[PATCH] D29989: [BasicAA] Take attributes into account when requesting modref info for a call site

[Daniel Berlin via llvm-commits]

On Mon, Feb 20, 2017 at 9:09 AM, Igor Laevsky via Phabricator <
reviews at reviews.llvm.org> wrote:

> igor-laevsky added inline comments.
>
>
> ================
> Comment at: lib/Analysis/BasicAliasAnalysis.cpp:886
>    // The AAResultBase base class has some smarts, lets use them.
> -  return AAResultBase::getModRefInfo(CS, Loc);
> +  return static_cast<ModRefInfo>(Result & AAResultBase::getModRefInfo(CS,
> Loc));
>  }
> ----------------
> dberlin wrote:
> > Uh?
> Idea here is to intersect smarts from AAResultBase with the result we've
> received in this function. I'm actually not certain if this is needed since
> "AAResults" does intersect results from all alias analyses. Plus
> AAResultBase currently doesn't have any smarts. So, yes, I'm not sure if we
> need this operation.
>
>
> ================
> Comment at: test/Transforms/NewGVN/readattrs.ll:1-3
> -; XFAIL: *
>  ; RUN: opt -newgvn -S -o - < %s | FileCheck %s
>
> ----------------
> davide wrote:
> > The fact this test passes now is actually good, but I'd like to
> understand why it was failing before and now it's not. Can you please
> elaborate?
> This test passes local non-escaping object as a readonly nocapture
> function argument. Before my change alias analysis would have ignored
> readonly argument and concluded that call modifies "%a". I suspect that old
> GVN might use some internal smarts to determine that "%a" is unmodified by
> the call, while NewGVN relies on the alias analysis.
>

GVN uses memdep, which uses getmodrefinfo, the same as memoryssa (which
newgvn uses).

I am going to analyze this further, as i believe it should already work as
well in both cases ..


>
>
> https://reviews.llvm.org/D29989
>
>
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20170220/8b6640e6/attachment.html>