[PATCH] D38190: Partial Inlining with multi-region outlining based on PGO information

[Graham Yiu via Phabricator via llvm-commits]

gyiu marked 6 inline comments as done.
gyiu added inline comments.


================
Comment at: lib/Transforms/IPO/PartialInlining.cpp:87
+    "cold-branch-ratio", cl::init(0.1), cl::Hidden,
+    cl::desc("Minimum BranchProbability to consider a region cold."));
 
----------------
davidxl wrote:
> Why using BP to determine region coldness instead of PSI and profile count?
No particular reason, mainly for convenience.  I figure I would still need to compare the profile count of a block to its sibling blocks to know its relative coldness.  Is PSI and profile count more accurate?


================
Comment at: lib/Transforms/IPO/PartialInlining.cpp:1060
+
+    Function *OutlinedFunc = CodeExtractor(RegionInfo.Region, &DT, /*AggregateArgs*/ false,
+                                 ClonedFuncBFI.get(), &BPI)
----------------
davidxl wrote:
> Is DT updated by the extractor?
I don't believe so.  You're thinking we should update the DT each time we extract a code region?


Repository:
  rL LLVM

https://reviews.llvm.org/D38190