stdarg: always have the __builtin_va_* available

This makes available the __builtin_va_list type and __builtin variants
of va_start, va_arg, va_copy and va_end.  We do this via a header file
that's prepended to all compilations always (except if merely
preprocessing): tcc_predefs.h.  That header could also be used
for predefining other builtins in the future.

We don't need the define hacks for musl anymore with this.

Also fix x86_64 gfunc_prologue to reserve enoug space for the
full va_list structure, not just 16 bytes.

include/stdarg.h

@@ -1,88 +1,11 @@
 #ifndef _STDARG_H
 #define _STDARG_H
 
-#ifdef __x86_64__
-#ifndef _WIN64
-
-#ifndef __APPLE__
-//This should be in sync with the declaration on our lib/libtcc1.c
-/* GCC compatible definition of va_list. */
-typedef struct {
-    unsigned int gp_offset;
-    unsigned int fp_offset;
-    union {
-        unsigned int overflow_offset;
-        char *overflow_arg_area;
-    };
-    char *reg_save_area;
-} __va_list_struct;
-typedef __va_list_struct va_list[1];
-#else
-/* This is sometimes a void* on TCC, which makes it unlikely to
-   work with va_copy, but until we have something better ... */
 typedef __builtin_va_list va_list;
-#endif
-
-void __va_start(va_list ap, void *fp);
-void *__va_arg(va_list ap, int arg_type, int size, int align);
-
-#define va_start(ap, last) __va_start(ap, __builtin_frame_address(0))
-#define va_arg(ap, type)                                                \
-    (*(type *)(__va_arg(ap, __builtin_va_arg_types(type), sizeof(type), __alignof__(type))))
-#define va_copy(dest, src) (*(dest) = *(src))
-#define va_end(ap) ((void)0)
-
-#else /* _WIN64 */
-typedef char *va_list;
-#define va_start(ap,last) __builtin_va_start(ap,last)
-#define va_arg(ap, t) ((sizeof(t) > 8 || (sizeof(t) & (sizeof(t) - 1))) \
-	? **(t **)((ap += 8) - 8) : *(t  *)((ap += 8) - 8))
-#define va_copy(dest, src) ((dest) = (src))
-#define va_end(ap) ((void)0)
-#endif
-
-#elif __arm__
-typedef char *va_list;
-#define _tcc_alignof(type) ((int)&((struct {char c;type x;} *)0)->x)
-#define _tcc_align(addr,type) (((unsigned)addr + _tcc_alignof(type) - 1) \
-                               & ~(_tcc_alignof(type) - 1))
-#define va_start(ap,last) ap = ((char *)&(last)) + ((sizeof(last)+3)&~3)
-#define va_arg(ap,type) (ap = (void *) ((_tcc_align(ap,type)+sizeof(type)+3) \
-                        &~3), *(type *)(ap - ((sizeof(type)+3)&~3)))
-#define va_copy(dest, src) (dest) = (src)
-#define va_end(ap) ((void)0)
-
-#elif defined(__aarch64__)
-typedef struct {
-    void *__stack;
-    void *__gr_top;
-    void *__vr_top;
-    int   __gr_offs;
-    int   __vr_offs;
-} va_list;
-#define va_start(ap, last) __va_start(ap, last)
-#define va_arg(ap, type) __va_arg(ap, type)
-#define va_end(ap) ((void)0)
-#define va_copy(dest, src) ((dest) = (src))
-
-#elif defined __riscv
-#define __va_reg_size (__riscv_xlen >> 3)
-#define _tcc_align(addr,type) (((unsigned long)addr + __alignof__(type) - 1) \
-                               & -(__alignof__(type)))
-typedef char *va_list;
 #define va_start __builtin_va_start
-#define va_arg(ap,type) (*(sizeof(type) > (2*__va_reg_size) ? *(type **)((ap += __va_reg_size) - __va_reg_size) : (ap = (va_list)(_tcc_align(ap,type) + (sizeof(type)+__va_reg_size - 1)& -__va_reg_size), (type *)(ap - ((sizeof(type)+ __va_reg_size - 1)& -__va_reg_size)))))
-#define va_copy(dest, src) (dest) = (src)
-#define va_end(ap) ((void)0)
-
-#else /* __i386__ */
-typedef char *va_list;
-/* only correct for i386 */
-#define va_start(ap,last) ap = ((char *)&(last)) + ((sizeof(last)+3)&~3)
-#define va_arg(ap,type) (ap += (sizeof(type)+3)&~3, *(type *)(ap - ((sizeof(type)+3)&~3)))
-#define va_copy(dest, src) (dest) = (src)
-#define va_end(ap) ((void)0)
-#endif
+#define va_arg __builtin_va_arg
+#define va_copy __builtin_va_copy
+#define va_end __builtin_va_end
 
 /* fix a buggy dependency on GCC in libio.h */
 typedef va_list __gnuc_va_list;

include/tcc_predefs.h

@@ -0,0 +1,64 @@
+#ifdef __x86_64__
+#ifndef _WIN64
+
+//This should be in sync with the declaration in our lib/libtcc1.c
+/* GCC compatible definition of va_list. */
+typedef struct {
+    unsigned int gp_offset;
+    unsigned int fp_offset;
+    union {
+        unsigned int overflow_offset;
+        char *overflow_arg_area;
+    };
+    char *reg_save_area;
+} __builtin_va_list[1];
+
+void *__va_arg(__builtin_va_list ap, int arg_type, int size, int align);
+
+#define __builtin_va_start(ap, last) \
+    (*(ap) = *(__builtin_va_list)((char*)__builtin_frame_address(0) - 24))
+#define __builtin_va_arg(ap, t)   \
+    (*(t *)(__va_arg(ap, __builtin_va_arg_types(t), sizeof(t), __alignof__(t))))
+#define __builtin_va_copy(dest, src) (*(dest) = *(src))
+
+#else /* _WIN64 */
+typedef char *__builtin_va_list;
+#define __builtin_va_arg(ap, t) ((sizeof(t) > 8 || (sizeof(t) & (sizeof(t) - 1))) \
+	? **(t **)((ap += 8) - 8) : *(t  *)((ap += 8) - 8))
+#endif
+
+#elif __arm__
+typedef char *__builtin_va_list;
+#define _tcc_alignof(type) ((int)&((struct {char c;type x;} *)0)->x)
+#define _tcc_align(addr,type) (((unsigned)addr + _tcc_alignof(type) - 1) \
+                               & ~(_tcc_alignof(type) - 1))
+#define __builtin_va_start(ap,last) ap = ((char *)&(last)) + ((sizeof(last)+3)&~3)
+#define __builtin_va_arg(ap,type) (ap = (void *) ((_tcc_align(ap,type)+sizeof(type)+3) \
+                        &~3), *(type *)(ap - ((sizeof(type)+3)&~3)))
+
+#elif defined(__aarch64__)
+typedef struct {
+    void *__stack;
+    void *__gr_top;
+    void *__vr_top;
+    int   __gr_offs;
+    int   __vr_offs;
+} __builtin_va_list;
+
+#elif defined __riscv
+typedef char *__builtin_va_list;
+#define __va_reg_size (__riscv_xlen >> 3)
+#define _tcc_align(addr,type) (((unsigned long)addr + __alignof__(type) - 1) \
+                               & -(__alignof__(type)))
+#define __builtin_va_arg(ap,type) (*(sizeof(type) > (2*__va_reg_size) ? *(type **)((ap += __va_reg_size) - __va_reg_size) : (ap = (va_list)(_tcc_align(ap,type) + (sizeof(type)+__va_reg_size - 1)& -__va_reg_size), (type *)(ap - ((sizeof(type)+ __va_reg_size - 1)& -__va_reg_size)))))
+
+#else /* __i386__ */
+typedef char *__builtin_va_list;
+#define __builtin_va_start(ap,last) ap = ((char *)&(last)) + ((sizeof(last)+3)&~3)
+#define __builtin_va_arg(ap,type) (ap += (sizeof(type)+3)&~3, *(type *)(ap - ((sizeof(type)+3)&~3)))
+#endif
+
+#ifndef __builtin_va_copy
+#define __builtin_va_copy(dest, src) (dest) = (src)
+#endif
+#define __builtin_va_end(ap) (void)(ap)

lib/va_list.c

@@ -3,7 +3,6 @@
 #if defined __x86_64__
 
 /* Avoid include files, they may not be available when cross compiling */
-extern void *memset(void *s, int c, __SIZE_TYPE__ n);
 extern void abort(void);
 
 /* This should be in sync with our include/stdarg.h */
@@ -12,6 +11,7 @@ enum __va_arg_type {
 };
 
 /* GCC compatible definition of va_list. */
+/*predefined by TCC (tcc_predefs.h):
 typedef struct {
     unsigned int gp_offset;
     unsigned int fp_offset;
@@ -20,23 +20,16 @@ typedef struct {
         char *overflow_arg_area;
     };
     char *reg_save_area;
-} __va_list_struct;
+} __builtin_va_list[1];
+*/
 
-void __va_start(__va_list_struct *ap, void *fp)
-{
-    memset(ap, 0, sizeof(__va_list_struct));
-    *ap = *(__va_list_struct *)((char *)fp - 16);
-    ap->overflow_arg_area = (char *)fp + ap->overflow_offset;
-    ap->reg_save_area = (char *)fp - 176 - 16;
-}
-
-void *__va_arg(__va_list_struct *ap,
+void *__va_arg(__builtin_va_list ap,
                int arg_type,
                int size, int align)
 {
     size = (size + 7) & ~7;
     align = (align + 7) & ~7;
-    switch (arg_type) {
+    switch ((enum __va_arg_type)arg_type) {
     case __va_gen_reg:
         if (ap->gp_offset + size <= 48) {
             ap->gp_offset += size;

libtcc.c

@@ -919,11 +919,6 @@ LIBTCCAPI TCCState *tcc_new(void)
     tcc_define_symbol(s, "__REDIRECT_NTH(name, proto, alias)",
         "name proto __asm__ (#alias) __THROW");
 # endif
-# if defined(TCC_MUSL)
-    tcc_define_symbol(s, "__DEFINED_va_list", "");
-    tcc_define_symbol(s, "__DEFINED___isoc_va_list", "");
-    tcc_define_symbol(s, "__isoc_va_list", "void *");
-# endif /* TCC_MUSL */
     /* Some GCC builtins that are simple to express as macros.  */
     tcc_define_symbol(s, "__builtin_extract_return_addr(x)", "x");
 #endif /* ndef TCC_TARGET_PE */
@@ -935,7 +930,6 @@ LIBTCCAPI TCCState *tcc_new(void)
     /* avoids usage of GCC/clang specific builtins in libc-headerfiles: */
     tcc_define_symbol(s, "__FINITE_MATH_ONLY__", "1");
     tcc_define_symbol(s, "_FORTIFY_SOURCE", "0");
-    tcc_define_symbol(s, "__builtin_va_list", "void *");
 #endif /* ndef TCC_TARGET_MACHO */
     return s;
 }

tccgen.c

@@ -5348,12 +5348,12 @@ ST_FUNC void unary(void)
 	vpushi(classify_x86_64_va_arg(&vtop->type));
 	vswap();
 	vpop();
-        break;
+	break;
 #endif
 #endif
 
 #ifdef TCC_TARGET_ARM64
-    case TOK___va_start: {
+    case TOK_builtin_va_start: {
 	parse_builtin_params(0, "ee");
         //xx check types
         gen_va_start();
@@ -5361,7 +5361,7 @@ ST_FUNC void unary(void)
         vtop->type.t = VT_VOID;
         break;
     }
-    case TOK___va_arg: {
+    case TOK_builtin_va_arg: {
 	parse_builtin_params(0, "et");
 	type = vtop->type;
 	vpop();

tccpp.c

@@ -3611,6 +3611,8 @@ ST_FUNC void preprocess_start(TCCState *s1, int is_asm)
         cstr_printf(&cstr, "#define __ASSEMBLER__ 1\n");
     if (s1->output_type == TCC_OUTPUT_MEMORY)
         cstr_printf(&cstr, "#define __TCC_RUN__ 1\n");
+    if (!is_asm && s1->output_type != TCC_OUTPUT_PREPROCESS)
+        cstr_cat(&cstr, "#include \"tcc_predefs.h\"\n", -1);
     if (s1->cmdline_incl.size)
         cstr_cat(&cstr, s1->cmdline_incl.data, s1->cmdline_incl.size);
     //printf("%s\n", (char*)cstr.data);

tcctok.h

@@ -160,8 +160,8 @@
 #elif defined TCC_TARGET_X86_64
      DEF(TOK_builtin_va_arg_types, "__builtin_va_arg_types")
 #elif defined TCC_TARGET_ARM64
-     DEF(TOK___va_start, "__va_start")
-     DEF(TOK___va_arg, "__va_arg")
+     DEF(TOK_builtin_va_start, "__builtin_va_start")
+     DEF(TOK_builtin_va_arg, "__builtin_va_arg")
 #elif defined TCC_TARGET_RISCV64
      DEF(TOK_builtin_va_start, "__builtin_va_start")
 #endif

x86_64-gen.c

@@ -1508,16 +1508,23 @@ void gfunc_prolog(Sym *func_sym)
             }
         }
 
-        loc -= 16;
-        /* movl $0x????????, -0x10(%rbp) */
-        o(0xf045c7);
+        loc -= 24;
+        /* movl $0x????????, -0x18(%rbp) */
+        o(0xe845c7);
         gen_le32(seen_reg_num * 8);
-        /* movl $0x????????, -0xc(%rbp) */
-        o(0xf445c7);
+        /* movl $0x????????, -0x14(%rbp) */
+        o(0xec45c7);
         gen_le32(seen_sse_num * 16 + 48);
-        /* movl $0x????????, -0x8(%rbp) */
-        o(0xf845c7);
-        gen_le32(seen_stack_size);
+	/* leaq $0x????????, %r11 */
+	o(0x9d8d4c);
+	gen_le32(seen_stack_size);
+	/* movq %r11, -0x10(%rbp) */
+	o(0xf05d894c);
+	/* leaq $-192(%rbp), %r11 */
+	o(0x9d8d4c);
+	gen_le32(-176 - 24);
+	/* movq %r11, -0x8(%rbp) */
+	o(0xf85d894c);
 
         /* save all register passing arguments */
         for (i = 0; i < 8; i++) {