1 | //===---- Delinearization.cpp - MultiDimensional Index Delinearization ----===// |
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 implements an analysis pass that tries to delinearize all GEP |
10 | // instructions in all loops using the SCEV analysis functionality. This pass is |
11 | // only used for testing purposes: if your pass needs delinearization, please |
12 | // use the on-demand SCEVAddRecExpr::delinearize() function. |
13 | // |
14 | //===----------------------------------------------------------------------===// |
15 | |
16 | #include "llvm/Analysis/Delinearization.h" |
17 | #include "llvm/Analysis/LoopInfo.h" |
18 | #include "llvm/Analysis/Passes.h" |
19 | #include "llvm/Analysis/ScalarEvolution.h" |
20 | #include "llvm/Analysis/ScalarEvolutionDivision.h" |
21 | #include "llvm/Analysis/ScalarEvolutionExpressions.h" |
22 | #include "llvm/IR/Constants.h" |
23 | #include "llvm/IR/DerivedTypes.h" |
24 | #include "llvm/IR/Function.h" |
25 | #include "llvm/IR/InstIterator.h" |
26 | #include "llvm/IR/Instructions.h" |
27 | #include "llvm/IR/PassManager.h" |
28 | #include "llvm/Support/Debug.h" |
29 | #include "llvm/Support/raw_ostream.h" |
30 | |
31 | using namespace llvm; |
32 | |
33 | #define DL_NAME "delinearize" |
34 | #define DEBUG_TYPE DL_NAME |
35 | |
36 | // Return true when S contains at least an undef value. |
37 | static inline bool containsUndefs(const SCEV *S) { |
38 | return SCEVExprContains(Root: S, Pred: [](const SCEV *S) { |
39 | if (const auto *SU = dyn_cast<SCEVUnknown>(Val: S)) |
40 | return isa<UndefValue>(Val: SU->getValue()); |
41 | return false; |
42 | }); |
43 | } |
44 | |
45 | namespace { |
46 | |
47 | // Collect all steps of SCEV expressions. |
48 | struct SCEVCollectStrides { |
49 | ScalarEvolution &SE; |
50 | SmallVectorImpl<const SCEV *> &Strides; |
51 | |
52 | SCEVCollectStrides(ScalarEvolution &SE, SmallVectorImpl<const SCEV *> &S) |
53 | : SE(SE), Strides(S) {} |
54 | |
55 | bool follow(const SCEV *S) { |
56 | if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Val: S)) |
57 | Strides.push_back(Elt: AR->getStepRecurrence(SE)); |
58 | return true; |
59 | } |
60 | |
61 | bool isDone() const { return false; } |
62 | }; |
63 | |
64 | // Collect all SCEVUnknown and SCEVMulExpr expressions. |
65 | struct SCEVCollectTerms { |
66 | SmallVectorImpl<const SCEV *> &Terms; |
67 | |
68 | SCEVCollectTerms(SmallVectorImpl<const SCEV *> &T) : Terms(T) {} |
69 | |
70 | bool follow(const SCEV *S) { |
71 | if (isa<SCEVUnknown>(Val: S) || isa<SCEVMulExpr>(Val: S) || |
72 | isa<SCEVSignExtendExpr>(Val: S)) { |
73 | if (!containsUndefs(S)) |
74 | Terms.push_back(Elt: S); |
75 | |
76 | // Stop recursion: once we collected a term, do not walk its operands. |
77 | return false; |
78 | } |
79 | |
80 | // Keep looking. |
81 | return true; |
82 | } |
83 | |
84 | bool isDone() const { return false; } |
85 | }; |
86 | |
87 | // Check if a SCEV contains an AddRecExpr. |
88 | struct SCEVHasAddRec { |
89 | bool &ContainsAddRec; |
90 | |
91 | SCEVHasAddRec(bool &ContainsAddRec) : ContainsAddRec(ContainsAddRec) { |
92 | ContainsAddRec = false; |
93 | } |
94 | |
95 | bool follow(const SCEV *S) { |
96 | if (isa<SCEVAddRecExpr>(Val: S)) { |
97 | ContainsAddRec = true; |
98 | |
99 | // Stop recursion: once we collected a term, do not walk its operands. |
100 | return false; |
101 | } |
102 | |
103 | // Keep looking. |
104 | return true; |
105 | } |
106 | |
107 | bool isDone() const { return false; } |
108 | }; |
109 | |
110 | // Find factors that are multiplied with an expression that (possibly as a |
111 | // subexpression) contains an AddRecExpr. In the expression: |
112 | // |
113 | // 8 * (100 + %p * %q * (%a + {0, +, 1}_loop)) |
114 | // |
115 | // "%p * %q" are factors multiplied by the expression "(%a + {0, +, 1}_loop)" |
116 | // that contains the AddRec {0, +, 1}_loop. %p * %q are likely to be array size |
117 | // parameters as they form a product with an induction variable. |
118 | // |
119 | // This collector expects all array size parameters to be in the same MulExpr. |
120 | // It might be necessary to later add support for collecting parameters that are |
121 | // spread over different nested MulExpr. |
122 | struct SCEVCollectAddRecMultiplies { |
123 | SmallVectorImpl<const SCEV *> &Terms; |
124 | ScalarEvolution &SE; |
125 | |
126 | SCEVCollectAddRecMultiplies(SmallVectorImpl<const SCEV *> &T, |
127 | ScalarEvolution &SE) |
128 | : Terms(T), SE(SE) {} |
129 | |
130 | bool follow(const SCEV *S) { |
131 | if (auto *Mul = dyn_cast<SCEVMulExpr>(Val: S)) { |
132 | bool HasAddRec = false; |
133 | SmallVector<const SCEV *, 0> Operands; |
134 | for (const auto *Op : Mul->operands()) { |
135 | const SCEVUnknown *Unknown = dyn_cast<SCEVUnknown>(Val: Op); |
136 | if (Unknown && !isa<CallInst>(Val: Unknown->getValue())) { |
137 | Operands.push_back(Elt: Op); |
138 | } else if (Unknown) { |
139 | HasAddRec = true; |
140 | } else { |
141 | bool ContainsAddRec = false; |
142 | SCEVHasAddRec ContiansAddRec(ContainsAddRec); |
143 | visitAll(Root: Op, Visitor&: ContiansAddRec); |
144 | HasAddRec |= ContainsAddRec; |
145 | } |
146 | } |
147 | if (Operands.size() == 0) |
148 | return true; |
149 | |
150 | if (!HasAddRec) |
151 | return false; |
152 | |
153 | Terms.push_back(Elt: SE.getMulExpr(Ops&: Operands)); |
154 | // Stop recursion: once we collected a term, do not walk its operands. |
155 | return false; |
156 | } |
157 | |
158 | // Keep looking. |
159 | return true; |
160 | } |
161 | |
162 | bool isDone() const { return false; } |
163 | }; |
164 | |
165 | } // end anonymous namespace |
166 | |
167 | /// Find parametric terms in this SCEVAddRecExpr. We first for parameters in |
168 | /// two places: |
169 | /// 1) The strides of AddRec expressions. |
170 | /// 2) Unknowns that are multiplied with AddRec expressions. |
171 | void llvm::collectParametricTerms(ScalarEvolution &SE, const SCEV *Expr, |
172 | SmallVectorImpl<const SCEV *> &Terms) { |
173 | SmallVector<const SCEV *, 4> Strides; |
174 | SCEVCollectStrides StrideCollector(SE, Strides); |
175 | visitAll(Root: Expr, Visitor&: StrideCollector); |
176 | |
177 | LLVM_DEBUG({ |
178 | dbgs() << "Strides:\n" ; |
179 | for (const SCEV *S : Strides) |
180 | dbgs() << *S << "\n" ; |
181 | }); |
182 | |
183 | for (const SCEV *S : Strides) { |
184 | SCEVCollectTerms TermCollector(Terms); |
185 | visitAll(Root: S, Visitor&: TermCollector); |
186 | } |
187 | |
188 | LLVM_DEBUG({ |
189 | dbgs() << "Terms:\n" ; |
190 | for (const SCEV *T : Terms) |
191 | dbgs() << *T << "\n" ; |
192 | }); |
193 | |
194 | SCEVCollectAddRecMultiplies MulCollector(Terms, SE); |
195 | visitAll(Root: Expr, Visitor&: MulCollector); |
196 | } |
197 | |
198 | static bool findArrayDimensionsRec(ScalarEvolution &SE, |
199 | SmallVectorImpl<const SCEV *> &Terms, |
200 | SmallVectorImpl<const SCEV *> &Sizes) { |
201 | int Last = Terms.size() - 1; |
202 | const SCEV *Step = Terms[Last]; |
203 | |
204 | // End of recursion. |
205 | if (Last == 0) { |
206 | if (const SCEVMulExpr *M = dyn_cast<SCEVMulExpr>(Val: Step)) { |
207 | SmallVector<const SCEV *, 2> Qs; |
208 | for (const SCEV *Op : M->operands()) |
209 | if (!isa<SCEVConstant>(Val: Op)) |
210 | Qs.push_back(Elt: Op); |
211 | |
212 | Step = SE.getMulExpr(Ops&: Qs); |
213 | } |
214 | |
215 | Sizes.push_back(Elt: Step); |
216 | return true; |
217 | } |
218 | |
219 | for (const SCEV *&Term : Terms) { |
220 | // Normalize the terms before the next call to findArrayDimensionsRec. |
221 | const SCEV *Q, *R; |
222 | SCEVDivision::divide(SE, Numerator: Term, Denominator: Step, Quotient: &Q, Remainder: &R); |
223 | |
224 | // Bail out when GCD does not evenly divide one of the terms. |
225 | if (!R->isZero()) |
226 | return false; |
227 | |
228 | Term = Q; |
229 | } |
230 | |
231 | // Remove all SCEVConstants. |
232 | erase_if(C&: Terms, P: [](const SCEV *E) { return isa<SCEVConstant>(Val: E); }); |
233 | |
234 | if (Terms.size() > 0) |
235 | if (!findArrayDimensionsRec(SE, Terms, Sizes)) |
236 | return false; |
237 | |
238 | Sizes.push_back(Elt: Step); |
239 | return true; |
240 | } |
241 | |
242 | // Returns true when one of the SCEVs of Terms contains a SCEVUnknown parameter. |
243 | static inline bool containsParameters(SmallVectorImpl<const SCEV *> &Terms) { |
244 | for (const SCEV *T : Terms) |
245 | if (SCEVExprContains(Root: T, Pred: [](const SCEV *S) { return isa<SCEVUnknown>(Val: S); })) |
246 | return true; |
247 | |
248 | return false; |
249 | } |
250 | |
251 | // Return the number of product terms in S. |
252 | static inline int numberOfTerms(const SCEV *S) { |
253 | if (const SCEVMulExpr *Expr = dyn_cast<SCEVMulExpr>(Val: S)) |
254 | return Expr->getNumOperands(); |
255 | return 1; |
256 | } |
257 | |
258 | static const SCEV *removeConstantFactors(ScalarEvolution &SE, const SCEV *T) { |
259 | if (isa<SCEVConstant>(Val: T)) |
260 | return nullptr; |
261 | |
262 | if (isa<SCEVUnknown>(Val: T)) |
263 | return T; |
264 | |
265 | if (const SCEVMulExpr *M = dyn_cast<SCEVMulExpr>(Val: T)) { |
266 | SmallVector<const SCEV *, 2> Factors; |
267 | for (const SCEV *Op : M->operands()) |
268 | if (!isa<SCEVConstant>(Val: Op)) |
269 | Factors.push_back(Elt: Op); |
270 | |
271 | return SE.getMulExpr(Ops&: Factors); |
272 | } |
273 | |
274 | return T; |
275 | } |
276 | |
277 | void llvm::findArrayDimensions(ScalarEvolution &SE, |
278 | SmallVectorImpl<const SCEV *> &Terms, |
279 | SmallVectorImpl<const SCEV *> &Sizes, |
280 | const SCEV *ElementSize) { |
281 | if (Terms.size() < 1 || !ElementSize) |
282 | return; |
283 | |
284 | // Early return when Terms do not contain parameters: we do not delinearize |
285 | // non parametric SCEVs. |
286 | if (!containsParameters(Terms)) |
287 | return; |
288 | |
289 | LLVM_DEBUG({ |
290 | dbgs() << "Terms:\n" ; |
291 | for (const SCEV *T : Terms) |
292 | dbgs() << *T << "\n" ; |
293 | }); |
294 | |
295 | // Remove duplicates. |
296 | array_pod_sort(Start: Terms.begin(), End: Terms.end()); |
297 | Terms.erase(CS: llvm::unique(R&: Terms), CE: Terms.end()); |
298 | |
299 | // Put larger terms first. |
300 | llvm::sort(C&: Terms, Comp: [](const SCEV *LHS, const SCEV *RHS) { |
301 | return numberOfTerms(S: LHS) > numberOfTerms(S: RHS); |
302 | }); |
303 | |
304 | // Try to divide all terms by the element size. If term is not divisible by |
305 | // element size, proceed with the original term. |
306 | for (const SCEV *&Term : Terms) { |
307 | const SCEV *Q, *R; |
308 | SCEVDivision::divide(SE, Numerator: Term, Denominator: ElementSize, Quotient: &Q, Remainder: &R); |
309 | if (!Q->isZero()) |
310 | Term = Q; |
311 | } |
312 | |
313 | SmallVector<const SCEV *, 4> NewTerms; |
314 | |
315 | // Remove constant factors. |
316 | for (const SCEV *T : Terms) |
317 | if (const SCEV *NewT = removeConstantFactors(SE, T)) |
318 | NewTerms.push_back(Elt: NewT); |
319 | |
320 | LLVM_DEBUG({ |
321 | dbgs() << "Terms after sorting:\n" ; |
322 | for (const SCEV *T : NewTerms) |
323 | dbgs() << *T << "\n" ; |
324 | }); |
325 | |
326 | if (NewTerms.empty() || !findArrayDimensionsRec(SE, Terms&: NewTerms, Sizes)) { |
327 | Sizes.clear(); |
328 | return; |
329 | } |
330 | |
331 | // The last element to be pushed into Sizes is the size of an element. |
332 | Sizes.push_back(Elt: ElementSize); |
333 | |
334 | LLVM_DEBUG({ |
335 | dbgs() << "Sizes:\n" ; |
336 | for (const SCEV *S : Sizes) |
337 | dbgs() << *S << "\n" ; |
338 | }); |
339 | } |
340 | |
341 | void llvm::computeAccessFunctions(ScalarEvolution &SE, const SCEV *Expr, |
342 | SmallVectorImpl<const SCEV *> &Subscripts, |
343 | SmallVectorImpl<const SCEV *> &Sizes) { |
344 | // Early exit in case this SCEV is not an affine multivariate function. |
345 | if (Sizes.empty()) |
346 | return; |
347 | |
348 | if (auto *AR = dyn_cast<SCEVAddRecExpr>(Val: Expr)) |
349 | if (!AR->isAffine()) |
350 | return; |
351 | |
352 | const SCEV *Res = Expr; |
353 | int Last = Sizes.size() - 1; |
354 | for (int i = Last; i >= 0; i--) { |
355 | const SCEV *Q, *R; |
356 | SCEVDivision::divide(SE, Numerator: Res, Denominator: Sizes[i], Quotient: &Q, Remainder: &R); |
357 | |
358 | LLVM_DEBUG({ |
359 | dbgs() << "Res: " << *Res << "\n" ; |
360 | dbgs() << "Sizes[i]: " << *Sizes[i] << "\n" ; |
361 | dbgs() << "Res divided by Sizes[i]:\n" ; |
362 | dbgs() << "Quotient: " << *Q << "\n" ; |
363 | dbgs() << "Remainder: " << *R << "\n" ; |
364 | }); |
365 | |
366 | Res = Q; |
367 | |
368 | // Do not record the last subscript corresponding to the size of elements in |
369 | // the array. |
370 | if (i == Last) { |
371 | |
372 | // Bail out if the byte offset is non-zero. |
373 | if (!R->isZero()) { |
374 | Subscripts.clear(); |
375 | Sizes.clear(); |
376 | return; |
377 | } |
378 | |
379 | continue; |
380 | } |
381 | |
382 | // Record the access function for the current subscript. |
383 | Subscripts.push_back(Elt: R); |
384 | } |
385 | |
386 | // Also push in last position the remainder of the last division: it will be |
387 | // the access function of the innermost dimension. |
388 | Subscripts.push_back(Elt: Res); |
389 | |
390 | std::reverse(first: Subscripts.begin(), last: Subscripts.end()); |
391 | |
392 | LLVM_DEBUG({ |
393 | dbgs() << "Subscripts:\n" ; |
394 | for (const SCEV *S : Subscripts) |
395 | dbgs() << *S << "\n" ; |
396 | }); |
397 | } |
398 | |
399 | /// Splits the SCEV into two vectors of SCEVs representing the subscripts and |
400 | /// sizes of an array access. Returns the remainder of the delinearization that |
401 | /// is the offset start of the array. The SCEV->delinearize algorithm computes |
402 | /// the multiples of SCEV coefficients: that is a pattern matching of sub |
403 | /// expressions in the stride and base of a SCEV corresponding to the |
404 | /// computation of a GCD (greatest common divisor) of base and stride. When |
405 | /// SCEV->delinearize fails, it returns the SCEV unchanged. |
406 | /// |
407 | /// For example: when analyzing the memory access A[i][j][k] in this loop nest |
408 | /// |
409 | /// void foo(long n, long m, long o, double A[n][m][o]) { |
410 | /// |
411 | /// for (long i = 0; i < n; i++) |
412 | /// for (long j = 0; j < m; j++) |
413 | /// for (long k = 0; k < o; k++) |
414 | /// A[i][j][k] = 1.0; |
415 | /// } |
416 | /// |
417 | /// the delinearization input is the following AddRec SCEV: |
418 | /// |
419 | /// AddRec: {{{%A,+,(8 * %m * %o)}<%for.i>,+,(8 * %o)}<%for.j>,+,8}<%for.k> |
420 | /// |
421 | /// From this SCEV, we are able to say that the base offset of the access is %A |
422 | /// because it appears as an offset that does not divide any of the strides in |
423 | /// the loops: |
424 | /// |
425 | /// CHECK: Base offset: %A |
426 | /// |
427 | /// and then SCEV->delinearize determines the size of some of the dimensions of |
428 | /// the array as these are the multiples by which the strides are happening: |
429 | /// |
430 | /// CHECK: ArrayDecl[UnknownSize][%m][%o] with elements of sizeof(double) |
431 | /// bytes. |
432 | /// |
433 | /// Note that the outermost dimension remains of UnknownSize because there are |
434 | /// no strides that would help identifying the size of the last dimension: when |
435 | /// the array has been statically allocated, one could compute the size of that |
436 | /// dimension by dividing the overall size of the array by the size of the known |
437 | /// dimensions: %m * %o * 8. |
438 | /// |
439 | /// Finally delinearize provides the access functions for the array reference |
440 | /// that does correspond to A[i][j][k] of the above C testcase: |
441 | /// |
442 | /// CHECK: ArrayRef[{0,+,1}<%for.i>][{0,+,1}<%for.j>][{0,+,1}<%for.k>] |
443 | /// |
444 | /// The testcases are checking the output of a function pass: |
445 | /// DelinearizationPass that walks through all loads and stores of a function |
446 | /// asking for the SCEV of the memory access with respect to all enclosing |
447 | /// loops, calling SCEV->delinearize on that and printing the results. |
448 | void llvm::delinearize(ScalarEvolution &SE, const SCEV *Expr, |
449 | SmallVectorImpl<const SCEV *> &Subscripts, |
450 | SmallVectorImpl<const SCEV *> &Sizes, |
451 | const SCEV *ElementSize) { |
452 | // First step: collect parametric terms. |
453 | SmallVector<const SCEV *, 4> Terms; |
454 | collectParametricTerms(SE, Expr, Terms); |
455 | |
456 | if (Terms.empty()) |
457 | return; |
458 | |
459 | // Second step: find subscript sizes. |
460 | findArrayDimensions(SE, Terms, Sizes, ElementSize); |
461 | |
462 | if (Sizes.empty()) |
463 | return; |
464 | |
465 | // Third step: compute the access functions for each subscript. |
466 | computeAccessFunctions(SE, Expr, Subscripts, Sizes); |
467 | |
468 | if (Subscripts.empty()) |
469 | return; |
470 | |
471 | LLVM_DEBUG({ |
472 | dbgs() << "succeeded to delinearize " << *Expr << "\n" ; |
473 | dbgs() << "ArrayDecl[UnknownSize]" ; |
474 | for (const SCEV *S : Sizes) |
475 | dbgs() << "[" << *S << "]" ; |
476 | |
477 | dbgs() << "\nArrayRef" ; |
478 | for (const SCEV *S : Subscripts) |
479 | dbgs() << "[" << *S << "]" ; |
480 | dbgs() << "\n" ; |
481 | }); |
482 | } |
483 | |
484 | bool llvm::getIndexExpressionsFromGEP(ScalarEvolution &SE, |
485 | const GetElementPtrInst *GEP, |
486 | SmallVectorImpl<const SCEV *> &Subscripts, |
487 | SmallVectorImpl<int> &Sizes) { |
488 | assert(Subscripts.empty() && Sizes.empty() && |
489 | "Expected output lists to be empty on entry to this function." ); |
490 | assert(GEP && "getIndexExpressionsFromGEP called with a null GEP" ); |
491 | Type *Ty = nullptr; |
492 | bool DroppedFirstDim = false; |
493 | for (unsigned i = 1; i < GEP->getNumOperands(); i++) { |
494 | const SCEV *Expr = SE.getSCEV(V: GEP->getOperand(i_nocapture: i)); |
495 | if (i == 1) { |
496 | Ty = GEP->getSourceElementType(); |
497 | if (auto *Const = dyn_cast<SCEVConstant>(Val: Expr)) |
498 | if (Const->getValue()->isZero()) { |
499 | DroppedFirstDim = true; |
500 | continue; |
501 | } |
502 | Subscripts.push_back(Elt: Expr); |
503 | continue; |
504 | } |
505 | |
506 | auto *ArrayTy = dyn_cast<ArrayType>(Val: Ty); |
507 | if (!ArrayTy) { |
508 | Subscripts.clear(); |
509 | Sizes.clear(); |
510 | return false; |
511 | } |
512 | |
513 | Subscripts.push_back(Elt: Expr); |
514 | if (!(DroppedFirstDim && i == 2)) |
515 | Sizes.push_back(Elt: ArrayTy->getNumElements()); |
516 | |
517 | Ty = ArrayTy->getElementType(); |
518 | } |
519 | return !Subscripts.empty(); |
520 | } |
521 | |
522 | bool llvm::tryDelinearizeFixedSizeImpl( |
523 | ScalarEvolution *SE, Instruction *Inst, const SCEV *AccessFn, |
524 | SmallVectorImpl<const SCEV *> &Subscripts, SmallVectorImpl<int> &Sizes) { |
525 | Value *SrcPtr = getLoadStorePointerOperand(V: Inst); |
526 | |
527 | // Check the simple case where the array dimensions are fixed size. |
528 | auto *SrcGEP = dyn_cast<GetElementPtrInst>(Val: SrcPtr); |
529 | if (!SrcGEP) |
530 | return false; |
531 | |
532 | getIndexExpressionsFromGEP(SE&: *SE, GEP: SrcGEP, Subscripts, Sizes); |
533 | |
534 | // Check that the two size arrays are non-empty and equal in length and |
535 | // value. |
536 | // TODO: it would be better to let the caller to clear Subscripts, similar |
537 | // to how we handle Sizes. |
538 | if (Sizes.empty() || Subscripts.size() <= 1) { |
539 | Subscripts.clear(); |
540 | return false; |
541 | } |
542 | |
543 | // Check that for identical base pointers we do not miss index offsets |
544 | // that have been added before this GEP is applied. |
545 | Value *SrcBasePtr = SrcGEP->getOperand(i_nocapture: 0)->stripPointerCasts(); |
546 | const SCEVUnknown *SrcBase = |
547 | dyn_cast<SCEVUnknown>(Val: SE->getPointerBase(V: AccessFn)); |
548 | if (!SrcBase || SrcBasePtr != SrcBase->getValue()) { |
549 | Subscripts.clear(); |
550 | return false; |
551 | } |
552 | |
553 | assert(Subscripts.size() == Sizes.size() + 1 && |
554 | "Expected equal number of entries in the list of size and " |
555 | "subscript." ); |
556 | |
557 | return true; |
558 | } |
559 | |
560 | namespace { |
561 | |
562 | void printDelinearization(raw_ostream &O, Function *F, LoopInfo *LI, |
563 | ScalarEvolution *SE) { |
564 | O << "Delinearization on function " << F->getName() << ":\n" ; |
565 | for (Instruction &Inst : instructions(F)) { |
566 | // Only analyze loads and stores. |
567 | if (!isa<StoreInst>(Val: &Inst) && !isa<LoadInst>(Val: &Inst) && |
568 | !isa<GetElementPtrInst>(Val: &Inst)) |
569 | continue; |
570 | |
571 | const BasicBlock *BB = Inst.getParent(); |
572 | // Delinearize the memory access as analyzed in all the surrounding loops. |
573 | // Do not analyze memory accesses outside loops. |
574 | for (Loop *L = LI->getLoopFor(BB); L != nullptr; L = L->getParentLoop()) { |
575 | const SCEV *AccessFn = SE->getSCEVAtScope(V: getPointerOperand(V: &Inst), L); |
576 | |
577 | const SCEVUnknown *BasePointer = |
578 | dyn_cast<SCEVUnknown>(Val: SE->getPointerBase(V: AccessFn)); |
579 | // Do not delinearize if we cannot find the base pointer. |
580 | if (!BasePointer) |
581 | break; |
582 | AccessFn = SE->getMinusSCEV(LHS: AccessFn, RHS: BasePointer); |
583 | |
584 | O << "\n" ; |
585 | O << "Inst:" << Inst << "\n" ; |
586 | O << "In Loop with Header: " << L->getHeader()->getName() << "\n" ; |
587 | O << "AccessFunction: " << *AccessFn << "\n" ; |
588 | |
589 | SmallVector<const SCEV *, 3> Subscripts, Sizes; |
590 | delinearize(SE&: *SE, Expr: AccessFn, Subscripts, Sizes, ElementSize: SE->getElementSize(Inst: &Inst)); |
591 | if (Subscripts.size() == 0 || Sizes.size() == 0 || |
592 | Subscripts.size() != Sizes.size()) { |
593 | O << "failed to delinearize\n" ; |
594 | continue; |
595 | } |
596 | |
597 | O << "Base offset: " << *BasePointer << "\n" ; |
598 | O << "ArrayDecl[UnknownSize]" ; |
599 | int Size = Subscripts.size(); |
600 | for (int i = 0; i < Size - 1; i++) |
601 | O << "[" << *Sizes[i] << "]" ; |
602 | O << " with elements of " << *Sizes[Size - 1] << " bytes.\n" ; |
603 | |
604 | O << "ArrayRef" ; |
605 | for (int i = 0; i < Size; i++) |
606 | O << "[" << *Subscripts[i] << "]" ; |
607 | O << "\n" ; |
608 | } |
609 | } |
610 | } |
611 | |
612 | } // end anonymous namespace |
613 | |
614 | DelinearizationPrinterPass::DelinearizationPrinterPass(raw_ostream &OS) |
615 | : OS(OS) {} |
616 | PreservedAnalyses DelinearizationPrinterPass::run(Function &F, |
617 | FunctionAnalysisManager &AM) { |
618 | printDelinearization(O&: OS, F: &F, LI: &AM.getResult<LoopAnalysis>(IR&: F), |
619 | SE: &AM.getResult<ScalarEvolutionAnalysis>(IR&: F)); |
620 | return PreservedAnalyses::all(); |
621 | } |
622 | |