7 files changed, 147 insertions, 37 deletions

diff --git a/src/host/buildvm_peobj.c b/src/host/buildvm_peobj.c
index 5bca6df8..e3e1026e 100644
--- a/src/host/buildvm_peobj.c
+++ b/src/host/buildvm_peobj.c
@@ -9,7 +9,7 @@
 #include "buildvm.h"
 #include "lj_bc.h"
 
-#if LJ_TARGET_X86ORX64
+#if LJ_TARGET_WINDOWS
 
 /* Context for PE object emitter. */
 static char *strtab;
@@ -93,6 +93,17 @@ typedef struct PEsymaux {
 #define PEOBJ_RELOC_ADDR32NB    0x03
 #define PEOBJ_RELOC_OFS         0
 #define PEOBJ_TEXT_FLAGS        0x60500020  /* 60=r+x, 50=align16, 20=code. */
+#define PEOBJ_PDATA_NRELOC      6
+#define PEOBJ_XDATA_SIZE        (8*2+4+6*2)
+#elif LJ_TARGET_ARM64
+#define PEOBJ_ARCH_TARGET       0xaa64
+#define PEOBJ_RELOC_REL32       0x03  /* MS: BRANCH26. */
+#define PEOBJ_RELOC_DIR32       0x01
+#define PEOBJ_RELOC_ADDR32NB    0x02
+#define PEOBJ_RELOC_OFS         (-4)
+#define PEOBJ_TEXT_FLAGS        0x60500020  /* 60=r+x, 50=align16, 20=code. */
+#define PEOBJ_PDATA_NRELOC      4
+#define PEOBJ_XDATA_SIZE        (4+24+4 +4+8)
 #endif
 
 /* Section numbers (0-based). */
@@ -100,7 +111,7 @@ enum {
   PEOBJ_SECT_ABS = -2,
   PEOBJ_SECT_UNDEF = -1,
   PEOBJ_SECT_TEXT,
-#if LJ_TARGET_X64
+#ifdef PEOBJ_PDATA_NRELOC
   PEOBJ_SECT_PDATA,
   PEOBJ_SECT_XDATA,
 #elif LJ_TARGET_X86
@@ -175,6 +186,9 @@ void emit_peobj(BuildCtx *ctx)
   uint32_t sofs;
   int i, nrsym;
   union { uint8_t b; uint32_t u; } host_endian;
+#ifdef PEOBJ_PDATA_NRELOC
+  uint32_t fcofs = (uint32_t)ctx->sym[ctx->nsym-1].ofs;
+#endif
 
   sofs = sizeof(PEheader) + PEOBJ_NSECTIONS*sizeof(PEsection);
 
@@ -188,18 +202,18 @@ void emit_peobj(BuildCtx *ctx)
   /* Flags: 60 = read+execute, 50 = align16, 20 = code. */
   pesect[PEOBJ_SECT_TEXT].flags = PEOBJ_TEXT_FLAGS;
 
-#if LJ_TARGET_X64
+#ifdef PEOBJ_PDATA_NRELOC
   memcpy(pesect[PEOBJ_SECT_PDATA].name, ".pdata", sizeof(".pdata")-1);
   pesect[PEOBJ_SECT_PDATA].ofs = sofs;
-  sofs += (pesect[PEOBJ_SECT_PDATA].size = 6*4);
+  sofs += (pesect[PEOBJ_SECT_PDATA].size = PEOBJ_PDATA_NRELOC*4);
   pesect[PEOBJ_SECT_PDATA].relocofs = sofs;
-  sofs += (pesect[PEOBJ_SECT_PDATA].nreloc = 6) * PEOBJ_RELOC_SIZE;
+  sofs += (pesect[PEOBJ_SECT_PDATA].nreloc = PEOBJ_PDATA_NRELOC) * PEOBJ_RELOC_SIZE;
   /* Flags: 40 = read, 30 = align4, 40 = initialized data. */
   pesect[PEOBJ_SECT_PDATA].flags = 0x40300040;
 
   memcpy(pesect[PEOBJ_SECT_XDATA].name, ".xdata", sizeof(".xdata")-1);
   pesect[PEOBJ_SECT_XDATA].ofs = sofs;
-  sofs += (pesect[PEOBJ_SECT_XDATA].size = 8*2+4+6*2);  /* See below. */
+  sofs += (pesect[PEOBJ_SECT_XDATA].size = PEOBJ_XDATA_SIZE);  /* See below. */
   pesect[PEOBJ_SECT_XDATA].relocofs = sofs;
   sofs += (pesect[PEOBJ_SECT_XDATA].nreloc = 1) * PEOBJ_RELOC_SIZE;
   /* Flags: 40 = read, 30 = align4, 40 = initialized data. */
@@ -234,7 +248,7 @@ void emit_peobj(BuildCtx *ctx)
   */
   nrsym = ctx->nrelocsym;
   pehdr.nsyms = 1+PEOBJ_NSECTIONS*2 + 1+ctx->nsym + nrsym;
-#if LJ_TARGET_X64
+#ifdef PEOBJ_PDATA_NRELOC
   pehdr.nsyms += 1;  /* Symbol for lj_err_unwind_win. */
 #endif
 
@@ -259,7 +273,6 @@ void emit_peobj(BuildCtx *ctx)
 
 #if LJ_TARGET_X64
   { /* Write .pdata section. */
-    uint32_t fcofs = (uint32_t)ctx->sym[ctx->nsym-1].ofs;
     uint32_t pdata[3];  /* Start of .text, end of .text and .xdata. */
     PEreloc reloc;
     pdata[0] = 0; pdata[1] = fcofs; pdata[2] = 0;
@@ -308,6 +321,88 @@ void emit_peobj(BuildCtx *ctx)
     reloc.type = PEOBJ_RELOC_ADDR32NB;
     owrite(ctx, &reloc, PEOBJ_RELOC_SIZE);
   }
+#elif LJ_TARGET_ARM64
+  /* https://learn.microsoft.com/en-us/cpp/build/arm64-exception-handling */
+  { /* Write .pdata section. */
+    uint32_t pdata[4];
+    PEreloc reloc;
+    pdata[0] = 0;
+    pdata[1] = 0;
+    pdata[2] = fcofs;
+    pdata[3] = 4+24+4;
+    owrite(ctx, &pdata, sizeof(pdata));
+    /* Start of .text and start of .xdata. */
+    reloc.vaddr = 0; reloc.symidx = 1+2+nrsym+2+2+1;
+    reloc.type = PEOBJ_RELOC_ADDR32NB;
+    owrite(ctx, &reloc, PEOBJ_RELOC_SIZE);
+    reloc.vaddr = 4; reloc.symidx = 1+2+nrsym+2;
+    reloc.type = PEOBJ_RELOC_ADDR32NB;
+    owrite(ctx, &reloc, PEOBJ_RELOC_SIZE);
+    /* Start of vm_ffi_call and start of second part of .xdata. */
+    reloc.vaddr = 8; reloc.symidx = 1+2+nrsym+2+2+1;
+    reloc.type = PEOBJ_RELOC_ADDR32NB;
+    owrite(ctx, &reloc, PEOBJ_RELOC_SIZE);
+    reloc.vaddr = 12; reloc.symidx = 1+2+nrsym+2;
+    reloc.type = PEOBJ_RELOC_ADDR32NB;
+    owrite(ctx, &reloc, PEOBJ_RELOC_SIZE);
+  }
+  { /* Write .xdata section. */
+    uint32_t u32;
+    uint8_t *p, uwc[24];
+    PEreloc reloc;
+
+#define CBE16(x)        (*p = ((x) >> 8) & 0xff, p[1] = (x) & 0xff, p += 2)
+#define CALLOC_S(s)     (*p++ = ((s) >> 4))  /* s < 512 */
+#define CSAVE_FPLR(o)   (*p++ = 0x40 | ((o) >> 3))  /* o <= 504 */
+#define CSAVE_REGP(r,o) CBE16(0xc800 | (((r)-19)<< 6) | ((o) >> 3))
+#define CSAVE_REGS(r1,r2,o1) do { \
+  int r, o; for (r = r1, o = o1; r <= r2; r += 2, o -= 16) CSAVE_REGP(r, o); \
+} while (0)
+#define CSAVE_FREGP(r,o) CBE16(0xd800 | (((r) - 8) << 6) | ((o) >> 3))
+#define CSAVE_FREGS(r1,r2,o1) do { \
+  int r, o; for (r = r1, o = o1; r <= r2; r += 2, o -= 16) CSAVE_FREGP(r, o); \
+} while (0)
+#define CSAVE_REGX(r,o) CBE16(0xd400 | (((r) - 19) << 5) | (~(o) >> 3))
+#define CADD_FP(s)      CBE16(0xe200 | ((s) >> 3))  /* s < 8*256 */
+#define CODE_NOP        0xe3
+#define CODE_END        0xe4
+#define CEND_ALIGN      do { \
+  *p++ = CODE_END; \
+  while ((p - uwc) & 3) *p++ = CODE_NOP; \
+} while (0)
+
+    /* Unwind codes for .text section with handler. */
+    p = uwc;
+    CALLOC_S(208);              /* +1 */
+    CSAVE_FPLR(192);            /* +1 */
+    CADD_FP(192);               /* +2 */
+    CSAVE_REGS(19, 28, 184);    /* +5*2 */
+    CSAVE_FREGS(8, 15, 104);    /* +4*2 */
+    CEND_ALIGN;                 /* +1 +1 -> 24 */
+
+    u32 = ((24u >> 2) << 27) | (1u << 20) | (fcofs >> 2);
+    owrite(ctx, &u32, 4);
+    owrite(ctx, &uwc, 24);
+
+    u32 = 0;  /* Handler RVA to be relocated at 4 + 24. */
+    owrite(ctx, &u32, 4);
+
+    /* Unwind codes for vm_ffi_call without handler. */
+    p = uwc;
+    CSAVE_FPLR(16);             /* +1 */
+    CADD_FP(16);                /* +2 */
+    CSAVE_REGX(19, -24);        /* +2 */
+    CSAVE_REGX(20, -32);        /* +2 */
+    CEND_ALIGN;                 /* +1 +0 -> 8 */
+
+    u32 = ((8u >> 2) << 27) | (((uint32_t)ctx->codesz - fcofs) >> 2);
+    owrite(ctx, &u32, 4);
+    owrite(ctx, &uwc, 8);
+
+    reloc.vaddr = 4 + 24; reloc.symidx = 1+2+nrsym+2+2;
+    reloc.type = PEOBJ_RELOC_ADDR32NB;
+    owrite(ctx, &reloc, PEOBJ_RELOC_SIZE);
+  }
 #elif LJ_TARGET_X86
   /* Write .sxdata section. */
   for (i = 0; i < nrsym; i++) {
@@ -339,7 +434,7 @@ void emit_peobj(BuildCtx *ctx)
       emit_peobj_sym(ctx, ctx->relocsym[i], 0,
                      PEOBJ_SECT_UNDEF, PEOBJ_TYPE_FUNC, PEOBJ_SCL_EXTERN);
 
-#if LJ_TARGET_X64
+#ifdef PEOBJ_PDATA_NRELOC
     emit_peobj_sym_sect(ctx, pesect, PEOBJ_SECT_PDATA);
     emit_peobj_sym_sect(ctx, pesect, PEOBJ_SECT_XDATA);
     emit_peobj_sym(ctx, "lj_err_unwind_win", 0,
diff --git a/src/lj_arch.h b/src/lj_arch.h
index 3e920f2a..026e741f 100644
--- a/src/lj_arch.h
+++ b/src/lj_arch.h
@@ -57,7 +57,7 @@
 #define LUAJIT_TARGET   LUAJIT_ARCH_X64
 #elif defined(__arm__) || defined(__arm) || defined(__ARM__) || defined(__ARM)
 #define LUAJIT_TARGET   LUAJIT_ARCH_ARM
-#elif defined(__aarch64__)
+#elif defined(__aarch64__) || defined(_M_ARM64)
 #define LUAJIT_TARGET   LUAJIT_ARCH_ARM64
 #elif defined(__ppc__) || defined(__ppc) || defined(__PPC__) || defined(__PPC) || defined(__powerpc__) || defined(__powerpc) || defined(__POWERPC__) || defined(__POWERPC) || defined(_M_PPC)
 #define LUAJIT_TARGET   LUAJIT_ARCH_PPC
diff --git a/src/lj_def.h b/src/lj_def.h
index 88bc6336..1461d3d7 100644
--- a/src/lj_def.h
+++ b/src/lj_def.h
@@ -146,15 +146,9 @@ typedef uintptr_t BloomFilter;
 #define LJ_UNLIKELY(x)  __builtin_expect(!!(x), 0)
 
 #define lj_ffs(x)       ((uint32_t)__builtin_ctz(x))
-/* Don't ask ... */
-#if defined(__INTEL_COMPILER) && (defined(__i386__) || defined(__x86_64__))
-static LJ_AINLINE uint32_t lj_fls(uint32_t x)
-{
-  uint32_t r; __asm__("bsrl %1, %0" : "=r" (r) : "rm" (x) : "cc"); return r;
-}
-#else
 #define lj_fls(x)       ((uint32_t)(__builtin_clz(x)^31))
-#endif
+#define lj_ffs64(x)     ((uint32_t)__builtin_ctzll(x))
+#define lj_fls64(x)     ((uint32_t)(__builtin_clzll(x)^63))
 
 #if defined(__arm__)
 static LJ_AINLINE uint32_t lj_bswap(uint32_t x)
@@ -265,8 +259,12 @@ static LJ_AINLINE uint32_t lj_fls(uint32_t x)
 #else
 unsigned char _BitScanForward(unsigned long *, unsigned long);
 unsigned char _BitScanReverse(unsigned long *, unsigned long);
+unsigned char _BitScanForward64(unsigned long *, uint64_t);
+unsigned char _BitScanReverse64(unsigned long *, uint64_t);
 #pragma intrinsic(_BitScanForward)
 #pragma intrinsic(_BitScanReverse)
+#pragma intrinsic(_BitScanForward64)
+#pragma intrinsic(_BitScanReverse64)
 
 static LJ_AINLINE uint32_t lj_ffs(uint32_t x)
 {
@@ -277,6 +275,16 @@ static LJ_AINLINE uint32_t lj_fls(uint32_t x)
 {
   unsigned long r; _BitScanReverse(&r, x); return (uint32_t)r;
 }
+
+static LJ_AINLINE uint32_t lj_ffs64(uint64_t x)
+{
+  unsigned long r; _BitScanForward64(&r, x); return (uint32_t)r;
+}
+
+static LJ_AINLINE uint32_t lj_fls64(uint64_t x)
+{
+  unsigned long r; _BitScanReverse64(&r, x); return (uint32_t)r;
+}
 #endif
 
 unsigned long _byteswap_ulong(unsigned long);
diff --git a/src/lj_emit_arm64.h b/src/lj_emit_arm64.h
index 9161c958..fef5d973 100644
--- a/src/lj_emit_arm64.h
+++ b/src/lj_emit_arm64.h
@@ -30,15 +30,15 @@ static uint32_t emit_isk12(int64_t n)
   uint64_t k = n < 0 ? ~(uint64_t)n+1u : (uint64_t)n;
   uint32_t m = n < 0 ? 0x40000000 : 0;
   if (k < 0x1000) {
-    return A64I_K12|m|A64F_U12(k);
+    return (uint32_t)(A64I_K12|m|A64F_U12(k));
   } else if ((k & 0xfff000) == k) {
-    return A64I_K12|m|0x400000|A64F_U12(k>>12);
+    return (uint32_t)(A64I_K12|m|0x400000|A64F_U12(k>>12));
   }
   return 0;
 }
 
-#define emit_clz64(n)   __builtin_clzll(n)
-#define emit_ctz64(n)   __builtin_ctzll(n)
+#define emit_clz64(n)   (lj_fls64(n)^63)
+#define emit_ctz64(n)   lj_ffs64(n)
 
 /* Encode constant in K13 format for logical data processing instructions. */
 static uint32_t emit_isk13(uint64_t n, int is64)
diff --git a/src/lj_mcode.c b/src/lj_mcode.c
index c8ed95e1..8a4851dd 100644
--- a/src/lj_mcode.c
+++ b/src/lj_mcode.c
@@ -29,6 +29,11 @@
 #include <valgrind/valgrind.h>
 #endif
 
+#if LJ_TARGET_WINDOWS
+#define WIN32_LEAN_AND_MEAN
+#include <windows.h>
+#endif
+
 #if LJ_TARGET_IOS
 void sys_icache_invalidate(void *start, size_t len);
 #endif
@@ -41,6 +46,8 @@ void lj_mcode_sync(void *start, void *end)
 #endif
 #if LJ_TARGET_X86ORX64
   UNUSED(start); UNUSED(end);
+#elif LJ_TARGET_WINDOWS
+  FlushInstructionCache(GetCurrentProcess(), start, (char *)end-(char *)start);
 #elif LJ_TARGET_IOS
   sys_icache_invalidate(start, (char *)end-(char *)start);
 #elif LJ_TARGET_PPC
@@ -58,9 +65,6 @@ void lj_mcode_sync(void *start, void *end)
 
 #if LJ_TARGET_WINDOWS
 
-#define WIN32_LEAN_AND_MEAN
-#include <windows.h>
-
 #define MCPROT_RW       PAGE_READWRITE
 #define MCPROT_RX       PAGE_EXECUTE_READ
 #define MCPROT_RWX      PAGE_EXECUTE_READWRITE
diff --git a/src/lj_target.h b/src/lj_target.h
index 09d19bd9..e7322c07 100644
--- a/src/lj_target.h
+++ b/src/lj_target.h
@@ -58,9 +58,13 @@ typedef uint32_t RegSP;
 #if LJ_TARGET_PPC || LJ_TARGET_MIPS || LJ_TARGET_ARM64
 typedef uint64_t RegSet;
 #define RSET_BITS               6
+#define rset_picktop_(rs)       ((Reg)lj_fls64(rs))
+#define rset_pickbot_(rs)       ((Reg)lj_ffs64(rs))
 #else
 typedef uint32_t RegSet;
 #define RSET_BITS               5
+#define rset_picktop_(rs)       ((Reg)lj_fls(rs))
+#define rset_pickbot_(rs)       ((Reg)lj_ffs(rs))
 #endif
 
 #define RID2RSET(r)             (((RegSet)1) << (r))
@@ -71,13 +75,6 @@ typedef uint32_t RegSet;
 #define rset_set(rs, r)         (rs |= RID2RSET(r))
 #define rset_clear(rs, r)       (rs &= ~RID2RSET(r))
 #define rset_exclude(rs, r)     (rs & ~RID2RSET(r))
-#if LJ_TARGET_PPC || LJ_TARGET_MIPS || LJ_TARGET_ARM64
-#define rset_picktop_(rs)       ((Reg)(__builtin_clzll(rs)^63))
-#define rset_pickbot_(rs)       ((Reg)__builtin_ctzll(rs))
-#else
-#define rset_picktop_(rs)       ((Reg)lj_fls(rs))
-#define rset_pickbot_(rs)       ((Reg)lj_ffs(rs))
-#endif
 
 /* -- Register allocation cost -------------------------------------------- */
 
diff --git a/src/msvcbuild.bat b/src/msvcbuild.bat
index f9bf2528..2cfcf26e 100644
--- a/src/msvcbuild.bat
+++ b/src/msvcbuild.bat
@@ -34,20 +34,26 @@
 if exist minilua.exe.manifest^

   %LJMT% -manifest minilua.exe.manifest -outputresource:minilua.exe

 

-@set DASMFLAGS=-D WIN -D JIT -D FFI -D P64

+@set DASMFLAGS=-D WIN -D JIT -D FFI -D ENDIAN_LE -D FPU -D P64

 @set LJARCH=x64

 @minilua

-@if errorlevel 8 goto :X64

+@if errorlevel 8 goto :NO32

 @set DASC=vm_x86.dasc

-@set DASMFLAGS=-D WIN -D JIT -D FFI

+@set DASMFLAGS=-D WIN -D JIT -D FFI -D ENDIAN_LE -D FPU

 @set LJARCH=x86

 @set LJCOMPILE=%LJCOMPILE% /arch:SSE2

+@goto :DA

+:NO32

+@if "%VSCMD_ARG_TGT_ARCH%" neq "arm64" goto :X64

+@set DASC=vm_arm64.dasc

+@set LJARCH=arm64

+@goto :DA

 :X64

-@if "%1" neq "nogc64" goto :GC64

+@if "%1" neq "nogc64" goto :DA

 @shift

 @set DASC=vm_x86.dasc

 @set LJCOMPILE=%LJCOMPILE% /DLUAJIT_DISABLE_GC64

-:GC64

+:DA

 minilua %DASM% -LN %DASMFLAGS% -o host\buildvm_arch.h %DASC%

 @if errorlevel 1 goto :BAD