SROA is incompatible with RFC 3943 semantics

[tmiasko]

#![feature(custom_mir, core_intrinsics)]
#![allow(internal_features)]
use std::intrinsics::mir::*;
use std::mem::MaybeUninit;
#[custom_mir(dialect = "runtime")]
pub fn main() {
    mir! {
        let _1: (usize, MaybeUninit<usize>);
        let _2: MaybeUninit<usize>;
        {
            _1.0 = 0;
            _2 = _1.1;
            Return()
        }
    }
}

rustc --crate-type=lib a.rs -Zunpretty=mir -Zmir-opt-level=0 -Zmir-enable-passes=+ScalarReplacementOfAggregates

// WARNING: This output format is intended for human consumers only
// and is subject to change without notice. Knock yourself out.
// HINT: See also -Z dump-mir for MIR at specific points during compilation.
fn main() -> () {
    let mut _0: ();
    let mut _1: (usize, std::mem::MaybeUninit<usize>);
    let mut _2: std::mem::MaybeUninit<usize>;
    let mut _3: usize;
    let mut _4: std::mem::MaybeUninit<usize>;

    bb0: {
        _3 = const 0_usize;
        _2 = copy _4;
        return;
    }
}

cc @Amanieu

[Amanieu]

I addressed this in #163335. The basic idea is that a write to a place based on _1 has the side-effect of allocating _1. This side effect must be preserved when splitting by emitting StorageAlloc for every field at that point (except the field being written to, since it gets that implicitly from the write anyways).