[WIP] traits/fulfill: track GenericArgss instead of InferTys in stalled_on.

[eddyb]

EDIT: superseded by #70213 for performance reasons.

This PR addresses the representation side of #70180, but only actually collects ty::Infers via Ty::walk into stalled_on (collecting ty::ConstKind::Infers requires #70164).

However, it should be enough to handle #70107's needs (WF obligations are stalled only on the outermost type/const being an inference variable, no walk-ing is involved).

This PR will make step-by-step changes in order to check the perf impact of each.

r? @nikomatsakis cc @nnethercote

[eddyb]

I'm doing a perf run with just this first commit, to assess the impact of having to match on Ty to get at the ty::InferTy inside each stalled_on entry:

traits/fulfill: track full Tys instead of InferTys in stalled_on.

@bors try @rust-timer queue

[rust-timer]

Awaiting bors try build completion

[bors]

⌛ Trying commit e63c410 with merge fc63e0e7ac01afe8ebf2cbb38466eb80c1d42922...

[bors]

☀️ Try build successful - checks-azure
Build commit: fc63e0e7ac01afe8ebf2cbb38466eb80c1d42922 (fc63e0e7ac01afe8ebf2cbb38466eb80c1d42922)

[rust-timer]

Queued fc63e0e7ac01afe8ebf2cbb38466eb80c1d42922 with parent f4c675c, future comparison URL.

[rust-timer]

Finished benchmarking try commit fc63e0e7ac01afe8ebf2cbb38466eb80c1d42922, comparison URL.

[Centril]

Perf seems mostly improved I'd say, but it's a bit of a wash.

[eddyb]

@Centril most of the perf improvements around 1% I've seen lately look dubious and unrelated (i.e. likely noise), so I focus on the regressions and try to find a relevant one.

Besides, it should be impossible for this PR to make anything faster, so there's that.

keccak-opt clean and inflate-check clean paint a pretty clear picture: type-checking is slower.

I think the efficient approach moving forward would be to make an enum StalledOnInferVar with these variants (Ty, Int and Float) plus a Const one:

rust/src/librustc_infer/infer/mod.rs

Lines 1631 to 1655 in 1057dc9

	ty::TyVar(v) => {
	use self::type_variable::TypeVariableValue;
	// If `inlined_probe` returns a `Known` value its `kind` never
	// matches `infer`.
	match self.infcx.inner.borrow_mut().type_variables.inlined_probe(v) {
	TypeVariableValue::Unknown { .. } => false,
	TypeVariableValue::Known { .. } => true,
	}
	}
	ty::IntVar(v) => {
	// If inlined_probe_value returns a value it's always a
	// `ty::Int(_)` or `ty::UInt(_)`, which never matches a
	// `ty::Infer(_)`.
	self.infcx.inner.borrow_mut().int_unification_table.inlined_probe_value(v).is_some()
	}
	ty::FloatVar(v) => {
	// If inlined_probe_value returns a value it's always a
	// `ty::Float(_)`, which never matches a `ty::Infer(_)`.
	//
	// Not `inlined_probe_value(v)` because this call site is colder.
	self.infcx.inner.borrow_mut().float_unification_table.probe_value(v).is_some()
	}

I'm closing this PR to serve as a tombstone for the inefficient approach.
EDIT: proposed approach submitted as #70213.

[eddyb]

Oops, I should've renamed this to trait_ref.