R600/SI: Lower 64-bit immediates using REG_SEQUENCE

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@205561 91177308-0d34-0410-b5e6-96231b3b80d8

lib/Target/R600/AMDGPUISelDAGToDAG.cpp

@@ -48,6 +48,7 @@ public:
   virtual void PostprocessISelDAG();
 
 private:
+  bool isInlineImmediate(SDNode *N) const;
   inline SDValue getSmallIPtrImm(unsigned Imm);
   bool FoldOperand(SDValue &Src, SDValue &Sel, SDValue &Neg, SDValue &Abs,
                    const R600InstrInfo *TII);
@@ -103,6 +104,12 @@ AMDGPUDAGToDAGISel::AMDGPUDAGToDAGISel(TargetMachine &TM)
 AMDGPUDAGToDAGISel::~AMDGPUDAGToDAGISel() {
 }
 
+bool AMDGPUDAGToDAGISel::isInlineImmediate(SDNode *N) const {
+  const SITargetLowering *TL
+      = static_cast<const SITargetLowering *>(getTargetLowering());
+  return TL->analyzeImmediate(N) == 0;
+}
+
 /// \brief Determine the register class for \p OpNo
 /// \returns The register class of the virtual register that will be used for
 /// the given operand number \OpNo or NULL if the register class cannot be
@@ -357,6 +364,37 @@ SDNode *AMDGPUDAGToDAGISel::Select(SDNode *N) {
     return CurDAG->getMachineNode(TargetOpcode::REG_SEQUENCE,
                                   SDLoc(N), N->getValueType(0), Ops);
   }
+
+  case ISD::Constant:
+  case ISD::ConstantFP: {
+    const AMDGPUSubtarget &ST = TM.getSubtarget<AMDGPUSubtarget>();
+    if (ST.getGeneration() < AMDGPUSubtarget::SOUTHERN_ISLANDS ||
+        N->getValueType(0).getSizeInBits() != 64 || isInlineImmediate(N))
+      break;
+
+    uint64_t Imm;
+    if (ConstantFPSDNode *FP = dyn_cast<ConstantFPSDNode>(N))
+      Imm = FP->getValueAPF().bitcastToAPInt().getZExtValue();
+    else {
+      ConstantSDNode *C = dyn_cast<ConstantSDNode>(N);
+      assert(C);
+      Imm = C->getZExtValue();
+    }
+
+    SDNode *Lo = CurDAG->getMachineNode(AMDGPU::S_MOV_B32, SDLoc(N), MVT::i32,
+                                CurDAG->getConstant(Imm & 0xFFFFFFFF, MVT::i32));
+    SDNode *Hi = CurDAG->getMachineNode(AMDGPU::S_MOV_B32, SDLoc(N), MVT::i32,
+                                CurDAG->getConstant(Imm >> 32, MVT::i32));
+    const SDValue Ops[] = {
+      CurDAG->getTargetConstant(AMDGPU::SReg_64RegClassID, MVT::i32),
+      SDValue(Lo, 0), CurDAG->getTargetConstant(AMDGPU::sub0, MVT::i32),
+      SDValue(Hi, 0), CurDAG->getTargetConstant(AMDGPU::sub1, MVT::i32)
+    };
+
+    return CurDAG->getMachineNode(TargetOpcode::REG_SEQUENCE, SDLoc(N),
+                                  N->getValueType(0), Ops);
+  }
+
   case AMDGPUISD::REGISTER_LOAD: {
     const AMDGPUSubtarget &ST = TM.getSubtarget<AMDGPUSubtarget>();
     if (ST.getGeneration() <= AMDGPUSubtarget::NORTHERN_ISLANDS)

lib/Target/R600/SIISelLowering.cpp

@@ -1029,9 +1029,11 @@ int32_t SITargetLowering::analyzeImmediate(const SDNode *N) const {
         return -1;
     }
     Imm.I = Node->getSExtValue();
-  } else if (const ConstantFPSDNode *Node = dyn_cast<ConstantFPSDNode>(N))
+  } else if (const ConstantFPSDNode *Node = dyn_cast<ConstantFPSDNode>(N)) {
+    if (N->getValueType(0) != MVT::f32)
+      return -1;
     Imm.F = Node->getValueAPF().convertToFloat();
-  else
+  } else
     return -1; // It isn't an immediate
 
   if ((Imm.I >= -16 && Imm.I <= 64) ||

lib/Target/R600/SIInstrInfo.td

@@ -117,8 +117,7 @@ def mubuf_vaddr_offset : PatFrag<
 >;
 
 class InlineImm <ValueType vt> : PatLeaf <(vt imm), [{
-  return
-    (*(const SITargetLowering *)getTargetLowering()).analyzeImmediate(N) == 0;
+  return isInlineImmediate(N);
 }]>;
 
 class SGPRImm <dag frag> : PatLeaf<frag, [{

lib/Target/R600/SIInstructions.td

@@ -1777,21 +1777,6 @@ def : Pat <
   (S_MOV_B64 InlineImm<i64>:$imm)
 >;
 
-// i64 immediates aren't supported in hardware, split it into two 32bit values
-def : Pat <
-  (i64 imm:$imm),
-  (INSERT_SUBREG (INSERT_SUBREG (i64 (IMPLICIT_DEF)),
-    (S_MOV_B32 (i32 (LO32 imm:$imm))), sub0),
-    (S_MOV_B32 (i32 (HI32 imm:$imm))), sub1)
->;
-
-def : Pat <
-  (f64 fpimm:$imm),
-  (INSERT_SUBREG (INSERT_SUBREG (f64 (IMPLICIT_DEF)),
-    (V_MOV_B32_e32 (f32 (LO32f fpimm:$imm))), sub0),
-    (V_MOV_B32_e32 (f32 (HI32f fpimm:$imm))), sub1)
->;
-
 /********** ===================== **********/
 /********** Interpolation Paterns **********/
 /********** ===================== **********/

test/CodeGen/R600/fconst64.ll

@@ -1,8 +1,8 @@
 ; RUN: llc < %s -march=r600 -mcpu=tahiti -verify-machineinstrs | FileCheck %s
 
 ; CHECK: @fconst_f64
-; CHECK: V_MOV_B32_e32 {{v[0-9]+}}, 0.000000e+00
-; CHECK-NEXT: V_MOV_B32_e32 {{v[0-9]+}}, 2.312500e+00
+; CHECK-DAG: S_MOV_B32 {{s[0-9]+}}, 1075052544
+; CHECK-DAG: S_MOV_B32 {{s[0-9]+}}, 0
 
 define void @fconst_f64(double addrspace(1)* %out, double addrspace(1)* %in) {
    %r1 = load double addrspace(1)* %in

test/CodeGen/R600/trunc.ll

@@ -31,8 +31,9 @@ define void @trunc_load_shl_i64(i32 addrspace(1)* %out, i64 %a) {
 
 ; SI-LABEL: @trunc_shl_i64:
 ; SI: S_LOAD_DWORDX2 s{{\[}}[[LO_SREG:[0-9]+]]:{{[0-9]+\]}},
-; SI: V_ADD_I32_e32 v[[LO_ADD:[0-9]+]], s[[LO_SREG]],
-; SI: V_LSHL_B64 v{{\[}}[[LO_VREG:[0-9]+]]:{{[0-9]+\]}}, v{{\[}}[[LO_ADD]]:{{[0-9]+\]}}, 2
+; SI: S_ADD_I32 s[[LO_ADD:[0-9]+]], s[[LO_SREG]],
+; SI: S_LSHL_B64 s{{\[}}[[LO_SREG2:[0-9]+]]:{{[0-9]+\]}}, s{{\[}}[[LO_ADD]]:{{[0-9]+\]}}, 2
+; SI: V_MOV_B32_e32 v[[LO_VREG:[0-9]+]], s[[LO_SREG2]]
 ; SI: BUFFER_STORE_DWORD v[[LO_VREG]],
 define void @trunc_shl_i64(i64 addrspace(1)* %out2, i32 addrspace(1)* %out, i64 %a) {
   %aa = add i64 %a, 234 ; Prevent shrinking store.