On Fri, 10 Feb 2023, Richard Sandiford wrote:
> Richard Biener via Gcc-patches <gcc-patches@gcc.gnu.org> writes:
> > This fixes an oversight to when removing the hard limits on using
> > generic vectors for the vectorizer to enable both SLP and BB
> > vectorization to use those. The vectorizer relies on vector lowering
> > to expand plus, minus and negate to bit operations but vector
> > lowering has a hard limit on the minimum number of elements per
> > work item. Vectorizer costs for the testcase at hand work out
> > to vectorize a loop with just two work items per vector and that
> > causes element wise expansion and spilling.
> >
> > The fix for now is to re-instantiate the hard limit, matching what
> > vector lowering does. For the future the way to go is to emit the
> > lowered sequence directly from the vectorizer instead.
> >
> > Bootstrapped and tested on x86_64-unknown-linux-gnu, OK?
>
> LGTM after reading the vector lowering stuff in the PR trail.
>
> TBH I don't remember when the hard limit was removed though.
I removed it as part of fixing PR101801 but didn't anticipate the
close connection with vector lowering.
Richard.
> Thanks,
> Richard
>
> >
> > Thanks,
> > Richard.
> >
> > PR tree-optimization/108724
> > * tree-vect-stmts.cc (vectorizable_operation): Avoid
> > using word_mode vectors when vector lowering will
> > decompose them to elementwise operations.
> >
> > * gcc.target/i386/pr108724.c: New testcase.
> > ---
> > gcc/testsuite/gcc.target/i386/pr108724.c | 15 +++++++++++++++
> > gcc/tree-vect-stmts.cc | 14 ++++++++++++++
> > 2 files changed, 29 insertions(+)
> > create mode 100644 gcc/testsuite/gcc.target/i386/pr108724.c
> >
> > diff --git a/gcc/testsuite/gcc.target/i386/pr108724.c
> > b/gcc/testsuite/gcc.target/i386/pr108724.c
> > new file mode 100644
> > index 00000000000..c4e0e918610
> > --- /dev/null
> > +++ b/gcc/testsuite/gcc.target/i386/pr108724.c
> > @@ -0,0 +1,15 @@
> > +/* { dg-do compile } */
> > +/* { dg-options "-O3 -mno-sse" } */
> > +
> > +int a[16], b[16], c[16];
> > +void foo()
> > +{
> > + for (int i = 0; i < 16; i++) {
> > + a[i] = b[i] + c[i];
> > + }
> > +}
> > +
> > +/* When this is vectorized this shouldn't be expanded piecewise again
> > + which will result in spilling for the upper half access. */
> > +
> > +/* { dg-final { scan-assembler-not "\\\[er\\\]sp" } } */
> > diff --git a/gcc/tree-vect-stmts.cc b/gcc/tree-vect-stmts.cc
> > index c86249adcc3..09b5af603d2 100644
> > --- a/gcc/tree-vect-stmts.cc
> > +++ b/gcc/tree-vect-stmts.cc
> > @@ -6315,6 +6315,20 @@ vectorizable_operation (vec_info *vinfo,
> > return false;
> > }
> >
> > + /* ??? We should instead expand the operations here, instead of
> > + relying on vector lowering which has this hard cap on the number
> > + of vector elements below it performs elementwise operations. */
> > + if (using_emulated_vectors_p
> > + && (code == PLUS_EXPR || code == MINUS_EXPR || code == NEGATE_EXPR)
> > + && ((BITS_PER_WORD / vector_element_bits (vectype)) < 4
> > + || maybe_lt (nunits_out, 4U)))
> > + {
> > + if (dump_enabled_p ())
> > + dump_printf (MSG_NOTE, "not using word mode for +- and less than "
> > + "four vector elements\n");
> > + return false;
> > + }
> > +
> > int reduc_idx = STMT_VINFO_REDUC_IDX (stmt_info);
> > vec_loop_masks *masks = (loop_vinfo ? &LOOP_VINFO_MASKS (loop_vinfo) :
> > NULL);
> > internal_fn cond_fn = get_conditional_internal_fn (code);
>
--
Richard Biener <rguent...@suse.de>
SUSE Software Solutions Germany GmbH, Frankenstrasse 146, 90461 Nuernberg,
Germany; GF: Ivo Totev, Andrew Myers, Andrew McDonald, Boudien Moerman;
HRB 36809 (AG Nuernberg)
