[clang] [llvm] [VectorCombine] Shrink loads used in shufflevector rebroadcasts (PR #128938)

[Simon Pilgrim via llvm-commits]

================
@@ -3483,6 +3487,125 @@ bool VectorCombine::foldInterleaveIntrinsics(Instruction &I) {
   return true;
 }
 
+// Attempt to shrink loads that are only used by shufflevector instructions.
+bool VectorCombine::shrinkLoadForShuffles(Instruction &I) {
+  auto *OldLoad = dyn_cast<LoadInst>(&I);
+  if (!OldLoad || !OldLoad->isSimple())
+    return false;
+
+  auto *OldLoadTy = dyn_cast<FixedVectorType>(OldLoad->getType());
+  if (!OldLoadTy)
+    return false;
+
+  unsigned const OldNumElements = OldLoadTy->getNumElements();
+
+  // Search all uses of load. If all uses are shufflevector instructions, and
+  // the second operands are all poison values, find the minimum and maximum
+  // indices of the vector elements referenced by all shuffle masks.
+  // Otherwise return `std::nullopt`.
+  using IndexRange = std::pair<int, int>;
+  auto GetIndexRangeInShuffles = [&]() -> std::optional<IndexRange> {
+    IndexRange OutputRange = IndexRange(OldNumElements, -1);
+    for (auto &Use : I.uses()) {
----------------
RKSimon wrote:

do you have test cases with multiple uses?

https://github.com/llvm/llvm-project/pull/128938