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.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@295788 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/lib/Target/AMDGPU/AMDGPU.td b/lib/Target/AMDGPU/AMDGPU.td
index 0de72734b71b985796f1c4b5bcf424fc7534d98c..17d6b460a80aeae263c556aa7a7eb81bcbf4383c 100644 (file)
--- a/lib/Target/AMDGPU/AMDGPU.td
+++ b/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/lib/Target/AMDGPU/AMDGPUAsmPrinter.cpp b/lib/Target/AMDGPU/AMDGPUAsmPrinter.cpp
index c3ff2e322546151a19d2ba86461ce3448aefd51f..577509d31ce94073252ffa6d803da737c222bd3e 100644 (file)
--- a/lib/Target/AMDGPU/AMDGPUAsmPrinter.cpp
+++ b/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/lib/Target/AMDGPU/AMDGPUISelLowering.cpp b/lib/Target/AMDGPU/AMDGPUISelLowering.cpp
index 12c17c3e9eb565b9e0fc1dc218c0de81c2861b31..9e3ac8b871e6340c8e24e0b941d598804a739d94 100644 (file)
--- a/lib/Target/AMDGPU/AMDGPUISelLowering.cpp
+++ b/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/lib/Target/AMDGPU/AMDGPUISelLowering.h b/lib/Target/AMDGPU/AMDGPUISelLowering.h
index 58ac09f2698a1bfdb95e2f2f7c61210d1365f1fa..fb487f914ad92497791cbc2c2ef1da2ac8bd9787 100644 (file)
--- a/lib/Target/AMDGPU/AMDGPUISelLowering.h
+++ b/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/lib/Target/AMDGPU/AMDGPUInstrInfo.td b/lib/Target/AMDGPU/AMDGPUInstrInfo.td
index 91341ef88ea695c9ce9bf2d86709fbd0f3bb44a0..a081c30ef815d676978d3d5ddef8fe046045225b 100644 (file)
--- a/lib/Target/AMDGPU/AMDGPUInstrInfo.td
+++ b/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/lib/Target/AMDGPU/AMDGPUInstructions.td b/lib/Target/AMDGPU/AMDGPUInstructions.td
index 7d2a52bccc8a4faa987bf32a3094e8a67597ef91..e76891c3bed09ea2bc320522bf272bb4ef30118b 100644 (file)
--- a/lib/Target/AMDGPU/AMDGPUInstructions.td
+++ b/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/lib/Target/AMDGPU/AMDGPUSubtarget.cpp b/lib/Target/AMDGPU/AMDGPUSubtarget.cpp
index 32b04fe7a1e162417e8b1f41df39d20595d68662..a719931eff0b3443ff23f2767cfbd4e199a4fac7 100644 (file)
--- a/lib/Target/AMDGPU/AMDGPUSubtarget.cpp
+++ b/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/lib/Target/AMDGPU/AMDGPUSubtarget.h b/lib/Target/AMDGPU/AMDGPUSubtarget.h
index 433c295feaf88bf5480595b0bc27d4b232692bdb..8f1aaa1d8939c0a4b8755cc6285425b469c3a7db 100644 (file)
--- a/lib/Target/AMDGPU/AMDGPUSubtarget.h
+++ b/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/lib/Target/AMDGPU/SIISelLowering.cpp b/lib/Target/AMDGPU/SIISelLowering.cpp
index 60acf4fb1c011824a7cca72b60efd3575570e02a..40a28203000603189087a2f2527f71f48bd93c57 100644 (file)
--- a/lib/Target/AMDGPU/SIISelLowering.cpp
+++ b/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/lib/Target/AMDGPU/SIISelLowering.h b/lib/Target/AMDGPU/SIISelLowering.h
index 037b6f730c5bf8cb9029c13b111923ef1ccd661a..fe86f8e2b65587ec27249b48ad803e2aa39ee815 100644 (file)
--- a/lib/Target/AMDGPU/SIISelLowering.h
+++ b/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/lib/Target/AMDGPU/SIInstrInfo.td b/lib/Target/AMDGPU/SIInstrInfo.td
index 94580543afde18016e1afbcb4330120d3578979d..aa6d677f466c6dfc1d16dc65cfe799a1423de7ff 100644 (file)
--- a/lib/Target/AMDGPU/SIInstrInfo.td
+++ b/lib/Target/AMDGPU/SIInstrInfo.td
@@ -625,6 +625,7 @@ def SRCMODS {
 
 def DSTCLAMP {
   int NONE = 0;
+  int ENABLE = 1;
 }
 
 def DSTOMOD {

diff --git a/lib/Target/AMDGPU/SIInstructions.td b/lib/Target/AMDGPU/SIInstructions.td
index 1d09880147d8a20c585727d27a911b1890b21812..4bd759012a395b90eb8401b1a7748cbe8ec2bff5 100644 (file)
--- a/lib/Target/AMDGPU/SIInstructions.td
+++ b/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 **********/
 /********** ================================ **********/

diff --git a/test/CodeGen/AMDGPU/clamp.ll b/test/CodeGen/AMDGPU/clamp.ll
new file mode 100644 (file)
index 0000000..5f9884e
--- /dev/null
+++ b/test/CodeGen/AMDGPU/clamp.ll
@@ -0,0 +1,535 @@
+; RUN: llc -march=amdgcn -verify-machineinstrs < %s | FileCheck -check-prefix=GCN -check-prefix=SI %s
+; RUN: llc -march=amdgcn -mcpu=fiji -verify-machineinstrs < %s | FileCheck -check-prefix=GCN -check-prefix=VI %s
+
+; GCN-LABEL: {{^}}v_clamp_f32:
+; GCN: {{buffer|flat}}_load_dword [[A:v[0-9]+]]
+; GCN: v_max_f32_e64 v{{[0-9]+}}, [[A]], [[A]] clamp{{$}}
+define amdgpu_kernel void @v_clamp_f32(float addrspace(1)* %out, float addrspace(1)* %aptr) #0 {
+  %tid = call i32 @llvm.amdgcn.workitem.id.x()
+  %gep0 = getelementptr float, float addrspace(1)* %aptr, i32 %tid
+  %out.gep = getelementptr float, float addrspace(1)* %out, i32 %tid
+  %a = load float, float addrspace(1)* %gep0
+  %max = call float @llvm.maxnum.f32(float %a, float 0.0)
+  %med = call float @llvm.minnum.f32(float %max, float 1.0)
+
+  store float %med, float addrspace(1)* %out.gep
+  ret void
+}
+
+; GCN-LABEL: {{^}}v_clamp_neg_f32:
+; GCN: {{buffer|flat}}_load_dword [[A:v[0-9]+]]
+; GCN: v_max_f32_e64 v{{[0-9]+}}, -[[A]], -[[A]] clamp{{$}}
+define amdgpu_kernel void @v_clamp_neg_f32(float addrspace(1)* %out, float addrspace(1)* %aptr) #0 {
+  %tid = call i32 @llvm.amdgcn.workitem.id.x()
+  %gep0 = getelementptr float, float addrspace(1)* %aptr, i32 %tid
+  %out.gep = getelementptr float, float addrspace(1)* %out, i32 %tid
+  %a = load float, float addrspace(1)* %gep0
+  %fneg.a = fsub float -0.0, %a
+  %max = call float @llvm.maxnum.f32(float %fneg.a, float 0.0)
+  %med = call float @llvm.minnum.f32(float %max, float 1.0)
+
+  store float %med, float addrspace(1)* %out.gep
+  ret void
+}
+
+; GCN-LABEL: {{^}}v_clamp_negabs_f32:
+; GCN: {{buffer|flat}}_load_dword [[A:v[0-9]+]]
+; GCN: v_max_f32_e64 v{{[0-9]+}}, -|[[A]]|, -|[[A]]| clamp{{$}}
+define amdgpu_kernel void @v_clamp_negabs_f32(float addrspace(1)* %out, float addrspace(1)* %aptr) #0 {
+  %tid = call i32 @llvm.amdgcn.workitem.id.x()
+  %gep0 = getelementptr float, float addrspace(1)* %aptr, i32 %tid
+  %out.gep = getelementptr float, float addrspace(1)* %out, i32 %tid
+  %a = load float, float addrspace(1)* %gep0
+  %fabs.a = call float @llvm.fabs.f32(float %a)
+  %fneg.fabs.a = fsub float -0.0, %fabs.a
+
+  %max = call float @llvm.maxnum.f32(float %fneg.fabs.a, float 0.0)
+  %med = call float @llvm.minnum.f32(float %max, float 1.0)
+
+  store float %med, float addrspace(1)* %out.gep
+  ret void
+}
+
+; GCN-LABEL: {{^}}v_clamp_negzero_f32:
+; GCN: {{buffer|flat}}_load_dword [[A:v[0-9]+]]
+; GCN: v_bfrev_b32_e32 [[SIGNBIT:v[0-9]+]], 1
+; GCN: v_med3_f32 v{{[0-9]+}}, [[A]], [[SIGNBIT]], 1.0
+define amdgpu_kernel void @v_clamp_negzero_f32(float addrspace(1)* %out, float addrspace(1)* %aptr) #0 {
+  %tid = call i32 @llvm.amdgcn.workitem.id.x()
+  %gep0 = getelementptr float, float addrspace(1)* %aptr, i32 %tid
+  %out.gep = getelementptr float, float addrspace(1)* %out, i32 %tid
+  %a = load float, float addrspace(1)* %gep0
+  %max = call float @llvm.maxnum.f32(float %a, float -0.0)
+  %med = call float @llvm.minnum.f32(float %max, float 1.0)
+
+  store float %med, float addrspace(1)* %out.gep
+  ret void
+}
+
+; GCN-LABEL: {{^}}v_clamp_multi_use_max_f32:
+; GCN: {{buffer|flat}}_load_dword [[A:v[0-9]+]]
+; GCN: v_max_f32_e32 [[MAX:v[0-9]+]], 0, [[A]]
+; GCN: v_min_f32_e32 [[MIN:v[0-9]+]], 1.0, [[MAX]]
+define amdgpu_kernel void @v_clamp_multi_use_max_f32(float addrspace(1)* %out, float addrspace(1)* %aptr) #0 {
+  %tid = call i32 @llvm.amdgcn.workitem.id.x()
+  %gep0 = getelementptr float, float addrspace(1)* %aptr, i32 %tid
+  %out.gep = getelementptr float, float addrspace(1)* %out, i32 %tid
+  %a = load float, float addrspace(1)* %gep0
+  %max = call float @llvm.maxnum.f32(float %a, float 0.0)
+  %med = call float @llvm.minnum.f32(float %max, float 1.0)
+
+  store float %med, float addrspace(1)* %out.gep
+  store volatile float %max, float addrspace(1)* undef
+  ret void
+}
+
+; GCN-LABEL: {{^}}v_clamp_f16:
+; GCN: {{buffer|flat}}_load_ushort [[A:v[0-9]+]]
+; VI: v_max_f16_e64 v{{[0-9]+}}, [[A]], [[A]] clamp{{$}}
+
+; SI: v_cvt_f32_f16_e32 [[CVT:v[0-9]+]], [[A]]
+; SI: v_max_f32_e64 v{{[0-9]+}}, [[CVT]], [[CVT]] clamp{{$}}
+; SI: v_cvt_f16_f32_e32
+define amdgpu_kernel void @v_clamp_f16(half addrspace(1)* %out, half addrspace(1)* %aptr) #0 {
+  %tid = call i32 @llvm.amdgcn.workitem.id.x()
+  %gep0 = getelementptr half, half addrspace(1)* %aptr, i32 %tid
+  %out.gep = getelementptr half, half addrspace(1)* %out, i32 %tid
+  %a = load half, half addrspace(1)* %gep0
+  %max = call half @llvm.maxnum.f16(half %a, half 0.0)
+  %med = call half @llvm.minnum.f16(half %max, half 1.0)
+
+  store half %med, half addrspace(1)* %out.gep
+  ret void
+}
+
+; GCN-LABEL: {{^}}v_clamp_neg_f16:
+; GCN: {{buffer|flat}}_load_ushort [[A:v[0-9]+]]
+; VI: v_max_f16_e64 v{{[0-9]+}}, -[[A]], -[[A]] clamp{{$}}
+
+; FIXME: Better to fold neg into max
+; SI: v_cvt_f32_f16_e64 [[CVT:v[0-9]+]], -[[A]]
+; SI: v_max_f32_e64 v{{[0-9]+}}, [[CVT]], [[CVT]] clamp{{$}}
+; SI: v_cvt_f16_f32
+define amdgpu_kernel void @v_clamp_neg_f16(half addrspace(1)* %out, half addrspace(1)* %aptr) #0 {
+  %tid = call i32 @llvm.amdgcn.workitem.id.x()
+  %gep0 = getelementptr half, half addrspace(1)* %aptr, i32 %tid
+  %out.gep = getelementptr half, half addrspace(1)* %out, i32 %tid
+  %a = load half, half addrspace(1)* %gep0
+  %fneg.a = fsub half -0.0, %a
+  %max = call half @llvm.maxnum.f16(half %fneg.a, half 0.0)
+  %med = call half @llvm.minnum.f16(half %max, half 1.0)
+
+  store half %med, half addrspace(1)* %out.gep
+  ret void
+}
+
+; GCN-LABEL: {{^}}v_clamp_negabs_f16:
+; GCN: {{buffer|flat}}_load_ushort [[A:v[0-9]+]]
+; VI: v_max_f16_e64 v{{[0-9]+}}, -|[[A]]|, -|[[A]]| clamp{{$}}
+
+; FIXME: Better to fold neg/abs into max
+
+; SI: v_cvt_f32_f16_e64 [[CVT:v[0-9]+]], -|[[A]]|
+; SI: v_max_f32_e64 v{{[0-9]+}}, [[CVT]], [[CVT]] clamp{{$}}
+; SI: v_cvt_f16_f32_e32
+define amdgpu_kernel void @v_clamp_negabs_f16(half addrspace(1)* %out, half addrspace(1)* %aptr) #0 {
+  %tid = call i32 @llvm.amdgcn.workitem.id.x()
+  %gep0 = getelementptr half, half addrspace(1)* %aptr, i32 %tid
+  %out.gep = getelementptr half, half addrspace(1)* %out, i32 %tid
+  %a = load half, half addrspace(1)* %gep0
+  %fabs.a = call half @llvm.fabs.f16(half %a)
+  %fneg.fabs.a = fsub half -0.0, %fabs.a
+
+  %max = call half @llvm.maxnum.f16(half %fneg.fabs.a, half 0.0)
+  %med = call half @llvm.minnum.f16(half %max, half 1.0)
+
+  store half %med, half addrspace(1)* %out.gep
+  ret void
+}
+
+; FIXME: Do f64 instructions support clamp?
+; GCN-LABEL: {{^}}v_clamp_f64:
+; GCN: {{buffer|flat}}_load_dwordx2 [[A:v\[[0-9]+:[0-9]+\]]]
+; GCN: v_max_f64
+; GCN: v_min_f64
+define amdgpu_kernel void @v_clamp_f64(double addrspace(1)* %out, double addrspace(1)* %aptr) #0 {
+  %tid = call i32 @llvm.amdgcn.workitem.id.x()
+  %gep0 = getelementptr double, double addrspace(1)* %aptr, i32 %tid
+  %out.gep = getelementptr double, double addrspace(1)* %out, i32 %tid
+  %a = load double, double addrspace(1)* %gep0
+  %max = call double @llvm.maxnum.f64(double %a, double 0.0)
+  %med = call double @llvm.minnum.f64(double %max, double 1.0)
+
+  store double %med, double addrspace(1)* %out.gep
+  ret void
+}
+
+; GCN-LABEL: {{^}}v_clamp_neg_f64:
+; GCN: {{buffer|flat}}_load_dwordx2 [[A:v\[[0-9]+:[0-9]+\]]]
+; GCN: v_max_f64
+; GCN: v_min_f64
+define amdgpu_kernel void @v_clamp_neg_f64(double addrspace(1)* %out, double addrspace(1)* %aptr) #0 {
+  %tid = call i32 @llvm.amdgcn.workitem.id.x()
+  %gep0 = getelementptr double, double addrspace(1)* %aptr, i32 %tid
+  %out.gep = getelementptr double, double addrspace(1)* %out, i32 %tid
+  %a = load double, double addrspace(1)* %gep0
+  %fneg.a = fsub double -0.0, %a
+  %max = call double @llvm.maxnum.f64(double %fneg.a, double 0.0)
+  %med = call double @llvm.minnum.f64(double %max, double 1.0)
+
+  store double %med, double addrspace(1)* %out.gep
+  ret void
+}
+
+; GCN-LABEL: {{^}}v_clamp_negabs_f64:
+; GCN: {{buffer|flat}}_load_dwordx2 [[A:v\[[0-9]+:[0-9]+\]]]
+; GCN: v_max_f64
+; GCN: v_min_f64
+define amdgpu_kernel void @v_clamp_negabs_f64(double addrspace(1)* %out, double addrspace(1)* %aptr) #0 {
+  %tid = call i32 @llvm.amdgcn.workitem.id.x()
+  %gep0 = getelementptr double, double addrspace(1)* %aptr, i32 %tid
+  %out.gep = getelementptr double, double addrspace(1)* %out, i32 %tid
+  %a = load double, double addrspace(1)* %gep0
+  %fabs.a = call double @llvm.fabs.f64(double %a)
+  %fneg.fabs.a = fsub double -0.0, %fabs.a
+
+  %max = call double @llvm.maxnum.f64(double %fneg.fabs.a, double 0.0)
+  %med = call double @llvm.minnum.f64(double %max, double 1.0)
+
+  store double %med, double addrspace(1)* %out.gep
+  ret void
+}
+
+; GCN-LABEL: {{^}}v_clamp_med3_aby_negzero_f32:
+; GCN: {{buffer|flat}}_load_dword [[A:v[0-9]+]]
+; GCN: v_med3_f32
+define amdgpu_kernel void @v_clamp_med3_aby_negzero_f32(float addrspace(1)* %out, float addrspace(1)* %aptr) #0 {
+  %tid = call i32 @llvm.amdgcn.workitem.id.x()
+  %gep0 = getelementptr float, float addrspace(1)* %aptr, i32 %tid
+  %out.gep = getelementptr float, float addrspace(1)* %out, i32 %tid
+  %a = load float, float addrspace(1)* %gep0
+  %med = call float @llvm.amdgcn.fmed3.f32(float -0.0, float 1.0, float %a)
+  store float %med, float addrspace(1)* %out.gep
+  ret void
+}
+
+; GCN-LABEL: {{^}}v_clamp_med3_aby_f32:
+; GCN: {{buffer|flat}}_load_dword [[A:v[0-9]+]]
+; GCN: v_max_f32_e64 v{{[0-9]+}}, [[A]], [[A]] clamp{{$}}
+define amdgpu_kernel void @v_clamp_med3_aby_f32(float addrspace(1)* %out, float addrspace(1)* %aptr) #0 {
+  %tid = call i32 @llvm.amdgcn.workitem.id.x()
+  %gep0 = getelementptr float, float addrspace(1)* %aptr, i32 %tid
+  %out.gep = getelementptr float, float addrspace(1)* %out, i32 %tid
+  %a = load float, float addrspace(1)* %gep0
+  %med = call float @llvm.amdgcn.fmed3.f32(float 0.0, float 1.0, float %a)
+  store float %med, float addrspace(1)* %out.gep
+  ret void
+}
+
+; GCN-LABEL: {{^}}v_clamp_med3_bay_f32:
+; GCN: {{buffer|flat}}_load_dword [[A:v[0-9]+]]
+; GCN: v_max_f32_e64 v{{[0-9]+}}, [[A]], [[A]] clamp{{$}}
+define amdgpu_kernel void @v_clamp_med3_bay_f32(float addrspace(1)* %out, float addrspace(1)* %aptr) #0 {
+  %tid = call i32 @llvm.amdgcn.workitem.id.x()
+  %gep0 = getelementptr float, float addrspace(1)* %aptr, i32 %tid
+  %out.gep = getelementptr float, float addrspace(1)* %out, i32 %tid
+  %a = load float, float addrspace(1)* %gep0
+  %med = call float @llvm.amdgcn.fmed3.f32(float 1.0, float 0.0, float %a)
+  store float %med, float addrspace(1)* %out.gep
+  ret void
+}
+
+; GCN-LABEL: {{^}}v_clamp_med3_yab_f32:
+; GCN: {{buffer|flat}}_load_dword [[A:v[0-9]+]]
+; GCN: v_max_f32_e64 v{{[0-9]+}}, [[A]], [[A]] clamp{{$}}
+define amdgpu_kernel void @v_clamp_med3_yab_f32(float addrspace(1)* %out, float addrspace(1)* %aptr) #0 {
+  %tid = call i32 @llvm.amdgcn.workitem.id.x()
+  %gep0 = getelementptr float, float addrspace(1)* %aptr, i32 %tid
+  %out.gep = getelementptr float, float addrspace(1)* %out, i32 %tid
+  %a = load float, float addrspace(1)* %gep0
+  %med = call float @llvm.amdgcn.fmed3.f32(float %a, float 0.0, float 1.0)
+  store float %med, float addrspace(1)* %out.gep
+  ret void
+}
+
+; GCN-LABEL: {{^}}v_clamp_med3_yba_f32:
+; GCN: {{buffer|flat}}_load_dword [[A:v[0-9]+]]
+; GCN: v_max_f32_e64 v{{[0-9]+}}, [[A]], [[A]] clamp{{$}}
+define amdgpu_kernel void @v_clamp_med3_yba_f32(float addrspace(1)* %out, float addrspace(1)* %aptr) #0 {
+  %tid = call i32 @llvm.amdgcn.workitem.id.x()
+  %gep0 = getelementptr float, float addrspace(1)* %aptr, i32 %tid
+  %out.gep = getelementptr float, float addrspace(1)* %out, i32 %tid
+  %a = load float, float addrspace(1)* %gep0
+  %med = call float @llvm.amdgcn.fmed3.f32(float %a, float 1.0, float 0.0)
+  store float %med, float addrspace(1)* %out.gep
+  ret void
+}
+
+; GCN-LABEL: {{^}}v_clamp_med3_ayb_f32:
+; GCN: {{buffer|flat}}_load_dword [[A:v[0-9]+]]
+; GCN: v_max_f32_e64 v{{[0-9]+}}, [[A]], [[A]] clamp{{$}}
+define amdgpu_kernel void @v_clamp_med3_ayb_f32(float addrspace(1)* %out, float addrspace(1)* %aptr) #0 {
+  %tid = call i32 @llvm.amdgcn.workitem.id.x()
+  %gep0 = getelementptr float, float addrspace(1)* %aptr, i32 %tid
+  %out.gep = getelementptr float, float addrspace(1)* %out, i32 %tid
+  %a = load float, float addrspace(1)* %gep0
+  %med = call float @llvm.amdgcn.fmed3.f32(float 0.0, float %a, float 1.0)
+  store float %med, float addrspace(1)* %out.gep
+  ret void
+}
+
+; GCN-LABEL: {{^}}v_clamp_med3_bya_f32:
+; GCN: {{buffer|flat}}_load_dword [[A:v[0-9]+]]
+; GCN: v_max_f32_e64 v{{[0-9]+}}, [[A]], [[A]] clamp{{$}}
+define amdgpu_kernel void @v_clamp_med3_bya_f32(float addrspace(1)* %out, float addrspace(1)* %aptr) #0 {
+  %tid = call i32 @llvm.amdgcn.workitem.id.x()
+  %gep0 = getelementptr float, float addrspace(1)* %aptr, i32 %tid
+  %out.gep = getelementptr float, float addrspace(1)* %out, i32 %tid
+  %a = load float, float addrspace(1)* %gep0
+  %med = call float @llvm.amdgcn.fmed3.f32(float 1.0, float %a, float 0.0)
+  store float %med, float addrspace(1)* %out.gep
+  ret void
+}
+
+; GCN-LABEL: {{^}}v_clamp_constants_to_one_f32:
+; GCN: v_mov_b32_e32 v{{[0-9]+}}, 1.0
+define amdgpu_kernel void @v_clamp_constants_to_one_f32(float addrspace(1)* %out) #0 {
+  %tid = call i32 @llvm.amdgcn.workitem.id.x()
+  %out.gep = getelementptr float, float addrspace(1)* %out, i32 %tid
+  %med = call float @llvm.amdgcn.fmed3.f32(float 0.0, float 1.0, float 4.0)
+  store float %med, float addrspace(1)* %out.gep
+  ret void
+}
+
+; GCN-LABEL: {{^}}v_clamp_constants_to_zero_f32:
+; GCN: v_mov_b32_e32 v{{[0-9]+}}, 0{{$}}
+define amdgpu_kernel void @v_clamp_constants_to_zero_f32(float addrspace(1)* %out) #0 {
+  %tid = call i32 @llvm.amdgcn.workitem.id.x()
+  %out.gep = getelementptr float, float addrspace(1)* %out, i32 %tid
+  %med = call float @llvm.amdgcn.fmed3.f32(float 0.0, float 1.0, float -4.0)
+  store float %med, float addrspace(1)* %out.gep
+  ret void
+}
+
+; GCN-LABEL: {{^}}v_clamp_constant_preserve_f32:
+; GCN: v_mov_b32_e32 v{{[0-9]+}}, 0.5
+define amdgpu_kernel void @v_clamp_constant_preserve_f32(float addrspace(1)* %out) #0 {
+  %tid = call i32 @llvm.amdgcn.workitem.id.x()
+  %out.gep = getelementptr float, float addrspace(1)* %out, i32 %tid
+  %med = call float @llvm.amdgcn.fmed3.f32(float 0.0, float 1.0, float 0.5)
+  store float %med, float addrspace(1)* %out.gep
+  ret void
+}
+
+; GCN-LABEL: {{^}}v_clamp_constant_preserve_denorm_f32:
+; GCN: v_mov_b32_e32 v{{[0-9]+}}, 0x7fffff{{$}}
+define amdgpu_kernel void @v_clamp_constant_preserve_denorm_f32(float addrspace(1)* %out) #0 {
+  %tid = call i32 @llvm.amdgcn.workitem.id.x()
+  %out.gep = getelementptr float, float addrspace(1)* %out, i32 %tid
+  %med = call float @llvm.amdgcn.fmed3.f32(float 0.0, float 1.0, float bitcast (i32 8388607 to float))
+  store float %med, float addrspace(1)* %out.gep
+  ret void
+}
+
+; GCN-LABEL: {{^}}v_clamp_constant_qnan_f32:
+; GCN: v_mov_b32_e32 v{{[0-9]+}}, 0{{$}}
+define amdgpu_kernel void @v_clamp_constant_qnan_f32(float addrspace(1)* %out) #0 {
+  %tid = call i32 @llvm.amdgcn.workitem.id.x()
+  %out.gep = getelementptr float, float addrspace(1)* %out, i32 %tid
+  %med = call float @llvm.amdgcn.fmed3.f32(float 0.0, float 1.0, float 0x7FF8000000000000)
+  store float %med, float addrspace(1)* %out.gep
+  ret void
+}
+
+; GCN-LABEL: {{^}}v_clamp_constant_snan_f32:
+; GCN: v_mov_b32_e32 v{{[0-9]+}}, 0{{$}}
+define amdgpu_kernel void @v_clamp_constant_snan_f32(float addrspace(1)* %out) #0 {
+  %tid = call i32 @llvm.amdgcn.workitem.id.x()
+  %out.gep = getelementptr float, float addrspace(1)* %out, i32 %tid
+  %med = call float @llvm.amdgcn.fmed3.f32(float 0.0, float 1.0, float bitcast (i32 2139095041 to float))
+  store float %med, float addrspace(1)* %out.gep
+  ret void
+}
+
+; ---------------------------------------------------------------------
+; Test non-default behaviors enabling snans and disabling dx10_clamp
+; ---------------------------------------------------------------------
+
+; GCN-LABEL: {{^}}v_clamp_f32_no_dx10_clamp:
+; GCN: {{buffer|flat}}_load_dword [[A:v[0-9]+]]
+; GCN: v_med3_f32 v{{[0-9]+}}, [[A]], 0, 1.0
+define amdgpu_kernel void @v_clamp_f32_no_dx10_clamp(float addrspace(1)* %out, float addrspace(1)* %aptr) #2 {
+  %tid = call i32 @llvm.amdgcn.workitem.id.x()
+  %gep0 = getelementptr float, float addrspace(1)* %aptr, i32 %tid
+  %out.gep = getelementptr float, float addrspace(1)* %out, i32 %tid
+  %a = load float, float addrspace(1)* %gep0
+  %max = call float @llvm.maxnum.f32(float %a, float 0.0)
+  %med = call float @llvm.minnum.f32(float %max, float 1.0)
+
+  store float %med, float addrspace(1)* %out.gep
+  ret void
+}
+
+; GCN-LABEL: {{^}}v_clamp_f32_snan_dx10clamp:
+; GCN: {{buffer|flat}}_load_dword [[A:v[0-9]+]]
+; GCN: v_max_f32_e64 v{{[0-9]+}}, [[A]], [[A]] clamp{{$}}
+define amdgpu_kernel void @v_clamp_f32_snan_dx10clamp(float addrspace(1)* %out, float addrspace(1)* %aptr) #3 {
+  %tid = call i32 @llvm.amdgcn.workitem.id.x()
+  %gep0 = getelementptr float, float addrspace(1)* %aptr, i32 %tid
+  %out.gep = getelementptr float, float addrspace(1)* %out, i32 %tid
+  %a = load float, float addrspace(1)* %gep0
+  %max = call float @llvm.maxnum.f32(float %a, float 0.0)
+  %med = call float @llvm.minnum.f32(float %max, float 1.0)
+
+  store float %med, float addrspace(1)* %out.gep
+  ret void
+}
+
+; GCN-LABEL: {{^}}v_clamp_f32_snan_no_dx10clamp:
+; GCN: {{buffer|flat}}_load_dword [[A:v[0-9]+]]
+; GCN: v_max_f32_e32 [[MAX:v[0-9]+]], 0, [[A]]
+; GCN: v_min_f32_e32 [[MIN:v[0-9]+]], 1.0, [[MAX]]
+define amdgpu_kernel void @v_clamp_f32_snan_no_dx10clamp(float addrspace(1)* %out, float addrspace(1)* %aptr) #4 {
+  %tid = call i32 @llvm.amdgcn.workitem.id.x()
+  %gep0 = getelementptr float, float addrspace(1)* %aptr, i32 %tid
+  %out.gep = getelementptr float, float addrspace(1)* %out, i32 %tid
+  %a = load float, float addrspace(1)* %gep0
+  %max = call float @llvm.maxnum.f32(float %a, float 0.0)
+  %med = call float @llvm.minnum.f32(float %max, float 1.0)
+
+  store float %med, float addrspace(1)* %out.gep
+  ret void
+}
+
+; GCN-LABEL: {{^}}v_clamp_f32_snan_no_dx10clamp_nnan_src:
+; GCN: {{buffer|flat}}_load_dword [[A:v[0-9]+]]
+; GCN: v_med3_f32 v{{[0-9]+}}, [[A]], 0, 1.0
+define amdgpu_kernel void @v_clamp_f32_snan_no_dx10clamp_nnan_src(float addrspace(1)* %out, float addrspace(1)* %aptr) #4 {
+  %tid = call i32 @llvm.amdgcn.workitem.id.x()
+  %gep0 = getelementptr float, float addrspace(1)* %aptr, i32 %tid
+  %out.gep = getelementptr float, float addrspace(1)* %out, i32 %tid
+  %a = load float, float addrspace(1)* %gep0
+  %add  = fadd nnan float %a, 1.0
+  %max = call float @llvm.maxnum.f32(float %add, float 0.0)
+  %med = call float @llvm.minnum.f32(float %max, float 1.0)
+
+  store float %med, float addrspace(1)* %out.gep
+  ret void
+}
+
+; GCN-LABEL: {{^}}v_clamp_med3_aby_f32_no_dx10_clamp:
+; GCN: {{buffer|flat}}_load_dword [[A:v[0-9]+]]
+; GCN: v_max_f32_e64 v{{[0-9]+}}, [[A]], [[A]] clamp{{$}}
+define amdgpu_kernel void @v_clamp_med3_aby_f32_no_dx10_clamp(float addrspace(1)* %out, float addrspace(1)* %aptr) #2 {
+  %tid = call i32 @llvm.amdgcn.workitem.id.x()
+  %gep0 = getelementptr float, float addrspace(1)* %aptr, i32 %tid
+  %out.gep = getelementptr float, float addrspace(1)* %out, i32 %tid
+  %a = load float, float addrspace(1)* %gep0
+  %med = call float @llvm.amdgcn.fmed3.f32(float 0.0, float 1.0, float %a)
+  store float %med, float addrspace(1)* %out.gep
+  ret void
+}
+
+; GCN-LABEL: {{^}}v_clamp_med3_bay_f32_no_dx10_clamp:
+; GCN: {{buffer|flat}}_load_dword [[A:v[0-9]+]]
+; GCN: v_max_f32_e64 v{{[0-9]+}}, [[A]], [[A]] clamp{{$}}
+define amdgpu_kernel void @v_clamp_med3_bay_f32_no_dx10_clamp(float addrspace(1)* %out, float addrspace(1)* %aptr) #2 {
+  %tid = call i32 @llvm.amdgcn.workitem.id.x()
+  %gep0 = getelementptr float, float addrspace(1)* %aptr, i32 %tid
+  %out.gep = getelementptr float, float addrspace(1)* %out, i32 %tid
+  %a = load float, float addrspace(1)* %gep0
+  %med = call float @llvm.amdgcn.fmed3.f32(float 1.0, float 0.0, float %a)
+  store float %med, float addrspace(1)* %out.gep
+  ret void
+}
+
+; GCN-LABEL: {{^}}v_clamp_med3_yab_f32_no_dx10_clamp:
+; GCN: {{buffer|flat}}_load_dword [[A:v[0-9]+]]
+; GCN: v_med3_f32 v{{[0-9]+}}, [[A]], 0, 1.0
+define amdgpu_kernel void @v_clamp_med3_yab_f32_no_dx10_clamp(float addrspace(1)* %out, float addrspace(1)* %aptr) #2 {
+  %tid = call i32 @llvm.amdgcn.workitem.id.x()
+  %gep0 = getelementptr float, float addrspace(1)* %aptr, i32 %tid
+  %out.gep = getelementptr float, float addrspace(1)* %out, i32 %tid
+  %a = load float, float addrspace(1)* %gep0
+  %med = call float @llvm.amdgcn.fmed3.f32(float %a, float 0.0, float 1.0)
+  store float %med, float addrspace(1)* %out.gep
+  ret void
+}
+
+; GCN-LABEL: {{^}}v_clamp_med3_yba_f32_no_dx10_clamp:
+; GCN: {{buffer|flat}}_load_dword [[A:v[0-9]+]]
+; GCN: v_med3_f32 v{{[0-9]+}}, [[A]], 1.0, 0
+define amdgpu_kernel void @v_clamp_med3_yba_f32_no_dx10_clamp(float addrspace(1)* %out, float addrspace(1)* %aptr) #2 {
+  %tid = call i32 @llvm.amdgcn.workitem.id.x()
+  %gep0 = getelementptr float, float addrspace(1)* %aptr, i32 %tid
+  %out.gep = getelementptr float, float addrspace(1)* %out, i32 %tid
+  %a = load float, float addrspace(1)* %gep0
+  %med = call float @llvm.amdgcn.fmed3.f32(float %a, float 1.0, float 0.0)
+  store float %med, float addrspace(1)* %out.gep
+  ret void
+}
+
+; GCN-LABEL: {{^}}v_clamp_med3_ayb_f32_no_dx10_clamp:
+; GCN: {{buffer|flat}}_load_dword [[A:v[0-9]+]]
+; GCN: v_med3_f32 v{{[0-9]+}}, 0, [[A]], 1.0
+define amdgpu_kernel void @v_clamp_med3_ayb_f32_no_dx10_clamp(float addrspace(1)* %out, float addrspace(1)* %aptr) #2 {
+  %tid = call i32 @llvm.amdgcn.workitem.id.x()
+  %gep0 = getelementptr float, float addrspace(1)* %aptr, i32 %tid
+  %out.gep = getelementptr float, float addrspace(1)* %out, i32 %tid
+  %a = load float, float addrspace(1)* %gep0
+  %med = call float @llvm.amdgcn.fmed3.f32(float 0.0, float %a, float 1.0)
+  store float %med, float addrspace(1)* %out.gep
+  ret void
+}
+
+; GCN-LABEL: {{^}}v_clamp_med3_bya_f32_no_dx10_clamp:
+; GCN: {{buffer|flat}}_load_dword [[A:v[0-9]+]]
+; GCN: v_med3_f32 v{{[0-9]+}}, 1.0, [[A]], 0
+define amdgpu_kernel void @v_clamp_med3_bya_f32_no_dx10_clamp(float addrspace(1)* %out, float addrspace(1)* %aptr) #2 {
+  %tid = call i32 @llvm.amdgcn.workitem.id.x()
+  %gep0 = getelementptr float, float addrspace(1)* %aptr, i32 %tid
+  %out.gep = getelementptr float, float addrspace(1)* %out, i32 %tid
+  %a = load float, float addrspace(1)* %gep0
+  %med = call float @llvm.amdgcn.fmed3.f32(float 1.0, float %a, float 0.0)
+  store float %med, float addrspace(1)* %out.gep
+  ret void
+}
+
+; GCN-LABEL: {{^}}v_clamp_constant_qnan_f32_no_dx10_clamp:
+; GCN: v_mov_b32_e32 v{{[0-9]+}}, 0x7fc00000
+define amdgpu_kernel void @v_clamp_constant_qnan_f32_no_dx10_clamp(float addrspace(1)* %out) #2 {
+  %tid = call i32 @llvm.amdgcn.workitem.id.x()
+  %out.gep = getelementptr float, float addrspace(1)* %out, i32 %tid
+  %med = call float @llvm.amdgcn.fmed3.f32(float 0.0, float 1.0, float 0x7FF8000000000000)
+  store float %med, float addrspace(1)* %out.gep
+  ret void
+}
+
+; GCN-LABEL: {{^}}v_clamp_constant_snan_f32_no_dx10_clamp:
+; GCN: v_mov_b32_e32 v{{[0-9]+}}, 0x7f800001
+define amdgpu_kernel void @v_clamp_constant_snan_f32_no_dx10_clamp(float addrspace(1)* %out) #2 {
+  %tid = call i32 @llvm.amdgcn.workitem.id.x()
+  %out.gep = getelementptr float, float addrspace(1)* %out, i32 %tid
+  %med = call float @llvm.amdgcn.fmed3.f32(float 0.0, float 1.0, float bitcast (i32 2139095041 to float))
+  store float %med, float addrspace(1)* %out.gep
+  ret void
+}
+
+declare i32 @llvm.amdgcn.workitem.id.x() #1
+declare float @llvm.fabs.f32(float) #1
+declare float @llvm.minnum.f32(float, float) #1
+declare float @llvm.maxnum.f32(float, float) #1
+declare float @llvm.amdgcn.fmed3.f32(float, float, float) #1
+declare double @llvm.fabs.f64(double) #1
+declare double @llvm.minnum.f64(double, double) #1
+declare double @llvm.maxnum.f64(double, double) #1
+declare half @llvm.fabs.f16(half) #1
+declare half @llvm.minnum.f16(half, half) #1
+declare half @llvm.maxnum.f16(half, half) #1
+
+
+attributes #0 = { nounwind }
+attributes #1 = { nounwind readnone }
+attributes #2 = { nounwind "target-features"="-dx10-clamp,-fp-exceptions" "no-nans-fp-math"="false" }
+attributes #3 = { nounwind "target-features"="+dx10-clamp,+fp-exceptions" "no-nans-fp-math"="false" }
+attributes #4 = { nounwind "target-features"="-dx10-clamp,+fp-exceptions" "no-nans-fp-math"="false" }

diff --git a/test/CodeGen/AMDGPU/hsa-fp-mode.ll b/test/CodeGen/AMDGPU/hsa-fp-mode.ll
index 977667cbe6c3148046bee7d2aac711420e66856a..ecb825a2f20c1765f16f55b3fe06695db82239da 100644 (file)
--- a/test/CodeGen/AMDGPU/hsa-fp-mode.ll
+++ b/test/CodeGen/AMDGPU/hsa-fp-mode.ll
@@ -60,9 +60,20 @@ define void @test_no_denormals(float addrspace(1)* %out0, double addrspace(1)* %
   ret void
 }
 
+; GCN-LABEL: {{^}}test_no_dx10_clamp_vi:
+; GCN: float_mode = 192
+; GCN: enable_dx10_clamp = 0
+; GCN: enable_ieee_mode = 1
+define void @test_no_dx10_clamp_vi(float addrspace(1)* %out0, double addrspace(1)* %out1) #6 {
+  store float 0.0, float addrspace(1)* %out0
+  store double 0.0, double addrspace(1)* %out1
+  ret void
+}
+
 attributes #0 = { nounwind "target-cpu"="kaveri" }
 attributes #1 = { nounwind "target-cpu"="fiji" }
 attributes #2 = { nounwind "target-features"="-fp32-denormals,+fp64-fp16-denormals" }
 attributes #3 = { nounwind "target-features"="+fp32-denormals,-fp64-fp16-denormals" }
 attributes #4 = { nounwind "target-features"="+fp32-denormals,+fp64-fp16-denormals" }
 attributes #5 = { nounwind "target-features"="-fp32-denormals,-fp64-fp16-denormals" }
+attributes #6 = { nounwind "target-cpu"="fiji" "target-features"="-dx10-clamp" }

diff --git a/test/CodeGen/AMDGPU/llvm.AMDGPU.clamp.ll b/test/CodeGen/AMDGPU/llvm.AMDGPU.clamp.ll
index 2336109f4dadf1ea9a2783845c42b8083b4dba93..1925a0b6fe4a08290134770e0f949ba375cda913 100644 (file)
--- a/test/CodeGen/AMDGPU/llvm.AMDGPU.clamp.ll
+++ b/test/CodeGen/AMDGPU/llvm.AMDGPU.clamp.ll
@@ -7,7 +7,7 @@ declare float @llvm.AMDGPU.clamp.f32(float, float, float) nounwind readnone
 
 ; FUNC-LABEL: {{^}}clamp_0_1_f32:
 ; SI: s_load_dword [[ARG:s[0-9]+]],
-; SI: v_add_f32_e64 [[RESULT:v[0-9]+]], [[ARG]], 0 clamp{{$}}
+; SI: v_max_f32_e64 [[RESULT:v[0-9]+]], [[ARG]], [[ARG]] clamp{{$}}
 ; SI: buffer_store_dword [[RESULT]]
 ; SI: s_endpgm
 
@@ -20,7 +20,7 @@ define void @clamp_0_1_f32(float addrspace(1)* %out, float %src) nounwind {
 
 ; FUNC-LABEL: {{^}}clamp_fabs_0_1_f32:
 ; SI: s_load_dword [[ARG:s[0-9]+]],
-; SI: v_add_f32_e64 [[RESULT:v[0-9]+]], |[[ARG]]|, 0 clamp{{$}}
+; SI: v_max_f32_e64 [[RESULT:v[0-9]+]], |[[ARG]]|, |[[ARG]]| clamp{{$}}
 ; SI: buffer_store_dword [[RESULT]]
 ; SI: s_endpgm
 define void @clamp_fabs_0_1_f32(float addrspace(1)* %out, float %src) nounwind {
@@ -32,7 +32,7 @@ define void @clamp_fabs_0_1_f32(float addrspace(1)* %out, float %src) nounwind {
 
 ; FUNC-LABEL: {{^}}clamp_fneg_0_1_f32:
 ; SI: s_load_dword [[ARG:s[0-9]+]],
-; SI: v_add_f32_e64 [[RESULT:v[0-9]+]], -[[ARG]], 0 clamp{{$}}
+; SI: v_max_f32_e64 [[RESULT:v[0-9]+]], -[[ARG]], -[[ARG]] clamp{{$}}
 ; SI: buffer_store_dword [[RESULT]]
 ; SI: s_endpgm
 define void @clamp_fneg_0_1_f32(float addrspace(1)* %out, float %src) nounwind {
@@ -44,7 +44,7 @@ define void @clamp_fneg_0_1_f32(float addrspace(1)* %out, float %src) nounwind {
 
 ; FUNC-LABEL: {{^}}clamp_fneg_fabs_0_1_f32:
 ; SI: s_load_dword [[ARG:s[0-9]+]],
-; SI: v_add_f32_e64 [[RESULT:v[0-9]+]], -|[[ARG]]|, 0 clamp{{$}}
+; SI: v_max_f32_e64 [[RESULT:v[0-9]+]], -|[[ARG]]|, -|[[ARG]]| clamp{{$}}
 ; SI: buffer_store_dword [[RESULT]]
 ; SI: s_endpgm
 define void @clamp_fneg_fabs_0_1_f32(float addrspace(1)* %out, float %src) nounwind {