[Bug rtl-optimization/19680] sub-optimial register allocation with sse

["tbptbp at gmail dot com" <gcc-bugzilla at gcc dot gnu dot org>]

------- Additional Comments From tbptbp at gmail dot com  2005-01-31 20:18 -------
-fno-gcse is a godsend, instant speedup and most of the sillyness when inlining
is gone.

Now i've applied both your patches, and while there's promising they also
triggers their own nastyness; gcc is so fond of memory inputs that it dumps
stuff on the stack only to address them some instructions latter (well, that's
my interpretation :).

For example,
  4010c3:       0f 28 6c 13 30          movaps 0x30(%ebx,%edx,1),%xmm5
  4010c8:       0f 28 f9                movaps %xmm1,%xmm7
  4010cb:       0f 28 cb                movaps %xmm3,%xmm1
  4010ce:       0f 29 6c 24 10          movaps %xmm5,0x10(%esp)
  4010d3:       0f 59 ce                mulps  %xmm6,%xmm1
  4010d6:       0f 59 c4                mulps  %xmm4,%xmm0
  4010d9:       0f 28 6c 16 30          movaps 0x30(%esi,%edx,1),%xmm5
  4010de:       0f 59 5c 24 10          mulps  0x10(%esp),%xmm3

or
  40119d:       0f c2 c1 01             cmpltps %xmm1,%xmm0
  4011a1:       0f 29 04 24             movaps %xmm0,(%esp)
  4011a5:       0f 28 c5                movaps %xmm5,%xmm0
  4011a8:       0f c2 c1 01             cmpltps %xmm1,%xmm0
  4011ac:       0f 28 c8                movaps %xmm0,%xmm1
  4011af:       0f 56 0c 24             orps   (%esp),%xmm1

Other than those quirks, it looks better to me.

Just to be sure i've tried that patched version on my app, and it's slower than
the unpatched version (both with -fno-gcse).


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=19680