diff --git a/src/host/hardware_divider/include/hardware/divider.h b/src/host/hardware_divider/include/hardware/divider.h
index c1266885..0a6eccc0 100644
--- a/src/host/hardware_divider/include/hardware/divider.h
+++ b/src/host/hardware_divider/include/hardware/divider.h
@@ -7,7 +7,11 @@
 #ifndef _HARDWARE_DIVIDER_H
 #define _HARDWARE_DIVIDER_H
 
-#include "pico/types.h"
+#include "pico.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
 
 typedef uint64_t divmod_result_t;
 
@@ -15,108 +19,313 @@ static inline int __sign_of(int32_t v) {
     return v > 0 ? 1 : (v < 0 ? -1 : 0);
 }
 
-// divides unsigned values a by b... (a/b) returned in low 32 bits, (a%b) in high 32 bits... results undefined for b==0
-static inline uint64_t hw_divider_divmod_u32(uint32_t a, uint32_t b) {
+/*! \brief Do an unsigned HW divide and wait for result
+ *  \ingroup hardware_divider
+ *
+ * Divide \p a by \p b, wait for calculation to complete, return result as a pair of 32-bit quotient/remainder values.
+ *
+ * \param a The dividend
+ * \param b The divisor
+ * \return Results of divide as a pair of 32-bit quotient/remainder values.
+ */
+static inline divmod_result_t hw_divider_divmod_u32(uint32_t a, uint32_t b)  {
     if (!b) return (((uint64_t)a)<<32u) | (uint32_t)(-1); // todo check this
     return (((uint64_t)(a%b))<<32u) | (a/b);
 }
 
-// divides signed values a by b... (a/b) returned in low 32 bits, (a%b) in high 32 bits... results undefined for b==0
-static inline uint64_t hw_divider_divmod_s32(int32_t a, int32_t b) {
+/*! \brief Do a signed HW divide and wait for result
+ *  \ingroup hardware_divider
+ *
+ * Divide \p a by \p b, wait for calculation to complete, return result as a pair of 32-bit quotient/remainder values.
+ *
+ * \param a The dividend
+ * \param b The divisor
+ * \return Results of divide as a pair of 32-bit quotient/remainder values.
+ */
+static inline divmod_result_t hw_divider_divmod_s32(int32_t a, int32_t b) {
     if (!b) return (((uint64_t)a)<<32u) | (uint32_t)(-__sign_of(a));
     return (((uint64_t)(a%b))<<32u) | (uint32_t)(a/b);
 }
 
 extern __thread divmod_result_t hw_divider_result_threadlocal;
 
+/*! \brief Start a signed asynchronous divide
+ *  \ingroup hardware_divider
+ *
+ * Start a divide of the specified signed parameters. You should wait for 8 cycles (__div_pause()) or wait for the ready bit to be set
+ * (hw_divider_wait_ready()) prior to reading the results.
+ *
+ * \param a The dividend
+ * \param b The divisor
+ */
 static inline void hw_divider_divmod_s32_start(int32_t a, int32_t b) {
     hw_divider_result_threadlocal = hw_divider_divmod_s32(a, b);
 }
 
+/*! \brief Start an unsigned asynchronous divide
+ *  \ingroup hardware_divider
+ *
+ * Start a divide of the specified unsigned parameters. You should wait for 8 cycles (__div_pause()) or wait for the ready bit to be set
+ * (hw_divider_wait_ready()) prior to reading the results.
+ *
+ * \param a The dividend
+ * \param b The divisor
+ */
 static inline void hw_divider_divmod_u32_start(uint32_t a, uint32_t b) {
     hw_divider_result_threadlocal = hw_divider_divmod_u32(a, b);
 }
 
-static inline divmod_result_t hw_divider_result_wait() {
+/*! \brief Return result of last asynchronous HW divide
+ *  \ingroup hardware_divider
+ *
+ * This function waits for the result to be ready by calling hw_divider_wait_ready().
+ *
+ * \return Current result. Most significant 32 bits are the remainder, lower 32 bits are the quotient.
+ */
+static inline divmod_result_t hw_divider_result_wait(void) {
     return hw_divider_result_threadlocal;
 }
 
-static inline uint64_t hw_divider_result_nowait() {
+/*! \brief Return result of HW divide, nowait
+ *  \ingroup hardware_divider
+ *
+ * \note This is UNSAFE in that the calculation may not have been completed.
+ *
+ * \return Current result. Most significant 32 bits are the remainder, lower 32 bits are the quotient.
+ */
+static inline divmod_result_t hw_divider_result_nowait(void) {
     return hw_divider_result_threadlocal;
 }
 
-inline static uint32_t to_quotient_u32(unsigned long long int r) {
+/*! \brief Wait for a divide to complete
+ *  \ingroup hardware_divider
+ *
+ * Wait for a divide to complete
+ */
+static inline void hw_divider_wait_ready(void) {}
+
+
+/*! \brief Efficient extraction of unsigned quotient from 32p32 fixed point
+ *  \ingroup hardware_divider
+ *
+ * \param r A pair of 32-bit quotient/remainder values.
+ * \return Unsigned quotient
+ */
+inline static uint32_t to_quotient_u32(divmod_result_t r) {
     return (uint32_t) r;
 }
 
-inline static int32_t to_quotient_s32(unsigned long long int r) {
+/*! \brief Efficient extraction of signed quotient from 32p32 fixed point
+ *  \ingroup hardware_divider
+ *
+ * \param r A pair of 32-bit quotient/remainder values.
+ * \return Unsigned quotient
+ */
+inline static int32_t to_quotient_s32(divmod_result_t r) {
     return (int32_t)(uint32_t)r;
 }
 
-inline static uint32_t to_remainder_u32(unsigned long long int r) {
+/*! \brief Efficient extraction of unsigned remainder from 32p32 fixed point
+ *  \ingroup hardware_divider
+ *
+ * \param r A pair of 32-bit quotient/remainder values.
+ * \return Unsigned remainder
+ *
+ * \note On Arm this is just a 32 bit register move or a nop
+ */
+inline static uint32_t to_remainder_u32(divmod_result_t r) {
     return (uint32_t)(r >> 32u);
 }
 
-inline static int32_t to_remainder_s32(unsigned long long int r) {
+/*! \brief Efficient extraction of signed remainder from 32p32 fixed point
+ *  \ingroup hardware_divider
+ *
+ * \param r A pair of 32-bit quotient/remainder values.
+ * \return Signed remainder
+ *
+ * \note On arm this is just a 32 bit register move or a nop
+ */
+inline static int32_t to_remainder_s32(divmod_result_t r) {
     return (int32_t)(r >> 32u);
 }
 
-static inline uint32_t hw_divider_u32_quotient_wait() {
-    return to_quotient_u32(hw_divider_result_wait());
-}
-
-static inline uint32_t hw_divider_u32_remainder_wait() {
-    return to_remainder_u32(hw_divider_result_wait());
-}
-
-static inline int32_t hw_divider_s32_quotient_wait() {
-    return to_quotient_s32(hw_divider_result_wait());
-}
-
-static inline int32_t hw_divider_s32_remainder_wait() {
-    return to_remainder_s32(hw_divider_result_wait());
-}
+static inline void hw_divider_pause(void) {}
 
+/*! \brief Do an unsigned HW divide, wait for result, return quotient
+ *  \ingroup hardware_divider
+ *
+ * Divide \p a by \p b, wait for calculation to complete, return quotient.
+ *
+ * \param a The dividend
+ * \param b The divisor
+ * \return Quotient results of the divide
+ */
 static inline uint32_t hw_divider_u32_quotient(uint32_t a, uint32_t b) {
     return b ? (a / b) : (uint32_t)(-1);
 }
 
+/*! \brief Do an unsigned HW divide, wait for result, return remainder
+ *  \ingroup hardware_divider
+ *
+ * Divide \p a by \p b, wait for calculation to complete, return remainder.
+ *
+ * \param a The dividend
+ * \param b The divisor
+ * \return Remainder results of the divide
+ */
 static inline uint32_t hw_divider_u32_remainder(uint32_t a, uint32_t b) {
     return b ? (a % b) : a;
 }
 
-static inline int32_t hw_divider_s32_quotient(int32_t a, int32_t b) {
-    return b ? (a / b) : -__sign_of(a);
+/*! \brief Do a signed HW divide, wait for result, return quotient
+ *  \ingroup hardware_divider
+ *
+ * Divide \p a by \p b, wait for calculation to complete, return quotient.
+ *
+ * \param a The dividend
+ * \param b The divisor
+ * \return Quotient results of the divide
+ */
+static inline int32_t hw_divider_quotient_s32(int32_t a, int32_t b) {
+    return to_quotient_s32(hw_divider_divmod_s32(a, b));
 }
 
-static inline int32_t hw_divider_s32_remainder(int32_t a, int32_t b) {
+/*! \brief Do a signed HW divide, wait for result, return remainder
+ *  \ingroup hardware_divider
+ *
+ * Divide \p a by \p b, wait for calculation to complete, return remainder.
+ *
+ * \param a The dividend
+ * \param b The divisor
+ * \return Remainder results of the divide
+ */
+static inline int32_t hw_divider_remainder_s32(int32_t a, int32_t b) {
     return b ? (a % b) : a;
 }
 
+/*! \brief Return result of last asynchronous HW divide, unsigned quotient only
+ *  \ingroup hardware_divider
+ *
+ * This function waits for the result to be ready by calling hw_divider_wait_ready().
+ *
+ * \return Current unsigned quotient result.
+ */
+static inline uint32_t hw_divider_u32_quotient_wait(void) {
+    return to_quotient_u32(hw_divider_result_wait());
+}
+
+/*! \brief Return result of last asynchronous HW divide, signed quotient only
+ *  \ingroup hardware_divider
+ *
+ * This function waits for the result to be ready by calling hw_divider_wait_ready().
+ *
+ * \return Current signed quotient result.
+ */
+static inline int32_t hw_divider_s32_quotient_wait(void) {
+    return to_remainder_u32(hw_divider_result_wait());
+}
+
+/*! \brief Return result of last asynchronous HW divide, unsigned remainder only
+ *  \ingroup hardware_divider
+ *
+ * This function waits for the result to be ready by calling hw_divider_wait_ready().
+ *
+ * \return Current unsigned remainder result.
+ */
+static inline uint32_t hw_divider_u32_remainder_wait(void) {
+    return to_quotient_s32(hw_divider_result_wait());
+}
+
+/*! \brief Return result of last asynchronous HW divide, signed remainder only
+ *  \ingroup hardware_divider
+ *
+ * This function waits for the result to be ready by calling hw_divider_wait_ready().
+ *
+ * \return Current remainder results.
+ */
+static inline int32_t hw_divider_s32_remainder_wait(void) {
+    return to_remainder_s32(hw_divider_result_wait());
+}
+
+/*! \brief Do a hardware unsigned HW divide, wait for result, return quotient
+ *  \ingroup hardware_divider
+ *
+ * Divide \p a by \p b, wait for calculation to complete, return quotient.
+ *
+ * \param a The dividend
+ * \param b The divisor
+ * \return Quotient result of the divide
+ */
 static inline uint32_t hw_divider_u32_quotient_inlined(uint32_t a, uint32_t b) {
     return hw_divider_u32_quotient(a,b);
 }
 
+/*! \brief Do a hardware unsigned HW divide, wait for result, return remainder
+ *  \ingroup hardware_divider
+ *
+ * Divide \p a by \p b, wait for calculation to complete, return remainder.
+ *
+ * \param a The dividend
+ * \param b The divisor
+ * \return Remainder result of the divide
+ */
 static inline uint32_t hw_divider_u32_remainder_inlined(uint32_t a, uint32_t b) {
     return hw_divider_u32_remainder(a,b);
 }
 
+/*! \brief Do a hardware signed HW divide, wait for result, return quotient
+ *  \ingroup hardware_divider
+ *
+ * Divide \p a by \p b, wait for calculation to complete, return quotient.
+ *
+ * \param a The dividend
+ * \param b The divisor
+ * \return Quotient result of the divide
+ */
 static inline int32_t hw_divider_s32_quotient_inlined(int32_t a, int32_t b) {
-    return hw_divider_s32_quotient(a,b);
+    return hw_divider_quotient_s32(a,b);
 }
 
+/*! \brief Do a hardware signed HW divide, wait for result, return remainder
+ *  \ingroup hardware_divider
+ *
+ * Divide \p a by \p b, wait for calculation to complete, return remainder.
+ *
+ * \param a The dividend
+ * \param b The divisor
+ * \return Remainder result of the divide
+ */
 static inline int32_t hw_divider_s32_remainder_inlined(int32_t a, int32_t b) {
-    return hw_divider_s32_remainder(a,b);
+    return hw_divider_remainder_s32(a,b);
 }
 
 typedef uint64_t hw_divider_state_t;
 
-static inline void hw_divider_save_state(hw_divider_state_t *dest) {
+/*! \brief Save the calling cores hardware divider state
+ *  \ingroup hardware_divider
+ *
+ * Copy the current core's hardware divider state into the provided structure. This method
+ * waits for the divider results to be stable, then copies them to memory.
+ * They can be restored via hw_divider_restore_state()
+ *
+ * \param dest the location to store the divider state
+ */
+ static inline void hw_divider_save_state(hw_divider_state_t *dest) {
     *dest = hw_divider_result_threadlocal;
 }
 
-static inline void hw_divider_restore_state(hw_divider_state_t *src) {
+/*! \brief Load a saved hardware divider state into the current core's hardware divider
+ *  \ingroup hardware_divider
+ *
+ * Copy the passed hardware divider state into the hardware divider.
+ *
+ * \param src the location to load the divider state from
+ */
+ static inline void hw_divider_restore_state(hw_divider_state_t *src) {
     hw_divider_result_threadlocal = *src;
 }
 
+#ifdef __cplusplus
+}
+#endif
 #endif // _HARDWARE_DIVIDER_H
diff --git a/src/host/pico_divider/divider.c b/src/host/pico_divider/divider.c
index df0e2751..25ede84b 100644
--- a/src/host/pico_divider/divider.c
+++ b/src/host/pico_divider/divider.c
@@ -8,7 +8,7 @@
 
 // These functions save/restore divider state, so are safe to call from interrupts
 int32_t div_s32s32(int32_t a, int32_t b) {
-    return hw_divider_s32_quotient(a, b);
+    return hw_divider_quotient_s32(a, b);
 }
 
 divmod_result_t divmod_s32s32(int32_t a, int32_t b) {
diff --git a/src/rp2_common/boot_stage2/boot2_at25sf128a.S b/src/rp2_common/boot_stage2/boot2_at25sf128a.S
index be232ff1..f013ea99 100644
--- a/src/rp2_common/boot_stage2/boot2_at25sf128a.S
+++ b/src/rp2_common/boot_stage2/boot2_at25sf128a.S
@@ -86,9 +86,7 @@
 // Start of 2nd Stage Boot Code
 // ----------------------------------------------------------------------------
 
-.syntax unified
-.cpu cortex-m0plus
-.thumb
+pico_default_asm_setup
 
 .section .text
 
@@ -97,10 +95,7 @@
 // Otherwise it will be a return address -- second stage being called as a
 // function by user code, after copying out of XIP region. r3 holds SSI base,
 // r0...2 used as temporaries. Other GPRs not used.
-.global _stage2_boot
-.type _stage2_boot,%function
-.thumb_func
-_stage2_boot:
+regular_func _stage2_boot
     push {lr}
 
     // Set pad configuration:
diff --git a/src/rp2_common/boot_stage2/boot2_generic_03h.S b/src/rp2_common/boot_stage2/boot2_generic_03h.S
index a10e66ab..3c8715f0 100644
--- a/src/rp2_common/boot_stage2/boot2_generic_03h.S
+++ b/src/rp2_common/boot_stage2/boot2_generic_03h.S
@@ -20,6 +20,8 @@
 #include "hardware/regs/addressmap.h"
 #include "hardware/regs/ssi.h"
 
+pico_default_asm_setup
+
 // ----------------------------------------------------------------------------
 // Config section
 // ----------------------------------------------------------------------------
@@ -53,25 +55,19 @@
 // Start of 2nd Stage Boot Code
 // ----------------------------------------------------------------------------
 
-.cpu cortex-m0
-.thumb
-
 .section .text
 
-.global _stage2_boot
-.type _stage2_boot,%function
-.thumb_func
-_stage2_boot:
+regular_func _stage2_boot
     push {lr}
 
     ldr r3, =XIP_SSI_BASE                // Use as base address where possible
 
     // Disable SSI to allow further config
-    mov r1, #0
+    movs r1, #0
     str r1, [r3, #SSI_SSIENR_OFFSET]
 
     // Set baud rate
-    mov r1, #PICO_FLASH_SPI_CLKDIV
+    movs r1, #PICO_FLASH_SPI_CLKDIV
     str r1, [r3, #SSI_BAUDR_OFFSET]
 
     ldr r1, =(CTRLR0_XIP)
@@ -82,11 +78,11 @@ _stage2_boot:
     str r1, [r0]
 
     // NDF=0 (single 32b read)
-    mov r1, #0x0
+    movs r1, #0x0
     str r1, [r3, #SSI_CTRLR1_OFFSET]
 
     // Re-enable SSI
-    mov r1, #1
+    movs r1, #1
     str r1, [r3, #SSI_SSIENR_OFFSET]
 
 // We are now in XIP mode. Any bus accesses to the XIP address window will be
diff --git a/src/rp2_common/boot_stage2/boot2_is25lp080.S b/src/rp2_common/boot_stage2/boot2_is25lp080.S
index 80bf9d11..e2c24d42 100644
--- a/src/rp2_common/boot_stage2/boot2_is25lp080.S
+++ b/src/rp2_common/boot_stage2/boot2_is25lp080.S
@@ -80,25 +80,20 @@
 // Start of 2nd Stage Boot Code
 // ----------------------------------------------------------------------------
 
-.cpu cortex-m0
-.thumb
+pico_default_asm_setup
 
-.section .text
-
-.global _stage2_boot
-.type _stage2_boot,%function
-.thumb_func
-_stage2_boot:
+.section text
+regular_func _stage2_boot
     push {lr}
 
     ldr r3, =XIP_SSI_BASE                // Use as base address where possible
 
     // Disable SSI to allow further config
-    mov r1, #0
+    movs r1, #0
     str r1, [r3, #SSI_SSIENR_OFFSET]
 
     // Set baud rate
-    mov r1, #PICO_FLASH_SPI_CLKDIV
+    movs r1, #PICO_FLASH_SPI_CLKDIV
     str r1, [r3, #SSI_BAUDR_OFFSET]
 
 // On QSPI parts we usually need a 01h SR-write command to enable QSPI mode
@@ -113,7 +108,7 @@ program_sregs:
     str r1, [r3, #SSI_CTRLR0_OFFSET]
 
      // Enable SSI and select slave 0
-    mov r1, #1
+    movs r1, #1
     str r1, [r3, #SSI_SSIENR_OFFSET]
 
     // Check whether SR needs updating
@@ -124,7 +119,7 @@ program_sregs:
     beq skip_sreg_programming
 
     // Send write enable command
-    mov r1, #CMD_WRITE_ENABLE
+    movs r1, #CMD_WRITE_ENABLE
     str r1, [r3, #SSI_DR0_OFFSET]
 
     // Poll for completion and discard RX
@@ -132,9 +127,9 @@ program_sregs:
     ldr r1, [r3, #SSI_DR0_OFFSET]
 
     // Send status write command followed by data bytes
-    mov r1, #CMD_WRITE_STATUS
+    movs r1, #CMD_WRITE_STATUS
     str r1, [r3, #SSI_DR0_OFFSET]
-    mov r0, #0
+    movs r0, #0
     str r2, [r3, #SSI_DR0_OFFSET]
 
     bl wait_ssi_ready
@@ -145,7 +140,7 @@ program_sregs:
 1:
     ldr r0, =CMD_READ_STATUS
     bl read_flash_sreg
-    mov r1, #1
+    movs r1, #1
     tst r0, r1
     bne 1b
 
@@ -157,7 +152,7 @@ skip_sreg_programming:
 //    bl wait_ssi_ready
 
     // Disable SSI again so that it can be reconfigured
-    mov r1, #0
+    movs r1, #0
     str r1, [r3, #SSI_SSIENR_OFFSET]
 #endif
 
@@ -182,7 +177,7 @@ dummy_read:
     ldr r1, =(CTRLR0_ENTER_XIP)
     str r1, [r3, #SSI_CTRLR0_OFFSET]
 
-    mov r1, #0x0                    // NDF=0 (single 32b read)
+    movs r1, #0x0                   // NDF=0 (single 32b read)
     str r1, [r3, #SSI_CTRLR1_OFFSET]
 
 #define SPI_CTRLR0_ENTER_XIP \
@@ -197,12 +192,12 @@ dummy_read:
     ldr r0, =(XIP_SSI_BASE + SSI_SPI_CTRLR0_OFFSET)  // SPI_CTRL0 Register
     str r1, [r0]
 
-    mov r1, #1                      // Re-enable SSI
+    movs r1, #1                     // Re-enable SSI
     str r1, [r3, #SSI_SSIENR_OFFSET]
 
-    mov r1, #CMD_READ 
+    movs r1, #CMD_READ
     str r1, [r3, #SSI_DR0_OFFSET]   // Push SPI command into TX FIFO
-    mov r1, #MODE_CONTINUOUS_READ   // 32-bit: 24 address bits (we don't care, so 0) and M[7:4]=1010
+    movs r1, #MODE_CONTINUOUS_READ  // 32-bit: 24 address bits (we don't care, so 0) and M[7:4]=1010
     str r1, [r3, #SSI_DR0_OFFSET]   // Push Address into TX FIFO - this will trigger the transaction
 
     // Poll for completion
@@ -218,7 +213,7 @@ dummy_read:
 // to APM mode and generate a 28-bit address phase with the extra nibble set
 // to 4'b0000).
 
-    mov r1, #0
+    movs r1, #0
     str r1, [r3, #SSI_SSIENR_OFFSET]   // Disable SSI (and clear FIFO) to allow further config
 
 // Note that the INST_L field is used to select what XIP data gets pushed into
@@ -240,7 +235,7 @@ configure_ssi:
     ldr r0, =(XIP_SSI_BASE + SSI_SPI_CTRLR0_OFFSET)
     str r1, [r0]
 
-    mov r1, #1
+    movs r1, #1
     str r1, [r3, #SSI_SSIENR_OFFSET]   // Re-enable SSI
 
 // We are now in XIP mode, with all transactions using Dual I/O and only
diff --git a/src/rp2_common/boot_stage2/boot2_usb_blinky.S b/src/rp2_common/boot_stage2/boot2_usb_blinky.S
index 74c47a3e..0249a455 100644
--- a/src/rp2_common/boot_stage2/boot2_usb_blinky.S
+++ b/src/rp2_common/boot_stage2/boot2_usb_blinky.S
@@ -4,6 +4,8 @@
  * SPDX-License-Identifier: BSD-3-Clause
  */
 
+#include "pico/asm_helper.S"
+
 // Stub second stage which calls into USB bootcode, with parameters.
 // USB boot takes two parameters:
 // - A GPIO mask for activity LED -- if mask is 0, don't touch GPIOs at all
@@ -19,17 +21,12 @@
 #define ACTIVITY_LED 0
 #define BOOT_MODE USB_BOOT_MSD_AND_PICOBOOT
 
-.cpu cortex-m0
-.thumb
+pico_default_asm_setup
 
 .section .text
 
-.global _stage2_boot
-.type _stage2_boot,%function
-
-.thumb_func
-_stage2_boot:
-    mov r7, #0x14               // Pointer to _well_known pointer table in ROM
+regular_func _stage2_boot
+    movs r7, #0x14               // Pointer to _well_known pointer table in ROM
     ldrh r0, [r7, #0]           // Offset 0 is 16 bit pointer to function table
     ldrh r7, [r7, #4]           // Offset 4 is 16 bit pointer to table lookup routine
     ldr r1, =('U' | ('B' << 8)) // Symbol for USB Boot
@@ -39,7 +36,7 @@ _stage2_boot:
 
     mov r7, r0
     ldr r0, =(1u << ACTIVITY_LED) // Mask of which GPIO (or GPIOs) to use
-    mov r1, #BOOT_MODE
+    movs r1, #BOOT_MODE
     blx r7
 
 dead:
diff --git a/src/rp2_common/boot_stage2/boot2_w25q080.S b/src/rp2_common/boot_stage2/boot2_w25q080.S
index ad3238e2..7309f53b 100644
--- a/src/rp2_common/boot_stage2/boot2_w25q080.S
+++ b/src/rp2_common/boot_stage2/boot2_w25q080.S
@@ -86,9 +86,7 @@
 // Start of 2nd Stage Boot Code
 // ----------------------------------------------------------------------------
 
-.syntax unified
-.cpu cortex-m0plus
-.thumb
+pico_default_asm_setup
 
 .section .text
 
@@ -97,10 +95,7 @@
 // Otherwise it will be a return address -- second stage being called as a
 // function by user code, after copying out of XIP region. r3 holds SSI base,
 // r0...2 used as temporaries. Other GPRs not used.
-.global _stage2_boot
-.type _stage2_boot,%function
-.thumb_func
-_stage2_boot:
+regular_func _stage2_boot
     push {lr}
 
     // Set pad configuration:
diff --git a/src/rp2_common/boot_stage2/boot2_w25x10cl.S b/src/rp2_common/boot_stage2/boot2_w25x10cl.S
index 02628d4e..89d2cd10 100644
--- a/src/rp2_common/boot_stage2/boot2_w25x10cl.S
+++ b/src/rp2_common/boot_stage2/boot2_w25x10cl.S
@@ -40,6 +40,8 @@
 #define PICO_FLASH_SPI_CLKDIV 4
 #endif
 
+pico_default_asm_setup
+
 // ----------------------------------------------------------------------------
 // The "System Control Block" is a set of internal Cortex-M0+ control registers
 // that are memory mapped and accessed like any other H/W register.  They have
@@ -69,31 +71,25 @@
 // Start of 2nd Stage Boot Code
 // ----------------------------------------------------------------------------
 
-.cpu cortex-m0
-.thumb
-
 .org 0
 
 .section .text
 
 // This code will get copied to 0x20000000 and then executed
 
-.global _stage2_boot
-.type _stage2_boot,%function
-.thumb_func
-_stage2_boot:
+regular_func _stage2_boot
     push {lr}
     ldr r3, =XIP_SSI_BASE                // Use as base address where possible
 
 // We are primarily interested in setting up Flash for DSPI XIP w/ continuous read
 
-    mov r1, #0
+    movs r1, #0
     str r1, [r3, #SSI_SSIENR_OFFSET] // Disable SSI to allow further config
 
 // The Boot ROM sets a very conservative SPI clock frequency to be sure it can
 // read the initial 256 bytes from any device.  Here we can be more aggressive.
 
-    mov r1, #PICO_FLASH_SPI_CLKDIV
+    movs r1, #PICO_FLASH_SPI_CLKDIV
     str r1, [r3, #SSI_BAUDR_OFFSET]  // Set SSI Clock
 
 // First we need to send the initial command to get us in to Fast Read Dual I/O
@@ -116,7 +112,7 @@ _stage2_boot:
     ldr r1, =(CTRLR0_ENTER_XIP)
     str r1, [r3, #SSI_CTRLR0_OFFSET]
 
-    mov r1, #0x0                    // NDF=0 (single 32b read)
+    movs r1, #0x0                   // NDF=0 (single 32b read)
     str r1, [r3, #SSI_CTRLR1_OFFSET]
 
 #define SPI_CTRLR0_ENTER_XIP \
@@ -131,18 +127,18 @@ _stage2_boot:
     ldr r0, =(XIP_SSI_BASE + SSI_SPI_CTRLR0_OFFSET)  // SPI_CTRL0 Register
     str r1, [r0]
 
-    mov r1, #1                      // Re-enable SSI
+    movs r1, #1                     // Re-enable SSI
     str r1, [r3, #SSI_SSIENR_OFFSET]
 
-    mov r1, #W25X10CL_CMD_READ_DATA_FAST_DUAL_IO    // 8b command = 0xBB
+    movs r1, #W25X10CL_CMD_READ_DATA_FAST_DUAL_IO   // 8b command = 0xBB
     str r1, [r3, #SSI_DR0_OFFSET]   // Push SPI command into TX FIFO
-    mov r1, #0x0000002              // 28-bit Address for dummy read = 0x000000 + 0x2 Mode bits to set M[5:4]=10
+    movs r1, #0x0000002             // 28-bit Address for dummy read = 0x000000 + 0x2 Mode bits to set M[5:4]=10
     str r1, [r3, #SSI_DR0_OFFSET]   // Push Address into TX FIFO - this will trigger the transaction
 
 // Now we wait for the read transaction to complete by monitoring the SSI
 // status register and checking for the "RX FIFO Not Empty" flag to assert.
 
-    mov r1, #SSI_SR_RFNE_BITS
+    movs r1, #SSI_SR_RFNE_BITS
 00:
     ldr r0, [r3, #SSI_SR_OFFSET]    // Read status register
     tst r0, r1                      // RFNE status flag set?
@@ -158,7 +154,7 @@ _stage2_boot:
 // to APM mode and generate a 28-bit address phase with the extra nibble set
 // to 4'b0000).
 
-    mov r1, #0
+    movs r1, #0
     str r1, [r3, #SSI_SSIENR_OFFSET]   // Disable SSI (and clear FIFO) to allow further config
 
 // Note that the INST_L field is used to select what XIP data gets pushed into
@@ -180,7 +176,7 @@ _stage2_boot:
     ldr r0, =(XIP_SSI_BASE + SSI_SPI_CTRLR0_OFFSET)
     str r1, [r0]
 
-    mov r1, #1
+    movs r1, #1
     str r1, [r3, #SSI_SSIENR_OFFSET]   // Re-enable SSI
 
 // We are now in XIP mode, with all transactions using Dual I/O and only
diff --git a/src/rp2_common/hardware_clocks/clocks.c b/src/rp2_common/hardware_clocks/clocks.c
index 75078931..152a9641 100644
--- a/src/rp2_common/hardware_clocks/clocks.c
+++ b/src/rp2_common/hardware_clocks/clocks.c
@@ -48,7 +48,7 @@ bool clock_configure(enum clock_index clk_index, uint32_t src, uint32_t auxsrc,
         return false;
 
     // Div register is 24.8 int.frac divider so multiply by 2^8 (left shift by 8)
-    div = (uint32_t) (((uint64_t) src_freq << 8) / freq);
+    div = (uint32_t) (((uint64_t) src_freq << CLOCKS_CLK_GPOUT0_DIV_INT_LSB) / freq);
 
     clock_hw_t *clock = &clocks_hw->clk[clk_index];
 
@@ -76,14 +76,9 @@ bool clock_configure(enum clock_index clk_index, uint32_t src, uint32_t auxsrc,
         if (configured_freq[clk_index] > 0) {
             // Delay for 3 cycles of the target clock, for ENABLE propagation.
             // Note XOSC_COUNT is not helpful here because XOSC is not
-            // necessarily running, nor is timer... so, 3 cycles per loop:
+            // necessarily running, nor is timer...:
             uint delay_cyc = configured_freq[clk_sys] / configured_freq[clk_index] + 1;
-            unified_asm (
-                "1: \n\t"
-                "subs %0, #1 \n\t"
-                "bne 1b"
-                : "+r" (delay_cyc)
-            );
+            busy_wait_at_least_cycles(delay_cyc * 3);
         }
     }
 
diff --git a/src/rp2_common/hardware_divider/divider.S b/src/rp2_common/hardware_divider/divider.S
index 55134f2c..9d90e24a 100644
--- a/src/rp2_common/hardware_divider/divider.S
+++ b/src/rp2_common/hardware_divider/divider.S
@@ -2,9 +2,7 @@
 #include "hardware/regs/addressmap.h"
 #include "hardware/regs/sio.h"
 
-.syntax unified
-.cpu cortex-m0plus
-.thumb
+pico_default_asm_setup
 
 // tag::hw_div_s32[]
 regular_func_with_section hw_divider_divmod_s32
diff --git a/src/rp2_common/hardware_divider/include/hardware/divider.h b/src/rp2_common/hardware_divider/include/hardware/divider.h
index c5b91dec..cab3e043 100644
--- a/src/rp2_common/hardware_divider/include/hardware/divider.h
+++ b/src/rp2_common/hardware_divider/include/hardware/divider.h
@@ -85,7 +85,7 @@ static inline void hw_divider_wait_ready(void) {
     // we use one less register and instruction than gcc which uses a TST instruction
 
     uint32_t tmp; // allow compiler to pick scratch register
-    unified_asm (
+    pico_default_asm (
     "hw_divider_result_loop_%=:"
     "ldr %0, [%1, %2]\n\t"
     "lsrs %0, %0, #1\n\t"
@@ -177,29 +177,29 @@ static inline int32_t hw_divider_s32_remainder_wait(void) {
 /*! \brief Do a signed HW divide and wait for result
  *  \ingroup hardware_divider
  *
- * Divide \p a by \p b, wait for calculation to complete, return result as a fixed point 32p32 value.
+ * Divide \p a by \p b, wait for calculation to complete, return result as a pair of 32-bit quotient/remainder values.
  *
  * \param a The dividend
  * \param b The divisor
- * \return Results of divide as a 32p32 fixed point value.
+ * \return Results of divide as a pair of 32-bit quotient/remainder values.
  */
 divmod_result_t hw_divider_divmod_s32(int32_t a, int32_t b);
 
 /*! \brief Do an unsigned HW divide and wait for result
  *  \ingroup hardware_divider
  *
- * Divide \p a by \p b, wait for calculation to complete, return result as a fixed point 32p32 value.
+ * Divide \p a by \p b, wait for calculation to complete, return result as a pair of 32-bit quotient/remainder values.
  *
  * \param a The dividend
  * \param b The divisor
- * \return Results of divide as a 32p32 fixed point value.
+ * \return Results of divide as a pair of 32-bit quotient/remainder values.
  */
 divmod_result_t hw_divider_divmod_u32(uint32_t a, uint32_t b);
 
 /*! \brief Efficient extraction of unsigned quotient from 32p32 fixed point
  *  \ingroup hardware_divider
  *
- * \param r 32p32 fixed point value.
+ * \param r A pair of 32-bit quotient/remainder values.
  * \return Unsigned quotient
  */
 inline static uint32_t to_quotient_u32(divmod_result_t r) {
@@ -209,7 +209,7 @@ inline static uint32_t to_quotient_u32(divmod_result_t r) {
 /*! \brief Efficient extraction of signed quotient from 32p32 fixed point
  *  \ingroup hardware_divider
  *
- * \param r 32p32 fixed point value.
+ * \param r A pair of 32-bit quotient/remainder values.
  * \return Unsigned quotient
  */
 inline static int32_t to_quotient_s32(divmod_result_t r) {
@@ -219,7 +219,7 @@ inline static int32_t to_quotient_s32(divmod_result_t r) {
 /*! \brief Efficient extraction of unsigned remainder from 32p32 fixed point
  *  \ingroup hardware_divider
  *
- * \param r 32p32 fixed point value.
+ * \param r A pair of 32-bit quotient/remainder values.
  * \return Unsigned remainder
  *
  * \note On Arm this is just a 32 bit register move or a nop
@@ -231,7 +231,7 @@ inline static uint32_t to_remainder_u32(divmod_result_t r) {
 /*! \brief Efficient extraction of signed remainder from 32p32 fixed point
  *  \ingroup hardware_divider
  *
- * \param r 32p32 fixed point value.
+ * \param r A pair of 32-bit quotient/remainder values.
  * \return Signed remainder
  *
  * \note On arm this is just a 32 bit register move or a nop
@@ -296,7 +296,7 @@ static inline int32_t hw_divider_remainder_s32(int32_t a, int32_t b) {
  *  \ingroup hardware_divider
  */
 static inline void hw_divider_pause(void) {
-    unified_asm (
+    pico_default_asm (
     "b _1_%=\n"
     "_1_%=:\n"
     "b _2_%=\n"
diff --git a/src/rp2_common/hardware_divider/include/hardware/divider_helper.S b/src/rp2_common/hardware_divider/include/hardware/divider_helper.S
index e69fab59..ff4ed919 100644
--- a/src/rp2_common/hardware_divider/include/hardware/divider_helper.S
+++ b/src/rp2_common/hardware_divider/include/hardware/divider_helper.S
@@ -4,6 +4,7 @@
  * SPDX-License-Identifier: BSD-3-Clause
  */
 
+// Note this file is always included by another, so does not do pico_default_asm_setup
 #include "hardware/regs/addressmap.h"
 #include "hardware/regs/sio.h"
 
diff --git a/src/rp2_common/hardware_exception/exception.c b/src/rp2_common/hardware_exception/exception.c
index dd89401f..8e141a69 100644
--- a/src/rp2_common/hardware_exception/exception.c
+++ b/src/rp2_common/hardware_exception/exception.c
@@ -5,7 +5,7 @@
  */
 
 #include "hardware/exception.h"
-#include "hardware/regs/m0plus.h"
+#include "hardware/platform_defs.h"
 #include "hardware/structs/scb.h"
 #include "hardware/sync.h"
 #include "pico/assert.h"
diff --git a/src/rp2_common/hardware_exception/include/hardware/exception.h b/src/rp2_common/hardware_exception/include/hardware/exception.h
index b8bddea2..63af843e 100644
--- a/src/rp2_common/hardware_exception/include/hardware/exception.h
+++ b/src/rp2_common/hardware_exception/include/hardware/exception.h
@@ -9,7 +9,6 @@
 
 #include "pico.h"
 #include "hardware/address_mapped.h"
-#include "hardware/regs/m0plus.h"
 
 /** \file exception.h
  *  \defgroup hardware_exception hardware_exception
diff --git a/src/rp2_common/hardware_gpio/gpio.c b/src/rp2_common/hardware_gpio/gpio.c
index 6e4261fc..4945738d 100644
--- a/src/rp2_common/hardware_gpio/gpio.c
+++ b/src/rp2_common/hardware_gpio/gpio.c
@@ -244,8 +244,8 @@ void gpio_set_input_enabled(uint gpio, bool enabled) {
 }
 
 void gpio_init(uint gpio) {
-    sio_hw->gpio_oe_clr = 1ul << gpio;
-    sio_hw->gpio_clr = 1ul << gpio;
+    gpio_set_dir(gpio, GPIO_IN);
+    gpio_put(gpio, 0);
     gpio_set_function(gpio, GPIO_FUNC_SIO);
 }
 
diff --git a/src/rp2_common/hardware_irq/irq.c b/src/rp2_common/hardware_irq/irq.c
index 3e66c189..1fbc3cc8 100644
--- a/src/rp2_common/hardware_irq/irq.c
+++ b/src/rp2_common/hardware_irq/irq.c
@@ -186,7 +186,7 @@ static inline int8_t slot_diff(struct irq_handler_chain_slot *to, struct irq_han
     int32_t result = 0xaaaa;
     // return (to - from);
     // note this implementation has limited range, but is fine for plenty more than -128->127 result
-    unified_asm (
+    pico_default_asm (
          "subs %1, %2\n"
          "adcs %1, %1\n" // * 2 (and + 1 if negative for rounding)
          "muls %0, %1\n"
diff --git a/src/rp2_common/hardware_irq/irq_handler_chain.S b/src/rp2_common/hardware_irq/irq_handler_chain.S
index bb65fb16..cd15ce77 100644
--- a/src/rp2_common/hardware_irq/irq_handler_chain.S
+++ b/src/rp2_common/hardware_irq/irq_handler_chain.S
@@ -6,11 +6,11 @@
 
 #include "pico.h"
 #include "hardware/irq.h"
+#include "pico/asm_helper.S"
+
+pico_default_asm_setup
 
 #if !PICO_DISABLE_SHARED_IRQ_HANDLERS
-.syntax unified
-.cpu cortex-m0plus
-.thumb
 
 .data
 .align 2
diff --git a/src/rp2_common/hardware_sync/include/hardware/sync.h b/src/rp2_common/hardware_sync/include/hardware/sync.h
index c5715421..2acf5a90 100644
--- a/src/rp2_common/hardware_sync/include/hardware/sync.h
+++ b/src/rp2_common/hardware_sync/include/hardware/sync.h
@@ -116,7 +116,7 @@ typedef volatile uint32_t spin_lock_t;
  */
 #if !__has_builtin(__sev)
 __force_inline static void __sev(void) {
-    unified_asm ("sev");
+    pico_default_asm ("sev");
 }
 #endif
 
@@ -128,7 +128,7 @@ __force_inline static void __sev(void) {
  */
 #if !__has_builtin(__wfe)
 __force_inline static void __wfe(void) {
-    unified_asm ("wfe");
+    pico_default_asm ("wfe");
 }
 #endif
 
@@ -139,7 +139,7 @@ __force_inline static void __wfe(void) {
  */
 #if !__has_builtin(__wfi)
 __force_inline static void __wfi(void) {
-    unified_asm ("wfi");
+    pico_default_asm ("wfi");
 }
 #endif
 
@@ -150,7 +150,7 @@ __force_inline static void __wfi(void) {
  * instruction will be observed before any explicit access after the instruction.
  */
 __force_inline static void __dmb(void) {
-    unified_asm ("dmb" : : : "memory");
+    pico_default_asm ("dmb" : : : "memory");
 }
 
 /*! \brief Insert a DSB instruction in to the code path.
@@ -161,7 +161,7 @@ __force_inline static void __dmb(void) {
  * accesses before this instruction complete.
  */
 __force_inline static void __dsb(void) {
-    unified_asm ("dsb" : : : "memory");
+    pico_default_asm ("dsb" : : : "memory");
 }
 
 /*! \brief Insert a ISB instruction in to the code path.
@@ -172,7 +172,7 @@ __force_inline static void __dsb(void) {
  * the ISB instruction has been completed.
  */
 __force_inline static void __isb(void) {
-    unified_asm ("isb");
+    pico_default_asm ("isb");
 }
 
 /*! \brief Acquire a memory fence
@@ -213,8 +213,8 @@ __force_inline static void __mem_fence_release(void) {
  */
 __force_inline static uint32_t save_and_disable_interrupts(void) {
     uint32_t status;
-    unified_asm ("mrs %0, PRIMASK" : "=r" (status)::);
-    unified_asm ("cpsid i");
+    pico_default_asm ("mrs %0, PRIMASK" : "=r" (status)::);
+    pico_default_asm ("cpsid i");
     return status;
 }
 
@@ -224,7 +224,7 @@ __force_inline static uint32_t save_and_disable_interrupts(void) {
  * \param status Previous interrupt status from save_and_disable_interrupts()
   */
 __force_inline static void restore_interrupts(uint32_t status) {
-    unified_asm ("msr PRIMASK,%0"::"r" (status) : );
+    pico_default_asm ("msr PRIMASK,%0"::"r" (status) : );
 }
 
 /*! \brief Get HW Spinlock instance from number
diff --git a/src/rp2_common/pico_bit_ops/bit_ops_aeabi.S b/src/rp2_common/pico_bit_ops/bit_ops_aeabi.S
index 7b83edb9..02837dcf 100644
--- a/src/rp2_common/pico_bit_ops/bit_ops_aeabi.S
+++ b/src/rp2_common/pico_bit_ops/bit_ops_aeabi.S
@@ -4,14 +4,10 @@
  * SPDX-License-Identifier: BSD-3-Clause
  */
 
-.syntax unified
-.cpu cortex-m0plus
-.thumb
-
 #include "pico/asm_helper.S"
 #include "pico/bootrom.h"
 
-__pre_init __aeabi_bits_init, 00010
+pico_default_asm_setup
 
 .macro bits_section name
 #if PICO_BITS_IN_RAM
@@ -21,6 +17,8 @@ __pre_init __aeabi_bits_init, 00010
 #endif
 .endm
 
+__pre_init __aeabi_bits_init, 00010
+
 .section .data.aeabi_bits_funcs
 .global aeabi_bits_funcs, aeabi_bits_funcs_end
 .equ BITS_FUNC_COUNT, 4
diff --git a/src/rp2_common/pico_cyw43_driver/cyw43_bus_pio_spi.c b/src/rp2_common/pico_cyw43_driver/cyw43_bus_pio_spi.c
index 6976d479..16993301 100644
--- a/src/rp2_common/pico_cyw43_driver/cyw43_bus_pio_spi.c
+++ b/src/rp2_common/pico_cyw43_driver/cyw43_bus_pio_spi.c
@@ -63,7 +63,7 @@ static uint32_t counter = 0;
 
 //#define SWAP32(A) ((((A) & 0xff000000U) >> 8) | (((A) & 0xff0000U) << 8) | (((A) & 0xff00U) >> 8) | (((A) & 0xffU) << 8))
 __force_inline static uint32_t __swap16x2(uint32_t a) {
-    unified_asm ("rev16 %0, %0" : "+l" (a) : : );
+    pico_default_asm ("rev16 %0, %0" : "+l" (a) : : );
     return a;
 }
 #define SWAP32(a) __swap16x2(a)
diff --git a/src/rp2_common/pico_divider/divider.S b/src/rp2_common/pico_divider/divider.S
index c095f350..493e39dd 100644
--- a/src/rp2_common/pico_divider/divider.S
+++ b/src/rp2_common/pico_divider/divider.S
@@ -4,17 +4,12 @@
  * SPDX-License-Identifier: BSD-3-Clause
  */
 
+#include "pico/asm_helper.S"
 #include "hardware/regs/addressmap.h"
 #include "hardware/divider_helper.S"
 
-.syntax unified
-.cpu cortex-m0plus
-.thumb
-
 // PICO_CONFIG: PICO_DIVIDER_DISABLE_INTERRUPTS, Disable interrupts around division such that divider state need not be saved/restored in exception handlers, default=0, group=pico_divider
 
-#include "pico/asm_helper.S"
-
 // PICO_CONFIG: PICO_DIVIDER_CALL_IDIV0, Whether 32 bit division by zero should call __aeabi_idiv0, default=1, group=pico_divider
 #ifndef PICO_DIVIDER_CALL_IDIV0
 #define PICO_DIVIDER_CALL_IDIV0 1
@@ -25,6 +20,8 @@
 #define PICO_DIVIDER_CALL_LDIV0 1
 #endif
 
+pico_default_asm_setup
+
 // PICO_CONFIG: PICO_DIVIDER_IN_RAM, Whether divider functions should be placed in RAM, default=0, group=pico_divider
 .macro div_section name
 #if PICO_DIVIDER_IN_RAM
@@ -37,11 +34,11 @@
 @ wait 8-n cycles for the hardware divider
 .macro wait_div n
 .rept (8-\n) / 2
-  b 9f
+    b 9f
 9:
 .endr
 .if (8-\n) % 2
- nop
+    nop
 .endif
 .endm
 
diff --git a/src/rp2_common/pico_double/double_aeabi.S b/src/rp2_common/pico_double/double_aeabi.S
index 04cf22c6..425385e1 100644
--- a/src/rp2_common/pico_double/double_aeabi.S
+++ b/src/rp2_common/pico_double/double_aeabi.S
@@ -8,11 +8,9 @@
 #include "pico/bootrom/sf_table.h"
 #include "hardware/divider_helper.S"
 
-__pre_init __aeabi_double_init, 00020
+pico_default_asm_setup
 
-.syntax unified
-.cpu cortex-m0plus
-.thumb
+__pre_init __aeabi_double_init, 00020
 
 .macro double_section name
 #if PICO_DOUBLE_IN_RAM
diff --git a/src/rp2_common/pico_double/double_none.S b/src/rp2_common/pico_double/double_none.S
index feded31c..89fb1177 100644
--- a/src/rp2_common/pico_double/double_none.S
+++ b/src/rp2_common/pico_double/double_none.S
@@ -7,9 +7,7 @@
 #include "pico/asm_helper.S"
 #include "pico/bootrom/sf_table.h"
 
-.syntax unified
-.cpu cortex-m0plus
-.thumb
+pico_default_asm_setup
 
     wrapper_func __aeabi_dadd
     wrapper_func __aeabi_ddiv
diff --git a/src/rp2_common/pico_double/double_v1_rom_shim.S b/src/rp2_common/pico_double/double_v1_rom_shim.S
index 6dac28ab..2e65170f 100644
--- a/src/rp2_common/pico_double/double_v1_rom_shim.S
+++ b/src/rp2_common/pico_double/double_v1_rom_shim.S
@@ -17,9 +17,7 @@
 
 #include "pico/asm_helper.S"
 
-.syntax unified
-.cpu cortex-m0plus
-.thumb
+pico_default_asm_setup
 
 .macro double_section name
 // todo separate flag for shims?
diff --git a/src/rp2_common/pico_float/float_aeabi.S b/src/rp2_common/pico_float/float_aeabi.S
index b3b6c04f..e2862b5c 100644
--- a/src/rp2_common/pico_float/float_aeabi.S
+++ b/src/rp2_common/pico_float/float_aeabi.S
@@ -10,9 +10,7 @@
 
 __pre_init __aeabi_float_init, 00020
 
-.syntax unified
-.cpu cortex-m0plus
-.thumb
+pico_default_asm_setup
 
 .macro float_section name
 #if PICO_FLOAT_IN_RAM
diff --git a/src/rp2_common/pico_float/float_none.S b/src/rp2_common/pico_float/float_none.S
index db8e1268..5b1916a7 100644
--- a/src/rp2_common/pico_float/float_none.S
+++ b/src/rp2_common/pico_float/float_none.S
@@ -7,9 +7,7 @@
 #include "pico/asm_helper.S"
 #include "pico/bootrom/sf_table.h"
 
-.syntax unified
-.cpu cortex-m0plus
-.thumb
+pico_default_asm_setup
 
 wrapper_func __aeabi_fadd
 wrapper_func __aeabi_fdiv
diff --git a/src/rp2_common/pico_float/float_v1_rom_shim.S b/src/rp2_common/pico_float/float_v1_rom_shim.S
index 50df612f..665a6103 100644
--- a/src/rp2_common/pico_float/float_v1_rom_shim.S
+++ b/src/rp2_common/pico_float/float_v1_rom_shim.S
@@ -7,14 +7,13 @@
 #include "pico/asm_helper.S"
 
 #if PICO_FLOAT_SUPPORT_ROM_V1 && PICO_RP2040_B0_SUPPORTED
-.syntax unified
-.cpu cortex-m0plus
-.thumb
 
 #ifndef PICO_FLOAT_IN_RAM
 #define PICO_FLOAT_IN_RAM 0
 #endif
 
+pico_default_asm_setup
+
 .macro float_section name
 // todo separate flag for shims?
 #if PICO_FLOAT_IN_RAM
diff --git a/src/rp2_common/pico_int64_ops/pico_int64_ops_aeabi.S b/src/rp2_common/pico_int64_ops/pico_int64_ops_aeabi.S
index 114a1aa8..a1ad7e21 100644
--- a/src/rp2_common/pico_int64_ops/pico_int64_ops_aeabi.S
+++ b/src/rp2_common/pico_int64_ops/pico_int64_ops_aeabi.S
@@ -4,12 +4,10 @@
  * SPDX-License-Identifier: BSD-3-Clause
  */
 
-.syntax unified
-.cpu cortex-m0plus
-.thumb
-
 #include "pico/asm_helper.S"
 
+pico_default_asm_setup
+
 .macro int64_section name
 #if PICO_INT64_OPS_IN_RAM
 .section RAM_SECTION_NAME(\name), "ax"
diff --git a/src/rp2_common/pico_mem_ops/mem_ops_aeabi.S b/src/rp2_common/pico_mem_ops/mem_ops_aeabi.S
index 85409053..cb9bc89f 100644
--- a/src/rp2_common/pico_mem_ops/mem_ops_aeabi.S
+++ b/src/rp2_common/pico_mem_ops/mem_ops_aeabi.S
@@ -4,13 +4,11 @@
  * SPDX-License-Identifier: BSD-3-Clause
  */
 
-.syntax unified
-.cpu cortex-m0plus
-.thumb
-
 #include "pico/asm_helper.S"
 #include "pico/bootrom.h"
 
+pico_default_asm_setup
+
 __pre_init __aeabi_mem_init, 00001
 
 .macro mem_section name
diff --git a/src/rp2_common/pico_multicore/multicore.c b/src/rp2_common/pico_multicore/multicore.c
index 1622145e..04856efc 100644
--- a/src/rp2_common/pico_multicore/multicore.c
+++ b/src/rp2_common/pico_multicore/multicore.c
@@ -75,7 +75,7 @@ bool multicore_fifo_pop_timeout_us(uint64_t timeout_us, uint32_t *out) {
 static uint32_t __attribute__((section(".stack1"))) core1_stack[PICO_CORE1_STACK_SIZE / sizeof(uint32_t)];
 
 static void __attribute__ ((naked)) core1_trampoline(void) {
-    unified_asm ("pop {r0, r1, pc}");
+    pico_default_asm ("pop {r0, r1, pc}");
 }
 
 int core1_wrapper(int (*entry)(void), void *stack_base) {
diff --git a/src/rp2_common/pico_platform/include/pico/asm_helper.S b/src/rp2_common/pico_platform/include/pico/asm_helper.S
index 3e01c665..23a92583 100644
--- a/src/rp2_common/pico_platform/include/pico/asm_helper.S
+++ b/src/rp2_common/pico_platform/include/pico/asm_helper.S
@@ -6,6 +6,15 @@
 
 #include "pico.h"
 
+# note we don't do this by default in this file for backwards comaptibility with user code
+# that may include this file, but not use unified syntax. Note that this macro does equivalent
+# setup to the pico_default_asm macro for inline assembly in C code.
+.macro pico_default_asm_setup
+.syntax unified
+.cpu cortex-m0plus
+.thumb
+.endm
+
 // do not put align in here as it is used mid function sometimes
 .macro regular_func x
 .global \x
diff --git a/src/rp2_common/pico_platform/include/pico/platform.h b/src/rp2_common/pico_platform/include/pico/platform.h
index 91606deb..fde69af7 100644
--- a/src/rp2_common/pico_platform/include/pico/platform.h
+++ b/src/rp2_common/pico_platform/include/pico/platform.h
@@ -335,13 +335,13 @@ extern "C" {
 #define MIN(a, b) ((b)>(a)?(a):(b))
 #endif
 
-#define unified_asm(...) __asm volatile (".syntax unified\n" __VA_ARGS__)
+#define pico_default_asm(...) __asm volatile (".syntax unified\n" __VA_ARGS__)
 
 /*! \brief Execute a breakpoint instruction
  *  \ingroup pico_platform
  */
 static inline void __breakpoint(void) {
-    unified_asm ("bkpt #0");
+    pico_default_asm ("bkpt #0");
 }
 
 /*! \brief Ensure that the compiler does not move memory access across this method call
@@ -357,7 +357,7 @@ static inline void __breakpoint(void) {
  * might - even above the memory store!)
  */
 __force_inline static void __compiler_memory_barrier(void) {
-    unified_asm ("" : : : "memory");
+    pico_default_asm ("" : : : "memory");
 }
 
 /*! \brief Macro for converting memory addresses to 32 bit addresses suitable for DMA
@@ -443,7 +443,7 @@ static __force_inline void tight_loop_contents(void) {}
  * \return a * b
  */
 __force_inline static int32_t __mul_instruction(int32_t a, int32_t b) {
-    unified_asm ("muls %0, %1" : "+l" (a) : "l" (b) : );
+    pico_default_asm ("muls %0, %1" : "+l" (a) : "l" (b) : );
     return a;
 }
 
@@ -477,19 +477,15 @@ __force_inline static int32_t __mul_instruction(int32_t a, int32_t b) {
  *
  * \return the exception number if the CPU is handling an exception, or 0 otherwise
  */
-static inline uint __get_current_exception(void) {
+static __force_inline uint __get_current_exception(void) {
     uint exception;
-    unified_asm ("mrs %0, ipsr" : "=l" (exception));
+    pico_default_asm ("mrs %0, ipsr" : "=l" (exception));
     return exception;
 }
 
 #define WRAPPER_FUNC(x) __wrap_ ## x
 #define REAL_FUNC(x) __real_ ## x
 
-#ifdef __cplusplus
-}
-#endif
-
 /*! \brief Helper method to busy-wait for at least the given number of cycles
  *  \ingroup pico_platform
  *
@@ -505,7 +501,7 @@ static inline uint __get_current_exception(void) {
  * \param minimum_cycles the minimum number of system clock cycles to delay for
  */
 static inline void busy_wait_at_least_cycles(uint32_t minimum_cycles) {
-    unified_asm (
+    pico_default_asm (
         "1: subs %0, #3\n"
         "bcs 1b\n"
         : "+r" (minimum_cycles) : : "memory"
@@ -521,6 +517,10 @@ __force_inline static uint get_core_num(void) {
     return (*(uint32_t *) (SIO_BASE + SIO_CPUID_OFFSET));
 }
 
+#ifdef __cplusplus
+}
+#endif
+
 #else // __ASSEMBLER__
 
 #define WRAPPER_FUNC_NAME(x) __wrap_##x
diff --git a/src/rp2_common/pico_printf/printf_none.S b/src/rp2_common/pico_printf/printf_none.S
index adc00ee5..212c1c50 100644
--- a/src/rp2_common/pico_printf/printf_none.S
+++ b/src/rp2_common/pico_printf/printf_none.S
@@ -7,9 +7,7 @@
 #include "pico/asm_helper.S"
 #include "pico/bootrom/sf_table.h"
 
-.syntax unified
-.cpu cortex-m0plus
-.thumb
+pico_default_asm_setup
 
 wrapper_func sprintf
 wrapper_func snprintf
diff --git a/src/rp2_common/pico_runtime/runtime.c b/src/rp2_common/pico_runtime/runtime.c
index b7076172..144ace1b 100644
--- a/src/rp2_common/pico_runtime/runtime.c
+++ b/src/rp2_common/pico_runtime/runtime.c
@@ -151,7 +151,7 @@ void runtime_init(void) {
 #ifndef NDEBUG
     if (__get_current_exception()) {
         // crap; started in exception handler
-        unified_asm ("bkpt #0");
+        __breakpoint();
     }
 #endif
 
@@ -286,7 +286,7 @@ extern void __attribute__((noreturn)) __printflike(1, 0) PICO_PANIC_FUNCTION(__u
 // Use a forwarding method here as it is a little simpler than renaming the symbol as it is used from assembler
 void __attribute__((naked, noreturn)) __printflike(1, 0) panic(__unused const char *fmt, ...) {
     // if you get an undefined reference here, you didn't define your PICO_PANIC_FUNCTION!
-    unified_asm (
+    pico_default_asm (
             "push {lr}\n"
 #if !PICO_PANIC_FUNCTION_EMPTY
             "bl " __XSTRING(PICO_PANIC_FUNCTION) "\n"
diff --git a/src/rp2_common/pico_standard_link/crt0.S b/src/rp2_common/pico_standard_link/crt0.S
index 43b069d0..1187a9ad 100644
--- a/src/rp2_common/pico_standard_link/crt0.S
+++ b/src/rp2_common/pico_standard_link/crt0.S
@@ -5,10 +5,11 @@
  */
 
 #include "pico.h"
+#include "pico/asm_helper.S"
+
 #include "hardware/regs/m0plus.h"
 #include "hardware/regs/addressmap.h"
 #include "hardware/regs/sio.h"
-#include "pico/asm_helper.S"
 #include "pico/binary_info/defs.h"
 
 #ifdef NDEBUG
@@ -17,9 +18,7 @@
 #endif
 #endif
 
-.syntax unified
-.cpu cortex-m0plus
-.thumb
+pico_default_asm_setup
 
 .section .vectors, "ax"
 .align 2
diff --git a/src/rp2_common/pico_stdio_semihosting/stdio_semihosting.c b/src/rp2_common/pico_stdio_semihosting/stdio_semihosting.c
index 291e9396..7a7910ff 100644
--- a/src/rp2_common/pico_stdio_semihosting/stdio_semihosting.c
+++ b/src/rp2_common/pico_stdio_semihosting/stdio_semihosting.c
@@ -25,7 +25,7 @@ static void stdio_semihosting_out_chars(const char *buf, int length) {
     args.buf = buf;
     args.len = length;
 
-    unified_asm (
+    pico_default_asm (
     // r1 must contain a pointer to the arguments
     "movs r1, %[args]\n"
     // semihosting call number 0x05 = SYS_WRITE
diff --git a/src/rp2_common/pico_stdlib/CMakeLists.txt b/src/rp2_common/pico_stdlib/CMakeLists.txt
index 98693458..183ca52b 100644
--- a/src/rp2_common/pico_stdlib/CMakeLists.txt
+++ b/src/rp2_common/pico_stdlib/CMakeLists.txt
@@ -1,9 +1,9 @@
 # PICO_CMAKE_CONFIG: PICO_STDIO_UART, OPTION: Globally enable stdio UART, default=1, group=pico_stdlib
-option(PICO_STDIO_UART "Globablly enable stdio UART" 1)
+option(PICO_STDIO_UART "Globally enable stdio UART" 1)
 # PICO_CMAKE_CONFIG: PICO_STDIO_USB, OPTION: Globally enable stdio USB, default=0, group=pico_stdlib
-option(PICO_STDIO_USB "Globablly enable stdio USB" 0)
+option(PICO_STDIO_USB "Globally enable stdio USB" 0)
 # PICO_CMAKE_CONFIG: PICO_STDIO_SEMIHOSTING, OPTION: Globally enable stdio semihosting, default=0, group=pico_stdlib
-option(PICO_STDIO_SEMIHOSTING "Globablly enable stdio semihosting" 0)
+option(PICO_STDIO_SEMIHOSTING "Globally enable stdio semi-hosting" 0)
 
 if (NOT TARGET pico_stdlib)
     pico_add_impl_library(pico_stdlib)
diff --git a/test/kitchen_sink/kitchen_sink.c b/test/kitchen_sink/kitchen_sink.c
index ab6118f1..3297dd05 100644
--- a/test/kitchen_sink/kitchen_sink.c
+++ b/test/kitchen_sink/kitchen_sink.c
@@ -138,6 +138,6 @@ int main(void) {
     hard_assert(recursive_mutex_try_enter(&recursive_mutex, NULL));
     hard_assert(recursive_mutex_try_enter(&recursive_mutex, NULL));
     // this should compile as we are Cortex M0+
-    unified_asm ("SVC #3");
+    pico_default_asm ("SVC #3");
 
 }
diff --git a/test/pico_divider_test/pico_divider_test.c b/test/pico_divider_test/pico_divider_test.c
index 97936f8d..6f0984ee 100644
--- a/test/pico_divider_test/pico_divider_test.c
+++ b/test/pico_divider_test/pico_divider_test.c
@@ -193,7 +193,7 @@ void test_random() {
 #endif
 
 uint32_t __attribute__((naked)) time_32(uint32_t a, uint32_t b, uint32_t (*func)(uint32_t a, uint32_t b)) {
-    unified_asm (
+    pico_default_asm (
         "push {r4, r5, lr}\n"
         "ldr r4, =0xe000e018\n"
         "ldr r5, [r4]\n"
@@ -207,7 +207,7 @@ uint32_t __attribute__((naked)) time_32(uint32_t a, uint32_t b, uint32_t (*func)
 }
 
 uint32_t __attribute__((naked)) time_64(uint64_t a, uint64_t b, uint64_t (*func64)(uint64_t a, uint64_t b)) {
-    unified_asm (
+    pico_default_asm (
     "push {r4-r6, lr}\n"
     "ldr r6, [sp, #16]\n"
     "ldr r4, =0xe000e018\n"
diff --git a/test/pico_float_test/llvm/call_apsr.S b/test/pico_float_test/llvm/call_apsr.S
index 6c7d3616..d24ccab4 100644
--- a/test/pico_float_test/llvm/call_apsr.S
+++ b/test/pico_float_test/llvm/call_apsr.S
@@ -12,9 +12,9 @@
 //
 //===-
 
-.syntax unified
-.cpu cortex-m0plus
-.thumb
+#include "pico/asm_helper.S"
+
+pico_default_asm_setup
 
 .align 2
 
diff --git a/test/pico_float_test/pico_float_test.c b/test/pico_float_test/pico_float_test.c
index 6cb034c3..4d45a9ab 100644
--- a/test/pico_float_test/pico_float_test.c
+++ b/test/pico_float_test/pico_float_test.c
@@ -518,7 +518,7 @@ int main() {
     }
     for(float x = 4294967296.f * 4294967296.f * 2.f; x>=0.5f; x/=2.f) {
         printf("f2i64 %f->%lld\n", x, (int64_t)x);
-        if (x >= INT64_MAX) {
+        if ((double)x >= (double)INT64_MAX) {
             // seems like there is a bug in the gcc version!
             assert(__aeabi_f2lz(x) == INT64_MAX);
         } else {