[llvm] [SLP]Model reduction_add(ext(<n x i1>)) as ext(ctpop(bitcast <n x i1> to int n)) (PR #116875)

[Simon Pilgrim via llvm-commits]

================
@@ -17885,24 +17897,37 @@ void BoUpSLP::computeMinimumValueSizes() {
   // Add reduction ops sizes, if any.
   if (UserIgnoreList &&
       isa<IntegerType>(VectorizableTree.front()->Scalars.front()->getType())) {
-    for (Value *V : *UserIgnoreList) {
-      auto NumSignBits = ComputeNumSignBits(V, *DL, 0, AC, nullptr, DT);
-      auto NumTypeBits = DL->getTypeSizeInBits(V->getType());
-      unsigned BitWidth1 = NumTypeBits - NumSignBits;
-      if (!isKnownNonNegative(V, SimplifyQuery(*DL)))
-        ++BitWidth1;
-      unsigned BitWidth2 = BitWidth1;
-      if (!RecurrenceDescriptor::isIntMinMaxRecurrenceKind(::getRdxKind(V))) {
-        auto Mask = DB->getDemandedBits(cast<Instruction>(V));
-        BitWidth2 = Mask.getBitWidth() - Mask.countl_zero();
+    // Convert vector_reduce_add(ZExt(<n x i1>)) to ZExtOrTrunc(ctpop(bitcast <n
+    // x i1> to in)).
+    if (all_of(*UserIgnoreList,
+               [](Value *V) {
+                 return cast<Instruction>(V)->getOpcode() == Instruction::Add;
+               }) &&
+        VectorizableTree.front()->State == TreeEntry::Vectorize &&
+        VectorizableTree.front()->getOpcode() == Instruction::ZExt &&
+        cast<CastInst>(VectorizableTree.front()->getMainOp())->getSrcTy() ==
+            Builder.getInt1Ty()) {
+      ReductionBitWidth = 1;
+    } else {
+      for (Value *V : *UserIgnoreList) {
+        auto NumSignBits = ComputeNumSignBits(V, *DL, 0, AC, nullptr, DT);
+        auto NumTypeBits = DL->getTypeSizeInBits(V->getType());
----------------
RKSimon wrote:

(style) avoid auto

https://github.com/llvm/llvm-project/pull/116875