[llvm] r226945 - R600/SI: Move i64 -> v2i32 load promotion into AMDGPUDAGToDAGISel::Select()
Hans Wennborg
hans at chromium.org
Mon Jan 26 09:27:25 PST 2015
On Mon, Jan 26, 2015 at 7:57 AM, Tom Stellard <tom at stellard.net> wrote:
> Hi Hans,
>
> Is this OK to merge to the 3.6 branch? It fixes and infinite loop
> when legalizing misaligned loads.
>
> I am the code owner and I approve this patch.
Yes, go ahead.
Thanks,
Hans
> On Fri, Jan 23, 2015 at 10:05:45PM -0000, Tom Stellard wrote:
>> Author: tstellar
>> Date: Fri Jan 23 16:05:45 2015
>> New Revision: 226945
>>
>> URL: http://llvm.org/viewvc/llvm-project?rev=226945&view=rev
>> Log:
>> R600/SI: Move i64 -> v2i32 load promotion into AMDGPUDAGToDAGISel::Select()
>>
>> We used to do this promotion during DAG legalization, but this
>> caused an infinite loop in ExpandUnalignedLoad() because it assumed
>> that i64 loads were legal if i64 was a legal type.
>>
>> It also seems better to report i64 loads as legal, since they actually
>> are and we were just promoting them to simplify our tablegen files.
>>
>> Added:
>> llvm/trunk/test/CodeGen/R600/misaligned-load.ll
>> Modified:
>> llvm/trunk/lib/Target/R600/AMDGPUISelDAGToDAG.cpp
>> llvm/trunk/lib/Target/R600/AMDGPUISelLowering.cpp
>>
>> Modified: llvm/trunk/lib/Target/R600/AMDGPUISelDAGToDAG.cpp
>> URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Target/R600/AMDGPUISelDAGToDAG.cpp?rev=226945&r1=226944&r2=226945&view=diff
>> ==============================================================================
>> --- llvm/trunk/lib/Target/R600/AMDGPUISelDAGToDAG.cpp (original)
>> +++ llvm/trunk/lib/Target/R600/AMDGPUISelDAGToDAG.cpp Fri Jan 23 16:05:45 2015
>> @@ -417,6 +417,28 @@ SDNode *AMDGPUDAGToDAGISel::Select(SDNod
>> N->getValueType(0), Ops);
>> }
>>
>> + case ISD::LOAD: {
>> + // To simplify the TableGen patters, we replace all i64 loads with
>> + // v2i32 loads. Alternatively, we could promote i64 loads to v2i32
>> + // during DAG legalization, however, so places (ExpandUnalignedLoad)
>> + // in the DAG legalizer assume that if i64 is legal, so doing this
>> + // promotion early can cause problems.
>> + EVT VT = N->getValueType(0);
>> + LoadSDNode *LD = cast<LoadSDNode>(N);
>> + if (VT != MVT::i64 || LD->getExtensionType() != ISD::NON_EXTLOAD)
>> + break;
>> +
>> + SDValue NewLoad = CurDAG->getLoad(MVT::v2i32, SDLoc(N), LD->getChain(),
>> + LD->getBasePtr(), LD->getMemOperand());
>> + SDValue BitCast = CurDAG->getNode(ISD::BITCAST, SDLoc(N),
>> + MVT::i64, NewLoad);
>> + CurDAG->ReplaceAllUsesOfValueWith(SDValue(N, 1), NewLoad.getValue(1));
>> + CurDAG->ReplaceAllUsesOfValueWith(SDValue(N, 0), BitCast);
>> + SelectCode(NewLoad.getNode());
>> + N = BitCast.getNode();
>> + break;
>> + }
>> +
>> case AMDGPUISD::REGISTER_LOAD: {
>> if (ST.getGeneration() <= AMDGPUSubtarget::NORTHERN_ISLANDS)
>> break;
>>
>> Modified: llvm/trunk/lib/Target/R600/AMDGPUISelLowering.cpp
>> URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Target/R600/AMDGPUISelLowering.cpp?rev=226945&r1=226944&r2=226945&view=diff
>> ==============================================================================
>> --- llvm/trunk/lib/Target/R600/AMDGPUISelLowering.cpp (original)
>> +++ llvm/trunk/lib/Target/R600/AMDGPUISelLowering.cpp Fri Jan 23 16:05:45 2015
>> @@ -189,9 +189,6 @@ AMDGPUTargetLowering::AMDGPUTargetLoweri
>> setOperationAction(ISD::LOAD, MVT::v2f32, Promote);
>> AddPromotedToType(ISD::LOAD, MVT::v2f32, MVT::v2i32);
>>
>> - setOperationAction(ISD::LOAD, MVT::i64, Promote);
>> - AddPromotedToType(ISD::LOAD, MVT::i64, MVT::v2i32);
>> -
>> setOperationAction(ISD::LOAD, MVT::v4f32, Promote);
>> AddPromotedToType(ISD::LOAD, MVT::v4f32, MVT::v4i32);
>>
>>
>> Added: llvm/trunk/test/CodeGen/R600/misaligned-load.ll
>> URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/test/CodeGen/R600/misaligned-load.ll?rev=226945&view=auto
>> ==============================================================================
>> --- llvm/trunk/test/CodeGen/R600/misaligned-load.ll (added)
>> +++ llvm/trunk/test/CodeGen/R600/misaligned-load.ll Fri Jan 23 16:05:45 2015
>> @@ -0,0 +1,18 @@
>> +; RUN: llc -march=amdgcn -mcpu=SI -verify-machineinstrs < %s | FileCheck -check-prefix=SI %s
>> +
>> +; SI: @byte_aligned_load64
>> +; SI: ds_read_u8
>> +; SI: ds_read_u8
>> +; SI: ds_read_u8
>> +; SI: ds_read_u8
>> +; SI: ds_read_u8
>> +; SI: ds_read_u8
>> +; SI: ds_read_u8
>> +; SI: ds_read_u8
>> +; SI: s_endpgm
>> +define void @byte_aligned_load64(i64 addrspace(1)* %out, i64 addrspace(3)* %in) {
>> +entry:
>> + %0 = load i64 addrspace(3)* %in, align 1
>> + store i64 %0, i64 addrspace(1)* %out
>> + ret void
>> +}
>>
>>
>> _______________________________________________
>> llvm-commits mailing list
>> llvm-commits at cs.uiuc.edu
>> http://lists.cs.uiuc.edu/mailman/listinfo/llvm-commits
More information about the llvm-commits
mailing list