arm: Move bootstage record for board_init_f() to after arch_cpu_init()

The timer may be inited in arch_cpu_init() so it is not safe to make a
bootstage mark before this is called. Arrange the code to fix this.

Note: The question was raised as to why we don't keep all archs in sync.
PowerPC doesn't have specific bootstage markers at present (although it
does use boot progress). I hope that the generic board series will solve
this problem in general, but in the meantime this is a real problem, and
only in ARM.

We now get a correct time for board_init_f:

Timer summary in microseconds:
       Mark    Elapsed  Stage
          0          0  reset
    100,000    100,000  spl_start
    848,530    748,530  board_init_f
    907,301     58,771  board_init_r
    910,478      3,177  board_init

Signed-off-by: Simon Glass <sjg@chromium.org>
diff --git a/arch/arm/lib/board.c b/arch/arm/lib/board.c
index ec5c35f..404ef4d 100644
--- a/arch/arm/lib/board.c
+++ b/arch/arm/lib/board.c
@@ -232,8 +232,17 @@
 int power_init_board(void)
 	__attribute__((weak, alias("__power_init_board")));
 
+	/* Record the board_init_f() bootstage (after arch_cpu_init()) */
+static int mark_bootstage(void)
+{
+	bootstage_mark_name(BOOTSTAGE_ID_START_UBOOT_F, "board_init_f");
+
+	return 0;
+}
+
 init_fnc_t *init_sequence[] = {
 	arch_cpu_init,		/* basic arch cpu dependent setup */
+	mark_bootstage,
 #ifdef CONFIG_OF_CONTROL
 	fdtdec_check_fdt,
 #endif
@@ -277,8 +286,6 @@
 	void *new_fdt = NULL;
 	size_t fdt_size = 0;
 
-	bootstage_mark_name(BOOTSTAGE_ID_START_UBOOT_F, "board_init_f");
-
 	memset((void *)gd, 0, sizeof(gd_t));
 
 	gd->mon_len = _bss_end_ofs;