perf(decode): 8-slot software prefetch pipeline for sequence execution (#208)#227

Merged

polaz merged 23 commits into

mainfrom

perf/#208-match-prefetch

May 22, 2026

Commits on May 22, 2026

perf(decode): 4-slot software prefetch pipeline for sequence execution
polaz
committed
docs(decode): clarify prefetch L2 deviation + comment direction
polaz
committed
perf(decode): gate prefetch on offset >= 32 KiB + drop dead prefill guard
polaz
committed
Merge branch 'main' into perf/#208-match-prefetch
polaz
authored
fix(decode): use wrapping_add for prefetch position math to stay fuzz-stable
polaz
committed
perf(decode): exact actual_offset via shadow_hist + short-offset fast path
polaz
committed
refactor(decode): align prefetch pipeline comments + drop redundant tail-slice guard
polaz
committed
perf(decode): gate prefetch pipeline by FSE-table long-offset share (donor parity)
polaz
committed
fix(decode): prefetch across s1/s2 boundary when match source spans the ring wrap
polaz
committed
perf(decode): bump prefetch ADVANCE 4→8 + switch lookahead hint T1→T0 (donor parity)
polaz
committed
perf(decode): short-circuit FSE offsets walk by num_sequences first
polaz
committed
refactor(decode): pointer-width-aware long-offset gate + accurate format-spec note
polaz
committed
docs(decode): align prefetch pipeline comments with ADVANCE=8 + T0/L1 hint
polaz
committed
fix(decode): prefetch the match-start cache line even when s1/s2 tail < 64B
polaz
committed
docs(decode): drop redundant 'STORED_SEQS = 8' echo in ADVANCE power-of-two guard comment
polaz
committed
perf(decode): cache offsets-long-share in FSEScratch, recompute only on table rebuild
polaz
committed
docs+test(decode): align review comments + add prefetch_lookahead_match_source coverage
polaz
committed
perf(decode): pipeline-aware repeat variant + pre-resolved offset in ring (A+B+C)
polaz
committed
fix(decode): keep buffer.reserve in pipelined repeat + scratch reset of offsets_long_share
polaz
committed
Merge branch 'main' into perf/#208-match-prefetch
polaz
authored
docs+test(decode): align pipeline comments with actual ring/reserve shape + offsets_long_share tests
polaz
committed
fix(decode): recompute offsets_long_share in FSEScratch::reinit_from
polaz
committed
fix(decode): rollback offset_hist + buffer on mid-loop pipelined Err
polaz
committed