Blackfin: turn off caches when self initializing

When bootstrapping ourselves on the fly at runtime (via "go"), we need to
turn off the caches to avoid taking software exceptions.  Since caches
need CPLBs and CPLBs need exception handlers, but we're about to rewrite
the code in memory where those exception handlers live, we need to turn
off caches first.

This new code also encourages a slight code optimization by storing the
MMR bases in dedicated registers so we don't have to fully load up the
pointer regs multiple times.

Signed-off-by: Mike Frysinger <vapier@gentoo.org>

arch/blackfin/cpu/start.S

@@ -52,6 +52,19 @@ ENTRY(_start)
 	sp.l = LO(L1_SRAM_SCRATCH_END - 20);
 	sp.h = HI(L1_SRAM_SCRATCH_END - 20);
 
+	/* Optimization register tricks: keep a base value in the
+	 * reserved P registers so we use the load/store with an
+	 * offset syntax.  R0 = [P5 + <constant>];
+	 *   P4 - system MMR base
+	 *   P5 - core MMR base
+	 */
+#ifdef CONFIG_HW_WATCHDOG
+	p4.l = 0;
+	p4.h = HI(SYSMMR_BASE);
+#endif
+	p5.l = 0;
+	p5.h = HI(COREMMR_BASE);
+
 #ifdef CONFIG_HW_WATCHDOG
 # ifndef CONFIG_HW_WATCHDOG_TIMEOUT_START
 #  define CONFIG_HW_WATCHDOG_TIMEOUT_START 5000
@@ -60,13 +73,11 @@ ENTRY(_start)
 	 * That should be long enough to bootstrap ourselves up and
 	 * then the common u-boot code can take over.
 	 */
-	P0.L = LO(WDOG_CNT);
-	P0.H = HI(WDOG_CNT);
-	R0.L = 0;
-	R0.H = HI(MSEC_TO_SCLK(CONFIG_HW_WATCHDOG_TIMEOUT_START));
-	[P0] = R0;
+	r0 = 0;
+	r0.h = HI(MSEC_TO_SCLK(CONFIG_HW_WATCHDOG_TIMEOUT_START));
+	[p4 + (WDOG_CNT - SYSMMR_BASE)] = r0;
 	/* fire up the watchdog - R0.L above needs to be 0x0000 */
-	W[P0 + (WDOG_CTL - WDOG_CNT)] = R0;
+	W[p4 + (WDOG_CTL - SYSMMR_BASE)] = r0;
 #endif
 
 	/* Turn on the serial for debugging the init process */
@@ -121,6 +132,18 @@ ENTRY(_start)
 	if cc jump .Lnorelocate;
 	r6 = 0 (x);
 
+	/* Turn off caches as they require CPLBs and a CPLB miss requires
+	 * a software exception handler to process it.  But we're about to
+	 * clobber any previous executing software (like U-Boot that just
+	 * launched a new U-Boot via 'go'), so any handler state will be
+	 * unreliable after the memcpy below.
+	 */
+	serial_early_puts("Kill Caches");
+	r0 = 0;
+	[p5 + (IMEM_CONTROL - COREMMR_BASE)] = r0;
+	[p5 + (DMEM_CONTROL - COREMMR_BASE)] = r0;
+	ssync;
+
 	/* In bypass mode, we don't have an LDR with an init block
 	 * so we need to explicitly call it ourselves.  This will
 	 * reprogram our clocks, memory, and setup our async banks.
@@ -204,17 +227,15 @@ ENTRY(_start)
 	serial_early_puts("Lower to 15");
 	r0 = r7;
 	r1 = r6;
-	p0.l = LO(EVT15);
-	p0.h = HI(EVT15);
 	p1.l = .Lenable_nested;
 	p1.h = .Lenable_nested;
-	[p0] = p1;
+	[p5 + (EVT15 - COREMMR_BASE)] = p1;
 	r7 = EVT_IVG15 (z);
 	sti r7;
 	raise 15;
-	p4.l = .LWAIT_HERE;
-	p4.h = .LWAIT_HERE;
-	reti = p4;
+	p3.l = .LWAIT_HERE;
+	p3.h = .LWAIT_HERE;
+	reti = p3;
 	rti;
 
 	/* Enable nested interrupts before continuing with cpu init */