LowerArcToLLVM.cpp

Go to the documentation of this file.

//===- LowerArcToLLVM.cpp -------------------------------------------------===//
//
// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
// See https://llvm.org/LICENSE.txt for license information.
// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
//
//===----------------------------------------------------------------------===//
 
#include "circt/Conversion/ArcToLLVM.h"
#include "circt/Conversion/CombToArith.h"
#include "circt/Conversion/CombToLLVM.h"
#include "circt/Conversion/HWToLLVM.h"
#include "circt/Dialect/Arc/ArcConstants.h"
#include "circt/Dialect/Arc/ArcOps.h"
#include "circt/Dialect/Arc/ModelInfo.h"
#include "circt/Dialect/Arc/Runtime/Common.h"
#include "circt/Dialect/Arc/Runtime/FmtDescriptor.h"
#include "circt/Dialect/Arc/Runtime/JITBind.h"
#include "circt/Dialect/Arc/Runtime/TraceTaps.h"
#include "circt/Dialect/Comb/CombOps.h"
#include "circt/Dialect/LLHD/LLHDOps.h"
#include "circt/Dialect/LLHD/LLHDTypes.h"
#include "circt/Dialect/Seq/SeqOps.h"
#include "circt/Dialect/Sim/SimOps.h"
#include "circt/Support/ConversionPatternSet.h"
#include "circt/Support/Namespace.h"
#include "mlir/Conversion/ArithToLLVM/ArithToLLVM.h"
#include "mlir/Conversion/ControlFlowToLLVM/ControlFlowToLLVM.h"
#include "mlir/Conversion/FuncToLLVM/ConvertFuncToLLVM.h"
#include "mlir/Conversion/IndexToLLVM/IndexToLLVM.h"
#include "mlir/Conversion/LLVMCommon/ConversionTarget.h"
#include "mlir/Conversion/LLVMCommon/TypeConverter.h"
#include "mlir/Conversion/SCFToControlFlow/SCFToControlFlow.h"
#include "mlir/Dialect/ControlFlow/IR/ControlFlowOps.h"
#include "mlir/Dialect/Func/IR/FuncOps.h"
#include "mlir/Dialect/Index/IR/IndexOps.h"
#include "mlir/Dialect/LLVMIR/FunctionCallUtils.h"
#include "mlir/Dialect/LLVMIR/LLVMAttrs.h"
#include "mlir/Dialect/LLVMIR/LLVMDialect.h"
#include "mlir/Dialect/SCF/IR/SCF.h"
#include "mlir/IR/Builders.h"
#include "mlir/IR/BuiltinDialect.h"
#include "mlir/Pass/Pass.h"
#include "mlir/Transforms/DialectConversion.h"
#include "llvm/Support/Debug.h"
#include "llvm/Support/FormatVariadic.h"
 
#include <cstddef>
 
#define DEBUG_TYPE "lower-arc-to-llvm"
 
namespace circt {
#define GEN_PASS_DEF_LOWERARCTOLLVM
#include "circt/Conversion/Passes.h.inc"
} // namespace circt
 
using namespace mlir;
using namespace circt;
using namespace arc;
using namespace hw;
using namespace runtime;
 
//===----------------------------------------------------------------------===//
// Lowering Patterns
//===----------------------------------------------------------------------===//
 
static llvm::Twine evalSymbolFromModelName(StringRef modelName) {
  return modelName + "_eval";
}
 
namespace {
 
struct ModelOpLowering : public OpConversionPattern<arc::ModelOp> {
  using OpConversionPattern::OpConversionPattern;
  LogicalResult
  matchAndRewrite(arc::ModelOp op, OpAdaptor adaptor,
                  ConversionPatternRewriter &rewriter) const final {
    {
      IRRewriter::InsertionGuard guard(rewriter);
      rewriter.setInsertionPointToEnd(&op.getBodyBlock());
      func::ReturnOp::create(rewriter, op.getLoc());
    }
    auto funcName =
        rewriter.getStringAttr(evalSymbolFromModelName(op.getName()));
    auto funcType =
        rewriter.getFunctionType(op.getBody().getArgumentTypes(), {});
    auto func =
        mlir::func::FuncOp::create(rewriter, op.getLoc(), funcName, funcType);
    rewriter.inlineRegionBefore(op.getRegion(), func.getBody(), func.end());
    rewriter.eraseOp(op);
    return success();
  }
};
 
struct AllocStorageOpLowering
    : public OpConversionPattern<arc::AllocStorageOp> {
  using OpConversionPattern::OpConversionPattern;
  LogicalResult
  matchAndRewrite(arc::AllocStorageOp op, OpAdaptor adaptor,
                  ConversionPatternRewriter &rewriter) const final {
    auto type = typeConverter->convertType(op.getType());
    if (!op.getOffset().has_value())
      return failure();
    rewriter.replaceOpWithNewOp<LLVM::GEPOp>(op, type, rewriter.getI8Type(),
                                             adaptor.getInput(),
                                             LLVM::GEPArg(*op.getOffset()));
    return success();
  }
};
 
template <class ConcreteOp>
struct AllocStateLikeOpLowering : public OpConversionPattern<ConcreteOp> {
  using OpConversionPattern<ConcreteOp>::OpConversionPattern;
  using OpConversionPattern<ConcreteOp>::typeConverter;
  using OpAdaptor = typename ConcreteOp::Adaptor;
 
  LogicalResult
  matchAndRewrite(ConcreteOp op, OpAdaptor adaptor,
                  ConversionPatternRewriter &rewriter) const final {
    // Get a pointer to the correct offset in the storage.
    auto offsetAttr = op->template getAttrOfType<IntegerAttr>("offset");
    if (!offsetAttr)
      return failure();
    Value ptr = LLVM::GEPOp::create(
        rewriter, op->getLoc(), adaptor.getStorage().getType(),
        rewriter.getI8Type(), adaptor.getStorage(),
        LLVM::GEPArg(offsetAttr.getValue().getZExtValue()));
    rewriter.replaceOp(op, ptr);
    return success();
  }
};
 
struct StateReadOpLowering : public OpConversionPattern<arc::StateReadOp> {
  using OpConversionPattern::OpConversionPattern;
  LogicalResult
  matchAndRewrite(arc::StateReadOp op, OpAdaptor adaptor,
                  ConversionPatternRewriter &rewriter) const final {
    // Loading an ArrayRef is a no-op as ArrayRefs are accessed by reference.
    if (isa<ArrayRefType>(op.getType())) {
      rewriter.replaceOp(op, adaptor.getState());
      return success();
    }
 
    auto type = typeConverter->convertType(op.getType());
    rewriter.replaceOpWithNewOp<LLVM::LoadOp>(op, type, adaptor.getState());
    return success();
  }
};
 
struct StateWriteOpLowering : public OpConversionPattern<arc::StateWriteOp> {
  using OpConversionPattern::OpConversionPattern;
  LogicalResult
  matchAndRewrite(arc::StateWriteOp op, OpAdaptor adaptor,
                  ConversionPatternRewriter &rewriter) const final {
    if (!isa<ArrayRefType>(op.getValue().getType())) {
      rewriter.replaceOpWithNewOp<LLVM::StoreOp>(op, adaptor.getValue(),
                                                 adaptor.getState());
      return success();
    }
 
    int numBytes = op.getState().getType().getByteWidth();
    Value size = LLVM::ConstantOp::create(rewriter, op.getLoc(),
                                          rewriter.getI64Type(), numBytes);
    rewriter.replaceOpWithNewOp<LLVM::MemcpyOp>(
        op, adaptor.getState(), adaptor.getValue(), size, /*volatile=*/false);
    return success();
  }
};
 
//===----------------------------------------------------------------------===//
// Time Operations Lowering
//===----------------------------------------------------------------------===//
 
struct CurrentTimeOpLowering : public OpConversionPattern<arc::CurrentTimeOp> {
  using OpConversionPattern::OpConversionPattern;
  LogicalResult
  matchAndRewrite(arc::CurrentTimeOp op, OpAdaptor adaptor,
                  ConversionPatternRewriter &rewriter) const final {
    // Time is stored at offset 0 in storage (no offset needed).
    Value ptr = adaptor.getStorage();
    rewriter.replaceOpWithNewOp<LLVM::LoadOp>(op, rewriter.getI64Type(), ptr);
    return success();
  }
};
 
// Lower `llhd.constant_time` to an `i64` LLVM constant holding the time in
// femtoseconds. Time attributes with non-zero delta or epsilon, units smaller
// than `fs`, or values that overflow `i64` femtoseconds are rejected.
struct ConstantTimeOpLowering
    : public OpConversionPattern<llhd::ConstantTimeOp> {
  using OpConversionPattern::OpConversionPattern;
  LogicalResult
  matchAndRewrite(llhd::ConstantTimeOp op, OpAdaptor adaptor,
                  ConversionPatternRewriter &rewriter) const final {
    auto attr = op.getValue();
    if (attr.getDelta() != 0 || attr.getEpsilon() != 0)
      return rewriter.notifyMatchFailure(
          op, "non-zero delta or epsilon time components are not supported");
    uint64_t value = attr.getTime();
    StringRef unit = attr.getTimeUnit();
    uint64_t scale;
    if (unit == "fs")
      scale = 1;
    else if (unit == "ps")
      scale = 1'000ULL;
    else if (unit == "ns")
      scale = 1'000'000ULL;
    else if (unit == "us")
      scale = 1'000'000'000ULL;
    else if (unit == "ms")
      scale = 1'000'000'000'000ULL;
    else if (unit == "s")
      scale = 1'000'000'000'000'000ULL;
    else
      return rewriter.notifyMatchFailure(
          op, "time units smaller than `fs` are not supported");
    if (value > std::numeric_limits<uint64_t>::max() / scale)
      return rewriter.notifyMatchFailure(
          op, "time value does not fit into `i64` femtoseconds");
    rewriter.replaceOpWithNewOp<LLVM::ConstantOp>(op, rewriter.getI64Type(),
                                                  value * scale);
    return success();
  }
};
 
// `llhd.int_to_time` is a no-op
struct IntToTimeOpLowering : public OpConversionPattern<llhd::IntToTimeOp> {
  using OpConversionPattern::OpConversionPattern;
  LogicalResult
  matchAndRewrite(llhd::IntToTimeOp op, OpAdaptor adaptor,
                  ConversionPatternRewriter &rewriter) const final {
    rewriter.replaceOp(op, adaptor.getInput());
    return success();
  }
};
 
// `llhd.time_to_int` is a no-op
struct TimeToIntOpLowering : public OpConversionPattern<llhd::TimeToIntOp> {
  using OpConversionPattern::OpConversionPattern;
  LogicalResult
  matchAndRewrite(llhd::TimeToIntOp op, OpAdaptor adaptor,
                  ConversionPatternRewriter &rewriter) const final {
    rewriter.replaceOp(op, adaptor.getInput());
    return success();
  }
};
 
//===----------------------------------------------------------------------===//
// Memory and Storage Lowering
//===----------------------------------------------------------------------===//
 
struct AllocMemoryOpLowering : public OpConversionPattern<arc::AllocMemoryOp> {
  using OpConversionPattern::OpConversionPattern;
  LogicalResult
  matchAndRewrite(arc::AllocMemoryOp op, OpAdaptor adaptor,
                  ConversionPatternRewriter &rewriter) const final {
    auto offsetAttr = op->getAttrOfType<IntegerAttr>("offset");
    if (!offsetAttr)
      return failure();
    Value ptr = LLVM::GEPOp::create(
        rewriter, op.getLoc(), adaptor.getStorage().getType(),
        rewriter.getI8Type(), adaptor.getStorage(),
        LLVM::GEPArg(offsetAttr.getValue().getZExtValue()));
 
    rewriter.replaceOp(op, ptr);
    return success();
  }
};
 
struct StorageGetOpLowering : public OpConversionPattern<arc::StorageGetOp> {
  using OpConversionPattern::OpConversionPattern;
  LogicalResult
  matchAndRewrite(arc::StorageGetOp op, OpAdaptor adaptor,
                  ConversionPatternRewriter &rewriter) const final {
    Value offset = LLVM::ConstantOp::create(
        rewriter, op.getLoc(), rewriter.getI32Type(), op.getOffsetAttr());
    Value ptr = LLVM::GEPOp::create(
        rewriter, op.getLoc(), adaptor.getStorage().getType(),
        rewriter.getI8Type(), adaptor.getStorage(), offset);
    rewriter.replaceOp(op, ptr);
    return success();
  }
};
 
struct MemoryAccess {
  Value ptr;
  Value withinBounds;
};
 
static MemoryAccess prepareMemoryAccess(Location loc, Value memory,
                                        Value address, MemoryType type,
                                        ConversionPatternRewriter &rewriter) {
  auto zextAddrType = rewriter.getIntegerType(
      cast<IntegerType>(address.getType()).getWidth() + 1);
  Value addr = LLVM::ZExtOp::create(rewriter, loc, zextAddrType, address);
  Value addrLimit =
      LLVM::ConstantOp::create(rewriter, loc, zextAddrType,
                               rewriter.getI32IntegerAttr(type.getNumWords()));
  Value withinBounds = LLVM::ICmpOp::create(
      rewriter, loc, LLVM::ICmpPredicate::ult, addr, addrLimit);
  Value ptr = LLVM::GEPOp::create(
      rewriter, loc, LLVM::LLVMPointerType::get(memory.getContext()),
      rewriter.getIntegerType(type.getStride() * 8), memory, ValueRange{addr});
  return {ptr, withinBounds};
}
 
struct MemoryReadOpLowering : public OpConversionPattern<arc::MemoryReadOp> {
  using OpConversionPattern::OpConversionPattern;
  LogicalResult
  matchAndRewrite(arc::MemoryReadOp op, OpAdaptor adaptor,
                  ConversionPatternRewriter &rewriter) const final {
    auto type = typeConverter->convertType(op.getType());
    auto memoryType = cast<MemoryType>(op.getMemory().getType());
    auto access =
        prepareMemoryAccess(op.getLoc(), adaptor.getMemory(),
                            adaptor.getAddress(), memoryType, rewriter);
 
    // Only attempt to read the memory if the address is within bounds,
    // otherwise produce a zero value.
    rewriter.replaceOpWithNewOp<scf::IfOp>(
        op, access.withinBounds,
        [&](auto &builder, auto loc) {
          Value loadOp = LLVM::LoadOp::create(
              builder, loc, memoryType.getWordType(), access.ptr);
          scf::YieldOp::create(builder, loc, loadOp);
        },
        [&](auto &builder, auto loc) {
          Value zeroValue = LLVM::ConstantOp::create(
              builder, loc, type, builder.getI64IntegerAttr(0));
          scf::YieldOp::create(builder, loc, zeroValue);
        });
    return success();
  }
};
 
struct MemoryWriteOpLowering : public OpConversionPattern<arc::MemoryWriteOp> {
  using OpConversionPattern::OpConversionPattern;
  LogicalResult
  matchAndRewrite(arc::MemoryWriteOp op, OpAdaptor adaptor,
                  ConversionPatternRewriter &rewriter) const final {
    auto access = prepareMemoryAccess(
        op.getLoc(), adaptor.getMemory(), adaptor.getAddress(),
        cast<MemoryType>(op.getMemory().getType()), rewriter);
    auto enable = access.withinBounds;
 
    // Only attempt to write the memory if the address is within bounds.
    rewriter.replaceOpWithNewOp<scf::IfOp>(
        op, enable, [&](auto &builder, auto loc) {
          LLVM::StoreOp::create(builder, loc, adaptor.getData(), access.ptr);
          scf::YieldOp::create(builder, loc);
        });
    return success();
  }
};
 
/// A dummy lowering for clock gates to an AND gate.
struct ClockGateOpLowering : public OpConversionPattern<seq::ClockGateOp> {
  using OpConversionPattern::OpConversionPattern;
  LogicalResult
  matchAndRewrite(seq::ClockGateOp op, OpAdaptor adaptor,
                  ConversionPatternRewriter &rewriter) const final {
    rewriter.replaceOpWithNewOp<LLVM::AndOp>(op, adaptor.getInput(),
                                             adaptor.getEnable());
    return success();
  }
};
 
/// Lower 'seq.clock_inv x' to 'llvm.xor x true'
struct ClockInvOpLowering : public OpConversionPattern<seq::ClockInverterOp> {
  using OpConversionPattern::OpConversionPattern;
  LogicalResult
  matchAndRewrite(seq::ClockInverterOp op, OpAdaptor adaptor,
                  ConversionPatternRewriter &rewriter) const final {
    auto constTrue = LLVM::ConstantOp::create(rewriter, op->getLoc(),
                                              rewriter.getI1Type(), 1);
    rewriter.replaceOpWithNewOp<LLVM::XOrOp>(op, adaptor.getInput(), constTrue);
    return success();
  }
};
 
struct ZeroCountOpLowering : public OpConversionPattern<arc::ZeroCountOp> {
  using OpConversionPattern::OpConversionPattern;
  LogicalResult
  matchAndRewrite(arc::ZeroCountOp op, OpAdaptor adaptor,
                  ConversionPatternRewriter &rewriter) const override {
    // Use poison when input is zero.
    IntegerAttr isZeroPoison = rewriter.getBoolAttr(true);
 
    if (op.getPredicate() == arc::ZeroCountPredicate::leading) {
      rewriter.replaceOpWithNewOp<LLVM::CountLeadingZerosOp>(
          op, adaptor.getInput().getType(), adaptor.getInput(), isZeroPoison);
      return success();
    }
 
    rewriter.replaceOpWithNewOp<LLVM::CountTrailingZerosOp>(
        op, adaptor.getInput().getType(), adaptor.getInput(), isZeroPoison);
    return success();
  }
};
 
struct SeqConstClockLowering : public OpConversionPattern<seq::ConstClockOp> {
  using OpConversionPattern::OpConversionPattern;
  LogicalResult
  matchAndRewrite(seq::ConstClockOp op, OpAdaptor adaptor,
                  ConversionPatternRewriter &rewriter) const override {
    rewriter.replaceOpWithNewOp<LLVM::ConstantOp>(
        op, rewriter.getI1Type(), static_cast<int64_t>(op.getValue()));
    return success();
  }
};
 
template <typename OpTy>
struct ReplaceOpWithInputPattern : public OpConversionPattern<OpTy> {
  using OpConversionPattern<OpTy>::OpConversionPattern;
  using OpAdaptor = typename OpTy::Adaptor;
  LogicalResult
  matchAndRewrite(OpTy op, OpAdaptor adaptor,
                  ConversionPatternRewriter &rewriter) const override {
    rewriter.replaceOp(op, adaptor.getInput());
    return success();
  }
};
 
} // namespace
 
//===----------------------------------------------------------------------===//
// Simulation Orchestration Lowering Patterns
//===----------------------------------------------------------------------===//
 
namespace {
 
struct ModelInfoMap {
  size_t numStateBytes;
  llvm::DenseMap<StringRef, StateInfo> states;
  mlir::FlatSymbolRefAttr initialFnSymbol;
  mlir::FlatSymbolRefAttr finalFnSymbol;
};
 
template <typename OpTy>
struct ModelAwarePattern : public OpConversionPattern<OpTy> {
  ModelAwarePattern(const TypeConverter &typeConverter, MLIRContext *context,
                    llvm::DenseMap<StringRef, ModelInfoMap> &modelInfo)
      : OpConversionPattern<OpTy>(typeConverter, context),
        modelInfo(modelInfo) {}
 
protected:
  Value createPtrToPortState(ConversionPatternRewriter &rewriter, Location loc,
                             Value state, const StateInfo &port) const {
    MLIRContext *ctx = rewriter.getContext();
    return LLVM::GEPOp::create(rewriter, loc, LLVM::LLVMPointerType::get(ctx),
                               IntegerType::get(ctx, 8), state,
                               LLVM::GEPArg(port.offset));
  }
 
  llvm::DenseMap<StringRef, ModelInfoMap> &modelInfo;
};
 
/// Lowers SimInstantiateOp to a malloc and memset call. This pattern will
/// mutate the global module.
struct SimInstantiateOpLowering
    : public ModelAwarePattern<arc::SimInstantiateOp> {
  using ModelAwarePattern::ModelAwarePattern;
 
  LogicalResult
  matchAndRewrite(arc::SimInstantiateOp op, OpAdaptor adaptor,
                  ConversionPatternRewriter &rewriter) const final {
    auto modelIt = modelInfo.find(
        cast<SimModelInstanceType>(op.getBody().getArgument(0).getType())
            .getModel()
            .getValue());
    ModelInfoMap &model = modelIt->second;
 
    bool useRuntime = op.getRuntimeModel().has_value();
 
    ModuleOp moduleOp = op->getParentOfType<ModuleOp>();
    if (!moduleOp)
      return failure();
 
    ConversionPatternRewriter::InsertionGuard guard(rewriter);
 
    // FIXME: like the rest of MLIR, this assumes sizeof(intptr_t) ==
    // sizeof(size_t) on the target architecture.
    Type convertedIndex = typeConverter->convertType(rewriter.getIndexType());
    Location loc = op.getLoc();
    Value allocated;
 
    if (useRuntime) {
      // The instance is using the runtime library
      auto ptrTy = LLVM::LLVMPointerType::get(getContext());
 
      Value runtimeArgs;
      // If present, materialize the runtime argument string on the stack
      if (op.getRuntimeArgs().has_value()) {
        SmallVector<int8_t> argStringVec(op.getRuntimeArgsAttr().begin(),
                                         op.getRuntimeArgsAttr().end());
        argStringVec.push_back('\0');
        auto strAttr = mlir::DenseElementsAttr::get(
            mlir::RankedTensorType::get({(int64_t)argStringVec.size()},
                                        rewriter.getI8Type()),
            llvm::ArrayRef(argStringVec));
 
        auto arrayCst = LLVM::ConstantOp::create(
            rewriter, loc,
            LLVM::LLVMArrayType::get(rewriter.getI8Type(), argStringVec.size()),
            strAttr);
        auto cst1 = LLVM::ConstantOp::create(rewriter, loc,
                                             rewriter.getI32IntegerAttr(1));
        runtimeArgs = LLVM::AllocaOp::create(rewriter, loc, ptrTy,
                                             arrayCst.getType(), cst1);
        LLVM::LifetimeStartOp::create(rewriter, loc, runtimeArgs);
        LLVM::StoreOp::create(rewriter, loc, arrayCst, runtimeArgs);
      } else {
        runtimeArgs = LLVM::ZeroOp::create(rewriter, loc, ptrTy).getResult();
      }
      // Call the state allocation function
      auto rtModelPtr = LLVM::AddressOfOp::create(rewriter, loc, ptrTy,
                                                  op.getRuntimeModelAttr())
                            .getResult();
      allocated =
          LLVM::CallOp::create(rewriter, loc, {ptrTy},
                               runtime::APICallbacks::symNameAllocInstance,
                               {rtModelPtr, runtimeArgs})
              .getResult();
 
      if (op.getRuntimeArgs().has_value())
        LLVM::LifetimeEndOp::create(rewriter, loc, runtimeArgs);
 
    } else {
      // The instance is not using the runtime library
      FailureOr<LLVM::LLVMFuncOp> mallocFunc =
          LLVM::lookupOrCreateMallocFn(rewriter, moduleOp, convertedIndex);
      if (failed(mallocFunc))
        return mallocFunc;
 
      Value numStateBytes = LLVM::ConstantOp::create(
          rewriter, loc, convertedIndex, model.numStateBytes);
      allocated = LLVM::CallOp::create(rewriter, loc, mallocFunc.value(),
                                       ValueRange{numStateBytes})
                      .getResult();
      Value zero =
          LLVM::ConstantOp::create(rewriter, loc, rewriter.getI8Type(), 0);
      LLVM::MemsetOp::create(rewriter, loc, allocated, zero, numStateBytes,
                             false);
    }
 
    // Call the model's 'initial' function if present.
    if (model.initialFnSymbol) {
      auto initialFnType = LLVM::LLVMFunctionType::get(
          LLVM::LLVMVoidType::get(op.getContext()),
          {LLVM::LLVMPointerType::get(op.getContext())});
      LLVM::CallOp::create(rewriter, loc, initialFnType, model.initialFnSymbol,
                           ValueRange{allocated});
    }
 
    // Call the runtime's 'onInitialized' function if present.
    if (useRuntime)
      LLVM::CallOp::create(rewriter, loc, TypeRange{},
                           runtime::APICallbacks::symNameOnInitialized,
                           {allocated});
 
    // Execute the body.
    rewriter.inlineBlockBefore(&adaptor.getBody().getBlocks().front(), op,
                               {allocated});
 
    // Call the model's 'final' function if present.
    if (model.finalFnSymbol) {
      auto finalFnType = LLVM::LLVMFunctionType::get(
          LLVM::LLVMVoidType::get(op.getContext()),
          {LLVM::LLVMPointerType::get(op.getContext())});
      LLVM::CallOp::create(rewriter, loc, finalFnType, model.finalFnSymbol,
                           ValueRange{allocated});
    }
 
    if (useRuntime) {
      LLVM::CallOp::create(rewriter, loc, TypeRange{},
                           runtime::APICallbacks::symNameDeleteInstance,
                           {allocated});
    } else {
      FailureOr<LLVM::LLVMFuncOp> freeFunc =
          LLVM::lookupOrCreateFreeFn(rewriter, moduleOp);
      if (failed(freeFunc))
        return freeFunc;
 
      LLVM::CallOp::create(rewriter, loc, freeFunc.value(),
                           ValueRange{allocated});
    }
 
    rewriter.eraseOp(op);
    return success();
  }
};
 
struct SimSetInputOpLowering : public ModelAwarePattern<arc::SimSetInputOp> {
  using ModelAwarePattern::ModelAwarePattern;
 
  LogicalResult
  matchAndRewrite(arc::SimSetInputOp op, OpAdaptor adaptor,
                  ConversionPatternRewriter &rewriter) const final {
    auto modelIt =
        modelInfo.find(cast<SimModelInstanceType>(op.getInstance().getType())
                           .getModel()
                           .getValue());
    ModelInfoMap &model = modelIt->second;
 
    auto portIt = model.states.find(op.getInput());
    if (portIt == model.states.end()) {
      // If the port is not found in the state, it means the model does not
      // actually use it. Thus this operation is a no-op.
      rewriter.eraseOp(op);
      return success();
    }
 
    StateInfo &port = portIt->second;
    Value statePtr = createPtrToPortState(rewriter, op.getLoc(),
                                          adaptor.getInstance(), port);
    rewriter.replaceOpWithNewOp<LLVM::StoreOp>(op, adaptor.getValue(),
                                               statePtr);
 
    return success();
  }
};
 
struct SimGetPortOpLowering : public ModelAwarePattern<arc::SimGetPortOp> {
  using ModelAwarePattern::ModelAwarePattern;
 
  LogicalResult
  matchAndRewrite(arc::SimGetPortOp op, OpAdaptor adaptor,
                  ConversionPatternRewriter &rewriter) const final {
    auto modelIt =
        modelInfo.find(cast<SimModelInstanceType>(op.getInstance().getType())
                           .getModel()
                           .getValue());
    ModelInfoMap &model = modelIt->second;
 
    auto type = typeConverter->convertType(op.getValue().getType());
    if (!type)
      return failure();
    auto portIt = model.states.find(op.getPort());
    if (portIt == model.states.end()) {
      // If the port is not found in the state, it means the model does not
      // actually set it. Thus this operation returns 0.
      rewriter.replaceOpWithNewOp<LLVM::ConstantOp>(op, type, 0);
      return success();
    }
 
    StateInfo &port = portIt->second;
    Value statePtr = createPtrToPortState(rewriter, op.getLoc(),
                                          adaptor.getInstance(), port);
    rewriter.replaceOpWithNewOp<LLVM::LoadOp>(op, type, statePtr);
 
    return success();
  }
};
 
struct SimStepOpLowering : public ModelAwarePattern<arc::SimStepOp> {
  using ModelAwarePattern::ModelAwarePattern;
 
  LogicalResult
  matchAndRewrite(arc::SimStepOp op, OpAdaptor adaptor,
                  ConversionPatternRewriter &rewriter) const final {
    StringRef modelName = cast<SimModelInstanceType>(op.getInstance().getType())
                              .getModel()
                              .getValue();
 
    if (adaptor.getTimePostIncrement()) {
      // Increment time after step
      OpBuilder::InsertionGuard g(rewriter);
      rewriter.setInsertionPointAfter(op);
      auto oldTime =
          arc::SimGetTimeOp::create(rewriter, op.getLoc(), op.getInstance());
      auto newTime = LLVM::AddOp::create(rewriter, op.getLoc(), oldTime,
                                         adaptor.getTimePostIncrement());
      arc::SimSetTimeOp::create(rewriter, op.getLoc(), op.getInstance(),
                                newTime);
    }
 
    StringAttr evalFunc =
        rewriter.getStringAttr(evalSymbolFromModelName(modelName));
    rewriter.replaceOpWithNewOp<LLVM::CallOp>(op, mlir::TypeRange(), evalFunc,
                                              adaptor.getInstance());
 
    return success();
  }
};
 
// Loads the simulation time (i64 femtoseconds) from byte offset 0 in the
// model instance's state storage.
struct SimGetTimeOpLowering : public OpConversionPattern<arc::SimGetTimeOp> {
  using OpConversionPattern::OpConversionPattern;
 
  LogicalResult
  matchAndRewrite(arc::SimGetTimeOp op, OpAdaptor adaptor,
                  ConversionPatternRewriter &rewriter) const final {
    // Time is stored at offset 0 in the instance storage.
    rewriter.replaceOpWithNewOp<LLVM::LoadOp>(op, rewriter.getI64Type(),
                                              adaptor.getInstance());
    return success();
  }
};
 
// Stores the simulation time (i64 femtoseconds) to byte offset 0 in the
// model instance's state storage.
struct SimSetTimeOpLowering : public OpConversionPattern<arc::SimSetTimeOp> {
  using OpConversionPattern::OpConversionPattern;
 
  LogicalResult
  matchAndRewrite(arc::SimSetTimeOp op, OpAdaptor adaptor,
                  ConversionPatternRewriter &rewriter) const final {
    // Time is stored at offset 0 in the instance storage.
    rewriter.replaceOpWithNewOp<LLVM::StoreOp>(op, adaptor.getTime(),
                                               adaptor.getInstance());
    return success();
  }
};
 
// Loads the next wakeup time (i64 femtoseconds) from `kNextWakeupOffset` of
// the model instance's state storage.
struct SimGetNextWakeupOpLowering
    : public OpConversionPattern<arc::SimGetNextWakeupOp> {
  using OpConversionPattern::OpConversionPattern;
 
  LogicalResult
  matchAndRewrite(arc::SimGetNextWakeupOp op, OpAdaptor adaptor,
                  ConversionPatternRewriter &rewriter) const final {
    auto loc = op.getLoc();
    auto ptrType = LLVM::LLVMPointerType::get(rewriter.getContext());
    Value slotPtr = LLVM::GEPOp::create(
        rewriter, loc, ptrType, rewriter.getI8Type(), adaptor.getInstance(),
        ArrayRef<LLVM::GEPArg>{arc::kNextWakeupOffset});
    rewriter.replaceOpWithNewOp<LLVM::LoadOp>(op, rewriter.getI64Type(),
                                              slotPtr);
    return success();
  }
};
 
// Global string constants in the module.
class StringCache {
public:
  Value getOrCreate(OpBuilder &b, StringRef formatStr) {
    auto it = cache.find(formatStr);
    if (it != cache.end()) {
      return LLVM::AddressOfOp::create(b, b.getUnknownLoc(), it->second);
    }
 
    Location loc = b.getUnknownLoc();
    LLVM::GlobalOp global;
    {
      OpBuilder::InsertionGuard guard(b);
      ModuleOp m =
          b.getInsertionBlock()->getParent()->getParentOfType<ModuleOp>();
      b.setInsertionPointToStart(m.getBody());
 
      SmallVector<char> strVec(formatStr.begin(), formatStr.end());
      strVec.push_back(0);
 
      auto name = llvm::formatv("_arc_str_{0}", cache.size()).str();
      auto globalType = LLVM::LLVMArrayType::get(b.getI8Type(), strVec.size());
      global = LLVM::GlobalOp::create(b, loc, globalType, /*isConstant=*/true,
                                      LLVM::Linkage::Internal,
                                      /*name=*/name, b.getStringAttr(strVec),
                                      /*alignment=*/0);
    }
 
    cache[formatStr] = global;
    return LLVM::AddressOfOp::create(b, loc, global);
  }
 
private:
  llvm::StringMap<LLVM::GlobalOp> cache;
};
 
FailureOr<LLVM::CallOp> emitPrintfCall(OpBuilder &builder, Location loc,
                                       StringCache &cache, StringRef formatStr,
                                       ValueRange args) {
  ModuleOp moduleOp =
      builder.getInsertionBlock()->getParent()->getParentOfType<ModuleOp>();
  // Lookup or create printf function symbol.
  MLIRContext *ctx = builder.getContext();
  auto printfFunc = LLVM::lookupOrCreateFn(builder, moduleOp, "printf",
                                           LLVM::LLVMPointerType::get(ctx),
                                           LLVM::LLVMVoidType::get(ctx), true);
  if (failed(printfFunc))
    return printfFunc;
 
  Value formatStrPtr = cache.getOrCreate(builder, formatStr);
  SmallVector<Value> argsVec(1, formatStrPtr);
  argsVec.append(args.begin(), args.end());
  return LLVM::CallOp::create(builder, loc, printfFunc.value(), argsVec);
}
 
/// Lowers SimEmitValueOp to a printf call. The integer will be printed in its
/// entirety if it is of size up to size_t, and explicitly truncated otherwise.
/// This pattern will mutate the global module.
struct SimEmitValueOpLowering
    : public OpConversionPattern<arc::SimEmitValueOp> {
  SimEmitValueOpLowering(const TypeConverter &typeConverter,
                         MLIRContext *context, StringCache &formatStringCache)
      : OpConversionPattern(typeConverter, context),
        formatStringCache(formatStringCache) {}
 
  LogicalResult
  matchAndRewrite(arc::SimEmitValueOp op, OpAdaptor adaptor,
                  ConversionPatternRewriter &rewriter) const final {
    auto valueType = dyn_cast<IntegerType>(adaptor.getValue().getType());
    if (!valueType)
      return failure();
 
    Location loc = op.getLoc();
 
    ModuleOp moduleOp = op->getParentOfType<ModuleOp>();
    if (!moduleOp)
      return failure();
 
    SmallVector<Value> printfVariadicArgs;
    SmallString<16> printfFormatStr;
    int remainingBits = valueType.getWidth();
    Value value = adaptor.getValue();
 
    // Assumes the target platform uses 64bit for long long ints (%llx
    // formatter).
    constexpr llvm::StringRef intFormatter = "llx";
    auto intType = IntegerType::get(getContext(), 64);
    Value shiftValue = LLVM::ConstantOp::create(
        rewriter, loc, rewriter.getIntegerAttr(valueType, intType.getWidth()));
 
    if (valueType.getWidth() < intType.getWidth()) {
      int width = llvm::divideCeil(valueType.getWidth(), 4);
      printfFormatStr = llvm::formatv("%0{0}{1}", width, intFormatter);
      printfVariadicArgs.push_back(
          LLVM::ZExtOp::create(rewriter, loc, intType, value));
    } else {
      // Process the value in 64 bit chunks, starting from the least significant
      // bits. Since we append chunks in low-to-high order, we reverse the
      // vector to print them in the correct high-to-low order.
      int otherChunkWidth = intType.getWidth() / 4;
      int firstChunkWidth =
          llvm::divideCeil(valueType.getWidth() % intType.getWidth(), 4);
      if (firstChunkWidth == 0) { // print the full 64-bit hex or a subset.
        firstChunkWidth = otherChunkWidth;
      }
 
      std::string firstChunkFormat =
          llvm::formatv("%0{0}{1}", firstChunkWidth, intFormatter);
      std::string otherChunkFormat =
          llvm::formatv("%0{0}{1}", otherChunkWidth, intFormatter);
 
      for (int i = 0; remainingBits > 0; ++i) {
        // Append 64-bit chunks to the printf arguments, in low-to-high
        // order. The integer is printed in hex format with zero padding.
        printfVariadicArgs.push_back(
            LLVM::TruncOp::create(rewriter, loc, intType, value));
 
        // Zero-padded format specifier for fixed width, e.g. %01llx for 4 bits.
        printfFormatStr.append(i == 0 ? firstChunkFormat : otherChunkFormat);
 
        value =
            LLVM::LShrOp::create(rewriter, loc, value, shiftValue).getResult();
        remainingBits -= intType.getWidth();
      }
    }
 
    std::reverse(printfVariadicArgs.begin(), printfVariadicArgs.end());
 
    SmallString<16> formatStr = adaptor.getValueName();
    formatStr.append(" = ");
    formatStr.append(printfFormatStr);
    formatStr.append("\n");
 
    auto callOp = emitPrintfCall(rewriter, op->getLoc(), formatStringCache,
                                 formatStr, printfVariadicArgs);
    if (failed(callOp))
      return failure();
    rewriter.replaceOp(op, *callOp);
 
    return success();
  }
 
  StringCache &formatStringCache;
};
 
//===----------------------------------------------------------------------===//
// `sim` dialect lowerings
//===----------------------------------------------------------------------===//
 
// Helper struct to hold the format string and arguments for arcRuntimeFormat.
struct FormatInfo {
  SmallVector<FmtDescriptor> descriptors;
  SmallVector<Value> args;
};
 
// Copies the given integer value into an alloca, returning a pointer to it.
//
// The alloca is rounded up to a 64-bit boundary and is written as little-endian
// words of size 64-bits, to be compatible with the constructor of APInt.
static Value reg2mem(ConversionPatternRewriter &rewriter, Location loc,
                     Value value) {
  // Round up the type size to a 64-bit boundary.
  int64_t origBitwidth = cast<IntegerType>(value.getType()).getWidth();
  int64_t bitwidth = llvm::divideCeil(origBitwidth, 64) * 64;
  int64_t numWords = bitwidth / 64;
 
  // Create an alloca for the rounded up type.
  LLVM::ConstantOp alloca_size =
      LLVM::ConstantOp::create(rewriter, loc, rewriter.getI32Type(), numWords);
  auto ptrType = LLVM::LLVMPointerType::get(rewriter.getContext());
  auto allocaOp = LLVM::AllocaOp::create(rewriter, loc, ptrType,
                                         rewriter.getI64Type(), alloca_size);
  LLVM::LifetimeStartOp::create(rewriter, loc, allocaOp);
 
  // Copy `value` into the alloca, 64-bits at a time from the least significant
  // bits first.
  for (int64_t wordIdx = 0; wordIdx < numWords; ++wordIdx) {
    Value cst = LLVM::ConstantOp::create(
        rewriter, loc, rewriter.getIntegerType(origBitwidth), wordIdx * 64);
    Value v = LLVM::LShrOp::create(rewriter, loc, value, cst);
    if (origBitwidth > 64) {
      v = LLVM::TruncOp::create(rewriter, loc, rewriter.getI64Type(), v);
    } else if (origBitwidth < 64) {
      v = LLVM::ZExtOp::create(rewriter, loc, rewriter.getI64Type(), v);
    }
    Value gep = LLVM::GEPOp::create(rewriter, loc, ptrType,
                                    rewriter.getI64Type(), allocaOp, {wordIdx});
    LLVM::StoreOp::create(rewriter, loc, v, gep);
  }
 
  return allocaOp;
}
 
// Statically folds a value of type sim::FormatStringType to a FormatInfo.
static FailureOr<FormatInfo>
foldFormatString(ConversionPatternRewriter &rewriter, Value fstringValue,
                 StringCache &cache) {
  Operation *op = fstringValue.getDefiningOp();
  return llvm::TypeSwitch<Operation *, FailureOr<FormatInfo>>(op)
      .Case<sim::FormatCharOp>(
          [&](sim::FormatCharOp op) -> FailureOr<FormatInfo> {
            FmtDescriptor d = FmtDescriptor::createChar();
            return FormatInfo{{d}, {op.getValue()}};
          })
      .Case<sim::FormatDecOp>([&](sim::FormatDecOp op)
                                  -> FailureOr<FormatInfo> {
        FmtDescriptor d = FmtDescriptor::createInt(
            op.getValue().getType().getWidth(), 10, op.getIsLeftAligned(),
            op.getSpecifierWidth().value_or(-1), op.getPaddingChar(), false,
            op.getIsSigned());
        return FormatInfo{{d}, {reg2mem(rewriter, op.getLoc(), op.getValue())}};
      })
      .Case<sim::FormatHexOp>([&](sim::FormatHexOp op)
                                  -> FailureOr<FormatInfo> {
        FmtDescriptor d = FmtDescriptor::createInt(
            op.getValue().getType().getWidth(), 16, op.getIsLeftAligned(),
            op.getSpecifierWidth().value_or(-1), op.getPaddingChar(),
            op.getIsHexUppercase(), false);
        return FormatInfo{{d}, {reg2mem(rewriter, op.getLoc(), op.getValue())}};
      })
      .Case<sim::FormatOctOp>([&](sim::FormatOctOp op)
                                  -> FailureOr<FormatInfo> {
        FmtDescriptor d = FmtDescriptor::createInt(
            op.getValue().getType().getWidth(), 8, op.getIsLeftAligned(),
            op.getSpecifierWidth().value_or(-1), op.getPaddingChar(), false,
            false);
        return FormatInfo{{d}, {reg2mem(rewriter, op.getLoc(), op.getValue())}};
      })
      .Case<sim::FormatLiteralOp>(
          [&](sim::FormatLiteralOp op) -> FailureOr<FormatInfo> {
            if (op.getLiteral().size() < 8 &&
                op.getLiteral().find('\0') == StringRef::npos) {
              // We can use the small string optimization.
              FmtDescriptor d =
                  FmtDescriptor::createSmallLiteral(op.getLiteral());
              return FormatInfo{{d}, {}};
            }
            FmtDescriptor d =
                FmtDescriptor::createLiteral(op.getLiteral().size());
            Value value = cache.getOrCreate(rewriter, op.getLiteral());
            return FormatInfo{{d}, {value}};
          })
      .Case<sim::FormatStringConcatOp>(
          [&](sim::FormatStringConcatOp op) -> FailureOr<FormatInfo> {
            auto fmt = foldFormatString(rewriter, op.getInputs()[0], cache);
            if (failed(fmt))
              return failure();
            for (auto input : op.getInputs().drop_front()) {
              auto next = foldFormatString(rewriter, input, cache);
              if (failed(next))
                return failure();
              fmt->descriptors.append(next->descriptors);
              fmt->args.append(next->args);
            }
            return fmt;
          })
      .Default(
          [](Operation *op) -> FailureOr<FormatInfo> { return failure(); });
}
 
FailureOr<LLVM::CallOp> emitFmtCall(OpBuilder &builder, Location loc,
                                    StringCache &stringCache,
                                    ArrayRef<FmtDescriptor> descriptors,
                                    ValueRange args) {
  ModuleOp moduleOp =
      builder.getInsertionBlock()->getParent()->getParentOfType<ModuleOp>();
  // Lookup or create the arcRuntimeFormat function symbol.
  MLIRContext *ctx = builder.getContext();
  auto func = LLVM::lookupOrCreateFn(
      builder, moduleOp, runtime::APICallbacks::symNameFormat,
      LLVM::LLVMPointerType::get(ctx), LLVM::LLVMVoidType::get(ctx), true);
  if (failed(func))
    return func;
 
  StringRef rawDescriptors(reinterpret_cast<const char *>(descriptors.data()),
                           descriptors.size() * sizeof(FmtDescriptor));
  Value fmtPtr = stringCache.getOrCreate(builder, rawDescriptors);
 
  SmallVector<Value> argsVec(1, fmtPtr);
  argsVec.append(args.begin(), args.end());
  auto result = LLVM::CallOp::create(builder, loc, func.value(), argsVec);
 
  for (Value arg : args) {
    Operation *definingOp = arg.getDefiningOp();
    if (auto alloca = dyn_cast_if_present<LLVM::AllocaOp>(definingOp)) {
      LLVM::LifetimeEndOp::create(builder, loc, arg);
    }
  }
 
  return result;
}
 
struct SimPrintFormattedProcOpLowering
    : public OpConversionPattern<sim::PrintFormattedProcOp> {
  SimPrintFormattedProcOpLowering(const TypeConverter &typeConverter,
                                  MLIRContext *context,
                                  StringCache &stringCache)
      : OpConversionPattern<sim::PrintFormattedProcOp>(typeConverter, context),
        stringCache(stringCache) {}
 
  LogicalResult
  matchAndRewrite(sim::PrintFormattedProcOp op, OpAdaptor adaptor,
                  ConversionPatternRewriter &rewriter) const override {
    auto formatInfo = foldFormatString(rewriter, op.getInput(), stringCache);
    if (failed(formatInfo))
      return rewriter.notifyMatchFailure(op, "unsupported format string");
 
    // Add the end descriptor.
    formatInfo->descriptors.push_back(FmtDescriptor());
 
    auto result = emitFmtCall(rewriter, op.getLoc(), stringCache,
                              formatInfo->descriptors, formatInfo->args);
    if (failed(result))
      return failure();
    rewriter.replaceOp(op, result.value());
 
    return success();
  }
 
  StringCache &stringCache;
};
 
struct TerminateOpLowering : public OpConversionPattern<arc::TerminateOp> {
  using OpConversionPattern::OpConversionPattern;
 
  LogicalResult
  matchAndRewrite(arc::TerminateOp op, OpAdaptor adaptor,
                  ConversionPatternRewriter &rewriter) const override {
    auto loc = op.getLoc();
 
    auto i8Type = rewriter.getI8Type();
    auto ptrType = LLVM::LLVMPointerType::get(rewriter.getContext());
 
    Value flagPtr = LLVM::GEPOp::create(
        rewriter, loc, ptrType, i8Type, adaptor.getStorage(),
        ArrayRef<LLVM::GEPArg>{arc::kTerminateFlagOffset});
 
    uint8_t statusCode = op.getSuccess() ? 1 : 2;
    Value codeVal = LLVM::ConstantOp::create(
        rewriter, loc, i8Type, rewriter.getI8IntegerAttr(statusCode));
 
    LLVM::StoreOp::create(rewriter, loc, codeVal, flagPtr);
 
    rewriter.eraseOp(op);
    return success();
  }
};
 
// Loads the next wakeup time (i64 femtoseconds) from the model's storage at
// `kNextWakeupOffset`.
struct GetNextWakeupOpLowering
    : public OpConversionPattern<arc::GetNextWakeupOp> {
  using OpConversionPattern::OpConversionPattern;
 
  LogicalResult
  matchAndRewrite(arc::GetNextWakeupOp op, OpAdaptor adaptor,
                  ConversionPatternRewriter &rewriter) const override {
    auto loc = op.getLoc();
    auto ptrType = LLVM::LLVMPointerType::get(rewriter.getContext());
    Value slotPtr = LLVM::GEPOp::create(
        rewriter, loc, ptrType, rewriter.getI8Type(), adaptor.getStorage(),
        ArrayRef<LLVM::GEPArg>{arc::kNextWakeupOffset});
    rewriter.replaceOpWithNewOp<LLVM::LoadOp>(op, rewriter.getI64Type(),
                                              slotPtr);
    return success();
  }
};
 
// Stores the next wakeup time (i64 femtoseconds) to the model's storage at
// `kNextWakeupOffset`.
struct SetNextWakeupOpLowering
    : public OpConversionPattern<arc::SetNextWakeupOp> {
  using OpConversionPattern::OpConversionPattern;
 
  LogicalResult
  matchAndRewrite(arc::SetNextWakeupOp op, OpAdaptor adaptor,
                  ConversionPatternRewriter &rewriter) const override {
    auto loc = op.getLoc();
    auto ptrType = LLVM::LLVMPointerType::get(rewriter.getContext());
    Value slotPtr = LLVM::GEPOp::create(
        rewriter, loc, ptrType, rewriter.getI8Type(), adaptor.getStorage(),
        ArrayRef<LLVM::GEPArg>{arc::kNextWakeupOffset});
    rewriter.replaceOpWithNewOp<LLVM::StoreOp>(op, adaptor.getTime(), slotPtr);
    return success();
  }
};
 
} // namespace
 
static LogicalResult convert(arc::ExecuteOp op, arc::ExecuteOp::Adaptor adaptor,
                             ConversionPatternRewriter &rewriter,
                             const TypeConverter &converter) {
  // Convert the argument types in the body blocks.
  if (failed(rewriter.convertRegionTypes(&op.getBody(), converter)))
    return failure();
 
  // Split the block at the current insertion point such that we can branch into
  // the `arc.execute` body region, and have `arc.output` branch back to the
  // point after the `arc.execute`.
  auto *blockBefore = rewriter.getInsertionBlock();
  auto *blockAfter =
      rewriter.splitBlock(blockBefore, rewriter.getInsertionPoint());
 
  // Branch to the entry block.
  rewriter.setInsertionPointToEnd(blockBefore);
  mlir::cf::BranchOp::create(rewriter, op.getLoc(), &op.getBody().front(),
                             adaptor.getInputs());
 
  // Make all `arc.output` terminators branch to the block after the
  // `arc.execute` op.
  for (auto &block : op.getBody()) {
    auto outputOp = dyn_cast<arc::OutputOp>(block.getTerminator());
    if (!outputOp)
      continue;
    rewriter.setInsertionPointToEnd(&block);
    rewriter.replaceOpWithNewOp<mlir::cf::BranchOp>(outputOp, blockAfter,
                                                    outputOp.getOperands());
  }
 
  // Inline the body region between the before and after blocks.
  rewriter.inlineRegionBefore(op.getBody(), blockAfter);
 
  // Add arguments to the block after the `arc.execute`, replace the op's
  // results with the arguments, then perform block signature conversion.
  SmallVector<Value> args;
  args.reserve(op.getNumResults());
  for (auto result : op.getResults())
    args.push_back(blockAfter->addArgument(result.getType(), result.getLoc()));
  rewriter.replaceOp(op, args);
  auto conversion = converter.convertBlockSignature(blockAfter);
  if (!conversion)
    return failure();
  rewriter.applySignatureConversion(blockAfter, *conversion, &converter);
  return success();
}
 
//===----------------------------------------------------------------------===//
// Runtime Implementation
//===----------------------------------------------------------------------===//
 
template <typename T, typename = std::enable_if_t<std::is_integral<T>::value>>
static LLVM::GlobalOp
buildGlobalConstantIntArray(OpBuilder &builder, Location loc, Twine symName,
                            SmallVectorImpl<T> &data,
                            unsigned alignment = alignof(T)) {
  auto intType = builder.getIntegerType(8 * sizeof(T));
  Attribute denseAttr = mlir::DenseElementsAttr::get(
      mlir::RankedTensorType::get({(int64_t)data.size()}, intType),
      llvm::ArrayRef(data));
  auto globalOp = LLVM::GlobalOp::create(
      builder, loc, LLVM::LLVMArrayType::get(intType, data.size()),
      /*isConstant=*/true, LLVM::Linkage::Internal,
      builder.getStringAttr(symName), denseAttr);
  globalOp.setAlignmentAttr(builder.getI64IntegerAttr(alignment));
  return globalOp;
}
 
// Construct a raw constant byte array from a vector of struct values
template <typename T>
static LLVM::GlobalOp
buildGlobalConstantRuntimeStructArray(OpBuilder &builder, Location loc,
                                      Twine symName,
                                      SmallVectorImpl<T> &array) {
  assert(!array.empty());
  static_assert(std::is_standard_layout<T>(),
                "Runtime struct must have standard layout");
  int64_t numBytes = sizeof(T) * array.size();
  Attribute denseAttr = mlir::DenseElementsAttr::get(
      mlir::RankedTensorType::get({numBytes}, builder.getI8Type()),
      llvm::ArrayRef(reinterpret_cast<uint8_t *>(array.data()), numBytes));
  auto globalOp = LLVM::GlobalOp::create(
      builder, loc, LLVM::LLVMArrayType::get(builder.getI8Type(), numBytes),
      /*isConstant=*/true, LLVM::Linkage::Internal,
      builder.getStringAttr(symName), denseAttr, alignof(T));
  return globalOp;
}
 
struct RuntimeModelOpLowering
    : public OpConversionPattern<arc::RuntimeModelOp> {
  using OpConversionPattern::OpConversionPattern;
 
  static constexpr uint64_t runtimeApiVersion = ARC_RUNTIME_API_VERSION;
 
  // Build the constant ArcModelTraceInfo struct and its members
  LLVM::GlobalOp
  buildTraceInfoStruct(arc::RuntimeModelOp &op,
                       ConversionPatternRewriter &rewriter) const {
    if (!op.getTraceTaps().has_value() || op.getTraceTaps()->empty())
      return {};
    // Construct the array of tap names/aliases
    SmallVector<char> namesArray;
    SmallVector<ArcTraceTap> tapArray;
    tapArray.reserve(op.getTraceTaps()->size());
    for (auto attr : op.getTraceTapsAttr()) {
      auto tap = cast<TraceTapAttr>(attr);
      assert(!tap.getNames().empty() &&
             "Expected trace tap to have at least one name");
      for (auto alias : tap.getNames()) {
        auto aliasStr = cast<StringAttr>(alias);
        namesArray.append(aliasStr.begin(), aliasStr.end());
        namesArray.push_back('\0');
      }
      ArcTraceTap tapStruct;
      tapStruct.stateOffset = tap.getStateOffset();
      tapStruct.nameOffset = namesArray.size() - 1;
      tapStruct.typeBits = tap.getSigType().getValue().getIntOrFloatBitWidth();
      tapStruct.reserved = 0;
      tapArray.emplace_back(tapStruct);
    }
    auto ptrTy = LLVM::LLVMPointerType::get(getContext());
    auto namesGlobal = buildGlobalConstantIntArray(
        rewriter, op.getLoc(), "_arc_tap_names_" + op.getName(), namesArray);
    auto traceTapsArrayGlobal = buildGlobalConstantRuntimeStructArray(
        rewriter, op.getLoc(), "_arc_trace_taps_" + op.getName(), tapArray);
 
    //
    //  struct ArcModelTraceInfo {
    //    uint64_t numTraceTaps;
    //    struct ArcTraceTap *traceTaps;
    //    const char *traceTapNames;
    //    uint64_t traceBufferCapacity;
    //  };
    //
    auto traceInfoStructType = LLVM::LLVMStructType::getLiteral(
        getContext(),
        {rewriter.getI64Type(), ptrTy, ptrTy, rewriter.getI64Type()});
    static_assert(sizeof(ArcModelTraceInfo) == 32 &&
                  "Unexpected size of ArcModelTraceInfo struct");
 
    auto globalSymName =
        rewriter.getStringAttr("_arc_trace_info_" + op.getName());
    auto traceInfoGlobalOp = LLVM::GlobalOp::create(
        rewriter, op.getLoc(), traceInfoStructType,
        /*isConstant=*/false, LLVM::Linkage::Internal, globalSymName,
        Attribute{}, alignof(ArcModelTraceInfo));
    OpBuilder::InsertionGuard g(rewriter);
 
    // Struct Initializer
    Region &initRegion = traceInfoGlobalOp.getInitializerRegion();
    Block *initBlock = rewriter.createBlock(&initRegion);
    rewriter.setInsertionPointToStart(initBlock);
 
    auto numTraceTapsCst = LLVM::ConstantOp::create(
        rewriter, op.getLoc(), rewriter.getI64IntegerAttr(tapArray.size()));
    auto traceTapArrayAddr =
        LLVM::AddressOfOp::create(rewriter, op.getLoc(), traceTapsArrayGlobal);
    auto tapNameArrayAddr =
        LLVM::AddressOfOp::create(rewriter, op.getLoc(), namesGlobal);
    auto bufferCapacityCst = LLVM::ConstantOp::create(
        rewriter, op.getLoc(),
        rewriter.getI64IntegerAttr(runtime::defaultTraceBufferCapacity));
 
    Value initStruct =
        LLVM::PoisonOp::create(rewriter, op.getLoc(), traceInfoStructType);
 
    // Field: uint64_t numTraceTaps
    initStruct =
        LLVM::InsertValueOp::create(rewriter, op.getLoc(), initStruct,
                                    numTraceTapsCst, ArrayRef<int64_t>{0});
    static_assert(offsetof(ArcModelTraceInfo, numTraceTaps) == 0,
                  "Unexpected offset of field numTraceTaps");
    // Field: struct ArcTraceTap *traceTaps
    initStruct =
        LLVM::InsertValueOp::create(rewriter, op.getLoc(), initStruct,
                                    traceTapArrayAddr, ArrayRef<int64_t>{1});
    static_assert(offsetof(ArcModelTraceInfo, traceTaps) == 8,
                  "Unexpected offset of field traceTaps");
    // Field: const char *traceTapNames
    initStruct =
        LLVM::InsertValueOp::create(rewriter, op.getLoc(), initStruct,
                                    tapNameArrayAddr, ArrayRef<int64_t>{2});
    static_assert(offsetof(ArcModelTraceInfo, traceTapNames) == 16,
                  "Unexpected offset of field traceTapNames");
    // Field: uint64_t traceBufferCapacity
    initStruct =
        LLVM::InsertValueOp::create(rewriter, op.getLoc(), initStruct,
                                    bufferCapacityCst, ArrayRef<int64_t>{3});
    static_assert(offsetof(ArcModelTraceInfo, traceBufferCapacity) == 24,
                  "Unexpected offset of field traceBufferCapacity");
    LLVM::ReturnOp::create(rewriter, op.getLoc(), initStruct);
 
    return traceInfoGlobalOp;
  }
 
  // Create a global LLVM struct containing the RuntimeModel metadata
  LogicalResult
  matchAndRewrite(arc::RuntimeModelOp op, OpAdaptor adaptor,
                  ConversionPatternRewriter &rewriter) const final {
 
    auto ptrTy = LLVM::LLVMPointerType::get(getContext());
    auto modelInfoStructType = LLVM::LLVMStructType::getLiteral(
        getContext(),
        {rewriter.getI64Type(), rewriter.getI64Type(), ptrTy, ptrTy});
    static_assert(sizeof(ArcRuntimeModelInfo) == 32 &&
                  "Unexpected size of ArcRuntimeModelInfo struct");
 
    rewriter.setInsertionPoint(op);
    auto traceInfoGlobal = buildTraceInfoStruct(op, rewriter);
 
    // Construct the Model Name String GlobalOp
    SmallVector<char, 16> modNameArray(op.getName().begin(),
                                       op.getName().end());
    modNameArray.push_back('\0');
    auto nameGlobalType =
        LLVM::LLVMArrayType::get(rewriter.getI8Type(), modNameArray.size());
    auto globalSymName =
        rewriter.getStringAttr("_arc_mod_name_" + op.getName());
    auto nameGlobal = LLVM::GlobalOp::create(
        rewriter, op.getLoc(), nameGlobalType, /*isConstant=*/true,
        LLVM::Linkage::Internal,
        /*name=*/globalSymName, rewriter.getStringAttr(modNameArray),
        /*alignment=*/0);
 
    // Construct the Model Info Struct GlobalOp
    // Note: The struct is supposed to be constant at runtime, but contains the
    // relocatable address of another symbol, so it should not be placed in the
    // "rodata" section.
    auto modInfoGlobalOp =
        LLVM::GlobalOp::create(rewriter, op.getLoc(), modelInfoStructType,
                               /*isConstant=*/false, LLVM::Linkage::External,
                               op.getSymName(), Attribute{});
 
    // Struct Initializer
    Region &initRegion = modInfoGlobalOp.getInitializerRegion();
    Block *initBlock = rewriter.createBlock(&initRegion);
    rewriter.setInsertionPointToStart(initBlock);
    auto apiVersionCst = LLVM::ConstantOp::create(
        rewriter, op.getLoc(), rewriter.getI64IntegerAttr(runtimeApiVersion));
    auto numStateBytesCst = LLVM::ConstantOp::create(rewriter, op.getLoc(),
                                                     op.getNumStateBytesAttr());
    auto nameAddr =
        LLVM::AddressOfOp::create(rewriter, op.getLoc(), nameGlobal);
    Value traceInfoPtr;
    if (traceInfoGlobal)
      traceInfoPtr =
          LLVM::AddressOfOp::create(rewriter, op.getLoc(), traceInfoGlobal);
    else
      traceInfoPtr = LLVM::ZeroOp::create(rewriter, op.getLoc(), ptrTy);
 
    Value initStruct =
        LLVM::PoisonOp::create(rewriter, op.getLoc(), modelInfoStructType);
 
    // Field: uint64_t apiVersion
    initStruct = LLVM::InsertValueOp::create(
        rewriter, op.getLoc(), initStruct, apiVersionCst, ArrayRef<int64_t>{0});
    static_assert(offsetof(ArcRuntimeModelInfo, apiVersion) == 0,
                  "Unexpected offset of field apiVersion");
    // Field: uint64_t numStateBytes
    initStruct =
        LLVM::InsertValueOp::create(rewriter, op.getLoc(), initStruct,
                                    numStateBytesCst, ArrayRef<int64_t>{1});
    static_assert(offsetof(ArcRuntimeModelInfo, numStateBytes) == 8,
                  "Unexpected offset of field numStateBytes");
    // Field: const char *modelName
    initStruct = LLVM::InsertValueOp::create(rewriter, op.getLoc(), initStruct,
                                             nameAddr, ArrayRef<int64_t>{2});
    static_assert(offsetof(ArcRuntimeModelInfo, modelName) == 16,
                  "Unexpected offset of field modelName");
    // Field: struct ArcModelTraceInfo *traceInfo
    initStruct = LLVM::InsertValueOp::create(
        rewriter, op.getLoc(), initStruct, traceInfoPtr, ArrayRef<int64_t>{3});
    static_assert(offsetof(ArcRuntimeModelInfo, traceInfo) == 24,
                  "Unexpected offset of field traceInfo");
 
    LLVM::ReturnOp::create(rewriter, op.getLoc(), initStruct);
 
    rewriter.replaceOp(op, modInfoGlobalOp);
    return success();
  }
};
 
//===----------------------------------------------------------------------===//
// ArrayRef patterns
//===----------------------------------------------------------------------===//
 
size_t computeByteWidth(ArrayRefType type) {
  auto bitWidth = computeLLVMBitWidth(type);
  assert(bitWidth.has_value());
  return llvm::divideCeil(*bitWidth, 8);
}
 
// Computes the padded bytewidth (stride) of each element.
size_t computeElementByteWidth(ArrayRefType arrayRefType) {
  auto arrayBitWidth = computeLLVMBitWidth(arrayRefType);
  assert(arrayBitWidth.has_value());
  assert(arrayRefType.getNumElements() > 0 &&
         "Cannot compute stride for zero sized array");
  size_t elementBitWidth = *arrayBitWidth / arrayRefType.getNumElements();
  return llvm::divideCeil(elementBitWidth, 8);
}
 
struct ArrayRefAllocOpLowering : public OpConversionPattern<ArrayRefAllocOp> {
  using OpConversionPattern::OpConversionPattern;
 
  LogicalResult
  matchAndRewrite(ArrayRefAllocOp op, OpAdaptor adaptor,
                  ConversionPatternRewriter &rewriter) const override {
    auto ptrTy = LLVM::LLVMPointerType::get(getContext());
    auto i8Ty = rewriter.getI8Type();
    ArrayRefType arrayRefType = op.getType();
    size_t byteWidth = computeByteWidth(arrayRefType);
    auto size = LLVM::ConstantOp::create(rewriter, op.getLoc(),
                                         rewriter.getI64Type(), byteWidth);
 
    size_t alignment = computeAllocaAlignment(arrayRefType, op);
    auto alloc = LLVM::AllocaOp::create(rewriter, op.getLoc(), ptrTy, i8Ty,
                                        size, alignment);
 
    if (op.getInitAttr()) {
      ArrayAttr initAttr = op.getInitAttr();
      if (isZero(initAttr)) {
        auto i8Ty = rewriter.getI8Type();
        auto zero = LLVM::ConstantOp::create(rewriter, op.getLoc(), i8Ty, 0);
        LLVM::MemsetOp::create(rewriter, op.getLoc(), alloc, zero, size,
                               /*isVolatile=*/false);
      } else {
        initializeArray(rewriter, op.getLoc(), alloc, initAttr, arrayRefType);
      }
    }
 
    rewriter.replaceOp(op, alloc);
    return success();
  }
 
  // Computes the required alignment for an AllocaOp of the given type.
  // c.f. HWToLLVM.cpp.
  size_t computeAllocaAlignment(ArrayRefType type, Operation *op) const {
    if (alignmentCache.count(type)) {
      return alignmentCache[type];
    }
    auto dl = DataLayout::closest(op);
    auto hwType =
        hw::ArrayType::get(type.getElementType(), type.getNumElements());
    auto llvmType = getTypeConverter()->convertType(hwType);
    auto alignment =
        static_cast<unsigned>(dl.getTypePreferredAlignment(llvmType));
    alignment = std::max(4u, alignment);
    alignmentCache[type] = alignment;
    return alignment;
  }
 
  bool isZero(ArrayAttr arrayAttr) const {
    return llvm::all_of(arrayAttr.getAsValueRange<IntegerAttr>(),
                        [](APInt i) { return i.isZero(); });
  }
 
  void initializeArray(ConversionPatternRewriter &rewriter, Location loc,
                       Value alloc, ArrayAttr initAttr,
                       ArrayRefType arrayRefType) const {
    size_t elemByteWidth = computeElementByteWidth(arrayRefType);
    Type ptrTy = LLVM::LLVMPointerType::get(getContext());
    Type i8Ty = rewriter.getI8Type();
    for (unsigned i = 0; i < arrayRefType.getNumElements(); ++i) {
      unsigned elemIndex = arrayRefType.getNumElements() - i - 1;
      Value elemOffset = LLVM::ConstantOp::create(
          rewriter, loc, rewriter.getI64Type(), elemIndex * elemByteWidth);
      auto elemAddr =
          LLVM::GEPOp::create(rewriter, loc, ptrTy, i8Ty, alloc, elemOffset);
      auto elem = LLVM::ConstantOp::create(
          rewriter, loc, arrayRefType.getElementType(), initAttr[i]);
      LLVM::StoreOp::create(rewriter, loc, elem, elemAddr);
    }
  }
 
private:
  mutable DenseMap<ArrayRefType, size_t> alignmentCache;
};
 
struct ArrayRefCreateOpLowering : public OpConversionPattern<ArrayRefCreateOp> {
  using OpConversionPattern::OpConversionPattern;
 
  LogicalResult
  matchAndRewrite(ArrayRefCreateOp op, OpAdaptor adaptor,
                  ConversionPatternRewriter &rewriter) const override {
    ArrayRefType arrayRefType = cast<ArrayRefType>(op.getType());
    Value alloc = adaptor.getInput();
    auto ptrTy = LLVM::LLVMPointerType::get(getContext());
    auto i8Ty = rewriter.getI8Type();
    size_t elemByteWidth = computeElementByteWidth(arrayRefType);
    auto elements = adaptor.getElements();
    for (unsigned i = 0; i < elements.size(); ++i) {
      // Note: hardcoded for little endian targets.
      unsigned elemIndex = arrayRefType.getNumElements() - i - 1;
      Value elemOffset =
          LLVM::ConstantOp::create(rewriter, op.getLoc(), rewriter.getI64Type(),
                                   elemIndex * elemByteWidth);
      auto elemAddr = LLVM::GEPOp::create(rewriter, op.getLoc(), ptrTy, i8Ty,
                                          alloc, elemOffset);
      LLVM::StoreOp::create(rewriter, op.getLoc(), elements[i], elemAddr);
    }
    rewriter.replaceOp(op, alloc);
    return success();
  }
};
 
struct ArrayRefGetOpLowering : public OpConversionPattern<ArrayRefGetOp> {
  using OpConversionPattern::OpConversionPattern;
 
  LogicalResult
  matchAndRewrite(ArrayRefGetOp op, OpAdaptor adaptor,
                  ConversionPatternRewriter &rewriter) const override {
    auto loc = op.getLoc();
    ArrayRefType arrayRefType = cast<ArrayRefType>(op.getInput().getType());
    auto ptrTy = LLVM::LLVMPointerType::get(getContext());
    auto i8Ty = rewriter.getI8Type();
    auto i64Ty = rewriter.getI64Type();
    size_t elemByteWidth = computeElementByteWidth(arrayRefType);
    assert(!isa<ArrayRefType>(arrayRefType.getElementType()));
 
    Value stride =
        LLVM::ConstantOp::create(rewriter, loc, i64Ty, elemByteWidth);
    Value byteOffset =
        LLVM::MulOp::create(rewriter, loc, adaptor.getIndex(), stride);
    // Defend against out-of-bounds accesses. What we return is undefined in the
    // case of OOB.
    size_t lastElementByteOffset =
        elemByteWidth * (arrayRefType.getNumElements() - 1);
    Value lastElementByteOffsetVal =
        LLVM::ConstantOp::create(rewriter, loc, i64Ty, lastElementByteOffset);
    Value clampedOffset = LLVM::UMinOp::create(rewriter, loc, i64Ty, byteOffset,
                                               lastElementByteOffsetVal);
    auto elemAddr = LLVM::GEPOp::create(rewriter, loc, ptrTy, i8Ty,
                                        adaptor.getInput(), clampedOffset);
    Value loaded = LLVM::LoadOp::create(
        rewriter, loc, typeConverter->convertType(op.getValue().getType()),
        elemAddr);
    rewriter.replaceOp(op, loaded);
    return success();
  }
};
 
struct ArrayRefInjectOpLowering : public OpConversionPattern<ArrayRefInjectOp> {
  using OpConversionPattern::OpConversionPattern;
 
  LogicalResult
  matchAndRewrite(ArrayRefInjectOp op, OpAdaptor adaptor,
                  ConversionPatternRewriter &rewriter) const override {
    auto loc = op.getLoc();
    ArrayRefType arrayRefType = cast<ArrayRefType>(op.getInput().getType());
    assert(!isa<ArrayRefType>(arrayRefType.getElementType()));
    auto ptrTy = LLVM::LLVMPointerType::get(getContext());
    auto i8Ty = rewriter.getI8Type();
    auto i64Ty = rewriter.getI64Type();
    size_t byteWidth = computeByteWidth(arrayRefType);
    size_t elemByteWidth = computeElementByteWidth(arrayRefType);
 
    Value stride =
        LLVM::ConstantOp::create(rewriter, loc, i64Ty, elemByteWidth);
    Value byteOffset =
        LLVM::MulOp::create(rewriter, loc, adaptor.getIndex(), stride);
    Value totalSize = LLVM::ConstantOp::create(rewriter, loc, i64Ty, byteWidth);
    // Defend against out-of-bounds accesses. We must avoid corrupting the
    // array.
    Value isInbounds = LLVM::ICmpOp::create(
        rewriter, loc, LLVM::ICmpPredicate::ult, byteOffset, totalSize);
    scf::IfOp::create(rewriter, loc, isInbounds, [&](OpBuilder &b, Location) {
      auto elemAddr = LLVM::GEPOp::create(b, loc, ptrTy, i8Ty,
                                          adaptor.getInput(), byteOffset);
      LLVM::StoreOp::create(b, loc, adaptor.getElement(), elemAddr);
      scf::YieldOp::create(b, loc);
    });
 
    // Inject is pure; returns the same pointer (input buffer is modified
    // in-place and the pointer is forwarded as the result).
    rewriter.replaceOp(op, adaptor.getInput());
    return success();
  }
};
 
struct ArrayRefSliceOpLowering : public OpConversionPattern<ArrayRefSliceOp> {
  using OpConversionPattern::OpConversionPattern;
 
  LogicalResult
  matchAndRewrite(ArrayRefSliceOp op, OpAdaptor adaptor,
                  ConversionPatternRewriter &rewriter) const override {
    auto loc = op.getLoc();
    // The result type is the sub-array type; use its element size.
    ArrayRefType inputType = cast<ArrayRefType>(op.getInput().getType());
    ArrayRefType resultType = cast<ArrayRefType>(op.getOutput().getType());
    auto ptrTy = LLVM::LLVMPointerType::get(getContext());
    auto i8Ty = rewriter.getI8Type();
    auto i64Ty = rewriter.getI64Type();
    size_t elemByteWidth = computeElementByteWidth(resultType);
 
    // Ensure the slice doesn't go out of bounds.
    size_t maxLowIndex =
        inputType.getNumElements() - resultType.getNumElements();
    Value maxLowIndexVal =
        LLVM::ConstantOp::create(rewriter, loc, i64Ty, maxLowIndex);
    Value clampedLowIndex = LLVM::UMinOp::create(
        rewriter, loc, i64Ty, adaptor.getLowIndex(), maxLowIndexVal);
 
    // Byte offset = lowIndex * elemByteWidth.
    Value stride =
        LLVM::ConstantOp::create(rewriter, loc, i64Ty, elemByteWidth);
    Value byteOffset =
        LLVM::MulOp::create(rewriter, loc, clampedLowIndex, stride);
    auto sliceAddr = LLVM::GEPOp::create(rewriter, loc, ptrTy, i8Ty,
                                         adaptor.getInput(), byteOffset);
    rewriter.replaceOp(op, sliceAddr);
    return success();
  }
};
 
struct ArrayRefCopyOpLowering : public OpConversionPattern<ArrayRefCopyOp> {
  using OpConversionPattern::OpConversionPattern;
 
  LogicalResult
  matchAndRewrite(ArrayRefCopyOp op, OpAdaptor adaptor,
                  ConversionPatternRewriter &rewriter) const override {
    auto loc = op.getLoc();
    ArrayRefType arrayRefType = cast<ArrayRefType>(op.getInput().getType());
    auto i64Ty = rewriter.getI64Type();
    size_t byteWidth = computeByteWidth(arrayRefType);
    Value size = LLVM::ConstantOp::create(rewriter, loc, i64Ty, byteWidth);
    // Use a memmove rather than a memcpy just in case the arrays alias.
    LLVM::MemmoveOp::create(rewriter, loc, adaptor.getInput(),
                            adaptor.getSource(), size,
                            /*isVolatile=*/false);
    rewriter.replaceOp(op, adaptor.getInput());
    return success();
  }
};
 
static Value loadArrayRefAsArray(ImplicitLocOpBuilder &builder, Value arrayRef,
                                 ArrayRefType arrayRefType,
                                 LLVM::LLVMArrayType llvmType) {
  auto i8Ty = builder.getI8Type();
  auto ptrTy = LLVM::LLVMPointerType::get(builder.getContext());
  size_t elemByteWidth = computeElementByteWidth(arrayRefType);
  Value v = LLVM::PoisonOp::create(builder, llvmType);
  int32_t size = arrayRefType.getNumElements();
  for (int32_t i = 0; i < size; i++) {
    int32_t byteOffset = i * elemByteWidth;
    Value gep = LLVM::GEPOp::create(builder, ptrTy, i8Ty, arrayRef,
                                    LLVM::GEPArg{byteOffset});
    Value load = LLVM::LoadOp::create(builder, llvmType.getElementType(), gep);
    v = LLVM::InsertValueOp::create(builder, v, load, i);
  }
  return v;
}
 
static void storeArrayAsArrayRef(ImplicitLocOpBuilder &builder, Value array,
                                 Value arrayRef, ArrayRefType arrayRefType) {
  auto i8Ty = builder.getI8Type();
  auto ptrTy = LLVM::LLVMPointerType::get(builder.getContext());
  size_t elemByteWidth = computeElementByteWidth(arrayRefType);
  int32_t size = arrayRefType.getNumElements();
  for (int32_t i = 0; i < size; i++) {
    int32_t byteOffset = i * elemByteWidth;
    Value gep = LLVM::GEPOp::create(builder, ptrTy, i8Ty, arrayRef,
                                    LLVM::GEPArg{byteOffset});
    Value val = LLVM::ExtractValueOp::create(builder, array, i);
    LLVM::StoreOp::create(builder, val, gep);
  }
}
 
struct ArrayRefToLLVMArrayOpLowering
    : public OpConversionPattern<UnrealizedConversionCastOp> {
  using OpConversionPattern::OpConversionPattern;
 
  LogicalResult
  matchAndRewrite(UnrealizedConversionCastOp op, OpAdaptor adaptor,
                  ConversionPatternRewriter &rewriter) const override {
    if (!isa<ArrayRefType>(op.getOperand(0).getType()) ||
        !isa<LLVM::LLVMArrayType>(op.getResult(0).getType())) {
      return failure();
    }
 
    ImplicitLocOpBuilder b(op.getLoc(), rewriter);
    Value loaded = loadArrayRefAsArray(
        b, adaptor.getInputs().front(),
        cast<ArrayRefType>(op.getOperand(0).getType()),
        cast<LLVM::LLVMArrayType>(op.getResult(0).getType()));
    rewriter.replaceOp(op, loaded);
    return success();
  }
};
 
struct ArrayRefToArrayOpLowering
    : public OpConversionPattern<ArrayRefToArrayOp> {
  using OpConversionPattern::OpConversionPattern;
 
  LogicalResult
  matchAndRewrite(ArrayRefToArrayOp op, OpAdaptor adaptor,
                  ConversionPatternRewriter &rewriter) const override {
    Type resultType = getTypeConverter()->convertType(op.getResult().getType());
    ImplicitLocOpBuilder b(op.getLoc(), rewriter);
    Value loaded = loadArrayRefAsArray(
        b, adaptor.getInput(), cast<ArrayRefType>(op.getInput().getType()),
        cast<LLVM::LLVMArrayType>(resultType));
    rewriter.replaceOp(op, loaded);
    return success();
  }
};
 
struct ArrayRefFromArrayOpLowering
    : public OpConversionPattern<ArrayRefFromArrayOp> {
  using OpConversionPattern::OpConversionPattern;
 
  LogicalResult
  matchAndRewrite(ArrayRefFromArrayOp op, OpAdaptor adaptor,
                  ConversionPatternRewriter &rewriter) const override {
    ImplicitLocOpBuilder b(op.getLoc(), rewriter);
    storeArrayAsArrayRef(b, adaptor.getArray(), adaptor.getInput(),
                         cast<ArrayRefType>(op.getInput().getType()));
    rewriter.replaceOp(op, adaptor.getInput());
    return success();
  }
};
 
//===----------------------------------------------------------------------===//
// Pass Implementation
//===----------------------------------------------------------------------===//
 
namespace {
struct LowerArcToLLVMPass
    : public circt::impl::LowerArcToLLVMBase<LowerArcToLLVMPass> {
  void runOnOperation() override;
};
} // namespace
 
void LowerArcToLLVMPass::runOnOperation() {
  // Add `dereferenceable(<N>)` attributes to all function arguments that take
  // ArrayRefTypes.
  for (func::FuncOp func : getOperation().getOps<func::FuncOp>()) {
    for (int i = 0, e = func.getNumArguments(); i != e; ++i) {
      if (auto arrayRefType =
              dyn_cast<ArrayRefType>(func.getArgumentTypes()[i])) {
        size_t byteWidth = computeByteWidth(arrayRefType);
        Builder builder(&getContext());
        func.setArgAttr(i, LLVM::LLVMDialect::getDereferenceableAttrName(),
                        builder.getI64IntegerAttr(byteWidth));
      }
    }
  }
 
  // Collect the symbols in the root op such that the HW-to-LLVM lowering can
  // create LLVM globals with non-colliding names.
  Namespace globals;
  SymbolCache cache;
  cache.addDefinitions(getOperation());
  globals.add(cache);
 
  // Setup the conversion target. Explicitly mark `scf.yield` legal since it
  // does not have a conversion itself, which would cause it to fail
  // legalization and for the conversion to abort. (It relies on its parent op's
  // conversion to remove it.)
  LLVMConversionTarget target(getContext());
  target.addLegalOp<mlir::ModuleOp>();
  target.addLegalOp<scf::YieldOp>(); // quirk of SCF dialect conversion
 
  // Mark sim::Format*Op as legal. These are not converted to LLVM, but the
  // lowering of sim::PrintFormattedOp walks them to build up its format string.
  // They are all marked Pure so are removed after the conversion.
  target.addLegalOp<sim::FormatLiteralOp, sim::FormatDecOp, sim::FormatHexOp,
                    sim::FormatBinOp, sim::FormatOctOp, sim::FormatCharOp,
                    sim::FormatStringConcatOp>();
 
  // Setup the arc dialect type conversion.
  LLVMTypeConverter converter(&getContext());
  converter.addConversion([&](seq::ClockType type) {
    return IntegerType::get(type.getContext(), 1);
  });
  converter.addConversion([&](StorageType type) {
    return LLVM::LLVMPointerType::get(type.getContext());
  });
  converter.addConversion([&](MemoryType type) {
    return LLVM::LLVMPointerType::get(type.getContext());
  });
  converter.addConversion([&](StateType type) {
    return LLVM::LLVMPointerType::get(type.getContext());
  });
  converter.addConversion([&](SimModelInstanceType type) {
    return LLVM::LLVMPointerType::get(type.getContext());
  });
  converter.addConversion([&](sim::FormatStringType type) {
    return LLVM::LLVMPointerType::get(type.getContext());
  });
  converter.addConversion([&](llhd::TimeType type) {
    // LLHD time is represented as i64 femtoseconds.
    return IntegerType::get(type.getContext(), 64);
  });
  converter.addConversion([&](ArrayRefType type) {
    return LLVM::LLVMPointerType::get(type.getContext());
  });
 
  // Convert an UnrealizedConversionCastOp from !arc.arrayref<T> to
  // !llvm.array<T>. These are inserted by the InsertRuntime pass.
  target.addDynamicallyLegalOp<UnrealizedConversionCastOp>([&](Operation *op) {
    Type src = op->getOperand(0).getType();
    Type dst = op->getResult(0).getType();
    bool needsConvert = isa<ArrayRefType>(src) && isa<LLVM::LLVMArrayType>(dst);
    return !needsConvert;
  });
 
  // Setup the conversion patterns.
  ConversionPatternSet patterns(&getContext(), converter);
 
  // MLIR patterns.
  populateSCFToControlFlowConversionPatterns(patterns);
  populateFuncToLLVMConversionPatterns(converter, patterns);
  cf::populateControlFlowToLLVMConversionPatterns(converter, patterns);
  arith::populateArithToLLVMConversionPatterns(converter, patterns);
  index::populateIndexToLLVMConversionPatterns(converter, patterns);
  populateAnyFunctionOpInterfaceTypeConversionPattern(patterns, converter);
 
  // CIRCT patterns.
  DenseMap<std::pair<Type, ArrayAttr>, LLVM::GlobalOp> constAggregateGlobalsMap;
  populateHWToLLVMTypeConversions(converter);
  std::optional<HWToLLVMArraySpillCache> spillCacheOpt =
      HWToLLVMArraySpillCache();
  {
    OpBuilder spillBuilder(getOperation());
    spillCacheOpt->spillNonHWOps(spillBuilder, converter, getOperation());
  }
  populateHWToLLVMConversionPatterns(converter, patterns, globals,
                                     constAggregateGlobalsMap, spillCacheOpt);
 
  populateCombToArithConversionPatterns(converter, patterns);
  populateCombToLLVMConversionPatterns(converter, patterns);
 
  // Arc patterns.
  // clang-format off
  patterns.add<
    AllocMemoryOpLowering,
    AllocStateLikeOpLowering<arc::AllocStateOp>,
    AllocStateLikeOpLowering<arc::RootInputOp>,
    AllocStateLikeOpLowering<arc::RootOutputOp>,
    AllocStorageOpLowering,
    ClockGateOpLowering,
    ClockInvOpLowering,
    ConstantTimeOpLowering,
    CurrentTimeOpLowering,
    GetNextWakeupOpLowering,
    IntToTimeOpLowering,
    MemoryReadOpLowering,
    MemoryWriteOpLowering,
    ModelOpLowering,
    ReplaceOpWithInputPattern<seq::ToClockOp>,
    ReplaceOpWithInputPattern<seq::FromClockOp>,
    RuntimeModelOpLowering,
    SeqConstClockLowering,
    SetNextWakeupOpLowering,
    SimGetNextWakeupOpLowering,
    SimGetTimeOpLowering,
    SimSetTimeOpLowering,
    StateReadOpLowering,
    StateWriteOpLowering,
    StorageGetOpLowering,
    TerminateOpLowering,
    TimeToIntOpLowering,
    ZeroCountOpLowering,
    ArrayRefCreateOpLowering,
    ArrayRefAllocOpLowering,
    ArrayRefGetOpLowering,
    ArrayRefInjectOpLowering,
    ArrayRefSliceOpLowering,
    ArrayRefCopyOpLowering,
    ArrayRefToLLVMArrayOpLowering,
    ArrayRefToArrayOpLowering,
    ArrayRefFromArrayOpLowering
  >(converter, &getContext());
  // clang-format on
  patterns.add<ExecuteOp>(convert);
 
  StringCache stringCache;
  patterns.add<SimEmitValueOpLowering, SimPrintFormattedProcOpLowering>(
      converter, &getContext(), stringCache);
 
  auto &modelInfo = getAnalysis<ModelInfoAnalysis>();
  llvm::DenseMap<StringRef, ModelInfoMap> modelMap(modelInfo.infoMap.size());
  for (auto &[_, modelInfo] : modelInfo.infoMap) {
    llvm::DenseMap<StringRef, StateInfo> states(modelInfo.states.size());
    for (StateInfo &stateInfo : modelInfo.states)
      states.insert({stateInfo.name, stateInfo});
    modelMap.insert(
        {modelInfo.name,
         ModelInfoMap{modelInfo.numStateBytes, std::move(states),
                      modelInfo.initialFnSym, modelInfo.finalFnSym}});
  }
 
  patterns.add<SimInstantiateOpLowering, SimSetInputOpLowering,
               SimGetPortOpLowering, SimStepOpLowering>(
      converter, &getContext(), modelMap);
 
  // Apply the conversion.
  ConversionConfig config;
  config.allowPatternRollback = false;
  if (failed(applyFullConversion(getOperation(), target, std::move(patterns),
                                 config)))
    signalPassFailure();
}
 
std::unique_ptr<OperationPass<ModuleOp>> circt::createLowerArcToLLVMPass() {
  return std::make_unique<LowerArcToLLVMPass>();
}