https://gcc.gnu.org/bugzilla/show_bug.cgi?id=96339
Bug ID: 96339
Summary: [SVE] Optimise svlast[ab]
Product: gcc
Version: 11.0
Status: UNCONFIRMED
Keywords: missed-optimization
Severity: enhancement
Priority: P3
Component: target
Assignee: unassigned at gcc dot gnu.org
Reporter: rsandifo at gcc dot gnu.org
Target Milestone: ---
For:
#include <arm_sve.h>
int8_t
foo (svint8_t x)
{
return svlasta (svptrue_pat_b8 (SV_VL1), x);
}
we generate:
ptrue p0.b, vl1
lasta w0, p0, z0.b
ret
But this is really a BIT_FIELD_REF with a constant index,
so we can generate:
umov w0, v0.b[0]
ret
(works even on big-endian targets).
We should make svlast_impl fold calls to a BIT_FIELD_REF if:
(1) the predicate input is a constant,
(2) we know at compile time which element the constant selects, and
(3) the selected element ends on or before byte 64
For other cases it is probably better to keep the original call.
(2) excludes variable-length predicate constants if (a) we can't
determine the index of the final 1 bit at compile time or
(b) that index might be beyond the end of the vector
(3) restricts the optimisation to cases that can be handled by
*vec_extract<mode><Vel>_v128 and *vec_extract<mode><Vel>_dup.
