llvm dev - Jul 2008 - [LLVMdev] Structs as first class values.

Hi,

I'm trying to implement structs as first class values. If I assemble

; ModuleID = 't0022.s.bc'
target datalayout = 
"e-p:32:32:32-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:32:64-f32:32:32-f64:32:64-v64:64:64-v128:128:128-a0:0:64-f80:32:32"
target triple = "i386-pc-linux-gnu"
         %struct.foo = type { i32, i32, i32 }

define %struct.foo @main() nounwind  {
entry:
         %foo = alloca %struct.foo               ; <%struct.foo*>
[#uses=1]
         %"alloca point" = bitcast i32 0 to i32          ; <i32>
[#uses=0]
         br label %return

return:         ; preds = %entry
         %t = load %struct.foo* %foo             ; <%struct.foo> [#uses=1]
         ret %struct.foo %t
}

It assembles without complaint.

If I do llc I get an assert:

x86:
Return operand #2 has unhandled type i32
llc[0x8b5981e]
...

arm:
llc: /home/rich/llvm-trunk-new/lib/VMCore/ValueTypes.cpp:109: static 
llvm::MVT llvm::MVT::getMVT(const llvm::Type*, bool): Assertion `0 && 
"Unknown type!"' failed.
...

My questions are:

  Is this supposed to work? (I understand it may be a work in progress.)
  Is the LLVM well formed at the source level?

-Rich

On Jul 20, 2008, at 6:25 AM, Richard Pennington wrote:
> Hi,
>
> I'm trying to implement structs as first class values. If I assemble
>
> It assembles without complaint.
Yep, the code looks fine.
> If I do llc I get an assert:
>
>
> My questions are:
>
>  Is this supposed to work? (I understand it may be a work in  
> progress.)
Yes, that is supposed to work in theory.  In practice, most targets  
only support returning up to two values in registers.  Eventually, we  
want the code generator to support returning excess return values on  
the stack, but we don't have that code yet.
>  Is the LLVM well formed at the source level?
Yep, this is just a limitation of the current code generators.  Prior  
to first-class aggregate support, we did have MRV support, which has  
the same limitation so far.

-Chris

Chris Lattner wrote:
> On Jul 20, 2008, at 6:25 AM, Richard Pennington wrote:
> Yes, that is supposed to work in theory.  In practice, most targets  
> only support returning up to two values in registers.  Eventually, we  
> want the code generator to support returning excess return values on  
> the stack, but we don't have that code yet.
Thanks for the info. Is the intention to implement this in an ABI 
compatible manner? That is, pass the address to put the return value as 
a hidden first parameter, for example?

-Rich