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.
 -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.
>
> Â -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)Â #0Â {
>> entry:
>> Â Â store i32 %val, i32* %arg
>> Â Â ret void
>> }
>>
>> define i32 @bar(i32 %n)Â {
>> entry:
>> Â Â store i32 1, i32* @g
>> Â Â 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
>> Â Â br i1 %cond, label %exit, label %loop
>>
>> exit: ; preds = %bb1
>> Â Â 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 #0Â {
>> 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 #0Â {
>> 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
>
--
Hal Finkel
Lead, Compiler Technology and Programming Languages
Leadership Computing Facility
Argonne National Laboratory
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20200714/2f453cc3/attachment.html>
RetroSearch is an open source project built by @garambo | Open a GitHub Issue
Search and Browse the WWW like it's 1997 | Search results from DuckDuckGo
HTML:
3.2
| Encoding:
UTF-8
| Version:
0.7.4