diff --git a/Makefile b/Makefile
index 62b7372..3a5c000 100644
--- a/Makefile
+++ b/Makefile
@@ -173,11 +173,12 @@ ifeq ($(ONE_SOURCE),yes)
 LIBTCC_OBJ = $(X)libtcc.o
 LIBTCC_INC = $($T_FILES)
 TCC_FILES = $(X)tcc.o
-$(X)libtcc.o $T-tcc$(EXESUF) : DEFINES += -DONE_SOURCE
+tcc.o : DEFINES += -DONE_SOURCE=0
 else
 LIBTCC_OBJ = $(patsubst %.c,$(X)%.o,$(LIBTCC_SRC))
 LIBTCC_INC = $(filter %.h %-gen.c %-link.c,$($T_FILES))
 TCC_FILES = $(X)tcc.o $(LIBTCC_OBJ)
+$(TCC_FILES) : DEFINES += -DONE_SOURCE=0
 endif
 
 # target specific object rule
@@ -200,7 +201,7 @@ $(CROSS_TARGET)-tcc$(EXESUF): $(TCC_FILES)
 
 # profiling version
 tcc_p$(EXESUF): $($T_FILES)
-	$(CC) -o $@ $< $(DEFINES) -DONE_SOURCE $(CFLAGS_P) $(LIBS_P) $(LDFLAGS_P)
+	$(CC) -o $@ $< $(DEFINES) $(CFLAGS_P) $(LIBS_P) $(LDFLAGS_P)
 
 # static libtcc library
 libtcc.a: $(LIBTCC_OBJ)
diff --git a/arm-asm.c b/arm-asm.c
index 15253ab..3b5ae66 100644
--- a/arm-asm.c
+++ b/arm-asm.c
@@ -70,7 +70,6 @@ ST_FUNC void asm_gen_code(ASMOperand *operands, int nb_operands,
                          uint8_t *clobber_regs,
                          int out_reg)
 {
-    asm_error();
 }
 
 ST_FUNC void asm_compute_constraints(ASMOperand *operands,
@@ -78,7 +77,6 @@ ST_FUNC void asm_compute_constraints(ASMOperand *operands,
                                     const uint8_t *clobber_regs,
                                     int *pout_reg)
 {
-    asm_error();
 }
 
 ST_FUNC void asm_clobber(uint8_t *clobber_regs, const char *str)
diff --git a/configure b/configure
index 4a337e7..78f2072 100755
--- a/configure
+++ b/configure
@@ -410,7 +410,7 @@ echo "Info directory      $infodir"
 echo "Doc directory       $docdir"
 fcho "Target root prefix  " "$sysroot"
 echo "Source path         $source_path"
-echo "C compiler          $cc"
+echo "C compiler          $cc ($gcc_major.$gcc_minor)"
 echo "Target OS           $targetos"
 echo "CPU                 $cpu"
 fcho "Triplet             " "$triplet"
diff --git a/lib/Makefile b/lib/Makefile
index 35046a4..3b582d4 100644
--- a/lib/Makefile
+++ b/lib/Makefile
@@ -29,7 +29,7 @@ endif
 # only for native compiler
 $(X)BCHECK_O = bcheck.o
 
-ifeq ($(CONFIG_musl),yes)
+ifeq ($(CONFIG_musl)$(CONFIG_uClibc),yes)
  BCHECK_O =
 endif
 
@@ -61,10 +61,10 @@ $(BIN) : $(patsubst %.o,$(X)%.o,$(OBJ-$T))
 	$(XAR) rcs $@ $^
 
 $(X)%.o : %.c
-	$(XCC) -c $< -o $@ $(DEFINES) $(XFLAGS)
+	$(XCC) -c $< -o $@ $(XFLAGS)
 
 $(X)%.o : %.S
-	$(XCC) -c $< -o $@ $(DEFINES) $(XFLAGS)
+	$(XCC) -c $< -o $@ $(XFLAGS)
 
 $(X)crt1w.o : crt1.c
 $(X)wincrt1w.o : wincrt1.c
diff --git a/lib/alloca86-bt.S b/lib/alloca86-bt.S
index 5215505..4f95cf1 100644
--- a/lib/alloca86-bt.S
+++ b/lib/alloca86-bt.S
@@ -11,7 +11,7 @@ __bound_alloca:
     and     $-4,%eax
     jz      p6
 
-#ifdef TCC_TARGET_PE
+#ifdef _WIN32
 p4:
     cmp     $4096,%eax
     jbe     p5
diff --git a/lib/alloca86.S b/lib/alloca86.S
index ec1e185..bb7a2c2 100644
--- a/lib/alloca86.S
+++ b/lib/alloca86.S
@@ -10,7 +10,7 @@ alloca:
     and     $-4,%eax
     jz      p3
 
-#ifdef TCC_TARGET_PE
+#ifdef _WIN32
 p1:
     cmp     $4096,%eax
     jbe     p2
diff --git a/lib/alloca86_64-bt.S b/lib/alloca86_64-bt.S
index 9b6c3c6..4cbad90 100644
--- a/lib/alloca86_64-bt.S
+++ b/lib/alloca86_64-bt.S
@@ -4,7 +4,7 @@
 .globl __bound_alloca
 __bound_alloca:
 
-#ifdef TCC_TARGET_PE
+#ifdef _WIN32
     # bound checking is not implemented
     pop     %rdx
     mov     %rcx,%rax
diff --git a/lib/alloca86_64.S b/lib/alloca86_64.S
index 4f4dad7..ae3c97d 100644
--- a/lib/alloca86_64.S
+++ b/lib/alloca86_64.S
@@ -5,7 +5,7 @@
 
 alloca:
     pop     %rdx
-#ifdef TCC_TARGET_PE
+#ifdef _WIN32
     mov     %rcx,%rax
 #else
     mov     %rdi,%rax
@@ -14,7 +14,7 @@ alloca:
     and     $-16,%rax
     jz      p3
 
-#ifdef TCC_TARGET_PE
+#ifdef _WIN32
 p1:
     cmp     $4096,%rax
     jbe     p2
diff --git a/lib/bcheck.c b/lib/bcheck.c
index 41c08aa..d0c1ae2 100644
--- a/lib/bcheck.c
+++ b/lib/bcheck.c
@@ -21,10 +21,15 @@
 #include <stdio.h>
 #include <stdarg.h>
 #include <string.h>
-#if !defined(__FreeBSD__) && !defined(__FreeBSD_kernel__) \
-    && !defined(__DragonFly__) && !defined(__OpenBSD__) && !defined(__NetBSD__)
+
+#if !defined(__FreeBSD__) \
+ && !defined(__FreeBSD_kernel__) \
+ && !defined(__DragonFly__) \
+ && !defined(__OpenBSD__) \
+ && !defined(__NetBSD__)
 #include <malloc.h>
 #endif
+
 #if !defined(_WIN32)
 #include <unistd.h>
 #endif
@@ -45,10 +50,13 @@
 #define CONFIG_TCC_MALLOC_HOOKS
 #define HAVE_MEMALIGN
 
-#if defined(__FreeBSD__) || defined(__FreeBSD_kernel__) \
-    || defined(__DragonFly__) || defined(__dietlibc__) \
-    || defined(__UCLIBC__) || defined(__OpenBSD__) || defined(__NetBSD__) \
-    || defined(_WIN32) || defined(TCC_UCLIBC) || defined(TCC_MUSL)
+#if defined(__FreeBSD__) \
+ || defined(__FreeBSD_kernel__) \
+ || defined(__DragonFly__) \
+ || defined(__OpenBSD__) \
+ || defined(__NetBSD__) \
+ || defined(__dietlibc__) \
+ || defined(_WIN32)
 //#warning Bound checking does not support malloc (etc.) in this environment.
 #undef CONFIG_TCC_MALLOC_HOOKS
 #undef HAVE_MEMALIGN
diff --git a/lib/libtcc1.c b/lib/libtcc1.c
index 0702a59..0e46618 100644
--- a/lib/libtcc1.c
+++ b/lib/libtcc1.c
@@ -107,10 +107,10 @@ union float_long {
 };
 
 /* XXX: we don't support several builtin supports for now */
-#if !defined(TCC_TARGET_X86_64) && !defined(TCC_TARGET_ARM)
+#if !defined __x86_64__ && !defined __arm__
 
 /* XXX: use gcc/tcc intrinsic ? */
-#if defined(TCC_TARGET_I386)
+#if defined __i386__
 #define sub_ddmmss(sh, sl, ah, al, bh, bl) \
   __asm__ ("subl %5,%1\n\tsbbl %3,%0"					\
 	   : "=r" ((USItype) (sh)),					\
@@ -589,7 +589,7 @@ long long __fixdfdi (double a1)
     return s ? ret : -ret;
 }
 
-#ifndef TCC_TARGET_ARM
+#ifndef __arm__
 unsigned long long __fixunsxfdi (long double a1)
 {
     register union ldouble_long dl1;
diff --git a/lib/va_list.c b/lib/va_list.c
index 57dce0d..8749f46 100644
--- a/lib/va_list.c
+++ b/lib/va_list.c
@@ -1,6 +1,6 @@
 /* va_list.c - tinycc support for va_list on X86_64 */
 
-#if defined TCC_TARGET_X86_64
+#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);
diff --git a/libtcc.c b/libtcc.c
index c00a909..ec81992 100644
--- a/libtcc.c
+++ b/libtcc.c
@@ -36,7 +36,7 @@ static int nb_states;
 
 /********************************************************/
 
-#ifdef ONE_SOURCE
+#if ONE_SOURCE
 #include "tccpp.c"
 #include "tccgen.c"
 #include "tccelf.c"
@@ -58,6 +58,7 @@ static int nb_states;
 #ifdef TCC_TARGET_C67
 #include "c67-gen.c"
 #include "c67-link.c"
+#include "tcccoff.c"
 #endif
 #ifdef TCC_TARGET_X86_64
 #include "x86_64-gen.c"
@@ -67,9 +68,6 @@ static int nb_states;
 #ifdef CONFIG_TCC_ASM
 #include "tccasm.c"
 #endif
-#ifdef TCC_TARGET_COFF
-#include "tcccoff.c"
-#endif
 #ifdef TCC_TARGET_PE
 #include "tccpe.c"
 #endif
diff --git a/tcc.c b/tcc.c
index bef02e7..4c7235a 100644
--- a/tcc.c
+++ b/tcc.c
@@ -18,10 +18,9 @@
  * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA
  */
 
-#ifdef ONE_SOURCE
-#include "libtcc.c"
-#else
 #include "tcc.h"
+#if ONE_SOURCE
+# include "libtcc.c"
 #endif
 #include "tcctools.c"
 
@@ -335,12 +334,14 @@ redo:
         }
         s->filetype = 0;
         s->alacarte_link = 1;
-        if (ret || --n == 0
+        if (--n == 0 || ret
             || (s->output_type == TCC_OUTPUT_OBJ && !s->option_r))
             break;
     }
 
-    if (s->output_type == TCC_OUTPUT_PREPROCESS) {
+    if (s->run_test) {
+        t = 0;
+    } else if (s->output_type == TCC_OUTPUT_PREPROCESS) {
         ;
     } else if (0 == ret) {
         if (s->output_type == TCC_OUTPUT_MEMORY) {
@@ -357,10 +358,6 @@ redo:
         }
     }
 
-    if (t)
-        ret = 0;
-    if (s->run_test)
-        t = 0;
     if (s->do_bench && (n | t | ret) == 0)
         tcc_print_stats(s, getclock_ms() - start_time);
     tcc_delete(s);
diff --git a/tcc.h b/tcc.h
index 05371bd..bfe6ab4 100644
--- a/tcc.h
+++ b/tcc.h
@@ -81,6 +81,14 @@ extern long double strtold (const char *__nptr, char **__endptr);
 # define O_BINARY 0
 #endif
 
+#ifndef offsetof
+#define offsetof(type, field) ((size_t) &((type *)0)->field)
+#endif
+
+#ifndef countof
+#define countof(tab) (sizeof(tab) / sizeof((tab)[0]))
+#endif
+
 #ifdef _MSC_VER
 # define NORETURN __declspec(noreturn)
 # define ALIGNED(x) __declspec(align(x))
@@ -99,12 +107,6 @@ extern long double strtold (const char *__nptr, char **__endptr);
 # define PATHCMP strcmp
 #endif
 
-#ifdef TCC_TARGET_PE
-#define PATHSEP ';'
-#else
-#define PATHSEP ':'
-#endif
-
 /* -------------------------------------------- */
 
 /* parser debug */
@@ -129,12 +131,20 @@ extern long double strtold (const char *__nptr, char **__endptr);
 #if !defined(TCC_TARGET_I386) && !defined(TCC_TARGET_ARM) && \
     !defined(TCC_TARGET_ARM64) && !defined(TCC_TARGET_C67) && \
     !defined(TCC_TARGET_X86_64)
-#define TCC_TARGET_I386
-#endif
-
-/* object format selection */
-#if defined(TCC_TARGET_C67)
-#define TCC_TARGET_COFF
+# if defined __x86_64__ || defined _AMD64_
+#  define TCC_TARGET_X86_64
+# elif defined __arm__
+#  define TCC_TARGET_ARM
+#  define TCC_ARM_EABI
+#  define TCC_ARM_HARDFLOAT
+# elif defined __aarch64__
+#  define TCC_TARGET_ARM64
+# else
+#  define TCC_TARGET_I386
+# endif
+# ifdef _WIN32
+#  define TCC_TARGET_PE 1
+# endif
 #endif
 
 /* only native compiler supports -run */
@@ -164,7 +174,7 @@ extern long double strtold (const char *__nptr, char **__endptr);
 # define CONFIG_SYSROOT ""
 #endif
 #ifndef CONFIG_TCCDIR
-# define CONFIG_TCCDIR "."
+# define CONFIG_TCCDIR "/usr/local/lib/tcc"
 #endif
 #ifndef CONFIG_LDDIR
 # define CONFIG_LDDIR "lib"
@@ -266,6 +276,12 @@ extern long double strtold (const char *__nptr, char **__endptr);
 #define TCC_LIBGCC USE_TRIPLET(CONFIG_SYSROOT "/" CONFIG_LDDIR) "/libgcc_s.so.1"
 #endif
 
+#ifdef TCC_TARGET_PE
+#define PATHSEP ';'
+#else
+#define PATHSEP ':'
+#endif
+
 /* -------------------------------------------- */
 
 #include "libtcc.h"
@@ -278,7 +294,11 @@ extern long double strtold (const char *__nptr, char **__endptr);
 # define PUB_FUNC
 #endif
 
-#ifdef ONE_SOURCE
+#ifndef ONE_SOURCE
+# define ONE_SOURCE 1
+#endif
+
+#if ONE_SOURCE
 #define ST_INLN static inline
 #define ST_FUNC static
 #define ST_DATA static
@@ -314,6 +334,7 @@ extern long double strtold (const char *__nptr, char **__endptr);
 # include "arm64-link.c"
 #endif
 #ifdef TCC_TARGET_C67
+# define TCC_TARGET_COFF
 # include "coff.h"
 # include "c67-gen.c"
 # include "c67-link.c"
@@ -965,14 +986,6 @@ struct filespec {
 #define TOK_A_SHL 0x81
 #define TOK_A_SAR 0x82
 
-#ifndef offsetof
-#define offsetof(type, field) ((size_t) &((type *)0)->field)
-#endif
-
-#ifndef countof
-#define countof(tab) (sizeof(tab) / sizeof((tab)[0]))
-#endif
-
 #define TOK_EOF       (-1)  /* end of file */
 #define TOK_LINEFEED  10    /* line feed */
 
@@ -1641,7 +1654,7 @@ ST_FUNC void gen_makedeps(TCCState *s, const char *target, const char *filename)
 
 /********************************************************/
 #undef ST_DATA
-#ifdef ONE_SOURCE
+#if ONE_SOURCE
 #define ST_DATA static
 #else
 #define ST_DATA
diff --git a/tccgen.c b/tccgen.c
index 2754192..0828820 100644
--- a/tccgen.c
+++ b/tccgen.c
@@ -1218,12 +1218,16 @@ ST_FUNC int gv(int rc)
     } else {
         if (is_float(vtop->type.t) && 
             (vtop->r & (VT_VALMASK | VT_LVAL)) == VT_CONST) {
+            unsigned long offset;
             /* CPUs usually cannot use float constants, so we store them
                generically in data segment */
             size = type_size(&vtop->type, &align);
-            vpush_ref(&vtop->type, data_section, data_section->data_offset, size);
+            if (NODATA_WANTED)
+                size = 0, align = 1;
+            offset = section_add(data_section, size, align);
+            vpush_ref(&vtop->type, data_section, offset, size);
 	    vswap();
-	    init_putv(&vtop->type, data_section, data_section->data_offset);
+	    init_putv(&vtop->type, data_section, offset);
 	    vtop->r |= VT_LVAL;
         }
 #ifdef CONFIG_TCC_BCHECK
@@ -3518,8 +3522,11 @@ static void struct_layout(CType *type, AttributeDef *ad)
 		/* In PCC layout named bit-fields influence the alignment
 		   of the containing struct using the base types alignment,
 		   except for packed fields (which here have correct align).  */
-		if (f->v & SYM_FIRST_ANOM)
+		if (f->v & SYM_FIRST_ANOM
+                    // && bit_size // ??? gcc on ARM/rpi does that
+                    )
 		    align = 1;
+
 	    } else {
 		bt = f->type.t & VT_BTYPE;
 		if ((bit_pos + bit_size > size * 8)
@@ -3610,7 +3617,7 @@ static void struct_layout(CType *type, AttributeDef *ad)
     if (a < maxalign)
         a = maxalign;
     type->ref->r = a;
-    if (pragma_pack && pragma_pack < maxalign) {
+    if (pragma_pack && pragma_pack < maxalign && 0 == pcc) {
         /* can happen if individual align for some member was given.  In
            this case MSVC ignores maxalign when aligning the size */
         a = pragma_pack;
diff --git a/tccpp.c b/tccpp.c
index 5209b39..4d5169e 100644
--- a/tccpp.c
+++ b/tccpp.c
@@ -49,6 +49,7 @@ static unsigned char isidnum_table[256 - CH_EOF];
 static int pp_debug_tok, pp_debug_symv;
 static int pp_once;
 static int pp_expr;
+static int pp_counter;
 static void tok_print(const char *msg, const int *str);
 
 static struct TinyAlloc *toksym_alloc;
@@ -1409,7 +1410,7 @@ ST_FUNC void label_pop(Sym **ptop, Sym *slast, int keep)
 }
 
 /* fake the nth "#if defined test_..." for tcc -dt -run */
-static void maybe_run_test(TCCState *s, int *c)
+static void maybe_run_test(TCCState *s)
 {
     const char *p;
     if (s->include_stack_ptr != s->include_stack)
@@ -1420,7 +1421,7 @@ static void maybe_run_test(TCCState *s, int *c)
     if (0 != --s->run_test)
         return;
     fprintf(s->ppfp, "\n[%s]\n" + !(s->dflag & 32), p), fflush(s->ppfp);
-    *c = 1;
+    define_push(tok, MACRO_OBJ, NULL, NULL);
 }
 
 /* eval an expression for #if/#elif */
@@ -1440,9 +1441,9 @@ static int expr_preprocess(void)
                 next_nomacro();
             if (tok < TOK_IDENT)
                 expect("identifier");
-            c = define_find(tok) != 0;
             if (tcc_state->run_test)
-                maybe_run_test(tcc_state, &c);
+                maybe_run_test(tcc_state);
+            c = define_find(tok) != 0;
             if (t == '(') {
                 next_nomacro();
                 if (tok != ')')
@@ -3222,18 +3223,17 @@ static int macro_subst_tok(
     Sym *s)
 {
     Sym *args, *sa, *sa1;
-    int parlevel, *mstr, t, t1, spc;
+    int parlevel, t, t1, spc;
     TokenString str;
     char *cstrval;
     CValue cval;
     CString cstr;
     char buf[32];
-    static int counter;
-    
+
     /* if symbol is a macro, prepare substitution */
     /* special macros */
     if (tok == TOK___LINE__ || tok == TOK___COUNTER__) {
-        t = tok == TOK___LINE__ ? file->line_num : counter++;
+        t = tok == TOK___LINE__ ? file->line_num : pp_counter++;
         snprintf(buf, sizeof(buf), "%d", t);
         cstrval = buf;
         t1 = TOK_PPNUM;
@@ -3264,11 +3264,11 @@ static int macro_subst_tok(
         cval.str.data = cstr.data;
         tok_str_add2(tok_str, t1, &cval);
         cstr_free(&cstr);
-    } else {
+    } else if (s->d) {
         int saved_parse_flags = parse_flags;
 	int *joined_str = NULL;
+        int *mstr = s->d;
 
-        mstr = s->d;
         if (s->type.t == MACRO_FUNC) {
             /* whitespace between macro name and argument list */
             TokenString ws_str;
@@ -3523,6 +3523,9 @@ ST_FUNC void preprocess_start(TCCState *s1, int is_asm)
     s1->include_stack_ptr = s1->include_stack;
     s1->ifdef_stack_ptr = s1->ifdef_stack;
     file->ifdef_stack_ptr = s1->ifdef_stack_ptr;
+    pp_expr = 0;
+    pp_counter = 0;
+    pp_debug_tok = pp_debug_symv = 0;
     pp_once++;
     pvtop = vtop = vstack - 1;
     s1->pack_stack[0] = 0;
diff --git a/tests/Makefile b/tests/Makefile
index 00d2c9e..39dee7f 100644
--- a/tests/Makefile
+++ b/tests/Makefile
@@ -55,7 +55,7 @@ ifdef CONFIG_WIN32
  PATH := $(CURDIR)/$(TOP)$(SEP)$(PATH) # for libtcc_test to find libtcc.dll
 endif
 
-RUN_TCC = $(NATIVE_DEFINES) -DONE_SOURCE -run $(TOPSRC)/tcc.c $(TCCFLAGS)
+RUN_TCC = $(NATIVE_DEFINES) -run $(TOPSRC)/tcc.c $(TCCFLAGS)
 DISAS = objdump -d
 DUMPTCC = (set -x; $(TOP)/tcc -vv; ldd $(TOP)/tcc; exit 1)
 
@@ -129,23 +129,23 @@ test4: tcctest.c test.ref
 # use tcc to create libtcc.so/.dll and the tcc(.exe) frontend and run them
 dlltest:
 	@echo ------------ $@ ------------
-	$(TCC) -DONE_SOURCE $(NATIVE_DEFINES) -DLIBTCC_AS_DLL $(TOPSRC)/libtcc.c $(LIBS) -shared -o libtcc2$(DLLSUF)
-	$(TCC) $(NATIVE_DEFINES) $(TOPSRC)/tcc.c libtcc2$(DLLSUF) $(LIBS) -Wl,-rpath=. -o tcc2$(EXESUF)
+	$(TCC) $(NATIVE_DEFINES) -DLIBTCC_AS_DLL $(TOPSRC)/libtcc.c $(LIBS) -shared -o libtcc2$(DLLSUF)
+	$(TCC) $(NATIVE_DEFINES) -DONE_SOURCE=0 $(TOPSRC)/tcc.c libtcc2$(DLLSUF) $(LIBS) -Wl,-rpath=. -o tcc2$(EXESUF)
 	./tcc2$(EXESUF) $(TCCFLAGS) $(RUN_TCC) -run $(TOPSRC)/examples/ex1.c
 ifndef CONFIG_WIN32
 	@echo ------------ $@ with PIC ------------
-	$(CC) $(CFLAGS) -fPIC -DONE_SOURCE $(NATIVE_DEFINES) -DLIBTCC_AS_DLL -c $(TOPSRC)/libtcc.c
+	$(CC) $(CFLAGS) -fPIC $(NATIVE_DEFINES) -DLIBTCC_AS_DLL -c $(TOPSRC)/libtcc.c
 	$(TCC) libtcc.o $(LIBS) -shared -o libtcc2$(DLLSUF)
-	$(TCC) $(NATIVE_DEFINES) $(TOPSRC)/tcc.c libtcc2$(DLLSUF) $(LIBS) -Wl,-rpath=. -o tcc2$(EXESUF)
+	$(TCC) $(NATIVE_DEFINES) -DONE_SOURCE=0 $(TOPSRC)/tcc.c libtcc2$(DLLSUF) $(LIBS) -Wl,-rpath=. -o tcc2$(EXESUF)
 	./tcc2$(EXESUF) $(TCCFLAGS) $(RUN_TCC) -run $(TOPSRC)/examples/ex1.c
 endif
 	@rm tcc2$(EXESUF) libtcc2$(DLLSUF)
 
 memtest:
 	@echo ------------ $@ ------------
-	$(CC) $(CFLAGS) $(NATIVE_DEFINES) -DONE_SOURCE -DMEM_DEBUG=2 $(TOPSRC)/tcc.c $(LIBS) -o memtest-tcc$(EXESUF)
-	./memtest-tcc$(EXESUF) $(TCCFLAGS) $(NATIVE_DEFINES) -DONE_SOURCE $(TOPSRC)/tcc.c $(LIBS)
-	./memtest-tcc$(EXESUF) $(TCCFLAGS) $(NATIVE_DEFINES) -DONE_SOURCE -run $(TOPSRC)/tcc.c $(TCCFLAGS) $(TOPSRC)/tests/tcctest.c
+	$(CC) $(CFLAGS) $(NATIVE_DEFINES) -DMEM_DEBUG=2 $(TOPSRC)/tcc.c $(LIBS) -o memtest-tcc$(EXESUF)
+	./memtest-tcc$(EXESUF) $(TCCFLAGS) $(NATIVE_DEFINES) $(TOPSRC)/tcc.c $(LIBS)
+	./memtest-tcc$(EXESUF) $(TCCFLAGS) $(NATIVE_DEFINES) -run $(TOPSRC)/tcc.c $(TCCFLAGS) $(TOPSRC)/tests/tcctest.c
 
 
 # memory and bound check auto test
@@ -210,7 +210,7 @@ abitest-cc$(EXESUF): abitest.c $(LIBTCC)
 	$(CC) -o $@ $^ $(CFLAGS) $(LIBS) -w
 
 abitest-tcc$(EXESUF): abitest.c libtcc.c
-	$(TCC) -o $@ $^ $(NATIVE_DEFINES) -DONE_SOURCE $(LIBS)
+	$(TCC) -o $@ $^ $(NATIVE_DEFINES) $(LIBS)
 
 ABITESTS := abitest-cc$(EXESUF)
 ifneq ($(CONFIG_arm_eabi),yes) # not ARM soft-float
diff --git a/tests/tests2/95_bitfields.c b/tests/tests2/95_bitfields.c
index 7edbeed..683becf 100644
--- a/tests/tests2/95_bitfields.c
+++ b/tests/tests2/95_bitfields.c
@@ -87,8 +87,10 @@
 # define P
 #endif
 
+printf("\n\n" + 2*top);
 #define TEST 1
 #include SELF
+top = 0;
 #define TEST 2
 #include SELF
 #define TEST 3
@@ -99,7 +101,6 @@
 #include SELF
 #define TEST 6
 #include SELF
-printf("\n\n");
 
 #if PACK
 # pragma pack(pop)
@@ -174,7 +175,7 @@ void dump(void *p, int s)
 
 #define TEST_STRUCT(v1,v2,v3,v4,v5) { \
         struct __s _s, *s = & _s; \
-        printf("---- TEST %d%s%s%s ----\n", \
+        printf("\n---- TEST %d%s%s%s ----\n" + top, \
             TEST, MS_BF?" - MS-BITFIELDS":"", \
             PACK?" - PACKED":"", \
             ALIGN?" - WITH ALIGN":""); \
@@ -184,7 +185,7 @@ void dump(void *p, int s)
         s->x = v1, s->y = v2, s->z = v3, s->a += v4, ++s->a, s->b = v5; \
         printf("bits as set : "), dump(s, sizeof *s); \
         printf("values      :"), pv(x), pv(y), pv(z), pv(a), pv(b), printf("\n"); \
-        printf("align/size  : %d %d\n\n", alignof(struct __s),sizeof(struct __s)); \
+        printf("align/size  : %d %d\n", alignof(struct __s),sizeof(struct __s)); \
     }
 
 #ifdef _MSC_VER
@@ -203,6 +204,8 @@ void dump(void *p, int s)
 
 #define SELF "95_bitfields.c"
 
+int top = 1;
+
 int main()
 {
 #define MS_BF MS_BITFIELDS
diff --git a/tests/tests2/95_bitfields.expect b/tests/tests2/95_bitfields.expect
new file mode 100644
index 0000000..6a8fd9a
--- /dev/null
+++ b/tests/tests2/95_bitfields.expect
@@ -0,0 +1,149 @@
+---- TEST 1 ----
+bits in use : 0000001FFFFFFFFF007F0FFF
+bits as set : 000000076055555500440333
+values      : 333 44 555555 06 07
+align/size  : 4 12
+
+---- TEST 2 ----
+bits in use : 000000000000003F7FFFFFFFFFFFFFFF00000000003F0FFF
+bits as set : 0000000000000025123456789ABCDEF000000000001E0003
+values      : 03 1e 123456789abcdef0 05 fffffffe
+align/size  : 8 24
+
+---- TEST 3 ----
+bits in use : 001F1F1F000003FF
+bits as set : 000E0619000002F5
+values      : 15 17 19 06 0e
+align/size  : 4 8
+
+---- TEST 4 ----
+bits in use : 0007FFFF00000027
+bits as set : 00078F0F00000023
+values      : 03 ffffffff 0f fffffff8 78
+align/size  : 4 8
+
+---- TEST 5 ----
+bits in use : FFFFFF3FFFFFFFFF000000003FFFFFFF00001FFFFFFFFFFF
+bits as set : 007744000000007800000000300000000000000123456789
+values      : 0000000123456789 f0000000 0000000000000078 44 77
+align/size  : 8 24
+
+---- TEST 6 ----
+bits in use : 0000007000FFFFFFFFFFFFFF
+bits as set : 00000030002001FD00000004
+values      : 01 02 03 04 fffffffd
+align/size  : 4 12
+
+
+
+---- TEST 1 - PACKED ----
+bits in use : FFFFFFFFFFFFFF
+bits as set : 3B02AAAAAC4333
+values      : 333 44 555555 06 07
+align/size  : 1 7
+
+---- TEST 2 - PACKED ----
+bits in use : 7FFFFFFFFFFFFFFFFFFFFF
+bits as set : 4A48D159E26AF37BC1E003
+values      : 03 1e 123456789abcdef0 05 fffffffe
+align/size  : 1 11
+
+---- TEST 3 - PACKED ----
+bits in use : 7FFF000003FF
+bits as set : 38D9000002F5
+values      : 15 17 19 06 0e
+align/size  : 1 6
+
+---- TEST 4 - PACKED ----
+bits in use : 07FFFF00000027
+bits as set : 078F0F00000023
+values      : 03 ffffffff 0f fffffff8 78
+align/size  : 1 7
+
+---- TEST 5 - PACKED ----
+bits in use : FFFFFF07FFFFFFFFFFFFFFFF9FFFFFFFFFFF
+bits as set : 007744000000000F18000000000123456789
+values      : 0000000123456789 f0000000 0000000000000078 44 77
+align/size  : 1 18
+
+---- TEST 6 - PACKED ----
+bits in use : 007000FFFFFFFFFFFFFF
+bits as set : 0030002001FD00000004
+values      : 01 02 03 04 fffffffd
+align/size  : 1 10
+
+
+
+---- TEST 1 - WITH ALIGN ----
+bits in use : 000000000000001FFFFFFFFF007F0FFF
+bits as set : 00000000000000076055555500440333
+values      : 333 44 555555 06 07
+align/size  : 16 16
+
+---- TEST 2 - WITH ALIGN ----
+bits in use : 0000000000000000000000000000003F7FFFFFFFFFFFFFFF00000000003F0FFF
+bits as set : 00000000000000000000000000000025123456789ABCDEF000000000001E0003
+values      : 03 1e 123456789abcdef0 05 fffffffe
+align/size  : 16 32
+
+---- TEST 3 - WITH ALIGN ----
+bits in use : 0000000000000000000000000000001F000000000000000000001F1F000003FF
+bits as set : 0000000000000000000000000000000E000000000000000000000619000002F5
+values      : 15 17 19 06 0e
+align/size  : 16 32
+
+---- TEST 4 - WITH ALIGN ----
+bits in use : 0007FFFF00000027
+bits as set : 00078F0F00000023
+values      : 03 ffffffff 0f fffffff8 78
+align/size  : 4 8
+
+---- TEST 5 - WITH ALIGN ----
+bits in use : FFFFFF3FFFFFFFFF000000003FFFFFFF00001FFFFFFFFFFF
+bits as set : 007744000000007800000000300000000000000123456789
+values      : 0000000123456789 f0000000 0000000000000078 44 77
+align/size  : 8 24
+
+---- TEST 6 - WITH ALIGN ----
+bits in use : 0000007000FFFFFFFFFFFFFF
+bits as set : 00000030002001FD00000004
+values      : 01 02 03 04 fffffffd
+align/size  : 4 12
+
+
+
+---- TEST 1 - PACKED - WITH ALIGN ----
+bits in use : 000000000000000000FFFFFFFFFFFFFF
+bits as set : 0000000000000000003B02AAAAAC4333
+values      : 333 44 555555 06 07
+align/size  : 16 16
+
+---- TEST 2 - PACKED - WITH ALIGN ----
+bits in use : 3F01FFFFFFFFFFFFFFFFFFFF
+bits as set : 250048D159E26AF37BC1E003
+values      : 03 1e 123456789abcdef0 05 fffffffe
+align/size  : 1 12
+
+---- TEST 3 - PACKED - WITH ALIGN ----
+bits in use : 1F03FF000003FF
+bits as set : 0E00D9000002F5
+values      : 15 17 19 06 0e
+align/size  : 1 7
+
+---- TEST 4 - PACKED - WITH ALIGN ----
+bits in use : 07FFFF00000027
+bits as set : 078F0F00000023
+values      : 03 ffffffff 0f fffffff8 78
+align/size  : 1 7
+
+---- TEST 5 - PACKED - WITH ALIGN ----
+bits in use : FFFFFF07FFFFFFFFFFFFFFFF9FFFFFFFFFFF
+bits as set : 007744000000000F18000000000123456789
+values      : 0000000123456789 f0000000 0000000000000078 44 77
+align/size  : 1 18
+
+---- TEST 6 - PACKED - WITH ALIGN ----
+bits in use : 007000FFFFFFFFFFFFFF
+bits as set : 0030002001FD00000004
+values      : 01 02 03 04 fffffffd
+align/size  : 1 10
diff --git a/tests/tests2/95_bitfields_ms.expect b/tests/tests2/95_bitfields_ms.expect
index 6b5c3f9..8ccafb7 100644
--- a/tests/tests2/95_bitfields_ms.expect
+++ b/tests/tests2/95_bitfields_ms.expect
@@ -147,6 +147,3 @@ bits in use : 00000000700000FFFFFFFFFFFFFF
 bits as set : 000000003000002001FD00000004
 values      : 01 02 03 04 fffffffd
 align/size  : 1 14
-
-
-
diff --git a/tests/tests2/96_nodata_wanted.c b/tests/tests2/96_nodata_wanted.c
index 1f022c9..b594938 100644
--- a/tests/tests2/96_nodata_wanted.c
+++ b/tests/tests2/96_nodata_wanted.c
@@ -25,50 +25,60 @@ void foo() {
     short w = &foo; /* 2 cast warnings */
 }
 
-#elif defined test_data_suppression
+#elif defined test_data_suppression_off || defined test_data_suppression_on
+
+#if defined test_data_suppression_on
+# define SKIP 1
+#else
+# define SKIP 0
+#endif
+
 #include <stdio.h>
-
-#define ASMLABELS(s) \
-    __asm__(".global d"#s",t"#s"\n.data\nd"#s":\n.text\nt"#s":\n")
-
-#define PROG \
-        static void *p = (void*)&main;\
-        static char cc[] = "static string";\
-        static double d = 8.0;\
-        static struct __attribute__((packed)) {\
-            unsigned x : 12;\
-            unsigned char y : 7;\
-            unsigned z : 28, a: 4, b: 5;\
-        } s = { 0x333,0x44,0x555555,6,7 };\
-        printf("  static data: %d - %.1f - %.1f - %s - %s\n",\
-            sizeof 8.0, 8.0, d, __FUNCTION__, cc);\
-        printf("  static bitfields: %x %x %x %x %x\n", s.x, s.y, s.z, s.a, s.b);
+/* some gcc headers #define __attribute__ to empty if it's not gcc */
+#undef __attribute__
 
 int main()
 {
-    extern char ds1[],ts1[];
-    extern char ds2[],ts2[];
-    extern char de1[],te1[];
-    extern char de2[],te2[];
+    __label__ ts0, te0, ts1, te1;
+    int tl, dl;
 
-    printf("suppression off\n");
-    if (1) {
-        ASMLABELS(s1);
-        PROG
-        ASMLABELS(e1);
-    }
-    printf("  data length is %s\n", de1 - ds1 ? "not 0":"0");
-    printf("  text length is %s\n", te1 - ts1 ? "not 0":"0");
+    static char ds0 = 0;
+    static char de0 = 0;
+    /* get reference size of empty jmp */
+ts0:;
+    if (!SKIP) {}
+te0:;
+    dl = -(&de0 - &ds0);
+    tl = -(&&te0 - &&ts0);
 
-    printf("suppression on\n");
-    if (0) {
-        ASMLABELS(s2);
-        PROG
-        ASMLABELS(e2);
+    /* test data and code suppression */
+    static char ds1 = 0;
+ts1:;
+    if (!SKIP) {
+        static void *p = (void*)&main;
+        static char cc[] = "static string";
+        static double d = 8.0;
+
+        static struct __attribute__((packed)) {
+            unsigned x : 12;
+            unsigned char y : 7;
+            unsigned z : 28, a: 4, b: 5;
+        } s = { 0x333,0x44,0x555555,6,7 };
+
+        printf("data:\n");
+        printf("  %d - %.1f - %.1f - %s - %s\n",
+            sizeof 8.0, 8.0, d, __FUNCTION__, cc);
+        printf("  %x %x %x %x %x\n",
+            s.x, s.y, s.z, s.a, s.b);
     }
-    printf("  data length is %x\n", de2 - ds2);
-    printf("  text length is %X\n", te2 - ts2);
-    return 0;
+te1:;
+    static char de1 = 0;
+
+    dl += &de1 - &ds1;
+    tl += &&te1 - &&ts1;
+    printf("size of data/text:\n  %s/%s\n",
+        dl ? "non-zero":"zero", tl ? "non-zero":"zero");
+    printf("# %d/%d\n", dl, tl);
 }
 
 #endif
diff --git a/tests/tests2/96_nodata_wanted.expect b/tests/tests2/96_nodata_wanted.expect
index 9fd35ca..6e54f24 100644
--- a/tests/tests2/96_nodata_wanted.expect
+++ b/tests/tests2/96_nodata_wanted.expect
@@ -11,12 +11,15 @@
 96_nodata_wanted.c:25: warning: assignment makes integer from pointer without a cast
 96_nodata_wanted.c:25: warning: nonportable conversion from pointer to char/short
 
-[test_data_suppression]
-suppression off
-  static data: 8 - 8.0 - 8.0 - main - static string
-  static bitfields: 333 44 555555 6 7
-  data length is not 0
-  text length is not 0
-suppression on
-  data length is 0
-  text length is 0
+[test_data_suppression_off]
+data:
+  8 - 8.0 - 8.0 - main - static string
+  333 44 555555 6 7
+size of data/text:
+  non-zero/non-zero
+# 111/193
+
+[test_data_suppression_on]
+size of data/text:
+  zero/zero
+# 0/0
diff --git a/tests/tests2/Makefile b/tests/tests2/Makefile
index 4dd1591..c75db8c 100644
--- a/tests/tests2/Makefile
+++ b/tests/tests2/Makefile
@@ -24,7 +24,7 @@ ifeq (-$(findstring gcc,$(CC))-,--)
  SKIP += $(patsubst %.expect,%.test,$(GEN-ALWAYS))
 endif
 ifeq (-$(CONFIG_WIN32)-$(CONFIG_i386)$(CONFIG_arm)-,--yes-)
- SKIP += 95_bitfields_ms.test # type_align is differnt on 32bit-non-windows
+ SKIP += 95_bitfields%.test # type_align is differnt on 32bit-non-windows
 endif
 
 # Some tests might need arguments
@@ -46,7 +46,7 @@ FLAGS =
 
 # Always generate certain .expects (don't put these in the GIT),
 GEN-ALWAYS =
-GEN-ALWAYS += 95_bitfields.expect
+# GEN-ALWAYS += 95_bitfields.expect # does not work
 
 # using the ms compiler for the really ms-compatible bitfields
 95_bitfields_ms.test : GEN = $(GEN-MSC)
@@ -67,7 +67,7 @@ all test tests2.all: $(filter-out $(SKIP),$(TESTS)) ;
 T1 = $(TCC) $(FLAGS) $< -o a.exe && ./a.exe $(ARGS)
 T2 = $(TCC) $(FLAGS) -run $< $(ARGS)
 T3 = $(FILTER) >$*.output 2>&1 || true \
-     && diff -Nbu $(filter %.expect,$^) $*.output \
+     && diff -Nbu -I "^\#" $(filter %.expect,$^) $*.output \
      && rm -f $*.output $(filter $*.expect,$(GEN-ALWAYS))
 
 # run single test and update .expect file, e.g. "make tests2.37+"
diff --git a/win32/build-tcc.bat b/win32/build-tcc.bat
index 9a9184b..b234178 100644
--- a/win32/build-tcc.bat
+++ b/win32/build-tcc.bat
@@ -111,10 +111,10 @@ echo>> ..\config.h #endif
 for %%f in (*tcc.exe *tcc.dll) do @del %%f
 
 :compiler
-%CC% -o libtcc.dll -shared ..\libtcc.c %D% -DONE_SOURCE -DLIBTCC_AS_DLL
+%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%
-%CC% -o %PX%-tcc.exe ..\tcc.c %DX% -DONE_SOURCE
+%CC% -o tcc.exe ..\tcc.c libtcc.dll %D% -DONE_SOURCE=0
+%CC% -o %PX%-tcc.exe ..\tcc.c %DX%
 
 @if (%TCC_FILES%)==(no) goto :files-done
 
@@ -131,30 +131,30 @@ copy>nul tcc-win32.txt doc
 
 :libtcc1.a
 @set O1=libtcc1.o crt1.o crt1w.o wincrt1.o wincrt1w.o dllcrt1.o dllmain.o chkstk.o bcheck.o
-.\tcc -m32 %D32% -c ../lib/libtcc1.c
-.\tcc -m32 %D32% -c lib/crt1.c
-.\tcc -m32 %D32% -c lib/crt1w.c
-.\tcc -m32 %D32% -c lib/wincrt1.c
-.\tcc -m32 %D32% -c lib/wincrt1w.c
-.\tcc -m32 %D32% -c lib/dllcrt1.c
-.\tcc -m32 %D32% -c lib/dllmain.c
-.\tcc -m32 %D32% -c lib/chkstk.S
-.\tcc -m32 %D32% -w -c ../lib/bcheck.c
-.\tcc -m32 %D32% -c ../lib/alloca86.S
-.\tcc -m32 %D32% -c ../lib/alloca86-bt.S
+.\tcc -m32 -c ../lib/libtcc1.c
+.\tcc -m32 -c lib/crt1.c
+.\tcc -m32 -c lib/crt1w.c
+.\tcc -m32 -c lib/wincrt1.c
+.\tcc -m32 -c lib/wincrt1w.c
+.\tcc -m32 -c lib/dllcrt1.c
+.\tcc -m32 -c lib/dllmain.c
+.\tcc -m32 -c lib/chkstk.S
+.\tcc -m32 -w -c ../lib/bcheck.c
+.\tcc -m32 -c ../lib/alloca86.S
+.\tcc -m32 -c ../lib/alloca86-bt.S
 .\tcc -m32 -ar lib/libtcc1-32.a %O1% alloca86.o alloca86-bt.o
 @if errorlevel 1 goto :the_end
-.\tcc -m64 %D64% -c ../lib/libtcc1.c
-.\tcc -m64 %D64% -c lib/crt1.c
-.\tcc -m64 %D64% -c lib/crt1w.c
-.\tcc -m64 %D64% -c lib/wincrt1.c
-.\tcc -m64 %D64% -c lib/wincrt1w.c
-.\tcc -m64 %D64% -c lib/dllcrt1.c
-.\tcc -m64 %D64% -c lib/dllmain.c
-.\tcc -m64 %D64% -c lib/chkstk.S
-.\tcc -m64 %D64% -w -c ../lib/bcheck.c
-.\tcc -m64 %D64% -c ../lib/alloca86_64.S
-.\tcc -m64 %D64% -c ../lib/alloca86_64-bt.S
+.\tcc -m64 -c ../lib/libtcc1.c
+.\tcc -m64 -c lib/crt1.c
+.\tcc -m64 -c lib/crt1w.c
+.\tcc -m64 -c lib/wincrt1.c
+.\tcc -m64 -c lib/wincrt1w.c
+.\tcc -m64 -c lib/dllcrt1.c
+.\tcc -m64 -c lib/dllmain.c
+.\tcc -m64 -c lib/chkstk.S
+.\tcc -m64 -w -c ../lib/bcheck.c
+.\tcc -m64 -c ../lib/alloca86_64.S
+.\tcc -m64 -c ../lib/alloca86_64-bt.S
 .\tcc -m64 -ar lib/libtcc1-64.a %O1% alloca86_64.o alloca86_64-bt.o
 @if errorlevel 1 goto :the_end
 
diff --git a/win32/lib/chkstk.S b/win32/lib/chkstk.S
index 2726061..ec5c07f 100644
--- a/win32/lib/chkstk.S
+++ b/win32/lib/chkstk.S
@@ -2,7 +2,7 @@
 /* chkstk86.s */
 
 /* ---------------------------------------------- */
-#ifndef TCC_TARGET_X86_64
+#ifndef __x86_64__
 /* ---------------------------------------------- */
 
 .globl __chkstk
@@ -68,7 +68,7 @@ tinyc_getbp:
 
 
 /* ---------------------------------------------- */
-#ifndef TCC_TARGET_X86_64
+#ifndef __x86_64__
 /* ---------------------------------------------- */
 
 /*