[LLVMdev] Enable changing UnsafeFPMath on a per-function basis

[Akira Hatanaka]

On Mon, Jan 12, 2015 at 2:26 PM, Eric Christopher <echristo at gmail.com>
wrote:

Hi Eric,

Hi Akira,
>
> On Thu Jan 08 2015 at 5:36:54 PM Akira Hatanaka <ahatanak at gmail.com>
> wrote:
>
>> To continue the discussion I started last year (see the link below) on
>> embedding command-line options in bitcode, I came up with a plan to improve
>> the way the backend changes UnsafeFPMath on a per-function basis. The code
>> in trunk currently resets TargetOptions::UnsafeFPMath at the beginning of
>> SelectionDAGISel::runOnMachineFunction to enable compiling one function
>> with “unsafe-fp-math=true” and another with “unsafe-fp-math=false”, but
>> this won’t work if we want to parallelize the backend in the future, which
>> is something Eric mentioned in his talk last year.
>>
>> Here is the link to the original discussion I started last year:
>> http://lists.cs.uiuc.edu/pipermail/llvmdev/2014-November/078785.html
>>
>>
> Thanks for coming back around to this. I'd hoped to have most of this
> handled, but as you can tell it's a nice herd of yaks to get everything
> playing nicely.
>
>
>> These are the changes I plan to make:
>>
>> 1. In llc.cpp (and any other tools that use the option), change the
>> function attribute in the IR based on the value of command line option
>> “enable-unsafe-pf-math” (EnableUnsafeFPMath, defined in CommandFlags.h).
>>
>> 2. Replace usages of TargetOptions::UnsafeFPMath with calls to a
>>
> function which gets the value of attribute “unsafe-fp-math” in the IR.
>>
>
> These two combined seems a little weird. Do you mean that you plan on,
> effectively, changing the IR if someone passes a command line flag to an
> existing set of IR?
>
>

Yes, my current plan is to change the existing attributes in the IR when
command line options are passed. I wasn't sure if changing the IR was a
good idea, but it seemed to be the simplest way to handle command line
options passed to llc using the current cl::opt infrastructure. This also
depends on how the command line infrastructure will look like after Chris
checks in his patches.

3. Stringify function attribute “unsafe-fp-math” and append it to the
>> string that is used as the lookup key in
>> TargetMachine::getSubtargetImpl(const Function&). Also, pass the function
>> attribute to the subtarget constructors that need it to construct itself
>> (e.g., ARM) or construct one of the backend objects (e.g., X86, which needs
>> it in the constructor of X86TargetLowering).
>>
>>
> Another thought is to put it as part of the feature string rather than
> adding it as a separate attribute. That way you won't have to add a lookup
> in each target.
>
>

Yes, that's another way to pass the option, which is probably less
error-prone than passing each function attribute. Does it mean that we
define SubtargetFeatures for unsafe-fp-math and other options too? The
downside of this approach is that it will needlessly create new subtarget
objects for targets that currently don't need to specialize subtargets
based on the value of "unsafe-fp-math", although I'm not sure how much
impact it will have on memory.

4. In MachineFunction’s constructor, call
>> TargetMachine::getSubtargetImpl(const Function &) to initialize STI
>> (pointer to the per-function subtarget object). Currently, the version of
>> TargetMachine::getSubtargetImpl that doesn’t take a const Function&
>> parameter is called, which returns the module-level subtarget object.
>>
>
> This can be done first if you like at the moment, though you may run into
> other problems and testing may be somewhat minimal.
>

OK. I'll spend a little more time investigating and testing and send a
patch for it if everything goes well.


>
>
>> 5.  Fix ARMAsmPrinter::emitAttributes to compute the value of
>> TargetOptions::UnsafeFPMath based on the function attributes of all the
>> functions in the module being compiled (see the link below).
>>
>> http://lists.cs.uiuc.edu/pipermail/llvmdev/2014-December/079904.html
>>
>> Yes.
>
>
>> 6. Move the code in CGCall.cpp that sets the function attributes to
>> BackendUtil.cpp. clang should set the function attributes regardless of
>> whether it is compiling from source code or from an IR file (e.g., clang
>> foo1.ll -o foo1.s -ffast-math), but currently this happens only if it’s
>> compiling from source.
>>
>
> This is an interesting solution, but involves changing IR files and I
> don't know that we want to do this (as I mentioned above). But I do admit I
> like it more than having to check both TargetOptions and the attribute to
> do it.
>
> Another thing you haven't brought up is the inliner (or other cross
> function optimizations)- what do you plan on doing there? Think two files,
> one compiled with and one compiled without and linked together.
>
>
I haven't thought much about how I should fix the inliner as I was thinking
this should be tackled as a separate problem. Since "unsafe-fp-math" is
already embedded in the IR as a function attribute, we don't have to wait
for my changes to be checked in and we can start fixing the problem today.
Is that right?


> -eric
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20150113/3e8f5f1e/attachment.html>