX = 4; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
22 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
|
23 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
|
24 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
|
25 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
|
26 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
27 else |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
28 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
29 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
|
30 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
|
31 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
|
32 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
|
33 } |
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 //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
|
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 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
|
38 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
39 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
|
40 } |
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 alias short T; |
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 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
|
45 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
46 /* ======================================================================== */ |
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 * Computes: |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
50 * 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
|
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 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
53 T[] _arraySliceExpAddSliceAssign_u(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
|
54 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
55 return _arraySliceExpAddSliceAssign_s(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
|
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 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
58 T[] _arraySliceExpAddSliceAssign_t(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
|
59 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
60 return _arraySliceExpAddSliceAssign_s(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
|
61 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
62 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
63 T[] _arraySliceExpAddSliceAssign_s(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
|
64 in |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
65 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
66 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
|
67 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
|
68 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
69 body |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
70 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
71 //printf("_arraySliceExpAddSliceAssign_s()\n"); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
72 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
|
73 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
|
74 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
|
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 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
|
77 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
78 // SSE2 aligned version is 3343% faster |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
79 if (sse2() && 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
|
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 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
|
82 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
83 uint l = cast(ushort) value; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
84 l |= (l << 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 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
86 if (((cast(uint) aptr | cast(uint) bptr) & 15) != 0) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
87 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
88 asm // 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
|
89 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
90 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
|
91 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
|
92 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
|
93 movd XMM2, l; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
94 pshufd XMM2, XMM2, 0; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
95 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
96 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
|
97 startaddsse2u: |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
98 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
|
99 movdqu 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
|
100 movdqu 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
|
101 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
|
102 paddw XMM0, XMM2; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
103 paddw XMM1, XMM2; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
104 movdqu [ESI -32], XMM0; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
105 movdqu [ESI+16-32], XMM1; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
106 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
|
107 jb startaddsse2u; |
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 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
|
110 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
|
111 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
112 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
113 else |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
114 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
115 asm // 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
|
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 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
|
118 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
|
119 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
|
120 movd XMM2, l; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
121 pshufd XMM2, XMM2, 0; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
122 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
123 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
|
124 startaddsse2a: |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
125 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
|
126 movdqa 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
|
127 movdqa 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
|
128 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
|
129 paddw XMM0, XMM2; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
130 paddw XMM1, XMM2; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
131 movdqa [ESI -32], XMM0; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
132 movdqa [ESI+16-32], XMM1; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
133 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
|
134 jb startaddsse2a; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
135 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
136 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
|
137 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
|
138 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
139 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
140 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
141 else |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
142 // MMX version is 3343% faster |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
143 if (mmx() && 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
|
144 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
145 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
|
146 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
147 uint l = cast(ushort) value; |
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 asm |
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 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
|
152 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
|
153 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
|
154 movd MM2, l; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
155 pshufw MM2, MM2, 0; |
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 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
|
158 startmmx: |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
159 add 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
|
160 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
|
161 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
|
162 add 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
|
163 paddw MM0, MM2; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
164 paddw MM1, MM2; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
165 movq [ESI -16], MM0; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
166 movq [ESI+8-16], MM1; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
167 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
|
168 jb startmmx; |
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 emms; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
171 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
|
172 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
|
173 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
174 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
175 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
176 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
177 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
|
178 *aptr++ = cast(T)(*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
|
179 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
180 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
|
181 } |
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 unittest |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
184 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
185 printf("_arraySliceExpAddSliceAssign_s 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
|
186 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
187 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
|
188 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
189 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
|
190 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
191 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
|
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 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
|
194 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
|
195 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
|
196 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
|
197 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
|
198 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
|
199 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
|
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 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
|
202 { 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
|
203 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
|
204 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
|
205 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
206 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
207 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
|
208 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
209 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
|
210 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
211 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
|
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 printf("[%d]: %d != %d + 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
|
214 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
|
215 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
216 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
217 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
218 } |
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 |
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 /* ======================================================================== */ |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
223 |
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 * Computes: |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
226 * 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
|
227 */ |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
228 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
229 T[] _arraySliceSliceAddSliceAssign_u(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
|
230 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
231 return _arraySliceSliceAddSliceAssign_s(a, 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
|
232 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
233 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
234 T[] _arraySliceSliceAddSliceAssign_t(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
|
235 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
236 return _arraySliceSliceAddSliceAssign_s(a, 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
|
237 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
238 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
239 T[] _arraySliceSliceAddSliceAssign_s(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
|
240 in |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
241 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
242 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
|
243 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
|
244 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
|
245 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
|
246 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
247 body |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
248 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
249 //printf("_arraySliceSliceAddSliceAssign_s()\n"); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
250 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
|
251 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
|
252 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
|
253 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
|
254 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
255 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
|
256 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
257 // SSE2 aligned version is 3777% faster |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
258 if (sse2() && 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
|
259 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
260 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
|
261 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
262 if (((cast(uint) aptr | cast(uint) bptr | cast(uint) cptr) & 15) != 0) |
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 asm // 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
|
265 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
266 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
|
267 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
|
268 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
|
269 mov ECX, cptr; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
270 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
271 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
|
272 startsse2u: |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
273 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
|
274 movdqu 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
|
275 movdqu 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
|
276 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
|
277 movdqu XMM2, [ECX]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
278 movdqu XMM3, [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
|
279 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
|
280 paddw XMM0, XMM2; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
281 paddw XMM1, XMM3; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
282 movdqu [ESI -32], XMM0; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
283 movdqu [ESI+16-32], XMM1; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
284 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
|
285 jb startsse2u; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
286 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
287 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
|
288 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
|
289 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
|
290 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
291 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
292 else |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
293 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
294 asm // 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
|
295 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
296 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
|
297 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
|
298 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
|
299 mov ECX, cptr; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
300 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
301 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
|
302 startsse2a: |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
303 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
|
304 movdqa 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
|
305 movdqa 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
|
306 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
|
307 movdqa XMM2, [ECX]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
308 movdqa XMM3, [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
|
309 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
|
310 paddw XMM0, XMM2; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
311 paddw XMM1, XMM3; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
312 movdqa [ESI -32], XMM0; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
313 movdqa [ESI+16-32], XMM1; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
314 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
|
315 jb startsse2a; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
316 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
317 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
|
318 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
|
319 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
|
320 } |
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 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
323 else |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
324 // MMX version is 2068% faster |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
325 if (mmx() && 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
|
326 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
327 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
|
328 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
329 asm |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
330 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
331 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
|
332 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
|
333 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
|
334 mov ECX, cptr; |
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 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
|
337 startmmx: |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
338 add 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
|
339 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
|
340 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
|
341 add 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
|
342 movq MM2, [ECX]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
343 movq MM3, [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
|
344 add 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
|
345 paddw MM0, MM2; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
346 paddw MM1, MM3; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
347 movq [ESI -16], MM0; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
348 movq [ESI+8-16], MM1; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
349 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
|
350 jb startmmx; |
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 emms; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
353 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
|
354 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
|
355 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
|
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 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
358 } |
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 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
|
361 *aptr++ = cast(T)(*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
|
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 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
|
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 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
366 unittest |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
367 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
368 printf("_arraySliceSliceAddSliceAssign_s 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
|
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 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
|
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 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
|
373 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
374 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
|
375 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
376 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
|
377 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
|
378 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
|
379 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
|
380 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
|
381 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
|
382 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
|
383 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
384 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
|
385 { 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
|
386 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
|
387 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
|
388 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
389 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
390 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
|
391 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
392 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
|
393 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
394 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
|
395 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
396 printf("[%d]: %d != %d + %d\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
|
397 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
|
398 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
399 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
400 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
401 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
402 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
403 |
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 /* ======================================================================== */ |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
406 |
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 * Computes: |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
409 * 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
|
410 */ |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
411 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
412 T[] _arrayExpSliceAddass_u(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
|
413 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
414 return _arrayExpSliceAddass_s(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
|
415 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
416 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
417 T[] _arrayExpSliceAddass_t(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
|
418 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
419 return _arrayExpSliceAddass_s(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
|
420 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
421 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
422 T[] _arrayExpSliceAddass_s(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
|
423 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
424 //printf("_arrayExpSliceAddass_s(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
|
425 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
|
426 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
|
427 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
428 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
|
429 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
430 // SSE2 aligned version is 832% faster |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
431 if (sse2() && 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
|
432 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
433 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
|
434 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
435 uint l = cast(ushort) value; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
436 l |= (l << 16); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
437 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
438 if (((cast(uint) aptr) & 15) != 0) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
439 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
440 asm // 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
|
441 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
442 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
|
443 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
|
444 movd XMM2, l; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
445 pshufd XMM2, XMM2, 0; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
446 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
447 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
|
448 startaddsse2u: |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
449 movdqu 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
|
450 movdqu 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
|
451 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
|
452 paddw XMM0, XMM2; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
453 paddw XMM1, XMM2; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
454 movdqu [ESI -32], XMM0; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
455 movdqu [ESI+16-32], XMM1; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
456 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
|
457 jb startaddsse2u; |
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 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
|
460 } |
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 else |
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 asm // 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
|
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 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
|
467 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
|
468 movd XMM2, l; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
469 pshufd XMM2, XMM2, 0; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
470 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
471 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
|
472 startaddsse2a: |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
473 movdqa 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
|
474 movdqa 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
|
475 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
|
476 paddw XMM0, XMM2; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
477 paddw XMM1, XMM2; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
478 movdqa [ESI -32], XMM0; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
479 movdqa [ESI+16-32], XMM1; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
480 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
|
481 jb startaddsse2a; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
482 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
483 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
|
484 } |
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 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
487 else |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
488 // MMX version is 826% faster |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
489 if (mmx() && 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
|
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 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
|
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 uint l = cast(ushort) value; |
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 asm |
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 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
|
498 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
|
499 movd MM2, l; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
500 pshufw MM2, MM2, 0; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
501 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
502 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
|
503 startmmx: |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
504 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
|
505 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
|
506 add 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
|
507 paddw MM0, MM2; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
508 paddw MM1, MM2; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
509 movq [ESI -16], MM0; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
510 movq [ESI+8-16], MM1; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
511 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
|
512 jb startmmx; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
513 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
514 emms; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
515 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
|
516 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
517 } |
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 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
520 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
|
521 *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
|
522 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
523 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
|
524 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
525 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
526 unittest |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
527 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
528 printf("_arrayExpSliceAddass_s 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
|
529 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
530 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
|
531 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
532 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
|
533 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
534 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
|
535 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
536 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
|
537 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
|
538 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
|
539 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
|
540 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
|
541 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
|
542 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
|
543 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
544 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
|
545 { 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
|
546 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
|
547 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
|
548 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
549 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
550 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
|
551 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
|
552 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
553 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
|
554 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
555 if (a[i] != cast(T)(c[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
|
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 printf("[%d]: %d != %d + 6\n", i, a[i], c[i]); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
558 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
|
559 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
560 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
561 } |
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 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
564 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
565 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
566 /* ======================================================================== */ |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
567 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
568 /*********************** |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
569 * Computes: |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
570 * 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
|
571 */ |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
572 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
573 T[] _arraySliceSliceAddass_u(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
|
574 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
575 return _arraySliceSliceAddass_s(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
|
576 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
577 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
578 T[] _arraySliceSliceAddass_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
|
579 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
580 return _arraySliceSliceAddass_s(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
|
581 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
582 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
583 T[] _arraySliceSliceAddass_s(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
|
584 in |
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 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
|
587 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
|
588 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
589 body |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
590 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
591 //printf("_arraySliceSliceAddass_s()\n"); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
592 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
|
593 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
|
594 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
|
595 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
596 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
|
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 // SSE2 aligned version is 2085% faster |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
599 if (sse2() && 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
|
600 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
601 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
|
602 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
603 if (((cast(uint) aptr | cast(uint) bptr) & 15) != 0) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
604 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
605 asm // 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
|
606 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
607 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
|
608 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
|
609 mov ECX, bptr; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
610 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
611 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
|
612 startsse2u: |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
613 movdqu 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
|
614 movdqu 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
|
615 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
|
616 movdqu XMM2, [ECX]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
617 movdqu XMM3, [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
|
618 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
|
619 paddw XMM0, XMM2; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
620 paddw XMM1, XMM3; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
621 movdqu [ESI -32], XMM0; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
622 movdqu [ESI+16-32], XMM1; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
623 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
|
624 jb startsse2u; |
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 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
|
627 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
|
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 else |
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 asm // 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
|
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 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
|
635 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
|
636 mov ECX, bptr; |
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 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
|
639 startsse2a: |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
640 movdqa 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
|
641 movdqa 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
|
642 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
|
643 movdqa XMM2, [ECX]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
644 movdqa XMM3, [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
|
645 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
|
646 paddw XMM0, XMM2; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
647 paddw XMM1, XMM3; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
648 movdqa [ESI -32], XMM0; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
649 movdqa [ESI+16-32], XMM1; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
650 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
|
651 jb startsse2a; |
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 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
|
654 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
|
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 } |
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 else |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
659 // MMX version is 1022% faster |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
660 if (mmx() && 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
|
661 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
662 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
|
663 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
664 asm |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
665 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
666 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
|
667 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
|
668 mov ECX, bptr; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
669 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
670 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
|
671 start: |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
672 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
|
673 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
|
674 add 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
|
675 movq MM2, [ECX]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
676 movq MM3, [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
|
677 add 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
|
678 paddw MM0, MM2; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
679 paddw MM1, MM3; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
680 movq [ESI -16], MM0; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
681 movq [ESI+8-16], MM1; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
682 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
|
683 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
|
684 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
685 emms; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
686 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
|
687 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
|
688 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
689 } |
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 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
|
693 *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
|
694 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
695 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
|
696 } |
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 unittest |
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 printf("_arraySliceSliceAddass_s 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
|
701 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
702 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
|
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 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
|
705 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
706 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
|
707 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
708 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
|
709 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
|
710 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
|
711 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
|
712 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
|
713 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
|
714 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
|
715 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
716 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
|
717 { 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
|
718 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
|
719 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
|
720 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
721 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
722 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
|
723 c[] += a[]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
724 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
725 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
|
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 if (c[i] != cast(T)(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
|
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 printf("[%d]: %d != %d + %d\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
|
730 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
|
731 } |
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 } |
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 |
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 /* ======================================================================== */ |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
739 |
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 * Computes: |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
742 * 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
|
743 */ |
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 T[] _arraySliceExpMinSliceAssign_u(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
|
746 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
747 return _arraySliceExpMinSliceAssign_s(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
|
748 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
749 |
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[] _arraySliceExpMinSliceAssign_t(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
|
751 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
752 return _arraySliceExpMinSliceAssign_s(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
|
753 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
754 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
755 T[] _arraySliceExpMinSliceAssign_s(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
|
756 in |
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 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
|
759 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
|
760 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
761 body |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
762 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
763 //printf("_arraySliceExpMinSliceAssign_s()\n"); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
764 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
|
765 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
|
766 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
|
767 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
768 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
|
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 // SSE2 aligned version is 3695% faster |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
771 if (sse2() && 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
|
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 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
|
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 uint l = cast(ushort) value; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
776 l |= (l << 16); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
777 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
778 if (((cast(uint) aptr | cast(uint) bptr) & 15) != 0) |
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 asm // 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
|
781 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
782 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
|
783 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
|
784 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
|
785 movd XMM2, l; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
786 pshufd XMM2, XMM2, 0; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
787 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
788 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
|
789 startaddsse2u: |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
790 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
|
791 movdqu 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
|
792 movdqu 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
|
793 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
|
794 psubw XMM0, XMM2; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
795 psubw XMM1, XMM2; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
796 movdqu [ESI -32], XMM0; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
797 movdqu [ESI+16-32], XMM1; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
798 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
|
799 jb startaddsse2u; |
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 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
|
802 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
|
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 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
805 else |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
806 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
807 asm // 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
|
808 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
809 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
|
810 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
|
811 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
|
812 movd XMM2, l; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
813 pshufd XMM2, XMM2, 0; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
814 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
815 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
|
816 startaddsse2a: |
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 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
|
818 movdqa 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
|
819 movdqa 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
|
820 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
|
821 psubw XMM0, XMM2; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
822 psubw XMM1, XMM2; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
823 movdqa [ESI -32], XMM0; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
824 movdqa [ESI+16-32], XMM1; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
825 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
|
826 jb startaddsse2a; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
827 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
828 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
|
829 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
|
830 } |
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 // MMX version is 3049% 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 (mmx() && 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 uint l = cast(ushort) value; |
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 asm |
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 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
|
844 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
|
845 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
|
846 movd MM2, l; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
847 pshufw MM2, MM2, 0; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
848 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
849 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
|
850 startmmx: |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
851 add 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
|
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 add 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 psubw MM0, MM2; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
856 psubw MM1, MM2; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
857 movq [ESI -16], MM0; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
858 movq [ESI+8-16], MM1; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
859 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
|
860 jb startmmx; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
861 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
862 emms; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
863 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
|
864 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
|
865 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
866 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
867 } |
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 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
|
870 *aptr++ = cast(T)(*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
|
871 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
872 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
|
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 unittest |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
876 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
877 printf("_arraySliceExpMinSliceAssign_s 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
|
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 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
|
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 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
|
882 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
883 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
|
884 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
885 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
|
886 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
|
887 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
|
888 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
|
889 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
|
890 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
|
891 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
|
892 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
893 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
|
894 { 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
|
895 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
|
896 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
|
897 } |
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 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
|
900 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
901 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
|
902 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
903 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
|
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 printf("[%d]: %d != %d - 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
|
906 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
|
907 } |
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 } |
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 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
912 |
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 * Computes: |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
918 * 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
|
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 T[] _arrayExpSliceMinSliceAssign_u(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
|
922 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
923 return _arrayExpSliceMinSliceAssign_s(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
|
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_t(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 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
928 return _arrayExpSliceMinSliceAssign_s(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
|
929 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
930 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
931 T[] _arrayExpSliceMinSliceAssign_s(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
|
932 in |
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 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
|
935 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
|
936 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
937 body |
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 //printf("_arrayExpSliceMinSliceAssign_s()\n"); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
940 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
|
941 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
|
942 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
|
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 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
|
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 // SSE2 aligned version is 4995% faster |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
947 if (sse2() && 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
|
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 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
|
950 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
951 uint l = cast(ushort) value; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
952 l |= (l << 16); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
953 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
954 if (((cast(uint) aptr | cast(uint) bptr) & 15) != 0) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
955 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
956 asm // 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
|
957 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
958 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
|
959 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
|
960 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
|
961 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
962 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
|
963 startaddsse2u: |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
964 movd XMM2, l; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
965 pshufd XMM2, XMM2, 0; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
966 movd XMM3, l; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
967 pshufd XMM3, XMM3, 0; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
968 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
|
969 movdqu 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
|
970 movdqu 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
|
971 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
|
972 psubw XMM2, XMM0; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
973 psubw XMM3, XMM1; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
974 movdqu [ESI -32], XMM2; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
975 movdqu [ESI+16-32], XMM3; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
976 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
|
977 jb startaddsse2u; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
978 |
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 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
|
980 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
|
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 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
983 else |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
984 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
985 asm // 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
|
986 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
987 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
|
988 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
|
989 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
|
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 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
|
992 startaddsse2a: |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
993 movd XMM2, l; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
994 pshufd XMM2, XMM2, 0; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
995 movd XMM3, l; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
996 pshufd XMM3, XMM3, 0; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
997 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
|
998 movdqa 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
|
999 movdqa 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
|
1000 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
|
1001 psubw XMM2, XMM0; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1002 psubw XMM3, XMM1; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1003 movdqa [ESI -32], XMM2; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1004 movdqa [ESI+16-32], XMM3; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1005 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
|
1006 jb startaddsse2a; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1007 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1008 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
|
1009 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
|
1010 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1011 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1012 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1013 else |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1014 // MMX version is 4562% faster |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1015 if (mmx() && 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
|
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 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
|
1018 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1019 uint l = cast(ushort) value; |
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 asm |
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 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
|
1024 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
|
1025 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
|
1026 movd MM4, l; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1027 pshufw MM4, MM4, 0; |
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 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
|
1030 startmmx: |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1031 add 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
|
1032 movq MM2, [EAX]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1033 movq MM3, [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
|
1034 movq 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
|
1035 movq 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
|
1036 add 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
|
1037 psubw MM0, MM2; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1038 psubw MM1, MM3; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1039 movq [ESI -16], MM0; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1040 movq [ESI+8-16], MM1; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1041 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
|
1042 jb startmmx; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1043 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1044 emms; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1045 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
|
1046 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
|
1047 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1048 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1049 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1050 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1051 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
|
1052 *aptr++ = cast(T)(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
|
1053 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1054 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
|
1055 } |
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 unittest |
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("_arrayExpSliceMinSliceAssign_s 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
|
1060 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1061 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
|
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 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
|
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 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
|
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 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
|
1068 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
|
1069 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
|
1070 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
|
1071 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
|
1072 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
|
1073 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
|
1074 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1075 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
|
1076 { 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
|
1077 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
|
1078 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
|
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 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1081 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
|
1082 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1083 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
|
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 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
|
1086 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1087 printf("[%d]: %d != 6 - %d\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
|
1088 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
|
1089 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1090 } |
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 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1093 } |
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 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1096 /* ======================================================================== */ |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1097 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1098 /*********************** |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1099 * Computes: |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1100 * 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
|
1101 */ |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1102 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1103 T[] _arraySliceSliceMinSliceAssign_u(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
|
1104 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1105 return _arraySliceSliceMinSliceAssign_s(a, 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
|
1106 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1107 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1108 T[] _arraySliceSliceMinSliceAssign_t(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
|
1109 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1110 return _arraySliceSliceMinSliceAssign_s(a, 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
|
1111 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1112 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1113 T[] _arraySliceSliceMinSliceAssign_s(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
|
1114 in |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1115 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1116 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
|
1117 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
|
1118 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
|
1119 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
|
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 body |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1122 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1123 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
|
1124 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
|
1125 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
|
1126 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
|
1127 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1128 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
|
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 // SSE2 aligned version is 4129% faster |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1131 if (sse2() && 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
|
1132 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1133 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
|
1134 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1135 if (((cast(uint) aptr | cast(uint) bptr | cast(uint) cptr) & 15) != 0) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1136 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1137 asm // 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
|
1138 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1139 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
|
1140 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
|
1141 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
|
1142 mov ECX, cptr; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1143 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1144 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
|
1145 startsse2u: |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1146 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
|
1147 movdqu 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
|
1148 movdqu 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
|
1149 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
|
1150 movdqu XMM2, [ECX]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1151 movdqu XMM3, [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
|
1152 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
|
1153 psubw XMM0, XMM2; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1154 psubw XMM1, XMM3; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1155 movdqu [ESI -32], XMM0; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1156 movdqu [ESI+16-32], XMM1; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1157 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
|
1158 jb startsse2u; |
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 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
|
1161 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
|
1162 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
|
1163 } |
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 else |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1166 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1167 asm // 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
|
1168 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1169 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
|
1170 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
|
1171 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
|
1172 mov ECX, cptr; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1173 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1174 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
|
1175 startsse2a: |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1176 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
|
1177 movdqa 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
|
1178 movdqa 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
|
1179 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
|
1180 movdqa XMM2, [ECX]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1181 movdqa XMM3, [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
|
1182 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
|
1183 psubw XMM0, XMM2; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1184 psubw XMM1, XMM3; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1185 movdqa [ESI -32], XMM0; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1186 movdqa [ESI+16-32], XMM1; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1187 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
|
1188 jb startsse2a; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1189 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1190 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
|
1191 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
|
1192 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
|
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 } |
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 else |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1197 // MMX version is 2018% faster |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1198 if (mmx() && 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
|
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 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
|
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 asm |
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 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
|
1205 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
|
1206 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
|
1207 mov ECX, cptr; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1208 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1209 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
|
1210 startmmx: |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1211 add 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
|
1212 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
|
1213 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
|
1214 add 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
|
1215 movq MM2, [ECX]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1216 movq MM3, [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
|
1217 add 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
|
1218 psubw MM0, MM2; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1219 psubw MM1, MM3; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1220 movq [ESI -16], MM0; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1221 movq [ESI+8-16], MM1; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1222 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
|
1223 jb startmmx; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1224 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1225 emms; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1226 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
|
1227 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
|
1228 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
|
1229 } |
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 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1232 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1233 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
|
1234 *aptr++ = cast(T)(*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
|
1235 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1236 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
|
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 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1239 unittest |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1240 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1241 printf("_arraySliceSliceMinSliceAssign_s 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
|
1242 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1243 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
|
1244 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1245 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
|
1246 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1247 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
|
1248 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1249 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
|
1250 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
|
1251 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
|
1252 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
|
1253 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
|
1254 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
|
1255 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
|
1256 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1257 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
|
1258 { 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
|
1259 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
|
1260 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
|
1261 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1262 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1263 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
|
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 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
|
1266 { |
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 (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
|
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 printf("[%d]: %d != %d - %d\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
|
1270 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
|
1271 } |
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 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1274 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1275 } |
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 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1278 /* ======================================================================== */ |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1279 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1280 /*********************** |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1281 * Computes: |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1282 * 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
|
1283 */ |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1284 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1285 T[] _arrayExpSliceMinass_u(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
|
1286 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1287 return _arrayExpSliceMinass_s(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
|
1288 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1289 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1290 T[] _arrayExpSliceMinass_t(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
|
1291 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1292 return _arrayExpSliceMinass_s(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
|
1293 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1294 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1295 T[] _arrayExpSliceMinass_s(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
|
1296 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1297 //printf("_arrayExpSliceMinass_s(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
|
1298 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
|
1299 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
|
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 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
|
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 // SSE2 aligned version is 835% faster |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1304 if (sse2() && 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
|
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 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
|
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 uint l = cast(ushort) value; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1309 l |= (l << 16); |
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 if (((cast(uint) aptr) & 15) != 0) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1312 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1313 asm // 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
|
1314 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1315 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
|
1316 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
|
1317 movd XMM2, l; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1318 pshufd XMM2, XMM2, 0; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1319 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1320 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
|
1321 startaddsse2u: |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1322 movdqu 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
|
1323 movdqu 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
|
1324 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
|
1325 psubw XMM0, XMM2; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1326 psubw XMM1, XMM2; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1327 movdqu [ESI -32], XMM0; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1328 movdqu [ESI+16-32], XMM1; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1329 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
|
1330 jb startaddsse2u; |
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 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
|
1333 } |
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 else |
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 asm // 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
|
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 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
|
1340 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
|
1341 movd XMM2, l; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1342 pshufd XMM2, XMM2, 0; |
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 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
|
1345 startaddsse2a: |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1346 movdqa 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
|
1347 movdqa 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
|
1348 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
|
1349 psubw XMM0, XMM2; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1350 psubw XMM1, XMM2; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1351 movdqa [ESI -32], XMM0; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1352 movdqa [ESI+16-32], XMM1; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1353 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
|
1354 jb startaddsse2a; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1355 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1356 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
|
1357 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1358 } |
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 else |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1361 // MMX version is 835% faster |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1362 if (mmx() && 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
|
1363 { |
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 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
|
1365 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1366 uint l = cast(ushort) value; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1367 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1368 asm |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1369 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1370 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
|
1371 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
|
1372 movd MM2, l; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1373 pshufw MM2, MM2, 0; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1374 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1375 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
|
1376 startmmx: |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1377 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
|
1378 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
|
1379 add 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
|
1380 psubw MM0, MM2; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1381 psubw MM1, MM2; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1382 movq [ESI -16], MM0; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1383 movq [ESI+8-16], MM1; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1384 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
|
1385 jb startmmx; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1386 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1387 emms; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1388 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
|
1389 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1390 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1391 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1392 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1393 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
|
1394 *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
|
1395 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1396 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
|
1397 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1398 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1399 unittest |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1400 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1401 printf("_arrayExpSliceMinass_s 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
|
1402 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1403 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
|
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 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
|
1406 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1407 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
|
1408 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1409 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
|
1410 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
|
1411 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
|
1412 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
|
1413 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
|
1414 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
|
1415 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
|
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 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
|
1418 { 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
|
1419 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
|
1420 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
|
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 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1423 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
|
1424 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
|
1425 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1426 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
|
1427 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1428 if (a[i] != cast(T)(c[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
|
1429 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1430 printf("[%d]: %d != %d - 6\n", i, a[i], c[i]); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1431 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
|
1432 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1433 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1434 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1435 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1436 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1437 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1438 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1439 /* ======================================================================== */ |
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 /*********************** |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1442 * Computes: |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1443 * 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
|
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 T[] _arraySliceSliceMinass_u(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
|
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 return _arraySliceSliceMinass_s(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
|
1449 } |
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 T[] _arraySliceSliceMinass_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
|
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 return _arraySliceSliceMinass_s(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
|
1454 } |
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 T[] _arraySliceSliceMinass_s(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
|
1457 in |
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 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
|
1460 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
|
1461 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1462 body |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1463 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1464 //printf("_arraySliceSliceMinass_s()\n"); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1465 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
|
1466 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
|
1467 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
|
1468 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1469 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
|
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 // SSE2 aligned version is 2121% faster |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1472 if (sse2() && 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
|
1473 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1474 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
|
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 if (((cast(uint) aptr | cast(uint) bptr) & 15) != 0) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1477 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1478 asm // 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
|
1479 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1480 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
|
1481 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
|
1482 mov ECX, bptr; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1483 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1484 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
|
1485 startsse2u: |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1486 movdqu 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
|
1487 movdqu 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
|
1488 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
|
1489 movdqu XMM2, [ECX]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1490 movdqu XMM3, [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
|
1491 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
|
1492 psubw XMM0, XMM2; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1493 psubw XMM1, XMM3; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1494 movdqu [ESI -32], XMM0; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1495 movdqu [ESI+16-32], XMM1; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1496 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
|
1497 jb startsse2u; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1498 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1499 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
|
1500 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
|
1501 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1502 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1503 else |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1504 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1505 asm // 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
|
1506 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1507 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
|
1508 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
|
1509 mov ECX, bptr; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1510 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1511 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
|
1512 startsse2a: |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1513 movdqa 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
|
1514 movdqa 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
|
1515 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
|
1516 movdqa XMM2, [ECX]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1517 movdqa XMM3, [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
|
1518 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
|
1519 psubw XMM0, XMM2; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1520 psubw XMM1, XMM3; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1521 movdqa [ESI -32], XMM0; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1522 movdqa [ESI+16-32], XMM1; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1523 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
|
1524 jb startsse2a; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1525 |
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 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
|
1527 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
|
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 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1530 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1531 else |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1532 // MMX version is 1116% faster |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1533 if (mmx() && 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
|
1534 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1535 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
|
1536 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1537 asm |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1538 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1539 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
|
1540 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
|
1541 mov ECX, bptr; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1542 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1543 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
|
1544 start: |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1545 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
|
1546 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
|
1547 add 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
|
1548 movq MM2, [ECX]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1549 movq MM3, [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
|
1550 add 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
|
1551 psubw MM0, MM2; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1552 psubw MM1, MM3; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1553 movq [ESI -16], MM0; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1554 movq [ESI+8-16], MM1; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1555 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
|
1556 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
|
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 emms; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1559 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
|
1560 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
|
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 } |
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 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1565 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
|
1566 *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
|
1567 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1568 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
|
1569 } |
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 unittest |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1572 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1573 printf("_arraySliceSliceMinass_s 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
|
1574 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1575 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
|
1576 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1577 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
|
1578 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1579 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
|
1580 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1581 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
|
1582 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
|
1583 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
|
1584 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
|
1585 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
|
1586 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
|
1587 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
|
1588 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1589 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
|
1590 { 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
|
1591 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
|
1592 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
|
1593 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1594 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1595 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
|
1596 c[] -= a[]; |
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 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
|
1599 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1600 if (c[i] != cast(T)(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
|
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 printf("[%d]: %d != %d - %d\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
|
1603 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
|
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 } |
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 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1608 } |
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 |
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 |
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 * Computes: |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1615 * 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
|
1616 */ |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1617 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1618 T[] _arraySliceExpMulSliceAssign_u(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
|
1619 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1620 return _arraySliceExpMulSliceAssign_s(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
|
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 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1623 T[] _arraySliceExpMulSliceAssign_t(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
|
1624 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1625 return _arraySliceExpMulSliceAssign_s(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
|
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 T[] _arraySliceExpMulSliceAssign_s(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
|
1629 in |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1630 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1631 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
|
1632 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
|
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 body |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1635 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1636 //printf("_arraySliceExpMulSliceAssign_s()\n"); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1637 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
|
1638 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
|
1639 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
|
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 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
|
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 // SSE2 aligned version is 3733% faster |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1644 if (sse2() && 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
|
1645 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1646 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
|
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 uint l = cast(ushort) value; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1649 l |= l << 16; |
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 if (((cast(uint) aptr | cast(uint) bptr) & 15) != 0) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1652 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1653 asm |
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 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
|
1656 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
|
1657 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
|
1658 movd XMM2, l; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1659 pshufd XMM2, XMM2, 0; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1660 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1661 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
|
1662 startsse2u: |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1663 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
|
1664 movdqu 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
|
1665 movdqu 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
|
1666 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
|
1667 pmullw XMM0, XMM2; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1668 pmullw XMM1, XMM2; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1669 movdqu [ESI -32], XMM0; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1670 movdqu [ESI+16-32], XMM1; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1671 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
|
1672 jb startsse2u; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1673 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1674 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
|
1675 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
|
1676 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1677 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1678 else |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1679 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1680 asm |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1681 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1682 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
|
1683 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
|
1684 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
|
1685 movd XMM2, l; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1686 pshufd XMM2, XMM2, 0; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1687 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1688 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
|
1689 startsse2a: |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1690 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
|
1691 movdqa 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
|
1692 movdqa 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
|
1693 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
|
1694 pmullw XMM0, XMM2; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1695 pmullw XMM1, XMM2; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1696 movdqa [ESI -32], XMM0; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1697 movdqa [ESI+16-32], XMM1; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1698 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
|
1699 jb startsse2a; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1700 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1701 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
|
1702 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
|
1703 } |
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 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1706 else |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1707 // MMX version is 3733% faster |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1708 if (mmx() && 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
|
1709 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1710 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
|
1711 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1712 uint l = cast(ushort) value; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1713 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1714 asm |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1715 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1716 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
|
1717 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
|
1718 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
|
1719 movd MM2, l; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1720 pshufw MM2, MM2, 0; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1721 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1722 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
|
1723 startmmx: |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1724 add 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
|
1725 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
|
1726 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
|
1727 add 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
|
1728 pmullw MM0, MM2; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1729 pmullw MM1, MM2; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1730 movq [ESI -16], MM0; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1731 movq [ESI+8-16], MM1; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1732 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
|
1733 jb startmmx; |
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 emms; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1736 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
|
1737 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
|
1738 } |
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 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1742 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
|
1743 *aptr++ = cast(T)(*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
|
1744 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1745 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
|
1746 } |
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 unittest |
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 printf("_arraySliceExpMulSliceAssign_s 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
|
1751 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1752 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
|
1753 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1754 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
|
1755 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1756 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
|
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 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
|
1759 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
|
1760 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
|
1761 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
|
1762 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
|
1763 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
|
1764 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
|
1765 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1766 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
|
1767 { 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
|
1768 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
|
1769 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
|
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 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1772 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
|
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 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
|
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 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
|
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 printf("[%d]: %d != %d * 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
|
1779 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
|
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 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1783 } |
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 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1787 /* ======================================================================== */ |
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 /*********************** |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1790 * Computes: |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1791 * 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
|
1792 */ |
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 T[] _arraySliceSliceMulSliceAssign_u(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
|
1795 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1796 return _arraySliceSliceMulSliceAssign_s(a, 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
|
1797 } |
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 T[] _arraySliceSliceMulSliceAssign_t(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
|
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 return _arraySliceSliceMulSliceAssign_s(a, 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
|
1802 } |
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 T[] _arraySliceSliceMulSliceAssign_s(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
|
1805 in |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1806 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1807 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
|
1808 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
|
1809 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
|
1810 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
|
1811 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1812 body |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1813 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1814 //printf("_arraySliceSliceMulSliceAssign_s()\n"); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1815 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
|
1816 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
|
1817 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
|
1818 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
|
1819 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1820 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
|
1821 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1822 // SSE2 aligned version is 2515% faster |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1823 if (sse2() && 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
|
1824 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1825 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
|
1826 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1827 if (((cast(uint) aptr | cast(uint) bptr | cast(uint) cptr) & 15) != 0) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1828 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1829 asm |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1830 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1831 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
|
1832 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
|
1833 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
|
1834 mov ECX, cptr; |
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 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
|
1837 startsse2u: |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1838 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
|
1839 movdqu 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
|
1840 movdqu XMM2, [ECX]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1841 movdqu 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
|
1842 movdqu XMM3, [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
|
1843 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
|
1844 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
|
1845 pmullw XMM0, XMM2; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1846 pmullw XMM1, XMM3; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1847 movdqu [ESI -32], XMM0; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1848 movdqu [ESI+16-32], XMM1; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1849 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
|
1850 jb startsse2u; |
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 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
|
1853 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
|
1854 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
|
1855 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1856 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1857 else |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1858 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1859 asm |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1860 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1861 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
|
1862 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
|
1863 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
|
1864 mov ECX, cptr; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1865 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1866 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
|
1867 startsse2a: |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1868 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
|
1869 movdqa 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
|
1870 movdqa XMM2, [ECX]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1871 movdqa 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
|
1872 movdqa XMM3, [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
|
1873 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
|
1874 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
|
1875 pmullw XMM0, XMM2; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1876 pmullw XMM1, XMM3; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1877 movdqa [ESI -32], XMM0; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1878 movdqa [ESI+16-32], XMM1; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1879 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
|
1880 jb startsse2a; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1881 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1882 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
|
1883 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
|
1884 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
|
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 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1887 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1888 else |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1889 // MMX version is 2515% faster |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1890 if (mmx() && 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
|
1891 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1892 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
|
1893 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1894 asm |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1895 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1896 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
|
1897 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
|
1898 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
|
1899 mov ECX, cptr; |
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 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
|
1902 startmmx: |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1903 add 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
|
1904 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
|
1905 movq MM2, [ECX]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1906 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
|
1907 movq MM3, [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
|
1908 add 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
|
1909 add 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
|
1910 pmullw MM0, MM2; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1911 pmullw MM1, MM3; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1912 movq [ESI -16], MM0; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1913 movq [ESI+8-16], MM1; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1914 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
|
1915 jb startmmx; |
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 emms; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1918 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
|
1919 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
|
1920 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
|
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 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
|
1926 *aptr++ = cast(T)(*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
|
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 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
|
1929 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1930 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1931 unittest |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1932 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1933 printf("_arraySliceSliceMulSliceAssign_s 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
|
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 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
|
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 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
|
1938 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1939 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
|
1940 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1941 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
|
1942 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
|
1943 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
|
1944 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
|
1945 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
|
1946 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
|
1947 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
|
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 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
|
1950 { 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
|
1951 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
|
1952 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
|
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 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1955 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
|
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 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
|
1958 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1959 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
|
1960 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1961 printf("[%d]: %d != %d * %d\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
|
1962 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
|
1963 } |
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 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1966 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1967 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1968 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1969 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1970 /* ======================================================================== */ |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1971 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1972 /*********************** |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1973 * Computes: |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1974 * 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
|
1975 */ |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1976 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1977 T[] _arrayExpSliceMulass_u(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
|
1978 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1979 return _arrayExpSliceMulass_s(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
|
1980 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1981 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1982 T[] _arrayExpSliceMulass_t(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
|
1983 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1984 return _arrayExpSliceMulass_s(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
|
1985 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1986 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1987 T[] _arrayExpSliceMulass_s(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
|
1988 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1989 //printf("_arrayExpSliceMulass_s(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
|
1990 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
|
1991 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
|
1992 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1993 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
|
1994 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1995 // SSE2 aligned version is 2044% faster |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1996 if (sse2() && 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
|
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 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
|
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 uint l = cast(ushort) value; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2001 l |= l << 16; |
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 if (((cast(uint) aptr) & 15) != 0) |
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 asm |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2006 { |
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 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
|
2008 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
|
2009 movd XMM2, l; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2010 pshufd XMM2, XMM2, 0; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2011 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2012 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
|
2013 startsse2u: |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2014 movdqu 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
|
2015 movdqu 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
|
2016 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
|
2017 pmullw XMM0, XMM2; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2018 pmullw XMM1, XMM2; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2019 movdqu [ESI -32], XMM0; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2020 movdqu [ESI+16-32], XMM1; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2021 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
|
2022 jb startsse2u; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2023 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2024 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
|
2025 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2026 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2027 else |
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 asm |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2030 { |
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 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
|
2032 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
|
2033 movd XMM2, l; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2034 pshufd XMM2, XMM2, 0; |
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 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
|
2037 startsse2a: |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2038 movdqa 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
|
2039 movdqa 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
|
2040 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
|
2041 pmullw XMM0, XMM2; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2042 pmullw XMM1, XMM2; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2043 movdqa [ESI -32], XMM0; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2044 movdqa [ESI+16-32], XMM1; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2045 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
|
2046 jb startsse2a; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2047 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2048 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
|
2049 } |
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 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2052 else |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2053 // MMX version is 2056% faster |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2054 if (mmx() && 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
|
2055 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2056 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
|
2057 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2058 uint l = cast(ushort) value; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2059 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2060 asm |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2061 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2062 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
|
2063 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
|
2064 movd MM2, l; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2065 pshufw MM2, MM2, 0; |
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 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
|
2068 startmmx: |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2069 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
|
2070 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
|
2071 add 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
|
2072 pmullw MM0, MM2; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2073 pmullw MM1, MM2; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2074 movq [ESI -16], MM0; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2075 movq [ESI+8-16], MM1; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2076 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
|
2077 jb startmmx; |
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 emms; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2080 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
|
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 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2083 } |
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 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
|
2086 *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
|
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 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
|
2089 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2090 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2091 unittest |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2092 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2093 printf("_arrayExpSliceMulass_s 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
|
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 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
|
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 (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
|
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 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
|
2100 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2101 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
|
2102 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
|
2103 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
|
2104 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
|
2105 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
|
2106 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
|
2107 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
|
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 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
|
2110 { 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
|
2111 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
|
2112 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
|
2113 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2114 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2115 b[] = a[]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2116 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
|
2117 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2118 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
|
2119 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2120 if (a[i] != cast(T)(b[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
|
2121 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2122 printf("[%d]: %d != %d * 6\n", 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
|
2123 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
|
2124 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2125 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2126 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2127 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2128 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2129 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2130 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2131 /* ======================================================================== */ |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2132 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2133 /*********************** |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2134 * Computes: |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2135 * 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
|
2136 */ |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2137 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2138 T[] _arraySliceSliceMulass_u(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
|
2139 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2140 return _arraySliceSliceMulass_s(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
|
2141 } |
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 T[] _arraySliceSliceMulass_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
|
2144 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2145 return _arraySliceSliceMulass_s(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
|
2146 } |
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 T[] _arraySliceSliceMulass_s(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
|
2149 in |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2150 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2151 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
|
2152 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
|
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 body |
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 //printf("_arraySliceSliceMulass_s()\n"); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2157 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
|
2158 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
|
2159 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
|
2160 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2161 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
|
2162 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2163 // SSE2 aligned version is 2519% faster |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2164 if (sse2() && 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
|
2165 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2166 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
|
2167 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2168 if (((cast(uint) aptr | cast(uint) bptr) & 15) != 0) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2169 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2170 asm |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2171 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2172 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
|
2173 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
|
2174 mov ECX, bptr; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2175 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2176 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
|
2177 startsse2u: |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2178 movdqu 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
|
2179 movdqu XMM2, [ECX]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2180 movdqu 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
|
2181 movdqu XMM3, [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
|
2182 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
|
2183 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
|
2184 pmullw XMM0, XMM2; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2185 pmullw XMM1, XMM3; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2186 movdqu [ESI -32], XMM0; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2187 movdqu [ESI+16-32], XMM1; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2188 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
|
2189 jb startsse2u; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2190 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2191 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
|
2192 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
|
2193 } |
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 else |
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 asm |
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 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
|
2200 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
|
2201 mov ECX, bptr; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2202 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2203 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
|
2204 startsse2a: |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2205 movdqa 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
|
2206 movdqa XMM2, [ECX]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2207 movdqa 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
|
2208 movdqa XMM3, [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
|
2209 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
|
2210 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
|
2211 pmullw XMM0, XMM2; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2212 pmullw XMM1, XMM3; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2213 movdqa [ESI -32], XMM0; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2214 movdqa [ESI+16-32], XMM1; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2215 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
|
2216 jb startsse2a; |
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 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
|
2219 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
|
2220 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2221 } |
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 else |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2224 // MMX version is 1712% faster |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2225 if (mmx() && 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
|
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 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
|
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 asm |
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 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
|
2232 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
|
2233 mov ECX, bptr; |
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 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
|
2236 startmmx: |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2237 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
|
2238 movq MM2, [ECX]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2239 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
|
2240 movq MM3, [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
|
2241 add 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
|
2242 add 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
|
2243 pmullw MM0, MM2; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2244 pmullw MM1, MM3; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2245 movq [ESI -16], MM0; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2246 movq [ESI+8-16], MM1; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2247 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
|
2248 jb startmmx; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2249 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2250 emms; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2251 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
|
2252 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
|
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 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2255 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2256 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2257 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
|
2258 *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
|
2259 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2260 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
|
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 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2263 unittest |
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 printf("_arraySliceSliceMulass_s 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
|
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 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
|
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 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
|
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 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
|
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 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
|
2274 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
|
2275 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
|
2276 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
|
2277 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
|
2278 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
|
2279 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
|
2280 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2281 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
|
2282 { 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
|
2283 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
|
2284 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
|
2285 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2286 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2287 b[] = a[]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2288 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
|
2289 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2290 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
|
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 if (a[i] != cast(T)(b[i] * c[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]: %d != %d * %d\n", i, a[i], b[i], c[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 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
|
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 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
2298 } |
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 } |