Add predefined macros for library use in defining __STDC_IEC_559*

ISO C99 and C11 specify various predefined macros for implementation
properties that describe library features, or features of the
combination of the compiler and library implementations.  In such
cases, GCC does not predefine the macros but provides a mechanism for
the library implementation to provide a stdc-predef.h header with
definitions of those macros, and implicitly preincludes that header in
compilations to meet the standard requirement for the macro
definitions to be constant throughout the translation unit.

Two such macros are __STDC_IEC_559__ and __STDC_IEC_559_COMPLEX__,
relating to IEEE 754 floating-point arithmetic support.  Now, for
neither of those are the relevant features fully implemented in GCC
(or most other C99 compilers), but as with __STDC_VERSION__ it's
appropriate to treat these macros as indicating intent rather than
completeness, given that the GCC documentation specifically disclaims
completeness of C99 or C11 support.

Intent can be unclear in some cases, but it's reasonably clear that if
the user used -ffast-math, for example, that has specified an intent
to enable optimizations that conflict with the C99/C11 Annex F
requirements.  And if the compiler configuration uses software
floating point without exceptions and rounding modes support (and
making software floating point support exceptions and rounding modes
depends on compiler and library cooperation, and likely isn't desired
for many configurations that use software floating point), that is
also prevents proper Annex F support.

<https://sourceware.org/ml/libc-alpha/2005-03/msg00196.html> suggested
a glibc sysdeps mechanism for controlling whether these macros are
defined, but that would only handle architecture-specific conditionals
for software floating point.  There are various GCC options that
indicate intent contrary to Annex F, many of which do not predefine
macros that could be used to control the __STDC_* definitions
appropriately.

This patch adds new predefined macros __GCC_IEC_559 and
__GCC_IEC_559_COMPLEX, whose values specify the intent of the compiler
configuration and command-line options and so can be used by
stdc-predef.h to determine whether to define the __STDC_* macros.  The
detailed semantics are given in the documentation changes;
distinguishing whether IEEE 754-2008 is supported is intended to allow
use in future to control the definition of __STDC_IEC_60559_BFP__ if
support for TS 18661-1 (latest draft: WG14 N1756) were to be added to
glibc (again, there are compiler features involved, quite complicated
to implement in the case of the static rounding direction pragmas, but
the macros are treated as being about intent rather than
completeness).

(It is quite likely there is *also* use for predefined macros for
various individual floating-point options that don't currently have
them, to facilitate optimization of library calls by indicating to the
library what the requirements are for accuracy, rounding mode
handling, exception handling etc. of library functions being called
from the current translatin unit, and so allow calls to be directed to
appropriate optimized function variants.  But that's independent of
the new macros in support of __STDC_*; I don't think __STDC_* can
readily be computed from macros for individual options, given the
possibility of new options for new optimizations that may also
conflict with Annex F requirements.)

The default macro definitions are based on whether there is an adddf3
insn pattern - except on powerpc*-*-linux*, where a separate target
hook implementation is used because of software floating point support
in glibc handling exceptions and rounding modes - and on various
command-line options.  (-ftrapping-math and -frounding-math are not
considered because in C99/C11 terms those relate to the default state
of the FENV_ACCESS pragma, which is permitted to be "off".)

There is a "???" comment regarding -ffp-contract, where logically
-ffp-contract=fast should be considered to conflict with Annex F but
options such as -std=c99 don't do anything to change the default to
-ffp-contract=on (= -ffp-contract=off at present).  I looked back at
some of the discussion (Oct/Nov 2010) from when -ffp-contract was
added, but didn't find anything giving a reason why conformance
options don't do this, as I suggested in
<http://gcc.gnu.org/ml/gcc-patches/2010-10/msg01951.html> that they
should.  So in the absence of someone coming up with such a reason I'd
be inclined to make such options enable -ffp-contract=on, just as they
enable -fexcess-precision=standard on 32-bit x86, and add a
corresponding check when setting the new macros.

Portable testcases for these macros are difficult, given the inherent
system-dependency of their values (which is part of the point of
adding them), but tests are added that various options force the
values of the macros to 0, along with a test enabled for i?86-*-linux*
x86_64-*-linux* powerpc*-*-linux* that the values are 2 on those
targets with -std=c11.

Bootstrapped with no regressions on x86_64-unknown-linux-gnu, and
tested with no regressions with cross to powerpc-linux-gnu
(soft-float).  OK to commit (hook addition, rs6000/powerpc changes)?

2013-10-15  Joseph Myers  <joseph@codesourcery.com>

	* doc/cpp.texi (__GCC_IEC_559, __GCC_IEC_559_COMPLEX): Document
	macros.
	* target.def (float_exceptions_rounding_supported_p): New hook.
	* targhooks.c (default_float_exceptions_rounding_supported_p): New
	function.
	* targhooks.h (default_float_exceptions_rounding_supported_p):
	Declare.
	* doc/tm.texi.in (TARGET_FLOAT_EXCEPTIONS_ROUNDING_SUPPORTED_P):
	New @hook.
	* doc/tm.texi: Regenerate.
	* config.gcc (powerpc*-*-linux*): Set extra_objs.
	* config/rs6000/rs6000-linux.c: New file.
	* config/rs6000/rs6000-protos.h
	(rs6000_linux_float_exceptions_rounding_supported_p): Declare.
	* config/rs6000/linux.h
	(TARGET_FLOAT_EXCEPTIONS_ROUNDING_SUPPORTED_P): New macro.
	* config/rs6000/linux64.h
	(TARGET_FLOAT_EXCEPTIONS_ROUNDING_SUPPORTED_P): Likewise.
	* config/rs6000/t-linux (rs6000-linux.o): New rule.
	* config/rs6000/t-linux64 (rs6000-linux.o): Likewise.

c-family:
2013-10-15  Joseph Myers  <joseph@codesourcery.com>

	* c-cppbuiltin.c (cpp_iec_559_value, cpp_iec_559_complex_value):
	New functions.
	(c_cpp_builtins): Define __GCC_IEC_559 and __GCC_IEC_559_COMPLEX.

testsuite:
2013-10-15  Joseph Myers  <joseph@codesourcery.com>

	* gcc.dg/iec-559-macros-1.c, gcc.dg/iec-559-macros-2.c,
	gcc.dg/iec-559-macros-3.c, gcc.dg/iec-559-macros-4.c,
	gcc.dg/iec-559-macros-5.c, gcc.dg/iec-559-macros-6.c,
	gcc.dg/iec-559-macros-7.c, gcc.dg/iec-559-macros-8.c,
	gcc.dg/iec-559-macros-9.c: New tests.

Add predefined macros for library use in defining __STDC_IEC_559*

Commit Message

Comments

Patch