[llvm-dev] Bug in pass 'ipsccp' on function attribute 'argmemonly'?

Johannes Doerfert via llvm-dev llvm-dev at lists.llvm.org
Tue Jul 14 17:00:39 PDT 2020


On 7/14/20 4:34 PM, Hal Finkel via llvm-dev wrote:
>
> On 7/14/20 11:28 AM, Fangqing Du wrote:
>> Thank you Hal and Stefan!
>>
>> Bug report is filed: https://bugs.llvm.org/show_bug.cgi?id=46717 
>> <https://bugs.llvm.org/show_bug.cgi?id=46717>
>>
>> And Stefan,
>> I think 'attributor' is a really nice pass, and can infer more 
>> precise and useful attributes, which may give more opportunities for 
>> optimization.
>> But we shouldn't depend on 'attributor' to correct wrong function 
>> attributes, because we cannot run 'attributor' after every pass, right?
>
> Correct. Each pass must be correct on its own. We need to fix ipsccp.
>
I don't think that was ever questioned ;)

I think the comment about the Attributor was more to show that this behavior

is not the same there which is yet another good indicator this is a bug.


~ Johannes


>  -Hal
>
>
>>
>> Thanks,
>> Fangqing
>>
>> On Sat, Jul 11, 2020 at 5:12 AM Hal Finkel <hfinkel at anl.gov 
>> <mailto:hfinkel at anl.gov>> wrote:
>>
>>     Hi, Fangqing,
>>
>>     Yes, this seems like a bug. Can you please file a bug report at
>>     https://bugs.llvm.org/ <https://bugs.llvm.org/> ? If you need
>>     someone else to do this on your behalf, please let us know.
>>
>>      -Hal
>>
>>     On 7/10/20 9:04 PM, Fangqing Du via llvm-dev wrote:
>>>     Hi all,
>>>
>>>     Let's see an example (from Alexandre Isoard) first:
>>> ****************************************************************************************
>>>     ; RUN: opt -ipsccp -deadargelim -licm -O2 -S < %s
>>>
>>>     @g = internal global i32 0
>>>
>>>     ; Function Attrs: argmemonly
>>>     define internal void @foo(i32* nonnull dereferenceable(4) %arg,
>>>     i32 %val) #0 {
>>>     entry:
>>>       store i32 %val, i32* %arg
>>>       ret void
>>>     }
>>>
>>>     define i32 @bar(i32 %n) {
>>>     entry:
>>>       store i32 1, i32* @g
>>>       br label %loop
>>>
>>>     loop: ; preds = %bb1, %bb
>>>     %i = phi i32 [ %i.inc, %loop ], [ 0, %entry ]
>>>     %g.val = load i32, i32* @g
>>>     %g.inc = add nuw i32 %g.val, 1
>>>     tail call void @foo(i32* @g, i32 %g.inc)
>>>     %i.inc = add nuw i32 %i, 1
>>>     %cond = icmp eq i32 %i.inc, %n
>>>       br i1 %cond, label %exit, label %loop
>>>
>>>     exit: ; preds = %bb1
>>>       ret i32 %g.val
>>>     }
>>>
>>>     declare void @llvm.assume(i1)
>>>
>>>     attributes #0 = { argmemonly }
>>> ****************************************************************************************
>>>
>>>     With opt cmd '-ipsccp -deadargelim -licm -O2', function @bar will
>>>     return constant value 1, instead of correct value %n. This is
>>>     crazy, right?
>>>     Let's see what happens here.
>>>     Due to pass 'ipsccp' replaced the argument of function '@foo'
>>>     with global variable '@g', but keeps attr 'argmemonly' there, so
>>>     pass 'licm' will hoist the load of '@g' before the loop (as the
>>>     value of @g isn't changed, but it is changed), and will cause
>>>     function return wrong constant value '1', instead of '%n' (the
>>>     correct value) , like following:
>>> ****************************************************************************************
>>>
>>>     ; Function Attrs: nofree norecurse nounwind writeonly
>>>     define i32 @bar(i32 %n) local_unnamed_addr #0 {
>>>     entry:
>>>     ret i32 1
>>>     }
>>>
>>>     attributes #0 = { nofree norecurse nounwind writeonly }
>>> ****************************************************************************************
>>>
>>>
>>>     And if remove the 'argmemonly' attribute, function @bar will
>>>     return the correct value:
>>> ****************************************************************************************
>>>
>>>     ; Function Attrs: nofree norecurse nounwind
>>>     define i32 @bar(i32 %n) local_unnamed_addr #0 {
>>>     entry:
>>>     ret i32 %n
>>>     }
>>> ****************************************************************************************
>>>
>>>
>>>     So if function attribute 'argmemonly' on function @foo is
>>>     correct, then there's a bug in pass 'ipsccp'. When it replaces
>>>     the function local value with global variable, then it shoud
>>>     remember to remove this function attribute.
>>>
>>>     Thanks,
>>>     Fangqing
>>>
>>>     _______________________________________________
>>>     LLVM Developers mailing list
>>>     llvm-dev at lists.llvm.org <mailto:llvm-dev at lists.llvm.org>
>>>     https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev 
>>> <https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev>
>>
>>     --     Hal Finkel
>>     Lead, Compiler Technology and Programming Languages
>>     Leadership Computing Facility
>>     Argonne National Laboratory
>>
>>
>> _______________________________________________
>> LLVM Developers mailing list
>> llvm-dev at lists.llvm.org
>> https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20200714/fd2f1356/attachment-0001.html>


More information about the llvm-dev mailing list