[RFC] #pragma ivdep

[Redmond, Paul]

On 2013-03-07, at 4:57 AM, Tobias Grosser wrote:

> On 03/07/2013 12:35 AM, Redmond, Paul wrote:
>> 
>> On 2013-03-06, at 3:42 PM, Pekka Jääskeläinen wrote:
>> 
>>> Hi Paul,
>>> 
>>> On 03/06/2013 10:11 PM, Redmond, Paul wrote:
>>>> I have updated the patch to not add metadata on loads and stores where the
>>>> pointer comes from an alloca. I wonder if the check should be more
>>>> conservative and only include pointers coming from Arguments and GlobalValues
>>>> (perhaps Constants too?)
>>> 
>>> I think it's safer that way around for now.
>> 
>> Hmm.. I'm not sure that this approach (or my original) is general enough. Here's a simple loop and the corresponding IR as generated by clang:
> 
> Before taking decisions on what exactly we need to do during code generation, I would appreciate if we could write down the exact semantics of #pragma ivdep on the C level. We should get a list of interesting test cases and investigate what other compilers gcc and icc do in such cases. Pekka and me gave a couple of test cases, that may be a good start.
> 

I haven't been able to find a detailed description of #pragma ivdep semantics other than "ignore assumed vector dependencies" which is pretty vague.

I think the most basic semantic requirement at the C level is that the loop:

for (i...) {
  A;
  B;
  C;
  ...
}

may be distributed as (without reordering statements):

for (i...) A;
for (i...) B;
for (i...) C;
for (i...) ...

to produce the same result.

Regarding the case:

#pragma ivdep
for (i = 0; i < 100; i++) {
  long t[100];
  t[0] = i + 2;
  A[i] = A[i+b] + t[0];
}

ICC is supposed to privatize t. I'm currently investigating how to handle this during codegen--how to decide when to add the metadata and when not to.

paul

> Regarding your current approach. I wonder if, instead of analyzing the IR that clang generates, we could take decisions according to how
> a certain memory access is/was represented at the C level. This is the level the #pragma ivdep semantics are defined and this should allow us to take the right conclusions.
> 
> A safe start would e.g. be to only annotate loads/stores that come from
> array style memory accesses ala 'A[i][j]', where A was declared outside
> of the loop that was marked with #pragma ivdep.
> 
> Cheers,
> Tobi