[Bug target/71460] Copying structs can trap (on x86-32) due to SNaN to QNaN

["ubizjak at gmail dot com" <gcc-bugzilla at gcc dot gnu dot org>]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=71460

--- Comment #9 from UroÅ Bizjak <ubizjak at gmail dot com> ---
(In reply to joseph@codesourcery.com from comment #8)
> It's not correct to use flag_signaling_nans for a fix.  
> flag_signaling_nans is only for cases where a bit-pattern for a signaling 
> NaN is interpreted as a floating-point value.  It's not where a struct or 
> union element has floating-point type but is never interpreted as a value 
> of that type, just the whole aggregate copied.  It's completely valid 
> without flag_signaling_nans to copy a struct that happens to have such a 
> bit pattern.

Indeed. I have removed this condition, and the patch I'm testing is:

--cut here--
diff --git a/gcc/config/i386/i386.c b/gcc/config/i386/i386.c
index b807a9a..a719d62 100644
--- a/gcc/config/i386/i386.c
+++ b/gcc/config/i386/i386.c
@@ -10311,9 +10311,25 @@ static bool
 ix86_member_type_forces_blk (const_tree field, machine_mode mode)
 {
   /* Union with XFmode must be in BLKmode.  */
-  return (mode == XFmode
-         && (TREE_CODE (DECL_FIELD_CONTEXT (field)) == UNION_TYPE
-             || TREE_CODE (DECL_FIELD_CONTEXT (field)) == QUAL_UNION_TYPE));
+  if (mode == XFmode
+      && (TREE_CODE (DECL_FIELD_CONTEXT (field)) == UNION_TYPE
+         || TREE_CODE (DECL_FIELD_CONTEXT (field)) == QUAL_UNION_TYPE))
+    return true;
+
+  /* Early exit for XFmode moves, they don't trap
+     on sNaNs and don't convert sNaNs into qNaNs.  */
+  if (mode == XFmode)
+    return false;
+
+  /* DFmode and SFmode moves through X87 stack trap on sNaNs if traps
+     are turned on or convert sNaNs to qNaNs if traps are turned off.
+     This violates C11, 6.2.6.1p6, that specifically says that
+     "[t]he value of a structure or union object is never a trap
+     representation, even though the value of a member of the structure
+     or union object may be a trap representation."  */
+
+  return (IS_STACK_MODE (mode)
+         && RECORD_OR_UNION_TYPE_P (DECL_FIELD_CONTEXT (field)));
 }

 rtx
--cut here--

> Does the patch (without flag_signaling_nans) help with the bug 58416 case 
> or not?

Using the patched compiler, I get:

$ ~/gcc-build/gcc/xgcc -B ~/gcc-build/gcc/ -O2 -m32 -mfpmath=387 pr58416.C 
$ ./a.out
magic=fff0000000000001
copy= fff0000000000001

... which looks OK to me.