* [PATCH] Fix reassoc bit test optimization (PR tree-optimization/65418)
@ 2015-03-14 9:10 Jakub Jelinek
2015-03-14 9:48 ` Richard Biener
0 siblings, 1 reply; 2+ messages in thread
From: Jakub Jelinek @ 2015-03-14 9:10 UTC (permalink / raw)
To: Richard Biener; +Cc: gcc-patches
Hi!
The first testcase shows a bug in my bit test reassoc optimization,
extract_bit_test_mask is (intentionally) stripping nops, but was setting
*totallowp and operating with tbias in the type of unstripped expression,
which then results in different signedness of types used and confusing the
optimization. In particular, -218 and -216 are already folded into (x is
signed int)
(((unsigned) x + 218U) & -2U) == 0
and thus without the patch we set lowi in the parent to -218U.
Then -146 and -132 are just
x == -146
and
x == -132
thus we were comparing -218U to -146 or -132. But we really want
to use -218 instead, as that is the type of x.
Fixed thusly, bootstrapped/regtested on
{x86_64,i686,aarch64,powerpc64{,le},s390{,x}}-linux, ok for trunk?
2015-03-14 Jakub Jelinek <jakub@redhat.com>
PR tree-optimization/65418
* tree-ssa-reassoc.c (extract_bit_test_mask): If there
are casts in the first PLUS_EXPR operand, ensure tbias and
*totallowp are in the inner type.
* gcc.c-torture/execute/pr65418-1.c: New test.
* gcc.c-torture/execute/pr65418-2.c: New test.
--- gcc/tree-ssa-reassoc.c.jj 2015-02-26 22:02:39.000000000 +0100
+++ gcc/tree-ssa-reassoc.c 2015-03-13 16:22:50.506295252 +0100
@@ -2439,26 +2439,25 @@ extract_bit_test_mask (tree exp, int pre
&& TREE_CODE (exp) == PLUS_EXPR
&& TREE_CODE (TREE_OPERAND (exp, 1)) == INTEGER_CST)
{
+ tree ret = TREE_OPERAND (exp, 0);
+ STRIP_NOPS (ret);
widest_int bias
= wi::neg (wi::sext (wi::to_widest (TREE_OPERAND (exp, 1)),
TYPE_PRECISION (TREE_TYPE (low))));
- tree tbias = wide_int_to_tree (TREE_TYPE (low), bias);
+ tree tbias = wide_int_to_tree (TREE_TYPE (ret), bias);
if (totallowp)
{
*totallowp = tbias;
- exp = TREE_OPERAND (exp, 0);
- STRIP_NOPS (exp);
- return exp;
+ return ret;
}
else if (!tree_int_cst_lt (totallow, tbias))
return NULL_TREE;
+ bias = wi::to_widest (tbias);
bias -= wi::to_widest (totallow);
if (wi::ges_p (bias, 0) && wi::lts_p (bias, prec - max))
{
*mask = wi::lshift (*mask, bias);
- exp = TREE_OPERAND (exp, 0);
- STRIP_NOPS (exp);
- return exp;
+ return ret;
}
}
}
--- gcc/testsuite/gcc.c-torture/execute/pr65418-1.c.jj 2015-03-13 16:49:07.973604649 +0100
+++ gcc/testsuite/gcc.c-torture/execute/pr65418-1.c 2015-03-13 16:48:28.000000000 +0100
@@ -0,0 +1,19 @@
+/* PR tree-optimization/65418 */
+
+__attribute__((noinline, noclone)) int
+foo (int x)
+{
+ if (x == -216 || x == -132 || x == -218 || x == -146)
+ return 1;
+ return 0;
+}
+
+int
+main ()
+{
+ volatile int i;
+ for (i = -230; i < -120; i++)
+ if (foo (i) != (i == -216 || i == -132 || i == -218 || i == -146))
+ __builtin_abort ();
+ return 0;
+}
--- gcc/testsuite/gcc.c-torture/execute/pr65418-2.c.jj 2015-03-13 16:49:10.992556110 +0100
+++ gcc/testsuite/gcc.c-torture/execute/pr65418-2.c 2015-03-13 16:48:44.000000000 +0100
@@ -0,0 +1,19 @@
+/* PR tree-optimization/65418 */
+
+__attribute__((noinline, noclone)) int
+foo (int x)
+{
+ if (x == -216 || x == -211 || x == -218 || x == -205 || x == -223)
+ return 1;
+ return 0;
+}
+
+int
+main ()
+{
+ volatile int i;
+ for (i = -230; i < -200; i++)
+ if (foo (i) != (i == -216 || i == -211 || i == -218 || i == -205 || i == -223))
+ __builtin_abort ();
+ return 0;
+}
Jakub
^ permalink raw reply [flat|nested] 2+ messages in thread
* Re: [PATCH] Fix reassoc bit test optimization (PR tree-optimization/65418)
2015-03-14 9:10 [PATCH] Fix reassoc bit test optimization (PR tree-optimization/65418) Jakub Jelinek
@ 2015-03-14 9:48 ` Richard Biener
0 siblings, 0 replies; 2+ messages in thread
From: Richard Biener @ 2015-03-14 9:48 UTC (permalink / raw)
To: Jakub Jelinek; +Cc: gcc-patches
On March 14, 2015 10:10:34 AM GMT+01:00, Jakub Jelinek <jakub@redhat.com> wrote:
>Hi!
>
>The first testcase shows a bug in my bit test reassoc optimization,
>extract_bit_test_mask is (intentionally) stripping nops, but was
>setting
>*totallowp and operating with tbias in the type of unstripped
>expression,
>which then results in different signedness of types used and confusing
>the
>optimization. In particular, -218 and -216 are already folded into (x
>is
>signed int)
>(((unsigned) x + 218U) & -2U) == 0
>and thus without the patch we set lowi in the parent to -218U.
>Then -146 and -132 are just
>x == -146
>and
>x == -132
>thus we were comparing -218U to -146 or -132. But we really want
>to use -218 instead, as that is the type of x.
>
>Fixed thusly, bootstrapped/regtested on
>{x86_64,i686,aarch64,powerpc64{,le},s390{,x}}-linux, ok for trunk?
OK.
Thanks,
Richard.
>2015-03-14 Jakub Jelinek <jakub@redhat.com>
>
> PR tree-optimization/65418
> * tree-ssa-reassoc.c (extract_bit_test_mask): If there
> are casts in the first PLUS_EXPR operand, ensure tbias and
> *totallowp are in the inner type.
>
> * gcc.c-torture/execute/pr65418-1.c: New test.
> * gcc.c-torture/execute/pr65418-2.c: New test.
>
>--- gcc/tree-ssa-reassoc.c.jj 2015-02-26 22:02:39.000000000 +0100
>+++ gcc/tree-ssa-reassoc.c 2015-03-13 16:22:50.506295252 +0100
>@@ -2439,26 +2439,25 @@ extract_bit_test_mask (tree exp, int pre
> && TREE_CODE (exp) == PLUS_EXPR
> && TREE_CODE (TREE_OPERAND (exp, 1)) == INTEGER_CST)
> {
>+ tree ret = TREE_OPERAND (exp, 0);
>+ STRIP_NOPS (ret);
> widest_int bias
> = wi::neg (wi::sext (wi::to_widest (TREE_OPERAND (exp, 1)),
> TYPE_PRECISION (TREE_TYPE (low))));
>- tree tbias = wide_int_to_tree (TREE_TYPE (low), bias);
>+ tree tbias = wide_int_to_tree (TREE_TYPE (ret), bias);
> if (totallowp)
> {
> *totallowp = tbias;
>- exp = TREE_OPERAND (exp, 0);
>- STRIP_NOPS (exp);
>- return exp;
>+ return ret;
> }
> else if (!tree_int_cst_lt (totallow, tbias))
> return NULL_TREE;
>+ bias = wi::to_widest (tbias);
> bias -= wi::to_widest (totallow);
> if (wi::ges_p (bias, 0) && wi::lts_p (bias, prec - max))
> {
> *mask = wi::lshift (*mask, bias);
>- exp = TREE_OPERAND (exp, 0);
>- STRIP_NOPS (exp);
>- return exp;
>+ return ret;
> }
> }
> }
>--- gcc/testsuite/gcc.c-torture/execute/pr65418-1.c.jj 2015-03-13
>16:49:07.973604649 +0100
>+++ gcc/testsuite/gcc.c-torture/execute/pr65418-1.c 2015-03-13
>16:48:28.000000000 +0100
>@@ -0,0 +1,19 @@
>+/* PR tree-optimization/65418 */
>+
>+__attribute__((noinline, noclone)) int
>+foo (int x)
>+{
>+ if (x == -216 || x == -132 || x == -218 || x == -146)
>+ return 1;
>+ return 0;
>+}
>+
>+int
>+main ()
>+{
>+ volatile int i;
>+ for (i = -230; i < -120; i++)
>+ if (foo (i) != (i == -216 || i == -132 || i == -218 || i == -146))
>+ __builtin_abort ();
>+ return 0;
>+}
>--- gcc/testsuite/gcc.c-torture/execute/pr65418-2.c.jj 2015-03-13
>16:49:10.992556110 +0100
>+++ gcc/testsuite/gcc.c-torture/execute/pr65418-2.c 2015-03-13
>16:48:44.000000000 +0100
>@@ -0,0 +1,19 @@
>+/* PR tree-optimization/65418 */
>+
>+__attribute__((noinline, noclone)) int
>+foo (int x)
>+{
>+ if (x == -216 || x == -211 || x == -218 || x == -205 || x == -223)
>+ return 1;
>+ return 0;
>+}
>+
>+int
>+main ()
>+{
>+ volatile int i;
>+ for (i = -230; i < -200; i++)
>+ if (foo (i) != (i == -216 || i == -211 || i == -218 || i == -205
>|| i == -223))
>+ __builtin_abort ();
>+ return 0;
>+}
>
> Jakub
^ permalink raw reply [flat|nested] 2+ messages in thread
end of thread, other threads:[~2015-03-14 9:48 UTC | newest]
Thread overview: 2+ messages (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2015-03-14 9:10 [PATCH] Fix reassoc bit test optimization (PR tree-optimization/65418) Jakub Jelinek
2015-03-14 9:48 ` 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).