gnu/157025: gcc Bug 28796 - __builtin_nan() and __builtin_unordered() inconsistent

[Pedro Giffuni]

>Number:         157025
>Category:       gnu
>Synopsis:       gcc Bug 28796 - __builtin_nan() and __builtin_unordered() inconsistent
>Confidential:   no
>Severity:       non-critical
>Priority:       low
>Responsible:    freebsd-bugs
>State:          open
>Quarter:        
>Keywords:       
>Date-Required:
>Class:          sw-bug
>Submitter-Id:   current-users
>Arrival-Date:   Sat May 14 02:30:06 UTC 2011
>Closed-Date:
>Last-Modified:
>Originator:     Pedro Giffuni
>Release:        8.2-Release
>Organization:
>Environment:
FreeBSD mogwai.giffuni.net 8.2-RELEASE FreeBSD 8.2-RELEASE #0: Fri Mar 18 15:35:09 UTC 2011     root at mogwai.giffuni.net:/usr/src/sys/amd64/compile/GENERIC  amd64

>Description:
As can be read in the top of Apple's gcc Changelog:
http://opensource.apple.com/source/gcc/gcc-5664/gcc/ChangeLog

The following commit was made:

2006-10-24  Richard Guenther

Radar 5675014
PR middle-end/28796
* builtins.c (fold_builtin_classify): Use HONOR_INFINITIES
and HONOR_NANS instead of MODE_HAS_INFINITIES and MODE_HAS_NANS
for deciding optimizations in consistency with fold-const.c
(fold_builtin_unordered_cmp): Likewise.
_______

Applying this will make our base compiler behave consistently
with -ffinite-math-only and is in accordance to what clang
and newer GCCs do.
>How-To-Repeat:
There is a testcase in gcc bug 28796
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=28796

FWIW, This was reported in gnu/153298 too but it was obviated
because the other patches were more important.

>Fix:
Patch attached.

Patch attached with submission follows:

--- gcc/builtins.c.orig	2007-05-19 02:12:19.000000000 +0000
+++ gcc/builtins.c	2011-05-13 20:58:56.000000000 +0000
@@ -8720,7 +8720,7 @@
   switch (builtin_index)
     {
     case BUILT_IN_ISINF:
-      if (!MODE_HAS_INFINITIES (TYPE_MODE (TREE_TYPE (arg))))
+      if (!HONOR_INFINITIES (TYPE_MODE (TREE_TYPE (arg))))
 	return omit_one_operand (type, integer_zero_node, arg);
 
       if (TREE_CODE (arg) == REAL_CST)
@@ -8736,8 +8736,8 @@
       return NULL_TREE;
 
     case BUILT_IN_FINITE:
-      if (!MODE_HAS_NANS (TYPE_MODE (TREE_TYPE (arg)))
-	  && !MODE_HAS_INFINITIES (TYPE_MODE (TREE_TYPE (arg))))
+      if (!HONOR_NANS (TYPE_MODE (TREE_TYPE (arg)))
+	  && !HONOR_INFINITIES (TYPE_MODE (TREE_TYPE (arg))))
 	return omit_one_operand (type, integer_zero_node, arg);
 
       if (TREE_CODE (arg) == REAL_CST)
@@ -8750,7 +8750,7 @@
       return NULL_TREE;
 
     case BUILT_IN_ISNAN:
-      if (!MODE_HAS_NANS (TYPE_MODE (TREE_TYPE (arg))))
+      if (!HONOR_NANS (TYPE_MODE (TREE_TYPE (arg))))
 	return omit_one_operand (type, integer_zero_node, arg);
 
       if (TREE_CODE (arg) == REAL_CST)
@@ -8833,13 +8833,13 @@
 
   if (unordered_code == UNORDERED_EXPR)
     {
-      if (!MODE_HAS_NANS (TYPE_MODE (TREE_TYPE (arg0))))
+      if (!HONOR_NANS (TYPE_MODE (TREE_TYPE (arg0))))
 	return omit_two_operands (type, integer_zero_node, arg0, arg1);
       return fold_build2 (UNORDERED_EXPR, type, arg0, arg1);
     }
 
-  code = MODE_HAS_NANS (TYPE_MODE (TREE_TYPE (arg0))) ? unordered_code
-						      : ordered_code;
+  code = HONOR_NANS (TYPE_MODE (TREE_TYPE (arg0))) ? unordered_code
+						   : ordered_code;
   return fold_build1 (TRUTH_NOT_EXPR, type,
 		      fold_build2 (code, type, arg0, arg1));
 }


>Release-Note:
>Audit-Trail:
>Unformatted: