Share one subcode inflater across the CD codecs

[rtissera]

cdzl, cdlz and cdfl each carry their own zlib_codec_data for the subcode stream, and each of those allocates an 8376-byte tinfl_decompressor. A CD image compressed with chdman's default trio therefore holds four inflaters — cdzl's sector-data one plus three identical subcode ones — even though a hunk is decoded by exactly one CD codec and only one subcode inflater is ever in flight.

The three now share a single instance owned by the chd_file. Nothing carries across, because zlib_codec_decompress() runs tinfl_init() on entry to every call. A borrowed flag keeps ownership explicit so only the chd_file frees it. cdzs is untouched — its subcode goes through zstd.

Seeding happens before each codec's own init, so nothing is allocated and then discarded, and it is insensitive to ordering — which matters because LOWRAM_TARGET readies codecs lazily, in whatever order the hunks ask for them. The system-zlib path is excluded: there zlib_codec_data holds a z_stream by value, with no pointer to share.

Effect

Peak heap under massif, same binary either side:

seed	before	after
cd_default (cdlz+cdzl+cdfl)	183532	166836	-16696, -9.1%
every single-codec seed			+56

The +56 is the shared zlib_codec_data now living in chd_file even when one CD codec would have sufficed. Worth noting that the RAM-budget firmware uses one codec per file, so it cannot see this saving by construction — only the +32 it costs there.

On the CPU cost

Sharing base and subcode inside cdzl was tried before and reverted at 2.6% CPU — see the note in codec_cdzs.h. That does not transfer: those two streams alternate within a hunk, these three never do. Measured with cachegrind over a real disc: +1089 instructions out of 5.1 billion, and 7689 fewer D1 misses — three codecs touching one 8KB object have better locality than three separate ones. Wall-clock is unchanged within the ±4% noise floor of this machine.

Verification

Decoded output is byte-identical across the seed corpus, the CHDv3/v4 fixtures, real CHDv4 images, six real discs and an AVHuff image, in the default build, with CHDR_CD_SCRATCH_BUFFER=OFF, CHDR_LOWRAM_TARGET=ON and WITH_SYSTEM_ZLIB=ON — 144 comparisons, no difference, and no divergence in error codes.

Under ASan/UBSan over 820 corrupt and fuzzed images: no memory errors, and the leak count is identical either side (57 of 220, all pre-existing in chd_open's failure path — worth its own issue, not this one).

rv32imafc under qemu with LOWRAM_TARGET=ON: all nine codecs decode, every threshold held.

🤖 Generated with Claude Code