[do not merge] Try turning on MIR inliner by default

[wesleywiser]

Try turning on the MIR inliner by default to measure the performance impact.

Rebased on top of #68170 to get that fix.

r? @ghost

[wesleywiser]

@bors try
@rust-timer queue

[rust-timer]

Awaiting bors try build completion

[bors]

⌛ Trying commit a3db620 with merge dc0f962...

[bors]

☀️ Try build successful - checks-azure
Build commit: dc0f962 (dc0f962493ef0ce61d14530aee40c9fe64de0caf)

[rust-timer]

Queued dc0f962 with parent c06e4ac, future comparison URL.

[rust-timer]

Finished benchmarking try commit dc0f962, comparison URL.

[wesleywiser]

Interestingly, we're already seeing a lot of improvements across most tests which is very nice!

Looking at the (major) regressions, I'm seeing the following patterns:

• LLVM taking much more time

  • This affects the deep-vector-opt, keccak-debug, regex-debug, encoding-debug, and most significantly deeply-nested-opt.
  • MIR optimization results in significantly increased compile time for a rustc-perf benchmark #68105 seems related

• normalize_ty_after_erasing_regions being called more times:

  • This affects the await-call-tree-check, await-call-tree-debug, and await-call-tree-opt tests.

[bjorn3]

normalize_ty_after_erasing_regions being called more times:

This may be the monomorphization of the mir body:

rust/src/librustc_mir/transform/inline.rs

Lines 125 to 133 in 632387f

	let callee_body = if self.consider_optimizing(callsite, callee_body) {
	self.tcx.subst_and_normalize_erasing_regions(
	&callsite.substs,
	param_env,
	callee_body,
	)
	} else {
	continue;
	};

[Zoxc]

#68606 may help with the await-call-tree regression.