/*========================== begin_copyright_notice ============================

Copyright (C) 2018-2021 Intel Corporation

SPDX-License-Identifier: MIT

============================= end_copyright_notice ===========================*/

#include "Compiler/Optimizer/OpenCLPasses/ErrorCheckPass/ErrorCheckPass.h"
#include "Compiler/Optimizer/OpenCLPasses/KernelArgs/KernelArgs.hpp"
#include "Compiler/IGCPassSupport.h"
#include "Compiler/CISACodeGen/OpenCLKernelCodeGen.hpp"

#include "common/LLVMWarningsPush.hpp"
#include <llvm/Pass.h>
#include <llvm/IR/InstVisitor.h>
#include <llvm/IR/Function.h>
#include <llvm/IR/Instructions.h>
#include "common/LLVMWarningsPop.hpp"

using namespace llvm;
using namespace IGC;

char ErrorCheck::ID = 0;

// Register pass to igc-opt
#define PASS_FLAG "igc-error-check"
#define PASS_DESCRIPTION "Check for input errors"
#define PASS_CFG_ONLY false
#define PASS_ANALYSIS false
IGC_INITIALIZE_PASS_BEGIN(ErrorCheck, PASS_FLAG, PASS_DESCRIPTION, PASS_CFG_ONLY, PASS_ANALYSIS)
IGC_INITIALIZE_PASS_DEPENDENCY(CodeGenContextWrapper)
IGC_INITIALIZE_PASS_DEPENDENCY(MetaDataUtilsWrapper)
IGC_INITIALIZE_PASS_END(ErrorCheck, PASS_FLAG, PASS_DESCRIPTION, PASS_CFG_ONLY, PASS_ANALYSIS)

ErrorCheck::ErrorCheck(void) : FunctionPass(ID) { initializeErrorCheckPass(*PassRegistry::getPassRegistry()); }

bool ErrorCheck::runOnFunction(Function &F) {
  // add more checks as needed later
  visit(F);

  if (isEntryFunc(getAnalysis<MetaDataUtilsWrapper>().getMetaDataUtils(), &F)) {
    checkArgsSize(F);
  }

  return m_hasError;
}

void ErrorCheck::checkArgsSize(Function &F) {
  auto Ctx = getAnalysis<CodeGenContextWrapper>().getCodeGenContext();
  auto MdUtils = getAnalysis<MetaDataUtilsWrapper>().getMetaDataUtils();
  auto ModMD = getAnalysis<MetaDataUtilsWrapper>().getModuleMetaData();

  auto DL = F.getParent()->getDataLayout();
  KernelArgs KernelArgs(F, &DL, MdUtils, ModMD, Ctx->platform.getGRFSize());
  auto MaxParamSize = Ctx->platform.getMaxOCLParameteSize();
  uint64_t TotalSize = 0;

  if (KernelArgs.empty()) {
    return;
  }

  for (auto &KernelArg : KernelArgs) {
    auto Arg = KernelArg.getArg();
    Type *ArgType = Arg->getType();
    ArgType = Arg->hasByValAttr() ? Arg->getParamByValType() : ArgType;
    if (!KernelArg.isImplicitArg()) {
      TotalSize += DL.getTypeAllocSize(ArgType);
    }
  }

  if (TotalSize > MaxParamSize) {
    std::string ErrorMsg =
        "Total size of kernel arguments exceeds limit! Total arguments size: " + std::to_string(TotalSize) +
        ", limit: " + std::to_string(MaxParamSize);

    Ctx->EmitError(ErrorMsg.c_str(), &F);
    m_hasError = true;
  }
}

static bool isFP64Operation(llvm::Instruction *I) {
  if (I->getType()->isDoubleTy())
    return true;

  for (int i = 0, numOpnd = (int)I->getNumOperands(); i < numOpnd; ++i)
    if (I->getOperand(i)->getType()->isDoubleTy())
      return true;

  return false;
}

static bool isFP64ArithmeticOperation(llvm::Instruction *I) {
  if (auto *II = dyn_cast<IntrinsicInst>(I)) {
    switch (II->getIntrinsicID()) {
    case Intrinsic::fabs:
    case Intrinsic::fma:
    case Intrinsic::sqrt:
      return true;
    default:
      return false;
    }
  } else if (auto *GII = dyn_cast<GenIntrinsicInst>(I)) {
    switch (GII->getIntrinsicID()) {
    case GenISAIntrinsic::GenISA_fma_rtn:
    case GenISAIntrinsic::GenISA_fma_rtp:
    case GenISAIntrinsic::GenISA_fma_rtz:
      return true;
    default:
      return false;
    }
  } else {
    switch (I->getOpcode()) {
    case Instruction::FAdd:
    case Instruction::FSub:
    case Instruction::FMul:
    case Instruction::FDiv:
      return true;
    default:
      return false;
    }
  }
}

static bool isValidFP64InstructionForDPConvEmu(llvm::Instruction *I) {
  switch (I->getOpcode()) {
  case Instruction::Load:
  case Instruction::Store:
  case Instruction::FPToSI:
  case Instruction::FPToUI:
  case Instruction::SIToFP:
  case Instruction::UIToFP:
  case Instruction::FPExt:
  case Instruction::FPTrunc:
  case Instruction::ExtractElement:
  case Instruction::InsertElement:
  case Instruction::ExtractValue:
  case Instruction::InsertValue:
  case Instruction::BitCast:
  case Instruction::PHI:
  case Instruction::Select:
  case Instruction::FCmp:
  // By default, we assume that Call instruction is valid for DPConvEmu,
  // because it can call e.g. other func or kernel.
  // Call instruction can be invalid for DPConvEmu when it calls intrinsic
  // which is arithmetic operation, but this case is handled in isFP64ArithmeticOperation()
  case Instruction::Call:
    return true;
  default:
    return false;
  }
}

// This function handles FP64 emulation mode. It investigates the following cases:
// if (ctx->m_hasDPConvEmu && !ctx->m_hasDPEmu
//    where AO == ArithmeticOperation, VI == ValidInstructionForDPConvEmu
//     1. poison=1 AO=1         => attr + warning
//     2. poison=1 AO=0 VI=0    => attr
//     3. poison=0 AO=1         => err msg
//     4. poison=0 AO=0         => emu
// if (!ctx->m_hasDPConvEmu && ctx->m_hasDPEmu)
//     1. poison=1              => emu
//     2. poison=0              => emu
// if (ctx->m_hasDPConvEmu && ctx->m_hasDPEmu)
//     1. poison=1              => emu
//     2. poison=0              => emu
// if (!ctx->m_hasDPConvEmu && !ctx->m_hasDPEmu)
//     1. poison=1              => attr
//     2. poison=0              => err msg
void ErrorCheck::handleFP64EmulationMode(llvm::Instruction &I) {
  auto ctx = getAnalysis<CodeGenContextWrapper>().getCodeGenContext();

  // This is the condition that double emulation is used.
  ctx->checkDPEmulationEnabled();

  bool poisonFP64KernelsEnabled = false;
  if (ctx->type == ShaderType::OPENCL_SHADER) {
    OpenCLProgramContext *OCLContext = static_cast<OpenCLProgramContext *>(ctx);
    poisonFP64KernelsEnabled = OCLContext->m_Options.EnableUnsupportedFP64Poisoning;
  }

  // check that has HW DP support and DP emu is disabled
  if (!ctx->platform.hasNoFP64Inst() && !ctx->m_hasDPEmu)
    return;

  // check that input does not use double
  const bool usesDouble = isFP64Operation(&I);
  if (!usesDouble)
    return;

  // emit error msg when platform don't support DP operations and poisonFP64Kernels is disabled
  if (!poisonFP64KernelsEnabled && !ctx->m_hasDPEmu && !ctx->m_hasDPConvEmu) {
    ctx->EmitError("Double type is not supported on this platform.", &I);
    m_hasError = true;
    return;
  }

  // emit error msg when platform can emulate DP conversion operations, but in the kernel are used DP arithmetic
  // operations (poisonFP64Kernels is disabled)
  if (!poisonFP64KernelsEnabled && !ctx->m_hasDPEmu && ctx->m_hasDPConvEmu && isFP64ArithmeticOperation(&I)) {
    ctx->EmitError("Double arithmetic operation is not supported on this platform with FP64 conversion emulation mode "
                   "(poison FP64 kernels is disabled).",
                   &I);
    m_hasError = true;
    return;
  }

  // emit error msg when platform can emulate DP conversion operations, but in the kernel are used instructions that are
  // not valid for DPConvEmu (poisonFP64Kernels is disabled)
  if (!poisonFP64KernelsEnabled && !ctx->m_hasDPEmu && ctx->m_hasDPConvEmu && !isValidFP64InstructionForDPConvEmu(&I)) {
    ctx->EmitError("Instruction is not valid on this platform with FP64 conversion emulation mode (poison FP64 kernels "
                   "is disabled).",
                   &I);
    m_hasError = true;
    return;
  }

  // emit warning when platform can emulate DP conversion operations, but in the kernel are used DP arithmetic
  // operations and poisonFP64Kernels is enabled we can add "uses-fp64-math" attr for functions here, because
  // poisonFP64Kernels is enabled and we may encounter Call instruction to intrinsic which is arithmetic operation
  if (poisonFP64KernelsEnabled && !ctx->m_hasDPEmu && ctx->m_hasDPConvEmu && isFP64ArithmeticOperation(&I)) {
    Function *F = I.getParent()->getParent();
    F->addFnAttr("uses-fp64-math");
    ctx->EmitWarning("Double arithmetic operation is not supported on this platform with FP64 conversion emulation "
                     "mode (poison FP64 kernels is enabled).");
  }

  // emit warning when used two options: for FP64 conv emu and for FP64 full emu,
  // the FP64 conversion emu flag will be ignored by the compiler, because is a subset of FP64 full emu.
  if (ctx->m_hasDPEmu && ctx->m_hasDPConvEmu) {
    ctx->EmitWarning(
        "Used two options for FP64 emulation: for FP64 conversion emu (-cl/-ze-fp64-gen-conv-emu) and FP64 full emu "
        "(-cl/-ze-fp64-gen-emu). The FP64 conversion emu flag will be ignored by the compiler.");
  }

  // Add "uses-fp64-math" attr for functions when poison fp64 kernels is required
  // on platform and DP emulation is disabled
  // When investigated instruction is valid for DPConvEmu, we don't need to add the attribute
  if ((poisonFP64KernelsEnabled && !ctx->m_hasDPEmu && !ctx->m_hasDPConvEmu) ||
      (poisonFP64KernelsEnabled && !ctx->m_hasDPEmu && ctx->m_hasDPConvEmu &&
       !isValidFP64InstructionForDPConvEmu(&I))) {
    Function *F = I.getParent()->getParent();
    F->addFnAttr("uses-fp64-math");
  }
}

void ErrorCheck::visitInstruction(llvm::Instruction &I) { handleFP64EmulationMode(I); }

void ErrorCheck::visitCallInst(CallInst &CI) {
  if (auto *GII = dyn_cast<GenIntrinsicInst>(&CI)) {
    switch (GII->getIntrinsicID()) {
    case GenISAIntrinsic::GenISA_dp4a_ss:
    case GenISAIntrinsic::GenISA_dp4a_su:
    case GenISAIntrinsic::GenISA_dp4a_us:
    case GenISAIntrinsic::GenISA_dp4a_uu: {
      CodeGenContext *Ctx = getAnalysis<CodeGenContextWrapper>().getCodeGenContext();
      if (!Ctx->platform.hasHWDp4AddSupport()) {
        std::string Msg = "Unsupported call to ";
        Msg += CI.getCalledFunction() ? CI.getCalledFunction()->getName() : "indirect function";
        Ctx->EmitError(Msg.c_str(), &CI);
        m_hasError = true;
      }
      break;
    }
    default:
      // Intrinsic supported.
      break;
    }
  }

  handleFP64EmulationMode(CI);
}