Mercurial > projects > ldc
annotate druntime/src/compiler/dmd/arraybyte.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 char, byte, |
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 ubyte ('a', 'g' and 'h' 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.arraybyte; |
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 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.amd3dnow amd3dnow; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
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 byte 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 /*********************** |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
50 * Computes: |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
51 * 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
|
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 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
54 T[] _arraySliceExpAddSliceAssign_a(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
|
55 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
56 return _arraySliceExpAddSliceAssign_g(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
|
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 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
59 T[] _arraySliceExpAddSliceAssign_h(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
|
60 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
61 return _arraySliceExpAddSliceAssign_g(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
|
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 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
64 T[] _arraySliceExpAddSliceAssign_g(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
|
65 in |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
66 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
67 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
|
68 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
|
69 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
70 body |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
71 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
72 //printf("_arraySliceExpAddSliceAssign_g()\n"); |
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 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
|
74 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
|
75 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
|
76 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
77 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
|
78 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
79 // SSE2 aligned version is 1088% faster |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
80 if (sse2() && a.length >= 64) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
81 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
82 auto n = aptr + (a.length & ~63); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
83 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
84 uint l = cast(ubyte) value; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
85 l |= (l << 8); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
86 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
|
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 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
|
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 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
|
91 { |
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 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
|
93 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
|
94 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
|
95 movd XMM4, l; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
96 pshufd XMM4, XMM4, 0; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
97 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
98 align 8; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
99 startaddsse2u: |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
100 add ESI, 64; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
101 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
|
102 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
|
103 movdqu XMM2, [EAX+32]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
104 movdqu XMM3, [EAX+48]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
105 add EAX, 64; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
106 paddb XMM0, XMM4; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
107 paddb XMM1, XMM4; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
108 paddb XMM2, XMM4; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
109 paddb XMM3, XMM4; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
110 movdqu [ESI -64], XMM0; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
111 movdqu [ESI+16-64], XMM1; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
112 movdqu [ESI+32-64], XMM2; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
113 movdqu [ESI+48-64], XMM3; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
114 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
|
115 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
|
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 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
|
118 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
|
119 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
120 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
121 else |
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 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
|
124 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
125 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
|
126 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
|
127 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
|
128 movd XMM4, l; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
129 pshufd XMM4, XMM4, 0; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
130 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
131 align 8; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
132 startaddsse2a: |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
133 add ESI, 64; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
134 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
|
135 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
|
136 movdqa XMM2, [EAX+32]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
137 movdqa XMM3, [EAX+48]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
138 add EAX, 64; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
139 paddb XMM0, XMM4; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
140 paddb XMM1, XMM4; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
141 paddb XMM2, XMM4; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
142 paddb XMM3, XMM4; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
143 movdqa [ESI -64], XMM0; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
144 movdqa [ESI+16-64], XMM1; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
145 movdqa [ESI+32-64], XMM2; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
146 movdqa [ESI+48-64], XMM3; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
147 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
|
148 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
|
149 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
150 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
|
151 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
|
152 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
153 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
154 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
155 else |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
156 // MMX version is 1000% faster |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
157 if (mmx() && a.length >= 32) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
158 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
159 auto n = aptr + (a.length & ~31); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
160 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
161 uint l = cast(ubyte) value; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
162 l |= (l << 8); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
163 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
164 asm |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
165 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
166 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
|
167 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
|
168 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
|
169 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
|
170 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
|
171 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
172 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
|
173 startaddmmx: |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
174 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
|
175 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
|
176 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
|
177 movq MM2, [EAX+16]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
178 movq MM3, [EAX+24]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
179 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
|
180 paddb 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
|
181 paddb 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
|
182 paddb MM2, MM4; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
183 paddb MM3, MM4; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
184 movq [ESI -32], MM0; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
185 movq [ESI+8 -32], MM1; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
186 movq [ESI+16-32], MM2; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
187 movq [ESI+24-32], MM3; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
188 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
|
189 jb startaddmmx; |
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 emms; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
192 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
|
193 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
|
194 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
195 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
196 /* trying to be fair and treat normal 32-bit cpu the same way as we do |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
197 * the SIMD units, with unrolled asm. There's not enough registers, |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
198 * really. |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
199 */ |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
200 else |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
201 if (a.length >= 4) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
202 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
203 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
204 auto n = aptr + (a.length & ~3); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
205 asm |
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 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
|
208 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
|
209 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
|
210 mov CL, value; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
211 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
212 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
|
213 startadd386: |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
214 add ESI, 4; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
215 mov DX, [EAX]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
216 mov BX, [EAX+2]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
217 add EAX, 4; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
218 add BL, CL; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
219 add BH, CL; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
220 add DL, CL; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
221 add DH, CL; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
222 mov [ESI -4], DX; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
223 mov [ESI+2 -4], BX; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
224 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
|
225 jb startadd386; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
226 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
227 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
|
228 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
|
229 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
230 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
231 } |
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 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
|
235 *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
|
236 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
237 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
|
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 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
240 unittest |
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 printf("_arraySliceExpAddSliceAssign_g 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
|
243 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
244 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
|
245 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
246 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
|
247 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
248 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
|
249 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
250 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
|
251 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
|
252 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
|
253 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
|
254 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
|
255 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
|
256 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
|
257 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
258 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
|
259 { 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
|
260 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
|
261 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
|
262 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
263 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
264 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
|
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 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
|
267 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
268 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
|
269 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
270 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
|
271 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
|
272 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
273 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
274 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
275 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
276 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
277 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
278 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
279 /* ======================================================================== */ |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
280 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
281 /*********************** |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
282 * Computes: |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
283 * 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
|
284 */ |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
285 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
286 T[] _arraySliceSliceAddSliceAssign_a(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
|
287 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
288 return _arraySliceSliceAddSliceAssign_g(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
|
289 } |
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 T[] _arraySliceSliceAddSliceAssign_h(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
|
292 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
293 return _arraySliceSliceAddSliceAssign_g(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
|
294 } |
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 T[] _arraySliceSliceAddSliceAssign_g(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
|
297 in |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
298 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
299 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
|
300 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
|
301 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
|
302 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
|
303 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
304 body |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
305 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
306 //printf("_arraySliceSliceAddSliceAssign_g()\n"); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
307 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
|
308 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
|
309 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
|
310 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
|
311 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
312 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
|
313 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
314 // SSE2 aligned version is 5739% faster |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
315 if (sse2() && a.length >= 64) |
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 auto n = aptr + (a.length & ~63); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
318 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
319 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
|
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 version (log) printf("\tsse2 unaligned\n"); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
322 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
|
323 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
324 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
|
325 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
|
326 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
|
327 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
|
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 align 8; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
330 startaddlsse2u: |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
331 add ESI, 64; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
332 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
|
333 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
|
334 movdqu XMM2, [EAX+32]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
335 movdqu XMM3, [EAX+48]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
336 add EAX, 64; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
337 movdqu XMM4, [ECX]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
338 movdqu XMM5, [ECX+16]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
339 movdqu XMM6, [ECX+32]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
340 movdqu XMM7, [ECX+48]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
341 add ECX, 64; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
342 paddb XMM0, XMM4; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
343 paddb XMM1, XMM5; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
344 paddb XMM2, XMM6; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
345 paddb XMM3, XMM7; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
346 movdqu [ESI -64], XMM0; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
347 movdqu [ESI+16-64], XMM1; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
348 movdqu [ESI+32-64], XMM2; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
349 movdqu [ESI+48-64], XMM3; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
350 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
|
351 jb startaddlsse2u; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
352 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
353 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 else |
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 version (log) printf("\tsse2 aligned\n"); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
361 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
|
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 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
|
364 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
|
365 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
|
366 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
|
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 align 8; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
369 startaddlsse2a: |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
370 add ESI, 64; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
371 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
|
372 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
|
373 movdqa XMM2, [EAX+32]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
374 movdqa XMM3, [EAX+48]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
375 add EAX, 64; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
376 movdqa XMM4, [ECX]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
377 movdqa XMM5, [ECX+16]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
378 movdqa XMM6, [ECX+32]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
379 movdqa XMM7, [ECX+48]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
380 add ECX, 64; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
381 paddb XMM0, XMM4; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
382 paddb XMM1, XMM5; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
383 paddb XMM2, XMM6; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
384 paddb XMM3, XMM7; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
385 movdqa [ESI -64], XMM0; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
386 movdqa [ESI+16-64], XMM1; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
387 movdqa [ESI+32-64], XMM2; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
388 movdqa [ESI+48-64], XMM3; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
389 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
|
390 jb startaddlsse2a; |
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 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
|
393 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
|
394 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
|
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 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
397 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
398 else |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
399 // MMX version is 4428% faster |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
400 if (mmx() && a.length >= 32) |
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 version (log) printf("\tmmx\n"); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
403 auto n = aptr + (a.length & ~31); |
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 asm |
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 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
|
408 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
|
409 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
|
410 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
|
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 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
|
413 startaddlmmx: |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
414 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
|
415 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
|
416 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
|
417 movq MM2, [EAX+16]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
418 movq MM3, [EAX+24]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
419 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
|
420 movq MM4, [ECX]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
421 movq MM5, [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
|
422 movq MM6, [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
|
423 movq MM7, [ECX+24]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
424 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
|
425 paddb 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
|
426 paddb MM1, MM5; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
427 paddb MM2, MM6; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
428 paddb MM3, MM7; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
429 movq [ESI -32], MM0; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
430 movq [ESI+8 -32], MM1; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
431 movq [ESI+16-32], MM2; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
432 movq [ESI+24-32], MM3; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
433 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
|
434 jb startaddlmmx; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
435 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
436 emms; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
437 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
|
438 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
|
439 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
|
440 } |
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 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
443 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
444 version (log) if (aptr < aend) printf("\tbase\n"); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
445 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
|
446 *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
|
447 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
448 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
|
449 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
450 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
451 unittest |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
452 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
453 printf("_arraySliceSliceAddSliceAssign_g 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
|
454 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
455 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
|
456 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
457 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
|
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 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
|
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 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
|
462 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
|
463 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
|
464 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
|
465 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
|
466 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
|
467 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
|
468 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
469 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
|
470 { 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
|
471 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
|
472 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
|
473 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
474 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
475 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
|
476 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
477 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
|
478 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
479 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
|
480 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
481 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
|
482 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
|
483 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
484 } |
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 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
488 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
489 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
490 /* ======================================================================== */ |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
491 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
492 /*********************** |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
493 * Computes: |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
494 * 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
|
495 */ |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
496 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
497 T[] _arrayExpSliceAddass_a(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
|
498 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
499 return _arrayExpSliceAddass_g(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
|
500 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
501 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
502 T[] _arrayExpSliceAddass_h(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
|
503 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
504 return _arrayExpSliceAddass_g(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
|
505 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
506 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
507 T[] _arrayExpSliceAddass_g(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
|
508 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
509 //printf("_arrayExpSliceAddass_g(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
|
510 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
|
511 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
|
512 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
513 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
|
514 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
515 // SSE2 aligned version is 1578% faster |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
516 if (sse2() && a.length >= 64) |
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 auto n = aptr + (a.length & ~63); |
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 uint l = cast(ubyte) value; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
521 l |= (l << 8); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
522 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
|
523 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
524 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
|
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 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
|
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 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
|
529 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
|
530 movd XMM4, l; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
531 pshufd XMM4, XMM4, 0; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
532 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
533 align 8; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
534 startaddasssse2u: |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
535 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
|
536 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
|
537 movdqu XMM2, [ESI+32]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
538 movdqu XMM3, [ESI+48]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
539 add ESI, 64; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
540 paddb XMM0, XMM4; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
541 paddb XMM1, XMM4; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
542 paddb XMM2, XMM4; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
543 paddb XMM3, XMM4; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
544 movdqu [ESI -64], XMM0; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
545 movdqu [ESI+16-64], XMM1; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
546 movdqu [ESI+32-64], XMM2; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
547 movdqu [ESI+48-64], XMM3; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
548 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
|
549 jb startaddasssse2u; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
550 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
551 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
|
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 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
554 else |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
555 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
556 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
|
557 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
558 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
|
559 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
|
560 movd XMM4, l; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
561 pshufd XMM4, XMM4, 0; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
562 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
563 align 8; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
564 startaddasssse2a: |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
565 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
|
566 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
|
567 movdqa XMM2, [ESI+32]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
568 movdqa XMM3, [ESI+48]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
569 add ESI, 64; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
570 paddb XMM0, XMM4; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
571 paddb XMM1, XMM4; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
572 paddb XMM2, XMM4; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
573 paddb XMM3, XMM4; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
574 movdqa [ESI -64], XMM0; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
575 movdqa [ESI+16-64], XMM1; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
576 movdqa [ESI+32-64], XMM2; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
577 movdqa [ESI+48-64], XMM3; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
578 cmp ESI, EDI; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
579 jb startaddasssse2a; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
580 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
581 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
|
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 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
584 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
585 else |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
586 // MMX version is 1721% faster |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
587 if (mmx() && a.length >= 32) |
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 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
590 auto n = aptr + (a.length & ~31); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
591 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
592 uint l = cast(ubyte) value; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
593 l |= (l << 8); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
594 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
595 asm |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
596 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
597 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
|
598 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
|
599 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
|
600 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
|
601 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
602 align 8; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
603 startaddassmmx: |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
604 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
|
605 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
|
606 movq MM2, [ESI+16]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
607 movq MM3, [ESI+24]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
608 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
|
609 paddb 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
|
610 paddb 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
|
611 paddb MM2, MM4; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
612 paddb MM3, MM4; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
613 movq [ESI -32], MM0; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
614 movq [ESI+8 -32], MM1; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
615 movq [ESI+16-32], MM2; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
616 movq [ESI+24-32], MM3; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
617 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
|
618 jb startaddassmmx; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
619 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
620 emms; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
621 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
|
622 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
623 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
624 } |
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 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
|
627 *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
|
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 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
|
630 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
631 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
632 unittest |
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 printf("_arrayExpSliceAddass_g 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
|
635 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
636 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
|
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 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
|
639 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
640 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
|
641 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
642 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
|
643 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
|
644 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
|
645 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
|
646 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
|
647 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
|
648 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
|
649 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
650 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
|
651 { 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
|
652 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
|
653 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
|
654 } |
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 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
|
657 c[] += 6; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
658 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
659 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
|
660 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
661 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
|
662 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
663 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
|
664 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
|
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 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
667 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
668 } |
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 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
671 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
672 /* ======================================================================== */ |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
673 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
674 /*********************** |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
675 * Computes: |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
676 * 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
|
677 */ |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
678 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
679 T[] _arraySliceSliceAddass_a(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
|
680 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
681 return _arraySliceSliceAddass_g(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
|
682 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
683 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
684 T[] _arraySliceSliceAddass_h(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
|
685 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
686 return _arraySliceSliceAddass_g(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
|
687 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
688 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
689 T[] _arraySliceSliceAddass_g(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
|
690 in |
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 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
|
693 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
|
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 body |
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 //printf("_arraySliceSliceAddass_g()\n"); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
698 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
|
699 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
|
700 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
|
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 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
|
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 // SSE2 aligned version is 4727% faster |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
705 if (sse2() && a.length >= 64) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
706 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
707 auto n = aptr + (a.length & ~63); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
708 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
709 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
|
710 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
711 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
|
712 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
713 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
|
714 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
|
715 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
|
716 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
717 align 8; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
718 startaddasslsse2u: |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
719 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
|
720 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
|
721 movdqu XMM2, [ESI+32]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
722 movdqu XMM3, [ESI+48]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
723 add ESI, 64; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
724 movdqu XMM4, [ECX]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
725 movdqu XMM5, [ECX+16]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
726 movdqu XMM6, [ECX+32]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
727 movdqu XMM7, [ECX+48]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
728 add ECX, 64; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
729 paddb XMM0, XMM4; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
730 paddb XMM1, XMM5; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
731 paddb XMM2, XMM6; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
732 paddb XMM3, XMM7; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
733 movdqu [ESI -64], XMM0; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
734 movdqu [ESI+16-64], XMM1; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
735 movdqu [ESI+32-64], XMM2; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
736 movdqu [ESI+48-64], XMM3; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
737 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
|
738 jb startaddasslsse2u; |
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 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
|
741 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
|
742 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
743 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
744 else |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
745 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
746 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
|
747 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
748 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
|
749 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
|
750 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
|
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 align 8; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
753 startaddasslsse2a: |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
754 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
|
755 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
|
756 movdqa XMM2, [ESI+32]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
757 movdqa XMM3, [ESI+48]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
758 add ESI, 64; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
759 movdqa XMM4, [ECX]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
760 movdqa XMM5, [ECX+16]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
761 movdqa XMM6, [ECX+32]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
762 movdqa XMM7, [ECX+48]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
763 add ECX, 64; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
764 paddb XMM0, XMM4; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
765 paddb XMM1, XMM5; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
766 paddb XMM2, XMM6; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
767 paddb XMM3, XMM7; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
768 movdqa [ESI -64], XMM0; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
769 movdqa [ESI+16-64], XMM1; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
770 movdqa [ESI+32-64], XMM2; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
771 movdqa [ESI+48-64], XMM3; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
772 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
|
773 jb startaddasslsse2a; |
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 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
|
776 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
|
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 } |
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 else |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
781 // MMX version is 3059% faster |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
782 if (mmx() && a.length >= 32) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
783 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
784 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
785 auto n = aptr + (a.length & ~31); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
786 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
787 asm |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
788 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
789 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
|
790 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
|
791 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
|
792 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
793 align 8; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
794 startaddasslmmx: |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
795 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
|
796 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
|
797 movq MM2, [ESI+16]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
798 movq MM3, [ESI+24]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
799 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
|
800 movq MM4, [ECX]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
801 movq MM5, [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
|
802 movq MM6, [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
|
803 movq MM7, [ECX+24]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
804 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
|
805 paddb 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
|
806 paddb MM1, MM5; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
807 paddb MM2, MM6; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
808 paddb MM3, MM7; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
809 movq [ESI -32], MM0; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
810 movq [ESI+8 -32], MM1; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
811 movq [ESI+16-32], MM2; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
812 movq [ESI+24-32], MM3; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
813 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
|
814 jb startaddasslmmx; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
815 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
816 emms; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
817 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
|
818 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
|
819 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
820 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
821 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
822 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
823 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
|
824 *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
|
825 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
826 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
|
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 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
829 unittest |
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 printf("_arraySliceSliceAddass_g 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
|
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 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
|
834 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
835 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
|
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 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
|
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 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
|
840 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
|
841 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
|
842 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
|
843 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
|
844 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
|
845 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
|
846 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
847 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
|
848 { 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
|
849 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
|
850 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
|
851 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
852 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
853 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
|
854 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
|
855 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
856 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
|
857 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
858 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
|
859 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
860 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
|
861 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
|
862 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
863 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
864 } |
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 /* ======================================================================== */ |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
870 |
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 /*********************** |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
873 * Computes: |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
874 * 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
|
875 */ |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
876 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
877 T[] _arraySliceExpMinSliceAssign_a(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
|
878 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
879 return _arraySliceExpMinSliceAssign_g(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
|
880 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
881 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
882 T[] _arraySliceExpMinSliceAssign_h(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
|
883 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
884 return _arraySliceExpMinSliceAssign_g(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
|
885 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
886 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
887 T[] _arraySliceExpMinSliceAssign_g(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
|
888 in |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
889 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
890 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
|
891 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
|
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 body |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
894 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
895 //printf("_arraySliceExpMinSliceAssign_g()\n"); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
896 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
|
897 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
|
898 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
|
899 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
900 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
|
901 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
902 // SSE2 aligned version is 1189% faster |
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 (sse2() && a.length >= 64) |
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 auto n = aptr + (a.length & ~63); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
906 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
907 uint l = cast(ubyte) value; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
908 l |= (l << 8); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
909 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
|
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 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
|
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 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
|
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 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
|
916 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
|
917 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
|
918 movd XMM4, l; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
919 pshufd XMM4, XMM4, 0; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
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 align 8; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
922 startsubsse2u: |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
923 add ESI, 64; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
924 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
|
925 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
|
926 movdqu XMM2, [EAX+32]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
927 movdqu XMM3, [EAX+48]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
928 add EAX, 64; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
929 psubb XMM0, XMM4; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
930 psubb XMM1, XMM4; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
931 psubb XMM2, XMM4; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
932 psubb XMM3, XMM4; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
933 movdqu [ESI -64], XMM0; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
934 movdqu [ESI+16-64], XMM1; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
935 movdqu [ESI+32-64], XMM2; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
936 movdqu [ESI+48-64], XMM3; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
937 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
|
938 jb startsubsse2u; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
939 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
940 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
|
941 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
|
942 } |
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 else |
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 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
|
947 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
948 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
|
949 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
|
950 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
|
951 movd XMM4, l; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
952 pshufd XMM4, XMM4, 0; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
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 align 8; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
955 startsubsse2a: |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
956 add ESI, 64; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
957 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
|
958 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
|
959 movdqa XMM2, [EAX+32]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
960 movdqa XMM3, [EAX+48]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
961 add EAX, 64; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
962 psubb XMM0, XMM4; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
963 psubb XMM1, XMM4; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
964 psubb XMM2, XMM4; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
965 psubb XMM3, XMM4; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
966 movdqa [ESI -64], XMM0; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
967 movdqa [ESI+16-64], XMM1; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
968 movdqa [ESI+32-64], XMM2; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
969 movdqa [ESI+48-64], XMM3; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
970 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
|
971 jb startsubsse2a; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
972 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
973 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
|
974 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
|
975 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
976 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
977 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
978 else |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
979 // MMX version is 1079% faster |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
980 if (mmx() && a.length >= 32) |
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 auto n = aptr + (a.length & ~31); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
983 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
984 uint l = cast(ubyte) value; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
985 l |= (l << 8); |
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 asm |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
988 { |
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 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
|
990 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
|
991 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
|
992 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
|
993 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
|
994 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
995 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
|
996 startsubmmx: |
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 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
|
999 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
|
1000 movq MM2, [EAX+16]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1001 movq MM3, [EAX+24]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1002 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
|
1003 psubb 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
|
1004 psubb 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
|
1005 psubb MM2, MM4; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1006 psubb MM3, MM4; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1007 movq [ESI -32], MM0; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1008 movq [ESI+8 -32], MM1; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1009 movq [ESI+16-32], MM2; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1010 movq [ESI+24-32], MM3; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1011 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
|
1012 jb startsubmmx; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1013 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1014 emms; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1015 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
|
1016 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
|
1017 } |
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 // trying to be fair and treat normal 32-bit cpu the same way as we do the SIMD units, with unrolled asm. There's not enough registers, really. |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1020 else |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1021 if (a.length >= 4) |
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 auto n = aptr + (a.length & ~3); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1024 asm |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1025 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1026 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
|
1027 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
|
1028 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
|
1029 mov CL, value; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1030 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1031 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
|
1032 startsub386: |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1033 add ESI, 4; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1034 mov DX, [EAX]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1035 mov BX, [EAX+2]; |
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, 4; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1037 sub BL, CL; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1038 sub BH, CL; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1039 sub DL, CL; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1040 sub DH, CL; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1041 mov [ESI -4], DX; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1042 mov [ESI+2 -4], BX; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1043 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
|
1044 jb startsub386; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1045 |
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 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
|
1047 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
|
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 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1052 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
|
1053 *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
|
1054 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1055 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
|
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 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1058 unittest |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1059 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1060 printf("_arraySliceExpMinSliceAssign_g 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
|
1061 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1062 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
|
1063 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1064 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
|
1065 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1066 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
|
1067 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1068 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
|
1069 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
|
1070 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
|
1071 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
|
1072 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
|
1073 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
|
1074 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
|
1075 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1076 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
|
1077 { 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
|
1078 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
|
1079 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
|
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 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1082 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
|
1083 c[] = b[] - 6; |
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 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
|
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 if (c[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
|
1088 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1089 printf("[%d]: %d != %d - 6\n", i, c[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
|
1090 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
|
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 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1100 /*********************** |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1101 * Computes: |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1102 * 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
|
1103 */ |
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 T[] _arrayExpSliceMinSliceAssign_a(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
|
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 return _arrayExpSliceMinSliceAssign_g(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
|
1108 } |
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 T[] _arrayExpSliceMinSliceAssign_h(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
|
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 return _arrayExpSliceMinSliceAssign_g(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
|
1113 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1114 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1115 T[] _arrayExpSliceMinSliceAssign_g(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
|
1116 in |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1117 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1118 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
|
1119 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
|
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 //printf("_arrayExpSliceMinSliceAssign_g()\n"); |
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 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
|
1125 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
|
1126 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
|
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 8748% 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 >= 64) |
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 & ~63); |
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 uint l = cast(ubyte) value; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1136 l |= (l << 8); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1137 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
|
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 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
|
1140 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1141 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
|
1142 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1143 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
|
1144 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
|
1145 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
|
1146 movd XMM4, l; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1147 pshufd XMM4, XMM4, 0; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1148 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1149 align 8; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1150 startsubrsse2u: |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1151 add ESI, 64; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1152 movdqa XMM5, XMM4; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1153 movdqa XMM6, XMM4; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1154 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
|
1155 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
|
1156 psubb XMM5, XMM0; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1157 psubb XMM6, XMM1; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1158 movdqu [ESI -64], XMM5; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1159 movdqu [ESI+16-64], XMM6; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1160 movdqa XMM5, XMM4; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1161 movdqa XMM6, XMM4; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1162 movdqu XMM2, [EAX+32]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1163 movdqu XMM3, [EAX+48]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1164 add EAX, 64; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1165 psubb XMM5, XMM2; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1166 psubb XMM6, XMM3; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1167 movdqu [ESI+32-64], XMM5; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1168 movdqu [ESI+48-64], XMM6; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1169 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
|
1170 jb startsubrsse2u; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1171 |
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 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
|
1173 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
|
1174 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1175 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1176 else |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1177 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1178 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
|
1179 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1180 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
|
1181 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
|
1182 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
|
1183 movd XMM4, l; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1184 pshufd XMM4, XMM4, 0; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1185 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1186 align 8; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1187 startsubrsse2a: |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1188 add ESI, 64; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1189 movdqa XMM5, XMM4; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1190 movdqa XMM6, XMM4; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1191 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
|
1192 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
|
1193 psubb XMM5, XMM0; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1194 psubb XMM6, XMM1; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1195 movdqa [ESI -64], XMM5; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1196 movdqa [ESI+16-64], XMM6; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1197 movdqa XMM5, XMM4; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1198 movdqa XMM6, XMM4; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1199 movdqa XMM2, [EAX+32]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1200 movdqa XMM3, [EAX+48]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1201 add EAX, 64; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1202 psubb XMM5, XMM2; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1203 psubb XMM6, XMM3; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1204 movdqa [ESI+32-64], XMM5; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1205 movdqa [ESI+48-64], XMM6; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1206 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
|
1207 jb startsubrsse2a; |
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 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
|
1210 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
|
1211 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1212 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1213 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1214 else |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1215 // MMX version is 7397% faster |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1216 if (mmx() && a.length >= 32) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1217 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1218 auto n = aptr + (a.length & ~31); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1219 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1220 uint l = cast(ubyte) value; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1221 l |= (l << 8); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1222 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1223 asm |
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 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
|
1226 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
|
1227 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
|
1228 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
|
1229 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
|
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 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
|
1232 startsubrmmx: |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1233 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
|
1234 movq MM5, MM4; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1235 movq MM6, MM4; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1236 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
|
1237 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
|
1238 psubb MM5, MM0; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1239 psubb MM6, MM1; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1240 movq [ESI -32], MM5; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1241 movq [ESI+8 -32], MM6; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1242 movq MM5, MM4; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1243 movq MM6, MM4; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1244 movq MM2, [EAX+16]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1245 movq MM3, [EAX+24]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1246 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
|
1247 psubb MM5, MM2; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1248 psubb MM6, MM3; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1249 movq [ESI+16-32], MM5; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1250 movq [ESI+24-32], MM6; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1251 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
|
1252 jb startsubrmmx; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1253 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1254 emms; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1255 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
|
1256 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
|
1257 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1258 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1259 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1260 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1261 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1262 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
|
1263 *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
|
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 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
|
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 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1268 unittest |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1269 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1270 printf("_arrayExpSliceMinSliceAssign_g 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
|
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 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
|
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 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
|
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 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
|
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 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
|
1279 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
|
1280 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
|
1281 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
|
1282 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
|
1283 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
|
1284 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
|
1285 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1286 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
|
1287 { 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
|
1288 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
|
1289 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
|
1290 } |
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 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
|
1293 c[] = 6 - b[]; |
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 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
|
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 if (c[i] != cast(T)(6 - 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
|
1298 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1299 printf("[%d]: %d != 6 - %d\n", i, c[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
|
1300 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
|
1301 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1302 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1303 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1304 } |
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 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1307 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1308 /* ======================================================================== */ |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1309 |
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 * Computes: |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1312 * 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
|
1313 */ |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1314 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1315 T[] _arraySliceSliceMinSliceAssign_a(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
|
1316 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1317 return _arraySliceSliceMinSliceAssign_g(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
|
1318 } |
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 T[] _arraySliceSliceMinSliceAssign_h(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
|
1321 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1322 return _arraySliceSliceMinSliceAssign_g(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
|
1323 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1324 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1325 T[] _arraySliceSliceMinSliceAssign_g(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
|
1326 in |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1327 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1328 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
|
1329 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
|
1330 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
|
1331 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
|
1332 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1333 body |
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 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
|
1336 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
|
1337 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
|
1338 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
|
1339 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1340 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
|
1341 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1342 // SSE2 aligned version is 5756% faster |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1343 if (sse2() && a.length >= 64) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1344 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1345 auto n = aptr + (a.length & ~63); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1346 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1347 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
|
1348 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1349 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
|
1350 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1351 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
|
1352 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
|
1353 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
|
1354 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
|
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 align 8; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1357 startsublsse2u: |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1358 add ESI, 64; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1359 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
|
1360 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
|
1361 movdqu XMM2, [EAX+32]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1362 movdqu XMM3, [EAX+48]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1363 add EAX, 64; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1364 movdqu XMM4, [ECX]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1365 movdqu XMM5, [ECX+16]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1366 movdqu XMM6, [ECX+32]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1367 movdqu XMM7, [ECX+48]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1368 add ECX, 64; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1369 psubb XMM0, XMM4; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1370 psubb XMM1, XMM5; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1371 psubb XMM2, XMM6; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1372 psubb XMM3, XMM7; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1373 movdqu [ESI -64], XMM0; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1374 movdqu [ESI+16-64], XMM1; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1375 movdqu [ESI+32-64], XMM2; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1376 movdqu [ESI+48-64], XMM3; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1377 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
|
1378 jb startsublsse2u; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1379 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1380 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
|
1381 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
|
1382 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
|
1383 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1384 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1385 else |
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 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
|
1388 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1389 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
|
1390 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
|
1391 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
|
1392 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
|
1393 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1394 align 8; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1395 startsublsse2a: |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1396 add ESI, 64; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1397 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
|
1398 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
|
1399 movdqa XMM2, [EAX+32]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1400 movdqa XMM3, [EAX+48]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1401 add EAX, 64; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1402 movdqa XMM4, [ECX]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1403 movdqa XMM5, [ECX+16]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1404 movdqa XMM6, [ECX+32]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1405 movdqa XMM7, [ECX+48]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1406 add ECX, 64; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1407 psubb XMM0, XMM4; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1408 psubb XMM1, XMM5; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1409 psubb XMM2, XMM6; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1410 psubb XMM3, XMM7; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1411 movdqa [ESI -64], XMM0; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1412 movdqa [ESI+16-64], XMM1; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1413 movdqa [ESI+32-64], XMM2; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1414 movdqa [ESI+48-64], XMM3; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1415 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
|
1416 jb startsublsse2a; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1417 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1418 mov 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
|
1419 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
|
1420 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
|
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 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1424 else |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1425 // MMX version is 4428% faster |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1426 if (mmx() && a.length >= 32) |
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 auto n = aptr + (a.length & ~31); |
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 asm |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1431 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1432 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
|
1433 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
|
1434 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
|
1435 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
|
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 align 8; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1438 startsublmmx: |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1439 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
|
1440 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
|
1441 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
|
1442 movq MM2, [EAX+16]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1443 movq MM3, [EAX+24]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1444 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
|
1445 movq MM4, [ECX]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1446 movq MM5, [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
|
1447 movq MM6, [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
|
1448 movq MM7, [ECX+24]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1449 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
|
1450 psubb 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
|
1451 psubb MM1, MM5; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1452 psubb MM2, MM6; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1453 psubb MM3, MM7; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1454 movq [ESI -32], MM0; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1455 movq [ESI+8 -32], MM1; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1456 movq [ESI+16-32], MM2; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1457 movq [ESI+24-32], MM3; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1458 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
|
1459 jb startsublmmx; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1460 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1461 emms; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1462 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
|
1463 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
|
1464 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
|
1465 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1466 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1467 } |
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 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
|
1470 *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
|
1471 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1472 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
|
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 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1475 unittest |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1476 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1477 printf("_arraySliceSliceMinSliceAssign_g 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
|
1478 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1479 for (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
|
1480 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1481 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
|
1482 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1483 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
|
1484 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1485 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
|
1486 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
|
1487 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
|
1488 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
|
1489 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
|
1490 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
|
1491 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
|
1492 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1493 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
|
1494 { 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
|
1495 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
|
1496 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
|
1497 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1498 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1499 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
|
1500 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1501 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
|
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 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
|
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 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
|
1506 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
|
1507 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1508 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1509 } |
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 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1512 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1513 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1514 /* ======================================================================== */ |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1515 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1516 /*********************** |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1517 * Computes: |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1518 * 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
|
1519 */ |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1520 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1521 T[] _arrayExpSliceMinass_a(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
|
1522 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1523 return _arrayExpSliceMinass_g(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
|
1524 } |
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 T[] _arrayExpSliceMinass_h(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
|
1527 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1528 return _arrayExpSliceMinass_g(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
|
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 T[] _arrayExpSliceMinass_g(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
|
1532 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1533 //printf("_arrayExpSliceMinass_g(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
|
1534 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
|
1535 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
|
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 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
|
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 // SSE2 aligned version is 1577% faster |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1540 if (sse2() && a.length >= 64) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1541 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1542 auto n = aptr + (a.length & ~63); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1543 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1544 uint l = cast(ubyte) value; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1545 l |= (l << 8); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1546 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
|
1547 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1548 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
|
1549 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1550 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
|
1551 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1552 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
|
1553 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
|
1554 movd XMM4, l; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1555 pshufd XMM4, XMM4, 0; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1556 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1557 align 8; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1558 startsubasssse2u: |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1559 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
|
1560 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
|
1561 movdqu XMM2, [ESI+32]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1562 movdqu XMM3, [ESI+48]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1563 add ESI, 64; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1564 psubb XMM0, XMM4; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1565 psubb XMM1, XMM4; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1566 psubb XMM2, XMM4; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1567 psubb XMM3, XMM4; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1568 movdqu [ESI -64], XMM0; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1569 movdqu [ESI+16-64], XMM1; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1570 movdqu [ESI+32-64], XMM2; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1571 movdqu [ESI+48-64], XMM3; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1572 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
|
1573 jb startsubasssse2u; |
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 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
|
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 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1578 else |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1579 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1580 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
|
1581 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1582 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
|
1583 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
|
1584 movd XMM4, l; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1585 pshufd XMM4, XMM4, 0; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1586 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1587 align 8; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1588 startsubasssse2a: |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1589 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
|
1590 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
|
1591 movdqa XMM2, [ESI+32]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1592 movdqa XMM3, [ESI+48]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1593 add ESI, 64; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1594 psubb XMM0, XMM4; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1595 psubb XMM1, XMM4; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1596 psubb XMM2, XMM4; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1597 psubb XMM3, XMM4; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1598 movdqa [ESI -64], XMM0; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1599 movdqa [ESI+16-64], XMM1; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1600 movdqa [ESI+32-64], XMM2; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1601 movdqa [ESI+48-64], XMM3; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1602 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
|
1603 jb startsubasssse2a; |
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 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
|
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 else |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1610 // MMX version is 1577% faster |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1611 if (mmx() && a.length >= 32) |
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 auto n = aptr + (a.length & ~31); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1615 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1616 uint l = cast(ubyte) value; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1617 l |= (l << 8); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1618 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1619 asm |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1620 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1621 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
|
1622 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
|
1623 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
|
1624 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
|
1625 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1626 align 8; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1627 startsubassmmx: |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1628 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
|
1629 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
|
1630 movq MM2, [ESI+16]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1631 movq MM3, [ESI+24]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1632 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
|
1633 psubb 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
|
1634 psubb 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
|
1635 psubb MM2, MM4; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1636 psubb MM3, MM4; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1637 movq [ESI -32], MM0; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1638 movq [ESI+8 -32], MM1; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1639 movq [ESI+16-32], MM2; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1640 movq [ESI+24-32], MM3; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1641 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
|
1642 jb startsubassmmx; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1643 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1644 emms; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1645 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
|
1646 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1647 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1648 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1649 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1650 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
|
1651 *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
|
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 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
|
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 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1656 unittest |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1657 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1658 printf("_arrayExpSliceMinass_g 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
|
1659 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1660 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
|
1661 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1662 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
|
1663 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1664 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
|
1665 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1666 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
|
1667 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
|
1668 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
|
1669 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
|
1670 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
|
1671 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
|
1672 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
|
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 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
|
1675 { 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
|
1676 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
|
1677 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
|
1678 } |
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 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
|
1681 c[] -= 6; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1682 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1683 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
|
1684 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1685 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
|
1686 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1687 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
|
1688 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
|
1689 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1690 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1691 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1692 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1693 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1694 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1695 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1696 /* ======================================================================== */ |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1697 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1698 /*********************** |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1699 * Computes: |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1700 * 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
|
1701 */ |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1702 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1703 T[] _arraySliceSliceMinass_a(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
|
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 return _arraySliceSliceMinass_g(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
|
1706 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1707 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1708 T[] _arraySliceSliceMinass_h(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
|
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 return _arraySliceSliceMinass_g(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
|
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 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1713 T[] _arraySliceSliceMinass_g(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
|
1714 in |
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 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
|
1717 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
|
1718 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1719 body |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1720 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1721 //printf("_arraySliceSliceMinass_g()\n"); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1722 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
|
1723 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
|
1724 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
|
1725 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1726 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
|
1727 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1728 // SSE2 aligned version is 4800% faster |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1729 if (sse2() && a.length >= 64) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1730 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1731 auto n = aptr + (a.length & ~63); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1732 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1733 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
|
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 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
|
1736 { |
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 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
|
1738 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
|
1739 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
|
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 align 8; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1742 startsubasslsse2u: |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1743 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
|
1744 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
|
1745 movdqu XMM2, [ESI+32]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1746 movdqu XMM3, [ESI+48]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1747 add ESI, 64; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1748 movdqu XMM4, [ECX]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1749 movdqu XMM5, [ECX+16]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1750 movdqu XMM6, [ECX+32]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1751 movdqu XMM7, [ECX+48]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1752 add ECX, 64; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1753 psubb XMM0, XMM4; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1754 psubb XMM1, XMM5; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1755 psubb XMM2, XMM6; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1756 psubb XMM3, XMM7; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1757 movdqu [ESI -64], XMM0; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1758 movdqu [ESI+16-64], XMM1; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1759 movdqu [ESI+32-64], XMM2; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1760 movdqu [ESI+48-64], XMM3; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1761 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
|
1762 jb startsubasslsse2u; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1763 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1764 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
|
1765 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
|
1766 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1767 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1768 else |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1769 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1770 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
|
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 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
|
1773 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
|
1774 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
|
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 align 8; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1777 startsubasslsse2a: |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1778 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
|
1779 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
|
1780 movdqa XMM2, [ESI+32]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1781 movdqa XMM3, [ESI+48]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1782 add ESI, 64; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1783 movdqa XMM4, [ECX]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1784 movdqa XMM5, [ECX+16]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1785 movdqa XMM6, [ECX+32]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1786 movdqa XMM7, [ECX+48]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1787 add ECX, 64; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1788 psubb XMM0, XMM4; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1789 psubb XMM1, XMM5; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1790 psubb XMM2, XMM6; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1791 psubb XMM3, XMM7; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1792 movdqa [ESI -64], XMM0; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1793 movdqa [ESI+16-64], XMM1; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1794 movdqa [ESI+32-64], XMM2; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1795 movdqa [ESI+48-64], XMM3; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1796 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
|
1797 jb startsubasslsse2a; |
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 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
|
1800 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
|
1801 } |
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 else |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1805 // MMX version is 3107% faster |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1806 if (mmx() && a.length >= 32) |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1807 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1808 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1809 auto n = aptr + (a.length & ~31); |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1810 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1811 asm |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1812 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1813 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
|
1814 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
|
1815 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
|
1816 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1817 align 8; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1818 startsubasslmmx: |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1819 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
|
1820 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
|
1821 movq MM2, [ESI+16]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1822 movq MM3, [ESI+24]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1823 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
|
1824 movq MM4, [ECX]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1825 movq MM5, [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
|
1826 movq MM6, [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
|
1827 movq MM7, [ECX+24]; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1828 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
|
1829 psubb 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
|
1830 psubb MM1, MM5; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1831 psubb MM2, MM6; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1832 psubb MM3, MM7; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1833 movq [ESI -32], MM0; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1834 movq [ESI+8 -32], MM1; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1835 movq [ESI+16-32], MM2; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1836 movq [ESI+24-32], MM3; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1837 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
|
1838 jb startsubasslmmx; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1839 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1840 emms; |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1841 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
|
1842 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
|
1843 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1844 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1845 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1846 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1847 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
|
1848 *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
|
1849 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1850 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
|
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 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1853 unittest |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1854 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1855 printf("_arraySliceSliceMinass_g 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
|
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 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
|
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 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
|
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 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
|
1862 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1863 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
|
1864 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
|
1865 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
|
1866 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
|
1867 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
|
1868 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
|
1869 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
|
1870 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1871 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
|
1872 { 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
|
1873 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
|
1874 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
|
1875 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1876 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1877 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
|
1878 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
|
1879 |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1880 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
|
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 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
|
1883 { |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1884 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
|
1885 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
|
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 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1889 } |
d3eb054172f9
Added copy of druntime from DMD 2.020 modified for LDC.
Tomas Lindquist Olsen <tomas.l.olsen@gmail.com>
parents:
diff
changeset
|
1890 } |