diff options
| author | Richard Biener <rguenther@suse.de> | 2021-09-15 10:20:34 +0200 |
|---|---|---|
| committer | Richard Biener <rguenther@suse.de> | 2021-09-15 11:13:21 +0200 |
| commit | b6d8fa66e1bf08756cb4134735b5034e171f49d1 (patch) | |
| tree | a3c40c775922f33cc28c3dafb3ffacb4b3fee4f7 /gcc/tree-vect-loop.c | |
| parent | cc1e28878a228b6c4a0872e56d97ac88971b7725 (diff) | |
| download | gcc-b6d8fa66e1bf08756cb4134735b5034e171f49d1.zip gcc-b6d8fa66e1bf08756cb4134735b5034e171f49d1.tar.gz gcc-b6d8fa66e1bf08756cb4134735b5034e171f49d1.tar.bz2 | |
tree-optimization/102318 - reduction epilogue re-use
This refines the fix for PR102226 to do the mode conversion
from V2DI to VNx2DI separately from the sign-conversion, retaining
the signedness of the saved accumulator as before the original fix.
2021-09-15 Richard Biener <rguenther@suse.de>
PR tree-optimization/102318
* tree-vect-loop.c (vect_transform_cycle_phi): Revert
previous change and do the mode conversion separately from
the sign conversion.
* gcc.dg/vect/pr102318.c: New testcase.
Diffstat (limited to 'gcc/tree-vect-loop.c')
| -rw-r--r-- | gcc/tree-vect-loop.c | 13 |
1 files changed, 11 insertions, 2 deletions
diff --git a/gcc/tree-vect-loop.c b/gcc/tree-vect-loop.c index c9dcc64..5a5b8da 100644 --- a/gcc/tree-vect-loop.c +++ b/gcc/tree-vect-loop.c @@ -7755,11 +7755,20 @@ vect_transform_cycle_phi (loop_vec_info loop_vinfo, (reduc_info), &stmts); } - if (!useless_type_conversion_p (vectype_out, TREE_TYPE (def))) - def = gimple_convert (&stmts, vectype_out, def); + /* The epilogue loop might use a different vector mode, like + VNx2DI vs. V2DI. */ + if (TYPE_MODE (vectype_out) != TYPE_MODE (TREE_TYPE (def))) + { + tree reduc_type = build_vector_type_for_mode + (TREE_TYPE (TREE_TYPE (def)), TYPE_MODE (vectype_out)); + def = gimple_convert (&stmts, reduc_type, def); + } /* Adjust the input so we pick up the partially reduced value for the skip edge in vect_create_epilog_for_reduction. */ accumulator->reduc_input = def; + /* And the reduction could be carried out using a different sign. */ + if (!useless_type_conversion_p (vectype_out, TREE_TYPE (def))) + def = gimple_convert (&stmts, vectype_out, def); if (loop_vinfo->main_loop_edge) { /* While we'd like to insert on the edge this will split |