I noticed that there was no meaningful test coverage for
fegetexceptflag and fesetexceptflag (one test ensures that calls to
them compile and link, but nothing to verify they work correctly).
This patch adds tests for these functions.
fesetexceptflag is meant to set the relevant exception flag bits to
the saved state without causing enabled traps to be taken. On some
architectures, it is not possible to set exception flag bits without
causing enabled traps to occur. Such architectures need to define
EXCEPTION_SET_FORCES_TRAP to 1 in their math-tests.h, as is done in
this patch for powerpc. x86 avoids needing to define this because the
traps resulting from setting exception bits don't occur until the next
floating-point operation or fwait instruction.
Tested for x86_64, x86 and powerpc. Note that test-fexcept fails for
powerpc because of a pre-existing bug in fesetexceptflag for powerpc,
which I'll fix separately.
* math/test-fexcept-traps.c: New file.
* math/test-fexcept.c: Likewise.
* math/Makefile (tests): Add test-fexcept and test-fexcept-traps.
* sysdeps/generic/math-tests.h (EXCEPTION_SET_FORCES_TRAP): New
macro.
* sysdeps/powerpc/math-tests.h [!__NO_FPRS__]
(EXCEPTION_SET_FORCES_TRAP): Likewise.
As pointer out on the mailing list, the inline assembly code in
sysdeps/powerpc/ifunc-sel.h doesn't have a list of clobbered registers
and used wrong constraints.
This patch fixes that. I verified it doesn't introduce any change in the
generated code.
Changelog:
* sysdeps/powerpc/ifunc-sel.h (ifunc_sel): Add "11", "12", "cr0" to the
clobber list. Use "i" constraint instead of "X".
(ifunc_one): Add "12" to the clobber list. Use "i" constraint instead
of "X".
On 32-bit PowerPC GCC 6 always saves the PIC register on the stack in
the prologue and adjust the stack in the epilogue. It is therefore not
possible anymore to just exit the function in the inline asm code,
otherwise it corrupts the stack pointer. This causes the following tests
to fail when using GCC 6:
FAIL: elf/ifuncmain1
FAIL: elf/ifuncmain1pic
FAIL: elf/ifuncmain1picstatic
FAIL: elf/ifuncmain1pie
FAIL: elf/ifuncmain1staticpic
FAIL: elf/ifuncmain1staticpie
FAIL: elf/ifuncmain1vis
FAIL: elf/ifuncmain1vispic
FAIL: elf/ifuncmain1vispie
FAIL: elf/ifuncmain2pic
FAIL: elf/ifuncmain2picstatic
FAIL: elf/ifuncmain3
FAIL: elf/ifuncmain4picstatic
FAIL: elf/ifuncmain5
FAIL: elf/ifuncmain5picstatic
FAIL: elf/ifuncmain5staticpic
The solution is to replace the beqlr instructions by a beq to the end
of the inline asm code. This fixes all the above failures.
ChangeLog:
* sysdeps/powerpc/ifunc-sel.h (ifunc_sel): Replace beqlr instructions
by beq instructions jumping to the end of the function.
Commit a6a4395d fixed modf implementation by compiling s_modf.c and
s_modff.c with -fsignaling-nans. However these files are also included
from the pre-POWER5+ implementation, and thus these files should also
be compiled with -fsignaling-nans.
Changelog:
[BZ #20240]
* sysdeps/powerpc/powerpc32/power4/fpu/multiarch/Makefile
(CFLAGS-s_modf-ppc32.c): New variable.
(CFLAGS-s_modff-ppc32.c): Likewise.
* sysdeps/powerpc/powerpc64/fpu/multiarch/Makefile
(CFLAGS-s_modf-ppc64.c): Likewise.
(CFLAGS-s_modff-ppc64.c): Likewise.
If the input values are unaligned and if there are null characters in the
memory before the starting address of the input values, strcasecmp
gives incorrect return code. Fixed it by adding mask the bits that
are not part of the string.
This implementation is based on the one already used at
sysdeps/x86_64/fpu/e_expf.S.
This implementation improves the performance by ~14% on average in synthetic
benchmarks at the cost of decreasing accuracy to 1 ULP.
atomic_compare_and_exchange_bool_rel and
catomic_compare_and_exchange_bool_rel are removed and replaced with the
new C11-like atomic_compare_exchange_weak_release. The concurrent code
in nscd/cache.c has not been reviewed yet, so this patch does not add
detailed comments.
* nscd/cache.c (cache_add): Use new C11-like atomic operation instead
of atomic_compare_and_exchange_bool_rel.
* nptl/pthread_mutex_unlock.c (__pthread_mutex_unlock_full): Likewise.
* include/atomic.h (atomic_compare_and_exchange_bool_rel,
catomic_compare_and_exchange_bool_rel): Remove.
* sysdeps/aarch64/atomic-machine.h
(atomic_compare_and_exchange_bool_rel): Likewise.
* sysdeps/alpha/atomic-machine.h
(atomic_compare_and_exchange_bool_rel): Likewise.
* sysdeps/arm/atomic-machine.h
(atomic_compare_and_exchange_bool_rel): Likewise.
* sysdeps/mips/atomic-machine.h
(atomic_compare_and_exchange_bool_rel): Likewise.
* sysdeps/tile/atomic-machine.h
(atomic_compare_and_exchange_bool_rel): Likewise.
Some architectures have their own versions of fdim functions, which
are missing errno setting (bug 6796) and may also return sNaN instead
of qNaN for sNaN input, in the case of the x86 / x86_64 long double
versions (bug 20256).
These versions are not actually doing anything that a compiler
couldn't generate, just straightforward comparisons / arithmetic (and,
in the x86 / x86_64 case, testing for NaNs with fxam, which isn't
actually needed once you use an unordered comparison and let the NaNs
pass through the same subtraction as non-NaN inputs). This patch
removes the x86 / x86_64 / powerpc versions, so that those
architectures use the generic C versions, which correctly handle
setting errno and deal properly with sNaN inputs. This seems better
than dealing with setting errno in lots of .S versions.
The i386 versions also return results with excess range and precision,
which is not appropriate for a function exactly defined by reference
to IEEE operations. For errno setting to work correctly on overflow,
it's necessary to remove excess range with math_narrow_eval, which
this patch duly does in the float and double versions so that the
tests can reliably pass on x86. For float, this avoids any double
rounding issues as the long double precision is more than twice that
of float. For double, double rounding issues will need to be
addressed separately, so this patch does not fully fix bug 20255.
Tested for x86_64, x86 and powerpc.
[BZ #6796]
[BZ #20255]
[BZ #20256]
* math/s_fdim.c: Include <math_private.h>.
(__fdim): Use math_narrow_eval on result.
* math/s_fdimf.c: Include <math_private.h>.
(__fdimf): Use math_narrow_eval on result.
* sysdeps/i386/fpu/s_fdim.S: Remove file.
* sysdeps/i386/fpu/s_fdimf.S: Likewise.
* sysdeps/i386/fpu/s_fdiml.S: Likewise.
* sysdeps/i386/i686/fpu/s_fdim.S: Likewise.
* sysdeps/i386/i686/fpu/s_fdimf.S: Likewise.
* sysdeps/i386/i686/fpu/s_fdiml.S: Likewise.
* sysdeps/powerpc/fpu/s_fdim.c: Likewise.
* sysdeps/powerpc/fpu/s_fdimf.c: Likewise.
* sysdeps/powerpc/powerpc32/fpu/s_fdim.c: Likewise.
* sysdeps/powerpc/powerpc64/fpu/s_fdim.c: Likewise.
* sysdeps/x86_64/fpu/s_fdiml.S: Likewise.
* math/libm-test.inc (fdim_test_data): Expect errno setting on
overflow. Add sNaN tests.
This implementation utilizes vectors to improve performance
compared to current byte by byte implementation for POWER7.
The performance improvement is upto 4x. This patch is tested
on powerpc64 and powerpc64le.
The powerpc64 versions of ceil, floor, round, trunc, rint, nearbyint
and their float versions return sNaN for sNaN input when they should
return qNaN. This patch fixes them to add a NaN argument to itself to
quiet sNaNs before returning.
Tested for powerpc64.
[BZ #20160]
* sysdeps/powerpc/powerpc64/fpu/s_ceil.S (__ceil): Add NaN
argument to itself before returning the result.
* sysdeps/powerpc/powerpc64/fpu/s_ceilf.S (__ceilf): Likewise.
* sysdeps/powerpc/powerpc64/fpu/s_floor.S (__floor): Likewise.
* sysdeps/powerpc/powerpc64/fpu/s_floorf.S (__floorf): Likewise.
* sysdeps/powerpc/powerpc64/fpu/s_nearbyint.S (__nearbyint):
Likewise.
* sysdeps/powerpc/powerpc64/fpu/s_nearbyintf.S (__nearbyintf):
Likewise.
* sysdeps/powerpc/powerpc64/fpu/s_rint.S (__rint): Likewise.
* sysdeps/powerpc/powerpc64/fpu/s_rintf.S (__rintf): Likewise.
* sysdeps/powerpc/powerpc64/fpu/s_round.S (__round): Likewise.
* sysdeps/powerpc/powerpc64/fpu/s_roundf.S (__roundf): Likewise.
* sysdeps/powerpc/powerpc64/fpu/s_trunc.S (__trunc): Likewise.
* sysdeps/powerpc/powerpc64/fpu/s_truncf.S (__truncf): Likewise.
The powerpc32 versions of ceil, floor, round, trunc, rint, nearbyint
and their float versions return sNaN for sNaN input when they should
return qNaN. This patch fixes them to add a NaN argument to itself to
quiet sNaNs before returning. The powerpc64 versions, which have the
same bug, will be addressed separately.
Tested for powerpc32.
[BZ #20160]
* sysdeps/powerpc/powerpc32/fpu/s_ceil.S (__ceil): Add NaN
argument to itself before returning the result.
* sysdeps/powerpc/powerpc32/fpu/s_ceilf.S (__ceilf): Likewise.
* sysdeps/powerpc/powerpc32/fpu/s_floor.S (__floor): Likewise.
* sysdeps/powerpc/powerpc32/fpu/s_floorf.S (__floorf): Likewise.
* sysdeps/powerpc/powerpc32/fpu/s_nearbyint.S (__nearbyint):
Likewise.
* sysdeps/powerpc/powerpc32/fpu/s_nearbyintf.S (__nearbyintf):
Likewise.
* sysdeps/powerpc/powerpc32/fpu/s_rint.S (__rint): Likewise.
* sysdeps/powerpc/powerpc32/fpu/s_rintf.S (__rintf): Likewise.
* sysdeps/powerpc/powerpc32/fpu/s_round.S (__round): Likewise.
* sysdeps/powerpc/powerpc32/fpu/s_roundf.S (__roundf): Likewise.
* sysdeps/powerpc/powerpc32/fpu/s_trunc.S (__trunc): Likewise.
* sysdeps/powerpc/powerpc32/fpu/s_truncf.S (__truncf): Likewise.
The powerpc implementations of fabsl for ldbl-128ibm (both powerpc32
and powerpc64) wrongly raise the "invalid" exception for sNaN
arguments. fabs functions should be quiet for all inputs including
signaling NaNs. The problem is the use of a comparison instruction
fcmpu to determine if the high part of the argument is negative and so
the low part needs to be negated; such instructions raise "invalid"
for sNaNs.
There is a pure integer implementation of fabsl in
sysdeps/ieee754/ldbl-128ibm/s_fabsl.c. However, it's not necessary to
use it to avoid such exceptions. The fsel instruction does not raise
exceptions for sNaNs, and can be used in place of the original
comparison. (Note that if the high part is zero or a NaN, it does not
matter whether the low part is negated; the choice of whether the low
part of a zero is +0 or -0 does not affect the value, and the low part
of a NaN does not affect the value / payload either.)
The condition in GCC for fsel to be available is TARGET_PPC_GFXOPT,
corresponding to the _ARCH_PPCGR predefined macro. fsel is available
on all 64-bit processors supported by GCC. A few 32-bit processors
supported by GCC do not have TARGET_PPC_GFXOPT despite having hard
float support. To support those processors, integer code (similar to
that in copysignl) is included for the !_ARCH_PPCGR case for
powerpc32.
Tested for powerpc32 (configurations with and without _ARCH_PPCGR) and
powerpc64.
[BZ #20157]
* sysdeps/powerpc/powerpc32/fpu/s_fabsl.S (__fabsl): Use fsel to
determine whether to negate low half if [_ARCH_PPCGR], and integer
comparison otherwise.
* sysdeps/powerpc/powerpc64/fpu/s_fabsl.S (__fabsl): Use fsel to
determine whether to negate low half.
Continuing fixes for ceil, floor and trunc functions not to raise the
"inexact" exception, this patch fixes the versions used on older
powerpc64 processors. As was done with the round implementations some
time ago, the save of floating-point state is moved after the first
floating-point operation on the input to ensure that any "invalid"
exception from signaling NaN input is included in the saved state, and
then the whole state gets restored rather than just the rounding mode.
This has no effect on configurations using the power5+ code, since
such processors can do these operations with a single instruction (and
those instructions do not set "inexact", so are correct for TS 18661-1
semantics).
Tested for powerpc64.
[BZ #15479]
* sysdeps/powerpc/powerpc64/fpu/s_ceil.S (__ceil): Move save of
floating-point state after first floating-point operation on
input. Restore full floating-point state instead of just rounding
mode.
* sysdeps/powerpc/powerpc64/fpu/s_ceilf.S (__ceilf): Likewise.
* sysdeps/powerpc/powerpc64/fpu/s_floor.S (__floor): Likewise.
* sysdeps/powerpc/powerpc64/fpu/s_floorf.S (__floorf): Likewise.
* sysdeps/powerpc/powerpc64/fpu/s_trunc.S (__trunc): Likewise.
* sysdeps/powerpc/powerpc64/fpu/s_truncf.S (__truncf): Likewise.
Continuing fixes for ceil, floor and trunc functions not to raise the
"inexact" exception, this patch fixes the versions used on older
powerpc32 processors. As was done with the round implementations some
time ago, the save of floating-point state is moved after the first
floating-point operation on the input to ensure that any "invalid"
exception from signaling NaN input is included in the saved state, and
then the whole state gets restored rather than just the rounding mode.
This has no effect on configurations using the power5+ code, since
such processors can do these operations with a single instruction (and
those instructions do not set "inexact", so are correct for TS 18661-1
semantics).
Tested for powerpc32.
[BZ #15479]
* sysdeps/powerpc/powerpc32/fpu/s_ceil.S (__ceil): Move save of
floating-point state after first floating-point operation on
input. Restore full floating-point state instead of just rounding
mode.
* sysdeps/powerpc/powerpc32/fpu/s_ceilf.S (__ceilf): Likewise.
* sysdeps/powerpc/powerpc32/fpu/s_floor.S (__floor): Likewise.
* sysdeps/powerpc/powerpc32/fpu/s_floorf.S (__floorf): Likewise.
* sysdeps/powerpc/powerpc32/fpu/s_trunc.S (__trunc): Likewise.
* sysdeps/powerpc/powerpc32/fpu/s_truncf.S (__truncf): Likewise.
The file sysdeps/powerpc/sysdeps.h defines aliases for condition register
operands. E.g.: 'cr7' means condition register 7. On the one hand, this
increases readability, as it makes it easier for readers to know whether the
operand is a condition register, a general purpose register or an immediate.
On the other hand, this permits that condition registers be written as if they
were general purpose, and vice-versa, thus reducing the readability of the
code.
This commit removes some of these unintentional misuses.
The changes have no effect on the final code. Checked with objdump.
Call __memset_power8 to pad, with zeros, the remaining bytes in the
dest string on __strncpy_power8 and __stpncpy_power8. This improves
performance when n is larger than the input string, giving ~30% gain for
larger strings without impacting much shorter strings.
This patch optimizes strcasestr function for power >= 8 systems. The average
improvement of this optimization is ~40% and compares 16 bytes at a time
using vector instructions. This patch is tested on powerpc64 and powerpc64le.
This utilizes vectors and bitmasks. For small needle, large
haystack, the performance improvement is upto 8x. For short
strings (0-4B), the cost of computing the bitmask dominates,
and is a tad slower.
This patch removes the powerpc64 optimized strspn, strcspn, and
strpbrk assembly implementation now that the default C one
implements the same strategy. On internal glibc benchtests
current implementations shows similar performance with -O2.
Tested on powerpc64le (POWER8).
* sysdeps/powerpc/powerpc64/strcspn.S: Remove file.
* sysdeps/powerpc/powerpc64/strpbrk.S: Remove file.
* sysdeps/powerpc/powerpc64/strspn.S: Remove file.
The ldbl-128ibm implementation of nearbyintl uses logic that only
works in round-to-nearest mode. This contrasts with rintl, which
works in all rounding modes.
Now, arguably nearbyintl could simply be aliased to rintl, given that
spurious "inexact" is generally allowed for ldbl-128ibm, even for the
underlying arithmetic operations. But given that the only point of
nearbyintl is to avoid "inexact", this patch follows the more
conservative approach of adding conditionals to the rintl
implementation to make it suitable for use to implement nearbyintl,
then builds it for nearbyintl with USE_AS_NEARBYINTL defined. The
test test-nearbyint-except-2 shows up issues when traps on "inexact"
are enabled, which turn out to be problems with the powerpc
fenv_private.h implementation (two functions that should disable
exception traps potentially failing to do so in some cases); this
patch duly fixes that as well (I don't see any other existing cases
where this would be user-visible; there isn't much use of *_NOEX,
*hold* etc. in libm that requires exceptions to be discarded and not
trapped on).
Tested for powerpc.
[BZ #19790]
* sysdeps/ieee754/ldbl-128ibm/s_rintl.c [USE_AS_NEARBYINTL]
(rintl): Define as macro.
[USE_AS_NEARBYINTL] (__rintl): Likewise.
(__rintl) [USE_AS_NEARBYINTL]: Use SET_RESTORE_ROUND_NOEX instead
of fesetround. Ensure results are evaluated before end of scope.
* sysdeps/ieee754/ldbl-128ibm/s_nearbyintl.c: Define
USE_AS_NEARBYINTL and include s_rintl.c.
* sysdeps/powerpc/fpu/fenv_private.h (libc_feholdsetround_ppc):
Disable exception traps in new environment.
(libc_feholdsetround_ppc_ctx): Likewise.
The operand modifier %s on powerpc is an undocumented internal implementation
detail of GCC. Besides that, the GCC community wants to remove it. This patch
rewrites the expressions that use this modifier with logically equivalent
expressions that don't require it.
Explanation for the substitution:
The %s modifier takes an immediate operand and prints 32 less such immediate.
Thus, in the previous code, the expression resulted in:
32 - __builtin_ffs(e)
where e was guaranteed to have exactly a single bit set, by the following
expressions:
(e & (e-1) == 0) : e has at most one bit set.
(e != 0) : e is not zero, thus it has at least one bit set.
Since we guarantee that there is exactly only one bit set, the following
statement is true:
32 - __builtin_ffs(e) == __builtin_clz(e)
Thus, we can replace __builtin_ffs with __builtin_clz and remove the %s operand
modifier.
HWCAP-related code should had been updated when the 32 bits of HWCAP were
used. This patch updates the code in dl-procinfo.h to loop through all
the 32 bits in HWCAP and updates _dl_powerpc_cap_flags accordingly.
On running tests after from-scratch ulps regeneration, I found that
some libm tests failed with ulps in excess of those recorded in the
from-scratch regeneration, which should never happen unless those ulps
exceed the limit on ulps that can go in libm-test-ulps files.
Failure: Test: atan2_upward (inf, -inf)
Result:
is: 2.35619498e+00 0x1.2d97ccp+1
should be: 2.35619450e+00 0x1.2d97c8p+1
difference: 4.76837159e-07 0x1.000000p-21
ulp : 2.0000
max.ulp : 1.0000
Maximal error of `atan2_upward'
is : 2 ulp
accepted: 1 ulp
Failure: Test: carg_upward (-inf + inf i)
Result:
is: 2.35619498e+00 0x1.2d97ccp+1
should be: 2.35619450e+00 0x1.2d97c8p+1
difference: 4.76837159e-07 0x1.000000p-21
ulp : 2.0000
max.ulp : 1.0000
Maximal error of `carg_upward'
is : 2 ulp
accepted: 1 ulp
The problem comes from the addition of tests for the finite-math-only
versions of libm functions. Those tests share ulps with the default
function variants. make regen-ulps runs the default tests before the
finite-math-only tests, concatenating the resulting ulps before
feeding them to gen-libm-test.pl to generate a new libm-test-ulps
file. But gen-libm-test.pl always takes the last ulps value given for
any (function, type) pair. So, if the largest ulps for a function
come from non-finite inputs, a from-scratch regeneration loses those
ulps.
This patch fixes gen-libm-test.pl, in the case where there are
multiple ulps values for a (function, type) pair - which can only
happen as part of a regeneration - to take the largest ulps value
rather than the last one.
Tested for ARM / MIPS / powerpc-nofpu.
* math/gen-libm-test.pl (parse_ulps): Do not reduce
already-recorded ulps.
* sysdeps/arm/libm-test-ulps: Regenerated.
* sysdeps/mips/mips32/libm-test-ulps: Likewise.
* sysdeps/mips/mips64/libm-test-ulps: Likewise.
* sysdeps/powerpc/nofpu/libm-test-ulps: Likewise.
Work around a GCC behavior with hardware transactional memory built-ins.
GCC doesn't treat the PowerPC transactional built-ins as compiler
barriers, moving instructions past the transaction boundaries and
altering their atomicity.
Commit 67385a01d2 added a new feature for
powerpc, where we store HWCAP/Platform bits in the TCB. In the dynamic
linking case, we use the versioned symbol
'__parse_hwcap_and_convert_at_platform' to verify if this feature is
available. However, the same symbol was not exported to libc.a, making
it not possible for GCC to check for it prior to link time.
The optimized POWER7 logb implementation does not use the absolute
value of the word extracted from the input to apply the leading 0-bits
builtin (to ignore the float sign). This patch fixes it by
clearing the signal bit in the resulting word.
It fixes the subnormal tests failures when running on POWER7 ou
newer chip.
Tested on powerpc64le (POWER8).
[BZ# 19375]
* sysdeps/powerpc/power7/fpu/s_logb.c (__logb): Fix return for
negative subnormals.
This patch adds a new feature for powerpc. In order to get faster access to
the HWCAP/HWCAP2 bits and platform number (i.e. for implementing
__builtin_cpu_is () / __builtin_cpu_supports () in GCC) without the overhead of
reading from the auxiliary vector, we now reserve space for them in the TCB.
This is an ABI change for GLIBC 2.23.
A new versioned symbol '__parse_hwcap_and_convert_at_platform' is available to
get the data from the auxiliary vector and parse it, and store it for later use
in the TLS initialization code. This function is called very early
(in _dl_sysdep_start () via DL_PLATFORM_INFO for the dynamic linking case, and
in __libc_start_main () for the static linking case) to make sure the data is
available at the time of TLS initialization.
* sysdeps/powerpc/Makefile (sysdep-dl-routines): Add hwcapinfo.
(sysdep_routines): Likewise.
(sysdep-rtld-routines): Likewise.
[$(subdir) = nptl](tests): Add test-get_hwcap and test-get_hwcap-static
[$(subdir) = nptl](tests-static): test-get_hwcap-static
* sysdeps/powerpc/Versions: Added new
__parse_hwcap_and_convert_at_platform symbol to GLIBC-2.23.
* sysdeps/powerpc/hwcapinfo.c: New file.
(__tcb_parse_hwcap_and_convert_at_platform): New function to initialize
and parse hwcap, hwcap2 and platform number information.
* sysdeps/powerpc/hwcapinfo.h: New file. Creates global variables
to store HWCAP+HWCAP2 and platform number.
* sysdeps/powerpc/nptl/tcb-offsets.sym: Added new offsets
for HWCAP+HWCAP2 and platform number in the TCB.
* sysdeps/powerpc/nptl/tls.h: New functionality. Stores
the HWCAP, HWCAP2 and platform number in the TCB.
(dtv): Added new fields for HWCAP+HWCAP2 and platform number.
(TLS_INIT_TP): Included calls to add the hwcap and
at_platform values in the TCB in TP initialization.
(TLS_DEFINE_INIT_TP): Likewise.
(THREAD_GET_HWCAP): New macro.
(THREAD_SET_HWCAP): Likewise.
(THREAD_GET_AT_PLATFORM): Likewise.
(THREAD_SET_AT_PLATFORM): Likewise.
* sysdeps/powerpc/powerpc32/dl-machine.h:
(dl_platform_init): New function that calls
__parse_hwcap_and_convert_at_platform for the dymanic linking case for
powerpc32.
* sysdeps/powerpc/powerpc64/dl-machine.h: Likewise, for powerpc64.
* sysdeps/powerpc/test-get_hwcap-static.c: New file. Testcase for
this functionality, static linking case.
* sysdeps/powerpc/test-get_hwcap.c: New file. Likewise, dynamic
linking case.
* sysdeps/unix/sysv/linux/powerpc/libc-start.c: Added call to
__parse_hwcap_and_convert_at_platform for the static linking case.
* sysdeps/unix/sysv/linux/powerpc/powerpc32/ld.abilist:
Included the new __parse_hwcap_and_convert_at_platform symbol in the
ABI list for GLIBC 2.23.
* sysdeps/unix/sysv/linux/powerpc/powerpc64/ld-le.abilist:
Likewise.
* sysdeps/unix/sysv/linux/powerpc/powerpc64/ld.abilist:
Likewise.
This patch optimizes powerpc spinlock implementation by:
* Use the correct EH hint bit on the larx for supported ISA. For lock
acquisition, the thread that acquired the lock with a successful stcx
does not want to give away the write ownership on the cacheline. The
idea is to make the load reservation "sticky" about retaining write
authority to the line. That way, the store that must inevitably come
to release the lock can succeed quickly and not contend with other
threads issuing lwarx. If another thread does a store to the line
(false sharing), the winning thread must give up write authority to
the proper value of EH for the larx for a lock acquisition is 1.
* Increase contented lock performance by up to 40%, and no measurable
impact on uncontended locks on P8.
Thanks to Adhemerval Zanella who did most of the work. I've run some
tests, and addressed some minor feedback.
* sysdeps/powerpc/nptl/pthread_spin_lock.c (pthread_spin_lock):
Add lwarx hint, and use macro for acquire instruction.
* sysdeps/powerpc/nptl/pthread_spin_trylock.c (pthread_spin_trylock):
Likewise.
* sysdep/unix/sysv/linux/powerpc/pthread_spin_unlock.c: Move to ...
* sysdeps/powerpc/nptl/pthread_spin_unlock.c: ... here, and
update to use new atomic macros.
The powerpc hard-float round and roundf functions, both 32-bit and
64-bit, raise spurious "inexact" exceptions for integer arguments from
adding 0.5 and rounding to integer toward zero.
Since these functions already save and restore the rounding mode, it's
natural to make them restore the full floating-point state instead to
fix this bug, which this patch does. The save of the state is moved
after the first floating-point operation on the input so that any
"invalid" exceptions from signaling NaN inputs are properly
preserved. As a consequence of this approach to the fix, "inexact"
for noninteger arguments (disallowed by TS 18661-1 but not by C99/C11,
see bug 15479) is also avoided for these implementations; this is
*not* a general fix for bug 15479 since plenty of other
implementations of various functions still raise spurious "inexact"
for noninteger arguments.
This issue and fix do not apply to builds using power5+ versions of
round and roundf, which use the frin instruction and avoid "inexact"
exceptions that way.
This patch should get hard-float powerpc32 and powerpc64 (default
function implementations) back to a state where test-float and
test-double will pass after ulps regeneration.
Tested for powerpc32 and powerpc64.
[BZ #15479]
[BZ #19238]
* sysdeps/powerpc/powerpc32/fpu/s_round.S (__round): Save
floating-point state after first operation on input. Restore full
state rather than just rounding mode.
* sysdeps/powerpc/powerpc32/fpu/s_roundf.S (__roundf): Likewise.
* sysdeps/powerpc/powerpc64/fpu/s_round.S (__round): Likewise.
* sysdeps/powerpc/powerpc64/fpu/s_roundf.S (__roundf): Likewise.
Similar to bug 19134 for powerpc32, the powerpc64 implementations of
lround, lroundf, llround, llroundf can raise spurious "inexact"
exceptions for integer arguments from adding 0.5 then converting to
integer (this does not apply to the power5+ version for double, which
uses the frin instruction which is defined never to raise "inexact"; I
don't know why power5+ doesn't use that version for float as well).
This patch fixes the bug in a similar way to the powerpc32 bug, by
testing for integers (adding and subtracting 2^52 and comparing with
the value before that addition and subtraction) and not adding 0.5 in
that case.
The powerpc maintainers may wish to look at making power5+ / power6x /
power8 use frin for float lround / llround as well as for double,
unless there's some reason I've missed that this isn't beneficial.
Tested for powerpc64.
[BZ #19235]
* sysdeps/powerpc/powerpc64/fpu/s_llround.S (__llround): Do not
add 0.5 to integer arguments.
* sysdeps/powerpc/powerpc64/fpu/s_llroundf.S (__llroundf):
Likewise.
(.LC2): New object.
Similar to bug 15491 recently fixed for x86_64 / x86, the powerpc
(both powerpc32 and powerpc64) hard-float implementations of
nearbyintf and nearbyint wrongly clear an "inexact" exception that was
raised before the function was called; this shows up as failure of the
test math/test-nearbyint-except added when that bug was fixed. They
also wrongly leave traps on "inexact" disabled if they were enabled
before the function was called.
This patch fixes the bugs similar to how the x86 bug was fixed: saving
and restoring the whole floating-point state, both to restore the
original "inexact" flag state and to restore the original state of
whether traps on "inexact" were enabled. Because there's a convenient
point in the powerpc implementations to save state after any sNaN
arguments will have raised "invalid" but before "inexact" traps need
to be disabled, no special handling for "invalid" is needed as in the
x86 version.
Tested for powerpc64 and powerpc32, where it fixes the
math/test-nearbyint-except failure as well as fixing the new test
math/test-nearbyint-except-2 added by this patch. Also tested for
x86_64 and x86 that the new test passes.
If powerpc experts see a more efficient way of doing this
(e.g. instruction positioning that's better for pipelines on typical
processors) then of course followups optimizing the fix are welcome.
[BZ #19228]
* sysdeps/powerpc/powerpc32/fpu/s_nearbyint.S (__nearbyint): Save
and restore full floating-point state.
* sysdeps/powerpc/powerpc32/fpu/s_nearbyintf.S (__nearbyintf):
Likewise.
* sysdeps/powerpc/powerpc64/fpu/s_nearbyint.S (__nearbyint):
Likewise.
* sysdeps/powerpc/powerpc64/fpu/s_nearbyintf.S (__nearbyintf):
Likewise.
* math/test-nearbyint-except-2.c: New file.
* math/Makefile (tests): Add test-nearbyint-except-2.
Since '--no-tls-optimize' is available for Power in ld, we need to provide
__tls_get_addr () in static libc in order to avoid undefined references to this
symbol when that flag is used.
* sysdeps/powerpc/libc-tls.c: New file. Provides __tls_get_addr () in
static libc.
The skip_lock_out_of_tbegin_retries adaptive parameter was
not being used correctly, nor as described. This prevents
a fallback for all users of the lock if a transient abort
occurs within the accepted number of retries.
[BZ #19174]
* sysdeps/powerpc/nptl/elide.h (__elide_lock): Fix usage of
.skip_lock_out_of_tbegin_retries.
* sysdeps/unix/sysv/linux/powerpc/elision-lock.c
(__lll_lock_elision): Likewise, and respect a value of
try_tbegin <= 0.
The previous code used to evaluate the preprocessor token is_lock_free to
a variable before starting a transaction. This behavior can cause an
error if another thread got the lock (without using a transaction)
between the evaluation of the token and the beginning of the transaction.
This bug can be triggered with the following order of events:
1. The lock accessed by is_lock_free is free.
2. Thread T1 evaluates is_lock_free and stores into register R1 that the
lock is free.
3. Thread T2 acquires the same lock used in is_lock_free.
4. T1 begins the transaction, creating a memory barrier where is_lock_free
is false, but R1 is true.
5. T1 reads R1 and doesn't abort the transaction.
6. T1 calls ELIDE_UNLOCK, which reads false from is_lock_free and decides
to unlock a lock acquired by T2, leading to undefined behavior.
This patch delays the evaluation of is_lock_free to inside a transaction
by moving this part of the code to the macro ELIDE_LOCK.
[BZ #18743]
* sysdeps/powerpc/nptl/elide.h (__elide_lock): Move most of this
code to...
(ELIDE_LOCK): ...here.
(__get_new_count): New function with part of the code from
__elide_lock that updates the value of adapt_count after a
transaction abort.
(__elided_trylock): Moved this code to...
(ELIDE_TRYLOCK): ...here.
The powerpc32 implementation of lround and lroundf can produce
spurious exceptions from adding 0.5 then converting to integer. This
includes "inexact" from the conversion to integer (not allowed for
integer arguments to these functions), and, for larger integer
arguments, "inexact", and "overflow" when rounding upward, from the
addition. In addition, "inexact" is not allowed together with
"invalid" and so inexact addition must be avoided when the integer
will be out of range of 32-bit long, whether or not the argument is an
integer.
This patch fixes these problems. As in the powerpc64 llround
implementation, a check is added for too-large arguments; in the
powerpc64 case that means arguments at least 2^52 in magnitude (so
that 0.5 cannot be added exactly), while in this case it means
arguments for which the result would overflow "long". In those cases
a suitable overflowing value is used for the integer conversion
without adding 0.5, while for smaller arguments it's tested whether
the argument is an integer (by adding and subtracting 2^52 to the
absolute value and comparing with the original absolute value) to
avoid adding 0.5 to integers and generating spurious "inexact".
This code is not used when the power5+ sysdeps directories are used,
as there's a separate power5+ version of these functions..
Tested for powerpc. This gets test-float (for a default powerpc32
hard-float build without any --with-cpu) back to the point where it
should pass once powerpc ulps are regenerated; test-double still needs
another problem with exceptions fixed to get back to that point (and I
haven't looked lately at what default powerpc64 results are like).
[BZ #19134]
* sysdeps/powerpc/powerpc32/fpu/s_lround.S (.LC1): New object.
(.LC2): Likewise.
(.LC3): Likewise.
(__lround): Do not add 0.5 to integer or out-of-range arguments.
The powerpc32 implementations of llroundf and llround produce spurious
and missing exceptions (some arising from such exceptions from
conversions to long long, some present even when fctidz is used).
This patch fixes those problems in a similar way to the llrint /
llrintf fixes. The spurious exceptions in the fctidz case for large
arguments arise from a converted value that saturated as LLONG_MAX
being converted back to float or double (the conversion back being
inexact, but "inexact" must not be raised together with "invalid"),
and from the subtraction x - xrf also being inexact for sufficiently
large arguments (whether the saturation was to LLONG_MAX or
LLONG_MIN); those are fixed by returning early if the argument is
large enough that no rounding is needed.
This code is not used for --with-cpu=power4 builds (I suspect the code
used in that case may also produce spurious "inexact" exceptions, but
that's something to investigate later).
Tested for powerpc.
[BZ #19125]
* sysdeps/powerpc/powerpc32/fpu/s_llround.c: Include <limits.h>,
<math_private.h> and <stdint.h>.
(__llround): Avoid conversions to and from long long int, and
subtractions, where those might raise spurious exceptions.
* sysdeps/powerpc/powerpc32/fpu/s_llroundf.c: Include
<math_private.h> and <stdint.h>.
(__llroundf): Avoid conversions to and from long long int, and
subtractions, where those might raise spurious exceptions.
