[llvm] [LoadStoreVectorizer] Fill gaps in load/store chains to enable vectorization (PR #159388)

[Drew Kersnar via llvm-commits]

================
@@ -623,6 +663,25 @@ std::vector<Chain> Vectorizer::splitChainByContiguity(Chain &C) {
     dumpChain(C);
   });
 
+  // If the chain is not contiguous, we try to fill the gap with "extra"
+  // elements to artificially make it contiguous, to try to enable
+  // vectorization. We only fill gaps if there is a potentially legal masked
+  // load/store for the target. If later on, we don't end up with a chain that
+  // could be vectorized into a legal masked load/store, the chains with extra
+  // elements will be filtered out in splitChainByAlignment.
+  bool TryFillGaps = shouldAttemptMaskedLoadStore(C);
+
+  unsigned ASPtrBits =
+      DL.getIndexSizeInBits(getLoadStoreAddressSpace(C[0].Inst));
+
+  // Compute the alignment of the leader of the chain (which every stored offset
+  // is based on) using the current first element of the chain. This is
+  // conservative, we may be able to derive better alignment by iterating over
+  // the chain and finding the leader.
+  Align LeaderOfChainAlign =
----------------
dakersnar wrote:

> The alignment of the leader is independent of the alignment of C[0] =>my claim is, we're only using it, combined with the offset of C[0] from the leader to make a conservative approximation of the leader's alignment.

You're 100% correct in this claim, but I was just trying to avoid changing the Chain data structure for such a narrow edge case. To be specific, this alignment deriving is _only_ useful in the rare event where an extra element ends up as the leader of a chain. Changing the data structure from a SmallVector to a struct that contains a SmallVector + a leader alignment seems like it is not worth the complication to the overall file.

> commonAlign(best possible alignment in the chain, offset of gap load/store from the best aligned load/store)

This might work better, but the code will be a bit more complex. Let me try to implement it and then get your opinion.

https://github.com/llvm/llvm-project/pull/159388