1 // Copyright (c) 2017 The Khronos Group Inc.
2 // Copyright (c) 2017 Valve Corporation
3 // Copyright (c) 2017 LunarG Inc.
4 //
5 // Licensed under the Apache License, Version 2.0 (the "License");
6 // you may not use this file except in compliance with the License.
7 // You may obtain a copy of the License at
8 //
9 // http://www.apache.org/licenses/LICENSE-2.0
10 //
11 // Unless required by applicable law or agreed to in writing, software
12 // distributed under the License is distributed on an "AS IS" BASIS,
13 // WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
14 // See the License for the specific language governing permissions and
15 // limitations under the License.
16
17 #include "source/opt/inline_pass.h"
18
19 #include <unordered_set>
20 #include <utility>
21
22 #include "source/cfa.h"
23 #include "source/opt/reflect.h"
24 #include "source/util/make_unique.h"
25
26 // Indices of operands in SPIR-V instructions
27
28 static const int kSpvFunctionCallFunctionId = 2;
29 static const int kSpvFunctionCallArgumentId = 3;
30 static const int kSpvReturnValueId = 0;
31
32 namespace spvtools {
33 namespace opt {
34
AddPointerToType(uint32_t type_id,SpvStorageClass storage_class)35 uint32_t InlinePass::AddPointerToType(uint32_t type_id,
36 SpvStorageClass storage_class) {
37 uint32_t resultId = context()->TakeNextId();
38 if (resultId == 0) {
39 return resultId;
40 }
41
42 std::unique_ptr<Instruction> type_inst(
43 new Instruction(context(), SpvOpTypePointer, 0, resultId,
44 {{spv_operand_type_t::SPV_OPERAND_TYPE_STORAGE_CLASS,
45 {uint32_t(storage_class)}},
46 {spv_operand_type_t::SPV_OPERAND_TYPE_ID, {type_id}}}));
47 context()->AddType(std::move(type_inst));
48 analysis::Type* pointeeTy;
49 std::unique_ptr<analysis::Pointer> pointerTy;
50 std::tie(pointeeTy, pointerTy) =
51 context()->get_type_mgr()->GetTypeAndPointerType(type_id,
52 SpvStorageClassFunction);
53 context()->get_type_mgr()->RegisterType(resultId, *pointerTy);
54 return resultId;
55 }
56
AddBranch(uint32_t label_id,std::unique_ptr<BasicBlock> * block_ptr)57 void InlinePass::AddBranch(uint32_t label_id,
58 std::unique_ptr<BasicBlock>* block_ptr) {
59 std::unique_ptr<Instruction> newBranch(
60 new Instruction(context(), SpvOpBranch, 0, 0,
61 {{spv_operand_type_t::SPV_OPERAND_TYPE_ID, {label_id}}}));
62 (*block_ptr)->AddInstruction(std::move(newBranch));
63 }
64
AddBranchCond(uint32_t cond_id,uint32_t true_id,uint32_t false_id,std::unique_ptr<BasicBlock> * block_ptr)65 void InlinePass::AddBranchCond(uint32_t cond_id, uint32_t true_id,
66 uint32_t false_id,
67 std::unique_ptr<BasicBlock>* block_ptr) {
68 std::unique_ptr<Instruction> newBranch(
69 new Instruction(context(), SpvOpBranchConditional, 0, 0,
70 {{spv_operand_type_t::SPV_OPERAND_TYPE_ID, {cond_id}},
71 {spv_operand_type_t::SPV_OPERAND_TYPE_ID, {true_id}},
72 {spv_operand_type_t::SPV_OPERAND_TYPE_ID, {false_id}}}));
73 (*block_ptr)->AddInstruction(std::move(newBranch));
74 }
75
AddLoopMerge(uint32_t merge_id,uint32_t continue_id,std::unique_ptr<BasicBlock> * block_ptr)76 void InlinePass::AddLoopMerge(uint32_t merge_id, uint32_t continue_id,
77 std::unique_ptr<BasicBlock>* block_ptr) {
78 std::unique_ptr<Instruction> newLoopMerge(new Instruction(
79 context(), SpvOpLoopMerge, 0, 0,
80 {{spv_operand_type_t::SPV_OPERAND_TYPE_ID, {merge_id}},
81 {spv_operand_type_t::SPV_OPERAND_TYPE_ID, {continue_id}},
82 {spv_operand_type_t::SPV_OPERAND_TYPE_LOOP_CONTROL, {0}}}));
83 (*block_ptr)->AddInstruction(std::move(newLoopMerge));
84 }
85
AddStore(uint32_t ptr_id,uint32_t val_id,std::unique_ptr<BasicBlock> * block_ptr,const Instruction * line_inst,const DebugScope & dbg_scope)86 void InlinePass::AddStore(uint32_t ptr_id, uint32_t val_id,
87 std::unique_ptr<BasicBlock>* block_ptr,
88 const Instruction* line_inst,
89 const DebugScope& dbg_scope) {
90 std::unique_ptr<Instruction> newStore(
91 new Instruction(context(), SpvOpStore, 0, 0,
92 {{spv_operand_type_t::SPV_OPERAND_TYPE_ID, {ptr_id}},
93 {spv_operand_type_t::SPV_OPERAND_TYPE_ID, {val_id}}}));
94 if (line_inst != nullptr) {
95 newStore->AddDebugLine(line_inst);
96 }
97 newStore->SetDebugScope(dbg_scope);
98 (*block_ptr)->AddInstruction(std::move(newStore));
99 }
100
AddLoad(uint32_t type_id,uint32_t resultId,uint32_t ptr_id,std::unique_ptr<BasicBlock> * block_ptr,const Instruction * line_inst,const DebugScope & dbg_scope)101 void InlinePass::AddLoad(uint32_t type_id, uint32_t resultId, uint32_t ptr_id,
102 std::unique_ptr<BasicBlock>* block_ptr,
103 const Instruction* line_inst,
104 const DebugScope& dbg_scope) {
105 std::unique_ptr<Instruction> newLoad(
106 new Instruction(context(), SpvOpLoad, type_id, resultId,
107 {{spv_operand_type_t::SPV_OPERAND_TYPE_ID, {ptr_id}}}));
108 if (line_inst != nullptr) {
109 newLoad->AddDebugLine(line_inst);
110 }
111 newLoad->SetDebugScope(dbg_scope);
112 (*block_ptr)->AddInstruction(std::move(newLoad));
113 }
114
NewLabel(uint32_t label_id)115 std::unique_ptr<Instruction> InlinePass::NewLabel(uint32_t label_id) {
116 std::unique_ptr<Instruction> newLabel(
117 new Instruction(context(), SpvOpLabel, 0, label_id, {}));
118 return newLabel;
119 }
120
GetFalseId()121 uint32_t InlinePass::GetFalseId() {
122 if (false_id_ != 0) return false_id_;
123 false_id_ = get_module()->GetGlobalValue(SpvOpConstantFalse);
124 if (false_id_ != 0) return false_id_;
125 uint32_t boolId = get_module()->GetGlobalValue(SpvOpTypeBool);
126 if (boolId == 0) {
127 boolId = context()->TakeNextId();
128 if (boolId == 0) {
129 return 0;
130 }
131 get_module()->AddGlobalValue(SpvOpTypeBool, boolId, 0);
132 }
133 false_id_ = context()->TakeNextId();
134 if (false_id_ == 0) {
135 return 0;
136 }
137 get_module()->AddGlobalValue(SpvOpConstantFalse, false_id_, boolId);
138 return false_id_;
139 }
140
MapParams(Function * calleeFn,BasicBlock::iterator call_inst_itr,std::unordered_map<uint32_t,uint32_t> * callee2caller)141 void InlinePass::MapParams(
142 Function* calleeFn, BasicBlock::iterator call_inst_itr,
143 std::unordered_map<uint32_t, uint32_t>* callee2caller) {
144 int param_idx = 0;
145 calleeFn->ForEachParam(
146 [&call_inst_itr, ¶m_idx, &callee2caller](const Instruction* cpi) {
147 const uint32_t pid = cpi->result_id();
148 (*callee2caller)[pid] = call_inst_itr->GetSingleWordOperand(
149 kSpvFunctionCallArgumentId + param_idx);
150 ++param_idx;
151 });
152 }
153
CloneAndMapLocals(Function * calleeFn,std::vector<std::unique_ptr<Instruction>> * new_vars,std::unordered_map<uint32_t,uint32_t> * callee2caller,analysis::DebugInlinedAtContext * inlined_at_ctx)154 bool InlinePass::CloneAndMapLocals(
155 Function* calleeFn, std::vector<std::unique_ptr<Instruction>>* new_vars,
156 std::unordered_map<uint32_t, uint32_t>* callee2caller,
157 analysis::DebugInlinedAtContext* inlined_at_ctx) {
158 auto callee_block_itr = calleeFn->begin();
159 auto callee_var_itr = callee_block_itr->begin();
160 while (callee_var_itr->opcode() == SpvOp::SpvOpVariable ||
161 callee_var_itr->GetCommonDebugOpcode() ==
162 CommonDebugInfoDebugDeclare) {
163 if (callee_var_itr->opcode() != SpvOp::SpvOpVariable) {
164 ++callee_var_itr;
165 continue;
166 }
167
168 std::unique_ptr<Instruction> var_inst(callee_var_itr->Clone(context()));
169 uint32_t newId = context()->TakeNextId();
170 if (newId == 0) {
171 return false;
172 }
173 get_decoration_mgr()->CloneDecorations(callee_var_itr->result_id(), newId);
174 var_inst->SetResultId(newId);
175 var_inst->UpdateDebugInlinedAt(
176 context()->get_debug_info_mgr()->BuildDebugInlinedAtChain(
177 callee_var_itr->GetDebugInlinedAt(), inlined_at_ctx));
178 (*callee2caller)[callee_var_itr->result_id()] = newId;
179 new_vars->push_back(std::move(var_inst));
180 ++callee_var_itr;
181 }
182 return true;
183 }
184
CreateReturnVar(Function * calleeFn,std::vector<std::unique_ptr<Instruction>> * new_vars)185 uint32_t InlinePass::CreateReturnVar(
186 Function* calleeFn, std::vector<std::unique_ptr<Instruction>>* new_vars) {
187 uint32_t returnVarId = 0;
188 const uint32_t calleeTypeId = calleeFn->type_id();
189 analysis::TypeManager* type_mgr = context()->get_type_mgr();
190 assert(type_mgr->GetType(calleeTypeId)->AsVoid() == nullptr &&
191 "Cannot create a return variable of type void.");
192 // Find or create ptr to callee return type.
193 uint32_t returnVarTypeId =
194 type_mgr->FindPointerToType(calleeTypeId, SpvStorageClassFunction);
195
196 if (returnVarTypeId == 0) {
197 returnVarTypeId = AddPointerToType(calleeTypeId, SpvStorageClassFunction);
198 if (returnVarTypeId == 0) {
199 return 0;
200 }
201 }
202
203 // Add return var to new function scope variables.
204 returnVarId = context()->TakeNextId();
205 if (returnVarId == 0) {
206 return 0;
207 }
208
209 std::unique_ptr<Instruction> var_inst(
210 new Instruction(context(), SpvOpVariable, returnVarTypeId, returnVarId,
211 {{spv_operand_type_t::SPV_OPERAND_TYPE_STORAGE_CLASS,
212 {SpvStorageClassFunction}}}));
213 new_vars->push_back(std::move(var_inst));
214 get_decoration_mgr()->CloneDecorations(calleeFn->result_id(), returnVarId);
215 return returnVarId;
216 }
217
IsSameBlockOp(const Instruction * inst) const218 bool InlinePass::IsSameBlockOp(const Instruction* inst) const {
219 return inst->opcode() == SpvOpSampledImage || inst->opcode() == SpvOpImage;
220 }
221
CloneSameBlockOps(std::unique_ptr<Instruction> * inst,std::unordered_map<uint32_t,uint32_t> * postCallSB,std::unordered_map<uint32_t,Instruction * > * preCallSB,std::unique_ptr<BasicBlock> * block_ptr)222 bool InlinePass::CloneSameBlockOps(
223 std::unique_ptr<Instruction>* inst,
224 std::unordered_map<uint32_t, uint32_t>* postCallSB,
225 std::unordered_map<uint32_t, Instruction*>* preCallSB,
226 std::unique_ptr<BasicBlock>* block_ptr) {
227 return (*inst)->WhileEachInId([&postCallSB, &preCallSB, &block_ptr,
228 this](uint32_t* iid) {
229 const auto mapItr = (*postCallSB).find(*iid);
230 if (mapItr == (*postCallSB).end()) {
231 const auto mapItr2 = (*preCallSB).find(*iid);
232 if (mapItr2 != (*preCallSB).end()) {
233 // Clone pre-call same-block ops, map result id.
234 const Instruction* inInst = mapItr2->second;
235 std::unique_ptr<Instruction> sb_inst(inInst->Clone(context()));
236 if (!CloneSameBlockOps(&sb_inst, postCallSB, preCallSB, block_ptr)) {
237 return false;
238 }
239
240 const uint32_t rid = sb_inst->result_id();
241 const uint32_t nid = context()->TakeNextId();
242 if (nid == 0) {
243 return false;
244 }
245 get_decoration_mgr()->CloneDecorations(rid, nid);
246 sb_inst->SetResultId(nid);
247 (*postCallSB)[rid] = nid;
248 *iid = nid;
249 (*block_ptr)->AddInstruction(std::move(sb_inst));
250 }
251 } else {
252 // Reset same-block op operand.
253 *iid = mapItr->second;
254 }
255 return true;
256 });
257 }
258
MoveInstsBeforeEntryBlock(std::unordered_map<uint32_t,Instruction * > * preCallSB,BasicBlock * new_blk_ptr,BasicBlock::iterator call_inst_itr,UptrVectorIterator<BasicBlock> call_block_itr)259 void InlinePass::MoveInstsBeforeEntryBlock(
260 std::unordered_map<uint32_t, Instruction*>* preCallSB,
261 BasicBlock* new_blk_ptr, BasicBlock::iterator call_inst_itr,
262 UptrVectorIterator<BasicBlock> call_block_itr) {
263 for (auto cii = call_block_itr->begin(); cii != call_inst_itr;
264 cii = call_block_itr->begin()) {
265 Instruction* inst = &*cii;
266 inst->RemoveFromList();
267 std::unique_ptr<Instruction> cp_inst(inst);
268 // Remember same-block ops for possible regeneration.
269 if (IsSameBlockOp(&*cp_inst)) {
270 auto* sb_inst_ptr = cp_inst.get();
271 (*preCallSB)[cp_inst->result_id()] = sb_inst_ptr;
272 }
273 new_blk_ptr->AddInstruction(std::move(cp_inst));
274 }
275 }
276
AddGuardBlock(std::vector<std::unique_ptr<BasicBlock>> * new_blocks,std::unordered_map<uint32_t,uint32_t> * callee2caller,std::unique_ptr<BasicBlock> new_blk_ptr,uint32_t entry_blk_label_id)277 std::unique_ptr<BasicBlock> InlinePass::AddGuardBlock(
278 std::vector<std::unique_ptr<BasicBlock>>* new_blocks,
279 std::unordered_map<uint32_t, uint32_t>* callee2caller,
280 std::unique_ptr<BasicBlock> new_blk_ptr, uint32_t entry_blk_label_id) {
281 const auto guard_block_id = context()->TakeNextId();
282 if (guard_block_id == 0) {
283 return nullptr;
284 }
285 AddBranch(guard_block_id, &new_blk_ptr);
286 new_blocks->push_back(std::move(new_blk_ptr));
287 // Start the next block.
288 new_blk_ptr = MakeUnique<BasicBlock>(NewLabel(guard_block_id));
289 // Reset the mapping of the callee's entry block to point to
290 // the guard block. Do this so we can fix up phis later on to
291 // satisfy dominance.
292 (*callee2caller)[entry_blk_label_id] = guard_block_id;
293 return new_blk_ptr;
294 }
295
AddStoresForVariableInitializers(const std::unordered_map<uint32_t,uint32_t> & callee2caller,analysis::DebugInlinedAtContext * inlined_at_ctx,std::unique_ptr<BasicBlock> * new_blk_ptr,UptrVectorIterator<BasicBlock> callee_first_block_itr)296 InstructionList::iterator InlinePass::AddStoresForVariableInitializers(
297 const std::unordered_map<uint32_t, uint32_t>& callee2caller,
298 analysis::DebugInlinedAtContext* inlined_at_ctx,
299 std::unique_ptr<BasicBlock>* new_blk_ptr,
300 UptrVectorIterator<BasicBlock> callee_first_block_itr) {
301 auto callee_itr = callee_first_block_itr->begin();
302 while (callee_itr->opcode() == SpvOp::SpvOpVariable ||
303 callee_itr->GetCommonDebugOpcode() == CommonDebugInfoDebugDeclare) {
304 if (callee_itr->opcode() == SpvOp::SpvOpVariable &&
305 callee_itr->NumInOperands() == 2) {
306 assert(callee2caller.count(callee_itr->result_id()) &&
307 "Expected the variable to have already been mapped.");
308 uint32_t new_var_id = callee2caller.at(callee_itr->result_id());
309
310 // The initializer must be a constant or global value. No mapped
311 // should be used.
312 uint32_t val_id = callee_itr->GetSingleWordInOperand(1);
313 AddStore(new_var_id, val_id, new_blk_ptr, callee_itr->dbg_line_inst(),
314 context()->get_debug_info_mgr()->BuildDebugScope(
315 callee_itr->GetDebugScope(), inlined_at_ctx));
316 }
317 if (callee_itr->GetCommonDebugOpcode() == CommonDebugInfoDebugDeclare) {
318 InlineSingleInstruction(
319 callee2caller, new_blk_ptr->get(), &*callee_itr,
320 context()->get_debug_info_mgr()->BuildDebugInlinedAtChain(
321 callee_itr->GetDebugScope().GetInlinedAt(), inlined_at_ctx));
322 }
323 ++callee_itr;
324 }
325 return callee_itr;
326 }
327
InlineSingleInstruction(const std::unordered_map<uint32_t,uint32_t> & callee2caller,BasicBlock * new_blk_ptr,const Instruction * inst,uint32_t dbg_inlined_at)328 bool InlinePass::InlineSingleInstruction(
329 const std::unordered_map<uint32_t, uint32_t>& callee2caller,
330 BasicBlock* new_blk_ptr, const Instruction* inst, uint32_t dbg_inlined_at) {
331 // If we have return, it must be at the end of the callee. We will handle
332 // it at the end.
333 if (inst->opcode() == SpvOpReturnValue || inst->opcode() == SpvOpReturn)
334 return true;
335
336 // Copy callee instruction and remap all input Ids.
337 std::unique_ptr<Instruction> cp_inst(inst->Clone(context()));
338 cp_inst->ForEachInId([&callee2caller](uint32_t* iid) {
339 const auto mapItr = callee2caller.find(*iid);
340 if (mapItr != callee2caller.end()) {
341 *iid = mapItr->second;
342 }
343 });
344
345 // If result id is non-zero, remap it.
346 const uint32_t rid = cp_inst->result_id();
347 if (rid != 0) {
348 const auto mapItr = callee2caller.find(rid);
349 if (mapItr == callee2caller.end()) {
350 return false;
351 }
352 uint32_t nid = mapItr->second;
353 cp_inst->SetResultId(nid);
354 get_decoration_mgr()->CloneDecorations(rid, nid);
355 }
356
357 cp_inst->UpdateDebugInlinedAt(dbg_inlined_at);
358 new_blk_ptr->AddInstruction(std::move(cp_inst));
359 return true;
360 }
361
InlineReturn(const std::unordered_map<uint32_t,uint32_t> & callee2caller,std::vector<std::unique_ptr<BasicBlock>> * new_blocks,std::unique_ptr<BasicBlock> new_blk_ptr,analysis::DebugInlinedAtContext * inlined_at_ctx,Function * calleeFn,const Instruction * inst,uint32_t returnVarId)362 std::unique_ptr<BasicBlock> InlinePass::InlineReturn(
363 const std::unordered_map<uint32_t, uint32_t>& callee2caller,
364 std::vector<std::unique_ptr<BasicBlock>>* new_blocks,
365 std::unique_ptr<BasicBlock> new_blk_ptr,
366 analysis::DebugInlinedAtContext* inlined_at_ctx, Function* calleeFn,
367 const Instruction* inst, uint32_t returnVarId) {
368 // Store return value to return variable.
369 if (inst->opcode() == SpvOpReturnValue) {
370 assert(returnVarId != 0);
371 uint32_t valId = inst->GetInOperand(kSpvReturnValueId).words[0];
372 const auto mapItr = callee2caller.find(valId);
373 if (mapItr != callee2caller.end()) {
374 valId = mapItr->second;
375 }
376 AddStore(returnVarId, valId, &new_blk_ptr, inst->dbg_line_inst(),
377 context()->get_debug_info_mgr()->BuildDebugScope(
378 inst->GetDebugScope(), inlined_at_ctx));
379 }
380
381 uint32_t returnLabelId = 0;
382 for (auto callee_block_itr = calleeFn->begin();
383 callee_block_itr != calleeFn->end(); ++callee_block_itr) {
384 if (spvOpcodeIsAbort(callee_block_itr->tail()->opcode())) {
385 returnLabelId = context()->TakeNextId();
386 break;
387 }
388 }
389 if (returnLabelId == 0) return new_blk_ptr;
390
391 if (inst->opcode() == SpvOpReturn || inst->opcode() == SpvOpReturnValue)
392 AddBranch(returnLabelId, &new_blk_ptr);
393 new_blocks->push_back(std::move(new_blk_ptr));
394 return MakeUnique<BasicBlock>(NewLabel(returnLabelId));
395 }
396
InlineEntryBlock(const std::unordered_map<uint32_t,uint32_t> & callee2caller,std::unique_ptr<BasicBlock> * new_blk_ptr,UptrVectorIterator<BasicBlock> callee_first_block,analysis::DebugInlinedAtContext * inlined_at_ctx)397 bool InlinePass::InlineEntryBlock(
398 const std::unordered_map<uint32_t, uint32_t>& callee2caller,
399 std::unique_ptr<BasicBlock>* new_blk_ptr,
400 UptrVectorIterator<BasicBlock> callee_first_block,
401 analysis::DebugInlinedAtContext* inlined_at_ctx) {
402 auto callee_inst_itr = AddStoresForVariableInitializers(
403 callee2caller, inlined_at_ctx, new_blk_ptr, callee_first_block);
404
405 while (callee_inst_itr != callee_first_block->end()) {
406 // Don't inline function definition links, the calling function is not a
407 // definition.
408 if (callee_inst_itr->GetShader100DebugOpcode() ==
409 NonSemanticShaderDebugInfo100DebugFunctionDefinition) {
410 ++callee_inst_itr;
411 continue;
412 }
413
414 if (!InlineSingleInstruction(
415 callee2caller, new_blk_ptr->get(), &*callee_inst_itr,
416 context()->get_debug_info_mgr()->BuildDebugInlinedAtChain(
417 callee_inst_itr->GetDebugScope().GetInlinedAt(),
418 inlined_at_ctx))) {
419 return false;
420 }
421 ++callee_inst_itr;
422 }
423 return true;
424 }
425
InlineBasicBlocks(std::vector<std::unique_ptr<BasicBlock>> * new_blocks,const std::unordered_map<uint32_t,uint32_t> & callee2caller,std::unique_ptr<BasicBlock> new_blk_ptr,analysis::DebugInlinedAtContext * inlined_at_ctx,Function * calleeFn)426 std::unique_ptr<BasicBlock> InlinePass::InlineBasicBlocks(
427 std::vector<std::unique_ptr<BasicBlock>>* new_blocks,
428 const std::unordered_map<uint32_t, uint32_t>& callee2caller,
429 std::unique_ptr<BasicBlock> new_blk_ptr,
430 analysis::DebugInlinedAtContext* inlined_at_ctx, Function* calleeFn) {
431 auto callee_block_itr = calleeFn->begin();
432 ++callee_block_itr;
433
434 while (callee_block_itr != calleeFn->end()) {
435 new_blocks->push_back(std::move(new_blk_ptr));
436 const auto mapItr =
437 callee2caller.find(callee_block_itr->GetLabelInst()->result_id());
438 if (mapItr == callee2caller.end()) return nullptr;
439 new_blk_ptr = MakeUnique<BasicBlock>(NewLabel(mapItr->second));
440
441 auto tail_inst_itr = callee_block_itr->end();
442 for (auto inst_itr = callee_block_itr->begin(); inst_itr != tail_inst_itr;
443 ++inst_itr) {
444 // Don't inline function definition links, the calling function is not a
445 // definition
446 if (inst_itr->GetShader100DebugOpcode() ==
447 NonSemanticShaderDebugInfo100DebugFunctionDefinition)
448 continue;
449 if (!InlineSingleInstruction(
450 callee2caller, new_blk_ptr.get(), &*inst_itr,
451 context()->get_debug_info_mgr()->BuildDebugInlinedAtChain(
452 inst_itr->GetDebugScope().GetInlinedAt(), inlined_at_ctx))) {
453 return nullptr;
454 }
455 }
456
457 ++callee_block_itr;
458 }
459 return new_blk_ptr;
460 }
461
MoveCallerInstsAfterFunctionCall(std::unordered_map<uint32_t,Instruction * > * preCallSB,std::unordered_map<uint32_t,uint32_t> * postCallSB,std::unique_ptr<BasicBlock> * new_blk_ptr,BasicBlock::iterator call_inst_itr,bool multiBlocks)462 bool InlinePass::MoveCallerInstsAfterFunctionCall(
463 std::unordered_map<uint32_t, Instruction*>* preCallSB,
464 std::unordered_map<uint32_t, uint32_t>* postCallSB,
465 std::unique_ptr<BasicBlock>* new_blk_ptr,
466 BasicBlock::iterator call_inst_itr, bool multiBlocks) {
467 // Copy remaining instructions from caller block.
468 for (Instruction* inst = call_inst_itr->NextNode(); inst;
469 inst = call_inst_itr->NextNode()) {
470 inst->RemoveFromList();
471 std::unique_ptr<Instruction> cp_inst(inst);
472 // If multiple blocks generated, regenerate any same-block
473 // instruction that has not been seen in this last block.
474 if (multiBlocks) {
475 if (!CloneSameBlockOps(&cp_inst, postCallSB, preCallSB, new_blk_ptr)) {
476 return false;
477 }
478
479 // Remember same-block ops in this block.
480 if (IsSameBlockOp(&*cp_inst)) {
481 const uint32_t rid = cp_inst->result_id();
482 (*postCallSB)[rid] = rid;
483 }
484 }
485 new_blk_ptr->get()->AddInstruction(std::move(cp_inst));
486 }
487
488 return true;
489 }
490
MoveLoopMergeInstToFirstBlock(std::vector<std::unique_ptr<BasicBlock>> * new_blocks)491 void InlinePass::MoveLoopMergeInstToFirstBlock(
492 std::vector<std::unique_ptr<BasicBlock>>* new_blocks) {
493 // Move the OpLoopMerge from the last block back to the first, where
494 // it belongs.
495 auto& first = new_blocks->front();
496 auto& last = new_blocks->back();
497 assert(first != last);
498
499 // Insert a modified copy of the loop merge into the first block.
500 auto loop_merge_itr = last->tail();
501 --loop_merge_itr;
502 assert(loop_merge_itr->opcode() == SpvOpLoopMerge);
503 std::unique_ptr<Instruction> cp_inst(loop_merge_itr->Clone(context()));
504 first->tail().InsertBefore(std::move(cp_inst));
505
506 // Remove the loop merge from the last block.
507 loop_merge_itr->RemoveFromList();
508 delete &*loop_merge_itr;
509 }
510
UpdateSingleBlockLoopContinueTarget(uint32_t new_id,std::vector<std::unique_ptr<BasicBlock>> * new_blocks)511 void InlinePass::UpdateSingleBlockLoopContinueTarget(
512 uint32_t new_id, std::vector<std::unique_ptr<BasicBlock>>* new_blocks) {
513 auto& header = new_blocks->front();
514 auto* merge_inst = header->GetLoopMergeInst();
515
516 // The back-edge block is split at the branch to create a new back-edge
517 // block. The old block is modified to branch to the new block. The loop
518 // merge instruction is updated to declare the new block as the continue
519 // target. This has the effect of changing the loop from being a large
520 // continue construct and an empty loop construct to being a loop with a loop
521 // construct and a trivial continue construct. This change is made to satisfy
522 // structural dominance.
523
524 // Add the new basic block.
525 std::unique_ptr<BasicBlock> new_block =
526 MakeUnique<BasicBlock>(NewLabel(new_id));
527 auto& old_backedge = new_blocks->back();
528 auto old_branch = old_backedge->tail();
529
530 // Move the old back edge into the new block.
531 std::unique_ptr<Instruction> br(&*old_branch);
532 new_block->AddInstruction(std::move(br));
533
534 // Add a branch to the new block from the old back-edge block.
535 AddBranch(new_id, &old_backedge);
536 new_blocks->push_back(std::move(new_block));
537
538 // Update the loop's continue target to the new block.
539 merge_inst->SetInOperand(1u, {new_id});
540 }
541
GenInlineCode(std::vector<std::unique_ptr<BasicBlock>> * new_blocks,std::vector<std::unique_ptr<Instruction>> * new_vars,BasicBlock::iterator call_inst_itr,UptrVectorIterator<BasicBlock> call_block_itr)542 bool InlinePass::GenInlineCode(
543 std::vector<std::unique_ptr<BasicBlock>>* new_blocks,
544 std::vector<std::unique_ptr<Instruction>>* new_vars,
545 BasicBlock::iterator call_inst_itr,
546 UptrVectorIterator<BasicBlock> call_block_itr) {
547 // Map from all ids in the callee to their equivalent id in the caller
548 // as callee instructions are copied into caller.
549 std::unordered_map<uint32_t, uint32_t> callee2caller;
550 // Pre-call same-block insts
551 std::unordered_map<uint32_t, Instruction*> preCallSB;
552 // Post-call same-block op ids
553 std::unordered_map<uint32_t, uint32_t> postCallSB;
554
555 analysis::DebugInlinedAtContext inlined_at_ctx(&*call_inst_itr);
556
557 // Invalidate the def-use chains. They are not kept up to date while
558 // inlining. However, certain calls try to keep them up-to-date if they are
559 // valid. These operations can fail.
560 context()->InvalidateAnalyses(IRContext::kAnalysisDefUse);
561
562 // If the caller is a loop header and the callee has multiple blocks, then the
563 // normal inlining logic will place the OpLoopMerge in the last of several
564 // blocks in the loop. Instead, it should be placed at the end of the first
565 // block. We'll wait to move the OpLoopMerge until the end of the regular
566 // inlining logic, and only if necessary.
567 bool caller_is_loop_header = call_block_itr->GetLoopMergeInst() != nullptr;
568
569 // Single-trip loop continue block
570 std::unique_ptr<BasicBlock> single_trip_loop_cont_blk;
571
572 Function* calleeFn = id2function_[call_inst_itr->GetSingleWordOperand(
573 kSpvFunctionCallFunctionId)];
574
575 // Map parameters to actual arguments.
576 MapParams(calleeFn, call_inst_itr, &callee2caller);
577
578 // Define caller local variables for all callee variables and create map to
579 // them.
580 if (!CloneAndMapLocals(calleeFn, new_vars, &callee2caller, &inlined_at_ctx)) {
581 return false;
582 }
583
584 // First block needs to use label of original block
585 // but map callee label in case of phi reference.
586 uint32_t entry_blk_label_id = calleeFn->begin()->GetLabelInst()->result_id();
587 callee2caller[entry_blk_label_id] = call_block_itr->id();
588 std::unique_ptr<BasicBlock> new_blk_ptr =
589 MakeUnique<BasicBlock>(NewLabel(call_block_itr->id()));
590
591 // Move instructions of original caller block up to call instruction.
592 MoveInstsBeforeEntryBlock(&preCallSB, new_blk_ptr.get(), call_inst_itr,
593 call_block_itr);
594
595 if (caller_is_loop_header &&
596 (*(calleeFn->begin())).GetMergeInst() != nullptr) {
597 // We can't place both the caller's merge instruction and
598 // another merge instruction in the same block. So split the
599 // calling block. Insert an unconditional branch to a new guard
600 // block. Later, once we know the ID of the last block, we
601 // will move the caller's OpLoopMerge from the last generated
602 // block into the first block. We also wait to avoid
603 // invalidating various iterators.
604 new_blk_ptr = AddGuardBlock(new_blocks, &callee2caller,
605 std::move(new_blk_ptr), entry_blk_label_id);
606 if (new_blk_ptr == nullptr) return false;
607 }
608
609 // Create return var if needed.
610 const uint32_t calleeTypeId = calleeFn->type_id();
611 uint32_t returnVarId = 0;
612 analysis::Type* calleeType = context()->get_type_mgr()->GetType(calleeTypeId);
613 if (calleeType->AsVoid() == nullptr) {
614 returnVarId = CreateReturnVar(calleeFn, new_vars);
615 if (returnVarId == 0) {
616 return false;
617 }
618 }
619
620 calleeFn->WhileEachInst([&callee2caller, this](const Instruction* cpi) {
621 // Create set of callee result ids. Used to detect forward references
622 const uint32_t rid = cpi->result_id();
623 if (rid != 0 && callee2caller.find(rid) == callee2caller.end()) {
624 const uint32_t nid = context()->TakeNextId();
625 if (nid == 0) return false;
626 callee2caller[rid] = nid;
627 }
628 return true;
629 });
630
631 // Inline DebugClare instructions in the callee's header.
632 calleeFn->ForEachDebugInstructionsInHeader(
633 [&new_blk_ptr, &callee2caller, &inlined_at_ctx, this](Instruction* inst) {
634 InlineSingleInstruction(
635 callee2caller, new_blk_ptr.get(), inst,
636 context()->get_debug_info_mgr()->BuildDebugInlinedAtChain(
637 inst->GetDebugScope().GetInlinedAt(), &inlined_at_ctx));
638 });
639
640 // Inline the entry block of the callee function.
641 if (!InlineEntryBlock(callee2caller, &new_blk_ptr, calleeFn->begin(),
642 &inlined_at_ctx)) {
643 return false;
644 }
645
646 // Inline blocks of the callee function other than the entry block.
647 new_blk_ptr =
648 InlineBasicBlocks(new_blocks, callee2caller, std::move(new_blk_ptr),
649 &inlined_at_ctx, calleeFn);
650 if (new_blk_ptr == nullptr) return false;
651
652 new_blk_ptr = InlineReturn(callee2caller, new_blocks, std::move(new_blk_ptr),
653 &inlined_at_ctx, calleeFn,
654 &*(calleeFn->tail()->tail()), returnVarId);
655
656 // Load return value into result id of call, if it exists.
657 if (returnVarId != 0) {
658 const uint32_t resId = call_inst_itr->result_id();
659 assert(resId != 0);
660 AddLoad(calleeTypeId, resId, returnVarId, &new_blk_ptr,
661 call_inst_itr->dbg_line_inst(), call_inst_itr->GetDebugScope());
662 }
663
664 // Move instructions of original caller block after call instruction.
665 if (!MoveCallerInstsAfterFunctionCall(&preCallSB, &postCallSB, &new_blk_ptr,
666 call_inst_itr,
667 calleeFn->begin() != calleeFn->end()))
668 return false;
669
670 // Finalize inline code.
671 new_blocks->push_back(std::move(new_blk_ptr));
672
673 if (caller_is_loop_header && (new_blocks->size() > 1)) {
674 MoveLoopMergeInstToFirstBlock(new_blocks);
675
676 // If the loop was a single basic block previously, update it's structure.
677 auto& header = new_blocks->front();
678 auto* merge_inst = header->GetLoopMergeInst();
679 if (merge_inst->GetSingleWordInOperand(1u) == header->id()) {
680 auto new_id = context()->TakeNextId();
681 if (new_id == 0) return false;
682 UpdateSingleBlockLoopContinueTarget(new_id, new_blocks);
683 }
684 }
685
686 // Update block map given replacement blocks.
687 for (auto& blk : *new_blocks) {
688 id2block_[blk->id()] = &*blk;
689 }
690
691 // We need to kill the name and decorations for the call, which will be
692 // deleted.
693 context()->KillNamesAndDecorates(&*call_inst_itr);
694
695 return true;
696 }
697
IsInlinableFunctionCall(const Instruction * inst)698 bool InlinePass::IsInlinableFunctionCall(const Instruction* inst) {
699 if (inst->opcode() != SpvOp::SpvOpFunctionCall) return false;
700 const uint32_t calleeFnId =
701 inst->GetSingleWordOperand(kSpvFunctionCallFunctionId);
702 const auto ci = inlinable_.find(calleeFnId);
703 if (ci == inlinable_.cend()) return false;
704
705 if (early_return_funcs_.find(calleeFnId) != early_return_funcs_.end()) {
706 // We rely on the merge-return pass to handle the early return case
707 // in advance.
708 std::string message =
709 "The function '" + id2function_[calleeFnId]->DefInst().PrettyPrint() +
710 "' could not be inlined because the return instruction "
711 "is not at the end of the function. This could be fixed by "
712 "running merge-return before inlining.";
713 consumer()(SPV_MSG_WARNING, "", {0, 0, 0}, message.c_str());
714 return false;
715 }
716
717 return true;
718 }
719
UpdateSucceedingPhis(std::vector<std::unique_ptr<BasicBlock>> & new_blocks)720 void InlinePass::UpdateSucceedingPhis(
721 std::vector<std::unique_ptr<BasicBlock>>& new_blocks) {
722 const auto firstBlk = new_blocks.begin();
723 const auto lastBlk = new_blocks.end() - 1;
724 const uint32_t firstId = (*firstBlk)->id();
725 const uint32_t lastId = (*lastBlk)->id();
726 const BasicBlock& const_last_block = *lastBlk->get();
727 const_last_block.ForEachSuccessorLabel(
728 [&firstId, &lastId, this](const uint32_t succ) {
729 BasicBlock* sbp = this->id2block_[succ];
730 sbp->ForEachPhiInst([&firstId, &lastId](Instruction* phi) {
731 phi->ForEachInId([&firstId, &lastId](uint32_t* id) {
732 if (*id == firstId) *id = lastId;
733 });
734 });
735 });
736 }
737
HasNoReturnInLoop(Function * func)738 bool InlinePass::HasNoReturnInLoop(Function* func) {
739 // If control not structured, do not do loop/return analysis
740 // TODO: Analyze returns in non-structured control flow
741 if (!context()->get_feature_mgr()->HasCapability(SpvCapabilityShader))
742 return false;
743 const auto structured_analysis = context()->GetStructuredCFGAnalysis();
744 // Search for returns in structured construct.
745 bool return_in_loop = false;
746 for (auto& blk : *func) {
747 auto terminal_ii = blk.cend();
748 --terminal_ii;
749 if (spvOpcodeIsReturn(terminal_ii->opcode()) &&
750 structured_analysis->ContainingLoop(blk.id()) != 0) {
751 return_in_loop = true;
752 break;
753 }
754 }
755 return !return_in_loop;
756 }
757
AnalyzeReturns(Function * func)758 void InlinePass::AnalyzeReturns(Function* func) {
759 // Analyze functions without a return in loop.
760 if (HasNoReturnInLoop(func)) {
761 no_return_in_loop_.insert(func->result_id());
762 }
763 // Analyze functions with a return before its tail basic block.
764 for (auto& blk : *func) {
765 auto terminal_ii = blk.cend();
766 --terminal_ii;
767 if (spvOpcodeIsReturn(terminal_ii->opcode()) && &blk != func->tail()) {
768 early_return_funcs_.insert(func->result_id());
769 break;
770 }
771 }
772 }
773
IsInlinableFunction(Function * func)774 bool InlinePass::IsInlinableFunction(Function* func) {
775 // We can only inline a function if it has blocks.
776 if (func->cbegin() == func->cend()) return false;
777
778 // Do not inline functions with DontInline flag.
779 if (func->control_mask() & SpvFunctionControlDontInlineMask) {
780 return false;
781 }
782
783 // Do not inline functions with returns in loops. Currently early return
784 // functions are inlined by wrapping them in a one trip loop and implementing
785 // the returns as a branch to the loop's merge block. However, this can only
786 // done validly if the return was not in a loop in the original function.
787 // Also remember functions with multiple (early) returns.
788 AnalyzeReturns(func);
789 if (no_return_in_loop_.find(func->result_id()) == no_return_in_loop_.cend()) {
790 return false;
791 }
792
793 if (func->IsRecursive()) {
794 return false;
795 }
796
797 // Do not inline functions with an abort instruction if they are called from a
798 // continue construct. If it is inlined into a continue construct the backedge
799 // will no longer post-dominate the continue target, which is invalid. An
800 // `OpUnreachable` is acceptable because it will not change post-dominance if
801 // it is statically unreachable.
802 bool func_is_called_from_continue =
803 funcs_called_from_continue_.count(func->result_id()) != 0;
804
805 if (func_is_called_from_continue && ContainsAbortOtherThanUnreachable(func)) {
806 return false;
807 }
808
809 return true;
810 }
811
ContainsAbortOtherThanUnreachable(Function * func) const812 bool InlinePass::ContainsAbortOtherThanUnreachable(Function* func) const {
813 return !func->WhileEachInst([](Instruction* inst) {
814 return inst->opcode() == SpvOpUnreachable ||
815 !spvOpcodeIsAbort(inst->opcode());
816 });
817 }
818
InitializeInline()819 void InlinePass::InitializeInline() {
820 false_id_ = 0;
821
822 // clear collections
823 id2function_.clear();
824 id2block_.clear();
825 inlinable_.clear();
826 no_return_in_loop_.clear();
827 early_return_funcs_.clear();
828 funcs_called_from_continue_ =
829 context()->GetStructuredCFGAnalysis()->FindFuncsCalledFromContinue();
830
831 for (auto& fn : *get_module()) {
832 // Initialize function and block maps.
833 id2function_[fn.result_id()] = &fn;
834 for (auto& blk : fn) {
835 id2block_[blk.id()] = &blk;
836 }
837 // Compute inlinability
838 if (IsInlinableFunction(&fn)) inlinable_.insert(fn.result_id());
839 }
840 }
841
InlinePass()842 InlinePass::InlinePass() {}
843
844 } // namespace opt
845 } // namespace spvtools
846