Stack layout in the x86 back-end

Project Infrastructure LLVM Dev List Archives
1

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

2

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

3

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