klibc - Apr 2012 - klibc 2.0~rc3-1 ./. mksh 40.9.20120414-2

Hi,

after changing mksh to no longer use mkstemp, I decided to try
and build mksh-static with klcc on most platforms, in Debian.
Here are the results by Debian architecture:

alpha	(4) fails
amd64	works
armel	works
armhf	(3) fails
i386	works
ia64	works
m68k	(1) workaround
mips	works
mipsel	works
powerpc	(2) fails
ppc64	works (!)
s390	works
s390x	works
sparc	works

Architectures not listed either do not have a working klibc 2.0~rc3,
or did not build mksh or one of its dependencies due to issues with
the port (such as debhelper being uninstallable), or do not have a
Linux kernel.

The ssize_t prototype is bogus at least on s390:
../../shf.c:453:3: warning: format '%zd' expects argument of type
'signed size_t', but argument 4 has type 'ssize_t' [-Wformat]
That is not a problem, though.

(1) m68k: fails, unless I do *not* add -O2 or -Os to CFLAGS
    and do always add -g (might be a toolchain bug?); this
    workaround was added to the Debian mksh source package

https://buildd.debian.org/status/fetch.php?pkg=mksh&arch=powerpc&ver=40.9.20120414-2&stamp=1334431629
(2) powerpc: amazingly, ppc64 does NOT fail? this looks like
    m68k without the above-mentioned workaround: the failure
    I already saw often: after running an external programme,
    control flow does not return to the shell, and the child
    is not cleaned up (sticks around as zombie); might be a
    signal delivery problem (SIGCHLD), signal mask stuff, etc.

https://buildd.debian.org/status/fetch.php?pkg=mksh&arch=armhf&ver=40.9.20120414-2&stamp=1334433386
(3) armhf: one of the tests has the memory pool allocator
    error out as one of its internal data structures got
    corrupted, or a wrong pointer passed to aresize/afree

http://buildd.debian-ports.org/status/fetch.php?pkg=mksh&arch=alpha&ver=40.9.20120414-2&stamp=1334504721
(4) alpha: random SIGSEGVs that happen with neither eglibc
    nor dietlibc so almost certainly (considering mksh also
    runs with other OSes on DEC Alpha systems, even with
    the Compaq C compiler) are problems in klibc


If someone wants to have a stab at these: there are Debian
porterboxen around for at least powerpc and armhf; alpha
might one get access to; for m68k, there is an up-to-date
(April 2012) VM base image and I can give instructions how
to run the VM. Since the failures happen when executing
certain shell code, tracking them down to C code is very
difficult; the gdb way (or printf debugging, well, use
shellf() inside mksh code instead) might be easier.

bye,
//mirabilos
-- 
I believe no one can invent an algorithm. One just happens to hit upon it
when God enlightens him. Or only God invents algorithms, we merely copy them.
If you don't believe in God, just consider God as Nature if you won't
deny
existence.		-- Coywolf Qi Hunt

Dixi quod?
>The ssize_t prototype is bogus at least on s390:
>../../shf.c:453:3: warning: format '%zd' expects argument of type
'signed size_t', but argument 4 has type 'ssize_t' [-Wformat]
And cris:

? klibc.git/usr/include/bits32/bitsize/stddef.h

    9 #if defined(__s390__) || defined(__cris__)
   10 typedef unsigned long size_t;
   11 #else
   12 typedef unsigned int size_t;
   13 #endif
   14
   15 #define _PTRDIFF_T
   16 typedef signed int ptrdiff_t;

? klibc.git/usr/include/sys/types.h

   12 #define _SSIZE_T
   13 typedef ptrdiff_t ssize_t;

Combined, these two definitions make the problem.

The fix applied to MirBSD, dietlibc and Android bionic libc is
to rely on GCC?s definitions for size_t, and hack that into an
ssize_t definition:

? mircvs://src/sys/sys/types.h

  116 #if !defined(_GCC_SIZE_T)
  117 #define _GCC_SIZE_T
  118 typedef __SIZE_TYPE__   size_t;
  119 #endif
  120
  121 /* e-eeeevil kludge, but apparently works */
  122 #define unsigned        signed
  123 typedef __SIZE_TYPE__   ssize_t;
  124 #undef unsigned

Something like this would also work:

	typedef __SIZE_TYPE__ size_t;
	#define unsigned /* nothing */
	typedef signed __SIZE_TYPE__ ssize_t;
	#undef unsigned

So far, the code from above MirBSD header has never been a problem
though, as GCC is _very_ picky about its type definitions, such as
size_t and wchar_t:

? mircvs://gcc/gcc/config/mirbsd.h

  135 #define WCHAR_TYPE "short unsigned int"
  141 #define WINT_TYPE "unsigned int"
  144 #define SIZE_TYPE "long unsigned int"
  147 #define PTRDIFF_TYPE "long int"

Even changing the order to, say, "unsigned short int" for WCHAR_TYPE,
led to errors later on, so this is pretty constrained.


If you want, I can prepare and submit a klibc patch to use GCC?s
definitions similar to what I?ve done for bionic and dietlibc.

bye,
//mirabilos
-- 
FWIW, I'm quite impressed with mksh interactively. I thought it was much
*much* more bare bones. But it turns out it beats the living hell out of
ksh93 in that respect. I'd even consider it for my daily use if I hadn't
wasted half my life on my zsh setup. :-) -- Frank Terbeck in #!/bin/mksh

On 04/20/2012 09:12 AM, Thorsten Glaser wrote:
> 
> If you want, I can prepare and submit a klibc patch to use GCC?s
> definitions similar to what I?ve done for bionic and dietlibc.
> 
Yes, that would be the right thing.

	-hpa