Workaround for slow codegen on x86 atomic load

[Alex Guteniev (AlexGuteniev)]

Compiler emits jumptable or jcc sequence that prevents inlining
of atomic load; separation of order check and barrier condition helps

[Alex Guteniev (AlexGuteniev)]

Not sure if it should be applied, or the compiler should be fixed instead.

[AZero13 (AZero13)]

Send feedback to the Visual Studio compiler team via visual studio feedback

[Alex Guteniev (AlexGuteniev)]

The feedback was already sent by Christian Fersch (@Chronial) , see DevCom-1491677

I'm not really sure if there should be workaround in code. There are at least three optimizer issues:

• Compiler barrier emits npad 1 (nop instruction) instead of being truly no-op
• switch does not collapse to if, even though there are only two possibilities
• Decision whether to inline a function is based on full function size, not the size when constant propagation happens

Fixing just one of them would help. The workaround is a particular solutiin for wider problems.

[Alex Guteniev (AlexGuteniev)]

Proof that it works: https://godbolt.org/z/5e7re6bGY

[Stephan T. Lavavej (StephanTLavavej)]

Looks good! The behavior is equivalent, and the code is simpler, so this is a good perma-workaround (no TRANSITION comment).

[Charlie Barto (barcharcraz)]

I also like the "workaround" code better than the old code.

[Stephan T. Lavavej (StephanTLavavej)]

I'm mirroring this to an MSVC-internal PR. Please notify me if any further changes are pushed.

[Stephan T. Lavavej (StephanTLavavej)]

Changed the title to say "slow codegen", as the compiler back-end team conventionally uses "bad codegen" to mean incorrect codegen.

[Stephan T. Lavavej (StephanTLavavej)]

Thanks... for... improving... this... slow... codegen... ! 🐌 😹 🐇