B2 changes

README.md

@@ -1,4 +1,4 @@
-This is the B1 version of the RP2040 bootrom.
+This is the B2 version of the RP2040 bootrom.
 
 The version on the chip was built in _Debug_ mode using GCC 9.3.1 (GNU Arm Embedded Toolchain 9-2020-q2-update).
 

bootrom/bootrom_main.c

@@ -16,6 +16,7 @@
 #include "hardware/sync.h"
 #include "hardware/resets.h"
 #include "usb_boot_device.h"
+#include "resets.h"
 
 #include "async_task.h"
 #include "bootrom_crc32.h"
@@ -152,8 +153,7 @@ static void _usb_clock_setup() {
     //
     // Total postdiv of 25 means that too-fast xtal will push VCO out of
     // lockable range *before* clk_sys goes out of closure (factor of 1.88)
-    reset_block(RESETS_RESET_PLL_SYS_BITS);
-    unreset_block_wait(RESETS_RESET_PLL_SYS_BITS);
+    reset_unreset_block_wait_noinline(RESETS_RESET_PLL_SYS_BITS);
     pll_sys_hw->cs = 1u << PLL_CS_REFDIV_LSB;
     pll_sys_hw->fbdiv_int = 100;
     pll_sys_hw->prim =
@@ -183,10 +183,10 @@ void __noinline __attribute__((noreturn)) async_task_worker_thunk();
 
 static __noinline __attribute__((noreturn)) void _usb_boot(uint32_t _usb_activity_gpio_pin_mask,
                                                                   uint32_t disable_interface_mask) {
-    reset_block(RESETS_RESET_USBCTRL_BITS);
+    reset_block_noinline(RESETS_RESET_USBCTRL_BITS);
     if (!running_on_fpga())
         _usb_clock_setup();
-    unreset_block_wait(RESETS_RESET_USBCTRL_BITS);
+    unreset_block_wait_noinline(RESETS_RESET_USBCTRL_BITS);
 
     // Ensure timer and watchdog are running at approximately correct speed
     // (can't switch clk_ref to xosc at this time, as we might lose ability to resus)
@@ -228,8 +228,7 @@ int main() {
             RESETS_RESET_IO_QSPI_BITS |
             RESETS_RESET_PADS_QSPI_BITS |
             RESETS_RESET_TIMER_BITS;
-    reset_block(rst_mask);
-    unreset_block(rst_mask);
+    reset_unreset_block_wait_noinline(rst_mask);
 
     // Workaround for behaviour of TXB0108 bidirectional level shifters on
     // FPGA platform (JIRA PRJMU-726), not used on ASIC

bootrom/bootrom_rt0.S

@@ -20,6 +20,8 @@
 // - Pass core 0 control over to the main flash boot sequence
 
 #include "hardware/regs/addressmap.h"
+#include "hardware/regs/pads_bank0.h"
+#include "hardware/regs/resets.h"
 #include "hardware/regs/sio.h"
 #include "hardware/regs/watchdog.h"
 #include "hardware/regs/syscfg.h"
@@ -47,7 +49,7 @@ _magic:
 # compatibility version (change if function table is incompatible, or functions are backwards incompatible)
 .byte 1
 # ROM version
-.byte 2
+.byte 3
 
 .global _well_known
 _well_known:
@@ -189,8 +191,6 @@ function_table:
 .hword flash_flush_cache + 1
 .byte 'C', 'X'
 .hword flash_enter_cmd_xip + 1
-.byte 'E', 'C'
-.hword enable_clocks + 1
 # end of function table marker
 .hword 0
 
@@ -247,15 +247,33 @@ check_core:
 // communicating with the processors).
 check_rescue:
     ldr r1, =(VREG_AND_CHIP_RESET_BASE + VREG_AND_CHIP_RESET_CHIP_RESET_OFFSET)
-    ldr r3, =(VREG_AND_CHIP_RESET_CHIP_RESET_PSM_RESTART_FLAG_BITS)
     ldr r2, [r1]
-    tst r2, r3
-    beq 1f
-// Acknowledge and halt
-    str r3, [r1]
+#if VREG_AND_CHIP_RESET_CHIP_RESET_PSM_RESTART_FLAG_BITS != 0x01000000
+#error
+#endif
+    lsr r3, r2, #25
+    bcc 1f
+// Acknowledge and halt (note we write all bits, but VREG_AND_CHIP_RESET_CHIP_RESET_PSM_RESTART_FLAG_BITS
+// is the only WC bit
+    str r2, [r1]
     b _dead
 1:
 
+disable_adc_ie:
+#if RESETS_RESET_PADS_BANK0_BITS != (0x80 << 1)
+#error
+#endif
+    mov r0, #0x80
+    lsl r0, #1
+    bl unreset_block_wait_noinline
+
+    ldr r1, =PADS_BANK0_BASE + REG_ALIAS_CLR_BITS
+    mov r2, #PADS_BANK0_GPIO0_IE_BITS
+    str r2, [r1, #PADS_BANK0_GPIO26_OFFSET]
+    str r2, [r1, #PADS_BANK0_GPIO27_OFFSET]
+    str r2, [r1, #PADS_BANK0_GPIO28_OFFSET]
+    str r2, [r1, #PADS_BANK0_GPIO29_OFFSET]
+
 // Check watchdog scratch for direct-boot magic numbers
 // This is useful in factory test for running ARM code without accessing DAP.
 // Probably also useful for imaginative software engineers
@@ -343,7 +361,7 @@ core_0_handshake_loop:
     adr r5, core1_launch
     // receive IP (0 sends us back into handshake loop)
     bl send_and_then
-    nop ;// .. for alignment
+//    nop // .. for alignment
 core1_launch:
     // Disable SLEEPDEEP before exiting, as it affects wake latency
     mov r1, #0
@@ -372,6 +390,35 @@ receive_and_check_zero:
     beq core_0_handshake_loop
     bx  lr
 
+.global reset_block_noinline
+.type reset_block_noinline,%function
+.thumb_func
+reset_block_noinline:
+    ldr r1, =RESETS_BASE + REG_ALIAS_SET_BITS
+    str r0, [r1]
+    bx lr
+
+.global reset_unreset_block_wait_noinline
+.type reset_unreset_block_wait_noinline,%function
+.thumb_func
+reset_unreset_block_wait_noinline:
+    ldr r1, =RESETS_BASE + REG_ALIAS_SET_BITS
+    str r0, [r1]
+    // fall thru
+
+.global unreset_block_wait_noinline
+.type unreset_block_wait_noinline,%function
+.thumb_func
+unreset_block_wait_noinline:
+    ldr r1, =RESETS_BASE
+    ldr r2, =RESETS_BASE + REG_ALIAS_CLR_BITS
+    str r0, [r2]
+1:
+    ldr r2, [r1, #RESETS_RESET_DONE_OFFSET]
+    bic r0, r2
+    bne 1b
+    bx lr
+
 // ----------------------------------------------------------------------------
 // Simple debugger trampoline for break-on-return
 // ----------------------------------------------------------------------------
@@ -419,7 +466,7 @@ soft_float_table:
     .word mufp_fcos
     .word mufp_fsin
     .word mufp_ftan
-    .word _dead // todo use this for storage?
+    .word mufp_fsincos
     .word mufp_fexp
     .word mufp_fln
 
@@ -455,7 +502,7 @@ soft_double_table:
     .word mufp_dcos
     .word mufp_dsin
     .word mufp_dtan
-    .word _dead
+    .word mufp_dsincos
     .word mufp_dexp
     .word mufp_dln
 
@@ -485,6 +532,9 @@ async_task_worker_thunk:
     ldr r0, =usb_boot_stack_end
     msr MSP, r0
     bl async_task_worker
+    // async_task_worker does not return
+
+_:
 
 .section .bss
 .align 2

bootrom/info_uf2.txt

@@ -1,3 +1,3 @@
-UF2 Bootloader v2.0
+UF2 Bootloader v3.0
 Model: Raspberry Pi RP2
 Board-ID: RPI-RP2

bootrom/mufplib-double.S

@@ -29,6 +29,7 @@
 .global mufp_dsqrt
 .global mufp_dcos
 .global mufp_dsin
+.global mufp_dsincos
 .global mufp_dtan
 .global mufp_datan2
 .global mufp_dexp
@@ -1791,6 +1792,18 @@ mufp_dsin:
  bl pop_r8_r11
  pop {r4-r7,r15}
 
+.thumb_func
+mufp_dsincos:
+     push {r4-r7,r14}
+     bl push_r8_r11
+     bl dsincos
+     mov r12,r0                    @ save ε
+     bl dcos_finish
+     push {r0,r1}
+     mov r0,r12
+     bl dsin_finish
+     pop {r2,r3}
+     b 1b
 
 @ unpack double θ in r0:r1, range reduce and calculate ε, cos α and sin α such that
 @ θ=α+ε and |ε|≤2^-32

bootrom/mufplib.S

@@ -44,6 +44,7 @@
 .global mufp_ufix642float
 .global mufp_fcos
 .global mufp_fsin
+.global mufp_fsincos
 .global mufp_ftan
 .global mufp_fatan2
 .global mufp_fexp
@@ -481,7 +482,9 @@ cordic_vec:
  b cordic_exit
 
 .thumb_func
-mufp_fsin:            @ calculate sin and cos using CORDIC rotation method
+mufp_fsin:
+.thumb_func
+mufp_fsincos:         @ calculate sin and cos using CORDIC rotation method
  push {r4,r5,r14}
  movs r1,#24
  bl mufp_float2fix    @ range reduction by repeated subtraction/addition in fixed point

bootrom/program_flash_generic.c

@@ -10,6 +10,7 @@
 #include "hardware/structs/xip_ctrl.h"
 #include "hardware/resets.h"
 #include "program_flash_generic.h"
+#include "resets.h"
 
 // These are supported by almost any SPI flash
 #define FLASHCMD_PAGE_PROGRAM     0x02
@@ -38,8 +39,7 @@ check_hw_layout(ssi_hw_t, spi_ctrlr0, SSI_SPI_CTRLR0_OFFSET);
 void __noinline connect_internal_flash() {
     // Use hard reset to force IO and pad controls to known state (don't touch
     // IO_BANK0 as that does not affect XIP signals)
-    reset_block(RESETS_RESET_IO_QSPI_BITS | RESETS_RESET_PADS_QSPI_BITS);
-    unreset_block_wait(RESETS_RESET_IO_QSPI_BITS | RESETS_RESET_PADS_QSPI_BITS);
+    reset_unreset_block_wait_noinline(RESETS_RESET_IO_QSPI_BITS | RESETS_RESET_PADS_QSPI_BITS);
 
     // Then mux XIP block onto internal QSPI flash pads
     io_rw_32 *iobank1 = (io_rw_32 *) IO_QSPI_BASE;

bootrom/resets.h

@@ -0,0 +1,16 @@
+/**
+ * Copyright (c) 2020 Raspberry Pi (Trading) Ltd.
+ *
+ * SPDX-License-Identifier: BSD-3-Clause
+ */
+
+#ifndef _RESETS_H
+#define _RESETS_H
+
+#include "pico/types.h"
+
+void reset_block_noinline(uint32_t mask);
+void unreset_block_wait_noinline(uint32_t mask);
+void reset_unreset_block_wait_noinline(uint32_t mask);
+
+#endif

bootrom/usb_boot_device.c

@@ -20,6 +20,7 @@ char serial_number_string[13];
 // todo make descriptor strings should probably belong/come from the configs
 static char *descriptor_strings[] =
         {
+                "",
                 "Raspberry Pi",
                 "RP2 Boot",
                 serial_number_string
@@ -164,11 +165,8 @@ uint32_t msc_get_serial_number32() {
 }
 
 const char *_get_descriptor_string(uint index) {
-    if (index <= count_of(descriptor_strings)) {
-        return descriptor_strings[index - 1];
-    } else {
-        return "";
-    }
+    if (index >= count_of(descriptor_strings)) index = 0;
+    return descriptor_strings[index];
 }
 
 

test/tc_rom_double.c

@@ -47,8 +47,7 @@ struct mufp_funcs {
     double (*mufp_dcos)(double);
     double (*mufp_dsin)(double);
     double (*mufp_dtan)(double);
-    // since float doesn't use this slot, we don't either
-    uint32_t _not_fatan2; // double (*mufp_datan2)(double, double);
+    double (*v3_mufp_dsincos)(double);
     double (*mufp_dexp)(double);
     double (*mufp_dln)(double);
 
@@ -235,7 +234,17 @@ int __attribute__((naked)) dcmp_from_dcmp_flags(double a, double b, int (*dmcp_f
     );
 }
 
-#define check_dcmp_flags(a,b) check_int(dcmp(a, b), dcmp_from_dcmp_flags(a, b, mufp_funcs->mufp_dcmp)) // dcmp is dcmp_flags now
+double __attribute__((naked)) call_dsincos(double v, double *cout, double (*func)(double)) {
+    asm(
+    "push {r2, r4, lr}\n"
+    "blx r3\n"
+    "pop {r4}\n"
+    "stmia r4!, {r2, r3}\n"
+    "pop {r4, pc}\n"
+    );
+}
+
+#define check_dcmp_flags(a,b) check_int(dcmp(a, b), dcmp_from_dcmp_flags(a, b, (void (*)(double,double))mufp_funcs->mufp_dcmp)) // dcmp is dcmp_flags now
 #define check_dcmp_fast_flags(a,b) check_int(dcmp_fast(a, b), dcmp_from_dcmp_flags(a, b, mufp_funcs->mufp_dcmp_fast_flags))
 
 int main()
@@ -522,6 +531,15 @@ int main()
     check_double_fn1(double2float, 3.0e68);
     check_double_fn1(double2float, -3.0e68);
 
+    if (rom_version >= 3) {
+        for(double a = -0.3f; a<7.f; a += 0.137) {
+            double s = mufp_funcs->mufp_dsin(a);
+            double c = mufp_funcs->mufp_dcos(a);
+            double co=0;
+            double so = call_dsincos(a, &co, mufp_funcs->v3_mufp_dsincos);
+            ASSERT(so == s && co == c);
+        }
+    }
     printf("DOUBLE OK\n");
-	return 0;
+    return 0;
 }

test/tc_rom_float.c

@@ -45,7 +45,7 @@ struct mufp_funcs {
     float (*mufp_fcos)(float);
     float (*mufp_fsin)(float);
     float (*mufp_ftan)(float);
-    uint32_t _broken_fatan2; //    float (*mufp_fatan2)(float, float);
+    float (*v3_mufp_fsincos)(float);
     float (*mufp_fexp)(float);
     float (*mufp_fln)(float);
 
@@ -232,7 +232,17 @@ int __attribute__((naked)) fcmp_from_fcmp_flags(float a, float b, int (*fmcp_fla
     );
 }
 
-#define check_fcmp_flags(a,b) check_int(fcmp(a, b), fcmp_from_fcmp_flags(a, b, mufp_funcs->mufp_fcmp)) // f_cmp is f_cmp_flags now
+float __attribute__((naked)) call_fsincos(float v, float *cout, float (*func)(float)) {
+    asm(
+    "push {r1, lr}\n"
+    "blx r2\n"
+    "pop {r2}\n"
+    "str r1, [r2]\n"
+    "pop {pc}\n"
+    );
+}
+
+#define check_fcmp_flags(a,b) check_int(fcmp(a, b), fcmp_from_fcmp_flags(a, b, (void (*)(float, float))mufp_funcs->mufp_fcmp)) // f_cmp is f_cmp_flags now
 #define check_fcmp_fast_flags(a,b) check_int(fcmp_fast(a, b), fcmp_from_fcmp_flags(a, b, mufp_funcs->mufp_fcmp_fast_flags))
 
 int main()
@@ -501,6 +511,15 @@ int main()
         check_float_fn1(float2double, -0.f);
     }
 
+    if (rom_version >= 3) {
+        for(float a = -0.3f; a<7.f; a += 0.137) {
+            float s = mufp_funcs->mufp_fsin(a);
+            float c = mufp_funcs->mufp_fcos(a);
+            float co=0;
+            float so = call_fsincos(a, &co, mufp_funcs->v3_mufp_fsincos);
+            ASSERT(so == s && co == c);
+        }
+    }
     printf("FLOAT OK\n");
 	return 0;
 }