Re: [AD] x color conversion again

["Eric Botcazou" <ebotcazou@xxxxxxxxxx>]

> After examining a bit the code, I found the possible reason for that
> "reserve for one push" comment; it's a fake push/pop pair in the 16->24
> i386 converter. Attached is a new version of the patch which should fix
> the issue. It's untested though, so it'd be nice if someone with a gfx
> card supporting 24 bpp modes could test it...

@@ -2344,7 +2356,7 @@
       /* 100% Pentium pairable loop */
       /* 22 cycles = 20 cycles/4 pixels + 1 cycle stack + 1 cycle loop */
       next_block_16_to_24_no_mmx:
-         movl %ecx, -16(%esp)           /* fake pushl %ecx
*/
+         movl %ecx, (%esp)              /* fake pushl %ecx
*/
          xorl %ebx, %ebx
          xorl %eax, %eax
          movb 7(%esi), %bl              /* bl = high byte pixel4
*/

I think the patch is wrong because the %edx value pushed on the stack is
overwritten.

The original code was:

     pushl %edx

      _align_
      /* 100% Pentium pairable loop */
      /* 22 cycles = 20 cycles/4 pixels + 1 cycle stack + 1 cycle loop */
      next_block_16_to_24_no_mmx:
         movl %ecx, -16(%esp)           /* fake pushl %ecx
*/

This worked because, while MYLOCAL3 was -16(%esp), the "push" means that the
value was actually written at -20(%esp), which was free.

So I think the correct solution is to get rid of the "push" and use another
spill slot.  Which means that 2 spill slots would be needed for the 24-bit
code, thus forcing the frame to be 32 bytes large.

--
Eric Botcazou