for loop desugaring treats the loop's trailing expression in unusual way

[petrochenkov]

Namely, it's treated as a semicolon-less statement in a middle of the block.

// Like this:
fn main() {
    { 0 } // <- semicolon-less statement in a middle of the block

    ()
}

• "Semicolon-less" means that its type should unify with () (Do not discern between statements with and without semicolon after lowering to HIR #61753).
• "In a middle of the block" means that it's not treated as trailing and its lifetime is not extended from the block outwards. That means "does not live long enough" errors like these do not happen (a minimized test case would be appreciated).

This is unusual because

• No other block construction or loop does this.
• The desugared form cannot be written in source code (it will either need a semicolon, or will be considered trailing).

Git blame says that the behavior was introduced by @Zoxc in #42265.

We cannot remove this behavior because it appears to be useful and code will break if it's removed.
We can, however, use it in other block constructions for which the block's trailing expression does not represent its result.

Block construction	Trailing expression is the result
if cond { ... } [else { ... }]	Yes
if let pat = expr { ... } [else { ... }]	Yes
while cond { ... }	No
while let pat = expr { ... }	No
for pat in expr { ... }	No
loop { ... }	No
[unsafe] { ... }	Yes
async { ... }	???
try { ... }	Yes

[petrochenkov]

We can, however, use it in other block constructions for which the block's trailing expression does not represent its result.

Or, perhaps, avoid the lifetime prolonging behavior for trailing non-result expressions (aka trailing expressions in loops) somewhere at MIR lowering time.

With #61988 merged that's going to be very simple:

Block construction	Trailing expression is the result
hir::ExprKind::Block	Yes
hir::ExprKind::Loop	No

[eddyb]

Or, perhaps, avoid the lifetime prolonging behavior for trailing non-result expressions (aka trailing expressions in loops) somewhere at MIR lowering time.

I think all of the scoping is still controlled by rustc::middle::region, FWIW (that module should've been renamed whenever the contents were renamed...)

[oskgo]

Triage:

The current desugaring of for can be written using loop (#90352), so there should no longer be any special behavior here.

I suspect that the regressions in the crater run would have been rejected by NLL, either because the originals were unsound, or because they hit NLL regressions (#62170), and those are no longer relevant with NLL being on by default. The crates I looked at were all using the 2015 edition, which would have been using Lexical Lifetimes before Rust 1.63.

I think this issue is outdated and should be closed, but I can't be 100% confident with there being no simple original repro.