phiopt: Fix up conditional_replacement [PR99305]
authorJakub Jelinek <jakub@redhat.com>
Tue, 9 Mar 2021 18:13:11 +0000 (19:13 +0100)
committerJakub Jelinek <jakub@redhat.com>
Tue, 9 Mar 2021 18:13:11 +0000 (19:13 +0100)
commitb610c30453d8e4cc88693d85a5a100d089640be5
treea7927f20db4d10668770625a9b0b59c447d676df
parentc60ad1c5fe0249f48362be0f989184ca447f9d17
phiopt: Fix up conditional_replacement [PR99305]

Before my PR97690 changes, conditional_replacement would not set neg
when the nonzero arg was boolean true.
I've simplified the testing, so that it first finds the zero argument
and then checks the other argument for all the handled cases
(1, -1 and 1 << X, where the last case is what the patch added support for).
But, unfortunately I've placed the integer_all_onesp test first.
For unsigned precision 1 types such as bool integer_all_onesp, integer_onep
and integer_pow2p can all be true and the code set neg to true in that case,
which is undesirable.

The following patch tests integer_pow2p first (which is trivially true
for integer_onep too and tree_log2 in that case gives shift == 0)
and only if that isn't the case, integer_all_onesp.

2021-03-09  Jakub Jelinek  <jakub@redhat.com>

PR tree-optimization/99305
* tree-ssa-phiopt.c (conditional_replacement): Test integer_pow2p
before integer_all_onesp instead of vice versa.

* g++.dg/opt/pr99305.C: New test.
gcc/testsuite/g++.dg/opt/pr99305.C [new file with mode: 0644]
gcc/tree-ssa-phiopt.c