3 <style> .red {color:red} </style>
14 Written by the `LLVM Team <http://llvm.org/>`_
16 :red:`These are in-progress notes for the upcoming LLVM 3.2 release. You may
17 prefer the` `LLVM 3.1 Release Notes <http://llvm.org/releases/3.1/docs
18 /ReleaseNotes.html>`_.
23 This document contains the release notes for the LLVM Compiler Infrastructure,
24 release 3.2. Here we describe the status of LLVM, including major improvements
25 from the previous release, improvements in various subprojects of LLVM, and
26 some of the current users of the code. All LLVM releases may be downloaded
27 from the `LLVM releases web site <http://llvm.org/releases/>`_.
29 For more information about LLVM, including information about the latest
30 release, please check out the `main LLVM web site <http://llvm.org/>`_. If you
31 have questions or comments, the `LLVM Developer's Mailing List
32 <http://lists.cs.uiuc.edu/mailman/listinfo/llvmdev>`_ is a good place to send
35 Note that if you are reading this file from a Subversion checkout or the main
36 LLVM web page, this document applies to the *next* release, not the current
37 one. To see the release notes for a specific release, please see the `releases
38 page <http://llvm.org/releases/>`_.
40 Sub-project Status Update
41 =========================
43 The LLVM 3.2 distribution currently consists of code from the core LLVM
44 repository, which roughly includes the LLVM optimizers, code generators and
45 supporting tools, and the Clang repository. In addition to this code, the LLVM
46 Project includes other sub-projects that are in development. Here we include
47 updates on these subprojects.
49 Clang: C/C++/Objective-C Frontend Toolkit
50 -----------------------------------------
52 `Clang <http://clang.llvm.org/>`_ is an LLVM front end for the C, C++, and
53 Objective-C languages. Clang aims to provide a better user experience through
54 expressive diagnostics, a high level of conformance to language standards, fast
55 compilation, and low memory use. Like LLVM, Clang provides a modular,
56 library-based architecture that makes it suitable for creating or integrating
57 with other development tools. Clang is considered a production-quality
58 compiler for C, Objective-C, C++ and Objective-C++ on x86 (32- and 64-bit), and
59 for Darwin/ARM targets.
61 In the LLVM 3.2 time-frame, the Clang team has made many improvements.
64 #. More powerful warnings, especially `-Wuninitialized`
65 #. Template type diffing in diagnostic messages
66 #. Higher quality and more efficient debug info generation
68 For more details about the changes to Clang since the 3.1 release, see the
69 `Clang release notes. <http://clang.llvm.org/docs/ReleaseNotes.html>`_
71 If Clang rejects your code but another compiler accepts it, please take a look
72 at the `language compatibility <http://clang.llvm.org/compatibility.html>`_
73 guide to make sure this is not intentional or a known issue.
75 DragonEgg: GCC front-ends, LLVM back-end
76 ----------------------------------------
78 `DragonEgg <http://dragonegg.llvm.org/>`_ is a `gcc plugin
79 <http://gcc.gnu.org/wiki/plugins>`_ that replaces GCC's optimizers and code
80 generators with LLVM's. It works with gcc-4.5 and gcc-4.6 (and partially with
81 gcc-4.7), can target the x86-32/x86-64 and ARM processor families, and has been
82 successfully used on the Darwin, FreeBSD, KFreeBSD, Linux and OpenBSD
83 platforms. It fully supports Ada, C, C++ and Fortran. It has partial support
84 for Go, Java, Obj-C and Obj-C++.
86 The 3.2 release has the following notable changes:
90 compiler-rt: Compiler Runtime Library
91 -------------------------------------
93 The new LLVM `compiler-rt project <http://compiler-rt.llvm.org/>`_ is a simple
94 library that provides an implementation of the low-level target-specific hooks
95 required by code generation and other runtime components. For example, when
96 compiling for a 32-bit target, converting a double to a 64-bit unsigned integer
97 is compiled into a runtime call to the ``__fixunsdfdi`` function. The
98 ``compiler-rt`` library provides highly optimized implementations of this and
99 other low-level routines (some are 3x faster than the equivalent libgcc
102 The 3.2 release has the following notable changes:
106 LLDB: Low Level Debugger
107 ------------------------
109 `LLDB <http://lldb.llvm.org>`_ is a ground-up implementation of a command line
110 debugger, as well as a debugger API that can be used from other applications.
111 LLDB makes use of the Clang parser to provide high-fidelity expression parsing
112 (particularly for C++) and uses the LLVM JIT for target support.
114 The 3.2 release has the following notable changes:
118 libc++: C++ Standard Library
119 ----------------------------
121 Like compiler_rt, libc++ is now :ref:`dual licensed
122 <copyright-license-patents>` under the MIT and UIUC license, allowing it to be
123 used more permissively.
125 Within the LLVM 3.2 time-frame there were the following highlights:
132 The `VMKit project <http://vmkit.llvm.org/>`_ is an implementation of a Java
133 Virtual Machine (Java VM or JVM) that uses LLVM for static and just-in-time
136 The 3.2 release has the following notable changes:
140 Polly: Polyhedral Optimizer
141 ---------------------------
143 `Polly <http://polly.llvm.org/>`_ is an *experimental* optimizer for data
144 locality and parallelism. It provides high-level loop optimizations and
145 automatic parallelisation.
147 Within the LLVM 3.2 time-frame there were the following highlights:
149 #. isl, the integer set library used by Polly, was relicensed to the MIT license
150 #. isl based code generation
151 #. MIT licensed replacement for CLooG (LGPLv2)
152 #. Fine grained option handling (separation of core and border computations,
153 control overhead vs. code size)
154 #. Support for FORTRAN and dragonegg
155 #. OpenMP code generation fixes
157 External Open Source Projects Using LLVM 3.2
158 ============================================
160 An exciting aspect of LLVM is that it is used as an enabling technology for a
161 lot of other language and tools projects. This section lists some of the
162 projects that have already been updated to work with LLVM 3.2.
167 `Crack <http://code.google.com/p/crack-language/>`_ aims to provide the ease of
168 development of a scripting language with the performance of a compiled
169 language. The language derives concepts from C++, Java and Python,
170 incorporating object-oriented programming, operator overloading and strong
176 `FAUST <http://faust.grame.fr/>`_ is a compiled language for real-time audio
177 signal processing. The name FAUST stands for Functional AUdio STream. Its
178 programming model combines two approaches: functional programming and block
179 diagram composition. In addition with the C, C++, Java, JavaScript output
180 formats, the Faust compiler can generate LLVM bitcode, and works with LLVM
183 Glasgow Haskell Compiler (GHC)
184 ------------------------------
186 `GHC <http://www.haskell.org/ghc/>`_ is an open source compiler and programming
187 suite for Haskell, a lazy functional programming language. It includes an
188 optimizing static compiler generating good code for a variety of platforms,
189 together with an interactive system for convenient, quick development.
191 GHC 7.0 and onwards include an LLVM code generator, supporting LLVM 2.8 and
197 `Julia <https://github.com/JuliaLang/julia>`_ is a high-level, high-performance
198 dynamic language for technical computing. It provides a sophisticated
199 compiler, distributed parallel execution, numerical accuracy, and an extensive
200 mathematical function library. The compiler uses type inference to generate
201 fast code without any type declarations, and uses LLVM's optimization passes
202 and JIT compiler. The `Julia Language <http://julialang.org/>`_ is designed
203 around multiple dispatch, giving programs a large degree of flexibility. It is
204 ready for use on many kinds of problems.
209 `LLVM D Compiler <https://github.com/ldc-developers/ldc>`_ (LDC) is a compiler
210 for the D programming Language. It is based on the DMD frontend and uses LLVM
213 Open Shading Language
214 ---------------------
216 `Open Shading Language (OSL)
217 <https://github.com/imageworks/OpenShadingLanguage/>`_ is a small but rich
218 language for programmable shading in advanced global illumination renderers and
219 other applications, ideal for describing materials, lights, displacement, and
220 pattern generation. It uses LLVM to JIT complex shader networks to x86 code at
223 OSL was developed by Sony Pictures Imageworks for use in its in-house renderer
224 used for feature film animation and visual effects, and is distributed as open
225 source software with the "New BSD" license.
227 Portable OpenCL (pocl)
228 ----------------------
230 In addition to producing an easily portable open source OpenCL implementation,
231 another major goal of `pocl <http://pocl.sourceforge.net/>`_ is improving
232 performance portability of OpenCL programs with compiler optimizations,
233 reducing the need for target-dependent manual optimizations. An important part
234 of pocl is a set of LLVM passes used to statically parallelize multiple
235 work-items with the kernel compiler, even in the presence of work-group
236 barriers. This enables static parallelization of the fine-grained static
237 concurrency in the work groups in multiple ways (SIMD, VLIW, superscalar, ...).
242 `Pure <http://pure-lang.googlecode.com/>`_ is an algebraic/functional
243 programming language based on term rewriting. Programs are collections of
244 equations which are used to evaluate expressions in a symbolic fashion. The
245 interpreter uses LLVM as a backend to JIT-compile Pure programs to fast native
246 code. Pure offers dynamic typing, eager and lazy evaluation, lexical closures,
247 a hygienic macro system (also based on term rewriting), built-in list and
248 matrix support (including list and matrix comprehensions) and an easy-to-use
249 interface to C and other programming languages (including the ability to load
250 LLVM bitcode modules, and inline C, C++, Fortran and Faust code in Pure
251 programs if the corresponding LLVM-enabled compilers are installed).
253 Pure version 0.54 has been tested and is known to work with LLVM 3.1 (and
254 continues to work with older LLVM releases >= 2.5).
256 TTA-based Co-design Environment (TCE)
257 -------------------------------------
259 `TCE <http://tce.cs.tut.fi/>`_ is a toolset for designing application-specific
260 processors (ASP) based on the Transport triggered architecture (TTA). The
261 toolset provides a complete co-design flow from C/C++ programs down to
262 synthesizable VHDL/Verilog and parallel program binaries. Processor
263 customization points include the register files, function units, supported
264 operations, and the interconnection network.
266 TCE uses Clang and LLVM for C/C++ language support, target independent
267 optimizations and also for parts of code generation. It generates new
268 LLVM-based code generators "on the fly" for the designed TTA processors and
269 loads them in to the compiler backend as runtime libraries to avoid per-target
270 recompilation of larger parts of the compiler chain.
272 Installation Instructions
273 =========================
275 See :doc:`GettingStarted`.
277 What's New in LLVM 3.2?
278 =======================
280 This release includes a huge number of bug fixes, performance tweaks and minor
281 improvements. Some of the major improvements and new features are listed in
289 Features that need text if they're finished for 3.2:
293 loop dependence analysis
294 CorrelatedValuePropagation
295 Integrated assembler on by default for arm/thumb?
298 Analysis/RegionInfo.h + Dom Frontiers
299 SparseBitVector: used in LiveVar.
300 llvm/lib/Archive - replace with lib object?
303 LLVM 3.2 includes several major changes and big features:
305 #. New NVPTX back-end (replacing existing PTX back-end) based on NVIDIA sources
308 LLVM IR and Core Improvements
309 -----------------------------
311 LLVM IR has several new features for better support of new targets and that
312 expose new optimization opportunities:
314 #. Thread local variables may have a specified TLS model. See the :ref:`Language
315 Reference Manual <globalvars>`.
318 Optimizer Improvements
319 ----------------------
321 In addition to many minor performance tweaks and bug fixes, this release
322 includes a few major enhancements and additions to the optimizers:
324 Loop Vectorizer - We've added a loop vectorizer and we are now able to
325 vectorize small loops. The loop vectorizer is disabled by default and can be
326 enabled using the ``-mllvm -vectorize-loops`` flag. The SIMD vector width can
327 be specified using the flag ``-mllvm -force-vector-width=4``. The default
328 value is ``0`` which means auto-select.
330 We can now vectorize this function:
334 unsigned sum_arrays(int *A, int *B, int start, int end) {
336 for (int i = start; i < end; ++i)
337 sum += A[i] + B[i] + i;
341 We vectorize under the following loops:
343 #. The inner most loops must have a single basic block.
344 #. The number of iterations are known before the loop starts to execute.
345 #. The loop counter needs to be incremented by one.
346 #. The loop trip count **can** be a variable.
347 #. Loops do **not** need to start at zero.
348 #. The induction variable can be used inside the loop.
349 #. Loop reductions are supported.
350 #. Arrays with affine access pattern do **not** need to be marked as
351 '``noalias``' and are checked at runtime.
354 SROA - We've re-written SROA to be significantly more powerful and generate
355 code which is much more friendly to the rest of the optimization pipeline.
356 Previously this pass had scaling problems that required it to only operate on
357 relatively small aggregates, and at times it would mistakenly replace a large
358 aggregate with a single very large integer in order to make it a scalar SSA
359 value. The result was a large number of i1024 and i2048 values representing any
360 small stack buffer. These in turn slowed down many subsequent optimization
363 The new SROA pass uses a different algorithm that allows it to only promote to
364 scalars the pieces of the aggregate actively in use. Because of this it doesn't
365 require any thresholds. It also always deduces the scalar values from the uses
366 of the aggregate rather than the specific LLVM type of the aggregate. These
367 features combine to both optimize more code with the pass but to improve the
368 compile time of many functions dramatically.
370 #. Branch weight metadata is preseved through more of the optimizer.
373 MC Level Improvements
374 ---------------------
376 The LLVM Machine Code (aka MC) subsystem was created to solve a number of
377 problems in the realm of assembly, disassembly, object file format handling,
378 and a number of other related areas that CPU instruction-set level tools work
379 in. For more information, please see the `Intro to the LLVM MC Project Blog
380 Post <http://blog.llvm.org/2010/04/intro-to-llvm-mc-project.html>`_.
386 Target Independent Code Generator Improvements
387 ----------------------------------------------
389 Stack Coloring - We have implemented a new optimization pass to merge stack
390 objects which are used in disjoin areas of the code. This optimization reduces
391 the required stack space significantly, in cases where it is clear to the
392 optimizer that the stack slot is not shared. We use the lifetime markers to
393 tell the codegen that a certain alloca is used within a region.
395 We now merge consecutive loads and stores.
397 We have put a significant amount of work into the code generator
398 infrastructure, which allows us to implement more aggressive algorithms and
403 We added new TableGen infrastructure to support bundling for Very Long
404 Instruction Word (VLIW) architectures. TableGen can now automatically generate
405 a deterministic finite automaton from a VLIW target's schedule description
406 which can be queried to determine legal groupings of instructions in a bundle.
408 We have added a new target independent VLIW packetizer based on the DFA
409 infrastructure to group machine instructions into bundles.
411 Basic Block Placement
412 ^^^^^^^^^^^^^^^^^^^^^
414 A probability based block placement and code layout algorithm was added to
415 LLVM's code generator. This layout pass supports probabilities derived from
416 static heuristics as well as source code annotations such as
417 ``__builtin_expect``.
419 X86-32 and X86-64 Target Improvements
420 -------------------------------------
422 New features and major changes in the X86 target include:
428 ARM Target Improvements
429 -----------------------
431 New features of the ARM target include:
435 .. _armintegratedassembler:
437 ARM Integrated Assembler
438 ^^^^^^^^^^^^^^^^^^^^^^^^
440 The ARM target now includes a full featured macro assembler, including
441 direct-to-object module support for clang. The assembler is currently enabled
442 by default for Darwin only pending testing and any additional necessary
443 platform specific support for Linux.
445 Full support is included for Thumb1, Thumb2 and ARM modes, along with subtarget
446 and CPU specific extensions for VFP2, VFP3 and NEON.
448 The assembler is Unified Syntax only (see ARM Architecural Reference Manual for
449 details). While there is some, and growing, support for pre-unfied (divided)
450 syntax, there are still significant gaps in that support.
452 MIPS Target Improvements
453 ------------------------
455 New features and major changes in the MIPS target include:
459 PowerPC Target Improvements
460 ---------------------------
462 Many fixes and changes across LLVM (and Clang) for better compliance with the
463 64-bit PowerPC ELF Application Binary Interface, interoperability with GCC, and
464 overall 64-bit PowerPC support. Some highlights include:
466 #. MCJIT support added.
467 #. PPC64 relocation support and (small code model) TOC handling added.
468 #. Parameter passing and return value fixes (alignment issues, padding, varargs
469 support, proper register usage, odd-sized structure support, float support,
470 extension of return values for i32 return values).
471 #. Fixes in spill and reload code for vector registers.
472 #. C++ exception handling enabled.
473 #. Changes to remediate double-rounding compatibility issues with respect to
475 #. Refactoring to disentangle ``ppc64-elf-linux`` ABI from Darwin ppc64 ABI
477 #. Assorted new test cases and test case fixes (endian and word size issues).
478 #. Fixes for big-endian codegen bugs, instruction encodings, and instruction
480 #. Implemented ``-integrated-as`` support.
481 #. Additional support for Altivec compare operations.
482 #. IBM long double support.
484 There have also been code generation improvements for both 32- and 64-bit code.
485 Instruction scheduling support for the Freescale e500mc and e5500 cores has
488 PTX/NVPTX Target Improvements
489 -----------------------------
491 The PTX back-end has been replaced by the NVPTX back-end, which is based on the
492 LLVM back-end used by NVIDIA in their CUDA (nvcc) and OpenCL compiler. Some
495 #. Compatibility with PTX 3.1 and SM 3.5.
496 #. Support for NVVM intrinsics as defined in the NVIDIA Compiler SDK.
497 #. Full compatibility with old PTX back-end, with much greater coverage of LLVM
500 Please submit any back-end bugs to the LLVM Bugzilla site.
502 Other Target Specific Improvements
503 ----------------------------------
507 Major Changes and Removed Features
508 ----------------------------------
510 If you're already an LLVM user or developer with out-of-tree changes based on
511 LLVM 3.2, this section lists some "gotchas" that you may run into upgrading
512 from the previous release.
514 #. The CellSPU port has been removed. It can still be found in older versions.
520 In addition, many APIs have changed in this release. Some of the major LLVM
523 We've added a new interface for allowing IR-level passes to access
524 target-specific information. A new IR-level pass, called
525 ``TargetTransformInfo`` provides a number of low-level interfaces. LSR and
526 LowerInvoke already use the new interface.
528 The ``TargetData`` structure has been renamed to ``DataLayout`` and moved to
529 ``VMCore`` to remove a dependency on ``Target``.
536 In addition, some tools have changed in this release. Some of the changes are:
543 Officially supported Python bindings have been added! Feature support is far
544 from complete. The current bindings support interfaces to:
551 LLVM is generally a production quality compiler, and is used by a broad range
552 of applications and shipping in many products. That said, not every subsystem
553 is as mature as the aggregate, particularly the more obscure1 targets. If you
554 run into a problem, please check the `LLVM bug database
555 <http://llvm.org/bugs/>`_ and submit a bug if there isn't already one or ask on
556 the `LLVMdev list <http://lists.cs.uiuc.edu/mailman/listinfo/llvmdev>`_.
558 Known problem areas include:
560 #. The CellSPU, MSP430, and XCore backends are experimental.
562 #. The integrated assembler, disassembler, and JIT is not supported by several
563 targets. If an integrated assembler is not supported, then a system
564 assembler is required. For more details, see the
565 :ref:`target-feature-matrix`.
567 Additional Information
568 ======================
570 A wide variety of additional information is available on the `LLVM web page
571 <http://llvm.org/>`_, in particular in the `documentation
572 <http://llvm.org/docs/>`_ section. The web page also contains versions of the
573 API documentation which is up-to-date with the Subversion version of the source
574 code. You can access versions of these documents specific to this release by
575 going into the ``llvm/docs/`` directory in the LLVM tree.
577 If you have any questions or comments about LLVM, please feel free to contact
578 us via the `mailing lists <http://llvm.org/docs/#maillist>`_.