llvm dev - May 2017 - Runtime-configurable LLVM_DEFAULT_TARGET_TRIPLE by env var

I have been working for extending test coverage for years.
Nowadays, I have several cross-testing (target != host). See
http://bb.pgr.jp/console
Each of them (test-*-linux) is doing;

  - Assume a preceding builder passes with warming ccache.
  - All compilation units will hit ccache whenever the tree is built before
lit.
  - Almost all compilation units will hit ccache except for Host.cpp when
the tree is built with different LLVM_DEFAULT_TARGET_TRIPLE.

FYI, I am doing a trick to let ccache recognize compilation units identical
for build.ninja;

  $ sed -i -r 's=(-I|_COMPILER\S*
)/home/bb/\w[^/]*/llvm-project=\1../llvm-project=g' build.ninja

In other words, testing builders have almost same build except for Host.cpp.
I have to build the tree for each testing target.

I propose an option to override LLVM_DEFAULT_TARGET_TRIPLE in run-time.
The change will be small but I wonder which option I could take.
Please let me know your suggestions.

* Should it be configurable option?

I think it shouldn't be required for release binaries, and it wouldn't
be
required for most developers.
I think it may be off by default.

* What is the name of environment variable?

I suggest the name itself could be configurable.
If I would introduce the envvar name as "LLVM_TARGET", clang users
would
complain why clang didn't have like "CLANG_TARGET".

  $ cmake -DLLVM_DEFAULT_TARGET_ENVVAR_NAME=LLVM_TARGET
  $ LLVM_TARGET=i686-win32 bin/llvm-lit path/to/tests

This feature may be disabled if LLVM_DEFAULT_TARGET_ENVVAR_NAME is blank.
I wonder the name LLVM_DEFAULT_TARGET_ENVVAR_NAME might be weird.
Better suggestions are welcome.

As a future plan, I will implement, in Lit, run tests with the matrix of
triples,
like {arm|mips|i686|x86_64}-{darwin|linux|win32}.
Some of tests should be skipped if it isn't affected to default target
triple.
For example, FileCheck

Regards, Takumi
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20170503/a8b8624d/attachment.html>

Hypothetically you could hack this in with our existing environment
override option, CCC_OVERRIDE_OPTIONS, like so:
ninja all
export CCC_OVERRIDE_OPTIONS=#^--target=$triple
# run lit tests

I'd rather not make a new, easier to use environment variable because it
will encourage people to use it for production cross-compilation. The
existing mechanism is probably good enough for testing purposes.

On Tue, May 2, 2017 at 5:26 PM, NAKAMURA Takumi via llvm-dev <
llvm-dev at lists.llvm.org> wrote:
> I have been working for extending test coverage for years.
> Nowadays, I have several cross-testing (target != host). See
> http://bb.pgr.jp/console
> Each of them (test-*-linux) is doing;
>
>   - Assume a preceding builder passes with warming ccache.
>   - All compilation units will hit ccache whenever the tree is built
> before lit.
>   - Almost all compilation units will hit ccache except for Host.cpp when
> the tree is built with different LLVM_DEFAULT_TARGET_TRIPLE.
>
> FYI, I am doing a trick to let ccache recognize compilation units
> identical for build.ninja;
>
>   $ sed -i -r 's=(-I|_COMPILER\S*
)/home/bb/\w[^/]*/llvm-project=\1../llvm-project=g'
> build.ninja
>
> In other words, testing builders have almost same build except for
> Host.cpp.
> I have to build the tree for each testing target.
>
> I propose an option to override LLVM_DEFAULT_TARGET_TRIPLE in run-time.
> The change will be small but I wonder which option I could take.
> Please let me know your suggestions.
>
> * Should it be configurable option?
>
> I think it shouldn't be required for release binaries, and it
wouldn't be
> required for most developers.
> I think it may be off by default.
>
> * What is the name of environment variable?
>
> I suggest the name itself could be configurable.
> If I would introduce the envvar name as "LLVM_TARGET", clang
users would
> complain why clang didn't have like "CLANG_TARGET".
>
>   $ cmake -DLLVM_DEFAULT_TARGET_ENVVAR_NAME=LLVM_TARGET
>   $ LLVM_TARGET=i686-win32 bin/llvm-lit path/to/tests
>
> This feature may be disabled if LLVM_DEFAULT_TARGET_ENVVAR_NAME is blank.
> I wonder the name LLVM_DEFAULT_TARGET_ENVVAR_NAME might be weird.
> Better suggestions are welcome.
>
> As a future plan, I will implement, in Lit, run tests with the matrix of
> triples,
> like {arm|mips|i686|x86_64}-{darwin|linux|win32}.
> Some of tests should be skipped if it isn't affected to default target
> triple.
> For example, FileCheck
>
> Regards, Takumi
>
> _______________________________________________
> LLVM Developers mailing list
> llvm-dev at lists.llvm.org
> http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20170502/62439303/attachment.html>

On Wed, May 3, 2017 at 9:49 AM Reid Kleckner <rnk at google.com> wrote:
> Hypothetically you could hack this in with our existing environment
> override option, CCC_OVERRIDE_OPTIONS, like so:
> ninja all
> export CCC_OVERRIDE_OPTIONS=#^--target=$triple
> # run lit tests
>
It's an option just in clang driver. It doesn't affect to other clang
tools
(including libclang), and LLVM.

I'd rather not make a new, easier to use environment variable because
it
> will encourage people to use it for production cross-compilation. The
> existing mechanism is probably good enough for testing purposes.
>
Agreed partially. I won't introduce such an option in release binaries.

> On Tue, May 2, 2017 at 5:26 PM, NAKAMURA Takumi via llvm-dev <
> llvm-dev at lists.llvm.org> wrote:
>
>> I have been working for extending test coverage for years.
>> Nowadays, I have several cross-testing (target != host). See
>> http://bb.pgr.jp/console
>> Each of them (test-*-linux) is doing;
>>
>>   - Assume a preceding builder passes with warming ccache.
>>   - All compilation units will hit ccache whenever the tree is built
>> before lit.
>>   - Almost all compilation units will hit ccache except for Host.cpp
when
>> the tree is built with different LLVM_DEFAULT_TARGET_TRIPLE.
>>
>> FYI, I am doing a trick to let ccache recognize compilation units
>> identical for build.ninja;
>>
>>   $ sed -i -r 's=(-I|_COMPILER\S*
>> )/home/bb/\w[^/]*/llvm-project=\1../llvm-project=g' build.ninja
>>
>> In other words, testing builders have almost same build except for
>> Host.cpp.
>> I have to build the tree for each testing target.
>>
>> I propose an option to override LLVM_DEFAULT_TARGET_TRIPLE in run-time.
>> The change will be small but I wonder which option I could take.
>> Please let me know your suggestions.
>>
>> * Should it be configurable option?
>>
>> I think it shouldn't be required for release binaries, and it
wouldn't be
>> required for most developers.
>> I think it may be off by default.
>>
>> * What is the name of environment variable?
>>
>> I suggest the name itself could be configurable.
>> If I would introduce the envvar name as "LLVM_TARGET", clang
users would
>> complain why clang didn't have like "CLANG_TARGET".
>>
>>   $ cmake -DLLVM_DEFAULT_TARGET_ENVVAR_NAME=LLVM_TARGET
>>   $ LLVM_TARGET=i686-win32 bin/llvm-lit path/to/tests
>>
>> This feature may be disabled if LLVM_DEFAULT_TARGET_ENVVAR_NAME is
blank.
>> I wonder the name LLVM_DEFAULT_TARGET_ENVVAR_NAME might be weird.
>> Better suggestions are welcome.
>>
>> As a future plan, I will implement, in Lit, run tests with the matrix
of
>> triples,
>> like {arm|mips|i686|x86_64}-{darwin|linux|win32}.
>> Some of tests should be skipped if it isn't affected to default
target
>> triple.
>> For example, FileCheck
>>
>> Regards, Takumi
>>
>> _______________________________________________
>> LLVM Developers mailing list
>> llvm-dev at lists.llvm.org
>> http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev
>>
>>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20170503/a7a26db9/attachment.html>

I have proposed a patch, thank you.

[CMake] Introduce LLVM_TARGET_TRIPLE_ENV as an option to override
LLVM_DEFAULT_TARGET_TRIPLE at runtime.
https://reviews.llvm.org/D33662

On Wed, May 3, 2017 at 9:26 AM NAKAMURA Takumi <geek4civic at gmail.com>
wrote:
> I have been working for extending test coverage for years.
> Nowadays, I have several cross-testing (target != host). See
> http://bb.pgr.jp/console
> Each of them (test-*-linux) is doing;
>
>   - Assume a preceding builder passes with warming ccache.
>   - All compilation units will hit ccache whenever the tree is built
> before lit.
>   - Almost all compilation units will hit ccache except for Host.cpp when
> the tree is built with different LLVM_DEFAULT_TARGET_TRIPLE.
>
> FYI, I am doing a trick to let ccache recognize compilation units
> identical for build.ninja;
>
>   $ sed -i -r 's=(-I|_COMPILER\S*
> )/home/bb/\w[^/]*/llvm-project=\1../llvm-project=g' build.ninja
>
> In other words, testing builders have almost same build except for
> Host.cpp.
> I have to build the tree for each testing target.
>
> I propose an option to override LLVM_DEFAULT_TARGET_TRIPLE in run-time.
> The change will be small but I wonder which option I could take.
> Please let me know your suggestions.
>
> * Should it be configurable option?
>
> I think it shouldn't be required for release binaries, and it
wouldn't be
> required for most developers.
> I think it may be off by default.
>
> * What is the name of environment variable?
>
> I suggest the name itself could be configurable.
> If I would introduce the envvar name as "LLVM_TARGET", clang
users would
> complain why clang didn't have like "CLANG_TARGET".
>
>   $ cmake -DLLVM_DEFAULT_TARGET_ENVVAR_NAME=LLVM_TARGET
>   $ LLVM_TARGET=i686-win32 bin/llvm-lit path/to/tests
>
> This feature may be disabled if LLVM_DEFAULT_TARGET_ENVVAR_NAME is blank.
> I wonder the name LLVM_DEFAULT_TARGET_ENVVAR_NAME might be weird.
> Better suggestions are welcome.
>
> As a future plan, I will implement, in Lit, run tests with the matrix of
> triples,
> like {arm|mips|i686|x86_64}-{darwin|linux|win32}.
> Some of tests should be skipped if it isn't affected to default target
> triple.
> For example, FileCheck
>
> Regards, Takumi
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20170530/5941d9a5/attachment.html>