Remove -fexpose-all-unfoldings and -fspecialise-aggressively #69

totbwf commented 2026年04月30日 18:35:40 +02:00

Enabling these flags makes the build way slower and can really bloat the binary size with useless specializations. This bloat can also cause performance issues, as we end up with more icache misses.

To this end, we should isolate the functions that actually need specialization, and mark them with SPECIALIZE pragmas.

Enabling these flags makes the build *way* slower and can really bloat the binary size with useless specializations. This bloat can also cause performance issues, as we end up with more icache misses. To this end, we should isolate the functions that actually need specialization, and mark them with `SPECIALIZE` pragmas.

totbwf commented 2026年05月03日 23:52:54 +02:00

Some relevant stats

-optimise-heavily

on my machine, 8e930b5e51 with -optimise-heavily +debug takes ~13min to compile, and allocates 740gb while doing so; full report generated by https://github.com/codedownio/time-ghc-modules linked below. The resulting binary is 142m. Stats for checking the 1lab are:

1,654,370,034,040 bytes allocated in the heap
 160,555,604,768 bytes copied during GC
 2,463,677,992 bytes maximum residency (133 sample(s))
 4,817,896 bytes maximum slop
 4707 MiB total memory in use (0 MiB lost due to fragmentation)
 Tot time (elapsed) Avg pause Max pause
 Gen 0 395052 colls, 0 par 74.991s 76.062s 0.0002s 0.0034s
 Gen 1 133 colls, 0 par 4.790s 5.209s 0.0392s 0.2543s
 TASKS: 4 (1 bound, 3 peak workers (3 total), using -N1)
 SPARKS: 0 (0 converted, 0 overflowed, 0 dud, 0 GC'd, 0 fizzled)
 INIT time 0.002s ( 0.002s elapsed)
 MUT time 274.165s (273.487s elapsed)
 GC time 79.781s ( 81.270s elapsed)
 EXIT time 0.187s ( 0.010s elapsed)
 Total time 354.135s (354.769s elapsed)
 Alloc rate 6,034,214,440 bytes per MUT second
 Productivity 77.4% of total user, 77.1% of total elapsed

+optimise-heavily

Conversely, a build with +optimise-heavily +debug takes ~16 minutes to typecheck while allocating 995gb of memory, and results in a 172m binary. Stats for checking the 1lab:

 898,004,982,600 bytes allocated in the heap
 126,166,773,328 bytes copied during GC
 2,455,545,752 bytes maximum residency (118 sample(s))
 4,801,968 bytes maximum slop
 4770 MiB total memory in use (0 MiB lost due to fragmentation)
 Tot time (elapsed) Avg pause Max pause
 Gen 0 213999 colls, 0 par 58.785s 61.142s 0.0003s 0.0083s
 Gen 1 118 colls, 0 par 4.144s 5.083s 0.0431s 0.5015s
 TASKS: 4 (1 bound, 3 peak workers (3 total), using -N1)
 SPARKS: 0 (0 converted, 0 overflowed, 0 dud, 0 GC'd, 0 fizzled)
 INIT time 0.003s ( 0.003s elapsed)
 MUT time 156.325s (154.849s elapsed)
 GC time 62.929s ( 66.225s elapsed)
 EXIT time 0.379s ( 0.013s elapsed)
 Total time 219.636s (221.090s elapsed)
 Alloc rate 5,744,468,220 bytes per MUT second
 Productivity 71.2% of total user, 70.0% of total elapsed

Some relevant stats ## `-optimise-heavily` on my machine, 8e930b5e51 with -optimise-heavily +debug takes ~13min to compile, and allocates 740gb while doing so; full report generated by https://github.com/codedownio/time-ghc-modules linked below. The resulting binary is 142m. Stats for checking the 1lab are: ``` 1,654,370,034,040 bytes allocated in the heap 160,555,604,768 bytes copied during GC 2,463,677,992 bytes maximum residency (133 sample(s)) 4,817,896 bytes maximum slop 4707 MiB total memory in use (0 MiB lost due to fragmentation) Tot time (elapsed) Avg pause Max pause Gen 0 395052 colls, 0 par 74.991s 76.062s 0.0002s 0.0034s Gen 1 133 colls, 0 par 4.790s 5.209s 0.0392s 0.2543s TASKS: 4 (1 bound, 3 peak workers (3 total), using -N1) SPARKS: 0 (0 converted, 0 overflowed, 0 dud, 0 GC'd, 0 fizzled) INIT time 0.002s ( 0.002s elapsed) MUT time 274.165s (273.487s elapsed) GC time 79.781s ( 81.270s elapsed) EXIT time 0.187s ( 0.010s elapsed) Total time 354.135s (354.769s elapsed) Alloc rate 6,034,214,440 bytes per MUT second Productivity 77.4% of total user, 77.1% of total elapsed ``` ## `+optimise-heavily` Conversely, a build with +optimise-heavily +debug takes ~16 minutes to typecheck while allocating 995gb of memory, and results in a 172m binary. Stats for checking the 1lab: ``` 898,004,982,600 bytes allocated in the heap 126,166,773,328 bytes copied during GC 2,455,545,752 bytes maximum residency (118 sample(s)) 4,801,968 bytes maximum slop 4770 MiB total memory in use (0 MiB lost due to fragmentation) Tot time (elapsed) Avg pause Max pause Gen 0 213999 colls, 0 par 58.785s 61.142s 0.0003s 0.0083s Gen 1 118 colls, 0 par 4.144s 5.083s 0.0431s 0.5015s TASKS: 4 (1 bound, 3 peak workers (3 total), using -N1) SPARKS: 0 (0 converted, 0 overflowed, 0 dud, 0 GC'd, 0 fizzled) INIT time 0.003s ( 0.003s elapsed) MUT time 156.325s (154.849s elapsed) GC time 62.929s ( 66.225s elapsed) EXIT time 0.379s ( 0.013s elapsed) Total time 219.636s (221.090s elapsed) Alloc rate 5,744,468,220 bytes per MUT second Productivity 71.2% of total user, 70.0% of total elapsed ```

AndrasKovacs commented 2026年05月13日 11:28:06 +02:00

To reduce binary sizes, independently of opt flags, you could try to use -split-sections on everything (including dependencies but I don't know ATM how to configure that outside of Stack).

To reduce binary sizes, independently of opt flags, you could try to use `-split-sections` on everything (including dependencies but I don't know ATM how to configure that outside of Stack).