[llvm] [InstCombine] Extend folding of aggregate construction to cases when source aggregates are partially available (PR #100828)

[Nikita Popov via llvm-commits]

================
@@ -1117,9 +1124,68 @@ Instruction *InstCombinerImpl::foldAggregateConstructionIntoAggregateReuse(
     // aggregate produced by OrigIVI must have been originally extracted from
     // the same aggregate. Is that so? Can we find said original aggregate?
     SourceAggregate = FindCommonSourceAggregate(UseBB, Pred);
-    if (Describe(SourceAggregate) != AggregateDescription::Found)
-      return nullptr; // Give up.
-    IV.first->second = *SourceAggregate;
+    if (Describe(SourceAggregate) == AggregateDescription::Found) {
+      FoundSrcAgg = true;
+      IV.first->second = *SourceAggregate;
+    } else {
+      // If UseBB is the single successor of Pred, we can add InsertValue to
+      // Pred.
+      auto *BI = dyn_cast<BranchInst>(Pred->getTerminator());
+      if (!BI || !BI->isUnconditional())
+        return nullptr;
+    }
+  }
+
+  if (!FoundSrcAgg)
+    return nullptr;
+
+  // Do some sanity check if we need to add insertvalue into predecessors.
+  auto OrigBB = OrigIVI.getParent();
+  for (auto &It : SourceAggregates) {
+    if (Describe(It.second) == AggregateDescription::Found)
+      continue;
+
+    // Element is defined in UseBB, so it can't be used in predecessors.
+    if (EltDefinedInUseBB)
+      return nullptr;
+
+    // Do this transformation cross loop boundary may cause dead loop. So we
+    // should avoid this situation. But LoopInfo is not generally available, we
+    // must be conservative here.
+    // If OrigIVI is in UseBB and it's the only successor of PredBB, PredBB
+    // can't be in inner loop.
+    if (UseBB != OrigBB)
+      return nullptr;
----------------
nikic wrote:

I have a hard time understanding why this check is sufficient to prevent cycles in all cases, or even how the UseBB != OrigBB condition is related. I guess this is supposed to check the case where OrigBB is a loop header with phis and UseBB is an exit block with inserts -- but couldn't it fail on non-sunk inserts?

https://github.com/llvm/llvm-project/pull/100828