On Mon, Nov 6, 2017 at 10:23 PM, Jakub Jelinek <ja...@redhat.com> wrote:
> Hi!
>
> Without the following patch we emit kmovb %k1, %eax; testb %al, %al
> when if just testing the Zero bit we can as well do ktestb %k1, %k1.
>
> Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk?
>
> 2017-11-06 Jakub Jelinek <ja...@redhat.com>
>
> PR target/82855
> * config/i386/i386.md (SWI1248_AVX512BWDQ2_64): New mode iterator.
> (*cmp<mode>_ccz_1): New insn with k alternative.
>
> * gcc.target/i386/avx512dq-pr82855.c: New test.
>
> --- gcc/config/i386/i386.md.jj 2017-10-28 09:00:44.000000000 +0200
> +++ gcc/config/i386/i386.md 2017-11-06 12:57:44.171215745 +0100
> @@ -1275,6 +1275,26 @@ (define_expand "cmp<mode>_1"
> (compare:CC (match_operand:SWI48 0 "nonimmediate_operand")
> (match_operand:SWI48 1 "<general_operand>")))])
>
> +(define_mode_iterator SWI1248_AVX512BWDQ2_64
> + [(QI "TARGET_AVX512DQ") (HI "TARGET_AVX512DQ")
> + (SI "TARGET_AVX512BW") (DI "TARGET_AVX512BW && TARGET_64BIT")])
> +
> +(define_insn "*cmp<mode>_ccz_1"
> + [(set (reg FLAGS_REG)
> + (compare (match_operand:SWI1248_AVX512BWDQ2_64 0
> + "nonimmediate_operand" "<r>,k,?m<r>")
Please put "k" alternative as the last alternative. I suggest to
decorate it with "*", so register allocator won't use mask registers
instead of integer registers. This would be the same solution as in
movdi_internal/movsi_internal patterns.
Uros.
> + (match_operand:SWI1248_AVX512BWDQ2_64 1 "const0_operand")))]
> + "ix86_match_ccmode (insn, CCZmode)"
> + "@
> + test{<imodesuffix>}\t%0, %0
> + ktest<mskmodesuffix>\t%0, %0
> + cmp{<imodesuffix>}\t{%1, %0|%0, %1}"
> + [(set_attr "type" "test,msklog,icmp")
> + (set_attr "length_immediate" "0,*,1")
> + (set_attr "modrm_class" "op0,*,unknown")
> + (set_attr "prefix" "*,vex,*")
> + (set_attr "mode" "<MODE>")])
> +
> (define_insn "*cmp<mode>_ccno_1"
> [(set (reg FLAGS_REG)
> (compare (match_operand:SWI 0 "nonimmediate_operand" "<r>,?m<r>")
> --- gcc/testsuite/gcc.target/i386/avx512dq-pr82855.c.jj 2017-11-06
> 13:51:21.380463119 +0100
> +++ gcc/testsuite/gcc.target/i386/avx512dq-pr82855.c 2017-11-06
> 13:52:07.164875064 +0100
> @@ -0,0 +1,14 @@
> +/* PR target/82855 */
> +/* { dg-do compile } */
> +/* { dg-options "-O2 -mavx512vl -mavx512dq" } */
> +/* { dg-final { scan-assembler {\mktestb\M} } } */
> +
> +#include <immintrin.h>
> +
> +int
> +foo (const __m256i *ptr)
> +{
> + __m256i v = _mm256_loadu_si256 (ptr);
> + __mmask8 m = _mm256_cmpeq_epi32_mask (v, _mm256_setzero_si256 ());
> + return 0 == m;
> +}
>
> Jakub
