On Fri, Jun 14, 2024 at 5:54 AM Richard Biener <rguent...@suse.de> wrote:
>
> Automatic arrays that are not address-taken should not be subject to
> store data races.
That seems conservative enough. Though I would think if the array
never escaped the function would be still correct and allow for more
arrays (but maybe that is not tracked).
Thanks,
Andrew Pinski
> This applies to OMP SIMD in-branch lowered
> functions result array which for the testcase otherwise prevents
> vectorization with SSE and for AVX and AVX512 ends up with spurious
> .MASK_STORE to the stack surviving.
>
> This inefficiency was noted in PR111793.
>
> Bootstrap and regtest running on x86_64-unknown-linux-gnu. Is my
> idea of store data races correct? At least phiopt uses the same
> check but for example LIM doesn't special-case locals.
>
> PR tree-optimization/111793
> * tree-if-conv.cc (ifcvt_memrefs_wont_trap): For stores
> that do not trap only consider -fstore-data-races when
> the underlying object is not automatic or has its address
> taken.
>
> * gcc.dg/vect/vect-simd-clone-21.c: New testcase.
> ---
> gcc/testsuite/gcc.dg/vect/vect-simd-clone-21.c | 16 ++++++++++++++++
> gcc/tree-if-conv.cc | 13 +++++++++++--
> 2 files changed, 27 insertions(+), 2 deletions(-)
> create mode 100644 gcc/testsuite/gcc.dg/vect/vect-simd-clone-21.c
>
> diff --git a/gcc/testsuite/gcc.dg/vect/vect-simd-clone-21.c
> b/gcc/testsuite/gcc.dg/vect/vect-simd-clone-21.c
> new file mode 100644
> index 00000000000..49c52fb59bd
> --- /dev/null
> +++ b/gcc/testsuite/gcc.dg/vect/vect-simd-clone-21.c
> @@ -0,0 +1,16 @@
> +/* { dg-do compile } */
> +/* { dg-require-effective-target vect_simd_clones } */
> +/* { dg-additional-options "-fopenmp-simd" } */
> +
> +#pragma omp declare simd simdlen(4) inbranch
> +__attribute__((noinline)) int
> +foo (int a, int b)
> +{
> + return a + b;
> +}
> +
> +/* { dg-final { scan-tree-dump-times "vectorized 1 loops" 4 "vect" { target
> i?86-*-* x86_64-*-* } } } */
> +/* if-conversion shouldn't need to resort to masked stores for the result
> + array created by OMP lowering since that's automatic and does not have
> + its address taken. */
> +/* { dg-final { scan-tree-dump-not "MASK_STORE" "vect" } } */
> diff --git a/gcc/tree-if-conv.cc b/gcc/tree-if-conv.cc
> index c4c3ed41a44..974c614edf3 100644
> --- a/gcc/tree-if-conv.cc
> +++ b/gcc/tree-if-conv.cc
> @@ -934,14 +934,23 @@ ifcvt_memrefs_wont_trap (gimple *stmt,
> vec<data_reference_p> drs)
> if (DR_IS_READ (a))
> return true;
>
> + bool ok = flag_store_data_races;
> + base = get_base_address (base);
> + if (DECL_P (base)
> + && auto_var_in_fn_p (base, cfun->decl)
> + && ! may_be_aliased (base))
> + /* Automatic variables not aliased are not subject to
> + data races. */
> + ok = true;
> +
> /* an unconditionaly write won't trap if the base is written
> to unconditionally. */
> if (base_master_dr
> && DR_BASE_W_UNCONDITIONALLY (*base_master_dr))
> - return flag_store_data_races;
> + return ok;
> /* or the base is known to be not readonly. */
> else if (base_object_writable (DR_REF (a)))
> - return flag_store_data_races;
> + return ok;
> }
>
> return false;
> --
> 2.35.3
