[Bug middle-end/64448] New: New middle-end pattern breaks vector BIF folding on AArch64.

[belagod at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=64448

            Bug ID: 64448
           Summary: New middle-end pattern breaks vector BIF folding on
                    AArch64.
           Product: gcc
           Version: 5.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: middle-end
          Assignee: unassigned at gcc dot gnu.org
          Reporter: belagod at gcc dot gnu.org

This new pattern

Author: mpolacek <mpolacek@138bc75d-0d04-0410-961f-82ee72b054a4>
Date: Wed Dec 17 11:48:33 2014 +0000

PR middle-end/63568

    match.pd: Add (x & ~m) | (y & m) -> ((x ^ y) & m) ^ x pattern.

    gcc.dg/pr63568.c: New test.

breaks BSL folding to a BIF on AArch64.

Causes this regression:

FAIL: gcc.target/aarch64/vbslq_u64_1.c scan-assembler-times bif\\tv 1


The code now generated is:

vbslq_dummy_u32:
    eor    v0.16b, v1.16b, v0.16b
    and    v0.16b, v0.16b, v2.16b
    eor    v0.16b, v1.16b, v0.16b
    ret
    .size    vbslq_dummy_u32, .-vbslq_dummy_

instead of:

vbslq_dummy_u32:
    bif    v0.16b, v1.16b, v2.16b
    ret
    .size    vbslq_dummy_u32, .-vbslq_dummy_u32

Optimized tree when folding happens:

vbslq_dummy_u32 (uint32x4_t a, uint32x4_t b, uint32x4_t mask)
{
  __Uint32x4_t _3;
  __Uint32x4_t _4;
  __Uint32x4_t _6;
  uint32x4_t _7;

  <bb 2>:
  _3 = mask_1(D) & a_2(D);
  _4 = ~mask_1(D);
  _6 = _4 & b_5(D);
  _7 = _3 | _6;
  return _7;

}

Optimized tree where folding does not happen:

vbslq_dummy_u32 (uint32x4_t a, uint32x4_t b, uint32x4_t mask)
{
  __Uint32x4_t _3;
  __Uint32x4_t _5;
  uint32x4_t _6;

  <bb 2>:
  _3 = b_1(D) ^ a_2(D);
  _5 = _3 & mask_4(D);
  _6 = b_1(D) ^ _5;
  return _6;

}

This will probably need another idiom to be caught by the BSL -> BIF folder.