[PATCH] Fix -march=knl wrong-code (PR target/84786)
Uros Bizjak
ubizjak@gmail.com
Tue Mar 13 07:54:00 GMT 2018
On Mon, Mar 12, 2018 at 9:52 PM, Jakub Jelinek <jakub@redhat.com> wrote:
> Hi!
>
> For TARGET_AVX512F && !TARGET_AVX512VL, the RA ensures that
> 128-bit and 256-bit vectors aren't allocated in %xmm16+ registers,
> but not so for scalar modes. In that case we use avx512vl isa attribute,
> or TARGET_AVX512VL guards, or Yv constraint.
>
> The following patch does the last one for sse2_loadhpd pattern where the
> last argument has DFmode and thus using v constraint for it is unsafe
> and on the testcase we emit
> vunpcklpd %xmm18, %xmm0, %xmm0
> instruction which needs avx512vl. Bootstrapped/regtested on x86_64-linux
> and i686-linux, ok for trunk?
>
> 2018-03-12 Jakub Jelinek <jakub@redhat.com>
>
> PR target/84786
> * config/i386/sse.md (sse2_loadhpd): Use Yv constraint rather than v
> on the last operand.
>
> * gcc.target/i386/avx512f-pr84786-1.c: New test.
> * gcc.target/i386/avx512f-pr84786-2.c: New test.
OK for trunk and backports.
Thanks,
Uros.
> --- gcc/config/i386/sse.md.jj 2018-03-05 17:00:24.568655800 +0100
> +++ gcc/config/i386/sse.md 2018-03-12 11:05:48.917401886 +0100
> @@ -9022,14 +9022,14 @@ (define_expand "sse2_loadhpd_exp"
> ;; see comment above inline_secondary_memory_needed function in i386.c
> (define_insn "sse2_loadhpd"
> [(set (match_operand:V2DF 0 "nonimmediate_operand"
> - "=x,v,x,v,o,o ,o")
> + "=x,v,x,v ,o,o ,o")
> (vec_concat:V2DF
> (vec_select:DF
> (match_operand:V2DF 1 "nonimmediate_operand"
> - " 0,v,0,v,0,0 ,0")
> + " 0,v,0,v ,0,0 ,0")
> (parallel [(const_int 0)]))
> (match_operand:DF 2 "nonimmediate_operand"
> - " m,m,x,v,x,*f,r")))]
> + " m,m,x,Yv,x,*f,r")))]
> "TARGET_SSE2 && !(MEM_P (operands[1]) && MEM_P (operands[2]))"
> "@
> movhpd\t{%2, %0|%0, %2}
> --- gcc/testsuite/gcc.target/i386/avx512f-pr84786-1.c.jj 2018-03-12 11:32:33.563665173 +0100
> +++ gcc/testsuite/gcc.target/i386/avx512f-pr84786-1.c 2018-03-12 11:35:33.964695384 +0100
> @@ -0,0 +1,25 @@
> +/* PR target/84786 */
> +/* { dg-do run { target { ! ia32 } } } */
> +/* { dg-options "-mavx512f -mno-avx512vl -O2" } */
> +/* { dg-require-effective-target avx512f } */
> +
> +#include "avx512f-check.h"
> +
> +typedef double V __attribute__((vector_size (16)));
> +
> +__attribute__((noipa)) V
> +foo (V x, double y)
> +{
> + register double z __asm ("xmm18");
> + asm volatile ("" : "=v" (z) : "0" (y));
> + x[1] = z;
> + return x;
> +}
> +
> +static void
> +avx512f_test (void)
> +{
> + V a = foo ((V) { 1.0, 2.0 }, 3.0);
> + if (a[0] != 1.0 || a[1] != 3.0)
> + abort ();
> +}
> --- gcc/testsuite/gcc.target/i386/avx512f-pr84786-2.c.jj 2018-03-12 11:32:43.445666826 +0100
> +++ gcc/testsuite/gcc.target/i386/avx512f-pr84786-2.c 2018-03-12 11:35:45.260697279 +0100
> @@ -0,0 +1,16 @@
> +/* PR target/84786 */
> +/* { dg-do compile { target { ! ia32 } } } */
> +/* { dg-options "-mavx512f -mno-avx512vl -O2" } */
> +
> +typedef double V __attribute__((vector_size (16)));
> +
> +__attribute__((noipa)) V
> +foo (V x, double y)
> +{
> + register double z __asm ("xmm18");
> + asm volatile ("" : "=v" (z) : "0" (y));
> + x[1] = z;
> + return x;
> +}
> +
> +/* { dg-final { scan-assembler-not "vunpcklpd\[\^\n\r]*xmm(1\[6-9]|\[23]\[0-9])" } } */
>
> Jakub
More information about the Gcc-patches
mailing list