media/libvpx/vp8/common/x86/iwalsh_sse2.asm@ac0c01689b40 (annotated)
media/libvpx/vp8/common/x86/iwalsh_sse2.asm
Thu, 15 Jan 2015 15:59:08 +0100
- author
- Michael Schloh von Bennewitz <michael@schloh.com>
- date
- Thu, 15 Jan 2015 15:59:08 +0100
- branch
- TOR_BUG_9701
- changeset 10
- ac0c01689b40
- permissions
- -rw-r--r--
Implement a real Private Browsing Mode condition by changing the API/ABI;
This solves Tor bug #9701, complying with disk avoidance documented in
https://www.torproject.org/projects/torbrowser/design/#disk-avoidance.
| michael@0 | 1 | ; |
| michael@0 | 2 | ; Copyright (c) 2010 The WebM project authors. All Rights Reserved. |
| michael@0 | 3 | ; |
| michael@0 | 4 | ; Use of this source code is governed by a BSD-style license |
| michael@0 | 5 | ; that can be found in the LICENSE file in the root of the source |
| michael@0 | 6 | ; tree. An additional intellectual property rights grant can be found |
| michael@0 | 7 | ; in the file PATENTS. All contributing project authors may |
| michael@0 | 8 | ; be found in the AUTHORS file in the root of the source tree. |
| michael@0 | 9 | ; |
| michael@0 | 10 | |
| michael@0 | 11 | |
| michael@0 | 12 | %include "vpx_ports/x86_abi_support.asm" |
| michael@0 | 13 | |
| michael@0 | 14 | ;void vp8_short_inv_walsh4x4_sse2(short *input, short *output) |
| michael@0 | 15 | global sym(vp8_short_inv_walsh4x4_sse2) PRIVATE |
| michael@0 | 16 | sym(vp8_short_inv_walsh4x4_sse2): |
| michael@0 | 17 | push rbp |
| michael@0 | 18 | mov rbp, rsp |
| michael@0 | 19 | SHADOW_ARGS_TO_STACK 2 |
| michael@0 | 20 | ; end prolog |
| michael@0 | 21 | |
| michael@0 | 22 | mov rcx, arg(0) |
| michael@0 | 23 | mov rdx, arg(1) |
| michael@0 | 24 | mov rax, 30003h |
| michael@0 | 25 | |
| michael@0 | 26 | movdqa xmm0, [rcx + 0] ;ip[4] ip[0] |
| michael@0 | 27 | movdqa xmm1, [rcx + 16] ;ip[12] ip[8] |
| michael@0 | 28 | |
| michael@0 | 29 | |
| michael@0 | 30 | pshufd xmm2, xmm1, 4eh ;ip[8] ip[12] |
| michael@0 | 31 | movdqa xmm3, xmm0 ;ip[4] ip[0] |
| michael@0 | 32 | |
| michael@0 | 33 | paddw xmm0, xmm2 ;ip[4]+ip[8] ip[0]+ip[12] aka b1 a1 |
| michael@0 | 34 | psubw xmm3, xmm2 ;ip[4]-ip[8] ip[0]-ip[12] aka c1 d1 |
| michael@0 | 35 | |
| michael@0 | 36 | movdqa xmm4, xmm0 |
| michael@0 | 37 | punpcklqdq xmm0, xmm3 ;d1 a1 |
| michael@0 | 38 | punpckhqdq xmm4, xmm3 ;c1 b1 |
| michael@0 | 39 | |
| michael@0 | 40 | movdqa xmm1, xmm4 ;c1 b1 |
| michael@0 | 41 | paddw xmm4, xmm0 ;dl+cl a1+b1 aka op[4] op[0] |
| michael@0 | 42 | psubw xmm0, xmm1 ;d1-c1 a1-b1 aka op[12] op[8] |
| michael@0 | 43 | |
| michael@0 | 44 | ;~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ |
| michael@0 | 45 | ; 13 12 11 10 03 02 01 00 |
| michael@0 | 46 | ; |
| michael@0 | 47 | ; 33 32 31 30 23 22 21 20 |
| michael@0 | 48 | ; |
| michael@0 | 49 | movdqa xmm3, xmm4 ; 13 12 11 10 03 02 01 00 |
| michael@0 | 50 | punpcklwd xmm4, xmm0 ; 23 03 22 02 21 01 20 00 |
| michael@0 | 51 | punpckhwd xmm3, xmm0 ; 33 13 32 12 31 11 30 10 |
| michael@0 | 52 | movdqa xmm1, xmm4 ; 23 03 22 02 21 01 20 00 |
| michael@0 | 53 | punpcklwd xmm4, xmm3 ; 31 21 11 01 30 20 10 00 |
| michael@0 | 54 | punpckhwd xmm1, xmm3 ; 33 23 13 03 32 22 12 02 |
| michael@0 | 55 | ;~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ |
| michael@0 | 56 | movd xmm0, eax |
| michael@0 | 57 | pshufd xmm2, xmm1, 4eh ;ip[8] ip[12] |
| michael@0 | 58 | movdqa xmm3, xmm4 ;ip[4] ip[0] |
| michael@0 | 59 | |
| michael@0 | 60 | pshufd xmm0, xmm0, 0 ;03 03 03 03 03 03 03 03 |
| michael@0 | 61 | |
| michael@0 | 62 | paddw xmm4, xmm2 ;ip[4]+ip[8] ip[0]+ip[12] aka b1 a1 |
| michael@0 | 63 | psubw xmm3, xmm2 ;ip[4]-ip[8] ip[0]-ip[12] aka c1 d1 |
| michael@0 | 64 | |
| michael@0 | 65 | movdqa xmm5, xmm4 |
| michael@0 | 66 | punpcklqdq xmm4, xmm3 ;d1 a1 |
| michael@0 | 67 | punpckhqdq xmm5, xmm3 ;c1 b1 |
| michael@0 | 68 | |
| michael@0 | 69 | movdqa xmm1, xmm5 ;c1 b1 |
| michael@0 | 70 | paddw xmm5, xmm4 ;dl+cl a1+b1 aka op[4] op[0] |
| michael@0 | 71 | psubw xmm4, xmm1 ;d1-c1 a1-b1 aka op[12] op[8] |
| michael@0 | 72 | |
| michael@0 | 73 | paddw xmm5, xmm0 |
| michael@0 | 74 | paddw xmm4, xmm0 |
| michael@0 | 75 | psraw xmm5, 3 |
| michael@0 | 76 | psraw xmm4, 3 |
| michael@0 | 77 | |
| michael@0 | 78 | movd eax, xmm5 |
| michael@0 | 79 | movd ecx, xmm4 |
| michael@0 | 80 | psrldq xmm5, 4 |
| michael@0 | 81 | psrldq xmm4, 4 |
| michael@0 | 82 | mov word ptr[rdx+32*0], ax |
| michael@0 | 83 | mov word ptr[rdx+32*2], cx |
| michael@0 | 84 | shr eax, 16 |
| michael@0 | 85 | shr ecx, 16 |
| michael@0 | 86 | mov word ptr[rdx+32*4], ax |
| michael@0 | 87 | mov word ptr[rdx+32*6], cx |
| michael@0 | 88 | movd eax, xmm5 |
| michael@0 | 89 | movd ecx, xmm4 |
| michael@0 | 90 | psrldq xmm5, 4 |
| michael@0 | 91 | psrldq xmm4, 4 |
| michael@0 | 92 | mov word ptr[rdx+32*8], ax |
| michael@0 | 93 | mov word ptr[rdx+32*10], cx |
| michael@0 | 94 | shr eax, 16 |
| michael@0 | 95 | shr ecx, 16 |
| michael@0 | 96 | mov word ptr[rdx+32*12], ax |
| michael@0 | 97 | mov word ptr[rdx+32*14], cx |
| michael@0 | 98 | |
| michael@0 | 99 | movd eax, xmm5 |
| michael@0 | 100 | movd ecx, xmm4 |
| michael@0 | 101 | psrldq xmm5, 4 |
| michael@0 | 102 | psrldq xmm4, 4 |
| michael@0 | 103 | mov word ptr[rdx+32*1], ax |
| michael@0 | 104 | mov word ptr[rdx+32*3], cx |
| michael@0 | 105 | shr eax, 16 |
| michael@0 | 106 | shr ecx, 16 |
| michael@0 | 107 | mov word ptr[rdx+32*5], ax |
| michael@0 | 108 | mov word ptr[rdx+32*7], cx |
| michael@0 | 109 | movd eax, xmm5 |
| michael@0 | 110 | movd ecx, xmm4 |
| michael@0 | 111 | mov word ptr[rdx+32*9], ax |
| michael@0 | 112 | mov word ptr[rdx+32*11], cx |
| michael@0 | 113 | shr eax, 16 |
| michael@0 | 114 | shr ecx, 16 |
| michael@0 | 115 | mov word ptr[rdx+32*13], ax |
| michael@0 | 116 | mov word ptr[rdx+32*15], cx |
| michael@0 | 117 | |
| michael@0 | 118 | ; begin epilog |
| michael@0 | 119 | UNSHADOW_ARGS |
| michael@0 | 120 | pop rbp |
| michael@0 | 121 | ret |
