On Fri, Mar 3, 2017 at 8:44 PM, Jakub Jelinek <ja...@redhat.com> wrote:
> On Fri, Mar 03, 2017 at 12:18:09PM +0100, Uros Bizjak wrote:
>> Yes. Although expander takes care not to generate two memory
>> references, combine can propagate memory to the other operand,
>> creating semi-invalid RTX that is later resolved by RA.
>
> Here is a patch which does that.
>
> Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk?
>
> 2017-03-03 Jakub Jelinek <ja...@redhat.com>
>
> * config/i386/sse.md (sse_storehps, sse_storelps,
> avx_<castmode><avxsizesuffix>_<castmode>,
> avx512f_<castmode><avxsizesuffix>_<castmode>,
> avx512f_<castmode><avxsizesuffix>_256<castmode>): Require
> in condition that at least one operand is not a MEM.
OK.
Thanks,
Uros.
> --- gcc/config/i386/sse.md.jj 2017-03-03 17:08:53.000000000 +0100
> +++ gcc/config/i386/sse.md 2017-03-03 17:39:05.907871328 +0100
> @@ -6637,7 +6637,7 @@ (define_insn "sse_storehps"
> (vec_select:V2SF
> (match_operand:V4SF 1 "nonimmediate_operand" "v,v,o")
> (parallel [(const_int 2) (const_int 3)])))]
> - "TARGET_SSE"
> + "TARGET_SSE && !(MEM_P (operands[0]) && MEM_P (operands[1]))"
> "@
> %vmovhps\t{%1, %0|%q0, %1}
> %vmovhlps\t{%1, %d0|%d0, %1}
> @@ -6690,7 +6690,7 @@ (define_insn "sse_storelps"
> (vec_select:V2SF
> (match_operand:V4SF 1 "nonimmediate_operand" " v,v,m")
> (parallel [(const_int 0) (const_int 1)])))]
> - "TARGET_SSE"
> + "TARGET_SSE && !(MEM_P (operands[0]) && MEM_P (operands[1]))"
> "@
> %vmovlps\t{%1, %0|%q0, %1}
> %vmovaps\t{%1, %0|%0, %1}
> @@ -18661,7 +18661,7 @@ (define_insn_and_split "avx_<castmode><a
> (unspec:AVX256MODE2P
> [(match_operand:<ssehalfvecmode> 1 "nonimmediate_operand" "xm,x")]
> UNSPEC_CAST))]
> - "TARGET_AVX"
> + "TARGET_AVX && !(MEM_P (operands[0]) && MEM_P (operands[1]))"
> "#"
> "&& reload_completed"
> [(set (match_dup 0) (match_dup 1))]
> @@ -19575,7 +19575,7 @@ (define_insn_and_split "avx512f_<castmod
> (unspec:AVX512MODE2P
> [(match_operand:<ssequartermode> 1 "nonimmediate_operand" "xm,x")]
> UNSPEC_CAST))]
> - "TARGET_AVX512F"
> + "TARGET_AVX512F && !(MEM_P (operands[0]) && MEM_P (operands[1]))"
> "#"
> "&& reload_completed"
> [(set (match_dup 0) (match_dup 1))]
> @@ -19592,7 +19592,7 @@ (define_insn_and_split "avx512f_<castmod
> (unspec:AVX512MODE2P
> [(match_operand:<ssehalfvecmode> 1 "nonimmediate_operand" "xm,x")]
> UNSPEC_CAST))]
> - "TARGET_AVX512F"
> + "TARGET_AVX512F && !(MEM_P (operands[0]) && MEM_P (operands[1]))"
> "#"
> "&& reload_completed"
> [(set (match_dup 0) (match_dup 1))]
>
>
> Jakub
