On Fri, Nov 10, 2023 at 2:52 AM liuhongt <hongtao....@intel.com> wrote:
>
> When I'm working on PR112443, I notice there's some misoptimizations: after we
> fold _mm{,256}_blendv_epi8/pd/ps into gimple, the backend fails to combine it
> back to v{,p}blendv{v,ps,pd} since the pattern is too complicated, so I think
> maybe we should hanlde it in the gimple level.
>
> The dump is like
>
> _1 = c_3(D) >= { 0, 0, 0, 0 };
> _2 = VEC_COND_EXPR <_1, { -1, -1, -1, -1 }, { 0, 0, 0, 0 }>;
> _7 = VIEW_CONVERT_EXPR<vector(32) char>(_2);
> _8 = VIEW_CONVERT_EXPR<vector(32) char>(b_6(D));
> _9 = VIEW_CONVERT_EXPR<vector(32) char>(a_5(D));
> _10 = _7 < { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
> 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 };
> _11 = VEC_COND_EXPR <_10, _8, _9>;
>
>
> It can be optimized to
>
> _6 = VIEW_CONVERT_EXPR<vector(32) char>(b_4(D));
> _7 = VIEW_CONVERT_EXPR<vector(32) char>(a_3(D));
> _10 = VIEW_CONVERT_EXPR<vector(32) char>(c_1(D));
> _5 = _10 >= { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
> 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 };
> _8 = VEC_COND_EXPR <_5, _6, _7>;
> _9 = VIEW_CONVERT_EXPR<__m256i>(_8);
>
> since _7 is either -1 or 0, _7 < 0 should is euqal to _1 = c_3(D) > { 0, 0,
> 0, 0 };
> The patch add a gimple pattern to handle that.
>
> Bootstrapped and regtested on x86_64-pc-linux-gnu{-m32,}
> Ok for trunk?
>
> gcc/ChangeLog:
>
> * match.pd (VCE:(a cmp b ? -1 : 0) < 0) ? c : d ---> (VCE:a cmp
> VCE:b) ? c : d): New gimple simplication.
>
> gcc/testsuite/ChangeLog:
>
> * gcc.target/i386/avx512vl-blendv-3.c: New test.
> * gcc.target/i386/blendv-3.c: New test.
> ---
> gcc/match.pd | 17 +++++++
> .../gcc.target/i386/avx512vl-blendv-3.c | 6 +++
> gcc/testsuite/gcc.target/i386/blendv-3.c | 46 +++++++++++++++++++
> 3 files changed, 69 insertions(+)
> create mode 100644 gcc/testsuite/gcc.target/i386/avx512vl-blendv-3.c
> create mode 100644 gcc/testsuite/gcc.target/i386/blendv-3.c
>
> diff --git a/gcc/match.pd b/gcc/match.pd
> index dbc811b2b38..e6f9c4fa1fd 100644
> --- a/gcc/match.pd
> +++ b/gcc/match.pd
> @@ -5170,6 +5170,23 @@ DEFINE_INT_AND_FLOAT_ROUND_FN (RINT)
> (if (optimize_vectors_before_lowering_p () && types_match (@0, @3))
> (vec_cond (bit_and @0 (bit_not @3)) @2 @1)))
Would be nice to have a comment here.
> +(for cmp (simple_comparison)
> + (simplify
> + (vec_cond
> + (lt@4 (view_convert?@5 (vec_cond (cmp @0 @1)
> + integer_all_onesp
> + integer_zerop))
> + integer_zerop) @2 @3)
> + (if (VECTOR_INTEGER_TYPE_P (TREE_TYPE (@0))
> + && VECTOR_INTEGER_TYPE_P (TREE_TYPE (@5))
> + && TYPE_SIGN (TREE_TYPE (@0)) == TYPE_SIGN (TREE_TYPE (@5))
> + && VECTOR_TYPE_P (type))
> + (with {
> + tree itype = TREE_TYPE (@5);
> + tree vbtype = TREE_TYPE (@4);}
> + (vec_cond (cmp:vbtype (view_convert:itype @0)
> + (view_convert:itype @1)) @2 @3)))))
It looks like the outer vec_cond isn't actually relevant to the simplification?
(lt (view_convert? (vec_cond (cmp @0 @1) integer_all_onesp
integer_zerop)) integer_zerop)
is the relevant part? I wonder what canonicalizes the inner vec_cond?
Did you ever see
the (view_convert ... missing?
> +
> /* c1 ? c2 ? a : b : b --> (c1 & c2) ? a : b */
> (simplify
> (vec_cond @0 (vec_cond:s @1 @2 @3) @3)
> diff --git a/gcc/testsuite/gcc.target/i386/avx512vl-blendv-3.c
> b/gcc/testsuite/gcc.target/i386/avx512vl-blendv-3.c
> new file mode 100644
> index 00000000000..2777e72ab5f
> --- /dev/null
> +++ b/gcc/testsuite/gcc.target/i386/avx512vl-blendv-3.c
> @@ -0,0 +1,6 @@
> +/* { dg-do compile } */
> +/* { dg-options "-mavx512vl -mavx512bw -O2" } */
> +/* { dg-final { scan-assembler-times {vp?blendv(?:b|p[sd])[ \t]*} 6 } } */
> +/* { dg-final { scan-assembler-not {vpcmp} } } */
> +
> +#include "blendv-3.c"
> diff --git a/gcc/testsuite/gcc.target/i386/blendv-3.c
> b/gcc/testsuite/gcc.target/i386/blendv-3.c
> new file mode 100644
> index 00000000000..fa0fb067a73
> --- /dev/null
> +++ b/gcc/testsuite/gcc.target/i386/blendv-3.c
> @@ -0,0 +1,46 @@
> +/* { dg-do compile } */
> +/* { dg-options "-mavx2 -O2" } */
> +/* { dg-final { scan-assembler-times {vp?blendv(?:b|p[sd])[ \t]*} 6 } } */
> +/* { dg-final { scan-assembler-not {vpcmp} } } */
> +
> +#include <immintrin.h>
> +
> +__m256i
> +foo (__m256i a, __m256i b, __m256i c)
> +{
> + return _mm256_blendv_epi8 (a, b, ~c < 0);
> +}
> +
> +__m256d
> +foo1 (__m256d a, __m256d b, __m256i c)
> +{
> + __m256i d = ~c < 0;
> + return _mm256_blendv_pd (a, b, (__m256d)d);
> +}
> +
> +__m256
> +foo2 (__m256 a, __m256 b, __m256i c)
> +{
> + __m256i d = ~c < 0;
> + return _mm256_blendv_ps (a, b, (__m256)d);
> +}
> +
> +__m128i
> +foo4 (__m128i a, __m128i b, __m128i c)
> +{
> + return _mm_blendv_epi8 (a, b, ~c < 0);
> +}
> +
> +__m128d
> +foo5 (__m128d a, __m128d b, __m128i c)
> +{
> + __m128i d = ~c < 0;
> + return _mm_blendv_pd (a, b, (__m128d)d);
> +}
> +
> +__m128
> +foo6 (__m128 a, __m128 b, __m128i c)
> +{
> + __m128i d = ~c < 0;
> + return _mm_blendv_ps (a, b, (__m128)d);
> +}
> --
> 2.31.1
>
