llvm dev - Apr 2014 - [LLVMdev] Proposal: Move host CPU auto-detection out of the TargetMachine

We already decide on a default. For these three targets is “whatever the user is
running on.” For all the other targets it’s something the backend selects. I’m
proposing we, as you say, remove the uncertainty and have the default always be
the same from one run to the next even (especially) when those runs are on
different machines.

For X86, there’s two options, the “generic” target that backend already supports
or “core2” which is what clang typically defaults to for x86_64h. I lean towards
the former for simplicity. We also don’t have to change all three of the ports.
I’m personally motivated primarily by x86, and was intending to leave it to the
maintainers of the other two targets to decide what they want to do here,
perhaps with some strong encouragement.

It’s very important that a run of “llc” on one machine produce the same output
on two heterogenous machines given the same input and command lines*. That’s not
true right now, leading to lots of bot failures that patch originators can’t
reproduce because they’re getting different code locally due to the
auto-detection. The recent “Test failures with 3.4.1” thread for examples.

-Jim

* With the caveat of platform differences like elf vs. macho. A similar argument
can be made there, but unlike CPU, there’s no sensible way to define a
least-common-denominator default, so it’s harder. I’m explicitly avoiding
opening that can of worms.

On Apr 9, 2014, at 10:19 PM, Eric Christopher <echristo at gmail.com>
wrote:
> I'm not a huge fan of this because then you get to decide on a default
> for all the ports, but I can understand if people want to move this
> way to reduce uncertainty.
> 
> Alternately have a way for the backend to pretty print out the results
> of the auto-detection when asked?
> 
> *shrug*
> 
> -eric
> 
> On Tue, Apr 8, 2014 at 1:59 AM, Jim Grosbach <grosbach at apple.com>
wrote:
>> All,
>> 
>> Currently the X86 backend does CPU auto-detection and subtarget feature
detection when the TargetMachine is created if no explicit CPU was specified.
It's counterintuitive for low level tools like 'llc' to do this, as
it means the same .ll file compiled on heterogenous machines generates different
results from the same 'llc' command line. It is still useful to be able
to opt-in to such behavior to, for example, replicate clang's behavior when
-mcpu=native is supplied to clang. My thought is to do something similar here
and teach 'llc' to recognize -mcpu=native and probe the host CPU if that
is given. The subtarget features will then be filled in according to the feature
string for that CPU. This (a) changes the auto-detection from opt-out to opt-in
and (b) moves the logic out of the core target backend and into the tools
drivers.
>> 
>> Attached are draft patches that do this for X86. Similar but smaller
cleanups can also be done for SystemZ and PowerPC if it's agreed this is a
good idea.
>> 
>> -Jim
>> 
>> 
>> 
>> _______________________________________________
>> LLVM Developers mailing list
>> LLVMdev at cs.uiuc.edu         http://llvm.cs.uiuc.edu
>> http://lists.cs.uiuc.edu/mailman/listinfo/llvmdev
>>

On Apr 10, 2014, at 21:32 , Jim Grosbach <grosbach at apple.com> wrote:
> We already decide on a default. For these three targets is “whatever the
user is running on.” For all the other targets it’s something the backend
selects. I’m proposing we, as you say, remove the uncertainty and have the
default always be the same from one run to the next even (especially) when those
runs are on different machines.
> 
> For X86, there’s two options, the “generic” target that backend already
supports or “core2” which is what clang typically defaults to for x86_64h. I
lean towards the former for simplicity. We also don’t have to change all three
of the ports. I’m personally motivated primarily by x86, and was intending to
leave it to the maintainers of the other two targets to decide what they want to
do here, perhaps with some strong encouragement.
> 
> It’s very important that a run of “llc” on one machine produce the same
output on two heterogenous machines given the same input and command lines*.
That’s not true right now, leading to lots of bot failures that patch
originators can’t reproduce because they’re getting different code locally due
to the auto-detection. The recent “Test failures with 3.4.1” thread for
examples.
As one of the contributors to pocl <http://pocl.sourceforge.net/>, I want
to emphatically chime in here. Before our 0.9 release, we had to discover that
running clang, llc, opt, etc. often give results that differ depending on which
machine they are run, and also between each other. We ended up specifying
exactly what kind of machine is to be targeted, via various -triple, -march, and
-mcpu options.

The fact that these options have a different syntax for different LLVM tools
didn't help...

Any step towards more unification among different machines and different
LLVM-based tools would be a great help.

-erik
> -Jim
> 
> * With the caveat of platform differences like elf vs. macho. A similar
argument can be made there, but unlike CPU, there’s no sensible way to define a
least-common-denominator default, so it’s harder. I’m explicitly avoiding
opening that can of worms.
> 
> On Apr 9, 2014, at 10:19 PM, Eric Christopher <echristo at gmail.com>
wrote:
> 
>> I'm not a huge fan of this because then you get to decide on a
default
>> for all the ports, but I can understand if people want to move this
>> way to reduce uncertainty.
>> 
>> Alternately have a way for the backend to pretty print out the results
>> of the auto-detection when asked?
>> 
>> *shrug*
>> 
>> -eric
>> 
>> On Tue, Apr 8, 2014 at 1:59 AM, Jim Grosbach <grosbach at
apple.com> wrote:
>>> All,
>>> 
>>> Currently the X86 backend does CPU auto-detection and subtarget
feature detection when the TargetMachine is created if no explicit CPU was
specified. It's counterintuitive for low level tools like 'llc' to
do this, as it means the same .ll file compiled on heterogenous machines
generates different results from the same 'llc' command line. It is
still useful to be able to opt-in to such behavior to, for example, replicate
clang's behavior when -mcpu=native is supplied to clang. My thought is to do
something similar here and teach 'llc' to recognize -mcpu=native and
probe the host CPU if that is given. The subtarget features will then be filled
in according to the feature string for that CPU. This (a) changes the
auto-detection from opt-out to opt-in and (b) moves the logic out of the core
target backend and into the tools drivers.
>>> 
>>> Attached are draft patches that do this for X86. Similar but
smaller cleanups can also be done for SystemZ and PowerPC if it's agreed
this is a good idea.
>>> 
>>> -Jim
>>> 
>>> 
>>> 
>>> _______________________________________________
>>> LLVM Developers mailing list
>>> LLVMdev at cs.uiuc.edu         http://llvm.cs.uiuc.edu
>>> http://lists.cs.uiuc.edu/mailman/listinfo/llvmdev
>>> 
> 
> 
> _______________________________________________
> LLVM Developers mailing list
> LLVMdev at cs.uiuc.edu         http://llvm.cs.uiuc.edu
> http://lists.cs.uiuc.edu/mailman/listinfo/llvmdev
-- 
Erik Schnetter <schnetter at gmail.com>
http://www.perimeterinstitute.ca/personal/eschnetter/

My email is as private as my paper mail. I therefore support encrypting
and signing email messages. Get my PGP key from http://pgp.mit.edu/.

-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 203 bytes
Desc: Message signed with OpenPGP using GPGMail
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20140410/063d9ec9/attachment.sig>

On Thu, Apr 10, 2014 at 6:32 PM, Jim Grosbach <grosbach at apple.com>
wrote:
> It’s very important that a run of “llc” on one machine produce the same
> output on two heterogenous machines given the same input and command
> lines*. That’s not true right now, leading to lots of bot failures that
> patch originators can’t reproduce because they’re getting different code
> locally due to the auto-detection. The recent “Test failures with 3.4.1”
> thread for examples.
>
I think we should do this, but only to make llc's behavior more
deterministic and predictable. I don't think we should start checking in
tests that rely on the default subtarget features without explicitly
requesting the relevant features.

Consider somebody who works on an ARM or x86 variant like atom.  Probably
what they'll want to do is set up a bot that runs the LLVM test suite in
such a way that their subtarget features are on by default.  Our test suite
currently "supports" that if the host CPU features happen to be the
ones
you want.  Instead, we should probably move to a world where bots can set
different defaults by configuring the tools appropriately.  For example,
they could rewrite 'llc' to 'llc -mcpu=blah' in lit if no mcpu
flags exist.

This would be similar to what we do in Clang for the default C++ ABI. If
the test actually needs the MSVC or Itanium C++ ABI, they ask for it
explicitly. Otherwise they test one or the other depending on the default
target triple. I don't think we should double the number of RUN lines to
keep that test coverage, and I don't think the cost of Windows-bot-only
test failures is too high.

2c
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20140410/44b0fcf0/attachment.html>

On Thu, Apr 10, 2014 at 7:41 PM, Reid Kleckner <rnk at google.com>
wrote:
> On Thu, Apr 10, 2014 at 6:32 PM, Jim Grosbach <grosbach at apple.com>
wrote:
>>
>> It's very important that a run of "llc" on one machine
produce the same
>> output on two heterogenous machines given the same input and command
lines*.
>> That's not true right now, leading to lots of bot failures that
patch
>> originators can't reproduce because they're getting different
code locally
>> due to the auto-detection. The recent "Test failures with
3.4.1" thread for
>> examples.
>
>
> I think we should do this, but only to make llc's behavior more
> deterministic and predictable. I don't think we should start checking
in
> tests that rely on the default subtarget features without explicitly
> requesting the relevant features.
So this is largely my concern here. I hate breaking the Atom bots all
the time. That said, breaking the Atom bots all the time keeps me
aware that I want to make my tests resilient against various different
cpu types or that I should be picking explicit options to test a
particular backend feature.

-eric
>
> Consider somebody who works on an ARM or x86 variant like atom.  Probably
> what they'll want to do is set up a bot that runs the LLVM test suite
in
> such a way that their subtarget features are on by default.  Our test suite
> currently "supports" that if the host CPU features happen to be
the ones you
> want.  Instead, we should probably move to a world where bots can set
> different defaults by configuring the tools appropriately.  For example,
> they could rewrite 'llc' to 'llc -mcpu=blah' in lit if no
mcpu flags exist.
>
> This would be similar to what we do in Clang for the default C++ ABI. If
the
> test actually needs the MSVC or Itanium C++ ABI, they ask for it
explicitly.
> Otherwise they test one or the other depending on the default target
triple.
> I don't think we should double the number of RUN lines to keep that
test
> coverage, and I don't think the cost of Windows-bot-only test failures
is
> too high.
>
> 2c