llvm dev - May 2013 - [LLVMdev] AArch64 build error

Hi,

I'm getting a "random" build failure related to AArch64 when
building with cmake. The build is running with -j32 so it seems that there is
some dependency issue that causes the build to fail when things build in a
certain order.

Is anyone else seeing this problem?

paul

3008	make[2]: *** No rule to make target
`lib/Target/AArch64/AArch64GenDAGISel.inc', needed by
`lib/Target/AArch64/Utils/CMakeFiles/LLVMAArch64Utils.dir/AArch64BaseInfo.cpp.o'.
Stop.
3009	make[2]: Leaving directory
`/home/pulseagent/.pulse2-agent/data/agents/109155/recipes/791203/base/build-cmake'
3010	make[1]: *** [lib/Target/AArch64/Utils/CMakeFiles/LLVMAArch64Utils.dir/all]
Error 2
3011	make[1]: /home/pulseagent/local/bin/cmake -E cmake_progress_report
/home/pulseagent/.pulse2-agent/data/agents/109155/recipes/791203/base/build-cmake/CMakeFiles
3012	*** Waiting for unfinished jobs....

Hi Paul,
> I'm getting a "random" build failure related to AArch64 when
building with cmake.
> The build is running with -j32 so it seems that there is some dependency
issue
> that causes the build to fail when things build in a certain order.
> 3008    make[2]: *** No rule to make target
`lib/Target/AArch64/AArch64GenDAGISel.inc', needed by
Hmm. That file's generated by TableGen (via some kind of custom make
magic) and included by the "regular" sources, so it's entirely
possible there's a race hiding. I always thought all TableGen commands
were run before the main sources got compiled, but at 32 threads
perhaps that doesn't matter so much.

Has this just started since AArch64 was added, or have you only
recently started building at -j32? As far as I can see AArch64 is
unexceptional in its dependencies, except for being lexicographically
the first target. I'd be interested in your results if AArch64 was
disabled.

There's clearly some issue though. I've only got 4 cores (+ 4 faked)
here so I'm not optimistic about reproducing any races you're seeing,
though I'll keep trying periodic builds. How many cores do you have
and how frequently does this seem to occur? Your host compiler and
configure command-line could be useful for controlling the environment
too.

Tim.

Hi Tim,

On 2013-05-17 4:18 PM, "Tim Northover" <t.p.northover at
gmail.com> wrote:
>Hi Paul,
>
>> I'm getting a "random" build failure related to AArch64
when building
>>with cmake.
>> The build is running with -j32 so it seems that there is some
>>dependency issue
>> that causes the build to fail when things build in a certain order.
>
>> 3008    make[2]: *** No rule to make target
>>`lib/Target/AArch64/AArch64GenDAGISel.inc', needed by
>
>Hmm. That file's generated by TableGen (via some kind of custom make
>magic) and included by the "regular" sources, so it's entirely
>possible there's a race hiding. I always thought all TableGen commands
>were run before the main sources got compiled, but at 32 threads
>perhaps that doesn't matter so much.
>
>Has this just started since AArch64 was added, or have you only
>recently started building at -j32? As far as I can see AArch64 is
>unexceptional in its dependencies, except for being lexicographically
>the first target. I'd be interested in your results if AArch64 was
>disabled.
I can't say for sure if this is a new problem but we've been building on
this machine for a few weeks now.

I commented out AArch64 from LLVM_ALL_TARGETS and rebuilt but I can't
reproduce the error in this case--it appears to be AArch64 specific.

One thing I forgot to mention is that our build machines use ccache. When
I don't use ccache I can't reproduce the error either. However, I
don't
think ccache is the problem it just makes the race easier to reproduce.
>
>There's clearly some issue though. I've only got 4 cores (+ 4 faked)
>here so I'm not optimistic about reproducing any races you're
seeing,
>though I'll keep trying periodic builds. How many cores do you have
>and how frequently does this seem to occur? Your host compiler and
>configure command-line could be useful for controlling the environment
>too.
The machine has 16 physical cores each hyper-threaded. I'm just using gcc
4.7.2 from Ubuntu 12.10.

I can reproduce this with the following:

$cmake -DCMAKE_CXX_COMPILER="ccache"
-DCMAKE_CXX_COMPILER_ARG1="c++"
../llvm/
$make -j32
$make clean
$make -j32

The second make will fail with the error I pasted before.

paul
>
>Tim.