We can fold svbrka/b to pfalse, if the governing predicate and the operand
are both ptrue.
We implemented this optimization during gimple folding in
svbrk_unary_impl::fold.
The patch was bootstrapped and tested on aarch64-linux-gnu, no regression.
OK for trunk?
Signed-off-by: Jennifer Schmitz <jschm...@nvidia.com>
gcc/
PR target/121604
* config/aarch64/aarch64-sve-builtins-base.cc
(svbrk_unary_impl::fold): Fold calls with ptrue predicate and
operand to pfalse.
gcc/testsuite/
PR target/121604
* gcc.target/aarch64/sve/pr121604_brk.c: Adjust expected outcome.
---
gcc/config/aarch64/aarch64-sve-builtins-base.cc | 12 ++++++++++++
gcc/testsuite/gcc.target/aarch64/sve/pr121604_brk.c | 6 ++----
2 files changed, 14 insertions(+), 4 deletions(-)
diff --git a/gcc/config/aarch64/aarch64-sve-builtins-base.cc
b/gcc/config/aarch64/aarch64-sve-builtins-base.cc
index ecc06877cac..a2877533c42 100644
--- a/gcc/config/aarch64/aarch64-sve-builtins-base.cc
+++ b/gcc/config/aarch64/aarch64-sve-builtins-base.cc
@@ -346,6 +346,18 @@ class svbrk_unary_impl : public function_base
public:
CONSTEXPR svbrk_unary_impl (int unspec) : m_unspec (unspec) {}
+ gimple *
+ fold (gimple_folder &f) const override
+ {
+ /* If the predicate and operand are ptrue, fold to pfalse. */
+ if (is_ptrue (gimple_call_arg (f.call, 1),
+ f.type_suffix (0).element_bytes)
+ && is_ptrue (gimple_call_arg (f.call, 2),
+ f.type_suffix (0).element_bytes))
+ return f.fold_call_to (build_zero_cst (TREE_TYPE (f.lhs)));
+ return NULL;
+ }
+
rtx
expand (function_expander &e) const override
{
diff --git a/gcc/testsuite/gcc.target/aarch64/sve/pr121604_brk.c
b/gcc/testsuite/gcc.target/aarch64/sve/pr121604_brk.c
index a474a20554d..c1965b034cd 100644
--- a/gcc/testsuite/gcc.target/aarch64/sve/pr121604_brk.c
+++ b/gcc/testsuite/gcc.target/aarch64/sve/pr121604_brk.c
@@ -6,8 +6,7 @@
/*
** foo:
-** ptrue p0\.b, all
-** brkb p0\.b, p0/z, p0\.b
+** pfalse p0\.b
** ret
*/
svbool_t foo () {
@@ -16,8 +15,7 @@ svbool_t foo () {
/*
** bar:
-** ptrue p0\.b, all
-** brka p0\.b, p0/z, p0\.b
+** pfalse p0\.b
** ret
*/
svbool_t bar () {
--
2.34.1
