unix: Add setjmp-based GC register helper implementation.

As I suspected for a long time, for x86, register helper doesn't really make
any difference - there's simply not enough register to keep anything in
them for any prolonged time. Anything gets pushed on stack anyway. So, on
x86, uPy passed all tests even with empty reg helper. So, this setjmp
implementation goes as "untested".

unix/gccollect.c

@@ -34,6 +34,17 @@
 
 extern void *stack_top;
 
+#if MICROPY_GCREGS_SETJMP
+#include <setjmp.h>
+
+typedef jmp_buf regs_t;
+
+void gc_helper_get_regs(regs_t arr) {
+    setjmp(arr);
+}
+
+#else // !MICROPY_GCREGS_SETJMP
+
 // We capture here callee-save registers, i.e. ones which may contain
 // interesting values held there by our callers. It doesn't make sense
 // to capture caller-saved registers, because they, well, put on the
@@ -112,6 +123,7 @@ void gc_helper_get_regs(regs_t arr) {
     arr[9] = r13;
 }
 #endif
+#endif // !MICROPY_GCREGS_SETJMP
 
 void gc_collect(void) {
     //gc_dump_info();

unix/mpconfigport.h

@@ -49,6 +49,9 @@
 // Define to MICROPY_ERROR_REPORTING_DETAILED to get function, etc.
 // names in exception messages (may require more RAM).
 #define MICROPY_ERROR_REPORTING     (MICROPY_ERROR_REPORTING_DETAILED)
+// Define to 1 to use untested inefficient GC helper implementation
+// (if more efficient arch-specific one is not available).
+#define MICROPY_GCREGS_SETJMP       (0)
 
 extern const struct _mp_obj_module_t mp_module_os;
 extern const struct _mp_obj_module_t mp_module_time;