[ASAN] Support memory checks on vp.load/store.

[fakepaper56]

Summary

The patch adds new member MaybeEVL into InterestingMemoryOperand to represent
the effective vector length for vp intrinsics. It may be extended for some target intrinsics in the future.

[reames]

llvm/lib/Transforms/Instrumentation/AddressSanitizer.cpp

1467

This looks wrong? What does the bitwidth have to do with the *value* of EVL?

1494

I may be missing something here, but isn't EVL just an upper bound on the loop? If so, maybe we shouldn't use for the ForEachLane variant and should instead have a variant which takes End explicitly?

[fakepaper56]

llvm/lib/Transforms/Instrumentation/AddressSanitizer.cpp

1494

Agree. I am too conservative to change the original code.

[fakepaper56]

Create a new interface SplitBlockAndInsertForLanes, similar as
SplitBlockAndInsertForEachLane but have input argument to control end index.

[fakepaper56]

Rebase and ping.

[reames]

llvm/lib/Transforms/Instrumentation/AddressSanitizer.cpp

1491

I believe this case should be a trunc.

1497

I don't believe there's anything preventing AVL from being 0 here right? If so, this code is wrong. The routine you're calling has a (documented!) assumption that End > 0. To reuse, you need to add the loop guard and remove that assumption.

llvm/lib/Transforms/Utils/BasicBlockUtils.cpp

1641

It's this call where a dynamically zero End becomes problematic.

[fakepaper56]

llvm/lib/Transforms/Instrumentation/AddressSanitizer.cpp

1497

Thank you for pointing my mistake. I am not positive to remove End > 0 assumption from SplitBlockAndInsertSimpleForLoop because it is used in MemorySanitizerVisitor::paintOrigin and there should not a loop guard to test End.

[fakepaper56]

Fix mistakes that Philip pointed.

[fakepaper56]

Rebase and fix test fails.

[fakepaper56]

Rebase and ping.

[fakepaper56]

Rebase and ping.

[fakepaper56]

This update does,

1. Only check whether EVL is zero when EVL is not nullptr. It helps not generate redundant zero check for non-EVL InterestingMemoryOperand.
2. Clean up my patch.

I am sorry for that the quality of the patch was terrible. I guess it was due to
that I was too rushed to land the revision before my military educational
recall. It's reasonable that no one wanted to review the terrible patch.
I promise that I will improve the quality of my succeeding patches in LLVM.

[craig.topper]

llvm/lib/Transforms/Instrumentation/AddressSanitizer.cpp

1477

Can we use IB.CreateZExtOrTrunc(EVL, IntptrTy) and remove the ifs?

1481

extacting -> extracting

llvm/lib/Transforms/Utils/BasicBlockUtils.cpp

1670

Why do we need to keep setting the insertion point every iteration? Will that cause the last index to be checked first?

Is there a test case for constant evl. I didn't see one.

[fakepaper56]

llvm/lib/Transforms/Utils/BasicBlockUtils.cpp

1670

Why do we need to keep setting the insertion point every iteration?

This is to follow the SplitBlockAndInsertForEachLane(ElementCount, Type *, Instruction *, std::function<void(IRBuilderBase&, Value*)>). And the old one is to follow the legacy code.

Will that cause the last index to be checked first?

I think it won't because the letter index will be inserted to the back of the former index.

Is there a test case for constant evl.

Fixed vector masked.load will have a constant evl here. store.v4f32.1110 in llvm/test/Instrumentation/AddressSanitizer/asan-masked-load-store.ll is an example.

[fakepaper56]

Use CreateZExtOrTrunc and fix typos.

[craig.topper]

llvm/lib/Transforms/Utils/BasicBlockUtils.cpp

1670

Why do we need to keep setting the insertion point every iteration?

This is to follow the SplitBlockAndInsertForEachLane(ElementCount, Type *, Instruction *, std::function<void(IRBuilderBase&, Value*)>). And the old one is to follow the legacy code.

Will that cause the last index to be checked first?

I think it won't because the letter index will be inserted to the back of the former index.

You're right. InsertBefore is still after whatever the previous iteration inserted so it didn't really change the insertion point.

Is there a test case for constant evl.

Fixed vector masked.load will have a constant evl here. store.v4f32.1110 in llvm/test/Instrumentation/AddressSanitizer/asan-masked-load-store.ll is an example.

[kito-cheng]

llvm/test/Instrumentation/AddressSanitizer/asan-vp-load-store.ll

30

Could you take a look to see why it use __asan_storeN, not __asan_store4? I saw D145198 can just generate __asan_store4?

[fakepaper56]

llvm/test/Instrumentation/AddressSanitizer/asan-vp-load-store.ll

30

__asan_load4 needs the input address is 4-bytes aligned. I will refine the test to check that.

[fakepaper56]

Update test to test the aligment of pointer.

[kito-cheng]

LGTM

[fakepaper56]

@reames Do you have another concern about the revision? I will land the revision this weekend if the revision does not have comment needed to address. Thank you.

llvm/include/llvm/Transforms/Instrumentation/AddressSanitizerCommon.h

 public:
   Use *PtrUse;
   bool IsWrite;
   Type *OpType;
   TypeSize TypeStoreSize = TypeSize::Fixed(0);
   MaybeAlign Alignment;
   // The mask Value, if we're looking at a masked load/store.
   Value *MaybeMask;
+  // The EVL Value, if we're looking at a vp intrinsic.
+  Value *MaybeEVL;
 
   InterestingMemoryOperand(Instruction *I, unsigned OperandNo, bool IsWrite,
                            class Type *OpType, MaybeAlign Alignment,
-                           Value *MaybeMask = nullptr)
+                           Value *MaybeMask = nullptr,
+                           Value *MaybeEVL = nullptr)
       : IsWrite(IsWrite), OpType(OpType), Alignment(Alignment),
-        MaybeMask(MaybeMask) {
+        MaybeMask(MaybeMask), MaybeEVL(MaybeEVL) {
     const DataLayout &DL = I->getModule()->getDataLayout();
     TypeStoreSize = DL.getTypeStoreSizeInBits(OpType);
     PtrUse = &I->getOperandUse(OperandNo);
   }
 
   Instruction *getInsn() { return cast<Instruction>(PtrUse->getUser()); }
 
   Value *getPtr() { return PtrUse->get(); }

llvm/include/llvm/Transforms/Utils/BasicBlockUtils.h

 /// expected to change in the future. The callback takes as arguments an
 /// IRBuilder whose insert point is correctly set for instantiating the
 /// given index, and a value which is (at runtime) the index to access.
 /// This index *may* be a constant.
 void SplitBlockAndInsertForEachLane(ElementCount EC, Type *IndexTy,
     Instruction *InsertBefore,
     std::function<void(IRBuilderBase&, Value*)> Func);
 
+/// Utility function for performing a given action on each lane of a vector
+/// with \p EVL effective length. EVL is assumed > 0. To simplify porting legacy
+/// code, this defaults to unrolling the implied loop for non-scalable element
+/// counts, but this is not considered to be part of the contract of this
+/// routine, and is expected to change in the future. The callback takes as
+/// arguments an IRBuilder whose insert point is correctly set for instantiating
+/// the given index, and a value which is (at runtime) the index to access. This
+/// index *may* be a constant.
+void SplitBlockAndInsertForEachLane(
+    Value *End, Instruction *InsertBefore,
+    std::function<void(IRBuilderBase &, Value *)> Func);
+
 /// Check whether BB is the merge point of a if-region.
 /// If so, return the branch instruction that determines which entry into
 /// BB will be taken.  Also, return by references the block that will be
 /// entered from if the condition is true, and the block that will be
 /// entered if the condition is false.
 ///
 /// This does no checking to see if the true/false blocks have large or unsavory
 /// instructions in them.

llvm/lib/Transforms/Instrumentation/AddressSanitizer.cpp

AddressSanitizer(Module &M, const StackSafetyGlobalInfo *SSGI,
       : CompileKernel(ClEnableKasan.getNumOccurrences() > 0 ? ClEnableKasan
                                                             : CompileKernel),
         Recover(ClRecover.getNumOccurrences() > 0 ? ClRecover : Recover),
         UseAfterScope(UseAfterScope || ClUseAfterScope),
         UseAfterReturn(ClUseAfterReturn.getNumOccurrences() ? ClUseAfterReturn
                                                             : UseAfterReturn),
         SSGI(SSGI) {
     C = &(M.getContext());
+    DL = &M.getDataLayout();
     LongSize = M.getDataLayout().getPointerSizeInBits();
     IntptrTy = Type::getIntNTy(*C, LongSize);
     Int8PtrTy = Type::getInt8PtrTy(*C);
     Int32Ty = Type::getInt32Ty(*C);
     TargetTriple = Triple(M.getTargetTriple());
 
     Mapping = getShadowMapping(TargetTriple, LongSize, this->CompileKernel);
 
struct FunctionStateRAII {
 
     ~FunctionStateRAII() {
       Pass->LocalDynamicShadow = nullptr;
       Pass->ProcessedAllocas.clear();
     }
   };
 
   LLVMContext *C;
+  const DataLayout *DL;
   Triple TargetTriple;
   int LongSize;
   bool CompileKernel;
   bool Recover;
   bool UseAfterScope;
   AsanDetectStackUseAfterReturnMode UseAfterReturn;
   Type *IntptrTy;
   Type *Int8PtrTy;
Interesting.emplace_back(I, RMW->getPointerOperandIndex(), true,
                              RMW->getValOperand()->getType(), std::nullopt);
   } else if (AtomicCmpXchgInst *XCHG = dyn_cast<AtomicCmpXchgInst>(I)) {
     if (!ClInstrumentAtomics || ignoreAccess(I, XCHG->getPointerOperand()))
       return;
     Interesting.emplace_back(I, XCHG->getPointerOperandIndex(), true,
                              XCHG->getCompareOperand()->getType(),
                              std::nullopt);
   } else if (auto CI = dyn_cast<CallInst>(I)) {
-    if (CI->getIntrinsicID() == Intrinsic::masked_load ||
-        CI->getIntrinsicID() == Intrinsic::masked_store) {
+    switch (CI->getIntrinsicID()) {
+    case Intrinsic::masked_load:
+    case Intrinsic::masked_store: {
       bool IsWrite = CI->getIntrinsicID() == Intrinsic::masked_store;
       // Masked store has an initial operand for the value.
       unsigned OpOffset = IsWrite ? 1 : 0;
       if (IsWrite ? !ClInstrumentWrites : !ClInstrumentReads)
         return;
 
       auto BasePtr = CI->getOperand(OpOffset);
       if (ignoreAccess(I, BasePtr))
         return;
       Type *Ty = IsWrite ? CI->getArgOperand(0)->getType() : CI->getType();
       MaybeAlign Alignment = Align(1);
       // Otherwise no alignment guarantees. We probably got Undef.
       if (auto *Op = dyn_cast<ConstantInt>(CI->getOperand(1 + OpOffset)))
         Alignment = Op->getMaybeAlignValue();
       Value *Mask = CI->getOperand(2 + OpOffset);
       Interesting.emplace_back(I, OpOffset, IsWrite, Ty, Alignment, Mask);
-    } else {
+      break;
+    }
+    case Intrinsic::vp_load:
+    case Intrinsic::vp_store: {
+      auto *VPI = cast<VPIntrinsic>(CI);
+      unsigned IID = CI->getIntrinsicID();
+      bool IsWrite = IID == Intrinsic::vp_store;
+      if (IsWrite ? !ClInstrumentWrites : !ClInstrumentReads)
+        return;
+      unsigned PtrOpNo = *VPI->getMemoryPointerParamPos(IID);
+      Type *Ty = IsWrite ? CI->getArgOperand(0)->getType() : CI->getType();
+      MaybeAlign Alignment = VPI->getOperand(PtrOpNo)->getPointerAlignment(*DL);
+      Interesting.emplace_back(I, PtrOpNo, IsWrite, Ty, Alignment,
+                               VPI->getMaskParam(),
+                               VPI->getVectorLengthParam());
+      break;
+    }
+    default:
       for (unsigned ArgNo = 0; ArgNo < CI->arg_size(); ArgNo++) {
         if (!ClInstrumentByval || !CI->isByValArgument(ArgNo) ||
             ignoreAccess(I, CI->getArgOperand(ArgNo)))
           continue;
         Type *Ty = CI->getParamByValType(ArgNo);
         Interesting.emplace_back(I, ArgNo, false, Ty, Align(1));
       }
     }
if (!TypeStoreSize.isScalable()) {
     }
   }
   Pass->instrumentUnusualSizeOrAlignment(I, InsertBefore, Addr, TypeStoreSize,
                                          IsWrite, nullptr, UseCalls, Exp);
 }
 
 static void instrumentMaskedLoadOrStore(AddressSanitizer *Pass,
                                         const DataLayout &DL, Type *IntptrTy,
-                                        Value *Mask, Instruction *I,
+                                        Value *Mask, Value *EVL, Instruction *I,
                                         Value *Addr, MaybeAlign Alignment,
                                         unsigned Granularity, Type *OpType,
                                         bool IsWrite, Value *SizeArgument,
                                         bool UseCalls, uint32_t Exp) {
   auto *VTy = cast<VectorType>(OpType);
-
   TypeSize ElemTypeSize = DL.getTypeStoreSizeInBits(VTy->getScalarType());
   auto Zero = ConstantInt::get(IntptrTy, 0);
 
-  SplitBlockAndInsertForEachLane(VTy->getElementCount(), IntptrTy, I,
+  IRBuilder IB(I);
+  Instruction *LoopInsertBefore = I;
+  if (EVL) {
+    // The end argument of SplitBlockAndInsertForLane is assumed bigger
+    // than zero, so we should check whether EVL is zero here.
+    Type *EVLType = EVL->getType();
+    Value *IsEVLZero = IB.CreateICmpNE(EVL, ConstantInt::get(EVLType, 0));
+    LoopInsertBefore = SplitBlockAndInsertIfThen(IsEVLZero, I, false);
+    IB.SetInsertPoint(LoopInsertBefore);
+    // Cast EVL to IntptrTy.
+    EVL = IB.CreateZExtOrTrunc(EVL, IntptrTy);
+    // To avoid undefined behavior for extracting with out of range index, use
+    // the minimum of evl and element count as trip count.
+    Value *EC = IB.CreateElementCount(IntptrTy, VTy->getElementCount());
+    EVL = IB.CreateBinaryIntrinsic(Intrinsic::umin, EVL, EC);
+  } else {
+    EVL = IB.CreateElementCount(IntptrTy, VTy->getElementCount());
+  }
+
+  SplitBlockAndInsertForEachLane(EVL, LoopInsertBefore,
                                  [&](IRBuilderBase &IRB, Value *Index) {
     Value *MaskElem = IRB.CreateExtractElement(Mask, Index);
     if (auto *MaskElemC = dyn_cast<ConstantInt>(MaskElem)) {
       if (MaskElemC->isZero())
         // No check
         return;
       // Unconditional check
     } else {
       // Conditional check
-      Instruction *ThenTerm = SplitBlockAndInsertIfThen(MaskElem, &*IRB.GetInsertPoint(), false);
+      Instruction *ThenTerm = SplitBlockAndInsertIfThen(
+          MaskElem, &*IRB.GetInsertPoint(), false);
       IRB.SetInsertPoint(ThenTerm);
     }
 
     Value *InstrumentedAddress = IRB.CreateGEP(VTy, Addr, {Zero, Index});
-    doInstrumentAddress(Pass, I, &*IRB.GetInsertPoint(), InstrumentedAddress, Alignment,
-                        Granularity, ElemTypeSize, IsWrite, SizeArgument,
-                        UseCalls, Exp);
+    doInstrumentAddress(Pass, I, &*IRB.GetInsertPoint(),
+                        InstrumentedAddress, Alignment, Granularity,
+                        ElemTypeSize, IsWrite, SizeArgument, UseCalls, Exp);
   });
 }
 
 void AddressSanitizer::instrumentMop(ObjectSizeOffsetVisitor &ObjSizeVis,
                                      InterestingMemoryOperand &O, bool UseCalls,
                                      const DataLayout &DL) {
   Value *Addr = O.getPtr();
 
void AddressSanitizer::instrumentMop(ObjectSizeOffsetVisitor &ObjSizeVis,
 
   if (O.IsWrite)
     NumInstrumentedWrites++;
   else
     NumInstrumentedReads++;
 
   unsigned Granularity = 1 << Mapping.Scale;
   if (O.MaybeMask) {
-    instrumentMaskedLoadOrStore(this, DL, IntptrTy, O.MaybeMask, O.getInsn(),
-                                Addr, O.Alignment, Granularity, O.OpType,
-                                O.IsWrite, nullptr, UseCalls, Exp);
+    instrumentMaskedLoadOrStore(this, DL, IntptrTy, O.MaybeMask, O.MaybeEVL,
+                                O.getInsn(), Addr, O.Alignment, Granularity,
+                                O.OpType, O.IsWrite, nullptr, UseCalls, Exp);
   } else {
     doInstrumentAddress(this, O.getInsn(), O.getInsn(), Addr, O.Alignment,
                         Granularity, O.TypeStoreSize, O.IsWrite, nullptr, UseCalls,
                         Exp);
   }
 }
 
 Instruction *AddressSanitizer::generateCrashCode(Instruction *InsertBefore,

llvm/lib/Transforms/Utils/BasicBlockUtils.cpp

void llvm::SplitBlockAndInsertForEachLane(ElementCount EC,
 
   IRBuilder<> IRB(InsertBefore);
 
   if (EC.isScalable()) {
     Value *NumElements = IRB.CreateElementCount(IndexTy, EC);
 
     auto [BodyIP, Index] =
       SplitBlockAndInsertSimpleForLoop(NumElements, InsertBefore);
 
     IRB.SetInsertPoint(BodyIP);
     Func(IRB, Index);
     return;
   }
 
   unsigned Num = EC.getFixedValue();
   for (unsigned Idx = 0; Idx < Num; ++Idx) {
     IRB.SetInsertPoint(InsertBefore);
     Func(IRB, ConstantInt::get(IndexTy, Idx));
   }
 }
 
+void llvm::SplitBlockAndInsertForEachLane(
+    Value *EVL, Instruction *InsertBefore,
+    std::function<void(IRBuilderBase &, Value *)> Func) {
+
+  IRBuilder<> IRB(InsertBefore);
+  Type *Ty = EVL->getType();
+
+  if (!isa<ConstantInt>(EVL)) {
+    auto [BodyIP, Index] = SplitBlockAndInsertSimpleForLoop(EVL, InsertBefore);
+    IRB.SetInsertPoint(BodyIP);
+    Func(IRB, Index);
+    return;
+  }
+
+  unsigned Num = cast<ConstantInt>(EVL)->getZExtValue();
+  for (unsigned Idx = 0; Idx < Num; ++Idx) {
+    IRB.SetInsertPoint(InsertBefore);
+    Func(IRB, ConstantInt::get(Ty, Idx));
+  }
+}
+
 BranchInst *llvm::GetIfCondition(BasicBlock *BB, BasicBlock *&IfTrue,
                                  BasicBlock *&IfFalse) {
   PHINode *SomePHI = dyn_cast<PHINode>(BB->begin());
   BasicBlock *Pred1 = nullptr;
   BasicBlock *Pred2 = nullptr;
 
   if (SomePHI) {
     if (SomePHI->getNumIncomingValues() != 2)

llvm/test/Instrumentation/AddressSanitizer/asan-vp-load-store.ll

+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
+; RUN: opt < %s -passes=asan -asan-instrumentation-with-call-threshold=0 -S \
+; RUN:     | FileCheck %s
+; RUN: opt < %s -passes=asan -asan-instrumentation-with-call-threshold=0 -asan-instrument-reads=0 -asan-instrument-writes=0 -S \
+; RUN:     | FileCheck %s -check-prefix=DISABLED
+
+; Support ASan instrumentation for constant-mask llvm.vp.{load,store}
+
+target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"
+
+;;;;;;;;;;;;;;;; STORE
+declare void @llvm.vp.store.v4f32.p0(<4 x float>, ptr, <4 x i1>, i32) argmemonly nounwind
+
+define void @store.v4f32.variable(ptr align 4 %p, <4 x float> %arg, <4 x i1> %mask, i32 %evl) sanitize_address {
+; CHECK-LABEL: @store.v4f32.variable(
+; CHECK-NEXT:    [[TMP1:%.*]] = icmp ne i32 [[EVL:%.*]], 0
+; CHECK-NEXT:    br i1 [[TMP1]], label [[TMP2:%.*]], label [[TMP10:%.*]]
+; CHECK:       2:
+; CHECK-NEXT:    [[TMP3:%.*]] = zext i32 [[EVL]] to i64
+; CHECK-NEXT:    [[TMP4:%.*]] = call i64 @llvm.umin.i64(i64 [[TMP3]], i64 4)
+; CHECK-NEXT:    br label [[DOTSPLIT:%.*]]
+; CHECK:       .split:
+; CHECK-NEXT:    [[IV:%.*]] = phi i64 [ 0, [[TMP2]] ], [ [[IV_NEXT:%.*]], [[TMP9:%.*]] ]
+; CHECK-NEXT:    [[TMP5:%.*]] = extractelement <4 x i1> [[MASK:%.*]], i64 [[IV]]
+; CHECK-NEXT:    br i1 [[TMP5]], label [[TMP6:%.*]], label [[TMP9]]
+; CHECK:       6:
+; CHECK-NEXT:    [[TMP7:%.*]] = getelementptr <4 x float>, ptr [[P:%.*]], i64 0, i64 [[IV]]
+; CHECK-NEXT:    [[TMP8:%.*]] = ptrtoint ptr [[TMP7]] to i64
+; CHECK-NEXT:    call void @__asan_store4(i64 [[TMP8]])
+; CHECK-NEXT:    br label [[TMP9]]
+; CHECK:       9:
+; CHECK-NEXT:    [[IV_NEXT]] = add nuw nsw i64 [[IV]], 1
+; CHECK-NEXT:    [[IV_CHECK:%.*]] = icmp eq i64 [[IV_NEXT]], [[TMP4]]
+; CHECK-NEXT:    br i1 [[IV_CHECK]], label [[DOTSPLIT_SPLIT:%.*]], label [[DOTSPLIT]]
+; CHECK:       .split.split:
+; CHECK-NEXT:    br label [[TMP10]]
+; CHECK:       10:
+; CHECK-NEXT:    tail call void @llvm.vp.store.v4f32.p0(<4 x float> [[ARG:%.*]], ptr [[P]], <4 x i1> [[MASK]], i32 [[EVL]])
+; CHECK-NEXT:    ret void
+;
+; DISABLED-LABEL: @store.v4f32.variable(
+; DISABLED-NEXT:    tail call void @llvm.vp.store.v4f32.p0(<4 x float> [[ARG:%.*]], ptr [[P:%.*]], <4 x i1> [[MASK:%.*]], i32 [[EVL:%.*]])
+; DISABLED-NEXT:    ret void
+;
+  tail call void @llvm.vp.store.v4f32.p0(<4 x float> %arg, ptr %p, <4 x i1> %mask, i32 %evl)
+  ret void
+}
+
+;; Store using two vp.stores, which should instrument them both.
+define void @store.v4f32.1010.split(ptr align 4 %p, <4 x float> %arg, i32 %evl) sanitize_address {
+; CHECK-LABEL: @store.v4f32.1010.split(
+; CHECK-NEXT:    [[TMP1:%.*]] = icmp ne i32 [[EVL:%.*]], 0
+; CHECK-NEXT:    br i1 [[TMP1]], label [[TMP2:%.*]], label [[TMP10:%.*]]
+; CHECK:       2:
+; CHECK-NEXT:    [[TMP3:%.*]] = zext i32 [[EVL]] to i64
+; CHECK-NEXT:    [[TMP4:%.*]] = call i64 @llvm.umin.i64(i64 [[TMP3]], i64 4)
+; CHECK-NEXT:    br label [[DOTSPLIT:%.*]]
+; CHECK:       .split:
+; CHECK-NEXT:    [[IV:%.*]] = phi i64 [ 0, [[TMP2]] ], [ [[IV_NEXT:%.*]], [[TMP9:%.*]] ]
+; CHECK-NEXT:    [[TMP5:%.*]] = extractelement <4 x i1> <i1 false, i1 false, i1 true, i1 true>, i64 [[IV]]
+; CHECK-NEXT:    br i1 [[TMP5]], label [[TMP6:%.*]], label [[TMP9]]
+; CHECK:       6:
+; CHECK-NEXT:    [[TMP7:%.*]] = getelementptr <4 x float>, ptr [[P:%.*]], i64 0, i64 [[IV]]
+; CHECK-NEXT:    [[TMP8:%.*]] = ptrtoint ptr [[TMP7]] to i64
+; CHECK-NEXT:    call void @__asan_store4(i64 [[TMP8]])
+; CHECK-NEXT:    br label [[TMP9]]
+; CHECK:       9:
+; CHECK-NEXT:    [[IV_NEXT]] = add nuw nsw i64 [[IV]], 1
+; CHECK-NEXT:    [[IV_CHECK:%.*]] = icmp eq i64 [[IV_NEXT]], [[TMP4]]
+; CHECK-NEXT:    br i1 [[IV_CHECK]], label [[DOTSPLIT_SPLIT:%.*]], label [[DOTSPLIT]]
+; CHECK:       .split.split:
+; CHECK-NEXT:    br label [[TMP10]]
+; CHECK:       10:
+; CHECK-NEXT:    tail call void @llvm.vp.store.v4f32.p0(<4 x float> [[ARG:%.*]], ptr [[P]], <4 x i1> <i1 false, i1 false, i1 true, i1 true>, i32 [[EVL]])
+; CHECK-NEXT:    [[TMP11:%.*]] = icmp ne i32 [[EVL]], 0
+; CHECK-NEXT:    br i1 [[TMP11]], label [[TMP12:%.*]], label [[TMP20:%.*]]
+; CHECK:       12:
+; CHECK-NEXT:    [[TMP13:%.*]] = zext i32 [[EVL]] to i64
+; CHECK-NEXT:    [[TMP14:%.*]] = call i64 @llvm.umin.i64(i64 [[TMP13]], i64 4)
+; CHECK-NEXT:    br label [[DOTSPLIT1:%.*]]
+; CHECK:       .split1:
+; CHECK-NEXT:    [[IV2:%.*]] = phi i64 [ 0, [[TMP12]] ], [ [[IV2_NEXT:%.*]], [[TMP19:%.*]] ]
+; CHECK-NEXT:    [[TMP15:%.*]] = extractelement <4 x i1> <i1 false, i1 false, i1 true, i1 false>, i64 [[IV2]]
+; CHECK-NEXT:    br i1 [[TMP15]], label [[TMP16:%.*]], label [[TMP19]]
+; CHECK:       16:
+; CHECK-NEXT:    [[TMP17:%.*]] = getelementptr <4 x float>, ptr [[P]], i64 0, i64 [[IV2]]
+; CHECK-NEXT:    [[TMP18:%.*]] = ptrtoint ptr [[TMP17]] to i64
+; CHECK-NEXT:    call void @__asan_store4(i64 [[TMP18]])
+; CHECK-NEXT:    br label [[TMP19]]
+; CHECK:       19:
+; CHECK-NEXT:    [[IV2_NEXT]] = add nuw nsw i64 [[IV2]], 1
+; CHECK-NEXT:    [[IV2_CHECK:%.*]] = icmp eq i64 [[IV2_NEXT]], [[TMP14]]
+; CHECK-NEXT:    br i1 [[IV2_CHECK]], label [[DOTSPLIT1_SPLIT:%.*]], label [[DOTSPLIT1]]
+; CHECK:       .split1.split:
+; CHECK-NEXT:    br label [[TMP20]]
+; CHECK:       20:
+; CHECK-NEXT:    tail call void @llvm.vp.store.v4f32.p0(<4 x float> [[ARG]], ptr [[P]], <4 x i1> <i1 false, i1 false, i1 true, i1 false>, i32 [[EVL]])
+; CHECK-NEXT:    ret void
+;
+; DISABLED-LABEL: @store.v4f32.1010.split(
+; DISABLED-NEXT:    tail call void @llvm.vp.store.v4f32.p0(<4 x float> [[ARG:%.*]], ptr [[P:%.*]], <4 x i1> <i1 false, i1 false, i1 true, i1 true>, i32 [[EVL:%.*]])
+; DISABLED-NEXT:    tail call void @llvm.vp.store.v4f32.p0(<4 x float> [[ARG]], ptr [[P]], <4 x i1> <i1 false, i1 false, i1 true, i1 false>, i32 [[EVL]])
+; DISABLED-NEXT:    ret void
+;
+  tail call void @llvm.vp.store.v4f32.p0(<4 x float> %arg, ptr %p, <4 x i1> <i1 false, i1 false, i1 true, i1 true>, i32 %evl)
+  tail call void @llvm.vp.store.v4f32.p0(<4 x float> %arg, ptr %p, <4 x i1> <i1 false, i1 false, i1 true, i1 false>, i32 %evl)
+  ret void
+}
+
+;; Store using a vp.store after a full store. Shouldn't instrument the second one.
+define void @store.v4f32.0010.after.full.store(ptr align 4 %p, <4 x float> %arg, i32 %evl) sanitize_address {
+; CHECK-LABEL: @store.v4f32.0010.after.full.store(
+; CHECK-NEXT:    [[TMP1:%.*]] = ptrtoint ptr [[P:%.*]] to i64
+; CHECK-NEXT:    call void @__asan_store16(i64 [[TMP1]])
+; CHECK-NEXT:    store <4 x float> [[ARG:%.*]], ptr [[P]], align 16
+; CHECK-NEXT:    tail call void @llvm.vp.store.v4f32.p0(<4 x float> [[ARG]], ptr [[P]], <4 x i1> <i1 false, i1 false, i1 true, i1 false>, i32 [[EVL:%.*]])
+; CHECK-NEXT:    ret void
+;
+; DISABLED-LABEL: @store.v4f32.0010.after.full.store(
+; DISABLED-NEXT:    store <4 x float> [[ARG:%.*]], ptr [[P:%.*]], align 16
+; DISABLED-NEXT:    tail call void @llvm.vp.store.v4f32.p0(<4 x float> [[ARG]], ptr [[P]], <4 x i1> <i1 false, i1 false, i1 true, i1 false>, i32 [[EVL:%.*]])
+; DISABLED-NEXT:    ret void
+;
+  store <4 x float> %arg, ptr %p
+  tail call void @llvm.vp.store.v4f32.p0(<4 x float> %arg, ptr %p, <4 x i1> <i1 false, i1 false, i1 true, i1 false>, i32 %evl)
+  ret void
+}
+
+;;;;;;;;;;;;;;;; LOAD
+declare <4 x float> @llvm.vp.load.v4f32.p0(ptr, <4 x i1>, i32) argmemonly nounwind
+declare <8 x i32> @llvm.vp.load.v8i32.p0(ptr, <8 x i1>, i32) argmemonly nounwind
+
+define <4 x float> @load.v4f32.variable(ptr align 4 %p, <4 x float> %arg, <4 x i1> %mask, i32 %evl) sanitize_address {
+; CHECK-LABEL: @load.v4f32.variable(
+; CHECK-NEXT:    [[TMP1:%.*]] = icmp ne i32 [[EVL:%.*]], 0
+; CHECK-NEXT:    br i1 [[TMP1]], label [[TMP2:%.*]], label [[TMP10:%.*]]
+; CHECK:       2:
+; CHECK-NEXT:    [[TMP3:%.*]] = zext i32 [[EVL]] to i64
+; CHECK-NEXT:    [[TMP4:%.*]] = call i64 @llvm.umin.i64(i64 [[TMP3]], i64 4)
+; CHECK-NEXT:    br label [[DOTSPLIT:%.*]]
+; CHECK:       .split:
+; CHECK-NEXT:    [[IV:%.*]] = phi i64 [ 0, [[TMP2]] ], [ [[IV_NEXT:%.*]], [[TMP9:%.*]] ]
+; CHECK-NEXT:    [[TMP5:%.*]] = extractelement <4 x i1> [[MASK:%.*]], i64 [[IV]]
+; CHECK-NEXT:    br i1 [[TMP5]], label [[TMP6:%.*]], label [[TMP9]]
+; CHECK:       6:
+; CHECK-NEXT:    [[TMP7:%.*]] = getelementptr <4 x float>, ptr [[P:%.*]], i64 0, i64 [[IV]]
+; CHECK-NEXT:    [[TMP8:%.*]] = ptrtoint ptr [[TMP7]] to i64
+; CHECK-NEXT:    call void @__asan_load4(i64 [[TMP8]])
+; CHECK-NEXT:    br label [[TMP9]]
+; CHECK:       9:
+; CHECK-NEXT:    [[IV_NEXT]] = add nuw nsw i64 [[IV]], 1
+; CHECK-NEXT:    [[IV_CHECK:%.*]] = icmp eq i64 [[IV_NEXT]], [[TMP4]]
+; CHECK-NEXT:    br i1 [[IV_CHECK]], label [[DOTSPLIT_SPLIT:%.*]], label [[DOTSPLIT]]
+; CHECK:       .split.split:
+; CHECK-NEXT:    br label [[TMP10]]
+; CHECK:       10:
+; CHECK-NEXT:    [[RES:%.*]] = tail call <4 x float> @llvm.vp.load.v4f32.p0(ptr [[P]], <4 x i1> [[MASK]], i32 [[EVL]])
+; CHECK-NEXT:    ret <4 x float> [[RES]]
+;
+; DISABLED-LABEL: @load.v4f32.variable(
+; DISABLED-NEXT:    [[RES:%.*]] = tail call <4 x float> @llvm.vp.load.v4f32.p0(ptr [[P:%.*]], <4 x i1> [[MASK:%.*]], i32 [[EVL:%.*]])
+; DISABLED-NEXT:    ret <4 x float> [[RES]]
+;
+  %res = tail call <4 x float> @llvm.vp.load.v4f32.p0(ptr %p, <4 x i1> %mask, i32 %evl)
+  ret <4 x float> %res
+}
+
+;; Load using two vp.loads, which should instrument them both.
+define <4 x float> @load.v4f32.1001.split(ptr align 4 %p, i32 %evl) sanitize_address {
+; CHECK-LABEL: @load.v4f32.1001.split(
+; CHECK-NEXT:    [[TMP1:%.*]] = icmp ne i32 [[EVL:%.*]], 0
+; CHECK-NEXT:    br i1 [[TMP1]], label [[TMP2:%.*]], label [[TMP10:%.*]]
+; CHECK:       2:
+; CHECK-NEXT:    [[TMP3:%.*]] = zext i32 [[EVL]] to i64
+; CHECK-NEXT:    [[TMP4:%.*]] = call i64 @llvm.umin.i64(i64 [[TMP3]], i64 4)
+; CHECK-NEXT:    br label [[DOTSPLIT:%.*]]
+; CHECK:       .split:
+; CHECK-NEXT:    [[IV:%.*]] = phi i64 [ 0, [[TMP2]] ], [ [[IV_NEXT:%.*]], [[TMP9:%.*]] ]
+; CHECK-NEXT:    [[TMP5:%.*]] = extractelement <4 x i1> <i1 true, i1 false, i1 false, i1 false>, i64 [[IV]]
+; CHECK-NEXT:    br i1 [[TMP5]], label [[TMP6:%.*]], label [[TMP9]]
+; CHECK:       6:
+; CHECK-NEXT:    [[TMP7:%.*]] = getelementptr <4 x float>, ptr [[P:%.*]], i64 0, i64 [[IV]]
+; CHECK-NEXT:    [[TMP8:%.*]] = ptrtoint ptr [[TMP7]] to i64
+; CHECK-NEXT:    call void @__asan_load4(i64 [[TMP8]])
+; CHECK-NEXT:    br label [[TMP9]]
+; CHECK:       9:
+; CHECK-NEXT:    [[IV_NEXT]] = add nuw nsw i64 [[IV]], 1
+; CHECK-NEXT:    [[IV_CHECK:%.*]] = icmp eq i64 [[IV_NEXT]], [[TMP4]]
+; CHECK-NEXT:    br i1 [[IV_CHECK]], label [[DOTSPLIT_SPLIT:%.*]], label [[DOTSPLIT]]
+; CHECK:       .split.split:
+; CHECK-NEXT:    br label [[TMP10]]
+; CHECK:       10:
+; CHECK-NEXT:    [[RES:%.*]] = tail call <4 x float> @llvm.vp.load.v4f32.p0(ptr [[P]], <4 x i1> <i1 true, i1 false, i1 false, i1 false>, i32 [[EVL]])
+; CHECK-NEXT:    [[TMP11:%.*]] = icmp ne i32 [[EVL]], 0
+; CHECK-NEXT:    br i1 [[TMP11]], label [[TMP12:%.*]], label [[TMP20:%.*]]
+; CHECK:       12:
+; CHECK-NEXT:    [[TMP13:%.*]] = zext i32 [[EVL]] to i64
+; CHECK-NEXT:    [[TMP14:%.*]] = call i64 @llvm.umin.i64(i64 [[TMP13]], i64 4)
+; CHECK-NEXT:    br label [[DOTSPLIT1:%.*]]
+; CHECK:       .split1:
+; CHECK-NEXT:    [[IV2:%.*]] = phi i64 [ 0, [[TMP12]] ], [ [[IV2_NEXT:%.*]], [[TMP19:%.*]] ]
+; CHECK-NEXT:    [[TMP15:%.*]] = extractelement <4 x i1> <i1 false, i1 false, i1 false, i1 true>, i64 [[IV2]]
+; CHECK-NEXT:    br i1 [[TMP15]], label [[TMP16:%.*]], label [[TMP19]]
+; CHECK:       16:
+; CHECK-NEXT:    [[TMP17:%.*]] = getelementptr <4 x float>, ptr [[P]], i64 0, i64 [[IV2]]
+; CHECK-NEXT:    [[TMP18:%.*]] = ptrtoint ptr [[TMP17]] to i64
+; CHECK-NEXT:    call void @__asan_load4(i64 [[TMP18]])
+; CHECK-NEXT:    br label [[TMP19]]
+; CHECK:       19:
+; CHECK-NEXT:    [[IV2_NEXT]] = add nuw nsw i64 [[IV2]], 1
+; CHECK-NEXT:    [[IV2_CHECK:%.*]] = icmp eq i64 [[IV2_NEXT]], [[TMP14]]
+; CHECK-NEXT:    br i1 [[IV2_CHECK]], label [[DOTSPLIT1_SPLIT:%.*]], label [[DOTSPLIT1]]
+; CHECK:       .split1.split:
+; CHECK-NEXT:    br label [[TMP20]]
+; CHECK:       20:
+; CHECK-NEXT:    [[RES2:%.*]] = tail call <4 x float> @llvm.vp.load.v4f32.p0(ptr [[P]], <4 x i1> <i1 false, i1 false, i1 false, i1 true>, i32 [[EVL]])
+; CHECK-NEXT:    ret <4 x float> [[RES2]]
+;
+; DISABLED-LABEL: @load.v4f32.1001.split(
+; DISABLED-NEXT:    [[RES:%.*]] = tail call <4 x float> @llvm.vp.load.v4f32.p0(ptr [[P:%.*]], <4 x i1> <i1 true, i1 false, i1 false, i1 false>, i32 [[EVL:%.*]])
+; DISABLED-NEXT:    [[RES2:%.*]] = tail call <4 x float> @llvm.vp.load.v4f32.p0(ptr [[P]], <4 x i1> <i1 false, i1 false, i1 false, i1 true>, i32 [[EVL]])
+; DISABLED-NEXT:    ret <4 x float> [[RES2]]
+;
+  %res = tail call <4 x float> @llvm.vp.load.v4f32.p0(ptr %p, <4 x i1> <i1 true, i1 false, i1 false, i1 false>, i32 %evl)
+  %res2 = tail call <4 x float> @llvm.vp.load.v4f32.p0(ptr %p, <4 x i1> <i1 false, i1 false, i1 false, i1 true>, i32 %evl)
+  ret <4 x float> %res2
+}
+
+;; Load using a vp.load after a full load. Shouldn't instrument the second one.
+define <4 x float> @load.v4f32.1001.after.full.load(ptr align 4 %p, i32 %evl) sanitize_address {
+; CHECK-LABEL: @load.v4f32.1001.after.full.load(
+; CHECK-NEXT:    [[TMP1:%.*]] = ptrtoint ptr [[P:%.*]] to i64
+; CHECK-NEXT:    call void @__asan_load16(i64 [[TMP1]])
+; CHECK-NEXT:    [[RES:%.*]] = load <4 x float>, ptr [[P]], align 16
+; CHECK-NEXT:    [[RES2:%.*]] = tail call <4 x float> @llvm.vp.load.v4f32.p0(ptr [[P]], <4 x i1> <i1 false, i1 false, i1 false, i1 true>, i32 [[EVL:%.*]])
+; CHECK-NEXT:    ret <4 x float> [[RES2]]
+;
+; DISABLED-LABEL: @load.v4f32.1001.after.full.load(
+; DISABLED-NEXT:    [[RES:%.*]] = load <4 x float>, ptr [[P:%.*]], align 16
+; DISABLED-NEXT:    [[RES2:%.*]] = tail call <4 x float> @llvm.vp.load.v4f32.p0(ptr [[P]], <4 x i1> <i1 false, i1 false, i1 false, i1 true>, i32 [[EVL:%.*]])
+; DISABLED-NEXT:    ret <4 x float> [[RES2]]
+;
+  %res = load <4 x float>, ptr %p
+  %res2 = tail call <4 x float> @llvm.vp.load.v4f32.p0(ptr %p, <4 x i1> <i1 false, i1 false, i1 false, i1 true>, i32 %evl)
+  ret <4 x float> %res2
+}
+
+;; Scalable vector tests
+;; ---------------------------
+declare <vscale x 4 x float> @llvm.vp.load.nxv4f32.p0(ptr, <vscale x 4 x i1>, i32)
+declare void @llvm.vp.store.nxv4f32.p0(<vscale x 4 x float>, ptr, <vscale x 4 x i1>, i32)
+
+define <vscale x 4 x float> @scalable.load.nxv4f32(ptr align 4 %p, <vscale x 4 x i1> %mask, i32 %evl) sanitize_address {
+; CHECK-LABEL: @scalable.load.nxv4f32(
+; CHECK-NEXT:    [[TMP1:%.*]] = icmp ne i32 [[EVL:%.*]], 0
+; CHECK-NEXT:    br i1 [[TMP1]], label [[TMP2:%.*]], label [[TMP12:%.*]]
+; CHECK:       2:
+; CHECK-NEXT:    [[TMP3:%.*]] = zext i32 [[EVL]] to i64
+; CHECK-NEXT:    [[TMP4:%.*]] = call i64 @llvm.vscale.i64()
+; CHECK-NEXT:    [[TMP5:%.*]] = mul i64 [[TMP4]], 4
+; CHECK-NEXT:    [[TMP6:%.*]] = call i64 @llvm.umin.i64(i64 [[TMP3]], i64 [[TMP5]])
+; CHECK-NEXT:    br label [[DOTSPLIT:%.*]]
+; CHECK:       .split:
+; CHECK-NEXT:    [[IV:%.*]] = phi i64 [ 0, [[TMP2]] ], [ [[IV_NEXT:%.*]], [[TMP11:%.*]] ]
+; CHECK-NEXT:    [[TMP7:%.*]] = extractelement <vscale x 4 x i1> [[MASK:%.*]], i64 [[IV]]
+; CHECK-NEXT:    br i1 [[TMP7]], label [[TMP8:%.*]], label [[TMP11]]
+; CHECK:       8:
+; CHECK-NEXT:    [[TMP9:%.*]] = getelementptr <vscale x 4 x float>, ptr [[P:%.*]], i64 0, i64 [[IV]]
+; CHECK-NEXT:    [[TMP10:%.*]] = ptrtoint ptr [[TMP9]] to i64
+; CHECK-NEXT:    call void @__asan_load4(i64 [[TMP10]])
+; CHECK-NEXT:    br label [[TMP11]]
+; CHECK:       11:
+; CHECK-NEXT:    [[IV_NEXT]] = add nuw nsw i64 [[IV]], 1
+; CHECK-NEXT:    [[IV_CHECK:%.*]] = icmp eq i64 [[IV_NEXT]], [[TMP6]]
+; CHECK-NEXT:    br i1 [[IV_CHECK]], label [[DOTSPLIT_SPLIT:%.*]], label [[DOTSPLIT]]
+; CHECK:       .split.split:
+; CHECK-NEXT:    br label [[TMP12]]
+; CHECK:       12:
+; CHECK-NEXT:    [[RES:%.*]] = tail call <vscale x 4 x float> @llvm.vp.load.nxv4f32.p0(ptr [[P]], <vscale x 4 x i1> [[MASK]], i32 [[EVL]])
+; CHECK-NEXT:    ret <vscale x 4 x float> [[RES]]
+;
+; DISABLED-LABEL: @scalable.load.nxv4f32(
+; DISABLED-NEXT:    [[RES:%.*]] = tail call <vscale x 4 x float> @llvm.vp.load.nxv4f32.p0(ptr [[P:%.*]], <vscale x 4 x i1> [[MASK:%.*]], i32 [[EVL:%.*]])
+; DISABLED-NEXT:    ret <vscale x 4 x float> [[RES]]
+;
+  %res = tail call <vscale x 4 x float> @llvm.vp.load.nxv4f32.p0(ptr %p, <vscale x 4 x i1> %mask, i32 %evl)
+  ret <vscale x 4 x float> %res
+}
+
+define void @scalable.store.nxv4f32(ptr align 4 %p, <vscale x 4 x float> %arg, <vscale x 4 x i1> %mask, i32 %evl) sanitize_address {
+; CHECK-LABEL: @scalable.store.nxv4f32(
+; CHECK-NEXT:    [[TMP1:%.*]] = icmp ne i32 [[EVL:%.*]], 0
+; CHECK-NEXT:    br i1 [[TMP1]], label [[TMP2:%.*]], label [[TMP12:%.*]]
+; CHECK:       2:
+; CHECK-NEXT:    [[TMP3:%.*]] = zext i32 [[EVL]] to i64
+; CHECK-NEXT:    [[TMP4:%.*]] = call i64 @llvm.vscale.i64()
+; CHECK-NEXT:    [[TMP5:%.*]] = mul i64 [[TMP4]], 4
+; CHECK-NEXT:    [[TMP6:%.*]] = call i64 @llvm.umin.i64(i64 [[TMP3]], i64 [[TMP5]])
+; CHECK-NEXT:    br label [[DOTSPLIT:%.*]]
+; CHECK:       .split:
+; CHECK-NEXT:    [[IV:%.*]] = phi i64 [ 0, [[TMP2]] ], [ [[IV_NEXT:%.*]], [[TMP11:%.*]] ]
+; CHECK-NEXT:    [[TMP7:%.*]] = extractelement <vscale x 4 x i1> [[MASK:%.*]], i64 [[IV]]
+; CHECK-NEXT:    br i1 [[TMP7]], label [[TMP8:%.*]], label [[TMP11]]
+; CHECK:       8:
+; CHECK-NEXT:    [[TMP9:%.*]] = getelementptr <vscale x 4 x float>, ptr [[P:%.*]], i64 0, i64 [[IV]]
+; CHECK-NEXT:    [[TMP10:%.*]] = ptrtoint ptr [[TMP9]] to i64
+; CHECK-NEXT:    call void @__asan_store4(i64 [[TMP10]])
+; CHECK-NEXT:    br label [[TMP11]]
+; CHECK:       11:
+; CHECK-NEXT:    [[IV_NEXT]] = add nuw nsw i64 [[IV]], 1
+; CHECK-NEXT:    [[IV_CHECK:%.*]] = icmp eq i64 [[IV_NEXT]], [[TMP6]]
+; CHECK-NEXT:    br i1 [[IV_CHECK]], label [[DOTSPLIT_SPLIT:%.*]], label [[DOTSPLIT]]
+; CHECK:       .split.split:
+; CHECK-NEXT:    br label [[TMP12]]
+; CHECK:       12:
+; CHECK-NEXT:    tail call void @llvm.vp.store.nxv4f32.p0(<vscale x 4 x float> [[ARG:%.*]], ptr [[P]], <vscale x 4 x i1> [[MASK]], i32 [[EVL]])
+; CHECK-NEXT:    ret void
+;
+; DISABLED-LABEL: @scalable.store.nxv4f32(
+; DISABLED-NEXT:    tail call void @llvm.vp.store.nxv4f32.p0(<vscale x 4 x float> [[ARG:%.*]], ptr [[P:%.*]], <vscale x 4 x i1> [[MASK:%.*]], i32 [[EVL:%.*]])
+; DISABLED-NEXT:    ret void
+;
+  tail call void @llvm.vp.store.nxv4f32.p0(<vscale x 4 x float> %arg, ptr %p, <vscale x 4 x i1> %mask, i32 %evl)
+  ret void
+}