The findidx functions used in implementing strcoll / wcscoll already
use DIAG_IGNORE_Os_NEEDS_COMMENT for spurious -Wmaybe-uninitialized
warnings that appear with -Os. In building with GCC 7 for x86_64 with
-Os, I find there are additional such warnings, for the same structure
elements, which are spurious for the same reasons given in the
existing comments (and this was also reported for MIPS with GCC 5 in
bug 21313). This patch adds corresponding uses of DIAG_* in the
places that get the additional warnings.
Tested for x86_64 with -Os that this eliminates those warnings and so
allows the build to progress further.
[BZ #21313]
* locale/weight.h (findidx): Disable -Wmaybe-uninitialized for -Os
in another place.
* locale/weightwc.h (findidx): Likewise.
Remove the slow paths from log. Like several other double precision math
functions, log is exactly rounded. This is not required from math functions
and causes major overheads as it requires multiple fallbacks using higher
precision arithmetic if a result is close to 0.5ULP. Ridiculous slowdowns
of up to 100000x have been reported when the highest precision path triggers.
Interestingly removing the slow paths makes hardly any difference in practice:
the worst case error is still ~0.502ULP, and exp(log(x)) shows identical results
before/after on many millions of random cases. All GLIBC math tests pass on
AArch64 and x64 with no change in ULP error. A simple test over a few hundred
million values shows log is now 18% faster on average.
* manual/probes.texi (slowlog): Delete documentation of removed probe.
(slowlog_inexact): Likewise
* sysdeps/ieee754/dbl-64/e_log.c (__ieee754_log): Remove slow paths.
* sysdeps/ieee754/dbl-64/ulog.h: Remove unused declarations.
We have a general principle of preferring optimizations for library
facilities to use compiler built-in functions rather than being
located in library headers, where the compiler can reasonably optimize
code without needing to know glibc implementation details.
This patch applies this principle to bits/byteswap.h, eliminating all
the architecture-specific variants and bits/byteswap-16.h. The
__bswap_16, __bswap_32 and __bswap_64 interfaces all become inline
functions, never macros, using the GCC built-in functions where
available and otherwise a single architecture-independent definition
using shifts and masking (which compilers may well be able to detect
and optimize; GCC has detection of various byte-swapping idioms).
The __bswap_constant_32 macro needs to stay around because of uses in
static initializers within glibc and its tests, and so for consistency
all __bswap_constant_* are kept rather than just being inlined into
the old-GCC-or-non-GCC parts of the __bswap_* inline function
definitions.
Various open bugs are addressed by this cleanup, with caveats about
exactly what is covered by those bugs and when the bugs applied at
all.
Bug 14508 reports -Wformat warnings building glibc because __bswap_*
sometimes returned the wrong types. Obviously we already don't have
such warnings any more or the build would be failing, given -Werror,
and I suspect that bug was originally for wrong types for x86_64, as
fixed by commit d394eb742a (glibc 2.17).
The only case I saw removed by this patch where the types would still
have been wrong was the non-__GNUC__ case of __bswap_64 in the s390
header (using unsigned long long int, but uint64_t would be unsigned
long int for 64-bit). In any case, the single header consistently
uses __uintN_t types after this patch, thereby eliminating all such
bugs. The existing string/test-endian-types.c test already suffices
to verify that the types are correct with the compiler used to build
glibc and its tests.
Bug 15512 reports an error from __bswap_constant_16 with -Werror
-Wsign-conversion. I am unable to reproduce this with any GCC version
supporting -Wsign-conversion - all seem to be able to avoid warning
for ((x) >> 8) & 0xffu, where x is uint16_t, which while it formally
does involve an implicit conversion from int to unsigned int, is also
a case where it should be easy for the compiler to see that the value
converted is never negative. But in this patch __bswap_constant_16 is
changed to use signed 0xff so that no such implicit conversion occurs
at all, and a test with -Werror -Wsign-conversion is added.
Bug 17082 objects to the use of ({}) statement expressions in these
macros preventing use at file scope (in C, that's in sizeof etc.; in
C++, more generally in static initializers). The particular case of
these interfaces is fixed by this patch as it changes them to inline
functions, eliminating all uses of ({}) in bits/byteswap.h, and a
corresponding testcase is added. The bug tries to raise a more
general policy question about use of ({}) in macros in installed
headers, referring to "many other libc functions" (unspecified which
functions are being considered).
Since such policy questions belong on libc-alpha, and since there
*are* macros in installed headers which can't really avoid using ({})
(where they are type-generic, so can't use an inline function, but
need a temporary variable, and a few where the interface involves
returning memory from alloca so can't use an inline function either),
I propose to consider that bug fixed with this change. That is
without prejudice to any other new bugs anyone wishes to file *for
precisely defined sets of macros* requesting moving away from ({})
*where it is clearly possible for those interfaces*. Where ({}) can
be avoided, typically by use of an inline function, I think that's a
good idea - that inline functions are typically to be preferred to
({}) for header interfaces where such optimizations are useful but the
interface is suited to being defined using an inline function.
Bug 20530 requests use of __builtin_bswap16 when available (GCC 4.8
and later), which this patch implements.
Tested for x86_64, and with build-many-glibcs.py. Also did an x86_64
test with the __GNUC_PREREQ conditionals changed to "#if 0" to verify
the old-GCC/non-GCC case in the headers. (There are already existing
tests for correctness of results of these interfaces.)
[BZ #14508]
[BZ #15512]
[BZ #17082]
[BZ #20530]
* bits/byteswap.h: Update file comment. Do not include
<bits/byteswap-16.h>.
(__bswap_constant_16): Cast result to __uint16_t. Use signed 0xff
constant.
(__bswap_16): Define as inline function.
(__bswap_constant_32): Reformat definition.
(__bswap_32): Always define as inline function, not macro, using
__uint32_t. Use __builtin_bswap32 if [__GNUC_PREREQ (4, 3)],
otherwise __bswap_constant_32.
(__bswap_constant_64): Reformat definition. Do not use
__extension__ here.
(__bswap_64): Always define as inline function, not macro. Use
__extension__ on function definition. Use __builtin_bswap64 if
[__GNUC_PREREQ (4, 3)], otherwise __bswap_constant_64.
* string/test-endian-file-scope.c: New file.
* string/test-endian-sign-conversion.c: Likewise.
* string/Makefile (headers): Remove bits/byteswap-16.h.
(tests): Add test-endian-file-scope and
test-endian-sign-conversion.
(CFLAGS-test-endian-sign-conversion.c): New variable.
* bits/byteswap-16.h: Remove file.
* sysdeps/ia64/bits/byteswap-16.h: Likewise.
* sysdeps/ia64/bits/byteswap.h: Likewise.
* sysdeps/m68k/bits/byteswap.h: Likewise.
* sysdeps/s390/bits/byteswap-16.h: Likewise.
* sysdeps/s390/bits/byteswap.h: Likewise.
* sysdeps/tile/bits/byteswap.h: Likewise.
* sysdeps/x86/bits/byteswap-16.h: Likewise.
* sysdeps/x86/bits/byteswap.h: Likewise.
Bug 17721 reports that the non-__GNUC__ definitions of __inline and
__restrict are suboptimal, in that they are defined to empty when they
could be defined to inline and restrict for appropriate language
versions. This patch makes those fixes.
Tested for x86_64 (however, I have not done any testing with an actual
non-__GNUC__ compiler and it's likely such compilers may have other
problems with glibc headers).
[BZ #17721]
* misc/sys/cdefs.h [!__GNUC__ && (__cplusplus || (__STDC_VERSION__
&& __STDC_VERSION__ >= 199901L))] (__inline): Define to inline.
[!__GNUC_PREREQ (2,92) && __STDC_VERSION__ && __STDC_VERSION__ >=
199901L] (__restrict): Define to restrict.
Bug 19667 reports unchecked malloc calls in the test
string/testcopy.c. This patch makes that test use xmalloc and the
support/test-driver.c test framework.
Tested for x86_64.
[BZ #19667]
* string/testcopy.c: Include <support/support.h>. Do not include
<malloc.h>. Use <support/test-driver.c>.
(main): Rename to do_test. Make static. Use xmalloc instead of
malloc.
Bug 13575 reports that SSIZE_MAX is wrongly defined as LONG_MAX on
32-bit systems where ssize_t is defined as int (which is most 32-bit
systems supported by glibc).
This patch fixes the definition, using a conditional on
__WORDSIZE32_SIZE_ULONG to determine the appropriate type in the
32-bit case. Formally ssize_t need not be the signed type
corresponding to size_t, but as it is for all current glibc
configurations, there is no need for a new macro different from the
one used for defining SIZE_MAX. A testcase is added for both the type
and the value of SSIZE_MAX.
There is a relevant peculiarity in
sysdeps/unix/sysv/linux/s390/bits/typesizes.h:
/* Compatibility with g++ 2.95.x. */
/* size_t is unsigned long int on s390 -m31. */
This has the effect that for GCC 2 for s390, ssize_t does not match
__WORDSIZE32_SIZE_ULONG. I don't think such a conditional on the GCC
version makes sense - to have a well-defined ABI, the choices of
standard types should not depend on the GCC version. It's also the
case that upstream GCC 2.95 did not support s390, and glibc headers
don't in general try to support past development GCC versions - only
actual releases and current mainline development. But whether or not
that GCC 2 case should be removed (with or without a NEWS entry for
such a change), this patch does not result in any changes for s390;
the value is always still LONG_MAX in the s390 case because
__WORDSIZE32_SIZE_ULONG is always defined for 32-bit s390. I don't
think any such oddity in code only active for unofficial or unreleased
old compiler versions should block closing the present bug as fixed
once this patch is in.
Tested for x86_64 and x86, and with build-many-glibcs.py.
[BZ #13575]
* posix/bits/posix1_lim.h: Include <bits/wordsize.h>.
[!SSIZE_MAX && !(__WORDSIZE == 64 || __WORDSIZE32_SIZE_ULONG)]
(SSIZE_MAX): Define to INT_MAX.
* posix/test-ssize-max.c: New file.
* posix/Makefile (tests): Add test-ssize-max.
Bug 19668 reports an unchecked malloc call in the test
sysdeps/powerpc/fpu/tst-setcontext-fpscr.c. This patch makes that
test use xmalloc. It does not otherwise move this test to the
support/ infrastructure or support/test-driver.c; the test has various
uses of exit and _exit on error cases, and uses atexit, and while I
think those things would all still work in the context of
test-driver.c, it's not an immediately obvious conversion the way it
would be for many tests that don't use test-driver.c.
Tested for powerpc.
[BZ #19668]
* sysdeps/powerpc/fpu/tst-setcontext-fpscr.c: Include
<support/support.h>. Do not include <malloc.h>.
(query_auxv): Use xmalloc instead of malloc.
Bug 14553 reports that sys/types.h defines loff_t unconditionally,
despite it not being part of any supported standard. This is
permitted by the POSIX *_t reservation, but as a
quality-of-implementation issue it's still best not to define it
except for __USE_MISC. This patch conditions the definition
accordingly, updating a macro in sysdeps/unix/sysv/linux/sys/quota.h
to use __loff_t so it still works even if __USE_MISC is not defined.
codesearch.debian.net suggests there are quite a lot of loff_t uses
outside glibc, but it might well make sense to change all (few) uses
of loff_t or __loff_t inside glibc to use off64_t or __off64_t
instead, leaving only the definitions, treating this name as
obsolescent.
Tested for x86_64.
[BZ #14553]
* posix/sys/types.h (loff_t): Only define for [__USE_MISC].
* sysdeps/unix/sysv/linux/sys/quota.h (dqoff): Use __loff_t
instead of loff_t.
The getc and putc macros in the public stdio.h expand to call _IO_getc
and _IO_putc respectively. As _IO_getc, fgetc, and getc are all aliases
for the same function, and _IO_putc, fputc, and putc are also all aliases
for the same function, the macros are pointless. The C standard does
not require getc and putc to be macros, so let's just not have macros.
All four symbols are exported from libc.so at the same, ancient symbol
version, so there should be no risks for binary compatibility. Similarly,
the getchar and putchar inlines in bits/stdio.h forward to getc and putc
instead of their _IO_ aliases.
As a change from longstanding historical practice, this does seem
like it might break _something_, so there is a note in NEWS, which
is also a convenient place to advise people that if they thought getc
and putc had reduced per-character overhead they should consider using
getc_unlocked and putc_unlocked instead. (These are also not macros,
but when optimizing, they are inlines.)
* libio/stdio.h: Don't define getc or putc as macros.
* libio/bits/stdio.h (getchar, putchar): Use getc and putc,
not _IO_getc and _IO_putc.
Two files in stdio-common were unnecessarily redefining some standard
symbols as their _IO_ aliases.
* stdio-common/vfprintf.c: Don't redefine FILE, va_list, or BUFSIZ.
* stdio-common/tstgetln.c: Don't redefine ssize_t.
This patch adds the TCP_FASTOPEN_KEY and TCP_FASTOPEN_NO_COOKIE macros
from Linux 4.15 to sysdeps/gnu/netinet/tcp.h.
Tested for x86_64.
* sysdeps/gnu/netinet/tcp.h (TCP_FASTOPEN_KEY): New macro.
(TCP_FASTOPEN_NO_COOKIE): Likewise.
This patch adds the IPV6_FREEBIND macro from Linux 4.15 to
sysdeps/unix/sysv/linux/bits/in.h.
Tested for x86_64.
* sysdeps/unix/sysv/linux/bits/in.h (IPV6_FREEBIND): New macro.
Linux 4.15 adds NT_S390_RI_CB to linux/elf.h (and NT_ARM_SVE, which we
already have in glibc). This shows up that various other ELF note
values from linux/elf.h are missing from glibc's elf.h.
This patch adds the missing values that are relevant to glibc
architectures. As elf.h is a general description of the ELF format,
not necessarily limited to glibc configurations, there's an argument
for having the remaining NT_* values that Linux uses for non-glibc
architectures in glibc's elf.h as well, but this patch does not add
them.
Adding the NT_PRFPREG name is bug 14890. That bug also requests
making the NT_FPREGSET name obsolete. Given that elf.h is not just
for Linux but can describe ELF for other operating systems, I don't
think that a change of name in the Linux kernel is sufficient
justification for declaring the other name obsolete; there can be
multiple names for the same note value, even with incompatible
semantics, if those reflect variants of the ELF format in actual use.
For example, FreeBSD appears still to have the name NT_FPREGSET
<https://github.com/freebsd/freebsd/blob/master/sys/sys/elf_common.h>
(note: I haven't checked whether the FreeBSD kernel actually generates
such notes or whether this is actually an other-OS definition present
in FreeBSD's header).
[BZ #14890]
* elf/elf.h (NT_PRFPREG): New macro.
(NT_S390_VXRS_LOW): Likewise.
(NT_S390_VXRS_HIGH): Likewise.
(NT_S390_GS_CB): Likewise.
(NT_S390_GS_BC): Likewise.
(NT_S390_RI_CB): Likewise.
This patch adds the MAP_SYNC macro from Linux 4.15 to various
bits/mman.h headers. Note that this is *not* added to all
architectures: in Linux 4.15, this macro is only in
asm-generic/mman.h, and only some architectures' asm/mman.h include
the asm-generic file - the architectures not using the asm-generic
file will need their own values of MAP_SYNC allocated to support this
functionality (some of them also already have conflicting mmap flags
so the value there will have to be different from the generic
0x80000). Specifically, for glibc architectures, alpha hppa mips
powerpc sparc tile lack allocations of values for MAP_SYNC.
Tested for x86_64.
* sysdeps/unix/sysv/linux/aarch64/bits/mman.h [__USE_MISC]
(MAP_SYNC): New macro.
* sysdeps/unix/sysv/linux/arm/bits/mman.h [__USE_MISC] (MAP_SYNC):
Likewise.
* sysdeps/unix/sysv/linux/ia64/bits/mman.h [__USE_MISC]
(MAP_SYNC): Likewise.
* sysdeps/unix/sysv/linux/m68k/bits/mman.h [__USE_MISC]
(MAP_SYNC): Likewise.
* sysdeps/unix/sysv/linux/microblaze/bits/mman.h [__USE_MISC]
(MAP_SYNC): Likewise.
* sysdeps/unix/sysv/linux/nios2/bits/mman.h [__USE_MISC]
(MAP_SYNC): Likewise.
* sysdeps/unix/sysv/linux/riscv/bits/mman.h [__USE_MISC]
(MAP_SYNC): Likewise.
* sysdeps/unix/sysv/linux/s390/bits/mman.h [__USE_MISC]
(MAP_SYNC): Likewise.
* sysdeps/unix/sysv/linux/sh/bits/mman.h [__USE_MISC] (MAP_SYNC):
Likewise.
* sysdeps/unix/sysv/linux/x86/bits/mman.h [__USE_MISC] (MAP_SYNC):
Likewise.
This patch adds the MAP_SHARED_VALIDATE macro from Linux 4.15 to
bits/mman-linux.h and the hppa bits/mman.h.
Tested for x86_64.
* sysdeps/unix/sysv/linux/bits/mman-linux.h [__USE_MISC]
(MAP_SHARED_VALIDATE): New macro.
* sysdeps/unix/sysv/linux/hppa/bits/mman.h [__USE_MISC]
(MAP_SHARED_VALIDATE): Likewise.
The only differences in ld.so are line numbers for asserts.
Reviewed-by: Jonathan Nieder <jrnieder@gmail.com>
* elf/dl-addr.c (determine_info): Use ADDRIDX with DT_GNU_HASH.
* elf/dl-lookup.c (_dl_setup_hash): Likewise.
* elf/get-dynamic-info.h (elf_get_dynamic_info): Likewise.
Since start.o may be compiled as PIC, we should check PIC instead of
SHARED.
[BZ #22638]
* sysdeps/sparc/sparc32/start.S (_start): Check PIC instead of
SHARED.
* sysdeps/sparc/sparc64/start.S (_start): Likewise.
Keeping the buffers the same across test runs gives later invocations
the advantage since they access cached data. Reallocate so that all
test runs are on equal grounds.
* benchtests/bench-memcmp.c (do_test): Call realloc_buf for
every test run.
This patch updates sysdeps/unix/sysv/linux/syscall-names.list for
Linux 4.15. There only appears to be one new syscall to add to the
list. (The riscv_flush_icache syscall is *not* added because for
whatever reason it doesn't appear in the uapi asm/unistd.h; only in
arch/riscv/include/uapi/asm/syscalls.h, which is only included by the
non-uapi asm/unistd.h - and only syscalls whose __NR_* macros are
defined in the uapi asm/unistd.h are relevant for this list.)
Tested for x86_64, and with build-many-glibcs.py.
* sysdeps/unix/sysv/linux/syscall-names.list: Update kernel
version to 4.15.
(s390_sthyi): New syscall.
The general rule in glibc is that it's better for a macro to be always
defined, and tested with #if, than for it to be tested with #ifdef,
because the latter is prone to typos in the macro name as well as to
the header with the macro accidentally not being included in a file
testing it. (Testing with an "if" statement is even better, in those
cases where it's possible to do things that way, as it then means both
cases in the code get checked for syntax in glibc builds with either
value of the condition.)
math_private.h has several different groups of macros, meaning that
architectures wanting to override some of them need to define those
then include the generic version, which then defines macros if not
already defined. It's hard to avoid that arrangement completely, but
various cases can be improved by splitting out macros or groups of
macros into separate files.
This patch splits out the LDBL_CLASSIFY_COMPAT macro into a separate
ldbl-classify-compat.h header. This macro is tested with #ifdef; this
patch changes it to testing with #if, with a default definition to 0
in the generic header and then architecture-specific headers defining
it to 1.
Tested with build-many-glibcs.py that installed stripped shared
libraries are unchanged by the patch.
* sysdeps/generic/ldbl-classify-compat.h: New file.
* sysdeps/arm/ldbl-classify-compat.h: Likewise.
* sysdeps/m68k/coldfire/ldbl-classify-compat.h: Likewise.
* sysdeps/microblaze/ldbl-classify-compat.h: Likewise.
* sysdeps/mips/ldbl-classify-compat.h: Likewise.
* sysdeps/nios2/ldbl-classify-compat.h: Likewise.
* sysdeps/sh/ldbl-classify-compat.h: Likewise.
* sysdeps/ieee754/dbl-64/s_finite.c: Include
<ldbl-classify-compat.h>.
[LDBL_CLASSIFY_COMPAT]: Test value, not whether defined.
* sysdeps/ieee754/dbl-64/s_isinf.c: Include
<ldbl-classify-compat.h>.
[LDBL_CLASSIFY_COMPAT]: Test value, not whether defined.
* sysdeps/ieee754/dbl-64/s_isnan.c: Include
<ldbl-classify-compat.h>.
[LDBL_CLASSIFY_COMPAT]: Test value, not whether defined.
* sysdeps/ieee754/dbl-64/wordsize-64/s_finite.c: Include
<ldbl-classify-compat.h>.
[LDBL_CLASSIFY_COMPAT]: Test value, not whether defined.
* sysdeps/ieee754/dbl-64/wordsize-64/s_isinf.c: Include
<ldbl-classify-compat.h>.
[LDBL_CLASSIFY_COMPAT]: Test value, not whether defined.
* sysdeps/ieee754/dbl-64/wordsize-64/s_isnan.c: Include
<ldbl-classify-compat.h>.
[LDBL_CLASSIFY_COMPAT]: Test value, not whether defined.
* sysdeps/arm/math_private.h (LDBL_CLASSIFY_COMPAT): Remove macro.
* sysdeps/mips/math_private.h (LDBL_CLASSIFY_COMPAT): Likewise.
* sysdeps/m68k/coldfire/math_private.h: Remove file.
* sysdeps/microblaze/math_private.h: Likewise.
* sysdeps/nios2/math_private.h: Likewise.
* sysdeps/sh/math_private.h: Likewise.
math_private.h headers for configurations lacking support for
floating-point exceptions and rounding modes define
libc_feholdexcept_setround to override the default version with one
that discards its rounding mode argument.
Unlike other such libc_fe* macros that I removed, this one is actually
used for such configurations (in dbl-64/e_sqrt.c). However, this does
not make the macro required. It's only used for such configurations
with FE_TONEAREST as the rounding mode (anything needing another mode
should not be used when that mode is unavailable), and the default
definition just calls __feholdexcept and __fesetround. Since we now
have suitable inline do-nothing definitions of __feholdexcept and
__fesetround for the cases of no exceptions and rounding modes, we can
just rely on those inlines to achieve the same optimization as this
macro definition. Thus, this patch removes those macro definitions
(and the math_private.h headers containing them, when no longer needed
after that removal).
Tested with build-many-glibcs.py that installed stripped shared
libraries are unchanged by the patch.
* sysdeps/m68k/coldfire/fpu/math_private.h: Move to ....
* sysdeps/m68k/coldfire/math_private.h: ... here.
* sysdeps/m68k/coldfire/nofpu/math_private.h: Remove file.
* sysdeps/tile/math_private.h: Likewise.
* sysdeps/microblaze/math_private.h (libc_feholdexcept_setround):
Remove macro.
* sysdeps/nios2/math_private.h (libc_feholdexcept_setround):
Likewise.
math_private.h headers for configurations lacking support for
floating-point exceptions and rounding modes define various libc_fe*
macros to override the default versions with ones that discard any
exception or rounding mode arguments.
Three of the four macros defined in these headers are no longer needed
there: those macros are only used in fma implementations that are not
used for such configurations, now all those configurations properly
use soft-fp fma implementations instead. (Effectively, those macros
were a workaround to allow glibc to build in the absence of a proper
fma implementation for this case - now there is such an
implementation, there is no need to support building the wrong
implementation for those configurations.) Thus, this patch removes
the unnecessary macros.
Tested with build-many-glibcs.py that installed stripped shared
libraries are unchanged by the patch.
* sysdeps/m68k/coldfire/nofpu/math_private.h (libc_fesetround):
Remove macro.
(libc_fetestexcept): Likewise.
(libc_feupdateenv_test): Likewise.
* sysdeps/microblaze/math_private.h (libc_fesetround): Likewise.
(libc_fetestexcept): Likewise.
(libc_feupdateenv_test): Likewise.
* sysdeps/nios2/math_private.h (libc_fesetround): Likewise.
(libc_fetestexcept): Likewise.
(libc_feupdateenv_test): Likewise.
* sysdeps/tile/math_private.h (libc_fesetround): Likewise.
(libc_fetestexcept): Likewise.
(libc_feupdateenv_test): Likewise.
Continuing the process of improving and cleaning up the handling of
configurations lacking support for floating-point exceptions and
rounding modes, this patch adds trivial inline definitions of
feholdexcept and __feholdexcept to the set of inlines for such
configurations in math_private.h. These inlines were missing from the
tile version used as a basis for the previous inlines, despite a few
such function calls ending up in libm.so.
Tested with build-many-glibcs.py. As expected, installed stripped
shared libraries are unchanged for architectures supporting exceptions
and rounding modes, but changed for architectures lacking such
support.
* sysdeps/generic/math_private.h
[!FE_HAVE_ROUNDING_MODES && FE_ALL_EXCEPT == 0] (feholdexcept):
New inline function.
[!FE_HAVE_ROUNDING_MODES && FE_ALL_EXCEPT == 0] (__feholdexcept):
Likewise.
The tile version of math_private.h defines some inline functions for
fenv.h functions, to optimize away internal calls to these functions
that do nothing given no support for floating-point exceptions and
rounding modes. (Some functions may have error cases for invalid
arguments, but those aren't applicable to the internal calls from
within glibc.) Other configurations lacking support for exceptions
and rounding modes lack such inline functions. This patch moves them
to the generic math_private.h, appropriately conditioned, so that all
such configurations can benefit from the.
include/fenv.h is made to check whether there are any non-default
rounding modes; that needs to be done there, rather than later,
because get-rounding-mode.h defines values for otherwise unsupported
FE_* rounding modes. It also gives an error for FE_TONEAREST
undefined, a case that already did not work for building the glibc
testsuite; the convention has by now been established that all
architectures need to provide a version of bits/fenv.h that at least
defines FE_TONEAREST.
Tested with build-many-glibcs.py. As expected, installed stripped
shared libraries are unchanged for tile and for architectures
supporting exceptions and rounding modes, but changed for non-tile
architectures not supporting exceptions and rounding modes that
previously lacked this optimization (e.g. Nios II libm.so is about 1kB
smaller).
The optimization is not in fact complete (does not cover feholdexcept
/ __feholdexcept, so a few calls to those remain unnecessarily within
libm even after this patch), but that can be dealt with separately.
* include/fenv.h [!_ISOMAC && !FE_TONEAREST]: Give #error.
[!_ISOMAC] (FE_HAVE_ROUNDING_MODES): New macro.
* sysdeps/generic/math_private.h
[!FE_HAVE_ROUNDING_MODES && FE_ALL_EXCEPT == 0] (fegetenv): New
inline function.
[!FE_HAVE_ROUNDING_MODES && FE_ALL_EXCEPT == 0] (__fegetenv):
Likewise.
[!FE_HAVE_ROUNDING_MODES && FE_ALL_EXCEPT == 0] (fesetenv):
Likewise.
[!FE_HAVE_ROUNDING_MODES && FE_ALL_EXCEPT == 0] (__fesetenv):
Likewise.
[!FE_HAVE_ROUNDING_MODES && FE_ALL_EXCEPT == 0] (feupdateenv):
Likewise.
[!FE_HAVE_ROUNDING_MODES && FE_ALL_EXCEPT == 0] (__feupdateenv):
Likewise.
[!FE_HAVE_ROUNDING_MODES] (fegetround): Likewise.
[!FE_HAVE_ROUNDING_MODES] (__fegetround): Likewise.
[!FE_HAVE_ROUNDING_MODES] (fesetround): Likewise.
[!FE_HAVE_ROUNDING_MODES] (__fesetround): Likewise.
* sysdeps/tile/math_private.h (fegetenv): Remove inline function.
(__fegetenv): Likewise.
(fesetenv): Likewise.
(__fesetenv): Likewise.
(feupdateenv): Likewise.
(__feupdateenv): Likewise.
(fegetround): Likewise.
(__fegetround): Likewise.
(fesetround): Likewise.
(__fesetround): Likewise.
Various configurations lacking support for floating-point exceptions
and rounding modes have a math_private.h that overrides certain
functions and macros, internal and external, to avoid references to
FE_* constants that are undefined in those configurations. For
example, there are unconditional feraiseexcept (FE_INVALID) calls in
generic libm code, and these macro definitions duly define
feraiseexcept to ignore its argument to avoid an error from FE_INVALID
being undefined.
In fact it is easy to tell in an architecture-independent way whether
this is needed, by testing whether FE_ALL_EXCEPT == 0. Thus, this
patch puts such a test, and feraiseexcept and __feraiseexcept macros,
in the generic math_private.h, so reducing the duplication between
architecture versions of this header. The feclearexcept macro present
in several versions of this header, and fetestexcept in the tile
version, are not needed; they would have been needed before there were
proper soft-fp fma implementations (when generic versions, that depend
on FE_TOWARDZERO and FE_INEXACT, were being used for configurations
not supporting those features), but aren't needed any more, and so are
removed.
The tile version of this header has several inline functions for
fenv.h functions to optimize calls to them away in such configurations
where they do nothing useful, and all these header versions also have
definitions of some of the libc_fe* internal macros. I intend to make
those generic in subsequent patches.
Tested with build-many-glibcs.py that installed stripped shared
libraries are unchanged by this patch.
* sysdeps/generic/math_private.h [FE_ALL_EXCEPT == 0]
(feraiseexcept): New macro.
[FE_ALL_EXCEPT == 0] (__feraiseexcept): Likewise.
* sysdeps/m68k/coldfire/nofpu/math_private.h (feraiseexcept):
Remove macro.
(__feraiseexcept): Likewise.
(feclearexcept): Likewise.
* sysdeps/microblaze/math_private.h (feraiseexcept): Likewise.
(__feraiseexcept): Likewise.
(feclearexcept): Likewise.
* sysdeps/nios2/math_private.h (feraiseexcept): Likewise.
(__feraiseexcept): Likewise.
(feclearexcept): Likewise.
* sysdeps/tile/math_private.h (feraiseexcept): Likewise.
(__feraiseexcept): Likewise.
(feclearexcept): Likewise.
(fetestexcept): Likewise.
Since I've been fixing build issues for ColdFire, this patch adds a
math-tests.h file for ColdFire, reflecting the lack of support for
exceptions and rounding modes for soft float. I think it is logically
correct, but have not tested it beyond build-many-glibcs.py for both
hard and soft float.
* sysdeps/m68k/coldfire/math-tests.h: New file.
The m68k bits/fenv.h is in sysdeps/m68k/fpu/, meaning that no-FPU
ColdFire instead gets the generic (top-level) bits/fenv.h.
That top-level bits/fenv.h defines no rounding mode constants. That
no longer works for building glibc tests: some tests fail to build (at
least with warnings) if no rounding mode macros are defined, so at
least FE_TONEAREST must be defined in all cases (as various
architectures without rounding mode support indeed do), while
__FE_UNDEFINED must be defined in the case where not all the standard
rounding modes are supported.
On general principles of supporting multilib toolchains with a single
set of headers shared between multilibs for a given architecture, it's
also desirable for the same bits/fenv.h header to work for both FPU
and no-FPU configurations. Thus, this patch moves the m68k
bits/fenv.h to sysdeps/m68k/bits/fenv.h, and inserts appropriate
conditionals to handle the no-FPU case. All the exception macros, and
FE_NOMASK_ENV, are disabled in the no-FPU case; FE_ALL_EXCEPT is
defined to 0 in that case. All rounding modes except FE_TONEAREST are
disabled in that case, and __FE_UNDEFINED is defined accordingly. To
avoid an unnecessary ABI change, fenv_t is defined in the no-FPU case
to match the definition it would have got from the generic
bits/fenv.h.
This suffices to get a clean glibc and testsuite build for this
configuration with build-many-glibcs.py (and keeps a clean build for
the other m68k configurations); it has not been otherwise tested.
* sysdeps/m68k/fpu/bits/fenv.h: Move to ....
* sysdeps/m68k/bits/fenv.h: ... here.
[!__HAVE_68881__ && !__HAVE_FPU__ && !__mcffpu__] (FE_INEXACT): Do
not define.
[!__HAVE_68881__ && !__HAVE_FPU__ && !__mcffpu__] (FE_DIVBYZERO):
Likewise.
[!__HAVE_68881__ && !__HAVE_FPU__ && !__mcffpu__] (FE_UNDERFLOW):
Likewise.
[!__HAVE_68881__ && !__HAVE_FPU__ && !__mcffpu__] (FE_OVERFLOW):
Likewise.
[!__HAVE_68881__ && !__HAVE_FPU__ && !__mcffpu__] (FE_INVALID):
Likewise.
[!__HAVE_68881__ && !__HAVE_FPU__ && !__mcffpu__] (FE_ALL_EXCEPT):
Define to 0.
[!__HAVE_68881__ && !__HAVE_FPU__ && !__mcffpu__]
(__FE_UNDEFINED): New enum constant.
[!__HAVE_68881__ && !__HAVE_FPU__ && !__mcffpu__] (FE_TOWARDZERO):
Do not define.
[!__HAVE_68881__ && !__HAVE_FPU__ && !__mcffpu__] (FE_DOWNWARD):
Likewise.
[!__HAVE_68881__ && !__HAVE_FPU__ && !__mcffpu__] (FE_UPWARD):
Likewise.
[!__HAVE_68881__ && !__HAVE_FPU__ && !__mcffpu__] (fenv_t): Define
to match generic bits/fenv.h.
[!__HAVE_68881__ && !__HAVE_FPU__ && !__mcffpu__] (FE_NOMASK_ENV):
Do not define.
Building for soft-float ColdFire produces an error in soft-fp:
In file included from ../sysdeps/ieee754/soft-fp/s_fmaf.c:42:
../soft-fp/single.h:85:3: error: 'packed' attribute ignored for field of type 'struct <anonymous>' [-Werror=attributes]
} bits __attribute__ ((packed));
^
While this error only appears in that particular case, this attribute
is in fact never useful, on any architecture. If you have
struct __attribute__ ((packed)) { ... } bits;
or
struct { ... } __attribute__ ((packed)) bits;
then the attribute affects the layout of the structure type. But with
the form used in this code
struct { ... } bits __attribute__ ((packed));
the field bits is being declared packed, but the layout of its type
has already been determined at that point. If on any platform the
layout of the sequence of bit-fields were wrong without the use of a
packed attribute, the attribute would need to be used via a definition
of _FP_STRUCT_LAYOUT, not in its present position.
So this patch removes the useless attribute to fix the build for
ColdFire soft-float. Tested with build-many-glibcs.py that installed
stripped shared libraries are unchanged by the patch.
* soft-fp/double.h (union _FP_UNION_D): Do not use attribute
packed on bits.
* soft-fp/extended.h (union _FP_UNION_E): Likewise.
* soft-fp/half.h (union _FP_UNION_H): Likewise.
* soft-fp/quad.h (union _FP_UNION_Q): Likewise.
* soft-fp/single.h (union _FP_UNION_S): Likewise.
As reported in bug 21314, building log1p and log1pf fails with -Os
because of a spurious -Wmaybe-uninitialized warning (reported there
for GCC 5 for MIPS, I see it also with GCC 7 for x86_64). This patch,
based on the patches in the bug, fixes this using the DIAG_* macros.
Tested for x86_64 with -Os that this eliminates those warnings and so
allows the build to progress further.
2018-02-01 Carlos O'Donell <carlos@redhat.com>
Ramin Seyed-Moussavi <lordrasmus@gmail.com>
Joseph Myers <joseph@codesourcery.com>
[BZ #21314]
* sysdeps/ieee754/dbl-64/s_log1p.c: Include <libc-diag.h>.
(__log1p): Disable -Wmaybe-uninitialized for -Os around
computation using c.
* sysdeps/ieee754/flt-32/s_log1pf.c: Include <libc-diag.h>.
(__log1pf): Disable -Wmaybe-uninitialized for -Os around
computation using c.
These tests require a new thread stack size set to a value (0x20000)
lower than the architecture minimum (0x30000). Set the stack size
to PTHREAD_STACK_MIN in this case.
Checked on ia64-linux-gnu.
* stdlib/test-atexit-race-common.c (do_test): Check stack size
against PTHREAD_STACK_MIN.
