search

noir-lang / noir

Noir is a domain specific language for zero knowledge proofs
https://noir-lang.org
Apache License 2.0
912 stars 206 forks source link

Reduce memory usage of tracking instruction callstacks in `dfg.locations` #6603

Open TomAFrench opened 1 week ago

TomAFrench commented 1 week ago

We currently track the callstacks for each instruction in the locations field of the DataflowGraph:

https://github.com/noir-lang/noir/blob/3c361c9f78a5d9de1b1bcb5a839d3bc481f89898/compiler/noirc_evaluator/src/ssa/ir/dfg.rs#L94

Heaptrack is reporting that inserting entries into this hashmap is using 17% of the compiler's peak memory.

Image

We should definitely not require 157MB of memory in order to track this as the number of unique callstacks is going to be relatively low. It should be relatively easy to rework this so that we don't track N different instances of the same callstack.

jfecher commented 1 week ago

Hmm, are we sure it's coming from N different instances of the same callstack? I thought a im::Vector should be shared behind the scenes. I suppose it's possible these call stacks are all so small that the small vector optimization to copy them instead of sharing them is hurting us rather than helping us.

TomAFrench commented 1 week ago

I'm speculating on the last bit but this memory is definitely associated with the callstacks for the instructions being inserted and I feel that something must be going awry here.