LLVM Weekly - #56, Jan 26th 2015
Welcome to the fifty-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 asb@asbradbury.org, or or @llvmweekly or @asbradbury on Twitter.
I'll be talking the lowRISC project to produce a fully open-source SoC at FOSDEM this coming weekend. Do come and see my main track talk and read my speaker interview for more background. There is of course an LLVM toolchain devroom on the Sunday.
News and articles from around the web
Stephen Diehl has written an absolutely fantastic tutorial on writing an LLVM specializer for Python, guiding you through the process of creating something like Numba.
A new tool, Dwgrep (DWARF Grep) may be of interest to many LLVM Weekly readers. This blog post gives an intro to using it.
Paul Smith has a blog post on getting started with the LLVM C API.
A post on the official LLVM Blog announces that LLDB is coming to Windows, announcing to a wider audience that it is now possible to debug simple programs with LLDB on Windows and giving a rationale for investing effort into porting LLDB to Windows and adding support for the MS debug format. The post also features a todo list indicating what's next for Windows support.
A draft version 0.1 of the IA-32 psABI (processor specific application binary interface) is available. This aims to supplement the existing System V ABI with conventions relevant to newer features such as SSE1-4 and AVX. Comments are welcome.
LLVM/Clang 3.6-rc1 is now available. Get testing and filing bugs.
ELLCC 0.1.8 has been released. ELLCC is an LLVM/Clang-based cross compilation toolchain.
LLDB now has it's own IRC channel. You'll want to join #lldb on irc.oftc.net.
On the mailing lists
-
Chandler Carruth has posted a canonicalization-related RFC. He demonstrates a case where a trivial function is compiled to two equivalent IR sequences. This was later committed.
-
Michael Zolotukhin has proposed an RFC on adding a heuristic for complete loop unrolling. Currently, the loop unrolling heuristics don't take account of any new optimisations that maybe enabled by unrolling the loop. Changing that would allow for the profitability of the unroll to be more accurately approximated.
-
Chandler Carruth is getting ready to turn on by default the next part of his vector shuffle work. Now is a good time to benchmark and report any regressions you see with
-x86-experimental-vector-shuffle-legality
on your codebases. -
Ahmed Bougacha has been having problems with the cost model calculations for saturation instructions. The cost is over-estimated because a number of the individual IR instructions fold-away later in lowering. He suggests adding a new method to TargetTransformInfo for multi-instruction cost computation. There hasn't been much feedback thus far.
-
Chandler Carruth has been looking through the LLD libraries and trying to work out the current layering, as well as what a potential future layering might be. He proposes offering a basic library offering basic functionality and a second library offering a higher-level interface for actually doing linking.
LLVM commits
-
A backend targeting the extended BPF (Berkeley Packet Filter) interpreter/JIT in the Linux kernel has been added. See this LWN article for more background. r227008.
-
The initial version of the new ORC JIT API has landed. r226940.
-
There's been a flurry of work on the new pass manager this week. One commit I will choose to pick out is the port of InstCombine to the new pass manager, which seems like a milestone or sorts. r226987.
-
LLVM learnt how to use the GHC calling convention on AArch64. r226473.
-
InstCombine will now canonicalize loads which are only ever stored to always use a legal integer type if one is available. r226781.
-
The
llvm_any_ty
type for intrinsics has been born. r226857. -
llvm-objdump now understands
-indirect-symbols
to dump the Mach-O indirect symbol table. r226848.
Clang commits
-
Clang now supports SPIR calling conventions. r226548.
-
It's now possible to set the stack probe size on the command line. r226601.
-
Clang gained initial support for Win64 SEH IR emission. r226760.