Re: PR 15492: floating-point arguments are loaded too early to x87 stack

[Steven Bosscher <stevenb at suse dot de>]

On Saturday 21 August 2004 03:29, James E Wilson wrote:
> Uros Bizjak wrote:
> > Is there everything OK with fsched-insn2 for pentum4 case?
>
> The problem here is that we have no DFA scheduler for pentium4, and we
> also have no default scheduling info.  As a result, the compiler assumes
> that almost everything takes 1 cycle.  This is unlikely to result in a
> good schedule.
>
> Someone will have to write a DFA scheduler for pentium4 if we want good
> code for that target.

Vlad has a pentium4 scheduling description but he said it doesn't help
much at all...


> I say almost everything, because in the ppro.md file, the reservation
> for call instruction is missing the cpu==pentiumpro check, and hence it
> applies to pentium4 also.

Ouch.  It applies to every CPU in i386.md.


>  I suspect this is a bug, but it could be an
> undocumented misfeature.

This is definitely a bug.  Very nice catch, it's a shame that genautomata
does not complain if an insn may match more than one reservation.  The
obvious fix is this:

Index: ppro.md
===================================================================
RCS file: /cvs/gcc/gcc/gcc/config/i386/ppro.md,v
retrieving revision 1.6
diff -c -3 -p -r1.6 ppro.md
*** ppro.md     11 Mar 2004 05:54:30 -0000      1.6
--- ppro.md     21 Aug 2004 10:05:49 -0000
***************
*** 137,143 ****
  ;; on decoder 0, and say that it takes a little while before the result
  ;; is available.
  (define_insn_reservation "ppro_complex_insn" 6
!                        (eq_attr "type" "other,multi,call,callv,str")
                         "decoder0")

  ;; imov with memory operands does not use the integer units.
--- 137,144 ----
  ;; on decoder 0, and say that it takes a little while before the result
  ;; is available.
  (define_insn_reservation "ppro_complex_insn" 6
!                        (and (eq_attr "cpu" "pentiumpro")
!                             (eq_attr "type" "other,multi,call,callv,str"))
                         "decoder0")

  ;; imov with memory operands does not use the integer units.

Gr.
Steven