this will allow the compiler to cache and reuse the result, meaning we
no longer have to take care not to load it more than once for the sake
of archs where the load may be expensive.

depends on commit 1c84c999 for
correctness, since otherwise the compiler could hoist loads during
stage 3 of dynamic linking before the initial thread-pointer setup.

versions of clang all the way back to 3.1 lack the bug this was
purportedly working around.

revert commit a603a75a.

as a result of commit 1c84c999 this is
now safe, assuming an interpretation of the somewhat-underspecified
attribute((const)) consistent with real-world usage.

commit a603a75a removed attribute
const from __errno_location and pthread_self, and the same reasoning
forced arch definitions of __pthread_self to use volatile asm,
significantly impacting code generation and imposing manual caching of
pointers where the impact might be noticable.

reorder the thread pointer setup and place it across a strong barrier
(symbolic function lookup) so that there is no assumed ordering
between the initialization and the accesses to the thread pointer in
stage 3.

don't repeat definition in two places.

the placement triggered -Wmisleading-indentation warnings if enabled,
and was gratuitously confusing to anyone reading the code.

fma is only available on recent x86_64 cpus and it is much faster than
a software fma, so this should be done with a runtime check, however
that requires more changes, this patch just adds the code so it can be
tested when musl is compiled with -mfma or -mfma4.

vfma is available in the vfpv4 fpu and above, the ACLE standard feature
test for double precision hardware fma support is
  __ARM_FEATURE_FMA && __ARM_FP&8
we need further checks to work around clang bugs (fixed in clang >=7.0)
  && !__SOFTFP__
because __ARM_FP is defined even with -mfloat-abi=soft
  && !BROKEN_VFP_ASM
to disable the single precision code when inline asm handling is broken.

For runtime selection the HWCAP_ARM_VFPv4 hwcap flag can be used, but
that requires further work.

These are only available on hard float target and sqrt is not available
in the base ISA, so further check is used.

These are available in the s390x baseline isa -march=z900.

previously (before and after rewrite), spurious escaping of path
separators as \/ was not treated the same as /, but rather got split
as an unpaired \ at the end of the fnmatch pattern and an unescaped /,
resulting in a mismatch/error.

for the case of \/ as part of the maximal literal prefix, remove the
explicit rejection of it and move the handling of / below escape
processing.

for the case of \/ after a proper glob pattern, it's hard to parse the
pattern, so don't. instead cheat and count repetitions of \ prior to
the already-found / character. if there are an odd number, the last is
escaping the /, so back up the split position by one. now the
char clobbered by null termination is variable, so save it and restore
as needed.

this code has been long overdue for a rewrite, but the immediate cause
that necessitated it was total failure to see past unreadable path
components. for example, A/B/* would fail to match anything, even
though it should succeed, when both A and A/B are searchable but only
A/B is readable. this problem both was caught in conformance testing,
and impacted users.

the old glob implementation insisted on searching the listing of each
path component for a match, even if the next component was a literal.
it also used considerable stack space, up to length of the pattern,
per recursion level, and relied on an artificial bound of the pattern
length by PATH_MAX, which was incorrect because a pattern can be much
longer than PATH_MAX while having matches shorter (for example, with
necessarily long bracket expressions, or with redundancy).

in the new implementation, each level of recursion starts by consuming
the maximal literal (possibly escaped-literal) path prefix remaining
in the pattern, and only opening a directory to read when there is a
proper glob pattern in the next path component. it then recurses into
each matching entry. the top-level glob function provided automatic
storage (up to PATH_MAX) for construction of candidate/result strings,
and allocates a duplicate of the pattern that can be modified in-place
with temporary null-termination to pass to fnmatch. this allocation is
not a big deal since glob already has to perform allocation, and has
to link free to clean up if it experiences an allocation failure or
other error after some results have already been allocated.

care is taken to use the d_type field from iterated dirents when
possible; stat is called only when there are literal path components
past the last proper-glob component, or when needed to disambiguate
symlinks for the purpose of GLOB_MARK.

one peculiarity with the new implementation is the manner in which the
error handling callback will be called. if attempting to match */B/C/D
where a directory A exists that is inaccessible, the error reported
will be a stat error for A/B/C/D rather than (previous and wrong
implementation) an opendir error for A, or (likely on other
implementations) a stat error for A/B. such behavior does not seem to
be non-conforming, but if it turns out to be undesirable for any
reason, backtracking could be done on error to report the first
component producing it.

also, redundant slashes are no longer normalized, but preserved as
they appear in the pattern; this is probably more correct, and falls
out naturally from the algorithm used. since trailing slashes (which
force all matches to be directories) are preserved as well, the
behavior of GLOB_MARK has been adjusted not to append an additional
slash to results that already end in slash.

commit 6ba5517a modified
__tls_get_addr to offset the address by +DTP_OFFSET (0x8000 on
powerpc, mips, etc.) and adjusted the result of DTPREL relocations by
-DTP_OFFSET to compensate, but missed changing the argument setup for
calls to __tls_get_addr from dlsym.

as explained in commit 6ba5517a, some
archs use an offset (typicaly -0x8000) with their DTPOFF relocations,
which __tls_get_addr needs to invert. on affected archs, which lack
direct support for large immediates, this can cost multiple extra
instructions in the hot path. instead, incorporate the DTP_OFFSET into
the DTV entries. this means they are no longer valid pointers, so
store them as an array of uintptr_t rather than void *; this also
makes it easier to access slot 0 as a valid slot count.

commit e75b16cf left behind cruft in
two places, __reset_tls and __tls_get_new, from back when it was
possible to have uninitialized gap slots indicated by a null pointer
in the DTV. since the concept of null pointer is no longer meaningful
with an offset applied, remove this cruft.

presently there are no archs with both TLSDESC and nonzero DTP_OFFSET,
but the dynamic TLSDESC relocation code is also updated to apply an
inverted offset to its offset field, so that the offset DTV would not
impose a runtime cost in TLSDESC resolver functions.

len was already passed as an argument, so don't use strcat, and use
memcpy instead of strcpy.

Rounding modes are not bit flags, but arbitrary non-negative integers.

when invoking the assembler, arm gcc does not always pass the right
flags to enable use of vfp instruction mnemonics. for C code it
produces, it emits the .fpu directive, but this does not help when
building asm source files, which tlsdesc needs to be. to fix, use an
explicit directive here.

commit 0beb9dfb introduced this
regression. it has not appeared in any release.

the specification for freeaddrinfo allows it to be used to free
"arbitrary sublists" of the list returned by getaddrinfo. it's not
clearly stated how such sublists come into existence, but the
interpretation seems to be that the application can edit the ai_next
pointers to cut off a portion of the list and then free it.

actual freeing of individual list slots is contrary to the design of
our getaddrinfo implementation, which has no failure paths after
making a single allocation, so that light callers can avoid linking
realloc/free. freeing individual slots is also incompatible with
sharing the string for ai_canonname, which the current implementation
does despite no requirement that it be present except on the first
result. so, rather than actually freeing individual slots, provide a
way to find the start of the allocated array, and reference-count it,
freeing the memory all at once after the last slot has been freed.

since the language in the spec is "arbitrary sublists", no provision
for handling other constructs like multiple lists glued together,
circular links, etc. is made. presumably passing such a construct to
freeaddrinfo produces undefined behavior.

the indirect function call is a significant portion of the code path
for the dynamic case, and most users are probably building for ISA
levels where it can be omitted.

we could drop at least one register save/restore (lr) with this
change, and possibly another (ip) with some clever shuffling, but it's
not clear whether there's a way to do it that's not more expensive, or
whether avoiding the save/restore would have any practical effect, so
in the interest of avoiding complexity it's omitted for now.

unlike other asm where the baseline ISA is used, these functions are
hot paths and use ISA-level specializations.

call-clobbered vfp registers are saved before calling __tls_get_new,
since there is no guarantee it won't use them. while setjmp/longjmp
have to use hwcap to decide whether to the fpu is in use, since
application code could be using vfp registers even if libc was
compiled as pure softfloat, __tls_get_new is part of libc and can be
assumed not to have access to vfp registers if tlsdesc.S does not.
thus it suffices just to check the predefined preprocessor macros. the
check for __ARM_PCS_VFP is redundant; !__SOFTFP__ must always be true
if the target ISA level includes fpu instructions/registers.

use the GNU C may_alias attribute if available, and fallback to naive
byte-by-byte loops if __GNUC__ is not defined.

this patch has been written to minimize changes so that history
remains reviewable; it does not attempt to bring the affected code
into a more consistent or elegant form.

the comparison must take place in the address space model as an
integer type, since comparing pointers that are not pointing into the
same array is undefined.

the subsequent d<s comparison however is valid, because it's only
reached in the case where the source and dest overlap, in which case
they are necessarily pointing to parts of the same array.

to make the comparison, use an unsigned range check for dist(s,d)>=n,
algebraically !(-n<s-d<n). subtracting n yields !(-2*n<s-d-n<0), which
mapped into unsigned modular arithmetic is !(-2*n<s-d-n) or rather
-2*n>=s-d-n.

Rewrote the AVL tree implementation:

- It is now non-recursive with fixed stack usage (large enough for
worst case tree height). twalk and tdestroy are still recursive as
that's smaller/simpler.

- Moved unrelated interfaces into separate translation units.

- The node structure is changed to use indexed children instead of
left/right pointers, this simplifies the balancing logic.

- Using void * pointers instead of struct node * in various places,
because this better fits the api (node address is passed in a void**
argument, so it is tempting to incorrectly cast it to struct node **).

- As a further performance improvement the rebalancing now stops
when it is not needed (subtree height is unchanged). Otherwise
the behaviour should be the same as before (checked over generated
random inputs that the resulting tree shape is equivalent).

- Removed the old copyright notice (including prng related one: it
should be licensed under the same terms as the rest of the project).

.text size of pic tsearch + tfind + tdelete + twalk:

   x86_64 i386 aarch64  arm mips powerpc ppc64le  sh4 m68k s390x
old   941  899    1220 1068 1852    1400    1600 1008 1008  1488
new   857  881    1040  976 1564    1192    1360  736  820  1408

These should have been added in commit
df6d9450
that added target specific PTRACE_ macros, but somehow got missed.

despite not being documented to do so in the standard or Linux
documentation, attempts to udp connect to 127.0.0.1 or ::1 generate
EADDRNOTAVAIL when the loopback device is not configured and there is
no default route for IPv6. this caused getaddrinfo with AI_ADDRCONFIG
to fail with EAI_SYSTEM and EADDRNOTAVAIL on some no-IPv6
configurations, rather than the intended behavior of detecting IPv6 as
unsuppported and producing IPv4-only results.

previously, only EAFNOSUPPORT was treated as unavailability of the
address family being probed. instead, treat all errors related to
inability to get an address or route as conclusive that the family
being probed is unsupported, and only fail with EAI_SYSTEM on other
errors.

further improvements may be desirable, such as reporting EAI_AGAIN
instead of EAI_SYSTEM for errors which are expected to be transient,
but this patch should suffice to fix the serious regression.

the clang internal assembler does not accept assembler options passed
via the usual -Wa mechanism, but it does accept -mimplicit-it directly
as an option to the compiler driver.

this facilitates building software that assumes a large default stack
size without any patching to call pthread_setattr_default_np or
pthread_attr_setstacksize at each thread creation site, using just
LDFLAGS.

normally the PT_GNU_STACK header is used only to reflect whether
executable stack is desired, but with GNU ld at least, passing
-Wl,-z,stack-size=N will set a size on the program header. with this
patch, that size will be incorporated into the default stack size
(subject to increase-only rule and DEFAULT_STACK_MAX limit).

both static and dynamic linking honor the program header. for dynamic
linking, all libraries loaded at program start, including preloaded
ones, are considered. dlopened libraries are not considered, for
several reasons. extra logic would be needed to defer processing until
the load of the new library is commited, synchronization woud be
needed since other threads may be running concurrently, and the
effectiveness woud be limited since the larger size would not apply to
threads that already existed at the time of dlopen. programs that will
dlopen code expecting a large stack need to declare the requirement
themselves, or pthread_setattr_default_np can be used.

stack size default is increased from 80k to 128k. this coincides with
Linux's hard-coded default stack for the main thread (128k is
initially committed; growth beyond that up to ulimit is contingent on
additional allocation succeeding) and GNU ld's default PT_GNU_STACK
size for FDPIC, at least on sh.

guard size default is increased from 4k to 8k to reduce the risk of
guard page jumping on overflow, since use of just over 4k of stack is
common (PATH_MAX buffers, etc.).

limit to 8MB/1MB, repectively. since the defaults cannot be reduced
once increased, excessively large settings would lead to an
unrecoverably broken state. this change is in preparation to allow
defaults to be increased via program headers at the linker level.

creation of threads that really need larger sizes needs to be done
with an explicit attribute.

these are now declared in pthread_impl.h.

access to defaults should be protected against concurrent changes.

per POSIX, deletion of a key for which some threads still have values
stored is permitted, and newly created keys must initially hold the
null value in all threads. these properties were not met by our
implementation; if a key was deleted with values left and a new key
was created in the same slot, the old values were still visible.

moreover, due to lack of any synchronization in pthread_key_delete,
there was a TOCTOU race whereby a concurrent pthread_exit could
attempt to call a null destructor pointer for the newly orphaned
value.

this commit introduces a solution based on __synccall, stopping the
world to zero out the values for deleted keys, but only does so lazily
when all key slots have been exhausted. pthread_key_delete is split
off into a separate translation unit so that static-linked programs
which only create keys but never delete them will not pull in the
__synccall machinery.

a global rwlock is added to synchronize creation and deletion of keys
with dtor execution. since the dtor execution loop now has to release
and retake the lock around its call to each dtor, checks are made not
to call the nodtor dummy function for keys which lack a dtor.

The condition occurs when
- thread #1 is holding the lock
- thread #2 is waiting for it on __futexwait
- thread #1 is about to release the lock and performs a_swap
- thread #3 enters the __lockfile function and manages to grab the lock
  before thread #1 calls __wake, resetting the MAYBE_WAITERS flag
- thread #1 calls __wake
- thread #2 wakes up but goes again to __futexwait as the lock is
  held by thread #3
- thread #3 releases the lock but does not call __wake as the
  MAYBE_WAITERS flag is not set

This condition results in thread #2 not being woken up. This patch fixes
the problem by making the woken up thread ensure that the flag is
properly set before going to sleep again.

Mainainer's note: This fixes a regression introduced in commit
c21f7507.

commit b114190b introduced spurious
realloc of the output buffer in cases where the result would exactly
fit in the caller-provided buffer. this is contrary to a strict
reading of the spec, which only allows realloc when the provided
buffer is "of insufficient size".

revert the adjustment of the realloc threshold, and instead push the
byte read by getc_unlocked (for which the adjustment was made) back
into the stdio buffer if it does not fit in the output buffer, to be
read in the next loop iteration.

in order not to leave a pushed-back byte in the stdio buffer if
realloc fails (which would violate the invariant that logical FILE
position and underlying open file description offset match for
unbuffered FILEs), the OOM code path must be changed. it would suffice
move just one byte in this case, but from a QoI perspective, in the
event of ENOMEM the entire output buffer (up to the allocated length
reported via *n) should contain bytes read from the FILE stream.
otherwise the caller has no way to distinguish trunated data from
uninitialized buffer space.

the SIZE_MAX/2 check is removed since the sum of disjoint object sizes
is assumed not to be able to overflow, leaving just one OOM code path.

morally, for null pointers a and b, a-b, a<b, and a>b should all be
defined as 0; however, C does not define any of them.

the stdio implementation makes heavy use of such pointer comparison
and subtraction for buffer logic, and also uses null pos/base/end
pointers to indicate that the FILE is not in the corresponding (read
or write) mode ready for accesses through the buffer.

all of the comparisons are fixed trivially by using != in place of the
relational operators, since the opposite relation (e.g. pos>end) is
logically impossible. the subtractions have been reviewed to check
that they are conditional the stream being in the appropriate reading-
or writing-through-buffer mode, with checks added where needed.

in fgets and getdelim, the checks added should improve performance for
unbuffered streams by avoiding a do-nothing call to memchr, and should
be negligible for buffered streams.

if EINVAL or ENOMEM happened before the first getc_unlocked, it was
possible that the stream orientation had not yet been set.

on some archs, linux support for futex operations (including
robust_list processing) that depend on kernelspace CAS is conditional
on a runtime check. as of linux 4.18, this check fails unconditionally
on nommu archs that perform it, and spurious failure on powerpc64 was
observed but not explained. it's also possible that futex support is
omitted entirely, or that the kernel is older than 2.6.17. for most
futex ops, ENOSYS does not yield hard breakage; userspace will just
spin at 100% cpu load. but for robust mutexes, correct behavior
depends on the kernel functionality.

use the get_robust_list syscall to probe for support at the first call
to pthread_mutexattr_setrobust, and block creation of robust mutexes
with a reportable error if they can't be supported.