Commit 6240654
committed
[NewPM] Don't preserve BlockFrequencyInfo in FunctionToLoopPassAdaptor
Function analyses in LoopStandardAnalysisResults are marked as preserved by the loop pass adaptor, because LoopAnalysisManagerFunctionProxy manually invalidates most of them.
However the proxy doesn't invalidate BFI, since it is only preserved on a "lossy" basis: see https://reviews.llvm.org/D86156 and https://reviews.llvm.org/D110438.
So any changes to the CFG will result in BFI giving incorrect results, which is fine for loop passes which deal with the lossiness.
But the loop pass adapator still marks it as preserved, which causes the lossy result to leak out into function passes.
This causes incorrect results when viewed from e.g. LoopVectorizer, where an innermost loop header may be reported to have a smaller frequency than its successors.
This fixes this by dropping the call to preserve, and adds a test with the -O1 pipeline which shows the effects whenever the CFG is changed and UseBlockFrequencyInfo is set.
I've also dropped it for BranchProbabilityAnalysis too, but I couldn't test for it since UseBranchProbabilityInfo always seems to be false? This may be dead code.1 parent d62b028 commit 6240654
File tree
2 files changed
+1
-4
lines changed- llvm
- lib/Transforms/Scalar
- test/Other
2 files changed
+1
-4
lines changed| Original file line number | Diff line number | Diff line change | |
|---|---|---|---|
| |||
349 | 349 | | |
350 | 350 | | |
351 | 351 | | |
352 | | - | |
353 | | - | |
354 | | - | |
355 | | - | |
356 | 352 | | |
357 | 353 | | |
358 | 354 | | |
| |||
| Original file line number | Diff line number | Diff line change | |
|---|---|---|---|
| |||
257 | 257 | | |
258 | 258 | | |
259 | 259 | | |
| 260 | + | |
260 | 261 | | |
261 | 262 | | |
262 | 263 | | |
| |||
0 commit comments