AuroraMiddleware/glibc
1
0
Fork 0
mirror of https://sourceware.org/git/glibc.git synced 2024-11-26 23:10:06 +00:00
Code Issues Packages Projects Releases Wiki Activity
33,235 Commits 334 Branches 1,646 Tags 278 MiB
de800d8305
Commit Graph

7 Commits

Author SHA1 Message Date
Joseph Myers
63716ab270 Add build infrastructure for narrowing libm functions. 
TS 18661-1 defines libm functions that carry out an operation (+ - * /
sqrt fma) on their arguments and return a result rounded to a
(usually) narrower type, as if the original result were computed to
infinite precision and then rounded directly to the result type
without any intermediate rounding to the argument type.  For example,
fadd, faddl and daddl for addition.  These are the last remaining TS
18661-1 functions left to be added to glibc.  TS 18661-3 extends this
to corresponding functions for _FloatN and _FloatNx types.

As functions parametrized by two rather than one varying
floating-point types, these functions require infrastructure in glibc
that was not required for previous libm functions.  This patch
provides such infrastructure - excluding test support, and actual
function implementations, which will be in subsequent patches.

Declaring the functions uses a header bits/mathcalls-narrow.h, which
is included many times, for each relevant pair of types.  This will
end up containing macro calls of the form

__MATHCALL_NARROW (__MATHCALL_NAME (add), __MATHCALL_REDIR_NAME (add), 2);

for each family of narrowing functions.  (The structure of this macro
call, with the calls to __MATHCALL_NAME and __MATHCALL_REDIR_NAME
there rather than in the definition of __MATHCALL_NARROW, arises from
the names such as "add" *not* themselves being reserved identifiers -
meaning it's necessary to avoid any indirection that would result in a
user-defined "add" macro being expanded.)  Whereas for existing
functions declaring long double functions is disabled if _LIBC in the
case where they alias double functions, to facilitate defining the
long double functions as aliases of the double ones, there is no such
logic for the narrowing functions in this patch.  Rather, the files
defining such functions are expected to use #define to hide the
original declarations of the alias names, to avoid errors about
defining aliases with incompatible types.

math/Makefile support is added for building the functions (listed in
libm-narrow-fns, currently empty) for all relevant pairs of types.  An
internal header math-narrow.h is added for macros shared between
multiple function implementations - currently a ROUND_TO_ODD macro to
facilitate writing functions using the round-to-odd implementation
approach, and alias macros to create all the required function
aliases.  libc_feholdexcept_setroundf128 and libc_feupdateenv_testf128
are added for use when required (only for x86_64).  float128_private.h
support is added for ldbl-128 narrowing functions to be used for
_Float128.

Certain things are specifically omitted from this patch and the
immediate followups.  tgmath.h support is deferred; there remain
unresolved questions about how the type-generic macros for these
functions are supposed to work, especially in the case of arguments of
integer type.  The math.h / bits/mathcalls-narrow.h logic, and the
logic for determining what functions / aliases to define, will need
some adjustments to support the sqrt and fma functions, where
e.g. f32xsqrtf64 can just be an alias for sqrt rather than a separate
function.  TS 18661-1 defines FP_FAST_* macros but no support is
included for defining them (they won't in general be true without
architecture-specific optimized function versions).

For each of the function groups (add sub mul div sqrt fma) there are
always six functions present (e.g. fadd, faddl, daddl, f32addf64,
f32addf32x, f32xaddf64).  When _Float64x and _Float128 are supported,
there are seven more (e.g. f32addf64x, f32addf128, f64addf64x,
f64addf128, f32xaddf64x, f32xaddf128, f64xaddf128).  In addition, in
the ldbl-opt case there are function names such as __nldbl_daddl (an
alias for f32xaddf64, which is not a reserved name in TS 18661-1, only
in TS 18661-3), for calls to daddl to be mapped to in the
-mlong-double-64 case.  (Calls to faddl just get mapped to fadd, and
for sqrt and fma there won't be __nldbl_* functions because dsqrtl and
dfmal can just be mapped to sqrt and fma with -mlong-double-64.)

While there are six or thirteen functions present in each group (plus
__nldbl_* names only as an ABI, not an API), not all are distinct;
they fall in various groups of aliases.  There are two distinct
versions built if long double has the same format as double; four if
they have distinct formats but there is no _Float64x or _Float128
support; five if long double has binary128 format; seven when
_Float128 is distinct from long double.

Architecture-specific optimized versions are possible, but not
included in my patches.  For example, IA64 generally supports
narrowing the result of most floating-point instructions; Power ISA
2.07 (POWER8) supports double values as arguments to float
instructions, with the results narrowed as expected; Power ISA 3
(POWER9) supports round-to-odd for float128 instructions, so meaning
that approach can be used without needing to set and restore the
rounding mode and test "inexact".  I intend to leave any such
optimized versions to the architecture maintainers.  Generally in such
cases it would also make sense for calls to these functions to be
expanded inline (given -fno-math-errno); I put a suggestion for TS
18661-1 built-in functions at <https://gcc.gnu.org/wiki/SummerOfCode>.

Tested for x86_64 (this patch in isolation, as well as testing for
various configurations in conjunction with further patches).

	* math/bits/mathcalls-narrow.h: New file.
	* include/bits/mathcalls-narrow.h: Likewise.
	* math/math-narrow.h: Likewise.
	* math/math.h (__MATHCALL_NARROW_ARGS_1): New macro.
	(__MATHCALL_NARROW_ARGS_2): Likewise.
	(__MATHCALL_NARROW_ARGS_3): Likewise.
	(__MATHCALL_NARROW_NORMAL): Likewise.
	(__MATHCALL_NARROW_REDIR): Likewise.
	(__MATHCALL_NARROW): Likewise.
	[__GLIBC_USE (IEC_60559_BFP_EXT)]: Repeatedly include
	<bits/mathcalls-narrow.h> with _Mret_, _Marg_ and __MATHCALL_NAME
	defined.
	[__GLIBC_USE (IEC_60559_TYPES_EXT)]: Likewise.
	* math/Makefile (headers): Add bits/mathcalls-narrow.h.
	(libm-narrow-fns): New variable.
	(libm-narrow-types-basic): Likewise.
	(libm-narrow-types-ldouble-yes): Likewise.
	(libm-narrow-types-float128-yes): Likewise.
	(libm-narrow-types-float128-alias-yes): Likewise.
	(libm-narrow-types): Likewise.
	(libm-routines): Add narrowing functions.
	* sysdeps/i386/fpu/fenv_private.h [__x86_64__]
	(libc_feholdexcept_setroundf128): New macro.
	[__x86_64__] (libc_feupdateenv_testf128): Likewise.
	* sysdeps/ieee754/float128/float128_private.h: Include
	<math/math-narrow.h>.
	[libc_feholdexcept_setroundf128] (libc_feholdexcept_setroundl):
	Undefine and redefine.
	[libc_feupdateenv_testf128] (libc_feupdateenv_testl): Likewise.
	(libm_alias_float_ldouble): Undefine and redefine.
	(libm_alias_double_ldouble): Likewise.
 2018-02-09 21:18:52 +00:00
Joseph Myers
c86ed71d63 Add float128 support for x86_64, x86. 
This patch enables float128 support for x86_64 and x86.  All GCC
versions that can build glibc provide the required support, but since
GCC 6 and before don't provide __builtin_nanq / __builtin_nansq, sNaN
tests and some tests of NaN payloads need to be disabled with such
compilers (this does not affect the generated glibc binaries at all,
just the tests).  bits/floatn.h declares float128 support to be
available for GCC versions that provide the required libgcc support
(4.3 for x86_64, 4.4 for i386 GNU/Linux, 4.5 for i386 GNU/Hurd);
compilation-only support was present some time before then, but not
really useful without the libgcc functions.

fenv_private.h needed updating to avoid trying to put _Float128 values
in registers.  I make no assertion of optimality of the
math_opt_barrier / math_force_eval definitions for this case; they are
simply intended to be sufficient to work correctly.

Tested for x86_64 and x86, with GCC 7 and GCC 6.  (Testing for x32 was
compilation tests only with build-many-glibcs.py to verify the ABI
baseline updates.  I have not done any testing for Hurd, although the
float128 support is enabled there as for GNU/Linux.)

	* sysdeps/i386/Implies: Add ieee754/float128.
	* sysdeps/x86_64/Implies: Likewise.
	* sysdeps/x86/bits/floatn.h: New file.
	* sysdeps/x86/float128-abi.h: Likewise.
	* manual/math.texi (Mathematics): Document support for _Float128
	on x86_64 and x86.
	* sysdeps/i386/fpu/fenv_private.h: Include <bits/floatn.h>.
	(math_opt_barrier): Do not put _Float128 values in floating-point
	registers.
	(math_force_eval): Likewise.
	[__x86_64__] (SET_RESTORE_ROUNDF128): New macro.
	* sysdeps/x86/fpu/Makefile [$(subdir) = math] (CPPFLAGS): Append
	to Makefile variable.
	* sysdeps/x86/fpu/e_sqrtf128.c: New file.
	* sysdeps/x86/fpu/sfp-machine.h: Likewise.  Based on libgcc.
	* sysdeps/x86/math-tests.h: New file.
	* math/libm-test-support.h (XFAIL_FLOAT128_PAYLOAD): New macro.
	* math/libm-test-getpayload.inc (getpayload_test_data): Use
	XFAIL_FLOAT128_PAYLOAD.
	* math/libm-test-setpayload.inc (setpayload_test_data): Likewise.
	* math/libm-test-totalorder.inc (totalorder_test_data): Likewise.
	* math/libm-test-totalordermag.inc (totalordermag_test_data):
	Likewise.
	* sysdeps/unix/sysv/linux/i386/libc.abilist: Update.
	* sysdeps/unix/sysv/linux/i386/libm.abilist: Likewise.
	* sysdeps/unix/sysv/linux/x86_64/64/libc.abilist: Likewise.
	* sysdeps/unix/sysv/linux/x86_64/64/libm.abilist: Likewise.
	* sysdeps/unix/sysv/linux/x86_64/x32/libc.abilist: Likewise.
	* sysdeps/unix/sysv/linux/x86_64/x32/libm.abilist: Likewise.
	* sysdeps/i386/fpu/libm-test-ulps: Likewise.
	* sysdeps/i386/i686/fpu/multiarch/libm-test-ulps: Likewise.
	* sysdeps/x86_64/fpu/libm-test-ulps: Likewise.
 2017-06-26 22:02:24 +00:00
Joseph Myers
be25493251 Fix yn overflow handling in non-default rounding modes (bug 16561, bug 16562). 
This patch fixes bugs 16561 and 16562, bad results of yn in overflow
cases in non-default rounding modes, both because an intermediate
overflow in the recurrence does not get detected if the result is not
an infinity and because an overflowing result may occur in the wrong
sign.  The fix is to set FE_TONEAREST mode internally for the parts of
the function where such overflows can occur (which includes the call
to y1 - where yn is used to compute a Bessel function of order -1,
negating the result of y1 isn't correct for overflowing results in
directed rounding modes) and then compute an overflowing value in the
original rounding mode if the to-nearest result was an infinity.

Tested x86_64 and x86 and ulps updated accordingly.  Also tested for
mips64 and powerpc32 to test the ldbl-128 and ldbl-128ibm changes.

(The tests for these bugs were added in my previous y1 patch, so the
only thing this patch has to do with the testsuite is enable yn
testing in all rounding modes.)

	[BZ #16561]
	[BZ #16562]
	* sysdeps/ieee754/dbl-64/e_jn.c: Include <float.h>.
	(__ieee754_yn): Set FE_TONEAREST mode internally and then
	recompute overflowing results in original rounding mode.
	* sysdeps/ieee754/flt-32/e_jnf.c: Include <float.h>.
	(__ieee754_ynf): Set FE_TONEAREST mode internally and then
	recompute overflowing results in original rounding mode.
	* sysdeps/ieee754/ldbl-128/e_jnl.c: Include <float.h>.
	(__ieee754_ynl): Set FE_TONEAREST mode internally and then
	recompute overflowing results in original rounding mode.
	* sysdeps/ieee754/ldbl-128ibm/e_jnl.c: Include <float.h>.
	(__ieee754_ynl): Set FE_TONEAREST mode internally and then
	recompute overflowing results in original rounding mode.
	* sysdeps/ieee754/ldbl-96/e_jnl.c: Include <float.h>.
	(__ieee754_ynl): Set FE_TONEAREST mode internally and then
	recompute overflowing results in original rounding mode.
	* sysdeps/i386/fpu/fenv_private.h [!__SSE2_MATH__]
	(libc_feholdsetround_ctx): New macro.
	* math/libm-test.inc (yn_test): Use ALL_RM_TEST.
	* sysdeps/i386/fpu/libm-test-ulps: Update.
	* sysdeps/x86_64/fpu/libm-test-ulps : Likewise.
 2014-06-27 14:52:13 +00:00
Siddhesh Poyarekar
2506109403 Set/restore rounding mode only when needed 
The most common use case of math functions is with default rounding
mode, i.e. rounding to nearest.  Setting and restoring rounding mode
is an unnecessary overhead for this, so I've added support for a
context, which does the set/restore only if the FP status needs a
change.  The code is written such that only x86 uses these.  Other
architectures should be unaffected by it, but would definitely benefit
if the set/restore has as much overhead relative to the rest of the
code, as the x86 bits do.

Here's a summary of the performance improvement due to these
improvements; I've only mentioned functions that use the set/restore
and have benchmark inputs for x86_64:

Before:

cos(): ITERS:4.69335e+08: TOTAL:28884.6Mcy, MAX:4080.28cy, MIN:57.562cy, 16248.6 calls/Mcy
exp(): ITERS:4.47604e+08: TOTAL:28796.2Mcy, MAX:207.721cy, MIN:62.385cy, 15543.9 calls/Mcy
pow(): ITERS:1.63485e+08: TOTAL:28879.9Mcy, MAX:362.255cy, MIN:172.469cy, 5660.86 calls/Mcy
sin(): ITERS:3.89578e+08: TOTAL:28900Mcy, MAX:704.859cy, MIN:47.583cy, 13480.2 calls/Mcy
tan(): ITERS:7.0971e+07: TOTAL:28902.2Mcy, MAX:1357.79cy, MIN:388.58cy, 2455.55 calls/Mcy

After:

cos(): ITERS:6.0014e+08: TOTAL:28875.9Mcy, MAX:364.283cy, MIN:45.716cy, 20783.4 calls/Mcy
exp(): ITERS:5.48578e+08: TOTAL:28764.9Mcy, MAX:191.617cy, MIN:51.011cy, 19071.1 calls/Mcy
pow(): ITERS:1.70013e+08: TOTAL:28873.6Mcy, MAX:689.522cy, MIN:163.989cy, 5888.18 calls/Mcy
sin(): ITERS:4.64079e+08: TOTAL:28891.5Mcy, MAX:6959.3cy, MIN:36.189cy, 16062.8 calls/Mcy
tan(): ITERS:7.2354e+07: TOTAL:28898.9Mcy, MAX:1295.57cy, MIN:380.698cy, 2503.7 calls/Mcy

So the improvements are:

cos: 27.9089%
exp: 22.6919%
pow: 4.01564%
sin: 19.1585%
tan: 1.96086%

The downside of the change is that it will have an adverse performance
impact on non-default rounding modes, but I think the tradeoff is
justified.
 2013-06-12 10:36:48 +05:30
Siddhesh Poyarekar
0b57daebab Fix application of the exception mask 
Fixes BZ #14496.
 2013-01-18 14:16:25 +05:30
Joseph Myers
5b5b04d628 Make fma use of Dekker and Knuth algorithms use round-to-nearest (bug 14796). 2012-11-03 19:48:53 +00:00
Richard Henderson
bd37f2ee31 Optimize private 387 fenv access; share code between i386 and x86_64. 2012-03-19 06:51:39 -07:00