LLVM 2.8 Release Notes

This document contains the release notes for the LLVM Compiler Infrastructure, release 2.8. 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 2.8 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.

Clang is an LLVM front end for the C, C++, and Objective-C languages. Clang aims to provide a better user experience through expressive diagnostics, a high level of conformance to language standards, fast compilation, and low memory use. Like LLVM, Clang provides a modular, library-based architecture that makes it suitable for creating or integrating with other development tools. Clang is considered a production-quality compiler for C, Objective-C, C++ and Objective-C++ on x86 (32- and 64-bit), and for darwin-arm targets.

In the LLVM 2.8 time-frame, the Clang team has made many improvements:

Clang C++ is now feature-complete with respect to the ISO C++ 1998 and 2003 standards.
Added support for Objective-C++.
Clang now uses LLVM-MC to directly generate object code and to parse inline assembly (on Darwin).
Introduced many new warnings, including -Wmissing-field-initializers, -Wshadow, -Wno-protocol, -Wtautological-compare, -Wstrict-selector-match, -Wcast-align, -Wunused improvements, and greatly improved format-string checking.
Introduced the "libclang" library, a C interface to Clang intended to support IDE clients.
Added support for #pragma GCC visibility, #pragma align, and others.
Added support for SSE, AVX, ARM NEON, and AltiVec.
Improved support for many Microsoft extensions.
Implemented support for blocks in C++.
Implemented precompiled headers for C++.
Improved abstract syntax trees to retain more accurate source information.
Added driver support for handling LLVM IR and bitcode files directly.
Major improvements to compiler correctness for exception handling.
Improved generated code quality in some areas:
- Good code generation for X86-32 and X86-64 ABI handling.
- Improved code generation for bit-fields, although important work remains.

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.

The LLVM 2.8 release fixes a number of bugs and slightly improves precision over 2.7, but there are no major new features in the release.

DragonEgg is a port of llvm-gcc to gcc-4.5. Unlike llvm-gcc, dragonegg in theory does not require any gcc-4.5 modifications whatsoever (currently one small patch is needed) thanks to the new gcc plugin architecture. DragonEgg is a gcc plugin that makes gcc-4.5 use the LLVM optimizers and code generators instead of gcc's, just like with llvm-gcc.

DragonEgg is still a work in progress, but it is able to compile a lot of code, for example all of gcc, LLVM and clang. Currently Ada, C, C++ and Fortran work well, while 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 may need additional gcc patches).

The 2.8 release has the following notable changes:

The plugin loads faster due to exporting fewer symbols.
Additional vector operations such as addps256 are now supported.
Ada global variables with no initial value are no longer zero initialized, resulting in better optimization.
The '-fplugin-arg-dragonegg-enable-gcc-optzns' flag now runs all gcc optimizers, rather than just a handful.
Fortran programs using common variables now link correctly.
GNU OMP constructs no longer crash the compiler.

The VMKit project is an implementation of a Java Virtual Machine (Java VM or JVM) that uses LLVM for static and just-in-time compilation. As of LLVM 2.8, VMKit now supports copying garbage collectors, and can be configured to use MMTk's copy mark-sweep garbage collector. In LLVM 2.8, the VMKit .NET VM is no longer being maintained.

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 2.8, compiler_rt now supports soft floating point (for targets that don't have a real floating point unit), and includes an extensive testsuite for the "blocks" language feature and the blocks runtime included in compiler_rt.

LLDB is a brand new member of the LLVM umbrella of projects. LLDB is a next generation, high-performance debugger. It is built as a set of reusable components which highly leverage existing libraries in the larger LLVM Project, such as the Clang expression parser, the LLVM disassembler and the LLVM JIT.

LLDB is in early development and not included as part of the LLVM 2.8 release, but is mature enough to support basic debugging scenarios on Mac OS X in C, Objective-C and C++. We'd really like help extending and expanding LLDB to support new platforms, new languages, new architectures, and new features.

libc++ is another new member of the LLVM family. It is an implementation of the C++ standard library, written from the ground up to specifically target the forthcoming C++'0X standard and focus on delivering great performance.

As of the LLVM 2.8 release, libc++ is virtually feature complete, but would benefit from more testing and better integration with Clang++. It is also looking forward to the C++ committee finalizing the C++'0x standard.

KLEE is a symbolic execution framework for programs in LLVM bitcode form. KLEE tries to symbolically evaluate "all" paths through the application and records state transitions that lead to fault states. This allows it to construct testcases that lead to faults and can even be used to verify some algorithms.

Although KLEE does not have any major new features as of 2.8, we have made various minor improvements, particular to ease development:

Added support for LLVM 2.8. KLEE currently maintains compatibility with LLVM 2.6, 2.7, and 2.8.
Added a buildbot for 2.6, 2.7, and trunk. A 2.8 buildbot will be coming soon following release.
Fixed many C++ code issues to allow building with Clang++. Mostly complete, except for the version of MiniSAT which is inside the KLEE STP version.
Improved support for building with separate source and build directories.
Added support for "long double" on x86.
Initial work on KLEE support for using 'lit' test runner instead of DejaGNU.
Added configure support for using an external version of STP.

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 2.8.

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.

Horizon is a bytecode language and compiler written on top of LLVM, intended for producing single-address-space managed code operating systems that run faster than the equivalent multiple-address-space C systems. More in-depth blurb is available on the wiki.

Clam AntiVirus is an open source (GPL) anti-virus toolkit for UNIX, designed especially for e-mail scanning on mail gateways. Since version 0.96 it has bytecode signatures that allow writing detections for complex malware. It uses LLVM's JIT to speed up the execution of bytecode on X86, X86-64, PPC32/64, falling back to its own interpreter otherwise. The git version was updated to work with LLVM 2.8.

The ClamAV bytecode compiler uses Clang and LLVM to compile a C-like language, insert runtime checks, and generate ClamAV bytecode.

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 0.44 and later have been tested and are known to work with LLVM 2.8 (and continue to work with older LLVM releases >= 2.5).

GHC is an open source, state-of-the-art programming suite for Haskell, a standard lazy functional programming language. It includes an optimizing static compiler generating good code for a variety of platforms, together with an interactive system for convenient, quick development.

In addition to the existing C and native code generators, GHC 7.0 now supports an LLVM code generator. GHC supports LLVM 2.7 and later.

Clay is a new systems programming language that is specifically designed for generic programming. It makes generic programming very concise thanks to whole program type propagation. It uses LLVM as its backend.

llvm-py has been updated to work with LLVM 2.8. llvm-py provides Python bindings for LLVM, allowing you to write a compiler backend or a VM in Python.

FAUST is a compiled language for real-time audio signal processing. The name FAUST stands for Functional AUdio STream. Its programming model combines two approaches: functional programming and block diagram composition. In addition with the C, C++, JAVA output formats, the Faust compiler can now generate LLVM bitcode, and works with LLVM 2.7 and 2.8.

Jade (Just-in-time Adaptive Decoder Engine) is a generic video decoder engine using LLVM for just-in-time compilation of video decoder configurations. Those configurations are designed by MPEG Reconfigurable Video Coding (RVC) committee. MPEG RVC standard is built on a stream-based dataflow representation of decoders. It is composed of a standard library of coding tools written in RVC-CAL language and a dataflow configuration — block diagram — of a decoder.

Jade project is hosted as part of the Open RVC-CAL Compiler and requires it to translate the RVC-CAL standard library of video coding tools into an LLVM assembly code.

Neko LLVM JIT replaces the standard Neko JIT with an LLVM-based implementation. While not fully complete, it is already providing a 1.5x speedup on 64-bit systems. Neko LLVM JIT requires LLVM 2.8 or later.

Crack aims to provide the ease of development of a scripting language with the performance of a compiled language. The language derives concepts from C++, Java and Python, incorporating object-oriented programming, operator overloading and strong typing. Crack 0.2 works with LLVM 2.7, and the forthcoming Crack 0.2.1 release builds on LLVM 2.8.

DTMC provides support for Transactional Memory, which is an easy-to-use and efficient way to synchronize accesses to shared memory. Transactions can contain normal C/C++ code (e.g., __transaction { list.remove(x); x.refCount--; }) and will be executed virtually atomically and isolated from other transactions.

Kai (Japanese 会 for meeting/gathering) is an experimental interpreter that provides a highly extensible runtime environment and explicit control over the compilation process. Programs are defined using nested symbolic expressions, which are all parsed into first-class values with minimal intrinsic semantics. Kai can generate optimised code at run-time (using LLVM) in order to exploit the nature of the underlying hardware and to integrate with external software libraries. It is a unique exploration into world of dynamic code compilation, and the interaction between high level and low level semantics.

OSL is a shading language designed for use in physically based renderers and in particular production rendering. By using LLVM instead of the interpreter, it was able to meet its performance goals (>= C-code) while retaining the benefits of runtime specialization and a portable high-level language.

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.

LLVM 2.8 includes several major new capabilities:

As mentioned above, libc++ and LLDB are major new additions to the LLVM collective.
LLVM 2.8 now has pretty decent support for debugging optimized code. You should be able to reliably get debug info for function arguments, assuming that the value is actually available where you have stopped.
A new 'llvm-diff' tool is available that does a semantic diff of .ll files.
The MC subproject has made major progress in this release. Direct .o file writing support for darwin/x86[-64] is now reliable and support for other targets and object file formats are in progress.

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

The memcpy, memmove, and memset intrinsics now take address space qualified pointers and a bit to indicate whether the transfer is "volatile" or not.
Per-instruction debug info metadata is much faster and uses less memory by using the new DebugLoc class.
LLVM IR now has a more formalized concept of "trap values", which allow the optimizer to optimize more aggressively in the presence of undefined behavior, while still producing predictable results.
LLVM IR now supports two new linkage types (linker_private_weak and linker_private_weak_def_auto) which map onto some obscure MachO concepts.

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:

As mentioned above, the optimizer now has support for updating debug information as it goes. A key aspect of this is the new llvm.dbg.value intrinsic. This intrinsic represents debug info for variables that are promoted to SSA values (typically by mem2reg or the -scalarrepl passes).
The JumpThreading pass is now much more aggressive about implied value relations, allowing it to thread conditions like "a == 4" when a is known to be 13 in one of the predecessors of a block. It does this in conjunction with the new LazyValueInfo analysis pass.
The new RegionInfo analysis pass identifies single-entry single-exit regions in the CFG. You can play with it with the "opt -regions -analyze" or "opt -view-regions" commands.
The loop optimizer has significantly improved strength reduction and analysis capabilities. Notably it is able to build on the trap value and signed integer overflow information to optimize <= and >= loops.
The CallGraphSCCPassManager now has some basic support for iterating within an SCC when a optimizer devirtualizes a function call. This allows inlining through indirect call sites that are devirtualized by store-load forwarding and other optimizations.
The new -loweratomic pass is available to lower atomic instructions into their non-atomic form. This can be useful to optimize generic code that expects to run in a single-threaded environment.

The LLVM Machine Code (aka MC) subsystem 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.

The MC subproject has made great leaps in LLVM 2.8. For example, support for directly writing .o files from LLC (and clang) now works reliably for darwin/x86[-64] (including inline assembly support) and the integrated assembler is turned on by default in Clang for these targets. This provides improved compile times among other things.

The entire compiler has converted over to using the MCStreamer assembler API instead of writing out a .s file textually.
The "assembler parser" is far more mature than in 2.7, supporting a full complement of directives, now supports assembler macros, etc.
The "assembler backend" has been completed, including support for relaxation relocation processing and all the other things that an assembler does.
The MachO file format support is now fully functional and works.
The MC disassembler now fully supports ARM and Thumb. ARM assembler support is still in early development though.
The X86 MC assembler now supports the X86 AES and AVX instruction set.
Work on ELF and COFF object files and ARM target support is well underway, but isn't useful yet in LLVM 2.8. Please contact the llvmdev mailing list if you're interested in this.

For more information, please see the Intro to the LLVM MC Project Blog Post.

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 clang/gcc -momit-leaf-frame-pointer argument is now supported.
The clang/gcc -ffunction-sections and -fdata-sections arguments are now supported on ELF targets (like GCC).
The MachineCSE pass is now tuned and on by default. It eliminates common subexpressions that are exposed when lowering to machine instructions.
The "local" register allocator was replaced by a new "fast" register allocator. This new allocator (which is often used at -O0) is substantially faster and produces better code than the old local register allocator.
A new LLC "-regalloc=default" option is available, which automatically chooses a register allocator based on the -O optimization level.
The common code generator code was modified to promote illegal argument and return value vectors to wider ones when possible instead of scalarizing them. For example, <3 x float> will now pass in one SSE register instead of 3 on X86. This generates substantially better code since the rest of the code generator was already expecting this.
The code generator uses a new "COPY" machine instruction. This speeds up the code generator and eliminates the need for targets to implement the isMoveInstr hook. Also, the copyRegToReg hook was renamed to copyPhysReg and simplified.
The code generator now has a "LocalStackSlotPass", which optimizes stack slot access for targets (like ARM) that have limited stack displacement addressing.
A new "PeepholeOptimizer" is available, which eliminates sign and zero extends, and optimizes away compare instructions when the condition result is available from a previous instruction.
Atomic operations now get legalized into simpler atomic operations if not natively supported, easing the implementation burden on targets.
We have added two new bottom-up pre-allocation register pressure aware schedulers:
1. The hybrid scheduler schedules aggressively to minimize schedule length when registers are available and avoid overscheduling in high pressure situations.
2. The instruction-level-parallelism scheduler schedules for maximum ILP when registers are available and avoid overscheduling in high pressure situations.
The tblgen type inference algorithm was rewritten to be more consistent and diagnose more target bugs. If you have an out-of-tree backend, you may find that it finds bugs in your target description. This support also allows limited support for writing patterns for instructions that return multiple results (e.g. a virtual register and a flag result). The 'parallel' modifier in tblgen was removed, you should use the new support for multiple results instead.
A new (experimental) "-rendermf" pass is available which renders a MachineFunction into HTML, showing live ranges and other useful details.
The new SubRegIndex tablegen class allows subregisters to be indexed symbolically instead of numerically. If your target uses subregisters you will need to adapt to use SubRegIndex when you upgrade to 2.8.
The -fast-isel instruction selection path (used at -O0 on X86) was rewritten to work bottom-up on basic blocks instead of top down. This makes it slightly faster (because the MachineDCE pass is not needed any longer) and allows it to generate better code in some cases.

New features and major changes in the X86 target include:

The X86 backend now supports holding X87 floating point stack values in registers across basic blocks, dramatically improving performance of code that uses long double, and when targeting CPUs that don't support SSE.
The X86 backend now uses a SSEDomainFix pass to optimize SSE operations. On Nehalem ("Core i7") and newer CPUs there is a 2 cycle latency penalty on using a register in a different domain than where it was defined. This pass optimizes away these stalls.
The X86 backend now promotes 16-bit integer operations to 32-bits when possible. This avoids 0x66 prefixes, which are slow on some microarchitectures and bloat the code on all of them.
The X86 backend now supports the Microsoft "thiscall" calling convention, and a calling convention to support ghc.
The X86 backend supports a new "llvm.x86.int" intrinsic, which maps onto the X86 "int $42" and "int3" instructions.
At the IR level, the <2 x float> datatype is now promoted and passed around as a <4 x float> instead of being passed and returned as an MMX vector. If you have a frontend that uses this, please pass and return a <2 x i32> instead (using bitcasts).
When printing .s files in verbose assembly mode (the default for clang -S), the X86 backend now decodes X86 shuffle instructions and prints human readable comments after the most inscrutable of them, e.g.:
```
  insertps $113, %xmm3, %xmm0 # xmm0 = zero,xmm0[1,2],xmm3[1]
  unpcklps %xmm1, %xmm0       # xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]
  pshufd   $1, %xmm1, %xmm1   # xmm1 = xmm1[1,0,0,0]
```

New features of the ARM target include:

The ARM backend now optimizes tail calls into jumps.
Scheduling is improved through the new list-hybrid scheduler as well as through better modeling of structural hazards.
Half float instructions are now supported.
NEON support has been improved to model instructions which operate onto multiple consecutive registers more aggressively. This avoids lots of extraneous register copies.
The ARM backend now uses a new "ARMGlobalMerge" pass, which merges several global variables into one, saving extra address computation (all the global variables can be accessed via same base address) and potentially reducing register pressure.
The ARM backend has received many minor improvements and tweaks which lead to substantially better performance in a wide range of different scenarios.
The ARM NEON intrinsics have been substantially reworked to reduce redundancy and improve code generation. Some of the major changes are:
1. All of the NEON load and store intrinsics (llvm.arm.neon.vld* and llvm.arm.neon.vst*) take an extra parameter to specify the alignment in bytes of the memory being accessed.
2. The llvm.arm.neon.vaba intrinsic (vector absolute difference and accumulate) has been removed. This operation is now represented using the llvm.arm.neon.vabd intrinsic (vector absolute difference) followed by a vector add.
3. The llvm.arm.neon.vabdl and llvm.arm.neon.vabal intrinsics (lengthening vector absolute difference with and without accumulation) have been removed. They are represented using the llvm.arm.neon.vabd intrinsic (vector absolute difference) followed by a vector zero-extend operation, and for vabal, a vector add.
4. The llvm.arm.neon.vmovn intrinsic has been removed. Calls of this intrinsic are now replaced by vector truncate operations.
5. The llvm.arm.neon.vmovls and llvm.arm.neon.vmovlu intrinsics have been removed. They are now represented as vector sign-extend (vmovls) and zero-extend (vmovlu) operations.
6. The llvm.arm.neon.vaddl*, llvm.arm.neon.vaddw*, llvm.arm.neon.vsubl*, and llvm.arm.neon.vsubw* intrinsics (lengthening vector add and subtract) have been removed. They are replaced by vector add and vector subtract operations where one (vaddw, vsubw) or both (vaddl, vsubl) of the operands are either sign-extended or zero-extended.
7. The llvm.arm.neon.vmulls, llvm.arm.neon.vmullu, llvm.arm.neon.vmlal*, and llvm.arm.neon.vmlsl* intrinsics (lengthening vector multiply with and without accumulation and subtraction) have been removed. These operations are now represented as vector multiplications where the operands are either sign-extended or zero-extended, followed by a vector add for vmlal or a vector subtract for vmlsl. Note that the polynomial vector multiply intrinsic, llvm.arm.neon.vmullp, remains unchanged.

If you're already an LLVM user or developer with out-of-tree changes based on LLVM 2.7, this section lists some "gotchas" that you may run into upgrading from the previous release.

The build configuration machinery changed the output directory names. It wasn't clear to many people that a "Release-Asserts" build was a release build without asserts. To make this more clear, "Release" does not include assertions and "Release+Asserts" does (likewise, "Debug" and "Debug+Asserts").
The MSIL Backend was removed, it was unsupported and broken.
The ABCD, SSI, and SCCVN passes were removed. These were not fully functional and their behavior has been or will be subsumed by the LazyValueInfo pass.
The LLVM IR 'Union' feature was removed. While this is a desirable feature for LLVM IR to support, the existing implementation was half baked and barely useful. We'd really like anyone interested to resurrect the work and finish it for a future release.
If you're used to reading .ll files, you'll probably notice that .ll file dumps don't produce #uses comments anymore. To get them, run a .bc file through "llvm-dis --show-annotations".
Target triples are now stored in a normalized form, and all inputs from humans are expected to be normalized by Triple::normalize before being stored in a module triple or passed to another library.

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

LLVM 2.8 changes the internal order of operands in InvokeInst and CallInst. To be portable across releases, please use the CallSite class and the high-level accessors, such as getCalledValue and setUnwindDest.
You can no longer pass use_iterators directly to cast<> (and similar), because these routines tend to perform costly dereference operations more than once. You have to dereference the iterators yourself and pass them in.
llvm.memcpy.*, llvm.memset.*, llvm.memmove.* intrinsics take an extra parameter now ("i1 isVolatile"), totaling 5 parameters, and the pointer operands are now address-space qualified. If you were creating these intrinsic calls and prototypes yourself (as opposed to using Intrinsic::getDeclaration), you can use UpgradeIntrinsicFunction/UpgradeIntrinsicCall to be portable across releases.
SetCurrentDebugLocation takes a DebugLoc now instead of a MDNode. Change your code to use SetCurrentDebugLocation(DebugLoc::getFromDILocation(...)).
The RegisterPass and RegisterAnalysisGroup templates are considered deprecated, but continue to function in LLVM 2.8. Clients are strongly advised to use the upcoming INITIALIZE_PASS() and INITIALIZE_AG_PASS() macros instead.
The constructor for the Triple class no longer tries to understand odd triple specifications. Frontends should ensure that they only pass valid triples to LLVM. The Triple::normalize utility method has been added to help front-ends deal with funky triples.
The signature of the GCMetadataPrinter::finishAssembly virtual function changed: the raw_ostream and MCAsmInfo arguments were dropped. GC plugins which compute stack maps must be updated to avoid having the old definition overload the new signature.
The signature of MemoryBuffer::getMemBuffer changed. Unfortunately calls intended for the old version still compile, but will not work correctly, leading to a confusing error about an invalid header in the bitcode.
Some APIs were renamed:
- llvm_report_error -> report_fatal_error
- llvm_install_error_handler -> install_fatal_error_handler
- llvm::DwarfExceptionHandling -> llvm::JITExceptionHandling
- VISIBILITY_HIDDEN -> LLVM_LIBRARY_VISIBILITY
Some public headers were renamed:
- llvm/Assembly/AsmAnnotationWriter.h was renamed to llvm/Assembly/AssemblyAnnotationWriter.h

This section lists changes to the LLVM development infrastructure. This mostly impacts users who actively work on LLVM or follow development on mainline, but may also impact users who leverage the LLVM build infrastructure or are interested in LLVM qualification.

The default for make check is now to use the lit testing tool, which is part of LLVM itself. You can use lit directly as well, or use the llvm-lit tool which is created as part of a Makefile or CMake build (and knows how to find the appropriate tools). See the lit documentation and the blog post, and PR5217 for more information.
The LLVM test-suite infrastructure has a new "simple" test format (make TEST=simple). The new format is intended to require only a compiler and not a full set of LLVM tools. This makes it useful for testing released compilers, for running the test suite with other compilers (for performance comparisons), and makes sure that we are testing the compiler as users would see it. The new format is also designed to work using reference outputs instead of comparison to a baseline compiler, which makes it run much faster and makes it less system dependent.
Significant progress has been made on a new interface to running the LLVM test-suite (aka the LLVM "nightly tests") using the LNT infrastructure. The LNT interface to the test-suite brings significantly improved reporting capabilities for monitoring the correctness and generated code quality produced by LLVM over time.

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.

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 Alpha, Blackfin, CellSPU, MicroBlaze, MSP430, MIPS, SystemZ and XCore backends are experimental.
llc "-filetype=obj" is experimental on all targets other than darwin-i386 and darwin-x86_64.

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'.
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 numerous problems and is not being actively maintained. Depending on it for anything serious is not advised.

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.

llvm-gcc is generally very stable for the C family of languages. 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. Note that llvm-gcc is missing major Fortran performance work in the frontend and library that went into GCC after 4.2. If you are interested in Fortran, we recommend that you consider using dragonegg instead.

The llvm-gcc 4.2 Ada compiler has basic functionality, but is no longer being actively maintained. If you are interested in Ada, we recommend that you consider using dragonegg instead.