dynarmic/README.md
Merry a37f3673f8 Squashed 'externals/oaknut/' changes from d0488d932..9d091109d
9d091109d oaknut: 2.0.1
7f3e9f600 oaknut: Support single argument constructor for CodeGenerator again

git-subtree-dir: externals/oaknut
git-subtree-split: 9d091109deb445bc6e9289c6195a282b7c993d49
2024-01-28 17:02:37 +00:00

9 KiB

Oaknut

A C++20 assembler for AArch64 (ARMv8.0 to ARMv8.2)

Oaknut is a header-only library that allows one to dynamically assemble code in-memory at runtime.

Usage

Give oaknut::CodeGenerator a pointer to a block of memory. Call functions on it to emit code.

Simple example:

#include <cstdio>
#include <oaknut/code_block.hpp>
#include <oaknut/oaknut.hpp>

using EmittedFunction = int (*)();

EmittedFunction EmitExample(oaknut::CodeGenerator& code, int value)
{
    using namespace oaknut::util;

    EmittedFunction result = code.xptr<EmittedFunction>();

    code.MOV(W0, value);
    code.RET();

    return result;
}

int main()
{
    oaknut::CodeBlock mem{4096};
    oaknut::CodeGenerator code{mem.ptr()};

    mem.unprotect();

    EmittedFunction fn = EmitExample(code, 42);

    mem.protect();
    mem.invalidate_all();

    std::printf("%i\n", fn());  // Output: 42

    return 0;
}

CodeGenerator also has a constructor taking two pointers. The first pointer is the memory address to write to, and the second pointer is the memory address that the code will be executing from. This allows you to write to a buffer before copying to the final destination for execution, or to have to use dual-mapped memory blocks to avoid memory protection overhead.

Below is an example of using the oaknut-provided utility header for dual-mapped memory blocks:

#include <cstdio>
#include <oaknut/dual_code_block.hpp>
#include <oaknut/oaknut.hpp>

using EmittedFunction = ;

int main()
{
    using namespace oaknut::util;

    oaknut::DualCodeBlock mem{4096};
    oaknut::CodeGenerator code{mem.wptr(), mem.xptr()};

    const auto result = code.xptr<int (*)()>();

    code.MOV(W0, value);
    code.RET();

    mem.invalidate_all();

    std::printf("%i\n", fn());  // Output: 42

    return 0;
}

Emit to std::vector

If you wish to merely emit code into memory without executing it, or if you are developing a cross-compiler that is not running on an ARM64 device, you can use oaknut::VectorCodeGenerator instead.

Provide oaknut::VectorCodeGenerator with a reference to a std::vector<std::uint32_t> and it will append to that vector.

The second pointer argument represents the destination address the code will eventually be executed from.

Simple example:

#include <cstdint>
#include <cstdio>
#include <oaknut/oaknut.hpp>
#include <vector>

int main()
{
    std::vector<std::uint32_t> vec;
    oaknut::VectorCodeGenerator code{vec, (uint32_t*)0x1000};

    code.MOV(W0, 42);
    code.RET();

    std::printf("%08x %08x\n", vec[0], vec[1]);  // Output: d2800540 d65f03c0

    return 0;
}

Headers

Header Compiles on non-ARM64 Contents
<oaknut/oaknut.hpp> Yes Provides CodeGenerator and VectorCodeGenerator for code emission, as well as the oaknut::util namespace.
<oaknut/code_block.hpp> No Utility header that provides CodeBlock, allocates, alters permissions of, and invalidates executable memory.
<oaknut/dual_code_block.hpp> No Utility header that provides DualCodeBlock, which allocates two mirrored memory blocks (with RW and RX permissions respectively).
<oaknut/oaknut_exception.hpp> Yes Provides OaknutException which is thrown on an error.
<oaknut/feature_detection/cpu_feature.hpp> Yes Utility header that provides CpuFeatures which can be used to describe AArch64 features.
<oaknut/feature_detection/feature_detection.hpp> No Utility header that provides detect_features and read_id_registers for determining available AArch64 features.

Instructions

Each AArch64 instruction corresponds to one emitter function. For a list of emitter functions see:

Operands

The oaknut::util namespace provides convenient names for operands for instructions. For example:

Name Class
W0, W1, ..., W30 WReg 32-bit general purpose registers
X0, X1, ..., X30 XReg 64-bit general purpose registers
WZR WzrReg (convertable to WReg) 32-bit zero register
XZR ZrReg (convertable to XReg) 64-bit zero register
WSP WspReg (convertable to WRegSp) 32-bit stack pointer
SP SpReg (convertable to XRegSp) 64-bit stack pointer
B0, B1, ..., B31 BReg 8-bit scalar SIMD register
H0, H1, ..., H31 HReg 16-bit scalar SIMD register
S0, S1, ..., S31 SReg 32-bit scalar SIMD register
D0, D1, ..., D31 DReg 64-bit scalar SIMD register
Q0, Q1, ..., Q31 QReg 128-bit scalar SIMD register

For vector operations, you can specify registers like so:

Name Class
V0.B8(), ... VReg_8B 8 elements each 8 bits in size
V0.B16(), ... VReg_16B 16 elements each 8 bits in size
V0.H4(), ... VReg_4H 4 elements each 16 bits in size
V0.H8(), ... VReg_8H 8 elements each 16 bits in size
V0.S2(), ... VReg_2S 2 elements each 32 bits in size
V0.S4(), ... VReg_4S 4 elements each 32 bits in size
V0.D1(), ... VReg_1D 1 elements each 64 bits in size
V0.D2(), ... VReg_2D 2 elements each 64 bits in size

And you can specify elements like so:

Name Class
V0.B()[0] BElem 0th 8-bit element of V0 register
V0.H()[0] HElem 0th 16-bit element of V0 register
V0.S()[0] SElem 0th 32-bit element of V0 register
V0.D()[0] DElem 0th 64-bit element of V0 register

Register lists are specified using List:

List{V0.B16(), V1.B16(), V2.B16()}  // This expression has type List<VReg_16B, 3>

And lists of elements similarly (both forms are equivalent):

List{V0.B()[1], V1.B()[1], V2.B()[1]}  // This expression has type List<BElem, 3>
List{V0.B(), V1.B(), V2.B()}[1]        // This expression has type List<BElem, 3>

You can find examples of instruction use in tests/general.cpp and tests/fpsimd.cpp.

Feature Detection

CPU features

This library also includes utility headers for CPU feature detection.

One just needs to include <oaknut/feature_detection/feature_detection.hpp>, then call detect_features to get a bitset of features in a cross-platform manner.

CPU feature detection is operating system specific, and some operating systems even have multiple methods. Here are a list of supported operating systems and implemented methods:

Operating system Default Method
Linux / Android ELF hwcaps
Apple sysctlbyname
Windows IsProcessorFeaturePresent
FreeBSD ELF hwcaps
NetBSD machdep.cpu%d.cpu_id sysctl
OpenBSD CTL_MACHDEP.CPU_ID_* sysctl

There are alternative methods available for advanced users to specify specific methods to detect features if they wish. (See detect_features_via_*.)

Simple example:

#include <cstdio>
#include <oaknut/feature_detection/feature_detection.hpp>

int main() {
    oaknut::CpuFeatures feats = oaknut::detect_features();

    std::printf("CPU supports JSCVT: %i\n", feats.has(oaknut::CpuFeature::JSCVT));
}

ID registers

We also provide a crossplatform way for ID registers to be read:

OAKNUT_SUPPORTS_READING_ID_REGISTERS Available functionality
0 Reading ID registers is not supported on this operating system.
1 This operating system provides a system-wide set of ID registers, use read_id_registers().
2 Per-core ID registers, use get_core_count() and read_id_registers(int index).

All of the above operating systems with the exception of apple also support reading ID registers, and if one prefers one can do feature detection via detect_features_via_id_registers(*read_id_registers()).

Simple example:

#include <cstddef>
#include <cstdio>
#include <oaknut/feature_detection/feature_detection.hpp>

int main() {
#if OAKNUT_SUPPORTS_READING_ID_REGISTERS == 1

    oaknut::id::IdRegisters id = oaknut::read_id_registers();

    std::printf("ISAR0 register: %08x\n", id.isar0.value);

#elif OAKNUT_SUPPORTS_READING_ID_REGISTERS == 2

    oaknut::id::IdRegisters id = oaknut::read_id_registers(0);

    const std::size_t core_count = oaknut::get_core_count();
    for (std::size_t core_index = 0; core_index < core_count; core_index++) {
        std::printf("ISAR0 register (for core %zu): %08x\n", core_index, id.isar0.value);
    }

#else

    std::printf("Reading ID registers not supported\n");

#endif
}

License

This project is MIT licensed.