aarch64: Save and restore VFP registers on update_mem_regions()

It seems to overwrite aarch64 VFP registers while recording args in
string format.

Here is the test program and the preparation.

  $ cat float.c
  #include <stdio.h>

  float float_add(float a, float b)
  {
          return a + b;
  }

  int main(int argc, char *argv[])
  {
          double c;

          c = float_add(-0.1, 0.2);
          fprintf(stderr, "%lf\n", c);

          return c > 0;
  }

  $ gcc -pg float.c

  $ uftrace -A fprintf@arg2 -F main a.out
  0.100000
  # DURATION     TID     FUNCTION
              [  1047] | main() {
     1.198 us [  1047] |   float_add();
    95.104 us [  1047] |   fprintf(0xaaaab814ca48);
   109.843 us [  1047] | } /* main */

However, the return value of 'float_add' is overwritten by some garbage
value if uftrace records a function in string format.

  $ uftrace -A fprintf@arg2/s -F main a.out
  121362470815289747852695694937146665636545397707008516 ... 912.000000
  # DURATION     TID     FUNCTION
              [  1058] | main() {
     1.250 us [  1058] |   float_add();
     1.278 ms [  1058] |   fprintf("%lf\n");
     1.292 ms [  1058] | } /* main */

This patch protects the context by saving d0-d7 registers before calling
it, and it makes the result correct.

Signed-off-by: Honggyu Kim <honggyu.kp@gmail.com>

arch/aarch64/mcount-arch.h

@@ -26,7 +26,9 @@ struct mcount_regs {
 #define ARCH_MAX_REG_ARGS  8
 #define ARCH_MAX_FLOAT_REGS  8
 
+#define HAVE_MCOUNT_ARCH_CONTEXT
 struct mcount_arch_context {
+	double d[ARCH_MAX_FLOAT_REGS];
 };
 
 #define ARCH_PLT0_SIZE  32

arch/aarch64/mcount-support.c

@@ -193,3 +193,27 @@ unsigned long mcount_arch_plthook_addr(struct plthook_data *pd, int idx)
 	sym = &pd->dsymtab.sym[0];
 	return sym->addr - ARCH_PLT0_SIZE;
 }
+
+void mcount_save_arch_context(struct mcount_arch_context *ctx)
+{
+	asm volatile ("str d0, %0\n" : "=m" (ctx->d[0]));
+	asm volatile ("str d1, %0\n" : "=m" (ctx->d[1]));
+	asm volatile ("str d2, %0\n" : "=m" (ctx->d[2]));
+	asm volatile ("str d3, %0\n" : "=m" (ctx->d[3]));
+	asm volatile ("str d4, %0\n" : "=m" (ctx->d[4]));
+	asm volatile ("str d5, %0\n" : "=m" (ctx->d[5]));
+	asm volatile ("str d6, %0\n" : "=m" (ctx->d[6]));
+	asm volatile ("str d7, %0\n" : "=m" (ctx->d[7]));
+}
+
+void mcount_restore_arch_context(struct mcount_arch_context *ctx)
+{
+	asm volatile ("ldr d0, %0\n" :: "m" (ctx->d[0]));
+	asm volatile ("ldr d1, %0\n" :: "m" (ctx->d[1]));
+	asm volatile ("ldr d2, %0\n" :: "m" (ctx->d[2]));
+	asm volatile ("ldr d3, %0\n" :: "m" (ctx->d[3]));
+	asm volatile ("ldr d4, %0\n" :: "m" (ctx->d[4]));
+	asm volatile ("ldr d5, %0\n" :: "m" (ctx->d[5]));
+	asm volatile ("ldr d6, %0\n" :: "m" (ctx->d[6]));
+	asm volatile ("ldr d7, %0\n" :: "m" (ctx->d[7]));
+}