[Bug c/67167] New: cilkplus vectorization problems

[Bug c/67167] New: cilkplus vectorization problems

[marcin.krotkiewski at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=67167

            Bug ID: 67167
           Summary: cilkplus vectorization problems
           Product: gcc
           Version: 5.1.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: c
          Assignee: unassigned at gcc dot gnu.org
          Reporter: marcin.krotkiewski at gmail dot com
  Target Milestone: ---

I think there is a problem with vectorization of arithmetic operations in the
cilkplus implementation in gcc. I have inspected generated asm of the following
two implementations of vector addition (a = a + b). The code is compiled with
'gcc -O3 -mavx -ftree-vectorize -fopt-info-vec -fcilkplus test.c'.


// ICC compatibility - alignment hint
#ifdef __GNUC__

#define __assume_aligned(lvalueptr, align) lvalueptr = __builtin_assume_aligned
(lvalueptr, align)

#endif
#define RESTRICT __restrict__

typedef double Double;

void test(Double * RESTRICT a, Double * RESTRICT b, int size)
{
  int i;

  __assume_aligned(a, 64);
  __assume_aligned(b, 64);

  for(i=0; i<size; i++)
    a[i] = a[i] + b[i];

}


void test_cilkplus1(Double * RESTRICT a, Double * RESTRICT b, int size)
{

  __assume_aligned(a, 64);
  __assume_aligned(b, 64);

  a[0:size] = a[0:size] + b[0:size];

}


The first code (test) is vectorized as expected - here comes the ASM:

.L4:
        vmovapd (%rdi,%r8), %ymm0
        addl    $1, %r9d
        vaddpd  (%rsi,%r8), %ymm0, %ymm0
        vmovapd %ymm0, (%rdi,%r8)
        addq    $32, %r8
        cmpl    %r9d, %ecx
        ja      .L4


On the contrary, the second function (test_cilkplus1) is not vectorized:

.L21:
        vmovsd  (%rdi,%rax), %xmm0
        movl    %ecx, %r8d
        addl    $1, %ecx
        vaddsd  (%rsi,%rax), %xmm0, %xmm0
        vmovsd  %xmm0, (%rdi,%rax)
        addq    $8, %rax
        cmpl    %r8d, %edx
        jg      .L21


Now I have made sure that the compiler understands that there is no aliasing
(restrict) and that the vectors are aligned in memory. Clearly this is enough
for the standard implementation to produce a vectorized code, but not for the
CilkPlus array notation.

[Bug c/67167] cilkplus vectorization problems

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=67167

Richard Biener <rguenth at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
           Keywords|                            |missed-optimization
             Status|UNCONFIRMED                 |NEW
   Last reconfirmed|                            |2015-08-11
                 CC|                            |kyukhin at gcc dot gnu.org
         Depends on|                            |66502
     Ever confirmed|0                           |1

--- Comment #1 from Richard Biener <rguenth at gcc dot gnu.org> ---
t.c:28:13: note: === get_loop_niters ===
t.c:28:13: note: not vectorized: number of iterations cannot be computed.
t.c:28:13: note: bad loop form.
t.c:22:6: note: vectorized 0 loops in function.

Which is probably because the loop has two IVs and a combined exit test:

  <bb 5>:
  # _32 = PHI <0(4), _21(6)>
  # _33 = PHI <0(4), _22(6)>
...
  _21 = _32 + 1;
  _22 = _33 + 1;
  _25 = size_9(D) > _22;
  _23 = size_9(D) > _21;
  _24 = _23 & _25;
  if (_24 != 0)
    goto <bb 6>;
  else
    goto <bb 7>;


related to PR66502 (proposed patch helps and makes the testcase vectorized,
but as noted in the PR regresses another case).

Not sure if this can be reasonably addressed in Cilk+ lowering as well?
(generate a single IV)


Referenced Bugs:

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=66502
[Bug 66502] SCCVN can't handle PHIs optimistically optimally

[Bug c/67167] cilkplus vectorization problems

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=67167

--- Comment #2 from Richard Biener <rguenth at gcc dot gnu.org> ---
Author: rguenth
Date: Thu Aug 13 07:06:10 2015
New Revision: 226850

URL: https://gcc.gnu.org/viewcvs?rev=226850&root=gcc&view=rev
Log:
2015-08-13  Richard Biener  <rguenther@suse.de>

        PR tree-optimization/66502
        PR tree-optimization/67167
        * tree-ssa-sccvn.c (vn_phi_compute_hash): Do not include
        backedge arguments.
        (vn_phi_lookup): Adjust.
        (vn_phi_insert): Likewise.
        (visit_phi): Prefer to value-number to another PHI node
        over value-numbering to a PHI argument.
        (init_scc_vn): Mark DFS back edges.

        * gcc.dg/tree-ssa/ssa-fre-46.c: New testcase.

Added:
    trunk/gcc/testsuite/gcc.dg/tree-ssa/ssa-fre-46.c
Modified:
    trunk/gcc/ChangeLog
    trunk/gcc/testsuite/ChangeLog
    trunk/gcc/tree-ssa-sccvn.c

[Bug c/67167] cilkplus vectorization problems

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=67167

Richard Biener <rguenth at gcc dot gnu.org> changed:

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

--- Comment #3 from Richard Biener <rguenth at gcc dot gnu.org> ---
Fixed.

[Bug c/67167] cilkplus vectorization problems

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=67167
Bug 67167 depends on bug 66502, which changed state.

Bug 66502 Summary: SCCVN can't handle PHIs optimistically optimally
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=66502

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