The following adds a pattern to elide a .REDUC_IOR operation when
the result is compared against zero with a cbranch. I've resorted
to using can_compare_p since that's what RTL expansion eventually
checks - while GIMPLE allowed whole vector equality compares for long
I'll notice vector lowering won't lower unsupported ones and RTL
expansion doesn't seem to try using [u]cmp<vector-mode> optabs
(and neither x86 nor aarch64 implements those). There's cstore
but no target implements that for vector modes either.
Bootstrapped on x86_64-unknown-linux-gnu, testing in progress.
If anybody has suggestions on how to better guard the pattern I'm
all ears.
Thanks,
Richard.
PR tree-optimization/117000
* match.pd (.REDUC_IOR !=/== 0): New pattern.
* gimple-match-head.cc: Include memmodel.h and optabs.h.
* generic-match-head.cc: Likewise.
* gcc.target/i386/pr117000.c: New testcase.
---
gcc/generic-match-head.cc | 2 ++
gcc/gimple-match-head.cc | 2 ++
gcc/match.pd | 9 +++++++++
gcc/testsuite/gcc.target/i386/pr117000.c | 13 +++++++++++++
4 files changed, 26 insertions(+)
create mode 100644 gcc/testsuite/gcc.target/i386/pr117000.c
diff --git a/gcc/generic-match-head.cc b/gcc/generic-match-head.cc
index 42dee626613..7d7e2a9f792 100644
--- a/gcc/generic-match-head.cc
+++ b/gcc/generic-match-head.cc
@@ -35,6 +35,8 @@ along with GCC; see the file COPYING3. If not see
#include "builtins.h"
#include "case-cfn-macros.h"
#include "gimplify.h"
+#include "memmodel.h"
+#include "optabs.h"
#include "optabs-tree.h"
#include "dbgcnt.h"
#include "tm.h"
diff --git a/gcc/gimple-match-head.cc b/gcc/gimple-match-head.cc
index 4147a0eb38a..b9d5f751b7c 100644
--- a/gcc/gimple-match-head.cc
+++ b/gcc/gimple-match-head.cc
@@ -41,6 +41,8 @@ along with GCC; see the file COPYING3. If not see
#include "internal-fn.h"
#include "case-cfn-macros.h"
#include "gimplify.h"
+#include "memmodel.h"
+#include "optabs.h"
#include "optabs-tree.h"
#include "tree-eh.h"
#include "dbgcnt.h"
diff --git a/gcc/match.pd b/gcc/match.pd
index ba83f0f29e6..c2efebd98bb 100644
--- a/gcc/match.pd
+++ b/gcc/match.pd
@@ -10332,6 +10332,15 @@ and,
(simplify (reduc (op @0 VECTOR_CST@1))
(op (reduc:type @0) (reduc:type @1))))
+/* Simplify .REDUC_IOR (@0) ==/!= 0 to @0 ==/!= 0. */
+(for cmp (eq ne)
+ (simplify
+ (cmp (IFN_REDUC_IOR @0) integer_zerop)
+ (if (VECTOR_MODE_P (TYPE_MODE (TREE_TYPE (@0)))
+ && can_compare_p (cmp == EQ_EXPR ? EQ : NE, TYPE_MODE (TREE_TYPE (@0)),
+ ccp_jump))
+ (cmp @0 { build_zero_cst (TREE_TYPE (@0)); }))))
+
/* Simplify vector floating point operations of alternating sub/add pairs
into using an fneg of a wider element type followed by a normal add.
under IEEE 754 the fneg of the wider type will negate every even entry
diff --git a/gcc/testsuite/gcc.target/i386/pr117000.c
b/gcc/testsuite/gcc.target/i386/pr117000.c
new file mode 100644
index 00000000000..04f94344eb1
--- /dev/null
+++ b/gcc/testsuite/gcc.target/i386/pr117000.c
@@ -0,0 +1,13 @@
+/* { dg-do compile } */
+/* { dg-options "-O2 -msse4.1" { target sse4 } } */
+
+int eq(unsigned long *x, unsigned long *y)
+{
+ unsigned long folded = 0;
+ for (int i = 0; i < 4; ++i)
+ folded |= x[i] ^ y[i];
+ return folded == 0;
+}
+
+/* We want to elide the .REDUC_IOR with the compare against zero. */
+/* { dg-final { scan-assembler "ptest" } } */
--
2.43.0
