diff options
| author | Richard Biener <rguenther@suse.de> | 2021-06-24 10:47:18 +0200 |
|---|---|---|
| committer | Richard Biener <rguenther@suse.de> | 2021-06-24 11:42:09 +0200 |
| commit | 82ab14927651e0eaf0f6d6964396048bb5471f23 (patch) | |
| tree | eadf872f28e67d1f9a94d278a8d922e6246fe3c2 /gcc | |
| parent | 7619d33471c10fe3d149dcbb701d99ed3dd23528 (diff) | |
| download | gcc-82ab14927651e0eaf0f6d6964396048bb5471f23.zip gcc-82ab14927651e0eaf0f6d6964396048bb5471f23.tar.gz gcc-82ab14927651e0eaf0f6d6964396048bb5471f23.tar.bz2 | |
Fix SLP permute propagation error
This fixes SLP permute propagation to not propagate across operations
that have different semantics on different lanes like for example
the recently added COMPLEX_ADD_ROT90.
2021-06-24 Richard Biener <rguenther@suse.de>
* tree-vect-slp.c (vect_optimize_slp): Do not propagate
across operations that have different semantics on different
lanes.
Diffstat (limited to 'gcc')
| -rw-r--r-- | gcc/tree-vect-slp.c | 30 |
1 files changed, 23 insertions, 7 deletions
diff --git a/gcc/tree-vect-slp.c b/gcc/tree-vect-slp.c index 29db56e..69ee8fa 100644 --- a/gcc/tree-vect-slp.c +++ b/gcc/tree-vect-slp.c @@ -3680,18 +3680,34 @@ vect_optimize_slp (vec_info *vinfo) { int idx = ipo[i-1]; slp_tree node = vertices[idx].node; - /* For leafs there's nothing to do - we've seeded permutes - on those above. */ - if (SLP_TREE_DEF_TYPE (node) != vect_internal_def) + + /* Handle externals and constants optimistically throughout the + iteration. */ + if (SLP_TREE_DEF_TYPE (node) == vect_external_def + || SLP_TREE_DEF_TYPE (node) == vect_constant_def) continue; vertices[idx].visited = 1; - /* We cannot move a permute across a store. */ - if (STMT_VINFO_DATA_REF (SLP_TREE_REPRESENTATIVE (node)) - && DR_IS_WRITE - (STMT_VINFO_DATA_REF (SLP_TREE_REPRESENTATIVE (node)))) + /* We do not handle stores with a permutation. */ + stmt_vec_info rep = SLP_TREE_REPRESENTATIVE (node); + if (STMT_VINFO_DATA_REF (rep) + && DR_IS_WRITE (STMT_VINFO_DATA_REF (rep))) continue; + /* We cannot move a permute across an operation that is + not independent on lanes. Note this is an explicit + negative list since that's much shorter than the respective + positive one but it's critical to keep maintaining it. */ + if (is_gimple_call (STMT_VINFO_STMT (rep))) + switch (gimple_call_combined_fn (STMT_VINFO_STMT (rep))) + { + case CFN_COMPLEX_ADD_ROT90: + case CFN_COMPLEX_ADD_ROT270: + case CFN_COMPLEX_MUL: + case CFN_COMPLEX_MUL_CONJ: + continue; + default:; + } int perm = -1; for (graph_edge *succ = slpg->vertices[idx].succ; |