AMDGPU: Fix invalid copies when copying i1 to phys reg

Insert a VReg_1 virtual register so the i1 workaround pass
can handle it.

llvm-svn: 300113

diff --git a/llvm/lib/Target/AMDGPU/AMDGPUISelDAGToDAG.cpp b/llvm/lib/Target/AMDGPU/AMDGPUISelDAGToDAG.cpp
index fc3ea67..318de7f 100644 (file)
--- a/llvm/lib/Target/AMDGPU/AMDGPUISelDAGToDAG.cpp
+++ b/llvm/lib/Target/AMDGPU/AMDGPUISelDAGToDAG.cpp
@@ -563,7 +563,7 @@ void AMDGPUDAGToDAGISel::Select(SDNode *N) {
   case ISD::CopyToReg: {
     const SITargetLowering& Lowering =
       *static_cast<const SITargetLowering*>(getTargetLowering());
-    Lowering.legalizeTargetIndependentNode(N, *CurDAG);
+    N = Lowering.legalizeTargetIndependentNode(N, *CurDAG);
     break;
   }
   case ISD::AND:

diff --git a/llvm/lib/Target/AMDGPU/SIISelLowering.cpp b/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
index eda825d..7268131 100644 (file)
--- a/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
+++ b/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
@@ -4885,8 +4885,33 @@ static bool isFrameIndexOp(SDValue Op) {
 /// \brief Legalize target independent instructions (e.g. INSERT_SUBREG)
 /// with frame index operands.
 /// LLVM assumes that inputs are to these instructions are registers.
-void SITargetLowering::legalizeTargetIndependentNode(SDNode *Node,
-                                                     SelectionDAG &DAG) const {
+SDNode *SITargetLowering::legalizeTargetIndependentNode(SDNode *Node,
+                                                        SelectionDAG &DAG) const {
+  if (Node->getOpcode() == ISD::CopyToReg) {
+    RegisterSDNode *DestReg = cast<RegisterSDNode>(Node->getOperand(1));
+    SDValue SrcVal = Node->getOperand(2);
+
+    // Insert a copy to a VReg_1 virtual register so LowerI1Copies doesn't have
+    // to try understanding copies to physical registers.
+    if (SrcVal.getValueType() == MVT::i1 &&
+        TargetRegisterInfo::isPhysicalRegister(DestReg->getReg())) {
+      SDLoc SL(Node);
+      MachineRegisterInfo &MRI = DAG.getMachineFunction().getRegInfo();
+      SDValue VReg = DAG.getRegister(
+        MRI.createVirtualRegister(&AMDGPU::VReg_1RegClass), MVT::i1);
+
+      SDNode *Glued = Node->getGluedNode();
+      SDValue ToVReg
+        = DAG.getCopyToReg(Node->getOperand(0), SL, VReg, SrcVal,
+                         SDValue(Glued, Glued ? Glued->getNumValues() - 1 : 0));
+      SDValue ToResultReg
+        = DAG.getCopyToReg(ToVReg, SL, SDValue(DestReg, 0),
+                           VReg, ToVReg.getValue(1));
+      DAG.ReplaceAllUsesWith(Node, ToResultReg.getNode());
+      DAG.RemoveDeadNode(Node);
+      return ToResultReg.getNode();
+    }
+  }
 
   SmallVector<SDValue, 8> Ops;
   for (unsigned i = 0; i < Node->getNumOperands(); ++i) {
@@ -4902,6 +4927,7 @@ void SITargetLowering::legalizeTargetIndependentNode(SDNode *Node,
   }
 
   DAG.UpdateNodeOperands(Node, Ops);
+  return Node;
 }
 
 /// \brief Fold the instructions after selecting them.

diff --git a/llvm/lib/Target/AMDGPU/SIISelLowering.h b/llvm/lib/Target/AMDGPU/SIISelLowering.h
index 452ee68..d177777 100644 (file)
--- a/llvm/lib/Target/AMDGPU/SIISelLowering.h
+++ b/llvm/lib/Target/AMDGPU/SIISelLowering.h
@@ -206,7 +206,7 @@ public:
 
   SDValue CreateLiveInRegister(SelectionDAG &DAG, const TargetRegisterClass *RC,
                                unsigned Reg, EVT VT) const override;
-  void legalizeTargetIndependentNode(SDNode *Node, SelectionDAG &DAG) const;
+  SDNode *legalizeTargetIndependentNode(SDNode *Node, SelectionDAG &DAG) const;
 
   MachineSDNode *wrapAddr64Rsrc(SelectionDAG &DAG, const SDLoc &DL,
                                 SDValue Ptr) const;

diff --git a/llvm/test/CodeGen/AMDGPU/inline-asm.ll b/llvm/test/CodeGen/AMDGPU/inline-asm.ll
index 5d49b11..0d7e07b 100644 (file)
--- a/llvm/test/CodeGen/AMDGPU/inline-asm.ll
+++ b/llvm/test/CodeGen/AMDGPU/inline-asm.ll
@@ -196,3 +196,39 @@ entry:
   call void asm sideeffect "; use $0 ", "{VGPR0_VGPR1}"(i64 123456)
   ret void
 }
+
+; CHECK-LABEL: {{^}}i1_imm_input_phys_vgpr:
+; CHECK: v_mov_b32_e32 v0, -1{{$}}
+; CHECK: ; use v0
+define amdgpu_kernel void @i1_imm_input_phys_vgpr() {
+entry:
+  call void asm sideeffect "; use $0 ", "{VGPR0}"(i1 true)
+  ret void
+}
+
+; CHECK-LABEL: {{^}}i1_input_phys_vgpr:
+; CHECK: {{buffer|flat}}_load_ubyte [[LOAD:v[0-9]+]]
+; CHECK: v_and_b32_e32 [[LOAD]], 1, [[LOAD]]
+; CHECK-NEXT: v_cmp_eq_u32_e32 vcc, 1, [[LOAD]]
+; CHECK-NEXT: v_cndmask_b32_e64 v0, 0, -1, vcc
+; CHECK: ; use v0
+define amdgpu_kernel void @i1_input_phys_vgpr() {
+entry:
+  %val = load i1, i1 addrspace(1)* undef
+  call void asm sideeffect "; use $0 ", "{VGPR0}"(i1 %val)
+  ret void
+}
+
+; FIXME: Should be scheduled to shrink vcc
+; CHECK-LABEL: {{^}}i1_input_phys_vgpr_x2:
+; CHECK: v_cmp_eq_u32_e32 vcc, 1, v0
+; CHECK: v_cmp_eq_u32_e64 s[0:1], 1, v1
+; CHECK: v_cndmask_b32_e64 v0, 0, -1, vcc
+; CHECK: v_cndmask_b32_e64 v1, 0, -1, s[0:1]
+define amdgpu_kernel void @i1_input_phys_vgpr_x2() {
+entry:
+  %val0 = load volatile i1, i1 addrspace(1)* undef
+  %val1 = load volatile i1, i1 addrspace(1)* undef
+  call void asm sideeffect "; use $0 $1 ", "{VGPR0}, {VGPR1}"(i1 %val0, i1 %val1)
+  ret void
+}