On Mon, 28 Oct 2024, Alex Coplan wrote: > This fixes a latent wrong code issue whereby vect_do_peeling determined > the wrong condition for inserting the vector skip guard. Specifically > in the case where the loop niters are unknown at compile time we used to > check: > > !LOOP_REQUIRES_VERSIONING (loop_vinfo) > > but LOOP_REQUIRES_VERSIONING is true for loops which we have versioned > for aliasing, and that has nothing to do with prolog peeling. I think > this condition should instead be checking specifically if we aren't > versioning for alignment. > > As it stands, when we version for alignment, we don't peel, so the > vector skip guard is indeed redundant in that case. > > With the testcase added (reduced from the Fortran frontend) we would > version for aliasing, omit the vector skip guard, and then at runtime we > would peel sufficient iterations for alignment that there wasn't a full > vector iteration left when we entered the vector body, thus overflowing > the output buffer.
OK. Thanks, Richard. > gcc/ChangeLog: > > * tree-vect-loop-manip.cc (vect_do_peeling): Adjust skip_vector > condition to only omit the edge if we're versioning for > alignment. > > gcc/testsuite/ChangeLog: > > * gcc.dg/vect/vect-early-break_130.c: New test. > --- > .../gcc.dg/vect/vect-early-break_130.c | 91 +++++++++++++++++++ > gcc/tree-vect-loop-manip.cc | 2 +- > 2 files changed, 92 insertions(+), 1 deletion(-) > create mode 100644 gcc/testsuite/gcc.dg/vect/vect-early-break_130.c > > -- Richard Biener <rguent...@suse.de> SUSE Software Solutions Germany GmbH, Frankenstrasse 146, 90461 Nuernberg, Germany; GF: Ivo Totev, Andrew McDonald, Werner Knoblich; (HRB 36809, AG Nuernberg)
