diff --git a/llvm/lib/Target/RISCV/RISCVISelLowering.h b/llvm/lib/Target/RISCV/RISCVISelLowering.h --- a/llvm/lib/Target/RISCV/RISCVISelLowering.h +++ b/llvm/lib/Target/RISCV/RISCVISelLowering.h @@ -383,6 +383,9 @@ EmitInstrWithCustomInserter(MachineInstr &MI, MachineBasicBlock *BB) const override; + void AdjustInstrPostInstrSelection(MachineInstr &MI, + SDNode *Node) const override; + EVT getSetCCResultType(const DataLayout &DL, LLVMContext &Context, EVT VT) const override; diff --git a/llvm/lib/Target/RISCV/RISCVISelLowering.cpp b/llvm/lib/Target/RISCV/RISCVISelLowering.cpp --- a/llvm/lib/Target/RISCV/RISCVISelLowering.cpp +++ b/llvm/lib/Target/RISCV/RISCVISelLowering.cpp @@ -8015,6 +8015,22 @@ } } +void RISCVTargetLowering::AdjustInstrPostInstrSelection(MachineInstr &MI, + SDNode *Node) const { + // Add FRM dependency to any instructions with dynamic rounding mode. + unsigned Opc = MI.getOpcode(); + auto Idx = RISCV::getNamedOperandIdx(Opc, RISCV::OpName::funct3); + if (Idx < 0) + return; + if (MI.getOperand(Idx).getImm() != RISCVFPRndMode::DYN) + return; + // If the instruction already reads FRM, don't add another read. + if (MI.readsRegister(RISCV::FRM)) + return; + MI.addOperand( + MachineOperand::CreateReg(RISCV::FRM, /*isDef*/ false, /*isImp*/ true)); +} + // Calling Convention Implementation. // The expectations for frontend ABI lowering vary from target to target. // Ideally, an LLVM frontend would be able to avoid worrying about many ABI diff --git a/llvm/lib/Target/RISCV/RISCVInstrInfo.h b/llvm/lib/Target/RISCV/RISCVInstrInfo.h --- a/llvm/lib/Target/RISCV/RISCVInstrInfo.h +++ b/llvm/lib/Target/RISCV/RISCVInstrInfo.h @@ -20,6 +20,9 @@ #define GET_INSTRINFO_HEADER #include "RISCVGenInstrInfo.inc" +#define GET_INSTRINFO_OPERAND_ENUM +#include "RISCVGenInstrInfo.inc" + namespace llvm { class RISCVSubtarget; @@ -181,6 +184,10 @@ }; namespace RISCV { + +// Implemented in RISCVGenInstrInfo.inc +int16_t getNamedOperandIdx(uint16_t Opcode, uint16_t NamedIndex); + // Special immediate for AVL operand of V pseudo instructions to indicate VLMax. static constexpr int64_t VLMaxSentinel = -1LL; } // namespace RISCV diff --git a/llvm/lib/Target/RISCV/RISCVInstrInfo.cpp b/llvm/lib/Target/RISCV/RISCVInstrInfo.cpp --- a/llvm/lib/Target/RISCV/RISCVInstrInfo.cpp +++ b/llvm/lib/Target/RISCV/RISCVInstrInfo.cpp @@ -35,6 +35,7 @@ #include "RISCVGenCompressInstEmitter.inc" #define GET_INSTRINFO_CTOR_DTOR +#define GET_INSTRINFO_NAMED_OPS #include "RISCVGenInstrInfo.inc" static cl::opt<bool> PreferWholeRegisterMove( diff --git a/llvm/lib/Target/RISCV/RISCVInstrInfoF.td b/llvm/lib/Target/RISCV/RISCVInstrInfoF.td --- a/llvm/lib/Target/RISCV/RISCVInstrInfoF.td +++ b/llvm/lib/Target/RISCV/RISCVInstrInfoF.td @@ -59,7 +59,8 @@ // Instruction class templates //===----------------------------------------------------------------------===// -let hasSideEffects = 0, mayLoad = 0, mayStore = 0, mayRaiseFPException = 1 in +let hasSideEffects = 0, mayLoad = 0, mayStore = 0, mayRaiseFPException = 1, + UseNamedOperandTable = 1, hasPostISelHook = 1 in class FPFMA_rrr_frm<RISCVOpcode opcode, bits<2> funct2, string opcodestr, RegisterClass rty> : RVInstR4Frm<funct2, opcode, (outs rty:$rd), @@ -77,7 +78,8 @@ : RVInstR<funct7, funct3, OPC_OP_FP, (outs rty:$rd), (ins rty:$rs1, rty:$rs2), opcodestr, "$rd, $rs1, $rs2">; -let hasSideEffects = 0, mayLoad = 0, mayStore = 0, mayRaiseFPException = 1 in +let hasSideEffects = 0, mayLoad = 0, mayStore = 0, mayRaiseFPException = 1, + UseNamedOperandTable = 1, hasPostISelHook = 1 in class FPALU_rr_frm<bits<7> funct7, string opcodestr, RegisterClass rty> : RVInstRFrm<funct7, OPC_OP_FP, (outs rty:$rd), (ins rty:$rs1, rty:$rs2, frmarg:$funct3), opcodestr, @@ -96,7 +98,8 @@ let rs2 = rs2val; } -let hasSideEffects = 0, mayLoad = 0, mayStore = 0, mayRaiseFPException = 1 in +let hasSideEffects = 0, mayLoad = 0, mayStore = 0, mayRaiseFPException = 1, + UseNamedOperandTable = 1, hasPostISelHook = 1 in class FPUnaryOp_r_frm<bits<7> funct7, bits<5> rs2val, RegisterClass rdty, RegisterClass rs1ty, string opcodestr> : RVInstRFrm<funct7, OPC_OP_FP, (outs rdty:$rd), diff --git a/llvm/test/CodeGen/RISCV/frm-dependency.ll b/llvm/test/CodeGen/RISCV/frm-dependency.ll new file mode 100644 --- /dev/null +++ b/llvm/test/CodeGen/RISCV/frm-dependency.ll @@ -0,0 +1,96 @@ +; NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py +; RUN: llc -mtriple=riscv32 -mattr=+f -stop-after=finalize-isel < %s \ +; RUN: | FileCheck -check-prefixes=RV32IF %s +; RUN: llc -mtriple=riscv64 -mattr=+f -stop-after=finalize-isel < %s \ +; RUN: | FileCheck -check-prefixes=RV64IF %s + +; Make sure an implicit FRM dependency is added to instructions with dynamic +; rounding. + +define float @fadd_s(float %a, float %b) nounwind { + ; RV32IF-LABEL: name: fadd_s + ; RV32IF: bb.0 (%ir-block.0): + ; RV32IF-NEXT: liveins: $x10, $x11 + ; RV32IF-NEXT: {{ $}} + ; RV32IF-NEXT: [[COPY:%[0-9]+]]:gpr = COPY $x11 + ; RV32IF-NEXT: [[COPY1:%[0-9]+]]:gpr = COPY $x10 + ; RV32IF-NEXT: %2:fpr32 = nofpexcept FMV_W_X [[COPY]] + ; RV32IF-NEXT: %3:fpr32 = nofpexcept FMV_W_X [[COPY1]] + ; RV32IF-NEXT: %4:fpr32 = nofpexcept FADD_S killed %3, killed %2, 7, implicit $frm + ; RV32IF-NEXT: %5:gpr = nofpexcept FMV_X_W killed %4 + ; RV32IF-NEXT: $x10 = COPY %5 + ; RV32IF-NEXT: PseudoRET implicit $x10 + ; RV64IF-LABEL: name: fadd_s + ; RV64IF: bb.0 (%ir-block.0): + ; RV64IF-NEXT: liveins: $x10, $x11 + ; RV64IF-NEXT: {{ $}} + ; RV64IF-NEXT: [[COPY:%[0-9]+]]:gpr = COPY $x11 + ; RV64IF-NEXT: [[COPY1:%[0-9]+]]:gpr = COPY $x10 + ; RV64IF-NEXT: %2:fpr32 = nofpexcept FMV_W_X [[COPY]] + ; RV64IF-NEXT: %3:fpr32 = nofpexcept FMV_W_X [[COPY1]] + ; RV64IF-NEXT: %4:fpr32 = nofpexcept FADD_S killed %3, killed %2, 7, implicit $frm + ; RV64IF-NEXT: %5:gpr = nofpexcept FMV_X_W killed %4 + ; RV64IF-NEXT: $x10 = COPY %5 + ; RV64IF-NEXT: PseudoRET implicit $x10 + %1 = fadd float %a, %b + ret float %1 +} + +declare float @llvm.fma.f32(float, float, float) + +define float @fmadd_s(float %a, float %b, float %c) nounwind { + ; RV32IF-LABEL: name: fmadd_s + ; RV32IF: bb.0 (%ir-block.0): + ; RV32IF-NEXT: liveins: $x10, $x11, $x12 + ; RV32IF-NEXT: {{ $}} + ; RV32IF-NEXT: [[COPY:%[0-9]+]]:gpr = COPY $x12 + ; RV32IF-NEXT: [[COPY1:%[0-9]+]]:gpr = COPY $x11 + ; RV32IF-NEXT: [[COPY2:%[0-9]+]]:gpr = COPY $x10 + ; RV32IF-NEXT: %3:fpr32 = nofpexcept FMV_W_X [[COPY]] + ; RV32IF-NEXT: %4:fpr32 = nofpexcept FMV_W_X [[COPY1]] + ; RV32IF-NEXT: %5:fpr32 = nofpexcept FMV_W_X [[COPY2]] + ; RV32IF-NEXT: %6:fpr32 = nofpexcept FMADD_S killed %5, killed %4, killed %3, 7, implicit $frm + ; RV32IF-NEXT: %7:gpr = nofpexcept FMV_X_W killed %6 + ; RV32IF-NEXT: $x10 = COPY %7 + ; RV32IF-NEXT: PseudoRET implicit $x10 + ; RV64IF-LABEL: name: fmadd_s + ; RV64IF: bb.0 (%ir-block.0): + ; RV64IF-NEXT: liveins: $x10, $x11, $x12 + ; RV64IF-NEXT: {{ $}} + ; RV64IF-NEXT: [[COPY:%[0-9]+]]:gpr = COPY $x12 + ; RV64IF-NEXT: [[COPY1:%[0-9]+]]:gpr = COPY $x11 + ; RV64IF-NEXT: [[COPY2:%[0-9]+]]:gpr = COPY $x10 + ; RV64IF-NEXT: %3:fpr32 = nofpexcept FMV_W_X [[COPY]] + ; RV64IF-NEXT: %4:fpr32 = nofpexcept FMV_W_X [[COPY1]] + ; RV64IF-NEXT: %5:fpr32 = nofpexcept FMV_W_X [[COPY2]] + ; RV64IF-NEXT: %6:fpr32 = nofpexcept FMADD_S killed %5, killed %4, killed %3, 7, implicit $frm + ; RV64IF-NEXT: %7:gpr = nofpexcept FMV_X_W killed %6 + ; RV64IF-NEXT: $x10 = COPY %7 + ; RV64IF-NEXT: PseudoRET implicit $x10 + %1 = call float @llvm.fma.f32(float %a, float %b, float %c) + ret float %1 +} + +; This uses rtz instead of dyn rounding mode so shouldn't have an FRM dependncy. +define i32 @fcvt_w_s(float %a) nounwind { + ; RV32IF-LABEL: name: fcvt_w_s + ; RV32IF: bb.0 (%ir-block.0): + ; RV32IF-NEXT: liveins: $x10 + ; RV32IF-NEXT: {{ $}} + ; RV32IF-NEXT: [[COPY:%[0-9]+]]:gpr = COPY $x10 + ; RV32IF-NEXT: %1:fpr32 = nofpexcept FMV_W_X [[COPY]] + ; RV32IF-NEXT: %2:gpr = nofpexcept FCVT_W_S killed %1, 1 + ; RV32IF-NEXT: $x10 = COPY %2 + ; RV32IF-NEXT: PseudoRET implicit $x10 + ; RV64IF-LABEL: name: fcvt_w_s + ; RV64IF: bb.0 (%ir-block.0): + ; RV64IF-NEXT: liveins: $x10 + ; RV64IF-NEXT: {{ $}} + ; RV64IF-NEXT: [[COPY:%[0-9]+]]:gpr = COPY $x10 + ; RV64IF-NEXT: %1:fpr32 = nofpexcept FMV_W_X [[COPY]] + ; RV64IF-NEXT: %2:gpr = nofpexcept FCVT_W_S killed %1, 1 + ; RV64IF-NEXT: $x10 = COPY %2 + ; RV64IF-NEXT: PseudoRET implicit $x10 + %1 = fptosi float %a to i32 + ret i32 %1 +}