1 //===- LowerMatrixIntrinsics.cpp -  Lower matrix intrinsics -----*- C++ -*-===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 //
9 // Lower matrix intrinsics to vector operations.
10 //
11 // TODO:
12 //  * Improve fusion:
13 //   * Support more cases, e.g. multiply-add, multiply-sub, operands/results
14 //     transposed.
15 //   * Improve cost-modeling, e.g. choose different number of rows/columns
16 //     columns for tiles, consider cost of copies on alias.
17 //
18 //===----------------------------------------------------------------------===//
19 
20 #include "llvm/Transforms/Scalar/LowerMatrixIntrinsics.h"
21 #include "llvm/ADT/PostOrderIterator.h"
22 #include "llvm/ADT/SmallVector.h"
23 #include "llvm/Analysis/AliasAnalysis.h"
24 #include "llvm/Analysis/DomTreeUpdater.h"
25 #include "llvm/Analysis/LoopInfo.h"
26 #include "llvm/Analysis/OptimizationRemarkEmitter.h"
27 #include "llvm/Analysis/TargetTransformInfo.h"
28 #include "llvm/Analysis/ValueTracking.h"
29 #include "llvm/Analysis/VectorUtils.h"
30 #include "llvm/IR/CFG.h"
31 #include "llvm/IR/DataLayout.h"
32 #include "llvm/IR/DebugInfoMetadata.h"
33 #include "llvm/IR/Function.h"
34 #include "llvm/IR/IRBuilder.h"
35 #include "llvm/IR/Instructions.h"
36 #include "llvm/IR/IntrinsicInst.h"
37 #include "llvm/IR/MatrixBuilder.h"
38 #include "llvm/IR/PatternMatch.h"
39 #include "llvm/InitializePasses.h"
40 #include "llvm/Pass.h"
41 #include "llvm/Support/Alignment.h"
42 #include "llvm/Support/CommandLine.h"
43 #include "llvm/Support/Debug.h"
44 #include "llvm/Transforms/Scalar.h"
45 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
46 #include "llvm/Transforms/Utils/LoopUtils.h"
47 #include "llvm/Transforms/Utils/MatrixUtils.h"
48 
49 #include <cmath>
50 
51 using namespace llvm;
52 using namespace PatternMatch;
53 
54 #define DEBUG_TYPE "lower-matrix-intrinsics"
55 
56 static cl::opt<bool>
57     FuseMatrix("fuse-matrix", cl::init(true), cl::Hidden,
58                cl::desc("Enable/disable fusing matrix instructions."));
59 // TODO: Allow and use non-square tiles.
60 static cl::opt<unsigned> TileSize(
61     "fuse-matrix-tile-size", cl::init(4), cl::Hidden,
62     cl::desc(
63         "Tile size for matrix instruction fusion using square-shaped tiles."));
64 static cl::opt<bool> TileUseLoops("fuse-matrix-use-loops", cl::init(false),
65                                   cl::Hidden,
66                                   cl::desc("Generate loop nest for tiling."));
67 static cl::opt<bool> ForceFusion(
68     "force-fuse-matrix", cl::init(false), cl::Hidden,
69     cl::desc("Force matrix instruction fusion even if not profitable."));
70 static cl::opt<bool> AllowContractEnabled(
71     "matrix-allow-contract", cl::init(false), cl::Hidden,
72     cl::desc("Allow the use of FMAs if available and profitable. This may "
73              "result in different results, due to less rounding error."));
74 
75 static cl::opt<bool>
76     VerifyShapeInfo("verify-matrix-shapes", cl::Hidden,
77                     cl::desc("Enable/disable matrix shape verification."),
78                     cl::init(false));
79 
80 enum class MatrixLayoutTy { ColumnMajor, RowMajor };
81 
82 static cl::opt<MatrixLayoutTy> MatrixLayout(
83     "matrix-default-layout", cl::init(MatrixLayoutTy::ColumnMajor),
84     cl::desc("Sets the default matrix layout"),
85     cl::values(clEnumValN(MatrixLayoutTy::ColumnMajor, "column-major",
86                           "Use column-major layout"),
87                clEnumValN(MatrixLayoutTy::RowMajor, "row-major",
88                           "Use row-major layout")));
89 
90 static cl::opt<bool> PrintAfterTransposeOpt("matrix-print-after-transpose-opt",
91                                             cl::init(false));
92 
93 /// Helper function to either return Scope, if it is a subprogram or the
94 /// attached subprogram for a local scope.
95 static DISubprogram *getSubprogram(DIScope *Scope) {
96   if (auto *Subprogram = dyn_cast<DISubprogram>(Scope))
97     return Subprogram;
98   return cast<DILocalScope>(Scope)->getSubprogram();
99 }
100 
101 /// Erase \p V from \p BB and move \II forward to avoid invalidating
102 /// iterators.
103 static void eraseFromParentAndMove(Value *V, BasicBlock::reverse_iterator &II,
104                                    BasicBlock &BB) {
105   auto *Inst = cast<Instruction>(V);
106   // Still used, don't erase.
107   if (!Inst->use_empty())
108     return;
109   if (II != BB.rend() && Inst == &*II)
110     ++II;
111   Inst->eraseFromParent();
112 }
113 
114 /// Return true if V is a splat of a value (which is used when multiplying a
115 /// matrix with a scalar).
116 static bool isSplat(Value *V) {
117   if (auto *SV = dyn_cast<ShuffleVectorInst>(V))
118     return SV->isZeroEltSplat();
119   return false;
120 }
121 
122 /// Match any mul operation (fp or integer).
123 template <typename LTy, typename RTy>
124 auto m_AnyMul(const LTy &L, const RTy &R) {
125   return m_CombineOr(m_Mul(L, R), m_FMul(L, R));
126 }
127 
128 /// Match any add operation (fp or integer).
129 template <typename LTy, typename RTy>
130 auto m_AnyAdd(const LTy &L, const RTy &R) {
131   return m_CombineOr(m_Add(L, R), m_FAdd(L, R));
132 }
133 
134 namespace {
135 
136 // Given an element pointer \p BasePtr to the start of a (sub) matrix, compute
137 // the start address of vector \p VecIdx with type (\p EltType x \p NumElements)
138 // assuming \p Stride elements between start two consecutive vectors.
139 // \p Stride must be >= \p NumElements.
140 // For column-major matrixes, the function computes the address of a column
141 // vectors and \p NumElements must be set to the number of elements in a column
142 // (= number of rows of the matrix). For row-major matrixes, the function
143 // computes the address of a row vector and \p NumElements must be set to the
144 // number of elements in a column (= number of columns of the matrix).
145 //
146 // Consider a 4x4 matrix in column-mjaor layout like below
147 //
148 //      0       1      2      3
149 // 0   v_0_0  v_0_1  v_0_2  v_0_3
150 // 1   v_1_0  v_1_1  v_1_2  v_1_3
151 // 2   v_2_0  v_2_1  v_2_2  v_2_3
152 // 3   v_3_0  v_3_1  v_3_2  v_3_3
153 
154 // To compute the column addresses for a 2x3 sub-matrix at row 1 and column 1,
155 // we need a pointer to the first element of the submatrix as base pointer.
156 // Then we can use computeVectorAddr to compute the addresses for the columns
157 // of the sub-matrix.
158 //
159 // Column 0: computeVectorAddr(Base, 0 (column), 4 (stride), 2 (num rows), ..)
160 //           -> just returns Base
161 // Column 1: computeVectorAddr(Base, 1 (column), 4 (stride), 2 (num rows), ..)
162 //           -> returns Base + (1 * 4)
163 // Column 2: computeVectorAddr(Base, 2 (column), 4 (stride), 2 (num rows), ..)
164 //           -> returns Base + (2 * 4)
165 //
166 // The graphic below illustrates the number of elements in a column (marked
167 // with |) and the number of skipped elements (marked with }).
168 //
169 //         v_0_0  v_0_1 {v_0_2 {v_0_3
170 //                Base   Col 1  Col 2
171 //                  |     |      |
172 //         v_1_0 |v_1_1 |v_1_2 |v_1_3
173 //         v_2_0 |v_2_1 |v_2_2 |v_2_3
174 //         v_3_0 {v_3_1 {v_3_2  v_3_3
175 //
176 Value *computeVectorAddr(Value *BasePtr, Value *VecIdx, Value *Stride,
177                          unsigned NumElements, Type *EltType,
178                          IRBuilder<> &Builder) {
179 
180   assert((!isa<ConstantInt>(Stride) ||
181           cast<ConstantInt>(Stride)->getZExtValue() >= NumElements) &&
182          "Stride must be >= the number of elements in the result vector.");
183   unsigned AS = cast<PointerType>(BasePtr->getType())->getAddressSpace();
184 
185   // Compute the start of the vector with index VecIdx as VecIdx * Stride.
186   Value *VecStart = Builder.CreateMul(VecIdx, Stride, "vec.start");
187 
188   // Get pointer to the start of the selected vector. Skip GEP creation,
189   // if we select vector 0.
190   if (isa<ConstantInt>(VecStart) && cast<ConstantInt>(VecStart)->isZero())
191     VecStart = BasePtr;
192   else
193     VecStart = Builder.CreateGEP(EltType, BasePtr, VecStart, "vec.gep");
194 
195   // Cast elementwise vector start pointer to a pointer to a vector
196   // (EltType x NumElements)*.
197   auto *VecType = FixedVectorType::get(EltType, NumElements);
198   Type *VecPtrType = PointerType::get(VecType, AS);
199   return Builder.CreatePointerCast(VecStart, VecPtrType, "vec.cast");
200 }
201 
202 /// LowerMatrixIntrinsics contains the methods used to lower matrix intrinsics.
203 ///
204 /// Currently, the lowering for each matrix intrinsic is done as follows:
205 /// 1. Propagate the shape information from intrinsics to connected
206 /// instructions.
207 /// 2. Lower instructions with shape information (assuming column-major layout).
208 ///  The lowering works similarly using row-major layout.
209 ///  2.1. Get column vectors for each argument. If we already lowered the
210 ///       definition of an argument, use the produced column vectors directly.
211 ///       If not, split the operand vector containing an embedded matrix into
212 ///       a set of column vectors,
213 ///  2.2. Lower the instruction in terms of column major operations, which
214 ///       yields a set of column vectors containing result matrix. Note that we
215 ///       lower all instructions that have shape information. Besides the
216 ///       intrinsics, this includes stores for example.
217 ///  2.3. Update uses of the lowered instruction. If we have shape information
218 ///       for a user, there is nothing to do, as we will look up the result
219 ///       column matrix when lowering the user. For other uses, we embed the
220 ///       result matrix in a flat vector and update the use.
221 ///  2.4. Cache the result column matrix for the instruction we lowered
222 /// 3. After we lowered all instructions in a function, remove the now
223 ///    obsolete instructions.
224 ///
225 class LowerMatrixIntrinsics {
226   Function &Func;
227   const DataLayout &DL;
228   const TargetTransformInfo &TTI;
229   AliasAnalysis *AA;
230   DominatorTree *DT;
231   LoopInfo *LI;
232   OptimizationRemarkEmitter *ORE;
233 
234   /// Contains estimates of the number of operations (loads, stores, compute) required to lower a matrix operation.
235   struct OpInfoTy {
236     /// Number of stores emitted to generate this matrix.
237     unsigned NumStores = 0;
238     /// Number of loads emitted to generate this matrix.
239     unsigned NumLoads = 0;
240     /// Number of compute operations emitted to generate this matrix.
241     unsigned NumComputeOps = 0;
242     /// Most of the time transposes can be fused with matrix multiplies or can
243     /// be folded away via algebraic simplifications.  This is the number of
244     /// transposes that we failed to make "free" via such optimizations.
245     unsigned NumExposedTransposes = 0;
246 
247     OpInfoTy &operator+=(const OpInfoTy &RHS) {
248       NumStores += RHS.NumStores;
249       NumLoads += RHS.NumLoads;
250       NumComputeOps += RHS.NumComputeOps;
251       NumExposedTransposes += RHS.NumExposedTransposes;
252       return *this;
253     }
254   };
255 
256   /// Wrapper class representing a matrix as a set of vectors, either in row or
257   /// column major layout. All vectors must have the same vector type.
258   class MatrixTy {
259     SmallVector<Value *, 16> Vectors;
260 
261     OpInfoTy OpInfo;
262 
263     bool IsColumnMajor = true;
264 
265   public:
266     MatrixTy() : IsColumnMajor(MatrixLayout == MatrixLayoutTy::ColumnMajor) {}
267     MatrixTy(ArrayRef<Value *> Vectors)
268         : Vectors(Vectors.begin(), Vectors.end()),
269           IsColumnMajor(MatrixLayout == MatrixLayoutTy::ColumnMajor) {}
270     MatrixTy(unsigned NumRows, unsigned NumColumns, Type *EltTy)
271         : IsColumnMajor(MatrixLayout == MatrixLayoutTy::ColumnMajor) {
272 
273       unsigned D = isColumnMajor() ? NumColumns : NumRows;
274       for (unsigned J = 0; J < D; ++J)
275         addVector(PoisonValue::get(FixedVectorType::get(
276             EltTy, isColumnMajor() ? NumRows : NumColumns)));
277     }
278 
279     Value *getVector(unsigned i) const { return Vectors[i]; }
280     Value *getColumn(unsigned i) const {
281       assert(isColumnMajor() && "only supported for column-major matrixes");
282       return Vectors[i];
283     }
284     Value *getRow(unsigned i) const {
285       assert(!isColumnMajor() && "only supported for row-major matrixes");
286       return Vectors[i];
287     }
288 
289     void setVector(unsigned i, Value *V) { Vectors[i] = V; }
290 
291     Type *getElementType() const { return getVectorTy()->getElementType(); }
292 
293     unsigned getNumVectors() const {
294       if (isColumnMajor())
295         return getNumColumns();
296       return getNumRows();
297     }
298 
299     unsigned getNumColumns() const {
300       if (isColumnMajor())
301         return Vectors.size();
302       else {
303         assert(Vectors.size() > 0 && "Cannot call getNumRows without columns");
304         return cast<FixedVectorType>(Vectors[0]->getType())->getNumElements();
305       }
306     }
307     unsigned getNumRows() const {
308       if (isColumnMajor()) {
309         assert(Vectors.size() > 0 && "Cannot call getNumRows without columns");
310         return cast<FixedVectorType>(Vectors[0]->getType())->getNumElements();
311       } else
312         return Vectors.size();
313     }
314 
315     void addVector(Value *V) { Vectors.push_back(V); }
316     VectorType *getColumnTy() {
317       assert(isColumnMajor() && "only supported for column-major matrixes");
318       return getVectorTy();
319     }
320 
321     VectorType *getVectorTy() const {
322       return cast<VectorType>(Vectors[0]->getType());
323     }
324 
325     iterator_range<SmallVector<Value *, 8>::iterator> columns() {
326       assert(isColumnMajor() &&
327              "columns() only supported for column-major matrixes");
328       return make_range(Vectors.begin(), Vectors.end());
329     }
330 
331     iterator_range<SmallVector<Value *, 8>::iterator> vectors() {
332       return make_range(Vectors.begin(), Vectors.end());
333     }
334 
335     /// Embed the vectors of the matrix into a flat vector by concatenating
336     /// them.
337     Value *embedInVector(IRBuilder<> &Builder) const {
338       return Vectors.size() == 1 ? Vectors[0]
339                                  : concatenateVectors(Builder, Vectors);
340     }
341 
342     MatrixTy &addNumLoads(unsigned N) {
343       OpInfo.NumLoads += N;
344       return *this;
345     }
346 
347     void setNumLoads(unsigned N) { OpInfo.NumLoads = N; }
348 
349     MatrixTy &addNumStores(unsigned N) {
350       OpInfo.NumStores += N;
351       return *this;
352     }
353 
354     MatrixTy &addNumExposedTransposes(unsigned N) {
355       OpInfo.NumExposedTransposes += N;
356       return *this;
357     }
358 
359     MatrixTy &addNumComputeOps(unsigned N) {
360       OpInfo.NumComputeOps += N;
361       return *this;
362     }
363 
364     unsigned getNumStores() const { return OpInfo.NumStores; }
365     unsigned getNumLoads() const { return OpInfo.NumLoads; }
366     unsigned getNumComputeOps() const { return OpInfo.NumComputeOps; }
367 
368     const OpInfoTy &getOpInfo() const { return OpInfo; }
369 
370     bool isColumnMajor() const { return IsColumnMajor; }
371 
372     unsigned getStride() const {
373       if (isColumnMajor())
374         return getNumRows();
375       return getNumColumns();
376     }
377 
378     /// Extract a vector of \p NumElts starting at index (\p I, \p J). If the
379     /// matrix is column-major, the result vector is extracted from a column
380     /// vector, otherwise from a row vector.
381     Value *extractVector(unsigned I, unsigned J, unsigned NumElts,
382                          IRBuilder<> &Builder) const {
383       Value *Vec = isColumnMajor() ? getColumn(J) : getRow(I);
384       assert(cast<FixedVectorType>(Vec->getType())->getNumElements() >=
385                  NumElts &&
386              "Extracted vector will contain poison values");
387       return Builder.CreateShuffleVector(
388           Vec, createSequentialMask(isColumnMajor() ? I : J, NumElts, 0),
389           "block");
390     }
391   };
392 
393   struct ShapeInfo {
394     unsigned NumRows;
395     unsigned NumColumns;
396 
397     bool IsColumnMajor;
398 
399     ShapeInfo(unsigned NumRows = 0, unsigned NumColumns = 0)
400         : NumRows(NumRows), NumColumns(NumColumns),
401           IsColumnMajor(MatrixLayout == MatrixLayoutTy::ColumnMajor) {}
402 
403     ShapeInfo(Value *NumRows, Value *NumColumns)
404         : ShapeInfo(cast<ConstantInt>(NumRows)->getZExtValue(),
405                     cast<ConstantInt>(NumColumns)->getZExtValue()) {}
406 
407     bool operator==(const ShapeInfo &other) {
408       return NumRows == other.NumRows && NumColumns == other.NumColumns;
409     }
410     bool operator!=(const ShapeInfo &other) { return !(*this == other); }
411 
412     /// Returns true if shape-information is defined, meaning both dimensions
413     /// are != 0.
414     operator bool() const {
415       assert(NumRows == 0 || NumColumns != 0);
416       return NumRows != 0;
417     }
418 
419     unsigned getStride() const {
420       if (IsColumnMajor)
421         return NumRows;
422       return NumColumns;
423     }
424 
425     unsigned getNumVectors() const {
426       if (IsColumnMajor)
427         return NumColumns;
428       return NumRows;
429     }
430 
431     /// Returns the transposed shape.
432     ShapeInfo t() const { return ShapeInfo(NumColumns, NumRows); }
433   };
434 
435   /// Maps instructions to their shape information. The shape information
436   /// describes the shape to be used while lowering. This matches the shape of
437   /// the result value of the instruction, with the only exceptions being store
438   /// instructions and the matrix_column_major_store intrinsics. For those, the
439   /// shape information indicates that those instructions should be lowered
440   /// using shape information as well.  A ValueMap is used so that when
441   /// sub-passes like optimizeTransposes performs RAUW the map stays
442   /// up-to-date.
443   ValueMap<Value *, ShapeInfo> ShapeMap;
444 
445   /// List of instructions to remove. While lowering, we are not replacing all
446   /// users of a lowered instruction, if shape information is available and
447   /// those need to be removed after we finished lowering.
448   SmallVector<Instruction *, 16> ToRemove;
449 
450   /// Map from instructions to their produced column matrix.
451   MapVector<Value *, MatrixTy> Inst2ColumnMatrix;
452 
453 private:
454   static FastMathFlags getFastMathFlags(Instruction *Inst) {
455     FastMathFlags FMF;
456 
457     if (isa<FPMathOperator>(*Inst))
458       FMF = Inst->getFastMathFlags();
459 
460     FMF.setAllowContract(AllowContractEnabled || FMF.allowContract());
461 
462     return FMF;
463   }
464 
465 public:
466   LowerMatrixIntrinsics(Function &F, TargetTransformInfo &TTI,
467                         AliasAnalysis *AA, DominatorTree *DT, LoopInfo *LI,
468                         OptimizationRemarkEmitter *ORE)
469       : Func(F), DL(F.getParent()->getDataLayout()), TTI(TTI), AA(AA), DT(DT),
470         LI(LI), ORE(ORE) {}
471 
472   unsigned getNumOps(Type *VT) {
473     assert(isa<VectorType>(VT) && "Expected vector type");
474     return getNumOps(VT->getScalarType(),
475                      cast<FixedVectorType>(VT)->getNumElements());
476   }
477 
478   /// Is this the minimal version executed in the backend pipelines.
479   bool isMinimal() const {
480     return !DT;
481   }
482 
483   /// Return the estimated number of vector ops required for an operation on
484   /// \p VT * N.
485   unsigned getNumOps(Type *ST, unsigned N) {
486     return std::ceil((ST->getPrimitiveSizeInBits() * N).getFixedValue() /
487                      double(TTI.getRegisterBitWidth(
488                                    TargetTransformInfo::RGK_FixedWidthVector)
489                                 .getFixedValue()));
490   }
491 
492   /// Return the set of vectors that a matrix value is lowered to.
493   ///
494   /// If we lowered \p MatrixVal, just return the cache result matrix. Otherwise
495   /// split the flat vector \p MatrixVal containing a matrix with shape \p SI
496   /// into vectors.
497   MatrixTy getMatrix(Value *MatrixVal, const ShapeInfo &SI,
498                      IRBuilder<> &Builder) {
499     VectorType *VType = dyn_cast<VectorType>(MatrixVal->getType());
500     assert(VType && "MatrixVal must be a vector type");
501     assert(cast<FixedVectorType>(VType)->getNumElements() ==
502                SI.NumRows * SI.NumColumns &&
503            "The vector size must match the number of matrix elements");
504 
505     // Check if we lowered MatrixVal using shape information. In that case,
506     // return the existing matrix, if it matches the requested shape
507     // information. If there is a mis-match, embed the result in a flat
508     // vector and split it later.
509     auto Found = Inst2ColumnMatrix.find(MatrixVal);
510     if (Found != Inst2ColumnMatrix.end()) {
511       MatrixTy &M = Found->second;
512       // Return the found matrix, if its shape matches the requested shape
513       // information
514       if (SI.NumRows == M.getNumRows() && SI.NumColumns == M.getNumColumns())
515         return M;
516 
517       MatrixVal = M.embedInVector(Builder);
518     }
519 
520     // Otherwise split MatrixVal.
521     SmallVector<Value *, 16> SplitVecs;
522     for (unsigned MaskStart = 0;
523          MaskStart < cast<FixedVectorType>(VType)->getNumElements();
524          MaskStart += SI.getStride()) {
525       Value *V = Builder.CreateShuffleVector(
526           MatrixVal, createSequentialMask(MaskStart, SI.getStride(), 0),
527           "split");
528       SplitVecs.push_back(V);
529     }
530 
531     return {SplitVecs};
532   }
533 
534   /// If \p V already has a known shape return false.  Otherwise set the shape
535   /// for instructions that support it.
536   bool setShapeInfo(Value *V, ShapeInfo Shape) {
537     assert(Shape && "Shape not set");
538     if (isa<UndefValue>(V) || !supportsShapeInfo(V))
539       return false;
540 
541     auto SIter = ShapeMap.find(V);
542     if (SIter != ShapeMap.end()) {
543       if (VerifyShapeInfo && (SIter->second.NumRows != Shape.NumRows ||
544                               SIter->second.NumColumns != Shape.NumColumns)) {
545         errs() << "Conflicting shapes (" << SIter->second.NumRows << "x"
546                << SIter->second.NumColumns << " vs " << Shape.NumRows << "x"
547                << Shape.NumColumns << ") for " << *V << "\n";
548         report_fatal_error(
549             "Matrix shape verification failed, compilation aborted!");
550       }
551 
552       LLVM_DEBUG(dbgs() << "  not overriding existing shape: "
553                         << SIter->second.NumRows << " "
554                         << SIter->second.NumColumns << " for " << *V << "\n");
555       return false;
556     }
557 
558     ShapeMap.insert({V, Shape});
559     LLVM_DEBUG(dbgs() << "  " << Shape.NumRows << " x " << Shape.NumColumns
560                       << " for " << *V << "\n");
561     return true;
562   }
563 
564   bool isUniformShape(Value *V) {
565     Instruction *I = dyn_cast<Instruction>(V);
566     if (!I)
567       return true;
568 
569     switch (I->getOpcode()) {
570     case Instruction::FAdd:
571     case Instruction::FSub:
572     case Instruction::FMul: // Scalar multiply.
573     case Instruction::FNeg:
574     case Instruction::Add:
575     case Instruction::Mul:
576     case Instruction::Sub:
577       return true;
578     default:
579       return false;
580     }
581   }
582 
583   /// Returns true if shape information can be used for \p V. The supported
584   /// instructions must match the instructions that can be lowered by this pass.
585   bool supportsShapeInfo(Value *V) {
586     Instruction *Inst = dyn_cast<Instruction>(V);
587     if (!Inst)
588       return false;
589 
590     IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst);
591     if (II)
592       switch (II->getIntrinsicID()) {
593       case Intrinsic::matrix_multiply:
594       case Intrinsic::matrix_transpose:
595       case Intrinsic::matrix_column_major_load:
596       case Intrinsic::matrix_column_major_store:
597         return true;
598       default:
599         return false;
600       }
601     return isUniformShape(V) || isa<StoreInst>(V) || isa<LoadInst>(V);
602   }
603 
604   /// Propagate the shape information of instructions to their users.
605   /// The work list contains instructions for which we can compute the shape,
606   /// either based on the information provided by matrix intrinsics or known
607   /// shapes of operands.
608   SmallVector<Instruction *, 32>
609   propagateShapeForward(SmallVectorImpl<Instruction *> &WorkList) {
610     SmallVector<Instruction *, 32> NewWorkList;
611     // Pop an element for which we guaranteed to have at least one of the
612     // operand shapes.  Add the shape for this and then add users to the work
613     // list.
614     LLVM_DEBUG(dbgs() << "Forward-propagate shapes:\n");
615     while (!WorkList.empty()) {
616       Instruction *Inst = WorkList.pop_back_val();
617 
618       // New entry, set the value and insert operands
619       bool Propagate = false;
620 
621       Value *MatrixA;
622       Value *MatrixB;
623       Value *M;
624       Value *N;
625       Value *K;
626       if (match(Inst, m_Intrinsic<Intrinsic::matrix_multiply>(
627                           m_Value(MatrixA), m_Value(MatrixB), m_Value(M),
628                           m_Value(N), m_Value(K)))) {
629         Propagate = setShapeInfo(Inst, {M, K});
630       } else if (match(Inst, m_Intrinsic<Intrinsic::matrix_transpose>(
631                                  m_Value(MatrixA), m_Value(M), m_Value(N)))) {
632         // Flip dimensions.
633         Propagate = setShapeInfo(Inst, {N, M});
634       } else if (match(Inst, m_Intrinsic<Intrinsic::matrix_column_major_store>(
635                                  m_Value(MatrixA), m_Value(), m_Value(),
636                                  m_Value(), m_Value(M), m_Value(N)))) {
637         Propagate = setShapeInfo(Inst, {N, M});
638       } else if (match(Inst, m_Intrinsic<Intrinsic::matrix_column_major_load>(
639                                  m_Value(), m_Value(), m_Value(), m_Value(M),
640                                  m_Value(N)))) {
641         Propagate = setShapeInfo(Inst, {M, N});
642       } else if (match(Inst, m_Store(m_Value(MatrixA), m_Value()))) {
643         auto OpShape = ShapeMap.find(MatrixA);
644         if (OpShape != ShapeMap.end())
645           setShapeInfo(Inst, OpShape->second);
646         continue;
647       } else if (isUniformShape(Inst)) {
648         // Find the first operand that has a known shape and use that.
649         for (auto &Op : Inst->operands()) {
650           auto OpShape = ShapeMap.find(Op.get());
651           if (OpShape != ShapeMap.end()) {
652             Propagate |= setShapeInfo(Inst, OpShape->second);
653             break;
654           }
655         }
656       }
657 
658       if (Propagate) {
659         NewWorkList.push_back(Inst);
660         for (auto *User : Inst->users())
661           if (ShapeMap.count(User) == 0)
662             WorkList.push_back(cast<Instruction>(User));
663       }
664     }
665 
666     return NewWorkList;
667   }
668 
669   /// Propagate the shape to operands of instructions with shape information.
670   /// \p Worklist contains the instruction for which we already know the shape.
671   SmallVector<Instruction *, 32>
672   propagateShapeBackward(SmallVectorImpl<Instruction *> &WorkList) {
673     SmallVector<Instruction *, 32> NewWorkList;
674 
675     auto pushInstruction = [](Value *V,
676                               SmallVectorImpl<Instruction *> &WorkList) {
677       Instruction *I = dyn_cast<Instruction>(V);
678       if (I)
679         WorkList.push_back(I);
680     };
681     // Pop an element with known shape.  Traverse the operands, if their shape
682     // derives from the result shape and is unknown, add it and add them to the
683     // worklist.
684     LLVM_DEBUG(dbgs() << "Backward-propagate shapes:\n");
685     while (!WorkList.empty()) {
686       Value *V = WorkList.pop_back_val();
687 
688       size_t BeforeProcessingV = WorkList.size();
689       if (!isa<Instruction>(V))
690         continue;
691 
692       Value *MatrixA;
693       Value *MatrixB;
694       Value *M;
695       Value *N;
696       Value *K;
697       if (match(V, m_Intrinsic<Intrinsic::matrix_multiply>(
698                        m_Value(MatrixA), m_Value(MatrixB), m_Value(M),
699                        m_Value(N), m_Value(K)))) {
700         if (setShapeInfo(MatrixA, {M, N}))
701           pushInstruction(MatrixA, WorkList);
702 
703         if (setShapeInfo(MatrixB, {N, K}))
704           pushInstruction(MatrixB, WorkList);
705 
706       } else if (match(V, m_Intrinsic<Intrinsic::matrix_transpose>(
707                               m_Value(MatrixA), m_Value(M), m_Value(N)))) {
708         // Flip dimensions.
709         if (setShapeInfo(MatrixA, {M, N}))
710           pushInstruction(MatrixA, WorkList);
711       } else if (match(V, m_Intrinsic<Intrinsic::matrix_column_major_store>(
712                               m_Value(MatrixA), m_Value(), m_Value(), m_Value(),
713                               m_Value(M), m_Value(N)))) {
714         if (setShapeInfo(MatrixA, {M, N})) {
715           pushInstruction(MatrixA, WorkList);
716         }
717       } else if (isa<LoadInst>(V) ||
718                  match(V, m_Intrinsic<Intrinsic::matrix_column_major_load>())) {
719         // Nothing to do, no matrix input.
720       } else if (isa<StoreInst>(V)) {
721         // Nothing to do.  We forward-propagated to this so we would just
722         // backward propagate to an instruction with an already known shape.
723       } else if (isUniformShape(V)) {
724         // Propagate to all operands.
725         ShapeInfo Shape = ShapeMap[V];
726         for (Use &U : cast<Instruction>(V)->operands()) {
727           if (setShapeInfo(U.get(), Shape))
728             pushInstruction(U.get(), WorkList);
729         }
730       }
731       // After we discovered new shape info for new instructions in the
732       // worklist, we use their users as seeds for the next round of forward
733       // propagation.
734       for (size_t I = BeforeProcessingV; I != WorkList.size(); I++)
735         for (User *U : WorkList[I]->users())
736           if (isa<Instruction>(U) && V != U)
737             NewWorkList.push_back(cast<Instruction>(U));
738     }
739     return NewWorkList;
740   }
741 
742   /// (Op0 op Op1)^T -> Op0^T op Op1^T
743   /// Transpose \p Op0 and \p Op1 of shape \p Shape0 and \p Shape1, then use
744   /// them on both sides of \p Operation.
745   Instruction *distributeTransposes(
746       Value *Op0, ShapeInfo Shape0, Value *Op1, ShapeInfo Shape1,
747       MatrixBuilder &Builder,
748       function_ref<Instruction *(Value *, ShapeInfo, Value *, ShapeInfo)>
749           Operation) {
750     Value *T0 = Builder.CreateMatrixTranspose(
751         Op0, Shape0.NumRows, Shape0.NumColumns, Op0->getName() + "_t");
752     // We are being run after shape prop, add shape for newly created
753     // instructions so that we lower them later.
754     setShapeInfo(T0, Shape0.t());
755     Value *T1 = Builder.CreateMatrixTranspose(
756         Op1, Shape1.NumRows, Shape1.NumColumns, Op1->getName() + "_t");
757     setShapeInfo(T1, Shape1.t());
758     return Operation(T0, Shape0.t(), T1, Shape1.t());
759   }
760 
761   void updateShapeAndReplaceAllUsesWith(Instruction &Old, Value *New) {
762     // We need to remove Old from the ShapeMap otherwise RAUW will replace it
763     // with New. We should only add New it it supportsShapeInfo so we insert
764     // it conditionally instead.
765     auto S = ShapeMap.find(&Old);
766     if (S != ShapeMap.end()) {
767       ShapeMap.erase(S);
768       if (supportsShapeInfo(New))
769         ShapeMap.insert({New, S->second});
770     }
771     Old.replaceAllUsesWith(New);
772   }
773 
774   /// Sink a top-level transpose inside matmuls and adds.
775   /// This creates and erases instructions as needed, and returns the newly
776   /// created instruction while updating the iterator to avoid invalidation. If
777   /// this returns nullptr, no new instruction was created.
778   Instruction *sinkTranspose(Instruction &I, BasicBlock::reverse_iterator &II) {
779     BasicBlock &BB = *I.getParent();
780     IRBuilder<> IB(&I);
781     MatrixBuilder Builder(IB);
782 
783     Value *TA, *TAMA, *TAMB;
784     ConstantInt *R, *K, *C;
785     if (!match(&I, m_Intrinsic<Intrinsic::matrix_transpose>(
786                        m_Value(TA), m_ConstantInt(R), m_ConstantInt(C))))
787       return nullptr;
788 
789     // Transpose of a transpose is a nop
790     Value *TATA;
791     if (match(TA, m_Intrinsic<Intrinsic::matrix_transpose>(m_Value(TATA)))) {
792       updateShapeAndReplaceAllUsesWith(I, TATA);
793       eraseFromParentAndMove(&I, II, BB);
794       eraseFromParentAndMove(TA, II, BB);
795       return nullptr;
796     }
797 
798     // k^T -> k
799     if (isSplat(TA)) {
800       updateShapeAndReplaceAllUsesWith(I, TA);
801       eraseFromParentAndMove(&I, II, BB);
802       return nullptr;
803     }
804 
805     // (A * B)^t -> B^t * A^t
806     // RxK KxC      CxK   KxR
807     if (match(TA, m_Intrinsic<Intrinsic::matrix_multiply>(
808                       m_Value(TAMA), m_Value(TAMB), m_ConstantInt(R),
809                       m_ConstantInt(K), m_ConstantInt(C)))) {
810       auto NewInst = distributeTransposes(
811           TAMB, {K, C}, TAMA, {R, K}, Builder,
812           [&](Value *T0, ShapeInfo Shape0, Value *T1, ShapeInfo Shape1) {
813             return Builder.CreateMatrixMultiply(T0, T1, Shape0.NumRows,
814                                                 Shape0.NumColumns,
815                                                 Shape1.NumColumns, "mmul");
816           });
817       updateShapeAndReplaceAllUsesWith(I, NewInst);
818       eraseFromParentAndMove(&I, II, BB);
819       eraseFromParentAndMove(TA, II, BB);
820       return NewInst;
821     }
822 
823     // Same as above, but with a mul, which occurs when multiplied
824     // with a scalar.
825     // (A * k)^t -> A^t * k
826     //  R  x  C     RxC
827     if (match(TA, m_AnyMul(m_Value(TAMA), m_Value(TAMB))) &&
828         (isSplat(TAMA) || isSplat(TAMB))) {
829       IRBuilder<> LocalBuilder(&I);
830       // We know that the transposed operand is of shape RxC.
831       // An when multiplied with a scalar, the shape is preserved.
832       auto NewInst = distributeTransposes(
833           TAMA, {R, C}, TAMB, {R, C}, Builder,
834           [&](Value *T0, ShapeInfo Shape0, Value *T1, ShapeInfo Shape1) {
835             bool IsFP = I.getType()->isFPOrFPVectorTy();
836             auto *Mul = IsFP ? LocalBuilder.CreateFMul(T0, T1, "mmul")
837                              : LocalBuilder.CreateMul(T0, T1, "mmul");
838             auto *Result = cast<Instruction>(Mul);
839             setShapeInfo(Result, Shape0);
840             return Result;
841           });
842       updateShapeAndReplaceAllUsesWith(I, NewInst);
843       eraseFromParentAndMove(&I, II, BB);
844       eraseFromParentAndMove(TA, II, BB);
845       return NewInst;
846     }
847 
848     // (A + B)^t -> A^t + B^t
849     // RxC RxC      CxR   CxR
850     if (match(TA, m_AnyAdd(m_Value(TAMA), m_Value(TAMB)))) {
851       IRBuilder<> LocalBuilder(&I);
852       auto NewInst = distributeTransposes(
853           TAMA, {R, C}, TAMB, {R, C}, Builder,
854           [&](Value *T0, ShapeInfo Shape0, Value *T1, ShapeInfo Shape1) {
855             bool IsFP = I.getType()->isFPOrFPVectorTy();
856             auto *Add = IsFP ? LocalBuilder.CreateFAdd(T0, T1, "madd")
857                              : LocalBuilder.CreateAdd(T0, T1, "madd");
858 
859             auto *Result = cast<Instruction>(Add);
860             setShapeInfo(Result, Shape0);
861             return Result;
862           });
863       updateShapeAndReplaceAllUsesWith(I, NewInst);
864       eraseFromParentAndMove(&I, II, BB);
865       eraseFromParentAndMove(TA, II, BB);
866       return NewInst;
867     }
868 
869     return nullptr;
870   }
871 
872   void liftTranspose(Instruction &I) {
873     // Erase dead Instructions after lifting transposes from binops.
874     auto CleanupBinOp = [](Instruction &T, Value *A, Value *B) {
875       if (T.use_empty())
876         T.eraseFromParent();
877       if (A->use_empty())
878         cast<Instruction>(A)->eraseFromParent();
879       if (A != B && B->use_empty())
880         cast<Instruction>(B)->eraseFromParent();
881     };
882 
883     Value *A, *B, *AT, *BT;
884     ConstantInt *R, *K, *C;
885     // A^t * B ^t -> (B * A)^t
886     if (match(&I, m_Intrinsic<Intrinsic::matrix_multiply>(
887                       m_Value(A), m_Value(B), m_ConstantInt(R),
888                       m_ConstantInt(K), m_ConstantInt(C))) &&
889         match(A, m_Intrinsic<Intrinsic::matrix_transpose>(m_Value(AT))) &&
890         match(B, m_Intrinsic<Intrinsic::matrix_transpose>(m_Value((BT))))) {
891       IRBuilder<> IB(&I);
892       MatrixBuilder Builder(IB);
893       Value *M = Builder.CreateMatrixMultiply(
894           BT, AT, C->getZExtValue(), K->getZExtValue(), R->getZExtValue());
895       setShapeInfo(M, {C, R});
896       Instruction *NewInst = Builder.CreateMatrixTranspose(M, C->getZExtValue(),
897                                                            R->getZExtValue());
898       updateShapeAndReplaceAllUsesWith(I, NewInst);
899       CleanupBinOp(I, A, B);
900     }
901     // A^t + B ^t -> (A + B)^t
902     else if (match(&I, m_FAdd(m_Value(A), m_Value(B))) &&
903              match(A, m_Intrinsic<Intrinsic::matrix_transpose>(
904                           m_Value(AT), m_ConstantInt(R), m_ConstantInt(C))) &&
905              match(B, m_Intrinsic<Intrinsic::matrix_transpose>(
906                           m_Value(BT), m_ConstantInt(R), m_ConstantInt(C)))) {
907       IRBuilder<> Builder(&I);
908       Value *Add = cast<Instruction>(Builder.CreateFAdd(AT, BT, "mfadd"));
909       setShapeInfo(Add, {C, R});
910       MatrixBuilder MBuilder(Builder);
911       Instruction *NewInst = MBuilder.CreateMatrixTranspose(
912           Add, C->getZExtValue(), R->getZExtValue(), "mfadd_t");
913       updateShapeAndReplaceAllUsesWith(I, NewInst);
914       CleanupBinOp(I, A, B);
915     }
916   }
917 
918   /// Try moving transposes in order to fold them away or into multiplies.
919   void optimizeTransposes() {
920     // First sink all transposes inside matmuls and adds, hoping that we end up
921     // with NN, NT or TN variants.
922     for (BasicBlock &BB : reverse(Func)) {
923       for (auto II = BB.rbegin(); II != BB.rend();) {
924         Instruction &I = *II;
925         // We may remove II.  By default continue on the next/prev instruction.
926         ++II;
927         if (Instruction *NewInst = sinkTranspose(I, II))
928           II = std::next(BasicBlock::reverse_iterator(NewInst));
929       }
930     }
931 
932     // If we have a TT matmul or a TT add, lift the transpose. We may be able
933     // to fold into consuming multiply or add.
934     for (BasicBlock &BB : Func) {
935       for (Instruction &I : llvm::make_early_inc_range(BB)) {
936         liftTranspose(I);
937       }
938     }
939   }
940 
941   bool Visit() {
942     SmallVector<Instruction *, 32> WorkList;
943 
944     // Initially only the shape of matrix intrinsics is known.
945     // Initialize the work list with ops carrying shape information.
946     for (BasicBlock &BB : Func)
947       for (Instruction &Inst : BB) {
948         IntrinsicInst *II = dyn_cast<IntrinsicInst>(&Inst);
949         if (!II)
950           continue;
951 
952         switch (II->getIntrinsicID()) {
953         case Intrinsic::matrix_multiply:
954         case Intrinsic::matrix_transpose:
955         case Intrinsic::matrix_column_major_load:
956         case Intrinsic::matrix_column_major_store:
957           WorkList.push_back(&Inst);
958           break;
959         default:
960           break;
961         }
962       }
963 
964     // Avoid unnecessary work if there are no matrix intrinsics in the function.
965     if (WorkList.empty())
966       return false;
967 
968     // Propagate shapes until nothing changes any longer.
969     while (!WorkList.empty()) {
970       WorkList = propagateShapeForward(WorkList);
971       WorkList = propagateShapeBackward(WorkList);
972     }
973 
974     if (!isMinimal()) {
975       optimizeTransposes();
976       if (PrintAfterTransposeOpt) {
977         dbgs() << "Dump after matrix transpose optimization:\n";
978         Func.print(dbgs());
979       }
980     }
981 
982     bool Changed = false;
983     SmallVector<CallInst *, 16> MaybeFusableInsts;
984     SmallVector<Instruction *, 16> MatrixInsts;
985 
986     // First, collect all instructions with shape information and candidates for
987     // fusion (currently only matrix multiplies).
988     ReversePostOrderTraversal<Function *> RPOT(&Func);
989     for (auto *BB : RPOT)
990       for (Instruction &I : *BB) {
991         if (ShapeMap.find(&I) == ShapeMap.end())
992           continue;
993         if (match(&I, m_Intrinsic<Intrinsic::matrix_multiply>()))
994           MaybeFusableInsts.push_back(cast<CallInst>(&I));
995         MatrixInsts.push_back(&I);
996       }
997 
998     // Second, try to lower any dot products
999     SmallPtrSet<Instruction *, 16> FusedInsts;
1000     for (CallInst *CI : MaybeFusableInsts)
1001       lowerDotProduct(CI, FusedInsts, getFastMathFlags(CI));
1002 
1003     // Third, try to fuse candidates.
1004     for (CallInst *CI : MaybeFusableInsts)
1005       LowerMatrixMultiplyFused(CI, FusedInsts);
1006 
1007     Changed = !FusedInsts.empty();
1008 
1009     // Fourth, lower remaining instructions with shape information.
1010     for (Instruction *Inst : MatrixInsts) {
1011       if (FusedInsts.count(Inst))
1012         continue;
1013 
1014       IRBuilder<> Builder(Inst);
1015 
1016       if (CallInst *CInst = dyn_cast<CallInst>(Inst))
1017         Changed |= VisitCallInst(CInst);
1018 
1019       Value *Op1;
1020       Value *Op2;
1021       if (auto *BinOp = dyn_cast<BinaryOperator>(Inst))
1022         Changed |= VisitBinaryOperator(BinOp);
1023       if (auto *UnOp = dyn_cast<UnaryOperator>(Inst))
1024         Changed |= VisitUnaryOperator(UnOp);
1025       if (match(Inst, m_Load(m_Value(Op1))))
1026         Changed |= VisitLoad(cast<LoadInst>(Inst), Op1, Builder);
1027       else if (match(Inst, m_Store(m_Value(Op1), m_Value(Op2))))
1028         Changed |= VisitStore(cast<StoreInst>(Inst), Op1, Op2, Builder);
1029     }
1030 
1031     if (ORE) {
1032       RemarkGenerator RemarkGen(Inst2ColumnMatrix, *ORE, Func);
1033       RemarkGen.emitRemarks();
1034     }
1035 
1036     // Delete the instructions backwards, as it has a reduced likelihood of
1037     // having to update as many def-use and use-def chains.
1038     //
1039     // Because we add to ToRemove during fusion we can't guarantee that defs
1040     // are before uses.  Change uses to poison temporarily as these should get
1041     // removed as well.
1042     //
1043     // For verification, we keep track of where we changed uses to poison in
1044     // PoisonedInsts and then check that we in fact remove them.
1045     SmallSet<Instruction *, 16> PoisonedInsts;
1046     for (auto *Inst : reverse(ToRemove)) {
1047       for (Use &U : llvm::make_early_inc_range(Inst->uses())) {
1048         if (auto *Poisoned = dyn_cast<Instruction>(U.getUser()))
1049           PoisonedInsts.insert(Poisoned);
1050         U.set(PoisonValue::get(Inst->getType()));
1051       }
1052       Inst->eraseFromParent();
1053       PoisonedInsts.erase(Inst);
1054     }
1055     if (!PoisonedInsts.empty()) {
1056       // If we didn't remove all poisoned instructions, it's a hard error.
1057       dbgs() << "Poisoned but present instructions:\n";
1058       for (auto *I : PoisonedInsts)
1059         dbgs() << *I << "\n";
1060       llvm_unreachable("Poisoned but instruction not removed");
1061     }
1062 
1063     return Changed;
1064   }
1065 
1066   /// Turns \p BasePtr into an elementwise pointer to \p EltType.
1067   Value *createElementPtr(Value *BasePtr, Type *EltType, IRBuilder<> &Builder) {
1068     unsigned AS = cast<PointerType>(BasePtr->getType())->getAddressSpace();
1069     Type *EltPtrType = PointerType::get(EltType, AS);
1070     return Builder.CreatePointerCast(BasePtr, EltPtrType);
1071   }
1072 
1073   /// Replace intrinsic calls
1074   bool VisitCallInst(CallInst *Inst) {
1075     if (!Inst->getCalledFunction() || !Inst->getCalledFunction()->isIntrinsic())
1076       return false;
1077 
1078     switch (Inst->getCalledFunction()->getIntrinsicID()) {
1079     case Intrinsic::matrix_multiply:
1080       LowerMultiply(Inst);
1081       break;
1082     case Intrinsic::matrix_transpose:
1083       LowerTranspose(Inst);
1084       break;
1085     case Intrinsic::matrix_column_major_load:
1086       LowerColumnMajorLoad(Inst);
1087       break;
1088     case Intrinsic::matrix_column_major_store:
1089       LowerColumnMajorStore(Inst);
1090       break;
1091     default:
1092       return false;
1093     }
1094     return true;
1095   }
1096 
1097   /// Compute the alignment for a column/row \p Idx with \p Stride between them.
1098   /// The address at \p Idx == 0 has alignment \p A. If \p Stride is a
1099   /// ConstantInt, reduce the initial alignment based on the byte offset. For
1100   /// non-ConstantInt strides, return the common alignment of the initial
1101   /// alignment and the element size in bytes.
1102   Align getAlignForIndex(unsigned Idx, Value *Stride, Type *ElementTy,
1103                          MaybeAlign A) const {
1104     Align InitialAlign = DL.getValueOrABITypeAlignment(A, ElementTy);
1105     if (Idx == 0)
1106       return InitialAlign;
1107 
1108     TypeSize ElementSizeInBits = DL.getTypeSizeInBits(ElementTy);
1109     if (auto *ConstStride = dyn_cast<ConstantInt>(Stride)) {
1110       uint64_t StrideInBytes =
1111           ConstStride->getZExtValue() * ElementSizeInBits / 8;
1112       return commonAlignment(InitialAlign, Idx * StrideInBytes);
1113     }
1114     return commonAlignment(InitialAlign, ElementSizeInBits / 8);
1115   }
1116 
1117   /// Load a matrix with \p Shape starting at \p Ptr and using \p Stride between
1118   /// vectors.
1119   MatrixTy loadMatrix(Type *Ty, Value *Ptr, MaybeAlign MAlign, Value *Stride,
1120                       bool IsVolatile, ShapeInfo Shape, IRBuilder<> &Builder) {
1121     auto *VType = cast<VectorType>(Ty);
1122     Type *EltTy = VType->getElementType();
1123     Type *VecTy = FixedVectorType::get(EltTy, Shape.getStride());
1124     Value *EltPtr = createElementPtr(Ptr, EltTy, Builder);
1125     MatrixTy Result;
1126     for (unsigned I = 0, E = Shape.getNumVectors(); I < E; ++I) {
1127       Value *GEP = computeVectorAddr(
1128           EltPtr, Builder.getIntN(Stride->getType()->getScalarSizeInBits(), I),
1129           Stride, Shape.getStride(), EltTy, Builder);
1130       Value *Vector = Builder.CreateAlignedLoad(
1131           VecTy, GEP, getAlignForIndex(I, Stride, EltTy, MAlign),
1132           IsVolatile, "col.load");
1133 
1134       Result.addVector(Vector);
1135     }
1136     return Result.addNumLoads(getNumOps(Result.getVectorTy()) *
1137                               Result.getNumVectors());
1138   }
1139 
1140   /// Loads a sub-matrix with shape \p ResultShape from a \p R x \p C matrix,
1141   /// starting at \p MatrixPtr[I][J].
1142   MatrixTy loadMatrix(Value *MatrixPtr, MaybeAlign Align, bool IsVolatile,
1143                       ShapeInfo MatrixShape, Value *I, Value *J,
1144                       ShapeInfo ResultShape, Type *EltTy,
1145                       IRBuilder<> &Builder) {
1146 
1147     Value *Offset = Builder.CreateAdd(
1148         Builder.CreateMul(J, Builder.getInt64(MatrixShape.getStride())), I);
1149 
1150     unsigned AS = cast<PointerType>(MatrixPtr->getType())->getAddressSpace();
1151     Value *EltPtr =
1152         Builder.CreatePointerCast(MatrixPtr, PointerType::get(EltTy, AS));
1153     Value *TileStart = Builder.CreateGEP(EltTy, EltPtr, Offset);
1154     auto *TileTy = FixedVectorType::get(EltTy, ResultShape.NumRows *
1155                                                    ResultShape.NumColumns);
1156     Type *TilePtrTy = PointerType::get(TileTy, AS);
1157     Value *TilePtr =
1158         Builder.CreatePointerCast(TileStart, TilePtrTy, "col.cast");
1159 
1160     return loadMatrix(TileTy, TilePtr, Align,
1161                       Builder.getInt64(MatrixShape.getStride()), IsVolatile,
1162                       ResultShape, Builder);
1163   }
1164 
1165   /// Lower a load instruction with shape information.
1166   void LowerLoad(Instruction *Inst, Value *Ptr, MaybeAlign Align, Value *Stride,
1167                  bool IsVolatile, ShapeInfo Shape) {
1168     IRBuilder<> Builder(Inst);
1169     finalizeLowering(Inst,
1170                      loadMatrix(Inst->getType(), Ptr, Align, Stride, IsVolatile,
1171                                 Shape, Builder),
1172                      Builder);
1173   }
1174 
1175   /// Lowers llvm.matrix.column.major.load.
1176   ///
1177   /// The intrinsic loads a matrix from memory using a stride between columns.
1178   void LowerColumnMajorLoad(CallInst *Inst) {
1179     assert(MatrixLayout == MatrixLayoutTy::ColumnMajor &&
1180            "Intrinsic only supports column-major layout!");
1181     Value *Ptr = Inst->getArgOperand(0);
1182     Value *Stride = Inst->getArgOperand(1);
1183     LowerLoad(Inst, Ptr, Inst->getParamAlign(0), Stride,
1184               cast<ConstantInt>(Inst->getArgOperand(2))->isOne(),
1185               {Inst->getArgOperand(3), Inst->getArgOperand(4)});
1186   }
1187 
1188   /// Stores a sub-matrix \p StoreVal into the \p R x \p C matrix starting at \p
1189   /// MatrixPtr[I][J].
1190   void storeMatrix(const MatrixTy &StoreVal, Value *MatrixPtr,
1191                    MaybeAlign MAlign, bool IsVolatile, ShapeInfo MatrixShape,
1192                    Value *I, Value *J, Type *EltTy, IRBuilder<> &Builder) {
1193     Value *Offset = Builder.CreateAdd(
1194         Builder.CreateMul(J, Builder.getInt64(MatrixShape.getStride())), I);
1195 
1196     unsigned AS = cast<PointerType>(MatrixPtr->getType())->getAddressSpace();
1197     Value *EltPtr =
1198         Builder.CreatePointerCast(MatrixPtr, PointerType::get(EltTy, AS));
1199     Value *TileStart = Builder.CreateGEP(EltTy, EltPtr, Offset);
1200     auto *TileTy = FixedVectorType::get(EltTy, StoreVal.getNumRows() *
1201                                                    StoreVal.getNumColumns());
1202     Type *TilePtrTy = PointerType::get(TileTy, AS);
1203     Value *TilePtr =
1204         Builder.CreatePointerCast(TileStart, TilePtrTy, "col.cast");
1205 
1206     storeMatrix(TileTy, StoreVal, TilePtr, MAlign,
1207                 Builder.getInt64(MatrixShape.getStride()), IsVolatile, Builder);
1208   }
1209 
1210   /// Store matrix \p StoreVal starting at \p Ptr and using \p Stride between
1211   /// vectors.
1212   MatrixTy storeMatrix(Type *Ty, MatrixTy StoreVal, Value *Ptr,
1213                        MaybeAlign MAlign, Value *Stride, bool IsVolatile,
1214                        IRBuilder<> &Builder) {
1215     auto VType = cast<VectorType>(Ty);
1216     Value *EltPtr = createElementPtr(Ptr, VType->getElementType(), Builder);
1217     for (auto Vec : enumerate(StoreVal.vectors())) {
1218       Value *GEP = computeVectorAddr(
1219           EltPtr,
1220           Builder.getIntN(Stride->getType()->getScalarSizeInBits(),
1221                           Vec.index()),
1222           Stride, StoreVal.getStride(), VType->getElementType(), Builder);
1223       Builder.CreateAlignedStore(Vec.value(), GEP,
1224                                  getAlignForIndex(Vec.index(), Stride,
1225                                                   VType->getElementType(),
1226                                                   MAlign),
1227                                  IsVolatile);
1228     }
1229     return MatrixTy().addNumStores(getNumOps(StoreVal.getVectorTy()) *
1230                                    StoreVal.getNumVectors());
1231   }
1232 
1233   /// Lower a store instruction with shape information.
1234   void LowerStore(Instruction *Inst, Value *Matrix, Value *Ptr, MaybeAlign A,
1235                   Value *Stride, bool IsVolatile, ShapeInfo Shape) {
1236     IRBuilder<> Builder(Inst);
1237     auto StoreVal = getMatrix(Matrix, Shape, Builder);
1238     finalizeLowering(Inst,
1239                      storeMatrix(Matrix->getType(), StoreVal, Ptr, A, Stride,
1240                                  IsVolatile, Builder),
1241                      Builder);
1242   }
1243 
1244   /// Lowers llvm.matrix.column.major.store.
1245   ///
1246   /// The intrinsic store a matrix back memory using a stride between columns.
1247   void LowerColumnMajorStore(CallInst *Inst) {
1248     assert(MatrixLayout == MatrixLayoutTy::ColumnMajor &&
1249            "Intrinsic only supports column-major layout!");
1250     Value *Matrix = Inst->getArgOperand(0);
1251     Value *Ptr = Inst->getArgOperand(1);
1252     Value *Stride = Inst->getArgOperand(2);
1253     LowerStore(Inst, Matrix, Ptr, Inst->getParamAlign(1), Stride,
1254                cast<ConstantInt>(Inst->getArgOperand(3))->isOne(),
1255                {Inst->getArgOperand(4), Inst->getArgOperand(5)});
1256   }
1257 
1258   // Set elements I..I+NumElts-1 to Block
1259   Value *insertVector(Value *Col, unsigned I, Value *Block,
1260                       IRBuilder<> &Builder) {
1261 
1262     // First, bring Block to the same size as Col
1263     unsigned BlockNumElts =
1264         cast<FixedVectorType>(Block->getType())->getNumElements();
1265     unsigned NumElts = cast<FixedVectorType>(Col->getType())->getNumElements();
1266     assert(NumElts >= BlockNumElts && "Too few elements for current block");
1267 
1268     Block = Builder.CreateShuffleVector(
1269         Block, createSequentialMask(0, BlockNumElts, NumElts - BlockNumElts));
1270 
1271     // If Col is 7 long and I is 2 and BlockNumElts is 2 the mask is: 0, 1, 7,
1272     // 8, 4, 5, 6
1273     SmallVector<int, 16> Mask;
1274     unsigned i;
1275     for (i = 0; i < I; i++)
1276       Mask.push_back(i);
1277 
1278     unsigned VecNumElts =
1279         cast<FixedVectorType>(Col->getType())->getNumElements();
1280     for (; i < I + BlockNumElts; i++)
1281       Mask.push_back(i - I + VecNumElts);
1282 
1283     for (; i < VecNumElts; i++)
1284       Mask.push_back(i);
1285 
1286     return Builder.CreateShuffleVector(Col, Block, Mask);
1287   }
1288 
1289   Value *createMulAdd(Value *Sum, Value *A, Value *B, bool UseFPOp,
1290                       IRBuilder<> &Builder, bool AllowContraction,
1291                       unsigned &NumComputeOps) {
1292     NumComputeOps += getNumOps(A->getType());
1293     if (!Sum)
1294       return UseFPOp ? Builder.CreateFMul(A, B) : Builder.CreateMul(A, B);
1295 
1296     if (UseFPOp) {
1297       if (AllowContraction) {
1298         // Use fmuladd for floating point operations and let the backend decide
1299         // if that's profitable.
1300         Function *FMulAdd = Intrinsic::getDeclaration(
1301             Func.getParent(), Intrinsic::fmuladd, A->getType());
1302         return Builder.CreateCall(FMulAdd, {A, B, Sum});
1303       }
1304       NumComputeOps += getNumOps(A->getType());
1305       Value *Mul = Builder.CreateFMul(A, B);
1306       return Builder.CreateFAdd(Sum, Mul);
1307     }
1308 
1309     NumComputeOps += getNumOps(A->getType());
1310     Value *Mul = Builder.CreateMul(A, B);
1311     return Builder.CreateAdd(Sum, Mul);
1312   }
1313 
1314   /// Cache \p Matrix as result of \p Inst and update the uses of \p Inst. For
1315   /// users with shape information, there's nothing to do: they will use the
1316   /// cached value when they are lowered. For other users, \p Matrix is
1317   /// flattened and the uses are updated to use it. Also marks \p Inst for
1318   /// deletion.
1319   void finalizeLowering(Instruction *Inst, MatrixTy Matrix,
1320                         IRBuilder<> &Builder) {
1321     auto inserted = Inst2ColumnMatrix.insert(std::make_pair(Inst, Matrix));
1322     (void)inserted;
1323     assert(inserted.second && "multiple matrix lowering mapping");
1324 
1325     ToRemove.push_back(Inst);
1326     Value *Flattened = nullptr;
1327     for (Use &U : llvm::make_early_inc_range(Inst->uses())) {
1328       if (ShapeMap.find(U.getUser()) == ShapeMap.end()) {
1329         if (!Flattened)
1330           Flattened = Matrix.embedInVector(Builder);
1331         U.set(Flattened);
1332       }
1333     }
1334   }
1335 
1336   /// Special case for MatMul lowering. Prevents scalar loads of row-major
1337   /// vectors Lowers to vector reduction add instead of sequential add if
1338   /// reassocation is enabled.
1339   void lowerDotProduct(CallInst *MatMul,
1340                        SmallPtrSet<Instruction *, 16> &FusedInsts,
1341                        FastMathFlags FMF) {
1342     if (FusedInsts.contains(MatMul) ||
1343         MatrixLayout != MatrixLayoutTy::ColumnMajor)
1344       return;
1345     ShapeInfo LShape(MatMul->getArgOperand(2), MatMul->getArgOperand(3));
1346     ShapeInfo RShape(MatMul->getArgOperand(3), MatMul->getArgOperand(4));
1347 
1348     if (LShape.NumRows != 1 || RShape.NumColumns != 1) // not a dot product
1349       return;
1350 
1351     Value *LHS = MatMul->getArgOperand(0);
1352     Value *RHS = MatMul->getArgOperand(1);
1353 
1354     Type *ElementType = cast<VectorType>(LHS->getType())->getElementType();
1355     bool IsIntVec = ElementType->isIntegerTy();
1356 
1357     // Floating point reductions require reassocation.
1358     if (!IsIntVec && !FMF.allowReassoc())
1359       return;
1360 
1361     auto CanBeFlattened = [this](Value *Op) {
1362       if (match(Op, m_BinOp()) && ShapeMap.find(Op) != ShapeMap.end())
1363         return true;
1364       return match(
1365           Op, m_OneUse(m_CombineOr(
1366                   m_Load(m_Value()),
1367                   m_CombineOr(m_Intrinsic<Intrinsic::matrix_transpose>(),
1368                               m_Intrinsic<Intrinsic::matrix_column_major_load>(
1369                                   m_Value(), m_SpecificInt(1))))));
1370     };
1371     // Returns the cost benefit of using \p Op with the dot product lowering. If
1372     // the returned cost is < 0, the argument is cheaper to use in the
1373     // dot-product lowering.
1374     auto GetCostForArg = [this, &CanBeFlattened](Value *Op, unsigned N) {
1375       if (!isa<Instruction>(Op))
1376         return InstructionCost(0);
1377 
1378       FixedVectorType *VecTy = cast<FixedVectorType>(Op->getType());
1379       Type *EltTy = VecTy->getElementType();
1380 
1381       if (!CanBeFlattened(Op)) {
1382         InstructionCost EmbedCost(0);
1383         // Roughly estimate the cost for embedding the columns into a vector.
1384         for (unsigned I = 1; I < N; ++I)
1385           EmbedCost -=
1386               TTI.getShuffleCost(TTI::SK_Splice, FixedVectorType::get(EltTy, 1),
1387                                  std::nullopt, TTI::TCK_RecipThroughput);
1388         return EmbedCost;
1389       }
1390 
1391       if (match(Op, m_BinOp()) && ShapeMap.find(Op) != ShapeMap.end()) {
1392         InstructionCost OriginalCost =
1393             TTI.getArithmeticInstrCost(cast<Instruction>(Op)->getOpcode(),
1394                                        EltTy) *
1395             N;
1396         InstructionCost NewCost = TTI.getArithmeticInstrCost(
1397             cast<Instruction>(Op)->getOpcode(), VecTy);
1398         return NewCost - OriginalCost;
1399       }
1400 
1401       if (match(Op, m_Intrinsic<Intrinsic::matrix_transpose>())) {
1402         // The transpose can be skipped for the dot product lowering, roughly
1403         // estimate the savings as the cost of embedding the columns in a
1404         // vector.
1405         InstructionCost EmbedCost(0);
1406         for (unsigned I = 1; I < N; ++I)
1407           EmbedCost +=
1408               TTI.getShuffleCost(TTI::SK_Splice, FixedVectorType::get(EltTy, 1),
1409                                  std::nullopt, TTI::TCK_RecipThroughput);
1410         return EmbedCost;
1411       }
1412 
1413       // Costs for loads.
1414       if (N == 1)
1415         return InstructionCost(0);
1416 
1417       return TTI.getMemoryOpCost(Instruction::Load, VecTy, Align(1), 0) -
1418              N * TTI.getMemoryOpCost(Instruction::Load, EltTy, Align(1), 0);
1419     };
1420     auto LHSCost = GetCostForArg(LHS, LShape.NumColumns);
1421 
1422     // We compare the costs of a vector.reduce.add to sequential add.
1423     int AddOpCode = IsIntVec ? Instruction::Add : Instruction::FAdd;
1424     int MulOpCode = IsIntVec ? Instruction::Mul : Instruction::FMul;
1425     InstructionCost ReductionCost =
1426         TTI.getArithmeticReductionCost(
1427             AddOpCode, cast<VectorType>(LHS->getType()),
1428             IsIntVec ? std::nullopt : std::optional(FMF)) +
1429         TTI.getArithmeticInstrCost(MulOpCode, LHS->getType());
1430     InstructionCost SequentialAddCost =
1431         TTI.getArithmeticInstrCost(AddOpCode, ElementType) *
1432             (LShape.NumColumns - 1) +
1433         TTI.getArithmeticInstrCost(MulOpCode, ElementType) *
1434             (LShape.NumColumns);
1435     if ((LHSCost + ReductionCost - SequentialAddCost) > InstructionCost(0))
1436       return;
1437 
1438     FusedInsts.insert(MatMul);
1439     IRBuilder<> Builder(MatMul);
1440     auto FlattenArg = [&Builder, &FusedInsts, &CanBeFlattened,
1441                        this](Value *Op) -> Value * {
1442       // Matmul must be the only user of loads because we don't use LowerLoad
1443       // for row vectors (LowerLoad results in scalar loads and shufflevectors
1444       // instead of single vector load).
1445       if (!CanBeFlattened(Op))
1446         return Op;
1447 
1448       if (match(Op, m_BinOp()) && ShapeMap.find(Op) != ShapeMap.end()) {
1449         ShapeMap[Op] = ShapeMap[Op].t();
1450         return Op;
1451       }
1452 
1453       FusedInsts.insert(cast<Instruction>(Op));
1454       // If vector uses the builtin load, lower to a LoadInst
1455       Value *Arg;
1456       if (match(Op, m_Intrinsic<Intrinsic::matrix_column_major_load>(
1457                         m_Value(Arg)))) {
1458         auto *NewLoad = Builder.CreateLoad(Op->getType(), Arg);
1459         Op->replaceAllUsesWith(NewLoad);
1460         cast<Instruction>(Op)->eraseFromParent();
1461         return NewLoad;
1462       } else if (match(Op, m_Intrinsic<Intrinsic::matrix_transpose>(
1463                                m_Value(Arg)))) {
1464         ToRemove.push_back(cast<Instruction>(Op));
1465         return Arg;
1466       }
1467 
1468       return Op;
1469     };
1470     LHS = FlattenArg(LHS);
1471 
1472     // Insert mul/fmul and llvm.vector.reduce.fadd
1473     Value *Mul =
1474         IsIntVec ? Builder.CreateMul(LHS, RHS) : Builder.CreateFMul(LHS, RHS);
1475 
1476     Value *Result;
1477     if (IsIntVec)
1478       Result = Builder.CreateAddReduce(Mul);
1479     else {
1480       Result = Builder.CreateFAddReduce(
1481           ConstantFP::get(cast<VectorType>(LHS->getType())->getElementType(),
1482                           0.0),
1483           Mul);
1484       cast<Instruction>(Result)->setFastMathFlags(FMF);
1485     }
1486 
1487     // pack scalar back into a matrix and then replace matmul inst
1488     Result = Builder.CreateInsertElement(PoisonValue::get(MatMul->getType()),
1489                                          Result, uint64_t(0));
1490     MatMul->replaceAllUsesWith(Result);
1491     FusedInsts.insert(MatMul);
1492     ToRemove.push_back(MatMul);
1493   }
1494 
1495   /// Compute \p Result += \p A * \p B for input matrices with left-associating
1496   /// addition.
1497   ///
1498   /// We can fold a transpose into the operand that is used to extract scalars.
1499   /// This is the first operands with row-major and the second with
1500   /// column-major.  If \p IsScalarMatrixTransposed we assume the appropriate
1501   /// operand is transposed.
1502   void emitMatrixMultiply(MatrixTy &Result, const MatrixTy &A,
1503                           const MatrixTy &B, IRBuilder<> &Builder, bool IsTiled,
1504                           bool IsScalarMatrixTransposed, FastMathFlags FMF) {
1505     const unsigned VF = std::max<unsigned>(
1506         TTI.getRegisterBitWidth(TargetTransformInfo::RGK_FixedWidthVector)
1507                 .getFixedValue() /
1508             Result.getElementType()->getPrimitiveSizeInBits().getFixedValue(),
1509         1U);
1510     unsigned R = Result.getNumRows();
1511     unsigned C = Result.getNumColumns();
1512     unsigned M = A.getNumColumns();
1513 
1514     bool IsFP = Result.getElementType()->isFloatingPointTy();
1515     assert(A.isColumnMajor() == B.isColumnMajor() &&
1516            Result.isColumnMajor() == A.isColumnMajor() &&
1517            "operands must agree on matrix layout");
1518     unsigned NumComputeOps = 0;
1519 
1520     Builder.setFastMathFlags(FMF);
1521 
1522     if (A.isColumnMajor()) {
1523       // Multiply columns from the first operand with scalars from the second
1524       // operand. Then move along the K axes and accumulate the columns.  With
1525       // this the adds can be vectorized without reassociation.
1526       for (unsigned J = 0; J < C; ++J) {
1527         unsigned BlockSize = VF;
1528         // If Result is zero, we don't need to accumulate in the K==0 iteration.
1529         bool isSumZero = isa<ConstantAggregateZero>(Result.getColumn(J));
1530 
1531         for (unsigned I = 0; I < R; I += BlockSize) {
1532           // Gradually lower the vectorization factor to cover the remainder.
1533           while (I + BlockSize > R)
1534             BlockSize /= 2;
1535 
1536           Value *Sum = IsTiled ? Result.extractVector(I, J, BlockSize, Builder)
1537                                : nullptr;
1538           for (unsigned K = 0; K < M; ++K) {
1539             Value *L = A.extractVector(I, K, BlockSize, Builder);
1540             Value *RH = Builder.CreateExtractElement(
1541                 B.getColumn(IsScalarMatrixTransposed ? K : J),
1542                 IsScalarMatrixTransposed ? J : K);
1543             Value *Splat = Builder.CreateVectorSplat(BlockSize, RH, "splat");
1544             Sum =
1545                 createMulAdd(isSumZero && K == 0 ? nullptr : Sum, L, Splat,
1546                              IsFP, Builder, FMF.allowContract(), NumComputeOps);
1547           }
1548           Result.setVector(J,
1549                            insertVector(Result.getVector(J), I, Sum, Builder));
1550         }
1551       }
1552     } else {
1553       // Multiply rows from the second operand with scalars from the first
1554       // operand. Then move along the K axes and accumulate the rows.  With this
1555       // the adds can be vectorized without reassociation.
1556       for (unsigned I = 0; I < R; ++I) {
1557         unsigned BlockSize = VF;
1558         bool isSumZero = isa<ConstantAggregateZero>(Result.getRow(I));
1559         for (unsigned J = 0; J < C; J += BlockSize) {
1560           // Gradually lower the vectorization factor to cover the remainder.
1561           while (J + BlockSize > C)
1562             BlockSize /= 2;
1563 
1564           Value *Sum = nullptr;
1565           for (unsigned K = 0; K < M; ++K) {
1566             Value *R = B.extractVector(K, J, BlockSize, Builder);
1567             Value *LH = Builder.CreateExtractElement(
1568                 A.getVector(IsScalarMatrixTransposed ? K : I),
1569                 IsScalarMatrixTransposed ? I : K);
1570             Value *Splat = Builder.CreateVectorSplat(BlockSize, LH, "splat");
1571             Sum =
1572                 createMulAdd(isSumZero && K == 0 ? nullptr : Sum, Splat, R,
1573                              IsFP, Builder, FMF.allowContract(), NumComputeOps);
1574           }
1575           Result.setVector(I,
1576                            insertVector(Result.getVector(I), J, Sum, Builder));
1577         }
1578       }
1579     }
1580     Result.addNumComputeOps(NumComputeOps);
1581   }
1582 
1583   /// Ensure that the memory in \p Load does not alias \p Store by potentially
1584   /// copying it to a new location.  This new or otherwise the original location
1585   /// is returned.
1586   Value *getNonAliasingPointer(LoadInst *Load, StoreInst *Store,
1587                                CallInst *MatMul) {
1588     MemoryLocation StoreLoc = MemoryLocation::get(Store);
1589     MemoryLocation LoadLoc = MemoryLocation::get(Load);
1590 
1591     // If we can statically determine noalias we're good.
1592     if (AA->isNoAlias(LoadLoc, StoreLoc))
1593       return Load->getPointerOperand();
1594 
1595     // Create code to check if the memory locations of the Load and Store
1596     // overlap and if they do, copy Load's operand to a new buffer.
1597 
1598     // First, create  new blocks for 2n part of the check and the copy.
1599     BasicBlock *Check0 = MatMul->getParent();
1600     // FIXME: Use lazy DTU and update SplitBlock to accept a DTU instead of a
1601     // DT. Manually collect dominator tree updates, to avoid unnecessary work,
1602     // as we adjust Check0 and Check1's branches.
1603     SmallVector<DominatorTree::UpdateType, 4> DTUpdates;
1604     for (BasicBlock *Succ : successors(Check0))
1605       DTUpdates.push_back({DT->Delete, Check0, Succ});
1606 
1607     BasicBlock *Check1 =
1608         SplitBlock(MatMul->getParent(), MatMul, (DomTreeUpdater *)nullptr, LI,
1609                    nullptr, "alias_cont");
1610     BasicBlock *Copy =
1611         SplitBlock(MatMul->getParent(), MatMul, (DomTreeUpdater *)nullptr, LI,
1612                    nullptr, "copy");
1613     BasicBlock *Fusion =
1614         SplitBlock(MatMul->getParent(), MatMul, (DomTreeUpdater *)nullptr, LI,
1615                    nullptr, "no_alias");
1616 
1617     // Check if the loaded memory location begins before the end of the store
1618     // location. If the condition holds, they might overlap, otherwise they are
1619     // guaranteed to not overlap.
1620     IRBuilder<> Builder(MatMul);
1621     Check0->getTerminator()->eraseFromParent();
1622     Builder.SetInsertPoint(Check0);
1623     Type *IntPtrTy = Builder.getIntPtrTy(Load->getModule()->getDataLayout());
1624     Value *StoreBegin = Builder.CreatePtrToInt(
1625         const_cast<Value *>(StoreLoc.Ptr), IntPtrTy, "store.begin");
1626     Value *StoreEnd = Builder.CreateAdd(
1627         StoreBegin, ConstantInt::get(IntPtrTy, StoreLoc.Size.getValue()),
1628         "store.end", true, true);
1629     Value *LoadBegin = Builder.CreatePtrToInt(const_cast<Value *>(LoadLoc.Ptr),
1630                                               IntPtrTy, "load.begin");
1631     Builder.CreateCondBr(Builder.CreateICmpULT(LoadBegin, StoreEnd), Check1,
1632                          Fusion);
1633 
1634     // Check if the store begins before the end of the load location. If the
1635     // condition holds, they alias, otherwise they are guaranteed to not
1636     // overlap.
1637     Check1->getTerminator()->eraseFromParent();
1638     Builder.SetInsertPoint(Check1, Check1->begin());
1639     Value *LoadEnd = Builder.CreateAdd(
1640         LoadBegin, ConstantInt::get(IntPtrTy, LoadLoc.Size.getValue()),
1641         "load.end", true, true);
1642     Builder.CreateCondBr(Builder.CreateICmpULT(StoreBegin, LoadEnd), Copy,
1643                          Fusion);
1644 
1645     // Copy load operand to new alloca.
1646     Builder.SetInsertPoint(Copy, Copy->begin());
1647     auto *VT = cast<FixedVectorType>(Load->getType());
1648     // Use an array type for the alloca, to avoid potentially huge alignment
1649     // requirements for large vector types.
1650     auto *ArrayTy = ArrayType::get(VT->getElementType(), VT->getNumElements());
1651     AllocaInst *Alloca =
1652         Builder.CreateAlloca(ArrayTy, Load->getPointerAddressSpace());
1653 
1654     Builder.CreateMemCpy(Alloca, Alloca->getAlign(), Load->getPointerOperand(),
1655                          Load->getAlign(), LoadLoc.Size.getValue());
1656     Builder.SetInsertPoint(Fusion, Fusion->begin());
1657     PHINode *PHI = Builder.CreatePHI(Load->getPointerOperandType(), 3);
1658     PHI->addIncoming(Load->getPointerOperand(), Check0);
1659     PHI->addIncoming(Load->getPointerOperand(), Check1);
1660     PHI->addIncoming(Alloca, Copy);
1661 
1662     // Adjust DT.
1663     DTUpdates.push_back({DT->Insert, Check0, Check1});
1664     DTUpdates.push_back({DT->Insert, Check0, Fusion});
1665     DTUpdates.push_back({DT->Insert, Check1, Copy});
1666     DTUpdates.push_back({DT->Insert, Check1, Fusion});
1667     DT->applyUpdates(DTUpdates);
1668     return PHI;
1669   }
1670 
1671   bool isFusionProfitable(CallInst *MatMul) {
1672     if (ForceFusion)
1673       return true;
1674 
1675     ShapeInfo LShape(MatMul->getArgOperand(2), MatMul->getArgOperand(3));
1676     ShapeInfo RShape(MatMul->getArgOperand(3), MatMul->getArgOperand(4));
1677 
1678     const unsigned R = LShape.NumRows;
1679     const unsigned C = RShape.NumColumns;
1680     const unsigned M = LShape.NumColumns;
1681     auto *EltType = cast<VectorType>(MatMul->getType())->getElementType();
1682 
1683     const unsigned VF = std::max<unsigned>(
1684         TTI.getRegisterBitWidth(TargetTransformInfo::RGK_FixedWidthVector)
1685                 .getFixedValue() /
1686             EltType->getPrimitiveSizeInBits().getFixedValue(),
1687         1U);
1688 
1689     // Cost model for tiling
1690     //
1691     // For tiling to be beneficial, we need reuse either along the R or
1692     // the C axis.  We vectorize along the R axis so that means at least
1693     // 3 elements.
1694     // TODO: Also consider cost of copying if operands alias.
1695     if (R <= VF && C == 1)
1696       return false;
1697     // Then we need enough elements to exceed the number of vector
1698     // registers we have.  Note that this is an oversimplification since
1699     // fusing also takes some extra loads which may exceed the number of
1700     // reloads necessary.
1701     unsigned Op0Regs = (R + VF - 1) / VF * M;
1702     unsigned Op1Regs = (M + VF - 1) / VF * C;
1703     return Op0Regs + Op1Regs >
1704            TTI.getNumberOfRegisters(TTI.getRegisterClassForType(true));
1705   }
1706 
1707   MatrixTy getZeroMatrix(Type *EltType, unsigned R, unsigned C) {
1708     MatrixTy Res;
1709     auto *ColumType = FixedVectorType::get(EltType, R);
1710     for (unsigned I = 0; I < C; ++I)
1711       Res.addVector(ConstantAggregateZero::get(ColumType));
1712     return Res;
1713   }
1714 
1715   void createTiledLoops(CallInst *MatMul, Value *LPtr, ShapeInfo LShape,
1716                         Value *RPtr, ShapeInfo RShape, StoreInst *Store) {
1717     auto *EltType = cast<VectorType>(MatMul->getType())->getElementType();
1718 
1719     // Create the main tiling loop nest.
1720     TileInfo TI(LShape.NumRows, RShape.NumColumns, LShape.NumColumns, TileSize);
1721     DomTreeUpdater DTU(DT, DomTreeUpdater::UpdateStrategy::Lazy);
1722     Instruction *InsertI = cast<Instruction>(MatMul);
1723     BasicBlock *Start = InsertI->getParent();
1724     BasicBlock *End =
1725         SplitBlock(InsertI->getParent(), InsertI, DT, LI, nullptr, "continue");
1726     IRBuilder<> Builder(MatMul);
1727     BasicBlock *InnerBody = TI.CreateTiledLoops(Start, End, Builder, DTU, *LI);
1728 
1729     Type *TileVecTy =
1730         FixedVectorType::get(MatMul->getType()->getScalarType(), TileSize);
1731     MatrixTy TileResult;
1732     // Insert in the inner loop header.
1733     Builder.SetInsertPoint(TI.KLoop.Header->getTerminator());
1734     // Create PHI nodes for the result columns to accumulate across iterations.
1735     SmallVector<PHINode *, 4> ColumnPhis;
1736     for (unsigned I = 0; I < TileSize; I++) {
1737       auto *Phi = Builder.CreatePHI(TileVecTy, 2, "result.vec." + Twine(I));
1738       Phi->addIncoming(ConstantAggregateZero::get(TileVecTy),
1739                        TI.RowLoop.Header->getSingleSuccessor());
1740       TileResult.addVector(Phi);
1741       ColumnPhis.push_back(Phi);
1742     }
1743 
1744     // Insert in the inner loop body, which computes
1745     //   Res += Load(CurrentRow, K) * Load(K, CurrentColumn)
1746     Builder.SetInsertPoint(InnerBody->getTerminator());
1747     // Load tiles of the operands.
1748     MatrixTy A =
1749         loadMatrix(LPtr, {}, false, LShape, TI.RowLoop.Index, TI.KLoop.Index,
1750                    {TileSize, TileSize}, EltType, Builder);
1751     MatrixTy B =
1752         loadMatrix(RPtr, {}, false, RShape, TI.KLoop.Index, TI.ColumnLoop.Index,
1753                    {TileSize, TileSize}, EltType, Builder);
1754     emitMatrixMultiply(TileResult, A, B, Builder, true, false,
1755                        getFastMathFlags(MatMul));
1756     // Store result after the inner loop is done.
1757     Builder.SetInsertPoint(TI.RowLoop.Latch->getTerminator());
1758     storeMatrix(TileResult, Store->getPointerOperand(), Store->getAlign(),
1759                 Store->isVolatile(), {LShape.NumRows, RShape.NumColumns},
1760                 TI.RowLoop.Index, TI.ColumnLoop.Index, EltType, Builder);
1761 
1762     for (unsigned I = 0; I < TileResult.getNumVectors(); I++)
1763       ColumnPhis[I]->addIncoming(TileResult.getVector(I), TI.KLoop.Latch);
1764 
1765     // Force unrolling of a few iterations of the inner loop, to make sure there
1766     // is enough work per iteration.
1767     // FIXME: The unroller should make this decision directly instead, but
1768     // currently the cost-model is not up to the task.
1769     unsigned InnerLoopUnrollCount = std::min(10u, LShape.NumColumns / TileSize);
1770     addStringMetadataToLoop(LI->getLoopFor(TI.KLoop.Header),
1771                             "llvm.loop.unroll.count", InnerLoopUnrollCount);
1772   }
1773 
1774   void emitSIMDTiling(CallInst *MatMul, LoadInst *LoadOp0, LoadInst *LoadOp1,
1775                       StoreInst *Store,
1776                       SmallPtrSetImpl<Instruction *> &FusedInsts) {
1777     assert(MatrixLayout == MatrixLayoutTy::ColumnMajor &&
1778            "Tiling only supported for column-major matrixes at the moment!");
1779     if (!isFusionProfitable(MatMul))
1780       return;
1781 
1782     ShapeInfo LShape(MatMul->getArgOperand(2), MatMul->getArgOperand(3));
1783     ShapeInfo RShape(MatMul->getArgOperand(3), MatMul->getArgOperand(4));
1784 
1785     const unsigned R = LShape.NumRows;
1786     const unsigned C = RShape.NumColumns;
1787     const unsigned M = LShape.NumColumns;
1788     auto *EltType = cast<VectorType>(MatMul->getType())->getElementType();
1789 
1790     Value *APtr = getNonAliasingPointer(LoadOp0, Store, MatMul);
1791     Value *BPtr = getNonAliasingPointer(LoadOp1, Store, MatMul);
1792     Value *CPtr = Store->getPointerOperand();
1793 
1794     if (TileUseLoops && (R % TileSize == 0 && C % TileSize == 0))
1795       createTiledLoops(MatMul, APtr, LShape, BPtr, RShape, Store);
1796     else {
1797       IRBuilder<> Builder(Store);
1798       for (unsigned J = 0; J < C; J += TileSize)
1799         for (unsigned I = 0; I < R; I += TileSize) {
1800           const unsigned TileR = std::min(R - I, unsigned(TileSize));
1801           const unsigned TileC = std::min(C - J, unsigned(TileSize));
1802           MatrixTy Res = getZeroMatrix(EltType, TileR, TileC);
1803 
1804           for (unsigned K = 0; K < M; K += TileSize) {
1805             const unsigned TileM = std::min(M - K, unsigned(TileSize));
1806             MatrixTy A =
1807                 loadMatrix(APtr, LoadOp0->getAlign(), LoadOp0->isVolatile(),
1808                            LShape, Builder.getInt64(I), Builder.getInt64(K),
1809                            {TileR, TileM}, EltType, Builder);
1810             MatrixTy B =
1811                 loadMatrix(BPtr, LoadOp1->getAlign(), LoadOp1->isVolatile(),
1812                            RShape, Builder.getInt64(K), Builder.getInt64(J),
1813                            {TileM, TileC}, EltType, Builder);
1814             emitMatrixMultiply(Res, A, B, Builder, true, false,
1815                                getFastMathFlags(MatMul));
1816           }
1817           storeMatrix(Res, CPtr, Store->getAlign(), Store->isVolatile(), {R, M},
1818                       Builder.getInt64(I), Builder.getInt64(J), EltType,
1819                       Builder);
1820         }
1821     }
1822 
1823     // Mark eliminated instructions as fused and remove them.
1824     FusedInsts.insert(Store);
1825     FusedInsts.insert(MatMul);
1826     Store->eraseFromParent();
1827     MatMul->eraseFromParent();
1828     if (LoadOp0->hasNUses(0)) {
1829       FusedInsts.insert(LoadOp0);
1830       LoadOp0->eraseFromParent();
1831     }
1832     if (LoadOp1 != LoadOp0 && LoadOp1->hasNUses(0)) {
1833       FusedInsts.insert(LoadOp1);
1834       LoadOp1->eraseFromParent();
1835     }
1836   }
1837 
1838   /// Try to lower matrix multiply chains by fusing operations.
1839   ///
1840   /// Call finalizeLowering on lowered instructions.  Instructions that are
1841   /// completely eliminated by fusion are added to \p FusedInsts.
1842   void LowerMatrixMultiplyFused(CallInst *MatMul,
1843                                 SmallPtrSetImpl<Instruction *> &FusedInsts) {
1844     if (!FuseMatrix || !DT)
1845       return;
1846 
1847     assert(AA && LI && "Analyses should be available");
1848 
1849     Value *A = MatMul->getArgOperand(0);
1850     Value *B = MatMul->getArgOperand(1);
1851 
1852     // We can fold the transpose into the operand that is used to fetch scalars.
1853     Value *T;
1854     if (MatrixLayout == MatrixLayoutTy::ColumnMajor
1855             ? match(B, m_Intrinsic<Intrinsic::matrix_transpose>(m_Value(T)))
1856             : match(A, m_Intrinsic<Intrinsic::matrix_transpose>(m_Value(T)))) {
1857       IRBuilder<> Builder(MatMul);
1858       auto *EltType = cast<VectorType>(MatMul->getType())->getElementType();
1859       ShapeInfo LShape(MatMul->getArgOperand(2), MatMul->getArgOperand(3));
1860       ShapeInfo RShape(MatMul->getArgOperand(3), MatMul->getArgOperand(4));
1861       const unsigned R = LShape.NumRows;
1862       const unsigned M = LShape.NumColumns;
1863       const unsigned C = RShape.NumColumns;
1864 
1865       MatrixTy MA;
1866       MatrixTy MB;
1867 
1868       Value *Transpose;
1869       if (MatrixLayout == MatrixLayoutTy::ColumnMajor) {
1870         MA = getMatrix(A, ShapeInfo(R, M), Builder);
1871         MB = getMatrix(T, ShapeInfo(C, M), Builder);
1872         Transpose = B;
1873       } else {
1874         MA = getMatrix(T, ShapeInfo(R, M), Builder);
1875         MB = getMatrix(B, ShapeInfo(C, M), Builder);
1876         Transpose = A;
1877       }
1878 
1879       // Initialize the output
1880       MatrixTy Result(R, C, EltType);
1881 
1882       emitMatrixMultiply(Result, MA, MB, Builder, false, true,
1883                          getFastMathFlags(MatMul));
1884 
1885       FusedInsts.insert(MatMul);
1886       if (Transpose->hasOneUse()) {
1887         FusedInsts.insert(cast<Instruction>(Transpose));
1888         ToRemove.push_back(cast<Instruction>(Transpose));
1889         // TODO: add a fake entry for the folded instruction so that this is
1890         // included in the expression in the remark.
1891         Inst2ColumnMatrix[Transpose] = MatrixTy(M, C, EltType);
1892       }
1893       finalizeLowering(MatMul, Result, Builder);
1894       return;
1895     }
1896 
1897     if (!MatMul->hasOneUse() || MatrixLayout != MatrixLayoutTy::ColumnMajor)
1898       return;
1899 
1900     // Lower {ld, ld} -> matmul -> st chains.  No need to call finalizeLowering
1901     // since the single store user will be lowered as part of this.
1902     auto *LoadOp0 = dyn_cast<LoadInst>(A);
1903     auto *LoadOp1 = dyn_cast<LoadInst>(B);
1904     auto *Store = dyn_cast<StoreInst>(*MatMul->user_begin());
1905     if (LoadOp0 && LoadOp1 && Store) {
1906       // The store address must dominate the MatMul instruction, otherwise
1907       // we create invalid IR.
1908       SetVector<Value *> WorkList;
1909       WorkList.insert(Store->getOperand(1));
1910       SmallVector<Instruction *> ToHoist;
1911       for (unsigned I = 0; I != WorkList.size(); ++I) {
1912         Value *Current = WorkList[I];
1913         auto *CurrI = dyn_cast<Instruction>(Current);
1914         if (!CurrI)
1915           continue;
1916         if (isa<PHINode>(CurrI))
1917           return;
1918         if (DT->dominates(CurrI, MatMul))
1919           continue;
1920         if (CurrI->mayHaveSideEffects() || CurrI->mayReadFromMemory())
1921           return;
1922         ToHoist.push_back(CurrI);
1923         WorkList.insert(CurrI->op_begin(), CurrI->op_end());
1924       }
1925 
1926       sort(ToHoist, [this](Instruction *A, Instruction *B) {
1927         return DT->dominates(A, B);
1928       });
1929       for (Instruction *I : ToHoist)
1930         I->moveBefore(MatMul);
1931 
1932       emitSIMDTiling(MatMul, LoadOp0, LoadOp1, Store, FusedInsts);
1933       return;
1934     }
1935   }
1936 
1937   /// Lowers llvm.matrix.multiply.
1938   void LowerMultiply(CallInst *MatMul) {
1939     IRBuilder<> Builder(MatMul);
1940     auto *EltType = cast<VectorType>(MatMul->getType())->getElementType();
1941     ShapeInfo LShape(MatMul->getArgOperand(2), MatMul->getArgOperand(3));
1942     ShapeInfo RShape(MatMul->getArgOperand(3), MatMul->getArgOperand(4));
1943 
1944     const MatrixTy &Lhs = getMatrix(MatMul->getArgOperand(0), LShape, Builder);
1945     const MatrixTy &Rhs = getMatrix(MatMul->getArgOperand(1), RShape, Builder);
1946     assert(Lhs.getElementType() == Rhs.getElementType() &&
1947            "Matrix multiply argument element types do not match.");
1948 
1949     const unsigned R = LShape.NumRows;
1950     const unsigned C = RShape.NumColumns;
1951     assert(LShape.NumColumns == RShape.NumRows);
1952 
1953     // Initialize the output
1954     MatrixTy Result(R, C, EltType);
1955     assert(Lhs.getElementType() == Result.getElementType() &&
1956            "Matrix multiply result element type does not match arguments.");
1957 
1958     emitMatrixMultiply(Result, Lhs, Rhs, Builder, false, false,
1959                        getFastMathFlags(MatMul));
1960     finalizeLowering(MatMul, Result, Builder);
1961   }
1962 
1963   /// Lowers llvm.matrix.transpose.
1964   void LowerTranspose(CallInst *Inst) {
1965     MatrixTy Result;
1966     IRBuilder<> Builder(Inst);
1967     Value *InputVal = Inst->getArgOperand(0);
1968     VectorType *VectorTy = cast<VectorType>(InputVal->getType());
1969     ShapeInfo ArgShape(Inst->getArgOperand(1), Inst->getArgOperand(2));
1970     MatrixTy InputMatrix = getMatrix(InputVal, ArgShape, Builder);
1971 
1972     const unsigned NewNumVecs =
1973         InputMatrix.isColumnMajor() ? ArgShape.NumRows : ArgShape.NumColumns;
1974     const unsigned NewNumElts =
1975         InputMatrix.isColumnMajor() ? ArgShape.NumColumns : ArgShape.NumRows;
1976 
1977     for (unsigned I = 0; I < NewNumVecs; ++I) {
1978       // Build a single result vector. First initialize it.
1979       Value *ResultVector = PoisonValue::get(
1980           FixedVectorType::get(VectorTy->getElementType(), NewNumElts));
1981       // Go through the old elements and insert it into the resulting vector.
1982       for (auto J : enumerate(InputMatrix.vectors())) {
1983         Value *Elt = Builder.CreateExtractElement(J.value(), I);
1984         // Row and column indices are transposed.
1985         ResultVector =
1986             Builder.CreateInsertElement(ResultVector, Elt, J.index());
1987       }
1988       Result.addVector(ResultVector);
1989     }
1990 
1991     // TODO: Improve estimate of operations needed for transposes. Currently we
1992     // just count the insertelement/extractelement instructions, but do not
1993     // account for later simplifications/combines.
1994     finalizeLowering(
1995         Inst,
1996         Result.addNumComputeOps(2 * ArgShape.NumRows * ArgShape.NumColumns)
1997             .addNumExposedTransposes(1),
1998         Builder);
1999   }
2000 
2001   /// Lower load instructions, if shape information is available.
2002   bool VisitLoad(LoadInst *Inst, Value *Ptr, IRBuilder<> &Builder) {
2003     auto I = ShapeMap.find(Inst);
2004     if (I == ShapeMap.end())
2005       return false;
2006 
2007     LowerLoad(Inst, Ptr, Inst->getAlign(),
2008               Builder.getInt64(I->second.getStride()), Inst->isVolatile(),
2009               I->second);
2010     return true;
2011   }
2012 
2013   bool VisitStore(StoreInst *Inst, Value *StoredVal, Value *Ptr,
2014                   IRBuilder<> &Builder) {
2015     auto I = ShapeMap.find(StoredVal);
2016     if (I == ShapeMap.end())
2017       return false;
2018 
2019     LowerStore(Inst, StoredVal, Ptr, Inst->getAlign(),
2020                Builder.getInt64(I->second.getStride()), Inst->isVolatile(),
2021                I->second);
2022     return true;
2023   }
2024 
2025   /// Lower binary operators, if shape information is available.
2026   bool VisitBinaryOperator(BinaryOperator *Inst) {
2027     auto I = ShapeMap.find(Inst);
2028     if (I == ShapeMap.end())
2029       return false;
2030 
2031     Value *Lhs = Inst->getOperand(0);
2032     Value *Rhs = Inst->getOperand(1);
2033 
2034     IRBuilder<> Builder(Inst);
2035     ShapeInfo &Shape = I->second;
2036 
2037     MatrixTy Result;
2038     MatrixTy A = getMatrix(Lhs, Shape, Builder);
2039     MatrixTy B = getMatrix(Rhs, Shape, Builder);
2040     assert(A.isColumnMajor() == B.isColumnMajor() &&
2041            Result.isColumnMajor() == A.isColumnMajor() &&
2042            "operands must agree on matrix layout");
2043 
2044     Builder.setFastMathFlags(getFastMathFlags(Inst));
2045 
2046     // Helper to perform binary op on vectors.
2047     auto BuildVectorOp = [&Builder, Inst](Value *LHS, Value *RHS) {
2048       switch (Inst->getOpcode()) {
2049       case Instruction::Add:
2050         return Builder.CreateAdd(LHS, RHS);
2051       case Instruction::Mul:
2052         return Builder.CreateMul(LHS, RHS);
2053       case Instruction::Sub:
2054         return Builder.CreateSub(LHS, RHS);
2055       case Instruction::FAdd:
2056         return Builder.CreateFAdd(LHS, RHS);
2057       case Instruction::FMul:
2058         return Builder.CreateFMul(LHS, RHS);
2059       case Instruction::FSub:
2060         return Builder.CreateFSub(LHS, RHS);
2061       default:
2062         llvm_unreachable("Unsupported binary operator for matrix");
2063       }
2064     };
2065 
2066     for (unsigned I = 0; I < Shape.getNumVectors(); ++I)
2067       Result.addVector(BuildVectorOp(A.getVector(I), B.getVector(I)));
2068 
2069     finalizeLowering(Inst,
2070                      Result.addNumComputeOps(getNumOps(Result.getVectorTy()) *
2071                                              Result.getNumVectors()),
2072                      Builder);
2073     return true;
2074   }
2075 
2076   /// Lower unary operators, if shape information is available.
2077   bool VisitUnaryOperator(UnaryOperator *Inst) {
2078     auto I = ShapeMap.find(Inst);
2079     if (I == ShapeMap.end())
2080       return false;
2081 
2082     Value *Op = Inst->getOperand(0);
2083 
2084     IRBuilder<> Builder(Inst);
2085     ShapeInfo &Shape = I->second;
2086 
2087     MatrixTy Result;
2088     MatrixTy M = getMatrix(Op, Shape, Builder);
2089 
2090     Builder.setFastMathFlags(getFastMathFlags(Inst));
2091 
2092     // Helper to perform unary op on vectors.
2093     auto BuildVectorOp = [&Builder, Inst](Value *Op) {
2094       switch (Inst->getOpcode()) {
2095       case Instruction::FNeg:
2096         return Builder.CreateFNeg(Op);
2097       default:
2098         llvm_unreachable("Unsupported unary operator for matrix");
2099       }
2100     };
2101 
2102     for (unsigned I = 0; I < Shape.getNumVectors(); ++I)
2103       Result.addVector(BuildVectorOp(M.getVector(I)));
2104 
2105     finalizeLowering(Inst,
2106                      Result.addNumComputeOps(getNumOps(Result.getVectorTy()) *
2107                                              Result.getNumVectors()),
2108                      Builder);
2109     return true;
2110   }
2111 
2112   /// Helper to linearize a matrix expression tree into a string. Currently
2113   /// matrix expressions are linarized by starting at an expression leaf and
2114   /// linearizing bottom up.
2115   struct ExprLinearizer {
2116     unsigned LengthToBreak = 100;
2117     std::string Str;
2118     raw_string_ostream Stream;
2119     unsigned LineLength = 0;
2120     const DataLayout &DL;
2121 
2122     /// Mapping from instructions to matrixes. It is used to identify
2123     /// matrix instructions.
2124     const MapVector<Value *, MatrixTy> &Inst2Matrix;
2125 
2126     /// Mapping from values to the leaves of all expressions that the value is
2127     /// part of.
2128     const DenseMap<Value *, SmallPtrSet<Value *, 2>> &Shared;
2129 
2130     /// Set of matrix expressions in the scope of a given DISubprogram.
2131     const SmallSetVector<Value *, 32> &ExprsInSubprogram;
2132 
2133     /// Leaf node of the expression to linearize.
2134     Value *Leaf;
2135 
2136     /// Used to keep track of sub-expressions that get reused while linearizing
2137     /// the expression. Re-used sub-expressions are marked as (reused).
2138     SmallPtrSet<Value *, 8> ReusedExprs;
2139 
2140     ExprLinearizer(const DataLayout &DL,
2141                    const MapVector<Value *, MatrixTy> &Inst2Matrix,
2142                    const DenseMap<Value *, SmallPtrSet<Value *, 2>> &Shared,
2143                    const SmallSetVector<Value *, 32> &ExprsInSubprogram,
2144                    Value *Leaf)
2145         : Stream(Str), DL(DL), Inst2Matrix(Inst2Matrix), Shared(Shared),
2146           ExprsInSubprogram(ExprsInSubprogram), Leaf(Leaf) {}
2147 
2148     void indent(unsigned N) {
2149       LineLength += N;
2150       for (unsigned i = 0; i < N; i++)
2151         Stream << " ";
2152     }
2153 
2154     void lineBreak() {
2155       Stream << "\n";
2156       LineLength = 0;
2157     }
2158 
2159     void maybeIndent(unsigned Indent) {
2160       if (LineLength >= LengthToBreak)
2161         lineBreak();
2162 
2163       if (LineLength == 0)
2164         indent(Indent);
2165     }
2166 
2167     void write(StringRef S) {
2168       LineLength += S.size();
2169       Stream << S;
2170     }
2171 
2172     Value *getUnderlyingObjectThroughLoads(Value *V) {
2173       if (Value *Ptr = getPointerOperand(V))
2174         return getUnderlyingObjectThroughLoads(Ptr);
2175       else if (V->getType()->isPointerTy())
2176         return getUnderlyingObject(V);
2177       return V;
2178     }
2179 
2180     /// Returns true if \p V is a matrix value in the given subprogram.
2181     bool isMatrix(Value *V) const { return ExprsInSubprogram.count(V); }
2182 
2183     /// If \p V is a matrix value, print its shape as as NumRows x NumColumns to
2184     /// \p SS.
2185     void prettyPrintMatrixType(Value *V, raw_string_ostream &SS) {
2186       auto M = Inst2Matrix.find(V);
2187       if (M == Inst2Matrix.end())
2188         SS << "unknown";
2189       else {
2190         SS << M->second.getNumRows();
2191         SS << "x";
2192         SS << M->second.getNumColumns();
2193       }
2194     }
2195 
2196     /// Write the called function name. Handles calls to llvm.matrix.*
2197     /// specially: we write the name, followed by the dimensions of the input
2198     /// matrixes, followed by the scalar type name.
2199     void writeFnName(CallInst *CI) {
2200       if (!CI->getCalledFunction())
2201         write("<no called fn>");
2202       else {
2203         StringRef Name = CI->getCalledFunction()->getName();
2204         if (!Name.startswith("llvm.matrix")) {
2205           write(Name);
2206           return;
2207         }
2208         auto *II = cast<IntrinsicInst>(CI);
2209         write(Intrinsic::getBaseName(II->getIntrinsicID())
2210                   .drop_front(StringRef("llvm.matrix.").size()));
2211         write(".");
2212         std::string Tmp;
2213         raw_string_ostream SS(Tmp);
2214 
2215         switch (II->getIntrinsicID()) {
2216         case Intrinsic::matrix_multiply:
2217           prettyPrintMatrixType(II->getOperand(0), SS);
2218           SS << ".";
2219           prettyPrintMatrixType(II->getOperand(1), SS);
2220           SS << "." << *II->getType()->getScalarType();
2221           break;
2222         case Intrinsic::matrix_transpose:
2223           prettyPrintMatrixType(II->getOperand(0), SS);
2224           SS << "." << *II->getType()->getScalarType();
2225           break;
2226         case Intrinsic::matrix_column_major_load:
2227           prettyPrintMatrixType(II, SS);
2228           SS << "." << *II->getType()->getScalarType();
2229           break;
2230         case Intrinsic::matrix_column_major_store:
2231           prettyPrintMatrixType(II->getOperand(0), SS);
2232           SS << "." << *II->getOperand(0)->getType()->getScalarType();
2233           break;
2234         default:
2235           llvm_unreachable("Unhandled case");
2236         }
2237         SS.flush();
2238         write(Tmp);
2239       }
2240     }
2241 
2242     unsigned getNumShapeArgs(CallInst *CI) const {
2243       if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(CI)) {
2244         switch (II->getIntrinsicID()) {
2245         case Intrinsic::matrix_multiply:
2246           return 3;
2247         case Intrinsic::matrix_transpose:
2248           return 2;
2249         case Intrinsic::matrix_column_major_load:
2250         case Intrinsic::matrix_column_major_store:
2251           return 3;
2252         default:
2253           return 0;
2254         }
2255       }
2256       return 0;
2257     }
2258 
2259     /// Special printing for values: for pointers, we print if they refer to an
2260     /// (function) external address or a stack address, for other values we
2261     /// either print the constant or "scalar"/"matrix" for other values.
2262     void write(Value *V) {
2263       V = getUnderlyingObjectThroughLoads(V);
2264       if (V->getType()->isPointerTy()) {
2265         if (isa<AllocaInst>(V)) {
2266           Stream << "stack addr";
2267           LineLength += StringRef("stack addr").size();
2268         } else {
2269           Stream << "addr";
2270           LineLength += StringRef("addr").size();
2271         }
2272         if (!V->getName().empty()) {
2273           Stream << " %" << V->getName() << "";
2274           LineLength += V->getName().size() + 2;
2275         }
2276         return;
2277       }
2278 
2279       std::string Tmp;
2280       raw_string_ostream TmpStream(Tmp);
2281 
2282       if (auto *CI = dyn_cast<ConstantInt>(V))
2283         TmpStream << CI->getValue();
2284       else if (isa<Constant>(V))
2285         TmpStream << "constant";
2286       else {
2287         if (isMatrix(V))
2288           TmpStream << "matrix";
2289         else
2290           TmpStream << "scalar";
2291       }
2292       TmpStream.flush();
2293       Tmp = std::string(StringRef(Tmp).trim());
2294       LineLength += Tmp.size();
2295       Stream << Tmp;
2296     }
2297 
2298     /// Linearize expression \p Expr starting at an indentation of \p Indent.
2299     /// Expressions that are re-used multiple times are prefixed with (reused)
2300     /// at the re-used root instruction.
2301     void linearizeExpr(Value *Expr, unsigned Indent, bool ParentReused,
2302                        bool ParentShared) {
2303       auto *I = cast<Instruction>(Expr);
2304       maybeIndent(Indent);
2305       SmallVector<Value *, 8> Ops;
2306 
2307       // Is Expr shared with other expression leaves?
2308       bool ExprShared = false;
2309 
2310       // Deal with shared subtrees. Mark them as shared, if required.
2311       if (!ParentShared) {
2312         auto SI = Shared.find(Expr);
2313         assert(SI != Shared.end() && SI->second.count(Leaf));
2314 
2315         for (Value *S : SI->second) {
2316           if (S == Leaf)
2317             continue;
2318           DebugLoc DL = cast<Instruction>(S)->getDebugLoc();
2319           write("shared with remark at line " + std::to_string(DL.getLine()) +
2320                 " column " + std::to_string(DL.getCol()) + " (");
2321         }
2322         ExprShared = SI->second.size() > 1;
2323       }
2324 
2325       bool Reused = !ReusedExprs.insert(Expr).second;
2326       if (Reused && !ParentReused)
2327         write("(reused) ");
2328 
2329       if (auto *CI = dyn_cast<CallInst>(I)) {
2330         writeFnName(CI);
2331 
2332         Ops.append(CI->arg_begin(), CI->arg_end() - getNumShapeArgs(CI));
2333       } else if (isa<BitCastInst>(Expr)) {
2334         // Special case bitcasts, which are used to materialize matrixes from
2335         // non-matrix ops.
2336         write("matrix");
2337         return;
2338       } else {
2339         Ops.append(I->value_op_begin(), I->value_op_end());
2340         write(std::string(I->getOpcodeName()));
2341       }
2342 
2343       write(std::string("("));
2344 
2345       unsigned NumOpsToBreak = 1;
2346       if (match(Expr, m_Intrinsic<Intrinsic::matrix_column_major_load>()))
2347         NumOpsToBreak = 2;
2348 
2349       for (Value *Op : Ops) {
2350         if (Ops.size() > NumOpsToBreak)
2351           lineBreak();
2352 
2353         maybeIndent(Indent + 1);
2354         if (isMatrix(Op))
2355           linearizeExpr(Op, Indent + 1, Reused, ExprShared);
2356         else
2357           write(Op);
2358         if (Op != Ops.back())
2359           write(", ");
2360       }
2361 
2362       write(")");
2363     }
2364 
2365     const std::string &getResult() {
2366       Stream.flush();
2367       return Str;
2368     }
2369   };
2370 
2371   /// Generate remarks for matrix operations in a function. To generate remarks
2372   /// for matrix expressions, the following approach is used:
2373   /// 1. Use the inlined-at debug information to group matrix operations to the
2374   ///    DISubprograms they are contained in.
2375   /// 2. Collect leaves of matrix expressions (done in
2376   ///    RemarkGenerator::getExpressionLeaves) for each subprogram - expression
2377   //     mapping.  Leaves are lowered matrix instructions without other matrix
2378   //     users (like stores) in the current subprogram.
2379   /// 3. For each leaf, create a remark containing a linearizied version of the
2380   ///    matrix expression. The expression is linearized by a recursive
2381   ///    bottom-up traversal of the matrix operands, starting at a leaf. Note
2382   ///    that multiple leaves can share sub-expressions. Shared subexpressions
2383   ///    are explicitly marked as shared().
2384   struct RemarkGenerator {
2385     const MapVector<Value *, MatrixTy> &Inst2Matrix;
2386     OptimizationRemarkEmitter &ORE;
2387     Function &Func;
2388     const DataLayout &DL;
2389 
2390     RemarkGenerator(const MapVector<Value *, MatrixTy> &Inst2Matrix,
2391                     OptimizationRemarkEmitter &ORE, Function &Func)
2392         : Inst2Matrix(Inst2Matrix), ORE(ORE), Func(Func),
2393           DL(Func.getParent()->getDataLayout()) {}
2394 
2395     /// Return all leaves of the expressions in \p ExprsInSubprogram. Those are
2396     /// instructions in Inst2Matrix returning void or without any users in
2397     /// \p ExprsInSubprogram. Currently that should only include stores.
2398     SmallVector<Value *, 4>
2399     getExpressionLeaves(const SmallSetVector<Value *, 32> &ExprsInSubprogram) {
2400       SmallVector<Value *, 4> Leaves;
2401       for (auto *Expr : ExprsInSubprogram)
2402         if (Expr->getType()->isVoidTy() ||
2403             !any_of(Expr->users(), [&ExprsInSubprogram](User *U) {
2404               return ExprsInSubprogram.count(U);
2405             }))
2406           Leaves.push_back(Expr);
2407       return Leaves;
2408     }
2409 
2410     /// Recursively traverse expression \p V starting at \p Leaf and add \p Leaf
2411     /// to all visited expressions in \p Shared. Limit the matrix operations to
2412     /// the ones in \p ExprsInSubprogram.
2413     void collectSharedInfo(Value *Leaf, Value *V,
2414                            const SmallSetVector<Value *, 32> &ExprsInSubprogram,
2415                            DenseMap<Value *, SmallPtrSet<Value *, 2>> &Shared) {
2416 
2417       if (!ExprsInSubprogram.count(V))
2418         return;
2419 
2420       auto I = Shared.insert({V, {}});
2421       I.first->second.insert(Leaf);
2422 
2423       for (Value *Op : cast<Instruction>(V)->operand_values())
2424         collectSharedInfo(Leaf, Op, ExprsInSubprogram, Shared);
2425     }
2426 
2427     /// Calculate the number of exclusive and shared op counts for expression
2428     /// starting at \p V. Expressions used multiple times are counted once.
2429     /// Limit the matrix operations to the ones in \p ExprsInSubprogram.
2430     std::pair<OpInfoTy, OpInfoTy>
2431     sumOpInfos(Value *Root, SmallPtrSetImpl<Value *> &ReusedExprs,
2432                const SmallSetVector<Value *, 32> &ExprsInSubprogram,
2433                DenseMap<Value *, SmallPtrSet<Value *, 2>> &Shared) const {
2434       if (!ExprsInSubprogram.count(Root))
2435         return {};
2436 
2437       // Already counted this expression. Stop.
2438       if (!ReusedExprs.insert(Root).second)
2439         return {};
2440 
2441       OpInfoTy SharedCount;
2442       OpInfoTy Count;
2443 
2444       auto I = Shared.find(Root);
2445       auto CM = Inst2Matrix.find(Root);
2446       if (I->second.size() == 1)
2447         Count = CM->second.getOpInfo();
2448       else
2449         SharedCount = CM->second.getOpInfo();
2450 
2451       for (Value *Op : cast<Instruction>(Root)->operand_values()) {
2452         auto C = sumOpInfos(Op, ReusedExprs, ExprsInSubprogram, Shared);
2453         Count += C.first;
2454         SharedCount += C.second;
2455       }
2456       return {Count, SharedCount};
2457     }
2458 
2459     void emitRemarks() {
2460       if (!ORE.allowExtraAnalysis(DEBUG_TYPE))
2461         return;
2462 
2463       // Map matrix operations to their containting subprograms, by traversing
2464       // the inlinedAt chain. If the function does not have a DISubprogram, we
2465       // only map them to the containing function.
2466       MapVector<DISubprogram *, SmallVector<Value *, 8>> Subprog2Exprs;
2467       for (const auto &KV : Inst2Matrix) {
2468         if (Func.getSubprogram()) {
2469           auto *I = cast<Instruction>(KV.first);
2470           DILocation *Context = I->getDebugLoc();
2471           while (Context) {
2472             auto I =
2473                 Subprog2Exprs.insert({getSubprogram(Context->getScope()), {}});
2474             I.first->second.push_back(KV.first);
2475             Context = DebugLoc(Context).getInlinedAt();
2476           }
2477         } else {
2478           auto I = Subprog2Exprs.insert({nullptr, {}});
2479           I.first->second.push_back(KV.first);
2480         }
2481       }
2482       for (auto &KV : Subprog2Exprs) {
2483         SmallSetVector<Value *, 32> ExprsInSubprogram(KV.second.begin(),
2484                                                       KV.second.end());
2485         auto Leaves = getExpressionLeaves(ExprsInSubprogram);
2486 
2487         DenseMap<Value *, SmallPtrSet<Value *, 2>> Shared;
2488         for (Value *Leaf : Leaves)
2489           collectSharedInfo(Leaf, Leaf, ExprsInSubprogram, Shared);
2490 
2491         // Generate remarks for each leaf.
2492         for (auto *L : Leaves) {
2493 
2494           DebugLoc Loc = cast<Instruction>(L)->getDebugLoc();
2495           DILocation *Context = cast<Instruction>(L)->getDebugLoc();
2496           while (Context) {
2497             if (getSubprogram(Context->getScope()) == KV.first) {
2498               Loc = Context;
2499               break;
2500             }
2501             Context = DebugLoc(Context).getInlinedAt();
2502           }
2503 
2504           SmallPtrSet<Value *, 8> ReusedExprs;
2505           OpInfoTy Counts, SharedCounts;
2506           std::tie(Counts, SharedCounts) =
2507               sumOpInfos(L, ReusedExprs, ExprsInSubprogram, Shared);
2508 
2509           OptimizationRemark Rem(DEBUG_TYPE, "matrix-lowered", Loc,
2510                                  cast<Instruction>(L)->getParent());
2511 
2512           Rem << "Lowered with ";
2513           Rem << ore::NV("NumStores", Counts.NumStores) << " stores, "
2514               << ore::NV("NumLoads", Counts.NumLoads) << " loads, "
2515               << ore::NV("NumComputeOps", Counts.NumComputeOps)
2516               << " compute ops, "
2517               << ore::NV("NumExposedTransposes", Counts.NumExposedTransposes)
2518               << " exposed transposes";
2519 
2520           if (SharedCounts.NumStores > 0 || SharedCounts.NumLoads > 0 ||
2521               SharedCounts.NumComputeOps > 0) {
2522             Rem << ",\nadditionally "
2523                 << ore::NV("NumStores", SharedCounts.NumStores) << " stores, "
2524                 << ore::NV("NumLoads", SharedCounts.NumLoads) << " loads, "
2525                 << ore::NV("NumFPOps", SharedCounts.NumComputeOps)
2526                 << " compute ops"
2527                 << " are shared with other expressions";
2528           }
2529 
2530           Rem << ("\n" + linearize(L, Shared, ExprsInSubprogram, DL));
2531           ORE.emit(Rem);
2532         }
2533       }
2534     }
2535 
2536     std::string
2537     linearize(Value *L,
2538               const DenseMap<Value *, SmallPtrSet<Value *, 2>> &Shared,
2539               const SmallSetVector<Value *, 32> &ExprsInSubprogram,
2540               const DataLayout &DL) {
2541       ExprLinearizer Lin(DL, Inst2Matrix, Shared, ExprsInSubprogram, L);
2542       Lin.linearizeExpr(L, 0, false, false);
2543       return Lin.getResult();
2544     }
2545   };
2546 };
2547 } // namespace
2548 
2549 PreservedAnalyses LowerMatrixIntrinsicsPass::run(Function &F,
2550                                                  FunctionAnalysisManager &AM) {
2551   auto &TTI = AM.getResult<TargetIRAnalysis>(F);
2552   OptimizationRemarkEmitter *ORE = nullptr;
2553   AAResults *AA = nullptr;
2554   DominatorTree *DT = nullptr;
2555   LoopInfo *LI = nullptr;
2556 
2557   if (!Minimal) {
2558     ORE = &AM.getResult<OptimizationRemarkEmitterAnalysis>(F);
2559     AA = &AM.getResult<AAManager>(F);
2560     DT = &AM.getResult<DominatorTreeAnalysis>(F);
2561     LI = &AM.getResult<LoopAnalysis>(F);
2562   }
2563 
2564   LowerMatrixIntrinsics LMT(F, TTI, AA, DT, LI, ORE);
2565   if (LMT.Visit()) {
2566     PreservedAnalyses PA;
2567     if (!Minimal) {
2568       PA.preserve<LoopAnalysis>();
2569       PA.preserve<DominatorTreeAnalysis>();
2570     }
2571     return PA;
2572   }
2573   return PreservedAnalyses::all();
2574 }
2575 
2576 void LowerMatrixIntrinsicsPass::printPipeline(
2577     raw_ostream &OS, function_ref<StringRef(StringRef)> MapClassName2PassName) {
2578   static_cast<PassInfoMixin<LowerMatrixIntrinsicsPass> *>(this)->printPipeline(
2579       OS, MapClassName2PassName);
2580   OS << '<';
2581   if (Minimal)
2582     OS << "minimal";
2583   OS << '>';
2584 }
2585