We can't copy trivial structs

[ChuanqiXu9]

Reproducer:

struct S {
    int a;
    int b;
};

S getS(S);

S foo() {
    S s;
    s = getS(s);
    return s;
}

when we try to emit CIR for it, we will get

clangir/clang/lib/CIR/CodeGen/CIRGenClass.cpp

Line 151 in 329aca9

llvm_unreachable("NYI");

[smeenai]

Yeah, this is part of a broader category of issues where we want to emit a constructor call instead of a memcpy to facilitate better analysis, but then we need to remove trivial constructors and convert to memcpy where possible to have parity with CodeGen. I ran into #1055 with a slightly more complex case, and #1098 is a current missed optimization. I think we'd want to emit copy constructors here and then convert to memcpy later in the pipeline somewhere. ABI lowering also comes into play for the function arguments and return values, which is another consideration.

[ChuanqiXu9]

Yeah, this is part of a broader category of issues where we want to emit a constructor call instead of a memcpy to facilitate better analysis, but then we need to remove trivial constructors and convert to memcpy where possible to have parity with CodeGen. I ran into #1055 with a slightly more complex case, and #1098 is a current missed optimization. I think we'd want to emit copy constructors here and then convert to memcpy later in the pipeline somewhere. ABI lowering also comes into play for the function arguments and return values, which is another consideration.

Thanks for replying. I feel currently CIR may be slightly low level. It should be good enough here to emit a constructor here. Such layering is helpful for analysis too.

[smeenai]

Yup, emitting a constructor here is good (and we've been trying to represent all constructor calls faithfully in CIR); any optimizations can come later.