llvm dev - Oct 2007 - [LLVMdev] Stack layout in the x86 back-end

Hi all,

I've been quiet for a while, but work on the new model checker has been 
going pretty well of late. I do have a question for the team, however.

I've implemented code that manages multiple stacks, with an ability to 
version them and roll back and forward between versions so I can 
implement backtracking and nontrivial search strategies -- this all 
works. However, I'm currently dumbly checking the entire stack (not just 
the currently relevant stack frame(s)) for changes, but I'd rather just 
be able to check a minimal area. Whenever I context switch, the code 
under test calls a function in my run-time-system (mcp_yield(), not that 
the details matter much). What I'd like to be able to do is pick up the 
address range of the stack frame of the function that called it -- 
ideally, I'd like a block that encompasses everything from the  function 
parameters up to the current top-of-stack including spilled registers. 
Is this feasible?

Thank you in advance,
Sarah

On Tue, 9 Oct 2007, Sarah Thompson wrote:
> ideally, I'd like a block that encompasses everything from the 
function
> parameters up to the current top-of-stack including spilled registers.
> Is this feasible?
Hi Sarah,

I'm not sure how much of this is possible.  You're guaranteed that each 
function stack frame is itself contiguous (including spilled regs), but 
there isn't really a notion of "top of stack" in the compiler.  To
get
this information, you'll need information from the runtime or OS.

-Chris

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

Hi,

what about the llvm.frameaddress, llvm.stacksave, and llvm.stackrestore 
intrinsics? According to the documentation, they give you the current stack 
frame, including the current top of the frame.
I currently use these to do a simple stack-rollback in our STM, and they seem 
to work okay (I haven't looked at the stack contents in example programs to 
validate this, though). Is this a valid approach for stack rollback (paired 
with a setjmp?

Torvald

On Tuesday 09 October 2007 22:53, Chris Lattner wrote:
> On Tue, 9 Oct 2007, Sarah Thompson wrote:
> > ideally, I'd like a block that encompasses everything from the 
function
> > parameters up to the current top-of-stack including spilled registers.
> > Is this feasible?
>
> Hi Sarah,
>
> I'm not sure how much of this is possible.  You're guaranteed that
each
> function stack frame is itself contiguous (including spilled regs), but
> there isn't really a notion of "top of stack" in the
compiler.  To get
> this information, you'll need information from the runtime or OS.
>
> -Chris