Split assignment rationalization into a separate phase

[SingleAccretion]

And start adding the support required for store nodes to replace assignments in HIR.

This change is the proper beginning to the ASG removal project. The plan as of currently is to slowly move the newly introduced phase up the phase list, adding support to shared code (aka morph) where necessary and rewriting the rest as needed. In other words, there will be no intermediate states where both ASG and store nodes can appear in IR at once - the work will be split across phases, but not node types.

Initially, I was concerned about the TP cost of introducing a new phase, and considered more complex schemes to achieve incrementality, however, this ended up not being a problem - with some judicious use of PIN, we should be about even or slightly up on the TP metrics even with this change. Needless to say, once the end goal here has been achieved, we expect significant cumulative gains as the IR will shrink and special-casing from a number of places removed.

[ghost]

Tagging subscribers to this area: @JulieLeeMSFT, @jakobbotsch
See info in area-owners.md if you want to be subscribed.

Issue Details

---

And start adding the support required for store nodes to replace assignments in HIR.

This change is the proper beginning to the ASG removal project. The plan as of currently is to slowly move the newly introduced phase up the phase list, adding support to shared code (aka morph) where necessary and rewriting the rest as needed. In other words, there will be no intermediate states where both ASG and store nodes can appear in IR at once - the work will be split across the phases, but not node types.

Initially, I was concerned about the TP cost of introducing a new phase, and considered more complex schemes to achieve instrumentality, however, this ended up not being a problem - with some judicious use of PIN, we should be about even or slightly up on the TP metrics even with this change. Needless to say, once the end goal here has been achieved, we expect significant cumulative gains as the IR will shrink and special-casing from a number of places removed.

Author:	SingleAccretion
Assignees:	-
Labels:	area-CodeGen-coreclr, community-contribution
Milestone:	-

[SingleAccretion]

Diffs.

The TP situation (on 64 bit, 32 bit sees 0.2%-0.4% TP improvements across the board) has been improved, but some MinOpts regressions still remain. The larger ones look to be in collections with a lack of MinOpts MCs, and where such MCs are well-represented, the average is +0.05%. We are adding a new (temporary, by nature) phase here after all, so I am inclined to not spend more energy on mitigating this unless there are objections. As a point of reference, follow-up changes to remove ASGs from the optimization phases have shown > 1% TP improvements (in optimized code, naturally).

@dotnet/jit-contrib

[jakobbotsch]

/azp run runtime-coreclr jitstress, runtime-coreclr libraries-jitstress

[azure-pipelines]

Azure Pipelines successfully started running 2 pipeline(s).

[SingleAccretion]

/azp run runtime-coreclr jitstress, runtime-coreclr libraries-jitstress

Ooops, did not notice this... I've removed the stress bit, since it looks like moving rationalization earlier (thus, testing morph) will be easier than originally anticipated.

[jakobbotsch]

Ooops, did not notice this... I've removed the stress bit, since it looks like moving rationalization earlier (thus, testing morph) will be easier than originally anticipated.

No worries, I'll rerun it just to be safe.

[jakobbotsch]

/azp run runtime-coreclr jitstress, runtime-coreclr libraries-jitstress

[azure-pipelines]

Azure Pipelines successfully started running 2 pipeline(s).

[jakobbotsch]

LGTM. Agree that it's fine to take the TP regressions for now.

[SingleAccretion]

One unique failure in libraries stress, though hard to see it being related:

===========================================================================================================
/root/helix/work/workitem/e /root/helix/work/workitem/e
  Discovering: System.Net.Quic.Functional.Tests (method display = ClassAndMethod, method display options = None)
MsQuic supported and using 'libmsquic.so 2.1.8.348060 (c6931cb1eae6d885a8ec04c2e5558129b616730d)'.
  Discovered:  System.Net.Quic.Functional.Tests (found 115 of 124 test cases)
  Starting:    System.Net.Quic.Functional.Tests (parallel test collections = on, max threads = 2)
    System.Net.Quic.Tests.MsQuicPlatformDetectionTests.UnsupportedPlatforms_ThrowsPlatformNotSupportedException [SKIP]
      Condition(s) not met: "IsQuicUnsupported"
    System.Net.Quic.Tests.QuicListenerTests.AcceptConnectionAsync_ListenerDisposed_Throws [FAIL]
      Assert.Throws() Failure
      Expected: typeof(System.Security.Authentication.AuthenticationException)
      Actual:   typeof(System.Net.Quic.QuicException): The connection timed out from inactivity.
      ---- System.Net.Quic.QuicException : The connection timed out from inactivity.

[am11]

One unique failure in libraries stress, though hard to see it being related:

Stack trace in logs resembles the one reported in #85371.