[PATCH] D126236: [CaptureTracking] Increase limit but use it for all visited uses.

[Nikita Popov via Phabricator via llvm-commits]

nikic added a comment.

> Good point. Below I added tables with NumCaptured, NumNotcaptured, the sum and NotCaptured/Sum. I've only included SPEC2006/SPEC2017, as those have the largest number of capture queries.

Thanks! I've put the data in a spreadsheet for easy comparison: https://docs.google.com/spreadsheets/d/13cwt-aWjAb_a1Ri9vqS4RgCslfDkEwSqYoYpEtuoeeI/edit?usp=sharing For Limit=60 we're doing slightly better than baseline on average (1.1x) with two outlier regressions. For Limit=80 we avoid one outlier and slightly improve the average (1.2x). With Limit=100 we avoid both outliers and again slightly improve the average (1.3x).

> It might be worth exploring different limits for different clients, but I think another major client is GVN (via MemDepAnalysis) and I think exploring a sufficient number of uses there can be quite important (the regression I was investigating was due to missed load elimination by GVN)

Unfortunately MemDepAnalysis is exactly the area where cost increases are most problematic :( In part because cross-BB MDA uses completely ridiculous cutoffs -- we should probably go ahead and reduce those, and not wait for the magic bullet of NewGVN or MSSA. But that's neither here nor there.


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D126236/new/

https://reviews.llvm.org/D126236