GitLab

Detail:
  c/cmodule - Try using OS_ReadSysInfo 6 to find IRQsema before falling back on the legacy address
  h/scheduler - Add IRQsema pointer extern, update ThreadResumed extern to cope with it now being in a data area
  s/scheduler - Use workspace IRQsema pointer instead of hardcoded legacy address. Moved PreEmptionRecovery & ThreadResumed into data area for easier access to IRQsema ptr.
Admin:
  Tested on rev A2 BB-xM


Version 0.06. Tagged as 'RTSupport-0_06'

Detail:
  - Makefile (hopefully temporarily) modified to use "-cpu 3" setting for compiling C files, to work around internal compiler error
  - RT_Deregister fixed (in s/scheduler) to behave properly if a routine attempts to deregister itself after being entered from the foreground via RT_Yield
  - Simple test program added to test correct behaviour of RT_Deregister
  - SVC stack allocation routines rewritten to cope with the possibility that OS_DynamicArea 9/10 may be unable to move pages (e.g. RT_Deregister called while IRQSema is nonzero)
  - Stack allocation routine also attempts to deal with kernels which have broken sparse area error reporting (anything below version 5.15), by falling back on a simpler allocation scheme that doesn't resize the dynamic area.
  - Erroneous comment in s/scheduler fixed
  - Erroneous #define in h/scheduler fixed
Admin:
  Tested on RISC OS 5.15 rev C2 beagleboard, RISC OS 5.12 Iyonix
  Makefile fix "fixes" bug #218


Version 0.05...

Detail:
  * First bug was that due to a careless slip-up in the very last stage of
    the process of switching to a routine (the loading of the SPSR), all
    routines were actually being entered in SYS mode with IRQs disabled. This
    was leading to horrendous interrupt latency - exactly the situation this
    module was written to avert! The most obvious reproducible symptom was
    that the real time clock ran slowly.
  * Once this was fixed, it revealed the presence of a more complicated bug:
    the module wasn't including the most-recently-executed routine along
    with the list of other routines in a non-sleeping state for a certain
    sub-set of actions. This was fixed by setting the pollword pointer to
    yet another magic address while it was executing, to minimise the
    disruption to existing code.
Admin:
  Tested with a deinterlacing routine.

Version 0.04. Tagged as 'RTSupport-0_04'

Version 0.03. Tagged as 'RTSupport-0_03'

Also fixed length calculation when setting system variable to exclude
trailing null byte (not that FileSwitch seems to mind!)

Version 0.02. Tagged as 'RTSupport-0_02'

This has so far been demonstrated to work with a single client under
RISC OS-STB 5.0.0 and 5.0.2.