[llvm-dev] [cfe-dev] RFC: change -fp-contract=off to actually disable FMAs

[Stephen Canon via llvm-dev]

Echoing what everyone else has said, keying on the word “fused” is a red herring here.

fp-contract refers to behavior governed by the STDC FP_CONTRACT pragma. “Contraction” has a formal definition in the C standard:

> A floating expression may be contracted, that is, evaluated as though it were a single operation, thereby omitting rounding errors implied by the source code and the way to disallow contracted expressions. 

Note that this definition is *purely* in terms of the rounding of arithmetic operations performed by the abstract machine; there is no notion of instructions generated. Formation of fused multiply-add instructions is one specific form of fusion licensed by this pragma, which happens to be the main one of interest from the standpoint of compiler performance optimization for FMA-based architectures.

There’s some imprecision in the documentation caused by a mismatch between what’s interesting for compiler writers (where rounding changes due to FMA formation are allowed) and the abstract specification. That should be cleaned up. However, fp-contract is not a knob to control whether or not abstract-machine operations generate a single arithmetic instruction—it definitely does not, and should not, enable or disable MAD formation.

– Steve

> On Jul 10, 2019, at 6:14 PM, Scott Manley via cfe-dev <cfe-dev at lists.llvm.org> wrote:
> 
> >  I think you have a different definition of fused then. Fused is a description of how the operation is computed/rounded, not an instruction count.  
> 
> "Only fuse FP ops when the result won't be affected" is what the existing comment says. So it can't be both a fused op and not a fused op if it's only meant to imply a difference in rounding. I'm just re-using the existing wording, and I agree it could be cleaned up if that's the intent of the -fp-contract option -- which I why I was asking for context.
> 
> >  For FMA, I think your example IR is correctly handled. The fast instruction flag should override the global FP option you’re providing. For the issue you are describing, this is more of a question of whether clang should be emitting the fast flag or not.  
> 
> I disagree. How does clang know what would ultimately form an FMA?  It would have to blanket remove 'fast' from all fadds. 
> 
>> 
>> On Wed, Jul 10, 2019 at 4:16 PM Matt Arsenault <arsenm2 at gmail.com> wrote:
>> 
>>> On Jul 10, 2019, at 16:56, Scott Manley <rscottmanley at gmail.com> wrote:
>>> 
>>> At any rate, I was only offering an additional reason. Personally I think it's strange for an option to say "this will never fuse ops" and then under the covers will fuse ops, regardless of how FMAD is defined. However, my primary concern is for FMAs. They have both numeric and performance implications and I do not think it's unreasonable that off means off.
>> 
>> I think you have a different definition of fused then. Fused is a description of how the operation is computed/rounded, not an instruction count. The F in FMAD is not fused (I know this naming scheme is not great. Every other FP node besides FMA has an F prefix)
>> 
>> For FMA, I think your example IR is correctly handled. The fast instruction flag should override the global FP option you’re providing. For the issue you are describing, this is more of a question of whether clang should be emitting the fast flag or not.