[llvm-commits] [PATCH] Adding x32 ABI triple and separate StackSlotSize in MCAsmInfo

[Evan Cheng]

On Jan 18, 2013, at 2:35 PM, Eli Bendersky <eliben at google.com> wrote:

>>> Following David's email on llvmdev
>>> (http://lists.cs.uiuc.edu/pipermail/llvmdev/2013-January/058209.html),
>>> here is an initial patch for adding the x32 ABI triple to LLVM.
>>> 
>>> The patch is deliberately very simple. All it does is add the new
>>> Triple::EnvironmentType value for the x32 ABI, and uses it on one
>>> place - separating stack slot size and pointer size in MCAsmInfo. All
>>> targets provide the same value there, but x86-64 has to behave
>>> differently based on the ABI.
>>> One place where this is important is the data alignment factor in
>>> DWARF frame section. I discussed this with Rafael previosuly and we
>>> agreed this solution makes sense.
>>> 
>>> A limitation in our llvm-dwarfdump tool currently precludes me from
>>> adding more tests that direct the slot size. I will try to see if I
>>> can patch up llvm-dwarfdump to allow this separately. Once that
>>> capability is in place, I will add more tests.
>>> 
>>> Note that changes based on the x32 ABI will have to eventually
>>> propagate into other places, but I will break it to small reviewable
>>> and testable chunks over time. For the full picture of where this is
>>> going, you can refer to the patches Michael Liao sent last year
>>> (http://lists.cs.uiuc.edu/pipermail/llvmdev/2012-June/050360.html and
>>> other threads).
> 
>>> On Fri, Jan 18, 2013 at 1:56 PM, Evan Cheng <evan.cheng at apple.com> wrote:
>> Hi Eli,
>> 
>> What does "stack slot size" mean? Stack slots may not have the same size, depending on what objects are stored in them.
>> 
> 
> Good question, and perhaps a more descriptive name for the
> member/method can be found. The essential concept we're alluding to
> here is the slot allotted for saving registers on the stack. The x32
> ABI decouples register size from pointer size (which are equivalent on
> x86 and on default x86-64). And yet, the x32 ABI requires to save the
> full %rXX registers on the stack (section 3.2.1):
> 
>    This subsection discusses usage of each register. Registers %rbp, %rbx and
>    %r12 through %r15 “belong” to the calling function and the called
> function is
>    required to preserve their values. In other words, a called
> function must preserve
>    these registers’ values for its caller.
> 
> The patch teaches MCDwarf to use the correct "slot" size in its frame
> information output. To conclude, while previously the x86 target could
> simply assume that the pointer size is what it needs, with x32 this is
> no longer so, so a new concept is added. Would "RegisterStackSlotSize"
> be a better / more descriptive name?

It's theoretically possible to store smaller registers in smaller slots. Perhaps CalleeSaveStackSlotSize? Is it too long?

This change requires updates to out of tree targets, right? It might be worthwhile to send an email to llvm-dev as well. 

Evan

> 
> Eli