[mir-opt] optimization idea: bubble `return` terminators to predecessor blocks

[oli-obk]

cc @rust-lang/wg-mir-opt

if we have a basic block with no statements and just a return terminator, any other basic block that just gotos into that block could just as well call return itself. Or do we have some rule that we can only have a single return terminator per mir::Body?

Maybe we can generalize this and bubble any terminator of a statement-less block into its predecessor terminators iff that predecessor terminator is a goto.

As an example take

rust/src/test/mir-opt/issue-62289/rustc.test.ElaborateDrops.before.mir

Line 84 in 7b80539

bb7: {

which is a basic block with no statements and just a return terminator, and both predecessor blocks have a goto to this block.

[hanna-kruppe]

if we have a basic block with no statements and just a return terminator, any other basic block that just gotos into that block could just as well call return itself. Or do we have some rule that we can only have a single return terminator per mir::Body?

I don't know if we have such a rule in (any part of) rustc, but in my experience this is a convenient assumption for some analyses and transforms. So depending on the expected value of "bubbling up" returns (at a glance, it seems low to me), I'd be cautious about doing it just because we can.

[ecstatic-morse]

The docs for TerminatorKind::Return do say that "this should occur at most once". I don't know how many places depend on this, however, and it shouldn't be too hard to fix up the ones that do.

[matthewjasper]

The built MIR only contains a single Return terminator. However the generator transform code appears to output MIR with multiple return terminators in src/test/mir-opt/generator-drop-cleanup/rustc.main-{{closure}}.generator_drop.0.mir

[nagisa]

this should occur at most once

I recall this being motivated by LLVM liking a singular return better than multiple returns. "Should" is not a "must", so it isn’t incorrect that we generate (sometimes) multiple returns. But it is still a guidance that stands.

As thus I don’t love what #72563 did /me shrugs.

[ecstatic-morse]

I concur with @nagisa. Canonicalizing a mistake made in the generator transform is wrongheaded IMO. If we want to do this, we should make a conscious decision. There's probably code that relies on only one Return terminator existing; we just haven't found those bugs yet.

[bjorn3]

Cranelift likes multiple returns better than single returns. It never duplicates ret instructions, even when it would decrease the amount of register moves and jumps.