tail calls: Passing Indirect Operands into LLVM hits a span_bug

[tnuha]

Essentially the meat of #144293 (comment) and #144232 (comment)

We want to resolve the span_bug in

rust/compiler/rustc_codegen_ssa/src/mir/block.rs

Line 1059 in 7cd9505

span_bug!(

.

LLVM is currently emitted wrongly for instances with indirect operands, mostly due to cleanup after the fact. Can see this by the MIR of tests/ui/explicit-tail-call/recursion-etc-u64wrapper.rs:

fn count(_1: U64Wrapper, _2: U64Wrapper) -> U64Wrapper {
    debug curr => _1;
    debug top => _2;
    let mut _0: U64Wrapper;
    let mut _3: bool;
    let mut _4: u64;
    let mut _5: u64;
    let mut _6: u64;
    let mut _7: U64Wrapper;
    let mut _8: u64;
    let mut _9: u64;
    let mut _10: (u64, bool);
    let mut _11: std::string::String;
    let mut _12: U64Wrapper;
    let mut _13: bool;

    bb0: {
        _13 = const false;
        _13 = const true;
        _5 = copy (_1.0: u64);
        _4 = std::hint::black_box::<u64>(move _5) -> [return: bb1, unwind: bb6];
    }

    bb1: {
        _6 = copy (_2.0: u64);
        _3 = Ge(move _4, move _6);
        switchInt(move _3) -> [0: bb3, otherwise: bb2];
    }

    bb2: {
        _13 = const false;
        _0 = move _1;
        drop(_2) -> [return: bb5, unwind: bb9];
    }

    bb3: {
        _9 = copy (_1.0: u64);
        _10 = AddWithOverflow(copy _9, const 1_u64);
        assert(!move (_10.1: bool), "attempt to compute `{} + {}`, which would overflow", move _9, const 1_u64) -> [success: bb4, unwind: bb6];
    }

    bb4: {
        _8 = move (_10.0: u64);
        _13 = const false;
        _11 = move (_1.1: std::string::String);
        _7 = U64Wrapper { x: move _8, arbitrary: move _11 };
        _12 = move _2;
        tailcall count(Spanned { node: move _7, span: r.rs:24:13: 27:14 (#0) }, Spanned { node: move _12, span: r.rs:28:13: 28:16 (#0) });
    }

    bb5: {
        return;
    }

    bb6 (cleanup): {
        drop(_2) -> [return: bb9, unwind terminate(cleanup)];
    }

    bb7 (cleanup): {
        resume;
    }

    bb8 (cleanup): {
        drop((_1.1: std::string::String)) -> [return: bb7, unwind terminate(cleanup)];
    }

    bb9 (cleanup): {
        switchInt(copy _13) -> [0: bb7, otherwise: bb8];
    }
}

I believe the only real work involved would be migrating the bbs related to cleanup before the tailcall, or placing the object explicitly on the stack for LLVM's byval to clean it up for us.

[tnuha]

@rustbot label: +F-explicit_tail_calls +E-medium +I-ICE +A-LLVM

[matthiaskrgr]

looks like a duplicate of #144754

[compiler-errors]

The discussion in #144232 (comment) suggests that this is a bit more complicated than just allocating the local onto the stack.

[tnuha]

The discussion in #144232 (comment) suggests that this is a bit more complicated than just allocating the local onto the stack.

Makes sense- I think my confusion was from comparing different places, and seeing the following:

https://github.com/rust-lang/rust/blob/master/compiler%2Frustc_codegen_llvm%2Fsrc%2Fabi.rs#L488

More or less implying, to me, at least, that byval is used with stack allocated arguments, and to my understanding byval does eventually free those objects regardless. But I haven't been successful in finding any stack allocated arguments with explicit TCE anyway.