• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 // Copyright (c) 2018 Google LLC.
2 //
3 // Licensed under the Apache License, Version 2.0 (the "License");
4 // you may not use this file except in compliance with the License.
5 // You may obtain a copy of the License at
6 //
7 //     http://www.apache.org/licenses/LICENSE-2.0
8 //
9 // Unless required by applicable law or agreed to in writing, software
10 // distributed under the License is distributed on an "AS IS" BASIS,
11 // WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
12 // See the License for the specific language governing permissions and
13 // limitations under the License.
14 
15 #include <algorithm>
16 #include <memory>
17 #include <unordered_map>
18 #include <unordered_set>
19 #include <utility>
20 #include <vector>
21 
22 #include "source/cfa.h"
23 #include "source/opt/cfg.h"
24 #include "source/opt/ir_builder.h"
25 #include "source/opt/ir_context.h"
26 #include "source/opt/loop_descriptor.h"
27 #include "source/opt/loop_utils.h"
28 
29 namespace spvtools {
30 namespace opt {
31 
32 namespace {
33 // Return true if |bb| is dominated by at least one block in |exits|
DominatesAnExit(BasicBlock * bb,const std::unordered_set<BasicBlock * > & exits,const DominatorTree & dom_tree)34 static inline bool DominatesAnExit(BasicBlock* bb,
35                                    const std::unordered_set<BasicBlock*>& exits,
36                                    const DominatorTree& dom_tree) {
37   for (BasicBlock* e_bb : exits)
38     if (dom_tree.Dominates(bb, e_bb)) return true;
39   return false;
40 }
41 
42 // Utility class to rewrite out-of-loop uses of an in-loop definition in terms
43 // of phi instructions to achieve a LCSSA form.
44 // For a given definition, the class user registers phi instructions using that
45 // definition in all loop exit blocks by which the definition escapes.
46 // Then, when rewriting a use of the definition, the rewriter walks the
47 // paths from the use the loop exits. At each step, it will insert a phi
48 // instruction to merge the incoming value according to exit blocks definition.
49 class LCSSARewriter {
50  public:
LCSSARewriter(IRContext * context,const DominatorTree & dom_tree,const std::unordered_set<BasicBlock * > & exit_bb,BasicBlock * merge_block)51   LCSSARewriter(IRContext* context, const DominatorTree& dom_tree,
52                 const std::unordered_set<BasicBlock*>& exit_bb,
53                 BasicBlock* merge_block)
54       : context_(context),
55         cfg_(context_->cfg()),
56         dom_tree_(dom_tree),
57         exit_bb_(exit_bb),
58         merge_block_id_(merge_block ? merge_block->id() : 0) {}
59 
60   struct UseRewriter {
UseRewriterspvtools::opt::__anoncf61d8e60111::LCSSARewriter::UseRewriter61     explicit UseRewriter(LCSSARewriter* base, const Instruction& def_insn)
62         : base_(base), def_insn_(def_insn) {}
63     // Rewrites the use of |def_insn_| by the instruction |user| at the index
64     // |operand_index| in terms of phi instruction. This recursively builds new
65     // phi instructions from |user| to the loop exit blocks' phis. The use of
66     // |def_insn_| in |user| is replaced by the relevant phi instruction at the
67     // end of the operation.
68     // It is assumed that |user| does not dominates any of the loop exit basic
69     // block. This operation does not update the def/use manager, instead it
70     // records what needs to be updated. The actual update is performed by
71     // UpdateManagers.
RewriteUsespvtools::opt::__anoncf61d8e60111::LCSSARewriter::UseRewriter72     void RewriteUse(BasicBlock* bb, Instruction* user, uint32_t operand_index) {
73       assert(
74           (user->opcode() != SpvOpPhi || bb != GetParent(user)) &&
75           "The root basic block must be the incoming edge if |user| is a phi "
76           "instruction");
77       assert((user->opcode() == SpvOpPhi || bb == GetParent(user)) &&
78              "The root basic block must be the instruction parent if |user| is "
79              "not "
80              "phi instruction");
81 
82       Instruction* new_def = GetOrBuildIncoming(bb->id());
83 
84       user->SetOperand(operand_index, {new_def->result_id()});
85       rewritten_.insert(user);
86     }
87 
88     // In-place update of some managers (avoid full invalidation).
UpdateManagersspvtools::opt::__anoncf61d8e60111::LCSSARewriter::UseRewriter89     inline void UpdateManagers() {
90       analysis::DefUseManager* def_use_mgr = base_->context_->get_def_use_mgr();
91       // Register all new definitions.
92       for (Instruction* insn : rewritten_) {
93         def_use_mgr->AnalyzeInstDef(insn);
94       }
95       // Register all new uses.
96       for (Instruction* insn : rewritten_) {
97         def_use_mgr->AnalyzeInstUse(insn);
98       }
99     }
100 
101    private:
102     // Return the basic block that |instr| belongs to.
GetParentspvtools::opt::__anoncf61d8e60111::LCSSARewriter::UseRewriter103     BasicBlock* GetParent(Instruction* instr) {
104       return base_->context_->get_instr_block(instr);
105     }
106 
107     // Builds a phi instruction for the basic block |bb|. The function assumes
108     // that |defining_blocks| contains the list of basic block that define the
109     // usable value for each predecessor of |bb|.
CreatePhiInstructionspvtools::opt::__anoncf61d8e60111::LCSSARewriter::UseRewriter110     inline Instruction* CreatePhiInstruction(
111         BasicBlock* bb, const std::vector<uint32_t>& defining_blocks) {
112       std::vector<uint32_t> incomings;
113       const std::vector<uint32_t>& bb_preds = base_->cfg_->preds(bb->id());
114       assert(bb_preds.size() == defining_blocks.size());
115       for (size_t i = 0; i < bb_preds.size(); i++) {
116         incomings.push_back(
117             GetOrBuildIncoming(defining_blocks[i])->result_id());
118         incomings.push_back(bb_preds[i]);
119       }
120       InstructionBuilder builder(base_->context_, &*bb->begin(),
121                                  IRContext::kAnalysisInstrToBlockMapping);
122       Instruction* incoming_phi =
123           builder.AddPhi(def_insn_.type_id(), incomings);
124 
125       rewritten_.insert(incoming_phi);
126       return incoming_phi;
127     }
128 
129     // Builds a phi instruction for the basic block |bb|, all incoming values
130     // will be |value|.
CreatePhiInstructionspvtools::opt::__anoncf61d8e60111::LCSSARewriter::UseRewriter131     inline Instruction* CreatePhiInstruction(BasicBlock* bb,
132                                              const Instruction& value) {
133       std::vector<uint32_t> incomings;
134       const std::vector<uint32_t>& bb_preds = base_->cfg_->preds(bb->id());
135       for (size_t i = 0; i < bb_preds.size(); i++) {
136         incomings.push_back(value.result_id());
137         incomings.push_back(bb_preds[i]);
138       }
139       InstructionBuilder builder(base_->context_, &*bb->begin(),
140                                  IRContext::kAnalysisInstrToBlockMapping);
141       Instruction* incoming_phi =
142           builder.AddPhi(def_insn_.type_id(), incomings);
143 
144       rewritten_.insert(incoming_phi);
145       return incoming_phi;
146     }
147 
148     // Return the new def to use for the basic block |bb_id|.
149     // If |bb_id| does not have a suitable def to use then we:
150     //   - return the common def used by all predecessors;
151     //   - if there is no common def, then we build a new phi instr at the
152     //     beginning of |bb_id| and return this new instruction.
GetOrBuildIncomingspvtools::opt::__anoncf61d8e60111::LCSSARewriter::UseRewriter153     Instruction* GetOrBuildIncoming(uint32_t bb_id) {
154       assert(base_->cfg_->block(bb_id) != nullptr && "Unknown basic block");
155 
156       Instruction*& incoming_phi = bb_to_phi_[bb_id];
157       if (incoming_phi) {
158         return incoming_phi;
159       }
160 
161       BasicBlock* bb = &*base_->cfg_->block(bb_id);
162       // If this is an exit basic block, look if there already is an eligible
163       // phi instruction. An eligible phi has |def_insn_| as all incoming
164       // values.
165       if (base_->exit_bb_.count(bb)) {
166         // Look if there is an eligible phi in this block.
167         if (!bb->WhileEachPhiInst([&incoming_phi, this](Instruction* phi) {
168               for (uint32_t i = 0; i < phi->NumInOperands(); i += 2) {
169                 if (phi->GetSingleWordInOperand(i) != def_insn_.result_id())
170                   return true;
171               }
172               incoming_phi = phi;
173               rewritten_.insert(incoming_phi);
174               return false;
175             })) {
176           return incoming_phi;
177         }
178         incoming_phi = CreatePhiInstruction(bb, def_insn_);
179         return incoming_phi;
180       }
181 
182       // Get the block that defines the value to use for each predecessor.
183       // If the vector has 1 value, then it means that this block does not need
184       // to build a phi instruction unless |bb_id| is the loop merge block.
185       const std::vector<uint32_t>& defining_blocks =
186           base_->GetDefiningBlocks(bb_id);
187 
188       // Special case for structured loops: merge block might be different from
189       // the exit block set. To maintain structured properties it will ease
190       // transformations if the merge block also holds a phi instruction like
191       // the exit ones.
192       if (defining_blocks.size() > 1 || bb_id == base_->merge_block_id_) {
193         if (defining_blocks.size() > 1) {
194           incoming_phi = CreatePhiInstruction(bb, defining_blocks);
195         } else {
196           assert(bb_id == base_->merge_block_id_);
197           incoming_phi =
198               CreatePhiInstruction(bb, *GetOrBuildIncoming(defining_blocks[0]));
199         }
200       } else {
201         incoming_phi = GetOrBuildIncoming(defining_blocks[0]);
202       }
203 
204       return incoming_phi;
205     }
206 
207     LCSSARewriter* base_;
208     const Instruction& def_insn_;
209     std::unordered_map<uint32_t, Instruction*> bb_to_phi_;
210     std::unordered_set<Instruction*> rewritten_;
211   };
212 
213  private:
214   // Return the new def to use for the basic block |bb_id|.
215   // If |bb_id| does not have a suitable def to use then we:
216   //   - return the common def used by all predecessors;
217   //   - if there is no common def, then we build a new phi instr at the
218   //     beginning of |bb_id| and return this new instruction.
GetDefiningBlocks(uint32_t bb_id)219   const std::vector<uint32_t>& GetDefiningBlocks(uint32_t bb_id) {
220     assert(cfg_->block(bb_id) != nullptr && "Unknown basic block");
221     std::vector<uint32_t>& defining_blocks = bb_to_defining_blocks_[bb_id];
222 
223     if (defining_blocks.size()) return defining_blocks;
224 
225     // Check if one of the loop exit basic block dominates |bb_id|.
226     for (const BasicBlock* e_bb : exit_bb_) {
227       if (dom_tree_.Dominates(e_bb->id(), bb_id)) {
228         defining_blocks.push_back(e_bb->id());
229         return defining_blocks;
230       }
231     }
232 
233     // Process parents, they will returns their suitable blocks.
234     // If they are all the same, this means this basic block is dominated by a
235     // common block, so we won't need to build a phi instruction.
236     for (uint32_t pred_id : cfg_->preds(bb_id)) {
237       const std::vector<uint32_t>& pred_blocks = GetDefiningBlocks(pred_id);
238       if (pred_blocks.size() == 1)
239         defining_blocks.push_back(pred_blocks[0]);
240       else
241         defining_blocks.push_back(pred_id);
242     }
243     assert(defining_blocks.size());
244     if (std::all_of(defining_blocks.begin(), defining_blocks.end(),
245                     [&defining_blocks](uint32_t id) {
246                       return id == defining_blocks[0];
247                     })) {
248       // No need for a phi.
249       defining_blocks.resize(1);
250     }
251 
252     return defining_blocks;
253   }
254 
255   IRContext* context_;
256   CFG* cfg_;
257   const DominatorTree& dom_tree_;
258   const std::unordered_set<BasicBlock*>& exit_bb_;
259   uint32_t merge_block_id_;
260   // This map represent the set of known paths. For each key, the vector
261   // represent the set of blocks holding the definition to be used to build the
262   // phi instruction.
263   // If the vector has 0 value, then the path is unknown yet, and must be built.
264   // If the vector has 1 value, then the value defined by that basic block
265   //   should be used.
266   // If the vector has more than 1 value, then a phi node must be created, the
267   //   basic block ordering is the same as the predecessor ordering.
268   std::unordered_map<uint32_t, std::vector<uint32_t>> bb_to_defining_blocks_;
269 };
270 
271 // Make the set |blocks| closed SSA. The set is closed SSA if all the uses
272 // outside the set are phi instructions in exiting basic block set (hold by
273 // |lcssa_rewriter|).
MakeSetClosedSSA(IRContext * context,Function * function,const std::unordered_set<uint32_t> & blocks,const std::unordered_set<BasicBlock * > & exit_bb,LCSSARewriter * lcssa_rewriter)274 inline void MakeSetClosedSSA(IRContext* context, Function* function,
275                              const std::unordered_set<uint32_t>& blocks,
276                              const std::unordered_set<BasicBlock*>& exit_bb,
277                              LCSSARewriter* lcssa_rewriter) {
278   CFG& cfg = *context->cfg();
279   DominatorTree& dom_tree =
280       context->GetDominatorAnalysis(function)->GetDomTree();
281   analysis::DefUseManager* def_use_manager = context->get_def_use_mgr();
282 
283   for (uint32_t bb_id : blocks) {
284     BasicBlock* bb = cfg.block(bb_id);
285     // If bb does not dominate an exit block, then it cannot have escaping defs.
286     if (!DominatesAnExit(bb, exit_bb, dom_tree)) continue;
287     for (Instruction& inst : *bb) {
288       LCSSARewriter::UseRewriter rewriter(lcssa_rewriter, inst);
289       def_use_manager->ForEachUse(
290           &inst, [&blocks, &rewriter, &exit_bb, context](
291                      Instruction* use, uint32_t operand_index) {
292             BasicBlock* use_parent = context->get_instr_block(use);
293             assert(use_parent);
294             if (blocks.count(use_parent->id())) return;
295 
296             if (use->opcode() == SpvOpPhi) {
297               // If the use is a Phi instruction and the incoming block is
298               // coming from the loop, then that's consistent with LCSSA form.
299               if (exit_bb.count(use_parent)) {
300                 return;
301               } else {
302                 // That's not an exit block, but the user is a phi instruction.
303                 // Consider the incoming branch only.
304                 use_parent = context->get_instr_block(
305                     use->GetSingleWordOperand(operand_index + 1));
306               }
307             }
308             // Rewrite the use. Note that this call does not invalidate the
309             // def/use manager. So this operation is safe.
310             rewriter.RewriteUse(use_parent, use, operand_index);
311           });
312       rewriter.UpdateManagers();
313     }
314   }
315 }
316 
317 }  // namespace
318 
CreateLoopDedicatedExits()319 void LoopUtils::CreateLoopDedicatedExits() {
320   Function* function = loop_->GetHeaderBlock()->GetParent();
321   LoopDescriptor& loop_desc = *context_->GetLoopDescriptor(function);
322   CFG& cfg = *context_->cfg();
323   analysis::DefUseManager* def_use_mgr = context_->get_def_use_mgr();
324 
325   const IRContext::Analysis PreservedAnalyses =
326       IRContext::kAnalysisDefUse | IRContext::kAnalysisInstrToBlockMapping;
327 
328   // Gathers the set of basic block that are not in this loop and have at least
329   // one predecessor in the loop and one not in the loop.
330   std::unordered_set<uint32_t> exit_bb_set;
331   loop_->GetExitBlocks(&exit_bb_set);
332 
333   std::unordered_set<BasicBlock*> new_loop_exits;
334   bool made_change = false;
335   // For each block, we create a new one that gathers all branches from
336   // the loop and fall into the block.
337   for (uint32_t non_dedicate_id : exit_bb_set) {
338     BasicBlock* non_dedicate = cfg.block(non_dedicate_id);
339     const std::vector<uint32_t>& bb_pred = cfg.preds(non_dedicate_id);
340     // Ignore the block if all the predecessors are in the loop.
341     if (std::all_of(bb_pred.begin(), bb_pred.end(),
342                     [this](uint32_t id) { return loop_->IsInsideLoop(id); })) {
343       new_loop_exits.insert(non_dedicate);
344       continue;
345     }
346 
347     made_change = true;
348     Function::iterator insert_pt = function->begin();
349     for (; insert_pt != function->end() && &*insert_pt != non_dedicate;
350          ++insert_pt) {
351     }
352     assert(insert_pt != function->end() && "Basic Block not found");
353 
354     // Create the dedicate exit basic block.
355     // TODO(1841): Handle id overflow.
356     BasicBlock& exit = *insert_pt.InsertBefore(std::unique_ptr<BasicBlock>(
357         new BasicBlock(std::unique_ptr<Instruction>(new Instruction(
358             context_, SpvOpLabel, 0, context_->TakeNextId(), {})))));
359     exit.SetParent(function);
360 
361     // Redirect in loop predecessors to |exit| block.
362     for (uint32_t exit_pred_id : bb_pred) {
363       if (loop_->IsInsideLoop(exit_pred_id)) {
364         BasicBlock* pred_block = cfg.block(exit_pred_id);
365         pred_block->ForEachSuccessorLabel([non_dedicate, &exit](uint32_t* id) {
366           if (*id == non_dedicate->id()) *id = exit.id();
367         });
368         // Update the CFG.
369         // |non_dedicate|'s predecessor list will be updated at the end of the
370         // loop.
371         cfg.RegisterBlock(pred_block);
372       }
373     }
374 
375     // Register the label to the def/use manager, requires for the phi patching.
376     def_use_mgr->AnalyzeInstDefUse(exit.GetLabelInst());
377     context_->set_instr_block(exit.GetLabelInst(), &exit);
378 
379     InstructionBuilder builder(context_, &exit, PreservedAnalyses);
380     // Now jump from our dedicate basic block to the old exit.
381     // We also reset the insert point so all instructions are inserted before
382     // the branch.
383     builder.SetInsertPoint(builder.AddBranch(non_dedicate->id()));
384     non_dedicate->ForEachPhiInst(
385         [&builder, &exit, def_use_mgr, this](Instruction* phi) {
386           // New phi operands for this instruction.
387           std::vector<uint32_t> new_phi_op;
388           // Phi operands for the dedicated exit block.
389           std::vector<uint32_t> exit_phi_op;
390           for (uint32_t i = 0; i < phi->NumInOperands(); i += 2) {
391             uint32_t def_id = phi->GetSingleWordInOperand(i);
392             uint32_t incoming_id = phi->GetSingleWordInOperand(i + 1);
393             if (loop_->IsInsideLoop(incoming_id)) {
394               exit_phi_op.push_back(def_id);
395               exit_phi_op.push_back(incoming_id);
396             } else {
397               new_phi_op.push_back(def_id);
398               new_phi_op.push_back(incoming_id);
399             }
400           }
401 
402           // Build the new phi instruction dedicated exit block.
403           Instruction* exit_phi = builder.AddPhi(phi->type_id(), exit_phi_op);
404           // Build the new incoming branch.
405           new_phi_op.push_back(exit_phi->result_id());
406           new_phi_op.push_back(exit.id());
407           // Rewrite operands.
408           uint32_t idx = 0;
409           for (; idx < new_phi_op.size(); idx++)
410             phi->SetInOperand(idx, {new_phi_op[idx]});
411           // Remove extra operands, from last to first (more efficient).
412           for (uint32_t j = phi->NumInOperands() - 1; j >= idx; j--)
413             phi->RemoveInOperand(j);
414           // Update the def/use manager for this |phi|.
415           def_use_mgr->AnalyzeInstUse(phi);
416         });
417     // Update the CFG.
418     cfg.RegisterBlock(&exit);
419     cfg.RemoveNonExistingEdges(non_dedicate->id());
420     new_loop_exits.insert(&exit);
421     // If non_dedicate is in a loop, add the new dedicated exit in that loop.
422     if (Loop* parent_loop = loop_desc[non_dedicate])
423       parent_loop->AddBasicBlock(&exit);
424   }
425 
426   if (new_loop_exits.size() == 1) {
427     loop_->SetMergeBlock(*new_loop_exits.begin());
428   }
429 
430   if (made_change) {
431     context_->InvalidateAnalysesExceptFor(
432         PreservedAnalyses | IRContext::kAnalysisCFG |
433         IRContext::Analysis::kAnalysisLoopAnalysis);
434   }
435 }
436 
MakeLoopClosedSSA()437 void LoopUtils::MakeLoopClosedSSA() {
438   CreateLoopDedicatedExits();
439 
440   Function* function = loop_->GetHeaderBlock()->GetParent();
441   CFG& cfg = *context_->cfg();
442   DominatorTree& dom_tree =
443       context_->GetDominatorAnalysis(function)->GetDomTree();
444 
445   std::unordered_set<BasicBlock*> exit_bb;
446   {
447     std::unordered_set<uint32_t> exit_bb_id;
448     loop_->GetExitBlocks(&exit_bb_id);
449     for (uint32_t bb_id : exit_bb_id) {
450       exit_bb.insert(cfg.block(bb_id));
451     }
452   }
453 
454   LCSSARewriter lcssa_rewriter(context_, dom_tree, exit_bb,
455                                loop_->GetMergeBlock());
456   MakeSetClosedSSA(context_, function, loop_->GetBlocks(), exit_bb,
457                    &lcssa_rewriter);
458 
459   // Make sure all defs post-dominated by the merge block have their last use no
460   // further than the merge block.
461   if (loop_->GetMergeBlock()) {
462     std::unordered_set<uint32_t> merging_bb_id;
463     loop_->GetMergingBlocks(&merging_bb_id);
464     merging_bb_id.erase(loop_->GetMergeBlock()->id());
465     // Reset the exit set, now only the merge block is the exit.
466     exit_bb.clear();
467     exit_bb.insert(loop_->GetMergeBlock());
468     // LCSSARewriter is reusable here only because it forces the creation of a
469     // phi instruction in the merge block.
470     MakeSetClosedSSA(context_, function, merging_bb_id, exit_bb,
471                      &lcssa_rewriter);
472   }
473 
474   context_->InvalidateAnalysesExceptFor(
475       IRContext::Analysis::kAnalysisCFG |
476       IRContext::Analysis::kAnalysisDominatorAnalysis |
477       IRContext::Analysis::kAnalysisLoopAnalysis);
478 }
479 
CloneLoop(LoopCloningResult * cloning_result) const480 Loop* LoopUtils::CloneLoop(LoopCloningResult* cloning_result) const {
481   // Compute the structured order of the loop basic blocks and store it in the
482   // vector ordered_loop_blocks.
483   std::vector<BasicBlock*> ordered_loop_blocks;
484   loop_->ComputeLoopStructuredOrder(&ordered_loop_blocks);
485 
486   // Clone the loop.
487   return CloneLoop(cloning_result, ordered_loop_blocks);
488 }
489 
CloneAndAttachLoopToHeader(LoopCloningResult * cloning_result)490 Loop* LoopUtils::CloneAndAttachLoopToHeader(LoopCloningResult* cloning_result) {
491   // Clone the loop.
492   Loop* new_loop = CloneLoop(cloning_result);
493 
494   // Create a new exit block/label for the new loop.
495   // TODO(1841): Handle id overflow.
496   std::unique_ptr<Instruction> new_label{new Instruction(
497       context_, SpvOp::SpvOpLabel, 0, context_->TakeNextId(), {})};
498   std::unique_ptr<BasicBlock> new_exit_bb{new BasicBlock(std::move(new_label))};
499   new_exit_bb->SetParent(loop_->GetMergeBlock()->GetParent());
500 
501   // Create an unconditional branch to the header block.
502   InstructionBuilder builder{context_, new_exit_bb.get()};
503   builder.AddBranch(loop_->GetHeaderBlock()->id());
504 
505   // Save the ids of the new and old merge block.
506   const uint32_t old_merge_block = loop_->GetMergeBlock()->id();
507   const uint32_t new_merge_block = new_exit_bb->id();
508 
509   // Replace the uses of the old merge block in the new loop with the new merge
510   // block.
511   for (std::unique_ptr<BasicBlock>& basic_block : cloning_result->cloned_bb_) {
512     for (Instruction& inst : *basic_block) {
513       // For each operand in each instruction check if it is using the old merge
514       // block and change it to be the new merge block.
515       auto replace_merge_use = [old_merge_block,
516                                 new_merge_block](uint32_t* id) {
517         if (*id == old_merge_block) *id = new_merge_block;
518       };
519       inst.ForEachInOperand(replace_merge_use);
520     }
521   }
522 
523   const uint32_t old_header = loop_->GetHeaderBlock()->id();
524   const uint32_t new_header = new_loop->GetHeaderBlock()->id();
525   analysis::DefUseManager* def_use = context_->get_def_use_mgr();
526 
527   def_use->ForEachUse(old_header,
528                       [new_header, this](Instruction* inst, uint32_t operand) {
529                         if (!this->loop_->IsInsideLoop(inst))
530                           inst->SetOperand(operand, {new_header});
531                       });
532 
533   // TODO(1841): Handle failure to create pre-header.
534   def_use->ForEachUse(
535       loop_->GetOrCreatePreHeaderBlock()->id(),
536       [new_merge_block, this](Instruction* inst, uint32_t operand) {
537         if (this->loop_->IsInsideLoop(inst))
538           inst->SetOperand(operand, {new_merge_block});
539 
540       });
541   new_loop->SetMergeBlock(new_exit_bb.get());
542 
543   new_loop->SetPreHeaderBlock(loop_->GetPreHeaderBlock());
544 
545   // Add the new block into the cloned instructions.
546   cloning_result->cloned_bb_.push_back(std::move(new_exit_bb));
547 
548   return new_loop;
549 }
550 
CloneLoop(LoopCloningResult * cloning_result,const std::vector<BasicBlock * > & ordered_loop_blocks) const551 Loop* LoopUtils::CloneLoop(
552     LoopCloningResult* cloning_result,
553     const std::vector<BasicBlock*>& ordered_loop_blocks) const {
554   analysis::DefUseManager* def_use_mgr = context_->get_def_use_mgr();
555 
556   std::unique_ptr<Loop> new_loop = MakeUnique<Loop>(context_);
557 
558   CFG& cfg = *context_->cfg();
559 
560   // Clone and place blocks in a SPIR-V compliant order (dominators first).
561   for (BasicBlock* old_bb : ordered_loop_blocks) {
562     // For each basic block in the loop, we clone it and register the mapping
563     // between old and new ids.
564     BasicBlock* new_bb = old_bb->Clone(context_);
565     new_bb->SetParent(&function_);
566     // TODO(1841): Handle id overflow.
567     new_bb->GetLabelInst()->SetResultId(context_->TakeNextId());
568     def_use_mgr->AnalyzeInstDef(new_bb->GetLabelInst());
569     context_->set_instr_block(new_bb->GetLabelInst(), new_bb);
570     cloning_result->cloned_bb_.emplace_back(new_bb);
571 
572     cloning_result->old_to_new_bb_[old_bb->id()] = new_bb;
573     cloning_result->new_to_old_bb_[new_bb->id()] = old_bb;
574     cloning_result->value_map_[old_bb->id()] = new_bb->id();
575 
576     if (loop_->IsInsideLoop(old_bb)) new_loop->AddBasicBlock(new_bb);
577 
578     for (auto new_inst = new_bb->begin(), old_inst = old_bb->begin();
579          new_inst != new_bb->end(); ++new_inst, ++old_inst) {
580       cloning_result->ptr_map_[&*new_inst] = &*old_inst;
581       if (new_inst->HasResultId()) {
582         // TODO(1841): Handle id overflow.
583         new_inst->SetResultId(context_->TakeNextId());
584         cloning_result->value_map_[old_inst->result_id()] =
585             new_inst->result_id();
586 
587         // Only look at the defs for now, uses are not updated yet.
588         def_use_mgr->AnalyzeInstDef(&*new_inst);
589       }
590     }
591   }
592 
593   // All instructions (including all labels) have been cloned,
594   // remap instruction operands id with the new ones.
595   for (std::unique_ptr<BasicBlock>& bb_ref : cloning_result->cloned_bb_) {
596     BasicBlock* bb = bb_ref.get();
597 
598     for (Instruction& insn : *bb) {
599       insn.ForEachInId([cloning_result](uint32_t* old_id) {
600         // If the operand is defined in the loop, remap the id.
601         auto id_it = cloning_result->value_map_.find(*old_id);
602         if (id_it != cloning_result->value_map_.end()) {
603           *old_id = id_it->second;
604         }
605       });
606       // Only look at what the instruction uses. All defs are register, so all
607       // should be fine now.
608       def_use_mgr->AnalyzeInstUse(&insn);
609       context_->set_instr_block(&insn, bb);
610     }
611     cfg.RegisterBlock(bb);
612   }
613 
614   PopulateLoopNest(new_loop.get(), *cloning_result);
615 
616   return new_loop.release();
617 }
618 
PopulateLoopNest(Loop * new_loop,const LoopCloningResult & cloning_result) const619 void LoopUtils::PopulateLoopNest(
620     Loop* new_loop, const LoopCloningResult& cloning_result) const {
621   std::unordered_map<Loop*, Loop*> loop_mapping;
622   loop_mapping[loop_] = new_loop;
623 
624   if (loop_->HasParent()) loop_->GetParent()->AddNestedLoop(new_loop);
625   PopulateLoopDesc(new_loop, loop_, cloning_result);
626 
627   for (Loop& sub_loop :
628        make_range(++TreeDFIterator<Loop>(loop_), TreeDFIterator<Loop>())) {
629     Loop* cloned = new Loop(context_);
630     if (Loop* parent = loop_mapping[sub_loop.GetParent()])
631       parent->AddNestedLoop(cloned);
632     loop_mapping[&sub_loop] = cloned;
633     PopulateLoopDesc(cloned, &sub_loop, cloning_result);
634   }
635 
636   loop_desc_->AddLoopNest(std::unique_ptr<Loop>(new_loop));
637 }
638 
639 // Populates |new_loop| descriptor according to |old_loop|'s one.
PopulateLoopDesc(Loop * new_loop,Loop * old_loop,const LoopCloningResult & cloning_result) const640 void LoopUtils::PopulateLoopDesc(
641     Loop* new_loop, Loop* old_loop,
642     const LoopCloningResult& cloning_result) const {
643   for (uint32_t bb_id : old_loop->GetBlocks()) {
644     BasicBlock* bb = cloning_result.old_to_new_bb_.at(bb_id);
645     new_loop->AddBasicBlock(bb);
646   }
647   new_loop->SetHeaderBlock(
648       cloning_result.old_to_new_bb_.at(old_loop->GetHeaderBlock()->id()));
649   if (old_loop->GetLatchBlock())
650     new_loop->SetLatchBlock(
651         cloning_result.old_to_new_bb_.at(old_loop->GetLatchBlock()->id()));
652   if (old_loop->GetContinueBlock())
653     new_loop->SetContinueBlock(
654         cloning_result.old_to_new_bb_.at(old_loop->GetContinueBlock()->id()));
655   if (old_loop->GetMergeBlock()) {
656     auto it =
657         cloning_result.old_to_new_bb_.find(old_loop->GetMergeBlock()->id());
658     BasicBlock* bb = it != cloning_result.old_to_new_bb_.end()
659                          ? it->second
660                          : old_loop->GetMergeBlock();
661     new_loop->SetMergeBlock(bb);
662   }
663   if (old_loop->GetPreHeaderBlock()) {
664     auto it =
665         cloning_result.old_to_new_bb_.find(old_loop->GetPreHeaderBlock()->id());
666     if (it != cloning_result.old_to_new_bb_.end()) {
667       new_loop->SetPreHeaderBlock(it->second);
668     }
669   }
670 }
671 
672 // Class to gather some metrics about a region of interest.
Analyze(const Loop & loop)673 void CodeMetrics::Analyze(const Loop& loop) {
674   CFG& cfg = *loop.GetContext()->cfg();
675 
676   roi_size_ = 0;
677   block_sizes_.clear();
678 
679   for (uint32_t id : loop.GetBlocks()) {
680     const BasicBlock* bb = cfg.block(id);
681     size_t bb_size = 0;
682     bb->ForEachInst([&bb_size](const Instruction* insn) {
683       if (insn->opcode() == SpvOpLabel) return;
684       if (insn->IsNop()) return;
685       if (insn->opcode() == SpvOpPhi) return;
686       bb_size++;
687     });
688     block_sizes_[bb->id()] = bb_size;
689     roi_size_ += bb_size;
690   }
691 }
692 
693 }  // namespace opt
694 }  // namespace spvtools
695