LLVM Weekly - #152, Nov 28th 2016
Welcome to the one hundred and fifty-second 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'm back in the Bay Area again this week for a very brief trip, where I'll be at the RISC-V workshop. I'll be talking about RISC-V LLVM and why we (the lowRISC project) see this as an important part of our efforts to produce a completely open source SoC. Hopefully I'll see some of you there.
News and articles from around the web
RV, the Region Vectorizer framework for LLVM is now publicly available. As described by one of the author, it provides a unified interface to vectorize code regions. The authors are hoping to upstream parts of RV.
Botond Ballo has written up a trip report from the recent C++ standards meeting in Issaquah.
The Transport Trigger Architecture (TTA) Code-design Environment v1.14 has been released. This adds support for LLVM 3.9, as well as variable-length local arrays and alloca amongst other changes.
mulle-objc, a new Objective-C compiler and runtime has been announced. The compiler is a fork of Clang. The discussion of the project on Hacker News has some more details.
On the mailing lists
-
Discussion has continued on the proposed patches for ARM's scalable vector extension (SVE). As Renato Golin points out, the fact the SVE spec hasn't yet been published makes reviewing the patch somewhat difficult. Some concerns have been raised that the choice to modify IR rather than rely on intrinsics may not have been fully justified.
-
Malcolm Parsons has proposed integrating support for constexpr StringRefs in order to reduce the number static initializers. A barrier to implementing this is that lack of a fully portable constexpr strlen. The thread discusses a number of workarounds.
-
Gadi Haber has posted an RFC on reducing X86 code size by using the VEX AVX512 encoding where possible. As pointed out by Hal Finkel, tablegen's InstrMapping feature makes this easier to implement.
-
Sean Silva has done some more analysis of the page fault costs of LLD on Linux.
LLVM commits
-
Integer min/max are now canonicalized in InstCombine. r287585.
-
The interface to TargetMachine::getSymbol has been simplified. r287645.
-
TargetLibraryInfo gained new functions to determine if integer parameters should be zeroext or signext. r287533.
-
New IntrInaccessibleMemOnly and IntrInaccessibleMemOrArgMemOnly properties were added for intrinsics. This indicates the intrinsic only accesses memory that isn't accessible by the module being compiled, or in the case of the latter property memory accessible via its pointer-typed arguments. r287680.
-
Analyses now have an AnalysisKey rather than a
static char PassID. r287783.
Clang commits
-
PPC gained various arithmetic builtins for vector arithmetic in altivec.h. r287872.
-
The OpenCL frontend now differentiates between read and write pipe types, meaning an error can be raised when passing a read-only pipe to write-only pipe argument. r287343.