[Ffmpeg-devel] fft_3dn2.c compile error

[Reimar Doeffinger]

Hello,
On Tue, Aug 08, 2006 at 11:47:00PM -0700, Loren Merritt wrote:
> Fixed sorta. Does anyone see a workaround that isn't slower?

No, I don't really. But just to collected ideas I attached the variant I
first thought of for the second part. I didn't test it at all, and it
certainly is incorrect if n is not divisible by 2.

Greetings,
Reimar Doeffinger
-------------- next part --------------
Index: libavcodec/i386/fft_3dn2.c
===================================================================
--- libavcodec/i386/fft_3dn2.c	(revision 5966)
+++ libavcodec/i386/fft_3dn2.c	(working copy)
@@ -199,9 +199,6 @@
         asm volatile(
             "movq         %0, %%mm0 \n\t"
             "pswapd       %1, %%mm1 \n\t"
-            ::"m"(z[k]), "m"(z[-1-k])
-        );
-        asm volatile(
             "movq      %%mm0, %%mm2 \n\t"
             "pxor      %%mm7, %%mm2 \n\t"
             "punpckldq %%mm1, %%mm2 \n\t"
@@ -210,13 +207,12 @@
             "pswapd    %%mm0, %%mm4 \n\t"
             "pxor      %%mm7, %%mm0 \n\t"
             "pxor      %%mm7, %%mm4 \n\t"
-            "movq      %%mm0, %0    \n\t" // { -z[n8+k].im, z[n8-1-k].re }
-            "movq      %%mm4, %1    \n\t" // { -z[n8-1-k].re, z[n8+k].im }
-            "movq      %%mm2, %2    \n\t" // { -z[n8+k].re, z[n8-1-k].im }
-            "movq      %%mm3, %3    \n\t" // { z[n8-1-k].im, -z[n8+k].re }
-            :"=m"(output[2*k]), "=m"(output[n2-2-2*k]),
-             "=m"(output[n2+2*k]), "=m"(output[n-2-2*k])
-            ::"memory"
+            "movq      %%mm0, (%2)    \n\t" // { -z[n8+k].im, z[n8-1-k].re }
+            "movq      %%mm4, (%3)    \n\t" // { -z[n8-1-k].re, z[n8+k].im }
+            "movq      %%mm2, (%2, %4)\n\t" // { -z[n8+k].re, z[n8-1-k].im }
+            "movq      %%mm3, (%3, %4)\n\t" // { z[n8-1-k].im, -z[n8+k].re }
+            ::"m"(z[k]), "m"(z[-1-k]), "r"(&output[2*k]), "r"(&output[n2-2-2*k]), "r"(n2)
+            :"memory"
         );
     }
     asm volatile("femms");