[AMDGPU] Fix clamp bit DAG operand

Summary:
- Should use `targetconstant` instead of `constant` operand for clamp
  bit, which is expected as an immediate operand. Under certain
  conditions, such as a common `i1 false` constant is used in other
  place and selected before the instruction with clamp bit, register
  operand may be added instead of immediate one. Use `targetcosntant` to
  enforce that.

Subscribers: arsenm, kzhuravl, jvesely, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye, hiraditya, llvm-commits

Tags: #llvm

Differential Revision: https://reviews.llvm.org/D59608

llvm-svn: 356608

llvm/lib/Target/AMDGPU/AMDGPUISelDAGToDAG.cpp

@@ -931,9 +931,10 @@ void AMDGPUDAGToDAGISel::SelectUADDO_USUBO(SDNode *N) {
   unsigned Opc = N->getOpcode() == ISD::UADDO ?
     AMDGPU::V_ADD_I32_e64 : AMDGPU::V_SUB_I32_e64;
 
-  CurDAG->SelectNodeTo(N, Opc, N->getVTList(),
+  CurDAG->SelectNodeTo(
-                       {N->getOperand(0), N->getOperand(1),
+      N, Opc, N->getVTList(),
-                        CurDAG->getConstant(0, {}, MVT::i1)/*clamp bit*/});
+      {N->getOperand(0), N->getOperand(1),
+       CurDAG->getTargetConstant(0, {}, MVT::i1) /*clamp bit*/});
 }
 
 void AMDGPUDAGToDAGISel::SelectFMA_W_CHAIN(SDNode *N) {
@@ -1041,7 +1042,8 @@ bool AMDGPUDAGToDAGISel::SelectDS1Addr1Offset(SDValue Addr, SDValue &Base,
           unsigned SubOp = AMDGPU::V_SUB_I32_e32;
           if (Subtarget->hasAddNoCarry()) {
             SubOp = AMDGPU::V_SUB_U32_e64;
-            Opnds.push_back(Zero); // clamp bit
+            Opnds.push_back(
+                CurDAG->getTargetConstant(0, {}, MVT::i1)); // clamp bit
           }
 
           MachineSDNode *MachineSub =
@@ -1119,7 +1121,8 @@ bool AMDGPUDAGToDAGISel::SelectDS64Bit4ByteAligned(SDValue Addr, SDValue &Base,
           unsigned SubOp = AMDGPU::V_SUB_I32_e32;
           if (Subtarget->hasAddNoCarry()) {
             SubOp = AMDGPU::V_SUB_U32_e64;
-            Opnds.push_back(Zero); // clamp bit
+            Opnds.push_back(
+                CurDAG->getTargetConstant(0, {}, MVT::i1)); // clamp bit
           }
 
           MachineSDNode *MachineSub

llvm/test/CodeGen/AMDGPU/uaddo.ll

@@ -170,6 +170,28 @@ define amdgpu_kernel void @v_uaddo_v2i32(<2 x i32> addrspace(1)* %out, <2 x i32>
   ret void
 }
 
+; FUNC-LABEL: {{^}}s_uaddo_clamp_bit:
+; GCN: v_add_{{i|u|co_u}}32_e32
+; GCN: s_endpgm
+define amdgpu_kernel void @s_uaddo_clamp_bit(i32 addrspace(1)* %out, i1 addrspace(1)* %carryout, i32 %a, i32 %b) #0 {
+entry:
+  %uadd = call { i32, i1 } @llvm.uadd.with.overflow.i32(i32 %a, i32 %b)
+  %val = extractvalue { i32, i1 } %uadd, 0
+  %carry = extractvalue { i32, i1 } %uadd, 1
+  %c2 = icmp eq i1 %carry, false
+  %cc = icmp eq i32 %a, %b
+  br i1 %cc, label %exit, label %if
+
+if:
+  br label %exit
+
+exit:
+  %cout = phi i1 [false, %entry], [%c2, %if]
+  store i32 %val, i32 addrspace(1)* %out, align 4
+  store i1 %cout, i1 addrspace(1)* %carryout
+  ret void
+}
+
 
 declare i32 @llvm.amdgcn.workitem.id.x() #1
 declare { i16, i1 } @llvm.uadd.with.overflow.i16(i16, i16) #1