Mikael Holmén via llvm-dev
2016-Jan-22 12:58 UTC
[llvm-dev] Return value from TargetLowering::LowerOperation?
Hi, I'm a litle bit puzzled by the TargetLowering::LowerOperation function, and what different callers of this function assumes about the returned value. In several places it seems like it is assumed that LowerOperation can return three kinds of values: * Something completely new. * SDValue() * The same SDValue as LowerOperation was called on. However in some places, e.g. in TargetLowering::LowerOperationWrapper, it seems like it is assumed only the first two of the above cases can happen because there we do: SDValue Res = LowerOperation(SDValue(N, 0), DAG); if (Res.getNode()) Results.push_back(Res); So, when LowerOperationWrapper is called from DAGTypeLegalizer::CustomLowerNode, and my target's LowerOperation returns the same SDValue as it was called on since we don't want to do anything special with this particular SDValue, then we get an assert "Potential legalization loop!" when DAGTypeLegalizer::CustomLowerNode tries to replace the SDValue with itself. So in some cases it's ok for LowerOperation to return the same SDValue it was called on, and sometimes it's not? Can anyone shed some light on this? Thanks, Mikael
Tom Stellard via llvm-dev
2016-Jan-22 16:02 UTC
[llvm-dev] Return value from TargetLowering::LowerOperation?
On Fri, Jan 22, 2016 at 01:58:49PM +0100, Mikael Holmén via llvm-dev wrote:> Hi, > > I'm a litle bit puzzled by the TargetLowering::LowerOperation function, > and what different callers of this function assumes about the returned > value. >SelectionDAGLegalize::LegalizeOp() is your best reference for this.> In several places it seems like it is assumed that LowerOperation can > return three kinds of values: > * Something completely new. > * SDValue()This tells the legalizer to treat the node as if it were marked Expand. For loads and stores it tells the legalizer to treat the node as if it were marked Promote.> * The same SDValue as LowerOperation was called on.This tells the legalizer that the input node is legal.> > However in some places, e.g. in TargetLowering::LowerOperationWrapper, > it seems like it is assumed only the first two of the above cases can > happen because there we do: > > SDValue Res = LowerOperation(SDValue(N, 0), DAG); > if (Res.getNode()) > Results.push_back(Res); > > So, when LowerOperationWrapper is called from > DAGTypeLegalizer::CustomLowerNode, and my target's LowerOperation > returns the same SDValue as it was called on since we don't want to do > anything special with this particular SDValue, then we get an assert > "Potential legalization loop!" when DAGTypeLegalizer::CustomLowerNode > tries to replace the SDValue with itself. >I think this error can only happen during type legalization, so my guess is that you are returning a node that has an illegal type. Can you provide more information about the node this is failing with? -Tom> So in some cases it's ok for LowerOperation to return the same SDValue > it was called on, and sometimes it's not? > > Can anyone shed some light on this? >> Thanks, > Mikael > _______________________________________________ > LLVM Developers mailing list > llvm-dev at lists.llvm.org > http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev
Matt Arsenault via llvm-dev
2016-Jan-22 18:27 UTC
[llvm-dev] Return value from TargetLowering::LowerOperation?
> On Jan 22, 2016, at 08:02, Tom Stellard via llvm-dev <llvm-dev at lists.llvm.org> wrote: > >> * The same SDValue as LowerOperation was called on. > > This tells the legalizer that the input node is legal.I thought SDValue() does this. The same op is supposed to defer to the generic legalization -------------- next part -------------- An HTML attachment was scrubbed... URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20160122/6102e61b/attachment.html>
Mikael Holmén via llvm-dev
2016-Jan-25 07:29 UTC
[llvm-dev] Return value from TargetLowering::LowerOperation?
Hi, On 01/22/2016 05:02 PM, Tom Stellard wrote:> On Fri, Jan 22, 2016 at 01:58:49PM +0100, Mikael Holmén via llvm-dev wrote: >> Hi, >> >> I'm a litle bit puzzled by the TargetLowering::LowerOperation function, >> and what different callers of this function assumes about the returned >> value. >> > SelectionDAGLegalize::LegalizeOp() is your best reference for this. > >> In several places it seems like it is assumed that LowerOperation can >> return three kinds of values: >> * Something completely new. >> * SDValue() > > This tells the legalizer to treat the node as if it were marked Expand. > For loads and stores it tells the legalizer to treat the node as if > it were marked Promote. > >> * The same SDValue as LowerOperation was called on. > > This tells the legalizer that the input node is legal. > >> >> However in some places, e.g. in TargetLowering::LowerOperationWrapper, >> it seems like it is assumed only the first two of the above cases can >> happen because there we do: >> >> SDValue Res = LowerOperation(SDValue(N, 0), DAG); >> if (Res.getNode()) >> Results.push_back(Res); >> >> So, when LowerOperationWrapper is called from >> DAGTypeLegalizer::CustomLowerNode, and my target's LowerOperation >> returns the same SDValue as it was called on since we don't want to do >> anything special with this particular SDValue, then we get an assert >> "Potential legalization loop!" when DAGTypeLegalizer::CustomLowerNode >> tries to replace the SDValue with itself. >> > > I think this error can only happen during type legalization, so my guess > is that you are returning a node that has an illegal type. Can you > provide more information about the node this is failing with?On my target v2i16, v4i16, v2i32, v4i32, v2f32, v4f32 are legal and all other vector types are not. Vectors of i16 are a bit special and we need to custom lower bitcasts to/from them. Therefore we do setOperationAction(ISD::BITCAST, VT, Custom); on all MVT:s, and in our target's LowerOperation/LowerBitcast we specifically handle when the source or target type is v2i16 or v4i16, and for other cases we just return the input SDValue and let the "normal" code handle it in whatever way it see fits. In this particular case, when it crashes, we have a bitcast from v2i64 to v4i32: t70: v4i32 = bitcast t27 and t27 is t27: v2i64 = or t15, t26 so it's a bitcast from v2i64 to v4i32 that we don't want to do anything special with so we return the input SDValue which TargetLowering::LowerOperationWrapper returns back to DAGTypeLegalizer::CustomLowerNode and then we get the assert in DAGTypeLegalizer::ReplaceValueWith. Thanks, Mikael> > -Tom > >> So in some cases it's ok for LowerOperation to return the same SDValue >> it was called on, and sometimes it's not? >> >> Can anyone shed some light on this? >> > > > > >> Thanks, >> Mikael >> _______________________________________________ >> LLVM Developers mailing list >> llvm-dev at lists.llvm.org >> http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev >
Apparently Analagous Threads
- Return value from TargetLowering::LowerOperation?
- Handling node through TargetLowering::LowerOperation vs TargetLowering::ReplaceNodeResults
- Handling node through TargetLowering::LowerOperation vs TargetLowering::ReplaceNodeResults
- [LLVMdev] r114523 (convert the last 4 X86ISD...) breaks clang
- [LLVMdev] r114523 (convert the last 4 X86ISD...) breaks clang