[PATCH] D91211: [mlir][Inliner] Refactor the inliner to use nested pass pipelines instead of just canonicalization

Mon Nov 30 12:24:58 PST 2020

mehdi_amini added a comment.

I don't quite get why you have to keep a separate builder callback for the default pass manager and build it on demand and cache it on every new encountered op name, why not just relying on the client adding it to the `opPipelines` `StringMap`with an empty key for instance?

================
Comment at: mlir/include/mlir/Pass/AnalysisManager.h:289

-  /// Get an analysis manager for the given child operation.
+  /// Get an analysis manager for the given operation.
   AnalysisManager nest(Operation *op);
----------------
I'd explain that the operation has to be a transitive child: it can't be an arbitrary op

================
Comment at: mlir/lib/Pass/Pass.cpp:871

 /// Get an analysis manager for the given child operation.
 AnalysisManager AnalysisManager::nest(Operation *op) {
----------------
Comment it out of sync

Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D91211/new/

https://reviews.llvm.org/D91211