diff --git a/Makefile b/Makefile
index 22e775c..b229075 100644
--- a/Makefile
+++ b/Makefile
@@ -195,14 +195,14 @@ $(X)arm-gen.o : arm-asm.c
 
 # Host Tiny C Compiler
 tcc$(EXESUF): tcc.o $(LIBTCC)
-	$(CC) -o $@ $^ $(DEFINES) $(CFLAGS) $(LIBS) $(LDFLAGS) $(LINK_LIBTCC)
+	$(CC) -o $@ $^ $(LIBS) $(LDFLAGS) $(LINK_LIBTCC)
 
 # Cross Tiny C Compilers
 %-tcc$(EXESUF): FORCE
 	@$(MAKE) --no-print-directory $@ CROSS_TARGET=$* ONE_SOURCE=$(or $(ONE_SOURCE),yes)
 
 $(CROSS_TARGET)-tcc$(EXESUF): $(TCC_FILES)
-	$(CC) -o $@ $^ $(DEFINES) $(CFLAGS) $(LIBS) $(LDFLAGS)
+	$(CC) -o $@ $^ $(LIBS) $(LDFLAGS)
 
 # profiling version
 tcc_p$(EXESUF): $($T_FILES)
@@ -214,13 +214,14 @@ libtcc.a: $(LIBTCC_OBJ)
 
 # dynamic libtcc library
 libtcc.so: $(LIBTCC_OBJ)
-	$(CC) -shared -Wl,-soname,$@ -o $@ $^ $(CFLAGS) $(LDFLAGS)
+	$(CC) -shared -Wl,-soname,$@ -o $@ $^ $(LDFLAGS)
 
 libtcc.so: CFLAGS+=-fPIC
+libtcc.so: LDFLAGS+=-fPIC
 
 # windows dynamic libtcc library
 libtcc.dll : $(LIBTCC_OBJ)
-	$(CC) -shared -o $@ $^ $(CFLAGS) $(LDFLAGS)
+	$(CC) -shared -o $@ $^ $(LDFLAGS)
 
 libtcc.def : libtcc.dll tcc$(EXESUF)
 	./tcc$(EXESUF) -impdef $< -o $@
@@ -229,12 +230,13 @@ libtcc.dll : DEFINES += -DLIBTCC_AS_DLL
 
 # TinyCC runtime libraries
 libtcc1.a : tcc$(EXESUF) FORCE
-	@$(MAKE) -C lib
+	@$(MAKE) -C lib DEFINES="$(DEF-$T)"
 
 # Cross libtcc1.a
 libtcc1-%.a : %-tcc$(EXESUF) FORCE
-	@$(MAKE) -C lib CROSS_TARGET=$*
+	@$(MAKE) -C lib DEFINES="$(DEF-$*)" CROSS_TARGET=$*
 
+.PRECIOUS: libtcc1-%.a
 FORCE:
 
 # --------------------------------------------------------------------------
diff --git a/arm-gen.c b/arm-gen.c
index 7512b61..a5dabd1 100644
--- a/arm-gen.c
+++ b/arm-gen.c
@@ -179,6 +179,7 @@ ST_FUNC void arm_init(struct TCCState *s)
 #define func_ldouble_type func_old_type
 ST_FUNC void arm_init(struct TCCState *s)
 {
+#if 0
 #if !defined (TCC_ARM_VFP)
     tcc_warning("Support for FPA is deprecated and will be removed in next"
                 " release");
@@ -187,6 +188,7 @@ ST_FUNC void arm_init(struct TCCState *s)
     tcc_warning("Support for OABI is deprecated and will be removed in next"
                 " release");
 #endif
+#endif
 }
 #endif
 
diff --git a/lib/Makefile b/lib/Makefile
index 00ab7dc..fe517e9 100644
--- a/lib/Makefile
+++ b/lib/Makefile
@@ -17,61 +17,54 @@ XFLAGS-win = -B$(TOPSRC)/win32 -I$(TOPSRC)/include
 XFLAGS = $(XFLAGS$(XCFG))
 XCFG = $(or $(findstring -win,$T),-unx)
 
-ifeq ($(X),)
- BCHECK_O = bcheck.o
- ifeq "$T" "arm"
-  XCC = $(CC)
-  XAR = $(AR)
-  XFLAGS = $(CFLAGS) -fPIC
- endif
+# using gcc ("make armlib-usegcc=no" to use tcc)
+armlib-usegcc ?= no
+
+ifeq "$($(X)$(T)lib-usegcc)" "yes"
+ XCC = $(CC)
+ XAR = $(AR)
+ XFLAGS = $(CFLAGS) -fPIC
+endif
+
+# only for native compiler
+$(X)BCHECK_O = bcheck.o
+
+ifeq ($(CONFIG_musl),yes)
+ BCHECK_O =
 endif
 
 ifdef CONFIG_OSX
  XFLAGS += -D_ANSI_SOURCE
 endif
 
-ifeq ($(CONFIG_musl),yes)
- BCHECK_O =
-endif
-
 I386_O = libtcc1.o alloca86.o alloca86-bt.o
 X86_64_O = libtcc1.o alloca86_64.o alloca86_64-bt.o
-ARM_O = libtcc1.o armeabi.o alloca-arm.o
+ARM_O = libtcc1.o armeabi.o alloca-arm.o armflush.o
 ARM64_O = lib-arm64.o
 WIN_O = crt1.o crt1w.o wincrt1.o wincrt1w.o dllcrt1.o dllmain.o
 
 OBJ-i386 = $(I386_O) $(BCHECK_O)
-TGT-i386 = -DTCC_TARGET_I386
-
-OBJ-x86_64 = $(X86_64_O) $(BCHECK_O)
-TGT-x86_64 = -DTCC_TARGET_X86_64
-
-OBJ-x86_64-osx = $(X86_64_O)
-TGT-x86_64-osx = -DTCC_TARGET_X86_64 -DTCC_TARGET_MACHO
-
-OBJ-arm = $(ARM_O)
-TGT-arm = -DTCC_TARGET_ARM
-
-OBJ-arm64 = $(ARM64_O)
-TGT-arm64 = -DTCC_TARGET_ARM64
-
+OBJ-x86_64 = $(X86_64_O) va_list.o $(BCHECK_O)
+OBJ-x86_64-osx = $(X86_64_O) va_list.o
 OBJ-i386-win32 = $(I386_O) chkstk.o bcheck.o $(WIN_O)
-TGT-i386-win32 = -DTCC_TARGET_I386 -DTCC_TARGET_PE
-
 OBJ-x86_64-win32 = $(X86_64_O) chkstk.o bcheck.o $(WIN_O)
-TGT-x86_64-win32 = -DTCC_TARGET_X86_64 -DTCC_TARGET_PE
-
+OBJ-arm64 = $(ARM64_O)
+OBJ-arm = $(ARM_O)
+OBJ-arm-fpa = $(ARM_O)
+OBJ-arm-fpa-ld = $(ARM_O)
+OBJ-arm-vfp = $(ARM_O)
+OBJ-arm-eabi = $(ARM_O)
+OBJ-arm-eabihf = $(ARM_O)
 OBJ-arm-wince = $(ARM_O) $(WIN_O)
-TGT-arm-wince = -DTCC_TARGET_ARM -DTCC_TARGET_PE
 
 $(BIN) : $(patsubst %.o,$(X)%.o,$(OBJ-$T))
 	$(XAR) rcs $@ $^
 
 $(X)%.o : %.c
-	$(XCC) -c $< -o $@ $(TGT-$T) $(XFLAGS)
+	$(XCC) -c $< -o $@ $(DEFINES) $(XFLAGS)
 
 $(X)%.o : %.S
-	$(XCC) -c $< -o $@ $(TGT-$T) $(XFLAGS)
+	$(XCC) -c $< -o $@ $(DEFINES) $(XFLAGS)
 
 $(X)crt1w.o : crt1.c
 $(X)wincrt1w.o : wincrt1.c
diff --git a/lib/armflush.c b/lib/armflush.c
new file mode 100644
index 0000000..eae3260
--- /dev/null
+++ b/lib/armflush.c
@@ -0,0 +1,58 @@
+/* armflush.c - flush the instruction cache
+
+   __clear_cache is used in tccrun.c,  It is a built-in
+   intrinsic with gcc.  However tcc in order to compile
+   itself needs this function */
+
+#ifdef __TINYC__
+
+/* syscall wrapper */
+unsigned syscall(unsigned syscall_nr, ...);
+
+/* arm-tcc supports only fake asm currently */
+__asm__(
+    ".global syscall\n"
+    "syscall:\n"
+    ".int 0xe92d4080\n"  // push    {r7, lr}
+    ".int 0xe1a07000\n"  // mov     r7, r0
+    ".int 0xe1a00001\n"  // mov     r0, r1
+    ".int 0xe1a01002\n"  // mov     r1, r2
+    ".int 0xe1a02003\n"  // mov     r2, r3
+    ".int 0xef000000\n"  // svc     0x00000000
+    ".int 0xe8bd8080\n"  // pop     {r7, pc}
+    );
+
+/* from unistd.h: */
+#if defined(__thumb__) || defined(__ARM_EABI__)
+# define __NR_SYSCALL_BASE      0x0
+#else
+# define __NR_SYSCALL_BASE      0x900000
+#endif
+#define __ARM_NR_BASE           (__NR_SYSCALL_BASE+0x0f0000)
+#define __ARM_NR_cacheflush     (__ARM_NR_BASE+2)
+
+#else
+
+#define _GNU_SOURCE
+#include <unistd.h>
+#include <sys/syscall.h>
+#include <stdio.h>
+
+#endif
+
+/* Flushing for tccrun */
+void __clear_cache(void *beginning, void *end)
+{
+/* __ARM_NR_cacheflush is kernel private and should not be used in user space.
+ * However, there is no ARM asm parser in tcc so we use it for now */
+#if 1
+    syscall(__ARM_NR_cacheflush, beginning, end, 0);
+#else
+    __asm__ ("push {r7}\n\t"
+             "mov r7, #0xf0002\n\t"
+             "mov r2, #0\n\t"
+             "swi 0\n\t"
+             "pop {r7}\n\t"
+             "ret");
+#endif
+}
diff --git a/lib/libtcc1.c b/lib/libtcc1.c
index 9195489..0702a59 100644
--- a/lib/libtcc1.c
+++ b/lib/libtcc1.c
@@ -550,6 +550,13 @@ unsigned long long __fixunssfdi (float a1)
         return 0;
 }
 
+long long __fixsfdi (float a1)
+{
+    long long ret; int s;
+    ret = __fixunssfdi((s = a1 >= 0) ? a1 : -a1);
+    return s ? ret : -ret;
+}
+
 unsigned long long __fixunsdfdi (double a1)
 {
     register union double_long dl1;
@@ -575,6 +582,14 @@ unsigned long long __fixunsdfdi (double a1)
         return 0;
 }
 
+long long __fixdfdi (double a1)
+{
+    long long ret; int s;
+    ret = __fixunsdfdi((s = a1 >= 0) ? a1 : -a1);
+    return s ? ret : -ret;
+}
+
+#ifndef TCC_TARGET_ARM
 unsigned long long __fixunsxfdi (long double a1)
 {
     register union ldouble_long dl1;
@@ -598,121 +613,10 @@ unsigned long long __fixunsxfdi (long double a1)
         return 0;
 }
 
-long long __fixsfdi (float a1)
-{
-    long long ret; int s;
-    ret = __fixunssfdi((s = a1 >= 0) ? a1 : -a1);
-    return s ? ret : -ret;
-}
-
-long long __fixdfdi (double a1)
-{
-    long long ret; int s;
-    ret = __fixunsdfdi((s = a1 >= 0) ? a1 : -a1);
-    return s ? ret : -ret;
-}
-
 long long __fixxfdi (long double a1)
 {
     long long ret; int s;
     ret = __fixunsxfdi((s = a1 >= 0) ? a1 : -a1);
     return s ? ret : -ret;
 }
-
-#if defined(TCC_TARGET_X86_64) && !defined(_WIN64)
-
-#ifndef __TINYC__
-# include <stdlib.h>
-# include <stdio.h>
-# include <string.h>
-# undef __va_start
-# undef __va_arg
-# undef __va_copy
-# undef __va_end
-#else
-/* 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);
-#endif
-
-/* This should be in sync with our include/stdarg.h */
-enum __va_arg_type {
-    __va_gen_reg, __va_float_reg, __va_stack
-};
-
-/* 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;
-
-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,
-               enum __va_arg_type arg_type,
-               int size, int align)
-{
-    size = (size + 7) & ~7;
-    align = (align + 7) & ~7;
-    switch (arg_type) {
-    case __va_gen_reg:
-        if (ap->gp_offset + size <= 48) {
-            ap->gp_offset += size;
-            return ap->reg_save_area + ap->gp_offset - size;
-        }
-        goto use_overflow_area;
-
-    case __va_float_reg:
-        if (ap->fp_offset < 128 + 48) {
-            ap->fp_offset += 16;
-            return ap->reg_save_area + ap->fp_offset - 16;
-        }
-        size = 8;
-        goto use_overflow_area;
-
-    case __va_stack:
-    use_overflow_area:
-        ap->overflow_arg_area += size;
-        ap->overflow_arg_area = (char*)((long long)(ap->overflow_arg_area + align - 1) & -align);
-        return ap->overflow_arg_area - size;
-
-    default: /* should never happen */
-        abort();
-    }
-}
-#endif /* __x86_64__ */
-
-#if defined TCC_TARGET_ARM && !defined __TINYC__
-#define _GNU_SOURCE
-#include <unistd.h>
-#include <sys/syscall.h>
-#include <stdio.h>
-
-/* Flushing for tccrun */
-void __clear_cache(void *beginning, void *end)
-{
-/* __ARM_NR_cacheflush is kernel private and should not be used in user space.
- * However, there is no ARM asm parser in tcc so we use it for now */
-#if 1
-    syscall(__ARM_NR_cacheflush, beginning, end, 0);
-#else
-    __asm__ ("push {r7}\n\t"
-             "mov r7, #0xf0002\n\t"
-             "mov r2, #0\n\t"
-             "swi 0\n\t"
-             "pop {r7}\n\t"
-             "ret");
-#endif
-}
-#endif /* arm */
+#endif /* !ARM */
diff --git a/lib/va_list.c b/lib/va_list.c
new file mode 100644
index 0000000..57dce0d
--- /dev/null
+++ b/lib/va_list.c
@@ -0,0 +1,65 @@
+/* va_list.c - tinycc support for va_list on X86_64 */
+
+#if defined TCC_TARGET_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 */
+enum __va_arg_type {
+    __va_gen_reg, __va_float_reg, __va_stack
+};
+
+/* 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;
+
+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,
+               enum __va_arg_type arg_type,
+               int size, int align)
+{
+    size = (size + 7) & ~7;
+    align = (align + 7) & ~7;
+    switch (arg_type) {
+    case __va_gen_reg:
+        if (ap->gp_offset + size <= 48) {
+            ap->gp_offset += size;
+            return ap->reg_save_area + ap->gp_offset - size;
+        }
+        goto use_overflow_area;
+
+    case __va_float_reg:
+        if (ap->fp_offset < 128 + 48) {
+            ap->fp_offset += 16;
+            return ap->reg_save_area + ap->fp_offset - 16;
+        }
+        size = 8;
+        goto use_overflow_area;
+
+    case __va_stack:
+    use_overflow_area:
+        ap->overflow_arg_area += size;
+        ap->overflow_arg_area = (char*)((long long)(ap->overflow_arg_area + align - 1) & -align);
+        return ap->overflow_arg_area - size;
+
+    default: /* should never happen */
+        abort();
+    }
+}
+#endif
diff --git a/tccgen.c b/tccgen.c
index 424c4e0..3bcfa94 100644
--- a/tccgen.c
+++ b/tccgen.c
@@ -4578,9 +4578,12 @@ ST_FUNC void unary(void)
 #ifdef TCC_TARGET_PE
     case TOK_builtin_va_start:
 	parse_builtin_params(0, "ee");
-	if ((vtop->r & VT_VALMASK) != VT_LOCAL)
-	  tcc_error("__builtin_va_start expects a local variable");
-	vtop->r &= ~VT_LVAL;
+        r = vtop->r & VT_VALMASK;
+        if (r == VT_LLOCAL)
+            r = VT_LOCAL;
+        if (r != VT_LOCAL)
+            tcc_error("__builtin_va_start expects a local variable");
+        vtop->r = r;
 	vtop->type = char_pointer_type;
 	vtop->c.i += 8;
 	vstore();
@@ -6169,7 +6172,11 @@ static void init_putv(CType *type, Section *sec, unsigned long c)
                 if (sizeof(long double) == LDOUBLE_SIZE)
 		    *(long double *)ptr = vtop->c.ld;
 		else if (sizeof(double) == LDOUBLE_SIZE)
-		    *(double *)ptr = vtop->c.ld;
+		    *(double *)ptr = (double)vtop->c.ld;
+#if (defined __i386__ || defined __x86_64__) && (defined TCC_TARGET_I386 || defined TCC_TARGET_X86_64)
+                else if (sizeof (long double) >= 10)
+                    memcpy(memset(ptr, 0, LDOUBLE_SIZE), &vtop->c.ld, 10);
+#endif
 		else
                     tcc_error("can't cross compile long double constants");
 		break;
diff --git a/tests/abitest.c b/tests/abitest.c
index 896e97b..580e921 100644
--- a/tests/abitest.c
+++ b/tests/abitest.c
@@ -637,10 +637,13 @@ int main(int argc, char **argv) {
   RUN_TEST(ret_longdouble_test);
   RUN_TEST(ret_2float_test);
   RUN_TEST(ret_2double_test);
-  /* RUN_TEST(ret_8plus2double_test); currently broken on x86_64 */
-  /* RUN_TEST(ret_6plus2longlong_test); currently broken on x86_64 */
-  /* RUN_TEST(ret_mixed_test); currently broken on x86_64 */
-  /* RUN_TEST(ret_mixed2_test); currently broken on x86_64 */
+#if !defined __x86_64__ || defined _WIN32
+  /* currently broken on x86_64 linux */
+  RUN_TEST(ret_8plus2double_test);
+  RUN_TEST(ret_6plus2longlong_test);
+  RUN_TEST(ret_mixed_test);
+  RUN_TEST(ret_mixed2_test);
+#endif
   RUN_TEST(ret_mixed3_test);
   RUN_TEST(reg_pack_test);
   RUN_TEST(reg_pack_longlong_test);
diff --git a/tests/tcctest.c b/tests/tcctest.c
index 3a4860a..bc08968 100644
--- a/tests/tcctest.c
+++ b/tests/tcctest.c
@@ -2044,13 +2044,6 @@ void bitfield_test(void)
     else 
         printf("st1.f2 != -1\n");
 
-#ifndef __i386__
-    /* on i386 we don't correctly support long long bit-fields.
-       The bitfields can straddle long long boundaries (at least with
-       GCC bitfield layout) and code generation isn't prepared for this
-       (would have to work with two words in that case).  */
-    /* bit sizes below must be bigger than 32 since GCC doesn't allow
-       long-long bitfields whose size is not bigger than int */
     struct sbf2 {
         long long f1 : 45;
         long long : 2;
@@ -2063,7 +2056,7 @@ void bitfield_test(void)
     st2.f3 = a;
     st2.f2++;
     printf("%lld %lld %lld\n", st2.f1, st2.f2, st2.f3);
-#endif
+
 #if 0
     Disabled for now until further clarification re GCC compatibility
     struct sbf3 {
@@ -2076,6 +2069,7 @@ void bitfield_test(void)
     } st3;
     printf("sizeof(st3) = %d\n", sizeof(st3));
 #endif
+
     struct sbf4 {
 	int x : 31;
 	char y : 2;