llvm dev - Aug 2009 - [LLVMdev] AliasSetTracker

I started using the AliasSetTracker, but I have a problem.

I am using -basicaa -anders-aa and optimizations are enabled for llvm-gcc
and llvm-ld.
I use an instruction visitor to populate the alias set tracker, but two
pointers end up in different alias sets.
While investigating this I notice at one point that if I call:

bool New = false;
AliasSet &X = AST->getAliasSetForPointer(Pointer, Size, &New);

After this call, New is true
but AST->containsPointer(Pointer, Size) is false! Shouldn't it be true at
this point?

If I call again getAliasSetForPointer after resetting New, it remains false,
as I expect it after the first call.

But AST->getAliasSetForPointerIfExists(Pointer, Size) returns null. Again,
shouldn't it return the same alias set as the first call?

Am I misunderstanding something? This seems to work without optimizations or
without -anders-aa.

Using revision 79271

Best regards,
Marcus
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20090818/020555f3/attachment.html>

On Aug 17, 2009, at 4:16 PM, Marcus Hicks wrote:
> I started using the AliasSetTracker, but I have a problem.
Hi Marcus,

At best, anders-aa is beta-quality.  It probably has several bugs and  
is not really ready to be widely used.

-Chris
>
> I am using -basicaa -anders-aa and optimizations are enabled for  
> llvm-gcc and llvm-ld.
> I use an instruction visitor to populate the alias set tracker, but  
> two pointers end up in different alias sets.
> While investigating this I notice at one point that if I call:
>
> bool New = false;
> AliasSet &X = AST->getAliasSetForPointer(Pointer, Size, &New);
>
> After this call, New is true
> but AST->containsPointer(Pointer, Size) is false! Shouldn't it be  
> true at this point?
>
> If I call again getAliasSetForPointer after resetting New, it  
> remains false, as I expect it after the first call.
>
> But AST->getAliasSetForPointerIfExists(Pointer, Size) returns null.  
> Again, shouldn't it return the same alias set as the first call?
>
> Am I misunderstanding something? This seems to work without  
> optimizations or without -anders-aa.
>
> Using revision 79271
>
> Best regards,
> Marcus
> _______________________________________________
> LLVM Developers mailing list
> LLVMdev at cs.uiuc.edu         http://llvm.cs.uiuc.edu
> http://lists.cs.uiuc.edu/mailman/listinfo/llvmdev

On Tue, Aug 18, 2009 at 7:48 AM, Chris Lattner <clattner at apple.com>
wrote:
> On Aug 17, 2009, at 4:16 PM, Marcus Hicks wrote:
>
>  I started using the AliasSetTracker, but I have a problem.
>>
>
> Hi Marcus,
>
> At best, anders-aa is beta-quality.  It probably has several bugs and is
> not really ready to be widely used.
>
Thanks, I'll keep that in mind.

One thing I notice is that pointers like
extern char *p;
extern char *q;
don't alias with each other.

This seems to be a bug, as the code that loops over globals in Andersens
does make these point to a universal set,
so their points-to should intersect.

Eventually I want all these externals to end up in the same AliasSet, and
then I need a way to check which alias set includes the universal set.

Anyway, I copied Andersens into my project and will try to fix these.

> -Chris
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20090822/3683579a/attachment.html>