First class aggregates of small size: split when used in function call

Hello,

In my LLVM frontend (CLR/MSIL), I am currently using first-class aggregates to represent loaded value types on the “CLR stack”.

However, I noticed that when calling external method taking those aggregate by value, they were not passed as I expected:

%COLORREF = type { i8, i8, i8, i8 }

declare i32 @SetLayeredWindowAttributes(i8*, %COLORREF, i8, i32)
I call this function with call x86_stdcallcc (it’s a Win32 function, loaded with GetProcAddress)

However, checking the assembly code, it seems that the %COLORREF gets split due to the calling convention: first i8 field go through %edx, but the 3 next fields go through the stacks.
I would like all of it to go through either a single 32bit register or a 32bit stack value (since all of the structure fits in a i32 and it is already packed in memory that way before the call).

I was thinking using alloca with sret/byval might help, but I am not even sure since it is enough, since clang also seems to actually use i16 or i32 (and even i32+i16 or i32+i32) to represent such struct <= 8 bytes when passing them to a method (even if they contain many smaller i8 fields).

Does somebody know if only alloca with sret/byval is enough or if I also need to concat myself smaller struct into i32 types like clang does to be sure it won’t be split across registers?
Any other hint or idea on how I can achieve this?

Also, I was wondering which is the current recommendation (sret/byval with alloca for every copy vs first-class aggregate) considering the current state of LLVM and supported optimizations. Since clang uses sret/byval, I expect it to be more optimized/mature, but I might be wrong.

I suppose LLVM will easily understand/optimize all those additional aggregate alloca/memcpy I will end up doing if I were to switch to a sret/byval approach?

Thanks,

The current recommendation for matching external C ABIs is actually “use Clang”, as in literally link it into your program and try to leverage the methods of clang::CodeGen::CodeGenFunction. This is the only way to get the lowering 100% correct, but it’s a lot of work, so depending on your needs, you may want to roll your own lowering of high-level function prototype to LLVM function prototype.

If you roll your own, then LLVM generally passes FCAs as though they were split into constituent elements. I think ARM does something different here, though. :frowning:

Clang sometimes uses integers of appropriate size on non-x86 architectures to try to model the usage of an integer register to pass the whole struct. For a small struct, this is good because element accesses can be transformed into shifts, masks, and truncs.

If you just want to match MSVC’s C ABI, byval is probably the way to go. LLVM is not very good at optimizing it, but it will do the right thing. Splitting into i32-sized chunks would also work.

I should mention this was the topic of a talk by John McCall at the LLVM
dev meeting:

See slides starting around #101 for why this is hard and the many different
ways of lowering this C function prototype into LLVM IR:

typedef struct {
float x, y;
} Point2f;
Point2f flipOverXAxis(Point2f point) {
// ...
}

Hi Virgile,

You might be interested in the interop code I wrote for my LLVM.NET binding: https://bitbucket.org/lost/llvm.net/src/d8014b07723c69571e188a453ab39c764252985c/LLVM/Interop/?at=default

  • Victor