Hi all, I'm trying to use LLVM on a 64-bit PowerPC Linux box. This platform has a weird way of dealing with function pointers which I don't think LLVM has support for. Can anyone point me in the right direction so I can start implementing it? The problem is this. On most platforms, a function pointer in C is the address of the entry point of the function. This is the case for 32-bit Linux PowerPC, and, from looking at the code in LLVM it seems to be the case for both 32- and 64-bit Darwin. A call on these platforms looks like this, assuming the function pointer is in rX, and all arguments to the call have been put in the appropriate places: mtctr rX bctrl Simple. On 64-bit Linux PowerPC, however, a function pointer in C is the address of a function descriptor, a structure containing three pointers: - The first pointer contains the address of the entry point of the function. - The second pointer contains the TOC base address for the function. - The third pointer contains the environment pointer (for languages that need one). The TOC base address needs to be in r2, and the environment pointer needs to be in r11. Starting with the C function pointer in rX, a call on 64-bit Linux PowerPC looks something like this: ld rY, 0(rX) ; where rY is some free register mtctr rY std r2, 40(r1) ; optional ld r2, 8(rX) ld r11, 16(rX) bctrl ld r2, 40(r1) ; optional The two lines marked "optional" could be omitted since LLVM doesn't generate or use a TOC. Any idea where I need to look to implement this? Cheers, Gary -- http://gbenson.net/
On Fri, Jul 3, 2009 at 7:29 AM, Gary Benson<gbenson at redhat.com> wrote:> I'm trying to use LLVM on a 64-bit PowerPC Linux box. This platform > has a weird way of dealing with function pointers which I don't think > LLVM has support for. Can anyone point me in the right direction so > I can start implementing it?You can support lowering a call any way you want for your target by custom-lowering the CALL instruction in CodeGen. The relevant code is PPCTargetLowering::LowerCALL in llvm/lib/Target/PowerPC/PPCISelLowering.cpp. I'd suggest starting there, and asking if there's anything you can't figure out. -Eli
btw, this PPC convention is not unique to 64-bit Linux; it is used on AIX 32-bit and I believe it was used on earlier versions of MacOSX (but is no longer). Neither of those is an immediate target but you might want to keep it in mind when deciding where to put the new code you're adding. On Jul 3, 2009, at 7:29 AMPDT, Gary Benson wrote:> Simple. On 64-bit Linux PowerPC, however, a function pointer in C > is the address of a function descriptor, a structure containing three > pointers: > > - The first pointer contains the address of the entry point of > the function. > - The second pointer contains the TOC base address for the function. > - The third pointer contains the environment pointer (for languages > that need one).
Seemingly Similar Threads
- [LLVMdev] Implementing llvm.atomic.cmp.swap.i32 on PowerPC
- [LLVMdev] Implementing llvm.atomic.cmp.swap.i32 on PowerPC
- [LLVMdev] Implementing llvm.atomic.cmp.swap.i32 on PowerPC
- [LLVMdev] Implementing llvm.atomic.cmp.swap.i32 on PowerPC
- [LLVMdev] Implementing llvm.atomic.cmp.swap.i32 on PowerPC