* [PATCH] Limit some match.pd conversion folding to GIMPLE (PR c++/70336)
@ 2016-04-05 16:47 Jakub Jelinek
2016-04-05 20:27 ` Richard Biener
0 siblings, 1 reply; 2+ messages in thread
From: Jakub Jelinek @ 2016-04-05 16:47 UTC (permalink / raw)
To: Richard Biener; +Cc: gcc-patches
Hi!
As the testcase shows, the folding which previously has been done
only in tree-ssa-forwprop.c breaks some cases with explicit casts
in -Wconversion, so this patch limits that folding to GIMPLE only.
Unfortunately that breaks a few cases in builtin-isinf_sign-1.c
test at -O0, but IMNSHO expecting all the folding at -O0 is wrong,
outside of constexpr and initializers we shouldn't guarantee any
folding.
Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk?
2016-04-05 Jakub Jelinek <jakub@redhat.com>
PR c++/70336
* match.pd (nested int casts): Limit to GIMPLE.
* c-c++-common/pr70336.c: New test.
* gcc.dg/torture/builtin-isinf_sign-1.c (foo): Guard tests
no longer optimized away at -O0 with #ifndef __OPTIMIZE__.
--- gcc/match.pd.jj 2016-03-23 19:25:56.000000000 +0100
+++ gcc/match.pd 2016-04-04 12:39:32.535498926 +0200
@@ -1554,7 +1554,8 @@ DEFINE_INT_AND_FLOAT_ROUND_FN (RINT)
/* A truncation to an unsigned type (a zero-extension) should be
canonicalized as bitwise and of a mask. */
- (if (final_int && inter_int && inside_int
+ (if (GIMPLE /* PR70366: doing this in GENERIC breaks -Wconversion. */
+ && final_int && inter_int && inside_int
&& final_prec == inside_prec
&& final_prec > inter_prec
&& inter_unsignedp)
--- gcc/testsuite/c-c++-common/pr70336.c.jj 2016-04-04 12:42:48.056806268 +0200
+++ gcc/testsuite/c-c++-common/pr70336.c 2016-04-04 12:44:19.343549096 +0200
@@ -0,0 +1,37 @@
+/* PR c++/70336 */
+/* { dg-do compile } */
+/* { dg-options "-Wconversion" } */
+
+void
+f1 (unsigned char * x, int y, int z)
+{
+ x[z / 8] |= (unsigned char) (0x80 >> y); /* { dg-bogus "may alter its value" } */
+}
+
+unsigned char
+f2 (unsigned char x, int y)
+{
+ x = x | (unsigned char) (0x80 >> y); /* { dg-bogus "may alter its value" } */
+ return x;
+}
+
+unsigned char
+f3 (unsigned char x, int y)
+{
+ x = x | (unsigned char) (y & 255); /* { dg-bogus "may alter its value" } */
+ return x;
+}
+
+unsigned char
+f4 (unsigned char x, unsigned char y)
+{
+ x = x | (unsigned char) (y & 255); /* { dg-bogus "may alter its value" } */
+ return x;
+}
+
+unsigned char
+f5 (unsigned char x, int y)
+{
+ x = (unsigned char) (y & 255); /* { dg-bogus "may alter its value" } */
+ return x;
+}
--- gcc/testsuite/gcc.dg/torture/builtin-isinf_sign-1.c.jj 2008-09-05 12:54:28.000000000 +0200
+++ gcc/testsuite/gcc.dg/torture/builtin-isinf_sign-1.c 2016-04-05 16:13:40.649320108 +0200
@@ -24,6 +24,7 @@ foo (float f, double d, long double ld)
!= (__builtin_isinf(ld) ? (__builtin_signbitl(ld) ? -1 : 1) : 0))
link_error (__LINE__);
+#ifdef __OPTIMIZE__
/* In boolean contexts, GCC will fold the inner conditional
expression to 1. So isinf_sign folds to plain isinf. */
@@ -33,6 +34,7 @@ foo (float f, double d, long double ld)
link_error (__LINE__);
if ((_Bool)__builtin_isinf_sign(ld) != (__builtin_isinf(ld) != 0))
link_error (__LINE__);
+#endif
if ((__builtin_isinf_sign(f) != 0) != (__builtin_isinf(f) != 0))
link_error (__LINE__);
Jakub
^ permalink raw reply [flat|nested] 2+ messages in thread
* Re: [PATCH] Limit some match.pd conversion folding to GIMPLE (PR c++/70336)
2016-04-05 16:47 [PATCH] Limit some match.pd conversion folding to GIMPLE (PR c++/70336) Jakub Jelinek
@ 2016-04-05 20:27 ` Richard Biener
0 siblings, 0 replies; 2+ messages in thread
From: Richard Biener @ 2016-04-05 20:27 UTC (permalink / raw)
To: Jakub Jelinek; +Cc: gcc-patches
On April 5, 2016 6:47:20 PM GMT+02:00, Jakub Jelinek <jakub@redhat.com> wrote:
>Hi!
>
>As the testcase shows, the folding which previously has been done
>only in tree-ssa-forwprop.c breaks some cases with explicit casts
>in -Wconversion, so this patch limits that folding to GIMPLE only.
>Unfortunately that breaks a few cases in builtin-isinf_sign-1.c
>test at -O0, but IMNSHO expecting all the folding at -O0 is wrong,
>outside of constexpr and initializers we shouldn't guarantee any
>folding.
>
>Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk?
Ok.
Richard.
>2016-04-05 Jakub Jelinek <jakub@redhat.com>
>
> PR c++/70336
> * match.pd (nested int casts): Limit to GIMPLE.
>
> * c-c++-common/pr70336.c: New test.
> * gcc.dg/torture/builtin-isinf_sign-1.c (foo): Guard tests
> no longer optimized away at -O0 with #ifndef __OPTIMIZE__.
>
>--- gcc/match.pd.jj 2016-03-23 19:25:56.000000000 +0100
>+++ gcc/match.pd 2016-04-04 12:39:32.535498926 +0200
>@@ -1554,7 +1554,8 @@ DEFINE_INT_AND_FLOAT_ROUND_FN (RINT)
>
> /* A truncation to an unsigned type (a zero-extension) should be
> canonicalized as bitwise and of a mask. */
>- (if (final_int && inter_int && inside_int
>+ (if (GIMPLE /* PR70366: doing this in GENERIC breaks -Wconversion.
> */
>+ && final_int && inter_int && inside_int
> && final_prec == inside_prec
> && final_prec > inter_prec
> && inter_unsignedp)
>--- gcc/testsuite/c-c++-common/pr70336.c.jj 2016-04-04
>12:42:48.056806268 +0200
>+++ gcc/testsuite/c-c++-common/pr70336.c 2016-04-04 12:44:19.343549096
>+0200
>@@ -0,0 +1,37 @@
>+/* PR c++/70336 */
>+/* { dg-do compile } */
>+/* { dg-options "-Wconversion" } */
>+
>+void
>+f1 (unsigned char * x, int y, int z)
>+{
>+ x[z / 8] |= (unsigned char) (0x80 >> y); /* { dg-bogus "may alter
>its value" } */
>+}
>+
>+unsigned char
>+f2 (unsigned char x, int y)
>+{
>+ x = x | (unsigned char) (0x80 >> y); /* { dg-bogus "may alter its
>value" } */
>+ return x;
>+}
>+
>+unsigned char
>+f3 (unsigned char x, int y)
>+{
>+ x = x | (unsigned char) (y & 255); /* { dg-bogus "may alter its
>value" } */
>+ return x;
>+}
>+
>+unsigned char
>+f4 (unsigned char x, unsigned char y)
>+{
>+ x = x | (unsigned char) (y & 255); /* { dg-bogus "may alter its
>value" } */
>+ return x;
>+}
>+
>+unsigned char
>+f5 (unsigned char x, int y)
>+{
>+ x = (unsigned char) (y & 255); /* { dg-bogus "may alter its value"
>} */
>+ return x;
>+}
>--- gcc/testsuite/gcc.dg/torture/builtin-isinf_sign-1.c.jj 2008-09-05
>12:54:28.000000000 +0200
>+++ gcc/testsuite/gcc.dg/torture/builtin-isinf_sign-1.c 2016-04-05
>16:13:40.649320108 +0200
>@@ -24,6 +24,7 @@ foo (float f, double d, long double ld)
> != (__builtin_isinf(ld) ? (__builtin_signbitl(ld) ? -1 : 1) : 0))
> link_error (__LINE__);
>
>+#ifdef __OPTIMIZE__
> /* In boolean contexts, GCC will fold the inner conditional
> expression to 1. So isinf_sign folds to plain isinf. */
>
>@@ -33,6 +34,7 @@ foo (float f, double d, long double ld)
> link_error (__LINE__);
> if ((_Bool)__builtin_isinf_sign(ld) != (__builtin_isinf(ld) != 0))
> link_error (__LINE__);
>+#endif
>
> if ((__builtin_isinf_sign(f) != 0) != (__builtin_isinf(f) != 0))
> link_error (__LINE__);
>
> Jakub
^ permalink raw reply [flat|nested] 2+ messages in thread
end of thread, other threads:[~2016-04-05 20:27 UTC | newest]
Thread overview: 2+ messages (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2016-04-05 16:47 [PATCH] Limit some match.pd conversion folding to GIMPLE (PR c++/70336) Jakub Jelinek
2016-04-05 20:27 ` Richard Biener
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).