/home/arjun/llvm-project/mlir/lib/IR/AffineExpr.cpp
Line | Count | Source (jump to first uncovered line) |
1 | | //===- AffineExpr.cpp - MLIR Affine Expr Classes --------------------------===// |
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 | | #include "mlir/IR/AffineExpr.h" |
10 | | #include "AffineExprDetail.h" |
11 | | #include "mlir/IR/AffineExprVisitor.h" |
12 | | #include "mlir/IR/AffineMap.h" |
13 | | #include "mlir/IR/IntegerSet.h" |
14 | | #include "mlir/Support/MathExtras.h" |
15 | | #include "llvm/ADT/STLExtras.h" |
16 | | |
17 | | using namespace mlir; |
18 | | using namespace mlir::detail; |
19 | | |
20 | 0 | MLIRContext *AffineExpr::getContext() const { return expr->context; } |
21 | | |
22 | 0 | AffineExprKind AffineExpr::getKind() const { |
23 | 0 | return static_cast<AffineExprKind>(expr->getKind()); |
24 | 0 | } |
25 | | |
26 | | /// Walk all of the AffineExprs in this subgraph in postorder. |
27 | 0 | void AffineExpr::walk(std::function<void(AffineExpr)> callback) const { |
28 | 0 | struct AffineExprWalker : public AffineExprVisitor<AffineExprWalker> { |
29 | 0 | std::function<void(AffineExpr)> callback; |
30 | 0 |
|
31 | 0 | AffineExprWalker(std::function<void(AffineExpr)> callback) |
32 | 0 | : callback(callback) {} |
33 | 0 |
|
34 | 0 | void visitAffineBinaryOpExpr(AffineBinaryOpExpr expr) { callback(expr); } |
35 | 0 | void visitConstantExpr(AffineConstantExpr expr) { callback(expr); } |
36 | 0 | void visitDimExpr(AffineDimExpr expr) { callback(expr); } |
37 | 0 | void visitSymbolExpr(AffineSymbolExpr expr) { callback(expr); } |
38 | 0 | }; |
39 | 0 |
|
40 | 0 | AffineExprWalker(callback).walkPostOrder(*this); |
41 | 0 | } |
42 | | |
43 | | // Dispatch affine expression construction based on kind. |
44 | | AffineExpr mlir::getAffineBinaryOpExpr(AffineExprKind kind, AffineExpr lhs, |
45 | 0 | AffineExpr rhs) { |
46 | 0 | if (kind == AffineExprKind::Add) |
47 | 0 | return lhs + rhs; |
48 | 0 | if (kind == AffineExprKind::Mul) |
49 | 0 | return lhs * rhs; |
50 | 0 | if (kind == AffineExprKind::FloorDiv) |
51 | 0 | return lhs.floorDiv(rhs); |
52 | 0 | if (kind == AffineExprKind::CeilDiv) |
53 | 0 | return lhs.ceilDiv(rhs); |
54 | 0 | if (kind == AffineExprKind::Mod) |
55 | 0 | return lhs % rhs; |
56 | 0 | |
57 | 0 | llvm_unreachable("unknown binary operation on affine expressions"); |
58 | 0 | } |
59 | | |
60 | | /// This method substitutes any uses of dimensions and symbols (e.g. |
61 | | /// dim#0 with dimReplacements[0]) and returns the modified expression tree. |
62 | | AffineExpr |
63 | | AffineExpr::replaceDimsAndSymbols(ArrayRef<AffineExpr> dimReplacements, |
64 | 0 | ArrayRef<AffineExpr> symReplacements) const { |
65 | 0 | switch (getKind()) { |
66 | 0 | case AffineExprKind::Constant: |
67 | 0 | return *this; |
68 | 0 | case AffineExprKind::DimId: { |
69 | 0 | unsigned dimId = cast<AffineDimExpr>().getPosition(); |
70 | 0 | if (dimId >= dimReplacements.size()) |
71 | 0 | return *this; |
72 | 0 | return dimReplacements[dimId]; |
73 | 0 | } |
74 | 0 | case AffineExprKind::SymbolId: { |
75 | 0 | unsigned symId = cast<AffineSymbolExpr>().getPosition(); |
76 | 0 | if (symId >= symReplacements.size()) |
77 | 0 | return *this; |
78 | 0 | return symReplacements[symId]; |
79 | 0 | } |
80 | 0 | case AffineExprKind::Add: |
81 | 0 | case AffineExprKind::Mul: |
82 | 0 | case AffineExprKind::FloorDiv: |
83 | 0 | case AffineExprKind::CeilDiv: |
84 | 0 | case AffineExprKind::Mod: |
85 | 0 | auto binOp = cast<AffineBinaryOpExpr>(); |
86 | 0 | auto lhs = binOp.getLHS(), rhs = binOp.getRHS(); |
87 | 0 | auto newLHS = lhs.replaceDimsAndSymbols(dimReplacements, symReplacements); |
88 | 0 | auto newRHS = rhs.replaceDimsAndSymbols(dimReplacements, symReplacements); |
89 | 0 | if (newLHS == lhs && newRHS == rhs) |
90 | 0 | return *this; |
91 | 0 | return getAffineBinaryOpExpr(getKind(), newLHS, newRHS); |
92 | 0 | } |
93 | 0 | llvm_unreachable("Unknown AffineExpr"); |
94 | 0 | } |
95 | | |
96 | | /// Returns true if this expression is made out of only symbols and |
97 | | /// constants (no dimensional identifiers). |
98 | 0 | bool AffineExpr::isSymbolicOrConstant() const { |
99 | 0 | switch (getKind()) { |
100 | 0 | case AffineExprKind::Constant: |
101 | 0 | return true; |
102 | 0 | case AffineExprKind::DimId: |
103 | 0 | return false; |
104 | 0 | case AffineExprKind::SymbolId: |
105 | 0 | return true; |
106 | 0 |
|
107 | 0 | case AffineExprKind::Add: |
108 | 0 | case AffineExprKind::Mul: |
109 | 0 | case AffineExprKind::FloorDiv: |
110 | 0 | case AffineExprKind::CeilDiv: |
111 | 0 | case AffineExprKind::Mod: { |
112 | 0 | auto expr = this->cast<AffineBinaryOpExpr>(); |
113 | 0 | return expr.getLHS().isSymbolicOrConstant() && |
114 | 0 | expr.getRHS().isSymbolicOrConstant(); |
115 | 0 | } |
116 | 0 | } |
117 | 0 | llvm_unreachable("Unknown AffineExpr"); |
118 | 0 | } |
119 | | |
120 | | /// Returns true if this is a pure affine expression, i.e., multiplication, |
121 | | /// floordiv, ceildiv, and mod is only allowed w.r.t constants. |
122 | 0 | bool AffineExpr::isPureAffine() const { |
123 | 0 | switch (getKind()) { |
124 | 0 | case AffineExprKind::SymbolId: |
125 | 0 | case AffineExprKind::DimId: |
126 | 0 | case AffineExprKind::Constant: |
127 | 0 | return true; |
128 | 0 | case AffineExprKind::Add: { |
129 | 0 | auto op = cast<AffineBinaryOpExpr>(); |
130 | 0 | return op.getLHS().isPureAffine() && op.getRHS().isPureAffine(); |
131 | 0 | } |
132 | 0 |
|
133 | 0 | case AffineExprKind::Mul: { |
134 | 0 | // TODO: Canonicalize the constants in binary operators to the RHS when |
135 | 0 | // possible, allowing this to merge into the next case. |
136 | 0 | auto op = cast<AffineBinaryOpExpr>(); |
137 | 0 | return op.getLHS().isPureAffine() && op.getRHS().isPureAffine() && |
138 | 0 | (op.getLHS().template isa<AffineConstantExpr>() || |
139 | 0 | op.getRHS().template isa<AffineConstantExpr>()); |
140 | 0 | } |
141 | 0 | case AffineExprKind::FloorDiv: |
142 | 0 | case AffineExprKind::CeilDiv: |
143 | 0 | case AffineExprKind::Mod: { |
144 | 0 | auto op = cast<AffineBinaryOpExpr>(); |
145 | 0 | return op.getLHS().isPureAffine() && |
146 | 0 | op.getRHS().template isa<AffineConstantExpr>(); |
147 | 0 | } |
148 | 0 | } |
149 | 0 | llvm_unreachable("Unknown AffineExpr"); |
150 | 0 | } |
151 | | |
152 | | // Returns the greatest known integral divisor of this affine expression. |
153 | 0 | int64_t AffineExpr::getLargestKnownDivisor() const { |
154 | 0 | AffineBinaryOpExpr binExpr(nullptr); |
155 | 0 | switch (getKind()) { |
156 | 0 | case AffineExprKind::SymbolId: |
157 | 0 | LLVM_FALLTHROUGH; |
158 | 0 | case AffineExprKind::DimId: |
159 | 0 | return 1; |
160 | 0 | case AffineExprKind::Constant: |
161 | 0 | return std::abs(this->cast<AffineConstantExpr>().getValue()); |
162 | 0 | case AffineExprKind::Mul: { |
163 | 0 | binExpr = this->cast<AffineBinaryOpExpr>(); |
164 | 0 | return binExpr.getLHS().getLargestKnownDivisor() * |
165 | 0 | binExpr.getRHS().getLargestKnownDivisor(); |
166 | 0 | } |
167 | 0 | case AffineExprKind::Add: |
168 | 0 | LLVM_FALLTHROUGH; |
169 | 0 | case AffineExprKind::FloorDiv: |
170 | 0 | case AffineExprKind::CeilDiv: |
171 | 0 | case AffineExprKind::Mod: { |
172 | 0 | binExpr = cast<AffineBinaryOpExpr>(); |
173 | 0 | return llvm::GreatestCommonDivisor64( |
174 | 0 | binExpr.getLHS().getLargestKnownDivisor(), |
175 | 0 | binExpr.getRHS().getLargestKnownDivisor()); |
176 | 0 | } |
177 | 0 | } |
178 | 0 | llvm_unreachable("Unknown AffineExpr"); |
179 | 0 | } |
180 | | |
181 | 0 | bool AffineExpr::isMultipleOf(int64_t factor) const { |
182 | 0 | AffineBinaryOpExpr binExpr(nullptr); |
183 | 0 | uint64_t l, u; |
184 | 0 | switch (getKind()) { |
185 | 0 | case AffineExprKind::SymbolId: |
186 | 0 | LLVM_FALLTHROUGH; |
187 | 0 | case AffineExprKind::DimId: |
188 | 0 | return factor * factor == 1; |
189 | 0 | case AffineExprKind::Constant: |
190 | 0 | return cast<AffineConstantExpr>().getValue() % factor == 0; |
191 | 0 | case AffineExprKind::Mul: { |
192 | 0 | binExpr = cast<AffineBinaryOpExpr>(); |
193 | 0 | // It's probably not worth optimizing this further (to not traverse the |
194 | 0 | // whole sub-tree under - it that would require a version of isMultipleOf |
195 | 0 | // that on a 'false' return also returns the largest known divisor). |
196 | 0 | return (l = binExpr.getLHS().getLargestKnownDivisor()) % factor == 0 || |
197 | 0 | (u = binExpr.getRHS().getLargestKnownDivisor()) % factor == 0 || |
198 | 0 | (l * u) % factor == 0; |
199 | 0 | } |
200 | 0 | case AffineExprKind::Add: |
201 | 0 | case AffineExprKind::FloorDiv: |
202 | 0 | case AffineExprKind::CeilDiv: |
203 | 0 | case AffineExprKind::Mod: { |
204 | 0 | binExpr = cast<AffineBinaryOpExpr>(); |
205 | 0 | return llvm::GreatestCommonDivisor64( |
206 | 0 | binExpr.getLHS().getLargestKnownDivisor(), |
207 | 0 | binExpr.getRHS().getLargestKnownDivisor()) % |
208 | 0 | factor == |
209 | 0 | 0; |
210 | 0 | } |
211 | 0 | } |
212 | 0 | llvm_unreachable("Unknown AffineExpr"); |
213 | 0 | } |
214 | | |
215 | 0 | bool AffineExpr::isFunctionOfDim(unsigned position) const { |
216 | 0 | if (getKind() == AffineExprKind::DimId) { |
217 | 0 | return *this == mlir::getAffineDimExpr(position, getContext()); |
218 | 0 | } |
219 | 0 | if (auto expr = this->dyn_cast<AffineBinaryOpExpr>()) { |
220 | 0 | return expr.getLHS().isFunctionOfDim(position) || |
221 | 0 | expr.getRHS().isFunctionOfDim(position); |
222 | 0 | } |
223 | 0 | return false; |
224 | 0 | } |
225 | | |
226 | | AffineBinaryOpExpr::AffineBinaryOpExpr(AffineExpr::ImplType *ptr) |
227 | 0 | : AffineExpr(ptr) {} |
228 | 0 | AffineExpr AffineBinaryOpExpr::getLHS() const { |
229 | 0 | return static_cast<ImplType *>(expr)->lhs; |
230 | 0 | } |
231 | 0 | AffineExpr AffineBinaryOpExpr::getRHS() const { |
232 | 0 | return static_cast<ImplType *>(expr)->rhs; |
233 | 0 | } |
234 | | |
235 | 0 | AffineDimExpr::AffineDimExpr(AffineExpr::ImplType *ptr) : AffineExpr(ptr) {} |
236 | 0 | unsigned AffineDimExpr::getPosition() const { |
237 | 0 | return static_cast<ImplType *>(expr)->position; |
238 | 0 | } |
239 | | |
240 | | static AffineExpr getAffineDimOrSymbol(AffineExprKind kind, unsigned position, |
241 | 0 | MLIRContext *context) { |
242 | 0 | auto assignCtx = [context](AffineDimExprStorage *storage) { |
243 | 0 | storage->context = context; |
244 | 0 | }; |
245 | 0 |
|
246 | 0 | StorageUniquer &uniquer = context->getAffineUniquer(); |
247 | 0 | return uniquer.get<AffineDimExprStorage>( |
248 | 0 | assignCtx, static_cast<unsigned>(kind), position); |
249 | 0 | } |
250 | | |
251 | 0 | AffineExpr mlir::getAffineDimExpr(unsigned position, MLIRContext *context) { |
252 | 0 | return getAffineDimOrSymbol(AffineExprKind::DimId, position, context); |
253 | 0 | } |
254 | | |
255 | | AffineSymbolExpr::AffineSymbolExpr(AffineExpr::ImplType *ptr) |
256 | 0 | : AffineExpr(ptr) {} |
257 | 0 | unsigned AffineSymbolExpr::getPosition() const { |
258 | 0 | return static_cast<ImplType *>(expr)->position; |
259 | 0 | } |
260 | | |
261 | 0 | AffineExpr mlir::getAffineSymbolExpr(unsigned position, MLIRContext *context) { |
262 | 0 | return getAffineDimOrSymbol(AffineExprKind::SymbolId, position, context); |
263 | 0 | ; |
264 | 0 | } |
265 | | |
266 | | AffineConstantExpr::AffineConstantExpr(AffineExpr::ImplType *ptr) |
267 | 0 | : AffineExpr(ptr) {} |
268 | 0 | int64_t AffineConstantExpr::getValue() const { |
269 | 0 | return static_cast<ImplType *>(expr)->constant; |
270 | 0 | } |
271 | | |
272 | 0 | bool AffineExpr::operator==(int64_t v) const { |
273 | 0 | return *this == getAffineConstantExpr(v, getContext()); |
274 | 0 | } |
275 | | |
276 | 0 | AffineExpr mlir::getAffineConstantExpr(int64_t constant, MLIRContext *context) { |
277 | 0 | auto assignCtx = [context](AffineConstantExprStorage *storage) { |
278 | 0 | storage->context = context; |
279 | 0 | }; |
280 | 0 |
|
281 | 0 | StorageUniquer &uniquer = context->getAffineUniquer(); |
282 | 0 | return uniquer.get<AffineConstantExprStorage>( |
283 | 0 | assignCtx, static_cast<unsigned>(AffineExprKind::Constant), constant); |
284 | 0 | } |
285 | | |
286 | | /// Simplify add expression. Return nullptr if it can't be simplified. |
287 | 0 | static AffineExpr simplifyAdd(AffineExpr lhs, AffineExpr rhs) { |
288 | 0 | auto lhsConst = lhs.dyn_cast<AffineConstantExpr>(); |
289 | 0 | auto rhsConst = rhs.dyn_cast<AffineConstantExpr>(); |
290 | 0 | // Fold if both LHS, RHS are a constant. |
291 | 0 | if (lhsConst && rhsConst) |
292 | 0 | return getAffineConstantExpr(lhsConst.getValue() + rhsConst.getValue(), |
293 | 0 | lhs.getContext()); |
294 | 0 | |
295 | 0 | // Canonicalize so that only the RHS is a constant. (4 + d0 becomes d0 + 4). |
296 | 0 | // If only one of them is a symbolic expressions, make it the RHS. |
297 | 0 | if (lhs.isa<AffineConstantExpr>() || |
298 | 0 | (lhs.isSymbolicOrConstant() && !rhs.isSymbolicOrConstant())) { |
299 | 0 | return rhs + lhs; |
300 | 0 | } |
301 | 0 | |
302 | 0 | // At this point, if there was a constant, it would be on the right. |
303 | 0 | |
304 | 0 | // Addition with a zero is a noop, return the other input. |
305 | 0 | if (rhsConst) { |
306 | 0 | if (rhsConst.getValue() == 0) |
307 | 0 | return lhs; |
308 | 0 | } |
309 | 0 | // Fold successive additions like (d0 + 2) + 3 into d0 + 5. |
310 | 0 | auto lBin = lhs.dyn_cast<AffineBinaryOpExpr>(); |
311 | 0 | if (lBin && rhsConst && lBin.getKind() == AffineExprKind::Add) { |
312 | 0 | if (auto lrhs = lBin.getRHS().dyn_cast<AffineConstantExpr>()) |
313 | 0 | return lBin.getLHS() + (lrhs.getValue() + rhsConst.getValue()); |
314 | 0 | } |
315 | 0 | |
316 | 0 | // Detect "c1 * expr + c_2 * expr" as "(c1 + c2) * expr". |
317 | 0 | // c1 is rRhsConst, c2 is rLhsConst; firstExpr, secondExpr are their |
318 | 0 | // respective multiplicands. |
319 | 0 | Optional<int64_t> rLhsConst, rRhsConst; |
320 | 0 | AffineExpr firstExpr, secondExpr; |
321 | 0 | AffineConstantExpr rLhsConstExpr; |
322 | 0 | auto lBinOpExpr = lhs.dyn_cast<AffineBinaryOpExpr>(); |
323 | 0 | if (lBinOpExpr && lBinOpExpr.getKind() == AffineExprKind::Mul && |
324 | 0 | (rLhsConstExpr = lBinOpExpr.getRHS().dyn_cast<AffineConstantExpr>())) { |
325 | 0 | rLhsConst = rLhsConstExpr.getValue(); |
326 | 0 | firstExpr = lBinOpExpr.getLHS(); |
327 | 0 | } else { |
328 | 0 | rLhsConst = 1; |
329 | 0 | firstExpr = lhs; |
330 | 0 | } |
331 | 0 |
|
332 | 0 | auto rBinOpExpr = rhs.dyn_cast<AffineBinaryOpExpr>(); |
333 | 0 | AffineConstantExpr rRhsConstExpr; |
334 | 0 | if (rBinOpExpr && rBinOpExpr.getKind() == AffineExprKind::Mul && |
335 | 0 | (rRhsConstExpr = rBinOpExpr.getRHS().dyn_cast<AffineConstantExpr>())) { |
336 | 0 | rRhsConst = rRhsConstExpr.getValue(); |
337 | 0 | secondExpr = rBinOpExpr.getLHS(); |
338 | 0 | } else { |
339 | 0 | rRhsConst = 1; |
340 | 0 | secondExpr = rhs; |
341 | 0 | } |
342 | 0 |
|
343 | 0 | if (rLhsConst && rRhsConst && firstExpr == secondExpr) |
344 | 0 | return getAffineBinaryOpExpr( |
345 | 0 | AffineExprKind::Mul, firstExpr, |
346 | 0 | getAffineConstantExpr(rLhsConst.getValue() + rRhsConst.getValue(), |
347 | 0 | lhs.getContext())); |
348 | 0 | |
349 | 0 | // When doing successive additions, bring constant to the right: turn (d0 + 2) |
350 | 0 | // + d1 into (d0 + d1) + 2. |
351 | 0 | if (lBin && lBin.getKind() == AffineExprKind::Add) { |
352 | 0 | if (auto lrhs = lBin.getRHS().dyn_cast<AffineConstantExpr>()) { |
353 | 0 | return lBin.getLHS() + rhs + lrhs; |
354 | 0 | } |
355 | 0 | } |
356 | 0 | |
357 | 0 | // Detect and transform "expr - c * (expr floordiv c)" to "expr mod c". This |
358 | 0 | // leads to a much more efficient form when 'c' is a power of two, and in |
359 | 0 | // general a more compact and readable form. |
360 | 0 | |
361 | 0 | // Process '(expr floordiv c) * (-c)'. |
362 | 0 | if (!rBinOpExpr) |
363 | 0 | return nullptr; |
364 | 0 | |
365 | 0 | auto lrhs = rBinOpExpr.getLHS(); |
366 | 0 | auto rrhs = rBinOpExpr.getRHS(); |
367 | 0 |
|
368 | 0 | // Process lrhs, which is 'expr floordiv c'. |
369 | 0 | AffineBinaryOpExpr lrBinOpExpr = lrhs.dyn_cast<AffineBinaryOpExpr>(); |
370 | 0 | if (!lrBinOpExpr || lrBinOpExpr.getKind() != AffineExprKind::FloorDiv) |
371 | 0 | return nullptr; |
372 | 0 | |
373 | 0 | auto llrhs = lrBinOpExpr.getLHS(); |
374 | 0 | auto rlrhs = lrBinOpExpr.getRHS(); |
375 | 0 |
|
376 | 0 | if (lhs == llrhs && rlrhs == -rrhs) { |
377 | 0 | return lhs % rlrhs; |
378 | 0 | } |
379 | 0 | return nullptr; |
380 | 0 | } |
381 | | |
382 | 0 | AffineExpr AffineExpr::operator+(int64_t v) const { |
383 | 0 | return *this + getAffineConstantExpr(v, getContext()); |
384 | 0 | } |
385 | 0 | AffineExpr AffineExpr::operator+(AffineExpr other) const { |
386 | 0 | if (auto simplified = simplifyAdd(*this, other)) |
387 | 0 | return simplified; |
388 | 0 | |
389 | 0 | StorageUniquer &uniquer = getContext()->getAffineUniquer(); |
390 | 0 | return uniquer.get<AffineBinaryOpExprStorage>( |
391 | 0 | /*initFn=*/{}, static_cast<unsigned>(AffineExprKind::Add), *this, other); |
392 | 0 | } |
393 | | |
394 | | /// Simplify a multiply expression. Return nullptr if it can't be simplified. |
395 | 0 | static AffineExpr simplifyMul(AffineExpr lhs, AffineExpr rhs) { |
396 | 0 | auto lhsConst = lhs.dyn_cast<AffineConstantExpr>(); |
397 | 0 | auto rhsConst = rhs.dyn_cast<AffineConstantExpr>(); |
398 | 0 |
|
399 | 0 | if (lhsConst && rhsConst) |
400 | 0 | return getAffineConstantExpr(lhsConst.getValue() * rhsConst.getValue(), |
401 | 0 | lhs.getContext()); |
402 | 0 | |
403 | 0 | assert(lhs.isSymbolicOrConstant() || rhs.isSymbolicOrConstant()); |
404 | 0 |
|
405 | 0 | // Canonicalize the mul expression so that the constant/symbolic term is the |
406 | 0 | // RHS. If both the lhs and rhs are symbolic, swap them if the lhs is a |
407 | 0 | // constant. (Note that a constant is trivially symbolic). |
408 | 0 | if (!rhs.isSymbolicOrConstant() || lhs.isa<AffineConstantExpr>()) { |
409 | 0 | // At least one of them has to be symbolic. |
410 | 0 | return rhs * lhs; |
411 | 0 | } |
412 | 0 | |
413 | 0 | // At this point, if there was a constant, it would be on the right. |
414 | 0 | |
415 | 0 | // Multiplication with a one is a noop, return the other input. |
416 | 0 | if (rhsConst) { |
417 | 0 | if (rhsConst.getValue() == 1) |
418 | 0 | return lhs; |
419 | 0 | // Multiplication with zero. |
420 | 0 | if (rhsConst.getValue() == 0) |
421 | 0 | return rhsConst; |
422 | 0 | } |
423 | 0 | |
424 | 0 | // Fold successive multiplications: eg: (d0 * 2) * 3 into d0 * 6. |
425 | 0 | auto lBin = lhs.dyn_cast<AffineBinaryOpExpr>(); |
426 | 0 | if (lBin && rhsConst && lBin.getKind() == AffineExprKind::Mul) { |
427 | 0 | if (auto lrhs = lBin.getRHS().dyn_cast<AffineConstantExpr>()) |
428 | 0 | return lBin.getLHS() * (lrhs.getValue() * rhsConst.getValue()); |
429 | 0 | } |
430 | 0 | |
431 | 0 | // When doing successive multiplication, bring constant to the right: turn (d0 |
432 | 0 | // * 2) * d1 into (d0 * d1) * 2. |
433 | 0 | if (lBin && lBin.getKind() == AffineExprKind::Mul) { |
434 | 0 | if (auto lrhs = lBin.getRHS().dyn_cast<AffineConstantExpr>()) { |
435 | 0 | return (lBin.getLHS() * rhs) * lrhs; |
436 | 0 | } |
437 | 0 | } |
438 | 0 | |
439 | 0 | return nullptr; |
440 | 0 | } |
441 | | |
442 | 0 | AffineExpr AffineExpr::operator*(int64_t v) const { |
443 | 0 | return *this * getAffineConstantExpr(v, getContext()); |
444 | 0 | } |
445 | 0 | AffineExpr AffineExpr::operator*(AffineExpr other) const { |
446 | 0 | if (auto simplified = simplifyMul(*this, other)) |
447 | 0 | return simplified; |
448 | 0 | |
449 | 0 | StorageUniquer &uniquer = getContext()->getAffineUniquer(); |
450 | 0 | return uniquer.get<AffineBinaryOpExprStorage>( |
451 | 0 | /*initFn=*/{}, static_cast<unsigned>(AffineExprKind::Mul), *this, other); |
452 | 0 | } |
453 | | |
454 | | // Unary minus, delegate to operator*. |
455 | 0 | AffineExpr AffineExpr::operator-() const { |
456 | 0 | return *this * getAffineConstantExpr(-1, getContext()); |
457 | 0 | } |
458 | | |
459 | | // Delegate to operator+. |
460 | 0 | AffineExpr AffineExpr::operator-(int64_t v) const { return *this + (-v); } |
461 | 0 | AffineExpr AffineExpr::operator-(AffineExpr other) const { |
462 | 0 | return *this + (-other); |
463 | 0 | } |
464 | | |
465 | 0 | static AffineExpr simplifyFloorDiv(AffineExpr lhs, AffineExpr rhs) { |
466 | 0 | auto lhsConst = lhs.dyn_cast<AffineConstantExpr>(); |
467 | 0 | auto rhsConst = rhs.dyn_cast<AffineConstantExpr>(); |
468 | 0 |
|
469 | 0 | // mlir floordiv by zero or negative numbers is undefined and preserved as is. |
470 | 0 | if (!rhsConst || rhsConst.getValue() < 1) |
471 | 0 | return nullptr; |
472 | 0 | |
473 | 0 | if (lhsConst) |
474 | 0 | return getAffineConstantExpr( |
475 | 0 | floorDiv(lhsConst.getValue(), rhsConst.getValue()), lhs.getContext()); |
476 | 0 | |
477 | 0 | // Fold floordiv of a multiply with a constant that is a multiple of the |
478 | 0 | // divisor. Eg: (i * 128) floordiv 64 = i * 2. |
479 | 0 | if (rhsConst == 1) |
480 | 0 | return lhs; |
481 | 0 | |
482 | 0 | // Simplify (expr * const) floordiv divConst when expr is known to be a |
483 | 0 | // multiple of divConst. |
484 | 0 | auto lBin = lhs.dyn_cast<AffineBinaryOpExpr>(); |
485 | 0 | if (lBin && lBin.getKind() == AffineExprKind::Mul) { |
486 | 0 | if (auto lrhs = lBin.getRHS().dyn_cast<AffineConstantExpr>()) { |
487 | 0 | // rhsConst is known to be a positive constant. |
488 | 0 | if (lrhs.getValue() % rhsConst.getValue() == 0) |
489 | 0 | return lBin.getLHS() * (lrhs.getValue() / rhsConst.getValue()); |
490 | 0 | } |
491 | 0 | } |
492 | 0 | |
493 | 0 | // Simplify (expr1 + expr2) floordiv divConst when either expr1 or expr2 is |
494 | 0 | // known to be a multiple of divConst. |
495 | 0 | if (lBin && lBin.getKind() == AffineExprKind::Add) { |
496 | 0 | int64_t llhsDiv = lBin.getLHS().getLargestKnownDivisor(); |
497 | 0 | int64_t lrhsDiv = lBin.getRHS().getLargestKnownDivisor(); |
498 | 0 | // rhsConst is known to be a positive constant. |
499 | 0 | if (llhsDiv % rhsConst.getValue() == 0 || |
500 | 0 | lrhsDiv % rhsConst.getValue() == 0) |
501 | 0 | return lBin.getLHS().floorDiv(rhsConst.getValue()) + |
502 | 0 | lBin.getRHS().floorDiv(rhsConst.getValue()); |
503 | 0 | } |
504 | 0 | |
505 | 0 | return nullptr; |
506 | 0 | } |
507 | | |
508 | 0 | AffineExpr AffineExpr::floorDiv(uint64_t v) const { |
509 | 0 | return floorDiv(getAffineConstantExpr(v, getContext())); |
510 | 0 | } |
511 | 0 | AffineExpr AffineExpr::floorDiv(AffineExpr other) const { |
512 | 0 | if (auto simplified = simplifyFloorDiv(*this, other)) |
513 | 0 | return simplified; |
514 | 0 | |
515 | 0 | StorageUniquer &uniquer = getContext()->getAffineUniquer(); |
516 | 0 | return uniquer.get<AffineBinaryOpExprStorage>( |
517 | 0 | /*initFn=*/{}, static_cast<unsigned>(AffineExprKind::FloorDiv), *this, |
518 | 0 | other); |
519 | 0 | } |
520 | | |
521 | 0 | static AffineExpr simplifyCeilDiv(AffineExpr lhs, AffineExpr rhs) { |
522 | 0 | auto lhsConst = lhs.dyn_cast<AffineConstantExpr>(); |
523 | 0 | auto rhsConst = rhs.dyn_cast<AffineConstantExpr>(); |
524 | 0 |
|
525 | 0 | if (!rhsConst || rhsConst.getValue() < 1) |
526 | 0 | return nullptr; |
527 | 0 | |
528 | 0 | if (lhsConst) |
529 | 0 | return getAffineConstantExpr( |
530 | 0 | ceilDiv(lhsConst.getValue(), rhsConst.getValue()), lhs.getContext()); |
531 | 0 | |
532 | 0 | // Fold ceildiv of a multiply with a constant that is a multiple of the |
533 | 0 | // divisor. Eg: (i * 128) ceildiv 64 = i * 2. |
534 | 0 | if (rhsConst.getValue() == 1) |
535 | 0 | return lhs; |
536 | 0 | |
537 | 0 | // Simplify (expr * const) ceildiv divConst when const is known to be a |
538 | 0 | // multiple of divConst. |
539 | 0 | auto lBin = lhs.dyn_cast<AffineBinaryOpExpr>(); |
540 | 0 | if (lBin && lBin.getKind() == AffineExprKind::Mul) { |
541 | 0 | if (auto lrhs = lBin.getRHS().dyn_cast<AffineConstantExpr>()) { |
542 | 0 | // rhsConst is known to be a positive constant. |
543 | 0 | if (lrhs.getValue() % rhsConst.getValue() == 0) |
544 | 0 | return lBin.getLHS() * (lrhs.getValue() / rhsConst.getValue()); |
545 | 0 | } |
546 | 0 | } |
547 | 0 | |
548 | 0 | return nullptr; |
549 | 0 | } |
550 | | |
551 | 0 | AffineExpr AffineExpr::ceilDiv(uint64_t v) const { |
552 | 0 | return ceilDiv(getAffineConstantExpr(v, getContext())); |
553 | 0 | } |
554 | 0 | AffineExpr AffineExpr::ceilDiv(AffineExpr other) const { |
555 | 0 | if (auto simplified = simplifyCeilDiv(*this, other)) |
556 | 0 | return simplified; |
557 | 0 | |
558 | 0 | StorageUniquer &uniquer = getContext()->getAffineUniquer(); |
559 | 0 | return uniquer.get<AffineBinaryOpExprStorage>( |
560 | 0 | /*initFn=*/{}, static_cast<unsigned>(AffineExprKind::CeilDiv), *this, |
561 | 0 | other); |
562 | 0 | } |
563 | | |
564 | 0 | static AffineExpr simplifyMod(AffineExpr lhs, AffineExpr rhs) { |
565 | 0 | auto lhsConst = lhs.dyn_cast<AffineConstantExpr>(); |
566 | 0 | auto rhsConst = rhs.dyn_cast<AffineConstantExpr>(); |
567 | 0 |
|
568 | 0 | // mod w.r.t zero or negative numbers is undefined and preserved as is. |
569 | 0 | if (!rhsConst || rhsConst.getValue() < 1) |
570 | 0 | return nullptr; |
571 | 0 | |
572 | 0 | if (lhsConst) |
573 | 0 | return getAffineConstantExpr(mod(lhsConst.getValue(), rhsConst.getValue()), |
574 | 0 | lhs.getContext()); |
575 | 0 | |
576 | 0 | // Fold modulo of an expression that is known to be a multiple of a constant |
577 | 0 | // to zero if that constant is a multiple of the modulo factor. Eg: (i * 128) |
578 | 0 | // mod 64 is folded to 0, and less trivially, (i*(j*4*(k*32))) mod 128 = 0. |
579 | 0 | if (lhs.getLargestKnownDivisor() % rhsConst.getValue() == 0) |
580 | 0 | return getAffineConstantExpr(0, lhs.getContext()); |
581 | 0 | |
582 | 0 | // Simplify (expr1 + expr2) mod divConst when either expr1 or expr2 is |
583 | 0 | // known to be a multiple of divConst. |
584 | 0 | auto lBin = lhs.dyn_cast<AffineBinaryOpExpr>(); |
585 | 0 | if (lBin && lBin.getKind() == AffineExprKind::Add) { |
586 | 0 | int64_t llhsDiv = lBin.getLHS().getLargestKnownDivisor(); |
587 | 0 | int64_t lrhsDiv = lBin.getRHS().getLargestKnownDivisor(); |
588 | 0 | // rhsConst is known to be a positive constant. |
589 | 0 | if (llhsDiv % rhsConst.getValue() == 0) |
590 | 0 | return lBin.getRHS() % rhsConst.getValue(); |
591 | 0 | if (lrhsDiv % rhsConst.getValue() == 0) |
592 | 0 | return lBin.getLHS() % rhsConst.getValue(); |
593 | 0 | } |
594 | 0 | |
595 | 0 | return nullptr; |
596 | 0 | } |
597 | | |
598 | 0 | AffineExpr AffineExpr::operator%(uint64_t v) const { |
599 | 0 | return *this % getAffineConstantExpr(v, getContext()); |
600 | 0 | } |
601 | 0 | AffineExpr AffineExpr::operator%(AffineExpr other) const { |
602 | 0 | if (auto simplified = simplifyMod(*this, other)) |
603 | 0 | return simplified; |
604 | 0 | |
605 | 0 | StorageUniquer &uniquer = getContext()->getAffineUniquer(); |
606 | 0 | return uniquer.get<AffineBinaryOpExprStorage>( |
607 | 0 | /*initFn=*/{}, static_cast<unsigned>(AffineExprKind::Mod), *this, other); |
608 | 0 | } |
609 | | |
610 | 0 | AffineExpr AffineExpr::compose(AffineMap map) const { |
611 | 0 | SmallVector<AffineExpr, 8> dimReplacements(map.getResults().begin(), |
612 | 0 | map.getResults().end()); |
613 | 0 | return replaceDimsAndSymbols(dimReplacements, {}); |
614 | 0 | } |
615 | 0 | raw_ostream &mlir::operator<<(raw_ostream &os, AffineExpr expr) { |
616 | 0 | expr.print(os); |
617 | 0 | return os; |
618 | 0 | } |
619 | | |
620 | | /// Constructs an affine expression from a flat ArrayRef. If there are local |
621 | | /// identifiers (neither dimensional nor symbolic) that appear in the sum of |
622 | | /// products expression, `localExprs` is expected to have the AffineExpr |
623 | | /// for it, and is substituted into. The ArrayRef `flatExprs` is expected to be |
624 | | /// in the format [dims, symbols, locals, constant term]. |
625 | | AffineExpr mlir::getAffineExprFromFlatForm(ArrayRef<int64_t> flatExprs, |
626 | | unsigned numDims, |
627 | | unsigned numSymbols, |
628 | | ArrayRef<AffineExpr> localExprs, |
629 | 0 | MLIRContext *context) { |
630 | 0 | // Assert expected numLocals = flatExprs.size() - numDims - numSymbols - 1. |
631 | 0 | assert(flatExprs.size() - numDims - numSymbols - 1 == localExprs.size() && |
632 | 0 | "unexpected number of local expressions"); |
633 | 0 |
|
634 | 0 | auto expr = getAffineConstantExpr(0, context); |
635 | 0 | // Dimensions and symbols. |
636 | 0 | for (unsigned j = 0; j < numDims + numSymbols; j++) { |
637 | 0 | if (flatExprs[j] == 0) |
638 | 0 | continue; |
639 | 0 | auto id = j < numDims ? getAffineDimExpr(j, context) |
640 | 0 | : getAffineSymbolExpr(j - numDims, context); |
641 | 0 | expr = expr + id * flatExprs[j]; |
642 | 0 | } |
643 | 0 |
|
644 | 0 | // Local identifiers. |
645 | 0 | for (unsigned j = numDims + numSymbols, e = flatExprs.size() - 1; j < e; |
646 | 0 | j++) { |
647 | 0 | if (flatExprs[j] == 0) |
648 | 0 | continue; |
649 | 0 | auto term = localExprs[j - numDims - numSymbols] * flatExprs[j]; |
650 | 0 | expr = expr + term; |
651 | 0 | } |
652 | 0 |
|
653 | 0 | // Constant term. |
654 | 0 | int64_t constTerm = flatExprs[flatExprs.size() - 1]; |
655 | 0 | if (constTerm != 0) |
656 | 0 | expr = expr + constTerm; |
657 | 0 | return expr; |
658 | 0 | } |
659 | | |
660 | | SimpleAffineExprFlattener::SimpleAffineExprFlattener(unsigned numDims, |
661 | | unsigned numSymbols) |
662 | 0 | : numDims(numDims), numSymbols(numSymbols), numLocals(0) { |
663 | 0 | operandExprStack.reserve(8); |
664 | 0 | } |
665 | | |
666 | 0 | void SimpleAffineExprFlattener::visitMulExpr(AffineBinaryOpExpr expr) { |
667 | 0 | assert(operandExprStack.size() >= 2); |
668 | 0 | // This is a pure affine expr; the RHS will be a constant. |
669 | 0 | assert(expr.getRHS().isa<AffineConstantExpr>()); |
670 | 0 | // Get the RHS constant. |
671 | 0 | auto rhsConst = operandExprStack.back()[getConstantIndex()]; |
672 | 0 | operandExprStack.pop_back(); |
673 | 0 | // Update the LHS in place instead of pop and push. |
674 | 0 | auto &lhs = operandExprStack.back(); |
675 | 0 | for (unsigned i = 0, e = lhs.size(); i < e; i++) { |
676 | 0 | lhs[i] *= rhsConst; |
677 | 0 | } |
678 | 0 | } |
679 | | |
680 | 0 | void SimpleAffineExprFlattener::visitAddExpr(AffineBinaryOpExpr expr) { |
681 | 0 | assert(operandExprStack.size() >= 2); |
682 | 0 | const auto &rhs = operandExprStack.back(); |
683 | 0 | auto &lhs = operandExprStack[operandExprStack.size() - 2]; |
684 | 0 | assert(lhs.size() == rhs.size()); |
685 | 0 | // Update the LHS in place. |
686 | 0 | for (unsigned i = 0, e = rhs.size(); i < e; i++) { |
687 | 0 | lhs[i] += rhs[i]; |
688 | 0 | } |
689 | 0 | // Pop off the RHS. |
690 | 0 | operandExprStack.pop_back(); |
691 | 0 | } |
692 | | |
693 | | // |
694 | | // t = expr mod c <=> t = expr - c*q and c*q <= expr <= c*q + c - 1 |
695 | | // |
696 | | // A mod expression "expr mod c" is thus flattened by introducing a new local |
697 | | // variable q (= expr floordiv c), such that expr mod c is replaced with |
698 | | // 'expr - c * q' and c * q <= expr <= c * q + c - 1 are added to localVarCst. |
699 | 0 | void SimpleAffineExprFlattener::visitModExpr(AffineBinaryOpExpr expr) { |
700 | 0 | assert(operandExprStack.size() >= 2); |
701 | 0 | // This is a pure affine expr; the RHS will be a constant. |
702 | 0 | assert(expr.getRHS().isa<AffineConstantExpr>()); |
703 | 0 | auto rhsConst = operandExprStack.back()[getConstantIndex()]; |
704 | 0 | operandExprStack.pop_back(); |
705 | 0 | auto &lhs = operandExprStack.back(); |
706 | 0 | // TODO(bondhugula): handle modulo by zero case when this issue is fixed |
707 | 0 | // at the other places in the IR. |
708 | 0 | assert(rhsConst > 0 && "RHS constant has to be positive"); |
709 | 0 |
|
710 | 0 | // Check if the LHS expression is a multiple of modulo factor. |
711 | 0 | unsigned i, e; |
712 | 0 | for (i = 0, e = lhs.size(); i < e; i++) |
713 | 0 | if (lhs[i] % rhsConst != 0) |
714 | 0 | break; |
715 | 0 | // If yes, modulo expression here simplifies to zero. |
716 | 0 | if (i == lhs.size()) { |
717 | 0 | std::fill(lhs.begin(), lhs.end(), 0); |
718 | 0 | return; |
719 | 0 | } |
720 | 0 | |
721 | 0 | // Add a local variable for the quotient, i.e., expr % c is replaced by |
722 | 0 | // (expr - q * c) where q = expr floordiv c. Do this while canceling out |
723 | 0 | // the GCD of expr and c. |
724 | 0 | SmallVector<int64_t, 8> floorDividend(lhs); |
725 | 0 | uint64_t gcd = rhsConst; |
726 | 0 | for (unsigned i = 0, e = lhs.size(); i < e; i++) |
727 | 0 | gcd = llvm::GreatestCommonDivisor64(gcd, std::abs(lhs[i])); |
728 | 0 | // Simplify the numerator and the denominator. |
729 | 0 | if (gcd != 1) { |
730 | 0 | for (unsigned i = 0, e = floorDividend.size(); i < e; i++) |
731 | 0 | floorDividend[i] = floorDividend[i] / static_cast<int64_t>(gcd); |
732 | 0 | } |
733 | 0 | int64_t floorDivisor = rhsConst / static_cast<int64_t>(gcd); |
734 | 0 |
|
735 | 0 | // Construct the AffineExpr form of the floordiv to store in localExprs. |
736 | 0 | MLIRContext *context = expr.getContext(); |
737 | 0 | auto dividendExpr = getAffineExprFromFlatForm( |
738 | 0 | floorDividend, numDims, numSymbols, localExprs, context); |
739 | 0 | auto divisorExpr = getAffineConstantExpr(floorDivisor, context); |
740 | 0 | auto floorDivExpr = dividendExpr.floorDiv(divisorExpr); |
741 | 0 | int loc; |
742 | 0 | if ((loc = findLocalId(floorDivExpr)) == -1) { |
743 | 0 | addLocalFloorDivId(floorDividend, floorDivisor, floorDivExpr); |
744 | 0 | // Set result at top of stack to "lhs - rhsConst * q". |
745 | 0 | lhs[getLocalVarStartIndex() + numLocals - 1] = -rhsConst; |
746 | 0 | } else { |
747 | 0 | // Reuse the existing local id. |
748 | 0 | lhs[getLocalVarStartIndex() + loc] = -rhsConst; |
749 | 0 | } |
750 | 0 | } |
751 | | |
752 | 0 | void SimpleAffineExprFlattener::visitCeilDivExpr(AffineBinaryOpExpr expr) { |
753 | 0 | visitDivExpr(expr, /*isCeil=*/true); |
754 | 0 | } |
755 | 0 | void SimpleAffineExprFlattener::visitFloorDivExpr(AffineBinaryOpExpr expr) { |
756 | 0 | visitDivExpr(expr, /*isCeil=*/false); |
757 | 0 | } |
758 | | |
759 | 0 | void SimpleAffineExprFlattener::visitDimExpr(AffineDimExpr expr) { |
760 | 0 | operandExprStack.emplace_back(SmallVector<int64_t, 32>(getNumCols(), 0)); |
761 | 0 | auto &eq = operandExprStack.back(); |
762 | 0 | assert(expr.getPosition() < numDims && "Inconsistent number of dims"); |
763 | 0 | eq[getDimStartIndex() + expr.getPosition()] = 1; |
764 | 0 | } |
765 | | |
766 | 0 | void SimpleAffineExprFlattener::visitSymbolExpr(AffineSymbolExpr expr) { |
767 | 0 | operandExprStack.emplace_back(SmallVector<int64_t, 32>(getNumCols(), 0)); |
768 | 0 | auto &eq = operandExprStack.back(); |
769 | 0 | assert(expr.getPosition() < numSymbols && "inconsistent number of symbols"); |
770 | 0 | eq[getSymbolStartIndex() + expr.getPosition()] = 1; |
771 | 0 | } |
772 | | |
773 | 0 | void SimpleAffineExprFlattener::visitConstantExpr(AffineConstantExpr expr) { |
774 | 0 | operandExprStack.emplace_back(SmallVector<int64_t, 32>(getNumCols(), 0)); |
775 | 0 | auto &eq = operandExprStack.back(); |
776 | 0 | eq[getConstantIndex()] = expr.getValue(); |
777 | 0 | } |
778 | | |
779 | | // t = expr floordiv c <=> t = q, c * q <= expr <= c * q + c - 1 |
780 | | // A floordiv is thus flattened by introducing a new local variable q, and |
781 | | // replacing that expression with 'q' while adding the constraints |
782 | | // c * q <= expr <= c * q + c - 1 to localVarCst (done by |
783 | | // FlatAffineConstraints::addLocalFloorDiv). |
784 | | // |
785 | | // A ceildiv is similarly flattened: |
786 | | // t = expr ceildiv c <=> t = (expr + c - 1) floordiv c |
787 | | void SimpleAffineExprFlattener::visitDivExpr(AffineBinaryOpExpr expr, |
788 | 0 | bool isCeil) { |
789 | 0 | assert(operandExprStack.size() >= 2); |
790 | 0 | assert(expr.getRHS().isa<AffineConstantExpr>()); |
791 | 0 |
|
792 | 0 | // This is a pure affine expr; the RHS is a positive constant. |
793 | 0 | int64_t rhsConst = operandExprStack.back()[getConstantIndex()]; |
794 | 0 | // TODO(bondhugula): handle division by zero at the same time the issue is |
795 | 0 | // fixed at other places. |
796 | 0 | assert(rhsConst > 0 && "RHS constant has to be positive"); |
797 | 0 | operandExprStack.pop_back(); |
798 | 0 | auto &lhs = operandExprStack.back(); |
799 | 0 |
|
800 | 0 | // Simplify the floordiv, ceildiv if possible by canceling out the greatest |
801 | 0 | // common divisors of the numerator and denominator. |
802 | 0 | uint64_t gcd = std::abs(rhsConst); |
803 | 0 | for (unsigned i = 0, e = lhs.size(); i < e; i++) |
804 | 0 | gcd = llvm::GreatestCommonDivisor64(gcd, std::abs(lhs[i])); |
805 | 0 | // Simplify the numerator and the denominator. |
806 | 0 | if (gcd != 1) { |
807 | 0 | for (unsigned i = 0, e = lhs.size(); i < e; i++) |
808 | 0 | lhs[i] = lhs[i] / static_cast<int64_t>(gcd); |
809 | 0 | } |
810 | 0 | int64_t divisor = rhsConst / static_cast<int64_t>(gcd); |
811 | 0 | // If the divisor becomes 1, the updated LHS is the result. (The |
812 | 0 | // divisor can't be negative since rhsConst is positive). |
813 | 0 | if (divisor == 1) |
814 | 0 | return; |
815 | 0 | |
816 | 0 | // If the divisor cannot be simplified to one, we will have to retain |
817 | 0 | // the ceil/floor expr (simplified up until here). Add an existential |
818 | 0 | // quantifier to express its result, i.e., expr1 div expr2 is replaced |
819 | 0 | // by a new identifier, q. |
820 | 0 | MLIRContext *context = expr.getContext(); |
821 | 0 | auto a = |
822 | 0 | getAffineExprFromFlatForm(lhs, numDims, numSymbols, localExprs, context); |
823 | 0 | auto b = getAffineConstantExpr(divisor, context); |
824 | 0 |
|
825 | 0 | int loc; |
826 | 0 | auto divExpr = isCeil ? a.ceilDiv(b) : a.floorDiv(b); |
827 | 0 | if ((loc = findLocalId(divExpr)) == -1) { |
828 | 0 | if (!isCeil) { |
829 | 0 | SmallVector<int64_t, 8> dividend(lhs); |
830 | 0 | addLocalFloorDivId(dividend, divisor, divExpr); |
831 | 0 | } else { |
832 | 0 | // lhs ceildiv c <=> (lhs + c - 1) floordiv c |
833 | 0 | SmallVector<int64_t, 8> dividend(lhs); |
834 | 0 | dividend.back() += divisor - 1; |
835 | 0 | addLocalFloorDivId(dividend, divisor, divExpr); |
836 | 0 | } |
837 | 0 | } |
838 | 0 | // Set the expression on stack to the local var introduced to capture the |
839 | 0 | // result of the division (floor or ceil). |
840 | 0 | std::fill(lhs.begin(), lhs.end(), 0); |
841 | 0 | if (loc == -1) |
842 | 0 | lhs[getLocalVarStartIndex() + numLocals - 1] = 1; |
843 | 0 | else |
844 | 0 | lhs[getLocalVarStartIndex() + loc] = 1; |
845 | 0 | } |
846 | | |
847 | | // Add a local identifier (needed to flatten a mod, floordiv, ceildiv expr). |
848 | | // The local identifier added is always a floordiv of a pure add/mul affine |
849 | | // function of other identifiers, coefficients of which are specified in |
850 | | // dividend and with respect to a positive constant divisor. localExpr is the |
851 | | // simplified tree expression (AffineExpr) corresponding to the quantifier. |
852 | | void SimpleAffineExprFlattener::addLocalFloorDivId(ArrayRef<int64_t> dividend, |
853 | | int64_t divisor, |
854 | | AffineExpr localExpr) { |
855 | | assert(divisor > 0 && "positive constant divisor expected"); |
856 | | for (auto &subExpr : operandExprStack) |
857 | | subExpr.insert(subExpr.begin() + getLocalVarStartIndex() + numLocals, 0); |
858 | | localExprs.push_back(localExpr); |
859 | | numLocals++; |
860 | | // dividend and divisor are not used here; an override of this method uses it. |
861 | | } |
862 | | |
863 | 0 | int SimpleAffineExprFlattener::findLocalId(AffineExpr localExpr) { |
864 | 0 | SmallVectorImpl<AffineExpr>::iterator it; |
865 | 0 | if ((it = llvm::find(localExprs, localExpr)) == localExprs.end()) |
866 | 0 | return -1; |
867 | 0 | return it - localExprs.begin(); |
868 | 0 | } |
869 | | |
870 | | /// Simplify the affine expression by flattening it and reconstructing it. |
871 | | AffineExpr mlir::simplifyAffineExpr(AffineExpr expr, unsigned numDims, |
872 | 0 | unsigned numSymbols) { |
873 | 0 | // TODO(bondhugula): only pure affine for now. The simplification here can |
874 | 0 | // be extended to semi-affine maps in the future. |
875 | 0 | if (!expr.isPureAffine()) |
876 | 0 | return expr; |
877 | 0 | |
878 | 0 | SimpleAffineExprFlattener flattener(numDims, numSymbols); |
879 | 0 | flattener.walkPostOrder(expr); |
880 | 0 | ArrayRef<int64_t> flattenedExpr = flattener.operandExprStack.back(); |
881 | 0 | auto simplifiedExpr = |
882 | 0 | getAffineExprFromFlatForm(flattenedExpr, numDims, numSymbols, |
883 | 0 | flattener.localExprs, expr.getContext()); |
884 | 0 | flattener.operandExprStack.pop_back(); |
885 | 0 | assert(flattener.operandExprStack.empty()); |
886 | 0 |
|
887 | 0 | return simplifiedExpr; |
888 | 0 | } |