diff options
| author | Richard Biener <rguenther@suse.de> | 2021-03-24 12:55:16 +0100 |
|---|---|---|
| committer | Richard Biener <rguenther@suse.de> | 2021-03-25 09:07:34 +0100 |
| commit | a29124d28253cdf603ba1977db2f09c9f233fea5 (patch) | |
| tree | 4e4f8ee0bf5a8eec84a69179283cd2410cdbc9f1 | |
| parent | 08103e4d6ada9b57366f2df2a2b745babfab914c (diff) | |
| download | gcc-a29124d28253cdf603ba1977db2f09c9f233fea5.zip gcc-a29124d28253cdf603ba1977db2f09c9f233fea5.tar.gz gcc-a29124d28253cdf603ba1977db2f09c9f233fea5.tar.bz2 | |
tree-optimization/99746 - avoid confusing hybrid code
This avoids confusing the hybrid vectorization code with SLP
patterns by not marking SLP pattern covered stmts as patterns
(they are marked as SLP patterns already). This means that loop
vectorization will vectorize the scalar stmt rather than the SLP
pattern stmt (which it can't anyway).
2021-03-24 Richard Biener <rguenther@suse.de>
PR tree-optimization/99746
* tree-vect-slp-patterns.c (complex_pattern::build): Do not mark
the scalar stmt as patterned. Instead set up required things
manually.
* gfortran.dg/vect/pr99746.f90: New testcase.
| -rw-r--r-- | gcc/testsuite/gfortran.dg/vect/pr99746.f90 | 45 | ||||
| -rw-r--r-- | gcc/tree-vect-slp-patterns.c | 10 |
2 files changed, 47 insertions, 8 deletions
diff --git a/gcc/testsuite/gfortran.dg/vect/pr99746.f90 b/gcc/testsuite/gfortran.dg/vect/pr99746.f90 new file mode 100644 index 0000000..fe947ae --- /dev/null +++ b/gcc/testsuite/gfortran.dg/vect/pr99746.f90 @@ -0,0 +1,45 @@ +! { dg-do compile } +! { dg-additional-options "-march=armv8.3-a" { target aarch64-*-* } } +SUBROUTINE CLAREF(A, WANTZ, Z, ICOL1, ITMP1, ITMP2, T1, T2, V2) +LOGICAL BLOCK, WANTZ +COMPLEX T1, T2, V2 +COMPLEX A(LDA, *), VECS, Z(LDA, *) +COMPLEX SUM +LOGICAL LSAME +IF (LSAME) THEN + DO 30 K = ITMP1, ITMP2, 3 + T1 = VECS0 +30 CONTINUE +ELSE + IF (BLOCK) THEN + DO 90 K = ITMP1, ITMP2 - 1, 3 + A(J, ICOL1) = ITMP1 + IF (WANTZ) THEN + DO 80 J = ITMP1, ITMP2 + SUM = ICOL1 + Z(J, 3) = V23 +80 CONTINUE + END IF +90 CONTINUE + DO 120 K = ITMP1, ITMP2 + V2 = VECS() + DO 100 J = ITMP1, ITMP2 + A(J, ICOL1) = A(J, ICOL1) - SUM +100 CONTINUE + IF (WANTZ) THEN + DO 110 J = 1, 3 + SUM = Z(J, ICOL1) + Z(J, ICOL1) = 0 +110 CONTINUE + END IF + ICOL1 = ICOL1 + 1 +120 CONTINUE + ELSE + DO 130 J = ITMP1, ITMP2 + SUM = T1 * A(J, ICOL1) + T2 * A(J, 1) + V2 * A(J, 2) + A(J, ICOL1) = SUM + A(J, ICOL1 + 2) = SUM * V1 +130 CONTINUE + END IF +END IF +END diff --git a/gcc/tree-vect-slp-patterns.c b/gcc/tree-vect-slp-patterns.c index 85f2d03..cdd61e9 100644 --- a/gcc/tree-vect-slp-patterns.c +++ b/gcc/tree-vect-slp-patterns.c @@ -571,14 +571,8 @@ complex_pattern::build (vec_info *vinfo) STMT_VINFO_RELEVANT (call_stmt_info) = vect_used_in_scope; STMT_SLP_TYPE (call_stmt_info) = pure_slp; - /* add_pattern_stmt can't be done in vect_mark_pattern_stmts because - the non-SLP pattern matchers already have added the statement to VINFO - by the time it is called. Some of them need to modify the returned - stmt_info. vect_mark_pattern_stmts is called by recog_pattern and it - would increase the size of each pattern with boilerplate code to make - the call there. */ - vect_mark_pattern_stmts (vinfo, stmt_info, call_stmt, - SLP_TREE_VECTYPE (node)); + gimple_set_bb (call_stmt, gimple_bb (stmt_info->stmt)); + STMT_VINFO_VECTYPE (call_stmt_info) = SLP_TREE_VECTYPE (node); STMT_VINFO_SLP_VECT_ONLY_PATTERN (call_stmt_info) = true; /* Since we are replacing all the statements in the group with the same |