diff options
| author | Richard Biener <rguenther@suse.de> | 2021-01-08 13:17:18 +0100 |
|---|---|---|
| committer | Richard Biener <rguenther@suse.de> | 2021-01-08 14:08:10 +0100 |
| commit | bdcde1504502719504a7a63ab10059e171694dc2 (patch) | |
| tree | 5a143e25895b14de5fb6abf1c73e78b6ea3873a5 /gcc/tree-vect-slp.c | |
| parent | 1b885264a48dcd71b7aeb26c0abeb91246724897 (diff) | |
| download | gcc-bdcde1504502719504a7a63ab10059e171694dc2.zip gcc-bdcde1504502719504a7a63ab10059e171694dc2.tar.gz gcc-bdcde1504502719504a7a63ab10059e171694dc2.tar.bz2 | |
tree-optimization/98544 - more permute optimization fixes
Permute nodes are not transparent to the permute of their children.
Instead we have to materialize child permutes always and in future
may treat permute nodes as the source of arbitrary permutes as
we can permute the lane permutation vector at will (as the target
supports in the end).
2021-01-08 Richard Biener <rguenther@suse.de>
PR tree-optimization/98544
* tree-vect-slp.c (vect_optimize_slp): Always materialize
permutes at a permute node.
* gcc.dg/vect/bb-slp-pr98544.c: New testcase.
Diffstat (limited to 'gcc/tree-vect-slp.c')
| -rw-r--r-- | gcc/tree-vect-slp.c | 34 |
1 files changed, 21 insertions, 13 deletions
diff --git a/gcc/tree-vect-slp.c b/gcc/tree-vect-slp.c index c9da845..e0f3539 100644 --- a/gcc/tree-vect-slp.c +++ b/gcc/tree-vect-slp.c @@ -3029,19 +3029,27 @@ vect_optimize_slp (vec_info *vinfo) /* Decide on permute materialization. Look whether there's a use (pred) edge that is permuted differently than us. - In that case mark ourselves so the permutation is applied. */ - bool all_preds_permuted = slpg->vertices[idx].pred != NULL; - for (graph_edge *pred = slpg->vertices[idx].pred; - pred; pred = pred->pred_next) - { - gcc_checking_assert (bitmap_bit_p (n_visited, pred->src)); - int pred_perm = n_perm[pred->src]; - if (!vect_slp_perms_eq (perms, perm, pred_perm)) - { - all_preds_permuted = false; - break; - } - } + In that case mark ourselves so the permutation is applied. + For VEC_PERM_EXPRs the permutation doesn't carry along + from children to parents so force materialization at the + point of the VEC_PERM_EXPR. In principle VEC_PERM_EXPRs + are a source of an arbitrary permutation again, similar + to constants/externals - that's something we do not yet + optimally handle. */ + bool all_preds_permuted = (SLP_TREE_CODE (node) != VEC_PERM_EXPR + && slpg->vertices[idx].pred != NULL); + if (all_preds_permuted) + for (graph_edge *pred = slpg->vertices[idx].pred; + pred; pred = pred->pred_next) + { + gcc_checking_assert (bitmap_bit_p (n_visited, pred->src)); + int pred_perm = n_perm[pred->src]; + if (!vect_slp_perms_eq (perms, perm, pred_perm)) + { + all_preds_permuted = false; + break; + } + } if (!all_preds_permuted) { if (!bitmap_bit_p (n_materialize, idx)) |