https://gcc.gnu.org/bugzilla/show_bug.cgi?id=100866
--- Comment #6 from luoxhu at gcc dot gnu.org ---
For V4SI, it is also better to use vector splat and vector rotate operations.
revb:
.LFB0:
.cfi_startproc
vspltish %v1,8
vspltisw %v0,-16
vrlh %v2,%v2,%v1
vrlw %v2,%v2,%v0
blr
Performance improved from 7.322s to 2.445s with a small benchmark due to load
instruction replaced.
But for V2DI, we don't have "vspltisd" to splat {32,32} to vector register
before Power9, so lvx is still required?
vector unsigned long long revb_pwr7_l(vector unsigned long long a)
{
return vec_rl(a, vec_splats((unsigned long long)32));
}
generates:
revb_pwr7_l:
.LFB1:
.cfi_startproc
.LCF1:
0: addis 2,12,.TOC.-.LCF1@ha
addi 2,2,.TOC.-.LCF1@l
.localentry revb_pwr7_l,.-revb_pwr7_l
addis %r9,%r2,.LC0@toc@ha
addi %r9,%r9,.LC0@toc@l
lvx %v0,0,%r9
vrld %v2,%v2,%v0
blr
.LC0:
.quad 32
.quad 32
.align 4
