LLVM Weekly - #23, Jun 9th 2014
Welcome to the twenty-third issue of LLVM Weekly, a weekly newsletter (published every Monday) covering developments in LLVM, Clang, and related projects. LLVM Weekly is brought to you by Alex Bradbury. Subscribe to future issues at http://llvmweekly.org and pass it on to anyone else you think may be interested. Please send any tips or feedback to email@example.com, or @llvmweekly or @asbradbury on Twitter.
The canonical home for this issue can be found here at llvmweekly.org.
News and articles from around the web
Philip Reames has announced that code for late safepoint placement is now available.
This is a set of patches to LLVM from Azul Systems that aim to support precise relocating garbage collection in LLVM. Phlip has a long list of questions where he is seeking feedback from the community on how to move forwards with these patches. There's not been much response so far, hopefully it will come soon as I know there are many communities who are very interested in seeing better GC support in LLVM (e.g. Rust, Ocaml).
The biggest LLVM news this week is of course the announcement of the Swift programming language from Apple. Its development was led by Chris Lattner, original author of LLVM. He has some more info about Swift on his website. There is no source release as of yet, and no indication from Apple as to whether it will remain proprietary. Either way, it's an interesting development. Chris Lattner is now on Twitter and has been passing out tidbits about the Swift implementation.
LunarG have announced the Glassy Mesa project. This project, funded by Valve, will explore increasing game performance in Mesa through improvements in the shader compiler. The current parser and optimisation layer are replaced with glslang and the LLVM-based LunarGlass. More technical details are available in the slide deck.
Sébastien Métrot has released xspray, a frontend for lldb on OS X. One of its interesting features is the inbuilt support for plotting your data.
With all the LLVM news recently, it seems search traffic for 'llvm' has skyrocketed.
On the mailing lists
Zachary Tuner has started a discussion on multi-threading and mutexes in LLVM, following from his patches (currently in review) that tries to replace LLVM's own mutex implementation with
std::recursive_mutex. The key questions are whether multi-threading should be a compile-time or tunetime parameter, what should happen if you attempt to acquire a mutex in an app with threading disabled, and whether debugging code for finding deadlocks should be included.
Sean Silva queries whether stack maps should be marked as stable in LLVM 3.5 given that the feature now has a number of users, including WebKit. Andrew Trick suggests another round of RFC before this, and intends to submit another proposal in the coming week.
Gerolf Hoflehner is asking for feedback on his proposed instruction combiner pass. This would run at the machine IR level and aims to overcome some of the deficiencies observed when combining instructions at the ISEL level.
A question has arisen over whether LLVM 3.3/3.4 IR is backward compatible with 3.2. This isn't a form of compatibility that LLVM offers, though Stephen Hines notes that Android maintains an LLVM 3.2 bitcode writer.
Quite some time ago, Dennis Luehring drew attention to a case where the GCC optimiser seems to do better than Clang. Benjamin Kramer and Richard Smith both submitted patches which now results in Clang acing this testcase.
The jumptable attribute has been introduced. If you mark a function with this attribute, references to it can be rewritten with a reference to the appropriate jump-instruction-table function pointer. r210280.
GlobalAlias can now point to an arbitrary ConstantExpression. See the commit message for a discussion of the consequences of this. r210062.
The subword level parallelism (SLP) vectorizer has been extended to support vectorization of getelementptr expressions. r210342.
The LLVM programmer's manual has been improved with an example of using IRBuilder. r210354.
Semantic analysis to make sure a loop is in OpenMP canonical form has been committed. r210095.
__builtin_operator_deletehave been added. Some optimisations are allowed on these which would not be on
::operator newand are intended for the implementation of things like std::allocator. r210137.
New pragmas have been introduced to give optimisation hints for vectorization and interleaving. You can now use
#pragma clang loop vectorize(enable)as well as
Support for the MSVC++ ABI continues with the addition of
dynamic_castfor MS. r210377.
Support for global named registers has been expanded slightly to allow pointer types to be held in these variables. r210274.
-Wframe-larger-than=bytesdiagnostic is now supported. r210293.
Other project commits
A benchmarking-only mode has been added to the testsuite r210251.
An initial set of Makefiles has been committed to lld. r210177.
lldb gained support for inspecting enum members. r210046.
Polly can now be built without any GPLed software. r210176.