[pooma-dev] Re: InlineEvaluator implementation question

[Jeffrey Oldham]

Richard Guenther wrote:
> On Mon, 16 Dec 2002, Mark Mitchell wrote:
> 
> 
>>--On Monday, December 16, 2002 04:55:52 PM +0100 Richard Guenther
>><rguenth at tat.physik.uni-tuebingen.de> wrote:
>>
>>
>>>Hi!
>>>
>>>Does anyone remember why we create copies of the LHS and RHS inside
>>>the KernelEvaluator<InlineKernelTag>::evaluate() methods (within
>>>ReductionEvaluator<InlineKernelTag>::evaluate() is similar code)? I.e.
>>>there is code like
>>>
>>>  template<class LHS,class Op,class RHS,class Domain>
>>>  inline static void evaluate(const LHS& lhs,const Op& op,const RHS& rhs,
>>>                              const Domain& domain,WrappedInt<1>)
>>>  {
>>>    CTAssert(Domain::unitStride);
>>>    PAssert(domain[0].first() == 0);
>>>    LHS localLHS(lhs);
>>>    RHS localRHS(rhs);
>>>    int e0 = domain[0].length();
>>>    for (int i0=0; i0<e0; ++i0)
>>>      op(localLHS(i0),localRHS.read(i0));
>>>  }
>>
>>I'm pretty sure that this copy allowed some C++ compilers (KCC) to see
>>that some parts of lhs/rhs were loop-invariant, and then hoist references
>>to those fields out of the loop.  (The compiler can see that nothing can
>>modify localLHS; it's less obvious to it that nothing can modify rhs
>>since it doesn't know what else might point to that location.)
> 
> 
> Hmm - as both, lhs and rhs are declared const, isnt this enough to tell
> the compiler? Or has the compiler to assume every function call can have
> a side-effect on any (but local) variable?
> 
> Well, at least gcc creates worse (larger) code with copying than without.

Using the copies measurably reduced execution time in previous experiments.

Thanks,
Jeffrey D. Oldham
oldham at codesourcery.com