Fix 97272

[EgorBo]

Fixes #97272

[EgorBo]

@MihuBot

[EgorBo]

/azp run runtime-nativeaot-outerloop

[azure-pipelines]

Azure Pipelines successfully started running 1 pipeline(s).

[Sergio0694]

[Sergio0694]

Me checking this PR before going to bed:

[EgorBo]

/azp list

[azure-pipelines]

CI/CD Pipelines for this repository:

• runtime-coreclr outerloop
• runtime-coreclr jitstress
• runtime-coreclr jitstressregs
• runtime-coreclr jitstress2-jitstressregs
• runtime-coreclr gcstress0x3-gcstress0xc
• runtime-coreclr gcstress-extra
• runtime-coreclr r2r-extra
• runtime-coreclr jitstress-isas-x86
• runtime-coreclr jitstress-isas-arm
• runtime-coreclr jitstressregs-x86
• runtime-coreclr libraries-jitstressregs
• runtime-coreclr libraries-jitstress2-jitstressregs
• runtime-coreclr r2r
• runtime-coreclr runincontext
• runtime-coreclr crossgen2
• runtime-libraries-coreclr outerloop
• runtime-libraries-coreclr outerloop-windows
• runtime-libraries-coreclr outerloop-linux
• runtime-libraries-coreclr outerloop-osx
• runtime
• runtime-libraries enterprise-linux
• runtime-libraries stress-http
• runtime-libraries stress-ssl
• runtime-dev-innerloop
• runtime-coreclr crossgen2 outerloop
• coreclr-release-outerloop-nightly
• runtime-coreclr crossgen2-composite
• runtime-jit-experimental
• runtime-coreclr libraries-jitstress
• dotnet-linker-tests
• runtime-coreclr ilasm
• runtime-coreclr crossgen2-composite gcstress
• runtime-staging
• runtime-coreclr pgo
• runtime-coreclr libraries-pgo
• Antigen
• runtime-community
• Fuzzlyn
• runtime-coreclr superpmi-replay
• runtime-wasm
• runtime-coreclr superpmi-diffs
• runtime-coreclr superpmi-asmdiffs-checked-release
• runtime-extra-platforms
• jit-cfg
• runtime-wasm-perf
• runtime-llvm
• runtime-coreclr jitstress-random
• runtime-coreclr libraries-jitstress-random
• runtime-android-grpc-client-tests
• runtime-wasm-libtests
• runtime-wasm-non-libtests
• runtime-android
• runtime-androidemulator
• runtime-ioslike
• runtime-ioslikesimulator
• runtime-linuxbionic
• runtime-maccatalyst
• runtime-coreclr pgostress
• runtime-coreclr jitstress-isas-avx512
• runtime-tools-tests
• runtime-libraries-mono outerloop
• runtime-sanitized
• runtime-wasm-dbgtests
• runtime-wasm-optional
• runtime-ioslike-coreclr
• runtime-ioslike-mono
• runtime-nativeaot-outerloop

[EgorBo]

/azp run runtime-coreclr outerloop, runtime-coreclr jitstress-isas-x86, runtime-coreclr jitstress-isas-arm

[azure-pipelines]

Azure Pipelines successfully started running 3 pipeline(s).

[Sergio0694]

[EgorBo]

/azp run runtime-coreclr jitstress-isas-x86, runtime-coreclr jitstress-isas-arm

[azure-pipelines]

Azure Pipelines successfully started running 2 pipeline(s).

[EgorBo]

@tannergooding or anyone from @dotnet/jit-contrib could you take a look?
tl;dr: we currently detect must-expand intrinsics as "does this call invoke current root method" while it should be - "does this call invoke itself" regardless of who is the root - this led to problems with [Intrinsic] attribute on top of virtual methods in Type type. The interesting side-effect from this is that we now can inline must-expand SIMD intrsinsics which we give up due to non-constant args for IMM etc. It allows jit to fold the fallback if late phases detect a constant. Example:

Vector128<int> Foo(Vector128<int> v)
{
    byte mask = 0b11110000;
    return Sse2.Shuffle(v, mask);
}

Main:

; Method MyBench:Foo
       mov      rcx, rdx
       mov      rdx, r8
       mov      r8d, 240
       tail.jmp [Sse2:Shuffle(Vector128`1[int],ubyte):Vector128`1[int]]
; Total bytes of code: 18

PR:

; Method MyBench:Foo
       vpshufd  xmm0, xmmword ptr [r8], -16
       vmovups  xmmword ptr [rdx], xmm0
       mov      rax, rdx
       ret      
; Total bytes of code: 14

Diffs

There a few regressions which are just PMI-artifacts, e.g. we inline the fallback here. but it shouldn't matter since the method is supposed to be inlined (AggressiveInlining).

• a couple of r2r regressions - I investigated those and looks like they're "cancel inlining because BlockNonDeterministicIntrinsics said so"

[tannergooding]

There's two libraries-pmi regressions that look interesting. I think in production they aren't actually hit for the cases highlighted because they are small methods that get inlined and the constant is propagated.

But, it showcases an issue where we aren't preserving it as a call for the non-constant case, so we end up expanding the jump table inline instead, this can lead to nasty code explosion in some cases and that is potentially concerning (particularly for T0 like code).

[Sergio0694]

Egor pls

[EgorBo]

@tannergooding PTAL again, your work to rewrite HWINTRINSIC as calls helped to remove the hacks I had in the previous version.

[tannergooding]

This looks a lot better now, thanks!

[Sergio0694]