Speed up OptMergePass by 1.75x. #3975
Conversation
The main speedup comes from swithing from using a SHA1 hash to std::hash<std::string>. There is no need to use an expensive cryptographic hash for fingerprinting in this context.
|
You could probably use an even cheaper hash than SHA1. |
The change here is away from SHA1. |
|
@Ravenslofty is there a setting to auto-squash my commits here? I noticed my commits didn't get squashed for the PRs you merged yesterday. |
|
There is, but I have the habit of "rebase and merge" to not lose history; if you want your commits squashed I can do that with my next merge. |
|
Btw @Ravenslofty we were discussing in the dev JF that doing merge commits is better since they have the PR number, which is otherwise lost. |
|
Well okay. |
|
OK, I'll leave that up to you. I'm just used to us squashing for Eigen. |
|
FWIW squash commits should have the PR number and are ok too, it's just the rebase that loses it. |
… and most hash function for uint64_t to hashlib.h to support this.
…atch with some compilers.
|
PTAL |
rmlarsen
changed the title
I just ran the measurements again, and it actually ended up slightly faster too. Go figure... :-) |
Ravenslofty
added
the
discuss
The main speedup comes from switching away from using SHA1 hash to using
std::hash<std::string>for fingerprinting cells. There is no need to use an expensive cryptographic hash for fingerprinting in this context.Flamegraph of benchmark timings from synthesis of a representative circuit:
Before:
After:
