Thanks, buildman -B localises it well, and there is an easy win in it. Almost all the growth is the new Huffman fast-decode path in 1.5.7. turris_mox and mvebu_espressobin-88f3720 give identical breakdowns, text +1896 rodata +24:
HUF_decompress4X1_usingDTable_internal - 1508 +1508 HUF_readDTableX1_wksp 8 1312 +1304 HUF_decompress4X1_usingDTable_internal_fast_c_loop - 1092 +1092 FSE_decompress_wksp_bmi2 4 316 +312 Upstream guards that path with HUF_DISABLE_FAST_DECODE. Adding it to the CONFIG_ZSTD_LIB_MINIFY flags in lib/zstd/Makefile, beside the HUF_FORCE_DECOMPRESS_X1 and DYNAMIC_BMI2=0 already there, takes turris_mox from +1920 to +236 bytes against 1.5.2. Sandbox ut lib still passes 133 tests with 0 failures, including both zstd cases. Trading decode speed for size is what MINIFY is for, so it looks like it belongs there rather than on by default. Want that as a separate patch on top? I did not fold it in because it changes behaviour for existing MINIFY users, not just the library version.
