X-Git-Url: https://git.gag.com/?p=fw%2Faltos;a=blobdiff_plain;f=src%2Fstm%2Fao_arch_funcs.h;h=d2cc11201a7e7c550d36985d9ef0df871d61764e;hp=87bbe73e3b926a6b964d181f6dc5a60926cb4305;hb=HEAD;hpb=688a9458bb03a81e71554c14295d1baacbbbd530

diff --git a/src/stm/ao_arch_funcs.h b/src/stm/ao_arch_funcs.h
index 87bbe73e..0f2f36a6 100644
--- a/src/stm/ao_arch_funcs.h
+++ b/src/stm/ao_arch_funcs.h
@@ -3,7 +3,8 @@
  *
  * This program is free software; you can redistribute it and/or modify
  * it under the terms of the GNU General Public License as published by
- * the Free Software Foundation; version 2 of the License.
+ * the Free Software Foundation; either version 2 of the License, or
+ * (at your option) any later version.
  *
  * This program is distributed in the hope that it will be useful, but
  * WITHOUT ANY WARRANTY; without even the implied warranty of
@@ -23,20 +24,30 @@
 
 /* PCLK is set to 16MHz (HCLK 32MHz, APB prescaler 2) */
 
-#define AO_SPI_SPEED_8MHz	STM_SPI_CR1_BR_PCLK_2	/* This doesn't appear to work */
-#define AO_SPI_SPEED_4MHz	STM_SPI_CR1_BR_PCLK_4
-#define AO_SPI_SPEED_2MHz	STM_SPI_CR1_BR_PCLK_8
-#define AO_SPI_SPEED_1MHz	STM_SPI_CR1_BR_PCLK_16
-#define AO_SPI_SPEED_500kHz	STM_SPI_CR1_BR_PCLK_32
-#define AO_SPI_SPEED_250kHz	STM_SPI_CR1_BR_PCLK_64
-#define AO_SPI_SPEED_125kHz	STM_SPI_CR1_BR_PCLK_128
-#define AO_SPI_SPEED_62500Hz	STM_SPI_CR1_BR_PCLK_256
+//#define AO_SPI_SPEED_8MHz	STM_SPI_CR1_BR_PCLK_2	/* too fast to use safely */
+#define _AO_SPI_SPEED_4MHz	STM_SPI_CR1_BR_PCLK_4
+#define _AO_SPI_SPEED_2MHz	STM_SPI_CR1_BR_PCLK_8
+#define _AO_SPI_SPEED_1MHz	STM_SPI_CR1_BR_PCLK_16
+#define _AO_SPI_SPEED_500kHz	STM_SPI_CR1_BR_PCLK_32
+#define _AO_SPI_SPEED_250kHz	STM_SPI_CR1_BR_PCLK_64
+#define _AO_SPI_SPEED_125kHz	STM_SPI_CR1_BR_PCLK_128
+#define _AO_SPI_SPEED_62500Hz	STM_SPI_CR1_BR_PCLK_256
 
-#define AO_SPI_SPEED_FAST	AO_SPI_SPEED_4MHz
-
-/* Companion bus wants something no faster than 200kHz */
+static inline uint32_t
+ao_spi_speed(int index, uint32_t hz)
+{
+	(void) index;
+	if (hz >= 4000000) return _AO_SPI_SPEED_4MHz;
+	if (hz >= 2000000) return _AO_SPI_SPEED_2MHz;
+	if (hz >= 1000000) return _AO_SPI_SPEED_1MHz;
+	if (hz >=  500000) return _AO_SPI_SPEED_500kHz;
+	if (hz >=  250000) return _AO_SPI_SPEED_250kHz;
+	if (hz >=  125000) return _AO_SPI_SPEED_125kHz;
+	return _AO_SPI_SPEED_62500Hz;
+}
 
-#define AO_SPI_SPEED_200kHz	AO_SPI_SPEED_125kHz
+#define AO_SPI_CPOL_BIT		4
+#define AO_SPI_CPHA_BIT		5
 
 #define AO_SPI_CONFIG_1		0x00
 #define AO_SPI_1_CONFIG_PA5_PA6_PA7	AO_SPI_CONFIG_1
@@ -63,6 +74,18 @@
 
 #define AO_SPI_INDEX(id)	((id) & AO_SPI_INDEX_MASK)
 #define AO_SPI_CONFIG(id)	((id) & AO_SPI_CONFIG_MASK)
+#define AO_SPI_PIN_CONFIG(id)	((id) & (AO_SPI_INDEX_MASK | AO_SPI_CONFIG_MASK))
+#define AO_SPI_CPOL(id)		((uint32_t) (((id) >> AO_SPI_CPOL_BIT) & 1))
+#define AO_SPI_CPHA(id)		((uint32_t) (((id) >> AO_SPI_CPHA_BIT) & 1))
+
+#define AO_SPI_MAKE_MODE(pol,pha)	(((pol) << AO_SPI_CPOL_BIT) | ((pha) << AO_SPI_CPHA_BIT))
+#define AO_SPI_MODE_0		AO_SPI_MAKE_MODE(0,0)
+#define AO_SPI_MODE_1		AO_SPI_MAKE_MODE(0,1)
+#define AO_SPI_MODE_2		AO_SPI_MAKE_MODE(1,0)
+#define AO_SPI_MODE_3		AO_SPI_MAKE_MODE(1,1)
+
+uint8_t
+ao_spi_try_get(uint8_t spi_index, uint32_t speed, uint8_t task_id);
 
 void
 ao_spi_get(uint8_t spi_index, uint32_t speed);
@@ -71,34 +94,100 @@ void
 ao_spi_put(uint8_t spi_index);
 
 void
-ao_spi_send(void *block, uint16_t len, uint8_t spi_index);
+ao_spi_put_pins(uint8_t spi_index);
+
+void
+ao_spi_send(const void *block, uint16_t len, uint8_t spi_index);
 
 void
 ao_spi_send_fixed(uint8_t value, uint16_t len, uint8_t spi_index);
 
 void
-ao_spi_recv(void *block, uint16_t len, uint8_t spi_index);
+ao_spi_send_sync(const void *block, uint16_t len, uint8_t spi_index);
+
+void
+ao_spi_start_bytes(uint8_t spi_index);
 
 void
-ao_spi_duplex(void *out, void *in, uint16_t len, uint8_t spi_index);
+ao_spi_stop_bytes(uint8_t spi_index);
 
-extern uint16_t	ao_spi_speed[STM_NUM_SPI];
+static inline void
+ao_spi_send_byte(uint8_t byte, uint8_t spi_index)
+{
+	struct stm_spi	*stm_spi;
+
+	switch (AO_SPI_INDEX(spi_index)) {
+	case 0:
+		stm_spi = &stm_spi1;
+		break;
+	case 1:
+		stm_spi = &stm_spi2;
+		break;
+	}
+
+	while (!(stm_spi->sr & (1 << STM_SPI_SR_TXE)))
+		;
+	stm_spi->dr = byte;
+	while (!(stm_spi->sr & (1 << STM_SPI_SR_RXNE)))
+		;
+	(void) stm_spi->dr;
+}
+
+static inline uint8_t
+ao_spi_recv_byte(uint8_t spi_index)
+{
+	struct stm_spi	*stm_spi;
+
+	switch (AO_SPI_INDEX(spi_index)) {
+	case 0:
+		stm_spi = &stm_spi1;
+		break;
+	case 1:
+		stm_spi = &stm_spi2;
+		break;
+	}
+
+	while (!(stm_spi->sr & (1 << STM_SPI_SR_TXE)))
+		;
+	stm_spi->dr = 0xff;
+	while (!(stm_spi->sr & (1 << STM_SPI_SR_RXNE)))
+		;
+	return (uint8_t) stm_spi->dr;
+}
+
+void
+ao_spi_recv(void *block, uint16_t len, uint8_t spi_index);
+
+void
+ao_spi_duplex(const void *out, void *in, uint16_t len, uint8_t spi_index);
 
 void
 ao_spi_init(void);
 
+#define ao_spi_set_cs(reg,mask) ((reg)->bsrr = ((uint32_t) (mask)) << 16)
+#define ao_spi_clr_cs(reg,mask) ((reg)->bsrr = (mask))
+
 #define ao_spi_get_mask(reg,mask,bus, speed) do {		\
 		ao_spi_get(bus, speed);				\
-		(reg)->bsrr = ((uint32_t) mask) << 16;	\
+		ao_spi_set_cs(reg,mask);			\
 	} while (0)
 
+static inline uint8_t
+ao_spi_try_get_mask(struct stm_gpio *reg, uint16_t mask, uint8_t bus, uint32_t speed, uint8_t task_id)
+{
+	if (!ao_spi_try_get(bus, speed, task_id))
+		return 0;
+	ao_spi_set_cs(reg, mask);
+	return 1;
+}
+
 #define ao_spi_put_mask(reg,mask,bus) do {	\
-		(reg)->bsrr = mask;		\
+		ao_spi_clr_cs(reg,mask);	\
 		ao_spi_put(bus);		\
 	} while (0)
 
-#define ao_spi_get_bit(reg,bit,pin,bus,speed) ao_spi_get_mask(reg,(1<<bit),bus,speed)
-#define ao_spi_put_bit(reg,bit,pin,bus) ao_spi_put_mask(reg,(1<<bit),bus)
+#define ao_spi_get_bit(reg,bit,bus,speed) ao_spi_get_mask(reg,1<<(bit),bus,speed)
+#define ao_spi_put_bit(reg,bit,bus) ao_spi_put_mask(reg,1<<(bit),bus)
 
 #define ao_enable_port(port) do {					\
 		if ((port) == &stm_gpioa)				\
@@ -113,20 +202,51 @@ ao_spi_init(void);
 			stm_rcc.ahbenr |= (1 << STM_RCC_AHBENR_GPIOEEN); \
 	} while (0)
 
+#define ao_disable_port(port) do {					\
+		if ((port) == &stm_gpioa)				\
+			stm_rcc.ahbenr &= ~(1UL << STM_RCC_AHBENR_GPIOAEN); \
+		else if ((port) == &stm_gpiob)				\
+			stm_rcc.ahbenr &= ~(1UL << STM_RCC_AHBENR_GPIOBEN); \
+		else if ((port) == &stm_gpioc)				\
+			stm_rcc.ahbenr &= ~(1UL << STM_RCC_AHBENR_GPIOCEN); \
+		else if ((port) == &stm_gpiod)				\
+			stm_rcc.ahbenr &= ~(1UL << STM_RCC_AHBENR_GPIODEN); \
+		else if ((port) == &stm_gpioe)				\
+			stm_rcc.ahbenr &= ~(1UL << STM_RCC_AHBENR_GPIOEEN); \
+	} while (0)
+
+
+#define ao_gpio_set(port, bit, v) stm_gpio_set(port, bit, v)
 
-#define ao_gpio_set(port, bit, pin, v) stm_gpio_set(port, bit, v)
+#define ao_gpio_get(port, bit) stm_gpio_get(port, bit)
 
-#define ao_gpio_get(port, bit, pin) stm_gpio_get(port, bit)
+#define ao_gpio_set_bits(port, bits) stm_gpio_set_bits(port, bits)
 
-#define ao_enable_output(port,bit,pin,v) do {			\
+#define ao_gpio_set_mask(port, bits, mask) stm_gpio_set_mask(port, bits, mask)
+
+#define ao_gpio_clr_bits(port, bits) stm_gpio_clr_bits(port, bits);
+
+#define ao_gpio_get_all(port) stm_gpio_get_all(port)
+
+#define ao_enable_output(port,bit,v) do {			\
 		ao_enable_port(port);				\
-		ao_gpio_set(port, bit, pin, v);			\
+		ao_gpio_set(port, bit, v);			\
 		stm_moder_set(port, bit, STM_MODER_OUTPUT);\
 	} while (0)
 
-#define ao_enable_input(port,bit,mode) do {				\
-		ao_enable_port(port);					\
-		stm_moder_set(port, bit, STM_MODER_INPUT);		\
+#define ao_enable_output_mask(port,bits,mask) do {		\
+		ao_enable_port(port);				\
+		ao_gpio_set_mask(port, bits, mask);		\
+		ao_set_output_mask(port, mask);			\
+	} while (0)
+
+#define AO_OUTPUT_PUSH_PULL	STM_OTYPER_PUSH_PULL
+#define AO_OUTPUT_OPEN_DRAIN	STM_OTYPER_OPEN_DRAIN
+
+#define ao_gpio_set_output_mode(port,pin,mode) \
+	stm_otyper_set(port, pin, mode)
+
+#define ao_gpio_set_mode(port,bit,mode) do {				\
 		if (mode == AO_EXTI_MODE_PULL_UP)			\
 			stm_pupdr_set(port, bit, STM_PUPDR_PULL_UP);	\
 		else if (mode == AO_EXTI_MODE_PULL_DOWN)		\
@@ -135,29 +255,72 @@ ao_spi_init(void);
 			stm_pupdr_set(port, bit, STM_PUPDR_NONE);	\
 	} while (0)
 
-#define ao_enable_cs(port,bit) do {				\
+#define ao_gpio_set_mode_mask(port,mask,mode) do {			\
+		if (mode == AO_EXTI_MODE_PULL_UP)			\
+			stm_pupdr_set_mask(port, mask, STM_PUPDR_PULL_UP); \
+		else if (mode == AO_EXTI_MODE_PULL_DOWN)		\
+			stm_pupdr_set_mask(port, mask, STM_PUPDR_PULL_DOWN); \
+		else							\
+			stm_pupdr_set_mask(port, mask, STM_PUPDR_NONE);	\
+	} while (0)
+
+#define ao_set_input(port, bit) do {				\
+		stm_moder_set(port, bit, STM_MODER_INPUT);	\
+	} while (0)
+
+#define ao_set_output(port, bit, v) do {			\
+		ao_gpio_set(port, bit, v);			\
+		stm_moder_set(port, bit, STM_MODER_OUTPUT);	\
+	} while (0)
+
+#define ao_set_output_mask(port, mask) do {			\
+		stm_moder_set_mask(port, mask, STM_MODER_OUTPUT);	\
+	} while (0)
+
+#define ao_set_input_mask(port, mask) do {				\
+		stm_moder_set_mask(port, mask, STM_MODER_INPUT);	\
+	} while (0)
+
+#define ao_enable_input(port,bit,mode) do {				\
+		ao_enable_port(port);					\
+		ao_set_input(port, bit);				\
+		ao_gpio_set_mode(port, bit, mode);			\
+	} while (0)
+
+#define ao_enable_input_mask(port,mask,mode) do {	\
+		ao_enable_port(port);			\
+		ao_gpio_set_mode_mask(port, mask, mode);	\
+		ao_set_input_mask(port, mask);		\
+	} while (0)
+
+#define _ao_enable_cs(port, bit) do {				\
 		stm_gpio_set((port), bit, 1);			\
 		stm_moder_set((port), bit, STM_MODER_OUTPUT);	\
 	} while (0)
 
+#define ao_enable_cs(port,bit) do {				\
+		ao_enable_port(port);				\
+		_ao_enable_cs(port, bit);			\
+	} while (0)
+
 #define ao_spi_init_cs(port, mask) do {				\
 		ao_enable_port(port);				\
-		if ((mask) & 0x0001) ao_enable_cs(port, 0);	\
-		if ((mask) & 0x0002) ao_enable_cs(port, 1);	\
-		if ((mask) & 0x0004) ao_enable_cs(port, 2);	\
-		if ((mask) & 0x0008) ao_enable_cs(port, 3);	\
-		if ((mask) & 0x0010) ao_enable_cs(port, 4);	\
-		if ((mask) & 0x0020) ao_enable_cs(port, 5);	\
-		if ((mask) & 0x0040) ao_enable_cs(port, 6);	\
-		if ((mask) & 0x0080) ao_enable_cs(port, 7);	\
-		if ((mask) & 0x0100) ao_enable_cs(port, 8);	\
-		if ((mask) & 0x0200) ao_enable_cs(port, 9);	\
-		if ((mask) & 0x0400) ao_enable_cs(port, 10);\
-		if ((mask) & 0x0800) ao_enable_cs(port, 11);\
-		if ((mask) & 0x1000) ao_enable_cs(port, 12);\
-		if ((mask) & 0x2000) ao_enable_cs(port, 13);\
-		if ((mask) & 0x4000) ao_enable_cs(port, 14);\
-		if ((mask) & 0x8000) ao_enable_cs(port, 15);\
+		if ((mask) & 0x0001) _ao_enable_cs(port, 0);	\
+		if ((mask) & 0x0002) _ao_enable_cs(port, 1);	\
+		if ((mask) & 0x0004) _ao_enable_cs(port, 2);	\
+		if ((mask) & 0x0008) _ao_enable_cs(port, 3);	\
+		if ((mask) & 0x0010) _ao_enable_cs(port, 4);	\
+		if ((mask) & 0x0020) _ao_enable_cs(port, 5);	\
+		if ((mask) & 0x0040) _ao_enable_cs(port, 6);	\
+		if ((mask) & 0x0080) _ao_enable_cs(port, 7);	\
+		if ((mask) & 0x0100) _ao_enable_cs(port, 8);	\
+		if ((mask) & 0x0200) _ao_enable_cs(port, 9);	\
+		if ((mask) & 0x0400) _ao_enable_cs(port, 10);\
+		if ((mask) & 0x0800) _ao_enable_cs(port, 11);\
+		if ((mask) & 0x1000) _ao_enable_cs(port, 12);\
+		if ((mask) & 0x2000) _ao_enable_cs(port, 13);\
+		if ((mask) & 0x4000) _ao_enable_cs(port, 14);\
+		if ((mask) & 0x8000) _ao_enable_cs(port, 15);\
 	} while (0)
 
 /* ao_dma_stm.c
@@ -181,9 +344,6 @@ ao_dma_start(uint8_t index);
 void
 ao_dma_done_transfer(uint8_t index);
 
-void
-ao_dma_abort(uint8_t index);
-
 void
 ao_dma_alloc(uint8_t index);
 
@@ -210,14 +370,40 @@ ao_i2c_recv(void *block, uint16_t len, uint8_t i2c_index, uint8_t stop);
 void
 ao_i2c_init(void);
 
+#if USE_SERIAL_1_SW_FLOW || USE_SERIAL_2_SW_FLOW || USE_SERIAL_3_SW_FLOW
+#define HAS_SERIAL_SW_FLOW 1
+#else
+#define HAS_SERIAL_SW_FLOW 0
+#endif
+
+#if USE_SERIAL_1_FLOW && !USE_SERIAL_1_SW_FLOW || USE_SERIAL_2_FLOW && !USE_SERIAL_2_SW_FLOW || USE_SERIAL_3_FLOW && !USE_SERIAL_3_SW_FLOW
+#define HAS_SERIAL_HW_FLOW 1
+#else
+#define HAS_SERIAL_HW_FLOW 0
+#endif
+
 /* ao_serial_stm.c */
 struct ao_stm_usart {
 	struct ao_fifo		rx_fifo;
 	struct ao_fifo		tx_fifo;
 	struct stm_usart	*reg;
-	uint8_t			tx_started;
+	uint8_t			tx_running;
+	uint8_t			draining;
+#if HAS_SERIAL_SW_FLOW
+	/* RTS - 0 if we have FIFO space, 1 if not
+	 * CTS - 0 if we can send, 0 if not
+	 */
+	struct stm_gpio		*gpio_rts;
+	struct stm_gpio		*gpio_cts;
+	uint8_t			pin_rts;
+	uint8_t			pin_cts;
+	uint8_t			rts;
+#endif
 };
 
+void
+ao_debug_out(char c);
+
 #if HAS_SERIAL_1
 extern struct ao_stm_usart	ao_stm_usart1;
 #endif
@@ -232,28 +418,71 @@ extern struct ao_stm_usart	ao_stm_usart3;
 
 #define ARM_PUSH32(stack, val)	(*(--(stack)) = (val))
 
+typedef uint32_t	ao_arch_irq_t;
+
+static inline void
+ao_arch_block_interrupts(void) {
+#ifdef AO_NONMASK_INTERRUPTS
+	asm("msr basepri,%0" : : "r" (AO_STM_NVIC_BASEPRI_MASK));
+#else
+	asm("cpsid i");
+#endif
+}
+
+static inline void
+ao_arch_release_interrupts(void) {
+#ifdef AO_NONMASK_INTERRUPTS
+	asm("msr basepri,%0" : : "r" (0x0));
+#else
+	asm("cpsie i");
+#endif
+}
+
 static inline uint32_t
 ao_arch_irqsave(void) {
-	uint32_t	primask;
-	asm("mrs %0,primask" : "=&r" (primask));
+	uint32_t	val;
+#ifdef AO_NONMASK_INTERRUPTS
+	asm("mrs %0,basepri" : "=r" (val));
+#else
+	asm("mrs %0,primask" : "=r" (val));
+#endif
 	ao_arch_block_interrupts();
-	return primask;
+	return val;
 }
 
 static inline void
-ao_arch_irqrestore(uint32_t primask) {
-	asm("msr primask,%0" : : "r" (primask));
+ao_arch_irqrestore(uint32_t basepri) {
+#ifdef AO_NONMASK_INTERRUPTS
+	asm("msr basepri,%0" : : "r" (basepri));
+#else
+	asm("msr primask,%0" : : "r" (basepri));
+#endif
 }
 
 static inline void
-ao_arch_memory_barrier() {
+ao_arch_memory_barrier(void) {
 	asm volatile("" ::: "memory");
 }
 
 static inline void
-ao_arch_init_stack(struct ao_task *task, void *start)
+ao_arch_irq_check(void) {
+#ifdef AO_NONMASK_INTERRUPTS
+	uint32_t	basepri;
+	asm("mrs %0,basepri" : "=r" (basepri));
+	if (basepri == 0)
+		ao_panic(AO_PANIC_IRQ);
+#else
+	uint32_t	primask;
+	asm("mrs %0,primask" : "=r" (primask));
+	if ((primask & 1) == 0)
+		ao_panic(AO_PANIC_IRQ);
+#endif
+}
+
+#if HAS_TASK
+static inline void
+ao_arch_init_stack(struct ao_task *task, uint32_t *sp, void *start)
 {
-	uint32_t	*sp = (uint32_t *) (task->stack + AO_STACK_SIZE);
 	uint32_t	a = (uint32_t) start;
 	int		i;
 
@@ -268,10 +497,10 @@ ao_arch_init_stack(struct ao_task *task, void *start)
 	/* APSR */
 	ARM_PUSH32(sp, 0);
 
-	/* PRIMASK with interrupts enabled */
+	/* BASEPRI with interrupts enabled */
 	ARM_PUSH32(sp, 0);
 
-	task->sp = sp;
+	task->sp32 = sp;
 }
 
 static inline void ao_arch_save_regs(void) {
@@ -282,33 +511,39 @@ static inline void ao_arch_save_regs(void) {
 	asm("mrs r0,apsr");
 	asm("push {r0}");
 
+#ifdef AO_NONMASK_INTERRUPTS
+	/* Save BASEPRI */
+	asm("mrs r0,basepri");
+#else
 	/* Save PRIMASK */
 	asm("mrs r0,primask");
+#endif
 	asm("push {r0}");
 }
 
 static inline void ao_arch_save_stack(void) {
 	uint32_t	*sp;
 	asm("mov %0,sp" : "=&r" (sp) );
-	ao_cur_task->sp = (sp);
-	if ((uint8_t *) sp < &ao_cur_task->stack[0])
-		ao_panic (AO_PANIC_STACK);
+	ao_cur_task->sp32 = (sp);
 }
 
 static inline void ao_arch_restore_stack(void) {
-	uint32_t	sp;
-	sp = (uint32_t) ao_cur_task->sp;
-
 	/* Switch stacks */
-	asm("mov sp, %0" : : "r" (sp) );
+	asm("mov sp, %0" : : "r" (ao_cur_task->sp32) );
 
+#ifdef AO_NONMASK_INTERRUPTS
+	/* Restore BASEPRI */
+	asm("pop {r0}");
+	asm("msr basepri,r0");
+#else
 	/* Restore PRIMASK */
 	asm("pop {r0}");
 	asm("msr primask,r0");
+#endif
 
 	/* Restore APSR */
 	asm("pop {r0}");
-	asm("msr apsr,r0");
+	asm("msr apsr_nczvq,r0");
 
 	/* Restore general registers */
 	asm("pop {r0-r12,lr}\n");
@@ -317,18 +552,71 @@ static inline void ao_arch_restore_stack(void) {
 	asm("bx lr");
 }
 
+#ifndef HAS_SAMPLE_PROFILE
+#define HAS_SAMPLE_PROFILE 0
+#endif
+
+#if DEBUG
+#define HAS_ARCH_VALIDATE_CUR_STACK	1
+
+static inline void
+ao_validate_cur_stack(void)
+{
+	uint8_t		*psp;
+
+	asm("mrs %0,psp" : "=&r" (psp));
+	if (ao_cur_task &&
+	    (psp <= ao_cur_task->stack8 ||
+	     psp >= ao_cur_task->stack8 + AO_STACK_SIZE))
+		ao_panic(AO_PANIC_STACK);
+}
+#endif
+
+#if !HAS_SAMPLE_PROFILE
+#define HAS_ARCH_START_SCHEDULER	1
+
+static inline void ao_arch_start_scheduler(void) {
+	uint32_t	sp;
+	uint32_t	control;
+
+	asm("mrs %0,msp" : "=&r" (sp));
+	asm("msr psp,%0" : : "r" (sp));
+	asm("mrs %0,control" : "=r" (control));
+	control |= (1 << 1);
+	asm("msr control,%0" : : "r" (control));
+	asm("isb");
+}
+#endif
+
 #define ao_arch_isr_stack()
 
-#define ao_arch_wait_interrupt() do {			\
-		asm(".global ao_idle_loc\n\twfi\nao_idle_loc:");	\
-		ao_arch_release_interrupts();				\
-		ao_arch_block_interrupts();				\
-	} while (0)
+#endif
 
-#define ao_arch_critical(b) do {				\
-		ao_arch_block_interrupts();			\
-		do { b } while (0);				\
-		ao_arch_release_interrupts();			\
+static inline void
+ao_arch_wait_interrupt(void) {
+#ifdef AO_NONMASK_INTERRUPTS
+	asm(
+	    "dsb\n"			/* Serialize data */
+	    "isb\n"			/* Serialize instructions */
+	    "cpsid i\n"			/* Block all interrupts */
+	    "msr basepri,%0\n"		/* Allow all interrupts through basepri */
+	    "wfi\n"			/* Wait for an interrupt */
+	    "cpsie i\n"			/* Allow all interrupts */
+	    "msr basepri,%1\n"		/* Block interrupts through basepri */
+	    : : "r" (0), "r" (AO_STM_NVIC_BASEPRI_MASK));
+#else
+	asm("\twfi\n");
+	ao_arch_release_interrupts();
+	ao_arch_block_interrupts();
+#endif
+}
+
+#define ao_arch_critical(b) do {			\
+		uint32_t __mask = ao_arch_irqsave();	\
+		do { b } while (0);			\
+		ao_arch_irqrestore(__mask);		\
 	} while (0)
 
+void start(void);
+
 #endif /* _AO_ARCH_FUNCS_H_ */