Misha Brukman | e633603 | 2004-02-28 03:26:20 +0000 | [diff] [blame] | 1 | //===- CodeExtractor.cpp - Pull code region into a new function -----------===// |
Misha Brukman | fd93908 | 2005-04-21 23:48:37 +0000 | [diff] [blame] | 2 | // |
Misha Brukman | e633603 | 2004-02-28 03:26:20 +0000 | [diff] [blame] | 3 | // The LLVM Compiler Infrastructure |
| 4 | // |
Chris Lattner | 4ee451d | 2007-12-29 20:36:04 +0000 | [diff] [blame] | 5 | // This file is distributed under the University of Illinois Open Source |
| 6 | // License. See LICENSE.TXT for details. |
Misha Brukman | fd93908 | 2005-04-21 23:48:37 +0000 | [diff] [blame] | 7 | // |
Misha Brukman | e633603 | 2004-02-28 03:26:20 +0000 | [diff] [blame] | 8 | //===----------------------------------------------------------------------===// |
| 9 | // |
| 10 | // This file implements the interface to tear out a code region, such as an |
| 11 | // individual loop or a parallel section, into a new function, replacing it with |
| 12 | // a call to the new function. |
| 13 | // |
| 14 | //===----------------------------------------------------------------------===// |
| 15 | |
Chris Lattner | 33e197b | 2004-03-18 03:15:29 +0000 | [diff] [blame] | 16 | #include "llvm/Transforms/Utils/FunctionUtils.h" |
Misha Brukman | e633603 | 2004-02-28 03:26:20 +0000 | [diff] [blame] | 17 | #include "llvm/Constants.h" |
| 18 | #include "llvm/DerivedTypes.h" |
| 19 | #include "llvm/Instructions.h" |
Chris Lattner | bf74936 | 2004-05-12 06:01:40 +0000 | [diff] [blame] | 20 | #include "llvm/Intrinsics.h" |
Misha Brukman | e633603 | 2004-02-28 03:26:20 +0000 | [diff] [blame] | 21 | #include "llvm/Module.h" |
| 22 | #include "llvm/Pass.h" |
Chris Lattner | b5c1dbf | 2004-03-18 03:49:40 +0000 | [diff] [blame] | 23 | #include "llvm/Analysis/Dominators.h" |
Misha Brukman | e633603 | 2004-02-28 03:26:20 +0000 | [diff] [blame] | 24 | #include "llvm/Analysis/LoopInfo.h" |
Chris Lattner | ffada93 | 2004-03-14 03:17:22 +0000 | [diff] [blame] | 25 | #include "llvm/Analysis/Verifier.h" |
Misha Brukman | e633603 | 2004-02-28 03:26:20 +0000 | [diff] [blame] | 26 | #include "llvm/Transforms/Utils/BasicBlockUtils.h" |
Reid Spencer | 551ccae | 2004-09-01 22:55:40 +0000 | [diff] [blame] | 27 | #include "llvm/Support/CommandLine.h" |
Reid Spencer | 9133fe2 | 2007-02-05 23:32:05 +0000 | [diff] [blame] | 28 | #include "llvm/Support/Compiler.h" |
Reid Spencer | 551ccae | 2004-09-01 22:55:40 +0000 | [diff] [blame] | 29 | #include "llvm/Support/Debug.h" |
| 30 | #include "llvm/ADT/StringExtras.h" |
Misha Brukman | e633603 | 2004-02-28 03:26:20 +0000 | [diff] [blame] | 31 | #include <algorithm> |
Chris Lattner | 0e06674 | 2004-03-14 22:34:55 +0000 | [diff] [blame] | 32 | #include <set> |
Misha Brukman | e633603 | 2004-02-28 03:26:20 +0000 | [diff] [blame] | 33 | using namespace llvm; |
| 34 | |
Misha Brukman | 22108fa | 2004-04-23 23:54:17 +0000 | [diff] [blame] | 35 | // Provide a command-line option to aggregate function arguments into a struct |
| 36 | // for functions produced by the code extrator. This is useful when converting |
| 37 | // extracted functions to pthread-based code, as only one argument (void*) can |
| 38 | // be passed in to pthread_create(). |
| 39 | static cl::opt<bool> |
| 40 | AggregateArgsOpt("aggregate-extracted-args", cl::Hidden, |
| 41 | cl::desc("Aggregate arguments to code-extracted functions")); |
| 42 | |
Misha Brukman | e633603 | 2004-02-28 03:26:20 +0000 | [diff] [blame] | 43 | namespace { |
Reid Spencer | 9133fe2 | 2007-02-05 23:32:05 +0000 | [diff] [blame] | 44 | class VISIBILITY_HIDDEN CodeExtractor { |
Misha Brukman | e633603 | 2004-02-28 03:26:20 +0000 | [diff] [blame] | 45 | typedef std::vector<Value*> Values; |
Chris Lattner | 0e06674 | 2004-03-14 22:34:55 +0000 | [diff] [blame] | 46 | std::set<BasicBlock*> BlocksToExtract; |
Owen Anderson | c6fcf29 | 2007-04-07 05:31:27 +0000 | [diff] [blame] | 47 | DominatorTree* DT; |
Misha Brukman | 22108fa | 2004-04-23 23:54:17 +0000 | [diff] [blame] | 48 | bool AggregateArgs; |
Chris Lattner | 346be7f | 2004-05-12 04:14:24 +0000 | [diff] [blame] | 49 | unsigned NumExitBlocks; |
| 50 | const Type *RetTy; |
Misha Brukman | e633603 | 2004-02-28 03:26:20 +0000 | [diff] [blame] | 51 | public: |
Devang Patel | 4b90e3a | 2007-06-07 22:17:16 +0000 | [diff] [blame] | 52 | CodeExtractor(DominatorTree* dt = 0, bool AggArgs = false) |
| 53 | : DT(dt), AggregateArgs(AggArgs||AggregateArgsOpt), NumExitBlocks(~0U) {} |
Chris Lattner | b5c1dbf | 2004-03-18 03:49:40 +0000 | [diff] [blame] | 54 | |
Misha Brukman | e633603 | 2004-02-28 03:26:20 +0000 | [diff] [blame] | 55 | Function *ExtractCodeRegion(const std::vector<BasicBlock*> &code); |
| 56 | |
Misha Brukman | 22108fa | 2004-04-23 23:54:17 +0000 | [diff] [blame] | 57 | bool isEligible(const std::vector<BasicBlock*> &code); |
| 58 | |
Misha Brukman | e633603 | 2004-02-28 03:26:20 +0000 | [diff] [blame] | 59 | private: |
Chris Lattner | bf74936 | 2004-05-12 06:01:40 +0000 | [diff] [blame] | 60 | /// definedInRegion - Return true if the specified value is defined in the |
| 61 | /// extracted region. |
| 62 | bool definedInRegion(Value *V) const { |
| 63 | if (Instruction *I = dyn_cast<Instruction>(V)) |
| 64 | if (BlocksToExtract.count(I->getParent())) |
| 65 | return true; |
| 66 | return false; |
| 67 | } |
Misha Brukman | fd93908 | 2005-04-21 23:48:37 +0000 | [diff] [blame] | 68 | |
Chris Lattner | bf74936 | 2004-05-12 06:01:40 +0000 | [diff] [blame] | 69 | /// definedInCaller - Return true if the specified value is defined in the |
| 70 | /// function being code extracted, but not in the region being extracted. |
| 71 | /// These values must be passed in as live-ins to the function. |
| 72 | bool definedInCaller(Value *V) const { |
| 73 | if (isa<Argument>(V)) return true; |
| 74 | if (Instruction *I = dyn_cast<Instruction>(V)) |
| 75 | if (!BlocksToExtract.count(I->getParent())) |
| 76 | return true; |
| 77 | return false; |
| 78 | } |
| 79 | |
| 80 | void severSplitPHINodes(BasicBlock *&Header); |
Chris Lattner | d99e1d3 | 2004-05-12 16:07:41 +0000 | [diff] [blame] | 81 | void splitReturnBlocks(); |
Chris Lattner | bf74936 | 2004-05-12 06:01:40 +0000 | [diff] [blame] | 82 | void findInputsOutputs(Values &inputs, Values &outputs); |
Misha Brukman | e633603 | 2004-02-28 03:26:20 +0000 | [diff] [blame] | 83 | |
Misha Brukman | e633603 | 2004-02-28 03:26:20 +0000 | [diff] [blame] | 84 | Function *constructFunction(const Values &inputs, |
| 85 | const Values &outputs, |
Chris Lattner | 5cc8ca6 | 2004-03-18 05:28:49 +0000 | [diff] [blame] | 86 | BasicBlock *header, |
Misha Brukman | e633603 | 2004-02-28 03:26:20 +0000 | [diff] [blame] | 87 | BasicBlock *newRootNode, BasicBlock *newHeader, |
Misha Brukman | e633603 | 2004-02-28 03:26:20 +0000 | [diff] [blame] | 88 | Function *oldFunction, Module *M); |
| 89 | |
Chris Lattner | 0e06674 | 2004-03-14 22:34:55 +0000 | [diff] [blame] | 90 | void moveCodeToFunction(Function *newFunction); |
Misha Brukman | e633603 | 2004-02-28 03:26:20 +0000 | [diff] [blame] | 91 | |
| 92 | void emitCallAndSwitchStatement(Function *newFunction, |
| 93 | BasicBlock *newHeader, |
Misha Brukman | e633603 | 2004-02-28 03:26:20 +0000 | [diff] [blame] | 94 | Values &inputs, |
| 95 | Values &outputs); |
| 96 | |
| 97 | }; |
| 98 | } |
| 99 | |
Chris Lattner | bf74936 | 2004-05-12 06:01:40 +0000 | [diff] [blame] | 100 | /// severSplitPHINodes - If a PHI node has multiple inputs from outside of the |
| 101 | /// region, we need to split the entry block of the region so that the PHI node |
| 102 | /// is easier to deal with. |
| 103 | void CodeExtractor::severSplitPHINodes(BasicBlock *&Header) { |
Chris Lattner | e746ad5 | 2004-05-12 15:29:13 +0000 | [diff] [blame] | 104 | bool HasPredsFromRegion = false; |
| 105 | unsigned NumPredsOutsideRegion = 0; |
Chris Lattner | bf74936 | 2004-05-12 06:01:40 +0000 | [diff] [blame] | 106 | |
Dan Gohman | ecb7a77 | 2007-03-22 16:38:57 +0000 | [diff] [blame] | 107 | if (Header != &Header->getParent()->getEntryBlock()) { |
Chris Lattner | e746ad5 | 2004-05-12 15:29:13 +0000 | [diff] [blame] | 108 | PHINode *PN = dyn_cast<PHINode>(Header->begin()); |
| 109 | if (!PN) return; // No PHI nodes. |
Chris Lattner | bf74936 | 2004-05-12 06:01:40 +0000 | [diff] [blame] | 110 | |
Chris Lattner | e746ad5 | 2004-05-12 15:29:13 +0000 | [diff] [blame] | 111 | // If the header node contains any PHI nodes, check to see if there is more |
| 112 | // than one entry from outside the region. If so, we need to sever the |
| 113 | // header block into two. |
| 114 | for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) |
| 115 | if (BlocksToExtract.count(PN->getIncomingBlock(i))) |
| 116 | HasPredsFromRegion = true; |
| 117 | else |
| 118 | ++NumPredsOutsideRegion; |
| 119 | |
| 120 | // If there is one (or fewer) predecessor from outside the region, we don't |
| 121 | // need to do anything special. |
| 122 | if (NumPredsOutsideRegion <= 1) return; |
| 123 | } |
| 124 | |
| 125 | // Otherwise, we need to split the header block into two pieces: one |
| 126 | // containing PHI nodes merging values from outside of the region, and a |
| 127 | // second that contains all of the code for the block and merges back any |
| 128 | // incoming values from inside of the region. |
| 129 | BasicBlock::iterator AfterPHIs = Header->begin(); |
| 130 | while (isa<PHINode>(AfterPHIs)) ++AfterPHIs; |
| 131 | BasicBlock *NewBB = Header->splitBasicBlock(AfterPHIs, |
| 132 | Header->getName()+".ce"); |
| 133 | |
| 134 | // We only want to code extract the second block now, and it becomes the new |
| 135 | // header of the region. |
| 136 | BasicBlock *OldPred = Header; |
| 137 | BlocksToExtract.erase(OldPred); |
| 138 | BlocksToExtract.insert(NewBB); |
| 139 | Header = NewBB; |
| 140 | |
| 141 | // Okay, update dominator sets. The blocks that dominate the new one are the |
| 142 | // blocks that dominate TIBB plus the new block itself. |
Devang Patel | 0e7f728 | 2007-06-21 17:23:45 +0000 | [diff] [blame] | 143 | if (DT) |
| 144 | DT->splitBlock(NewBB); |
Chris Lattner | e746ad5 | 2004-05-12 15:29:13 +0000 | [diff] [blame] | 145 | |
| 146 | // Okay, now we need to adjust the PHI nodes and any branches from within the |
| 147 | // region to go to the new header block instead of the old header block. |
| 148 | if (HasPredsFromRegion) { |
| 149 | PHINode *PN = cast<PHINode>(OldPred->begin()); |
| 150 | // Loop over all of the predecessors of OldPred that are in the region, |
| 151 | // changing them to branch to NewBB instead. |
| 152 | for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) |
| 153 | if (BlocksToExtract.count(PN->getIncomingBlock(i))) { |
| 154 | TerminatorInst *TI = PN->getIncomingBlock(i)->getTerminator(); |
| 155 | TI->replaceUsesOfWith(OldPred, NewBB); |
| 156 | } |
| 157 | |
| 158 | // Okay, everthing within the region is now branching to the right block, we |
| 159 | // just have to update the PHI nodes now, inserting PHI nodes into NewBB. |
Reid Spencer | 2da5c3d | 2004-09-15 17:06:42 +0000 | [diff] [blame] | 160 | for (AfterPHIs = OldPred->begin(); isa<PHINode>(AfterPHIs); ++AfterPHIs) { |
| 161 | PHINode *PN = cast<PHINode>(AfterPHIs); |
Chris Lattner | e746ad5 | 2004-05-12 15:29:13 +0000 | [diff] [blame] | 162 | // Create a new PHI node in the new region, which has an incoming value |
| 163 | // from OldPred of PN. |
Gabor Greif | 051a950 | 2008-04-06 20:25:17 +0000 | [diff] [blame] | 164 | PHINode *NewPN = PHINode::Create(PN->getType(), PN->getName()+".ce", |
| 165 | NewBB->begin()); |
Chris Lattner | e746ad5 | 2004-05-12 15:29:13 +0000 | [diff] [blame] | 166 | NewPN->addIncoming(PN, OldPred); |
| 167 | |
| 168 | // Loop over all of the incoming value in PN, moving them to NewPN if they |
| 169 | // are from the extracted region. |
| 170 | for (unsigned i = 0; i != PN->getNumIncomingValues(); ++i) { |
| 171 | if (BlocksToExtract.count(PN->getIncomingBlock(i))) { |
| 172 | NewPN->addIncoming(PN->getIncomingValue(i), PN->getIncomingBlock(i)); |
| 173 | PN->removeIncomingValue(i); |
| 174 | --i; |
| 175 | } |
| 176 | } |
| 177 | } |
| 178 | } |
Chris Lattner | d99e1d3 | 2004-05-12 16:07:41 +0000 | [diff] [blame] | 179 | } |
Chris Lattner | e746ad5 | 2004-05-12 15:29:13 +0000 | [diff] [blame] | 180 | |
Chris Lattner | d99e1d3 | 2004-05-12 16:07:41 +0000 | [diff] [blame] | 181 | void CodeExtractor::splitReturnBlocks() { |
| 182 | for (std::set<BasicBlock*>::iterator I = BlocksToExtract.begin(), |
| 183 | E = BlocksToExtract.end(); I != E; ++I) |
| 184 | if (ReturnInst *RI = dyn_cast<ReturnInst>((*I)->getTerminator())) |
| 185 | (*I)->splitBasicBlock(RI, (*I)->getName()+".ret"); |
Chris Lattner | bf74936 | 2004-05-12 06:01:40 +0000 | [diff] [blame] | 186 | } |
| 187 | |
| 188 | // findInputsOutputs - Find inputs to, outputs from the code region. |
| 189 | // |
| 190 | void CodeExtractor::findInputsOutputs(Values &inputs, Values &outputs) { |
Chris Lattner | 346be7f | 2004-05-12 04:14:24 +0000 | [diff] [blame] | 191 | std::set<BasicBlock*> ExitBlocks; |
Misha Brukman | fd93908 | 2005-04-21 23:48:37 +0000 | [diff] [blame] | 192 | for (std::set<BasicBlock*>::const_iterator ci = BlocksToExtract.begin(), |
Chris Lattner | 0e06674 | 2004-03-14 22:34:55 +0000 | [diff] [blame] | 193 | ce = BlocksToExtract.end(); ci != ce; ++ci) { |
Misha Brukman | e633603 | 2004-02-28 03:26:20 +0000 | [diff] [blame] | 194 | BasicBlock *BB = *ci; |
Chris Lattner | bf74936 | 2004-05-12 06:01:40 +0000 | [diff] [blame] | 195 | |
Chris Lattner | 0de632b | 2004-03-15 01:18:23 +0000 | [diff] [blame] | 196 | for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) { |
| 197 | // If a used value is defined outside the region, it's an input. If an |
| 198 | // instruction is used outside the region, it's an output. |
Chris Lattner | bf74936 | 2004-05-12 06:01:40 +0000 | [diff] [blame] | 199 | for (User::op_iterator O = I->op_begin(), E = I->op_end(); O != E; ++O) |
| 200 | if (definedInCaller(*O)) |
| 201 | inputs.push_back(*O); |
Misha Brukman | fd93908 | 2005-04-21 23:48:37 +0000 | [diff] [blame] | 202 | |
Chris Lattner | bf74936 | 2004-05-12 06:01:40 +0000 | [diff] [blame] | 203 | // Consider uses of this instruction (outputs). |
Chris Lattner | 0de632b | 2004-03-15 01:18:23 +0000 | [diff] [blame] | 204 | for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); |
| 205 | UI != E; ++UI) |
Chris Lattner | bf74936 | 2004-05-12 06:01:40 +0000 | [diff] [blame] | 206 | if (!definedInRegion(*UI)) { |
Chris Lattner | b5c1dbf | 2004-03-18 03:49:40 +0000 | [diff] [blame] | 207 | outputs.push_back(I); |
| 208 | break; |
| 209 | } |
Chris Lattner | 0de632b | 2004-03-15 01:18:23 +0000 | [diff] [blame] | 210 | } // for: insts |
Chris Lattner | 346be7f | 2004-05-12 04:14:24 +0000 | [diff] [blame] | 211 | |
Chris Lattner | bf74936 | 2004-05-12 06:01:40 +0000 | [diff] [blame] | 212 | // Keep track of the exit blocks from the region. |
Chris Lattner | 346be7f | 2004-05-12 04:14:24 +0000 | [diff] [blame] | 213 | TerminatorInst *TI = BB->getTerminator(); |
| 214 | for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i) |
| 215 | if (!BlocksToExtract.count(TI->getSuccessor(i))) |
| 216 | ExitBlocks.insert(TI->getSuccessor(i)); |
Chris Lattner | 0de632b | 2004-03-15 01:18:23 +0000 | [diff] [blame] | 217 | } // for: basic blocks |
Chris Lattner | 346be7f | 2004-05-12 04:14:24 +0000 | [diff] [blame] | 218 | |
| 219 | NumExitBlocks = ExitBlocks.size(); |
Chris Lattner | 5879927 | 2004-05-12 16:26:18 +0000 | [diff] [blame] | 220 | |
| 221 | // Eliminate duplicates. |
| 222 | std::sort(inputs.begin(), inputs.end()); |
| 223 | inputs.erase(std::unique(inputs.begin(), inputs.end()), inputs.end()); |
| 224 | std::sort(outputs.begin(), outputs.end()); |
| 225 | outputs.erase(std::unique(outputs.begin(), outputs.end()), outputs.end()); |
Misha Brukman | e633603 | 2004-02-28 03:26:20 +0000 | [diff] [blame] | 226 | } |
| 227 | |
Misha Brukman | e633603 | 2004-02-28 03:26:20 +0000 | [diff] [blame] | 228 | /// constructFunction - make a function based on inputs and outputs, as follows: |
| 229 | /// f(in0, ..., inN, out0, ..., outN) |
| 230 | /// |
| 231 | Function *CodeExtractor::constructFunction(const Values &inputs, |
| 232 | const Values &outputs, |
Chris Lattner | 5cc8ca6 | 2004-03-18 05:28:49 +0000 | [diff] [blame] | 233 | BasicBlock *header, |
Misha Brukman | e633603 | 2004-02-28 03:26:20 +0000 | [diff] [blame] | 234 | BasicBlock *newRootNode, |
| 235 | BasicBlock *newHeader, |
Chris Lattner | 5cc8ca6 | 2004-03-18 05:28:49 +0000 | [diff] [blame] | 236 | Function *oldFunction, |
| 237 | Module *M) { |
Bill Wendling | 0d45a09 | 2006-11-26 10:17:54 +0000 | [diff] [blame] | 238 | DOUT << "inputs: " << inputs.size() << "\n"; |
| 239 | DOUT << "outputs: " << outputs.size() << "\n"; |
Misha Brukman | e633603 | 2004-02-28 03:26:20 +0000 | [diff] [blame] | 240 | |
| 241 | // This function returns unsigned, outputs will go back by reference. |
Chris Lattner | 346be7f | 2004-05-12 04:14:24 +0000 | [diff] [blame] | 242 | switch (NumExitBlocks) { |
| 243 | case 0: |
| 244 | case 1: RetTy = Type::VoidTy; break; |
Reid Spencer | 4fe16d6 | 2007-01-11 18:21:29 +0000 | [diff] [blame] | 245 | case 2: RetTy = Type::Int1Ty; break; |
Reid Spencer | c5b206b | 2006-12-31 05:48:39 +0000 | [diff] [blame] | 246 | default: RetTy = Type::Int16Ty; break; |
Chris Lattner | 346be7f | 2004-05-12 04:14:24 +0000 | [diff] [blame] | 247 | } |
| 248 | |
Misha Brukman | e633603 | 2004-02-28 03:26:20 +0000 | [diff] [blame] | 249 | std::vector<const Type*> paramTy; |
| 250 | |
| 251 | // Add the types of the input values to the function's argument list |
| 252 | for (Values::const_iterator i = inputs.begin(), |
| 253 | e = inputs.end(); i != e; ++i) { |
| 254 | const Value *value = *i; |
Bill Wendling | 0d45a09 | 2006-11-26 10:17:54 +0000 | [diff] [blame] | 255 | DOUT << "value used in func: " << *value << "\n"; |
Misha Brukman | e633603 | 2004-02-28 03:26:20 +0000 | [diff] [blame] | 256 | paramTy.push_back(value->getType()); |
| 257 | } |
| 258 | |
Chris Lattner | b5c1dbf | 2004-03-18 03:49:40 +0000 | [diff] [blame] | 259 | // Add the types of the output values to the function's argument list. |
| 260 | for (Values::const_iterator I = outputs.begin(), E = outputs.end(); |
| 261 | I != E; ++I) { |
Bill Wendling | 0d45a09 | 2006-11-26 10:17:54 +0000 | [diff] [blame] | 262 | DOUT << "instr used in func: " << **I << "\n"; |
Misha Brukman | 22108fa | 2004-04-23 23:54:17 +0000 | [diff] [blame] | 263 | if (AggregateArgs) |
| 264 | paramTy.push_back((*I)->getType()); |
| 265 | else |
Christopher Lamb | 43ad6b3 | 2007-12-17 01:12:55 +0000 | [diff] [blame] | 266 | paramTy.push_back(PointerType::getUnqual((*I)->getType())); |
Misha Brukman | e633603 | 2004-02-28 03:26:20 +0000 | [diff] [blame] | 267 | } |
| 268 | |
Bill Wendling | 0d45a09 | 2006-11-26 10:17:54 +0000 | [diff] [blame] | 269 | DOUT << "Function type: " << *RetTy << " f("; |
| 270 | for (std::vector<const Type*>::iterator i = paramTy.begin(), |
| 271 | e = paramTy.end(); i != e; ++i) |
| 272 | DOUT << **i << ", "; |
| 273 | DOUT << ")\n"; |
Misha Brukman | e633603 | 2004-02-28 03:26:20 +0000 | [diff] [blame] | 274 | |
Misha Brukman | 22108fa | 2004-04-23 23:54:17 +0000 | [diff] [blame] | 275 | if (AggregateArgs && (inputs.size() + outputs.size() > 0)) { |
Christopher Lamb | 43ad6b3 | 2007-12-17 01:12:55 +0000 | [diff] [blame] | 276 | PointerType *StructPtr = PointerType::getUnqual(StructType::get(paramTy)); |
Misha Brukman | 22108fa | 2004-04-23 23:54:17 +0000 | [diff] [blame] | 277 | paramTy.clear(); |
| 278 | paramTy.push_back(StructPtr); |
| 279 | } |
Chris Lattner | 346be7f | 2004-05-12 04:14:24 +0000 | [diff] [blame] | 280 | const FunctionType *funcType = FunctionType::get(RetTy, paramTy, false); |
Misha Brukman | e633603 | 2004-02-28 03:26:20 +0000 | [diff] [blame] | 281 | |
| 282 | // Create the new function |
Gabor Greif | 051a950 | 2008-04-06 20:25:17 +0000 | [diff] [blame] | 283 | Function *newFunction = Function::Create(funcType, |
| 284 | GlobalValue::InternalLinkage, |
| 285 | oldFunction->getName() + "_" + |
| 286 | header->getName(), M); |
Misha Brukman | e633603 | 2004-02-28 03:26:20 +0000 | [diff] [blame] | 287 | newFunction->getBasicBlockList().push_back(newRootNode); |
| 288 | |
Chris Lattner | b5c1dbf | 2004-03-18 03:49:40 +0000 | [diff] [blame] | 289 | // Create an iterator to name all of the arguments we inserted. |
Chris Lattner | e4d5c44 | 2005-03-15 04:54:21 +0000 | [diff] [blame] | 290 | Function::arg_iterator AI = newFunction->arg_begin(); |
Chris Lattner | b5c1dbf | 2004-03-18 03:49:40 +0000 | [diff] [blame] | 291 | |
| 292 | // Rewrite all users of the inputs in the extracted region to use the |
Misha Brukman | 22108fa | 2004-04-23 23:54:17 +0000 | [diff] [blame] | 293 | // arguments (or appropriate addressing into struct) instead. |
| 294 | for (unsigned i = 0, e = inputs.size(); i != e; ++i) { |
| 295 | Value *RewriteVal; |
| 296 | if (AggregateArgs) { |
David Greene | b8f7479 | 2007-09-04 15:46:09 +0000 | [diff] [blame] | 297 | Value *Idx[2]; |
| 298 | Idx[0] = Constant::getNullValue(Type::Int32Ty); |
| 299 | Idx[1] = ConstantInt::get(Type::Int32Ty, i); |
Misha Brukman | 22108fa | 2004-04-23 23:54:17 +0000 | [diff] [blame] | 300 | std::string GEPname = "gep_" + inputs[i]->getName(); |
| 301 | TerminatorInst *TI = newFunction->begin()->getTerminator(); |
Gabor Greif | 051a950 | 2008-04-06 20:25:17 +0000 | [diff] [blame] | 302 | GetElementPtrInst *GEP = GetElementPtrInst::Create(AI, Idx, Idx+2, |
| 303 | GEPname, TI); |
Misha Brukman | 22108fa | 2004-04-23 23:54:17 +0000 | [diff] [blame] | 304 | RewriteVal = new LoadInst(GEP, "load" + GEPname, TI); |
| 305 | } else |
| 306 | RewriteVal = AI++; |
| 307 | |
Misha Brukman | e633603 | 2004-02-28 03:26:20 +0000 | [diff] [blame] | 308 | std::vector<User*> Users(inputs[i]->use_begin(), inputs[i]->use_end()); |
| 309 | for (std::vector<User*>::iterator use = Users.begin(), useE = Users.end(); |
Chris Lattner | ffada93 | 2004-03-14 03:17:22 +0000 | [diff] [blame] | 310 | use != useE; ++use) |
| 311 | if (Instruction* inst = dyn_cast<Instruction>(*use)) |
Chris Lattner | 0e06674 | 2004-03-14 22:34:55 +0000 | [diff] [blame] | 312 | if (BlocksToExtract.count(inst->getParent())) |
Misha Brukman | 22108fa | 2004-04-23 23:54:17 +0000 | [diff] [blame] | 313 | inst->replaceUsesOfWith(inputs[i], RewriteVal); |
Misha Brukman | e633603 | 2004-02-28 03:26:20 +0000 | [diff] [blame] | 314 | } |
| 315 | |
Misha Brukman | 22108fa | 2004-04-23 23:54:17 +0000 | [diff] [blame] | 316 | // Set names for input and output arguments. |
| 317 | if (!AggregateArgs) { |
Chris Lattner | e4d5c44 | 2005-03-15 04:54:21 +0000 | [diff] [blame] | 318 | AI = newFunction->arg_begin(); |
Misha Brukman | 22108fa | 2004-04-23 23:54:17 +0000 | [diff] [blame] | 319 | for (unsigned i = 0, e = inputs.size(); i != e; ++i, ++AI) |
Owen Anderson | 6bc41e8 | 2008-04-14 17:38:21 +0000 | [diff] [blame^] | 320 | AI->setName(inputs[i]->getName()); |
Misha Brukman | 22108fa | 2004-04-23 23:54:17 +0000 | [diff] [blame] | 321 | for (unsigned i = 0, e = outputs.size(); i != e; ++i, ++AI) |
Misha Brukman | fd93908 | 2005-04-21 23:48:37 +0000 | [diff] [blame] | 322 | AI->setName(outputs[i]->getName()+".out"); |
Misha Brukman | 22108fa | 2004-04-23 23:54:17 +0000 | [diff] [blame] | 323 | } |
Chris Lattner | b5c1dbf | 2004-03-18 03:49:40 +0000 | [diff] [blame] | 324 | |
Misha Brukman | e633603 | 2004-02-28 03:26:20 +0000 | [diff] [blame] | 325 | // Rewrite branches to basic blocks outside of the loop to new dummy blocks |
| 326 | // within the new function. This must be done before we lose track of which |
| 327 | // blocks were originally in the code region. |
| 328 | std::vector<User*> Users(header->use_begin(), header->use_end()); |
Chris Lattner | 5cc8ca6 | 2004-03-18 05:28:49 +0000 | [diff] [blame] | 329 | for (unsigned i = 0, e = Users.size(); i != e; ++i) |
| 330 | // The BasicBlock which contains the branch is not in the region |
| 331 | // modify the branch target to a new block |
| 332 | if (TerminatorInst *TI = dyn_cast<TerminatorInst>(Users[i])) |
| 333 | if (!BlocksToExtract.count(TI->getParent()) && |
| 334 | TI->getParent()->getParent() == oldFunction) |
| 335 | TI->replaceUsesOfWith(header, newHeader); |
Misha Brukman | e633603 | 2004-02-28 03:26:20 +0000 | [diff] [blame] | 336 | |
| 337 | return newFunction; |
| 338 | } |
| 339 | |
Chris Lattner | bf74936 | 2004-05-12 06:01:40 +0000 | [diff] [blame] | 340 | /// emitCallAndSwitchStatement - This method sets up the caller side by adding |
| 341 | /// the call instruction, splitting any PHI nodes in the header block as |
| 342 | /// necessary. |
| 343 | void CodeExtractor:: |
| 344 | emitCallAndSwitchStatement(Function *newFunction, BasicBlock *codeReplacer, |
| 345 | Values &inputs, Values &outputs) { |
| 346 | // Emit a call to the new function, passing in: *pointer to struct (if |
| 347 | // aggregating parameters), or plan inputs and allocated memory for outputs |
Misha Brukman | 22108fa | 2004-04-23 23:54:17 +0000 | [diff] [blame] | 348 | std::vector<Value*> params, StructValues, ReloadOutputs; |
Chris Lattner | 04229c1 | 2004-03-18 04:12:05 +0000 | [diff] [blame] | 349 | |
Misha Brukman | 22108fa | 2004-04-23 23:54:17 +0000 | [diff] [blame] | 350 | // Add inputs as params, or to be filled into the struct |
| 351 | for (Values::iterator i = inputs.begin(), e = inputs.end(); i != e; ++i) |
| 352 | if (AggregateArgs) |
| 353 | StructValues.push_back(*i); |
| 354 | else |
| 355 | params.push_back(*i); |
| 356 | |
| 357 | // Create allocas for the outputs |
| 358 | for (Values::iterator i = outputs.begin(), e = outputs.end(); i != e; ++i) { |
| 359 | if (AggregateArgs) { |
| 360 | StructValues.push_back(*i); |
| 361 | } else { |
| 362 | AllocaInst *alloca = |
| 363 | new AllocaInst((*i)->getType(), 0, (*i)->getName()+".loc", |
| 364 | codeReplacer->getParent()->begin()->begin()); |
| 365 | ReloadOutputs.push_back(alloca); |
| 366 | params.push_back(alloca); |
| 367 | } |
| 368 | } |
| 369 | |
| 370 | AllocaInst *Struct = 0; |
| 371 | if (AggregateArgs && (inputs.size() + outputs.size() > 0)) { |
| 372 | std::vector<const Type*> ArgTypes; |
| 373 | for (Values::iterator v = StructValues.begin(), |
| 374 | ve = StructValues.end(); v != ve; ++v) |
| 375 | ArgTypes.push_back((*v)->getType()); |
| 376 | |
| 377 | // Allocate a struct at the beginning of this function |
| 378 | Type *StructArgTy = StructType::get(ArgTypes); |
Misha Brukman | fd93908 | 2005-04-21 23:48:37 +0000 | [diff] [blame] | 379 | Struct = |
| 380 | new AllocaInst(StructArgTy, 0, "structArg", |
Chris Lattner | b5c1dbf | 2004-03-18 03:49:40 +0000 | [diff] [blame] | 381 | codeReplacer->getParent()->begin()->begin()); |
Misha Brukman | 22108fa | 2004-04-23 23:54:17 +0000 | [diff] [blame] | 382 | params.push_back(Struct); |
| 383 | |
| 384 | for (unsigned i = 0, e = inputs.size(); i != e; ++i) { |
David Greene | b8f7479 | 2007-09-04 15:46:09 +0000 | [diff] [blame] | 385 | Value *Idx[2]; |
| 386 | Idx[0] = Constant::getNullValue(Type::Int32Ty); |
| 387 | Idx[1] = ConstantInt::get(Type::Int32Ty, i); |
Misha Brukman | 22108fa | 2004-04-23 23:54:17 +0000 | [diff] [blame] | 388 | GetElementPtrInst *GEP = |
Gabor Greif | 051a950 | 2008-04-06 20:25:17 +0000 | [diff] [blame] | 389 | GetElementPtrInst::Create(Struct, Idx, Idx + 2, |
| 390 | "gep_" + StructValues[i]->getName()); |
Misha Brukman | 22108fa | 2004-04-23 23:54:17 +0000 | [diff] [blame] | 391 | codeReplacer->getInstList().push_back(GEP); |
| 392 | StoreInst *SI = new StoreInst(StructValues[i], GEP); |
| 393 | codeReplacer->getInstList().push_back(SI); |
| 394 | } |
Misha Brukman | fd93908 | 2005-04-21 23:48:37 +0000 | [diff] [blame] | 395 | } |
Misha Brukman | 22108fa | 2004-04-23 23:54:17 +0000 | [diff] [blame] | 396 | |
| 397 | // Emit the call to the function |
Gabor Greif | 051a950 | 2008-04-06 20:25:17 +0000 | [diff] [blame] | 398 | CallInst *call = CallInst::Create(newFunction, params.begin(), params.end(), |
| 399 | NumExitBlocks > 1 ? "targetBlock" : ""); |
Misha Brukman | 22108fa | 2004-04-23 23:54:17 +0000 | [diff] [blame] | 400 | codeReplacer->getInstList().push_back(call); |
| 401 | |
Chris Lattner | e4d5c44 | 2005-03-15 04:54:21 +0000 | [diff] [blame] | 402 | Function::arg_iterator OutputArgBegin = newFunction->arg_begin(); |
Misha Brukman | 22108fa | 2004-04-23 23:54:17 +0000 | [diff] [blame] | 403 | unsigned FirstOut = inputs.size(); |
| 404 | if (!AggregateArgs) |
| 405 | std::advance(OutputArgBegin, inputs.size()); |
| 406 | |
| 407 | // Reload the outputs passed in by reference |
| 408 | for (unsigned i = 0, e = outputs.size(); i != e; ++i) { |
| 409 | Value *Output = 0; |
| 410 | if (AggregateArgs) { |
David Greene | b8f7479 | 2007-09-04 15:46:09 +0000 | [diff] [blame] | 411 | Value *Idx[2]; |
| 412 | Idx[0] = Constant::getNullValue(Type::Int32Ty); |
| 413 | Idx[1] = ConstantInt::get(Type::Int32Ty, FirstOut + i); |
Misha Brukman | fd93908 | 2005-04-21 23:48:37 +0000 | [diff] [blame] | 414 | GetElementPtrInst *GEP |
Gabor Greif | 051a950 | 2008-04-06 20:25:17 +0000 | [diff] [blame] | 415 | = GetElementPtrInst::Create(Struct, Idx, Idx + 2, |
| 416 | "gep_reload_" + outputs[i]->getName()); |
Misha Brukman | 22108fa | 2004-04-23 23:54:17 +0000 | [diff] [blame] | 417 | codeReplacer->getInstList().push_back(GEP); |
| 418 | Output = GEP; |
| 419 | } else { |
| 420 | Output = ReloadOutputs[i]; |
| 421 | } |
| 422 | LoadInst *load = new LoadInst(Output, outputs[i]->getName()+".reload"); |
Chris Lattner | b5c1dbf | 2004-03-18 03:49:40 +0000 | [diff] [blame] | 423 | codeReplacer->getInstList().push_back(load); |
| 424 | std::vector<User*> Users(outputs[i]->use_begin(), outputs[i]->use_end()); |
| 425 | for (unsigned u = 0, e = Users.size(); u != e; ++u) { |
| 426 | Instruction *inst = cast<Instruction>(Users[u]); |
| 427 | if (!BlocksToExtract.count(inst->getParent())) |
| 428 | inst->replaceUsesOfWith(outputs[i], load); |
Misha Brukman | e633603 | 2004-02-28 03:26:20 +0000 | [diff] [blame] | 429 | } |
| 430 | } |
Chris Lattner | 12f390e | 2004-03-14 23:05:49 +0000 | [diff] [blame] | 431 | |
Misha Brukman | e633603 | 2004-02-28 03:26:20 +0000 | [diff] [blame] | 432 | // Now we can emit a switch statement using the call as a value. |
Chris Lattner | 346be7f | 2004-05-12 04:14:24 +0000 | [diff] [blame] | 433 | SwitchInst *TheSwitch = |
Gabor Greif | 051a950 | 2008-04-06 20:25:17 +0000 | [diff] [blame] | 434 | SwitchInst::Create(ConstantInt::getNullValue(Type::Int16Ty), |
| 435 | codeReplacer, 0, codeReplacer); |
Misha Brukman | e633603 | 2004-02-28 03:26:20 +0000 | [diff] [blame] | 436 | |
| 437 | // Since there may be multiple exits from the original region, make the new |
Chris Lattner | 12f390e | 2004-03-14 23:05:49 +0000 | [diff] [blame] | 438 | // function return an unsigned, switch on that number. This loop iterates |
| 439 | // over all of the blocks in the extracted region, updating any terminator |
| 440 | // instructions in the to-be-extracted region that branch to blocks that are |
| 441 | // not in the region to be extracted. |
| 442 | std::map<BasicBlock*, BasicBlock*> ExitBlockMap; |
| 443 | |
Misha Brukman | e633603 | 2004-02-28 03:26:20 +0000 | [diff] [blame] | 444 | unsigned switchVal = 0; |
Chris Lattner | 0e06674 | 2004-03-14 22:34:55 +0000 | [diff] [blame] | 445 | for (std::set<BasicBlock*>::const_iterator i = BlocksToExtract.begin(), |
| 446 | e = BlocksToExtract.end(); i != e; ++i) { |
Chris Lattner | 12f390e | 2004-03-14 23:05:49 +0000 | [diff] [blame] | 447 | TerminatorInst *TI = (*i)->getTerminator(); |
| 448 | for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i) |
| 449 | if (!BlocksToExtract.count(TI->getSuccessor(i))) { |
| 450 | BasicBlock *OldTarget = TI->getSuccessor(i); |
| 451 | // add a new basic block which returns the appropriate value |
| 452 | BasicBlock *&NewTarget = ExitBlockMap[OldTarget]; |
| 453 | if (!NewTarget) { |
| 454 | // If we don't already have an exit stub for this non-extracted |
| 455 | // destination, create one now! |
Gabor Greif | 051a950 | 2008-04-06 20:25:17 +0000 | [diff] [blame] | 456 | NewTarget = BasicBlock::Create(OldTarget->getName() + ".exitStub", |
| 457 | newFunction); |
Chris Lattner | 346be7f | 2004-05-12 04:14:24 +0000 | [diff] [blame] | 458 | unsigned SuccNum = switchVal++; |
Misha Brukman | e633603 | 2004-02-28 03:26:20 +0000 | [diff] [blame] | 459 | |
Chris Lattner | 346be7f | 2004-05-12 04:14:24 +0000 | [diff] [blame] | 460 | Value *brVal = 0; |
| 461 | switch (NumExitBlocks) { |
| 462 | case 0: |
| 463 | case 1: break; // No value needed. |
| 464 | case 2: // Conditional branch, return a bool |
Reid Spencer | 579dca1 | 2007-01-12 04:24:46 +0000 | [diff] [blame] | 465 | brVal = ConstantInt::get(Type::Int1Ty, !SuccNum); |
Chris Lattner | 346be7f | 2004-05-12 04:14:24 +0000 | [diff] [blame] | 466 | break; |
| 467 | default: |
Reid Spencer | c5b206b | 2006-12-31 05:48:39 +0000 | [diff] [blame] | 468 | brVal = ConstantInt::get(Type::Int16Ty, SuccNum); |
Chris Lattner | 346be7f | 2004-05-12 04:14:24 +0000 | [diff] [blame] | 469 | break; |
| 470 | } |
| 471 | |
Gabor Greif | 051a950 | 2008-04-06 20:25:17 +0000 | [diff] [blame] | 472 | ReturnInst *NTRet = ReturnInst::Create(brVal, NewTarget); |
Misha Brukman | e633603 | 2004-02-28 03:26:20 +0000 | [diff] [blame] | 473 | |
Chris Lattner | 12f390e | 2004-03-14 23:05:49 +0000 | [diff] [blame] | 474 | // Update the switch instruction. |
Reid Spencer | c5b206b | 2006-12-31 05:48:39 +0000 | [diff] [blame] | 475 | TheSwitch->addCase(ConstantInt::get(Type::Int16Ty, SuccNum), |
Chris Lattner | 346be7f | 2004-05-12 04:14:24 +0000 | [diff] [blame] | 476 | OldTarget); |
Misha Brukman | e633603 | 2004-02-28 03:26:20 +0000 | [diff] [blame] | 477 | |
Chris Lattner | 12f390e | 2004-03-14 23:05:49 +0000 | [diff] [blame] | 478 | // Restore values just before we exit |
Chris Lattner | e4d5c44 | 2005-03-15 04:54:21 +0000 | [diff] [blame] | 479 | Function::arg_iterator OAI = OutputArgBegin; |
Misha Brukman | 22108fa | 2004-04-23 23:54:17 +0000 | [diff] [blame] | 480 | for (unsigned out = 0, e = outputs.size(); out != e; ++out) { |
| 481 | // For an invoke, the normal destination is the only one that is |
| 482 | // dominated by the result of the invocation |
| 483 | BasicBlock *DefBlock = cast<Instruction>(outputs[out])->getParent(); |
Chris Lattner | 9ecc046 | 2004-11-13 00:06:45 +0000 | [diff] [blame] | 484 | |
| 485 | bool DominatesDef = true; |
| 486 | |
Chris Lattner | 68c311a | 2004-11-12 23:50:44 +0000 | [diff] [blame] | 487 | if (InvokeInst *Invoke = dyn_cast<InvokeInst>(outputs[out])) { |
Misha Brukman | 22108fa | 2004-04-23 23:54:17 +0000 | [diff] [blame] | 488 | DefBlock = Invoke->getNormalDest(); |
Chris Lattner | 68c311a | 2004-11-12 23:50:44 +0000 | [diff] [blame] | 489 | |
| 490 | // Make sure we are looking at the original successor block, not |
| 491 | // at a newly inserted exit block, which won't be in the dominator |
| 492 | // info. |
| 493 | for (std::map<BasicBlock*, BasicBlock*>::iterator I = |
| 494 | ExitBlockMap.begin(), E = ExitBlockMap.end(); I != E; ++I) |
| 495 | if (DefBlock == I->second) { |
| 496 | DefBlock = I->first; |
| 497 | break; |
| 498 | } |
Chris Lattner | 9ecc046 | 2004-11-13 00:06:45 +0000 | [diff] [blame] | 499 | |
| 500 | // In the extract block case, if the block we are extracting ends |
| 501 | // with an invoke instruction, make sure that we don't emit a |
| 502 | // store of the invoke value for the unwind block. |
Devang Patel | 4b90e3a | 2007-06-07 22:17:16 +0000 | [diff] [blame] | 503 | if (!DT && DefBlock != OldTarget) |
Chris Lattner | 9ecc046 | 2004-11-13 00:06:45 +0000 | [diff] [blame] | 504 | DominatesDef = false; |
Chris Lattner | 68c311a | 2004-11-12 23:50:44 +0000 | [diff] [blame] | 505 | } |
| 506 | |
Devang Patel | 4b90e3a | 2007-06-07 22:17:16 +0000 | [diff] [blame] | 507 | if (DT) |
| 508 | DominatesDef = DT->dominates(DefBlock, OldTarget); |
Chris Lattner | 9ecc046 | 2004-11-13 00:06:45 +0000 | [diff] [blame] | 509 | |
| 510 | if (DominatesDef) { |
Misha Brukman | 22108fa | 2004-04-23 23:54:17 +0000 | [diff] [blame] | 511 | if (AggregateArgs) { |
David Greene | b8f7479 | 2007-09-04 15:46:09 +0000 | [diff] [blame] | 512 | Value *Idx[2]; |
| 513 | Idx[0] = Constant::getNullValue(Type::Int32Ty); |
| 514 | Idx[1] = ConstantInt::get(Type::Int32Ty,FirstOut+out); |
Misha Brukman | 22108fa | 2004-04-23 23:54:17 +0000 | [diff] [blame] | 515 | GetElementPtrInst *GEP = |
Gabor Greif | 051a950 | 2008-04-06 20:25:17 +0000 | [diff] [blame] | 516 | GetElementPtrInst::Create(OAI, Idx, Idx + 2, |
| 517 | "gep_" + outputs[out]->getName(), |
| 518 | NTRet); |
Misha Brukman | 22108fa | 2004-04-23 23:54:17 +0000 | [diff] [blame] | 519 | new StoreInst(outputs[out], GEP, NTRet); |
Chris Lattner | 9ecc046 | 2004-11-13 00:06:45 +0000 | [diff] [blame] | 520 | } else { |
Misha Brukman | 22108fa | 2004-04-23 23:54:17 +0000 | [diff] [blame] | 521 | new StoreInst(outputs[out], OAI, NTRet); |
Chris Lattner | 9ecc046 | 2004-11-13 00:06:45 +0000 | [diff] [blame] | 522 | } |
| 523 | } |
Misha Brukman | 22108fa | 2004-04-23 23:54:17 +0000 | [diff] [blame] | 524 | // Advance output iterator even if we don't emit a store |
| 525 | if (!AggregateArgs) ++OAI; |
| 526 | } |
Misha Brukman | e633603 | 2004-02-28 03:26:20 +0000 | [diff] [blame] | 527 | } |
Chris Lattner | 12f390e | 2004-03-14 23:05:49 +0000 | [diff] [blame] | 528 | |
| 529 | // rewrite the original branch instruction with this new target |
| 530 | TI->setSuccessor(i, NewTarget); |
Misha Brukman | e633603 | 2004-02-28 03:26:20 +0000 | [diff] [blame] | 531 | } |
Misha Brukman | e633603 | 2004-02-28 03:26:20 +0000 | [diff] [blame] | 532 | } |
Chris Lattner | 65826bf | 2004-03-14 23:43:24 +0000 | [diff] [blame] | 533 | |
Chris Lattner | 5b01e29 | 2004-05-12 03:22:33 +0000 | [diff] [blame] | 534 | // Now that we've done the deed, simplify the switch instruction. |
Chris Lattner | b519efb | 2004-08-12 03:17:02 +0000 | [diff] [blame] | 535 | const Type *OldFnRetTy = TheSwitch->getParent()->getParent()->getReturnType(); |
Chris Lattner | 346be7f | 2004-05-12 04:14:24 +0000 | [diff] [blame] | 536 | switch (NumExitBlocks) { |
| 537 | case 0: |
Chris Lattner | b519efb | 2004-08-12 03:17:02 +0000 | [diff] [blame] | 538 | // There are no successors (the block containing the switch itself), which |
Misha Brukman | 22108fa | 2004-04-23 23:54:17 +0000 | [diff] [blame] | 539 | // means that previously this was the last part of the function, and hence |
| 540 | // this should be rewritten as a `ret' |
Misha Brukman | fd93908 | 2005-04-21 23:48:37 +0000 | [diff] [blame] | 541 | |
Misha Brukman | 22108fa | 2004-04-23 23:54:17 +0000 | [diff] [blame] | 542 | // Check if the function should return a value |
Chris Lattner | b519efb | 2004-08-12 03:17:02 +0000 | [diff] [blame] | 543 | if (OldFnRetTy == Type::VoidTy) { |
Gabor Greif | 051a950 | 2008-04-06 20:25:17 +0000 | [diff] [blame] | 544 | ReturnInst::Create(0, TheSwitch); // Return void |
Chris Lattner | b519efb | 2004-08-12 03:17:02 +0000 | [diff] [blame] | 545 | } else if (OldFnRetTy == TheSwitch->getCondition()->getType()) { |
Misha Brukman | 22108fa | 2004-04-23 23:54:17 +0000 | [diff] [blame] | 546 | // return what we have |
Gabor Greif | 051a950 | 2008-04-06 20:25:17 +0000 | [diff] [blame] | 547 | ReturnInst::Create(TheSwitch->getCondition(), TheSwitch); |
Chris Lattner | b519efb | 2004-08-12 03:17:02 +0000 | [diff] [blame] | 548 | } else { |
| 549 | // Otherwise we must have code extracted an unwind or something, just |
| 550 | // return whatever we want. |
Gabor Greif | 051a950 | 2008-04-06 20:25:17 +0000 | [diff] [blame] | 551 | ReturnInst::Create(Constant::getNullValue(OldFnRetTy), TheSwitch); |
Chris Lattner | b519efb | 2004-08-12 03:17:02 +0000 | [diff] [blame] | 552 | } |
Misha Brukman | 22108fa | 2004-04-23 23:54:17 +0000 | [diff] [blame] | 553 | |
| 554 | TheSwitch->getParent()->getInstList().erase(TheSwitch); |
Chris Lattner | 346be7f | 2004-05-12 04:14:24 +0000 | [diff] [blame] | 555 | break; |
| 556 | case 1: |
| 557 | // Only a single destination, change the switch into an unconditional |
| 558 | // branch. |
Gabor Greif | 051a950 | 2008-04-06 20:25:17 +0000 | [diff] [blame] | 559 | BranchInst::Create(TheSwitch->getSuccessor(1), TheSwitch); |
Chris Lattner | 346be7f | 2004-05-12 04:14:24 +0000 | [diff] [blame] | 560 | TheSwitch->getParent()->getInstList().erase(TheSwitch); |
| 561 | break; |
| 562 | case 2: |
Gabor Greif | 051a950 | 2008-04-06 20:25:17 +0000 | [diff] [blame] | 563 | BranchInst::Create(TheSwitch->getSuccessor(1), TheSwitch->getSuccessor(2), |
| 564 | call, TheSwitch); |
Chris Lattner | 346be7f | 2004-05-12 04:14:24 +0000 | [diff] [blame] | 565 | TheSwitch->getParent()->getInstList().erase(TheSwitch); |
| 566 | break; |
| 567 | default: |
| 568 | // Otherwise, make the default destination of the switch instruction be one |
| 569 | // of the other successors. |
| 570 | TheSwitch->setOperand(0, call); |
| 571 | TheSwitch->setSuccessor(0, TheSwitch->getSuccessor(NumExitBlocks)); |
| 572 | TheSwitch->removeCase(NumExitBlocks); // Remove redundant case |
| 573 | break; |
Chris Lattner | 65826bf | 2004-03-14 23:43:24 +0000 | [diff] [blame] | 574 | } |
Misha Brukman | e633603 | 2004-02-28 03:26:20 +0000 | [diff] [blame] | 575 | } |
| 576 | |
Chris Lattner | bf74936 | 2004-05-12 06:01:40 +0000 | [diff] [blame] | 577 | void CodeExtractor::moveCodeToFunction(Function *newFunction) { |
| 578 | Function *oldFunc = (*BlocksToExtract.begin())->getParent(); |
| 579 | Function::BasicBlockListType &oldBlocks = oldFunc->getBasicBlockList(); |
| 580 | Function::BasicBlockListType &newBlocks = newFunction->getBasicBlockList(); |
| 581 | |
| 582 | for (std::set<BasicBlock*>::const_iterator i = BlocksToExtract.begin(), |
| 583 | e = BlocksToExtract.end(); i != e; ++i) { |
| 584 | // Delete the basic block from the old function, and the list of blocks |
| 585 | oldBlocks.remove(*i); |
| 586 | |
| 587 | // Insert this basic block into the new function |
| 588 | newBlocks.push_back(*i); |
| 589 | } |
| 590 | } |
Misha Brukman | e633603 | 2004-02-28 03:26:20 +0000 | [diff] [blame] | 591 | |
| 592 | /// ExtractRegion - Removes a loop from a function, replaces it with a call to |
| 593 | /// new function. Returns pointer to the new function. |
| 594 | /// |
| 595 | /// algorithm: |
| 596 | /// |
| 597 | /// find inputs and outputs for the region |
| 598 | /// |
Misha Brukman | fd93908 | 2005-04-21 23:48:37 +0000 | [diff] [blame] | 599 | /// for inputs: add to function as args, map input instr* to arg# |
| 600 | /// for outputs: add allocas for scalars, |
Misha Brukman | e633603 | 2004-02-28 03:26:20 +0000 | [diff] [blame] | 601 | /// add to func as args, map output instr* to arg# |
| 602 | /// |
| 603 | /// rewrite func to use argument #s instead of instr* |
| 604 | /// |
Misha Brukman | fd93908 | 2005-04-21 23:48:37 +0000 | [diff] [blame] | 605 | /// for each scalar output in the function: at every exit, store intermediate |
Misha Brukman | e633603 | 2004-02-28 03:26:20 +0000 | [diff] [blame] | 606 | /// computed result back into memory. |
| 607 | /// |
Chris Lattner | b519efb | 2004-08-12 03:17:02 +0000 | [diff] [blame] | 608 | Function *CodeExtractor:: |
| 609 | ExtractCodeRegion(const std::vector<BasicBlock*> &code) { |
Misha Brukman | 22108fa | 2004-04-23 23:54:17 +0000 | [diff] [blame] | 610 | if (!isEligible(code)) |
| 611 | return 0; |
| 612 | |
Misha Brukman | e633603 | 2004-02-28 03:26:20 +0000 | [diff] [blame] | 613 | // 1) Find inputs, outputs |
| 614 | // 2) Construct new function |
| 615 | // * Add allocas for defs, pass as args by reference |
| 616 | // * Pass in uses as args |
| 617 | // 3) Move code region, add call instr to func |
Chris Lattner | 0e06674 | 2004-03-14 22:34:55 +0000 | [diff] [blame] | 618 | // |
| 619 | BlocksToExtract.insert(code.begin(), code.end()); |
Misha Brukman | e633603 | 2004-02-28 03:26:20 +0000 | [diff] [blame] | 620 | |
| 621 | Values inputs, outputs; |
| 622 | |
| 623 | // Assumption: this is a single-entry code region, and the header is the first |
Chris Lattner | 0de632b | 2004-03-15 01:18:23 +0000 | [diff] [blame] | 624 | // block in the region. |
Misha Brukman | e633603 | 2004-02-28 03:26:20 +0000 | [diff] [blame] | 625 | BasicBlock *header = code[0]; |
Chris Lattner | bf74936 | 2004-05-12 06:01:40 +0000 | [diff] [blame] | 626 | |
Chris Lattner | 0de632b | 2004-03-15 01:18:23 +0000 | [diff] [blame] | 627 | for (unsigned i = 1, e = code.size(); i != e; ++i) |
| 628 | for (pred_iterator PI = pred_begin(code[i]), E = pred_end(code[i]); |
| 629 | PI != E; ++PI) |
| 630 | assert(BlocksToExtract.count(*PI) && |
| 631 | "No blocks in this region may have entries from outside the region" |
| 632 | " except for the first block!"); |
Misha Brukman | fd93908 | 2005-04-21 23:48:37 +0000 | [diff] [blame] | 633 | |
Chris Lattner | d99e1d3 | 2004-05-12 16:07:41 +0000 | [diff] [blame] | 634 | // If we have to split PHI nodes or the entry block, do so now. |
Chris Lattner | e746ad5 | 2004-05-12 15:29:13 +0000 | [diff] [blame] | 635 | severSplitPHINodes(header); |
| 636 | |
Chris Lattner | d99e1d3 | 2004-05-12 16:07:41 +0000 | [diff] [blame] | 637 | // If we have any return instructions in the region, split those blocks so |
| 638 | // that the return is not in the region. |
| 639 | splitReturnBlocks(); |
| 640 | |
Misha Brukman | e633603 | 2004-02-28 03:26:20 +0000 | [diff] [blame] | 641 | Function *oldFunction = header->getParent(); |
Misha Brukman | e633603 | 2004-02-28 03:26:20 +0000 | [diff] [blame] | 642 | |
| 643 | // This takes place of the original loop |
Gabor Greif | 051a950 | 2008-04-06 20:25:17 +0000 | [diff] [blame] | 644 | BasicBlock *codeReplacer = BasicBlock::Create("codeRepl", oldFunction, header); |
Misha Brukman | e633603 | 2004-02-28 03:26:20 +0000 | [diff] [blame] | 645 | |
| 646 | // The new function needs a root node because other nodes can branch to the |
Chris Lattner | bf74936 | 2004-05-12 06:01:40 +0000 | [diff] [blame] | 647 | // head of the region, but the entry node of a function cannot have preds. |
Gabor Greif | 051a950 | 2008-04-06 20:25:17 +0000 | [diff] [blame] | 648 | BasicBlock *newFuncRoot = BasicBlock::Create("newFuncRoot"); |
| 649 | newFuncRoot->getInstList().push_back(BranchInst::Create(header)); |
Misha Brukman | e633603 | 2004-02-28 03:26:20 +0000 | [diff] [blame] | 650 | |
Chris Lattner | bf74936 | 2004-05-12 06:01:40 +0000 | [diff] [blame] | 651 | // Find inputs to, outputs from the code region. |
| 652 | findInputsOutputs(inputs, outputs); |
Misha Brukman | e633603 | 2004-02-28 03:26:20 +0000 | [diff] [blame] | 653 | |
Chris Lattner | bf74936 | 2004-05-12 06:01:40 +0000 | [diff] [blame] | 654 | // Construct new function based on inputs/outputs & add allocas for all defs. |
Chris Lattner | e746ad5 | 2004-05-12 15:29:13 +0000 | [diff] [blame] | 655 | Function *newFunction = constructFunction(inputs, outputs, header, |
Misha Brukman | fd93908 | 2005-04-21 23:48:37 +0000 | [diff] [blame] | 656 | newFuncRoot, |
Chris Lattner | 0de632b | 2004-03-15 01:18:23 +0000 | [diff] [blame] | 657 | codeReplacer, oldFunction, |
| 658 | oldFunction->getParent()); |
Misha Brukman | e633603 | 2004-02-28 03:26:20 +0000 | [diff] [blame] | 659 | |
Chris Lattner | 0e06674 | 2004-03-14 22:34:55 +0000 | [diff] [blame] | 660 | emitCallAndSwitchStatement(newFunction, codeReplacer, inputs, outputs); |
Misha Brukman | e633603 | 2004-02-28 03:26:20 +0000 | [diff] [blame] | 661 | |
Chris Lattner | 0e06674 | 2004-03-14 22:34:55 +0000 | [diff] [blame] | 662 | moveCodeToFunction(newFunction); |
Misha Brukman | e633603 | 2004-02-28 03:26:20 +0000 | [diff] [blame] | 663 | |
Chris Lattner | e746ad5 | 2004-05-12 15:29:13 +0000 | [diff] [blame] | 664 | // Loop over all of the PHI nodes in the header block, and change any |
Chris Lattner | 5cc8ca6 | 2004-03-18 05:28:49 +0000 | [diff] [blame] | 665 | // references to the old incoming edge to be the new incoming edge. |
Reid Spencer | 2da5c3d | 2004-09-15 17:06:42 +0000 | [diff] [blame] | 666 | for (BasicBlock::iterator I = header->begin(); isa<PHINode>(I); ++I) { |
| 667 | PHINode *PN = cast<PHINode>(I); |
Chris Lattner | 5cc8ca6 | 2004-03-18 05:28:49 +0000 | [diff] [blame] | 668 | for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) |
| 669 | if (!BlocksToExtract.count(PN->getIncomingBlock(i))) |
| 670 | PN->setIncomingBlock(i, newFuncRoot); |
Reid Spencer | 2da5c3d | 2004-09-15 17:06:42 +0000 | [diff] [blame] | 671 | } |
Misha Brukman | fd93908 | 2005-04-21 23:48:37 +0000 | [diff] [blame] | 672 | |
Chris Lattner | d07f899 | 2004-03-18 05:38:31 +0000 | [diff] [blame] | 673 | // Look at all successors of the codeReplacer block. If any of these blocks |
| 674 | // had PHI nodes in them, we need to update the "from" block to be the code |
| 675 | // replacer, not the original block in the extracted region. |
| 676 | std::vector<BasicBlock*> Succs(succ_begin(codeReplacer), |
| 677 | succ_end(codeReplacer)); |
| 678 | for (unsigned i = 0, e = Succs.size(); i != e; ++i) |
Reid Spencer | 2da5c3d | 2004-09-15 17:06:42 +0000 | [diff] [blame] | 679 | for (BasicBlock::iterator I = Succs[i]->begin(); isa<PHINode>(I); ++I) { |
| 680 | PHINode *PN = cast<PHINode>(I); |
Chris Lattner | a670c68 | 2004-08-13 03:27:07 +0000 | [diff] [blame] | 681 | std::set<BasicBlock*> ProcessedPreds; |
Chris Lattner | d07f899 | 2004-03-18 05:38:31 +0000 | [diff] [blame] | 682 | for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) |
Anton Korobeynikov | 07e6e56 | 2008-02-20 11:26:25 +0000 | [diff] [blame] | 683 | if (BlocksToExtract.count(PN->getIncomingBlock(i))) { |
Chris Lattner | a670c68 | 2004-08-13 03:27:07 +0000 | [diff] [blame] | 684 | if (ProcessedPreds.insert(PN->getIncomingBlock(i)).second) |
| 685 | PN->setIncomingBlock(i, codeReplacer); |
| 686 | else { |
| 687 | // There were multiple entries in the PHI for this block, now there |
| 688 | // is only one, so remove the duplicated entries. |
| 689 | PN->removeIncomingValue(i, false); |
| 690 | --i; --e; |
| 691 | } |
Anton Korobeynikov | 07e6e56 | 2008-02-20 11:26:25 +0000 | [diff] [blame] | 692 | } |
Chris Lattner | a670c68 | 2004-08-13 03:27:07 +0000 | [diff] [blame] | 693 | } |
Misha Brukman | fd93908 | 2005-04-21 23:48:37 +0000 | [diff] [blame] | 694 | |
Bill Wendling | e815619 | 2006-12-07 01:30:32 +0000 | [diff] [blame] | 695 | //cerr << "NEW FUNCTION: " << *newFunction; |
Chris Lattner | e746ad5 | 2004-05-12 15:29:13 +0000 | [diff] [blame] | 696 | // verifyFunction(*newFunction); |
| 697 | |
Bill Wendling | e815619 | 2006-12-07 01:30:32 +0000 | [diff] [blame] | 698 | // cerr << "OLD FUNCTION: " << *oldFunction; |
Chris Lattner | e746ad5 | 2004-05-12 15:29:13 +0000 | [diff] [blame] | 699 | // verifyFunction(*oldFunction); |
Chris Lattner | d07f899 | 2004-03-18 05:38:31 +0000 | [diff] [blame] | 700 | |
Chris Lattner | ffada93 | 2004-03-14 03:17:22 +0000 | [diff] [blame] | 701 | DEBUG(if (verifyFunction(*newFunction)) abort()); |
Misha Brukman | e633603 | 2004-02-28 03:26:20 +0000 | [diff] [blame] | 702 | return newFunction; |
| 703 | } |
| 704 | |
Misha Brukman | 22108fa | 2004-04-23 23:54:17 +0000 | [diff] [blame] | 705 | bool CodeExtractor::isEligible(const std::vector<BasicBlock*> &code) { |
Chris Lattner | bf74936 | 2004-05-12 06:01:40 +0000 | [diff] [blame] | 706 | // Deny code region if it contains allocas or vastarts. |
Misha Brukman | 22108fa | 2004-04-23 23:54:17 +0000 | [diff] [blame] | 707 | for (std::vector<BasicBlock*>::const_iterator BB = code.begin(), e=code.end(); |
| 708 | BB != e; ++BB) |
| 709 | for (BasicBlock::const_iterator I = (*BB)->begin(), Ie = (*BB)->end(); |
| 710 | I != Ie; ++I) |
| 711 | if (isa<AllocaInst>(*I)) |
| 712 | return false; |
Chris Lattner | bf74936 | 2004-05-12 06:01:40 +0000 | [diff] [blame] | 713 | else if (const CallInst *CI = dyn_cast<CallInst>(I)) |
| 714 | if (const Function *F = CI->getCalledFunction()) |
| 715 | if (F->getIntrinsicID() == Intrinsic::vastart) |
| 716 | return false; |
Misha Brukman | 22108fa | 2004-04-23 23:54:17 +0000 | [diff] [blame] | 717 | return true; |
| 718 | } |
| 719 | |
| 720 | |
Misha Brukman | 0256e4b | 2004-03-02 00:20:57 +0000 | [diff] [blame] | 721 | /// ExtractCodeRegion - slurp a sequence of basic blocks into a brand new |
| 722 | /// function |
| 723 | /// |
Devang Patel | 4b90e3a | 2007-06-07 22:17:16 +0000 | [diff] [blame] | 724 | Function* llvm::ExtractCodeRegion(DominatorTree &DT, |
Misha Brukman | 22108fa | 2004-04-23 23:54:17 +0000 | [diff] [blame] | 725 | const std::vector<BasicBlock*> &code, |
| 726 | bool AggregateArgs) { |
Devang Patel | 4b90e3a | 2007-06-07 22:17:16 +0000 | [diff] [blame] | 727 | return CodeExtractor(&DT, AggregateArgs).ExtractCodeRegion(code); |
Misha Brukman | 0256e4b | 2004-03-02 00:20:57 +0000 | [diff] [blame] | 728 | } |
| 729 | |
Misha Brukman | b97fce5 | 2004-03-01 18:28:34 +0000 | [diff] [blame] | 730 | /// ExtractBasicBlock - slurp a natural loop into a brand new function |
| 731 | /// |
Devang Patel | 4b90e3a | 2007-06-07 22:17:16 +0000 | [diff] [blame] | 732 | Function* llvm::ExtractLoop(DominatorTree &DT, Loop *L, bool AggregateArgs) { |
| 733 | return CodeExtractor(&DT, AggregateArgs).ExtractCodeRegion(L->getBlocks()); |
Misha Brukman | e633603 | 2004-02-28 03:26:20 +0000 | [diff] [blame] | 734 | } |
| 735 | |
Misha Brukman | b97fce5 | 2004-03-01 18:28:34 +0000 | [diff] [blame] | 736 | /// ExtractBasicBlock - slurp a basic block into a brand new function |
| 737 | /// |
Misha Brukman | 22108fa | 2004-04-23 23:54:17 +0000 | [diff] [blame] | 738 | Function* llvm::ExtractBasicBlock(BasicBlock *BB, bool AggregateArgs) { |
Misha Brukman | b97fce5 | 2004-03-01 18:28:34 +0000 | [diff] [blame] | 739 | std::vector<BasicBlock*> Blocks; |
| 740 | Blocks.push_back(BB); |
Devang Patel | 4b90e3a | 2007-06-07 22:17:16 +0000 | [diff] [blame] | 741 | return CodeExtractor(0, AggregateArgs).ExtractCodeRegion(Blocks); |
Misha Brukman | b97fce5 | 2004-03-01 18:28:34 +0000 | [diff] [blame] | 742 | } |