tccrun: win64: add unwind function table for dynamic code

This works only when tcc.exe is compiled using MSC.  MinGW does
something in the startup code that defeats it.

tcc.h

@@ -566,10 +566,11 @@ struct TCCState {
     int pe_subsystem;
     unsigned long pe_file_align;
     unsigned long pe_stack_size;
-    struct pe_uw {
-        Section *pdata;
-        int sym_1, sym_2, offs_1;
-    } pe_unwind;
+#ifdef TCC_TARGET_X86_64
+    Section *uw_pdata;
+    int uw_sym;
+    unsigned uw_offs;
+#endif
 #endif
 
 #ifndef TCC_TARGET_PE

tccpe.c

@@ -1694,54 +1694,65 @@ ST_FUNC int pe_add_dll(struct TCCState *s, const char *libname)
 
 /* ------------------------------------------------------------- */
 #ifdef TCC_TARGET_X86_64
-ST_FUNC void pe_add_unwind_data(unsigned start, unsigned end, unsigned stack)
+static unsigned pe_add_uwwind_info(TCCState *s1)
 {
-    static const char uw_info[] = {
-        0x01, // UBYTE: 3 Version , UBYTE: 5 Flags
-        0x04, // UBYTE Size of prolog
-        0x02, // UBYTE Count of unwind codes
-        0x05, // UBYTE: 4 Frame Register (rbp), UBYTE: 4 Frame Register offset (scaled)
-        // USHORT * n Unwind codes array
-        // 0x0b, 0x01, 0xff, 0xff, // stack size
-        0x04, 0x03, // set frame ptr (mov rsp -> rbp)
-        0x01, 0x50  // push reg (rbp)
-    };
-
-    struct pe_uw *pe_uw = &tcc_state->pe_unwind;
-
-    Section *uw, *pd;
-    WORD *p1;
-    DWORD *p2;
-    unsigned o2;
-
-    uw = data_section;
-    pd = pe_uw->pdata;
-    if (NULL == pd)
-    {
-        pe_uw->pdata = pd = find_section(tcc_state, ".pdata");
-        pe_uw->pdata->sh_addralign = 4;
-        section_ptr_add(uw, -uw->data_offset & 3);
-        pe_uw->offs_1 = uw->data_offset;
-        p1 = section_ptr_add(uw, sizeof uw_info);
-        /* use one common entry for all functions */
-        memcpy(p1, uw_info, sizeof uw_info);
-        pe_uw->sym_1 = put_elf_sym(symtab_section, 0, 0, 0, 0, text_section->sh_num, NULL);
-        pe_uw->sym_2 = put_elf_sym(symtab_section, 0, 0, 0, 0, uw->sh_num, NULL);
+    if (NULL == s1->uw_pdata) {
+        s1->uw_pdata = find_section(tcc_state, ".pdata");
+        s1->uw_pdata->sh_addralign = 4;
+        s1->uw_sym = put_elf_sym(symtab_section, 0, 0, 0, 0, text_section->sh_num, NULL);
     }
 
-    o2 = pd->data_offset;
-    p2 = section_ptr_add(pd, 3 * sizeof (DWORD));
+    if (0 == s1->uw_offs) {
+        /* As our functions all have the same stackframe, we use one entry for all */
+        static const unsigned char uw_info[] = {
+            0x01, // UBYTE: 3 Version , UBYTE: 5 Flags
+            0x04, // UBYTE Size of prolog
+            0x02, // UBYTE Count of unwind codes
+            0x05, // UBYTE: 4 Frame Register (rbp), UBYTE: 4 Frame Register offset (scaled)
+            // USHORT * n Unwind codes array
+            // 0x0b, 0x01, 0xff, 0xff, // stack size
+            0x04, 0x03, // set frame ptr (mov rsp -> rbp)
+            0x01, 0x50  // push reg (rbp)
+        };
+
+        Section *s = text_section;
+        unsigned char *p;
+
+        section_ptr_add(s, -s->data_offset & 3); /* align */
+        s1->uw_offs = s->data_offset;
+        p = section_ptr_add(s, sizeof uw_info);
+        memcpy(p, uw_info, sizeof uw_info);
+    }
+
+    return s1->uw_offs;
+}
+
+ST_FUNC void pe_add_unwind_data(unsigned start, unsigned end, unsigned stack)
+{
+    TCCState *s1 = tcc_state;
+    Section *pd;
+    unsigned o, n, d;
+    struct /* _RUNTIME_FUNCTION */ {
+      DWORD BeginAddress;
+      DWORD EndAddress;
+      DWORD UnwindData;
+    } *p;
+
+    d = pe_add_uwwind_info(s1);
+    pd = s1->uw_pdata;
+    o = pd->data_offset;
+    p = section_ptr_add(pd, sizeof *p);
+
     /* record this function */
-    p2[0] = start;
-    p2[1] = end;
-    p2[2] = pe_uw->offs_1;
+    p->BeginAddress = start;
+    p->EndAddress = end;
+    p->UnwindData = d;
+
     /* put relocations on it */
-    put_elf_reloc(symtab_section, pd, o2,   R_X86_64_RELATIVE, pe_uw->sym_1);
-    put_elf_reloc(symtab_section, pd, o2+4, R_X86_64_RELATIVE, pe_uw->sym_1);
-    put_elf_reloc(symtab_section, pd, o2+8, R_X86_64_RELATIVE, pe_uw->sym_2);
+    for (n = o + sizeof *p; o < n; o += sizeof p->BeginAddress)
+        put_elf_reloc(symtab_section, pd, o,  R_X86_64_RELATIVE, s1->uw_sym);
 }
 #endif
-
 /* ------------------------------------------------------------- */
 #ifdef TCC_TARGET_X86_64
 #define PE_STDSYM(n,s) n

tccrun.c

@@ -30,6 +30,10 @@ static int rt_get_caller_pc(uplong *paddr, ucontext_t *uc, int level);
 static void rt_error(ucontext_t *uc, const char *fmt, ...);
 static int tcc_relocate_ex(TCCState *s1, void *ptr);
 
+#ifdef _WIN64
+static void win64_add_function_table(TCCState *s1);
+#endif
+
 /* ------------------------------------------------------------- */
 /* Do all relocations (needed before using tcc_get_symbol())
    Returns -1 on error. */
@@ -189,6 +193,10 @@ static int tcc_relocate_ex(TCCState *s1, void *ptr)
     set_pages_executable(s1->runtime_plt_and_got,
                          s1->runtime_plt_and_got_offset);
 #endif
+
+#ifdef _WIN64
+    win64_add_function_table(s1);
+#endif
     return 0;
 }
 
@@ -578,6 +586,17 @@ static void set_exception_handler(void)
     SetUnhandledExceptionFilter(cpu_exception_handler);
 }
 
+#ifdef _WIN64
+static void win64_add_function_table(TCCState *s1)
+{
+    RtlAddFunctionTable(
+        (RUNTIME_FUNCTION*)(uplong)s1->uw_pdata->sh_addr,
+        s1->uw_pdata->data_offset / sizeof (RUNTIME_FUNCTION),
+        (uplong)text_section->sh_addr
+        );
+}
+#endif
+
 #ifdef _WIN64
 #define Eip Rip
 #define Ebp Rbp

win32/lib/chkstk.S

@@ -57,17 +57,9 @@ P0:
 /* ---------------------------------------------- */
 /* setjmp/longjmp support */
 
-.globl tinyc_no_getbp
-tinyc_no_getbp:
-    .byte 0x90
-
 .globl tinyc_getbp
 tinyc_getbp:
-    xor %rax,%rax
-    cmp %al,tinyc_no_getbp(%rax)
-    je t1
     mov %rbp,%rax
-t1:
     ret
 
 /* ---------------------------------------------- */

win32/lib/kernel32.def

@@ -549,7 +549,9 @@ ResetEvent
 ResetNLSUserInfoCache
 ResetWriteWatch
 ResumeThread
+RtlAddFunctionTable
 RtlFillMemory
+RtlInstallFunctionTableCallback
 RtlMoveMemory
 RtlUnwind
 RtlZeroMemory

x86_64-gen.c

@@ -772,8 +772,6 @@ void gfunc_epilog(void)
     /* align local size to word & save local variables */
     v = (func_scratch + -loc + 15) & -16;
 
-    pe_add_unwind_data(ind, saved_ind, v);
-
     if (v >= 4096) {
         Sym *sym = external_global_sym(TOK___chkstk, &func_old_type, 0);
         oad(0xb8, v); /* mov stacksize, %eax */
@@ -785,7 +783,10 @@ void gfunc_epilog(void)
         o(0xec8148);  /* sub rsp, stacksize */
         gen_le32(v);
     }
-    ind = saved_ind;
+
+    cur_text_section->data_offset = saved_ind;
+    pe_add_unwind_data(ind, saved_ind, v);
+    ind = cur_text_section->data_offset;
 }
 
 #else