This is the mail archive of the
gcc-patches@gcc.gnu.org
mailing list for the GCC project.
Re: [PATCH, i386] Fix alignment check for AVX-512 masked store
- From: Kirill Yukhin <kirill dot yukhin at gmail dot com>
- To: Ilya Enkovich <enkovich dot gnu at gmail dot com>
- Cc: GCC Patches <gcc-patches at gcc dot gnu dot org>, Uros Bizjak <ubizjak at gmail dot com>
- Date: Fri, 4 Dec 2015 16:45:52 +0300
- Subject: Re: [PATCH, i386] Fix alignment check for AVX-512 masked store
- Authentication-results: sourceware.org; auth=none
- References: <20151202135154 dot GA4176 at msticlxl57 dot ims dot intel dot com>
Hi Ilya,
On 02 Dec 16:51, Ilya Enkovich wrote:
> Hi,
>
> This patch fixes wrong alignment check in <avx512>_store<mode>_mask
> pattern. Currently we check a register operand instead of a memory
> one. This fixes segfault on 481.wrf compiled at -O3 for KNL target.
> I bootstrapped and tested this patch on x86_64-unknown-linux-gnu.
>
> I got a bunch of new failures:
>
> FAIL: gcc.target/i386/avx512vl-vmovapd-1.c scan-assembler-times vmovapd[ \\t]+[^{\n]*%xmm[0-9]+[^\n]*\\){%k[1-7]}(?:\n|[ \\t]+#) 1
> FAIL: gcc.target/i386/avx512vl-vmovapd-1.c scan-assembler-times vmovapd[ \\t]+[^{\n]*%xmm[0-9]+[^\n]*\\){%k[1-7]}(?:\n|[ \\t]+#) 1
> FAIL: gcc.target/i386/avx512vl-vmovapd-1.c scan-assembler-times vmovapd[ \\t]+[^{\n]*%ymm[0-9]+[^\n]*\\){%k[1-7]}(?:\n|[ \\t]+#) 1
> FAIL: gcc.target/i386/avx512vl-vmovapd-1.c scan-assembler-times vmovapd[ \\t]+[^{\n]*%ymm[0-9]+[^\n]*\\){%k[1-7]}(?:\n|[ \\t]+#) 1
> FAIL: gcc.target/i386/avx512vl-vmovaps-1.c scan-assembler-times vmovaps[ \\t]+[^{\n]*%xmm[0-9]+[^\n]*\\){%k[1-7]}(?:\n|[ \\t]+#) 1
> FAIL: gcc.target/i386/avx512vl-vmovaps-1.c scan-assembler-times vmovaps[ \\t]+[^{\n]*%xmm[0-9]+[^\n]*\\){%k[1-7]}(?:\n|[ \\t]+#) 1
> FAIL: gcc.target/i386/avx512vl-vmovaps-1.c scan-assembler-times vmovaps[ \\t]+[^{\n]*%ymm[0-9]+[^\n]*\\){%k[1-7]}(?:\n|[ \\t]+#) 1
> FAIL: gcc.target/i386/avx512vl-vmovaps-1.c scan-assembler-times vmovaps[ \\t]+[^{\n]*%ymm[0-9]+[^\n]*\\){%k[1-7]}(?:\n|[ \\t]+#) 1
>
> With patch applied test generates vmovup[sd] because memory
> references don't have proper alignment set. Since this is
> another bug and it's actually a performance one, I think
> this patch should go to trunk.
Your patch definetely fixes stability issue.
New fails are performance issues.
OK for main trunk. Pls, file regression bug in Bugzilla against new fails.
--
Thanks, K
>
> Thanks,
> Ilya
> --
> gcc/
>
> 2015-12-02 Ilya Enkovich <enkovich.gnu@gmail.com>
>
> * config/i386/sse.md (<avx512>_store<mode>_mask): Fix
> operand checked for alignment.
>
>
> diff --git a/gcc/config/i386/sse.md b/gcc/config/i386/sse.md
> index e7b517a..d65ed0c 100644
> --- a/gcc/config/i386/sse.md
> +++ b/gcc/config/i386/sse.md
> @@ -1051,7 +1051,7 @@
> sse_suffix = "<ssescalarsize>";
> }
>
> - if (misaligned_operand (operands[1], <MODE>mode))
> + if (misaligned_operand (operands[0], <MODE>mode))
> align = "u";
> else
> align = "a";