GCC 4.6 Release Series Changes, New Features, and Fixes
Caveats
The options -b <machine> and
-V <version> have been removed because
they were unreliable. Instead, users should directly run
<machine>-gcc when cross-compiling, or
<machine>-gcc-<version>
to run a different version of gcc.
GCC now has stricter checks for invalid command-line options.
In particular, when gcc was called to link object
files rather than compile source code, it would previously accept
and ignore all options starting with --, including
linker options such as --as-needed
and --export-dynamic, although such options would
result in errors if any source code was compiled. Such options,
if unknown to the compiler, are now rejected in all cases; if the
intent was to pass them to the linker, options such
as -Wl,--as-needed should be used.
Versions of the GNU C library up to and including 2.11.1
included an incorrect
implementation of the cproj function. GCC
optimizes its builtin cproj according to the behavior
specified and allowed by the ISO C99 standard. If you want to
avoid discrepancies between the C library and GCC's builtin
transformations when using cproj in your code, use
GLIBC 2.12 or later. If you are using an older GLIBC and actually
rely on the incorrect behavior of cproj, then you can disable
GCC's transformations using -fno-builtin-cproj.
The C-only intermodule optimization framework (IMA, enabled by
-combine) has been removed in favor of the new
generic link-time optimization framework (LTO) introduced
in GCC 4.5.0.
GCC now ships with the LGPL-licensed
libquadmath library, which provides quad-precision
mathematical functions for targets with a __float128
datatype. __float128 is available for targets on
32-bit x86, x86-64 and Itanium architectures. The
libquadmath library is automatically built on
such targets when building the Fortran compiler.
New -Wunused-but-set-variable and
-Wunused-but-set-parameter warnings were added
for C, C++, Objective-C and Objective-C++.
These warnings diagnose variables respective parameters which
are only set in the code and never otherwise used.
Usually such variables are useless and often even the value
assigned to them is computed needlessly, sometimes expensively.
The -Wunused-but-set-variable warning is enabled by
default by -Wall flag and -Wunused-but-set-parameter
by -Wall -Wextra flags.
Support for a number of older systems and recently
unmaintained or untested target ports of GCC has been declared
obsolete in GCC 4.6. Unless there is activity to revive them, the
next release of GCC will have their sources permanently
removed.
All GCC ports for the following processor
architectures have been declared obsolete:
Argonaut ARC (arc-*)
National Semiconductor CRX (crx-*)
Motorola 68HC11 and 68HC12
(m68hc11-*-*, m6811-*-*,
m68hc12-*-*, m6812-*-*)
Sunplus S+core (score-*)
The following ports for individual systems on
particular architectures have been obsoleted:
Interix (i[34567]86-*-interix3*)
Generic ARM PE (arm-*-pe* other
than arm*-wince-pe*)
MCore PE (mcore-*-pe*)
SH SymbianOS (sh*-*-symbianelf*)
GNU Hurd on Alpha and PowerPC
(alpha*-*-gnu*, powerpc*-*-gnu*)
M68K uClinux old ABI
(m68k-*-uclinuxoldabi*)
a.out NetBSD
(arm*-*-netbsd*, i[34567]86-*-netbsd*,
vax-*-netbsd*, but
not *-*-netbsdelf*)
The i[34567]86-*-pe alias for Cygwin targets has
also been obsoleted; users should configure
for i[34567]86-*-cygwin* instead.
Certain configure options to control the set of libraries built
with GCC on some targets have been obsoleted. On ARM targets, the
options --disable-fpu, --disable-26bit,
--disable-underscore, --disable-interwork,
--disable-biendian and --disable-nofmult
have been obsoleted. On MIPS targets, the options
--disable-single-float, --disable-biendian
and --disable-softfloat have been obsoleted.
A new general optimization level, -Ofast, has been
introduced. It combines the existing optimization level -O3
with options that can affect standards compliance but result in
better optimized code. For example, -Ofast enables
-ffast-math.
Link-time optimization improvements:
The Scalable Whole
Program Optimizer (WHOPR) project has stabilized to the
point of being usable. It has become the default mode when
using the LTO optimization model. Link time optimization can
now split itself into multiple parallel compilations. Parallelism
is controlled with -flto=n (where
n specifies the number of compilations to execute in
parallel). GCC can also cooperate with a GNU make job server
by specifying the -flto=jobserver option and
adding + to the beginning of the
Makefile rule executing the linker.
Classical LTO mode can be enforced by
-flto-partition=none. This may result in small code
quality improvements.
A large number of bugs were fixed. GCC itself, Mozilla
Firefox and other large applications can be built with
LTO enabled.
The linker plugin support improvements
Linker plugin is now enabled by default when the linker is
detected to have plugin support. This is the case for GNU
ld 2.21.51 or newer (on ELF and Cygwin targets) and the Gold
linker on ELF targets. Plugin support of the Apple linker on
Darwin is not compatible with GCC.
The linker plugin can also be controlled by the
-fuse-linker-plugin command line option.
Resolution information from the linker plugin is used to drive
whole program assumptions. Use of the linker plugin results in
more aggressive optimization on binaries and on shared libraries
that use the hidden visibility
attribute. Consequently the use
of -fwhole-program is not neccesary in addition to
LTO.
Hidden symbols used from non-LTO objects now have to be
explicitly annotated with externally_visible when
the linker plugin is not used.
C++ inline functions and virtual tables are now privatized more
aggressively, leading to better inter-procedural optimization
and faster dynamic linking.
Memory usage and intermediate language streaming performance
have been improved.
Static constructors and destructors from individual units are
inlined into a single function.
This can significantly improve startup times of large C++
applications where static constructors are very common. For
example, static constructors are used when including the
iostream header.
Support for the Ada language has been added.
Interprocedural optimization improvements
The interprocedural framework was re-tuned for link time
optimization. Several scalability issues were resolved.
Improved auto-detection of const and pure
functions. Newly, noreturn functions are auto-detected.
A number of inlining heuristic improvements. In particular:
Partial inlining is now supported and enabled by default
at -O2 and greater. The feature can be
controlled via -fpartial-inlining.
Partial inlining splits functions with short hot path
to return. This allows more aggressive inlining of the
hot path leading to better performance and often to
code size reductions (because cold parts of functions
are not duplicated).
Scalability for large compilation units was improved
significantly.
Inlining of callbacks is now more aggressive.
Virtual methods are considered for inlining when the
caller is
inlined and devirtualization is then possible.
Inlining when optimizing for size (either in cold
regions of a program or when compiling with
-Os) was improved to better handle C++
programs with larger abstraction penalty, leading
to smaller and faster code.
The IPA reference optimization pass detecting global
variables used or modified by functions was strengthened
and sped up.
Functions whose address was taken are now optimized out
when all references to them are dead.
A new inter-procedural static profile estimation pass detects
functions that are executed once or unlikely to be executed.
Unlikely executed functions are optimized for size. Functions
executed once are optimized for size except for the inner
loops.
On most targets with named section support, functions used only
at startup (static constructors and main), functions
used only at exit and functions detected to be cold are placed into
separate text segment subsections.
This extends the -freorder-functions feature and is
controlled by the same switch. The goal is to improve the startup
time of large C++ programs.
Proper function placement requires linker support.
GNU ld 2.21.51 on ELF targets was updated to place
those functions together within the text section leading to better code
locality and faster startup times of large C++ programs. The feature is
also supported in the Apple linker.
Support in the gold linker is planned.
A new switch -fstack-usage has been added. It makes
the compiler output stack usage information for the program, on a
per-function basis, in an auxiliary file.
A new switch -fcombine-stack-adjustments has been added.
It can be used to enable or disable the compiler's stack-slot combining
pass which before was enabled automatically at -O1 and above,
but could not be controlled on its own.
A new switch -fstrict-volatile-bitfields has been
added. Using it indicates that accesses to volatile bitfields
should use a single access of the width of the field's type.
This option can be useful for precisely defining and accessing
memory-mapped peripheral registers from C or C++.
Compile time and memory usage improvements
Datastructures used by the dataflow framework in GCC were reorganized
for better memory usage and more cache locality. Compile
time is improved especially on units with large functions (possibly
resulting from a lot of inlining) not fitting into the processor cache.
The compile time of the GCC C compiler binary with link-time
optimization went down by over 10% (benchmarked on x86-64 target).
New Languages and Language specific improvements
Ada
Stack checking has been improved on selected architectures (Alpha,
IA-32/x86-64, RS/6000 and SPARC): it now will detect stack overflows
in all cases on these architectures.
Initial support for Ada 2012 has been added.
C family
A new warning, enabled by -Wdouble-promotion,
has been added that warns about cases where a value of type
float is implicitly promoted to double.
This is especially helpful for CPUs that handle the former in
hardware, but emulate the latter in software.
A new function attribute leaf was introduced.
This attribute allows better inter-procedural optimization across
calls to functions that return to the current unit only via returning
or exception handling. This is the case for most library functions
that have no callbacks.
Support for a new data type __int128 for targets having
wide enough machine-mode support.
The new function attribute callee_pop_aggregate allows
to specify if the caller or callee is responsible for popping the
aggregate return pointer value from the stack.
Support for selectively enabling and disabling warnings
via #pragma GCC diagnostic has been added. For instance:
#pragma GCC diagnostic error "-Wuninitialized"
foo(a); /* error is given for this one */
#pragma GCC diagnostic push
#pragma GCC diagnostic ignored "-Wuninitialized"
foo(b); /* no diagnostic for this one */
#pragma GCC diagnostic pop
foo(c); /* error is given for this one */
#pragma GCC diagnostic pop
foo(d); /* depends on command line options */
The -fmax-errors=N option is now supported. Using
this option causes the compiler to exit after N errors
have been issued.
C
There is now experimental support for some features from the
upcoming C1X revision of the ISO C standard. This support may be
selected with -std=c1x, or -std=gnu1x
for C1X with GNU extensions. Note that this support is
experimental and may change incompatibly in future releases for
consistency with changes to the C1X standard draft. The following
features are newly supported as described in the N1539 draft of
C1X (with changes agreed at the March 2011 WG14 meeting); some
other features were already supported with no compiler
changes being needed, or have some support but not in full accord
with N1539 (as amended).
Static assertions (_Static_assert keyword)
Typedef redefinition
New macros in <float.h>
Anonymous structures and unions
The new -fplan9-extensions option directs the
compiler to support some extensions for anonymous struct fields
which are implemented by the Plan 9 compiler. A pointer to a
struct may be automatically converted to a pointer to an
anonymous field when calling a function, in order to make the
types match. An anonymous struct field whose type is a typedef
name may be referred to using the typedef name.
C++
Improved experimental support for the
upcoming C++0x ISO C++ standard, including support for
constexpr (thanks to Gabriel Dos Reis and Jason Merrill),
nullptr (thanks to Magnus Fromreide), noexcept,
unrestricted unions, range-based for loops (thanks to Rodrigo Rivas Costa),
opaque enum declarations (thanks also to Rodrigo), implicitly deleted
functions and implicit move constructors.
When an extern declaration within a function does not match a
declaration in the enclosing context, G++ now properly declares the
name within the namespace of the function rather than the namespace
which was open just before the function definition
(c++/43145).
GCC now warns by default when casting integers to larger
pointer types. These warnings can be disabled with the option
-Wno-int-to-pointer-cast, which is now also available
in C++.
G++ no longer optimizes using the assumption that a value of
enumeration type will fall within the range specified by the standard,
since that assumption is easily violated with a conversion from integer
type (c++/43680).
The old behavior can be restored with -fstrict-enums.
The new -fnothrow-opt flag changes the semantics of
a throw() exception specification to match the proposed
semantics of the noexcept specification: just call
terminate if an exception tries to propagate out of a
function with such an exception specification. This dramatically
reduces or eliminates the code size overhead from adding the exception
specification.
The new -Wnoexcept flag will suggest adding
a noexcept qualifier to a function that the compiler can
tell doesn't throw if it would change the value of
a noexcept expression.
The -Wshadow option now warns if a local variable or
type declaration shadows another type in C++. Note that the compiler will
not warn if a local variable shadows a struct/class/enum, but will warn
if it shadows an explicit typedef.
When an identifier is not found in the current scope, G++ now
offers suggestions about which identifier might have been
intended.
G++ now issues clearer diagnostics for missing semicolons
after class, struct,
and union definitions.
G++ now issues clearer diagnostics for missing semicolons after
class member declarations.
G++ now issues clearer diagnostics when a colon is used in a
place where a double-colon was intended.
G++ no longer accepts mutable on reference members
(c++/33558).
Use -fpermissive to allow the old, non-conforming behaviour.
A few mangling fixes have been made, to attribute const/volatile on
function pointer types, decltype of a plain decl, and use of a
function parameter in the declaration of another parameter. By
default the compiler still uses the old mangling, but emits aliases
with the new mangling on targets that support strong aliases. Users
can switch over entirely to the new mangling
with -fabi-version=5 or -fabi-version=0.
-Wabi will now warn about code that uses the old
mangling.
G++ no longer allows objects of const-qualified type to be default
initialized unless the type has a user-declared default constructor.
Code that fails to compile can be fixed by providing an initializer e.g.
struct A { A(); };
struct B : A { };
const B b = B();
Use -fpermissive to allow the old, non-conforming behaviour.
Performance improvements to the
Debug
Mode, thanks to François Dumont.
Atomic operations used for reference-counting are annotated so that
they can be understood by race detectors such as Helgrind, see
Data
Race Hunting.
Most libstdc++ standard headers have been changed to no longer include
the cstddef header as an implementation detail. Code that
relied on that header being included as side-effect of including other
standard headers will need to include cstddef explicitly.
Fortran
On systems supporting the libquadmath library, GNU Fortran
now also supports a quad-precision, kind=16 floating-point
data type (REAL(16), COMPLEX(16)). As the data
type is not fully supported in hardware, calculations might be one to
two orders of magnitude slower than with the 4, 8 or 10 bytes
floating-point data types. This change does not affect systems which
support REAL(16) in hardware nor those which do not support
libquadmath.
Much improved compile time for large array constructors.
In order to reduce execution time and memory consumption, use of
temporary arrays in assignment expressions is avoided for
many cases. The compiler now reverses loops in order to avoid
generating a temporary array where possible.
Improved diagnostics, especially with
-fwhole-file.
The -fwhole-file flag is now enabled by default. This
improves code generation and diagnostics. It can be
disabled using the deprecated -fno-whole-file flag.
Support the generation of Makefile dependencies via the -M... flags of GCC; you may need to specify the
-cpp option in addition. The dependencies take
modules, Fortran's include, and CPP's #include
into account. Note: Using -M for the module path is no
longer supported, use -J instead.
The flag -Wconversion has been modified to only issue
warnings where a conversion leads to information loss. This drastically
reduces the number of warnings; -Wconversion is thus now
enabled with -Wall. The flag -Wconversion-extra
has been added and also warns about other conversions;
-Wconversion-extra typically issues a huge number of
warnings, most of which can be ignored.
A new command-line option -Wunused-dummy-argument warns
about unused dummy arguments and is included in -Wall.
Before, -Wunused-variable also warned about unused dummy
arguments.
Fortran 2003 support has been extended:
Improved support for polymorphism between libraries and
programs and for complicated inheritance patterns (cf. object-oriented programming).
Experimental support of the ASSOCIATE construct.
In pointer assignments it is now possible to specify the lower
bounds of the pointer and, for a rank-1 or a simply contiguous
data-target, to remap the bounds.
Automatic (re)allocation: In intrinsic assignments to
allocatable variables the left-hand side will be automatically
allocated (if unallocated) or reallocated (if the shape or type
parameter is different). To avoid the small performance penalty,
you can use a(:) = ... instead of a = ...
for arrays and character strings – or disable the feature using
-std=f95 or -fno-realloc-lhs.
Deferred type parameter: For scalar allocatable and pointer
variables the character length can be deferred.
Namelist variables with allocatable and pointer attribute and
nonconstant length type parameter are supported.
The STOP and the new ERROR STOP
statements now support all constant expressions.
Support for the CONTIGUOUS attribute.
Support for ALLOCATE with MOLD.
Support for the STORAGE_SIZE intrinsic inquiry
function.
Support of the NORM2 and PARITY
intrinsic functions.
The following bit intrinsics were added: POPCNT
and POPPAR for counting the number of 1 bits and
returning the parity; BGE, BGT,
BLE, and BLT for bitwise comparisons;
DSHIFTL and DSHIFTR for combined left
and right shifts, MASKL and MASKR for
simple left and right justified masks, MERGE_BITS
for a bitwise merge using a mask, SHIFTA,
SHIFTL and SHIFTR for shift operations,
and the transformational bit intrinsics IALL,
IANY and IPARITY.
Support of the EXECUTE_COMMAND_LINE intrinsic
subroutine.
Support for the IMPURE attribute for procedures,
which allows for ELEMENTAL procedures without the
restrictions of PURE.
Null pointers (including NULL()) and not
allocated variables can be used as actual argument to optional
non-pointer, non-allocatable dummy arguments, denoting an absent
argument.
Non-pointer variables with TARGET attribute can
be used as actual argument to POINTER dummies with
INTENT(IN)
Pointers including procedure pointers and those in a derived
type (pointer components) can now be initialized by a target
instead of only by NULL.
The EXIT statement (with construct-name) can
now be used to leave not only the DO but also the
ASSOCIATE, BLOCK, IF,
SELECT CASE and SELECT TYPE constructs.
Internal procedures can now be used as actual argument.
The named constants INTEGER_KINDS,
LOGICAL_KINDS, REAL_KINDS and
CHARACTER_KINDS of the intrinsic module
ISO_FORTRAN_ENV have been added; these arrays contain
the supported kind values for the respective types.
The module procedures C_SIZEOF of the intrinsic
module ISO_C_BINDINGS and COMPILER_VERSION
and COMPILER_OPTIONS of ISO_FORTRAN_ENV
have been implemented.
Minor changes: obsolescence diagnostics for ENTRY
was added for -std=f2008;
a line may start with a semicolon;
for internal and module procedures END can be used
instead of END SUBROUTINE and END
FUNCTION; SELECTED_REAL_KIND now also takes a
RADIX argument; intrinsic types are supported for
TYPE(intrinsic-type-spec); multiple type-bound
procedures can be declared in a single PROCEDURE
statement; implied-shape arrays are supported for named constants
(PARAMETER). The transformational, three argument
versions of BESSEL_JN and BESSEL_YN
were added – the elemental, two-argument version had been
added in GCC 4.4; note that the transformational functions use
a recurrence algorithm.
Go
Support for the Go programming
language has been added to GCC. It is not enabled by default
when you build GCC; use the --enable-languages
configure option to build it. The driver program for compiling Go
code is gccgo.
Go is currently known to work on GNU/Linux and RTEMS. Solaris
support is in progress. It may or may not work on other
platforms.
Java (GCJ)
Objective-C and Objective-C++
The -fobjc-exceptions flag is now required to
enable Objective-C exception and synchronization syntax
(introduced by the keywords @try,
@catch, @finally and
@synchronized).
A number of Objective-C 2.0 features and extensions are now
supported by GCC. These features are enabled by default; you can
disable them by using the new -fobjc-std=objc1
command-line option.
The Objective-C 2.0 dot-syntax is now supported. It is an
alternative syntax for using getters and setters;
object.count is automatically converted into
[object count] or [object setCount: ...]
depending on context; for example if (object.count >
0) is automatically compiled into the equivalent of
if ([object count] > 0) while object.count =
0; is automatically compiled into the equivalent ot
[object setCount: 0];. The dot-syntax can be used
with instance and class objects and with any setters or getters,
no matter if they are part of a declared property or not.
Objective-C 2.0 declared properties are now supported. They
are declared using the new @property keyword, and are
most commonly used in conjunction with the new Objective-C 2.0
dot-syntax. The nonatomic, readonly,
readwrite, assign, retain,
copy, setter and getter
attributes are all supported. Marking declared properties with
__attribute__ ((deprecated)) is supported too.
The Objective-C 2.0 @synthesize and
@dynamic keywords are supported.
@synthesize causes the compiler to automatically
synthesize a declared property, while @dynamic is
used to disable all warnings for a declared property for which no
implementation is provided at compile time. Synthesizing declared
properties requires runtime support in most useful cases; to be
able to use it with the GNU runtime, appropriate helper functions
have been added to the GNU Objective-C runtime ABI, and are
implemented by the GNU Objective-C runtime library shipped with
GCC.
The Objective-C 2.0 fast enumeration syntax is supported in
Objective-C. This is currently not yet available in
Objective-C++. Fast enumeration requires support in the runtime,
and such support has been added to the GNU Objective-C runtime
library (shipped with GCC).
The Objective-C 2.0 @optional keyword is
supported. It allows you to mark methods or properties in a
protocol as optional as opposed to required.
The Objective-C 2.0 @package keyword is
supported. It has currently the same effect as the
@public keyword.
Objective-C 2.0 method attributes are supported. Currently
the supported attributes are deprecated,
sentinel, noreturn and
format.
Objective-C 2.0 method argument attributes are supported. The
most widely used attribute is unused, to mark an
argument as unused in the implementation.
Objective-C 2.0 class and protocol attributes are supported.
Currently the only supported attribute is
deprecated.
Objective-C 2.0 class extensions are supported. A class
extension has the same syntax as a category declaration with no
category name, and the methods and properties declared in it are
added directly to the main class. It is mostly used as an
alternative to a category to add methods to a class without
advertising them in the public headers, with the advantage that
for class extensions the compiler checks that all the privately
declared methods are actually implemented.
As a result of these enhancements, GCC can now be used to
build Objective-C and Objective-C++ software that uses Foundation
and other important system frameworks with the NeXT runtime on
Darwin 9 and Darwin 10 (Mac OS X 10.5 and 10.6). Currently this is for
m32 code only.
Many bugs in the compiler have been fixed in this release; in
particular, LTO can now be used when compiling Objective-C and
Objective-C++ and the parser is much more robust in dealing with
invalid code.
Runtime Library (libobjc)
The GNU Objective-C runtime library now defines the macro
__GNU_LIBOBJC__ (with a value that is increased at
every release where there is any change to the API) in
objc/objc.h, making it easy to determine if the GNU
Objective-C runtime library is being used, and if so, which
version. Previous versions of the GNU Objective-C runtime library
(and other Objective-C runtime libraries such as the Apple one) do
not define this macro.
A new Objective-C 2.0 API, almost identical to the one
implemented by the Apple Objective-C runtime, has been implemented
in the GNU Objective-C runtime library. The new API hides the
internals of most runtime structures but provides a more extensive
set of functions to operate on them. It is much easier, for
example, to create or modify classes at runtime. The new API also
makes it easier to port software from Apple to GNU as almost no
changes should be required. The old API is still supported for
backwards compatibility; including the old
objc/objc-api.h header file automatically selects the
old API, while including the new objc/runtime.h
header file automatically selects the new API. Support for the
old API is being phased out and upgrading the software to use the
new API is strongly recommended. To check for the availability of
the new API, the __GNU_LIBOBJC__ macro can be used as
older versions of the GNU Objective-C runtime library, which do
not support the new API, do not define such a macro.
Runtime support for @synchronized has been added.
Runtime support for Objective-C 2.0 synthesized property
accessors has been added.
Runtime support for Objective-C 2.0 fast enumeration has been
added.
New Targets and Target Specific Improvements
ARM
GCC now supports the Cortex-M4 processor implementing
the v7-em version of the architecture using the option
-mcpu=cortex-m4.
Scheduling descriptions for the Cortex-M4, the Neon and
the floating point units of the Cortex-A9 and a pipeline
description for the Cortex-A5 have been added.
Synchronization primitives such as __sync_fetch_and_add
and friends are now inlined for supported architectures
rather than calling into a kernel helper function.
SSA loop prefetching is enabled by default for the
Cortex-A9 at -O3.
Several improvements were committed to improve code
generation for the ARM architecture including a rewritten
implementation for load and store multiples.
Several enhancements were committed to improve SIMD code
generation for NEON by adding support for widening instructions,
misaligned loads and stores, vector conditionals and
support for 64 bit arithmetic.
Support was added for the Faraday cores fa526, fa606te,
fa626te, fmp626te, fmp626 and fa726te and can be used with the
respective names as parameters to the -mcpu=
option.
Basic support was added for Cortex-A15 and is available through
-mcpu=cortex-a15.
GCC for AAPCS configurations now more closely adheres to the AAPCS
specification by enabling -fstrict-volatile-bitfields by
default.
IA-32/x86-64
The new -fsplit-stack option permits programs to
use a discontiguous stack. This is useful for threaded
programs, in that it is no longer necessary to specify the
maximum stack size when creating a thread. This feature is
currently only implemented for 32-bit and 64-bit x86 GNU/Linux
targets.
Support for emitting profiler counter calls before function
prologues. This is enabled via a new command-line option
-mfentry.
Optimization for the Intel Core 2 processors is now available through
the -march=core2 and -mtune=core2
options.
Support for Intel Core i3/i5/i7 processors is now available through
the -march=corei7 and -mtune=corei7
options.
Support for Intel Core i3/i5/i7 processors with AVX is now
available through the -march=corei7-avx and
-mtune=corei7-avx options.
Support for AMD Bobcat (family 14) processors is now available through
the -march=btver1 and -mtune=btver1
options.
The default setting (when not optimizing for size) for 32-bit
GNU/Linux and Darwin x86 targets has been changed to
-fomit-frame-pointer. The default can be reverted
to -fno-omit-frame-pointer by configuring GCC with
the --enable-frame-pointer configure option.
Darwin, FreeBSD, Solaris 2, MinGW and Cygwin now all support
__float128 on 32-bit and 64-bit x86 targets.
AVX floating-point arithmetic can now be enabled by default at
configure time with the new --with-fpmath=avx option.
The SSA loop prefetching pass is enabled when
using -O3 when optimizing for CPUs where prefetching
is beneficial (AMD CPUs newer than K6).
Support for TBM (Trailing Bit Manipulation) built-in functions
and code generation is available via -mtbm.
Support for AMD's BMI (Bit Manipulation) built-in functions and
code generation is available via -mbmi.
MicroBlaze
Support has been added for the Xilinx MicroBlaze softcore processor
(microblaze-elf) embedded target. This configurable processor is
supported on several Xilinx Spartan and Virtex FPGAs.
MIPS
GCC now supports the Loongson 3A processor. Its canonical
-march= and -mtune= name is
loongson3a.
MN10300 / AM33
The inline assembly register constraint "A" has
been renamed "c". This constraint is used to
select a floating-point register that can be used as the
destination of a multiply-accumulate instruction.
New inline assembly register constraints "A" and
"D" have been added. These constraint letters
resolve to all general registers when compiling for AM33, and
resolve to address registers only or data registers only when
compiling for MN10300.
The MDR register is represented in the compiler.
One can access the register via the "z" constraint
in inline assembly. It can be marked as clobbered or used as
a local register variable via the "mdr" name.
The compiler uses the RETF instruction if the
function does not modify the MDR register, so it
is important that inline assembly properly annotate any usage
of the register.
PowerPC/PowerPC64
GCC now supports the Applied Micro Titan processor
with -mcpu=titan.
The -mrecip option has been added, which indicates
whether the reciprocal and reciprocal square root instructions
should be used.
The -mveclibabi=mass option can be used to enable
the compiler to autovectorize mathematical functions using the
Mathematical Acceleration Subsystem library.
The -msingle-pic-base option has been added, which
instructs the compiler to avoid loading the PIC base register in
function prologues. The PIC base register must be initialized by
the runtime system.
The -mblock-move-inline-limit option has been
added, which enables the user to control the maximum size of
inlined memcpy calls and similar.
PowerPC64 GNU/Linux support for applications requiring a large
TOC section has been improved. A new command-line option,
-mcmodel=MODEL, controls this feature; valid values
for MODEL
are small, medium,
or large.
The altivec builtin functions vec_ld and vec_st
have been modified to generate the Altivec memory instructions
LVX and STVX, even if the -mvsx
option is used. In the initial GCC 4.5 release, these builtin functions
were changed to generate VSX memory reference instructions instead of
Altivec memory instructions, but there are differences between the two
instructions. If the VSX instruction set is available, you can now use
the new builtin functions vec_vsx_ld and vec_vsx_st
which always generates the VSX memory instructions.
The GCC compiler on AIX now defaults to a process layout with a
larger data space allowing larger programs to be compiled.
The GCC long double type on AIX 6.1 and above has reverted to 64 bit
double precision, matching the AIX XL compiler default, because of
missing C99 symbols required by the GCC runtime.
The default processor scheduling model and tuning for PowerPC64
GNU/Linux and for AIX 6.1 and above now is POWER7.
S/390, zSeries and System z9/z10, IBM zEnterprise z196
Support for the zEnterprise z196 processor has been added.
When using the -march=z196 option, the compiler
will generate code making use of the following instruction
facilities:
Conditional load/store
Distinct-operands
Floating-point-extension
Interlocked-access
Population-count
The -mtune=z196 option avoids the compare and
branch instructions as well as the load address instruction
with an index register as much as possible and performs
instruction scheduling appropriate for the new out-of-order
pipeline architecture.
When using the -m31 -mzarch options the generated
code still conforms to the 32-bit ABI but uses the general
purpose registers as 64-bit registers internally. This
requires a Linux kernel saving the whole 64-bit registers when
doing a context switch. Kernels providing that feature
indicate that by the 'highgprs' string
in /proc/cpuinfo.
The SSA loop prefetching pass is enabled when
using -O3.
SPARC
GCC now supports the LEON series of SPARC V8 processors. The
code generated by the compiler can either be tuned to it by means
of the --with-tune=leon configure option and
-mtune=leon compilation option, or the compiler can
be built for the sparc-leon-{elf,linux} and
sparc-leon3-{elf,linux} targets directly.
GCC has stopped sign/zero-extending parameter registers in the
callee for functions taking parameters with sub-word size in 32-bit
mode, since this is redundant with the specification of the ABI.
GCC has never done so in 64-bit mode since this is also redundant.
Operating Systems
Android
GCC now supports the Bionic C library and provides a convenient
way of building native libraries and applications for the Android
platform.
Refer to the documentation of the -mandroid and
-mbionic options for details on building native code.
At the moment, Android support is enabled only for ARM.
Darwin/Mac OS X
General
Initial support for CFString types has been
added. This allows GCC to build projects including the system
Core Foundation frameworks. The GCC Objective-C family
supports CFString "toll-free bridged" as per the Mac
OS X system tools. CFString is also recognized in the
context of format attributes and arguments (see the
documentation for format attributes for limitations).
At present, 8-bit character types are supported.
LTO-support. Darwin has benefited from ongoing work on
LTO; support for this is now stable and enabled by default.
Object file size reduction. The Darwin zeroed memory
allocators have been re-written to make more use of
.zerofill sections. For non-debug code, this can
reduce object file size significantly.
x86 Architecture
The -mdynamic-no-pic option has been
enabled. Code supporting -mdynamic-no-pic
optimization has been added and is applicable to -m32
builds. The compiler bootstrap uses the option where
appropriate.
The default value for -mtune= has been
changed. Since Darwin systems are primarily Xeon, Core-2 or
similar the default tuning has been changed to
-mtune=core2.
PPC Architecture
Darwin64 ABI. Several significant bugs have been fixed,
such that GCC now produces code compatible with the Darwin64
PowerPC ABI.
libffi and boehm-gc. The Darwin ports of the libffi and
boehm-gc libraries have been upgraded to include a Darwin64
implementation. This means that powerpc*-*-darwin9 platforms may
now, for example, build Java applications with -m64
enabled.
Plug-in support has been enabled.
The -fsection-anchors option is now available
although, presently, not heavily tested.
Solaris 2
New Features
Support symbol versioning with the Sun linker.
Allow libstdc++ to leverage full ISO C99 support on
Solaris 10+.
Support thread-local storage (TLS) with the Sun assembler on
Solaris 2/x86.
Support TLS on Solaris 8/9 if prerequisites are met.
Support COMDAT group with the GNU assembler and recent Sun
linker.
Support the Sun assembler visibility syntax.
Default Solaris 2/x86 to -march=pentium4 (Solaris
10+) resp. -march=pentiumpro (Solaris 8/9).
Don't use SSE on Solaris 8/9 x86 by default.
Enable 128-bit long double (__float128) support on
Solaris 2/x86.
ABI Change
Change the ABI for returning 8-byte vectors like
__m64 in MMX registers on Solaris 10+/x86 to match the
Sun Studio 12.1+ compilers. This is an incompatible change.
If you use such types, you must either recompile all your code with
the new compiler or use the new -mvect8-ret-in-mem
option to remain compatible with previous versions of GCC and
Sun Studio.
Windows x86/x86_64
Initial support for decimal floating point.
Support for the __thiscall calling-convention.
Support for hot-patchable function prologues via the
ms_hook_prologue attribute for x86_64 in addition to 32-bit x86.
Improvements of stack-probing and stack-allocation mechanisms.
Support of push/pop-macro pragma as preprocessor command.
With #pragma push_macro("macro-name") the
current definition of macro-name is saved and can be
restored with #pragma pop_macro("macro-name")
to its saved definition.
Enable 128-bit long double (__float128) support on
MinGW and Cygwin.
Documentation improvements
Other significant improvements
Installation changes
An install-stripmake target is provided
that installs stripped executables, and may install libraries with
unneeded or debugging sections stripped.
On Power7 systems, there is a potential problem if you build the GCC
compiler with a host compiler using options that enables the VSX
instruction set generation. If the host compiler has been patched so that
the vec_ld and vec_st builtin functions
generate Altivec memory instructions instead of VSX memory instructions,
then you should be able to build the compiler with VSX instruction
generation.
Changes for GCC Developers
Note: these changes concern developers that develop GCC itself or
software that integrates with GCC, such as plugins, and not the
general GCC users.
The gengtype utility, which previously was internal to
the GCC build process, has been enchanced to provide GC root
information for plugins as necessary.
The old GC allocation interface of ggc_alloc and
friends was replaced with a type-safe alternative.
For questions related to the use of GCC, please consult these web
pages and the GCC manuals. If
that fails, the gcc-help@gcc.gnu.org
mailing list might help.
Please send comments on these web pages and the development of GCC to our
developer list at gcc@gcc.gnu.org.
All of our lists have
public archives.
Copyright (C) Free Software Foundation, Inc.
Verbatim copying and distribution of this entire article is
permitted in any medium, provided this notice is preserved.
