|
|
Fixes for errors like:gimp-composite-mmx.c: In function 'gimp_composite_addition_rgba8_rgba8_rgba8_mmx':gimp-composite-mmx.c:94: error: unknown register name '%mm0' in 'asm'gimp-composite-mmx.c:101: error: unknown register name '%mm4' in 'asm'gimp-composite-mmx.c:101: error: unknown register name '%mm3' in 'asm'gimp-composite-mmx.c:101: error: unknown register name '%mm2' in 'asm'gimp-composite-mmx.c:101: error: unknown register name '%mm1' in 'asm'
diff -dur gimp-2.2.8/app/composite/gimp-composite-mmx.c gimp-2.2.8-p/app/composite/gimp-composite-mmx.c
--- gimp-2.2.8/app/composite/gimp-composite-mmx.c 2005-01-11 12:27:25.000000000 +0100
+++ gimp-2.2.8-p/app/composite/gimp-composite-mmx.c 2005-10-22 13:37:05.000000000 +0200
@@ -94,7 +94,7 @@
asm volatile ("movq %0,%%mm0" : /* empty */ : "m" (*rgba8_alpha_mask_64)- : "%mm0");
+ );
for (; n_pixels >= 2; n_pixels -= 2) {@@ -110,7 +110,7 @@
"\tmovq %%mm1, %0\n" : "=m" (*d) : "m" (*a), "m" (*b)- : "%mm0", "%mm1", "%mm2", "%mm3", "%mm4");
+ );
a++; b++; d++;@@ -130,7 +130,7 @@
"\tmovd %%mm1, %0\n" : "=m" (*d) : "m" (*a), "m" (*b)- : "%mm0", "%mm1", "%mm2", "%mm3", "%mm4");
+ );
} asm("emms");@@ -192,7 +192,7 @@
"\tmovq %%mm7,%0\n" : "=m" (*d) : "m" (*a), "m" (*b), "m" (*rgba8_b255_64), "m" (*rgba8_w1_64), "m" (*rgba8_w255_64), "m" (*rgba8_alpha_mask_64)- : pdivwqX_clobber, "%mm0", "%mm1", "%mm2", "%mm3", "%mm4", "%mm5", "%mm6", "%mm7");
+ );
d++; b++; a++;@@ -246,7 +246,7 @@
"\tmovd %%mm7,%0\n" : "=m" (*d) : "m" (*a), "m" (*b), "m" (*rgba8_b255_64), "m" (*rgba8_w1_64), "m" (*rgba8_w255_64), "m" (*rgba8_alpha_mask_64)- : pdivwqX_clobber, "%mm0", "%mm1", "%mm2", "%mm3", "%mm4", "%mm5", "%mm6", "%mm7");
+ : pdivwqX_clobber);
} asm("emms");@@ -269,7 +269,7 @@
"\tmovq %%mm2, %0\n" : "=m" (*d) : "m" (*a), "m" (*b)- : "%mm1", "%mm2", "%mm3", "%mm4");
+ );
a++; b++; d++;@@ -283,7 +283,7 @@
"\tmovd %%mm2, %0\n" : "=m" (*d) : "m" (*a), "m" (*b)- : "%mm2", "%mm3", "%mm4");
+ );
} asm("emms");@@ -297,7 +297,7 @@
uint64 *b = (uint64 *) _op->B; gulong n_pixels = _op->n_pixels; - asm volatile ("movq %0,%%mm0" : : "m" (*rgba8_alpha_mask_64) : "%mm0");
+ asm volatile ("movq %0,%%mm0" : : "m" (*rgba8_alpha_mask_64) );
for (; n_pixels >= 2; n_pixels -= 2) {@@ -316,7 +316,7 @@
"\tmovq %%mm1, %0\n" : "=m" (*d) : "m" (*a), "m" (*b)- : "%mm1", "%mm2", "%mm3", "%mm4");
+ );
a++; b++; d++;@@ -339,7 +339,7 @@
"\tmovd %%mm1, %0\n" : "=m" (*d) : "m" (*a), "m" (*b)- : "%mm1", "%mm2", "%mm3", "%mm4");
+ );
} asm("emms");@@ -358,7 +358,7 @@
"\tmovq %1, %%mm7\n" : : "m" (*rgba8_alpha_mask_64), "m" (*rgba8_w1_64)- : "%mm0", "%mm7");
+ );
for (; n_pixels >= 2; n_pixels -= 2) {@@ -398,7 +398,7 @@
"\tmovq %%mm3,%0\n" : "=m" (*d) : "m" (*a), "m" (*b), "m" (*rgba8_alpha_mask_64)- : pdivwuqX_clobber, "%mm1", "%mm2", "%mm3", "%mm4", "%mm5");
+ : pdivwuqX_clobber);
a++; b++; d++;@@ -442,7 +442,7 @@
"\tmovd %%mm3,%0\n" : "=m" (*d) : "m" (*a), "m" (*b), "m" (*rgba8_alpha_mask_64)- : pdivwuqX_clobber, "%mm1", "%mm2", "%mm3", "%mm4", "%mm5");
+ : pdivwuqX_clobber);
} asm("emms");@@ -495,7 +495,7 @@
"\tmovq %%mm7,%0\n" : "=m" (*d) : "m" (*a), "m" (*b), "m" (*rgba8_w256_64), "m" (*rgba8_alpha_mask_64)- : pdivwuqX_clobber, "%mm0", "%mm1", "%mm2", "%mm3", "%mm4", "%mm5", "%mm6", "%mm7");
+ : pdivwuqX_clobber);
a++; b++; d++;@@ -538,7 +538,7 @@
"\tmovd %%mm7,%2\n" : /* empty */ : "m" (*a), "m" (*b), "m" (*d), "m" (*rgba8_w256_64), "m" (*rgba8_alpha_mask_64)- : pdivwuqX_clobber, "%mm1", "%mm2", "%mm3", "%mm4", "%mm5");
+ : pdivwuqX_clobber);
} asm("emms");@@ -558,7 +558,7 @@
"movq %1,%%mm7\n" : /* no outputs */ : "m" (*rgba8_alpha_mask_64), "m" (*rgba8_w128_64)- : "%mm0", "%mm7", "%mm6");
+ );
for (; n_pixels >= 2; n_pixels -= 2) {@@ -589,7 +589,7 @@
"\tmovq %%mm1,%0\n" : "=m" (*d) : "m" (*a), "m" (*b)- : "%mm1", "%mm2", "%mm3", "%mm4");
+ );
a++; b++; d++;@@ -620,7 +620,7 @@
"\tmovd %%mm1, %0\n" : "=m" (*d) : "m" (*a), "m" (*b)- : "%mm1", "%mm2", "%mm3", "%mm4");
+ );
} asm("emms");@@ -639,7 +639,7 @@
"movq %1, %%mm7\n" : /* empty */ : "m" (*rgba8_alpha_mask_64), "m" (*rgba8_w128_64)- : "%mm0", "%mm6", "%mm7");
+);
for (; n_pixels >= 2; n_pixels -= 2) {@@ -667,7 +668,7 @@
"\tmovq %%mm1, %0\n" : "=m" (*d) : "m" (*a), "m" (*b)- : "%mm1", "%mm2", "%mm3", "%mm4");
+);
a++; b++; d++;@@ -697,7 +699,7 @@
"\tmovd %%mm1, %0\n" : "=m" (*d) : "m" (*a), "m" (*b)- : "%mm1", "%mm2", "%mm3", "%mm4");
+ );
} asm("emms");@@ -711,7 +713,7 @@
uint64 *b = (uint64 *) _op->B; gulong n_pixels = _op->n_pixels; - asm volatile ("movq %0,%%mm0" : : "m" (*rgba8_alpha_mask_64) : "%mm0");
+ asm volatile ("movq %0,%%mm0" : : "m" (*rgba8_alpha_mask_64) );
for (; n_pixels >= 2; n_pixels -= 2) {@@ -727,7 +729,7 @@
"\tmovq %%mm1, %0\n" : "=m" (*d) : "m" (*a), "m" (*b)- : "%mm1", "%mm2", "%mm3", "%mm4", "%mm5");
+ );
a++; b++; d++;@@ -750,7 +752,7 @@
"\tmovd %%mm1, %0\n" : "=m" (*d) : "m" (*a), "m" (*b)- : "%mm1", "%mm2", "%mm3", "%mm4", "%mm5");
+ );
} asm("emms");@@ -770,7 +772,7 @@
"pxor %%mm6,%%mm6\n" : /* empty */ : "m" (*rgba8_alpha_mask_64), "m" (*rgba8_w128_64)- : "%mm6", "%mm7", "%mm0");
+ );
for (; n_pixels >= 2; n_pixels -= 2) {@@ -797,7 +799,7 @@
"\tmovq %%mm1, %0\n" : "=m" (*d) : "m" (*a), "m" (*b)- : "%mm1", "%mm2", "%mm3", "%mm4", "%mm5");
+ );
a++; b++; d++;@@ -824,7 +826,7 @@
"\tmovd %%mm1, %0\n" : "=m" (*d) : "m" (*a), "m" (*b)- : "%mm1", "%mm2", "%mm3", "%mm4", "%mm5");
+ );
} asm("emms");@@ -892,7 +894,7 @@
asm volatile ("pxor %%mm0,%%mm0\n" "movq %0,%%mm7" : /* empty */- : "m" (*rgba8_w128_64) : "%mm0");
+ : "m" (*rgba8_w128_64) );
for (; n_pixels >= 2; n_pixels -= 2) {@@ -943,7 +945,7 @@
"\tmovq %%mm1,%2\n" : "+m" (*a), "+m" (*b), "+m" (*d) : "m" (*rgba8_w2_64), "m" (*rgba8_alpha_mask_64)- : "%mm1", "%mm2", "%mm3", "%mm4");
+ );
a++; b++; d++;@@ -998,7 +1000,7 @@
"\tmovd %%mm1,%0\n" : "=m" (*d) : "m" (*a), "m" (*b), "m" (*rgba8_w2_64), "m" (*rgba8_alpha_mask_64)- : "%mm1", "%mm2", "%mm3", "%mm4");
+ );
} asm("emms");@@ -1024,7 +1026,7 @@
"\tmovq %1,%%mm7\n" : /* empty */ : "m" (_op->scale.scale), "m" (*rgba8_w128_64)- : "%eax", "%ebx", "%mm0", "%mm5", "%mm6", "%mm7");
+ );
for (; n_pixels >= 2; n_pixels -= 2) {@@ -1046,7 +1048,7 @@
"\tmovq %%mm1,%0\n" : "=m" (*d) : "m" (*a)- : "%mm1", "%mm2", "%mm3", "%mm4", "%mm5", "%mm6", "%mm7");
+ );
a++; d++; }@@ -1064,7 +1066,7 @@
"\tmovd %%mm1,%0\n" : "=m" (*d) : "m" (*a)- : "%mm1", "%mm2", "%mm3", "%mm4", "%mm5", "%mm6", "%mm7");
+ );
} asm("emms");@@ -1083,7 +1085,7 @@
"movq %1,%%mm7\n" : /* empty */ : "m" (*rgba8_alpha_mask_64), "m" (*rgba8_w128_64)- : "%mm0", "%mm6", "%mm7");
+ );
for (; n_pixels >= 2; n_pixels -= 2) {@@ -1134,7 +1136,7 @@
"\tmovq %%mm1,%0\n" : "=m" (*d) : "m" (*a), "m" (*b)- : "%mm1", "%mm2", "%mm3", "%mm4", "%mm5");
+ );
a++; b++; d++;@@ -1189,7 +1191,7 @@
"\tmovd %%mm1,%0\n" : "=m" (*d) : "m" (*a), "m" (*b)- : "%mm1", "%mm2", "%mm3", "%mm4", "%mm5");
+ );
} asm volatile ("emms");@@ -1204,7 +1206,7 @@
uint64 *b = (uint64 *) _op->B; gulong n_pixels = _op->n_pixels; - asm volatile ("movq %0,%%mm0" : : "m" (*rgba8_alpha_mask_64) : "%mm0");
+ asm volatile ("movq %0,%%mm0" : : "m" (*rgba8_alpha_mask_64) );
for (; n_pixels >= 2; n_pixels -= 2) {@@ -1224,7 +1226,7 @@
"\tmovq %%mm1,%0\n" : "=m" (*d) : "m" (*a), "m" (*b)- : "%mm1", "%mm2", "%mm3", "%mm4", "%mm5");
+ );
a++; b++; d++;@@ -1248,7 +1250,7 @@
"\tmovd %%mm1,%0\n" : "=m" (*d) : "m" (*a), "m" (*b)- : "%mm1", "%mm2", "%mm3", "%mm4", "%mm5");
+ );
} asm volatile ("emms");@@ -1269,7 +1271,7 @@
"\tmovq %%mm2,%1\n" : "+m" (*a), "+m" (*b) : - : "%mm1", "%mm2", "%mm3", "%mm4");
+ );
a++; b++; }@@ -1282,7 +1284,7 @@
"\tmovd %%mm2,%1\n" : "+m" (*a), "+m" (*b) :- : "%mm1", "%mm2", "%mm3", "%mm4");
+ );
} asm("emms");diff -dur gimp-2.2.8/app/composite/gimp-composite-sse2.c gimp-2.2.8-p/app/composite/gimp-composite-sse2.c
--- gimp-2.2.8/app/composite/gimp-composite-sse2.c 2005-05-06 14:27:50.000000000 +0200
+++ gimp-2.2.8-p/app/composite/gimp-composite-sse2.c 2005-10-22 13:49:02.000000000 +0200
@@ -100,7 +100,7 @@
"\tmovq %1,%%mm0" : /* empty */ : "m" (*rgba8_alpha_mask_128), "m" (*rgba8_alpha_mask_64)- : "%xmm0", "%mm0");
+ );
for (; n_pixels >= 4; n_pixels -= 4) {@@ -117,7 +117,7 @@
"\tmovdqu %%xmm1,%0\n" : "=m" (*D) : "m" (*A), "m" (*B)- : "%xmm0", "%xmm1", "%xmm2", "%xmm3", "%xmm4", "%xmm5", "%xmm6", "%xmm7");
+ );
A++; B++; D++;@@ -141,7 +141,7 @@
"\tmovq %%mm1,%0\n" : "=m" (*d) : "m" (*a), "m" (*b)- : "%mm0", "%mm1", "%mm2", "%mm3", "%mm4", "%mm5", "%mm6", "%mm7");
+ );
a++; b++; d++;@@ -161,7 +161,7 @@
"\tmovd %%mm1,%0\n" : "=m" (*d) : "m" (*a), "m" (*b)- : "%mm0", "%mm1", "%mm2", "%mm3", "%mm4", "%mm5", "%mm6", "%mm7");
+ );
} asm("emms");@@ -195,7 +195,7 @@
"\tmovdqu %%xmm2,%0\n" : "=m" (*D) : "m" (*A), "m" (*B)- : "%xmm1", "%xmm2", "%xmm3", "%xmm4");
+ );
A++; B++; D++;@@ -212,7 +212,7 @@
"\tmovq %%mm2, %0\n" : "=m" (*d) : "m" (*a), "m" (*b)- : "%mm1", "%mm2", "%mm3", "%mm4");
+ );
a++; b++; d++;@@ -226,7 +226,7 @@
"\tmovd %%mm2, %0\n" : "=m" (*d) : "m" (*a), "m" (*b)- : "%mm2", "%mm3", "%mm4");
+ );
} asm("emms");@@ -247,7 +247,7 @@
"\tmovdqu %1,%%xmm0" : /* */ : "m" (*rgba8_alpha_mask_64), "m" (*rgba8_alpha_mask_128)- : "%mm0", "%xmm0");
+ );
for (; n_pixels >= 4; n_pixels -= 4) {@@ -266,7 +266,7 @@
"\tmovdqu %%xmm1,%0\n" : "=m" (*D) : "m" (*A), "m" (*B)- : "%xmm1", "%xmm2", "%xmm3", "%xmm4", "%xmm5");
+ );
A++; B++; D++;@@ -293,7 +293,7 @@
"\tmovq %%mm1, %0\n" : "=m" (*d) : "m" (*a), "m" (*b)- : "%mm1", "%mm2", "%mm3", "%mm4", "%mm5");
+ );
a++; b++; d++;@@ -316,7 +316,7 @@
"\tmovd %%mm1, %0\n" : "=m" (*d) : "m" (*a), "m" (*b)- : "%mm1", "%mm2", "%mm3", "%mm4", "%mm5");
+ );
} asm("emms");@@ -366,7 +366,7 @@
"\tmovdqu %%xmm7,%0\n" : "=m" (*op.D) : "m" (*op.A), "m" (*op.B), "m" (*rgba8_w256_128), "m" (*rgba8_alpha_mask_128)- : "%eax", "%ecx", "%edx", "%xmm0", "%xmm1", "%xmm2", "%xmm3", "%xmm4", "%xmm5", "%xmm6", "%xmm7");
+ );
op.A += 16; op.B += 16; op.D += 16;@@ -409,7 +409,7 @@
"\tmovq %%mm7,%0\n" : (*op.D) : "m" (*op.A), "m" (*op.B), "m" (*rgba8_w256_64), "m" (*rgba8_alpha_mask_64)- : "%eax", "%ecx", "%edx", "%mm0", "%mm1", "%mm2", "%mm3", "%mm4", "%mm5", "%mm6", "%mm7");
+ );
op.A += 8; op.B += 8; op.D += 8;@@ -452,7 +452,7 @@
"\tmovd %%mm7,%0\n" : "=m" (*op.D) : "m" (*op.A), "m" (*op.B), "m" (*rgba8_w256_64), "m" (*rgba8_alpha_mask_64)- : "%eax", "%ecx", "%edx", "%mm1", "%mm2", "%mm3", "%mm4", "%mm5");
+ );
} asm("emms");@@ -478,7 +478,7 @@
"\tmovdqu %3,%%xmm7\n" : /* empty */ : "m" (*rgba8_alpha_mask_64), "m" (*rgba8_w128_64), "m" (*rgba8_alpha_mask_128), "m" (*rgba8_w128_128)- : "%mm0", "%mm6", "%mm7", "%xmm0", "%xmm6", "%xmm7");
+ );
for (; n_pixels >= 4; n_pixels -= 4) {@@ -509,7 +509,7 @@
"\tmovdqu %%xmm1,%0\n" : "=m" (*D) : "m" (*A), "m" (*B)- : "%xmm1", "%xmm2", "%xmm3", "%xmm4");
+ );
A++; B++; D++;@@ -548,7 +548,7 @@
"\tmovq %%mm1,%0\n" : "=m" (*d) : "m" (*a), "m" (*b)- : "%mm1", "%mm2", "%mm3", "%mm4");
+ );
a++; b++; d++;@@ -573,7 +573,7 @@
"\tmovd %%mm1, %0\n" : "=m" (*d) : "m" (*a), "m" (*b)- : "%mm1", "%mm2", "%mm3", "%mm4");
+ );
} asm("emms");@@ -590,7 +590,7 @@
uint128 *B = (uint128 *) _op->B; gulong n_pixels = _op->n_pixels; - asm volatile ("movdqu %0,%%xmm0" : : "m" (*rgba8_alpha_mask_64) : "%xmm0");
+ asm volatile ("movdqu %0,%%xmm0" : : "m" (*rgba8_alpha_mask_64) );
for (; n_pixels >= 4; n_pixels -= 4) {@@ -606,7 +606,7 @@
"\tmovdqu %%xmm1, %0\n" : "=m" (*D) : "m" (*A), "m" (*B)- : "%xmm1", "%xmm2", "%xmm3", "%xmm4");
+ );
A++; B++; D++;@@ -630,7 +630,7 @@
"\tmovq %%mm1, %0\n" : "=m" (*d) : "m" (*a), "m" (*b)- : "%mm1", "%mm2", "%mm3", "%mm4");
+ );
a++; b++; d++;@@ -650,7 +650,7 @@
"\tmovd %%mm1, %0\n" : "=m" (*d) : "m" (*a), "m" (*b)- : "%mm1", "%mm2", "%mm3", "%mm4");
+ );
} asm("emms");@@ -671,7 +671,7 @@
"\tmovdqu %1,%%xmm0\n" : /* empty */ : "m" (*rgba8_alpha_mask_64), "m" (*rgba8_alpha_mask_128)- : "%mm0", "%xmm0");
+ );
for (; n_pixels >= 4; n_pixels -= 4) {@@ -688,7 +688,7 @@
"\tmovdqu %%xmm1,%0\n" : "=m" (*D) : "m" (*A), "m" (*B)- : "%xmm1", "%xmm2", "%xmm3", "%xmm4");
+ );
A++; B++; D++;@@ -712,7 +712,7 @@
"\tmovq %%mm1,%0\n" : "=m" (*d) : "m" (*a), "m" (*b)- : "%mm1", "%mm2", "%mm3", "%mm4");
+ );
a++; b++; d++;@@ -732,7 +732,7 @@
"\tmovd %%mm1,%0\n" : "=m" (*d) : "m" (*a), "m" (*b)- : "%mm1", "%mm2", "%mm3", "%mm4");
+ );
} asm("emms");@@ -772,7 +772,6 @@
"+m" (op.A[2]), "+m" (op.B[2]), "+m" (op.A[3]), "+m" (op.B[3]) : /* empty */- : "%xmm0", "%xmm1", "%xmm2", "%xmm3", "%xmm4", "%xmm5", "%xmm6", "%xmm7"
); #else asm volatile (" movdqu %0,%%xmm0\n"@@ -828,7 +827,7 @@
"\tmovdqu %%xmm2,%1\n" : "+m" (*op.A), "+m" (*op.B) : /* empty */- : "%xmm2", "%xmm3");
+ );
op.A += 16; op.B += 16; }@@ -841,7 +840,7 @@
"\tmovq %%mm2,%1\n" : "+m" (*op.A), "+m" (*op.B) : /* empty */- : "%mm2", "%mm3");
+ );
op.A += 8; op.B += 8; }@@ -854,7 +853,7 @@
"\tmovd %%mm2,%1\n" : "+m" (*op.A), "+m" (*op.B) : /* empty */- : "%mm1", "%mm2", "%mm3", "%mm4");
+ );
} asm("emms");diff -dur gimp-2.2.8/app/composite/gimp-composite-sse.c gimp-2.2.8-p/app/composite/gimp-composite-sse.c
--- gimp-2.2.8/app/composite/gimp-composite-sse.c 2005-01-08 00:58:33.000000000 +0100
+++ gimp-2.2.8-p/app/composite/gimp-composite-sse.c 2005-10-22 13:42:12.000000000 +0200
@@ -72,7 +72,7 @@
asm volatile ("movq %0,%%mm0" : /* empty */ : "m" (*rgba8_alpha_mask_64)- : "%mm0");
+ );
for (; n_pixels >= 2; n_pixels -= 2) {@@ -88,7 +88,7 @@
"\tmovq %%mm1, %0\n" : "=m" (*d) : "m" (*a), "m" (*b)- : "%mm0", "%mm1", "%mm2", "%mm3", "%mm4");
+ );
a++; b++; d++;@@ -108,7 +108,7 @@
"\tmovd %%mm1, %0\n" : "=m" (*d) : "m" (*a), "m" (*b)- : "%mm0", "%mm1", "%mm2", "%mm3", "%mm4");
+ );
} asm("emms");@@ -171,7 +171,7 @@
"\tmovq %%mm7,%0\n" : "=m" (*d) : "m" (*a), "m" (*b), "m" (*rgba8_b255_64), "m" (*rgba8_w1_64), "m" (*rgba8_w255_64), "m" (*rgba8_alpha_mask_64)- : pdivwqX_clobber, "%mm0", "%mm1", "%mm2", "%mm3", "%mm4", "%mm5", "%mm6", "%mm7");
+ : pdivwqX_clobber);
d++; b++; a++;@@ -225,7 +225,7 @@
"\tmovd %%mm7,%0\n" : "=m" (*d) : "m" (*a), "m" (*b), "m" (*rgba8_b255_64), "m" (*rgba8_w1_64), "m" (*rgba8_w255_64), "m" (*rgba8_alpha_mask_64)- : pdivwqX_clobber, "%mm0", "%mm1", "%mm2", "%mm3", "%mm4", "%mm5", "%mm6", "%mm7");
+ : pdivwqX_clobber);
} asm("emms");@@ -248,7 +248,7 @@
"\tmovq %%mm2, %0\n" : "=m" (*d) : "m" (*a), "m" (*b)- : "%mm1", "%mm2", "%mm3", "%mm4");
+ );
a++; b++; d++;@@ -262,7 +262,7 @@
"\tmovd %%mm2, %0\n" : "=m" (*d) : "m" (*a), "m" (*b)- : "%mm2", "%mm3", "%mm4");
+ );
} asm("emms");@@ -276,7 +276,7 @@
uint64 *b = (uint64 *) _op->B; gulong n_pixels = _op->n_pixels; - asm volatile ("movq %0,%%mm0" : : "m" (*rgba8_alpha_mask_64) : "%mm0");
+ asm volatile ("movq %0,%%mm0" : : "m" (*rgba8_alpha_mask_64) );
for (; n_pixels >= 2; n_pixels -= 2) {@@ -295,7 +295,7 @@
"\tmovq %%mm1, %0\n" : "=m" (*d) : "m" (*a), "m" (*b)- : "%mm1", "%mm2", "%mm3", "%mm4");
+ );
a++; b++; d++;@@ -318,7 +318,7 @@
"\tmovd %%mm1, %0\n" : "=m" (*d) : "m" (*a), "m" (*b)- : "%mm1", "%mm2", "%mm3", "%mm4");
+ );
} asm("emms");@@ -334,7 +334,7 @@
"\tmovq %1, %%mm7\n" : : "m" (*rgba8_alpha_mask_64), "m" (*rgba8_w1_64)- : "%mm0", "%mm7");
+ );
for (; op.n_pixels >= 2; op.n_pixels -= 2) {@@ -374,7 +374,7 @@
"\tmovq %%mm3,%0\n" : "=m" (*op.D) : "m" (*op.A), "m" (*op.B), "m" (*rgba8_alpha_mask_64)- : "%eax", "%ecx", "%edx", "%mm1", "%mm2", "%mm3", "%mm4", "%mm5");
+ );
op.A += 8; op.B += 8; op.D += 8;@@ -419,7 +419,7 @@
"\tmovd %%mm3,%0\n" : "=m" (*op.D) : "m" (*op.A), "m" (*op.B), "m" (*rgba8_alpha_mask_64)- : "%eax", "%ecx", "%edx", "%mm1", "%mm2", "%mm3", "%mm4", "%mm5");
+ );
} asm("emms");@@ -469,7 +469,7 @@
"\tmovq %%mm7,%0\n" : "=m" (*op.D) : "m" (*op.A), "m" (*op.B), "m" (*rgba8_w256_64), "m" (*rgba8_alpha_mask_64)- : "%eax", "%ecx", "%edx", "%mm0", "%mm1", "%mm2", "%mm3", "%mm4", "%mm5", "%mm6", "%mm7");
+ );
op.A += 8; op.B += 8; op.D += 8;@@ -512,7 +512,7 @@
"\tmovd %%mm7,%2\n" : "=m" (*op.D) : "m" (*op.A), "m" (*op.B), "m" (*rgba8_w256_64), "m" (*rgba8_alpha_mask_64)- : "%eax", "%ecx", "%edx", "%mm1", "%mm2", "%mm3", "%mm4", "%mm5");
+ );
} asm("emms");@@ -533,7 +533,7 @@
"\tmovq %1,%%mm7\n" : /* empty */ : "m" (*rgba8_alpha_mask_64), "m" (*rgba8_w128_64)- : "%mm0", "%mm6", "%mm7");
+ );
for (; n_pixels >= 2; n_pixels -= 2)@@ -565,7 +565,7 @@
"\tmovq %%mm1,%0\n" : "=m" (*d) : "m" (*a), "m" (*b)- : "%mm1", "%mm2", "%mm3", "%mm4");
+ );
a++; b++; d++;@@ -597,7 +597,7 @@
"\tmovd %%mm1, %0\n" : "=m" (*d) : "m" (*a), "m" (*b)- : "%mm1", "%mm2", "%mm3", "%mm4");
+ );
} asm("emms");@@ -616,7 +616,7 @@
"movq %1, %%mm7\n" : /* empty */ : "m" (*rgba8_alpha_mask_64), "m" (*rgba8_w128_64)- : "%mm0", "%mm6", "%mm7");
+ );
for (; n_pixels >= 2; n_pixels -= 2) {@@ -644,7 +644,7 @@
"\tmovq %%mm1, %0\n" : "=m" (*d) : "m" (*a), "m" (*b)- : "%mm1", "%mm2", "%mm3", "%mm4");
+ );
a++; b++; d++;@@ -675,7 +675,7 @@
"\tmovd %%mm1, %0\n" : "+m" (*d) : "m" (*a), "m" (*b)- : "%mm1", "%mm2", "%mm3", "%mm4");
+ );
} asm("emms");@@ -689,7 +689,7 @@
uint64 *b = (uint64 *) _op->B; gulong n_pixels = _op->n_pixels; - asm volatile ("movq %0,%%mm0" : : "m" (*rgba8_alpha_mask_64) : "%mm0");
+ asm volatile ("movq %0,%%mm0" : : "m" (*rgba8_alpha_mask_64) );
for (; n_pixels >= 2; n_pixels -= 2) {@@ -705,7 +705,7 @@
"\tmovq %%mm1, %0\n" : "=m" (*d) : "m" (*a), "m" (*b)- : "%mm1", "%mm2", "%mm3", "%mm4", "%mm5");
+ );
a++; b++; d++;@@ -728,7 +728,7 @@
"\tmovd %%mm1, %0\n" : "=m" (*d) : "m" (*a), "m" (*b)- : "%mm1", "%mm2", "%mm3", "%mm4", "%mm5");
+ );
} asm("emms");@@ -742,9 +742,9 @@
uint64 *b = (uint64 *) _op->B; gulong n_pixels = _op->n_pixels; - asm volatile ("movq %0,%%mm0" : : "m" (*rgba8_alpha_mask_64) : "%mm0");
- asm volatile ("movq %0,%%mm7" : : "m" (*rgba8_w128_64) : "%mm7");
- asm volatile ("pxor %%mm6,%%mm6" : : : "%mm6");
+ asm volatile ("movq %0,%%mm0" : : "m" (*rgba8_alpha_mask_64) );
+ asm volatile ("movq %0,%%mm7" : : "m" (*rgba8_w128_64) );
+ asm volatile ("pxor %%mm6,%%mm6" : : );
for (; n_pixels >= 2; n_pixels -= 2) {@@ -771,7 +771,7 @@
"\tmovq %%mm1, %0\n" : "=m" (*d) : "m" (*a), "m" (*b)- : "%mm1", "%mm2", "%mm3", "%mm4", "%mm5");
+ );
a++; b++; d++;@@ -798,7 +798,7 @@
"\tmovd %%mm1, %0\n" : "=m" (*d) : "m" (*a), "m" (*b)- : "%mm1", "%mm2", "%mm3", "%mm4", "%mm5");
+ );
} asm("emms");@@ -993,7 +993,7 @@
"\tmovq %1,%%mm7\n" : /* empty */ : "m" (_op->scale.scale), "m" (*rgba8_w128_64)- : "%eax", "%ebx", "%mm0", "%mm3", "%mm5", "%mm6", "%mm7");
+ );
for (; n_pixels >= 2; n_pixels -= 2) {@@ -1015,7 +1015,7 @@
"\tmovq %%mm1,%0\n" : "=m" (*d) : "m" (*a)- : "%mm1", "%mm2", "%mm4", "%mm5", "%mm7");
+ );
a++; d++; }@@ -1033,7 +1033,7 @@
"\tmovd %%mm1,%0\n" : "=m" (*d) : "m" (*a)- : "%mm1", "%mm2", "%mm4", "%mm5", "%mm6", "%mm7");
+ );
} asm("emms");@@ -1047,8 +1047,8 @@
uint64 *b = (uint64 *) _op->B; gulong n_pixels = _op->n_pixels; - asm volatile ("movq %0,%%mm0" : : "m" (*rgba8_alpha_mask_64) : "%mm0");
- asm volatile ("movq %0,%%mm7" : : "m" (*rgba8_w128_64) : "%mm7");
+ asm volatile ("movq %0,%%mm0" : : "m" (*rgba8_alpha_mask_64) );
+ asm volatile ("movq %0,%%mm7" : : "m" (*rgba8_w128_64) );
asm volatile ("pxor %mm6, %mm6"); for (; n_pixels >= 2; n_pixels -= 2)@@ -1100,7 +1100,7 @@
"\tmovq %%mm1,%0\n" : "=m" (*d) : "m" (*a), "m" (*b)- : "%mm1", "%mm2", "%mm3", "%mm4", "%mm5");
+ );
a++; b++; d++;@@ -1155,7 +1155,7 @@
"\tmovd %%mm1,%0\n" : "=m" (*d) : "m" (*a), "m" (*b)- : "%mm1", "%mm2", "%mm3", "%mm4", "%mm5");
+ );
} asm("emms");@@ -1170,7 +1170,7 @@
uint64 *b = (uint64 *) _op->B; gulong n_pixels = _op->n_pixels; - asm volatile ("movq %0,%%mm0" : : "m" (*rgba8_alpha_mask_64) : "%mm0");
+ asm volatile ("movq %0,%%mm0" : : "m" (*rgba8_alpha_mask_64) );
for (; n_pixels >= 2; n_pixels -= 2) {@@ -1186,7 +1186,7 @@
"\tmovq %%mm1,%0\n" : "=m" (*d) : "m" (*a), "m" (*b)- : "%mm1", "%mm2", "%mm3", "%mm4");
+ );
a++; b++; d++;@@ -1206,7 +1206,7 @@
"\tmovd %%mm1,%0\n" : "=m" (*d) : "m" (*a), "m" (*b)- : "%mm1", "%mm2", "%mm3", "%mm4");
+ );
} asm("emms");@@ -1227,7 +1227,7 @@
"\tmovq %%mm2,%1\n" : "+m" (*a), "+m" (*b) :- : "%mm1", "%mm2", "%mm3", "%mm4");
+ );
a++; b++; }@@ -1240,7 +1240,7 @@
"\tmovd %%mm2,%1\n" : "+m" (*a), "+m" (*b) : /* empty */- : "%mm1", "%mm2", "%mm3", "%mm4");
+ );
} asm("emms");@@ -1390,7 +1390,7 @@
"\tmovq %%mm7,(%2); addl $8,%2\n" : "+r" (op.A), "+r" (op.B), "+r" (op.D) : "m" (*va8_b255), "m" (*va8_w1), "m" (*va8_w255), "m" (*va8_alpha_mask)- : "%mm1", "%mm2", "%mm3", "%mm4");
+ );
} if (op.n_pixels)@@ -1441,7 +1441,7 @@
"\tmovd %%mm7,(%2)\n" : /* empty */ : "r" (op.A), "r" (op.B), "r" (op.D), "m" (*va8_b255), "m" (*va8_w1), "m" (*va8_w255), "m" (*va8_alpha_mask)- : "%mm0", "%mm1", "%mm2", "%mm3", "%mm4", "%mm5", "%mm6", "%mm7");
+ );
} asm("emms");
|
