[Bug c/59084] New: Sub-optimal vector moves in AVX2 vectorized loop for unaligned loads.

[Bug c/59084] New: Sub-optimal vector moves in AVX2 vectorized loop for unaligned loads.

[hendrik.greving.intel at gmail dot com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=59084

            Bug ID: 59084
           Summary: Sub-optimal vector moves in AVX2 vectorized loop for
                    unaligned loads.
           Product: gcc
           Version: 4.9.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: c
          Assignee: unassigned at gcc dot gnu.org
          Reporter: hendrik.greving.intel at gmail dot com

The simple test case below produces sub-optimal split load/stores (AVX1/16
bytes), apparently due to the fact that g_a, g_b, g_c are put in a common
section which doesn't guarantee alignment. Compiling with -fno-common actually
produces good code. Only affects C, due to the described alignment issue above.

This bug might be related to or be a duplicate of #41464.

Sub-optimal code:

compiled with gcc -S -O3 -march=core-avx2 foo.c -ftree-vectorizer-verbose=1 -dp
-v -da

    vmovdqu    (%rsi,%rax), %xmm0    # 160    sse2_loaddquv16qi    [length = 5]
    vinserti128    $0x1, 16(%rsi,%rax), %ymm0, %ymm0    # 161   
avx_vec_concatv32qi/1    [length = 8]
    addl    $1, %edx    # 165    *addsi_1/1    [length = 3]
    vpaddd    (%r8,%rax), %ymm0, %ymm0    # 162    *addv8si3/2    [length = 6]
    vmovups    %xmm0, (%rcx,%rax)    # 410    *movv16qi_internal/3    [length =
5]
    vextracti128    $0x1, %ymm0, 16(%rcx,%rax)    # 164   
vec_extract_hi_v32qi/2    [leng

Good code:

compiled with gcc -S -O3 -march=core-avx2 foo.c -ftree-vectorizer-verbose=1 -dp
-v -da -fno-common

    vmovdqa    g_a(%rax), %ymm0    # 26    *movv8si_internal/2    [length = 8]
    vpaddd    g_b(%rax), %ymm0, %ymm0    # 27    *addv8si3/2    [length = 8]
    addq    $32, %rax    # 29    *adddi_1/1    [length = 4]
    vmovaps    %ymm0, g_c-32(%rax)    # 28    *movv8si_internal/3    [length =
8]

Test case:

#include <stdio.h>
#include <stdint.h>

#define LENGTH 10000

int g_a[LENGTH];
int g_b[LENGTH];
int g_c[LENGTH];

void
foo()
{
    int i ;
    for (i = 0; i < LENGTH; i++) {
        g_c[i] = g_a[i] + g_b[i];
    }
}

[Bug c/59084] Sub-optimal vector moves in AVX2 vectorized loop for unaligned loads.

[hjl at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=59084

--- Comment #1 from hjl at gcc dot gnu.org <hjl at gcc dot gnu.org> ---
Author: hjl
Date: Tue Nov 12 13:26:51 2013
New Revision: 204700

URL: http://gcc.gnu.org/viewcvs?rev=204700&root=gcc&view=rev
Log:
Properly handle AVX256 unaligned load and store

    PR target/59084
    * config/i386/i386.c (ix86_option_override_internal): Check
    X86_TUNE_AVX256_UNALIGNED_LOAD_OPTIMAL and
    X86_TUNE_AVX256_UNALIGNED_STORE_OPTIMAL for
    MASK_AVX256_SPLIT_UNALIGNED_LOAD and
    MASK_AVX256_SPLIT_UNALIGNED_STORE.

    * config/i386/x86-tune.def (X86_TUNE_AVX256_UNALIGNED_LOAD_OPTIMAL):
    Clear m_COREI7_AVX and update comments.
    (X86_TUNE_AVX256_UNALIGNED_STORE_OPTIMAL): Likewise.

Modified:
    trunk/gcc/ChangeLog
    trunk/gcc/config/i386/i386.c
    trunk/gcc/config/i386/x86-tune.def

[Bug c/59084] Sub-optimal vector moves in AVX2 vectorized loop for unaligned loads.

[hjl.tools at gmail dot com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=59084

H.J. Lu <hjl.tools at gmail dot com> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |RESOLVED
         Resolution|---                         |FIXED

--- Comment #2 from H.J. Lu <hjl.tools at gmail dot com> ---
Fixed.