py/nlrx64: Mark nlr_push() as naked function when possible.

Supported from GCC 8 and up, and Compiler Explorer suggests it works as
expected with Clang since 3.6 (2014).

- Fixes situation where building embedded MicroPython with -O0 and
  MICROPY_NLR_X64 crashes at runtime (due to nlr_push pushing the
  frame pointer register EBP). Closes #12421.

- Allows removing the macOS tweak to undo pushing EBP onto the stack
  in the generated function prelude.

Signed-off-by: Angus Gratton <angus@redyak.com.au>

py/nlrx64.c

@@ -35,8 +35,27 @@
 
 __attribute__((used)) unsigned int nlr_push_tail(nlr_buf_t *nlr);
 
+#if !MICROPY_NLR_OS_WINDOWS
+#if defined(__clang__) || (defined(__GNUC__) && __GNUC__ >= 8)
+#define USE_NAKED 1
+#else
+// On older gcc the equivalent here is to force omit-frame-pointer
+__attribute__((optimize("omit-frame-pointer")))
+#endif
+#endif
+
+#if !defined(USE_NAKED)
+#define USE_NAKED 0
+#endif
+
+#if USE_NAKED
+// nlr_push prelude should never push frame pointer register ebp onto the stack
+__attribute__((naked))
+#endif
 unsigned int nlr_push(nlr_buf_t *nlr) {
+    #if !USE_NAKED
     (void)nlr;
+    #endif
 
     #if MICROPY_NLR_OS_WINDOWS
 
@@ -58,9 +77,6 @@ unsigned int nlr_push(nlr_buf_t *nlr) {
     #else
 
     __asm volatile (
-        #if defined(__APPLE__) && defined(__MACH__)
-        "pop    %rbp                \n" // undo function's prelude
-        #endif
         "movq   (%rsp), %rax        \n" // load return %rip
         "movq   %rax, 16(%rdi)      \n" // store %rip into nlr_buf
         "movq   %rbp, 24(%rdi)      \n" // store %rbp into nlr_buf
@@ -79,7 +95,9 @@ unsigned int nlr_push(nlr_buf_t *nlr) {
 
     #endif
 
+    #if !USE_NAKED
     return 0; // needed to silence compiler warning
+    #endif
 }
 
 NORETURN void nlr_jump(void *val) {