CakeML
The encoders are _big_
Some numbers from a --target=riscv --exp_cut=99999 build (9a0180e sexpr-bootstrap):
1069060 cml_x64Prog_x64_enc_5306
990976 cml_arm7Prog_arm7_enc_5296
938348 cml_riscvProg_riscv_enc_5348
784956 cml_arm8Prog_arm8_enc_5342
373572 cml_mipsProg_mips_enc_5352
291168 cml_ag32_enc_5324
Some of this is probably unavoidable with heap-allocated words but looking at the HOL input to the translator in riscvProgScript and x64ProgScript we seem to be forcing the encoders into a highly canonical form where no logic can be shared between different instructions; particularly notable are the 16 reg/imm (not)(lower, less, equal, test) cases of JumpCmp which result in 16 copies of the logic to encode a conditional jump displacement. Improving the pre-translate conversions here could make the bootstrap significantly faster with no effect whatsoever on compiling things in the logic.
That latter issue you mention (16 copies of same code) seems like it might be a result of too much manual inlining on my part when I was setting up those translations.
It should certainly be possible to extract out that logic again into a separate function if it is indeed shared code.
