src/ispc.h - main - redgpu/ispc

src/ispc.h
1 032 строки34 KB
Jérôme Richard
Make fast-math optimizations more aggressive (#3489)
18 мар 2026, 07:20
18 мар 2026, 07:201921a43
/*
  Copyright (c) 2010-2026, Intel Corporation

  SPDX-License-Identifier: BSD-3-Clause
*/

/** @file ispc.h
    @brief Main ispc.header file. Defines Target, Globals and Opt classes.
*/

#pragma once

#include "ispc_version.h"
#include "target_capabilities.h"
#include "target_enums.h"
#include "target_registry.h"

#if defined(_WIN32) || defined(_WIN64)
#define ISPC_HOST_IS_WINDOWS
#elif defined(__linux__)
#define ISPC_HOST_IS_LINUX
#elif defined(__FreeBSD__)
#define ISPC_HOST_IS_FREEBSD
#elif defined(__APPLE__)
#define ISPC_HOST_IS_APPLE
#endif

#include <bitset>
#include <map>
#include <memory>
#include <set>
#include <stdint.h>
#include <stdio.h>
#include <stdlib.h>
#include <string>
#include <vector>

#include <llvm/ADT/APFloat.h>
#include <llvm/ADT/StringRef.h>
#include <llvm/Support/VersionTuple.h>

/** @def ISPC_MAX_NVEC maximum vector size of any of the compliation
    targets.
 */
#define ISPC_MAX_NVEC 64

// Number of first construction phase
#define PRE_OPT_NUMBER 0

// Number of initial optimization phase
#define INIT_OPT_NUMBER 10

// Number of final optimization phase
#define LAST_OPT_NUMBER 1000

// Forward declarations of a number of widely-used LLVM types
namespace llvm {

class AttrBuilder;
class BasicBlock;
class Constant;
class ConstantValue;
class DataLayout;
class DIBuilder;
class Function;
class FunctionType;
class LLVMContext;
class Module;
class Target;
class TargetMachine;
class Type;
class Value;
class DIFile;
class DINamespace;
class DIType;

class DIScope;
} // namespace llvm

namespace ispc {

class ArrayType;
class AST;
class ASTNode;
class AtomicType;
class FunctionEmitContext;
class Expr;
class ExprList;
class Function;
class FunctionTemplate;
class FunctionType;
class Module;
class PointerType;
class Stmt;
class Symbol;
class SymbolTable;
class TemplateArg;
class TemplateInstantiation;
class TemplateParam;
class TemplateParms;
class TemplateSymbol;
class Type;
struct VariableDeclaration;
typedef std::vector<TemplateArg> TemplateArgs;

class StorageClass {
  public:
    enum class Kind {
        NONE,
        EXT,      // extern
        STATIC,   // static
        TYPEDEF,  // typedef
        EXT_C,    // extern "C"
        EXT_SYCL, // extern "SYCL"
    };

    // The constructor, operators, and constants are required to use
    // StorageClass in the same way as an enum class (which StorageClass was
    // previously) in switch statements, comparisons, and so on.
    StorageClass(Kind kind);
    operator Kind() const;
    static constexpr Kind NONE{Kind::NONE};
    static constexpr Kind EXT{Kind::EXT};
    static constexpr Kind STATIC{Kind::STATIC};
    static constexpr Kind TYPEDEF{Kind::TYPEDEF};
    static constexpr Kind EXT_C{Kind::EXT_C};
    static constexpr Kind EXT_SYCL{Kind::EXT_SYCL};

    std::string GetString() const;
    bool IsNone() const;
    bool IsExtern() const;
    bool IsStatic() const;
    bool IsTypedef() const;
    bool IsExternC() const;
    bool IsExternSYCL() const;
    bool IsAnyExtern() const;

  private:
    Kind m_kind = Kind::NONE;
};

// Stack smash protection
enum class SSPKind { SSPNone, SSPOn, SSPStrong, SSPReq };

// Enumerant for address spaces.
enum class AddressSpace {
    ispc_default,  // 0 = ispc_private
    ispc_global,   // 1
    ispc_constant, // 2
    ispc_local,    // 3
    ispc_generic,  // 4
};

namespace dispatch {
// This would create unnecessary unused copies of functions defined in isa.h
// in each translation unit that includes the current header (ispc.h). However,
// all unused ones will be removed by the compiler because they are static.
// We need to include this header here because we use the `dispatch::ISA` enum
// later in the current header.
#include "isa.h"
} // namespace dispatch

/** @brief Representation of a range of positions in a source file.

    This class represents a range of characters in a source file
    (e.g. those that span a token's definition), from starting line and
    column to ending line and column.  (These values are tracked by the
    lexing code).  Both lines and columns are counted starting from one.
 */
struct SourcePos {
    SourcePos(const char *n = nullptr, int fl = 0, int fc = 0, int ll = 0, int lc = 0);

    const char *name;
    int first_line;
    int first_column;
    int last_line;
    int last_column;

    /** Prints the filename and line/column range to standard output. */
    void Print() const;

    /** Returns a LLVM DIFile object that represents the SourcePos's file */
    llvm::DIFile *GetDIFile() const;

    /** Returns a LLVM DINamespace object that represents 'ispc' namespace. */
    llvm::DINamespace *GetDINamespace() const;

    bool operator==(const SourcePos &p2) const;
};

/** Returns a SourcePos that encompasses the extent of both of the given
    extents. */
SourcePos Union(const SourcePos &p1, const SourcePos &p2);

typedef unsigned int PerfWarningTypeUnderlyingType;

/** An enum to represent different types of perfarmance warnings that should be triggered for specific target */
enum class PerfWarningType : PerfWarningTypeUnderlyingType {
    // x86, SSE2/SSE4/AVX/AVX2.
    // Converts between [float|double] and uint[32|64] types (both directions) are much more expensive than similar
    // converts involving signed integers.
    CVTUIntFloat = 0x1,
    // x86, AVX2.
    // Converts between float16 and uint[32|64] types (both directions) are much more expensive than similar
    // converts involving signed integers.
    // SSE2/SSE4/AVX does not warn about FP16, as it is extreamly slow due to emulation of any FP16 ops.
    CVTUIntFloat16 = 0x2,
    // x86
    // Varying integer division and modulo operations are not supported in hardware and are scalarized.
    DIVModInt = 0x4,
    // x86, SSE2, SSE4.
    // Shift right by variable amount.
    VariableShiftRight = 0x8,
};

/** Code model */
enum class MCModel {
    Default, /** default model - i.e. not specified on the command line */
    Small,   /** small model */
    Large,   /** large model */
};

/** PIC level. It corresponds to llvm::PICLevel. */
enum class PICLevel {
    Default,  /** default model - i.e. not specified on the command line */
    NotPIC,   /** not PIC */
    SmallPIC, /** small PIC */
    BigPIC,   /** big PIC */
};

/** @brief Structure that defines a compilation target

    This structure defines a compilation target for the ispc compiler.
*/
class Target {
  public:
    /** Enumerator giving the instruction sets that the compiler can
        target.  These should be ordered from "worse" to "better" in that
        if a processor supports multiple target ISAs, then the most
        flexible/performant of them will appear last in the enumeration.  Note
        also that __best_available_isa() needs to be updated if ISAs are
        added or the enumerant values are reordered.  */
    enum ISA {
        INVALID = dispatch::ISA::INVALID,
        SSE2 = dispatch::ISA::SSE2,
        SSE41 = dispatch::ISA::SSE41,
        SSE42 = dispatch::ISA::SSE42,
        AVX = dispatch::ISA::AVX,
        // AVX11 is not supported anymore. Use either AVX or AVX2.
        AVX11 = dispatch::ISA::AVX11,
        AVX2 = dispatch::ISA::AVX2,
        AVX2VNNI = dispatch::ISA::AVX2VNNI,
        // KNL is not supported anymore.
        KNL_AVX512 = dispatch::ISA::KNL_AVX512,
        SKX_AVX512 = dispatch::ISA::SKX_AVX512,
        ICL_AVX512 = dispatch::ISA::ICL_AVX512,
        SPR_AVX512 = dispatch::ISA::SPR_AVX512,
        GNR_AVX512 = dispatch::ISA::GNR_AVX512,
        DMR_AVX10_2 = dispatch::ISA::DMR_AVX10_2,
#ifdef ISPC_ARM_ENABLED
        NEON,
#endif
#ifdef ISPC_PPC64_ENABLED
        VSX,
#endif
#ifdef ISPC_RISCV_ENABLED
        RV64GCV,
#endif
#ifdef ISPC_WASM_ENABLED
        WASM,
#endif
#ifdef ISPC_XE_ENABLED
        XELP,
        XEHPG,
        XEHPC,
        XELPG,
        XE2HPG,
        XE2LPG,
#endif
        NUM_ISAS
    };

#ifdef ISPC_XE_ENABLED
    enum class XePlatform {
        xe_lp,
        xe_hpg,
        xe_lpg,
        xe2_hpg,
        xe2_lpg,
        xe_hpc,
    };
#endif

    /** Initializes the given Target pointer for a target of the given
        name, if the name is a known target.  Returns true if the
        target was initialized and false if the name is unknown. */
    Target(Arch arch, const char *cpu, ISPCTarget isa, PICLevel picLevel, MCModel code_model, bool printTarget);

    ~Target();

    static std::unique_ptr<Target> Create(Arch arch, const char *cpu, ISPCTarget target, PICLevel picLevel,
                                          MCModel codeModel, bool printTarget);

    // We don't copy Target objects at the moment. If we will then proper
    // implementations are needed considering the ownership of heap-allocated
    // fields like m_dataLayout.
    Target(const Target &) = delete;
    Target &operator=(const Target &) = delete;

    /** Check if LLVM intrinsic is supported for the current target. */
    bool checkIntrinsticSupport(llvm::StringRef name, SourcePos pos);

    /** Returns a comma-delimited string giving the names of the currently
        supported CPUs. */
    static std::string SupportedCPUs();

    /** Returns a triple specifying the target architecture, vendor,
        and environment. */
    llvm::Triple GetTriple() const;

    /** Returns the LLVM TargetMachine object corresponding to this
        target. */
    llvm::TargetMachine *GetTargetMachine() const { return m_targetMachine; }

    /** Convert ISA enum to string */
    static const char *ISAToString(Target::ISA isa);

    /** Returns a string like "avx" encoding the target. Good for mangling. */
    const char *GetISAString() const;

    /** Convert ISA enum to string */
    static const char *ISAToTargetString(Target::ISA isa);

    /** Returns a string like "avx2-i32x8" encoding the target.
        This may be used for Target initialization. */
    const char *GetISATargetString() const;

    /** Return the suffix to use for target-specific functions.  */
    std::string GetTargetSuffix();

    /** Returns the size of the given type */
    llvm::Value *SizeOf(llvm::Type *type, llvm::BasicBlock *insertAtEnd);

    /** Given a structure type and an element number in the structure,
        returns a value corresponding to the number of bytes from the start
        of the structure where the element is located. */
    llvm::Value *StructOffset(llvm::Type *type, int element, llvm::BasicBlock *insertAtEnd);

    /** Update function name with __regcall3_ prefix. */
    void markFuncNameWithRegCallPrefix(std::string &funcName) const;

    /** Mark LLVM function with target specific attribute, if required. */
    void markFuncWithTargetAttr(llvm::Function *func);

    /** Set LLVM function with Calling Convention.
        The usage of this function is deprecated. The preferred way is to use
        llvm::Function::setCallingConv(llvm::CallingConv) and FunctionType::GetCallingConv())
    */
    void markFuncWithCallingConv(llvm::Function *func);

    const llvm::Target *getTarget() const { return m_target; }

    // Note the same name of method for 3.1 and 3.2+, this allows
    // to reduce number ifdefs on client side.
    const llvm::DataLayout *getDataLayout() const { return m_dataLayout; }

    /** Reports if Target object has valid state. */
    bool isValid() const { return m_valid; }

    ISPCTarget getISPCTarget() const { return m_ispc_target; }

    ISA getISA() const { return m_isa; }

    /** Converts an ISPC target to the corresponding Target ISA. */
    static ISA TargetToISA(ISPCTarget target);

    bool isXeTarget() {
#ifdef ISPC_XE_ENABLED
        return m_isa == Target::XELP || m_isa == Target::XEHPG || m_isa == Target::XEHPC || m_isa == Target::XELPG ||
               m_isa == Target::XE2HPG || m_isa == Target::XE2LPG;
#else
        return false;
#endif
    }

#ifdef ISPC_XE_ENABLED
    XePlatform getXePlatform() const;
    uint32_t getXeGrfSize() const;
    bool hasXePrefetch() const;
#endif

    Arch getArch() const { return m_arch; }

    bool is32Bit() const { return m_is32Bit; }

    std::string getCPU() const { return m_cpu; }

    int getNativeVectorWidth() const { return m_nativeVectorWidth; }

    int getNativeVectorAlignment() const { return m_nativeVectorAlignment; }

    int getDataTypeWidth() const { return m_dataTypeWidth; }

    int getVectorWidth() const { return m_vectorWidth; }

    PICLevel getPICLevel() const { return m_picLevel; }

    MCModel getMCModel() const { return m_codeModel; }

    bool getMaskingIsFree() const { return m_maskingIsFree; }

    int getMaskBitCount() const { return m_maskBitCount; }

    bool hasGather() const { return m_hasGather; }

    bool useGather() const;

    bool hasScatter() const { return m_hasScatter; }

    bool useScatter() const;

    bool hasVecPrefetch() const { return m_hasVecPrefetch; }

    /** Check if target has a specific capability */
    bool hasCapability(TargetCapability cap) const { return m_capabilities[static_cast<size_t>(cap)]; }

    /** Set a specific capability (default is to enable it) */
    void setCapability(TargetCapability cap, bool value = true) { m_capabilities[static_cast<size_t>(cap)] = value; }

    /** Set multiple capabilities at once */
    void setCapabilities(std::initializer_list<TargetCapability> caps) {
        for (auto cap : caps) {
            setCapability(cap);
        }
    }

    void setWarning(PerfWarningType warningType) { m_warnings |= static_cast<unsigned int>(warningType); }

    bool shouldWarn(PerfWarningType warningType) { return (m_warnings & static_cast<unsigned int>(warningType)) != 0; }

  private:
    /** llvm Target object representing this target. */
    const llvm::Target *m_target;

    /** llvm TargetMachine. Deconstrcted in ~Target. */
    llvm::TargetMachine *m_targetMachine;

    /** This is deconstructed in ~Target. */
    llvm::DataLayout *m_dataLayout;

    /** flag to report invalid state after construction
        (due to bad parameters passed to constructor). */
    bool m_valid;

    /** ISPC target being used */
    ISPCTarget m_ispc_target;

    /** Instruction set being compiled to. */
    ISA m_isa;

    /** Target system architecture.  (e.g. "x86-64", "x86"). */
    Arch m_arch;

    /** Is the target architecture 32 or 64 bit */
    bool m_is32Bit;

    /** Target CPU. (e.g. "corei7", "corei7-avx", ..) */
    std::string m_cpu;

    /** Target-specific attribute string to pass along to the LLVM backend */
    std::string m_attributes;

    /** Target-specific function attributes */
    std::vector<std::pair<std::string, std::string>> m_funcAttributes;

    /** Target-specific LLVM attribute, which has to be attached to every
        function to ensure that it is generated for correct target architecture.
        This is requirement was introduced in LLVM 3.3 */
    llvm::AttrBuilder *m_tf_attributes;

    /** Native vector width of the vector instruction set.  Note that this
        value is directly derived from the ISA being used (e.g. it's 4 for
        SSE, 8 for AVX, etc.) */
    int m_nativeVectorWidth;

    /** Native vector alignment in bytes. Theoretically this may be derived
        from the vector size, but it's better to manage directly the alignement.
        It allows easier experimenting and better fine tuning for particular
        platform. This information is primatily used when
        --opt=force-aligned-memory is used. */
    int m_nativeVectorAlignment;

    /** Data type width in bits. Typically it's 32, but could be 8, 16 or 64. */
    int m_dataTypeWidth;

    /** Actual vector width currently being compiled to.  This may be an
        integer multiple of the native vector width, for example if we're
        "doubling up" and compiling 8-wide on a 4-wide SSE system. */
    int m_vectorWidth;

    /** Indicates whether position independent code should be generated. */
    PICLevel m_picLevel;

    /** Code model */
    MCModel m_codeModel;

    /** Is there overhead associated with masking on the target
        architecture; e.g. there is on SSE, due to extra blends and the
        like, but there isn't with an ISA that supports masking
        natively. */
    bool m_maskingIsFree;

    /** How many bits are used to store each element of the mask: e.g. this
        is 32 on SSE/AVX, since that matches the HW better. */
    int m_maskBitCount;

    /** Bitset storing target capabilities for efficient querying.
     *  Use hasCapability()/setCapability() to access/modify capabilities.
     */
    std::bitset<static_cast<size_t>(TargetCapability::COUNT)> m_capabilities;

    /** Indicates whether the target has a native gather instruction */
    bool m_hasGather;

    /** Indicates whether the target has a native scatter instruction */
    bool m_hasScatter;

    /** Indicates whether the target has hardware instruction for vector prefetch. */
    bool m_hasVecPrefetch;

    /** A bitset of PerfWarningType values indicating the warnings that are relevant for the target. */
    PerfWarningTypeUnderlyingType m_warnings;
};

/** @brief Structure that collects optimization options

    This structure collects all of the options related to optimization of
    generated code.
*/
struct Opt {
    Opt();

    /** Optimization level.  Currently, the only valid values are 0,
        indicating essentially no optimization, 1, indicating optimization for size,
        and 2, indicating as much optimization as possible. */
    int level;

    /** Indicates whether "fast and loose" numerically unsafe optimizations
        should be performed.  This is None by default. */
    enum class FastMathMode { None, Legacy, Balanced, Aggressive };
    FastMathMode fastMath;

    /** Indicates whether an vector load should be issued for masked loads
        on platforms that don't have a native masked vector load.  (This may
        lead to accessing memory up to programCount-1 elements past the end of
        arrays, so is unsafe in general.) */
    bool fastMaskedVload;

    /** Indicates when loops should be unrolled (when doing so seems like
        it will make sense. */
    bool unrollLoops;

    /** Indicates if addressing math will be done with 32-bit math, even on
        64-bit systems.  (This is generally noticably more efficient,
        though at the cost of addressing >2GB).
     */
    bool force32BitAddressing;

    /** Indicates whether Assert() statements should be ignored (for
        performance in the generated code). */
    bool disableAsserts;

    /** Indicates whether gathers should be disabled for the targets that support them (for
        performance in the generated code). */
    bool disableGathers;

    /** Indicates whether scatters should be disabled for the targets that support them (for
        performance in the generated code). */
    bool disableScatters;

    /** Indicates whether FMA instructions should be disabled (on targets
        that support them). */
    bool disableFMA;

    /** Always generate aligned vector load/store instructions; this
        implies a guarantee that all dynamic access through pointers that
        becomes a vector load/store will be a cache-aligned sequence of
        locations. */
    bool forceAlignedMemory;

    /** If enabled, enables LoadStoreVectorizerPass in ISPC optimization pipeline. */
    bool enableLoadStoreVectorizer;

    /** If enabled, enables SLPVectorizerPass in ISPC optimization pipeline. */
    bool enableSLPVectorizer;

    /** If enabled, disables the various optimizations that kick in when
        the execution mask can be determined to be "all on" at compile
        time. */
    bool disableMaskAllOnOptimizations;

    /** If enabled, the various __pseudo* memory ops (gather/scatter,
        masked load/store) are left in their __pseudo* form, for better
        understanding of the structure of generated code when reading
        it. */
    bool disableHandlePseudoMemoryOps;

    /** On targets that don't have a masked store instruction but do have a
        blending instruction, by default, we simulate masked stores by
        loading the old value, blending, and storing the result.  This can
        potentially be unsafe in multi-threaded code, in that it writes to
        locations that aren't supposed to be written to.  Setting this
        value to true disables this work-around, and instead implements
        masked stores by 'scalarizing' them, so that we iterate over the
        ISIMD lanes and do a scalar write for the ones that are running. */
    bool disableBlendedMaskedStores;

    /** Disables the 'coherent control flow' constructs in the
        language. (e.g. this causes "cif" statements to be demoted to "if"
        statements.)  This is likely only useful for measuring the impact
        of coherent control flow. */
    bool disableCoherentControlFlow;

    /** Disables uniform control flow optimizations (e.g. this changes an
        "if" statement with a uniform condition to have a varying
        condition).  This is likely only useful for measuring the impact of
        uniform control flow. */
    bool disableUniformControlFlow;

    /** Disables the backend optimizations related to gather/scatter
        (e.g. transforming gather from sequential locations to an unaligned
        load, etc.)  This is likely only useful for measuring the impact of
        these optimizations. */
    bool disableGatherScatterOptimizations;

    /** Disables the optimization that demotes masked stores to regular
        stores when the store is happening at the same control flow level
        where the variable was declared.  This is likely only useful for
        measuring the impact of this optimization. */
    bool disableMaskedStoreToStore;

    /** Disables the optimization that detects when the execution mask is
        all on and emits code for gathers and scatters that doesn't loop
        over the SIMD lanes but just does the scalar loads and stores
        directly. */
    bool disableGatherScatterFlattening;

    /** Disables the optimizations that detect when arrays are being
        indexed with 'uniform' values and issue scalar loads/stores rather
        than gathers/scatters.  This is likely only useful for measuring
        the impact of this optimization. */
    bool disableUniformMemoryOptimizations;

    /** Disables optimizations that coalesce incoherent scalar memory
        access from gathers into wider vector operations, when possible. */
    bool disableCoalescing;

    /** Disable using zmm registers for avx512 target in favour of ymm.
        Affects only >= 512 bit wide targets and only if avx512vl is available */
    bool disableZMM;

    /** Set FTZ/DAZ flags on the extern function entrance and restore them.
        upon return to "host" code.*/
    bool resetFTZ_DAZ;

#ifdef ISPC_XE_ENABLED
    /** Disables optimization that coalesce gathers on Xe. This is
        likely only useful for measuring the impact of this optimization */
    bool disableXeGatherCoalescing;

    /** Minimal difference between the number of eliminated loads and
        the number of newly created mem insts. Default value is zero:
        assuming that gather is more dufficult due to address calculations.
        the default value should be adjusted with some experiments. */
    int thresholdForXeGatherCoalescing;

    /** Enables experimental support of foreach statement inside varying CF.
        Current implementation brings performance degradation due to ineffective
        implementation of unmasked.*/
    bool enableForeachInsideVarying;

    /** Enables emitting of genx.any intrinsics and the control flow which is
        based on impliit hardware mask. Forces generation of goto/join instructions
        in assembly.*/
    bool emitXeHardwareMask;

    /** Enables generation of masked loads implemented using svm loads which
     * may lead to out of bound reads but bring prformance improvement in
     * most of the cases.
     */
    bool enableXeUnsafeMaskedLoad;
#endif
};

/** @brief This structure collects together a number of global variables.

    This structure collects a number of global variables that mostly
    represent parameter settings for this compilation run.  In particular,
    none of these values should change after compilation befins; their
    values are all set during command-line argument processing or very
    early during the compiler's execution, before any files are parsed.
  */
struct Globals {
    Globals();
    ~Globals();

    /** TargetRegistry holding all stdlib bitcode. */
    TargetLibRegistry *target_registry;

    /** Optimization option settings */
    Opt opt;

    /** Compilation target information */
    Target *target;

    /** Target OS */
    TargetOS target_os;

    /** Function Calling Convention */
    CallingConv calling_conv;

    /** There are a number of math libraries that can be used for
        transcendentals and the like during program compilation. */
    enum class MathLib { Math_ISPC, Math_ISPCFast, Math_SVML, Math_System };
    MathLib mathLib;

    /** Optimization level to be specified while creating TargetMachine. */
    enum class CodegenOptLevel { None, Default, Aggressive };
    CodegenOptLevel codegenOptLevel;

    /** Records whether the ispc standard library should be made available
        to the program during compilations. (Default is true.) */
    bool includeStdlib;

    /** Indicates whether the C pre-processor should be run over the
        program source before compiling it.  (Default is true.) */
    bool runCPP;

    /** Indicates whether the compiler places each function in its own
        section. */
    bool functionSections;

    /** When \c true, only runs the C pre-processor. (Default is false.) */
    bool onlyCPP;

    /** When \c true, suppresses errors from the C pre-processor.
        (Default is false.) */
    bool ignoreCPPErrors;

    /** When \c true, voluminous debugging output will be printed during
        ispc's execution. */
    bool debugPrint;

    /** When \c true, print verbose output from PassManager.
     * (Default is false.)
     */
    bool debugPM;

    /** When \c true, print time trace output from PassManager.
     * (Default is false.)
     */
    bool debugPMTimeTrace;

    /** When \c true, dump AST.
        None - don't dump AST
        All - dump AST for all the code
    */
    enum class ASTDumpKind { None, All };
    ASTDumpKind astDump;

    /** When \c true, target ISA will be printed during ispc's execution. */
    bool printTarget;

    /** When \c true, LLVM won't omit frame pointer. */
    bool NoOmitFramePointer;

    /** Indicates which stages of optimization we want to dump. */
    std::set<int> debug_stages;

    /** Whether to dump IR to file. */
    bool dumpFile;

    /** Whether we are in special mode of generating IR for stdlib. */
    bool genStdlib;

    /** Whether ISPC binary is slim. */
    bool isSlimBinary;

    /** Store the path to directory for IR file dumps. */
    std::string dumpFilePath;

    /** Store the absolute path to share/ispc directory. */
    std::string shareDirPath;

    /** Indicates after which optimization we want to generate
        DebugIR information. */
    int debugIR;

    /** Indicates which phases of optimization we want to switch off. */
    std::set<int> off_stages;

    /** Indicates whether all warning messages should be surpressed. */
    bool disableWarnings;

    /** Indicates whether warnings should be issued as errors. */
    bool warningsAsErrors;

    /** Preserve wrap-around on signed integer overflow by disabling the
        nsw attribute when emitting arithmetic for signed integer
        expressions.  Without this disabled, the compiler may rely on UB
        behavior for optimizations on signed integer types.  */
    bool wrapSignedInt;

    /** Indicates whether line wrapping of error messages to the terminal
        width should be disabled. */
    bool disableLineWrap;

    /** Indicates whether additional warnings should be issued about
        possible performance pitfalls. */
    bool emitPerfWarnings;

    /** Indicates whether all printed output should be surpressed. */
    bool quiet;

    /** Always use ANSI escape sequences to colorize warning and error
        messages, even if piping output to a file, etc. */
    bool forceColoredOutput;

    /** Indicates whether calls should be emitted in the program to an
        externally-defined program instrumentation function. (See the
        "Instrumenting your ispc programs" section in the user's
        manual.) */
    bool emitInstrumentation;

#ifdef ISPC_XE_ENABLED
    /** Arguments to pass to Vector Compiler backend for offline
    compilation to L0 binary */
    std::string vcOpts;

    /* Stateless stack memory size in VC backend */
    unsigned int stackMemSize;
#endif

    bool noPragmaOnce;

    /** Indicates whether ispc should generate debugging symbols for the
        program in its output. */
    bool generateDebuggingSymbols;

    /** Debug info type to generate. */
    enum class DebugInfoType { None = 0, DWARF, CodeView };
    DebugInfoType debugInfoType;

    /** Require generation of DWARF of certain version (2, 3, 4, 5). For
        default version, this field is set to 5. */
    // Hint: to verify dwarf version in the object file, run on Linux:
    // readelf --debug-dump=info object.o | grep -A 2 'Compilation Unit @'
    // on Mac:
    // xcrun dwarfdump -r0 object.o
    int generateDWARFVersion;

    /** Enable sample-based profiling debug info generation. */
    bool sampleProfilingDebugInfo;

    /** Path to sample profile data file for profile-guided optimization.
        When provided, enables sample-based profile-guided optimization using
        the specified profile data to guide optimization decisions. */
    std::string profileSampleUse;

    /** If true, function names are mangled by appending the target ISA and
        vector width to them. */
    bool mangleFunctionsWithTarget;

    /** Indicates whether to generate internal (ISPC-callable) versions of
        exported functions. When true (default), both internal and external
        versions are generated. When false, only external (C/C++-callable)
        versions are generated, matching the behavior of the external_only
        attribute. */
    bool generateInternalExportFunctions;

    /* If enabled, allows the user to directly call LLVM intrinsics. */
    bool enableLLVMIntrinsics;

    /** Global LLVMContext object */
    llvm::LLVMContext *ctx;

    /** Current working directory when the ispc compiler starts
        execution. */
    char currentDirectory[1024];

    /** Arguments to pass along to the C pre-processor, if it is run on the
        program before compilation. */
    std::vector<std::string> cppArgs;

    /** Additional user-provided directories to search when processing
        #include directives in the preprocessor. */
    std::vector<std::string> includePath;

    /** Indicates that alignment in memory allocation routines should be
        forced to have given value. -1 value means natural alignment for the platforms. */
    int forceAlignment;

    /** When true, flag non-static functions with dllexport attribute on Windows. */
    bool dllExport;

    /** Lines for which warnings are turned off. */
    std::map<std::pair<int, std::string>, bool> turnOffWarnings;

    enum pragmaUnrollType { none, nounroll, unroll, count };

    /** Preprocessor Output Types, process with -E/-dD/-dM options */
    enum class PreprocessorOutputType { Cpp, WithMacros, MacrosOnly };

    PreprocessorOutputType preprocessorOutputType = PreprocessorOutputType::Cpp;

    /* If true, we are compiling for more than one target. */
    bool isMultiTargetCompilation;

    /* Number of errors to show in ISPC. */
    int errorLimit;

    /* When true, enable compile time tracing. */
    bool enableTimeTrace;

    /* If true, disable validation of target support in Target constructor. */
    bool disableTargetValidation;

    /* When compile time tracing is enabled, set time granularity. */
    int timeTraceGranularity;

    /* Set macOS/iOS deployment target. The version will be propagated to the triple.
       This address the new linker introduced in Xcode 15 that issues a warning if version when no version is provided.
       https://github.com/ispc/ispc/issues/3143  */
    llvm::VersionTuple darwinVersionMin;

    /** When true, include float16 conversion functions permanently to the compiled module */
    bool includeFloat16Conversions;

    /* Stores the stack smash protection setting. */
    SSPKind SSPLevel;
};

// This is used when empty string is used for "--darwin-version-min"
constexpr llvm::VersionTuple darwinUnspecifiedVersion(INT_MAX);

enum {
    COST_ASSIGN = 1,
    COST_COMPLEX_ARITH_OP = 4,
    COST_DELETE = 32,
    COST_DEREF = 4,
    COST_FUNCALL = 4,
    COST_FUNPTR_UNIFORM = 12,
    COST_FUNPTR_VARYING = 24,
    COST_GATHER = 8,
    COST_GOTO = 4,
    COST_LOAD = 2,
    COST_NEW = 32,
    COST_BREAK_CONTINUE = 3,
    COST_RETURN = 4,
    COST_SELECT = 4,
    COST_SIMPLE_ARITH_LOGIC_OP = 1,
    COST_SYNC = 32,
    COST_TASK_LAUNCH = 32,
    COST_INVOKE = 32,
    COST_TYPECAST_COMPLEX = 4,
    COST_TYPECAST_SIMPLE = 1,
    COST_UNIFORM_IF = 2,
    COST_VARYING_IF = 3,
    COST_UNIFORM_LOOP = 4,
    COST_VARYING_LOOP = 6,
    COST_UNIFORM_SWITCH = 4,
    COST_VARYING_SWITCH = 12,
    COST_ASSERT = 8,

    CHECK_MASK_AT_FUNCTION_START_COST = 16,
    PREDICATE_SAFE_IF_STATEMENT_COST = 6,
    // For Xe target we want to avoid branches as much as possible
    // so we use increased cost here
    PREDICATE_SAFE_SHORT_CIRC_XE_STATEMENT_COST = 10,
};

extern Globals *g;
extern Module *m;

// Singleton object for bookkeeping heap objects to destroy them later to
// avoid memory leak.
class BookKeeper {
  private:
    BookKeeper() {}

    template <typename T> std::vector<T *> &getStorage() {
        // Vector to store bookkeeped objects.
        static std::vector<T *> v;
        return v;
    }

    template <typename T> void freeOne() {
        std::vector<T *> &v = getStorage<T>();
        for (auto e : v)
            delete e;
        v.clear();
    }

  public:
    static BookKeeper &in();
    BookKeeper(BookKeeper const &) = delete;
    void operator=(BookKeeper const &) = delete;

    template <typename T> void *add(T *p) {
        std::vector<T *> &v = getStorage<T>();
        v.push_back(p);
        return p;
    }

    // Free all bookkeeped objects.
    void freeAll();
};

// Base class to inherit for objects needed to be bookkeeped.
class Traceable {
  public:
    void *operator new(size_t size) { return BookKeeper::in().add(static_cast<Traceable *>(::operator new(size))); }
    virtual ~Traceable() = default;
};

// An enum class enumerating wrap semantic settings for use with BinaryOperator and other
// signed arithmetic IR emitters in cases of signed overflow
enum class WrapSemantics {
    NSW = 0, // Do not preserve wraparound behavior
    None = 1 // Preserve wraparound behavior
};
} // namespace ispc