This is the mail archive of the mailing list for the GCC project.

Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]
Other format: [Raw text]

Notes from GNU Cauldron 2014

First of all, thanks to the organisers and the speakers for a very
enjoyable and informative event. I took notes during many of the talks
and someone suggested I share these notes to the list in case anyone
is interested. I'm afraid I missed the steering committee Q+A and
didn't take notes on the release management Q+A.

You can find a markdown version of my notes pasted below, or else in
HTML at the end of Details on the
speakers can be found at

### Glibc BoF
* 2.20 is in "slushy" freeze mode. What else is left? fmemopen, fd locking,
some `-Wundef` work
* Anyone planning to check in something big for 2.21?
    * Mentor Graphics planning to check in a NIOS II port. They won't be
accepted until Linux kernel patches are in a kernel release.
    * A desire for AArch64 ILP32 ABI to get in. Kernel patches currently in
review, compiler work is ready.
    * OpenRISC
    * NaCl (nptl)
* Benchmarking glibc? Does anyone have a good approach. There is a preload
library approach (see notes from Ondrej's talk).
* Glibc has been built with AddressSanitizer, help needed to get it integrated
into the build system. There was a comment this would be nice to get in to
* Red Hat are working on supporting alternate libm implementations, including
a low-precision and high-precision implementation. Intel are looking to add
math functions that work on small vectors.

### Abigail: toward ABI taming
* Want to determine if changes to your shared library break apps for users,
and users want to know whether an updated library remains compatible with
their code. The bidiff tool will tell you the differences in terms of ABI
given two object files as its input.
* libabi consists of modules such as a DWARF reader, the comparison engine.
Tools such as bidiff are built on this API
* What's next for libabigail?
    * bicompat will help application authors determine whether their
    application A is still compatibile with an updated version of a given
    library L by examining the undefined symbols of A that are resolved by L.
    * More amenable to automation (such as integration into build systems)
    * Support for un-instantiated templates. This would require declarations
    of uninstantiated templates to be represented in DWARF.
* A first official release (though source is available at

### Writing VMs in Java and debugging them with GDB
* Oracle Labs have been working on various dynamic language implementations in
Java (e.g. Ruby, Python, R, JS, ...).
* FastR is a reimplementation of R in Java featuring an interpreter (Truffle)
and dynamic compiler (Graal).
* Truffle and Graal starts with an AST interpreter. The first time a node is
evaluated it is specialised to the type that was seen at runtime. Later the
tree is compiled using partial evaluation.
* It may be deployed on standard HotSpot (no compilation), GraalVM, or the
SubstrateVM (SVM) which uses Graal to ahead-of-time compile the language
implementation. Debugging the SVM is difficult as Java debugging tools are not
available. The solution is to generate DWARF information in the SVM's output.
* Truffle and Graal are open source, the SubstrateVM is not (yet?).

### GCC and LLVM collaboration
* Good news: license issues, personal grudges and performance are off-topic.
* Users should be protected from whatever disagreements take place. In the
future we should have more pro-active discussions on various issues as opposed
to reactive discussions regarding e.g. compiler flags that have been noticed
to be arbitrarily different after the fact.
* Renato lists common projects that we may collaborate on: binutils, glibc,
sanitizers. Sanitizers are a collaboration success story.
* Can we agree on a (new?) common user interface?
* There's a surprising amount of confusion about `-march`, `-mtune`, and
`-mcpu` considering we're in a room of compiler developers. It sounds like
there's not much support for re-engineering the set of compiler flags as the
potential gain is not seen as being great enough.
* Can we agree to standardise on attributes, C/C++ extensions, builtins, ASM,
the linker API?
* GCC docs have just been rewritten, so some criticisms about how difficult it
is to dig in are no longer valid.

### Machine Guided Energy Efficient Compilation
* Initial investigations in 2012 found that compiler flags can have a
meaningful effect on energy consumption. This raises the question of how to
determine which flags to use.
* MAGEEC will target both GCC and LLVM initially. It is implemented as a
compiler plugin which performs feature extraction and allows the output of the
machine learning algorithm to change the sequence of passes which are run.
Fractional Factorial Design is used to reduce the optimisation space to
* Turning passes on/off arbitrarily can often result in internal compiler
errors. Should the machine learning algorithm learn this, or should GCC better
document pass requirements?
* It would be useful to MAGEEC if the (currently internal) plugin API could be
stabilized. They also currently have to use a hacked up Clang as it doesn't
provide plugin hooks.
* The project has produced a low cost energy measurement board as well as
their own benchmark suite (Bristol/Embecosm Embedded Benchmark Suite, or
BEEBS). BEEBS 2.0 is schedule for release by 31st August 2014 with a much
wider range of benchmarks (currently 93). Jeremy showed a rather pleasing live
demo where you can run a benchmark on a microcontroller development board and
immediately find the number of mJ consumed in running it.
* The current state of the project has it not achieving better results than
GCC O2, but this is expected to change over the coming months.

### Just-in-time compilation using GCC
* []( is an experimental branch of
GCC which allows you to build GCC as a shared library and embed it in other
programs in order to allow in-process code generation at runtime.
* A dedicated API for JIT will allow better stability guarantees. It provides
a high-level API designed for ease of use.
* The API doesn't offer solutions for type inference, escape analysis,
unboxing, inline caching, etc.
* It has a C++ API wich includes some cunning operator overloading to
massively reduce verbosity, and a Python API.
* David Malcolm has written
[Coconut](, a JIT compiler for Python
using It is incomplete and experimental.
* Drawback: currently have to write out a .s to a file and invoke gcc on it.
Some might make a cheeky comment about the benefits of architecting a compiler
so it can be used as a library, but I of course wouldn't dare. The good news
is the speaker is actively looking at what would be needed to use GAS and GNU
ld as a library.

### Introduction to new Intel SIMD ISA and its impact on GCC
* AVX-512 offers 64 simple precision or 32 double precision floating point
operations per cycle. It also has 8x64-bit mask registers.
* Rounding modes can be set on a per-instruction process
* Basic support is available from GCC 4.9.x.

### News from Sanitizers
* MemorySanitizer detects use of uninitialized memory. Increases CPU by about
2.5x and RAM by 2x. Was released in LLVM in 2013. It is currently Linux/x86-64
* History growth is limited by limiting the history depth and the number of
new history nodes per stack trace.
* MSan has found hundreds of bugs across Google internal code, Chromium, LLVM,
etc. It was more challenging for Chromium due to the number of system libs
that had to be rebuilt.
* AddressSanitizer annotations allows you to detect access to the regions of
e.g. `std::vector<>` which has been allocated as part of its capacity but not
yet been used (i.e. will start to be used in the next `push_back`). Next is to
do the same for `std::string` and `std::deque`.
* Glibc uses GNU-C instead of ANSI C which currently prevents compilation with
Clang (nested functions in particular are problematic). It can however be
built with ASan by GCC.
* Evgeniy comments that the lack of standardisation between Clang and GCC for
things like `__has_feature(address_sanitizer)` vs `__SANITIZE_ADDRESS__` is
irritating. This is just the sort of thing Renato was talking about yesterday
of course.

### glibc performance tuning
* Use memset as an example. Look at 3 variants.
* Writing a useful benchmark is more difficult than you might think.
Simply running `memset` many times in a loop is not a good benchmark when
using the same memory locations due to the processor's load-store forwarding.
Even when fixing this, the branch predictor may perform much better than it
would when memset is used in a real world scenario and lead to
unrepresentative results.
* To move beyond microbenchmarks, Ondrej has been using `LD_PRELOAD` to link
against instrumented versions of the functions which record details about the
time taken.
* See
for memset benchmarks and
for more background.
* strcmp was the most frequently called glibc function in Ondrej's testing
(when running Firefox).

### Devirtualization in GCC
* This is a special case of indirect call removal, and although the talk is
given in the context of C++ the techniques apply to other languages too. Some
basic cases are handled in the front-end and even specified by the language
* It is a special case of constant propagation across aggregates, which is
already done by Global Value Numbering and Interprocedural Constant
Propagation. But these passes only catch a tiny number of possible cases.
* Loss of information between the frontend and middle end can make some cases
almost impossible. The intermediate language can be extended with explicit
representations of base types, locations of virtual table pointers, and
vtables. Also annotate polymorphic calls specifying instance and polymorphic
call type and flags to denote constructors/destructors.
* I'm not able to summarise details on the GCC devirt implementation better
than the slides do. Hopefully they'll be made available online.
* A particular challenge is to match types between different compilation
units. The C++ One Definition Rule is used.
* It can be used to strengthen unreachable function removal.
* Feedback-directed devirtualization was extended in GCC 4.9 to work
inter-module with LTO.


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]