This is the mail archive of the
gcc-patches@gcc.gnu.org
mailing list for the GCC project.
Avoid memory missmatches on i386
- From: Jan Hubicka <jh at suse dot cz>
- To: gcc-patches at gcc dot gnu dot org
- Date: Sun, 4 Dec 2005 00:41:48 +0100
- Subject: Avoid memory missmatches on i386
Hi,
the code promoting 8bit/16bit reads to 32bit can cause memory missmatch
faults on modern i386 chips. This actually happens for SPEC version of
Mesa benchmark, so I think it is better to disable the optimization that
is only saving byte of prefix anyway.
I can measure 0x58% SPECfp speedup and no changes on SPECint side on K8.
Bootstrapping/regtesting in progress on i686-linux OK assuming it passes?
Honza
2005-12-04 Jan Hubicka <jh@suse.cz>
* config/i386/predicates.md (aligned_operand): Return 0 when memory
missmatches are prohibited.
Index: config/i386/predicates.md
===================================================================
*** config/i386/predicates.md (revision 107998)
--- config/i386/predicates.md (working copy)
***************
*** 698,703 ****
--- 698,708 ----
if (GET_CODE (op) != MEM)
return 1;
+ /* All patterns using algined_operand on memory operands ends up
+ in promoting memory operand to 64bit and thus causing memory missmatch. */
+ if (TARGET_MEMORY_MISMATCH_STALL && !optimize_size)
+ return 0;
+
/* Don't even try to do any aligned optimizations with volatiles. */
if (MEM_VOLATILE_P (op))
return 0;