Move the PWM, ADC, and SPI pin tables closer to their corresponding
code. This is code movement only - no code changes.
Signed-off-by: Kevin O'Connor <kevin@koconnor.net>
Extend the endstop code so that more than one stepper can be halted
during endstop homing. Some kinematic setups (eg, corexy) require an
endstop to support this.
Signed-off-by: Kevin O'Connor <kevin@koconnor.net>
The relax option corrupts the compilation on at least some versions of
gcc/binutils (eg, on fedora's avr-gcc 6.2.0 / binutils 2.27) due to
corruption of switch tables that use jump offsets. This issue is also
the root cause that resulted in commit d67f962a. Since the --relax
option provides minimal size / performance improvements it can simply
be dropped.
Signed-off-by: Kevin O'Connor <kevin@koconnor.net>
Commmit a1c61563 renamed timer_shutdown() to timer_reset() but
neglected to delete timer_shutdown().
Signed-off-by: Kevin O'Connor <kevin@koconnor.net>
Commit 16e3dbb1 changed the avr implementation of timer_read_time().
Unfortunately, it raised the possibility for timer_read_time() to be
out of sync with the scheduler's understanding of the current time.
In particular, it was common for the timer irq to overflow the 16bit
hardware counter once at startup, and this would lead to
timer_read_time() always returning a time ~4ms ahead of the scheduler
on 16Mhz chips. This resulted in "Move queue empty" errors.
To resolve this issue, only increment timer_high from timer_periodic()
and make sure the timer irqs start immediately after timer_init().
Signed-off-by: Kevin O'Connor <kevin@koconnor.net>
The code alignment of the TC0_Handler function seems to noticeably
impact performance benchmarks. Set the function alignment to 16 bytes
to improve testing consistency. An alignment of 16 doesn't
necessarily improve performance, but it seems to improve testing
consistency on code changes unrelated to timer dispatch.
Signed-off-by: Kevin O'Connor <kevin@koconnor.net>
Rename the file to make it clear that the code is helper functions for
boards with irq based timers.
Signed-off-by: Kevin O'Connor <kevin@koconnor.net>
Rename sched_timer_kick() to sched_timer_dispatch() and move its loop
into its callers in the board code. This eliminates the need to
export timer_try_set_next() from the board code.
Signed-off-by: Kevin O'Connor <kevin@koconnor.net>
Reset the timer repeat checks on shutdown, otherwise it is possible to
get into an infinite shutdown loop.
Signed-off-by: Kevin O'Connor <kevin@koconnor.net>
On faster MCUs where a delay is needed between step and unstep use a
"busy loop" in the scheduler instead of trying to schedule to the
unstep time. This reduces the chance of jitter in the scheduler
accumulating.
Signed-off-by: Kevin O'Connor <kevin@koconnor.net>
Directly use the Kconfig defined CONFIG_CLOCK_FREQ in the code and
avoid defining F_CPU. Also, remove the unnecessary O2 option - that
is already the default from the main makefile.
Signed-off-by: Kevin O'Connor <kevin@koconnor.net>
Explicilty load the timer waketime variable into local variables in
sched_timer_kick(). Change the optimization level from Os to O2.
This helps gcc to avoid unnecessary reloads from memory in the common
stepper_event() case.
Signed-off-by: Kevin O'Connor <kevin@koconnor.net>
Provide hand-coded assembler for timer_is_before() on AVR as that code
is used frequently in the time-critical timer dispatch loop and gcc
doesn't do a good job at compiling that comparison code. Remove the
no longer needed waketime+1 hack from reschedule_timer().
Signed-off-by: Kevin O'Connor <kevin@koconnor.net>
Move sched_is_before() from sched.c to timer_is_before() in the board
specific timer code. This allows the board code to provide its own
definition.
Also, remove the sched_from_us() and sched_read_time() wrapper
functions and change the callers to directly invoke timer_from_us() /
timer_read_time().
Signed-off-by: Kevin O'Connor <kevin@koconnor.net>
Commit f28eb902 reworked the switch to fix int16 encoding. However,
at least one version of avr gcc doesn't like that switch layout (it
uses a jump table). Reorg the switch to avoid that issue.
Signed-off-by: Kevin O'Connor <kevin@koconnor.net>
The code wasn't properly sign-extending 16bit integers which caused
int16_t reports in output() to appear as uint16_t.
Signed-off-by: Kevin O'Connor <kevin@koconnor.net>
Most of sam3x8e/timer.c is going to be platform agnostic for any board
with standard irq handling. Move the generic code into a new file
generic/timer.c.
Signed-off-by: Kevin O'Connor <kevin@koconnor.net>
If a shutdown occurs while the machine is already shutdown, then keep
the original shutdown reason code.
Signed-off-by: Kevin O'Connor <kevin@koconnor.net>
It's tricky to reschedule the timer irq correctly (due to race
conditions with the irq) and in practice it's very rarely needed.
Handle the special cases in the generic sched.c code so that the board
code doesn't have to handle it.
Signed-off-by: Kevin O'Connor <kevin@koconnor.net>
Introduce a dummy sentinel timer object that is always the last item
on timer_list. This optimizes the timer_list walking code as it no
longer needs to check for NULL when traversing the list.
Signed-off-by: Kevin O'Connor <kevin@koconnor.net>
The hardware timer overflow bit can be used to optimize the conversion
from 16bit timers to 32bit timers.
Signed-off-by: Kevin O'Connor <kevin@koconnor.net>
Move the logic to calculate and report the stepper's current position
from endstop.c to stepper.c. This localizes the stepper code into
stepper.c.
Signed-off-by: Kevin O'Connor <kevin@koconnor.net>
It's better to not have the unimplemented command defined so that the
host can detect when it is actually implemented.
Signed-off-by: Kevin O'Connor <kevin@koconnor.net>
Consistently use an "oid_" prefix on the oid functions - this makes
them similar to other functions with a common prefix.
Signed-off-by: Kevin O'Connor <kevin@koconnor.net>
Detect the maximum size of each "move queue" item during the
configuration phase instead of using the stepper move struct. This
allows the stepper code to be contained entirely in stepper.c and it
allows for future run time allocations from other types of objects.
Signed-off-by: Kevin O'Connor <kevin@koconnor.net>
Store a full 64bit uptime in the mcu and query it at the start of each
connection. This ensures the host's 64bit clock is always in synch
with the mcu's clock.
Signed-off-by: Kevin O'Connor <kevin@koconnor.net>
The irq handling in sam3x8e isn't specific to the sam3x8e proccessor -
it's generic for all armcm type machines. So, move the definitions
into a new file generic/armcm-irq.c
Signed-off-by: Kevin O'Connor <kevin@koconnor.net>
It's possible for sched_del_timer() to be called on a timer that fires
just after sched_del_timer disables irqs but before the next timer is
scheduled. In this case be sure to clear the irq pending status flag
after scheduling the next timer so that a delayed irq doesn't cause
the wrong timer to be run. For the same reason, make sure to check
the irq pending status flag at the start of the timer irq.
Also, as a safety check, make sure timer_set_next() isn't called from
within the timer irq dispatch loop.
Signed-off-by: Kevin O'Connor <kevin@koconnor.net>
There is a small possibility that a shutdown could occur between
clearing transmit_max and clearing transmit_pos - so make sure to
handle the case where transmit_pos > transmit_max.
Signed-off-by: Kevin O'Connor <kevin@koconnor.net>
As long as the code is careful when writing the in_sendf variable it
should be safe to update it without having to disable irqs.
Also, make sure in_sendf is cleared on shutdown.
Signed-off-by: Kevin O'Connor <kevin@koconnor.net>
Allow sendf() to be called from irq and timer context - check for the
case where sendf() is called while already in sendf() and simply
discard those messages. This makes it safe to use output() debugging
calls even in irq and timer context.
Signed-off-by: Kevin O'Connor <kevin@koconnor.net>
Use a base of 256 instead of 65536 when calculating the sum of the
square of the clock differences in the stats. This makes the
calculation more accurate. Export the new base via DECL_CONSTANT for
the host to access. Use DIV_ROUND_UP() when adjusting for the base to
ensure no lost ticks. Do the division after multiplication in the
common case where the time between stats_task() invocations is less
than 64K ticks.
Signed-off-by: Kevin O'Connor <kevin@koconnor.net>
Kevin O'Connor