[PATCH] D109632: [clang] de-duplicate methods from AST files

Wed Oct 20 15:16:35 PDT 2021

vsapsai added a comment.

Sorry for the delay, I haven't finished testing more projects yet but here are my preliminary numbers

| **File**        | **Baseline (s)** | **Set Dedupe (s)** | **No external (s)** | **Set Dedupe (percentage of baseline)** | **No external (percentage of baseline)** |
| Project1. File1 | 1.26146575              | 0.931755875                | 0.943674125                 | 73.863%                                 | 74.808%                                  |
| Project1. File2 | 1.444321875              | 1.0868395                | 1.095805875                 | 75.249%                                 | 75.870%                                  |
| Project1. File3 | 1.2475025              | 0.94032725                | 0.942943875                 | 75.377%                                 | 75.587%                                  |
| Project1. File4 | 1.197715         | 0.98331625                | 0.9898685                 | 82.099%                                 | 82.646%                                  |
| Project1. File5 | 1.252481375              | 0.94339925                | 0.945766            | 75.322%                                 | 75.511%                                  |
|

Methodology: clear a modules cache, compile a file once to pre-populate the cache, compile file 8 times and measure elapsed time, take the time average. Didn't include files using a precompiled header because that makes cleaning a module cache trickier. Standard deviation of the measurements is consistent between different approaches and doesn't go above 0.015, so I haven't included it here. "Project1" is an app, so it should have a fairly long chain of dependencies, though I haven't checked it in details.

So it looks like "no external" approach is slightly but consistently slower 😢 than "set dedupe" approach. I'm curious to get the results for an empty module cache because clean builds are also important for us. And if you see anything suspicious in the methodology, please let me know.

Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D109632/new/

https://reviews.llvm.org/D109632