This document contains the release notes for the LLVM compiler infrastructure, release 2.1. Here we describe the status of LLVM, including major improvements from the previous release and any known problems. All LLVM releases may be downloaded from the LLVM releases web site.

For more information about LLVM, including information about the latest release, please check out the main LLVM web site. If you have questions or comments, the LLVM developer's mailing list is a good place to send them.

Note that if you are reading this file from a Subversion checkout or the main LLVM web page, this document applies to the next release, not the current one. To see the release notes for a specific releases, please see the releases page.

This is the twelfth public release of the LLVM Compiler Infrastructure. It includes many features and refinements from LLVM 2.0.

LLVM 2.1 brings two new beta C front-ends. First, a new version of llvm-gcc based on GCC 4.2, innovatively called "llvm-gcc-4.2". This promises to bring FORTRAN and Ada support to LLVM as well as features like atomic builtins and OpenMP. None of these actually work yet, but don't let that stop you checking it out!

Second, LLVM now includes its own native C and Objective-C front-end (C++ is in progress, but is not very far along) code named "clang". This front-end has a number of great features, primarily aimed at source-level analysis and speeding up compile-time. At this point though, the LLVM Code Generator component is still very early in development, so it's mostly useful for people looking to build source-level analysis tools or source-to-source translators.

Some of the most noticable feature improvements this release have been in the optimizer, speeding it up and making it more aggressive. For example:

• Owen Anderson wrote the new MemoryDependenceAnalysis pass, which provides a lazy, caching layer on top of AliasAnalysis. He then used it to rewrite DeadStoreElimination which resulted in significantly better compile time in common cases,
• Owen implemented the new GVN pass, which is also based on MemoryDependenceAnalysis. This pass replaces GCSE/LoadVN in the standard set of passes, providing more aggressive optimization at a some-what improved compile-time cost.
• Owen implemented GVN-PRE, a partial redundancy elimination algorithm that shares some details with the new GVN pass. It is still in need of compile time tuning, and is not turned on by default.
• Devang merged ETForest and DomTree into a single easier to use data structure. This makes it more obvious which datastructure to choose (because there is only one) and makes the compiler more memory and time efficient (less stuff to keep up-to-date).
• Nick Lewycky improved loop trip count analysis to handle many more common cases.

One of the main focuses of this release was performance tuning and bug fixing. In addition to these, several new major changes occurred:

• Dale finished up the Tail Merging optimization in the code generator, and enabled it by default. This produces smaller code that is also faster in some cases.
• Christopher Lamb implemented support for virtual register sub-registers, which can be used to better model many forms of subregisters. As an example use, he modified the X86 backend to use this to model truncates and extends more accurately (leading to better code).
• Dan Gohman changed the way we represent vectors before legalization, significantly simplifying the SelectionDAG representation for these and making the code generator faster for vector code.
• Evan contributed a new target independent if-converter. While it is target independent, so far only the ARM backend uses it.
• Evan rewrote the way the register allocator handles rematerialization, allowing it to be much more effective on two-address targets like X86, and taught it to fold loads away when possible (also a big win on X86).
• Dan Gohman contributed support for better alignment and volatility handling in the code generator, and significantly enhanced alignment analysis for SSE load/store instructions. With his changes, an insufficiently-aligned SSE load instruction turns into movups, for example.
• Duraid Madina contributed a new "bigblock" register allocator, and Roman Levenstein contributed several big improvements. BigBlock is optimized for code that uses very large basic blocks. It is slightly slower than the "local" allocator, but produces much better code.
• David Greene refactored the register allocator to split coalescing out from allocation, making coalescers pluggable.

New features include:

• Bruno Cardoso Lopes contributed initial MIPS support. It is sufficient to run many small programs, but is still incomplete and is not yet fully performant.
• Bill Wendling added SSSE3 support to the X86 backend.
• Nicholas Geoffray contributed improved linux/ppc ABI and JIT support.
• Dale Johannesen rewrote handling of 32-bit float values in the X86 backend when using the floating point stack, fixing several nasty bugs.
• Dan contributed rematerialization support for the X86 backend, in addition to several X86-specific micro optimizations.

New features include:

• Duncan and Anton made significant progress chasing down a number of problems with C++ Zero-Cost exception handling in llvm-gcc 4.0 and 4.2. It is now at the point where it "just works" on linux/X86-32 and has partial support on other targets.
• Devang and Duncan fixed a huge number of bugs relating to bitfields, pragma pack, and variable sized fields in structures.
• Tanya implemented support for __attribute__((noinline)) in llvm-gcc, and added support for generic variable annotations which are propagated into the LLVM IR, e.g. "int X __attribute__((annotate("myproperty")));".
• Sheng Zhou and Christopher Lamb implemented alias analysis support for "restrict" pointer arguments to functions.
• Duncan contributed support for trampolines (taking the address of a nested function). Currently this is only supported on the X86-32 target.
• Lauro Ramos Venancio contributed support to encode alignment info in load and store instructions, the foundation for other alignment-related work.

New features include:

• Neil Booth contributed a new "APFloat" class, which ensures that floating point representation and constant folding is not dependent on the host architecture that builds the application. This support is the foundation for "long double" support that will be wrapped up in LLVM 2.2.
• Based on the APFloat class, Dale redesigned the internals of the ConstantFP class and has been working on extending the core and optimizer components to support various target-specific 'long double's. We expect this work to be completed in LLVM 2.2.
• LLVM now provides an LLVMBuilder class, which makes it significantly easier to create LLVM IR instructions.
• Reid contributed support for intrinsics that take arbitrary integer typed arguments. Dan Gohman and Chandler extended it to support arbitrary floating point arguments and vectors.

New features include:

• Sterling Stein contributed a new BrainF frontend, located in llvm/examples. This shows a some of the more modern APIs for building a front-end, and demonstrates JIT compiler support.
• David Green contributed a new --enable-expensive-checks configure option which enables STL checking, and fixed several bugs exposed by it.

LLVM is known to work on the following platforms:

• Intel and AMD machines running Red Hat Linux, Fedora Core and FreeBSD (and probably other unix-like systems).
• PowerPC and X86-based Mac OS X systems, running 10.2 and above in 32-bit and 64-bit modes.
• Intel and AMD machines running on Win32 using MinGW libraries (native)
• Intel and AMD machines running on Win32 with the Cygwin libraries (limited support is available for native builds with Visual C++).
• Sun UltraSPARC workstations running Solaris 8.
• Alpha-based machines running Debian GNU/Linux.
• Itanium-based machines running Linux and HP-UX.

The core LLVM infrastructure uses GNU autoconf to adapt itself to the machine and operating system on which it is built. However, minor porting may be required to get LLVM to work on new platforms. We welcome your portability patches and reports of successful builds or error messages.

This section contains all known problems with the LLVM system, listed by component. As new problems are discovered, they will be added to these sections. If you run into a problem, please check the LLVM bug database and submit a bug if there isn't already one.

The following components of this LLVM release are either untested, known to be broken or unreliable, or are in early development. These components should not be relied on, and bugs should not be filed against them, but they may be useful to some people. In particular, if you would like to work on one of these components, please contact us on the LLVMdev list.

• The -cee pass is known to be buggy, and may be removed in a future release.
• The MSIL backend is experimental.
• The IA64 code generator is experimental.
• The Alpha backend is experimental.
• "-filetype=asm" (the default) is the only supported value for the -filetype llc option.

• The X86 backend does not yet support inline assembly that uses the X86 floating point stack.
• The X86 backend occasionally has alignment problems on operating systems that don't require 16-byte stack alignment (including most non-darwin OS's like linux).

• PowerPC backend does not correctly implement ordered FP comparisons.
• The Linux PPC32/ABI support needs testing for the interpreter and static compilation, and lacks support for debug information.

• Thumb mode works only on ARMv6 or higher processors. On sub-ARMv6 processors, thumb programs can crash or produce wrong results (PR1388).
• Compilation for ARM Linux OABI (old ABI) is supported, but not fully tested.
• There is a bug in QEMU-ARM (<= 0.9.0) which causes it to incorrectly execute programs compiled with LLVM. Please use more recent versions of QEMU.

• The SPARC backend only supports the 32-bit SPARC ABI (-m32), it does not support the 64-bit SPARC ABI (-m64).

• On 21164s, some rare FP arithmetic sequences which may trap do not have the appropriate nops inserted to ensure restartability.

• C++ programs are likely to fail on IA64, as calls to setjmp are made where the argument is not 16-byte aligned, as required on IA64. (Strictly speaking this is not a bug in the IA64 back-end; it will also be encountered when building C++ programs using the C back-end.)
• The C++ front-end does not use IA64 ABI compliant layout of v-tables. In particular, it just stores function pointers instead of function descriptors in the vtable. This bug prevents mixing C++ code compiled with LLVM with C++ objects compiled by other C++ compilers.
• There are a few ABI violations which will lead to problems when mixing LLVM output with code built with other compilers, particularly for floating-point programs.
• Defining vararg functions is not supported (but calling them is ok).
• The Itanium backend has bitrotted somewhat.

• The C backend does not support inline assembly code.
• The C backend does not support vectors yet.
• The C backend violates the ABI of common C++ programs, preventing intermixing between C++ compiled by the CBE and C++ code compiled with LLC or native compilers.

llvm-gcc4 does not currently support Link-Time Optimization on most platforms "out-of-the-box". Please inquire on the llvmdev mailing list if you are interested.

• "long double" is silently transformed by the front-end into "double". There is no support for floating point data types of any size other than 32 and 64 bits.

• llvm-gcc does not support __builtin_apply yet. See Constructing Calls: Dispatching a call to another function.

• llvm-gcc partially supports these GCC extensions:

  1. Nested Functions: As in Algol and Pascal, lexical scoping of functions. Nested functions are supported, but llvm-gcc does not support taking the address of a nested function (except on the X86-32 target) or non-local gotos.
  2. Function Attributes: Declaring that functions have no side effects or that they can never return.
     Supported: alias, always_inline, cdecl, const, constructor, destructor, deprecated, fastcall, format, format_arg, non_null, noinline, noreturn, pure, regparm section, stdcall, unused, used, visibility, warn_unused_result, weak
     Ignored: nothrow, malloc, no_instrument_function

• llvm-gcc supports the vast majority of GCC extensions, including:

  1. Pragmas: Pragmas accepted by GCC.
  2. Local Labels: Labels local to a block.
  3. Other Builtins: Other built-in functions.
  4. Variable Attributes: Specifying attributes of variables.
  5. Type Attributes: Specifying attributes of types.
  6. Thread-Local: Per-thread variables.
  7. Variable Length: Arrays whose length is computed at run time.
  8. Labels as Values: Getting pointers to labels and computed gotos.
  9. Statement Exprs: Putting statements and declarations inside expressions.
  10. Typeof: typeof: referring to the type of an expression.
  11. Lvalues: Using ?:, "," and casts in lvalues.
  12. Conditionals: Omitting the middle operand of a ?: expression.
  13. Long Long: Double-word integers.
  14. Complex: Data types for complex numbers.
  15. Hex Floats:Hexadecimal floating-point constants.
  16. Zero Length: Zero-length arrays.
  17. Empty Structures: Structures with no members.
  18. Variadic Macros: Macros with a variable number of arguments.
  19. Escaped Newlines: Slightly looser rules for escaped newlines.
  20. Extended Asm: Assembler instructions with C expressions as operands.
  21. Constraints: Constraints for asm operands.
  22. Asm Labels: Specifying the assembler name to use for a C symbol.
  23. Explicit Reg Vars: Defining variables residing in specified registers.
  24. Vector Extensions: Using vector instructions through built-in functions.
  25. Target Builtins: Built-in functions specific to particular targets.
  26. Subscripting: Any array can be subscripted, even if not an lvalue.
  27. Pointer Arith: Arithmetic on void-pointers and function pointers.
  28. Initializers: Non-constant initializers.
  29. Compound Literals: Compound literals give structures, unions, or arrays as values.
  30. Designated Inits: Labeling elements of initializers.
  31. Cast to Union: Casting to union type from any member of the union.
  32. Case Ranges: `case 1 ... 9' and such.
  33. Mixed Declarations: Mixing declarations and code.
  34. Function Prototypes: Prototype declarations and old-style definitions.
  35. C++ Comments: C++ comments are recognized.
  36. Dollar Signs: Dollar sign is allowed in identifiers.
  37. Character Escapes: \e stands for the character <ESC>.
  38. Alignment: Inquiring about the alignment of a type or variable.
  39. Inline: Defining inline functions (as fast as macros).
  40. Alternate Keywords:__const__, __asm__, etc., for header files.
  41. Incomplete Enums: enum foo;, with details to follow.
  42. Function Names: Printable strings which are the name of the current function.
  43. Return Address: Getting the return or frame address of a function.
  44. Unnamed Fields: Unnamed struct/union fields within structs/unions.
  45. Attribute Syntax: Formal syntax for attributes.

If you run into GCC extensions which have not been included in any of these lists, please let us know (also including whether or not they work).

The C++ front-end is considered to be fully tested and works for a number of non-trivial programs, including LLVM itself, Qt, Mozilla, etc.

• Exception handling only works well on the linux/X86-32 target. In some cases, illegally throwing an exception does not result in a call to terminate.

A wide variety of additional information is available on the LLVM web page, in particular in the documentation section. The web page also contains versions of the API documentation which is up-to-date with the Subversion version of the source code. You can access versions of these documents specific to this release by going into the "llvm/doc/" directory in the LLVM tree.

If you have any questions or comments about LLVM, please feel free to contact us via the mailing lists.