RE: PATCH: testsuite/37074: gcc.dg/torture/stackalign/builtin-apply-4.c failed with SSE2

["Ye, Joey" <joey dot ye at intel dot com>]

Joey wrote on Wednesday, August 13, 2008 11:17 AM:
> Following patch fixes builtin-apply-4.c. But the changes in 
> i386.md is ugly and inefficient, which insert EMMS after 
> every mmx movd insn. Any suggestion?
Unless someone think it worthwhile to continue exploring and post a
patch that is applicable (my previous patch wasn't even close to there),
I somewhat believe that avoid mixing MMX and x87 in builtin_apply and
mention it in user manual is a good idea. 

How about this patch to doc?

Thanks - Joey

Index: gcc/doc/extend.texi
===================================================================
--- gcc/doc/extend.texi (revision 138935)
+++ gcc/doc/extend.texi (working copy)
@@ -527,6 +527,24 @@ sophisticated features or other extensio
 is, therefore, not recommended to use them outside very simple
 functions acting as mere forwarders for their arguments.
 
+In x86 (not x86_64) platform, these built-in function won't work
+properly if both MMX and x87 floating point are supported (either
+in application code or in library). Following rules can be applied
+to avoid malfunction while using these functions:
+@itemize @bullet
+@item When floating point data types are used, don't enable MMX
+via -mmmx, -msse or -msseX options.
+@end itemize
+@itemize @bullet
+@item When floating point data types are used and it is required to
+enable SSE, use -mno-mmx to disable MMX.
+@end itemize
+@itemize @bullet
+@item When MMX instruction or data type are used, avoid using floating
+point data type and calling libraries that may involve floating
+point.
+@end itemize
+
 @deftypefn {Built-in Function} {void *} __builtin_apply_args ()
 This built-in function returns a pointer to data
 describing how to perform a call with the same arguments as were passed