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

[David Li via Phabricator via llvm-commits]

davidxl added a comment.

In D158689#4629243 <https://reviews.llvm.org/D158689#4629243>, @huangjd wrote:

> In D158689#4629032 <https://reviews.llvm.org/D158689#4629032>, @davidxl wrote:
>
>> Interesting. @williamjhuang can you double check your measurement ? If the improvement can be reproduced, do some profiling and understand where the performance come from is useful.
>
> I measured the time on reading the function offset table section and reading the entire profile, not the compilation time. The other steps in the compilation pipeline are expected to take up a much longer time than reading the profile (actually only the function table is always being read, the samples are read only if they match some functions in the source code), so we need to have a common ground on what's being measured.
>
> The objective of this series of patch is to improve the profile reader (of course need to make sure the overall compilation time is improved as well).

The e2e compilation should still see reduced instructions and cycles.  Can you collect e2e data for more compilations (and also report profile size)?  While the slowest compilation determines the build latency, the goal here is to save build pool's utilization, so you don't have to pick the slowest compilations for measurement.


Repository:
  rG LLVM Github Monorepo

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

https://reviews.llvm.org/D158689