1//===-- LegalizeTypes.cpp - Common code for DAG type legalizer ------------===//
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 the SelectionDAG::LegalizeTypes method. It transforms
10// an arbitrary well-formed SelectionDAG to only consist of legal types. This
11// is common code shared among the LegalizeTypes*.cpp files.
12//
13//===----------------------------------------------------------------------===//
14
15#include "LegalizeTypes.h"
16#include "llvm/ADT/SetVector.h"
17#include "llvm/IR/DataLayout.h"
18#include "llvm/Support/CommandLine.h"
19#include "llvm/Support/ErrorHandling.h"
20#include "llvm/Support/raw_ostream.h"
21using namespace llvm;
22
23#define DEBUG_TYPE "legalize-types"
24
25static cl::opt<bool>
26EnableExpensiveChecks("enable-legalize-types-checking", cl::Hidden);
27
28/// Do extensive, expensive, basic correctness checking.
29void DAGTypeLegalizer::PerformExpensiveChecks() {
30 // If a node is not processed, then none of its values should be mapped by any
31 // of PromotedIntegers, ExpandedIntegers, ..., ReplacedValues.
32
33 // If a node is processed, then each value with an illegal type must be mapped
34 // by exactly one of PromotedIntegers, ExpandedIntegers, ..., ReplacedValues.
35 // Values with a legal type may be mapped by ReplacedValues, but not by any of
36 // the other maps.
37
38 // Note that these invariants may not hold momentarily when processing a node:
39 // the node being processed may be put in a map before being marked Processed.
40
41 // Note that it is possible to have nodes marked NewNode in the DAG. This can
42 // occur in two ways. Firstly, a node may be created during legalization but
43 // never passed to the legalization core. This is usually due to the implicit
44 // folding that occurs when using the DAG.getNode operators. Secondly, a new
45 // node may be passed to the legalization core, but when analyzed may morph
46 // into a different node, leaving the original node as a NewNode in the DAG.
47 // A node may morph if one of its operands changes during analysis. Whether
48 // it actually morphs or not depends on whether, after updating its operands,
49 // it is equivalent to an existing node: if so, it morphs into that existing
50 // node (CSE). An operand can change during analysis if the operand is a new
51 // node that morphs, or it is a processed value that was mapped to some other
52 // value (as recorded in ReplacedValues) in which case the operand is turned
53 // into that other value. If a node morphs then the node it morphed into will
54 // be used instead of it for legalization, however the original node continues
55 // to live on in the DAG.
56 // The conclusion is that though there may be nodes marked NewNode in the DAG,
57 // all uses of such nodes are also marked NewNode: the result is a fungus of
58 // NewNodes growing on top of the useful nodes, and perhaps using them, but
59 // not used by them.
60
61 // If a value is mapped by ReplacedValues, then it must have no uses, except
62 // by nodes marked NewNode (see above).
63
64 // The final node obtained by mapping by ReplacedValues is not marked NewNode.
65 // Note that ReplacedValues should be applied iteratively.
66
67 // Note that the ReplacedValues map may also map deleted nodes (by iterating
68 // over the DAG we never dereference deleted nodes). This means that it may
69 // also map nodes marked NewNode if the deallocated memory was reallocated as
70 // another node, and that new node was not seen by the LegalizeTypes machinery
71 // (for example because it was created but not used). In general, we cannot
72 // distinguish between new nodes and deleted nodes.
73 SmallVector<SDNode*, 16> NewNodes;
74 for (SDNode &Node : DAG.allnodes()) {
75 // Remember nodes marked NewNode - they are subject to extra checking below.
76 if (Node.getNodeId() == NewNode)
77 NewNodes.push_back(Elt: &Node);
78
79 for (unsigned i = 0, e = Node.getNumValues(); i != e; ++i) {
80 SDValue Res(&Node, i);
81 bool Failed = false;
82 // Don't create a value in map.
83 auto ResId = ValueToIdMap.lookup(Val: Res);
84
85 unsigned Mapped = 0;
86 if (ResId) {
87 auto I = ReplacedValues.find(Val: ResId);
88 if (I != ReplacedValues.end()) {
89 Mapped |= 1;
90 // Check that remapped values are only used by nodes marked NewNode.
91 for (SDNode::use_iterator UI = Node.use_begin(), UE = Node.use_end();
92 UI != UE; ++UI)
93 if (UI.getUse().getResNo() == i)
94 assert(UI->getNodeId() == NewNode &&
95 "Remapped value has non-trivial use!");
96
97 // Check that the final result of applying ReplacedValues is not
98 // marked NewNode.
99 auto NewValId = I->second;
100 I = ReplacedValues.find(Val: NewValId);
101 while (I != ReplacedValues.end()) {
102 NewValId = I->second;
103 I = ReplacedValues.find(Val: NewValId);
104 }
105 SDValue NewVal = getSDValue(Id&: NewValId);
106 (void)NewVal;
107 assert(NewVal.getNode()->getNodeId() != NewNode &&
108 "ReplacedValues maps to a new node!");
109 }
110 if (PromotedIntegers.count(Val: ResId))
111 Mapped |= 2;
112 if (SoftenedFloats.count(Val: ResId))
113 Mapped |= 4;
114 if (ScalarizedVectors.count(Val: ResId))
115 Mapped |= 8;
116 if (ExpandedIntegers.count(Val: ResId))
117 Mapped |= 16;
118 if (ExpandedFloats.count(Val: ResId))
119 Mapped |= 32;
120 if (SplitVectors.count(Val: ResId))
121 Mapped |= 64;
122 if (WidenedVectors.count(Val: ResId))
123 Mapped |= 128;
124 if (PromotedFloats.count(Val: ResId))
125 Mapped |= 256;
126 if (SoftPromotedHalfs.count(Val: ResId))
127 Mapped |= 512;
128 }
129
130 if (Node.getNodeId() != Processed) {
131 // Since we allow ReplacedValues to map deleted nodes, it may map nodes
132 // marked NewNode too, since a deleted node may have been reallocated as
133 // another node that has not been seen by the LegalizeTypes machinery.
134 if ((Node.getNodeId() == NewNode && Mapped > 1) ||
135 (Node.getNodeId() != NewNode && Mapped != 0)) {
136 dbgs() << "Unprocessed value in a map!";
137 Failed = true;
138 }
139 } else if (isTypeLegal(VT: Res.getValueType()) || IgnoreNodeResults(N: &Node)) {
140 if (Mapped > 1) {
141 dbgs() << "Value with legal type was transformed!";
142 Failed = true;
143 }
144 } else {
145 if (Mapped == 0) {
146 SDValue NodeById = IdToValueMap.lookup(Val: ResId);
147 // It is possible the node has been remapped to another node and had
148 // its Id updated in the Value to Id table. The node it remapped to
149 // may not have been processed yet. Look up the Id in the Id to Value
150 // table and re-check the Processed state. If the node hasn't been
151 // remapped we'll get the same state as we got earlier.
152 if (NodeById->getNodeId() == Processed) {
153 dbgs() << "Processed value not in any map!";
154 Failed = true;
155 }
156 } else if (Mapped & (Mapped - 1)) {
157 dbgs() << "Value in multiple maps!";
158 Failed = true;
159 }
160 }
161
162 if (Failed) {
163 if (Mapped & 1)
164 dbgs() << " ReplacedValues";
165 if (Mapped & 2)
166 dbgs() << " PromotedIntegers";
167 if (Mapped & 4)
168 dbgs() << " SoftenedFloats";
169 if (Mapped & 8)
170 dbgs() << " ScalarizedVectors";
171 if (Mapped & 16)
172 dbgs() << " ExpandedIntegers";
173 if (Mapped & 32)
174 dbgs() << " ExpandedFloats";
175 if (Mapped & 64)
176 dbgs() << " SplitVectors";
177 if (Mapped & 128)
178 dbgs() << " WidenedVectors";
179 if (Mapped & 256)
180 dbgs() << " PromotedFloats";
181 if (Mapped & 512)
182 dbgs() << " SoftPromoteHalfs";
183 dbgs() << "\n";
184 llvm_unreachable(nullptr);
185 }
186 }
187 }
188
189#ifndef NDEBUG
190 // Checked that NewNodes are only used by other NewNodes.
191 for (unsigned i = 0, e = NewNodes.size(); i != e; ++i) {
192 SDNode *N = NewNodes[i];
193 for (SDNode *U : N->uses())
194 assert(U->getNodeId() == NewNode && "NewNode used by non-NewNode!");
195 }
196#endif
197}
198
199/// This is the main entry point for the type legalizer. This does a top-down
200/// traversal of the dag, legalizing types as it goes. Returns "true" if it made
201/// any changes.
202bool DAGTypeLegalizer::run() {
203 bool Changed = false;
204
205 // Create a dummy node (which is not added to allnodes), that adds a reference
206 // to the root node, preventing it from being deleted, and tracking any
207 // changes of the root.
208 HandleSDNode Dummy(DAG.getRoot());
209 Dummy.setNodeId(Unanalyzed);
210
211 // The root of the dag may dangle to deleted nodes until the type legalizer is
212 // done. Set it to null to avoid confusion.
213 DAG.setRoot(SDValue());
214
215 // Walk all nodes in the graph, assigning them a NodeId of 'ReadyToProcess'
216 // (and remembering them) if they are leaves and assigning 'Unanalyzed' if
217 // non-leaves.
218 for (SDNode &Node : DAG.allnodes()) {
219 if (Node.getNumOperands() == 0) {
220 Node.setNodeId(ReadyToProcess);
221 Worklist.push_back(Elt: &Node);
222 } else {
223 Node.setNodeId(Unanalyzed);
224 }
225 }
226
227 // Now that we have a set of nodes to process, handle them all.
228 while (!Worklist.empty()) {
229#ifndef EXPENSIVE_CHECKS
230 if (EnableExpensiveChecks)
231#endif
232 PerformExpensiveChecks();
233
234 SDNode *N = Worklist.pop_back_val();
235 assert(N->getNodeId() == ReadyToProcess &&
236 "Node should be ready if on worklist!");
237
238 LLVM_DEBUG(dbgs() << "\nLegalizing node: "; N->dump(&DAG));
239 if (IgnoreNodeResults(N)) {
240 LLVM_DEBUG(dbgs() << "Ignoring node results\n");
241 goto ScanOperands;
242 }
243
244 // Scan the values produced by the node, checking to see if any result
245 // types are illegal.
246 for (unsigned i = 0, NumResults = N->getNumValues(); i < NumResults; ++i) {
247 EVT ResultVT = N->getValueType(ResNo: i);
248 LLVM_DEBUG(dbgs() << "Analyzing result type: " << ResultVT << "\n");
249 switch (getTypeAction(VT: ResultVT)) {
250 case TargetLowering::TypeLegal:
251 LLVM_DEBUG(dbgs() << "Legal result type\n");
252 break;
253 case TargetLowering::TypeScalarizeScalableVector:
254 report_fatal_error(
255 reason: "Scalarization of scalable vectors is not supported.");
256 // The following calls must take care of *all* of the node's results,
257 // not just the illegal result they were passed (this includes results
258 // with a legal type). Results can be remapped using ReplaceValueWith,
259 // or their promoted/expanded/etc values registered in PromotedIntegers,
260 // ExpandedIntegers etc.
261 case TargetLowering::TypePromoteInteger:
262 PromoteIntegerResult(N, ResNo: i);
263 Changed = true;
264 goto NodeDone;
265 case TargetLowering::TypeExpandInteger:
266 ExpandIntegerResult(N, ResNo: i);
267 Changed = true;
268 goto NodeDone;
269 case TargetLowering::TypeSoftenFloat:
270 SoftenFloatResult(N, ResNo: i);
271 Changed = true;
272 goto NodeDone;
273 case TargetLowering::TypeExpandFloat:
274 ExpandFloatResult(N, ResNo: i);
275 Changed = true;
276 goto NodeDone;
277 case TargetLowering::TypeScalarizeVector:
278 ScalarizeVectorResult(N, ResNo: i);
279 Changed = true;
280 goto NodeDone;
281 case TargetLowering::TypeSplitVector:
282 SplitVectorResult(N, ResNo: i);
283 Changed = true;
284 goto NodeDone;
285 case TargetLowering::TypeWidenVector:
286 WidenVectorResult(N, ResNo: i);
287 Changed = true;
288 goto NodeDone;
289 case TargetLowering::TypePromoteFloat:
290 PromoteFloatResult(N, ResNo: i);
291 Changed = true;
292 goto NodeDone;
293 case TargetLowering::TypeSoftPromoteHalf:
294 SoftPromoteHalfResult(N, ResNo: i);
295 Changed = true;
296 goto NodeDone;
297 }
298 }
299
300ScanOperands:
301 // Scan the operand list for the node, handling any nodes with operands that
302 // are illegal.
303 {
304 unsigned NumOperands = N->getNumOperands();
305 bool NeedsReanalyzing = false;
306 unsigned i;
307 for (i = 0; i != NumOperands; ++i) {
308 if (IgnoreNodeResults(N: N->getOperand(Num: i).getNode()))
309 continue;
310
311 const auto &Op = N->getOperand(Num: i);
312 LLVM_DEBUG(dbgs() << "Analyzing operand: "; Op.dump(&DAG));
313 EVT OpVT = Op.getValueType();
314 switch (getTypeAction(VT: OpVT)) {
315 case TargetLowering::TypeLegal:
316 LLVM_DEBUG(dbgs() << "Legal operand\n");
317 continue;
318 case TargetLowering::TypeScalarizeScalableVector:
319 report_fatal_error(
320 reason: "Scalarization of scalable vectors is not supported.");
321 // The following calls must either replace all of the node's results
322 // using ReplaceValueWith, and return "false"; or update the node's
323 // operands in place, and return "true".
324 case TargetLowering::TypePromoteInteger:
325 NeedsReanalyzing = PromoteIntegerOperand(N, OpNo: i);
326 Changed = true;
327 break;
328 case TargetLowering::TypeExpandInteger:
329 NeedsReanalyzing = ExpandIntegerOperand(N, OpNo: i);
330 Changed = true;
331 break;
332 case TargetLowering::TypeSoftenFloat:
333 NeedsReanalyzing = SoftenFloatOperand(N, OpNo: i);
334 Changed = true;
335 break;
336 case TargetLowering::TypeExpandFloat:
337 NeedsReanalyzing = ExpandFloatOperand(N, OpNo: i);
338 Changed = true;
339 break;
340 case TargetLowering::TypeScalarizeVector:
341 NeedsReanalyzing = ScalarizeVectorOperand(N, OpNo: i);
342 Changed = true;
343 break;
344 case TargetLowering::TypeSplitVector:
345 NeedsReanalyzing = SplitVectorOperand(N, OpNo: i);
346 Changed = true;
347 break;
348 case TargetLowering::TypeWidenVector:
349 NeedsReanalyzing = WidenVectorOperand(N, OpNo: i);
350 Changed = true;
351 break;
352 case TargetLowering::TypePromoteFloat:
353 NeedsReanalyzing = PromoteFloatOperand(N, OpNo: i);
354 Changed = true;
355 break;
356 case TargetLowering::TypeSoftPromoteHalf:
357 NeedsReanalyzing = SoftPromoteHalfOperand(N, OpNo: i);
358 Changed = true;
359 break;
360 }
361 break;
362 }
363
364 // The sub-method updated N in place. Check to see if any operands are new,
365 // and if so, mark them. If the node needs revisiting, don't add all users
366 // to the worklist etc.
367 if (NeedsReanalyzing) {
368 assert(N->getNodeId() == ReadyToProcess && "Node ID recalculated?");
369
370 N->setNodeId(NewNode);
371 // Recompute the NodeId and correct processed operands, adding the node to
372 // the worklist if ready.
373 SDNode *M = AnalyzeNewNode(N);
374 if (M == N)
375 // The node didn't morph - nothing special to do, it will be revisited.
376 continue;
377
378 // The node morphed - this is equivalent to legalizing by replacing every
379 // value of N with the corresponding value of M. So do that now.
380 assert(N->getNumValues() == M->getNumValues() &&
381 "Node morphing changed the number of results!");
382 for (unsigned i = 0, e = N->getNumValues(); i != e; ++i)
383 // Replacing the value takes care of remapping the new value.
384 ReplaceValueWith(From: SDValue(N, i), To: SDValue(M, i));
385 assert(N->getNodeId() == NewNode && "Unexpected node state!");
386 // The node continues to live on as part of the NewNode fungus that
387 // grows on top of the useful nodes. Nothing more needs to be done
388 // with it - move on to the next node.
389 continue;
390 }
391
392 if (i == NumOperands) {
393 LLVM_DEBUG(dbgs() << "Legally typed node: "; N->dump(&DAG));
394 }
395 }
396NodeDone:
397
398 // If we reach here, the node was processed, potentially creating new nodes.
399 // Mark it as processed and add its users to the worklist as appropriate.
400 assert(N->getNodeId() == ReadyToProcess && "Node ID recalculated?");
401 N->setNodeId(Processed);
402
403 for (SDNode *User : N->uses()) {
404 int NodeId = User->getNodeId();
405
406 // This node has two options: it can either be a new node or its Node ID
407 // may be a count of the number of operands it has that are not ready.
408 if (NodeId > 0) {
409 User->setNodeId(NodeId-1);
410
411 // If this was the last use it was waiting on, add it to the ready list.
412 if (NodeId-1 == ReadyToProcess)
413 Worklist.push_back(Elt: User);
414 continue;
415 }
416
417 // If this is an unreachable new node, then ignore it. If it ever becomes
418 // reachable by being used by a newly created node then it will be handled
419 // by AnalyzeNewNode.
420 if (NodeId == NewNode)
421 continue;
422
423 // Otherwise, this node is new: this is the first operand of it that
424 // became ready. Its new NodeId is the number of operands it has minus 1
425 // (as this node is now processed).
426 assert(NodeId == Unanalyzed && "Unknown node ID!");
427 User->setNodeId(User->getNumOperands() - 1);
428
429 // If the node only has a single operand, it is now ready.
430 if (User->getNumOperands() == 1)
431 Worklist.push_back(Elt: User);
432 }
433 }
434
435#ifndef EXPENSIVE_CHECKS
436 if (EnableExpensiveChecks)
437#endif
438 PerformExpensiveChecks();
439
440 // If the root changed (e.g. it was a dead load) update the root.
441 DAG.setRoot(Dummy.getValue());
442
443 // Remove dead nodes. This is important to do for cleanliness but also before
444 // the checking loop below. Implicit folding by the DAG.getNode operators and
445 // node morphing can cause unreachable nodes to be around with their flags set
446 // to new.
447 DAG.RemoveDeadNodes();
448
449 // In a debug build, scan all the nodes to make sure we found them all. This
450 // ensures that there are no cycles and that everything got processed.
451#ifndef NDEBUG
452 for (SDNode &Node : DAG.allnodes()) {
453 bool Failed = false;
454
455 // Check that all result types are legal.
456 if (!IgnoreNodeResults(N: &Node))
457 for (unsigned i = 0, NumVals = Node.getNumValues(); i < NumVals; ++i)
458 if (!isTypeLegal(VT: Node.getValueType(ResNo: i))) {
459 dbgs() << "Result type " << i << " illegal: ";
460 Node.dump(G: &DAG);
461 Failed = true;
462 }
463
464 // Check that all operand types are legal.
465 for (unsigned i = 0, NumOps = Node.getNumOperands(); i < NumOps; ++i)
466 if (!IgnoreNodeResults(N: Node.getOperand(Num: i).getNode()) &&
467 !isTypeLegal(VT: Node.getOperand(Num: i).getValueType())) {
468 dbgs() << "Operand type " << i << " illegal: ";
469 Node.getOperand(Num: i).dump(G: &DAG);
470 Failed = true;
471 }
472
473 if (Node.getNodeId() != Processed) {
474 if (Node.getNodeId() == NewNode)
475 dbgs() << "New node not analyzed?\n";
476 else if (Node.getNodeId() == Unanalyzed)
477 dbgs() << "Unanalyzed node not noticed?\n";
478 else if (Node.getNodeId() > 0)
479 dbgs() << "Operand not processed?\n";
480 else if (Node.getNodeId() == ReadyToProcess)
481 dbgs() << "Not added to worklist?\n";
482 Failed = true;
483 }
484
485 if (Failed) {
486 Node.dump(G: &DAG); dbgs() << "\n";
487 llvm_unreachable(nullptr);
488 }
489 }
490#endif
491
492 return Changed;
493}
494
495/// The specified node is the root of a subtree of potentially new nodes.
496/// Correct any processed operands (this may change the node) and calculate the
497/// NodeId. If the node itself changes to a processed node, it is not remapped -
498/// the caller needs to take care of this. Returns the potentially changed node.
499SDNode *DAGTypeLegalizer::AnalyzeNewNode(SDNode *N) {
500 // If this was an existing node that is already done, we're done.
501 if (N->getNodeId() != NewNode && N->getNodeId() != Unanalyzed)
502 return N;
503
504 // Okay, we know that this node is new. Recursively walk all of its operands
505 // to see if they are new also. The depth of this walk is bounded by the size
506 // of the new tree that was constructed (usually 2-3 nodes), so we don't worry
507 // about revisiting of nodes.
508 //
509 // As we walk the operands, keep track of the number of nodes that are
510 // processed. If non-zero, this will become the new nodeid of this node.
511 // Operands may morph when they are analyzed. If so, the node will be
512 // updated after all operands have been analyzed. Since this is rare,
513 // the code tries to minimize overhead in the non-morphing case.
514
515 std::vector<SDValue> NewOps;
516 unsigned NumProcessed = 0;
517 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
518 SDValue OrigOp = N->getOperand(Num: i);
519 SDValue Op = OrigOp;
520
521 AnalyzeNewValue(Val&: Op); // Op may morph.
522
523 if (Op.getNode()->getNodeId() == Processed)
524 ++NumProcessed;
525
526 if (!NewOps.empty()) {
527 // Some previous operand changed. Add this one to the list.
528 NewOps.push_back(x: Op);
529 } else if (Op != OrigOp) {
530 // This is the first operand to change - add all operands so far.
531 NewOps.insert(position: NewOps.end(), first: N->op_begin(), last: N->op_begin() + i);
532 NewOps.push_back(x: Op);
533 }
534 }
535
536 // Some operands changed - update the node.
537 if (!NewOps.empty()) {
538 SDNode *M = DAG.UpdateNodeOperands(N, Ops: NewOps);
539 if (M != N) {
540 // The node morphed into a different node. Normally for this to happen
541 // the original node would have to be marked NewNode. However this can
542 // in theory momentarily not be the case while ReplaceValueWith is doing
543 // its stuff. Mark the original node NewNode to help basic correctness
544 // checking.
545 N->setNodeId(NewNode);
546 if (M->getNodeId() != NewNode && M->getNodeId() != Unanalyzed)
547 // It morphed into a previously analyzed node - nothing more to do.
548 return M;
549
550 // It morphed into a different new node. Do the equivalent of passing
551 // it to AnalyzeNewNode: expunge it and calculate the NodeId. No need
552 // to remap the operands, since they are the same as the operands we
553 // remapped above.
554 N = M;
555 }
556 }
557
558 // Calculate the NodeId.
559 N->setNodeId(N->getNumOperands() - NumProcessed);
560 if (N->getNodeId() == ReadyToProcess)
561 Worklist.push_back(Elt: N);
562
563 return N;
564}
565
566/// Call AnalyzeNewNode, updating the node in Val if needed.
567/// If the node changes to a processed node, then remap it.
568void DAGTypeLegalizer::AnalyzeNewValue(SDValue &Val) {
569 Val.setNode(AnalyzeNewNode(N: Val.getNode()));
570 if (Val.getNode()->getNodeId() == Processed)
571 // We were passed a processed node, or it morphed into one - remap it.
572 RemapValue(V&: Val);
573}
574
575/// If the specified value was already legalized to another value,
576/// replace it by that value.
577void DAGTypeLegalizer::RemapValue(SDValue &V) {
578 auto Id = getTableId(V);
579 V = getSDValue(Id);
580}
581
582void DAGTypeLegalizer::RemapId(TableId &Id) {
583 auto I = ReplacedValues.find(Val: Id);
584 if (I != ReplacedValues.end()) {
585 assert(Id != I->second && "Id is mapped to itself.");
586 // Use path compression to speed up future lookups if values get multiply
587 // replaced with other values.
588 RemapId(Id&: I->second);
589 Id = I->second;
590
591 // Note that N = IdToValueMap[Id] it is possible to have
592 // N.getNode()->getNodeId() == NewNode at this point because it is possible
593 // for a node to be put in the map before being processed.
594 }
595}
596
597namespace {
598 /// This class is a DAGUpdateListener that listens for updates to nodes and
599 /// recomputes their ready state.
600 class NodeUpdateListener : public SelectionDAG::DAGUpdateListener {
601 DAGTypeLegalizer &DTL;
602 SmallSetVector<SDNode*, 16> &NodesToAnalyze;
603 public:
604 explicit NodeUpdateListener(DAGTypeLegalizer &dtl,
605 SmallSetVector<SDNode*, 16> &nta)
606 : SelectionDAG::DAGUpdateListener(dtl.getDAG()),
607 DTL(dtl), NodesToAnalyze(nta) {}
608
609 void NodeDeleted(SDNode *N, SDNode *E) override {
610 assert(N->getNodeId() != DAGTypeLegalizer::ReadyToProcess &&
611 N->getNodeId() != DAGTypeLegalizer::Processed &&
612 "Invalid node ID for RAUW deletion!");
613 // It is possible, though rare, for the deleted node N to occur as a
614 // target in a map, so note the replacement N -> E in ReplacedValues.
615 assert(E && "Node not replaced?");
616 DTL.NoteDeletion(Old: N, New: E);
617
618 // In theory the deleted node could also have been scheduled for analysis.
619 // So remove it from the set of nodes which will be analyzed.
620 NodesToAnalyze.remove(X: N);
621
622 // In general nothing needs to be done for E, since it didn't change but
623 // only gained new uses. However N -> E was just added to ReplacedValues,
624 // and the result of a ReplacedValues mapping is not allowed to be marked
625 // NewNode. So if E is marked NewNode, then it needs to be analyzed.
626 if (E->getNodeId() == DAGTypeLegalizer::NewNode)
627 NodesToAnalyze.insert(X: E);
628 }
629
630 void NodeUpdated(SDNode *N) override {
631 // Node updates can mean pretty much anything. It is possible that an
632 // operand was set to something already processed (f.e.) in which case
633 // this node could become ready. Recompute its flags.
634 assert(N->getNodeId() != DAGTypeLegalizer::ReadyToProcess &&
635 N->getNodeId() != DAGTypeLegalizer::Processed &&
636 "Invalid node ID for RAUW deletion!");
637 N->setNodeId(DAGTypeLegalizer::NewNode);
638 NodesToAnalyze.insert(X: N);
639 }
640 };
641}
642
643
644/// The specified value was legalized to the specified other value.
645/// Update the DAG and NodeIds replacing any uses of From to use To instead.
646void DAGTypeLegalizer::ReplaceValueWith(SDValue From, SDValue To) {
647 assert(From.getNode() != To.getNode() && "Potential legalization loop!");
648
649 // If expansion produced new nodes, make sure they are properly marked.
650 AnalyzeNewValue(Val&: To);
651
652 // Anything that used the old node should now use the new one. Note that this
653 // can potentially cause recursive merging.
654 SmallSetVector<SDNode*, 16> NodesToAnalyze;
655 NodeUpdateListener NUL(*this, NodesToAnalyze);
656 do {
657
658 // The old node may be present in a map like ExpandedIntegers or
659 // PromotedIntegers. Inform maps about the replacement.
660 auto FromId = getTableId(V: From);
661 auto ToId = getTableId(V: To);
662
663 if (FromId != ToId)
664 ReplacedValues[FromId] = ToId;
665 DAG.ReplaceAllUsesOfValueWith(From, To);
666
667 // Process the list of nodes that need to be reanalyzed.
668 while (!NodesToAnalyze.empty()) {
669 SDNode *N = NodesToAnalyze.pop_back_val();
670 if (N->getNodeId() != DAGTypeLegalizer::NewNode)
671 // The node was analyzed while reanalyzing an earlier node - it is safe
672 // to skip. Note that this is not a morphing node - otherwise it would
673 // still be marked NewNode.
674 continue;
675
676 // Analyze the node's operands and recalculate the node ID.
677 SDNode *M = AnalyzeNewNode(N);
678 if (M != N) {
679 // The node morphed into a different node. Make everyone use the new
680 // node instead.
681 assert(M->getNodeId() != NewNode && "Analysis resulted in NewNode!");
682 assert(N->getNumValues() == M->getNumValues() &&
683 "Node morphing changed the number of results!");
684 for (unsigned i = 0, e = N->getNumValues(); i != e; ++i) {
685 SDValue OldVal(N, i);
686 SDValue NewVal(M, i);
687 if (M->getNodeId() == Processed)
688 RemapValue(V&: NewVal);
689 // OldVal may be a target of the ReplacedValues map which was marked
690 // NewNode to force reanalysis because it was updated. Ensure that
691 // anything that ReplacedValues mapped to OldVal will now be mapped
692 // all the way to NewVal.
693 auto OldValId = getTableId(V: OldVal);
694 auto NewValId = getTableId(V: NewVal);
695 DAG.ReplaceAllUsesOfValueWith(From: OldVal, To: NewVal);
696 if (OldValId != NewValId)
697 ReplacedValues[OldValId] = NewValId;
698 }
699 // The original node continues to exist in the DAG, marked NewNode.
700 }
701 }
702 // When recursively update nodes with new nodes, it is possible to have
703 // new uses of From due to CSE. If this happens, replace the new uses of
704 // From with To.
705 } while (!From.use_empty());
706}
707
708void DAGTypeLegalizer::SetPromotedInteger(SDValue Op, SDValue Result) {
709 assert(Result.getValueType() ==
710 TLI.getTypeToTransformTo(*DAG.getContext(), Op.getValueType()) &&
711 "Invalid type for promoted integer");
712 AnalyzeNewValue(Val&: Result);
713
714 auto &OpIdEntry = PromotedIntegers[getTableId(V: Op)];
715 assert((OpIdEntry == 0) && "Node is already promoted!");
716 OpIdEntry = getTableId(V: Result);
717
718 DAG.transferDbgValues(From: Op, To: Result);
719}
720
721void DAGTypeLegalizer::SetSoftenedFloat(SDValue Op, SDValue Result) {
722#ifndef NDEBUG
723 EVT VT = Result.getValueType();
724 LLVMContext &Ctx = *DAG.getContext();
725 assert((VT == EVT::getIntegerVT(Ctx, 80) ||
726 VT == TLI.getTypeToTransformTo(Ctx, Op.getValueType())) &&
727 "Invalid type for softened float");
728#endif
729 AnalyzeNewValue(Val&: Result);
730
731 auto &OpIdEntry = SoftenedFloats[getTableId(V: Op)];
732 assert((OpIdEntry == 0) && "Node is already converted to integer!");
733 OpIdEntry = getTableId(V: Result);
734}
735
736void DAGTypeLegalizer::SetPromotedFloat(SDValue Op, SDValue Result) {
737 assert(Result.getValueType() ==
738 TLI.getTypeToTransformTo(*DAG.getContext(), Op.getValueType()) &&
739 "Invalid type for promoted float");
740 AnalyzeNewValue(Val&: Result);
741
742 auto &OpIdEntry = PromotedFloats[getTableId(V: Op)];
743 assert((OpIdEntry == 0) && "Node is already promoted!");
744 OpIdEntry = getTableId(V: Result);
745}
746
747void DAGTypeLegalizer::SetSoftPromotedHalf(SDValue Op, SDValue Result) {
748 assert(Result.getValueType() == MVT::i16 &&
749 "Invalid type for soft-promoted half");
750 AnalyzeNewValue(Val&: Result);
751
752 auto &OpIdEntry = SoftPromotedHalfs[getTableId(V: Op)];
753 assert((OpIdEntry == 0) && "Node is already promoted!");
754 OpIdEntry = getTableId(V: Result);
755}
756
757void DAGTypeLegalizer::SetScalarizedVector(SDValue Op, SDValue Result) {
758 // Note that in some cases vector operation operands may be greater than
759 // the vector element type. For example BUILD_VECTOR of type <1 x i1> with
760 // a constant i8 operand.
761
762 // We don't currently support the scalarization of scalable vector types.
763 assert(Result.getValueSizeInBits().getFixedValue() >=
764 Op.getScalarValueSizeInBits() &&
765 "Invalid type for scalarized vector");
766 AnalyzeNewValue(Val&: Result);
767
768 auto &OpIdEntry = ScalarizedVectors[getTableId(V: Op)];
769 assert((OpIdEntry == 0) && "Node is already scalarized!");
770 OpIdEntry = getTableId(V: Result);
771}
772
773void DAGTypeLegalizer::GetExpandedInteger(SDValue Op, SDValue &Lo,
774 SDValue &Hi) {
775 std::pair<TableId, TableId> &Entry = ExpandedIntegers[getTableId(V: Op)];
776 assert((Entry.first != 0) && "Operand isn't expanded");
777 Lo = getSDValue(Id&: Entry.first);
778 Hi = getSDValue(Id&: Entry.second);
779}
780
781void DAGTypeLegalizer::SetExpandedInteger(SDValue Op, SDValue Lo,
782 SDValue Hi) {
783 assert(Lo.getValueType() ==
784 TLI.getTypeToTransformTo(*DAG.getContext(), Op.getValueType()) &&
785 Hi.getValueType() == Lo.getValueType() &&
786 "Invalid type for expanded integer");
787 // Lo/Hi may have been newly allocated, if so, add nodeid's as relevant.
788 AnalyzeNewValue(Val&: Lo);
789 AnalyzeNewValue(Val&: Hi);
790
791 // Transfer debug values. Don't invalidate the source debug value until it's
792 // been transferred to the high and low bits.
793 if (DAG.getDataLayout().isBigEndian()) {
794 DAG.transferDbgValues(From: Op, To: Hi, OffsetInBits: 0, SizeInBits: Hi.getValueSizeInBits(), InvalidateDbg: false);
795 DAG.transferDbgValues(From: Op, To: Lo, OffsetInBits: Hi.getValueSizeInBits(),
796 SizeInBits: Lo.getValueSizeInBits());
797 } else {
798 DAG.transferDbgValues(From: Op, To: Lo, OffsetInBits: 0, SizeInBits: Lo.getValueSizeInBits(), InvalidateDbg: false);
799 DAG.transferDbgValues(From: Op, To: Hi, OffsetInBits: Lo.getValueSizeInBits(),
800 SizeInBits: Hi.getValueSizeInBits());
801 }
802
803 // Remember that this is the result of the node.
804 std::pair<TableId, TableId> &Entry = ExpandedIntegers[getTableId(V: Op)];
805 assert((Entry.first == 0) && "Node already expanded");
806 Entry.first = getTableId(V: Lo);
807 Entry.second = getTableId(V: Hi);
808}
809
810void DAGTypeLegalizer::GetExpandedFloat(SDValue Op, SDValue &Lo,
811 SDValue &Hi) {
812 std::pair<TableId, TableId> &Entry = ExpandedFloats[getTableId(V: Op)];
813 assert((Entry.first != 0) && "Operand isn't expanded");
814 Lo = getSDValue(Id&: Entry.first);
815 Hi = getSDValue(Id&: Entry.second);
816}
817
818void DAGTypeLegalizer::SetExpandedFloat(SDValue Op, SDValue Lo,
819 SDValue Hi) {
820 assert(Lo.getValueType() ==
821 TLI.getTypeToTransformTo(*DAG.getContext(), Op.getValueType()) &&
822 Hi.getValueType() == Lo.getValueType() &&
823 "Invalid type for expanded float");
824 // Lo/Hi may have been newly allocated, if so, add nodeid's as relevant.
825 AnalyzeNewValue(Val&: Lo);
826 AnalyzeNewValue(Val&: Hi);
827
828 std::pair<TableId, TableId> &Entry = ExpandedFloats[getTableId(V: Op)];
829 assert((Entry.first == 0) && "Node already expanded");
830 Entry.first = getTableId(V: Lo);
831 Entry.second = getTableId(V: Hi);
832}
833
834void DAGTypeLegalizer::GetSplitVector(SDValue Op, SDValue &Lo,
835 SDValue &Hi) {
836 std::pair<TableId, TableId> &Entry = SplitVectors[getTableId(V: Op)];
837 Lo = getSDValue(Id&: Entry.first);
838 Hi = getSDValue(Id&: Entry.second);
839 assert(Lo.getNode() && "Operand isn't split");
840 ;
841}
842
843void DAGTypeLegalizer::SetSplitVector(SDValue Op, SDValue Lo,
844 SDValue Hi) {
845 assert(Lo.getValueType().getVectorElementType() ==
846 Op.getValueType().getVectorElementType() &&
847 Lo.getValueType().getVectorElementCount() * 2 ==
848 Op.getValueType().getVectorElementCount() &&
849 Hi.getValueType() == Lo.getValueType() &&
850 "Invalid type for split vector");
851 // Lo/Hi may have been newly allocated, if so, add nodeid's as relevant.
852 AnalyzeNewValue(Val&: Lo);
853 AnalyzeNewValue(Val&: Hi);
854
855 // Remember that this is the result of the node.
856 std::pair<TableId, TableId> &Entry = SplitVectors[getTableId(V: Op)];
857 assert((Entry.first == 0) && "Node already split");
858 Entry.first = getTableId(V: Lo);
859 Entry.second = getTableId(V: Hi);
860}
861
862void DAGTypeLegalizer::SetWidenedVector(SDValue Op, SDValue Result) {
863 assert(Result.getValueType() ==
864 TLI.getTypeToTransformTo(*DAG.getContext(), Op.getValueType()) &&
865 "Invalid type for widened vector");
866 AnalyzeNewValue(Val&: Result);
867
868 auto &OpIdEntry = WidenedVectors[getTableId(V: Op)];
869 assert((OpIdEntry == 0) && "Node already widened!");
870 OpIdEntry = getTableId(V: Result);
871}
872
873
874//===----------------------------------------------------------------------===//
875// Utilities.
876//===----------------------------------------------------------------------===//
877
878/// Convert to an integer of the same size.
879SDValue DAGTypeLegalizer::BitConvertToInteger(SDValue Op) {
880 unsigned BitWidth = Op.getValueSizeInBits();
881 return DAG.getNode(Opcode: ISD::BITCAST, DL: SDLoc(Op),
882 VT: EVT::getIntegerVT(Context&: *DAG.getContext(), BitWidth), Operand: Op);
883}
884
885/// Convert to a vector of integers of the same size.
886SDValue DAGTypeLegalizer::BitConvertVectorToIntegerVector(SDValue Op) {
887 assert(Op.getValueType().isVector() && "Only applies to vectors!");
888 unsigned EltWidth = Op.getScalarValueSizeInBits();
889 EVT EltNVT = EVT::getIntegerVT(Context&: *DAG.getContext(), BitWidth: EltWidth);
890 auto EltCnt = Op.getValueType().getVectorElementCount();
891 return DAG.getNode(Opcode: ISD::BITCAST, DL: SDLoc(Op),
892 VT: EVT::getVectorVT(Context&: *DAG.getContext(), VT: EltNVT, EC: EltCnt), Operand: Op);
893}
894
895SDValue DAGTypeLegalizer::CreateStackStoreLoad(SDValue Op,
896 EVT DestVT) {
897 SDLoc dl(Op);
898 // Create the stack frame object. Make sure it is aligned for both
899 // the source and destination types.
900
901 // In cases where the vector is illegal it will be broken down into parts
902 // and stored in parts - we should use the alignment for the smallest part.
903 Align DestAlign = DAG.getReducedAlign(VT: DestVT, /*UseABI=*/false);
904 Align OpAlign = DAG.getReducedAlign(VT: Op.getValueType(), /*UseABI=*/false);
905 Align Align = std::max(a: DestAlign, b: OpAlign);
906 SDValue StackPtr =
907 DAG.CreateStackTemporary(Bytes: Op.getValueType().getStoreSize(), Alignment: Align);
908 // Emit a store to the stack slot.
909 SDValue Store = DAG.getStore(Chain: DAG.getEntryNode(), dl, Val: Op, Ptr: StackPtr,
910 PtrInfo: MachinePointerInfo(), Alignment: Align);
911 // Result is a load from the stack slot.
912 return DAG.getLoad(VT: DestVT, dl, Chain: Store, Ptr: StackPtr, PtrInfo: MachinePointerInfo(), Alignment: Align);
913}
914
915/// Replace the node's results with custom code provided by the target and
916/// return "true", or do nothing and return "false".
917/// The last parameter is FALSE if we are dealing with a node with legal
918/// result types and illegal operand. The second parameter denotes the type of
919/// illegal OperandNo in that case.
920/// The last parameter being TRUE means we are dealing with a
921/// node with illegal result types. The second parameter denotes the type of
922/// illegal ResNo in that case.
923bool DAGTypeLegalizer::CustomLowerNode(SDNode *N, EVT VT, bool LegalizeResult) {
924 // See if the target wants to custom lower this node.
925 if (TLI.getOperationAction(Op: N->getOpcode(), VT) != TargetLowering::Custom)
926 return false;
927
928 SmallVector<SDValue, 8> Results;
929 if (LegalizeResult)
930 TLI.ReplaceNodeResults(N, Results, DAG);
931 else
932 TLI.LowerOperationWrapper(N, Results, DAG);
933
934 if (Results.empty())
935 // The target didn't want to custom lower it after all.
936 return false;
937
938 // Make everything that once used N's values now use those in Results instead.
939 assert(Results.size() == N->getNumValues() &&
940 "Custom lowering returned the wrong number of results!");
941 for (unsigned i = 0, e = Results.size(); i != e; ++i) {
942 ReplaceValueWith(From: SDValue(N, i), To: Results[i]);
943 }
944 return true;
945}
946
947
948/// Widen the node's results with custom code provided by the target and return
949/// "true", or do nothing and return "false".
950bool DAGTypeLegalizer::CustomWidenLowerNode(SDNode *N, EVT VT) {
951 // See if the target wants to custom lower this node.
952 if (TLI.getOperationAction(Op: N->getOpcode(), VT) != TargetLowering::Custom)
953 return false;
954
955 SmallVector<SDValue, 8> Results;
956 TLI.ReplaceNodeResults(N, Results, DAG);
957
958 if (Results.empty())
959 // The target didn't want to custom widen lower its result after all.
960 return false;
961
962 // Update the widening map.
963 assert(Results.size() == N->getNumValues() &&
964 "Custom lowering returned the wrong number of results!");
965 for (unsigned i = 0, e = Results.size(); i != e; ++i) {
966 // If this is a chain output or already widened just replace it.
967 bool WasWidened = SDValue(N, i).getValueType() != Results[i].getValueType();
968 if (WasWidened)
969 SetWidenedVector(Op: SDValue(N, i), Result: Results[i]);
970 else
971 ReplaceValueWith(From: SDValue(N, i), To: Results[i]);
972 }
973 return true;
974}
975
976SDValue DAGTypeLegalizer::DisintegrateMERGE_VALUES(SDNode *N, unsigned ResNo) {
977 for (unsigned i = 0, e = N->getNumValues(); i != e; ++i)
978 if (i != ResNo)
979 ReplaceValueWith(From: SDValue(N, i), To: SDValue(N->getOperand(Num: i)));
980 return SDValue(N->getOperand(Num: ResNo));
981}
982
983/// Use ISD::EXTRACT_ELEMENT nodes to extract the low and high parts of the
984/// given value.
985void DAGTypeLegalizer::GetPairElements(SDValue Pair,
986 SDValue &Lo, SDValue &Hi) {
987 SDLoc dl(Pair);
988 EVT NVT = TLI.getTypeToTransformTo(Context&: *DAG.getContext(), VT: Pair.getValueType());
989 std::tie(args&: Lo, args&: Hi) = DAG.SplitScalar(N: Pair, DL: dl, LoVT: NVT, HiVT: NVT);
990}
991
992/// Build an integer with low bits Lo and high bits Hi.
993SDValue DAGTypeLegalizer::JoinIntegers(SDValue Lo, SDValue Hi) {
994 // Arbitrarily use dlHi for result SDLoc
995 SDLoc dlHi(Hi);
996 SDLoc dlLo(Lo);
997 EVT LVT = Lo.getValueType();
998 EVT HVT = Hi.getValueType();
999 EVT NVT = EVT::getIntegerVT(Context&: *DAG.getContext(),
1000 BitWidth: LVT.getSizeInBits() + HVT.getSizeInBits());
1001
1002 EVT ShiftAmtVT = TLI.getShiftAmountTy(LHSTy: NVT, DL: DAG.getDataLayout());
1003 Lo = DAG.getNode(Opcode: ISD::ZERO_EXTEND, DL: dlLo, VT: NVT, Operand: Lo);
1004 Hi = DAG.getNode(Opcode: ISD::ANY_EXTEND, DL: dlHi, VT: NVT, Operand: Hi);
1005 Hi = DAG.getNode(Opcode: ISD::SHL, DL: dlHi, VT: NVT, N1: Hi,
1006 N2: DAG.getConstant(Val: LVT.getSizeInBits(), DL: dlHi, VT: ShiftAmtVT));
1007 return DAG.getNode(Opcode: ISD::OR, DL: dlHi, VT: NVT, N1: Lo, N2: Hi);
1008}
1009
1010/// Promote the given target boolean to a target boolean of the given type.
1011/// A target boolean is an integer value, not necessarily of type i1, the bits
1012/// of which conform to getBooleanContents.
1013///
1014/// ValVT is the type of values that produced the boolean.
1015SDValue DAGTypeLegalizer::PromoteTargetBoolean(SDValue Bool, EVT ValVT) {
1016 return TLI.promoteTargetBoolean(DAG, Bool, ValVT);
1017}
1018
1019/// Return the lower LoVT bits of Op in Lo and the upper HiVT bits in Hi.
1020void DAGTypeLegalizer::SplitInteger(SDValue Op,
1021 EVT LoVT, EVT HiVT,
1022 SDValue &Lo, SDValue &Hi) {
1023 SDLoc dl(Op);
1024 assert(LoVT.getSizeInBits() + HiVT.getSizeInBits() ==
1025 Op.getValueSizeInBits() && "Invalid integer splitting!");
1026 Lo = DAG.getNode(Opcode: ISD::TRUNCATE, DL: dl, VT: LoVT, Operand: Op);
1027 unsigned ReqShiftAmountInBits =
1028 Log2_32_Ceil(Value: Op.getValueType().getSizeInBits());
1029 MVT ShiftAmountTy =
1030 TLI.getScalarShiftAmountTy(DAG.getDataLayout(), Op.getValueType());
1031 if (ReqShiftAmountInBits > ShiftAmountTy.getSizeInBits())
1032 ShiftAmountTy = MVT::getIntegerVT(BitWidth: NextPowerOf2(A: ReqShiftAmountInBits));
1033 Hi = DAG.getNode(Opcode: ISD::SRL, DL: dl, VT: Op.getValueType(), N1: Op,
1034 N2: DAG.getConstant(Val: LoVT.getSizeInBits(), DL: dl, VT: ShiftAmountTy));
1035 Hi = DAG.getNode(Opcode: ISD::TRUNCATE, DL: dl, VT: HiVT, Operand: Hi);
1036}
1037
1038/// Return the lower and upper halves of Op's bits in a value type half the
1039/// size of Op's.
1040void DAGTypeLegalizer::SplitInteger(SDValue Op,
1041 SDValue &Lo, SDValue &Hi) {
1042 EVT HalfVT =
1043 EVT::getIntegerVT(Context&: *DAG.getContext(), BitWidth: Op.getValueSizeInBits() / 2);
1044 SplitInteger(Op, LoVT: HalfVT, HiVT: HalfVT, Lo, Hi);
1045}
1046
1047
1048//===----------------------------------------------------------------------===//
1049// Entry Point
1050//===----------------------------------------------------------------------===//
1051
1052/// This transforms the SelectionDAG into a SelectionDAG that only uses types
1053/// natively supported by the target. Returns "true" if it made any changes.
1054///
1055/// Note that this is an involved process that may invalidate pointers into
1056/// the graph.
1057bool SelectionDAG::LegalizeTypes() {
1058 return DAGTypeLegalizer(*this).run();
1059}
1060

source code of llvm/lib/CodeGen/SelectionDAG/LegalizeTypes.cpp