JIT ARM64-SVE: Move SVE-specific emitIns logic to emitarm64sve.cpp

[amanasifkhalid]

We already had a few SVE-specific emitIns_* methods; this PR strips out the remaining SVE instructions from our general emitIns methods, and moves them to corresponding emitInsSve methods in emitarm64sve.cpp. I plan to move over our SVE-specific helper methods, and extract the SVE cases in other shared emitter methods to SVE methods, but this PR is big enough that this seemed like a natural stopping point. There are no diffs in the SVE unit tests from this change.

cc @dotnet/arm64-contrib

[dotnet-policy-service]

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

[a74nh]

It's fairly tricky for me to confirm the code in emitarm64sve.cpp is just a copy/paste of the code from emitarm64.cpp. However, the changes good.

I'm assuming:

• No difference in codegen test output
• The contents of the functions in emitarm64sve.cpp are identical to that cut from emitarm64.cpp, plus potentially some repeated parts/boilerplate.

Additional thoughts:

• We should merge this asap
• In another PR, we could move the other SVE parts from the sanity check, disp instr, perf score etc functions into here too.
• We have the choice to eventually use emitInsSve_ functions directly instead of calling from emitIns_. No need to make that choice yet - depends on the IR code.

[amanasifkhalid]

@a74nh thanks for the review! Your assumptions are correct. This is a no-diff change, and I did a diff of the SVE unit test output with and without this change, and didn't see any diffs there either. I plan to do similar transition work for other methods (emitOutputInstr is my next priority, since that should yield TP improvements too).

We have the choice to eventually use emitInsSve_ functions directly instead of calling from emitIns_. No need to make that choice yet - depends on the IR code.

Agreed. It would be slightly more performant if we could call the emitInsSve methods directly instead of falling into the default case in emitIns, but I didn't want to change the API surface just yet.

cc @kunalspathak @TIHan

[kunalspathak]

We have the choice to eventually use emitInsSve_ functions directly instead of calling from emitIns_

Agree.

I am hoping to see 0 or very few references of "sve_" text in emitarm64.cpp. Currently, there are still methods like insEncodeSveElemsize_tszh_tszl_and_imm , insGetSveReg1ListSize, etc. in emitarm64.cpp. Is there a reason why they can't be moved to emitarm64sve.cpp?

[amanasifkhalid]

I am hoping to see 0 or very few references of "sve_" text in emitarm64.cpp. Currently, there are still methods like insEncodeSveElemsize_tszh_tszl_and_imm , insGetSveReg1ListSize, etc. in emitarm64.cpp. Is there a reason why they can't be moved to emitarm64sve.cpp?

I plan to move those too. I just thought that would be better suited for a follow-up PR, since this one is already pretty big.

[kunalspathak]

LGTM

[kunalspathak]

[amanasifkhalid]

For reference, the MinOpts regression is only in the smoke_tests.nativeaot collection (diffs). I imagine it's some outlier method...