[Bug middle-end/65962] Missed vectorization of strided stores

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

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=65962

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|NEW                         |ASSIGNED
           Assignee|unassigned at gcc dot gnu.org      |rguenth at gcc dot gnu.org

--- Comment #3 from Richard Biener <rguenth at gcc dot gnu.org> ---
While strided stores are now implemented the case is still not handled because
single-element interleaving takes precedence (and single-element interleaving
isn't supported for stores as that always produces gaps).

I have a patch that produces

.L2:
        movdqu  16(%rax), %xmm1
        addq    $32, %rax
        movdqu  -32(%rax), %xmm0
        shufps  $136, %xmm1, %xmm0
        paddd   %xmm2, %xmm0
        pshufd  $85, %xmm0, %xmm1
        movd    %xmm0, -32(%rax)
        movd    %xmm1, -24(%rax)
        movdqa  %xmm0, %xmm1
        punpckhdq       %xmm0, %xmm1
        pshufd  $255, %xmm0, %xmm0
        movd    %xmm1, -16(%rax)
        movd    %xmm0, -8(%rax)
        cmpq    %rdx, %rax
        jne     .L2

when you disable the cost model.  Otherwise it's deemed not profitable.  Using
scatters for AVX could in theory make it profitable (not sure).

t.c:5:3: note: Cost model analysis:
  Vector inside of loop cost: 13
  Vector prologue cost: 1
  Vector epilogue cost: 12
  Scalar iteration cost: 3
  Scalar outside cost: 0
  Vector outside cost: 13
  prologue iterations: 0
  epilogue iterations: 4
t.c:5:3: note: cost model: the vector iteration cost = 13 divided by the scalar
iteration cost = 3 is greater or equal to the vectorization factor = 4.
t.c:5:3: note: not vectorized: vectorization not profitable.
t.c:5:3: note: not vectorized: vector version will never be profitable.

t.c:5:3: note: ==> examining statement: *_8 = _10;
t.c:5:3: note: vect_is_simple_use: operand _10
t.c:5:3: note: def_stmt: _10 = _9 + 7;
t.c:5:3: note: type of def: internal
t.c:5:3: note: vect_model_store_cost: inside_cost = 8, prologue_cost = 0 .

so the strided store has cost 8, that's 4 extracts plus 4 scalar stores.
With AVX we generate

        vmovd   %xmm0, -32(%rax)
        vpextrd $1, %xmm0, -24(%rax)
        vpextrd $2, %xmm0, -16(%rax)
        vpextrd $3, %xmm0, -8(%rax)

so it can combine extract and store, with SSE2 we get

        pshufd  $85, %xmm0, %xmm1
        movd    %xmm0, -32(%rax)
        movd    %xmm1, -24(%rax)
        movdqa  %xmm0, %xmm1
        punpckhdq       %xmm0, %xmm1
        pshufd  $255, %xmm0, %xmm0
        movd    %xmm1, -16(%rax)
        movd    %xmm0, -8(%rax)

which is even worse than expected ;)

As usual the cost model isn't target aware enough here (and it errs on the
conservative side here)