xref: /freebsd/contrib/llvm-project/llvm/lib/Transforms/Scalar/LowerMatrixIntrinsics.cpp (revision 9f23cbd6cae82fd77edfad7173432fa8dccd0a95)
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 enum class MatrixLayoutTy { ColumnMajor, RowMajor };
76 
77 static cl::opt<MatrixLayoutTy> MatrixLayout(
78     "matrix-default-layout", cl::init(MatrixLayoutTy::ColumnMajor),
79     cl::desc("Sets the default matrix layout"),
80     cl::values(clEnumValN(MatrixLayoutTy::ColumnMajor, "column-major",
81                           "Use column-major layout"),
82                clEnumValN(MatrixLayoutTy::RowMajor, "row-major",
83                           "Use row-major layout")));
84 
85 static cl::opt<bool> PrintAfterTransposeOpt("matrix-print-after-transpose-opt",
86                                             cl::init(false));
87 
88 /// Helper function to either return Scope, if it is a subprogram or the
89 /// attached subprogram for a local scope.
90 static DISubprogram *getSubprogram(DIScope *Scope) {
91   if (auto *Subprogram = dyn_cast<DISubprogram>(Scope))
92     return Subprogram;
93   return cast<DILocalScope>(Scope)->getSubprogram();
94 }
95 
96 /// Erase \p V from \p BB and move \II forward to avoid invalidating
97 /// iterators.
98 static void eraseFromParentAndMove(Value *V, BasicBlock::reverse_iterator &II,
99                                    BasicBlock &BB) {
100   auto *Inst = cast<Instruction>(V);
101   // Still used, don't erase.
102   if (!Inst->use_empty())
103     return;
104   if (II != BB.rend() && Inst == &*II)
105     ++II;
106   Inst->eraseFromParent();
107 }
108 
109 /// Return true if V is a splat of a value (which is used when multiplying a
110 /// matrix with a scalar).
111 static bool isSplat(Value *V) {
112   if (auto *SV = dyn_cast<ShuffleVectorInst>(V))
113     return SV->isZeroEltSplat();
114   return false;
115 }
116 
117 /// Match any mul operation (fp or integer).
118 template <typename LTy, typename RTy>
119 auto m_AnyMul(const LTy &L, const RTy &R) {
120   return m_CombineOr(m_Mul(L, R), m_FMul(L, R));
121 }
122 
123 /// Match any add operation (fp or integer).
124 template <typename LTy, typename RTy>
125 auto m_AnyAdd(const LTy &L, const RTy &R) {
126   return m_CombineOr(m_Add(L, R), m_FAdd(L, R));
127 }
128 
129 namespace {
130 
131 // Given an element pointer \p BasePtr to the start of a (sub) matrix, compute
132 // the start address of vector \p VecIdx with type (\p EltType x \p NumElements)
133 // assuming \p Stride elements between start two consecutive vectors.
134 // \p Stride must be >= \p NumElements.
135 // For column-major matrixes, the function computes the address of a column
136 // vectors and \p NumElements must be set to the number of elements in a column
137 // (= number of rows of the matrix). For row-major matrixes, the function
138 // computes the address of a row vector and \p NumElements must be set to the
139 // number of elements in a column (= number of columns of the matrix).
140 //
141 // Consider a 4x4 matrix in column-mjaor layout like below
142 //
143 //      0       1      2      3
144 // 0   v_0_0  v_0_1  v_0_2  v_0_3
145 // 1   v_1_0  v_1_1  v_1_2  v_1_3
146 // 2   v_2_0  v_2_1  v_2_2  v_2_3
147 // 3   v_3_0  v_3_1  v_3_2  v_3_3
148 
149 // To compute the column addresses for a 2x3 sub-matrix at row 1 and column 1,
150 // we need a pointer to the first element of the submatrix as base pointer.
151 // Then we can use computeVectorAddr to compute the addresses for the columns
152 // of the sub-matrix.
153 //
154 // Column 0: computeVectorAddr(Base, 0 (column), 4 (stride), 2 (num rows), ..)
155 //           -> just returns Base
156 // Column 1: computeVectorAddr(Base, 1 (column), 4 (stride), 2 (num rows), ..)
157 //           -> returns Base + (1 * 4)
158 // Column 2: computeVectorAddr(Base, 2 (column), 4 (stride), 2 (num rows), ..)
159 //           -> returns Base + (2 * 4)
160 //
161 // The graphic below illustrates the number of elements in a column (marked
162 // with |) and the number of skipped elements (marked with }).
163 //
164 //         v_0_0  v_0_1 {v_0_2 {v_0_3
165 //                Base   Col 1  Col 2
166 //                  |     |      |
167 //         v_1_0 |v_1_1 |v_1_2 |v_1_3
168 //         v_2_0 |v_2_1 |v_2_2 |v_2_3
169 //         v_3_0 {v_3_1 {v_3_2  v_3_3
170 //
171 Value *computeVectorAddr(Value *BasePtr, Value *VecIdx, Value *Stride,
172                          unsigned NumElements, Type *EltType,
173                          IRBuilder<> &Builder) {
174 
175   assert((!isa<ConstantInt>(Stride) ||
176           cast<ConstantInt>(Stride)->getZExtValue() >= NumElements) &&
177          "Stride must be >= the number of elements in the result vector.");
178   unsigned AS = cast<PointerType>(BasePtr->getType())->getAddressSpace();
179 
180   // Compute the start of the vector with index VecIdx as VecIdx * Stride.
181   Value *VecStart = Builder.CreateMul(VecIdx, Stride, "vec.start");
182 
183   // Get pointer to the start of the selected vector. Skip GEP creation,
184   // if we select vector 0.
185   if (isa<ConstantInt>(VecStart) && cast<ConstantInt>(VecStart)->isZero())
186     VecStart = BasePtr;
187   else
188     VecStart = Builder.CreateGEP(EltType, BasePtr, VecStart, "vec.gep");
189 
190   // Cast elementwise vector start pointer to a pointer to a vector
191   // (EltType x NumElements)*.
192   auto *VecType = FixedVectorType::get(EltType, NumElements);
193   Type *VecPtrType = PointerType::get(VecType, AS);
194   return Builder.CreatePointerCast(VecStart, VecPtrType, "vec.cast");
195 }
196 
197 /// LowerMatrixIntrinsics contains the methods used to lower matrix intrinsics.
198 ///
199 /// Currently, the lowering for each matrix intrinsic is done as follows:
200 /// 1. Propagate the shape information from intrinsics to connected
201 /// instructions.
202 /// 2. Lower instructions with shape information (assuming column-major layout).
203 ///  The lowering works similarly using row-major layout.
204 ///  2.1. Get column vectors for each argument. If we already lowered the
205 ///       definition of an argument, use the produced column vectors directly.
206 ///       If not, split the operand vector containing an embedded matrix into
207 ///       a set of column vectors,
208 ///  2.2. Lower the instruction in terms of column major operations, which
209 ///       yields a set of column vectors containing result matrix. Note that we
210 ///       lower all instructions that have shape information. Besides the
211 ///       intrinsics, this includes stores for example.
212 ///  2.3. Update uses of the lowered instruction. If we have shape information
213 ///       for a user, there is nothing to do, as we will look up the result
214 ///       column matrix when lowering the user. For other uses, we embed the
215 ///       result matrix in a flat vector and update the use.
216 ///  2.4. Cache the result column matrix for the instruction we lowered
217 /// 3. After we lowered all instructions in a function, remove the now
218 ///    obsolete instructions.
219 ///
220 class LowerMatrixIntrinsics {
221   Function &Func;
222   const DataLayout &DL;
223   const TargetTransformInfo &TTI;
224   AliasAnalysis *AA;
225   DominatorTree *DT;
226   LoopInfo *LI;
227   OptimizationRemarkEmitter *ORE;
228 
229   /// Contains estimates of the number of operations (loads, stores, compute) required to lower a matrix operation.
230   struct OpInfoTy {
231     /// Number of stores emitted to generate this matrix.
232     unsigned NumStores = 0;
233     /// Number of loads emitted to generate this matrix.
234     unsigned NumLoads = 0;
235     /// Number of compute operations emitted to generate this matrix.
236     unsigned NumComputeOps = 0;
237     /// Most of the time transposes can be fused with matrix multiplies or can
238     /// be folded away via algebraic simplifications.  This is the number of
239     /// transposes that we failed to make "free" via such optimizations.
240     unsigned NumExposedTransposes = 0;
241 
242     OpInfoTy &operator+=(const OpInfoTy &RHS) {
243       NumStores += RHS.NumStores;
244       NumLoads += RHS.NumLoads;
245       NumComputeOps += RHS.NumComputeOps;
246       NumExposedTransposes += RHS.NumExposedTransposes;
247       return *this;
248     }
249   };
250 
251   /// Wrapper class representing a matrix as a set of vectors, either in row or
252   /// column major layout. All vectors must have the same vector type.
253   class MatrixTy {
254     SmallVector<Value *, 16> Vectors;
255 
256     OpInfoTy OpInfo;
257 
258     bool IsColumnMajor = true;
259 
260   public:
261     MatrixTy() : IsColumnMajor(MatrixLayout == MatrixLayoutTy::ColumnMajor) {}
262     MatrixTy(ArrayRef<Value *> Vectors)
263         : Vectors(Vectors.begin(), Vectors.end()),
264           IsColumnMajor(MatrixLayout == MatrixLayoutTy::ColumnMajor) {}
265     MatrixTy(unsigned NumRows, unsigned NumColumns, Type *EltTy)
266         : IsColumnMajor(MatrixLayout == MatrixLayoutTy::ColumnMajor) {
267 
268       unsigned D = isColumnMajor() ? NumColumns : NumRows;
269       for (unsigned J = 0; J < D; ++J)
270         addVector(UndefValue::get(FixedVectorType::get(
271             EltTy, isColumnMajor() ? NumRows : NumColumns)));
272     }
273 
274     Value *getVector(unsigned i) const { return Vectors[i]; }
275     Value *getColumn(unsigned i) const {
276       assert(isColumnMajor() && "only supported for column-major matrixes");
277       return Vectors[i];
278     }
279     Value *getRow(unsigned i) const {
280       assert(!isColumnMajor() && "only supported for row-major matrixes");
281       return Vectors[i];
282     }
283 
284     void setVector(unsigned i, Value *V) { Vectors[i] = V; }
285 
286     Type *getElementType() const { return getVectorTy()->getElementType(); }
287 
288     unsigned getNumVectors() const {
289       if (isColumnMajor())
290         return getNumColumns();
291       return getNumRows();
292     }
293 
294     unsigned getNumColumns() const {
295       if (isColumnMajor())
296         return Vectors.size();
297       else {
298         assert(Vectors.size() > 0 && "Cannot call getNumRows without columns");
299         return cast<FixedVectorType>(Vectors[0]->getType())->getNumElements();
300       }
301     }
302     unsigned getNumRows() const {
303       if (isColumnMajor()) {
304         assert(Vectors.size() > 0 && "Cannot call getNumRows without columns");
305         return cast<FixedVectorType>(Vectors[0]->getType())->getNumElements();
306       } else
307         return Vectors.size();
308     }
309 
310     void addVector(Value *V) { Vectors.push_back(V); }
311     VectorType *getColumnTy() {
312       assert(isColumnMajor() && "only supported for column-major matrixes");
313       return getVectorTy();
314     }
315 
316     VectorType *getVectorTy() const {
317       return cast<VectorType>(Vectors[0]->getType());
318     }
319 
320     iterator_range<SmallVector<Value *, 8>::iterator> columns() {
321       assert(isColumnMajor() &&
322              "columns() only supported for column-major matrixes");
323       return make_range(Vectors.begin(), Vectors.end());
324     }
325 
326     iterator_range<SmallVector<Value *, 8>::iterator> vectors() {
327       return make_range(Vectors.begin(), Vectors.end());
328     }
329 
330     /// Embed the vectors of the matrix into a flat vector by concatenating
331     /// them.
332     Value *embedInVector(IRBuilder<> &Builder) const {
333       return Vectors.size() == 1 ? Vectors[0]
334                                  : concatenateVectors(Builder, Vectors);
335     }
336 
337     MatrixTy &addNumLoads(unsigned N) {
338       OpInfo.NumLoads += N;
339       return *this;
340     }
341 
342     void setNumLoads(unsigned N) { OpInfo.NumLoads = N; }
343 
344     MatrixTy &addNumStores(unsigned N) {
345       OpInfo.NumStores += N;
346       return *this;
347     }
348 
349     MatrixTy &addNumExposedTransposes(unsigned N) {
350       OpInfo.NumExposedTransposes += N;
351       return *this;
352     }
353 
354     MatrixTy &addNumComputeOps(unsigned N) {
355       OpInfo.NumComputeOps += N;
356       return *this;
357     }
358 
359     unsigned getNumStores() const { return OpInfo.NumStores; }
360     unsigned getNumLoads() const { return OpInfo.NumLoads; }
361     unsigned getNumComputeOps() const { return OpInfo.NumComputeOps; }
362 
363     const OpInfoTy &getOpInfo() const { return OpInfo; }
364 
365     bool isColumnMajor() const { return IsColumnMajor; }
366 
367     unsigned getStride() const {
368       if (isColumnMajor())
369         return getNumRows();
370       return getNumColumns();
371     }
372 
373     /// Extract a vector of \p NumElts starting at index (\p I, \p J). If the
374     /// matrix is column-major, the result vector is extracted from a column
375     /// vector, otherwise from a row vector.
376     Value *extractVector(unsigned I, unsigned J, unsigned NumElts,
377                          IRBuilder<> &Builder) const {
378       Value *Vec = isColumnMajor() ? getColumn(J) : getRow(I);
379       assert(cast<FixedVectorType>(Vec->getType())->getNumElements() >=
380                  NumElts &&
381              "Extracted vector will contain poison values");
382       return Builder.CreateShuffleVector(
383           Vec, createSequentialMask(isColumnMajor() ? I : J, NumElts, 0),
384           "block");
385     }
386   };
387 
388   struct ShapeInfo {
389     unsigned NumRows;
390     unsigned NumColumns;
391 
392     bool IsColumnMajor;
393 
394     ShapeInfo(unsigned NumRows = 0, unsigned NumColumns = 0)
395         : NumRows(NumRows), NumColumns(NumColumns),
396           IsColumnMajor(MatrixLayout == MatrixLayoutTy::ColumnMajor) {}
397 
398     ShapeInfo(Value *NumRows, Value *NumColumns)
399         : ShapeInfo(cast<ConstantInt>(NumRows)->getZExtValue(),
400                     cast<ConstantInt>(NumColumns)->getZExtValue()) {}
401 
402     bool operator==(const ShapeInfo &other) {
403       return NumRows == other.NumRows && NumColumns == other.NumColumns;
404     }
405     bool operator!=(const ShapeInfo &other) { return !(*this == other); }
406 
407     /// Returns true if shape-information is defined, meaning both dimensions
408     /// are != 0.
409     operator bool() const {
410       assert(NumRows == 0 || NumColumns != 0);
411       return NumRows != 0;
412     }
413 
414     unsigned getStride() const {
415       if (IsColumnMajor)
416         return NumRows;
417       return NumColumns;
418     }
419 
420     unsigned getNumVectors() const {
421       if (IsColumnMajor)
422         return NumColumns;
423       return NumRows;
424     }
425 
426     /// Returns the transposed shape.
427     ShapeInfo t() const { return ShapeInfo(NumColumns, NumRows); }
428   };
429 
430   /// Maps instructions to their shape information. The shape information
431   /// describes the shape to be used while lowering. This matches the shape of
432   /// the result value of the instruction, with the only exceptions being store
433   /// instructions and the matrix_column_major_store intrinsics. For those, the
434   /// shape information indicates that those instructions should be lowered
435   /// using shape information as well.  A ValueMap is used so that when
436   /// sub-passes like optimizeTransposes performs RAUW the map stays
437   /// up-to-date.
438   ValueMap<Value *, ShapeInfo> ShapeMap;
439 
440   /// List of instructions to remove. While lowering, we are not replacing all
441   /// users of a lowered instruction, if shape information is available and
442   /// those need to be removed after we finished lowering.
443   SmallVector<Instruction *, 16> ToRemove;
444 
445   /// Map from instructions to their produced column matrix.
446   MapVector<Value *, MatrixTy> Inst2ColumnMatrix;
447 
448 private:
449   static FastMathFlags getFastMathFlags(Instruction *Inst) {
450     FastMathFlags FMF;
451 
452     if (isa<FPMathOperator>(*Inst))
453       FMF = Inst->getFastMathFlags();
454 
455     FMF.setAllowContract(AllowContractEnabled || FMF.allowContract());
456 
457     return FMF;
458   }
459 
460 public:
461   LowerMatrixIntrinsics(Function &F, TargetTransformInfo &TTI,
462                         AliasAnalysis *AA, DominatorTree *DT, LoopInfo *LI,
463                         OptimizationRemarkEmitter *ORE)
464       : Func(F), DL(F.getParent()->getDataLayout()), TTI(TTI), AA(AA), DT(DT),
465         LI(LI), ORE(ORE) {}
466 
467   unsigned getNumOps(Type *VT) {
468     assert(isa<VectorType>(VT) && "Expected vector type");
469     return getNumOps(VT->getScalarType(),
470                      cast<FixedVectorType>(VT)->getNumElements());
471   }
472 
473   /// Is this the minimal version executed in the backend pipelines.
474   bool isMinimal() const {
475     return !DT;
476   }
477 
478   /// Return the estimated number of vector ops required for an operation on
479   /// \p VT * N.
480   unsigned getNumOps(Type *ST, unsigned N) {
481     return std::ceil((ST->getPrimitiveSizeInBits() * N).getFixedValue() /
482                      double(TTI.getRegisterBitWidth(
483                                    TargetTransformInfo::RGK_FixedWidthVector)
484                                 .getFixedValue()));
485   }
486 
487   /// Return the set of vectors that a matrix value is lowered to.
488   ///
489   /// If we lowered \p MatrixVal, just return the cache result matrix. Otherwise
490   /// split the flat vector \p MatrixVal containing a matrix with shape \p SI
491   /// into vectors.
492   MatrixTy getMatrix(Value *MatrixVal, const ShapeInfo &SI,
493                      IRBuilder<> &Builder) {
494     VectorType *VType = dyn_cast<VectorType>(MatrixVal->getType());
495     assert(VType && "MatrixVal must be a vector type");
496     assert(cast<FixedVectorType>(VType)->getNumElements() ==
497                SI.NumRows * SI.NumColumns &&
498            "The vector size must match the number of matrix elements");
499 
500     // Check if we lowered MatrixVal using shape information. In that case,
501     // return the existing matrix, if it matches the requested shape
502     // information. If there is a mis-match, embed the result in a flat
503     // vector and split it later.
504     auto Found = Inst2ColumnMatrix.find(MatrixVal);
505     if (Found != Inst2ColumnMatrix.end()) {
506       MatrixTy &M = Found->second;
507       // Return the found matrix, if its shape matches the requested shape
508       // information
509       if (SI.NumRows == M.getNumRows() && SI.NumColumns == M.getNumColumns())
510         return M;
511 
512       MatrixVal = M.embedInVector(Builder);
513     }
514 
515     // Otherwise split MatrixVal.
516     SmallVector<Value *, 16> SplitVecs;
517     for (unsigned MaskStart = 0;
518          MaskStart < cast<FixedVectorType>(VType)->getNumElements();
519          MaskStart += SI.getStride()) {
520       Value *V = Builder.CreateShuffleVector(
521           MatrixVal, createSequentialMask(MaskStart, SI.getStride(), 0),
522           "split");
523       SplitVecs.push_back(V);
524     }
525 
526     return {SplitVecs};
527   }
528 
529   /// If \p V already has a known shape return false.  Otherwise set the shape
530   /// for instructions that support it.
531   bool setShapeInfo(Value *V, ShapeInfo Shape) {
532     assert(Shape && "Shape not set");
533     if (isa<UndefValue>(V) || !supportsShapeInfo(V))
534       return false;
535 
536     auto SIter = ShapeMap.find(V);
537     if (SIter != ShapeMap.end()) {
538       LLVM_DEBUG(dbgs() << "  not overriding existing shape: "
539                         << SIter->second.NumRows << " "
540                         << SIter->second.NumColumns << " for " << *V << "\n");
541       return false;
542     }
543 
544     ShapeMap.insert({V, Shape});
545     LLVM_DEBUG(dbgs() << "  " << Shape.NumRows << " x " << Shape.NumColumns
546                       << " for " << *V << "\n");
547     return true;
548   }
549 
550   bool isUniformShape(Value *V) {
551     Instruction *I = dyn_cast<Instruction>(V);
552     if (!I)
553       return true;
554 
555     switch (I->getOpcode()) {
556     case Instruction::FAdd:
557     case Instruction::FSub:
558     case Instruction::FMul: // Scalar multiply.
559     case Instruction::FNeg:
560     case Instruction::Add:
561     case Instruction::Mul:
562     case Instruction::Sub:
563       return true;
564     default:
565       return false;
566     }
567   }
568 
569   /// Returns true if shape information can be used for \p V. The supported
570   /// instructions must match the instructions that can be lowered by this pass.
571   bool supportsShapeInfo(Value *V) {
572     Instruction *Inst = dyn_cast<Instruction>(V);
573     if (!Inst)
574       return false;
575 
576     IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst);
577     if (II)
578       switch (II->getIntrinsicID()) {
579       case Intrinsic::matrix_multiply:
580       case Intrinsic::matrix_transpose:
581       case Intrinsic::matrix_column_major_load:
582       case Intrinsic::matrix_column_major_store:
583         return true;
584       default:
585         return false;
586       }
587     return isUniformShape(V) || isa<StoreInst>(V) || isa<LoadInst>(V);
588   }
589 
590   /// Propagate the shape information of instructions to their users.
591   /// The work list contains instructions for which we can compute the shape,
592   /// either based on the information provided by matrix intrinsics or known
593   /// shapes of operands.
594   SmallVector<Instruction *, 32>
595   propagateShapeForward(SmallVectorImpl<Instruction *> &WorkList) {
596     SmallVector<Instruction *, 32> NewWorkList;
597     // Pop an element for which we guaranteed to have at least one of the
598     // operand shapes.  Add the shape for this and then add users to the work
599     // list.
600     LLVM_DEBUG(dbgs() << "Forward-propagate shapes:\n");
601     while (!WorkList.empty()) {
602       Instruction *Inst = WorkList.pop_back_val();
603 
604       // New entry, set the value and insert operands
605       bool Propagate = false;
606 
607       Value *MatrixA;
608       Value *MatrixB;
609       Value *M;
610       Value *N;
611       Value *K;
612       if (match(Inst, m_Intrinsic<Intrinsic::matrix_multiply>(
613                           m_Value(MatrixA), m_Value(MatrixB), m_Value(M),
614                           m_Value(N), m_Value(K)))) {
615         Propagate = setShapeInfo(Inst, {M, K});
616       } else if (match(Inst, m_Intrinsic<Intrinsic::matrix_transpose>(
617                                  m_Value(MatrixA), m_Value(M), m_Value(N)))) {
618         // Flip dimensions.
619         Propagate = setShapeInfo(Inst, {N, M});
620       } else if (match(Inst, m_Intrinsic<Intrinsic::matrix_column_major_store>(
621                                  m_Value(MatrixA), m_Value(), m_Value(),
622                                  m_Value(), m_Value(M), m_Value(N)))) {
623         Propagate = setShapeInfo(Inst, {N, M});
624       } else if (match(Inst, m_Intrinsic<Intrinsic::matrix_column_major_load>(
625                                  m_Value(), m_Value(), m_Value(), m_Value(M),
626                                  m_Value(N)))) {
627         Propagate = setShapeInfo(Inst, {M, N});
628       } else if (match(Inst, m_Store(m_Value(MatrixA), m_Value()))) {
629         auto OpShape = ShapeMap.find(MatrixA);
630         if (OpShape != ShapeMap.end())
631           setShapeInfo(Inst, OpShape->second);
632         continue;
633       } else if (isUniformShape(Inst)) {
634         // Find the first operand that has a known shape and use that.
635         for (auto &Op : Inst->operands()) {
636           auto OpShape = ShapeMap.find(Op.get());
637           if (OpShape != ShapeMap.end()) {
638             Propagate |= setShapeInfo(Inst, OpShape->second);
639             break;
640           }
641         }
642       }
643 
644       if (Propagate) {
645         NewWorkList.push_back(Inst);
646         for (auto *User : Inst->users())
647           if (ShapeMap.count(User) == 0)
648             WorkList.push_back(cast<Instruction>(User));
649       }
650     }
651 
652     return NewWorkList;
653   }
654 
655   /// Propagate the shape to operands of instructions with shape information.
656   /// \p Worklist contains the instruction for which we already know the shape.
657   SmallVector<Instruction *, 32>
658   propagateShapeBackward(SmallVectorImpl<Instruction *> &WorkList) {
659     SmallVector<Instruction *, 32> NewWorkList;
660 
661     auto pushInstruction = [](Value *V,
662                               SmallVectorImpl<Instruction *> &WorkList) {
663       Instruction *I = dyn_cast<Instruction>(V);
664       if (I)
665         WorkList.push_back(I);
666     };
667     // Pop an element with known shape.  Traverse the operands, if their shape
668     // derives from the result shape and is unknown, add it and add them to the
669     // worklist.
670     LLVM_DEBUG(dbgs() << "Backward-propagate shapes:\n");
671     while (!WorkList.empty()) {
672       Value *V = WorkList.pop_back_val();
673 
674       size_t BeforeProcessingV = WorkList.size();
675       if (!isa<Instruction>(V))
676         continue;
677 
678       Value *MatrixA;
679       Value *MatrixB;
680       Value *M;
681       Value *N;
682       Value *K;
683       if (match(V, m_Intrinsic<Intrinsic::matrix_multiply>(
684                        m_Value(MatrixA), m_Value(MatrixB), m_Value(M),
685                        m_Value(N), m_Value(K)))) {
686         if (setShapeInfo(MatrixA, {M, N}))
687           pushInstruction(MatrixA, WorkList);
688 
689         if (setShapeInfo(MatrixB, {N, K}))
690           pushInstruction(MatrixB, WorkList);
691 
692       } else if (match(V, m_Intrinsic<Intrinsic::matrix_transpose>(
693                               m_Value(MatrixA), m_Value(M), m_Value(N)))) {
694         // Flip dimensions.
695         if (setShapeInfo(MatrixA, {M, N}))
696           pushInstruction(MatrixA, WorkList);
697       } else if (match(V, m_Intrinsic<Intrinsic::matrix_column_major_store>(
698                               m_Value(MatrixA), m_Value(), m_Value(), m_Value(),
699                               m_Value(M), m_Value(N)))) {
700         if (setShapeInfo(MatrixA, {M, N})) {
701           pushInstruction(MatrixA, WorkList);
702         }
703       } else if (isa<LoadInst>(V) ||
704                  match(V, m_Intrinsic<Intrinsic::matrix_column_major_load>())) {
705         // Nothing to do, no matrix input.
706       } else if (isa<StoreInst>(V)) {
707         // Nothing to do.  We forward-propagated to this so we would just
708         // backward propagate to an instruction with an already known shape.
709       } else if (isUniformShape(V)) {
710         // Propagate to all operands.
711         ShapeInfo Shape = ShapeMap[V];
712         for (Use &U : cast<Instruction>(V)->operands()) {
713           if (setShapeInfo(U.get(), Shape))
714             pushInstruction(U.get(), WorkList);
715         }
716       }
717       // After we discovered new shape info for new instructions in the
718       // worklist, we use their users as seeds for the next round of forward
719       // propagation.
720       for (size_t I = BeforeProcessingV; I != WorkList.size(); I++)
721         for (User *U : WorkList[I]->users())
722           if (isa<Instruction>(U) && V != U)
723             NewWorkList.push_back(cast<Instruction>(U));
724     }
725     return NewWorkList;
726   }
727 
728   /// (Op0 op Op1)^T -> Op0^T op Op1^T
729   /// Transpose \p Op0 and \p Op1 of shape \p Shape0 and \p Shape1, then use
730   /// them on both sides of \p Operation.
731   Instruction *distributeTransposes(
732       Value *Op0, ShapeInfo Shape0, Value *Op1, ShapeInfo Shape1,
733       MatrixBuilder &Builder,
734       function_ref<Instruction *(Value *, ShapeInfo, Value *, ShapeInfo)>
735           Operation) {
736     Value *T0 = Builder.CreateMatrixTranspose(
737         Op0, Shape0.NumRows, Shape0.NumColumns, Op0->getName() + "_t");
738     // We are being run after shape prop, add shape for newly created
739     // instructions so that we lower them later.
740     setShapeInfo(T0, Shape0.t());
741     Value *T1 = Builder.CreateMatrixTranspose(
742         Op1, Shape1.NumRows, Shape1.NumColumns, Op1->getName() + "_t");
743     setShapeInfo(T1, Shape1.t());
744     return Operation(T0, Shape0.t(), T1, Shape1.t());
745   }
746 
747   void updateShapeAndReplaceAllUsesWith(Instruction &Old, Value *New) {
748     // We need to remove Old from the ShapeMap otherwise RAUW will replace it
749     // with New. We should only add New it it supportsShapeInfo so we insert
750     // it conditionally instead.
751     auto S = ShapeMap.find(&Old);
752     if (S != ShapeMap.end()) {
753       ShapeMap.erase(S);
754       if (supportsShapeInfo(New))
755         ShapeMap.insert({New, S->second});
756     }
757     Old.replaceAllUsesWith(New);
758   }
759 
760   /// Sink a top-level transpose inside matmuls and adds.
761   /// This creates and erases instructions as needed, and returns the newly
762   /// created instruction while updating the iterator to avoid invalidation. If
763   /// this returns nullptr, no new instruction was created.
764   Instruction *sinkTranspose(Instruction &I, BasicBlock::reverse_iterator &II) {
765     BasicBlock &BB = *I.getParent();
766     IRBuilder<> IB(&I);
767     MatrixBuilder Builder(IB);
768 
769     Value *TA, *TAMA, *TAMB;
770     ConstantInt *R, *K, *C;
771     if (!match(&I, m_Intrinsic<Intrinsic::matrix_transpose>(
772                        m_Value(TA), m_ConstantInt(R), m_ConstantInt(C))))
773       return nullptr;
774 
775     // Transpose of a transpose is a nop
776     Value *TATA;
777     if (match(TA, m_Intrinsic<Intrinsic::matrix_transpose>(m_Value(TATA)))) {
778       updateShapeAndReplaceAllUsesWith(I, TATA);
779       eraseFromParentAndMove(&I, II, BB);
780       eraseFromParentAndMove(TA, II, BB);
781       return nullptr;
782     }
783 
784     // k^T -> k
785     if (isSplat(TA)) {
786       updateShapeAndReplaceAllUsesWith(I, TA);
787       eraseFromParentAndMove(&I, II, BB);
788       return nullptr;
789     }
790 
791     // (A * B)^t -> B^t * A^t
792     // RxK KxC      CxK   KxR
793     if (match(TA, m_Intrinsic<Intrinsic::matrix_multiply>(
794                       m_Value(TAMA), m_Value(TAMB), m_ConstantInt(R),
795                       m_ConstantInt(K), m_ConstantInt(C)))) {
796       auto NewInst = distributeTransposes(
797           TAMB, {K, C}, TAMA, {R, K}, Builder,
798           [&](Value *T0, ShapeInfo Shape0, Value *T1, ShapeInfo Shape1) {
799             return Builder.CreateMatrixMultiply(T0, T1, Shape0.NumRows,
800                                                 Shape0.NumColumns,
801                                                 Shape1.NumColumns, "mmul");
802           });
803       updateShapeAndReplaceAllUsesWith(I, NewInst);
804       eraseFromParentAndMove(&I, II, BB);
805       eraseFromParentAndMove(TA, II, BB);
806       return NewInst;
807     }
808 
809     // Same as above, but with a mul, which occurs when multiplied
810     // with a scalar.
811     // (A * k)^t -> A^t * k
812     //  R  x  C     RxC
813     if (match(TA, m_AnyMul(m_Value(TAMA), m_Value(TAMB))) &&
814         (isSplat(TAMA) || isSplat(TAMB))) {
815       IRBuilder<> LocalBuilder(&I);
816       // We know that the transposed operand is of shape RxC.
817       // An when multiplied with a scalar, the shape is preserved.
818       auto NewInst = distributeTransposes(
819           TAMA, {R, C}, TAMB, {R, C}, Builder,
820           [&](Value *T0, ShapeInfo Shape0, Value *T1, ShapeInfo Shape1) {
821             bool IsFP = I.getType()->isFPOrFPVectorTy();
822             auto *Mul = IsFP ? LocalBuilder.CreateFMul(T0, T1, "mmul")
823                              : LocalBuilder.CreateMul(T0, T1, "mmul");
824             auto *Result = cast<Instruction>(Mul);
825             setShapeInfo(Result, Shape0);
826             return Result;
827           });
828       updateShapeAndReplaceAllUsesWith(I, NewInst);
829       eraseFromParentAndMove(&I, II, BB);
830       eraseFromParentAndMove(TA, II, BB);
831       return NewInst;
832     }
833 
834     // (A + B)^t -> A^t + B^t
835     // RxC RxC      CxR   CxR
836     if (match(TA, m_AnyAdd(m_Value(TAMA), m_Value(TAMB)))) {
837       IRBuilder<> LocalBuilder(&I);
838       auto NewInst = distributeTransposes(
839           TAMA, {R, C}, TAMB, {R, C}, Builder,
840           [&](Value *T0, ShapeInfo Shape0, Value *T1, ShapeInfo Shape1) {
841             auto *FAdd =
842                 cast<Instruction>(LocalBuilder.CreateFAdd(T0, T1, "mfadd"));
843             setShapeInfo(FAdd, Shape0);
844             return FAdd;
845           });
846       updateShapeAndReplaceAllUsesWith(I, NewInst);
847       eraseFromParentAndMove(&I, II, BB);
848       eraseFromParentAndMove(TA, II, BB);
849       return NewInst;
850     }
851 
852     return nullptr;
853   }
854 
855   void liftTranspose(Instruction &I) {
856     // Erase dead Instructions after lifting transposes from binops.
857     auto CleanupBinOp = [](Instruction &T, Value *A, Value *B) {
858       if (T.use_empty())
859         T.eraseFromParent();
860       if (A->use_empty())
861         cast<Instruction>(A)->eraseFromParent();
862       if (A != B && B->use_empty())
863         cast<Instruction>(B)->eraseFromParent();
864     };
865 
866     Value *A, *B, *AT, *BT;
867     ConstantInt *R, *K, *C;
868     // A^t * B ^t -> (B * A)^t
869     if (match(&I, m_Intrinsic<Intrinsic::matrix_multiply>(
870                       m_Value(A), m_Value(B), m_ConstantInt(R),
871                       m_ConstantInt(K), m_ConstantInt(C))) &&
872         match(A, m_Intrinsic<Intrinsic::matrix_transpose>(m_Value(AT))) &&
873         match(B, m_Intrinsic<Intrinsic::matrix_transpose>(m_Value((BT))))) {
874       IRBuilder<> IB(&I);
875       MatrixBuilder Builder(IB);
876       Value *M = Builder.CreateMatrixMultiply(
877           BT, AT, C->getZExtValue(), K->getZExtValue(), R->getZExtValue());
878       setShapeInfo(M, {C, R});
879       Instruction *NewInst = Builder.CreateMatrixTranspose(M, C->getZExtValue(),
880                                                            R->getZExtValue());
881       updateShapeAndReplaceAllUsesWith(I, NewInst);
882       CleanupBinOp(I, A, B);
883     }
884     // A^t + B ^t -> (A + B)^t
885     else if (match(&I, m_FAdd(m_Value(A), m_Value(B))) &&
886              match(A, m_Intrinsic<Intrinsic::matrix_transpose>(
887                           m_Value(AT), m_ConstantInt(R), m_ConstantInt(C))) &&
888              match(B, m_Intrinsic<Intrinsic::matrix_transpose>(
889                           m_Value(BT), m_ConstantInt(R), m_ConstantInt(C)))) {
890       IRBuilder<> Builder(&I);
891       Value *Add = cast<Instruction>(Builder.CreateFAdd(AT, BT, "mfadd"));
892       setShapeInfo(Add, {C, R});
893       MatrixBuilder MBuilder(Builder);
894       Instruction *NewInst = MBuilder.CreateMatrixTranspose(
895           Add, C->getZExtValue(), R->getZExtValue(), "mfadd_t");
896       updateShapeAndReplaceAllUsesWith(I, NewInst);
897       CleanupBinOp(I, A, B);
898     }
899   }
900 
901   /// Try moving transposes in order to fold them away or into multiplies.
902   void optimizeTransposes() {
903     // First sink all transposes inside matmuls and adds, hoping that we end up
904     // with NN, NT or TN variants.
905     for (BasicBlock &BB : reverse(Func)) {
906       for (auto II = BB.rbegin(); II != BB.rend();) {
907         Instruction &I = *II;
908         // We may remove II.  By default continue on the next/prev instruction.
909         ++II;
910         if (Instruction *NewInst = sinkTranspose(I, II))
911           II = std::next(BasicBlock::reverse_iterator(NewInst));
912       }
913     }
914 
915     // If we have a TT matmul or a TT add, lift the transpose. We may be able
916     // to fold into consuming multiply or add.
917     for (BasicBlock &BB : Func) {
918       for (Instruction &I : llvm::make_early_inc_range(BB)) {
919         liftTranspose(I);
920       }
921     }
922   }
923 
924   bool Visit() {
925     SmallVector<Instruction *, 32> WorkList;
926 
927     // Initially only the shape of matrix intrinsics is known.
928     // Initialize the work list with ops carrying shape information.
929     for (BasicBlock &BB : Func)
930       for (Instruction &Inst : BB) {
931         IntrinsicInst *II = dyn_cast<IntrinsicInst>(&Inst);
932         if (!II)
933           continue;
934 
935         switch (II->getIntrinsicID()) {
936         case Intrinsic::matrix_multiply:
937         case Intrinsic::matrix_transpose:
938         case Intrinsic::matrix_column_major_load:
939         case Intrinsic::matrix_column_major_store:
940           WorkList.push_back(&Inst);
941           break;
942         default:
943           break;
944         }
945       }
946 
947     // Avoid unnecessary work if there are no matrix intrinsics in the function.
948     if (WorkList.empty())
949       return false;
950 
951     // Propagate shapes until nothing changes any longer.
952     while (!WorkList.empty()) {
953       WorkList = propagateShapeForward(WorkList);
954       WorkList = propagateShapeBackward(WorkList);
955     }
956 
957     if (!isMinimal()) {
958       optimizeTransposes();
959       if (PrintAfterTransposeOpt) {
960         dbgs() << "Dump after matrix transpose optimization:\n";
961         Func.print(dbgs());
962       }
963     }
964 
965     bool Changed = false;
966     SmallVector<CallInst *, 16> MaybeFusableInsts;
967     SmallVector<Instruction *, 16> MatrixInsts;
968 
969     // First, collect all instructions with shape information and candidates for
970     // fusion (currently only matrix multiplies).
971     ReversePostOrderTraversal<Function *> RPOT(&Func);
972     for (auto *BB : RPOT)
973       for (Instruction &I : *BB) {
974         if (ShapeMap.find(&I) == ShapeMap.end())
975           continue;
976         if (match(&I, m_Intrinsic<Intrinsic::matrix_multiply>()))
977           MaybeFusableInsts.push_back(cast<CallInst>(&I));
978         MatrixInsts.push_back(&I);
979       }
980 
981     // Second, try to fuse candidates.
982     SmallPtrSet<Instruction *, 16> FusedInsts;
983     for (CallInst *CI : MaybeFusableInsts)
984       LowerMatrixMultiplyFused(CI, FusedInsts);
985     Changed = !FusedInsts.empty();
986 
987     // Third, lower remaining instructions with shape information.
988     for (Instruction *Inst : MatrixInsts) {
989       if (FusedInsts.count(Inst))
990         continue;
991 
992       IRBuilder<> Builder(Inst);
993 
994       if (CallInst *CInst = dyn_cast<CallInst>(Inst))
995         Changed |= VisitCallInst(CInst);
996 
997       Value *Op1;
998       Value *Op2;
999       if (auto *BinOp = dyn_cast<BinaryOperator>(Inst))
1000         Changed |= VisitBinaryOperator(BinOp);
1001       if (auto *UnOp = dyn_cast<UnaryOperator>(Inst))
1002         Changed |= VisitUnaryOperator(UnOp);
1003       if (match(Inst, m_Load(m_Value(Op1))))
1004         Changed |= VisitLoad(cast<LoadInst>(Inst), Op1, Builder);
1005       else if (match(Inst, m_Store(m_Value(Op1), m_Value(Op2))))
1006         Changed |= VisitStore(cast<StoreInst>(Inst), Op1, Op2, Builder);
1007     }
1008 
1009     if (ORE) {
1010       RemarkGenerator RemarkGen(Inst2ColumnMatrix, *ORE, Func);
1011       RemarkGen.emitRemarks();
1012     }
1013 
1014     // Delete the instructions backwards, as it has a reduced likelihood of
1015     // having to update as many def-use and use-def chains.
1016     //
1017     // Because we add to ToRemove during fusion we can't guarantee that defs
1018     // are before uses.  Change uses to poison temporarily as these should get
1019     // removed as well.
1020     //
1021     // For verification, we keep track of where we changed uses to poison in
1022     // PoisonedInsts and then check that we in fact remove them.
1023     SmallSet<Instruction *, 16> PoisonedInsts;
1024     for (auto *Inst : reverse(ToRemove)) {
1025       for (Use &U : llvm::make_early_inc_range(Inst->uses())) {
1026         if (auto *Poisoned = dyn_cast<Instruction>(U.getUser()))
1027           PoisonedInsts.insert(Poisoned);
1028         U.set(PoisonValue::get(Inst->getType()));
1029       }
1030       Inst->eraseFromParent();
1031       PoisonedInsts.erase(Inst);
1032     }
1033     if (!PoisonedInsts.empty()) {
1034       // If we didn't remove all poisoned instructions, it's a hard error.
1035       dbgs() << "Poisoned but present instructions:\n";
1036       for (auto *I : PoisonedInsts)
1037         dbgs() << *I << "\n";
1038       llvm_unreachable("Poisoned but instruction not removed");
1039     }
1040 
1041     return Changed;
1042   }
1043 
1044   /// Turns \p BasePtr into an elementwise pointer to \p EltType.
1045   Value *createElementPtr(Value *BasePtr, Type *EltType, IRBuilder<> &Builder) {
1046     unsigned AS = cast<PointerType>(BasePtr->getType())->getAddressSpace();
1047     Type *EltPtrType = PointerType::get(EltType, AS);
1048     return Builder.CreatePointerCast(BasePtr, EltPtrType);
1049   }
1050 
1051   /// Replace intrinsic calls
1052   bool VisitCallInst(CallInst *Inst) {
1053     if (!Inst->getCalledFunction() || !Inst->getCalledFunction()->isIntrinsic())
1054       return false;
1055 
1056     switch (Inst->getCalledFunction()->getIntrinsicID()) {
1057     case Intrinsic::matrix_multiply:
1058       LowerMultiply(Inst);
1059       break;
1060     case Intrinsic::matrix_transpose:
1061       LowerTranspose(Inst);
1062       break;
1063     case Intrinsic::matrix_column_major_load:
1064       LowerColumnMajorLoad(Inst);
1065       break;
1066     case Intrinsic::matrix_column_major_store:
1067       LowerColumnMajorStore(Inst);
1068       break;
1069     default:
1070       return false;
1071     }
1072     return true;
1073   }
1074 
1075   /// Compute the alignment for a column/row \p Idx with \p Stride between them.
1076   /// The address at \p Idx == 0 has alignment \p A. If \p Stride is a
1077   /// ConstantInt, reduce the initial alignment based on the byte offset. For
1078   /// non-ConstantInt strides, return the common alignment of the initial
1079   /// alignment and the element size in bytes.
1080   Align getAlignForIndex(unsigned Idx, Value *Stride, Type *ElementTy,
1081                          MaybeAlign A) const {
1082     Align InitialAlign = DL.getValueOrABITypeAlignment(A, ElementTy);
1083     if (Idx == 0)
1084       return InitialAlign;
1085 
1086     TypeSize ElementSizeInBits = DL.getTypeSizeInBits(ElementTy);
1087     if (auto *ConstStride = dyn_cast<ConstantInt>(Stride)) {
1088       uint64_t StrideInBytes =
1089           ConstStride->getZExtValue() * ElementSizeInBits / 8;
1090       return commonAlignment(InitialAlign, Idx * StrideInBytes);
1091     }
1092     return commonAlignment(InitialAlign, ElementSizeInBits / 8);
1093   }
1094 
1095   /// Load a matrix with \p Shape starting at \p Ptr and using \p Stride between
1096   /// vectors.
1097   MatrixTy loadMatrix(Type *Ty, Value *Ptr, MaybeAlign MAlign, Value *Stride,
1098                       bool IsVolatile, ShapeInfo Shape, IRBuilder<> &Builder) {
1099     auto *VType = cast<VectorType>(Ty);
1100     Type *EltTy = VType->getElementType();
1101     Type *VecTy = FixedVectorType::get(EltTy, Shape.getStride());
1102     Value *EltPtr = createElementPtr(Ptr, EltTy, Builder);
1103     MatrixTy Result;
1104     for (unsigned I = 0, E = Shape.getNumVectors(); I < E; ++I) {
1105       Value *GEP = computeVectorAddr(
1106           EltPtr, Builder.getIntN(Stride->getType()->getScalarSizeInBits(), I),
1107           Stride, Shape.getStride(), EltTy, Builder);
1108       Value *Vector = Builder.CreateAlignedLoad(
1109           VecTy, GEP, getAlignForIndex(I, Stride, EltTy, MAlign),
1110           IsVolatile, "col.load");
1111 
1112       Result.addVector(Vector);
1113     }
1114     return Result.addNumLoads(getNumOps(Result.getVectorTy()) *
1115                               Result.getNumVectors());
1116   }
1117 
1118   /// Loads a sub-matrix with shape \p ResultShape from a \p R x \p C matrix,
1119   /// starting at \p MatrixPtr[I][J].
1120   MatrixTy loadMatrix(Value *MatrixPtr, MaybeAlign Align, bool IsVolatile,
1121                       ShapeInfo MatrixShape, Value *I, Value *J,
1122                       ShapeInfo ResultShape, Type *EltTy,
1123                       IRBuilder<> &Builder) {
1124 
1125     Value *Offset = Builder.CreateAdd(
1126         Builder.CreateMul(J, Builder.getInt64(MatrixShape.getStride())), I);
1127 
1128     unsigned AS = cast<PointerType>(MatrixPtr->getType())->getAddressSpace();
1129     Value *EltPtr =
1130         Builder.CreatePointerCast(MatrixPtr, PointerType::get(EltTy, AS));
1131     Value *TileStart = Builder.CreateGEP(EltTy, EltPtr, Offset);
1132     auto *TileTy = FixedVectorType::get(EltTy, ResultShape.NumRows *
1133                                                    ResultShape.NumColumns);
1134     Type *TilePtrTy = PointerType::get(TileTy, AS);
1135     Value *TilePtr =
1136         Builder.CreatePointerCast(TileStart, TilePtrTy, "col.cast");
1137 
1138     return loadMatrix(TileTy, TilePtr, Align,
1139                       Builder.getInt64(MatrixShape.getStride()), IsVolatile,
1140                       ResultShape, Builder);
1141   }
1142 
1143   /// Lower a load instruction with shape information.
1144   void LowerLoad(Instruction *Inst, Value *Ptr, MaybeAlign Align, Value *Stride,
1145                  bool IsVolatile, ShapeInfo Shape) {
1146     IRBuilder<> Builder(Inst);
1147     finalizeLowering(Inst,
1148                      loadMatrix(Inst->getType(), Ptr, Align, Stride, IsVolatile,
1149                                 Shape, Builder),
1150                      Builder);
1151   }
1152 
1153   /// Lowers llvm.matrix.column.major.load.
1154   ///
1155   /// The intrinsic loads a matrix from memory using a stride between columns.
1156   void LowerColumnMajorLoad(CallInst *Inst) {
1157     assert(MatrixLayout == MatrixLayoutTy::ColumnMajor &&
1158            "Intrinsic only supports column-major layout!");
1159     Value *Ptr = Inst->getArgOperand(0);
1160     Value *Stride = Inst->getArgOperand(1);
1161     LowerLoad(Inst, Ptr, Inst->getParamAlign(0), Stride,
1162               cast<ConstantInt>(Inst->getArgOperand(2))->isOne(),
1163               {Inst->getArgOperand(3), Inst->getArgOperand(4)});
1164   }
1165 
1166   /// Stores a sub-matrix \p StoreVal into the \p R x \p C matrix starting at \p
1167   /// MatrixPtr[I][J].
1168   void storeMatrix(const MatrixTy &StoreVal, Value *MatrixPtr,
1169                    MaybeAlign MAlign, bool IsVolatile, ShapeInfo MatrixShape,
1170                    Value *I, Value *J, Type *EltTy, IRBuilder<> &Builder) {
1171     Value *Offset = Builder.CreateAdd(
1172         Builder.CreateMul(J, Builder.getInt64(MatrixShape.getStride())), I);
1173 
1174     unsigned AS = cast<PointerType>(MatrixPtr->getType())->getAddressSpace();
1175     Value *EltPtr =
1176         Builder.CreatePointerCast(MatrixPtr, PointerType::get(EltTy, AS));
1177     Value *TileStart = Builder.CreateGEP(EltTy, EltPtr, Offset);
1178     auto *TileTy = FixedVectorType::get(EltTy, StoreVal.getNumRows() *
1179                                                    StoreVal.getNumColumns());
1180     Type *TilePtrTy = PointerType::get(TileTy, AS);
1181     Value *TilePtr =
1182         Builder.CreatePointerCast(TileStart, TilePtrTy, "col.cast");
1183 
1184     storeMatrix(TileTy, StoreVal, TilePtr, MAlign,
1185                 Builder.getInt64(MatrixShape.getStride()), IsVolatile, Builder);
1186   }
1187 
1188   /// Store matrix \p StoreVal starting at \p Ptr and using \p Stride between
1189   /// vectors.
1190   MatrixTy storeMatrix(Type *Ty, MatrixTy StoreVal, Value *Ptr,
1191                        MaybeAlign MAlign, Value *Stride, bool IsVolatile,
1192                        IRBuilder<> &Builder) {
1193     auto VType = cast<VectorType>(Ty);
1194     Value *EltPtr = createElementPtr(Ptr, VType->getElementType(), Builder);
1195     for (auto Vec : enumerate(StoreVal.vectors())) {
1196       Value *GEP = computeVectorAddr(
1197           EltPtr,
1198           Builder.getIntN(Stride->getType()->getScalarSizeInBits(),
1199                           Vec.index()),
1200           Stride, StoreVal.getStride(), VType->getElementType(), Builder);
1201       Builder.CreateAlignedStore(Vec.value(), GEP,
1202                                  getAlignForIndex(Vec.index(), Stride,
1203                                                   VType->getElementType(),
1204                                                   MAlign),
1205                                  IsVolatile);
1206     }
1207     return MatrixTy().addNumStores(getNumOps(StoreVal.getVectorTy()) *
1208                                    StoreVal.getNumVectors());
1209   }
1210 
1211   /// Lower a store instruction with shape information.
1212   void LowerStore(Instruction *Inst, Value *Matrix, Value *Ptr, MaybeAlign A,
1213                   Value *Stride, bool IsVolatile, ShapeInfo Shape) {
1214     IRBuilder<> Builder(Inst);
1215     auto StoreVal = getMatrix(Matrix, Shape, Builder);
1216     finalizeLowering(Inst,
1217                      storeMatrix(Matrix->getType(), StoreVal, Ptr, A, Stride,
1218                                  IsVolatile, Builder),
1219                      Builder);
1220   }
1221 
1222   /// Lowers llvm.matrix.column.major.store.
1223   ///
1224   /// The intrinsic store a matrix back memory using a stride between columns.
1225   void LowerColumnMajorStore(CallInst *Inst) {
1226     assert(MatrixLayout == MatrixLayoutTy::ColumnMajor &&
1227            "Intrinsic only supports column-major layout!");
1228     Value *Matrix = Inst->getArgOperand(0);
1229     Value *Ptr = Inst->getArgOperand(1);
1230     Value *Stride = Inst->getArgOperand(2);
1231     LowerStore(Inst, Matrix, Ptr, Inst->getParamAlign(1), Stride,
1232                cast<ConstantInt>(Inst->getArgOperand(3))->isOne(),
1233                {Inst->getArgOperand(4), Inst->getArgOperand(5)});
1234   }
1235 
1236   // Set elements I..I+NumElts-1 to Block
1237   Value *insertVector(Value *Col, unsigned I, Value *Block,
1238                       IRBuilder<> &Builder) {
1239 
1240     // First, bring Block to the same size as Col
1241     unsigned BlockNumElts =
1242         cast<FixedVectorType>(Block->getType())->getNumElements();
1243     unsigned NumElts = cast<FixedVectorType>(Col->getType())->getNumElements();
1244     assert(NumElts >= BlockNumElts && "Too few elements for current block");
1245 
1246     Block = Builder.CreateShuffleVector(
1247         Block, createSequentialMask(0, BlockNumElts, NumElts - BlockNumElts));
1248 
1249     // If Col is 7 long and I is 2 and BlockNumElts is 2 the mask is: 0, 1, 7,
1250     // 8, 4, 5, 6
1251     SmallVector<int, 16> Mask;
1252     unsigned i;
1253     for (i = 0; i < I; i++)
1254       Mask.push_back(i);
1255 
1256     unsigned VecNumElts =
1257         cast<FixedVectorType>(Col->getType())->getNumElements();
1258     for (; i < I + BlockNumElts; i++)
1259       Mask.push_back(i - I + VecNumElts);
1260 
1261     for (; i < VecNumElts; i++)
1262       Mask.push_back(i);
1263 
1264     return Builder.CreateShuffleVector(Col, Block, Mask);
1265   }
1266 
1267   Value *createMulAdd(Value *Sum, Value *A, Value *B, bool UseFPOp,
1268                       IRBuilder<> &Builder, bool AllowContraction,
1269                       unsigned &NumComputeOps) {
1270     NumComputeOps += getNumOps(A->getType());
1271     if (!Sum)
1272       return UseFPOp ? Builder.CreateFMul(A, B) : Builder.CreateMul(A, B);
1273 
1274     if (UseFPOp) {
1275       if (AllowContraction) {
1276         // Use fmuladd for floating point operations and let the backend decide
1277         // if that's profitable.
1278         Function *FMulAdd = Intrinsic::getDeclaration(
1279             Func.getParent(), Intrinsic::fmuladd, A->getType());
1280         return Builder.CreateCall(FMulAdd, {A, B, Sum});
1281       }
1282       NumComputeOps += getNumOps(A->getType());
1283       Value *Mul = Builder.CreateFMul(A, B);
1284       return Builder.CreateFAdd(Sum, Mul);
1285     }
1286 
1287     NumComputeOps += getNumOps(A->getType());
1288     Value *Mul = Builder.CreateMul(A, B);
1289     return Builder.CreateAdd(Sum, Mul);
1290   }
1291 
1292   /// Cache \p Matrix as result of \p Inst and update the uses of \p Inst. For
1293   /// users with shape information, there's nothing to do: they will use the
1294   /// cached value when they are lowered. For other users, \p Matrix is
1295   /// flattened and the uses are updated to use it. Also marks \p Inst for
1296   /// deletion.
1297   void finalizeLowering(Instruction *Inst, MatrixTy Matrix,
1298                         IRBuilder<> &Builder) {
1299     auto inserted = Inst2ColumnMatrix.insert(std::make_pair(Inst, Matrix));
1300     (void)inserted;
1301     assert(inserted.second && "multiple matrix lowering mapping");
1302 
1303     ToRemove.push_back(Inst);
1304     Value *Flattened = nullptr;
1305     for (Use &U : llvm::make_early_inc_range(Inst->uses())) {
1306       if (ShapeMap.find(U.getUser()) == ShapeMap.end()) {
1307         if (!Flattened)
1308           Flattened = Matrix.embedInVector(Builder);
1309         U.set(Flattened);
1310       }
1311     }
1312   }
1313 
1314   /// Compute \p Result += \p A * \p B for input matrices with left-associating
1315   /// addition.
1316   ///
1317   /// We can fold a transpose into the operand that is used to extract scalars.
1318   /// This is the first operands with row-major and the second with
1319   /// column-major.  If \p IsScalarMatrixTransposed we assume the appropriate
1320   /// operand is transposed.
1321   void emitMatrixMultiply(MatrixTy &Result, const MatrixTy &A,
1322                           const MatrixTy &B, IRBuilder<> &Builder, bool IsTiled,
1323                           bool IsScalarMatrixTransposed, FastMathFlags FMF) {
1324     const unsigned VF = std::max<unsigned>(
1325         TTI.getRegisterBitWidth(TargetTransformInfo::RGK_FixedWidthVector)
1326                 .getFixedValue() /
1327             Result.getElementType()->getPrimitiveSizeInBits().getFixedValue(),
1328         1U);
1329     unsigned R = Result.getNumRows();
1330     unsigned C = Result.getNumColumns();
1331     unsigned M = A.getNumColumns();
1332 
1333     bool IsFP = Result.getElementType()->isFloatingPointTy();
1334     assert(A.isColumnMajor() == B.isColumnMajor() &&
1335            Result.isColumnMajor() == A.isColumnMajor() &&
1336            "operands must agree on matrix layout");
1337     unsigned NumComputeOps = 0;
1338 
1339     Builder.setFastMathFlags(FMF);
1340 
1341     if (A.isColumnMajor()) {
1342       // Multiply columns from the first operand with scalars from the second
1343       // operand. Then move along the K axes and accumulate the columns.  With
1344       // this the adds can be vectorized without reassociation.
1345       for (unsigned J = 0; J < C; ++J) {
1346         unsigned BlockSize = VF;
1347         // If Result is zero, we don't need to accumulate in the K==0 iteration.
1348         bool isSumZero = isa<ConstantAggregateZero>(Result.getColumn(J));
1349 
1350         for (unsigned I = 0; I < R; I += BlockSize) {
1351           // Gradually lower the vectorization factor to cover the remainder.
1352           while (I + BlockSize > R)
1353             BlockSize /= 2;
1354 
1355           Value *Sum = IsTiled ? Result.extractVector(I, J, BlockSize, Builder)
1356                                : nullptr;
1357           for (unsigned K = 0; K < M; ++K) {
1358             Value *L = A.extractVector(I, K, BlockSize, Builder);
1359             Value *RH = Builder.CreateExtractElement(
1360                 B.getColumn(IsScalarMatrixTransposed ? K : J),
1361                 IsScalarMatrixTransposed ? J : K);
1362             Value *Splat = Builder.CreateVectorSplat(BlockSize, RH, "splat");
1363             Sum =
1364                 createMulAdd(isSumZero && K == 0 ? nullptr : Sum, L, Splat,
1365                              IsFP, Builder, FMF.allowContract(), NumComputeOps);
1366           }
1367           Result.setVector(J,
1368                            insertVector(Result.getVector(J), I, Sum, Builder));
1369         }
1370       }
1371     } else {
1372       // Multiply rows from the second operand with scalars from the first
1373       // operand. Then move along the K axes and accumulate the rows.  With this
1374       // the adds can be vectorized without reassociation.
1375       for (unsigned I = 0; I < R; ++I) {
1376         unsigned BlockSize = VF;
1377         bool isSumZero = isa<ConstantAggregateZero>(Result.getRow(I));
1378         for (unsigned J = 0; J < C; J += BlockSize) {
1379           // Gradually lower the vectorization factor to cover the remainder.
1380           while (J + BlockSize > C)
1381             BlockSize /= 2;
1382 
1383           Value *Sum = nullptr;
1384           for (unsigned K = 0; K < M; ++K) {
1385             Value *R = B.extractVector(K, J, BlockSize, Builder);
1386             Value *LH = Builder.CreateExtractElement(
1387                 A.getVector(IsScalarMatrixTransposed ? K : I),
1388                 IsScalarMatrixTransposed ? I : K);
1389             Value *Splat = Builder.CreateVectorSplat(BlockSize, LH, "splat");
1390             Sum =
1391                 createMulAdd(isSumZero && K == 0 ? nullptr : Sum, Splat, R,
1392                              IsFP, Builder, FMF.allowContract(), NumComputeOps);
1393           }
1394           Result.setVector(I,
1395                            insertVector(Result.getVector(I), J, Sum, Builder));
1396         }
1397       }
1398     }
1399     Result.addNumComputeOps(NumComputeOps);
1400   }
1401 
1402   /// Ensure that the memory in \p Load does not alias \p Store by potentially
1403   /// copying it to a new location.  This new or otherwise the original location
1404   /// is returned.
1405   Value *getNonAliasingPointer(LoadInst *Load, StoreInst *Store,
1406                                CallInst *MatMul) {
1407     MemoryLocation StoreLoc = MemoryLocation::get(Store);
1408     MemoryLocation LoadLoc = MemoryLocation::get(Load);
1409 
1410     // If we can statically determine noalias we're good.
1411     if (AA->isNoAlias(LoadLoc, StoreLoc))
1412       return Load->getPointerOperand();
1413 
1414     // Create code to check if the memory locations of the Load and Store
1415     // overlap and if they do, copy Load's operand to a new buffer.
1416 
1417     // First, create  new blocks for 2n part of the check and the copy.
1418     BasicBlock *Check0 = MatMul->getParent();
1419     // FIXME: Use lazy DTU and update SplitBlock to accept a DTU instead of a
1420     // DT. Manually collect dominator tree updates, to avoid unnecessary work,
1421     // as we adjust Check0 and Check1's branches.
1422     SmallVector<DominatorTree::UpdateType, 4> DTUpdates;
1423     for (BasicBlock *Succ : successors(Check0))
1424       DTUpdates.push_back({DT->Delete, Check0, Succ});
1425 
1426     BasicBlock *Check1 =
1427         SplitBlock(MatMul->getParent(), MatMul, (DomTreeUpdater *)nullptr, LI,
1428                    nullptr, "alias_cont");
1429     BasicBlock *Copy =
1430         SplitBlock(MatMul->getParent(), MatMul, (DomTreeUpdater *)nullptr, LI,
1431                    nullptr, "copy");
1432     BasicBlock *Fusion =
1433         SplitBlock(MatMul->getParent(), MatMul, (DomTreeUpdater *)nullptr, LI,
1434                    nullptr, "no_alias");
1435 
1436     // Check if the loaded memory location begins before the end of the store
1437     // location. If the condition holds, they might overlap, otherwise they are
1438     // guaranteed to not overlap.
1439     IRBuilder<> Builder(MatMul);
1440     Check0->getTerminator()->eraseFromParent();
1441     Builder.SetInsertPoint(Check0);
1442     Type *IntPtrTy = Builder.getIntPtrTy(Load->getModule()->getDataLayout());
1443     Value *StoreBegin = Builder.CreatePtrToInt(
1444         const_cast<Value *>(StoreLoc.Ptr), IntPtrTy, "store.begin");
1445     Value *StoreEnd = Builder.CreateAdd(
1446         StoreBegin, ConstantInt::get(IntPtrTy, StoreLoc.Size.getValue()),
1447         "store.end", true, true);
1448     Value *LoadBegin = Builder.CreatePtrToInt(const_cast<Value *>(LoadLoc.Ptr),
1449                                               IntPtrTy, "load.begin");
1450     Builder.CreateCondBr(Builder.CreateICmpULT(LoadBegin, StoreEnd), Check1,
1451                          Fusion);
1452 
1453     // Check if the store begins before the end of the load location. If the
1454     // condition holds, they alias, otherwise they are guaranteed to not
1455     // overlap.
1456     Check1->getTerminator()->eraseFromParent();
1457     Builder.SetInsertPoint(Check1, Check1->begin());
1458     Value *LoadEnd = Builder.CreateAdd(
1459         LoadBegin, ConstantInt::get(IntPtrTy, LoadLoc.Size.getValue()),
1460         "load.end", true, true);
1461     Builder.CreateCondBr(Builder.CreateICmpULT(StoreBegin, LoadEnd), Copy,
1462                          Fusion);
1463 
1464     // Copy load operand to new alloca.
1465     Builder.SetInsertPoint(Copy, Copy->begin());
1466     auto *VT = cast<FixedVectorType>(Load->getType());
1467     // Use an array type for the alloca, to avoid potentially huge alignment
1468     // requirements for large vector types.
1469     auto *ArrayTy = ArrayType::get(VT->getElementType(), VT->getNumElements());
1470     AllocaInst *Alloca =
1471         Builder.CreateAlloca(ArrayTy, Load->getPointerAddressSpace());
1472     Value *BC = Builder.CreateBitCast(Alloca, VT->getPointerTo());
1473 
1474     Builder.CreateMemCpy(BC, Alloca->getAlign(), Load->getPointerOperand(),
1475                          Load->getAlign(), LoadLoc.Size.getValue());
1476     Builder.SetInsertPoint(Fusion, Fusion->begin());
1477     PHINode *PHI = Builder.CreatePHI(Load->getPointerOperandType(), 3);
1478     PHI->addIncoming(Load->getPointerOperand(), Check0);
1479     PHI->addIncoming(Load->getPointerOperand(), Check1);
1480     PHI->addIncoming(BC, Copy);
1481 
1482     // Adjust DT.
1483     DTUpdates.push_back({DT->Insert, Check0, Check1});
1484     DTUpdates.push_back({DT->Insert, Check0, Fusion});
1485     DTUpdates.push_back({DT->Insert, Check1, Copy});
1486     DTUpdates.push_back({DT->Insert, Check1, Fusion});
1487     DT->applyUpdates(DTUpdates);
1488     return PHI;
1489   }
1490 
1491   bool isFusionProfitable(CallInst *MatMul) {
1492     if (ForceFusion)
1493       return true;
1494 
1495     ShapeInfo LShape(MatMul->getArgOperand(2), MatMul->getArgOperand(3));
1496     ShapeInfo RShape(MatMul->getArgOperand(3), MatMul->getArgOperand(4));
1497 
1498     const unsigned R = LShape.NumRows;
1499     const unsigned C = RShape.NumColumns;
1500     const unsigned M = LShape.NumColumns;
1501     auto *EltType = cast<VectorType>(MatMul->getType())->getElementType();
1502 
1503     const unsigned VF = std::max<unsigned>(
1504         TTI.getRegisterBitWidth(TargetTransformInfo::RGK_FixedWidthVector)
1505                 .getFixedValue() /
1506             EltType->getPrimitiveSizeInBits().getFixedValue(),
1507         1U);
1508 
1509     // Cost model for tiling
1510     //
1511     // For tiling to be beneficial, we need reuse either along the R or
1512     // the C axis.  We vectorize along the R axis so that means at least
1513     // 3 elements.
1514     // TODO: Also consider cost of copying if operands alias.
1515     if (R <= VF && C == 1)
1516       return false;
1517     // Then we need enough elements to exceed the number of vector
1518     // registers we have.  Note that this is an oversimplification since
1519     // fusing also takes some extra loads which may exceed the number of
1520     // reloads necessary.
1521     unsigned Op0Regs = (R + VF - 1) / VF * M;
1522     unsigned Op1Regs = (M + VF - 1) / VF * C;
1523     return Op0Regs + Op1Regs >
1524            TTI.getNumberOfRegisters(TTI.getRegisterClassForType(true));
1525   }
1526 
1527   MatrixTy getZeroMatrix(Type *EltType, unsigned R, unsigned C) {
1528     MatrixTy Res;
1529     auto *ColumType = FixedVectorType::get(EltType, R);
1530     for (unsigned I = 0; I < C; ++I)
1531       Res.addVector(ConstantAggregateZero::get(ColumType));
1532     return Res;
1533   }
1534 
1535   void createTiledLoops(CallInst *MatMul, Value *LPtr, ShapeInfo LShape,
1536                         Value *RPtr, ShapeInfo RShape, StoreInst *Store) {
1537     auto *EltType = cast<VectorType>(MatMul->getType())->getElementType();
1538 
1539     // Create the main tiling loop nest.
1540     TileInfo TI(LShape.NumRows, RShape.NumColumns, LShape.NumColumns, TileSize);
1541     DomTreeUpdater DTU(DT, DomTreeUpdater::UpdateStrategy::Lazy);
1542     Instruction *InsertI = cast<Instruction>(MatMul);
1543     BasicBlock *Start = InsertI->getParent();
1544     BasicBlock *End =
1545         SplitBlock(InsertI->getParent(), InsertI, DT, LI, nullptr, "continue");
1546     IRBuilder<> Builder(MatMul);
1547     BasicBlock *InnerBody = TI.CreateTiledLoops(Start, End, Builder, DTU, *LI);
1548 
1549     Type *TileVecTy =
1550         FixedVectorType::get(MatMul->getType()->getScalarType(), TileSize);
1551     MatrixTy TileResult;
1552     // Insert in the inner loop header.
1553     Builder.SetInsertPoint(TI.KLoop.Header->getTerminator());
1554     // Create PHI nodes for the result columns to accumulate across iterations.
1555     SmallVector<PHINode *, 4> ColumnPhis;
1556     for (unsigned I = 0; I < TileSize; I++) {
1557       auto *Phi = Builder.CreatePHI(TileVecTy, 2, "result.vec." + Twine(I));
1558       Phi->addIncoming(ConstantAggregateZero::get(TileVecTy),
1559                        TI.RowLoop.Header->getSingleSuccessor());
1560       TileResult.addVector(Phi);
1561       ColumnPhis.push_back(Phi);
1562     }
1563 
1564     // Insert in the inner loop body, which computes
1565     //   Res += Load(CurrentRow, K) * Load(K, CurrentColumn)
1566     Builder.SetInsertPoint(InnerBody->getTerminator());
1567     // Load tiles of the operands.
1568     MatrixTy A =
1569         loadMatrix(LPtr, {}, false, LShape, TI.RowLoop.Index, TI.KLoop.Index,
1570                    {TileSize, TileSize}, EltType, Builder);
1571     MatrixTy B =
1572         loadMatrix(RPtr, {}, false, RShape, TI.KLoop.Index, TI.ColumnLoop.Index,
1573                    {TileSize, TileSize}, EltType, Builder);
1574     emitMatrixMultiply(TileResult, A, B, Builder, true, false,
1575                        getFastMathFlags(MatMul));
1576     // Store result after the inner loop is done.
1577     Builder.SetInsertPoint(TI.RowLoop.Latch->getTerminator());
1578     storeMatrix(TileResult, Store->getPointerOperand(), Store->getAlign(),
1579                 Store->isVolatile(), {LShape.NumRows, RShape.NumColumns},
1580                 TI.RowLoop.Index, TI.ColumnLoop.Index, EltType, Builder);
1581 
1582     for (unsigned I = 0; I < TileResult.getNumVectors(); I++)
1583       ColumnPhis[I]->addIncoming(TileResult.getVector(I), TI.KLoop.Latch);
1584 
1585     // Force unrolling of a few iterations of the inner loop, to make sure there
1586     // is enough work per iteration.
1587     // FIXME: The unroller should make this decision directly instead, but
1588     // currently the cost-model is not up to the task.
1589     unsigned InnerLoopUnrollCount = std::min(10u, LShape.NumColumns / TileSize);
1590     addStringMetadataToLoop(LI->getLoopFor(TI.KLoop.Header),
1591                             "llvm.loop.unroll.count", InnerLoopUnrollCount);
1592   }
1593 
1594   void emitSIMDTiling(CallInst *MatMul, LoadInst *LoadOp0, LoadInst *LoadOp1,
1595                       StoreInst *Store,
1596                       SmallPtrSetImpl<Instruction *> &FusedInsts) {
1597     assert(MatrixLayout == MatrixLayoutTy::ColumnMajor &&
1598            "Tiling only supported for column-major matrixes at the moment!");
1599     if (!isFusionProfitable(MatMul))
1600       return;
1601 
1602     ShapeInfo LShape(MatMul->getArgOperand(2), MatMul->getArgOperand(3));
1603     ShapeInfo RShape(MatMul->getArgOperand(3), MatMul->getArgOperand(4));
1604 
1605     const unsigned R = LShape.NumRows;
1606     const unsigned C = RShape.NumColumns;
1607     const unsigned M = LShape.NumColumns;
1608     auto *EltType = cast<VectorType>(MatMul->getType())->getElementType();
1609 
1610     Value *APtr = getNonAliasingPointer(LoadOp0, Store, MatMul);
1611     Value *BPtr = getNonAliasingPointer(LoadOp1, Store, MatMul);
1612     Value *CPtr = Store->getPointerOperand();
1613 
1614     if (TileUseLoops && (R % TileSize == 0 && C % TileSize == 0))
1615       createTiledLoops(MatMul, APtr, LShape, BPtr, RShape, Store);
1616     else {
1617       IRBuilder<> Builder(Store);
1618       for (unsigned J = 0; J < C; J += TileSize)
1619         for (unsigned I = 0; I < R; I += TileSize) {
1620           const unsigned TileR = std::min(R - I, unsigned(TileSize));
1621           const unsigned TileC = std::min(C - J, unsigned(TileSize));
1622           MatrixTy Res = getZeroMatrix(EltType, TileR, TileC);
1623 
1624           for (unsigned K = 0; K < M; K += TileSize) {
1625             const unsigned TileM = std::min(M - K, unsigned(TileSize));
1626             MatrixTy A =
1627                 loadMatrix(APtr, LoadOp0->getAlign(), LoadOp0->isVolatile(),
1628                            LShape, Builder.getInt64(I), Builder.getInt64(K),
1629                            {TileR, TileM}, EltType, Builder);
1630             MatrixTy B =
1631                 loadMatrix(BPtr, LoadOp1->getAlign(), LoadOp1->isVolatile(),
1632                            RShape, Builder.getInt64(K), Builder.getInt64(J),
1633                            {TileM, TileC}, EltType, Builder);
1634             emitMatrixMultiply(Res, A, B, Builder, true, false,
1635                                getFastMathFlags(MatMul));
1636           }
1637           storeMatrix(Res, CPtr, Store->getAlign(), Store->isVolatile(), {R, M},
1638                       Builder.getInt64(I), Builder.getInt64(J), EltType,
1639                       Builder);
1640         }
1641     }
1642 
1643     // Mark eliminated instructions as fused and remove them.
1644     FusedInsts.insert(Store);
1645     FusedInsts.insert(MatMul);
1646     Store->eraseFromParent();
1647     MatMul->eraseFromParent();
1648     if (LoadOp0->hasNUses(0)) {
1649       FusedInsts.insert(LoadOp0);
1650       LoadOp0->eraseFromParent();
1651     }
1652     if (LoadOp1 != LoadOp0 && LoadOp1->hasNUses(0)) {
1653       FusedInsts.insert(LoadOp1);
1654       LoadOp1->eraseFromParent();
1655     }
1656   }
1657 
1658   /// Try to lower matrix multiply chains by fusing operations.
1659   ///
1660   /// Call finalizeLowering on lowered instructions.  Instructions that are
1661   /// completely eliminated by fusion are added to \p FusedInsts.
1662   void LowerMatrixMultiplyFused(CallInst *MatMul,
1663                                 SmallPtrSetImpl<Instruction *> &FusedInsts) {
1664     if (!FuseMatrix || !DT)
1665       return;
1666 
1667     assert(AA && LI && "Analyses should be available");
1668 
1669     Value *A = MatMul->getArgOperand(0);
1670     Value *B = MatMul->getArgOperand(1);
1671 
1672     // We can fold the transpose into the operand that is used to fetch scalars.
1673     Value *T;
1674     if (MatrixLayout == MatrixLayoutTy::ColumnMajor
1675             ? match(B, m_Intrinsic<Intrinsic::matrix_transpose>(m_Value(T)))
1676             : match(A, m_Intrinsic<Intrinsic::matrix_transpose>(m_Value(T)))) {
1677       IRBuilder<> Builder(MatMul);
1678       auto *EltType = cast<VectorType>(MatMul->getType())->getElementType();
1679       ShapeInfo LShape(MatMul->getArgOperand(2), MatMul->getArgOperand(3));
1680       ShapeInfo RShape(MatMul->getArgOperand(3), MatMul->getArgOperand(4));
1681       const unsigned R = LShape.NumRows;
1682       const unsigned M = LShape.NumColumns;
1683       const unsigned C = RShape.NumColumns;
1684 
1685       MatrixTy MA;
1686       MatrixTy MB;
1687 
1688       Value *Transpose;
1689       if (MatrixLayout == MatrixLayoutTy::ColumnMajor) {
1690         MA = getMatrix(A, ShapeInfo(R, M), Builder);
1691         MB = getMatrix(T, ShapeInfo(C, M), Builder);
1692         Transpose = B;
1693       } else {
1694         MA = getMatrix(T, ShapeInfo(R, M), Builder);
1695         MB = getMatrix(B, ShapeInfo(C, M), Builder);
1696         Transpose = A;
1697       }
1698 
1699       // Initialize the output
1700       MatrixTy Result(R, C, EltType);
1701 
1702       emitMatrixMultiply(Result, MA, MB, Builder, false, true,
1703                          getFastMathFlags(MatMul));
1704 
1705       FusedInsts.insert(MatMul);
1706       if (Transpose->hasOneUse()) {
1707         FusedInsts.insert(cast<Instruction>(Transpose));
1708         ToRemove.push_back(cast<Instruction>(Transpose));
1709         // TODO: add a fake entry for the folded instruction so that this is
1710         // included in the expression in the remark.
1711         Inst2ColumnMatrix[Transpose] = MatrixTy(M, C, EltType);
1712       }
1713       finalizeLowering(MatMul, Result, Builder);
1714       return;
1715     }
1716 
1717     if (!MatMul->hasOneUse() || MatrixLayout != MatrixLayoutTy::ColumnMajor)
1718       return;
1719 
1720     // Lower {ld, ld} -> matmul -> st chains.  No need to call finalizeLowering
1721     // since the single store user will be lowered as part of this.
1722     auto *LoadOp0 = dyn_cast<LoadInst>(A);
1723     auto *LoadOp1 = dyn_cast<LoadInst>(B);
1724     auto *Store = dyn_cast<StoreInst>(*MatMul->user_begin());
1725     if (LoadOp0 && LoadOp1 && Store) {
1726       // The store address must dominate the MatMul instruction, otherwise
1727       // we create invalid IR.
1728       SetVector<Value *> WorkList;
1729       WorkList.insert(Store->getOperand(1));
1730       SmallVector<Instruction *> ToHoist;
1731       for (unsigned I = 0; I != WorkList.size(); ++I) {
1732         Value *Current = WorkList[I];
1733         auto *CurrI = dyn_cast<Instruction>(Current);
1734         if (!CurrI)
1735           continue;
1736         if (isa<PHINode>(CurrI))
1737           return;
1738         if (DT->dominates(CurrI, MatMul))
1739           continue;
1740         if (CurrI->mayHaveSideEffects() || CurrI->mayReadFromMemory())
1741           return;
1742         ToHoist.push_back(CurrI);
1743         WorkList.insert(CurrI->op_begin(), CurrI->op_end());
1744       }
1745 
1746       sort(ToHoist, [this](Instruction *A, Instruction *B) {
1747         return DT->dominates(A, B);
1748       });
1749       for (Instruction *I : ToHoist)
1750         I->moveBefore(MatMul);
1751 
1752       emitSIMDTiling(MatMul, LoadOp0, LoadOp1, Store, FusedInsts);
1753       return;
1754     }
1755   }
1756 
1757   /// Lowers llvm.matrix.multiply.
1758   void LowerMultiply(CallInst *MatMul) {
1759     IRBuilder<> Builder(MatMul);
1760     auto *EltType = cast<VectorType>(MatMul->getType())->getElementType();
1761     ShapeInfo LShape(MatMul->getArgOperand(2), MatMul->getArgOperand(3));
1762     ShapeInfo RShape(MatMul->getArgOperand(3), MatMul->getArgOperand(4));
1763 
1764     const MatrixTy &Lhs = getMatrix(MatMul->getArgOperand(0), LShape, Builder);
1765     const MatrixTy &Rhs = getMatrix(MatMul->getArgOperand(1), RShape, Builder);
1766     assert(Lhs.getElementType() == Rhs.getElementType() &&
1767            "Matrix multiply argument element types do not match.");
1768 
1769     const unsigned R = LShape.NumRows;
1770     const unsigned C = RShape.NumColumns;
1771     assert(LShape.NumColumns == RShape.NumRows);
1772 
1773     // Initialize the output
1774     MatrixTy Result(R, C, EltType);
1775     assert(Lhs.getElementType() == Result.getElementType() &&
1776            "Matrix multiply result element type does not match arguments.");
1777 
1778     emitMatrixMultiply(Result, Lhs, Rhs, Builder, false, false,
1779                        getFastMathFlags(MatMul));
1780     finalizeLowering(MatMul, Result, Builder);
1781   }
1782 
1783   /// Lowers llvm.matrix.transpose.
1784   void LowerTranspose(CallInst *Inst) {
1785     MatrixTy Result;
1786     IRBuilder<> Builder(Inst);
1787     Value *InputVal = Inst->getArgOperand(0);
1788     VectorType *VectorTy = cast<VectorType>(InputVal->getType());
1789     ShapeInfo ArgShape(Inst->getArgOperand(1), Inst->getArgOperand(2));
1790     MatrixTy InputMatrix = getMatrix(InputVal, ArgShape, Builder);
1791 
1792     const unsigned NewNumVecs =
1793         InputMatrix.isColumnMajor() ? ArgShape.NumRows : ArgShape.NumColumns;
1794     const unsigned NewNumElts =
1795         InputMatrix.isColumnMajor() ? ArgShape.NumColumns : ArgShape.NumRows;
1796 
1797     for (unsigned I = 0; I < NewNumVecs; ++I) {
1798       // Build a single result vector. First initialize it.
1799       Value *ResultVector = PoisonValue::get(
1800           FixedVectorType::get(VectorTy->getElementType(), NewNumElts));
1801       // Go through the old elements and insert it into the resulting vector.
1802       for (auto J : enumerate(InputMatrix.vectors())) {
1803         Value *Elt = Builder.CreateExtractElement(J.value(), I);
1804         // Row and column indices are transposed.
1805         ResultVector =
1806             Builder.CreateInsertElement(ResultVector, Elt, J.index());
1807       }
1808       Result.addVector(ResultVector);
1809     }
1810 
1811     // TODO: Improve estimate of operations needed for transposes. Currently we
1812     // just count the insertelement/extractelement instructions, but do not
1813     // account for later simplifications/combines.
1814     finalizeLowering(
1815         Inst,
1816         Result.addNumComputeOps(2 * ArgShape.NumRows * ArgShape.NumColumns)
1817             .addNumExposedTransposes(1),
1818         Builder);
1819   }
1820 
1821   /// Lower load instructions, if shape information is available.
1822   bool VisitLoad(LoadInst *Inst, Value *Ptr, IRBuilder<> &Builder) {
1823     auto I = ShapeMap.find(Inst);
1824     if (I == ShapeMap.end())
1825       return false;
1826 
1827     LowerLoad(Inst, Ptr, Inst->getAlign(),
1828               Builder.getInt64(I->second.getStride()), Inst->isVolatile(),
1829               I->second);
1830     return true;
1831   }
1832 
1833   bool VisitStore(StoreInst *Inst, Value *StoredVal, Value *Ptr,
1834                   IRBuilder<> &Builder) {
1835     auto I = ShapeMap.find(StoredVal);
1836     if (I == ShapeMap.end())
1837       return false;
1838 
1839     LowerStore(Inst, StoredVal, Ptr, Inst->getAlign(),
1840                Builder.getInt64(I->second.getStride()), Inst->isVolatile(),
1841                I->second);
1842     return true;
1843   }
1844 
1845   /// Lower binary operators, if shape information is available.
1846   bool VisitBinaryOperator(BinaryOperator *Inst) {
1847     auto I = ShapeMap.find(Inst);
1848     if (I == ShapeMap.end())
1849       return false;
1850 
1851     Value *Lhs = Inst->getOperand(0);
1852     Value *Rhs = Inst->getOperand(1);
1853 
1854     IRBuilder<> Builder(Inst);
1855     ShapeInfo &Shape = I->second;
1856 
1857     MatrixTy Result;
1858     MatrixTy A = getMatrix(Lhs, Shape, Builder);
1859     MatrixTy B = getMatrix(Rhs, Shape, Builder);
1860     assert(A.isColumnMajor() == B.isColumnMajor() &&
1861            Result.isColumnMajor() == A.isColumnMajor() &&
1862            "operands must agree on matrix layout");
1863 
1864     Builder.setFastMathFlags(getFastMathFlags(Inst));
1865 
1866     // Helper to perform binary op on vectors.
1867     auto BuildVectorOp = [&Builder, Inst](Value *LHS, Value *RHS) {
1868       switch (Inst->getOpcode()) {
1869       case Instruction::Add:
1870         return Builder.CreateAdd(LHS, RHS);
1871       case Instruction::Mul:
1872         return Builder.CreateMul(LHS, RHS);
1873       case Instruction::Sub:
1874         return Builder.CreateSub(LHS, RHS);
1875       case Instruction::FAdd:
1876         return Builder.CreateFAdd(LHS, RHS);
1877       case Instruction::FMul:
1878         return Builder.CreateFMul(LHS, RHS);
1879       case Instruction::FSub:
1880         return Builder.CreateFSub(LHS, RHS);
1881       default:
1882         llvm_unreachable("Unsupported binary operator for matrix");
1883       }
1884     };
1885 
1886     for (unsigned I = 0; I < Shape.getNumVectors(); ++I)
1887       Result.addVector(BuildVectorOp(A.getVector(I), B.getVector(I)));
1888 
1889     finalizeLowering(Inst,
1890                      Result.addNumComputeOps(getNumOps(Result.getVectorTy()) *
1891                                              Result.getNumVectors()),
1892                      Builder);
1893     return true;
1894   }
1895 
1896   /// Lower unary operators, if shape information is available.
1897   bool VisitUnaryOperator(UnaryOperator *Inst) {
1898     auto I = ShapeMap.find(Inst);
1899     if (I == ShapeMap.end())
1900       return false;
1901 
1902     Value *Op = Inst->getOperand(0);
1903 
1904     IRBuilder<> Builder(Inst);
1905     ShapeInfo &Shape = I->second;
1906 
1907     MatrixTy Result;
1908     MatrixTy M = getMatrix(Op, Shape, Builder);
1909 
1910     Builder.setFastMathFlags(getFastMathFlags(Inst));
1911 
1912     // Helper to perform unary op on vectors.
1913     auto BuildVectorOp = [&Builder, Inst](Value *Op) {
1914       switch (Inst->getOpcode()) {
1915       case Instruction::FNeg:
1916         return Builder.CreateFNeg(Op);
1917       default:
1918         llvm_unreachable("Unsupported unary operator for matrix");
1919       }
1920     };
1921 
1922     for (unsigned I = 0; I < Shape.getNumVectors(); ++I)
1923       Result.addVector(BuildVectorOp(M.getVector(I)));
1924 
1925     finalizeLowering(Inst,
1926                      Result.addNumComputeOps(getNumOps(Result.getVectorTy()) *
1927                                              Result.getNumVectors()),
1928                      Builder);
1929     return true;
1930   }
1931 
1932   /// Helper to linearize a matrix expression tree into a string. Currently
1933   /// matrix expressions are linarized by starting at an expression leaf and
1934   /// linearizing bottom up.
1935   struct ExprLinearizer {
1936     unsigned LengthToBreak = 100;
1937     std::string Str;
1938     raw_string_ostream Stream;
1939     unsigned LineLength = 0;
1940     const DataLayout &DL;
1941 
1942     /// Mapping from instructions to matrixes. It is used to identify
1943     /// matrix instructions.
1944     const MapVector<Value *, MatrixTy> &Inst2Matrix;
1945 
1946     /// Mapping from values to the leaves of all expressions that the value is
1947     /// part of.
1948     const DenseMap<Value *, SmallPtrSet<Value *, 2>> &Shared;
1949 
1950     /// Set of matrix expressions in the scope of a given DISubprogram.
1951     const SmallSetVector<Value *, 32> &ExprsInSubprogram;
1952 
1953     /// Leaf node of the expression to linearize.
1954     Value *Leaf;
1955 
1956     /// Used to keep track of sub-expressions that get reused while linearizing
1957     /// the expression. Re-used sub-expressions are marked as (reused).
1958     SmallPtrSet<Value *, 8> ReusedExprs;
1959 
1960     ExprLinearizer(const DataLayout &DL,
1961                    const MapVector<Value *, MatrixTy> &Inst2Matrix,
1962                    const DenseMap<Value *, SmallPtrSet<Value *, 2>> &Shared,
1963                    const SmallSetVector<Value *, 32> &ExprsInSubprogram,
1964                    Value *Leaf)
1965         : Stream(Str), DL(DL), Inst2Matrix(Inst2Matrix), Shared(Shared),
1966           ExprsInSubprogram(ExprsInSubprogram), Leaf(Leaf) {}
1967 
1968     void indent(unsigned N) {
1969       LineLength += N;
1970       for (unsigned i = 0; i < N; i++)
1971         Stream << " ";
1972     }
1973 
1974     void lineBreak() {
1975       Stream << "\n";
1976       LineLength = 0;
1977     }
1978 
1979     void maybeIndent(unsigned Indent) {
1980       if (LineLength >= LengthToBreak)
1981         lineBreak();
1982 
1983       if (LineLength == 0)
1984         indent(Indent);
1985     }
1986 
1987     void write(StringRef S) {
1988       LineLength += S.size();
1989       Stream << S;
1990     }
1991 
1992     Value *getUnderlyingObjectThroughLoads(Value *V) {
1993       if (Value *Ptr = getPointerOperand(V))
1994         return getUnderlyingObjectThroughLoads(Ptr);
1995       else if (V->getType()->isPointerTy())
1996         return getUnderlyingObject(V);
1997       return V;
1998     }
1999 
2000     /// Returns true if \p V is a matrix value in the given subprogram.
2001     bool isMatrix(Value *V) const { return ExprsInSubprogram.count(V); }
2002 
2003     /// If \p V is a matrix value, print its shape as as NumRows x NumColumns to
2004     /// \p SS.
2005     void prettyPrintMatrixType(Value *V, raw_string_ostream &SS) {
2006       auto M = Inst2Matrix.find(V);
2007       if (M == Inst2Matrix.end())
2008         SS << "unknown";
2009       else {
2010         SS << M->second.getNumRows();
2011         SS << "x";
2012         SS << M->second.getNumColumns();
2013       }
2014     }
2015 
2016     /// Write the called function name. Handles calls to llvm.matrix.*
2017     /// specially: we write the name, followed by the dimensions of the input
2018     /// matrixes, followed by the scalar type name.
2019     void writeFnName(CallInst *CI) {
2020       if (!CI->getCalledFunction())
2021         write("<no called fn>");
2022       else {
2023         StringRef Name = CI->getCalledFunction()->getName();
2024         if (!Name.startswith("llvm.matrix")) {
2025           write(Name);
2026           return;
2027         }
2028         auto *II = cast<IntrinsicInst>(CI);
2029         write(Intrinsic::getBaseName(II->getIntrinsicID())
2030                   .drop_front(StringRef("llvm.matrix.").size()));
2031         write(".");
2032         std::string Tmp;
2033         raw_string_ostream SS(Tmp);
2034 
2035         switch (II->getIntrinsicID()) {
2036         case Intrinsic::matrix_multiply:
2037           prettyPrintMatrixType(II->getOperand(0), SS);
2038           SS << ".";
2039           prettyPrintMatrixType(II->getOperand(1), SS);
2040           SS << "." << *II->getType()->getScalarType();
2041           break;
2042         case Intrinsic::matrix_transpose:
2043           prettyPrintMatrixType(II->getOperand(0), SS);
2044           SS << "." << *II->getType()->getScalarType();
2045           break;
2046         case Intrinsic::matrix_column_major_load:
2047           prettyPrintMatrixType(II, SS);
2048           SS << "." << *II->getType()->getScalarType();
2049           break;
2050         case Intrinsic::matrix_column_major_store:
2051           prettyPrintMatrixType(II->getOperand(0), SS);
2052           SS << "." << *II->getOperand(0)->getType()->getScalarType();
2053           break;
2054         default:
2055           llvm_unreachable("Unhandled case");
2056         }
2057         SS.flush();
2058         write(Tmp);
2059       }
2060     }
2061 
2062     unsigned getNumShapeArgs(CallInst *CI) const {
2063       if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(CI)) {
2064         switch (II->getIntrinsicID()) {
2065         case Intrinsic::matrix_multiply:
2066           return 3;
2067         case Intrinsic::matrix_transpose:
2068           return 2;
2069         case Intrinsic::matrix_column_major_load:
2070         case Intrinsic::matrix_column_major_store:
2071           return 3;
2072         default:
2073           return 0;
2074         }
2075       }
2076       return 0;
2077     }
2078 
2079     /// Special printing for values: for pointers, we print if they refer to an
2080     /// (function) external address or a stack address, for other values we
2081     /// either print the constant or "scalar"/"matrix" for other values.
2082     void write(Value *V) {
2083       V = getUnderlyingObjectThroughLoads(V);
2084       if (V->getType()->isPointerTy()) {
2085         if (isa<AllocaInst>(V)) {
2086           Stream << "stack addr";
2087           LineLength += StringRef("stack addr").size();
2088         } else {
2089           Stream << "addr";
2090           LineLength += StringRef("addr").size();
2091         }
2092         if (!V->getName().empty()) {
2093           Stream << " %" << V->getName() << "";
2094           LineLength += V->getName().size() + 2;
2095         }
2096         return;
2097       }
2098 
2099       std::string Tmp;
2100       raw_string_ostream TmpStream(Tmp);
2101 
2102       if (auto *CI = dyn_cast<ConstantInt>(V))
2103         TmpStream << CI->getValue();
2104       else if (isa<Constant>(V))
2105         TmpStream << "constant";
2106       else {
2107         if (isMatrix(V))
2108           TmpStream << "matrix";
2109         else
2110           TmpStream << "scalar";
2111       }
2112       TmpStream.flush();
2113       Tmp = std::string(StringRef(Tmp).trim());
2114       LineLength += Tmp.size();
2115       Stream << Tmp;
2116     }
2117 
2118     /// Linearize expression \p Expr starting at an indentation of \p Indent.
2119     /// Expressions that are re-used multiple times are prefixed with (reused)
2120     /// at the re-used root instruction.
2121     void linearizeExpr(Value *Expr, unsigned Indent, bool ParentReused,
2122                        bool ParentShared) {
2123       auto *I = cast<Instruction>(Expr);
2124       maybeIndent(Indent);
2125       SmallVector<Value *, 8> Ops;
2126 
2127       // Is Expr shared with other expression leaves?
2128       bool ExprShared = false;
2129 
2130       // Deal with shared subtrees. Mark them as shared, if required.
2131       if (!ParentShared) {
2132         auto SI = Shared.find(Expr);
2133         assert(SI != Shared.end() && SI->second.count(Leaf));
2134 
2135         for (Value *S : SI->second) {
2136           if (S == Leaf)
2137             continue;
2138           DebugLoc DL = cast<Instruction>(S)->getDebugLoc();
2139           write("shared with remark at line " + std::to_string(DL.getLine()) +
2140                 " column " + std::to_string(DL.getCol()) + " (");
2141         }
2142         ExprShared = SI->second.size() > 1;
2143       }
2144 
2145       bool Reused = !ReusedExprs.insert(Expr).second;
2146       if (Reused && !ParentReused)
2147         write("(reused) ");
2148 
2149       if (auto *CI = dyn_cast<CallInst>(I)) {
2150         writeFnName(CI);
2151 
2152         Ops.append(CI->arg_begin(), CI->arg_end() - getNumShapeArgs(CI));
2153       } else if (isa<BitCastInst>(Expr)) {
2154         // Special case bitcasts, which are used to materialize matrixes from
2155         // non-matrix ops.
2156         write("matrix");
2157         return;
2158       } else {
2159         Ops.append(I->value_op_begin(), I->value_op_end());
2160         write(std::string(I->getOpcodeName()));
2161       }
2162 
2163       write(std::string("("));
2164 
2165       unsigned NumOpsToBreak = 1;
2166       if (match(Expr, m_Intrinsic<Intrinsic::matrix_column_major_load>()))
2167         NumOpsToBreak = 2;
2168 
2169       for (Value *Op : Ops) {
2170         if (Ops.size() > NumOpsToBreak)
2171           lineBreak();
2172 
2173         maybeIndent(Indent + 1);
2174         if (isMatrix(Op))
2175           linearizeExpr(Op, Indent + 1, Reused, ExprShared);
2176         else
2177           write(Op);
2178         if (Op != Ops.back())
2179           write(", ");
2180       }
2181 
2182       write(")");
2183     }
2184 
2185     const std::string &getResult() {
2186       Stream.flush();
2187       return Str;
2188     }
2189   };
2190 
2191   /// Generate remarks for matrix operations in a function. To generate remarks
2192   /// for matrix expressions, the following approach is used:
2193   /// 1. Use the inlined-at debug information to group matrix operations to the
2194   ///    DISubprograms they are contained in.
2195   /// 2. Collect leaves of matrix expressions (done in
2196   ///    RemarkGenerator::getExpressionLeaves) for each subprogram - expression
2197   //     mapping.  Leaves are lowered matrix instructions without other matrix
2198   //     users (like stores) in the current subprogram.
2199   /// 3. For each leaf, create a remark containing a linearizied version of the
2200   ///    matrix expression. The expression is linearized by a recursive
2201   ///    bottom-up traversal of the matrix operands, starting at a leaf. Note
2202   ///    that multiple leaves can share sub-expressions. Shared subexpressions
2203   ///    are explicitly marked as shared().
2204   struct RemarkGenerator {
2205     const MapVector<Value *, MatrixTy> &Inst2Matrix;
2206     OptimizationRemarkEmitter &ORE;
2207     Function &Func;
2208     const DataLayout &DL;
2209 
2210     RemarkGenerator(const MapVector<Value *, MatrixTy> &Inst2Matrix,
2211                     OptimizationRemarkEmitter &ORE, Function &Func)
2212         : Inst2Matrix(Inst2Matrix), ORE(ORE), Func(Func),
2213           DL(Func.getParent()->getDataLayout()) {}
2214 
2215     /// Return all leaves of the expressions in \p ExprsInSubprogram. Those are
2216     /// instructions in Inst2Matrix returning void or without any users in
2217     /// \p ExprsInSubprogram. Currently that should only include stores.
2218     SmallVector<Value *, 4>
2219     getExpressionLeaves(const SmallSetVector<Value *, 32> &ExprsInSubprogram) {
2220       SmallVector<Value *, 4> Leaves;
2221       for (auto *Expr : ExprsInSubprogram)
2222         if (Expr->getType()->isVoidTy() ||
2223             !any_of(Expr->users(), [&ExprsInSubprogram](User *U) {
2224               return ExprsInSubprogram.count(U);
2225             }))
2226           Leaves.push_back(Expr);
2227       return Leaves;
2228     }
2229 
2230     /// Recursively traverse expression \p V starting at \p Leaf and add \p Leaf
2231     /// to all visited expressions in \p Shared. Limit the matrix operations to
2232     /// the ones in \p ExprsInSubprogram.
2233     void collectSharedInfo(Value *Leaf, Value *V,
2234                            const SmallSetVector<Value *, 32> &ExprsInSubprogram,
2235                            DenseMap<Value *, SmallPtrSet<Value *, 2>> &Shared) {
2236 
2237       if (!ExprsInSubprogram.count(V))
2238         return;
2239 
2240       auto I = Shared.insert({V, {}});
2241       I.first->second.insert(Leaf);
2242 
2243       for (Value *Op : cast<Instruction>(V)->operand_values())
2244         collectSharedInfo(Leaf, Op, ExprsInSubprogram, Shared);
2245     }
2246 
2247     /// Calculate the number of exclusive and shared op counts for expression
2248     /// starting at \p V. Expressions used multiple times are counted once.
2249     /// Limit the matrix operations to the ones in \p ExprsInSubprogram.
2250     std::pair<OpInfoTy, OpInfoTy>
2251     sumOpInfos(Value *Root, SmallPtrSetImpl<Value *> &ReusedExprs,
2252                const SmallSetVector<Value *, 32> &ExprsInSubprogram,
2253                DenseMap<Value *, SmallPtrSet<Value *, 2>> &Shared) const {
2254       if (!ExprsInSubprogram.count(Root))
2255         return {};
2256 
2257       // Already counted this expression. Stop.
2258       if (!ReusedExprs.insert(Root).second)
2259         return {};
2260 
2261       OpInfoTy SharedCount;
2262       OpInfoTy Count;
2263 
2264       auto I = Shared.find(Root);
2265       auto CM = Inst2Matrix.find(Root);
2266       if (I->second.size() == 1)
2267         Count = CM->second.getOpInfo();
2268       else
2269         SharedCount = CM->second.getOpInfo();
2270 
2271       for (Value *Op : cast<Instruction>(Root)->operand_values()) {
2272         auto C = sumOpInfos(Op, ReusedExprs, ExprsInSubprogram, Shared);
2273         Count += C.first;
2274         SharedCount += C.second;
2275       }
2276       return {Count, SharedCount};
2277     }
2278 
2279     void emitRemarks() {
2280       if (!ORE.allowExtraAnalysis(DEBUG_TYPE))
2281         return;
2282 
2283       // Map matrix operations to their containting subprograms, by traversing
2284       // the inlinedAt chain. If the function does not have a DISubprogram, we
2285       // only map them to the containing function.
2286       MapVector<DISubprogram *, SmallVector<Value *, 8>> Subprog2Exprs;
2287       for (const auto &KV : Inst2Matrix) {
2288         if (Func.getSubprogram()) {
2289           auto *I = cast<Instruction>(KV.first);
2290           DILocation *Context = I->getDebugLoc();
2291           while (Context) {
2292             auto I =
2293                 Subprog2Exprs.insert({getSubprogram(Context->getScope()), {}});
2294             I.first->second.push_back(KV.first);
2295             Context = DebugLoc(Context).getInlinedAt();
2296           }
2297         } else {
2298           auto I = Subprog2Exprs.insert({nullptr, {}});
2299           I.first->second.push_back(KV.first);
2300         }
2301       }
2302       for (auto &KV : Subprog2Exprs) {
2303         SmallSetVector<Value *, 32> ExprsInSubprogram(KV.second.begin(),
2304                                                       KV.second.end());
2305         auto Leaves = getExpressionLeaves(ExprsInSubprogram);
2306 
2307         DenseMap<Value *, SmallPtrSet<Value *, 2>> Shared;
2308         for (Value *Leaf : Leaves)
2309           collectSharedInfo(Leaf, Leaf, ExprsInSubprogram, Shared);
2310 
2311         // Generate remarks for each leaf.
2312         for (auto *L : Leaves) {
2313 
2314           DebugLoc Loc = cast<Instruction>(L)->getDebugLoc();
2315           DILocation *Context = cast<Instruction>(L)->getDebugLoc();
2316           while (Context) {
2317             if (getSubprogram(Context->getScope()) == KV.first) {
2318               Loc = Context;
2319               break;
2320             }
2321             Context = DebugLoc(Context).getInlinedAt();
2322           }
2323 
2324           SmallPtrSet<Value *, 8> ReusedExprs;
2325           OpInfoTy Counts, SharedCounts;
2326           std::tie(Counts, SharedCounts) =
2327               sumOpInfos(L, ReusedExprs, ExprsInSubprogram, Shared);
2328 
2329           OptimizationRemark Rem(DEBUG_TYPE, "matrix-lowered", Loc,
2330                                  cast<Instruction>(L)->getParent());
2331 
2332           Rem << "Lowered with ";
2333           Rem << ore::NV("NumStores", Counts.NumStores) << " stores, "
2334               << ore::NV("NumLoads", Counts.NumLoads) << " loads, "
2335               << ore::NV("NumComputeOps", Counts.NumComputeOps)
2336               << " compute ops, "
2337               << ore::NV("NumExposedTransposes", Counts.NumExposedTransposes)
2338               << " exposed transposes";
2339 
2340           if (SharedCounts.NumStores > 0 || SharedCounts.NumLoads > 0 ||
2341               SharedCounts.NumComputeOps > 0) {
2342             Rem << ",\nadditionally "
2343                 << ore::NV("NumStores", SharedCounts.NumStores) << " stores, "
2344                 << ore::NV("NumLoads", SharedCounts.NumLoads) << " loads, "
2345                 << ore::NV("NumFPOps", SharedCounts.NumComputeOps)
2346                 << " compute ops"
2347                 << " are shared with other expressions";
2348           }
2349 
2350           Rem << ("\n" + linearize(L, Shared, ExprsInSubprogram, DL));
2351           ORE.emit(Rem);
2352         }
2353       }
2354     }
2355 
2356     std::string
2357     linearize(Value *L,
2358               const DenseMap<Value *, SmallPtrSet<Value *, 2>> &Shared,
2359               const SmallSetVector<Value *, 32> &ExprsInSubprogram,
2360               const DataLayout &DL) {
2361       ExprLinearizer Lin(DL, Inst2Matrix, Shared, ExprsInSubprogram, L);
2362       Lin.linearizeExpr(L, 0, false, false);
2363       return Lin.getResult();
2364     }
2365   };
2366 };
2367 } // namespace
2368 
2369 PreservedAnalyses LowerMatrixIntrinsicsPass::run(Function &F,
2370                                                  FunctionAnalysisManager &AM) {
2371   auto &TTI = AM.getResult<TargetIRAnalysis>(F);
2372   OptimizationRemarkEmitter *ORE = nullptr;
2373   AAResults *AA = nullptr;
2374   DominatorTree *DT = nullptr;
2375   LoopInfo *LI = nullptr;
2376 
2377   if (!Minimal) {
2378     ORE = &AM.getResult<OptimizationRemarkEmitterAnalysis>(F);
2379     AA = &AM.getResult<AAManager>(F);
2380     DT = &AM.getResult<DominatorTreeAnalysis>(F);
2381     LI = &AM.getResult<LoopAnalysis>(F);
2382   }
2383 
2384   LowerMatrixIntrinsics LMT(F, TTI, AA, DT, LI, ORE);
2385   if (LMT.Visit()) {
2386     PreservedAnalyses PA;
2387     if (!Minimal) {
2388       PA.preserve<LoopAnalysis>();
2389       PA.preserve<DominatorTreeAnalysis>();
2390     }
2391     return PA;
2392   }
2393   return PreservedAnalyses::all();
2394 }
2395 
2396 void LowerMatrixIntrinsicsPass::printPipeline(
2397     raw_ostream &OS, function_ref<StringRef(StringRef)> MapClassName2PassName) {
2398   static_cast<PassInfoMixin<LowerMatrixIntrinsicsPass> *>(this)->printPipeline(
2399       OS, MapClassName2PassName);
2400   OS << "<";
2401   if (Minimal)
2402     OS << "minimal";
2403   OS << ">";
2404 }
2405 
2406 namespace {
2407 
2408 class LowerMatrixIntrinsicsLegacyPass : public FunctionPass {
2409 public:
2410   static char ID;
2411 
2412   LowerMatrixIntrinsicsLegacyPass() : FunctionPass(ID) {
2413     initializeLowerMatrixIntrinsicsLegacyPassPass(
2414         *PassRegistry::getPassRegistry());
2415   }
2416 
2417   bool runOnFunction(Function &F) override {
2418     auto &TTI = getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
2419     auto &ORE = getAnalysis<OptimizationRemarkEmitterWrapperPass>().getORE();
2420     auto &AA = getAnalysis<AAResultsWrapperPass>().getAAResults();
2421     auto &DT = getAnalysis<DominatorTreeWrapperPass>().getDomTree();
2422     auto &LI = getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
2423     LowerMatrixIntrinsics LMT(F, TTI, &AA, &DT, &LI, &ORE);
2424     bool C = LMT.Visit();
2425     return C;
2426   }
2427 
2428   void getAnalysisUsage(AnalysisUsage &AU) const override {
2429     AU.addRequired<TargetTransformInfoWrapperPass>();
2430     AU.addRequired<OptimizationRemarkEmitterWrapperPass>();
2431     AU.addRequired<AAResultsWrapperPass>();
2432     AU.addRequired<DominatorTreeWrapperPass>();
2433     AU.addPreserved<DominatorTreeWrapperPass>();
2434     AU.addRequired<LoopInfoWrapperPass>();
2435     AU.addPreserved<LoopInfoWrapperPass>();
2436   }
2437 };
2438 } // namespace
2439 
2440 static const char pass_name[] = "Lower the matrix intrinsics";
2441 char LowerMatrixIntrinsicsLegacyPass::ID = 0;
2442 INITIALIZE_PASS_BEGIN(LowerMatrixIntrinsicsLegacyPass, DEBUG_TYPE, pass_name,
2443                       false, false)
2444 INITIALIZE_PASS_DEPENDENCY(OptimizationRemarkEmitterWrapperPass)
2445 INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass)
2446 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
2447 INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)
2448 INITIALIZE_PASS_END(LowerMatrixIntrinsicsLegacyPass, DEBUG_TYPE, pass_name,
2449                     false, false)
2450 
2451 Pass *llvm::createLowerMatrixIntrinsicsPass() {
2452   return new LowerMatrixIntrinsicsLegacyPass();
2453 }
2454 
2455 namespace {
2456 
2457 /// A lightweight version of the matrix lowering pass that only requires TTI.
2458 /// Advanced features that require DT, AA or ORE like tiling are disabled. This
2459 /// is used to lower matrix intrinsics if the main lowering pass is not run, for
2460 /// example with -O0.
2461 class LowerMatrixIntrinsicsMinimalLegacyPass : public FunctionPass {
2462 public:
2463   static char ID;
2464 
2465   LowerMatrixIntrinsicsMinimalLegacyPass() : FunctionPass(ID) {
2466     initializeLowerMatrixIntrinsicsMinimalLegacyPassPass(
2467         *PassRegistry::getPassRegistry());
2468   }
2469 
2470   bool runOnFunction(Function &F) override {
2471     auto &TTI = getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
2472     LowerMatrixIntrinsics LMT(F, TTI, nullptr, nullptr, nullptr, nullptr);
2473     bool C = LMT.Visit();
2474     return C;
2475   }
2476 
2477   void getAnalysisUsage(AnalysisUsage &AU) const override {
2478     AU.addRequired<TargetTransformInfoWrapperPass>();
2479     AU.setPreservesCFG();
2480   }
2481 };
2482 } // namespace
2483 
2484 static const char pass_name_minimal[] = "Lower the matrix intrinsics (minimal)";
2485 char LowerMatrixIntrinsicsMinimalLegacyPass::ID = 0;
2486 INITIALIZE_PASS_BEGIN(LowerMatrixIntrinsicsMinimalLegacyPass,
2487                       "lower-matrix-intrinsics-minimal", pass_name_minimal,
2488                       false, false)
2489 INITIALIZE_PASS_END(LowerMatrixIntrinsicsMinimalLegacyPass,
2490                     "lower-matrix-intrinsics-minimal", pass_name_minimal, false,
2491                     false)
2492 
2493 Pass *llvm::createLowerMatrixIntrinsicsMinimalPass() {
2494   return new LowerMatrixIntrinsicsMinimalLegacyPass();
2495 }
2496