diff options
| author | Jakub Jelinek <jakub@redhat.com> | 2023-03-29 08:33:30 +0200 |
|---|---|---|
| committer | Jakub Jelinek <jakub@redhat.com> | 2023-03-29 08:37:48 +0200 |
| commit | 8dd57939c20b5e218404a838f514429f8e414dea (patch) | |
| tree | 8d0bcdff7ce8588399da07ad56a6d6d1793b0095 | |
| parent | c9954996cd647daf0ba03e34dd279b97982f671f (diff) | |
| download | gcc-8dd57939c20b5e218404a838f514429f8e414dea.zip gcc-8dd57939c20b5e218404a838f514429f8e414dea.tar.gz gcc-8dd57939c20b5e218404a838f514429f8e414dea.tar.bz2 | |
tree-ssa-math-opts: Move PROP_gimple_opt_math from sincos pass to powcabs [PR109301]
The following testcase ICEs since the sincos and vect pass order has
been swapped. It is not valid to replace vector sqrt (sqrt (x)) with
pow (x, 0.25) because build_real on vector type is invalid (could be
handled by using build_uniform_cst and adjusting type passed to
build_real) but more importantly because nothing checks if we can
actually do vector pow.
While we have pow_optab, apparently no target defines it, so it doesn't
seem to be worth bothering with for now and the patch just punts on
non-scalar sqrts.
I think the other simplifications next to it are fine, as they mostly
use CBRT which doesn't even have internal function (so is a builtin
only and therefore always scalar), or have already pow in the IL (which
doesn't have optab and shouldn't be thus vector either).
It is true that with <bits/math-vector.h> we do vectorize some calls to
pow or cbrt (but don't handle others strangely), but those aren't using
internal functions but simd clones and so match.pd doesn't know anything
about those (at least for now).
The following patch fixes it by mostly restoring the state before
r13-1763 where canonicalize_math_p () was true only until the end of the
pass which transformed pow or pow-like calls before vectorization (formerly
sincos pass, now it is powcabs pass).
powcabs is a pass in the spot sincos was happening before, so the
only change was defer the sin+cos simplification into cexpi to a later
new pass (except for the name moving with it) and none of the
canonicalize_math_p () guarded simplification in match.pd seem to rely
on those sin+cos -> cexpi simplifications and canonicalize_math_p is
the only user of this property.
2023-03-29 Jakub Jelinek <jakub@redhat.com>
Richard Biener <rguenther@suse.de>
PR tree-optimization/109301
* tree-ssa-math-opts.cc (pass_data_cse_sincos): Change
properties_provided from PROP_gimple_opt_math to 0.
(pass_data_expand_powcabs): Change properties_provided from 0 to
PROP_gimple_opt_math.
* gcc.dg/pr109301.c: New test.
| -rw-r--r-- | gcc/testsuite/gcc.dg/pr109301.c | 13 | ||||
| -rw-r--r-- | gcc/tree-ssa-math-opts.cc | 4 |
2 files changed, 15 insertions, 2 deletions
diff --git a/gcc/testsuite/gcc.dg/pr109301.c b/gcc/testsuite/gcc.dg/pr109301.c new file mode 100644 index 0000000..ab26ea6 --- /dev/null +++ b/gcc/testsuite/gcc.dg/pr109301.c @@ -0,0 +1,13 @@ +/* PR tree-optimization/109301 */ +/* { dg-do compile } */ +/* { dg-options "-O3 -ffast-math" } */ + +double x[256]; + +void +foo (void) +{ + for (int i = 0; i < 256; ++i) + for (int j = 0; j < 8; ++j) + x[i] = __builtin_pow (x[i], 0.5); +} diff --git a/gcc/tree-ssa-math-opts.cc b/gcc/tree-ssa-math-opts.cc index 26ed91d..15eed3e 100644 --- a/gcc/tree-ssa-math-opts.cc +++ b/gcc/tree-ssa-math-opts.cc @@ -2237,7 +2237,7 @@ const pass_data pass_data_cse_sincos = OPTGROUP_NONE, /* optinfo_flags */ TV_TREE_SINCOS, /* tv_id */ PROP_ssa, /* properties_required */ - PROP_gimple_opt_math, /* properties_provided */ + 0, /* properties_provided */ 0, /* properties_destroyed */ 0, /* todo_flags_start */ TODO_update_ssa, /* todo_flags_finish */ @@ -2331,7 +2331,7 @@ const pass_data pass_data_expand_powcabs = OPTGROUP_NONE, /* optinfo_flags */ TV_TREE_POWCABS, /* tv_id */ PROP_ssa, /* properties_required */ - 0, /* properties_provided */ + PROP_gimple_opt_math, /* properties_provided */ 0, /* properties_destroyed */ 0, /* todo_flags_start */ TODO_update_ssa, /* todo_flags_finish */ |