[ARM64] Initial commit for arm64 backend

Based entirely off existing xbyak x86 implementation and available
tests. Still needs a lot of optimization and testing on non-Windows
platforms.

So far passes all tests and boots at least some games on Windows.
This commit is contained in:
Herman S.
2026-03-21 01:17:18 +09:00
parent c6e112bcd8
commit 883c2030d0
44 changed files with 13428 additions and 140 deletions

View File

@@ -7,6 +7,27 @@ else()
add_compile_options(-w)
endif()
# ==============================================================================
# xbyak_aarch64 (ARM64 JIT assembler)
# ==============================================================================
if(XE_TARGET_AARCH64)
add_library(xbyak_aarch64 STATIC
xbyak_aarch64/src/xbyak_aarch64_impl.cpp
xbyak_aarch64/src/util_impl.cpp
)
target_include_directories(xbyak_aarch64 PUBLIC
${CMAKE_CURRENT_SOURCE_DIR}/xbyak_aarch64
${CMAKE_CURRENT_SOURCE_DIR}/xbyak_aarch64/src
${CMAKE_CURRENT_SOURCE_DIR}/xbyak_aarch64/xbyak_aarch64
)
if(MSVC)
target_compile_definitions(xbyak_aarch64 PRIVATE NOMINMAX)
target_compile_options(xbyak_aarch64 PRIVATE /w)
else()
target_compile_options(xbyak_aarch64 PRIVATE -w)
endif()
endif()
# ==============================================================================
# aes_128
# ==============================================================================
@@ -21,25 +42,37 @@ target_include_directories(aes_128 PUBLIC aes_128)
add_library(capstone STATIC
capstone/cs.c
capstone/MCInst.c
capstone/MCInstPrinter.c
capstone/MCInstrDesc.c
capstone/MCRegisterInfo.c
capstone/SStream.c
capstone/utils.c
capstone/Mapping.c
)
# Glob X86 arch files
file(GLOB _capstone_x86_c "capstone/arch/X86/*.c")
file(GLOB _capstone_x86_h "capstone/arch/X86/*.h")
file(GLOB _capstone_x86_inc "capstone/arch/X86/*.inc")
# Remove excluded files
list(FILTER _capstone_x86_c EXCLUDE REGEX "X86ATTInstPrinter\\.c$")
list(FILTER _capstone_x86_inc EXCLUDE REGEX "reduce\\.inc$")
target_sources(capstone PRIVATE ${_capstone_x86_c} ${_capstone_x86_h} ${_capstone_x86_inc})
target_compile_definitions(capstone PRIVATE
CAPSTONE_X86_ATT_DISABLE
CAPSTONE_HAS_X86
CAPSTONE_USE_SYS_DYN_MEM
)
# Architecture-specific capstone files
if(XE_TARGET_AARCH64)
file(GLOB _capstone_arch_c "capstone/arch/AArch64/*.c")
file(GLOB _capstone_arch_h "capstone/arch/AArch64/*.h")
file(GLOB _capstone_arch_inc "capstone/arch/AArch64/*.inc")
target_sources(capstone PRIVATE ${_capstone_arch_c} ${_capstone_arch_h} ${_capstone_arch_inc})
target_compile_definitions(capstone PRIVATE
CAPSTONE_HAS_ARM64
CAPSTONE_HAS_AARCH64
CAPSTONE_USE_SYS_DYN_MEM
)
else()
file(GLOB _capstone_x86_c "capstone/arch/X86/*.c")
file(GLOB _capstone_x86_h "capstone/arch/X86/*.h")
file(GLOB _capstone_x86_inc "capstone/arch/X86/*.inc")
list(FILTER _capstone_x86_c EXCLUDE REGEX "X86ATTInstPrinter\\.c$")
list(FILTER _capstone_x86_inc EXCLUDE REGEX "reduce\\.inc$")
target_sources(capstone PRIVATE ${_capstone_x86_c} ${_capstone_x86_h} ${_capstone_x86_inc})
target_compile_definitions(capstone PRIVATE
CAPSTONE_X86_ATT_DISABLE
CAPSTONE_HAS_X86
CAPSTONE_USE_SYS_DYN_MEM
)
endif()
target_include_directories(capstone PUBLIC capstone/include PRIVATE capstone)
# Force all capstone sources to compile as C
file(GLOB_RECURSE _capstone_all_c "capstone/*.c")
@@ -77,9 +110,9 @@ add_library(discord-rpc STATIC
discord-rpc/src/serialization.cpp
discord-rpc/src/serialization.h
)
if(CMAKE_SYSTEM_PROCESSOR MATCHES "x86_64|AMD64")
if(XE_TARGET_X86_64)
target_compile_definitions(discord-rpc PRIVATE RAPIDJSON_SSE42)
elseif(CMAKE_SYSTEM_PROCESSOR MATCHES "aarch64|arm64|ARM64")
elseif(XE_TARGET_ARM64)
target_compile_definitions(discord-rpc PRIVATE RAPIDJSON_NEON)
endif()
target_include_directories(discord-rpc PUBLIC discord-rpc/include PRIVATE rapidjson/include)
@@ -156,11 +189,16 @@ target_include_directories(pugixml PUBLIC pugixml/src)
# ==============================================================================
# Generate snappy-stubs-public.h if missing
if(NOT EXISTS "${CMAKE_CURRENT_SOURCE_DIR}/snappy/snappy-stubs-public.h")
if(XE_TARGET_X86_64)
set(_snappy_avx_flag "-DSNAPPY_REQUIRE_AVX=ON")
else()
set(_snappy_avx_flag "-DSNAPPY_REQUIRE_AVX=OFF")
endif()
execute_process(
COMMAND ${CMAKE_COMMAND}
-DSNAPPY_BUILD_TESTS=OFF
-DSNAPPY_BUILD_BENCHMARKS=OFF
-DSNAPPY_REQUIRE_AVX=ON
${_snappy_avx_flag}
${CMAKE_CURRENT_SOURCE_DIR}/snappy
-B${CMAKE_CURRENT_SOURCE_DIR}/snappy
WORKING_DIRECTORY ${CMAKE_CURRENT_SOURCE_DIR}
@@ -244,44 +282,56 @@ if(_zlibng_needs_configure)
)
endif()
file(GLOB _zlibng_c "zlib-ng/*.c")
file(GLOB _zlibng_x86 "zlib-ng/arch/x86/*.c")
file(GLOB _zlibng_generic "zlib-ng/arch/generic/*.c")
add_library(zlib-ng STATIC
${_zlibng_c}
${_zlibng_x86}
${_zlibng_generic}
)
target_compile_definitions(zlib-ng PRIVATE
X86_FEATURES
X86_HAVE_XSAVE_INTRIN
X86_SSSE3
X86_SSE42
WITH_GZFILEOP
)
if(WIN32)
if(XE_TARGET_AARCH64)
file(GLOB _zlibng_arch "zlib-ng/arch/arm/*.c")
add_library(zlib-ng STATIC
${_zlibng_c}
${_zlibng_arch}
${_zlibng_generic}
)
target_compile_definitions(zlib-ng PRIVATE
X86_SSE2
X86_AVX2
X86_AVX512
X86_AVX512VNNI
X86_PCLMULQDQ_CRC
X86_VPCLMULQDQ_CRC
WITH_GZFILEOP
)
else()
# Remove AVX2/AVX512 files on non-Windows
set(_zlibng_exclude_patterns
"adler32_avx2" "adler32_avx512" "adler32_avx512_vnni"
"chunkset_avx2" "compare256_avx2"
"crc32_pclmulqdq" "crc32_vpclmulqdq"
"slide_hash_avx2"
file(GLOB _zlibng_x86 "zlib-ng/arch/x86/*.c")
add_library(zlib-ng STATIC
${_zlibng_c}
${_zlibng_x86}
${_zlibng_generic}
)
foreach(_pat ${_zlibng_exclude_patterns})
list(FILTER _zlibng_x86 EXCLUDE REGEX "${_pat}")
endforeach()
# Re-set sources without excluded files
get_target_property(_zlibng_srcs zlib-ng SOURCES)
set_target_properties(zlib-ng PROPERTIES SOURCES "")
target_sources(zlib-ng PRIVATE ${_zlibng_c} ${_zlibng_x86} ${_zlibng_generic})
target_compile_definitions(zlib-ng PRIVATE
X86_FEATURES
X86_HAVE_XSAVE_INTRIN
X86_SSSE3
X86_SSE42
WITH_GZFILEOP
)
if(WIN32)
target_compile_definitions(zlib-ng PRIVATE
X86_SSE2
X86_AVX2
X86_AVX512
X86_AVX512VNNI
X86_PCLMULQDQ_CRC
X86_VPCLMULQDQ_CRC
)
else()
# Remove AVX2/AVX512 files on non-Windows
set(_zlibng_exclude_patterns
"adler32_avx2" "adler32_avx512" "adler32_avx512_vnni"
"chunkset_avx2" "compare256_avx2"
"crc32_pclmulqdq" "crc32_vpclmulqdq"
"slide_hash_avx2"
)
foreach(_pat ${_zlibng_exclude_patterns})
list(FILTER _zlibng_x86 EXCLUDE REGEX "${_pat}")
endforeach()
# Re-set sources without excluded files
get_target_property(_zlibng_srcs zlib-ng SOURCES)
set_target_properties(zlib-ng PROPERTIES SOURCES "")
target_sources(zlib-ng PRIVATE ${_zlibng_c} ${_zlibng_x86} ${_zlibng_generic})
endif()
endif()
target_include_directories(zlib-ng PUBLIC zlib-ng)
@@ -389,6 +439,66 @@ endif()
# ==============================================================================
# FFmpeg (source in FFmpeg submodule, config overlay in ffmpeg-xenia)
# ==============================================================================
set(XE_HAS_GAS_ASM FALSE)
if(XE_TARGET_AARCH64)
if(MSVC)
# MSVC can't assemble GAS-syntax .S files. Use clang as the ASM compiler
# if available — the resulting .obj files link fine with MSVC-compiled code.
# Pick clang from the host architecture directory so it can actually
# run on this machine (clang is a cross-compiler — it handles the
# target arch via --target regardless of host).
if(CMAKE_HOST_SYSTEM_PROCESSOR STREQUAL "ARM64")
set(_clang_host_dir "ARM64")
else()
set(_clang_host_dir "x64")
endif()
find_program(CLANG_EXE clang HINTS
"$ENV{VCINSTALLDIR}/Tools/Llvm/${_clang_host_dir}/bin"
"$ENV{VCINSTALLDIR}/Tools/Llvm/bin"
"C:/Program Files/LLVM/bin"
)
if(CLANG_EXE)
set(XE_HAS_GAS_ASM TRUE)
set(XE_CLANG_ASM "${CLANG_EXE}")
message(STATUS "Using clang for ARM64 assembly: ${CLANG_EXE}")
else()
message(WARNING "clang not found — FFmpeg NEON assembly disabled on MSVC ARM64. "
"Install LLVM/clang to enable optimized audio decoding.")
endif()
else()
enable_language(ASM)
set(XE_HAS_GAS_ASM TRUE)
endif()
endif()
# Helper: compile GAS-syntax .S files with clang on MSVC, or natively on
# GCC/Clang. Adds the resulting objects to the target.
function(xe_add_gas_sources target)
if(NOT XE_HAS_GAS_ASM)
return()
endif()
foreach(src ${ARGN})
if(XE_CLANG_ASM)
# MSVC: use clang to assemble .S -> .obj
get_filename_component(src_name "${src}" NAME_WE)
get_filename_component(src_abs "${src}" ABSOLUTE)
set(obj "${CMAKE_CURRENT_BINARY_DIR}/${src_name}.obj")
add_custom_command(
OUTPUT "${obj}"
COMMAND "${XE_CLANG_ASM}" --target=aarch64-pc-windows-msvc
-I "${CMAKE_CURRENT_SOURCE_DIR}/FFmpeg"
-I "${CMAKE_CURRENT_SOURCE_DIR}/ffmpeg-xenia"
-c "${src_abs}" -o "${obj}"
DEPENDS "${src_abs}"
COMMENT "Assembling ${src} with clang"
)
target_sources(${target} PRIVATE "${obj}")
else()
# GCC/Clang: native ASM support
target_sources(${target} PRIVATE "${src}")
endif()
endforeach()
endfunction()
# Common function to apply shared FFmpeg settings to a target.
function(ffmpeg_common target)
@@ -515,7 +625,7 @@ add_library(libavutil STATIC
FFmpeg/libavutil/timestamp.c
)
# x86 platform files
if(NOT CMAKE_SYSTEM_PROCESSOR MATCHES "aarch64|arm64|ARM64")
if(XE_TARGET_X86_64)
target_sources(libavutil PRIVATE
FFmpeg/libavutil/x86/cpu.c
FFmpeg/libavutil/x86/fixed_dsp_init.c
@@ -525,6 +635,24 @@ if(NOT CMAKE_SYSTEM_PROCESSOR MATCHES "aarch64|arm64|ARM64")
FFmpeg/libavutil/x86/tx_float_init.c
)
endif()
# aarch64 platform files
if(XE_TARGET_AARCH64)
target_sources(libavutil PRIVATE
FFmpeg/libavutil/aarch64/cpu.c
)
if(XE_HAS_GAS_ASM)
# NEON init files reference symbols from the .S files, so both must
# be included together. Requires a GAS-compatible assembler (gcc/clang).
target_sources(libavutil PRIVATE
FFmpeg/libavutil/aarch64/float_dsp_init.c
FFmpeg/libavutil/aarch64/tx_float_init.c
)
xe_add_gas_sources(libavutil
FFmpeg/libavutil/aarch64/float_dsp_neon.S
FFmpeg/libavutil/aarch64/tx_float_neon.S
)
endif()
endif()
ffmpeg_common(libavutil)
# --- libavcodec ---
@@ -602,7 +730,7 @@ if(WIN32)
target_sources(libavcodec PRIVATE FFmpeg/libavcodec/file_open.c)
endif()
# x86 platform files
if(NOT CMAKE_SYSTEM_PROCESSOR MATCHES "aarch64|arm64|ARM64")
if(XE_TARGET_X86_64)
target_sources(libavcodec PRIVATE
FFmpeg/libavcodec/x86/constants.c
FFmpeg/libavcodec/x86/fdctdsp_init.c
@@ -611,6 +739,21 @@ if(NOT CMAKE_SYSTEM_PROCESSOR MATCHES "aarch64|arm64|ARM64")
FFmpeg/libavcodec/x86/fdct.c
)
endif()
# aarch64 platform files
if(XE_TARGET_AARCH64)
if(XE_HAS_GAS_ASM)
target_sources(libavcodec PRIVATE
FFmpeg/libavcodec/aarch64/idctdsp_init_aarch64.c
FFmpeg/libavcodec/aarch64/mpegaudiodsp_init.c
)
xe_add_gas_sources(libavcodec
FFmpeg/libavcodec/aarch64/idctdsp_neon.S
FFmpeg/libavcodec/aarch64/mpegaudiodsp_neon.S
FFmpeg/libavcodec/aarch64/neon.S
FFmpeg/libavcodec/aarch64/simple_idct_neon.S
)
endif()
endif()
target_include_directories(libavcodec PRIVATE
${CMAKE_CURRENT_SOURCE_DIR}/FFmpeg/libavcodec
)

View File

@@ -32,9 +32,13 @@
#define ARCH_X86 0
#define ARCH_X86_64 0
#if defined(__aarch64__)
/* ARM64 (Android) */
#define SLIBSUF ".so"
#if defined(__aarch64__) || defined(_M_ARM64)
/* ARM64 */
#if defined(_WIN32)
#define SLIBSUF ".dll"
#else
#define SLIBSUF ".so"
#endif
#undef ARCH_AARCH64
#define ARCH_AARCH64 1
#define HAVE_ARMV8 1
@@ -43,9 +47,15 @@
#define HAVE_ARMV8_EXTERNAL 1
#define HAVE_NEON_EXTERNAL 1
#define HAVE_VFP_EXTERNAL 1
#define HAVE_ARMV8_INLINE 1
#define HAVE_NEON_INLINE 1
#define HAVE_VFP_INLINE 1
#if defined(_MSC_VER)
#define HAVE_ARMV8_INLINE 0
#define HAVE_NEON_INLINE 0
#define HAVE_VFP_INLINE 0
#else
#define HAVE_ARMV8_INLINE 1
#define HAVE_NEON_INLINE 1
#define HAVE_VFP_INLINE 1
#endif
#define HAVE_INTRINSICS_NEON 1
#define HAVE_AS_FUNC 0
#define HAVE_AS_ARCH_DIRECTIVE 0