[Matrix] Preserve volatile when loading loads/stores.

[fhahn]

Summary

Currently the matrix lowering turns volatile loads/stores into
non-volatile ones. This patch updates the lowering to preserve the
volatile bit.

[LuoYuanke]

llvm/lib/Transforms/Scalar/LowerMatrixIntrinsics.cpp

1218–1219

Why we set volatile as false? Is it possible that the original load instruction is volatile load?
Do we allow fusion if the load and store instruction is volatile?

[fhahn]

Pass through volatile flag for matrix fusion, thanks @LuoYuanke!

[fhahn]

llvm/lib/Transforms/Scalar/LowerMatrixIntrinsics.cpp

1218–1219

Yes indeed! I wanted to share the code ASAP, fixed now (with additional test case)

[anemet]

LGTM

[fhahn]

Rebase after cleaning up test diffs.

[fhahn]

Rebase after adding another test

[fhahn]

Rebased after parent patches landed upstream. Will commit shortly.

llvm/lib/Transforms/Scalar/LowerMatrixIntrinsics.cpp

for (Instruction *Inst : MatrixInsts) {
       if (CallInst *CInst = dyn_cast<CallInst>(Inst))
         Changed |= VisitCallInst(CInst);
 
       Value *Op1;
       Value *Op2;
       if (auto *BinOp = dyn_cast<BinaryOperator>(Inst))
         Changed |= VisitBinaryOperator(BinOp);
       if (match(Inst, m_Load(m_Value(Op1))))
-        Changed |= VisitLoad(Inst, Op1, Builder);
+        Changed |= VisitLoad(cast<LoadInst>(Inst), Op1, Builder);
       else if (match(Inst, m_Store(m_Value(Op1), m_Value(Op2))))
-        Changed |= VisitStore(Inst, Op1, Op2, Builder);
+        Changed |= VisitStore(cast<StoreInst>(Inst), Op1, Op2, Builder);
     }
 
     RemarkGenerator RemarkGen(Inst2ColumnMatrix, ORE, Func);
     RemarkGen.emitRemarks();
 
     for (Instruction *Inst : reverse(ToRemove))
       Inst->eraseFromParent();
 
     return Changed;
   }
 
-  LoadInst *createVectorLoad(Value *ColumnPtr, Type *EltType,
+  LoadInst *createVectorLoad(Value *ColumnPtr, Type *EltType, bool IsVolatile,
                              IRBuilder<> &Builder) {
-    return Builder.CreateAlignedLoad(
-        ColumnPtr, Align(DL.getABITypeAlignment(EltType)), "col.load");
+    return Builder.CreateAlignedLoad(ColumnPtr,
+                                     Align(DL.getABITypeAlignment(EltType)),
+                                     IsVolatile, "col.load");
   }
 
   StoreInst *createVectorStore(Value *ColumnValue, Value *ColumnPtr,
-                               Type *EltType, IRBuilder<> &Builder) {
+                               Type *EltType, bool IsVolatile,
+                               IRBuilder<> &Builder) {
     return Builder.CreateAlignedStore(ColumnValue, ColumnPtr,
-                                      DL.getABITypeAlign(EltType));
+                                      DL.getABITypeAlign(EltType), IsVolatile);
   }
 
   /// Turns \p BasePtr into an elementwise pointer to \p EltType.
   Value *createElementPtr(Value *BasePtr, Type *EltType, IRBuilder<> &Builder) {
     unsigned AS = cast<PointerType>(BasePtr->getType())->getAddressSpace();
     Type *EltPtrType = PointerType::get(EltType, AS);
     return Builder.CreatePointerCast(BasePtr, EltPtrType);
   }
bool VisitCallInst(CallInst *Inst) {
     default:
       return false;
     }
     return true;
   }
 
   /// Load a matrix with \p Shape starting at \p Ptr and using \p Stride between
   /// vectors.
-  MatrixTy loadMatrix(Type *Ty, Value *Ptr, Value *Stride, ShapeInfo Shape,
-                      IRBuilder<> &Builder) {
+  MatrixTy loadMatrix(Type *Ty, Value *Ptr, Value *Stride, bool IsVolatile,
+                      ShapeInfo Shape, IRBuilder<> &Builder) {
     auto VType = cast<VectorType>(Ty);
     Value *EltPtr = createElementPtr(Ptr, VType->getElementType(), Builder);
     MatrixTy Result;
     for (unsigned I = 0, E = Shape.getNumVectors(); I < E; ++I) {
       Value *GEP = computeVectorAddr(EltPtr, Builder.getInt64(I), Stride,
                                      Shape.getStride(), VType->getElementType(),
                                      Builder);
-      Value *Vector = createVectorLoad(GEP, VType->getElementType(), Builder);
+      Value *Vector =
+          createVectorLoad(GEP, VType->getElementType(), IsVolatile, Builder);
       Result.addVector(Vector);
     }
     return Result.addNumLoads(getNumOps(Result.getVectorTy()) *
                               Result.getNumVectors());
   }
 
   /// Loads a sub-matrix with shape \p ResultShape from a \p R x \p C matrix,
   /// starting at \p MatrixPtr[I][J].
-  MatrixTy loadMatrix(Value *MatrixPtr, ShapeInfo MatrixShape, Value *I,
-                      Value *J, ShapeInfo ResultShape, Type *EltTy,
+  MatrixTy loadMatrix(Value *MatrixPtr, bool IsVolatile, ShapeInfo MatrixShape,
+                      Value *I, Value *J, ShapeInfo ResultShape, Type *EltTy,
                       IRBuilder<> &Builder) {
 
     Value *Offset = Builder.CreateAdd(
         Builder.CreateMul(J, Builder.getInt64(MatrixShape.getStride())), I);
 
     unsigned AS = cast<PointerType>(MatrixPtr->getType())->getAddressSpace();
     Value *EltPtr =
         Builder.CreatePointerCast(MatrixPtr, PointerType::get(EltTy, AS));
     Value *TileStart = Builder.CreateGEP(EltTy, EltPtr, Offset);
     auto *TileTy = FixedVectorType::get(EltTy, ResultShape.NumRows *
                                                    ResultShape.NumColumns);
     Type *TilePtrTy = PointerType::get(TileTy, AS);
     Value *TilePtr =
         Builder.CreatePointerCast(TileStart, TilePtrTy, "col.cast");
 
     return loadMatrix(TileTy, TilePtr,
-                      Builder.getInt64(MatrixShape.getStride()), ResultShape,
-                      Builder);
+                      Builder.getInt64(MatrixShape.getStride()), IsVolatile,
+                      ResultShape, Builder);
   }
 
   /// Lower a load instruction with shape information.
-  void LowerLoad(Instruction *Inst, Value *Ptr, Value *Stride,
+  void LowerLoad(Instruction *Inst, Value *Ptr, Value *Stride, bool IsVolatile,
                  ShapeInfo Shape) {
     IRBuilder<> Builder(Inst);
-    finalizeLowering(Inst,
-                     loadMatrix(Inst->getType(), Ptr, Stride, Shape, Builder),
-                     Builder);
+    finalizeLowering(
+        Inst,
+        loadMatrix(Inst->getType(), Ptr, Stride, IsVolatile, Shape, Builder),
   }
 
   /// Lowers llvm.matrix.column.major.load.
   ///
   /// The intrinsic loads a matrix from memory using a stride between columns.
   void LowerColumnMajorLoad(CallInst *Inst) {
     assert(MatrixLayout == MatrixLayoutTy::ColumnMajor &&
            "Intrinsic only supports column-major layout!");
     Value *Ptr = Inst->getArgOperand(0);
     Value *Stride = Inst->getArgOperand(1);
     LowerLoad(Inst, Ptr, Stride,
+              cast<ConstantInt>(Inst->getArgOperand(2))->isOne(),
               {Inst->getArgOperand(3), Inst->getArgOperand(4)});
   }
 
   /// Stores a sub-matrix \p StoreVal into the \p R x \p C matrix starting at \p
   /// MatrixPtr[I][J].
-  void storeMatrix(const MatrixTy &StoreVal, Value *MatrixPtr,
+  void storeMatrix(const MatrixTy &StoreVal, Value *MatrixPtr, bool IsVolatile,
                    ShapeInfo MatrixShape, Value *I, Value *J, Type *EltTy,
                    IRBuilder<> &Builder) {
     Value *Offset = Builder.CreateAdd(
         Builder.CreateMul(J, Builder.getInt64(MatrixShape.getStride())), I);
 
     unsigned AS = cast<PointerType>(MatrixPtr->getType())->getAddressSpace();
     Value *EltPtr =
         Builder.CreatePointerCast(MatrixPtr, PointerType::get(EltTy, AS));
     Value *TileStart = Builder.CreateGEP(EltTy, EltPtr, Offset);
     auto *TileTy = FixedVectorType::get(EltTy, StoreVal.getNumRows() *
                                                    StoreVal.getNumColumns());
     Type *TilePtrTy = PointerType::get(TileTy, AS);
     Value *TilePtr =
         Builder.CreatePointerCast(TileStart, TilePtrTy, "col.cast");
 
     storeMatrix(TileTy, StoreVal, TilePtr,
-                Builder.getInt64(MatrixShape.getStride()), Builder);
+                Builder.getInt64(MatrixShape.getStride()), IsVolatile, Builder);
   }
 
   /// Store matrix \p StoreVal starting at \p Ptr and using \p Stride between
   /// vectors.
   MatrixTy storeMatrix(Type *Ty, MatrixTy StoreVal, Value *Ptr, Value *Stride,
-                       IRBuilder<> &Builder) {
+                       bool IsVolatile, IRBuilder<> &Builder) {
     auto VType = cast<VectorType>(Ty);
     Value *EltPtr = createElementPtr(Ptr, VType->getElementType(), Builder);
     for (auto Vec : enumerate(StoreVal.vectors())) {
       Value *GEP = computeVectorAddr(EltPtr, Builder.getInt64(Vec.index()),
                                      Stride, StoreVal.getStride(),
                                      VType->getElementType(), Builder);
-      createVectorStore(Vec.value(), GEP, VType->getElementType(), Builder);
+      createVectorStore(Vec.value(), GEP, VType->getElementType(), IsVolatile,
+                        Builder);
     }
     return MatrixTy().addNumStores(getNumOps(StoreVal.getVectorTy()) *
                                    StoreVal.getNumVectors());
   }
 
   /// Lower a store instruction with shape information.
   void LowerStore(Instruction *Inst, Value *Matrix, Value *Ptr, Value *Stride,
-                  ShapeInfo Shape) {
+                  bool IsVolatile, ShapeInfo Shape) {
     IRBuilder<> Builder(Inst);
     auto StoreVal = getMatrix(Matrix, Shape, Builder);
-    finalizeLowering(
-        Inst, storeMatrix(Matrix->getType(), StoreVal, Ptr, Stride, Builder),
+    finalizeLowering(Inst,
+                     storeMatrix(Matrix->getType(), StoreVal, Ptr, Stride,
+                                 IsVolatile, Builder),
                      Builder);
   }
 
   /// Lowers llvm.matrix.column.major.store.
   ///
   /// The intrinsic store a matrix back memory using a stride between columns.
   void LowerColumnMajorStore(CallInst *Inst) {
     assert(MatrixLayout == MatrixLayoutTy::ColumnMajor &&
            "Intrinsic only supports column-major layout!");
     Value *Matrix = Inst->getArgOperand(0);
     Value *Ptr = Inst->getArgOperand(1);
     Value *Stride = Inst->getArgOperand(2);
     LowerStore(Inst, Matrix, Ptr, Stride,
+               cast<ConstantInt>(Inst->getArgOperand(3))->isOne(),
                {Inst->getArgOperand(4), Inst->getArgOperand(5)});
   }
 
   // Set elements I..I+NumElts-1 to Block
   Value *insertVector(Value *Col, unsigned I, Value *Block,
                       IRBuilder<> &Builder) {
 
     // First, bring Block to the same size as Col
void emitSIMDTiling(CallInst *MatMul, LoadInst *LoadOp0, LoadInst *LoadOp1,
     IRBuilder<> Builder(Store);
     for (unsigned J = 0; J < C; J += TileSize)
       for (unsigned I = 0; I < R; I += TileSize) {
         const unsigned TileR = std::min(R - I, unsigned(TileSize));
         const unsigned TileC = std::min(C - J, unsigned(TileSize));
         MatrixTy Res = getZeroMatrix(EltType, TileR, TileC);
 
         for (unsigned K = 0; K < M; K += TileSize) {
           const unsigned TileM = std::min(M - K, unsigned(TileSize));
-          MatrixTy A =
-              loadMatrix(APtr, LShape, Builder.getInt64(I), Builder.getInt64(K),
+          MatrixTy A = loadMatrix(APtr, LoadOp0->isVolatile(), LShape,
+                                  Builder.getInt64(I), Builder.getInt64(K),
                                   {TileR, TileM}, EltType, Builder);
-          MatrixTy B =
-              loadMatrix(BPtr, RShape, Builder.getInt64(K), Builder.getInt64(J),
+          MatrixTy B = loadMatrix(BPtr, LoadOp1->isVolatile(), RShape,
+                                  Builder.getInt64(K), Builder.getInt64(J),
                                   {TileM, TileC}, EltType, Builder);
           emitMatrixMultiply(Res, A, B, AllowContract, Builder, true);
         }
-        storeMatrix(Res, CPtr, {R, M}, Builder.getInt64(I), Builder.getInt64(J),
-                    EltType, Builder);
+        storeMatrix(Res, CPtr, Store->isVolatile(), {R, M}, Builder.getInt64(I),
+                    Builder.getInt64(J), EltType, Builder);
       }
 
     // Mark eliminated instructions as fused and remove them.
     FusedInsts.insert(Store);
     FusedInsts.insert(MatMul);
     Store->eraseFromParent();
     MatMul->eraseFromParent();
     if (LoadOp0->hasNUses(0)) {
void LowerTranspose(CallInst *Inst) {
     // account for later simplifications/combines.
     finalizeLowering(
         Inst,
         Result.addNumComputeOps(2 * ArgShape.NumRows * ArgShape.NumColumns),
         Builder);
   }
 
   /// Lower load instructions, if shape information is available.
-  bool VisitLoad(Instruction *Inst, Value *Ptr, IRBuilder<> &Builder) {
+  bool VisitLoad(LoadInst *Inst, Value *Ptr, IRBuilder<> &Builder) {
     auto I = ShapeMap.find(Inst);
     if (I == ShapeMap.end())
       return false;
 
-    LowerLoad(Inst, Ptr, Builder.getInt64(I->second.getStride()), I->second);
+    LowerLoad(Inst, Ptr, Builder.getInt64(I->second.getStride()),
+              Inst->isVolatile(), I->second);
     return true;
   }
 
-  bool VisitStore(Instruction *Inst, Value *StoredVal, Value *Ptr,
+  bool VisitStore(StoreInst *Inst, Value *StoredVal, Value *Ptr,
                   IRBuilder<> &Builder) {
     auto I = ShapeMap.find(StoredVal);
     if (I == ShapeMap.end())
       return false;
 
     LowerStore(Inst, StoredVal, Ptr, Builder.getInt64(I->second.getStride()),
-               I->second);
+               Inst->isVolatile(), I->second);
     return true;
   }
 
   /// Lower binary operators, if shape information is available.
   bool VisitBinaryOperator(BinaryOperator *Inst) {
     auto I = ShapeMap.find(Inst);
     if (I == ShapeMap.end())
       return false;

llvm/test/Transforms/LowerMatrixIntrinsics/load-align-volatile.ll

 ; RUN: opt -lower-matrix-intrinsics -S < %s | FileCheck %s
 ; RUN: opt -passes='lower-matrix-intrinsics' -S < %s | FileCheck %s
 
 define <9 x double> @strided_load_3x3_volatile(<9 x double>* %in, i64 %stride) {
 ; CHECK-LABEL: @strided_load_3x3_volatile(
 ; CHECK-NEXT:  entry:
 ; CHECK-NEXT:    [[TMP0:%.*]] = bitcast <9 x double>* [[IN:%.*]] to double*
 ; CHECK-NEXT:    [[VEC_START:%.*]] = mul i64 0, [[STRIDE:%.*]]
 ; CHECK-NEXT:    [[VEC_GEP:%.*]] = getelementptr double, double* [[TMP0]], i64 [[VEC_START]]
 ; CHECK-NEXT:    [[VEC_CAST:%.*]] = bitcast double* [[VEC_GEP]] to <3 x double>*
-; CHECK-NEXT:    load <3 x double>, <3 x double>* [[VEC_CAST]], align 8
+; CHECK-NEXT:    load volatile <3 x double>, <3 x double>* [[VEC_CAST]], align 8
 ; CHECK-NEXT:    [[VEC_START1:%.*]] = mul i64 1, [[STRIDE]]
 ; CHECK-NEXT:    [[VEC_GEP2:%.*]] = getelementptr double, double* [[TMP0]], i64 [[VEC_START1]]
 ; CHECK-NEXT:    [[VEC_CAST3:%.*]] = bitcast double* [[VEC_GEP2]] to <3 x double>*
-; CHECK-NEXT:    load <3 x double>, <3 x double>* [[VEC_CAST3]], align 8
+; CHECK-NEXT:    load volatile <3 x double>, <3 x double>* [[VEC_CAST3]], align 8
 ; CHECK-NEXT:    [[VEC_START5:%.*]] = mul i64 2, [[STRIDE]]
 ; CHECK-NEXT:    [[VEC_GEP6:%.*]] = getelementptr double, double* [[TMP0]], i64 [[VEC_START5]]
 ; CHECK-NEXT:    [[VEC_CAST7:%.*]] = bitcast double* [[VEC_GEP6]] to <3 x double>*
-; CHECK-NEXT:    load <3 x double>, <3 x double>* [[VEC_CAST7]], align 8
+; CHECK-NEXT:    load volatile <3 x double>, <3 x double>* [[VEC_CAST7]], align 8
 ; CHECK-NOT:     = load
 ;
 entry:
   %load = call <9 x double> @llvm.matrix.column.major.load(<9 x double>* %in, i64 %stride, i1 true, i64 3, i64 3)
   ret <9 x double> %load
 }
 
 declare <9 x double> @llvm.matrix.column.major.load(<9 x double>*, i64, i1, i64, i64)
 
 define <4 x double> @load_volatile_multiply(<4 x double>* %in) {
 ; CHECK-LABEL: @load_volatile_multiply(
 ; CHECK-NEXT:    [[TMP1:%.*]] = bitcast <4 x double>* [[IN:%.*]] to double*
 ; CHECK-NEXT:    [[VEC_CAST:%.*]] = bitcast double* [[TMP1]] to <2 x double>*
-; CHECK-NEXT:    load <2 x double>, <2 x double>* [[VEC_CAST]], align 8
+; CHECK-NEXT:    load volatile <2 x double>, <2 x double>* [[VEC_CAST]], align 8
 ; CHECK-NEXT:    [[VEC_GEP:%.*]] = getelementptr double, double* [[TMP1]], i64 2
 ; CHECK-NEXT:    [[VEC_CAST1:%.*]] = bitcast double* [[VEC_GEP]] to <2 x double>*
-; CHECK-NEXT:    load <2 x double>, <2 x double>* [[VEC_CAST1]], align 8
+; CHECK-NEXT:    load volatile <2 x double>, <2 x double>* [[VEC_CAST1]], align 8
 ; CHECK-NOT:     = load
 ;
   %in.m = load volatile <4 x double>, <4 x double>* %in
   %res = call <4 x double> @llvm.matrix.multiply(<4 x double> %in.m, <4 x double> %in.m, i32 2, i32 2, i32 2)
   ret <4 x double> %res
 }
 
 declare <4 x double> @llvm.matrix.multiply(<4 x double>, <4 x double>, i32, i32, i32)

llvm/test/Transforms/LowerMatrixIntrinsics/multiply-fused-volatile.ll

+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
+; RUN: opt -lower-matrix-intrinsics -fuse-matrix-tile-size=2 -matrix-allow-contract -force-fuse-matrix -verify-dom-info %s -S | FileCheck %s
+
+; REQUIRES: aarch64-registered-target
+
+target datalayout = "e-m:o-i64:64-f80:128-n8:16:32:64-S128"
+target triple = "aarch64-apple-ios"
+
+define void @multiply_all_volatile(<4 x double>* noalias %A, <4 x double>* noalias %B, <4 x double>* noalias %C) {
+; CHECK-LABEL: @multiply_all_volatile(
+; CHECK-NEXT:  entry:
+; CHECK-NEXT:    [[TMP0:%.*]] = bitcast <4 x double>* [[A:%.*]] to double*
+; CHECK-NEXT:    [[TMP1:%.*]] = getelementptr double, double* [[TMP0]], i64 0
+; CHECK-NEXT:    [[COL_CAST:%.*]] = bitcast double* [[TMP1]] to <4 x double>*
+; CHECK-NEXT:    [[TMP2:%.*]] = bitcast <4 x double>* [[COL_CAST]] to double*
+; CHECK-NEXT:    [[VEC_CAST:%.*]] = bitcast double* [[TMP2]] to <2 x double>*
+; CHECK-NEXT:    load volatile <2 x double>, <2 x double>* [[VEC_CAST]], align 8
+; CHECK-NEXT:    [[VEC_GEP:%.*]] = getelementptr double, double* [[TMP2]], i64 2
+; CHECK-NEXT:    [[VEC_CAST1:%.*]] = bitcast double* [[VEC_GEP]] to <2 x double>*
+; CHECK-NEXT:    load volatile <2 x double>, <2 x double>* [[VEC_CAST1]], align 8
+; CHECK-NEXT:    [[TMP3:%.*]] = bitcast <4 x double>* [[B:%.*]] to double*
+; CHECK-NEXT:    [[TMP4:%.*]] = getelementptr double, double* [[TMP3]], i64 0
+; CHECK-NEXT:    [[COL_CAST3:%.*]] = bitcast double* [[TMP4]] to <4 x double>*
+; CHECK-NEXT:    [[TMP5:%.*]] = bitcast <4 x double>* [[COL_CAST3]] to double*
+; CHECK-NEXT:    [[VEC_CAST4:%.*]] = bitcast double* [[TMP5]] to <2 x double>*
+; CHECK-NEXT:    load volatile <2 x double>, <2 x double>* [[VEC_CAST4]], align 8
+; CHECK-NEXT:    [[VEC_GEP6:%.*]] = getelementptr double, double* [[TMP5]], i64 2
+; CHECK-NEXT:    [[VEC_CAST7:%.*]] = bitcast double* [[VEC_GEP6]] to <2 x double>*
+; CHECK-NEXT:    load volatile <2 x double>, <2 x double>* [[VEC_CAST7]], align 8
+
+; CHECK:         [[TMP18:%.*]] = bitcast <4 x double>* [[C:%.*]] to double*
+; CHECK-NEXT:    [[TMP19:%.*]] = getelementptr double, double* [[TMP18]], i64 0
+; CHECK-NEXT:    [[COL_CAST18:%.*]] = bitcast double* [[TMP19]] to <4 x double>*
+; CHECK-NEXT:    [[TMP20:%.*]] = bitcast <4 x double>* [[COL_CAST18]] to double*
+; CHECK-NEXT:    [[VEC_CAST19:%.*]] = bitcast double* [[TMP20]] to <2 x double>*
+; CHECK-NEXT:    store volatile <2 x double> {{.*}}, <2 x double>* [[VEC_CAST19]], align 8
+; CHECK-NEXT:    [[VEC_GEP20:%.*]] = getelementptr double, double* [[TMP20]], i64 2
+; CHECK-NEXT:    [[VEC_CAST21:%.*]] = bitcast double* [[VEC_GEP20]] to <2 x double>*
+; CHECK-NEXT:    store volatile <2 x double> {{.*}}, <2 x double>* [[VEC_CAST21]], align 8
+; CHECK-NEXT:    ret void
+;
+
+entry:
+  %a = load volatile <4 x double>, <4 x double>* %A, align 16
+  %b = load volatile <4 x double>, <4 x double>* %B, align 16
+
+  %c = call <4 x double> @llvm.matrix.multiply(<4 x double> %a, <4 x double> %b, i32 2, i32 2, i32 2)
+
+  store volatile <4 x double> %c, <4 x double>* %C, align 16
+  ret void
+}
+
+
+define void @multiply_load0_volatile(<4 x double>* noalias %A, <4 x double>* noalias %B, <4 x double>* noalias %C) {
+; CHECK-LABEL: @multiply_load0_volatile(
+; CHECK-NEXT:  entry:
+; CHECK-NEXT:    [[TMP0:%.*]] = bitcast <4 x double>* [[A:%.*]] to double*
+; CHECK-NEXT:    [[TMP1:%.*]] = getelementptr double, double* [[TMP0]], i64 0
+; CHECK-NEXT:    [[COL_CAST:%.*]] = bitcast double* [[TMP1]] to <4 x double>*
+; CHECK-NEXT:    [[TMP2:%.*]] = bitcast <4 x double>* [[COL_CAST]] to double*
+; CHECK-NEXT:    [[VEC_CAST:%.*]] = bitcast double* [[TMP2]] to <2 x double>*
+; CHECK-NEXT:    load volatile <2 x double>, <2 x double>* [[VEC_CAST]], align 8
+; CHECK-NEXT:    [[VEC_GEP:%.*]] = getelementptr double, double* [[TMP2]], i64 2
+; CHECK-NEXT:    [[VEC_CAST1:%.*]] = bitcast double* [[VEC_GEP]] to <2 x double>*
+; CHECK-NEXT:    load volatile <2 x double>, <2 x double>* [[VEC_CAST1]], align 8
+; CHECK-NEXT:    [[TMP3:%.*]] = bitcast <4 x double>* [[B:%.*]] to double*
+; CHECK-NEXT:    [[TMP4:%.*]] = getelementptr double, double* [[TMP3]], i64 0
+; CHECK-NEXT:    [[COL_CAST3:%.*]] = bitcast double* [[TMP4]] to <4 x double>*
+; CHECK-NEXT:    [[TMP5:%.*]] = bitcast <4 x double>* [[COL_CAST3]] to double*
+; CHECK-NEXT:    [[VEC_CAST4:%.*]] = bitcast double* [[TMP5]] to <2 x double>*
+; CHECK-NEXT:    load <2 x double>, <2 x double>* [[VEC_CAST4]], align 8
+; CHECK-NEXT:    [[VEC_GEP6:%.*]] = getelementptr double, double* [[TMP5]], i64 2
+; CHECK-NEXT:    [[VEC_CAST7:%.*]] = bitcast double* [[VEC_GEP6]] to <2 x double>*
+; CHECK-NEXT:    load <2 x double>, <2 x double>* [[VEC_CAST7]], align 8
+
+; CHECK:         [[TMP18:%.*]] = bitcast <4 x double>* [[C:%.*]] to double*
+; CHECK-NEXT:    [[TMP19:%.*]] = getelementptr double, double* [[TMP18]], i64 0
+; CHECK-NEXT:    [[COL_CAST18:%.*]] = bitcast double* [[TMP19]] to <4 x double>*
+; CHECK-NEXT:    [[TMP20:%.*]] = bitcast <4 x double>* [[COL_CAST18]] to double*
+; CHECK-NEXT:    [[VEC_CAST19:%.*]] = bitcast double* [[TMP20]] to <2 x double>*
+; CHECK-NEXT:    store <2 x double> {{.*}}, <2 x double>* [[VEC_CAST19]], align 8
+; CHECK-NEXT:    [[VEC_GEP20:%.*]] = getelementptr double, double* [[TMP20]], i64 2
+; CHECK-NEXT:    [[VEC_CAST21:%.*]] = bitcast double* [[VEC_GEP20]] to <2 x double>*
+; CHECK-NEXT:    store <2 x double> {{.*}}, <2 x double>* [[VEC_CAST21]], align 8
+; CHECK-NEXT:    ret void
+;
+
+entry:
+  %a = load volatile <4 x double>, <4 x double>* %A, align 16
+  %b = load <4 x double>, <4 x double>* %B, align 16
+
+  %c = call <4 x double> @llvm.matrix.multiply(<4 x double> %a, <4 x double> %b, i32 2, i32 2, i32 2)
+
+  store <4 x double> %c, <4 x double>* %C, align 16
+  ret void
+}
+
+define void @multiply_load1_volatile(<4 x double>* noalias %A, <4 x double>* noalias %B, <4 x double>* noalias %C) {
+; CHECK-LABEL: @multiply_load1_volatile(
+; CHECK-NEXT:  entry:
+; CHECK-NEXT:    [[TMP0:%.*]] = bitcast <4 x double>* [[A:%.*]] to double*
+; CHECK-NEXT:    [[TMP1:%.*]] = getelementptr double, double* [[TMP0]], i64 0
+; CHECK-NEXT:    [[COL_CAST:%.*]] = bitcast double* [[TMP1]] to <4 x double>*
+; CHECK-NEXT:    [[TMP2:%.*]] = bitcast <4 x double>* [[COL_CAST]] to double*
+; CHECK-NEXT:    [[VEC_CAST:%.*]] = bitcast double* [[TMP2]] to <2 x double>*
+; CHECK-NEXT:    load <2 x double>, <2 x double>* [[VEC_CAST]], align 8
+; CHECK-NEXT:    [[VEC_GEP:%.*]] = getelementptr double, double* [[TMP2]], i64 2
+; CHECK-NEXT:    [[VEC_CAST1:%.*]] = bitcast double* [[VEC_GEP]] to <2 x double>*
+; CHECK-NEXT:    load <2 x double>, <2 x double>* [[VEC_CAST1]], align 8
+; CHECK-NEXT:    [[TMP3:%.*]] = bitcast <4 x double>* [[B:%.*]] to double*
+; CHECK-NEXT:    [[TMP4:%.*]] = getelementptr double, double* [[TMP3]], i64 0
+; CHECK-NEXT:    [[COL_CAST3:%.*]] = bitcast double* [[TMP4]] to <4 x double>*
+; CHECK-NEXT:    [[TMP5:%.*]] = bitcast <4 x double>* [[COL_CAST3]] to double*
+; CHECK-NEXT:    [[VEC_CAST4:%.*]] = bitcast double* [[TMP5]] to <2 x double>*
+; CHECK-NEXT:    load volatile <2 x double>, <2 x double>* [[VEC_CAST4]], align 8
+; CHECK-NEXT:    [[VEC_GEP6:%.*]] = getelementptr double, double* [[TMP5]], i64 2
+; CHECK-NEXT:    [[VEC_CAST7:%.*]] = bitcast double* [[VEC_GEP6]] to <2 x double>*
+; CHECK-NEXT:    load volatile <2 x double>, <2 x double>* [[VEC_CAST7]], align 8
+
+; CHECK:         [[TMP18:%.*]] = bitcast <4 x double>* [[C:%.*]] to double*
+; CHECK-NEXT:    [[TMP19:%.*]] = getelementptr double, double* [[TMP18]], i64 0
+; CHECK-NEXT:    [[COL_CAST18:%.*]] = bitcast double* [[TMP19]] to <4 x double>*
+; CHECK-NEXT:    [[TMP20:%.*]] = bitcast <4 x double>* [[COL_CAST18]] to double*
+; CHECK-NEXT:    [[VEC_CAST19:%.*]] = bitcast double* [[TMP20]] to <2 x double>*
+; CHECK-NEXT:    store <2 x double> {{.*}}, <2 x double>* [[VEC_CAST19]], align 8
+; CHECK-NEXT:    [[VEC_GEP20:%.*]] = getelementptr double, double* [[TMP20]], i64 2
+; CHECK-NEXT:    [[VEC_CAST21:%.*]] = bitcast double* [[VEC_GEP20]] to <2 x double>*
+; CHECK-NEXT:    store <2 x double> {{.*}}, <2 x double>* [[VEC_CAST21]], align 8
+; CHECK-NEXT:    ret void
+;
+
+entry:
+  %a = load <4 x double>, <4 x double>* %A, align 16
+  %b = load volatile <4 x double>, <4 x double>* %B, align 16
+
+  %c = call <4 x double> @llvm.matrix.multiply(<4 x double> %a, <4 x double> %b, i32 2, i32 2, i32 2)
+
+  store <4 x double> %c, <4 x double>* %C, align 16
+  ret void
+}
+
+define void @multiply_store_volatile(<4 x double>* noalias %A, <4 x double>* noalias %B, <4 x double>* noalias %C) {
+; CHECK-LABEL: @multiply_store_volatile(
+; CHECK-NEXT:  entry:
+; CHECK-NEXT:    [[TMP0:%.*]] = bitcast <4 x double>* [[A:%.*]] to double*
+; CHECK-NEXT:    [[TMP1:%.*]] = getelementptr double, double* [[TMP0]], i64 0
+; CHECK-NEXT:    [[COL_CAST:%.*]] = bitcast double* [[TMP1]] to <4 x double>*
+; CHECK-NEXT:    [[TMP2:%.*]] = bitcast <4 x double>* [[COL_CAST]] to double*
+; CHECK-NEXT:    [[VEC_CAST:%.*]] = bitcast double* [[TMP2]] to <2 x double>*
+; CHECK-NEXT:    load <2 x double>, <2 x double>* [[VEC_CAST]], align 8
+; CHECK-NEXT:    [[VEC_GEP:%.*]] = getelementptr double, double* [[TMP2]], i64 2
+; CHECK-NEXT:    [[VEC_CAST1:%.*]] = bitcast double* [[VEC_GEP]] to <2 x double>*
+; CHECK-NEXT:    load <2 x double>, <2 x double>* [[VEC_CAST1]], align 8
+; CHECK-NEXT:    [[TMP3:%.*]] = bitcast <4 x double>* [[B:%.*]] to double*
+; CHECK-NEXT:    [[TMP4:%.*]] = getelementptr double, double* [[TMP3]], i64 0
+; CHECK-NEXT:    [[COL_CAST3:%.*]] = bitcast double* [[TMP4]] to <4 x double>*
+; CHECK-NEXT:    [[TMP5:%.*]] = bitcast <4 x double>* [[COL_CAST3]] to double*
+; CHECK-NEXT:    [[VEC_CAST4:%.*]] = bitcast double* [[TMP5]] to <2 x double>*
+; CHECK-NEXT:    load <2 x double>, <2 x double>* [[VEC_CAST4]], align 8
+; CHECK-NEXT:    [[VEC_GEP6:%.*]] = getelementptr double, double* [[TMP5]], i64 2
+; CHECK-NEXT:    [[VEC_CAST7:%.*]] = bitcast double* [[VEC_GEP6]] to <2 x double>*
+; CHECK-NEXT:    load <2 x double>, <2 x double>* [[VEC_CAST7]], align 8
+
+; CHECK:         [[TMP18:%.*]] = bitcast <4 x double>* [[C:%.*]] to double*
+; CHECK-NEXT:    [[TMP19:%.*]] = getelementptr double, double* [[TMP18]], i64 0
+; CHECK-NEXT:    [[COL_CAST18:%.*]] = bitcast double* [[TMP19]] to <4 x double>*
+; CHECK-NEXT:    [[TMP20:%.*]] = bitcast <4 x double>* [[COL_CAST18]] to double*
+; CHECK-NEXT:    [[VEC_CAST19:%.*]] = bitcast double* [[TMP20]] to <2 x double>*
+; CHECK-NEXT:    store volatile <2 x double> {{.*}}, <2 x double>* [[VEC_CAST19]], align 8
+; CHECK-NEXT:    [[VEC_GEP20:%.*]] = getelementptr double, double* [[TMP20]], i64 2
+; CHECK-NEXT:    [[VEC_CAST21:%.*]] = bitcast double* [[VEC_GEP20]] to <2 x double>*
+; CHECK-NEXT:    store volatile <2 x double> {{.*}}, <2 x double>* [[VEC_CAST21]], align 8
+; CHECK-NEXT:    ret void
+;
+entry:
+  %a = load <4 x double>, <4 x double>* %A, align 16
+  %b = load <4 x double>, <4 x double>* %B, align 16
+
+  %c = call <4 x double> @llvm.matrix.multiply(<4 x double> %a, <4 x double> %b, i32 2, i32 2, i32 2)
+
+  store volatile <4 x double> %c, <4 x double>* %C, align 16
+  ret void
+}
+
+declare <4 x double> @llvm.matrix.multiply(<4 x double>, <4 x double>, i32, i32, i32)

llvm/test/Transforms/LowerMatrixIntrinsics/store-align-volatile.ll

 ; RUN: opt -lower-matrix-intrinsics -S < %s | FileCheck %s
 ; RUN: opt -passes='lower-matrix-intrinsics' -S < %s | FileCheck %s
 
 define void @strided_store_volatile(<6 x i32> %in, i32* %out) {
 ; CHECK-LABEL: @strided_store_volatile(
 ; CHECK-NEXT:    [[SPLIT:%.*]] = shufflevector <6 x i32> [[IN:%.*]], <6 x i32> undef, <3 x i32> <i32 0, i32 1, i32 2>
 ; CHECK-NEXT:    [[SPLIT1:%.*]] = shufflevector <6 x i32> [[IN]], <6 x i32> undef, <3 x i32> <i32 3, i32 4, i32 5>
 ; CHECK-NEXT:    [[VEC_CAST:%.*]] = bitcast i32* [[OUT:%.*]] to <3 x i32>*
-; CHECK-NEXT:    store <3 x i32> [[SPLIT]], <3 x i32>* [[VEC_CAST]], align 4
+; CHECK-NEXT:    store volatile <3 x i32> [[SPLIT]], <3 x i32>* [[VEC_CAST]], align 4
 ; CHECK-NEXT:    [[VEC_GEP:%.*]] = getelementptr i32, i32* [[OUT]], i64 5
 ; CHECK-NEXT:    [[VEC_CAST2:%.*]] = bitcast i32* [[VEC_GEP]] to <3 x i32>*
-; CHECK-NEXT:    store <3 x i32> [[SPLIT1]], <3 x i32>* [[VEC_CAST2]], align 4
+; CHECK-NEXT:    store volatile <3 x i32> [[SPLIT1]], <3 x i32>* [[VEC_CAST2]], align 4
 ; CHECK-NEXT:    ret void
 ;
   call void @llvm.matrix.column.major.store(<6 x i32> %in, i32* %out, i64 5, i1 true, i64 3, i64 2)
   ret void
 }
 
 declare void @llvm.matrix.column.major.store(<6 x i32>, i32*, i64, i1, i64, i64)
 
 
 define void @multiply_store_volatile(<4 x i32> %in, <4 x i32>* %out) {
 ; CHECK-LABEL: @multiply_store_volatile(
 ; CHECK:         [[TMP29:%.*]] = bitcast <4 x i32>* %out to i32*
 ; CHECK-NEXT:    [[VEC_CAST:%.*]] = bitcast i32* [[TMP29]] to <2 x i32>*
-; CHECK-NEXT:    store <2 x i32> {{.*}}, <2 x i32>* [[VEC_CAST]], align 4
+; CHECK-NEXT:    store volatile <2 x i32> {{.*}}, <2 x i32>* [[VEC_CAST]], align 4
 ; CHECK-NEXT:    [[VEC_GEP:%.*]] = getelementptr i32, i32* [[TMP29]], i64 2
 ; CHECK-NEXT:    [[VEC_CAST25:%.*]] = bitcast i32* [[VEC_GEP]] to <2 x i32>*
-; CHECK-NEXT:    store <2 x i32> {{.*}}, <2 x i32>* [[VEC_CAST25]], align 4
+; CHECK-NEXT:    store volatile <2 x i32> {{.*}}, <2 x i32>* [[VEC_CAST25]], align 4
 ; CHECK-NEXT:    ret void
 ;
   %res = call <4 x i32> @llvm.matrix.multiply(<4 x i32> %in, <4 x i32> %in, i32 2, i32 2, i32 2)
   store volatile <4 x i32> %res, <4 x i32>* %out
   ret void
 }
 
 declare <4 x i32> @llvm.matrix.multiply(<4 x i32>, <4 x i32>, i32, i32, i32)