LLVM 3.1 Release Notes

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 3.1 time-frame, the Clang team has made many improvements. Highlights include:

• Greatly expanded C++11 support including lambdas, initializer lists, constexpr, user-defined literals, and atomics.
• A new tooling library to ease building of clang-based standalone tools.
• Extended support for literals in Objective C.

For more details about the changes to Clang since the 3.0 release, see the Clang release notes.

If Clang rejects your code but another compiler accepts it, please take a look at the language compatibility guide to make sure this is not intentional or a known issue.

DragonEgg is a gcc plugin that replaces GCC's optimizers and code generators with LLVM's. It works with gcc-4.5 and gcc-4.6 (and partially with gcc-4.7), can target the x86-32/x86-64 and ARM processor families, and has been successfully used on the Darwin, FreeBSD, KFreeBSD, Linux and OpenBSD platforms. It fully supports Ada, C, C++ and Fortran. It has partial support for Go, Java, Obj-C and Obj-C++.

The 3.1 release has the following notable changes:

• Partial support for gcc-4.7. Ada support is poor, but other languages work fairly well.
• Support for ARM processors. Some essential gcc headers that are needed to build DragonEgg for ARM are not installed by gcc. To work around this, copy the missing headers from the gcc source tree.
• Better optimization for Fortran by exploiting the fact that Fortran scalar arguments have 'restrict' semantics.
• Better optimization for all languages by passing information about type aliasing and type ranges to the LLVM optimizers.
• A regression test-suite was added.

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).

As of 3.1, compiler-rt includes the helper functions for atomic operations, allowing atomic operations on arbitrary-sized quantities to work. These functions follow the specification defined by gcc and are used by clang.

LLDB is a ground-up implementation of a command line debugger, as well as a debugger API that can be used from other applications. LLDB makes use of the Clang parser to provide high-fidelity expression parsing (particularly for C++) and uses the LLVM JIT for target support.

Like compiler_rt, libc++ is now dual licensed under the MIT and UIUC license, allowing it to be used more permissively.

Within the LLVM 3.1 time-frame there were the following highlights:

• The <atomic> header is now passing all tests, when compiling with clang and linking against the support code from compiler-rt.
• FreeBSD now includes libc++ as part of the base system.
• libc++ has been ported to Solaris and, in combination with libcxxrt and clang, is working with a large body of existing code.

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.

In the LLVM 3.1 time-frame, VMKit has had significant improvements on both runtime and startup performance.

Polly is an experimental optimizer for data locality and parallelism. It currently provides high-level loop optimizations and automatic parallelisation (using the OpenMP run time). Work in the area of automatic SIMD and accelerator code generation was started.

Within the LLVM 3.1 time-frame there were the following highlights:

• Polly became an official LLVM project
• Polly can be loaded directly into clang (enabled by '-O3 -mllvm -polly')
• An automatic scheduling optimizer (derived from Pluto) was integrated. It performs loop transformations to optimize for data-locality and parallelism. The transformations include, but are not limited to interchange, fusion, fission, skewing and tiling.

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.

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, JavaScript output formats, the Faust compiler can generate LLVM bitcode, and works with LLVM 2.7-3.1.

GHC is an open source compiler and programming suite for Haskell, a 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.

GHC 7.0 and onwards include an LLVM code generator, supporting LLVM 2.8 and later.

Julia is a high-level, high-performance dynamic language for technical computing. It provides a sophisticated compiler, distributed parallel execution, numerical accuracy, and an extensive mathematical function library. The compiler uses type inference to generate fast code without any type declarations, and uses LLVM's optimization passes and JIT compiler. The Julia Language is designed around multiple dispatch, giving programs a large degree of flexibility. It is ready for use on many kinds of problems.

LLVM D Compiler (LDC) is a compiler for the D programming Language. It is based on the DMD frontend and uses LLVM as backend.

Open Shading Language (OSL) is a small but rich language for programmable shading in advanced global illumination renderers and other applications, ideal for describing materials, lights, displacement, and pattern generation. It uses LLVM to JIT complex shader networks to x86 code at runtime.

OSL was developed by Sony Pictures Imageworks for use in its in-house renderer used for feature film animation and visual effects, and is distributed as open source software with the "New BSD" license.

In addition to producing an easily portable open source OpenCL implementation, another major goal of pocl is improving performance portability of OpenCL programs with compiler optimizations, reducing the need for target-dependent manual optimizations. An important part of pocl is a set of LLVM passes used to statically parallelize multiple work-items with the kernel compiler, even in the presence of work-group barriers. This enables static parallelization of the fine-grained static concurrency in the work groups in multiple ways (SIMD, VLIW, superscalar,...).

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. The interpreter uses LLVM as a backend to JIT-compile Pure programs to fast native code. 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 interface to C and other programming languages (including the ability to load LLVM bitcode modules, and inline C, C++, Fortran and Faust code in Pure programs if the corresponding LLVM-enabled compilers are installed).

Pure version 0.54 has been tested and is known to work with LLVM 3.1 (and continues to work with older LLVM releases >= 2.5).

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/Verilog and parallel program binaries. Processor customization points include the register files, function units, supported operations, and the interconnection network.

TCE uses 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.

LLVM 3.1 includes several major changes and big features:

• AddressSanitizer, a fast memory error detector.
• MachineInstr Bundles, Support to model instruction bundling / packing.
• ARM Integrated Assembler, A full featured assembler and direct-to-object support for ARM.
• Basic Block Placement Probability driven basic block placement.

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

• A new type representing 16 bit half floating point values has been added.
• IR now supports vectors of pointers, including vector GEPs.
• Module flags have been introduced. They convey information about the module as a whole to LLVM subsystems. This is currently used to encode Objective C ABI information.
• Loads can now have range metadata attached to them to describe the possible values being loaded.
• The llvm.ctlz and llvm.cttz intrinsics now have an additional argument which indicates whether the behavior of the intrinsic is undefined on a zero input. This can be used to generate more efficient code on platforms that only have instructions which don't return the type size when counting bits in 0.

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

• The loop unroll pass now is able to unroll loops with run-time trip counts. This feature is turned off by default, and is enabled with the -unroll-runtime flag.
• A new basic-block autovectorization pass is available. Pass -vectorize to run this pass along with some associated post-vectorization cleanup passes. For more information, see the EuroLLVM 2012 slides: Autovectorization with LLVM.
• Inline cost heuristics have been completely overhauled and now closely model constant propagation through call sites, disregard trivially dead code costs, and can model C++ STL iterator patterns.

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

• The integrated assembler can optionally emit debug information when assembling a .s file. It can be enabled by passing the -g option to llvm-mc.

We have changed the way that the Type Legalizer legalizes vectors. The type legalizer now attempts to promote integer elements. This enabled the implementation of vector-select. Additionally, we see a performance boost on workloads which use vectors of chars and shorts, since they are now promoted to 32-bit types, which are better supported by the SIMD instruction set. Floating point types are still widened as before.

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:

• TableGen can now synthesize register classes that are only needed to represent combinations of constraints from instructions and sub-registers. The synthetic register classes inherit most of their properties form their closest user-defined super-class.
• MachineRegisterInfo now allows the reserved registers to be frozen when register allocation starts. Target hooks should use the MRI->canReserveReg(FramePtr) method to avoid accidentally disabling frame pointer elimination during register allocation.
• A new kind of MachineOperand provides a compact representation of large clobber lists on call instructions. The register mask operand references a bit mask of preserved registers. Everything else is clobbered.
• The DWARF debug info writer gained support for emitting data for the name accelerator tables DWARF extension. It is used by LLDB to speed up name lookup.

We added new TableGen infrastructure to support bundling for Very Long Instruction Word (VLIW) architectures. TableGen can now automatically generate a deterministic finite automaton from a VLIW target's schedule description which can be queried to determine legal groupings of instructions in a bundle.

We have added a new target independent VLIW packetizer based on the DFA infrastructure to group machine instructions into bundles.

A probability based block placement and code layout algorithm was added to LLVM's code generator. This layout pass supports probabilities derived from static heuristics as well as source code annotations such as __builtin_expect.

New features and major changes in the X86 target include:

• Greatly improved support for AVX2.
• Lots of bug fixes and improvements for AVX1.
• Support for the FMA4 and XOP instruction set extensions.
• Call instructions use the new register mask operands for faster compile times and better support for different calling conventions. The old WINCALL instructions are no longer needed.
• DW2 Exception Handling is enabled on Cygwin and MinGW.
• Support for implicit TLS model used with MSVC runtime.

New features of the ARM target include:

• The constant island pass now supports basic block and constant pool entry alignments greater than 4 bytes.
• On Darwin, the ARM target now has a full-featured integrated assembler.

ARM Integrated Assembler

• MIPS32 little-endian direct object code emission is functional.
• MIPS64 little-endian code generation is largely functional for N64 ABI in assembly printing mode with the exception of handling of long double (f128) type.
• Support for new instructions has been added, which includes swap-bytes instructions (WSBH and DSBH), floating point multiply-add/subtract and negative multiply-add/subtract instructions, and floating point load/store instructions with reg+reg addressing (LWXC1, etc.)
• Various fixes to improve performance have been implemented.
• Post-RA scheduling is now enabled at -O3.
• Support for soft-float code generation has been added.
• clang driver's support for MIPS 64-bits targets.
• Support for MIPS floating point ABI option in clang driver.

An outstanding conditional inversion bug was fixed in this release.

NOTE: LLVM 3.1 marks the last release of the PTX back-end, in its current form. The back-end is currently being replaced by the NVPTX back-end, currently in SVN ToT.

• Support for Qualcomm's Hexagon VLIW processor has been added.

If you're already an LLVM user or developer with out-of-tree changes based on LLVM 3.1, this section lists some "gotchas" that you may run into upgrading from the previous release.

• LLVM's build system now requires a python 2 interpreter to be present at build time. A perl interpreter is no longer required.
• The C backend has been removed. It had numerous problems, to the point of not being able to compile any nontrivial program.
• The Alpha, Blackfin and SystemZ targets have been removed due to lack of maintenance.
• LLVM 3.1 removes support for reading LLVM 2.9 bitcode files. Going forward, we aim for all future versions of LLVM to read bitcode files and .ll files produced by LLVM 3.0 and later.
• The unwind instruction is now gone. With the introduction of the new exception handling system in LLVM 3.0, the unwind instruction became obsolete.
• LLVM 3.0 and earlier automatically added the returns_twice fo functions like setjmp based on the name. This functionality was removed in 3.1. This affects Clang users, if -ffreestanding is used.

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

• Target specific options have been moved from global variables to members on the new TargetOptions class, which is local to each TargetMachine. As a consequence, the associated flags will no longer be accepted by clang -mllvm. This includes:
  • llvm::PrintMachineCode
  • llvm::NoFramePointerElim
  • llvm::NoFramePointerElimNonLeaf
  • llvm::DisableFramePointerElim(const MachineFunction &)
  • llvm::LessPreciseFPMADOption
  • llvm::LessPrecideFPMAD()
  • llvm::NoExcessFPPrecision
  • llvm::UnsafeFPMath
  • llvm::NoInfsFPMath
  • llvm::NoNaNsFPMath
  • llvm::HonorSignDependentRoundingFPMathOption
  • llvm::HonorSignDependentRoundingFPMath()
  • llvm::UseSoftFloat
  • llvm::FloatABIType
  • llvm::NoZerosInBSS
  • llvm::JITExceptionHandling
  • llvm::JITEmitDebugInfo
  • llvm::JITEmitDebugInfoToDisk
  • llvm::GuaranteedTailCallOpt
  • llvm::StackAlignmentOverride
  • llvm::RealignStack
  • llvm::DisableJumpTables
  • llvm::EnableFastISel
  • llvm::getTrapFunctionName()
  • llvm::EnableSegmentedStacks
• The MDBuilder class has been added to simplify the creation of metadata.

In addition, some tools have changed in this release. Some of the changes are:

• llvm-stress is a command line tool for generating random .ll files to fuzz different LLVM components.
• The llvm-ld tool has been removed. The clang driver provides a more reliable solution for turning a set of bitcode files into a binary. To merge bitcode files llvm-link can be used instead.

Officially supported Python bindings have been added! Feature support is far from complete. The current bindings support interfaces to:

• Object File Interface
• Disassembler

Using the Object File Interface, it is possible to inspect binary object files. Think of it as a Python version of readelf or llvm-objdump.

Support for additional features is currently being developed by community contributors. If you are interested in shaping the direction of the Python bindings, please express your intent on IRC or the developers list.