R600/SI: Fix unreachable with a sext_in_reg to an illegal type.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@204945 91177308-0d34-0410-b5e6-96231b3b80d8

lib/Target/R600/AMDGPUISelLowering.cpp

@@ -333,6 +333,24 @@ SDValue AMDGPUTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG)
   return Op;
 }
 
+void AMDGPUTargetLowering::ReplaceNodeResults(SDNode *N,
+                                              SmallVectorImpl<SDValue> &Results,
+                                              SelectionDAG &DAG) const {
+  switch (N->getOpcode()) {
+  case ISD::SIGN_EXTEND_INREG:
+    // Different parts of legalization seem to interpret which type of
+    // sign_extend_inreg is the one to check for custom lowering. The extended
+    // from type is what really matters, but some places check for custom
+    // lowering of the result type. This results in trying to use
+    // ReplaceNodeResults to sext_in_reg to an illegal type, so we'll just do
+    // nothing here and let the illegal result integer be handled normally.
+    return;
+
+  default:
+    return;
+  }
+}
+
 SDValue AMDGPUTargetLowering::LowerConstantInitializer(const Constant* Init,
                                                        const GlobalValue *GV,
                                                        const SDValue &InitPtr,

lib/Target/R600/AMDGPUISelLowering.h

@@ -103,6 +103,10 @@ public:
   }
 
   virtual SDValue LowerOperation(SDValue Op, SelectionDAG &DAG) const;
+  virtual void ReplaceNodeResults(SDNode * N,
+                                  SmallVectorImpl<SDValue> &Results,
+                                  SelectionDAG &DAG) const override;
+
   SDValue LowerIntrinsicIABS(SDValue Op, SelectionDAG &DAG) const;
   SDValue LowerIntrinsicLRP(SDValue Op, SelectionDAG &DAG) const;
   SDValue LowerMinMax(SDValue Op, SelectionDAG &DAG) const;

lib/Target/R600/R600ISelLowering.cpp

@@ -762,7 +762,9 @@ void R600TargetLowering::ReplaceNodeResults(SDNode *N,
                                             SmallVectorImpl<SDValue> &Results,
                                             SelectionDAG &DAG) const {
   switch (N->getOpcode()) {
-  default: return;
+  default:
+    AMDGPUTargetLowering::ReplaceNodeResults(N, Results, DAG);
+    return;
   case ISD::FP_TO_UINT: Results.push_back(LowerFPTOUINT(N->getOperand(0), DAG));
     return;
   case ISD::LOAD: {

lib/Target/R600/R600ISelLowering.h

@@ -28,9 +28,9 @@ public:
       MachineBasicBlock * BB) const;
   virtual SDValue LowerOperation(SDValue Op, SelectionDAG &DAG) const;
   virtual SDValue PerformDAGCombine(SDNode *N, DAGCombinerInfo &DCI) const;
-  void ReplaceNodeResults(SDNode * N,
+  virtual void ReplaceNodeResults(SDNode * N,
-      SmallVectorImpl<SDValue> &Results,
+                                  SmallVectorImpl<SDValue> &Results,
-      SelectionDAG &DAG) const;
+                                  SelectionDAG &DAG) const override;
   virtual SDValue LowerFormalArguments(
                                       SDValue Chain,
                                       CallingConv::ID CallConv,

test/CodeGen/R600/sext-in-reg.ll

@@ -1,6 +1,9 @@
 ; RUN: llc < %s -march=r600 -mcpu=SI | FileCheck -check-prefix=SI -check-prefix=FUNC %s
 ; RUN: llc < %s -march=r600 -mcpu=cypress | FileCheck -check-prefix=EG -check-prefix=FUNC %s
 
+declare i32 @llvm.AMDGPU.imax(i32, i32) nounwind readnone
+
+
 ; FUNC-LABEL: @sext_in_reg_i1_i32
 ; SI: S_LOAD_DWORD [[ARG:s[0-9]+]],
 ; SI: V_BFE_I32 [[EXTRACT:v[0-9]+]], [[ARG]], 0, 1
@@ -248,3 +251,21 @@ define void @testcase_3(i8 addrspace(1)* %out, i8 %a) nounwind {
   store i8 %xor, i8 addrspace(1)* %out
   ret void
 }
+
+; FIXME: The BFE should really be eliminated. I think it should happen
+; when computeMaskedBitsForTargetNode is implemented for imax.
+
+; FUNC-LABEL: @sext_in_reg_to_illegal_type
+; SI: BUFFER_LOAD_SBYTE
+; SI: V_MAX_I32
+; SI: V_BFE_I32
+; SI: BUFFER_STORE_SHORT
+define void @sext_in_reg_to_illegal_type(i16 addrspace(1)* nocapture %out, i8 addrspace(1)* nocapture %src) nounwind {
+  %tmp5 = load i8 addrspace(1)* %src, align 1
+  %tmp2 = sext i8 %tmp5 to i32
+  %tmp3 = tail call i32 @llvm.AMDGPU.imax(i32 %tmp2, i32 0) nounwind readnone
+  %tmp4 = trunc i32 %tmp3 to i8
+  %tmp6 = sext i8 %tmp4 to i16
+  store i16 %tmp6, i16 addrspace(1)* %out, align 2
+  ret void
+}