unrolling for single-element search

This commit is contained in:
dzaima 2023-04-06 15:10:02 +03:00
parent 876914d053
commit 5492eae8c5

View File

@ -10,13 +10,35 @@ if (hasarch{'AVX2'}) {
}
include './mask'
def findFirst{C, M, F, ...v1} = {
def exit = makelabel{}
def args = each{{c:T}=>{ a:=undefined{T} }, M{...each{{c}=>tupsel{0,c}, v1}}}
def am = tuplen{tupsel{0,v1}}
each{{last, ...v2} => {
if (last or C{...v2}) {
each{=, args, M{...v2}}
goto{exit}
}
}, iota{am} == am-1, ...v1}
unreachable{}
setlabel{exit}
F{...args}
}
fn search{A, E}(l:*void, e0:A, n:u64) : u64 = {
def e = if (A==E) e0 else cast_i{E, e0}
def bulk = arch_defvw/width{E}
def VT = [bulk]E
maskedLoop{bulk, n, {i, M} => {
eq:= loadBatch{*E~~l, i, VT} == VT**e
if (homAny{M{eq}}) return{i*bulk + promote{u64, ctz{homMask{eq}}}}
muLoop{bulk, tern{arch_defvw>=256, 1, 2}, n, {is, M} => {
eq:= eachx{==, loadBatch{*E~~l, is, VT}, VT**e}
if (homAny{M{tree_fold{|, eq}}}) {
findFirst{
{i,c} => homAny{c},
{i,c} => tup{i,c},
{i,c} => return{i*bulk + promote{u64, ctz{homMask{c}}}},
is, eq
}
}
}}
n
}