On Fri, Apr 3, 2020 at 7:06 PM Jakub Jelinek <ja...@redhat.com> wrote:
>
> Hi!
>
> The following testcase is miscompiled, because the AVX2 patterns don't
> describe correctly what the insn does. E.g. vphaddd with %ymm* operands
> (the second pattern) instruction as per:
> https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm256_hadd_epi32&expand=2941
> does { a0+a1, a2+a3, b0+b1, b2+b3, a4+a5, a6+a7, b4+b5, b6+b7 }
> but our RTL pattern did
> { a0+a1, a2+a3, a4+a5, a6+a7, b0+b1, b2+b3, b4+b5, b6+b7 }
> where the first and last 64 bits are the same and two middle 64 bits
> swapped.
> https://software.intel.com/sites/landingpage/IntrinsicsGuide/#text=_mm256_hadd_epi16&expand=2939
> similarly, insn does:
> { a0+a1, a2+a3, a4+a5, a6+a7, b0+b1, b2+b3, b4+b5, b6+b7,
> a8+a9, a10+a11, a12+a13, a14+a15, b8+b9, b10+b11, b12+b13, b14+b15 }
> but RTL pattern did
> { a0+a1, a2+a3, a4+a5, a6+a7, a8+a9, a10+a11, a12+a13, a14+a15,
> b0+b1, b2+b3, b4+b5, b6+b7, b8+b9, b10+b11, b12+b13, b14+b15 }
> again, first and last 64 bits are the same and the two middle 64 bits
> swapped.
>
> Fixed thusly, bootstrapped/regtested on x86_64-linux and i686-linux, ok for
> trunk?
>
> As a follow-up for GCC11, I have simplification for the patterns.
>
> 2020-04-03 Jakub Jelinek <ja...@redhat.com>
>
> PR target/94460
> * config/i386/sse.md (avx2_ph<plusminus_mnemonic>wv16hi3,
> avx2_ph<plusminus_mnemonic>dv8si3): Fix up RTL pattern to do
> second half of first lane from first lane of second operand and
> first half of second lane from second lane of first operand.
>
> * gcc.target/i386/avx2-pr94460.c: New test.
OK for trunk and backports.
Thanks,
Uros.
> --- gcc/config/i386/sse.md.jj 2020-03-30 18:04:31.942435289 +0200
> +++ gcc/config/i386/sse.md 2020-04-03 10:21:51.110564277 +0200
> @@ -16060,22 +16060,6 @@ (define_insn "avx2_ph<plusminus_mnemonic
> (vec_concat:V4HI
> (vec_concat:V2HI
> (ssse3_plusminus:HI
> - (vec_select:HI (match_dup 1) (parallel [(const_int 8)]))
> - (vec_select:HI (match_dup 1) (parallel [(const_int 9)])))
> - (ssse3_plusminus:HI
> - (vec_select:HI (match_dup 1) (parallel [(const_int 10)]))
> - (vec_select:HI (match_dup 1) (parallel [(const_int 11)]))))
> - (vec_concat:V2HI
> - (ssse3_plusminus:HI
> - (vec_select:HI (match_dup 1) (parallel [(const_int 12)]))
> - (vec_select:HI (match_dup 1) (parallel [(const_int 13)])))
> - (ssse3_plusminus:HI
> - (vec_select:HI (match_dup 1) (parallel [(const_int 14)]))
> - (vec_select:HI (match_dup 1) (parallel [(const_int
> 15)]))))))
> - (vec_concat:V8HI
> - (vec_concat:V4HI
> - (vec_concat:V2HI
> - (ssse3_plusminus:HI
> (vec_select:HI
> (match_operand:V16HI 2 "nonimmediate_operand" "xm")
> (parallel [(const_int 0)]))
> @@ -16089,7 +16073,23 @@ (define_insn "avx2_ph<plusminus_mnemonic
> (vec_select:HI (match_dup 2) (parallel [(const_int 5)])))
> (ssse3_plusminus:HI
> (vec_select:HI (match_dup 2) (parallel [(const_int 6)]))
> - (vec_select:HI (match_dup 2) (parallel [(const_int 7)])))))
> + (vec_select:HI (match_dup 2) (parallel [(const_int 7)]))))))
> + (vec_concat:V8HI
> + (vec_concat:V4HI
> + (vec_concat:V2HI
> + (ssse3_plusminus:HI
> + (vec_select:HI (match_dup 1) (parallel [(const_int 8)]))
> + (vec_select:HI (match_dup 1) (parallel [(const_int 9)])))
> + (ssse3_plusminus:HI
> + (vec_select:HI (match_dup 1) (parallel [(const_int 10)]))
> + (vec_select:HI (match_dup 1) (parallel [(const_int 11)]))))
> + (vec_concat:V2HI
> + (ssse3_plusminus:HI
> + (vec_select:HI (match_dup 1) (parallel [(const_int 12)]))
> + (vec_select:HI (match_dup 1) (parallel [(const_int 13)])))
> + (ssse3_plusminus:HI
> + (vec_select:HI (match_dup 1) (parallel [(const_int 14)]))
> + (vec_select:HI (match_dup 1) (parallel [(const_int 15)])))))
> (vec_concat:V4HI
> (vec_concat:V2HI
> (ssse3_plusminus:HI
> @@ -16224,21 +16224,21 @@ (define_insn "avx2_ph<plusminus_mnemonic
> (vec_select:SI (match_dup 1) (parallel [(const_int 3)]))))
> (vec_concat:V2SI
> (plusminus:SI
> - (vec_select:SI (match_dup 1) (parallel [(const_int 4)]))
> - (vec_select:SI (match_dup 1) (parallel [(const_int 5)])))
> - (plusminus:SI
> - (vec_select:SI (match_dup 1) (parallel [(const_int 6)]))
> - (vec_select:SI (match_dup 1) (parallel [(const_int 7)])))))
> - (vec_concat:V4SI
> - (vec_concat:V2SI
> - (plusminus:SI
> (vec_select:SI
> (match_operand:V8SI 2 "nonimmediate_operand" "xm")
> (parallel [(const_int 0)]))
> (vec_select:SI (match_dup 2) (parallel [(const_int 1)])))
> (plusminus:SI
> (vec_select:SI (match_dup 2) (parallel [(const_int 2)]))
> - (vec_select:SI (match_dup 2) (parallel [(const_int 3)]))))
> + (vec_select:SI (match_dup 2) (parallel [(const_int 3)])))))
> + (vec_concat:V4SI
> + (vec_concat:V2SI
> + (plusminus:SI
> + (vec_select:SI (match_dup 1) (parallel [(const_int 4)]))
> + (vec_select:SI (match_dup 1) (parallel [(const_int 5)])))
> + (plusminus:SI
> + (vec_select:SI (match_dup 1) (parallel [(const_int 6)]))
> + (vec_select:SI (match_dup 1) (parallel [(const_int 7)]))))
> (vec_concat:V2SI
> (plusminus:SI
> (vec_select:SI (match_dup 2) (parallel [(const_int 4)]))
> --- gcc/testsuite/gcc.target/i386/avx2-pr94460.c.jj 2020-04-03
> 11:21:20.932237701 +0200
> +++ gcc/testsuite/gcc.target/i386/avx2-pr94460.c 2020-04-03
> 11:20:55.566617063 +0200
> @@ -0,0 +1,31 @@
> +/* PR target/94460 */
> +/* { dg-do run { target { avx2 && int128 } } } */
> +/* { dg-options "-O2 -mavx2" } */
> +
> +#include <x86intrin.h>
> +#include "avx2-check.h"
> +
> +typedef __int128 v2ti __attribute__ ((__vector_size__ (32)));
> +
> +static inline v2ti
> +foo (__v16hi b)
> +{
> + return (v2ti) _mm256_hsub_epi16 ((__m256i) b, (__m256i) b);
> +}
> +
> +static inline v2ti
> +bar (__v8si b)
> +{
> + return (v2ti) _mm256_hsub_epi32 ((__m256i) b, (__m256i) b);
> +}
> +
> +static void
> +avx2_test (void)
> +{
> + v2ti x = foo ((__v16hi) { 1 });
> + if (x[0] != ((__int128)1 << 64 | 1) || x[1] != 0)
> + abort ();
> + x = bar ((__v8si) { 1 });
> + if (x[0] != ((__int128)1 << 64 | 1) || x[1] != 0)
> + abort ();
> +}
>
> Jakub
>
