Convert be sinked into a vec_cond if both sides
fold. Unlike other unary operations, we need to check that we still can handle
this vec_cond's first operand is the same as the new truth type.
I tried a few different versions of this patch:
view_convert to the new truth_type but that does not work as we always support
all vec_cond
afterwards.
using expand_vec_cond_expr_p; but that would allow too much.
I also tried to see if view_convert can be handled here but we end up with:
_3 = VEC_COND_EXPR <_2, { Nan(-1), Nan(-1), Nan(-1), Nan(-1) }, { 0.0,
0.0, 0.0, 0.0 }>;
Which isel does not know how to handle as just being a view_convert from
`vector(4) <signed-boolean:32>`
to `vector(4) float` and causes a regression with `g++.target/i386/pr88152.C`
Note, in the case of the SVE testcase, we will sink negate after the convert
and be able
to remove a few extra instructions in the end.
Also with this change gcc.target/aarch64/sve/cond_unary_5.c will now pass.
OK? Bootstrapped and tested on x86_64-linux-gnu and aarch64-linux-gnu.
gcc/ChangeLog:
PR tree-optimization/111006
PR tree-optimization/110986
* match.pd: (op(vec_cond(a,b,c))): Handle convert for op.
gcc/testsuite/ChangeLog:
PR tree-optimization/111006
* gcc.target/aarch64/sve/cond_convert_7.c: New test.
---
gcc/match.pd | 9 ++++++++
.../gcc.target/aarch64/sve/cond_convert_7.c | 23 +++++++++++++++++++
2 files changed, 32 insertions(+)
create mode 100644 gcc/testsuite/gcc.target/aarch64/sve/cond_convert_7.c
diff --git a/gcc/match.pd b/gcc/match.pd
index acd2a964917..ca5ab6f289d 100644
--- a/gcc/match.pd
+++ b/gcc/match.pd
@@ -4704,6 +4704,15 @@ DEFINE_INT_AND_FLOAT_ROUND_FN (RINT)
(op (vec_cond:s @0 @1 @2))
(vec_cond @0 (op! @1) (op! @2))))
+/* Sink unary conversions to branches, but only if we do fold both
+ and the target's truth type is the same as we already have. */
+(for op (convert)
+ (simplify
+ (op (vec_cond:s @0 @1 @2))
+ (if (VECTOR_TYPE_P (type)
+ && types_match (TREE_TYPE (@0), truth_type_for (type)))
+ (vec_cond @0 (op! @1) (op! @2)))))
+
/* Sink binary operation to branches, but only if we can fold it. */
(for op (tcc_comparison plus minus mult bit_and bit_ior bit_xor
lshift rshift rdiv trunc_div ceil_div floor_div round_div
diff --git a/gcc/testsuite/gcc.target/aarch64/sve/cond_convert_7.c
b/gcc/testsuite/gcc.target/aarch64/sve/cond_convert_7.c
new file mode 100644
index 00000000000..4bb95b92195
--- /dev/null
+++ b/gcc/testsuite/gcc.target/aarch64/sve/cond_convert_7.c
@@ -0,0 +1,23 @@
+/* { dg-do compile } */
+/* { dg-options "-O2 -ftree-vectorize -moverride=sve_width=256
-fdump-tree-optimized" } */
+
+/* This is a modified reduced version of cond_unary_5.c */
+
+void __attribute__ ((noipa))
+f0 (unsigned short *__restrict r,
+ int *__restrict a,
+ int *__restrict pred)
+{
+ for (int i = 0; i < 1024; ++i)
+ {
+ int p = pred[i]?-1:0;
+ r[i] = p ;
+ }
+}
+
+/* { dg-final { scan-assembler-times {\tmov\tz[0-9]+\.h, p[0-7]+/z, #-1} 1 } }
*/
+/* { dg-final { scan-assembler-not {\tmov\tz[0-9]+\.[hs], p[0-7]+/z, #1} } } */
+
+/* { dg-final { scan-tree-dump-not "VIEW_CONVERT_EXPR " "optimized" } } */
+/* { dg-final { scan-tree-dump-not " = -" "optimized" } } */
+/* { dg-final { scan-tree-dump-not " = \\\(vector" "optimized" } } */
--
2.31.1
