[3.3 PATCH] Fix PR optimization/12510

[Eric Botcazou <ebotcazou at libertysurf dot fr>]

Hi,

This is the ICE caused by the wrong ordering of splitting passes on x86, a 
regression introduced by the following patch:
   http://gcc.gnu.org/ml/gcc-patches/2003-02/msg02260.html

Basically one splitting pass is missing at -O1 on x86 between flow2 and 
reg-stack. This was detected and fixed on mainline by Jakub with:
   http://gcc.gnu.org/ml/gcc-patches/2003-09/msg00457.html
but it is still present on the 3.3 branch as of today.

I find Jakub's patch convoluted so I wrote my own version for the branch.
Bootstrapped/regtested on i586-redhat-linux-gnu (3.3 branch except Ada).


2003-10-07 Eric Botcazou  <ebotcazou@libertysurf.fr>

        PR optimization/12510
        * toplev.c (rest_of_compilation) [insns_split_for_shed2]: New
	automatic variable to record whether insns were split for sched2.
	Split only when optimizing before flow2 when STACK_REGS is defined.
	Split before reg-stack if insns were not split for sched2.


2003-10-07  Volker Reichelt <reichelt@igpm.rwth-aachen.de>

	* gcc.dg/20031007-1.c: New test.


-- 
Eric Botcazou

Index: toplev.c
===================================================================
RCS file: /cvs/gcc/gcc/gcc/toplev.c,v
retrieving revision 1.690.2.22
diff -u -p -r1.690.2.22 toplev.c
--- toplev.c	17 Sep 2003 07:17:09 -0000	1.690.2.22
+++ toplev.c	7 Oct 2003 09:48:33 -0000
@@ -2364,6 +2364,7 @@ rest_of_compilation (decl)
   int failure = 0;
   int rebuild_label_notes_after_reload;
   int register_life_up_to_date;
+  int insns_split_for_shed2;
 
   timevar_push (TV_REST_OF_COMPILATION);
 
@@ -3361,9 +3362,7 @@ rest_of_compilation (decl)
 #endif
 
   /* If optimizing, then go ahead and split insns now.  */
-#ifndef STACK_REGS
   if (optimize > 0)
-#endif
     split_all_insns (0);
 
   if (optimize)
@@ -3436,6 +3435,8 @@ rest_of_compilation (decl)
       timevar_pop (TV_IFCVT2);
     }
 
+  insns_split_for_shed2 = 0;
+
 #ifdef INSN_SCHEDULING
   if (optimize > 0 && flag_schedule_insns_after_reload)
     {
@@ -3446,6 +3447,7 @@ rest_of_compilation (decl)
 	 and write some more of the results to dump file.  */
 
       split_all_insns (1);
+      insns_split_for_shed2 = 1;
 
       schedule_insns (rtl_dump_file);
 
@@ -3464,6 +3466,11 @@ rest_of_compilation (decl)
 #ifdef STACK_REGS
   timevar_push (TV_REG_STACK);
   open_dump_file (DFI_stack, decl);
+
+  /* We have to split now because we won't split
+     before shorten_branches below.  */
+  if (! insns_split_for_shed2)
+    split_all_insns (optimize > 0);
 
   reg_to_stack (insns, rtl_dump_file);
 

/* PR optimization/12510 */
/* Origin: Lars Skovlund <lskovlun@image.dk>  */
/* Reduced testcase by Volker Reichelt <reichelt@igpm.rwth-aachen.de> */

/* Verify that one splitting pass is not missing on x86 at -O1 */

/* { dg-do compile } */
/* { dg-options "-O -mcpu=i686" { target i?86-*-* x86_64-*-* } } */

extern foo(double);

void bar(double x, double y)
{
  foo (x);
  if (y) x = y ? 0 : 1/y;
  else if (y) x = y < 1 ? 1 : y;
  else x = 1/y < 1 ? 1 : x;
  foo (x);
}