Page MenuHomePhabricator

[mlir] Relax sign extend to work with signed or signless integers.
AbandonedPublic

Authored by nicolasvasilache on Nov 27 2020, 8:16 AM.

Details

Summary

SignExtendIOp only allows Signless operands but it seems reasonable to also allow
signed operands.
This revision relaxes the semantics.

Diff Detail

Event Timeline

nicolasvasilache requested review of this revision.Nov 27 2020, 8:16 AM
rriddle requested changes to this revision.Nov 27 2020, 1:23 PM

This is a major shift behind the rationale of the standard dialect, I would like to see an RFC for this. We have an explicit part of the rationale for only using signless integers.

This revision now requires changes to proceed.Nov 27 2020, 1:23 PM

This is a major shift behind the rationale of the standard dialect, I would like to see an RFC for this. We have an explicit part of the rationale for only using signless integers.

See the rationale here: https://mlir.llvm.org/docs/Rationale/Rationale/#integer-signedness-semantics

This has been the rationale since the beginning, so if we change it we should have discussion as to why.

This is a major shift behind the rationale of the standard dialect, I would like to see an RFC for this. We have an explicit part of the rationale for only using signless integers.

See the rationale here: https://mlir.llvm.org/docs/Rationale/Rationale/#integer-signedness-semantics

This has been the rationale since the beginning, so if we change it we should have discussion as to why.

Thanks, I added a post about this to the discussion in https://llvm.discourse.group/t/rfc-vector-dialects-neon-and-sve/2284/7 which triggered this PR.

nicolasvasilache abandoned this revision.Nov 30 2020, 1:14 AM