[Bug regression/54390] New: [AVX] FAIL: gcc.dg/vect/no-tree-sra-bb-slp-pr50730.c

[Bug regression/54390] New: [AVX] FAIL: gcc.dg/vect/no-tree-sra-bb-slp-pr50730.c

[izamyatin at gmail dot com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=54390

             Bug #: 54390
           Summary: [AVX] FAIL: gcc.dg/vect/no-tree-sra-bb-slp-pr50730.c
    Classification: Unclassified
           Product: gcc
           Version: 4.8.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: regression
        AssignedTo: unassigned@gcc.gnu.org
        ReportedBy: izamyatin@gmail.com


For avx target on x86_64 there is one fail introduced by r190492:

FAIL: gcc.dg/vect/no-tree-sra-bb-slp-pr50730.c scan-tree-dump-times slp "not
vectorized: more than one data ref in stmt" 0

r190492:

Author: hjl <hjl@138bc75d-0d04-0410-961f-82ee72b054a4>
Date:   Fri Aug 17 19:59:46 2012 +0000

    Define MAX_FIXED_MODE_SIZE/TARGET_MEMBER_TYPE_FORCES_BLK for i386

[Bug regression/54390] [AVX] FAIL: gcc.dg/vect/no-tree-sra-bb-slp-pr50730.c

[hjl.tools at gmail dot com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=54390

--- Comment #1 from H.J. Lu <hjl.tools at gmail dot com> 2012-08-28 13:21:29 UTC ---
Without AVX, we got

    xorps    %xmm4, %xmm4
    shufps    $0xe4, %xmm4, %xmm0
    movlhps    %xmm1, %xmm0
    movaps    %xmm4, %xmm1
    movaps    %xmm2, %xmm4
    shufps    $0xe4, %xmm1, %xmm4
    movaps    %xmm4, %xmm1
    movlhps    %xmm3, %xmm1
    addps    %xmm1, %xmm0
    movhps    %xmm0, -16(%rsp)
    movq    -16(%rsp), %rax
    movlps    %xmm0, -24(%rsp)
    movq    %rax, -48(%rsp)
    movq    -48(%rsp), %xmm1
    movq    -24(%rsp), %xmm0
    ret

Adding -mtune=corei7, we got

    movq    %xmm1, -16(%rsp)
    movq    %xmm0, -24(%rsp)
    movq    %xmm2, -40(%rsp)
    movups    -24(%rsp), %xmm0
    movq    %xmm3, -32(%rsp)
    movups    -40(%rsp), %xmm1
    addps    %xmm1, %xmm0
    movups    %xmm0, -24(%rsp)
    movq    -16(%rsp), %rax
    movq    -24(%rsp), %xmm0
    movd    %rax, %xmm1
    ret

With AVX, we got

    vmovq    %xmm0, -24(%rsp)
    vmovq    %xmm1, -16(%rsp)
    vmovq    %xmm2, -40(%rsp)
    vmovq    %xmm3, -32(%rsp)
    vmovups    -24(%rsp), %xmm1
    vmovups    -40(%rsp), %xmm0
    vaddps    %xmm0, %xmm1, %xmm0
    vmovups    %xmm0, -24(%rsp)
    movq    -16(%rsp), %rax
    movq    %rax, -48(%rsp)
    vmovq    -24(%rsp), %xmm0
    vmovq    -48(%rsp), %xmm1
    ret

I think -mavx/-mtune=corei7 enables unaligned load/store which
improves vectorizer.

[Bug regression/54390] [AVX] FAIL: gcc.dg/vect/no-tree-sra-bb-slp-pr50730.c

[vbyakovl23 at gmail dot com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=54390

Vladimir Yakovlev <vbyakovl23 at gmail dot com> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |vbyakovl23 at gmail dot com

--- Comment #2 from Vladimir Yakovlev <vbyakovl23 at gmail dot com> 2012-12-28 15:40:53 UTC ---
Compiler has different behavior on the test
gcc.dg/vect/no-tree-sra-bb-slp-pr50730.c with –mavx and –mno-avx. With –mno-avx
routine get_vectype_for_scalar_type (scalar_type) at tree-vect-data-refs.c:3265
returns NULL for scalar_type “struct A” whereas with –mavx returns “vector(2)
__int128 unsigned”. 
The test is passed if constant 16 at line 6 of the test is replaced by 32 or 64
(better 64 otherwise we will have the problem with avx2 in future).

diff --git a/gcc/testsuite/gcc.dg/vect/no-tree-sra-bb-slp-pr50730.c
b/gcc/testsuite/gcc.dg/vect/no-tree-sra-bb-slp-pr50730.c
index 90dcd84..68e0bf1 100644
--- a/gcc/testsuite/gcc.dg/vect/no-tree-sra-bb-slp-pr50730.c
+++ b/gcc/testsuite/gcc.dg/vect/no-tree-sra-bb-slp-pr50730.c
@@ -3,7 +3,7 @@

 typedef __complex__ float Value;
 typedef struct {
-  Value a[16 / sizeof (Value)];
+  Value a[64 / sizeof (Value)];
 } A;

 A sum(A a,A b)

[Bug testsuite/54390] [AVX] FAIL: gcc.dg/vect/no-tree-sra-bb-slp-pr50730.c

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=54390

Andrew Pinski <pinskia at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |RESOLVED
         Resolution|---                         |FIXED
          Component|target                      |testsuite
   Target Milestone|---                         |6.0

--- Comment #3 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
The testcase was fixed in r6-4653 to match based on the basic block and then
changed again in r11-3723 to be really fixed.

Anyways fixed. It was just a testcase failure rather than a missed
optimization.