This feature is specifically for the C++ compiler to offload calling
thread_local object destructors on thread program exit, to glibc.
This is to overcome the possible complication of destructors of
thread_local objects getting called after the DSO in which they're
defined is unloaded by the dynamic linker. The DSO is marked as
'unloadable' if it has a constructed thread_local object and marked as
'unloadable' again when all the constructed thread_local objects
defined in it are destroyed.
Previously, we would see a bad frame in the gdb backtrace output, e.g.:
(gdb) bt
#0 foo () at foo.c:5
#1 0x000000aaaab68ee8 in start_thread () from /lib/libpthread.so.0
#2 0x000000aaaad01c88 in clone () from /lib/libc.so.6
#3 0x0000000000000000 in ?? ()
With this change the bogus frame #3 is gone and we have the
same output as x86 does for the same program.
Atomic ops are issued directly from the core, rather than
potentially sitting in the write buffer, so can improve the
performance of other waiters. In addition, if we didn't end
up pulling a copy of the cache line where the lock is into cache,
by using an atomic op we don't have to acquire the cache line
before we can unlock.
With gcc 4.8 tilegx has support for -mcmodel=large, to tolerate very
large shared objects. This option changes the compiler output to
not include direct jump instructions, which have a range of only
2^30, i.e +/- 512MB. Instead the compiler marshalls the target PCs
into registers and then uses jump- or call-to-register instructions.
For glibc, the upshot is that we need to arrange for a few functions
to tolerate the possibility of a large range between the PC and
the target. In particular, the crti.S and start.S code needs
to be able to reach from .init to the PLT, as does gmon-start.c.
The elf-init.c code has the reverse problem, needing to call from
libc_nonshared.a (linked at the end of shared objects) back to the
_init section at the beginning.
No other functions in *_nonshared.a need to be built this way, as
they only call the PLT (or potentially each other), but all of that
code is linked at the very end of the shared object.
We don't build the standard -static archives with this option as the
performance cost is high enough and the use case is rare enough that
it doesn't seem worthwhile. Instead, we would encourage developers
who need the -static model with huge executables to build a private
copy of glibc and configure it with -mcmodel=large.
Note that libc.so et al don't need any changes; the only changes
are for code that is statically linked into user code built with
-mcmodel=large.
For the assembly code, I just rewrote it so that it unconditionally
uses the large model. To be able to pass -mcmodel=large to
csu/elf-init.c and csu/gmon-start.c, I need to check to see if the
compiler supports that flag, since gcc 4.7 doesn't; I added the
support by creating a small Makefile fragment that just runs the
compiler to check.
Normally, the simulator is notified of absolute pathnames by the
_dl_load_hook hook. However, when a relative pathname is used, the
simulator may not know that the relative path matches a path that
it could figure out in the file system that it has access to.
Instead we provide a simplified version of the realpath function
so we can pass a plausible absolute pathname to the simulator.
Since we're now doing more work at object load time, we also add
a guard so we do no work at all if we're not running on the simulator.
- Override <memcopy.h> so we use full 8-byte word copies on tilegx32
for memmove, then use op_t in memcpy instead of the previous
locally-defined word_t just to avoid proliferating identical types.
- Fix bug in memcpy prefetch that caused us to never prefetch past
the first cache line.
- Optimize misaligned memcpy by inlining _wordcopy_fwd_dest_aligned
instead of just doing a dumb word-at-a-time copy.
- Make memcpy safe for forward copies by doing all the loads from
a given cache line prior to doing a wh64 (cache line zero-fill)
on the destination. Remove now-redundant src == dst check.
- Copy and optimize the generic wordcopy.c routines to use the tile
"double align" instruction instead of the MERGE macro; to avoid
offset addressing mode (which tile doesn't have) by rewriting the
pointer math to load and store with a zero index; and to use
post-increment addresses in the inner loops to improve scheduling.
This hook is useful for any arch-specific functionality that
should be done on loaded objects. For the tile architecture,
the hook is already provided (though we switch to using the new
macro name with this commit) and implements a simulator notifier
so that the simulator can load Elf symbols to match the object
and generate better error messages for PC's.
Also, remove a spurious definition of DL_UNMAP in dl-runtime.c
We can discover our x,y coordinate in the core mesh with an
mfspr instruction, multiply y by the core mesh width, and have
the core number without needing to ask the kernel.
The previous dummy definition (as type int) was fine in general, since
tile doesn't have floating-point registers, but it confused gdb's
configure, leading to later compile errors. This change also makes
prfpregset_t parallel to prgregset_t, which seems like generally the
right thing regardless of the non-existence of the actual registers :-)
(__ptrace_eventcodes): Add new value PTRACE_EVENT_SECCOMP from Linux 3.5.
(__ptrace_setoptions): Add new value PTRACE_O_TRACESECCOMP, adjust PTRACE_O_MASK.
Andreas Schwab