diff options
| author |  diego <diego@b3059339-0415-0410-9bf9-f77b7e298cf2> | 2010-04-12 10:56:17 +0000 |
|---|---|---|
| committer | diego <diego@b3059339-0415-0410-9bf9-f77b7e298cf2> | 2010-04-12 10:56:17 +0000 |
| commit | 7573c29480850d715e2f06cae70f252573098123 (patch) | |
| tree | a5a2f498ad3a19806957e1d7e01f913c1650b33d /mp3lib | |
| parent | 86ea8d4f4abf23672516fa0ca3378aa19c44bf2c (diff) | |
the great MPlayer tab removal: part I
git-svn-id: svn://svn.mplayerhq.hu/mplayer/trunk@31032 b3059339-0415-0410-9bf9-f77b7e298cf2
Diffstat (limited to 'mp3lib')
| -rw-r--r-- | mp3lib/dct36.c | 28 | ||||
| -rw-r--r-- | mp3lib/dct36_3dnow.c | 902 | ||||
| -rw-r--r-- | mp3lib/dct64_3dnow.c | 1666 | ||||
| -rw-r--r-- | mp3lib/dct64_k7.c | 1404 | ||||
| -rw-r--r-- | mp3lib/dct64_mmx.c | 1892 | ||||
| -rw-r--r-- | mp3lib/decod386.c | 10 | ||||
| -rw-r--r-- | mp3lib/decode_i586.c | 6 | ||||
| -rw-r--r-- | mp3lib/decode_mmx.c | 382 | ||||
| -rw-r--r-- | mp3lib/equalizer.c | 119 | ||||
| -rw-r--r-- | mp3lib/l2tables.h | 258 | ||||
| -rw-r--r-- | mp3lib/layer1.c | 2 | ||||
| -rw-r--r-- | mp3lib/layer2.c | 6 | ||||
| -rw-r--r-- | mp3lib/layer3.c | 28 | ||||
| -rw-r--r-- | mp3lib/mpg123.h | 4 | ||||
| -rw-r--r-- | mp3lib/sr1.c | 36 |
15 files changed, 3371 insertions, 3372 deletions
diff --git a/mp3lib/dct36.c b/mp3lib/dct36.c index b884bb949d..22ec273277 100644 --- a/mp3lib/dct36.c +++ b/mp3lib/dct36.c @@ -189,21 +189,21 @@ static void dct36(real *inbuf,real *o1,real *o2,real *wintab,real *tsbuf) ts[SBLIMIT*(8-(v))] = out1[8-(v)] + sum0 * w[8-(v)]; \ ts[SBLIMIT*(9+(v))] = out1[9+(v)] + sum0 * w[9+(v)]; #define MACRO1(v) { \ - real sum0,sum1; \ + real sum0, sum1; \ sum0 = tmp1a + tmp2a; \ - sum1 = (tmp1b + tmp2b) * tfcos36[(v)]; \ - MACRO0(v); } + sum1 = (tmp1b + tmp2b) * tfcos36[(v)]; \ + MACRO0(v); } #define MACRO2(v) { \ - real sum0,sum1; \ + real sum0, sum1; \ sum0 = tmp2a - tmp1a; \ sum1 = (tmp2b - tmp1b) * tfcos36[(v)]; \ - MACRO0(v); } + MACRO0(v); } register const real *c = COS9; register real *out2 = o2; - register real *w = wintab; - register real *out1 = o1; - register real *ts = tsbuf; + register real *w = wintab; + register real *out1 = o1; + register real *ts = tsbuf; real ta33,ta66,tb33,tb66; @@ -256,12 +256,12 @@ static void dct36(real *inbuf,real *o1,real *o2,real *wintab,real *tsbuf) MACRO2(5); } - { - real sum0,sum1; - sum0 = in[2*0+0] - in[2*2+0] + in[2*4+0] - in[2*6+0] + in[2*8+0]; - sum1 = (in[2*0+1] - in[2*2+1] + in[2*4+1] - in[2*6+1] + in[2*8+1] ) * tfcos36[4]; - MACRO0(4); - } + { + real sum0,sum1; + sum0 = in[2*0+0] - in[2*2+0] + in[2*4+0] - in[2*6+0] + in[2*8+0]; + sum1 = (in[2*0+1] - in[2*2+1] + in[2*4+1] - in[2*6+1] + in[2*8+1] ) * tfcos36[4]; + MACRO0(4); + } } #endif diff --git a/mp3lib/dct36_3dnow.c b/mp3lib/dct36_3dnow.c index 8263639149..68cc33014b 100644 --- a/mp3lib/dct36_3dnow.c +++ b/mp3lib/dct36_3dnow.c @@ -40,462 +40,462 @@ void dct36_3dnow(real *inbuf, real *o1, #endif { __asm__ volatile( - "movq (%%eax),%%mm0\n\t" - "movq 4(%%eax),%%mm1\n\t" - "pfadd %%mm1,%%mm0\n\t" - "movq %%mm0,4(%%eax)\n\t" - "psrlq $32,%%mm1\n\t" - "movq 12(%%eax),%%mm2\n\t" - "punpckldq %%mm2,%%mm1\n\t" - "pfadd %%mm2,%%mm1\n\t" - "movq %%mm1,12(%%eax)\n\t" - "psrlq $32,%%mm2\n\t" - "movq 20(%%eax),%%mm3\n\t" - "punpckldq %%mm3,%%mm2\n\t" - "pfadd %%mm3,%%mm2\n\t" - "movq %%mm2,20(%%eax)\n\t" - "psrlq $32,%%mm3\n\t" - "movq 28(%%eax),%%mm4\n\t" - "punpckldq %%mm4,%%mm3\n\t" - "pfadd %%mm4,%%mm3\n\t" - "movq %%mm3,28(%%eax)\n\t" - "psrlq $32,%%mm4\n\t" - "movq 36(%%eax),%%mm5\n\t" - "punpckldq %%mm5,%%mm4\n\t" - "pfadd %%mm5,%%mm4\n\t" - "movq %%mm4,36(%%eax)\n\t" - "psrlq $32,%%mm5\n\t" - "movq 44(%%eax),%%mm6\n\t" - "punpckldq %%mm6,%%mm5\n\t" - "pfadd %%mm6,%%mm5\n\t" - "movq %%mm5,44(%%eax)\n\t" - "psrlq $32,%%mm6\n\t" - "movq 52(%%eax),%%mm7\n\t" - "punpckldq %%mm7,%%mm6\n\t" - "pfadd %%mm7,%%mm6\n\t" - "movq %%mm6,52(%%eax)\n\t" - "psrlq $32,%%mm7\n\t" - "movq 60(%%eax),%%mm0\n\t" - "punpckldq %%mm0,%%mm7\n\t" - "pfadd %%mm0,%%mm7\n\t" - "movq %%mm7,60(%%eax)\n\t" - "psrlq $32,%%mm0\n\t" - "movd 68(%%eax),%%mm1\n\t" - "pfadd %%mm1,%%mm0\n\t" - "movd %%mm0,68(%%eax)\n\t" - "movd 4(%%eax),%%mm0\n\t" - "movd 12(%%eax),%%mm1\n\t" - "punpckldq %%mm1,%%mm0\n\t" - "punpckldq 20(%%eax),%%mm1\n\t" - "pfadd %%mm1,%%mm0\n\t" - "movd %%mm0,12(%%eax)\n\t" - "psrlq $32,%%mm0\n\t" - "movd %%mm0,20(%%eax)\n\t" - "psrlq $32,%%mm1\n\t" - "movd 28(%%eax),%%mm2\n\t" - "punpckldq %%mm2,%%mm1\n\t" - "punpckldq 36(%%eax),%%mm2\n\t" - "pfadd %%mm2,%%mm1\n\t" - "movd %%mm1,28(%%eax)\n\t" - "psrlq $32,%%mm1\n\t" - "movd %%mm1,36(%%eax)\n\t" - "psrlq $32,%%mm2\n\t" - "movd 44(%%eax),%%mm3\n\t" - "punpckldq %%mm3,%%mm2\n\t" - "punpckldq 52(%%eax),%%mm3\n\t" - "pfadd %%mm3,%%mm2\n\t" - "movd %%mm2,44(%%eax)\n\t" - "psrlq $32,%%mm2\n\t" - "movd %%mm2,52(%%eax)\n\t" - "psrlq $32,%%mm3\n\t" - "movd 60(%%eax),%%mm4\n\t" - "punpckldq %%mm4,%%mm3\n\t" - "punpckldq 68(%%eax),%%mm4\n\t" - "pfadd %%mm4,%%mm3\n\t" - "movd %%mm3,60(%%eax)\n\t" - "psrlq $32,%%mm3\n\t" - "movd %%mm3,68(%%eax)\n\t" + "movq (%%eax),%%mm0\n\t" + "movq 4(%%eax),%%mm1\n\t" + "pfadd %%mm1,%%mm0\n\t" + "movq %%mm0,4(%%eax)\n\t" + "psrlq $32,%%mm1\n\t" + "movq 12(%%eax),%%mm2\n\t" + "punpckldq %%mm2,%%mm1\n\t" + "pfadd %%mm2,%%mm1\n\t" + "movq %%mm1,12(%%eax)\n\t" + "psrlq $32,%%mm2\n\t" + "movq 20(%%eax),%%mm3\n\t" + "punpckldq %%mm3,%%mm2\n\t" + "pfadd %%mm3,%%mm2\n\t" + "movq %%mm2,20(%%eax)\n\t" + "psrlq $32,%%mm3\n\t" + "movq 28(%%eax),%%mm4\n\t" + "punpckldq %%mm4,%%mm3\n\t" + "pfadd %%mm4,%%mm3\n\t" + "movq %%mm3,28(%%eax)\n\t" + "psrlq $32,%%mm4\n\t" + "movq 36(%%eax),%%mm5\n\t" + "punpckldq %%mm5,%%mm4\n\t" + "pfadd %%mm5,%%mm4\n\t" + "movq %%mm4,36(%%eax)\n\t" + "psrlq $32,%%mm5\n\t" + "movq 44(%%eax),%%mm6\n\t" + "punpckldq %%mm6,%%mm5\n\t" + "pfadd %%mm6,%%mm5\n\t" + "movq %%mm5,44(%%eax)\n\t" + "psrlq $32,%%mm6\n\t" + "movq 52(%%eax),%%mm7\n\t" + "punpckldq %%mm7,%%mm6\n\t" + "pfadd %%mm7,%%mm6\n\t" + "movq %%mm6,52(%%eax)\n\t" + "psrlq $32,%%mm7\n\t" + "movq 60(%%eax),%%mm0\n\t" + "punpckldq %%mm0,%%mm7\n\t" + "pfadd %%mm0,%%mm7\n\t" + "movq %%mm7,60(%%eax)\n\t" + "psrlq $32,%%mm0\n\t" + "movd 68(%%eax),%%mm1\n\t" + "pfadd %%mm1,%%mm0\n\t" + "movd %%mm0,68(%%eax)\n\t" + "movd 4(%%eax),%%mm0\n\t" + "movd 12(%%eax),%%mm1\n\t" + "punpckldq %%mm1,%%mm0\n\t" + "punpckldq 20(%%eax),%%mm1\n\t" + "pfadd %%mm1,%%mm0\n\t" + "movd %%mm0,12(%%eax)\n\t" + "psrlq $32,%%mm0\n\t" + "movd %%mm0,20(%%eax)\n\t" + "psrlq $32,%%mm1\n\t" + "movd 28(%%eax),%%mm2\n\t" + "punpckldq %%mm2,%%mm1\n\t" + "punpckldq 36(%%eax),%%mm2\n\t" + "pfadd %%mm2,%%mm1\n\t" + "movd %%mm1,28(%%eax)\n\t" + "psrlq $32,%%mm1\n\t" + "movd %%mm1,36(%%eax)\n\t" + "psrlq $32,%%mm2\n\t" + "movd 44(%%eax),%%mm3\n\t" + "punpckldq %%mm3,%%mm2\n\t" + "punpckldq 52(%%eax),%%mm3\n\t" + "pfadd %%mm3,%%mm2\n\t" + "movd %%mm2,44(%%eax)\n\t" + "psrlq $32,%%mm2\n\t" + "movd %%mm2,52(%%eax)\n\t" + "psrlq $32,%%mm3\n\t" + "movd 60(%%eax),%%mm4\n\t" + "punpckldq %%mm4,%%mm3\n\t" + "punpckldq 68(%%eax),%%mm4\n\t" + "pfadd %%mm4,%%mm3\n\t" + "movd %%mm3,60(%%eax)\n\t" + "psrlq $32,%%mm3\n\t" + "movd %%mm3,68(%%eax)\n\t" - "movq 24(%%eax),%%mm0\n\t" - "movq 48(%%eax),%%mm1\n\t" - "movd "MANGLE(COS9)"+12,%%mm2\n\t" - "punpckldq %%mm2,%%mm2\n\t" - "movd "MANGLE(COS9)"+24,%%mm3\n\t" - "punpckldq %%mm3,%%mm3\n\t" - "pfmul %%mm2,%%mm0\n\t" - "pfmul %%mm3,%%mm1\n\t" - "pushl %%eax\n\t" - "movl $1,%%eax\n\t" - "movd %%eax,%%mm7\n\t" - "pi2fd %%mm7,%%mm7\n\t" - "popl %%eax\n\t" - "movq 8(%%eax),%%mm2\n\t" - "movd "MANGLE(COS9)"+4,%%mm3\n\t" - "punpckldq %%mm3,%%mm3\n\t" - "pfmul %%mm3,%%mm2\n\t" - "pfadd %%mm0,%%mm2\n\t" - "movq 40(%%eax),%%mm3\n\t" - "movd "MANGLE(COS9)"+20,%%mm4\n\t" - "punpckldq %%mm4,%%mm4\n\t" - "pfmul %%mm4,%%mm3\n\t" - "pfadd %%mm3,%%mm2\n\t" - "movq 56(%%eax),%%mm3\n\t" - "movd "MANGLE(COS9)"+28,%%mm4\n\t" - "punpckldq %%mm4,%%mm4\n\t" - "pfmul %%mm4,%%mm3\n\t" - "pfadd %%mm3,%%mm2\n\t" - "movq (%%eax),%%mm3\n\t" - "movq 16(%%eax),%%mm4\n\t" - "movd "MANGLE(COS9)"+8,%%mm5\n\t" - "punpckldq %%mm5,%%mm5\n\t" - "pfmul %%mm5,%%mm4\n\t" - "pfadd %%mm4,%%mm3\n\t" - "movq 32(%%eax),%%mm4\n\t" - "movd "MANGLE(COS9)"+16,%%mm5\n\t" - "punpckldq %%mm5,%%mm5\n\t" - "pfmul %%mm5,%%mm4\n\t" - "pfadd %%mm4,%%mm3\n\t" - "pfadd %%mm1,%%mm3\n\t" - "movq 64(%%eax),%%mm4\n\t" - "movd "MANGLE(COS9)"+32,%%mm5\n\t" - "punpckldq %%mm5,%%mm5\n\t" - "pfmul %%mm5,%%mm4\n\t" - "pfadd %%mm4,%%mm3\n\t" - "movq %%mm2,%%mm4\n\t" - "pfadd %%mm3,%%mm4\n\t" - "movq %%mm7,%%mm5\n\t" - "punpckldq "MANGLE(tfcos36)"+0,%%mm5\n\t" - "pfmul %%mm5,%%mm4\n\t" - "movq %%mm4,%%mm5\n\t" - "pfacc %%mm5,%%mm5\n\t" - "movd 108(%%edx),%%mm6\n\t" - "punpckldq 104(%%edx),%%mm6\n\t" - "pfmul %%mm6,%%mm5\n\t" + "movq 24(%%eax),%%mm0\n\t" + "movq 48(%%eax),%%mm1\n\t" + "movd "MANGLE(COS9)"+12,%%mm2\n\t" + "punpckldq %%mm2,%%mm2\n\t" + "movd "MANGLE(COS9)"+24,%%mm3\n\t" + "punpckldq %%mm3,%%mm3\n\t" + "pfmul %%mm2,%%mm0\n\t" + "pfmul %%mm3,%%mm1\n\t" + "pushl %%eax\n\t" + "movl $1,%%eax\n\t" + "movd %%eax,%%mm7\n\t" + "pi2fd %%mm7,%%mm7\n\t" + "popl %%eax\n\t" + "movq 8(%%eax),%%mm2\n\t" + "movd "MANGLE(COS9)"+4,%%mm3\n\t" + "punpckldq %%mm3,%%mm3\n\t" + "pfmul %%mm3,%%mm2\n\t" + "pfadd %%mm0,%%mm2\n\t" + "movq 40(%%eax),%%mm3\n\t" + "movd "MANGLE(COS9)"+20,%%mm4\n\t" + "punpckldq %%mm4,%%mm4\n\t" + "pfmul %%mm4,%%mm3\n\t" + "pfadd %%mm3,%%mm2\n\t" + "movq 56(%%eax),%%mm3\n\t" + "movd "MANGLE(COS9)"+28,%%mm4\n\t" + "punpckldq %%mm4,%%mm4\n\t" + "pfmul %%mm4,%%mm3\n\t" + "pfadd %%mm3,%%mm2\n\t" + "movq (%%eax),%%mm3\n\t" + "movq 16(%%eax),%%mm4\n\t" + "movd "MANGLE(COS9)"+8,%%mm5\n\t" + "punpckldq %%mm5,%%mm5\n\t" + "pfmul %%mm5,%%mm4\n\t" + "pfadd %%mm4,%%mm3\n\t" + "movq 32(%%eax),%%mm4\n\t" + "movd "MANGLE(COS9)"+16,%%mm5\n\t" + "punpckldq %%mm5,%%mm5\n\t" + "pfmul %%mm5,%%mm4\n\t" + "pfadd %%mm4,%%mm3\n\t" + "pfadd %%mm1,%%mm3\n\t" + "movq 64(%%eax),%%mm4\n\t" + "movd "MANGLE(COS9)"+32,%%mm5\n\t" + "punpckldq %%mm5,%%mm5\n\t" + "pfmul %%mm5,%%mm4\n\t" + "pfadd %%mm4,%%mm3\n\t" + "movq %%mm2,%%mm4\n\t" + "pfadd %%mm3,%%mm4\n\t" + "movq %%mm7,%%mm5\n\t" + "punpckldq "MANGLE(tfcos36)"+0,%%mm5\n\t" + "pfmul %%mm5,%%mm4\n\t" + "movq %%mm4,%%mm5\n\t" + "pfacc %%mm5,%%mm5\n\t" + "movd 108(%%edx),%%mm6\n\t" + "punpckldq 104(%%edx),%%mm6\n\t" + "pfmul %%mm6,%%mm5\n\t" #ifdef DCT36_OPTIMIZE_FOR_K7 - "pswapd %%mm5,%%mm5\n\t" - "movq %%mm5,32(%%ecx)\n\t" + "pswapd %%mm5,%%mm5\n\t" + "movq %%mm5,32(%%ecx)\n\t" #else - "movd %%mm5,36(%%ecx)\n\t" - "psrlq $32,%%mm5\n\t" - "movd %%mm5,32(%%ecx)\n\t" + "movd %%mm5,36(%%ecx)\n\t" + "psrlq $32,%%mm5\n\t" + "movd %%mm5,32(%%ecx)\n\t" #endif - "movq %%mm4,%%mm6\n\t" - "punpckldq %%mm6,%%mm5\n\t" - "pfsub %%mm6,%%mm5\n\t" - "punpckhdq %%mm5,%%mm5\n\t" - "movd 32(%%edx),%%mm6\n\t" - "punpckldq 36(%%edx),%%mm6\n\t" - "pfmul %%mm6,%%mm5\n\t" - "movd 32(%%esi),%%mm6\n\t" - "punpckldq 36(%%esi),%%mm6\n\t" - "pfadd %%mm6,%%mm5\n\t" - "movd %%mm5,1024(%%ebx)\n\t" - "psrlq $32,%%mm5\n\t" - "movd %%mm5,1152(%%ebx)\n\t" - "movq %%mm3,%%mm4\n\t" - "pfsub %%mm2,%%mm4\n\t" - "movq %%mm7,%%mm5\n\t" - "punpckldq "MANGLE(tfcos36)"+32,%%mm5\n\t" - "pfmul %%mm5,%%mm4\n\t" - "movq %%mm4,%%mm5\n\t" - "pfacc %%mm5,%%mm5\n\t" - "movd 140(%%edx),%%mm6\n\t" - "punpckldq 72(%%edx),%%mm6\n\t" - "pfmul %%mm6,%%mm5\n\t" - "movd %%mm5,68(%%ecx)\n\t" - "psrlq $32,%%mm5\n\t" - "movd %%mm5,0(%%ecx)\n\t" - "movq %%mm4,%%mm6\n\t" - "punpckldq %%mm6,%%mm5\n\t" - "pfsub %%mm6,%%mm5\n\t" - "punpckhdq %%mm5,%%mm5\n\t" - "movd 0(%%edx),%%mm6\n\t" - "punpckldq 68(%%edx),%%mm6\n\t" - "pfmul %%mm6,%%mm5\n\t" - "movd 0(%%esi),%%mm6\n\t" - "punpckldq 68(%%esi),%%mm6\n\t" - "pfadd %%mm6,%%mm5\n\t" - "movd %%mm5,0(%%ebx)\n\t" - "psrlq $32,%%mm5\n\t" - "movd %%mm5,2176(%%ebx)\n\t" - "movq 8(%%eax),%%mm2\n\t" - "movq 40(%%eax),%%mm3\n\t" - "pfsub %%mm3,%%mm2\n\t" - "movq 56(%%eax),%%mm3\n\t" - "pfsub %%mm3,%%mm2\n\t" - "movd "MANGLE(COS9)"+12,%%mm3\n\t" - "punpckldq %%mm3,%%mm3\n\t" - "pfmul %%mm3,%%mm2\n\t" - "movq 16(%%eax),%%mm3\n\t" - "movq 32(%%eax),%%mm4\n\t" - "pfsub %%mm4,%%mm3\n\t" - "movq 64(%%eax),%%mm4\n\t" - "pfsub %%mm4,%%mm3\n\t" - "movd "MANGLE(COS9)"+24,%%mm4\n\t" - "punpckldq %%mm4,%%mm4\n\t" - "pfmul %%mm4,%%mm3\n\t" - "movq 48(%%eax),%%mm4\n\t" - "pfsub %%mm4,%%mm3\n\t" - "movq (%%eax),%%mm4\n\t" - "pfadd %%mm4,%%mm3\n\t" - "movq %%mm2,%%mm4\n\t" - "pfadd %%mm3,%%mm4\n\t" - "movq %%mm7,%%mm5\n\t" - "punpckldq "MANGLE(tfcos36)"+4,%%mm5\n\t" - "pfmul %%mm5,%%mm4\n\t" - "movq %%mm4,%%mm5\n\t" - "pfacc %%mm5,%%mm5\n\t" - "movd 112(%%edx),%%mm6\n\t" - "punpckldq 100(%%edx),%%mm6\n\t" - "pfmul %%mm6,%%mm5\n\t" - "movd %%mm5,40(%%ecx)\n\t" - "psrlq $32,%%mm5\n\t" - "movd %%mm5,28(%%ecx)\n\t" - "movq %%mm4,%%mm6\n\t" - "punpckldq %%mm6,%%mm5\n\t" - "pfsub %%mm6,%%mm5\n\t" - "punpckhdq %%mm5,%%mm5\n\t" - "movd 28(%%edx),%%mm6\n\t" - "punpckldq 40(%%edx),%%mm6\n\t" - "pfmul %%mm6,%%mm5\n\t" - "movd 28(%%esi),%%mm6\n\t" - "punpckldq 40(%%esi),%%mm6\n\t" - "pfadd %%mm6,%%mm5\n\t" - "movd %%mm5,896(%%ebx)\n\t" - "psrlq $32,%%mm5\n\t" - "movd %%mm5,1280(%%ebx)\n\t" - "movq %%mm3,%%mm4\n\t" - "pfsub %%mm2,%%mm4\n\t" - "movq %%mm7,%%mm5\n\t" - "punpckldq "MANGLE(tfcos36)"+28,%%mm5\n\t" - "pfmul %%mm5,%%mm4\n\t" - "movq %%mm4,%%mm5\n\t" - "pfacc %%mm5,%%mm5\n\t" - "movd 136(%%edx),%%mm6\n\t" - "punpckldq 76(%%edx),%%mm6\n\t" - "pfmul %%mm6,%%mm5\n\t" - "movd %%mm5,64(%%ecx)\n\t" - "psrlq $32,%%mm5\n\t" - "movd %%mm5,4(%%ecx)\n\t" - "movq %%mm4,%%mm6\n\t" - "punpckldq %%mm6,%%mm5\n\t" - "pfsub %%mm6,%%mm5\n\t" - "punpckhdq %%mm5,%%mm5\n\t" - "movd 4(%%edx),%%mm6\n\t" - "punpckldq 64(%%edx),%%mm6\n\t" - "pfmul %%mm6,%%mm5\n\t" - "movd 4(%%esi),%%mm6\n\t" - "punpckldq 64(%%esi),%%mm6\n\t" - "pfadd %%mm6,%%mm5\n\t" - "movd %%mm5,128(%%ebx)\n\t" - "psrlq $32,%%mm5\n\t" - "movd %%mm5,2048(%%ebx)\n\t" + "movq %%mm4,%%mm6\n\t" + "punpckldq %%mm6,%%mm5\n\t" + "pfsub %%mm6,%%mm5\n\t" + "punpckhdq %%mm5,%%mm5\n\t" + "movd 32(%%edx),%%mm6\n\t" + "punpckldq 36(%%edx),%%mm6\n\t" + "pfmul %%mm6,%%mm5\n\t" + "movd 32(%%esi),%%mm6\n\t" + "punpckldq 36(%%esi),%%mm6\n\t" + "pfadd %%mm6,%%mm5\n\t" + "movd %%mm5,1024(%%ebx)\n\t" + "psrlq $32,%%mm5\n\t" + "movd %%mm5,1152(%%ebx)\n\t" + "movq %%mm3,%%mm4\n\t" + "pfsub %%mm2,%%mm4\n\t" + "movq %%mm7,%%mm5\n\t" + "punpckldq "MANGLE(tfcos36)"+32,%%mm5\n\t" + "pfmul %%mm5,%%mm4\n\t" + "movq %%mm4,%%mm5\n\t" + "pfacc %%mm5,%%mm5\n\t" + "movd 140(%%edx),%%mm6\n\t" + "punpckldq 72(%%edx),%%mm6\n\t" + "pfmul %%mm6,%%mm5\n\t" + "movd %%mm5,68(%%ecx)\n\t" + "psrlq $32,%%mm5\n\t" + "movd %%mm5,0(%%ecx)\n\t" + "movq %%mm4,%%mm6\n\t" + "punpckldq %%mm6,%%mm5\n\t" + "pfsub %%mm6,%%mm5\n\t" + "punpckhdq %%mm5,%%mm5\n\t" + "movd 0(%%edx),%%mm6\n\t" + "punpckldq 68(%%edx),%%mm6\n\t" + "pfmul %%mm6,%%mm5\n\t" + "movd 0(%%esi),%%mm6\n\t" + "punpckldq 68(%%esi),%%mm6\n\t" + "pfadd %%mm6,%%mm5\n\t" + "movd %%mm5,0(%%ebx)\n\t" + "psrlq $32,%%mm5\n\t" + "movd %%mm5,2176(%%ebx)\n\t" + "movq 8(%%eax),%%mm2\n\t" + "movq 40(%%eax),%%mm3\n\t" + "pfsub %%mm3,%%mm2\n\t" + "movq 56(%%eax),%%mm3\n\t" + "pfsub %%mm3,%%mm2\n\t" + "movd "MANGLE(COS9)"+12,%%mm3\n\t" + "punpckldq %%mm3,%%mm3\n\t" + "pfmul %%mm3,%%mm2\n\t" + "movq 16(%%eax),%%mm3\n\t" + "movq 32(%%eax),%%mm4\n\t" + "pfsub %%mm4,%%mm3\n\t" + "movq 64(%%eax),%%mm4\n\t" + "pfsub %%mm4,%%mm3\n\t" + "movd "MANGLE(COS9)"+24,%%mm4\n\t" + "punpckldq %%mm4,%%mm4\n\t" + "pfmul %%mm4,%%mm3\n\t" + "movq 48(%%eax),%%mm4\n\t" + "pfsub %%mm4,%%mm3\n\t" + "movq (%%eax),%%mm4\n\t" + "pfadd %%mm4,%%mm3\n\t" + "movq %%mm2,%%mm4\n\t" + "pfadd %%mm3,%%mm4\n\t" + "movq %%mm7,%%mm5\n\t" + "punpckldq "MANGLE(tfcos36)"+4,%%mm5\n\t" + "pfmul %%mm5,%%mm4\n\t" + "movq %%mm4,%%mm5\n\t" + "pfacc %%mm5,%%mm5\n\t" + "movd 112(%%edx),%%mm6\n\t" + "punpckldq 100(%%edx),%%mm6\n\t" + "pfmul %%mm6,%%mm5\n\t" + "movd %%mm5,40(%%ecx)\n\t" + "psrlq $32,%%mm5\n\t" + "movd %%mm5,28(%%ecx)\n\t" + "movq %%mm4,%%mm6\n\t" + "punpckldq %%mm6,%%mm5\n\t" + "pfsub %%mm6,%%mm5\n\t" + "punpckhdq %%mm5,%%mm5\n\t" + "movd 28(%%edx),%%mm6\n\t" + "punpckldq 40(%%edx),%%mm6\n\t" + "pfmul %%mm6,%%mm5\n\t" + "movd 28(%%esi),%%mm6\n\t" + "punpckldq 40(%%esi),%%mm6\n\t" + "pfadd %%mm6,%%mm5\n\t" + "movd %%mm5,896(%%ebx)\n\t" + "psrlq $32,%%mm5\n\t" + "movd %%mm5,1280(%%ebx)\n\t" + "movq %%mm3,%%mm4\n\t" + "pfsub %%mm2,%%mm4\n\t" + "movq %%mm7,%%mm5\n\t" + "punpckldq "MANGLE(tfcos36)"+28,%%mm5\n\t" + "pfmul %%mm5,%%mm4\n\t" + "movq %%mm4,%%mm5\n\t" + "pfacc %%mm5,%%mm5\n\t" + "movd 136(%%edx),%%mm6\n\t" + "punpckldq 76(%%edx),%%mm6\n\t" + "pfmul %%mm6,%%mm5\n\t" + "movd %%mm5,64(%%ecx)\n\t" + "psrlq $32,%%mm5\n\t" + "movd %%mm5,4(%%ecx)\n\t" + "movq %%mm4,%%mm6\n\t" + "punpckldq %%mm6,%%mm5\n\t" + "pfsub %%mm6,%%mm5\n\t" + "punpckhdq %%mm5,%%mm5\n\t" + "movd 4(%%edx),%%mm6\n\t" + "punpckldq 64(%%edx),%%mm6\n\t" + "pfmul %%mm6,%%mm5\n\t" + "movd 4(%%esi),%%mm6\n\t" + "punpckldq 64(%%esi),%%mm6\n\t" + "pfadd %%mm6,%%mm5\n\t" + "movd %%mm5,128(%%ebx)\n\t" + "psrlq $32,%%mm5\n\t" + "movd %%mm5,2048(%%ebx)\n\t" - "movq 8(%%eax),%%mm2\n\t" - "movd "MANGLE(COS9)"+20,%%mm3\n\t" - "punpckldq %%mm3,%%mm3\n\t" - "pfmul %%mm3,%%mm2\n\t" - "pfsub %%mm0,%%mm2\n\t" - "movq 40(%%eax),%%mm3\n\t" - "movd "MANGLE(COS9)"+28,%%mm4\n\t" - "punpckldq %%mm4,%%mm4\n\t" - "pfmul %%mm4,%%mm3\n\t" - "pfsub %%mm3,%%mm2\n\t" - "movq 56(%%eax),%%mm3\n\t" - "movd "MANGLE(COS9)"+4,%%mm4\n\t" - "punpckldq %%mm4,%%mm4\n\t" - "pfmul %%mm4,%%mm3\n\t" - "pfadd %%mm3,%%mm2\n\t" - "movq (%%eax),%%mm3\n\t" - "movq 16(%%eax),%%mm4\n\t" - "movd "MANGLE(COS9)"+32,%%mm5\n\t" - "punpckldq %%mm5,%%mm5\n\t" - "pfmul %%mm5,%%mm4\n\t" - "pfsub %%mm4,%%mm3\n\t" - "movq 32(%%eax),%%mm4\n\t" - "movd "MANGLE(COS9)"+8,%%mm5\n\t" - "punpckldq %%mm5,%%mm5\n\t" - "pfmul %%mm5,%%mm4\n\t" - "pfsub %%mm4,%%mm3\n\t" - "pfadd %%mm1,%%mm3\n\t" - "movq 64(%%eax),%%mm4\n\t" - "movd "MANGLE(COS9)"+16,%%mm5\n\t" - "punpckldq %%mm5,%%mm5\n\t" - "pfmul %%mm5,%%mm4\n\t" - "pfadd %%mm4,%%mm3\n\t" - "movq %%mm2,%%mm4\n\t" - "pfadd %%mm3,%%mm4\n\t" - "movq %%mm7,%%mm5\n\t" - "punpckldq "MANGLE(tfcos36)"+8,%%mm5\n\t" - "pfmul %%mm5,%%mm4\n\t" - "movq %%mm4,%%mm5\n\t" - "pfacc %%mm5,%%mm5\n\t" - "movd 116(%%edx),%%mm6\n\t" - "punpckldq 96(%%edx),%%mm6\n\t" - "pfmul %%mm6,%%mm5\n\t" - "movd %%mm5,44(%%ecx)\n\t" - "psrlq $32,%%mm5\n\t" - "movd %%mm5,24(%%ecx)\n\t" - "movq %%mm4,%%mm6\n\t" - "punpckldq %%mm6,%%mm5\n\t" - "pfsub %%mm6,%%mm5\n\t" - "punpckhdq %%mm5,%%mm5\n\t" - "movd 24(%%edx),%%mm6\n\t" - "punpckldq 44(%%edx),%%mm6\n\t" - "pfmul %%mm6,%%mm5\n\t" - "movd 24(%%esi),%%mm6\n\t" - "punpckldq 44(%%esi),%%mm6\n\t" - "pfadd %%mm6,%%mm5\n\t" - "movd %%mm5,768(%%ebx)\n\t" - "psrlq $32,%%mm5\n\t" - "movd %%mm5,1408(%%ebx)\n\t" - "movq %%mm3,%%mm4\n\t" - "pfsub %%mm2,%%mm4\n\t" - "movq %%mm7,%%mm5\n\t" - "punpckldq "MANGLE(tfcos36)"+24,%%mm5\n\t" - "pfmul %%mm5,%%mm4\n\t" - "movq %%mm4,%%mm5\n\t" - "pfacc %%mm5,%%mm5\n\t" - "movd 132(%%edx),%%mm6\n\t" - "punpckldq 80(%%edx),%%mm6\n\t" - "pfmul %%mm6,%%mm5\n\t" - "movd %%mm5,60(%%ecx)\n\t" - "psrlq $32,%%mm5\n\t" - "movd %%mm5,8(%%ecx)\n\t" - "movq %%mm4,%%mm6\n\t" - "punpckldq %%mm6,%%mm5\n\t" - "pfsub %%mm6,%%mm5\n\t" - "punpckhdq %%mm5,%%mm5\n\t" - "movd 8(%%edx),%%mm6\n\t" - "punpckldq 60(%%edx),%%mm6\n\t" - "pfmul %%mm6,%%mm5\n\t" - "movd 8(%%esi),%%mm6\n\t" - "punpckldq 60(%%esi),%%mm6\n\t" - "pfadd %%mm6,%%mm5\n\t" - "movd %%mm5,256(%%ebx)\n\t" - "psrlq $32,%%mm5\n\t" - "movd %%mm5,1920(%%ebx)\n\t" - "movq 8(%%eax),%%mm2\n\t" - "movd "MANGLE(COS9)"+28,%%mm3\n\t" - "punpckldq %%mm3,%%mm3\n\t" - "pfmul %%mm3,%%mm2\n\t" - "pfsub %%mm0,%%mm2\n\t" - "movq 40(%%eax),%%mm3\n\t" - "movd "MANGLE(COS9)"+4,%%mm4\n\t" - "punpckldq %%mm4,%%mm4\n\t" - "pfmul %%mm4,%%mm3\n\t" - "pfadd %%mm3,%%mm2\n\t" - "movq 56(%%eax),%%mm3\n\t" - "movd "MANGLE(COS9)"+20,%%mm4\n\t" - "punpckldq %%mm4,%%mm4\n\t" - "pfmul %%mm4,%%mm3\n\t" - "pfsub %%mm3,%%mm2\n\t" - "movq (%%eax),%%mm3\n\t" - "movq 16(%%eax),%%mm4\n\t" - "movd "MANGLE(COS9)"+16,%%mm5\n\t" - "punpckldq %%mm5,%%mm5\n\t" - "pfmul %%mm5,%%mm4\n\t" - "pfsub %%mm4,%%mm3\n\t" - "movq 32(%%eax),%%mm4\n\t" - "movd "MANGLE(COS9)"+32,%%mm5\n\t" - "punpckldq %%mm5,%%mm5\n\t" - "pfmul %%mm5,%%mm4\n\t" - "pfadd %%mm4,%%mm3\n\t" - "pfadd %%mm1,%%mm3\n\t" - "movq 64(%%eax),%%mm4\n\t" - "movd "MANGLE(COS9)"+8,%%mm5\n\t" - "punpckldq %%mm5,%%mm5\n\t" - "pfmul %%mm5,%%mm4\n\t" - "pfsub %%mm4,%%mm3\n\t" - "movq %%mm2,%%mm4\n\t" - "pfadd %%mm3,%%mm4\n\t" - "movq %%mm7,%%mm5\n\t" - "punpckldq "MANGLE(tfcos36)"+12,%%mm5\n\t" - "pfmul %%mm5,%%mm4\n\t" - "movq %%mm4,%%mm5\n\t" - "pfacc %%mm5,%%mm5\n\t" - "movd 120(%%edx),%%mm6\n\t" - "punpckldq 92(%%edx),%%mm6\n\t" - "pfmul %%mm6,%%mm5\n\t" - "movd %%mm5,48(%%ecx)\n\t" - "psrlq $32,%%mm5\n\t" - "movd %%mm5,20(%%ecx)\n\t" - "movq %%mm4,%%mm6\n\t" - "punpckldq %%mm6,%%mm5\n\t" - "pfsub %%mm6,%%mm5\n\t" - "punpckhdq %%mm5,%%mm5\n\t" - "movd 20(%%edx),%%mm6\n\t" - "punpckldq 48(%%edx),%%mm6\n\t" - "pfmul %%mm6,%%mm5\n\t" - "movd 20(%%esi),%%mm6\n\t" - "punpckldq 48(%%esi),%%mm6\n\t" - "pfadd %%mm6,%%mm5\n\t" - "movd %%mm5,640(%%ebx)\n\t" - "psrlq $32,%%mm5\n\t" - "movd %%mm5,1536(%%ebx)\n\t" - "movq %%mm3,%%mm4\n\t" - "pfsub %%mm2,%%mm4\n\t" - "movq %%mm7,%%mm5\n\t" - "punpckldq "MANGLE(tfcos36)"+20,%%mm5\n\t" - "pfmul %%mm5,%%mm4\n\t" - "movq %%mm4,%%mm5\n\t" - "pfacc %%mm5,%%mm5\n\t" - "movd 128(%%edx),%%mm6\n\t" - "punpckldq 84(%%edx),%%mm6\n\t" - "pfmul %%mm6,%%mm5\n\t" - "movd %%mm5,56(%%ecx)\n\t" - "psrlq $32,%%mm5\n\t" - "movd %%mm5,12(%%ecx)\n\t" - "movq %%mm4,%%mm6\n\t" - "punpckldq %%mm6,%%mm5\n\t" - "pfsub %%mm6,%%mm5\n\t" - "punpckhdq %%mm5,%%mm5\n\t" - "movd 12(%%edx),%%mm6\n\t" - "punpckldq 56(%%edx),%%mm6\n\t" - "pfmul %%mm6,%%mm5\n\t" - "movd 12(%%esi),%%mm6\n\t" - "punpckldq 56(%%esi),%%mm6\n\t" - "pfadd %%mm6,%%mm5\n\t" - "movd %%mm5,384(%%ebx)\n\t" - "psrlq $32,%%mm5\n\t" - "movd %%mm5,1792(%%ebx)\n\t" + "movq 8(%%eax),%%mm2\n\t" + "movd "MANGLE(COS9)"+20,%%mm3\n\t" + "punpckldq %%mm3,%%mm3\n\t" + "pfmul %%mm3,%%mm2\n\t" + "pfsub %%mm0,%%mm2\n\t" + "movq 40(%%eax),%%mm3\n\t" + "movd "MANGLE(COS9)"+28,%%mm4\n\t" + "punpckldq %%mm4,%%mm4\n\t" + "pfmul %%mm4,%%mm3\n\t" + "pfsub %%mm3,%%mm2\n\t" + "movq 56(%%eax),%%mm3\n\t" + "movd "MANGLE(COS9)"+4,%%mm4\n\t" + "punpckldq %%mm4,%%mm4\n\t" + "pfmul %%mm4,%%mm3\n\t" + "pfadd %%mm3,%%mm2\n\t" + "movq (%%eax),%%mm3\n\t" + "movq 16(%%eax),%%mm4\n\t" + "movd "MANGLE(COS9)"+32,%%mm5\n\t" + "punpckldq %%mm5,%%mm5\n\t" + "pfmul %%mm5,%%mm4\n\t" + "pfsub %%mm4,%%mm3\n\t" + "movq 32(%%eax),%%mm4\n\t" + "movd "MANGLE(COS9)"+8,%%mm5\n\t" + "punpckldq %%mm5,%%mm5\n\t" + "pfmul %%mm5,%%mm4\n\t" + "pfsub %%mm4,%%mm3\n\t" + "pfadd %%mm1,%%mm3\n\t" + "movq 64(%%eax),%%mm4\n\t" + "movd "MANGLE(COS9)"+16,%%mm5\n\t" + "punpckldq %%mm5,%%mm5\n\t" + "pfmul %%mm5,%%mm4\n\t" + "pfadd %%mm4,%%mm3\n\t" + "movq %%mm2,%%mm4\n\t" + "pfadd %%mm3,%%mm4\n\t" + "movq %%mm7,%%mm5\n\t" + "punpckldq "MANGLE(tfcos36)"+8,%%mm5\n\t" + "pfmul %%mm5,%%mm4\n\t" + "movq %%mm4,%%mm5\n\t" + "pfacc %%mm5,%%mm5\n\t" + "movd 116(%%edx),%%mm6\n\t" + "punpckldq 96(%%edx),%%mm6\n\t" + "pfmul %%mm6,%%mm5\n\t" + "movd %%mm5,44(%%ecx)\n\t" + "psrlq $32,%%mm5\n\t" + "movd %%mm5,24(%%ecx)\n\t" + "movq %%mm4,%%mm6\n\t" + "punpckldq %%mm6,%%mm5\n\t" + "pfsub %%mm6,%%mm5\n\t" + "punpckhdq %%mm5,%%mm5\n\t" + "movd 24(%%edx),%%mm6\n\t" + "punpckldq 44(%%edx),%%mm6\n\t" + "pfmul %%mm6,%%mm5\n\t" + "movd 24(%%esi),%%mm6\n\t" + "punpckldq 44(%%esi),%%mm6\n\t" + "pfadd %%mm6,%%mm5\n\t" + "movd %%mm5,768(%%ebx)\n\t" + "psrlq $32,%%mm5\n\t" + "movd %%mm5,1408(%%ebx)\n\t" + "movq %%mm3,%%mm4\n\t" + "pfsub %%mm2,%%mm4\n\t" + "movq %%mm7,%%mm5\n\t" + "punpckldq "MANGLE(tfcos36)"+24,%%mm5\n\t" + "pfmul %%mm5,%%mm4\n\t" + "movq %%mm4,%%mm5\n\t" + "pfacc %%mm5,%%mm5\n\t" + "movd 132(%%edx),%%mm6\n\t" + "punpckldq 80(%%edx),%%mm6\n\t" + "pfmul %%mm6,%%mm5\n\t" + "movd %%mm5,60(%%ecx)\n\t" + "psrlq $32,%%mm5\n\t" + "movd %%mm5,8(%%ecx)\n\t" + "movq %%mm4,%%mm6\n\t" + "punpckldq %%mm6,%%mm5\n\t" + "pfsub %%mm6,%%mm5\n\t" + "punpckhdq %%mm5,%%mm5\n\t" + "movd 8(%%edx),%%mm6\n\t" + "punpckldq 60(%%edx),%%mm6\n\t" + "pfmul %%mm6,%%mm5\n\t" + "movd 8(%%esi),%%mm6\n\t" + "punpckldq 60(%%esi),%%mm6\n\t" + "pfadd %%mm6,%%mm5\n\t" + "movd %%mm5,256(%%ebx)\n\t" + "psrlq $32,%%mm5\n\t" + "movd %%mm5,1920(%%ebx)\n\t" + "movq 8(%%eax),%%mm2\n\t" + "movd "MANGLE(COS9)"+28,%%mm3\n\t" + "punpckldq %%mm3,%%mm3\n\t" + "pfmul %%mm3,%%mm2\n\t" + "pfsub %%mm0,%%mm2\n\t" + "movq 40(%%eax),%%mm3\n\t" + "movd "MANGLE(COS9)"+4,%%mm4\n\t" + "punpckldq %%mm4,%%mm4\n\t" + "pfmul %%mm4,%%mm3\n\t" + "pfadd %%mm3,%%mm2\n\t" + "movq 56(%%eax),%%mm3\n\t" + "movd "MANGLE(COS9)"+20,%%mm4\n\t" + "punpckldq %%mm4,%%mm4\n\t" + "pfmul %%mm4,%%mm3\n\t" + "pfsub %%mm3,%%mm2\n\t" + "movq (%%eax),%%mm3\n\t" + "movq 16(%%eax),%%mm4\n\t" + "movd "MANGLE(COS9)"+16,%%mm5\n\t" + "punpckldq %%mm5,%%mm5\n\t" + "pfmul %%mm5,%%mm4\n\t" + "pfsub %%mm4,%%mm3\n\t" + "movq 32(%%eax),%%mm4\n\t" + "movd "MANGLE(COS9)"+32,%%mm5\n\t" + "punpckldq %%mm5,%%mm5\n\t" + "pfmul %%mm5,%%mm4\n\t" + "pfadd %%mm4,%%mm3\n\t" + "pfadd %%mm1,%%mm3\n\t" + "movq 64(%%eax),%%mm4\n\t" + "movd "MANGLE(COS9)"+8,%%mm5\n\t" + "punpckldq %%mm5,%%mm5\n\t" + "pfmul %%mm5,%%mm4\n\t" + "pfsub %%mm4,%%mm3\n\t" + "movq %%mm2,%%mm4\n\t" + "pfadd %%mm3,%%mm4\n\t" + "movq %%mm7,%%mm5\n\t" + "punpckldq "MANGLE(tfcos36)"+12,%%mm5\n\t" + "pfmul %%mm5,%%mm4\n\t" + "movq %%mm4,%%mm5\n\t" + "pfacc %%mm5,%%mm5\n\t" + "movd 120(%%edx),%%mm6\n\t" + "punpckldq 92(%%edx),%%mm6\n\t" + "pfmul %%mm6,%%mm5\n\t" + "movd %%mm5,48(%%ecx)\n\t" + "psrlq $32,%%mm5\n\t" + "movd %%mm5,20(%%ecx)\n\t" + "movq %%mm4,%%mm6\n\t" + "punpckldq %%mm6,%%mm5\n\t" + "pfsub %%mm6,%%mm5\n\t" + "punpckhdq %%mm5,%%mm5\n\t" + "movd 20(%%edx),%%mm6\n\t" + "punpckldq 48(%%edx),%%mm6\n\t" + "pfmul %%mm6,%%mm5\n\t" + "movd 20(%%esi),%%mm6\n\t" + "punpckldq 48(%%esi),%%mm6\n\t" + "pfadd %%mm6,%%mm5\n\t" + "movd %%mm5,640(%%ebx)\n\t" + "psrlq $32,%%mm5\n\t" + "movd %%mm5,1536(%%ebx)\n\t" + "movq %%mm3,%%mm4\n\t" + "pfsub %%mm2,%%mm4\n\t" + "movq %%mm7,%%mm5\n\t" + "punpckldq "MANGLE(tfcos36)"+20,%%mm5\n\t" + "pfmul %%mm5,%%mm4\n\t" + "movq %%mm4,%%mm5\n\t" + "pfacc %%mm5,%%mm5\n\t" + "movd 128(%%edx),%%mm6\n\t" + "punpckldq 84(%%edx),%%mm6\n\t" + "pfmul %%mm6,%%mm5\n\t" + "movd %%mm5,56(%%ecx)\n\t" + "psrlq $32,%%mm5\n\t" + "movd %%mm5,12(%%ecx)\n\t" + "movq %%mm4,%%mm6\n\t" + "punpckldq %%mm6,%%mm5\n\t" + "pfsub %%mm6,%%mm5\n\t" + "punpckhdq %%mm5,%%mm5\n\t" + "movd 12(%%edx),%%mm6\n\t" + "punpckldq 56(%%edx),%%mm6\n\t" + "pfmul %%mm6,%%mm5\n\t" + "movd 12(%%esi),%%mm6\n\t" + "punpckldq 56(%%esi),%%mm6\n\t" + "pfadd %%mm6,%%mm5\n\t" + "movd %%mm5,384(%%ebx)\n\t" + "psrlq $32,%%mm5\n\t" + "movd %%mm5,1792(%%ebx)\n\t" - "movq (%%eax),%%mm4\n\t" - "movq 16(%%eax),%%mm3\n\t" - "pfsub %%mm3,%%mm4\n\t" - "movq 32(%%eax),%%mm3\n\t" - "pfadd %%mm3,%%mm4\n\t" - "movq 48(%%eax),%%mm3\n\t" - "pfsub %%mm3,%%mm4\n\t" - "movq 64(%%eax),%%mm3\n\t" - "pfadd %%mm3,%%mm4\n\t" - "movq %%mm7,%%mm5\n\t" - "punpckldq "MANGLE(tfcos36)"+16,%%mm5\n\t" - "pfmul %%mm5,%%mm4\n\t" - "movq %%mm4,%%mm5\n\t" - "pfacc %%mm5,%%mm5\n\t" - "movd 124(%%edx),%%mm6\n\t" - "punpckldq 88(%%edx),%%mm6\n\t" - "pfmul %%mm6,%%mm5\n\t" - "movd %%mm5,52(%%ecx)\n\t" - "psrlq $32,%%mm5\n\t" - "movd %%mm5,16(%%ecx)\n\t" - "movq %%mm4,%%mm6\n\t" - "punpckldq %%mm6,%%mm5\n\t" - "pfsub %%mm6,%%mm5\n\t" - "punpckhdq %%mm5,%%mm5\n\t" - "movd 16(%%edx),%%mm6\n\t" - "punpckldq 52(%%edx),%%mm6\n\t" - "pfmul %%mm6,%%mm5\n\t" - "movd 16(%%esi),%%mm6\n\t" - "punpckldq 52(%%esi),%%mm6\n\t" - "pfadd %%mm6,%%mm5\n\t" - "movd %%mm5,512(%%ebx)\n\t" - "psrlq $32,%%mm5\n\t" - "movd %%mm5,1664(%%ebx)\n\t" + "movq (%%eax),%%mm4\n\t" + "movq 16(%%eax),%%mm3\n\t" + "pfsub %%mm3,%%mm4\n\t" + "movq 32(%%eax),%%mm3\n\t" + "pfadd %%mm3,%%mm4\n\t" + "movq 48(%%eax),%%mm3\n\t" + "pfsub %%mm3,%%mm4\n\t" + "movq 64(%%eax),%%mm3\n\t" + "pfadd %%mm3,%%mm4\n\t" + "movq %%mm7,%%mm5\n\t" + "punpckldq "MANGLE(tfcos36)"+16,%%mm5\n\t" + "pfmul %%mm5,%%mm4\n\t" + "movq %%mm4,%%mm5\n\t" + "pfacc %%mm5,%%mm5\n\t" + "movd 124(%%edx),%%mm6\n\t" + "punpckldq 88(%%edx),%%mm6\n\t" + "pfmul %%mm6,%%mm5\n\t" + "movd %%mm5,52(%%ecx)\n\t" + "psrlq $32,%%mm5\n\t" + "movd %%mm5,16(%%ecx)\n\t" + "movq %%mm4,%%mm6\n\t" + "punpckldq %%mm6,%%mm5\n\t" + "pfsub %%mm6,%%mm5\n\t" + "punpckhdq %%mm5,%%mm5\n\t" + "movd 16(%%edx),%%mm6\n\t" + "punpckldq 52(%%edx),%%mm6\n\t" + "pfmul %%mm6,%%mm5\n\t" + "movd 16(%%esi),%%mm6\n\t" + "punpckldq 52(%%esi),%%mm6\n\t" + "pfadd %%mm6,%%mm5\n\t" + "movd %%mm5,512(%%ebx)\n\t" + "psrlq $32,%%mm5\n\t" + "movd %%mm5,1664(%%ebx)\n\t" - "femms\n\t" - : - : "a" (inbuf), "S" (o1), "c" (o2), "d" (wintab), "b" (tsbuf) - : "memory"); + "femms\n\t" + : + : "a" (inbuf), "S" (o1), "c" (o2), "d" (wintab), "b" (tsbuf) + : "memory"); } diff --git a/mp3lib/dct64_3dnow.c b/mp3lib/dct64_3dnow.c index 55340b0512..d528654207 100644 --- a/mp3lib/dct64_3dnow.c +++ b/mp3lib/dct64_3dnow.c @@ -19,910 +19,910 @@ void dct64_MMX_3dnow(short *a,short *b,real *c) { char tmp[256]; __asm__ volatile( -" movl %2,%%eax\n\t" +" movl %2,%%eax\n\t" -" leal 128+%3,%%edx\n\t" -" movl %0,%%esi\n\t" -" movl %1,%%edi\n\t" -" movl $"MANGLE(costab_mmx)",%%ebx\n\t" -" leal %3,%%ecx\n\t" +" leal 128+%3,%%edx\n\t" +" movl %0,%%esi\n\t" +" movl %1,%%edi\n\t" +" movl $"MANGLE(costab_mmx)",%%ebx\n\t" +" leal %3,%%ecx\n\t" /* Phase 1*/ -" movq (%%eax), %%mm0\n\t" -" movq 8(%%eax), %%mm4\n\t" -" movq %%mm0, %%mm3\n\t" -" movq %%mm4, %%mm7\n\t" -" movq 120(%%eax), %%mm1\n\t" -" movq 112(%%eax), %%mm5\n\t" +" movq (%%eax), %%mm0\n\t" +" movq 8(%%eax), %%mm4\n\t" +" movq %%mm0, %%mm3\n\t" +" movq %%mm4, %%mm7\n\t" +" movq 120(%%eax), %%mm1\n\t" +" movq 112(%%eax), %%mm5\n\t" /* n.b.: pswapd*/ -" movq %%mm1, %%mm2\n\t" -" movq %%mm5, %%mm6\n\t" -" psrlq $32, %%mm1\n\t" -" psrlq $32, %%mm5\n\t" -" punpckldq %%mm2, %%mm1\n\t" -" punpckldq %%mm6, %%mm5\n\t" +" movq %%mm1, %%mm2\n\t" +" movq %%mm5, %%mm6\n\t" +" psrlq $32, %%mm1\n\t" +" psrlq $32, %%mm5\n\t" +" punpckldq %%mm2, %%mm1\n\t" +" punpckldq %%mm6, %%mm5\n\t" /**/ -" pfadd %%mm1, %%mm0\n\t" -" pfadd %%mm5, %%mm4\n\t" -" movq %%mm0, (%%edx)\n\t" -" movq %%mm4, 8(%%edx)\n\t" -" pfsub %%mm1, %%mm3\n\t" -" pfsub %%mm5, %%mm7\n\t" -" pfmul (%%ebx), %%mm3\n\t" -" pfmul 8(%%ebx), %%mm7\n\t" -" movd %%mm3, 124(%%edx)\n\t" -" movd %%mm7, 116(%%edx)\n\t" -" psrlq $32, %%mm3\n\t" -" psrlq $32, %%mm7\n\t" -" movd %%mm3, 120(%%edx)\n\t" -" movd %%mm7, 112(%%edx)\n\t" - -" movq 16(%%eax), %%mm0\n\t" -" movq 24(%%eax), %%mm4\n\t" -" movq %%mm0, %%mm3\n\t" -" movq %%mm4, %%mm7\n\t" -" movq 104(%%eax), %%mm1\n\t" -" movq 96(%%eax), %%mm5\n\t" +" pfadd %%mm1, %%mm0\n\t" +" pfadd %%mm5, %%mm4\n\t" +" movq %%mm0, (%%edx)\n\t" +" movq %%mm4, 8(%%edx)\n\t" +" pfsub %%mm1, %%mm3\n\t" +" pfsub %%mm5, %%mm7\n\t" +" pfmul (%%ebx), %%mm3\n\t" +" pfmul 8(%%ebx), %%mm7\n\t" +" movd %%mm3, 124(%%edx)\n\t" +" movd %%mm7, 116(%%edx)\n\t" +" psrlq $32, %%mm3\n\t" +" psrlq $32, %%mm7\n\t" +" movd %%mm3, 120(%%edx)\n\t" +" movd %%mm7, 112(%%edx)\n\t" + +" movq 16(%%eax), %%mm0\n\t" +" movq 24(%%eax), %%mm4\n\t" +" movq %%mm0, %%mm3\n\t" +" movq %%mm4, %%mm7\n\t" +" movq 104(%%eax), %%mm1\n\t" +" movq 96(%%eax), %%mm5\n\t" /* n.b.: pswapd*/ -" movq %%mm1, %%mm2\n\t" -" movq %%mm5, %%mm6\n\t" -" psrlq $32, %%mm1\n\t" -" psrlq $32, %%mm5\n\t" -" punpckldq %%mm2, %%mm1\n\t" -" punpckldq %%mm6, %%mm5\n\t" +" movq %%mm1, %%mm2\n\t" +" movq %%mm5, %%mm6\n\t" +" psrlq $32, %%mm1\n\t" +" psrlq $32, %%mm5\n\t" +" punpckldq %%mm2, %%mm1\n\t" +" punpckldq %%mm6, %%mm5\n\t" /**/ -" pfadd %%mm1, %%mm0\n\t" -" pfadd %%mm5, %%mm4\n\t" -" movq %%mm0, 16(%%edx)\n\t" -" movq %%mm4, 24(%%edx)\n\t" -" pfsub %%mm1, %%mm3\n\t" -" pfsub %%mm5, %%mm7\n\t" -" pfmul 16(%%ebx), %%mm3\n\t" -" pfmul 24(%%ebx), %%mm7\n\t" -" movd %%mm3, 108(%%edx)\n\t" -" movd %%mm7, 100(%%edx)\n\t" -" psrlq $32, %%mm3\n\t" -" psrlq $32, %%mm7\n\t" -" movd %%mm3, 104(%%edx)\n\t" -" movd %%mm7, 96(%%edx)\n\t" - -" movq 32(%%eax), %%mm0\n\t" -" movq 40(%%eax), %%mm4\n\t" -" movq %%mm0, %%mm3\n\t" -" movq %%mm4, %%mm7\n\t" -" movq 88(%%eax), %%mm1\n\t" -" movq 80(%%eax), %%mm5\n\t" +" pfadd %%mm1, %%mm0\n\t" +" pfadd %%mm5, %%mm4\n\t" +" movq %%mm0, 16(%%edx)\n\t" +" movq %%mm4, 24(%%edx)\n\t" +" pfsub %%mm1, %%mm3\n\t" +" pfsub %%mm5, %%mm7\n\t" +" pfmul 16(%%ebx), %%mm3\n\t" +" pfmul 24(%%ebx), %%mm7\n\t" +" movd %%mm3, 108(%%edx)\n\t" +" movd %%mm7, 100(%%edx)\n\t" +" psrlq $32, %%mm3\n\t" +" psrlq $32, %%mm7\n\t" +" movd %%mm3, 104(%%edx)\n\t" +" movd %%mm7, 96(%%edx)\n\t" + +" movq 32(%%eax), %%mm0\n\t" +" movq 40(%%eax), %%mm4\n\t" +" movq %%mm0, %%mm3\n\t" +" movq %%mm4, %%mm7\n\t" +" movq 88(%%eax), %%mm1\n\t" +" movq 80(%%eax), %%mm5\n\t" /* n.b.: pswapd*/ -" movq %%mm1, %%mm2\n\t" -" movq %%mm5, %%mm6\n\t" -" psrlq $32, %%mm1\n\t" -" psrlq $32, %%mm5\n\t" -" punpckldq %%mm2, %%mm1\n\t" -" punpckldq %%mm6, %%mm5\n\t" +" movq %%mm1, %%mm2\n\t" +" movq %%mm5, %%mm6\n\t" +" psrlq $32, %%mm1\n\t" +" psrlq $32, %%mm5\n\t" +" punpckldq %%mm2, %%mm1\n\t" +" punpckldq %%mm6, %%mm5\n\t" /**/ -" pfadd %%mm1, %%mm0\n\t" -" pfadd %%mm5, %%mm4\n\t" -" movq %%mm0, 32(%%edx)\n\t" -" movq %%mm4, 40(%%edx)\n\t" -" pfsub %%mm1, %%mm3\n\t" -" pfsub %%mm5, %%mm7\n\t" -" pfmul 32(%%ebx), %%mm3\n\t" -" pfmul 40(%%ebx), %%mm7\n\t" -" movd %%mm3, 92(%%edx)\n\t" -" movd %%mm7, 84(%%edx)\n\t" -" psrlq $32, %%mm3\n\t" -" psrlq $32, %%mm7\n\t" -" movd %%mm3, 88(%%edx)\n\t" -" movd %%mm7, 80(%%edx)\n\t" - -" movq 48(%%eax), %%mm0\n\t" -" movq 56(%%eax), %%mm4\n\t" -" movq %%mm0, %%mm3\n\t" -" movq %%mm4, %%mm7\n\t" -" movq 72(%%eax), %%mm1\n\t" -" movq 64(%%eax), %%mm5\n\t" +" pfadd %%mm1, %%mm0\n\t" +" pfadd %%mm5, %%mm4\n\t" +" movq %%mm0, 32(%%edx)\n\t" +" movq %%mm4, 40(%%edx)\n\t" +" pfsub %%mm1, %%mm3\n\t" +" pfsub %%mm5, %%mm7\n\t" +" pfmul 32(%%ebx), %%mm3\n\t" +" pfmul 40(%%ebx), %%mm7\n\t" +" movd %%mm3, 92(%%edx)\n\t" +" movd %%mm7, 84(%%edx)\n\t" +" psrlq $32, %%mm3\n\t" +" psrlq $32, %%mm7\n\t" +" movd %%mm3, 88(%%edx)\n\t" +" movd %%mm7, 80(%%edx)\n\t" + +" movq 48(%%eax), %%mm0\n\t" +" movq 56(%%eax), %%mm4\n\t" +" movq %%mm0, %%mm3\n\t" +" movq %%mm4, %%mm7\n\t" +" movq 72(%%eax), %%mm1\n\t" +" movq 64(%%eax), %%mm5\n\t" /* n.b.: pswapd*/ -" movq %%mm1, %%mm2\n\t" -" movq %%mm5, %%mm6\n\t" -" psrlq $32, %%mm1\n\t" -" psrlq $32, %%mm5\n\t" -" punpckldq %%mm2, %%mm1\n\t" -" punpckldq %%mm6, %%mm5\n\t" +" movq %%mm1, %%mm2\n\t" +" movq %%mm5, %%mm6\n\t" +" psrlq $32, %%mm1\n\t" +" psrlq $32, %%mm5\n\t" +" punpckldq %%mm2, %%mm1\n\t" +" punpckldq %%mm6, %%mm5\n\t" /**/ -" pfadd %%mm1, %%mm0\n\t" -" pfadd %%mm5, %%mm4\n\t" -" movq %%mm0, 48(%%edx)\n\t" -" movq %%mm4, 56(%%edx)\n\t" -" pfsub %%mm1, %%mm3\n\t" -" pfsub %%mm5, %%mm7\n\t" -" pfmul 48(%%ebx), %%mm3\n\t" -" pfmul 56(%%ebx), %%mm7\n\t" -" movd %%mm3, 76(%%edx)\n\t" -" movd %%mm7, 68(%%edx)\n\t" -" psrlq $32, %%mm3\n\t" -" psrlq $32, %%mm7\n\t" -" movd %%mm3, 72(%%edx)\n\t" -" movd %%mm7, 64(%%edx)\n\t" +" pfadd %%mm1, %%mm0\n\t" +" pfadd %%mm5, %%mm4\n\t" +" movq %%mm0, 48(%%edx)\n\t" +" movq %%mm4, 56(%%edx)\n\t" +" pfsub %%mm1, %%mm3\n\t" +" pfsub %%mm5, %%mm7\n\t" +" pfmul 48(%%ebx), %%mm3\n\t" +" pfmul 56(%%ebx), %%mm7\n\t" +" movd %%mm3, 76(%%edx)\n\t" +" movd %%mm7, 68(%%edx)\n\t" +" psrlq $32, %%mm3\n\t" +" psrlq $32, %%mm7\n\t" +" movd %%mm3, 72(%%edx)\n\t" +" movd %%mm7, 64(%%edx)\n\t" /* Phase 2*/ -" movq (%%edx), %%mm0\n\t" -" movq 8(%%edx), %%mm4\n\t" -" movq %%mm0, %%mm3\n\t" -" movq %%mm4, %%mm7\n\t" -" movq 56(%%edx), %%mm1\n\t" -" movq 48(%%edx), %%mm5\n\t" +" movq (%%edx), %%mm0\n\t" +" movq 8(%%edx), %%mm4\n\t" +" movq %%mm0, %%mm3\n\t" +" movq %%mm4, %%mm7\n\t" +" movq 56(%%edx), %%mm1\n\t" +" movq 48(%%edx), %%mm5\n\t" /* n.b.: pswapd*/ -" movq %%mm1, %%mm2\n\t" -" movq %%mm5, %%mm6\n\t" -" psrlq $32, %%mm1\n\t" -" psrlq $32, %%mm5\n\t" -" punpckldq %%mm2, %%mm1\n\t" -" punpckldq %%mm6, %%mm5\n\t" +" movq %%mm1, %%mm2\n\t" +" movq %%mm5, %%mm6\n\t" +" psrlq $32, %%mm1\n\t" +" psrlq $32, %%mm5\n\t" +" punpckldq %%mm2, %%mm1\n\t" +" punpckldq %%mm6, %%mm5\n\t" /**/ -" pfadd %%mm1, %%mm0\n\t" -" pfadd %%mm5, %%mm4\n\t" -" movq %%mm0, (%%ecx)\n\t" -" movq %%mm4, 8(%%ecx)\n\t" -" pfsub %%mm1, %%mm3\n\t" -" pfsub %%mm5, %%mm7\n\t" -" pfmul 64(%%ebx), %%mm3\n\t" -" pfmul 72(%%ebx), %%mm7\n\t" -" movd %%mm3, 60(%%ecx)\n\t" -" movd %%mm7, 52(%%ecx)\n\t" -" psrlq $32, %%mm3\n\t" -" psrlq $32, %%mm7\n\t" -" movd %%mm3, 56(%%ecx)\n\t" -" movd %%mm7, 48(%%ecx)\n\t" - -" movq 16(%%edx), %%mm0\n\t" -" movq 24(%%edx), %%mm4\n\t" -" movq %%mm0, %%mm3\n\t" -" movq %%mm4, %%mm7\n\t" -" movq 40(%%edx), %%mm1\n\t" -" movq 32(%%edx), %%mm5\n\t" +" pfadd %%mm1, %%mm0\n\t" +" pfadd %%mm5, %%mm4\n\t" +" movq %%mm0, (%%ecx)\n\t" +" movq %%mm4, 8(%%ecx)\n\t" +" pfsub %%mm1, %%mm3\n\t" +" pfsub %%mm5, %%mm7\n\t" +" pfmul 64(%%ebx), %%mm3\n\t" +" pfmul 72(%%ebx), %%mm7\n\t" +" movd %%mm3, 60(%%ecx)\n\t" +" movd %%mm7, 52(%%ecx)\n\t" +" psrlq $32, %%mm3\n\t" +" psrlq $32, %%mm7\n\t" +" movd %%mm3, 56(%%ecx)\n\t" +" movd %%mm7, 48(%%ecx)\n\t" + +" movq 16(%%edx), %%mm0\n\t" +" movq 24(%%edx), %%mm4\n\t" +" movq %%mm0, %%mm3\n\t" +" movq %%mm4, %%mm7\n\t" +" movq 40(%%edx), %%mm1\n\t" +" movq 32(%%edx), %%mm5\n\t" /* n.b.: pswapd*/ -" movq %%mm1, %%mm2\n\t" -" movq %%mm5, %%mm6\n\t" -" psrlq $32, %%mm1\n\t" -" psrlq $32, %%mm5\n\t" -" punpckldq %%mm2, %%mm1\n\t" -" punpckldq %%mm6, %%mm5\n\t" +" movq %%mm1, %%mm2\n\t" +" movq %%mm5, %%mm6\n\t" +" psrlq $32, %%mm1\n\t" +" psrlq $32, %%mm5\n\t" +" punpckldq %%mm2, %%mm1\n\t" +" punpckldq %%mm6, %%mm5\n\t" /**/ -" pfadd %%mm1, %%mm0\n\t" -" pfadd %%mm5, %%mm4\n\t" -" movq %%mm0, 16(%%ecx)\n\t" -" movq %%mm4, 24(%%ecx)\n\t" -" pfsub %%mm1, %%mm3\n\t" -" pfsub %%mm5, %%mm7\n\t" -" pfmul 80(%%ebx), %%mm3\n\t" -" pfmul 88(%%ebx), %%mm7\n\t" -" movd %%mm3, 44(%%ecx)\n\t" -" movd %%mm7, 36(%%ecx)\n\t" -" psrlq $32, %%mm3\n\t" -" psrlq $32, %%mm7\n\t" -" movd %%mm3, 40(%%ecx)\n\t" -" movd %%mm7, 32(%%ecx)\n\t" +" pfadd %%mm1, %%mm0\n\t" +" pfadd %%mm5, %%mm4\n\t" +" movq %%mm0, 16(%%ecx)\n\t" +" movq %%mm4, 24(%%ecx)\n\t" +" pfsub %%mm1, %%mm3\n\t" +" pfsub %%mm5, %%mm7\n\t" +" pfmul 80(%%ebx), %%mm3\n\t" +" pfmul 88(%%ebx), %%mm7\n\t" +" movd %%mm3, 44(%%ecx)\n\t" +" movd %%mm7, 36(%%ecx)\n\t" +" psrlq $32, %%mm3\n\t" +" psrlq $32, %%mm7\n\t" +" movd %%mm3, 40(%%ecx)\n\t" +" movd %%mm7, 32(%%ecx)\n\t" /* Phase 3*/ -" movq 64(%%edx), %%mm0\n\t" -" movq 72(%%edx), %%mm4\n\t" -" movq %%mm0, %%mm3\n\t" -" movq %%mm4, %%mm7\n\t" -" movq 120(%%edx), %%mm1\n\t" -" movq 112(%%edx), %%mm5\n\t" +" movq 64(%%edx), %%mm0\n\t" +" movq 72(%%edx), %%mm4\n\t" +" movq %%mm0, %%mm3\n\t" +" movq %%mm4, %%mm7\n\t" +" movq 120(%%edx), %%mm1\n\t" +" movq 112(%%edx), %%mm5\n\t" /* n.b.: pswapd*/ -" movq %%mm1, %%mm2\n\t" -" movq %%mm5, %%mm6\n\t" -" psrlq $32, %%mm1\n\t" -" psrlq $32, %%mm5\n\t" -" punpckldq %%mm2, %%mm1\n\t" -" punpckldq %%mm6, %%mm5\n\t" +" movq %%mm1, %%mm2\n\t" +" movq %%mm5, %%mm6\n\t" +" psrlq $32, %%mm1\n\t" +" psrlq $32, %%mm5\n\t" +" punpckldq %%mm2, %%mm1\n\t" +" punpckldq %%mm6, %%mm5\n\t" /**/ -" pfadd %%mm1, %%mm0\n\t" -" pfadd %%mm5, %%mm4\n\t" -" movq %%mm0, 64(%%ecx)\n\t" -" movq %%mm4, 72(%%ecx)\n\t" -" pfsubr %%mm1, %%mm3\n\t" -" pfsubr %%mm5, %%mm7\n\t" -" pfmul 64(%%ebx), %%mm3\n\t" -" pfmul 72(%%ebx), %%mm7\n\t" -" movd %%mm3, 124(%%ecx)\n\t" -" movd %%mm7, 116(%%ecx)\n\t" -" psrlq $32, %%mm3\n\t" -" psrlq $32, %%mm7\n\t" -" movd %%mm3, 120(%%ecx)\n\t" -" movd %%mm7, 112(%%ecx)\n\t" - -" movq 80(%%edx), %%mm0\n\t" -" movq 88(%%edx), %%mm4\n\t" -" movq %%mm0, %%mm3\n\t" -" movq %%mm4, %%mm7\n\t" -" movq 104(%%edx), %%mm1\n\t" -" movq 96(%%edx), %%mm5\n\t" +" pfadd %%mm1, %%mm0\n\t" +" pfadd %%mm5, %%mm4\n\t" +" movq %%mm0, 64(%%ecx)\n\t" +" movq %%mm4, 72(%%ecx)\n\t" +" pfsubr %%mm1, %%mm3\n\t" +" pfsubr %%mm5, %%mm7\n\t" +" pfmul 64(%%ebx), %%mm3\n\t" +" pfmul 72(%%ebx), %%mm7\n\t" +" movd %%mm3, 124(%%ecx)\n\t" +" movd %%mm7, 116(%%ecx)\n\t" +" psrlq $32, %%mm3\n\t" +" psrlq $32, %%mm7\n\t" +" movd %%mm3, 120(%%ecx)\n\t" +" movd %%mm7, 112(%%ecx)\n\t" + +" movq 80(%%edx), %%mm0\n\t" +" movq 88(%%edx), %%mm4\n\t" +" movq %%mm0, %%mm3\n\t" +" movq %%mm4, %%mm7\n\t" +" movq 104(%%edx), %%mm1\n\t" +" movq 96(%%edx), %%mm5\n\t" /* n.b.: pswapd*/ -" movq %%mm1, %%mm2\n\t" -" movq %%mm5, %%mm6\n\t" -" psrlq $32, %%mm1\n\t" -" psrlq $32, %%mm5\n\t" -" punpckldq %%mm2, %%mm1\n\t" -" punpckldq %%mm6, %%mm5\n\t" +" movq %%mm1, %%mm2\n\t" +" movq %%mm5, %%mm6\n\t" +" psrlq $32, %%mm1\n\t" +" psrlq $32, %%mm5\n\t" +" punpckldq %%mm2, %%mm1\n\t" +" punpckldq %%mm6, %%mm5\n\t" /**/ -" pfadd %%mm1, %%mm0\n\t" -" pfadd %%mm5, %%mm4\n\t" -" movq %%mm0, 80(%%ecx)\n\t" -" movq %%mm4, 88(%%ecx)\n\t" -" pfsubr %%mm1, %%mm3\n\t" -" pfsubr %%mm5, %%mm7\n\t" -" pfmul 80(%%ebx), %%mm3\n\t" -" pfmul 88(%%ebx), %%mm7\n\t" -" movd %%mm3, 108(%%ecx)\n\t" -" movd %%mm7, 100(%%ecx)\n\t" -" psrlq $32, %%mm3\n\t" -" psrlq $32, %%mm7\n\t" -" movd %%mm3, 104(%%ecx)\n\t" -" movd %%mm7, 96(%%ecx)\n\t" +" pfadd %%mm1, %%mm0\n\t" +" pfadd %%mm5, %%mm4\n\t" +" movq %%mm0, 80(%%ecx)\n\t" +" movq %%mm4, 88(%%ecx)\n\t" +" pfsubr %%mm1, %%mm3\n\t" +" pfsubr %%mm5, %%mm7\n\t" +" pfmul 80(%%ebx), %%mm3\n\t" +" pfmul 88(%%ebx), %%mm7\n\t" +" movd %%mm3, 108(%%ecx)\n\t" +" movd %%mm7, 100(%%ecx)\n\t" +" psrlq $32, %%mm3\n\t" +" psrlq $32, %%mm7\n\t" +" movd %%mm3, 104(%%ecx)\n\t" +" movd %%mm7, 96(%%ecx)\n\t" /* Phase 4*/ -" movq (%%ecx), %%mm0\n\t" -" movq 8(%%ecx), %%mm4\n\t" -" movq %%mm0, %%mm3\n\t" -" movq %%mm4, %%mm7\n\t" -" movq 24(%%ecx), %%mm1\n\t" -" movq 16(%%ecx), %%mm5\n\t" +" movq (%%ecx), %%mm0\n\t" +" movq 8(%%ecx), %%mm4\n\t" +" movq %%mm0, %%mm3\n\t" +" movq %%mm4, %%mm7\n\t" +" movq 24(%%ecx), %%mm1\n\t" +" movq 16(%%ecx), %%mm5\n\t" /* n.b.: pswapd*/ -" movq %%mm1, %%mm2\n\t" -" movq %%mm5, %%mm6\n\t" -" psrlq $32, %%mm1\n\t" -" psrlq $32, %%mm5\n\t" -" punpckldq %%mm2, %%mm1\n\t" -" punpckldq %%mm6, %%mm5\n\t" +" movq %%mm1, %%mm2\n\t" +" movq %%mm5, %%mm6\n\t" +" psrlq $32, %%mm1\n\t" +" psrlq $32, %%mm5\n\t" +" punpckldq %%mm2, %%mm1\n\t" +" punpckldq %%mm6, %%mm5\n\t" /**/ -" pfadd %%mm1, %%mm0\n\t" -" pfadd %%mm5, %%mm4\n\t" -" movq %%mm0, (%%edx)\n\t" -" movq %%mm4, 8(%%edx)\n\t" -" pfsub %%mm1, %%mm3\n\t" -" pfsub %%mm5, %%mm7\n\t" -" pfmul 96(%%ebx), %%mm3\n\t" -" pfmul 104(%%ebx), %%mm7\n\t" -" movd %%mm3, 28(%%edx)\n\t" -" movd %%mm7, 20(%%edx)\n\t" -" psrlq $32, %%mm3\n\t" -" psrlq $32, %%mm7\n\t" -" movd %%mm3, 24(%%edx)\n\t" -" movd %%mm7, 16(%%edx)\n\t" - -" movq 32(%%ecx), %%mm0\n\t" -" movq 40(%%ecx), %%mm4\n\t" -" movq %%mm0, %%mm3\n\t" -" movq %%mm4, %%mm7\n\t" -" movq 56(%%ecx), %%mm1\n\t" -" movq 48(%%ecx), %%mm5\n\t" +" pfadd %%mm1, %%mm0\n\t" +" pfadd %%mm5, %%mm4\n\t" +" movq %%mm0, (%%edx)\n\t" +" movq %%mm4, 8(%%edx)\n\t" +" pfsub %%mm1, %%mm3\n\t" +" pfsub %%mm5, %%mm7\n\t" +" pfmul 96(%%ebx), %%mm3\n\t" +" pfmul 104(%%ebx), %%mm7\n\t" +" movd %%mm3, 28(%%edx)\n\t" +" movd %%mm7, 20(%%edx)\n\t" +" psrlq $32, %%mm3\n\t" +" psrlq $32, %%mm7\n\t" +" movd %%mm3, 24(%%edx)\n\t" +" movd %%mm7, 16(%%edx)\n\t" + +" movq 32(%%ecx), %%mm0\n\t" +" movq 40(%%ecx), %%mm4\n\t" +" movq %%mm0, %%mm3\n\t" +" movq %%mm4, %%mm7\n\t" +" movq 56(%%ecx), %%mm1\n\t" +" movq 48(%%ecx), %%mm5\n\t" /* n.b.: pswapd*/ -" movq %%mm1, %%mm2\n\t" -" movq %%mm5, %%mm6\n\t" -" psrlq $32, %%mm1\n\t" -" psrlq $32, %%mm5\n\t" -" punpckldq %%mm2, %%mm1\n\t" -" punpckldq %%mm6, %%mm5\n\t" +" movq %%mm1, %%mm2\n\t" +" movq %%mm5, %%mm6\n\t" +" psrlq $32, %%mm1\n\t" +" psrlq $32, %%mm5\n\t" +" punpckldq %%mm2, %%mm1\n\t" +" punpckldq %%mm6, %%mm5\n\t" /**/ -" pfadd %%mm1, %%mm0\n\t" -" pfadd %%mm5, %%mm4\n\t" -" movq %%mm0, 32(%%edx)\n\t" -" movq %%mm4, 40(%%edx)\n\t" -" pfsubr %%mm1, %%mm3\n\t" -" pfsubr %%mm5, %%mm7\n\t" -" pfmul 96(%%ebx), %%mm3\n\t" -" pfmul 104(%%ebx), %%mm7\n\t" -" movd %%mm3, 60(%%edx)\n\t" -" movd %%mm7, 52(%%edx)\n\t" -" psrlq $32, %%mm3\n\t" -" psrlq $32, %%mm7\n\t" -" movd %%mm3, 56(%%edx)\n\t" -" movd %%mm7, 48(%%edx)\n\t" - -" movq 64(%%ecx), %%mm0\n\t" -" movq 72(%%ecx), %%mm4\n\t" -" movq %%mm0, %%mm3\n\t" -" movq %%mm4, %%mm7\n\t" -" movq 88(%%ecx), %%mm1\n\t" -" movq 80(%%ecx), %%mm5\n\t" +" pfadd %%mm1, %%mm0\n\t" +" pfadd %%mm5, %%mm4\n\t" +" movq %%mm0, 32(%%edx)\n\t" +" movq %%mm4, 40(%%edx)\n\t" +" pfsubr %%mm1, %%mm3\n\t" +" pfsubr %%mm5, %%mm7\n\t" +" pfmul 96(%%ebx), %%mm3\n\t" +" pfmul 104(%%ebx), %%mm7\n\t" +" movd %%mm3, 60(%%edx)\n\t" +" movd %%mm7, 52(%%edx)\n\t" +" psrlq $32, %%mm3\n\t" +" psrlq $32, %%mm7\n\t" +" movd %%mm3, 56(%%edx)\n\t" +" movd %%mm7, 48(%%edx)\n\t" + +" movq 64(%%ecx), %%mm0\n\t" +" movq 72(%%ecx), %%mm4\n\t" +" movq %%mm0, %%mm3\n\t" +" movq %%mm4, %%mm7\n\t" +" movq 88(%%ecx), %%mm1\n\t" +" movq 80(%%ecx), %%mm5\n\t" /* n.b.: pswapd*/ -" movq %%mm1, %%mm2\n\t" -" movq %%mm5, %%mm6\n\t" -" psrlq $32, %%mm1\n\t" -" psrlq $32, %%mm5\n\t" -" punpckldq %%mm2, %%mm1\n\t" -" punpckldq %%mm6, %%mm5\n\t" +" movq %%mm1, %%mm2\n\t" +" movq %%mm5, %%mm6\n\t" +" psrlq $32, %%mm1\n\t" +" psrlq $32, %%mm5\n\t" +" punpckldq %%mm2, %%mm1\n\t" +" punpckldq %%mm6, %%mm5\n\t" /**/ -" pfadd %%mm1, %%mm0\n\t" -" pfadd %%mm5, %%mm4\n\t" -" movq %%mm0, 64(%%edx)\n\t" -" movq %%mm4, 72(%%edx)\n\t" -" pfsub %%mm1, %%mm3\n\t" -" pfsub %%mm5, %%mm7\n\t" -" pfmul 96(%%ebx), %%mm3\n\t" -" pfmul 104(%%ebx), %%mm7\n\t" -" movd %%mm3, 92(%%edx)\n\t" -" movd %%mm7, 84(%%edx)\n\t" -" psrlq $32, %%mm3\n\t" -" psrlq $32, %%mm7\n\t" -" movd %%mm3, 88(%%edx)\n\t" -" movd %%mm7, 80(%%edx)\n\t" - -" movq 96(%%ecx), %%mm0\n\t" -" movq 104(%%ecx), %%mm4\n\t" -" movq %%mm0, %%mm3\n\t" -" movq %%mm4, %%mm7\n\t" -" movq 120(%%ecx), %%mm1\n\t" -" movq 112(%%ecx), %%mm5\n\t" +" pfadd %%mm1, %%mm0\n\t" +" pfadd %%mm5, %%mm4\n\t" +" movq %%mm0, 64(%%edx)\n\t" +" movq %%mm4, 72(%%edx)\n\t" +" pfsub %%mm1, %%mm3\n\t" +" pfsub %%mm5, %%mm7\n\t" +" pfmul 96(%%ebx), %%mm3\n\t" +" pfmul 104(%%ebx), %%mm7\n\t" +" movd %%mm3, 92(%%edx)\n\t" +" movd %%mm7, 84(%%edx)\n\t" +" psrlq $32, %%mm3\n\t" +" psrlq $32, %%mm7\n\t" +" movd %%mm3, 88(%%edx)\n\t" +" movd %%mm7, 80(%%edx)\n\t" + +" movq 96(%%ecx), %%mm0\n\t" +" movq 104(%%ecx), %%mm4\n\t" +" movq %%mm0, %%mm3\n\t" +" movq %%mm4, %%mm7\n\t" +" movq 120(%%ecx), %%mm1\n\t" +" movq 112(%%ecx), %%mm5\n\t" /* n.b.: pswapd*/ -" movq %%mm1, %%mm2\n\t" -" movq %%mm5, %%mm6\n\t" -" psrlq $32, %%mm1\n\t" -" psrlq $32, %%mm5\n\t" -" punpckldq %%mm2, %%mm1\n\t" -" punpckldq %%mm6, %%mm5\n\t" +" movq %%mm1, %%mm2\n\t" +" movq %%mm5, %%mm6\n\t" +" psrlq $32, %%mm1\n\t" +" psrlq $32, %%mm5\n\t" +" punpckldq %%mm2, %%mm1\n\t" +" punpckldq %%mm6, %%mm5\n\t" /**/ -" pfadd %%mm1, %%mm0\n\t" -" pfadd %%mm5, %%mm4\n\t" -" movq %%mm0, 96(%%edx)\n\t" -" movq %%mm4, 104(%%edx)\n\t" -" pfsubr %%mm1, %%mm3\n\t" -" pfsubr %%mm5, %%mm7\n\t" -" pfmul 96(%%ebx), %%mm3\n\t" -" pfmul 104(%%ebx), %%mm7\n\t" -" movd %%mm3, 124(%%edx)\n\t" -" movd %%mm7, 116(%%edx)\n\t" -" psrlq $32, %%mm3\n\t" -" psrlq $32, %%mm7\n\t" -" movd %%mm3, 120(%%edx)\n\t" -" movd %%mm7, 112(%%edx)\n\t" +" pfadd %%mm1, %%mm0\n\t" +" pfadd %%mm5, %%mm4\n\t" +" movq %%mm0, 96(%%edx)\n\t" +" movq %%mm4, 104(%%edx)\n\t" +" pfsubr %%mm1, %%mm3\n\t" +" pfsubr %%mm5, %%mm7\n\t" +" pfmul 96(%%ebx), %%mm3\n\t" +" pfmul 104(%%ebx), %%mm7\n\t" +" movd %%mm3, 124(%%edx)\n\t" +" movd %%mm7, 116(%%edx)\n\t" +" psrlq $32, %%mm3\n\t" +" psrlq $32, %%mm7\n\t" +" movd %%mm3, 120(%%edx)\n\t" +" movd %%mm7, 112(%%edx)\n\t" /* Phase 5 */ -" movq (%%edx), %%mm0\n\t" -" movq 16(%%edx), %%mm4\n\t" -" movq %%mm0, %%mm3\n\t" -" movq %%mm4, %%mm7\n\t" -" movq 8(%%edx), %%mm1\n\t" -" movq 24(%%edx), %%mm5\n\t" +" movq (%%edx), %%mm0\n\t" +" movq 16(%%edx), %%mm4\n\t" +" movq %%mm0, %%mm3\n\t" +" movq %%mm4, %%mm7\n\t" +" movq 8(%%edx), %%mm1\n\t" +" movq 24(%%edx), %%mm5\n\t" /* n.b.: pswapd*/ -" movq %%mm1, %%mm2\n\t" -" movq %%mm5, %%mm6\n\t" -" psrlq $32, %%mm1\n\t" -" psrlq $32, %%mm5\n\t" -" punpckldq %%mm2, %%mm1\n\t" -" punpckldq %%mm6, %%mm5\n\t" +" movq %%mm1, %%mm2\n\t" +" movq %%mm5, %%mm6\n\t" +" psrlq $32, %%mm1\n\t" +" psrlq $32, %%mm5\n\t" +" punpckldq %%mm2, %%mm1\n\t" +" punpckldq %%mm6, %%mm5\n\t" /**/ -" pfadd %%mm1, %%mm0\n\t" -" pfadd %%mm5, %%mm4\n\t" -" movq %%mm0, (%%ecx)\n\t" -" movq %%mm4, 16(%%ecx)\n\t" -" pfsub %%mm1, %%mm3\n\t" -" pfsubr %%mm5, %%mm7\n\t" -" pfmul 112(%%ebx), %%mm3\n\t" -" pfmul 112(%%ebx), %%mm7\n\t" -" movd %%mm3, 12(%%ecx)\n\t" -" movd %%mm7, 28(%%ecx)\n\t" -" psrlq $32, %%mm3\n\t" -" psrlq $32, %%mm7\n\t" -" movd %%mm3, 8(%%ecx)\n\t" -" movd %%mm7, 24(%%ecx)\n\t" - -" movq 32(%%edx), %%mm0\n\t" -" movq 48(%%edx), %%mm4\n\t" -" movq %%mm0, %%mm3\n\t" -" movq %%mm4, %%mm7\n\t" -" movq 40(%%edx), %%mm1\n\t" -" movq 56(%%edx), %%mm5\n\t" +" pfadd %%mm1, %%mm0\n\t" +" pfadd %%mm5, %%mm4\n\t" +" movq %%mm0, (%%ecx)\n\t" +" movq %%mm4, 16(%%ecx)\n\t" +" pfsub %%mm1, %%mm3\n\t" +" pfsubr %%mm5, %%mm7\n\t" +" pfmul 112(%%ebx), %%mm3\n\t" +" pfmul 112(%%ebx), %%mm7\n\t" +" movd %%mm3, 12(%%ecx)\n\t" +" movd %%mm7, 28(%%ecx)\n\t" +" psrlq $32, %%mm3\n\t" +" psrlq $32, %%mm7\n\t" +" movd %%mm3, 8(%%ecx)\n\t" +" movd %%mm7, 24(%%ecx)\n\t" + +" movq 32(%%edx), %%mm0\n\t" +" movq 48(%%edx), %%mm4\n\t" +" movq %%mm0, %%mm3\n\t" +" movq %%mm4, %%mm7\n\t" +" movq 40(%%edx), %%mm1\n\t" +" movq 56(%%edx), %%mm5\n\t" /* n.b.: pswapd*/ -" movq %%mm1, %%mm2\n\t" -" movq %%mm5, %%mm6\n\t" -" psrlq $32, %%mm1\n\t" -" psrlq $32, %%mm5\n\t" -" punpckldq %%mm2, %%mm1\n\t" -" punpckldq %%mm6, %%mm5\n\t" +" movq %%mm1, %%mm2\n\t" +" movq %%mm5, %%mm6\n\t" +" psrlq $32, %%mm1\n\t" +" psrlq $32, %%mm5\n\t" +" punpckldq %%mm2, %%mm1\n\t" +" punpckldq %%mm6, %%mm5\n\t" /**/ -" pfadd %%mm1, %%mm0\n\t" -" pfadd %%mm5, %%mm4\n\t" -" movq %%mm0, 32(%%ecx)\n\t" -" movq %%mm4, 48(%%ecx)\n\t" -" pfsub %%mm1, %%mm3\n\t" -" pfsubr %%mm5, %%mm7\n\t" -" pfmul 112(%%ebx), %%mm3\n\t" -" pfmul 112(%%ebx), %%mm7\n\t" -" movd %%mm3, 44(%%ecx)\n\t" -" movd %%mm7, 60(%%ecx)\n\t" -" psrlq $32, %%mm3\n\t" -" psrlq $32, %%mm7\n\t" -" movd %%mm3, 40(%%ecx)\n\t" -" movd %%mm7, 56(%%ecx)\n\t" - -" movq 64(%%edx), %%mm0\n\t" -" movq 80(%%edx), %%mm4\n\t" -" movq %%mm0, %%mm3\n\t" -" movq %%mm4, %%mm7\n\t" -" movq 72(%%edx), %%mm1\n\t" -" movq 88(%%edx), %%mm5\n\t" +" pfadd %%mm1, %%mm0\n\t" +" pfadd %%mm5, %%mm4\n\t" +" movq %%mm0, 32(%%ecx)\n\t" +" movq %%mm4, 48(%%ecx)\n\t" +" pfsub %%mm1, %%mm3\n\t" +" pfsubr %%mm5, %%mm7\n\t" +" pfmul 112(%%ebx), %%mm3\n\t" +" pfmul 112(%%ebx), %%mm7\n\t" +" movd %%mm3, 44(%%ecx)\n\t" +" movd %%mm7, 60(%%ecx)\n\t" +" psrlq $32, %%mm3\n\t" +" psrlq $32, %%mm7\n\t" +" movd %%mm3, 40(%%ecx)\n\t" +" movd %%mm7, 56(%%ecx)\n\t" + +" movq 64(%%edx), %%mm0\n\t" +" movq 80(%%edx), %%mm4\n\t" +" movq %%mm0, %%mm3\n\t" +" movq %%mm4, %%mm7\n\t" +" movq 72(%%edx), %%mm1\n\t" +" movq 88(%%edx), %%mm5\n\t" /* n.b.: pswapd*/ -" movq %%mm1, %%mm2\n\t" -" movq %%mm5, %%mm6\n\t" -" psrlq $32, %%mm1\n\t" -" psrlq $32, %%mm5\n\t" -" punpckldq %%mm2, %%mm1\n\t" -" punpckldq %%mm6, %%mm5\n\t" +" movq %%mm1, %%mm2\n\t" +" movq %%mm5, %%mm6\n\t" +" psrlq $32, %%mm1\n\t" +" psrlq $32, %%mm5\n\t" +" punpckldq %%mm2, %%mm1\n\t" +" punpckldq %%mm6, %%mm5\n\t" /**/ -" pfadd %%mm1, %%mm0\n\t" -" pfadd %%mm5, %%mm4\n\t" -" movq %%mm0, 64(%%ecx)\n\t" -" movq %%mm4, 80(%%ecx)\n\t" -" pfsub %%mm1, %%mm3\n\t" -" pfsubr %%mm5, %%mm7\n\t" -" pfmul 112(%%ebx), %%mm3\n\t" -" pfmul 112(%%ebx), %%mm7\n\t" -" movd %%mm3, 76(%%ecx)\n\t" -" movd %%mm7, 92(%%ecx)\n\t" -" psrlq $32, %%mm3\n\t" -" psrlq $32, %%mm7\n\t" -" movd %%mm3, 72(%%ecx)\n\t" -" movd %%mm7, 88(%%ecx)\n\t" - -" movq 96(%%edx), %%mm0\n\t" -" movq 112(%%edx), %%mm4\n\t" -" movq %%mm0, %%mm3\n\t" -" movq %%mm4, %%mm7\n\t" -" movq 104(%%edx), %%mm1\n\t" -" movq 120(%%edx), %%mm5\n\t" +" pfadd %%mm1, %%mm0\n\t" +" pfadd %%mm5, %%mm4\n\t" +" movq %%mm0, 64(%%ecx)\n\t" +" movq %%mm4, 80(%%ecx)\n\t" +" pfsub %%mm1, %%mm3\n\t" +" pfsubr %%mm5, %%mm7\n\t" +" pfmul 112(%%ebx), %%mm3\n\t" +" pfmul 112(%%ebx), %%mm7\n\t" +" movd %%mm3, 76(%%ecx)\n\t" +" movd %%mm7, 92(%%ecx)\n\t" +" psrlq $32, %%mm3\n\t" +" psrlq $32, %%mm7\n\t" +" movd %%mm3, 72(%%ecx)\n\t" +" movd %%mm7, 88(%%ecx)\n\t" + +" movq 96(%%edx), %%mm0\n\t" +" movq 112(%%edx), %%mm4\n\t" +" movq %%mm0, %%mm3\n\t" +" movq %%mm4, %%mm7\n\t" +" movq 104(%%edx), %%mm1\n\t" +" movq 120(%%edx), %%mm5\n\t" /* n.b.: pswapd*/ -" movq %%mm1, %%mm2\n\t" -" movq %%mm5, %%mm6\n\t" -" psrlq $32, %%mm1\n\t" -" psrlq $32, %%mm5\n\t" -" punpckldq %%mm2, %%mm1\n\t" -" punpckldq %%mm6, %%mm5\n\t" +" movq %%mm1, %%mm2\n\t" +" movq %%mm5, %%mm6\n\t" +" psrlq $32, %%mm1\n\t" +" psrlq $32, %%mm5\n\t" +" punpckldq %%mm2, %%mm1\n\t" +" punpckldq %%mm6, %%mm5\n\t" /**/ -" pfadd %%mm1, %%mm0\n\t" -" pfadd %%mm5, %%mm4\n\t" -" movq %%mm0, 96(%%ecx)\n\t" -" movq %%mm4, 112(%%ecx)\n\t" -" pfsub %%mm1, %%mm3\n\t" -" pfsubr %%mm5, %%mm7\n\t" -" pfmul 112(%%ebx), %%mm3\n\t" -" pfmul 112(%%ebx), %%mm7\n\t" -" movd %%mm3, 108(%%ecx)\n\t" -" movd %%mm7, 124(%%ecx)\n\t" -" psrlq $32, %%mm3\n\t" -" psrlq $32, %%mm7\n\t" -" movd %%mm3, 104(%%ecx)\n\t" -" movd %%mm7, 120(%%ecx)\n\t" +" pfadd %%mm1, %%mm0\n\t" +" pfadd %%mm5, %%mm4\n\t" +" movq %%mm0, 96(%%ecx)\n\t" +" movq %%mm4, 112(%%ecx)\n\t" +" pfsub %%mm1, %%mm3\n\t" +" pfsubr %%mm5, %%mm7\n\t" +" pfmul 112(%%ebx), %%mm3\n\t" +" pfmul 112(%%ebx), %%mm7\n\t" +" movd %%mm3, 108(%%ecx)\n\t" +" movd %%mm7, 124(%%ecx)\n\t" +" psrlq $32, %%mm3\n\t" +" psrlq $32, %%mm7\n\t" +" movd %%mm3, 104(%%ecx)\n\t" +" movd %%mm7, 120(%%ecx)\n\t" /* Phase 6. This is the end of easy road. */ /* Code below is coded in scalar mode. Should be optimized */ -" movd "MANGLE(plus_1f)", %%mm6\n\t" -" punpckldq 120(%%ebx), %%mm6\n\t" /* mm6 = 1.0 | 120(%%ebx)*/ -" movq "MANGLE(x_plus_minus_3dnow)", %%mm7\n\t" /* mm7 = +1 | -1 */ - -" movq 32(%%ecx), %%mm0\n\t" -" movq 64(%%ecx), %%mm2\n\t" -" movq %%mm0, %%mm1\n\t" -" movq %%mm2, %%mm3\n\t" -" pxor %%mm7, %%mm1\n\t" -" pxor %%mm7, %%mm3\n\t" -" pfacc %%mm1, %%mm0\n\t" -" pfacc %%mm3, %%mm2\n\t" -" pfmul %%mm6, %%mm0\n\t" -" pfmul %%mm6, %%mm2\n\t" -" movq %%mm0, 32(%%edx)\n\t" -" movq %%mm2, 64(%%edx)\n\t" - -" movd 44(%%ecx), %%mm0\n\t" -" movd 40(%%ecx), %%mm2\n\t" -" movd 120(%%ebx), %%mm3\n\t" -" punpckldq 76(%%ecx), %%mm0\n\t" -" punpckldq 72(%%ecx), %%mm2\n\t" -" punpckldq %%mm3, %%mm3\n\t" -" movq %%mm0, %%mm4\n\t" -" movq %%mm2, %%mm5\n\t" -" pfsub %%mm2, %%mm0\n\t" -" pfmul %%mm3, %%mm0\n\t" -" movq %%mm0, %%mm1\n\t" -" pfadd %%mm5, %%mm0\n\t" -" pfadd %%mm4, %%mm0\n\t" -" movq %%mm0, %%mm2\n\t" -" punpckldq %%mm1, %%mm0\n\t" -" punpckhdq %%mm1, %%mm2\n\t" -" movq %%mm0, 40(%%edx)\n\t" -" movq %%mm2, 72(%%edx)\n\t" - -" movd 48(%%ecx), %%mm3\n\t" -" movd 60(%%ecx), %%mm2\n\t" -" pfsub 52(%%ecx), %%mm3\n\t" -" pfsub 56(%%ecx), %%mm2\n\t" -" pfmul 120(%%ebx), %%mm3\n\t" -" pfmul 120(%%ebx), %%mm2\n\t" -" movq %%mm2, %%mm1\n\t" - -" pfadd 56(%%ecx), %%mm1\n\t" -" pfadd 60(%%ecx), %%mm1\n\t" -" movq %%mm1, %%mm0\n\t" - -" pfadd 48(%%ecx), %%mm0\n\t" -" pfadd 52(%%ecx), %%mm0\n\t" -" pfadd %%mm3, %%mm1\n\t" -" punpckldq %%mm2, %%mm1\n\t" -" pfadd %%mm3, %%mm2\n\t" -" punpckldq %%mm2, %%mm0\n\t" -" movq %%mm1, 56(%%edx)\n\t" -" movq %%mm0, 48(%%edx)\n\t" +" movd "MANGLE(plus_1f)", %%mm6\n\t" +" punpckldq 120(%%ebx), %%mm6\n\t" /* mm6 = 1.0 | 120(%%ebx)*/ +" movq "MANGLE(x_plus_minus_3dnow)", %%mm7\n\t" /* mm7 = +1 | -1 */ + +" movq 32(%%ecx), %%mm0\n\t" +" movq 64(%%ecx), %%mm2\n\t" +" movq %%mm0, %%mm1\n\t" +" movq %%mm2, %%mm3\n\t" +" pxor %%mm7, %%mm1\n\t" +" pxor %%mm7, %%mm3\n\t" +" pfacc %%mm1, %%mm0\n\t" +" pfacc %%mm3, %%mm2\n\t" +" pfmul %%mm6, %%mm0\n\t" +" pfmul %%mm6, %%mm2\n\t" +" movq %%mm0, 32(%%edx)\n\t" +" movq %%mm2, 64(%%edx)\n\t" + +" movd 44(%%ecx), %%mm0\n\t" +" movd 40(%%ecx), %%mm2\n\t" +" movd 120(%%ebx), %%mm3\n\t" +" punpckldq 76(%%ecx), %%mm0\n\t" +" punpckldq 72(%%ecx), %%mm2\n\t" +" punpckldq %%mm3, %%mm3\n\t" +" movq %%mm0, %%mm4\n\t" +" movq %%mm2, %%mm5\n\t" +" pfsub %%mm2, %%mm0\n\t" +" pfmul %%mm3, %%mm0\n\t" +" movq %%mm0, %%mm1\n\t" +" pfadd %%mm5, %%mm0\n\t" +" pfadd %%mm4, %%mm0\n\t" +" movq %%mm0, %%mm2\n\t" +" punpckldq %%mm1, %%mm0\n\t" +" punpckhdq %%mm1, %%mm2\n\t" +" movq %%mm0, 40(%%edx)\n\t" +" movq %%mm2, 72(%%edx)\n\t" + +" movd 48(%%ecx), %%mm3\n\t" +" movd 60(%%ecx), %%mm2\n\t" +" pfsub 52(%%ecx), %%mm3\n\t" +" pfsub 56(%%ecx), %%mm2\n\t" +" pfmul 120(%%ebx), %%mm3\n\t" +" pfmul 120(%%ebx), %%mm2\n\t" +" movq %%mm2, %%mm1\n\t" + +" pfadd 56(%%ecx), %%mm1\n\t" +" pfadd 60(%%ecx), %%mm1\n\t" +" movq %%mm1, %%mm0\n\t" + +" pfadd 48(%%ecx), %%mm0\n\t" +" pfadd 52(%%ecx), %%mm0\n\t" +" pfadd %%mm3, %%mm1\n\t" +" punpckldq %%mm2, %%mm1\n\t" +" pfadd %%mm3, %%mm2\n\t" +" punpckldq %%mm2, %%mm0\n\t" +" movq %%mm1, 56(%%edx)\n\t" +" movq %%mm0, 48(%%edx)\n\t" /*---*/ -" movd 92(%%ecx), %%mm1\n\t" -" pfsub 88(%%ecx), %%mm1\n\t" -" pfmul 120(%%ebx), %%mm1\n\t" -" movd %%mm1, 92(%%edx)\n\t" -" pfadd 92(%%ecx), %%mm1\n\t" -" pfadd 88(%%ecx), %%mm1\n\t" -" movq %%mm1, %%mm0\n\t" - -" pfadd 80(%%ecx), %%mm0\n\t" -" pfadd 84(%%ecx), %%mm0\n\t" -" movd %%mm0, 80(%%edx)\n\t" - -" movd 80(%%ecx), %%mm0\n\t" -" pfsub 84(%%ecx), %%mm0\n\t" -" pfmul 120(%%ebx), %%mm0\n\t" -" pfadd %%mm0, %%mm1\n\t" -" pfadd 92(%%edx), %%mm0\n\t" -" punpckldq %%mm1, %%mm0\n\t" -" movq %%mm0, 84(%%edx)\n\t" - -" movq 96(%%ecx), %%mm0\n\t" -" movq %%mm0, %%mm1\n\t" -" pxor %%mm7, %%mm1\n\t" -" pfacc %%mm1, %%mm0\n\t" -" pfmul %%mm6, %%mm0\n\t" -" movq %%mm0, 96(%%edx)\n\t" - -" movd 108(%%ecx), %%mm0\n\t" -" pfsub 104(%%ecx), %%mm0\n\t" -" pfmul 120(%%ebx), %%mm0\n\t" -" movd %%mm0, 108(%%edx)\n\t" -" pfadd 104(%%ecx), %%mm0\n\t" -" pfadd 108(%%ecx), %%mm0\n\t" -" movd %%mm0, 104(%%edx)\n\t" - -" movd 124(%%ecx), %%mm1\n\t" -" pfsub 120(%%ecx), %%mm1\n\t" -" pfmul 120(%%ebx), %%mm1\n\t" -" movd %%mm1, 124(%%edx)\n\t" -" pfadd 120(%%ecx), %%mm1\n\t" -" pfadd 124(%%ecx), %%mm1\n\t" -" movq %%mm1, %%mm0\n\t" - -" pfadd 112(%%ecx), %%mm0\n\t" -" pfadd 116(%%ecx), %%mm0\n\t" -" movd %%mm0, 112(%%edx)\n\t" - -" movd 112(%%ecx), %%mm0\n\t" -" pfsub 116(%%ecx), %%mm0\n\t" -" pfmul 120(%%ebx), %%mm0\n\t" -" pfadd %%mm0,%%mm1\n\t" -" pfadd 124(%%edx), %%mm0\n\t" -" punpckldq %%mm1, %%mm0\n\t" -" movq %%mm0, 116(%%edx)\n\t" +" movd 92(%%ecx), %%mm1\n\t" +" pfsub 88(%%ecx), %%mm1\n\t" +" pfmul 120(%%ebx), %%mm1\n\t" +" movd %%mm1, 92(%%edx)\n\t" +" pfadd 92(%%ecx), %%mm1\n\t" +" pfadd 88(%%ecx), %%mm1\n\t" +" movq %%mm1, %%mm0\n\t" + +" pfadd 80(%%ecx), %%mm0\n\t" +" pfadd 84(%%ecx), %%mm0\n\t" +" movd %%mm0, 80(%%edx)\n\t" + +" movd 80(%%ecx), %%mm0\n\t" +" pfsub 84(%%ecx), %%mm0\n\t" +" pfmul 120(%%ebx), %%mm0\n\t" +" pfadd %%mm0, %%mm1\n\t" +" pfadd 92(%%edx), %%mm0\n\t" +" punpckldq %%mm1, %%mm0\n\t" +" movq %%mm0, 84(%%edx)\n\t" + +" movq 96(%%ecx), %%mm0\n\t" +" movq %%mm0, %%mm1\n\t" +" pxor %%mm7, %%mm1\n\t" +" pfacc %%mm1, %%mm0\n\t" +" pfmul %%mm6, %%mm0\n\t" +" movq %%mm0, 96(%%edx)\n\t" + +" movd 108(%%ecx), %%mm0\n\t" +" pfsub 104(%%ecx), %%mm0\n\t" +" pfmul 120(%%ebx), %%mm0\n\t" +" movd %%mm0, 108(%%edx)\n\t" +" pfadd 104(%%ecx), %%mm0\n\t" +" pfadd 108(%%ecx), %%mm0\n\t" +" movd %%mm0, 104(%%edx)\n\t" + +" movd 124(%%ecx), %%mm1\n\t" +" pfsub 120(%%ecx), %%mm1\n\t" +" pfmul 120(%%ebx), %%mm1\n\t" +" movd %%mm1, 124(%%edx)\n\t" +" pfadd 120(%%ecx), %%mm1\n\t" +" pfadd 124(%%ecx), %%mm1\n\t" +" movq %%mm1, %%mm0\n\t" + +" pfadd 112(%%ecx), %%mm0\n\t" +" pfadd 116(%%ecx), %%mm0\n\t" +" movd %%mm0, 112(%%edx)\n\t" + +" movd 112(%%ecx), %%mm0\n\t" +" pfsub 116(%%ecx), %%mm0\n\t" +" pfmul 120(%%ebx), %%mm0\n\t" +" pfadd %%mm0,%%mm1\n\t" +" pfadd 124(%%edx), %%mm0\n\t" +" punpckldq %%mm1, %%mm0\n\t" +" movq %%mm0, 116(%%edx)\n\t" // this code is broken, there is nothing modifying the z flag above. #if 0 -" jnz .L01\n\t" +" jnz .L01\n\t" /* Phase 7*/ /* Code below is coded in scalar mode. Should be optimized */ -" movd (%%ecx), %%mm0\n\t" -" pfadd 4(%%ecx), %%mm0\n\t" -" movd %%mm0, 1024(%%esi)\n\t" - -" movd (%%ecx), %%mm0\n\t" -" pfsub 4(%%ecx), %%mm0\n\t" -" pfmul 120(%%ebx), %%mm0\n\t" -" movd %%mm0, (%%esi)\n\t" -" movd %%mm0, (%%edi)\n\t" - -" movd 12(%%ecx), %%mm0\n\t" -" pfsub 8(%%ecx), %%mm0\n\t" -" pfmul 120(%%ebx), %%mm0\n\t" -" movd %%mm0, 512(%%edi)\n\t" -" pfadd 12(%%ecx), %%mm0\n\t" -" pfadd 8(%%ecx), %%mm0\n\t" -" movd %%mm0, 512(%%esi)\n\t" - -" movd 16(%%ecx), %%mm0\n\t" -" pfsub 20(%%ecx), %%mm0\n\t" -" pfmul 120(%%ebx), %%mm0\n\t" -" movq %%mm0, %%mm3\n\t" - -" movd 28(%%ecx), %%mm0\n\t" -" pfsub 24(%%ecx), %%mm0\n\t" -" pfmul 120(%%ebx), %%mm0\n\t" -" movd %%mm0, 768(%%edi)\n\t" -" movq %%mm0, %%mm2\n\t" - -" pfadd 24(%%ecx), %%mm0\n\t" -" pfadd 28(%%ecx), %%mm0\n\t" -" movq %%mm0, %%mm1\n\t" - -" pfadd 16(%%ecx), %%mm0\n\t" -" pfadd 20(%%ecx), %%mm0\n\t" -" movd %%mm0, 768(%%esi)\n\t" -" pfadd %%mm3, %%mm1\n\t" -" movd %%mm1, 256(%%esi)\n\t" -" pfadd %%mm3, %%mm2\n\t" -" movd %%mm2, 256(%%edi)\n\t" +" movd (%%ecx), %%mm0\n\t" +" pfadd 4(%%ecx), %%mm0\n\t" +" movd %%mm0, 1024(%%esi)\n\t" + +" movd (%%ecx), %%mm0\n\t" +" pfsub 4(%%ecx), %%mm0\n\t" +" pfmul 120(%%ebx), %%mm0\n\t" +" movd %%mm0, (%%esi)\n\t" +" movd %%mm0, (%%edi)\n\t" + +" movd 12(%%ecx), %%mm0\n\t" +" pfsub 8(%%ecx), %%mm0\n\t" +" pfmul 120(%%ebx), %%mm0\n\t" +" movd %%mm0, 512(%%edi)\n\t" +" pfadd 12(%%ecx), %%mm0\n\t" +" pfadd 8(%%ecx), %%mm0\n\t" +" movd %%mm0, 512(%%esi)\n\t" + +" movd 16(%%ecx), %%mm0\n\t" +" pfsub 20(%%ecx), %%mm0\n\t" +" pfmul 120(%%ebx), %%mm0\n\t" +" movq %%mm0, %%mm3\n\t" + +" movd 28(%%ecx), %%mm0\n\t" +" pfsub 24(%%ecx), %%mm0\n\t" +" pfmul 120(%%ebx), %%mm0\n\t" +" movd %%mm0, 768(%%edi)\n\t" +" movq %%mm0, %%mm2\n\t" + +" pfadd 24(%%ecx), %%mm0\n\t" +" pfadd 28(%%ecx), %%mm0\n\t" +" movq %%mm0, %%mm1\n\t" + +" pfadd 16(%%ecx), %%mm0\n\t" +" pfadd 20(%%ecx), %%mm0\n\t" +" movd %%mm0, 768(%%esi)\n\t" +" pfadd %%mm3, %%mm1\n\t" +" movd %%mm1, 256(%%esi)\n\t" +" pfadd %%mm3, %%mm2\n\t" +" movd %%mm2, 256(%%edi)\n\t" /* Phase 8*/ -" movq 32(%%edx), %%mm0\n\t" -" movq 48(%%edx), %%mm1\n\t" -" pfadd 48(%%edx), %%mm0\n\t" -" pfadd 40(%%edx), %%mm1\n\t" -" movd %%mm0, 896(%%esi)\n\t" -" movd %%mm1, 640(%%esi)\n\t" -" psrlq $32, %%mm0\n\t" -" psrlq $32, %%mm1\n\t" -" movd %%mm0, 128(%%edi)\n\t" -" movd %%mm1, 384(%%edi)\n\t" - -" movd 40(%%edx), %%mm0\n\t" -" pfadd 56(%%edx), %%mm0\n\t" -" movd %%mm0, 384(%%esi)\n\t" - -" movd 56(%%edx), %%mm0\n\t" -" pfadd 36(%%edx), %%mm0\n\t" -" movd %%mm0, 128(%%esi)\n\t" - -" movd 60(%%edx), %%mm0\n\t" -" movd %%mm0, 896(%%edi)\n\t" -" pfadd 44(%%edx), %%mm0\n\t" -" movd %%mm0, 640(%%edi)\n\t" - -" movq 96(%%edx), %%mm0\n\t" -" movq 112(%%edx), %%mm2\n\t" -" movq 104(%%edx), %%mm4\n\t" -" pfadd 112(%%edx), %%mm0\n\t" -" pfadd 104(%%edx), %%mm2\n\t" -" pfadd 120(%%edx), %%mm4\n\t" -" movq %%mm0, %%mm1\n\t" -" movq %%mm2, %%mm3\n\t" -" movq %%mm4, %%mm5\n\t" -" pfadd 64(%%edx), %%mm0\n\t" -" pfadd 80(%%edx), %%mm2\n\t" -" pfadd 72(%%edx), %%mm4\n\t" -" movd %%mm0, 960(%%esi)\n\t" -" movd %%mm2, 704(%%esi)\n\t" -" movd %%mm4, 448(%%esi)\n\t" -" psrlq $32, %%mm0\n\t" -" psrlq $32, %%mm2\n\t" -" psrlq $32, %%mm4\n\t" -" movd %%mm0, 64(%%edi)\n\t" -" movd %%mm2, 320(%%edi)\n\t" -" movd %%mm4, 576(%%edi)\n\t" -" pfadd 80(%%edx), %%mm1\n\t" -" pfadd 72(%%edx), %%mm3\n\t" -" pfadd 88(%%edx), %%mm5\n\t" -" movd %%mm1, 832(%%esi)\n\t" -" movd %%mm3, 576(%%esi)\n\t" -" movd %%mm5, 320(%%esi)\n\t" -" psrlq $32, %%mm1\n\t" -" psrlq $32, %%mm3\n\t" -" psrlq $32, %%mm5\n\t" -" movd %%mm1, 192(%%edi)\n\t" -" movd %%mm3, 448(%%edi)\n\t" -" movd %%mm5, 704(%%edi)\n\t" - -" movd 120(%%edx), %%mm0\n\t" -" pfadd 100(%%edx), %%mm0\n\t" -" movq %%mm0, %%mm1\n\t" -" pfadd 88(%%edx), %%mm0\n\t" -" movd %%mm0, 192(%%esi)\n\t" -" pfadd 68(%%edx), %%mm1\n\t" -" movd %%mm1, 64(%%esi)\n\t" - -" movd 124(%%edx), %%mm0\n\t" -" movd %%mm0, 960(%%edi)\n\t" -" pfadd 92(%%edx), %%mm0\n\t" -" movd %%mm0, 832(%%edi)\n\t" - -" jmp .L_bye\n\t" +" movq 32(%%edx), %%mm0\n\t" +" movq 48(%%edx), %%mm1\n\t" +" pfadd 48(%%edx), %%mm0\n\t" +" pfadd 40(%%edx), %%mm1\n\t" +" movd %%mm0, 896(%%esi)\n\t" +" movd %%mm1, 640(%%esi)\n\t" +" psrlq $32, %%mm0\n\t" +" psrlq $32, %%mm1\n\t" +" movd %%mm0, 128(%%edi)\n\t" +" movd %%mm1, 384(%%edi)\n\t" + +" movd 40(%%edx), %%mm0\n\t" +" pfadd 56(%%edx), %%mm0\n\t" +" movd %%mm0, 384(%%esi)\n\t" + +" movd 56(%%edx), %%mm0\n\t" +" pfadd 36(%%edx), %%mm0\n\t" +" movd %%mm0, 128(%%esi)\n\t" + +" movd 60(%%edx), %%mm0\n\t" +" movd %%mm0, 896(%%edi)\n\t" +" pfadd 44(%%edx), %%mm0\n\t" +" movd %%mm0, 640(%%edi)\n\t" + +" movq 96(%%edx), %%mm0\n\t" +" movq 112(%%edx), %%mm2\n\t" +" movq 104(%%edx), %%mm4\n\t" +" pfadd 112(%%edx), %%mm0\n\t" +" pfadd 104(%%edx), %%mm2\n\t" +" pfadd 120(%%edx), %%mm4\n\t" +" movq %%mm0, %%mm1\n\t" +" movq %%mm2, %%mm3\n\t" +" movq %%mm4, %%mm5\n\t" +" pfadd 64(%%edx), %%mm0\n\t" +" pfadd 80(%%edx), %%mm2\n\t" +" pfadd 72(%%edx), %%mm4\n\t" +" movd %%mm0, 960(%%esi)\n\t" +" movd %%mm2, 704(%%esi)\n\t" +" movd %%mm4, 448(%%esi)\n\t" +" psrlq $32, %%mm0\n\t" +" psrlq $32, %%mm2\n\t" +" psrlq $32, %%mm4\n\t" +" movd %%mm0, 64(%%edi)\n\t" +" movd %%mm2, 320(%%edi)\n\t" +" movd %%mm4, 576(%%edi)\n\t" +" pfadd 80(%%edx), %%mm1\n\t" +" pfadd 72(%%edx), %%mm3\n\t" +" pfadd 88(%%edx), %%mm5\n\t" +" movd %%mm1, 832(%%esi)\n\t" +" movd %%mm3, 576(%%esi)\n\t" +" movd %%mm5, 320(%%esi)\n\t" +" psrlq $32, %%mm1\n\t" +" psrlq $32, %%mm3\n\t" +" psrlq $32, %%mm5\n\t" +" movd %%mm1, 192(%%edi)\n\t" +" movd %%mm3, 448(%%edi)\n\t" +" movd %%mm5, 704(%%edi)\n\t" + +" movd 120(%%edx), %%mm0\n\t" +" pfadd 100(%%edx), %%mm0\n\t" +" movq %%mm0, %%mm1\n\t" +" pfadd 88(%%edx), %%mm0\n\t" +" movd %%mm0, 192(%%esi)\n\t" +" pfadd 68(%%edx), %%mm1\n\t" +" movd %%mm1, 64(%%esi)\n\t" + +" movd 124(%%edx), %%mm0\n\t" +" movd %%mm0, 960(%%edi)\n\t" +" pfadd 92(%%edx), %%mm0\n\t" +" movd %%mm0, 832(%%edi)\n\t" + +" jmp .L_bye\n\t" ".L01:\n\t" #endif /* Phase 9*/ -" movq (%%ecx), %%mm0\n\t" -" movq %%mm0, %%mm1\n\t" -" pxor %%mm7, %%mm1\n\t" -" pfacc %%mm1, %%mm0\n\t" -" pfmul %%mm6, %%mm0\n\t" -" pf2id %%mm0, %%mm0\n\t" -" packssdw %%mm0, %%mm0\n\t" -" movd %%mm0, %%eax\n\t" -" movw %%ax, 512(%%esi)\n\t" -" shrl $16, %%eax\n\t" -" movw %%ax, (%%esi)\n\t" - -" movd 12(%%ecx), %%mm0\n\t" -" pfsub 8(%%ecx), %%mm0\n\t" -" pfmul 120(%%ebx), %%mm0\n\t" -" pf2id %%mm0, %%mm7\n\t" -" packssdw %%mm7, %%mm7\n\t" -" movd %%mm7, %%eax\n\t" -" movw %%ax, 256(%%edi)\n\t" -" pfadd 12(%%ecx), %%mm0\n\t" -" pfadd 8(%%ecx), %%mm0\n\t" -" pf2id %%mm0, %%mm0\n\t" -" packssdw %%mm0, %%mm0\n\t" -" movd %%mm0, %%eax\n\t" -" movw %%ax, 256(%%esi)\n\t" - -" movd 16(%%ecx), %%mm3\n\t" -" pfsub 20(%%ecx), %%mm3\n\t" -" pfmul 120(%%ebx), %%mm3\n\t" -" movq %%mm3, %%mm2\n\t" - -" movd 28(%%ecx), %%mm2\n\t" -" pfsub 24(%%ecx), %%mm2\n\t" -" pfmul 120(%%ebx), %%mm2\n\t" -" movq %%mm2, %%mm1\n\t" - -" pf2id %%mm2, %%mm7\n\t" -" packssdw %%mm7, %%mm7\n\t" -" movd %%mm7, %%eax\n\t" -" movw %%ax, 384(%%edi)\n\t" - -" pfadd 24(%%ecx), %%mm1\n\t" -" pfadd 28(%%ecx), %%mm1\n\t" -" movq %%mm1, %%mm0\n\t" - -" pfadd 16(%%ecx), %%mm0\n\t" -" pfadd 20(%%ecx), %%mm0\n\t" -" pf2id %%mm0, %%mm0\n\t" -" packssdw %%mm0, %%mm0\n\t" -" movd %%mm0, %%eax\n\t" -" movw %%ax, 384(%%esi)\n\t" -" pfadd %%mm3, %%mm1\n\t" -" pf2id %%mm1, %%mm1\n\t" -" packssdw %%mm1, %%mm1\n\t" -" movd %%mm1, %%eax\n\t" -" movw %%ax, 128(%%esi)\n\t" -" pfadd %%mm3, %%mm2\n\t" -" pf2id %%mm2, %%mm2\n\t" -" packssdw %%mm2, %%mm2\n\t" -" movd %%mm2, %%eax\n\t" -" movw %%ax, 128(%%edi)\n\t" +" movq (%%ecx), %%mm0\n\t" +" movq %%mm0, %%mm1\n\t" +" pxor %%mm7, %%mm1\n\t" +" pfacc %%mm1, %%mm0\n\t" +" pfmul %%mm6, %%mm0\n\t" +" pf2id %%mm0, %%mm0\n\t" +" packssdw %%mm0, %%mm0\n\t" +" movd %%mm0, %%eax\n\t" +" movw %%ax, 512(%%esi)\n\t" +" shrl $16, %%eax\n\t" +" movw %%ax, (%%esi)\n\t" + +" movd 12(%%ecx), %%mm0\n\t" +" pfsub 8(%%ecx), %%mm0\n\t" +" pfmul 120(%%ebx), %%mm0\n\t" +" pf2id %%mm0, %%mm7\n\t" +" packssdw %%mm7, %%mm7\n\t" +" movd %%mm7, %%eax\n\t" +" movw %%ax, 256(%%edi)\n\t" +" pfadd 12(%%ecx), %%mm0\n\t" +" pfadd 8(%%ecx), %%mm0\n\t" +" pf2id %%mm0, %%mm0\n\t" +" packssdw %%mm0, %%mm0\n\t" +" movd %%mm0, %%eax\n\t" +" movw %%ax, 256(%%esi)\n\t" + +" movd 16(%%ecx), %%mm3\n\t" +" pfsub 20(%%ecx), %%mm3\n\t" +" pfmul 120(%%ebx), %%mm3\n\t" +" movq %%mm3, %%mm2\n\t" + +" movd 28(%%ecx), %%mm2\n\t" +" pfsub 24(%%ecx), %%mm2\n\t" +" pfmul 120(%%ebx), %%mm2\n\t" +" movq %%mm2, %%mm1\n\t" + +" pf2id %%mm2, %%mm7\n\t" +" packssdw %%mm7, %%mm7\n\t" +" movd %%mm7, %%eax\n\t" +" movw %%ax, 384(%%edi)\n\t" + +" pfadd 24(%%ecx), %%mm1\n\t" +" pfadd 28(%%ecx), %%mm1\n\t" +" movq %%mm1, %%mm0\n\t" + +" pfadd 16(%%ecx), %%mm0\n\t" +" pfadd 20(%%ecx), %%mm0\n\t" +" pf2id %%mm0, %%mm0\n\t" +" packssdw %%mm0, %%mm0\n\t" +" movd %%mm0, %%eax\n\t" +" movw %%ax, 384(%%esi)\n\t" +" pfadd %%mm3, %%mm1\n\t" +" pf2id %%mm1, %%mm1\n\t" +" packssdw %%mm1, %%mm1\n\t" +" movd %%mm1, %%eax\n\t" +" movw %%ax, 128(%%esi)\n\t" +" pfadd %%mm3, %%mm2\n\t" +" pf2id %%mm2, %%mm2\n\t" +" packssdw %%mm2, %%mm2\n\t" +" movd %%mm2, %%eax\n\t" +" movw %%ax, 128(%%edi)\n\t" /* Phase 10*/ -" movq 32(%%edx), %%mm0\n\t" -" movq 48(%%edx), %%mm1\n\t" -" pfadd 48(%%edx), %%mm0\n\t" -" pfadd 40(%%edx), %%mm1\n\t" -" pf2id %%mm0, %%mm0\n\t" -" pf2id %%mm1, %%mm1\n\t" -" packssdw %%mm0, %%mm0\n\t" -" packssdw %%mm1, %%mm1\n\t" -" movd %%mm0, %%eax\n\t" -" movd %%mm1, %%ecx\n\t" -" movw %%ax, 448(%%esi)\n\t" -" movw %%cx, 320(%%esi)\n\t" -" shrl $16, %%eax\n\t" -" shrl $16, %%ecx\n\t" -" movw %%ax, 64(%%edi)\n\t" -" movw %%cx, 192(%%edi)\n\t" - -" movd 40(%%edx), %%mm3\n\t" -" movd 56(%%edx), %%mm4\n\t" -" movd 60(%%edx), %%mm0\n\t" -" movd 44(%%edx), %%mm2\n\t" -" movd 120(%%edx), %%mm5\n\t" -" punpckldq %%mm4, %%mm3\n\t" -" punpckldq 124(%%edx), %%mm0\n\t" -" pfadd 100(%%edx), %%mm5\n\t" -" punpckldq 36(%%edx), %%mm4\n\t" -" punpckldq 92(%%edx), %%mm2\n\t" -" movq %%mm5, %%mm6\n\t" -" pfadd %%mm4, %%mm3\n\t" -" pf2id %%mm0, %%mm1\n\t" -" pf2id %%mm3, %%mm3\n\t" -" packssdw %%mm1, %%mm1\n\t" -" packssdw %%mm3, %%mm3\n\t" -" pfadd 88(%%edx), %%mm5\n\t" -" movd %%mm1, %%eax\n\t" -" movd %%mm3, %%ecx\n\t" -" movw %%ax, 448(%%edi)\n\t" -" movw %%cx, 192(%%esi)\n\t" -" pf2id %%mm5, %%mm5\n\t" -" packssdw %%mm5, %%mm5\n\t" -" shrl $16, %%eax\n\t" -" shrl $16, %%ecx\n\t" -" movd %%mm5, %%ebx\n\t" -" movw %%bx, 96(%%esi)\n\t" -" movw %%ax, 480(%%edi)\n\t" -" movw %%cx, 64(%%esi)\n\t" -" pfadd %%mm2, %%mm0\n\t" -" pf2id %%mm0, %%mm0\n\t" -" packssdw %%mm0, %%mm0\n\t" -" movd %%mm0, %%eax\n\t" -" pfadd 68(%%edx), %%mm6\n\t" -" movw %%ax, 320(%%edi)\n\t" -" shr $16, %%eax\n\t" -" pf2id %%mm6, %%mm6\n\t" -" packssdw %%mm6, %%mm6\n\t" -" movd %%mm6, %%ebx\n\t" -" movw %%ax, 416(%%edi)\n\t" -" movw %%bx, 32(%%esi)\n\t" - -" movq 96(%%edx), %%mm0\n\t" -" movq 112(%%edx), %%mm2\n\t" -" movq 104(%%edx), %%mm4\n\t" -" pfadd %%mm2, %%mm0\n\t" -" pfadd %%mm4, %%mm2\n\t" -" pfadd 120(%%edx), %%mm4\n\t" -" movq %%mm0, %%mm1\n\t" -" movq %%mm2, %%mm3\n\t" -" movq %%mm4, %%mm5\n\t" -" pfadd 64(%%edx), %%mm0\n\t" -" pfadd 80(%%edx), %%mm2\n\t" -" pfadd 72(%%edx), %%mm4\n\t" -" pf2id %%mm0, %%mm0\n\t" -" pf2id %%mm2, %%mm2\n\t" -" pf2id %%mm4, %%mm4\n\t" -" packssdw %%mm0, %%mm0\n\t" -" packssdw %%mm2, %%mm2\n\t" -" packssdw %%mm4, %%mm4\n\t" -" movd %%mm0, %%eax\n\t" -" movd %%mm2, %%ecx\n\t" -" movd %%mm4, %%ebx\n\t" -" movw %%ax, 480(%%esi)\n\t" -" movw %%cx, 352(%%esi)\n\t" -" movw %%bx, 224(%%esi)\n\t" -" shrl $16, %%eax\n\t" -" shrl $16, %%ecx\n\t" -" shrl $16, %%ebx\n\t" -" movw %%ax, 32(%%edi)\n\t" -" movw %%cx, 160(%%edi)\n\t" -" movw %%bx, 288(%%edi)\n\t" -" pfadd 80(%%edx), %%mm1\n\t" -" pfadd 72(%%edx), %%mm3\n\t" -" pfadd 88(%%edx), %%mm5\n\t" -" pf2id %%mm1, %%mm1\n\t" -" pf2id %%mm3, %%mm3\n\t" -" pf2id %%mm5, %%mm5\n\t" -" packssdw %%mm1, %%mm1\n\t" -" packssdw %%mm3, %%mm3\n\t" -" packssdw %%mm5, %%mm5\n\t" -" movd %%mm1, %%eax\n\t" -" movd %%mm3, %%ecx\n\t" -" movd %%mm5, %%ebx\n\t" -" movw %%ax, 416(%%esi)\n\t" -" movw %%cx, 288(%%esi)\n\t" -" movw %%bx, 160(%%esi)\n\t" -" shrl $16, %%eax\n\t" -" shrl $16, %%ecx\n\t" -" shrl $16, %%ebx\n\t" -" movw %%ax, 96(%%edi)\n\t" -" movw %%cx, 224(%%edi)\n\t" -" movw %%bx, 352(%%edi)\n\t" - -" movsw\n\t" +" movq 32(%%edx), %%mm0\n\t" +" movq 48(%%edx), %%mm1\n\t" +" pfadd 48(%%edx), %%mm0\n\t" +" pfadd 40(%%edx), %%mm1\n\t" +" pf2id %%mm0, %%mm0\n\t" +" pf2id %%mm1, %%mm1\n\t" +" packssdw %%mm0, %%mm0\n\t" +" packssdw %%mm1, %%mm1\n\t" +" movd %%mm0, %%eax\n\t" +" movd %%mm1, %%ecx\n\t" +" movw %%ax, 448(%%esi)\n\t" +" movw %%cx, 320(%%esi)\n\t" +" shrl $16, %%eax\n\t" +" shrl $16, %%ecx\n\t" +" movw %%ax, 64(%%edi)\n\t" +" movw %%cx, 192(%%edi)\n\t" + +" movd 40(%%edx), %%mm3\n\t" +" movd 56(%%edx), %%mm4\n\t" +" movd 60(%%edx), %%mm0\n\t" +" movd 44(%%edx), %%mm2\n\t" +" movd 120(%%edx), %%mm5\n\t" +" punpckldq %%mm4, %%mm3\n\t" +" punpckldq 124(%%edx), %%mm0\n\t" +" pfadd 100(%%edx), %%mm5\n\t" +" punpckldq 36(%%edx), %%mm4\n\t" +" punpckldq 92(%%edx), %%mm2\n\t" +" movq %%mm5, %%mm6\n\t" +" pfadd %%mm4, %%mm3\n\t" +" pf2id %%mm0, %%mm1\n\t" +" pf2id %%mm3, %%mm3\n\t" +" packssdw %%mm1, %%mm1\n\t" +" packssdw %%mm3, %%mm3\n\t" +" pfadd 88(%%edx), %%mm5\n\t" +" movd %%mm1, %%eax\n\t" +" movd %%mm3, %%ecx\n\t" +" movw %%ax, 448(%%edi)\n\t" +" movw %%cx, 192(%%esi)\n\t" +" pf2id %%mm5, %%mm5\n\t" +" packssdw %%mm5, %%mm5\n\t" +" shrl $16, %%eax\n\t" +" shrl $16, %%ecx\n\t" +" movd %%mm5, %%ebx\n\t" +" movw %%bx, 96(%%esi)\n\t" +" movw %%ax, 480(%%edi)\n\t" +" movw %%cx, 64(%%esi)\n\t" +" pfadd %%mm2, %%mm0\n\t" +" pf2id %%mm0, %%mm0\n\t" +" packssdw %%mm0, %%mm0\n\t" +" movd %%mm0, %%eax\n\t" +" pfadd 68(%%edx), %%mm6\n\t" +" movw %%ax, 320(%%edi)\n\t" +" shr $16, %%eax\n\t" +" pf2id %%mm6, %%mm6\n\t" +" packssdw %%mm6, %%mm6\n\t" +" movd %%mm6, %%ebx\n\t" +" movw %%ax, 416(%%edi)\n\t" +" movw %%bx, 32(%%esi)\n\t" + +" movq 96(%%edx), %%mm0\n\t" +" movq 112(%%edx), %%mm2\n\t" +" movq 104(%%edx), %%mm4\n\t" +" pfadd %%mm2, %%mm0\n\t" +" pfadd %%mm4, %%mm2\n\t" +" pfadd 120(%%edx), %%mm4\n\t" +" movq %%mm0, %%mm1\n\t" +" movq %%mm2, %%mm3\n\t" +" movq %%mm4, %%mm5\n\t" +" pfadd 64(%%edx), %%mm0\n\t" +" pfadd 80(%%edx), %%mm2\n\t" +" pfadd 72(%%edx), %%mm4\n\t" +" pf2id %%mm0, %%mm0\n\t" +" pf2id %%mm2, %%mm2\n\t" +" pf2id %%mm4, %%mm4\n\t" +" packssdw %%mm0, %%mm0\n\t" +" packssdw %%mm2, %%mm2\n\t" +" packssdw %%mm4, %%mm4\n\t" +" movd %%mm0, %%eax\n\t" +" movd %%mm2, %%ecx\n\t" +" movd %%mm4, %%ebx\n\t" +" movw %%ax, 480(%%esi)\n\t" +" movw %%cx, 352(%%esi)\n\t" +" movw %%bx, 224(%%esi)\n\t" +" shrl $16, %%eax\n\t" +" shrl $16, %%ecx\n\t" +" shrl $16, %%ebx\n\t" +" movw %%ax, 32(%%edi)\n\t" +" movw %%cx, 160(%%edi)\n\t" +" movw %%bx, 288(%%edi)\n\t" +" pfadd 80(%%edx), %%mm1\n\t" +" pfadd 72(%%edx), %%mm3\n\t" +" pfadd 88(%%edx), %%mm5\n\t" +" pf2id %%mm1, %%mm1\n\t" +" pf2id %%mm3, %%mm3\n\t" +" pf2id %%mm5, %%mm5\n\t" +" packssdw %%mm1, %%mm1\n\t" +" packssdw %%mm3, %%mm3\n\t" +" packssdw %%mm5, %%mm5\n\t" +" movd %%mm1, %%eax\n\t" +" movd %%mm3, %%ecx\n\t" +" movd %%mm5, %%ebx\n\t" +" movw %%ax, 416(%%esi)\n\t" +" movw %%cx, 288(%%esi)\n\t" +" movw %%bx, 160(%%esi)\n\t" +" shrl $16, %%eax\n\t" +" shrl $16, %%ecx\n\t" +" shrl $16, %%ebx\n\t" +" movw %%ax, 96(%%edi)\n\t" +" movw %%cx, 224(%%edi)\n\t" +" movw %%bx, 352(%%edi)\n\t" + +" movsw\n\t" ".L_bye:\n\t" -" femms\n\t" - : - :"m"(a),"m"(b),"m"(c),"m"(tmp[0]) - :"memory","%eax","%ebx","%ecx","%edx","%esi","%edi"); +" femms\n\t" + : + :"m"(a),"m"(b),"m"(c),"m"(tmp[0]) + :"memory","%eax","%ebx","%ecx","%edx","%esi","%edi"); } diff --git a/mp3lib/dct64_k7.c b/mp3lib/dct64_k7.c index edd41bd7a2..8204877cad 100644 --- a/mp3lib/dct64_k7.c +++ b/mp3lib/dct64_k7.c @@ -19,748 +19,748 @@ void dct64_MMX_3dnowex(short *a,short *b,real *c) { char tmp[256]; __asm__ volatile( -" movl %2,%%eax\n\t" +" movl %2,%%eax\n\t" -" leal 128+%3,%%edx\n\t" -" movl %0,%%esi\n\t" -" movl %1,%%edi\n\t" -" movl $"MANGLE(costab_mmx)",%%ebx\n\t" -" leal %3,%%ecx\n\t" +" leal 128+%3,%%edx\n\t" +" movl %0,%%esi\n\t" +" movl %1,%%edi\n\t" +" movl $"MANGLE(costab_mmx)",%%ebx\n\t" +" leal %3,%%ecx\n\t" /* Phase 1*/ -" movq (%%eax), %%mm0\n\t" -" movq 8(%%eax), %%mm4\n\t" -" movq %%mm0, %%mm3\n\t" -" movq %%mm4, %%mm7\n\t" -" pswapd 120(%%eax), %%mm1\n\t" -" pswapd 112(%%eax), %%mm5\n\t" -" pfadd %%mm1, %%mm0\n\t" -" pfadd %%mm5, %%mm4\n\t" -" movq %%mm0, (%%edx)\n\t" -" movq %%mm4, 8(%%edx)\n\t" -" pfsub %%mm1, %%mm3\n\t" -" pfsub %%mm5, %%mm7\n\t" -" pfmul (%%ebx), %%mm3\n\t" -" pfmul 8(%%ebx), %%mm7\n\t" -" pswapd %%mm3, %%mm3\n\t" -" pswapd %%mm7, %%mm7\n\t" -" movq %%mm3, 120(%%edx)\n\t" -" movq %%mm7, 112(%%edx)\n\t" - -" movq 16(%%eax), %%mm0\n\t" -" movq 24(%%eax), %%mm4\n\t" -" movq %%mm0, %%mm3\n\t" -" movq %%mm4, %%mm7\n\t" -" pswapd 104(%%eax), %%mm1\n\t" -" pswapd 96(%%eax), %%mm5\n\t" -" pfadd %%mm1, %%mm0\n\t" -" pfadd %%mm5, %%mm4\n\t" -" movq %%mm0, 16(%%edx)\n\t" -" movq %%mm4, 24(%%edx)\n\t" -" pfsub %%mm1, %%mm3\n\t" -" pfsub %%mm5, %%mm7\n\t" -" pfmul 16(%%ebx), %%mm3\n\t" -" pfmul 24(%%ebx), %%mm7\n\t" -" pswapd %%mm3, %%mm3\n\t" -" pswapd %%mm7, %%mm7\n\t" -" movq %%mm3, 104(%%edx)\n\t" -" movq %%mm7, 96(%%edx)\n\t" - -" movq 32(%%eax), %%mm0\n\t" -" movq 40(%%eax), %%mm4\n\t" -" movq %%mm0, %%mm3\n\t" -" movq %%mm4, %%mm7\n\t" -" pswapd 88(%%eax), %%mm1\n\t" -" pswapd 80(%%eax), %%mm5\n\t" -" pfadd %%mm1, %%mm0\n\t" -" pfadd %%mm5, %%mm4\n\t" -" movq %%mm0, 32(%%edx)\n\t" -" movq %%mm4, 40(%%edx)\n\t" -" pfsub %%mm1, %%mm3\n\t" -" pfsub %%mm5, %%mm7\n\t" -" pfmul 32(%%ebx), %%mm3\n\t" -" pfmul 40(%%ebx), %%mm7\n\t" -" pswapd %%mm3, %%mm3\n\t" -" pswapd %%mm7, %%mm7\n\t" -" movq %%mm3, 88(%%edx)\n\t" -" movq %%mm7, 80(%%edx)\n\t" - -" movq 48(%%eax), %%mm0\n\t" -" movq 56(%%eax), %%mm4\n\t" -" movq %%mm0, %%mm3\n\t" -" movq %%mm4, %%mm7\n\t" -" pswapd 72(%%eax), %%mm1\n\t" -" pswapd 64(%%eax), %%mm5\n\t" -" pfadd %%mm1, %%mm0\n\t" -" pfadd %%mm5, %%mm4\n\t" -" movq %%mm0, 48(%%edx)\n\t" -" movq %%mm4, 56(%%edx)\n\t" -" pfsub %%mm1, %%mm3\n\t" -" pfsub %%mm5, %%mm7\n\t" -" pfmul 48(%%ebx), %%mm3\n\t" -" pfmul 56(%%ebx), %%mm7\n\t" -" pswapd %%mm3, %%mm3\n\t" -" pswapd %%mm7, %%mm7\n\t" -" movq %%mm3, 72(%%edx)\n\t" -" movq %%mm7, 64(%%edx)\n\t" +" movq (%%eax), %%mm0\n\t" +" movq 8(%%eax), %%mm4\n\t" +" movq %%mm0, %%mm3\n\t" +" movq %%mm4, %%mm7\n\t" +" pswapd 120(%%eax), %%mm1\n\t" +" pswapd 112(%%eax), %%mm5\n\t" +" pfadd %%mm1, %%mm0\n\t" +" pfadd %%mm5, %%mm4\n\t" +" movq %%mm0, (%%edx)\n\t" +" movq %%mm4, 8(%%edx)\n\t" +" pfsub %%mm1, %%mm3\n\t" +" pfsub %%mm5, %%mm7\n\t" +" pfmul (%%ebx), %%mm3\n\t" +" pfmul 8(%%ebx), %%mm7\n\t" +" pswapd %%mm3, %%mm3\n\t" +" pswapd %%mm7, %%mm7\n\t" +" movq %%mm3, 120(%%edx)\n\t" +" movq %%mm7, 112(%%edx)\n\t" + +" movq 16(%%eax), %%mm0\n\t" +" movq 24(%%eax), %%mm4\n\t" +" movq %%mm0, %%mm3\n\t" +" movq %%mm4, %%mm7\n\t" +" pswapd 104(%%eax), %%mm1\n\t" +" pswapd 96(%%eax), %%mm5\n\t" +" pfadd %%mm1, %%mm0\n\t" +" pfadd %%mm5, %%mm4\n\t" +" movq %%mm0, 16(%%edx)\n\t" +" movq %%mm4, 24(%%edx)\n\t" +" pfsub %%mm1, %%mm3\n\t" +" pfsub %%mm5, %%mm7\n\t" +" pfmul 16(%%ebx), %%mm3\n\t" +" pfmul 24(%%ebx), %%mm7\n\t" +" pswapd %%mm3, %%mm3\n\t" +" pswapd %%mm7, %%mm7\n\t" +" movq %%mm3, 104(%%edx)\n\t" +" movq %%mm7, 96(%%edx)\n\t" + +" movq 32(%%eax), %%mm0\n\t" +" movq 40(%%eax), %%mm4\n\t" +" movq %%mm0, %%mm3\n\t" +" movq %%mm4, %%mm7\n\t" +" pswapd 88(%%eax), %%mm1\n\t" +" pswapd 80(%%eax), %%mm5\n\t" +" pfadd %%mm1, %%mm0\n\t" +" pfadd %%mm5, %%mm4\n\t" +" movq %%mm0, 32(%%edx)\n\t" +" movq %%mm4, 40(%%edx)\n\t" +" pfsub %%mm1, %%mm3\n\t" +" pfsub %%mm5, %%mm7\n\t" +" pfmul 32(%%ebx), %%mm3\n\t" +" pfmul 40(%%ebx), %%mm7\n\t" +" pswapd %%mm3, %%mm3\n\t" +" pswapd %%mm7, %%mm7\n\t" +" movq %%mm3, 88(%%edx)\n\t" +" movq %%mm7, 80(%%edx)\n\t" + +" movq 48(%%eax), %%mm0\n\t" +" movq 56(%%eax), %%mm4\n\t" +" movq %%mm0, %%mm3\n\t" +" movq %%mm4, %%mm7\n\t" +" pswapd 72(%%eax), %%mm1\n\t" +" pswapd 64(%%eax), %%mm5\n\t" +" pfadd %%mm1, %%mm0\n\t" +" pfadd %%mm5, %%mm4\n\t" +" movq %%mm0, 48(%%edx)\n\t" +" movq %%mm4, 56(%%edx)\n\t" +" pfsub %%mm1, %%mm3\n\t" +" pfsub %%mm5, %%mm7\n\t" +" pfmul 48(%%ebx), %%mm3\n\t" +" pfmul 56(%%ebx), %%mm7\n\t" +" pswapd %%mm3, %%mm3\n\t" +" pswapd %%mm7, %%mm7\n\t" +" movq %%mm3, 72(%%edx)\n\t" +" movq %%mm7, 64(%%edx)\n\t" /* Phase 2*/ -" movq (%%edx), %%mm0\n\t" -" movq 8(%%edx), %%mm4\n\t" -" movq %%mm0, %%mm3\n\t" -" movq %%mm4, %%mm7\n\t" -" pswapd 56(%%edx), %%mm1\n\t" -" pswapd 48(%%edx), %%mm5\n\t" -" pfadd %%mm1, %%mm0\n\t" -" pfadd %%mm5, %%mm4\n\t" -" movq %%mm0, (%%ecx)\n\t" -" movq %%mm4, 8(%%ecx)\n\t" -" pfsub %%mm1, %%mm3\n\t" -" pfsub %%mm5, %%mm7\n\t" -" pfmul 64(%%ebx), %%mm3\n\t" -" pfmul 72(%%ebx), %%mm7\n\t" -" pswapd %%mm3, %%mm3\n\t" -" pswapd %%mm7, %%mm7\n\t" -" movq %%mm3, 56(%%ecx)\n\t" -" movq %%mm7, 48(%%ecx)\n\t" - -" movq 16(%%edx), %%mm0\n\t" -" movq 24(%%edx), %%mm4\n\t" -" movq %%mm0, %%mm3\n\t" -" movq %%mm4, %%mm7\n\t" -" pswapd 40(%%edx), %%mm1\n\t" -" pswapd 32(%%edx), %%mm5\n\t" -" pfadd %%mm1, %%mm0\n\t" -" pfadd %%mm5, %%mm4\n\t" -" movq %%mm0, 16(%%ecx)\n\t" -" movq %%mm4, 24(%%ecx)\n\t" -" pfsub %%mm1, %%mm3\n\t" -" pfsub %%mm5, %%mm7\n\t" -" pfmul 80(%%ebx), %%mm3\n\t" -" pfmul 88(%%ebx), %%mm7\n\t" -" pswapd %%mm3, %%mm3\n\t" -" pswapd %%mm7, %%mm7\n\t" -" movq %%mm3, 40(%%ecx)\n\t" -" movq %%mm7, 32(%%ecx)\n\t" +" movq (%%edx), %%mm0\n\t" +" movq 8(%%edx), %%mm4\n\t" +" movq %%mm0, %%mm3\n\t" +" movq %%mm4, %%mm7\n\t" +" pswapd 56(%%edx), %%mm1\n\t" +" pswapd 48(%%edx), %%mm5\n\t" +" pfadd %%mm1, %%mm0\n\t" +" pfadd %%mm5, %%mm4\n\t" +" movq %%mm0, (%%ecx)\n\t" +" movq %%mm4, 8(%%ecx)\n\t" +" pfsub %%mm1, %%mm3\n\t" +" pfsub %%mm5, %%mm7\n\t" +" pfmul 64(%%ebx), %%mm3\n\t" +" pfmul 72(%%ebx), %%mm7\n\t" +" pswapd %%mm3, %%mm3\n\t" +" pswapd %%mm7, %%mm7\n\t" +" movq %%mm3, 56(%%ecx)\n\t" +" movq %%mm7, 48(%%ecx)\n\t" + +" movq 16(%%edx), %%mm0\n\t" +" movq 24(%%edx), %%mm4\n\t" +" movq %%mm0, %%mm3\n\t" +" movq %%mm4, %%mm7\n\t" +" pswapd 40(%%edx), %%mm1\n\t" +" pswapd 32(%%edx), %%mm5\n\t" +" pfadd %%mm1, %%mm0\n\t" +" pfadd %%mm5, %%mm4\n\t" +" movq %%mm0, 16(%%ecx)\n\t" +" movq %%mm4, 24(%%ecx)\n\t" +" pfsub %%mm1, %%mm3\n\t" +" pfsub %%mm5, %%mm7\n\t" +" pfmul 80(%%ebx), %%mm3\n\t" +" pfmul 88(%%ebx), %%mm7\n\t" +" pswapd %%mm3, %%mm3\n\t" +" pswapd %%mm7, %%mm7\n\t" +" movq %%mm3, 40(%%ecx)\n\t" +" movq %%mm7, 32(%%ecx)\n\t" /* Phase 3*/ -" movq 64(%%edx), %%mm0\n\t" -" movq 72(%%edx), %%mm4\n\t" -" movq %%mm0, %%mm3\n\t" -" movq %%mm4, %%mm7\n\t" -" pswapd 120(%%edx), %%mm1\n\t" -" pswapd 112(%%edx), %%mm5\n\t" -" pfadd %%mm1, %%mm0\n\t" -" pfadd %%mm5, %%mm4\n\t" -" movq %%mm0, 64(%%ecx)\n\t" -" movq %%mm4, 72(%%ecx)\n\t" -" pfsubr %%mm1, %%mm3\n\t" -" pfsubr %%mm5, %%mm7\n\t" -" pfmul 64(%%ebx), %%mm3\n\t" -" pfmul 72(%%ebx), %%mm7\n\t" -" pswapd %%mm3, %%mm3\n\t" -" pswapd %%mm7, %%mm7\n\t" -" movq %%mm3, 120(%%ecx)\n\t" -" movq %%mm7, 112(%%ecx)\n\t" - -" movq 80(%%edx), %%mm0\n\t" -" movq 88(%%edx), %%mm4\n\t" -" movq %%mm0, %%mm3\n\t" -" movq %%mm4, %%mm7\n\t" -" pswapd 104(%%edx), %%mm1\n\t" -" pswapd 96(%%edx), %%mm5\n\t" -" pfadd %%mm1, %%mm0\n\t" -" pfadd %%mm5, %%mm4\n\t" -" movq %%mm0, 80(%%ecx)\n\t" -" movq %%mm4, 88(%%ecx)\n\t" -" pfsubr %%mm1, %%mm3\n\t" -" pfsubr %%mm5, %%mm7\n\t" -" pfmul 80(%%ebx), %%mm3\n\t" -" pfmul 88(%%ebx), %%mm7\n\t" -" pswapd %%mm3, %%mm3\n\t" -" pswapd %%mm7, %%mm7\n\t" -" movq %%mm3, 104(%%ecx)\n\t" -" movq %%mm7, 96(%%ecx)\n\t" +" movq 64(%%edx), %%mm0\n\t" +" movq 72(%%edx), %%mm4\n\t" +" movq %%mm0, %%mm3\n\t" +" movq %%mm4, %%mm7\n\t" +" pswapd 120(%%edx), %%mm1\n\t" +" pswapd 112(%%edx), %%mm5\n\t" +" pfadd %%mm1, %%mm0\n\t" +" pfadd %%mm5, %%mm4\n\t" +" movq %%mm0, 64(%%ecx)\n\t" +" movq %%mm4, 72(%%ecx)\n\t" +" pfsubr %%mm1, %%mm3\n\t" +" pfsubr %%mm5, %%mm7\n\t" +" pfmul 64(%%ebx), %%mm3\n\t" +" pfmul 72(%%ebx), %%mm7\n\t" +" pswapd %%mm3, %%mm3\n\t" +" pswapd %%mm7, %%mm7\n\t" +" movq %%mm3, 120(%%ecx)\n\t" +" movq %%mm7, 112(%%ecx)\n\t" + +" movq 80(%%edx), %%mm0\n\t" +" movq 88(%%edx), %%mm4\n\t" +" movq %%mm0, %%mm3\n\t" +" movq %%mm4, %%mm7\n\t" +" pswapd 104(%%edx), %%mm1\n\t" +" pswapd 96(%%edx), %%mm5\n\t" +" pfadd %%mm1, %%mm0\n\t" +" pfadd %%mm5, %%mm4\n\t" +" movq %%mm0, 80(%%ecx)\n\t" +" movq %%mm4, 88(%%ecx)\n\t" +" pfsubr %%mm1, %%mm3\n\t" +" pfsubr %%mm5, %%mm7\n\t" +" pfmul 80(%%ebx), %%mm3\n\t" +" pfmul 88(%%ebx), %%mm7\n\t" +" pswapd %%mm3, %%mm3\n\t" +" pswapd %%mm7, %%mm7\n\t" +" movq %%mm3, 104(%%ecx)\n\t" +" movq %%mm7, 96(%%ecx)\n\t" /* Phase 4*/ -" movq 96(%%ebx), %%mm2\n\t" -" movq 104(%%ebx), %%mm6\n\t" - -" movq (%%ecx), %%mm0\n\t" -" movq 8(%%ecx), %%mm4\n\t" -" movq %%mm0, %%mm3\n\t" -" movq %%mm4, %%mm7\n\t" -" pswapd 24(%%ecx), %%mm1\n\t" -" pswapd 16(%%ecx), %%mm5\n\t" -" pfadd %%mm1, %%mm0\n\t" -" pfadd %%mm5, %%mm4\n\t" -" movq %%mm0, (%%edx)\n\t" -" movq %%mm4, 8(%%edx)\n\t" -" pfsub %%mm1, %%mm3\n\t" -" pfsub %%mm5, %%mm7\n\t" -" pfmul %%mm2, %%mm3\n\t" -" pfmul %%mm6, %%mm7\n\t" -" pswapd %%mm3, %%mm3\n\t" -" pswapd %%mm7, %%mm7\n\t" -" movq %%mm3, 24(%%edx)\n\t" -" movq %%mm7, 16(%%edx)\n\t" - -" movq 32(%%ecx), %%mm0\n\t" -" movq 40(%%ecx), %%mm4\n\t" -" movq %%mm0, %%mm3\n\t" -" movq %%mm4, %%mm7\n\t" -" pswapd 56(%%ecx), %%mm1\n\t" -" pswapd 48(%%ecx), %%mm5\n\t" -" pfadd %%mm1, %%mm0\n\t" -" pfadd %%mm5, %%mm4\n\t" -" movq %%mm0, 32(%%edx)\n\t" -" movq %%mm4, 40(%%edx)\n\t" -" pfsubr %%mm1, %%mm3\n\t" -" pfsubr %%mm5, %%mm7\n\t" -" pfmul %%mm2, %%mm3\n\t" -" pfmul %%mm6, %%mm7\n\t" -" pswapd %%mm3, %%mm3\n\t" -" pswapd %%mm7, %%mm7\n\t" -" movq %%mm3, 56(%%edx)\n\t" -" movq %%mm7, 48(%%edx)\n\t" - -" movq 64(%%ecx), %%mm0\n\t" -" movq 72(%%ecx), %%mm4\n\t" -" movq %%mm0, %%mm3\n\t" -" movq %%mm4, %%mm7\n\t" -" pswapd 88(%%ecx), %%mm1\n\t" -" pswapd 80(%%ecx), %%mm5\n\t" -" pfadd %%mm1, %%mm0\n\t" -" pfadd %%mm5, %%mm4\n\t" -" movq %%mm0, 64(%%edx)\n\t" -" movq %%mm4, 72(%%edx)\n\t" -" pfsub %%mm1, %%mm3\n\t" -" pfsub %%mm5, %%mm7\n\t" -" pfmul %%mm2, %%mm3\n\t" -" pfmul %%mm6, %%mm7\n\t" -" pswapd %%mm3, %%mm3\n\t" -" pswapd %%mm7, %%mm7\n\t" -" movq %%mm3, 88(%%edx)\n\t" -" movq %%mm7, 80(%%edx)\n\t" - -" movq 96(%%ecx), %%mm0\n\t" -" movq 104(%%ecx), %%mm4\n\t" -" movq %%mm0, %%mm3\n\t" -" movq %%mm4, %%mm7\n\t" -" pswapd 120(%%ecx), %%mm1\n\t" -" pswapd 112(%%ecx), %%mm5\n\t" -" pfadd %%mm1, %%mm0\n\t" -" pfadd %%mm5, %%mm4\n\t" -" movq %%mm0, 96(%%edx)\n\t" -" movq %%mm4, 104(%%edx)\n\t" -" pfsubr %%mm1, %%mm3\n\t" -" pfsubr %%mm5, %%mm7\n\t" -" pfmul %%mm2, %%mm3\n\t" -" pfmul %%mm6, %%mm7\n\t" -" pswapd %%mm3, %%mm3\n\t" -" pswapd %%mm7, %%mm7\n\t" -" movq %%mm3, 120(%%edx)\n\t" -" movq %%mm7, 112(%%edx)\n\t" +" movq 96(%%ebx), %%mm2\n\t" +" movq 104(%%ebx), %%mm6\n\t" + +" movq (%%ecx), %%mm0\n\t" +" movq 8(%%ecx), %%mm4\n\t" +" movq %%mm0, %%mm3\n\t" +" movq %%mm4, %%mm7\n\t" +" pswapd 24(%%ecx), %%mm1\n\t" +" pswapd 16(%%ecx), %%mm5\n\t" +" pfadd %%mm1, %%mm0\n\t" +" pfadd %%mm5, %%mm4\n\t" +" movq %%mm0, (%%edx)\n\t" +" movq %%mm4, 8(%%edx)\n\t" +" pfsub %%mm1, %%mm3\n\t" +" pfsub %%mm5, %%mm7\n\t" +" pfmul %%mm2, %%mm3\n\t" +" pfmul %%mm6, %%mm7\n\t" +" pswapd %%mm3, %%mm3\n\t" +" pswapd %%mm7, %%mm7\n\t" +" movq %%mm3, 24(%%edx)\n\t" +" movq %%mm7, 16(%%edx)\n\t" + +" movq 32(%%ecx), %%mm0\n\t" +" movq 40(%%ecx), %%mm4\n\t" +" movq %%mm0, %%mm3\n\t" +" movq %%mm4, %%mm7\n\t" +" pswapd 56(%%ecx), %%mm1\n\t" +" pswapd 48(%%ecx), %%mm5\n\t" +" pfadd %%mm1, %%mm0\n\t" +" pfadd %%mm5, %%mm4\n\t" +" movq %%mm0, 32(%%edx)\n\t" +" movq %%mm4, 40(%%edx)\n\t" +" pfsubr %%mm1, %%mm3\n\t" +" pfsubr %%mm5, %%mm7\n\t" +" pfmul %%mm2, %%mm3\n\t" +" pfmul %%mm6, %%mm7\n\t" +" pswapd %%mm3, %%mm3\n\t" +" pswapd %%mm7, %%mm7\n\t" +" movq %%mm3, 56(%%edx)\n\t" +" movq %%mm7, 48(%%edx)\n\t" + +" movq 64(%%ecx), %%mm0\n\t" +" movq 72(%%ecx), %%mm4\n\t" +" movq %%mm0, %%mm3\n\t" +" movq %%mm4, %%mm7\n\t" +" pswapd 88(%%ecx), %%mm1\n\t" +" pswapd 80(%%ecx), %%mm5\n\t" +" pfadd %%mm1, %%mm0\n\t" +" pfadd %%mm5, %%mm4\n\t" +" movq %%mm0, 64(%%edx)\n\t" +" movq %%mm4, 72(%%edx)\n\t" +" pfsub %%mm1, %%mm3\n\t" +" pfsub %%mm5, %%mm7\n\t" +" pfmul %%mm2, %%mm3\n\t" +" pfmul %%mm6, %%mm7\n\t" +" pswapd %%mm3, %%mm3\n\t" +" pswapd %%mm7, %%mm7\n\t" +" movq %%mm3, 88(%%edx)\n\t" +" movq %%mm7, 80(%%edx)\n\t" + +" movq 96(%%ecx), %%mm0\n\t" +" movq 104(%%ecx), %%mm4\n\t" +" movq %%mm0, %%mm3\n\t" +" movq %%mm4, %%mm7\n\t" +" pswapd 120(%%ecx), %%mm1\n\t" +" pswapd 112(%%ecx), %%mm5\n\t" +" pfadd %%mm1, %%mm0\n\t" +" pfadd %%mm5, %%mm4\n\t" +" movq %%mm0, 96(%%edx)\n\t" +" movq %%mm4, 104(%%edx)\n\t" +" pfsubr %%mm1, %%mm3\n\t" +" pfsubr %%mm5, %%mm7\n\t" +" pfmul %%mm2, %%mm3\n\t" +" pfmul %%mm6, %%mm7\n\t" +" pswapd %%mm3, %%mm3\n\t" +" pswapd %%mm7, %%mm7\n\t" +" movq %%mm3, 120(%%edx)\n\t" +" movq %%mm7, 112(%%edx)\n\t" /* Phase 5 */ -" movq 112(%%ebx), %%mm2\n\t" - -" movq (%%edx), %%mm0\n\t" -" movq 16(%%edx), %%mm4\n\t" -" movq %%mm0, %%mm3\n\t" -" movq %%mm4, %%mm7\n\t" -" pswapd 8(%%edx), %%mm1\n\t" -" pswapd 24(%%edx), %%mm5\n\t" -" pfadd %%mm1, %%mm0\n\t" -" pfadd %%mm5, %%mm4\n\t" -" movq %%mm0, (%%ecx)\n\t" -" movq %%mm4, 16(%%ecx)\n\t" -" pfsub %%mm1, %%mm3\n\t" -" pfsubr %%mm5, %%mm7\n\t" -" pfmul %%mm2, %%mm3\n\t" -" pfmul %%mm2, %%mm7\n\t" -" pswapd %%mm3, %%mm3\n\t" -" pswapd %%mm7, %%mm7\n\t" -" movq %%mm3, 8(%%ecx)\n\t" -" movq %%mm7, 24(%%ecx)\n\t" - -" movq 32(%%edx), %%mm0\n\t" -" movq 48(%%edx), %%mm4\n\t" -" movq %%mm0, %%mm3\n\t" -" movq %%mm4, %%mm7\n\t" -" pswapd 40(%%edx), %%mm1\n\t" -" pswapd 56(%%edx), %%mm5\n\t" -" pfadd %%mm1, %%mm0\n\t" -" pfadd %%mm5, %%mm4\n\t" -" movq %%mm0, 32(%%ecx)\n\t" -" movq %%mm4, 48(%%ecx)\n\t" -" pfsub %%mm1, %%mm3\n\t" -" pfsubr %%mm5, %%mm7\n\t" -" pfmul %%mm2, %%mm3\n\t" -" pfmul %%mm2, %%mm7\n\t" -" pswapd %%mm3, %%mm3\n\t" -" pswapd %%mm7, %%mm7\n\t" -" movq %%mm3, 40(%%ecx)\n\t" -" movq %%mm7, 56(%%ecx)\n\t" - -" movq 64(%%edx), %%mm0\n\t" -" movq 80(%%edx), %%mm4\n\t" -" movq %%mm0, %%mm3\n\t" -" movq %%mm4, %%mm7\n\t" -" pswapd 72(%%edx), %%mm1\n\t" -" pswapd 88(%%edx), %%mm5\n\t" -" pfadd %%mm1, %%mm0\n\t" -" pfadd %%mm5, %%mm4\n\t" -" movq %%mm0, 64(%%ecx)\n\t" -" movq %%mm4, 80(%%ecx)\n\t" -" pfsub %%mm1, %%mm3\n\t" -" pfsubr %%mm5, %%mm7\n\t" -" pfmul %%mm2, %%mm3\n\t" -" pfmul %%mm2, %%mm7\n\t" -" pswapd %%mm3, %%mm3\n\t" -" pswapd %%mm7, %%mm7\n\t" -" movq %%mm3, 72(%%ecx)\n\t" -" movq %%mm7, 88(%%ecx)\n\t" - -" movq 96(%%edx), %%mm0\n\t" -" movq 112(%%edx), %%mm4\n\t" -" movq %%mm0, %%mm3\n\t" -" movq %%mm4, %%mm7\n\t" -" pswapd 104(%%edx), %%mm1\n\t" -" pswapd 120(%%edx), %%mm5\n\t" -" pfadd %%mm1, %%mm0\n\t" -" pfadd %%mm5, %%mm4\n\t" -" movq %%mm0, 96(%%ecx)\n\t" -" movq %%mm4, 112(%%ecx)\n\t" -" pfsub %%mm1, %%mm3\n\t" -" pfsubr %%mm5, %%mm7\n\t" -" pfmul %%mm2, %%mm3\n\t" -" pfmul %%mm2, %%mm7\n\t" -" pswapd %%mm3, %%mm3\n\t" -" pswapd %%mm7, %%mm7\n\t" -" movq %%mm3, 104(%%ecx)\n\t" -" movq %%mm7, 120(%%ecx)\n\t" +" movq 112(%%ebx), %%mm2\n\t" + +" movq (%%edx), %%mm0\n\t" +" movq 16(%%edx), %%mm4\n\t" +" movq %%mm0, %%mm3\n\t" +" movq %%mm4, %%mm7\n\t" +" pswapd 8(%%edx), %%mm1\n\t" +" pswapd 24(%%edx), %%mm5\n\t" +" pfadd %%mm1, %%mm0\n\t" +" pfadd %%mm5, %%mm4\n\t" +" movq %%mm0, (%%ecx)\n\t" +" movq %%mm4, 16(%%ecx)\n\t" +" pfsub %%mm1, %%mm3\n\t" +" pfsubr %%mm5, %%mm7\n\t" +" pfmul %%mm2, %%mm3\n\t" +" pfmul %%mm2, %%mm7\n\t" +" pswapd %%mm3, %%mm3\n\t" +" pswapd %%mm7, %%mm7\n\t" +" movq %%mm3, 8(%%ecx)\n\t" +" movq %%mm7, 24(%%ecx)\n\t" + +" movq 32(%%edx), %%mm0\n\t" +" movq 48(%%edx), %%mm4\n\t" +" movq %%mm0, %%mm3\n\t" +" movq %%mm4, %%mm7\n\t" +" pswapd 40(%%edx), %%mm1\n\t" +" pswapd 56(%%edx), %%mm5\n\t" +" pfadd %%mm1, %%mm0\n\t" +" pfadd %%mm5, %%mm4\n\t" +" movq %%mm0, 32(%%ecx)\n\t" +" movq %%mm4, 48(%%ecx)\n\t" +" pfsub %%mm1, %%mm3\n\t" +" pfsubr %%mm5, %%mm7\n\t" +" pfmul %%mm2, %%mm3\n\t" +" pfmul %%mm2, %%mm7\n\t" +" pswapd %%mm3, %%mm3\n\t" +" pswapd %%mm7, %%mm7\n\t" +" movq %%mm3, 40(%%ecx)\n\t" +" movq %%mm7, 56(%%ecx)\n\t" + +" movq 64(%%edx), %%mm0\n\t" +" movq 80(%%edx), %%mm4\n\t" +" movq %%mm0, %%mm3\n\t" +" movq %%mm4, %%mm7\n\t" +" pswapd 72(%%edx), %%mm1\n\t" +" pswapd 88(%%edx), %%mm5\n\t" +" pfadd %%mm1, %%mm0\n\t" +" pfadd %%mm5, %%mm4\n\t" +" movq %%mm0, 64(%%ecx)\n\t" +" movq %%mm4, 80(%%ecx)\n\t" +" pfsub %%mm1, %%mm3\n\t" +" pfsubr %%mm5, %%mm7\n\t" +" pfmul %%mm2, %%mm3\n\t" +" pfmul %%mm2, %%mm7\n\t" +" pswapd %%mm3, %%mm3\n\t" +" pswapd %%mm7, %%mm7\n\t" +" movq %%mm3, 72(%%ecx)\n\t" +" movq %%mm7, 88(%%ecx)\n\t" + +" movq 96(%%edx), %%mm0\n\t" +" movq 112(%%edx), %%mm4\n\t" +" movq %%mm0, %%mm3\n\t" +" movq %%mm4, %%mm7\n\t" +" pswapd 104(%%edx), %%mm1\n\t" +" pswapd 120(%%edx), %%mm5\n\t" +" pfadd %%mm1, %%mm0\n\t" +" pfadd %%mm5, %%mm4\n\t" +" movq %%mm0, 96(%%ecx)\n\t" +" movq %%mm4, 112(%%ecx)\n\t" +" pfsub %%mm1, %%mm3\n\t" +" pfsubr %%mm5, %%mm7\n\t" +" pfmul %%mm2, %%mm3\n\t" +" pfmul %%mm2, %%mm7\n\t" +" pswapd %%mm3, %%mm3\n\t" +" pswapd %%mm7, %%mm7\n\t" +" movq %%mm3, 104(%%ecx)\n\t" +" movq %%mm7, 120(%%ecx)\n\t" /* Phase 6. This is the end of easy road. */ /* Code below is coded in scalar mode. Should be optimized */ -" movd "MANGLE(plus_1f)", %%mm6\n\t" -" punpckldq 120(%%ebx), %%mm6\n\t" /* mm6 = 1.0 | 120(%%ebx)*/ -" movq "MANGLE(x_plus_minus_3dnow)", %%mm7\n\t" /* mm7 = +1 | -1 */ - -" movq 32(%%ecx), %%mm0\n\t" -" movq 64(%%ecx), %%mm2\n\t" -" movq %%mm0, %%mm1\n\t" -" movq %%mm2, %%mm3\n\t" -" pxor %%mm7, %%mm1\n\t" -" pxor %%mm7, %%mm3\n\t" -" pfacc %%mm1, %%mm0\n\t" -" pfacc %%mm3, %%mm2\n\t" -" pfmul %%mm6, %%mm0\n\t" -" pfmul %%mm6, %%mm2\n\t" -" movq %%mm0, 32(%%edx)\n\t" -" movq %%mm2, 64(%%edx)\n\t" - -" movd 44(%%ecx), %%mm0\n\t" -" movd 40(%%ecx), %%mm2\n\t" -" movd 120(%%ebx), %%mm3\n\t" -" punpckldq 76(%%ecx), %%mm0\n\t" -" punpckldq 72(%%ecx), %%mm2\n\t" -" punpckldq %%mm3, %%mm3\n\t" -" movq %%mm0, %%mm4\n\t" -" movq %%mm2, %%mm5\n\t" -" pfsub %%mm2, %%mm0\n\t" -" pfmul %%mm3, %%mm0\n\t" -" movq %%mm0, %%mm1\n\t" -" pfadd %%mm5, %%mm0\n\t" -" pfadd %%mm4, %%mm0\n\t" -" movq %%mm0, %%mm2\n\t" -" punpckldq %%mm1, %%mm0\n\t" -" punpckhdq %%mm1, %%mm2\n\t" -" movq %%mm0, 40(%%edx)\n\t" -" movq %%mm2, 72(%%edx)\n\t" - -" movd 48(%%ecx), %%mm3\n\t" -" movd 60(%%ecx), %%mm2\n\t" -" pfsub 52(%%ecx), %%mm3\n\t" -" pfsub 56(%%ecx), %%mm2\n\t" -" pfmul 120(%%ebx), %%mm3\n\t" -" pfmul 120(%%ebx), %%mm2\n\t" -" movq %%mm2, %%mm1\n\t" - -" pfadd 56(%%ecx), %%mm1\n\t" -" pfadd 60(%%ecx), %%mm1\n\t" -" movq %%mm1, %%mm0\n\t" - -" pfadd 48(%%ecx), %%mm0\n\t" -" pfadd 52(%%ecx), %%mm0\n\t" -" pfadd %%mm3, %%mm1\n\t" -" punpckldq %%mm2, %%mm1\n\t" -" pfadd %%mm3, %%mm2\n\t" -" punpckldq %%mm2, %%mm0\n\t" -" movq %%mm1, 56(%%edx)\n\t" -" movq %%mm0, 48(%%edx)\n\t" +" movd "MANGLE(plus_1f)", %%mm6\n\t" +" punpckldq 120(%%ebx), %%mm6\n\t" /* mm6 = 1.0 | 120(%%ebx)*/ +" movq "MANGLE(x_plus_minus_3dnow)", %%mm7\n\t" /* mm7 = +1 | -1 */ + +" movq 32(%%ecx), %%mm0\n\t" +" movq 64(%%ecx), %%mm2\n\t" +" movq %%mm0, %%mm1\n\t" +" movq %%mm2, %%mm3\n\t" +" pxor %%mm7, %%mm1\n\t" +" pxor %%mm7, %%mm3\n\t" +" pfacc %%mm1, %%mm0\n\t" +" pfacc %%mm3, %%mm2\n\t" +" pfmul %%mm6, %%mm0\n\t" +" pfmul %%mm6, %%mm2\n\t" +" movq %%mm0, 32(%%edx)\n\t" +" movq %%mm2, 64(%%edx)\n\t" + +" movd 44(%%ecx), %%mm0\n\t" +" movd 40(%%ecx), %%mm2\n\t" +" movd 120(%%ebx), %%mm3\n\t" +" punpckldq 76(%%ecx), %%mm0\n\t" +" punpckldq 72(%%ecx), %%mm2\n\t" +" punpckldq %%mm3, %%mm3\n\t" +" movq %%mm0, %%mm4\n\t" +" movq %%mm2, %%mm5\n\t" +" pfsub %%mm2, %%mm0\n\t" +" pfmul %%mm3, %%mm0\n\t" +" movq %%mm0, %%mm1\n\t" +" pfadd %%mm5, %%mm0\n\t" +" pfadd %%mm4, %%mm0\n\t" +" movq %%mm0, %%mm2\n\t" +" punpckldq %%mm1, %%mm0\n\t" +" punpckhdq %%mm1, %%mm2\n\t" +" movq %%mm0, 40(%%edx)\n\t" +" movq %%mm2, 72(%%edx)\n\t" + +" movd 48(%%ecx), %%mm3\n\t" +" movd 60(%%ecx), %%mm2\n\t" +" pfsub 52(%%ecx), %%mm3\n\t" +" pfsub 56(%%ecx), %%mm2\n\t" +" pfmul 120(%%ebx), %%mm3\n\t" +" pfmul 120(%%ebx), %%mm2\n\t" +" movq %%mm2, %%mm1\n\t" + +" pfadd 56(%%ecx), %%mm1\n\t" +" pfadd 60(%%ecx), %%mm1\n\t" +" movq %%mm1, %%mm0\n\t" + +" pfadd 48(%%ecx), %%mm0\n\t" +" pfadd 52(%%ecx), %%mm0\n\t" +" pfadd %%mm3, %%mm1\n\t" +" punpckldq %%mm2, %%mm1\n\t" +" pfadd %%mm3, %%mm2\n\t" +" punpckldq %%mm2, %%mm0\n\t" +" movq %%mm1, 56(%%edx)\n\t" +" movq %%mm0, 48(%%edx)\n\t" /*---*/ -" movd 92(%%ecx), %%mm1\n\t" -" pfsub 88(%%ecx), %%mm1\n\t" -" pfmul 120(%%ebx), %%mm1\n\t" -" movd %%mm1, 92(%%edx)\n\t" -" pfadd 92(%%ecx), %%mm1\n\t" -" pfadd 88(%%ecx), %%mm1\n\t" -" movq %%mm1, %%mm0\n\t" - -" pfadd 80(%%ecx), %%mm0\n\t" -" pfadd 84(%%ecx), %%mm0\n\t" -" movd %%mm0, 80(%%edx)\n\t" - -" movd 80(%%ecx), %%mm0\n\t" -" pfsub 84(%%ecx), %%mm0\n\t" -" pfmul 120(%%ebx), %%mm0\n\t" -" pfadd %%mm0, %%mm1\n\t" -" pfadd 92(%%edx), %%mm0\n\t" -" punpckldq %%mm1, %%mm0\n\t" -" movq %%mm0, 84(%%edx)\n\t" - -" movq 96(%%ecx), %%mm0\n\t" -" movq %%mm0, %%mm1\n\t" -" pxor %%mm7, %%mm1\n\t" -" pfacc %%mm1, %%mm0\n\t" -" pfmul %%mm6, %%mm0\n\t" -" movq %%mm0, 96(%%edx)\n\t" - -" movd 108(%%ecx), %%mm0\n\t" -" pfsub 104(%%ecx), %%mm0\n\t" -" pfmul 120(%%ebx), %%mm0\n\t" -" movd %%mm0, 108(%%edx)\n\t" -" pfadd 104(%%ecx), %%mm0\n\t" -" pfadd 108(%%ecx), %%mm0\n\t" -" movd %%mm0, 104(%%edx)\n\t" - -" movd 124(%%ecx), %%mm1\n\t" -" pfsub 120(%%ecx), %%mm1\n\t" -" pfmul 120(%%ebx), %%mm1\n\t" -" movd %%mm1, 124(%%edx)\n\t" -" pfadd 120(%%ecx), %%mm1\n\t" -" pfadd 124(%%ecx), %%mm1\n\t" -" movq %%mm1, %%mm0\n\t" - -" pfadd 112(%%ecx), %%mm0\n\t" -" pfadd 116(%%ecx), %%mm0\n\t" -" movd %%mm0, 112(%%edx)\n\t" - -" movd 112(%%ecx), %%mm0\n\t" -" pfsub 116(%%ecx), %%mm0\n\t" -" pfmul 120(%%ebx), %%mm0\n\t" -" pfadd %%mm0,%%mm1\n\t" -" pfadd 124(%%edx), %%mm0\n\t" -" punpckldq %%mm1, %%mm0\n\t" -" movq %%mm0, 116(%%edx)\n\t" +" movd 92(%%ecx), %%mm1\n\t" +" pfsub 88(%%ecx), %%mm1\n\t" +" pfmul 120(%%ebx), %%mm1\n\t" +" movd %%mm1, 92(%%edx)\n\t" +" pfadd 92(%%ecx), %%mm1\n\t" +" pfadd 88(%%ecx), %%mm1\n\t" +" movq %%mm1, %%mm0\n\t" + +" pfadd 80(%%ecx), %%mm0\n\t" +" pfadd 84(%%ecx), %%mm0\n\t" +" movd %%mm0, 80(%%edx)\n\t" + +" movd 80(%%ecx), %%mm0\n\t" +" pfsub 84(%%ecx), %%mm0\n\t" +" pfmul 120(%%ebx), %%mm0\n\t" +" pfadd %%mm0, %%mm1\n\t" +" pfadd 92(%%edx), %%mm0\n\t" +" punpckldq %%mm1, %%mm0\n\t" +" movq %%mm0, 84(%%edx)\n\t" + +" movq 96(%%ecx), %%mm0\n\t" +" movq %%mm0, %%mm1\n\t" +" pxor %%mm7, %%mm1\n\t" +" pfacc %%mm1, %%mm0\n\t" +" pfmul %%mm6, %%mm0\n\t" +" movq %%mm0, 96(%%edx)\n\t" + +" movd 108(%%ecx), %%mm0\n\t" +" pfsub 104(%%ecx), %%mm0\n\t" +" pfmul 120(%%ebx), %%mm0\n\t" +" movd %%mm0, 108(%%edx)\n\t" +" pfadd 104(%%ecx), %%mm0\n\t" +" pfadd 108(%%ecx), %%mm0\n\t" +" movd %%mm0, 104(%%edx)\n\t" + +" movd 124(%%ecx), %%mm1\n\t" +" pfsub 120(%%ecx), %%mm1\n\t" +" pfmul 120(%%ebx), %%mm1\n\t" +" movd %%mm1, 124(%%edx)\n\t" +" pfadd 120(%%ecx), %%mm1\n\t" +" pfadd 124(%%ecx), %%mm1\n\t" +" movq %%mm1, %%mm0\n\t" + +" pfadd 112(%%ecx), %%mm0\n\t" +" pfadd 116(%%ecx), %%mm0\n\t" +" movd %%mm0, 112(%%edx)\n\t" + +" movd 112(%%ecx), %%mm0\n\t" +" pfsub 116(%%ecx), %%mm0\n\t" +" pfmul 120(%%ebx), %%mm0\n\t" +" pfadd %%mm0,%%mm1\n\t" +" pfadd 124(%%edx), %%mm0\n\t" +" punpckldq %%mm1, %%mm0\n\t" +" movq %%mm0, 116(%%edx)\n\t" // this code is broken, there is nothing modifying the z flag above. #if 0 -" jnz .L01\n\t" +" jnz .L01\n\t" /* Phase 7*/ /* Code below is coded in scalar mode. Should be optimized */ -" movd (%%ecx), %%mm0\n\t" -" pfadd 4(%%ecx), %%mm0\n\t" -" movd %%mm0, 1024(%%esi)\n\t" - -" movd (%%ecx), %%mm0\n\t" -" pfsub 4(%%ecx), %%mm0\n\t" -" pfmul 120(%%ebx), %%mm0\n\t" -" movd %%mm0, (%%esi)\n\t" -" movd %%mm0, (%%edi)\n\t" - -" movd 12(%%ecx), %%mm0\n\t" -" pfsub 8(%%ecx), %%mm0\n\t" -" pfmul 120(%%ebx), %%mm0\n\t" -" movd %%mm0, 512(%%edi)\n\t" -" pfadd 12(%%ecx), %%mm0\n\t" -" pfadd 8(%%ecx), %%mm0\n\t" -" movd %%mm0, 512(%%esi)\n\t" - -" movd 16(%%ecx), %%mm0\n\t" -" pfsub 20(%%ecx), %%mm0\n\t" -" pfmul 120(%%ebx), %%mm0\n\t" -" movq %%mm0, %%mm3\n\t" - -" movd 28(%%ecx), %%mm0\n\t" -" pfsub 24(%%ecx), %%mm0\n\t" -" pfmul 120(%%ebx), %%mm0\n\t" -" movd %%mm0, 768(%%edi)\n\t" -" movq %%mm0, %%mm2\n\t" - -" pfadd 24(%%ecx), %%mm0\n\t" -" pfadd 28(%%ecx), %%mm0\n\t" -" movq %%mm0, %%mm1\n\t" - -" pfadd 16(%%ecx), %%mm0\n\t" -" pfadd 20(%%ecx), %%mm0\n\t" -" movd %%mm0, 768(%%esi)\n\t" -" pfadd %%mm3, %%mm1\n\t" -" movd %%mm1, 256(%%esi)\n\t" -" pfadd %%mm3, %%mm2\n\t" -" movd %%mm2, 256(%%edi)\n\t" +" movd (%%ecx), %%mm0\n\t" +" pfadd 4(%%ecx), %%mm0\n\t" +" movd %%mm0, 1024(%%esi)\n\t" + +" movd (%%ecx), %%mm0\n\t" +" pfsub 4(%%ecx), %%mm0\n\t" +" pfmul 120(%%ebx), %%mm0\n\t" +" movd %%mm0, (%%esi)\n\t" +" movd %%mm0, (%%edi)\n\t" + +" movd 12(%%ecx), %%mm0\n\t" +" pfsub 8(%%ecx), %%mm0\n\t" +" pfmul 120(%%ebx), %%mm0\n\t" +" movd %%mm0, 512(%%edi)\n\t" +" pfadd 12(%%ecx), %%mm0\n\t" +" pfadd 8(%%ecx), %%mm0\n\t" +" movd %%mm0, 512(%%esi)\n\t" + +" movd 16(%%ecx), %%mm0\n\t" +" pfsub 20(%%ecx), %%mm0\n\t" +" pfmul 120(%%ebx), %%mm0\n\t" +" movq %%mm0, %%mm3\n\t" + +" movd 28(%%ecx), %%mm0\n\t" +" pfsub 24(%%ecx), %%mm0\n\t" +" pfmul 120(%%ebx), %%mm0\n\t" +" movd %%mm0, 768(%%edi)\n\t" +" movq %%mm0, %%mm2\n\t" + +" pfadd 24(%%ecx), %%mm0\n\t" +" pfadd 28(%%ecx), %%mm0\n\t" +" movq %%mm0, %%mm1\n\t" + +" pfadd 16(%%ecx), %%mm0\n\t" +" pfadd 20(%%ecx), %%mm0\n\t" +" movd %%mm0, 768(%%esi)\n\t" +" pfadd %%mm3, %%mm1\n\t" +" movd %%mm1, 256(%%esi)\n\t" +" pfadd %%mm3, %%mm2\n\t" +" movd %%mm2, 256(%%edi)\n\t" /* Phase 8*/ -" movq 32(%%edx), %%mm0\n\t" -" movq 48(%%edx), %%mm1\n\t" -" pfadd 48(%%edx), %%mm0\n\t" -" pfadd 40(%%edx), %%mm1\n\t" -" movd %%mm0, 896(%%esi)\n\t" -" movd %%mm1, 640(%%esi)\n\t" -" psrlq $32, %%mm0\n\t" -" psrlq $32, %%mm1\n\t" -" movd %%mm0, 128(%%edi)\n\t" -" movd %%mm1, 384(%%edi)\n\t" - -" movd 40(%%edx), %%mm0\n\t" -" pfadd 56(%%edx), %%mm0\n\t" -" movd %%mm0, 384(%%esi)\n\t" - -" movd 56(%%edx), %%mm0\n\t" -" pfadd 36(%%edx), %%mm0\n\t" -" movd %%mm0, 128(%%esi)\n\t" - -" movd 60(%%edx), %%mm0\n\t" -" movd %%mm0, 896(%%edi)\n\t" -" pfadd 44(%%edx), %%mm0\n\t" -" movd %%mm0, 640(%%edi)\n\t" - -" movq 96(%%edx), %%mm0\n\t" -" movq 112(%%edx), %%mm2\n\t" -" movq 104(%%edx), %%mm4\n\t" -" pfadd 112(%%edx), %%mm0\n\t" -" pfadd 104(%%edx), %%mm2\n\t" -" pfadd 120(%%edx), %%mm4\n\t" -" movq %%mm0, %%mm1\n\t" -" movq %%mm2, %%mm3\n\t" -" movq %%mm4, %%mm5\n\t" -" pfadd 64(%%edx), %%mm0\n\t" -" pfadd 80(%%edx), %%mm2\n\t" -" pfadd 72(%%edx), %%mm4\n\t" -" movd %%mm0, 960(%%esi)\n\t" -" movd %%mm2, 704(%%esi)\n\t" -" movd %%mm4, 448(%%esi)\n\t" -" psrlq $32, %%mm0\n\t" -" psrlq $32, %%mm2\n\t" -" psrlq $32, %%mm4\n\t" -" movd %%mm0, 64(%%edi)\n\t" -" movd %%mm2, 320(%%edi)\n\t" -" movd %%mm4, 576(%%edi)\n\t" -" pfadd 80(%%edx), %%mm1\n\t" -" pfadd 72(%%edx), %%mm3\n\t" -" pfadd 88(%%edx), %%mm5\n\t" -" movd %%mm1, 832(%%esi)\n\t" -" movd %%mm3, 576(%%esi)\n\t" -" movd %%mm5, 320(%%esi)\n\t" -" psrlq $32, %%mm1\n\t" -" psrlq $32, %%mm3\n\t" -" psrlq $32, %%mm5\n\t" -" movd %%mm1, 192(%%edi)\n\t" -" movd %%mm3, 448(%%edi)\n\t" -" movd %%mm5, 704(%%edi)\n\t" - -" movd 120(%%edx), %%mm0\n\t" -" pfadd 100(%%edx), %%mm0\n\t" -" movq %%mm0, %%mm1\n\t" -" pfadd 88(%%edx), %%mm0\n\t" -" movd %%mm0, 192(%%esi)\n\t" -" pfadd 68(%%edx), %%mm1\n\t" -" movd %%mm1, 64(%%esi)\n\t" - -" movd 124(%%edx), %%mm0\n\t" -" movd %%mm0, 960(%%edi)\n\t" -" pfadd 92(%%edx), %%mm0\n\t" -" movd %%mm0, 832(%%edi)\n\t" - -" jmp .L_bye\n\t" -".L01: \n\t" +" movq 32(%%edx), %%mm0\n\t" +" movq 48(%%edx), %%mm1\n\t" +" pfadd 48(%%edx), %%mm0\n\t" +" pfadd 40(%%edx), %%mm1\n\t" +" movd %%mm0, 896(%%esi)\n\t" +" movd %%mm1, 640(%%esi)\n\t" +" psrlq $32, %%mm0\n\t" +" psrlq $32, %%mm1\n\t" +" movd %%mm0, 128(%%edi)\n\t" +" movd %%mm1, 384(%%edi)\n\t" + +" movd 40(%%edx), %%mm0\n\t" +" pfadd 56(%%edx), %%mm0\n\t" +" movd %%mm0, 384(%%esi)\n\t" + +" movd 56(%%edx), %%mm0\n\t" +" pfadd 36(%%edx), %%mm0\n\t" +" movd %%mm0, 128(%%esi)\n\t" + +" movd 60(%%edx), %%mm0\n\t" +" movd %%mm0, 896(%%edi)\n\t" +" pfadd 44(%%edx), %%mm0\n\t" +" movd %%mm0, 640(%%edi)\n\t" + +" movq 96(%%edx), %%mm0\n\t" +" movq 112(%%edx), %%mm2\n\t" +" movq 104(%%edx), %%mm4\n\t" +" pfadd 112(%%edx), %%mm0\n\t" +" pfadd 104(%%edx), %%mm2\n\t" +" pfadd 120(%%edx), %%mm4\n\t" +" movq %%mm0, %%mm1\n\t" +" movq %%mm2, %%mm3\n\t" +" movq %%mm4, %%mm5\n\t" +" pfadd 64(%%edx), %%mm0\n\t" +" pfadd 80(%%edx), %%mm2\n\t" +" pfadd 72(%%edx), %%mm4\n\t" +" movd %%mm0, 960(%%esi)\n\t" +" movd %%mm2, 704(%%esi)\n\t" +" movd %%mm4, 448(%%esi)\n\t" +" psrlq $32, %%mm0\n\t" +" psrlq $32, %%mm2\n\t" +" psrlq $32, %%mm4\n\t" +" movd %%mm0, 64(%%edi)\n\t" +" movd %%mm2, 320(%%edi)\n\t" +" movd %%mm4, 576(%%edi)\n\t" +" pfadd 80(%%edx), %%mm1\n\t" +" pfadd 72(%%edx), %%mm3\n\t" +" pfadd 88(%%edx), %%mm5\n\t" +" movd %%mm1, 832(%%esi)\n\t" +" movd %%mm3, 576(%%esi)\n\t" +" movd %%mm5, 320(%%esi)\n\t" +" psrlq $32, %%mm1\n\t" +" psrlq $32, %%mm3\n\t" +" psrlq $32, %%mm5\n\t" +" movd %%mm1, 192(%%edi)\n\t" +" movd %%mm3, 448(%%edi)\n\t" +" movd %%mm5, 704(%%edi)\n\t" + +" movd 120(%%edx), %%mm0\n\t" +" pfadd 100(%%edx), %%mm0\n\t" +" movq %%mm0, %%mm1\n\t" +" pfadd 88(%%edx), %%mm0\n\t" +" movd %%mm0, 192(%%esi)\n\t" +" pfadd 68(%%edx), %%mm1\n\t" +" movd %%mm1, 64(%%esi)\n\t" + +" movd 124(%%edx), %%mm0\n\t" +" movd %%mm0, 960(%%edi)\n\t" +" pfadd 92(%%edx), %%mm0\n\t" +" movd %%mm0, 832(%%edi)\n\t" + +" jmp .L_bye\n\t" +".L01: \n\t" #endif /* Phase 9*/ -" movq (%%ecx), %%mm0\n\t" -" movq %%mm0, %%mm1\n\t" -" pxor %%mm7, %%mm1\n\t" -" pfacc %%mm1, %%mm0\n\t" -" pfmul %%mm6, %%mm0\n\t" -" pf2iw %%mm0, %%mm0\n\t" -" movd %%mm0, %%eax\n\t" -" movw %%ax, 512(%%esi)\n\t" -" psrlq $32, %%mm0\n\t" -" movd %%mm0, %%eax\n\t" -" movw %%ax, (%%esi)\n\t" - -" movd 12(%%ecx), %%mm0\n\t" -" pfsub 8(%%ecx), %%mm0\n\t" -" pfmul 120(%%ebx), %%mm0\n\t" -" pf2iw %%mm0, %%mm7\n\t" -" movd %%mm7, %%eax\n\t" -" movw %%ax, 256(%%edi)\n\t" -" pfadd 12(%%ecx), %%mm0\n\t" -" pfadd 8(%%ecx), %%mm0\n\t" -" pf2iw %%mm0, %%mm0\n\t" -" movd %%mm0, %%eax\n\t" -" movw %%ax, 256(%%esi)\n\t" - -" movd 16(%%ecx), %%mm3\n\t" -" pfsub 20(%%ecx), %%mm3\n\t" -" pfmul 120(%%ebx), %%mm3\n\t" -" movq %%mm3, %%mm2\n\t" - -" movd 28(%%ecx), %%mm2\n\t" -" pfsub 24(%%ecx), %%mm2\n\t" -" pfmul 120(%%ebx), %%mm2\n\t" -" movq %%mm2, %%mm1\n\t" - -" pf2iw %%mm2, %%mm7\n\t" -" movd %%mm7, %%eax\n\t" -" movw %%ax, 384(%%edi)\n\t" - -" pfadd 24(%%ecx), %%mm1\n\t" -" pfadd 28(%%ecx), %%mm1\n\t" -" movq %%mm1, %%mm0\n\t" - -" pfadd 16(%%ecx), %%mm0\n\t" -" pfadd 20(%%ecx), %%mm0\n\t" -" pf2iw %%mm0, %%mm0\n\t" -" movd %%mm0, %%eax\n\t" -" movw %%ax, 384(%%esi)\n\t" -" pfadd %%mm3, %%mm1\n\t" -" pf2iw %%mm1, %%mm1\n\t" -" movd %%mm1, %%eax\n\t" -" movw %%ax, 128(%%esi)\n\t" -" pfadd %%mm3, %%mm2\n\t" -" pf2iw %%mm2, %%mm2\n\t" -" movd %%mm2, %%eax\n\t" -" movw %%ax, 128(%%edi)\n\t" +" movq (%%ecx), %%mm0\n\t" +" movq %%mm0, %%mm1\n\t" +" pxor %%mm7, %%mm1\n\t" +" pfacc %%mm1, %%mm0\n\t" +" pfmul %%mm6, %%mm0\n\t" +" pf2iw %%mm0, %%mm0\n\t" +" movd %%mm0, %%eax\n\t" +" movw %%ax, 512(%%esi)\n\t" +" psrlq $32, %%mm0\n\t" +" movd %%mm0, %%eax\n\t" +" movw %%ax, (%%esi)\n\t" + +" movd 12(%%ecx), %%mm0\n\t" +" pfsub 8(%%ecx), %%mm0\n\t" +" pfmul 120(%%ebx), %%mm0\n\t" +" pf2iw %%mm0, %%mm7\n\t" +" movd %%mm7, %%eax\n\t" +" movw %%ax, 256(%%edi)\n\t" +" pfadd 12(%%ecx), %%mm0\n\t" +" pfadd 8(%%ecx), %%mm0\n\t" +" pf2iw %%mm0, %%mm0\n\t" +" movd %%mm0, %%eax\n\t" +" movw %%ax, 256(%%esi)\n\t" + +" movd 16(%%ecx), %%mm3\n\t" +" pfsub 20(%%ecx), %%mm3\n\t" +" pfmul 120(%%ebx), %%mm3\n\t" +" movq %%mm3, %%mm2\n\t" + +" movd 28(%%ecx), %%mm2\n\t" +" pfsub 24(%%ecx), %%mm2\n\t" +" pfmul 120(%%ebx), %%mm2\n\t" +" movq %%mm2, %%mm1\n\t" + +" pf2iw %%mm2, %%mm7\n\t" +" movd %%mm7, %%eax\n\t" +" movw %%ax, 384(%%edi)\n\t" + +" pfadd 24(%%ecx), %%mm1\n\t" +" pfadd 28(%%ecx), %%mm1\n\t" +" movq %%mm1, %%mm0\n\t" + +" pfadd 16(%%ecx), %%mm0\n\t" +" pfadd 20(%%ecx), %%mm0\n\t" +" pf2iw %%mm0, %%mm0\n\t" +" movd %%mm0, %%eax\n\t" +" movw %%ax, 384(%%esi)\n\t" +" pfadd %%mm3, %%mm1\n\t" +" pf2iw %%mm1, %%mm1\n\t" +" movd %%mm1, %%eax\n\t" +" movw %%ax, 128(%%esi)\n\t" +" pfadd %%mm3, %%mm2\n\t" +" pf2iw %%mm2, %%mm2\n\t" +" movd %%mm2, %%eax\n\t" +" movw %%ax, 128(%%edi)\n\t" /* Phase 10*/ -" movq 32(%%edx), %%mm0\n\t" -" movq 48(%%edx), %%mm1\n\t" -" pfadd 48(%%edx), %%mm0\n\t" -" pfadd 40(%%edx), %%mm1\n\t" -" pf2iw %%mm0, %%mm0\n\t" -" pf2iw %%mm1, %%mm1\n\t" -" movd %%mm0, %%eax\n\t" -" movd %%mm1, %%ecx\n\t" -" movw %%ax, 448(%%esi)\n\t" -" movw %%cx, 320(%%esi)\n\t" -" psrlq $32, %%mm0\n\t" -" psrlq $32, %%mm1\n\t" -" movd %%mm0, %%eax\n\t" -" movd %%mm1, %%ecx\n\t" -" movw %%ax, 64(%%edi)\n\t" -" movw %%cx, 192(%%edi)\n\t" - -" movd 40(%%edx), %%mm3\n\t" -" movd 56(%%edx), %%mm4\n\t" -" movd 60(%%edx), %%mm0\n\t" -" movd 44(%%edx), %%mm2\n\t" -" movd 120(%%edx), %%mm5\n\t" -" punpckldq %%mm4, %%mm3\n\t" -" punpckldq 124(%%edx), %%mm0\n\t" -" pfadd 100(%%edx), %%mm5\n\t" -" punpckldq 36(%%edx), %%mm4\n\t" -" punpckldq 92(%%edx), %%mm2\n\t" -" movq %%mm5, %%mm6\n\t" -" pfadd %%mm4, %%mm3\n\t" -" pf2iw %%mm0, %%mm1\n\t" -" pf2iw %%mm3, %%mm3\n\t" -" pfadd 88(%%edx), %%mm5\n\t" -" movd %%mm1, %%eax\n\t" -" movd %%mm3, %%ecx\n\t" -" movw %%ax, 448(%%edi)\n\t" -" movw %%cx, 192(%%esi)\n\t" -" pf2iw %%mm5, %%mm5\n\t" -" psrlq $32, %%mm1\n\t" +" movq 32(%%edx), %%mm0\n\t" +" movq 48(%%edx), %%mm1\n\t" +" pfadd 48(%%edx), %%mm0\n\t" +" pfadd 40(%%edx), %%mm1\n\t" +" pf2iw %%mm0, %%mm0\n\t" +" pf2iw %%mm1, %%mm1\n\t" +" movd %%mm0, %%eax\n\t" +" movd %%mm1, %%ecx\n\t" +" movw %%ax, 448(%%esi)\n\t" +" movw %%cx, 320(%%esi)\n\t" +" psrlq $32, %%mm0\n\t" +" psrlq $32, %%mm1\n\t" +" movd %%mm0, %%eax\n\t" +" movd %%mm1, %%ecx\n\t" +" movw %%ax, 64(%%edi)\n\t" +" movw %%cx, 192(%%edi)\n\t" + +" movd 40(%%edx), %%mm3\n\t" +" movd 56(%%edx), %%mm4\n\t" +" movd 60(%%edx), %%mm0\n\t" +" movd 44(%%edx), %%mm2\n\t" +" movd 120(%%edx), %%mm5\n\t" +" punpckldq %%mm4, %%mm3\n\t" +" punpckldq 124(%%edx), %%mm0\n\t" +" pfadd 100(%%edx), %%mm5\n\t" +" punpckldq 36(%%edx), %%mm4\n\t" +" punpckldq 92(%%edx), %%mm2\n\t" +" movq %%mm5, %%mm6\n\t" +" pfadd %%mm4, %%mm3\n\t" +" pf2iw %%mm0, %%mm1\n\t" +" pf2iw %%mm3, %%mm3\n\t" +" pfadd 88(%%edx), %%mm5\n\t" +" movd %%mm1, %%eax\n\t" +" movd %%mm3, %%ecx\n\t" +" movw %%ax, 448(%%edi)\n\t" +" movw %%cx, 192(%%esi)\n\t" +" pf2iw %%mm5, %%mm5\n\t" +" psrlq $32, %%mm1\n\t" +" psrlq $32, %%mm3\n\t" +" movd %%mm5, %%ebx\n\t" +" movd %%mm1, %%eax\n\t" +" movd %%mm3, %%ecx\n\t" +" movw %%bx, 96(%%esi)\n\t" +" movw %%ax, 480(%%edi)\n\t" +" movw %%cx, 64(%%esi)\n\t" +" pfadd %%mm2, %%mm0\n\t" +" pf2iw %%mm0, %%mm0\n\t" +" movd %%mm0, %%eax\n\t" +" pfadd 68(%%edx), %%mm6\n\t" +" movw %%ax, 320(%%edi)\n\t" +" psrlq $32, %%mm0\n\t" +" pf2iw %%mm6, %%mm6\n\t" +" movd %%mm0, %%eax\n\t" +" movd %%mm6, %%ebx\n\t" +" movw %%ax, 416(%%edi)\n\t" +" movw %%bx, 32(%%esi)\n\t" + +" movq 96(%%edx), %%mm0\n\t" +" movq 112(%%edx), %%mm2\n\t" +" movq 104(%%edx), %%mm4\n\t" +" pfadd %%mm2, %%mm0\n\t" +" pfadd %%mm4, %%mm2\n\t" +" pfadd 120(%%edx), %%mm4\n\t" +" movq %%mm0, %%mm1\n\t" +" movq %%mm2, %%mm3\n\t" +" movq %%mm4, %%mm5\n\t" +" pfadd 64(%%edx), %%mm0\n\t" +" pfadd 80(%%edx), %%mm2\n\t" +" pfadd 72(%%edx), %%mm4\n\t" +" pf2iw %%mm0, %%mm0\n\t" +" pf2iw %%mm2, %%mm2\n\t" +" pf2iw %%mm4, %%mm4\n\t" +" movd %%mm0, %%eax\n\t" +" movd %%mm2, %%ecx\n\t" +" movd %%mm4, %%ebx\n\t" +" movw %%ax, 480(%%esi)\n\t" +" movw %%cx, 352(%%esi)\n\t" +" movw %%bx, 224(%%esi)\n\t" +" psrlq $32, %%mm0\n\t" +" psrlq $32, %%mm2\n\t" +" psrlq $32, %%mm4\n\t" +" movd %%mm0, %%eax\n\t" +" movd %%mm2, %%ecx\n\t" +" movd %%mm4, %%ebx\n\t" +" movw %%ax, 32(%%edi)\n\t" +" movw %%cx, 160(%%edi)\n\t" +" movw %%bx, 288(%%edi)\n\t" +" pfadd 80(%%edx), %%mm1\n\t" +" pfadd 72(%%edx), %%mm3\n\t" +" pfadd 88(%%edx), %%mm5\n\t" +" pf2iw %%mm1, %%mm1\n\t" +" pf2iw %%mm3, %%mm3\n\t" +" pf2iw %%mm5, %%mm5\n\t" +" movd %%mm1, %%eax\n\t" +" movd %%mm3, %%ecx\n\t" +" movd %%mm5, %%ebx\n\t" +" movw %%ax, 416(%%esi)\n\t" +" movw %%cx, 288(%%esi)\n\t" +" movw %%bx, 160(%%esi)\n\t" +" psrlq $32, %%mm1\n\t" " psrlq $32, %%mm3\n\t" -" movd %%mm5, %%ebx\n\t" -" movd %%mm1, %%eax\n\t" -" movd %%mm3, %%ecx\n\t" -" movw %%bx, 96(%%esi)\n\t" -" movw %%ax, 480(%%edi)\n\t" -" movw %%cx, 64(%%esi)\n\t" -" pfadd %%mm2, %%mm0\n\t" -" pf2iw %%mm0, %%mm0\n\t" -" movd %%mm0, %%eax\n\t" -" pfadd 68(%%edx), %%mm6\n\t" -" movw %%ax, 320(%%edi)\n\t" -" psrlq $32, %%mm0\n\t" -" pf2iw %%mm6, %%mm6\n\t" -" movd %%mm0, %%eax\n\t" -" movd %%mm6, %%ebx\n\t" -" movw %%ax, 416(%%edi)\n\t" -" movw %%bx, 32(%%esi)\n\t" - -" movq 96(%%edx), %%mm0\n\t" -" movq 112(%%edx), %%mm2\n\t" -" movq 104(%%edx), %%mm4\n\t" -" pfadd %%mm2, %%mm0\n\t" -" pfadd %%mm4, %%mm2\n\t" -" pfadd 120(%%edx), %%mm4\n\t" -" movq %%mm0, %%mm1\n\t" -" movq %%mm2, %%mm3\n\t" -" movq %%mm4, %%mm5\n\t" -" pfadd 64(%%edx), %%mm0\n\t" -" pfadd 80(%%edx), %%mm2\n\t" -" pfadd 72(%%edx), %%mm4\n\t" -" pf2iw %%mm0, %%mm0\n\t" -" pf2iw %%mm2, %%mm2\n\t" -" pf2iw %%mm4, %%mm4\n\t" -" movd %%mm0, %%eax\n\t" -" movd %%mm2, %%ecx\n\t" -" movd %%mm4, %%ebx\n\t" -" movw %%ax, 480(%%esi)\n\t" -" movw %%cx, 352(%%esi)\n\t" -" movw %%bx, 224(%%esi)\n\t" -" psrlq $32, %%mm0\n\t" -" psrlq $32, %%mm2\n\t" -" psrlq $32, %%mm4\n\t" -" movd %%mm0, %%eax\n\t" -" movd %%mm2, %%ecx\n\t" -" movd %%mm4, %%ebx\n\t" -" movw %%ax, 32(%%edi)\n\t" -" movw %%cx, 160(%%edi)\n\t" -" movw %%bx, 288(%%edi)\n\t" -" pfadd 80(%%edx), %%mm1\n\t" -" pfadd 72(%%edx), %%mm3\n\t" -" pfadd 88(%%edx), %%mm5\n\t" -" pf2iw %%mm1, %%mm1\n\t" -" pf2iw %%mm3, %%mm3\n\t" -" pf2iw %%mm5, %%mm5\n\t" -" movd %%mm1, %%eax\n\t" -" movd %%mm3, %%ecx\n\t" -" movd %%mm5, %%ebx\n\t" -" movw %%ax, 416(%%esi)\n\t" -" movw %%cx, 288(%%esi)\n\t" -" movw %%bx, 160(%%esi)\n\t" -" psrlq $32, %%mm1\n\t" -" psrlq $32, %%mm3\n\t" -" psrlq $32, %%mm5\n\t" -" movd %%mm1, %%eax\n\t" -" movd %%mm3, %%ecx\n\t" -" movd %%mm5, %%ebx\n\t" -" movw %%ax, 96(%%edi)\n\t" -" movw %%cx, 224(%%edi)\n\t" -" movw %%bx, 352(%%edi)\n\t" - -" movsw\n\t" +" psrlq $32, %%mm5\n\t" +" movd %%mm1, %%eax\n\t" +" movd %%mm3, %%ecx\n\t" +" movd %%mm5, %%ebx\n\t" +" movw %%ax, 96(%%edi)\n\t" +" movw %%cx, 224(%%edi)\n\t" +" movw %%bx, 352(%%edi)\n\t" + +" movsw\n\t" ".L_bye:\n\t" -" femms\n\t" - : - :"m"(a),"m"(b),"m"(c),"m"(tmp[0]) - :"memory","%eax","%ebx","%ecx","%edx","%esi","%edi"); +" femms\n\t" + : + :"m"(a),"m"(b),"m"(c),"m"(tmp[0]) + :"memory","%eax","%ebx","%ecx","%edx","%esi","%edi"); } diff --git a/mp3lib/dct64_mmx.c b/mp3lib/dct64_mmx.c index 817913a016..6231ebf723 100644 --- a/mp3lib/dct64_mmx.c +++ b/mp3lib/dct64_mmx.c @@ -11,976 +11,976 @@ void dct64_MMX(short *a,short *b,real *c) { char tmp[256]; __asm__ volatile( -" movl %2,%%eax\n\t" +" movl %2,%%eax\n\t" /* Phase 1*/ -" flds (%%eax)\n\t" -" leal 128+%3,%%edx\n\t" -" fadds 124(%%eax)\n\t" -" movl %0,%%esi\n\t" -" fstps (%%edx)\n\t" -" movl %1,%%edi\n\t" - -" flds 4(%%eax)\n\t" -" movl $"MANGLE(costab_mmx)",%%ebx\n\t" -" fadds 120(%%eax)\n\t" -" orl %%ecx,%%ecx\n\t" -" fstps 4(%%edx)\n\t" - -" flds (%%eax)\n\t" -" leal %3,%%ecx\n\t" -" fsubs 124(%%eax)\n\t" -" fmuls (%%ebx)\n\t" -" fstps 124(%%edx)\n\t" - -" flds 4(%%eax)\n\t" -" fsubs 120(%%eax)\n\t" -" fmuls 4(%%ebx)\n\t" -" fstps 120(%%edx)\n\t" - -" flds 8(%%eax)\n\t" -" fadds 116(%%eax)\n\t" -" fstps 8(%%edx)\n\t" - -" flds 12(%%eax)\n\t" -" fadds 112(%%eax)\n\t" -" fstps 12(%%edx)\n\t" - -" flds 8(%%eax)\n\t" -" fsubs 116(%%eax)\n\t" -" fmuls 8(%%ebx)\n\t" -" fstps 116(%%edx)\n\t" - -" flds 12(%%eax)\n\t" -" fsubs 112(%%eax)\n\t" -" fmuls 12(%%ebx)\n\t" -" fstps 112(%%edx)\n\t" - -" flds 16(%%eax)\n\t" -" fadds 108(%%eax)\n\t" -" fstps 16(%%edx)\n\t" - -" flds 20(%%eax)\n\t" -" fadds 104(%%eax)\n\t" -" fstps 20(%%edx)\n\t" - -" flds 16(%%eax)\n\t" -" fsubs 108(%%eax)\n\t" -" fmuls 16(%%ebx)\n\t" -" fstps 108(%%edx)\n\t" - -" flds 20(%%eax)\n\t" -" fsubs 104(%%eax)\n\t" -" fmuls 20(%%ebx)\n\t" -" fstps 104(%%edx)\n\t" - -" flds 24(%%eax)\n\t" -" fadds 100(%%eax)\n\t" -" fstps 24(%%edx)\n\t" - -" flds 28(%%eax)\n\t" -" fadds 96(%%eax)\n\t" -" fstps 28(%%edx)\n\t" - -" flds 24(%%eax)\n\t" -" fsubs 100(%%eax)\n\t" -" fmuls 24(%%ebx)\n\t" -" fstps 100(%%edx)\n\t" - -" flds 28(%%eax)\n\t" -" fsubs 96(%%eax)\n\t" -" fmuls 28(%%ebx)\n\t" -" fstps 96(%%edx)\n\t" - -" flds 32(%%eax)\n\t" -" fadds 92(%%eax)\n\t" -" fstps 32(%%edx)\n\t" - -" flds 36(%%eax)\n\t" -" fadds 88(%%eax)\n\t" -" fstps 36(%%edx)\n\t" - -" flds 32(%%eax)\n\t" -" fsubs 92(%%eax)\n\t" -" fmuls 32(%%ebx)\n\t" -" fstps 92(%%edx)\n\t" - -" flds 36(%%eax)\n\t" -" fsubs 88(%%eax)\n\t" -" fmuls 36(%%ebx)\n\t" -" fstps 88(%%edx)\n\t" - -" flds 40(%%eax)\n\t" -" fadds 84(%%eax)\n\t" -" fstps 40(%%edx)\n\t" - -" flds 44(%%eax)\n\t" -" fadds 80(%%eax)\n\t" -" fstps 44(%%edx)\n\t" - -" flds 40(%%eax)\n\t" -" fsubs 84(%%eax)\n\t" -" fmuls 40(%%ebx)\n\t" -" fstps 84(%%edx)\n\t" - -" flds 44(%%eax)\n\t" -" fsubs 80(%%eax)\n\t" -" fmuls 44(%%ebx)\n\t" -" fstps 80(%%edx)\n\t" - -" flds 48(%%eax)\n\t" -" fadds 76(%%eax)\n\t" -" fstps 48(%%edx)\n\t" - -" flds 52(%%eax)\n\t" -" fadds 72(%%eax)\n\t" -" fstps 52(%%edx)\n\t" - -" flds 48(%%eax)\n\t" -" fsubs 76(%%eax)\n\t" -" fmuls 48(%%ebx)\n\t" -" fstps 76(%%edx)\n\t" - -" flds 52(%%eax)\n\t" -" fsubs 72(%%eax)\n\t" -" fmuls 52(%%ebx)\n\t" -" fstps 72(%%edx)\n\t" - -" flds 56(%%eax)\n\t" -" fadds 68(%%eax)\n\t" -" fstps 56(%%edx)\n\t" - -" flds 60(%%eax)\n\t" -" fadds 64(%%eax)\n\t" -" fstps 60(%%edx)\n\t" - -" flds 56(%%eax)\n\t" -" fsubs 68(%%eax)\n\t" -" fmuls 56(%%ebx)\n\t" -" fstps 68(%%edx)\n\t" - -" flds 60(%%eax)\n\t" -" fsubs 64(%%eax)\n\t" -" fmuls 60(%%ebx)\n\t" -" fstps 64(%%edx)\n\t" +" flds (%%eax)\n\t" +" leal 128+%3,%%edx\n\t" +" fadds 124(%%eax)\n\t" +" movl %0,%%esi\n\t" +" fstps (%%edx)\n\t" +" movl %1,%%edi\n\t" + +" flds 4(%%eax)\n\t" +" movl $"MANGLE(costab_mmx)",%%ebx\n\t" +" fadds 120(%%eax)\n\t" +" orl %%ecx,%%ecx\n\t" +" fstps 4(%%edx)\n\t" + +" flds (%%eax)\n\t" +" leal %3,%%ecx\n\t" +" fsubs 124(%%eax)\n\t" +" fmuls (%%ebx)\n\t" +" fstps 124(%%edx)\n\t" + +" flds 4(%%eax)\n\t" +" fsubs 120(%%eax)\n\t" +" fmuls 4(%%ebx)\n\t" +" fstps 120(%%edx)\n\t" + +" flds 8(%%eax)\n\t" +" fadds 116(%%eax)\n\t" +" fstps 8(%%edx)\n\t" + +" flds 12(%%eax)\n\t" +" fadds 112(%%eax)\n\t" +" fstps 12(%%edx)\n\t" + +" flds 8(%%eax)\n\t" +" fsubs 116(%%eax)\n\t" +" fmuls 8(%%ebx)\n\t" +" fstps 116(%%edx)\n\t" + +" flds 12(%%eax)\n\t" +" fsubs 112(%%eax)\n\t" +" fmuls 12(%%ebx)\n\t" +" fstps 112(%%edx)\n\t" + +" flds 16(%%eax)\n\t" +" fadds 108(%%eax)\n\t" +" fstps 16(%%edx)\n\t" + +" flds 20(%%eax)\n\t" +" fadds 104(%%eax)\n\t" +" fstps 20(%%edx)\n\t" + +" flds 16(%%eax)\n\t" +" fsubs 108(%%eax)\n\t" +" fmuls 16(%%ebx)\n\t" +" fstps 108(%%edx)\n\t" + +" flds 20(%%eax)\n\t" +" fsubs 104(%%eax)\n\t" +" fmuls 20(%%ebx)\n\t" +" fstps 104(%%edx)\n\t" + +" flds 24(%%eax)\n\t" +" fadds 100(%%eax)\n\t" +" fstps 24(%%edx)\n\t" + +" flds 28(%%eax)\n\t" +" fadds 96(%%eax)\n\t" +" fstps 28(%%edx)\n\t" + +" flds 24(%%eax)\n\t" +" fsubs 100(%%eax)\n\t" +" fmuls 24(%%ebx)\n\t" +" fstps 100(%%edx)\n\t" + +" flds 28(%%eax)\n\t" +" fsubs 96(%%eax)\n\t" +" fmuls 28(%%ebx)\n\t" +" fstps 96(%%edx)\n\t" + +" flds 32(%%eax)\n\t" +" fadds 92(%%eax)\n\t" +" fstps 32(%%edx)\n\t" + +" flds 36(%%eax)\n\t" +" fadds 88(%%eax)\n\t" +" fstps 36(%%edx)\n\t" + +" flds 32(%%eax)\n\t" +" fsubs 92(%%eax)\n\t" +" fmuls 32(%%ebx)\n\t" +" fstps 92(%%edx)\n\t" + +" flds 36(%%eax)\n\t" +" fsubs 88(%%eax)\n\t" +" fmuls 36(%%ebx)\n\t" +" fstps 88(%%edx)\n\t" + +" flds 40(%%eax)\n\t" +" fadds 84(%%eax)\n\t" +" fstps 40(%%edx)\n\t" + +" flds 44(%%eax)\n\t" +" fadds 80(%%eax)\n\t" +" fstps 44(%%edx)\n\t" + +" flds 40(%%eax)\n\t" +" fsubs 84(%%eax)\n\t" +" fmuls 40(%%ebx)\n\t" +" fstps 84(%%edx)\n\t" + +" flds 44(%%eax)\n\t" +" fsubs 80(%%eax)\n\t" +" fmuls 44(%%ebx)\n\t" +" fstps 80(%%edx)\n\t" + +" flds 48(%%eax)\n\t" +" fadds 76(%%eax)\n\t" +" fstps 48(%%edx)\n\t" + +" flds 52(%%eax)\n\t" +" fadds 72(%%eax)\n\t" +" fstps 52(%%edx)\n\t" + +" flds 48(%%eax)\n\t" +" fsubs 76(%%eax)\n\t" +" fmuls 48(%%ebx)\n\t" +" fstps 76(%%edx)\n\t" + +" flds 52(%%eax)\n\t" +" fsubs 72(%%eax)\n\t" +" fmuls 52(%%ebx)\n\t" +" fstps 72(%%edx)\n\t" + +" flds 56(%%eax)\n\t" +" fadds 68(%%eax)\n\t" +" fstps 56(%%edx)\n\t" + +" flds 60(%%eax)\n\t" +" fadds 64(%%eax)\n\t" +" fstps 60(%%edx)\n\t" + +" flds 56(%%eax)\n\t" +" fsubs 68(%%eax)\n\t" +" fmuls 56(%%ebx)\n\t" +" fstps 68(%%edx)\n\t" + +" flds 60(%%eax)\n\t" +" fsubs 64(%%eax)\n\t" +" fmuls 60(%%ebx)\n\t" +" fstps 64(%%edx)\n\t" /* Phase 2*/ -" flds (%%edx)\n\t" -" fadds 60(%%edx)\n\t" -" fstps (%%ecx)\n\t" - -" flds 4(%%edx)\n\t" -" fadds 56(%%edx)\n\t" -" fstps 4(%%ecx)\n\t" - -" flds (%%edx)\n\t" -" fsubs 60(%%edx)\n\t" -" fmuls 64(%%ebx)\n\t" -" fstps 60(%%ecx)\n\t" - -" flds 4(%%edx)\n\t" -" fsubs 56(%%edx)\n\t" -" fmuls 68(%%ebx)\n\t" -" fstps 56(%%ecx)\n\t" - -" flds 8(%%edx)\n\t" -" fadds 52(%%edx)\n\t" -" fstps 8(%%ecx)\n\t" - -" flds 12(%%edx)\n\t" -" fadds 48(%%edx)\n\t" -" fstps 12(%%ecx)\n\t" - -" flds 8(%%edx)\n\t" -" fsubs 52(%%edx)\n\t" -" fmuls 72(%%ebx)\n\t" -" fstps 52(%%ecx)\n\t" - -" flds 12(%%edx)\n\t" -" fsubs 48(%%edx)\n\t" -" fmuls 76(%%ebx)\n\t" -" fstps 48(%%ecx)\n\t" - -" flds 16(%%edx)\n\t" -" fadds 44(%%edx)\n\t" -" fstps 16(%%ecx)\n\t" - -" flds 20(%%edx)\n\t" -" fadds 40(%%edx)\n\t" -" fstps 20(%%ecx)\n\t" - -" flds 16(%%edx)\n\t" -" fsubs 44(%%edx)\n\t" -" fmuls 80(%%ebx)\n\t" -" fstps 44(%%ecx)\n\t" - -" flds 20(%%edx)\n\t" -" fsubs 40(%%edx)\n\t" -" fmuls 84(%%ebx)\n\t" -" fstps 40(%%ecx)\n\t" - -" flds 24(%%edx)\n\t" -" fadds 36(%%edx)\n\t" -" fstps 24(%%ecx)\n\t" - -" flds 28(%%edx)\n\t" -" fadds 32(%%edx)\n\t" -" fstps 28(%%ecx)\n\t" - -" flds 24(%%edx)\n\t" -" fsubs 36(%%edx)\n\t" -" fmuls 88(%%ebx)\n\t" -" fstps 36(%%ecx)\n\t" - -" flds 28(%%edx)\n\t" -" fsubs 32(%%edx)\n\t" -" fmuls 92(%%ebx)\n\t" -" fstps 32(%%ecx)\n\t" +" flds (%%edx)\n\t" +" fadds 60(%%edx)\n\t" +" fstps (%%ecx)\n\t" + +" flds 4(%%edx)\n\t" +" fadds 56(%%edx)\n\t" +" fstps 4(%%ecx)\n\t" + +" flds (%%edx)\n\t" +" fsubs 60(%%edx)\n\t" +" fmuls 64(%%ebx)\n\t" +" fstps 60(%%ecx)\n\t" + +" flds 4(%%edx)\n\t" +" fsubs 56(%%edx)\n\t" +" fmuls 68(%%ebx)\n\t" +" fstps 56(%%ecx)\n\t" + +" flds 8(%%edx)\n\t" +" fadds 52(%%edx)\n\t" +" fstps 8(%%ecx)\n\t" + +" flds 12(%%edx)\n\t" +" fadds 48(%%edx)\n\t" +" fstps 12(%%ecx)\n\t" + +" flds 8(%%edx)\n\t" +" fsubs 52(%%edx)\n\t" +" fmuls 72(%%ebx)\n\t" +" fstps 52(%%ecx)\n\t" + +" flds 12(%%edx)\n\t" +" fsubs 48(%%edx)\n\t" +" fmuls 76(%%ebx)\n\t" +" fstps 48(%%ecx)\n\t" + +" flds 16(%%edx)\n\t" +" fadds 44(%%edx)\n\t" +" fstps 16(%%ecx)\n\t" + +" flds 20(%%edx)\n\t" +" fadds 40(%%edx)\n\t" +" fstps 20(%%ecx)\n\t" + +" flds 16(%%edx)\n\t" +" fsubs 44(%%edx)\n\t" +" fmuls 80(%%ebx)\n\t" +" fstps 44(%%ecx)\n\t" + +" flds 20(%%edx)\n\t" +" fsubs 40(%%edx)\n\t" +" fmuls 84(%%ebx)\n\t" +" fstps 40(%%ecx)\n\t" + +" flds 24(%%edx)\n\t" +" fadds 36(%%edx)\n\t" +" fstps 24(%%ecx)\n\t" + +" flds 28(%%edx)\n\t" +" fadds 32(%%edx)\n\t" +" fstps 28(%%ecx)\n\t" + +" flds 24(%%edx)\n\t" +" fsubs 36(%%edx)\n\t" +" fmuls 88(%%ebx)\n\t" +" fstps 36(%%ecx)\n\t" + +" flds 28(%%edx)\n\t" +" fsubs 32(%%edx)\n\t" +" fmuls 92(%%ebx)\n\t" +" fstps 32(%%ecx)\n\t" /* Phase 3*/ -" flds 64(%%edx)\n\t" -" fadds 124(%%edx)\n\t" -" fstps 64(%%ecx)\n\t" - -" flds 68(%%edx)\n\t" -" fadds 120(%%edx)\n\t" -" fstps 68(%%ecx)\n\t" - -" flds 124(%%edx)\n\t" -" fsubs 64(%%edx)\n\t" -" fmuls 64(%%ebx)\n\t" -" fstps 124(%%ecx)\n\t" - -" flds 120(%%edx)\n\t" -" fsubs 68(%%edx)\n\t" -" fmuls 68(%%ebx)\n\t" -" fstps 120(%%ecx)\n\t" - -" flds 72(%%edx)\n\t" -" fadds 116(%%edx)\n\t" -" fstps 72(%%ecx)\n\t" - -" flds 76(%%edx)\n\t" -" fadds 112(%%edx)\n\t" -" fstps 76(%%ecx)\n\t" - -" flds 116(%%edx)\n\t" -" fsubs 72(%%edx)\n\t" -" fmuls 72(%%ebx)\n\t" -" fstps 116(%%ecx)\n\t" - -" flds 112(%%edx)\n\t" -" fsubs 76(%%edx)\n\t" -" fmuls 76(%%ebx)\n\t" -" fstps 112(%%ecx)\n\t" - -" flds 80(%%edx)\n\t" -" fadds 108(%%edx)\n\t" -" fstps 80(%%ecx)\n\t" - -" flds 84(%%edx)\n\t" -" fadds 104(%%edx)\n\t" -" fstps 84(%%ecx)\n\t" - -" flds 108(%%edx)\n\t" -" fsubs 80(%%edx)\n\t" -" fmuls 80(%%ebx)\n\t" -" fstps 108(%%ecx)\n\t" - -" flds 104(%%edx)\n\t" -" fsubs 84(%%edx)\n\t" -" fmuls 84(%%ebx)\n\t" -" fstps 104(%%ecx)\n\t" - -" flds 88(%%edx)\n\t" -" fadds 100(%%edx)\n\t" -" fstps 88(%%ecx)\n\t" - -" flds 92(%%edx)\n\t" -" fadds 96(%%edx)\n\t" -" fstps 92(%%ecx)\n\t" - -" flds 100(%%edx)\n\t" -" fsubs 88(%%edx)\n\t" -" fmuls 88(%%ebx)\n\t" -" fstps 100(%%ecx)\n\t" - -" flds 96(%%edx)\n\t" -" fsubs 92(%%edx)\n\t" -" fmuls 92(%%ebx)\n\t" -" fstps 96(%%ecx)\n\t" +" flds 64(%%edx)\n\t" +" fadds 124(%%edx)\n\t" +" fstps 64(%%ecx)\n\t" + +" flds 68(%%edx)\n\t" +" fadds 120(%%edx)\n\t" +" fstps 68(%%ecx)\n\t" + +" flds 124(%%edx)\n\t" +" fsubs 64(%%edx)\n\t" +" fmuls 64(%%ebx)\n\t" +" fstps 124(%%ecx)\n\t" + +" flds 120(%%edx)\n\t" +" fsubs 68(%%edx)\n\t" +" fmuls 68(%%ebx)\n\t" +" fstps 120(%%ecx)\n\t" + +" flds 72(%%edx)\n\t" +" fadds 116(%%edx)\n\t" +" fstps 72(%%ecx)\n\t" + +" flds 76(%%edx)\n\t" +" fadds 112(%%edx)\n\t" +" fstps 76(%%ecx)\n\t" + +" flds 116(%%edx)\n\t" +" fsubs 72(%%edx)\n\t" +" fmuls 72(%%ebx)\n\t" +" fstps 116(%%ecx)\n\t" + +" flds 112(%%edx)\n\t" +" fsubs 76(%%edx)\n\t" +" fmuls 76(%%ebx)\n\t" +" fstps 112(%%ecx)\n\t" + +" flds 80(%%edx)\n\t" +" fadds 108(%%edx)\n\t" +" fstps 80(%%ecx)\n\t" + +" flds 84(%%edx)\n\t" +" fadds 104(%%edx)\n\t" +" fstps 84(%%ecx)\n\t" + +" flds 108(%%edx)\n\t" +" fsubs 80(%%edx)\n\t" +" fmuls 80(%%ebx)\n\t" +" fstps 108(%%ecx)\n\t" + +" flds 104(%%edx)\n\t" +" fsubs 84(%%edx)\n\t" +" fmuls 84(%%ebx)\n\t" +" fstps 104(%%ecx)\n\t" + +" flds 88(%%edx)\n\t" +" fadds 100(%%edx)\n\t" +" fstps 88(%%ecx)\n\t" + +" flds 92(%%edx)\n\t" +" fadds 96(%%edx)\n\t" +" fstps 92(%%ecx)\n\t" + +" flds 100(%%edx)\n\t" +" fsubs 88(%%edx)\n\t" +" fmuls 88(%%ebx)\n\t" +" fstps 100(%%ecx)\n\t" + +" flds 96(%%edx)\n\t" +" fsubs 92(%%edx)\n\t" +" fmuls 92(%%ebx)\n\t" +" fstps 96(%%ecx)\n\t" /* Phase 4*/ -" flds (%%ecx)\n\t" -" fadds 28(%%ecx)\n\t" -" fstps (%%edx)\n\t" - -" flds (%%ecx)\n\t" -" fsubs 28(%%ecx)\n\t" -" fmuls 96(%%ebx)\n\t" -" fstps 28(%%edx)\n\t" - -" flds 4(%%ecx)\n\t" -" fadds 24(%%ecx)\n\t" -" fstps 4(%%edx)\n\t" - -" flds 4(%%ecx)\n\t" -" fsubs 24(%%ecx)\n\t" -" fmuls 100(%%ebx)\n\t" -" fstps 24(%%edx)\n\t" - -" flds 8(%%ecx)\n\t" -" fadds 20(%%ecx)\n\t" -" fstps 8(%%edx)\n\t" - -" flds 8(%%ecx)\n\t" -" fsubs 20(%%ecx)\n\t" -" fmuls 104(%%ebx)\n\t" -" fstps 20(%%edx)\n\t" - -" flds 12(%%ecx)\n\t" -" fadds 16(%%ecx)\n\t" -" fstps 12(%%edx)\n\t" - -" flds 12(%%ecx)\n\t" -" fsubs 16(%%ecx)\n\t" -" fmuls 108(%%ebx)\n\t" -" fstps 16(%%edx)\n\t" - -" flds 32(%%ecx)\n\t" -" fadds 60(%%ecx)\n\t" -" fstps 32(%%edx)\n\t" - -" flds 60(%%ecx)\n\t" -" fsubs 32(%%ecx)\n\t" -" fmuls 96(%%ebx)\n\t" -" fstps 60(%%edx)\n\t" - -" flds 36(%%ecx)\n\t" -" fadds 56(%%ecx)\n\t" -" fstps 36(%%edx)\n\t" - -" flds 56(%%ecx)\n\t" -" fsubs 36(%%ecx)\n\t" -" fmuls 100(%%ebx)\n\t" -" fstps 56(%%edx)\n\t" - -" flds 40(%%ecx)\n\t" -" fadds 52(%%ecx)\n\t" -" fstps 40(%%edx)\n\t" - -" flds 52(%%ecx)\n\t" -" fsubs 40(%%ecx)\n\t" -" fmuls 104(%%ebx)\n\t" -" fstps 52(%%edx)\n\t" - -" flds 44(%%ecx)\n\t" -" fadds 48(%%ecx)\n\t" -" fstps 44(%%edx)\n\t" - -" flds 48(%%ecx)\n\t" -" fsubs 44(%%ecx)\n\t" -" fmuls 108(%%ebx)\n\t" -" fstps 48(%%edx)\n\t" - -" flds 64(%%ecx)\n\t" -" fadds 92(%%ecx)\n\t" -" fstps 64(%%edx)\n\t" - -" flds 64(%%ecx)\n\t" -" fsubs 92(%%ecx)\n\t" -" fmuls 96(%%ebx)\n\t" -" fstps 92(%%edx)\n\t" - -" flds 68(%%ecx)\n\t" -" fadds 88(%%ecx)\n\t" -" fstps 68(%%edx)\n\t" - -" flds 68(%%ecx)\n\t" -" fsubs 88(%%ecx)\n\t" -" fmuls 100(%%ebx)\n\t" -" fstps 88(%%edx)\n\t" - -" flds 72(%%ecx)\n\t" -" fadds 84(%%ecx)\n\t" -" fstps 72(%%edx)\n\t" - -" flds 72(%%ecx)\n\t" -" fsubs 84(%%ecx)\n\t" -" fmuls 104(%%ebx)\n\t" -" fstps 84(%%edx)\n\t" - -" flds 76(%%ecx)\n\t" -" fadds 80(%%ecx)\n\t" -" fstps 76(%%edx)\n\t" - -" flds 76(%%ecx)\n\t" -" fsubs 80(%%ecx)\n\t" -" fmuls 108(%%ebx)\n\t" -" fstps 80(%%edx)\n\t" - -" flds 96(%%ecx)\n\t" -" fadds 124(%%ecx)\n\t" -" fstps 96(%%edx)\n\t" - -" flds 124(%%ecx)\n\t" -" fsubs 96(%%ecx)\n\t" -" fmuls 96(%%ebx)\n\t" -" fstps 124(%%edx)\n\t" - -" flds 100(%%ecx)\n\t" -" fadds 120(%%ecx)\n\t" -" fstps 100(%%edx)\n\t" - -" flds 120(%%ecx)\n\t" -" fsubs 100(%%ecx)\n\t" -" fmuls 100(%%ebx)\n\t" -" fstps 120(%%edx)\n\t" - -" flds 104(%%ecx)\n\t" -" fadds 116(%%ecx)\n\t" -" fstps 104(%%edx)\n\t" - -" flds 116(%%ecx)\n\t" -" fsubs 104(%%ecx)\n\t" -" fmuls 104(%%ebx)\n\t" -" fstps 116(%%edx)\n\t" - -" flds 108(%%ecx)\n\t" -" fadds 112(%%ecx)\n\t" -" fstps 108(%%edx)\n\t" - -" flds 112(%%ecx)\n\t" -" fsubs 108(%%ecx)\n\t" -" fmuls 108(%%ebx)\n\t" -" fstps 112(%%edx)\n\t" - -" flds (%%edx)\n\t" -" fadds 12(%%edx)\n\t" -" fstps (%%ecx)\n\t" - -" flds (%%edx)\n\t" -" fsubs 12(%%edx)\n\t" -" fmuls 112(%%ebx)\n\t" -" fstps 12(%%ecx)\n\t" - -" flds 4(%%edx)\n\t" -" fadds 8(%%edx)\n\t" -" fstps 4(%%ecx)\n\t" - -" flds 4(%%edx)\n\t" -" fsubs 8(%%edx)\n\t" -" fmuls 116(%%ebx)\n\t" -" fstps 8(%%ecx)\n\t" - -" flds 16(%%edx)\n\t" -" fadds 28(%%edx)\n\t" -" fstps 16(%%ecx)\n\t" - -" flds 28(%%edx)\n\t" -" fsubs 16(%%edx)\n\t" -" fmuls 112(%%ebx)\n\t" -" fstps 28(%%ecx)\n\t" - -" flds 20(%%edx)\n\t" -" fadds 24(%%edx)\n\t" -" fstps 20(%%ecx)\n\t" - -" flds 24(%%edx)\n\t" -" fsubs 20(%%edx)\n\t" -" fmuls 116(%%ebx)\n\t" -" fstps 24(%%ecx)\n\t" - -" flds 32(%%edx)\n\t" -" fadds 44(%%edx)\n\t" -" fstps 32(%%ecx)\n\t" - -" flds 32(%%edx)\n\t" -" fsubs 44(%%edx)\n\t" -" fmuls 112(%%ebx)\n\t" -" fstps 44(%%ecx)\n\t" - -" flds 36(%%edx)\n\t" -" fadds 40(%%edx)\n\t" -" fstps 36(%%ecx)\n\t" - -" flds 36(%%edx)\n\t" -" fsubs 40(%%edx)\n\t" -" fmuls 116(%%ebx)\n\t" -" fstps 40(%%ecx)\n\t" - -" flds 48(%%edx)\n\t" -" fadds 60(%%edx)\n\t" -" fstps 48(%%ecx)\n\t" - -" flds 60(%%edx)\n\t" -" fsubs 48(%%edx)\n\t" -" fmuls 112(%%ebx)\n\t" -" fstps 60(%%ecx)\n\t" - -" flds 52(%%edx)\n\t" -" fadds 56(%%edx)\n\t" -" fstps 52(%%ecx)\n\t" - -" flds 56(%%edx)\n\t" -" fsubs 52(%%edx)\n\t" -" fmuls 116(%%ebx)\n\t" -" fstps 56(%%ecx)\n\t" - -" flds 64(%%edx)\n\t" -" fadds 76(%%edx)\n\t" -" fstps 64(%%ecx)\n\t" - -" flds 64(%%edx)\n\t" -" fsubs 76(%%edx)\n\t" -" fmuls 112(%%ebx)\n\t" -" fstps 76(%%ecx)\n\t" - -" flds 68(%%edx)\n\t" -" fadds 72(%%edx)\n\t" -" fstps 68(%%ecx)\n\t" - -" flds 68(%%edx)\n\t" -" fsubs 72(%%edx)\n\t" -" fmuls 116(%%ebx)\n\t" -" fstps 72(%%ecx)\n\t" - -" flds 80(%%edx)\n\t" -" fadds 92(%%edx)\n\t" -" fstps 80(%%ecx)\n\t" - -" flds 92(%%edx)\n\t" -" fsubs 80(%%edx)\n\t" -" fmuls 112(%%ebx)\n\t" -" fstps 92(%%ecx)\n\t" - -" flds 84(%%edx)\n\t" -" fadds 88(%%edx)\n\t" -" fstps 84(%%ecx)\n\t" - -" flds 88(%%edx)\n\t" -" fsubs 84(%%edx)\n\t" -" fmuls 116(%%ebx)\n\t" -" fstps 88(%%ecx)\n\t" - -" flds 96(%%edx)\n\t" -" fadds 108(%%edx)\n\t" -" fstps 96(%%ecx)\n\t" - -" flds 96(%%edx)\n\t" -" fsubs 108(%%edx)\n\t" -" fmuls 112(%%ebx)\n\t" -" fstps 108(%%ecx)\n\t" - -" flds 100(%%edx)\n\t" -" fadds 104(%%edx)\n\t" -" fstps 100(%%ecx)\n\t" - -" flds 100(%%edx)\n\t" -" fsubs 104(%%edx)\n\t" -" fmuls 116(%%ebx)\n\t" -" fstps 104(%%ecx)\n\t" - -" flds 112(%%edx)\n\t" -" fadds 124(%%edx)\n\t" -" fstps 112(%%ecx)\n\t" - -" flds 124(%%edx)\n\t" -" fsubs 112(%%edx)\n\t" -" fmuls 112(%%ebx)\n\t" -" fstps 124(%%ecx)\n\t" - -" flds 116(%%edx)\n\t" -" fadds 120(%%edx)\n\t" -" fstps 116(%%ecx)\n\t" - -" flds 120(%%edx)\n\t" -" fsubs 116(%%edx)\n\t" -" fmuls 116(%%ebx)\n\t" -" fstps 120(%%ecx)\n\t" +" flds (%%ecx)\n\t" +" fadds 28(%%ecx)\n\t" +" fstps (%%edx)\n\t" + +" flds (%%ecx)\n\t" +" fsubs 28(%%ecx)\n\t" +" fmuls 96(%%ebx)\n\t" +" fstps 28(%%edx)\n\t" + +" flds 4(%%ecx)\n\t" +" fadds 24(%%ecx)\n\t" +" fstps 4(%%edx)\n\t" + +" flds 4(%%ecx)\n\t" +" fsubs 24(%%ecx)\n\t" +" fmuls 100(%%ebx)\n\t" +" fstps 24(%%edx)\n\t" + +" flds 8(%%ecx)\n\t" +" fadds 20(%%ecx)\n\t" +" fstps 8(%%edx)\n\t" + +" flds 8(%%ecx)\n\t" +" fsubs 20(%%ecx)\n\t" +" fmuls 104(%%ebx)\n\t" +" fstps 20(%%edx)\n\t" + +" flds 12(%%ecx)\n\t" +" fadds 16(%%ecx)\n\t" +" fstps 12(%%edx)\n\t" + +" flds 12(%%ecx)\n\t" +" fsubs 16(%%ecx)\n\t" +" fmuls 108(%%ebx)\n\t" +" fstps 16(%%edx)\n\t" + +" flds 32(%%ecx)\n\t" +" fadds 60(%%ecx)\n\t" +" fstps 32(%%edx)\n\t" + +" flds 60(%%ecx)\n\t" +" fsubs 32(%%ecx)\n\t" +" fmuls 96(%%ebx)\n\t" +" fstps 60(%%edx)\n\t" + +" flds 36(%%ecx)\n\t" +" fadds 56(%%ecx)\n\t" +" fstps 36(%%edx)\n\t" + +" flds 56(%%ecx)\n\t" +" fsubs 36(%%ecx)\n\t" +" fmuls 100(%%ebx)\n\t" +" fstps 56(%%edx)\n\t" + +" flds 40(%%ecx)\n\t" +" fadds 52(%%ecx)\n\t" +" fstps 40(%%edx)\n\t" + +" flds 52(%%ecx)\n\t" +" fsubs 40(%%ecx)\n\t" +" fmuls 104(%%ebx)\n\t" +" fstps 52(%%edx)\n\t" + +" flds 44(%%ecx)\n\t" +" fadds 48(%%ecx)\n\t" +" fstps 44(%%edx)\n\t" + +" flds 48(%%ecx)\n\t" +" fsubs 44(%%ecx)\n\t" +" fmuls 108(%%ebx)\n\t" +" fstps 48(%%edx)\n\t" + +" flds 64(%%ecx)\n\t" +" fadds 92(%%ecx)\n\t" +" fstps 64(%%edx)\n\t" + +" flds 64(%%ecx)\n\t" +" fsubs 92(%%ecx)\n\t" +" fmuls 96(%%ebx)\n\t" +" fstps 92(%%edx)\n\t" + +" flds 68(%%ecx)\n\t" +" fadds 88(%%ecx)\n\t" +" fstps 68(%%edx)\n\t" + +" flds 68(%%ecx)\n\t" +" fsubs 88(%%ecx)\n\t" +" fmuls 100(%%ebx)\n\t" +" fstps 88(%%edx)\n\t" + +" flds 72(%%ecx)\n\t" +" fadds 84(%%ecx)\n\t" +" fstps 72(%%edx)\n\t" + +" flds 72(%%ecx)\n\t" +" fsubs 84(%%ecx)\n\t" +" fmuls 104(%%ebx)\n\t" +" fstps 84(%%edx)\n\t" + +" flds 76(%%ecx)\n\t" +" fadds 80(%%ecx)\n\t" +" fstps 76(%%edx)\n\t" + +" flds 76(%%ecx)\n\t" +" fsubs 80(%%ecx)\n\t" +" fmuls 108(%%ebx)\n\t" +" fstps 80(%%edx)\n\t" + +" flds 96(%%ecx)\n\t" +" fadds 124(%%ecx)\n\t" +" fstps 96(%%edx)\n\t" + +" flds 124(%%ecx)\n\t" +" fsubs 96(%%ecx)\n\t" +" fmuls 96(%%ebx)\n\t" +" fstps 124(%%edx)\n\t" + +" flds 100(%%ecx)\n\t" +" fadds 120(%%ecx)\n\t" +" fstps 100(%%edx)\n\t" + +" flds 120(%%ecx)\n\t" +" fsubs 100(%%ecx)\n\t" +" fmuls 100(%%ebx)\n\t" +" fstps 120(%%edx)\n\t" + +" flds 104(%%ecx)\n\t" +" fadds 116(%%ecx)\n\t" +" fstps 104(%%edx)\n\t" + +" flds 116(%%ecx)\n\t" +" fsubs 104(%%ecx)\n\t" +" fmuls 104(%%ebx)\n\t" +" fstps 116(%%edx)\n\t" + +" flds 108(%%ecx)\n\t" +" fadds 112(%%ecx)\n\t" +" fstps 108(%%edx)\n\t" + +" flds 112(%%ecx)\n\t" +" fsubs 108(%%ecx)\n\t" +" fmuls 108(%%ebx)\n\t" +" fstps 112(%%edx)\n\t" + +" flds (%%edx)\n\t" +" fadds 12(%%edx)\n\t" +" fstps (%%ecx)\n\t" + +" flds (%%edx)\n\t" +" fsubs 12(%%edx)\n\t" +" fmuls 112(%%ebx)\n\t" +" fstps 12(%%ecx)\n\t" + +" flds 4(%%edx)\n\t" +" fadds 8(%%edx)\n\t" +" fstps 4(%%ecx)\n\t" + +" flds 4(%%edx)\n\t" +" fsubs 8(%%edx)\n\t" +" fmuls 116(%%ebx)\n\t" +" fstps 8(%%ecx)\n\t" + +" flds 16(%%edx)\n\t" +" fadds 28(%%edx)\n\t" +" fstps 16(%%ecx)\n\t" + +" flds 28(%%edx)\n\t" +" fsubs 16(%%edx)\n\t" +" fmuls 112(%%ebx)\n\t" +" fstps 28(%%ecx)\n\t" + +" flds 20(%%edx)\n\t" +" fadds 24(%%edx)\n\t" +" fstps 20(%%ecx)\n\t" + +" flds 24(%%edx)\n\t" +" fsubs 20(%%edx)\n\t" +" fmuls 116(%%ebx)\n\t" +" fstps 24(%%ecx)\n\t" + +" flds 32(%%edx)\n\t" +" fadds 44(%%edx)\n\t" +" fstps 32(%%ecx)\n\t" + +" flds 32(%%edx)\n\t" +" fsubs 44(%%edx)\n\t" +" fmuls 112(%%ebx)\n\t" +" fstps 44(%%ecx)\n\t" + +" flds 36(%%edx)\n\t" +" fadds 40(%%edx)\n\t" +" fstps 36(%%ecx)\n\t" + +" flds 36(%%edx)\n\t" +" fsubs 40(%%edx)\n\t" +" fmuls 116(%%ebx)\n\t" +" fstps 40(%%ecx)\n\t" + +" flds 48(%%edx)\n\t" +" fadds 60(%%edx)\n\t" +" fstps 48(%%ecx)\n\t" + +" flds 60(%%edx)\n\t" +" fsubs 48(%%edx)\n\t" +" fmuls 112(%%ebx)\n\t" +" fstps 60(%%ecx)\n\t" + +" flds 52(%%edx)\n\t" +" fadds 56(%%edx)\n\t" +" fstps 52(%%ecx)\n\t" + +" flds 56(%%edx)\n\t" +" fsubs 52(%%edx)\n\t" +" fmuls 116(%%ebx)\n\t" +" fstps 56(%%ecx)\n\t" + +" flds 64(%%edx)\n\t" +" fadds 76(%%edx)\n\t" +" fstps 64(%%ecx)\n\t" + +" flds 64(%%edx)\n\t" +" fsubs 76(%%edx)\n\t" +" fmuls 112(%%ebx)\n\t" +" fstps 76(%%ecx)\n\t" + +" flds 68(%%edx)\n\t" +" fadds 72(%%edx)\n\t" +" fstps 68(%%ecx)\n\t" + +" flds 68(%%edx)\n\t" +" fsubs 72(%%edx)\n\t" +" fmuls 116(%%ebx)\n\t" +" fstps 72(%%ecx)\n\t" + +" flds 80(%%edx)\n\t" +" fadds 92(%%edx)\n\t" +" fstps 80(%%ecx)\n\t" + +" flds 92(%%edx)\n\t" +" fsubs 80(%%edx)\n\t" +" fmuls 112(%%ebx)\n\t" +" fstps 92(%%ecx)\n\t" + +" flds 84(%%edx)\n\t" +" fadds 88(%%edx)\n\t" +" fstps 84(%%ecx)\n\t" + +" flds 88(%%edx)\n\t" +" fsubs 84(%%edx)\n\t" +" fmuls 116(%%ebx)\n\t" +" fstps 88(%%ecx)\n\t" + +" flds 96(%%edx)\n\t" +" fadds 108(%%edx)\n\t" +" fstps 96(%%ecx)\n\t" + +" flds 96(%%edx)\n\t" +" fsubs 108(%%edx)\n\t" +" fmuls 112(%%ebx)\n\t" +" fstps 108(%%ecx)\n\t" + +" flds 100(%%edx)\n\t" +" fadds 104(%%edx)\n\t" +" fstps 100(%%ecx)\n\t" + +" flds 100(%%edx)\n\t" +" fsubs 104(%%edx)\n\t" +" fmuls 116(%%ebx)\n\t" +" fstps 104(%%ecx)\n\t" + +" flds 112(%%edx)\n\t" +" fadds 124(%%edx)\n\t" +" fstps 112(%%ecx)\n\t" + +" flds 124(%%edx)\n\t" +" fsubs 112(%%edx)\n\t" +" fmuls 112(%%ebx)\n\t" +" fstps 124(%%ecx)\n\t" + +" flds 116(%%edx)\n\t" +" fadds 120(%%edx)\n\t" +" fstps 116(%%ecx)\n\t" + +" flds 120(%%edx)\n\t" +" fsubs 116(%%edx)\n\t" +" fmuls 116(%%ebx)\n\t" +" fstps 120(%%ecx)\n\t" /* Phase 5*/ -" flds 32(%%ecx)\n\t" -" fadds 36(%%ecx)\n\t" -" fstps 32(%%edx)\n\t" - -" flds 32(%%ecx)\n\t" -" fsubs 36(%%ecx)\n\t" -" fmuls 120(%%ebx)\n\t" -" fstps 36(%%edx)\n\t" - -" flds 44(%%ecx)\n\t" -" fsubs 40(%%ecx)\n\t" -" fmuls 120(%%ebx)\n\t" -" fsts 44(%%edx)\n\t" -" fadds 40(%%ecx)\n\t" -" fadds 44(%%ecx)\n\t" -" fstps 40(%%edx)\n\t" - -" flds 48(%%ecx)\n\t" -" fsubs 52(%%ecx)\n\t" -" fmuls 120(%%ebx)\n\t" - -" flds 60(%%ecx)\n\t" -" fsubs 56(%%ecx)\n\t" -" fmuls 120(%%ebx)\n\t" -" fld %%st(0)\n\t" -" fadds 56(%%ecx)\n\t" -" fadds 60(%%ecx)\n\t" -" fld %%st(0)\n\t" -" fadds 48(%%ecx)\n\t" -" fadds 52(%%ecx)\n\t" -" fstps 48(%%edx)\n\t" -" fadd %%st(2)\n\t" -" fstps 56(%%edx)\n\t" -" fsts 60(%%edx)\n\t" -" faddp %%st(1)\n\t" -" fstps 52(%%edx)\n\t" - -" flds 64(%%ecx)\n\t" -" fadds 68(%%ecx)\n\t" -" fstps 64(%%edx)\n\t" - -" flds 64(%%ecx)\n\t" -" fsubs 68(%%ecx)\n\t" -" fmuls 120(%%ebx)\n\t" -" fstps 68(%%edx)\n\t" - -" flds 76(%%ecx)\n\t" -" fsubs 72(%%ecx)\n\t" -" fmuls 120(%%ebx)\n\t" -" fsts 76(%%edx)\n\t" -" fadds 72(%%ecx)\n\t" -" fadds 76(%%ecx)\n\t" -" fstps 72(%%edx)\n\t" - -" flds 92(%%ecx)\n\t" -" fsubs 88(%%ecx)\n\t" -" fmuls 120(%%ebx)\n\t" -" fsts 92(%%edx)\n\t" -" fadds 92(%%ecx)\n\t" -" fadds 88(%%ecx)\n\t" -" fld %%st(0)\n\t" -" fadds 80(%%ecx)\n\t" -" fadds 84(%%ecx)\n\t" -" fstps 80(%%edx)\n\t" - -" flds 80(%%ecx)\n\t" -" fsubs 84(%%ecx)\n\t" -" fmuls 120(%%ebx)\n\t" -" fadd %%st(0), %%st(1)\n\t" -" fadds 92(%%edx)\n\t" -" fstps 84(%%edx)\n\t" -" fstps 88(%%edx)\n\t" - -" flds 96(%%ecx)\n\t" -" fadds 100(%%ecx)\n\t" -" fstps 96(%%edx)\n\t" - -" flds 96(%%ecx)\n\t" -" fsubs 100(%%ecx)\n\t" -" fmuls 120(%%ebx)\n\t" -" fstps 100(%%edx)\n\t" - -" flds 108(%%ecx)\n\t" -" fsubs 104(%%ecx)\n\t" -" fmuls 120(%%ebx)\n\t" -" fsts 108(%%edx)\n\t" -" fadds 104(%%ecx)\n\t" -" fadds 108(%%ecx)\n\t" -" fstps 104(%%edx)\n\t" - -" flds 124(%%ecx)\n\t" -" fsubs 120(%%ecx)\n\t" -" fmuls 120(%%ebx)\n\t" -" fsts 124(%%edx)\n\t" -" fadds 120(%%ecx)\n\t" -" fadds 124(%%ecx)\n\t" -" fld %%st(0)\n\t" -" fadds 112(%%ecx)\n\t" -" fadds 116(%%ecx)\n\t" -" fstps 112(%%edx)\n\t" - -" flds 112(%%ecx)\n\t" -" fsubs 116(%%ecx)\n\t" -" fmuls 120(%%ebx)\n\t" -" fadd %%st(0),%%st(1)\n\t" -" fadds 124(%%edx)\n\t" -" fstps 116(%%edx)\n\t" -" fstps 120(%%edx)\n\t" -" jnz .L01\n\t" +" flds 32(%%ecx)\n\t" +" fadds 36(%%ecx)\n\t" +" fstps 32(%%edx)\n\t" + +" flds 32(%%ecx)\n\t" +" fsubs 36(%%ecx)\n\t" +" fmuls 120(%%ebx)\n\t" +" fstps 36(%%edx)\n\t" + +" flds 44(%%ecx)\n\t" +" fsubs 40(%%ecx)\n\t" +" fmuls 120(%%ebx)\n\t" +" fsts 44(%%edx)\n\t" +" fadds 40(%%ecx)\n\t" +" fadds 44(%%ecx)\n\t" +" fstps 40(%%edx)\n\t" + +" flds 48(%%ecx)\n\t" +" fsubs 52(%%ecx)\n\t" +" fmuls 120(%%ebx)\n\t" + +" flds 60(%%ecx)\n\t" +" fsubs 56(%%ecx)\n\t" +" fmuls 120(%%ebx)\n\t" +" fld %%st(0)\n\t" +" fadds 56(%%ecx)\n\t" +" fadds 60(%%ecx)\n\t" +" fld %%st(0)\n\t" +" fadds 48(%%ecx)\n\t" +" fadds 52(%%ecx)\n\t" +" fstps 48(%%edx)\n\t" +" fadd %%st(2)\n\t" +" fstps 56(%%edx)\n\t" +" fsts 60(%%edx)\n\t" +" faddp %%st(1)\n\t" +" fstps 52(%%edx)\n\t" + +" flds 64(%%ecx)\n\t" +" fadds 68(%%ecx)\n\t" +" fstps 64(%%edx)\n\t" + +" flds 64(%%ecx)\n\t" +" fsubs 68(%%ecx)\n\t" +" fmuls 120(%%ebx)\n\t" +" fstps 68(%%edx)\n\t" + +" flds 76(%%ecx)\n\t" +" fsubs 72(%%ecx)\n\t" +" fmuls 120(%%ebx)\n\t" +" fsts 76(%%edx)\n\t" +" fadds 72(%%ecx)\n\t" +" fadds 76(%%ecx)\n\t" +" fstps 72(%%edx)\n\t" + +" flds 92(%%ecx)\n\t" +" fsubs 88(%%ecx)\n\t" +" fmuls 120(%%ebx)\n\t" +" fsts 92(%%edx)\n\t" +" fadds 92(%%ecx)\n\t" +" fadds 88(%%ecx)\n\t" +" fld %%st(0)\n\t" +" fadds 80(%%ecx)\n\t" +" fadds 84(%%ecx)\n\t" +" fstps 80(%%edx)\n\t" + +" flds 80(%%ecx)\n\t" +" fsubs 84(%%ecx)\n\t" +" fmuls 120(%%ebx)\n\t" +" fadd %%st(0), %%st(1)\n\t" +" fadds 92(%%edx)\n\t" +" fstps 84(%%edx)\n\t" +" fstps 88(%%edx)\n\t" + +" flds 96(%%ecx)\n\t" +" fadds 100(%%ecx)\n\t" +" fstps 96(%%edx)\n\t" + +" flds 96(%%ecx)\n\t" +" fsubs 100(%%ecx)\n\t" +" fmuls 120(%%ebx)\n\t" +" fstps 100(%%edx)\n\t" + +" flds 108(%%ecx)\n\t" +" fsubs 104(%%ecx)\n\t" +" fmuls 120(%%ebx)\n\t" +" fsts 108(%%edx)\n\t" +" fadds 104(%%ecx)\n\t" +" fadds 108(%%ecx)\n\t" +" fstps 104(%%edx)\n\t" + +" flds 124(%%ecx)\n\t" +" fsubs 120(%%ecx)\n\t" +" fmuls 120(%%ebx)\n\t" +" fsts 124(%%edx)\n\t" +" fadds 120(%%ecx)\n\t" +" fadds 124(%%ecx)\n\t" +" fld %%st(0)\n\t" +" fadds 112(%%ecx)\n\t" +" fadds 116(%%ecx)\n\t" +" fstps 112(%%edx)\n\t" + +" flds 112(%%ecx)\n\t" +" fsubs 116(%%ecx)\n\t" +" fmuls 120(%%ebx)\n\t" +" fadd %%st(0),%%st(1)\n\t" +" fadds 124(%%edx)\n\t" +" fstps 116(%%edx)\n\t" +" fstps 120(%%edx)\n\t" +" jnz .L01\n\t" /* Phase 6*/ -" flds (%%ecx)\n\t" -" fadds 4(%%ecx)\n\t" -" fstps 1024(%%esi)\n\t" - -" flds (%%ecx)\n\t" -" fsubs 4(%%ecx)\n\t" -" fmuls 120(%%ebx)\n\t" -" fsts (%%esi)\n\t" -" fstps (%%edi)\n\t" - -" flds 12(%%ecx)\n\t" -" fsubs 8(%%ecx)\n\t" -" fmuls 120(%%ebx)\n\t" -" fsts 512(%%edi)\n\t" -" fadds 12(%%ecx)\n\t" -" fadds 8(%%ecx)\n\t" -" fstps 512(%%esi)\n\t" - -" flds 16(%%ecx)\n\t" -" fsubs 20(%%ecx)\n\t" -" fmuls 120(%%ebx)\n\t" - -" flds 28(%%ecx)\n\t" -" fsubs 24(%%ecx)\n\t" -" fmuls 120(%%ebx)\n\t" -" fsts 768(%%edi)\n\t" -" fld %%st(0)\n\t" -" fadds 24(%%ecx)\n\t" -" fadds 28(%%ecx)\n\t" -" fld %%st(0)\n\t" -" fadds 16(%%ecx)\n\t" -" fadds 20(%%ecx)\n\t" -" fstps 768(%%esi)\n\t" -" fadd %%st(2)\n\t" -" fstps 256(%%esi)\n\t" -" faddp %%st(1)\n\t" -" fstps 256(%%edi)\n\t" +" flds (%%ecx)\n\t" +" fadds 4(%%ecx)\n\t" +" fstps 1024(%%esi)\n\t" + +" flds (%%ecx)\n\t" +" fsubs 4(%%ecx)\n\t" +" fmuls 120(%%ebx)\n\t" +" fsts (%%esi)\n\t" +" fstps (%%edi)\n\t" + +" flds 12(%%ecx)\n\t" +" fsubs 8(%%ecx)\n\t" +" fmuls 120(%%ebx)\n\t" +" fsts 512(%%edi)\n\t" +" fadds 12(%%ecx)\n\t" +" fadds 8(%%ecx)\n\t" +" fstps 512(%%esi)\n\t" + +" flds 16(%%ecx)\n\t" +" fsubs 20(%%ecx)\n\t" +" fmuls 120(%%ebx)\n\t" + +" flds 28(%%ecx)\n\t" +" fsubs 24(%%ecx)\n\t" +" fmuls 120(%%ebx)\n\t" +" fsts 768(%%edi)\n\t" +" fld %%st(0)\n\t" +" fadds 24(%%ecx)\n\t" +" fadds 28(%%ecx)\n\t" +" fld %%st(0)\n\t" +" fadds 16(%%ecx)\n\t" +" fadds 20(%%ecx)\n\t" +" fstps 768(%%esi)\n\t" +" fadd %%st(2)\n\t" +" fstps 256(%%esi)\n\t" +" faddp %%st(1)\n\t" +" fstps 256(%%edi)\n\t" /* Phase 7*/ -" flds 32(%%edx)\n\t" -" fadds 48(%%edx)\n\t" -" fstps 896(%%esi)\n\t" - -" flds 48(%%edx)\n\t" -" fadds 40(%%edx)\n\t" -" fstps 640(%%esi)\n\t" - -" flds 40(%%edx)\n\t" -" fadds 56(%%edx)\n\t" -" fstps 384(%%esi)\n\t" - -" flds 56(%%edx)\n\t" -" fadds 36(%%edx)\n\t" -" fstps 128(%%esi)\n\t" - -" flds 36(%%edx)\n\t" -" fadds 52(%%edx)\n\t" -" fstps 128(%%edi)\n\t" - -" flds 52(%%edx)\n\t" -" fadds 44(%%edx)\n\t" -" fstps 384(%%edi)\n\t" - -" flds 60(%%edx)\n\t" -" fsts 896(%%edi)\n\t" -" fadds 44(%%edx)\n\t" -" fstps 640(%%edi)\n\t" - -" flds 96(%%edx)\n\t" -" fadds 112(%%edx)\n\t" -" fld %%st(0)\n\t" -" fadds 64(%%edx)\n\t" -" fstps 960(%%esi)\n\t" -" fadds 80(%%edx)\n\t" -" fstps 832(%%esi)\n\t" - -" flds 112(%%edx)\n\t" -" fadds 104(%%edx)\n\t" -" fld %%st(0)\n\t" -" fadds 80(%%edx)\n\t" -" fstps 704(%%esi)\n\t" -" fadds 72(%%edx)\n\t" -" fstps 576(%%esi)\n\t" - -" flds 104(%%edx)\n\t" -" fadds 120(%%edx)\n\t" -" fld %%st(0)\n\t" -" fadds 72(%%edx)\n\t" -" fstps 448(%%esi)\n\t" -" fadds 88(%%edx)\n\t" -" fstps 320(%%esi)\n\t" - -" flds 120(%%edx)\n\t" -" fadds 100(%%edx)\n\t" -" fld %%st(0)\n\t" -" fadds 88(%%edx)\n\t" -" fstps 192(%%esi)\n\t" -" fadds 68(%%edx)\n\t" -" fstps 64(%%esi)\n\t" - -" flds 100(%%edx)\n\t" -" fadds 116(%%edx)\n\t" -" fld %%st(0)\n\t" -" fadds 68(%%edx)\n\t" -" fstps 64(%%edi)\n\t" -" fadds 84(%%edx)\n\t" -" fstps 192(%%edi)\n\t" - -" flds 116(%%edx)\n\t" -" fadds 108(%%edx)\n\t" -" fld %%st(0)\n\t" -" fadds 84(%%edx)\n\t" -" fstps 320(%%edi)\n\t" -" fadds 76(%%edx)\n\t" -" fstps 448(%%edi)\n\t" - -" flds 108(%%edx)\n\t" -" fadds 124(%%edx)\n\t" -" fld %%st(0)\n\t" -" fadds 76(%%edx)\n\t" -" fstps 576(%%edi)\n\t" -" fadds 92(%%edx)\n\t" -" fstps 704(%%edi)\n\t" - -" flds 124(%%edx)\n\t" -" fsts 960(%%edi)\n\t" -" fadds 92(%%edx)\n\t" -" fstps 832(%%edi)\n\t" -" jmp .L_bye\n\t" +" flds 32(%%edx)\n\t" +" fadds 48(%%edx)\n\t" +" fstps 896(%%esi)\n\t" + +" flds 48(%%edx)\n\t" +" fadds 40(%%edx)\n\t" +" fstps 640(%%esi)\n\t" + +" flds 40(%%edx)\n\t" +" fadds 56(%%edx)\n\t" +" fstps 384(%%esi)\n\t" + +" flds 56(%%edx)\n\t" +" fadds 36(%%edx)\n\t" +" fstps 128(%%esi)\n\t" + +" flds 36(%%edx)\n\t" +" fadds 52(%%edx)\n\t" +" fstps 128(%%edi)\n\t" + +" flds 52(%%edx)\n\t" +" fadds 44(%%edx)\n\t" +" fstps 384(%%edi)\n\t" + +" flds 60(%%edx)\n\t" +" fsts 896(%%edi)\n\t" +" fadds 44(%%edx)\n\t" +" fstps 640(%%edi)\n\t" + +" flds 96(%%edx)\n\t" +" fadds 112(%%edx)\n\t" +" fld %%st(0)\n\t" +" fadds 64(%%edx)\n\t" +" fstps 960(%%esi)\n\t" +" fadds 80(%%edx)\n\t" +" fstps 832(%%esi)\n\t" + +" flds 112(%%edx)\n\t" +" fadds 104(%%edx)\n\t" +" fld %%st(0)\n\t" +" fadds 80(%%edx)\n\t" +" fstps 704(%%esi)\n\t" +" fadds 72(%%edx)\n\t" +" fstps 576(%%esi)\n\t" + +" flds 104(%%edx)\n\t" +" fadds 120(%%edx)\n\t" +" fld %%st(0)\n\t" +" fadds 72(%%edx)\n\t" +" fstps 448(%%esi)\n\t" +" fadds 88(%%edx)\n\t" +" fstps 320(%%esi)\n\t" + +" flds 120(%%edx)\n\t" +" fadds 100(%%edx)\n\t" +" fld %%st(0)\n\t" +" fadds 88(%%edx)\n\t" +" fstps 192(%%esi)\n\t" +" fadds 68(%%edx)\n\t" +" fstps 64(%%esi)\n\t" + +" flds 100(%%edx)\n\t" +" fadds 116(%%edx)\n\t" +" fld %%st(0)\n\t" +" fadds 68(%%edx)\n\t" +" fstps 64(%%edi)\n\t" +" fadds 84(%%edx)\n\t" +" fstps 192(%%edi)\n\t" + +" flds 116(%%edx)\n\t" +" fadds 108(%%edx)\n\t" +" fld %%st(0)\n\t" +" fadds 84(%%edx)\n\t" +" fstps 320(%%edi)\n\t" +" fadds 76(%%edx)\n\t" +" fstps 448(%%edi)\n\t" + +" flds 108(%%edx)\n\t" +" fadds 124(%%edx)\n\t" +" fld %%st(0)\n\t" +" fadds 76(%%edx)\n\t" +" fstps 576(%%edi)\n\t" +" fadds 92(%%edx)\n\t" +" fstps 704(%%edi)\n\t" + +" flds 124(%%edx)\n\t" +" fsts 960(%%edi)\n\t" +" fadds 92(%%edx)\n\t" +" fstps 832(%%edi)\n\t" +" jmp .L_bye\n\t" ".L01:\n\t" /* Phase 8*/ -" flds (%%ecx)\n\t" -" fadds 4(%%ecx)\n\t" -" fistp 512(%%esi)\n\t" - -" flds (%%ecx)\n\t" -" fsubs 4(%%ecx)\n\t" -" fmuls 120(%%ebx)\n\t" - -" fistp (%%esi)\n\t" - - -" flds 12(%%ecx)\n\t" -" fsubs 8(%%ecx)\n\t" -" fmuls 120(%%ebx)\n\t" -" fist 256(%%edi)\n\t" -" fadds 12(%%ecx)\n\t" -" fadds 8(%%ecx)\n\t" -" fistp 256(%%esi)\n\t" - -" flds 16(%%ecx)\n\t" -" fsubs 20(%%ecx)\n\t" -" fmuls 120(%%ebx)\n\t" - -" flds 28(%%ecx)\n\t" -" fsubs 24(%%ecx)\n\t" -" fmuls 120(%%ebx)\n\t" -" fist 384(%%edi)\n\t" -" fld %%st(0)\n\t" -" fadds 24(%%ecx)\n\t" -" fadds 28(%%ecx)\n\t" -" fld %%st(0)\n\t" -" fadds 16(%%ecx)\n\t" -" fadds 20(%%ecx)\n\t" -" fistp 384(%%esi)\n\t" -" fadd %%st(2)\n\t" -" fistp 128(%%esi)\n\t" -" faddp %%st(1)\n\t" -" fistp 128(%%edi)\n\t" +" flds (%%ecx)\n\t" +" fadds 4(%%ecx)\n\t" +" fistp 512(%%esi)\n\t" + +" flds (%%ecx)\n\t" +" fsubs 4(%%ecx)\n\t" +" fmuls 120(%%ebx)\n\t" + +" fistp (%%esi)\n\t" + + +" flds 12(%%ecx)\n\t" +" fsubs 8(%%ecx)\n\t" +" fmuls 120(%%ebx)\n\t" +" fist 256(%%edi)\n\t" +" fadds 12(%%ecx)\n\t" +" fadds 8(%%ecx)\n\t" +" fistp 256(%%esi)\n\t" + +" flds 16(%%ecx)\n\t" +" fsubs 20(%%ecx)\n\t" +" fmuls 120(%%ebx)\n\t" + +" flds 28(%%ecx)\n\t" +" fsubs 24(%%ecx)\n\t" +" fmuls 120(%%ebx)\n\t" +" fist 384(%%edi)\n\t" +" fld %%st(0)\n\t" +" fadds 24(%%ecx)\n\t" +" fadds 28(%%ecx)\n\t" +" fld %%st(0)\n\t" +" fadds 16(%%ecx)\n\t" +" fadds 20(%%ecx)\n\t" +" fistp 384(%%esi)\n\t" +" fadd %%st(2)\n\t" +" fistp 128(%%esi)\n\t" +" faddp %%st(1)\n\t" +" fistp 128(%%edi)\n\t" /* Phase 9*/ -" flds 32(%%edx)\n\t" -" fadds 48(%%edx)\n\t" -" fistp 448(%%esi)\n\t" - -" flds 48(%%edx)\n\t" -" fadds 40(%%edx)\n\t" -" fistp 320(%%esi)\n\t" - -" flds 40(%%edx)\n\t" -" fadds 56(%%edx)\n\t" -" fistp 192(%%esi)\n\t" - -" flds 56(%%edx)\n\t" -" fadds 36(%%edx)\n\t" -" fistp 64(%%esi)\n\t" - -" flds 36(%%edx)\n\t" -" fadds 52(%%edx)\n\t" -" fistp 64(%%edi)\n\t" - -" flds 52(%%edx)\n\t" -" fadds 44(%%edx)\n\t" -" fistp 192(%%edi)\n\t" - -" flds 60(%%edx)\n\t" -" fist 448(%%edi)\n\t" -" fadds 44(%%edx)\n\t" -" fistp 320(%%edi)\n\t" - -" flds 96(%%edx)\n\t" -" fadds 112(%%edx)\n\t" -" fld %%st(0)\n\t" -" fadds 64(%%edx)\n\t" -" fistp 480(%%esi)\n\t" -" fadds 80(%%edx)\n\t" -" fistp 416(%%esi)\n\t" - -" flds 112(%%edx)\n\t" -" fadds 104(%%edx)\n\t" -" fld %%st(0)\n\t" -" fadds 80(%%edx)\n\t" -" fistp 352(%%esi)\n\t" -" fadds 72(%%edx)\n\t" -" fistp 288(%%esi)\n\t" - -" flds 104(%%edx)\n\t" -" fadds 120(%%edx)\n\t" -" fld %%st(0)\n\t" -" fadds 72(%%edx)\n\t" -" fistp 224(%%esi)\n\t" -" fadds 88(%%edx)\n\t" -" fistp 160(%%esi)\n\t" - -" flds 120(%%edx)\n\t" -" fadds 100(%%edx)\n\t" -" fld %%st(0)\n\t" -" fadds 88(%%edx)\n\t" -" fistp 96(%%esi)\n\t" -" fadds 68(%%edx)\n\t" -" fistp 32(%%esi)\n\t" - -" flds 100(%%edx)\n\t" -" fadds 116(%%edx)\n\t" -" fld %%st(0)\n\t" -" fadds 68(%%edx)\n\t" -" fistp 32(%%edi)\n\t" -" fadds 84(%%edx)\n\t" -" fistp 96(%%edi)\n\t" - -" flds 116(%%edx)\n\t" -" fadds 108(%%edx)\n\t" -" fld %%st(0)\n\t" -" fadds 84(%%edx)\n\t" -" fistp 160(%%edi)\n\t" -" fadds 76(%%edx)\n\t" -" fistp 224(%%edi)\n\t" - -" flds 108(%%edx)\n\t" -" fadds 124(%%edx)\n\t" -" fld %%st(0)\n\t" -" fadds 76(%%edx)\n\t" -" fistp 288(%%edi)\n\t" -" fadds 92(%%edx)\n\t" -" fistp 352(%%edi)\n\t" - -" flds 124(%%edx)\n\t" -" fist 480(%%edi)\n\t" -" fadds 92(%%edx)\n\t" -" fistp 416(%%edi)\n\t" -" movsw\n\t" +" flds 32(%%edx)\n\t" +" fadds 48(%%edx)\n\t" +" fistp 448(%%esi)\n\t" + +" flds 48(%%edx)\n\t" +" fadds 40(%%edx)\n\t" +" fistp 320(%%esi)\n\t" + +" flds 40(%%edx)\n\t" +" fadds 56(%%edx)\n\t" +" fistp 192(%%esi)\n\t" + +" flds 56(%%edx)\n\t" +" fadds 36(%%edx)\n\t" +" fistp 64(%%esi)\n\t" + +" flds 36(%%edx)\n\t" +" fadds 52(%%edx)\n\t" +" fistp 64(%%edi)\n\t" + +" flds 52(%%edx)\n\t" +" fadds 44(%%edx)\n\t" +" fistp 192(%%edi)\n\t" + +" flds 60(%%edx)\n\t" +" fist 448(%%edi)\n\t" +" fadds 44(%%edx)\n\t" +" fistp 320(%%edi)\n\t" + +" flds 96(%%edx)\n\t" +" fadds 112(%%edx)\n\t" +" fld %%st(0)\n\t" +" fadds 64(%%edx)\n\t" +" fistp 480(%%esi)\n\t" +" fadds 80(%%edx)\n\t" +" fistp 416(%%esi)\n\t" + +" flds 112(%%edx)\n\t" +" fadds 104(%%edx)\n\t" +" fld %%st(0)\n\t" +" fadds 80(%%edx)\n\t" +" fistp 352(%%esi)\n\t" +" fadds 72(%%edx)\n\t" +" fistp 288(%%esi)\n\t" + +" flds 104(%%edx)\n\t" +" fadds 120(%%edx)\n\t" +" fld %%st(0)\n\t" +" fadds 72(%%edx)\n\t" +" fistp 224(%%esi)\n\t" +" fadds 88(%%edx)\n\t" +" fistp 160(%%esi)\n\t" + +" flds 120(%%edx)\n\t" +" fadds 100(%%edx)\n\t" +" fld %%st(0)\n\t" +" fadds 88(%%edx)\n\t" +" fistp 96(%%esi)\n\t" +" fadds 68(%%edx)\n\t" +" fistp 32(%%esi)\n\t" + +" flds 100(%%edx)\n\t" +" fadds 116(%%edx)\n\t" +" fld %%st(0)\n\t" +" fadds 68(%%edx)\n\t" +" fistp 32(%%edi)\n\t" +" fadds 84(%%edx)\n\t" +" fistp 96(%%edi)\n\t" + +" flds 116(%%edx)\n\t" +" fadds 108(%%edx)\n\t" +" fld %%st(0)\n\t" +" fadds 84(%%edx)\n\t" +" fistp 160(%%edi)\n\t" +" fadds 76(%%edx)\n\t" +" fistp 224(%%edi)\n\t" + +" flds 108(%%edx)\n\t" +" fadds 124(%%edx)\n\t" +" fld %%st(0)\n\t" +" fadds 76(%%edx)\n\t" +" fistp 288(%%edi)\n\t" +" fadds 92(%%edx)\n\t" +" fistp 352(%%edi)\n\t" + +" flds 124(%%edx)\n\t" +" fist 480(%%edi)\n\t" +" fadds 92(%%edx)\n\t" +" fistp 416(%%edi)\n\t" +" movsw\n\t" ".L_bye:" - : - :"m"(a),"m"(b),"m"(c),"m"(tmp[0]) - :"memory","%eax","%ebx","%ecx","%edx","%esi","%edi"); + : + :"m"(a),"m"(b),"m"(c),"m"(tmp[0]) + :"memory","%eax","%ebx","%ecx","%edx","%esi","%edi"); } diff --git a/mp3lib/decod386.c b/mp3lib/decod386.c index a2cfb56ce8..82657aedc3 100644 --- a/mp3lib/decod386.c +++ b/mp3lib/decod386.c @@ -22,10 +22,10 @@ #if 0 /* old WRITE_SAMPLE */ /* is portable */ -#define WRITE_SAMPLE(samples,sum,clip) { \ - if( (sum) > 32767.0) { *(samples) = 0x7fff; (clip)++; } \ +#define WRITE_SAMPLE(samples,sum,clip) { \ + if( (sum) > 32767.0) { *(samples) = 0x7fff; (clip)++; } \ else if( (sum) < -32768.0) { *(samples) = -0x8000; (clip)++; }\ - else { *(samples) = sum; } \ + else { *(samples) = sum; } \ } #else /* new WRITE_SAMPLE */ @@ -55,9 +55,9 @@ * it doesn't work for you, just enable the "old WRITE_SAMPLE" macro. */ #if HAVE_BIGENDIAN -#define MANTISSA_OFFSET 1 +#define MANTISSA_OFFSET 1 #else -#define MANTISSA_OFFSET 0 +#define MANTISSA_OFFSET 0 #endif /* sizeof(int) == 4 */ diff --git a/mp3lib/decode_i586.c b/mp3lib/decode_i586.c index d4cea638c1..83933d7ad4 100644 --- a/mp3lib/decode_i586.c +++ b/mp3lib/decode_i586.c @@ -311,8 +311,8 @@ int synth_1to1_pent(real *bandPtr, int channel, short *samples) " jnz .L68\n\t" " movl %%edi,%%eax\n\t" " movl "MANGLE(saved_ebp)",%%ebp\n\t" - :"=a"(retval) - :"m"(bandPtr),"m"(channel),"m"(samples),"m"(tmp[0]) - :"memory","%edi","%esi","%ebx","%ecx","%edx"); + :"=a"(retval) + :"m"(bandPtr),"m"(channel),"m"(samples),"m"(tmp[0]) + :"memory","%edi","%esi","%ebx","%ecx","%edx"); return retval; } diff --git a/mp3lib/decode_mmx.c b/mp3lib/decode_mmx.c index 211630f219..3027657421 100644 --- a/mp3lib/decode_mmx.c +++ b/mp3lib/decode_mmx.c @@ -15,37 +15,37 @@ static const unsigned long long attribute_used __attribute__((aligned(8))) null_ static const unsigned long long attribute_used __attribute__((aligned(8))) one_null = 0xffff0000ffff0000ULL; const unsigned int __attribute__((aligned(16))) costab_mmx[] = { - 1056974725, - 1057056395, - 1057223771, - 1057485416, - 1057855544, - 1058356026, - 1059019886, - 1059897405, - 1061067246, - 1062657950, - 1064892987, - 1066774581, - 1069414683, - 1073984175, - 1079645762, - 1092815430, - 1057005197, - 1057342072, - 1058087743, - 1059427869, - 1061799040, - 1065862217, - 1071413542, - 1084439708, - 1057128951, - 1058664893, - 1063675095, - 1076102863, - 1057655764, - 1067924853, - 1060439283, + 1056974725, + 1057056395, + 1057223771, + 1057485416, + 1057855544, + 1058356026, + 1059019886, + 1059897405, + 1061067246, + 1062657950, + 1064892987, + 1066774581, + 1069414683, + 1073984175, + 1079645762, + 1092815430, + 1057005197, + 1057342072, + 1058087743, + 1059427869, + 1061799040, + 1065862217, + 1071413542, + 1084439708, + 1057128951, + 1058664893, + 1063675095, + 1076102863, + 1057655764, + 1067924853, + 1060439283, }; /** @@ -55,134 +55,134 @@ const unsigned int __attribute__((aligned(16))) costab_mmx[] = */ static const short __attribute__((aligned(8))) mp3lib_decwins[] = { - 0, 7, 54, 114, 510, 1288, 1644, 9372, - 18760, -9373, 1644, -1289, 510, -115, 54, -8, - 0, 7, 54, 114, 510, 1288, 1644, 9372, - 18760, -9373, 1644, -1289, 510, -115, 54, -8, - 0, 7, 55, 129, 500, 1379, 1490, 9834, - 18748, -8910, 1784, -1197, 516, -101, 52, -7, - 0, 7, 55, 129, 500, 1379, 1490, 9834, - 18748, -8910, 1784, -1197, 516, -101, 52, -7, - 0, 8, 56, 145, 488, 1469, 1322, 10294, - 18714, -8448, 1910, -1107, 520, -87, 51, -6, - 0, 8, 56, 145, 488, 1469, 1322, 10294, - 18714, -8448, 1910, -1107, 520, -87, 51, -6, - 0, 9, 57, 161, 474, 1559, 1141, 10751, - 18658, -7987, 2023, -1016, 522, -74, 49, -6, - 0, 9, 57, 161, 474, 1559, 1141, 10751, - 18658, -7987, 2023, -1016, 522, -74, 49, -6, - 0, 10, 57, 177, 456, 1647, 944, 11205, - 18579, -7528, 2123, -927, 522, -61, 48, -5, - 0, 10, 57, 177, 456, 1647, 944, 11205, - 18579, -7528, 2123, -927, 522, -61, 48, -5, - 0, 11, 57, 194, 435, 1733, 734, 11654, - 18477, -7073, 2210, -838, 519, -50, 46, -5, - 0, 11, 57, 194, 435, 1733, 734, 11654, - 18477, -7073, 2210, -838, 519, -50, 46, -5, - 0, 12, 57, 212, 411, 1817, 510, 12097, - 18354, -6621, 2285, -751, 515, -39, 44, -4, - 0, 12, 57, 212, 411, 1817, 510, 12097, - 18354, -6621, 2285, -751, 515, -39, 44, -4, - 0, 13, 57, 229, 384, 1899, 271, 12534, - 18209, -6174, 2348, -666, 508, -28, 43, -4, - 0, 13, 57, 229, 384, 1899, 271, 12534, - 18209, -6174, 2348, -666, 508, -28, 43, -4, - 0, 14, 56, 247, 354, 1977, 18, 12963, - 18043, -5733, 2398, -583, 501, -18, 41, -4, - 0, 14, 56, 247, 354, 1977, 18, 12963, - 18043, -5733, 2398, -583, 501, -18, 41, -4, - 0, 15, 56, 266, 320, 2052, -249, 13383, - 17855, -5298, 2438, -502, 491, -9, 39, -3, - 0, 15, 56, 266, 320, 2052, -249, 13383, - 17855, -5298, 2438, -502, 491, -9, 39, -3, - 0, 17, 54, 284, 283, 2122, -530, 13794, - 17648, -4870, 2466, -423, 480, -1, 37, -3, - 0, 17, 54, 284, 283, 2122, -530, 13794, - 17648, -4870, 2466, -423, 480, -1, 37, -3, - 0, 18, 52, 302, 243, 2188, -825, 14194, - 17420, -4450, 2484, -347, 468, 7, 35, -3, - 0, 18, 52, 302, 243, 2188, -825, 14194, - 17420, -4450, 2484, -347, 468, 7, 35, -3, - 0, 19, 50, 320, 199, 2249, -1133, 14583, - 17173, -4039, 2492, -274, 455, 14, 33, -2, - 0, 19, 50, 320, 199, 2249, -1133, 14583, - 17173, -4039, 2492, -274, 455, 14, 33, -2, - -1, 21, 48, 339, 152, 2304, -1454, 14959, - 16908, -3637, 2490, -204, 440, 20, 32, -2, - -1, 21, 48, 339, 152, 2304, -1454, 14959, - 16908, -3637, 2490, -204, 440, 20, 32, -2, - -1, 22, 45, 357, 101, 2354, -1788, 15322, - 16624, -3245, 2479, -137, 425, 26, 30, -2, - -1, 22, 45, 357, 101, 2354, -1788, 15322, - 16624, -3245, 2479, -137, 425, 26, 30, -2, - -1, 24, 41, 374, 47, 2396, -2135, 15671, - 16323, -2864, 2460, -72, 409, 31, 28, -2, - -1, 24, 41, 374, 47, 2396, -2135, 15671, - 16323, -2864, 2460, -72, 409, 31, 28, -2, - -1, 26, 37, 391, -11, 2431, -2493, 16004, - 16005, -2494, 2432, -12, 392, 36, 26, -2, - -1, 26, 37, 391, -11, 2431, -2493, 16004, - 16005, -2494, 2432, -12, 392, 36, 26, -2, - -2, -28, 31, -409, -72, -2460, -2864, -16323, - 15671, 2135, 2396, -47, 374, -41, 24, 1, - -2, -28, 31, -409, -72, -2460, -2864, -16323, - 15671, 2135, 2396, -47, 374, -41, 24, 1, - -2, -30, 26, -425, -137, -2479, -3245, -16624, - 15322, 1788, 2354, -101, 357, -45, 22, 1, - -2, -30, 26, -425, -137, -2479, -3245, -16624, - 15322, 1788, 2354, -101, 357, -45, 22, 1, - -2, -32, 20, -440, -204, -2490, -3637, -16908, - 14959, 1454, 2304, -152, 339, -48, 21, 1, - -2, -32, 20, -440, -204, -2490, -3637, -16908, - 14959, 1454, 2304, -152, 339, -48, 21, 1, - -2, -33, 14, -455, -274, -2492, -4039, -17173, - 14583, 1133, 2249, -199, 320, -50, 19, 0, - -2, -33, 14, -455, -274, -2492, -4039, -17173, - 14583, 1133, 2249, -199, 320, -50, 19, 0, - -3, -35, 7, -468, -347, -2484, -4450, -17420, - 14194, 825, 2188, -243, 302, -52, 18, 0, - -3, -35, 7, -468, -347, -2484, -4450, -17420, - 14194, 825, 2188, -243, 302, -52, 18, 0, - -3, -37, -1, -480, -423, -2466, -4870, -17648, - 13794, 530, 2122, -283, 284, -54, 17, 0, - -3, -37, -1, -480, -423, -2466, -4870, -17648, - 13794, 530, 2122, -283, 284, -54, 17, 0, - -3, -39, -9, -491, -502, -2438, -5298, -17855, - 13383, 249, 2052, -320, 266, -56, 15, 0, - -3, -39, -9, -491, -502, -2438, -5298, -17855, - 13383, 249, 2052, -320, 266, -56, 15, 0, - -4, -41, -18, -501, -583, -2398, -5733, -18043, - 12963, -18, 1977, -354, 247, -56, 14, 0, - -4, -41, -18, -501, -583, -2398, -5733, -18043, - 12963, -18, 1977, -354, 247, -56, 14, 0, - -4, -43, -28, -508, -666, -2348, -6174, -18209, - 12534, -271, 1899, -384, 229, -57, 13, 0, - -4, -43, -28, -508, -666, -2348, -6174, -18209, - 12534, -271, 1899, -384, 229, -57, 13, 0, - -4, -44, -39, -515, -751, -2285, -6621, -18354, - 12097, -510, 1817, -411, 212, -57, 12, 0, - -4, -44, -39, -515, -751, -2285, -6621, -18354, - 12097, -510, 1817, -411, 212, -57, 12, 0, - -5, -46, -50, -519, -838, -2210, -7073, -18477, - 11654, -734, 1733, -435, 194, -57, 11, 0, - -5, -46, -50, -519, -838, -2210, -7073, -18477, - 11654, -734, 1733, -435, 194, -57, 11, 0, - -5, -48, -61, -522, -927, -2123, -7528, -18579, - 11205, -944, 1647, -456, 177, -57, 10, 0, - -5, -48, -61, -522, -927, -2123, -7528, -18579, - 11205, -944, 1647, -456, 177, -57, 10, 0, - -6, -49, -74, -522, -1016, -2023, -7987, -18658, - 10751, -1141, 1559, -474, 161, -57, 9, 0, - -6, -49, -74, -522, -1016, -2023, -7987, -18658, - 10751, -1141, 1559, -474, 161, -57, 9, 0, - -6, -51, -87, -520, -1107, -1910, -8448, -18714, - 10294, -1322, 1469, -488, 145, -56, 8, 0, - -6, -51, -87, -520, -1107, -1910, -8448, -18714, - 10294, -1322, 1469, -488, 145, -56, 8, 0, - -7, -52, -101, -516, -1197, -1784, -8910, -18748, - 9834, -1490, 1379, -500, 129, -55, 7, 0, - -7, -52, -101, -516, -1197, -1784, -8910, -18748, - 9834, -1490, 1379, -500, 129, -55, 7, 0, + 0, 7, 54, 114, 510, 1288, 1644, 9372, + 18760, -9373, 1644, -1289, 510, -115, 54, -8, + 0, 7, 54, 114, 510, 1288, 1644, 9372, + 18760, -9373, 1644, -1289, 510, -115, 54, -8, + 0, 7, 55, 129, 500, 1379, 1490, 9834, + 18748, -8910, 1784, -1197, 516, -101, 52, -7, + 0, 7, 55, 129, 500, 1379, 1490, 9834, + 18748, -8910, 1784, -1197, 516, -101, 52, -7, + 0, 8, 56, 145, 488, 1469, 1322, 10294, + 18714, -8448, 1910, -1107, 520, -87, 51, -6, + 0, 8, 56, 145, 488, 1469, 1322, 10294, + 18714, -8448, 1910, -1107, 520, -87, 51, -6, + 0, 9, 57, 161, 474, 1559, 1141, 10751, + 18658, -7987, 2023, -1016, 522, -74, 49, -6, + 0, 9, 57, 161, 474, 1559, 1141, 10751, + 18658, -7987, 2023, -1016, 522, -74, 49, -6, + 0, 10, 57, 177, 456, 1647, 944, 11205, + 18579, -7528, 2123, -927, 522, -61, 48, -5, + 0, 10, 57, 177, 456, 1647, 944, 11205, + 18579, -7528, 2123, -927, 522, -61, 48, -5, + 0, 11, 57, 194, 435, 1733, 734, 11654, + 18477, -7073, 2210, -838, 519, -50, 46, -5, + 0, 11, 57, 194, 435, 1733, 734, 11654, + 18477, -7073, 2210, -838, 519, -50, 46, -5, + 0, 12, 57, 212, 411, 1817, 510, 12097, + 18354, -6621, 2285, -751, 515, -39, 44, -4, + 0, 12, 57, 212, 411, 1817, 510, 12097, + 18354, -6621, 2285, -751, 515, -39, 44, -4, + 0, 13, 57, 229, 384, 1899, 271, 12534, + 18209, -6174, 2348, -666, 508, -28, 43, -4, + 0, 13, 57, 229, 384, 1899, 271, 12534, + 18209, -6174, 2348, -666, 508, -28, 43, -4, + 0, 14, 56, 247, 354, 1977, 18, 12963, + 18043, -5733, 2398, -583, 501, -18, 41, -4, + 0, 14, 56, 247, 354, 1977, 18, 12963, + 18043, -5733, 2398, -583, 501, -18, 41, -4, + 0, 15, 56, 266, 320, 2052, -249, 13383, + 17855, -5298, 2438, -502, 491, -9, 39, -3, + 0, 15, 56, 266, 320, 2052, -249, 13383, + 17855, -5298, 2438, -502, 491, -9, 39, -3, + 0, 17, 54, 284, 283, 2122, -530, 13794, + 17648, -4870, 2466, -423, 480, -1, 37, -3, + 0, 17, 54, 284, 283, 2122, -530, 13794, + 17648, -4870, 2466, -423, 480, -1, 37, -3, + 0, 18, 52, 302, 243, 2188, -825, 14194, + 17420, -4450, 2484, -347, 468, 7, 35, -3, + 0, 18, 52, 302, 243, 2188, -825, 14194, + 17420, -4450, 2484, -347, 468, 7, 35, -3, + 0, 19, 50, 320, 199, 2249, -1133, 14583, + 17173, -4039, 2492, -274, 455, 14, 33, -2, + 0, 19, 50, 320, 199, 2249, -1133, 14583, + 17173, -4039, 2492, -274, 455, 14, 33, -2, + -1, 21, 48, 339, 152, 2304, -1454, 14959, + 16908, -3637, 2490, -204, 440, 20, 32, -2, + -1, 21, 48, 339, 152, 2304, -1454, 14959, + 16908, -3637, 2490, -204, 440, 20, 32, -2, + -1, 22, 45, 357, 101, 2354, -1788, 15322, + 16624, -3245, 2479, -137, 425, 26, 30, -2, + -1, 22, 45, 357, 101, 2354, -1788, 15322, + 16624, -3245, 2479, -137, 425, 26, 30, -2, + -1, 24, 41, 374, 47, 2396, -2135, 15671, + 16323, -2864, 2460, -72, 409, 31, 28, -2, + -1, 24, 41, 374, 47, 2396, -2135, 15671, + 16323, -2864, 2460, -72, 409, 31, 28, -2, + -1, 26, 37, 391, -11, 2431, -2493, 16004, + 16005, -2494, 2432, -12, 392, 36, 26, -2, + -1, 26, 37, 391, -11, 2431, -2493, 16004, + 16005, -2494, 2432, -12, 392, 36, 26, -2, + -2, -28, 31, -409, -72, -2460, -2864, -16323, + 15671, 2135, 2396, -47, 374, -41, 24, 1, + -2, -28, 31, -409, -72, -2460, -2864, -16323, + 15671, 2135, 2396, -47, 374, -41, 24, 1, + -2, -30, 26, -425, -137, -2479, -3245, -16624, + 15322, 1788, 2354, -101, 357, -45, 22, 1, + -2, -30, 26, -425, -137, -2479, -3245, -16624, + 15322, 1788, 2354, -101, 357, -45, 22, 1, + -2, -32, 20, -440, -204, -2490, -3637, -16908, + 14959, 1454, 2304, -152, 339, -48, 21, 1, + -2, -32, 20, -440, -204, -2490, -3637, -16908, + 14959, 1454, 2304, -152, 339, -48, 21, 1, + -2, -33, 14, -455, -274, -2492, -4039, -17173, + 14583, 1133, 2249, -199, 320, -50, 19, 0, + -2, -33, 14, -455, -274, -2492, -4039, -17173, + 14583, 1133, 2249, -199, 320, -50, 19, 0, + -3, -35, 7, -468, -347, -2484, -4450, -17420, + 14194, 825, 2188, -243, 302, -52, 18, 0, + -3, -35, 7, -468, -347, -2484, -4450, -17420, + 14194, 825, 2188, -243, 302, -52, 18, 0, + -3, -37, -1, -480, -423, -2466, -4870, -17648, + 13794, 530, 2122, -283, 284, -54, 17, 0, + -3, -37, -1, -480, -423, -2466, -4870, -17648, + 13794, 530, 2122, -283, 284, -54, 17, 0, + -3, -39, -9, -491, -502, -2438, -5298, -17855, + 13383, 249, 2052, -320, 266, -56, 15, 0, + -3, -39, -9, -491, -502, -2438, -5298, -17855, + 13383, 249, 2052, -320, 266, -56, 15, 0, + -4, -41, -18, -501, -583, -2398, -5733, -18043, + 12963, -18, 1977, -354, 247, -56, 14, 0, + -4, -41, -18, -501, -583, -2398, -5733, -18043, + 12963, -18, 1977, -354, 247, -56, 14, 0, + -4, -43, -28, -508, -666, -2348, -6174, -18209, + 12534, -271, 1899, -384, 229, -57, 13, 0, + -4, -43, -28, -508, -666, -2348, -6174, -18209, + 12534, -271, 1899, -384, 229, -57, 13, 0, + -4, -44, -39, -515, -751, -2285, -6621, -18354, + 12097, -510, 1817, -411, 212, -57, 12, 0, + -4, -44, -39, -515, -751, -2285, -6621, -18354, + 12097, -510, 1817, -411, 212, -57, 12, 0, + -5, -46, -50, -519, -838, -2210, -7073, -18477, + 11654, -734, 1733, -435, 194, -57, 11, 0, + -5, -46, -50, -519, -838, -2210, -7073, -18477, + 11654, -734, 1733, -435, 194, -57, 11, 0, + -5, -48, -61, -522, -927, -2123, -7528, -18579, + 11205, -944, 1647, -456, 177, -57, 10, 0, + -5, -48, -61, -522, -927, -2123, -7528, -18579, + 11205, -944, 1647, -456, 177, -57, 10, 0, + -6, -49, -74, -522, -1016, -2023, -7987, -18658, + 10751, -1141, 1559, -474, 161, -57, 9, 0, + -6, -49, -74, -522, -1016, -2023, -7987, -18658, + 10751, -1141, 1559, -474, 161, -57, 9, 0, + -6, -51, -87, -520, -1107, -1910, -8448, -18714, + 10294, -1322, 1469, -488, 145, -56, 8, 0, + -6, -51, -87, -520, -1107, -1910, -8448, -18714, + 10294, -1322, 1469, -488, 145, -56, 8, 0, + -7, -52, -101, -516, -1197, -1784, -8910, -18748, + 9834, -1490, 1379, -500, 129, -55, 7, 0, + -7, -52, -101, -516, -1197, -1784, -8910, -18748, + 9834, -1490, 1379, -500, 129, -55, 7, 0, }; int synth_1to1_MMX(real *bandPtr, int channel, short *samples) @@ -194,23 +194,23 @@ int synth_1to1_MMX(real *bandPtr, int channel, short *samples) int bo1, i = 8; if (channel == 0) { - bo = (bo - 1) & 0xf; - buf = buffs[1]; + bo = (bo - 1) & 0xf; + buf = buffs[1]; } else { - samples++; - buf = buffs[0]; + samples++; + buf = buffs[0]; } if (bo & 1) { - b0 = buf[1]; - bo1 = bo + 1; - a = buf[0] + bo; - b = buf[1] + ((bo + 1) & 0xf); + b0 = buf[1]; + bo1 = bo + 1; + a = buf[0] + bo; + b = buf[1] + ((bo + 1) & 0xf); } else { - b0 = buf[0]; - bo1 = bo; - b = buf[0] + bo; - a = buf[1] + ((bo + 1) & 0xf); + b0 = buf[0]; + bo1 = bo; + b = buf[0] + bo; + a = buf[1] + ((bo + 1) & 0xf); } dct64_MMX_func(a, b, bandPtr); @@ -252,18 +252,18 @@ ASMALIGN(4) "packssdw %%mm0,%%mm0\n\t" "packssdw %%mm4,%%mm4\n\t" - "movq (%3), %%mm1\n\t" - "punpckldq %%mm4, %%mm0\n\t" - "pand "MANGLE(one_null)", %%mm1\n\t" - "pand "MANGLE(null_one)", %%mm0\n\t" - "por %%mm0, %%mm1\n\t" - "movq %%mm1,(%3)\n\t" + "movq (%3), %%mm1\n\t" + "punpckldq %%mm4, %%mm0\n\t" + "pand "MANGLE(one_null)", %%mm1\n\t" + "pand "MANGLE(null_one)", %%mm0\n\t" + "por %%mm0, %%mm1\n\t" + "movq %%mm1,(%3)\n\t" "add $64,%2\n\t" "add $128,%1\n\t" "add $8,%3\n\t" - "decl %0\n\t" + "decl %0\n\t" "jnz 0b\n\t" "movq (%1),%%mm0\n\t" @@ -283,7 +283,7 @@ ASMALIGN(4) "psrad $13,%%mm0\n\t" "packssdw %%mm0,%%mm0\n\t" "movd %%mm0,%%eax\n\t" - "movw %%ax, (%3)\n\t" + "movw %%ax, (%3)\n\t" "sub $32,%2\n\t" "add $64,%1\n\t" "add $4,%3\n\t" @@ -328,18 +328,18 @@ ASMALIGN(4) "psubsw %%mm1,%%mm0\n\t" "psubsw %%mm5,%%mm4\n\t" - "movq (%3), %%mm1\n\t" - "punpckldq %%mm4, %%mm0\n\t" - "pand "MANGLE(one_null)", %%mm1\n\t" - "pand "MANGLE(null_one)", %%mm0\n\t" - "por %%mm0, %%mm1\n\t" - "movq %%mm1,(%3)\n\t" + "movq (%3), %%mm1\n\t" + "punpckldq %%mm4, %%mm0\n\t" + "pand "MANGLE(one_null)", %%mm1\n\t" + "pand "MANGLE(null_one)", %%mm0\n\t" + "por %%mm0, %%mm1\n\t" + "movq %%mm1,(%3)\n\t" "sub $64,%2\n\t" "add $128,%1\n\t" "add $8,%3\n\t" "decl %0\n\t" - "jnz 1b\n\t" + "jnz 1b\n\t" "movq (%1),%%mm0\n\t" "pmaddwd (%2),%%mm0\n\t" @@ -360,10 +360,10 @@ ASMALIGN(4) "psubd %%mm0,%%mm0\n\t" "psubsw %%mm1,%%mm0\n\t" "movd %%mm0,%%eax\n\t" - "movw %%ax,(%3)\n\t" - "emms\n\t" - :"+r"(i), "+r"(window), "+r"(b0), "+r"(samples) - : - :"memory", "%eax"); + "movw %%ax,(%3)\n\t" + "emms\n\t" + :"+r"(i), "+r"(window), "+r"(b0), "+r"(samples) + : + :"memory", "%eax"); return 0; } diff --git a/mp3lib/equalizer.c b/mp3lib/equalizer.c index 78db17ef67..1c1a3f38ed 100644 --- a/mp3lib/equalizer.c +++ b/mp3lib/equalizer.c @@ -2,78 +2,77 @@ void init_spline(float *x, float *y, int n, float *y2) { + int i, k; + float p, qn, sig, un, *u; + u = malloc(n * sizeof (float)); - int i, k; - float p, qn, sig, un, *u; - u = malloc(n * sizeof (float)); + y2[0] = u[0] = 0.0; - y2[0] = u[0] = 0.0; + for (i = 1; i < n - 1; i++) + { + sig = ((float) x[i] - x[i - 1]) / ((float) x[i + 1] - x[i - 1]); + p = sig * y2[i - 1] + 2.0; + y2[i] = (sig - 1.0) / p; + u[i] = (((float) y[i + 1] - y[i]) / (x[i + 1] - x[i])) - + (((float) y[i] - y[i - 1]) / (x[i] - x[i - 1])); + u[i] = (6.0 * u[i] / (x[i + 1] - x[i - 1]) - sig * u[i - 1]) / p; + } + qn = un = 0.0; - for (i = 1; i < n - 1; i++) - { - sig = ((float) x[i] - x[i - 1]) / ((float) x[i + 1] - x[i - 1]); - p = sig * y2[i - 1] + 2.0; - y2[i] = (sig - 1.0) / p; - u[i] = (((float) y[i + 1] - y[i]) / (x[i + 1] - x[i])) - - (((float) y[i] - y[i - 1]) / (x[i] - x[i - 1])); - u[i] = (6.0 * u[i] / (x[i + 1] - x[i - 1]) - sig * u[i - 1]) / p; - } - qn = un = 0.0; - - y2[n - 1] = (un - qn * u[n - 2]) / (qn * y2[n - 2] + 1.0); - for (k = n - 2; k >= 0; k--) - y2[k] = y2[k] * y2[k + 1] + u[k]; - free(u); + y2[n - 1] = (un - qn * u[n - 2]) / (qn * y2[n - 2] + 1.0); + for (k = n - 2; k >= 0; k--) + y2[k] = y2[k] * y2[k + 1] + u[k]; + free(u); } float eval_spline(float xa[], float ya[], float y2a[], int n, float x) { - int klo, khi, k; - float h, b, a; + int klo, khi, k; + float h, b, a; - klo = 0; - khi = n - 1; - while (khi - klo > 1) - { - k = (khi + klo) >> 1; - if (xa[k] > x) - khi = k; - else - klo = k; - } - h = xa[khi] - xa[klo]; - a = (xa[khi] - x) / h; - b = (x - xa[klo]) / h; - return (a * ya[klo] + b * ya[khi] + ((a * a * a - a) * y2a[klo] + (b * b * b - b) * y2a[khi]) - * (h * h) / 6.0); + klo = 0; + khi = n - 1; + while (khi - klo > 1) + { + k = (khi + klo) >> 1; + if (xa[k] > x) + khi = k; + else + klo = k; + } + h = xa[khi] - xa[klo]; + a = (xa[khi] - x) / h; + b = (x - xa[klo]) / h; + return (a * ya[klo] + b * ya[khi] + ((a * a * a - a) * y2a[klo] + (b * b * b - b) * y2a[khi]) + * (h * h) / 6.0); } void mpg123_set_eq(int on, float preamp, float *b) { - float x[] = - {0, 1, 2, 3, 4, 5, 6, 7, 8, 9}, yf[10], val, band[10]; - int bands[] = - {0, 4, 8, 16, 26, 78, 157, 313, 366, 418}; - int i, j; + float x[] = + {0, 1, 2, 3, 4, 5, 6, 7, 8, 9}, yf[10], val, band[10]; + int bands[] = + {0, 4, 8, 16, 26, 78, 157, 313, 366, 418}; + int i, j; - mpg123_info->eq_active = on; - if (mpg123_info->eq_active) - { - for (i = 0; i < 10; i++) - { - band[i] = b[i] + preamp; - } + mpg123_info->eq_active = on; + if (mpg123_info->eq_active) + { + for (i = 0; i < 10; i++) + { + band[i] = b[i] + preamp; + } - init_spline(x, band, 10, yf); - for (i = 0; i < 9; i++) - { - for (j = bands[i]; j < bands[i + 1]; j++) - { - val = eval_spline(x, band, yf, 10, i + ((float) (j - bands[i]) * (1.0 / (bands[i + 1] - bands[i])))); - mpg123_info->eq_mul[j] = pow(2, val / 10.0); - } - } - for (i = bands[9]; i < 576; i++) - mpg123_info->eq_mul[i] = mpg123_info->eq_mul[bands[9] - 1]; - } + init_spline(x, band, 10, yf); + for (i = 0; i < 9; i++) + { + for (j = bands[i]; j < bands[i + 1]; j++) + { + val = eval_spline(x, band, yf, 10, i + ((float) (j - bands[i]) * (1.0 / (bands[i + 1] - bands[i])))); + mpg123_info->eq_mul[j] = pow(2, val / 10.0); + } + } + for (i = bands[9]; i < 576; i++) + mpg123_info->eq_mul[i] = mpg123_info->eq_mul[bands[9] - 1]; + } } diff --git a/mp3lib/l2tables.h b/mp3lib/l2tables.h index cf6299e88d..feef3d1cdb 100644 --- a/mp3lib/l2tables.h +++ b/mp3lib/l2tables.h @@ -17,143 +17,143 @@ */ static struct al_table alloc_0[] = { - {4,0},{5,3},{3,-3},{4,-7},{5,-15},{6,-31},{7,-63},{8,-127},{9,-255},{10,-511}, - {11,-1023},{12,-2047},{13,-4095},{14,-8191},{15,-16383},{16,-32767}, - {4,0},{5,3},{3,-3},{4,-7},{5,-15},{6,-31},{7,-63},{8,-127},{9,-255},{10,-511}, - {11,-1023},{12,-2047},{13,-4095},{14,-8191},{15,-16383},{16,-32767}, - {4,0},{5,3},{3,-3},{4,-7},{5,-15},{6,-31},{7,-63},{8,-127},{9,-255},{10,-511}, - {11,-1023},{12,-2047},{13,-4095},{14,-8191},{15,-16383},{16,-32767}, - {4,0},{5,3},{7,5},{3,-3},{10,9},{4,-7},{5,-15},{6,-31},{7,-63},{8,-127}, - {9,-255},{10,-511},{11,-1023},{12,-2047},{13,-4095},{16,-32767}, - {4,0},{5,3},{7,5},{3,-3},{10,9},{4,-7},{5,-15},{6,-31},{7,-63},{8,-127}, - {9,-255},{10,-511},{11,-1023},{12,-2047},{13,-4095},{16,-32767}, - {4,0},{5,3},{7,5},{3,-3},{10,9},{4,-7},{5,-15},{6,-31},{7,-63},{8,-127}, - {9,-255},{10,-511},{11,-1023},{12,-2047},{13,-4095},{16,-32767}, - {4,0},{5,3},{7,5},{3,-3},{10,9},{4,-7},{5,-15},{6,-31},{7,-63},{8,-127}, - {9,-255},{10,-511},{11,-1023},{12,-2047},{13,-4095},{16,-32767}, - {4,0},{5,3},{7,5},{3,-3},{10,9},{4,-7},{5,-15},{6,-31},{7,-63},{8,-127}, - {9,-255},{10,-511},{11,-1023},{12,-2047},{13,-4095},{16,-32767}, - {4,0},{5,3},{7,5},{3,-3},{10,9},{4,-7},{5,-15},{6,-31},{7,-63},{8,-127}, - {9,-255},{10,-511},{11,-1023},{12,-2047},{13,-4095},{16,-32767}, - {4,0},{5,3},{7,5},{3,-3},{10,9},{4,-7},{5,-15},{6,-31},{7,-63},{8,-127}, - {9,-255},{10,-511},{11,-1023},{12,-2047},{13,-4095},{16,-32767}, - {4,0},{5,3},{7,5},{3,-3},{10,9},{4,-7},{5,-15},{6,-31},{7,-63},{8,-127}, - {9,-255},{10,-511},{11,-1023},{12,-2047},{13,-4095},{16,-32767}, - {3,0},{5,3},{7,5},{3,-3},{10,9},{4,-7},{5,-15},{16,-32767}, - {3,0},{5,3},{7,5},{3,-3},{10,9},{4,-7},{5,-15},{16,-32767}, - {3,0},{5,3},{7,5},{3,-3},{10,9},{4,-7},{5,-15},{16,-32767}, - {3,0},{5,3},{7,5},{3,-3},{10,9},{4,-7},{5,-15},{16,-32767}, - {3,0},{5,3},{7,5},{3,-3},{10,9},{4,-7},{5,-15},{16,-32767}, - {3,0},{5,3},{7,5},{3,-3},{10,9},{4,-7},{5,-15},{16,-32767}, - {3,0},{5,3},{7,5},{3,-3},{10,9},{4,-7},{5,-15},{16,-32767}, - {3,0},{5,3},{7,5},{3,-3},{10,9},{4,-7},{5,-15},{16,-32767}, - {3,0},{5,3},{7,5},{3,-3},{10,9},{4,-7},{5,-15},{16,-32767}, - {3,0},{5,3},{7,5},{3,-3},{10,9},{4,-7},{5,-15},{16,-32767}, - {3,0},{5,3},{7,5},{3,-3},{10,9},{4,-7},{5,-15},{16,-32767}, - {3,0},{5,3},{7,5},{3,-3},{10,9},{4,-7},{5,-15},{16,-32767}, - {2,0},{5,3},{7,5},{16,-32767}, - {2,0},{5,3},{7,5},{16,-32767}, - {2,0},{5,3},{7,5},{16,-32767}, - {2,0},{5,3},{7,5},{16,-32767} }; + {4,0},{5,3},{3,-3},{4,-7},{5,-15},{6,-31},{7,-63},{8,-127},{9,-255},{10,-511}, + {11,-1023},{12,-2047},{13,-4095},{14,-8191},{15,-16383},{16,-32767}, + {4,0},{5,3},{3,-3},{4,-7},{5,-15},{6,-31},{7,-63},{8,-127},{9,-255},{10,-511}, + {11,-1023},{12,-2047},{13,-4095},{14,-8191},{15,-16383},{16,-32767}, + {4,0},{5,3},{3,-3},{4,-7},{5,-15},{6,-31},{7,-63},{8,-127},{9,-255},{10,-511}, + {11,-1023},{12,-2047},{13,-4095},{14,-8191},{15,-16383},{16,-32767}, + {4,0},{5,3},{7,5},{3,-3},{10,9},{4,-7},{5,-15},{6,-31},{7,-63},{8,-127}, + {9,-255},{10,-511},{11,-1023},{12,-2047},{13,-4095},{16,-32767}, + {4,0},{5,3},{7,5},{3,-3},{10,9},{4,-7},{5,-15},{6,-31},{7,-63},{8,-127}, + {9,-255},{10,-511},{11,-1023},{12,-2047},{13,-4095},{16,-32767}, + {4,0},{5,3},{7,5},{3,-3},{10,9},{4,-7},{5,-15},{6,-31},{7,-63},{8,-127}, + {9,-255},{10,-511},{11,-1023},{12,-2047},{13,-4095},{16,-32767}, + {4,0},{5,3},{7,5},{3,-3},{10,9},{4,-7},{5,-15},{6,-31},{7,-63},{8,-127}, + {9,-255},{10,-511},{11,-1023},{12,-2047},{13,-4095},{16,-32767}, + {4,0},{5,3},{7,5},{3,-3},{10,9},{4,-7},{5,-15},{6,-31},{7,-63},{8,-127}, + {9,-255},{10,-511},{11,-1023},{12,-2047},{13,-4095},{16,-32767}, + {4,0},{5,3},{7,5},{3,-3},{10,9},{4,-7},{5,-15},{6,-31},{7,-63},{8,-127}, + {9,-255},{10,-511},{11,-1023},{12,-2047},{13,-4095},{16,-32767}, + {4,0},{5,3},{7,5},{3,-3},{10,9},{4,-7},{5,-15},{6,-31},{7,-63},{8,-127}, + {9,-255},{10,-511},{11,-1023},{12,-2047},{13,-4095},{16,-32767}, + {4,0},{5,3},{7,5},{3,-3},{10,9},{4,-7},{5,-15},{6,-31},{7,-63},{8,-127}, + {9,-255},{10,-511},{11,-1023},{12,-2047},{13,-4095},{16,-32767}, + {3,0},{5,3},{7,5},{3,-3},{10,9},{4,-7},{5,-15},{16,-32767}, + {3,0},{5,3},{7,5},{3,-3},{10,9},{4,-7},{5,-15},{16,-32767}, + {3,0},{5,3},{7,5},{3,-3},{10,9},{4,-7},{5,-15},{16,-32767}, + {3,0},{5,3},{7,5},{3,-3},{10,9},{4,-7},{5,-15},{16,-32767}, + {3,0},{5,3},{7,5},{3,-3},{10,9},{4,-7},{5,-15},{16,-32767}, + {3,0},{5,3},{7,5},{3,-3},{10,9},{4,-7},{5,-15},{16,-32767}, + {3,0},{5,3},{7,5},{3,-3},{10,9},{4,-7},{5,-15},{16,-32767}, + {3,0},{5,3},{7,5},{3,-3},{10,9},{4,-7},{5,-15},{16,-32767}, + {3,0},{5,3},{7,5},{3,-3},{10,9},{4,-7},{5,-15},{16,-32767}, + {3,0},{5,3},{7,5},{3,-3},{10,9},{4,-7},{5,-15},{16,-32767}, + {3,0},{5,3},{7,5},{3,-3},{10,9},{4,-7},{5,-15},{16,-32767}, + {3,0},{5,3},{7,5},{3,-3},{10,9},{4,-7},{5,-15},{16,-32767}, + {2,0},{5,3},{7,5},{16,-32767}, + {2,0},{5,3},{7,5},{16,-32767}, + {2,0},{5,3},{7,5},{16,-32767}, + {2,0},{5,3},{7,5},{16,-32767} }; static struct al_table alloc_1[] = { - {4,0},{5,3},{3,-3},{4,-7},{5,-15},{6,-31},{7,-63},{8,-127},{9,-255},{10,-511}, - {11,-1023},{12,-2047},{13,-4095},{14,-8191},{15,-16383},{16,-32767}, - {4,0},{5,3},{3,-3},{4,-7},{5,-15},{6,-31},{7,-63},{8,-127},{9,-255},{10,-511}, - {11,-1023},{12,-2047},{13,-4095},{14,-8191},{15,-16383},{16,-32767}, - {4,0},{5,3},{3,-3},{4,-7},{5,-15},{6,-31},{7,-63},{8,-127},{9,-255},{10,-511}, - {11,-1023},{12,-2047},{13,-4095},{14,-8191},{15,-16383},{16,-32767}, - {4,0},{5,3},{7,5},{3,-3},{10,9},{4,-7},{5,-15},{6,-31},{7,-63},{8,-127}, - {9,-255},{10,-511},{11,-1023},{12,-2047},{13,-4095},{16,-32767}, - {4,0},{5,3},{7,5},{3,-3},{10,9},{4,-7},{5,-15},{6,-31},{7,-63},{8,-127}, - {9,-255},{10,-511},{11,-1023},{12,-2047},{13,-4095},{16,-32767}, - {4,0},{5,3},{7,5},{3,-3},{10,9},{4,-7},{5,-15},{6,-31},{7,-63},{8,-127}, - {9,-255},{10,-511},{11,-1023},{12,-2047},{13,-4095},{16,-32767}, - {4,0},{5,3},{7,5},{3,-3},{10,9},{4,-7},{5,-15},{6,-31},{7,-63},{8,-127}, - {9,-255},{10,-511},{11,-1023},{12,-2047},{13,-4095},{16,-32767}, - {4,0},{5,3},{7,5},{3,-3},{10,9},{4,-7},{5,-15},{6,-31},{7,-63},{8,-127}, - {9,-255},{10,-511},{11,-1023},{12,-2047},{13,-4095},{16,-32767}, - {4,0},{5,3},{7,5},{3,-3},{10,9},{4,-7},{5,-15},{6,-31},{7,-63},{8,-127}, - {9,-255},{10,-511},{11,-1023},{12,-2047},{13,-4095},{16,-32767}, - {4,0},{5,3},{7,5},{3,-3},{10,9},{4,-7},{5,-15},{6,-31},{7,-63},{8,-127}, - {9,-255},{10,-511},{11,-1023},{12,-2047},{13,-4095},{16,-32767}, - {4,0},{5,3},{7,5},{3,-3},{10,9},{4,-7},{5,-15},{6,-31},{7,-63},{8,-127}, - {9,-255},{10,-511},{11,-1023},{12,-2047},{13,-4095},{16,-32767}, - {3,0},{5,3},{7,5},{3,-3},{10,9},{4,-7},{5,-15},{16,-32767}, - {3,0},{5,3},{7,5},{3,-3},{10,9},{4,-7},{5,-15},{16,-32767}, - {3,0},{5,3},{7,5},{3,-3},{10,9},{4,-7},{5,-15},{16,-32767}, - {3,0},{5,3},{7,5},{3,-3},{10,9},{4,-7},{5,-15},{16,-32767}, - {3,0},{5,3},{7,5},{3,-3},{10,9},{4,-7},{5,-15},{16,-32767}, - {3,0},{5,3},{7,5},{3,-3},{10,9},{4,-7},{5,-15},{16,-32767}, - {3,0},{5,3},{7,5},{3,-3},{10,9},{4,-7},{5,-15},{16,-32767}, - {3,0},{5,3},{7,5},{3,-3},{10,9},{4,-7},{5,-15},{16,-32767}, - {3,0},{5,3},{7,5},{3,-3},{10,9},{4,-7},{5,-15},{16,-32767}, - {3,0},{5,3},{7,5},{3,-3},{10,9},{4,-7},{5,-15},{16,-32767}, - {3,0},{5,3},{7,5},{3,-3},{10,9},{4,-7},{5,-15},{16,-32767}, - {3,0},{5,3},{7,5},{3,-3},{10,9},{4,-7},{5,-15},{16,-32767}, - {2,0},{5,3},{7,5},{16,-32767}, - {2,0},{5,3},{7,5},{16,-32767}, - {2,0},{5,3},{7,5},{16,-32767}, - {2,0},{5,3},{7,5},{16,-32767}, - {2,0},{5,3},{7,5},{16,-32767}, - {2,0},{5,3},{7,5},{16,-32767}, - {2,0},{5,3},{7,5},{16,-32767} }; + {4,0},{5,3},{3,-3},{4,-7},{5,-15},{6,-31},{7,-63},{8,-127},{9,-255},{10,-511}, + {11,-1023},{12,-2047},{13,-4095},{14,-8191},{15,-16383},{16,-32767}, + {4,0},{5,3},{3,-3},{4,-7},{5,-15},{6,-31},{7,-63},{8,-127},{9,-255},{10,-511}, + {11,-1023},{12,-2047},{13,-4095},{14,-8191},{15,-16383},{16,-32767}, + {4,0},{5,3},{3,-3},{4,-7},{5,-15},{6,-31},{7,-63},{8,-127},{9,-255},{10,-511}, + {11,-1023},{12,-2047},{13,-4095},{14,-8191},{15,-16383},{16,-32767}, + {4,0},{5,3},{7,5},{3,-3},{10,9},{4,-7},{5,-15},{6,-31},{7,-63},{8,-127}, + {9,-255},{10,-511},{11,-1023},{12,-2047},{13,-4095},{16,-32767}, + {4,0},{5,3},{7,5},{3,-3},{10,9},{4,-7},{5,-15},{6,-31},{7,-63},{8,-127}, + {9,-255},{10,-511},{11,-1023},{12,-2047},{13,-4095},{16,-32767}, + {4,0},{5,3},{7,5},{3,-3},{10,9},{4,-7},{5,-15},{6,-31},{7,-63},{8,-127}, + {9,-255},{10,-511},{11,-1023},{12,-2047},{13,-4095},{16,-32767}, + {4,0},{5,3},{7,5},{3,-3},{10,9},{4,-7},{5,-15},{6,-31},{7,-63},{8,-127}, + {9,-255},{10,-511},{11,-1023},{12,-2047},{13,-4095},{16,-32767}, + {4,0},{5,3},{7,5},{3,-3},{10,9},{4,-7},{5,-15},{6,-31},{7,-63},{8,-127}, + {9,-255},{10,-511},{11,-1023},{12,-2047},{13,-4095},{16,-32767}, + {4,0},{5,3},{7,5},{3,-3},{10,9},{4,-7},{5,-15},{6,-31},{7,-63},{8,-127}, + {9,-255},{10,-511},{11,-1023},{12,-2047},{13,-4095},{16,-32767}, + {4,0},{5,3},{7,5},{3,-3},{10,9},{4,-7},{5,-15},{6,-31},{7,-63},{8,-127}, + {9,-255},{10,-511},{11,-1023},{12,-2047},{13,-4095},{16,-32767}, + {4,0},{5,3},{7,5},{3,-3},{10,9},{4,-7},{5,-15},{6,-31},{7,-63},{8,-127}, + {9,-255},{10,-511},{11,-1023},{12,-2047},{13,-4095},{16,-32767}, + {3,0},{5,3},{7,5},{3,-3},{10,9},{4,-7},{5,-15},{16,-32767}, + {3,0},{5,3},{7,5},{3,-3},{10,9},{4,-7},{5,-15},{16,-32767}, + {3,0},{5,3},{7,5},{3,-3},{10,9},{4,-7},{5,-15},{16,-32767}, + {3,0},{5,3},{7,5},{3,-3},{10,9},{4,-7},{5,-15},{16,-32767}, + {3,0},{5,3},{7,5},{3,-3},{10,9},{4,-7},{5,-15},{16,-32767}, + {3,0},{5,3},{7,5},{3,-3},{10,9},{4,-7},{5,-15},{16,-32767}, + {3,0},{5,3},{7,5},{3,-3},{10,9},{4,-7},{5,-15},{16,-32767}, + {3,0},{5,3},{7,5},{3,-3},{10,9},{4,-7},{5,-15},{16,-32767}, + {3,0},{5,3},{7,5},{3,-3},{10,9},{4,-7},{5,-15},{16,-32767}, + {3,0},{5,3},{7,5},{3,-3},{10,9},{4,-7},{5,-15},{16,-32767}, + {3,0},{5,3},{7,5},{3,-3},{10,9},{4,-7},{5,-15},{16,-32767}, + {3,0},{5,3},{7,5},{3,-3},{10,9},{4,-7},{5,-15},{16,-32767}, + {2,0},{5,3},{7,5},{16,-32767}, + {2,0},{5,3},{7,5},{16,-32767}, + {2,0},{5,3},{7,5},{16,-32767}, + {2,0},{5,3},{7,5},{16,-32767}, + {2,0},{5,3},{7,5},{16,-32767}, + {2,0},{5,3},{7,5},{16,-32767}, + {2,0},{5,3},{7,5},{16,-32767} }; static struct al_table alloc_2[] = { - {4,0},{5,3},{7,5},{10,9},{4,-7},{5,-15},{6,-31},{7,-63},{8,-127},{9,-255}, - {10,-511},{11,-1023},{12,-2047},{13,-4095},{14,-8191},{15,-16383}, - {4,0},{5,3},{7,5},{10,9},{4,-7},{5,-15},{6,-31},{7,-63},{8,-127},{9,-255}, - {10,-511},{11,-1023},{12,-2047},{13,-4095},{14,-8191},{15,-16383}, - {3,0},{5,3},{7,5},{10,9},{4,-7},{5,-15},{6,-31},{7,-63}, - {3,0},{5,3},{7,5},{10,9},{4,-7},{5,-15},{6,-31},{7,-63}, - {3,0},{5,3},{7,5},{10,9},{4,-7},{5,-15},{6,-31},{7,-63}, - {3,0},{5,3},{7,5},{10,9},{4,-7},{5,-15},{6,-31},{7,-63}, - {3,0},{5,3},{7,5},{10,9},{4,-7},{5,-15},{6,-31},{7,-63}, - {3,0},{5,3},{7,5},{10,9},{4,-7},{5,-15},{6,-31},{7,-63} }; + {4,0},{5,3},{7,5},{10,9},{4,-7},{5,-15},{6,-31},{7,-63},{8,-127},{9,-255}, + {10,-511},{11,-1023},{12,-2047},{13,-4095},{14,-8191},{15,-16383}, + {4,0},{5,3},{7,5},{10,9},{4,-7},{5,-15},{6,-31},{7,-63},{8,-127},{9,-255}, + {10,-511},{11,-1023},{12,-2047},{13,-4095},{14,-8191},{15,-16383}, + {3,0},{5,3},{7,5},{10,9},{4,-7},{5,-15},{6,-31},{7,-63}, + {3,0},{5,3},{7,5},{10,9},{4,-7},{5,-15},{6,-31},{7,-63}, + {3,0},{5,3},{7,5},{10,9},{4,-7},{5,-15},{6,-31},{7,-63}, + {3,0},{5,3},{7,5},{10,9},{4,-7},{5,-15},{6,-31},{7,-63}, + {3,0},{5,3},{7,5},{10,9},{4,-7},{5,-15},{6,-31},{7,-63}, + {3,0},{5,3},{7,5},{10,9},{4,-7},{5,-15},{6,-31},{7,-63} }; static struct al_table alloc_3[] = { - {4,0},{5,3},{7,5},{10,9},{4,-7},{5,-15},{6,-31},{7,-63},{8,-127},{9,-255}, - {10,-511},{11,-1023},{12,-2047},{13,-4095},{14,-8191},{15,-16383}, - {4,0},{5,3},{7,5},{10,9},{4,-7},{5,-15},{6,-31},{7,-63},{8,-127},{9,-255}, - {10,-511},{11,-1023},{12,-2047},{13,-4095},{14,-8191},{15,-16383}, - {3,0},{5,3},{7,5},{10,9},{4,-7},{5,-15},{6,-31},{7,-63}, - {3,0},{5,3},{7,5},{10,9},{4,-7},{5,-15},{6,-31},{7,-63}, - {3,0},{5,3},{7,5},{10,9},{4,-7},{5,-15},{6,-31},{7,-63}, - {3,0},{5,3},{7,5},{10,9},{4,-7},{5,-15},{6,-31},{7,-63}, - {3,0},{5,3},{7,5},{10,9},{4,-7},{5,-15},{6,-31},{7,-63}, - {3,0},{5,3},{7,5},{10,9},{4,-7},{5,-15},{6,-31},{7,-63}, - {3,0},{5,3},{7,5},{10,9},{4,-7},{5,-15},{6,-31},{7,-63}, - {3,0},{5,3},{7,5},{10,9},{4,-7},{5,-15},{6,-31},{7,-63}, - {3,0},{5,3},{7,5},{10,9},{4,-7},{5,-15},{6,-31},{7,-63}, - {3,0},{5,3},{7,5},{10,9},{4,-7},{5,-15},{6,-31},{7,-63} }; + {4,0},{5,3},{7,5},{10,9},{4,-7},{5,-15},{6,-31},{7,-63},{8,-127},{9,-255}, + {10,-511},{11,-1023},{12,-2047},{13,-4095},{14,-8191},{15,-16383}, + {4,0},{5,3},{7,5},{10,9},{4,-7},{5,-15},{6,-31},{7,-63},{8,-127},{9,-255}, + {10,-511},{11,-1023},{12,-2047},{13,-4095},{14,-8191},{15,-16383}, + {3,0},{5,3},{7,5},{10,9},{4,-7},{5,-15},{6,-31},{7,-63}, + {3,0},{5,3},{7,5},{10,9},{4,-7},{5,-15},{6,-31},{7,-63}, + {3,0},{5,3},{7,5},{10,9},{4,-7},{5,-15},{6,-31},{7,-63}, + {3,0},{5,3},{7,5},{10,9},{4,-7},{5,-15},{6,-31},{7,-63}, + {3,0},{5,3},{7,5},{10,9},{4,-7},{5,-15},{6,-31},{7,-63}, + {3,0},{5,3},{7,5},{10,9},{4,-7},{5,-15},{6,-31},{7,-63}, + {3,0},{5,3},{7,5},{10,9},{4,-7},{5,-15},{6,-31},{7,-63}, + {3,0},{5,3},{7,5},{10,9},{4,-7},{5,-15},{6,-31},{7,-63}, + {3,0},{5,3},{7,5},{10,9},{4,-7},{5,-15},{6,-31},{7,-63}, + {3,0},{5,3},{7,5},{10,9},{4,-7},{5,-15},{6,-31},{7,-63} }; static struct al_table alloc_4[] = { - {4,0},{5,3},{7,5},{3,-3},{10,9},{4,-7},{5,-15},{6,-31},{7,-63},{8,-127}, - {9,-255},{10,-511},{11,-1023},{12,-2047},{13,-4095},{14,-8191}, - {4,0},{5,3},{7,5},{3,-3},{10,9},{4,-7},{5,-15},{6,-31},{7,-63},{8,-127}, - {9,-255},{10,-511},{11,-1023},{12,-2047},{13,-4095},{14,-8191}, - {4,0},{5,3},{7,5},{3,-3},{10,9},{4,-7},{5,-15},{6,-31},{7,-63},{8,-127}, - {9,-255},{10,-511},{11,-1023},{12,-2047},{13,-4095},{14,-8191}, - {4,0},{5,3},{7,5},{3,-3},{10,9},{4,-7},{5,-15},{6,-31},{7,-63},{8,-127}, - {9,-255},{10,-511},{11,-1023},{12,-2047},{13,-4095},{14,-8191}, - {3,0},{5,3},{7,5},{10,9},{4,-7},{5,-15},{6,-31},{7,-63}, - {3,0},{5,3},{7,5},{10,9},{4,-7},{5,-15},{6,-31},{7,-63}, - {3,0},{5,3},{7,5},{10,9},{4,-7},{5,-15},{6,-31},{7,-63}, - {3,0},{5,3},{7,5},{10,9},{4,-7},{5,-15},{6,-31},{7,-63}, - {3,0},{5,3},{7,5},{10,9},{4,-7},{5,-15},{6,-31},{7,-63}, - {3,0},{5,3},{7,5},{10,9},{4,-7},{5,-15},{6,-31},{7,-63}, - {3,0},{5,3},{7,5},{10,9},{4,-7},{5,-15},{6,-31},{7,-63}, - {2,0},{5,3},{7,5},{10,9}, - {2,0},{5,3},{7,5},{10,9}, - {2,0},{5,3},{7,5},{10,9}, - {2,0},{5,3},{7,5},{10,9}, - {2,0},{5,3},{7,5},{10,9}, - {2,0},{5,3},{7,5},{10,9}, - {2,0},{5,3},{7,5},{10,9}, - {2,0},{5,3},{7,5},{10,9}, - {2,0},{5,3},{7,5},{10,9}, - {2,0},{5,3},{7,5},{10,9}, - {2,0},{5,3},{7,5},{10,9}, + {4,0},{5,3},{7,5},{3,-3},{10,9},{4,-7},{5,-15},{6,-31},{7,-63},{8,-127}, + {9,-255},{10,-511},{11,-1023},{12,-2047},{13,-4095},{14,-8191}, + {4,0},{5,3},{7,5},{3,-3},{10,9},{4,-7},{5,-15},{6,-31},{7,-63},{8,-127}, + {9,-255},{10,-511},{11,-1023},{12,-2047},{13,-4095},{14,-8191}, + {4,0},{5,3},{7,5},{3,-3},{10,9},{4,-7},{5,-15},{6,-31},{7,-63},{8,-127}, + {9,-255},{10,-511},{11,-1023},{12,-2047},{13,-4095},{14,-8191}, + {4,0},{5,3},{7,5},{3,-3},{10,9},{4,-7},{5,-15},{6,-31},{7,-63},{8,-127}, + {9,-255},{10,-511},{11,-1023},{12,-2047},{13,-4095},{14,-8191}, + {3,0},{5,3},{7,5},{10,9},{4,-7},{5,-15},{6,-31},{7,-63}, + {3,0},{5,3},{7,5},{10,9},{4,-7},{5,-15},{6,-31},{7,-63}, + {3,0},{5,3},{7,5},{10,9},{4,-7},{5,-15},{6,-31},{7,-63}, + {3,0},{5,3},{7,5},{10,9},{4,-7},{5,-15},{6,-31},{7,-63}, + {3,0},{5,3},{7,5},{10,9},{4,-7},{5,-15},{6,-31},{7,-63}, + {3,0},{5,3},{7,5},{10,9},{4,-7},{5,-15},{6,-31},{7,-63}, + {3,0},{5,3},{7,5},{10,9},{4,-7},{5,-15},{6,-31},{7,-63}, + {2,0},{5,3},{7,5},{10,9}, + {2,0},{5,3},{7,5},{10,9}, + {2,0},{5,3},{7,5},{10,9}, + {2,0},{5,3},{7,5},{10,9}, + {2,0},{5,3},{7,5},{10,9}, + {2,0},{5,3},{7,5},{10,9}, + {2,0},{5,3},{7,5},{10,9}, + {2,0},{5,3},{7,5},{10,9}, + {2,0},{5,3},{7,5},{10,9}, + {2,0},{5,3},{7,5},{10,9}, + {2,0},{5,3},{7,5},{10,9}, {2,0},{5,3},{7,5},{10,9}, {2,0},{5,3},{7,5},{10,9}, {2,0},{5,3},{7,5},{10,9}, diff --git a/mp3lib/layer1.c b/mp3lib/layer1.c index d53c628314..a4be24677a 100644 --- a/mp3lib/layer1.c +++ b/mp3lib/layer1.c @@ -61,7 +61,7 @@ static void I_step_one(unsigned int balloc[], unsigned int scale_index[2][SBLIMI } static void I_step_two(real fraction[2][SBLIMIT],unsigned int balloc[2*SBLIMIT], - unsigned int scale_index[2][SBLIMIT],struct frame *fr) + unsigned int scale_index[2][SBLIMIT],struct frame *fr) { int i,n; int smpb[2*SBLIMIT]; /* values: 0-65535 */ diff --git a/mp3lib/layer2.c b/mp3lib/layer2.c index 867da2f2dc..f86d716882 100644 --- a/mp3lib/layer2.c +++ b/mp3lib/layer2.c @@ -18,7 +18,7 @@ static int grp_3tab[32 * 3] = { 0, }; /* used: 27 */ static int grp_5tab[128 * 3] = { 0, }; /* used: 125 */ static int grp_9tab[1024 * 3] = { 0, }; /* used: 729 */ -static real muls[27][64]; /* also used by layer 1 */ +static real muls[27][64]; /* also used by layer 1 */ static void init_layer2(void) { @@ -59,7 +59,7 @@ static void init_layer2(void) if(_has_mmx) { for(j=3,i=0;i<63;i++,j--) - *table++ = 16384 * m * pow(2.0,(double) j / 3.0); + *table++ = 16384 * m * pow(2.0,(double) j / 3.0); } else for(j=3,i=0;i<63;i++,j--) @@ -189,7 +189,7 @@ static void II_step_two(unsigned int *bit_alloc,real fraction[2][4][SBLIMIT],int for (i=jsbound;i<sblimit;i++,alloc1+=(1<<step)) { step = alloc1->bits; - bita++; /* channel 1 and channel 2 bitalloc are the same */ + bita++; /* channel 1 and channel 2 bitalloc are the same */ if ( (ba=*bita++) ) { k=(alloc2 = alloc1+ba)->bits; diff --git a/mp3lib/layer3.c b/mp3lib/layer3.c index 16dfb06bb8..91943a8f18 100644 --- a/mp3lib/layer3.c +++ b/mp3lib/layer3.c @@ -25,7 +25,7 @@ #define CUT_HF #endif -# define REAL_MUL(x, y) ((x) * (y)) +#define REAL_MUL(x, y) ((x) * (y)) static real ispow[8207]; static real aa_ca[8],aa_cs[8]; @@ -734,8 +734,8 @@ static int III_dequantize_sample(real xr[SBLIMIT][SSLIMIT],int *scf, mask <<= 1; } if(part2remain+num <= 0) { - num -= part2remain+num; - break; + num -= part2remain+num; + break; } for(i=0;i<4;i++) { @@ -851,7 +851,7 @@ static int III_dequantize_sample(real xr[SBLIMIT][SSLIMIT],int *scf, if (x == 15 && h->linbits) { max = cb; - REFRESH_MASK; + REFRESH_MASK; x += ((unsigned long) mask) >> (BITSHIFT+8-h->linbits); num -= h->linbits+1; mask <<= h->linbits; @@ -875,7 +875,7 @@ static int III_dequantize_sample(real xr[SBLIMIT][SSLIMIT],int *scf, if (y == 15 && h->linbits) { max = cb; - REFRESH_MASK; + REFRESH_MASK; y += ((unsigned long) mask) >> (BITSHIFT+8-h->linbits); num -= h->linbits+1; mask <<= h->linbits; @@ -914,7 +914,7 @@ static int III_dequantize_sample(real xr[SBLIMIT][SSLIMIT],int *scf, mask <<= 1; } if(part2remain+num <= 0) { - num -= part2remain+num; + num -= part2remain+num; break; } @@ -1003,7 +1003,7 @@ static void III_i_stereo(real xr_buf[2][SBLIMIT][SSLIMIT],int *scalefac, #if 0 if(lsf) { int p = gr_info->scalefac_compress & 0x1; - if(ms_stereo) { + if(ms_stereo) { tab1 = pow1_2[p]; tab2 = pow2_2[p]; } else { @@ -1329,17 +1329,17 @@ static int do_layer3(struct frame *fr,int single){ for(ch=0;ch<stereo1;ch++) { struct gr_info_s *gr_info = &(sideinfo.ch[ch].gr[gr]); - III_antialias(hybridIn[ch],gr_info); - III_hybrid(hybridIn[ch], hybridOut[ch], ch,gr_info); + III_antialias(hybridIn[ch],gr_info); + III_hybrid(hybridIn[ch], hybridOut[ch], ch,gr_info); } for(ss=0;ss<SSLIMIT;ss++) { if(single >= 0) { - clip += (fr->synth_mono)(hybridOut[0][ss],pcm_sample,&pcm_point); - } else { - int p1 = pcm_point; - clip += (fr->synth)(hybridOut[0][ss],0,pcm_sample,&p1); - clip += (fr->synth)(hybridOut[1][ss],1,pcm_sample,&pcm_point); + clip += (fr->synth_mono)(hybridOut[0][ss],pcm_sample,&pcm_point); + } else { + int p1 = pcm_point; + clip += (fr->synth)(hybridOut[0][ss],0,pcm_sample,&p1); + clip += (fr->synth)(hybridOut[1][ss],1,pcm_sample,&pcm_point); } } diff --git a/mp3lib/mpg123.h b/mp3lib/mpg123.h index 5a7548250a..0b9f781f80 100644 --- a/mp3lib/mpg123.h +++ b/mp3lib/mpg123.h @@ -15,10 +15,10 @@ #include "config.h" #ifndef M_PI -#define M_PI 3.141592653589793238462 +#define M_PI 3.141592653589793238462 #endif #ifndef M_SQRT2 -#define M_SQRT2 1.414213562373095048802 +#define M_SQRT2 1.414213562373095048802 #endif #define REAL_IS_FLOAT #define NEW_DCT9 diff --git a/mp3lib/sr1.c b/mp3lib/sr1.c index 694961ebc0..caed5f8c90 100644 --- a/mp3lib/sr1.c +++ b/mp3lib/sr1.c @@ -412,34 +412,34 @@ void MP3_Init(void){ #if HAVE_MMX if (gCpuCaps.hasMMX) { - _has_mmx = 1; - synth_func = synth_1to1_MMX; + _has_mmx = 1; + synth_func = synth_1to1_MMX; } #endif #if HAVE_AMD3DNOWEXT if (gCpuCaps.has3DNowExt) { - dct36_func=dct36_3dnowex; - dct64_MMX_func= dct64_MMX_3dnowex; - mp_msg(MSGT_DECAUDIO,MSGL_V,"mp3lib: using 3DNow!Ex optimized decore!\n"); + dct36_func=dct36_3dnowex; + dct64_MMX_func= dct64_MMX_3dnowex; + mp_msg(MSGT_DECAUDIO,MSGL_V,"mp3lib: using 3DNow!Ex optimized decore!\n"); } else #endif #if HAVE_AMD3DNOW if (gCpuCaps.has3DNow) { - dct36_func = dct36_3dnow; - dct64_MMX_func = dct64_MMX_3dnow; - mp_msg(MSGT_DECAUDIO,MSGL_V,"mp3lib: using 3DNow! optimized decore!\n"); + dct36_func = dct36_3dnow; + dct64_MMX_func = dct64_MMX_3dnow; + mp_msg(MSGT_DECAUDIO,MSGL_V,"mp3lib: using 3DNow! optimized decore!\n"); } else #endif #if HAVE_SSE if (gCpuCaps.hasSSE) { - dct64_MMX_func = dct64_sse; - mp_msg(MSGT_DECAUDIO,MSGL_V,"mp3lib: using SSE optimized decore!\n"); + dct64_MMX_func = dct64_sse; + mp_msg(MSGT_DECAUDIO,MSGL_V,"mp3lib: using SSE optimized decore!\n"); } else #endif @@ -447,28 +447,28 @@ void MP3_Init(void){ #if HAVE_MMX if (gCpuCaps.hasMMX) { - dct64_MMX_func = dct64_MMX; - mp_msg(MSGT_DECAUDIO,MSGL_V,"mp3lib: using MMX optimized decore!\n"); + dct64_MMX_func = dct64_MMX; + mp_msg(MSGT_DECAUDIO,MSGL_V,"mp3lib: using MMX optimized decore!\n"); } else #endif if (gCpuCaps.cpuType >= CPUTYPE_I586) { - synth_func = synth_1to1_pent; - mp_msg(MSGT_DECAUDIO,MSGL_V,"mp3lib: using Pentium optimized decore!\n"); + synth_func = synth_1to1_pent; + mp_msg(MSGT_DECAUDIO,MSGL_V,"mp3lib: using Pentium optimized decore!\n"); } else #endif /* ARCH_X86_32 */ #if HAVE_ALTIVEC if (gCpuCaps.hasAltiVec) { - mp_msg(MSGT_DECAUDIO,MSGL_V,"mp3lib: using AltiVec optimized decore!\n"); + mp_msg(MSGT_DECAUDIO,MSGL_V,"mp3lib: using AltiVec optimized decore!\n"); } else #endif { - synth_func = NULL; /* use default c version */ - mp_msg(MSGT_DECAUDIO,MSGL_V,"mp3lib: using generic C decore!\n"); + synth_func = NULL; /* use default c version */ + mp_msg(MSGT_DECAUDIO,MSGL_V,"mp3lib: using generic C decore!\n"); } #ifdef CONFIG_FAKE_MONO @@ -535,7 +535,7 @@ int MP3_DecodeFrame(unsigned char *hova,short single){ case 3: do_layer3(&fr,single);break; case 1: do_layer1(&fr,single);break; default: - return 0; // unsupported + return 0; // unsupported } // ++MP3_frames; return pcm_point ? pcm_point : 2; |