[LLVMdev] questions about byval argument passing

[Eli Friedman]

On Thu, Oct 20, 2011 at 2:21 PM, Akira Hatanaka <ahatanak at gmail.com> wrote:
> Suppose this structure is passed by value to a function foo(struct S1),
>
> struct S1 {
>  int i0;
>  float f0;
>  double d0;
>  long long l0;
> };
>
> and the ABI dictates the elements of the structure are copied to both
> integer and floating pointer argument registers in the following way:
>
> (i0, f0) => 1st 64-bit integer argument register
> d0 => 2nd double float argument register (doubles are passed in FP registers)
> l0 => 3rd 64-bit integer argument register
>
> What is the best way to find out whether integer registers or floating
> pointer registers should be used for argument passing inside
> TargetLowering::LowerCall? I can get basic information of the
> structure being passed (for example, by calling
> ArgFlagsTy::getByValSize() or ArgFlagsTy::getByValAlign() to get size
> and alignment), but there doesn't seem to be an easy way to find more
> detailed information. Do I have to get the instance of StructType for
> structure S1 and analyze it?

The way we do stuff like this for x86-64 is adding handling to clang;
for example, to handle your case, make clang lower "void f(struct S1
x)" to the following:

define void @x(i64 %x.1, double %x.2, i64 %x.3)

Basically, an LLVM struct type doesn't really encode enough
information to accurately compute the C calling convention, so we
handle the tricky cases in the frontend.   The relevant code is in
clang/lib/CodeGen/TargetInfo.cpp .

-Eli