1 //===- subzero/src/IceCfg.cpp - Control flow graph implementation ---------===//
2 //
3 // The Subzero Code Generator
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 ///
10 /// \file
11 /// \brief Implements the Cfg class.
12 ///
13 //===----------------------------------------------------------------------===//
14
15 #include "IceCfg.h"
16
17 #include "IceAssembler.h"
18 #include "IceBitVector.h"
19 #include "IceCfgNode.h"
20 #include "IceClFlags.h"
21 #include "IceDefs.h"
22 #include "IceELFObjectWriter.h"
23 #include "IceGlobalInits.h"
24 #include "IceInst.h"
25 #include "IceInstVarIter.h"
26 #include "IceInstrumentation.h"
27 #include "IceLiveness.h"
28 #include "IceLoopAnalyzer.h"
29 #include "IceOperand.h"
30 #include "IceTargetLowering.h"
31
32 #include <memory>
33 #include <utility>
34
35 namespace Ice {
36
Cfg(GlobalContext * Ctx,uint32_t SequenceNumber)37 Cfg::Cfg(GlobalContext *Ctx, uint32_t SequenceNumber)
38 : Allocator(createAllocator()), Ctx(Ctx), SequenceNumber(SequenceNumber),
39 VMask(getFlags().getVerbose()), FunctionName(),
40 NextInstNumber(Inst::NumberInitial), Live(nullptr) {
41 NodeStrings.reset(new StringPool);
42 VarStrings.reset(new StringPool);
43 CfgLocalAllocatorScope _(this);
44 Target = TargetLowering::createLowering(getFlags().getTargetArch(), this);
45 VMetadata.reset(new VariablesMetadata(this));
46 TargetAssembler = Target->createAssembler();
47
48 if (getFlags().getRandomizeAndPoolImmediatesOption() == RPI_Randomize) {
49 // If -randomize-pool-immediates=randomize, create a random number
50 // generator to generate a cookie for constant blinding.
51 RandomNumberGenerator RNG(getFlags().getRandomSeed(), RPE_ConstantBlinding,
52 this->SequenceNumber);
53 ConstantBlindingCookie =
54 (uint32_t)RNG.next((uint64_t)std::numeric_limits<uint32_t>::max() + 1);
55 }
56 }
57
~Cfg()58 Cfg::~Cfg() {
59 assert(CfgAllocatorTraits::current() == nullptr);
60 if (getFlags().getDumpStrings()) {
61 OstreamLocker _(Ctx);
62 Ostream &Str = Ctx->getStrDump();
63 getNodeStrings()->dump(Str);
64 getVarStrings()->dump(Str);
65 }
66 }
67
68 // Called in the initalizer list of Cfg's constructor to create the Allocator
69 // and set it as TLS before any other member fields are constructed, since they
70 // may depend on it.
createAllocator()71 ArenaAllocator *Cfg::createAllocator() {
72 ArenaAllocator *Allocator = new ArenaAllocator();
73 CfgAllocatorTraits::set_current(Allocator);
74 return Allocator;
75 }
76
77 /// Create a string like "foo(i=123:b=9)" indicating the function name, number
78 /// of high-level instructions, and number of basic blocks. This string is only
79 /// used for dumping and other diagnostics, and the idea is that given a set of
80 /// functions to debug a problem on, it's easy to find the smallest or simplest
81 /// function to attack. Note that the counts may change somewhat depending on
82 /// what point it is called during the translation passes.
getFunctionNameAndSize() const83 std::string Cfg::getFunctionNameAndSize() const {
84 if (!BuildDefs::dump())
85 return getFunctionName().toString();
86 SizeT NodeCount = 0;
87 SizeT InstCount = 0;
88 for (CfgNode *Node : getNodes()) {
89 ++NodeCount;
90 // Note: deleted instructions are *not* ignored.
91 InstCount += Node->getPhis().size();
92 for (Inst &I : Node->getInsts()) {
93 if (!llvm::isa<InstTarget>(&I))
94 ++InstCount;
95 }
96 }
97 return getFunctionName() + "(i=" + std::to_string(InstCount) + ":b=" +
98 std::to_string(NodeCount) + ")";
99 }
100
setError(const std::string & Message)101 void Cfg::setError(const std::string &Message) {
102 HasError = true;
103 ErrorMessage = Message;
104 }
105
makeNode()106 CfgNode *Cfg::makeNode() {
107 SizeT LabelIndex = Nodes.size();
108 auto *Node = CfgNode::create(this, LabelIndex);
109 Nodes.push_back(Node);
110 return Node;
111 }
112
swapNodes(NodeList & NewNodes)113 void Cfg::swapNodes(NodeList &NewNodes) {
114 assert(Nodes.size() == NewNodes.size());
115 Nodes.swap(NewNodes);
116 for (SizeT I = 0, NumNodes = getNumNodes(); I < NumNodes; ++I)
117 Nodes[I]->resetIndex(I);
118 }
119
makeVariable(Type Ty)120 template <> Variable *Cfg::makeVariable<Variable>(Type Ty) {
121 SizeT Index = Variables.size();
122 Variable *Var;
123 if (Target->shouldSplitToVariableVecOn32(Ty)) {
124 Var = VariableVecOn32::create(this, Ty, Index);
125 } else if (Target->shouldSplitToVariable64On32(Ty)) {
126 Var = Variable64On32::create(this, Ty, Index);
127 } else {
128 Var = Variable::create(this, Ty, Index);
129 }
130 Variables.push_back(Var);
131 return Var;
132 }
133
addArg(Variable * Arg)134 void Cfg::addArg(Variable *Arg) {
135 Arg->setIsArg();
136 Args.push_back(Arg);
137 }
138
addImplicitArg(Variable * Arg)139 void Cfg::addImplicitArg(Variable *Arg) {
140 Arg->setIsImplicitArg();
141 ImplicitArgs.push_back(Arg);
142 }
143
144 // Returns whether the stack frame layout has been computed yet. This is used
145 // for dumping the stack frame location of Variables.
hasComputedFrame() const146 bool Cfg::hasComputedFrame() const { return getTarget()->hasComputedFrame(); }
147
148 namespace {
149 constexpr char BlockNameGlobalPrefix[] = ".L$profiler$block_name$";
150 constexpr char BlockStatsGlobalPrefix[] = ".L$profiler$block_info$";
151 } // end of anonymous namespace
152
createNodeNameDeclaration(const std::string & NodeAsmName)153 void Cfg::createNodeNameDeclaration(const std::string &NodeAsmName) {
154 auto *Var = VariableDeclaration::create(GlobalInits.get());
155 Var->setName(Ctx, BlockNameGlobalPrefix + NodeAsmName);
156 Var->setIsConstant(true);
157 Var->addInitializer(VariableDeclaration::DataInitializer::create(
158 GlobalInits.get(), NodeAsmName.data(), NodeAsmName.size() + 1));
159 const SizeT Int64ByteSize = typeWidthInBytes(IceType_i64);
160 Var->setAlignment(Int64ByteSize); // Wasteful, 32-bit could use 4 bytes.
161 GlobalInits->push_back(Var);
162 }
163
createBlockProfilingInfoDeclaration(const std::string & NodeAsmName,VariableDeclaration * NodeNameDeclaration)164 void Cfg::createBlockProfilingInfoDeclaration(
165 const std::string &NodeAsmName, VariableDeclaration *NodeNameDeclaration) {
166 auto *Var = VariableDeclaration::create(GlobalInits.get());
167 Var->setName(Ctx, BlockStatsGlobalPrefix + NodeAsmName);
168 const SizeT Int64ByteSize = typeWidthInBytes(IceType_i64);
169 Var->addInitializer(VariableDeclaration::ZeroInitializer::create(
170 GlobalInits.get(), Int64ByteSize));
171
172 const RelocOffsetT NodeNameDeclarationOffset = 0;
173 Var->addInitializer(VariableDeclaration::RelocInitializer::create(
174 GlobalInits.get(), NodeNameDeclaration,
175 {RelocOffset::create(Ctx, NodeNameDeclarationOffset)}));
176 Var->setAlignment(Int64ByteSize);
177 GlobalInits->push_back(Var);
178 }
179
profileBlocks()180 void Cfg::profileBlocks() {
181 if (GlobalInits == nullptr)
182 GlobalInits.reset(new VariableDeclarationList());
183
184 for (CfgNode *Node : Nodes) {
185 const std::string NodeAsmName = Node->getAsmName();
186 createNodeNameDeclaration(NodeAsmName);
187 createBlockProfilingInfoDeclaration(NodeAsmName, GlobalInits->back());
188 Node->profileExecutionCount(GlobalInits->back());
189 }
190 }
191
isProfileGlobal(const VariableDeclaration & Var)192 bool Cfg::isProfileGlobal(const VariableDeclaration &Var) {
193 if (!Var.getName().hasStdString())
194 return false;
195 return Var.getName().toString().find(BlockStatsGlobalPrefix) == 0;
196 }
197
addCallToProfileSummary()198 void Cfg::addCallToProfileSummary() {
199 // The call(s) to __Sz_profile_summary are added by the profiler in functions
200 // that cause the program to exit. This function is defined in
201 // runtime/szrt_profiler.c.
202 Constant *ProfileSummarySym =
203 Ctx->getConstantExternSym(Ctx->getGlobalString("__Sz_profile_summary"));
204 constexpr SizeT NumArgs = 0;
205 constexpr Variable *Void = nullptr;
206 constexpr bool HasTailCall = false;
207 auto *Call =
208 InstCall::create(this, NumArgs, Void, ProfileSummarySym, HasTailCall);
209 getEntryNode()->getInsts().push_front(Call);
210 }
211
translate()212 void Cfg::translate() {
213 if (hasError())
214 return;
215 // Cache the possibly-overridden optimization level once translation begins.
216 // It would be nicer to do this in the constructor, but we need to wait until
217 // after setFunctionName() has a chance to be called.
218 OptimizationLevel =
219 getFlags().matchForceO2(getFunctionName(), getSequenceNumber())
220 ? Opt_2
221 : getFlags().getOptLevel();
222 if (BuildDefs::timers()) {
223 if (getFlags().matchTimingFocus(getFunctionName(), getSequenceNumber())) {
224 setFocusedTiming();
225 getContext()->resetTimer(GlobalContext::TSK_Default);
226 }
227 }
228 if (BuildDefs::dump()) {
229 if (isVerbose(IceV_Status) &&
230 getFlags().matchTestStatus(getFunctionName(), getSequenceNumber())) {
231 getContext()->getStrDump() << ">>>Translating "
232 << getFunctionNameAndSize()
233 << " seq=" << getSequenceNumber() << "\n";
234 }
235 }
236 TimerMarker T_func(getContext(), getFunctionName().toStringOrEmpty());
237 TimerMarker T(TimerStack::TT_translate, this);
238
239 dump("Initial CFG");
240
241 if (getFlags().getEnableBlockProfile()) {
242 profileBlocks();
243 // TODO(jpp): this is fragile, at best. Figure out a better way of
244 // detecting exit functions.
245 if (getFunctionName().toStringOrEmpty() == "exit") {
246 addCallToProfileSummary();
247 }
248 dump("Profiled CFG");
249 }
250
251 // Create the Hi and Lo variables where a split was needed
252 for (Variable *Var : Variables) {
253 if (auto *Var64On32 = llvm::dyn_cast<Variable64On32>(Var)) {
254 Var64On32->initHiLo(this);
255 } else if (auto *VarVecOn32 = llvm::dyn_cast<VariableVecOn32>(Var)) {
256 VarVecOn32->initVecElement(this);
257 }
258 }
259
260 // Instrument the Cfg, e.g. with AddressSanitizer
261 if (!BuildDefs::minimal() && getFlags().getSanitizeAddresses()) {
262 getContext()->instrumentFunc(this);
263 dump("Instrumented CFG");
264 }
265
266 // The set of translation passes and their order are determined by the
267 // target.
268 getTarget()->translate();
269
270 dump("Final output");
271 if (getFocusedTiming()) {
272 getContext()->dumpLocalTimers(getFunctionName().toString());
273 }
274 }
275
fixPhiNodes()276 void Cfg::fixPhiNodes() {
277 for (auto *Node : Nodes) {
278 // Fix all the phi edges since WASM can't tell how to make them correctly at
279 // the beginning.
280 assert(Node);
281 const auto &InEdges = Node->getInEdges();
282 for (auto &Instr : Node->getPhis()) {
283 auto *Phi = llvm::cast<InstPhi>(&Instr);
284 assert(Phi);
285 for (SizeT i = 0; i < InEdges.size(); ++i) {
286 Phi->setLabel(i, InEdges[i]);
287 }
288 }
289 }
290 }
291
computeInOutEdges()292 void Cfg::computeInOutEdges() {
293 // Compute the out-edges.
294 for (CfgNode *Node : Nodes) {
295 Node->computeSuccessors();
296 }
297
298 // Prune any unreachable nodes before computing in-edges.
299 SizeT NumNodes = getNumNodes();
300 BitVector Reachable(NumNodes);
301 BitVector Pending(NumNodes);
302 Pending.set(getEntryNode()->getIndex());
303 while (true) {
304 int Index = Pending.find_first();
305 if (Index == -1)
306 break;
307 Pending.reset(Index);
308 Reachable.set(Index);
309 CfgNode *Node = Nodes[Index];
310 assert(Node->getIndex() == (SizeT)Index);
311 for (CfgNode *Succ : Node->getOutEdges()) {
312 SizeT SuccIndex = Succ->getIndex();
313 if (!Reachable.test(SuccIndex))
314 Pending.set(SuccIndex);
315 }
316 }
317 SizeT Dest = 0;
318 for (SizeT Source = 0; Source < NumNodes; ++Source) {
319 if (Reachable.test(Source)) {
320 Nodes[Dest] = Nodes[Source];
321 Nodes[Dest]->resetIndex(Dest);
322 // Compute the in-edges.
323 Nodes[Dest]->computePredecessors();
324 ++Dest;
325 }
326 }
327 Nodes.resize(Dest);
328
329 TimerMarker T(TimerStack::TT_phiValidation, this);
330 for (CfgNode *Node : Nodes)
331 Node->enforcePhiConsistency();
332 }
333
renumberInstructions()334 void Cfg::renumberInstructions() {
335 TimerMarker T(TimerStack::TT_renumberInstructions, this);
336 NextInstNumber = Inst::NumberInitial;
337 for (CfgNode *Node : Nodes)
338 Node->renumberInstructions();
339 // Make sure the entry node is the first node and therefore got the lowest
340 // instruction numbers, to facilitate live range computation of function
341 // arguments. We want to model function arguments as being live on entry to
342 // the function, otherwise an argument whose only use is in the first
343 // instruction will be assigned a trivial live range and the register
344 // allocator will not recognize its live range as overlapping another
345 // variable's live range.
346 assert(Nodes.empty() || (*Nodes.begin() == getEntryNode()));
347 }
348
349 // placePhiLoads() must be called before placePhiStores().
placePhiLoads()350 void Cfg::placePhiLoads() {
351 TimerMarker T(TimerStack::TT_placePhiLoads, this);
352 for (CfgNode *Node : Nodes)
353 Node->placePhiLoads();
354 }
355
356 // placePhiStores() must be called after placePhiLoads().
placePhiStores()357 void Cfg::placePhiStores() {
358 TimerMarker T(TimerStack::TT_placePhiStores, this);
359 for (CfgNode *Node : Nodes)
360 Node->placePhiStores();
361 }
362
deletePhis()363 void Cfg::deletePhis() {
364 TimerMarker T(TimerStack::TT_deletePhis, this);
365 for (CfgNode *Node : Nodes)
366 Node->deletePhis();
367 }
368
advancedPhiLowering()369 void Cfg::advancedPhiLowering() {
370 TimerMarker T(TimerStack::TT_advancedPhiLowering, this);
371 // Clear all previously computed live ranges (but not live-in/live-out bit
372 // vectors or last-use markers), because the follow-on register allocation is
373 // only concerned with live ranges across the newly created blocks.
374 for (Variable *Var : Variables) {
375 Var->getLiveRange().reset();
376 }
377 // This splits edges and appends new nodes to the end of the node list. This
378 // can invalidate iterators, so don't use an iterator.
379 SizeT NumNodes = getNumNodes();
380 SizeT NumVars = getNumVariables();
381 for (SizeT I = 0; I < NumNodes; ++I)
382 Nodes[I]->advancedPhiLowering();
383
384 TimerMarker TT(TimerStack::TT_lowerPhiAssignments, this);
385 if (true) {
386 // The following code does an in-place update of liveness and live ranges
387 // as a result of adding the new phi edge split nodes.
388 getLiveness()->initPhiEdgeSplits(Nodes.begin() + NumNodes,
389 Variables.begin() + NumVars);
390 TimerMarker TTT(TimerStack::TT_liveness, this);
391 // Iterate over the newly added nodes to add their liveness info.
392 for (auto I = Nodes.begin() + NumNodes, E = Nodes.end(); I != E; ++I) {
393 InstNumberT FirstInstNum = getNextInstNumber();
394 (*I)->renumberInstructions();
395 InstNumberT LastInstNum = getNextInstNumber() - 1;
396 (*I)->liveness(getLiveness());
397 (*I)->livenessAddIntervals(getLiveness(), FirstInstNum, LastInstNum);
398 }
399 } else {
400 // The following code does a brute-force recalculation of live ranges as a
401 // result of adding the new phi edge split nodes. The liveness calculation
402 // is particularly expensive because the new nodes are not yet in a proper
403 // topological order and so convergence is slower.
404 //
405 // This code is kept here for reference and can be temporarily enabled in
406 // case the incremental code is under suspicion.
407 renumberInstructions();
408 liveness(Liveness_Intervals);
409 getVMetadata()->init(VMK_All);
410 }
411 Target->regAlloc(RAK_Phi);
412 }
413
414 // Find a reasonable placement for nodes that have not yet been placed, while
415 // maintaining the same relative ordering among already placed nodes.
reorderNodes()416 void Cfg::reorderNodes() {
417 // TODO(ascull): it would be nice if the switch tests were always followed by
418 // the default case to allow for fall through.
419 using PlacedList = CfgList<CfgNode *>;
420 PlacedList Placed; // Nodes with relative placement locked down
421 PlacedList Unreachable; // Unreachable nodes
422 PlacedList::iterator NoPlace = Placed.end();
423 // Keep track of where each node has been tentatively placed so that we can
424 // manage insertions into the middle.
425 CfgVector<PlacedList::iterator> PlaceIndex(Nodes.size(), NoPlace);
426 for (CfgNode *Node : Nodes) {
427 // The "do ... while(0);" construct is to factor out the --PlaceIndex and
428 // assert() statements before moving to the next node.
429 do {
430 if (Node != getEntryNode() && Node->getInEdges().empty()) {
431 // The node has essentially been deleted since it is not a successor of
432 // any other node.
433 Unreachable.push_back(Node);
434 PlaceIndex[Node->getIndex()] = Unreachable.end();
435 Node->setNeedsPlacement(false);
436 continue;
437 }
438 if (!Node->needsPlacement()) {
439 // Add to the end of the Placed list.
440 Placed.push_back(Node);
441 PlaceIndex[Node->getIndex()] = Placed.end();
442 continue;
443 }
444 Node->setNeedsPlacement(false);
445 // Assume for now that the unplaced node is from edge-splitting and
446 // therefore has 1 in-edge and 1 out-edge (actually, possibly more than 1
447 // in-edge if the predecessor node was contracted). If this changes in
448 // the future, rethink the strategy.
449 assert(Node->getInEdges().size() >= 1);
450 assert(Node->hasSingleOutEdge());
451
452 // If it's a (non-critical) edge where the successor has a single
453 // in-edge, then place it before the successor.
454 CfgNode *Succ = Node->getOutEdges().front();
455 if (Succ->getInEdges().size() == 1 &&
456 PlaceIndex[Succ->getIndex()] != NoPlace) {
457 Placed.insert(PlaceIndex[Succ->getIndex()], Node);
458 PlaceIndex[Node->getIndex()] = PlaceIndex[Succ->getIndex()];
459 continue;
460 }
461
462 // Otherwise, place it after the (first) predecessor.
463 CfgNode *Pred = Node->getInEdges().front();
464 auto PredPosition = PlaceIndex[Pred->getIndex()];
465 // It shouldn't be the case that PredPosition==NoPlace, but if that
466 // somehow turns out to be true, we just insert Node before
467 // PredPosition=NoPlace=Placed.end() .
468 if (PredPosition != NoPlace)
469 ++PredPosition;
470 Placed.insert(PredPosition, Node);
471 PlaceIndex[Node->getIndex()] = PredPosition;
472 } while (0);
473
474 --PlaceIndex[Node->getIndex()];
475 assert(*PlaceIndex[Node->getIndex()] == Node);
476 }
477
478 // Reorder Nodes according to the built-up lists.
479 NodeList Reordered;
480 Reordered.reserve(Placed.size() + Unreachable.size());
481 for (CfgNode *Node : Placed)
482 Reordered.push_back(Node);
483 for (CfgNode *Node : Unreachable)
484 Reordered.push_back(Node);
485 assert(getNumNodes() == Reordered.size());
486 swapNodes(Reordered);
487 }
488
489 namespace {
getRandomPostOrder(CfgNode * Node,BitVector & ToVisit,Ice::NodeList & PostOrder,Ice::RandomNumberGenerator * RNG)490 void getRandomPostOrder(CfgNode *Node, BitVector &ToVisit,
491 Ice::NodeList &PostOrder,
492 Ice::RandomNumberGenerator *RNG) {
493 assert(ToVisit[Node->getIndex()]);
494 ToVisit[Node->getIndex()] = false;
495 NodeList Outs = Node->getOutEdges();
496 Ice::RandomShuffle(Outs.begin(), Outs.end(),
497 [RNG](int N) { return RNG->next(N); });
498 for (CfgNode *Next : Outs) {
499 if (ToVisit[Next->getIndex()])
500 getRandomPostOrder(Next, ToVisit, PostOrder, RNG);
501 }
502 PostOrder.push_back(Node);
503 }
504 } // end of anonymous namespace
505
shuffleNodes()506 void Cfg::shuffleNodes() {
507 if (!getFlags().getReorderBasicBlocks())
508 return;
509
510 NodeList ReversedReachable;
511 NodeList Unreachable;
512 BitVector ToVisit(Nodes.size(), true);
513 // Create Random number generator for function reordering
514 RandomNumberGenerator RNG(getFlags().getRandomSeed(),
515 RPE_BasicBlockReordering, SequenceNumber);
516 // Traverse from entry node.
517 getRandomPostOrder(getEntryNode(), ToVisit, ReversedReachable, &RNG);
518 // Collect the unreachable nodes.
519 for (CfgNode *Node : Nodes)
520 if (ToVisit[Node->getIndex()])
521 Unreachable.push_back(Node);
522 // Copy the layout list to the Nodes.
523 NodeList Shuffled;
524 Shuffled.reserve(ReversedReachable.size() + Unreachable.size());
525 for (CfgNode *Node : reverse_range(ReversedReachable))
526 Shuffled.push_back(Node);
527 for (CfgNode *Node : Unreachable)
528 Shuffled.push_back(Node);
529 assert(Nodes.size() == Shuffled.size());
530 swapNodes(Shuffled);
531
532 dump("After basic block shuffling");
533 }
534
localCSE(bool AssumeSSA)535 void Cfg::localCSE(bool AssumeSSA) {
536 // Performs basic-block local common-subexpression elimination
537 // If we have
538 // t1 = op b c
539 // t2 = op b c
540 // This pass will replace future references to t2 in a basic block by t1
541 // Points to note:
542 // 1. Assumes SSA by default. To change this, use -lcse=no-ssa
543 // This is needed if this pass is moved to a point later in the pipeline.
544 // If variables have a single definition (in the node), CSE can work just
545 // on the basis of an equality compare on instructions (sans Dest). When
546 // variables can be updated (hence, non-SSA) the result of a previous
547 // instruction which used that variable as an operand can not be reused.
548 // 2. Leaves removal of instructions to DCE.
549 // 3. Only enabled on arithmetic instructions. pnacl-clang (-O2) is expected
550 // to take care of cases not arising from GEP simplification.
551 // 4. By default, a single pass is made over each basic block. Control this
552 // with -lcse-max-iters=N
553
554 TimerMarker T(TimerStack::TT_localCse, this);
555 struct VariableHash {
556 size_t operator()(const Variable *Var) const { return Var->hashValue(); }
557 };
558
559 struct InstHash {
560 size_t operator()(const Inst *Instr) const {
561 auto Kind = Instr->getKind();
562 auto Result =
563 std::hash<typename std::underlying_type<Inst::InstKind>::type>()(
564 Kind);
565 for (SizeT i = 0; i < Instr->getSrcSize(); ++i) {
566 Result ^= Instr->getSrc(i)->hashValue();
567 }
568 return Result;
569 }
570 };
571 struct InstEq {
572 bool srcEq(const Operand *A, const Operand *B) const {
573 if (llvm::isa<Variable>(A) || llvm::isa<Constant>(A))
574 return (A == B);
575 return false;
576 }
577 bool operator()(const Inst *InstrA, const Inst *InstrB) const {
578 if ((InstrA->getKind() != InstrB->getKind()) ||
579 (InstrA->getSrcSize() != InstrB->getSrcSize()))
580 return false;
581
582 if (auto *A = llvm::dyn_cast<InstArithmetic>(InstrA)) {
583 auto *B = llvm::cast<InstArithmetic>(InstrB);
584 // A, B are guaranteed to be of the same 'kind' at this point
585 // So, dyn_cast is not needed
586 if (A->getOp() != B->getOp())
587 return false;
588 }
589 // Does not enter loop if different kind or number of operands
590 for (SizeT i = 0; i < InstrA->getSrcSize(); ++i) {
591 if (!srcEq(InstrA->getSrc(i), InstrB->getSrc(i)))
592 return false;
593 }
594 return true;
595 }
596 };
597
598 for (CfgNode *Node : getNodes()) {
599 CfgUnorderedSet<Inst *, InstHash, InstEq> Seen;
600 // Stores currently available instructions.
601
602 CfgUnorderedMap<Variable *, Variable *, VariableHash> Replacements;
603 // Combining the above two into a single data structure might consume less
604 // memory but will be slower i.e map of Instruction -> Set of Variables
605
606 CfgUnorderedMap<Variable *, std::vector<Inst *>, VariableHash> Dependency;
607 // Maps a variable to the Instructions that depend on it.
608 // a = op1 b c
609 // x = op2 c d
610 // Will result in the map : b -> {a}, c -> {a, x}, d -> {x}
611 // Not necessary for SSA as dependencies will never be invalidated, and the
612 // container will use minimal memory when left unused.
613
614 auto IterCount = getFlags().getLocalCseMaxIterations();
615
616 for (uint32_t i = 0; i < IterCount; ++i) {
617 // TODO(manasijm): Stats on IterCount -> performance
618 for (Inst &Instr : Node->getInsts()) {
619 if (Instr.isDeleted() || !llvm::isa<InstArithmetic>(&Instr))
620 continue;
621 if (!AssumeSSA) {
622 // Invalidate replacements
623 auto Iter = Replacements.find(Instr.getDest());
624 if (Iter != Replacements.end()) {
625 Replacements.erase(Iter);
626 }
627
628 // Invalidate 'seen' instructions whose operands were just updated.
629 auto DepIter = Dependency.find(Instr.getDest());
630 if (DepIter != Dependency.end()) {
631 for (auto *DepInst : DepIter->second) {
632 Seen.erase(DepInst);
633 }
634 }
635 }
636
637 // Replace - doing this before checking for repetitions might enable
638 // more optimizations
639 for (SizeT i = 0; i < Instr.getSrcSize(); ++i) {
640 auto *Opnd = Instr.getSrc(i);
641 if (auto *Var = llvm::dyn_cast<Variable>(Opnd)) {
642 if (Replacements.find(Var) != Replacements.end()) {
643 Instr.replaceSource(i, Replacements[Var]);
644 }
645 }
646 }
647
648 // Check for repetitions
649 auto SeenIter = Seen.find(&Instr);
650 if (SeenIter != Seen.end()) { // seen before
651 const Inst *Found = *SeenIter;
652 Replacements[Instr.getDest()] = Found->getDest();
653 } else { // new
654 Seen.insert(&Instr);
655
656 if (!AssumeSSA) {
657 // Update dependencies
658 for (SizeT i = 0; i < Instr.getSrcSize(); ++i) {
659 auto *Opnd = Instr.getSrc(i);
660 if (auto *Var = llvm::dyn_cast<Variable>(Opnd)) {
661 Dependency[Var].push_back(&Instr);
662 }
663 }
664 }
665 }
666 }
667 }
668 }
669 }
670
loopInvariantCodeMotion()671 void Cfg::loopInvariantCodeMotion() {
672 TimerMarker T(TimerStack::TT_loopInvariantCodeMotion, this);
673 // Does not introduce new nodes as of now.
674 for (auto &Loop : LoopInfo) {
675 CfgNode *Header = Loop.Header;
676 assert(Header);
677 if (Header->getLoopNestDepth() < 1)
678 return;
679 CfgNode *PreHeader = Loop.PreHeader;
680 if (PreHeader == nullptr || PreHeader->getInsts().size() == 0) {
681 return; // try next loop
682 }
683
684 auto &Insts = PreHeader->getInsts();
685 auto &LastInst = Insts.back();
686 Insts.pop_back();
687
688 for (auto *Inst : findLoopInvariantInstructions(Loop.Body)) {
689 PreHeader->appendInst(Inst);
690 }
691 PreHeader->appendInst(&LastInst);
692 }
693 }
694
695 CfgVector<Inst *>
findLoopInvariantInstructions(const CfgUnorderedSet<SizeT> & Body)696 Cfg::findLoopInvariantInstructions(const CfgUnorderedSet<SizeT> &Body) {
697 CfgUnorderedSet<Inst *> InvariantInsts;
698 CfgUnorderedSet<Variable *> InvariantVars;
699 for (auto *Var : getArgs()) {
700 InvariantVars.insert(Var);
701 }
702 bool Changed = false;
703 do {
704 Changed = false;
705 for (auto NodeIndex : Body) {
706 auto *Node = Nodes[NodeIndex];
707 CfgVector<std::reference_wrapper<Inst>> Insts(Node->getInsts().begin(),
708 Node->getInsts().end());
709
710 for (auto &InstRef : Insts) {
711 auto &Inst = InstRef.get();
712 if (Inst.isDeleted() ||
713 InvariantInsts.find(&Inst) != InvariantInsts.end())
714 continue;
715 switch (Inst.getKind()) {
716 case Inst::InstKind::Alloca:
717 case Inst::InstKind::Br:
718 case Inst::InstKind::Ret:
719 case Inst::InstKind::Phi:
720 case Inst::InstKind::Call:
721 case Inst::InstKind::IntrinsicCall:
722 case Inst::InstKind::Load:
723 case Inst::InstKind::Store:
724 case Inst::InstKind::Switch:
725 continue;
726 default:
727 break;
728 }
729
730 bool IsInvariant = true;
731 for (SizeT i = 0; i < Inst.getSrcSize(); ++i) {
732 if (auto *Var = llvm::dyn_cast<Variable>(Inst.getSrc(i))) {
733 if (InvariantVars.find(Var) == InvariantVars.end()) {
734 IsInvariant = false;
735 }
736 }
737 }
738 if (IsInvariant) {
739 Changed = true;
740 InvariantInsts.insert(&Inst);
741 Node->getInsts().remove(Inst);
742 if (Inst.getDest() != nullptr) {
743 InvariantVars.insert(Inst.getDest());
744 }
745 }
746 }
747 }
748 } while (Changed);
749
750 CfgVector<Inst *> InstVector(InvariantInsts.begin(), InvariantInsts.end());
751 std::sort(InstVector.begin(), InstVector.end(),
752 [](Inst *A, Inst *B) { return A->getNumber() < B->getNumber(); });
753 return InstVector;
754 }
755
shortCircuitJumps()756 void Cfg::shortCircuitJumps() {
757 // Split Nodes whenever an early jump is possible.
758 // __N :
759 // a = <something>
760 // Instruction 1 without side effect
761 // ... b = <something> ...
762 // Instruction N without side effect
763 // t1 = or a b
764 // br t1 __X __Y
765 //
766 // is transformed into:
767 // __N :
768 // a = <something>
769 // br a __X __N_ext
770 //
771 // __N_ext :
772 // Instruction 1 without side effect
773 // ... b = <something> ...
774 // Instruction N without side effect
775 // br b __X __Y
776 // (Similar logic for AND, jump to false instead of true target.)
777
778 TimerMarker T(TimerStack::TT_shortCircuit, this);
779 getVMetadata()->init(VMK_Uses);
780 auto NodeStack = this->getNodes();
781 CfgUnorderedMap<SizeT, CfgVector<CfgNode *>> Splits;
782 while (!NodeStack.empty()) {
783 auto *Node = NodeStack.back();
784 NodeStack.pop_back();
785 auto NewNode = Node->shortCircuit();
786 if (NewNode) {
787 NodeStack.push_back(NewNode);
788 NodeStack.push_back(Node);
789 Splits[Node->getIndex()].push_back(NewNode);
790 }
791 }
792
793 // Insert nodes in the right place
794 NodeList NewList;
795 NewList.reserve(Nodes.size());
796 CfgUnorderedSet<SizeT> Inserted;
797 for (auto *Node : Nodes) {
798 if (Inserted.find(Node->getIndex()) != Inserted.end())
799 continue; // already inserted
800 NodeList Stack{Node};
801 while (!Stack.empty()) {
802 auto *Current = Stack.back();
803 Stack.pop_back();
804 Inserted.insert(Current->getIndex());
805 NewList.push_back(Current);
806 for (auto *Next : Splits[Current->getIndex()]) {
807 Stack.push_back(Next);
808 }
809 }
810 }
811
812 SizeT NodeIndex = 0;
813 for (auto *Node : NewList) {
814 Node->resetIndex(NodeIndex++);
815 }
816 Nodes = NewList;
817 }
818
floatConstantCSE()819 void Cfg::floatConstantCSE() {
820 // Load multiple uses of a floating point constant (between two call
821 // instructions or block start/end) into a variable before its first use.
822 // t1 = b + 1.0
823 // t2 = c + 1.0
824 // Gets transformed to:
825 // t0 = 1.0
826 // t0_1 = t0
827 // t1 = b + t0_1
828 // t2 = c + t0_1
829 // Call instructions reset the procedure, but use the same variable, just in
830 // case it got a register. We are assuming floating point registers are not
831 // callee saved in general. Example, continuing from before:
832 // result = call <some function>
833 // t3 = d + 1.0
834 // Gets transformed to:
835 // result = call <some function>
836 // t0_2 = t0
837 // t3 = d + t0_2
838 // TODO(manasijm, stichnot): Figure out how to 'link' t0 to the stack slot of
839 // 1.0. When t0 does not get a register, introducing an extra assignment
840 // statement does not make sense. The relevant portion is marked below.
841
842 TimerMarker _(TimerStack::TT_floatConstantCse, this);
843 for (CfgNode *Node : getNodes()) {
844
845 CfgUnorderedMap<Constant *, Variable *> ConstCache;
846 auto Current = Node->getInsts().begin();
847 auto End = Node->getInsts().end();
848 while (Current != End) {
849 CfgUnorderedMap<Constant *, CfgVector<InstList::iterator>> FloatUses;
850 if (llvm::isa<InstCall>(iteratorToInst(Current))) {
851 ++Current;
852 assert(Current != End);
853 // Block should not end with a call
854 }
855 while (Current != End && !llvm::isa<InstCall>(iteratorToInst(Current))) {
856 if (!Current->isDeleted()) {
857 for (SizeT i = 0; i < Current->getSrcSize(); ++i) {
858 if (auto *Const = llvm::dyn_cast<Constant>(Current->getSrc(i))) {
859 if (Const->getType() == IceType_f32 ||
860 Const->getType() == IceType_f64) {
861 FloatUses[Const].push_back(Current);
862 }
863 }
864 }
865 }
866 ++Current;
867 }
868 for (auto &Pair : FloatUses) {
869 static constexpr SizeT MinUseThreshold = 3;
870 if (Pair.second.size() < MinUseThreshold)
871 continue;
872 // Only consider constants with at least `MinUseThreshold` uses
873 auto &Insts = Node->getInsts();
874
875 if (ConstCache.find(Pair.first) == ConstCache.end()) {
876 // Saw a constant (which is used at least twice) for the first time
877 auto *NewVar = makeVariable(Pair.first->getType());
878 // NewVar->setLinkedTo(Pair.first);
879 // TODO(manasijm): Plumbing for linking to an Operand.
880 auto *Assign = InstAssign::create(Node->getCfg(), NewVar, Pair.first);
881 Insts.insert(Pair.second[0], Assign);
882 ConstCache[Pair.first] = NewVar;
883 }
884
885 auto *NewVar = makeVariable(Pair.first->getType());
886 NewVar->setLinkedTo(ConstCache[Pair.first]);
887 auto *Assign =
888 InstAssign::create(Node->getCfg(), NewVar, ConstCache[Pair.first]);
889
890 Insts.insert(Pair.second[0], Assign);
891 for (auto InstUse : Pair.second) {
892 for (SizeT i = 0; i < InstUse->getSrcSize(); ++i) {
893 if (auto *Const = llvm::dyn_cast<Constant>(InstUse->getSrc(i))) {
894 if (Const == Pair.first) {
895 InstUse->replaceSource(i, NewVar);
896 }
897 }
898 }
899 }
900 }
901 }
902 }
903 }
904
doArgLowering()905 void Cfg::doArgLowering() {
906 TimerMarker T(TimerStack::TT_doArgLowering, this);
907 getTarget()->lowerArguments();
908 }
909
sortAndCombineAllocas(CfgVector<InstAlloca * > & Allocas,uint32_t CombinedAlignment,InstList & Insts,AllocaBaseVariableType BaseVariableType)910 void Cfg::sortAndCombineAllocas(CfgVector<InstAlloca *> &Allocas,
911 uint32_t CombinedAlignment, InstList &Insts,
912 AllocaBaseVariableType BaseVariableType) {
913 if (Allocas.empty())
914 return;
915 // Sort by decreasing alignment.
916 std::sort(Allocas.begin(), Allocas.end(), [](InstAlloca *A1, InstAlloca *A2) {
917 uint32_t Align1 = A1->getAlignInBytes();
918 uint32_t Align2 = A2->getAlignInBytes();
919 if (Align1 == Align2)
920 return A1->getNumber() < A2->getNumber();
921 else
922 return Align1 > Align2;
923 });
924 // Process the allocas in order of decreasing stack alignment. This allows
925 // us to pack less-aligned pieces after more-aligned ones, resulting in less
926 // stack growth. It also allows there to be at most one stack alignment "and"
927 // instruction for a whole list of allocas.
928 uint32_t CurrentOffset = 0;
929 CfgVector<int32_t> Offsets;
930 for (Inst *Instr : Allocas) {
931 auto *Alloca = llvm::cast<InstAlloca>(Instr);
932 // Adjust the size of the allocation up to the next multiple of the
933 // object's alignment.
934 uint32_t Alignment = std::max(Alloca->getAlignInBytes(), 1u);
935 auto *ConstSize =
936 llvm::dyn_cast<ConstantInteger32>(Alloca->getSizeInBytes());
937 uint32_t Size = Utils::applyAlignment(ConstSize->getValue(), Alignment);
938 if (BaseVariableType == BVT_FramePointer) {
939 // Addressing is relative to the frame pointer. Subtract the offset after
940 // adding the size of the alloca, because it grows downwards from the
941 // frame pointer.
942 Offsets.push_back(Target->getFramePointerOffset(CurrentOffset, Size));
943 } else {
944 // Addressing is relative to the stack pointer or to a user pointer. Add
945 // the offset before adding the size of the object, because it grows
946 // upwards from the stack pointer. In addition, if the addressing is
947 // relative to the stack pointer, we need to add the pre-computed max out
948 // args size bytes.
949 const uint32_t OutArgsOffsetOrZero =
950 (BaseVariableType == BVT_StackPointer)
951 ? getTarget()->maxOutArgsSizeBytes()
952 : 0;
953 Offsets.push_back(CurrentOffset + OutArgsOffsetOrZero);
954 }
955 // Update the running offset of the fused alloca region.
956 CurrentOffset += Size;
957 }
958 // Round the offset up to the alignment granularity to use as the size.
959 uint32_t TotalSize = Utils::applyAlignment(CurrentOffset, CombinedAlignment);
960 // Ensure every alloca was assigned an offset.
961 assert(Allocas.size() == Offsets.size());
962
963 switch (BaseVariableType) {
964 case BVT_UserPointer: {
965 Variable *BaseVariable = makeVariable(IceType_i32);
966 for (SizeT i = 0; i < Allocas.size(); ++i) {
967 auto *Alloca = llvm::cast<InstAlloca>(Allocas[i]);
968 // Emit a new addition operation to replace the alloca.
969 Operand *AllocaOffset = Ctx->getConstantInt32(Offsets[i]);
970 InstArithmetic *Add =
971 InstArithmetic::create(this, InstArithmetic::Add, Alloca->getDest(),
972 BaseVariable, AllocaOffset);
973 Insts.push_front(Add);
974 Alloca->setDeleted();
975 }
976 Operand *AllocaSize = Ctx->getConstantInt32(TotalSize);
977 InstAlloca *CombinedAlloca =
978 InstAlloca::create(this, BaseVariable, AllocaSize, CombinedAlignment);
979 CombinedAlloca->setKnownFrameOffset();
980 Insts.push_front(CombinedAlloca);
981 } break;
982 case BVT_StackPointer:
983 case BVT_FramePointer: {
984 for (SizeT i = 0; i < Allocas.size(); ++i) {
985 auto *Alloca = llvm::cast<InstAlloca>(Allocas[i]);
986 // Emit a fake definition of the rematerializable variable.
987 Variable *Dest = Alloca->getDest();
988 auto *Def = InstFakeDef::create(this, Dest);
989 if (BaseVariableType == BVT_StackPointer)
990 Dest->setRematerializable(getTarget()->getStackReg(), Offsets[i]);
991 else
992 Dest->setRematerializable(getTarget()->getFrameReg(), Offsets[i]);
993 Insts.push_front(Def);
994 Alloca->setDeleted();
995 }
996 // Allocate the fixed area in the function prolog.
997 getTarget()->reserveFixedAllocaArea(TotalSize, CombinedAlignment);
998 } break;
999 }
1000 }
1001
processAllocas(bool SortAndCombine)1002 void Cfg::processAllocas(bool SortAndCombine) {
1003 TimerMarker _(TimerStack::TT_alloca, this);
1004 const uint32_t StackAlignment = getTarget()->getStackAlignment();
1005 CfgNode *EntryNode = getEntryNode();
1006 assert(EntryNode);
1007 // LLVM enforces power of 2 alignment.
1008 assert(llvm::isPowerOf2_32(StackAlignment));
1009 // If the ABI's stack alignment is smaller than the vector size (16 bytes),
1010 // conservatively use a frame pointer to allow for explicit alignment of the
1011 // stack pointer. This needs to happen before register allocation so the frame
1012 // pointer can be reserved.
1013 if (getTarget()->needsStackPointerAlignment()) {
1014 getTarget()->setHasFramePointer();
1015 }
1016 // Determine if there are large alignment allocations in the entry block or
1017 // dynamic allocations (variable size in the entry block).
1018 bool HasLargeAlignment = false;
1019 bool HasDynamicAllocation = false;
1020 for (Inst &Instr : EntryNode->getInsts()) {
1021 if (Instr.isDeleted())
1022 continue;
1023 if (auto *Alloca = llvm::dyn_cast<InstAlloca>(&Instr)) {
1024 uint32_t AlignmentParam = Alloca->getAlignInBytes();
1025 if (AlignmentParam > StackAlignment)
1026 HasLargeAlignment = true;
1027 if (llvm::isa<Constant>(Alloca->getSizeInBytes()))
1028 Alloca->setKnownFrameOffset();
1029 else {
1030 HasDynamicAllocation = true;
1031 // If Allocas are not sorted, the first dynamic allocation causes
1032 // later Allocas to be at unknown offsets relative to the stack/frame.
1033 if (!SortAndCombine)
1034 break;
1035 }
1036 }
1037 }
1038 // Don't do the heavyweight sorting and layout for low optimization levels.
1039 if (!SortAndCombine)
1040 return;
1041 // Any alloca outside the entry block is a dynamic allocation.
1042 for (CfgNode *Node : Nodes) {
1043 if (Node == EntryNode)
1044 continue;
1045 for (Inst &Instr : Node->getInsts()) {
1046 if (Instr.isDeleted())
1047 continue;
1048 if (llvm::isa<InstAlloca>(&Instr)) {
1049 // Allocations outside the entry block require a frame pointer.
1050 HasDynamicAllocation = true;
1051 break;
1052 }
1053 }
1054 if (HasDynamicAllocation)
1055 break;
1056 }
1057 // Mark the target as requiring a frame pointer.
1058 if (HasLargeAlignment || HasDynamicAllocation)
1059 getTarget()->setHasFramePointer();
1060 // Collect the Allocas into the two vectors.
1061 // Allocas in the entry block that have constant size and alignment less
1062 // than or equal to the function's stack alignment.
1063 CfgVector<InstAlloca *> FixedAllocas;
1064 // Allocas in the entry block that have constant size and alignment greater
1065 // than the function's stack alignment.
1066 CfgVector<InstAlloca *> AlignedAllocas;
1067 // Maximum alignment used by any alloca.
1068 uint32_t MaxAlignment = StackAlignment;
1069 for (Inst &Instr : EntryNode->getInsts()) {
1070 if (Instr.isDeleted())
1071 continue;
1072 if (auto *Alloca = llvm::dyn_cast<InstAlloca>(&Instr)) {
1073 if (!llvm::isa<Constant>(Alloca->getSizeInBytes()))
1074 continue;
1075 uint32_t AlignmentParam = Alloca->getAlignInBytes();
1076 // For default align=0, set it to the real value 1, to avoid any
1077 // bit-manipulation problems below.
1078 AlignmentParam = std::max(AlignmentParam, 1u);
1079 assert(llvm::isPowerOf2_32(AlignmentParam));
1080 if (HasDynamicAllocation && AlignmentParam > StackAlignment) {
1081 // If we have both dynamic allocations and large stack alignments,
1082 // high-alignment allocations are pulled out with their own base.
1083 AlignedAllocas.push_back(Alloca);
1084 } else {
1085 FixedAllocas.push_back(Alloca);
1086 }
1087 MaxAlignment = std::max(AlignmentParam, MaxAlignment);
1088 }
1089 }
1090 // Add instructions to the head of the entry block in reverse order.
1091 InstList &Insts = getEntryNode()->getInsts();
1092 if (HasDynamicAllocation && HasLargeAlignment) {
1093 // We are using a frame pointer, but fixed large-alignment alloca addresses
1094 // do not have a known offset from either the stack or frame pointer.
1095 // They grow up from a user pointer from an alloca.
1096 sortAndCombineAllocas(AlignedAllocas, MaxAlignment, Insts, BVT_UserPointer);
1097 // Fixed size allocas are addressed relative to the frame pointer.
1098 sortAndCombineAllocas(FixedAllocas, StackAlignment, Insts,
1099 BVT_FramePointer);
1100 } else {
1101 // Otherwise, fixed size allocas are addressed relative to the stack unless
1102 // there are dynamic allocas.
1103 const AllocaBaseVariableType BasePointerType =
1104 (HasDynamicAllocation ? BVT_FramePointer : BVT_StackPointer);
1105 sortAndCombineAllocas(FixedAllocas, MaxAlignment, Insts, BasePointerType);
1106 }
1107 if (!FixedAllocas.empty() || !AlignedAllocas.empty())
1108 // No use calling findRematerializable() unless there is some
1109 // rematerializable alloca instruction to seed it.
1110 findRematerializable();
1111 }
1112
1113 namespace {
1114
1115 // Helpers for findRematerializable(). For each of them, if a suitable
1116 // rematerialization is found, the instruction's Dest variable is set to be
1117 // rematerializable and it returns true, otherwise it returns false.
1118
rematerializeArithmetic(const Inst * Instr)1119 bool rematerializeArithmetic(const Inst *Instr) {
1120 // Check that it's an Arithmetic instruction with an Add operation.
1121 auto *Arith = llvm::dyn_cast<InstArithmetic>(Instr);
1122 if (Arith == nullptr || Arith->getOp() != InstArithmetic::Add)
1123 return false;
1124 // Check that Src(0) is rematerializable.
1125 auto *Src0Var = llvm::dyn_cast<Variable>(Arith->getSrc(0));
1126 if (Src0Var == nullptr || !Src0Var->isRematerializable())
1127 return false;
1128 // Check that Src(1) is an immediate.
1129 auto *Src1Imm = llvm::dyn_cast<ConstantInteger32>(Arith->getSrc(1));
1130 if (Src1Imm == nullptr)
1131 return false;
1132 Arith->getDest()->setRematerializable(
1133 Src0Var->getRegNum(), Src0Var->getStackOffset() + Src1Imm->getValue());
1134 return true;
1135 }
1136
rematerializeAssign(const Inst * Instr)1137 bool rematerializeAssign(const Inst *Instr) {
1138 // An InstAssign only originates from an inttoptr or ptrtoint instruction,
1139 // which never occurs in a MINIMAL build.
1140 if (BuildDefs::minimal())
1141 return false;
1142 // Check that it's an Assign instruction.
1143 if (!llvm::isa<InstAssign>(Instr))
1144 return false;
1145 // Check that Src(0) is rematerializable.
1146 auto *Src0Var = llvm::dyn_cast<Variable>(Instr->getSrc(0));
1147 if (Src0Var == nullptr || !Src0Var->isRematerializable())
1148 return false;
1149 Instr->getDest()->setRematerializable(Src0Var->getRegNum(),
1150 Src0Var->getStackOffset());
1151 return true;
1152 }
1153
rematerializeCast(const Inst * Instr)1154 bool rematerializeCast(const Inst *Instr) {
1155 // An pointer-type bitcast never occurs in a MINIMAL build.
1156 if (BuildDefs::minimal())
1157 return false;
1158 // Check that it's a Cast instruction with a Bitcast operation.
1159 auto *Cast = llvm::dyn_cast<InstCast>(Instr);
1160 if (Cast == nullptr || Cast->getCastKind() != InstCast::Bitcast)
1161 return false;
1162 // Check that Src(0) is rematerializable.
1163 auto *Src0Var = llvm::dyn_cast<Variable>(Cast->getSrc(0));
1164 if (Src0Var == nullptr || !Src0Var->isRematerializable())
1165 return false;
1166 // Check that Dest and Src(0) have the same type.
1167 Variable *Dest = Cast->getDest();
1168 if (Dest->getType() != Src0Var->getType())
1169 return false;
1170 Dest->setRematerializable(Src0Var->getRegNum(), Src0Var->getStackOffset());
1171 return true;
1172 }
1173
1174 } // end of anonymous namespace
1175
1176 /// Scan the function to find additional rematerializable variables. This is
1177 /// possible when the source operand of an InstAssignment is a rematerializable
1178 /// variable, or the same for a pointer-type InstCast::Bitcast, or when an
1179 /// InstArithmetic is an add of a rematerializable variable and an immediate.
1180 /// Note that InstAssignment instructions and pointer-type InstCast::Bitcast
1181 /// instructions generally only come about from the IceConverter's treatment of
1182 /// inttoptr, ptrtoint, and bitcast instructions. TODO(stichnot): Consider
1183 /// other possibilities, however unlikely, such as InstArithmetic::Sub, or
1184 /// commutativity.
findRematerializable()1185 void Cfg::findRematerializable() {
1186 // Scan the instructions in order, and repeat until no new opportunities are
1187 // found. It may take more than one iteration because a variable's defining
1188 // block may happen to come after a block where it is used, depending on the
1189 // CfgNode linearization order.
1190 bool FoundNewAssignment;
1191 do {
1192 FoundNewAssignment = false;
1193 for (CfgNode *Node : getNodes()) {
1194 // No need to process Phi instructions.
1195 for (Inst &Instr : Node->getInsts()) {
1196 if (Instr.isDeleted())
1197 continue;
1198 Variable *Dest = Instr.getDest();
1199 if (Dest == nullptr || Dest->isRematerializable())
1200 continue;
1201 if (rematerializeArithmetic(&Instr) || rematerializeAssign(&Instr) ||
1202 rematerializeCast(&Instr)) {
1203 FoundNewAssignment = true;
1204 }
1205 }
1206 }
1207 } while (FoundNewAssignment);
1208 }
1209
doAddressOpt()1210 void Cfg::doAddressOpt() {
1211 TimerMarker T(TimerStack::TT_doAddressOpt, this);
1212 for (CfgNode *Node : Nodes)
1213 Node->doAddressOpt();
1214 }
1215
1216 namespace {
1217 // ShuffleVectorUtils implements helper functions for rematerializing
1218 // shufflevector instructions from a sequence of extractelement/insertelement
1219 // instructions. It looks for the following pattern:
1220 //
1221 // %t0 = extractelement A, %n0
1222 // %t1 = extractelement B, %n1
1223 // %t2 = extractelement C, %n2
1224 // ...
1225 // %tN = extractelement N, %nN
1226 // %d0 = insertelement undef, %t0, 0
1227 // %d1 = insertelement %d0, %t1, 1
1228 // %d2 = insertelement %d1, %t2, 2
1229 // ...
1230 // %dest = insertelement %d_N-1, %tN, N
1231 //
1232 // where N is num_element(typeof(%dest)), and A, B, C, ... N are at most two
1233 // distinct variables.
1234 namespace ShuffleVectorUtils {
1235 // findAllInserts is used when searching for all the insertelements that are
1236 // used in a shufflevector operation. This function works recursively, when
1237 // invoked with I = i, the function assumes Insts[i] is the last found
1238 // insertelement in the chain. The next insertelement insertruction is saved in
1239 // Insts[i+1].
findAllInserts(Cfg * Func,GlobalContext * Ctx,VariablesMetadata * VM,CfgVector<const Inst * > * Insts,SizeT I=0)1240 bool findAllInserts(Cfg *Func, GlobalContext *Ctx, VariablesMetadata *VM,
1241 CfgVector<const Inst *> *Insts, SizeT I = 0) {
1242 const bool Verbose = BuildDefs::dump() && Func->isVerbose(IceV_ShufMat);
1243
1244 if (I > Insts->size()) {
1245 if (Verbose) {
1246 Ctx->getStrDump() << "\tToo many inserts.\n";
1247 }
1248 return false;
1249 }
1250
1251 const auto *LastInsert = Insts->at(I);
1252 assert(llvm::isa<InstInsertElement>(LastInsert));
1253
1254 if (I == Insts->size() - 1) {
1255 // Matching against undef is not really needed because the value in Src(0)
1256 // will be totally overwritten. We still enforce it anyways because the
1257 // PNaCl toolchain generates the bitcode with it.
1258 if (!llvm::isa<ConstantUndef>(LastInsert->getSrc(0))) {
1259 if (Verbose) {
1260 Ctx->getStrDump() << "\tSrc0 is not undef: " << I << " "
1261 << Insts->size();
1262 LastInsert->dump(Func);
1263 Ctx->getStrDump() << "\n";
1264 }
1265 return false;
1266 }
1267
1268 // The following loop ensures that the insertelements are sorted. In theory,
1269 // we could relax this restriction and allow any order. As long as each
1270 // index appears exactly once, this chain is still a candidate for becoming
1271 // a shufflevector. The Insts vector is traversed backwards because the
1272 // instructions are "enqueued" in reverse order.
1273 int32_t ExpectedElement = 0;
1274 for (const auto *I : reverse_range(*Insts)) {
1275 if (llvm::cast<ConstantInteger32>(I->getSrc(2))->getValue() !=
1276 ExpectedElement) {
1277 return false;
1278 }
1279 ++ExpectedElement;
1280 }
1281 return true;
1282 }
1283
1284 const auto *Src0V = llvm::cast<Variable>(LastInsert->getSrc(0));
1285 const auto *Def = VM->getSingleDefinition(Src0V);
1286
1287 // Only optimize if the first operand in
1288 //
1289 // Dest = insertelement A, B, 10
1290 //
1291 // is singly-def'ed.
1292 if (Def == nullptr) {
1293 if (Verbose) {
1294 Ctx->getStrDump() << "\tmulti-def: ";
1295 (*Insts)[I]->dump(Func);
1296 Ctx->getStrDump() << "\n";
1297 }
1298 return false;
1299 }
1300
1301 // We also require the (single) definition to come from an insertelement
1302 // instruction.
1303 if (!llvm::isa<InstInsertElement>(Def)) {
1304 if (Verbose) {
1305 Ctx->getStrDump() << "\tnot insert element: ";
1306 Def->dump(Func);
1307 Ctx->getStrDump() << "\n";
1308 }
1309 return false;
1310 }
1311
1312 // Everything seems fine, so we save Def in Insts, and delegate the decision
1313 // to findAllInserts.
1314 (*Insts)[I + 1] = Def;
1315
1316 return findAllInserts(Func, Ctx, VM, Insts, I + 1);
1317 }
1318
1319 // insertsLastElement returns true if Insert is inserting an element in the last
1320 // position of a vector.
insertsLastElement(const Inst & Insert)1321 bool insertsLastElement(const Inst &Insert) {
1322 const Type DestTy = Insert.getDest()->getType();
1323 assert(isVectorType(DestTy));
1324 const SizeT Elem =
1325 llvm::cast<ConstantInteger32>(Insert.getSrc(2))->getValue();
1326 return Elem == typeNumElements(DestTy) - 1;
1327 }
1328
1329 // findAllExtracts goes over all the insertelement instructions that are
1330 // candidates to be replaced by a shufflevector, and searches for all the
1331 // definitions of the elements being inserted. If all of the elements are the
1332 // result of an extractelement instruction, and all of the extractelements
1333 // operate on at most two different sources, than the instructions can be
1334 // replaced by a shufflevector.
findAllExtracts(Cfg * Func,GlobalContext * Ctx,VariablesMetadata * VM,const CfgVector<const Inst * > & Insts,Variable ** Src0,Variable ** Src1,CfgVector<const Inst * > * Extracts)1335 bool findAllExtracts(Cfg *Func, GlobalContext *Ctx, VariablesMetadata *VM,
1336 const CfgVector<const Inst *> &Insts, Variable **Src0,
1337 Variable **Src1, CfgVector<const Inst *> *Extracts) {
1338 const bool Verbose = BuildDefs::dump() && Func->isVerbose(IceV_ShufMat);
1339
1340 *Src0 = nullptr;
1341 *Src1 = nullptr;
1342 assert(Insts.size() > 0);
1343 for (SizeT I = 0; I < Insts.size(); ++I) {
1344 const auto *Insert = Insts.at(I);
1345 const auto *Src1V = llvm::dyn_cast<Variable>(Insert->getSrc(1));
1346 if (Src1V == nullptr) {
1347 if (Verbose) {
1348 Ctx->getStrDump() << "src(1) is not a variable: ";
1349 Insert->dump(Func);
1350 Ctx->getStrDump() << "\n";
1351 }
1352 return false;
1353 }
1354
1355 const auto *Def = VM->getSingleDefinition(Src1V);
1356 if (Def == nullptr) {
1357 if (Verbose) {
1358 Ctx->getStrDump() << "multi-def src(1): ";
1359 Insert->dump(Func);
1360 Ctx->getStrDump() << "\n";
1361 }
1362 return false;
1363 }
1364
1365 if (!llvm::isa<InstExtractElement>(Def)) {
1366 if (Verbose) {
1367 Ctx->getStrDump() << "not extractelement: ";
1368 Def->dump(Func);
1369 Ctx->getStrDump() << "\n";
1370 }
1371 return false;
1372 }
1373
1374 auto *Src = llvm::cast<Variable>(Def->getSrc(0));
1375 if (*Src0 == nullptr) {
1376 // No sources yet. Save Src to Src0.
1377 *Src0 = Src;
1378 } else if (*Src1 == nullptr) {
1379 // We already have a source, so we might save Src in Src1 -- but only if
1380 // Src0 is not Src.
1381 if (*Src0 != Src) {
1382 *Src1 = Src;
1383 }
1384 } else if (Src != *Src0 && Src != *Src1) {
1385 // More than two sources, so we can't rematerialize the shufflevector
1386 // instruction.
1387 if (Verbose) {
1388 Ctx->getStrDump() << "Can't shuffle more than two sources.\n";
1389 }
1390 return false;
1391 }
1392
1393 (*Extracts)[I] = Def;
1394 }
1395
1396 // We should have seen at least one source operand.
1397 assert(*Src0 != nullptr);
1398
1399 // If a second source was not seen, then we just make Src1 = Src0 to simplify
1400 // things down stream. This should not matter, as all of the indexes in the
1401 // shufflevector instruction will point to Src0.
1402 if (*Src1 == nullptr) {
1403 *Src1 = *Src0;
1404 }
1405
1406 return true;
1407 }
1408
1409 } // end of namespace ShuffleVectorUtils
1410 } // end of anonymous namespace
1411
materializeVectorShuffles()1412 void Cfg::materializeVectorShuffles() {
1413 const bool Verbose = BuildDefs::dump() && isVerbose(IceV_ShufMat);
1414
1415 std::unique_ptr<OstreamLocker> L;
1416 if (Verbose) {
1417 L.reset(new OstreamLocker(getContext()));
1418 getContext()->getStrDump() << "\nShuffle materialization:\n";
1419 }
1420
1421 // MaxVectorElements is the maximum number of elements in the vector types
1422 // handled by Subzero. We use it to create the Inserts and Extracts vectors
1423 // with the appropriate size, thus avoiding resize() calls.
1424 const SizeT MaxVectorElements = typeNumElements(IceType_v16i8);
1425 CfgVector<const Inst *> Inserts(MaxVectorElements);
1426 CfgVector<const Inst *> Extracts(MaxVectorElements);
1427
1428 TimerMarker T(TimerStack::TT_materializeVectorShuffles, this);
1429 for (CfgNode *Node : Nodes) {
1430 for (auto &Instr : Node->getInsts()) {
1431 if (!llvm::isa<InstInsertElement>(Instr)) {
1432 continue;
1433 }
1434 if (!ShuffleVectorUtils::insertsLastElement(Instr)) {
1435 // To avoid wasting time, we only start the pattern match at the last
1436 // insertelement instruction -- and go backwards from there.
1437 continue;
1438 }
1439 if (Verbose) {
1440 getContext()->getStrDump() << "\tCandidate: ";
1441 Instr.dump(this);
1442 getContext()->getStrDump() << "\n";
1443 }
1444 Inserts.resize(typeNumElements(Instr.getDest()->getType()));
1445 Inserts[0] = &Instr;
1446 if (!ShuffleVectorUtils::findAllInserts(this, getContext(),
1447 VMetadata.get(), &Inserts)) {
1448 // If we fail to find a sequence of insertelements, we stop the
1449 // optimization.
1450 if (Verbose) {
1451 getContext()->getStrDump() << "\tFalse alarm.\n";
1452 }
1453 continue;
1454 }
1455 if (Verbose) {
1456 getContext()->getStrDump() << "\tFound the following insertelement: \n";
1457 for (auto *I : reverse_range(Inserts)) {
1458 getContext()->getStrDump() << "\t\t";
1459 I->dump(this);
1460 getContext()->getStrDump() << "\n";
1461 }
1462 }
1463 Extracts.resize(Inserts.size());
1464 Variable *Src0;
1465 Variable *Src1;
1466 if (!ShuffleVectorUtils::findAllExtracts(this, getContext(),
1467 VMetadata.get(), Inserts, &Src0,
1468 &Src1, &Extracts)) {
1469 // If we fail to match the definitions of the insertelements' sources
1470 // with extractelement instructions -- or if those instructions operate
1471 // on more than two different variables -- we stop the optimization.
1472 if (Verbose) {
1473 getContext()->getStrDump() << "\tFailed to match extractelements.\n";
1474 }
1475 continue;
1476 }
1477 if (Verbose) {
1478 getContext()->getStrDump()
1479 << "\tFound the following insert/extract element pairs: \n";
1480 for (SizeT I = 0; I < Inserts.size(); ++I) {
1481 const SizeT Pos = Inserts.size() - I - 1;
1482 getContext()->getStrDump() << "\t\tInsert : ";
1483 Inserts[Pos]->dump(this);
1484 getContext()->getStrDump() << "\n\t\tExtract: ";
1485 Extracts[Pos]->dump(this);
1486 getContext()->getStrDump() << "\n";
1487 }
1488 }
1489
1490 assert(Src0 != nullptr);
1491 assert(Src1 != nullptr);
1492
1493 auto *ShuffleVector =
1494 InstShuffleVector::create(this, Instr.getDest(), Src0, Src1);
1495 assert(ShuffleVector->getSrc(0) == Src0);
1496 assert(ShuffleVector->getSrc(1) == Src1);
1497 for (SizeT I = 0; I < Extracts.size(); ++I) {
1498 const SizeT Pos = Extracts.size() - I - 1;
1499 auto *Index = llvm::cast<ConstantInteger32>(Extracts[Pos]->getSrc(1));
1500 if (Src0 == Extracts[Pos]->getSrc(0)) {
1501 ShuffleVector->addIndex(Index);
1502 } else {
1503 ShuffleVector->addIndex(llvm::cast<ConstantInteger32>(
1504 Ctx->getConstantInt32(Index->getValue() + Extracts.size())));
1505 }
1506 }
1507
1508 if (Verbose) {
1509 getContext()->getStrDump() << "Created: ";
1510 ShuffleVector->dump(this);
1511 getContext()->getStrDump() << "\n";
1512 }
1513
1514 Instr.setDeleted();
1515 auto &LoweringContext = getTarget()->getContext();
1516 LoweringContext.setInsertPoint(instToIterator(&Instr));
1517 LoweringContext.insert(ShuffleVector);
1518 }
1519 }
1520 }
1521
doNopInsertion()1522 void Cfg::doNopInsertion() {
1523 if (!getFlags().getShouldDoNopInsertion())
1524 return;
1525 TimerMarker T(TimerStack::TT_doNopInsertion, this);
1526 RandomNumberGenerator RNG(getFlags().getRandomSeed(), RPE_NopInsertion,
1527 SequenceNumber);
1528 for (CfgNode *Node : Nodes)
1529 Node->doNopInsertion(RNG);
1530 }
1531
genCode()1532 void Cfg::genCode() {
1533 TimerMarker T(TimerStack::TT_genCode, this);
1534 for (CfgNode *Node : Nodes)
1535 Node->genCode();
1536 }
1537
1538 // Compute the stack frame layout.
genFrame()1539 void Cfg::genFrame() {
1540 TimerMarker T(TimerStack::TT_genFrame, this);
1541 getTarget()->addProlog(Entry);
1542 for (CfgNode *Node : Nodes)
1543 if (Node->getHasReturn())
1544 getTarget()->addEpilog(Node);
1545 }
1546
generateLoopInfo()1547 void Cfg::generateLoopInfo() {
1548 TimerMarker T(TimerStack::TT_computeLoopNestDepth, this);
1549 LoopInfo = ComputeLoopInfo(this);
1550 }
1551
1552 // This is a lightweight version of live-range-end calculation. Marks the last
1553 // use of only those variables whose definition and uses are completely with a
1554 // single block. It is a quick single pass and doesn't need to iterate until
1555 // convergence.
livenessLightweight()1556 void Cfg::livenessLightweight() {
1557 TimerMarker T(TimerStack::TT_livenessLightweight, this);
1558 getVMetadata()->init(VMK_Uses);
1559 for (CfgNode *Node : Nodes)
1560 Node->livenessLightweight();
1561 }
1562
liveness(LivenessMode Mode)1563 void Cfg::liveness(LivenessMode Mode) {
1564 TimerMarker T(TimerStack::TT_liveness, this);
1565 // Destroying the previous (if any) Liveness information clears the Liveness
1566 // allocator TLS pointer.
1567 Live = nullptr;
1568 Live = Liveness::create(this, Mode);
1569
1570 getVMetadata()->init(VMK_Uses);
1571 Live->init();
1572
1573 // Initialize with all nodes needing to be processed.
1574 BitVector NeedToProcess(Nodes.size(), true);
1575 while (NeedToProcess.any()) {
1576 // Iterate in reverse topological order to speed up convergence.
1577 for (CfgNode *Node : reverse_range(Nodes)) {
1578 if (NeedToProcess[Node->getIndex()]) {
1579 NeedToProcess[Node->getIndex()] = false;
1580 bool Changed = Node->liveness(getLiveness());
1581 if (Changed) {
1582 // If the beginning-of-block liveness changed since the last
1583 // iteration, mark all in-edges as needing to be processed.
1584 for (CfgNode *Pred : Node->getInEdges())
1585 NeedToProcess[Pred->getIndex()] = true;
1586 }
1587 }
1588 }
1589 }
1590 if (Mode == Liveness_Intervals) {
1591 // Reset each variable's live range.
1592 for (Variable *Var : Variables)
1593 Var->resetLiveRange();
1594 }
1595 // Make a final pass over each node to delete dead instructions, collect the
1596 // first and last instruction numbers, and add live range segments for that
1597 // node.
1598 for (CfgNode *Node : Nodes) {
1599 InstNumberT FirstInstNum = Inst::NumberSentinel;
1600 InstNumberT LastInstNum = Inst::NumberSentinel;
1601 for (Inst &I : Node->getPhis()) {
1602 I.deleteIfDead();
1603 if (Mode == Liveness_Intervals && !I.isDeleted()) {
1604 if (FirstInstNum == Inst::NumberSentinel)
1605 FirstInstNum = I.getNumber();
1606 assert(I.getNumber() > LastInstNum);
1607 LastInstNum = I.getNumber();
1608 }
1609 }
1610 for (Inst &I : Node->getInsts()) {
1611 I.deleteIfDead();
1612 if (Mode == Liveness_Intervals && !I.isDeleted()) {
1613 if (FirstInstNum == Inst::NumberSentinel)
1614 FirstInstNum = I.getNumber();
1615 assert(I.getNumber() > LastInstNum);
1616 LastInstNum = I.getNumber();
1617 }
1618 }
1619 if (Mode == Liveness_Intervals) {
1620 // Special treatment for live in-args. Their liveness needs to extend
1621 // beyond the beginning of the function, otherwise an arg whose only use
1622 // is in the first instruction will end up having the trivial live range
1623 // [2,2) and will *not* interfere with other arguments. So if the first
1624 // instruction of the method is "r=arg1+arg2", both args may be assigned
1625 // the same register. This is accomplished by extending the entry block's
1626 // instruction range from [2,n) to [1,n) which will transform the
1627 // problematic [2,2) live ranges into [1,2). This extension works because
1628 // the entry node is guaranteed to have the lowest instruction numbers.
1629 if (Node == getEntryNode()) {
1630 FirstInstNum = Inst::NumberExtended;
1631 // Just in case the entry node somehow contains no instructions...
1632 if (LastInstNum == Inst::NumberSentinel)
1633 LastInstNum = FirstInstNum;
1634 }
1635 // If this node somehow contains no instructions, don't bother trying to
1636 // add liveness intervals for it, because variables that are live-in and
1637 // live-out will have a bogus interval added.
1638 if (FirstInstNum != Inst::NumberSentinel)
1639 Node->livenessAddIntervals(getLiveness(), FirstInstNum, LastInstNum);
1640 }
1641 }
1642 }
1643
1644 // Traverse every Variable of every Inst and verify that it appears within the
1645 // Variable's computed live range.
validateLiveness() const1646 bool Cfg::validateLiveness() const {
1647 TimerMarker T(TimerStack::TT_validateLiveness, this);
1648 bool Valid = true;
1649 OstreamLocker L(Ctx);
1650 Ostream &Str = Ctx->getStrDump();
1651 for (CfgNode *Node : Nodes) {
1652 Inst *FirstInst = nullptr;
1653 for (Inst &Instr : Node->getInsts()) {
1654 if (Instr.isDeleted())
1655 continue;
1656 if (FirstInst == nullptr)
1657 FirstInst = &Instr;
1658 InstNumberT InstNumber = Instr.getNumber();
1659 if (Variable *Dest = Instr.getDest()) {
1660 if (!Dest->getIgnoreLiveness()) {
1661 bool Invalid = false;
1662 constexpr bool IsDest = true;
1663 if (!Dest->getLiveRange().containsValue(InstNumber, IsDest))
1664 Invalid = true;
1665 // Check that this instruction actually *begins* Dest's live range,
1666 // by checking that Dest is not live in the previous instruction. As
1667 // a special exception, we don't check this for the first instruction
1668 // of the block, because a Phi temporary may be live at the end of
1669 // the previous block, and if it is also assigned in the first
1670 // instruction of this block, the adjacent live ranges get merged.
1671 if (&Instr != FirstInst && !Instr.isDestRedefined() &&
1672 Dest->getLiveRange().containsValue(InstNumber - 1, IsDest))
1673 Invalid = true;
1674 if (Invalid) {
1675 Valid = false;
1676 Str << "Liveness error: inst " << Instr.getNumber() << " dest ";
1677 Dest->dump(this);
1678 Str << " live range " << Dest->getLiveRange() << "\n";
1679 }
1680 }
1681 }
1682 FOREACH_VAR_IN_INST(Var, Instr) {
1683 static constexpr bool IsDest = false;
1684 if (!Var->getIgnoreLiveness() &&
1685 !Var->getLiveRange().containsValue(InstNumber, IsDest)) {
1686 Valid = false;
1687 Str << "Liveness error: inst " << Instr.getNumber() << " var ";
1688 Var->dump(this);
1689 Str << " live range " << Var->getLiveRange() << "\n";
1690 }
1691 }
1692 }
1693 }
1694 return Valid;
1695 }
1696
contractEmptyNodes()1697 void Cfg::contractEmptyNodes() {
1698 // If we're decorating the asm output with register liveness info, this
1699 // information may become corrupted or incorrect after contracting nodes that
1700 // contain only redundant assignments. As such, we disable this pass when
1701 // DecorateAsm is specified. This may make the resulting code look more
1702 // branchy, but it should have no effect on the register assignments.
1703 if (getFlags().getDecorateAsm())
1704 return;
1705 for (CfgNode *Node : Nodes) {
1706 Node->contractIfEmpty();
1707 }
1708 }
1709
doBranchOpt()1710 void Cfg::doBranchOpt() {
1711 TimerMarker T(TimerStack::TT_doBranchOpt, this);
1712 for (auto I = Nodes.begin(), E = Nodes.end(); I != E; ++I) {
1713 auto NextNode = I + 1;
1714 (*I)->doBranchOpt(NextNode == E ? nullptr : *NextNode);
1715 }
1716 }
1717
markNodesForSandboxing()1718 void Cfg::markNodesForSandboxing() {
1719 for (const InstJumpTable *JT : JumpTables)
1720 for (SizeT I = 0; I < JT->getNumTargets(); ++I)
1721 JT->getTarget(I)->setNeedsAlignment();
1722 }
1723
1724 // ======================== Dump routines ======================== //
1725
1726 // emitTextHeader() is not target-specific (apart from what is abstracted by
1727 // the Assembler), so it is defined here rather than in the target lowering
1728 // class.
emitTextHeader(GlobalString Name,GlobalContext * Ctx,const Assembler * Asm)1729 void Cfg::emitTextHeader(GlobalString Name, GlobalContext *Ctx,
1730 const Assembler *Asm) {
1731 if (!BuildDefs::dump())
1732 return;
1733 Ostream &Str = Ctx->getStrEmit();
1734 Str << "\t.text\n";
1735 if (getFlags().getFunctionSections())
1736 Str << "\t.section\t.text." << Name << ",\"ax\",%progbits\n";
1737 if (!Asm->getInternal() || getFlags().getDisableInternal()) {
1738 Str << "\t.globl\t" << Name << "\n";
1739 Str << "\t.type\t" << Name << ",%function\n";
1740 }
1741 Str << "\t" << Asm->getAlignDirective() << " "
1742 << Asm->getBundleAlignLog2Bytes() << ",0x";
1743 for (uint8_t I : Asm->getNonExecBundlePadding())
1744 Str.write_hex(I);
1745 Str << "\n";
1746 Str << Name << ":\n";
1747 }
1748
emitJumpTables()1749 void Cfg::emitJumpTables() {
1750 switch (getFlags().getOutFileType()) {
1751 case FT_Elf:
1752 case FT_Iasm: {
1753 // The emission needs to be delayed until the after the text section so
1754 // save the offsets in the global context.
1755 for (const InstJumpTable *JumpTable : JumpTables) {
1756 Ctx->addJumpTableData(JumpTable->toJumpTableData(getAssembler()));
1757 }
1758 } break;
1759 case FT_Asm: {
1760 // Emit the assembly directly so we don't need to hang on to all the names
1761 for (const InstJumpTable *JumpTable : JumpTables)
1762 getTarget()->emitJumpTable(this, JumpTable);
1763 } break;
1764 }
1765 }
1766
emit()1767 void Cfg::emit() {
1768 if (!BuildDefs::dump())
1769 return;
1770 TimerMarker T(TimerStack::TT_emitAsm, this);
1771 if (getFlags().getDecorateAsm()) {
1772 renumberInstructions();
1773 getVMetadata()->init(VMK_Uses);
1774 liveness(Liveness_Basic);
1775 dump("After recomputing liveness for -decorate-asm");
1776 }
1777 OstreamLocker L(Ctx);
1778 Ostream &Str = Ctx->getStrEmit();
1779 const Assembler *Asm = getAssembler<>();
1780 const bool NeedSandboxing = getFlags().getUseSandboxing();
1781
1782 emitTextHeader(FunctionName, Ctx, Asm);
1783 if (getFlags().getDecorateAsm()) {
1784 for (Variable *Var : getVariables()) {
1785 if (Var->hasKnownStackOffset() && !Var->isRematerializable()) {
1786 Str << "\t" << Var->getSymbolicStackOffset() << " = "
1787 << Var->getStackOffset() << "\n";
1788 }
1789 }
1790 }
1791 for (CfgNode *Node : Nodes) {
1792 if (NeedSandboxing && Node->needsAlignment()) {
1793 Str << "\t" << Asm->getAlignDirective() << " "
1794 << Asm->getBundleAlignLog2Bytes() << "\n";
1795 }
1796 Node->emit(this);
1797 }
1798 emitJumpTables();
1799 Str << "\n";
1800 }
1801
emitIAS()1802 void Cfg::emitIAS() {
1803 TimerMarker T(TimerStack::TT_emitAsm, this);
1804 // The emitIAS() routines emit into the internal assembler buffer, so there's
1805 // no need to lock the streams.
1806 const bool NeedSandboxing = getFlags().getUseSandboxing();
1807 for (CfgNode *Node : Nodes) {
1808 if (NeedSandboxing && Node->needsAlignment())
1809 getAssembler()->alignCfgNode();
1810 Node->emitIAS(this);
1811 }
1812 emitJumpTables();
1813 }
1814
getTotalMemoryMB() const1815 size_t Cfg::getTotalMemoryMB() const {
1816 constexpr size_t _1MB = 1024 * 1024;
1817 assert(Allocator != nullptr);
1818 assert(CfgAllocatorTraits::current() == Allocator.get());
1819 return Allocator->getTotalMemory() / _1MB;
1820 }
1821
getLivenessMemoryMB() const1822 size_t Cfg::getLivenessMemoryMB() const {
1823 constexpr size_t _1MB = 1024 * 1024;
1824 if (Live == nullptr) {
1825 return 0;
1826 }
1827 return Live->getAllocator()->getTotalMemory() / _1MB;
1828 }
1829
1830 // Dumps the IR with an optional introductory message.
dump(const char * Message)1831 void Cfg::dump(const char *Message) {
1832 if (!BuildDefs::dump())
1833 return;
1834 if (!isVerbose())
1835 return;
1836 OstreamLocker L(Ctx);
1837 Ostream &Str = Ctx->getStrDump();
1838 if (Message[0])
1839 Str << "================ " << Message << " ================\n";
1840 if (isVerbose(IceV_Mem)) {
1841 Str << "Memory size = " << getTotalMemoryMB() << " MB\n";
1842 }
1843 setCurrentNode(getEntryNode());
1844 // Print function name+args
1845 if (isVerbose(IceV_Instructions)) {
1846 Str << "define ";
1847 if (getInternal() && !getFlags().getDisableInternal())
1848 Str << "internal ";
1849 Str << ReturnType << " @" << getFunctionName() << "(";
1850 for (SizeT i = 0; i < Args.size(); ++i) {
1851 if (i > 0)
1852 Str << ", ";
1853 Str << Args[i]->getType() << " ";
1854 Args[i]->dump(this);
1855 }
1856 // Append an extra copy of the function name here, in order to print its
1857 // size stats but not mess up lit tests.
1858 Str << ") { # " << getFunctionNameAndSize() << "\n";
1859 }
1860 resetCurrentNode();
1861 if (isVerbose(IceV_Liveness)) {
1862 // Print summary info about variables
1863 for (Variable *Var : Variables) {
1864 Str << "// multiblock=";
1865 if (getVMetadata()->isTracked(Var))
1866 Str << getVMetadata()->isMultiBlock(Var);
1867 else
1868 Str << "?";
1869 Str << " defs=";
1870 bool FirstPrint = true;
1871 if (VMetadata->getKind() != VMK_Uses) {
1872 if (const Inst *FirstDef = VMetadata->getFirstDefinition(Var)) {
1873 Str << FirstDef->getNumber();
1874 FirstPrint = false;
1875 }
1876 }
1877 if (VMetadata->getKind() == VMK_All) {
1878 for (const Inst *Instr : VMetadata->getLatterDefinitions(Var)) {
1879 if (!FirstPrint)
1880 Str << ",";
1881 Str << Instr->getNumber();
1882 FirstPrint = false;
1883 }
1884 }
1885 Str << " weight=" << Var->getWeight(this) << " ";
1886 Var->dump(this);
1887 Str << " LIVE=" << Var->getLiveRange() << "\n";
1888 }
1889 }
1890 // Print each basic block
1891 for (CfgNode *Node : Nodes)
1892 Node->dump(this);
1893 if (isVerbose(IceV_Instructions))
1894 Str << "}\n";
1895 }
1896
1897 } // end of namespace Ice
1898