LLVM 2.7 Release Notes

This document contains the release notes for the LLVM Compiler Infrastructure, release 2.7. Here we describe the status of LLVM, including major improvements from the previous release and significant 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 release, please see the releases page.

The LLVM 2.7 distribution currently consists of code from the core LLVM repository (which roughly includes the LLVM optimizers, code generators and supporting tools), the Clang repository and the llvm-gcc repository. In addition to this code, the LLVM Project includes other sub-projects that are in development. Here we include updates on these subprojects.

The Clang project is ...

In the LLVM 2.7 time-frame, the Clang team has made many improvements:

FIXME: C++! Include a link to cxx_compatibility.html Clang 2.7 can bootstrap???
CIndex API and Python bindings: Clang now includes a C API as part of the CIndex library. Although we make make some changes to the API in the future, it is intended to be stable and has been designed for use by external projects. See the Clang doxygen CIndex documentation for more details. The CIndex API also includes a preliminary set of Python bindings.
ARM Support: Clang now has ABI support for both the Darwin and Linux ARM ABIs. Coupled with many improvements to the LLVM ARM backend, Clang is now suitable for use as a a beta quality ARM compiler.

The Clang Static Analyzer project is an effort to use static source code analysis techniques to automatically find bugs in C and Objective-C programs (and hopefully C++ in the future!). The tool is very good at finding bugs that occur on specific paths through code, such as on error conditions.

In the LLVM 2.7 time-frame, the analyzer core has made several major and minor improvements, including better support for tracking the fields of structures, initial support (not enabled by default yet) for doing interprocedural (cross-function) analysis, and new checks have been added.

The VMKit project is an implementation of a JVM and a CLI Virtual Machine (Microsoft .NET is an implementation of the CLI) using LLVM for static and just-in-time compilation.

With the release of LLVM 2.7, VMKit has shifted to a great framework for writing virtual machines. VMKit now offers precise and efficient garbage collection with multi-threading support, thanks to the MMTk memory management toolkit, as well as just in time and ahead of time compilation with LLVM. The major changes in VMKit 0.27 are:

Garbage collection: VMKit now uses the MMTk toolkit for garbage collectors. The first collector to be ported is the MarkSweep collector, which is precise, and drastically improves the performance of VMKit.
Line number information in the JVM: by using the debug metadata of LLVM, the JVM now supports precise line number information, useful when printing a stack trace.
Interface calls in the JVM: we implemented a variant of the Interface Method Table technique for interface calls in the JVM.

The new LLVM compiler-rt project is a simple library that provides an implementation of the low-level target-specific hooks required by code generation and other runtime components. For example, when compiling for a 32-bit target, converting a double to a 64-bit unsigned integer is compiled into a runtime call to the "__fixunsdfdi" function. The compiler-rt library provides highly optimized implementations of this and other low-level routines (some are 3x faster than the equivalent libgcc routines).

All of the code in the compiler-rt project is available under the standard LLVM License, a "BSD-style" license. New in LLVM 2.7: compiler_rt now supports ARM targets.

DragonEgg is a port of llvm-gcc to gcc-4.5. Unlike llvm-gcc, which makes many intrusive changes to the underlying gcc-4.2 code, dragonegg in theory does not require any gcc-4.5 modifications whatsoever (currently one small patch is needed). This is thanks to the new gcc plugin architecture, which makes it possible to modify the behaviour of gcc at runtime by loading a plugin, which is nothing more than a dynamic library which conforms to the gcc plugin interface. DragonEgg is a gcc plugin that causes the LLVM optimizers to be run instead of the gcc optimizers, and the LLVM code generators instead of the gcc code generators, just like llvm-gcc. To use it, you add "-fplugin=path/dragonegg.so" to the gcc-4.5 command line, and gcc-4.5 magically becomes llvm-gcc-4.5!

DragonEgg is still a work in progress. Currently C works very well, while C++, Ada and Fortran work fairly well. All other languages either don't work at all, or only work poorly. For the moment only the x86-32 and x86-64 targets are supported, and only on linux and darwin (darwin needs an additional gcc patch).

DragonEgg is a new project which is seeing its first release with llvm-2.7.

The LLVM Machine Code (aka MC) sub-project of LLVM was created to solve a number of problems in the realm of assembly, disassembly, object file format handling, and a number of other related areas that CPU instruction-set level tools work in. It is a sub-project of LLVM which provides it with a number of advantages over other compilers that do not have tightly integrated assembly-level tools. For a gentle introduction, please see the Intro to the LLVM MC Project Blog Post.

2.7 includes major parts of the work required by the new MC Project. A few targets have been refactored to support it, and work is underway to support a native assembler in LLVM. This work is not complete in LLVM 2.7, but you has made substantially more progress on LLVM mainline.

One minor example of what MC can do is to transcode an AT&T syntax X86 .s file into intel syntax. You can do this with something like:

  llvm-mc foo.s -output-asm-variant=1 -o foo-intel.s

An exciting aspect of LLVM is that it is used as an enabling technology for a lot of other language and tools projects. This section lists some of the projects that have already been updated to work with LLVM 2.7.

Pure is an algebraic/functional programming language based on term rewriting. Programs are collections of equations which are used to evaluate expressions in a symbolic fashion. Pure offers dynamic typing, eager and lazy evaluation, lexical closures, a hygienic macro system (also based on term rewriting), built-in list and matrix support (including list and matrix comprehensions) and an easy-to-use C interface. The interpreter uses LLVM as a backend to JIT-compile Pure programs to fast native code.

Pure versions 0.43 and later have been tested and are known to work with LLVM 2.7 (and continue to work with older LLVM releases >= 2.5).

Roadsend PHP (rphp) is an open source implementation of the PHP programming language that uses LLVM for its optimizer, JIT and static compiler. This is a reimplementation of an earlier project that is now based on LLVM.

Unladen Swallow is a branch of Python intended to be fully compatible and significantly faster. It uses LLVM's optimization passes and JIT compiler.

TCE is a toolset for designing application-specific processors (ASP) based on the Transport triggered architecture (TTA). The toolset provides a complete co-design flow from C/C++ programs down to synthesizable VHDL and parallel program binaries. Processor customization points include the register files, function units, supported operations, and the interconnection network.

TCE uses llvm-gcc/Clang and LLVM for C/C++ language support, target independent optimizations and also for parts of code generation. It generates new LLVM-based code generators "on the fly" for the designed TTA processors and loads them in to the compiler backend as runtime libraries to avoid per-target recompilation of larger parts of the compiler chain.

SAFECode is a memory safe C compiler built using LLVM. It takes standard, unannotated C code, analyzes the code to ensure that memory accesses and array indexing operations are safe, and instruments the code with run-time checks when safety cannot be proven statically.

This release includes a huge number of bug fixes, performance tweaks and minor improvements. Some of the major improvements and new features are listed in this section.

In addition to changes to the code, between LLVM 2.6 and 2.7, a number of organization changes have happened:

LLVM has a new official logo!
Ted Kremenek and Doug Gregor have stepped forward as Code Owners of the Clang static analyzer and the Clang frontend, respectively.
LLVM now has an official Blog at http://blog.llvm.org. This is a great way to learn about new LLVM-related features as they are implemented. Several features in this release are already explained on the blog.
The LLVM web pages are now checked into the SVN server, in the "www", "www-pubs" and "www-releases" SVN modules. Previously they were hidden in a largely inaccessible old CVS server.
llvm.org is now hosted on a new (and much faster) server. It is still graciously hosted at the University of Illinois of Urbana Champaign.

LLVM 2.7 includes several major new capabilities:

2.7 includes initial support for the MicroBlaze target. MicroBlaze is a soft processor core designed for Xilinx FPGAs.
2.7 includes a new LLVM IR "extensible metadata" feature. This feature supports many different use cases, including allowing front-end authors to encode source level information into LLVM IR, which is consumed by later language-specific passes. This is a great way to do high-level optimizations like devirtualization, type-based alias analysis, etc. See the Extensible Metadata Blog Post for more information.
2.7 encodes debug information in a completely new way, built on extensible metadata. The new implementation is much more memory efficient and paves the way for improvements to optimized code debugging experience.
2.7 now directly supports taking the address of a label and doing an indirect branch through a pointer. This is particularly useful for interpreter loops, and is used to implement the GCC "address of label" extension. For more information, see the Address of Label and Indirect Branches in LLVM IR Blog Post.
2.7 is the first release to start supporting APIs for assembling and disassembling target machine code. These APIs are useful for a variety of low level clients, and are surfaced in the new "enhanced disassembly" API. For more information see the The X86 Disassembler Blog Post for more information.
2.7 includes major parts of the work required by the new MC Project, see the MC update above for more information.

LLVM IR has several new features for better support of new targets and that expose new optimization opportunities:

LLVM IR now supports a 16-bit "half float" data type through two new intrinsics and APFloat support.
LLVM IR supports two new function attributes: inlinehint and alignstack(n). The former is a hint to the optimizer that a function was declared 'inline' and thus the inliner should weight it higher when considering inlining it. The later indicates to the code generator that the function diverges from the platform ABI on stack alignment.
The new llvm.objectsize intrinsic allows the optimizer to infer the sizes of memory objects in some cases. This intrinsic is used to implement the GCC __builtin_object_size extension.
LLVM IR now supports marking load and store instructions with "non-temporal" hints (building on the new metadata feature). This hint encourages the code generator to generate non-temporal accesses when possible, which are useful for code that is carefully managing cache behavior. Currently, only the X86 backend provides target support for this feature.
LLVM 2.7 has pre-alpha support for unions in LLVM IR. Unfortunately, this support is not really usable in 2.7, so if you're interested in pushing it forward, please help contribute to LLVM mainline.

In addition to a large array of minor performance tweaks and bug fixes, this release includes a few major enhancements and additions to the optimizers:

The inliner reuses now merges arrays stack objects in different callees when inlining multiple call sites into one function. This reduces the stack size of the resultant function.
The -basicaa alias analysis pass (which is the default) has been improved to be less dependent on "type safe" pointers. It can now look through bitcasts and other constructs more aggressively, allowing better load/store optimization.
The load elimination optimization in the GVN Pass [intro blog post] has been substantially improved to be more aggressive about partial redundancy elimination and do more aggressive phi translation. Please see the Advanced Topics in Redundant Load Elimination with a Focus on PHI Translation Blog Post for more details.
The module target data string now includes a notion of 'native' integer data types for the target. This helps mid-level optimizations avoid promoting complex sequences of operations to data types that are not natively supported (e.g. converting i32 operations to i64 on 32-bit chips).
The mid-level optimizer is now conservative when operating on a module with no target data. Previously, it would default to SparcV9 settings, which is not what most people expected.
Jump threading is now much more aggressive at simplifying correlated conditionals and threading blocks with otherwise complex logic. It has subsumed the old "Conditional Propagation" pass, and -condprop has been removed from LLVM 2.7.
The -instcombine pass has been refactored from being one huge file to being a library of its own. Internally, it uses a customized IRBuilder to clean it up and simplify it.
The optimal edge profiling pass is reliable and much more complete than in 2.6. It can be used with the llvm-prof tool but isn't wired up to the llvm-gcc and clang command line options yet.
A new experimental alias analysis implementation, -scev-aa, has been added. It uses LLVM's Scalar Evolution implementation to do symbolic analysis of pointer offset expressions to disambiguate pointers. It can catch a few cases that basicaa cannot, particularly in complex loop nests.
The default pass ordering has been tweaked for improved optimization effectiveness.

The JIT now supports generating debug information and is compatible with the new GDB 7.0 (and later) interfaces for registering dynamically generated debug info.
The JIT now defaults to compiling eagerly to avoid a race condition in the lazy JIT. Clients that still want the lazy JIT can switch it on by calling ExecutionEngine::DisableLazyCompilation(false).
It is now possible to create more than one JIT instance in the same process. These JITs can generate machine code in parallel, although you still have to obey the other threading restrictions.

We have put a significant amount of work into the code generator infrastructure, which allows us to implement more aggressive algorithms and make it run faster:

The 'llc -asm-verbose' option (which is now the default) has been enhanced to emit many useful comments to .s files indicating information about spill slots and loop nest structure. This should make it much easier to read and understand assembly files. This is wired up in llvm-gcc and clang to the -fverbose-asm option.
New LSR with "full strength reduction" mode. FIXME: Description?
A new codegen level Common Subexpression Elimination pass (MachineCSE) is available and enabled by default. It catches redundancies exposed by lowering.
A new pre-register-allocation tail duplication pass is available and enabled by default, it can substantially improve branch prediction quality in some cases.
A new sign and zero extension optimization pass (OptimizeExtsPass) is available and enabled by default. This pass can takes advantage architecture features like x86-64 implicit zero extension behavior and sub-registers.
The code generator now supports a mode where it attempts to preserve the order of instructions in the input code. This is important for source that is hand scheduled and extremely sensitive to scheduling. It is compatible with the GCC -fno-schedule-insns option.
The target-independent code generator now supports generating code with arbitrary numbers of result values. Returning more values than was previously supported is handled by returning through a hidden pointer. In 2.7, only the X86 and XCore targets have adopted support for this though.
The code generator now supports generating code that follows the Glasgow Haskell Compiler Calling Convention and ABI.
The "DAG instruction selection" phase of the code generator has been largely rewritten for 2.7. Previously, tblgen spit out tons of C++ code which was compiled and linked into the target to do the pattern matching, now it emits a much smaller table which is read by the target-independent code. The primary advantages of this approach is that the size and compile time of various targets is much improved. The X86 code generator shrunk by 1.5MB of code, for example.
Almost the entire code generator has switched to emitting code through the MC interfaces instead of printing textually to the .s file. This led to a number of cleanups and speedups. In 2.7, debug an exception handling information does not go through MC yet.

New features of the X86 target include:

The X86 backend now optimizes tails calls much more aggressively for functions that use the standard C calling convention.
The X86 backend now models scalar SSE registers as subregs of the SSE vector registers, making the code generator more aggressive in cases where scalars and vector types are mixed.
PostRA scheduler for X86? FIXME: is this on by default in 2.7?

New features of the ARM target include:

The ARM backend now generates instructions in unified assembly syntax.
llvm-gcc now has complete support for the ARM v7 NEON instruction set. This support differs slightly from the GCC implementation. Please see the ARM Advanced SIMD (NEON) Intrinsics and Types in LLVM Blog Post for helpful information if migrating code from GCC to LLVM-GCC.
The ARM and Thumb code generators now use register scavenging for stack object address materialization.(FIXME: WHAT BENEFIT DOES THIS PROVIDE?)
The ARM backend now has good support for ARMv4 targets and has been tested on StrongARM hardware. Previously, LLVM only supported ARMv4T and newer chips.

This release includes a number of new APIs that are used internally, which may also be useful for external clients.

The optimizer uses the new CodeMetrics class to measure the size of code. Various passes that use thing (like the inliner, loop unswitcher, etc) all use this to make more accurate estimates of the code size impact of various optimizations.
A new llvm/Analysis/InstructionSimplify.h interface is available for doing symbolic simplification of instructions (e.g. a+0 -> a) without requiring the instruction to exist. This centralizes a lot of ad-hoc symbolic manipulation code scattered in various passes.
The optimizer now uses a new SSAUpdater class which efficiently supports doing unstructured SSA update operations. This centralized a bunch of code scattered throughout various passes (e.g. jump threading, lcssa, loop rotate, etc) for doing this sort of thing. The code generator has a similar MachineSSAUpdater class.
The llvm/Support/Regex.h header exposes a platform independent regular expression API. Building on this, the FileCheck utility now supports regular exressions.
raw_ostream now supports a circular "debug stream" accessed with "dbgs()". By default, this stream works the same way as "errs()", but if you pass -debug-buffer-size=1000 to opt, the debug stream is capped to a fixed sized circular buffer and the output is printed at the end of the program's execution. This is helpful if you have a long lived compiler process and you're interested in seeing snapshots in time.

Other miscellaneous features include:

You can now build LLVM as a big dynamic library (e.g. "libllvm2.7.so"). To get this, configure LLVM with the --enable-shared option.
LLVM command line tools now overwrite their output by default. Previously, they would only do this with -f. This makes them more convenient to use, and behave more like standard unix tools.
The opt and llc tools now autodetect whether their input is a .ll or .bc file, and automatically do the right thing. This means you don't need to explicitly use the llvm-as tool for most things.

If you're already an LLVM user or developer with out-of-tree changes based on LLVM 2.6, this section lists some "gotchas" that you may run into upgrading from the previous release.

The Andersen's alias analysis ("anders-aa") pass, the Predicate Simplifier ("predsimplify") pass, the LoopVR pass, the GVNPRE pass, and the random sampling profiling ("rsprofiling") passes have all been removed. They were not being actively maintained and had substantial problems. If you are interested in these components, you are welcome to ressurect them from SVN, fix the correctness problems, and resubmit them to mainline.
LLVM now defaults to building most libraries with RTTI turned off, providing a code size reduction. Packagers who are interested in building LLVM to support plugins that require RTTI information should build with "make REQUIRE_RTTI=1" and should read the new Advice on Packaging LLVM document.
The LLVM interpreter now defaults to not using libffi even if you have it installed. This makes it more likely that an LLVM built on one system will work when copied to a similar system. To use libffi, configure with --enable-libffi.
Debug information uses a completely different representation, an LLVM 2.6 .bc file should work with LLVM 2.7, but debug info won't come forward.
The LLVM 2.6 (and earlier) "malloc" and "free" instructions got removed, along with LowerAllocations pass. Now you should just use a call to the malloc and free functions in libc. These calls are optimized as well as the old instructions were.

In addition, many APIs have changed in this release. Some of the major LLVM API changes are:

Just about everything has been converted to use raw_ostream instead of std::ostream.
llvm/ADT/iterator.h has been removed, just use <iterator> instead.
The Streams.h file and "DOUT" got removed, use "DEBUG(errs() << ...);" instead.
The TargetAsmInfo interface was renamed to MCAsmInfo.
ModuleProvider has been removed and its methods moved to Module and GlobalValue. Most clients can remove uses of ExistingModuleProvider, replace getBitcodeModuleProvider with getLazyBitcodeModule, and pass their Module to functions that used to accept ModuleProvider. Clients who wrote their own ModuleProviders will need to derive from GVMaterializer instead and use Module::setMaterializer to attach it to a Module.
GhostLinkage has given up the ghost. GlobalValues that have not yet been read from their backing storage have the same linkage they will have after being read in. Clients must replace calls to GlobalValue::hasNotBeenReadFromBitcode with GlobalValue::isMaterializable.
The llvm/Support/DataTypes.h header has moved to llvm/System/DataTypes.h.
The isInteger, isIntOrIntVector, isFloatingPoint, isFPOrFPVector and isFPOrFPVector methods have been renamed isIntegerTy, isIntOrIntVectorTy, isFloatingPointTy, isFPOrFPVectorTy and isFPOrFPVectorTy respectively.
llvm::Instruction::clone() no longer takes argument.
raw_fd_ostream's constructor now takes a flag argument, not individual booleans (see include/llvm/Support/raw_ostream.h for details).
Some header files have been renamed:
- llvm/Support/AIXDataTypesFix.h to llvm/System/AIXDataTypesFix.h
- llvm/Support/DataTypes.h to llvm/System/DataTypes.h
- llvm/Transforms/Utils/InlineCost.h to llvm/Analysis/InlineCost.h
- llvm/Support/Mangler.h to llvm/Target/Mangler.h
- llvm/Analysis/Passes.h to llvm/CodeGen/Passes.h

LLVM is known to work on the following platforms:

Intel and AMD machines (IA32, X86-64, AMD64, EMT-64) running Red Hat Linux, Fedora Core, FreeBSD and AuroraUX (and probably other unix-like systems).
PowerPC and X86-based Mac OS X systems, running 10.4 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 x86 and AMD64 machines running Solaris 10, OpenSolaris 0906.
Alpha-based machines running Debian GNU/Linux.

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 significant known problems with the LLVM system, listed by component. If you run into a problem, please check the LLVM bug database and submit a bug if there isn't already one.

LLVM will not correctly compile on Solaris and/or OpenSolaris using the stock GCC 3.x.x series 'out the box', See: Broken versions of GCC and other tools. However, A Modern GCC Build for x86/x86-64 has been made available from the third party AuroraUX Project that has been meticulously tested for bootstrapping LLVM & Clang.

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 MSIL, Alpha, SPU, MIPS, PIC16, Blackfin, MSP430, SystemZ and MicroBlaze backends are experimental.
llc "-filetype=asm" (the default) is the only supported value for this option. The MachO writer is experimental, and works much better in mainline SVN.

The X86 backend does not yet support all inline assembly that uses the X86 floating point stack. It supports the 'f' and 't' constraints, but not 'u'.
The X86 backend generates inefficient floating point code when configured to generate code for systems that don't have SSE2.
Win64 code generation wasn't widely tested. Everything should work, but we expect small issues to happen. Also, llvm-gcc cannot build the mingw64 runtime currently due to lack of support for the 'u' inline assembly constraint and for X87 floating point inline assembly.
The X86-64 backend does not yet support the LLVM IR instruction va_arg. Currently, front-ends support variadic argument constructs on X86-64 by lowering them manually.

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.

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

64-bit MIPS targets are not supported yet.

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

The C backend has only basic support for inline assembly code.
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.
The C backend does not support all exception handling constructs.
The C backend does not support arbitrary precision integers.

The only major language feature of GCC not supported by llvm-gcc is the __builtin_apply family of builtins. However, some extensions are only supported on some targets. For example, trampolines are only supported on some targets (these are used when you take the address of a nested function).

Fortran support generally works, but there are still several unresolved bugs in Bugzilla. Please see the tools/gfortran component for details.

The llvm-gcc 4.2 Ada compiler works fairly well; however, this is not a mature technology, and problems should be expected.

The Ada front-end currently only builds on X86-32. This is mainly due to lack of trampoline support (pointers to nested functions) on other platforms. However, it also fails to build on X86-64 which does support trampolines.
The Ada front-end fails to bootstrap. This is due to lack of LLVM support for setjmp/longjmp style exception handling, which is used internally by the compiler. Workaround: configure with --disable-bootstrap.
The c380004, c393010 and cxg2021 ACATS tests fail (c380004 also fails with gcc-4.2 mainline). If the compiler is built with checks disabled then c393010 causes the compiler to go into an infinite loop, using up all system memory.
Some GCC specific Ada tests continue to crash the compiler.
The -E binder option (exception backtraces) does not work and will result in programs crashing if an exception is raised. Workaround: do not use -E.
Only discrete types are allowed to start or finish at a non-byte offset in a record. Workaround: do not pack records or use representation clauses that result in a field of a non-discrete type starting or finishing in the middle of a byte.
The lli interpreter considers 'main' as generated by the Ada binder to be invalid. Workaround: hand edit the file to use pointers for argv and envp rather than integers.
The -fstack-check option is ignored.