[llvm-commits] [llvm] r170503 - in /llvm/trunk/docs: Vectorizers.rst subsystems.rst

[Nadav Rotem]

Hi Shuxin, 

Yes, I agree. At the moment the check is too conservative and, as you said, we can check that P and Q are "vector-len" bytes apart.

Thanks,
Nadav 

On Dec 19, 2012, at 12:36 AM, Shuxin Yang <shuxin.llvm at gmail.com> wrote:

> Hi, Nadav:
> 
>  I guess we are bit conservative in this case.
> 
> +
> +  void bar(float *A, float* B, float K, int start, int end) {
> +   for (int i = start; i < end; ++i)
> +     A[i] *= B[i] + K;
> +  }
> +
>  Let us use a simpler case:
> 
>  void foo(float *P, float *Q) {
>     for (i = 0; i < symbolic-val; i++)
>       P[i] += Q[i];
>  }
> 
>  We don't need examine if P[*] overlap Q[*]. I guess condition
> "Q <= P || (Q - P) >= Vector-length" is enough. The reasoning is following:
> 
>  Break the statement in questions into following simpler stmts:
> 
>  S1: t1 = Q[i]
>  S2: t2 = P[i]
>  S3: P[i] = t1 + t2
> 
> There are bunch of dep here. We only care if the dependences form a SCC.
> SCC could exist only if there is a dep from S3 to S1 arising from mem-access.
> 
> So if Q <= P, there is mem-flow-dep at all.
> if Q - P = m (m> 0), there is mem-flow-dep between S3 and S1; however, if
> m >= vector-length, we can still go ahead vectorize in the presence of SCC.
> 
> 
>      
> 
>