GCC 3.4 Release Series
Changes, New Features, and Fixes

The latest release in the 3.4 release series is GCC 3.4.0.

Caveats

GNU Make is now required to build GCC.
With -nostdinc the preprocessor used to ignore both standard include paths and include paths contained in environment variables. It was neither documented nor intended that environment variable paths be ignored, so this has been corrected.
GCC no longer accepts the options -fvolatile, -fvolatile-global and -fvolatile-static. It is unlikely that they worked correctly in any 3.x release.
GCC no longer ships <varargs.h>. Use <stdarg.h> instead.
Support for all the systems obsoleted in GCC 3.3 has been removed from GCC 3.4. See below for a list of systems which are obsoleted in this release.
GCC now requires an ISO C90 (ANSI C89) C compiler to build. K&R C compilers will not work.
The implementation of the MIPS ABIs has changed. As a result, the code generated for certain MIPS targets will not be binary compatible with earlier releases.
In previous releases, the MIPS port had a fake "hilo" register with the user-visible name accum. This register has been removed.
The implementation of the SPARC ABIs has changed. As a result, the code generated will not be binary compatible with earlier releases in certain cases.
The configure option --enable-threads=pthreads has been removed; use --enable-threads=posix instead, which should have the same effect.
Code size estimates used by inlining heuristics for C, Objective-C, C++ and Java have been redesigned significantly. As a result the parameters of -finline-insns, --param max-inline-insns-single and --param max-inline-insns-auto need to be reconsidered.
--param max-inline-slope and --param min-inline-insns have been removed; they are not needed for the new bottom-up inlining heuristics.
The new unit-at-a-time compilation scheme has several compatibility issues:
- The order in which functions, variables, and top-level asm statements are emitted may have changed. Code relying on some particular ordering needs to be updated. The majority of such top-level asm statements can be replaced by section attributes.
- Unreferenced static variables and functions are removed. This may result in undefined references when an asm statement refers to the variable/function directly. In that case either the variable/function shall be listed in asm statement operand or in the case of top-level asm statements the attribute used shall be used to force function/variable to be always output and considered as a possibly used by unknown code.
  For variables the attribute is accepted only by GCC 3.4 and newer, while for earlier versions it is sufficient to use unused to silence warnings about the variables not being referenced. To keep code portable across different GCC versions, you can use appropriate preprocessor conditionals.
- Static functions now can use non-standard passing conventions that may break asm statements calling functions directly. Again the attribute used shall be used to prevent this behavior.
As a temporary workaround, -fno-unit-at-a-time can be used, but this scheme may not be supported by future releases of GCC.
GCC 3.4 automatically places zero-initialized variables in the .bss section on some operating systems. Versions of GNU Emacs up to (and including) 21.3 will not work correctly when using this optimization; you can use -fno-zero-initialized-in-bss to disable it.

General Optimizer Improvements

Usability of the profile feedback and coverage testing has been improved.
- Performance of profiled programs has been improved by faster profile merging code.
- Better use of the profile feedback for optimization (loop unrolling and loop peeling).
- File locking support allowing fork() calls and parallel runs of profiled programs.
- Coverage file format has been redesigned.
- gcov coverage tool has been improved.
- make profiledbootstrap available to build a faster compiler.
  Experiments made on i386 hardware showed an 11% speedup on -O0 and a 7.5% speedup on -O2 compilation of a large C++ testcase.
- New value profiling pass enabled via -fprofile-values
- New value profile transformations pass enabled via -fvpt aims to optimize some code sequences by exploiting knowledge about value ranges or other properties of the operands. At the moment a conversion of expensive divisions into cheaper operations has been implemented.
- New -fprofile-generate and -fprofile-use command line options to simplify the use of profile feedback.
A new unit-at-a-time compilation scheme for C, Objective-C, C++ and Java which is enabled via -funit-at-a-time (and implied by -O2). In this scheme a whole file is parsed first and optimized later. The following basic inter-procedural optimizations are implemented:
- Removal of unreachable functions and variables
- Discovery of local functions (functions with static linkage whose address is never taken)
- On i386, these local functions use register parameter passing conventions.
- Reordering of functions in topological order of the call graph to enable better propagation of optimizing hints (such as the stack alignments needed by functions) in the back end.
- Call graph based out-of-order inlining heuristics which allows to limit overall compilation unit growth (--param inline-unit-growth).
Overall, the unit-at-a-time scheme produces a 1.3% improvement for the SPECint2000 benchmark on the i386 architecture (AMD Athlon CPU).
More realistic code size estimates used by inlining for C, Objective-C, C++ and Java. The growth of large functions can now be limited via --param large-function-insns and --param large-function-growth.
A new cfg-level loop optimizer pass replaces the old loop unrolling pass and adds two other loop transformations -- loop peeling and loop unswitching -- and also uses the profile feedback to limit code growth. (The three optimizations are enabled by -funroll-loops, -fpeel-loops and -funswitch-loops flags, respectively).
The old loop unroller still can be enabled by -fold-unroll-loops and may produce better code in some cases, especially when the webizer optimization pass is not run.
A new web construction pass enabled via -fweb (and implied by -O3) improves the quality of register allocation, CSE, first scheduling pass and some other optimization passes by avoiding re-use of pseudo registers with non-overlapping live ranges. The pass almost always improves code quality but does make debugging difficult and thus is not enabled by default by -O2
The pass is especially effective as cleanup after code duplication passes, such as the loop unroller or the tracer.
Experimental implementations of superblock or trace scheduling in the second scheduling pass can be enabled via -fsched2-use-superblocks and -fsched2-use-traces, respectively.

New Languages and Language specific improvements

Ada

The Ada front end has been updated to include numerous bug fixes and enhancements. These include:
- Improved project file support
- Additional set of warnings about potential wrong code
- Improved error messages
- Improved code generation
- Improved cross reference information
- Improved inlining
- Better run-time check elimination
- Better error recovery
- More efficient implementation of unbounded strings
- Added features in GNAT.Sockets, GNAT.OS_Lib, GNAT.Debug_Pools, ...
- New GNAT.xxxx packages (e.g. GNAT.Strings, GNAT.Exception_Action)
- New pragmas
- New -gnatS switch replacing gnatpsta
- Implementation of new Ada features (in particular limited with, limited aggregates)

C/Objective-C/C++

Precompiled headers are now supported. Precompiled headers can dramatically speed up compilation of some projects. There are some known defects in the current precompiled header implementation that will result in compiler crashes in relatively rare situations. Therefore, precompiled headers should be considered a "technology preview" in this release. Read the manual for details about how to use precompiled headers.
File handling in the preprocessor has been rewritten. GCC no longer gets confused by symlinks and hardlinks, and now has a correct implementation of #import and #pragma once. These two directives have therefore been un-deprecated.
The undocumented extension that allowed C programs to have a label at the end of a compound statement, which has been deprecated since GCC 3.0, has been removed.
The cast-as-lvalue extension has been removed for C++ and deprecated for C and Objective-C. In particular, code like this:
```
        int i;
        (char) i = 5;
        
```
or this:
```
        char *p;
        ((int *) p)++;
        
```
is no longer accepted for C++ and will not be accepted for C and Objective-C in a future version.
The conditional-expression-as-lvalue extension has been deprecated for C and Objective-C. In particular, code like this:
```
        int a, b, c;
        (a ? b : c) = 2;
        
```
will not be accepted for C and Objective-C in a future version.
The compound-expression-as-lvalue extension has been deprecated for C and Objective-C. In particular, code like this:
```
        int a, b;
        (a, b) = 2;
        
```
will not be accepted for C and Objective-C in a future version. A possible non-intrusive workaround is the following:
```
        (*(a, &b)) = 2;
        
```
Several built-in functions such as __builtin_popcount for counting bits, finding the highest and lowest bit in a word, and parity have been added.
The -fwritable-strings option has been deprecated and will be removed.
Many C math library functions are now recognized as built-ins and optimized.
The C, C++, and Objective-C compilers can now handle source files written in any character encoding supported by the host C library. The default input character set is taken from the current locale, and may be overridden with the -finput-charset command line option. In the future we will add support for inline encoding markers.

C++

G++ is now much closer to full conformance to the ISO/ANSI C++ standard. This means, among other things, that a lot of invalid constructs which used to be accepted in previous versions will now be rejected. It is very likely that existing C++ code will need to be fixed. This document lists some of the most common issues.
A hand-written recursive-descent C++ parser has replaced the YACC-derived C++ parser from previous GCC releases. The new parser contains much improved infrastructure needed for better parsing of C++ source codes, handling of extensions, and clean separation (where possible) between proper semantics analysis and parsing. The new parser fixes many bugs that were found in the old parser.

You must now use the typename and template keywords to disambiguate dependent names, as required by the C++ standard.

	struct K {
	  typedef int mytype_t;
	};
	
	template <class T1> struct A {
	  template <class T2> struct B { 
	      void callme(void);
	    };

	  template <int N> void bar(void) 
	  {
	    // Use 'typename' to tell the parser that T1::mytype_t names
	    //  a type. This is needed because the name is dependent (in
	    //  this case, on template parameter T1).
	    typename T1::mytype_t x;
	    x = 0;
	  }
	};

	template <class T> void template_func(void)
	{
	  // Use 'template' to prefix member templates within 
	  //  dependent types (a has type A<T>, which depends on
	  //  the template parameter T).
	  A<T> a;
	  a.template bar<0>();

	  // Use 'template' to tell the parser that B is a nested 
	  //  template class (dependent on template parameter T), and 
	  //  'typename' because the whole A<T>::B<int> is 
	  //  the name of a type (again, dependent).
	  typename A<T>::template B<int> b;
	  b.callme();
	}
	
	void non_template_func(void)
	{
	  // Outside of any template class or function, no names can be
	  //  dependent, so the use of the keyword 'typename' and 'template'
	  //  is not needed (and actually forbidden).
	  A<K> a;
	  a.bar<0>();
	  A<K>::B<float> b;
	  b.callme();
	}

In a template definition, unqualified names will no longer find members of a dependent base. For example,

        template <typename T> struct B {
	  int m;
	  int n;
	  int f ();
	  int g ();
	};
	int n;
	int g ();
	template <typename T> struct C : B<T> {
	  void g ()
	  {
	    m = 0; // error
	    f ();  // error
	    n = 0; // ::n is modified
	    g ();  // ::g is called
	  }
	};

You must make the names dependent by prefixing them with this->. Here is the corrected definition of C<T>::g,

	template <typename T> void C<T>::g ()
	{
	  this->m = 0;
	  this->f ();
	  this->n = 0
	  this->g ();
	}

In templates, all non-dependent names are now looked up and bound at definition time (while parsing the code), instead of later when the template is instantiated. For instance:

	void foo(int);

	template <int> struct A {
	  static void bar(void){
	    foo('a');
	  }
	};

	void foo(char);

	int main()
	{
	  A<0>::bar();    // Calls foo(int), used to call foo(char).
	}

In an explicit instantiation of a class template, you must use class or struct before the template-id:

	template <int N>
	class A {};
	
	template A<0>;         // error, not accepted anymore
	template class A<0>;   // OK

The "named return value" and "implicit typename" extensions have been removed.
Default arguments in function types have been deprecated and will be removed.
ARM-style name-injection of friend declarations has been deprecated and will be removed. For example: struct S { friend void f(); }; void g() { f(); } will not be accepted by future versions of G++; instead a declaration of "f" will need to be present outside of the scope of "S".
Covariant returns are implemented for all but varadic functions that require an adjustment.

When -pedantic is used, G++ now issues errors about spurious semicolons. For example,

        namespace N {}; // Invalid semicolon.
        void f() {}; // Invalid semicolon.

G++ no longer accepts attributes for a declarator after the initializer associated with that declarator. For example,
```
        X x(1) __attribute__((...));
```
is no longer accepted. Instead, use:
```
        X x __attribute__((...)) (1);
```
Inside the scope of a template class, the name of the class itself can be treated as either a class or a template. So GCC used to accept the class name as argument of type template, and template template parameter. However this is not C++ standard compliant. Now the name is not treated as a valid template template argument unless you qualify the name by its scope. For example, the code below no longer compiles.
```
        template <template <class> class TT> class X {};
        template <class T> class Y {
          X<Y> x; // Invalid, Y is always a type template parameter.
        };
```
The valid code for the above example is
```
          X< ::Y> x; // Valid.
```
(Notice the space between < and : to prevent GCC to interpret this as a digraph for [.)
Friend declarations that refer to template specializations are rejected if the template has not already been declared. For example,
```
	template <typename T>
	class C {
	  friend void f<> (C&);
	};
```
is rejected. You must first declare f as a template,
```
	template <typename T>
	void f(T);
```
In case of friend declarations, every name used in the friend declaration must be accessible at the point of that declaration. Previous versions of G++ used to be less strict about this and allowed friend declarations for private class members, for example. See the ISO C++ Standard Committee's defect report #209 for details.

Declaration of member functions of class templates as friends are supported. For example,

	template <typename T> struct A {
	  void f();
	};
	class C {
	  template <typename T> friend void A<T>::f();
	};

You must use template <> to introduce template specializations, as required by the standard. For example,
```
	template <typename T>
	struct S;

	struct S<int> { };
```
is rejected. You must write,
```
	template <> struct S<int> {};
```
G++ used to accept code like this,
```
	struct S {
	  int h();
	  void f(int i = g());
	  int g(int i = h());
	};
```
This behavior is not mandated by the standard. Now G++ issues an error about this code. To avoid the error, you must move the declaration of g before the declaration of f. The default arguments for g must be visible at the point where it is called.
The C++ ABI Section 3.3.3 specifications for the array construction routines __cxa_vec_new2 and __cxa_vec_new3 were changed to return NULL when the allocator argument returns NULL. These changes are incorporated into the libstdc++ runtime library.

Using a name introduced by a typedef in a friend declaration or in an explicit instantiation is now rejected, as specified by the ISO C++ standard.

	class A;
	typedef A B;
	class C { 
	  friend class B;      // error, no typedef name here
	  friend B;            // error, friend always needs class/struct/enum
	  friend class A;      // OK
	};
	
	template <int> class Q {};
	typedef Q<0> R;
	template class R;      // error, no typedef name here
	template class Q<0>;   // OK

When allocating an array with a new expression, GCC used to allow parentheses around the type name. This is actually ill-formed and it is now rejected:
```
	int* a = new (int)[10];    // error, not accepted anymore
	int* a = new int[10];      // OK
```

When binding an rvalue of class type to a reference, the copy constructor of the class must be accessible. For instance, consider the following code:

	class A 
	{
	public:
	  A();
	  
	private:
	  A(const A&);   // private copy ctor
	};
	
	A makeA(void);
	void foo(const A&);
	
	void bar(void)
	{
	  foo(A());       // error, copy ctor is not accessible
	  foo(makeA());   // error, copy ctor is not accessible
	  
	  A a1;
	  foo(a1);        // OK, a1 is a lvalue
	}

This might be surprising at first sight, especially since most popular compilers do not correctly implement this rule (further details).

Runtime Library (libstdc++)

Optimization work:
- Streamlined streambuf, filebuf, separate synched with C Standard I/O streambuf.
- All formatted I/O now uses cached locale information.
- STL optimizations (memory/speed for list, red-black trees as used by sets and maps).
- More use of GCC builtins.
- String optimizations (avoid contention on increment/decrement-and-test of the reference count in the empty-string object, constructor from input_iterators speedup).
Static linkage size reductions.
Large File Support (files larger than 2 GB on 32-bit systems).
Wide character and variable encoding filebuf work (UTF-8, Unicode).
Generic character traits.
Also support wchar_t specializations on Mac OS 10.3.x, FreeBSD 5.x, Solaris 2.7 and above, AIX 5.x, Irix 6.5.
The allocator class is now standard-conformant, and two additional extension allocators have been added, mt_alloc and bitmap_allocator.
PCH support: -include bits/stdc++.h (2x compile speedup).
Rewrote __cxa_demangle with support for C++ style allocators.
New debug modes for STL containers and iterators.
Testsuite rewrite: five times as many tests, plus increasingly sophisticated tests, including I/O, MT, multi-locale, wide and narrow characters.
Use current versions of GNU "autotools" for build/configuration.

Objective-C

The Objective-C front end has been updated to include the numerous bug fixes and enhancements previously available only in Apple's version of GCC. These include:
- Structured exception (@try... @catch... @finally, @throw) and synchronization (@synchronized) support. These are accessible via the -fobjc-exceptions switch; as of this writing, they may only be used in conjunction with -fnext-runtime on Mac OS X 10.3 and later. See Options Controlling Objective-C Dialect for more information.
- An overhaul of @encode logic. The C99 _Bool and C++ bool type may now be encoded as 'B'. In addition, the back-end/codegen dependencies have been removed.
- An overhaul of message dispatch construction, ensuring that the various receiver types (and casts thereof) are handled properly, and that correct diagnostics are issued.
- Support for "Zero-Link" (-fzero-link) and "Fix-and-Continue" (-freplace-objc-classes) debugging modes, currently available on Mac OS X 10.3 and later. See Options Controlling Objective-C Dialect for more information.
- Access to optimized runtime entry points (-fno-nil-receivers) on the assumption that message receivers are never nil. This is currently available on Mac OS X 10.3 and later. See Options Controlling Objective-C Dialect for more information.

Java

Compiling a .jar file will now cause non-.class entries to be automatically compiled as resources.
libgcj has been ported to Darwin.
Jeff Sturm has adapted Jan Hubicka's call graph optimization code to gcj.
libgcj has a new gcjlib URL type; this lets URLClassLoader load code from shared libraries.
libgcj has been much more completely merged with GNU Classpath.
Class loading is now much more correct; in particular the caller's class loader is now used when that is required.
Eclipse 2.x will run out of the box using gij.
Parts of java.nio have been implemented. Direct and indirect buffers work, as do fundamental file and socket operations.
java.awt has been improved, though it is still not ready for general use.
The HTTP protocol handler now uses HTTP/1.1 and can handle the POST method.
The MinGW port has matured. Enhancements include socket timeout support, thread interruption, improved Runtime.exec() handling and support for accented characters in filenames.

Fortran

Fortran improvements are listed in the Fortran documentation.

New Targets and Target Specific Improvements

Alpha

Several built-in functions have been added such as __builtin_alpha_zap to allow utilizing the more obscure instructions of the CPU.
Parameter passing of complex arguments has changed to match the ABI. This change is incompatible with previous GCC versions, but does fix compatibility with the Tru64 compiler and several corner cases where GCC was incompatible with itself.

ARM

Nicolas Pitre has contributed his hand-coded floating-point support code for ARM. It is both significantly smaller and faster than the existing C-based implementation, even when building applications for Thumb. The arm-elf configuration has been converted to use the new code.
Support for the Intel's iWMMXt architecture, a second generation XScale processor, has been added. Enabled at run time with the -mcpu=iwmmxt command line switch.
A new ARM target has been added: arm-wince-pe. This is similar to the arm-pe target, but it defaults to using the APCS32 ABI.
The existing ARM pipeline description has been converted to the use the DFA processor pipeline model. There is not much change in code performance, but the description is now easier to understand.
Support for the Cirrus EP9312 Maverick floating point co-processor added. Enabled at run time with the -mcpu=ep9312 command line switch. Note however that the multilibs to support this chip are currently disabled in gcc/config/arm/t-arm-elf, so if you want to enable their production you will have to uncomment the entries in that file.

H8/300

Support for long long has been added.
Support for saveall attribute has been added.
Pavel Pisa contributed hand-written 32-bit-by-32-bit division code for H8/300H and H8S, which is much faster than the previous implementation.
A lot of small performance improvements.

IA-32/AMD64 (x86-64)

Tuning for K8 (AMD Opteron/Athlon64) core is available via -march=k8 and -mcpu=k8.
Scalar SSE code generation carefully avoids reformatting penalties, hidden dependencies and minimizes the number of uops generated on both Intel and AMD CPUs.
Vector MMX and SSE operands are now passed in registers to improve performance and match the argument passing convention used by the Intel C++ Compiler. As a result it is not possible to call functions accepting vector arguments compiled by older GCC version.
Conditional jump elimination is now more aggressive on modern CPUs.
The Athlon ports has been converted to use the DFA processor pipeline description.
Optimization of indirect tail calls is now possible in a similar fashion as direct sibcall optimization.
Further small performance improvements.
-m128bit-long-double is now less buggy.
__float128 support in 64-bit compilation.
Support for data structures exceeding 2GB in 64-bit mode.
-mcpu has been renamed to -mtune.

IA-64

Tuning code for the Itanium 2 processor has been added. The generation of code tuned for Itanium 2 (option -mtune=itanium2) is enabled by default now. To generate code tuned for Itanium 1 the option -mtune=itanium1 should be used.
DFA processor pipeline descriptions for the IA-64 processors have been added. This resulted in about 3% improvement on the SPECInt2000 benchmark for Itanium 2.
Instruction bundling for the IA-64 processors has been rewritten using the DFA pipeline hazard recognizer. It resulted in about 60% compiler speedup on the SPECInt2000 C programs.

M32R

Support for the M32R/2 processor has been added by Renesas.
Support for an M32R Linux target and PIC code generation has been added by Renesas.

M68000

Bernardo Innocenti (Develer S.r.l.) has contributed the m68k-uclinux target, based on former work done by Paul Dale (SnapGear Inc.). Code generation for the ColdFire processors family has been enhanced and extended to support the MCF 53xx and MCF 54xx cores, integrating former work done by Peter Barada (Motorola).

MIPS

Processor-specific changes

Support for the RM7000 and RM9000 processors has been added. It can be selected using the -march compiler option and should work with any MIPS I (mips-*) or MIPS III (mips64-*) configuration.
Support for revision 2 of the MIPS32 ISA has been added. It can be selected with the command-line option -march=mips32r2.
There is a new option, -mfix-sb1, to work around certain SB-1 errata.

Configuration

It is possible to customize GCC using the following configure-time options:
- --with-arch, which specifies the default value of the -march option.
- --with-tune, which specifies the default value of the -mtune option.
- --with-abi, which specifies the default ABI.
- --with-float=soft, which tells GCC to use software floating point by default.
- --with-float=hard, which tells GCC to use hardware floating point by default.
A 64-bit GNU/Linux port has been added. The associated configurations are mips64-linux-gnu and mips64el-linux-gnu.
The 32-bit GNU/Linux port now supports Java.
The IRIX 6 configuration now supports the o32 ABI and will build o32 multilibs by default. This support is compatible with both binutils and the SGI tools, but note that several features, including debugging information and DWARF2 exception handling, are only available when using the GNU assembler. Use of the GNU assembler and linker (version 2.15 or above) is strongly recommended.
The IRIX 6 configuration now supports 128-bit long doubles.
There are two new RTEMS-specific configurations, mips-rtems and mipsel-rtems.
There are two new *-elf configurations, mipsisa32r2-elf and mipsisa32r2el-elf.

General

Several ABI bugs have been fixed. Unfortunately, these changes will break binary compatibility with earlier releases.
GCC can now use explicit relocation operators when generating -mabicalls code. This behavior is controlled by -mexplicit-relocs and can have several performance benefits. For example:
- It allows for more optimization of GOT accesses, including better scheduling and redundancy elimination.
- It allows sibling calls to be implemented as jumps.
- n32 and n64 leaf functions can use a call-clobbered global pointer instead of $28.
- The code to set up $gp can be removed from functions that don't need it.
A new option, -mxgot, allows the GOT to be bigger than 64k. This option is equivalent to the assembler's -xgot option and should be used instead of -Wa,-xgot.
Frame pointer elimination is now supported when generating 64-bit MIPS16 code.
Inline block moves have been optimized to take more account of alignment information.
Many internal changes have been made to the MIPS port, mostly aimed at reducing the reliance on assembler macros.

PowerPC Darwin

Support for shared/dylib gcc libraries has been added. It is enabled by default on powerpc-apple-darwin7.0.0 and up.
Libgcj is enabled by default. On systems older than powerpc-apple-darwin7.0.0 you need to install dlcompat.

S/390 and zSeries

New command-line options allow to specify the intended execution environment for generated code:
- -mesa/-mzarch allows to specify whether to generate code running in ESA/390 mode or in z/Architecture mode (this is applicable to 31-bit code only).
- -march allows to specify a minimum processor architecture level (g5, g6, z900, or z990).
- -mtune allows to specify which processor to tune for.
It is possible to customize GCC using the following configure-time options:
- --with-mode, which specifies whether to default to assuming ESA/390 or z/Architecture mode.
- --with-arch, which specifies the default value of the -march option.
- --with-tune, which specifies the default value of the -mtune option.
Support for the z990 processor has been added, and can be selected using -march=z990 or -mtune=z990. This includes instruction scheduling tuned for the superscalar instruction pipeline of the z990 processor as well as support for all new instructions provided by the long-displacement facility.
Support to generate 31-bit code optimized for zSeries processors (running in ESA/390 or in z/Architecture mode) has been added. This can be selected using -march=z900 and -mzarch respectively.
Instruction scheduling for the z900 and z990 processors now uses the DFA pipeline hazard recognizer.
GCC no longer generates code to maintain a stack backchain, previously used to generate stack backtraces for debugging purposes. As replacement that does not incur runtime overhead, DWARF-2 call frame information is provided by GCC; this is supported by GDB 6.1. The old behavior can be restored using the -mbackchain option.
The stack frame size of functions may now exceed 2 GB in 64-bit code.
A port for the 64-bit IBM TPF operating system has been added; the configuration is s390x-ibm-tpf. This configuration is supported as cross-compilation target only.
Various changes to improve the generated code have been implemented, including:
- GCC now uses the MULTIPLY AND ADD and MULTIPLY AND SUBTRACT instructions to significantly speed up many floating-point applications.
- GCC now uses the ADD LOGICAL WITH CARRY and SUBTRACT LOGICAL WITH BORROW instructions to speed up long long arithmetic.
- GCC now uses the SEARCH STRING instruction to implement strlen().
- In many cases, function call overhead for 31-bit code has been reduced by placing the literal pool after the function code instead of after the function prolog.
- Register 14 is no longer reserved in 64-bit code.
- Handling of global register variables has been improved.

SPARC

The option -mflat is deprecated.
Support for large (> 2GB) frames has been added to the 64-bit port.
Several ABI bugs have been fixed. Unfortunately, these changes will break binary compatibility with earlier releases.
The default debugging format has been switched from STABS to DWARF-2 for 32-bit code on Solaris 7 and later. DWARF-2 is already the default debugging format for 64-bit code on Solaris.

SuperH

Support for the SH2E processor has been added. Enabled at run time with the -m2e command line switch, or at configure time by specifying sh2e as the machine part of the target triple.

V850

Support for the Mitsubishi V850E1 processor has been added. This is a variant of the V850E processor with some additional debugging instructions.

Xtensa

Several ABI bugs have been fixed. Unfortunately, these changes break binary compatibility with earlier releases.
- For big-endian processors, the padding of aggregate return values larger than a word has changed. If the size of an aggregate return value is not a multiple of 32 bits, previous versions of GCC inserted padding in the most-significant bytes of the first return value register. Aggregates larger than a word are now padded in the least-significant bytes of the last return value register used. Aggregates smaller than a word are still padded in the most-significant bytes. The return value padding has not changed for little-endian processors.
- Function arguments with 16-byte alignment are now properly aligned.
- The implementation of the va_list type has changed. A va_list value created by va_start from a previous release cannot be used with va_arg from this release, or vice versa.
More processor configuration options for Xtensa processors are supported:
- the ABS instruction is now optional;
- the ADDX* and SUBX* instructions are now optional;
- an experimental CONST16 instruction can be used to synthesize constants instead of loading them from constant pools.
These and other Xtensa processor configuration options can no longer be enabled or disabled by command-line options; the processor configuration must be specified by the xtensa-config.h header file when building GCC. Additionally, the -mno-serialize-volatile option is no longer supported.

Obsolete Systems

Support for a number of older systems has been declared obsolete in GCC 3.4. Unless there is activity to revive them, the next release of GCC will have their sources permanently removed.

All configurations of the following processor architectures have been declared obsolete:

Mitsubishi D30V, d30v-*
AT&T DSP1600 and DSP1610, dsp16xx-*
Intel 80960, i960

Also, some individual systems have been obsoleted:

ARM Family
- Support for generating code for operation in APCS/26 mode (-mapcs-26).
IBM ESA/390
- "Bigfoot" port, i370-*. (The other port, s390-*, is actively maintained and supported.)
Intel 386 family
- MOSS, i?86-moss-msdos and i?86-*-moss*
- NCR 3000 running System V r.4, i?86-ncr-sysv4*
- FreeBSD with a.out object format, i?86-*-freebsd*aout* and i?86-*-freebsd2*
- Linux with a.out object format, i?86-linux*aout*
- Linux with libc5, a.k.a. glibc1, i?86-linux*libc1*
- Interix versions before Interix 3, i?86-*-interix
- Mach microkernel, i?86-mach*
- SCO UnixWare with UDK, i?86-*-udk*
- Generic System V releases 1, 2, and 3, i?86-*-sysv[123]*
- VSTa microkernel, i386-*-vsta
Motorola M68000 family
- HPUX, m68k-hp-hpux* and m68000-hp-hpux*
- NetBSD with a.out object format (before NetBSD 1.4), m68k-*-*-netbsd* except m68k-*-*-netbsdelf*
- Generic System V r.4, m68k-*-sysv4*
VAX
- Generic VAX, vax-*-* (This is generic VAX only; we have not obsoleted any VAX triples for specific operating systems.)

Documentation improvements

Other significant improvements

The build system has undergone several significant cleanups. Subdirectories will only be configured if they are being built, and all subdirectory configures are run from the make command. The top level has been autoconfiscated.
Building GCC no longer writes to its source directory. This should help those wishing to share a read-only source directory over NFS or build from a CD. The exceptions to this feature are if you configure with either --enable-maintainer-mode or --enable-generated-files-in-srcdir.
The -W warning option has been renamed to -Wextra, which is more easily understood. The older spelling will be retained for backwards compatibility.
Substantial improvements in compile time have been made, particularly for non-optimizing compilations.

GCC 3.4.0

Detailed release notes for the GCC 3.4.0 release follow.

Bug Fixes

A vast number of bugs have been fixed in 3.4.0, too many to publish a complete list here. Follow this link to query the Bugzilla database for the list of over 900 bugs fixed in 3.4.0. This is the list of all bugs marked as resolved and fixed in 3.4.0 that are not flagged as 3.4 regressions.

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 public developer mailing list at gcc@gnu.org or gcc@gcc.gnu.org.

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

Last modified 2004-04-20

GCC 3.4 Release SeriesChanges, New Features, and Fixes