[PATCH] D24104: Make GlobalsAA ignore dead constant expressions.

[Friedman, Eli via llvm-commits]

On 9/11/2016 10:09 PM, Chandler Carruth wrote:
> On Thu, Sep 1, 2016 at 10:21 AM Friedman, Eli via llvm-commits 
> <llvm-commits at lists.llvm.org <mailto:llvm-commits at lists.llvm.org>> wrote:
>
>     On 8/31/2016 4:19 PM, Daniel Berlin wrote:
>     > In theory yes, in practice, it probably shouldn't.
>     >
>     > In particular, i would never expect *this* pass to modify the IR.
>     >
>     > Is there some reason things don't get cleaned up that means this
>     pass
>     > has to clean them up?
>
>     This isn't really modifying the IR: dead Constants aren't actually
>     part
>     of the IR in the sense that they're not relevant to the semantics
>     of IR,
>     and they're implicitly discarded if you dump the IR to a file.
>
>     Existing passes don't clean up dead Constants simply because it isn't
>     convenient; the standard accessors for modifying and erasing
>     instructions don't clean them up because it would lead to
>     use-after-free, and passes generally don't clean them up because it
>     would be mostly useless boilerplate.
>
>     I can come up with some other solution if you want to insist on the
>     invariant that GlobalsAA isn't allowed to modify any data structure
>     associated with the module, but I don't really see a need.
>
>
> I really don't think we should mutate the set of constants in an analysis.
>
> I think it is reasonable for other analyses to hold pointers to 
> constants and expect them to never go away even if they become "dead". 
> I suspect people view constants much like types -- once they get a 
> pointer to them, the pointer remains immutable and reliable. I would 
> want a *really* compelling motivation before we remove that invariant.

We already have a bunch of passes which call removeDeadConstantUsers... 
they're not analysis passes, but I don't think that makes much of a 
difference in practice.

That said, it's probably not a big deal to come up with a version which 
uses Constant::isConstantUsed() instead; I'll try to post a new version 
soon.

-Eli

-- 
Employee of Qualcomm Innovation Center, Inc.
Qualcomm Innovation Center, Inc. is a member of Code Aurora Forum, a Linux Foundation Collaborative Project

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20160912/c4d1af88/attachment.html>