[cfe-dev] computed goto/labels as values in interpreters

[Carsten Mattner]

On Thu, Aug 23, 2012 at 6:14 PM, Carsten Mattner
<carstenmattner at gmail.com> wrote:
> On Sat, Aug 18, 2012 at 1:55 PM, Carsten Mattner
> <carstenmattner at gmail.com> wrote:
>> On Fri, Aug 17, 2012 at 8:50 PM, Eli Friedman <eli.friedman at gmail.com> wrote:
>>> On Fri, Aug 17, 2012 at 11:42 AM, Carsten Mattner
>>> <carstenmattner at gmail.com> wrote:
>>>> On Fri, Aug 17, 2012 at 7:40 PM, Eli Friedman <eli.friedman at gmail.com> wrote:
>>>>> On Fri, Aug 17, 2012 at 8:14 AM, Carsten Mattner
>>>>> <carstenmattner at gmail.com> wrote:
>>>>>> After searching list archives and bug reports I wasn't able to conclude
>>>>>> whether computed goto/labels as values as used in Erlang and other (Guile)
>>>>>> interpreters is supported in clang 3.1. With some BSDs working to make
>>>>>> clang/clang++ the base cc/c++ this is going to be even more important.
>>>>>>
>>>>>> Is clang 3.1 supposed to compile beam_emu.c correctly and with the
>>>>>> same or comparable optimization result?
>>>>>>
>>>>>> https://github.com/erlang/otp/blob/master/erts/emulator/beam/beam_emu.c
>>>>>
>>>>> Yes, it's supposed to.
>>>>
>>>> Yes, seems to work but with some interesting differences compared to
>>>> gcc. While gcc-4.7 takes 6s to recompile Erlang's emulator if I
>>>> 'touch beam_emu.c', clang takes 2m11s and consumes around 141MB RES vs
>>>> 40MB gcc.
>>>
>>> IIRC, there's been some work in this area since 3.1.  If you're still
>>> seeing issues on trunk, please file a bug.
>>
>> Thanks for the info. From what I can see beam_emu did get built and
>> I was able to run the emulator. I didn't do any performance test because
>> AFAIR previously it didn't result in a runnable executable at all. This
>> result with enough success for now.
>>
>>>> Not sure why but clang also seemed to result in a longer
>>>> time to run ar. Tests were conducted on i386 linux 3.5.1 with
>>>> glibc-2.15 against erlang's maint branch. Next, a process which is
>>>> known to be slow is building some of the generated C++ files in lib/wx
>>>> (wxWidgets Erlang bindings). Now, when linking erl_gl.so with clang
>>>> it took forever to finish (didn't time). This is similar to the ar
>>>> archive step in erts/emulator. Is there anything inherently slow with
>>>> gcc-4.7, glibc-2.15, and binutils-2.22? ar and ld from binutils at
>>>> fault? While linking erl_gl.so RES of clang was at 265MB. It seemed
>>>> to be an infinite loop and I stopped it after more than 5 minutes.
>>>
>>> I'm not following this... linking shouldn't be spending any
>>> substantial amount of time in the "clang" process.  clang just invokes
>>> ld.
>>
>> Me neither, it's totally unexpected and surprising.
>>
>>> If you have a reproducible infinite loop, please file a bug.
>>
>> I won't be able to conduct another set of tests including clang trunk
>> before the middle of next week, sorry.
>>
>> If anyone wants to try in the meantime here are the steps:
>>
>> $ git clone git://github.com/erlang/otp.git
>> $ cd otp
>> $ export ERL_TOP=$PWD
>> $ export PATH=$ERL_TOP/bin:$PATH
>> $ export CC=clang
>> $ export CXX=clang++
>> $ ./otp_build setup -a --prefix=$PWD/localinstall
>> $ make install
>> You can run and test the emulator without installing locally.
>> If you want to build the emulator only to see about beam_emu.c
>> $ ./otp_build autoconf && ./otp_build configure --prefix=$PWD/localinstall
>> $ cd erts && make
>> erl_gl.so is built in lib/wx
>> $ cd lib/wx && make
>> This naturally requires wxWidgets
>
> Could the long and in wxErlang's case never ending link phase be explained
> by the use of the gold linker in this linux distro's clang 3.1 package?
> Are there known issues/incompatibilities with that?

I've built llvm and clang from git today and built otp.git maint branch
with both clang and gcc and didn't forget to time it.

clang:
$ export CC=clang
$ export CXX=clang++
$ export MAKEFLAGS=-j2
$ /usr/bin/time ./otp_build setup -a
1549.29user 90.32system 17:42.24elapsed 154%CPU (0avgtext+0avgdata
167100maxresident)k
8968inputs+314328outputs (36major+21665287minor)pagefaults 0swaps

gcc in a new shell after dropping/flusing the kernel's fs cache:
$ export MAKEFLAGS=-j2
$ /usr/bin/time ./otp_build setup -a
1355.43user 89.78system 14:22.01elapsed 167%CPU (0avgtext+0avgdata
73656maxresident)k
218288inputs+335936outputs (333major+19829177minor)pagefaults 0swaps

This seems to support the previous results.

I don't know whether this clang build uses the gold linker, it's
all just defaults detected by ./configure.

Can Eli or someone else confirm that clang/llvm is slower and uses
more memory when building Erlang/OTP?