1 | //===- ARMConstantIslandPass.cpp - ARM constant islands -------------------===// |
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 file contains a pass that splits the constant pool up into 'islands' |
10 | // which are scattered through-out the function. This is required due to the |
11 | // limited pc-relative displacements that ARM has. |
12 | // |
13 | //===----------------------------------------------------------------------===// |
14 | |
15 | #include "ARM.h" |
16 | #include "ARMBaseInstrInfo.h" |
17 | #include "ARMBasicBlockInfo.h" |
18 | #include "ARMMachineFunctionInfo.h" |
19 | #include "ARMSubtarget.h" |
20 | #include "MCTargetDesc/ARMBaseInfo.h" |
21 | #include "MVETailPredUtils.h" |
22 | #include "Thumb2InstrInfo.h" |
23 | #include "Utils/ARMBaseInfo.h" |
24 | #include "llvm/ADT/DenseMap.h" |
25 | #include "llvm/ADT/STLExtras.h" |
26 | #include "llvm/ADT/SmallSet.h" |
27 | #include "llvm/ADT/SmallVector.h" |
28 | #include "llvm/ADT/Statistic.h" |
29 | #include "llvm/ADT/StringRef.h" |
30 | #include "llvm/CodeGen/LivePhysRegs.h" |
31 | #include "llvm/CodeGen/MachineBasicBlock.h" |
32 | #include "llvm/CodeGen/MachineConstantPool.h" |
33 | #include "llvm/CodeGen/MachineDominators.h" |
34 | #include "llvm/CodeGen/MachineFunction.h" |
35 | #include "llvm/CodeGen/MachineFunctionPass.h" |
36 | #include "llvm/CodeGen/MachineInstr.h" |
37 | #include "llvm/CodeGen/MachineJumpTableInfo.h" |
38 | #include "llvm/CodeGen/MachineOperand.h" |
39 | #include "llvm/CodeGen/MachineRegisterInfo.h" |
40 | #include "llvm/Config/llvm-config.h" |
41 | #include "llvm/IR/DataLayout.h" |
42 | #include "llvm/IR/DebugLoc.h" |
43 | #include "llvm/MC/MCInstrDesc.h" |
44 | #include "llvm/Pass.h" |
45 | #include "llvm/Support/CommandLine.h" |
46 | #include "llvm/Support/Compiler.h" |
47 | #include "llvm/Support/Debug.h" |
48 | #include "llvm/Support/ErrorHandling.h" |
49 | #include "llvm/Support/Format.h" |
50 | #include "llvm/Support/MathExtras.h" |
51 | #include "llvm/Support/raw_ostream.h" |
52 | #include <algorithm> |
53 | #include <cassert> |
54 | #include <cstdint> |
55 | #include <iterator> |
56 | #include <utility> |
57 | #include <vector> |
58 | |
59 | using namespace llvm; |
60 | |
61 | #define DEBUG_TYPE "arm-cp-islands" |
62 | |
63 | #define ARM_CP_ISLANDS_OPT_NAME \ |
64 | "ARM constant island placement and branch shortening pass" |
65 | STATISTIC(NumCPEs, "Number of constpool entries" ); |
66 | STATISTIC(NumSplit, "Number of uncond branches inserted" ); |
67 | STATISTIC(NumCBrFixed, "Number of cond branches fixed" ); |
68 | STATISTIC(NumUBrFixed, "Number of uncond branches fixed" ); |
69 | STATISTIC(NumTBs, "Number of table branches generated" ); |
70 | STATISTIC(NumT2CPShrunk, "Number of Thumb2 constantpool instructions shrunk" ); |
71 | STATISTIC(NumT2BrShrunk, "Number of Thumb2 immediate branches shrunk" ); |
72 | STATISTIC(NumCBZ, "Number of CBZ / CBNZ formed" ); |
73 | STATISTIC(NumJTMoved, "Number of jump table destination blocks moved" ); |
74 | STATISTIC(NumJTInserted, "Number of jump table intermediate blocks inserted" ); |
75 | STATISTIC(NumLEInserted, "Number of LE backwards branches inserted" ); |
76 | |
77 | static cl::opt<bool> |
78 | AdjustJumpTableBlocks("arm-adjust-jump-tables" , cl::Hidden, cl::init(Val: true), |
79 | cl::desc("Adjust basic block layout to better use TB[BH]" )); |
80 | |
81 | static cl::opt<unsigned> |
82 | CPMaxIteration("arm-constant-island-max-iteration" , cl::Hidden, cl::init(Val: 30), |
83 | cl::desc("The max number of iteration for converge" )); |
84 | |
85 | static cl::opt<bool> SynthesizeThumb1TBB( |
86 | "arm-synthesize-thumb-1-tbb" , cl::Hidden, cl::init(Val: true), |
87 | cl::desc("Use compressed jump tables in Thumb-1 by synthesizing an " |
88 | "equivalent to the TBB/TBH instructions" )); |
89 | |
90 | namespace { |
91 | |
92 | /// ARMConstantIslands - Due to limited PC-relative displacements, ARM |
93 | /// requires constant pool entries to be scattered among the instructions |
94 | /// inside a function. To do this, it completely ignores the normal LLVM |
95 | /// constant pool; instead, it places constants wherever it feels like with |
96 | /// special instructions. |
97 | /// |
98 | /// The terminology used in this pass includes: |
99 | /// Islands - Clumps of constants placed in the function. |
100 | /// Water - Potential places where an island could be formed. |
101 | /// CPE - A constant pool entry that has been placed somewhere, which |
102 | /// tracks a list of users. |
103 | class ARMConstantIslands : public MachineFunctionPass { |
104 | std::unique_ptr<ARMBasicBlockUtils> BBUtils = nullptr; |
105 | |
106 | /// WaterList - A sorted list of basic blocks where islands could be placed |
107 | /// (i.e. blocks that don't fall through to the following block, due |
108 | /// to a return, unreachable, or unconditional branch). |
109 | std::vector<MachineBasicBlock*> WaterList; |
110 | |
111 | /// NewWaterList - The subset of WaterList that was created since the |
112 | /// previous iteration by inserting unconditional branches. |
113 | SmallSet<MachineBasicBlock*, 4> NewWaterList; |
114 | |
115 | using water_iterator = std::vector<MachineBasicBlock *>::iterator; |
116 | |
117 | /// CPUser - One user of a constant pool, keeping the machine instruction |
118 | /// pointer, the constant pool being referenced, and the max displacement |
119 | /// allowed from the instruction to the CP. The HighWaterMark records the |
120 | /// highest basic block where a new CPEntry can be placed. To ensure this |
121 | /// pass terminates, the CP entries are initially placed at the end of the |
122 | /// function and then move monotonically to lower addresses. The |
123 | /// exception to this rule is when the current CP entry for a particular |
124 | /// CPUser is out of range, but there is another CP entry for the same |
125 | /// constant value in range. We want to use the existing in-range CP |
126 | /// entry, but if it later moves out of range, the search for new water |
127 | /// should resume where it left off. The HighWaterMark is used to record |
128 | /// that point. |
129 | struct CPUser { |
130 | MachineInstr *MI; |
131 | MachineInstr *CPEMI; |
132 | MachineBasicBlock *HighWaterMark; |
133 | unsigned MaxDisp; |
134 | bool NegOk; |
135 | bool IsSoImm; |
136 | bool KnownAlignment = false; |
137 | |
138 | CPUser(MachineInstr *mi, MachineInstr *cpemi, unsigned maxdisp, |
139 | bool neg, bool soimm) |
140 | : MI(mi), CPEMI(cpemi), MaxDisp(maxdisp), NegOk(neg), IsSoImm(soimm) { |
141 | HighWaterMark = CPEMI->getParent(); |
142 | } |
143 | |
144 | /// getMaxDisp - Returns the maximum displacement supported by MI. |
145 | /// Correct for unknown alignment. |
146 | /// Conservatively subtract 2 bytes to handle weird alignment effects. |
147 | unsigned getMaxDisp() const { |
148 | return (KnownAlignment ? MaxDisp : MaxDisp - 2) - 2; |
149 | } |
150 | }; |
151 | |
152 | /// CPUsers - Keep track of all of the machine instructions that use various |
153 | /// constant pools and their max displacement. |
154 | std::vector<CPUser> CPUsers; |
155 | |
156 | /// CPEntry - One per constant pool entry, keeping the machine instruction |
157 | /// pointer, the constpool index, and the number of CPUser's which |
158 | /// reference this entry. |
159 | struct CPEntry { |
160 | MachineInstr *CPEMI; |
161 | unsigned CPI; |
162 | unsigned RefCount; |
163 | |
164 | CPEntry(MachineInstr *cpemi, unsigned cpi, unsigned rc = 0) |
165 | : CPEMI(cpemi), CPI(cpi), RefCount(rc) {} |
166 | }; |
167 | |
168 | /// CPEntries - Keep track of all of the constant pool entry machine |
169 | /// instructions. For each original constpool index (i.e. those that existed |
170 | /// upon entry to this pass), it keeps a vector of entries. Original |
171 | /// elements are cloned as we go along; the clones are put in the vector of |
172 | /// the original element, but have distinct CPIs. |
173 | /// |
174 | /// The first half of CPEntries contains generic constants, the second half |
175 | /// contains jump tables. Use getCombinedIndex on a generic CPEMI to look up |
176 | /// which vector it will be in here. |
177 | std::vector<std::vector<CPEntry>> CPEntries; |
178 | |
179 | /// Maps a JT index to the offset in CPEntries containing copies of that |
180 | /// table. The equivalent map for a CONSTPOOL_ENTRY is the identity. |
181 | DenseMap<int, int> JumpTableEntryIndices; |
182 | |
183 | /// Maps a JT index to the LEA that actually uses the index to calculate its |
184 | /// base address. |
185 | DenseMap<int, int> JumpTableUserIndices; |
186 | |
187 | // Maps a MachineBasicBlock to the number of jump tables entries. |
188 | DenseMap<const MachineBasicBlock *, int> BlockJumpTableRefCount; |
189 | |
190 | /// ImmBranch - One per immediate branch, keeping the machine instruction |
191 | /// pointer, conditional or unconditional, the max displacement, |
192 | /// and (if isCond is true) the corresponding unconditional branch |
193 | /// opcode. |
194 | struct ImmBranch { |
195 | MachineInstr *MI; |
196 | unsigned MaxDisp : 31; |
197 | bool isCond : 1; |
198 | unsigned UncondBr; |
199 | |
200 | ImmBranch(MachineInstr *mi, unsigned maxdisp, bool cond, unsigned ubr) |
201 | : MI(mi), MaxDisp(maxdisp), isCond(cond), UncondBr(ubr) {} |
202 | }; |
203 | |
204 | /// ImmBranches - Keep track of all the immediate branch instructions. |
205 | std::vector<ImmBranch> ImmBranches; |
206 | |
207 | /// PushPopMIs - Keep track of all the Thumb push / pop instructions. |
208 | SmallVector<MachineInstr*, 4> PushPopMIs; |
209 | |
210 | /// T2JumpTables - Keep track of all the Thumb2 jumptable instructions. |
211 | SmallVector<MachineInstr*, 4> T2JumpTables; |
212 | |
213 | MachineFunction *MF; |
214 | MachineConstantPool *MCP; |
215 | const ARMBaseInstrInfo *TII; |
216 | const ARMSubtarget *STI; |
217 | ARMFunctionInfo *AFI; |
218 | MachineDominatorTree *DT = nullptr; |
219 | bool isThumb; |
220 | bool isThumb1; |
221 | bool isThumb2; |
222 | bool isPositionIndependentOrROPI; |
223 | |
224 | public: |
225 | static char ID; |
226 | |
227 | ARMConstantIslands() : MachineFunctionPass(ID) {} |
228 | |
229 | bool runOnMachineFunction(MachineFunction &MF) override; |
230 | |
231 | void getAnalysisUsage(AnalysisUsage &AU) const override { |
232 | AU.addRequired<MachineDominatorTreeWrapperPass>(); |
233 | MachineFunctionPass::getAnalysisUsage(AU); |
234 | } |
235 | |
236 | MachineFunctionProperties getRequiredProperties() const override { |
237 | return MachineFunctionProperties().set( |
238 | MachineFunctionProperties::Property::NoVRegs); |
239 | } |
240 | |
241 | StringRef getPassName() const override { |
242 | return ARM_CP_ISLANDS_OPT_NAME; |
243 | } |
244 | |
245 | private: |
246 | void doInitialConstPlacement(std::vector<MachineInstr *> &CPEMIs); |
247 | void doInitialJumpTablePlacement(std::vector<MachineInstr *> &CPEMIs); |
248 | bool BBHasFallthrough(MachineBasicBlock *MBB); |
249 | CPEntry *findConstPoolEntry(unsigned CPI, const MachineInstr *CPEMI); |
250 | Align getCPEAlign(const MachineInstr *CPEMI); |
251 | void scanFunctionJumpTables(); |
252 | void initializeFunctionInfo(const std::vector<MachineInstr*> &CPEMIs); |
253 | MachineBasicBlock *splitBlockBeforeInstr(MachineInstr *MI); |
254 | void updateForInsertedWaterBlock(MachineBasicBlock *NewBB); |
255 | bool decrementCPEReferenceCount(unsigned CPI, MachineInstr* CPEMI); |
256 | unsigned getCombinedIndex(const MachineInstr *CPEMI); |
257 | int findInRangeCPEntry(CPUser& U, unsigned UserOffset); |
258 | bool findAvailableWater(CPUser&U, unsigned UserOffset, |
259 | water_iterator &WaterIter, bool CloserWater); |
260 | void createNewWater(unsigned CPUserIndex, unsigned UserOffset, |
261 | MachineBasicBlock *&NewMBB); |
262 | bool handleConstantPoolUser(unsigned CPUserIndex, bool CloserWater); |
263 | void removeDeadCPEMI(MachineInstr *CPEMI); |
264 | bool removeUnusedCPEntries(); |
265 | bool isCPEntryInRange(MachineInstr *MI, unsigned UserOffset, |
266 | MachineInstr *CPEMI, unsigned Disp, bool NegOk, |
267 | bool DoDump = false); |
268 | bool isWaterInRange(unsigned UserOffset, MachineBasicBlock *Water, |
269 | CPUser &U, unsigned &Growth); |
270 | bool fixupImmediateBr(ImmBranch &Br); |
271 | bool fixupConditionalBr(ImmBranch &Br); |
272 | bool fixupUnconditionalBr(ImmBranch &Br); |
273 | bool optimizeThumb2Instructions(); |
274 | bool optimizeThumb2Branches(); |
275 | bool reorderThumb2JumpTables(); |
276 | bool preserveBaseRegister(MachineInstr *JumpMI, MachineInstr *LEAMI, |
277 | unsigned &DeadSize, bool &CanDeleteLEA, |
278 | bool &BaseRegKill); |
279 | bool optimizeThumb2JumpTables(); |
280 | MachineBasicBlock *adjustJTTargetBlockForward(unsigned JTI, |
281 | MachineBasicBlock *BB, |
282 | MachineBasicBlock *JTBB); |
283 | |
284 | unsigned getUserOffset(CPUser&) const; |
285 | void dumpBBs(); |
286 | void verify(); |
287 | |
288 | bool isOffsetInRange(unsigned UserOffset, unsigned TrialOffset, |
289 | unsigned Disp, bool NegativeOK, bool IsSoImm = false); |
290 | bool isOffsetInRange(unsigned UserOffset, unsigned TrialOffset, |
291 | const CPUser &U) { |
292 | return isOffsetInRange(UserOffset, TrialOffset, |
293 | Disp: U.getMaxDisp(), NegativeOK: U.NegOk, IsSoImm: U.IsSoImm); |
294 | } |
295 | }; |
296 | |
297 | } // end anonymous namespace |
298 | |
299 | char ARMConstantIslands::ID = 0; |
300 | |
301 | /// verify - check BBOffsets, BBSizes, alignment of islands |
302 | void ARMConstantIslands::verify() { |
303 | #ifndef NDEBUG |
304 | BBInfoVector &BBInfo = BBUtils->getBBInfo(); |
305 | assert(is_sorted(*MF, [&BBInfo](const MachineBasicBlock &LHS, |
306 | const MachineBasicBlock &RHS) { |
307 | return BBInfo[LHS.getNumber()].postOffset() < |
308 | BBInfo[RHS.getNumber()].postOffset(); |
309 | })); |
310 | LLVM_DEBUG(dbgs() << "Verifying " << CPUsers.size() << " CP users.\n" ); |
311 | for (CPUser &U : CPUsers) { |
312 | unsigned UserOffset = getUserOffset(U); |
313 | // Verify offset using the real max displacement without the safety |
314 | // adjustment. |
315 | if (isCPEntryInRange(U.MI, UserOffset, U.CPEMI, U.getMaxDisp()+2, U.NegOk, |
316 | /* DoDump = */ true)) { |
317 | LLVM_DEBUG(dbgs() << "OK\n" ); |
318 | continue; |
319 | } |
320 | LLVM_DEBUG(dbgs() << "Out of range.\n" ); |
321 | dumpBBs(); |
322 | LLVM_DEBUG(MF->dump()); |
323 | llvm_unreachable("Constant pool entry out of range!" ); |
324 | } |
325 | #endif |
326 | } |
327 | |
328 | #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) |
329 | /// print block size and offset information - debugging |
330 | LLVM_DUMP_METHOD void ARMConstantIslands::dumpBBs() { |
331 | LLVM_DEBUG({ |
332 | BBInfoVector &BBInfo = BBUtils->getBBInfo(); |
333 | for (unsigned J = 0, E = BBInfo.size(); J !=E; ++J) { |
334 | const BasicBlockInfo &BBI = BBInfo[J]; |
335 | dbgs() << format("%08x %bb.%u\t" , BBI.Offset, J) |
336 | << " kb=" << unsigned(BBI.KnownBits) |
337 | << " ua=" << unsigned(BBI.Unalign) << " pa=" << Log2(BBI.PostAlign) |
338 | << format(" size=%#x\n" , BBInfo[J].Size); |
339 | } |
340 | }); |
341 | } |
342 | #endif |
343 | |
344 | // Align blocks where the previous block does not fall through. This may add |
345 | // extra NOP's but they will not be executed. It uses the PrefLoopAlignment as a |
346 | // measure of how much to align, and only runs at CodeGenOptLevel::Aggressive. |
347 | static bool AlignBlocks(MachineFunction *MF, const ARMSubtarget *STI) { |
348 | if (MF->getTarget().getOptLevel() != CodeGenOptLevel::Aggressive || |
349 | MF->getFunction().hasOptSize()) |
350 | return false; |
351 | |
352 | auto *TLI = STI->getTargetLowering(); |
353 | const Align Alignment = TLI->getPrefLoopAlignment(); |
354 | if (Alignment < 4) |
355 | return false; |
356 | |
357 | bool Changed = false; |
358 | bool PrevCanFallthough = true; |
359 | for (auto &MBB : *MF) { |
360 | if (!PrevCanFallthough) { |
361 | Changed = true; |
362 | MBB.setAlignment(Alignment); |
363 | } |
364 | |
365 | PrevCanFallthough = MBB.canFallThrough(); |
366 | |
367 | // For LOB's, the ARMLowOverheadLoops pass may remove the unconditional |
368 | // branch later in the pipeline. |
369 | if (STI->hasLOB()) { |
370 | for (const auto &MI : reverse(C: MBB.terminators())) { |
371 | if (MI.getOpcode() == ARM::t2B && |
372 | MI.getOperand(i: 0).getMBB() == MBB.getNextNode()) |
373 | continue; |
374 | if (isLoopStart(MI) || MI.getOpcode() == ARM::t2LoopEnd || |
375 | MI.getOpcode() == ARM::t2LoopEndDec) { |
376 | PrevCanFallthough = true; |
377 | break; |
378 | } |
379 | // Any other terminator - nothing to do |
380 | break; |
381 | } |
382 | } |
383 | } |
384 | |
385 | return Changed; |
386 | } |
387 | |
388 | bool ARMConstantIslands::runOnMachineFunction(MachineFunction &mf) { |
389 | MF = &mf; |
390 | MCP = mf.getConstantPool(); |
391 | BBUtils = std::make_unique<ARMBasicBlockUtils>(args&: mf); |
392 | |
393 | LLVM_DEBUG(dbgs() << "***** ARMConstantIslands: " |
394 | << MCP->getConstants().size() << " CP entries, aligned to " |
395 | << MCP->getConstantPoolAlign().value() << " bytes *****\n" ); |
396 | |
397 | STI = &MF->getSubtarget<ARMSubtarget>(); |
398 | TII = STI->getInstrInfo(); |
399 | isPositionIndependentOrROPI = |
400 | STI->getTargetLowering()->isPositionIndependent() || STI->isROPI(); |
401 | AFI = MF->getInfo<ARMFunctionInfo>(); |
402 | DT = &getAnalysis<MachineDominatorTreeWrapperPass>().getDomTree(); |
403 | |
404 | isThumb = AFI->isThumbFunction(); |
405 | isThumb1 = AFI->isThumb1OnlyFunction(); |
406 | isThumb2 = AFI->isThumb2Function(); |
407 | |
408 | bool GenerateTBB = isThumb2 || (isThumb1 && SynthesizeThumb1TBB); |
409 | // TBB generation code in this constant island pass has not been adapted to |
410 | // deal with speculation barriers. |
411 | if (STI->hardenSlsRetBr()) |
412 | GenerateTBB = false; |
413 | |
414 | // Renumber all of the machine basic blocks in the function, guaranteeing that |
415 | // the numbers agree with the position of the block in the function. |
416 | MF->RenumberBlocks(); |
417 | |
418 | // Try to reorder and otherwise adjust the block layout to make good use |
419 | // of the TB[BH] instructions. |
420 | bool MadeChange = false; |
421 | if (GenerateTBB && AdjustJumpTableBlocks) { |
422 | scanFunctionJumpTables(); |
423 | MadeChange |= reorderThumb2JumpTables(); |
424 | // Data is out of date, so clear it. It'll be re-computed later. |
425 | T2JumpTables.clear(); |
426 | // Blocks may have shifted around. Keep the numbering up to date. |
427 | MF->RenumberBlocks(); |
428 | } |
429 | |
430 | // Align any non-fallthrough blocks |
431 | MadeChange |= AlignBlocks(MF, STI); |
432 | |
433 | // Perform the initial placement of the constant pool entries. To start with, |
434 | // we put them all at the end of the function. |
435 | std::vector<MachineInstr*> CPEMIs; |
436 | if (!MCP->isEmpty()) |
437 | doInitialConstPlacement(CPEMIs); |
438 | |
439 | if (MF->getJumpTableInfo()) |
440 | doInitialJumpTablePlacement(CPEMIs); |
441 | |
442 | /// The next UID to take is the first unused one. |
443 | AFI->initPICLabelUId(UId: CPEMIs.size()); |
444 | |
445 | // Do the initial scan of the function, building up information about the |
446 | // sizes of each block, the location of all the water, and finding all of the |
447 | // constant pool users. |
448 | initializeFunctionInfo(CPEMIs); |
449 | CPEMIs.clear(); |
450 | LLVM_DEBUG(dumpBBs()); |
451 | |
452 | // Functions with jump tables need an alignment of 4 because they use the ADR |
453 | // instruction, which aligns the PC to 4 bytes before adding an offset. |
454 | if (!T2JumpTables.empty()) |
455 | MF->ensureAlignment(A: Align(4)); |
456 | |
457 | /// Remove dead constant pool entries. |
458 | MadeChange |= removeUnusedCPEntries(); |
459 | |
460 | // Iteratively place constant pool entries and fix up branches until there |
461 | // is no change. |
462 | unsigned NoCPIters = 0, NoBRIters = 0; |
463 | while (true) { |
464 | LLVM_DEBUG(dbgs() << "Beginning CP iteration #" << NoCPIters << '\n'); |
465 | bool CPChange = false; |
466 | for (unsigned i = 0, e = CPUsers.size(); i != e; ++i) |
467 | // For most inputs, it converges in no more than 5 iterations. |
468 | // If it doesn't end in 10, the input may have huge BB or many CPEs. |
469 | // In this case, we will try different heuristics. |
470 | CPChange |= handleConstantPoolUser(CPUserIndex: i, CloserWater: NoCPIters >= CPMaxIteration / 2); |
471 | if (CPChange && ++NoCPIters > CPMaxIteration) |
472 | report_fatal_error(reason: "Constant Island pass failed to converge!" ); |
473 | LLVM_DEBUG(dumpBBs()); |
474 | |
475 | // Clear NewWaterList now. If we split a block for branches, it should |
476 | // appear as "new water" for the next iteration of constant pool placement. |
477 | NewWaterList.clear(); |
478 | |
479 | LLVM_DEBUG(dbgs() << "Beginning BR iteration #" << NoBRIters << '\n'); |
480 | bool BRChange = false; |
481 | for (unsigned i = 0, e = ImmBranches.size(); i != e; ++i) |
482 | BRChange |= fixupImmediateBr(Br&: ImmBranches[i]); |
483 | if (BRChange && ++NoBRIters > 30) |
484 | report_fatal_error(reason: "Branch Fix Up pass failed to converge!" ); |
485 | LLVM_DEBUG(dumpBBs()); |
486 | |
487 | if (!CPChange && !BRChange) |
488 | break; |
489 | MadeChange = true; |
490 | } |
491 | |
492 | // Shrink 32-bit Thumb2 load and store instructions. |
493 | if (isThumb2 && !STI->prefers32BitThumb()) |
494 | MadeChange |= optimizeThumb2Instructions(); |
495 | |
496 | // Shrink 32-bit branch instructions. |
497 | if (isThumb && STI->hasV8MBaselineOps()) |
498 | MadeChange |= optimizeThumb2Branches(); |
499 | |
500 | // Optimize jump tables using TBB / TBH. |
501 | if (GenerateTBB && !STI->genExecuteOnly()) |
502 | MadeChange |= optimizeThumb2JumpTables(); |
503 | |
504 | // After a while, this might be made debug-only, but it is not expensive. |
505 | verify(); |
506 | |
507 | // Save the mapping between original and cloned constpool entries. |
508 | for (unsigned i = 0, e = CPEntries.size(); i != e; ++i) { |
509 | for (unsigned j = 0, je = CPEntries[i].size(); j != je; ++j) { |
510 | const CPEntry & CPE = CPEntries[i][j]; |
511 | if (CPE.CPEMI && CPE.CPEMI->getOperand(i: 1).isCPI()) |
512 | AFI->recordCPEClone(CPIdx: i, CPCloneIdx: CPE.CPI); |
513 | } |
514 | } |
515 | |
516 | LLVM_DEBUG(dbgs() << '\n'; dumpBBs()); |
517 | |
518 | BBUtils->clear(); |
519 | WaterList.clear(); |
520 | CPUsers.clear(); |
521 | CPEntries.clear(); |
522 | JumpTableEntryIndices.clear(); |
523 | JumpTableUserIndices.clear(); |
524 | BlockJumpTableRefCount.clear(); |
525 | ImmBranches.clear(); |
526 | PushPopMIs.clear(); |
527 | T2JumpTables.clear(); |
528 | |
529 | return MadeChange; |
530 | } |
531 | |
532 | /// Perform the initial placement of the regular constant pool entries. |
533 | /// To start with, we put them all at the end of the function. |
534 | void |
535 | ARMConstantIslands::doInitialConstPlacement(std::vector<MachineInstr*> &CPEMIs) { |
536 | // Create the basic block to hold the CPE's. |
537 | MachineBasicBlock *BB = MF->CreateMachineBasicBlock(); |
538 | MF->push_back(MBB: BB); |
539 | |
540 | // MachineConstantPool measures alignment in bytes. |
541 | const Align MaxAlign = MCP->getConstantPoolAlign(); |
542 | const unsigned MaxLogAlign = Log2(A: MaxAlign); |
543 | |
544 | // Mark the basic block as required by the const-pool. |
545 | BB->setAlignment(MaxAlign); |
546 | |
547 | // The function needs to be as aligned as the basic blocks. The linker may |
548 | // move functions around based on their alignment. |
549 | // Special case: halfword literals still need word alignment on the function. |
550 | Align FuncAlign = MaxAlign; |
551 | if (MaxAlign == 2) |
552 | FuncAlign = Align(4); |
553 | MF->ensureAlignment(A: FuncAlign); |
554 | |
555 | // Order the entries in BB by descending alignment. That ensures correct |
556 | // alignment of all entries as long as BB is sufficiently aligned. Keep |
557 | // track of the insertion point for each alignment. We are going to bucket |
558 | // sort the entries as they are created. |
559 | SmallVector<MachineBasicBlock::iterator, 8> InsPoint(MaxLogAlign + 1, |
560 | BB->end()); |
561 | |
562 | // Add all of the constants from the constant pool to the end block, use an |
563 | // identity mapping of CPI's to CPE's. |
564 | const std::vector<MachineConstantPoolEntry> &CPs = MCP->getConstants(); |
565 | |
566 | const DataLayout &TD = MF->getDataLayout(); |
567 | for (unsigned i = 0, e = CPs.size(); i != e; ++i) { |
568 | unsigned Size = CPs[i].getSizeInBytes(DL: TD); |
569 | Align Alignment = CPs[i].getAlign(); |
570 | // Verify that all constant pool entries are a multiple of their alignment. |
571 | // If not, we would have to pad them out so that instructions stay aligned. |
572 | assert(isAligned(Alignment, Size) && "CP Entry not multiple of 4 bytes!" ); |
573 | |
574 | // Insert CONSTPOOL_ENTRY before entries with a smaller alignment. |
575 | unsigned LogAlign = Log2(A: Alignment); |
576 | MachineBasicBlock::iterator InsAt = InsPoint[LogAlign]; |
577 | MachineInstr *CPEMI = |
578 | BuildMI(BB&: *BB, I: InsAt, MIMD: DebugLoc(), MCID: TII->get(Opcode: ARM::CONSTPOOL_ENTRY)) |
579 | .addImm(Val: i).addConstantPoolIndex(Idx: i).addImm(Val: Size); |
580 | CPEMIs.push_back(x: CPEMI); |
581 | |
582 | // Ensure that future entries with higher alignment get inserted before |
583 | // CPEMI. This is bucket sort with iterators. |
584 | for (unsigned a = LogAlign + 1; a <= MaxLogAlign; ++a) |
585 | if (InsPoint[a] == InsAt) |
586 | InsPoint[a] = CPEMI; |
587 | |
588 | // Add a new CPEntry, but no corresponding CPUser yet. |
589 | CPEntries.emplace_back(args: 1, args: CPEntry(CPEMI, i)); |
590 | ++NumCPEs; |
591 | LLVM_DEBUG(dbgs() << "Moved CPI#" << i << " to end of function, size = " |
592 | << Size << ", align = " << Alignment.value() << '\n'); |
593 | } |
594 | LLVM_DEBUG(BB->dump()); |
595 | } |
596 | |
597 | /// Do initial placement of the jump tables. Because Thumb2's TBB and TBH |
598 | /// instructions can be made more efficient if the jump table immediately |
599 | /// follows the instruction, it's best to place them immediately next to their |
600 | /// jumps to begin with. In almost all cases they'll never be moved from that |
601 | /// position. |
602 | void ARMConstantIslands::doInitialJumpTablePlacement( |
603 | std::vector<MachineInstr *> &CPEMIs) { |
604 | unsigned i = CPEntries.size(); |
605 | auto MJTI = MF->getJumpTableInfo(); |
606 | const std::vector<MachineJumpTableEntry> &JT = MJTI->getJumpTables(); |
607 | |
608 | // Only inline jump tables are placed in the function. |
609 | if (MJTI->getEntryKind() != MachineJumpTableInfo::EK_Inline) |
610 | return; |
611 | |
612 | MachineBasicBlock *LastCorrectlyNumberedBB = nullptr; |
613 | for (MachineBasicBlock &MBB : *MF) { |
614 | auto MI = MBB.getLastNonDebugInstr(); |
615 | // Look past potential SpeculationBarriers at end of BB. |
616 | while (MI != MBB.end() && |
617 | (isSpeculationBarrierEndBBOpcode(Opc: MI->getOpcode()) || |
618 | MI->isDebugInstr())) |
619 | --MI; |
620 | |
621 | if (MI == MBB.end()) |
622 | continue; |
623 | |
624 | unsigned JTOpcode; |
625 | switch (MI->getOpcode()) { |
626 | default: |
627 | continue; |
628 | case ARM::BR_JTadd: |
629 | case ARM::BR_JTr: |
630 | case ARM::tBR_JTr: |
631 | case ARM::BR_JTm_i12: |
632 | case ARM::BR_JTm_rs: |
633 | // These instructions are emitted only in ARM or Thumb1 modes which do not |
634 | // support PACBTI. Hence we don't add BTI instructions in the destination |
635 | // blocks. |
636 | assert(!MF->getInfo<ARMFunctionInfo>()->branchTargetEnforcement() && |
637 | "Branch protection must not be enabled for Arm or Thumb1 modes" ); |
638 | JTOpcode = ARM::JUMPTABLE_ADDRS; |
639 | break; |
640 | case ARM::t2BR_JT: |
641 | JTOpcode = ARM::JUMPTABLE_INSTS; |
642 | break; |
643 | case ARM::tTBB_JT: |
644 | case ARM::t2TBB_JT: |
645 | JTOpcode = ARM::JUMPTABLE_TBB; |
646 | break; |
647 | case ARM::tTBH_JT: |
648 | case ARM::t2TBH_JT: |
649 | JTOpcode = ARM::JUMPTABLE_TBH; |
650 | break; |
651 | } |
652 | |
653 | unsigned NumOps = MI->getDesc().getNumOperands(); |
654 | MachineOperand JTOp = |
655 | MI->getOperand(i: NumOps - (MI->isPredicable() ? 2 : 1)); |
656 | unsigned JTI = JTOp.getIndex(); |
657 | unsigned Size = JT[JTI].MBBs.size() * sizeof(uint32_t); |
658 | MachineBasicBlock *JumpTableBB = MF->CreateMachineBasicBlock(); |
659 | MF->insert(MBBI: std::next(x: MachineFunction::iterator(MBB)), MBB: JumpTableBB); |
660 | MachineInstr *CPEMI = BuildMI(BB&: *JumpTableBB, I: JumpTableBB->begin(), |
661 | MIMD: DebugLoc(), MCID: TII->get(Opcode: JTOpcode)) |
662 | .addImm(Val: i++) |
663 | .addJumpTableIndex(Idx: JTI) |
664 | .addImm(Val: Size); |
665 | CPEMIs.push_back(x: CPEMI); |
666 | CPEntries.emplace_back(args: 1, args: CPEntry(CPEMI, JTI)); |
667 | JumpTableEntryIndices.insert(KV: std::make_pair(x&: JTI, y: CPEntries.size() - 1)); |
668 | if (!LastCorrectlyNumberedBB) |
669 | LastCorrectlyNumberedBB = &MBB; |
670 | } |
671 | |
672 | // If we did anything then we need to renumber the subsequent blocks. |
673 | if (LastCorrectlyNumberedBB) |
674 | MF->RenumberBlocks(MBBFrom: LastCorrectlyNumberedBB); |
675 | } |
676 | |
677 | /// BBHasFallthrough - Return true if the specified basic block can fallthrough |
678 | /// into the block immediately after it. |
679 | bool ARMConstantIslands::BBHasFallthrough(MachineBasicBlock *MBB) { |
680 | // Get the next machine basic block in the function. |
681 | MachineFunction::iterator MBBI = MBB->getIterator(); |
682 | // Can't fall off end of function. |
683 | if (std::next(x: MBBI) == MBB->getParent()->end()) |
684 | return false; |
685 | |
686 | MachineBasicBlock *NextBB = &*std::next(x: MBBI); |
687 | if (!MBB->isSuccessor(MBB: NextBB)) |
688 | return false; |
689 | |
690 | // Try to analyze the end of the block. A potential fallthrough may already |
691 | // have an unconditional branch for whatever reason. |
692 | MachineBasicBlock *TBB, *FBB; |
693 | SmallVector<MachineOperand, 4> Cond; |
694 | bool TooDifficult = TII->analyzeBranch(MBB&: *MBB, TBB, FBB, Cond); |
695 | return TooDifficult || FBB == nullptr; |
696 | } |
697 | |
698 | /// findConstPoolEntry - Given the constpool index and CONSTPOOL_ENTRY MI, |
699 | /// look up the corresponding CPEntry. |
700 | ARMConstantIslands::CPEntry * |
701 | ARMConstantIslands::findConstPoolEntry(unsigned CPI, |
702 | const MachineInstr *CPEMI) { |
703 | std::vector<CPEntry> &CPEs = CPEntries[CPI]; |
704 | // Number of entries per constpool index should be small, just do a |
705 | // linear search. |
706 | for (CPEntry &CPE : CPEs) |
707 | if (CPE.CPEMI == CPEMI) |
708 | return &CPE; |
709 | return nullptr; |
710 | } |
711 | |
712 | /// getCPEAlign - Returns the required alignment of the constant pool entry |
713 | /// represented by CPEMI. |
714 | Align ARMConstantIslands::getCPEAlign(const MachineInstr *CPEMI) { |
715 | switch (CPEMI->getOpcode()) { |
716 | case ARM::CONSTPOOL_ENTRY: |
717 | break; |
718 | case ARM::JUMPTABLE_TBB: |
719 | return isThumb1 ? Align(4) : Align(1); |
720 | case ARM::JUMPTABLE_TBH: |
721 | return isThumb1 ? Align(4) : Align(2); |
722 | case ARM::JUMPTABLE_INSTS: |
723 | return Align(2); |
724 | case ARM::JUMPTABLE_ADDRS: |
725 | return Align(4); |
726 | default: |
727 | llvm_unreachable("unknown constpool entry kind" ); |
728 | } |
729 | |
730 | unsigned CPI = getCombinedIndex(CPEMI); |
731 | assert(CPI < MCP->getConstants().size() && "Invalid constant pool index." ); |
732 | return MCP->getConstants()[CPI].getAlign(); |
733 | } |
734 | |
735 | // Exception landing pads, blocks that has their adress taken, and function |
736 | // entry blocks will always be (potential) indirect jump targets, regardless of |
737 | // whether they are referenced by or not by jump tables. |
738 | static bool isAlwaysIndirectTarget(const MachineBasicBlock &MBB) { |
739 | return MBB.isEHPad() || MBB.hasAddressTaken() || |
740 | &MBB == &MBB.getParent()->front(); |
741 | } |
742 | |
743 | /// scanFunctionJumpTables - Do a scan of the function, building up |
744 | /// information about the sizes of each block and the locations of all |
745 | /// the jump tables. |
746 | void ARMConstantIslands::scanFunctionJumpTables() { |
747 | for (MachineBasicBlock &MBB : *MF) { |
748 | for (MachineInstr &I : MBB) |
749 | if (I.isBranch() && |
750 | (I.getOpcode() == ARM::t2BR_JT || I.getOpcode() == ARM::tBR_JTr)) |
751 | T2JumpTables.push_back(Elt: &I); |
752 | } |
753 | |
754 | if (!MF->getInfo<ARMFunctionInfo>()->branchTargetEnforcement()) |
755 | return; |
756 | |
757 | if (const MachineJumpTableInfo *JTI = MF->getJumpTableInfo()) |
758 | for (const MachineJumpTableEntry &JTE : JTI->getJumpTables()) |
759 | for (const MachineBasicBlock *MBB : JTE.MBBs) { |
760 | if (isAlwaysIndirectTarget(MBB: *MBB)) |
761 | // Set the reference count essentially to infinity, it will never |
762 | // reach zero and the BTI Instruction will never be removed. |
763 | BlockJumpTableRefCount[MBB] = std::numeric_limits<int>::max(); |
764 | else |
765 | ++BlockJumpTableRefCount[MBB]; |
766 | } |
767 | } |
768 | |
769 | /// initializeFunctionInfo - Do the initial scan of the function, building up |
770 | /// information about the sizes of each block, the location of all the water, |
771 | /// and finding all of the constant pool users. |
772 | void ARMConstantIslands:: |
773 | initializeFunctionInfo(const std::vector<MachineInstr*> &CPEMIs) { |
774 | |
775 | BBUtils->computeAllBlockSizes(); |
776 | BBInfoVector &BBInfo = BBUtils->getBBInfo(); |
777 | // The known bits of the entry block offset are determined by the function |
778 | // alignment. |
779 | BBInfo.front().KnownBits = Log2(A: MF->getAlignment()); |
780 | |
781 | // Compute block offsets and known bits. |
782 | BBUtils->adjustBBOffsetsAfter(MBB: &MF->front()); |
783 | |
784 | // We only care about jump table instructions when jump tables are inline. |
785 | MachineJumpTableInfo *MJTI = MF->getJumpTableInfo(); |
786 | bool InlineJumpTables = |
787 | MJTI && MJTI->getEntryKind() == MachineJumpTableInfo::EK_Inline; |
788 | |
789 | // Now go back through the instructions and build up our data structures. |
790 | for (MachineBasicBlock &MBB : *MF) { |
791 | // If this block doesn't fall through into the next MBB, then this is |
792 | // 'water' that a constant pool island could be placed. |
793 | if (!BBHasFallthrough(MBB: &MBB)) |
794 | WaterList.push_back(x: &MBB); |
795 | |
796 | for (MachineInstr &I : MBB) { |
797 | if (I.isDebugInstr()) |
798 | continue; |
799 | |
800 | unsigned Opc = I.getOpcode(); |
801 | if (I.isBranch()) { |
802 | bool isCond = false; |
803 | unsigned Bits = 0; |
804 | unsigned Scale = 1; |
805 | int UOpc = Opc; |
806 | switch (Opc) { |
807 | default: |
808 | continue; // Ignore other JT branches |
809 | case ARM::t2BR_JT: |
810 | case ARM::tBR_JTr: |
811 | if (InlineJumpTables) |
812 | T2JumpTables.push_back(Elt: &I); |
813 | continue; // Does not get an entry in ImmBranches |
814 | case ARM::Bcc: |
815 | isCond = true; |
816 | UOpc = ARM::B; |
817 | [[fallthrough]]; |
818 | case ARM::B: |
819 | Bits = 24; |
820 | Scale = 4; |
821 | break; |
822 | case ARM::tBcc: |
823 | isCond = true; |
824 | UOpc = ARM::tB; |
825 | Bits = 8; |
826 | Scale = 2; |
827 | break; |
828 | case ARM::tB: |
829 | Bits = 11; |
830 | Scale = 2; |
831 | break; |
832 | case ARM::t2Bcc: |
833 | isCond = true; |
834 | UOpc = ARM::t2B; |
835 | Bits = 20; |
836 | Scale = 2; |
837 | break; |
838 | case ARM::t2B: |
839 | Bits = 24; |
840 | Scale = 2; |
841 | break; |
842 | } |
843 | |
844 | // Record this immediate branch. |
845 | unsigned MaxOffs = ((1 << (Bits-1))-1) * Scale; |
846 | ImmBranches.push_back(x: ImmBranch(&I, MaxOffs, isCond, UOpc)); |
847 | } |
848 | |
849 | if (Opc == ARM::tPUSH || Opc == ARM::tPOP_RET) |
850 | PushPopMIs.push_back(Elt: &I); |
851 | |
852 | if (Opc == ARM::CONSTPOOL_ENTRY || Opc == ARM::JUMPTABLE_ADDRS || |
853 | Opc == ARM::JUMPTABLE_INSTS || Opc == ARM::JUMPTABLE_TBB || |
854 | Opc == ARM::JUMPTABLE_TBH) |
855 | continue; |
856 | |
857 | // Scan the instructions for constant pool operands. |
858 | for (unsigned op = 0, e = I.getNumOperands(); op != e; ++op) |
859 | if (I.getOperand(i: op).isCPI() || |
860 | (I.getOperand(i: op).isJTI() && InlineJumpTables)) { |
861 | // We found one. The addressing mode tells us the max displacement |
862 | // from the PC that this instruction permits. |
863 | |
864 | // Basic size info comes from the TSFlags field. |
865 | unsigned Bits = 0; |
866 | unsigned Scale = 1; |
867 | bool NegOk = false; |
868 | bool IsSoImm = false; |
869 | |
870 | switch (Opc) { |
871 | default: |
872 | llvm_unreachable("Unknown addressing mode for CP reference!" ); |
873 | |
874 | // Taking the address of a CP entry. |
875 | case ARM::LEApcrel: |
876 | case ARM::LEApcrelJT: { |
877 | // This takes a SoImm, which is 8 bit immediate rotated. We'll |
878 | // pretend the maximum offset is 255 * 4. Since each instruction |
879 | // 4 byte wide, this is always correct. We'll check for other |
880 | // displacements that fits in a SoImm as well. |
881 | Bits = 8; |
882 | NegOk = true; |
883 | IsSoImm = true; |
884 | unsigned CPI = I.getOperand(i: op).getIndex(); |
885 | assert(CPI < CPEMIs.size()); |
886 | MachineInstr *CPEMI = CPEMIs[CPI]; |
887 | const Align CPEAlign = getCPEAlign(CPEMI); |
888 | const unsigned LogCPEAlign = Log2(A: CPEAlign); |
889 | if (LogCPEAlign >= 2) |
890 | Scale = 4; |
891 | else |
892 | // For constants with less than 4-byte alignment, |
893 | // we'll pretend the maximum offset is 255 * 1. |
894 | Scale = 1; |
895 | } |
896 | break; |
897 | case ARM::t2LEApcrel: |
898 | case ARM::t2LEApcrelJT: |
899 | Bits = 12; |
900 | NegOk = true; |
901 | break; |
902 | case ARM::tLEApcrel: |
903 | case ARM::tLEApcrelJT: |
904 | Bits = 8; |
905 | Scale = 4; |
906 | break; |
907 | |
908 | case ARM::LDRBi12: |
909 | case ARM::LDRi12: |
910 | case ARM::LDRcp: |
911 | case ARM::t2LDRpci: |
912 | case ARM::t2LDRHpci: |
913 | case ARM::t2LDRSHpci: |
914 | case ARM::t2LDRBpci: |
915 | case ARM::t2LDRSBpci: |
916 | Bits = 12; // +-offset_12 |
917 | NegOk = true; |
918 | break; |
919 | |
920 | case ARM::tLDRpci: |
921 | Bits = 8; |
922 | Scale = 4; // +(offset_8*4) |
923 | break; |
924 | |
925 | case ARM::VLDRD: |
926 | case ARM::VLDRS: |
927 | Bits = 8; |
928 | Scale = 4; // +-(offset_8*4) |
929 | NegOk = true; |
930 | break; |
931 | case ARM::VLDRH: |
932 | Bits = 8; |
933 | Scale = 2; // +-(offset_8*2) |
934 | NegOk = true; |
935 | break; |
936 | } |
937 | |
938 | // Remember that this is a user of a CP entry. |
939 | unsigned CPI = I.getOperand(i: op).getIndex(); |
940 | if (I.getOperand(i: op).isJTI()) { |
941 | JumpTableUserIndices.insert(KV: std::make_pair(x&: CPI, y: CPUsers.size())); |
942 | CPI = JumpTableEntryIndices[CPI]; |
943 | } |
944 | |
945 | MachineInstr *CPEMI = CPEMIs[CPI]; |
946 | unsigned MaxOffs = ((1 << Bits)-1) * Scale; |
947 | CPUsers.push_back(x: CPUser(&I, CPEMI, MaxOffs, NegOk, IsSoImm)); |
948 | |
949 | // Increment corresponding CPEntry reference count. |
950 | CPEntry *CPE = findConstPoolEntry(CPI, CPEMI); |
951 | assert(CPE && "Cannot find a corresponding CPEntry!" ); |
952 | CPE->RefCount++; |
953 | |
954 | // Instructions can only use one CP entry, don't bother scanning the |
955 | // rest of the operands. |
956 | break; |
957 | } |
958 | } |
959 | } |
960 | } |
961 | |
962 | /// CompareMBBNumbers - Little predicate function to sort the WaterList by MBB |
963 | /// ID. |
964 | static bool CompareMBBNumbers(const MachineBasicBlock *LHS, |
965 | const MachineBasicBlock *RHS) { |
966 | return LHS->getNumber() < RHS->getNumber(); |
967 | } |
968 | |
969 | /// updateForInsertedWaterBlock - When a block is newly inserted into the |
970 | /// machine function, it upsets all of the block numbers. Renumber the blocks |
971 | /// and update the arrays that parallel this numbering. |
972 | void ARMConstantIslands::updateForInsertedWaterBlock(MachineBasicBlock *NewBB) { |
973 | // Renumber the MBB's to keep them consecutive. |
974 | NewBB->getParent()->RenumberBlocks(MBBFrom: NewBB); |
975 | |
976 | // Insert an entry into BBInfo to align it properly with the (newly |
977 | // renumbered) block numbers. |
978 | BBUtils->insert(BBNum: NewBB->getNumber(), BBI: BasicBlockInfo()); |
979 | |
980 | // Next, update WaterList. Specifically, we need to add NewMBB as having |
981 | // available water after it. |
982 | water_iterator IP = llvm::lower_bound(Range&: WaterList, Value&: NewBB, C: CompareMBBNumbers); |
983 | WaterList.insert(position: IP, x: NewBB); |
984 | } |
985 | |
986 | /// Split the basic block containing MI into two blocks, which are joined by |
987 | /// an unconditional branch. Update data structures and renumber blocks to |
988 | /// account for this change and returns the newly created block. |
989 | MachineBasicBlock *ARMConstantIslands::splitBlockBeforeInstr(MachineInstr *MI) { |
990 | MachineBasicBlock *OrigBB = MI->getParent(); |
991 | |
992 | // Collect liveness information at MI. |
993 | LivePhysRegs LRs(*MF->getSubtarget().getRegisterInfo()); |
994 | LRs.addLiveOuts(MBB: *OrigBB); |
995 | auto LivenessEnd = ++MachineBasicBlock::iterator(MI).getReverse(); |
996 | for (MachineInstr &LiveMI : make_range(x: OrigBB->rbegin(), y: LivenessEnd)) |
997 | LRs.stepBackward(MI: LiveMI); |
998 | |
999 | // Create a new MBB for the code after the OrigBB. |
1000 | MachineBasicBlock *NewBB = |
1001 | MF->CreateMachineBasicBlock(BB: OrigBB->getBasicBlock()); |
1002 | MachineFunction::iterator MBBI = ++OrigBB->getIterator(); |
1003 | MF->insert(MBBI, MBB: NewBB); |
1004 | |
1005 | // Splice the instructions starting with MI over to NewBB. |
1006 | NewBB->splice(Where: NewBB->end(), Other: OrigBB, From: MI, To: OrigBB->end()); |
1007 | |
1008 | // Add an unconditional branch from OrigBB to NewBB. |
1009 | // Note the new unconditional branch is not being recorded. |
1010 | // There doesn't seem to be meaningful DebugInfo available; this doesn't |
1011 | // correspond to anything in the source. |
1012 | unsigned Opc = isThumb ? (isThumb2 ? ARM::t2B : ARM::tB) : ARM::B; |
1013 | if (!isThumb) |
1014 | BuildMI(BB: OrigBB, MIMD: DebugLoc(), MCID: TII->get(Opcode: Opc)).addMBB(MBB: NewBB); |
1015 | else |
1016 | BuildMI(BB: OrigBB, MIMD: DebugLoc(), MCID: TII->get(Opcode: Opc)) |
1017 | .addMBB(MBB: NewBB) |
1018 | .add(MOs: predOps(Pred: ARMCC::AL)); |
1019 | ++NumSplit; |
1020 | |
1021 | // Update the CFG. All succs of OrigBB are now succs of NewBB. |
1022 | NewBB->transferSuccessors(FromMBB: OrigBB); |
1023 | |
1024 | // OrigBB branches to NewBB. |
1025 | OrigBB->addSuccessor(Succ: NewBB); |
1026 | |
1027 | // Update live-in information in the new block. |
1028 | MachineRegisterInfo &MRI = MF->getRegInfo(); |
1029 | for (MCPhysReg L : LRs) |
1030 | if (!MRI.isReserved(PhysReg: L)) |
1031 | NewBB->addLiveIn(PhysReg: L); |
1032 | |
1033 | // Update internal data structures to account for the newly inserted MBB. |
1034 | // This is almost the same as updateForInsertedWaterBlock, except that |
1035 | // the Water goes after OrigBB, not NewBB. |
1036 | MF->RenumberBlocks(MBBFrom: NewBB); |
1037 | |
1038 | // Insert an entry into BBInfo to align it properly with the (newly |
1039 | // renumbered) block numbers. |
1040 | BBUtils->insert(BBNum: NewBB->getNumber(), BBI: BasicBlockInfo()); |
1041 | |
1042 | // Next, update WaterList. Specifically, we need to add OrigMBB as having |
1043 | // available water after it (but not if it's already there, which happens |
1044 | // when splitting before a conditional branch that is followed by an |
1045 | // unconditional branch - in that case we want to insert NewBB). |
1046 | water_iterator IP = llvm::lower_bound(Range&: WaterList, Value&: OrigBB, C: CompareMBBNumbers); |
1047 | MachineBasicBlock* WaterBB = *IP; |
1048 | if (WaterBB == OrigBB) |
1049 | WaterList.insert(position: std::next(x: IP), x: NewBB); |
1050 | else |
1051 | WaterList.insert(position: IP, x: OrigBB); |
1052 | NewWaterList.insert(Ptr: OrigBB); |
1053 | |
1054 | // Figure out how large the OrigBB is. As the first half of the original |
1055 | // block, it cannot contain a tablejump. The size includes |
1056 | // the new jump we added. (It should be possible to do this without |
1057 | // recounting everything, but it's very confusing, and this is rarely |
1058 | // executed.) |
1059 | BBUtils->computeBlockSize(MBB: OrigBB); |
1060 | |
1061 | // Figure out how large the NewMBB is. As the second half of the original |
1062 | // block, it may contain a tablejump. |
1063 | BBUtils->computeBlockSize(MBB: NewBB); |
1064 | |
1065 | // All BBOffsets following these blocks must be modified. |
1066 | BBUtils->adjustBBOffsetsAfter(MBB: OrigBB); |
1067 | |
1068 | return NewBB; |
1069 | } |
1070 | |
1071 | /// getUserOffset - Compute the offset of U.MI as seen by the hardware |
1072 | /// displacement computation. Update U.KnownAlignment to match its current |
1073 | /// basic block location. |
1074 | unsigned ARMConstantIslands::getUserOffset(CPUser &U) const { |
1075 | unsigned UserOffset = BBUtils->getOffsetOf(MI: U.MI); |
1076 | |
1077 | SmallVectorImpl<BasicBlockInfo> &BBInfo = BBUtils->getBBInfo(); |
1078 | const BasicBlockInfo &BBI = BBInfo[U.MI->getParent()->getNumber()]; |
1079 | unsigned KnownBits = BBI.internalKnownBits(); |
1080 | |
1081 | // The value read from PC is offset from the actual instruction address. |
1082 | UserOffset += (isThumb ? 4 : 8); |
1083 | |
1084 | // Because of inline assembly, we may not know the alignment (mod 4) of U.MI. |
1085 | // Make sure U.getMaxDisp() returns a constrained range. |
1086 | U.KnownAlignment = (KnownBits >= 2); |
1087 | |
1088 | // On Thumb, offsets==2 mod 4 are rounded down by the hardware for |
1089 | // purposes of the displacement computation; compensate for that here. |
1090 | // For unknown alignments, getMaxDisp() constrains the range instead. |
1091 | if (isThumb && U.KnownAlignment) |
1092 | UserOffset &= ~3u; |
1093 | |
1094 | return UserOffset; |
1095 | } |
1096 | |
1097 | /// isOffsetInRange - Checks whether UserOffset (the location of a constant pool |
1098 | /// reference) is within MaxDisp of TrialOffset (a proposed location of a |
1099 | /// constant pool entry). |
1100 | /// UserOffset is computed by getUserOffset above to include PC adjustments. If |
1101 | /// the mod 4 alignment of UserOffset is not known, the uncertainty must be |
1102 | /// subtracted from MaxDisp instead. CPUser::getMaxDisp() does that. |
1103 | bool ARMConstantIslands::isOffsetInRange(unsigned UserOffset, |
1104 | unsigned TrialOffset, unsigned MaxDisp, |
1105 | bool NegativeOK, bool IsSoImm) { |
1106 | if (UserOffset <= TrialOffset) { |
1107 | // User before the Trial. |
1108 | if (TrialOffset - UserOffset <= MaxDisp) |
1109 | return true; |
1110 | // FIXME: Make use full range of soimm values. |
1111 | } else if (NegativeOK) { |
1112 | if (UserOffset - TrialOffset <= MaxDisp) |
1113 | return true; |
1114 | // FIXME: Make use full range of soimm values. |
1115 | } |
1116 | return false; |
1117 | } |
1118 | |
1119 | /// isWaterInRange - Returns true if a CPE placed after the specified |
1120 | /// Water (a basic block) will be in range for the specific MI. |
1121 | /// |
1122 | /// Compute how much the function will grow by inserting a CPE after Water. |
1123 | bool ARMConstantIslands::isWaterInRange(unsigned UserOffset, |
1124 | MachineBasicBlock* Water, CPUser &U, |
1125 | unsigned &Growth) { |
1126 | BBInfoVector &BBInfo = BBUtils->getBBInfo(); |
1127 | const Align CPEAlign = getCPEAlign(CPEMI: U.CPEMI); |
1128 | const unsigned CPEOffset = BBInfo[Water->getNumber()].postOffset(Alignment: CPEAlign); |
1129 | unsigned NextBlockOffset; |
1130 | Align NextBlockAlignment; |
1131 | MachineFunction::const_iterator NextBlock = Water->getIterator(); |
1132 | if (++NextBlock == MF->end()) { |
1133 | NextBlockOffset = BBInfo[Water->getNumber()].postOffset(); |
1134 | } else { |
1135 | NextBlockOffset = BBInfo[NextBlock->getNumber()].Offset; |
1136 | NextBlockAlignment = NextBlock->getAlignment(); |
1137 | } |
1138 | unsigned Size = U.CPEMI->getOperand(i: 2).getImm(); |
1139 | unsigned CPEEnd = CPEOffset + Size; |
1140 | |
1141 | // The CPE may be able to hide in the alignment padding before the next |
1142 | // block. It may also cause more padding to be required if it is more aligned |
1143 | // that the next block. |
1144 | if (CPEEnd > NextBlockOffset) { |
1145 | Growth = CPEEnd - NextBlockOffset; |
1146 | // Compute the padding that would go at the end of the CPE to align the next |
1147 | // block. |
1148 | Growth += offsetToAlignment(Value: CPEEnd, Alignment: NextBlockAlignment); |
1149 | |
1150 | // If the CPE is to be inserted before the instruction, that will raise |
1151 | // the offset of the instruction. Also account for unknown alignment padding |
1152 | // in blocks between CPE and the user. |
1153 | if (CPEOffset < UserOffset) |
1154 | UserOffset += Growth + UnknownPadding(Alignment: MF->getAlignment(), KnownBits: Log2(A: CPEAlign)); |
1155 | } else |
1156 | // CPE fits in existing padding. |
1157 | Growth = 0; |
1158 | |
1159 | return isOffsetInRange(UserOffset, TrialOffset: CPEOffset, U); |
1160 | } |
1161 | |
1162 | /// isCPEntryInRange - Returns true if the distance between specific MI and |
1163 | /// specific ConstPool entry instruction can fit in MI's displacement field. |
1164 | bool ARMConstantIslands::isCPEntryInRange(MachineInstr *MI, unsigned UserOffset, |
1165 | MachineInstr *CPEMI, unsigned MaxDisp, |
1166 | bool NegOk, bool DoDump) { |
1167 | unsigned CPEOffset = BBUtils->getOffsetOf(MI: CPEMI); |
1168 | |
1169 | if (DoDump) { |
1170 | LLVM_DEBUG({ |
1171 | BBInfoVector &BBInfo = BBUtils->getBBInfo(); |
1172 | unsigned Block = MI->getParent()->getNumber(); |
1173 | const BasicBlockInfo &BBI = BBInfo[Block]; |
1174 | dbgs() << "User of CPE#" << CPEMI->getOperand(0).getImm() |
1175 | << " max delta=" << MaxDisp |
1176 | << format(" insn address=%#x" , UserOffset) << " in " |
1177 | << printMBBReference(*MI->getParent()) << ": " |
1178 | << format("%#x-%x\t" , BBI.Offset, BBI.postOffset()) << *MI |
1179 | << format("CPE address=%#x offset=%+d: " , CPEOffset, |
1180 | int(CPEOffset - UserOffset)); |
1181 | }); |
1182 | } |
1183 | |
1184 | return isOffsetInRange(UserOffset, TrialOffset: CPEOffset, MaxDisp, NegativeOK: NegOk); |
1185 | } |
1186 | |
1187 | #ifndef NDEBUG |
1188 | /// BBIsJumpedOver - Return true of the specified basic block's only predecessor |
1189 | /// unconditionally branches to its only successor. |
1190 | static bool BBIsJumpedOver(MachineBasicBlock *MBB) { |
1191 | if (MBB->pred_size() != 1 || MBB->succ_size() != 1) |
1192 | return false; |
1193 | |
1194 | MachineBasicBlock *Succ = *MBB->succ_begin(); |
1195 | MachineBasicBlock *Pred = *MBB->pred_begin(); |
1196 | MachineInstr *PredMI = &Pred->back(); |
1197 | if (PredMI->getOpcode() == ARM::B || PredMI->getOpcode() == ARM::tB |
1198 | || PredMI->getOpcode() == ARM::t2B) |
1199 | return PredMI->getOperand(0).getMBB() == Succ; |
1200 | return false; |
1201 | } |
1202 | #endif // NDEBUG |
1203 | |
1204 | /// decrementCPEReferenceCount - find the constant pool entry with index CPI |
1205 | /// and instruction CPEMI, and decrement its refcount. If the refcount |
1206 | /// becomes 0 remove the entry and instruction. Returns true if we removed |
1207 | /// the entry, false if we didn't. |
1208 | bool ARMConstantIslands::decrementCPEReferenceCount(unsigned CPI, |
1209 | MachineInstr *CPEMI) { |
1210 | // Find the old entry. Eliminate it if it is no longer used. |
1211 | CPEntry *CPE = findConstPoolEntry(CPI, CPEMI); |
1212 | assert(CPE && "Unexpected!" ); |
1213 | if (--CPE->RefCount == 0) { |
1214 | removeDeadCPEMI(CPEMI); |
1215 | CPE->CPEMI = nullptr; |
1216 | --NumCPEs; |
1217 | return true; |
1218 | } |
1219 | return false; |
1220 | } |
1221 | |
1222 | unsigned ARMConstantIslands::getCombinedIndex(const MachineInstr *CPEMI) { |
1223 | if (CPEMI->getOperand(i: 1).isCPI()) |
1224 | return CPEMI->getOperand(i: 1).getIndex(); |
1225 | |
1226 | return JumpTableEntryIndices[CPEMI->getOperand(i: 1).getIndex()]; |
1227 | } |
1228 | |
1229 | /// LookForCPEntryInRange - see if the currently referenced CPE is in range; |
1230 | /// if not, see if an in-range clone of the CPE is in range, and if so, |
1231 | /// change the data structures so the user references the clone. Returns: |
1232 | /// 0 = no existing entry found |
1233 | /// 1 = entry found, and there were no code insertions or deletions |
1234 | /// 2 = entry found, and there were code insertions or deletions |
1235 | int ARMConstantIslands::findInRangeCPEntry(CPUser& U, unsigned UserOffset) { |
1236 | MachineInstr *UserMI = U.MI; |
1237 | MachineInstr *CPEMI = U.CPEMI; |
1238 | |
1239 | // Check to see if the CPE is already in-range. |
1240 | if (isCPEntryInRange(MI: UserMI, UserOffset, CPEMI, MaxDisp: U.getMaxDisp(), NegOk: U.NegOk, |
1241 | DoDump: true)) { |
1242 | LLVM_DEBUG(dbgs() << "In range\n" ); |
1243 | return 1; |
1244 | } |
1245 | |
1246 | // No. Look for previously created clones of the CPE that are in range. |
1247 | unsigned CPI = getCombinedIndex(CPEMI); |
1248 | std::vector<CPEntry> &CPEs = CPEntries[CPI]; |
1249 | for (CPEntry &CPE : CPEs) { |
1250 | // We already tried this one |
1251 | if (CPE.CPEMI == CPEMI) |
1252 | continue; |
1253 | // Removing CPEs can leave empty entries, skip |
1254 | if (CPE.CPEMI == nullptr) |
1255 | continue; |
1256 | if (isCPEntryInRange(MI: UserMI, UserOffset, CPEMI: CPE.CPEMI, MaxDisp: U.getMaxDisp(), |
1257 | NegOk: U.NegOk)) { |
1258 | LLVM_DEBUG(dbgs() << "Replacing CPE#" << CPI << " with CPE#" << CPE.CPI |
1259 | << "\n" ); |
1260 | // Point the CPUser node to the replacement |
1261 | U.CPEMI = CPE.CPEMI; |
1262 | // Change the CPI in the instruction operand to refer to the clone. |
1263 | for (MachineOperand &MO : UserMI->operands()) |
1264 | if (MO.isCPI()) { |
1265 | MO.setIndex(CPE.CPI); |
1266 | break; |
1267 | } |
1268 | // Adjust the refcount of the clone... |
1269 | CPE.RefCount++; |
1270 | // ...and the original. If we didn't remove the old entry, none of the |
1271 | // addresses changed, so we don't need another pass. |
1272 | return decrementCPEReferenceCount(CPI, CPEMI) ? 2 : 1; |
1273 | } |
1274 | } |
1275 | return 0; |
1276 | } |
1277 | |
1278 | /// getUnconditionalBrDisp - Returns the maximum displacement that can fit in |
1279 | /// the specific unconditional branch instruction. |
1280 | static inline unsigned getUnconditionalBrDisp(int Opc) { |
1281 | switch (Opc) { |
1282 | case ARM::tB: |
1283 | return ((1<<10)-1)*2; |
1284 | case ARM::t2B: |
1285 | return ((1<<23)-1)*2; |
1286 | default: |
1287 | break; |
1288 | } |
1289 | |
1290 | return ((1<<23)-1)*4; |
1291 | } |
1292 | |
1293 | /// findAvailableWater - Look for an existing entry in the WaterList in which |
1294 | /// we can place the CPE referenced from U so it's within range of U's MI. |
1295 | /// Returns true if found, false if not. If it returns true, WaterIter |
1296 | /// is set to the WaterList entry. For Thumb, prefer water that will not |
1297 | /// introduce padding to water that will. To ensure that this pass |
1298 | /// terminates, the CPE location for a particular CPUser is only allowed to |
1299 | /// move to a lower address, so search backward from the end of the list and |
1300 | /// prefer the first water that is in range. |
1301 | bool ARMConstantIslands::findAvailableWater(CPUser &U, unsigned UserOffset, |
1302 | water_iterator &WaterIter, |
1303 | bool CloserWater) { |
1304 | if (WaterList.empty()) |
1305 | return false; |
1306 | |
1307 | unsigned BestGrowth = ~0u; |
1308 | // The nearest water without splitting the UserBB is right after it. |
1309 | // If the distance is still large (we have a big BB), then we need to split it |
1310 | // if we don't converge after certain iterations. This helps the following |
1311 | // situation to converge: |
1312 | // BB0: |
1313 | // Big BB |
1314 | // BB1: |
1315 | // Constant Pool |
1316 | // When a CP access is out of range, BB0 may be used as water. However, |
1317 | // inserting islands between BB0 and BB1 makes other accesses out of range. |
1318 | MachineBasicBlock *UserBB = U.MI->getParent(); |
1319 | BBInfoVector &BBInfo = BBUtils->getBBInfo(); |
1320 | const Align CPEAlign = getCPEAlign(CPEMI: U.CPEMI); |
1321 | unsigned MinNoSplitDisp = BBInfo[UserBB->getNumber()].postOffset(Alignment: CPEAlign); |
1322 | if (CloserWater && MinNoSplitDisp > U.getMaxDisp() / 2) |
1323 | return false; |
1324 | for (water_iterator IP = std::prev(x: WaterList.end()), B = WaterList.begin();; |
1325 | --IP) { |
1326 | MachineBasicBlock* WaterBB = *IP; |
1327 | // Check if water is in range and is either at a lower address than the |
1328 | // current "high water mark" or a new water block that was created since |
1329 | // the previous iteration by inserting an unconditional branch. In the |
1330 | // latter case, we want to allow resetting the high water mark back to |
1331 | // this new water since we haven't seen it before. Inserting branches |
1332 | // should be relatively uncommon and when it does happen, we want to be |
1333 | // sure to take advantage of it for all the CPEs near that block, so that |
1334 | // we don't insert more branches than necessary. |
1335 | // When CloserWater is true, we try to find the lowest address after (or |
1336 | // equal to) user MI's BB no matter of padding growth. |
1337 | unsigned Growth; |
1338 | if (isWaterInRange(UserOffset, Water: WaterBB, U, Growth) && |
1339 | (WaterBB->getNumber() < U.HighWaterMark->getNumber() || |
1340 | NewWaterList.count(Ptr: WaterBB) || WaterBB == U.MI->getParent()) && |
1341 | Growth < BestGrowth) { |
1342 | // This is the least amount of required padding seen so far. |
1343 | BestGrowth = Growth; |
1344 | WaterIter = IP; |
1345 | LLVM_DEBUG(dbgs() << "Found water after " << printMBBReference(*WaterBB) |
1346 | << " Growth=" << Growth << '\n'); |
1347 | |
1348 | if (CloserWater && WaterBB == U.MI->getParent()) |
1349 | return true; |
1350 | // Keep looking unless it is perfect and we're not looking for the lowest |
1351 | // possible address. |
1352 | if (!CloserWater && BestGrowth == 0) |
1353 | return true; |
1354 | } |
1355 | if (IP == B) |
1356 | break; |
1357 | } |
1358 | return BestGrowth != ~0u; |
1359 | } |
1360 | |
1361 | /// createNewWater - No existing WaterList entry will work for |
1362 | /// CPUsers[CPUserIndex], so create a place to put the CPE. The end of the |
1363 | /// block is used if in range, and the conditional branch munged so control |
1364 | /// flow is correct. Otherwise the block is split to create a hole with an |
1365 | /// unconditional branch around it. In either case NewMBB is set to a |
1366 | /// block following which the new island can be inserted (the WaterList |
1367 | /// is not adjusted). |
1368 | void ARMConstantIslands::createNewWater(unsigned CPUserIndex, |
1369 | unsigned UserOffset, |
1370 | MachineBasicBlock *&NewMBB) { |
1371 | CPUser &U = CPUsers[CPUserIndex]; |
1372 | MachineInstr *UserMI = U.MI; |
1373 | MachineInstr *CPEMI = U.CPEMI; |
1374 | const Align CPEAlign = getCPEAlign(CPEMI); |
1375 | MachineBasicBlock *UserMBB = UserMI->getParent(); |
1376 | BBInfoVector &BBInfo = BBUtils->getBBInfo(); |
1377 | const BasicBlockInfo &UserBBI = BBInfo[UserMBB->getNumber()]; |
1378 | |
1379 | // If the block does not end in an unconditional branch already, and if the |
1380 | // end of the block is within range, make new water there. (The addition |
1381 | // below is for the unconditional branch we will be adding: 4 bytes on ARM + |
1382 | // Thumb2, 2 on Thumb1. |
1383 | if (BBHasFallthrough(MBB: UserMBB)) { |
1384 | // Size of branch to insert. |
1385 | unsigned Delta = isThumb1 ? 2 : 4; |
1386 | // Compute the offset where the CPE will begin. |
1387 | unsigned CPEOffset = UserBBI.postOffset(Alignment: CPEAlign) + Delta; |
1388 | |
1389 | if (isOffsetInRange(UserOffset, TrialOffset: CPEOffset, U)) { |
1390 | LLVM_DEBUG(dbgs() << "Split at end of " << printMBBReference(*UserMBB) |
1391 | << format(", expected CPE offset %#x\n" , CPEOffset)); |
1392 | NewMBB = &*++UserMBB->getIterator(); |
1393 | // Add an unconditional branch from UserMBB to fallthrough block. Record |
1394 | // it for branch lengthening; this new branch will not get out of range, |
1395 | // but if the preceding conditional branch is out of range, the targets |
1396 | // will be exchanged, and the altered branch may be out of range, so the |
1397 | // machinery has to know about it. |
1398 | int UncondBr = isThumb ? ((isThumb2) ? ARM::t2B : ARM::tB) : ARM::B; |
1399 | if (!isThumb) |
1400 | BuildMI(BB: UserMBB, MIMD: DebugLoc(), MCID: TII->get(Opcode: UncondBr)).addMBB(MBB: NewMBB); |
1401 | else |
1402 | BuildMI(BB: UserMBB, MIMD: DebugLoc(), MCID: TII->get(Opcode: UncondBr)) |
1403 | .addMBB(MBB: NewMBB) |
1404 | .add(MOs: predOps(Pred: ARMCC::AL)); |
1405 | unsigned MaxDisp = getUnconditionalBrDisp(Opc: UncondBr); |
1406 | ImmBranches.push_back(x: ImmBranch(&UserMBB->back(), |
1407 | MaxDisp, false, UncondBr)); |
1408 | BBUtils->computeBlockSize(MBB: UserMBB); |
1409 | BBUtils->adjustBBOffsetsAfter(MBB: UserMBB); |
1410 | return; |
1411 | } |
1412 | } |
1413 | |
1414 | // What a big block. Find a place within the block to split it. This is a |
1415 | // little tricky on Thumb1 since instructions are 2 bytes and constant pool |
1416 | // entries are 4 bytes: if instruction I references island CPE, and |
1417 | // instruction I+1 references CPE', it will not work well to put CPE as far |
1418 | // forward as possible, since then CPE' cannot immediately follow it (that |
1419 | // location is 2 bytes farther away from I+1 than CPE was from I) and we'd |
1420 | // need to create a new island. So, we make a first guess, then walk through |
1421 | // the instructions between the one currently being looked at and the |
1422 | // possible insertion point, and make sure any other instructions that |
1423 | // reference CPEs will be able to use the same island area; if not, we back |
1424 | // up the insertion point. |
1425 | |
1426 | // Try to split the block so it's fully aligned. Compute the latest split |
1427 | // point where we can add a 4-byte branch instruction, and then align to |
1428 | // Align which is the largest possible alignment in the function. |
1429 | const Align Align = MF->getAlignment(); |
1430 | assert(Align >= CPEAlign && "Over-aligned constant pool entry" ); |
1431 | unsigned KnownBits = UserBBI.internalKnownBits(); |
1432 | unsigned UPad = UnknownPadding(Alignment: Align, KnownBits); |
1433 | unsigned BaseInsertOffset = UserOffset + U.getMaxDisp() - UPad; |
1434 | LLVM_DEBUG(dbgs() << format("Split in middle of big block before %#x" , |
1435 | BaseInsertOffset)); |
1436 | |
1437 | // The 4 in the following is for the unconditional branch we'll be inserting |
1438 | // (allows for long branch on Thumb1). Alignment of the island is handled |
1439 | // inside isOffsetInRange. |
1440 | BaseInsertOffset -= 4; |
1441 | |
1442 | LLVM_DEBUG(dbgs() << format(", adjusted to %#x" , BaseInsertOffset) |
1443 | << " la=" << Log2(Align) << " kb=" << KnownBits |
1444 | << " up=" << UPad << '\n'); |
1445 | |
1446 | // This could point off the end of the block if we've already got constant |
1447 | // pool entries following this block; only the last one is in the water list. |
1448 | // Back past any possible branches (allow for a conditional and a maximally |
1449 | // long unconditional). |
1450 | if (BaseInsertOffset + 8 >= UserBBI.postOffset()) { |
1451 | // Ensure BaseInsertOffset is larger than the offset of the instruction |
1452 | // following UserMI so that the loop which searches for the split point |
1453 | // iterates at least once. |
1454 | BaseInsertOffset = |
1455 | std::max(a: UserBBI.postOffset() - UPad - 8, |
1456 | b: UserOffset + TII->getInstSizeInBytes(MI: *UserMI) + 1); |
1457 | // If the CP is referenced(ie, UserOffset) is in first four instructions |
1458 | // after IT, this recalculated BaseInsertOffset could be in the middle of |
1459 | // an IT block. If it is, change the BaseInsertOffset to just after the |
1460 | // IT block. This still make the CP Entry is in range becuase of the |
1461 | // following reasons. |
1462 | // 1. The initial BaseseInsertOffset calculated is (UserOffset + |
1463 | // U.getMaxDisp() - UPad). |
1464 | // 2. An IT block is only at most 4 instructions plus the "it" itself (18 |
1465 | // bytes). |
1466 | // 3. All the relevant instructions support much larger Maximum |
1467 | // displacement. |
1468 | MachineBasicBlock::iterator I = UserMI; |
1469 | ++I; |
1470 | Register PredReg; |
1471 | for (unsigned Offset = UserOffset + TII->getInstSizeInBytes(MI: *UserMI); |
1472 | I->getOpcode() != ARM::t2IT && |
1473 | getITInstrPredicate(MI: *I, PredReg) != ARMCC::AL; |
1474 | Offset += TII->getInstSizeInBytes(MI: *I), I = std::next(x: I)) { |
1475 | BaseInsertOffset = |
1476 | std::max(a: BaseInsertOffset, b: Offset + TII->getInstSizeInBytes(MI: *I) + 1); |
1477 | assert(I != UserMBB->end() && "Fell off end of block" ); |
1478 | } |
1479 | LLVM_DEBUG(dbgs() << format("Move inside block: %#x\n" , BaseInsertOffset)); |
1480 | } |
1481 | unsigned EndInsertOffset = BaseInsertOffset + 4 + UPad + |
1482 | CPEMI->getOperand(i: 2).getImm(); |
1483 | MachineBasicBlock::iterator MI = UserMI; |
1484 | ++MI; |
1485 | unsigned CPUIndex = CPUserIndex+1; |
1486 | unsigned NumCPUsers = CPUsers.size(); |
1487 | MachineInstr *LastIT = nullptr; |
1488 | for (unsigned Offset = UserOffset + TII->getInstSizeInBytes(MI: *UserMI); |
1489 | Offset < BaseInsertOffset; |
1490 | Offset += TII->getInstSizeInBytes(MI: *MI), MI = std::next(x: MI)) { |
1491 | assert(MI != UserMBB->end() && "Fell off end of block" ); |
1492 | if (CPUIndex < NumCPUsers && CPUsers[CPUIndex].MI == &*MI) { |
1493 | CPUser &U = CPUsers[CPUIndex]; |
1494 | if (!isOffsetInRange(UserOffset: Offset, TrialOffset: EndInsertOffset, U)) { |
1495 | // Shift intertion point by one unit of alignment so it is within reach. |
1496 | BaseInsertOffset -= Align.value(); |
1497 | EndInsertOffset -= Align.value(); |
1498 | } |
1499 | // This is overly conservative, as we don't account for CPEMIs being |
1500 | // reused within the block, but it doesn't matter much. Also assume CPEs |
1501 | // are added in order with alignment padding. We may eventually be able |
1502 | // to pack the aligned CPEs better. |
1503 | EndInsertOffset += U.CPEMI->getOperand(i: 2).getImm(); |
1504 | CPUIndex++; |
1505 | } |
1506 | |
1507 | // Remember the last IT instruction. |
1508 | if (MI->getOpcode() == ARM::t2IT) |
1509 | LastIT = &*MI; |
1510 | } |
1511 | |
1512 | --MI; |
1513 | |
1514 | // Avoid splitting an IT block. |
1515 | if (LastIT) { |
1516 | Register PredReg; |
1517 | ARMCC::CondCodes CC = getITInstrPredicate(MI: *MI, PredReg); |
1518 | if (CC != ARMCC::AL) |
1519 | MI = LastIT; |
1520 | } |
1521 | |
1522 | // Avoid splitting a MOVW+MOVT pair with a relocation on Windows. |
1523 | // On Windows, this instruction pair is covered by one single |
1524 | // IMAGE_REL_ARM_MOV32T relocation which covers both instructions. If a |
1525 | // constant island is injected inbetween them, the relocation will clobber |
1526 | // the instruction and fail to update the MOVT instruction. |
1527 | // (These instructions are bundled up until right before the ConstantIslands |
1528 | // pass.) |
1529 | if (STI->isTargetWindows() && isThumb && MI->getOpcode() == ARM::t2MOVTi16 && |
1530 | (MI->getOperand(i: 2).getTargetFlags() & ARMII::MO_OPTION_MASK) == |
1531 | ARMII::MO_HI16) { |
1532 | --MI; |
1533 | assert(MI->getOpcode() == ARM::t2MOVi16 && |
1534 | (MI->getOperand(1).getTargetFlags() & ARMII::MO_OPTION_MASK) == |
1535 | ARMII::MO_LO16); |
1536 | } |
1537 | |
1538 | // We really must not split an IT block. |
1539 | #ifndef NDEBUG |
1540 | Register PredReg; |
1541 | assert(!isThumb || getITInstrPredicate(*MI, PredReg) == ARMCC::AL); |
1542 | #endif |
1543 | NewMBB = splitBlockBeforeInstr(MI: &*MI); |
1544 | } |
1545 | |
1546 | /// handleConstantPoolUser - Analyze the specified user, checking to see if it |
1547 | /// is out-of-range. If so, pick up the constant pool value and move it some |
1548 | /// place in-range. Return true if we changed any addresses (thus must run |
1549 | /// another pass of branch lengthening), false otherwise. |
1550 | bool ARMConstantIslands::handleConstantPoolUser(unsigned CPUserIndex, |
1551 | bool CloserWater) { |
1552 | CPUser &U = CPUsers[CPUserIndex]; |
1553 | MachineInstr *UserMI = U.MI; |
1554 | MachineInstr *CPEMI = U.CPEMI; |
1555 | unsigned CPI = getCombinedIndex(CPEMI); |
1556 | unsigned Size = CPEMI->getOperand(i: 2).getImm(); |
1557 | // Compute this only once, it's expensive. |
1558 | unsigned UserOffset = getUserOffset(U); |
1559 | |
1560 | // See if the current entry is within range, or there is a clone of it |
1561 | // in range. |
1562 | int result = findInRangeCPEntry(U, UserOffset); |
1563 | if (result==1) return false; |
1564 | else if (result==2) return true; |
1565 | |
1566 | // No existing clone of this CPE is within range. |
1567 | // We will be generating a new clone. Get a UID for it. |
1568 | unsigned ID = AFI->createPICLabelUId(); |
1569 | |
1570 | // Look for water where we can place this CPE. |
1571 | MachineBasicBlock *NewIsland = MF->CreateMachineBasicBlock(); |
1572 | MachineBasicBlock *NewMBB; |
1573 | water_iterator IP; |
1574 | if (findAvailableWater(U, UserOffset, WaterIter&: IP, CloserWater)) { |
1575 | LLVM_DEBUG(dbgs() << "Found water in range\n" ); |
1576 | MachineBasicBlock *WaterBB = *IP; |
1577 | |
1578 | // If the original WaterList entry was "new water" on this iteration, |
1579 | // propagate that to the new island. This is just keeping NewWaterList |
1580 | // updated to match the WaterList, which will be updated below. |
1581 | if (NewWaterList.erase(Ptr: WaterBB)) |
1582 | NewWaterList.insert(Ptr: NewIsland); |
1583 | |
1584 | // The new CPE goes before the following block (NewMBB). |
1585 | NewMBB = &*++WaterBB->getIterator(); |
1586 | } else { |
1587 | // No water found. |
1588 | LLVM_DEBUG(dbgs() << "No water found\n" ); |
1589 | createNewWater(CPUserIndex, UserOffset, NewMBB); |
1590 | |
1591 | // splitBlockBeforeInstr adds to WaterList, which is important when it is |
1592 | // called while handling branches so that the water will be seen on the |
1593 | // next iteration for constant pools, but in this context, we don't want |
1594 | // it. Check for this so it will be removed from the WaterList. |
1595 | // Also remove any entry from NewWaterList. |
1596 | MachineBasicBlock *WaterBB = &*--NewMBB->getIterator(); |
1597 | IP = find(Range&: WaterList, Val: WaterBB); |
1598 | if (IP != WaterList.end()) |
1599 | NewWaterList.erase(Ptr: WaterBB); |
1600 | |
1601 | // We are adding new water. Update NewWaterList. |
1602 | NewWaterList.insert(Ptr: NewIsland); |
1603 | } |
1604 | // Always align the new block because CP entries can be smaller than 4 |
1605 | // bytes. Be careful not to decrease the existing alignment, e.g. NewMBB may |
1606 | // be an already aligned constant pool block. |
1607 | const Align Alignment = isThumb ? Align(2) : Align(4); |
1608 | if (NewMBB->getAlignment() < Alignment) |
1609 | NewMBB->setAlignment(Alignment); |
1610 | |
1611 | // Remove the original WaterList entry; we want subsequent insertions in |
1612 | // this vicinity to go after the one we're about to insert. This |
1613 | // considerably reduces the number of times we have to move the same CPE |
1614 | // more than once and is also important to ensure the algorithm terminates. |
1615 | if (IP != WaterList.end()) |
1616 | WaterList.erase(position: IP); |
1617 | |
1618 | // Okay, we know we can put an island before NewMBB now, do it! |
1619 | MF->insert(MBBI: NewMBB->getIterator(), MBB: NewIsland); |
1620 | |
1621 | // Update internal data structures to account for the newly inserted MBB. |
1622 | updateForInsertedWaterBlock(NewBB: NewIsland); |
1623 | |
1624 | // Now that we have an island to add the CPE to, clone the original CPE and |
1625 | // add it to the island. |
1626 | U.HighWaterMark = NewIsland; |
1627 | U.CPEMI = BuildMI(BB: NewIsland, MIMD: DebugLoc(), MCID: CPEMI->getDesc()) |
1628 | .addImm(Val: ID) |
1629 | .add(MO: CPEMI->getOperand(i: 1)) |
1630 | .addImm(Val: Size); |
1631 | CPEntries[CPI].push_back(x: CPEntry(U.CPEMI, ID, 1)); |
1632 | ++NumCPEs; |
1633 | |
1634 | // Decrement the old entry, and remove it if refcount becomes 0. |
1635 | decrementCPEReferenceCount(CPI, CPEMI); |
1636 | |
1637 | // Mark the basic block as aligned as required by the const-pool entry. |
1638 | NewIsland->setAlignment(getCPEAlign(CPEMI: U.CPEMI)); |
1639 | |
1640 | // Increase the size of the island block to account for the new entry. |
1641 | BBUtils->adjustBBSize(MBB: NewIsland, Size); |
1642 | BBUtils->adjustBBOffsetsAfter(MBB: &*--NewIsland->getIterator()); |
1643 | |
1644 | // Finally, change the CPI in the instruction operand to be ID. |
1645 | for (MachineOperand &MO : UserMI->operands()) |
1646 | if (MO.isCPI()) { |
1647 | MO.setIndex(ID); |
1648 | break; |
1649 | } |
1650 | |
1651 | LLVM_DEBUG( |
1652 | dbgs() << " Moved CPE to #" << ID << " CPI=" << CPI |
1653 | << format(" offset=%#x\n" , |
1654 | BBUtils->getBBInfo()[NewIsland->getNumber()].Offset)); |
1655 | |
1656 | return true; |
1657 | } |
1658 | |
1659 | /// removeDeadCPEMI - Remove a dead constant pool entry instruction. Update |
1660 | /// sizes and offsets of impacted basic blocks. |
1661 | void ARMConstantIslands::removeDeadCPEMI(MachineInstr *CPEMI) { |
1662 | MachineBasicBlock *CPEBB = CPEMI->getParent(); |
1663 | unsigned Size = CPEMI->getOperand(i: 2).getImm(); |
1664 | CPEMI->eraseFromParent(); |
1665 | BBInfoVector &BBInfo = BBUtils->getBBInfo(); |
1666 | BBUtils->adjustBBSize(MBB: CPEBB, Size: -Size); |
1667 | // All succeeding offsets have the current size value added in, fix this. |
1668 | if (CPEBB->empty()) { |
1669 | BBInfo[CPEBB->getNumber()].Size = 0; |
1670 | |
1671 | // This block no longer needs to be aligned. |
1672 | CPEBB->setAlignment(Align(1)); |
1673 | } else { |
1674 | // Entries are sorted by descending alignment, so realign from the front. |
1675 | CPEBB->setAlignment(getCPEAlign(CPEMI: &*CPEBB->begin())); |
1676 | } |
1677 | |
1678 | BBUtils->adjustBBOffsetsAfter(MBB: CPEBB); |
1679 | // An island has only one predecessor BB and one successor BB. Check if |
1680 | // this BB's predecessor jumps directly to this BB's successor. This |
1681 | // shouldn't happen currently. |
1682 | assert(!BBIsJumpedOver(CPEBB) && "How did this happen?" ); |
1683 | // FIXME: remove the empty blocks after all the work is done? |
1684 | } |
1685 | |
1686 | /// removeUnusedCPEntries - Remove constant pool entries whose refcounts |
1687 | /// are zero. |
1688 | bool ARMConstantIslands::removeUnusedCPEntries() { |
1689 | unsigned MadeChange = false; |
1690 | for (std::vector<CPEntry> &CPEs : CPEntries) { |
1691 | for (CPEntry &CPE : CPEs) { |
1692 | if (CPE.RefCount == 0 && CPE.CPEMI) { |
1693 | removeDeadCPEMI(CPEMI: CPE.CPEMI); |
1694 | CPE.CPEMI = nullptr; |
1695 | MadeChange = true; |
1696 | } |
1697 | } |
1698 | } |
1699 | return MadeChange; |
1700 | } |
1701 | |
1702 | |
1703 | /// fixupImmediateBr - Fix up an immediate branch whose destination is too far |
1704 | /// away to fit in its displacement field. |
1705 | bool ARMConstantIslands::fixupImmediateBr(ImmBranch &Br) { |
1706 | MachineInstr *MI = Br.MI; |
1707 | MachineBasicBlock *DestBB = MI->getOperand(i: 0).getMBB(); |
1708 | |
1709 | // Check to see if the DestBB is already in-range. |
1710 | if (BBUtils->isBBInRange(MI, DestBB, MaxDisp: Br.MaxDisp)) |
1711 | return false; |
1712 | |
1713 | if (!Br.isCond) |
1714 | return fixupUnconditionalBr(Br); |
1715 | return fixupConditionalBr(Br); |
1716 | } |
1717 | |
1718 | /// fixupUnconditionalBr - Fix up an unconditional branch whose destination is |
1719 | /// too far away to fit in its displacement field. If the LR register has been |
1720 | /// spilled in the epilogue, then we can use BL to implement a far jump. |
1721 | /// Otherwise, add an intermediate branch instruction to a branch. |
1722 | bool |
1723 | ARMConstantIslands::fixupUnconditionalBr(ImmBranch &Br) { |
1724 | MachineInstr *MI = Br.MI; |
1725 | MachineBasicBlock *MBB = MI->getParent(); |
1726 | if (!isThumb1) |
1727 | llvm_unreachable("fixupUnconditionalBr is Thumb1 only!" ); |
1728 | |
1729 | if (!AFI->isLRSpilled()) |
1730 | report_fatal_error(reason: "underestimated function size" ); |
1731 | |
1732 | // Use BL to implement far jump. |
1733 | Br.MaxDisp = (1 << 21) * 2; |
1734 | MI->setDesc(TII->get(Opcode: ARM::tBfar)); |
1735 | BBInfoVector &BBInfo = BBUtils->getBBInfo(); |
1736 | BBInfo[MBB->getNumber()].Size += 2; |
1737 | BBUtils->adjustBBOffsetsAfter(MBB); |
1738 | ++NumUBrFixed; |
1739 | |
1740 | LLVM_DEBUG(dbgs() << " Changed B to long jump " << *MI); |
1741 | |
1742 | return true; |
1743 | } |
1744 | |
1745 | /// fixupConditionalBr - Fix up a conditional branch whose destination is too |
1746 | /// far away to fit in its displacement field. It is converted to an inverse |
1747 | /// conditional branch + an unconditional branch to the destination. |
1748 | bool |
1749 | ARMConstantIslands::fixupConditionalBr(ImmBranch &Br) { |
1750 | MachineInstr *MI = Br.MI; |
1751 | MachineBasicBlock *DestBB = MI->getOperand(i: 0).getMBB(); |
1752 | |
1753 | // Add an unconditional branch to the destination and invert the branch |
1754 | // condition to jump over it: |
1755 | // blt L1 |
1756 | // => |
1757 | // bge L2 |
1758 | // b L1 |
1759 | // L2: |
1760 | ARMCC::CondCodes CC = (ARMCC::CondCodes)MI->getOperand(i: 1).getImm(); |
1761 | CC = ARMCC::getOppositeCondition(CC); |
1762 | Register CCReg = MI->getOperand(i: 2).getReg(); |
1763 | |
1764 | // If the branch is at the end of its MBB and that has a fall-through block, |
1765 | // direct the updated conditional branch to the fall-through block. Otherwise, |
1766 | // split the MBB before the next instruction. |
1767 | MachineBasicBlock *MBB = MI->getParent(); |
1768 | MachineInstr *BMI = &MBB->back(); |
1769 | bool NeedSplit = (BMI != MI) || !BBHasFallthrough(MBB); |
1770 | |
1771 | ++NumCBrFixed; |
1772 | if (BMI != MI) { |
1773 | if (std::next(x: MachineBasicBlock::iterator(MI)) == std::prev(x: MBB->end()) && |
1774 | BMI->getOpcode() == Br.UncondBr) { |
1775 | // Last MI in the BB is an unconditional branch. Can we simply invert the |
1776 | // condition and swap destinations: |
1777 | // beq L1 |
1778 | // b L2 |
1779 | // => |
1780 | // bne L2 |
1781 | // b L1 |
1782 | MachineBasicBlock *NewDest = BMI->getOperand(i: 0).getMBB(); |
1783 | if (BBUtils->isBBInRange(MI, DestBB: NewDest, MaxDisp: Br.MaxDisp)) { |
1784 | LLVM_DEBUG( |
1785 | dbgs() << " Invert Bcc condition and swap its destination with " |
1786 | << *BMI); |
1787 | BMI->getOperand(i: 0).setMBB(DestBB); |
1788 | MI->getOperand(i: 0).setMBB(NewDest); |
1789 | MI->getOperand(i: 1).setImm(CC); |
1790 | return true; |
1791 | } |
1792 | } |
1793 | } |
1794 | |
1795 | if (NeedSplit) { |
1796 | splitBlockBeforeInstr(MI); |
1797 | // No need for the branch to the next block. We're adding an unconditional |
1798 | // branch to the destination. |
1799 | int delta = TII->getInstSizeInBytes(MI: MBB->back()); |
1800 | BBUtils->adjustBBSize(MBB, Size: -delta); |
1801 | MBB->back().eraseFromParent(); |
1802 | |
1803 | // The conditional successor will be swapped between the BBs after this, so |
1804 | // update CFG. |
1805 | MBB->addSuccessor(Succ: DestBB); |
1806 | std::next(x: MBB->getIterator())->removeSuccessor(Succ: DestBB); |
1807 | |
1808 | // BBInfo[SplitBB].Offset is wrong temporarily, fixed below |
1809 | } |
1810 | MachineBasicBlock *NextBB = &*++MBB->getIterator(); |
1811 | |
1812 | LLVM_DEBUG(dbgs() << " Insert B to " << printMBBReference(*DestBB) |
1813 | << " also invert condition and change dest. to " |
1814 | << printMBBReference(*NextBB) << "\n" ); |
1815 | |
1816 | // Insert a new conditional branch and a new unconditional branch. |
1817 | // Also update the ImmBranch as well as adding a new entry for the new branch. |
1818 | BuildMI(BB: MBB, MIMD: DebugLoc(), MCID: TII->get(Opcode: MI->getOpcode())) |
1819 | .addMBB(MBB: NextBB).addImm(Val: CC).addReg(RegNo: CCReg); |
1820 | Br.MI = &MBB->back(); |
1821 | BBUtils->adjustBBSize(MBB, Size: TII->getInstSizeInBytes(MI: MBB->back())); |
1822 | if (isThumb) |
1823 | BuildMI(BB: MBB, MIMD: DebugLoc(), MCID: TII->get(Opcode: Br.UncondBr)) |
1824 | .addMBB(MBB: DestBB) |
1825 | .add(MOs: predOps(Pred: ARMCC::AL)); |
1826 | else |
1827 | BuildMI(BB: MBB, MIMD: DebugLoc(), MCID: TII->get(Opcode: Br.UncondBr)).addMBB(MBB: DestBB); |
1828 | BBUtils->adjustBBSize(MBB, Size: TII->getInstSizeInBytes(MI: MBB->back())); |
1829 | unsigned MaxDisp = getUnconditionalBrDisp(Opc: Br.UncondBr); |
1830 | ImmBranches.push_back(x: ImmBranch(&MBB->back(), MaxDisp, false, Br.UncondBr)); |
1831 | |
1832 | // Remove the old conditional branch. It may or may not still be in MBB. |
1833 | BBUtils->adjustBBSize(MBB: MI->getParent(), Size: -TII->getInstSizeInBytes(MI: *MI)); |
1834 | MI->eraseFromParent(); |
1835 | BBUtils->adjustBBOffsetsAfter(MBB); |
1836 | return true; |
1837 | } |
1838 | |
1839 | bool ARMConstantIslands::optimizeThumb2Instructions() { |
1840 | bool MadeChange = false; |
1841 | |
1842 | // Shrink ADR and LDR from constantpool. |
1843 | for (CPUser &U : CPUsers) { |
1844 | unsigned Opcode = U.MI->getOpcode(); |
1845 | unsigned NewOpc = 0; |
1846 | unsigned Scale = 1; |
1847 | unsigned Bits = 0; |
1848 | switch (Opcode) { |
1849 | default: break; |
1850 | case ARM::t2LEApcrel: |
1851 | if (isARMLowRegister(Reg: U.MI->getOperand(i: 0).getReg())) { |
1852 | NewOpc = ARM::tLEApcrel; |
1853 | Bits = 8; |
1854 | Scale = 4; |
1855 | } |
1856 | break; |
1857 | case ARM::t2LDRpci: |
1858 | if (isARMLowRegister(Reg: U.MI->getOperand(i: 0).getReg())) { |
1859 | NewOpc = ARM::tLDRpci; |
1860 | Bits = 8; |
1861 | Scale = 4; |
1862 | } |
1863 | break; |
1864 | } |
1865 | |
1866 | if (!NewOpc) |
1867 | continue; |
1868 | |
1869 | unsigned UserOffset = getUserOffset(U); |
1870 | unsigned MaxOffs = ((1 << Bits) - 1) * Scale; |
1871 | |
1872 | // Be conservative with inline asm. |
1873 | if (!U.KnownAlignment) |
1874 | MaxOffs -= 2; |
1875 | |
1876 | // FIXME: Check if offset is multiple of scale if scale is not 4. |
1877 | if (isCPEntryInRange(MI: U.MI, UserOffset, CPEMI: U.CPEMI, MaxDisp: MaxOffs, NegOk: false, DoDump: true)) { |
1878 | LLVM_DEBUG(dbgs() << "Shrink: " << *U.MI); |
1879 | U.MI->setDesc(TII->get(Opcode: NewOpc)); |
1880 | MachineBasicBlock *MBB = U.MI->getParent(); |
1881 | BBUtils->adjustBBSize(MBB, Size: -2); |
1882 | BBUtils->adjustBBOffsetsAfter(MBB); |
1883 | ++NumT2CPShrunk; |
1884 | MadeChange = true; |
1885 | } |
1886 | } |
1887 | |
1888 | return MadeChange; |
1889 | } |
1890 | |
1891 | |
1892 | bool ARMConstantIslands::optimizeThumb2Branches() { |
1893 | |
1894 | auto TryShrinkBranch = [this](ImmBranch &Br) { |
1895 | unsigned Opcode = Br.MI->getOpcode(); |
1896 | unsigned NewOpc = 0; |
1897 | unsigned Scale = 1; |
1898 | unsigned Bits = 0; |
1899 | switch (Opcode) { |
1900 | default: break; |
1901 | case ARM::t2B: |
1902 | NewOpc = ARM::tB; |
1903 | Bits = 11; |
1904 | Scale = 2; |
1905 | break; |
1906 | case ARM::t2Bcc: |
1907 | NewOpc = ARM::tBcc; |
1908 | Bits = 8; |
1909 | Scale = 2; |
1910 | break; |
1911 | } |
1912 | if (NewOpc) { |
1913 | unsigned MaxOffs = ((1 << (Bits-1))-1) * Scale; |
1914 | MachineBasicBlock *DestBB = Br.MI->getOperand(i: 0).getMBB(); |
1915 | if (BBUtils->isBBInRange(MI: Br.MI, DestBB, MaxDisp: MaxOffs)) { |
1916 | LLVM_DEBUG(dbgs() << "Shrink branch: " << *Br.MI); |
1917 | Br.MI->setDesc(TII->get(Opcode: NewOpc)); |
1918 | MachineBasicBlock *MBB = Br.MI->getParent(); |
1919 | BBUtils->adjustBBSize(MBB, Size: -2); |
1920 | BBUtils->adjustBBOffsetsAfter(MBB); |
1921 | ++NumT2BrShrunk; |
1922 | return true; |
1923 | } |
1924 | } |
1925 | return false; |
1926 | }; |
1927 | |
1928 | struct ImmCompare { |
1929 | MachineInstr* MI = nullptr; |
1930 | unsigned NewOpc = 0; |
1931 | }; |
1932 | |
1933 | auto FindCmpForCBZ = [this](ImmBranch &Br, ImmCompare &ImmCmp, |
1934 | MachineBasicBlock *DestBB) { |
1935 | ImmCmp.MI = nullptr; |
1936 | ImmCmp.NewOpc = 0; |
1937 | |
1938 | // If the conditional branch doesn't kill CPSR, then CPSR can be liveout |
1939 | // so this transformation is not safe. |
1940 | if (!Br.MI->killsRegister(Reg: ARM::CPSR, /*TRI=*/nullptr)) |
1941 | return false; |
1942 | |
1943 | Register PredReg; |
1944 | unsigned NewOpc = 0; |
1945 | ARMCC::CondCodes Pred = getInstrPredicate(MI: *Br.MI, PredReg); |
1946 | if (Pred == ARMCC::EQ) |
1947 | NewOpc = ARM::tCBZ; |
1948 | else if (Pred == ARMCC::NE) |
1949 | NewOpc = ARM::tCBNZ; |
1950 | else |
1951 | return false; |
1952 | |
1953 | // Check if the distance is within 126. Subtract starting offset by 2 |
1954 | // because the cmp will be eliminated. |
1955 | unsigned BrOffset = BBUtils->getOffsetOf(MI: Br.MI) + 4 - 2; |
1956 | BBInfoVector &BBInfo = BBUtils->getBBInfo(); |
1957 | unsigned DestOffset = BBInfo[DestBB->getNumber()].Offset; |
1958 | if (BrOffset >= DestOffset || (DestOffset - BrOffset) > 126) |
1959 | return false; |
1960 | |
1961 | // Search backwards to find a tCMPi8 |
1962 | auto *TRI = STI->getRegisterInfo(); |
1963 | MachineInstr *CmpMI = findCMPToFoldIntoCBZ(Br: Br.MI, TRI); |
1964 | if (!CmpMI || CmpMI->getOpcode() != ARM::tCMPi8) |
1965 | return false; |
1966 | |
1967 | ImmCmp.MI = CmpMI; |
1968 | ImmCmp.NewOpc = NewOpc; |
1969 | return true; |
1970 | }; |
1971 | |
1972 | auto TryConvertToLE = [this](ImmBranch &Br, ImmCompare &Cmp) { |
1973 | if (Br.MI->getOpcode() != ARM::t2Bcc || !STI->hasLOB() || |
1974 | STI->hasMinSize()) |
1975 | return false; |
1976 | |
1977 | MachineBasicBlock *MBB = Br.MI->getParent(); |
1978 | MachineBasicBlock *DestBB = Br.MI->getOperand(i: 0).getMBB(); |
1979 | if (BBUtils->getOffsetOf(MBB) < BBUtils->getOffsetOf(MBB: DestBB) || |
1980 | !BBUtils->isBBInRange(MI: Br.MI, DestBB, MaxDisp: 4094)) |
1981 | return false; |
1982 | |
1983 | if (!DT->dominates(A: DestBB, B: MBB)) |
1984 | return false; |
1985 | |
1986 | // We queried for the CBN?Z opcode based upon the 'ExitBB', the opposite |
1987 | // target of Br. So now we need to reverse the condition. |
1988 | Cmp.NewOpc = Cmp.NewOpc == ARM::tCBZ ? ARM::tCBNZ : ARM::tCBZ; |
1989 | |
1990 | MachineInstrBuilder MIB = BuildMI(BB&: *MBB, I: Br.MI, MIMD: Br.MI->getDebugLoc(), |
1991 | MCID: TII->get(Opcode: ARM::t2LE)); |
1992 | // Swapped a t2Bcc for a t2LE, so no need to update the size of the block. |
1993 | MIB.add(MO: Br.MI->getOperand(i: 0)); |
1994 | Br.MI->eraseFromParent(); |
1995 | Br.MI = MIB; |
1996 | ++NumLEInserted; |
1997 | return true; |
1998 | }; |
1999 | |
2000 | bool MadeChange = false; |
2001 | |
2002 | // The order in which branches appear in ImmBranches is approximately their |
2003 | // order within the function body. By visiting later branches first, we reduce |
2004 | // the distance between earlier forward branches and their targets, making it |
2005 | // more likely that the cbn?z optimization, which can only apply to forward |
2006 | // branches, will succeed. |
2007 | for (ImmBranch &Br : reverse(C&: ImmBranches)) { |
2008 | MachineBasicBlock *DestBB = Br.MI->getOperand(i: 0).getMBB(); |
2009 | MachineBasicBlock *MBB = Br.MI->getParent(); |
2010 | MachineBasicBlock *ExitBB = &MBB->back() == Br.MI ? |
2011 | MBB->getFallThrough() : |
2012 | MBB->back().getOperand(i: 0).getMBB(); |
2013 | |
2014 | ImmCompare Cmp; |
2015 | if (FindCmpForCBZ(Br, Cmp, ExitBB) && TryConvertToLE(Br, Cmp)) { |
2016 | DestBB = ExitBB; |
2017 | MadeChange = true; |
2018 | } else { |
2019 | FindCmpForCBZ(Br, Cmp, DestBB); |
2020 | MadeChange |= TryShrinkBranch(Br); |
2021 | } |
2022 | |
2023 | unsigned Opcode = Br.MI->getOpcode(); |
2024 | if ((Opcode != ARM::tBcc && Opcode != ARM::t2LE) || !Cmp.NewOpc) |
2025 | continue; |
2026 | |
2027 | Register Reg = Cmp.MI->getOperand(i: 0).getReg(); |
2028 | |
2029 | // Check for Kill flags on Reg. If they are present remove them and set kill |
2030 | // on the new CBZ. |
2031 | auto *TRI = STI->getRegisterInfo(); |
2032 | MachineBasicBlock::iterator KillMI = Br.MI; |
2033 | bool RegKilled = false; |
2034 | do { |
2035 | --KillMI; |
2036 | if (KillMI->killsRegister(Reg, TRI)) { |
2037 | KillMI->clearRegisterKills(Reg, RegInfo: TRI); |
2038 | RegKilled = true; |
2039 | break; |
2040 | } |
2041 | } while (KillMI != Cmp.MI); |
2042 | |
2043 | // Create the new CBZ/CBNZ |
2044 | LLVM_DEBUG(dbgs() << "Fold: " << *Cmp.MI << " and: " << *Br.MI); |
2045 | MachineInstr *NewBR = |
2046 | BuildMI(BB&: *MBB, I: Br.MI, MIMD: Br.MI->getDebugLoc(), MCID: TII->get(Opcode: Cmp.NewOpc)) |
2047 | .addReg(RegNo: Reg, flags: getKillRegState(B: RegKilled) | |
2048 | getRegState(RegOp: Cmp.MI->getOperand(i: 0))) |
2049 | .addMBB(MBB: DestBB, TargetFlags: Br.MI->getOperand(i: 0).getTargetFlags()); |
2050 | |
2051 | Cmp.MI->eraseFromParent(); |
2052 | |
2053 | if (Br.MI->getOpcode() == ARM::tBcc) { |
2054 | Br.MI->eraseFromParent(); |
2055 | Br.MI = NewBR; |
2056 | BBUtils->adjustBBSize(MBB, Size: -2); |
2057 | } else if (MBB->back().getOpcode() != ARM::t2LE) { |
2058 | // An LE has been generated, but it's not the terminator - that is an |
2059 | // unconditional branch. However, the logic has now been reversed with the |
2060 | // CBN?Z being the conditional branch and the LE being the unconditional |
2061 | // branch. So this means we can remove the redundant unconditional branch |
2062 | // at the end of the block. |
2063 | MachineInstr *LastMI = &MBB->back(); |
2064 | BBUtils->adjustBBSize(MBB, Size: -LastMI->getDesc().getSize()); |
2065 | LastMI->eraseFromParent(); |
2066 | } |
2067 | BBUtils->adjustBBOffsetsAfter(MBB); |
2068 | ++NumCBZ; |
2069 | MadeChange = true; |
2070 | } |
2071 | |
2072 | return MadeChange; |
2073 | } |
2074 | |
2075 | static bool isSimpleIndexCalc(MachineInstr &I, unsigned EntryReg, |
2076 | unsigned BaseReg) { |
2077 | if (I.getOpcode() != ARM::t2ADDrs) |
2078 | return false; |
2079 | |
2080 | if (I.getOperand(i: 0).getReg() != EntryReg) |
2081 | return false; |
2082 | |
2083 | if (I.getOperand(i: 1).getReg() != BaseReg) |
2084 | return false; |
2085 | |
2086 | // FIXME: what about CC and IdxReg? |
2087 | return true; |
2088 | } |
2089 | |
2090 | /// While trying to form a TBB/TBH instruction, we may (if the table |
2091 | /// doesn't immediately follow the BR_JT) need access to the start of the |
2092 | /// jump-table. We know one instruction that produces such a register; this |
2093 | /// function works out whether that definition can be preserved to the BR_JT, |
2094 | /// possibly by removing an intervening addition (which is usually needed to |
2095 | /// calculate the actual entry to jump to). |
2096 | bool ARMConstantIslands::preserveBaseRegister(MachineInstr *JumpMI, |
2097 | MachineInstr *LEAMI, |
2098 | unsigned &DeadSize, |
2099 | bool &CanDeleteLEA, |
2100 | bool &BaseRegKill) { |
2101 | if (JumpMI->getParent() != LEAMI->getParent()) |
2102 | return false; |
2103 | |
2104 | // Now we hope that we have at least these instructions in the basic block: |
2105 | // BaseReg = t2LEA ... |
2106 | // [...] |
2107 | // EntryReg = t2ADDrs BaseReg, ... |
2108 | // [...] |
2109 | // t2BR_JT EntryReg |
2110 | // |
2111 | // We have to be very conservative about what we recognise here though. The |
2112 | // main perturbing factors to watch out for are: |
2113 | // + Spills at any point in the chain: not direct problems but we would |
2114 | // expect a blocking Def of the spilled register so in practice what we |
2115 | // can do is limited. |
2116 | // + EntryReg == BaseReg: this is the one situation we should allow a Def |
2117 | // of BaseReg, but only if the t2ADDrs can be removed. |
2118 | // + Some instruction other than t2ADDrs computing the entry. Not seen in |
2119 | // the wild, but we should be careful. |
2120 | Register EntryReg = JumpMI->getOperand(i: 0).getReg(); |
2121 | Register BaseReg = LEAMI->getOperand(i: 0).getReg(); |
2122 | |
2123 | CanDeleteLEA = true; |
2124 | BaseRegKill = false; |
2125 | MachineInstr *RemovableAdd = nullptr; |
2126 | MachineBasicBlock::iterator I(LEAMI); |
2127 | for (++I; &*I != JumpMI; ++I) { |
2128 | if (isSimpleIndexCalc(I&: *I, EntryReg, BaseReg)) { |
2129 | RemovableAdd = &*I; |
2130 | break; |
2131 | } |
2132 | |
2133 | for (const MachineOperand &MO : I->operands()) { |
2134 | if (!MO.isReg() || !MO.getReg()) |
2135 | continue; |
2136 | if (MO.isDef() && MO.getReg() == BaseReg) |
2137 | return false; |
2138 | if (MO.isUse() && MO.getReg() == BaseReg) { |
2139 | BaseRegKill = BaseRegKill || MO.isKill(); |
2140 | CanDeleteLEA = false; |
2141 | } |
2142 | } |
2143 | } |
2144 | |
2145 | if (!RemovableAdd) |
2146 | return true; |
2147 | |
2148 | // Check the add really is removable, and that nothing else in the block |
2149 | // clobbers BaseReg. |
2150 | for (++I; &*I != JumpMI; ++I) { |
2151 | for (const MachineOperand &MO : I->operands()) { |
2152 | if (!MO.isReg() || !MO.getReg()) |
2153 | continue; |
2154 | if (MO.isDef() && MO.getReg() == BaseReg) |
2155 | return false; |
2156 | if (MO.isUse() && MO.getReg() == EntryReg) |
2157 | RemovableAdd = nullptr; |
2158 | } |
2159 | } |
2160 | |
2161 | if (RemovableAdd) { |
2162 | RemovableAdd->eraseFromParent(); |
2163 | DeadSize += isThumb2 ? 4 : 2; |
2164 | } else if (BaseReg == EntryReg) { |
2165 | // The add wasn't removable, but clobbered the base for the TBB. So we can't |
2166 | // preserve it. |
2167 | return false; |
2168 | } |
2169 | |
2170 | // We reached the end of the block without seeing another definition of |
2171 | // BaseReg (except, possibly the t2ADDrs, which was removed). BaseReg can be |
2172 | // used in the TBB/TBH if necessary. |
2173 | return true; |
2174 | } |
2175 | |
2176 | /// Returns whether CPEMI is the first instruction in the block |
2177 | /// immediately following JTMI (assumed to be a TBB or TBH terminator). If so, |
2178 | /// we can switch the first register to PC and usually remove the address |
2179 | /// calculation that preceded it. |
2180 | static bool jumpTableFollowsTB(MachineInstr *JTMI, MachineInstr *CPEMI) { |
2181 | MachineFunction::iterator MBB = JTMI->getParent()->getIterator(); |
2182 | MachineFunction *MF = MBB->getParent(); |
2183 | ++MBB; |
2184 | |
2185 | return MBB != MF->end() && !MBB->empty() && &*MBB->begin() == CPEMI; |
2186 | } |
2187 | |
2188 | static void RemoveDeadAddBetweenLEAAndJT(MachineInstr *LEAMI, |
2189 | MachineInstr *JumpMI, |
2190 | unsigned &DeadSize) { |
2191 | // Remove a dead add between the LEA and JT, which used to compute EntryReg, |
2192 | // but the JT now uses PC. Finds the last ADD (if any) that def's EntryReg |
2193 | // and is not clobbered / used. |
2194 | MachineInstr *RemovableAdd = nullptr; |
2195 | Register EntryReg = JumpMI->getOperand(i: 0).getReg(); |
2196 | |
2197 | // Find the last ADD to set EntryReg |
2198 | MachineBasicBlock::iterator I(LEAMI); |
2199 | for (++I; &*I != JumpMI; ++I) { |
2200 | if (I->getOpcode() == ARM::t2ADDrs && I->getOperand(i: 0).getReg() == EntryReg) |
2201 | RemovableAdd = &*I; |
2202 | } |
2203 | |
2204 | if (!RemovableAdd) |
2205 | return; |
2206 | |
2207 | // Ensure EntryReg is not clobbered or used. |
2208 | MachineBasicBlock::iterator J(RemovableAdd); |
2209 | for (++J; &*J != JumpMI; ++J) { |
2210 | for (const MachineOperand &MO : J->operands()) { |
2211 | if (!MO.isReg() || !MO.getReg()) |
2212 | continue; |
2213 | if (MO.isDef() && MO.getReg() == EntryReg) |
2214 | return; |
2215 | if (MO.isUse() && MO.getReg() == EntryReg) |
2216 | return; |
2217 | } |
2218 | } |
2219 | |
2220 | LLVM_DEBUG(dbgs() << "Removing Dead Add: " << *RemovableAdd); |
2221 | RemovableAdd->eraseFromParent(); |
2222 | DeadSize += 4; |
2223 | } |
2224 | |
2225 | /// optimizeThumb2JumpTables - Use tbb / tbh instructions to generate smaller |
2226 | /// jumptables when it's possible. |
2227 | bool ARMConstantIslands::optimizeThumb2JumpTables() { |
2228 | bool MadeChange = false; |
2229 | |
2230 | // FIXME: After the tables are shrunk, can we get rid some of the |
2231 | // constantpool tables? |
2232 | MachineJumpTableInfo *MJTI = MF->getJumpTableInfo(); |
2233 | if (!MJTI) return false; |
2234 | |
2235 | const std::vector<MachineJumpTableEntry> &JT = MJTI->getJumpTables(); |
2236 | for (MachineInstr *MI : T2JumpTables) { |
2237 | const MCInstrDesc &MCID = MI->getDesc(); |
2238 | unsigned NumOps = MCID.getNumOperands(); |
2239 | unsigned JTOpIdx = NumOps - (MI->isPredicable() ? 2 : 1); |
2240 | MachineOperand JTOP = MI->getOperand(i: JTOpIdx); |
2241 | unsigned JTI = JTOP.getIndex(); |
2242 | assert(JTI < JT.size()); |
2243 | |
2244 | bool ByteOk = true; |
2245 | bool HalfWordOk = true; |
2246 | unsigned JTOffset = BBUtils->getOffsetOf(MI) + 4; |
2247 | const std::vector<MachineBasicBlock*> &JTBBs = JT[JTI].MBBs; |
2248 | BBInfoVector &BBInfo = BBUtils->getBBInfo(); |
2249 | for (MachineBasicBlock *MBB : JTBBs) { |
2250 | unsigned DstOffset = BBInfo[MBB->getNumber()].Offset; |
2251 | // Negative offset is not ok. FIXME: We should change BB layout to make |
2252 | // sure all the branches are forward. |
2253 | if (ByteOk && (DstOffset - JTOffset) > ((1<<8)-1)*2) |
2254 | ByteOk = false; |
2255 | unsigned TBHLimit = ((1<<16)-1)*2; |
2256 | if (HalfWordOk && (DstOffset - JTOffset) > TBHLimit) |
2257 | HalfWordOk = false; |
2258 | if (!ByteOk && !HalfWordOk) |
2259 | break; |
2260 | } |
2261 | |
2262 | if (!ByteOk && !HalfWordOk) |
2263 | continue; |
2264 | |
2265 | CPUser &User = CPUsers[JumpTableUserIndices[JTI]]; |
2266 | MachineBasicBlock *MBB = MI->getParent(); |
2267 | if (!MI->getOperand(i: 0).isKill()) // FIXME: needed now? |
2268 | continue; |
2269 | |
2270 | unsigned DeadSize = 0; |
2271 | bool CanDeleteLEA = false; |
2272 | bool BaseRegKill = false; |
2273 | |
2274 | unsigned IdxReg = ~0U; |
2275 | bool IdxRegKill = true; |
2276 | if (isThumb2) { |
2277 | IdxReg = MI->getOperand(i: 1).getReg(); |
2278 | IdxRegKill = MI->getOperand(i: 1).isKill(); |
2279 | |
2280 | bool PreservedBaseReg = |
2281 | preserveBaseRegister(JumpMI: MI, LEAMI: User.MI, DeadSize, CanDeleteLEA, BaseRegKill); |
2282 | if (!jumpTableFollowsTB(JTMI: MI, CPEMI: User.CPEMI) && !PreservedBaseReg) |
2283 | continue; |
2284 | } else { |
2285 | // We're in thumb-1 mode, so we must have something like: |
2286 | // %idx = tLSLri %idx, 2 |
2287 | // %base = tLEApcrelJT |
2288 | // %t = tLDRr %base, %idx |
2289 | Register BaseReg = User.MI->getOperand(i: 0).getReg(); |
2290 | |
2291 | MachineBasicBlock *UserMBB = User.MI->getParent(); |
2292 | MachineBasicBlock::iterator Shift = User.MI->getIterator(); |
2293 | if (Shift == UserMBB->begin()) |
2294 | continue; |
2295 | |
2296 | Shift = prev_nodbg(It: Shift, Begin: UserMBB->begin()); |
2297 | if (Shift->getOpcode() != ARM::tLSLri || |
2298 | Shift->getOperand(i: 3).getImm() != 2 || |
2299 | !Shift->getOperand(i: 2).isKill()) |
2300 | continue; |
2301 | IdxReg = Shift->getOperand(i: 2).getReg(); |
2302 | Register ShiftedIdxReg = Shift->getOperand(i: 0).getReg(); |
2303 | |
2304 | // It's important that IdxReg is live until the actual TBB/TBH. Most of |
2305 | // the range is checked later, but the LEA might still clobber it and not |
2306 | // actually get removed. |
2307 | if (BaseReg == IdxReg && !jumpTableFollowsTB(JTMI: MI, CPEMI: User.CPEMI)) |
2308 | continue; |
2309 | |
2310 | MachineInstr *Load = User.MI->getNextNode(); |
2311 | if (Load->getOpcode() != ARM::tLDRr) |
2312 | continue; |
2313 | if (Load->getOperand(i: 1).getReg() != BaseReg || |
2314 | Load->getOperand(i: 2).getReg() != ShiftedIdxReg || |
2315 | !Load->getOperand(i: 2).isKill()) |
2316 | continue; |
2317 | |
2318 | // If we're in PIC mode, there should be another ADD following. |
2319 | auto *TRI = STI->getRegisterInfo(); |
2320 | |
2321 | // %base cannot be redefined after the load as it will appear before |
2322 | // TBB/TBH like: |
2323 | // %base = |
2324 | // %base = |
2325 | // tBB %base, %idx |
2326 | if (registerDefinedBetween(Reg: BaseReg, From: Load->getNextNode(), To: MBB->end(), TRI)) |
2327 | continue; |
2328 | |
2329 | if (isPositionIndependentOrROPI) { |
2330 | MachineInstr *Add = Load->getNextNode(); |
2331 | if (Add->getOpcode() != ARM::tADDrr || |
2332 | Add->getOperand(i: 2).getReg() != BaseReg || |
2333 | Add->getOperand(i: 3).getReg() != Load->getOperand(i: 0).getReg() || |
2334 | !Add->getOperand(i: 3).isKill()) |
2335 | continue; |
2336 | if (Add->getOperand(i: 0).getReg() != MI->getOperand(i: 0).getReg()) |
2337 | continue; |
2338 | if (registerDefinedBetween(Reg: IdxReg, From: Add->getNextNode(), To: MI, TRI)) |
2339 | // IdxReg gets redefined in the middle of the sequence. |
2340 | continue; |
2341 | Add->eraseFromParent(); |
2342 | DeadSize += 2; |
2343 | } else { |
2344 | if (Load->getOperand(i: 0).getReg() != MI->getOperand(i: 0).getReg()) |
2345 | continue; |
2346 | if (registerDefinedBetween(Reg: IdxReg, From: Load->getNextNode(), To: MI, TRI)) |
2347 | // IdxReg gets redefined in the middle of the sequence. |
2348 | continue; |
2349 | } |
2350 | |
2351 | // Now safe to delete the load and lsl. The LEA will be removed later. |
2352 | CanDeleteLEA = true; |
2353 | Shift->eraseFromParent(); |
2354 | Load->eraseFromParent(); |
2355 | DeadSize += 4; |
2356 | } |
2357 | |
2358 | LLVM_DEBUG(dbgs() << "Shrink JT: " << *MI); |
2359 | MachineInstr *CPEMI = User.CPEMI; |
2360 | unsigned Opc = ByteOk ? ARM::t2TBB_JT : ARM::t2TBH_JT; |
2361 | if (!isThumb2) |
2362 | Opc = ByteOk ? ARM::tTBB_JT : ARM::tTBH_JT; |
2363 | |
2364 | MachineBasicBlock::iterator MI_JT = MI; |
2365 | MachineInstr *NewJTMI = |
2366 | BuildMI(BB&: *MBB, I: MI_JT, MIMD: MI->getDebugLoc(), MCID: TII->get(Opcode: Opc)) |
2367 | .addReg(RegNo: User.MI->getOperand(i: 0).getReg(), |
2368 | flags: getKillRegState(B: BaseRegKill)) |
2369 | .addReg(RegNo: IdxReg, flags: getKillRegState(B: IdxRegKill)) |
2370 | .addJumpTableIndex(Idx: JTI, TargetFlags: JTOP.getTargetFlags()) |
2371 | .addImm(Val: CPEMI->getOperand(i: 0).getImm()); |
2372 | LLVM_DEBUG(dbgs() << printMBBReference(*MBB) << ": " << *NewJTMI); |
2373 | |
2374 | unsigned JTOpc = ByteOk ? ARM::JUMPTABLE_TBB : ARM::JUMPTABLE_TBH; |
2375 | CPEMI->setDesc(TII->get(Opcode: JTOpc)); |
2376 | |
2377 | if (jumpTableFollowsTB(JTMI: MI, CPEMI: User.CPEMI)) { |
2378 | NewJTMI->getOperand(i: 0).setReg(ARM::PC); |
2379 | NewJTMI->getOperand(i: 0).setIsKill(false); |
2380 | |
2381 | if (CanDeleteLEA) { |
2382 | if (isThumb2) |
2383 | RemoveDeadAddBetweenLEAAndJT(LEAMI: User.MI, JumpMI: MI, DeadSize); |
2384 | |
2385 | User.MI->eraseFromParent(); |
2386 | DeadSize += isThumb2 ? 4 : 2; |
2387 | |
2388 | // The LEA was eliminated, the TBB instruction becomes the only new user |
2389 | // of the jump table. |
2390 | User.MI = NewJTMI; |
2391 | User.MaxDisp = 4; |
2392 | User.NegOk = false; |
2393 | User.IsSoImm = false; |
2394 | User.KnownAlignment = false; |
2395 | } else { |
2396 | // The LEA couldn't be eliminated, so we must add another CPUser to |
2397 | // record the TBB or TBH use. |
2398 | int CPEntryIdx = JumpTableEntryIndices[JTI]; |
2399 | auto &CPEs = CPEntries[CPEntryIdx]; |
2400 | auto Entry = |
2401 | find_if(Range&: CPEs, P: [&](CPEntry &E) { return E.CPEMI == User.CPEMI; }); |
2402 | ++Entry->RefCount; |
2403 | CPUsers.emplace_back(args: CPUser(NewJTMI, User.CPEMI, 4, false, false)); |
2404 | } |
2405 | } |
2406 | |
2407 | unsigned NewSize = TII->getInstSizeInBytes(MI: *NewJTMI); |
2408 | unsigned OrigSize = TII->getInstSizeInBytes(MI: *MI); |
2409 | MI->eraseFromParent(); |
2410 | |
2411 | int Delta = OrigSize - NewSize + DeadSize; |
2412 | BBInfo[MBB->getNumber()].Size -= Delta; |
2413 | BBUtils->adjustBBOffsetsAfter(MBB); |
2414 | |
2415 | ++NumTBs; |
2416 | MadeChange = true; |
2417 | } |
2418 | |
2419 | return MadeChange; |
2420 | } |
2421 | |
2422 | /// reorderThumb2JumpTables - Adjust the function's block layout to ensure that |
2423 | /// jump tables always branch forwards, since that's what tbb and tbh need. |
2424 | bool ARMConstantIslands::reorderThumb2JumpTables() { |
2425 | bool MadeChange = false; |
2426 | |
2427 | MachineJumpTableInfo *MJTI = MF->getJumpTableInfo(); |
2428 | if (!MJTI) return false; |
2429 | |
2430 | const std::vector<MachineJumpTableEntry> &JT = MJTI->getJumpTables(); |
2431 | for (MachineInstr *MI : T2JumpTables) { |
2432 | const MCInstrDesc &MCID = MI->getDesc(); |
2433 | unsigned NumOps = MCID.getNumOperands(); |
2434 | unsigned JTOpIdx = NumOps - (MI->isPredicable() ? 2 : 1); |
2435 | MachineOperand JTOP = MI->getOperand(i: JTOpIdx); |
2436 | unsigned JTI = JTOP.getIndex(); |
2437 | assert(JTI < JT.size()); |
2438 | |
2439 | // We prefer if target blocks for the jump table come after the jump |
2440 | // instruction so we can use TB[BH]. Loop through the target blocks |
2441 | // and try to adjust them such that that's true. |
2442 | int JTNumber = MI->getParent()->getNumber(); |
2443 | const std::vector<MachineBasicBlock*> &JTBBs = JT[JTI].MBBs; |
2444 | for (MachineBasicBlock *MBB : JTBBs) { |
2445 | int DTNumber = MBB->getNumber(); |
2446 | |
2447 | if (DTNumber < JTNumber) { |
2448 | // The destination precedes the switch. Try to move the block forward |
2449 | // so we have a positive offset. |
2450 | MachineBasicBlock *NewBB = |
2451 | adjustJTTargetBlockForward(JTI, BB: MBB, JTBB: MI->getParent()); |
2452 | if (NewBB) |
2453 | MJTI->ReplaceMBBInJumpTable(Idx: JTI, Old: MBB, New: NewBB); |
2454 | MadeChange = true; |
2455 | } |
2456 | } |
2457 | } |
2458 | |
2459 | return MadeChange; |
2460 | } |
2461 | |
2462 | MachineBasicBlock *ARMConstantIslands::adjustJTTargetBlockForward( |
2463 | unsigned JTI, MachineBasicBlock *BB, MachineBasicBlock *JTBB) { |
2464 | // If the destination block is terminated by an unconditional branch, |
2465 | // try to move it; otherwise, create a new block following the jump |
2466 | // table that branches back to the actual target. This is a very simple |
2467 | // heuristic. FIXME: We can definitely improve it. |
2468 | MachineBasicBlock *TBB = nullptr, *FBB = nullptr; |
2469 | SmallVector<MachineOperand, 4> Cond; |
2470 | SmallVector<MachineOperand, 4> CondPrior; |
2471 | MachineFunction::iterator BBi = BB->getIterator(); |
2472 | MachineFunction::iterator OldPrior = std::prev(x: BBi); |
2473 | MachineFunction::iterator OldNext = std::next(x: BBi); |
2474 | |
2475 | // If the block terminator isn't analyzable, don't try to move the block |
2476 | bool B = TII->analyzeBranch(MBB&: *BB, TBB, FBB, Cond); |
2477 | |
2478 | // If the block ends in an unconditional branch, move it. The prior block |
2479 | // has to have an analyzable terminator for us to move this one. Be paranoid |
2480 | // and make sure we're not trying to move the entry block of the function. |
2481 | if (!B && Cond.empty() && BB != &MF->front() && |
2482 | !TII->analyzeBranch(MBB&: *OldPrior, TBB, FBB, Cond&: CondPrior)) { |
2483 | BB->moveAfter(NewBefore: JTBB); |
2484 | OldPrior->updateTerminator(PreviousLayoutSuccessor: BB); |
2485 | BB->updateTerminator(PreviousLayoutSuccessor: OldNext != MF->end() ? &*OldNext : nullptr); |
2486 | // Update numbering to account for the block being moved. |
2487 | MF->RenumberBlocks(); |
2488 | ++NumJTMoved; |
2489 | return nullptr; |
2490 | } |
2491 | |
2492 | // Create a new MBB for the code after the jump BB. |
2493 | MachineBasicBlock *NewBB = |
2494 | MF->CreateMachineBasicBlock(BB: JTBB->getBasicBlock()); |
2495 | MachineFunction::iterator MBBI = ++JTBB->getIterator(); |
2496 | MF->insert(MBBI, MBB: NewBB); |
2497 | |
2498 | // Copy live-in information to new block. |
2499 | for (const MachineBasicBlock::RegisterMaskPair &RegMaskPair : BB->liveins()) |
2500 | NewBB->addLiveIn(RegMaskPair); |
2501 | |
2502 | // Add an unconditional branch from NewBB to BB. |
2503 | // There doesn't seem to be meaningful DebugInfo available; this doesn't |
2504 | // correspond directly to anything in the source. |
2505 | if (isThumb2) |
2506 | BuildMI(BB: NewBB, MIMD: DebugLoc(), MCID: TII->get(Opcode: ARM::t2B)) |
2507 | .addMBB(MBB: BB) |
2508 | .add(MOs: predOps(Pred: ARMCC::AL)); |
2509 | else |
2510 | BuildMI(BB: NewBB, MIMD: DebugLoc(), MCID: TII->get(Opcode: ARM::tB)) |
2511 | .addMBB(MBB: BB) |
2512 | .add(MOs: predOps(Pred: ARMCC::AL)); |
2513 | |
2514 | // Update internal data structures to account for the newly inserted MBB. |
2515 | MF->RenumberBlocks(MBBFrom: NewBB); |
2516 | |
2517 | // Update the CFG. |
2518 | NewBB->addSuccessor(Succ: BB); |
2519 | JTBB->replaceSuccessor(Old: BB, New: NewBB); |
2520 | |
2521 | ++NumJTInserted; |
2522 | return NewBB; |
2523 | } |
2524 | |
2525 | /// createARMConstantIslandPass - returns an instance of the constpool |
2526 | /// island pass. |
2527 | FunctionPass *llvm::createARMConstantIslandPass() { |
2528 | return new ARMConstantIslands(); |
2529 | } |
2530 | |
2531 | INITIALIZE_PASS(ARMConstantIslands, "arm-cp-islands" , ARM_CP_ISLANDS_OPT_NAME, |
2532 | false, false) |
2533 | |