[Bug tree-optimization/89049] [8/9/10/11 Regression] Unexpected vectorization

["rguenth at gcc dot gnu.org" <gcc-bugzilla@gcc.gnu.org>]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=89049

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
           Assignee|rguenth at gcc dot gnu.org         |unassigned at gcc dot gnu.org
      Known to fail|                            |11.0
   Last reconfirmed|2019-01-25 00:00:00         |2021-3-11
             Status|ASSIGNED                    |NEW

--- Comment #13 from Richard Biener <rguenth at gcc dot gnu.org> ---
Re-confirmed.  Note the same in-order reduction is profitable with SSE:

0x3704d70 *_3 1 times scalar_load costs 12 in body
0x3704d70 _4 + r_16 1 times scalar_stmt costs 12 in body

0x3825d80 _4 + r_16 4 times vec_to_scalar costs 16 in body
0x3825d80 _4 + r_16 4 times scalar_stmt costs 48 in body
0x3825d80 *_3 1 times unaligned_load (misalign -1) costs 12 in body

t4.c:1:53: note:  Cost model analysis:
  Vector inside of loop cost: 76
  Vector prologue cost: 0
  Vector epilogue cost: 0
  Scalar iteration cost: 24
  Scalar outside cost: 0
  Vector outside cost: 0
  prologue iterations: 0
  epilogue iterations: 0
  Calculated minimum iters for profitability: 0


bar:
.LFB0:
        .cfi_startproc
        leaq    4096(%rdi), %rax
        pxor    %xmm1, %xmm1
        .p2align 4,,10
        .p2align 3
.L2:
        movups  (%rdi), %xmm0
        addq    $16, %rdi
        addss   %xmm0, %xmm1
        movaps  %xmm0, %xmm2
        shufps  $85, %xmm0, %xmm2
        addss   %xmm2, %xmm1
        movaps  %xmm0, %xmm2
        unpckhps        %xmm0, %xmm2
        shufps  $255, %xmm0, %xmm0
        addss   %xmm2, %xmm1
        addss   %xmm0, %xmm1
        cmpq    %rdi, %rax
        jne     .L2
        cvttss2sil      %xmm1, %eax
        ret

iff the code in forwprop that decomposes loads of BLKmode vectors used only
by BIT_FIELD_REFs would be relaxed to cover all modes and TARGET_MEM_REFs
that are simple we'd get

.L2:
        addl    $1, %eax
        addss   (%rdi), %xmm0
        addss   4(%rdi), %xmm0
        addq    $16, %rdi
        addss   -8(%rdi), %xmm0
        addss   -4(%rdi), %xmm0
        cmpl    $256, %eax
        jne     .L2