procyberian/gcc
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

MATCH: Sink convert for vec_cond

Browse source 
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.

Committed as approved after a 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.
This commit is contained in:
Andrew Pinski 2023-08-14 18:35:53 -07:00
parent 1e3003ca08
commit 70c50c8727
2 changed files with 31 additions and 0 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

8
gcc/match.pd
View file

@ -4704,6 +4704,14 @@ 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. */
(simplify
(convert (vec_cond:s @0 @1 @2))
(if (VECTOR_TYPE_P (type)
&& types_match (TREE_TYPE (@0), truth_type_for (type)))
(vec_cond @0 (convert! @1) (convert! @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

23
gcc/testsuite/gcc.target/aarch64/sve/cond_convert_7.c Normal file
View file

@ -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" } } */