vm: Make the speed-size trade-off compile time settable

.. and enable for all samd21 boards

ports/atmel-samd/mpconfigport.mk

@@ -36,6 +36,7 @@ CIRCUITPY_AUDIOMIXER ?= 0
 CIRCUITPY_BINASCII ?= 0
 CIRCUITPY_AUDIOMP3 ?= 0
 CIRCUITPY_BUILTINS_POW3 ?= 0
+CIRCUITPY_COMPUTED_GOTO_SAVE_SPACE ?= 1
 CIRCUITPY_FREQUENCYIO ?= 0
 CIRCUITPY_JSON ?= 0
 CIRCUITPY_TOUCHIO_USE_NATIVE ?= 1

py/circuitpy_mpconfig.h

@@ -75,6 +75,7 @@
 #define MICROPY_MODULE_BUILTIN_INIT      (1)
 #define MICROPY_NONSTANDARD_TYPECODES    (0)
 #define MICROPY_OPT_COMPUTED_GOTO        (1)
+#define MICROPY_OPT_COMPUTED_GOTO_SAVE_SPACE (CIRCUITPY_COMPUTED_GOTO_SAVE_SPACE)
 #define MICROPY_PERSISTENT_CODE_LOAD     (1)
 
 #define MICROPY_PY_ARRAY                 (1)

py/circuitpy_mpconfig.mk

@@ -130,6 +130,9 @@ CFLAGS += -DCIRCUITPY_CANIO=$(CIRCUITPY_CANIO)
 CIRCUITPY_DIGITALIO ?= 1
 CFLAGS += -DCIRCUITPY_DIGITALIO=$(CIRCUITPY_DIGITALIO)
 
+CIRCUITPY_COMPUTED_GOTO_SAVE_SPACE ?= 0
+CFLAGS += -DCIRCUITPY_COMPUTED_GOTO_SAVE_SPACE=$(CIRCUITPY_COMPUTED_GOTO_SAVE_SPACE)
+
 CIRCUITPY_COUNTIO ?= $(CIRCUITPY_FULL_BUILD)
 CFLAGS += -DCIRCUITPY_COUNTIO=$(CIRCUITPY_COUNTIO)
 

py/mpconfig.h

@@ -413,6 +413,14 @@
 #define MICROPY_OPT_COMPUTED_GOTO (0)
 #endif
 
+// Whether to save trade flash space for speed in MICROPY_OPT_COMPUTED_GOTO.
+// Costs about 3% speed, saves about 1500 bytes space.  In addition to the assumptions
+// of MICROPY_OPT_COMPUTED_GOTO, also assumes that mp_execute_bytecode is less than
+// 32kB in size.
+#ifndef MICROPY_OPT_COMPUTED_GOTO_SAVE_SPACE
+#define MICROPY_OPT_COMPUTED_GOTO_SAVE_SPACE (0)
+#endif
+
 // Whether to cache result of map lookups in LOAD_NAME, LOAD_GLOBAL, LOAD_ATTR,
 // STORE_ATTR bytecodes.  Uses 1 byte extra RAM for each of these opcodes and
 // uses a bit of extra code ROM, but greatly improves lookup speed.

py/vm.c

@@ -129,12 +129,21 @@ mp_vm_return_kind_t PLACE_IN_ITCM(mp_execute_bytecode)(mp_code_state_t *code_sta
 #endif
 #if MICROPY_OPT_COMPUTED_GOTO
     #include "py/vmentrytable.h"
+#if MICROPY_OPT_COMPUTED_GOTO_SAVE_SPACE
     #define ONE_TRUE_DISPATCH() one_true_dispatch: do { \
         TRACE(ip); \
         MARK_EXC_IP_GLOBAL(); \
         goto *(void*)((char*)&&entry_MP_BC_LOAD_CONST_FALSE + entry_table[*ip++]); \
     } while (0)
     #define DISPATCH() do { goto one_true_dispatch; } while(0)
+#else
+    #define DISPATCH() do { \
+        TRACE(ip); \
+        MARK_EXC_IP_GLOBAL(); \
+        goto *entry_table[*ip++]; \
+    } while (0)
+    #define ONE_TRUE_DISPATCH() DISPATCH()
+#endif
     #define DISPATCH_WITH_PEND_EXC_CHECK() goto pending_exception_check
     #define ENTRY(op) entry_##op
     #define ENTRY_DEFAULT entry_default

py/vmentrytable.h

@@ -31,9 +31,15 @@
 
 #include "supervisor/linker.h"
 
+#if MICROPY_OPT_COMPUTED_GOTO_SAVE_SPACE
 #define COMPUTE_ENTRY(x) ((char*)(x) - (char*)&&entry_MP_BC_LOAD_CONST_FALSE)
+typedef int16_t entry_table_type;
+#else
+#define COMPUTE_ENTRY(x) (x)
+typedef void *entry_table_type;
+#endif
 
-static int16_t const PLACE_IN_DTCM_DATA(entry_table[256]) = {
+static entry_table_type const PLACE_IN_DTCM_DATA(entry_table[256]) = {
     [0 ... 255] = COMPUTE_ENTRY(&&entry_default),
     [MP_BC_LOAD_CONST_FALSE] = COMPUTE_ENTRY(&&entry_MP_BC_LOAD_CONST_FALSE),
     [MP_BC_LOAD_CONST_NONE] = COMPUTE_ENTRY(&&entry_MP_BC_LOAD_CONST_NONE),