[PATCH] D83177: [llvm-reduce] Reducing call operand bundles

[Tyker via Phabricator via llvm-commits]

Tyker added inline comments.


================
Comment at: llvm/tools/llvm-reduce/deltas/ReduceOperandBundles.cpp:138
+
+  for_each(R.CallsToRefine, [](const auto &P) {
+    return maybeRewriteCallWithDifferentBundles(P.first, P.second);
----------------
lebedev.ri wrote:
> Tyker wrote:
> > Maybe we should make CallsToRefine a MapVector since the association from a index to a Bundle depends on its order in the map.
> > and the key depends on a pointer value that will change when the Module gets cloned.
> No, the current logic is correct, that map doesn't live that long.
here is the situation i think fails.

extractOperandBundesFromModule is called a first time and generate a reduction.
the reduction isn't considered interesting.
extractOperandBundesFromModule is called a second time with different chunks to keep.
but the association from an index to a Bundle is different from the first time because the module isn't the same
the association from a index to a Bundle isn't the same,
so extractOperandBundesFromModule can remove the same operand bundle a second time and not have tried every operand bundles at the end of the passe.




Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D83177/new/

https://reviews.llvm.org/D83177