On Tue, Oct 8, 2024 at 10:19 AM Richard Biener <rguent...@suse.de> wrote:
>
> 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 } } */
FYI, you don't have to guard compile tests with ISA target selectors.
This is intended for assembly and execute tests, where assembler can't
handle instructions from selected ISA.
Uros.
