[LV] Create in-loop sub reductions

This PR allows the loop vectorizer to handle sub reductions by forming a
normal add reduction with a negated input.

Author: SamTebbs33

llvm/lib/Analysis/IVDescriptors.cpp

@@ -1263,6 +1263,10 @@ RecurrenceDescriptor::getReductionOpChain(PHINode *Phi, Loop *L) const {
     if (isFMulAddIntrinsic(Cur))
       return true;
 
+    // Recognize a sub reduction. It gets canonicalized to add(sub (0, ...)).
+    if (Cur->getOpcode() == Instruction::Sub && getOpcode() == Instruction::Add)
+      return true;
+
     return Cur->getOpcode() == getOpcode();
   };
 

llvm/lib/Transforms/Vectorize/LoopVectorize.cpp

@@ -9144,6 +9144,14 @@ void LoopVectorizationPlanner::adjustRecipesForReductions(
             CurrentLinkI->getFastMathFlags());
         LinkVPBB->insert(FMulRecipe, CurrentLink->getIterator());
         VecOp = FMulRecipe;
+      } else if (PhiR->isInLoop() && Kind == RecurKind::Add &&
+                 CurrentLinkI->getOpcode() == Instruction::Sub) {
+        Type *PhiTy = PhiR->getUnderlyingValue()->getType();
+        auto *Zero = Plan->getOrAddLiveIn(ConstantInt::get(PhiTy, 0));
+        VPWidenRecipe *Sub = new VPWidenRecipe(
+            *CurrentLinkI, {Zero, CurrentLink->getOperand(1)});
+        LinkVPBB->insert(Sub, CurrentLink->getIterator());
+        VecOp = Sub;
       } else {
         if (RecurrenceDescriptor::isMinMaxRecurrenceKind(Kind)) {
           if (isa<VPWidenRecipe>(CurrentLink)) {

llvm/test/Transforms/LoopVectorize/reduction-inloop.ll

@@ -627,7 +627,7 @@ for.end:                                          ; preds = %for.body, %entry
   ret float %result.0.lcssa
 }
 
-; Sub we can create a reduction, but not inloop
+; Sub we can create a reduction inloop
 define i32 @reduction_sub_lhs(ptr noalias nocapture %A) {
 ; CHECK-LABEL: @reduction_sub_lhs(
 ; CHECK-NEXT:  entry:
@@ -636,15 +636,16 @@ define i32 @reduction_sub_lhs(ptr noalias nocapture %A) {
 ; CHECK-NEXT:    br label [[VECTOR_BODY:%.*]]
 ; CHECK:       vector.body:
 ; CHECK-NEXT:    [[INDEX:%.*]] = phi i64 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]
-; CHECK-NEXT:    [[VEC_PHI:%.*]] = phi <4 x i32> [ zeroinitializer, [[VECTOR_PH]] ], [ [[TMP1:%.*]], [[VECTOR_BODY]] ]
+; CHECK-NEXT:    [[VEC_PHI:%.*]] = phi i32 [ 0, [[VECTOR_PH]] ], [ [[TMP3:%.*]], [[VECTOR_BODY]] ]
 ; CHECK-NEXT:    [[TMP0:%.*]] = getelementptr inbounds i32, ptr [[A:%.*]], i64 [[INDEX]]
 ; CHECK-NEXT:    [[WIDE_LOAD:%.*]] = load <4 x i32>, ptr [[TMP0]], align 4
-; CHECK-NEXT:    [[TMP1]] = sub <4 x i32> [[VEC_PHI]], [[WIDE_LOAD]]
+; CHECK-NEXT:    [[TMP4:%.*]] = sub nsw <4 x i32> zeroinitializer, [[WIDE_LOAD]]
+; CHECK-NEXT:    [[TMP1:%.*]] = call i32 @llvm.vector.reduce.add.v4i32(<4 x i32> [[TMP4]])
+; CHECK-NEXT:    [[TMP3]] = add i32 [[TMP1]], [[VEC_PHI]]
 ; CHECK-NEXT:    [[INDEX_NEXT]] = add nuw i64 [[INDEX]], 4
 ; CHECK-NEXT:    [[TMP2:%.*]] = icmp eq i64 [[INDEX_NEXT]], 256
 ; CHECK-NEXT:    br i1 [[TMP2]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP26:![0-9]+]]
 ; CHECK:       middle.block:
-; CHECK-NEXT:    [[TMP3:%.*]] = call i32 @llvm.vector.reduce.add.v4i32(<4 x i32> [[TMP1]])
 ; CHECK-NEXT:    br i1 true, label [[FOR_END:%.*]], label [[SCALAR_PH]]
 ; CHECK:       scalar.ph:
 ; CHECK-NEXT:    br label [[FOR_BODY:%.*]]