https://gcc.gnu.org/bugzilla/show_bug.cgi?id=103568
Bug ID: 103568
Summary: sub-optimal vector construction with two loaded
doubles on Power10
Product: gcc
Version: 12.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: target
Assignee: unassigned at gcc dot gnu.org
Reporter: linkw at gcc dot gnu.org
Target Milestone: ---
For the test case:
vector double test(double *a, double *b) {
return (vector double) { *a, *b };
}
On Power10, we generate the code like:
ld 10,0(3)
ld 9,0(4)
mtvsrdd 34,9,10
As Power10 latency table, we can get better code with xxlor like:
lxsd 0, 0(4)
lxvrdx 1, 0, 3
xxlor 34, 1, 32
As to the prerequisites "if we can assume the doubleword 1 of a vsx register
after an lfd is zero", as Segher pointed out "ISA 3.1 section 7.1.1.1 says this
already".
SPEC2017 510.parest_r may be one benchmark to evaluate the effect (with
vectorization turned on).