[PATCH] Fix for wrong instcombine on vector insert/extract

[Andrea_DiBiagio at sn.scee.net]

Ping

-- Andrea DiBiagio

Andrea DiBiagio/SN R&D/BS/UK/SCEE wrote on 03/04/2013 13:24:02:

> From: Andrea DiBiagio/SN R&D/BS/UK/SCEE
> To: Andrea DiBiagio/SN R&D/BS/UK/SCEE at SCEE
> Cc: llvm-commits at cs.uiuc.edu
> Date: 03/04/2013 13:24
> Subject: Re: [PATCH] Fix for wrong instcombine on vector insert/extract
>
> Oops. I just noticed that my last email didn't thread correctly.
> I am reattaching an updated patch against the latest revision for
> completeness.
>
> [attachment "InstCombinePatch.diff" deleted by Andrea DiBiagio/SN
> R&D/BS/UK/SCEE]
>
> Sorry for the spam.
> Andrea  Di Biagio
>
> Andrea DiBiagio/SN R&D/BS/UK/SCEE wrote on 03/04/2013 13:06:44:
>
> > From: Andrea DiBiagio/SN R&D/BS/UK/SCEE
> > To: llvm-commits at cs.uiuc.edu
> > Date: 03/04/2013 13:06
> > Subject: Re: [PATCH] Fix for wrong instcombine on vector insert/extract
> >
> > Friendly Ping.
> >
> > Andrea DiBiagio/SN R&D/BS/UK/SCEE wrote on 22/03/2013 19:53:00:
> >
> > > From: Andrea DiBiagio/SN R&D/BS/UK/SCEE
> > > To: llvm-commits at cs.uiuc.edu
> > > Date: 22/03/2013 19:53
> > > Subject: [PATCH] Fix for wrong instcombine on vector insert/extract
> > >
> > > Hello,
> > >
> > > I would like to contribute a fix for a wrong transformation performed
> > > by the Instruction Combiner on sequences of
insertelement/extractelement
> > > instructions.
> > >
> > > When trying to collapse sequences of insertelement/extractelement
> > > instructions into single shuffle instructions, there is one specific
> > > case where the Instruction Combiner wrongly updates the resulting
> > > Mask of shuffle indexes.
> > >
> > > The problem is in function CollectShuffleElments.
> > >
> > > If we have a sequence of insert/extract element instructions
> > > like the one below:
> > >
> > >   %tmp1 = extractelement <4 x float> %LHS, i32 0
> > >   %tmp2 = insertelement <4 x float> %RHS, float %tmp1, i32 1
> > >   %tmp3 = extractelement <4 x float> %RHS, i32 2
> > >   %tmp4 = insertelement <4 x float> %tmp2, float %tmp3, i32 3
> > >
> > > Where:
> > >  . %RHS will have a mask of [4,5,6,7]
> > >  . %LHS will have a mask of [0,1,2,3]
> > >
> > > The Mask of shuffle indexes is wrongly computed to [4,1,6,7]
> > > instead of [4,0,6,7].
> > > When analyzing %tmp2 in order to compute the Mask for the
> > > resulting shuffle instruction, the algorithm forgets to update
> > > the mask index at position 1 with the index associated to the
> > > element extracted from %LHS by instruction %tmp1.
> > >
> > > The patch in attachment fixes the problem with the wrong
> > > propagation of shuffle indexes.
> > > It also adds two extra cases to test
> > > Transforms/InstCombine/vec_shuffle.ll to check that now we
> > > do the correct thing when collapsing sequences of insert/extract
> > > element instruction in that problematic scenario.
> > >
> > > Thanks!
> > > Andrea Di Biagio
> > >