When if-conversion was changed to use .COND_ADD/SUB for conditional
reduction it was forgotten to update reduction path handling to
canonicalize .COND_SUB to .COND_ADD for vectorizable_reduction
similar to what we do for MINUS_EXPR. The following adds this
and testcases exercising this at runtime and looking for the
appropriate masked subtraction in the vectorized code on x86.
Bootstrapped and tested on x86_64-unknown-linux-gnu, pushed.
PR tree-optimization/113078
* tree-vect-loop.cc (check_reduction_path): Canonicalize
.COND_SUB to .COND_ADD.
* gcc.dg/vect/vect-reduc-cond-sub.c: New testcase.
* gcc.target/i386/vect-pr113078.c: Likewise.
---
.../gcc.dg/vect/vect-reduc-cond-sub.c | 29 +++++++++++++++++++
gcc/testsuite/gcc.target/i386/vect-pr113078.c | 16 ++++++++++
gcc/tree-vect-loop.cc | 7 +++++
3 files changed, 52 insertions(+)
create mode 100644 gcc/testsuite/gcc.dg/vect/vect-reduc-cond-sub.c
create mode 100644 gcc/testsuite/gcc.target/i386/vect-pr113078.c
diff --git a/gcc/testsuite/gcc.dg/vect/vect-reduc-cond-sub.c
b/gcc/testsuite/gcc.dg/vect/vect-reduc-cond-sub.c
new file mode 100644
index 00000000000..0213a0ab4fd
--- /dev/null
+++ b/gcc/testsuite/gcc.dg/vect/vect-reduc-cond-sub.c
@@ -0,0 +1,29 @@
+/* { dg-require-effective-target vect_int } */
+
+#include "tree-vect.h"
+
+int __attribute__((noipa))
+foo (int n, int* p, int* pi)
+{
+ int sum = 0;
+ for (int i = 0; i != n; i++)
+ {
+ if (pi[i] > 0)
+ sum -= p[i];
+ }
+ return sum;
+}
+
+int p[16] __attribute__((aligned(__BIGGEST_ALIGNMENT__)))
+ = { 7, 3, 1, 4, 9, 10, 14, 7, -10, -55, 20, 9, 1, 2, 0, -17 };
+int pi[16] __attribute__((aligned(__BIGGEST_ALIGNMENT__)))
+ = { 0, 1, 1, 1, 0, 0, 1, 0, 1, 1, 0, 0, 1, 1, 1, 1 };
+int
+main()
+{
+ check_vect ();
+
+ if (foo (16, p, pi) != 57)
+ abort ();
+ return 0;
+}
diff --git a/gcc/testsuite/gcc.target/i386/vect-pr113078.c
b/gcc/testsuite/gcc.target/i386/vect-pr113078.c
new file mode 100644
index 00000000000..e7666054324
--- /dev/null
+++ b/gcc/testsuite/gcc.target/i386/vect-pr113078.c
@@ -0,0 +1,16 @@
+/* { dg-do compile } */
+/* { dg-options "-O3 -mavx512vl" } */
+
+int
+foo (int n, int* p, int* pi)
+{
+ int sum = 0;
+ for (int i = 0; i != n; i++)
+ {
+ if (pi[i] > 0)
+ sum -= p[i];
+ }
+ return sum;
+}
+
+/* { dg-final { scan-assembler-times "vpsub\[^\r\n\]*%k" 2 } } */
diff --git a/gcc/tree-vect-loop.cc b/gcc/tree-vect-loop.cc
index c5b2799be23..1bdad0fbe0f 100644
--- a/gcc/tree-vect-loop.cc
+++ b/gcc/tree-vect-loop.cc
@@ -4116,6 +4116,13 @@ pop:
if (op.ops[1] == op.ops[opi])
neg = ! neg;
}
+ else if (op.code == IFN_COND_SUB)
+ {
+ op.code = IFN_COND_ADD;
+ /* Track whether we negate the reduction value each iteration. */
+ if (op.ops[2] == op.ops[opi])
+ neg = ! neg;
+ }
if (CONVERT_EXPR_CODE_P (op.code)
&& tree_nop_conversion_p (op.type, TREE_TYPE (op.ops[0])))
;
--
2.35.3
