PING (expand^3, tree-ssa)

Paolo Bonzini paolo.bonzini@gmail.com
Sun Jun 28 14:04:00 GMT 2009


Richard Guenther wrote:
> On Fri, Jun 26, 2009 at 11:47 PM, Paolo Bonzini<bonzini@gnu.org> wrote:
>> [PATCH] post-cond-optab improvements
>> http://gcc.gnu.org/ml/gcc-patches/2009-05/msg00815.html
>> http://gcc.gnu.org/ml/gcc-patches/2009-05/msg00734.html
>> http://gcc.gnu.org/ml/gcc-patches/2009-05/msg00733.html
> 
> These are all ok if the still bootstrap / regtest ok. 

They do with this squashed in:

Index: gcc/dojump.c
===================================================================
--- gcc/dojump.c	(branch cond-optab2)
+++ gcc/dojump.c	(working copy)
@@ -799,7 +799,8 @@ do_compare_rtx_and_jump (rtx op0, rtx op
       implement it.  */
    if ((! if_true_label
         || ! can_compare_p (code, mode, ccp_jump))
-      && (code == ORDERED || code == UNORDERED
+      && (! FLOAT_MODE_P (mode)
+	  || code == ORDERED || code == UNORDERED
  	  || (! HONOR_NANS (mode) && (code == LTGT || code == UNEQ))
  	  || (! HONOR_SNANS (mode) && (code == EQ || code == NE))))
      {

It's an obvious bugfix and it is triggered only by 
builtin-unreachable-2.c.  This is quite expected; __builtin_unreachable 
is is the only case in which we do something special at RTL-expansion 
time that could cause additional jump threading.  Any other missed 
optimization with integers would be a missed jump threading on the 
tree-level.

I'll rerun regtesting with the above change and commit.

Paolo



More information about the Gcc-patches mailing list