On Mon, 3 Jul 2023, Andre Vieira (lists) wrote:
> Hi,
>
> This patch makes the vectorizer treat any vector widening IFN as simple, like
> it did with the tree codes VEC_WIDEN_*.
>
> I wasn't sure whether I should make all IFN's simple and then exclude some
> (like GOMP_ ones), or include more than just the new widening IFNs. But since
> this is the only behaviour that changed with the ifn patch, I decided to only
> special case the widening IFNs for now. Let me know if you have different
> thoughts on this.
>
> Bootstrapped and regression tested on aarch64-unknow-linux-gnu.
But could we expand all VEC_WIDEN_* with scalar operands properly? Can
we do that for the IFNs? I think that's what we will end up doing then?
How do we end up with !STMT_VINFO_LIVE_P here anyway?
vectorizable_live_operation does
/* If STMT is not relevant and it is a simple assignment and its inputs
are
invariant then it can remain in place, unvectorized. The original
last
scalar value that it computes will be used. */
if (!STMT_VINFO_RELEVANT_P (stmt_info))
{
gcc_assert (is_simple_and_all_uses_invariant (stmt_info,
loop_vinfo));
if (dump_enabled_p ())
dump_printf_loc (MSG_NOTE, vect_location,
"statement is simple and uses invariant. Leaving
in "
"place.\n");
return true;
> gcc/ChangeLog:
>
> PR tree-optimization/110436
> * tree-vect-stmts.cc (is_simple_and_all_uses_invariant): Treat
> widening
> IFN's as simple.
>
> gcc/testsuite/ChangeLog:
>
> * gcc.dg/pr110436.c: New test.
>
--
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)
