My understanding is that input_iterators do not have to return a reference, and may return a value type or some other proxy object.

Updated. This was actually a missed change from the split of DenseSet here(https://reviews.llvm.org/D25628) that results in build errors on MSVC when forward declaring DenseSet. This friend decl does not forward declare the template arguments as well, which creates a mismatch.

I was actually responsible for the target dependent attrs patch, Sean was nice enough to commit for me at the time.

LGTM

Has there been any sort of discussion on expanding/using the existing IR level code sinking pass? I am referring to the SinkingPass in scalar/sink.cpp. AFAIK it's only used in the AMDGPU preisel pipeline. I don't know it's current state/usability but the description of the pass is:

We will have more discussion before moving forward.

Finish moving outlining utilities to the interface.

Going to define a custom DenseMapInfo for my use case for now.

Fix patch format.

Slight NFC.

It never landed and load combine isn't really used, so it wasn't a big priority. The problem with load combine in general, AFAIK, is the cost analysis for if it is even profitable to combine loads.

Manual computation of BFI and BPI in partial inliner.

Updated to reflect Sean's comments

Updated to reflect Sean's comments

Updated after David's comments

Fixed test description

Updated after Sean's comments

Updated after Sean and David's comments

Updated after Davids comments

Formatting

Hard exit if an exit frequency is zero.

Moving assumptions to later patch

Updated from Sean's comments

Updated to reflect r276515

Formatting

Updated formatting and fixed error in test case

Fix the formatting of the implementation and added a test case for the assumption cache.

The new pass manager doesn't use AssumptionCacheTracker, providing a replacement instead with AssumptionAnalysis.

Explicit dependency for LegacyPass

Also, the pass was already ported to the new pass manager, this just moves the code into a different structure.

Sorry, I was under the impression that the new pass managed had a functor for the ACT. I guess that was a pass that hadn't landed yet.

It was already ported to the new pass manager, but given that InlineFunction directly accesses the ACT it will crash if run via legacy.

Sorry about that.

Updated tests

Refactor CombineLoads

Fixed load sorting

Redundant checks and comments. Updated test arguments.

Updated Memory location creation

Fixed argument aliasing and updated test suite

Hey David, This is my first patch so Im not too sure about the process for committing. Accordingly, Im fairly certain that I don't have commit access.
Thanks River

Fixed the size type change mistake. Changed the iteration variable to be a separate value.

Removed the other patch fixes and added a test to the test suite