https://gcc.gnu.org/bugzilla/show_bug.cgi?id=127885
--- Comment #2 from Richard Biener <rguenth at gcc dot gnu.org> ---
t.c:37:59: note: Cost model analysis for whole subgraph:
Vector cost: 216
Scalar cost: 208
t.c:37:59: missed: not vectorized: vectorization is not profitable.
We are now appropriately costing
t.c:37:59: note: op template: mask_18 = foo_ptr.0_1->explicit_mask[0];
t.c:37:59: note: [l*]stmt 0 mask_18 = foo_ptr.0_1->explicit_mask[0];
t.c:37:59: note: [l] stmt 1 mask_18 = foo_ptr.0_1->explicit_mask[0];
t.c:37:59: note: [l] stmt 2 mask_18 = foo_ptr.0_1->explicit_mask[0];
t.c:37:59: note: [l] stmt 3 mask_18 = foo_ptr.0_1->explicit_mask[0];
t.c:37:59: note: load permutation { 0 0 0 0 }
as
foo_ptr.0_1->explicit_mask[0] 1 times vec_to_scalar costs 16 in body
and code generated as
vect_mask_18.8_39 = MEM <vector(4) long unsigned int> [(long unsigned int
*)vectp.7_38];
vect_mask_18.9_40 = VEC_PERM_EXPR <vect_mask_18.8_39, vect_mask_18.8_39, { 0,
0, 0, 0 }>;
_41 = BIT_FIELD_REF <vect_mask_18.9_40, 64, 0>;
_2 = (int) _41;
x_param_align_loop_iterations = _2;
and
vmovdqu (%rcx), %ymm2
vpermq $0, %ymm2, %ymm0
vmovd %xmm2, x_param_align_loop_iterations(%rip)
that's the missed anticipation of a store-use I noted in the patch submission.
But I costing shows this was only profitable by a very low margin, 204 vs 208.
Costing a single splat more appropriately or rather code-generating it as such
node 0x5866240 foo_ptr.0_1->explicit_mask[0] 1 times vec_perm costs 4 in body
node 0x5866240 foo_ptr.0_1->explicit_mask[0] 1 times unaligned_load (misalign
-1) costs 20 in body
node 0x5866240 foo_ptr.0_1->explicit_mask[0] 1 times vec_to_scalar costs 16 in
body
can recover that. But all cost adjustments come with regressions.
