[Bug target/19506] [4.0 Regression] PovRay produces wrong pictures with -mfpmath=sse -ffast-math.

["uros at kss-loka dot si" <gcc-bugzilla@gcc.gnu.org>]

------- Additional Comments From uros at kss-loka dot si  2005-01-18 15:58 -------
This patch unfortunatelly didn't help:
http://gcc.gnu.org/ml/gcc-cvs/2005-01/msg00651.html

With the patch, tracing abyss.pov generates these warnings:
Parsing.............File: ../scenes/advanced/abyss.pov  Line: 434
Warning: Degenerate CSG bounding box (not used!).

......File: ../scenes/advanced/abyss.pov  Line: 642
Warning: Degenerate CSG bounding box (not used!).

...............................................File:
../scenes/advanced/abyss.pov  Line: 728
Warning: Degenerate CSG bounding box (not used!).

Trying to get a smaller testcase, I have found that this problem can be somehow
fixed by following change in ix86_fp_comparison_codes():

Index: i386.c
===================================================================
RCS file: /cvs/gcc/gcc/gcc/config/i386/i386.c,v
retrieving revision 1.780
diff -u -p -r1.780 i386.c
--- i386.c      18 Jan 2005 11:08:33 -0000      1.780
+++ i386.c      18 Jan 2005 15:51:18 -0000
@@ -8446,6 +8446,8 @@ ix86_fp_comparison_codes (enum rtx_code 
     case LT:                   /* LTU - CF=1 - fails on unordered */
       *first_code = UNLT;
       *bypass_code = UNORDERED;
+      return;
+
       break;
     case LE:                   /* LEU - CF=1 | ZF=1 - fails on unordered */
       *first_code = UNLE;

This change protects *bypass_code from being cleared back to UNKNOWN by
following code at the end of function:
  ...
  if (!TARGET_IEEE_FP)
    {
      *second_code = UNKNOWN;
      *bypass_code = UNKNOWN;
    }
}

It looks that something is wrong with LT compares.

-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=19506