Allow different vector types for stmt groups
authorRichard Biener <rguenther@suse.de>
Wed, 18 Nov 2020 08:36:57 +0000 (09:36 +0100)
committerRichard Biener <rguenther@suse.de>
Mon, 27 Sep 2021 08:24:12 +0000 (10:24 +0200)
commit6390c5047adb75960f86d56582e6322aaa4d9281
treeb3f3817652f28f26ddf2bc354c87f37c68736fb4
parente7b8d7020052110e5717230104e647f6235dd2c1
Allow different vector types for stmt groups

This allows vectorization (in practice non-loop vectorization) to
have a stmt participate in different vector type vectorizations.
It allows us to remove vect_update_shared_vectype and replace it
by pushing/popping STMT_VINFO_VECTYPE from SLP_TREE_VECTYPE around
vect_analyze_stmt and vect_transform_stmt.

For data-ref the situation is a bit more complicated since we
analyze alignment info with a specific vector type in mind which
doesn't play well when that changes.

So the bulk of the change is passing down the actual vector type
used for a vectorized access to the various accessors of alignment
info, first and foremost dr_misalignment but also aligned_access_p,
known_alignment_for_access_p, vect_known_alignment_in_bytes and
vect_supportable_dr_alignment.  I took the liberty to replace
ALL_CAPS macro accessors with the lower-case function invocations.

The actual changes to the behavior are in dr_misalignment which now
is the place factoring in the negative step adjustment as well as
handling alignment queries for a vector type with bigger alignment
requirements than what we can (or have) analyze(d).

vect_slp_analyze_node_alignment makes use of this and upon receiving
a vector type with a bigger alingment desire re-analyzes the DR
with respect to it but keeps an older more precise result if possible.
In this context it might be possible to do the analysis just once
but instead of analyzing with respect to a specific desired alignment
look for the biggest alignment we can compute a not unknown alignment.

The ChangeLog includes the functional changes but not the bulk due
to the alignment accessor API changes - I hope that's something good.

2021-09-17  Richard Biener  <rguenther@suse.de>

PR tree-optimization/97351
PR tree-optimization/97352
PR tree-optimization/82426
* tree-vectorizer.h (dr_misalignment): Add vector type
argument.
(aligned_access_p): Likewise.
(known_alignment_for_access_p): Likewise.
(vect_supportable_dr_alignment): Likewise.
(vect_known_alignment_in_bytes): Likewise.  Refactor.
(DR_MISALIGNMENT): Remove.
(vect_update_shared_vectype): Likewise.
* tree-vect-data-refs.c (dr_misalignment): Refactor, handle
a vector type with larger alignment requirement and apply
the negative step adjustment here.
(vect_calculate_target_alignment): Remove.
(vect_compute_data_ref_alignment): Get explicit vector type
argument, do not apply a negative step alignment adjustment
here.
(vect_slp_analyze_node_alignment): Re-analyze alignment
when we re-visit the DR with a bigger desired alignment but
keep more precise results from smaller alignments.
* tree-vect-slp.c (vect_update_shared_vectype): Remove.
(vect_slp_analyze_node_operations_1): Do not update the
shared vector type on stmts.
* tree-vect-stmts.c (vect_analyze_stmt): Push/pop the
vector type of an SLP node to the representative stmt-info.
(vect_transform_stmt): Likewise.

* gcc.target/i386/vect-pr82426.c: New testcase.
* gcc.target/i386/vect-pr97352.c: Likewise.
gcc/testsuite/gcc.target/i386/vect-pr82426.c [new file with mode: 0644]
gcc/testsuite/gcc.target/i386/vect-pr97352.c [new file with mode: 0644]
gcc/tree-vect-data-refs.c
gcc/tree-vect-slp.c
gcc/tree-vect-stmts.c
gcc/tree-vectorizer.h