Michael Niedermayr gave me the hint, that gcc might address memory
operands through %esp, so modifying the stack pointer with push/pop
is not gueranteed to work under all conditions.
One way would be, to backup the register to the stack manually, without
modifiying %esp, but this might not be necessary.
#if defined(PIC) && !defined(ARCH_X86_64)
"push %%"REG_a" \n\t"
"movl %2, %%"REG_a" \n\t"
"movq (%%"REG_a"), %%mm5 \n\t" // qmat[0]
"pop %%"REG_a" \n\t"
#else
"movq (%2), %%mm5 \n\t" // qmat[0]
#endif
"pxor %%mm6, %%mm6 \n\t"
"psubw (%3), %%mm6 \n\t" // -bias[0]
"mov $-128, %%"REG_a" \n\t"
push/pop seem to be not necessary at all here, because REG_a gets
loaded with $-128 and isn't used before, therefore no need to back
it up. This can be reduced to (couldn't test this yet):
#if defined(PIC) && !defined(ARCH_X86_64)
"movl %2, %%"REG_a" \n\t"
"movq (%%"REG_a"), %%mm5 \n\t" // qmat[0]
#else
"movq (%2), %%mm5 \n\t" // qmat[0]
#endif
"pxor %%mm6, %%mm6 \n\t"
"psubw (%3), %%mm6 \n\t" // -bias[0]
"mov $-128, %%"REG_a" \n\t"
In my previous suggestion to fix the bug in the code, usage of REG_b
and push/pop can be avoided this way (tested and working):
#if defined(PIC) && !defined(ARCH_X86_64)
"addl %2, %%"REG_a" \n\t"
"movq (%%"REG_a"), %%mm5 \n\t" // qmat[i]
"subl %2, %%"REG_a" \n\t"
#else
"movq (%2, %%"REG_a"), %%mm5 \n\t" // qmat[i]
#endif
Instead of loading %2 to reg_b and adding reg_b to reg_a to
refrence qmat[i], %2 can be added to reg_a directly and subtracted
again from it afterwards.
The complete diff to the upstream sources and to the previous debian
package is attached.
bye,
Tobias
diff -Nur orig/libavcodec/i386/mpegvideo_mmx_template.c ffmpeg-0.cvs20050626/libavcodec/i386/mpegvideo_mmx_template.c
--- orig/libavcodec/i386/mpegvideo_mmx_template.c 2005-07-17 10:39:07.000000000 +0200
+++ ffmpeg-0.cvs20050626/libavcodec/i386/mpegvideo_mmx_template.c 2005-07-17 12:14:50.000000000 +0200
@@ -96,10 +96,8 @@
"pxor %%mm7, %%mm7 \n\t" // 0
"pxor %%mm4, %%mm4 \n\t" // 0
#if defined(PIC) && !defined(ARCH_X86_64)
- "push %%"REG_a" \n\t"
"movl %2, %%"REG_a" \n\t"
"movq (%%"REG_a"), %%mm5 \n\t" // qmat[0]
- "pop %%"REG_a" \n\t"
#else
"movq (%2), %%mm5 \n\t" // qmat[0]
#endif
@@ -169,12 +167,11 @@
"movq (%3, %%"REG_a"), %%mm6 \n\t" // bias[0]
"paddusw %%mm6, %%mm0 \n\t" // ABS(block[i]) + bias[0]
#if defined(PIC) && !defined(ARCH_X86_64)
- "push %%"REG_a" \n\t"
- "movl %2, %%"REG_a" \n\t"
- "movq (%%"REG_a", %%"REG_a"), %%mm5 \n\t" // qmat[i]
- "pop %%"REG_a" \n\t"
+ "addl %2, %%"REG_a" \n\t"
+ "movq (%%"REG_a"), %%mm5 \n\t" // qmat[i]
+ "subl %2, %%"REG_a" \n\t"
#else
- "movq (%2, %%"REG_a"), %%mm5 \n\t" // qmat[i]
+ "movq (%2, %%"REG_a"), %%mm5 \n\t" // qmat[i]
#endif
"pmulhw %%mm5, %%mm0 \n\t" // (ABS(block[i])*qmat[0] + bias[0]*qmat[0])>>16
"por %%mm0, %%mm4 \n\t"
--- ffmpeg-0.cvs20050626.orig/libavcodec/i386/mpegvideo_mmx_template.c
+++ ffmpeg-0.cvs20050626/libavcodec/i386/mpegvideo_mmx_template.c
@@ -95,7 +95,12 @@
SPREADW(%%mm3)
"pxor %%mm7, %%mm7 \n\t" // 0
"pxor %%mm4, %%mm4 \n\t" // 0
+#if defined(PIC) && !defined(ARCH_X86_64)
+ "movl %2, %%"REG_a" \n\t"
+ "movq (%%"REG_a"), %%mm5 \n\t" // qmat[0]
+#else
"movq (%2), %%mm5 \n\t" // qmat[0]
+#endif
"pxor %%mm6, %%mm6 \n\t"
"psubw (%3), %%mm6 \n\t" // -bias[0]
"mov $-128, %%"REG_a" \n\t"
@@ -128,7 +133,11 @@
"movd %%mm3, %%"REG_a" \n\t"
"movzb %%al, %%"REG_a" \n\t" // last_non_zero_p1
: "+a" (last_non_zero_p1)
+#if defined(PIC) && !defined(ARCH_X86_64)
+ : "r" (block+64), "m" (qmat), "r" (bias),
+#else
: "r" (block+64), "r" (qmat), "r" (bias),
+#endif
"r" (inv_zigzag_direct16+64), "r" (temp_block+64)
);
// note the asm is split cuz gcc doesnt like that many operands ...
@@ -157,7 +166,13 @@
"psubw %%mm1, %%mm0 \n\t" // ABS(block[i])
"movq (%3, %%"REG_a"), %%mm6 \n\t" // bias[0]
"paddusw %%mm6, %%mm0 \n\t" // ABS(block[i]) + bias[0]
- "movq (%2, %%"REG_a"), %%mm5 \n\t" // qmat[i]
+#if defined(PIC) && !defined(ARCH_X86_64)
+ "addl %2, %%"REG_a" \n\t"
+ "movq (%%"REG_a"), %%mm5 \n\t" // qmat[i]
+ "subl %2, %%"REG_a" \n\t"
+#else
+ "movq (%2, %%"REG_a"), %%mm5 \n\t" // qmat[i]
+#endif
"pmulhw %%mm5, %%mm0 \n\t" // (ABS(block[i])*qmat[0] + bias[0]*qmat[0])>>16
"por %%mm0, %%mm4 \n\t"
"pxor %%mm1, %%mm0 \n\t"
@@ -179,7 +194,11 @@
"movd %%mm3, %%"REG_a" \n\t"
"movzb %%al, %%"REG_a" \n\t" // last_non_zero_p1
: "+a" (last_non_zero_p1)
+#if defined(PIC) && !defined(ARCH_X86_64)
+ : "r" (block+64), "m" (qmat+64), "r" (bias+64),
+#else
: "r" (block+64), "r" (qmat+64), "r" (bias+64),
+#endif
"r" (inv_zigzag_direct16+64), "r" (temp_block+64)
);
// note the asm is split cuz gcc doesnt like that many operands ...
