[LLVMdev] Parallel Loop Metadata

[Pekka Jääskeläinen]

On 02/08/2013 08:20 PM, Tobias Grosser wrote:
> That sounds elegant and seems to solve the correctness problems.

There is no big difference here except that the memory instructions
would not need the metadata.

I do not think the abundance of metadata is really the main problem, but
what to do with passes that do not know about parallel loops. How to minimize 
the changes needed to make the most relevant passes retain the information,
and more importantly, how to avoid code breakage if they are unaware of
its meaning.

Also, I tried something like this when I tried to find a nice way to couple
the memory and the loop metadata. It seems not possible to directly refer to
an Instruction from MDNode, or is it? I got crashes when serializing the
bitcode with MDNodes pointing to Instructions and thought that it's not meant
to work that way around.

Thus, I created the loop identifier metadata "anchor" to which the both
metadata types refer.

Even if it was possible, retaining that information across inlining etc. means
one has to update those Instruction references in the metadata to the new
matching ones in the inlined location, etc. The maintenance burden of the data
in several passes, which I thought was the actual concern of Nadav, gets AFAIU
even harder. E.g., if you remove memory instructions (e.g. DCE) you get
dangling pointers in the metadata or a null there which renders the parallel 
info invalid (and might even cause crashes?), which doesn't happen in the
current state of the patch.

All in all, if the number of metadata is what pains the most, this
is an improvement. Otherwise, I see it more fragile, and it's also uncertain if
it's possible to implement it.

IMHO, the best alternative presented so far is to have both metadata types
and potentially exploit the mem.parallel_loop_access MD even if there were
non-annotated instructions in the loop to *help* the dependency analysis.
isParallel() would stay as is now in the patch, it can be used as a short cut
check for a loop of which accesses all have the metadata (to skip the whole
dependency checking).

-- 
--Pekka