• file
• revisions
• annotate
• diff
• comparison
• raw

media/libvpx/vp8/common/x86/iwalsh_mmx.asm@ac0c01689b40 (annotated)

media/libvpx/vp8/common/x86/iwalsh_mmx.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_mmx(short *input, short *output)
michael@0	15	global sym(vp8_short_inv_walsh4x4_mmx) PRIVATE
michael@0	16	sym(vp8_short_inv_walsh4x4_mmx):
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 rdx, arg(0)
michael@0	23	mov rax, 30003h
michael@0	24
michael@0	25	movq mm0, [rdx + 0] ;ip[0]
michael@0	26	movq mm1, [rdx + 8] ;ip[4]
michael@0	27	movq mm7, rax
michael@0	28
michael@0	29	movq mm2, [rdx + 16] ;ip[8]
michael@0	30	movq mm3, [rdx + 24] ;ip[12]
michael@0	31	punpcklwd mm7, mm7 ;0003000300030003h
michael@0	32	mov rdx, arg(1)
michael@0	33
michael@0	34	movq mm4, mm0
michael@0	35	movq mm5, mm1
michael@0	36
michael@0	37	paddw mm4, mm3 ;ip[0] + ip[12] aka al
michael@0	38	paddw mm5, mm2 ;ip[4] + ip[8] aka bl
michael@0	39
michael@0	40	movq mm6, mm4 ;temp al
michael@0	41	paddw mm4, mm5 ;al + bl
michael@0	42	psubw mm6, mm5 ;al - bl
michael@0	43
michael@0	44	psubw mm0, mm3 ;ip[0] - ip[12] aka d1
michael@0	45	psubw mm1, mm2 ;ip[4] - ip[8] aka c1
michael@0	46
michael@0	47	movq mm5, mm0 ;temp dl
michael@0	48	paddw mm0, mm1 ;dl + cl
michael@0	49	psubw mm5, mm1 ;dl - cl
michael@0	50
michael@0	51	; 03 02 01 00
michael@0	52	; 13 12 11 10
michael@0	53	; 23 22 21 20
michael@0	54	; 33 32 31 30
michael@0	55
michael@0	56	movq mm3, mm4 ; 03 02 01 00
michael@0	57	punpcklwd mm4, mm0 ; 11 01 10 00
michael@0	58	punpckhwd mm3, mm0 ; 13 03 12 02
michael@0	59
michael@0	60	movq mm1, mm6 ; 23 22 21 20
michael@0	61	punpcklwd mm6, mm5 ; 31 21 30 20
michael@0	62	punpckhwd mm1, mm5 ; 33 23 32 22
michael@0	63
michael@0	64	movq mm0, mm4 ; 11 01 10 00
michael@0	65	movq mm2, mm3 ; 13 03 12 02
michael@0	66
michael@0	67	punpckldq mm0, mm6 ; 30 20 10 00 aka ip[0]
michael@0	68	punpckhdq mm4, mm6 ; 31 21 11 01 aka ip[4]
michael@0	69
michael@0	70	punpckldq mm2, mm1 ; 32 22 12 02 aka ip[8]
michael@0	71	punpckhdq mm3, mm1 ; 33 23 13 03 aka ip[12]
michael@0	72	;~~~~~~~~~~~~~~~~~~~~~
michael@0	73	movq mm1, mm0
michael@0	74	movq mm5, mm4
michael@0	75	paddw mm1, mm3 ;ip[0] + ip[12] aka al
michael@0	76	paddw mm5, mm2 ;ip[4] + ip[8] aka bl
michael@0	77
michael@0	78	movq mm6, mm1 ;temp al
michael@0	79	paddw mm1, mm5 ;al + bl
michael@0	80	psubw mm6, mm5 ;al - bl
michael@0	81	paddw mm1, mm7
michael@0	82	paddw mm6, mm7
michael@0	83	psraw mm1, 3
michael@0	84	psraw mm6, 3
michael@0	85
michael@0	86	psubw mm0, mm3 ;ip[0] - ip[12] aka d1
michael@0	87	psubw mm4, mm2 ;ip[4] - ip[8] aka c1
michael@0	88
michael@0	89	movq mm5, mm0 ;temp dl
michael@0	90	paddw mm0, mm4 ;dl + cl
michael@0	91	psubw mm5, mm4 ;dl - cl
michael@0	92	paddw mm0, mm7
michael@0	93	paddw mm5, mm7
michael@0	94	psraw mm0, 3
michael@0	95	psraw mm5, 3
michael@0	96	;~~~~~~~~~~~~~~~~~~~~~
michael@0	97
michael@0	98	movd eax, mm1
michael@0	99	movd ecx, mm0
michael@0	100	psrlq mm0, 32
michael@0	101	psrlq mm1, 32
michael@0	102	mov word ptr[rdx+32*0], ax
michael@0	103	mov word ptr[rdx+32*1], cx
michael@0	104	shr eax, 16
michael@0	105	shr ecx, 16
michael@0	106	mov word ptr[rdx+32*4], ax
michael@0	107	mov word ptr[rdx+32*5], cx
michael@0	108	movd eax, mm1
michael@0	109	movd ecx, mm0
michael@0	110	mov word ptr[rdx+32*8], ax
michael@0	111	mov word ptr[rdx+32*9], cx
michael@0	112	shr eax, 16
michael@0	113	shr ecx, 16
michael@0	114	mov word ptr[rdx+32*12], ax
michael@0	115	mov word ptr[rdx+32*13], cx
michael@0	116
michael@0	117	movd eax, mm6
michael@0	118	movd ecx, mm5
michael@0	119	psrlq mm5, 32
michael@0	120	psrlq mm6, 32
michael@0	121	mov word ptr[rdx+32*2], ax
michael@0	122	mov word ptr[rdx+32*3], cx
michael@0	123	shr eax, 16
michael@0	124	shr ecx, 16
michael@0	125	mov word ptr[rdx+32*6], ax
michael@0	126	mov word ptr[rdx+32*7], cx
michael@0	127	movd eax, mm6
michael@0	128	movd ecx, mm5
michael@0	129	mov word ptr[rdx+32*10], ax
michael@0	130	mov word ptr[rdx+32*11], cx
michael@0	131	shr eax, 16
michael@0	132	shr ecx, 16
michael@0	133	mov word ptr[rdx+32*14], ax
michael@0	134	mov word ptr[rdx+32*15], cx
michael@0	135
michael@0	136	; begin epilog
michael@0	137	UNSHADOW_ARGS
michael@0	138	pop rbp
michael@0	139	ret
michael@0	140