mpy-cross: Set number of registers in nlr_buf_t based on native arch.

Fixes micropython#5059.  Done in collaboration with Jim Mussared.

Author: dpgeorge

mpy-cross/main.c

@@ -209,12 +209,16 @@ MP_NOINLINE int main_(int argc, char **argv) {
     mp_dynamic_compiler.py_builtins_str_unicode = 1;
     #if defined(__i386__)
     mp_dynamic_compiler.native_arch = MP_NATIVE_ARCH_X86;
+    mp_dynamic_compiler.nlr_buf_num_regs = MICROPY_NLR_NUM_REGS_X86;
     #elif defined(__x86_64__)
     mp_dynamic_compiler.native_arch = MP_NATIVE_ARCH_X64;
+    mp_dynamic_compiler.nlr_buf_num_regs = MAX(MICROPY_NLR_NUM_REGS_X64, MICROPY_NLR_NUM_REGS_X64_WIN);
     #elif defined(__arm__) && !defined(__thumb2__)
     mp_dynamic_compiler.native_arch = MP_NATIVE_ARCH_ARMV6;
+    mp_dynamic_compiler.nlr_buf_num_regs = MICROPY_NLR_NUM_REGS_ARM_THUMB_FP;
     #else
     mp_dynamic_compiler.native_arch = MP_NATIVE_ARCH_NONE;
+    mp_dynamic_compiler.nlr_buf_num_regs = 0;
     #endif
 
     const char *input_file = NULL;
@@ -271,14 +275,19 @@ MP_NOINLINE int main_(int argc, char **argv) {
                 const char *arch = argv[a] + sizeof("-march=") - 1;
                 if (strcmp(arch, "x86") == 0) {
                     mp_dynamic_compiler.native_arch = MP_NATIVE_ARCH_X86;
+                    mp_dynamic_compiler.nlr_buf_num_regs = MICROPY_NLR_NUM_REGS_X86;
                 } else if (strcmp(arch, "x64") == 0) {
                     mp_dynamic_compiler.native_arch = MP_NATIVE_ARCH_X64;
+                    mp_dynamic_compiler.nlr_buf_num_regs = MAX(MICROPY_NLR_NUM_REGS_X64, MICROPY_NLR_NUM_REGS_X64_WIN);
                 } else if (strcmp(arch, "armv6") == 0) {
                     mp_dynamic_compiler.native_arch = MP_NATIVE_ARCH_ARMV6;
+                    mp_dynamic_compiler.nlr_buf_num_regs = MICROPY_NLR_NUM_REGS_ARM_THUMB_FP;
                 } else if (strcmp(arch, "armv7m") == 0) {
                     mp_dynamic_compiler.native_arch = MP_NATIVE_ARCH_ARMV7M;
+                    mp_dynamic_compiler.nlr_buf_num_regs = MICROPY_NLR_NUM_REGS_ARM_THUMB_FP;
                 } else if (strcmp(arch, "xtensa") == 0) {
                     mp_dynamic_compiler.native_arch = MP_NATIVE_ARCH_XTENSA;
+                    mp_dynamic_compiler.nlr_buf_num_regs = MICROPY_NLR_NUM_REGS_XTENSA;
                 } else {
                     return usage(argv);
                 }

py/emitnative.c

@@ -81,7 +81,11 @@
 //                              (L0-L2 may be in regs instead)
 
 // Native emitter needs to know the following sizes and offsets of C structs (on the target):
+#if MICROPY_DYNAMIC_COMPILER
+#define SIZEOF_NLR_BUF (2 + mp_dynamic_compiler.nlr_buf_num_regs + 1) // the +1 is conservative in case MICROPY_ENABLE_PYSTACK enabled
+#else
 #define SIZEOF_NLR_BUF (sizeof(nlr_buf_t) / sizeof(uintptr_t))
+#endif
 #define SIZEOF_CODE_STATE (sizeof(mp_code_state_t) / sizeof(uintptr_t))
 #define OFFSETOF_CODE_STATE_STATE (offsetof(mp_code_state_t, state) / sizeof(uintptr_t))
 #define OFFSETOF_CODE_STATE_FUN_BC (offsetof(mp_code_state_t, fun_bc) / sizeof(uintptr_t))

py/mpstate.h

@@ -47,6 +47,7 @@ typedef struct mp_dynamic_compiler_t {
     bool opt_cache_map_lookup_in_bytecode;
     bool py_builtins_str_unicode;
     uint8_t native_arch;
+    uint8_t nlr_buf_num_regs;
 } mp_dynamic_compiler_t;
 extern mp_dynamic_compiler_t mp_dynamic_compiler;
 #endif