[VPlan] Use op from underlying call in computeCost if needed.

This fixes a divergence between legacy and VPlan-based cost model, e.g.
if one of the operands has an first-order recurrence phi as operand.

Author: fhahn

llvm/lib/Transforms/Vectorize/VPlanRecipes.cpp

@@ -939,13 +939,19 @@ InstructionCost VPWidenCallRecipe::computeCost(ElementCount VF,
   if (auto *FPMO = dyn_cast_or_null<FPMathOperator>(getUnderlyingValue()))
     FMF = FPMO->getFastMathFlags();
 
-  // Some backends analyze intrinsic arguments to determine cost. If all
-  // operands are VPValues with an underlying IR value, use the original IR
-  // values for cost computations.
+  // Some backends analyze intrinsic arguments to determine cost. Use the
+  // underlying value for the operand if it has one. Otherwise try to use the
+  // operand of the underlying call instruction, if there is one. Otherwise
+  // clear Arguments.
+  // TODO: Rework TTI interface to be independent of concrete IR values.
   SmallVector<const Value *> Arguments;
-  for (VPValue *Op : operands()) {
+  for (const auto &[Idx, Op] : enumerate(operands())) {
     auto *V = Op->getUnderlyingValue();
     if (!V) {
+      if (auto *UI = dyn_cast_or_null<CallBase>(getUnderlyingValue())) {
+        Arguments.push_back(UI->getArgOperand(Idx));
+        continue;
+      }
       Arguments.clear();
       break;
     }

llvm/test/Transforms/LoopVectorize/AArch64/call-costs.ll

@@ -0,0 +1,83 @@
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 5
+; RUN: opt -p loop-vectorize -S %s | FileCheck %s
+
+target triple = "arm64-apple-macosx11.0.0"
+
+define void @fshl_operand_first_order_recurrence(ptr %dst, ptr noalias %src) {
+; CHECK-LABEL: define void @fshl_operand_first_order_recurrence(
+; CHECK-SAME: ptr [[DST:%.*]], ptr noalias [[SRC:%.*]]) {
+; CHECK-NEXT:  [[ENTRY:.*]]:
+; CHECK-NEXT:    br i1 false, label %[[SCALAR_PH:.*]], label %[[VECTOR_PH:.*]]
+; CHECK:       [[VECTOR_PH]]:
+; CHECK-NEXT:    br label %[[VECTOR_BODY:.*]]
+; CHECK:       [[VECTOR_BODY]]:
+; CHECK-NEXT:    [[INDEX:%.*]] = phi i64 [ 0, %[[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], %[[VECTOR_BODY]] ]
+; CHECK-NEXT:    [[VECTOR_RECUR:%.*]] = phi <2 x i64> [ <i64 poison, i64 0>, %[[VECTOR_PH]] ], [ [[WIDE_LOAD1:%.*]], %[[VECTOR_BODY]] ]
+; CHECK-NEXT:    [[TMP0:%.*]] = add i64 [[INDEX]], 0
+; CHECK-NEXT:    [[TMP1:%.*]] = add i64 [[INDEX]], 2
+; CHECK-NEXT:    [[TMP2:%.*]] = getelementptr inbounds i64, ptr [[SRC]], i64 [[TMP0]]
+; CHECK-NEXT:    [[TMP3:%.*]] = getelementptr inbounds i64, ptr [[SRC]], i64 [[TMP1]]
+; CHECK-NEXT:    [[TMP4:%.*]] = getelementptr inbounds i64, ptr [[TMP2]], i32 0
+; CHECK-NEXT:    [[TMP5:%.*]] = getelementptr inbounds i64, ptr [[TMP2]], i32 2
+; CHECK-NEXT:    [[WIDE_LOAD:%.*]] = load <2 x i64>, ptr [[TMP4]], align 8
+; CHECK-NEXT:    [[WIDE_LOAD1]] = load <2 x i64>, ptr [[TMP5]], align 8
+; CHECK-NEXT:    [[TMP6:%.*]] = shufflevector <2 x i64> [[VECTOR_RECUR]], <2 x i64> [[WIDE_LOAD]], <2 x i32> <i32 1, i32 2>
+; CHECK-NEXT:    [[TMP7:%.*]] = shufflevector <2 x i64> [[WIDE_LOAD]], <2 x i64> [[WIDE_LOAD1]], <2 x i32> <i32 1, i32 2>
+; CHECK-NEXT:    [[TMP8:%.*]] = call <2 x i64> @llvm.fshl.v2i64(<2 x i64> <i64 1, i64 1>, <2 x i64> [[TMP6]], <2 x i64> <i64 1, i64 1>)
+; CHECK-NEXT:    [[TMP9:%.*]] = call <2 x i64> @llvm.fshl.v2i64(<2 x i64> <i64 1, i64 1>, <2 x i64> [[TMP7]], <2 x i64> <i64 1, i64 1>)
+; CHECK-NEXT:    [[TMP10:%.*]] = getelementptr inbounds i64, ptr [[DST]], i64 [[TMP0]]
+; CHECK-NEXT:    [[TMP11:%.*]] = getelementptr inbounds i64, ptr [[DST]], i64 [[TMP1]]
+; CHECK-NEXT:    [[TMP12:%.*]] = getelementptr inbounds i64, ptr [[TMP10]], i32 0
+; CHECK-NEXT:    [[TMP13:%.*]] = getelementptr inbounds i64, ptr [[TMP10]], i32 2
+; CHECK-NEXT:    store <2 x i64> [[TMP8]], ptr [[TMP12]], align 8
+; CHECK-NEXT:    store <2 x i64> [[TMP9]], ptr [[TMP13]], align 8
+; CHECK-NEXT:    [[INDEX_NEXT]] = add nuw i64 [[INDEX]], 4
+; CHECK-NEXT:    [[TMP14:%.*]] = icmp eq i64 [[INDEX_NEXT]], 100
+; CHECK-NEXT:    br i1 [[TMP14]], label %[[MIDDLE_BLOCK:.*]], label %[[VECTOR_BODY]], !llvm.loop [[LOOP0:![0-9]+]]
+; CHECK:       [[MIDDLE_BLOCK]]:
+; CHECK-NEXT:    [[VECTOR_RECUR_EXTRACT:%.*]] = extractelement <2 x i64> [[WIDE_LOAD1]], i32 1
+; CHECK-NEXT:    br i1 false, label %[[EXIT:.*]], label %[[SCALAR_PH]]
+; CHECK:       [[SCALAR_PH]]:
+; CHECK-NEXT:    [[BC_RESUME_VAL:%.*]] = phi i64 [ 100, %[[MIDDLE_BLOCK]] ], [ 0, %[[ENTRY]] ]
+; CHECK-NEXT:    [[SCALAR_RECUR_INIT:%.*]] = phi i64 [ [[VECTOR_RECUR_EXTRACT]], %[[MIDDLE_BLOCK]] ], [ 0, %[[ENTRY]] ]
+; CHECK-NEXT:    br label %[[LOOP:.*]]
+; CHECK:       [[LOOP]]:
+; CHECK-NEXT:    [[IV:%.*]] = phi i64 [ [[BC_RESUME_VAL]], %[[SCALAR_PH]] ], [ [[IV_NEXT:%.*]], %[[LOOP]] ]
+; CHECK-NEXT:    [[RECUR:%.*]] = phi i64 [ [[SCALAR_RECUR_INIT]], %[[SCALAR_PH]] ], [ [[L:%.*]], %[[LOOP]] ]
+; CHECK-NEXT:    [[GEP_SRC:%.*]] = getelementptr inbounds i64, ptr [[SRC]], i64 [[IV]]
+; CHECK-NEXT:    [[L]] = load i64, ptr [[GEP_SRC]], align 8
+; CHECK-NEXT:    [[OR:%.*]] = tail call i64 @llvm.fshl.i64(i64 1, i64 [[RECUR]], i64 1)
+; CHECK-NEXT:    [[GEP_DST:%.*]] = getelementptr inbounds i64, ptr [[DST]], i64 [[IV]]
+; CHECK-NEXT:    store i64 [[OR]], ptr [[GEP_DST]], align 8
+; CHECK-NEXT:    [[IV_NEXT]] = add i64 [[IV]], 1
+; CHECK-NEXT:    [[EXITCOND_NOT:%.*]] = icmp eq i64 [[IV]], 100
+; CHECK-NEXT:    br i1 [[EXITCOND_NOT]], label %[[EXIT]], label %[[LOOP]], !llvm.loop [[LOOP3:![0-9]+]]
+; CHECK:       [[EXIT]]:
+; CHECK-NEXT:    ret void
+;
+entry:
+  br label %loop
+
+loop:
+  %iv = phi i64 [ 0, %entry ], [ %iv.next, %loop ]
+  %recur = phi i64 [ 0, %entry ], [ %l, %loop ]
+  %gep.src = getelementptr inbounds i64, ptr %src, i64 %iv
+  %l = load i64, ptr %gep.src, align 8
+  %or = tail call i64 @llvm.fshl.i64(i64 1, i64 %recur, i64 1)
+  %gep.dst = getelementptr inbounds i64, ptr %dst, i64 %iv
+  store i64 %or, ptr %gep.dst, align 8
+  %iv.next = add i64 %iv, 1
+  %exitcond.not = icmp eq i64 %iv, 100
+  br i1 %exitcond.not, label %exit, label %loop
+
+exit:
+  ret void
+}
+
+declare i64 @llvm.fshl.i64(i64, i64, i64)
+;.
+; CHECK: [[LOOP0]] = distinct !{[[LOOP0]], [[META1:![0-9]+]], [[META2:![0-9]+]]}
+; CHECK: [[META1]] = !{!"llvm.loop.isvectorized", i32 1}
+; CHECK: [[META2]] = !{!"llvm.loop.unroll.runtime.disable"}
+; CHECK: [[LOOP3]] = distinct !{[[LOOP3]], [[META2]], [[META1]]}
+;.