[LLVMdev] Upcoming Changes/Additions to Scoped-NoAlias metadata

[Hal Finkel]

Hi everyone,

As many of you might know, LLVM now has scoped noalias metadata (http://llvm.org/docs/LangRef.html#noalias-and-alias-scope-metadata) -- it allows us to preserve noalias function argument attributes when inlining, in addition to allowing frontends to add otherwise non-deducible aliasing properties to memory accesses. This currently works well, but needs a change and an intrinsic, as I'll explain below.

First, the problem: Currently, ScopedNoAliasAA.cpp is a little bit too much like TypeBasedAliasAnalysis.cpp in that when the metadata is used to return an aliasing result, the Size passed in the AliasAnalysis::Location object is ignored. This is not a problem if the Size is equal to (or less than) the size of the access with which the metadata was originally associated, but if the Size is larger than that associated with the original access, the result might not be correct. Fixing this is not hard for regular accesses (we can also store the size of the original access), although is harder for arbitrary function calls. 

Now you might think that constructing an AliasAnalysis::Location object with a size larger than the original is uncommon, and in some sense it is, but this is exactly what the loop vectorizer does when partitioning memory accesses in a loop into potential aliasing sets. The loop vectorizer partitions a loop's memory accesses by taking the access's natural Location object, and setting the size to UnknownSize (the largest representable size) and querying using that infinite-size Location. The logic is that if two access don't alias with both sizes set to infinity, then they must always come from disjoint sets of underlying objects, and because the vectorizer only considers access pointers that are linear recurrences, implies that the accesses don't alias both within and across loop iterations.

So how does this work now? Consider this function:

void foo(float * restrict a, float * restrict b) {
  for (int i = 0; i < 1600; ++i)
    a[i] = b[i] + 1;
}
(note that restrict at the C level becomes a noalias function argument attribute at the IR level)

With the scoped noalias metadata, the aliasing information from the 'a' and 'b' function arguments is preserved such that 'a[i]' and 'b[i]' are tagged as not aliasing. There is nothing wrong with this, and in fact, when the loop vectorizer uses its infinite-sized queries on 'a[i]' and 'b[i]', and that returns NoAlias, that result is also not wrong. It is not wrong, however, because the noalias function argument attributes, logically associated with the function entry block, dominate the loop ('a[i]' is based on 'a' for all 'i', and similarly for 'b[i]').

Now here's the problematic case:

void inner(float * restrict a, float * restrict b) {
  *a = *b + 1;
}
void foo(float * a, float * b) {
  for (int i = 0; i < 1600; ++i)
    inner(a+i, b+i);
}

The problem here is that the noalias function argument attributes on 'inner', when inlined into the loop body, do not dominate the loop; they apply to pointers within each loop iteration, not across loop iterations. But because the generated noalias metadata is essentially the same as in the previous case, and as noted ScopedNoAliasAA.cpp ignores the Location's Size, the loop vectorizer's infinite-sized alias queries return NoAlias as in the previous case. This is a bug. Making ScopedNoAliasAA.cpp check the size of the original access (as it should) would fix this problem, but it will also prevent ScopedNoAliasAA.cpp from returning NoAlias for the first case were we'd like it to do so.

So to summarize, scoped-noalias metadata does not preserve enough dominance information to truly capture the full semantics of the noalias function argument attributes.

After discussing this with Chandler offline last week, here's the proposed solution: instead of having both !alias.scope and !noalias metadata, we'll have only !alias.scope metadata and an intrinsic: i8* @llvm.noalias(i8* ptr, !metadata !?) where the metadata argument corresponds to a list of !alias.scopes. The idea being that the pointer returned by this intrinsic, and all pointers derived from it, are assumed not to alias with memory accesses tagged with any of the associated !alias.scope metadata entries. This intrinsic needs to carry control dependencies (it cannot be hoisted out of a loop, for example) -- in this sense it is very much like @llvm.assume. And like @llvm.assume, we'll need to add logic to various passes to ignore it as appropriate so that it does not block optimizations unnecessarily. I was hoping this avoid this part of the design space, but I don't see any way around it -- only some non-hoistable instruction can model a control dependence.

With this in place, it will be possible to build a loop-aware AA interface capable of answering questions like: Does location A in some loop iteration alias with location B in any iteration (which is really what the vectorizer wants to know). This interface can check whether the @llvm.noalias dominates the loop while answering this query.

On the bright side, the metadata will be simpler (or at least less verbose) than the current design.

If anyone has suggestions on an alternative design, please share! :-) Regardless, the current design needs to change: Just fixing it to check the size of the location will make it non-useful for the loop vectorizer and otherwise cripple it for mod-ref queries against arbitrary function calls.

Thanks again,
Hal

-- 
Hal Finkel
Assistant Computational Scientist
Leadership Computing Facility
Argonne National Laboratory