https://gcc.gnu.org/bugzilla/show_bug.cgi?id=96461
Bug ID: 96461
Summary: [SVE] Use the HISTCNT instruction for simple histogram
loops
Product: gcc
Version: 11.0
Status: UNCONFIRMED
Severity: enhancement
Priority: P3
Component: tree-optimization
Assignee: unassigned at gcc dot gnu.org
Reporter: rsandifo at gcc dot gnu.org
Blocks: 53947
Target Milestone: ---
Target: aarch64*-*-*
SVE2 has a HISTCNT instruction that, for each element, counts
the number of matching elements in that lane and previous lanes.
It allows things like:
void
update (uint32_t *histogram, uint32_t *records, uint32_t num_records)
{
for (uint32_t i = 0; i < num_records; i++)
histogram[records[i]] += 1;
}
to be vectorised using a gather load, HISTCNT and scatter store.
The instruction is intended for loops that do more than the bare
minimum above, but as far as autovec goes, we need to start somewhere.
This PR is therefore about recognising the:
histogram[…] += 1;
gather/HISTCNT/scatter pattern and generating something like the
asm below for the function above:
mov x3, #0
whilelo p0.s, xzr, x2
.loop
ld1w z1.s, p0/z, [x1, x3, lsl #2]
ld1w z2.s, p0/z, [x0, z1.s, uxtw #2]
histcnt z0.s, p0/z, z1.s, z1.s
add z2.s, p0/m, z2.s, z0.s
st1w z2.s, p0, [x0, z1.s, uxtw #2]
incw x3
whilelo p0.s, x3, x2
b.first .loop
ret
Once that's done we could move on to fancier cases.
Referenced Bugs:
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=53947
[Bug 53947] [meta-bug] vectorizer missed-optimizations
