[PATCH] D158689: [llvm-profdata] Use llvm::DenseMap in SampleProfileMap

[Wei Wang via Phabricator via llvm-commits]

weiwang added a comment.

We did some measurement internally on `DenseMap` vs. `std::unordered_map`.

The test was setup on a dedicated devserver as follows:

- Pick the top 100 slow compilations from a major internal build target.
- The profile (csspgo) size is 774MB including 170K profiled functions and 80M samples.
- Run every compilation using either version for 5 runs and calculate the average.
- Report both instruction count (very low variance) and elapsed wall clock time (relatively high variance).

The result

- 0.02% less instructions from the `unordered_map` version.
- 1.2% less wall clock time from the `unordered_map` version.

We didn't measure the time spent within sample profile loader though, the above numbers are all e2e compilations.


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D158689/new/

https://reviews.llvm.org/D158689