https://gcc.gnu.org/bugzilla/show_bug.cgi?id=120922
Jeffrey A. Law <law at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
CC| |tamar.christina at arm dot com
--- Comment #1 from Jeffrey A. Law <law at gcc dot gnu.org> ---
Bisects to:
309dbcea2cabb31bde1a65cdfd30bb7f87b170a2 is the first bad commit
commit 309dbcea2cabb31bde1a65cdfd30bb7f87b170a2
Author: Tamar Christina <tamar.christ...@arm.com>
Date: Tue Jun 24 07:13:22 2025 +0100
middle-end: replace log_vf usages with vf to allow support for non-power of
two vf
This patch fixes a bug where the current code assumed that exact_log2
returns
NULL on failure, but it instead returns -1. So there are some cases where
the
right shift could shift out the entire value.
Secondly it also removes the requirement that VF be a power of two. With
an
uneven unroll factor we can easily end up with a non-power of two VF which
SLP
can handle. This replaces shifts with multiplication and division.
The 32-bit x86 testcase from PR64110 was always wrong, it used to match by
pure
coincidence a vmovd inside the vector loop. What it intended to match was
that
the argument to the function isn't spilled and then reloaded from the stack
for
no reason.
But on 32-bit x86 all arguments are passed on the stack anyway and so the
match
would have never worked. The patch seems to simplify the loop preheader
which
gets it to remove an intermediate zero extend which causes the match to now
properly fail.
As such I'm skipping the test on 32-bit x86.
gcc/ChangeLog:
* tree-vect-loop-manip.cc (vect_gen_vector_loop_niters,
vect_gen_vector_loop_niters_mult_vf): Remove uses of log_vf.
gcc/testsuite/ChangeLog:
* gcc.target/i386/pr64110.c: Update testcase.
gcc/testsuite/gcc.target/i386/pr64110.c | 2 +-
gcc/tree-vect-loop-manip.cc | 36
+++++++++++++++++++-----------------
