ARM Compiler 6.13 testing revealed linker errors pointing out
conflicting use of `__user_setup_stackheap` and
`__user_initial_stackheap` in some targets. Remove the unwanted
`__user_initial_stackheap` from the targets - the setup is
centralised in the common platform code.
Looking into this, a number of other issues were highlighted
* Almost all targets had `__initial_sp` hardcoded in assembler,
rather than getting it from the scatter file. This was behind
issue #11313. Fix this generally.
* A few targets' `__initial_sp` values did not match the scatter
file layout, in some cases meaning they were overlapping heap
space. They now all use the area reserved in the scatter file.
If any problems are seen, then there is an error in the
scatter file.
* A number of targets were reserving unneeded space for heap and
stack in their startup assembler, on top of the space reserved in
the scatter file, so wasting a few K. A couple were using that
space for the stack, rather than the space in the scatter file.
To clarify expected behaviour:
* Each scatter file contains empty regions `ARM_LIB_HEAP` and
`ARM_LIB_STACK` to reserve space. `ARM_LIB_STACK` is sized
by the macro `MBED_BOOT_STACK_SIZE`, which is set by the tools.
`ARM_LIB_HEAP` is generally the space left over after static
RAM and stack.
* The address of the end of `ARM_LIB_STACK` is written into the
vector table and on reset the CPU sets MSP to that address.
* The common platform code in Mbed OS provides `__user_setup_stackheap`
for the ARM library. The ARM library calls this during startup, and
it calls `__mbed_user_setup_stackheap`.
* The default weak definition of `__mbed_user_setup_stackheap` does not
modify SP, so we remain on the boot stack, and the heap is set to
the region described by `ARM_LIB_HEAP`. If `ARM_LIB_HEAP` doesn't
exist, then the heap is the space from the end of the used data in
`RW_IRAM1` to the start of `ARM_LIB_STACK`.
* Targets can override `__mbed_user_setup_stackheap` if they want.
Currently only Renesas (ARMv7-A class) devices do.
* If microlib is in use, then it doesn't call `__user_setup_stackheap`.
Instead it just finds and uses `ARM_LIB_STACK` and `ARM_LIB_HEAP`
itself.
Instead of user defined symbols in assembly files or C files,
use linker scripts to add heap and stack - this is inconsistent
with ARM std linker scripts
HEAP memory should be 4K aligned for GCC newlib, with ISR stack at the end of
RAM memory we loose 3K of RAM memory. This fix is for device with <16K RAM to
use RAM entirely.
--legacyalign, --no_legacyalign are deprecated from ARMC6 compiler, in order to
remove deprecated flags all linker files should strictly align to 8-byte boundary
Set the ISR stack to be 1KB. https://github.com/ARMmbed/mbed-os/pull/7238
Set the heap size to 3KB(2KB + overhead + spare) so that atleast 2KB free ram is
available for testing.
With dynamic heap size, explicit size is not required. IAR 7.8 supports
static heap, hence the change is needed in IAR linker files.
Decrease the interrupt stack from 2k down to 1k so there is enough
ram to build all the tests with tickless enabled. In general, targets
should not need an interrupt stack greater than 1k with mbed-os.
- default value is the same as before patch
- system_stm32f0xx.c file is copied to family level with all other ST cube files
- specific clock configuration is now in a new file: system_clock.c (target level)
Remove HAL_Init and related code from SystemInit and move it to
mbed_sdk_init. The function SystemInit is called early in the boot
sequence before RAM is initialized or the VTOR is setup, so it should
not be used to perform the HAL initialization.
This fixes crashes due the vector table being used before it has been
relocated.
Restore cmsis_nvic (cmsis_nvic.c and cmsis_nvic.h) files for the
implementations which use a mechanism other than the VTOR to set
interrupts. These are vendor specific and were done for M0 devices
which do not have a VTOR.
Note - There were two cmsis_nvic files which did not use the VTOR that
which not restored in this patch. This is because these targets were
not M0 devices and could use the new unified implementation instead.
These files are:
targets\TARGET_ARM_SSG\TARGET_MPS2\TARGET_MPS2_M0P\device\cmsis_nvic.c
targets\TARGET_ONSEMI\TARGET_NCS36510\device\cmsis_nvic.c
Note - cmsis_nvic.c and cmsis_nvic.h were initial removed in
(and restored from) the commit:
b97ffe8fdc -
"CMSIS5: Replace target defined NVIC_Set/GetVector with CMSIS implementation"
For STM32 targets using a 32-bit timer for the microsecond ticker, the
driver did not properly handle timestamps that are in the past. It
would just blindly set the compare register to the requested timestamp,
resulting in the interrupt being serviced up to 4295 seconds late
(i.e. after the 32-bit timer counts all the way around to hit the
timestamp again).
This problem can easily be reproduced by creating a Timeout object
then calling the timeout's attach_us() member function to attach a
callback with a timeout of 0 us. The callback will not get called for
over 2147 seconds, and possibly up to 4295 seconds late if no other
microsecond ticker events are getting scheduled in the meantime.
Now, after the compare register has been set, the timestamp is checked
against the current time to see if the timestamp is in the past, and
if so, the compare event is manually set.
NOTE: By checking if the timestamp is in the past after configuring the
capture register, we ensure proper handling in the case where the timer
updates past the timestamp while setting the capture register.
Kevin Bracey
jeromecoutant
deepikabhavnani
Przemyslaw Stekiel
Anna Bridge
bcostm
Jimmy Brisson
Russ Butler
Sam Grove
Bradley Scott
Bartek Szatkowski
Laurent MEUNIER
Michel Jaouen
Christopher Haster