view ToDo @ 306:9f2aff560fa7

- todo update reflecting mips softfloat status more precisely (as only lacking for mips64/n32 callconv)
author Tassilo Philipp
date Sat, 09 Jun 2018 18:03:18 +0200
parents a6963fd4a8cc
children 8b0fc583ce62
line wrap: on
line source

1.1:
----
- plain_cpp only supports windows; fix
- ./configure on macOS does not pass CFLAGS etc.. to Makefile.config
- support platforms without shared libraries so they can be built by just typing `./configure;make` (e.g. minix < 3.2.1)
  * also make sure minix with shared library support builds and works in general (>= 3.2.1) as well as ELF on Minix (>= 3.2.0) to begin with (as that's the base, and we only test on 3.1.8 so far)
- dyncall supports calls to functions where prototype exists (with and without varargs); however not when no
  prototype exist (e.g. just calling some pointer), without any declaration... even possible in C, as one has
  to cast to the function type? need to support?


- dyncallback's "stack_cleanup" can be removed from non x86 impls... test if they still work
- target a minimum c version, for use with -pedantic, etc.? Right now we use e.g. extensions
  for anon structs, etc.
- the CallVM-free functions are per VTable, however the latter can be changed out
  at runtime... which basically changes potentially the deallocator...
  same goes for functions setting the mode. think about how to make this safer - currently
  we have a higher flexibility and we need to potentially clean up from the mode we are in,
  however, this is open to misuse
- check all vararg calling conv implementations if they promote correctly
  (e.g. float to double); we used to not care as it was the user's job to
  think of that, but we allow users to use ArgFloat, etc. on "..." now,
  taking care of the promotion; so it should be consistent all across the
  implementations -> also write testcode specifically for that
- autovar was never really pushed all the way through and still plays second
  fiddle to dyncall_macros.h, clean this all up, eventually (Tassilo: for now, I
  removed out of sync files to avoid confusion; but still probably duped code)
- ppc32 callback fixes on bsd @@@
- add portable "make install"
  * make install should also install manpages
- mailing list announcements (html email?)
- fix for NDS:
  * ARM:
    - callback_plain retval wrong (not only platform), area on stack where
      it's stored is probably not zeroed, so half-words, etc., come with garbage
  * THUMB:
    - ellipsis (might be test itself, not respecting fixed part of args)
    - callback_plain retval, see above under ARM
    - callf
- support Solaris11 CFLAGS with -D_FILE_OFFSET_BITS=64
- make sure selinux works (esp. regarding NX bits)
  Daniel has a patch for it, if needed, implementing allocwx with two memory blocks, one for w^x, mmaped to one for r|x
- support /SAFESEH on cl/win32
- test code functions should be in .so files, optionally, so the suites can directly
  be used to test dycnall bindings
- pkg-config support?
- quadmath support (long double)
- add tests using alloca(), as this might depend on the frame pointer being correctly
  set; we don't check for this, currently, at all

portasm:
--------
- add solaris support for x64

dyncall:
--------
- merge arm32_thumb_{gas/apple} and include thumb support for portasm
- x64 verification: return values are passed via RAX and RDX, and XMM0 and XMM1.
- find big endian arm box (also hardfloat), e.g. cubietruck: http://blog.netbsd.org/tnf/entry/working_arm_multiprocessor_support
- test if armhf/thumb is actually working... pulls in arm mode code in dyncall_callvm.c
- dyncall_call_mips_n32.h and dyncall_call_mips_n64.h are pretty much the same, share code
- support for return values: aggregate return values
- support for argument values: structures, half-precision, vector types, long double
  * make sure that struct support for ellipsis calls are not forgotten (copy everything by value)
- consider automatic type-promotion for arguments passed through DC_CALL_C_ELLIPSIS_VARARGS;
  this would make it easier to just pass arguments instead of having to know about the C
  type promotions (this is partly done, see todo-item above under 1.1)
- syscalls for x64
  * bsd
  * linux
- other syscalls
- test MIPS32 eabi big endian (current port works on psp, which is little endian)
- implement MIPS64 N32 (gcc -mabi=n32); both, little and big-endian (looks like NetNBSD on and
  EdgeRouter lite uses this: https://blog.netbsd.org/tnf/entry/hands_on_experience_with_edgerouter)
- MIPS64 n32 softfloat support (-msoft-float)
- implement MIPS64 EABI (gcc -mabi=eabi); both, little and big-endian
- support for Minix/arm
- improve ellipsis test:
  * test not just with one fixed arg, to also simulate alignment problems
  * test return values
  * convert from c++ to c, to have it also on plan9, and more portability

dynload:
--------
- RTLD_LAZY in darwin, why not in unix?
- dlGetLibraryPath: dl_iterate_phdr() used on OpenBSD, but not available on version < 3.7
- bug: test/nm crashes on qemu mips linux o32 debian etch
- test/resolve_self crashes/asserts on windows and bsds
- rename SymsInit to InitSyms and similar for cleanup function to reflect naming
  in dyncall
- get rid of Dl_info and friends (_GNU_SOURCE) in dynload/dynload_syms_elf.c
  for implementation of dlSymsNameFromValue(..)
- check if PSP can support dynload (missing elf.h, dlfcn.h, etc.); fixup readme if successful
- some functionality like java's System.mapLibraryName('name') might be helpful?
- Failed On Open Solaris 11:
    In file included from dynload_syms.c:28:
    dynload_syms_elf.c:56: error: syntax error before "Elf_Dyn"
    dynload_syms_elf.c:56: warning: data definition has no type or storage class
    dynload_syms_elf.c: In function `dlSymsInit':
    dynload_syms_elf.c:155: error: `pDyn' undeclared (first use in this function)
    dynload_syms_elf.c:155: error: (Each undeclared identifier is reported only once
    dynload_syms_elf.c:155: error: for each function it appears in.)
    dynload_syms_elf.c:155: error: syntax error before ')' token
    dynload_syms_elf.c:160: error: `DT_NULL' undeclared (first use in this function)
    dynload_syms_elf.c:163: error: `DT_STRTAB' undeclared (first use in this function)
    dynload_syms_elf.c:164: error: `DT_SYMTAB' undeclared (first use in this function)
    dynload_syms_elf.c:165: error: `DT_HASH' undeclared (first use in this function)
    make: *** [dynload_syms.o] Error 1

dyncallback:
------------
- callback_plain's return value not correct anymore on NDS (maybe just broken testcode?),
  see above under 1.1 items
  * check other platforms also, if asm stub initializes retval space, correctly
- finish PPC32 callbacks (see bugs section, below, BSD not working)
- test MIPS32 eabi big endian (current port works on psp, which is little endian)
- implement MIPS64 N32 (gcc -mabi=n32); both, little and big-endian (looks like NetNBSD on and
  EdgeRouter lite uses this: https://blog.netbsd.org/tnf/entry/hands_on_experience_with_edgerouter)
- MIPS64 n32 softfloat support (-msoft-float)
- implement MIPS64 EABI (gcc -mabi=eabi); both, little and big-endian
- support for Minix/arm

bindings:
---------
- release bindings as standalone packages (already done for rbdc as a gem and rdyncall on cran)
  * add note to documentation, where bindings are and that some bindings are svn-only ATM
- add rdoc documentation to ruby gem
- add pydoc for python
- add manpage for shdc
- expose callf-convenience functions to go
- add godoc for godc
- add javadoc to java
- add more bindings
- add callback support for erlang, go, java, lua, python and ruby (and shell? how?)
- add call-conv-mode-support for rbdc through signature

manual:
-------
- introduce doc overview tables and common vocab for calling convention specifics, following example of table on page 13 of https://www.linux-mips.org/pub/linux/mips/doc/ABI/MIPS-N32-ABI-Handbook.pdf; also as inspiration, e.g. user 'preserved' instead of 'permanent', etc. https://developer.apple.com/library/content/documentation/Xcode/Conceptual/iPhoneOSABIReference/Articles/ARMv6FunctionCallingConventions.html#//apple_ref/doc/uid/TP40009021-SW4
- make stack layout diagrams better, e.g. add grow direction (example: https://developer.apple.com/library/content/documentation/Xcode/Conceptual/iPhoneOSABIReference/art/arm_stack.jpg on https://developer.apple.com/library/content/documentation/Xcode/Conceptual/iPhoneOSABIReference/Articles/ARMv6FunctionCallingConventions.html#//apple_ref/doc/uid/TP40009021-SW1)
- callconv diagrams:
  * good idea to review it all, anyways, if the calling conventions are actually
    represented correctly (and if they are in sync with the .pdf, we used to have
    problems with multirow.4ht)
- provide more examples on how to use dyncall, dynload and dyncallback
  * e.g. enhance manual with a couple of examples (e.g. calling MessageBoxA on windows, etc.)
- manuals for bindings (as stated in bindings-section, above)
- under "dyncall support", in each calling convention appendix, list crucial details of e.g. big/little endian support, etc.
- specify integer models in use, use common ILP32, LLP64, etc. terms

regressions:
------------
- Haiku/gmake build doesn't build dynload due to the fact, that elf.h is not found
  on default installations (one has to install the system headers)
  * add to doc
- Solaris/gmake build doesn't build dynload and tests anymore (maybe not even dyncallback)
- test arm32/ATPCS/THUMB (and then change Status in doc Appendix)
- look over code and fix TODO or @@@ marks
- add ppc-support to doc (both, 32 and 64, as well as little and big endian remarks)

nice to have:
-------------
- update: microsoft visual c++ build files - integrate all test suites in solution file
- consistency: on x64 platforms: make both ABIs available ('win64','sysv') for all x64 OS ports
- microsoft build enhancement: autodetect assembly via C preprocessor (like with .S gcc files)
- test/thunk: some platforms do not allow for allocation of executable code on heap and stack -- 
  currently this test will crash, maybe we should collect information which platforms allow and which not.

bugs:
-----
- setting callconv, specifically for platforms that have only one (default)) excepts on some
  platforms, eg. setting X64_sysv on amd64/Linux - keeping cdecl which is the same in
  this case, works (this should work now...)
- vararg function calls don't work for ARM THUMB mode (Daniel's new interface for modesetting for varag
  args should be used); adapt test/ellipsis and/or retest

missing implementations (descending priority):
----------------------------------------------
- riscv
- itanium
- superh/j-core
- 68k, 88k, 6502, etc.
- arc
- alpha
- cell
- blackfin
- seaforth
- s/390
- z80
- hppa
- vax
- atmel avr
- propeller
- amd29k