[PATCH] R600/SI: Fix losing chain when fixing reg class of loads.

[Tom Stellard]

On Tue, Sep 09, 2014 at 09:08:30PM +0000, Matt Arsenault wrote:
> The lost chain resulting in earlier side effecting nodes being deleted.
> 

LGTM.
> http://reviews.llvm.org/D5277
> 
> Files:
>   lib/Target/R600/SIISelLowering.cpp
>   test/CodeGen/R600/missing-store.ll

> Index: lib/Target/R600/SIISelLowering.cpp
> ===================================================================
> --- lib/Target/R600/SIISelLowering.cpp
> +++ lib/Target/R600/SIISelLowering.cpp
> @@ -2004,12 +2004,20 @@
>        return N;
>      }
>      ConstantSDNode *Offset = cast<ConstantSDNode>(N->getOperand(1));
> -    SDValue Ops[] = {
> -      SDValue(DAG.getMachineNode(AMDGPU::SI_ADDR64_RSRC, DL, MVT::i128,
> -                                 DAG.getConstant(0, MVT::i64)), 0),
> -      N->getOperand(0),
> -      DAG.getConstant(Offset->getSExtValue() << 2, MVT::i32)
> -    };
> +    MachineSDNode *RSrc = DAG.getMachineNode(AMDGPU::SI_ADDR64_RSRC, DL,
> +                                             MVT::i128,
> +                                             DAG.getConstant(0, MVT::i64));
> +
> +    SmallVector<SDValue, 8> Ops;
> +    Ops.push_back(SDValue(RSrc, 0));
> +    Ops.push_back(N->getOperand(0));
> +    Ops.push_back(DAG.getConstant(Offset->getSExtValue() << 2, MVT::i32));
> +
> +    // Copy remaining operands so we keep any chain and glue nodes that follow
> +    // the normal operands.
> +    for (unsigned I = 2, E = N->getNumOperands(); I != E; ++I)
> +      Ops.push_back(N->getOperand(I));
> +
>      return DAG.getMachineNode(NewOpcode, DL, N->getVTList(), Ops);
>    }
>    }
> Index: test/CodeGen/R600/missing-store.ll
> ===================================================================
> --- /dev/null
> +++ test/CodeGen/R600/missing-store.ll
> @@ -0,0 +1,26 @@
> +; RUN: llc -march=r600 -mcpu=SI -verify-machineinstrs < %s | FileCheck -check-prefix=FUNC -check-prefix=SI %s
> +
> + at ptr_load = addrspace(3) global i32 addrspace(2)* null, align 8
> +
> +; Make sure when the load from %ptr2 is folded the chain isn't lost,
> +; resulting in losing the store to gptr
> +
> +; FUNC-LABEL: @missing_store_reduced
> +; SI: DS_READ_B64
> +; SI: BUFFER_STORE_DWORD
> +; SI: BUFFER_LOAD_DWORD
> +; SI: BUFFER_STORE_DWORD
> +; SI: S_ENDPGM
> +define void @missing_store_reduced(i32 addrspace(1)* %out, i32 addrspace(1)* %gptr) #0 {
> +  %ptr0 = load i32 addrspace(2)* addrspace(3)* @ptr_load, align 8
> +  %ptr2 = getelementptr inbounds i32 addrspace(2)* %ptr0, i64 2
> +
> +  store i32 99, i32 addrspace(1)* %gptr, align 4
> +  %tmp2 = load i32 addrspace(2)* %ptr2, align 4
> +
> +  store i32 %tmp2, i32 addrspace(1)* %out, align 4
> +  ret void
> +}
> +
> +attributes #0 = { nounwind }
> +

> _______________________________________________
> llvm-commits mailing list
> llvm-commits at cs.uiuc.edu
> http://lists.cs.uiuc.edu/mailman/listinfo/llvm-commits