Re: [PATCH v2] add powerpc64 port

[Rich Felker <dalias@libc.org>]

On Thu, Apr 14, 2016 at 01:05:07AM +0200, Szabolcs Nagy wrote:
> * Bobby Bingham <koorogi@koorogi.info> [2016-04-04 00:26:11 -0500]:
> > +++ b/arch/powerpc64/bits/setjmp.h
> > @@ -0,0 +1 @@
> > +typedef unsigned long long __jmp_buf[66];
> 
> hm glibc seems to use long[64] with 16byte alignment,
> is the size diff because of alignment?

Good catch.

> > +#if defined(_XOPEN_SOURCE) || defined(_GNU_SOURCE) || defined(_BSD_SOURCE)
> > +#define MINSIGSTKSZ 2048
> > +#define SIGSTKSZ    8192
> > +#endif
> 
> i think these should be bigger, e.g. follow ppc 4k, 10k

Agreed.

> > +typedef struct {
> > +	unsigned long vrregs[32][2];
> > +	unsigned _pad[3];
> > +	unsigned vrsave;
> > +	unsigned vscr;
> > +	unsigned _pad2[3];
> > +} vrregset_t;
> 
> it seems this type should be 16 byte aligned like elf_vrreg_t
> 
> (aarch64 has a similar issue i plan to fix: there we use
> 128bit long double to get correct alignment, but kernel
> and glibc uses __int128 which is visible in public headers.
> in powerpc64 at least __vector128 is not exposed in glibc,
> only 16byte aligned structs.)

I don't know what the best solution here is.

> > +++ b/arch/powerpc64/bits/syscall.h
> > @@ -0,0 +1,718 @@
> > +#define __NR_restart_syscall          0
> > +#define __NR_exit                     1
> > +#define __NR_fork                     2
> ....
> > +#define SYS_restart_syscall __NR_restart_syscall
> > +#define SYS_exit __NR_exit
> > +#define SYS_fork __NR_fork
> 
> i prefer SYS_ to use numbers too (easier to update for me),
> but it should be probably fixed together across archs.

I want to replace arch/*/bits/syscall.h with .in files that get
processed as something like:

	sed -e p -e s/__NR_/SYS_/ < $< > $@

This would reduce the source tree size a lot and eliminate the risk of
inconsistency.

> > +++ b/arch/powerpc64/reloc.h
> > @@ -0,0 +1,32 @@
> > +#include <endian.h>
> > +
> > +#if __BYTE_ORDER == __LITTLE_ENDIAN
> > +#define ENDIAN_SUFFIX "le"
> > +#else
> > +#define ENDIAN_SUFFIX ""
> > +#endif
> > +
> > +#define LDSO_ARCH "powerpc64" ENDIAN_SUFFIX
> > +
> 
> gcc also has -sf, i'm not sure if we should care about
> ppc64 soft-float

Probably not, but perhaps we should test in configure and reject it if
it's not supported correctly.

> > +#define TPOFF_K (-0x7000)
> > +
> > +#define REL_SYMBOLIC    R_PPC64_ADDR64
> > +#define REL_GOT         R_PPC64_GLOB_DAT
> > +#define REL_PLT         R_PPC64_JMP_SLOT
> > +#define REL_RELATIVE    R_PPC64_RELATIVE
> > +#define REL_COPY        R_PPC64_COPY
> > +#define REL_DTPMOD      R_PPC64_DTPMOD64
> > +#define REL_DTPOFF      R_PPC64_DTPREL64
> > +#define REL_TPOFF       R_PPC64_TPREL64
> 
> this reminds me that ppc(64) now has a tls optimization
> if the module is tagged with DT_PPC(64)_OPT, we don't
> need to implement it yet, but eventually elf.h should
> be updated.

Details?

> > +++ b/src/fenv/powerpc64/fenv.c
> > @@ -0,0 +1,68 @@
> > +#define _GNU_SOURCE
> > +#include <fenv.h>
> > +
> > +static inline double get_fpscr_f(void)
> > +{
> > +	double d;
> > +	__asm__ __volatile__("mffs %0" : "=d"(d));
> > +	return d;
> > +}
> > +
> > +static inline long get_fpscr(void)
> > +{
> > +	return (union {double f; long i;}) {get_fpscr_f()}.i;
> > +}
> > +
> > +static inline void set_fpscr_f(double fpscr)
> > +{
> > +	__asm__ __volatile__("mtfsf 255, %0" : : "d"(fpscr));
> > +}
> > +
> > +static void set_fpscr(long fpscr)
> > +{
> > +	set_fpscr_f((union {long i; double f;}) {fpscr}.f);
> > +}
> > +
> 
> yes now that .c is allowed under arch/ dirs, it
> makes sense to do fenv with such set/get_fpscr

Yes. At some point it would be nice to make it so most archs can just
define inline functions to load/store their fpu control/status
registers, and a generic fenv.c can implement all the functions. A few
weird archs (x86) would still need custom versions to be efficient but
most could use the generic framework. But this change should not be a
blocker for ppc64. I like the above approach fine for now.

Rich