[llvm-commits] [llvm] r99549 - /llvm/trunk/lib/Target/ARM/ARMSubtarget.cpp

[Evan Cheng]

On Mar 25, 2010, at 1:57 PM, Jim Grosbach wrote:

> I'll look at the x86 stuff and see if there's a better approach that'll work here. Thanks for the pointer. I agree having the conditional like this is a bit "eww."
> 
> That said, I went with this for two reasons: it's consistent with what we already do for the NEON-for-scalar-math flag, and we only want to disable the instructions for specific CPUs (cortex-a8 for now), not whole architecture variants, and as far as I know, the CPUString flag is the only place we have that information.

You can specify features for individual cpu's. I think that's a cleaner way to go about it.

Evan

> 
> 
> On Mar 25, 2010, at 1:54 PM, Evan Cheng wrote:
> 
>> Thanks Jim. But perhaps this can be a subtarget feature specified in ARM.td? For example, x86 has FeatureSlowBTMem which disables uses of certain instructions for specific variants.
>> 
>> Evan
>> 
>> On Mar 25, 2010, at 1:48 PM, Jim Grosbach wrote:
>> 
>>> Author: grosbach
>>> Date: Thu Mar 25 15:48:50 2010
>>> New Revision: 99549
>>> 
>>> URL: http://llvm.org/viewvc/llvm-project?rev=99549&view=rev
>>> Log:
>>> ARM cortex-a8 doesn't do vmla/vmls well. disable them by default for that cpu
>>> 
>>> Modified:
>>>  llvm/trunk/lib/Target/ARM/ARMSubtarget.cpp
>>> 
>>> Modified: llvm/trunk/lib/Target/ARM/ARMSubtarget.cpp
>>> URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Target/ARM/ARMSubtarget.cpp?rev=99549&r1=99548&r2=99549&view=diff
>>> ==============================================================================
>>> --- llvm/trunk/lib/Target/ARM/ARMSubtarget.cpp (original)
>>> +++ llvm/trunk/lib/Target/ARM/ARMSubtarget.cpp Thu Mar 25 15:48:50 2010
>>> @@ -127,6 +127,12 @@
>>>   // operations with NEON instructions.
>>>   if (UseNEONFP.getPosition() == 0)
>>>     UseNEONForSinglePrecisionFP = true;
>>> +    // The VFP vlma and vlms instructions don't play nicely with others;
>>> +    // disable them.
>>> +    // FIXME: This may be true for other variants as well. Get benchmark
>>> +    // numbers and add them if determined that's the case.
>>> +    if (UseVMLxInstructions.getPosition() == 0)
>>> +      UseVMLx = false;
>>> }
>>> }
>>> 
>>> 
>>> 
>>> _______________________________________________
>>> llvm-commits mailing list
>>> llvm-commits at cs.uiuc.edu
>>> http://lists.cs.uiuc.edu/mailman/listinfo/llvm-commits
>> 
>