GCC 4.0 Release Series
Changes, New Features, and Fixes

The latest release in the 4.0 release series is GCC 4.0.4.

Caveats

• GCC now generates location lists by default when compiling with debug info and optimization.
  • GDB 6.0 and older crashes when it sees location lists. GDB 6.1 or later is needed to debug binaries containing location lists.
  • When you are trying to view a value of a variable in a part of a function where it has no location (for example when the variable is no longer used and thus its location was used for something else) GDB will say that it is not available.
  You can disable generating location lists by -fno-var-tracking.
• GCC no longer accepts the -fwritable-strings option. Use named character arrays when you need a writable string.
• The options -freduce-all-givs and -fmove-all-movables have been discontinued. They were used to circumvent a shortcoming in the heuristics of the old loop optimization code with respect to common Fortran constructs. The new (tree) loop optimizer works differently and doesn't need those work-arounds.
• The graph-coloring register allocator, formerly enabled by the option -fnew-ra, has been discontinued.
• -I- has been deprecated. -iquote is meant to replace the need for this option.
• The MIPS -membedded-pic and -mrnames options have been removed.
• All MIPS targets now require the GNU assembler. In particular, IRIX configurations can no longer use the MIPSpro assemblers, although they do still support the MIPSpro linkers.
• The SPARC option -mflat has been removed.
• English-language diagnostic messages will now use Unicode quotation marks in UTF-8 locales. (Non-English messages already used the quotes appropriate for the language in previous releases.) If your terminal does not support UTF-8 but you are using a UTF-8 locale (such locales are the default on many GNU/Linux systems) then you should set LC_CTYPE=C in the environment to disable that locale. Programs that parse diagnostics and expect plain ASCII English-language messages should set LC_ALL=C. See Markus Kuhn's explanation of Unicode quotation marks for more information.
• The specs file is no longer installed on most platforms. Most users will be totally unaffected. However, if you are accustomed to editing the specs file yourself, you will now have to use the -dumpspecs option to generate the specs file, and then edit the resulting file.

General Optimizer Improvements

• The tree ssa branch has been merged. This merge has brought in a completely new optimization framework based on a higher level intermediate representation than the existing RTL representation. Numerous new code transformations based on the new framework are available in GCC 4.0, including:
  
  • Scalar replacement of aggregates
  • Constant propagation
  • Value range propagation
  • Partial redundancy elimination
  • Load and store motion
  • Strength reduction
  • Dead store elimination
  • Dead and unreachable code elimination
  • Autovectorization
  • Loop interchange
  • Tail recursion by accumulation
  
  Many of these passes outperform their counterparts from previous GCC releases.
• Swing Modulo Scheduling (SMS). An RTL level instruction scheduling optimization intended for loops that perform heavy computations.

New Languages and Language specific improvements

C family

• The sentinel attribute has been added to GCC. This function attribute allows GCC to warn when variadic functions such as execl are not NULL terminated. See the GCC manual for a complete description of its behavior.
• Given __attribute__((alias("target"))) it is now an error if target is not a symbol, defined in the same translation unit. This also applies to aliases created by #pragma weak alias=target. This is because it's meaningless to define an alias to an undefined symbol. On Solaris, the native assembler would have caught this error, but GNU as does not.

C and Objective-C

• The -Wstrict-aliasing=2 option has been added. This warning catches all unsafe cases, but it may also give a warning for some cases that are safe.
• The cast-as-lvalue, conditional-expression-as-lvalue and compound-expression-as-lvalue extensions, which were deprecated in 3.3.4 and 3.4, have been removed.
• The -fwritable-strings option, which was deprecated in 3.4, has been removed.
• #pragma pack() semantics have been brought closer to those used by other compilers. This also applies to C++.
• Taking the address of a variable with register storage is invalid in C. GCC now issues an error instead of a warning.
• Arrays of incomplete element type are invalid in C. GCC now issues an error for such arrays. Declarations such as extern struct s x[]; (where struct s has not been defined) can be moved after the definition of struct s. Function parameters declared as arrays of incomplete type can instead be declared as pointers.

C++

• When compiling without optimizations (-O0), the C++ frontend is much faster than in any previous versions of GCC. Independent testers have measured speed-ups up to 25% in real-world production code, compared to the 3.4 family (which was already the fastest version to date). Upgrading from older versions might show even bigger improvements.
• ELF visibility attributes can now be applied to a class type, so that it affects every member function of a class at once, without having to specify each individually:

  class __attribute__ ((visibility("hidden"))) Foo
  {
     int foo1();
     void foo2();
  };

  The syntax is deliberately similar to the __declspec() system used by Microsoft Windows based compilers, allowing cross-platform projects to easily reuse their existing macro system for denoting exports and imports. By explicitly marking internal classes never used outside a binary as hidden, one can completely avoid PLT indirection overheads during their usage by the compiler. You can find out more about the advantages of this at http://people.redhat.com/drepper/dsohowto.pdf
• The -fvisibility-inlines-hidden option has been added which marks all inlineable functions as having hidden ELF visibility, thus removing their symbol and typeinfo from the exported symbol table of the output ELF binary. Using this option can reduce the exported symbol count of template-heavy code by up to 40% with no code change at all, thus notably improving link and load times for the binary as well as a reduction in size of up to 10%. Also, check the new -fvisibility option.
• The compiler now uses the library interface specified by the C++ ABI for thread-safe initialization of function-scope static variables. Most users should leave this alone, but embedded programmers may want to disable this by specifying -fno-threadsafe-statics for a small savings in code size.
• Taking the address of an explicit register variable is no longer supported. Note that C++ allows taking the address of variables with register storage so this will continue to compile with a warning. For example, assuming that r0 is a machine register:

  register int foo asm ("r0");
  register int bar;
  &foo; // error, no longer accepted
  &bar; // OK, with a warning

• G++ has an undocumented extension to virtual function covariancy rules that allowed the overrider to return a type that was implicitly convertable to the overridden function's return type. For instance a function returning void * could be overridden by a function returning T *. This is now deprecated and will be removed in a future release.
• The G++ minimum and maximum operators (<? and >?) and their compound forms (<?=) and >?=) have been deprecated and will be removed in a future version. Code using these operators should be modified to use std::min and std::max instead.
• Declaration of nested classes of class templates as friends are supported:

  template <typename T> struct A {
    class B {};
  };
  class C {
    template <typename T> friend class A<T>::B;
  };

  This complements the feature member functions of class templates as friends introduced in GCC 3.4.0.
• When declaring a friend class using an unqualified name, classes outside the innermost non-class scope are not searched:

  class A;
  namespace N {
    class B {
      friend class A;   // Refer to N::A which has not been declared yet
                        // because name outside namespace N are not searched
      friend class ::A; // Refer to ::A
    };
  }

  Hiding the friend name until declaration is still not implemented.
• Friends of classes defined outside their namespace are correctly handled:

  namespace N {
    class A;
  }
  class N::A {
    friend class B; // Refer to N::B in GCC 4.0.0
                    // but ::B in earlier versions of GCC
  };

Runtime Library (libstdc++)

• Optimization work:
  • Added efficient specializations of istream functions for char and wchar_t.
  • Further performance tuning of strings, in particular wrt single-char append and getline.
  • iter_swap - and therefore most of the mutating algorithms - now makes an unqualified call to swap when the value_type of the two iterators is the same.
• A large subset of the features in Technical Report 1 (TR1 for short) is experimentally delivered (i.e., no guarantees about the implementation are provided. In particular it is not promised that the library will remain link-compatible when code using TR1 is used):
  • General utilities such as reference_wrapper and shared_ptr.
  • Function objects, i.e., result_of, mem_fn, bind, function.
  • Support for metaprogramming.
  • New containers such as tuple, array, unordered_set, unordered_map, unordered_multiset, unordered_multimap.
• As usual, many bugs have been fixed and LWG resolutions implemented for the first time (e.g., DR 409).

Java

• In order to prevent naming conflicts with other implementations of these tools, some GCJ binaries have been renamed:
  • rmic is now grmic,
  • rmiregistry is now grmiregistry, and
  • jar is now fastjar.
  In particular, these names were problematic for the jpackage.org packaging conventions which install symlinks in /usr/bin that point to the preferred versions of these tools.
• The -findirect-dispatch argument to the compiler now works and generates code following a new "binary compatibility" ABI. Code compiled this way follows the binary compatibility rules of the Java Language Specification.
• libgcj now has support for using GCJ as a JIT, using the gnu.gcj.jit family of system properties.
• libgcj can now find a shared library corresponding to the bytecode representation of a class. See the documentation for the new gcj-dbtool program, and the new gnu.gcj.precompiled.db.path system property.
• There have been many improvements to the class library. Here are some highlights:
  • Much more of AWT and Swing exist.
  • Many new packages and classes were added, including java.util.regex, java.net.URI, javax.crypto, javax.crypto.interfaces, javax.crypto.spec, javax.net, javax.net.ssl, javax.security.auth, javax.security.auth.callback, javax.security.auth.login, javax.security.auth.x500, javax.security.sasl, org.ietf.jgss, javax.imageio, javax.imageio.event, javax.imageio.spi, javax.print, javax.print.attribute, javax.print.attribute.standard, javax.print.event, and javax.xml
  • Updated SAX and DOM, and imported GNU JAXP

Fortran

• A new Fortran front end has replaced the aging GNU Fortran 77 front end. The new front end supports Fortran 90 and Fortran 95. It may not yet be as stable as the old Fortran front end.

Ada

• Ada (with tasking and Zero Cost Exceptions) is now available on many more targets, including but not limited to: alpha-linux, hppa-hpux, hppa-linux, powerpc-darwin, powerpc-linux, s390-linux, s390x-linux, sparc-linux.
• Some of the new Ada 2005 features are now implemented like Wide_Wide_Character and Ada.Containers.
• Many bugs have been fixed, tools and documentation improved.
• To compile Ada from the sources, install an older working Ada compiler and then use --enable-languages=ada at configuration time, since the Ada frontend is not currently activated by default. See the Installing GCC for details.

New Targets and Target Specific Improvements

H8/300

• The frame layout has changed. In the new layout, the prologue of a function first saves registers and then allocate space for locals, resulting in an 1% improvement on code size.

IA-32/x86-64 (AMD64)

• The acos, asin, drem, exp10, exp2, expm1, fmod, ilogb, log10, log1p, log2, logb and tan mathematical builtins (and their float and long double variants) are now implemented as inline x87 intrinsics when using -ffast-math.
• The ceil, floor, nearbyint, rint and trunc mathematical builtins (and their float and long double variants) are now implemented as inline x87 intrinsics when using -ffast-math.
• The x87's fsincos instruction is now used automatically with -ffast-math when calculating both the sin and cos of the same argument.
• Instruction selection for multiplication and division by constants has been improved.

IA-64

• Floating point division, integer division and sqrt are now inlined, resulting in significant performance improvements on some codes.

MIPS

• Division by zero checks now use conditional traps if the target processor supports them. This decreases code size by one word per division operation. The old behavior (branch and break) can be obtained either at configure time by passing --with-divide=breaks to configure or at runtime by passing -mdivide-breaks to GCC.
• Support for MIPS64 paired-single instructions has been added. It is enabled by -mpaired-single and can be accessed using both the target-independent vector extensions and new MIPS-specific built-in functions.
• Support for the MIPS-3D ASE has been added. It is enabled by -mips3d and provides new MIPS-3D-specific built-in functions.
• The -mexplicit-relocs option now supports static n64 code (as is used, for example, in 64-bit linux kernels). -mexplicit-relocs should now be feature-complete and is enabled by default when GCC is configured to use a compatible assembler.
• Support for the NEC VR4130 series has been added. This support includes the use of VR-specific instructions and a new VR4130 scheduler. Full VR4130 support can be selected with -march=vr4130 while code for any ISA can be tuned for the VR4130 using -mtune=vr4130. There is also a new -mvr4130-align option that produces better schedules at the cost of increased code size.
• Support for the Broadcom SB-1 has been extended. There is now an SB-1 scheduler as well as support for the SB-1-specific paired-single instructions. Full SB-1 support can be selected with -march=sb1 while code for any ISA can be optimized for the SB-1 using -mtune=sb1.
• The compiler can now work around errata in R4000, R4400, VR4120 and VR4130 processors. These workarounds are enabled by -mfix-r4000, -mfix-r4400, -mfix-vr4120 and -mfix-vr4130 respectively. The VR4120 and VR4130 workarounds need binutils 2.16 or above.
• IRIX shared libraries are now installed into the standard library directories: o32 libraries go into lib/, n32 libraries go into lib32/ and n64 libraries go into lib64/.
• The compiler supports a new -msym32 option. It can be used to optimize n64 code in which all symbols are known to have 32-bit values.

S/390 and zSeries

• New command line options help to generate code intended to run in an environment where stack space is restricted, e.g. Linux kernel code:
  • -mwarn-framesize and -mwarn-dynamicstack trigger compile-time warnings for single functions that require large or dynamic stack frames.
  • -mstack-size and -mstack-guard generate code that checks for stack overflow at run time.
  • -mpacked-stack generates code that reduces the stack frame size of many functions by reusing unneeded parts of the stack bias area.
• The -msoft-float option now ensures that generated code never accesses floating point registers.
• The s390x-ibm-tpf target now fully supports C++, including exceptions and threads.
• Various changes to improve performance of the generated code have been implemented, including:
  • GCC now uses sibling calls where possible.
  • Condition code handling has been optimized, allowing GCC to omit redundant comparisons in certain cases.
  • The cost function guiding many optimizations has been refined to more accurately represent the z900 and z990 processors.
  • The ADD LOGICAL WITH CARRY and SUBTRACT LOGICAL WITH BORROW instructions are now used to avoid conditional branches in certain cases.
  • The back end now uses the LEGITIMIZE_RELOAD_ADDRESS feature to optimize address arithmetic required to access large stack frames.
  • GCC now makes more efficient use of memory-to-memory type instructions (MVC, CLC, ...).
  • More precise tracking of special register use allows better instruction scheduling, in particular of the function prologue and epilogue sequences.
  • The Java front end now generates inline code to implement integer division, instead of calling library routines.

SPARC

• The options -mv8, -msparclite, -mcypress, -msupersparc, -mf930 and -mf934 have been removed. They have been replaced with -mcpu=xxx.
• The internal model used to estimate the relative cost of each instruction has been updated. It is expected to give better results on recent UltraSPARC processors.
• Code generation for function prologues and epilogues has been improved, resulting in better scheduling and allowing multiple exit points in functions.
• Support for Sun's Visual Instruction Set (VIS) has been enhanced. It is enabled by -mvis and provides new built-in functions for VIS instructions on UltraSPARC processors.
• The option -mapp-regs has been turned on by default on Solaris too.

NetWare

• Novell NetWare (on ix86, no other hardware platform was ever really supported by this OS) has been re-enabled and the ABI supported by GCC has been brought into sync with that of MetroWerks CodeWarrior (the ABI previously supported was that of some Unix systems, which NetWare never tried to support).

Obsolete Systems

Support for a number of older systems has been declared obsolete in GCC 4.0. 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:

• Intel i860
• Ubicom IP2022
• National Semiconductor NS32K
• Texas Instruments TMS320C[34]x

Also, those for some individual systems have been obsoleted:

• SPARC family
  • SPARClite-based systems (sparclite-*-coff, sparclite-*-elf, sparc86x-*-elf)
  • OpenBSD 32-bit (sparc-*-openbsd*)

Documentation improvements

Other significant improvements

• Location lists are now generated by default when compiling with debug info and optimization. Location lists provide more accurate debug info about locations of variables and they allow debugging code compiled with -fomit-frame-pointer.
• The -fvisibility option has been added which allows the default ELF visibility of all symbols to be set per compilation and the new #pragma GCC visibility preprocessor command allows the setting of default ELF visibility for a region of code. Using -fvisibility=hidden especially in combination with the new -fvisibility-inlines-hidden can yield substantial improvements in output binary quality including avoiding PLT indirection overheads, reduction of the exported symbol count by up to 60% (with resultant improvements to link and load times), better scope for the optimizer to improve code and up to a 20% reduction in binary size. Using these options correctly yields a binary with a similar symbol count to a Windows DLL.
  Perhaps more importantly, this new feature finally allows (with careful planning) complete avoidance of symbol clashes when manually loading shared objects with RTLD_GLOBAL, thus finally solving problems many projects such as python were forced to use RTLD_LOCAL for (with its resulting issues for C++ correctness). You can find more information about using these options at http://gcc.gnu.org/wiki/Visibility.

GCC 4.0.1

This is the list of problem reports (PRs) from GCC's bug tracking system that are known to be fixed in the 4.0.1 release. This list might not be complete (that is, it is possible that some PRs that have been fixed are not listed here).

GCC 4.0.2

This is the list of problem reports (PRs) from GCC's bug tracking system that are known to be fixed in the 4.0.2 release. This list might not be complete (that is, it is possible that some PRs that have been fixed are not listed here).

Unfortunately, due to a release engineering failure, this release has a regression on Solaris that will affect some C++ programs. We suggest that Solaris users apply a patch that corrects the problem. Users who do not wish to apply the patch should explicitly link C++ programs with the -pthreads option, even if they do not use threads. This problem has been corrected in the current 4.0 branch sources and will not be present in GCC 4.0.3.

GCC 4.0.3

Starting with this release, the function getcontext is recognized by the compiler as having the same semantics as the setjmp function. In particular, the compiler will ensure that all registers are dead before calling such a function and will emit a warning about the variables that may be clobbered after the second return from the function.

GCC 4.0.4

This is the list of problem reports (PRs) from GCC's bug tracking system that are known to be fixed in the 4.0.4 release. This list might not be complete (that is, it is possible that some PRs that have been fixed are not listed here).

The 4.0.4 release is provided for those that require a high degree of binary compatibility with previous 4.0.x releases. For most users, the GCC team recommends that version 4.1.1 or later be used instead."

Please send FSF & GNU inquiries & questions to gnu@gnu.org. There are also other ways to contact the FSF.

These pages are maintained by the GCC team.

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 mailing list at gcc@gnu.org or gcc@gcc.gnu.org. All of our lists have public archives.

Copyright (C) Free Software Foundation, Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110, USA.

Verbatim copying and distribution of this entire article is permitted in any medium, provided this notice is preserved.

Last modified 2007-02-04