GCC 4.9 Release Series — Changes, New Features, and Fixes - GNU Project

GCC 4.9 Release Series
Changes, New Features, and Fixes

Caveats

The mudflap run time checker has been removed. The mudflap options remain, but do nothing.
Support for a number of older systems and recently unmaintained or untested target ports of GCC has been declared obsolete in GCC 4.9. Unless there is activity to revive them, the next release of GCC will have their sources permanently removed.

The following ports for individual systems on particular architectures have been obsoleted:
- Solaris 9 (*-*-solaris2.9). Details can be found in the announcement.
On AArch64, the singleton vector types int64x1_t, uint64x1_t and float64x1_t exported by arm_neon.h are defined to be the same as their base types. This results in incorrect application of parameter passing rules to arguments of types int64x1_t and uint64x1_t, with respect to the AAPCS64 ABI specification. In addition, names of C++ functions with parameters of these types (including float64x1_t) are not mangled correctly. The current typedef declarations also unintentionally allow implicit casting between singleton vector types and their base types. These issues will be resolved in a near future release. See PR60825 for more information.

More information on porting to GCC 4.9 from previous versions of GCC can be found in the porting guide for this release.

General Optimizer Improvements

AddressSanitizer, a fast memory error detector, is now available on ARM.
UndefinedBehaviorSanitizer (ubsan), a fast undefined behavior detector, has been added and can be enabled via -fsanitize=undefined. Various computations will be instrumented to detect undefined behavior at runtime. UndefinedBehaviorSanitizer is currently available for the C and C++ languages.
Link-time optimization (LTO) improvements:
- Type merging was rewritten. The new implementation is significantly faster and uses less memory.
- Better partitioning algorithm resulting in less streaming during link time.
- Early removal of virtual methods reduces the size of object files and improves link-time memory usage and compile time.
- Function bodies are now loaded on-demand and released early improving overall memory usage at link time.
- C++ hidden keyed methods can now be optimized out.
- When using a linker plugin, compiling with the -flto option now generates slim object files (.o) which only contain intermediate language representation for LTO. Use -ffat-lto-objects to create files which contain additionally the object code. To generate static libraries suitable for LTO processing, use gcc-ar and gcc-ranlib; to list symbols from a slim object file use gcc-nm. (This requires that ar, ranlib and nm have been compiled with plugin support.)
Memory usage building Firefox with debug enabled was reduced from 15GB to 3.5GB; link time from 1700 seconds to 350 seconds.
Inter-procedural optimization improvements:
- New type inheritance analysis module improving devirtualization. Devirtualization now takes into account anonymous name-spaces and the C++11 final keyword.
- New speculative devirtualization pass (controlled by -fdevirtualize-speculatively.
- Calls that were speculatively made direct are turned back to indirect where direct call is not cheaper.
- Local aliases are introduced for symbols that are known to be semantically equivalent across shared libraries improving dynamic linking times.
Feedback directed optimization improvements:
- Profiling of programs using C++ inline functions is now more reliable.
- New time profiling determines typical order in which functions are executed.
- A new function reordering pass (controlled by -freorder-functions) significantly reduces startup time of large applications. Until binutils support is completed, it is effective only with link-time optimization.
- Feedback driven indirect call removal and devirtualization now handle cross-module calls when link-time optimization is enabled.

New Languages and Language specific improvements

Version 4.0 of the OpenMP specification is now supported in the C and C++ compilers and starting with the 4.9.1 release also in the Fortran compiler. The new -fopenmp-simd option can be used to enable OpenMP's SIMD directives while ignoring other OpenMP directives. The new -fsimd-cost-model= option permits to tune the vectorization cost model for loops annotated with OpenMP and Cilk Plus simd directives. -Wopenmp-simd warns when the current cost model overrides simd directives set by the user.
The -Wdate-time option has been added for the C, C++ and Fortran compilers, which warns when the __DATE__, __TIME__ or __TIMESTAMP__ macros are used. Those macros might prevent bit-wise-identical reproducible compilations.

Ada

GNAT switched to Ada 2012 instead of Ada 2005 by default.

Support for colorizing diagnostics emitted by GCC has been added. The -fdiagnostics-color=auto will enable it when outputting to terminals, -fdiagnostics-color=always unconditionally. The GCC_COLORS environment variable can be used to customize the colors or disable coloring. If GCC_COLORS variable is present in the environment, the default is -fdiagnostics-color=auto, otherwise -fdiagnostics-color=never.
Sample diagnostics output:

    $ g++ -fdiagnostics-color=always -S -Wall test.C
    test.C: In function ‘int foo()’:
    test.C:1:14: warning: no return statement in function returning non-void [-Wreturn-type]
     int foo () { }
                  ^
    test.C:2:46: error: template instantiation depth exceeds maximum of 900 (use -ftemplate-depth= to increase the maximum) instantiating ‘struct X<100>’
     template <int N> struct X { static const int value = X<N-1>::value; }; template struct X<1000>;
                                                  ^
    test.C:2:46:   recursively required from ‘const int X<999>::value’
    test.C:2:46:   required from ‘const int X<1000>::value’
    test.C:2:88:   required from here

    test.C:2:46: error: incomplete type ‘X<100>’ used in nested name specifier

With the new #pragma GCC ivdep, the user can assert that there are no loop-carried dependencies which would prevent concurrent execution of consecutive iterations using SIMD (single instruction multiple data) instructions.
Support for Cilk Plus has been added and can be enabled with the -fcilkplus option. Cilk Plus is an extension to the C and C++ languages to support data and task parallelism. The present implementation follows ABI version 1.2; all features but _Cilk_for have been implemented.

C

ISO C11 atomics (the _Atomic type specifier and qualifier and the <stdatomic.h> header) are now supported.
ISO C11 generic selections (_Generic keyword) are now supported.
ISO C11 thread-local storage (_Thread_local, similar to GNU C __thread) is now supported.
ISO C11 support is now at a similar level of completeness to ISO C99 support: substantially complete modulo bugs, extended identifiers (supported except for corner cases when -fextended-identifiers is used), floating-point issues (mainly but not entirely relating to optional C99 features from Annexes F and G) and the optional Annexes K (Bounds-checking interfaces) and L (Analyzability).
A new C extension __auto_type provides a subset of the functionality of C++11 auto in GNU C.

C++

The G++ implementation of C++1y return type deduction for normal functions has been updated to conform to N3638, the proposal accepted into the working paper. Most notably, it adds decltype(auto) for getting decltype semantics rather than the template argument deduction semantics of plain auto:
```
int& f();
         auto  i1 = f(); // int
decltype(auto) i2 = f(); // int&
```
G++ supports C++1y lambda capture initializers:
```
[x = 42]{ ... };
```
Actually, they have been accepted since GCC 4.5, but now the compiler doesn't warn about them with -std=c++1y, and supports parenthesized and brace-enclosed initializers as well.
G++ supports C++1y variable length arrays. G++ has supported GNU/C99-style VLAs for a long time, but now additionally supports initializers and lambda capture by reference. In C++1y mode G++ will complain about VLA uses that are not permitted by the draft standard, such as forming a pointer to VLA type or applying sizeof to a VLA variable. Note that it now appears that VLAs will not be part of C++14, but will be part of a separate document and then perhaps C++17.
```
void f(int n) {
  int a[n] = { 1, 2, 3 }; // throws std::bad_array_length if n < 3
  [&a]{ for (int i : a) { cout << i << endl; } }();
  &a; // error, taking address of VLA
}
```

G++ supports the C++1y [[deprecated]] attribute modulo bugs in the underlying [[gnu::deprecated]] attribute. Classes and functions can be marked deprecated and a diagnostic message added:

class A;
int bar(int n);
#if __cplusplus > 201103
class [[deprecated("A is deprecated in C++14; Use B instead")]] A;
[[deprecated("bar is unsafe; use foo() instead")]]
int bar(int n);

int foo(int n);
class B;
#endif
A aa; // warning: 'A' is deprecated : A is deprecated in C++14; Use B instead
int j = bar(2); // warning: 'int bar(int)' is deprecated : bar is unsafe; use foo() instead

G++ supports C++1y digit separators. Long numeric literals can be subdivided with a single quote ' to enhance readability:

int i = 1048576;
int j = 1'048'576;
int k = 0x10'0000;
int m = 0'004'000'000;
int n = 0b0001'0000'0000'0000'0000'0000;

double x = 1.602'176'565e-19;
double y = 1.602'176'565e-1'9;

G++ supports C++1y generic (polymorphic) lambdas.

// a functional object that will increment any type
auto incr = [](auto x) { return x++; };

As a GNU extension, G++ supports explicit template parameter syntax for generic lambdas. This can be combined in the expected way with the standard auto syntax.
```
// a functional object that will add two like-type objects
auto add = [] <typename T> (T a, T b) { return a + b; };
```
G++ supports unconstrained generic functions as specified by §4.1.2 and §5.1.1 of N3889: Concepts Lite Specification. Briefly, auto may be used as a type-specifier in a parameter declaration of any function declarator in order to introduce an implicit function template parameter, akin to generic lambdas.
```
// the following two function declarations are equivalent
auto incr(auto x) { return x++; }
template <typename T>
auto incr(T x) { return x++; }
```

Runtime Library (libstdc++)

Improved support for C++11, including:
- support for <regex>;
- The associative containers in <map> and <set> and the unordered associative containers in <unordered_map> and <unordered_set> meet the allocator-aware container requirements;
Improved experimental support for the upcoming ISO C++ standard, C++14, including:
- fixing constexpr member functions without const;
- implementation of the std::exchange() utility function;
- addressing tuples by type;
- implemention of std::make_unique;
- implemention of std::shared_lock;
- making std::result_of SFINAE-friendly;
- adding operator() to std::integral_constant;
- adding user-defined literals for standard library types std::basic_string, std::chrono::duration, and std::complex;
- adding two range overloads to non-modifying sequence oprations std::equal and std::mismatch;
- adding IO manipulators for quoted strings;
- adding constexpr members to <utility>, <complex>, <chrono>, and some containers;
- adding compile-time std::integer_sequence;
- adding cleaner transformation traits;
- making <functional>s operator functors easier to use and more generic;
An implementation of std::experimental::optional.
An implementation of std::experimental::string_view.
The non-standard function std::copy_exception has been deprecated and will be removed in a future version. std::make_exception_ptr should be used instead.

Fortran

Compatibility notice:
- Module files: The version of the module files (.mod) has been incremented; additionally, module files are now compressed. Fortran MODULEs compiled by earlier GCC versions have to be recompiled, when they are USEd by files compiled with GCC 4.9. GCC 4.9 is not able to read .mod files of earlier GCC versions; attempting to do so gives an error message. Note: The ABI of the produced assembler data itself has not changed: object files and libraries are fully compatible with older versions (except as stated below).
- ABI changes:
  - The argument passing ABI has changed for scalar dummy arguments of type INTEGER, REAL, COMPLEX and LOGICAL, which have both the VALUE and the OPTIONAL attributes.
  - To support finalization the virtual table associated with polymorphic variables has changed. Code containing CLASS should be recompiled, including all files which define derived types involved in the type definition used by polymorphic variables. (Note: Due to the incremented module version, trying to mix old code with new code will usually give an error message.)
- GNU Fortran no longer deallocates allocatable variables or allocatable components of variables declared in the main program. Since Fortran 2008, the standard explicitly states that variables declared in the Fortran main program automatically have the SAVE attribute.
- When opening files, the close-on-exec flag is set if the system supports such a feature. This is generally considered good practice these days, but if there is a need to pass file descriptors to child processes the parent process must now remember to clear the close-on-exec flag by calling fcntl(), e.g. via ISO_C_BINDING, before executing the child process.
The deprecated command-line option -fno-whole-file has been removed. (-fwhole-file is the default since GCC 4.6.) -fwhole-file/-fno-whole-file continue to be accepted but do not influence the code generation.
The compiler no longer unconditionally warns about DO loops with zero iterations. This warning is now controlled by the -Wzerotrip option, which is implied by -Wall.
The new NO_ARG_CHECK attribute of the !GCC$ directive can be used to disable the type-kind-rank (TKR) argument check for a dummy argument. The feature is similar to ISO/IEC TS 29133:2012's TYPE(*), except that it additionally also disables the rank check. Variables with NO_ARG_CHECK have to be dummy arguments and may only be used as argument to ISO_C_BINDING's C_LOC and as actual argument to another NO_ARG_CHECK dummy argument; also the other constraints of TYPE(*) apply. The dummy arguments should be declared as scalar or assumed-size variable of type type(*) (recommended) – or of type integer, real, complex or logical. With NO_ARG_CHECK, a pointer to the data without further type or shape information is passed, similar to C's void*. Note that also TS 29113's type(*),dimension(..) accepts arguments of any type and rank; contrary to NO_ARG_CHECK assumed-rank arguments pass an array descriptor which contains the array shape and stride of the argument.
Fortran 2003:
- Finalization is now supported. It is currently only done for a subset of those situations in which it should occur.
- Experimental support for scalar character components with deferred length (i.e. allocatable string length) in derived types has been added. (Deferred-length character variables are supported since GCC 4.6.)
Fortran 2008:
- When STOP or ERROR STOP are used to terminate the execution and any exception (but inexact) is signaling, a warning is printed to ERROR_UNIT, indicating which exceptions are signaling. The -ffpe-summary= command-line option can be used to fine-tune for which exceptions the warning should be shown.
- Rounding on input (READ) is now handled on systems where strtod honours the rounding mode. (For output, rounding is supported since GCC 4.5.) Note that for input, the compatible rounding mode is handled as nearest (i.e., rounding to an even least significant [cf. IEC 60559:1989] for a tie, while compatible rounds away from zero in that case).

Go

GCC 4.9 provides a complete implementation of the Go 1.2.1 release.

New Targets and Target Specific Improvements

AArch64

The ARMv8-A crypto and CRC instructions are now supported through intrinsics. These are enabled when the architecture supports these and are available through the -march=armv8-a+crc and -march=armv8-a+crypto options.
Initial support for ILP32 has now been added to the compiler. This is now available through the command-line option -mabi=ilp32. Support for ILP32 is considered experimental as the ABI specification is still beta.
Coverage of more of the ISA including the SIMD extensions has been added. The Advanced SIMD intrinsics have also been improved.
The new local register allocator (LRA) is now on by default for the AArch64 back end.
The REE (Redundant extension elimination) pass has now been enabled by default for the AArch64 back end.
Tuning for the Cortex-A53 and Cortex-A57 has been improved.
Initial big.LITTLE tuning support for the combination of Cortex-A57 and Cortex-A53 was added through the -mcpu=cortex-a57.cortex-a53 option.
A number of structural changes have been made to both the ARM and AArch64 back ends to facilitate improved code-generation.
As of GCC 4.9.2 a workaround for the ARM Cortex-A53 erratum 835769 has been added and can be enabled by giving the -mfix-cortex-a53-835769 option. Alternatively it can be enabled by default by configuring GCC with the --enable-fix-cortex-a53-835769 option.

ARC

A port for Synopsys Designware ARC has been contributed by Embecosm and Synopsys Inc.

ARM

Use of Advanced SIMD (Neon) for 64-bit scalar computations has been disabled by default. This was found to generate better code in only a small number of cases. It can be turned back on with the -mneon-for-64bits option.
Further support for the ARMv8-A architecture, notably implementing the restriction around IT blocks in the Thumb32 instruction set has been added. The -mrestrict-it option can be used with -march=armv7-a or the -march=armv7ve options to make code generation fully compatible with the deprecated instructions in ARMv8-A.
Support has now been added for the ARMv7ve variant of the architecture. This can be used by the -march=armv7ve option.
The ARMv8-A crypto and CRC instructions are now supported through intrinsics and are available through the -march=armv8-a+crc and mfpu=crypto-neon-fp-armv8 options.
LRA is now on by default for the ARM target. This can be turned off using the -mno-lra option. This option is a purely transitionary command-line option and will be removed in a future release. We are interested in any bug reports regarding functional and performance regressions with LRA.
A new option -mslow-flash-data to improve performance of programs fetching data on slow flash memory has now been introduced for the ARMv7-M profile cores.
A new option -mpic-data-is-text-relative for targets that allows data segments to be relative to text segments has been added. This is on by default for all targets except VxWorks RTP.
A number of infrastructural changes have been made to both the ARM and AArch64 back ends to facilitate improved code-generation.
GCC now supports Cortex-A12 and the Cortex-R7 through the -mcpu=cortex-a12 and -mcpu=cortex-r7 options.
GCC now has tuning for the Cortex-A57 and Cortex-A53 through the -mcpu=cortex-a57 and -mcpu=cortex-a53 options.
Initial big.LITTLE tuning support for the combination of Cortex-A57 and Cortex-A53 was added through the -mcpu=cortex-a57.cortex-a53 option. Similar support was added for the combination of Cortex-A15 and Cortex-A7 through the -mcpu=cortex-a15.cortex-a7 option.
Further performance optimizations for the Cortex-A15 and the Cortex-M4 have been added.
A number of code generation improvements for Thumb2 to reduce code size when compiling for the M-profile processors.

AVR

A new command-line option -mfract-convert-truncate has been added. It allows compiler to use truncation instead of rounding towards zero for fractional fixed-point types.

IA-32/x86-64

-mfpmath=sse is now implied by -ffast-math on all targets where SSE2 is supported.
Intel AVX-512 support was added to GCC. That includes inline assembly support, new registers and extending existing ones, new intrinsics (covered by corresponding testsuite), and basic autovectorization. AVX-512 instructions are available via the following GCC switches: AVX-512 foundation instructions: -mavx512f, AVX-512 prefetch instructions: -mavx512pf, AVX-512 exponential and reciprocal instructions: -mavx512er, AVX-512 conflict detection instructions: -mavx512cd.
It is now possible to call x86 intrinsics from select functions in a file that are tagged with the corresponding target attribute without having to compile the entire file with the -mxxx option. This improves the usability of x86 intrinsics and is particularly useful when doing Function Multiversioning.
GCC now supports the new Intel microarchitecture named Silvermont through -march=silvermont.
GCC now supports the new Intel microarchitecture named Broadwell through -march=broadwell.
Optimizing for other Intel microarchitectures have been renamed to -march=nehalem, westmere, sandybridge, ivybridge, haswell, bonnell.
-march=generic has been retuned for better support of Intel core and AMD Bulldozer architectures. Performance of AMD K7, K8, Intel Pentium-M, and Pentium4 based CPUs is no longer considered important for generic.
-mtune=intel can now be used to generate code running well on the most current Intel processors, which are Haswell and Silvermont for GCC 4.9.
Support to encode 32-bit assembly instructions in 16-bit format is now available through the -m16 command-line option.
Better inlining of memcpy and memset that is aware of value ranges and produces shorter alignment prologues.
-mno-accumulate-outgoing-args is now honored when unwind information is output. Argument accumulation is also now turned off for portions of programs optimized for size.
Support for new AMD family 15h processors (Excavator core) is now available through the -march=bdver4 and -mtune=bdver4 options.

MSP430

A new command-line option -mcpu= has been added to the MSP430 back end. This option is used to specify the ISA to be used. Accepted values are msp430 (the default), msp430x and msp430xv2. The ISA is no longer deduced from the -mmcu= option as there are far too many different MCU names. The -mmcu= option is still supported, and this is still used to select linker scripts and generate a C preprocessor symbol that will be recognised by the msp430.h header file.

NDS32

A new nds32 port supports the 32-bit architecture from Andes Technology Corporation.
The port provides initial support for the V2, V3, V3m instruction set architectures.

Nios II

A port for the Altera Nios II has been contributed by Mentor Graphics.

PowerPC / PowerPC64 / RS6000

GCC now supports Power ISA 2.07, which includes support for Hardware Transactional Memory (HTM), Quadword atomics and several VMX and VSX additions, including Crypto, 64-bit integer, 128-bit integer and decimal integer operations.
Support for the POWER8 processor is now available through the -mcpu=power8 and -mtune=power8 options.
The libitm library has been modified to add a HTM fastpath that automatically uses POWER's HTM hardware instructions when it is executing on a HTM enabled processor.
Support for the new powerpc64le-linux platform has been added. It defaults to generating code that conforms to the ELFV2 ABI.

S/390, System z

Support for the Transactional Execution Facility included with the IBM zEnterprise zEC12 processor has been added. A set of GCC style builtins as well as XLC style builtins are provided. The builtins are enabled by default when using the -march=zEC12 option but can explicitly be disabled with -mno-htm. Using the GCC builtins also libitm supports hardware transactions on S/390.
The hotpatch features allows to prepare functions for hotpatching. A certain amount of bytes is reserved before the function entry label plus a NOP is inserted at its very beginning to implement a backward jump when applying a patch. The feature can either be enabled per compilation unit via the command-line option -mhotpatch or per function using the hotpatch attribute.
The shrink wrap optimization is now supported on S/390 and enabled by default.
A major rework of the routines to determine which registers need to be saved and restored in function prologue/epilogue now allow to use floating point registers as save slots. This will happen for certain leaf function with -march=z10 or higher.
The LRA rtl pass replaces reload by default on S/390.

RX

The port now allows to specify the RX100, RX200, and RX600 processors with the command-line options -mcpu=rx100, -mcpu=rx200 and -mcpu=rx600.

SH

Minor improvements to code generated for integer arithmetic and code that involves the T bit.
Added support for the SH2A clips and clipu instructions. The compiler will now try to utilize them for min/max expressions such as max (-128, min (127, x)).
Added support for the cmp/str instruction through built-in functions such as __builtin_strlen. When not optimizing for size, the compiler will now expand calls to e.g. strlen as an inlined sequences which utilize the cmp/str instruction.
Improved code generated around volatile memory loads and stores.
The option -mcbranchdi has been deprecated. Specifying it will result in a warning and will not influence code generation.
The option -mcmpeqdi has been deprecated. Specifying it will result in a warning and will not influence code generation.

This is the list of problem reports (PRs) from GCC's bug tracking system that are known to be fixed in the 4.9.1 release. This list might not be complete (that is, it is possible that some PRs that have been fixed are not listed here).

Version 4.0 of the OpenMP specification is supported even in Fortran, not just C and C++.

GCC 4.9 Release Series — Changes, New Features, and Fixes - GNU Project