py/nlrthumb: Save and restore VFP registers s16-s21 when CPU has them.

These s16-s21 registers are used by gcc so need to be saved.  Future
versions of gcc (beyond v9.1.0), or other compilers, may eventually need
additional registers saved/restored.

See issue #4844.

py/nlr.h

@@ -54,7 +54,14 @@
         #endif
     #elif defined(__thumb2__) || defined(__thumb__) || defined(__arm__)
         #define MICROPY_NLR_THUMB (1)
-        #define MICROPY_NLR_NUM_REGS (10)
+        #if defined(__SOFTFP__)
+            #define MICROPY_NLR_NUM_REGS (10)
+        #else
+            // With hardware FP registers s16-s31 are callee save so in principle
+            // should be saved and restored by the NLR code.  gcc only uses s16-s21
+            // so only save/restore those as an optimisation.
+            #define MICROPY_NLR_NUM_REGS (10 + 6)
+        #endif
     #elif defined(__xtensa__)
         #define MICROPY_NLR_XTENSA (1)
         #define MICROPY_NLR_NUM_REGS (10)

py/nlrthumb.c

@@ -63,6 +63,11 @@ __attribute__((naked)) unsigned int nlr_push(nlr_buf_t *nlr) {
     "str    r10, [r0, #36]      \n" // store r10 into nlr_buf
     "str    r11, [r0, #40]      \n" // store r11 into nlr_buf
     "str    r13, [r0, #44]      \n" // store r13=sp into nlr_buf
+    #if MICROPY_NLR_NUM_REGS == 16
+    "vstr   d8, [r0, #48]       \n" // store s16-s17 into nlr_buf
+    "vstr   d9, [r0, #56]       \n" // store s18-s19 into nlr_buf
+    "vstr   d10, [r0, #64]      \n" // store s20-s21 into nlr_buf
+    #endif
     "str    lr, [r0, #8]        \n" // store lr into nlr_buf
 #endif
 
@@ -118,6 +123,11 @@ NORETURN void nlr_jump(void *val) {
     "ldr    r10, [r0, #36]      \n" // load r10 from nlr_buf
     "ldr    r11, [r0, #40]      \n" // load r11 from nlr_buf
     "ldr    r13, [r0, #44]      \n" // load r13=sp from nlr_buf
+    #if MICROPY_NLR_NUM_REGS == 16
+    "vldr   d8, [r0, #48]       \n" // load s16-s17 from nlr_buf
+    "vldr   d9, [r0, #56]       \n" // load s18-s19 from nlr_buf
+    "vldr   d10, [r0, #64]      \n" // load s20-s21 from nlr_buf
+    #endif
     "ldr    lr, [r0, #8]        \n" // load lr from nlr_buf
 #endif
     "movs   r0, #1              \n" // return 1, non-local return