reduced.swift:

// RUN: %swiftc -frontend -target thumbv7-unknown-windows-msvc -parse-stdlib -parse-as-library -I Inputs/usr/include -S -o - %s

import SRoA

public func g(_ s: S) {
{{ return f(s)}}
}

Inputs/usr/include/SRoA.h:

typedef struct S {
{{ float f;}}
{{ float g;}}
} S;

void f(S);

Inputs/usr/include/module.map:

module SRoA {
{{ header "SRoA.h"}}
}

cc @rjmccall, aschwaighofer@apple.com (JIRA User)

Currently, IRGen (in GenCall.cpp) will unconditionally explode llvm::StructTypes if the clang::CodeGen::ABIArgInfo type is Direct.

John explained to me that clang::CodeGen::ABIArgInfo has extra state whether this expansion is allowed to happen (ABIArgInfo::getCanBeFlattened).

I can see three places in GenCall.cpp that need to be modified to respect this flag: emitDirectForeignParameter, expandExternalSignatureTypes, emitDirectExternalArgument.

Yeah, I was using `getCanBeFlattened`. However, I am having some issues with the threading of the `ArgInfo`. Note that this also causes some issues since I saw cases of `getCanBeFlattened` returning `true` for `i8*`.

'i8*' is an llvm::PointerType not an llvm::StructType. clang's lib/CodeGen/CGCall.cpp is instructive.

  llvm::StructType *STy = dyn_cast<llvm::StructType>(AI.getCoerceToType());
  if (AI.isDirect() && AI.getCanBeFlattened() && STy) {
    ...

Agreed; I figured that out later, I just meant that we need a bit more checking 🙂

BTW, changing just one site was sufficient; however, that would still cause issues in `visitFullApplySite`. IRGenSIL is where I was missing access to the the `ArgABIInfo`.

#12525 should address this issue, but there are a couple of failing tests it seems.

Thank you!