LLVM Weekly - #6, Feb 10th 2014
Welcome to the sixth 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 https://llvmweekly.org and pass it on to anyone else you think may be interested. Please send any tips or feedback to [email protected], or @llvmweekly or @asbradbury on Twitter. I've been keeping the @llvmweekly Twitter account updated throughout the week, so follow that if you want more frequent news updates.
News and articles from around the web
Alexi Starovoitov has published an LLVM backend targeting an extended version of the Linux Kernel's BPF. An example of the sort of program that might be compiled and run via BPF can be found here.
There is now under a week to go to submit proposals for presentations, tutorials, posters, etc for the upcoming EuroLLVM 2014. Get writing!
LWN's coverage of the recent discussion about LLVM and its licensing on the GCC mailing list is now available to non-subscribers.
Renato Golin posted to the GCC mailing list suggesting there be more collaboration where possible on issues such as standardisation of command line interfaces, language extensions, or just general technical discussion.I know a mailing list GCC developers who want to keep abreast of LLVM/Clang developments should subscribe to...
Phoronix has published a benchmark comparing GCC 4.8.2, a GCC 4.9 snapshot and Clang 3.4 on an Intel Core i5-4670 system.
On the mailing lists
-
Sony Computer Entertainment America are looking for an intern for their compiler engineering team.
-
Venkatraman Govindaraju reports that Clang can now self host on Linux/sparc64 and FreeBSD/sparc64.
-
Paul Vario asks for more information on the algorithm used in RegionInfo, noting that the algorithm implemented seems substantially different to the one referenced in the paper in the header's comment. The original author of that code, Tobias Grosser follows up with an explanation for differences from that paper.
-
Gordon Keiser helpfully reminds us that clang will happily take LLVM bitcode as input.
-
Stephen Kelly, an upstream CMake developer dropped by to offer advice on LLVM's current CMake usage and learn what problems with CMake the project is facing. A number of responses are interested in the ability to use CMake to build multiple LLVM subprojects at once, or to build clang against an installed LLVM tree.
-
Last week's RFC questioning the use of BlockFrequency has now had a number of followups. I won't summarise here for fear of mischaracterising people, but do take a look if the issue interests you.
-
Tom Stellard is looking for bugfix patches that should be merged in to a future 3.4.1 release.
-
Renato Goling started a discussion about unwind and exception handling behaviour in LLVM/Clang, listing four problems he sees with the code as-is.
-
Vassil Vassilev asks whether anyone has implemented a copy and paste detector for Clang's static analyzer. Nick Lewycky shares his implementation of something in that direction.
-
Ben Langmuir posted an RFC on a virtual filesystem for clang. This aims to "allow a build system to provide a file/directory layout to clang without having to construct it 'for real' on disk".
LLVM commits
-
The x86 backend was slightly simplified by moving some matching for x86 bit manipulation instructions from X86ISelLowering.cpp to X86InstrInfo.td. I mention this commit mainly as it's a useful reference for those of you working on LLVM backend code. r200824.
-
The register allocator gained a new 'last chance recoloring mechanism'. Sadly the commit message doesn't include any data of how this improves register allocation for a given codebase. r200883.
-
The old SmallPtrSetImpl was renamed to SmallPtrSetImplBase, and a new SmallPtrSetImpl was introduced. This new SmallPtrSetImpl doesn't require a specific set size to be specified in its templated parameter. r200688.
-
A whole bunch of code was added to CodeGenPrepare which attempts to move sign extensions away from loads in order to increase the chance that the address computation can be folded in to the load on architectures like x86 with complex addressing modes. r200947.
-
strchr(p, 0)is now simplified top + strlen(p). r200736. -
Information on handling minor ('dot') releases was added to the HowToReleaseLLVM documentation. r200772.
-
The MIPS assembler learned to understand
%hi(sym1 - sym2)and%hi(sym1 - sym2)expressions. r200783. -
Mips gained a NaCl target. r200855.
-
LLVM now assumes the assembler supports the
.locdirective for specifying debug line numbers. r200862. -
The inliner was modified to consider the cold attribute on a function when deciding whether to inline. r200886. A later commit set the inlinecold-threshold to the same as the inline-threshold so that current inlining behaviour is maintained for now. r200898.
-
Initial implementation for a lazy call graph analysis pass (for use with the upcoming new pass manager) was committed. r200903.
-
The allowsUnalignedMemoryAccess function in TargetLowering now takes an address space argument. This was added for architectures like the R600 where different address spaces have different alignment requirements. r200887.
Clang commits
-
More support was MS ABI-compatible mangling was added. r200857.
-
The behaviour suggested by the C++ Defect Report 329 was implemented. r200673.
-
The ARM target gained support for crypto intrinsics defined in
arm_neon.h. r200708. -
The forRangeStmt AST matcher gained a handy hasLoopVariable sub-matcher. r200850.
-
The -verify-pch CC1 option is now supported. r200884.
-
The -fhiding-week-vtables CC1 option has been removed. r201011.
-
LLVM's new diagnostic system is now wired into clang's diagnostic system. r200931.