File: | lib/CodeGen/BranchFolding.cpp |
Warning: | line 1165, column 16 The left operand of '!=' is a garbage value |
Press '?' to see keyboard shortcuts
Keyboard shortcuts:
1 | //===- BranchFolding.cpp - Fold machine code branch instructions ----------===// | |||
2 | // | |||
3 | // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. | |||
4 | // See https://llvm.org/LICENSE.txt for license information. | |||
5 | // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception | |||
6 | // | |||
7 | //===----------------------------------------------------------------------===// | |||
8 | // | |||
9 | // This pass forwards branches to unconditional branches to make them branch | |||
10 | // directly to the target block. This pass often results in dead MBB's, which | |||
11 | // it then removes. | |||
12 | // | |||
13 | // Note that this pass must be run after register allocation, it cannot handle | |||
14 | // SSA form. It also must handle virtual registers for targets that emit virtual | |||
15 | // ISA (e.g. NVPTX). | |||
16 | // | |||
17 | //===----------------------------------------------------------------------===// | |||
18 | ||||
19 | #include "BranchFolding.h" | |||
20 | #include "llvm/ADT/BitVector.h" | |||
21 | #include "llvm/ADT/DenseMap.h" | |||
22 | #include "llvm/ADT/STLExtras.h" | |||
23 | #include "llvm/ADT/SmallPtrSet.h" | |||
24 | #include "llvm/ADT/SmallSet.h" | |||
25 | #include "llvm/ADT/SmallVector.h" | |||
26 | #include "llvm/ADT/Statistic.h" | |||
27 | #include "llvm/CodeGen/Analysis.h" | |||
28 | #include "llvm/CodeGen/LivePhysRegs.h" | |||
29 | #include "llvm/CodeGen/MachineBasicBlock.h" | |||
30 | #include "llvm/CodeGen/MachineBlockFrequencyInfo.h" | |||
31 | #include "llvm/CodeGen/MachineBranchProbabilityInfo.h" | |||
32 | #include "llvm/CodeGen/MachineFunction.h" | |||
33 | #include "llvm/CodeGen/MachineFunctionPass.h" | |||
34 | #include "llvm/CodeGen/MachineInstr.h" | |||
35 | #include "llvm/CodeGen/MachineInstrBuilder.h" | |||
36 | #include "llvm/CodeGen/MachineJumpTableInfo.h" | |||
37 | #include "llvm/CodeGen/MachineLoopInfo.h" | |||
38 | #include "llvm/CodeGen/MachineModuleInfo.h" | |||
39 | #include "llvm/CodeGen/MachineOperand.h" | |||
40 | #include "llvm/CodeGen/MachineRegisterInfo.h" | |||
41 | #include "llvm/CodeGen/TargetInstrInfo.h" | |||
42 | #include "llvm/CodeGen/TargetOpcodes.h" | |||
43 | #include "llvm/CodeGen/TargetPassConfig.h" | |||
44 | #include "llvm/CodeGen/TargetRegisterInfo.h" | |||
45 | #include "llvm/CodeGen/TargetSubtargetInfo.h" | |||
46 | #include "llvm/IR/DebugInfoMetadata.h" | |||
47 | #include "llvm/IR/DebugLoc.h" | |||
48 | #include "llvm/IR/Function.h" | |||
49 | #include "llvm/MC/LaneBitmask.h" | |||
50 | #include "llvm/MC/MCRegisterInfo.h" | |||
51 | #include "llvm/Pass.h" | |||
52 | #include "llvm/Support/BlockFrequency.h" | |||
53 | #include "llvm/Support/BranchProbability.h" | |||
54 | #include "llvm/Support/CommandLine.h" | |||
55 | #include "llvm/Support/Debug.h" | |||
56 | #include "llvm/Support/ErrorHandling.h" | |||
57 | #include "llvm/Support/raw_ostream.h" | |||
58 | #include "llvm/Target/TargetMachine.h" | |||
59 | #include <cassert> | |||
60 | #include <cstddef> | |||
61 | #include <iterator> | |||
62 | #include <numeric> | |||
63 | #include <vector> | |||
64 | ||||
65 | using namespace llvm; | |||
66 | ||||
67 | #define DEBUG_TYPE"branch-folder" "branch-folder" | |||
68 | ||||
69 | STATISTIC(NumDeadBlocks, "Number of dead blocks removed")static llvm::Statistic NumDeadBlocks = {"branch-folder", "NumDeadBlocks" , "Number of dead blocks removed", {0}, {false}}; | |||
70 | STATISTIC(NumBranchOpts, "Number of branches optimized")static llvm::Statistic NumBranchOpts = {"branch-folder", "NumBranchOpts" , "Number of branches optimized", {0}, {false}}; | |||
71 | STATISTIC(NumTailMerge , "Number of block tails merged")static llvm::Statistic NumTailMerge = {"branch-folder", "NumTailMerge" , "Number of block tails merged", {0}, {false}}; | |||
72 | STATISTIC(NumHoist , "Number of times common instructions are hoisted")static llvm::Statistic NumHoist = {"branch-folder", "NumHoist" , "Number of times common instructions are hoisted", {0}, {false }}; | |||
73 | STATISTIC(NumTailCalls, "Number of tail calls optimized")static llvm::Statistic NumTailCalls = {"branch-folder", "NumTailCalls" , "Number of tail calls optimized", {0}, {false}}; | |||
74 | ||||
75 | static cl::opt<cl::boolOrDefault> FlagEnableTailMerge("enable-tail-merge", | |||
76 | cl::init(cl::BOU_UNSET), cl::Hidden); | |||
77 | ||||
78 | // Throttle for huge numbers of predecessors (compile speed problems) | |||
79 | static cl::opt<unsigned> | |||
80 | TailMergeThreshold("tail-merge-threshold", | |||
81 | cl::desc("Max number of predecessors to consider tail merging"), | |||
82 | cl::init(150), cl::Hidden); | |||
83 | ||||
84 | // Heuristic for tail merging (and, inversely, tail duplication). | |||
85 | // TODO: This should be replaced with a target query. | |||
86 | static cl::opt<unsigned> | |||
87 | TailMergeSize("tail-merge-size", | |||
88 | cl::desc("Min number of instructions to consider tail merging"), | |||
89 | cl::init(3), cl::Hidden); | |||
90 | ||||
91 | namespace { | |||
92 | ||||
93 | /// BranchFolderPass - Wrap branch folder in a machine function pass. | |||
94 | class BranchFolderPass : public MachineFunctionPass { | |||
95 | public: | |||
96 | static char ID; | |||
97 | ||||
98 | explicit BranchFolderPass(): MachineFunctionPass(ID) {} | |||
99 | ||||
100 | bool runOnMachineFunction(MachineFunction &MF) override; | |||
101 | ||||
102 | void getAnalysisUsage(AnalysisUsage &AU) const override { | |||
103 | AU.addRequired<MachineBlockFrequencyInfo>(); | |||
104 | AU.addRequired<MachineBranchProbabilityInfo>(); | |||
105 | AU.addRequired<TargetPassConfig>(); | |||
106 | MachineFunctionPass::getAnalysisUsage(AU); | |||
107 | } | |||
108 | }; | |||
109 | ||||
110 | } // end anonymous namespace | |||
111 | ||||
112 | char BranchFolderPass::ID = 0; | |||
113 | ||||
114 | char &llvm::BranchFolderPassID = BranchFolderPass::ID; | |||
115 | ||||
116 | INITIALIZE_PASS(BranchFolderPass, DEBUG_TYPE,static void *initializeBranchFolderPassPassOnce(PassRegistry & Registry) { PassInfo *PI = new PassInfo( "Control Flow Optimizer" , "branch-folder", &BranchFolderPass::ID, PassInfo::NormalCtor_t (callDefaultCtor<BranchFolderPass>), false, false); Registry .registerPass(*PI, true); return PI; } static llvm::once_flag InitializeBranchFolderPassPassFlag; void llvm::initializeBranchFolderPassPass (PassRegistry &Registry) { llvm::call_once(InitializeBranchFolderPassPassFlag , initializeBranchFolderPassPassOnce, std::ref(Registry)); } | |||
117 | "Control Flow Optimizer", false, false)static void *initializeBranchFolderPassPassOnce(PassRegistry & Registry) { PassInfo *PI = new PassInfo( "Control Flow Optimizer" , "branch-folder", &BranchFolderPass::ID, PassInfo::NormalCtor_t (callDefaultCtor<BranchFolderPass>), false, false); Registry .registerPass(*PI, true); return PI; } static llvm::once_flag InitializeBranchFolderPassPassFlag; void llvm::initializeBranchFolderPassPass (PassRegistry &Registry) { llvm::call_once(InitializeBranchFolderPassPassFlag , initializeBranchFolderPassPassOnce, std::ref(Registry)); } | |||
118 | ||||
119 | bool BranchFolderPass::runOnMachineFunction(MachineFunction &MF) { | |||
120 | if (skipFunction(MF.getFunction())) | |||
| ||||
121 | return false; | |||
122 | ||||
123 | TargetPassConfig *PassConfig = &getAnalysis<TargetPassConfig>(); | |||
124 | // TailMerge can create jump into if branches that make CFG irreducible for | |||
125 | // HW that requires structurized CFG. | |||
126 | bool EnableTailMerge = !MF.getTarget().requiresStructuredCFG() && | |||
127 | PassConfig->getEnableTailMerge(); | |||
128 | BranchFolder::MBFIWrapper MBBFreqInfo( | |||
129 | getAnalysis<MachineBlockFrequencyInfo>()); | |||
130 | BranchFolder Folder(EnableTailMerge, /*CommonHoist=*/true, MBBFreqInfo, | |||
131 | getAnalysis<MachineBranchProbabilityInfo>()); | |||
132 | return Folder.OptimizeFunction(MF, MF.getSubtarget().getInstrInfo(), | |||
133 | MF.getSubtarget().getRegisterInfo(), | |||
134 | getAnalysisIfAvailable<MachineModuleInfo>()); | |||
135 | } | |||
136 | ||||
137 | BranchFolder::BranchFolder(bool defaultEnableTailMerge, bool CommonHoist, | |||
138 | MBFIWrapper &FreqInfo, | |||
139 | const MachineBranchProbabilityInfo &ProbInfo, | |||
140 | unsigned MinTailLength) | |||
141 | : EnableHoistCommonCode(CommonHoist), MinCommonTailLength(MinTailLength), | |||
142 | MBBFreqInfo(FreqInfo), MBPI(ProbInfo) { | |||
143 | if (MinCommonTailLength == 0) | |||
144 | MinCommonTailLength = TailMergeSize; | |||
145 | switch (FlagEnableTailMerge) { | |||
146 | case cl::BOU_UNSET: EnableTailMerge = defaultEnableTailMerge; break; | |||
147 | case cl::BOU_TRUE: EnableTailMerge = true; break; | |||
148 | case cl::BOU_FALSE: EnableTailMerge = false; break; | |||
149 | } | |||
150 | } | |||
151 | ||||
152 | void BranchFolder::RemoveDeadBlock(MachineBasicBlock *MBB) { | |||
153 | assert(MBB->pred_empty() && "MBB must be dead!")((MBB->pred_empty() && "MBB must be dead!") ? static_cast <void> (0) : __assert_fail ("MBB->pred_empty() && \"MBB must be dead!\"" , "/build/llvm-toolchain-snapshot-9~svn362543/lib/CodeGen/BranchFolding.cpp" , 153, __PRETTY_FUNCTION__)); | |||
154 | LLVM_DEBUG(dbgs() << "\nRemoving MBB: " << *MBB)do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("branch-folder")) { dbgs() << "\nRemoving MBB: " << *MBB; } } while (false); | |||
155 | ||||
156 | MachineFunction *MF = MBB->getParent(); | |||
157 | // drop all successors. | |||
158 | while (!MBB->succ_empty()) | |||
159 | MBB->removeSuccessor(MBB->succ_end()-1); | |||
160 | ||||
161 | // Avoid matching if this pointer gets reused. | |||
162 | TriedMerging.erase(MBB); | |||
163 | ||||
164 | // Remove the block. | |||
165 | MF->erase(MBB); | |||
166 | EHScopeMembership.erase(MBB); | |||
167 | if (MLI) | |||
168 | MLI->removeBlock(MBB); | |||
169 | } | |||
170 | ||||
171 | bool BranchFolder::OptimizeFunction(MachineFunction &MF, | |||
172 | const TargetInstrInfo *tii, | |||
173 | const TargetRegisterInfo *tri, | |||
174 | MachineModuleInfo *mmi, | |||
175 | MachineLoopInfo *mli, bool AfterPlacement) { | |||
176 | if (!tii) return false; | |||
177 | ||||
178 | TriedMerging.clear(); | |||
179 | ||||
180 | MachineRegisterInfo &MRI = MF.getRegInfo(); | |||
181 | AfterBlockPlacement = AfterPlacement; | |||
182 | TII = tii; | |||
183 | TRI = tri; | |||
184 | MMI = mmi; | |||
185 | MLI = mli; | |||
186 | this->MRI = &MRI; | |||
187 | ||||
188 | UpdateLiveIns = MRI.tracksLiveness() && TRI->trackLivenessAfterRegAlloc(MF); | |||
189 | if (!UpdateLiveIns) | |||
190 | MRI.invalidateLiveness(); | |||
191 | ||||
192 | // Fix CFG. The later algorithms expect it to be right. | |||
193 | bool MadeChange = false; | |||
194 | for (MachineBasicBlock &MBB : MF) { | |||
195 | MachineBasicBlock *TBB = nullptr, *FBB = nullptr; | |||
196 | SmallVector<MachineOperand, 4> Cond; | |||
197 | if (!TII->analyzeBranch(MBB, TBB, FBB, Cond, true)) | |||
198 | MadeChange |= MBB.CorrectExtraCFGEdges(TBB, FBB, !Cond.empty()); | |||
199 | } | |||
200 | ||||
201 | // Recalculate EH scope membership. | |||
202 | EHScopeMembership = getEHScopeMembership(MF); | |||
203 | ||||
204 | bool MadeChangeThisIteration = true; | |||
205 | while (MadeChangeThisIteration) { | |||
206 | MadeChangeThisIteration = TailMergeBlocks(MF); | |||
207 | // No need to clean up if tail merging does not change anything after the | |||
208 | // block placement. | |||
209 | if (!AfterBlockPlacement || MadeChangeThisIteration) | |||
210 | MadeChangeThisIteration |= OptimizeBranches(MF); | |||
211 | if (EnableHoistCommonCode) | |||
212 | MadeChangeThisIteration |= HoistCommonCode(MF); | |||
213 | MadeChange |= MadeChangeThisIteration; | |||
214 | } | |||
215 | ||||
216 | // See if any jump tables have become dead as the code generator | |||
217 | // did its thing. | |||
218 | MachineJumpTableInfo *JTI = MF.getJumpTableInfo(); | |||
219 | if (!JTI) | |||
220 | return MadeChange; | |||
221 | ||||
222 | // Walk the function to find jump tables that are live. | |||
223 | BitVector JTIsLive(JTI->getJumpTables().size()); | |||
224 | for (const MachineBasicBlock &BB : MF) { | |||
225 | for (const MachineInstr &I : BB) | |||
226 | for (const MachineOperand &Op : I.operands()) { | |||
227 | if (!Op.isJTI()) continue; | |||
228 | ||||
229 | // Remember that this JT is live. | |||
230 | JTIsLive.set(Op.getIndex()); | |||
231 | } | |||
232 | } | |||
233 | ||||
234 | // Finally, remove dead jump tables. This happens when the | |||
235 | // indirect jump was unreachable (and thus deleted). | |||
236 | for (unsigned i = 0, e = JTIsLive.size(); i != e; ++i) | |||
237 | if (!JTIsLive.test(i)) { | |||
238 | JTI->RemoveJumpTable(i); | |||
239 | MadeChange = true; | |||
240 | } | |||
241 | ||||
242 | return MadeChange; | |||
243 | } | |||
244 | ||||
245 | //===----------------------------------------------------------------------===// | |||
246 | // Tail Merging of Blocks | |||
247 | //===----------------------------------------------------------------------===// | |||
248 | ||||
249 | /// HashMachineInstr - Compute a hash value for MI and its operands. | |||
250 | static unsigned HashMachineInstr(const MachineInstr &MI) { | |||
251 | unsigned Hash = MI.getOpcode(); | |||
252 | for (unsigned i = 0, e = MI.getNumOperands(); i != e; ++i) { | |||
253 | const MachineOperand &Op = MI.getOperand(i); | |||
254 | ||||
255 | // Merge in bits from the operand if easy. We can't use MachineOperand's | |||
256 | // hash_code here because it's not deterministic and we sort by hash value | |||
257 | // later. | |||
258 | unsigned OperandHash = 0; | |||
259 | switch (Op.getType()) { | |||
260 | case MachineOperand::MO_Register: | |||
261 | OperandHash = Op.getReg(); | |||
262 | break; | |||
263 | case MachineOperand::MO_Immediate: | |||
264 | OperandHash = Op.getImm(); | |||
265 | break; | |||
266 | case MachineOperand::MO_MachineBasicBlock: | |||
267 | OperandHash = Op.getMBB()->getNumber(); | |||
268 | break; | |||
269 | case MachineOperand::MO_FrameIndex: | |||
270 | case MachineOperand::MO_ConstantPoolIndex: | |||
271 | case MachineOperand::MO_JumpTableIndex: | |||
272 | OperandHash = Op.getIndex(); | |||
273 | break; | |||
274 | case MachineOperand::MO_GlobalAddress: | |||
275 | case MachineOperand::MO_ExternalSymbol: | |||
276 | // Global address / external symbol are too hard, don't bother, but do | |||
277 | // pull in the offset. | |||
278 | OperandHash = Op.getOffset(); | |||
279 | break; | |||
280 | default: | |||
281 | break; | |||
282 | } | |||
283 | ||||
284 | Hash += ((OperandHash << 3) | Op.getType()) << (i & 31); | |||
285 | } | |||
286 | return Hash; | |||
287 | } | |||
288 | ||||
289 | /// HashEndOfMBB - Hash the last instruction in the MBB. | |||
290 | static unsigned HashEndOfMBB(const MachineBasicBlock &MBB) { | |||
291 | MachineBasicBlock::const_iterator I = MBB.getLastNonDebugInstr(); | |||
292 | if (I == MBB.end()) | |||
293 | return 0; | |||
294 | ||||
295 | return HashMachineInstr(*I); | |||
296 | } | |||
297 | ||||
298 | /// Whether MI should be counted as an instruction when calculating common tail. | |||
299 | static bool countsAsInstruction(const MachineInstr &MI) { | |||
300 | return !(MI.isDebugInstr() || MI.isCFIInstruction()); | |||
301 | } | |||
302 | ||||
303 | /// ComputeCommonTailLength - Given two machine basic blocks, compute the number | |||
304 | /// of instructions they actually have in common together at their end. Return | |||
305 | /// iterators for the first shared instruction in each block. | |||
306 | static unsigned ComputeCommonTailLength(MachineBasicBlock *MBB1, | |||
307 | MachineBasicBlock *MBB2, | |||
308 | MachineBasicBlock::iterator &I1, | |||
309 | MachineBasicBlock::iterator &I2) { | |||
310 | I1 = MBB1->end(); | |||
311 | I2 = MBB2->end(); | |||
312 | ||||
313 | unsigned TailLen = 0; | |||
314 | while (I1 != MBB1->begin() && I2 != MBB2->begin()) { | |||
315 | --I1; --I2; | |||
316 | // Skip debugging pseudos; necessary to avoid changing the code. | |||
317 | while (!countsAsInstruction(*I1)) { | |||
318 | if (I1==MBB1->begin()) { | |||
319 | while (!countsAsInstruction(*I2)) { | |||
320 | if (I2==MBB2->begin()) { | |||
321 | // I1==DBG at begin; I2==DBG at begin | |||
322 | goto SkipTopCFIAndReturn; | |||
323 | } | |||
324 | --I2; | |||
325 | } | |||
326 | ++I2; | |||
327 | // I1==DBG at begin; I2==non-DBG, or first of DBGs not at begin | |||
328 | goto SkipTopCFIAndReturn; | |||
329 | } | |||
330 | --I1; | |||
331 | } | |||
332 | // I1==first (untested) non-DBG preceding known match | |||
333 | while (!countsAsInstruction(*I2)) { | |||
334 | if (I2==MBB2->begin()) { | |||
335 | ++I1; | |||
336 | // I1==non-DBG, or first of DBGs not at begin; I2==DBG at begin | |||
337 | goto SkipTopCFIAndReturn; | |||
338 | } | |||
339 | --I2; | |||
340 | } | |||
341 | // I1, I2==first (untested) non-DBGs preceding known match | |||
342 | if (!I1->isIdenticalTo(*I2) || | |||
343 | // FIXME: This check is dubious. It's used to get around a problem where | |||
344 | // people incorrectly expect inline asm directives to remain in the same | |||
345 | // relative order. This is untenable because normal compiler | |||
346 | // optimizations (like this one) may reorder and/or merge these | |||
347 | // directives. | |||
348 | I1->isInlineAsm()) { | |||
349 | ++I1; ++I2; | |||
350 | break; | |||
351 | } | |||
352 | ++TailLen; | |||
353 | } | |||
354 | // Back past possible debugging pseudos at beginning of block. This matters | |||
355 | // when one block differs from the other only by whether debugging pseudos | |||
356 | // are present at the beginning. (This way, the various checks later for | |||
357 | // I1==MBB1->begin() work as expected.) | |||
358 | if (I1 == MBB1->begin() && I2 != MBB2->begin()) { | |||
359 | --I2; | |||
360 | while (I2->isDebugInstr()) { | |||
361 | if (I2 == MBB2->begin()) | |||
362 | return TailLen; | |||
363 | --I2; | |||
364 | } | |||
365 | ++I2; | |||
366 | } | |||
367 | if (I2 == MBB2->begin() && I1 != MBB1->begin()) { | |||
368 | --I1; | |||
369 | while (I1->isDebugInstr()) { | |||
370 | if (I1 == MBB1->begin()) | |||
371 | return TailLen; | |||
372 | --I1; | |||
373 | } | |||
374 | ++I1; | |||
375 | } | |||
376 | ||||
377 | SkipTopCFIAndReturn: | |||
378 | // Ensure that I1 and I2 do not point to a CFI_INSTRUCTION. This can happen if | |||
379 | // I1 and I2 are non-identical when compared and then one or both of them ends | |||
380 | // up pointing to a CFI instruction after being incremented. For example: | |||
381 | /* | |||
382 | BB1: | |||
383 | ... | |||
384 | INSTRUCTION_A | |||
385 | ADD32ri8 <- last common instruction | |||
386 | ... | |||
387 | BB2: | |||
388 | ... | |||
389 | INSTRUCTION_B | |||
390 | CFI_INSTRUCTION | |||
391 | ADD32ri8 <- last common instruction | |||
392 | ... | |||
393 | */ | |||
394 | // When INSTRUCTION_A and INSTRUCTION_B are compared as not equal, after | |||
395 | // incrementing the iterators, I1 will point to ADD, however I2 will point to | |||
396 | // the CFI instruction. Later on, this leads to BB2 being 'hacked off' at the | |||
397 | // wrong place (in ReplaceTailWithBranchTo()) which results in losing this CFI | |||
398 | // instruction. | |||
399 | while (I1 != MBB1->end() && I1->isCFIInstruction()) { | |||
400 | ++I1; | |||
401 | } | |||
402 | ||||
403 | while (I2 != MBB2->end() && I2->isCFIInstruction()) { | |||
404 | ++I2; | |||
405 | } | |||
406 | ||||
407 | return TailLen; | |||
408 | } | |||
409 | ||||
410 | void BranchFolder::replaceTailWithBranchTo(MachineBasicBlock::iterator OldInst, | |||
411 | MachineBasicBlock &NewDest) { | |||
412 | if (UpdateLiveIns) { | |||
413 | // OldInst should always point to an instruction. | |||
414 | MachineBasicBlock &OldMBB = *OldInst->getParent(); | |||
415 | LiveRegs.clear(); | |||
416 | LiveRegs.addLiveOuts(OldMBB); | |||
417 | // Move backward to the place where will insert the jump. | |||
418 | MachineBasicBlock::iterator I = OldMBB.end(); | |||
419 | do { | |||
420 | --I; | |||
421 | LiveRegs.stepBackward(*I); | |||
422 | } while (I != OldInst); | |||
423 | ||||
424 | // Merging the tails may have switched some undef operand to non-undef ones. | |||
425 | // Add IMPLICIT_DEFS into OldMBB as necessary to have a definition of the | |||
426 | // register. | |||
427 | for (MachineBasicBlock::RegisterMaskPair P : NewDest.liveins()) { | |||
428 | // We computed the liveins with computeLiveIn earlier and should only see | |||
429 | // full registers: | |||
430 | assert(P.LaneMask == LaneBitmask::getAll() &&((P.LaneMask == LaneBitmask::getAll() && "Can only handle full register." ) ? static_cast<void> (0) : __assert_fail ("P.LaneMask == LaneBitmask::getAll() && \"Can only handle full register.\"" , "/build/llvm-toolchain-snapshot-9~svn362543/lib/CodeGen/BranchFolding.cpp" , 431, __PRETTY_FUNCTION__)) | |||
431 | "Can only handle full register.")((P.LaneMask == LaneBitmask::getAll() && "Can only handle full register." ) ? static_cast<void> (0) : __assert_fail ("P.LaneMask == LaneBitmask::getAll() && \"Can only handle full register.\"" , "/build/llvm-toolchain-snapshot-9~svn362543/lib/CodeGen/BranchFolding.cpp" , 431, __PRETTY_FUNCTION__)); | |||
432 | MCPhysReg Reg = P.PhysReg; | |||
433 | if (!LiveRegs.available(*MRI, Reg)) | |||
434 | continue; | |||
435 | DebugLoc DL; | |||
436 | BuildMI(OldMBB, OldInst, DL, TII->get(TargetOpcode::IMPLICIT_DEF), Reg); | |||
437 | } | |||
438 | } | |||
439 | ||||
440 | TII->ReplaceTailWithBranchTo(OldInst, &NewDest); | |||
441 | ++NumTailMerge; | |||
442 | } | |||
443 | ||||
444 | MachineBasicBlock *BranchFolder::SplitMBBAt(MachineBasicBlock &CurMBB, | |||
445 | MachineBasicBlock::iterator BBI1, | |||
446 | const BasicBlock *BB) { | |||
447 | if (!TII->isLegalToSplitMBBAt(CurMBB, BBI1)) | |||
448 | return nullptr; | |||
449 | ||||
450 | MachineFunction &MF = *CurMBB.getParent(); | |||
451 | ||||
452 | // Create the fall-through block. | |||
453 | MachineFunction::iterator MBBI = CurMBB.getIterator(); | |||
454 | MachineBasicBlock *NewMBB = MF.CreateMachineBasicBlock(BB); | |||
455 | CurMBB.getParent()->insert(++MBBI, NewMBB); | |||
456 | ||||
457 | // Move all the successors of this block to the specified block. | |||
458 | NewMBB->transferSuccessors(&CurMBB); | |||
459 | ||||
460 | // Add an edge from CurMBB to NewMBB for the fall-through. | |||
461 | CurMBB.addSuccessor(NewMBB); | |||
462 | ||||
463 | // Splice the code over. | |||
464 | NewMBB->splice(NewMBB->end(), &CurMBB, BBI1, CurMBB.end()); | |||
465 | ||||
466 | // NewMBB belongs to the same loop as CurMBB. | |||
467 | if (MLI) | |||
468 | if (MachineLoop *ML = MLI->getLoopFor(&CurMBB)) | |||
469 | ML->addBasicBlockToLoop(NewMBB, MLI->getBase()); | |||
470 | ||||
471 | // NewMBB inherits CurMBB's block frequency. | |||
472 | MBBFreqInfo.setBlockFreq(NewMBB, MBBFreqInfo.getBlockFreq(&CurMBB)); | |||
473 | ||||
474 | if (UpdateLiveIns) | |||
475 | computeAndAddLiveIns(LiveRegs, *NewMBB); | |||
476 | ||||
477 | // Add the new block to the EH scope. | |||
478 | const auto &EHScopeI = EHScopeMembership.find(&CurMBB); | |||
479 | if (EHScopeI != EHScopeMembership.end()) { | |||
480 | auto n = EHScopeI->second; | |||
481 | EHScopeMembership[NewMBB] = n; | |||
482 | } | |||
483 | ||||
484 | return NewMBB; | |||
485 | } | |||
486 | ||||
487 | /// EstimateRuntime - Make a rough estimate for how long it will take to run | |||
488 | /// the specified code. | |||
489 | static unsigned EstimateRuntime(MachineBasicBlock::iterator I, | |||
490 | MachineBasicBlock::iterator E) { | |||
491 | unsigned Time = 0; | |||
492 | for (; I != E; ++I) { | |||
493 | if (!countsAsInstruction(*I)) | |||
494 | continue; | |||
495 | if (I->isCall()) | |||
496 | Time += 10; | |||
497 | else if (I->mayLoad() || I->mayStore()) | |||
498 | Time += 2; | |||
499 | else | |||
500 | ++Time; | |||
501 | } | |||
502 | return Time; | |||
503 | } | |||
504 | ||||
505 | // CurMBB needs to add an unconditional branch to SuccMBB (we removed these | |||
506 | // branches temporarily for tail merging). In the case where CurMBB ends | |||
507 | // with a conditional branch to the next block, optimize by reversing the | |||
508 | // test and conditionally branching to SuccMBB instead. | |||
509 | static void FixTail(MachineBasicBlock *CurMBB, MachineBasicBlock *SuccBB, | |||
510 | const TargetInstrInfo *TII) { | |||
511 | MachineFunction *MF = CurMBB->getParent(); | |||
512 | MachineFunction::iterator I = std::next(MachineFunction::iterator(CurMBB)); | |||
513 | MachineBasicBlock *TBB = nullptr, *FBB = nullptr; | |||
514 | SmallVector<MachineOperand, 4> Cond; | |||
515 | DebugLoc dl = CurMBB->findBranchDebugLoc(); | |||
516 | if (I != MF->end() && !TII->analyzeBranch(*CurMBB, TBB, FBB, Cond, true)) { | |||
517 | MachineBasicBlock *NextBB = &*I; | |||
518 | if (TBB == NextBB && !Cond.empty() && !FBB) { | |||
519 | if (!TII->reverseBranchCondition(Cond)) { | |||
520 | TII->removeBranch(*CurMBB); | |||
521 | TII->insertBranch(*CurMBB, SuccBB, nullptr, Cond, dl); | |||
522 | return; | |||
523 | } | |||
524 | } | |||
525 | } | |||
526 | TII->insertBranch(*CurMBB, SuccBB, nullptr, | |||
527 | SmallVector<MachineOperand, 0>(), dl); | |||
528 | } | |||
529 | ||||
530 | bool | |||
531 | BranchFolder::MergePotentialsElt::operator<(const MergePotentialsElt &o) const { | |||
532 | if (getHash() < o.getHash()) | |||
533 | return true; | |||
534 | if (getHash() > o.getHash()) | |||
535 | return false; | |||
536 | if (getBlock()->getNumber() < o.getBlock()->getNumber()) | |||
537 | return true; | |||
538 | if (getBlock()->getNumber() > o.getBlock()->getNumber()) | |||
539 | return false; | |||
540 | // _GLIBCXX_DEBUG checks strict weak ordering, which involves comparing | |||
541 | // an object with itself. | |||
542 | #ifndef _GLIBCXX_DEBUG | |||
543 | llvm_unreachable("Predecessor appears twice")::llvm::llvm_unreachable_internal("Predecessor appears twice" , "/build/llvm-toolchain-snapshot-9~svn362543/lib/CodeGen/BranchFolding.cpp" , 543); | |||
544 | #else | |||
545 | return false; | |||
546 | #endif | |||
547 | } | |||
548 | ||||
549 | BlockFrequency | |||
550 | BranchFolder::MBFIWrapper::getBlockFreq(const MachineBasicBlock *MBB) const { | |||
551 | auto I = MergedBBFreq.find(MBB); | |||
552 | ||||
553 | if (I != MergedBBFreq.end()) | |||
554 | return I->second; | |||
555 | ||||
556 | return MBFI.getBlockFreq(MBB); | |||
557 | } | |||
558 | ||||
559 | void BranchFolder::MBFIWrapper::setBlockFreq(const MachineBasicBlock *MBB, | |||
560 | BlockFrequency F) { | |||
561 | MergedBBFreq[MBB] = F; | |||
562 | } | |||
563 | ||||
564 | raw_ostream & | |||
565 | BranchFolder::MBFIWrapper::printBlockFreq(raw_ostream &OS, | |||
566 | const MachineBasicBlock *MBB) const { | |||
567 | return MBFI.printBlockFreq(OS, getBlockFreq(MBB)); | |||
568 | } | |||
569 | ||||
570 | raw_ostream & | |||
571 | BranchFolder::MBFIWrapper::printBlockFreq(raw_ostream &OS, | |||
572 | const BlockFrequency Freq) const { | |||
573 | return MBFI.printBlockFreq(OS, Freq); | |||
574 | } | |||
575 | ||||
576 | void BranchFolder::MBFIWrapper::view(const Twine &Name, bool isSimple) { | |||
577 | MBFI.view(Name, isSimple); | |||
578 | } | |||
579 | ||||
580 | uint64_t | |||
581 | BranchFolder::MBFIWrapper::getEntryFreq() const { | |||
582 | return MBFI.getEntryFreq(); | |||
583 | } | |||
584 | ||||
585 | /// CountTerminators - Count the number of terminators in the given | |||
586 | /// block and set I to the position of the first non-terminator, if there | |||
587 | /// is one, or MBB->end() otherwise. | |||
588 | static unsigned CountTerminators(MachineBasicBlock *MBB, | |||
589 | MachineBasicBlock::iterator &I) { | |||
590 | I = MBB->end(); | |||
591 | unsigned NumTerms = 0; | |||
592 | while (true) { | |||
593 | if (I == MBB->begin()) { | |||
594 | I = MBB->end(); | |||
595 | break; | |||
596 | } | |||
597 | --I; | |||
598 | if (!I->isTerminator()) break; | |||
599 | ++NumTerms; | |||
600 | } | |||
601 | return NumTerms; | |||
602 | } | |||
603 | ||||
604 | /// A no successor, non-return block probably ends in unreachable and is cold. | |||
605 | /// Also consider a block that ends in an indirect branch to be a return block, | |||
606 | /// since many targets use plain indirect branches to return. | |||
607 | static bool blockEndsInUnreachable(const MachineBasicBlock *MBB) { | |||
608 | if (!MBB->succ_empty()) | |||
609 | return false; | |||
610 | if (MBB->empty()) | |||
611 | return true; | |||
612 | return !(MBB->back().isReturn() || MBB->back().isIndirectBranch()); | |||
613 | } | |||
614 | ||||
615 | /// ProfitableToMerge - Check if two machine basic blocks have a common tail | |||
616 | /// and decide if it would be profitable to merge those tails. Return the | |||
617 | /// length of the common tail and iterators to the first common instruction | |||
618 | /// in each block. | |||
619 | /// MBB1, MBB2 The blocks to check | |||
620 | /// MinCommonTailLength Minimum size of tail block to be merged. | |||
621 | /// CommonTailLen Out parameter to record the size of the shared tail between | |||
622 | /// MBB1 and MBB2 | |||
623 | /// I1, I2 Iterator references that will be changed to point to the first | |||
624 | /// instruction in the common tail shared by MBB1,MBB2 | |||
625 | /// SuccBB A common successor of MBB1, MBB2 which are in a canonical form | |||
626 | /// relative to SuccBB | |||
627 | /// PredBB The layout predecessor of SuccBB, if any. | |||
628 | /// EHScopeMembership map from block to EH scope #. | |||
629 | /// AfterPlacement True if we are merging blocks after layout. Stricter | |||
630 | /// thresholds apply to prevent undoing tail-duplication. | |||
631 | static bool | |||
632 | ProfitableToMerge(MachineBasicBlock *MBB1, MachineBasicBlock *MBB2, | |||
633 | unsigned MinCommonTailLength, unsigned &CommonTailLen, | |||
634 | MachineBasicBlock::iterator &I1, | |||
635 | MachineBasicBlock::iterator &I2, MachineBasicBlock *SuccBB, | |||
636 | MachineBasicBlock *PredBB, | |||
637 | DenseMap<const MachineBasicBlock *, int> &EHScopeMembership, | |||
638 | bool AfterPlacement) { | |||
639 | // It is never profitable to tail-merge blocks from two different EH scopes. | |||
640 | if (!EHScopeMembership.empty()) { | |||
641 | auto EHScope1 = EHScopeMembership.find(MBB1); | |||
642 | assert(EHScope1 != EHScopeMembership.end())((EHScope1 != EHScopeMembership.end()) ? static_cast<void> (0) : __assert_fail ("EHScope1 != EHScopeMembership.end()", "/build/llvm-toolchain-snapshot-9~svn362543/lib/CodeGen/BranchFolding.cpp" , 642, __PRETTY_FUNCTION__)); | |||
643 | auto EHScope2 = EHScopeMembership.find(MBB2); | |||
644 | assert(EHScope2 != EHScopeMembership.end())((EHScope2 != EHScopeMembership.end()) ? static_cast<void> (0) : __assert_fail ("EHScope2 != EHScopeMembership.end()", "/build/llvm-toolchain-snapshot-9~svn362543/lib/CodeGen/BranchFolding.cpp" , 644, __PRETTY_FUNCTION__)); | |||
645 | if (EHScope1->second != EHScope2->second) | |||
646 | return false; | |||
647 | } | |||
648 | ||||
649 | CommonTailLen = ComputeCommonTailLength(MBB1, MBB2, I1, I2); | |||
650 | if (CommonTailLen == 0) | |||
651 | return false; | |||
652 | LLVM_DEBUG(dbgs() << "Common tail length of " << printMBBReference(*MBB1)do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("branch-folder")) { dbgs() << "Common tail length of " << printMBBReference(*MBB1) << " and " << printMBBReference (*MBB2) << " is " << CommonTailLen << '\n'; } } while (false) | |||
653 | << " and " << printMBBReference(*MBB2) << " is "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("branch-folder")) { dbgs() << "Common tail length of " << printMBBReference(*MBB1) << " and " << printMBBReference (*MBB2) << " is " << CommonTailLen << '\n'; } } while (false) | |||
654 | << CommonTailLen << '\n')do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("branch-folder")) { dbgs() << "Common tail length of " << printMBBReference(*MBB1) << " and " << printMBBReference (*MBB2) << " is " << CommonTailLen << '\n'; } } while (false); | |||
655 | ||||
656 | // It's almost always profitable to merge any number of non-terminator | |||
657 | // instructions with the block that falls through into the common successor. | |||
658 | // This is true only for a single successor. For multiple successors, we are | |||
659 | // trading a conditional branch for an unconditional one. | |||
660 | // TODO: Re-visit successor size for non-layout tail merging. | |||
661 | if ((MBB1 == PredBB || MBB2 == PredBB) && | |||
662 | (!AfterPlacement || MBB1->succ_size() == 1)) { | |||
663 | MachineBasicBlock::iterator I; | |||
664 | unsigned NumTerms = CountTerminators(MBB1 == PredBB ? MBB2 : MBB1, I); | |||
665 | if (CommonTailLen > NumTerms) | |||
666 | return true; | |||
667 | } | |||
668 | ||||
669 | // If these are identical non-return blocks with no successors, merge them. | |||
670 | // Such blocks are typically cold calls to noreturn functions like abort, and | |||
671 | // are unlikely to become a fallthrough target after machine block placement. | |||
672 | // Tail merging these blocks is unlikely to create additional unconditional | |||
673 | // branches, and will reduce the size of this cold code. | |||
674 | if (I1 == MBB1->begin() && I2 == MBB2->begin() && | |||
675 | blockEndsInUnreachable(MBB1) && blockEndsInUnreachable(MBB2)) | |||
676 | return true; | |||
677 | ||||
678 | // If one of the blocks can be completely merged and happens to be in | |||
679 | // a position where the other could fall through into it, merge any number | |||
680 | // of instructions, because it can be done without a branch. | |||
681 | // TODO: If the blocks are not adjacent, move one of them so that they are? | |||
682 | if (MBB1->isLayoutSuccessor(MBB2) && I2 == MBB2->begin()) | |||
683 | return true; | |||
684 | if (MBB2->isLayoutSuccessor(MBB1) && I1 == MBB1->begin()) | |||
685 | return true; | |||
686 | ||||
687 | // If both blocks are identical and end in a branch, merge them unless they | |||
688 | // both have a fallthrough predecessor and successor. | |||
689 | // We can only do this after block placement because it depends on whether | |||
690 | // there are fallthroughs, and we don't know until after layout. | |||
691 | if (AfterPlacement && I1 == MBB1->begin() && I2 == MBB2->begin()) { | |||
692 | auto BothFallThrough = [](MachineBasicBlock *MBB) { | |||
693 | if (MBB->succ_size() != 0 && !MBB->canFallThrough()) | |||
694 | return false; | |||
695 | MachineFunction::iterator I(MBB); | |||
696 | MachineFunction *MF = MBB->getParent(); | |||
697 | return (MBB != &*MF->begin()) && std::prev(I)->canFallThrough(); | |||
698 | }; | |||
699 | if (!BothFallThrough(MBB1) || !BothFallThrough(MBB2)) | |||
700 | return true; | |||
701 | } | |||
702 | ||||
703 | // If both blocks have an unconditional branch temporarily stripped out, | |||
704 | // count that as an additional common instruction for the following | |||
705 | // heuristics. This heuristic is only accurate for single-succ blocks, so to | |||
706 | // make sure that during layout merging and duplicating don't crash, we check | |||
707 | // for that when merging during layout. | |||
708 | unsigned EffectiveTailLen = CommonTailLen; | |||
709 | if (SuccBB && MBB1 != PredBB && MBB2 != PredBB && | |||
710 | (MBB1->succ_size() == 1 || !AfterPlacement) && | |||
711 | !MBB1->back().isBarrier() && | |||
712 | !MBB2->back().isBarrier()) | |||
713 | ++EffectiveTailLen; | |||
714 | ||||
715 | // Check if the common tail is long enough to be worthwhile. | |||
716 | if (EffectiveTailLen >= MinCommonTailLength) | |||
717 | return true; | |||
718 | ||||
719 | // If we are optimizing for code size, 2 instructions in common is enough if | |||
720 | // we don't have to split a block. At worst we will be introducing 1 new | |||
721 | // branch instruction, which is likely to be smaller than the 2 | |||
722 | // instructions that would be deleted in the merge. | |||
723 | MachineFunction *MF = MBB1->getParent(); | |||
724 | return EffectiveTailLen >= 2 && MF->getFunction().hasOptSize() && | |||
725 | (I1 == MBB1->begin() || I2 == MBB2->begin()); | |||
726 | } | |||
727 | ||||
728 | unsigned BranchFolder::ComputeSameTails(unsigned CurHash, | |||
729 | unsigned MinCommonTailLength, | |||
730 | MachineBasicBlock *SuccBB, | |||
731 | MachineBasicBlock *PredBB) { | |||
732 | unsigned maxCommonTailLength = 0U; | |||
733 | SameTails.clear(); | |||
734 | MachineBasicBlock::iterator TrialBBI1, TrialBBI2; | |||
735 | MPIterator HighestMPIter = std::prev(MergePotentials.end()); | |||
736 | for (MPIterator CurMPIter = std::prev(MergePotentials.end()), | |||
737 | B = MergePotentials.begin(); | |||
738 | CurMPIter != B && CurMPIter->getHash() == CurHash; --CurMPIter) { | |||
739 | for (MPIterator I = std::prev(CurMPIter); I->getHash() == CurHash; --I) { | |||
740 | unsigned CommonTailLen; | |||
741 | if (ProfitableToMerge(CurMPIter->getBlock(), I->getBlock(), | |||
742 | MinCommonTailLength, | |||
743 | CommonTailLen, TrialBBI1, TrialBBI2, | |||
744 | SuccBB, PredBB, | |||
745 | EHScopeMembership, | |||
746 | AfterBlockPlacement)) { | |||
747 | if (CommonTailLen > maxCommonTailLength) { | |||
748 | SameTails.clear(); | |||
749 | maxCommonTailLength = CommonTailLen; | |||
750 | HighestMPIter = CurMPIter; | |||
751 | SameTails.push_back(SameTailElt(CurMPIter, TrialBBI1)); | |||
752 | } | |||
753 | if (HighestMPIter == CurMPIter && | |||
754 | CommonTailLen == maxCommonTailLength) | |||
755 | SameTails.push_back(SameTailElt(I, TrialBBI2)); | |||
756 | } | |||
757 | if (I == B) | |||
758 | break; | |||
759 | } | |||
760 | } | |||
761 | return maxCommonTailLength; | |||
762 | } | |||
763 | ||||
764 | void BranchFolder::RemoveBlocksWithHash(unsigned CurHash, | |||
765 | MachineBasicBlock *SuccBB, | |||
766 | MachineBasicBlock *PredBB) { | |||
767 | MPIterator CurMPIter, B; | |||
768 | for (CurMPIter = std::prev(MergePotentials.end()), | |||
769 | B = MergePotentials.begin(); | |||
770 | CurMPIter->getHash() == CurHash; --CurMPIter) { | |||
771 | // Put the unconditional branch back, if we need one. | |||
772 | MachineBasicBlock *CurMBB = CurMPIter->getBlock(); | |||
773 | if (SuccBB && CurMBB != PredBB) | |||
774 | FixTail(CurMBB, SuccBB, TII); | |||
775 | if (CurMPIter == B) | |||
776 | break; | |||
777 | } | |||
778 | if (CurMPIter->getHash() != CurHash) | |||
779 | CurMPIter++; | |||
780 | MergePotentials.erase(CurMPIter, MergePotentials.end()); | |||
781 | } | |||
782 | ||||
783 | bool BranchFolder::CreateCommonTailOnlyBlock(MachineBasicBlock *&PredBB, | |||
784 | MachineBasicBlock *SuccBB, | |||
785 | unsigned maxCommonTailLength, | |||
786 | unsigned &commonTailIndex) { | |||
787 | commonTailIndex = 0; | |||
788 | unsigned TimeEstimate = ~0U; | |||
789 | for (unsigned i = 0, e = SameTails.size(); i != e; ++i) { | |||
790 | // Use PredBB if possible; that doesn't require a new branch. | |||
791 | if (SameTails[i].getBlock() == PredBB) { | |||
792 | commonTailIndex = i; | |||
793 | break; | |||
794 | } | |||
795 | // Otherwise, make a (fairly bogus) choice based on estimate of | |||
796 | // how long it will take the various blocks to execute. | |||
797 | unsigned t = EstimateRuntime(SameTails[i].getBlock()->begin(), | |||
798 | SameTails[i].getTailStartPos()); | |||
799 | if (t <= TimeEstimate) { | |||
800 | TimeEstimate = t; | |||
801 | commonTailIndex = i; | |||
802 | } | |||
803 | } | |||
804 | ||||
805 | MachineBasicBlock::iterator BBI = | |||
806 | SameTails[commonTailIndex].getTailStartPos(); | |||
807 | MachineBasicBlock *MBB = SameTails[commonTailIndex].getBlock(); | |||
808 | ||||
809 | LLVM_DEBUG(dbgs() << "\nSplitting " << printMBBReference(*MBB) << ", size "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("branch-folder")) { dbgs() << "\nSplitting " << printMBBReference (*MBB) << ", size " << maxCommonTailLength; } } while (false) | |||
810 | << maxCommonTailLength)do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("branch-folder")) { dbgs() << "\nSplitting " << printMBBReference (*MBB) << ", size " << maxCommonTailLength; } } while (false); | |||
811 | ||||
812 | // If the split block unconditionally falls-thru to SuccBB, it will be | |||
813 | // merged. In control flow terms it should then take SuccBB's name. e.g. If | |||
814 | // SuccBB is an inner loop, the common tail is still part of the inner loop. | |||
815 | const BasicBlock *BB = (SuccBB && MBB->succ_size() == 1) ? | |||
816 | SuccBB->getBasicBlock() : MBB->getBasicBlock(); | |||
817 | MachineBasicBlock *newMBB = SplitMBBAt(*MBB, BBI, BB); | |||
818 | if (!newMBB) { | |||
819 | LLVM_DEBUG(dbgs() << "... failed!")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("branch-folder")) { dbgs() << "... failed!"; } } while (false); | |||
820 | return false; | |||
821 | } | |||
822 | ||||
823 | SameTails[commonTailIndex].setBlock(newMBB); | |||
824 | SameTails[commonTailIndex].setTailStartPos(newMBB->begin()); | |||
825 | ||||
826 | // If we split PredBB, newMBB is the new predecessor. | |||
827 | if (PredBB == MBB) | |||
828 | PredBB = newMBB; | |||
829 | ||||
830 | return true; | |||
831 | } | |||
832 | ||||
833 | static void | |||
834 | mergeOperations(MachineBasicBlock::iterator MBBIStartPos, | |||
835 | MachineBasicBlock &MBBCommon) { | |||
836 | MachineBasicBlock *MBB = MBBIStartPos->getParent(); | |||
837 | // Note CommonTailLen does not necessarily matches the size of | |||
838 | // the common BB nor all its instructions because of debug | |||
839 | // instructions differences. | |||
840 | unsigned CommonTailLen = 0; | |||
841 | for (auto E = MBB->end(); MBBIStartPos != E; ++MBBIStartPos) | |||
842 | ++CommonTailLen; | |||
843 | ||||
844 | MachineBasicBlock::reverse_iterator MBBI = MBB->rbegin(); | |||
845 | MachineBasicBlock::reverse_iterator MBBIE = MBB->rend(); | |||
846 | MachineBasicBlock::reverse_iterator MBBICommon = MBBCommon.rbegin(); | |||
847 | MachineBasicBlock::reverse_iterator MBBIECommon = MBBCommon.rend(); | |||
848 | ||||
849 | while (CommonTailLen--) { | |||
850 | assert(MBBI != MBBIE && "Reached BB end within common tail length!")((MBBI != MBBIE && "Reached BB end within common tail length!" ) ? static_cast<void> (0) : __assert_fail ("MBBI != MBBIE && \"Reached BB end within common tail length!\"" , "/build/llvm-toolchain-snapshot-9~svn362543/lib/CodeGen/BranchFolding.cpp" , 850, __PRETTY_FUNCTION__)); | |||
851 | (void)MBBIE; | |||
852 | ||||
853 | if (!countsAsInstruction(*MBBI)) { | |||
854 | ++MBBI; | |||
855 | continue; | |||
856 | } | |||
857 | ||||
858 | while ((MBBICommon != MBBIECommon) && !countsAsInstruction(*MBBICommon)) | |||
859 | ++MBBICommon; | |||
860 | ||||
861 | assert(MBBICommon != MBBIECommon &&((MBBICommon != MBBIECommon && "Reached BB end within common tail length!" ) ? static_cast<void> (0) : __assert_fail ("MBBICommon != MBBIECommon && \"Reached BB end within common tail length!\"" , "/build/llvm-toolchain-snapshot-9~svn362543/lib/CodeGen/BranchFolding.cpp" , 862, __PRETTY_FUNCTION__)) | |||
862 | "Reached BB end within common tail length!")((MBBICommon != MBBIECommon && "Reached BB end within common tail length!" ) ? static_cast<void> (0) : __assert_fail ("MBBICommon != MBBIECommon && \"Reached BB end within common tail length!\"" , "/build/llvm-toolchain-snapshot-9~svn362543/lib/CodeGen/BranchFolding.cpp" , 862, __PRETTY_FUNCTION__)); | |||
863 | assert(MBBICommon->isIdenticalTo(*MBBI) && "Expected matching MIIs!")((MBBICommon->isIdenticalTo(*MBBI) && "Expected matching MIIs!" ) ? static_cast<void> (0) : __assert_fail ("MBBICommon->isIdenticalTo(*MBBI) && \"Expected matching MIIs!\"" , "/build/llvm-toolchain-snapshot-9~svn362543/lib/CodeGen/BranchFolding.cpp" , 863, __PRETTY_FUNCTION__)); | |||
864 | ||||
865 | // Merge MMOs from memory operations in the common block. | |||
866 | if (MBBICommon->mayLoad() || MBBICommon->mayStore()) | |||
867 | MBBICommon->cloneMergedMemRefs(*MBB->getParent(), {&*MBBICommon, &*MBBI}); | |||
868 | // Drop undef flags if they aren't present in all merged instructions. | |||
869 | for (unsigned I = 0, E = MBBICommon->getNumOperands(); I != E; ++I) { | |||
870 | MachineOperand &MO = MBBICommon->getOperand(I); | |||
871 | if (MO.isReg() && MO.isUndef()) { | |||
872 | const MachineOperand &OtherMO = MBBI->getOperand(I); | |||
873 | if (!OtherMO.isUndef()) | |||
874 | MO.setIsUndef(false); | |||
875 | } | |||
876 | } | |||
877 | ||||
878 | ++MBBI; | |||
879 | ++MBBICommon; | |||
880 | } | |||
881 | } | |||
882 | ||||
883 | void BranchFolder::mergeCommonTails(unsigned commonTailIndex) { | |||
884 | MachineBasicBlock *MBB = SameTails[commonTailIndex].getBlock(); | |||
885 | ||||
886 | std::vector<MachineBasicBlock::iterator> NextCommonInsts(SameTails.size()); | |||
887 | for (unsigned int i = 0 ; i != SameTails.size() ; ++i) { | |||
888 | if (i != commonTailIndex) { | |||
889 | NextCommonInsts[i] = SameTails[i].getTailStartPos(); | |||
890 | mergeOperations(SameTails[i].getTailStartPos(), *MBB); | |||
891 | } else { | |||
892 | assert(SameTails[i].getTailStartPos() == MBB->begin() &&((SameTails[i].getTailStartPos() == MBB->begin() && "MBB is not a common tail only block") ? static_cast<void > (0) : __assert_fail ("SameTails[i].getTailStartPos() == MBB->begin() && \"MBB is not a common tail only block\"" , "/build/llvm-toolchain-snapshot-9~svn362543/lib/CodeGen/BranchFolding.cpp" , 893, __PRETTY_FUNCTION__)) | |||
893 | "MBB is not a common tail only block")((SameTails[i].getTailStartPos() == MBB->begin() && "MBB is not a common tail only block") ? static_cast<void > (0) : __assert_fail ("SameTails[i].getTailStartPos() == MBB->begin() && \"MBB is not a common tail only block\"" , "/build/llvm-toolchain-snapshot-9~svn362543/lib/CodeGen/BranchFolding.cpp" , 893, __PRETTY_FUNCTION__)); | |||
894 | } | |||
895 | } | |||
896 | ||||
897 | for (auto &MI : *MBB) { | |||
898 | if (!countsAsInstruction(MI)) | |||
899 | continue; | |||
900 | DebugLoc DL = MI.getDebugLoc(); | |||
901 | for (unsigned int i = 0 ; i < NextCommonInsts.size() ; i++) { | |||
902 | if (i == commonTailIndex) | |||
903 | continue; | |||
904 | ||||
905 | auto &Pos = NextCommonInsts[i]; | |||
906 | assert(Pos != SameTails[i].getBlock()->end() &&((Pos != SameTails[i].getBlock()->end() && "Reached BB end within common tail" ) ? static_cast<void> (0) : __assert_fail ("Pos != SameTails[i].getBlock()->end() && \"Reached BB end within common tail\"" , "/build/llvm-toolchain-snapshot-9~svn362543/lib/CodeGen/BranchFolding.cpp" , 907, __PRETTY_FUNCTION__)) | |||
907 | "Reached BB end within common tail")((Pos != SameTails[i].getBlock()->end() && "Reached BB end within common tail" ) ? static_cast<void> (0) : __assert_fail ("Pos != SameTails[i].getBlock()->end() && \"Reached BB end within common tail\"" , "/build/llvm-toolchain-snapshot-9~svn362543/lib/CodeGen/BranchFolding.cpp" , 907, __PRETTY_FUNCTION__)); | |||
908 | while (!countsAsInstruction(*Pos)) { | |||
909 | ++Pos; | |||
910 | assert(Pos != SameTails[i].getBlock()->end() &&((Pos != SameTails[i].getBlock()->end() && "Reached BB end within common tail" ) ? static_cast<void> (0) : __assert_fail ("Pos != SameTails[i].getBlock()->end() && \"Reached BB end within common tail\"" , "/build/llvm-toolchain-snapshot-9~svn362543/lib/CodeGen/BranchFolding.cpp" , 911, __PRETTY_FUNCTION__)) | |||
911 | "Reached BB end within common tail")((Pos != SameTails[i].getBlock()->end() && "Reached BB end within common tail" ) ? static_cast<void> (0) : __assert_fail ("Pos != SameTails[i].getBlock()->end() && \"Reached BB end within common tail\"" , "/build/llvm-toolchain-snapshot-9~svn362543/lib/CodeGen/BranchFolding.cpp" , 911, __PRETTY_FUNCTION__)); | |||
912 | } | |||
913 | assert(MI.isIdenticalTo(*Pos) && "Expected matching MIIs!")((MI.isIdenticalTo(*Pos) && "Expected matching MIIs!" ) ? static_cast<void> (0) : __assert_fail ("MI.isIdenticalTo(*Pos) && \"Expected matching MIIs!\"" , "/build/llvm-toolchain-snapshot-9~svn362543/lib/CodeGen/BranchFolding.cpp" , 913, __PRETTY_FUNCTION__)); | |||
914 | DL = DILocation::getMergedLocation(DL, Pos->getDebugLoc()); | |||
915 | NextCommonInsts[i] = ++Pos; | |||
916 | } | |||
917 | MI.setDebugLoc(DL); | |||
918 | } | |||
919 | ||||
920 | if (UpdateLiveIns) { | |||
921 | LivePhysRegs NewLiveIns(*TRI); | |||
922 | computeLiveIns(NewLiveIns, *MBB); | |||
923 | LiveRegs.init(*TRI); | |||
924 | ||||
925 | // The flag merging may lead to some register uses no longer using the | |||
926 | // <undef> flag, add IMPLICIT_DEFs in the predecessors as necessary. | |||
927 | for (MachineBasicBlock *Pred : MBB->predecessors()) { | |||
928 | LiveRegs.clear(); | |||
929 | LiveRegs.addLiveOuts(*Pred); | |||
930 | MachineBasicBlock::iterator InsertBefore = Pred->getFirstTerminator(); | |||
931 | for (unsigned Reg : NewLiveIns) { | |||
932 | if (!LiveRegs.available(*MRI, Reg)) | |||
933 | continue; | |||
934 | DebugLoc DL; | |||
935 | BuildMI(*Pred, InsertBefore, DL, TII->get(TargetOpcode::IMPLICIT_DEF), | |||
936 | Reg); | |||
937 | } | |||
938 | } | |||
939 | ||||
940 | MBB->clearLiveIns(); | |||
941 | addLiveIns(*MBB, NewLiveIns); | |||
942 | } | |||
943 | } | |||
944 | ||||
945 | // See if any of the blocks in MergePotentials (which all have SuccBB as a | |||
946 | // successor, or all have no successor if it is null) can be tail-merged. | |||
947 | // If there is a successor, any blocks in MergePotentials that are not | |||
948 | // tail-merged and are not immediately before Succ must have an unconditional | |||
949 | // branch to Succ added (but the predecessor/successor lists need no | |||
950 | // adjustment). The lone predecessor of Succ that falls through into Succ, | |||
951 | // if any, is given in PredBB. | |||
952 | // MinCommonTailLength - Except for the special cases below, tail-merge if | |||
953 | // there are at least this many instructions in common. | |||
954 | bool BranchFolder::TryTailMergeBlocks(MachineBasicBlock *SuccBB, | |||
955 | MachineBasicBlock *PredBB, | |||
956 | unsigned MinCommonTailLength) { | |||
957 | bool MadeChange = false; | |||
958 | ||||
959 | LLVM_DEBUG(do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("branch-folder")) { dbgs() << "\nTryTailMergeBlocks: " ; for (unsigned i = 0, e = MergePotentials.size(); i != e; ++ i) dbgs() << printMBBReference(*MergePotentials[i].getBlock ()) << (i == e - 1 ? "" : ", "); dbgs() << "\n"; if (SuccBB) { dbgs() << " with successor " << printMBBReference (*SuccBB) << '\n'; if (PredBB) dbgs() << " which has fall-through from " << printMBBReference(*PredBB) << "\n"; } dbgs() << "Looking for common tails of at least " << MinCommonTailLength << " instruction" << (MinCommonTailLength == 1 ? "" : "s") << '\n';; } } while (false) | |||
960 | dbgs() << "\nTryTailMergeBlocks: ";do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("branch-folder")) { dbgs() << "\nTryTailMergeBlocks: " ; for (unsigned i = 0, e = MergePotentials.size(); i != e; ++ i) dbgs() << printMBBReference(*MergePotentials[i].getBlock ()) << (i == e - 1 ? "" : ", "); dbgs() << "\n"; if (SuccBB) { dbgs() << " with successor " << printMBBReference (*SuccBB) << '\n'; if (PredBB) dbgs() << " which has fall-through from " << printMBBReference(*PredBB) << "\n"; } dbgs() << "Looking for common tails of at least " << MinCommonTailLength << " instruction" << (MinCommonTailLength == 1 ? "" : "s") << '\n';; } } while (false) | |||
961 | for (unsigned i = 0, e = MergePotentials.size(); i != e; ++i) dbgs()do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("branch-folder")) { dbgs() << "\nTryTailMergeBlocks: " ; for (unsigned i = 0, e = MergePotentials.size(); i != e; ++ i) dbgs() << printMBBReference(*MergePotentials[i].getBlock ()) << (i == e - 1 ? "" : ", "); dbgs() << "\n"; if (SuccBB) { dbgs() << " with successor " << printMBBReference (*SuccBB) << '\n'; if (PredBB) dbgs() << " which has fall-through from " << printMBBReference(*PredBB) << "\n"; } dbgs() << "Looking for common tails of at least " << MinCommonTailLength << " instruction" << (MinCommonTailLength == 1 ? "" : "s") << '\n';; } } while (false) | |||
962 | << printMBBReference(*MergePotentials[i].getBlock())do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("branch-folder")) { dbgs() << "\nTryTailMergeBlocks: " ; for (unsigned i = 0, e = MergePotentials.size(); i != e; ++ i) dbgs() << printMBBReference(*MergePotentials[i].getBlock ()) << (i == e - 1 ? "" : ", "); dbgs() << "\n"; if (SuccBB) { dbgs() << " with successor " << printMBBReference (*SuccBB) << '\n'; if (PredBB) dbgs() << " which has fall-through from " << printMBBReference(*PredBB) << "\n"; } dbgs() << "Looking for common tails of at least " << MinCommonTailLength << " instruction" << (MinCommonTailLength == 1 ? "" : "s") << '\n';; } } while (false) | |||
963 | << (i == e - 1 ? "" : ", ");do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("branch-folder")) { dbgs() << "\nTryTailMergeBlocks: " ; for (unsigned i = 0, e = MergePotentials.size(); i != e; ++ i) dbgs() << printMBBReference(*MergePotentials[i].getBlock ()) << (i == e - 1 ? "" : ", "); dbgs() << "\n"; if (SuccBB) { dbgs() << " with successor " << printMBBReference (*SuccBB) << '\n'; if (PredBB) dbgs() << " which has fall-through from " << printMBBReference(*PredBB) << "\n"; } dbgs() << "Looking for common tails of at least " << MinCommonTailLength << " instruction" << (MinCommonTailLength == 1 ? "" : "s") << '\n';; } } while (false) | |||
964 | dbgs() << "\n"; if (SuccBB) {do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("branch-folder")) { dbgs() << "\nTryTailMergeBlocks: " ; for (unsigned i = 0, e = MergePotentials.size(); i != e; ++ i) dbgs() << printMBBReference(*MergePotentials[i].getBlock ()) << (i == e - 1 ? "" : ", "); dbgs() << "\n"; if (SuccBB) { dbgs() << " with successor " << printMBBReference (*SuccBB) << '\n'; if (PredBB) dbgs() << " which has fall-through from " << printMBBReference(*PredBB) << "\n"; } dbgs() << "Looking for common tails of at least " << MinCommonTailLength << " instruction" << (MinCommonTailLength == 1 ? "" : "s") << '\n';; } } while (false) | |||
965 | dbgs() << " with successor " << printMBBReference(*SuccBB) << '\n';do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("branch-folder")) { dbgs() << "\nTryTailMergeBlocks: " ; for (unsigned i = 0, e = MergePotentials.size(); i != e; ++ i) dbgs() << printMBBReference(*MergePotentials[i].getBlock ()) << (i == e - 1 ? "" : ", "); dbgs() << "\n"; if (SuccBB) { dbgs() << " with successor " << printMBBReference (*SuccBB) << '\n'; if (PredBB) dbgs() << " which has fall-through from " << printMBBReference(*PredBB) << "\n"; } dbgs() << "Looking for common tails of at least " << MinCommonTailLength << " instruction" << (MinCommonTailLength == 1 ? "" : "s") << '\n';; } } while (false) | |||
966 | if (PredBB)do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("branch-folder")) { dbgs() << "\nTryTailMergeBlocks: " ; for (unsigned i = 0, e = MergePotentials.size(); i != e; ++ i) dbgs() << printMBBReference(*MergePotentials[i].getBlock ()) << (i == e - 1 ? "" : ", "); dbgs() << "\n"; if (SuccBB) { dbgs() << " with successor " << printMBBReference (*SuccBB) << '\n'; if (PredBB) dbgs() << " which has fall-through from " << printMBBReference(*PredBB) << "\n"; } dbgs() << "Looking for common tails of at least " << MinCommonTailLength << " instruction" << (MinCommonTailLength == 1 ? "" : "s") << '\n';; } } while (false) | |||
967 | dbgs() << " which has fall-through from "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("branch-folder")) { dbgs() << "\nTryTailMergeBlocks: " ; for (unsigned i = 0, e = MergePotentials.size(); i != e; ++ i) dbgs() << printMBBReference(*MergePotentials[i].getBlock ()) << (i == e - 1 ? "" : ", "); dbgs() << "\n"; if (SuccBB) { dbgs() << " with successor " << printMBBReference (*SuccBB) << '\n'; if (PredBB) dbgs() << " which has fall-through from " << printMBBReference(*PredBB) << "\n"; } dbgs() << "Looking for common tails of at least " << MinCommonTailLength << " instruction" << (MinCommonTailLength == 1 ? "" : "s") << '\n';; } } while (false) | |||
968 | << printMBBReference(*PredBB) << "\n";do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("branch-folder")) { dbgs() << "\nTryTailMergeBlocks: " ; for (unsigned i = 0, e = MergePotentials.size(); i != e; ++ i) dbgs() << printMBBReference(*MergePotentials[i].getBlock ()) << (i == e - 1 ? "" : ", "); dbgs() << "\n"; if (SuccBB) { dbgs() << " with successor " << printMBBReference (*SuccBB) << '\n'; if (PredBB) dbgs() << " which has fall-through from " << printMBBReference(*PredBB) << "\n"; } dbgs() << "Looking for common tails of at least " << MinCommonTailLength << " instruction" << (MinCommonTailLength == 1 ? "" : "s") << '\n';; } } while (false) | |||
969 | } dbgs() << "Looking for common tails of at least "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("branch-folder")) { dbgs() << "\nTryTailMergeBlocks: " ; for (unsigned i = 0, e = MergePotentials.size(); i != e; ++ i) dbgs() << printMBBReference(*MergePotentials[i].getBlock ()) << (i == e - 1 ? "" : ", "); dbgs() << "\n"; if (SuccBB) { dbgs() << " with successor " << printMBBReference (*SuccBB) << '\n'; if (PredBB) dbgs() << " which has fall-through from " << printMBBReference(*PredBB) << "\n"; } dbgs() << "Looking for common tails of at least " << MinCommonTailLength << " instruction" << (MinCommonTailLength == 1 ? "" : "s") << '\n';; } } while (false) | |||
970 | << MinCommonTailLength << " instruction"do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("branch-folder")) { dbgs() << "\nTryTailMergeBlocks: " ; for (unsigned i = 0, e = MergePotentials.size(); i != e; ++ i) dbgs() << printMBBReference(*MergePotentials[i].getBlock ()) << (i == e - 1 ? "" : ", "); dbgs() << "\n"; if (SuccBB) { dbgs() << " with successor " << printMBBReference (*SuccBB) << '\n'; if (PredBB) dbgs() << " which has fall-through from " << printMBBReference(*PredBB) << "\n"; } dbgs() << "Looking for common tails of at least " << MinCommonTailLength << " instruction" << (MinCommonTailLength == 1 ? "" : "s") << '\n';; } } while (false) | |||
971 | << (MinCommonTailLength == 1 ? "" : "s") << '\n';)do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("branch-folder")) { dbgs() << "\nTryTailMergeBlocks: " ; for (unsigned i = 0, e = MergePotentials.size(); i != e; ++ i) dbgs() << printMBBReference(*MergePotentials[i].getBlock ()) << (i == e - 1 ? "" : ", "); dbgs() << "\n"; if (SuccBB) { dbgs() << " with successor " << printMBBReference (*SuccBB) << '\n'; if (PredBB) dbgs() << " which has fall-through from " << printMBBReference(*PredBB) << "\n"; } dbgs() << "Looking for common tails of at least " << MinCommonTailLength << " instruction" << (MinCommonTailLength == 1 ? "" : "s") << '\n';; } } while (false); | |||
972 | ||||
973 | // Sort by hash value so that blocks with identical end sequences sort | |||
974 | // together. | |||
975 | array_pod_sort(MergePotentials.begin(), MergePotentials.end()); | |||
976 | ||||
977 | // Walk through equivalence sets looking for actual exact matches. | |||
978 | while (MergePotentials.size() > 1) { | |||
979 | unsigned CurHash = MergePotentials.back().getHash(); | |||
980 | ||||
981 | // Build SameTails, identifying the set of blocks with this hash code | |||
982 | // and with the maximum number of instructions in common. | |||
983 | unsigned maxCommonTailLength = ComputeSameTails(CurHash, | |||
984 | MinCommonTailLength, | |||
985 | SuccBB, PredBB); | |||
986 | ||||
987 | // If we didn't find any pair that has at least MinCommonTailLength | |||
988 | // instructions in common, remove all blocks with this hash code and retry. | |||
989 | if (SameTails.empty()) { | |||
990 | RemoveBlocksWithHash(CurHash, SuccBB, PredBB); | |||
991 | continue; | |||
992 | } | |||
993 | ||||
994 | // If one of the blocks is the entire common tail (and not the entry | |||
995 | // block, which we can't jump to), we can treat all blocks with this same | |||
996 | // tail at once. Use PredBB if that is one of the possibilities, as that | |||
997 | // will not introduce any extra branches. | |||
998 | MachineBasicBlock *EntryBB = | |||
999 | &MergePotentials.front().getBlock()->getParent()->front(); | |||
1000 | unsigned commonTailIndex = SameTails.size(); | |||
1001 | // If there are two blocks, check to see if one can be made to fall through | |||
1002 | // into the other. | |||
1003 | if (SameTails.size() == 2 && | |||
1004 | SameTails[0].getBlock()->isLayoutSuccessor(SameTails[1].getBlock()) && | |||
1005 | SameTails[1].tailIsWholeBlock()) | |||
1006 | commonTailIndex = 1; | |||
1007 | else if (SameTails.size() == 2 && | |||
1008 | SameTails[1].getBlock()->isLayoutSuccessor( | |||
1009 | SameTails[0].getBlock()) && | |||
1010 | SameTails[0].tailIsWholeBlock()) | |||
1011 | commonTailIndex = 0; | |||
1012 | else { | |||
1013 | // Otherwise just pick one, favoring the fall-through predecessor if | |||
1014 | // there is one. | |||
1015 | for (unsigned i = 0, e = SameTails.size(); i != e; ++i) { | |||
1016 | MachineBasicBlock *MBB = SameTails[i].getBlock(); | |||
1017 | if (MBB == EntryBB && SameTails[i].tailIsWholeBlock()) | |||
1018 | continue; | |||
1019 | if (MBB == PredBB) { | |||
1020 | commonTailIndex = i; | |||
1021 | break; | |||
1022 | } | |||
1023 | if (SameTails[i].tailIsWholeBlock()) | |||
1024 | commonTailIndex = i; | |||
1025 | } | |||
1026 | } | |||
1027 | ||||
1028 | if (commonTailIndex == SameTails.size() || | |||
1029 | (SameTails[commonTailIndex].getBlock() == PredBB && | |||
1030 | !SameTails[commonTailIndex].tailIsWholeBlock())) { | |||
1031 | // None of the blocks consist entirely of the common tail. | |||
1032 | // Split a block so that one does. | |||
1033 | if (!CreateCommonTailOnlyBlock(PredBB, SuccBB, | |||
1034 | maxCommonTailLength, commonTailIndex)) { | |||
1035 | RemoveBlocksWithHash(CurHash, SuccBB, PredBB); | |||
1036 | continue; | |||
1037 | } | |||
1038 | } | |||
1039 | ||||
1040 | MachineBasicBlock *MBB = SameTails[commonTailIndex].getBlock(); | |||
1041 | ||||
1042 | // Recompute common tail MBB's edge weights and block frequency. | |||
1043 | setCommonTailEdgeWeights(*MBB); | |||
1044 | ||||
1045 | // Merge debug locations, MMOs and undef flags across identical instructions | |||
1046 | // for common tail. | |||
1047 | mergeCommonTails(commonTailIndex); | |||
1048 | ||||
1049 | // MBB is common tail. Adjust all other BB's to jump to this one. | |||
1050 | // Traversal must be forwards so erases work. | |||
1051 | LLVM_DEBUG(dbgs() << "\nUsing common tail in " << printMBBReference(*MBB)do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("branch-folder")) { dbgs() << "\nUsing common tail in " << printMBBReference(*MBB) << " for "; } } while (false) | |||
1052 | << " for ")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("branch-folder")) { dbgs() << "\nUsing common tail in " << printMBBReference(*MBB) << " for "; } } while (false); | |||
1053 | for (unsigned int i=0, e = SameTails.size(); i != e; ++i) { | |||
1054 | if (commonTailIndex == i) | |||
1055 | continue; | |||
1056 | LLVM_DEBUG(dbgs() << printMBBReference(*SameTails[i].getBlock())do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("branch-folder")) { dbgs() << printMBBReference(*SameTails [i].getBlock()) << (i == e - 1 ? "" : ", "); } } while ( false) | |||
1057 | << (i == e - 1 ? "" : ", "))do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("branch-folder")) { dbgs() << printMBBReference(*SameTails [i].getBlock()) << (i == e - 1 ? "" : ", "); } } while ( false); | |||
1058 | // Hack the end off BB i, making it jump to BB commonTailIndex instead. | |||
1059 | replaceTailWithBranchTo(SameTails[i].getTailStartPos(), *MBB); | |||
1060 | // BB i is no longer a predecessor of SuccBB; remove it from the worklist. | |||
1061 | MergePotentials.erase(SameTails[i].getMPIter()); | |||
1062 | } | |||
1063 | LLVM_DEBUG(dbgs() << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("branch-folder")) { dbgs() << "\n"; } } while (false); | |||
1064 | // We leave commonTailIndex in the worklist in case there are other blocks | |||
1065 | // that match it with a smaller number of instructions. | |||
1066 | MadeChange = true; | |||
1067 | } | |||
1068 | return MadeChange; | |||
1069 | } | |||
1070 | ||||
1071 | bool BranchFolder::TailMergeBlocks(MachineFunction &MF) { | |||
1072 | bool MadeChange = false; | |||
1073 | if (!EnableTailMerge) return MadeChange; | |||
1074 | ||||
1075 | // First find blocks with no successors. | |||
1076 | // Block placement does not create new tail merging opportunities for these | |||
1077 | // blocks. | |||
1078 | if (!AfterBlockPlacement) { | |||
1079 | MergePotentials.clear(); | |||
1080 | for (MachineBasicBlock &MBB : MF) { | |||
1081 | if (MergePotentials.size() == TailMergeThreshold) | |||
1082 | break; | |||
1083 | if (!TriedMerging.count(&MBB) && MBB.succ_empty()) | |||
1084 | MergePotentials.push_back(MergePotentialsElt(HashEndOfMBB(MBB), &MBB)); | |||
1085 | } | |||
1086 | ||||
1087 | // If this is a large problem, avoid visiting the same basic blocks | |||
1088 | // multiple times. | |||
1089 | if (MergePotentials.size() == TailMergeThreshold) | |||
1090 | for (unsigned i = 0, e = MergePotentials.size(); i != e; ++i) | |||
1091 | TriedMerging.insert(MergePotentials[i].getBlock()); | |||
1092 | ||||
1093 | // See if we can do any tail merging on those. | |||
1094 | if (MergePotentials.size() >= 2) | |||
1095 | MadeChange |= TryTailMergeBlocks(nullptr, nullptr, MinCommonTailLength); | |||
1096 | } | |||
1097 | ||||
1098 | // Look at blocks (IBB) with multiple predecessors (PBB). | |||
1099 | // We change each predecessor to a canonical form, by | |||
1100 | // (1) temporarily removing any unconditional branch from the predecessor | |||
1101 | // to IBB, and | |||
1102 | // (2) alter conditional branches so they branch to the other block | |||
1103 | // not IBB; this may require adding back an unconditional branch to IBB | |||
1104 | // later, where there wasn't one coming in. E.g. | |||
1105 | // Bcc IBB | |||
1106 | // fallthrough to QBB | |||
1107 | // here becomes | |||
1108 | // Bncc QBB | |||
1109 | // with a conceptual B to IBB after that, which never actually exists. | |||
1110 | // With those changes, we see whether the predecessors' tails match, | |||
1111 | // and merge them if so. We change things out of canonical form and | |||
1112 | // back to the way they were later in the process. (OptimizeBranches | |||
1113 | // would undo some of this, but we can't use it, because we'd get into | |||
1114 | // a compile-time infinite loop repeatedly doing and undoing the same | |||
1115 | // transformations.) | |||
1116 | ||||
1117 | for (MachineFunction::iterator I = std::next(MF.begin()), E = MF.end(); | |||
1118 | I != E; ++I) { | |||
1119 | if (I->pred_size() < 2) continue; | |||
1120 | SmallPtrSet<MachineBasicBlock *, 8> UniquePreds; | |||
1121 | MachineBasicBlock *IBB = &*I; | |||
1122 | MachineBasicBlock *PredBB = &*std::prev(I); | |||
1123 | MergePotentials.clear(); | |||
1124 | MachineLoop *ML; | |||
1125 | ||||
1126 | // Bail if merging after placement and IBB is the loop header because | |||
1127 | // -- If merging predecessors that belong to the same loop as IBB, the | |||
1128 | // common tail of merged predecessors may become the loop top if block | |||
1129 | // placement is called again and the predecessors may branch to this common | |||
1130 | // tail and require more branches. This can be relaxed if | |||
1131 | // MachineBlockPlacement::findBestLoopTop is more flexible. | |||
1132 | // --If merging predecessors that do not belong to the same loop as IBB, the | |||
1133 | // loop info of IBB's loop and the other loops may be affected. Calling the | |||
1134 | // block placement again may make big change to the layout and eliminate the | |||
1135 | // reason to do tail merging here. | |||
1136 | if (AfterBlockPlacement && MLI) { | |||
1137 | ML = MLI->getLoopFor(IBB); | |||
1138 | if (ML && IBB == ML->getHeader()) | |||
1139 | continue; | |||
1140 | } | |||
1141 | ||||
1142 | for (MachineBasicBlock *PBB : I->predecessors()) { | |||
1143 | if (MergePotentials.size() == TailMergeThreshold) | |||
1144 | break; | |||
1145 | ||||
1146 | if (TriedMerging.count(PBB)) | |||
1147 | continue; | |||
1148 | ||||
1149 | // Skip blocks that loop to themselves, can't tail merge these. | |||
1150 | if (PBB == IBB) | |||
1151 | continue; | |||
1152 | ||||
1153 | // Visit each predecessor only once. | |||
1154 | if (!UniquePreds.insert(PBB).second) | |||
1155 | continue; | |||
1156 | ||||
1157 | // Skip blocks which may jump to a landing pad. Can't tail merge these. | |||
1158 | if (PBB->hasEHPadSuccessor()) | |||
1159 | continue; | |||
1160 | ||||
1161 | // After block placement, only consider predecessors that belong to the | |||
1162 | // same loop as IBB. The reason is the same as above when skipping loop | |||
1163 | // header. | |||
1164 | if (AfterBlockPlacement && MLI) | |||
1165 | if (ML != MLI->getLoopFor(PBB)) | |||
| ||||
1166 | continue; | |||
1167 | ||||
1168 | MachineBasicBlock *TBB = nullptr, *FBB = nullptr; | |||
1169 | SmallVector<MachineOperand, 4> Cond; | |||
1170 | if (!TII->analyzeBranch(*PBB, TBB, FBB, Cond, true)) { | |||
1171 | // Failing case: IBB is the target of a cbr, and we cannot reverse the | |||
1172 | // branch. | |||
1173 | SmallVector<MachineOperand, 4> NewCond(Cond); | |||
1174 | if (!Cond.empty() && TBB == IBB) { | |||
1175 | if (TII->reverseBranchCondition(NewCond)) | |||
1176 | continue; | |||
1177 | // This is the QBB case described above | |||
1178 | if (!FBB) { | |||
1179 | auto Next = ++PBB->getIterator(); | |||
1180 | if (Next != MF.end()) | |||
1181 | FBB = &*Next; | |||
1182 | } | |||
1183 | } | |||
1184 | ||||
1185 | // Remove the unconditional branch at the end, if any. | |||
1186 | if (TBB && (Cond.empty() || FBB)) { | |||
1187 | DebugLoc dl = PBB->findBranchDebugLoc(); | |||
1188 | TII->removeBranch(*PBB); | |||
1189 | if (!Cond.empty()) | |||
1190 | // reinsert conditional branch only, for now | |||
1191 | TII->insertBranch(*PBB, (TBB == IBB) ? FBB : TBB, nullptr, | |||
1192 | NewCond, dl); | |||
1193 | } | |||
1194 | ||||
1195 | MergePotentials.push_back(MergePotentialsElt(HashEndOfMBB(*PBB), PBB)); | |||
1196 | } | |||
1197 | } | |||
1198 | ||||
1199 | // If this is a large problem, avoid visiting the same basic blocks multiple | |||
1200 | // times. | |||
1201 | if (MergePotentials.size() == TailMergeThreshold) | |||
1202 | for (unsigned i = 0, e = MergePotentials.size(); i != e; ++i) | |||
1203 | TriedMerging.insert(MergePotentials[i].getBlock()); | |||
1204 | ||||
1205 | if (MergePotentials.size() >= 2) | |||
1206 | MadeChange |= TryTailMergeBlocks(IBB, PredBB, MinCommonTailLength); | |||
1207 | ||||
1208 | // Reinsert an unconditional branch if needed. The 1 below can occur as a | |||
1209 | // result of removing blocks in TryTailMergeBlocks. | |||
1210 | PredBB = &*std::prev(I); // this may have been changed in TryTailMergeBlocks | |||
1211 | if (MergePotentials.size() == 1 && | |||
1212 | MergePotentials.begin()->getBlock() != PredBB) | |||
1213 | FixTail(MergePotentials.begin()->getBlock(), IBB, TII); | |||
1214 | } | |||
1215 | ||||
1216 | return MadeChange; | |||
1217 | } | |||
1218 | ||||
1219 | void BranchFolder::setCommonTailEdgeWeights(MachineBasicBlock &TailMBB) { | |||
1220 | SmallVector<BlockFrequency, 2> EdgeFreqLs(TailMBB.succ_size()); | |||
1221 | BlockFrequency AccumulatedMBBFreq; | |||
1222 | ||||
1223 | // Aggregate edge frequency of successor edge j: | |||
1224 | // edgeFreq(j) = sum (freq(bb) * edgeProb(bb, j)), | |||
1225 | // where bb is a basic block that is in SameTails. | |||
1226 | for (const auto &Src : SameTails) { | |||
1227 | const MachineBasicBlock *SrcMBB = Src.getBlock(); | |||
1228 | BlockFrequency BlockFreq = MBBFreqInfo.getBlockFreq(SrcMBB); | |||
1229 | AccumulatedMBBFreq += BlockFreq; | |||
1230 | ||||
1231 | // It is not necessary to recompute edge weights if TailBB has less than two | |||
1232 | // successors. | |||
1233 | if (TailMBB.succ_size() <= 1) | |||
1234 | continue; | |||
1235 | ||||
1236 | auto EdgeFreq = EdgeFreqLs.begin(); | |||
1237 | ||||
1238 | for (auto SuccI = TailMBB.succ_begin(), SuccE = TailMBB.succ_end(); | |||
1239 | SuccI != SuccE; ++SuccI, ++EdgeFreq) | |||
1240 | *EdgeFreq += BlockFreq * MBPI.getEdgeProbability(SrcMBB, *SuccI); | |||
1241 | } | |||
1242 | ||||
1243 | MBBFreqInfo.setBlockFreq(&TailMBB, AccumulatedMBBFreq); | |||
1244 | ||||
1245 | if (TailMBB.succ_size() <= 1) | |||
1246 | return; | |||
1247 | ||||
1248 | auto SumEdgeFreq = | |||
1249 | std::accumulate(EdgeFreqLs.begin(), EdgeFreqLs.end(), BlockFrequency(0)) | |||
1250 | .getFrequency(); | |||
1251 | auto EdgeFreq = EdgeFreqLs.begin(); | |||
1252 | ||||
1253 | if (SumEdgeFreq > 0) { | |||
1254 | for (auto SuccI = TailMBB.succ_begin(), SuccE = TailMBB.succ_end(); | |||
1255 | SuccI != SuccE; ++SuccI, ++EdgeFreq) { | |||
1256 | auto Prob = BranchProbability::getBranchProbability( | |||
1257 | EdgeFreq->getFrequency(), SumEdgeFreq); | |||
1258 | TailMBB.setSuccProbability(SuccI, Prob); | |||
1259 | } | |||
1260 | } | |||
1261 | } | |||
1262 | ||||
1263 | //===----------------------------------------------------------------------===// | |||
1264 | // Branch Optimization | |||
1265 | //===----------------------------------------------------------------------===// | |||
1266 | ||||
1267 | bool BranchFolder::OptimizeBranches(MachineFunction &MF) { | |||
1268 | bool MadeChange = false; | |||
1269 | ||||
1270 | // Make sure blocks are numbered in order | |||
1271 | MF.RenumberBlocks(); | |||
1272 | // Renumbering blocks alters EH scope membership, recalculate it. | |||
1273 | EHScopeMembership = getEHScopeMembership(MF); | |||
1274 | ||||
1275 | for (MachineFunction::iterator I = std::next(MF.begin()), E = MF.end(); | |||
1276 | I != E; ) { | |||
1277 | MachineBasicBlock *MBB = &*I++; | |||
1278 | MadeChange |= OptimizeBlock(MBB); | |||
1279 | ||||
1280 | // If it is dead, remove it. | |||
1281 | if (MBB->pred_empty()) { | |||
1282 | RemoveDeadBlock(MBB); | |||
1283 | MadeChange = true; | |||
1284 | ++NumDeadBlocks; | |||
1285 | } | |||
1286 | } | |||
1287 | ||||
1288 | return MadeChange; | |||
1289 | } | |||
1290 | ||||
1291 | // Blocks should be considered empty if they contain only debug info; | |||
1292 | // else the debug info would affect codegen. | |||
1293 | static bool IsEmptyBlock(MachineBasicBlock *MBB) { | |||
1294 | return MBB->getFirstNonDebugInstr() == MBB->end(); | |||
1295 | } | |||
1296 | ||||
1297 | // Blocks with only debug info and branches should be considered the same | |||
1298 | // as blocks with only branches. | |||
1299 | static bool IsBranchOnlyBlock(MachineBasicBlock *MBB) { | |||
1300 | MachineBasicBlock::iterator I = MBB->getFirstNonDebugInstr(); | |||
1301 | assert(I != MBB->end() && "empty block!")((I != MBB->end() && "empty block!") ? static_cast <void> (0) : __assert_fail ("I != MBB->end() && \"empty block!\"" , "/build/llvm-toolchain-snapshot-9~svn362543/lib/CodeGen/BranchFolding.cpp" , 1301, __PRETTY_FUNCTION__)); | |||
1302 | return I->isBranch(); | |||
1303 | } | |||
1304 | ||||
1305 | /// IsBetterFallthrough - Return true if it would be clearly better to | |||
1306 | /// fall-through to MBB1 than to fall through into MBB2. This has to return | |||
1307 | /// a strict ordering, returning true for both (MBB1,MBB2) and (MBB2,MBB1) will | |||
1308 | /// result in infinite loops. | |||
1309 | static bool IsBetterFallthrough(MachineBasicBlock *MBB1, | |||
1310 | MachineBasicBlock *MBB2) { | |||
1311 | // Right now, we use a simple heuristic. If MBB2 ends with a call, and | |||
1312 | // MBB1 doesn't, we prefer to fall through into MBB1. This allows us to | |||
1313 | // optimize branches that branch to either a return block or an assert block | |||
1314 | // into a fallthrough to the return. | |||
1315 | MachineBasicBlock::iterator MBB1I = MBB1->getLastNonDebugInstr(); | |||
1316 | MachineBasicBlock::iterator MBB2I = MBB2->getLastNonDebugInstr(); | |||
1317 | if (MBB1I == MBB1->end() || MBB2I == MBB2->end()) | |||
1318 | return false; | |||
1319 | ||||
1320 | // If there is a clear successor ordering we make sure that one block | |||
1321 | // will fall through to the next | |||
1322 | if (MBB1->isSuccessor(MBB2)) return true; | |||
1323 | if (MBB2->isSuccessor(MBB1)) return false; | |||
1324 | ||||
1325 | return MBB2I->isCall() && !MBB1I->isCall(); | |||
1326 | } | |||
1327 | ||||
1328 | /// getBranchDebugLoc - Find and return, if any, the DebugLoc of the branch | |||
1329 | /// instructions on the block. | |||
1330 | static DebugLoc getBranchDebugLoc(MachineBasicBlock &MBB) { | |||
1331 | MachineBasicBlock::iterator I = MBB.getLastNonDebugInstr(); | |||
1332 | if (I != MBB.end() && I->isBranch()) | |||
1333 | return I->getDebugLoc(); | |||
1334 | return DebugLoc(); | |||
1335 | } | |||
1336 | ||||
1337 | static void copyDebugInfoToPredecessor(const TargetInstrInfo *TII, | |||
1338 | MachineBasicBlock &MBB, | |||
1339 | MachineBasicBlock &PredMBB) { | |||
1340 | auto InsertBefore = PredMBB.getFirstTerminator(); | |||
1341 | for (MachineInstr &MI : MBB.instrs()) | |||
1342 | if (MI.isDebugInstr()) { | |||
1343 | TII->duplicate(PredMBB, InsertBefore, MI); | |||
1344 | LLVM_DEBUG(dbgs() << "Copied debug entity from empty block to pred: "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("branch-folder")) { dbgs() << "Copied debug entity from empty block to pred: " << MI; } } while (false) | |||
1345 | << MI)do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("branch-folder")) { dbgs() << "Copied debug entity from empty block to pred: " << MI; } } while (false); | |||
1346 | } | |||
1347 | } | |||
1348 | ||||
1349 | static void copyDebugInfoToSuccessor(const TargetInstrInfo *TII, | |||
1350 | MachineBasicBlock &MBB, | |||
1351 | MachineBasicBlock &SuccMBB) { | |||
1352 | auto InsertBefore = SuccMBB.SkipPHIsAndLabels(SuccMBB.begin()); | |||
1353 | for (MachineInstr &MI : MBB.instrs()) | |||
1354 | if (MI.isDebugInstr()) { | |||
1355 | TII->duplicate(SuccMBB, InsertBefore, MI); | |||
1356 | LLVM_DEBUG(dbgs() << "Copied debug entity from empty block to succ: "do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("branch-folder")) { dbgs() << "Copied debug entity from empty block to succ: " << MI; } } while (false) | |||
1357 | << MI)do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("branch-folder")) { dbgs() << "Copied debug entity from empty block to succ: " << MI; } } while (false); | |||
1358 | } | |||
1359 | } | |||
1360 | ||||
1361 | // Try to salvage DBG_VALUE instructions from an otherwise empty block. If such | |||
1362 | // a basic block is removed we would lose the debug information unless we have | |||
1363 | // copied the information to a predecessor/successor. | |||
1364 | // | |||
1365 | // TODO: This function only handles some simple cases. An alternative would be | |||
1366 | // to run a heavier analysis, such as the LiveDebugValues pass, before we do | |||
1367 | // branch folding. | |||
1368 | static void salvageDebugInfoFromEmptyBlock(const TargetInstrInfo *TII, | |||
1369 | MachineBasicBlock &MBB) { | |||
1370 | assert(IsEmptyBlock(&MBB) && "Expected an empty block (except debug info).")((IsEmptyBlock(&MBB) && "Expected an empty block (except debug info)." ) ? static_cast<void> (0) : __assert_fail ("IsEmptyBlock(&MBB) && \"Expected an empty block (except debug info).\"" , "/build/llvm-toolchain-snapshot-9~svn362543/lib/CodeGen/BranchFolding.cpp" , 1370, __PRETTY_FUNCTION__)); | |||
1371 | // If this MBB is the only predecessor of a successor it is legal to copy | |||
1372 | // DBG_VALUE instructions to the beginning of the successor. | |||
1373 | for (MachineBasicBlock *SuccBB : MBB.successors()) | |||
1374 | if (SuccBB->pred_size() == 1) | |||
1375 | copyDebugInfoToSuccessor(TII, MBB, *SuccBB); | |||
1376 | // If this MBB is the only successor of a predecessor it is legal to copy the | |||
1377 | // DBG_VALUE instructions to the end of the predecessor (just before the | |||
1378 | // terminators, assuming that the terminator isn't affecting the DBG_VALUE). | |||
1379 | for (MachineBasicBlock *PredBB : MBB.predecessors()) | |||
1380 | if (PredBB->succ_size() == 1) | |||
1381 | copyDebugInfoToPredecessor(TII, MBB, *PredBB); | |||
1382 | } | |||
1383 | ||||
1384 | bool BranchFolder::OptimizeBlock(MachineBasicBlock *MBB) { | |||
1385 | bool MadeChange = false; | |||
1386 | MachineFunction &MF = *MBB->getParent(); | |||
1387 | ReoptimizeBlock: | |||
1388 | ||||
1389 | MachineFunction::iterator FallThrough = MBB->getIterator(); | |||
1390 | ++FallThrough; | |||
1391 | ||||
1392 | // Make sure MBB and FallThrough belong to the same EH scope. | |||
1393 | bool SameEHScope = true; | |||
1394 | if (!EHScopeMembership.empty() && FallThrough != MF.end()) { | |||
1395 | auto MBBEHScope = EHScopeMembership.find(MBB); | |||
1396 | assert(MBBEHScope != EHScopeMembership.end())((MBBEHScope != EHScopeMembership.end()) ? static_cast<void > (0) : __assert_fail ("MBBEHScope != EHScopeMembership.end()" , "/build/llvm-toolchain-snapshot-9~svn362543/lib/CodeGen/BranchFolding.cpp" , 1396, __PRETTY_FUNCTION__)); | |||
1397 | auto FallThroughEHScope = EHScopeMembership.find(&*FallThrough); | |||
1398 | assert(FallThroughEHScope != EHScopeMembership.end())((FallThroughEHScope != EHScopeMembership.end()) ? static_cast <void> (0) : __assert_fail ("FallThroughEHScope != EHScopeMembership.end()" , "/build/llvm-toolchain-snapshot-9~svn362543/lib/CodeGen/BranchFolding.cpp" , 1398, __PRETTY_FUNCTION__)); | |||
1399 | SameEHScope = MBBEHScope->second == FallThroughEHScope->second; | |||
1400 | } | |||
1401 | ||||
1402 | // If this block is empty, make everyone use its fall-through, not the block | |||
1403 | // explicitly. Landing pads should not do this since the landing-pad table | |||
1404 | // points to this block. Blocks with their addresses taken shouldn't be | |||
1405 | // optimized away. | |||
1406 | if (IsEmptyBlock(MBB) && !MBB->isEHPad() && !MBB->hasAddressTaken() && | |||
1407 | SameEHScope) { | |||
1408 | salvageDebugInfoFromEmptyBlock(TII, *MBB); | |||
1409 | // Dead block? Leave for cleanup later. | |||
1410 | if (MBB->pred_empty()) return MadeChange; | |||
1411 | ||||
1412 | if (FallThrough == MF.end()) { | |||
1413 | // TODO: Simplify preds to not branch here if possible! | |||
1414 | } else if (FallThrough->isEHPad()) { | |||
1415 | // Don't rewrite to a landing pad fallthough. That could lead to the case | |||
1416 | // where a BB jumps to more than one landing pad. | |||
1417 | // TODO: Is it ever worth rewriting predecessors which don't already | |||
1418 | // jump to a landing pad, and so can safely jump to the fallthrough? | |||
1419 | } else if (MBB->isSuccessor(&*FallThrough)) { | |||
1420 | // Rewrite all predecessors of the old block to go to the fallthrough | |||
1421 | // instead. | |||
1422 | while (!MBB->pred_empty()) { | |||
1423 | MachineBasicBlock *Pred = *(MBB->pred_end()-1); | |||
1424 | Pred->ReplaceUsesOfBlockWith(MBB, &*FallThrough); | |||
1425 | } | |||
1426 | // If MBB was the target of a jump table, update jump tables to go to the | |||
1427 | // fallthrough instead. | |||
1428 | if (MachineJumpTableInfo *MJTI = MF.getJumpTableInfo()) | |||
1429 | MJTI->ReplaceMBBInJumpTables(MBB, &*FallThrough); | |||
1430 | MadeChange = true; | |||
1431 | } | |||
1432 | return MadeChange; | |||
1433 | } | |||
1434 | ||||
1435 | // Check to see if we can simplify the terminator of the block before this | |||
1436 | // one. | |||
1437 | MachineBasicBlock &PrevBB = *std::prev(MachineFunction::iterator(MBB)); | |||
1438 | ||||
1439 | MachineBasicBlock *PriorTBB = nullptr, *PriorFBB = nullptr; | |||
1440 | SmallVector<MachineOperand, 4> PriorCond; | |||
1441 | bool PriorUnAnalyzable = | |||
1442 | TII->analyzeBranch(PrevBB, PriorTBB, PriorFBB, PriorCond, true); | |||
1443 | if (!PriorUnAnalyzable) { | |||
1444 | // If the CFG for the prior block has extra edges, remove them. | |||
1445 | MadeChange |= PrevBB.CorrectExtraCFGEdges(PriorTBB, PriorFBB, | |||
1446 | !PriorCond.empty()); | |||
1447 | ||||
1448 | // If the previous branch is conditional and both conditions go to the same | |||
1449 | // destination, remove the branch, replacing it with an unconditional one or | |||
1450 | // a fall-through. | |||
1451 | if (PriorTBB && PriorTBB == PriorFBB) { | |||
1452 | DebugLoc dl = getBranchDebugLoc(PrevBB); | |||
1453 | TII->removeBranch(PrevBB); | |||
1454 | PriorCond.clear(); | |||
1455 | if (PriorTBB != MBB) | |||
1456 | TII->insertBranch(PrevBB, PriorTBB, nullptr, PriorCond, dl); | |||
1457 | MadeChange = true; | |||
1458 | ++NumBranchOpts; | |||
1459 | goto ReoptimizeBlock; | |||
1460 | } | |||
1461 | ||||
1462 | // If the previous block unconditionally falls through to this block and | |||
1463 | // this block has no other predecessors, move the contents of this block | |||
1464 | // into the prior block. This doesn't usually happen when SimplifyCFG | |||
1465 | // has been used, but it can happen if tail merging splits a fall-through | |||
1466 | // predecessor of a block. | |||
1467 | // This has to check PrevBB->succ_size() because EH edges are ignored by | |||
1468 | // AnalyzeBranch. | |||
1469 | if (PriorCond.empty() && !PriorTBB && MBB->pred_size() == 1 && | |||
1470 | PrevBB.succ_size() == 1 && | |||
1471 | !MBB->hasAddressTaken() && !MBB->isEHPad()) { | |||
1472 | LLVM_DEBUG(dbgs() << "\nMerging into block: " << PrevBBdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("branch-folder")) { dbgs() << "\nMerging into block: " << PrevBB << "From MBB: " << *MBB; } } while (false) | |||
1473 | << "From MBB: " << *MBB)do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("branch-folder")) { dbgs() << "\nMerging into block: " << PrevBB << "From MBB: " << *MBB; } } while (false); | |||
1474 | // Remove redundant DBG_VALUEs first. | |||
1475 | if (PrevBB.begin() != PrevBB.end()) { | |||
1476 | MachineBasicBlock::iterator PrevBBIter = PrevBB.end(); | |||
1477 | --PrevBBIter; | |||
1478 | MachineBasicBlock::iterator MBBIter = MBB->begin(); | |||
1479 | // Check if DBG_VALUE at the end of PrevBB is identical to the | |||
1480 | // DBG_VALUE at the beginning of MBB. | |||
1481 | while (PrevBBIter != PrevBB.begin() && MBBIter != MBB->end() | |||
1482 | && PrevBBIter->isDebugInstr() && MBBIter->isDebugInstr()) { | |||
1483 | if (!MBBIter->isIdenticalTo(*PrevBBIter)) | |||
1484 | break; | |||
1485 | MachineInstr &DuplicateDbg = *MBBIter; | |||
1486 | ++MBBIter; -- PrevBBIter; | |||
1487 | DuplicateDbg.eraseFromParent(); | |||
1488 | } | |||
1489 | } | |||
1490 | PrevBB.splice(PrevBB.end(), MBB, MBB->begin(), MBB->end()); | |||
1491 | PrevBB.removeSuccessor(PrevBB.succ_begin()); | |||
1492 | assert(PrevBB.succ_empty())((PrevBB.succ_empty()) ? static_cast<void> (0) : __assert_fail ("PrevBB.succ_empty()", "/build/llvm-toolchain-snapshot-9~svn362543/lib/CodeGen/BranchFolding.cpp" , 1492, __PRETTY_FUNCTION__)); | |||
1493 | PrevBB.transferSuccessors(MBB); | |||
1494 | MadeChange = true; | |||
1495 | return MadeChange; | |||
1496 | } | |||
1497 | ||||
1498 | // If the previous branch *only* branches to *this* block (conditional or | |||
1499 | // not) remove the branch. | |||
1500 | if (PriorTBB == MBB && !PriorFBB) { | |||
1501 | TII->removeBranch(PrevBB); | |||
1502 | MadeChange = true; | |||
1503 | ++NumBranchOpts; | |||
1504 | goto ReoptimizeBlock; | |||
1505 | } | |||
1506 | ||||
1507 | // If the prior block branches somewhere else on the condition and here if | |||
1508 | // the condition is false, remove the uncond second branch. | |||
1509 | if (PriorFBB == MBB) { | |||
1510 | DebugLoc dl = getBranchDebugLoc(PrevBB); | |||
1511 | TII->removeBranch(PrevBB); | |||
1512 | TII->insertBranch(PrevBB, PriorTBB, nullptr, PriorCond, dl); | |||
1513 | MadeChange = true; | |||
1514 | ++NumBranchOpts; | |||
1515 | goto ReoptimizeBlock; | |||
1516 | } | |||
1517 | ||||
1518 | // If the prior block branches here on true and somewhere else on false, and | |||
1519 | // if the branch condition is reversible, reverse the branch to create a | |||
1520 | // fall-through. | |||
1521 | if (PriorTBB == MBB) { | |||
1522 | SmallVector<MachineOperand, 4> NewPriorCond(PriorCond); | |||
1523 | if (!TII->reverseBranchCondition(NewPriorCond)) { | |||
1524 | DebugLoc dl = getBranchDebugLoc(PrevBB); | |||
1525 | TII->removeBranch(PrevBB); | |||
1526 | TII->insertBranch(PrevBB, PriorFBB, nullptr, NewPriorCond, dl); | |||
1527 | MadeChange = true; | |||
1528 | ++NumBranchOpts; | |||
1529 | goto ReoptimizeBlock; | |||
1530 | } | |||
1531 | } | |||
1532 | ||||
1533 | // If this block has no successors (e.g. it is a return block or ends with | |||
1534 | // a call to a no-return function like abort or __cxa_throw) and if the pred | |||
1535 | // falls through into this block, and if it would otherwise fall through | |||
1536 | // into the block after this, move this block to the end of the function. | |||
1537 | // | |||
1538 | // We consider it more likely that execution will stay in the function (e.g. | |||
1539 | // due to loops) than it is to exit it. This asserts in loops etc, moving | |||
1540 | // the assert condition out of the loop body. | |||
1541 | if (MBB->succ_empty() && !PriorCond.empty() && !PriorFBB && | |||
1542 | MachineFunction::iterator(PriorTBB) == FallThrough && | |||
1543 | !MBB->canFallThrough()) { | |||
1544 | bool DoTransform = true; | |||
1545 | ||||
1546 | // We have to be careful that the succs of PredBB aren't both no-successor | |||
1547 | // blocks. If neither have successors and if PredBB is the second from | |||
1548 | // last block in the function, we'd just keep swapping the two blocks for | |||
1549 | // last. Only do the swap if one is clearly better to fall through than | |||
1550 | // the other. | |||
1551 | if (FallThrough == --MF.end() && | |||
1552 | !IsBetterFallthrough(PriorTBB, MBB)) | |||
1553 | DoTransform = false; | |||
1554 | ||||
1555 | if (DoTransform) { | |||
1556 | // Reverse the branch so we will fall through on the previous true cond. | |||
1557 | SmallVector<MachineOperand, 4> NewPriorCond(PriorCond); | |||
1558 | if (!TII->reverseBranchCondition(NewPriorCond)) { | |||
1559 | LLVM_DEBUG(dbgs() << "\nMoving MBB: " << *MBBdo { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("branch-folder")) { dbgs() << "\nMoving MBB: " << *MBB << "To make fallthrough to: " << *PriorTBB << "\n"; } } while (false) | |||
1560 | << "To make fallthrough to: " << *PriorTBB << "\n")do { if (::llvm::DebugFlag && ::llvm::isCurrentDebugType ("branch-folder")) { dbgs() << "\nMoving MBB: " << *MBB << "To make fallthrough to: " << *PriorTBB << "\n"; } } while (false); | |||
1561 | ||||
1562 | DebugLoc dl = getBranchDebugLoc(PrevBB); | |||
1563 | TII->removeBranch(PrevBB); | |||
1564 | TII->insertBranch(PrevBB, MBB, nullptr, NewPriorCond, dl); | |||
1565 | ||||
1566 | // Move this block to the end of the function. | |||
1567 | MBB->moveAfter(&MF.back()); | |||
1568 | MadeChange = true; | |||
1569 | ++NumBranchOpts; | |||
1570 | return MadeChange; | |||
1571 | } | |||
1572 | } | |||
1573 | } | |||
1574 | } | |||
1575 | ||||
1576 | if (!IsEmptyBlock(MBB) && MBB->pred_size() == 1 && | |||
1577 | MF.getFunction().hasOptSize()) { | |||
1578 | // Changing "Jcc foo; foo: jmp bar;" into "Jcc bar;" might change the branch | |||
1579 | // direction, thereby defeating careful block placement and regressing | |||
1580 | // performance. Therefore, only consider this for optsize functions. | |||
1581 | MachineInstr &TailCall = *MBB->getFirstNonDebugInstr(); | |||
1582 | if (TII->isUnconditionalTailCall(TailCall)) { | |||
1583 | MachineBasicBlock *Pred = *MBB->pred_begin(); | |||
1584 | MachineBasicBlock *PredTBB = nullptr, *PredFBB = nullptr; | |||
1585 | SmallVector<MachineOperand, 4> PredCond; | |||
1586 | bool PredAnalyzable = | |||
1587 | !TII->analyzeBranch(*Pred, PredTBB, PredFBB, PredCond, true); | |||
1588 | ||||
1589 | if (PredAnalyzable && !PredCond.empty() && PredTBB == MBB && | |||
1590 | PredTBB != PredFBB) { | |||
1591 | // The predecessor has a conditional branch to this block which consists | |||
1592 | // of only a tail call. Try to fold the tail call into the conditional | |||
1593 | // branch. | |||
1594 | if (TII->canMakeTailCallConditional(PredCond, TailCall)) { | |||
1595 | // TODO: It would be nice if analyzeBranch() could provide a pointer | |||
1596 | // to the branch instruction so replaceBranchWithTailCall() doesn't | |||
1597 | // have to search for it. | |||
1598 | TII->replaceBranchWithTailCall(*Pred, PredCond, TailCall); | |||
1599 | ++NumTailCalls; | |||
1600 | Pred->removeSuccessor(MBB); | |||
1601 | MadeChange = true; | |||
1602 | return MadeChange; | |||
1603 | } | |||
1604 | } | |||
1605 | // If the predecessor is falling through to this block, we could reverse | |||
1606 | // the branch condition and fold the tail call into that. However, after | |||
1607 | // that we might have to re-arrange the CFG to fall through to the other | |||
1608 | // block and there is a high risk of regressing code size rather than | |||
1609 | // improving it. | |||
1610 | } | |||
1611 | } | |||
1612 | ||||
1613 | // Analyze the branch in the current block. | |||
1614 | MachineBasicBlock *CurTBB = nullptr, *CurFBB = nullptr; | |||
1615 | SmallVector<MachineOperand, 4> CurCond; | |||
1616 | bool CurUnAnalyzable = | |||
1617 | TII->analyzeBranch(*MBB, CurTBB, CurFBB, CurCond, true); | |||
1618 | if (!CurUnAnalyzable) { | |||
1619 | // If the CFG for the prior block has extra edges, remove them. | |||
1620 | MadeChange |= MBB->CorrectExtraCFGEdges(CurTBB, CurFBB, !CurCond.empty()); | |||
1621 | ||||
1622 | // If this is a two-way branch, and the FBB branches to this block, reverse | |||
1623 | // the condition so the single-basic-block loop is faster. Instead of: | |||
1624 | // Loop: xxx; jcc Out; jmp Loop | |||
1625 | // we want: | |||
1626 | // Loop: xxx; jncc Loop; jmp Out | |||
1627 | if (CurTBB && CurFBB && CurFBB == MBB && CurTBB != MBB) { | |||
1628 | SmallVector<MachineOperand, 4> NewCond(CurCond); | |||
1629 | if (!TII->reverseBranchCondition(NewCond)) { | |||
1630 | DebugLoc dl = getBranchDebugLoc(*MBB); | |||
1631 | TII->removeBranch(*MBB); | |||
1632 | TII->insertBranch(*MBB, CurFBB, CurTBB, NewCond, dl); | |||
1633 | MadeChange = true; | |||
1634 | ++NumBranchOpts; | |||
1635 | goto ReoptimizeBlock; | |||
1636 | } | |||
1637 | } | |||
1638 | ||||
1639 | // If this branch is the only thing in its block, see if we can forward | |||
1640 | // other blocks across it. | |||
1641 | if (CurTBB && CurCond.empty() && !CurFBB && | |||
1642 | IsBranchOnlyBlock(MBB) && CurTBB != MBB && | |||
1643 | !MBB->hasAddressTaken() && !MBB->isEHPad()) { | |||
1644 | DebugLoc dl = getBranchDebugLoc(*MBB); | |||
1645 | // This block may contain just an unconditional branch. Because there can | |||
1646 | // be 'non-branch terminators' in the block, try removing the branch and | |||
1647 | // then seeing if the block is empty. | |||
1648 | TII->removeBranch(*MBB); | |||
1649 | // If the only things remaining in the block are debug info, remove these | |||
1650 | // as well, so this will behave the same as an empty block in non-debug | |||
1651 | // mode. | |||
1652 | if (IsEmptyBlock(MBB)) { | |||
1653 | // Make the block empty, losing the debug info (we could probably | |||
1654 | // improve this in some cases.) | |||
1655 | MBB->erase(MBB->begin(), MBB->end()); | |||
1656 | } | |||
1657 | // If this block is just an unconditional branch to CurTBB, we can | |||
1658 | // usually completely eliminate the block. The only case we cannot | |||
1659 | // completely eliminate the block is when the block before this one | |||
1660 | // falls through into MBB and we can't understand the prior block's branch | |||
1661 | // condition. | |||
1662 | if (MBB->empty()) { | |||
1663 | bool PredHasNoFallThrough = !PrevBB.canFallThrough(); | |||
1664 | if (PredHasNoFallThrough || !PriorUnAnalyzable || | |||
1665 | !PrevBB.isSuccessor(MBB)) { | |||
1666 | // If the prior block falls through into us, turn it into an | |||
1667 | // explicit branch to us to make updates simpler. | |||
1668 | if (!PredHasNoFallThrough && PrevBB.isSuccessor(MBB) && | |||
1669 | PriorTBB != MBB && PriorFBB != MBB) { | |||
1670 | if (!PriorTBB) { | |||
1671 | assert(PriorCond.empty() && !PriorFBB &&((PriorCond.empty() && !PriorFBB && "Bad branch analysis" ) ? static_cast<void> (0) : __assert_fail ("PriorCond.empty() && !PriorFBB && \"Bad branch analysis\"" , "/build/llvm-toolchain-snapshot-9~svn362543/lib/CodeGen/BranchFolding.cpp" , 1672, __PRETTY_FUNCTION__)) | |||
1672 | "Bad branch analysis")((PriorCond.empty() && !PriorFBB && "Bad branch analysis" ) ? static_cast<void> (0) : __assert_fail ("PriorCond.empty() && !PriorFBB && \"Bad branch analysis\"" , "/build/llvm-toolchain-snapshot-9~svn362543/lib/CodeGen/BranchFolding.cpp" , 1672, __PRETTY_FUNCTION__)); | |||
1673 | PriorTBB = MBB; | |||
1674 | } else { | |||
1675 | assert(!PriorFBB && "Machine CFG out of date!")((!PriorFBB && "Machine CFG out of date!") ? static_cast <void> (0) : __assert_fail ("!PriorFBB && \"Machine CFG out of date!\"" , "/build/llvm-toolchain-snapshot-9~svn362543/lib/CodeGen/BranchFolding.cpp" , 1675, __PRETTY_FUNCTION__)); | |||
1676 | PriorFBB = MBB; | |||
1677 | } | |||
1678 | DebugLoc pdl = getBranchDebugLoc(PrevBB); | |||
1679 | TII->removeBranch(PrevBB); | |||
1680 | TII->insertBranch(PrevBB, PriorTBB, PriorFBB, PriorCond, pdl); | |||
1681 | } | |||
1682 | ||||
1683 | // Iterate through all the predecessors, revectoring each in-turn. | |||
1684 | size_t PI = 0; | |||
1685 | bool DidChange = false; | |||
1686 | bool HasBranchToSelf = false; | |||
1687 | while(PI != MBB->pred_size()) { | |||
1688 | MachineBasicBlock *PMBB = *(MBB->pred_begin() + PI); | |||
1689 | if (PMBB == MBB) { | |||
1690 | // If this block has an uncond branch to itself, leave it. | |||
1691 | ++PI; | |||
1692 | HasBranchToSelf = true; | |||
1693 | } else { | |||
1694 | DidChange = true; | |||
1695 | PMBB->ReplaceUsesOfBlockWith(MBB, CurTBB); | |||
1696 | // If this change resulted in PMBB ending in a conditional | |||
1697 | // branch where both conditions go to the same destination, | |||
1698 | // change this to an unconditional branch (and fix the CFG). | |||
1699 | MachineBasicBlock *NewCurTBB = nullptr, *NewCurFBB = nullptr; | |||
1700 | SmallVector<MachineOperand, 4> NewCurCond; | |||
1701 | bool NewCurUnAnalyzable = TII->analyzeBranch( | |||
1702 | *PMBB, NewCurTBB, NewCurFBB, NewCurCond, true); | |||
1703 | if (!NewCurUnAnalyzable && NewCurTBB && NewCurTBB == NewCurFBB) { | |||
1704 | DebugLoc pdl = getBranchDebugLoc(*PMBB); | |||
1705 | TII->removeBranch(*PMBB); | |||
1706 | NewCurCond.clear(); | |||
1707 | TII->insertBranch(*PMBB, NewCurTBB, nullptr, NewCurCond, pdl); | |||
1708 | MadeChange = true; | |||
1709 | ++NumBranchOpts; | |||
1710 | PMBB->CorrectExtraCFGEdges(NewCurTBB, nullptr, false); | |||
1711 | } | |||
1712 | } | |||
1713 | } | |||
1714 | ||||
1715 | // Change any jumptables to go to the new MBB. | |||
1716 | if (MachineJumpTableInfo *MJTI = MF.getJumpTableInfo()) | |||
1717 | MJTI->ReplaceMBBInJumpTables(MBB, CurTBB); | |||
1718 | if (DidChange) { | |||
1719 | ++NumBranchOpts; | |||
1720 | MadeChange = true; | |||
1721 | if (!HasBranchToSelf) return MadeChange; | |||
1722 | } | |||
1723 | } | |||
1724 | } | |||
1725 | ||||
1726 | // Add the branch back if the block is more than just an uncond branch. | |||
1727 | TII->insertBranch(*MBB, CurTBB, nullptr, CurCond, dl); | |||
1728 | } | |||
1729 | } | |||
1730 | ||||
1731 | // If the prior block doesn't fall through into this block, and if this | |||
1732 | // block doesn't fall through into some other block, see if we can find a | |||
1733 | // place to move this block where a fall-through will happen. | |||
1734 | if (!PrevBB.canFallThrough()) { | |||
1735 | // Now we know that there was no fall-through into this block, check to | |||
1736 | // see if it has a fall-through into its successor. | |||
1737 | bool CurFallsThru = MBB->canFallThrough(); | |||
1738 | ||||
1739 | if (!MBB->isEHPad()) { | |||
1740 | // Check all the predecessors of this block. If one of them has no fall | |||
1741 | // throughs, move this block right after it. | |||
1742 | for (MachineBasicBlock *PredBB : MBB->predecessors()) { | |||
1743 | // Analyze the branch at the end of the pred. | |||
1744 | MachineBasicBlock *PredTBB = nullptr, *PredFBB = nullptr; | |||
1745 | SmallVector<MachineOperand, 4> PredCond; | |||
1746 | if (PredBB != MBB && !PredBB->canFallThrough() && | |||
1747 | !TII->analyzeBranch(*PredBB, PredTBB, PredFBB, PredCond, true) && | |||
1748 | (!CurFallsThru || !CurTBB || !CurFBB) && | |||
1749 | (!CurFallsThru || MBB->getNumber() >= PredBB->getNumber())) { | |||
1750 | // If the current block doesn't fall through, just move it. | |||
1751 | // If the current block can fall through and does not end with a | |||
1752 | // conditional branch, we need to append an unconditional jump to | |||
1753 | // the (current) next block. To avoid a possible compile-time | |||
1754 | // infinite loop, move blocks only backward in this case. | |||
1755 | // Also, if there are already 2 branches here, we cannot add a third; | |||
1756 | // this means we have the case | |||
1757 | // Bcc next | |||
1758 | // B elsewhere | |||
1759 | // next: | |||
1760 | if (CurFallsThru) { | |||
1761 | MachineBasicBlock *NextBB = &*std::next(MBB->getIterator()); | |||
1762 | CurCond.clear(); | |||
1763 | TII->insertBranch(*MBB, NextBB, nullptr, CurCond, DebugLoc()); | |||
1764 | } | |||
1765 | MBB->moveAfter(PredBB); | |||
1766 | MadeChange = true; | |||
1767 | goto ReoptimizeBlock; | |||
1768 | } | |||
1769 | } | |||
1770 | } | |||
1771 | ||||
1772 | if (!CurFallsThru) { | |||
1773 | // Check all successors to see if we can move this block before it. | |||
1774 | for (MachineBasicBlock *SuccBB : MBB->successors()) { | |||
1775 | // Analyze the branch at the end of the block before the succ. | |||
1776 | MachineFunction::iterator SuccPrev = --SuccBB->getIterator(); | |||
1777 | ||||
1778 | // If this block doesn't already fall-through to that successor, and if | |||
1779 | // the succ doesn't already have a block that can fall through into it, | |||
1780 | // and if the successor isn't an EH destination, we can arrange for the | |||
1781 | // fallthrough to happen. | |||
1782 | if (SuccBB != MBB && &*SuccPrev != MBB && | |||
1783 | !SuccPrev->canFallThrough() && !CurUnAnalyzable && | |||
1784 | !SuccBB->isEHPad()) { | |||
1785 | MBB->moveBefore(SuccBB); | |||
1786 | MadeChange = true; | |||
1787 | goto ReoptimizeBlock; | |||
1788 | } | |||
1789 | } | |||
1790 | ||||
1791 | // Okay, there is no really great place to put this block. If, however, | |||
1792 | // the block before this one would be a fall-through if this block were | |||
1793 | // removed, move this block to the end of the function. There is no real | |||
1794 | // advantage in "falling through" to an EH block, so we don't want to | |||
1795 | // perform this transformation for that case. | |||
1796 | // | |||
1797 | // Also, Windows EH introduced the possibility of an arbitrary number of | |||
1798 | // successors to a given block. The analyzeBranch call does not consider | |||
1799 | // exception handling and so we can get in a state where a block | |||
1800 | // containing a call is followed by multiple EH blocks that would be | |||
1801 | // rotated infinitely at the end of the function if the transformation | |||
1802 | // below were performed for EH "FallThrough" blocks. Therefore, even if | |||
1803 | // that appears not to be happening anymore, we should assume that it is | |||
1804 | // possible and not remove the "!FallThrough()->isEHPad" condition below. | |||
1805 | MachineBasicBlock *PrevTBB = nullptr, *PrevFBB = nullptr; | |||
1806 | SmallVector<MachineOperand, 4> PrevCond; | |||
1807 | if (FallThrough != MF.end() && | |||
1808 | !FallThrough->isEHPad() && | |||
1809 | !TII->analyzeBranch(PrevBB, PrevTBB, PrevFBB, PrevCond, true) && | |||
1810 | PrevBB.isSuccessor(&*FallThrough)) { | |||
1811 | MBB->moveAfter(&MF.back()); | |||
1812 | MadeChange = true; | |||
1813 | return MadeChange; | |||
1814 | } | |||
1815 | } | |||
1816 | } | |||
1817 | ||||
1818 | return MadeChange; | |||
1819 | } | |||
1820 | ||||
1821 | //===----------------------------------------------------------------------===// | |||
1822 | // Hoist Common Code | |||
1823 | //===----------------------------------------------------------------------===// | |||
1824 | ||||
1825 | bool BranchFolder::HoistCommonCode(MachineFunction &MF) { | |||
1826 | bool MadeChange = false; | |||
1827 | for (MachineFunction::iterator I = MF.begin(), E = MF.end(); I != E; ) { | |||
1828 | MachineBasicBlock *MBB = &*I++; | |||
1829 | MadeChange |= HoistCommonCodeInSuccs(MBB); | |||
1830 | } | |||
1831 | ||||
1832 | return MadeChange; | |||
1833 | } | |||
1834 | ||||
1835 | /// findFalseBlock - BB has a fallthrough. Find its 'false' successor given | |||
1836 | /// its 'true' successor. | |||
1837 | static MachineBasicBlock *findFalseBlock(MachineBasicBlock *BB, | |||
1838 | MachineBasicBlock *TrueBB) { | |||
1839 | for (MachineBasicBlock *SuccBB : BB->successors()) | |||
1840 | if (SuccBB != TrueBB) | |||
1841 | return SuccBB; | |||
1842 | return nullptr; | |||
1843 | } | |||
1844 | ||||
1845 | template <class Container> | |||
1846 | static void addRegAndItsAliases(unsigned Reg, const TargetRegisterInfo *TRI, | |||
1847 | Container &Set) { | |||
1848 | if (TargetRegisterInfo::isPhysicalRegister(Reg)) { | |||
1849 | for (MCRegAliasIterator AI(Reg, TRI, true); AI.isValid(); ++AI) | |||
1850 | Set.insert(*AI); | |||
1851 | } else { | |||
1852 | Set.insert(Reg); | |||
1853 | } | |||
1854 | } | |||
1855 | ||||
1856 | /// findHoistingInsertPosAndDeps - Find the location to move common instructions | |||
1857 | /// in successors to. The location is usually just before the terminator, | |||
1858 | /// however if the terminator is a conditional branch and its previous | |||
1859 | /// instruction is the flag setting instruction, the previous instruction is | |||
1860 | /// the preferred location. This function also gathers uses and defs of the | |||
1861 | /// instructions from the insertion point to the end of the block. The data is | |||
1862 | /// used by HoistCommonCodeInSuccs to ensure safety. | |||
1863 | static | |||
1864 | MachineBasicBlock::iterator findHoistingInsertPosAndDeps(MachineBasicBlock *MBB, | |||
1865 | const TargetInstrInfo *TII, | |||
1866 | const TargetRegisterInfo *TRI, | |||
1867 | SmallSet<unsigned,4> &Uses, | |||
1868 | SmallSet<unsigned,4> &Defs) { | |||
1869 | MachineBasicBlock::iterator Loc = MBB->getFirstTerminator(); | |||
1870 | if (!TII->isUnpredicatedTerminator(*Loc)) | |||
1871 | return MBB->end(); | |||
1872 | ||||
1873 | for (const MachineOperand &MO : Loc->operands()) { | |||
1874 | if (!MO.isReg()) | |||
1875 | continue; | |||
1876 | unsigned Reg = MO.getReg(); | |||
1877 | if (!Reg) | |||
1878 | continue; | |||
1879 | if (MO.isUse()) { | |||
1880 | addRegAndItsAliases(Reg, TRI, Uses); | |||
1881 | } else { | |||
1882 | if (!MO.isDead()) | |||
1883 | // Don't try to hoist code in the rare case the terminator defines a | |||
1884 | // register that is later used. | |||
1885 | return MBB->end(); | |||
1886 | ||||
1887 | // If the terminator defines a register, make sure we don't hoist | |||
1888 | // the instruction whose def might be clobbered by the terminator. | |||
1889 | addRegAndItsAliases(Reg, TRI, Defs); | |||
1890 | } | |||
1891 | } | |||
1892 | ||||
1893 | if (Uses.empty()) | |||
1894 | return Loc; | |||
1895 | // If the terminator is the only instruction in the block and Uses is not | |||
1896 | // empty (or we would have returned above), we can still safely hoist | |||
1897 | // instructions just before the terminator as long as the Defs/Uses are not | |||
1898 | // violated (which is checked in HoistCommonCodeInSuccs). | |||
1899 | if (Loc == MBB->begin()) | |||
1900 | return Loc; | |||
1901 | ||||
1902 | // The terminator is probably a conditional branch, try not to separate the | |||
1903 | // branch from condition setting instruction. | |||
1904 | MachineBasicBlock::iterator PI = | |||
1905 | skipDebugInstructionsBackward(std::prev(Loc), MBB->begin()); | |||
1906 | ||||
1907 | bool IsDef = false; | |||
1908 | for (const MachineOperand &MO : PI->operands()) { | |||
1909 | // If PI has a regmask operand, it is probably a call. Separate away. | |||
1910 | if (MO.isRegMask()) | |||
1911 | return Loc; | |||
1912 | if (!MO.isReg() || MO.isUse()) | |||
1913 | continue; | |||
1914 | unsigned Reg = MO.getReg(); | |||
1915 | if (!Reg) | |||
1916 | continue; | |||
1917 | if (Uses.count(Reg)) { | |||
1918 | IsDef = true; | |||
1919 | break; | |||
1920 | } | |||
1921 | } | |||
1922 | if (!IsDef) | |||
1923 | // The condition setting instruction is not just before the conditional | |||
1924 | // branch. | |||
1925 | return Loc; | |||
1926 | ||||
1927 | // Be conservative, don't insert instruction above something that may have | |||
1928 | // side-effects. And since it's potentially bad to separate flag setting | |||
1929 | // instruction from the conditional branch, just abort the optimization | |||
1930 | // completely. | |||
1931 | // Also avoid moving code above predicated instruction since it's hard to | |||
1932 | // reason about register liveness with predicated instruction. | |||
1933 | bool DontMoveAcrossStore = true; | |||
1934 | if (!PI->isSafeToMove(nullptr, DontMoveAcrossStore) || TII->isPredicated(*PI)) | |||
1935 | return MBB->end(); | |||
1936 | ||||
1937 | // Find out what registers are live. Note this routine is ignoring other live | |||
1938 | // registers which are only used by instructions in successor blocks. | |||
1939 | for (const MachineOperand &MO : PI->operands()) { | |||
1940 | if (!MO.isReg()) | |||
1941 | continue; | |||
1942 | unsigned Reg = MO.getReg(); | |||
1943 | if (!Reg) | |||
1944 | continue; | |||
1945 | if (MO.isUse()) { | |||
1946 | addRegAndItsAliases(Reg, TRI, Uses); | |||
1947 | } else { | |||
1948 | if (Uses.erase(Reg)) { | |||
1949 | if (TargetRegisterInfo::isPhysicalRegister(Reg)) { | |||
1950 | for (MCSubRegIterator SubRegs(Reg, TRI); SubRegs.isValid(); ++SubRegs) | |||
1951 | Uses.erase(*SubRegs); // Use sub-registers to be conservative | |||
1952 | } | |||
1953 | } | |||
1954 | addRegAndItsAliases(Reg, TRI, Defs); | |||
1955 | } | |||
1956 | } | |||
1957 | ||||
1958 | return PI; | |||
1959 | } | |||
1960 | ||||
1961 | bool BranchFolder::HoistCommonCodeInSuccs(MachineBasicBlock *MBB) { | |||
1962 | MachineBasicBlock *TBB = nullptr, *FBB = nullptr; | |||
1963 | SmallVector<MachineOperand, 4> Cond; | |||
1964 | if (TII->analyzeBranch(*MBB, TBB, FBB, Cond, true) || !TBB || Cond.empty()) | |||
1965 | return false; | |||
1966 | ||||
1967 | if (!FBB) FBB = findFalseBlock(MBB, TBB); | |||
1968 | if (!FBB) | |||
1969 | // Malformed bcc? True and false blocks are the same? | |||
1970 | return false; | |||
1971 | ||||
1972 | // Restrict the optimization to cases where MBB is the only predecessor, | |||
1973 | // it is an obvious win. | |||
1974 | if (TBB->pred_size() > 1 || FBB->pred_size() > 1) | |||
1975 | return false; | |||
1976 | ||||
1977 | // Find a suitable position to hoist the common instructions to. Also figure | |||
1978 | // out which registers are used or defined by instructions from the insertion | |||
1979 | // point to the end of the block. | |||
1980 | SmallSet<unsigned, 4> Uses, Defs; | |||
1981 | MachineBasicBlock::iterator Loc = | |||
1982 | findHoistingInsertPosAndDeps(MBB, TII, TRI, Uses, Defs); | |||
1983 | if (Loc == MBB->end()) | |||
1984 | return false; | |||
1985 | ||||
1986 | bool HasDups = false; | |||
1987 | SmallSet<unsigned, 4> ActiveDefsSet, AllDefsSet; | |||
1988 | MachineBasicBlock::iterator TIB = TBB->begin(); | |||
1989 | MachineBasicBlock::iterator FIB = FBB->begin(); | |||
1990 | MachineBasicBlock::iterator TIE = TBB->end(); | |||
1991 | MachineBasicBlock::iterator FIE = FBB->end(); | |||
1992 | while (TIB != TIE && FIB != FIE) { | |||
1993 | // Skip dbg_value instructions. These do not count. | |||
1994 | TIB = skipDebugInstructionsForward(TIB, TIE); | |||
1995 | FIB = skipDebugInstructionsForward(FIB, FIE); | |||
1996 | if (TIB == TIE || FIB == FIE) | |||
1997 | break; | |||
1998 | ||||
1999 | if (!TIB->isIdenticalTo(*FIB, MachineInstr::CheckKillDead)) | |||
2000 | break; | |||
2001 | ||||
2002 | if (TII->isPredicated(*TIB)) | |||
2003 | // Hard to reason about register liveness with predicated instruction. | |||
2004 | break; | |||
2005 | ||||
2006 | bool IsSafe = true; | |||
2007 | for (MachineOperand &MO : TIB->operands()) { | |||
2008 | // Don't attempt to hoist instructions with register masks. | |||
2009 | if (MO.isRegMask()) { | |||
2010 | IsSafe = false; | |||
2011 | break; | |||
2012 | } | |||
2013 | if (!MO.isReg()) | |||
2014 | continue; | |||
2015 | unsigned Reg = MO.getReg(); | |||
2016 | if (!Reg) | |||
2017 | continue; | |||
2018 | if (MO.isDef()) { | |||
2019 | if (Uses.count(Reg)) { | |||
2020 | // Avoid clobbering a register that's used by the instruction at | |||
2021 | // the point of insertion. | |||
2022 | IsSafe = false; | |||
2023 | break; | |||
2024 | } | |||
2025 | ||||
2026 | if (Defs.count(Reg) && !MO.isDead()) { | |||
2027 | // Don't hoist the instruction if the def would be clobber by the | |||
2028 | // instruction at the point insertion. FIXME: This is overly | |||
2029 | // conservative. It should be possible to hoist the instructions | |||
2030 | // in BB2 in the following example: | |||
2031 | // BB1: | |||
2032 | // r1, eflag = op1 r2, r3 | |||
2033 | // brcc eflag | |||
2034 | // | |||
2035 | // BB2: | |||
2036 | // r1 = op2, ... | |||
2037 | // = op3, killed r1 | |||
2038 | IsSafe = false; | |||
2039 | break; | |||
2040 | } | |||
2041 | } else if (!ActiveDefsSet.count(Reg)) { | |||
2042 | if (Defs.count(Reg)) { | |||
2043 | // Use is defined by the instruction at the point of insertion. | |||
2044 | IsSafe = false; | |||
2045 | break; | |||
2046 | } | |||
2047 | ||||
2048 | if (MO.isKill() && Uses.count(Reg)) | |||
2049 | // Kills a register that's read by the instruction at the point of | |||
2050 | // insertion. Remove the kill marker. | |||
2051 | MO.setIsKill(false); | |||
2052 | } | |||
2053 | } | |||
2054 | if (!IsSafe) | |||
2055 | break; | |||
2056 | ||||
2057 | bool DontMoveAcrossStore = true; | |||
2058 | if (!TIB->isSafeToMove(nullptr, DontMoveAcrossStore)) | |||
2059 | break; | |||
2060 | ||||
2061 | // Remove kills from ActiveDefsSet, these registers had short live ranges. | |||
2062 | for (const MachineOperand &MO : TIB->operands()) { | |||
2063 | if (!MO.isReg() || !MO.isUse() || !MO.isKill()) | |||
2064 | continue; | |||
2065 | unsigned Reg = MO.getReg(); | |||
2066 | if (!Reg) | |||
2067 | continue; | |||
2068 | if (!AllDefsSet.count(Reg)) { | |||
2069 | continue; | |||
2070 | } | |||
2071 | if (TargetRegisterInfo::isPhysicalRegister(Reg)) { | |||
2072 | for (MCRegAliasIterator AI(Reg, TRI, true); AI.isValid(); ++AI) | |||
2073 | ActiveDefsSet.erase(*AI); | |||
2074 | } else { | |||
2075 | ActiveDefsSet.erase(Reg); | |||
2076 | } | |||
2077 | } | |||
2078 | ||||
2079 | // Track local defs so we can update liveins. | |||
2080 | for (const MachineOperand &MO : TIB->operands()) { | |||
2081 | if (!MO.isReg() || !MO.isDef() || MO.isDead()) | |||
2082 | continue; | |||
2083 | unsigned Reg = MO.getReg(); | |||
2084 | if (!Reg || TargetRegisterInfo::isVirtualRegister(Reg)) | |||
2085 | continue; | |||
2086 | addRegAndItsAliases(Reg, TRI, ActiveDefsSet); | |||
2087 | addRegAndItsAliases(Reg, TRI, AllDefsSet); | |||
2088 | } | |||
2089 | ||||
2090 | HasDups = true; | |||
2091 | ++TIB; | |||
2092 | ++FIB; | |||
2093 | } | |||
2094 | ||||
2095 | if (!HasDups) | |||
2096 | return false; | |||
2097 | ||||
2098 | MBB->splice(Loc, TBB, TBB->begin(), TIB); | |||
2099 | FBB->erase(FBB->begin(), FIB); | |||
2100 | ||||
2101 | if (UpdateLiveIns) { | |||
2102 | recomputeLiveIns(*TBB); | |||
2103 | recomputeLiveIns(*FBB); | |||
2104 | } | |||
2105 | ||||
2106 | ++NumHoist; | |||
2107 | return true; | |||
2108 | } |