GitLab

Detail:
  The kernel has a 'callback postponement flag' which will be set whenever an X SWI, called from user mode, returns a pointer to an error (and the error block looks like it's in RAM).
  Although this flag will help prevent the error buffer from being overwritten by any errors generated by callbacks, it will also have the effect of blocking Alt-Break and CLib's Escape handler, due to them both being reliant on callbacks.
  Since the flag may persist for a long time, it can be dangerous to leave the OS in this state (the flag can only be cleared by a SWI call from user mode - which may never happen if the program is stuck in a bad state).
  So to combat this, CLib will now make some effort to try and avoid leaving the postponement flag in situations where it shouldn't be needed.
  File changes:
  - c/armsys:
    - Change _sys_flen to return 0 for TTY streams, since calling OS_Args for them isn't going to generate any useful result (previously it was generating an error, causing the postponement flag to be left set on entry to main())
    - Be extra vigilant when entering main() to make sure the callback postponement flag isn't set, just in case other checks don't resolve it
  - kernel/s/k_body:
    - Change CopyError to attempt to clear the callback postponement flag if we suspect it's set. We've just copied the error to our own buffer, so it should be our job to worry about preventing that buffer being prematurely overwritten, not the kernel's.
Admin:
  Tested on BB-xM
  'while (1) {}' (and similar variants) now multitasks in task windows and is no longer unkillable
  Fixes issue reported on forums:
  https://www.riscosopen.org/forum/forums/4/topics/11542#posts-79767


Version 5.96. Tagged as 'RISC_OSLib-5_96'

Detail:
  kernel/s/k_body - Update AbortFindHandler to check that SP & SL are both word aligned, especially because the code will attempt to load from SL to check for the stack chunk magic marker.
Admin:
  Tested on BB-xM
  Avoids some "abort while in abort handler" type scenarios if an abort occurs while SP/SL are invalid


Version 5.95. Tagged as 'RISC_OSLib-5_95'

Detail:
  kernel/s/k_body - Eliminate a couple of global OS_SynchroniseCodeAreas calls in _kernel_init and _kernel_system
  s/initmodule - Eliminate a global OS_SynchroniseCodeAreas call in _Shared_Lib_Module_SWI_Code
Admin:
  Tested on iMx6
  Improves performance with SMP ROMs, where global D-cache clean isn't really possible


Version 5.93. Tagged as 'RISC_OSLib-5_93'

Detail:
  kernel/s/k_body, s/initmodule - Remove NOPs from after PSR manipulation macros; the macros now fully take care of ARM2 & StrongARM compatibility for us
Admin:
  Tested PlingSystem build on (ARM3) RISC OS 3.1
  Requires HdrSrc-2_63 for ARM2 compatibility


Version 5.92. Tagged as 'RISC_OSLib-5_92'

When multiple CPU architectures are supported, allow dynamically linked clients to be given optimal versions of routines where possible

Detail:
  This set of changes adds support for representing architecture-specific variants of routines within the export tables which the shared C library uses to fill in the client's stubs.
  This allows builds of the module which target multiple architectures (e.g. IOMD, Raspberry Pi, or softloads) to offer the most optimal versions of routines to clients wherever multiple variants exist.
  If only one architecture is supported, only one version of each routine will be built, and no variant tables will be generated.
  Currently routine selection for staticly linked clients isn't supported - staticly linked targets will only use the most backwards-compatible version of the routines, as before. Also not all routines are catered for yet (e.g. _ll_udiv)
  File changes:
  - s/h_common - New file containing macros used during variant table construction
  - Makefile - Set SHARED_C_LIBRARY to {TRUE} for SCL builds of assembler code, as per the SHARED_C_LIBRARY #define
  - kernel/s/k_body, s/longlong - For routines which have multiple variants available for the target architectures, build all variants, and generate variant table entries using the new RoutineVariant macro. This has required some reordering of the different variants to make sure that when both versions are built, the most backwards-compatible one will be first (and thus will be called by staticly linked clients)
  - s/h_modmacro - "Entry" macro (when building SCL export tables) changed to use DCD instead of B, to allow byte-aligned symbols to be represented (low bits of symbol address is now used as 'is a variant table' flag)
  - s/initmodule - Client stub initialisation rewritten to take into account the B -> DCD change, and to detect and evaluate variant entries
  - clib/s/cl_entries, kernel/s/k_entries, kernel/s/k_entries2 - Update library entry tables to reference the variant forms of routines where relevant
Admin:
  Tested on Raspberry Pi 1, 2, 3 (ROM), RISC OS 3.1 & StrongARM 3.7 (softload 26bit build)
  Requires Kernel-5_35-4_79_2_325


Version 5.91. Tagged as 'RISC_OSLib-5_91'

Prefer CPS over MSR for PSR manipulation. Use UDIV/SDIV for general-purpose division, and UMULL for /10.

Detail:
  s/k_body:
  - _kernel_irqs_off and _kernel_irqs_on now use CPS for disabling/enabling IRQs as opposed to MSR. Apart from being shorter code sequences, it's generally a faster instruction.
  - __rt_udiv and __rt_sdiv (and aliases) now use the UDIV and SDIV instructions if building for ARMv7VE
  - __rt_udiv10 and __rt_sdiv10 (and aliases) now use UMULL to multiply by 1/10 when building for targets with long multiply support, as this is faster than the old method. UDIV/SDIV can be fast too, but only for small numbers, making UMULL the best for the general case.
Admin:
  Tested on Cortex-A15
  Prototyping of division routines on assorted CPUs shows that UDIV/SDIV is generally between 20% and 400% faster than the old routine (Cortex-A7, Cortex-A53), or up to 1300% faster on Cortex-A15 (the CPU does not like the old routine!)
  Division by 10 is now about 20% faster across all appropriate CPUs


Version 5.89. Tagged as 'RISC_OSLib-5_89'

Detail:
  AcquireMutex doesn't have the same functionality as Arm2Swp, best not to
  make people think they might do. Also add an extra literal pool because
  otherwise some instructions don't reach when doing APCS-32 softload builds.

Retagged as 'RISC_OSLib-5_88'

Retagged as RISC_OSLib-5_88.

Detail:
  The SWP and SWPB instructions have finally been removed in ARMv8, after
  having been deprecated for a very long time. This version adds alternative
  versions of code that used to use them with ones that use LDREX/STREX and
  LDREXB/STREXB instead. Soft-loadable C libraries will choose between
  implementations at runtime using OS_PlatformFeatures; ROM builds only
  include the appropriate version for the target hardware.
Admin:
  Tested on Raspberry Pi 3.

Version 5.88. Tagged as 'RISC_OSLib-5_88'

Detail:
  kernel/s/k_body, s/initmodule - Escape some dollars contained in strings to avoid warnings from objasm
Admin:
  Resulting binary unchanged


Version 5.83. Retagged as 'RISC_OSLib-5_83'

Detail:
  kernel/s/k_body - At CopyUpDone in _kernel_system(), a big block of conditional code was converted to unconditional as part of the 32bit conversion process. However one line for setting up the error handler remained conditional, potentially preventing the correct error buffer pointer being set when about to start the child task.
Admin:
  Tested on BB-xM


Version 5.79. Tagged as 'RISC_OSLib-5_79'

Detail:
  kernel/s/k_body - Now ignores unaligned PC values, as they either indicate stack corruption or Thumb use (which the unwind handler doesn't support anyway)
Admin:
  Tested on Raspberry Pi with high processor vectors


Version 5.70. Tagged as 'RISC_OSLib-5_70'

Detail:
  kernel/s/k_body - NOP in Aborted2 is only needed if we're targeting StrongARM
  s/longlong - Make ReadCPUArch StrongARM & ARM2 safe
Admin:
  Tested in IOMD ROM softload


Version 5.67. Tagged as 'RISC_OSLib-5_67'

Was being used to conditionalise things which aren't really StrongARM related, now should be read as "support split I+D caches" switch.

Version 5.65. Tagged as 'RISC_OSLib-5_65'

Detail:
  * Added 64-bit file pointer API support, following the LFS spec v1.5
    (see Docs directory). Internally, now uses 64-bit file pointers
    exclusively. For now, the backend still limits you to 4GB-1 files, but
    clients of the C library won't need recompiling again to support larger
    files than this once they use the new API.
  * Fixed a large number of warnings produced in compilation, assembly and
    linking, many of them new and resulting from the improved checks in the
    latest toolchain.
  * Increased the maximum stream buffer size from 16MB to 1GB.
  * Added Hdr:SDFS and Hdr:SDIO to the list of headers used to build swis.h
    if they are present.
Admin:
  Verified that the new 64-bit file pointer buffering code works using a
  ROM build (since the toolchain makes extensive use of reading, writing
  and seeking internally).

Version 5.64. Tagged as 'RISC_OSLib-5_64'

Detail:
  clib/s/cl_stub, kernel/s/k_body, s/initmodule - Use OS_ReadSysInfo 6 to look up CLibWord/RISCOSLibWord/CLibCounter location each time it's needed instead of using hardcoded addresses. SWI overhead shouldn't be too bad since the zero page areas are rarely used.
  rlib/c/menu - Fix null pointer access when attempting to clear a submenu pointer via menu_submenu(...,...,NULL)
  s/h_workspc - Remove old local definitions of CLibCounter, RISCOSLibWorkSpace and CLibWorkSpace. Use Hdr:PublicWS instead.
Admin:
  Tested on rev A2 BB-xM


Version 5.60. Tagged as 'RISC_OSLib-5_60'

Detail:
  CallBackHandler in kernel.s.k_body now ensures that the stored 'sp' and 'sl' values are word-aligned before attempting to read from 'sl' to verify that it points to a stack chunk. This avoids an exception being triggered in certain situations when alignment exceptions are turned on and a piece of hand-written assembler is using the sl register to store data.
Admin:
  Further checks should be added to check that 'fp' is valid before calling _kernel_unwind, as _kernel_unwind currently only performs minimal checks that are not always sufficient.
  Tested on rev C2 beagleboard with KinoAmp pre-release.


Version 5.56. Tagged as 'RISC_OSLib-5_56'

Detail:
 This changes all the C and assembler includes to be a canoncial Unix format.
 Also match include paths to previous commit for EditIntern/DrawIntern/VerIntern
 Finally, also include some minor type fixes (NULL vs 0)
Admin:
 May be some other paths elsewhere in the source I'm not immediately able to fix.  Will address any issues ASAP, since this is a huge change.

Version 5.54. Not tagged

Detail:
  * Fixed signed pointer comparison: most top-bit-set pointers passed to
    _kernel_raise_error() were being translated into "Exit called"
  * Stack extension is now permitted in SYS mode; includes support in longjmp()
Admin:
  Verified that stack extension works.

Version 5.51. Tagged as 'RISC_OSLib-5_51'

Detail:
  * No longer uses aasm to build h.swis.
  * (Only) compatible with new C compilers: assumes const static data is
    placed in separate read-only areas, and suitable command-line switch
    is used to ensure that library static data is never placed in a zero-
    init area.
  * Stack extension code now thread-safe.
  * Heap thread-safety code (in clib) now uses SWP instruction test results
    from stack extension code (in kernel) - as a side effect, the exported
    symbol _swp_available now exists, if you need to use it elsewhere.
  * Slightly closer to having top-bit-set heap addresses working.
Admin:
  Tested in a Tungsten build, and with Customer W's test suite.

Version 5.48. Tagged as 'RISC_OSLib-5_48'

Detail:
  * _kernel_irqs_disabled() is included in ansilib build, now needed for
    ARM 2 SWP replacement in c.alloc
  * _kernel_entrypoint() for ansilib now initialises relocation offsets in
    root stack chunk, necessary if linking with -zm object code
  * system() implementation now calls DDEUtils_FlushCL if starting a
    short command line - necessary in case previous system() call was a
    long command line directed at a DDE-unaware application
Admin:
  Ansilib changes tested, DDE change not expected to cause problems.

Version 5.47. Not tagged

Detail:
  * Implemented the thread-safety mutex macros for the user mode heap.
  * Created ansilibm, a version of ansilib suitable for building modules.
  * Fixed an overestimation of the size of the workspace passed to
    SharedCLibrary_LibInitModule[APCS_32] by the stubs.
  * Fixed bug in SharedCLibrary_LibInitModule[APCS_32] regarding handling of
    non-reentrant modules - if module code was at a higher address than the
    module data, then all the intervening data would be pointlessly copied
    back over itself with no allowance for volatility.
  * Removed the bsearch source file, since it was only required for APCS-A
    which the library has not supported since the year 2000, and which the
    compiler is now no longer able to build.
Admin:
  ansilibm builds working modules, but thread safety has not been
  extensively tested.

Version 5.47. Tagged as 'RISC_OSLib-5_47'

* _Exit() added.
* Lots of new <math.h> functions (acosh, asinh, atanh, exp2, expm1,
  ilogb, log1p, log2, logb, scalbn, scalbln, cbrt, erf, erfc,
  lgamma, tgamma, nexttoward, fmaf). Float and long double forms
  of every function added; long double forms are included as another
  library object in the stubs rather than the shared library, as they
  just branch to the double form.
* Subnormal/NaN/infinity cases in various <math.h> functions improved.
* Added <tgmath.h>.
* Headers brought into line with CC 5.54.
* RMEnsures added to C library initialisation to try to load minimum
  CallASWI, FPEmulator, CLib. No errors reported if load fails.
* A few pointless inter-file dependencies removed to reduce minimum
  size of included ANSILib.

Version 5.46. Tagged as 'RISC_OSLib-5_46'

PCI added to swis.h
alloc.c updated to handle bigger slots (new code merged from ARM libraries)
Various 32-bit fixes for backtracing, and general trap handling.
Polite "Application is not 32-bit compatible" message.
Headers <stdint.h> and <inttypes.h> fixed to work in non-C99 mode.
txt changed to do new-style Delete behaviour

Version 5.44. Tagged as 'RISC_OSLib-5_44'

* Added two new library chunks, 4 and 5, which contain extensions to the kernel and C library respectively. These have no static data associated with them, just being extensions of the stub tables. The reason for this is to minimise wasted space in programs that don't use the C99 facilities; o.stubs is now a library split into 3 pieces - basic kernel and CLib, extra kernel and extra

  CLib; only the bits a program needs get included.

* Previous extensions to the C library stubs revoked - they now stop at _swix;
  all the new C99 functions now live in chunk 4. Anyone using those new
  functions should relink with new stubs and ensure this C library version.

* printf/scanf now support 64-bit types through "ll" and "j" length modifiers.

* Run-time support for VLAs (__rt_allocauto and __rt_freeauto) added. No
  attempt is currently made to clear up on longjmp or to cope with someone
  changing the kernel allocator while a VLA is active. These would be a
  future enhancement.

* Added complete 64-bit run-time support (48 functions) to kernel library;
  these functions are compatible with the ones used by the ARM ADS. Many of
  the simpler functions will not normally be used by the compiler, as it
  will generate inline code. There is scope for improvement by switching
  in MULL and CLZ-using forms of multiply and divide when possible.

* llabs and lldiv added to C library.

* Header files corrected in a few areas, and changed to match the C compiler.
  <stdint.h> and <stdbool.h> now require the compiler to be in C99 mode
  (as detected using __STDC_VERSION__).


Version 5.41. Tagged as 'RISC_OSLib-5_41'

  Aborts now give standard error messages (Abort on Data Transfer at... etc)
  *ShowRegs now filled in after aborts
  assert(), abort(), "free failed" and standard signal handlers now use Wimp
    error boxes if in the desktop
  Postmortem button on error boxes to view the postmortem

Also, x$multiply, x$divide, __rt_sdiv, x$remainder, x$udivide, __rt_udiv and
x$uremainder optimised.

Version 5.35. Tagged as 'RISC_OSLib-5_35'

Detail:
  * One check for 26-bit mode wan't being made, leading to the runtime reporting
    'Uncaught trap:' when a trap should have been catching the abort.  This
    prevented signal handlers from trapping some aborts.
  * When an abort was occuring in SVC mode, the stack wasn't being flattened,
    resulting in all sorts of nasty things when you tried to carry on from a
    signal handler (probably the main reason why you ever see 'No stack for trap
    handler'). SVC and undefined stacks are now flattened.  Further work should
    be done to restore the state of the OS after an abort, but this probably
    requires a new OS call to tidy itself up adequately.
Admin:
  Built 26-bit versions of the library with new 32-bit compatible entry points and checked abort handling with both sharedclibrary and ansilib.

Version 5.22. Tagged as 'RISC_OSLib-5_22'

Second attempt at getting 32-bit overlay manager to work.

Version 5.21. Tagged as 'RISC_OSLib-5_21'

Detail:
  This code would only work in 26-bit modes, since it stored the output
    flags using STR pc, [sp, #4*4]!.  When {CONFIG} is 32, it stores the
    CPSR instead.
Admin:
  Tested in 32-bit Lazarus build - this fixes the RCMMDriver stiffing
    (because it sits with interrupts off waiting for the C flag to be
    set on exit from an OS_Byte call it calls through _swix)

Version 5.19. Tagged as 'RISC_OSLib-5_19'

Detail:
  _kernel_setenv(varname, NULL) should delete the variables.  It now does.
Admin:
  Tested in module.

Version 5.18. Tagged as 'RISC_OSLib-5_18'

* Correction to static base location in modulewrap - was affecting some
  DSL 4000 video software.

Version 5.15. Tagged as 'RISC_OSLib-5_15'

OS_ReadSysInfo 6.

Version 5.08. Tagged as 'RISC_OSLib-5_08'

Version 5.07. Tagged as 'RISC_OSLib-5_07'

Version 5.06. Tagged as 'RISC_OSLib-5_06'

ROM patches.
fp and lr are now set correctly before calling a SWI. This allows much better
recovery if the SWI aborts or generates an error.

Version 4.95. Tagged as 'RISC_OSLib-4_95'

StrongARMfudge flag removed.
StrongARM compatibility set according to AMBKernel flag in Hdr:RISCOS
strftime %U and %W specifiers fixed for week 0.

Version 4.89. Tagged as 'RISC_OSLib-4_89'