R600/SI: Fix f64 inline immediates

llvm-svn: 224458

8 files changed, 126 insertions, 46 deletions

diff --git a/llvm/lib/Target/R600/AMDGPUMCInstLower.cpp b/llvm/lib/Target/R600/AMDGPUMCInstLower.cpp
index 2dcebf5f7e1..84ee7842a72 100644
--- a/llvm/lib/Target/R600/AMDGPUMCInstLower.cpp
+++ b/llvm/lib/Target/R600/AMDGPUMCInstLower.cpp
@@ -70,9 +70,14 @@ void AMDGPUMCInstLower::lower(const MachineInstr *MI, MCInst &OutMI) const {
       llvm_unreachable("unknown operand type");
     case MachineOperand::MO_FPImmediate: {
       const APFloat &FloatValue = MO.getFPImm()->getValueAPF();
-      assert(&FloatValue.getSemantics() == &APFloat::IEEEsingle &&
-             "Only floating point immediates are supported at the moment.");
-      MCOp = MCOperand::CreateFPImm(FloatValue.convertToFloat());
+
+      if (&FloatValue.getSemantics() == &APFloat::IEEEsingle)
+        MCOp = MCOperand::CreateFPImm(FloatValue.convertToFloat());
+      else if (&FloatValue.getSemantics() == &APFloat::IEEEdouble)
+        MCOp = MCOperand::CreateFPImm(FloatValue.convertToDouble());
+      else
+        llvm_unreachable("Unhandled floating point type");
+
       break;
     }
     case MachineOperand::MO_Immediate:
diff --git a/llvm/lib/Target/R600/InstPrinter/AMDGPUInstPrinter.cpp b/llvm/lib/Target/R600/InstPrinter/AMDGPUInstPrinter.cpp
index 12d077c0464..cb44c3e409f 100644
--- a/llvm/lib/Target/R600/InstPrinter/AMDGPUInstPrinter.cpp
+++ b/llvm/lib/Target/R600/InstPrinter/AMDGPUInstPrinter.cpp
@@ -14,6 +14,7 @@
 #include "MCTargetDesc/AMDGPUMCTargetDesc.h"
 #include "llvm/MC/MCExpr.h"
 #include "llvm/MC/MCInst.h"
+#include "llvm/MC/MCInstrInfo.h"
 #include "llvm/MC/MCRegisterInfo.h"
 #include "llvm/Support/MathExtras.h"
 
@@ -208,7 +209,7 @@ void AMDGPUInstPrinter::printRegOperand(unsigned reg, raw_ostream &O) {
   O << Type << '[' << RegIdx << ':' << (RegIdx + NumRegs - 1) << ']';
 }
 
-void AMDGPUInstPrinter::printImmediate(uint32_t Imm, raw_ostream &O) {
+void AMDGPUInstPrinter::printImmediate32(uint32_t Imm, raw_ostream &O) {
   int32_t SImm = static_cast<int32_t>(Imm);
   if (SImm >= -16 && SImm <= 64) {
     O << SImm;
@@ -233,9 +234,37 @@ void AMDGPUInstPrinter::printImmediate(uint32_t Imm, raw_ostream &O) {
     O << "4.0";
   else if (Imm == FloatToBits(-4.0f))
     O << "-4.0";
-  else {
+  else
     O << formatHex(static_cast<uint64_t>(Imm));
+}
+
+void AMDGPUInstPrinter::printImmediate64(uint64_t Imm, raw_ostream &O) {
+  int64_t SImm = static_cast<int64_t>(Imm);
+  if (SImm >= -16 && SImm <= 64) {
+    O << SImm;
+    return;
   }
+
+  if (Imm == DoubleToBits(0.0))
+    O << "0.0";
+  else if (Imm == DoubleToBits(1.0))
+    O << "1.0";
+  else if (Imm == DoubleToBits(-1.0))
+    O << "-1.0";
+  else if (Imm == DoubleToBits(0.5))
+    O << "0.5";
+  else if (Imm == DoubleToBits(-0.5))
+    O << "-0.5";
+  else if (Imm == DoubleToBits(2.0))
+    O << "2.0";
+  else if (Imm == DoubleToBits(-2.0))
+    O << "-2.0";
+  else if (Imm == DoubleToBits(4.0))
+    O << "4.0";
+  else if (Imm == DoubleToBits(-4.0))
+    O << "-4.0";
+  else
+    llvm_unreachable("64-bit literal constants not supported");
 }
 
 void AMDGPUInstPrinter::printOperand(const MCInst *MI, unsigned OpNo,
@@ -253,14 +282,37 @@ void AMDGPUInstPrinter::printOperand(const MCInst *MI, unsigned OpNo,
       break;
     }
   } else if (Op.isImm()) {
-    printImmediate(Op.getImm(), O);
+    const MCInstrDesc &Desc = MII.get(MI->getOpcode());
+    int RCID = Desc.OpInfo[OpNo].RegClass;
+    if (RCID != -1) {
+      const MCRegisterClass &ImmRC = MRI.getRegClass(RCID);
+      if (ImmRC.getSize() == 4)
+        printImmediate32(Op.getImm(), O);
+      else if (ImmRC.getSize() == 8)
+        printImmediate64(Op.getImm(), O);
+      else
+        llvm_unreachable("Invalid register class size");
+    } else {
+      // We hit this for the immediate instruction bits that don't yet have a
+      // custom printer.
+      // TODO: Eventually this should be unnecessary.
+      O << formatDec(Op.getImm());
+    }
   } else if (Op.isFPImm()) {
-
     // We special case 0.0 because otherwise it will be printed as an integer.
     if (Op.getFPImm() == 0.0)
       O << "0.0";
-    else
-      printImmediate(FloatToBits(Op.getFPImm()), O);
+    else {
+      const MCInstrDesc &Desc = MII.get(MI->getOpcode());
+      const MCRegisterClass &ImmRC = MRI.getRegClass(Desc.OpInfo[OpNo].RegClass);
+
+      if (ImmRC.getSize() == 4)
+        printImmediate32(FloatToBits(Op.getFPImm()), O);
+      else if (ImmRC.getSize() == 8)
+        printImmediate64(DoubleToBits(Op.getFPImm()), O);
+      else
+        llvm_unreachable("Invalid register class size");
+    }
   } else if (Op.isExpr()) {
     const MCExpr *Exp = Op.getExpr();
     Exp->print(O);
diff --git a/llvm/lib/Target/R600/InstPrinter/AMDGPUInstPrinter.h b/llvm/lib/Target/R600/InstPrinter/AMDGPUInstPrinter.h
index 4c06ac0b36b..1d43c7acbe7 100644
--- a/llvm/lib/Target/R600/InstPrinter/AMDGPUInstPrinter.h
+++ b/llvm/lib/Target/R600/InstPrinter/AMDGPUInstPrinter.h
@@ -48,7 +48,8 @@ private:
   void printSLC(const MCInst *MI, unsigned OpNo, raw_ostream &O);
   void printTFE(const MCInst *MI, unsigned OpNo, raw_ostream &O);
   void printRegOperand(unsigned RegNo, raw_ostream &O);
-  void printImmediate(uint32_t Imm, raw_ostream &O);
+  void printImmediate32(uint32_t I, raw_ostream &O);
+  void printImmediate64(uint64_t I, raw_ostream &O);
   void printOperand(const MCInst *MI, unsigned OpNo, raw_ostream &O);
   void printOperandAndMods(const MCInst *MI, unsigned OpNo, raw_ostream &O);
   static void printInterpSlot(const MCInst *MI, unsigned OpNum, raw_ostream &O);
diff --git a/llvm/lib/Target/R600/SIISelLowering.cpp b/llvm/lib/Target/R600/SIISelLowering.cpp
index e41cad23de2..5702ee67123 100644
--- a/llvm/lib/Target/R600/SIISelLowering.cpp
+++ b/llvm/lib/Target/R600/SIISelLowering.cpp
@@ -1578,31 +1578,30 @@ static bool isSSrc(unsigned RegClass) {
 /// and the immediate value if it's a literal immediate
 int32_t SITargetLowering::analyzeImmediate(const SDNode *N) const {
 
-  union {
-    int32_t I;
-    float F;
-  } Imm;
+  const SIInstrInfo *TII = static_cast<const SIInstrInfo *>(
+    getTargetMachine().getSubtargetImpl()->getInstrInfo());
 
   if (const ConstantSDNode *Node = dyn_cast<ConstantSDNode>(N)) {
-    if (Node->getZExtValue() >> 32) {
-        return -1;
-    }
-    Imm.I = Node->getSExtValue();
-  } else if (const ConstantFPSDNode *Node = dyn_cast<ConstantFPSDNode>(N)) {
-    if (N->getValueType(0) != MVT::f32)
+    if (Node->getZExtValue() >> 32)
       return -1;
-    Imm.F = Node->getValueAPF().convertToFloat();
-  } else
-    return -1; // It isn't an immediate
-
-  if ((Imm.I >= -16 && Imm.I <= 64) ||
-      Imm.F == 0.5f || Imm.F == -0.5f ||
-      Imm.F == 1.0f || Imm.F == -1.0f ||
-      Imm.F == 2.0f || Imm.F == -2.0f ||
-      Imm.F == 4.0f || Imm.F == -4.0f)
-    return 0; // It's an inline immediate
-
-  return Imm.I; // It's a literal immediate
+
+    if (TII->isInlineConstant(Node->getAPIntValue()))
+      return 0;
+
+    return Node->getZExtValue();
+  }
+
+  if (const ConstantFPSDNode *Node = dyn_cast<ConstantFPSDNode>(N)) {
+    if (TII->isInlineConstant(Node->getValueAPF().bitcastToAPInt()))
+      return 0;
+
+    if (Node->getValueType(0) == MVT::f32)
+      return FloatToBits(Node->getValueAPF().convertToFloat());
+
+    return -1;
+  }
+
+  return -1;
 }
 
 /// \brief Try to fold an immediate directly into an instruction
diff --git a/llvm/lib/Target/R600/SIInstrInfo.cpp b/llvm/lib/Target/R600/SIInstrInfo.cpp
index 9328c2c2c23..08bfc5e1422 100644
--- a/llvm/lib/Target/R600/SIInstrInfo.cpp
+++ b/llvm/lib/Target/R600/SIInstrInfo.cpp
@@ -896,10 +896,23 @@ bool SIInstrInfo::areMemAccessesTriviallyDisjoint(MachineInstr *MIa,
 }
 
 bool SIInstrInfo::isInlineConstant(const APInt &Imm) const {
-  int32_t Val = Imm.getSExtValue();
-  if (Val >= -16 && Val <= 64)
+  int64_t SVal = Imm.getSExtValue();
+  if (SVal >= -16 && SVal <= 64)
     return true;
 
+  if (Imm.getBitWidth() == 64) {
+    uint64_t Val = Imm.getZExtValue();
+    return (DoubleToBits(0.0) == Val) ||
+           (DoubleToBits(1.0) == Val) ||
+           (DoubleToBits(-1.0) == Val) ||
+           (DoubleToBits(0.5) == Val) ||
+           (DoubleToBits(-0.5) == Val) ||
+           (DoubleToBits(2.0) == Val) ||
+           (DoubleToBits(-2.0) == Val) ||
+           (DoubleToBits(4.0) == Val) ||
+           (DoubleToBits(-4.0) == Val);
+  }
+
   // The actual type of the operand does not seem to matter as long
   // as the bits match one of the inline immediate values.  For example:
   //
@@ -908,16 +921,17 @@ bool SIInstrInfo::isInlineConstant(const APInt &Imm) const {
   //
   // 1065353216 has the hexadecimal encoding 0x3f800000 which is 1.0f in
   // floating-point, so it is a legal inline immediate.
-
-  return (APInt::floatToBits(0.0f) == Imm) ||
-         (APInt::floatToBits(1.0f) == Imm) ||
-         (APInt::floatToBits(-1.0f) == Imm) ||
-         (APInt::floatToBits(0.5f) == Imm) ||
-         (APInt::floatToBits(-0.5f) == Imm) ||
-         (APInt::floatToBits(2.0f) == Imm) ||
-         (APInt::floatToBits(-2.0f) == Imm) ||
-         (APInt::floatToBits(4.0f) == Imm) ||
-         (APInt::floatToBits(-4.0f) == Imm);
+  uint32_t Val = Imm.getZExtValue();
+
+  return (FloatToBits(0.0f) == Val) ||
+         (FloatToBits(1.0f) == Val) ||
+         (FloatToBits(-1.0f) == Val) ||
+         (FloatToBits(0.5f) == Val) ||
+         (FloatToBits(-0.5f) == Val) ||
+         (FloatToBits(2.0f) == Val) ||
+         (FloatToBits(-2.0f) == Val) ||
+         (FloatToBits(4.0f) == Val) ||
+         (FloatToBits(-4.0f) == Val);
 }
 
 bool SIInstrInfo::isInlineConstant(const MachineOperand &MO) const {
diff --git a/llvm/lib/Target/R600/SIInstrInfo.td b/llvm/lib/Target/R600/SIInstrInfo.td
index 6323591359d..8122ccb65cc 100644
--- a/llvm/lib/Target/R600/SIInstrInfo.td
+++ b/llvm/lib/Target/R600/SIInstrInfo.td
@@ -188,6 +188,10 @@ class InlineImm <ValueType vt> : PatLeaf <(vt imm), [{
   return isInlineImmediate(N);
 }]>;
 
+class InlineFPImm <ValueType vt> : PatLeaf <(vt fpimm), [{
+  return isInlineImmediate(N);
+}]>;
+
 class SGPRImm <dag frag> : PatLeaf<frag, [{
   if (TM.getSubtarget<AMDGPUSubtarget>().getGeneration() <
       AMDGPUSubtarget::SOUTHERN_ISLANDS) {
diff --git a/llvm/lib/Target/R600/SIInstructions.td b/llvm/lib/Target/R600/SIInstructions.td
index 8c3f1403ca2..463287eaa97 100644
--- a/llvm/lib/Target/R600/SIInstructions.td
+++ b/llvm/lib/Target/R600/SIInstructions.td
@@ -2572,6 +2572,11 @@ def : Pat <
   (S_MOV_B64 (i64 (as_i64imm $imm)))
 >;
 
+def : Pat <
+  (f64 InlineFPImm<f64>:$imm),
+  (S_MOV_B64 InlineFPImm<f64>:$imm)
+>;
+
 /********** ===================== **********/
 /********** Interpolation Paterns **********/
 /********** ===================== **********/
diff --git a/llvm/lib/Target/R600/SIRegisterInfo.td b/llvm/lib/Target/R600/SIRegisterInfo.td
index 45c2b414a3e..a79da004def 100644
--- a/llvm/lib/Target/R600/SIRegisterInfo.td
+++ b/llvm/lib/Target/R600/SIRegisterInfo.td
@@ -184,9 +184,9 @@ def SReg_32 : RegisterClass<"AMDGPU", [f32, i32], 32,
   (add SGPR_32, M0Reg, VCC_LO, VCC_HI, EXEC_LO, EXEC_HI, FLAT_SCR_LO, FLAT_SCR_HI)
 >;
 
-def SGPR_64 : RegisterClass<"AMDGPU", [v2i32, i64], 64, (add SGPR_64Regs)>;
+def SGPR_64 : RegisterClass<"AMDGPU", [v2i32, i64, f64], 64, (add SGPR_64Regs)>;
 
-def SReg_64 : RegisterClass<"AMDGPU", [v2i32, i64, i1], 64,
+def SReg_64 : RegisterClass<"AMDGPU", [v2i32, i64, f64, i1], 64,
   (add SGPR_64, VCCReg, EXECReg, FLAT_SCR)
 >;