"a" is "b" is true

[FeepingCreature]

Repro:

extern(C) void printf(const char*, ...);

void main()
{
    if ("a" is "b")
    {
        printf("oops.\n");
    }
}

The issue goes away if I switch to IRBuilder<llvm::NoFolder>. I suspect this is a side effect of strings being unnamed_addr, meaning LLVM not-entirely-unreasonably concludes "well, I can interpret icmp eq i8* @.str.1, @.str.2 however I feel like". Of course, "a" will never actually be merged with "b", but...

[JohanEngelen]

This bug is not related to LLVM or wrong optimization by LLVM. The generated code does not even contain an icmp because LDC recognizes that the condition is a compile-time constant and thus is known to be either true or false at compile-time and no branching is needed. (#3134, since 1.19)
The bug arises since 1.25, where for some reason "a" is "b" evaluates to true (when it is the condition of if).

Edit: Actually this is related to LLVM. The bug shows up since LLVM 11.0.1, not with 11.0.0. LDC 1.25 happens to be released with LLVM 11.0.1, LDC 1.24 with LLVM 11.0.0.

Reproducer for change of LLVM behavior: https://llvm.godbolt.org/z/T4G9jcs1e
Notice how 11.0.0 folds the icmp to false, but 11.0.1 does not constfold the if-condition at all.

[JohanEngelen]

Related LLVM change: https://reviews.llvm.org/D87123