diff --git a/.github/workflows/build.yml b/.github/workflows/build.yml
index 2fbddce1..6ffa502f 100644
--- a/.github/workflows/build.yml
+++ b/.github/workflows/build.yml
@@ -34,56 +34,51 @@ jobs:
     timeout-minutes: 6
     steps:
     - uses: actions/checkout@v4
-    - name: make & test tcc (x86_64-win32)
-      shell: cmd
-      run: |
-        echo ::group:: setup msys mingw64-gcc
-        set MSYS2_PATH_TYPE=inherit
-        set MSYSTEM=MINGW64
-        set CHERE_INVOKING=yes
-        C:\msys64\usr\bin\bash -l -c "pacman -S --noconfirm mingw-w64-x86_64-gcc"
-        echo ::endgroup::
-        C:\msys64\usr\bin\bash -l -c "./configure && make && make test -k"
-    - uses: ilammy/msvc-dev-cmd@v1
-      with:
-        arch: amd64
-    - name: build with MSVC (x86_64-win32)
+    - name: test (x86_64-win32)
       shell: cmd
       run: |
         echo ::group:: run build-tcc.bat
         cd win32
-        call build-tcc.bat -t 64 -c cl
+        call "%ProgramFiles%\Microsoft Visual Studio\2022\Enterprise\Common7\Tools\VsDevCmd.bat" -arch=amd64
+        call build-tcc.bat -c cl -t x86_64
+        @echo off
         echo ::endgroup::
-        .\tcc -I.. libtcc.dll -v ../tests/libtcc_test.c -o libtest.exe && .\libtest.exe
-        .\tcc -I.. libtcc.dll -run ../tests/libtcc_test.c
+        cd ..\tests
+        call test-win32.bat all -k
 
   test-i386-win32:
     runs-on: windows-2025
     timeout-minutes: 6
     steps:
     - uses: actions/checkout@v4
-    - name: make & test tcc (i386-win32)
-      shell: cmd
-      run: |
-        echo ::group:: setup msys mingw32-gcc
-        set MSYS2_PATH_TYPE=inherit
-        set MSYSTEM=MINGW32
-        set CHERE_INVOKING=yes
-        C:\msys64\usr\bin\bash -l -c "pacman -S --noconfirm mingw-w64-i686-gcc"
-        echo ::endgroup::
-        C:\msys64\usr\bin\bash -l -c "./configure && make && make test -k"
-    - uses: ilammy/msvc-dev-cmd@v1
-      with:
-        arch: x86
-    - name: build with MSVC (i386-win32)
+    - name: test (i386-win32)
       shell: cmd
       run: |
         echo ::group:: run build-tcc.bat
         cd win32
-        call build-tcc.bat -t 32 -c cl
+        call "%ProgramFiles%\Microsoft Visual Studio\2022\Enterprise\Common7\Tools\VsDevCmd.bat" -arch=x86
+        call build-tcc.bat -c cl -t i386
+        @echo off
         echo ::endgroup::
-        .\tcc -I.. libtcc.dll -v ../tests/libtcc_test.c -o libtest.exe && .\libtest.exe
-        .\tcc -I.. libtcc.dll -run ../tests/libtcc_test.c
+        cd ..\tests
+        call test-win32.bat -p c:\mingw32\bin all -k
+
+  test-arm64-win32:
+    runs-on: windows-11-arm
+    timeout-minutes: 6
+    steps:
+    - uses: actions/checkout@v4
+    - name: test (arm64-win32)
+      shell: cmd
+      run: |
+        echo ::group:: run build-tcc.bat
+        cd win32
+        call "%ProgramFiles%\Microsoft Visual Studio\2022\Enterprise\Common7\Tools\VsDevCmd.bat" -arch=arm64
+        call build-tcc.bat -c cl -t arm64
+        @echo off
+        echo ::endgroup::
+        cd ..\tests
+        call test-win32.bat -c clang all -k
 
   test-armv7-linux:
     runs-on: ubuntu-22.04
diff --git a/Makefile b/Makefile
index 4491bd9d..125cc918 100644
--- a/Makefile
+++ b/Makefile
@@ -31,7 +31,9 @@ ifdef CONFIG_WIN32
   LIBTCCDEF = libtcc.def
  endif
  ifneq ($(CONFIG_debug),yes)
-  LDFLAGS += -s
+  ifneq ($(CC_NAME),clang)
+   LDFLAGS += -s
+  endif
  endif
  NATIVE_TARGET = $(if $(findstring arm64,$(ARCH)),arm64-win32,$(ARCH)-win$(if $(findstring arm,$(ARCH)),ce,32))
 else
diff --git a/arm64-asm.c b/arm64-asm.c
index 63c98a21..d34ff214 100644
--- a/arm64-asm.c
+++ b/arm64-asm.c
@@ -86,11 +86,7 @@ enum {
 #define TREG_X30 30
 #define TREG_SP  31
 
-#ifdef TCC_TARGET_PE
-#define ARM64_FREG_BASE 19
-#else
 #define ARM64_FREG_BASE 20
-#endif
 #define ARM64_FREG_LAST (ARM64_FREG_BASE + 7)
 
 typedef struct Operand {
diff --git a/arm64-gen.c b/arm64-gen.c
index 643aaccf..9157a9d2 100644
--- a/arm64-gen.c
+++ b/arm64-gen.c
@@ -579,7 +579,7 @@ ST_FUNC void load(int r, SValue *sv)
     if (svr < VT_CONST) {
         if (IS_FREG(r) && IS_FREG(svr))
             if (svtt == VT_LDOUBLE)
-                o(ARM64_MOV_V16B | fltr(r) | fltr(svr) << 5);
+                o(ARM64_MOV_V16B | fltr(r) | fltr(svr) * 0x10020);
                     // mov v(r).16b,v(svr).16b
             else
                 o(ARM64_FMOV_SCALAR | fltr(r) | fltr(svr) << 5); // fmov d(r),d(svr)
@@ -794,7 +794,7 @@ static int arm64_hfa_aux(CType *type, int *fsize, int num)
             return num;
         }
     }
-    else if ((type->t & VT_ARRAY) && ((type->t & VT_BTYPE) != VT_PTR)) {
+    else if (type->t & VT_ARRAY) { /* handle float array within struct */
         int num1;
         if (!type->ref->c)
             return num;
@@ -811,8 +811,7 @@ static int arm64_hfa_aux(CType *type, int *fsize, int num)
 
 static int arm64_hfa(CType *type, unsigned *fsize)
 {
-    if ((type->t & VT_BTYPE) == VT_STRUCT ||
-        ((type->t & VT_ARRAY) && ((type->t & VT_BTYPE) != VT_PTR))) {
+    if ((type->t & VT_BTYPE) == VT_STRUCT) {
         int sz = 0;
         int n = arm64_hfa_aux(type, &sz, 0);
         if (0 < n && n <= 4) {
@@ -1037,6 +1036,13 @@ static void arm64_sub_sp(uint64_t diff)
     }
 }
 
+static int gv_addr(int r)
+{
+    gaddrof();
+    vtop->type.t = VT_PTR;
+    return gv(r);
+}
+
 ST_FUNC void gfunc_call(int nb_args)
 {
     CType *return_type;
@@ -1147,9 +1153,7 @@ ST_FUNC void gfunc_call(int nb_args)
             else if ((vtop->type.t & VT_BTYPE) == VT_STRUCT) {
                 int align, size = type_size(&vtop->type, &align);
                 if (size) {
-                    vtop->type.t = VT_PTR;
-                    gaddrof();
-                    gv(RC_R(a[i] / 2));
+                    gv_addr(RC_R(a[i] / 2));
                     arm64_ldrs(a[i] / 2, size);
                 }
             }
@@ -1165,9 +1169,7 @@ ST_FUNC void gfunc_call(int nb_args)
                 uint32_t j, sz, n = arm64_hfa(&vtop->type, &sz);
                 if (n > 0) {
                     /* HFA struct - load from memory into float registers */
-                    vtop->type.t = VT_PTR;
-                    gaddrof();
-                    gv(RC_R30);
+                    gv_addr(RC_R30);
                     for (j = 0; j < n; j++)
                         o(0x3d4003c0 |
                           (sz & 16) << 19 | -(sz & 8) << 27 | (sz & 4) << 29 |
@@ -1186,7 +1188,7 @@ ST_FUNC void gfunc_call(int nb_args)
     if ((return_type->t & VT_BTYPE) == VT_STRUCT) {
         if (a[0] == 1) {
             // indirect return: set x8 and discard the stack value
-            gv(RC_R(8));
+            gv_addr(RC_R(8));
             --vtop;
         }
         else
@@ -1206,7 +1208,7 @@ ST_FUNC void gfunc_call(int nb_args)
         int bt = rt & VT_BTYPE;
         if (bt == VT_STRUCT && !(a[0] & 1)) {
             // A struct was returned in registers, so write it out:
-            gv(RC_R(8));
+            gv_addr(RC_R(8));
             --vtop;
             if (a[0] == 0) {
                 int align, size = type_size(return_type, &align);
@@ -1271,6 +1273,7 @@ ST_FUNC void gfunc_prolog(Sym *func_sym)
 
     for (sym = func_type->ref; sym; sym = sym->next)
         ++n;
+
     pcs_n = n - 1;
     c = n + variadic;
     t = tcc_malloc(c * sizeof(*t));
@@ -1321,6 +1324,8 @@ ST_FUNC void gfunc_prolog(Sym *func_sym)
 
     arm64_func_start_offset = ind;
     o(0xa9b27bfd); // stp x29,x30,[sp,#-224]!
+    o(0x910003fd); // mov x29,sp
+
     for (i = 0; i < last_float; i++)
         // stp q0,q1,[sp,#16], stp q2,q3,[sp,#48]
         // stp q4,q5,[sp,#80], stp q6,q7,[sp,#112]
@@ -1368,7 +1373,6 @@ ST_FUNC void gfunc_prolog(Sym *func_sym)
     tcc_free(a);
     tcc_free(t);
 
-    o(0x910003fd); // mov x29,sp
     arm64_func_sub_sp_offset = ind;
     /* In gfunc_epilog these will be replaced with stack setup code. */
     for (i = 0; i < ARM64_FUNC_STACK_SETUP_SLOTS; ++i)
@@ -1384,8 +1388,7 @@ ST_FUNC void gen_va_start(void)
 {
     int r;
     --vtop; // we don't need the "arg"
-    gaddrof();
-    r = intr(gv(RC_INT));
+    r = intr(gv_addr(RC_INT));
 
 #ifdef TCC_TARGET_PE
     if (arm64_func_va_list_stack) {
@@ -1433,13 +1436,12 @@ ST_FUNC void gen_va_arg(CType *t)
     uint32_t r0, r1;
 
 #ifdef TCC_TARGET_PE
-    int indirect = 0, slot = size + 7 & -8;
+    int indirect = 0, slot = (size + 7) & -8;
 
     if (size > 16)
         indirect = 1, slot = 8;
 
-    gaddrof();
-    r0 = intr(gv(RC_INT));
+    r0 = intr(gv_addr(RC_INT));
     r1 = get_reg(RC_INT);
     vtop[0].r = r1 | VT_LVAL;
     r1 = intr(r1);
@@ -1459,7 +1461,6 @@ ST_FUNC void gen_va_arg(CType *t)
         o(0x9100001e | r1 << 5 | slot << 10); // add x30,x(r1),#(slot)
         o(0xf900001e | r0 << 5); // str x30,[x(r0)] // ap += slot
     }
-
     if (indirect)
         o(ARM64_LDR_X | ARM64_RN(r1) | r1); // ldr x(r1),[x(r1)]
 
@@ -1469,8 +1470,7 @@ ST_FUNC void gen_va_arg(CType *t)
     if (!is_float(t->t))
         hfa = arm64_hfa(t, &fsize);
 
-    gaddrof();
-    r0 = intr(gv(RC_INT));
+    r0 = intr(gv_addr(RC_INT));
     r1 = get_reg(RC_INT);
     vtop[0].r = r1 | VT_LVAL;
     r1 = intr(r1);
@@ -1570,8 +1570,7 @@ ST_FUNC void gfunc_return(CType *func_type)
     case 0:
         if ((func_type->t & VT_BTYPE) == VT_STRUCT) {
             int align, size = type_size(func_type, &align);
-            gaddrof();
-            gv(RC_R(0));
+            gv_addr(RC_R(0));
             arm64_ldrs(0, size);
         }
         else
@@ -1590,8 +1589,7 @@ ST_FUNC void gfunc_return(CType *func_type)
         if ((func_type->t & VT_BTYPE) == VT_STRUCT) {
           /* HFA struct return - load from the address on vtop into float registers */
           uint32_t j, sz, n = arm64_hfa(func_type, &sz);
-          gaddrof();
-          gv(RC_R(0));
+          gv_addr(RC_R(0));
           for (j = 0; j < n; j++)
               o(0x3d400000 |
                 (sz & 16) << 19 | -(sz & 8) << 27 | (sz & 4) << 29 |
diff --git a/i386-asm.c b/i386-asm.c
index 64e44ce9..884b18cb 100644
--- a/i386-asm.c
+++ b/i386-asm.c
@@ -494,13 +494,6 @@ ST_FUNC void gen_expr32(ExprValue *pe)
 	gen_addr32(pe->sym ? VT_SYM : 0, pe->sym, pe->v);
 }
 
-#ifdef TCC_TARGET_X86_64
-ST_FUNC void gen_expr64(ExprValue *pe)
-{
-    gen_addr64(pe->sym ? VT_SYM : 0, pe->sym, pe->v);
-}
-#endif
-
 /* XXX: unify with C code output ? */
 static void gen_disp32(ExprValue *pe)
 {
diff --git a/lib/Makefile b/lib/Makefile
index 98126ea4..0afadff5 100644
--- a/lib/Makefile
+++ b/lib/Makefile
@@ -41,7 +41,7 @@ ARM_O = libtcc1.o armeabi.o armflush.o $(COMMON_O)
 ARM64_O = lib-arm64.o $(COMMON_O)
 RISCV64_O = lib-arm64.o $(COMMON_O)
 COMMON_O = stdatomic.o atomic.o builtin.o alloca.o alloca-bt.o
-WIN_O = crt1.o crt1w.o wincrt1.o wincrt1w.o dllcrt1.o dllmain.o
+WIN_O = crt1.o crt1w.o wincrt1.o wincrt1w.o dllcrt1.o dllmain.o winex.o
 LIN_O = dsohandle.o
 OSX_O =
 
@@ -63,7 +63,7 @@ OBJ-x86_64 = $(X86_64_O) va_list.o $(LIN_O)
 OBJ-x86_64-osx = $(X86_64_O) va_list.o $(OSX_O)
 OBJ-i386-win32 = $(I386_O) chkstk.o $(WIN_O)
 OBJ-x86_64-win32 = $(X86_64_O) chkstk.o $(WIN_O)
-OBJ-arm64 = $(ARM64_O) $(LIN_O)
+OBJ-arm64 = $(ARM64_O) armflush.o $(LIN_O)
 OBJ-arm64-osx = $(ARM64_O) $(OSX_O)
 OBJ-arm64-win32 = $(ARM64_O) chkstk.o $(WIN_O)
 OBJ-arm = $(ARM_O) $(LIN_O)
diff --git a/lib/armflush.c b/lib/armflush.c
index c379e436..d3ee99e7 100644
--- a/lib/armflush.c
+++ b/lib/armflush.c
@@ -4,6 +4,9 @@
    intrinsic with gcc.  However tcc in order to compile
    itself needs this function */
 
+/* ------------------------------------------------------------- */
+#if defined __arm__
+
 #ifdef __TINYC__
 
 /* syscall wrapper */
@@ -49,3 +52,13 @@ void __clear_cache(void *beginning, void *end)
  * However, there is no ARM asm parser in tcc so we use it for now */
     syscall(__ARM_NR_cacheflush, beginning, end, 0);
 }
+
+/* ------------------------------------------------------------- */
+#elif defined __aarch64__
+void __clear_cache(void *beg, void *end)
+{
+    __arm64_clear_cache(beg, end);
+}
+
+/* ------------------------------------------------------------- */
+#endif
diff --git a/lib/bt-dll.c b/lib/bt-dll.c
index 7c62cefa..e22f3776 100644
--- a/lib/bt-dll.c
+++ b/lib/bt-dll.c
@@ -34,7 +34,11 @@
   REDIR(__bound_strncmp) \
   REDIR(__bound_strcat) \
   REDIR(__bound_strchr) \
-  REDIR(__bound_strdup)
+  REDIR(__bound_strdup) \
+  REDIR(__bound_strncat) \
+  REDIR(__bound_strrchr) \
+  REDIR(__bound_setjmp) \
+  REDIR(__bound_longjmp)
 
 #ifdef __leading_underscore
 #define _(s) "_"#s
@@ -45,11 +49,28 @@
 #define REDIR(s) void *s;
 static struct { REDIR_ALL } all_ptrs;
 #undef REDIR
+
 #define REDIR(s) #s"\0"
 static const char all_names[] = REDIR_ALL;
 #undef REDIR
-#define REDIR(s) __asm__(".global " _(s) ";" _(s) ": jmp *%0" : : "m" (all_ptrs.s) );
-static void all_jmps() { REDIR_ALL }
+
+#if __aarch64__
+# define REDIR(s) \
+    __asm__(".global "_(s)";"_(s)":"); \
+    __asm__(".int 0x58000090"); /* ldr x16, [pc, #16] */ \
+    __asm__(".int 0xf9400210"); /* ldr x16, [x16] */ \
+    __asm__(".int 0xd61f0200"); /* br x16 */ \
+    __asm__(".int 0xd503201f"); /* nop for alignment */ \
+    __asm__(".quad all_ptrs + (. - all_jmps - 16) / 24 * 8"); \
+    __asm__(".type "_(s)",function\n.size "_(s)",.-"_(s));
+
+    __asm__(".text\n.align 8\nall_jmps:");
+    REDIR_ALL
+#else
+# define REDIR(s) \
+    __asm__(".global "_(s)";"_(s)":"); goto *all_ptrs.s;
+    static void all_jmps() { REDIR_ALL }
+#endif
 #undef REDIR
 
 void __bt_init_dll(int bcheck)
diff --git a/lib/lib-arm64.c b/lib/lib-arm64.c
index 5637358e..3c30f97c 100644
--- a/lib/lib-arm64.c
+++ b/lib/lib-arm64.c
@@ -23,13 +23,6 @@ typedef unsigned long long uint64_t;
 #include <string.h>
 #endif
 
-#if !defined __riscv && !defined __APPLE__
-void __clear_cache(void *beg, void *end)
-{
-    __arm64_clear_cache(beg, end);
-}
-#endif
-
 typedef union {
     struct { uint64_t x0, x1; };
     long double f;
diff --git a/lib/libtcc1.c b/lib/libtcc1.c
index 4bbd1cc7..878f9512 100644
--- a/lib/libtcc1.c
+++ b/lib/libtcc1.c
@@ -107,10 +107,9 @@ union float_long {
 };
 
 /* XXX: we don't support several builtin supports for now */
-#if !defined __x86_64__ && !defined __arm__ && !defined __riscv && !defined __aarch64__
+#if defined __i386__
 
 /* XXX: use gcc/tcc intrinsic ? */
-#if defined __i386__
 #define sub_ddmmss(sh, sl, ah, al, bh, bl) \
   __asm__ ("subl %5,%1\n\tsbbl %3,%0"					\
 	   : "=r" ((USItype) (sh)),					\
@@ -139,9 +138,6 @@ union float_long {
 	     : "=r" (__cbtmp) : "rm" ((USItype) (x)));			\
     (count) = __cbtmp ^ 31;						\
   } while (0)
-#else
-#error unsupported CPU type
-#endif
 
 /* most of this code is taken from libgcc2.c from gcc */
 
@@ -478,7 +474,7 @@ long long __ashldi3(long long a, int b)
 #endif
 }
 
-#endif /* !__x86_64__ */
+#endif /* __i386__ */
 
 /* XXX: fix tcc's code generator to do this instead */
 float __floatundisf(unsigned long long a)
@@ -625,17 +621,3 @@ long long __fixxfdi (long double a1)
     return s ? ret : -ret;
 }
 #endif /* !ARM */
-
-#if defined _WIN64
-/* MSVC x64 intrinsic */
-void __faststorefence(void)
-{
-#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
diff --git a/libtcc.c b/libtcc.c
index ce3a9796..9fd92e2d 100644
--- a/libtcc.c
+++ b/libtcc.c
@@ -124,8 +124,15 @@ static void tcc_add_systemdir(TCCState *s)
     tcc_add_library_path(s, normalize_slashes(buf));
 }
 #endif
+/* for tcc -E : On windows (depending on compiler) a FILE*
+   must be created by the same module where it is used. */
+PUB_FUNC FILE *tcc_fopen(const char *f, const char *m) {
+    return fopen(f, m);
+}
+PUB_FUNC int tcc_fclose(FILE *f) {
+    return fclose(f);
+}
 #endif
-
 /********************************************************/
 
 PUB_FUNC void tcc_enter_state(TCCState *s1)
diff --git a/tcc.c b/tcc.c
index e1819239..e59d8fbb 100644
--- a/tcc.c
+++ b/tcc.c
@@ -293,7 +293,7 @@ int main(int argc, char **argv)
     const char *first_file;
     int argc0 = argc;
     char **argv0 = argv;
-    FILE *ppfp = stdout;
+    FILE *ppfp = NULL;
 
 redo:
     argc = argc0, argv = argv0;
@@ -335,7 +335,7 @@ redo:
             tcc_error_noabort("no input files");
         } else if (s->output_type == TCC_OUTPUT_PREPROCESS) {
             if (s->outfile && 0!=strcmp("-",s->outfile)) {
-                ppfp = fopen(s->outfile, "wb");
+                ppfp = tcc_fopen(s->outfile, "wb");
                 if (!ppfp)
                     tcc_error_noabort("could not write '%s'", s->outfile);
             }
@@ -355,8 +355,8 @@ redo:
     if (s->output_type == 0)
         s->output_type = TCC_OUTPUT_EXE;
     tcc_set_output_type(s, s->output_type);
-    s->ppfp = ppfp;
-
+    if (ppfp)
+        s->ppfp = ppfp;
     if ((s->output_type == TCC_OUTPUT_MEMORY
       || s->output_type == TCC_OUTPUT_PREPROCESS)
         && (s->dflag & 16)) { /* -dt option */
@@ -422,7 +422,7 @@ redo:
     tcc_delete(s);
     if (!done)
         goto redo;
-    if (ppfp && ppfp != stdout)
-        fclose(ppfp);
+    if (ppfp)
+        tcc_fclose(ppfp);
     return ret;
 }
diff --git a/tcc.h b/tcc.h
index 04befa20..27ce1d2b 100644
--- a/tcc.h
+++ b/tcc.h
@@ -52,6 +52,9 @@ extern long double strtold (const char *__nptr, char **__endptr);
 
 #ifdef _WIN32
 # define WIN32_LEAN_AND_MEAN 1
+# ifndef _WIN32_WINNT
+#  define _WIN32_WINNT 0x502 /* AddVectoredExceptionHandler */
+# endif
 # include <windows.h>
 # include <io.h> /* open, close etc. */
 # include <direct.h> /* getcwd */
@@ -86,6 +89,9 @@ extern long double strtold (const char *__nptr, char **__endptr);
 #   define __x86_64__ 1
 #  endif
 # endif
+# if defined(_M_ARM64) && !defined(__aarch64__)
+#  define __aarch64__ 1
+# endif
 # ifndef va_copy
 #  define va_copy(a,b) a = b
 # endif
@@ -228,8 +234,7 @@ extern long double strtold (const char *__nptr, char **__endptr);
 /* No ten-byte long doubles on window and macos except in
    cross-compilers made by a mingw-GCC */
 #if defined TCC_TARGET_PE \
-    || (defined TCC_TARGET_MACHO && defined TCC_TARGET_ARM64) \
-    || (defined _WIN32 && !defined __GNUC__)
+    || (defined TCC_TARGET_MACHO && defined TCC_TARGET_ARM64)
 # define TCC_USING_DOUBLE_FOR_LDOUBLE 1
 #endif
 
@@ -1312,6 +1317,11 @@ PUB_FUNC void tcc_print_stats(TCCState *s, unsigned total_time);
 PUB_FUNC int tcc_parse_args(TCCState *s, int *argc, char ***argv);
 #ifdef _WIN32
 ST_FUNC char *normalize_slashes(char *path);
+PUB_FUNC FILE *tcc_fopen(const char *f, const char *m);
+PUB_FUNC int tcc_fclose(FILE *f);
+#else
+# define tcc_fopen fopen
+# define tcc_fclose fclose
 #endif
 ST_FUNC DLLReference *tcc_add_dllref(TCCState *s1, const char *dllname, int level);
 ST_FUNC char *tcc_load_text(int fd);
@@ -1683,7 +1693,6 @@ ST_FUNC void gen_increment_tcov (SValue *sv);
 
 /* ------------ x86_64-gen.c ------------ */
 #ifdef TCC_TARGET_X86_64
-ST_FUNC void gen_addr64(int r, Sym *sym, int64_t c);
 ST_FUNC void gen_opl(int op);
 #ifdef TCC_TARGET_PE
 ST_FUNC void gen_vla_result(int addr);
@@ -1742,11 +1751,11 @@ ST_FUNC int find_constraint(ASMOperand *operands, int nb_operands, const char *n
 ST_FUNC Sym* get_asm_sym(int name, Sym *csym);
 ST_FUNC void asm_expr(TCCState *s1, ExprValue *pe);
 ST_FUNC int asm_int_expr(TCCState *s1);
-/* ------------ i386-asm.c ------------ */
-ST_FUNC void gen_expr32(ExprValue *pe);
-#ifdef TCC_TARGET_X86_64
+#if PTR_SIZE == 8
 ST_FUNC void gen_expr64(ExprValue *pe);
 #endif
+/* ------------ i386-asm.c ------------ */
+ST_FUNC void gen_expr32(ExprValue *pe);
 ST_FUNC void asm_opcode(TCCState *s1, int opcode);
 ST_FUNC int asm_parse_regvar(int t);
 ST_FUNC void asm_compute_constraints(ASMOperand *operands, int nb_operands, int nb_outputs, const uint8_t *clobber_regs, int *pout_reg);
diff --git a/tccasm.c b/tccasm.c
index 29602317..508712ec 100644
--- a/tccasm.c
+++ b/tccasm.c
@@ -25,6 +25,26 @@
 static Section *last_text_section; /* to handle .previous asm directive */
 static int asmgoto_n;
 
+static int tcc_assemble_internal(TCCState *s1, int do_preprocess, int global);
+static Sym* asm_new_label(TCCState *s1, int label, int is_local);
+static Sym* asm_new_label1(TCCState *s1, int label, int is_local, int sh_num, int value);
+
+#if PTR_SIZE == 8
+/* output constant with relocation if 'r & VT_SYM' is true */
+ST_FUNC void gen_addr64(int r, Sym *sym, int64_t c)
+{
+    if (r & VT_SYM)
+        greloca(cur_text_section, sym, ind, R_DATA_PTR, c), c=0;
+    gen_le32(c);
+    gen_le32(c>>32);
+}
+
+ST_FUNC void gen_expr64(ExprValue *pe)
+{
+    gen_addr64(pe->sym ? VT_SYM : 0, pe->sym, pe->v);
+}
+#endif
+
 static int asm_get_prefix_name(TCCState *s1, const char *prefix, unsigned int n)
 {
     char buf[64];
@@ -37,10 +57,6 @@ ST_FUNC int asm_get_local_label_name(TCCState *s1, unsigned int n)
     return asm_get_prefix_name(s1, "L..", n);
 }
 
-static int tcc_assemble_internal(TCCState *s1, int do_preprocess, int global);
-static Sym* asm_new_label(TCCState *s1, int label, int is_local);
-static Sym* asm_new_label1(TCCState *s1, int label, int is_local, int sh_num, int value);
-
 /* If a C name has an _ prepended then only asm labels that start
    with _ are representable in C, by removing the first _.  ASM names
    without _ at the beginning don't correspond to C names, but we use
@@ -328,12 +344,12 @@ static inline void asm_expr_sum(TCCState *s1, ExprValue *pe)
 		if (esym1 && esym1->st_shndx == esym2->st_shndx
 		    && esym1->st_shndx != SHN_UNDEF) {
 		    /* we also accept defined symbols in the same section */
-		    pe->v += esym1->st_value - esym2->st_value;
+		    pe->v += (int)(esym1->st_value - esym2->st_value);
 		    pe->sym = NULL;
 		} else if (esym2->st_shndx == cur_text_section->sh_num) {
 		    /* When subtracting a defined symbol in current section
 		       this actually makes the value PC-relative.  */
-		    pe->v += 0 - esym2->st_value;
+		    pe->v += (int)(0 - esym2->st_value);
 		    pe->pcrel = 1;
 		    e2.sym = NULL;
 		} else {
@@ -543,7 +559,7 @@ static void asm_parse_directive(TCCState *s1, int global)
         ind += size;
         break;
     case TOK_ASMDIR_quad:
-#ifdef TCC_TARGET_X86_64
+#if PTR_SIZE == 8
 	size = 8;
 	goto asm_data;
 #else
@@ -592,7 +608,7 @@ static void asm_parse_directive(TCCState *s1, int global)
             if (sec->sh_type != SHT_NOBITS) {
                 if (size == 4) {
                     gen_expr32(&e);
-#ifdef TCC_TARGET_X86_64
+#if PTR_SIZE == 8
 		} else if (size == 8) {
 		    gen_expr64(&e);
 #endif
@@ -812,6 +828,7 @@ static void asm_parse_directive(TCCState *s1, int global)
     case TOK_ASMDIR_size:
         { 
             Sym *sym;
+            ElfSym *esym;
 
             next();
 	    if (tok < TOK_IDENT)
@@ -823,8 +840,10 @@ static void asm_parse_directive(TCCState *s1, int global)
             tcc_warning_c(warn_unsupported)("ignoring .size %s,*", get_tok_str(tok, NULL));
             next();
             skip(',');
-            while (tok != TOK_LINEFEED && tok != ';' && tok != CH_EOF) {
-                next();
+            n = asm_int_expr(s1);
+            esym = elfsym(sym);
+            if (esym) {
+                esym->st_size = n;
             }
         }
         break;
@@ -952,7 +971,7 @@ static void asm_parse_directive(TCCState *s1, int global)
         }
         break;
 #endif
-#ifdef TCC_TARGET_X86_64
+#if PTR_SIZE == 8
     /* added for compatibility with GAS */
     case TOK_ASMDIR_code64:
         next();
diff --git a/tcclib.h b/tcclib.h
index 8d59e4c9..b99cf69a 100644
--- a/tcclib.h
+++ b/tcclib.h
@@ -19,6 +19,7 @@ int atoi(const char *nptr);
 long int strtol(const char *nptr, char **endptr, int base);
 unsigned long int strtoul(const char *nptr, char **endptr, int base);
 void exit(int);
+void *alloca(size_t);
 
 /* stdio.h */
 typedef struct __FILE FILE;
@@ -39,6 +40,7 @@ int getchar(void);
 char *gets(char *s);
 int ungetc(int c, FILE *stream);
 int fflush(FILE *stream);
+int puts(const char *s);
 int putchar (int c);
 
 int printf(const char *format, ...);
diff --git a/tccpe.c b/tccpe.c
index 11817e33..918ba47f 100644
--- a/tccpe.c
+++ b/tccpe.c
@@ -1958,6 +1958,7 @@ ST_FUNC void pe_add_unwind_data(unsigned start, unsigned end, unsigned stack)
     for (n = o + sizeof *p; o < n; o += sizeof p->BeginAddress)
         put_elf_reloc(symtab_section, pd, o, R_XXX_RELATIVE, s1->uw_sym);
 }
+
 #elif defined(TCC_TARGET_ARM64)
 /* ARM64 unwind codes:
    save_fplr_x: 10iiiiii  - stp x29,lr,[sp,#-(i+1)*8]!
@@ -1975,10 +1976,7 @@ static Section *pe_add_unwind_info(TCCState *s1)
         s1->uw_pdata->sh_addralign = 4;
     }
     s = find_section(s1, ".xdata");
-    if (NULL == s) {
-        s = new_section(s1, ".xdata", SHT_PROGBITS, SHF_ALLOC);
-        s->sh_addralign = 4;
-    }
+    s->sh_addralign = 4;
     if (0 == s1->uw_sym)
         s1->uw_sym = put_elf_sym(symtab_section, 0, 0, 0, 0,
                                   text_section->sh_num, ".uw_text_base");
@@ -1992,31 +1990,22 @@ ST_FUNC void pe_add_unwind_data(unsigned start, unsigned end, unsigned stack)
 {
     TCCState *s1 = tcc_state;
     Section *pd, *xd;
-    unsigned o, n, d, code_bytes, func_len, stack_slots;
+    unsigned o, d, code_bytes, func_len;
     unsigned char *q;
     uint32_t header;
-    struct {
+    struct /* _RUNTIME_FUNCTION */ {
         DWORD BeginAddress;
-        DWORD EndAddress;
         DWORD UnwindData;
     } *p;
 
+    int epilog;
+
     xd = pe_add_unwind_info(s1);
     pd = s1->uw_pdata;
 
-    stack = (stack + 15) & ~15;
-    stack_slots = stack >> 4;
     func_len = (end - start) >> 2;
     code_bytes = 0;
-    if (stack_slots) {
-        if (stack_slots <= 31) {
-            code_bytes += 1;
-        } else if (stack_slots <= 0x7ff) {
-            code_bytes += 2;
-        } else {
-            code_bytes += 4;
-        }
-    }
+    epilog = code_bytes;
     code_bytes += 3; /* set_fp, save_fplr_x, end */
     code_bytes = (code_bytes + 3) & ~3;
 
@@ -2025,23 +2014,13 @@ ST_FUNC void pe_add_unwind_data(unsigned start, unsigned end, unsigned stack)
     q = section_ptr_add(xd, 4 + code_bytes);
 
     /* Full ARM64 xdata header: E=1 with one epilog and no exception handler. */
-    header = (func_len & 0x3ffff) | (1u << 21) | ((code_bytes >> 2) << 27);
+    header = (func_len & 0x3ffff)
+        | 1 << 21
+        | (epilog & 0x1F) << 22
+        | (code_bytes >> 2) << 27
+        ;
     write32le(q, header);
     q += 4;
-
-    if (stack_slots) {
-        if (stack_slots <= 31) {
-            *q++ = stack_slots; /* alloc_s */
-        } else if (stack_slots <= 0x7ff) {
-            *q++ = 0xC0 | (stack_slots >> 8); /* alloc_m */
-            *q++ = stack_slots & 0xff;
-        } else {
-            *q++ = 0xE0; /* alloc_l */
-            *q++ = (stack_slots >> 16) & 0xff;
-            *q++ = (stack_slots >> 8) & 0xff;
-            *q++ = stack_slots & 0xff;
-        }
-    }
     *q++ = 0xE1; /* set_fp */
     *q++ = 0x9B; /* save_fplr_x: stp x29,lr,[sp,#-224]! */
     *q++ = 0xE4; /* end */
@@ -2050,14 +2029,10 @@ ST_FUNC void pe_add_unwind_data(unsigned start, unsigned end, unsigned stack)
 
     o = pd->data_offset;
     p = section_ptr_add(pd, sizeof *p);
-
     p->BeginAddress = start;
-    p->EndAddress = end;
     p->UnwindData = d;
-
-    for (n = o + 2 * sizeof p->BeginAddress; o < n; o += sizeof p->BeginAddress)
-        put_elf_reloc(symtab_section, pd, o, R_XXX_RELATIVE, s1->uw_sym);
-    put_elf_reloc(symtab_section, pd, n, R_XXX_RELATIVE, s1->uw_xsym);
+    put_elf_reloc(symtab_section, pd, o, R_XXX_RELATIVE, s1->uw_sym);
+    put_elf_reloc(symtab_section, pd, o + 4, R_XXX_RELATIVE, s1->uw_xsym);
 }
 #endif
 /* ------------------------------------------------------------- */
@@ -2252,20 +2227,22 @@ ST_FUNC int pe_output_file(TCCState *s1, const char *filename)
     resolve_common_syms(s1);
     pe_set_options(s1, &pe);
     pe_check_symbols(&pe);
-
     if (s1->nb_errors)
-        ;
-    else if (filename) {
+        goto done;
+    if (filename) {
         pe_assign_addresses(&pe);
         relocate_syms(s1, s1->symtab, 0);
+        if (s1->nb_errors)
+            goto done;
         s1->pe_imagebase = pe.imagebase;
         relocate_sections(s1);
         pe.start_addr = (DWORD)
             (get_sym_addr(s1, pe.start_symbol, 1, 1) - pe.imagebase);
-        if (0 == s1->nb_errors)
-            pe_write(&pe);
-        dynarray_reset(&pe.sec_info, &pe.sec_count);
+        if (s1->nb_errors)
+            goto done;
+        pe_write(&pe);
     } else {
+        /* -run */
 #ifdef TCC_IS_NATIVE
         pe.thunk = data_section;
         pe_build_imports(&pe);
@@ -2275,6 +2252,8 @@ ST_FUNC int pe_output_file(TCCState *s1, const char *filename)
 #endif
 #endif
     }
+done:
+    dynarray_reset(&pe.sec_info, &pe.sec_count);
     pe_free_imports(&pe);
 #if PE_PRINT_SECTIONS
     if (g_debug & 8)
diff --git a/tccrun.c b/tccrun.c
index efec3228..8c9a6126 100644
--- a/tccrun.c
+++ b/tccrun.c
@@ -434,6 +434,8 @@ redo:
 
     /* relocate symbols */
     relocate_syms(s1, s1->symtab, 1);
+    if (s1->nb_errors)
+        goto redo;
     /* relocate sections */
 #ifdef TCC_TARGET_PE
     s1->pe_imagebase = mem;
diff --git a/tcctok.h b/tcctok.h
index 2abbe090..ec44b244 100644
--- a/tcctok.h
+++ b/tcctok.h
@@ -402,12 +402,13 @@
  DEF_ASMDIR(endr)
  DEF_ASMDIR(org)
  DEF_ASMDIR(quad)
-#if defined(TCC_TARGET_I386)
+#if PTR_SIZE == 4
  DEF_ASMDIR(code16)
  DEF_ASMDIR(code32)
-#elif defined(TCC_TARGET_X86_64)
+#else
  DEF_ASMDIR(code64)
-#elif defined(TCC_TARGET_RISCV64)
+#endif
+#if defined(TCC_TARGET_RISCV64)
  DEF_ASMDIR(option)
 #endif
  DEF_ASMDIR(short)
diff --git a/tests/Makefile b/tests/Makefile
index 410bbc8e..adb5733c 100644
--- a/tests/Makefile
+++ b/tests/Makefile
@@ -72,9 +72,9 @@ endif
 RUN_TCC = -run $(TOPSRC)/tcc.c $(TCCFLAGS)
 DISAS = objdump -d
 ifdef CONFIG_OSX
-DUMPTCC = (set -x; $(TOP)/tcc -vv; otool -L $(TOP)/tcc; exit 1)
+DUMPTCC = (set -x; $(TCC_LOCAL) -vv; otool -L $(TCC_LOCAL); exit 1)
 else
-DUMPTCC = (set -x; $(TOP)/tcc -vv; ldd $(TOP)/tcc; exit 1)
+DUMPTCC = (set -x; $(TCC_LOCAL) -vv; ldd $(TCC_LOCAL); exit 1)
 endif
 
 all test :
@@ -82,6 +82,7 @@ all test :
 	@$(TCC_LOCAL) -v
 	@$(MAKE) --no-print-directory -s clean
 	@$(MAKE) --no-print-directory -s -r _all
+	@echo ------- ALL TESTS PASSED --------
 
 _all : $(TESTS)
 
@@ -109,8 +110,8 @@ libtcc_test_mt$(EXESUF): libtcc_test_mt.c
 
 # test.ref - generate using cc
 test.ref: tcctest.c
-	$(CC) -o tcctest.gcc $< $(CFLAGS) -w -O0 -std=gnu99 -fno-omit-frame-pointer
-	./tcctest.gcc > $@
+	$(CC) -o tcctest.gcc$(EXESUF) $< $(CFLAGS) -w -O0 -std=gnu99 -fno-omit-frame-pointer
+	./tcctest.gcc$(EXESUF) > $@
 
 # auto test
 test1 test1b: tcctest.c test.ref
diff --git a/tests/libtcc_test_mt.c b/tests/libtcc_test_mt.c
index 23764637..396e4e25 100644
--- a/tests/libtcc_test_mt.c
+++ b/tests/libtcc_test_mt.c
@@ -162,6 +162,7 @@ void *reloc_state(TCCState *s, const char *entry)
 {
     void *func;
     tcc_add_symbol(s, "add", add);
+    tcc_add_symbol(s, "printf", printf);
     if (tcc_relocate(s) < 0) {
         fprintf(stderr, __FILE__ ": could not relocate tcc state.\n");
         return NULL;
@@ -337,16 +338,15 @@ int main(int argc, char **argv)
 
 #else
 #include <tcclib.h>
-
 #ifdef _WIN32
-# ifdef __i386__
-#  define LIBTCC_TEST_WINAPI __attribute__((__stdcall__))
-# else
-#  define LIBTCC_TEST_WINAPI
+# ifndef _WIN64
+    __declspec(stdcall)
 # endif
-void LIBTCC_TEST_WINAPI Sleep(unsigned int milliseconds);
+    void Sleep(unsigned);
+# define sleep_ms Sleep
 #else
-unsigned int sleep(unsigned int seconds);
+    int usleep(unsigned long);
+# define sleep_ms(x) usleep((x)*1000);
 #endif
 
 int fib(n)
@@ -356,11 +356,7 @@ int fib(n)
 
 int main(int argc, char **argv)
 {
-#ifdef _WIN32
-    Sleep(1000);
-#else
-    sleep(1);
-#endif
+    sleep_ms(333);
     printf(" %d", fib(atoi(argv[1])));
     return 0;
 }
diff --git a/tests/msvcrt_start.c b/tests/msvcrt_start.c
new file mode 100644
index 00000000..5ca2c5ac
--- /dev/null
+++ b/tests/msvcrt_start.c
@@ -0,0 +1,164 @@
+/* ------------------------------------------------------------- */
+/* minimal startup with runtime linker to msvcrt */
+
+#if 0
+
+#define REDIR_ALL \
+ REDIR(__set_app_type)\
+ REDIR(__getmainargs)\
+ REDIR(_controlfp)\
+ REDIR(_vsnprintf)\
+ REDIR(exit)\
+ \
+ REDIR(puts)\
+ REDIR(printf)\
+ REDIR(putchar)\
+ REDIR(strtod)\
+ REDIR(memset)\
+ REDIR(strcpy)\
+ REDIR(strlen)\
+ REDIR(malloc)\
+ REDIR(free)\
+
+#if defined __i386__ && !defined __TINYC__
+# define __leading_underscore 1
+#endif
+
+#ifdef __leading_underscore
+# define _(s) "_"#s
+#else
+# define _(s) #s
+#endif
+
+#define REDIR(s) void *s;
+static struct { REDIR_ALL } all_ptrs;
+#undef REDIR
+
+#define REDIR(s) #s"\0"
+static const char all_names[] = REDIR_ALL;
+#undef REDIR
+
+#if __aarch64__
+  #if defined __TINYC__
+  # define ALIGN ".align 8"
+  #else
+  # define ALIGN ".align 3" /* .align is power of 2 on non-ELF platforms */
+  #endif
+# define REDIR(s) \
+    __asm__("\n"_(s)":"); \
+    __asm__(".int 0x58000090"); /* ldr x16, [pc, #16] */ \
+    __asm__(".int 0xf9400210"); /* ldr x16, [x16] */ \
+    __asm__(".int 0xd61f0200"); /* br x16 */ \
+    __asm__(".int 0xd503201f"); /* nop for alignment */ \
+    __asm__(".quad all_ptrs + (. - all_jmps - 16) / 24 * 8"); \
+    __asm__(".global "_(s));
+
+    __asm__("\t.text\n\t"ALIGN"\nall_jmps:");
+    REDIR_ALL
+#else
+# define REDIR(s) \
+    __asm__("\n"_(s)":");\
+    __asm__("jmp *%0"::"m"(all_ptrs.s));\
+    __asm__(".global "_(s));
+
+    static void all_jmps() { REDIR_ALL }
+#endif
+#undef REDIR
+
+#if 0
+# include <windows.h>
+#else
+# if __i386__
+#  define STDCALL __declspec(stdcall)
+# else
+#  define STDCALL
+# endif
+# define DWORD long unsigned
+# define HMODULE void*
+# define HANDLE void*
+S TDCALL HMODULE LoadLibraryA(const char *);
+S TDCALL HMODULE GetProcAddress(HMODULE , char*);
+S TDCALL void ExitProcess(int);
+S TDCALL int WriteFile(HANDLE, const void*, DWORD, DWORD*, void*);
+S TDCALL HANDLE GetStdHandle(DWORD);
+S TDCALL int FlushFileBuffers(HANDLE);
+# define STD_ERROR_HANDLE -12
+#endif
+
+static void eput(const char *s)
+{
+    DWORD n_out;
+    int n = 0;
+    while (s[n])
+        ++n;
+    WriteFile(GetStdHandle(STD_ERROR_HANDLE), s, n, &n_out, 0);
+}
+
+static void rt_reloc()
+{
+    const char *s = all_names;
+    void **p = (void**)&all_ptrs;
+    void *dll = LoadLibraryA("msvcrt.dll");
+    do {
+        char buf[100], *d = buf;
+        *p = (void*)GetProcAddress(dll, (char*)s);
+        *d++ = '_'; do *d++ = *s; while (*s++);
+        if (0 == *p)
+            *p = (void*)GetProcAddress(dll, buf);
+        if (0 == *p) {
+            eput("MSVCRT_START.C: RUNTIME RELOCATION ERROR: '");
+            eput(buf+1);
+            eput("'\n");
+            ExitProcess(-1);
+        }
+        ++p;
+    } while (*s);
+}
+
+#else
+# define rt_reloc()
+#endif
+
+int main(int argc, char **argv, char **env);
+void exit(int);
+void __set_app_type(int apptype);
+
+typedef struct { int newmode; } _startupinfo;
+int __getmainargs(int *pargc, char ***pargv, char ***penv, int globb, _startupinfo*);
+
+void _controlfp(unsigned a, unsigned b);
+#define _MCW_PC 0x00030000 // Precision Control
+#define _PC_53 0x00010000 // 53 bits
+
+int __argc;
+char **__argv;
+char **environ;
+_startupinfo start_info = {0};
+
+void mainCRTStartup(void)
+{
+    rt_reloc();
+#if defined __i386__ || defined __x86_64__
+    _controlfp(_PC_53, _MCW_PC);
+#endif
+    __set_app_type(1);
+    __getmainargs(&__argc, &__argv, &environ, 0, &start_info);
+    exit(main(__argc, __argv, environ));
+}
+
+#include <stdarg.h>
+#define size_t __SIZE_TYPE__
+int printf(const char *, ...);
+int _vsnprintf(char *, size_t, const char *, va_list);
+
+/* undefined on windows-11-arm64 */
+int vprintf(const char *format, va_list ap)
+{
+    char buf[1000];
+    _vsnprintf(buf, sizeof buf, format, ap);
+    return printf("%s", buf);
+}
+
+void __main() {} /* for gcc */
+void _pei386_runtime_relocator(void) {} /* for gcc */
+void __chkstk(unsigned n) {} /* for clang */
diff --git a/tests/tcctest.c b/tests/tcctest.c
index 49959417..0bb392e4 100644
--- a/tests/tcctest.c
+++ b/tests/tcctest.c
@@ -81,9 +81,6 @@ typedef __SIZE_TYPE__ uintptr_t;
 #include incname
 #include stringify(funnyname)
 
-int puts(const char *s);
-void *alloca(size_t size);
-
 int fib(int n);
 void num(int n);
 void forward_ref(void);
@@ -287,6 +284,7 @@ comment
 
     printf("basefromheader %s\n", get_basefile_from_header());
     printf("base %s\n", __BASE_FILE__);
+#if !(defined _WIN32 && CC_NAME == CC_clang)
     {
       /* Some compilers (clang) prepend './' to __FILE__ from included
          files.  */
@@ -295,6 +293,8 @@ comment
         fn += 2;
       printf("filefromheader %s\n", fn);
     }
+#endif
+
     printf("file %s\n", __FILE__);
 
     /* Check that funnily named include was in fact included */
@@ -1095,8 +1095,10 @@ void struct_test()
            sizeof(struct aligntest2), __alignof__(struct aligntest2));
     printf("aligntest3 sizeof=%d alignof=%d\n",
            sizeof(struct aligntest3), __alignof__(struct aligntest3));
+#if !(defined _WIN32 && CC_NAME == CC_clang)
     printf("aligntest4 sizeof=%d alignof=%d\n",
            sizeof(struct aligntest4), __alignof__(struct aligntest4));
+#endif
     printf("aligntest5 sizeof=%d alignof=%d\n",
            sizeof(struct aligntest5), __alignof__(struct aligntest5));
     printf("aligntest6 sizeof=%d alignof=%d\n",
@@ -1105,8 +1107,10 @@ void struct_test()
            sizeof(struct aligntest7), __alignof__(struct aligntest7));
     printf("aligntest8 sizeof=%d alignof=%d\n",
            sizeof(struct aligntest8), __alignof__(struct aligntest8));
+#if !(defined _WIN32 && CC_NAME == CC_clang)
     printf("aligntest9 sizeof=%d alignof=%d\n",
            sizeof(struct aligntest9), __alignof__(struct aligntest9));
+#endif
     printf("aligntest10 sizeof=%d alignof=%d\n",
            sizeof(struct aligntest10), __alignof__(struct aligntest10));
     printf("altest5 sizeof=%d alignof=%d\n",
@@ -1117,7 +1121,9 @@ void struct_test()
            sizeof(altest7), __alignof__(altest7));
            
     /* empty structures (GCC extension) */
+#if !(defined _WIN32 && CC_NAME == CC_clang)
     printf("sizeof(struct empty) = %d\n", sizeof(struct empty));
+#endif
     printf("alignof(struct empty) = %d\n", __alignof__(struct empty));
 
     printf("Large: sizeof=%d\n", sizeof(ls));
@@ -2176,15 +2182,6 @@ float strtof(const char *nptr, char **endptr);
 LONG_DOUBLE strtold(const char *nptr, char **endptr);
 #endif
 
-#if CC_NAME == CC_clang
-/* In clang 0.0/0.0 is nan and not -nan.
-   Also some older clang version do v=-v
-   as v = -0 - v */
-static char enable_nan_test = 0;
-#else
-static char enable_nan_test = 1;
-#endif
-
 #define FTEST(prefix, typename, type, fmt)\
 void prefix ## cmp(type a, type b)\
 {\
@@ -2240,7 +2237,7 @@ void prefix ## fcast(type a)\
     b = llia;\
     printf("lltof: " fmt "\n", b);\
     b = llua;\
-    printf("ulltof: " fmt "\n", b);\
+    if (CC_NAME != CC_clang) printf("ulltof: " fmt "\n", b);\
 }\
 \
 float prefix ## retf(type a) { return a; }\
@@ -2300,7 +2297,7 @@ void prefix ## test(void)\
     prefix ## fcast(-2334.6);\
     prefix ## call();\
     prefix ## signed_zeros();\
-    if (enable_nan_test) prefix ## nan();\
+    if (CC_NAME != CC_clang) prefix ## nan();\
 }
 
 FTEST(f, float, float, "%f")
@@ -2556,8 +2553,8 @@ void longlong_test(void)
     a = ia;
     b = ua;
     printf(LONG_LONG_FORMAT " " LONG_LONG_FORMAT "\n", a, b);
-    printf(LONG_LONG_FORMAT " " LONG_LONG_FORMAT " " LONG_LONG_FORMAT " %Lx\n", 
-           (long long)1, 
+    printf(LONG_LONG_FORMAT " " LONG_LONG_FORMAT " " LONG_LONG_FORMAT " "XLONG_LONG_FORMAT"\n",
+           (long long)1,
            (long long)-2,
            1LL,
            0x1234567812345679);
@@ -2867,12 +2864,14 @@ void stdarg_test(void)
     stdarg_for_struct(bob, bob2, bob3, bob4, bob, bob, bob.profile);
     stdarg_for_libc("stdarg_for_libc: %s %.2f %d\n", "string", 1.23, 456);
     stdarg_syntax(1, 17);
+#if !(defined _WIN32 && CC_NAME == CC_clang) /* broken clang */
     stdarg_double_struct(6,-1,pts[0],pts[1],pts[2],pts[3],pts[4],pts[5]);
     stdarg_double_struct(7,1,pts[0],-1.0,pts[1],pts[2],pts[3],pts[4],pts[5]);
     stdarg_double_struct(7,2,pts[0],pts[1],-1.0,pts[2],pts[3],pts[4],pts[5]);
     stdarg_double_struct(7,3,pts[0],pts[1],pts[2],-1.0,pts[3],pts[4],pts[5]);
     stdarg_double_struct(7,4,pts[0],pts[1],pts[2],pts[3],-1.0,pts[4],pts[5]);
     stdarg_double_struct(7,5,pts[0],pts[1],pts[2],pts[3],pts[4],-1.0,pts[5]);
+#endif
 }
 
 int reltab[3] = { 1, 2, 3 };
@@ -3289,7 +3288,7 @@ void local_label_test(void)
 }
 
 /* inline assembler test */
-#if defined(__i386__) || defined(__x86_64__)
+#if (defined(__i386__) || defined(__x86_64__)) && !(defined _WIN32 && CC_NAME == CC_clang)
 
 typedef __SIZE_TYPE__ word;
 
@@ -3551,8 +3550,8 @@ void asm_local_label_diff (void)
 {
   printf ("asm_local_label_diff: %d %d\n", alld_stuff[0], alld_stuff[1]);
 }
-#endif
-#endif
+#endif //!__APPLE__
+#endif //!_WIN32
 
 /* This checks that static local variables are available from assembler.  */
 void asm_local_statics (void)
diff --git a/tests/test-win32.bat b/tests/test-win32.bat
new file mode 100644
index 00000000..bf593cbf
--- /dev/null
+++ b/tests/test-win32.bat
@@ -0,0 +1,74 @@
+@echo off
+setlocal
+set CC=gcc
+set TESTS=
+if "%1"=="/?" goto :usage
+goto :p0
+
+:usage
+echo usage: test-win32.bat [options...] [tests...]
+echo options:
+echo   -c compiler  reference compiler (gcc or clang)
+echo   -p path      prepend path to PATH
+echo tests          tests to run (default all)
+echo requires:      make, gcc/clang, and sh in the PATH
+exit /B 1
+
+:p2
+shift
+:p1
+shift
+:p0
+if "%1"=="-c" set CC=%2&&goto p2
+if "%1"=="-p" set PATH=%2;%PATH%&&goto p2
+if not "%1"=="" set TESTS=%TESTS% %1&&goto p1
+if "%TESTS%"=="" set TESTS=all -k
+
+set PATH=%CD%\..\win32;%PATH%
+
+for /f "delims=-" %%a in ('tcc.exe -dumpmachine') do set ARCH=%%a
+set ARCH=%ARCH:aarch=arm%
+set MACH=%ARCH:i386=x86%
+set MACH=%MACH:x86_64=amd64%
+rem echo ARCH:%ARCH%  MACHINE:%MACH%
+
+set CRTLIB=c:/windows/system32/msvcrt.dll
+set LIBTCC=libtcc.dll
+set LGCC=-lgcc
+if %CC%==gcc goto :c2
+
+set CRTLIB=msvcrt.lib
+set LIBTCC=libtcc.lib
+set LGCC=
+if exist %CRTLIB% goto :c3
+tcc -impdef msvcrt.dll
+lib >nul /def:msvcrt.def /out:%CRTLIB% /machine:%MACH%
+:c2
+if not exist %CRTLIB% echo test-win32.bat: error: %CRTLIB% not found&&exit /b 1
+:c3
+
+set REF_LINK=-nostdlib msvcrt_start.c %LGCC% -lkernel32 %CRTLIB%
+set CFG_MAK=..\config.mak
+set CFG_H=..\config.h
+
+echo>>%CFG_H% #define CC_NAME CC_%CC%
+echo>>%CFG_H% #define GCC_MAJOR 15
+
+if exist %CFG_MAK% del %CFG_MAK%
+echo>>%CFG_MAK% CC = %CC%.exe
+echo>>%CFG_MAK% CC_NAME = %CC%
+echo>>%CFG_MAK% ARCH = %ARCH%
+echo>>%CFG_MAK% CFLAGS = -Wall -O0
+echo>>%CFG_MAK% LDFLAGS =
+echo>>%CFG_MAK% LIBSUF = .lib
+echo>>%CFG_MAK% prefix = $(TOP)/bin
+echo>>%CFG_MAK% EXESUF = .exe
+echo>>%CFG_MAK% DLLSUF = .dll
+echo>>%CFG_MAK% TARGETOS = WIN32
+echo>>%CFG_MAK% CONFIG_WIN32 = yes
+echo>>%CFG_MAK% TOPSRC = $(TOP)
+echo>>%CFG_MAK% SHELL = sh
+echo>>%CFG_MAK% test.ref: CFLAGS += %REF_LINK%
+
+set GMAKE=make
+%GMAKE% TCC_LOCAL=tcc.exe LIBTCC=win32/%LIBTCC% %TESTS%
diff --git a/tests/tests2/Makefile b/tests/tests2/Makefile
index b01009c1..620ed77e 100644
--- a/tests/tests2/Makefile
+++ b/tests/tests2/Makefile
@@ -48,26 +48,16 @@ ifeq (-$(CONFIG_WIN32)-,-yes-)
  SKIP += 117_builtins.test # win32 port doesn't define __builtins
  SKIP += 124_atomic_counter.test # No pthread support
 endif
-ifneq (,$(filter arm% riscv%,$(ARCH)))
- SKIP += 85_asm-outside-function.test
- SKIP += 98_al_ax_extend.test
- SKIP += 99_fastcall.test
- SKIP += 127_asm_goto.test
-endif
-ifeq (,$(filter arm64 aarch64,$(ARCH)))
- SKIP += 138_arm64_encoding.test
- SKIP += 139_arm64_errors.test
-endif
-ifneq (,$(findstring win32,$(CROSS_TARGET)))
- SKIP += 106_versym.test
- SKIP += 114_bound_signal.test
- SKIP += 124_atomic_counter.test
-endif
 ifneq (,$(filter OpenBSD FreeBSD NetBSD,$(TARGETOS)))
  SKIP += 106_versym.test # no pthread_condattr_setpshared
  SKIP += 114_bound_signal.test # libc problem signal/fork
  SKIP += 116_bound_setjmp2.test # No TLS_FUNC/TLS_VAR in bcheck.c
 endif
+ifeq (,$(filter arm64 aarch64,$(ARCH)))
+ SKIP += 138_arm64_encoding.test
+ SKIP += 139_arm64_errors.test
+ SKIP += 140_arm64_extasm.test
+endif
 
 # Some tests might need arguments
 ARGS =
@@ -88,8 +78,8 @@ endif
 
 # These tests run several snippets from the same file one by one
 60_errors_and_warnings.test : FLAGS += -dt
-139_arm64_errors.test : FLAGS += -dt
 96_nodata_wanted.test : FLAGS += -dt
+139_arm64_errors.test : FLAGS += -dt
 
 # Always generate certain .expects (don't put these in the GIT),
 GEN-ALWAYS =
@@ -141,6 +131,7 @@ endif
 126_bound_global.test: NORUN = true
 128_run_atexit.test: FLAGS += -dt
 132_bound_test.test: FLAGS += -b
+140_arm64_extasm.test: GEN = $(GEN-TCC)
 
 # Filter source directory in warnings/errors (out-of-tree builds)
 FILTER = 2>&1 | sed -e 's,$(SRC)/,,g'
diff --git a/win32/build-tcc.bat b/win32/build-tcc.bat
index e5b576bc..16f69e17 100644
--- a/win32/build-tcc.bat
+++ b/win32/build-tcc.bat
@@ -5,13 +5,15 @@
 @echo off
 setlocal
 if (%1)==(-clean) goto :cleanup
-set CC=gcc
+set CC=gcc -O2 -Wall
 set /p VERSION= < ..\VERSION
 set TCCDIR=
 set BINDIR=
 set DOC=no
-set XCC=no
+set TX=
+set SELF=%~nx0
 goto :a0
+
 :a2
 shift
 :a3
@@ -27,20 +29,22 @@ if (%1)==(-v) set VERSION=%~2&& goto :a2
 if (%1)==(-i) set TCCDIR=%2&& goto :a2
 if (%1)==(-b) set BINDIR=%2&& goto :a2
 if (%1)==(-d) set DOC=yes&& goto :a3
-if (%1)==(-x) set XCC=yes&& goto :a3
+if (%1)==(-x) set TX=%2&& goto :a2
 if (%1)==() goto :p1
+
 :usage
 echo usage: build-tcc.bat [ options ... ]
 echo options:
 echo   -c prog              use prog (gcc/tcc/cl) to compile tcc
 echo   -c "prog options"    use prog with options to compile tcc
-echo   -t 32/64             force 32/64 bit default target
+echo   -t target            set target
+echo   -x target            build tcc cross-compiler for target
 echo   -v "version"         set tcc version
 echo   -i tccdir            install tcc into tccdir
 echo   -b bindir            but install tcc.exe and libtcc.dll into bindir
 echo   -d                   create tcc-doc.html too (needs makeinfo)
-echo   -x                   build the cross compiler too
 echo   -clean               delete all previously produced files and directories
+echo   supported targets    i386 x86_64 arm64
 exit /B 1
 
 @rem ------------------------------------------------------
@@ -65,6 +69,7 @@ exit /B 0
 if exist %1 rmdir /Q/S %1 && %LOG%   %1
 exit /B 0
 
+@rem ------------------------------------------------------
 :cl
 @echo off
 set CMD=cl
@@ -84,36 +89,19 @@ echo on
 @rem main program
 
 :p1
-if not %T%_==_ goto :p2
-set T=32
-if %PROCESSOR_ARCHITECTURE%_==AMD64_ set T=64
-if %PROCESSOR_ARCHITEW6432%_==AMD64_ set T=64
-:p2
-if "%CC:~-3%"=="gcc" set CC=%CC% -O2 -s -static
-if (%BINDIR%)==() set BINDIR=%TCCDIR%
+if not _%TX%_==__ set T=%TX%&&set TX=%TX%-win32-
+if _%T%_%PROCESSOR_ARCHITECTURE%_==__x86_ set T=i386
+if _%T%_%PROCESSOR_ARCHITECTURE%_==__ARM64_ set T=arm64
+if _%T%_==__ set T=x86_64
+if %T%==i386 set D=-DTCC_TARGET_PE -DTCC_TARGET_I386
+if %T%==x86_64 set D=-DTCC_TARGET_PE -DTCC_TARGET_X86_64
+if %T%==arm64 set D=-DTCC_TARGET_PE -DTCC_TARGET_ARM64
+if "%D%"=="" echo %SELF%: error: unknown target '%T%'&&exit /B 1
 
-set D32=-DTCC_TARGET_PE -DTCC_TARGET_I386
-set D64=-DTCC_TARGET_PE -DTCC_TARGET_X86_64
-set P32=i386-win32
-set P64=x86_64-win32
+@if (%CC:~0,3%)==(gcc) set CC=%CC% -s -static
+@if (%BINDIR%)==() set BINDIR=%TCCDIR%
 
-if %T%==64 goto :t64
-set D=%D32%
-set P=%P32%
-set DX=%D64%
-set PX=%P64%
-set TX=64
-goto :p3
-
-:t64
-set D=%D64%
-set P=%P64%
-set DX=%D32%
-set PX=%P32%
-set TX=32
-goto :p3
-
-:p3
+:git_hash
 git.exe --version 2>nul
 if not %ERRORLEVEL%==0 goto :git_done
 for /f %%b in ('git.exe rev-parse --abbrev-ref HEAD') do set GITHASH=%%b
@@ -125,34 +113,37 @@ if %ERRORLEVEL%==1 set GITHASH=%GITHASH%*
 
 :config.h
 echo>..\config.h #define TCC_VERSION "%VERSION%"
-if not (%GITHASH%)==() echo>> ..\config.h #define TCC_GITHASH "%GITHASH%"
-@if not (%BINDIR%)==(%TCCDIR%) echo>> ..\config.h #define CONFIG_TCCDIR "%TCCDIR:\=/%"
-if %TX%==64 echo>> ..\config.h #ifdef TCC_TARGET_X86_64
-if %TX%==32 echo>> ..\config.h #ifdef TCC_TARGET_I386
-echo>> ..\config.h #define CONFIG_TCC_CROSSPREFIX "%PX%-"
-echo>> ..\config.h #endif
+@if not "%GITHASH%"=="" echo>>..\config.h #define TCC_GITHASH "%GITHASH%"
+@if not _%BINDIR%_==_%TCCDIR%_ echo>>..\config.h #define CONFIG_TCCDIR "%TCCDIR:\=/%"
+@if not _%TX%_==__ @echo>>..\config.h #define CONFIG_TCC_CROSSPREFIX "%TX%"
 
 @rem echo>> ..\config.h #define CONFIG_TCC_PREDEFS 1
 @rem %CC% -DC2STR ..\conftest.c -o c2str.exe
 @rem .\c2str.exe ../include/tccdefs.h ../tccdefs_.h
 
-for %%f in (*tcc.exe *tcc.dll) do @del %%f
+@if not _%TX%_==__ goto :tcc_cross
+@if not _%TCC_C%_==__ goto :tcc_only
 
-@if _%TCC_C%_==__ goto compiler_2parts
-@rem if TCC_C was defined then build only tcc.exe
+@if _%LIBTCC_C%_==__ set LIBTCC_C=..\libtcc.c
+@set IMPLIB=libtcc.dll
+@if "%CC:~0,5%"=="clang" set IMPLIB=libtcc.lib
+
+%CC% -o libtcc.dll -shared %LIBTCC_C% %D% -DLIBTCC_AS_DLL
+@if errorlevel 1 goto :the_end
+%CC% -o tcc.exe ..\tcc.c %IMPLIB% %D% -DONE_SOURCE"=0"
+@if errorlevel 1 goto :the_end
+@goto :compiler_done
+
+:tcc_only
 %CC% -o tcc.exe %TCC_C% %D%
 @if errorlevel 1 goto :the_end
 @goto :compiler_done
 
-:compiler_2parts
-@if _%LIBTCC_C%_==__ set LIBTCC_C=..\libtcc.c
-%CC% -o libtcc.dll -shared %LIBTCC_C% %D% -DLIBTCC_AS_DLL
-@if errorlevel 1 goto :the_end
-%CC% -o tcc.exe ..\tcc.c libtcc.dll %D% -DONE_SOURCE"=0"
-@if errorlevel 1 goto :the_end
-if not _%XCC%_==_yes_ goto :compiler_done
-%CC% -o %PX%-tcc.exe ..\tcc.c %DX%
+:tcc_cross
+%CC% -o %TX%tcc.exe ..\tcc.c %D%
 @if errorlevel 1 goto :the_end
+@goto :compiler_done
+
 :compiler_done
 @if (%EXES_ONLY%)==(yes) goto :files_done
 
@@ -168,8 +159,7 @@ if exist libtcc.dll .\tcc -impdef libtcc.dll -o libtcc\libtcc.def
 @if errorlevel 1 goto :the_end
 
 :lib
-call :make_lib %T% || goto :the_end
-@if exist %PX%-tcc.exe call :make_lib %TX% %PX%- || goto :the_end
+@call :make_lib %TX% || goto :the_end
 
 :tcc-doc.html
 @if not (%DOC%)==(yes) goto :doc-done
@@ -193,23 +183,26 @@ for %%f in (include examples libtcc doc) do @xcopy>nul /s/i/q/y %%f %TCCDIR%\%%f
 exit /B %ERRORLEVEL%
 
 :make_lib
-.\tcc -B. -m%1 -c ../lib/libtcc1.c
-.\tcc -B. -m%1 -c lib/crt1.c
-.\tcc -B. -m%1 -c lib/crt1w.c
-.\tcc -B. -m%1 -c lib/wincrt1.c
-.\tcc -B. -m%1 -c lib/wincrt1w.c
-.\tcc -B. -m%1 -c lib/dllcrt1.c
-.\tcc -B. -m%1 -c lib/dllmain.c
-.\tcc -B. -m%1 -c lib/chkstk.S
-.\tcc -B. -m%1 -c ../lib/alloca.S
-.\tcc -B. -m%1 -c ../lib/alloca-bt.S
-.\tcc -B. -m%1 -c ../lib/stdatomic.c
-.\tcc -B. -m%1 -c ../lib/atomic.S
-.\tcc -B. -m%1 -c ../lib/builtin.c
-.\tcc -B. -m%1 -ar lib/%2libtcc1.a libtcc1.o crt1.o crt1w.o wincrt1.o wincrt1w.o dllcrt1.o dllmain.o chkstk.o alloca.o alloca-bt.o stdatomic.o atomic.o builtin.o
-.\tcc -B. -m%1 -c ../lib/bcheck.c -o lib/%2bcheck.o -bt -I..
-.\tcc -B. -m%1 -c ../lib/bt-exe.c -o lib/%2bt-exe.o
-.\tcc -B. -m%1 -c ../lib/bt-log.c -o lib/%2bt-log.o
-.\tcc -B. -m%1 -c ../lib/bt-dll.c -o lib/%2bt-dll.o
-.\tcc -B. -m%1 -c ../lib/runmain.c -o lib/%2runmain.o
+@set LIBTCC1=libtcc1
+@if _%1_==_arm64_ set LIBTCC1=lib-arm64
+.\%1tcc -B. -c ../lib/%LIBTCC1%.c
+.\%1tcc -B. -c lib/crt1.c
+.\%1tcc -B. -c lib/crt1w.c
+.\%1tcc -B. -c lib/wincrt1.c
+.\%1tcc -B. -c lib/wincrt1w.c
+.\%1tcc -B. -c lib/dllcrt1.c
+.\%1tcc -B. -c lib/dllmain.c
+.\%1tcc -B. -c lib/winex.c
+.\%1tcc -B. -c lib/chkstk.S
+.\%1tcc -B. -c ../lib/alloca.S
+.\%1tcc -B. -c ../lib/alloca-bt.S
+.\%1tcc -B. -c ../lib/stdatomic.c
+.\%1tcc -B. -c ../lib/atomic.S
+.\%1tcc -B. -c ../lib/builtin.c
+.\%1tcc -ar lib/%1libtcc1.a %LIBTCC1%.o crt1.o crt1w.o wincrt1.o wincrt1w.o dllcrt1.o dllmain.o winex.o chkstk.o alloca.o alloca-bt.o stdatomic.o atomic.o builtin.o
+.\%1tcc -B. -c ../lib/bcheck.c -o lib/%1bcheck.o -bt -I..
+.\%1tcc -B. -c ../lib/bt-exe.c -o lib/%1bt-exe.o
+.\%1tcc -B. -c ../lib/bt-log.c -o lib/%1bt-log.o
+.\%1tcc -B. -c ../lib/bt-dll.c -o lib/%1bt-dll.o
+.\%1tcc -B. -c ../lib/runmain.c -o lib/%1runmain.o
 exit /B %ERRORLEVEL%
diff --git a/win32/include/_mingw.h b/win32/include/_mingw.h
index 77a3bce6..3a677202 100644
--- a/win32/include/_mingw.h
+++ b/win32/include/_mingw.h
@@ -70,20 +70,14 @@
 
 #ifdef _WIN64
 #define __stdcall
-#if defined(__aarch64__) || defined(_M_ARM64) || defined(_ARM64_)
-#ifndef _M_ARM64
+#if defined(__aarch64__)
 #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__))
diff --git a/win32/include/setjmp.h b/win32/include/setjmp.h
index dc971c77..ab6a4248 100644
--- a/win32/include/setjmp.h
+++ b/win32/include/setjmp.h
@@ -170,7 +170,6 @@ extern "C" {
     double D[8];
   } _JUMP_BUFFER;
 #else
-
 #define _JBLEN 1
 #define _JBTYPE int
 #endif
@@ -179,31 +178,22 @@ extern "C" {
 #define _JMP_BUF_DEFINED
 #endif
 
-  void * __cdecl __attribute__ ((__nothrow__)) mingw_getsp(void);
+#pragma pack(pop)
 
-#ifdef USE_MINGW_SETJMP_TWO_ARGS
-#ifndef _INC_SETJMPEX
-#define setjmp(BUF) _setjmp((BUF),mingw_getsp())
-  int __cdecl __attribute__ ((__nothrow__)) _setjmp(jmp_buf _Buf,void *_Ctx);
-#else
-#undef setjmp
-#define setjmp(BUF) _setjmpex((BUF),mingw_getsp())
-#define setjmpex(BUF) _setjmpex((BUF),mingw_getsp())
-  int __cdecl __attribute__ ((__nothrow__)) _setjmpex(jmp_buf _Buf,void *_Ctx);
-#endif
-#else
-#ifndef _INC_SETJMPEX
-#define setjmp _setjmp
-#endif
-  int __cdecl __attribute__ ((__nothrow__)) setjmp(jmp_buf _Buf);
+#if defined __aarch64__
+  int _setjmpex(jmp_buf _Buf, void *frame);
+  #define setjmp(BUF) _setjmpex((BUF), (char*)__builtin_frame_address(0) + 224)
+#elif defined __x86_64__
+  int _setjmp(jmp_buf _Buf, void *frame);
+  #define setjmp(BUF) _setjmp((BUF), __builtin_frame_address(0))
+#else /* __i386__ */
+  int _setjmp(jmp_buf _Buf);
+  #define setjmp _setjmp
 #endif
 
-  __declspec(noreturn) __attribute__ ((__nothrow__)) void __cdecl ms_longjmp(jmp_buf _Buf,int _Value)/* throw(...)*/;
-  __declspec(noreturn) __attribute__ ((__nothrow__)) void __cdecl longjmp(jmp_buf _Buf,int _Value);
+__declspec(noreturn) void longjmp(jmp_buf _Buf,int _Value);
 
 #ifdef __cplusplus
 }
 #endif
-
-#pragma pack(pop)
 #endif
diff --git a/win32/include/stdlib.h b/win32/include/stdlib.h
index 21a0fbd9..b710ec13 100644
--- a/win32/include/stdlib.h
+++ b/win32/include/stdlib.h
@@ -207,6 +207,21 @@ extern "C" {
 #endif
 #endif
 #endif
+
+#if defined __aarch64__
+/* something does not work using those from msvcrt.dll */
+# undef __argc
+# undef __argv
+# undef __wargv
+# undef _wenviron
+# undef _environ
+extern int __argc;
+extern char **__argv;
+extern wchar_t **__wargv;
+extern char **_environ;
+extern wchar_t **_wenviron;
+#endif
+
 #ifndef _pgmptr
 #ifdef _MSVCRT_
   extern char *_pgmptr;
diff --git a/win32/include/winapi/winnt.h b/win32/include/winapi/winnt.h
index 01a850a8..405413fb 100644
--- a/win32/include/winapi/winnt.h
+++ b/win32/include/winapi/winnt.h
@@ -47,9 +47,6 @@ 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)
@@ -832,8 +829,6 @@ typedef DWORD LCID;
   typedef ULONG_PTR KSPIN_LOCK;
   typedef KSPIN_LOCK *PKSPIN_LOCK;
 
-#if defined(_AMD64_) || defined(_ARM64_)
-
 #if defined(__x86_64) && !defined(RC_INVOKED)
 
 #ifdef __cplusplus
@@ -1285,7 +1280,6 @@ typedef DWORD LCID;
 #ifdef __cplusplus
   }
 #endif
-#endif
 
 #define EXCEPTION_READ_FAULT 0
 #define EXCEPTION_WRITE_FAULT 1
@@ -1339,7 +1333,6 @@ 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;
@@ -1411,7 +1404,6 @@ typedef DWORD LCID;
     DWORD64 LastExceptionToRip;
     DWORD64 LastExceptionFromRip;
   } CONTEXT,*PCONTEXT;
-#endif /* defined(__x86_64) || defined(_AMD64_) */
 
 #define RUNTIME_FUNCTION_INDIRECT 0x1
 
@@ -1422,41 +1414,24 @@ typedef DWORD LCID;
   } RUNTIME_FUNCTION,*PRUNTIME_FUNCTION;
 
   typedef PRUNTIME_FUNCTION (*PGET_RUNTIME_FUNCTION_CALLBACK)(DWORD64 ControlPc,PVOID Context);
+  typedef DWORD (*POUT_OF_PROCESS_FUNCTION_TABLE_CALLBACK)(HANDLE Process,PVOID TableAddress,PDWORD Entries,PRUNTIME_FUNCTION *Functions);
 
-#if defined(_ARM64_) || defined(__aarch64__)
+#define OUT_OF_PROCESS_FUNCTION_TABLE_CALLBACK_EXPORT_NAME "OutOfProcessFunctionTableCallback"
+#endif /* defined(__x86_64) && !defined(RC_INVOKED) */
+
+#if defined(_ARM64_)
 
 /* 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;
@@ -1467,10 +1442,7 @@ typedef DWORD LCID;
     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;
@@ -1522,27 +1494,20 @@ typedef DWORD LCID;
   } 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 struct _RUNTIME_FUNCTION {
+    DWORD BeginAddress;
+    DWORD UnwindData;
+  } RUNTIME_FUNCTION,*PRUNTIME_FUNCTION;
+
+  typedef PRUNTIME_FUNCTION (*PGET_RUNTIME_FUNCTION_CALLBACK)(DWORD64 ControlPc,PVOID Context);
   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"
+#endif /* _ARM64_ */
 
+#if (defined _ARM64_ || defined _AMD64_) && !defined RC_INVOKED
   NTSYSAPI VOID __cdecl RtlRestoreContext (PCONTEXT ContextRecord,struct _EXCEPTION_RECORD *ExceptionRecord);
   NTSYSAPI BOOLEAN __cdecl RtlAddFunctionTable(PRUNTIME_FUNCTION FunctionTable,DWORD EntryCount,DWORD64 BaseAddress);
   NTSYSAPI BOOLEAN __cdecl RtlInstallFunctionTableCallback(DWORD64 TableIdentifier,DWORD64 BaseAddress,DWORD Length,PGET_RUNTIME_FUNCTION_CALLBACK Callback,PVOID Context,PCWSTR OutOfProcessCallbackDll);
diff --git a/win32/lib/chkstk.S b/win32/lib/chkstk.S
index 43e21634..2b5b6175 100644
--- a/win32/lib/chkstk.S
+++ b/win32/lib/chkstk.S
@@ -9,7 +9,6 @@
 
 /* ---------------------------------------------- */
 #if defined(__aarch64__)
-/* ---------------------------------------------- */
 
 .globl __chkstk
 __chkstk:
@@ -33,68 +32,8 @@ L_chkstk_tail:
 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__)
 /* ---------------------------------------------- */
+#elif defined(__i386__)
 
 .globl _(__chkstk)
 _(__chkstk):
@@ -118,8 +57,7 @@ P0:
     jmp     *4(%eax)
 
 /* ---------------------------------------------- */
-#else
-/* ---------------------------------------------- */
+#else /* __x86_64__ */
 
 .globl _(__chkstk)
 _(__chkstk):
@@ -143,14 +81,6 @@ P0:
     mov     (%rax),%rcx     /* restore ecx */
     jmp     *8(%rax)
 
-/* ---------------------------------------------- */
-/* setjmp/longjmp support */
-
-.globl _(tinyc_getbp)
-_(tinyc_getbp):
-    mov %rbp,%rax
-    ret
-
 /* ---------------------------------------------- */
 #endif
 /* ---------------------------------------------- */
diff --git a/win32/lib/crt1.c b/win32/lib/crt1.c
index e1910813..7022872a 100644
--- a/win32/lib/crt1.c
+++ b/win32/lib/crt1.c
@@ -75,10 +75,11 @@ __attribute__((weak)) extern int __run_on_exit();
 int _runtmain(int argc, /* as tcc passed in */ char **argv)
 {
     int ret;
-#ifdef UNICODE
+#if defined UNICODE || defined __aarch64__
     _startupinfo start_info = {0};
-
-    __tgetmainargs(&__argc, &__targv, &_tenviron, _dowildcard, &start_info);
+    __tgetmainargs(&__argc, &__targv, &_tenviron, 0, &start_info);
+#endif
+#ifdef UNICODE
     /* may be wrong when tcc has received wildcards (*.c) */
     if (argc < __argc) {
         __targv += __argc - argc;
@@ -93,6 +94,8 @@ int _runtmain(int argc, /* as tcc passed in */ char **argv)
 #endif
     run_ctors(__argc, __targv, _tenviron);
     ret = _tmain(__argc, __targv, _tenviron);
+    fflush(stdout);
+    fflush(stderr);
     run_dtors();
     __run_on_exit(ret);
     return ret;
diff --git a/win32/lib/winex.c b/win32/lib/winex.c
new file mode 100644
index 00000000..d84d8d75
--- /dev/null
+++ b/win32/lib/winex.c
@@ -0,0 +1,27 @@
+/* ------------------------------------------------------------------------- */
+/* winex.c : extra stuff */
+
+#if __aarch64__
+#include <stdlib.h>
+/* replaces environ from arm64-msvcrt.dll which does not exist */
+char **_environ;
+wchar_t **_wenviron;
+/* those do exist but have problems */
+int __argc;
+char **__argv;
+wchar_t **__wargv;
+#endif
+
+#if __aarch64__ || __x86_64__
+/* MSVC x64 intrinsic */
+void __faststorefence(void)
+{
+#if __aarch64__
+    /* ARM64: Data Memory Barrier (Inner Shareable) */
+    __asm__("dmb ish");
+#elif __x86_64__
+    /* x86-64: lock prefix to flush store buffer */
+    __asm__("lock; orl $0,(%%rsp)" ::: "memory");
+#endif
+}
+#endif
diff --git a/x86_64-gen.c b/x86_64-gen.c
index e1f4ae70..06eb90c0 100644
--- a/x86_64-gen.c
+++ b/x86_64-gen.c
@@ -266,14 +266,6 @@ ST_FUNC void gen_addr32(int r, Sym *sym, int c)
     gen_le32(c);
 }
 
-/* output constant with relocation if 'r & VT_SYM' is true */
-ST_FUNC void gen_addr64(int r, Sym *sym, int64_t c)
-{
-    if (r & VT_SYM)
-        greloca(cur_text_section, sym, ind, R_X86_64_64, c), c=0;
-    gen_le64(c);
-}
-
 /* output constant with relocation if 'r & VT_SYM' is true */
 ST_FUNC void gen_addrpc32(int r, Sym *sym, int c)
 {