To...
authorRichard Sandiford <richard.sandiford@arm.com>
Wed, 18 May 2016 14:01:31 +0000 (14:01 +0000)
committerRichard Sandiford <rsandifo@gcc.gnu.org>
Wed, 18 May 2016 14:01:31 +0000 (14:01 +0000)
commitcf8be00de9a9e31a2f05894d429fe38948a0c255
treea8b6d0e0b1940214a97b2c5e39a0e3fd1e099cf9
parent29799e9db1ee2bcb8113d96adc965c7a1a284aaa
To...

To: gcc-patches@gcc.gnu.org
Subject: PR 71020: Handle abnormal PHIs in tree-call-cdce.c
From: Richard Sandiford <richard.sandiford@arm.com>
Gcc: private.sent
--text follows this line--
The PR is about a case where tree-call-cdce.c causes two abnormal
PHIs for the same variable to be live at the same time, leading to
a coalescing failure.  It seemed like getting rid of these kinds of
input would be generally useful, so I added a utility to tree-dfa.c.

Tested on x86_64-linux-gnu.

gcc/
PR middle-end/71020
* tree-dfa.h (replace_abnormal_ssa_names): Declare.
* tree-dfa.c (replace_abnormal_ssa_names): New function.
* tree-call-cdce.c: Include tree-dfa.h.
(can_guard_call_p): New function, extracted from...
(can_use_internal_fn): ...here.
(shrink_wrap_one_built_in_call_with_conds): Remove failure path
and return void.
(shrink_wrap_one_built_in_call): Likewise.
(use_internal_fn): Likewise.
(shrink_wrap_conditional_dead_built_in_calls): Update accordingly
and return void.  Call replace_abnormal_ssa_names.
(pass_call_cdce::execute): Check can_guard_call_p during the
initial walk.  Assume shrink_wrap_conditional_dead_built_in_calls
will always change something.

gcc/testsuite/
* gcc.dg/torture/pr71020.c: New test.

From-SVN: r236393
gcc/ChangeLog
gcc/testsuite/ChangeLog
gcc/testsuite/gcc.dg/torture/pr71020.c [new file with mode: 0644]
gcc/tree-call-cdce.c
gcc/tree-dfa.c
gcc/tree-dfa.h