arm64-win32 support : runtime

include/tccdefs.h

@@ -224,7 +224,9 @@
                            &~3), *(type *)(ap - ((sizeof(type)+3)&~3)))
 
 #elif defined __aarch64__
-#if defined __APPLE__
+#if defined _WIN32
+    typedef char *__builtin_va_list;
+#elif defined __APPLE__
     typedef struct {
         void *__stack;
     } __builtin_va_list;

lib/bt-exe.c

@@ -11,6 +11,21 @@
 # define __declspec(n)
 #endif
 
+#ifdef _WIN64
+static void bt_init_pe_prog_base(rt_context *p)
+{
+    MEMORY_BASIC_INFORMATION mbi;
+    addr_t imagebase;
+
+    if (!p->prog_base)
+        return;
+    if (!VirtualQuery(p, &mbi, sizeof(mbi)) || !mbi.AllocationBase)
+        return;
+    imagebase = (addr_t)mbi.AllocationBase - p->prog_base;
+    p->prog_base = (addr_t)mbi.AllocationBase - (imagebase & 0xffffffffu);
+}
+#endif
+
 __declspec(dllexport)
 void __bt_init(rt_context *p, int is_exe)
 {
@@ -24,6 +39,10 @@ void __bt_init(rt_context *p, int is_exe)
     if (p->bounds_start)
 	__bound_init(p->bounds_start, -1);
 
+#ifdef _WIN64
+    bt_init_pe_prog_base(p);
+#endif
+
     /* add to chain */
     rt_wait_sem();
     p->next = g_rc, g_rc = p;
@@ -66,3 +85,13 @@ ST_FUNC char *pstrcpy(char *buf, size_t buf_size, const char *s)
     buf[l] = 0;
     return buf;
 }
+
+#if defined(_WIN64) && defined(__aarch64__)
+/* The bt-only Windows ARM64 build should not rely on importing this helper. */
+LONG InterlockedExchange(LONG volatile *Target, LONG Value)
+{
+    LONG Old = *Target;
+    *Target = Value;
+    return Old;
+}
+#endif

lib/libtcc1.c

@@ -630,6 +630,12 @@ long long __fixxfdi (long double a1)
 /* MSVC x64 intrinsic */
 void __faststorefence(void)
 {
-    __asm__("lock; orl $0,(%rsp)");
+#if defined(__aarch64__)
+    /* ARM64: Data Memory Barrier (Inner Shareable) */
+    __asm__("dmb ish");
+#else
+    /* x86-64: lock prefix to flush store buffer */
+    __asm__("lock; orl $0,(%%rsp)" ::: "memory");
+#endif
 }
 #endif

win32/include/_mingw.h

@@ -70,12 +70,21 @@
 
 #ifdef _WIN64
 #define __stdcall
+#if defined(__aarch64__) || defined(_M_ARM64) || defined(_ARM64_)
+#ifndef _M_ARM64
+#define _M_ARM64 1
+#endif
+#ifndef _ARM64_
+#define _ARM64_ 1
+#endif
+#else
 #define _AMD64_ 1
 #define __x86_64 1
 #define _M_X64 100 /* Visual Studio */
 #define _M_AMD64 100 /* Visual Studio */
 #define USE_MINGW_SETJMP_TWO_ARGS
 #define mingw_getsp tinyc_getbp
+#endif
 #else
 #define __stdcall __attribute__((__stdcall__))
 #define _X86_ 1

win32/include/setjmp.h

@@ -124,6 +124,55 @@ extern "C" {
     SETJMP_FLOAT128 Xmm14;
     SETJMP_FLOAT128 Xmm15;
   } _JUMP_BUFFER;
+#elif defined(_ARM_)
+
+#define _JBLEN 28
+#define _JBTYPE int
+
+  typedef struct __JUMP_BUFFER {
+    unsigned long Frame;
+    unsigned long R4;
+    unsigned long R5;
+    unsigned long R6;
+    unsigned long R7;
+    unsigned long R8;
+    unsigned long R9;
+    unsigned long R10;
+    unsigned long R11;
+    unsigned long Sp;
+    unsigned long Pc;
+    unsigned long Fpscr;
+    unsigned long long D[8];
+  } _JUMP_BUFFER;
+#elif defined(_ARM64_)
+
+#define _JBLEN 24
+#define _JBTYPE unsigned __int64
+
+  typedef struct __JUMP_BUFFER {
+    unsigned __int64 Frame;
+    unsigned __int64 Reserved;
+    unsigned __int64 X19;
+    unsigned __int64 X20;
+    unsigned __int64 X21;
+    unsigned __int64 X22;
+    unsigned __int64 X23;
+    unsigned __int64 X24;
+    unsigned __int64 X25;
+    unsigned __int64 X26;
+    unsigned __int64 X27;
+    unsigned __int64 X28;
+    unsigned __int64 Fp;
+    unsigned __int64 Lr;
+    unsigned __int64 Sp;
+    unsigned long Fpcr;
+    unsigned long Fpsr;
+    double D[8];
+  } _JUMP_BUFFER;
+#else
+
+#define _JBLEN 1
+#define _JBTYPE int
 #endif
 #ifndef _JMP_BUF_DEFINED
   typedef _JBTYPE jmp_buf[_JBLEN];

win32/include/winapi/winnt.h

@@ -21,7 +21,7 @@ extern "C" {
 #define __CRT_UNALIGNED
 #endif
 
-#if defined(__ia64__) || defined(__x86_64)
+#if defined(__ia64__) || defined(__x86_64) || defined(__aarch64__)
 #define UNALIGNED __CRT_UNALIGNED
 #ifdef _WIN64
 #define UNALIGNED64 __CRT_UNALIGNED
@@ -47,6 +47,9 @@ extern "C" {
 #endif
 #endif
 
+#if !defined(I_X86_) && !defined(_IA64_) && !defined(_AMD64_) && defined(__aarch64__) && !defined(_ARM64_)
+#define _ARM64_
+#endif
 
 #ifdef _WIN64
 #define MAX_NATURAL_ALIGNMENT sizeof(ULONGLONG)
@@ -65,7 +68,7 @@ extern "C" {
 #ifdef _WIN64
 #ifdef _AMD64_
 #define PROBE_ALIGNMENT(_s) TYPE_ALIGNMENT(DWORD)
-#elif defined(_IA64_)
+#elif defined(_IA64_) || defined(_ARM64_)
 #define PROBE_ALIGNMENT(_s) (TYPE_ALIGNMENT(_s) > TYPE_ALIGNMENT(DWORD) ? TYPE_ALIGNMENT(_s) : TYPE_ALIGNMENT(DWORD))
 #else
 #error No Target Architecture
@@ -79,7 +82,7 @@ extern "C" {
 
 #include <basetsd.h>
 
-#if defined(_X86_) || defined(__ia64__) || defined(__x86_64)
+#if defined(_X86_) || defined(__ia64__) || defined(__x86_64) || defined(__aarch64__)
 #define DECLSPEC_IMPORT __declspec(dllimport)
 #else
 #define DECLSPEC_IMPORT
@@ -321,7 +324,7 @@ typedef DWORD LCID;
 #define Int32x32To64(a,b) (LONGLONG)((LONGLONG)(LONG)(a) *(LONG)(b))
 #define UInt32x32To64(a,b) (ULONGLONG)((ULONGLONG)(DWORD)(a) *(DWORD)(b))
 #define Int64ShrlMod32(a,b) ((DWORDLONG)(a)>>(b))
-#elif defined(__ia64__) || defined(__x86_64)
+#elif defined(__ia64__) || defined(__x86_64) || defined(__aarch64__)
 #define Int32x32To64(a,b) ((LONGLONG)((LONG)(a)) *(LONGLONG)((LONG)(b)))
 #define UInt32x32To64(a,b) ((ULONGLONG)((DWORD)(a)) *(ULONGLONG)((DWORD)(b)))
 #define Int64ShrlMod32(a,b) ((ULONGLONG)(a) >> (b))
@@ -829,7 +832,7 @@ typedef DWORD LCID;
   typedef ULONG_PTR KSPIN_LOCK;
   typedef KSPIN_LOCK *PKSPIN_LOCK;
 
-#ifdef _AMD64_
+#if defined(_AMD64_) || defined(_ARM64_)
 
 #if defined(__x86_64) && !defined(RC_INVOKED)
 
@@ -1336,6 +1339,7 @@ typedef DWORD LCID;
 
 #define LEGACY_SAVE_AREA_LENGTH sizeof(XMM_SAVE_AREA32)
 
+#if defined(__x86_64) || defined(_AMD64_)
   typedef struct DECLSPEC_ALIGN(16) _CONTEXT {
     DWORD64 P1Home;
     DWORD64 P2Home;
@@ -1407,6 +1411,7 @@ typedef DWORD LCID;
     DWORD64 LastExceptionToRip;
     DWORD64 LastExceptionFromRip;
   } CONTEXT,*PCONTEXT;
+#endif /* defined(__x86_64) || defined(_AMD64_) */
 
 #define RUNTIME_FUNCTION_INDIRECT 0x1
 
@@ -1417,6 +1422,123 @@ typedef DWORD LCID;
   } RUNTIME_FUNCTION,*PRUNTIME_FUNCTION;
 
   typedef PRUNTIME_FUNCTION (*PGET_RUNTIME_FUNCTION_CALLBACK)(DWORD64 ControlPc,PVOID Context);
+
+#if defined(_ARM64_) || defined(__aarch64__)
+
+/* ARM64 Context Definition */
+#define CONTEXT_ARM64 0x00400000
+
+#ifndef CONTEXT_CONTROL
+#define CONTEXT_CONTROL (CONTEXT_ARM64 | 0x00000001L)
+#endif
+#ifndef CONTEXT_INTEGER
+#define CONTEXT_INTEGER (CONTEXT_ARM64 | 0x00000002L)
+#endif
+#ifndef CONTEXT_FLOATING_POINT
+#define CONTEXT_FLOATING_POINT (CONTEXT_ARM64 | 0x00000004L)
+#endif
+#ifndef CONTEXT_DEBUG
+#define CONTEXT_DEBUG (CONTEXT_ARM64 | 0x00000008L)
+#endif
+
+#ifndef CONTEXT_FULL
+#define CONTEXT_FULL (CONTEXT_CONTROL | CONTEXT_INTEGER | CONTEXT_FLOATING_POINT)
+#endif
+#ifndef CONTEXT_ALL
+#define CONTEXT_ALL (CONTEXT_CONTROL | CONTEXT_INTEGER | CONTEXT_FLOATING_POINT | CONTEXT_DEBUG)
+#endif
+
+#ifndef ARM64_MAX_BREAKPOINTS
+#define ARM64_MAX_BREAKPOINTS 8
+#endif
+#ifndef ARM64_MAX_WATCHPOINTS
+#define ARM64_MAX_WATCHPOINTS 2
+#endif
+
+#ifndef _ARM64_NT_NEON128_DECLARED
+#define _ARM64_NT_NEON128_DECLARED
+  typedef union _ARM64_NT_NEON128 {
+    struct {
+      ULONGLONG Low;
+      LONGLONG High;
+    } DUMMYSTRUCTNAME;
+    double D[2];
+    float S[4];
+    WORD H[8];
+    BYTE B[16];
+  } ARM64_NT_NEON128,*PARM64_NT_NEON128;
+#endif
+
+#ifndef _ARM64_CONTEXT_DECLARED
+#define _ARM64_CONTEXT_DECLARED
+  typedef struct DECLSPEC_ALIGN(16) _ARM64_NT_CONTEXT {
+    ULONG ContextFlags;
+    ULONG Cpsr;
+    union {
+      struct {
+        DWORD64 X0;
+        DWORD64 X1;
+        DWORD64 X2;
+        DWORD64 X3;
+        DWORD64 X4;
+        DWORD64 X5;
+        DWORD64 X6;
+        DWORD64 X7;
+        DWORD64 X8;
+        DWORD64 X9;
+        DWORD64 X10;
+        DWORD64 X11;
+        DWORD64 X12;
+        DWORD64 X13;
+        DWORD64 X14;
+        DWORD64 X15;
+        DWORD64 X16;
+        DWORD64 X17;
+        DWORD64 X18;
+        DWORD64 X19;
+        DWORD64 X20;
+        DWORD64 X21;
+        DWORD64 X22;
+        DWORD64 X23;
+        DWORD64 X24;
+        DWORD64 X25;
+        DWORD64 X26;
+        DWORD64 X27;
+        DWORD64 X28;
+        DWORD64 Fp;
+        DWORD64 Lr;
+      } DUMMYSTRUCTNAME;
+      DWORD64 X[31];
+    } DUMMYUNIONNAME;
+    DWORD64 Sp;
+    DWORD64 Pc;
+    ARM64_NT_NEON128 V[32];
+    DWORD Fpcr;
+    DWORD Fpsr;
+    DWORD Bcr[ARM64_MAX_BREAKPOINTS];
+    DWORD64 Bvr[ARM64_MAX_BREAKPOINTS];
+    DWORD Wcr[ARM64_MAX_WATCHPOINTS];
+    DWORD64 Wvr[ARM64_MAX_WATCHPOINTS];
+  } ARM64_NT_CONTEXT,*PARM64_NT_CONTEXT;
+
+  C_ASSERT(sizeof(ARM64_NT_CONTEXT) == 0x390);
+  C_ASSERT(offsetof(ARM64_NT_CONTEXT, ContextFlags) == 0x000);
+  C_ASSERT(offsetof(ARM64_NT_CONTEXT, X) == 0x008);
+  C_ASSERT(offsetof(ARM64_NT_CONTEXT, Fp) == 0x0f0);
+  C_ASSERT(offsetof(ARM64_NT_CONTEXT, Lr) == 0x0f8);
+  C_ASSERT(offsetof(ARM64_NT_CONTEXT, Sp) == 0x100);
+  C_ASSERT(offsetof(ARM64_NT_CONTEXT, Pc) == 0x108);
+  C_ASSERT(offsetof(ARM64_NT_CONTEXT, V) == 0x110);
+  C_ASSERT(sizeof(((ARM64_NT_CONTEXT *)0)->V[0]) == 16);
+  C_ASSERT(offsetof(ARM64_NT_CONTEXT, Fpcr) == 0x310);
+  C_ASSERT(offsetof(ARM64_NT_CONTEXT, Fpsr) == 0x314);
+  C_ASSERT(offsetof(ARM64_NT_CONTEXT, Bvr) == 0x338);
+  C_ASSERT(offsetof(ARM64_NT_CONTEXT, Wvr) == 0x380);
+#endif
+
+  typedef ARM64_NT_CONTEXT CONTEXT,*PCONTEXT;
+
+#endif /* _ARM64_ || __aarch64__ */
   typedef DWORD (*POUT_OF_PROCESS_FUNCTION_TABLE_CALLBACK)(HANDLE Process,PVOID TableAddress,PDWORD Entries,PRUNTIME_FUNCTION *Functions);
 
 #define OUT_OF_PROCESS_FUNCTION_TABLE_CALLBACK_EXPORT_NAME "OutOfProcessFunctionTableCallback"
@@ -3701,6 +3823,7 @@ typedef DWORD LCID;
 #define IMAGE_FILE_MACHINE_CEF 0x0CEF
 #define IMAGE_FILE_MACHINE_EBC 0x0EBC
 #define IMAGE_FILE_MACHINE_AMD64 0x8664
+#define IMAGE_FILE_MACHINE_ARM64 0xAA64
 #define IMAGE_FILE_MACHINE_M32R 0x9041
 #define IMAGE_FILE_MACHINE_CEE 0xC0EE
 
@@ -3857,10 +3980,16 @@ typedef DWORD LCID;
 #define IMAGE_SUBSYSTEM_EFI_ROM 13
 #define IMAGE_SUBSYSTEM_XBOX 14
 
+#define IMAGE_DLLCHARACTERISTICS_HIGH_ENTROPY_VA 0x0020
+#define IMAGE_DLLCHARACTERISTICS_DYNAMIC_BASE 0x0040
+#define IMAGE_DLLCHARACTERISTICS_FORCE_INTEGRITY 0x0080
+#define IMAGE_DLLCHARACTERISTICS_NX_COMPAT 0x0100
 #define IMAGE_DLLCHARACTERISTICS_NO_ISOLATION 0x0200
 #define IMAGE_DLLCHARACTERISTICS_NO_SEH 0x0400
 #define IMAGE_DLLCHARACTERISTICS_NO_BIND 0x0800
+#define IMAGE_DLLCHARACTERISTICS_APPCONTAINER 0x1000
 #define IMAGE_DLLCHARACTERISTICS_WDM_DRIVER 0x2000
+#define IMAGE_DLLCHARACTERISTICS_GUARD_CF 0x4000
 #define IMAGE_DLLCHARACTERISTICS_TERMINAL_SERVER_AWARE 0x8000
 
 #define IMAGE_DIRECTORY_ENTRY_EXPORT 0

win32/lib/chkstk.S

@@ -8,7 +8,92 @@
 #endif
 
 /* ---------------------------------------------- */
-#ifndef __x86_64__
+#if defined(__aarch64__)
+/* ---------------------------------------------- */
+
+.globl __chkstk
+__chkstk:
+    /* Windows ARM64 stack probing helper.
+       arm64-gen.c passes the requested frame size in x15, scaled in 16-byte
+       units. Probe one 4 KiB page at a time and leave SP unchanged; the caller
+       subtracts SP after the probe returns. */
+    mov     x16, sp
+    lsl     x17, x15, 4
+    cbz     x17, L_chkstk_done
+L_chkstk_loop:
+    subs    x0, x17, 4096
+    bls     L_chkstk_tail
+    sub     x16, x16, 4096
+    ldr     xzr, [x16]
+    sub     x17, x17, 4096
+    b       L_chkstk_loop
+L_chkstk_tail:
+    sub     x16, x16, x17
+    ldr     xzr, [x16]
+L_chkstk_done:
+    ret
+
+.globl _(tinyc_getbp)
+_(tinyc_getbp):
+    mov     x0, x29
+    ret
+
+.globl _(mingw_getsp)
+_(mingw_getsp):
+    mov     x0, sp
+    ret
+
+.globl _(__mingw_setjmp)
+_(__mingw_setjmp):
+    /* _JUMP_BUFFER layout matches win32/include/setjmp.h for _ARM64_:
+       0x00 Frame, 0x08 Reserved, 0x10-0x68 X19-X30, 0x70 Sp,
+       0x78 Fpcr/Fpsr, 0x80-0xB8 D8-D15. */
+    str     xzr, [x0]       /* Frame = 0 */
+    stp     x19, x20, [x0, 16]
+    stp     x21, x22, [x0, 32]
+    stp     x23, x24, [x0, 48]
+    stp     x25, x26, [x0, 64]
+    stp     x27, x28, [x0, 80]
+    stp     x29, x30, [x0, 96]
+    mov     x2, sp
+    str     x2, [x0, 112]   /* Sp */
+    mrs     x2, FPCR
+    str     w2, [x0, 120]   /* Fpcr */
+    mrs     x2, FPSR
+    str     w2, [x0, 124]   /* Fpsr */
+    stp     d8, d9, [x0, 128]
+    stp     d10, d11, [x0, 144]
+    stp     d12, d13, [x0, 160]
+    stp     d14, d15, [x0, 176]
+    mov     x0, 0
+    ret
+
+.globl _(__mingw_longjmp)
+_(__mingw_longjmp):
+    ldp     x19, x20, [x0, 16]
+    ldp     x21, x22, [x0, 32]
+    ldp     x23, x24, [x0, 48]
+    ldp     x25, x26, [x0, 64]
+    ldp     x27, x28, [x0, 80]
+    ldp     x29, x30, [x0, 96]
+    ldr     x2, [x0, 112]   /* Sp */
+    mov     sp, x2
+    ldr     w2, [x0, 120]   /* Fpcr */
+    msr     FPCR, x2
+    ldr     w2, [x0, 124]   /* Fpsr */
+    msr     FPSR, x2
+    ldp     d8, d9, [x0, 128]
+    ldp     d10, d11, [x0, 144]
+    ldp     d12, d13, [x0, 160]
+    ldp     d14, d15, [x0, 176]
+    mov     x0, x1
+    cbnz    x0, L_longjmp_done
+    mov     x0, 1
+L_longjmp_done:
+    ret
+
+/* ---------------------------------------------- */
+#elif !defined(__x86_64__)
 /* ---------------------------------------------- */
 
 .globl _(__chkstk)
@@ -69,5 +154,3 @@ _(tinyc_getbp):
 /* ---------------------------------------------- */
 #endif
 /* ---------------------------------------------- */
-
-