[llvm-dev] Bug in pass 'ipsccp' on function attribute 'argmemonly'? (original) (raw)

Alexandre Isoard via llvm-dev llvm-dev at lists.llvm.org
Tue Jul 14 14:28:32 PDT 2020

I think here, I would fix ipsccp.

It transform an access of a function argument based on that argument into an access of a function argument not based on that argument.

While one could argue it is still argmemonly (the documentation is ambiguous on that) it makes it significantly harder to trace that correlation for deadargelim or any form of alias analysis.

In general I wonder if we would not benefit from an attribute that list the globals that a function may access, and having argmemonly include those globals in what it consider "arg".

On Tue, Jul 14, 2020, 09:29 Fangqing Du via llvm-dev < llvm-dev at lists.llvm.org> wrote:

Thank you Hal and Stefan!

Bug report is filed: https://bugs.llvm.org/showbug.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? Thanks, Fangqing On Sat, Jul 11, 2020 at 5:12 AM Hal Finkel <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/ ? 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) localunnamedaddr #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) localunnamedaddr #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 listllvm-dev at lists.llvm.orghttps://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/b7984b32/attachment.html>

More information about the llvm-dev mailing list