typeck: use NoExpectation to check return type of diverging fn

[durka]

Fixes#35849.

Thanks @eddyb.

[rust-highfive]

r? @eddyb

(rust_highfive has picked a reviewer for you, use r? to override)

[eddyb]

I'd put this in a variable or call a different version of check_block in the TyNever case.

[eddyb]

r=me with nit fixed.

[durka]

Nit fixed. This also needs to be beta-nominated.

[eddyb]

@bors r+

[bors]

📌 Commit 9d53faf has been approved by eddyb

[canndrew]

This won't allow something like fn foo() -> ! { 3 } to compile will it?

[eddyb]

@canndrew It's the same logic as before, I believe liveness is what prevents control-flow from exiting the function if the function is supposed to be divergent (in a more sound manner).

[durka]

@canndrew@eddyb yep we got a problem here:

[root@li1424-173 gh35849]# build/x86_64-unknown-linux-gnu/stage1/bin/rustc - -o bad <<<'fn foo() -> ! { 42 } fn main() { foo(); }'
warning: broken MIR (return = const 42i32): bad assignment (! = i32): Sorts(ExpectedFound { expected: !, found: i32 })
--> <anon>:1:17
|
1 | fn foo() -> ! { 42 } fn main() { foo(); }
| ^^
[root@li1424-173 gh35849]# ./bad
Segmentation fault (core dumped)

[eddyb]

Whoa how did that happen?! That's bad, because the code erroring before wasn't in typeck AFAIK.

@bors r-

[durka]

IRC investigation reveals that #35162 removed an important check from the liveness pass, which I will reinstate whenever LLVM finishes building. Now I'm wondering if that PR removed anything else important.

[canndrew]

Are you talking about the check in check_ret which was removed? If so, I think I was assuming that any block which diverges would get typed ! (including ones ending in an expression). Would it better to fix that than to make a special case for ! not having to be type-check correctly and then make another special case for it further down the pipeline?

[eddyb]

@canndrew Maybe that can work, but we'll always need a sanity check - if typeck ends up being able to handle divergence correctly, we'd just have a proper CFG-based check ICE instead of erroring.

[arielb1]

Maybe just add a "redundant" return type check always, like MIR does?

[canndrew]

I'm generally uncomfortable making special cases for ! - part of the motivation for this RFC is that a lot of this sort of thing can be handled naturally by the type system. Plus ! isn't the only diverging type.

[eddyb]

@canndrew The special cases need to be there until there's a RFC that says [x: !] is typed ! instead of [!; 1]or that [!; 1] is-a-subtype-of/coerces-to !, and until the definition of "diverging function" changes.
We risk too much if we don't have sanity checks for the CFG of a diverging function - the MIR warning being one of them, we should probably make those warnings ICE now.

[durka]

we should probably make those warnings ICE now

cf #35499

[canndrew]

I don't think that liveness should have special cases for ! at all. As far as its is concerned ! isn't a special case, it's just one of many uninhabited types. If liveness can handle Void it should be able to handle ! as well. The only reason it can't is that we just added another special case for ! which broke typechecking and allowed clear nonsense like fn foo() -> ! { 3 } to get through. It's one special case begetting another.

The special cases need to be there until there's a RFC that says [x: !] is typed ! instead of [!; 1] or that [!; 1] is-a-subtype-of/coerces-to !

Why? This can get typechecked correctly:

fnfoo() -> ! {[panic!()];}

So there's some logic that can see the ! in the middle of that expression. Can we make that logic run on the block's final expression as well and not just it's statements?

[eddyb]

@canndrew No, that doesn't get type-checked either. The error you get when the function is declared to return T and you have a semicolon and no divergence is not all control paths return a value, also from liveness.
That is, if you don't have a trailing expression, it's not checked at all, although this seems to only happen at the outermost level, a type error is produced if you nest blocks.

[durka]

Amended to restore the special case in Liveness::check_ret.

[durka]

Nits knitted.

[eddyb]

These should not be combined, i.e. the else below should be to fn_ret.is_never() alone, to match the old semantics.

[durka]

Knit nitted.

[eddyb]

@bors r+

[bors]

📌 Commit ddebea0 has been approved by eddyb

[durka]

Fixed the travis failure and added some FIXME comments.

We unfortunately needed to ignore a test added for #![feature(never_type)] because it can't be done with the current structure of the liveness code. A fix may be possible, but the consensus was that we shouldn't try to get it into this PR, to keep it backportable. I filed a separate bug.

[eddyb]

@bors r+

[bors]

📌 Commit 702ea71 has been approved by eddyb

[bors]

⌛ Testing commit 702ea71 with merge 5f31fda...

[alexcrichton]

@bors: retry force clean

• restarted buildbot

[bors]

⌛ Testing commit 702ea71 with merge 03e23c7...

[bors]

☀️ Test successful - auto-linux-32-nopt-t, auto-linux-32-opt, auto-linux-64-cargotest, auto-linux-64-cross-freebsd, auto-linux-64-debug-opt, auto-linux-64-nopt-t, auto-linux-64-opt, auto-linux-64-opt-no-mir, auto-linux-64-opt-rustbuild, auto-linux-64-x-android-t, auto-linux-cross-opt, auto-linux-musl-64-opt, auto-mac-32-opt, auto-mac-64-opt, auto-mac-64-opt-rustbuild, auto-mac-cross-ios-opt, auto-win-gnu-32-opt, auto-win-gnu-32-opt-rustbuild, auto-win-gnu-64-opt, auto-win-msvc-32-opt, auto-win-msvc-64-cargotest, auto-win-msvc-64-opt, auto-win-msvc-64-opt-no-mir, auto-win-msvc-64-opt-rustbuild
Approved by: eddyb
Pushing 03e23c7 to master...