1 //===- LinalgTransforms.cpp - Linalg transformations as patterns ----------===//
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 // This file implements logic and helpers to expose Linalg transforms as rewrite
10 // patterns.
11 //
12 //===----------------------------------------------------------------------===//
13 
14 #include "mlir/Dialect/Linalg/Transforms/Transforms.h"
15 #include "mlir/Dialect/Affine/Utils.h"
16 #include "mlir/Dialect/Linalg/Analysis/DependenceAnalysis.h"
17 #include "mlir/Dialect/Linalg/IR/LinalgOps.h"
18 #include "mlir/Dialect/Linalg/Utils/Utils.h"
19 #include "mlir/Dialect/StandardOps/EDSC/Intrinsics.h"
20 #include "mlir/Dialect/Utils/StructuredOpsUtils.h"
21 #include "mlir/Dialect/Vector/EDSC/Intrinsics.h"
22 #include "mlir/Dialect/Vector/VectorOps.h"
23 #include "mlir/IR/AffineExpr.h"
24 #include "mlir/IR/Matchers.h"
25 #include "mlir/Pass/Pass.h"
26 #include "mlir/Support/LLVM.h"
27 #include "mlir/Transforms/GreedyPatternRewriteDriver.h"
28 #include "llvm/ADT/ScopeExit.h"
29 #include "llvm/Support/Debug.h"
30 #include "llvm/Support/raw_ostream.h"
31 #include <type_traits>
32 
33 #define DEBUG_TYPE "linalg-transforms"
34 
35 using namespace mlir;
36 using namespace mlir::edsc;
37 using namespace mlir::edsc::intrinsics;
38 using namespace mlir::linalg;
39 
40 #define DBGS() (llvm::dbgs() << "[" DEBUG_TYPE << "]: ")
41 
42 //===----------------------------------------------------------------------===//
43 // Transformations exposed as rewrite patterns.
44 //===----------------------------------------------------------------------===//
45 // Marker used as attribute name in generated Linalg rewriting transformations.
46 const StringLiteral mlir::linalg::LinalgTransforms::kLinalgTransformMarker =
47     "__internal_linalg_transform__";
48 
49 mlir::linalg::LinalgTransformationFilter::LinalgTransformationFilter(
50     ArrayRef<Identifier> matchDisjunction, Optional<Identifier> replacement)
51     : LinalgTransformationFilter([](Operation *) { return success(); },
52                                  matchDisjunction, replacement) {}
53 
54 mlir::linalg::LinalgTransformationFilter::LinalgTransformationFilter(
55     FilterFunction f, ArrayRef<Identifier> matchDisjunction,
56     Optional<Identifier> replacement)
57     : filter(f),
58       matchDisjunction(matchDisjunction.begin(), matchDisjunction.end()),
59       replacement(replacement) {}
60 
61 LogicalResult mlir::linalg::LinalgTransformationFilter::checkAndNotify(
62     PatternRewriter &rewriter, Operation *op) const {
63   if (filter && failed(filter(op)))
64     return failure();
65 
66   auto attr = op->template getAttrOfType<StringAttr>(
67       LinalgTransforms::kLinalgTransformMarker);
68 
69   if (!attr) {
70     // 1. Has no marker case and matchDisjunction is empty.
71     if (matchDisjunction.empty())
72       return success();
73 
74     // 2. Has no marker but was expecting a marker.
75     return rewriter.notifyMatchFailure(op, [&](Diagnostic &diag) {
76       diag << " does not have any marker from list: ";
77       interleaveComma(matchDisjunction, diag);
78     });
79   }
80 
81   // 4. Match explicit marker.
82   for (auto marker : matchDisjunction)
83     if (attr.getValue() == marker)
84       return success();
85 
86   // 5. Fail to match.
87   return rewriter.notifyMatchFailure(op, [&](Diagnostic &diag) {
88     diag << " does not have any marker from list: ";
89     interleaveComma(matchDisjunction, diag);
90   });
91 }
92 
93 void mlir::linalg::LinalgTransformationFilter::
94     replaceLinalgTransformationFilter(PatternRewriter &rewriter,
95                                       Operation *op) const {
96   if (replacement.hasValue())
97     op->setAttr(LinalgTransforms::kLinalgTransformMarker,
98                 rewriter.getStringAttr(replacement.getValue()));
99   else
100     op->removeAttr(Identifier::get(LinalgTransforms::kLinalgTransformMarker,
101                                    rewriter.getContext()));
102 }
103 
104 LinalgTilingOptions &
105 mlir::linalg::LinalgTilingOptions::setTileSizes(ArrayRef<int64_t> ts) {
106   SmallVector<int64_t, 4> tileSizes(ts.begin(), ts.end());
107   tileSizeComputationFunction = [tileSizes](OpBuilder &b, Operation *op) {
108     OpBuilder::InsertionGuard guard(b);
109     b.setInsertionPointToStart(
110         &op->getParentOfType<FuncOp>().getBody().front());
111     return llvm::to_vector<4>(map_range(tileSizes, [&](int64_t s) {
112       Value v = b.create<ConstantIndexOp>(op->getLoc(), s);
113       return v;
114     }));
115   };
116   return *this;
117 }
118 
119 /// Try to compute a static bounding box for `operand`
120 /// Return success if either:
121 ///   1. The operand is already statically shaped, `result` is left unchanged.
122 ///   2. The operand is (partially) dynamic, `result` is the result of a freshly
123 ///      created PadTensorOp.
124 /// Return failure if the operand cannot be padded to a static shape.
125 static LogicalResult padOperandToSmallestStaticBoundingBox(
126     PatternRewriter &rewriter, linalg::LinalgOp opToPad, Value operand,
127     const LinalgTilingOptions &options, Value &result) {
128   auto tensorType = operand.getType().cast<RankedTensorType>();
129   // Already static shape, no need to pad.
130   if (tensorType.hasStaticShape())
131     return success();
132   auto subtensor = operand.getDefiningOp<SubTensorOp>();
133   // Not a subtensor, cannot construct a static bounding box.
134   if (!subtensor)
135     return failure();
136   SmallVector<int64_t> staticSizes;
137   staticSizes.reserve(tensorType.getRank());
138   auto shapedOp =
139       cast<OffsetSizeAndStrideOpInterface>(subtensor.getOperation());
140   for (auto size : shapedOp.getMixedSizes()) {
141     auto indexAttr = size.is<Attribute>()
142                          ? size.get<Attribute>().dyn_cast<IntegerAttr>()
143                          : linalg::getSmallestBoundingIndex(size.get<Value>());
144     // SmallestBoundingIndex must exist for all sizes.
145     // For now return an error if we can't find it.
146     if (!indexAttr)
147       return rewriter.notifyMatchFailure(
148           opToPad, "No constant bounding box can be found for padding");
149     staticSizes.push_back(indexAttr.getInt());
150   }
151   Value pad = options.paddingValueComputationFunction(rewriter, opToPad);
152   auto staticTensorType =
153       RankedTensorType::get(staticSizes, tensorType.getElementType());
154   result = linalg::PadTensorOp::createPadHighOp(staticTensorType, operand, pad,
155                                                 opToPad->getLoc(), rewriter);
156   return success();
157 }
158 
159 // Try to create a static bounding box around each operand of `res.op`.
160 // If successful, `res.op` is rewritten in static form with padded operands.
161 // `res.op` is updated to the cloned static form of the op on success.
162 static LogicalResult rewriteAsPaddedOp(PatternRewriter &rewriter,
163                                        TiledLinalgOp &res,
164                                        const LinalgTilingOptions &options) {
165   LinalgOp opToPad = res.op;
166   Location loc = opToPad->getLoc();
167 
168   // If the op is fully static, it does not need padding.
169   // TODO: there are cases where we may still want to pad to larger sizes.
170   if (llvm::all_of(opToPad.getShapedOperands(), [](Value v) {
171         return v.getType().cast<RankedTensorType>().hasStaticShape();
172       }))
173     return success();
174 
175   OpBuilder::InsertionGuard g(rewriter);
176   // Set IP after op because we also take the dims of the original output.
177   rewriter.setInsertionPointAfter(opToPad);
178   // Make a copy of the shaped operands and update it.
179   SmallVector<Value> operands = opToPad.getShapedOperands();
180   for (Value &v : operands) {
181     Value paddedOperand;
182     // If padding was requested but the shape cannot be bounded statically then
183     // the pattern fails to apply.
184     if (failed(padOperandToSmallestStaticBoundingBox(rewriter, opToPad, v,
185                                                      options, paddedOperand))) {
186       return failure();
187     }
188     // Update v if we indeed got a padded operand.
189     v = paddedOperand ? paddedOperand : v;
190   }
191 
192   // Clone `opToPad` to operate on the statically padded shapes.
193   auto resultTensorTypes =
194       ValueRange(operands).take_back(opToPad.getNumOutputs()).getTypes();
195   ValueRange otherOperands = opToPad.getAssumedNonShapedOperands();
196   operands.append(otherOperands.begin(), otherOperands.end());
197   linalg::LinalgOp paddedOp =
198       opToPad.clone(rewriter, loc, resultTensorTypes, operands);
199 
200   // Recover the subtensor out of the new static results. This keeps the
201   // original linalg op around because it uses the dims of the original results.
202   // This later folds away.
203   SmallVector<Value> paddedSubviewResults;
204   paddedSubviewResults.reserve(opToPad->getNumResults());
205   llvm::SetVector<Operation *> newUsersOfOpToPad;
206   for (auto it : llvm::zip(opToPad->getResults(), paddedOp->getResults())) {
207     auto rank = std::get<0>(it).getType().cast<RankedTensorType>().getRank();
208     SmallVector<OpFoldResult> offsets(rank, rewriter.getIndexAttr(0));
209     auto sizes = llvm::to_vector<4>(llvm::map_range(
210         llvm::seq<unsigned>(0, rank), [&](unsigned d) -> OpFoldResult {
211           auto dimOp = rewriter.create<DimOp>(loc, std::get<0>(it), d);
212           newUsersOfOpToPad.insert(dimOp);
213           return dimOp.getResult();
214         }));
215     SmallVector<OpFoldResult> strides(rank, rewriter.getIndexAttr(1));
216     paddedSubviewResults.push_back(rewriter.create<SubTensorOp>(
217         loc, std::get<1>(it), offsets, sizes, strides));
218   }
219   // Replace the transient `opToPad` locally, except for uses that we just
220   // created for the purpose of extracting the dims.
221   rewriter.replaceOpWithIf(opToPad, paddedSubviewResults, [&](OpOperand &opOp) {
222     return !newUsersOfOpToPad.contains(opOp.getOwner());
223   });
224 
225   res = TiledLinalgOp{paddedOp, res.loops, res.tensorResults};
226   return success();
227 }
228 
229 /// Linalg base tiling pattern.
230 mlir::linalg::LinalgBaseTilingPattern::LinalgBaseTilingPattern(
231     StringRef opName, MLIRContext *context, LinalgTilingOptions options,
232     LinalgTransformationFilter marker, PatternBenefit benefit)
233     : RewritePattern(opName, {}, benefit, context), marker(marker),
234       options(options) {}
235 
236 mlir::linalg::LinalgBaseTilingPattern::LinalgBaseTilingPattern(
237     LinalgTilingOptions options, LinalgTransformationFilter marker,
238     PatternBenefit benefit)
239     : RewritePattern(benefit, MatchAnyOpTypeTag()), marker(marker),
240       options(options) {}
241 
242 LogicalResult mlir::linalg::LinalgBaseTilingPattern::matchAndRewriteBase(
243     Operation *op, PatternRewriter &rewriter, TiledLinalgOp &result) const {
244   LinalgOp linalgOp = dyn_cast<LinalgOp>(op);
245   if (!linalgOp)
246     return failure();
247   if (failed(marker.checkAndNotify(rewriter, linalgOp)))
248     return failure();
249 
250   Optional<TiledLinalgOp> res = tileLinalgOp(rewriter, linalgOp, options);
251 
252   if (!res)
253     return failure();
254 
255   // Setup RAII guard to return properly.
256   bool succeeded = true;
257   LinalgOp tiledOp = res->op;
258   auto guard = llvm::make_scope_exit([&]() {
259     if (!succeeded)
260       return;
261     // Return relevant information to derived pattern.
262     result = *res;
263     // Replace marker on both tiledOp and tiledAndPaddedOp, if necessary.
264     marker.replaceLinalgTransformationFilter(rewriter, tiledOp);
265     if (tiledOp != res->op)
266       marker.replaceLinalgTransformationFilter(rewriter, res->op);
267   });
268 
269   // Consider padding on the fly only if the op has tensor semantics.
270   if (!options.paddingValueComputationFunction ||
271       !linalgOp.hasTensorSemantics())
272     return success();
273 
274   // Try to pad on the fly by rewriting res->op as a padded op.
275   if (failed(rewriteAsPaddedOp(rewriter, *res, options))) {
276     // Set so RAII guard does not propagate TiledLinalgOp to `result`.
277     succeeded = false;
278     return failure();
279   }
280 
281   // Do not perform replacement of `linalgOp`, let the derived patterns
282   // do this as they see fit, from the resulting TiledLinalgOp.
283   return success();
284 }
285 
286 mlir::linalg::LinalgBaseTileAndFusePattern::LinalgBaseTileAndFusePattern(
287     StringRef opName, MLIRContext *context,
288     const LinalgDependenceGraph &dependenceGraph,
289     LinalgTilingOptions tilingOptions, LinalgFusionOptions fusionOptions,
290     LinalgTransformationFilter marker, LinalgTransformationFilter fusedOpMarker,
291     LinalgTransformationFilter originalOpMarker, PatternBenefit benefit)
292     : RewritePattern(opName, {}, benefit, context),
293       dependenceGraph(dependenceGraph), tilingOptions(tilingOptions),
294       fusionOptions(fusionOptions), marker(marker),
295       fusedOpMarker(fusedOpMarker), originalOpMarker(originalOpMarker) {}
296 
297 LogicalResult mlir::linalg::LinalgBaseTileAndFusePattern::matchAndRewrite(
298     Operation *op, PatternRewriter &rewriter) const {
299   LinalgOp linalgOp = dyn_cast<LinalgOp>(op);
300   if (!linalgOp)
301     return failure();
302   if (failed(marker.checkAndNotify(rewriter, linalgOp)))
303     return failure();
304   if (!linalgOp.hasBufferSemantics())
305     return failure();
306 
307   DenseSet<Operation *> producers;
308   producers.insert(linalgOp);
309   for (auto dependence : dependenceGraph.getDependentOperationsInto(linalgOp)) {
310     Optional<unsigned> operandNumber = dependence.getIndexingOpViewOperandNum();
311     // When looking at dependences into, indexingOp is always OpOperand. We
312     // could assert, but continue if this is not the case.
313     if (!operandNumber)
314       continue;
315     if (!fusionOptions.indicesToFuse.count(operandNumber.getValue()))
316       continue;
317     if (isa<LinalgOp>(dependence.getDependentOp()))
318       producers.insert(dependence.getDependentOp());
319   }
320 
321   SmallVector<LinalgOp, 1> fusionOps;
322   for (auto it = op->getBlock()->begin(), ie = Block::iterator(op); it != ie;
323        ++it) {
324     auto producerLinalgOp = dyn_cast<LinalgOp>(&(*it));
325     if (producerLinalgOp && producers.count(producerLinalgOp))
326       fusionOps.push_back(producerLinalgOp);
327   }
328   fusionOps.push_back(linalgOp);
329 
330   SmallVector<Value, 4> tileSizes =
331       tilingOptions.tileSizeComputationFunction(rewriter, op);
332   LinalgTilingOptions instanceTilingOptions = tilingOptions;
333   instanceTilingOptions.setTileSizes(tileSizes);
334   Optional<TiledAndFusedLinalgOps> tiledAndFusedOps = tileAndFuseLinalgOps(
335       rewriter, fusionOps, dependenceGraph, instanceTilingOptions);
336   if (!tiledAndFusedOps)
337     return failure();
338 
339   // Tile the unfused loops;
340   SmallVector<Value, 4> unfusedLoopTileSizes;
341   Value zero = rewriter.create<ConstantIndexOp>(op->getLoc(), 0);
342   for (auto tileSize : enumerate(tileSizes)) {
343     if (tiledAndFusedOps->fusedLoopDims.count(tileSize.index()))
344       unfusedLoopTileSizes.push_back(zero);
345     else
346       unfusedLoopTileSizes.push_back(tileSize.value());
347   }
348   // Tile the loop only if there is a non-zero tile size.
349   if (unfusedLoopTileSizes.size() > linalgOp.getNumLoops())
350     unfusedLoopTileSizes.resize(linalgOp.getNumLoops());
351   if (llvm::any_of(unfusedLoopTileSizes, [](Value val) {
352         if (auto cst = val.getDefiningOp<ConstantIndexOp>())
353           return cst.getValue() != 0;
354         return true;
355       })) {
356     LinalgTilingOptions unfusedTilingOptions = tilingOptions;
357     unfusedTilingOptions.setTileSizes(unfusedLoopTileSizes);
358     Optional<TiledLinalgOp> unfusedTiledOp =
359         tileLinalgOp(rewriter, tiledAndFusedOps->op, unfusedTilingOptions);
360     if (!unfusedTiledOp)
361       return failure();
362     rewriter.eraseOp(tiledAndFusedOps->op);
363     tiledAndFusedOps->op = unfusedTiledOp->op;
364   }
365 
366   marker.replaceLinalgTransformationFilter(rewriter,
367                                            tiledAndFusedOps->op.getOperation());
368   for (auto fusedOp : tiledAndFusedOps->fusedProducers) {
369     fusedOpMarker.replaceLinalgTransformationFilter(rewriter,
370                                                     fusedOp.getOperation());
371   }
372   for (auto origProducerOp : ArrayRef<LinalgOp>(fusionOps).drop_back()) {
373     originalOpMarker.replaceLinalgTransformationFilter(
374         rewriter, origProducerOp.getOperation());
375   }
376   rewriter.updateRootInPlace(op, [&]() {
377     originalOpMarker.replaceLinalgTransformationFilter(rewriter, op);
378   });
379   return success();
380 }
381 
382 /// Linalg base interchange pattern.
383 mlir::linalg::LinalgBaseInterchangePattern::LinalgBaseInterchangePattern(
384     StringRef opName, MLIRContext *context,
385     ArrayRef<unsigned> interchangeVector, LinalgTransformationFilter marker,
386     PatternBenefit benefit)
387     : RewritePattern(opName, {}, benefit, context), marker(marker),
388       interchangeVector(interchangeVector.begin(), interchangeVector.end()) {}
389 
390 LogicalResult mlir::linalg::LinalgBaseInterchangePattern::matchAndRewrite(
391     Operation *op, PatternRewriter &rewriter) const {
392   LinalgOp linalgOp = dyn_cast<LinalgOp>(op);
393   if (!linalgOp)
394     return failure();
395   if (failed(marker.checkAndNotify(rewriter, linalgOp)))
396     return failure();
397   if (failed(interchangeGenericLinalgOpPrecondition(op, interchangeVector)))
398     return failure();
399 
400   // TODO: figure out how this interplays with named ops. In particular this
401   // should break the named op property.
402   rewriter.updateRootInPlace(op, [&]() {
403     interchange(linalgOp, interchangeVector);
404     // New marker if specified.
405     marker.replaceLinalgTransformationFilter(rewriter, op);
406   });
407   return success();
408 }
409 
410 mlir::linalg::LinalgBasePromotionPattern::LinalgBasePromotionPattern(
411     StringRef opName, MLIRContext *context, LinalgPromotionOptions options,
412     LinalgTransformationFilter marker, PatternBenefit benefit)
413     : RewritePattern(opName, {}, benefit, context), marker(marker),
414       options(options) {}
415 
416 LogicalResult mlir::linalg::LinalgBasePromotionPattern::matchAndRewrite(
417     Operation *op, PatternRewriter &rewriter) const {
418   if (failed(marker.checkAndNotify(rewriter, op)))
419     return failure();
420   if (failed(promoteSubviewsPrecondition(op, options)))
421     return failure();
422 
423   // TODO: We cannot use root update here. This pattern is creating other ops,
424   // so if the promotion fails, those need to be cleaned up, which doesnt seem
425   // to be happening here. So to fail properly, we should be cloning the op and
426   // deleting the previous op. This needs more investigation.
427   rewriter.startRootUpdate(op);
428   Optional<LinalgOp> promotedOp = promoteSubViews(rewriter, op, options);
429   if (!promotedOp) {
430     rewriter.cancelRootUpdate(op);
431     return op->emitError("subview promotion failed");
432   }
433   rewriter.finalizeRootUpdate(op);
434   marker.replaceLinalgTransformationFilter(rewriter, op);
435   return success();
436 }
437 
438 mlir::linalg::LinalgBaseVectorizationPattern::LinalgBaseVectorizationPattern(
439     StringRef opName, MLIRContext *context, LinalgTransformationFilter marker,
440     PatternBenefit benefit)
441     : RewritePattern(opName, {}, benefit, context), marker(marker) {}
442 
443 LogicalResult mlir::linalg::LinalgBaseVectorizationPattern::matchAndRewrite(
444     Operation *op, PatternRewriter &rewriter) const {
445   LinalgOp linalgOp = dyn_cast<LinalgOp>(op);
446   if (!linalgOp)
447     return failure();
448   if (failed(marker.checkAndNotify(rewriter, linalgOp)))
449     return failure();
450   if (failed(vectorizeLinalgOpPrecondition(op)))
451     return failure();
452   vectorizeLinalgOp(rewriter, op);
453   rewriter.eraseOp(op);
454   return success();
455 }
456 
457 LogicalResult mlir::linalg::applyStagedPatterns(
458     Operation *op, ArrayRef<FrozenRewritePatternList> stage1Patterns,
459     const FrozenRewritePatternList &stage2Patterns,
460     function_ref<LogicalResult(Operation *)> stage3Lambda) {
461   unsigned iteration = 0;
462   (void)iteration;
463   for (const auto &patterns : stage1Patterns) {
464     LLVM_DEBUG(DBGS() << "Before 1st stage, iter: " << ++iteration << "\n"
465                       << *op);
466     if (failed(applyPatternsAndFoldGreedily(op, patterns))) {
467       LLVM_DEBUG(DBGS() << "Underlying first stage rewrite did not converge");
468       return failure();
469     }
470     LLVM_DEBUG(DBGS() << "After 1st stage, iter: " << ++iteration << "\n"
471                       << *op);
472     if (failed(applyPatternsAndFoldGreedily(op, stage2Patterns))) {
473       LLVM_DEBUG(DBGS() << "Underlying 2nd stage rewrite did not converge");
474       return failure();
475     }
476     LLVM_DEBUG(DBGS() << "After 2nd stage, iter : " << iteration << "\n"
477                       << *op);
478     if (stage3Lambda) {
479       if (failed(stage3Lambda(op)))
480         return failure();
481       LLVM_DEBUG(DBGS() << "After 3rd stage, iter : " << iteration << "\n"
482                         << *op);
483     }
484   }
485   return success();
486 }
487 
488 /// Given the `lbVal`, `ubVal` and `stepVal` of a loop, append `lbVal` and
489 /// `ubVal` to `dims` and `stepVal` to `symbols`.
490 /// Create new AffineDimExpr (`%lb` and `%ub`) and AffineSymbolExpr (`%step`)
491 /// with positions matching the newly appended values. Substitute occurrences of
492 /// `dimExpr` by either the min expression (i.e. `%lb`) or the max expression
493 /// (i.e. `%lb + %step * floordiv(%ub -1 - %lb, %step)`), depending on whether
494 /// the induction variable is used with a positive or negative  coefficient.
495 static AffineExpr substituteLoopInExpr(AffineExpr expr, AffineExpr dimExpr,
496                                        Value lbVal, Value ubVal, Value stepVal,
497                                        SmallVectorImpl<Value> &dims,
498                                        SmallVectorImpl<Value> &symbols) {
499   MLIRContext *ctx = lbVal.getContext();
500   AffineExpr lb = getAffineDimExpr(dims.size(), ctx);
501   dims.push_back(lbVal);
502   AffineExpr ub = getAffineDimExpr(dims.size(), ctx);
503   dims.push_back(ubVal);
504   AffineExpr step = getAffineSymbolExpr(symbols.size(), ctx);
505   symbols.push_back(stepVal);
506   LLVM_DEBUG(DBGS() << "Before: " << expr << "\n");
507   AffineExpr ee = substWithMin(expr, dimExpr, lb,
508                                lb + step * ((ub - 1) - lb).floorDiv(step));
509   LLVM_DEBUG(DBGS() << "After: " << expr << "\n");
510   return ee;
511 }
512 
513 /// Traverse the `dims` and substitute known min or max expressions in place of
514 /// induction variables in `exprs`.
515 static AffineMap substitute(AffineMap map, SmallVectorImpl<Value> &dims,
516                             SmallVectorImpl<Value> &symbols) {
517   auto exprs = llvm::to_vector<4>(map.getResults());
518   for (AffineExpr &expr : exprs) {
519     bool substituted = true;
520     while (substituted) {
521       substituted = false;
522       for (unsigned dimIdx = 0; dimIdx < dims.size(); ++dimIdx) {
523         Value dim = dims[dimIdx];
524         AffineExpr dimExpr = getAffineDimExpr(dimIdx, expr.getContext());
525         LLVM_DEBUG(DBGS() << "Subst: " << dim << " @ " << dimExpr << "\n");
526         AffineExpr substitutedExpr;
527         if (auto forOp = scf::getForInductionVarOwner(dim))
528           substitutedExpr = substituteLoopInExpr(
529               expr, dimExpr, forOp.lowerBound(), forOp.upperBound(),
530               forOp.step(), dims, symbols);
531 
532         if (auto parallelForOp = scf::getParallelForInductionVarOwner(dim))
533           for (unsigned idx = 0, e = parallelForOp.getNumLoops(); idx < e;
534                ++idx)
535             substitutedExpr = substituteLoopInExpr(
536                 expr, dimExpr, parallelForOp.lowerBound()[idx],
537                 parallelForOp.upperBound()[idx], parallelForOp.step()[idx],
538                 dims, symbols);
539 
540         if (!substitutedExpr)
541           continue;
542 
543         substituted = (substitutedExpr != expr);
544         expr = substitutedExpr;
545       }
546     }
547 
548     // Cleanup and simplify the results.
549     // This needs to happen outside of the loop iterating on dims.size() since
550     // it modifies dims.
551     SmallVector<Value, 4> operands(dims.begin(), dims.end());
552     operands.append(symbols.begin(), symbols.end());
553     auto map = AffineMap::get(dims.size(), symbols.size(), exprs,
554                               exprs.front().getContext());
555 
556     LLVM_DEBUG(DBGS() << "Map to simplify: " << map << "\n");
557 
558     // Pull in affine.apply operations and compose them fully into the
559     // result.
560     fullyComposeAffineMapAndOperands(&map, &operands);
561     canonicalizeMapAndOperands(&map, &operands);
562     map = simplifyAffineMap(map);
563     // Assign the results.
564     exprs.assign(map.getResults().begin(), map.getResults().end());
565     dims.assign(operands.begin(), operands.begin() + map.getNumDims());
566     symbols.assign(operands.begin() + map.getNumDims(), operands.end());
567 
568     LLVM_DEBUG(DBGS() << "Map simplified: " << map << "\n");
569   }
570 
571   assert(!exprs.empty() && "Unexpected empty exprs");
572   return AffineMap::get(dims.size(), symbols.size(), exprs, map.getContext());
573 }
574 
575 LogicalResult AffineMinSCFCanonicalizationPattern::matchAndRewrite(
576     AffineMinOp minOp, PatternRewriter &rewriter) const {
577   LLVM_DEBUG(DBGS() << "Canonicalize AffineMinSCF: " << *minOp.getOperation()
578                     << "\n");
579 
580   SmallVector<Value, 4> dims(minOp.getDimOperands()),
581       symbols(minOp.getSymbolOperands());
582   AffineMap map = substitute(minOp.getAffineMap(), dims, symbols);
583 
584   LLVM_DEBUG(DBGS() << "Resulting map: " << map << "\n");
585 
586   // Check whether any of the expressions, when subtracted from all other
587   // expressions, produces only >= 0 constants. If so, it is the min.
588   for (auto e : minOp.getAffineMap().getResults()) {
589     LLVM_DEBUG(DBGS() << "Candidate min: " << e << "\n");
590     if (!e.isSymbolicOrConstant())
591       continue;
592 
593     auto isNonPositive = [](AffineExpr e) {
594       if (auto cst = e.dyn_cast<AffineConstantExpr>())
595         return cst.getValue() < 0;
596       return true;
597     };
598 
599     // Build the subMap and check everything is statically known to be
600     // positive.
601     SmallVector<AffineExpr, 4> subExprs;
602     subExprs.reserve(map.getNumResults());
603     for (auto ee : map.getResults())
604       subExprs.push_back(ee - e);
605     MLIRContext *ctx = minOp.getContext();
606     AffineMap subMap = simplifyAffineMap(
607         AffineMap::get(map.getNumDims(), map.getNumSymbols(), subExprs, ctx));
608     LLVM_DEBUG(DBGS() << "simplified subMap: " << subMap << "\n");
609     if (llvm::any_of(subMap.getResults(), isNonPositive))
610       continue;
611 
612     // Static min found.
613     if (auto cst = e.dyn_cast<AffineConstantExpr>()) {
614       rewriter.replaceOpWithNewOp<ConstantIndexOp>(minOp, cst.getValue());
615     } else {
616       auto resultMap = AffineMap::get(0, map.getNumSymbols(), {e}, ctx);
617       SmallVector<Value, 4> resultOperands = dims;
618       resultOperands.append(symbols.begin(), symbols.end());
619       canonicalizeMapAndOperands(&resultMap, &resultOperands);
620       resultMap = simplifyAffineMap(resultMap);
621       rewriter.replaceOpWithNewOp<AffineApplyOp>(minOp, resultMap,
622                                                  resultOperands);
623     }
624     return success();
625   }
626 
627   return failure();
628 }
629