perf(present): zero-dispatch RGBA pack + zero-alloc fallback (measured 4-4.6x)

[tannevaled]

Performance audit of the per-frame present/pack path

Audit horodate: 2026-08-09 12:25 CEST. Measure-first: benchmarked before changing anything, fixed only proven costs, behaviour byte-identical.

Findings

Both live present paths were already zero-alloc per frame — X11 MIT-SHM packs directly into the mmap'd shared segment, Wayland directly into the wl_shm pool. The double-buffer swap allocates nothing. So the only remaining cost was CPU in the RGBA->wire conversion:

1. Wayland PackARGB8888 wrote each pixel through the binary.ByteOrder interface (NativeOrder.PutUint32) — ~2M indirect calls/frame at 1080p. Now uses the concrete binary.NativeEndian (inlines to a word store) + per-row reslicing for BCE.
2. X11 EncodeRectInto/encodeRegion packed per pixel via pixel()+putValue() method calls with per-byte inner loops. New shared packRect picks one of three byte-identical strategies once at NewPresenter time: zero-arithmetic fast path for the ubiquitous 32-bpp 8:8:8 TrueColor visual, word-wise generic-shift path for other 32-bpp visuals, per-channel path for non-32-bpp depths. Writes keyed by server image byte order (host-independent; s390x-safe).
3. Non-SHM PutImage fallback allocated ~8.3 MB/frame; now reuses a persistent per-Presenter scratch -> 0 allocs/present.

Measured (Apple M4 Max, 1920x1080, -benchmem), before -> after

bench	before	after	gain	allocs
x11 EncodeRectInto full	7.50 ms	1.64 ms	4.6x	0
x11 EncodeRectInto 64x64	14.9 us	3.43 us	4.3x	0
x11 encodeRegion full (fallback)	7.96 ms + 8.3 MB	1.66 ms	4.8x	8.3 MB/present -> 0
wl PackARGB8888 full	4.35 ms	1.09 ms	4.0x	0
wl PackARGB8888 64x64	8.43 us	2.22 us	3.8x	0

Committed Benchmark* (full-frame + damage-rect) added for both backends.

Conformance

• internal/x11 + internal/wayland stay at 100% statement coverage.
• -race (CGO=1) clean; whole module tests pass.
• CGO=0 cross-build clean on all six supported arches (amd64/arm64/riscv64/loong64/ppc64le/s390x) + darwin stub.
• No syscall usage touched — stays behind existing linux gating; the v0.3.1 syscall-leak fix is untouched.

🤖 Generated with Claude Code