brintos

brintos / llvm-project-archived public Read only

0
0
Text · 9.3 KiB · cc46159 Raw
227 lines · cpp
1//=== AArch64CallingConvention.cpp - AArch64 CC impl ------------*- C++ -*-===//2//3// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.4// See https://llvm.org/LICENSE.txt for license information.5// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception6//7//===----------------------------------------------------------------------===//8//9// This file contains the table-generated and custom routines for the AArch6410// Calling Convention.11//12//===----------------------------------------------------------------------===//13 14#include "AArch64CallingConvention.h"15#include "AArch64.h"16#include "AArch64InstrInfo.h"17#include "AArch64Subtarget.h"18#include "llvm/CodeGen/CallingConvLower.h"19#include "llvm/CodeGen/TargetInstrInfo.h"20using namespace llvm;21 22static const MCPhysReg XRegList[] = {AArch64::X0, AArch64::X1, AArch64::X2,23                                     AArch64::X3, AArch64::X4, AArch64::X5,24                                     AArch64::X6, AArch64::X7};25static const MCPhysReg HRegList[] = {AArch64::H0, AArch64::H1, AArch64::H2,26                                     AArch64::H3, AArch64::H4, AArch64::H5,27                                     AArch64::H6, AArch64::H7};28static const MCPhysReg SRegList[] = {AArch64::S0, AArch64::S1, AArch64::S2,29                                     AArch64::S3, AArch64::S4, AArch64::S5,30                                     AArch64::S6, AArch64::S7};31static const MCPhysReg DRegList[] = {AArch64::D0, AArch64::D1, AArch64::D2,32                                     AArch64::D3, AArch64::D4, AArch64::D5,33                                     AArch64::D6, AArch64::D7};34static const MCPhysReg QRegList[] = {AArch64::Q0, AArch64::Q1, AArch64::Q2,35                                     AArch64::Q3, AArch64::Q4, AArch64::Q5,36                                     AArch64::Q6, AArch64::Q7};37static const MCPhysReg ZRegList[] = {AArch64::Z0, AArch64::Z1, AArch64::Z2,38                                     AArch64::Z3, AArch64::Z4, AArch64::Z5,39                                     AArch64::Z6, AArch64::Z7};40static const MCPhysReg PRegList[] = {AArch64::P0, AArch64::P1, AArch64::P2,41                                     AArch64::P3};42 43static bool finishStackBlock(SmallVectorImpl<CCValAssign> &PendingMembers,44                             MVT LocVT, ISD::ArgFlagsTy &ArgFlags,45                             CCState &State, Align SlotAlign) {46  if (LocVT.isScalableVector()) {47    const AArch64Subtarget &Subtarget = static_cast<const AArch64Subtarget &>(48        State.getMachineFunction().getSubtarget());49    const AArch64TargetLowering *TLI = Subtarget.getTargetLowering();50 51    // We are about to reinvoke the CCAssignFn auto-generated handler. If we52    // don't unset these flags we will get stuck in an infinite loop forever53    // invoking the custom handler.54    ArgFlags.setInConsecutiveRegs(false);55    ArgFlags.setInConsecutiveRegsLast(false);56 57    // The calling convention for passing SVE tuples states that in the event58    // we cannot allocate enough registers for the tuple we should still leave59    // any remaining registers unallocated. However, when we call the60    // CCAssignFn again we want it to behave as if all remaining registers are61    // allocated. This will force the code to pass the tuple indirectly in62    // accordance with the PCS.63    bool ZRegsAllocated[8];64    for (int I = 0; I < 8; I++) {65      ZRegsAllocated[I] = State.isAllocated(ZRegList[I]);66      State.AllocateReg(ZRegList[I]);67    }68    // The same applies to P registers.69    bool PRegsAllocated[4];70    for (int I = 0; I < 4; I++) {71      PRegsAllocated[I] = State.isAllocated(PRegList[I]);72      State.AllocateReg(PRegList[I]);73    }74 75    auto &It = PendingMembers[0];76    CCAssignFn *AssignFn =77        TLI->CCAssignFnForCall(State.getCallingConv(), /*IsVarArg=*/false);78    // FIXME: Get the correct original type.79    Type *OrigTy = EVT(It.getValVT()).getTypeForEVT(State.getContext());80    if (AssignFn(It.getValNo(), It.getValVT(), It.getValVT(), CCValAssign::Full,81                 ArgFlags, OrigTy, State))82      llvm_unreachable("Call operand has unhandled type");83 84    // Return the flags to how they were before.85    ArgFlags.setInConsecutiveRegs(true);86    ArgFlags.setInConsecutiveRegsLast(true);87 88    // Return the register state back to how it was before, leaving any89    // unallocated registers available for other smaller types.90    for (int I = 0; I < 8; I++)91      if (!ZRegsAllocated[I])92        State.DeallocateReg(ZRegList[I]);93    for (int I = 0; I < 4; I++)94      if (!PRegsAllocated[I])95        State.DeallocateReg(PRegList[I]);96 97    // All pending members have now been allocated98    PendingMembers.clear();99    return true;100  }101 102  unsigned Size = LocVT.getSizeInBits() / 8;103  for (auto &It : PendingMembers) {104    It.convertToMem(State.AllocateStack(Size, SlotAlign));105    State.addLoc(It);106    SlotAlign = Align(1);107  }108 109  // All pending members have now been allocated110  PendingMembers.clear();111  return true;112}113 114/// The Darwin variadic PCS places anonymous arguments in 8-byte stack slots. An115/// [N x Ty] type must still be contiguous in memory though.116static bool CC_AArch64_Custom_Stack_Block(117      unsigned &ValNo, MVT &ValVT, MVT &LocVT, CCValAssign::LocInfo &LocInfo,118      ISD::ArgFlagsTy &ArgFlags, CCState &State) {119  SmallVectorImpl<CCValAssign> &PendingMembers = State.getPendingLocs();120 121  // Add the argument to the list to be allocated once we know the size of the122  // block.123  PendingMembers.push_back(124      CCValAssign::getPending(ValNo, ValVT, LocVT, LocInfo));125 126  if (!ArgFlags.isInConsecutiveRegsLast())127    return true;128 129  return finishStackBlock(PendingMembers, LocVT, ArgFlags, State, Align(8));130}131 132/// Given an [N x Ty] block, it should be passed in a consecutive sequence of133/// registers. If no such sequence is available, mark the rest of the registers134/// of that type as used and place the argument on the stack.135static bool CC_AArch64_Custom_Block(unsigned &ValNo, MVT &ValVT, MVT &LocVT,136                                    CCValAssign::LocInfo &LocInfo,137                                    ISD::ArgFlagsTy &ArgFlags, CCState &State) {138  const AArch64Subtarget &Subtarget = static_cast<const AArch64Subtarget &>(139      State.getMachineFunction().getSubtarget());140  bool IsDarwinILP32 = Subtarget.isTargetILP32() && Subtarget.isTargetMachO();141 142  // Try to allocate a contiguous block of registers, each of the correct143  // size to hold one member.144  ArrayRef<MCPhysReg> RegList;145  if (LocVT.SimpleTy == MVT::i64 || (IsDarwinILP32 && LocVT.SimpleTy == MVT::i32))146    RegList = XRegList;147  else if (LocVT.SimpleTy == MVT::f16 || LocVT.SimpleTy == MVT::bf16)148    RegList = HRegList;149  else if (LocVT.SimpleTy == MVT::f32 || LocVT.is32BitVector())150    RegList = SRegList;151  else if (LocVT.SimpleTy == MVT::f64 || LocVT.is64BitVector())152    RegList = DRegList;153  else if (LocVT.SimpleTy == MVT::f128 || LocVT.is128BitVector())154    RegList = QRegList;155  else if (LocVT.isScalableVector()) {156    // Scalable masks should be pass by Predicate registers.157    if (LocVT == MVT::nxv1i1 || LocVT == MVT::nxv2i1 || LocVT == MVT::nxv4i1 ||158        LocVT == MVT::nxv8i1 || LocVT == MVT::nxv16i1 ||159        LocVT == MVT::aarch64svcount)160      RegList = PRegList;161    else162      RegList = ZRegList;163  } else {164    // Not an array we want to split up after all.165    return false;166  }167 168  SmallVectorImpl<CCValAssign> &PendingMembers = State.getPendingLocs();169 170  // Add the argument to the list to be allocated once we know the size of the171  // block.172  PendingMembers.push_back(173      CCValAssign::getPending(ValNo, ValVT, LocVT, LocInfo));174 175  if (!ArgFlags.isInConsecutiveRegsLast())176    return true;177 178  // [N x i32] arguments get packed into x-registers on Darwin's arm64_32179  // because that's how the armv7k Clang front-end emits small structs.180  unsigned EltsPerReg = (IsDarwinILP32 && LocVT.SimpleTy == MVT::i32) ? 2 : 1;181  ArrayRef<MCPhysReg> RegResult = State.AllocateRegBlock(182      RegList, alignTo(PendingMembers.size(), EltsPerReg) / EltsPerReg);183  if (!RegResult.empty() && EltsPerReg == 1) {184    for (const auto &[It, Reg] : zip(PendingMembers, RegResult)) {185      It.convertToReg(Reg);186      State.addLoc(It);187    }188    PendingMembers.clear();189    return true;190  } else if (!RegResult.empty()) {191    assert(EltsPerReg == 2 && "unexpected ABI");192    bool UseHigh = false;193    CCValAssign::LocInfo Info;194    unsigned RegIdx = 0;195    for (auto &It : PendingMembers) {196      Info = UseHigh ? CCValAssign::AExtUpper : CCValAssign::ZExt;197      State.addLoc(CCValAssign::getReg(It.getValNo(), MVT::i32,198                                       RegResult[RegIdx], MVT::i64, Info));199      UseHigh = !UseHigh;200      if (!UseHigh)201        ++RegIdx;202    }203    PendingMembers.clear();204    return true;205  }206 207  if (!LocVT.isScalableVector()) {208    // Mark all regs in the class as unavailable209    for (auto Reg : RegList)210      State.AllocateReg(Reg);211  }212 213  const MaybeAlign StackAlign =214      State.getMachineFunction().getDataLayout().getStackAlignment();215  assert(StackAlign && "data layout string is missing stack alignment");216  const Align MemAlign = ArgFlags.getNonZeroMemAlign();217  Align SlotAlign = std::min(MemAlign, *StackAlign);218  if (!Subtarget.isTargetDarwin())219    SlotAlign = std::max(SlotAlign, Align(8));220 221  return finishStackBlock(PendingMembers, LocVT, ArgFlags, State, SlotAlign);222}223 224// TableGen provides definitions of the calling convention analysis entry225// points.226#include "AArch64GenCallingConv.inc"227