kripken
commented
2 years ago Wow, that orders of magnitude difference (3 to 112 seconds) on duplicate-function-elimination is very shocking... at least it should be easy to debug since it's so big (and also convenient it is the first pass to run).
I would try using the linux perf command on it (perf record wasm-opt [..], stop it after that pass, then perf report). That will hopefully show something useful.
One possible guess is that that pass hashes function contents. Perhaps the hash function in the STL on that machine is very slow, somehow? We do depend on system libraries there, so system differences can matter.
Another thing to check is whether the slow build is a debug build - maybe wasm-pack builds it at -O0 by mistake? But I wouldn't expect that to cause such a big difference on a specific pass and not everything.
Finally, I'd make sure the machine has enough memory, as maybe that pass uses more than others, and it starts to swap. Can try to build with BINARYEN_CORES=1 or 2 or 4, etc., which will limit parallelism but also save memory (if there are more cores).
Relevant to #4165
MichaelMauderer
wdanilo
It seems that
wasm-opton my machine has serious differences in performance when running either on Windows 10 or Manjaro.I'm using
wasm-optthroughwasm-packon a non-trivial Rust project. On Windows, thewasm-optpass takes 133.157 seconds, on the same machine, different partition running Manjaro the same compile takes 728.146 seconds.This is the output when setting
BINARYEN_PASS_DEBUG=1for Windowsand on Manjaro
The specs of the machine used: Processor AMD Ryzen Threadripper 2950X 16-Core Processor 3.50 GHz Installed RAM 64.0 GB (63.9 GB usable)
Is there a known reason for this performance difference? My preferred work environment is Manjaro, but the time it takes to run
wasm-optthere is killing my workflow, and I do need to work with the fully optimized output at the moment.