[X86] Scheduler models should use UnsupportedFeatures to indicate unsupported ISAs

[RKSimon]

Bugzilla Link	37601
Version	trunk
OS	Windows NT
Blocks	#31672 #36950
CC	@adibiagio,@legrosbuffle,@topperc,@gchatelet,@LebedevRI,@rotateright

Extended Description

We've mentioned this several times: [Bug #​17367] [Bug #​26792], [Bug #​35608]

But we still don't make use of this field to indicate unsupported ISAs and reduce test ranges.

NOTE: We must be careful with some models are they are reused on CPUs that DO support extra ISAs (Haswell on KnightsLanding/KnightsMill), plus SandyBridge is used as our current 'modern Generic cpu model' so needs to touch all ISAs.

[LebedevRI]

As far as i can tell, since "UnsupportedFeatures" is already used in other targets, only thing missing is actually adding correct
list UnsupportedFeatures = [..]
into "each" SchedMachineModel?

[topperc]

So every time we had a new ISA we would have to update every scheduler model? That seems like a maintenance burden and something we are likely to forget to do.

[LebedevRI]

So every time we had a new ISA we would have to update every scheduler
model? That seems like a maintenance burden and something we are likely to
forget to do.

Yes, this design seems backwards.

[RKSimon]

Hmm, it shouldn't be too awkward to support SupportedFeatures as an EXCLUSIVE alternative to UnsupportedFeatures - what do you think?

IIRC there was some chat about improving model checks (CodeGenSchedModels::checkCompleteness etc.) in tblgen as well, but I can't find the ticket.

[LebedevRI]

Hmm, it shouldn't be too awkward to support SupportedFeatures as an
EXCLUSIVE alternative to UnsupportedFeatures - what do you think?

I was actually trying to hack something like that yesterday, but got nowhere.
What i don't like is if there will be more than a single source of truth.
There already are all these Feature* in X86.td, and even a ProcessorFeatures
thingy.

That being said, as far as i can tell, that is not sufficient?
The UnsupportedFeatures is only used in CodeGenSchedModels::checkCompleteness,
so it won't magically 'blacklist' unsupported instructions.

IIRC there was some chat about improving model checks
(CodeGenSchedModels::checkCompleteness etc.) in tblgen as well, but I can't
find the ticket.

[RKSimon]

mentioned in issue #36950

[adibiagio]

mentioned in issue llvm/llvm-bugzilla-archive#42286