from https://github.com/ARM-software/optimized-routines,
commit 04884bd04eac4b251da4026900010ea7d8850edc

TOINT_INTRINSICS and EXP_USE_TOINT_NARROW cases are unused.

The underflow exception is signaled if the result is in the subnormal
range even if the result is exact (e.g. exp2(-1023.0)).

code size change: -1672 bytes.
benchmark on x86_64 before, after, speedup:

-Os:
   exp rthruput:  12.73 ns/call  6.68 ns/call 1.91x
    exp latency:  45.78 ns/call 21.79 ns/call 2.1x
  exp2 rthruput:   6.35 ns/call  5.26 ns/call 1.21x
   exp2 latency:  26.00 ns/call 16.58 ns/call 1.57x
-O3:
   exp rthruput:  12.75 ns/call  6.73 ns/call 1.89x
    exp latency:  45.91 ns/call 21.80 ns/call 2.11x
  exp2 rthruput:   6.47 ns/call  5.40 ns/call 1.2x
   exp2 latency:  26.03 ns/call 16.54 ns/call 1.57x

from https://github.com/ARM-software/optimized-routines,
commit 04884bd04eac4b251da4026900010ea7d8850edc

code size change: +2458 bytes (+1524 bytes with fma).
benchmark on x86_64 before, after, speedup:

-Os:
  log2 rthruput:  16.08 ns/call 10.49 ns/call 1.53x
   log2 latency:  44.54 ns/call 25.55 ns/call 1.74x
-O3:
  log2 rthruput:  15.92 ns/call 10.11 ns/call 1.58x
   log2 latency:  44.66 ns/call 26.16 ns/call 1.71x

from https://github.com/ARM-software/optimized-routines,
commit 04884bd04eac4b251da4026900010ea7d8850edc

Assume __FP_FAST_FMA implies __builtin_fma is inlined as a single
instruction.

code size change: +4588 bytes (+2540 bytes with fma).
benchmark on x86_64 before, after, speedup:

-Os:
   log rthruput:  12.61 ns/call  7.95 ns/call 1.59x
    log latency:  41.64 ns/call 23.38 ns/call 1.78x
-O3:
   log rthruput:  12.51 ns/call  7.75 ns/call 1.61x
    log latency:  41.82 ns/call 23.55 ns/call 1.78x

from https://github.com/ARM-software/optimized-routines,
commit 04884bd04eac4b251da4026900010ea7d8850edc

POWF_SCALE != 1.0 case only matters if TOINT_INTRINSICS is set, which
is currently not supported for any target.

SNaN is not supported, it would require an issignalingf
implementation.

code size change: -816 bytes.
benchmark on x86_64 before, after, speedup:

-Os:
  powf rthruput:  95.14 ns/call 20.04 ns/call 4.75x
   powf latency: 137.00 ns/call 34.98 ns/call 3.92x
-O3:
  powf rthruput:  92.48 ns/call 13.67 ns/call 6.77x
   powf latency: 131.11 ns/call 35.15 ns/call 3.73x

from https://github.com/ARM-software/optimized-routines,
commit 04884bd04eac4b251da4026900010ea7d8850edc

In expf TOINT_INTRINSICS is kept, but is unused, it would require support
for __builtin_round and __builtin_lround as single instruction.

code size change: +94 bytes.
benchmark on x86_64 before, after, speedup:

-Os:
  expf rthruput:   9.19 ns/call  8.11 ns/call 1.13x
   expf latency:  34.19 ns/call 18.77 ns/call 1.82x
 exp2f rthruput:   5.59 ns/call  6.52 ns/call 0.86x
  exp2f latency:  17.93 ns/call 16.70 ns/call 1.07x
-O3:
  expf rthruput:   9.12 ns/call  4.92 ns/call 1.85x
   expf latency:  34.44 ns/call 18.99 ns/call 1.81x
 exp2f rthruput:   5.58 ns/call  4.49 ns/call 1.24x
  exp2f latency:  17.95 ns/call 16.94 ns/call 1.06x

from https://github.com/ARM-software/optimized-routines,
commit 04884bd04eac4b251da4026900010ea7d8850edc

code size change: +177 bytes.
benchmark on x86_64 before, after, speedup:

-Os:
 log2f rthruput:  11.38 ns/call  5.99 ns/call 1.9x
  log2f latency:  35.01 ns/call 22.57 ns/call 1.55x
-O3:
 log2f rthruput:  10.82 ns/call  5.58 ns/call 1.94x
  log2f latency:  35.13 ns/call 21.04 ns/call 1.67x

from https://github.com/ARM-software/optimized-routines,
commit 04884bd04eac4b251da4026900010ea7d8850edc,
with minor changes to better fit into musl.

code size change: +289 bytes.
benchmark on x86_64 before, after, speedup:

-Os:
  logf rthruput:   8.40 ns/call  6.14 ns/call 1.37x
   logf latency:  31.79 ns/call 24.33 ns/call 1.31x
-O3:
  logf rthruput:   8.43 ns/call  5.58 ns/call 1.51x
   logf latency:  32.04 ns/call 20.88 ns/call 1.53x

Musl currently aims to support non-nearest rounding mode and does not
support SNaNs. These macros allow marking relevant code paths in case
these decisions are changed later (they also help documenting the
corner cases involved).

These don't have an effectw with -Os so not useful with default settings
other than documenting the expectation.

With --enable-optimize=internal,malloc,string,math the libc.so code size
increases by 18K on x86_64 and performance varies in -2% .. +10%.

These are supposed to be used in tail call positions when handling
special cases in new code. (fp exceptions may be raised "naturally"
by the common code path if special casing is more effort.)

This implements the error handling apis used in
https://github.com/ARM-software/optimized-routines
without errno setting.

Previously type casts or assignments were used for handling excess
precision, which assumed standard C99 semantics, but since it's a
rarely needed obscure detail, it's better to use explicit helper
functions to document where we rely on this.  It also helps if the
code is used outside of the libc in non-C99 compilation mode: with the
default excess precision handling of gcc, explicit inline asm barriers
are needed for narrowing on FLT_EVAL_METHOD!=0 targets.

I plan to use this in new code with the existing style that uses
double_t and float_t as much as possible.

One ugliness is that it is required for almost every return statement
since that does not drop excess precision (the standard changed this
in C11 annex F, but that does not help in non-standard compilation
modes or with old compilers).

C99 has ways to support fenv access, but compilers don't implement it
and assume nearest rounding mode and no fp status flag access. (gcc has
-frounding-math and then it does not assume nearest rounding mode, but
it still assumes the compiled code itself does not change the mode.
Even if the C99 mechanism was implemented it is not ideal: it requires
all code in the library to be compiled with FENV_ACCESS "on" to make it
usable in non-nearest rounding mode, but that limits optimizations more
than necessary.)

The math functions should give reasonable results in all rounding modes
(but the quality may be degraded in non-nearest rounding modes) and the
fp status flag settings should follow the spec, so fenv side-effects are
important and code transformations that break them should be prevented.

Unfortunately compilers don't give any help with this, the best we can
do is to add fp barriers to the code using volatile local variables
(they create a stack frame and undesirable memory accesses to it) or
inline asm (gcc specific, requires target specific fp reg constraints,
often creates unnecessary reg moves and multiple barriers are needed to
express that an operation has side-effects) or extern call (only useful
in tail-call position to avoid stack-frame creation and does not work
with lto).

We assume that in a math function if an operation depends on the input
and the output depends on it, then the operation will be evaluated at
runtime when the function is called, producing all the expected fenv
side-effects (this is not true in case of lto and in case the operation
is evaluated with excess precision that is not rounded away). So fp
barriers are needed (1) to prevent the move of an operation within a
function (in case it may be moved from an unevaluated code path into an
evaluated one or if it may be moved across a fenv access), (2) force the
evaluation of an operation for its side-effect when it has no input
dependency (may be constant folded) or (3) when its output is unused. I
belive that fp_barrier and fp_force_eval can take care of these and they
should not be needed in hot code paths.

Nothing is left from the original fdlibm header nor from the bsd
modifications to it other than some internal api declarations.

Comments are dropped that may be copyrightable content.

Code generation for SET_HIGH_WORD slightly changes, but it only affects
pow, otherwise the generated code is unchanged.

This makes it easier to build musl math code with a compiler that
does not support complex types (tcc) and in general more sensible
factorization of the internal headers.

FP_FAST_FMA can be defined if "the fma function generally executes about
as fast as, or faster than, a multiply and an add of double operands",
which can only be true if the fma call is inlined as an instruction.

gcc sets __FP_FAST_FMA if __builtin_fma is inlined as an instruction,
but that does not mean an fma call will be inlined (e.g. it is defined
with -fno-builtin-fma), other compilers (clang) don't even have such
macro, but this is the closest we can get.

(even if the libc fma implementation is a single instruction, the extern
call overhead is already too big when the macro is used to decide between
x*y+z and fma(x,y,z) so it cannot be based on libc only, defining the
macro unconditionally on targets which have fma in the base isa is also
incorrect: the compiler might not inline fma anyway.)

this solution works with gcc unless fma inlining is explicitly turned off.

POSIX: "[If] either O_TTY_INIT is set in oflag or O_TTY_INIT has the
value zero, open() shall set any non-standard termios structure
terminal parameters to a state that provides conforming behavior."

The Linux kernel tty drivers always perform initialisation on their
devices to set known good termios values during the open(2) call.  This
means that setting O_TTY_INIT to zero is conforming.

the weak version of __syscall_cp_c was using a tail call to __syscall
to avoid duplicating the 6-argument syscall code inline in small
static-linked programs, but now that __syscall no longer exists, the
inline expansion is no longer duplication.

the syscall.h machinery suppported up to 7 syscall arguments, only via
an external __syscall function, but we presently have no syscall call
points that actually make use of that many, and the kernel only
defines 7-argument calling conventions for arm, powerpc (32-bit), and
sh. if it turns out we need them in the future, they can easily be
added.

n32 and n64 ABIs add new argument registers vs o32, so that passing on
the stack is not necessary, so it's not clear why the 5- and
6-argument versions were special-cased to begin with; it seems to have
been pattern-copying from arch/mips (o32).

i've treated the new argument registers like the first 4 in terms of
clobber status (non-clobbered). hopefully this is correct.

the OABI passes these on the stack, using the convention that their
position on the stack is as if the first four arguments (in registers)
also had stack slots. originally this was deemed too awkward to do
inline, falling back to external __syscall, but it's not that bad and
now that external __syscall is being removed, it's necessary.

the inline syscall code is copied directly from powerpc64. the extent
of register clobber specifiers may be excessive on both; if that turns
out to be the case it can be fixed later.

it was never demonstrated to me that this workaround was needed, and
seems likely that, if there ever was any clang version for which it
was needed, it's old enough to be unusably buggy in other ways. if it
turns out some compilers actually can't do the register allocation
right, we'll need to replace this with inline shuffling code, since
the external __syscall dependency is being removed.

this is the first part of a series of patches intended to make
__syscall fully self-contained in the object file produced using
syscall.h, which will make it possible for crt1 code to perform
syscalls.

the (confusingly named) i386 __vsyscall mechanism, which this commit
removes, was introduced before the presence of a valid thread pointer
was mandatory; back then the thread pointer was setup lazily only if
threads were used. the intent was to be able to perform syscalls using
the kernel's fast entry point in the VDSO, which can use the sysenter
(Intel) or syscall (AMD) instruction instead of int $128, but without
inlining an access to the __syscall global at the point of each
syscall, which would incur a significant size cost from PIC setup
everywhere. the mechanism also shuffled registers/calling convention
around to avoid spills of call-saved registers, and to avoid
allocating ebx or ebp via asm constraints, since there are plenty of
broken-but-supported compiler versions which are incapable of
allocating ebx with -fPIC or ebp with -fno-omit-frame-pointer.

the new mechanism preserves the properties of avoiding spills and
avoiding allocation of ebx/ebp in constraints, but does it inline,
using some fairly simple register shuffling, and uses a field of the
thread structure rather than global data for the vdso-provided syscall
code address.

for now, the external __syscall function is refactored not to use the
old __vsyscall so it can be kept, but the intent is to remove it too.

this is a workaround to avoid a crashing regression on qemu-user when
dynamic TLS is installed at dlopen time. the sigaction syscall should
not be able to fail, but it does fail for implementation-internal
signals under qemu user-level emulation if the host libc qemu is
running under reserves the same signals for implementation-internal
use, since qemu makes no provision to redirect/emulate them. after
sigaction fails, the subsequent tkill would terminate the process
abnormally as the default action.

no provision to account for membarrier failing is made in the dynamic
linker code that installs new TLS. at the formal level, the missing
barrier in this case is incorrect, and perhaps we should fail the
dlopen operation, but in practice all the archs we support (and
probably all real-world archs except alpha, which isn't yet supported)
should give the right behavior with no barrier at all as a consequence
of consume-order properties.

in the long term, this workaround should be supplemented or replaced
by something better -- a different fallback approach to ensuring
memory consistency, or dynamic allocation of implementation-internal
signals. the latter is appealing in that it would allow cancellation
to work under qemu-user too, and would even allow many levels of
nested emulation.

This parameter was incorrectly declared to be a pointer to a function
accepting zero parameters.  The intent of makecontext is that it is
possible to pass integer parameters to the function, so this should
have been a pointer to a function accepting an unspecified set of
parameters.

Mark atanhi, atanlo, and aT in atanl.c as static, as they're not
intended to be part of the public API.

These are already static in the LDBL_MANT_DIG == 64 code, so this
patch is just making the LDBL_MANT_DIG == 113 code do the same thing.

The result is the same but takes less code.
Note that __execvpe calls getenv which calls __strchrnul so even
using static output the size of the executable won't grow.

commit 54ca6779 inadvertently
introduced bitwise and where logical and was intended. since the
right-hand operand is always 0 or -1 whenever the left-hand operand is
nonzero, the behavior happened to be equivalent.

priority inheritance is a feature to mitigate priority inversion
situations, where a execution of a medium-priority thread can
unboundedly block forward progress of a high-priority thread when a
lock it needs is held by a low-priority thread.

the natural way to do priority inheritance would be with a simple
futex flag to donate the calling thread's priority to a target thread
while it waits on the futex. unfortunately, linux does not offer such
an interface, but instead insists on implementing the whole locking
protocol in kernelspace with special futex commands that exist solely
for the purpose of doing PI mutexes. this would require the entire
"trylock" logic to be duplicated in the timedlock code path for PI
mutexes, since, once the previous lock holder releases the lock and
the futex call returns, the lock is already held by the caller.
obviously such code duplication is undesirable.

instead, I've made the PI timedlock success path set the mutex lock
count to -1, which can be thought of as "not yet complete", since a
lock count of 0 is "locked, with no recursive references". a simple
branch in a non-hot path of pthread_mutex_trylock can then see and act
on this state, skipping past the code that would check and take the
lock to the same code path that runs after the lock is obtained for a
non-PI mutex.

because we're forced to let the kernel perform the actual lock and
unlock operations whenever the mutex is contended, we have to patch
things up when it does the wrong thing:

1. the lock operation is not aware of whether the mutex is
   error-checking, so it will always fail with EDEADLK rather than
   deadlocking.

2. the lock operation is not aware of whether the mutex is robust, so
   it will successfully obtain mutexes in the owner-died state even if
   they're non-robust, whereas this operation should deadlock.

3. the unlock operation always sets the lock value to zero, whereas
   for robust mutexes, we want to set it to a special value indicating
   that the mutex obtained after its owner died was unlocked without
   marking it consistent, so that future operations all fail with
   ENOTRECOVERABLE.

the first of these is easy to solve, just by performing a futex wait
on a dummy futex address to simulate deadlock or ETIMEDOUT as
appropriate. but problems 2 and 3 interact in a nasty way. to solve
problem 2, we need to back out the spurious success. but if waiters
are present -- which we can't just ignore, because even if we don't
want to wake them, the calling thread is incorrectly inheriting their
priorities -- this requires using the kernel's unlock operation, which
will zero the lock value, thereby losing the "owner died with lock
held" state.

to solve these problems, we overload the mutex's waiters field, which
is unused for PI mutexes since they don't call the normal futex wait
functions, as an indicator that the PI mutex is permanently
non-lockable. originally I wanted to use the count field, but there is
one code path that needs to access this flag without synchronization:
trylock's CAS failure path needs to be able to decide whether to fail
with EBUSY or ENOTRECOVERABLE, the waiters field is already treated as
a relaxed-order atomic in our memory model, so this works out nicely.

there was no point in masking off the pshared bit when first loading
the type, since every subsequent access involves a mask anyway. not
masking it may avoid a subsequent load to check the pshared flag, and
it's just simpler.

commit 84d061d5 wrongly moved the
access to the global next_key outside of the scope of the lock. the
error manifested as spurious failure to find an available key slot
under concurrent calls to pthread_key_create, since the stopping
condition could be met after only a small number of slots were
examined.

commit d6c855ca caused this
"regression", though the behavior was undefined before, overlooking
that f->shend=0 was being used as a sentinel for "EOF" status (actual
EOF or hitting the scanf field width) of the stream helper (shgetc)
functions.

obviously the shgetc macro could be adjusted to check for a null
pointer in addition to the != comparison, but it's the hot path, and
adding extra code/branches to it begins to defeat the purpose.

so instead of setting shend to a null pointer to block further reads,
which no longer works, set it to the current position (rpos). this
makes the shgetc macro work with no change, but it breaks shunget,
which can no longer look at the value of shend to determine whether to
back up. Szabolcs Nagy suggested a solution which I'm using here:
setting shlim to a negative value is inexpensive to test at shunget
time, and automatically re-trips the cnt>=shlim stop condition in
__shgetc no matter what the original limit was.

commit 2de29bc9 left behind one
reference to pthread_mutex_trylock. fixing this also improves code
generation due to the namespace-safe version being hidde.