[ptweezy]

compress_icf_map_g() parses the level 3 match map greedily: from each position it takes the match there, a literal and the match after it, or two literals. Whether a position starts a match is close to random, so the branches of the C loop mispredict often. This PR moves the parse into a dispatched parse_icf_map_g() and adds parse_icf_map_g_06, which computes the advance from each position, 16 at a time, then follows the advances in a loop that writes the codes and counts the histogram with conditional moves. The loop's only branches are its exits.

Which CPUs run it

x86 CPUs with AVX-512 run parse_icf_map_g_06, at the _06 dispatch level like the other level 3 kernels. Everything else, including aarch64, riscv64 and arch=noarch, runs parse_icf_map_g_base: the C loops, with the output pointer kept in a local. Only level 3 calls the parse.

Output

The output does not change.

Results

Level 3 MB/s, master → this PR, in-process A/B on GitHub runners (harness; runs 1, 2, 3) on the Silesia files and on 256 slices of silesia.tar per size. Each cell is the median over all runs (two per runner) of the best of 11 rounds and, in parentheses, the median over runs of the change in median-round throughput. The benchmarked build has this PR's loop code with two extra setup instructions (a dead store and a redundant xor).

CPU (runners)	Whole files	64 KB	16 KB	4 KB	1 KB
AMD EPYC 9V45, Zen 5 (6)	432 → 478 (+11.9%)	400 → 446 (+11.7%)	338 → 369 (+9.6%)	230 → 240 (+4.3%)	116 → 116 (+0.1%)
AMD EPYC 9V74, Zen 4 (3)	302 → 310 (+2.4%)	289 → 296 (+2.3%)	250 → 253 (+1.3%)	182 → 181 (−0.9%)	96 → 95 (−1.4%)
Xeon Platinum 8370C, Ice Lake (3)	232 → 241 (+3.9%)	222 → 231 (+4.0%)	194 → 200 (+3.2%)	142 → 144 (+1.2%)	78 → 78 (−0.3%)
Xeon Platinum 8573C, Emerald Rapids (1)	402 → 445 (+10.9%)	381 → 419 (+9.4%)	311 → 335 (+7.9%)	206 → 211 (+2.0%)	95 → 96 (−0.3%)
Xeon 6973P-C, Granite Rapids (3)	431 → 464 (+7.5%)	413 → 444 (+7.1%)	337 → 357 (+5.4%)	219 → 223 (+1.4%)	103 → 103 (−0.3%)

The Zen 3 runners (EPYC 7763, no AVX-512) run the C version and are within 1% of master at every size.

Testing

• make check passes on x86-64, with arch=noarch, on aarch64 under qemu, and natively on Windows (mingw-w64) on a Ryzen 7 9800X3D, which dispatches parse_icf_map_g_06. The igzip tests pass on riscv64 under qemu.
• parse_icf_map_g_06 matches parse_icf_map_g_base (return value, codes, histogram) on random match maps, as SysV code and as win64 code called through a thunk, with guard pages after the output and 64 bytes past the end of the map. The vector pass reads up to 16 entries past the end, which hash_map.overflow[] makes room for.
• Level 3 output is byte-identical to master in random stateless and stateful runs (all flush types, dictionaries, gzip/zlib, small level buffers, Silesia slices), on x86 with AVX-512 and with arch=noarch.
• On native Windows, every call of parse_icf_map_g_06 in level 3 deflate preserves xmm6-xmm15, the callee-saved general purpose registers, rsp and DF.