[LLVMdev] passing allocas into tail calls

[Chris Lattner]

I can't think of any way that this would occur, at least by clang compiled code, so we get to choose what it should mean.

B or D make sense to me.

-Chris

> On Oct 27, 2014, at 12:26 PM, Nick Lewycky <nicholas at mxc.ca> wrote:
> 
> I have a LangRef question, what does this code mean:
> 
>  %A = alloca i32
>  tail call @external(i32* %A)
> 
> The 'tail' marker means that the callee doesn't access the caller's stack frame. So what's the right interpretation of this. Is it:
> 
> a)
>  %A = alloca i32
>  unreachable
> 
> b)
>  %A = alloca i32
>  tail call @external(i32* undef)
> 
> c)
>  %A = alloca i32
>  tail call @external(i32* readnone %A)
> 
> d) like (b) except that the 'undef' is really a poison value
> 
> I think the question boils down to two other questions, what does it mean to "access" the caller's stack frame? Is it accessing to look at the pointer but not to dereference it? And the second question is what exactly is the penalty for violating this rule. Does it happen at the call site, in which case conditionally ignoring the argument would still be invalid? Or is it only full UB once you dereference it? What about comparisons performed on such a pointer?
> 
> Nick
> 
> _______________________________________________
> LLVM Developers mailing list
> LLVMdev at cs.uiuc.edu         http://llvm.cs.uiuc.edu
> http://lists.cs.uiuc.edu/mailman/listinfo/llvmdev