i386 performance problem due to local-alloc.c change

Jeffrey A Law law@cygnus.com
Wed Jun 24 00:16:00 GMT 1998


  In message < 199806240650.CAA13269@jwlab.FEITH.COM >you write:
  > I seem to be seeing around a 10 percent performance hit when
  > running the crafty chess benchmark and the Brl-Cad raytracing
  > benchmarks on an Intel Pentium II with:
  > 
  > Tue Jun  9 12:36:16 1998  Jeffrey A Law  (law@cygnus.com)
  > 
  > 	* Makefile.in (local-alloc.o): Depend on insn-attr.h.
  > 	* local-alloc.c (block_alloc): Avoid creating false 
  > 	dependencies for targets which use instruction scheduling.
  > 
  > installed.  If I remove this patch then life gets better (not to
  > mention faster :-).
Hmmm, I thought the x86 didn't do any local register allocation due to
the definition to CLASS_LIKELY_SPILLED_P.  So I'm a little suprised.

One thought would be to disable that code if SMALL_REGISTER_CLASSES
is nonzero.  That's probably the right thing to do.

jeff

ps.  It's as much as a 25% improvement for ugly Fortran code on my PAs :-)




More information about the Gcc-patches mailing list