1.42.0-beta1 regression - LLVM ERROR: Broken module found

[pinver]

Compiles fine in 1.41.0

// LDC2 1.42.0 LLVM bug minimal reproduction
// Compile: ldc2 --preview=dip1000 standalone_test.d
// Error: "Referring to an instruction in another function!" + "LLVM ERROR: Broken module"
module standalone_test;

struct Column { string name; }
struct Table { Column[string] columns; }
struct Schema { Table[string] tables; }
struct Database(int E) { Schema[string] schemas; }

auto makeDB() => immutable Database!1(["s": immutable Schema(["t": immutable Table(["c": immutable Column("c")])])]);
enum immutable(Schema) schema = makeDB().schemas["s"];

struct S(const Schema s) {
    auto f(string x)() {
        static string g() { return (*("t" in s.tables)).columns.length > 0 ? "y" : "n"; }
        enum v = g();
        return R();
    }
    struct R { string front() => ""; }
}

alias Row = typeof(S!schema().f!"x".front);
void main() {}

[thewilsonator]

Reduced further to

struct Schema { int[int] tables; }
struct S(Schema s) {
    void g() { cast(void)s.tables; }
}
auto makeSchema() => Schema([0: 0]);
alias Row = S!(makeSchema());

yielding

Referring to an instruction in another function!
  store [1 x i32] zeroinitializer, ptr %2, align 4
Referring to an instruction in another function!
  %2 = insertvalue { i64, ptr } { i64 1, ptr undef }, ptr %2, 1
Referring to an instruction in another function!
  store [1 x i32] zeroinitializer, ptr %3, align 4
Referring to an instruction in another function!
  %3 = insertvalue { i64, ptr } { i64 1, ptr undef }, ptr %3, 1

the two values referred to in the error are the [0: 0] in auto makeSchema() => Schema([0: 0]); as changing some of the ints in int[int] to floats changes the type in the error message.

[thewilsonator]

Last known good commit was 3883e04 the merge of #5017
An upper bound on the first known bad commit that builds it is 2e07f51

Unfortunately this commit range doesn't build inside due to adaptations from the new frontend version so it is very difficult to bisect much further.
I do note that there were significant changes to the AA implementation over that commit range and the above test case fails to fail if int[int] is replaced by some other type, so those changes are likely to blame.

[thewilsonator]

CC @rainers as you authored those AA changes

[rikkimax]

Okay, I've been able to replicate this on Windows.

It appears to be something is wrong coming up against DIP1000.

[thewilsonator]

forcing https://github.com/ldc-developers/ldc/blob/master/dmd/expressionsem.d#L3660 to not be taken fixes the issue (although it causes druntime tests to fail)