Redo RP2040 flash settings

This switches stage2 to C and uses Jinja to change the C code based
on flash settings from https://github.com/adafruit/nvm.toml. It
produces the fastest settings for the given set of external flashes.
Flash size is no longer hard coded so switching flashes with similar
capabilities but different sizes should *just work*.

This PR also places "ITCM" code in RAM to save the XIP cache for
code execution. Further optimization is possible. A blink code.py
still requires a number of flash fetches every blink.

Fixes adafruit#4041

.gitmodules

@@ -173,4 +173,8 @@
 	url = https://github.com/adafruit/Adafruit_CircuitPython_LC709203F
 [submodule "ports/raspberrypi/sdk"]
 	path = ports/raspberrypi/sdk
-	url = https://github.com/raspberrypi/pico-sdk.git
+	url = https://github.com/adafruit/pico-sdk.git
+[submodule "data/nvm.toml"]
+	path = data/nvm.toml
+	url = https://github.com/adafruit/nvm.toml.git
+	branch = main

ports/raspberrypi/Makefile

@@ -106,7 +106,7 @@ INC += -I. \
        -I$(BUILD)
 
 # Pico specific configuration
-CFLAGS += -DPICO_ON_DEVICE=1 -DPICO_NO_BINARY_INFO=0 -DPICO_TIME_DEFAULT_ALARM_POOL_DISABLED=1 -DPICO_DIVIDER_CALL_IDIV0=0 -DPICO_DIVIDER_CALL_LDIV0=0 -DPICO_DIVIDER_HARDWARE=1 -DPICO_DOUBLE_ROM=1 -DPICO_FLOAT_ROM=1 -DPICO_MULTICORE=1 -DPICO_BITS_IN_RAM=0 -DPICO_DIVIDER_IN_RAM=0 -DPICO_DOUBLE_PROPAGATE_NANS=0 -DPICO_DOUBLE_IN_RAM=0 -DPICO_MEM_IN_RAM=0 -DPICO_FLOAT_IN_RAM=0 -DPICO_FLOAT_PROPAGATE_NANS=1 -DPICO_NO_FLASH=0 -DPICO_COPY_TO_RAM=0 -DPICO_DISABLE_SHARED_IRQ_HANDLERS=0 -DPICO_NO_BI_BOOTSEL_VIA_DOUBLE_RESET=0
+CFLAGS += -DRASPBERRYPI -DPICO_ON_DEVICE=1 -DPICO_NO_BINARY_INFO=0 -DPICO_TIME_DEFAULT_ALARM_POOL_DISABLED=1 -DPICO_DIVIDER_CALL_IDIV0=0 -DPICO_DIVIDER_CALL_LDIV0=0 -DPICO_DIVIDER_HARDWARE=1 -DPICO_DOUBLE_ROM=1 -DPICO_FLOAT_ROM=1 -DPICO_MULTICORE=1 -DPICO_BITS_IN_RAM=0 -DPICO_DIVIDER_IN_RAM=0 -DPICO_DOUBLE_PROPAGATE_NANS=0 -DPICO_DOUBLE_IN_RAM=0 -DPICO_MEM_IN_RAM=0 -DPICO_FLOAT_IN_RAM=0 -DPICO_FLOAT_PROPAGATE_NANS=1 -DPICO_NO_FLASH=0 -DPICO_COPY_TO_RAM=0 -DPICO_DISABLE_SHARED_IRQ_HANDLERS=0 -DPICO_NO_BI_BOOTSEL_VIA_DOUBLE_RESET=0
 OPTIMIZATION_FLAGS ?= -O3
 # TinyUSB defines
 CFLAGS += -DTUD_OPT_RP2040_USB_DEVICE_ENUMERATION_FIX=1 -DCFG_TUSB_MCU=OPT_MCU_RP2040 -DCFG_TUD_MIDI_RX_BUFSIZE=128 -DCFG_TUD_CDC_RX_BUFSIZE=256 -DCFG_TUD_MIDI_TX_BUFSIZE=128 -DCFG_TUD_CDC_TX_BUFSIZE=256 -DCFG_TUD_MSC_BUFSIZE=1024
@@ -169,7 +169,7 @@ SRC_SDK := \
 	src/rp2_common/hardware_irq/irq.c \
 	src/rp2_common/hardware_pio/pio.c \
 	src/rp2_common/hardware_pll/pll.c \
-  src/rp2_common/hardware_rtc/rtc.c \
+	src/rp2_common/hardware_rtc/rtc.c \
 	src/rp2_common/hardware_spi/spi.c \
 	src/rp2_common/hardware_sync/sync.c \
 	src/rp2_common/hardware_timer/timer.c \
@@ -270,9 +270,15 @@ $(BUILD)/boot2.bin: $(BUILD)/boot2.elf
 	$(STEPECHO) "OBJCOPY $<"
 	$(Q)$(OBJCOPY) -O binary $< $@
 
-$(BUILD)/boot2.elf: $(BOOT2_S_UPPER) | $(BUILD)/
+
+$(BUILD)/stage2.c: stage2.c.jinja gen_stage2.py | $(BUILD)/
+	$(STEPECHO) "GEN $<"
+	$(Q)$(PYTHON3) gen_stage2.py $< $@ $(EXTERNAL_FLASH_DEVICES)
+
+$(BUILD)/boot2.elf: $(BUILD)/stage2.c
 	$(STEPECHO) "BOOT $<"
-	$(Q)$(CC) $(CFLAGS) $(BOOT2_S_CFLAGS) -Isdk/src/rp2_common/boot_stage2/asminclude --specs=nosys.specs -nostartfiles -Wl,-T,sdk/src/rp2_common/boot_stage2/boot_stage2.ld -o $@ $<
+	$(Q)$(CC) $(CFLAGS) $(BOOT2_S_CFLAGS) -Os -ggdb3 -I. -fPIC --specs=nosys.specs -nostartfiles -Wl,-T,boot_stage2.ld  -Wl,-Map=$@.map -o $@ $<
+	$(Q)$(SIZE) $@
 
 SRC_QSTR += $(SRC_C) $(SRC_SUPERVISOR) $(SRC_COMMON_HAL_EXPANDED) $(SRC_SHARED_MODULE_EXPANDED)
 
@@ -285,7 +291,7 @@ $(BUILD)/firmware.elf: $(OBJ) link.ld
 
 $(BUILD)/firmware.bin: $(BUILD)/firmware.elf
 	$(STEPECHO) "Create $@"
-	$(Q)$(OBJCOPY) -O binary $^ $@
+	$(Q)$(OBJCOPY) -O binary -R .dtcm_bss $^ $@
 
 $(BUILD)/firmware.uf2: $(BUILD)/firmware.bin
 	$(STEPECHO) "Create $@"

ports/raspberrypi/boards/adafruit_feather_rp2040/mpconfigboard.h

@@ -12,6 +12,3 @@
 
 #define DEFAULT_UART_BUS_RX (&pin_GPIO1)
 #define DEFAULT_UART_BUS_TX (&pin_GPIO0)
-
-// Flash chip is GD25Q64 connected over QSPI
-#define TOTAL_FLASH_SIZE (8 * 1024 * 1024)

ports/raspberrypi/boards/adafruit_feather_rp2040/mpconfigboard.mk

@@ -6,4 +6,4 @@ USB_MANUFACTURER = "Adafruit"
 CHIP_VARIANT = RP2040
 CHIP_FAMILY = rp2
 
-INTERNAL_FLASH_FILESYSTEM = 1
+EXTERNAL_FLASH_DEVICES = "GD25Q64C"

ports/raspberrypi/boards/pimoroni_keybow2040/mpconfigboard.h

@@ -28,8 +28,6 @@
 #define DEFAULT_UART_BUS_RX (&pin_GPIO1)
 #define DEFAULT_UART_BUS_TX (&pin_GPIO0)
 
-#define TOTAL_FLASH_SIZE (2 * 1024 * 1024)
-
 // These pins are unconnected
 #define IGNORE_PIN_GPIO2    1
 #define IGNORE_PIN_GPIO22   1

ports/raspberrypi/boards/pimoroni_keybow2040/mpconfigboard.mk

@@ -6,6 +6,6 @@ USB_MANUFACTURER = "Pimoroni"
 CHIP_VARIANT = RP2040
 CHIP_FAMILY = rp2
 
-INTERNAL_FLASH_FILESYSTEM = 1
+EXTERNAL_FLASH_DEVICES = "W25Q16JV_IQ"
 
 CIRCUITPY__EVE = 1

ports/raspberrypi/boards/pimoroni_picosystem/mpconfigboard.h

@@ -38,8 +38,6 @@
 #define DEFAULT_UART_BUS_RX (&pin_GPIO1)
 #define DEFAULT_UART_BUS_TX (&pin_GPIO0)
 
-#define TOTAL_FLASH_SIZE (16 * 1024 * 1024)
-
 // These pins are unconnected
 #define IGNORE_PIN_GPIO3    1
 #define IGNORE_PIN_GPIO10   1

ports/raspberrypi/boards/pimoroni_picosystem/mpconfigboard.mk

@@ -6,6 +6,6 @@ USB_MANUFACTURER = "Pimoroni"
 CHIP_VARIANT = RP2040
 CHIP_FAMILY = rp2
 
-INTERNAL_FLASH_FILESYSTEM = 1
+EXTERNAL_FLASH_DEVICES = "W25Q128JV_IQ"
 
 CIRCUITPY__EVE = 1

ports/raspberrypi/boards/pimoroni_tiny2040/mpconfigboard.h

@@ -7,8 +7,6 @@
 
 #define MICROPY_HW_USER_SW (&pin_GPIO23)
 
-#define TOTAL_FLASH_SIZE (8 * 1024 * 1024)
-
 // These pins are unconnected
 #define IGNORE_PIN_GPIO8    1
 #define IGNORE_PIN_GPIO9    1

ports/raspberrypi/boards/pimoroni_tiny2040/mpconfigboard.mk

@@ -6,6 +6,6 @@ USB_MANUFACTURER = "Pimoroni"
 CHIP_VARIANT = RP2040
 CHIP_FAMILY = rp2
 
-INTERNAL_FLASH_FILESYSTEM = 1
+EXTERNAL_FLASH_DEVICES = "W25Q64JV_IQ"
 
 CIRCUITPY__EVE = 1

ports/raspberrypi/boards/qtpy_rp2040/mpconfigboard.h

@@ -12,6 +12,3 @@
 
 // #define DEFAULT_UART_BUS_RX (&pin_PA11)
 // #define DEFAULT_UART_BUS_TX (&pin_PA10)
-
-// Flash chip is GD25Q64 connected over QSPI
-#define TOTAL_FLASH_SIZE (8 * 1024 * 1024)

ports/raspberrypi/boards/qtpy_rp2040/mpconfigboard.mk

@@ -6,4 +6,4 @@ USB_MANUFACTURER = "Adafruit"
 CHIP_VARIANT = RP2040
 CHIP_FAMILY = rp2
 
-INTERNAL_FLASH_FILESYSTEM = 1
+EXTERNAL_FLASH_DEVICES = "W25Q32JV_IQ"

ports/raspberrypi/boards/raspberry_pi_pico/mpconfigboard.h

@@ -1,18 +1,2 @@
-// LEDs
-// #define MICROPY_HW_LED_STATUS   (&pin_PA17)
-
 #define MICROPY_HW_BOARD_NAME "Raspberry Pi Pico"
 #define MICROPY_HW_MCU_NAME "rp2040"
-
-// #define DEFAULT_I2C_BUS_SCL (&pin_PA23)
-// #define DEFAULT_I2C_BUS_SDA (&pin_PA22)
-
-// #define DEFAULT_SPI_BUS_SCK (&pin_PB11)
-// #define DEFAULT_SPI_BUS_MOSI (&pin_PB10)
-// #define DEFAULT_SPI_BUS_MISO (&pin_PA12)
-
-// #define DEFAULT_UART_BUS_RX (&pin_PA11)
-// #define DEFAULT_UART_BUS_TX (&pin_PA10)
-
-// Flash chip is W25Q16JVUXIQ connected over QSPI
-#define TOTAL_FLASH_SIZE (2 * 1024 * 1024)

ports/raspberrypi/boards/raspberry_pi_pico/mpconfigboard.mk

@@ -6,8 +6,6 @@ USB_MANUFACTURER = "Raspberry Pi"
 CHIP_VARIANT = RP2040
 CHIP_FAMILY = rp2
 
-INTERNAL_FLASH_FILESYSTEM = 1
+EXTERNAL_FLASH_DEVICES = "W25Q16JV_IQ"
 
 CIRCUITPY__EVE = 1
-
-BOOT2_S_UPPER = sdk/src/rp2_common/boot_stage2/boot2_w25q080.S

ports/raspberrypi/boards/sparkfun_pro_micro_rp2040/mpconfigboard.h

@@ -12,6 +12,3 @@
 
 #define DEFAULT_UART_BUS_RX (&pin_GPIO1)
 #define DEFAULT_UART_BUS_TX (&pin_GPIO0)
-
-// Flash chip is GD25Q128 connected over QSPI
-#define TOTAL_FLASH_SIZE (16 * 1024 * 1024)

ports/raspberrypi/boards/sparkfun_pro_micro_rp2040/mpconfigboard.mk

@@ -6,4 +6,4 @@ USB_MANUFACTURER = "SparkFun"
 CHIP_VARIANT = RP2040
 CHIP_FAMILY = rp2
 
-INTERNAL_FLASH_FILESYSTEM = 1
+EXTERNAL_FLASH_DEVICES = "W25Q128JV_IM"

ports/raspberrypi/boards/sparkfun_thing_plus_rp2040/mpconfigboard.h

@@ -12,6 +12,3 @@
 
 #define DEFAULT_UART_BUS_RX (&pin_GPIO1)
 #define DEFAULT_UART_BUS_TX (&pin_GPIO0)
-
-// Flash chip is GD25Q128 connected over QSPI
-#define TOTAL_FLASH_SIZE (16 * 1024 * 1024)

ports/raspberrypi/boards/sparkfun_thing_plus_rp2040/mpconfigboard.mk

@@ -6,4 +6,4 @@ USB_MANUFACTURER = "SparkFun"
 CHIP_VARIANT = RP2040
 CHIP_FAMILY = rp2
 
-INTERNAL_FLASH_FILESYSTEM = 1
+EXTERNAL_FLASH_DEVICES = "W25Q128JV_IM"

ports/raspberrypi/boot_stage2.ld

@@ -0,0 +1,13 @@
+MEMORY {
+    /* We are loaded to the top 256 bytes of SRAM, which is above the bootrom
+       stack. Note 4 bytes occupied by checksum. */
+    SRAM(rx) : ORIGIN = 0x20041f00, LENGTH = 252
+}
+
+SECTIONS {
+    . = ORIGIN(SRAM);
+    .text : {
+        *(.entry.*)
+        *(.text.*)
+    } >SRAM
+}

ports/raspberrypi/gen_stage2.py

@@ -0,0 +1,82 @@
+import sys
+import cascadetoml
+import pathlib
+import typer
+from jinja2 import Template
+
+
+def main(input_template: pathlib.Path, output_path: pathlib.Path, skus: str = typer.Argument("")):
+    if ", " in skus:
+        skus = skus.split(", ")
+    else:
+        skus = [skus]
+    skus = ['sku="{}"'.format(f) for f in skus]
+    flashes = cascadetoml.filter_toml(pathlib.Path("../../data/nvm.toml"), skus)
+
+    if len(skus) == 0:
+        print("Set EXTERNAL_FLASH_DEVICES in mpconfigboard.mk with all possible flash skus")
+        raise typer.Exit(code=1)
+
+    def all_have(nvms, key):
+        for nvm in nvms:
+            if not nvm.get(key, False):
+                return False
+        return True
+
+    def all_match(nvms, key, default=None):
+        shared_value = nvms[0].get(key, default)
+        for nvm in nvms:
+            this_value = nvm.get(key, default)
+            if this_value != shared_value:
+                print(
+                    "{}.{} = {} does not match {}".format(
+                        nvm["sku"], key, this_value, shared_value
+                    )
+                )
+                return None
+        return shared_value
+
+    quad_enable_status_byte = all_match(flashes["nvm"], "quad_enable_status_byte", 0)
+    quad_enable_bit_mask = all_match(flashes["nvm"], "quad_enable_bit_mask")
+    continuous_status_write = all_have(flashes["nvm"], "01_continuous_status_write")
+    split_status_write = all_have(flashes["nvm"], "write_status_register_split")
+    e7_quad_word_read = all_have(flashes["nvm"], "e7_quad_word_read")
+
+    quad_ok = quad_enable_status_byte is not None and quad_enable_bit_mask is not None
+
+    max_clock_speed_mhz = min((x.get("max_clock_speed_mhz", 1000) for x in flashes["nvm"]))
+
+    # Check that we have a consistent way to set quad enable.
+    if continuous_status_write is None and split_status_write is None:
+        print("quad not ok", continuous_status_write, split_status_write)
+        quad_ok = False
+
+    clock_divider = 4
+
+    read_command = 0x03
+    wait_cycles = 0
+    if quad_ok:
+        if e7_quad_word_read:
+            read_command = 0xE7
+            wait_cycles = 2
+        else:
+            read_command = 0xEB
+            wait_cycles = 4
+
+    flash_settings = {
+        "quad_ok": quad_ok,
+        "quad_enable_status_byte": quad_enable_status_byte,
+        "quad_enable_bit_mask": quad_enable_bit_mask,
+        "split_status_write": split_status_write,
+        "clock_divider": clock_divider,
+        "read_command": read_command,
+        "wait_cycles": wait_cycles,
+    }
+
+    template = Template(input_template.read_text())
+
+    output_path.write_text(template.render(flash_settings))
+
+
+if __name__ == "__main__":
+    typer.run(main)

ports/raspberrypi/link.ld

@@ -174,6 +174,40 @@ SECTIONS
         __data_end__ = .;
     } > RAM AT> FLASH_FIRMWARE
 
+    .itcm :
+    {
+        . = ALIGN(4);
+        *(.itcm.*)
+
+        . = ALIGN(4);
+    } > RAM AT> FLASH_FIRMWARE
+    _ld_itcm_destination = ADDR(.itcm);
+    _ld_itcm_flash_copy = LOADADDR(.itcm);
+    _ld_itcm_size = SIZEOF(.itcm);
+
+    .dtcm_data :
+    {
+        . = ALIGN(4);
+
+        *(.dtcm_data.*)
+
+        . = ALIGN(4);
+    } > RAM AT> FLASH_FIRMWARE
+    _ld_dtcm_data_destination = ADDR(.dtcm_data);
+    _ld_dtcm_data_flash_copy = LOADADDR(.dtcm_data);
+    _ld_dtcm_data_size = SIZEOF(.dtcm_data);
+
+    .dtcm_bss :
+    {
+        . = ALIGN(4);
+
+        *(.dtcm_bss.*)
+
+        . = ALIGN(4);
+    } > RAM AT> RAM
+    _ld_dtcm_bss_start = ADDR(.dtcm_bss);
+    _ld_dtcm_bss_size = SIZEOF(.dtcm_bss);
+
     .uninitialized_data (COPY): {
         . = ALIGN(4);
         *(.uninitialized_data*)

ports/raspberrypi/mpconfigport.mk

@@ -55,3 +55,5 @@ USB_SERIAL_NUMBER_LENGTH = 16
 
 # Number of USB endpoint pairs.
 USB_NUM_EP = 8
+
+INTERNAL_FLASH_FILESYSTEM = 1