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/GraphTraits.h"
22 #include "llvm/ADT/PostOrderIterator.h"
23 #include "llvm/ADT/SmallVector.h"
24 #include "llvm/Analysis/AliasAnalysis.h"
25 #include "llvm/Analysis/DomTreeUpdater.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/PatternMatch.h"
38 #include "llvm/InitializePasses.h"
39 #include "llvm/Pass.h"
40 #include "llvm/Support/Alignment.h"
41 #include "llvm/Support/CommandLine.h"
42 #include "llvm/Support/Debug.h"
43 #include "llvm/Transforms/Scalar.h"
44 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
45 
46 using namespace llvm;
47 using namespace PatternMatch;
48 
49 #define DEBUG_TYPE "lower-matrix-intrinsics"
50 
51 static cl::opt<bool> EnableShapePropagation(
52     "matrix-propagate-shape", cl::init(true), cl::Hidden,
53     cl::desc("Enable/disable shape propagation from matrix intrinsics to other "
54              "instructions."));
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> ForceFusion(
65     "force-fuse-matrix", cl::init(false), cl::Hidden,
66     cl::desc("Force matrix instruction fusion even if not profitable."));
67 static cl::opt<bool> AllowContractEnabled(
68     "matrix-allow-contract", cl::init(false), cl::Hidden,
69     cl::desc("Allow the use of FMAs if available and profitable. This may "
70              "result in different results, due to less rounding error."));
71 
72 enum class MatrixLayoutTy { ColumnMajor, RowMajor };
73 
74 static cl::opt<MatrixLayoutTy> MatrixLayout(
75     "matrix-default-layout", cl::init(MatrixLayoutTy::ColumnMajor),
76     cl::desc("Sets the default matrix layout"),
77     cl::values(clEnumValN(MatrixLayoutTy::ColumnMajor, "column-major",
78                           "Use column-major layout"),
79                clEnumValN(MatrixLayoutTy::RowMajor, "row-major",
80                           "Use row-major layout")));
81 
82 /// Helper function to either return Scope, if it is a subprogram or the
83 /// attached subprogram for a local scope.
getSubprogram(DIScope * Scope)84 static DISubprogram *getSubprogram(DIScope *Scope) {
85   if (auto *Subprogram = dyn_cast<DISubprogram>(Scope))
86     return Subprogram;
87   return cast<DILocalScope>(Scope)->getSubprogram();
88 }
89 
90 namespace {
91 
92 // Given an element pointer \p BasePtr to the start of a (sub) matrix, compute
93 // the start address of vector \p VecIdx with type (\p EltType x \p NumElements)
94 // assuming \p Stride elements between start two consecutive vectors.
95 // \p Stride must be >= \p NumElements.
96 // For column-major matrixes, the function computes the address of a column
97 // vectors and \p NumElements must be set to the number of elements in a column
98 // (= number of rows of the matrix). For row-major matrixes, the function
99 // computes the address of a row vector and \p NumElements must be set to the
100 // number of elements in a column (= number of columns of the matrix).
101 //
102 // Consider a 4x4 matrix in column-mjaor layout like below
103 //
104 //      0       1      2      3
105 // 0   v_0_0  v_0_1  v_0_2  v_0_3
106 // 1   v_1_0  v_1_1  v_1_2  v_1_3
107 // 2   v_2_0  v_2_1  v_2_2  v_2_3
108 // 3   v_3_0  v_3_1  v_3_2  v_3_3
109 
110 // To compute the column addresses for a 2x3 sub-matrix at row 1 and column 1,
111 // we need a pointer to the first element of the submatrix as base pointer.
112 // Then we can use computeVectorAddr to compute the addresses for the columns
113 // of the sub-matrix.
114 //
115 // Column 0: computeVectorAddr(Base, 0 (column), 4 (stride), 2 (num rows), ..)
116 //           -> just returns Base
117 // Column 1: computeVectorAddr(Base, 1 (column), 4 (stride), 2 (num rows), ..)
118 //           -> returns Base + (1 * 4)
119 // Column 2: computeVectorAddr(Base, 2 (column), 4 (stride), 2 (num rows), ..)
120 //           -> returns Base + (2 * 4)
121 //
122 // The graphic below illustrates the number of elements in a column (marked
123 // with |) and the number of skipped elements (marked with }).
124 //
125 //         v_0_0  v_0_1 {v_0_2 {v_0_3
126 //                Base   Col 1  Col 2
127 //                  |     |      |
128 //         v_1_0 |v_1_1 |v_1_2 |v_1_3
129 //         v_2_0 |v_2_1 |v_2_2 |v_2_3
130 //         v_3_0 {v_3_1 {v_3_2  v_3_3
131 //
computeVectorAddr(Value * BasePtr,Value * VecIdx,Value * Stride,unsigned NumElements,Type * EltType,IRBuilder<> & Builder)132 Value *computeVectorAddr(Value *BasePtr, Value *VecIdx, Value *Stride,
133                          unsigned NumElements, Type *EltType,
134                          IRBuilder<> &Builder) {
135 
136   assert((!isa<ConstantInt>(Stride) ||
137           cast<ConstantInt>(Stride)->getZExtValue() >= NumElements) &&
138          "Stride must be >= the number of elements in the result vector.");
139   unsigned AS = cast<PointerType>(BasePtr->getType())->getAddressSpace();
140 
141   // Compute the start of the vector with index VecIdx as VecIdx * Stride.
142   Value *VecStart = Builder.CreateMul(VecIdx, Stride, "vec.start");
143 
144   // Get pointer to the start of the selected vector. Skip GEP creation,
145   // if we select vector 0.
146   if (isa<ConstantInt>(VecStart) && cast<ConstantInt>(VecStart)->isZero())
147     VecStart = BasePtr;
148   else
149     VecStart = Builder.CreateGEP(EltType, BasePtr, VecStart, "vec.gep");
150 
151   // Cast elementwise vector start pointer to a pointer to a vector
152   // (EltType x NumElements)*.
153   auto *VecType = FixedVectorType::get(EltType, NumElements);
154   Type *VecPtrType = PointerType::get(VecType, AS);
155   return Builder.CreatePointerCast(VecStart, VecPtrType, "vec.cast");
156 }
157 
158 /// LowerMatrixIntrinsics contains the methods used to lower matrix intrinsics.
159 ///
160 /// Currently, the lowering for each matrix intrinsic is done as follows:
161 /// 1. Propagate the shape information from intrinsics to connected
162 /// instructions.
163 /// 2. Lower instructions with shape information (assuming column-major layout).
164 ///  The lowering works similarly using row-major layout.
165 ///  2.1. Get column vectors for each argument. If we already lowered the
166 ///       definition of an argument, use the produced column vectors directly.
167 ///       If not, split the operand vector containing an embedded matrix into
168 ///       a set of column vectors,
169 ///  2.2. Lower the instruction in terms of column major operations, which
170 ///       yields a set of column vectors containing result matrix. Note that we
171 ///       lower all instructions that have shape information. Besides the
172 ///       intrinsics, this includes stores for example.
173 ///  2.3. Update uses of the lowered instruction. If we have shape information
174 ///       for a user, there is nothing to do, as we will look up the result
175 ///       column matrix when lowering the user. For other uses, we embed the
176 ///       result matrix in a flat vector and update the use.
177 ///  2.4. Cache the result column matrix for the instruction we lowered
178 /// 3. After we lowered all instructions in a function, remove the now
179 ///    obsolete instructions.
180 ///
181 class LowerMatrixIntrinsics {
182   Function &Func;
183   const DataLayout &DL;
184   const TargetTransformInfo &TTI;
185   AliasAnalysis &AA;
186   DominatorTree &DT;
187   LoopInfo &LI;
188   OptimizationRemarkEmitter &ORE;
189 
190   /// Contains estimates of the number of operations (loads, stores, compute) required to lower a matrix operation.
191   struct OpInfoTy {
192     /// Number of stores emitted to generate this matrix.
193     unsigned NumStores = 0;
194     /// Number of loads emitted to generate this matrix.
195     unsigned NumLoads = 0;
196     /// Number of compute operations emitted to generate this matrix.
197     unsigned NumComputeOps = 0;
198 
operator +=__anon643663df0111::LowerMatrixIntrinsics::OpInfoTy199     OpInfoTy &operator+=(const OpInfoTy &RHS) {
200       NumStores += RHS.NumStores;
201       NumLoads += RHS.NumLoads;
202       NumComputeOps += RHS.NumComputeOps;
203       return *this;
204     }
205   };
206 
207   /// Wrapper class representing a matrix as a set of vectors, either in row or
208   /// column major layout. All vectors must have the same vector type.
209   class MatrixTy {
210     SmallVector<Value *, 16> Vectors;
211 
212     OpInfoTy OpInfo;
213 
214     bool IsColumnMajor = true;
215 
216   public:
MatrixTy()217     MatrixTy()
218         : Vectors(),
219           IsColumnMajor(MatrixLayout == MatrixLayoutTy::ColumnMajor) {}
MatrixTy(ArrayRef<Value * > Vectors)220     MatrixTy(ArrayRef<Value *> Vectors)
221         : Vectors(Vectors.begin(), Vectors.end()),
222           IsColumnMajor(MatrixLayout == MatrixLayoutTy::ColumnMajor) {}
MatrixTy(unsigned NumRows,unsigned NumColumns,Type * EltTy)223     MatrixTy(unsigned NumRows, unsigned NumColumns, Type *EltTy)
224         : IsColumnMajor(MatrixLayout == MatrixLayoutTy::ColumnMajor) {
225 
226       unsigned D = isColumnMajor() ? NumColumns : NumRows;
227       for (unsigned J = 0; J < D; ++J)
228         addVector(UndefValue::get(FixedVectorType::get(
229             EltTy, isColumnMajor() ? NumRows : NumColumns)));
230     }
231 
getVector(unsigned i) const232     Value *getVector(unsigned i) const { return Vectors[i]; }
getColumn(unsigned i) const233     Value *getColumn(unsigned i) const {
234       assert(isColumnMajor() && "only supported for column-major matrixes");
235       return Vectors[i];
236     }
getRow(unsigned i) const237     Value *getRow(unsigned i) const {
238       assert(!isColumnMajor() && "only supported for row-major matrixes");
239       return Vectors[i];
240     }
241 
setVector(unsigned i,Value * V)242     void setVector(unsigned i, Value *V) { Vectors[i] = V; }
243 
getElementType()244     Type *getElementType() { return getVectorTy()->getElementType(); }
245 
getNumVectors() const246     unsigned getNumVectors() const {
247       if (isColumnMajor())
248         return getNumColumns();
249       return getNumRows();
250     }
251 
getNumColumns() const252     unsigned getNumColumns() const {
253       if (isColumnMajor())
254         return Vectors.size();
255       else {
256         assert(Vectors.size() > 0 && "Cannot call getNumRows without columns");
257         return cast<FixedVectorType>(Vectors[0]->getType())->getNumElements();
258       }
259     }
getNumRows() const260     unsigned getNumRows() const {
261       if (isColumnMajor()) {
262         assert(Vectors.size() > 0 && "Cannot call getNumRows without columns");
263         return cast<FixedVectorType>(Vectors[0]->getType())->getNumElements();
264       } else
265         return Vectors.size();
266     }
267 
addVector(Value * V)268     void addVector(Value *V) { Vectors.push_back(V); }
getColumnTy()269     VectorType *getColumnTy() {
270       assert(isColumnMajor() && "only supported for column-major matrixes");
271       return getVectorTy();
272     }
273 
getVectorTy()274     VectorType *getVectorTy() {
275       return cast<VectorType>(Vectors[0]->getType());
276     }
277 
columns()278     iterator_range<SmallVector<Value *, 8>::iterator> columns() {
279       assert(isColumnMajor() &&
280              "columns() only supported for column-major matrixes");
281       return make_range(Vectors.begin(), Vectors.end());
282     }
283 
vectors()284     iterator_range<SmallVector<Value *, 8>::iterator> vectors() {
285       return make_range(Vectors.begin(), Vectors.end());
286     }
287 
288     /// Embed the vectors of the matrix into a flat vector by concatenating
289     /// them.
embedInVector(IRBuilder<> & Builder) const290     Value *embedInVector(IRBuilder<> &Builder) const {
291       return Vectors.size() == 1 ? Vectors[0]
292                                  : concatenateVectors(Builder, Vectors);
293     }
294 
addNumLoads(unsigned N)295     MatrixTy &addNumLoads(unsigned N) {
296       OpInfo.NumLoads += N;
297       return *this;
298     }
299 
setNumLoads(unsigned N)300     void setNumLoads(unsigned N) { OpInfo.NumLoads = N; }
301 
addNumStores(unsigned N)302     MatrixTy &addNumStores(unsigned N) {
303       OpInfo.NumStores += N;
304       return *this;
305     }
306 
addNumComputeOps(unsigned N)307     MatrixTy &addNumComputeOps(unsigned N) {
308       OpInfo.NumComputeOps += N;
309       return *this;
310     }
311 
getNumStores() const312     unsigned getNumStores() const { return OpInfo.NumStores; }
getNumLoads() const313     unsigned getNumLoads() const { return OpInfo.NumLoads; }
getNumComputeOps() const314     unsigned getNumComputeOps() const { return OpInfo.NumComputeOps; }
315 
getOpInfo() const316     const OpInfoTy &getOpInfo() const { return OpInfo; }
317 
isColumnMajor() const318     bool isColumnMajor() const { return IsColumnMajor; }
319 
getStride() const320     unsigned getStride() const {
321       if (isColumnMajor())
322         return getNumRows();
323       return getNumColumns();
324     }
325 
326     /// Extract a vector of \p NumElts starting at index (\p I, \p J). If the
327     /// matrix is column-major, the result vector is extracted from a column
328     /// vector, otherwise from a row vector.
extractVector(unsigned I,unsigned J,unsigned NumElts,IRBuilder<> & Builder) const329     Value *extractVector(unsigned I, unsigned J, unsigned NumElts,
330                          IRBuilder<> &Builder) const {
331       Value *Vec = isColumnMajor() ? getColumn(J) : getRow(I);
332       Value *Undef = UndefValue::get(Vec->getType());
333       return Builder.CreateShuffleVector(
334           Vec, Undef, createSequentialMask(isColumnMajor() ? I : J, NumElts, 0),
335           "block");
336     }
337   };
338 
339   struct ShapeInfo {
340     unsigned NumRows;
341     unsigned NumColumns;
342 
343     bool IsColumnMajor;
344 
ShapeInfo__anon643663df0111::LowerMatrixIntrinsics::ShapeInfo345     ShapeInfo(unsigned NumRows = 0, unsigned NumColumns = 0)
346         : NumRows(NumRows), NumColumns(NumColumns),
347           IsColumnMajor(MatrixLayout == MatrixLayoutTy::ColumnMajor) {}
348 
ShapeInfo__anon643663df0111::LowerMatrixIntrinsics::ShapeInfo349     ShapeInfo(Value *NumRows, Value *NumColumns)
350         : ShapeInfo(cast<ConstantInt>(NumRows)->getZExtValue(),
351                     cast<ConstantInt>(NumColumns)->getZExtValue()) {}
352 
operator ==__anon643663df0111::LowerMatrixIntrinsics::ShapeInfo353     bool operator==(const ShapeInfo &other) {
354       return NumRows == other.NumRows && NumColumns == other.NumColumns;
355     }
operator !=__anon643663df0111::LowerMatrixIntrinsics::ShapeInfo356     bool operator!=(const ShapeInfo &other) { return !(*this == other); }
357 
358     /// Returns true if shape-information is defined, meaning both dimensions
359     /// are != 0.
operator bool__anon643663df0111::LowerMatrixIntrinsics::ShapeInfo360     operator bool() const {
361       assert(NumRows == 0 || NumColumns != 0);
362       return NumRows != 0;
363     }
364 
getStride__anon643663df0111::LowerMatrixIntrinsics::ShapeInfo365     unsigned getStride() const {
366       if (IsColumnMajor)
367         return NumRows;
368       return NumColumns;
369     }
370 
getNumVectors__anon643663df0111::LowerMatrixIntrinsics::ShapeInfo371     unsigned getNumVectors() const {
372       if (IsColumnMajor)
373         return NumColumns;
374       return NumRows;
375     }
376   };
377 
378   /// Maps instructions to their shape information. The shape information
379   /// describes the shape to be used while lowering. This matches the shape of
380   /// the result value of the instruction, with the only exceptions being store
381   /// instructions and the matrix_column_major_store intrinsics. For those, the
382   /// shape information indicates that those instructions should be lowered
383   /// using shape information as well.
384   DenseMap<Value *, ShapeInfo> ShapeMap;
385 
386   /// List of instructions to remove. While lowering, we are not replacing all
387   /// users of a lowered instruction, if shape information is available and
388   /// those need to be removed after we finished lowering.
389   SmallVector<Instruction *, 16> ToRemove;
390 
391   /// Map from instructions to their produced column matrix.
392   MapVector<Value *, MatrixTy> Inst2ColumnMatrix;
393 
394 public:
LowerMatrixIntrinsics(Function & F,TargetTransformInfo & TTI,AliasAnalysis & AA,DominatorTree & DT,LoopInfo & LI,OptimizationRemarkEmitter & ORE)395   LowerMatrixIntrinsics(Function &F, TargetTransformInfo &TTI,
396                         AliasAnalysis &AA, DominatorTree &DT, LoopInfo &LI,
397                         OptimizationRemarkEmitter &ORE)
398       : Func(F), DL(F.getParent()->getDataLayout()), TTI(TTI), AA(AA), DT(DT),
399         LI(LI), ORE(ORE) {}
400 
getNumOps(Type * VT)401   unsigned getNumOps(Type *VT) {
402     assert(isa<VectorType>(VT) && "Expected vector type");
403     return getNumOps(VT->getScalarType(),
404                      cast<FixedVectorType>(VT)->getNumElements());
405   }
406 
407   //
408   /// Return the estimated number of vector ops required for an operation on
409   /// \p VT * N.
getNumOps(Type * ST,unsigned N)410   unsigned getNumOps(Type *ST, unsigned N) {
411     return std::ceil((ST->getPrimitiveSizeInBits() * N).getFixedSize() /
412                      double(TTI.getRegisterBitWidth(true)));
413   }
414 
415   /// Return the set of vectors that a matrix value is lowered to.
416   ///
417   /// If we lowered \p MatrixVal, just return the cache result matrix. Otherwise
418   /// split the flat vector \p MatrixVal containing a matrix with shape \p SI
419   /// into vectors.
getMatrix(Value * MatrixVal,const ShapeInfo & SI,IRBuilder<> & Builder)420   MatrixTy getMatrix(Value *MatrixVal, const ShapeInfo &SI,
421                      IRBuilder<> &Builder) {
422     VectorType *VType = dyn_cast<VectorType>(MatrixVal->getType());
423     assert(VType && "MatrixVal must be a vector type");
424     assert(cast<FixedVectorType>(VType)->getNumElements() ==
425                SI.NumRows * SI.NumColumns &&
426            "The vector size must match the number of matrix elements");
427 
428     // Check if we lowered MatrixVal using shape information. In that case,
429     // return the existing matrix, if it matches the requested shape
430     // information. If there is a mis-match, embed the result in a flat
431     // vector and split it later.
432     auto Found = Inst2ColumnMatrix.find(MatrixVal);
433     if (Found != Inst2ColumnMatrix.end()) {
434       MatrixTy &M = Found->second;
435       // Return the found matrix, if its shape matches the requested shape
436       // information
437       if (SI.NumRows == M.getNumRows() && SI.NumColumns == M.getNumColumns())
438         return M;
439 
440       MatrixVal = M.embedInVector(Builder);
441     }
442 
443     // Otherwise split MatrixVal.
444     SmallVector<Value *, 16> SplitVecs;
445     Value *Undef = UndefValue::get(VType);
446     for (unsigned MaskStart = 0;
447          MaskStart < cast<FixedVectorType>(VType)->getNumElements();
448          MaskStart += SI.getStride()) {
449       Value *V = Builder.CreateShuffleVector(
450           MatrixVal, Undef, createSequentialMask(MaskStart, SI.getStride(), 0),
451           "split");
452       SplitVecs.push_back(V);
453     }
454 
455     return {SplitVecs};
456   }
457 
458   /// If \p V already has a known shape return false.  Otherwise set the shape
459   /// for instructions that support it.
setShapeInfo(Value * V,ShapeInfo Shape)460   bool setShapeInfo(Value *V, ShapeInfo Shape) {
461     assert(Shape && "Shape not set");
462     if (isa<UndefValue>(V) || !supportsShapeInfo(V))
463       return false;
464 
465     auto SIter = ShapeMap.find(V);
466     if (SIter != ShapeMap.end()) {
467       LLVM_DEBUG(dbgs() << "  not overriding existing shape: "
468                         << SIter->second.NumRows << " "
469                         << SIter->second.NumColumns << " for " << *V << "\n");
470       return false;
471     }
472 
473     ShapeMap.insert({V, Shape});
474     LLVM_DEBUG(dbgs() << "  " << Shape.NumRows << " x " << Shape.NumColumns
475                       << " for " << *V << "\n");
476     return true;
477   }
478 
isUniformShape(Value * V)479   bool isUniformShape(Value *V) {
480     Instruction *I = dyn_cast<Instruction>(V);
481     if (!I)
482       return true;
483 
484     switch (I->getOpcode()) {
485     case Instruction::FAdd:
486     case Instruction::FSub:
487     case Instruction::FMul: // Scalar multiply.
488     case Instruction::Add:
489     case Instruction::Mul:
490     case Instruction::Sub:
491       return true;
492     default:
493       return false;
494     }
495   }
496 
497   /// Returns true if shape information can be used for \p V. The supported
498   /// instructions must match the instructions that can be lowered by this pass.
supportsShapeInfo(Value * V)499   bool supportsShapeInfo(Value *V) {
500     Instruction *Inst = dyn_cast<Instruction>(V);
501     if (!Inst)
502       return false;
503 
504     IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst);
505     if (II)
506       switch (II->getIntrinsicID()) {
507       case Intrinsic::matrix_multiply:
508       case Intrinsic::matrix_transpose:
509       case Intrinsic::matrix_column_major_load:
510       case Intrinsic::matrix_column_major_store:
511         return true;
512       default:
513         return false;
514       }
515     return isUniformShape(V) || isa<StoreInst>(V) || isa<LoadInst>(V);
516   }
517 
518   /// Propagate the shape information of instructions to their users.
519   /// The work list contains instructions for which we can compute the shape,
520   /// either based on the information provided by matrix intrinsics or known
521   /// shapes of operands.
522   SmallVector<Instruction *, 32>
propagateShapeForward(SmallVectorImpl<Instruction * > & WorkList)523   propagateShapeForward(SmallVectorImpl<Instruction *> &WorkList) {
524     SmallVector<Instruction *, 32> NewWorkList;
525     // Pop an element for which we guaranteed to have at least one of the
526     // operand shapes.  Add the shape for this and then add users to the work
527     // list.
528     LLVM_DEBUG(dbgs() << "Forward-propagate shapes:\n");
529     while (!WorkList.empty()) {
530       Instruction *Inst = WorkList.back();
531       WorkList.pop_back();
532 
533       // New entry, set the value and insert operands
534       bool Propagate = false;
535 
536       Value *MatrixA;
537       Value *MatrixB;
538       Value *M;
539       Value *N;
540       Value *K;
541       if (match(Inst, m_Intrinsic<Intrinsic::matrix_multiply>(
542                           m_Value(MatrixA), m_Value(MatrixB), m_Value(M),
543                           m_Value(N), m_Value(K)))) {
544         Propagate = setShapeInfo(Inst, {M, K});
545       } else if (match(Inst, m_Intrinsic<Intrinsic::matrix_transpose>(
546                                  m_Value(MatrixA), m_Value(M), m_Value(N)))) {
547         // Flip dimensions.
548         Propagate = setShapeInfo(Inst, {N, M});
549       } else if (match(Inst, m_Intrinsic<Intrinsic::matrix_column_major_store>(
550                                  m_Value(MatrixA), m_Value(), m_Value(),
551                                  m_Value(), m_Value(M), m_Value(N)))) {
552         Propagate = setShapeInfo(Inst, {N, M});
553       } else if (match(Inst, m_Intrinsic<Intrinsic::matrix_column_major_load>(
554                                  m_Value(), m_Value(), m_Value(), m_Value(M),
555                                  m_Value(N)))) {
556         Propagate = setShapeInfo(Inst, {M, N});
557       } else if (match(Inst, m_Store(m_Value(MatrixA), m_Value()))) {
558         auto OpShape = ShapeMap.find(MatrixA);
559         if (OpShape != ShapeMap.end())
560           setShapeInfo(Inst, OpShape->second);
561         continue;
562       } else if (isUniformShape(Inst)) {
563         // Find the first operand that has a known shape and use that.
564         for (auto &Op : Inst->operands()) {
565           auto OpShape = ShapeMap.find(Op.get());
566           if (OpShape != ShapeMap.end()) {
567             Propagate |= setShapeInfo(Inst, OpShape->second);
568             break;
569           }
570         }
571       }
572 
573       if (Propagate) {
574         NewWorkList.push_back(Inst);
575         for (auto *User : Inst->users())
576           if (ShapeMap.count(User) == 0)
577             WorkList.push_back(cast<Instruction>(User));
578       }
579     }
580 
581     return NewWorkList;
582   }
583 
584   /// Propagate the shape to operands of instructions with shape information.
585   /// \p Worklist contains the instruction for which we already know the shape.
586   SmallVector<Instruction *, 32>
propagateShapeBackward(SmallVectorImpl<Instruction * > & WorkList)587   propagateShapeBackward(SmallVectorImpl<Instruction *> &WorkList) {
588     SmallVector<Instruction *, 32> NewWorkList;
589 
590     auto pushInstruction = [](Value *V,
591                               SmallVectorImpl<Instruction *> &WorkList) {
592       Instruction *I = dyn_cast<Instruction>(V);
593       if (I)
594         WorkList.push_back(I);
595     };
596     // Pop an element with known shape.  Traverse the operands, if their shape
597     // derives from the result shape and is unknown, add it and add them to the
598     // worklist.
599     LLVM_DEBUG(dbgs() << "Backward-propagate shapes:\n");
600     while (!WorkList.empty()) {
601       Value *V = WorkList.back();
602       WorkList.pop_back();
603 
604       size_t BeforeProcessingV = WorkList.size();
605       if (!isa<Instruction>(V))
606         continue;
607 
608       Value *MatrixA;
609       Value *MatrixB;
610       Value *M;
611       Value *N;
612       Value *K;
613       if (match(V, m_Intrinsic<Intrinsic::matrix_multiply>(
614                        m_Value(MatrixA), m_Value(MatrixB), m_Value(M),
615                        m_Value(N), m_Value(K)))) {
616         if (setShapeInfo(MatrixA, {M, N}))
617           pushInstruction(MatrixA, WorkList);
618 
619         if (setShapeInfo(MatrixB, {N, K}))
620           pushInstruction(MatrixB, WorkList);
621 
622       } else if (match(V, m_Intrinsic<Intrinsic::matrix_transpose>(
623                               m_Value(MatrixA), m_Value(M), m_Value(N)))) {
624         // Flip dimensions.
625         if (setShapeInfo(MatrixA, {M, N}))
626           pushInstruction(MatrixA, WorkList);
627       } else if (match(V, m_Intrinsic<Intrinsic::matrix_column_major_store>(
628                               m_Value(MatrixA), m_Value(), m_Value(), m_Value(),
629                               m_Value(M), m_Value(N)))) {
630         if (setShapeInfo(MatrixA, {M, N})) {
631           pushInstruction(MatrixA, WorkList);
632         }
633       } else if (isa<LoadInst>(V) ||
634                  match(V, m_Intrinsic<Intrinsic::matrix_column_major_load>())) {
635         // Nothing to do, no matrix input.
636       } else if (isa<StoreInst>(V)) {
637         // Nothing to do.  We forward-propagated to this so we would just
638         // backward propagate to an instruction with an already known shape.
639       } else if (isUniformShape(V)) {
640         // Propagate to all operands.
641         ShapeInfo Shape = ShapeMap[V];
642         for (Use &U : cast<Instruction>(V)->operands()) {
643           if (setShapeInfo(U.get(), Shape))
644             pushInstruction(U.get(), WorkList);
645         }
646       }
647       // After we discovered new shape info for new instructions in the
648       // worklist, we use their users as seeds for the next round of forward
649       // propagation.
650       for (size_t I = BeforeProcessingV; I != WorkList.size(); I++)
651         for (User *U : WorkList[I]->users())
652           if (isa<Instruction>(U) && V != U)
653             NewWorkList.push_back(cast<Instruction>(U));
654     }
655     return NewWorkList;
656   }
657 
Visit()658   bool Visit() {
659     if (EnableShapePropagation) {
660       SmallVector<Instruction *, 32> WorkList;
661 
662       // Initially only the shape of matrix intrinsics is known.
663       // Initialize the work list with ops carrying shape information.
664       for (BasicBlock &BB : Func)
665         for (Instruction &Inst : BB) {
666           IntrinsicInst *II = dyn_cast<IntrinsicInst>(&Inst);
667           if (!II)
668             continue;
669 
670           switch (II->getIntrinsicID()) {
671           case Intrinsic::matrix_multiply:
672           case Intrinsic::matrix_transpose:
673           case Intrinsic::matrix_column_major_load:
674           case Intrinsic::matrix_column_major_store:
675             WorkList.push_back(&Inst);
676             break;
677           default:
678             break;
679           }
680         }
681       // Propagate shapes until nothing changes any longer.
682       while (!WorkList.empty()) {
683         WorkList = propagateShapeForward(WorkList);
684         WorkList = propagateShapeBackward(WorkList);
685       }
686     }
687 
688     bool Changed = false;
689     SmallVector<CallInst *, 16> MaybeFusableInsts;
690     SmallVector<Instruction *, 16> MatrixInsts;
691 
692     // First, collect all instructions with shape information and candidates for
693     // fusion (currently only matrix multiplies).
694     ReversePostOrderTraversal<Function *> RPOT(&Func);
695     for (auto *BB : RPOT)
696       for (Instruction &I : *BB) {
697         if (ShapeMap.find(&I) == ShapeMap.end())
698           continue;
699         if (match(&I, m_Intrinsic<Intrinsic::matrix_multiply>()))
700           MaybeFusableInsts.push_back(cast<CallInst>(&I));
701         MatrixInsts.push_back(&I);
702       }
703 
704     // Second, try to fuse candidates.
705     SmallPtrSet<Instruction *, 16> FusedInsts;
706     for (CallInst *CI : MaybeFusableInsts)
707       LowerMatrixMultiplyFused(CI, FusedInsts);
708     Changed = !FusedInsts.empty();
709 
710     // Third, lower remaining instructions with shape information.
711     for (Instruction *Inst : MatrixInsts) {
712       if (FusedInsts.count(Inst))
713         continue;
714 
715       IRBuilder<> Builder(Inst);
716 
717       if (CallInst *CInst = dyn_cast<CallInst>(Inst))
718         Changed |= VisitCallInst(CInst);
719 
720       Value *Op1;
721       Value *Op2;
722       if (auto *BinOp = dyn_cast<BinaryOperator>(Inst))
723         Changed |= VisitBinaryOperator(BinOp);
724       if (match(Inst, m_Load(m_Value(Op1))))
725         Changed |= VisitLoad(cast<LoadInst>(Inst), Op1, Builder);
726       else if (match(Inst, m_Store(m_Value(Op1), m_Value(Op2))))
727         Changed |= VisitStore(cast<StoreInst>(Inst), Op1, Op2, Builder);
728     }
729 
730     RemarkGenerator RemarkGen(Inst2ColumnMatrix, ORE, Func);
731     RemarkGen.emitRemarks();
732 
733     for (Instruction *Inst : reverse(ToRemove))
734       Inst->eraseFromParent();
735 
736     return Changed;
737   }
738 
739   /// Turns \p BasePtr into an elementwise pointer to \p EltType.
createElementPtr(Value * BasePtr,Type * EltType,IRBuilder<> & Builder)740   Value *createElementPtr(Value *BasePtr, Type *EltType, IRBuilder<> &Builder) {
741     unsigned AS = cast<PointerType>(BasePtr->getType())->getAddressSpace();
742     Type *EltPtrType = PointerType::get(EltType, AS);
743     return Builder.CreatePointerCast(BasePtr, EltPtrType);
744   }
745 
746   /// Replace intrinsic calls
VisitCallInst(CallInst * Inst)747   bool VisitCallInst(CallInst *Inst) {
748     if (!Inst->getCalledFunction() || !Inst->getCalledFunction()->isIntrinsic())
749       return false;
750 
751     switch (Inst->getCalledFunction()->getIntrinsicID()) {
752     case Intrinsic::matrix_multiply:
753       LowerMultiply(Inst);
754       break;
755     case Intrinsic::matrix_transpose:
756       LowerTranspose(Inst);
757       break;
758     case Intrinsic::matrix_column_major_load:
759       LowerColumnMajorLoad(Inst);
760       break;
761     case Intrinsic::matrix_column_major_store:
762       LowerColumnMajorStore(Inst);
763       break;
764     default:
765       return false;
766     }
767     return true;
768   }
769 
770   /// Compute the alignment for a column/row \p Idx with \p Stride between them.
771   /// The address at \p Idx == 0 has alignment \p A. If \p Stride is a
772   /// ConstantInt, reduce the initial alignment based on the byte offset. For
773   /// non-ConstantInt strides, return the common alignment of the initial
774   /// alignment and the element size in bytes.
getAlignForIndex(unsigned Idx,Value * Stride,Type * ElementTy,MaybeAlign A) const775   Align getAlignForIndex(unsigned Idx, Value *Stride, Type *ElementTy,
776                          MaybeAlign A) const {
777     Align InitialAlign = DL.getValueOrABITypeAlignment(A, ElementTy);
778     if (Idx == 0)
779       return InitialAlign;
780 
781     TypeSize ElementSizeInBits = DL.getTypeSizeInBits(ElementTy);
782     if (auto *ConstStride = dyn_cast<ConstantInt>(Stride)) {
783       uint64_t StrideInBytes =
784           ConstStride->getZExtValue() * ElementSizeInBits / 8;
785       return commonAlignment(InitialAlign, Idx * StrideInBytes);
786     }
787     return commonAlignment(InitialAlign, ElementSizeInBits / 8);
788   }
789 
790   /// Load a matrix with \p Shape starting at \p Ptr and using \p Stride between
791   /// vectors.
loadMatrix(Type * Ty,Value * Ptr,MaybeAlign MAlign,Value * Stride,bool IsVolatile,ShapeInfo Shape,IRBuilder<> & Builder)792   MatrixTy loadMatrix(Type *Ty, Value *Ptr, MaybeAlign MAlign, Value *Stride,
793                       bool IsVolatile, ShapeInfo Shape, IRBuilder<> &Builder) {
794     auto VType = cast<VectorType>(Ty);
795     Value *EltPtr = createElementPtr(Ptr, VType->getElementType(), Builder);
796     MatrixTy Result;
797     for (unsigned I = 0, E = Shape.getNumVectors(); I < E; ++I) {
798       Value *GEP = computeVectorAddr(EltPtr, Builder.getInt64(I), Stride,
799                                      Shape.getStride(), VType->getElementType(),
800                                      Builder);
801       Value *Vector = Builder.CreateAlignedLoad(
802           GEP, getAlignForIndex(I, Stride, VType->getElementType(), MAlign),
803           IsVolatile, "col.load");
804 
805       Result.addVector(Vector);
806     }
807     return Result.addNumLoads(getNumOps(Result.getVectorTy()) *
808                               Result.getNumVectors());
809   }
810 
811   /// Loads a sub-matrix with shape \p ResultShape from a \p R x \p C matrix,
812   /// starting at \p MatrixPtr[I][J].
loadMatrix(Value * MatrixPtr,MaybeAlign Align,bool IsVolatile,ShapeInfo MatrixShape,Value * I,Value * J,ShapeInfo ResultShape,Type * EltTy,IRBuilder<> & Builder)813   MatrixTy loadMatrix(Value *MatrixPtr, MaybeAlign Align, bool IsVolatile,
814                       ShapeInfo MatrixShape, Value *I, Value *J,
815                       ShapeInfo ResultShape, Type *EltTy,
816                       IRBuilder<> &Builder) {
817 
818     Value *Offset = Builder.CreateAdd(
819         Builder.CreateMul(J, Builder.getInt64(MatrixShape.getStride())), I);
820 
821     unsigned AS = cast<PointerType>(MatrixPtr->getType())->getAddressSpace();
822     Value *EltPtr =
823         Builder.CreatePointerCast(MatrixPtr, PointerType::get(EltTy, AS));
824     Value *TileStart = Builder.CreateGEP(EltTy, EltPtr, Offset);
825     auto *TileTy = FixedVectorType::get(EltTy, ResultShape.NumRows *
826                                                    ResultShape.NumColumns);
827     Type *TilePtrTy = PointerType::get(TileTy, AS);
828     Value *TilePtr =
829         Builder.CreatePointerCast(TileStart, TilePtrTy, "col.cast");
830 
831     return loadMatrix(TileTy, TilePtr, Align,
832                       Builder.getInt64(MatrixShape.getStride()), IsVolatile,
833                       ResultShape, Builder);
834   }
835 
836   /// Lower a load instruction with shape information.
LowerLoad(Instruction * Inst,Value * Ptr,MaybeAlign Align,Value * Stride,bool IsVolatile,ShapeInfo Shape)837   void LowerLoad(Instruction *Inst, Value *Ptr, MaybeAlign Align, Value *Stride,
838                  bool IsVolatile, ShapeInfo Shape) {
839     IRBuilder<> Builder(Inst);
840     finalizeLowering(Inst,
841                      loadMatrix(Inst->getType(), Ptr, Align, Stride, IsVolatile,
842                                 Shape, Builder),
843                      Builder);
844   }
845 
846   /// Lowers llvm.matrix.column.major.load.
847   ///
848   /// The intrinsic loads a matrix from memory using a stride between columns.
LowerColumnMajorLoad(CallInst * Inst)849   void LowerColumnMajorLoad(CallInst *Inst) {
850     assert(MatrixLayout == MatrixLayoutTy::ColumnMajor &&
851            "Intrinsic only supports column-major layout!");
852     Value *Ptr = Inst->getArgOperand(0);
853     Value *Stride = Inst->getArgOperand(1);
854     LowerLoad(Inst, Ptr, Inst->getParamAlign(0), Stride,
855               cast<ConstantInt>(Inst->getArgOperand(2))->isOne(),
856               {Inst->getArgOperand(3), Inst->getArgOperand(4)});
857   }
858 
859   /// Stores a sub-matrix \p StoreVal into the \p R x \p C matrix starting at \p
860   /// MatrixPtr[I][J].
storeMatrix(const MatrixTy & StoreVal,Value * MatrixPtr,MaybeAlign MAlign,bool IsVolatile,ShapeInfo MatrixShape,Value * I,Value * J,Type * EltTy,IRBuilder<> & Builder)861   void storeMatrix(const MatrixTy &StoreVal, Value *MatrixPtr,
862                    MaybeAlign MAlign, bool IsVolatile, ShapeInfo MatrixShape,
863                    Value *I, Value *J, Type *EltTy, IRBuilder<> &Builder) {
864     Value *Offset = Builder.CreateAdd(
865         Builder.CreateMul(J, Builder.getInt64(MatrixShape.getStride())), I);
866 
867     unsigned AS = cast<PointerType>(MatrixPtr->getType())->getAddressSpace();
868     Value *EltPtr =
869         Builder.CreatePointerCast(MatrixPtr, PointerType::get(EltTy, AS));
870     Value *TileStart = Builder.CreateGEP(EltTy, EltPtr, Offset);
871     auto *TileTy = FixedVectorType::get(EltTy, StoreVal.getNumRows() *
872                                                    StoreVal.getNumColumns());
873     Type *TilePtrTy = PointerType::get(TileTy, AS);
874     Value *TilePtr =
875         Builder.CreatePointerCast(TileStart, TilePtrTy, "col.cast");
876 
877     storeMatrix(TileTy, StoreVal, TilePtr, MAlign,
878                 Builder.getInt64(MatrixShape.getStride()), IsVolatile, Builder);
879   }
880 
881   /// Store matrix \p StoreVal starting at \p Ptr and using \p Stride between
882   /// vectors.
storeMatrix(Type * Ty,MatrixTy StoreVal,Value * Ptr,MaybeAlign MAlign,Value * Stride,bool IsVolatile,IRBuilder<> & Builder)883   MatrixTy storeMatrix(Type *Ty, MatrixTy StoreVal, Value *Ptr,
884                        MaybeAlign MAlign, Value *Stride, bool IsVolatile,
885                        IRBuilder<> &Builder) {
886     auto VType = cast<VectorType>(Ty);
887     Value *EltPtr = createElementPtr(Ptr, VType->getElementType(), Builder);
888     for (auto Vec : enumerate(StoreVal.vectors())) {
889       Value *GEP = computeVectorAddr(EltPtr, Builder.getInt64(Vec.index()),
890                                      Stride, StoreVal.getStride(),
891                                      VType->getElementType(), Builder);
892       Builder.CreateAlignedStore(Vec.value(), GEP,
893                                  getAlignForIndex(Vec.index(), Stride,
894                                                   VType->getElementType(),
895                                                   MAlign),
896                                  IsVolatile);
897     }
898     return MatrixTy().addNumStores(getNumOps(StoreVal.getVectorTy()) *
899                                    StoreVal.getNumVectors());
900   }
901 
902   /// Lower a store instruction with shape information.
LowerStore(Instruction * Inst,Value * Matrix,Value * Ptr,MaybeAlign A,Value * Stride,bool IsVolatile,ShapeInfo Shape)903   void LowerStore(Instruction *Inst, Value *Matrix, Value *Ptr, MaybeAlign A,
904                   Value *Stride, bool IsVolatile, ShapeInfo Shape) {
905     IRBuilder<> Builder(Inst);
906     auto StoreVal = getMatrix(Matrix, Shape, Builder);
907     finalizeLowering(Inst,
908                      storeMatrix(Matrix->getType(), StoreVal, Ptr, A, Stride,
909                                  IsVolatile, Builder),
910                      Builder);
911   }
912 
913   /// Lowers llvm.matrix.column.major.store.
914   ///
915   /// The intrinsic store a matrix back memory using a stride between columns.
LowerColumnMajorStore(CallInst * Inst)916   void LowerColumnMajorStore(CallInst *Inst) {
917     assert(MatrixLayout == MatrixLayoutTy::ColumnMajor &&
918            "Intrinsic only supports column-major layout!");
919     Value *Matrix = Inst->getArgOperand(0);
920     Value *Ptr = Inst->getArgOperand(1);
921     Value *Stride = Inst->getArgOperand(2);
922     LowerStore(Inst, Matrix, Ptr, Inst->getParamAlign(1), Stride,
923                cast<ConstantInt>(Inst->getArgOperand(3))->isOne(),
924                {Inst->getArgOperand(4), Inst->getArgOperand(5)});
925   }
926 
927   // Set elements I..I+NumElts-1 to Block
insertVector(Value * Col,unsigned I,Value * Block,IRBuilder<> & Builder)928   Value *insertVector(Value *Col, unsigned I, Value *Block,
929                       IRBuilder<> &Builder) {
930 
931     // First, bring Block to the same size as Col
932     unsigned BlockNumElts =
933         cast<FixedVectorType>(Block->getType())->getNumElements();
934     unsigned NumElts = cast<FixedVectorType>(Col->getType())->getNumElements();
935     assert(NumElts >= BlockNumElts && "Too few elements for current block");
936 
937     Value *Undef = UndefValue::get(Block->getType());
938     Block = Builder.CreateShuffleVector(
939         Block, Undef,
940         createSequentialMask(0, BlockNumElts, NumElts - BlockNumElts));
941 
942     // If Col is 7 long and I is 2 and BlockNumElts is 2 the mask is: 0, 1, 7,
943     // 8, 4, 5, 6
944     SmallVector<int, 16> Mask;
945     unsigned i;
946     for (i = 0; i < I; i++)
947       Mask.push_back(i);
948 
949     unsigned VecNumElts =
950         cast<FixedVectorType>(Col->getType())->getNumElements();
951     for (; i < I + BlockNumElts; i++)
952       Mask.push_back(i - I + VecNumElts);
953 
954     for (; i < VecNumElts; i++)
955       Mask.push_back(i);
956 
957     return Builder.CreateShuffleVector(Col, Block, Mask);
958   }
959 
createMulAdd(Value * Sum,Value * A,Value * B,bool UseFPOp,IRBuilder<> & Builder,bool AllowContraction,unsigned & NumComputeOps)960   Value *createMulAdd(Value *Sum, Value *A, Value *B, bool UseFPOp,
961                       IRBuilder<> &Builder, bool AllowContraction,
962                       unsigned &NumComputeOps) {
963     NumComputeOps += getNumOps(A->getType());
964     if (!Sum)
965       return UseFPOp ? Builder.CreateFMul(A, B) : Builder.CreateMul(A, B);
966 
967     if (UseFPOp) {
968       if (AllowContraction) {
969         // Use fmuladd for floating point operations and let the backend decide
970         // if that's profitable.
971         Function *FMulAdd = Intrinsic::getDeclaration(
972             Func.getParent(), Intrinsic::fmuladd, A->getType());
973         return Builder.CreateCall(FMulAdd, {A, B, Sum});
974       }
975       NumComputeOps += getNumOps(A->getType());
976       Value *Mul = Builder.CreateFMul(A, B);
977       return Builder.CreateFAdd(Sum, Mul);
978     }
979 
980     NumComputeOps += getNumOps(A->getType());
981     Value *Mul = Builder.CreateMul(A, B);
982     return Builder.CreateAdd(Sum, Mul);
983   }
984 
985   /// Cache \p Matrix as result of \p Inst and update the uses of \p Inst. For
986   /// users with shape information, there's nothing to do: the will use the
987   /// cached value when they are lowered. For other users, \p Matrix is
988   /// flattened and the uses are updated to use it. Also marks \p Inst for
989   /// deletion.
finalizeLowering(Instruction * Inst,MatrixTy Matrix,IRBuilder<> & Builder)990   void finalizeLowering(Instruction *Inst, MatrixTy Matrix,
991                         IRBuilder<> &Builder) {
992     Inst2ColumnMatrix.insert(std::make_pair(Inst, Matrix));
993 
994     ToRemove.push_back(Inst);
995     Value *Flattened = nullptr;
996     for (auto I = Inst->use_begin(), E = Inst->use_end(); I != E;) {
997       Use &U = *I++;
998       if (ShapeMap.find(U.getUser()) == ShapeMap.end()) {
999         if (!Flattened)
1000           Flattened = Matrix.embedInVector(Builder);
1001         U.set(Flattened);
1002       }
1003     }
1004   }
1005 
1006   /// Compute \p Result += \p A * \p B for input matrices with left-associating
1007   /// addition.
emitMatrixMultiply(MatrixTy & Result,const MatrixTy & A,const MatrixTy & B,bool AllowContraction,IRBuilder<> & Builder,bool isTiled)1008   void emitMatrixMultiply(MatrixTy &Result, const MatrixTy &A,
1009                           const MatrixTy &B, bool AllowContraction,
1010                           IRBuilder<> &Builder, bool isTiled) {
1011     const unsigned VF = std::max<unsigned>(
1012         TTI.getRegisterBitWidth(true) /
1013             Result.getElementType()->getPrimitiveSizeInBits().getFixedSize(),
1014         1U);
1015     unsigned R = Result.getNumRows();
1016     unsigned C = Result.getNumColumns();
1017     unsigned M = A.getNumColumns();
1018 
1019     bool IsFP = Result.getElementType()->isFloatingPointTy();
1020     assert(A.isColumnMajor() == B.isColumnMajor() &&
1021            Result.isColumnMajor() == A.isColumnMajor() &&
1022            "operands must agree on matrix layout");
1023     unsigned NumComputeOps = 0;
1024     if (A.isColumnMajor()) {
1025       // Multiply columns from the first operand with scalars from the second
1026       // operand. Then move along the K axes and accumulate the columns.  With
1027       // this the adds can be vectorized without reassociation.
1028       for (unsigned J = 0; J < C; ++J) {
1029         unsigned BlockSize = VF;
1030         // If Result is zero, we don't need to accumulate in the K==0 iteration.
1031         bool isSumZero = isa<ConstantAggregateZero>(Result.getColumn(J));
1032 
1033         for (unsigned I = 0; I < R; I += BlockSize) {
1034           // Gradually lower the vectorization factor to cover the remainder.
1035           while (I + BlockSize > R)
1036             BlockSize /= 2;
1037 
1038           Value *Sum = isTiled ? Result.extractVector(I, J, BlockSize, Builder)
1039                                : nullptr;
1040           for (unsigned K = 0; K < M; ++K) {
1041             Value *L = A.extractVector(I, K, BlockSize, Builder);
1042             Value *RH = Builder.CreateExtractElement(B.getColumn(J), K);
1043             Value *Splat = Builder.CreateVectorSplat(BlockSize, RH, "splat");
1044             Sum = createMulAdd(isSumZero && K == 0 ? nullptr : Sum, L, Splat,
1045                                Result.getElementType()->isFloatingPointTy(),
1046                                Builder, AllowContraction, NumComputeOps);
1047           }
1048           Result.setVector(J,
1049                            insertVector(Result.getVector(J), I, Sum, Builder));
1050         }
1051       }
1052     } else {
1053       // Multiply rows from the second operand with scalars from the first
1054       // operand. Then move along the K axes and accumulate the rows.  With this
1055       // the adds can be vectorized without reassociation.
1056       for (unsigned I = 0; I < R; ++I) {
1057         unsigned BlockSize = VF;
1058         bool isSumZero = isa<ConstantAggregateZero>(Result.getRow(I));
1059         for (unsigned J = 0; J < C; J += BlockSize) {
1060           // Gradually lower the vectorization factor to cover the remainder.
1061           while (J + BlockSize > C)
1062             BlockSize /= 2;
1063 
1064           Value *Sum = nullptr;
1065           for (unsigned K = 0; K < M; ++K) {
1066             Value *R = B.extractVector(K, J, BlockSize, Builder);
1067             Value *LH = Builder.CreateExtractElement(A.getVector(I), K);
1068             Value *Splat = Builder.CreateVectorSplat(BlockSize, LH, "splat");
1069             Sum = createMulAdd(isSumZero && K == 0 ? nullptr : Sum, Splat, R,
1070                                IsFP, Builder, AllowContraction, NumComputeOps);
1071           }
1072           Result.setVector(I,
1073                            insertVector(Result.getVector(I), J, Sum, Builder));
1074         }
1075       }
1076     }
1077     Result.addNumComputeOps(NumComputeOps);
1078   }
1079 
1080   /// Ensure that the memory in \p Load does not alias \p Store by potentially
1081   /// copying it to a new location.  This new or otherwise the original location
1082   /// is returned.
getNonAliasingPointer(LoadInst * Load,StoreInst * Store,CallInst * MatMul)1083   Value *getNonAliasingPointer(LoadInst *Load, StoreInst *Store,
1084                                CallInst *MatMul) {
1085     MemoryLocation StoreLoc = MemoryLocation::get(Store);
1086     MemoryLocation LoadLoc = MemoryLocation::get(Load);
1087 
1088     AliasResult LdAliased = AA.alias(LoadLoc, StoreLoc);
1089 
1090     // If we can statically determine noalias we're good.
1091     if (!LdAliased)
1092       return Load->getPointerOperand();
1093 
1094     // Create code to check if the memory locations of the Load and Store
1095     // overlap and if they do, copy Load's operand to a new buffer.
1096 
1097     // First, create  new blocks for 2n part of the check and the copy.
1098     BasicBlock *Check0 = MatMul->getParent();
1099     // FIXME: Use lazy DTU and update SplitBlock to accept a DTU instead of a
1100     // DT. Manually collect dominator tree updates, to avoid unnecessary work,
1101     // as we adjust Check0 and Check1's branches.
1102     SmallVector<DominatorTree::UpdateType, 4> DTUpdates;
1103     for (BasicBlock *Succ : successors(Check0))
1104       DTUpdates.push_back({DT.Delete, Check0, Succ});
1105 
1106     BasicBlock *Check1 = SplitBlock(MatMul->getParent(), MatMul, nullptr, &LI,
1107                                     nullptr, "alias_cont");
1108     BasicBlock *Copy =
1109         SplitBlock(MatMul->getParent(), MatMul, nullptr, &LI, nullptr, "copy");
1110     BasicBlock *Fusion = SplitBlock(MatMul->getParent(), MatMul, nullptr, &LI,
1111                                     nullptr, "no_alias");
1112 
1113     // Check if the loaded memory location begins before the end of the store
1114     // location. If the condition holds, they might overlap, otherwise they are
1115     // guaranteed to not overlap.
1116     IRBuilder<> Builder(MatMul);
1117     Check0->getTerminator()->eraseFromParent();
1118     Builder.SetInsertPoint(Check0);
1119     Type *IntPtrTy = Builder.getIntPtrTy(Load->getModule()->getDataLayout());
1120     Value *StoreBegin = Builder.CreatePtrToInt(
1121         const_cast<Value *>(StoreLoc.Ptr), IntPtrTy, "store.begin");
1122     Value *StoreEnd = Builder.CreateAdd(
1123         StoreBegin, ConstantInt::get(IntPtrTy, StoreLoc.Size.getValue()),
1124         "store.end", true, true);
1125     Value *LoadBegin = Builder.CreatePtrToInt(const_cast<Value *>(LoadLoc.Ptr),
1126                                               IntPtrTy, "load.begin");
1127     Builder.CreateCondBr(Builder.CreateICmpULT(LoadBegin, StoreEnd), Check1,
1128                          Fusion);
1129 
1130     // Check if the store begins before the end of the load location. If the
1131     // condition holds, they alias, otherwise they are guaranteed to not
1132     // overlap.
1133     Check1->getTerminator()->eraseFromParent();
1134     Builder.SetInsertPoint(Check1, Check1->begin());
1135     Value *LoadEnd = Builder.CreateAdd(
1136         LoadBegin, ConstantInt::get(IntPtrTy, LoadLoc.Size.getValue()),
1137         "load.end", true, true);
1138     Builder.CreateCondBr(Builder.CreateICmpULT(StoreBegin, LoadEnd), Copy,
1139                          Fusion);
1140 
1141     // Copy load operand to new alloca.
1142     Builder.SetInsertPoint(Copy, Copy->begin());
1143     AllocaInst *NewLd =
1144         Builder.CreateAlloca(Load->getType(), Load->getPointerAddressSpace());
1145     Builder.CreateMemCpy(NewLd, NewLd->getAlign(),
1146                          Load->getPointerOperand(), Load->getAlign(),
1147                          LoadLoc.Size.getValue());
1148     Builder.SetInsertPoint(Fusion, Fusion->begin());
1149     PHINode *PHI = Builder.CreatePHI(Load->getPointerOperandType(), 3);
1150     PHI->addIncoming(Load->getPointerOperand(), Check0);
1151     PHI->addIncoming(Load->getPointerOperand(), Check1);
1152     PHI->addIncoming(NewLd, Copy);
1153 
1154     // Adjust DT.
1155     DTUpdates.push_back({DT.Insert, Check0, Check1});
1156     DTUpdates.push_back({DT.Insert, Check0, Fusion});
1157     DTUpdates.push_back({DT.Insert, Check1, Copy});
1158     DTUpdates.push_back({DT.Insert, Check1, Fusion});
1159     DT.applyUpdates(DTUpdates);
1160     return PHI;
1161   }
1162 
isFusionProfitable(CallInst * MatMul)1163   bool isFusionProfitable(CallInst *MatMul) {
1164     if (ForceFusion)
1165       return true;
1166 
1167     ShapeInfo LShape(MatMul->getArgOperand(2), MatMul->getArgOperand(3));
1168     ShapeInfo RShape(MatMul->getArgOperand(3), MatMul->getArgOperand(4));
1169 
1170     const unsigned R = LShape.NumRows;
1171     const unsigned C = RShape.NumColumns;
1172     const unsigned M = LShape.NumColumns;
1173     auto *EltType = cast<VectorType>(MatMul->getType())->getElementType();
1174 
1175     const unsigned VF =
1176         std::max<unsigned>(TTI.getRegisterBitWidth(true) /
1177                                EltType->getPrimitiveSizeInBits().getFixedSize(),
1178                            1U);
1179 
1180     // Cost model for tiling
1181     //
1182     // For tiling to be beneficial, we need reuse either along the R or
1183     // the C axis.  We vectorize along the R axis so that means at least
1184     // 3 elements.
1185     // TODO: Also consider cost of copying if operands alias.
1186     if (R <= VF && C == 1)
1187       return false;
1188     // Then we need enough elements to exceed the number of vector
1189     // registers we have.  Note that this is an oversimplification since
1190     // fusing also takes some extra loads which may exceed the number of
1191     // reloads necessary.
1192     unsigned Op0Regs = (R + VF - 1) / VF * M;
1193     unsigned Op1Regs = (M + VF - 1) / VF * C;
1194     return Op0Regs + Op1Regs > TTI.getNumberOfRegisters(true);
1195   }
1196 
getZeroMatrix(Type * EltType,unsigned R,unsigned C)1197   MatrixTy getZeroMatrix(Type *EltType, unsigned R, unsigned C) {
1198     MatrixTy Res;
1199     auto *ColumType = FixedVectorType::get(EltType, R);
1200     for (unsigned I = 0; I < C; ++I)
1201       Res.addVector(ConstantAggregateZero::get(ColumType));
1202     return Res;
1203   }
1204 
emitSIMDTiling(CallInst * MatMul,LoadInst * LoadOp0,LoadInst * LoadOp1,StoreInst * Store,SmallPtrSetImpl<Instruction * > & FusedInsts)1205   void emitSIMDTiling(CallInst *MatMul, LoadInst *LoadOp0, LoadInst *LoadOp1,
1206                       StoreInst *Store,
1207                       SmallPtrSetImpl<Instruction *> &FusedInsts) {
1208     assert(MatrixLayout == MatrixLayoutTy::ColumnMajor &&
1209            "Tiling only supported for column-major matrixes at the moment!");
1210     if (!isFusionProfitable(MatMul))
1211       return;
1212 
1213     ShapeInfo LShape(MatMul->getArgOperand(2), MatMul->getArgOperand(3));
1214     ShapeInfo RShape(MatMul->getArgOperand(3), MatMul->getArgOperand(4));
1215 
1216     const unsigned R = LShape.NumRows;
1217     const unsigned C = RShape.NumColumns;
1218     const unsigned M = LShape.NumColumns;
1219     auto *EltType = cast<VectorType>(MatMul->getType())->getElementType();
1220 
1221     Value *APtr = getNonAliasingPointer(LoadOp0, Store, MatMul);
1222     Value *BPtr = getNonAliasingPointer(LoadOp1, Store, MatMul);
1223     Value *CPtr = Store->getPointerOperand();
1224 
1225     bool AllowContract = AllowContractEnabled || (isa<FPMathOperator>(MatMul) &&
1226                                                   MatMul->hasAllowContract());
1227     IRBuilder<> Builder(Store);
1228     for (unsigned J = 0; J < C; J += TileSize)
1229       for (unsigned I = 0; I < R; I += TileSize) {
1230         const unsigned TileR = std::min(R - I, unsigned(TileSize));
1231         const unsigned TileC = std::min(C - J, unsigned(TileSize));
1232         MatrixTy Res = getZeroMatrix(EltType, TileR, TileC);
1233 
1234         for (unsigned K = 0; K < M; K += TileSize) {
1235           const unsigned TileM = std::min(M - K, unsigned(TileSize));
1236           MatrixTy A =
1237               loadMatrix(APtr, LoadOp0->getAlign(), LoadOp0->isVolatile(),
1238                          LShape, Builder.getInt64(I), Builder.getInt64(K),
1239                          {TileR, TileM}, EltType, Builder);
1240           MatrixTy B =
1241               loadMatrix(BPtr, LoadOp1->getAlign(), LoadOp1->isVolatile(),
1242                          RShape, Builder.getInt64(K), Builder.getInt64(J),
1243                          {TileM, TileC}, EltType, Builder);
1244           emitMatrixMultiply(Res, A, B, AllowContract, Builder, true);
1245         }
1246         storeMatrix(Res, CPtr, Store->getAlign(), Store->isVolatile(), {R, M},
1247                     Builder.getInt64(I), Builder.getInt64(J), EltType, Builder);
1248       }
1249 
1250     // Mark eliminated instructions as fused and remove them.
1251     FusedInsts.insert(Store);
1252     FusedInsts.insert(MatMul);
1253     Store->eraseFromParent();
1254     MatMul->eraseFromParent();
1255     if (LoadOp0->hasNUses(0)) {
1256       FusedInsts.insert(LoadOp0);
1257       LoadOp0->eraseFromParent();
1258     }
1259     if (LoadOp1->hasNUses(0)) {
1260       FusedInsts.insert(LoadOp1);
1261       LoadOp1->eraseFromParent();
1262     }
1263   }
1264 
1265   /// Try to lower matrix multiply chains by fusing operations.
1266   ///
1267   /// Currently we only lower {ld, ld} -> matmul -> st chains.
1268   //
1269   /// No need to return a MatrixTy object for the result of the operation, since
1270   /// the single store user will be lowered as part of this. Instructions that
1271   /// are completely eliminated by fusion are added to \p FusedInsts.
LowerMatrixMultiplyFused(CallInst * MatMul,SmallPtrSetImpl<Instruction * > & FusedInsts)1272   void LowerMatrixMultiplyFused(CallInst *MatMul,
1273                                 SmallPtrSetImpl<Instruction *> &FusedInsts) {
1274     if (!FuseMatrix || !MatMul->hasOneUse() ||
1275         MatrixLayout != MatrixLayoutTy::ColumnMajor)
1276       return;
1277 
1278     auto *LoadOp0 = dyn_cast<LoadInst>(MatMul->getOperand(0));
1279     auto *LoadOp1 = dyn_cast<LoadInst>(MatMul->getOperand(1));
1280     auto *Store = dyn_cast<StoreInst>(*MatMul->user_begin());
1281     if (LoadOp0 && LoadOp1 && Store) {
1282       // The store address must dominate the MatMul instruction, otherwise
1283       // we create invalid IR.
1284       // FIXME: See if we can hoist the store address computation.
1285       auto *AddrI = dyn_cast<Instruction>(Store->getOperand(1));
1286       if (AddrI && (!DT.dominates(AddrI, MatMul)))
1287         return;
1288 
1289       emitSIMDTiling(MatMul, LoadOp0, LoadOp1, Store, FusedInsts);
1290       return;
1291     }
1292   }
1293 
1294   /// Lowers llvm.matrix.multiply.
LowerMultiply(CallInst * MatMul)1295   void LowerMultiply(CallInst *MatMul) {
1296     IRBuilder<> Builder(MatMul);
1297     auto *EltType = cast<VectorType>(MatMul->getType())->getElementType();
1298     ShapeInfo LShape(MatMul->getArgOperand(2), MatMul->getArgOperand(3));
1299     ShapeInfo RShape(MatMul->getArgOperand(3), MatMul->getArgOperand(4));
1300 
1301     const MatrixTy &Lhs = getMatrix(MatMul->getArgOperand(0), LShape, Builder);
1302     const MatrixTy &Rhs = getMatrix(MatMul->getArgOperand(1), RShape, Builder);
1303 
1304     const unsigned R = LShape.NumRows;
1305     const unsigned C = RShape.NumColumns;
1306     assert(LShape.NumColumns == RShape.NumRows);
1307 
1308     // Initialize the output
1309     MatrixTy Result(R, C, EltType);
1310 
1311     bool AllowContract = AllowContractEnabled || (isa<FPMathOperator>(MatMul) &&
1312                                                   MatMul->hasAllowContract());
1313     emitMatrixMultiply(Result, Lhs, Rhs, AllowContract, Builder, false);
1314     finalizeLowering(MatMul, Result, Builder);
1315   }
1316 
1317   /// Lowers llvm.matrix.transpose.
LowerTranspose(CallInst * Inst)1318   void LowerTranspose(CallInst *Inst) {
1319     MatrixTy Result;
1320     IRBuilder<> Builder(Inst);
1321     Value *InputVal = Inst->getArgOperand(0);
1322     VectorType *VectorTy = cast<VectorType>(InputVal->getType());
1323     ShapeInfo ArgShape(Inst->getArgOperand(1), Inst->getArgOperand(2));
1324     MatrixTy InputMatrix = getMatrix(InputVal, ArgShape, Builder);
1325 
1326     const unsigned NewNumVecs =
1327         InputMatrix.isColumnMajor() ? ArgShape.NumRows : ArgShape.NumColumns;
1328     const unsigned NewNumElts =
1329         InputMatrix.isColumnMajor() ? ArgShape.NumColumns : ArgShape.NumRows;
1330 
1331     for (unsigned I = 0; I < NewNumVecs; ++I) {
1332       // Build a single result vector. First initialize it.
1333       Value *ResultVector = UndefValue::get(
1334           FixedVectorType::get(VectorTy->getElementType(), NewNumElts));
1335       // Go through the old elements and insert it into the resulting vector.
1336       for (auto J : enumerate(InputMatrix.vectors())) {
1337         Value *Elt = Builder.CreateExtractElement(J.value(), I);
1338         // Row and column indices are transposed.
1339         ResultVector =
1340             Builder.CreateInsertElement(ResultVector, Elt, J.index());
1341       }
1342       Result.addVector(ResultVector);
1343     }
1344 
1345     // TODO: Improve estimate of operations needed for transposes. Currently we
1346     // just count the insertelement/extractelement instructions, but do not
1347     // account for later simplifications/combines.
1348     finalizeLowering(
1349         Inst,
1350         Result.addNumComputeOps(2 * ArgShape.NumRows * ArgShape.NumColumns),
1351         Builder);
1352   }
1353 
1354   /// Lower load instructions, if shape information is available.
VisitLoad(LoadInst * Inst,Value * Ptr,IRBuilder<> & Builder)1355   bool VisitLoad(LoadInst *Inst, Value *Ptr, IRBuilder<> &Builder) {
1356     auto I = ShapeMap.find(Inst);
1357     if (I == ShapeMap.end())
1358       return false;
1359 
1360     LowerLoad(Inst, Ptr, Inst->getAlign(),
1361               Builder.getInt64(I->second.getStride()), Inst->isVolatile(),
1362               I->second);
1363     return true;
1364   }
1365 
VisitStore(StoreInst * Inst,Value * StoredVal,Value * Ptr,IRBuilder<> & Builder)1366   bool VisitStore(StoreInst *Inst, Value *StoredVal, Value *Ptr,
1367                   IRBuilder<> &Builder) {
1368     auto I = ShapeMap.find(StoredVal);
1369     if (I == ShapeMap.end())
1370       return false;
1371 
1372     LowerStore(Inst, StoredVal, Ptr, Inst->getAlign(),
1373                Builder.getInt64(I->second.getStride()), Inst->isVolatile(),
1374                I->second);
1375     return true;
1376   }
1377 
1378   /// Lower binary operators, if shape information is available.
VisitBinaryOperator(BinaryOperator * Inst)1379   bool VisitBinaryOperator(BinaryOperator *Inst) {
1380     auto I = ShapeMap.find(Inst);
1381     if (I == ShapeMap.end())
1382       return false;
1383 
1384     Value *Lhs = Inst->getOperand(0);
1385     Value *Rhs = Inst->getOperand(1);
1386 
1387     IRBuilder<> Builder(Inst);
1388     ShapeInfo &Shape = I->second;
1389 
1390     MatrixTy Result;
1391     MatrixTy A = getMatrix(Lhs, Shape, Builder);
1392     MatrixTy B = getMatrix(Rhs, Shape, Builder);
1393     assert(A.isColumnMajor() == B.isColumnMajor() &&
1394            Result.isColumnMajor() == A.isColumnMajor() &&
1395            "operands must agree on matrix layout");
1396 
1397     // Helper to perform binary op on vectors.
1398     auto BuildVectorOp = [&Builder, Inst](Value *LHS, Value *RHS) {
1399       switch (Inst->getOpcode()) {
1400       case Instruction::Add:
1401         return Builder.CreateAdd(LHS, RHS);
1402       case Instruction::Mul:
1403         return Builder.CreateMul(LHS, RHS);
1404       case Instruction::Sub:
1405         return Builder.CreateSub(LHS, RHS);
1406       case Instruction::FAdd:
1407         return Builder.CreateFAdd(LHS, RHS);
1408       case Instruction::FMul:
1409         return Builder.CreateFMul(LHS, RHS);
1410       case Instruction::FSub:
1411         return Builder.CreateFSub(LHS, RHS);
1412       default:
1413         llvm_unreachable("Unsupported binary operator for matrix");
1414       }
1415     };
1416 
1417     for (unsigned I = 0; I < Shape.getNumVectors(); ++I)
1418       Result.addVector(BuildVectorOp(A.getVector(I), B.getVector(I)));
1419 
1420     finalizeLowering(Inst,
1421                      Result.addNumComputeOps(getNumOps(Result.getVectorTy()) *
1422                                              Result.getNumVectors()),
1423                      Builder);
1424     return true;
1425   }
1426 
1427   /// Helper to linearize a matrix expression tree into a string. Currently
1428   /// matrix expressions are linarized by starting at an expression leaf and
1429   /// linearizing bottom up.
1430   struct ExprLinearizer {
1431     unsigned LengthToBreak = 100;
1432     std::string Str;
1433     raw_string_ostream Stream;
1434     unsigned LineLength = 0;
1435     const DataLayout &DL;
1436 
1437     /// Mapping from instructions to matrixes. It is used to identify
1438     /// matrix instructions.
1439     const MapVector<Value *, MatrixTy> &Inst2Matrix;
1440 
1441     /// Mapping from values to the leaves of all expressions that the value is
1442     /// part of.
1443     const DenseMap<Value *, SmallPtrSet<Value *, 2>> &Shared;
1444 
1445     /// Set of matrix expressions in the scope of a given DISubprogram.
1446     const SmallSetVector<Value *, 32> &ExprsInSubprogram;
1447 
1448     /// Leaf node of the expression to linearize.
1449     Value *Leaf;
1450 
1451     /// Used to keep track of sub-expressions that get reused while linearizing
1452     /// the expression. Re-used sub-expressions are marked as (reused).
1453     SmallPtrSet<Value *, 8> ReusedExprs;
1454 
ExprLinearizer__anon643663df0111::LowerMatrixIntrinsics::ExprLinearizer1455     ExprLinearizer(const DataLayout &DL,
1456                    const MapVector<Value *, MatrixTy> &Inst2Matrix,
1457                    const DenseMap<Value *, SmallPtrSet<Value *, 2>> &Shared,
1458                    const SmallSetVector<Value *, 32> &ExprsInSubprogram,
1459                    Value *Leaf)
1460         : Str(), Stream(Str), DL(DL), Inst2Matrix(Inst2Matrix), Shared(Shared),
1461           ExprsInSubprogram(ExprsInSubprogram), Leaf(Leaf) {}
1462 
indent__anon643663df0111::LowerMatrixIntrinsics::ExprLinearizer1463     void indent(unsigned N) {
1464       LineLength += N;
1465       for (unsigned i = 0; i < N; i++)
1466         Stream << " ";
1467     }
1468 
lineBreak__anon643663df0111::LowerMatrixIntrinsics::ExprLinearizer1469     void lineBreak() {
1470       Stream << "\n";
1471       LineLength = 0;
1472     }
1473 
maybeIndent__anon643663df0111::LowerMatrixIntrinsics::ExprLinearizer1474     void maybeIndent(unsigned Indent) {
1475       if (LineLength >= LengthToBreak)
1476         lineBreak();
1477 
1478       if (LineLength == 0)
1479         indent(Indent);
1480     }
1481 
write__anon643663df0111::LowerMatrixIntrinsics::ExprLinearizer1482     void write(StringRef S) {
1483       LineLength += S.size();
1484       Stream << S;
1485     }
1486 
getUnderlyingObjectThroughLoads__anon643663df0111::LowerMatrixIntrinsics::ExprLinearizer1487     Value *getUnderlyingObjectThroughLoads(Value *V) {
1488       if (Value *Ptr = getPointerOperand(V))
1489         return getUnderlyingObjectThroughLoads(Ptr);
1490       else if (V->getType()->isPointerTy())
1491         return GetUnderlyingObject(V, DL);
1492       return V;
1493     }
1494 
1495     /// Returns true if \p V is a matrix value in the given subprogram.
isMatrix__anon643663df0111::LowerMatrixIntrinsics::ExprLinearizer1496     bool isMatrix(Value *V) const { return ExprsInSubprogram.count(V); }
1497 
1498     /// If \p V is a matrix value, print its shape as as NumRows x NumColumns to
1499     /// \p SS.
prettyPrintMatrixType__anon643663df0111::LowerMatrixIntrinsics::ExprLinearizer1500     void prettyPrintMatrixType(Value *V, raw_string_ostream &SS) {
1501       auto M = Inst2Matrix.find(V);
1502       if (M == Inst2Matrix.end())
1503         SS << "unknown";
1504       else {
1505         SS << M->second.getNumRows();
1506         SS << "x";
1507         SS << M->second.getNumColumns();
1508       }
1509     }
1510 
1511     /// Write the called function name. Handles calls to llvm.matrix.*
1512     /// specially: we write the name, followed by the dimensions of the input
1513     /// matrixes, followed by the scalar type name.
writeFnName__anon643663df0111::LowerMatrixIntrinsics::ExprLinearizer1514     void writeFnName(CallInst *CI) {
1515       if (!CI->getCalledFunction())
1516         write("<no called fn>");
1517       else {
1518         StringRef Name = CI->getCalledFunction()->getName();
1519         if (!Name.startswith("llvm.matrix")) {
1520           write(Name);
1521           return;
1522         }
1523         IntrinsicInst *II = dyn_cast<IntrinsicInst>(CI);
1524         write(StringRef(Intrinsic::getName(II->getIntrinsicID(), {}))
1525                   .drop_front(StringRef("llvm.matrix.").size()));
1526         write(".");
1527         std::string Tmp = "";
1528         raw_string_ostream SS(Tmp);
1529 
1530         switch (II->getIntrinsicID()) {
1531         case Intrinsic::matrix_multiply:
1532           prettyPrintMatrixType(II->getOperand(0), SS);
1533           SS << ".";
1534           prettyPrintMatrixType(II->getOperand(1), SS);
1535           SS << "." << *II->getType()->getScalarType();
1536           break;
1537         case Intrinsic::matrix_transpose:
1538           prettyPrintMatrixType(II->getOperand(0), SS);
1539           SS << "." << *II->getType()->getScalarType();
1540           break;
1541         case Intrinsic::matrix_column_major_load:
1542           prettyPrintMatrixType(II, SS);
1543           SS << "." << *II->getType()->getScalarType();
1544           break;
1545         case Intrinsic::matrix_column_major_store:
1546           prettyPrintMatrixType(II->getOperand(0), SS);
1547           SS << "." << *II->getOperand(0)->getType()->getScalarType();
1548           break;
1549         default:
1550           llvm_unreachable("Unhandled case");
1551         }
1552         SS.flush();
1553         write(Tmp);
1554       }
1555     }
1556 
getNumShapeArgs__anon643663df0111::LowerMatrixIntrinsics::ExprLinearizer1557     unsigned getNumShapeArgs(CallInst *CI) const {
1558       if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(CI)) {
1559         switch (II->getIntrinsicID()) {
1560         case Intrinsic::matrix_multiply:
1561           return 3;
1562         case Intrinsic::matrix_transpose:
1563           return 2;
1564         case Intrinsic::matrix_column_major_load:
1565         case Intrinsic::matrix_column_major_store:
1566           return 3;
1567         default:
1568           return 0;
1569         }
1570       }
1571       return 0;
1572     }
1573 
1574     /// Special printing for values: for pointers, we print if they refer to an
1575     /// (function) external address or a stack address, for other values we
1576     /// either print the constant or "scalar"/"matrix" for other values.
write__anon643663df0111::LowerMatrixIntrinsics::ExprLinearizer1577     void write(Value *V) {
1578       V = getUnderlyingObjectThroughLoads(V);
1579       if (V->getType()->isPointerTy()) {
1580         if (isa<AllocaInst>(V)) {
1581           Stream << "stack addr";
1582           LineLength += StringRef("stack addr").size();
1583         } else {
1584           Stream << "addr";
1585           LineLength += StringRef("addr").size();
1586         }
1587         if (!V->getName().empty()) {
1588           Stream << " %" << V->getName() << "";
1589           LineLength += V->getName().size() + 2;
1590         }
1591         return;
1592       }
1593 
1594       std::string Tmp;
1595       raw_string_ostream TmpStream(Tmp);
1596 
1597       if (auto *CI = dyn_cast<ConstantInt>(V))
1598         TmpStream << CI->getValue();
1599       else if (isa<Constant>(V))
1600         TmpStream << "constant";
1601       else {
1602         if (isMatrix(V))
1603           TmpStream << "matrix";
1604         else
1605           TmpStream << "scalar";
1606       }
1607       TmpStream.flush();
1608       Tmp = std::string(StringRef(Tmp).trim());
1609       LineLength += Tmp.size();
1610       Stream << Tmp;
1611     }
1612 
1613     /// Linearize expression \p Expr starting at an indentation of \p Indent.
1614     /// Expressions that are re-used multiple times are prefixed with (reused)
1615     /// at the re-used root instruction.
linearizeExpr__anon643663df0111::LowerMatrixIntrinsics::ExprLinearizer1616     void linearizeExpr(Value *Expr, unsigned Indent, bool ParentReused,
1617                        bool ParentShared) {
1618       auto *I = cast<Instruction>(Expr);
1619       maybeIndent(Indent);
1620       SmallVector<Value *, 8> Ops;
1621 
1622       // Is Expr shared with other expression leaves?
1623       bool ExprShared = false;
1624 
1625       // Deal with shared subtrees. Mark them as shared, if required.
1626       if (!ParentShared) {
1627         auto SI = Shared.find(Expr);
1628         assert(SI != Shared.end() && SI->second.count(Leaf));
1629 
1630         for (Value *S : SI->second) {
1631           if (S == Leaf)
1632             continue;
1633           DebugLoc DL = cast<Instruction>(S)->getDebugLoc();
1634           write("shared with remark at line " + std::to_string(DL.getLine()) +
1635                 " column " + std::to_string(DL.getCol()) + " (");
1636         }
1637         ExprShared = SI->second.size() > 1;
1638       }
1639 
1640       bool Reused = !ReusedExprs.insert(Expr).second;
1641       if (Reused && !ParentReused)
1642         write("(reused) ");
1643 
1644       if (auto *CI = dyn_cast<CallInst>(I)) {
1645         writeFnName(CI);
1646 
1647         Ops.append(CI->arg_begin(), CI->arg_end() - getNumShapeArgs(CI));
1648       } else if (isa<BitCastInst>(Expr)) {
1649         // Special case bitcasts, which are used to materialize matrixes from
1650         // non-matrix ops.
1651         write("matrix");
1652         return;
1653       } else {
1654         Ops.append(I->value_op_begin(), I->value_op_end());
1655         write(std::string(I->getOpcodeName()));
1656       }
1657 
1658       write(std::string("("));
1659 
1660       unsigned NumOpsToBreak = 1;
1661       if (match(Expr, m_Intrinsic<Intrinsic::matrix_column_major_load>()))
1662         NumOpsToBreak = 2;
1663 
1664       for (Value *Op : Ops) {
1665         if (Ops.size() > NumOpsToBreak)
1666           lineBreak();
1667 
1668         maybeIndent(Indent + 1);
1669         if (isMatrix(Op))
1670           linearizeExpr(Op, Indent + 1, Reused, ExprShared);
1671         else
1672           write(Op);
1673         if (Op != Ops.back())
1674           write(", ");
1675       }
1676 
1677       write(")");
1678     }
1679 
getResult__anon643663df0111::LowerMatrixIntrinsics::ExprLinearizer1680     const std::string &getResult() {
1681       Stream.flush();
1682       return Str;
1683     }
1684   };
1685 
1686   /// Generate remarks for matrix operations in a function. To generate remarks
1687   /// for matrix expressions, the following approach is used:
1688   /// 1. Use the inlined-at debug information to group matrix operations to the
1689   ///    DISubprograms they are contained in.
1690   /// 2. Collect leaves of matrix expressions (done in
1691   ///    RemarkGenerator::getExpressionLeaves) for each subprogram - expression
1692   //     mapping.  Leaves are lowered matrix instructions without other matrix
1693   //     users (like stores) in the current subprogram.
1694   /// 3. For each leaf, create a remark containing a linearizied version of the
1695   ///    matrix expression. The expression is linearized by a recursive
1696   ///    bottom-up traversal of the matrix operands, starting at a leaf. Note
1697   ///    that multiple leaves can share sub-expressions. Shared subexpressions
1698   ///    are explicitly marked as shared().
1699   struct RemarkGenerator {
1700     const MapVector<Value *, MatrixTy> &Inst2Matrix;
1701     OptimizationRemarkEmitter &ORE;
1702     Function &Func;
1703     const DataLayout &DL;
1704 
RemarkGenerator__anon643663df0111::LowerMatrixIntrinsics::RemarkGenerator1705     RemarkGenerator(const MapVector<Value *, MatrixTy> &Inst2Matrix,
1706                     OptimizationRemarkEmitter &ORE, Function &Func)
1707         : Inst2Matrix(Inst2Matrix), ORE(ORE), Func(Func),
1708           DL(Func.getParent()->getDataLayout()) {}
1709 
1710     /// Return all leaves of the expressions in \p ExprsInSubprogram. Those are
1711     /// instructions in Inst2Matrix returning void or without any users in
1712     /// \p ExprsInSubprogram. Currently that should only include stores.
1713     SmallVector<Value *, 4>
getExpressionLeaves__anon643663df0111::LowerMatrixIntrinsics::RemarkGenerator1714     getExpressionLeaves(const SmallSetVector<Value *, 32> &ExprsInSubprogram) {
1715       SmallVector<Value *, 4> Leaves;
1716       for (auto *Expr : ExprsInSubprogram)
1717         if (Expr->getType()->isVoidTy() ||
1718             !any_of(Expr->users(), [&ExprsInSubprogram](User *U) {
1719               return ExprsInSubprogram.count(U);
1720             }))
1721           Leaves.push_back(Expr);
1722       return Leaves;
1723     }
1724 
1725     /// Recursively traverse expression \p V starting at \p Leaf and add \p Leaf
1726     /// to all visited expressions in \p Shared. Limit the matrix operations to
1727     /// the ones in \p ExprsInSubprogram.
collectSharedInfo__anon643663df0111::LowerMatrixIntrinsics::RemarkGenerator1728     void collectSharedInfo(Value *Leaf, Value *V,
1729                            const SmallSetVector<Value *, 32> &ExprsInSubprogram,
1730                            DenseMap<Value *, SmallPtrSet<Value *, 2>> &Shared) {
1731 
1732       if (!ExprsInSubprogram.count(V))
1733         return;
1734 
1735       auto I = Shared.insert({V, {}});
1736       I.first->second.insert(Leaf);
1737 
1738       for (Value *Op : cast<Instruction>(V)->operand_values())
1739         collectSharedInfo(Leaf, Op, ExprsInSubprogram, Shared);
1740       return;
1741     }
1742 
1743     /// Calculate the number of exclusive and shared op counts for expression
1744     /// starting at \p V. Expressions used multiple times are counted once.
1745     /// Limit the matrix operations to the ones in \p ExprsInSubprogram.
1746     std::pair<OpInfoTy, OpInfoTy>
sumOpInfos__anon643663df0111::LowerMatrixIntrinsics::RemarkGenerator1747     sumOpInfos(Value *Root, SmallPtrSetImpl<Value *> &ReusedExprs,
1748                const SmallSetVector<Value *, 32> &ExprsInSubprogram,
1749                DenseMap<Value *, SmallPtrSet<Value *, 2>> &Shared) const {
1750       if (!ExprsInSubprogram.count(Root))
1751         return {};
1752 
1753       // Already counted this expression. Stop.
1754       if (!ReusedExprs.insert(Root).second)
1755         return {};
1756 
1757       OpInfoTy SharedCount;
1758       OpInfoTy Count;
1759 
1760       auto I = Shared.find(Root);
1761       auto CM = Inst2Matrix.find(Root);
1762       if (I->second.size() == 1)
1763         Count = CM->second.getOpInfo();
1764       else
1765         SharedCount = CM->second.getOpInfo();
1766 
1767       for (Value *Op : cast<Instruction>(Root)->operand_values()) {
1768         auto C = sumOpInfos(Op, ReusedExprs, ExprsInSubprogram, Shared);
1769         Count += C.first;
1770         SharedCount += C.second;
1771       }
1772       return {Count, SharedCount};
1773     }
1774 
emitRemarks__anon643663df0111::LowerMatrixIntrinsics::RemarkGenerator1775     void emitRemarks() {
1776       if (!ORE.allowExtraAnalysis(DEBUG_TYPE))
1777         return;
1778 
1779       // Map matrix operations to their containting subprograms, by traversing
1780       // the inlinedAt chain. If the function does not have a DISubprogram, we
1781       // only map them to the containing function.
1782       MapVector<DISubprogram *, SmallVector<Value *, 8>> Subprog2Exprs;
1783       for (auto &KV : Inst2Matrix) {
1784         if (Func.getSubprogram()) {
1785           auto *I = cast<Instruction>(KV.first);
1786           DILocation *Context = I->getDebugLoc();
1787           while (Context) {
1788             auto I =
1789                 Subprog2Exprs.insert({getSubprogram(Context->getScope()), {}});
1790             I.first->second.push_back(KV.first);
1791             Context = DebugLoc(Context).getInlinedAt();
1792           }
1793         } else {
1794           auto I = Subprog2Exprs.insert({nullptr, {}});
1795           I.first->second.push_back(KV.first);
1796         }
1797       }
1798       for (auto &KV : Subprog2Exprs) {
1799         SmallSetVector<Value *, 32> ExprsInSubprogram(KV.second.begin(),
1800                                                       KV.second.end());
1801         auto Leaves = getExpressionLeaves(ExprsInSubprogram);
1802 
1803         DenseMap<Value *, SmallPtrSet<Value *, 2>> Shared;
1804         for (Value *Leaf : Leaves)
1805           collectSharedInfo(Leaf, Leaf, ExprsInSubprogram, Shared);
1806 
1807         // Generate remarks for each leaf.
1808         for (auto *L : Leaves) {
1809 
1810           DebugLoc Loc = cast<Instruction>(L)->getDebugLoc();
1811           DILocation *Context = cast<Instruction>(L)->getDebugLoc();
1812           while (Context) {
1813             if (getSubprogram(Context->getScope()) == KV.first) {
1814               Loc = Context;
1815               break;
1816             }
1817             Context = DebugLoc(Context).getInlinedAt();
1818           }
1819 
1820           SmallPtrSet<Value *, 8> ReusedExprs;
1821           OpInfoTy Counts, SharedCounts;
1822           std::tie(Counts, SharedCounts) =
1823               sumOpInfos(L, ReusedExprs, ExprsInSubprogram, Shared);
1824 
1825           OptimizationRemark Rem(DEBUG_TYPE, "matrix-lowered", Loc,
1826                                  cast<Instruction>(L)->getParent());
1827 
1828           Rem << "Lowered with ";
1829           Rem << ore::NV("NumStores", Counts.NumStores) << " stores, "
1830               << ore::NV("NumLoads", Counts.NumLoads) << " loads, "
1831               << ore::NV("NumComputeOps", Counts.NumComputeOps)
1832               << " compute ops";
1833 
1834           if (SharedCounts.NumStores > 0 || SharedCounts.NumLoads > 0 ||
1835               SharedCounts.NumComputeOps > 0) {
1836             Rem << ",\nadditionally "
1837                 << ore::NV("NumStores", SharedCounts.NumStores) << " stores, "
1838                 << ore::NV("NumLoads", SharedCounts.NumLoads) << " loads, "
1839                 << ore::NV("NumFPOps", SharedCounts.NumComputeOps)
1840                 << " compute ops"
1841                 << " are shared with other expressions";
1842           }
1843 
1844           Rem << ("\n" + linearize(L, Shared, ExprsInSubprogram, DL));
1845           ORE.emit(Rem);
1846         }
1847       }
1848     }
1849 
1850     std::string
linearize__anon643663df0111::LowerMatrixIntrinsics::RemarkGenerator1851     linearize(Value *L,
1852               const DenseMap<Value *, SmallPtrSet<Value *, 2>> &Shared,
1853               const SmallSetVector<Value *, 32> &ExprsInSubprogram,
1854               const DataLayout &DL) {
1855       ExprLinearizer Lin(DL, Inst2Matrix, Shared, ExprsInSubprogram, L);
1856       Lin.linearizeExpr(L, 0, false, false);
1857       return Lin.getResult();
1858     }
1859   };
1860 };
1861 } // namespace
1862 
run(Function & F,FunctionAnalysisManager & AM)1863 PreservedAnalyses LowerMatrixIntrinsicsPass::run(Function &F,
1864                                                  FunctionAnalysisManager &AM) {
1865   auto &TTI = AM.getResult<TargetIRAnalysis>(F);
1866   auto &ORE = AM.getResult<OptimizationRemarkEmitterAnalysis>(F);
1867   auto &AA = AM.getResult<AAManager>(F);
1868   auto &DT = AM.getResult<DominatorTreeAnalysis>(F);
1869   auto &LI = AM.getResult<LoopAnalysis>(F);
1870 
1871   LowerMatrixIntrinsics LMT(F, TTI, AA, DT, LI, ORE);
1872   if (LMT.Visit()) {
1873     PreservedAnalyses PA;
1874     PA.preserveSet<CFGAnalyses>();
1875     return PA;
1876   }
1877   return PreservedAnalyses::all();
1878 }
1879 
1880 namespace {
1881 
1882 class LowerMatrixIntrinsicsLegacyPass : public FunctionPass {
1883 public:
1884   static char ID;
1885 
LowerMatrixIntrinsicsLegacyPass()1886   LowerMatrixIntrinsicsLegacyPass() : FunctionPass(ID) {
1887     initializeLowerMatrixIntrinsicsLegacyPassPass(
1888         *PassRegistry::getPassRegistry());
1889   }
1890 
runOnFunction(Function & F)1891   bool runOnFunction(Function &F) override {
1892     auto &TTI = getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
1893     auto &ORE = getAnalysis<OptimizationRemarkEmitterWrapperPass>().getORE();
1894     auto &AA = getAnalysis<AAResultsWrapperPass>().getAAResults();
1895     auto &DT = getAnalysis<DominatorTreeWrapperPass>().getDomTree();
1896     auto &LI = getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
1897     LowerMatrixIntrinsics LMT(F, TTI, AA, DT, LI, ORE);
1898     bool C = LMT.Visit();
1899     return C;
1900   }
1901 
getAnalysisUsage(AnalysisUsage & AU) const1902   void getAnalysisUsage(AnalysisUsage &AU) const override {
1903     AU.addRequired<TargetTransformInfoWrapperPass>();
1904     AU.addRequired<OptimizationRemarkEmitterWrapperPass>();
1905     AU.addRequired<AAResultsWrapperPass>();
1906     AU.addRequired<DominatorTreeWrapperPass>();
1907     AU.addPreserved<DominatorTreeWrapperPass>();
1908     AU.addRequired<LoopInfoWrapperPass>();
1909     AU.addPreserved<LoopInfoWrapperPass>();
1910   }
1911 };
1912 } // namespace
1913 
1914 static const char pass_name[] = "Lower the matrix intrinsics";
1915 char LowerMatrixIntrinsicsLegacyPass::ID = 0;
INITIALIZE_PASS_BEGIN(LowerMatrixIntrinsicsLegacyPass,DEBUG_TYPE,pass_name,false,false)1916 INITIALIZE_PASS_BEGIN(LowerMatrixIntrinsicsLegacyPass, DEBUG_TYPE, pass_name,
1917                       false, false)
1918 INITIALIZE_PASS_DEPENDENCY(OptimizationRemarkEmitterWrapperPass)
1919 INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass)
1920 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
1921 INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)
1922 INITIALIZE_PASS_END(LowerMatrixIntrinsicsLegacyPass, DEBUG_TYPE, pass_name,
1923                     false, false)
1924 
1925 Pass *llvm::createLowerMatrixIntrinsicsPass() {
1926   return new LowerMatrixIntrinsicsLegacyPass();
1927 }
1928