On Mon, 2021-11-29 at 22:19 +0100, Samuel Thibault wrote:
> Hello,
>
> Svante Signell, le lun. 29 nov. 2021 18:40:48 +0100, a ecrit:
> > I've been working lately with on how to port valgrind to GNU/Hurd, and found
> > out that this is not a trivial task.
>
> It is indeed very far from a trivial task. I'd rather recommend to port
> gcc's lsan support, it will be most probably much easier to port.
Maybe it is. I might look into that later on. Thanks for the tip.
> > Seems like one have to make a suitable mix of linux and darwin code
> > (and to some extent freebsd/solaris code).
>
> Also it'd need a fair amount of code to reimplement the basic posixish
> needs of valgrind in terms of the Hurd RPCs.
Do you mean that Darwin does not use RPCs? Seems like valgrind has some
RPC-code
for Darwin already. In fact there is a directory containing mach code:
coregrind/m_mach/ but most *.defs code is absent. Basically only mach_msg and
mach_msg_trap are defined (and nothing about Hurd RPCs of course).
> >
> > movl $$__NR___pthread_sigmask, %eax
> > int $$0x80 /* should be sysenter? */
> > jc L_$0_7 /* __pthread_sigmask failed */
> Mach doesn't use int 0x80 but an lcall. See for instance
> glibc/sysdeps/mach/i386/syscall.S
Can't 0x80 just be replaced by lcall? Probably not. Looking at the x86-specific
assembly code for all architectures 0x80 seems to be used everywhere.
Additionally: My question was about the macro definition with an option
argument:
/* DO_SYSCALL MACH|MDEP|UNIX */
#define MACH 1
#define MDEP 2
#define UNIX 3
.macro DO_SYSCALL <what to add here?>
<and here>
.endm
.globl ML_(do_syscall_for_client_unix_WRK)
ML_(do_syscall_for_client_unix_WRK):
DO_SYSCALL UNIX
.globl ML_(do_syscall_for_client_mach_WRK)
ML_(do_syscall_for_client_mach_WRK):
DO_SYSCALL MACH
.globl ML_(do_syscall_for_client_mdep_WRK)
ML_(do_syscall_for_client_mdep_WRK):
DO_SYSCALL MDEP
