After r15-3436-gb2b20b277988ab, poor_ifcvt_pred returns false for
the case where the statement could trap but in this case trapping
instructions cannot be made unconditional so it is a poor ifcvt.
This fixes a small preformance regression with TSVC s258 at
`-O3 -ftrapping-math` on aarch64 where ifcvt would not happen
and we would still have a branch.
On a specific aarch64, we go from 0.145s down to 0.118s.
Bootstrapped and tested on x86_64-linux-gnu.
gcc/ChangeLog:
PR tree-optimization/118505
* gimple-ssa-split-paths.cc (poor_ifcvt_pred): Return
true for trapping statements.
Signed-off-by: Andrew Pinski <quic_apin...@quicinc.com>
---
gcc/gimple-ssa-split-paths.cc | 5 +++++
1 file changed, 5 insertions(+)
diff --git a/gcc/gimple-ssa-split-paths.cc b/gcc/gimple-ssa-split-paths.cc
index 7c5bc1d842c..9db73fdcc6d 100644
--- a/gcc/gimple-ssa-split-paths.cc
+++ b/gcc/gimple-ssa-split-paths.cc
@@ -160,6 +160,11 @@ poor_ifcvt_pred (basic_block pred, basic_block bb)
gimple *stmt = last_and_only_stmt (pred);
if (!stmt || gimple_code (stmt) != GIMPLE_ASSIGN)
return true;
+
+ /* If the statement could trap, then this is a poor ifcvt candidate. */
+ if (gimple_could_trap_p (stmt))
+ return true;
+
tree_code code = gimple_assign_rhs_code (stmt);
if (poor_ifcvt_candidate_code (code))
return true;
--
2.43.0
