https://gcc.gnu.org/bugzilla/show_bug.cgi?id=88839
Bug ID: 88839
Summary: [SVE] Poor implementation of blend-like permutes
Product: gcc
Version: unknown
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: target
Assignee: unassigned at gcc dot gnu.org
Reporter: rsandifo at gcc dot gnu.org
Target Milestone: ---
Compiling this code with -O3 -msve-vector-bits=256:
typedef int v8si __attribute__((vector_size(32)));
v8si
f (v8si x, v8si y, v8si sel)
{
return __builtin_shuffle (x, y, (v8si) { 0, 9, 2, 11, 4, 13, 6, 15 });
}
produces an inefficient TBL-based sequence.
In these blend-like cases, where index I of the output comes from index I of
one of the inputs, we should be able to use a SEL with an appropriate predicate
constant. The preferred implementation of the above would be:
ptrue p0.d, vl4 // { 1, 0, 1, 0, ... } when used as p0.s
sel res, p0, y, x
This will also be useful for the default VL-agnostic mode when implementing
support for 2-operation SLP.
