[Bug target/114323] [14 Regression] MVE vector load intrinsic miscompiled since r14-5622-g4d7647edfd7d98

[prathamesh3492 at gcc dot gnu.org via Gcc-bugs]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=114323

prathamesh3492 at gcc dot gnu.org changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |prathamesh3492 at gcc dot 
gnu.org

--- Comment #3 from prathamesh3492 at gcc dot gnu.org ---
Just to expand on previous comments:
Before patch, input to dse is:

  uint32x4_t D.13560;
  const uint32_t D.13545[4];
  uint32x4_t V0;
  __simd128_uint32_t _7;

  <bb 2> :
  # .MEM_2 = VDEF <.MEM_1(D)>
  D.13545 = *.LC0;
  # .MEM_8 = VDEF <.MEM_2>
  _7 = __builtin_mve_vld1q_uv4si (&D.13545);
  # .MEM_6 = VDEF <.MEM_8>
  D.13545 ={v} {CLOBBER(eos)};
  # VUSE <.MEM_6>
  return _7;

In this case, we have following virtual def-use chain:
.MEM_1(D) -> .MEM_2 -> .MEM_8 -> .MEM_6


However after patch, input to dse is:
  const uint32_t D.13539[4];
  uint32x4_t V0;

  <bb 2> :
  # .MEM_2 = VDEF <.MEM_1(D)>
  D.13539 = *.LC0;
  V0_3 = vld1q_u32 (&D.13539);
  # .MEM_5 = VDEF <.MEM_2>
  D.13539 ={v} {CLOBBER(eos)};
  # VUSE <.MEM_5>
  return V0_3;

There's a missing use of MEM_2 in call to vld1q_u32, and
since the only use of MEM_2 now is in clobber statement,
dse considers it as a dead store, and simplifies it to:

  <bb 2> :
  V0_3 = vld1q_u32 (&D.13539);
  # .MEM_5 = VDEF <.MEM_1(D)>
  D.13539 ={v} {CLOBBER(eos)};
  # VUSE <.MEM_5>
  return V0_3;

thus passing uninitialized pointer to vld1q_u32.

Thanks,
Prathamesh