LLVM 3.5 Release Notes


This document contains the release notes for the LLVM Compiler Infrastructure, release 3.5. Here we describe the status of LLVM, including major improvements from the previous release, improvements in various subprojects of LLVM, and some of the current users of the code. 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.

Non-comprehensive list of changes in this release

  • All backends have been changed to use the MC asm printer and support for the non MC one has been removed.
  • Clang can now successfully self-host itself on Linux/Sparc64 and on FreeBSD/Sparc64.
  • LLVM now assumes the assembler supports .loc for generating debug line numbers. The old support for printing the debug line info directly was only used by llc and has been removed.
  • All inline assembly is parsed by the integrated assembler when it is enabled. Previously this was only the case for object-file output. It is now the case for assembly output as well. The integrated assembler can be disabled with the -no-integrated-as option.
  • llvm-ar now handles IR files like regular object files. In particular, a regular symbol table is created for symbols defined in IR files, including those in file scope inline assembly.
  • LLVM now always uses cfi directives for producing most stack unwinding information.
  • The prefix for loop vectorizer hint metadata has been changed from llvm.vectorizer to llvm.loop.vectorize. In addition, llvm.vectorizer.unroll metadata has been renamed llvm.loop.interleave.count.
  • Some backends previously implemented Atomic NAND(x,y) as x & ~y. Now all backends implement it as ~(x & y), matching the semantics of GCC 4.4 and later.
  • ... next change ...

Changes to the ARM Backend

Since release 3.3, a lot of new features have been included in the ARM back-end but weren’t production ready (ie. well tested) on release 3.4. Just after the 3.4 release, we started heavily testing two major parts of the back-end: the integrated assembler (IAS) and the ARM exception handling (EHABI), and now they are enabled by default on LLVM/Clang.

The IAS received a lot of GNU extensions and directives, as well as some specific pre-UAL instructions. Not all remaining directives will be implemented, as we made judgement calls on the need versus the complexity, and have chosen simplicity and future compatibility where hard decisions had to be made. The major difference is, as stated above, the IAS validates all inline ASM, not just for object emission, and that cause trouble with some uses of inline ASM as pre-processor magic.

So, while the IAS is good enough to compile large projects (including most of the Linux kernel), there are a few things that we can’t (and probably won’t) do. For those cases, please use -fno-integrated-as in Clang.

Exception handling is another big change. After extensive testing and changes to cooperate with Dwarf unwinding, EHABI is enabled by default. The options -arm-enable-ehabi and -arm-enable-ehabi-descriptors, which were used to enable EHABI in the previous releases, are removed now.

This means all ARM code will emit EH unwind tables, or CFI unwinding (for debug/profiling), or both. To avoid run-time inconsistencies, C code will also emit EH tables (in case they interoperate with C++ code), as is the case for other architectures (ex. x86_64).

Changes to the MIPS Target

There has been a large amount of improvements to the MIPS target which can be broken down into subtarget, ABI, and Integrated Assembler changes.


Added support for Release 6 of the MIPS32 and MIPS64 architecture (MIPS32r6 and MIPS64r6). Release 6 makes a number of significant changes to the MIPS32 and MIPS64 architectures. For example, FPU registers are always 64-bits wide, FPU NaN values conform to IEEE 754 (2008), and the unaligned memory instructions (such as lwl and lwr) have been replaced with a requirement for ordinary memory operations to support unaligned operations. Full details of MIPS32 and MIPS64 Release 6 can be found on the MIPS64 Architecture page at Imagination Technologies.

This release also adds experimental support for MIPS-IV, cnMIPS, and Cavium Octeon CPU’s.

Support for the MIPS SIMD Architecture (MSA) has been improved to support MSA on MIPS64.

Support for IEEE 754 (2008) NaN values has been added.

ABI and ABI extensions

There has also been considerable ABI work since the 3.4 release. This release adds support for the N32 ABI, the O32-FPXX ABI Extension, the O32-FP64 ABI Extension, and the O32-FP64A ABI Extension.

The N32 ABI is an existing ABI that has now been implemented in LLVM. It is a 64-bit ABI that is similar to N64 but retains 32-bit pointers. N64 remains the default 64-bit ABI in LLVM. This differs from GCC where N32 is the default 64-bit ABI.

The O32-FPXX ABI Extension is 100% compatible with the O32-ABI and the O32-FP64 ABI Extension and may be linked with either but may not be linked with both of these simultaneously. It extends the O32 ABI to allow the same code to execute without modification on processors with 32-bit FPU registers as well as 64-bit FPU registers. The O32-FPXX ABI Extension is enabled by default for the O32 ABI on mips*-img-linux-gnu and mips*-mti-linux-gnu triples and is selected with -mfpxx. It is expected that future releases of LLVM will enable the FPXX Extension for O32 on all triples.

The O32-FP64 ABI Extension is an extension to the O32 ABI to fully exploit FPU’s with 64-bit registers and is enabled with -mfp64. This replaces an undocumented and unsupported O32 extension which was previously enabled with -mfp64. It is 100% compatible with the O32-FPXX ABI Extension.

The O32-FP64A ABI Extension is a restricted form of the O32-FP64 ABI Extension which allows interlinking with unmodified binaries that use the base O32 ABI.

Integrated Assembler

The MIPS Integrated Assembler has undergone a substantial overhaul including a rewrite of the assembly parser. It’s not ready for general use in this release but adventurous users may wish to enable it using -fintegrated-as.

In this release, the integrated assembler supports the majority of MIPS-I, MIPS-II, MIPS-III, MIPS-IV, MIPS-V, MIPS32, MIPS32r2, MIPS32r6, MIPS64, MIPS64r2, and MIPS64r6 as well as some of the Application Specific Extensions such as MSA. It also supports several of the MIPS specific assembler directives such as .set, .module, .cpload, etc.

Changes to the AArch64 Target

The AArch64 target in LLVM 3.5 is based on substantially different code to the one in LLVM 3.4, having been created as the result of merging code released by Apple for targetting iOS with the previously existing backend.

We hope the result is a general improvement in the project. Particularly notable changes are:

  • We should produce faster code, having combined optimisations and ideas from both sources in the final backend.
  • We have a FastISel for AArch64, which should compile time for debug builds (at -O0).
  • We can now target iOS platforms (using the triple arm64-apple-ios7.0).


During the 3.5 release cycle, Apple released the source used to generate 64-bit ARM programs on iOS platforms. This took the form of a separate backend that had been developed in parallel to, and largely isolation from, the existing code.

We decided that maintaining the two backends indefinitely was not an option, since their features almost entirely overlapped. However, the implementation details in both were different enough that any merge had to firmly start with one backend as the core and cherry-pick the best features and optimisations from the other.

After discussion, we decided to start with the Apple backend (called ARM64 at the time) since it was older, more thoroughly tested in production use, and had fewer idiosyncracies in the implementation details.

Many people from across the community worked throughout April and May to ensure that this merge destination had all the features we wanted, from both sources. In many cases we could simply copy code across; others needed heavy modification for the new host; in the most worthwhile, we looked at both implementations and combined the best features of each in an entirely new way.

We had also decided that the name of the combined backend should be AArch64, following ARM’s official documentation. So, at the end of May the old AArch64 directory was removed, and ARM64 renamed into its place.

Changes to the PowerPC Target

The PowerPC 64-bit Little Endian subtarget (powerpc64le-unknown-linux-gnu) is now fully supported. This includes support for the Altivec instruction set.

The Power Architecture 64-Bit ELFv2 ABI Specification is now supported, and is the default ABI for Little Endian. The ELFv1 ABI remains the default ABI for Big Endian. Currently, it is not possible to override these defaults. That capability will be available (albeit not recommended) in a future release.

Links to the ELFv2 ABI specification and to the Power ISA Version 2.07 specification may be found here (free registration required). Efforts are underway to move this to a location that doesn’t require registration, but the planned site isn’t ready yet.

Experimental support for the VSX instruction set introduced with ISA 2.06 is now available using the -mvsx switch. Work remains on this, so it is not recommended for production use. VSX is disabled for Little Endian regardless of this switch setting.

Load/store cost estimates have been improved.

Constant hoisting has been enabled.

Global named register support has been enabled.

Initial support for PIC code has been added for the 32-bit ELF subtarget. Further support will be available in a future release.

Changes to CMake build system

  • Building and installing LLVM, Clang and lld sphinx documentation can now be done in CMake builds. If LLVM_ENABLE_SPHINX is enabled the “docs-<project>-html” and “docs-<project>-man” targets (e.g. docs-llvm-html) become available which can be invoked directly (e.g. make docs-llvm-html) to build only the relevant sphinx documentation. If LLVM_BUILD_DOCS is enabled then the sphinx documentation will also be built as part of the normal build. Enabling this variable also means that if the install target is invoked then the built documentation will be installed. See LLVM-specific variables.
  • Both the Autoconf/Makefile and CMake build systems now generate LLVMConfig.cmake (and other files) to export installed libraries. This means that projects using CMake to build against LLVM libraries can now build against an installed LLVM built by the Autoconf/Makefile system. See Embedding LLVM in your project for details.
  • Use of llvm_map_components_to_libraries() by external projects is deprecated and the new llvm_map_components_to_libnames() should be used instead.

External Open Source Projects Using LLVM 3.5

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 3.5.

LDC - the LLVM-based D compiler

D is a language with C-like syntax and static typing. It pragmatically combines efficiency, control, and modeling power, with safety and programmer productivity. D supports powerful concepts like Compile-Time Function Execution (CTFE) and Template Meta-Programming, provides an innovative approach to concurrency and offers many classical paradigms.

LDC uses the frontend from the reference compiler combined with LLVM as backend to produce efficient native code. LDC targets x86/x86_64 systems like Linux, OS X, FreeBSD and Windows and also Linux/PPC64. Ports to other architectures like ARM, AArch64 and MIPS64 are underway.

Portable Computing Language (pocl)

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.

TTA-based Co-design Environment (TCE)

TCE is a toolset for designing new exposed datapath processors 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++/OpenCL 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 processors and loads them in to the compiler backend as runtime libraries to avoid per-target recompilation of larger parts of the compiler chain.


ISPC is a C-based language based on the SPMD (single program, multiple data) programming model that generates efficient SIMD code for modern processors without the need for complex analysis and autovectorization. The language exploits the concept of “varying” data types, which ensure vector-friendly data layout, explicit vectorization and compact representation of the program. The project uses the LLVM infrastructure for optimization and code generation.


Likely is an embeddable just-in-time Lisp for image recognition and heterogenous architectures. Algorithms are just-in-time compiled using LLVM’s MCJIT infrastructure to execute on single or multi-threaded CPUs and potentially OpenCL SPIR or CUDA enabled GPUs. Likely exploits the observation that while image processing and statistical learning kernels must be written generically to handle any matrix datatype, at runtime they tend to be executed repeatedly on the same type. Likely also seeks to explore new optimizations for statistical learning algorithms by moving them from an offline model generation step to a compile-time simplification of a function (the learning algorithm) with constant arguments (the training set).

Additional Information

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/docs/ directory in the LLVM tree.

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