[AMDGPU] Implement 'llvm.get.fpenv' and 'llvm.set.fpenv'

jhuber6 · jhuber6 · commit 169f89142707 · 2024-03-05T07:38:15.000-06:00
Summary:
This patch implements the LLVM floating point environment control
intrinsics and also exposes it through clang. We encode the floating
point environment as a 64-bit value that simply concatenates the values
of the mode registers and the current trap status. We only fetch the
bits relevant for floating point instructions. That is, rounding mode,
denormalization mode, ieee, dx10 clamp, debug, enabled traps, f16
overflow, and active exceptions.
diff --git a/clang/include/clang/Basic/BuiltinsAMDGPU.def b/clang/include/clang/Basic/BuiltinsAMDGPU.def
@@ -325,6 +325,9 @@ BUILTIN(__builtin_amdgcn_read_exec_hi, "Ui", "nc")
 
 BUILTIN(__builtin_amdgcn_endpgm, "v", "nr")
 
+BUILTIN(__builtin_amdgcn_get_fpenv, "WUi", "n")
+BUILTIN(__builtin_amdgcn_set_fpenv, "vWUi", "n")
+
 //===----------------------------------------------------------------------===//
 // R600-NI only builtins.
 //===----------------------------------------------------------------------===//
diff --git a/clang/lib/CodeGen/CGBuiltin.cpp b/clang/lib/CodeGen/CGBuiltin.cpp
@@ -18406,6 +18406,17 @@ Value *CodeGenFunction::EmitAMDGPUBuiltinExpr(unsigned BuiltinID,
         CGM.getIntrinsic(Intrinsic::amdgcn_global_load_tr, {ArgTy});
     return Builder.CreateCall(F, {Addr});
   }
+  case AMDGPU::BI__builtin_amdgcn_get_fpenv: {
+    Function *F = CGM.getIntrinsic(Intrinsic::get_fpenv,
+                                   {llvm::Type::getInt64Ty(getLLVMContext())});
+    return Builder.CreateCall(F);
+  }
+  case AMDGPU::BI__builtin_amdgcn_set_fpenv: {
+    Function *F = CGM.getIntrinsic(Intrinsic::set_fpenv,
+                                   {llvm::Type::getInt64Ty(getLLVMContext())});
+    llvm::Value *Env = EmitScalarExpr(E->getArg(0));
+    return Builder.CreateCall(F, {Env});
+  }
   case AMDGPU::BI__builtin_amdgcn_read_exec:
     return EmitAMDGCNBallotForExec(*this, E, Int64Ty, Int64Ty, false);
   case AMDGPU::BI__builtin_amdgcn_read_exec_lo:
diff --git a/clang/test/CodeGenOpenCL/builtins-amdgcn.cl b/clang/test/CodeGenOpenCL/builtins-amdgcn.cl
@@ -839,6 +839,18 @@ unsigned test_wavefrontsize() {
   return __builtin_amdgcn_wavefrontsize();
 }
 
+// CHECK-LABEL test_get_fpenv(
+unsigned long test_get_fpenv() {
+  // CHECK: call i64 @llvm.get.fpenv.i64()
+  return __builtin_amdgcn_get_fpenv();
+}
+
+// CHECK-LABEL test_set_fpenv(
+void test_set_fpenv(unsigned long env) {
+  // CHECK: call void @llvm.set.fpenv.i64(i64 %[[ENV:.+]])
+  __builtin_amdgcn_set_fpenv(env);
+}
+
 // CHECK-DAG: [[$WI_RANGE]] = !{i32 0, i32 1024}
 // CHECK-DAG: [[$WS_RANGE]] = !{i16 1, i16 1025}
 // CHECK-DAG: attributes #[[$NOUNWIND_READONLY]] = { convergent mustprogress nocallback nofree nounwind willreturn memory(none) }
diff --git a/llvm/docs/AMDGPUUsage.rst b/llvm/docs/AMDGPUUsage.rst
@@ -1151,6 +1151,13 @@ The AMDGPU backend implements the following LLVM IR intrinsics.
                                                    register do not exactly match the FLT_ROUNDS values,
                                                    so a conversion is performed.
 
+  :ref:`llvm.get.fpenv<int_get_fpenv>`             Returns the current value of the AMDGPU floating point environment.
+                                                   This stores information related to the current rounding mode,
+                                                   denormalization mode, enabled traps, and floating point exceptions.
+                                                   The format is a 64-bit concatenation of the MODE and TRAPSTS registers.
+
+  :ref:`llvm.set.fpenv<int_set_fpenv>`             Sets the floating point environment to the specifies state.
+
   llvm.amdgcn.wave.reduce.umin                     Performs an arithmetic unsigned min reduction on the unsigned values
                                                    provided by each lane in the wavefront.
                                                    Intrinsic takes a hint for reduction strategy using second operand
diff --git a/llvm/docs/LangRef.rst b/llvm/docs/LangRef.rst
@@ -26432,6 +26432,7 @@ similar to C library function 'fesetround', however this intrinsic does not
 return any value and uses platform-independent representation of IEEE rounding
 modes.
 
+.. _int_get_fpenv:
 
 '``llvm.get.fpenv``' Intrinsic
 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
@@ -26455,6 +26456,7 @@ Semantics:
 The '``llvm.get.fpenv``' intrinsic reads the current floating-point environment
 and returns it as an integer value.
 
+.. _int_set_fpenv:
 
 '``llvm.set.fpenv``' Intrinsic
 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
diff --git a/llvm/docs/ReleaseNotes.rst b/llvm/docs/ReleaseNotes.rst
@@ -70,6 +70,8 @@ Changes to the AArch64 Backend
 Changes to the AMDGPU Backend
 -----------------------------
 
+* Implemented the ``llvm.get.fpenv`` and ``llvm.set.fpenv`` intrinsics.
+
 Changes to the ARM Backend
 --------------------------
 
diff --git a/llvm/lib/Target/AMDGPU/AMDGPULegalizerInfo.cpp b/llvm/lib/Target/AMDGPU/AMDGPULegalizerInfo.cpp
@@ -905,6 +905,8 @@ AMDGPULegalizerInfo::AMDGPULegalizerInfo(const GCNSubtarget &ST_,
   getActionDefinitionsBuilder(G_STACKRESTORE)
     .legalFor({PrivatePtr});
 
+  getActionDefinitionsBuilder({G_GET_FPENV, G_SET_FPENV}).customFor({S64});
+
   getActionDefinitionsBuilder(G_GLOBAL_VALUE)
     .customIf(typeIsNot(0, PrivatePtr));
 
@@ -2128,6 +2130,10 @@ bool AMDGPULegalizerInfo::legalizeCustom(
     return legalizeFPTruncRound(MI, B);
   case TargetOpcode::G_STACKSAVE:
     return legalizeStackSave(MI, B);
+  case TargetOpcode::G_GET_FPENV:
+    return legalizeGetFPEnv(MI, B);
+  case TargetOpcode::G_SET_FPENV:
+    return legalizeSetFPEnv(MI, B);
   default:
     return false;
   }
@@ -6940,6 +6946,42 @@ bool AMDGPULegalizerInfo::legalizeWaveID(MachineInstr &MI,
   return true;
 }
 
+static constexpr unsigned FPEnvModeBitField =
+    AMDGPU::Hwreg::HwregEncoding::encode(AMDGPU::Hwreg::ID_MODE, 0, 23);
+
+static constexpr unsigned FPEnvTrapBitField =
+    AMDGPU::Hwreg::HwregEncoding::encode(AMDGPU::Hwreg::ID_TRAPSTS, 0, 5);
+
+bool AMDGPULegalizerInfo::legalizeGetFPEnv(MachineInstr &MI,
+                                           MachineIRBuilder &B) const {
+  auto ModeReg =
+      B.buildIntrinsic(Intrinsic::amdgcn_s_getreg, {S32},
+                       /*HasSideEffects=*/true, /*isConvergent=*/false)
+          .addImm(FPEnvModeBitField);
+  auto TrapReg =
+      B.buildIntrinsic(Intrinsic::amdgcn_s_getreg, {S32},
+                       /*HasSideEffects=*/true, /*isConvergent=*/false)
+          .addImm(FPEnvTrapBitField);
+  B.buildMergeLikeInstr(MI.getOperand(0).getReg(), {ModeReg, TrapReg});
+  MI.eraseFromParent();
+  return true;
+}
+
+bool AMDGPULegalizerInfo::legalizeSetFPEnv(MachineInstr &MI,
+                                           MachineIRBuilder &B) const {
+  auto Unmerge = B.buildUnmerge({S32, S32}, MI.getOperand(0));
+  B.buildIntrinsic(Intrinsic::amdgcn_s_setreg, ArrayRef<DstOp>(),
+                   /*HasSideEffects=*/true, /*isConvergent=*/false)
+      .addImm(static_cast<int16_t>(FPEnvModeBitField))
+      .addReg(Unmerge.getReg(0));
+  B.buildIntrinsic(Intrinsic::amdgcn_s_setreg, ArrayRef<DstOp>(),
+                   /*HasSideEffects=*/true, /*isConvergent=*/false)
+      .addImm(static_cast<int16_t>(FPEnvTrapBitField))
+      .addReg(Unmerge.getReg(1));
+  MI.eraseFromParent();
+  return true;
+}
+
 bool AMDGPULegalizerInfo::legalizeIntrinsic(LegalizerHelper &Helper,
                                             MachineInstr &MI) const {
   MachineIRBuilder &B = Helper.MIRBuilder;
diff --git a/llvm/lib/Target/AMDGPU/AMDGPULegalizerInfo.h b/llvm/lib/Target/AMDGPU/AMDGPULegalizerInfo.h
@@ -214,6 +214,9 @@ class AMDGPULegalizerInfo final : public LegalizerInfo {
   bool legalizeStackSave(MachineInstr &MI, MachineIRBuilder &B) const;
   bool legalizeWaveID(MachineInstr &MI, MachineIRBuilder &B) const;
 
+  bool legalizeGetFPEnv(MachineInstr &MI, MachineIRBuilder &B) const;
+  bool legalizeSetFPEnv(MachineInstr &MI, MachineIRBuilder &B) const;
+
   bool legalizeImageIntrinsic(
       MachineInstr &MI, MachineIRBuilder &B,
       GISelChangeObserver &Observer,
diff --git a/llvm/lib/Target/AMDGPU/SIISelLowering.cpp b/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
@@ -876,6 +876,8 @@ SITargetLowering::SITargetLowering(const TargetMachine &TM,
 
   setOperationAction(ISD::STACKSAVE, MVT::Other, Custom);
   setOperationAction(ISD::GET_ROUNDING, MVT::i32, Custom);
+  setOperationAction(ISD::GET_FPENV, MVT::i64, Custom);
+  setOperationAction(ISD::SET_FPENV, MVT::i64, Custom);
 
   // TODO: Could move this to custom lowering, could benefit from combines on
   // extract of relevant bits.
@@ -4079,6 +4081,71 @@ SDValue SITargetLowering::lowerFP_EXTEND(SDValue Op, SelectionDAG &DAG) const {
   return DAG.getNode(ISD::BF16_TO_FP, SL, DstVT, BitCast);
 }
 
+SDValue SITargetLowering::lowerGET_FPENV(SDValue Op, SelectionDAG &DAG) const {
+  SDLoc SL(Op);
+  assert(Op.getValueType() == MVT::i64);
+
+  uint32_t ModeHwReg =
+      AMDGPU::Hwreg::HwregEncoding::encode(AMDGPU::Hwreg::ID_MODE, 0, 23);
+  SDValue ModeHwRegImm = DAG.getTargetConstant(ModeHwReg, SL, MVT::i32);
+  uint32_t TrapHwReg =
+      AMDGPU::Hwreg::HwregEncoding::encode(AMDGPU::Hwreg::ID_TRAPSTS, 0, 5);
+  SDValue TrapHwRegImm = DAG.getTargetConstant(TrapHwReg, SL, MVT::i32);
+
+  SDVTList VTList = DAG.getVTList(MVT::i32, MVT::Other);
+  SDValue IntrinID =
+      DAG.getTargetConstant(Intrinsic::amdgcn_s_getreg, SL, MVT::i32);
+  SDValue GetModeReg = DAG.getNode(ISD::INTRINSIC_W_CHAIN, SL, VTList,
+                                   Op.getOperand(0), IntrinID, ModeHwRegImm);
+  SDValue GetTrapReg = DAG.getNode(ISD::INTRINSIC_W_CHAIN, SL, VTList,
+                                   Op.getOperand(0), IntrinID, TrapHwRegImm);
+  SDValue TokenReg =
+      DAG.getNode(ISD::TokenFactor, SL, MVT::Other, GetModeReg.getValue(1),
+                  GetTrapReg.getValue(1));
+
+  SDValue CvtPtr =
+      DAG.getNode(ISD::BUILD_VECTOR, SL, MVT::v2i32, GetModeReg, GetTrapReg);
+  SDValue Result = DAG.getNode(ISD::BITCAST, SL, MVT::i64, CvtPtr);
+
+  return DAG.getMergeValues({Result, TokenReg}, SL);
+}
+
+SDValue SITargetLowering::lowerSET_FPENV(SDValue Op, SelectionDAG &DAG) const {
+  SDLoc SL(Op);
+  assert(Op.getOperand(1).getValueType() == MVT::i64);
+
+  SDValue Input = DAG.getNode(ISD::BITCAST, SL, MVT::v2i32, Op.getOperand(1));
+  SDValue NewModeReg = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SL, MVT::i32, Input,
+                                   DAG.getConstant(0, SL, MVT::i32));
+  SDValue NewTrapReg = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SL, MVT::i32, Input,
+                                   DAG.getConstant(1, SL, MVT::i32));
+
+  SDValue ReadFirstLaneID =
+      DAG.getTargetConstant(Intrinsic::amdgcn_readfirstlane, SL, MVT::i32);
+  NewModeReg = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, SL, MVT::i32,
+                           ReadFirstLaneID, NewModeReg);
+  NewTrapReg = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, SL, MVT::i32,
+                           ReadFirstLaneID, NewTrapReg);
+
+  unsigned ModeHwReg =
+      AMDGPU::Hwreg::HwregEncoding::encode(AMDGPU::Hwreg::ID_MODE, 0, 23);
+  SDValue ModeHwRegImm = DAG.getTargetConstant(ModeHwReg, SL, MVT::i32);
+  unsigned TrapHwReg =
+      AMDGPU::Hwreg::HwregEncoding::encode(AMDGPU::Hwreg::ID_TRAPSTS, 0, 5);
+  SDValue TrapHwRegImm = DAG.getTargetConstant(TrapHwReg, SL, MVT::i32);
+
+  SDVTList VTList = DAG.getVTList(MVT::Other);
+  SDValue IntrinID =
+      DAG.getTargetConstant(Intrinsic::amdgcn_s_setreg, SL, MVT::i32);
+  SDValue SetModeReg =
+      DAG.getNode(ISD::INTRINSIC_VOID, SL, VTList, Op.getOperand(0), IntrinID,
+                  ModeHwRegImm, NewModeReg);
+  SDValue SetTrapReg =
+      DAG.getNode(ISD::INTRINSIC_VOID, SL, VTList, Op.getOperand(0), IntrinID,
+                  TrapHwRegImm, NewTrapReg);
+  return DAG.getNode(ISD::TokenFactor, SL, MVT::Other, SetTrapReg, SetModeReg);
+}
+
 Register SITargetLowering::getRegisterByName(const char* RegName, LLT VT,
                                              const MachineFunction &MF) const {
   Register Reg = StringSwitch<Register>(RegName)
@@ -5663,6 +5730,10 @@ SDValue SITargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
   case ISD::FP_EXTEND:
   case ISD::STRICT_FP_EXTEND:
     return lowerFP_EXTEND(Op, DAG);
+  case ISD::GET_FPENV:
+    return lowerGET_FPENV(Op, DAG);
+  case ISD::SET_FPENV:
+    return lowerSET_FPENV(Op, DAG);
   }
   return SDValue();
 }
diff --git a/llvm/lib/Target/AMDGPU/SIISelLowering.h b/llvm/lib/Target/AMDGPU/SIISelLowering.h
@@ -425,6 +425,8 @@ class SITargetLowering final : public AMDGPUTargetLowering {
 
   SDValue lowerPREFETCH(SDValue Op, SelectionDAG &DAG) const;
   SDValue lowerFP_EXTEND(SDValue Op, SelectionDAG &DAG) const;
+  SDValue lowerGET_FPENV(SDValue Op, SelectionDAG &DAG) const;
+  SDValue lowerSET_FPENV(SDValue Op, SelectionDAG &DAG) const;
 
   Register getRegisterByName(const char* RegName, LLT VT,
                              const MachineFunction &MF) const override;
diff --git a/llvm/test/CodeGen/AMDGPU/fpenv.ll b/llvm/test/CodeGen/AMDGPU/fpenv.ll
