llvm dev - Aug 2014 - [LLVMdev] setAlreadyVectorized does not delete obsolete metadata?

I noticed that LoopVectorizeHints::setAlreadyVectorized never deletes old
"llvm.loop...." metadata.  It just appends more, possibly
contradicting the old metadata.  E.g., after vectorization, a loop previously
marked with llvm.loop.vectorize.width ends up with *two* such annotations, like
this:

    br i1 %exitcond.1, label %for.end.loopexit.unr-lcssa, label %for.body,
!llvm.loop !8
    ...
    !6 = metadata !{metadata !"llvm.loop.vectorize.width", i32 1}
    !7 = metadata !{metadata !"llvm.loop.interleave.count", i32 1}
    !8 = metadata !{metadata !8, metadata !9, metadata !6, metadata !7}
    !9 = metadata !{metadata !"llvm.loop.vectorize.width", i32 4}

!6 and !9 are in the list for loopID !8 and they specify conflicting widths.  Is
this a bug, or is there a deliberate convention that if there are multiple
llvm.loop.vectorize.width annotations, only the last one counts?

- Arch D. Robison
  Intel Corporation

On 13 August 2014 21:16, Robison, Arch <arch.robison at intel.com>
wrote:
> !6 and !9 are in the list for loopID !8 and they specify conflicting
widths.  Is this a bug, or is there a deliberate convention that if there are
multiple llvm.loop.vectorize.width annotations, only the last one counts?
It's a bug.

The vectorizer reads the metadata and writes it back with width = 1 to
stop it from being vectorized again (also avoids wasting time if it
wasn't worth it then), but it should not have two of the same and
especially not two *different* values. It could be a simple bug in the
writing back of the data, and should be simple to fix. Or it could be
that the loop was fused from two other loops with different metadata,
or inlining of code, etc, making it a bit more challenging.

Creating a bug with the appropriate sources and ways to reproduce it
is the best course of action.

Thanks!
-renato

Filed as Bugzilla 20655, with a .ll source file and instructions for reproducing
it.

- Arch

-----Original Message-----
From: Renato Golin [mailto:renato.golin at linaro.org] 
Sent: Wednesday, August 13, 2014 3:31 PM
To: Robison, Arch
Cc: llvmdev at cs.uiuc.edu
Subject: Re: [LLVMdev] setAlreadyVectorized does not delete obsolete metadata?

On 13 August 2014 21:16, Robison, Arch <arch.robison at intel.com>
wrote:
> !6 and !9 are in the list for loopID !8 and they specify conflicting
widths.  Is this a bug, or is there a deliberate convention that if there are
multiple llvm.loop.vectorize.width annotations, only the last one counts?
It's a bug.

The vectorizer reads the metadata and writes it back with width = 1 to stop it
from being vectorized again (also avoids wasting time if it wasn't worth it
then), but it should not have two of the same and especially not two *different*
values. It could be a simple bug in the writing back of the data, and should be
simple to fix. Or it could be that the loop was fused from two other loops with
different metadata, or inlining of code, etc, making it a bit more challenging.

Creating a bug with the appropriate sources and ways to reproduce it is the best
course of action.

Thanks!
-renato