[llvm-commits] [llvm] r89211 - /llvm/trunk/CMakeLists.txt

[Daniel Dunbar]

On Sun, Nov 22, 2009 at 10:51 AM, Nick Lewycky <nicholas at mxc.ca> wrote:
> Daniel Dunbar wrote:
>>
>> On Thu, Nov 19, 2009 at 1:15 PM, Óscar Fuentes<ofv at wanadoo.es>  wrote:
>>>
>>> Hello Daniel.
>>>
>>> Daniel Dunbar<daniel at zuster.org>  writes:
>>>
>>>> URL: http://llvm.org/viewvc/llvm-project?rev=89211&view=rev
>>>> Log:
>>>> Turn LLVM_BUILD_EXAMPLES off by default in CMake builds, to match
>>>> Makefiles&
>>>> Clang.
>>>
>>> What's the rationale for this?
>>>
>>> (And wouldn't it warrant the corresponding update of docs/CMake.html,
>>> where default values for variables are listed?)
>>
>> The rationale for changing the default (for cmake and make, I assume
>> we agree they should have the same default) is that most people
>> building LLVM don't care about the examples (because they are
>> developers), and this greatly increases cycle times because of the
>> sheer number of link steps it adds. Having examples be "optional"
>> builds for people exploring LLVM seems reasonable to me, the only
>> value I see in building them regularly is to prevent them from being
>> out-dated, but I will change (one, fast) buildbot to maintain that.
>>
>> I will update the doc.
>
> The rationale for copying each Kaleidoscope chapter into the examples/
> directory was to make sure that our examples on the website don't get out of
> date when someone changes the API.

I agree this is good.

> This just means that people will run make and make check, commit, and then
> discover that their change somehow broke the buildbots. It'd be nice to
> notice and resolve the breakage before committing.

I explained my view on this with my reply to Duncan, and there is no
counterargument here. I can quite easily argue that the increased
cycle time of the buildbots (and the corresponding improvement in
blamelist accuracy) as a group more than makes up for the rare times
when someone breaks an example.

 - Daniel