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