Re: [PATCH 3/3] add s390x port

[Rich Felker <dalias@libc.org>]

Sorry for taking so long to give this the attention it deserves. I
think it looks good and is mostly ready to commit. A few
questions/comments though:

On Mon, Jul 25, 2016 at 10:53:00PM -0500, Bobby Bingham wrote:
> diff --git a/arch/s390x/bits/alltypes.h.in b/arch/s390x/bits/alltypes.h.in
> new file mode 100644
> index 0000000..1a83846
> --- /dev/null
> +++ b/arch/s390x/bits/alltypes.h.in
> @@ -0,0 +1,26 @@
> +#define _Addr long
> +#define _Int64 long
> +#define _Reg long
> +
> +TYPEDEF __builtin_va_list va_list;
> +TYPEDEF __builtin_va_list __isoc_va_list;
> +
> +#ifndef __cplusplus
> +TYPEDEF int wchar_t;
> +#endif
> +
> +TYPEDEF double float_t;

Reportedly this is wrong (a bug copied from glibc) and floats are
actually evaluated as float on s390. Can you check and confirm?

> diff --git a/arch/s390x/bits/ipc.h b/arch/s390x/bits/ipc.h
> new file mode 100644
> index 0000000..4710c12
> --- /dev/null
> +++ b/arch/s390x/bits/ipc.h
> @@ -0,0 +1,14 @@
> +struct ipc_perm {
> +	key_t __ipc_perm_key;
> +	uid_t uid;
> +	gid_t gid;
> +	uid_t cuid;
> +	gid_t cgid;
> +	mode_t mode;
> +	unsigned short __pad1;
> +	unsigned short __ipc_perm_seq;
> +	unsigned long __pad2;
> +	unsigned long __pad3;
> +};

__ipc_perm_seq is int on other archs, without the padding. It's not a
standard type so I don't think we're forced to have matching type, but
I wonder why this is. OTOH maybe some current archs are buggy because
of this..

> diff --git a/arch/s390x/syscall_arch.h b/arch/s390x/syscall_arch.h
> [...]
> +#define SYSCALL_USE_SOCKETCALL

Is this intentional? s390x does not (or did not always) have dedicated
syscalls for socket ops?

> diff --git a/src/thread/s390x/__tls_get_offset.s b/src/thread/s390x/__tls_get_offset.s
> new file mode 100644
> index 0000000..8ee92de
> --- /dev/null
> +++ b/src/thread/s390x/__tls_get_offset.s
> @@ -0,0 +1,17 @@
> +	.global __tls_get_offset
> +	.type __tls_get_offset,%function
> +__tls_get_offset:
> +	stmg  %r14, %r15, 112(%r15)
> +	aghi  %r15, -160
> +
> +	la    %r2, 0(%r2, %r12)
> +	brasl %r14, __tls_get_addr
> +
> +	ear   %r1, %a0
> +	sllg  %r1, %r1, 32
> +	ear   %r1, %a1
> +
> +	sgr   %r2, %r1
> +
> +	lmg   %r14, %r15, 272(%r15)
> +	br    %r14

Can you explain this function and why a tail-call to __tls_get_addr
doesn't work? Based on the (albeit 32-bit) info in
https://www.akkadia.org/drepper/tls.pdf it seems like it could but
maybe I'm missing something. Is there a special contract about
clobbers?

Rich