Ensure that we lookup the correct instruction for embedded masking/broadcast scenarios

[tannergooding]

This resolves #117794

When doing the embedded masking/broadcast checks, we sometimes allow for changing the base type under the presumption we'd pick a different instruction. Even if we didn't get a new instruction, we'd still end up changing the base type which caused a codegen bug.

To resolve this, there is now an assert to validate we got a different instruction for the alternative base type. This is achieved by ensuring the existing lookup support returns the instruction optimistically.

We then peephole this back to the original instruction in codegen if we don't end up using any of the embedded features that would require it, since that allows us to have the smaller emitter output.

[Copilot]

Pull Request Overview

This PR resolves a codegen bug (#117794) related to embedded masking/broadcast scenarios in hardware intrinsics. The issue occurred when the embedded masking/broadcast checks would change the base type to pick a different instruction, but sometimes the same instruction would be returned, leading to incorrect base type usage in code generation.

• Introduces a new lookupIns method in CodeGen that handles instruction selection with embedded features and peephole optimization
• Adds assertions to validate that instruction changes occur when expected for embedded features
• Refactors instruction lookup calls throughout the codebase to use the new method

Reviewed Changes

Copilot reviewed 9 out of 9 changed files in this pull request and generated no comments.

Show a summary per file

File	Description
Runtime_117794.cs	Regression test case for the bug fix
Runtime_117794.csproj	Project file for the regression test
hwintrinsiccodegenxarch.cpp	Main implementation of new lookupIns method and refactored codegen
hwintrinsic.cpp	Updated instruction lookup to support optimistic EVEX instruction selection
hwintrinsic.h	Modified lookupIns signature to remove compiler parameter
instr.cpp	Removed embedded broadcast instruction mapping logic
gentree.cpp	Updated calls to use new instruction lookup method
codegenxarch.cpp	Updated instruction lookup calls
codegen.h	Updated method signatures

Comments suppressed due to low confidence (2)

src/coreclr/jit/hwintrinsiccodegenxarch.cpp:433

• The instruction name INS_movdqa32 is inconsistent with the EVEX naming pattern. It should be INS_vmovdqa32 to match the pattern used for other EVEX instructions in this switch statement.

                ins = INS_movdqa32;

src/coreclr/jit/hwintrinsiccodegenxarch.cpp:439

• The instruction name INS_movdqu32 is inconsistent with the EVEX naming pattern. It should be INS_vmovdqu32 to match the pattern used for other EVEX instructions in this switch statement.

                ins = INS_movdqu32;

[dotnet-policy-service]

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

[tannergooding]

CC. @dotnet/jit-contrib, @jakobbotsch

[tannergooding]

(Trying to minimize the throughput impact, the overall code isn't really changing)

[tannergooding]

Happy with the changes now. TP impact is minimized (+0.01% in the worst case), no regressions for any disasm output (only improvements), and the tests are passing as expected.

[tannergooding]

/azp run Fuzzlyn

[azure-pipelines]

Azure Pipelines successfully started running 1 pipeline(s).

[saucecontrol]

Nice change! I'll take another look at optimizing for broadcast size after this lands.

[tannergooding]

/ba-g unrelated arm64 timeouts