1;
2; jsimdcpu.asm - SIMD instruction support check
3;
4; Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
5;
6; Based on the x86 SIMD extension for IJG JPEG library
7; Copyright (C) 1999-2006, MIYASAKA Masaru.
8; For conditions of distribution and use, see copyright notice in jsimdext.inc
9;
10; This file should be assembled with NASM (Netwide Assembler),
11; can *not* be assembled with Microsoft's MASM or any compatible
12; assembler (including Borland's Turbo Assembler).
13; NASM is available from http://nasm.sourceforge.net/ or
14; http://sourceforge.net/project/showfiles.php?group_id=6208
15;
16; [TAB8]
17
18%include "jsimdext.inc"
19
20; --------------------------------------------------------------------------
21        SECTION SEG_TEXT
22        BITS    32
23;
24; Check if the CPU supports SIMD instructions
25;
26; GLOBAL(unsigned int)
27; jpeg_simd_cpu_support (void)
28;
29
30        align   16
31        global  EXTN(jpeg_simd_cpu_support)
32
33EXTN(jpeg_simd_cpu_support):
34        push    ebx
35;       push    ecx             ; need not be preserved
36;       push    edx             ; need not be preserved
37;       push    esi             ; unused
38        push    edi
39
40        xor     edi,edi                 ; simd support flag
41
42        pushfd
43        pop     eax
44        mov     edx,eax
45        xor     eax, 1<<21              ; flip ID bit in EFLAGS
46        push    eax
47        popfd
48        pushfd
49        pop     eax
50        xor     eax,edx
51        jz      short .return           ; CPUID is not supported
52
53        ; Check for MMX instruction support
54        xor     eax,eax
55        cpuid
56        test    eax,eax
57        jz      short .return
58
59        xor     eax,eax
60        inc     eax
61        cpuid
62        mov     eax,edx                 ; eax = Standard feature flags
63
64        test    eax, 1<<23              ; bit23:MMX
65        jz      short .no_mmx
66        or      edi, byte JSIMD_MMX
67.no_mmx:
68        test    eax, 1<<25              ; bit25:SSE
69        jz      short .no_sse
70        or      edi, byte JSIMD_SSE
71.no_sse:
72        test    eax, 1<<26              ; bit26:SSE2
73        jz      short .no_sse2
74        or      edi, byte JSIMD_SSE2
75.no_sse2:
76
77        ; Check for 3DNow! instruction support
78        mov     eax, 0x80000000
79        cpuid
80        cmp     eax, 0x80000000
81        jbe     short .return
82
83        mov     eax, 0x80000001
84        cpuid
85        mov     eax,edx                 ; eax = Extended feature flags
86
87        test    eax, 1<<31              ; bit31:3DNow!(vendor independent)
88        jz      short .no_3dnow
89        or      edi, byte JSIMD_3DNOW
90.no_3dnow:
91
92.return:
93        mov     eax,edi
94
95        pop     edi
96;       pop     esi             ; unused
97;       pop     edx             ; need not be preserved
98;       pop     ecx             ; need not be preserved
99        pop     ebx
100        ret
101
102; For some reason, the OS X linker does not honor the request to align the
103; segment unless we do this.
104        align   16
105