LLVM Weekly - #109, Feb 1st 2016
Welcome to the one hundred and ninth 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 firstname.lastname@example.org, 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
The GNU Tools Cauldron 2016 has been announced for the 9th-11th of September 2016, in Hebden Bridge, UK.
The Sulong project has been announced. It is an LLVM IR interpreter using the Truffle framework and Graal on the JVM to support JIT compilation.
Ehsan Akhgari has posted an updated on building Firefox with clang-cl. It is now possible to build a complete Firefox with Clang without using the MSVC fallback once.
I've mentioned it down below in the list of notable commits, but it's worth calling out here too: the old autoconf build-system has now been removed from LLVM. 3.8 will be the last release to include it. Time to switch to CMake if you haven't already.
John Regehr gave a talk about undefined behaviour in LLVM at the Paris LLVM meetup, and you can find the slides here.
On the mailing lists
James Knight has written to the list to get feedback on approaches to cleaning up Clang's handling of atomics. There seems to be widespread support for the cleanup. James followed up again to slightly revise his plan.
Matt Arsenault proposes that all libcalls be canonicalized to intrinsics. All responses so far are in favour.
Ke Bai has shared a proposal on representing multiple memory scopes in LLVM IR. There hasn't been any feedback yet.
Dmitree Kuvaiskii asks if anyone has implemented a pass utilizing Intel's new MPX memory protection. The answer appears to be no, and in addition David and Kostya are sceptical about how worthwhile it would be.
Peter Collingbourne has proposed a new optimisation, virtual constant propagation. The original motivation was to reduce the overhead added by enabling control-flow integrity in certain Chromium benchmarks. Constants will be devirtualized at LTO time.
The autoconf build system for LLVM has been removed. r258861.
The WebAssembly backend gained support for unaligned loads and stores. r258779.
LLVM's MCAsmSreamer will now always use .p2align rather than .align, because .align's behaviour can differ between targets. r258750.
Intrinsic IDs are now looked up by binary search rather than the previous more complex mechanism. This improves the compile time of Function.cpp. r258774.
TargetSelectionDAGInfo has been renamed to SelectionDAGTargetInfo and now lives in CodeGen rather than Target. r258939.
A LoopSimplifyCFG pass was added to canonicalise loops before running through passes such as LoopRotate and LoopUnroll. r259256.
The clang-cl driver will now warn for unknown arguments rather than erroring, to match the behaviour of MSVC. r258720.
The old autoconf build system was removed from Clang. r258862.
The 'sancov' (SanitizerCoverage) tool gained some documentation. r259000.
Other project commits
lld gained a new error function which won't cause process exit. The hope is this can be used to provide a gradual path towards lld-as-a-library. r259069.
The lit runner for the LLVM test suite can now be passed
--param=profile=perfwhich will cause each test to be run under
perf record. r259051.