[glennhickey]

The DP's block is sized for every row at full width, though a banded DP writes only the front of it. Linux, under its default overcommit policy, refuses an allocation larger than its memory and swap however little of it is touched, so pgrealign's banded alignments of long tandem-repeat alleles failed with posix_memalign fail on machines that could have run them: 62 kb against 70 kb reserves 128 GB and writes 5.5 GB.

A block of 64 MB or more is now mapped with MAP_NORESERVE (abpoa_matrix_alloc, abpoa_matrix_free in abpoa_simd.c), which Linux does not count against memory and swap; a smaller one, or one on a platform without MAP_NORESERVE, is allocated as before. Which one a block is follows from its size, so the struct and every other file are unchanged. A failed mapping exits as a failed posix_memalign did, with [abpoa_matrix_alloc] mmap fail! and the size.

• The rows written, and so the alignments and the memory they take, are the same: byte-identical output and peak memory on 24 pgrealign alignments of 5 to 49 kb in every mode it uses (unbanded, banded, seeded, several sequences), and on cactus's evolver primates with abPOA in BAR (MAF and halStats byte-identical, 103 s both).
• tests/run_all.sh passes.
• The 62 kb pair now runs in 4 s and 5.5 GB.
• pgrealign no longer refuses a region for abPOA's reservation (realign: reference allele as centre; outcomes that do not depend on the machine; long pairs aligned past abPOA's band benedictpaten/pgrealign#3), so the HPRC rejoins now align regions whose banded alignments reserve far more than a node's memory and swap. pgrealign: the same realignment on every machine; long repeat alleles aligned past abPOA's band cactus#2053 pins this commit; its CI passed with it.
• Built with avx2=1, as cactus builds it, it ran the 30-way HPRC rejoin (897,416 regions, up to a 1.74 Mb allele and a 69 GB region) and the 466-way chrX + chr15 rejoin on the cluster with no failure.

🤖 Generated with Claude Code