OLD | NEW |
1 ; | 1 ; |
2 ; jcclrss2-64.asm - colorspace conversion (64-bit SSE2) | 2 ; jcclrss2-64.asm - colorspace conversion (64-bit SSE2) |
3 ; | 3 ; |
4 ; x86 SIMD extension for IJG JPEG library | 4 ; x86 SIMD extension for IJG JPEG library |
5 ; Copyright (C) 1999-2006, MIYASAKA Masaru. | 5 ; Copyright (C) 1999-2006, MIYASAKA Masaru. |
6 ; Copyright (C) 2009, D. R. Commander. | 6 ; Copyright (C) 2009, D. R. Commander. |
7 ; For conditions of distribution and use, see copyright notice in jsimdext.inc | 7 ; For conditions of distribution and use, see copyright notice in jsimdext.inc |
8 ; | 8 ; |
9 ; This file should be assembled with NASM (Netwide Assembler), | 9 ; This file should be assembled with NASM (Netwide Assembler), |
10 ; can *not* be assembled with Microsoft's MASM or any compatible | 10 ; can *not* be assembled with Microsoft's MASM or any compatible |
(...skipping 32 matching lines...) Expand 10 before | Expand all | Expand 10 after Loading... |
43 push rbp | 43 push rbp |
44 mov rax,rsp ; rax = original rbp | 44 mov rax,rsp ; rax = original rbp |
45 sub rsp, byte 4 | 45 sub rsp, byte 4 |
46 and rsp, byte (-SIZEOF_XMMWORD) ; align to 128 bits | 46 and rsp, byte (-SIZEOF_XMMWORD) ; align to 128 bits |
47 mov [rsp],rax | 47 mov [rsp],rax |
48 mov rbp,rsp ; rbp = aligned rbp | 48 mov rbp,rsp ; rbp = aligned rbp |
49 lea rsp, [wk(0)] | 49 lea rsp, [wk(0)] |
50 collect_args | 50 collect_args |
51 push rbx | 51 push rbx |
52 | 52 |
53 » mov» rcx, r10 | 53 » mov» ecx, r10d |
54 test rcx,rcx | 54 test rcx,rcx |
55 jz near .return | 55 jz near .return |
56 | 56 |
57 push rcx | 57 push rcx |
58 | 58 |
59 mov rsi, r12 | 59 mov rsi, r12 |
60 » mov rcx, r13 | 60 » mov ecx, r13d |
61 mov rdi, JSAMPARRAY [rsi+0*SIZEOF_JSAMPARRAY] | 61 mov rdi, JSAMPARRAY [rsi+0*SIZEOF_JSAMPARRAY] |
62 mov rbx, JSAMPARRAY [rsi+1*SIZEOF_JSAMPARRAY] | 62 mov rbx, JSAMPARRAY [rsi+1*SIZEOF_JSAMPARRAY] |
63 mov rdx, JSAMPARRAY [rsi+2*SIZEOF_JSAMPARRAY] | 63 mov rdx, JSAMPARRAY [rsi+2*SIZEOF_JSAMPARRAY] |
64 lea rdi, [rdi+rcx*SIZEOF_JSAMPROW] | 64 lea rdi, [rdi+rcx*SIZEOF_JSAMPROW] |
65 lea rbx, [rbx+rcx*SIZEOF_JSAMPROW] | 65 lea rbx, [rbx+rcx*SIZEOF_JSAMPROW] |
66 lea rdx, [rdx+rcx*SIZEOF_JSAMPROW] | 66 lea rdx, [rdx+rcx*SIZEOF_JSAMPROW] |
67 | 67 |
68 pop rcx | 68 pop rcx |
69 | 69 |
70 mov rsi, r11 | 70 mov rsi, r11 |
(...skipping 405 matching lines...) Expand 10 before | Expand all | Expand 10 after Loading... |
476 pop rbx | 476 pop rbx |
477 uncollect_args | 477 uncollect_args |
478 mov rsp,rbp ; rsp <- aligned rbp | 478 mov rsp,rbp ; rsp <- aligned rbp |
479 pop rsp ; rsp <- original rbp | 479 pop rsp ; rsp <- original rbp |
480 pop rbp | 480 pop rbp |
481 ret | 481 ret |
482 | 482 |
483 ; For some reason, the OS X linker does not honor the request to align the | 483 ; For some reason, the OS X linker does not honor the request to align the |
484 ; segment unless we do this. | 484 ; segment unless we do this. |
485 align 16 | 485 align 16 |
OLD | NEW |