On Wed, Jun 26, 2019 at 10:11 AM Jakub Jelinek <ja...@redhat.com> wrote:
>
> Hi!
>
> The following testcase is miscompiled starting with my PR85480 change.
> While it is perfectly fine to use "xm" or "vm" constraints for the source
> operand when the other operand is "C", we rely on the AVX/AVX512 behavior
> that most 128-bit or 256-bit vector instructions clear the upper bits of the
> 512-bit vectors, but if the source is in memory, we need to check if it is
> aligned or maybe misaligned and use corresponding aligned or unaligned loads
> accordingly, rather than always aligned ones.
>
> Fixed thusly, bootstrapped/regtested on x86_64-linux and i686-linux, ok for
> trunk and 9.2?
>
> Note, I have a follow-up patch to improve avx_vec_concat<mode>, just will
> need to test it.
>
> 2019-06-26 Jakub Jelinek <ja...@redhat.com>
>
> PR target/90991
> * config/i386/sse.md
> (*<extract_type>_vinsert<shuffletype><extract_suf>_0): Use vmovupd,
> vmovups, vmovdqu, vmovdqu32 or vmovdqu64 instead of the aligned
> insns if operands[2] is misaligned_operand.
>
> * gcc.target/i386/avx512dq-pr90991-1.c: New test.
OK, looks even obvious to me.
Thanks,
Uros.
> --- gcc/config/i386/sse.md.jj 2019-06-21 08:43:17.734263742 +0200
> +++ gcc/config/i386/sse.md 2019-06-25 22:36:12.955476294 +0200
> @@ -13744,15 +13744,29 @@ (define_insn "*<extract_type>_vinsert<sh
> switch (<MODE>mode)
> {
> case E_V8DFmode:
> - return "vmovapd\t{%2, %x0|%x0, %2}";
> + if (misaligned_operand (operands[2], <ssequartermode>mode))
> + return "vmovupd\t{%2, %x0|%x0, %2}";
> + else
> + return "vmovapd\t{%2, %x0|%x0, %2}";
> case E_V16SFmode:
> - return "vmovaps\t{%2, %x0|%x0, %2}";
> + if (misaligned_operand (operands[2], <ssequartermode>mode))
> + return "vmovups\t{%2, %x0|%x0, %2}";
> + else
> + return "vmovaps\t{%2, %x0|%x0, %2}";
> case E_V8DImode:
> - return which_alternative == 2 ? "vmovdqa64\t{%2, %x0|%x0, %2}"
> - : "vmovdqa\t{%2, %x0|%x0, %2}";
> + if (misaligned_operand (operands[2], <ssequartermode>mode))
> + return which_alternative == 2 ? "vmovdqu64\t{%2, %x0|%x0, %2}"
> + : "vmovdqu\t{%2, %x0|%x0, %2}";
> + else
> + return which_alternative == 2 ? "vmovdqa64\t{%2, %x0|%x0, %2}"
> + : "vmovdqa\t{%2, %x0|%x0, %2}";
> case E_V16SImode:
> - return which_alternative == 2 ? "vmovdqa32\t{%2, %x0|%x0, %2}"
> - : "vmovdqa\t{%2, %x0|%x0, %2}";
> + if (misaligned_operand (operands[2], <ssequartermode>mode))
> + return which_alternative == 2 ? "vmovdqu32\t{%2, %x0|%x0, %2}"
> + : "vmovdqu\t{%2, %x0|%x0, %2}";
> + else
> + return which_alternative == 2 ? "vmovdqa32\t{%2, %x0|%x0, %2}"
> + : "vmovdqa\t{%2, %x0|%x0, %2}";
> default:
> gcc_unreachable ();
> }
> --- gcc/testsuite/gcc.target/i386/avx512dq-pr90991-1.c.jj 2019-06-25
> 23:17:44.831272146 +0200
> +++ gcc/testsuite/gcc.target/i386/avx512dq-pr90991-1.c 2019-06-25
> 23:27:27.732357035 +0200
> @@ -0,0 +1,47 @@
> +/* PR target/90991 */
> +/* { dg-do compile } */
> +/* { dg-options "-O2 -mavx512dq -masm=att" } */
> +/* { dg-final { scan-assembler-times "vmovaps\[ \t]\+\\(\[^\n\r]*\\), %xmm0"
> 1 } } */
> +/* { dg-final { scan-assembler-times "vmovapd\[ \t]\+\\(\[^\n\r]*\\), %xmm0"
> 1 } } */
> +/* { dg-final { scan-assembler-times "vmovdqa\[ \t]\+\\(\[^\n\r]*\\), %xmm0"
> 1 } } */
> +/* { dg-final { scan-assembler-times "vmovups\[ \t]\+\\(\[^\n\r]*\\), %xmm0"
> 1 } } */
> +/* { dg-final { scan-assembler-times "vmovupd\[ \t]\+\\(\[^\n\r]*\\), %xmm0"
> 1 } } */
> +/* { dg-final { scan-assembler-times "vmovdqu\[ \t]\+\\(\[^\n\r]*\\), %xmm0"
> 1 } } */
> +
> +#include <x86intrin.h>
> +
> +__m512
> +f1 (void *a)
> +{
> + return _mm512_insertf32x4 (_mm512_set1_ps (0.0f), _mm_load_ps (a), 0);
> +}
> +
> +__m512d
> +f2 (void *a)
> +{
> + return _mm512_insertf64x2 (_mm512_set1_pd (0.0), _mm_load_pd (a), 0);
> +}
> +
> +__m512i
> +f3 (void *a)
> +{
> + return _mm512_inserti32x4 (_mm512_set1_epi32 (0), _mm_load_si128 (a), 0);
> +}
> +
> +__m512
> +f4 (void *a)
> +{
> + return _mm512_insertf32x4 (_mm512_set1_ps (0.0f), _mm_loadu_ps (a), 0);
> +}
> +
> +__m512d
> +f5 (void *a)
> +{
> + return _mm512_insertf64x2 (_mm512_set1_pd (0.0), _mm_loadu_pd (a), 0);
> +}
> +
> +__m512i
> +f6 (void *a)
> +{
> + return _mm512_inserti32x4 (_mm512_set1_epi32 (0), _mm_loadu_si128 (a), 0);
> +}
>
> Jakub
