[llvm-dev] Use case of undef

[Sean Silva via llvm-dev]

On Sun, Dec 4, 2016 at 3:22 PM, LYU YH <lyh.kernel at gmail.com> wrote:

> Hello Sean,
>
> I read several LLVM undef materials[1][2][3] but still have some questions
> of it. I can understand "add %X, undef -> undef". But why the equation "and
> %X, undef -> 0" is valid? In LLVM Language Reference Manual, the
> explanation of such equation is that it is safe to assume that all bits of
> the ‘undef‘ could be 0. If so, why can't we assume all bits of the 'undef'
> is 0 in the former equation. Therefore we can get "add %X, 0 -> %X" instead?
>

This is a good question and shows that you have a fairly good understanding
already. "add %X, undef -> %X" is legal as you predict. The reason it isn't
done is that  "add %X, undef -> undef" is a stronger simplification since
it propagates the undef.

To go further, I can recommend the thread "[llvm-dev] RFC: Killing undef
and spreading poison". Nuno also did a presentation at the latest LLVM
developer meeting:
http://llvm.org/devmtg/2016-11/Slides/Lopes-LongLivePoison.pdf


>
> By the way, is there any document that describe how 'undef' should be
> handled with other LLVM value? Is the 'undef' in LLVM the same as the
> UNDEF[4] (the top element in lattice) in constant propagation?
>

No, the UNDEF lattice value in constant propagation (and any other lattice
framework) basically just means "I don't know anything about the set of
dynamic values that this value can take".

-- Sean Silva


>
> [1] http://www.nondot.org/sabre/LLVMNotes/UndefinedValue.txt
> [2] https://www.cs.utah.edu/~regehr/llvm-ub.pdf
> [3] http://sunfishcode.github.io/blog/2014/07/14/undef-introduction.html
> [4] http://infolab.stanford.edu/~ullman/dragon/w06/lectures/cp.pdf
>
> Thanks for your reply
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20161204/3ac94bc9/attachment.html>