[LLVMdev] Stack alignment in leaf functions

[Chris Lattner]

On Tue, 16 May 2006, Vladimir Prus wrote:
> right at the moment, LLVM won't align the stack for leaf functions. So, if
> stack alignment is 8 bytes, and leaf function has 1 variable of 4 bytes,
> the size of frame will be computed as 4 bytes, making stack pointer inside
> function non-aligned.

Ok.

> I've mentioned this before:
>   http://lists.cs.uiuc.edu/pipermail/llvmdev/2005-March/003701.html
> but did not pursue the issue back then.
>
> For my target, the stack should be aligned even for leaf functions. In
> particular, hardware interrupt can arrive at any time, and stack should be
> 8-byte aligned at the time of interrupt, otherwise things will break.

Gotcha.

> I can address this issue by using two tricks together:
>
> 1. Calling MachineFrameInfo::setHasCalls in my backend
> 2. Using the attached patch, so that value of setHasCalls is not reset to
> false later.
> Comments?

Unfortunately, I don't *think* this is the right approach.  I 
really think that HasCalls should be an accurate indication of whether or 
not there are calls, it shouldn't be overloaded if possible.

I *think* that this should be an issue that you can handle completely in 
your target code.  In particular, the stack size set in the FrameInfo is 
just a required minimum stack size.  In your <Target>RegisterInfo.cpp 
file, you can use this size as a starting point, then adjust it as you see 
fit.  In the X86 backend, for example, it does do explicit realignment of 
the size in certain cases, I think you should be able to do it as well.

> BTW, I've checked other uses of MachineFrameInfo::hasCalls. For example,
> PPCRegisterInfo uses it to check if if callee save registers can be stored
> in something called "red zone", whatever that is.

Right.  In the case of PPC/X86-64, a red zone is a small region that 
you are allowed to to offset below the stack pointer.  Signals are 
delivered at some offset below it.

> I wonder how this relates to interrupts. Say, function X is running.
> Interrupt arrives, and the interrupt handlers calls function X again. So, if
> that "red zone" is some static memory area, the values there are
> overwritten. Am I missing something? Should there be some "interrupts safe"
> mode?

For ABIs with red zones, the signal is delivered some offset below the 
stack pointer, making it okay to use it.

-Chris

-- 
http://nondot.org/sabre/
http://llvm.org/