AMDGPU: Redefine clamp node as clamp 0.0-1.0

Change implementation to use max instead of add.
min/max/med3 do not flush denormals regardless of the mode,
so it is OK to use it whether or not they are enabled.

Also allow using clamp with f16, and use knowledge
of dx10_clamp.

llvm-svn: 295788

12 files changed, 163 insertions, 29 deletions

diff --git a/llvm/lib/Target/AMDGPU/AMDGPU.td b/llvm/lib/Target/AMDGPU/AMDGPU.td
index 0de72734b71..17d6b460a80 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPU.td
+++ b/llvm/lib/Target/AMDGPU/AMDGPU.td
@@ -257,6 +257,12 @@ def FeatureFP16Denormals : SubtargetFeature<"fp16-denormals",
   [FeatureFP64FP16Denormals]
 >;
 
+def FeatureDX10Clamp : SubtargetFeature<"dx10-clamp",
+  "DX10Clamp",
+  "true",
+  "clamp modifier clamps NaNs to 0.0"
+>;
+
 def FeatureFPExceptions : SubtargetFeature<"fp-exceptions",
   "FPExceptions",
   "true",
diff --git a/llvm/lib/Target/AMDGPU/AMDGPUAsmPrinter.cpp b/llvm/lib/Target/AMDGPU/AMDGPUAsmPrinter.cpp
index c3ff2e32254..577509d31ce 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPUAsmPrinter.cpp
+++ b/llvm/lib/Target/AMDGPU/AMDGPUAsmPrinter.cpp
@@ -586,7 +586,7 @@ void AMDGPUAsmPrinter::getSIProgramInfo(SIProgramInfo &ProgInfo,
   ProgInfo.IEEEMode = STM.enableIEEEBit(MF);
 
   // Make clamp modifier on NaN input returns 0.
-  ProgInfo.DX10Clamp = 1;
+  ProgInfo.DX10Clamp = STM.enableDX10Clamp();
 
   const MachineFrameInfo &FrameInfo = MF.getFrameInfo();
   ProgInfo.ScratchSize = FrameInfo.getStackSize();
diff --git a/llvm/lib/Target/AMDGPU/AMDGPUISelLowering.cpp b/llvm/lib/Target/AMDGPU/AMDGPUISelLowering.cpp
index 12c17c3e9eb..9e3ac8b871e 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPUISelLowering.cpp
+++ b/llvm/lib/Target/AMDGPU/AMDGPUISelLowering.cpp
@@ -1012,22 +1012,29 @@ SDValue AMDGPUTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op,
   EVT VT = Op.getValueType();
 
   switch (IntrinsicID) {
-    default: return Op;
-    case AMDGPUIntrinsic::AMDGPU_clamp: // Legacy name.
-      return DAG.getNode(AMDGPUISD::CLAMP, DL, VT,
-                         Op.getOperand(1), Op.getOperand(2), Op.getOperand(3));
+  default: return Op;
+  case AMDGPUIntrinsic::AMDGPU_clamp: {
+    // Deprecated in favor of emitting min/max combo or fmed3.
+    ConstantFPSDNode *CSrc1 = dyn_cast<ConstantFPSDNode>(Op.getOperand(2));
+    ConstantFPSDNode *CSrc2 = dyn_cast<ConstantFPSDNode>(Op.getOperand(3));
+    if (CSrc1 && CSrc2 && CSrc1->isZero() && CSrc2->isExactlyValue(1.0))
+      return DAG.getNode(AMDGPUISD::CLAMP, DL, VT, Op.getOperand(1));
 
-    case AMDGPUIntrinsic::AMDGPU_bfe_i32:
-      return DAG.getNode(AMDGPUISD::BFE_I32, DL, VT,
-                         Op.getOperand(1),
-                         Op.getOperand(2),
-                         Op.getOperand(3));
+    SDValue Max = DAG.getNode(ISD::FMAXNUM, DL, VT, Op.getOperand(1),
+                              Op.getOperand(2));
+    return DAG.getNode(ISD::FMINNUM, DL, VT, Max, Op.getOperand(3));
+  }
+  case AMDGPUIntrinsic::AMDGPU_bfe_i32:
+    return DAG.getNode(AMDGPUISD::BFE_I32, DL, VT,
+                       Op.getOperand(1),
+                       Op.getOperand(2),
+                       Op.getOperand(3));
 
-    case AMDGPUIntrinsic::AMDGPU_bfe_u32:
-      return DAG.getNode(AMDGPUISD::BFE_U32, DL, VT,
-                         Op.getOperand(1),
-                         Op.getOperand(2),
-                         Op.getOperand(3));
+  case AMDGPUIntrinsic::AMDGPU_bfe_u32:
+    return DAG.getNode(AMDGPUISD::BFE_U32, DL, VT,
+                       Op.getOperand(1),
+                       Op.getOperand(2),
+                       Op.getOperand(3));
   }
 }
 
@@ -2445,6 +2452,28 @@ SDValue AMDGPUTargetLowering::performStoreCombine(SDNode *N,
                       SN->getBasePtr(), SN->getMemOperand());
 }
 
+SDValue AMDGPUTargetLowering::performClampCombine(SDNode *N,
+                                                  DAGCombinerInfo &DCI) const {
+  ConstantFPSDNode *CSrc = dyn_cast<ConstantFPSDNode>(N->getOperand(0));
+  if (!CSrc)
+    return SDValue();
+
+  const APFloat &F = CSrc->getValueAPF();
+  APFloat Zero = APFloat::getZero(F.getSemantics());
+  APFloat::cmpResult Cmp0 = F.compare(Zero);
+  if (Cmp0 == APFloat::cmpLessThan ||
+      (Cmp0 == APFloat::cmpUnordered && Subtarget->enableDX10Clamp())) {
+    return DCI.DAG.getConstantFP(Zero, SDLoc(N), N->getValueType(0));
+  }
+
+  APFloat One(F.getSemantics(), "1.0");
+  APFloat::cmpResult Cmp1 = F.compare(One);
+  if (Cmp1 == APFloat::cmpGreaterThan)
+    return DCI.DAG.getConstantFP(One, SDLoc(N), N->getValueType(0));
+
+  return SDValue(CSrc, 0);
+}
+
 /// Split the 64-bit value \p LHS into two 32-bit components, and perform the
 /// binary operation \p Opc to it with the corresponding constant operands.
 SDValue AMDGPUTargetLowering::splitBinaryBitConstantOpImpl(
@@ -3323,6 +3352,8 @@ SDValue AMDGPUTargetLowering::PerformDAGCombine(SDNode *N,
     return performLoadCombine(N, DCI);
   case ISD::STORE:
     return performStoreCombine(N, DCI);
+  case AMDGPUISD::CLAMP:
+    return performClampCombine(N, DCI);
   }
   return SDValue();
 }
diff --git a/llvm/lib/Target/AMDGPU/AMDGPUISelLowering.h b/llvm/lib/Target/AMDGPU/AMDGPUISelLowering.h
index 58ac09f2698..fb487f914ad 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPUISelLowering.h
+++ b/llvm/lib/Target/AMDGPU/AMDGPUISelLowering.h
@@ -70,6 +70,7 @@ protected:
   bool shouldCombineMemoryType(EVT VT) const;
   SDValue performLoadCombine(SDNode *N, DAGCombinerInfo &DCI) const;
   SDValue performStoreCombine(SDNode *N, DAGCombinerInfo &DCI) const;
+  SDValue performClampCombine(SDNode *N, DAGCombinerInfo &DCI) const;
 
   SDValue splitBinaryBitConstantOpImpl(DAGCombinerInfo &DCI, const SDLoc &SL,
                                        unsigned Opc, SDValue LHS,
@@ -238,7 +239,11 @@ enum NodeType : unsigned {
   RETURN,
   DWORDADDR,
   FRACT,
+
+  /// CLAMP value between 0.0 and 1.0. NaN clamped to 0, following clamp output
+  /// modifier behavior with dx10_enable.
   CLAMP,
+
   // This is SETCC with the full mask result which is used for a compare with a
   // result bit per item in the wavefront.
   SETCC,
diff --git a/llvm/lib/Target/AMDGPU/AMDGPUInstrInfo.td b/llvm/lib/Target/AMDGPU/AMDGPUInstrInfo.td
index 91341ef88ea..a081c30ef81 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPUInstrInfo.td
+++ b/llvm/lib/Target/AMDGPU/AMDGPUInstrInfo.td
@@ -92,7 +92,7 @@ def AMDGPUfmul_legacy : SDNode<"AMDGPUISD::FMUL_LEGACY", SDTFPBinOp,
   [SDNPCommutative, SDNPAssociative]
 >;
 
-def AMDGPUclamp : SDNode<"AMDGPUISD::CLAMP", SDTFPTernaryOp, []>;
+def AMDGPUclamp : SDNode<"AMDGPUISD::CLAMP", SDTFPUnaryOp>;
 
 // out = min(a, b) a and b are floats, where a nan comparison fails.
 def AMDGPUfmin_legacy : SDNode<"AMDGPUISD::FMIN_LEGACY", SDTFPBinOp,
diff --git a/llvm/lib/Target/AMDGPU/AMDGPUInstructions.td b/llvm/lib/Target/AMDGPU/AMDGPUInstructions.td
index 7d2a52bccc8..e76891c3bed 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPUInstructions.td
+++ b/llvm/lib/Target/AMDGPU/AMDGPUInstructions.td
@@ -452,7 +452,7 @@ class CLAMP <RegisterClass rc> : AMDGPUShaderInst <
   (outs rc:$dst),
   (ins rc:$src0),
   "CLAMP $dst, $src0",
-  [(set f32:$dst, (AMDGPUclamp f32:$src0, (f32 FP_ZERO), (f32 FP_ONE)))]
+  [(set f32:$dst, (AMDGPUclamp f32:$src0))]
 >;
 
 class FABS <RegisterClass rc> : AMDGPUShaderInst <
diff --git a/llvm/lib/Target/AMDGPU/AMDGPUSubtarget.cpp b/llvm/lib/Target/AMDGPU/AMDGPUSubtarget.cpp
index 32b04fe7a1e..a719931eff0 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPUSubtarget.cpp
+++ b/llvm/lib/Target/AMDGPU/AMDGPUSubtarget.cpp
@@ -42,7 +42,7 @@ AMDGPUSubtarget::initializeSubtargetDependencies(const Triple &TT,
   // for SI has the unhelpful behavior that it unsets everything else if you
   // disable it.
 
-  SmallString<256> FullFS("+promote-alloca,+fp64-fp16-denormals,+load-store-opt,");
+  SmallString<256> FullFS("+promote-alloca,+fp64-fp16-denormals,+dx10-clamp,+load-store-opt,");
   if (isAmdHsaOS()) // Turn on FlatForGlobal for HSA.
     FullFS += "+flat-for-global,+unaligned-buffer-access,+trap-handler,";
 
@@ -89,6 +89,7 @@ AMDGPUSubtarget::AMDGPUSubtarget(const Triple &TT, StringRef GPU, StringRef FS,
     FP32Denormals(false),
     FP64FP16Denormals(false),
     FPExceptions(false),
+    DX10Clamp(false),
     FlatForGlobal(false),
     UnalignedScratchAccess(false),
     UnalignedBufferAccess(false),
diff --git a/llvm/lib/Target/AMDGPU/AMDGPUSubtarget.h b/llvm/lib/Target/AMDGPU/AMDGPUSubtarget.h
index 433c295feaf..8f1aaa1d893 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPUSubtarget.h
+++ b/llvm/lib/Target/AMDGPU/AMDGPUSubtarget.h
@@ -103,6 +103,7 @@ protected:
   bool FP32Denormals;
   bool FP64FP16Denormals;
   bool FPExceptions;
+  bool DX10Clamp;
   bool FlatForGlobal;
   bool UnalignedScratchAccess;
   bool UnalignedBufferAccess;
@@ -294,10 +295,6 @@ public:
     return DumpCode;
   }
 
-  bool enableIEEEBit(const MachineFunction &MF) const {
-    return AMDGPU::isCompute(MF.getFunction()->getCallingConv());
-  }
-
   /// Return the amount of LDS that can be used that will not restrict the
   /// occupancy lower than WaveCount.
   unsigned getMaxLocalMemSizeWithWaveCount(unsigned WaveCount,
@@ -323,6 +320,14 @@ public:
     return FPExceptions;
   }
 
+  bool enableDX10Clamp() const {
+    return DX10Clamp;
+  }
+
+  bool enableIEEEBit(const MachineFunction &MF) const {
+    return AMDGPU::isCompute(MF.getFunction()->getCallingConv());
+  }
+
   bool useFlatForGlobal() const {
     return FlatForGlobal;
   }
diff --git a/llvm/lib/Target/AMDGPU/SIISelLowering.cpp b/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
index 60acf4fb1c0..40a28203000 100644
--- a/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
+++ b/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
@@ -3995,8 +3995,10 @@ static bool isKnownNeverSNan(SelectionDAG &DAG, SDValue Op) {
   return DAG.isKnownNeverNaN(Op);
 }
 
-static SDValue performFPMed3ImmCombine(SelectionDAG &DAG, const SDLoc &SL,
-                                       SDValue Op0, SDValue Op1) {
+SDValue SITargetLowering::performFPMed3ImmCombine(SelectionDAG &DAG,
+                                                  const SDLoc &SL,
+                                                  SDValue Op0,
+                                                  SDValue Op1) const {
   ConstantFPSDNode *K1 = dyn_cast<ConstantFPSDNode>(Op1);
   if (!K1)
     return SDValue();
@@ -4010,6 +4012,20 @@ static SDValue performFPMed3ImmCombine(SelectionDAG &DAG, const SDLoc &SL,
   if (Cmp == APFloat::cmpGreaterThan)
     return SDValue();
 
+  // TODO: Check IEEE bit enabled?
+  EVT VT = K0->getValueType(0);
+  if (Subtarget->enableDX10Clamp()) {
+    // If dx10_clamp is enabled, NaNs clamp to 0.0. This is the same as the
+    // hardware fmed3 behavior converting to a min.
+    // FIXME: Should this be allowing -0.0?
+    if (K1->isExactlyValue(1.0) && K0->isExactlyValue(0.0))
+      return DAG.getNode(AMDGPUISD::CLAMP, SL, VT, Op0.getOperand(0));
+  }
+
+  // No med3 for f16, but clamp is possible.
+  if (VT == MVT::f16)
+    return SDValue();
+
   // This isn't safe with signaling NaNs because in IEEE mode, min/max on a
   // signaling NaN gives a quiet NaN. The quiet NaN input to the min would then
   // give the other result, which is different from med3 with a NaN input.
@@ -4074,7 +4090,9 @@ SDValue SITargetLowering::performMinMaxCombine(SDNode *N,
   if (((Opc == ISD::FMINNUM && Op0.getOpcode() == ISD::FMAXNUM) ||
        (Opc == AMDGPUISD::FMIN_LEGACY &&
         Op0.getOpcode() == AMDGPUISD::FMAX_LEGACY)) &&
-      N->getValueType(0) == MVT::f32 && Op0.hasOneUse()) {
+      (N->getValueType(0) == MVT::f32 ||
+       (N->getValueType(0) == MVT::f16 && Subtarget->has16BitInsts())) &&
+      Op0.hasOneUse()) {
     if (SDValue Res = performFPMed3ImmCombine(DAG, SDLoc(N), Op0, Op1))
       return Res;
   }
@@ -4082,6 +4100,60 @@ SDValue SITargetLowering::performMinMaxCombine(SDNode *N,
   return SDValue();
 }
 
+static bool isClampZeroToOne(SDValue A, SDValue B) {
+  if (ConstantFPSDNode *CA = dyn_cast<ConstantFPSDNode>(A)) {
+    if (ConstantFPSDNode *CB = dyn_cast<ConstantFPSDNode>(B)) {
+      // FIXME: Should this be allowing -0.0?
+      return (CA->isExactlyValue(0.0) && CB->isExactlyValue(1.0)) ||
+             (CA->isExactlyValue(1.0) && CB->isExactlyValue(0.0));
+    }
+  }
+
+  return false;
+}
+
+// FIXME: Should only worry about snans for version with chain.
+SDValue SITargetLowering::performFMed3Combine(SDNode *N,
+                                              DAGCombinerInfo &DCI) const {
+  EVT VT = N->getValueType(0);
+  // v_med3_f32 and v_max_f32 behave identically wrt denorms, exceptions and
+  // NaNs. With a NaN input, the order of the operands may change the result.
+
+  SelectionDAG &DAG = DCI.DAG;
+  SDLoc SL(N);
+
+  SDValue Src0 = N->getOperand(0);
+  SDValue Src1 = N->getOperand(1);
+  SDValue Src2 = N->getOperand(2);
+
+  if (isClampZeroToOne(Src0, Src1)) {
+    // const_a, const_b, x -> clamp is safe in all cases including signaling
+    // nans.
+    // FIXME: Should this be allowing -0.0?
+    return DAG.getNode(AMDGPUISD::CLAMP, SL, VT, Src2);
+  }
+
+  // FIXME: dx10_clamp behavior assumed in instcombine. Should we really bother
+  // handling no dx10-clamp?
+  if (Subtarget->enableDX10Clamp()) {
+    // If NaNs is clamped to 0, we are free to reorder the inputs.
+
+    if (isa<ConstantFPSDNode>(Src0) && !isa<ConstantFPSDNode>(Src1))
+      std::swap(Src0, Src1);
+
+    if (isa<ConstantFPSDNode>(Src1) && !isa<ConstantFPSDNode>(Src2))
+      std::swap(Src1, Src2);
+
+    if (isa<ConstantFPSDNode>(Src0) && !isa<ConstantFPSDNode>(Src1))
+      std::swap(Src0, Src1);
+
+    if (isClampZeroToOne(Src1, Src2))
+      return DAG.getNode(AMDGPUISD::CLAMP, SL, VT, Src0);
+  }
+
+  return SDValue();
+}
+
 unsigned SITargetLowering::getFusedOpcode(const SelectionDAG &DAG,
                                           const SDNode *N0,
                                           const SDNode *N1) const {
@@ -4348,6 +4420,8 @@ SDValue SITargetLowering::PerformDAGCombine(SDNode *N,
   case AMDGPUISD::CVT_F32_UBYTE2:
   case AMDGPUISD::CVT_F32_UBYTE3:
     return performCvtF32UByteNCombine(N, DCI);
+  case AMDGPUISD::FMED3:
+    return performFMed3Combine(N, DCI);
   }
   return AMDGPUTargetLowering::PerformDAGCombine(N, DCI);
 }
diff --git a/llvm/lib/Target/AMDGPU/SIISelLowering.h b/llvm/lib/Target/AMDGPU/SIISelLowering.h
index 037b6f730c5..fe86f8e2b65 100644
--- a/llvm/lib/Target/AMDGPU/SIISelLowering.h
+++ b/llvm/lib/Target/AMDGPU/SIISelLowering.h
@@ -84,7 +84,10 @@ class SITargetLowering final : public AMDGPUTargetLowering {
   SDValue performClassCombine(SDNode *N, DAGCombinerInfo &DCI) const;
   SDValue performFCanonicalizeCombine(SDNode *N, DAGCombinerInfo &DCI) const;
 
+  SDValue performFPMed3ImmCombine(SelectionDAG &DAG, const SDLoc &SL,
+                                  SDValue Op0, SDValue Op1) const;
   SDValue performMinMaxCombine(SDNode *N, DAGCombinerInfo &DCI) const;
+  SDValue performFMed3Combine(SDNode *N, DAGCombinerInfo &DCI) const;
 
   unsigned getFusedOpcode(const SelectionDAG &DAG,
                           const SDNode *N0, const SDNode *N1) const;
diff --git a/llvm/lib/Target/AMDGPU/SIInstrInfo.td b/llvm/lib/Target/AMDGPU/SIInstrInfo.td
index 94580543afd..aa6d677f466 100644
--- a/llvm/lib/Target/AMDGPU/SIInstrInfo.td
+++ b/llvm/lib/Target/AMDGPU/SIInstrInfo.td
@@ -625,6 +625,7 @@ def SRCMODS {
 
 def DSTCLAMP {
   int NONE = 0;
+  int ENABLE = 1;
 }
 
 def DSTOMOD {
diff --git a/llvm/lib/Target/AMDGPU/SIInstructions.td b/llvm/lib/Target/AMDGPU/SIInstructions.td
index 1d09880147d..4bd759012a3 100644
--- a/llvm/lib/Target/AMDGPU/SIInstructions.td
+++ b/llvm/lib/Target/AMDGPU/SIInstructions.td
@@ -647,12 +647,20 @@ def : BitConvert <v16f32, v16i32, VReg_512>;
 /********** Src & Dst modifiers **********/
 /********** =================== **********/
 
-def : Pat <
-  (AMDGPUclamp (VOP3Mods0Clamp f32:$src0, i32:$src0_modifiers, i32:$omod),
-               (f32 FP_ZERO), (f32 FP_ONE)),
-  (V_ADD_F32_e64 $src0_modifiers, $src0, 0, (i32 0), 1, $omod)
+
+// If denormals are not enabled, it only impacts the compare of the
+// inputs. The output result is not flushed.
+class ClampPat<Instruction inst, ValueType vt> : Pat <
+  (vt (AMDGPUclamp
+        (VOP3Mods0Clamp vt:$src0, i32:$src0_modifiers, i32:$omod))),
+  (inst i32:$src0_modifiers, vt:$src0,
+        i32:$src0_modifiers, vt:$src0, DSTCLAMP.ENABLE, $omod)
 >;
 
+// TODO: Does f64 support clamp?
+def : ClampPat<V_MAX_F32_e64, f32>;
+def : ClampPat<V_MAX_F16_e64, f16>;
+
 /********** ================================ **********/
 /********** Floating point absolute/negative **********/
 /********** ================================ **********/