* [PATCH]: Backport fix for PR41963 (miscompilation of 177.mesa)
@ 2009-11-06 23:49 Uros Bizjak
2009-11-07 13:47 ` Richard Guenther
0 siblings, 1 reply; 2+ messages in thread
From: Uros Bizjak @ 2009-11-06 23:49 UTC (permalink / raw)
To: GCC Patches
[-- Attachment #1: Type: text/plain, Size: 735 bytes --]
Hello!
2009-11-07 Uros Bizjak <ubizjak@gmail.com>
Backport from mainline:
2009-11-06 Michael Matz <matz@suse.de>
PR middle-end/41963
* tree-ssa-math-opts.c (execute_cse_reciprocals): Check all uses
of a potential reciprocal to really be reciprocals.
testsuite/ChangeLog:
2009-11-07 Uros Bizjak <ubizjak@gmail.com>
Backport from mainline:
2009-11-06 Michael Matz <matz@suse.de>
PR middle-end/41963
* gcc.target/i386/pr41963.c: New test.
Since this transformation is still activated by -mrecip only in the
release branches, I have moved the test to i386 target directory, where
this option can be activated.
Patch was tested on x86_64-pc-linux-gnu. OK for 4.4 and 4.3 ?
Uros.
[-- Attachment #2: p.diff.txt --]
[-- Type: text/plain, Size: 2437 bytes --]
Index: tree-ssa-math-opts.c
===================================================================
--- tree-ssa-math-opts.c (revision 153979)
+++ tree-ssa-math-opts.c (working copy)
@@ -528,7 +528,9 @@ execute_cse_reciprocals (void)
|| DECL_BUILT_IN_CLASS (fndecl) == BUILT_IN_MD))
{
enum built_in_function code;
- bool md_code;
+ bool md_code, fail;
+ imm_use_iterator ui;
+ use_operand_p use_p;
code = DECL_FUNCTION_CODE (fndecl);
md_code = DECL_BUILT_IN_CLASS (fndecl) == BUILT_IN_MD;
@@ -537,12 +539,34 @@ execute_cse_reciprocals (void)
if (!fndecl)
continue;
+ /* Check that all uses of the SSA name are divisions,
+ otherwise replacing the defining statement will do
+ the wrong thing. */
+ fail = false;
+ FOR_EACH_IMM_USE_FAST (use_p, ui, arg1)
+ {
+ gimple stmt2 = USE_STMT (use_p);
+ if (!is_gimple_assign (stmt2)
+ || gimple_assign_rhs_code (stmt2) != RDIV_EXPR
+ || gimple_assign_rhs1 (stmt2) == arg1
+ || gimple_assign_rhs2 (stmt2) != arg1)
+ {
+ fail = true;
+ break;
+ }
+ }
+ if (fail)
+ continue;
+
gimple_call_set_fndecl (stmt1, fndecl);
update_stmt (stmt1);
- gimple_assign_set_rhs_code (stmt, MULT_EXPR);
- fold_stmt_inplace (stmt);
- update_stmt (stmt);
+ FOR_EACH_IMM_USE_STMT (stmt, ui, arg1)
+ {
+ gimple_assign_set_rhs_code (stmt, MULT_EXPR);
+ fold_stmt_inplace (stmt);
+ update_stmt (stmt);
+ }
}
}
}
Index: testsuite/gcc.target/i386/pr41963.c
===================================================================
--- testsuite/gcc.target/i386/pr41963.c (revision 0)
+++ testsuite/gcc.target/i386/pr41963.c (revision 0)
@@ -0,0 +1,36 @@
+/* { dg-do run } */
+/* { dg-options "-O2 -ffast-math -mrecip" } */
+#include <math.h>
+
+extern float sqrtf(float);
+
+static __attribute__((noinline)) void f (float *dst, float *src)
+{
+ int i, j;
+ for (i = 0; i < 2; i++)
+ {
+ float len;
+ dst[0] = src[0];
+ dst[1] = src[1];
+ len = sqrtf (dst[0] * dst[0] + dst[1] * dst[1]);
+ if (len > 0.5f)
+ {
+ len = 1.0f / len;
+ dst[0] *= len;
+ dst[1] *= len;
+ }
+ }
+}
+
+extern void abort (void);
+
+int main()
+{
+ float dst[2], src[2];
+ src[0] = 2.0f;
+ src[1] = 5.0f;
+ f (dst, src);
+ if (fabsf (dst[0] * dst[0] + dst[1] * dst[1] - 1.0f) > 0.01f)
+ abort ();
+ return 0;
+}
^ permalink raw reply [flat|nested] 2+ messages in thread
* Re: [PATCH]: Backport fix for PR41963 (miscompilation of 177.mesa)
2009-11-06 23:49 [PATCH]: Backport fix for PR41963 (miscompilation of 177.mesa) Uros Bizjak
@ 2009-11-07 13:47 ` Richard Guenther
0 siblings, 0 replies; 2+ messages in thread
From: Richard Guenther @ 2009-11-07 13:47 UTC (permalink / raw)
To: Uros Bizjak; +Cc: GCC Patches
2009/11/7 Uros Bizjak <ubizjak@gmail.com>:
> Hello!
>
> 2009-11-07 Uros Bizjak <ubizjak@gmail.com>
>
> Backport from mainline:
> 2009-11-06 Michael Matz <matz@suse.de>
>
> PR middle-end/41963
> * tree-ssa-math-opts.c (execute_cse_reciprocals): Check all uses
> of a potential reciprocal to really be reciprocals.
>
> testsuite/ChangeLog:
>
> 2009-11-07 Uros Bizjak <ubizjak@gmail.com>
>
> Backport from mainline:
> 2009-11-06 Michael Matz <matz@suse.de>
>
> PR middle-end/41963
> * gcc.target/i386/pr41963.c: New test.
>
> Since this transformation is still activated by -mrecip only in the release
> branches, I have moved the test to i386 target directory, where this option
> can be activated.
>
> Patch was tested on x86_64-pc-linux-gnu. OK for 4.4 and 4.3 ?
Ok for 4.4. 4.3 needs a different patch as it doesn't have tuples.
Richard.
> Uros.
>
^ permalink raw reply [flat|nested] 2+ messages in thread
end of thread, other threads:[~2009-11-07 13:46 UTC | newest]
Thread overview: 2+ messages (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2009-11-06 23:49 [PATCH]: Backport fix for PR41963 (miscompilation of 177.mesa) Uros Bizjak
2009-11-07 13:47 ` Richard Guenther
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for read-only IMAP folder(s) and NNTP newsgroup(s).