[llvm-dev] GISel using deprecated vector_extract for G_EXTRACT_VECTOR_ELT mapping

[Björn Pettersson A via llvm-dev]

> -----Original Message-----
> From: Daniel Sanders <daniel_l_sanders at apple.com>
> Sent: den 16 december 2020 04:13
> To: Björn Pettersson A <bjorn.a.pettersson at ericsson.com>
> Cc: llvm-dev <llvm-dev at lists.llvm.org>
> Subject: Re: [llvm-dev] GISel using deprecated vector_extract for
> G_EXTRACT_VECTOR_ELT mapping
> 
> 
> 
> > On Dec 9, 2020, at 06:51, Björn Pettersson A via llvm-dev <llvm-
> dev at lists.llvm.org> wrote:
> >
> > According to code comment the definition of vector_extract has
> > been deprecated for ~5 years (at least since Dec 11, 2015).
> >
> > Seems like some targets still are using it, and even the GlobalISel
> > implementation is supporting this deprecated functionality by
> > adding a mapping between G_EXTRACT_VECTOR_ELT and vector_extract
> > in SlectionDAGCompat.td.
> 
> GlobalISel only supports them because the first targets implemented used
> vector_extract

Ok I see.

I did put up a patch here https://reviews.llvm.org/D93416
to replace vector_extract by extractelt in SelectionDAGCompat.td.
That includes adding a complementary mapping for vector_extract for
the AArch target since it still uses the deprecated node type.

> 
> > My target is not using the deprecated vector_extract, and it is using
> > the preferred extractelt definition instead in tablegen patterns.
> >
> > The problem is that there is no similar mapping between
> > extractelt and G_EXTRACT_VECTOR_ELT.
> >
> > Is it allowed to have several GINodeEquiv statements, mapping the
> > same GISel instruction to different SDNodes?
> >
> > (i.e. would it be wrong to also add a mapping between
> > G_EXTRACT_VECTOR_ELT and extractelt)
> 
> Off-hand, I don't think there's a reason that would be a problem so long
> as you only have one or the other in your target. The GINodeEquiv
> relations are there to map operators in the rule like
> vector_extract/extractelt to their GISel counterparts so it can generate
> code in GISel terms. If you somehow have both and they have overlapping
> rules then it would be undefined which rule is used.
> 
> > Or should we stop using the deprecated nodes in GISel mappings?
> >
> > (then I assume someone need to finalize the work of getting rid
> > of vector_extract, at least in targets that support GISel, but
> > after 5 years I think it is about time to either remove the
> > deprecated definition or replace the "deprecated" comment with
> > something that explains why we need both vector_extract and
> > extractelt)
> >
> >
> > PS. The above questions also apply to vector_insert vs insertelt.
> >
> > /Björn
> > _______________________________________________
> > LLVM Developers mailing list
> > llvm-dev at lists.llvm.org
> > https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev