tree-optimization/116258 - do not lower PAREN_EXPR of vectors
Checks
Commit Message
The following avoids lowering of PAREN_EXPR of vectors as unsupported
to scalars. Instead PAREN_EXPR is like a plain move or a VIEW_CONVERT.
Bootstrapped and tested on x86_64-unknown-linux-gnu, pushed. I plan
to push to branches, __builtin_assoc_barrier is new in GCC 12.
PR tree-optimization/116258
* tree-vect-generic.cc (expand_vector_operations_1): Do not
lower PAREN_EXPR.
* gcc.target/i386/pr116258.c: New testcase.
---
gcc/testsuite/gcc.target/i386/pr116258.c | 14 ++++++++++++++
gcc/tree-vect-generic.cc | 9 +++++++--
2 files changed, 21 insertions(+), 2 deletions(-)
create mode 100644 gcc/testsuite/gcc.target/i386/pr116258.c
new file mode 100644
@@ -0,0 +1,14 @@
+/* { dg-do compile } */
+/* { dg-options "-O2 -msse2" } */
+
+#define vect16 __attribute__((vector_size(16)))
+#define h(a) __builtin_assoc_barrier((a))
+
+ vect16 float f( vect16 float x, vect16 float vconstants0)
+{
+ vect16 float t = (x * (vconstants0[0]));
+ return (x + h(t));
+}
+
+/* { dg-final { scan-assembler-times "shufps" 1 } } */
+/* { dg-final { scan-assembler-not "unpck" } } */
@@ -2206,10 +2206,15 @@ expand_vector_operations_1 (gimple_stmt_iterator *gsi,
}
}
+ /* Plain moves do not need lowering. */
+ if (code == SSA_NAME
+ || code == VIEW_CONVERT_EXPR
+ || code == PAREN_EXPR)
+ return;
+
if (CONVERT_EXPR_CODE_P (code)
|| code == FLOAT_EXPR
- || code == FIX_TRUNC_EXPR
- || code == VIEW_CONVERT_EXPR)
+ || code == FIX_TRUNC_EXPR)
return;
/* The signedness is determined from input argument. */