diff --git a/arch/arm/mach-ns9xxx/board-a9m9750dev.c b/arch/arm/mach-ns9xxx/board-a9m9750dev.c
index 7fa3fb1..0f65177 100644
--- a/arch/arm/mach-ns9xxx/board-a9m9750dev.c
+++ b/arch/arm/mach-ns9xxx/board-a9m9750dev.c
@@ -45,7 +45,13 @@
 
 static void a9m9750dev_fpga_mask_irq(unsigned int irq)
 {
-	FPGA_IER &= ~(1 << (irq - FPGA_IRQ(0)));
+	u8 ier;
+
+	ier = __raw_readb(FPGA_IER);
+
+	ier &= ~(1 << (irq - FPGA_IRQ(0)));
+
+	__raw_writeb(ier, FPGA_IER);
 }
 
 static void a9m9750dev_fpga_maskack_irq(unsigned int irq)
@@ -56,7 +62,13 @@
 
 static void a9m9750dev_fpga_unmask_irq(unsigned int irq)
 {
-	FPGA_IER |= 1 << (irq - FPGA_IRQ(0));
+	u8 ier;
+
+	ier = __raw_readb(FPGA_IER);
+
+	ier |= 1 << (irq - FPGA_IRQ(0));
+
+	__raw_writeb(ier, FPGA_IER);
 }
 
 static struct irq_chip a9m9750dev_fpga_chip = {
@@ -69,7 +81,7 @@
 static void a9m9750dev_fpga_demux_handler(unsigned int irq,
 		struct irq_desc *desc)
 {
-	int stat = FPGA_ISR;
+	u8 stat = __raw_readb(FPGA_ISR);
 
 	desc->chip->mask_ack(irq);
 
@@ -89,7 +101,7 @@
 
 void __init board_a9m9750dev_init_irq(void)
 {
-	u32 reg;
+	u32 eic;
 	int i;
 
 	if (gpio_request(11, "board a9m9750dev extirq2") == 0)
@@ -105,10 +117,10 @@
 	}
 
 	/* IRQ_EXT2: level sensitive + active low */
-	reg = SYS_EIC(2);
-	REGSET(reg, SYS_EIC, PLTY, AL);
-	REGSET(reg, SYS_EIC, LVEDG, LEVEL);
-	SYS_EIC(2) = reg;
+	eic = __raw_readl(SYS_EIC(2));
+	REGSET(eic, SYS_EIC, PLTY, AL);
+	REGSET(eic, SYS_EIC, LVEDG, LEVEL);
+	__raw_writel(eic, SYS_EIC(2));
 
 	set_irq_chained_handler(IRQ_EXT2,
 			a9m9750dev_fpga_demux_handler);
@@ -172,17 +184,18 @@
 	u32 reg;
 
 	/* setup static CS0: memory base ... */
-	REGSETIM(SYS_SMCSSMB(0), SYS_SMCSSMB, CSxB,
-			NS9XXX_CSxSTAT_PHYS(0) >> 12);
+	reg = __raw_readl(SYS_SMCSSMB(0));
+	REGSETIM(reg, SYS_SMCSSMB, CSxB, NS9XXX_CSxSTAT_PHYS(0) >> 12);
+	__raw_writel(reg, SYS_SMCSSMB(0));
 
 	/* ... and mask */
-	reg = SYS_SMCSSMM(0);
+	reg = __raw_readl(SYS_SMCSSMM(0));
 	REGSETIM(reg, SYS_SMCSSMM, CSxM, 0xfffff);
 	REGSET(reg, SYS_SMCSSMM, CSEx, EN);
-	SYS_SMCSSMM(0) = reg;
+	__raw_writel(reg, SYS_SMCSSMM(0));
 
 	/* setup static CS0: memory configuration */
-	reg = MEM_SMC(0);
+	reg = __raw_readl(MEM_SMC(0));
 	REGSET(reg, MEM_SMC, PSMC, OFF);
 	REGSET(reg, MEM_SMC, BSMC, OFF);
 	REGSET(reg, MEM_SMC, EW, OFF);
@@ -190,13 +203,13 @@
 	REGSET(reg, MEM_SMC, PC, AL);
 	REGSET(reg, MEM_SMC, PM, DIS);
 	REGSET(reg, MEM_SMC, MW, 8);
-	MEM_SMC(0) = reg;
+	__raw_writel(reg, MEM_SMC(0));
 
 	/* setup static CS0: timing */
-	MEM_SMWED(0) = 0x2;
-	MEM_SMOED(0) = 0x2;
-	MEM_SMRD(0) = 0x6;
-	MEM_SMWD(0) = 0x6;
+	__raw_writel(0x2, MEM_SMWED(0));
+	__raw_writel(0x2, MEM_SMOED(0));
+	__raw_writel(0x6, MEM_SMRD(0));
+	__raw_writel(0x6, MEM_SMWD(0));
 
 	platform_add_devices(board_a9m9750dev_devices,
 			ARRAY_SIZE(board_a9m9750dev_devices));
diff --git a/arch/arm/mach-ns9xxx/gpio.c b/arch/arm/mach-ns9xxx/gpio.c
index 87b872f..b223021 100644
--- a/arch/arm/mach-ns9xxx/gpio.c
+++ b/arch/arm/mach-ns9xxx/gpio.c
@@ -16,6 +16,7 @@
 #include <asm/arch-ns9xxx/gpio.h>
 #include <asm/arch-ns9xxx/processor.h>
 #include <asm/arch-ns9xxx/regs-bbu.h>
+#include <asm/io.h>
 #include <asm/bug.h>
 #include <asm/types.h>
 #include <asm/bitops.h>
@@ -47,38 +48,38 @@
 		BUG();
 }
 
-static inline volatile u32 *ns9xxx_gpio_get_gconfaddr(unsigned gpio)
+static inline void __iomem *ns9xxx_gpio_get_gconfaddr(unsigned gpio)
 {
 	if (gpio < 56)
-		return &BBU_GCONFb1(gpio / 8);
+		return BBU_GCONFb1(gpio / 8);
 	else
 		/*
 		 * this could be optimised away on
 		 * ns9750 only builds, but it isn't ...
 		 */
-		return &BBU_GCONFb2((gpio - 56) / 8);
+		return BBU_GCONFb2((gpio - 56) / 8);
 }
 
-static inline volatile u32 *ns9xxx_gpio_get_gctrladdr(unsigned gpio)
+static inline void __iomem *ns9xxx_gpio_get_gctrladdr(unsigned gpio)
 {
 	if (gpio < 32)
-		return &BBU_GCTRL1;
+		return BBU_GCTRL1;
 	else if (gpio < 64)
-		return &BBU_GCTRL2;
+		return BBU_GCTRL2;
 	else
 		/* this could be optimised away on ns9750 only builds */
-		return &BBU_GCTRL3;
+		return BBU_GCTRL3;
 }
 
-static inline volatile u32 *ns9xxx_gpio_get_gstataddr(unsigned gpio)
+static inline void __iomem *ns9xxx_gpio_get_gstataddr(unsigned gpio)
 {
 	if (gpio < 32)
-		return &BBU_GSTAT1;
+		return BBU_GSTAT1;
 	else if (gpio < 64)
-		return &BBU_GSTAT2;
+		return BBU_GSTAT2;
 	else
 		/* this could be optimised away on ns9750 only builds */
-		return &BBU_GSTAT3;
+		return BBU_GSTAT3;
 }
 
 int gpio_request(unsigned gpio, const char *label)
@@ -105,17 +106,17 @@
  */
 static int __ns9xxx_gpio_configure(unsigned gpio, int dir, int inv, int func)
 {
-	volatile u32 *conf = ns9xxx_gpio_get_gconfaddr(gpio);
+	void __iomem *conf = ns9xxx_gpio_get_gconfaddr(gpio);
 	u32 confval;
 	unsigned long flags;
 
 	spin_lock_irqsave(&gpio_lock, flags);
 
-	confval = *conf;
+	confval = __raw_readl(conf);
 	REGSETIM_IDX(confval, BBU_GCONFx, DIR, gpio & 7, dir);
 	REGSETIM_IDX(confval, BBU_GCONFx, INV, gpio & 7, inv);
 	REGSETIM_IDX(confval, BBU_GCONFx, FUNC, gpio & 7, func);
-	*conf = confval;
+	__raw_writel(confval, conf);
 
 	spin_unlock_irqrestore(&gpio_lock, flags);
 
@@ -158,10 +159,10 @@
 
 int gpio_get_value(unsigned gpio)
 {
-	volatile u32 *stat = ns9xxx_gpio_get_gstataddr(gpio);
+	void __iomem *stat = ns9xxx_gpio_get_gstataddr(gpio);
 	int ret;
 
-	ret = 1 & (*stat >> (gpio & 31));
+	ret = 1 & (__raw_readl(stat) >> (gpio & 31));
 
 	return ret;
 }
@@ -169,15 +170,20 @@
 
 void gpio_set_value(unsigned gpio, int value)
 {
-	volatile u32 *ctrl = ns9xxx_gpio_get_gctrladdr(gpio);
+	void __iomem *ctrl = ns9xxx_gpio_get_gctrladdr(gpio);
+	u32 ctrlval;
 	unsigned long flags;
 
 	spin_lock_irqsave(&gpio_lock, flags);
 
+	ctrlval = __raw_readl(ctrl);
+
 	if (value)
-		*ctrl |= 1 << (gpio & 31);
+		ctrlval |= 1 << (gpio & 31);
 	else
-		*ctrl &= ~(1 << (gpio & 31));
+		ctrlval &= ~(1 << (gpio & 31));
+
+	__raw_writel(ctrlval, ctrl);
 
 	spin_unlock_irqrestore(&gpio_lock, flags);
 }
diff --git a/arch/arm/mach-ns9xxx/irq.c b/arch/arm/mach-ns9xxx/irq.c
index 24d424e..00001b8 100644
--- a/arch/arm/mach-ns9xxx/irq.c
+++ b/arch/arm/mach-ns9xxx/irq.c
@@ -9,6 +9,7 @@
  * the Free Software Foundation.
  */
 #include <linux/interrupt.h>
+#include <asm/io.h>
 #include <asm/mach/irq.h>
 #include <asm/mach-types.h>
 #include <asm/arch-ns9xxx/regs-sys.h>
@@ -20,12 +21,14 @@
 static void ns9xxx_mask_irq(unsigned int irq)
 {
 	/* XXX: better use cpp symbols */
-	SYS_IC(irq / 4) &= ~(1 << (7 + 8 * (3 - (irq & 3))));
+	u32 ic = __raw_readl(SYS_IC(irq / 4));
+	ic &= ~(1 << (7 + 8 * (3 - (irq & 3))));
+	__raw_writel(ic, SYS_IC(irq / 4));
 }
 
 static void ns9xxx_ack_irq(unsigned int irq)
 {
-	SYS_ISRADDR = 0;
+	__raw_writel(0, SYS_ISRADDR);
 }
 
 static void ns9xxx_maskack_irq(unsigned int irq)
@@ -37,7 +40,9 @@
 static void ns9xxx_unmask_irq(unsigned int irq)
 {
 	/* XXX: better use cpp symbols */
-	SYS_IC(irq / 4) |= 1 << (7 + 8 * (3 - (irq & 3)));
+	u32 ic = __raw_readl(SYS_IC(irq / 4));
+	ic |= 1 << (7 + 8 * (3 - (irq & 3)));
+	__raw_writel(ic, SYS_IC(irq / 4));
 }
 
 static struct irq_chip ns9xxx_chip = {
@@ -53,14 +58,14 @@
 
 	/* disable all IRQs */
 	for (i = 0; i < 8; ++i)
-		SYS_IC(i) = (4 * i) << 24 | (4 * i + 1) << 16 |
-			(4 * i + 2) << 8 | (4 * i + 3);
+		__raw_writel((4 * i) << 24 | (4 * i + 1) << 16 |
+				(4 * i + 2) << 8 | (4 * i + 3), SYS_IC(i));
 
 	/* simple interrupt prio table:
 	 * prio(x) < prio(y) <=> x < y
 	 */
 	for (i = 0; i < 32; ++i)
-		SYS_IVA(i) = i;
+		__raw_writel(i, SYS_IVA(i));
 
 	for (i = IRQ_WATCHDOG; i <= IRQ_EXT3; ++i) {
 		set_irq_chip(i, &ns9xxx_chip);
diff --git a/arch/arm/mach-ns9xxx/time.c b/arch/arm/mach-ns9xxx/time.c
index 6f29515..c3dd1f4 100644
--- a/arch/arm/mach-ns9xxx/time.c
+++ b/arch/arm/mach-ns9xxx/time.c
@@ -27,7 +27,7 @@
 
 static cycle_t ns9xxx_clocksource_read(void)
 {
-	return SYS_TR(TIMER_CLOCKSOURCE);
+	return __raw_readl(SYS_TR(TIMER_CLOCKSOURCE));
 }
 
 static struct clocksource ns9xxx_clocksource = {
@@ -42,11 +42,11 @@
 static void ns9xxx_clockevent_setmode(enum clock_event_mode mode,
 		struct clock_event_device *clk)
 {
-	u32 tc = SYS_TC(TIMER_CLOCKEVENT);
+	u32 tc = __raw_readl(SYS_TC(TIMER_CLOCKEVENT));
 
 	switch(mode) {
 	case CLOCK_EVT_MODE_PERIODIC:
-		SYS_TRC(TIMER_CLOCKEVENT) = latch;
+		__raw_writel(latch, SYS_TRC(TIMER_CLOCKEVENT));
 		REGSET(tc, SYS_TCx, REN, EN);
 		REGSET(tc, SYS_TCx, INTS, EN);
 		REGSET(tc, SYS_TCx, TEN, EN);
@@ -66,24 +66,24 @@
 		break;
 	}
 
-	SYS_TC(TIMER_CLOCKEVENT) = tc;
+	__raw_writel(tc, SYS_TC(TIMER_CLOCKEVENT));
 }
 
 static int ns9xxx_clockevent_setnextevent(unsigned long evt,
 		struct clock_event_device *clk)
 {
-	u32 tc = SYS_TC(TIMER_CLOCKEVENT);
+	u32 tc = __raw_readl(SYS_TC(TIMER_CLOCKEVENT));
 
 	if (REGGET(tc, SYS_TCx, TEN)) {
 		REGSET(tc, SYS_TCx, TEN, DIS);
-		SYS_TC(TIMER_CLOCKEVENT) = tc;
+		__raw_writel(tc, SYS_TC(TIMER_CLOCKEVENT));
 	}
 
 	REGSET(tc, SYS_TCx, TEN, EN);
 
-	SYS_TRC(TIMER_CLOCKEVENT) = evt;
+	__raw_writel(evt, SYS_TRC(TIMER_CLOCKEVENT));
 
-	SYS_TC(TIMER_CLOCKEVENT) = tc;
+	__raw_writel(tc, SYS_TC(TIMER_CLOCKEVENT));
 
 	return 0;
 }
@@ -104,15 +104,15 @@
 	struct clock_event_device *evt = &ns9xxx_clockevent_device;
 
 	/* clear irq */
-	tc = SYS_TC(timerno);
+	tc = __raw_readl(SYS_TC(timerno));
 	if (REGGET(tc, SYS_TCx, REN) == SYS_TCx_REN_DIS) {
 		REGSET(tc, SYS_TCx, TEN, DIS);
-		SYS_TC(timerno) = tc;
+		__raw_writel(tc, SYS_TC(timerno));
 	}
 	REGSET(tc, SYS_TCx, INTC, SET);
-	SYS_TC(timerno) = tc;
+	__raw_writel(tc, SYS_TC(timerno));
 	REGSET(tc, SYS_TCx, INTC, UNSET);
-	SYS_TC(timerno) = tc;
+	__raw_writel(tc, SYS_TC(timerno));
 
 	evt->event_handler(evt);
 
@@ -129,13 +129,13 @@
 {
 	int tc;
 
-	tc = SYS_TC(TIMER_CLOCKSOURCE);
+	tc = __raw_readl(SYS_TC(TIMER_CLOCKSOURCE));
 	if (REGGET(tc, SYS_TCx, TEN)) {
 		REGSET(tc, SYS_TCx, TEN, DIS);
-		SYS_TC(TIMER_CLOCKSOURCE) = tc;
+		__raw_writel(tc, SYS_TC(TIMER_CLOCKSOURCE));
 	}
 
-	SYS_TRC(TIMER_CLOCKSOURCE) = 0;
+	__raw_writel(0, SYS_TRC(TIMER_CLOCKSOURCE));
 
 	REGSET(tc, SYS_TCx, TEN, EN);
 	REGSET(tc, SYS_TCx, TDBG, STOP);
@@ -146,7 +146,7 @@
 	REGSET(tc, SYS_TCx, TSZ, 32);
 	REGSET(tc, SYS_TCx, REN, EN);
 
-	SYS_TC(TIMER_CLOCKSOURCE) = tc;
+	__raw_writel(tc, SYS_TC(TIMER_CLOCKSOURCE));
 
 	ns9xxx_clocksource.mult = clocksource_hz2mult(ns9xxx_cpuclock(),
 			ns9xxx_clocksource.shift);
@@ -155,7 +155,7 @@
 
 	latch = SH_DIV(ns9xxx_cpuclock(), HZ, 0);
 
-	tc = SYS_TC(TIMER_CLOCKEVENT);
+	tc = __raw_readl(SYS_TC(TIMER_CLOCKEVENT));
 	REGSET(tc, SYS_TCx, TEN, DIS);
 	REGSET(tc, SYS_TCx, TDBG, STOP);
 	REGSET(tc, SYS_TCx, TLCS, CPU);
@@ -164,7 +164,7 @@
 	REGSET(tc, SYS_TCx, UDS, DOWN);
 	REGSET(tc, SYS_TCx, TSZ, 32);
 	REGSET(tc, SYS_TCx, REN, EN);
-	SYS_TC(TIMER_CLOCKEVENT) = tc;
+	__raw_writel(tc, SYS_TC(TIMER_CLOCKEVENT));
 
 	ns9xxx_clockevent_device.mult = div_sc(ns9xxx_cpuclock(),
 			NSEC_PER_SEC, ns9xxx_clockevent_device.shift);
diff --git a/include/asm-arm/arch-ns9xxx/clock.h b/include/asm-arm/arch-ns9xxx/clock.h
index 90604d4..b943d3a 100644
--- a/include/asm-arm/arch-ns9xxx/clock.h
+++ b/include/asm-arm/arch-ns9xxx/clock.h
@@ -19,7 +19,7 @@
 static inline u32 ns9xxx_systemclock(void) __attribute__((const));
 static inline u32 ns9xxx_systemclock(void)
 {
-	u32 pll = SYS_PLL;
+	u32 pll = __raw_readl(SYS_PLL);
 
 	/*
 	 * The system clock should be a multiple of HZ * TIMERCLOCKSELECT (in
diff --git a/include/asm-arm/arch-ns9xxx/hardware.h b/include/asm-arm/arch-ns9xxx/hardware.h
index b5d66f9..0b7b346 100644
--- a/include/asm-arm/arch-ns9xxx/hardware.h
+++ b/include/asm-arm/arch-ns9xxx/hardware.h
@@ -35,11 +35,8 @@
 
 #ifndef __ASSEMBLY__
 
-#  define __REG(x)	(*((volatile u32 *)io_p2v((x))))
-#  define __REG2(x, y)	(*((volatile u32 *)io_p2v((x)) + (y)))
-
-#  define __REGB(x)	(*((volatile u8 *)io_p2v((x))))
-#  define __REGB2(x)	(*((volatile u8 *)io_p2v((x)) + (y)))
+#  define __REG(x)	((void __iomem __force *)io_p2v((x)))
+#  define __REG2(x, y)	((void __iomem __force *)(io_p2v((x)) + 4 * (y)))
 
 #  define __REGSET(var, field, value)					\
 	((var) = (((var) & ~((field) & ~(value))) | (value)))
@@ -77,9 +74,6 @@
 #  define __REG(x)	io_p2v(x)
 #  define __REG2(x, y)	io_p2v((x) + (y))
 
-#  define __REGB(x)	__REG((x))
-#  define __REGB2(x, y)	__REG2((x), (y))
-
 #endif
 
 #endif /* ifndef __ASM_ARCH_HARDWARE_H */
diff --git a/include/asm-arm/arch-ns9xxx/regs-board-a9m9750dev.h b/include/asm-arm/arch-ns9xxx/regs-board-a9m9750dev.h
index c3dc532..afa3a9d 100644
--- a/include/asm-arm/arch-ns9xxx/regs-board-a9m9750dev.h
+++ b/include/asm-arm/arch-ns9xxx/regs-board-a9m9750dev.h
@@ -18,7 +18,7 @@
 #define FPGA_UARTC_BASE	io_p2v(NS9XXX_CSxSTAT_PHYS(0) + 0x10)
 #define FPGA_UARTD_BASE	io_p2v(NS9XXX_CSxSTAT_PHYS(0) + 0x18)
 
-#define FPGA_IER	__REGB(NS9XXX_CSxSTAT_PHYS(0) + 0x50)
-#define FPGA_ISR	__REGB(NS9XXX_CSxSTAT_PHYS(0) + 0x60)
+#define FPGA_IER	__REG(NS9XXX_CSxSTAT_PHYS(0) + 0x50)
+#define FPGA_ISR	__REG(NS9XXX_CSxSTAT_PHYS(0) + 0x60)
 
 #endif /* ifndef __ASM_ARCH_REGSBOARDA9M9750_H */
diff --git a/include/asm-arm/arch-ns9xxx/system.h b/include/asm-arm/arch-ns9xxx/system.h
index e3cd4d3..c1082bd 100644
--- a/include/asm-arm/arch-ns9xxx/system.h
+++ b/include/asm-arm/arch-ns9xxx/system.h
@@ -24,9 +24,9 @@
 {
 	u32 reg;
 
-	reg = SYS_PLL >> 16;
+	reg = __raw_readl(SYS_PLL) >> 16;
 	REGSET(reg, SYS_PLL, SWC, YES);
-	SYS_PLL = reg;
+	__raw_writel(reg, SYS_PLL);
 
 	BUG();
 }