the existing implementation of case mappings was very small (typically
around 1.5k), but unmaintainable, requiring manual addition of new
case mappings with each new edition of Unicode. often, it turned out
that newly-added case mappings were not easily representable in the
existing tightly-constrained table structures, requiring new hacks to
be invented and delaying support for new characters.

the new implementation added here follows the pattern used for
character class membership, with a two-level table allowing Unicode
blocks for which no data is needed to be elided. however, rather than
single-bit data, each character maps to a one of up to 6 case-mapping
rules available to its block, where 6 is floor(cbrt(256)) and allow 3
characters to be represented per byte (vs 8 with bit tables). blocks
that would need more than 6 rules designate one as an exception and
let lookup pass into a binary search of exceptional cases for the
block.

the number 6 was chosen empirically; many blocks would be ok with 4
rules (uncased, lower, upper, possible exceptions), some even just
with 2, but the latter are rare and fitting 4 characters per byte
rather than 3 does not save significant space. moreover, somewhat
surprisingly, there are sufficiently many blocks where even 4 rules
don't suffice without a lot of exceptions (blocks where some case
pairs are laced, others offset) that originally I was looking at
supporting variable-width tables, with 1-, 2-, or 3-bit entries,
thereby allowing blocks with 8 rules. as implemented in my
experiments, that version was significantly larger and involved more
memory accesses/cache lines.

improvements in size at the expense of some performance might be
possible by utilizing iswalpha data or merging the table of case
mapping identity with alphabetic identity. these were explored
somewhat when the code was first written, and might be worth
revisiting in the future.

somehow this seems to have been overlooked. add it now so that
subsequent overhaul of case mapping implementation will not introduce
a functional change at the same time.

linux fails the open with ENOSPC, but POSIX mandates EAGAIN.

for time64 support on 32-bit archs, the kernel interfaces use a
timespec layout padded to match the representation of a pair of 64-bit
values, which requires endian-specific padding.

use of an ordinary, non-bitfield, named member for the padding is
undesirable because, on big endian archs, it would alter the
interpretation of traditional (non-designated) initializers of the
form {s,ns}, initializing the padding instead of the tv_nsec member.
unnamed bitfield members solve this problem by not taking part in
initialization, and were the expected solution when the kernel
interfaces were designed. however, they also have further advantages
which we take advantage of here:

positioning of the padding could be controlled by having a
preprocessor conditional with separate definitions of struct timespec
for little and big endian, but whether padding should appear at all is
a function of whether time_t is larger than long. this condition is
not something the preprocessor can determine unless we were to define
a new macro specifically for that purpose.

by using unnamed bitfield members instead of ordinary named members,
we can arrange for the size of the padding to collapse to zero when it
should not be present, just by using sizeof(time_t) and sizeof(long)
in the bitfield width expression, which can be any integer constant
expression.

commit 2b4fd6f7 added time64 for this
function, but did so with a hidden assumption that the new time64
version of struct timex will be layout-compatible with the old one.
however, there is little benefit to doing it that way, and the cost is
permanent special-casing of 32-bit archs with 64-bit time_t in the
public interface definitions.

instead, do a full translation of the structure going in and out. this
commit is actually a revision to an earlier uncommited version of the
code.

presently the kernel does not actually define time64 versions of these
syscalls, and they're not really needed except to represent extreme
cpu time usage. however, x32's versions of the syscalls already behave
as time64 ones, meaning the functions were broken on x32 if the caller
used any part of the rusage result other than ru_utime and ru_stime.
commit 7e817114 made it possible to
fix this by treating x32's syscalls as time64 versions.

in the non-time64-syscall case, make the syscall with the rusage
destination pointer adjusted so that all members but the timevals line
up between the libc and kernel structures. on 64-bit archs, or present
32-bit archs with 32-bit time_t, the timevals will line up too and no
further work is needed. for future 32-bit archs with 64-bit time_t,
the timevals are copied into place, contingent on time_t being larger
than long.

this is analogous to commit 40aa18d5.
so far, there are not any actual time64 versions of the rusage
syscalls (getrusage and wait4) and might never be. however, the
existing x32 ones behave the way time64 versions would if they
existed: using 64-bit slots in place of all longs.

presently, wait4 and getrusage are broken on x32, storing the timevals
correctly but messing up everything else due to the long/kernel-long
mismatch. this would be a huge buffer overflow if not for the 16
reserved slots we left long ago, which suffice to prevent 14
double-sized longs from overflowing into unrelated memory. this commit
will make it possible to fix them.

this is to match the kernel and glibc interfaces. here, struct pt_regs
is an incomplete type, but that's harmless, and if it's completed by
inclusion of another header then members of the struct pointed to by
the regs member can be accessed directly without going through a cast
or intermediate pointer object.

the userspace ucontext API has this as an array rather than a
structure.

commit 3c59a868 fixed the
corresponding mistake for vrregset_t, namely that the original
powerpc64 port used a mix of types from 32-bit powerpc and powerpc64
rather than matching the 64-bit types.

aside from the special value EOF, ungetc is specified to accept and
convert values outside the range of unsigned char. conversion takes
place automatically as part of assignment when storing into the
buffer, but the return value is also required to be the resulting
converted value, and this requirement was not satisfied.

simplified from patch by Wang Jianjian.

based on patch by Dan Gohman, who caught this via compiler warnings.
analysis by Szabolcs Nagy determined that it's a bug, whereby errno
can be set incorrectly for values where the coercion from long double
to double causes rounding. it seems likely that floating point status
flags may be set incorrectly as a result too.

at the same time, clean up use of preprocessor concatenation involving
LDBL_MANT_DIG, which spuriously depends on it being a single unadorned
decimal integer literal, and instead use the equivalent formulation
2/LDBL_EPSILON. an equivalent change on the printf side was made in
commit bff6095d.

policy has long been that these definitions are purely a function of
whether long/pointer is 32- or 64-bit, and that they are not allowed
to vary per-arch. move the definition to the shared alltypes.h.in
fragment, using integer constant expressions in terms of sizeof to
vary the array dimensions appropriately. I'm not sure whether this is
more or less ugly than using preprocessor conditionals and two sets of
definitions here, but either way is a lot less ugly than repeating the
same thing for every arch.

LLONG_MAX is uniform for all archs we support and plenty of header and
code level logic assumes it is, so it does not make sense for limits.h
bits mechanism to pretend it's variable.

LONG_BIT can be defined in terms of LONG_MAX; there's no reason to put
it in bits.

by moving LONG_MAX definition to __LONG_MAX in alltypes.h and moving
LLONG_MAX out of bits, there are now no plain-C limits that are
defined in the bits header, so the bits header only needs to be
included in the POSIX or extended profiles. this allows the feature
test macro logic to be removed from the bits header, facilitating a
long-term goal of getting such logic out of bits.

having __LONG_MAX in alltypes.h will allow further generalization of
headers.

archs without a constant PAGESIZE no longer need bits/limits.h at all.

the resolution of Austin Group issue #162 adds endian.h as a standard
header for future versions of the standard, making it no longer
acceptable for some of the functionality to be hidden behind
_BSD_SOURCE or _GNU_SOURCE. the definitions of the [lb]etoh{16,32,64}
function-like macros are kept conditional since they are alternate
names which the standard did not adopt.

building on commit 97d35a55,
__BYTE_ORDER is now available wherever alltypes.h is included. since
reloc.h is only used from src/internal/dynlink.h, it can be assumed
that __BYTE_ORDER is exposed. reloc.h is not permitted to be included
in other contexts, and generally, like most arch headers, lacks
inclusion guards that would allow such usage. the mips64 version
mistakenly included such guards; they are removed for consistency.

building on commit 97d35a55,
__BYTE_ORDER is now available wherever alltypes.h is included.
endian.h should not be used since, in the future, it will expose
identifiers that are not in the reserved namespace for the headers
which were previously using it.

this change is motivated by the intersection of several factors.
presently, despite being a nonstandard header, endian.h is exposing
the unprefixed byte order macros and functions only if _BSD_SOURCE or
_GNU_SOURCE is defined. this is to accommodate use of endian.h from
other headers, including bits headers, which need to define structure
layout in terms of endianness. with time64 switch-over, even more
headers will need to do this.

at the same time, the resolution of Austin Group issue 162 makes
endian.h a standard header for POSIX-future, requiring that it expose
the unprefixed macros and the functions even in standards-conforming
profiles. changes to meet this new requirement would break existing
internal usage of endian.h by causing it to violate namespace where
it's used.

instead, have the arch's alltypes.h define __BYTE_ORDER, either as a
fixed constant or depending on the right arch-specific predefined
macros for determining endianness. explicit literals 1234 and 4321 are
used instead of __LITTLE_ENDIAN and __BIG_ENDIAN so that there's no
danger of getting the wrong result if a macro is undefined and
implicitly evaluates to 0 at the preprocessor level.

the powerpc (32-bit) bits/endian.h being removed had logic for varying
endianness, but our powerpc arch has never supported that and has
always been big-endian-only. this logic is not carried over to the new
__BYTE_ORDER definition in alltypes.h.

now that commit f7f10797 removed the
legacy i386 conditional definition, va_list is in no way
arch-specific, and has no reason to be in the future. move it to the
shared part of alltypes.h.in

commit ffaaa6d2 removed the
corresponding stdarg.h support for compilers without va_list builtins,
but failed to remove the alternate type definition, leaving incorrect
va_list definitions in place with compilers that don't define __GNUC__
with a value >= 3.

SQRT.fmt exists on MIPS II+ (float), MIPS III+ (double).

ABS.fmt exists on MIPS I+ but only cores with ABS2008 flag in FCSR
implement the required behaviour.

Both sqrt and sqrtf shifted the signed exponent as signed int to adjust
the bit representation of the result. There are signed right shifts too
in the code but those are implementation defined and are expected to
compile to arithmetic shift on supported compilers and targets.

mbsrtowcs contains "vectorized" loops to quickly step over bytes
without the high bit set; these have undefined behavior by virtue of
aliasing uint32_t over top of char data for the accesses.

commit 4d0a8217 fixed the
corresponding usage in string functions by using the may_alias
attribute conditional on __GNUC__ and disabled the vectorized code in
its absence. do the same for mbsrtowcs.

Several math functions are now from the ARM optimized-routines repo
licensed under standard MIT terms and copyrighted by Arm Limited,
so mention this in the COPYRIGHT too.

commit ab3eb89a removed it as part of
correcting the mcontext_t definition, but there is still code using
struct sigcontext and expecting the member names present in it, most
notably libgcc_eh. almost all such usage is incorrect, but bring back
struct sigcontext at least for now so as not to introduce regressions.

in order for sys/procfs.h (provided by sys/user.h) to be useful, it
needs to match the API its consumers (gdb, etc.) expect, including the
member names established by glibc.

this partly reverts commit 29e8737f,
which partly reverted d493206d,
eliminating struct user_fpregs_struct which seems to have had no
precedent and using union __riscv_mc_fp_state for elf_fpregset_t. this
requires indirect inclusion of signal.h to make union
__riscv_mc_fp_state visible, but being that these are nonstandard
"junk" headers with no official restrictions on what they can pull in,
that's no big deal.

split off and expanded from patch by Khem Raj.

the top-level mcontext_t member names were namespace-violating in
standards profiles before, and nested-level member names (some of them
single-letter) were egregiously bad namespace impositions even in
non-strict profiles. moreover, they mismatched those used in the
public API first defined in glibc, breaking any code making use of
them.

unlike most archs, the public API used in glibc for riscv mcontext_t
members was designed to be namespace-safe, so we can and should expose
the members regardless of feature test macros. only the typedefs for
greg_t, gregset_t, and fpregset_t need to be protected behind FTMs.

the struct tags for mcontext_t and ucontext_t are also changed. for
mcontext_t this is necessary to make the common definition across
profiles namespace-safe. for ucontext_t, it's just a matter of
matching the tag from the glibc-defined API.

these changes are split off and expanded from a patch by Khem Raj.

Some declarations of __tls_get_new were left in the code, even
though the definition got removed in

  commit 9d44b646
  install dynamic tls synchronously at dlopen, streamline access

this can make the build fail with

  ld: lib/libc.so: hidden symbol `__tls_get_new' isn't defined

when libc.so is linked without --gc-sections, because a .hidden
declaration in asm code creates a reference even if the symbol
is not actually used.

lrint in (LONG_MAX, 1/DBL_EPSILON) and in (-1/DBL_EPSILON, LONG_MIN)
is not trivial: rounding to int may be inexact, but the conversion to
int may overflow and then the inexact flag must not be raised. (the
overflow threshold is rounding mode dependent).

this matters on 32bit targets (without single instruction lrint or
rint), so the common case (when there is no overflow) is optimized by
inlining the lrint logic, otherwise the old code is kept as a fallback.

on my laptop an i486 lrint call is asm:10ns, old c:30ns, new c:21ns
on a smaller arm core: old c:71ns, new c:34ns
on a bigger arm core: old c:27ns, new c:19ns

analogous to commit ddc7c4f9 for
mips64 and n32, remove the hack to load the syscall number into $2 via
asm, and use a constraint to let the compiler load it instead.

now, only $4, $5, and $6 are potential input-only registers. $2 is
always input and output, and $7 is both when it's an argument,
otherwise output-only. previously, $7 was treated as an input (with a
"1" constraint matching its output position) even when it was not an
input, which was arguably undefined behavior (asm input from
indeterminate value). this is corrected.

as before, $8, $9, and $10 are conditionally input-output registers
for 5-, 6-, and 7-argument syscalls. their role in input is carrying
in the values that will be stored on the stack for arguments 5-7.
their role in output is carrying back whatever the kernel has
clobbered them with, so that the compiler cannot assume they still
contain the input values.

mips32 has two fpu register file variants: FR=0 with 32 32-bit
registers, where pairs of neighboring even/odd registers are used to
represent doubles, and FR=1 with 32 64-bit registers, each of which
can store a single or double.

up through r5 (our "mips" arch), the supported ABI uses FR=0, but
modern compilers generate "fpxx" model code that can safely operate
with either model. r6, which is an incompatible but similar ISA, drops
FR=0 and only provides the FR=1 model. as such, setjmp and longjmp,
which depended on being able to save and restore call-saved doubles by
storing and loading their 32-bit halves, were completely broken in the
presence of floating point code on mips r6.

to fix this, use the s.d and l.d mnemonics to store and load fpu
registers. these expand to the existing swc1 and lwc1 instructions for
pairs of 32-bit fpu registers on mips1, but on mips2 and later they
translate directly to the 64-bit sdc1 and ldc1.

with FR=0, sdc1 and ldc1 behave just like the pairs of swc1 and lwc1
instructions they replace, storing or loading the even/odd pair of fpu
registers that can be treated as separate single-precision floats or
as a unit representing a double. but with FR=1, they store/load
individual 64-bit registers. this yields the ABI-correct behavior on
mips r6, and should make linking of pre-r6 (plain "mips") code with
"fp64" model code workable, although this is and will likely remain
unsupported usage.

in addition to the mips r6 problem this change fixes, reportedly
clang's internal assembler refuses to assemble swc1 and lwc1
instructions for odd register indices when building for "fpxx" model
(the default). this caused setjmp and longjmp not to build. by using
the s.d and l.d forms, this problem is avoided too.

as a bonus, code size is reduced everywhere but mips1.

mips r6 (an incompatible isa from traditional mips) removes the hi and
lo registers used for mul/div results. older gcc versions accepted
them in the clobber list for asm, but their presence is incorrect and
breaks on later versions.

in the process of fixing this, the clobber list for 32-bit mips
syscalls has been deduplicated via a macro like on mips64 and n32.

armv8 removed the coprocessor instructions other than cp14, so
on an armv8 system the related hwcaps should never be set.

new llvm complains about the use of coprocessor instructions in
armv8-a mode (even though they are never executed at runtime),
so ifdef them out when musl is built for armv8.

in the timer thread start function, self->timer_id was accessed
without synchronization; the timer thread could fail to see the store
from the calling thread, resulting in timer_delete failing to delete
the correct kernel-level timer.

this fix is based on a patch by changdiankang, but with the load moved
to after receiving the timer_delete signal rather than just after the
start barrier, so as not to retain the possibility of data race with
timer_delete.

The operand sepcifiers in a_cas and a_cas_p for riscv64 were incorrect:
there's a backwards branch in the routine, so despite tmp being written
at the end of the assembly fragment it cannot be allocated in one of the
input registers because the input values may be needed for another trip
around the loop.

For code that follows the guaranteed forward progress requirements, the
backwards branch is rarely taken: SiFive's hardware only fails a store
conditional on execptional cases (ie, instruction cache misses inside
the loop), and until recently a bug in QEMU allowed back-to-back
store conditionals to succeed.  The bug has been fixed in the latest
QEMU release, but it turns out that the fix caused this latent bug in
musl to manifest.

commit 8a544ee3 introduced a
dependency of the failure path for explicit scheduling at thread
creation on __clone's handling of the start function returning, which
should result in SYS_exit.

as noted in commit 05870abe, the arm
version of __clone was broken in this case. in the past, the mips
version was also broken; it was fixed in commit
8b2b61e0.

since this code path is pretty much entirely untested (previously only
reachable in applications that call the public clone() and return from
the start function) and consists of fragile per-arch asm, don't assume
it works, at least not until it's been thoroughly tested. instead make
the SYS_exit syscall from the start function's failure path.