AArch64 backend for the IR-based JIT

.github/workflows/build.yml

@@ -79,11 +79,6 @@ jobs:
             build_system: cmake-clang-cl
             cmake_generator: Visual Studio 17 2022
 
-          - target: windows
-            architecture: 32
-            ffmpeg_url: https://github.com/HaxeFoundation/hashlink/files/5648055/ffmpeg-3.4.2-win32-dev.zip
-            architecture_string: Win32
-
           - target: windows
             architecture: 64
             ffmpeg_url: https://github.com/HaxeFoundation/hashlink/files/5648056/ffmpeg-3.4.2-win64-dev.zip
@@ -326,23 +321,21 @@ jobs:
             ${{ env.WINDOWS_BUILD_FOLDER }}/hl.exe --version
             ;;
           make*)
-            if [[ ${{ matrix.architecture }} != arm64 ]]; then
-              ./hl --version
-              case ${{ matrix.target }} in
-                linux)  ldd -v ./hl ;;
-                darwin) otool -L ./hl ;;
-              esac
-
-              haxe -hl hello.hl -cp other/tests -main HelloWorld -D interp \
-                ${{ matrix.architecture == 32 && '-D hl-legacy32' || '' }}
-              ./hl hello.hl
-
-              # ensure the executable still works when installed globally
-              cp hello.hl /tmp
-              pushd /tmp
-              hl hello.hl
-              popd
-            fi
+            ./hl --version
+            case ${{ matrix.target }} in
+              linux)  ldd -v ./hl ;;
+              darwin) otool -L ./hl ;;
+            esac
+
+            haxe -hl hello.hl -cp other/tests -main HelloWorld -D interp \
+              ${{ matrix.architecture == 32 && '-D hl-legacy32' || '' }}
+            ./hl hello.hl
+
+            # ensure the executable still works when installed globally
+            cp hello.hl /tmp
+            pushd /tmp
+            hl hello.hl
+            popd
 
             haxe -hl src/_main.c -cp other/tests -main HelloWorld
             make hlc
@@ -431,12 +424,6 @@ jobs:
         os: [darwin, linux, windows]
         architecture: [x86_64, arm64]
         include:
-        - architecture: arm64
-          test-flags: --skip-hl-jit # not yet supported
-
-        - architecture: x86_32
-          test-flags: --skip-hl-jit --msbuild-platform x86 -D hlgen.build.architecture=x86_32 # not stable
-
         - os: linux
           runner: ubuntu-latest
 
@@ -511,12 +498,7 @@ jobs:
         tar -xvzf hashlink-*.tar.gz
         cd $(echo hashlink-*/)
         sudo mkdir -p /usr/local/{bin,lib,include}
-        if [[ "${{ matrix.architecture }}" == "arm64" ]]; then
-          printf "#!/bin/sh\necho Jit is not supported on arm64\n" > /usr/local/bin/hl
-          chmod +x /usr/local/bin/hl
-        else
-          sudo cp hl /usr/local/bin/
-        fi
+        sudo cp hl /usr/local/bin/
         sudo cp libhl.* *.hdll /usr/local/lib/
         sudo cp include/* /usr/local/include/
 

CMakeLists.txt

@@ -20,7 +20,8 @@ include(FindPkgConfig)
 include(CTest)
 
 set(WITH_VM_DEFAULT ON)
-if(CMAKE_SYSTEM_PROCESSOR MATCHES "arm|aarch64" AND (NOT CMAKE_OSX_ARCHITECTURES MATCHES "x86_64"))
+# 32-bit ARM has no JIT backend; aarch64/arm64 uses src/jit_aarch64.c.
+if(CMAKE_SYSTEM_PROCESSOR MATCHES "^arm$|^armv[0-9]" AND (NOT CMAKE_OSX_ARCHITECTURES MATCHES "x86_64"))
     set(WITH_VM_DEFAULT OFF)
 endif()
 
@@ -225,12 +226,17 @@ else()
 endif()
 
 if (WITH_VM)
+    if(CMAKE_SYSTEM_PROCESSOR MATCHES "aarch64|arm64|ARM64")
+        set(HL_JIT_BACKEND src/jit_aarch64.c src/jit_aarch64_emit.c)
+    else()
+        set(HL_JIT_BACKEND src/jit_x86_64.c)
+    endif()
     add_executable(hl
         src/code.c
         src/jit.c
         src/jit_emit.c
         src/jit_regs.c
-        src/jit_x86_64.c
+        ${HL_JIT_BACKEND}
         src/jit_dump.c
         src/main.c
         src/module.c

Makefile

@@ -41,7 +41,13 @@ STD = src/std/array.o src/std/buffer.o src/std/bytes.o src/std/cast.o src/std/da
 	src/std/socket.o src/std/string.o src/std/sys.o src/std/types.o src/std/ucs2.o src/std/thread.o src/std/process.o \
 	src/std/track.o
 
-HL_OBJ = src/code.o src/jit.o src/jit_emit.o src/jit_regs.o src/jit_x86_64.o src/jit_dump.o src/main.o src/module.o src/debugger.o src/profile.o
+ifneq (,$(filter aarch64 arm64,$(ARCH)))
+HL_JIT_BACKEND_OBJ = src/jit_aarch64.o src/jit_aarch64_emit.o
+else
+HL_JIT_BACKEND_OBJ = src/jit_x86_64.o
+endif
+
+HL_OBJ = src/code.o src/jit.o src/jit_emit.o src/jit_regs.o $(HL_JIT_BACKEND_OBJ) src/jit_dump.o src/main.o src/module.o src/debugger.o src/profile.o
 
 FMT_CPPFLAGS = -I include/mikktspace -I include/minimp3
 
@@ -240,19 +246,12 @@ LIBHL = libhl.$(LIBEXT)
 HL = hl$(EXE_SUFFIX)
 HLC = hlc$(EXE_SUFFIX)
 
-all: $(LIBHL) libs
-ifeq ($(ARCH),arm64)
-	$(warning HashLink vm is not supported on arm64, skipping)
-else
-all: $(HL)
-endif
+all: $(LIBHL) libs $(HL)
 
 install:
 	$(UNAME)==Darwin && ${MAKE} uninstall
-ifneq ($(ARCH),arm64)
 	mkdir -p $(INSTALL_BIN_DIR)
 	cp $(HL) $(INSTALL_BIN_DIR)
-endif
 	mkdir -p $(INSTALL_LIB_DIR)
 	cp *.hdll $(INSTALL_LIB_DIR)
 	cp $(LIBHL) $(INSTALL_LIB_DIR)
@@ -365,11 +364,7 @@ release_win:
 	rm -rf $(PACKAGE_NAME)
 
 release_linux release_osx:
-ifeq ($(ARCH),arm64)
-	cp $(LIBHL) *.hdll $(PACKAGE_NAME)
-else
 	cp $(HL) $(LIBHL) *.hdll $(PACKAGE_NAME)
-endif
 	tar -cvzf $(PACKAGE_NAME).tar.gz $(PACKAGE_NAME)
 	rm -rf $(PACKAGE_NAME)
 

src/gc.c

@@ -28,6 +28,10 @@
 #	include <sys/mman.h>
 #endif
 
+#if defined(__APPLE__) && defined(__aarch64__)
+#	include <pthread.h>
+#endif
+
 #if defined(HL_EMSCRIPTEN)
 #	include <emscripten/heap.h>
 #endif
@@ -1151,11 +1155,35 @@ HL_PRIM void *hl_alloc_executable_memory( int size ) {
 	return NULL;
 #else
 	void *p;
-	p = mmap(NULL,size,PROT_READ|PROT_WRITE|PROT_EXEC,(MAP_PRIVATE|MAP_ANONYMOUS),-1,0);
+	int flags = MAP_PRIVATE | MAP_ANONYMOUS;
+#	if defined(__APPLE__) && defined(__aarch64__)
+	// Apple Silicon requires MAP_JIT for W^X-protected JIT pages; the binary
+	// must also carry the com.apple.security.cs.allow-jit entitlement.
+#		ifndef MAP_JIT
+#			define MAP_JIT 0x800
+#		endif
+	flags |= MAP_JIT;
+#	endif
+	p = mmap(NULL,size,PROT_READ|PROT_WRITE|PROT_EXEC,flags,-1,0);
+	if( p == MAP_FAILED ) return NULL;
+#	if defined(__APPLE__) && defined(__aarch64__)
+	// Leave the caller's thread in write mode so it can populate the page;
+	// hl_flush_executable_memory flips back to exec when emission is done.
+	pthread_jit_write_protect_np(false);
+#	endif
 	return p;
 #endif
 }
 
+HL_PRIM void hl_flush_executable_memory( void *code, int size ) {
+#if defined(__GNUC__) || defined(__clang__)
+	__builtin___clear_cache((char*)code, (char*)code + size);
+#endif
+#if defined(__APPLE__) && defined(__aarch64__)
+	pthread_jit_write_protect_np(true);
+#endif
+}
+
 HL_PRIM void hl_free_executable_memory( void *c, int size ) {
 #if defined(HL_WIN)
 	VirtualFree(c,0,MEM_RELEASE);

src/hl.h

@@ -824,6 +824,7 @@ HL_API void hl_global_lock( bool lock );
 
 HL_API void *hl_alloc_executable_memory( int size );
 HL_API void hl_free_executable_memory( void *ptr, int size );
+HL_API void hl_flush_executable_memory( void *code, int size );
 
 // ----------------------- BUFFER --------------------------------------------------
 

src/jit.c

@@ -331,6 +331,7 @@ void *hl_jit_code( jit_ctx *ctx, hl_module *m, int *codesize, hl_debug_infos **d
 	ctx->final_code = code;
 	hl_emit_final(ctx);
 	hl_codegen_final(ctx);
+	hl_flush_executable_memory(code, size);
 	arg_reg_count = ctx->cfg.regs.nargs;
 	arg_fp_count = ctx->cfg.floats.nargs;
 	call_jit_c2hl = ctx->final_code + ctx->code_funs.c2hl;

src/jit.h

@@ -187,12 +187,31 @@ typedef struct {
 	ereg *arg;
 } reg_config;
 
+// HL→native stack-arg layout. HL→HL calls always use the HL convention
+// (cfg.stack_arg_size stride, PUSH-based). HL→native calls follow the
+// platform C ABI, which on AArch64 differs from HL's convention:
+//   - Linux/Windows AArch64 (AAPCS64): each stack arg occupies an 8-byte slot.
+//   - Apple ARM64: stack args are packed at their natural size with
+//     natural alignment.
+typedef enum {
+	NATIVE_STACK_LAYOUT_HL = 0,        // default: same as HL stride (x86_64, etc.)
+	NATIVE_STACK_LAYOUT_AAPCS64,       // 8-byte slots (Linux/Win AArch64 native)
+	NATIVE_STACK_LAYOUT_APPLE_ARM64,   // natural-size packed (Apple ARM64 native)
+} native_stack_layout_kind;
+
 typedef struct {
 	reg_config regs;
 	reg_config floats;
 	ereg stack_reg;
 	ereg stack_pos;
 	int stack_align;
+	// Minimum bytes consumed by each stack argument. Defaults to HL_WSIZE
+	// when 0. Backends like AArch64 set this to 16 because each PUSH must
+	// move SP by 16 bytes to keep SP 16-byte aligned (any [SP, ...] access
+	// with a misaligned SP traps under EL0).
+	int stack_arg_size;
+	// Layout for HL→native (CALL_PTR) stack args. See native_stack_layout_kind.
+	int native_stack_layout;
 	int debug_prefix_size;
 	ereg req_bit_shifts;
 	ereg req_div_a;