1 | //===-- X86FixupLEAs.cpp - use or replace LEA instructions -----------===// |
2 | // |
3 | // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. |
4 | // See https://llvm.org/LICENSE.txt for license information. |
5 | // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception |
6 | // |
7 | //===----------------------------------------------------------------------===// |
8 | // |
9 | // This file defines the pass that finds instructions that can be |
10 | // re-written as LEA instructions in order to reduce pipeline delays. |
11 | // It replaces LEAs with ADD/INC/DEC when that is better for size/speed. |
12 | // |
13 | //===----------------------------------------------------------------------===// |
14 | |
15 | #include "X86.h" |
16 | #include "X86InstrInfo.h" |
17 | #include "X86Subtarget.h" |
18 | #include "llvm/ADT/Statistic.h" |
19 | #include "llvm/Analysis/ProfileSummaryInfo.h" |
20 | #include "llvm/CodeGen/LazyMachineBlockFrequencyInfo.h" |
21 | #include "llvm/CodeGen/MachineFunctionPass.h" |
22 | #include "llvm/CodeGen/MachineInstrBuilder.h" |
23 | #include "llvm/CodeGen/MachineSizeOpts.h" |
24 | #include "llvm/CodeGen/Passes.h" |
25 | #include "llvm/CodeGen/TargetSchedule.h" |
26 | #include "llvm/Support/Debug.h" |
27 | #include "llvm/Support/raw_ostream.h" |
28 | using namespace llvm; |
29 | |
30 | #define FIXUPLEA_DESC "X86 LEA Fixup" |
31 | #define FIXUPLEA_NAME "x86-fixup-LEAs" |
32 | |
33 | #define DEBUG_TYPE FIXUPLEA_NAME |
34 | |
35 | STATISTIC(NumLEAs, "Number of LEA instructions created" ); |
36 | |
37 | namespace { |
38 | class FixupLEAPass : public MachineFunctionPass { |
39 | enum RegUsageState { RU_NotUsed, RU_Write, RU_Read }; |
40 | |
41 | /// Given a machine register, look for the instruction |
42 | /// which writes it in the current basic block. If found, |
43 | /// try to replace it with an equivalent LEA instruction. |
44 | /// If replacement succeeds, then also process the newly created |
45 | /// instruction. |
46 | void seekLEAFixup(MachineOperand &p, MachineBasicBlock::iterator &I, |
47 | MachineBasicBlock &MBB); |
48 | |
49 | /// Given a memory access or LEA instruction |
50 | /// whose address mode uses a base and/or index register, look for |
51 | /// an opportunity to replace the instruction which sets the base or index |
52 | /// register with an equivalent LEA instruction. |
53 | void processInstruction(MachineBasicBlock::iterator &I, |
54 | MachineBasicBlock &MBB); |
55 | |
56 | /// Given a LEA instruction which is unprofitable |
57 | /// on SlowLEA targets try to replace it with an equivalent ADD instruction. |
58 | void processInstructionForSlowLEA(MachineBasicBlock::iterator &I, |
59 | MachineBasicBlock &MBB); |
60 | |
61 | /// Given a LEA instruction which is unprofitable |
62 | /// on SNB+ try to replace it with other instructions. |
63 | /// According to Intel's Optimization Reference Manual: |
64 | /// " For LEA instructions with three source operands and some specific |
65 | /// situations, instruction latency has increased to 3 cycles, and must |
66 | /// dispatch via port 1: |
67 | /// - LEA that has all three source operands: base, index, and offset |
68 | /// - LEA that uses base and index registers where the base is EBP, RBP, |
69 | /// or R13 |
70 | /// - LEA that uses RIP relative addressing mode |
71 | /// - LEA that uses 16-bit addressing mode " |
72 | /// This function currently handles the first 2 cases only. |
73 | void processInstrForSlow3OpLEA(MachineBasicBlock::iterator &I, |
74 | MachineBasicBlock &MBB, bool OptIncDec); |
75 | |
76 | /// Look for LEAs that are really two address LEAs that we might be able to |
77 | /// turn into regular ADD instructions. |
78 | bool optTwoAddrLEA(MachineBasicBlock::iterator &I, |
79 | MachineBasicBlock &MBB, bool OptIncDec, |
80 | bool UseLEAForSP) const; |
81 | |
82 | /// Look for and transform the sequence |
83 | /// lea (reg1, reg2), reg3 |
84 | /// sub reg3, reg4 |
85 | /// to |
86 | /// sub reg1, reg4 |
87 | /// sub reg2, reg4 |
88 | /// It can also optimize the sequence lea/add similarly. |
89 | bool optLEAALU(MachineBasicBlock::iterator &I, MachineBasicBlock &MBB) const; |
90 | |
91 | /// Step forwards in MBB, looking for an ADD/SUB instruction which uses |
92 | /// the dest register of LEA instruction I. |
93 | MachineBasicBlock::iterator searchALUInst(MachineBasicBlock::iterator &I, |
94 | MachineBasicBlock &MBB) const; |
95 | |
96 | /// Check instructions between LeaI and AluI (exclusively). |
97 | /// Set BaseIndexDef to true if base or index register from LeaI is defined. |
98 | /// Set AluDestRef to true if the dest register of AluI is used or defined. |
99 | /// *KilledBase is set to the killed base register usage. |
100 | /// *KilledIndex is set to the killed index register usage. |
101 | void checkRegUsage(MachineBasicBlock::iterator &LeaI, |
102 | MachineBasicBlock::iterator &AluI, bool &BaseIndexDef, |
103 | bool &AluDestRef, MachineOperand **KilledBase, |
104 | MachineOperand **KilledIndex) const; |
105 | |
106 | /// Determine if an instruction references a machine register |
107 | /// and, if so, whether it reads or writes the register. |
108 | RegUsageState usesRegister(MachineOperand &p, MachineBasicBlock::iterator I); |
109 | |
110 | /// Step backwards through a basic block, looking |
111 | /// for an instruction which writes a register within |
112 | /// a maximum of INSTR_DISTANCE_THRESHOLD instruction latency cycles. |
113 | MachineBasicBlock::iterator searchBackwards(MachineOperand &p, |
114 | MachineBasicBlock::iterator &I, |
115 | MachineBasicBlock &MBB); |
116 | |
117 | /// if an instruction can be converted to an |
118 | /// equivalent LEA, insert the new instruction into the basic block |
119 | /// and return a pointer to it. Otherwise, return zero. |
120 | MachineInstr *postRAConvertToLEA(MachineBasicBlock &MBB, |
121 | MachineBasicBlock::iterator &MBBI) const; |
122 | |
123 | public: |
124 | static char ID; |
125 | |
126 | StringRef getPassName() const override { return FIXUPLEA_DESC; } |
127 | |
128 | FixupLEAPass() : MachineFunctionPass(ID) { } |
129 | |
130 | /// Loop over all of the basic blocks, |
131 | /// replacing instructions by equivalent LEA instructions |
132 | /// if needed and when possible. |
133 | bool runOnMachineFunction(MachineFunction &MF) override; |
134 | |
135 | // This pass runs after regalloc and doesn't support VReg operands. |
136 | MachineFunctionProperties getRequiredProperties() const override { |
137 | return MachineFunctionProperties().set( |
138 | MachineFunctionProperties::Property::NoVRegs); |
139 | } |
140 | |
141 | void getAnalysisUsage(AnalysisUsage &AU) const override { |
142 | AU.addRequired<ProfileSummaryInfoWrapperPass>(); |
143 | AU.addRequired<LazyMachineBlockFrequencyInfoPass>(); |
144 | MachineFunctionPass::getAnalysisUsage(AU); |
145 | } |
146 | |
147 | private: |
148 | TargetSchedModel TSM; |
149 | const X86InstrInfo *TII = nullptr; |
150 | const X86RegisterInfo *TRI = nullptr; |
151 | }; |
152 | } |
153 | |
154 | char FixupLEAPass::ID = 0; |
155 | |
156 | INITIALIZE_PASS(FixupLEAPass, FIXUPLEA_NAME, FIXUPLEA_DESC, false, false) |
157 | |
158 | MachineInstr * |
159 | FixupLEAPass::postRAConvertToLEA(MachineBasicBlock &MBB, |
160 | MachineBasicBlock::iterator &MBBI) const { |
161 | MachineInstr &MI = *MBBI; |
162 | switch (MI.getOpcode()) { |
163 | case X86::MOV32rr: |
164 | case X86::MOV64rr: { |
165 | const MachineOperand &Src = MI.getOperand(i: 1); |
166 | const MachineOperand &Dest = MI.getOperand(i: 0); |
167 | MachineInstr *NewMI = |
168 | BuildMI(BB&: MBB, I: MBBI, MIMD: MI.getDebugLoc(), |
169 | MCID: TII->get(Opcode: MI.getOpcode() == X86::MOV32rr ? X86::LEA32r |
170 | : X86::LEA64r)) |
171 | .add(MO: Dest) |
172 | .add(MO: Src) |
173 | .addImm(Val: 1) |
174 | .addReg(RegNo: 0) |
175 | .addImm(Val: 0) |
176 | .addReg(RegNo: 0); |
177 | return NewMI; |
178 | } |
179 | } |
180 | |
181 | if (!MI.isConvertibleTo3Addr()) |
182 | return nullptr; |
183 | |
184 | switch (MI.getOpcode()) { |
185 | default: |
186 | // Only convert instructions that we've verified are safe. |
187 | return nullptr; |
188 | case X86::ADD64ri32: |
189 | case X86::ADD64ri32_DB: |
190 | case X86::ADD32ri: |
191 | case X86::ADD32ri_DB: |
192 | if (!MI.getOperand(i: 2).isImm()) { |
193 | // convertToThreeAddress will call getImm() |
194 | // which requires isImm() to be true |
195 | return nullptr; |
196 | } |
197 | break; |
198 | case X86::SHL64ri: |
199 | case X86::SHL32ri: |
200 | case X86::INC64r: |
201 | case X86::INC32r: |
202 | case X86::DEC64r: |
203 | case X86::DEC32r: |
204 | case X86::ADD64rr: |
205 | case X86::ADD64rr_DB: |
206 | case X86::ADD32rr: |
207 | case X86::ADD32rr_DB: |
208 | // These instructions are all fine to convert. |
209 | break; |
210 | } |
211 | return TII->convertToThreeAddress(MI, LV: nullptr, LIS: nullptr); |
212 | } |
213 | |
214 | FunctionPass *llvm::createX86FixupLEAs() { return new FixupLEAPass(); } |
215 | |
216 | static bool isLEA(unsigned Opcode) { |
217 | return Opcode == X86::LEA32r || Opcode == X86::LEA64r || |
218 | Opcode == X86::LEA64_32r; |
219 | } |
220 | |
221 | bool FixupLEAPass::runOnMachineFunction(MachineFunction &MF) { |
222 | if (skipFunction(F: MF.getFunction())) |
223 | return false; |
224 | |
225 | const X86Subtarget &ST = MF.getSubtarget<X86Subtarget>(); |
226 | bool IsSlowLEA = ST.slowLEA(); |
227 | bool IsSlow3OpsLEA = ST.slow3OpsLEA(); |
228 | bool LEAUsesAG = ST.leaUsesAG(); |
229 | |
230 | bool OptIncDec = !ST.slowIncDec() || MF.getFunction().hasOptSize(); |
231 | bool UseLEAForSP = ST.useLeaForSP(); |
232 | |
233 | TSM.init(TSInfo: &ST); |
234 | TII = ST.getInstrInfo(); |
235 | TRI = ST.getRegisterInfo(); |
236 | auto *PSI = &getAnalysis<ProfileSummaryInfoWrapperPass>().getPSI(); |
237 | auto *MBFI = (PSI && PSI->hasProfileSummary()) |
238 | ? &getAnalysis<LazyMachineBlockFrequencyInfoPass>().getBFI() |
239 | : nullptr; |
240 | |
241 | LLVM_DEBUG(dbgs() << "Start X86FixupLEAs\n" ;); |
242 | for (MachineBasicBlock &MBB : MF) { |
243 | // First pass. Try to remove or optimize existing LEAs. |
244 | bool OptIncDecPerBB = |
245 | OptIncDec || llvm::shouldOptimizeForSize(MBB: &MBB, PSI, MBFI); |
246 | for (MachineBasicBlock::iterator I = MBB.begin(); I != MBB.end(); ++I) { |
247 | if (!isLEA(Opcode: I->getOpcode())) |
248 | continue; |
249 | |
250 | if (optTwoAddrLEA(I, MBB, OptIncDec: OptIncDecPerBB, UseLEAForSP)) |
251 | continue; |
252 | |
253 | if (IsSlowLEA) |
254 | processInstructionForSlowLEA(I, MBB); |
255 | else if (IsSlow3OpsLEA) |
256 | processInstrForSlow3OpLEA(I, MBB, OptIncDec: OptIncDecPerBB); |
257 | } |
258 | |
259 | // Second pass for creating LEAs. This may reverse some of the |
260 | // transformations above. |
261 | if (LEAUsesAG) { |
262 | for (MachineBasicBlock::iterator I = MBB.begin(); I != MBB.end(); ++I) |
263 | processInstruction(I, MBB); |
264 | } |
265 | } |
266 | |
267 | LLVM_DEBUG(dbgs() << "End X86FixupLEAs\n" ;); |
268 | |
269 | return true; |
270 | } |
271 | |
272 | FixupLEAPass::RegUsageState |
273 | FixupLEAPass::usesRegister(MachineOperand &p, MachineBasicBlock::iterator I) { |
274 | RegUsageState RegUsage = RU_NotUsed; |
275 | MachineInstr &MI = *I; |
276 | |
277 | for (const MachineOperand &MO : MI.operands()) { |
278 | if (MO.isReg() && MO.getReg() == p.getReg()) { |
279 | if (MO.isDef()) |
280 | return RU_Write; |
281 | RegUsage = RU_Read; |
282 | } |
283 | } |
284 | return RegUsage; |
285 | } |
286 | |
287 | /// getPreviousInstr - Given a reference to an instruction in a basic |
288 | /// block, return a reference to the previous instruction in the block, |
289 | /// wrapping around to the last instruction of the block if the block |
290 | /// branches to itself. |
291 | static inline bool getPreviousInstr(MachineBasicBlock::iterator &I, |
292 | MachineBasicBlock &MBB) { |
293 | if (I == MBB.begin()) { |
294 | if (MBB.isPredecessor(MBB: &MBB)) { |
295 | I = --MBB.end(); |
296 | return true; |
297 | } else |
298 | return false; |
299 | } |
300 | --I; |
301 | return true; |
302 | } |
303 | |
304 | MachineBasicBlock::iterator |
305 | FixupLEAPass::searchBackwards(MachineOperand &p, MachineBasicBlock::iterator &I, |
306 | MachineBasicBlock &MBB) { |
307 | int InstrDistance = 1; |
308 | MachineBasicBlock::iterator CurInst; |
309 | static const int INSTR_DISTANCE_THRESHOLD = 5; |
310 | |
311 | CurInst = I; |
312 | bool Found; |
313 | Found = getPreviousInstr(I&: CurInst, MBB); |
314 | while (Found && I != CurInst) { |
315 | if (CurInst->isCall() || CurInst->isInlineAsm()) |
316 | break; |
317 | if (InstrDistance > INSTR_DISTANCE_THRESHOLD) |
318 | break; // too far back to make a difference |
319 | if (usesRegister(p, I: CurInst) == RU_Write) { |
320 | return CurInst; |
321 | } |
322 | InstrDistance += TSM.computeInstrLatency(MI: &*CurInst); |
323 | Found = getPreviousInstr(I&: CurInst, MBB); |
324 | } |
325 | return MachineBasicBlock::iterator(); |
326 | } |
327 | |
328 | static inline bool isInefficientLEAReg(unsigned Reg) { |
329 | return Reg == X86::EBP || Reg == X86::RBP || |
330 | Reg == X86::R13D || Reg == X86::R13; |
331 | } |
332 | |
333 | /// Returns true if this LEA uses base and index registers, and the base |
334 | /// register is known to be inefficient for the subtarget. |
335 | // TODO: use a variant scheduling class to model the latency profile |
336 | // of LEA instructions, and implement this logic as a scheduling predicate. |
337 | static inline bool hasInefficientLEABaseReg(const MachineOperand &Base, |
338 | const MachineOperand &Index) { |
339 | return Base.isReg() && isInefficientLEAReg(Reg: Base.getReg()) && Index.isReg() && |
340 | Index.getReg() != X86::NoRegister; |
341 | } |
342 | |
343 | static inline bool hasLEAOffset(const MachineOperand &Offset) { |
344 | return (Offset.isImm() && Offset.getImm() != 0) || Offset.isGlobal() || |
345 | Offset.isBlockAddress(); |
346 | } |
347 | |
348 | static inline unsigned getADDrrFromLEA(unsigned LEAOpcode) { |
349 | switch (LEAOpcode) { |
350 | default: |
351 | llvm_unreachable("Unexpected LEA instruction" ); |
352 | case X86::LEA32r: |
353 | case X86::LEA64_32r: |
354 | return X86::ADD32rr; |
355 | case X86::LEA64r: |
356 | return X86::ADD64rr; |
357 | } |
358 | } |
359 | |
360 | static inline unsigned getSUBrrFromLEA(unsigned LEAOpcode) { |
361 | switch (LEAOpcode) { |
362 | default: |
363 | llvm_unreachable("Unexpected LEA instruction" ); |
364 | case X86::LEA32r: |
365 | case X86::LEA64_32r: |
366 | return X86::SUB32rr; |
367 | case X86::LEA64r: |
368 | return X86::SUB64rr; |
369 | } |
370 | } |
371 | |
372 | static inline unsigned getADDriFromLEA(unsigned LEAOpcode, |
373 | const MachineOperand &Offset) { |
374 | switch (LEAOpcode) { |
375 | default: |
376 | llvm_unreachable("Unexpected LEA instruction" ); |
377 | case X86::LEA32r: |
378 | case X86::LEA64_32r: |
379 | return X86::ADD32ri; |
380 | case X86::LEA64r: |
381 | return X86::ADD64ri32; |
382 | } |
383 | } |
384 | |
385 | static inline unsigned getINCDECFromLEA(unsigned LEAOpcode, bool IsINC) { |
386 | switch (LEAOpcode) { |
387 | default: |
388 | llvm_unreachable("Unexpected LEA instruction" ); |
389 | case X86::LEA32r: |
390 | case X86::LEA64_32r: |
391 | return IsINC ? X86::INC32r : X86::DEC32r; |
392 | case X86::LEA64r: |
393 | return IsINC ? X86::INC64r : X86::DEC64r; |
394 | } |
395 | } |
396 | |
397 | MachineBasicBlock::iterator |
398 | FixupLEAPass::searchALUInst(MachineBasicBlock::iterator &I, |
399 | MachineBasicBlock &MBB) const { |
400 | const int InstrDistanceThreshold = 5; |
401 | int InstrDistance = 1; |
402 | MachineBasicBlock::iterator CurInst = std::next(x: I); |
403 | |
404 | unsigned LEAOpcode = I->getOpcode(); |
405 | unsigned AddOpcode = getADDrrFromLEA(LEAOpcode); |
406 | unsigned SubOpcode = getSUBrrFromLEA(LEAOpcode); |
407 | Register DestReg = I->getOperand(i: 0).getReg(); |
408 | |
409 | while (CurInst != MBB.end()) { |
410 | if (CurInst->isCall() || CurInst->isInlineAsm()) |
411 | break; |
412 | if (InstrDistance > InstrDistanceThreshold) |
413 | break; |
414 | |
415 | // Check if the lea dest register is used in an add/sub instruction only. |
416 | for (unsigned I = 0, E = CurInst->getNumOperands(); I != E; ++I) { |
417 | MachineOperand &Opnd = CurInst->getOperand(i: I); |
418 | if (Opnd.isReg()) { |
419 | if (Opnd.getReg() == DestReg) { |
420 | if (Opnd.isDef() || !Opnd.isKill()) |
421 | return MachineBasicBlock::iterator(); |
422 | |
423 | unsigned AluOpcode = CurInst->getOpcode(); |
424 | if (AluOpcode != AddOpcode && AluOpcode != SubOpcode) |
425 | return MachineBasicBlock::iterator(); |
426 | |
427 | MachineOperand &Opnd2 = CurInst->getOperand(i: 3 - I); |
428 | MachineOperand AluDest = CurInst->getOperand(i: 0); |
429 | if (Opnd2.getReg() != AluDest.getReg()) |
430 | return MachineBasicBlock::iterator(); |
431 | |
432 | // X - (Y + Z) may generate different flags than (X - Y) - Z when |
433 | // there is overflow. So we can't change the alu instruction if the |
434 | // flags register is live. |
435 | if (!CurInst->registerDefIsDead(Reg: X86::EFLAGS, TRI)) |
436 | return MachineBasicBlock::iterator(); |
437 | |
438 | return CurInst; |
439 | } |
440 | if (TRI->regsOverlap(RegA: DestReg, RegB: Opnd.getReg())) |
441 | return MachineBasicBlock::iterator(); |
442 | } |
443 | } |
444 | |
445 | InstrDistance++; |
446 | ++CurInst; |
447 | } |
448 | return MachineBasicBlock::iterator(); |
449 | } |
450 | |
451 | void FixupLEAPass::checkRegUsage(MachineBasicBlock::iterator &LeaI, |
452 | MachineBasicBlock::iterator &AluI, |
453 | bool &BaseIndexDef, bool &AluDestRef, |
454 | MachineOperand **KilledBase, |
455 | MachineOperand **KilledIndex) const { |
456 | BaseIndexDef = AluDestRef = false; |
457 | *KilledBase = *KilledIndex = nullptr; |
458 | Register BaseReg = LeaI->getOperand(i: 1 + X86::AddrBaseReg).getReg(); |
459 | Register IndexReg = LeaI->getOperand(i: 1 + X86::AddrIndexReg).getReg(); |
460 | Register AluDestReg = AluI->getOperand(i: 0).getReg(); |
461 | |
462 | for (MachineInstr &CurInst : llvm::make_range(x: std::next(x: LeaI), y: AluI)) { |
463 | for (MachineOperand &Opnd : CurInst.operands()) { |
464 | if (!Opnd.isReg()) |
465 | continue; |
466 | Register Reg = Opnd.getReg(); |
467 | if (TRI->regsOverlap(RegA: Reg, RegB: AluDestReg)) |
468 | AluDestRef = true; |
469 | if (TRI->regsOverlap(RegA: Reg, RegB: BaseReg)) { |
470 | if (Opnd.isDef()) |
471 | BaseIndexDef = true; |
472 | else if (Opnd.isKill()) |
473 | *KilledBase = &Opnd; |
474 | } |
475 | if (TRI->regsOverlap(RegA: Reg, RegB: IndexReg)) { |
476 | if (Opnd.isDef()) |
477 | BaseIndexDef = true; |
478 | else if (Opnd.isKill()) |
479 | *KilledIndex = &Opnd; |
480 | } |
481 | } |
482 | } |
483 | } |
484 | |
485 | bool FixupLEAPass::optLEAALU(MachineBasicBlock::iterator &I, |
486 | MachineBasicBlock &MBB) const { |
487 | // Look for an add/sub instruction which uses the result of lea. |
488 | MachineBasicBlock::iterator AluI = searchALUInst(I, MBB); |
489 | if (AluI == MachineBasicBlock::iterator()) |
490 | return false; |
491 | |
492 | // Check if there are any related register usage between lea and alu. |
493 | bool BaseIndexDef, AluDestRef; |
494 | MachineOperand *KilledBase, *KilledIndex; |
495 | checkRegUsage(LeaI&: I, AluI, BaseIndexDef, AluDestRef, KilledBase: &KilledBase, KilledIndex: &KilledIndex); |
496 | |
497 | MachineBasicBlock::iterator InsertPos = AluI; |
498 | if (BaseIndexDef) { |
499 | if (AluDestRef) |
500 | return false; |
501 | InsertPos = I; |
502 | KilledBase = KilledIndex = nullptr; |
503 | } |
504 | |
505 | // Check if there are same registers. |
506 | Register AluDestReg = AluI->getOperand(i: 0).getReg(); |
507 | Register BaseReg = I->getOperand(i: 1 + X86::AddrBaseReg).getReg(); |
508 | Register IndexReg = I->getOperand(i: 1 + X86::AddrIndexReg).getReg(); |
509 | if (I->getOpcode() == X86::LEA64_32r) { |
510 | BaseReg = TRI->getSubReg(Reg: BaseReg, Idx: X86::sub_32bit); |
511 | IndexReg = TRI->getSubReg(Reg: IndexReg, Idx: X86::sub_32bit); |
512 | } |
513 | if (AluDestReg == IndexReg) { |
514 | if (BaseReg == IndexReg) |
515 | return false; |
516 | std::swap(a&: BaseReg, b&: IndexReg); |
517 | std::swap(a&: KilledBase, b&: KilledIndex); |
518 | } |
519 | if (BaseReg == IndexReg) |
520 | KilledBase = nullptr; |
521 | |
522 | // Now it's safe to change instructions. |
523 | MachineInstr *NewMI1, *NewMI2; |
524 | unsigned NewOpcode = AluI->getOpcode(); |
525 | NewMI1 = BuildMI(BB&: MBB, I: InsertPos, MIMD: AluI->getDebugLoc(), MCID: TII->get(Opcode: NewOpcode), |
526 | DestReg: AluDestReg) |
527 | .addReg(RegNo: AluDestReg, flags: RegState::Kill) |
528 | .addReg(RegNo: BaseReg, flags: KilledBase ? RegState::Kill : 0); |
529 | NewMI1->addRegisterDead(Reg: X86::EFLAGS, RegInfo: TRI); |
530 | NewMI2 = BuildMI(BB&: MBB, I: InsertPos, MIMD: AluI->getDebugLoc(), MCID: TII->get(Opcode: NewOpcode), |
531 | DestReg: AluDestReg) |
532 | .addReg(RegNo: AluDestReg, flags: RegState::Kill) |
533 | .addReg(RegNo: IndexReg, flags: KilledIndex ? RegState::Kill : 0); |
534 | NewMI2->addRegisterDead(Reg: X86::EFLAGS, RegInfo: TRI); |
535 | |
536 | // Clear the old Kill flags. |
537 | if (KilledBase) |
538 | KilledBase->setIsKill(false); |
539 | if (KilledIndex) |
540 | KilledIndex->setIsKill(false); |
541 | |
542 | MBB.getParent()->substituteDebugValuesForInst(Old: *AluI, New&: *NewMI2, MaxOperand: 1); |
543 | MBB.erase(I); |
544 | MBB.erase(I: AluI); |
545 | I = NewMI1; |
546 | return true; |
547 | } |
548 | |
549 | bool FixupLEAPass::optTwoAddrLEA(MachineBasicBlock::iterator &I, |
550 | MachineBasicBlock &MBB, bool OptIncDec, |
551 | bool UseLEAForSP) const { |
552 | MachineInstr &MI = *I; |
553 | |
554 | const MachineOperand &Base = MI.getOperand(i: 1 + X86::AddrBaseReg); |
555 | const MachineOperand &Scale = MI.getOperand(i: 1 + X86::AddrScaleAmt); |
556 | const MachineOperand &Index = MI.getOperand(i: 1 + X86::AddrIndexReg); |
557 | const MachineOperand &Disp = MI.getOperand(i: 1 + X86::AddrDisp); |
558 | const MachineOperand &Segment = MI.getOperand(i: 1 + X86::AddrSegmentReg); |
559 | |
560 | if (Segment.getReg() != 0 || !Disp.isImm() || Scale.getImm() > 1 || |
561 | MBB.computeRegisterLiveness(TRI, Reg: X86::EFLAGS, Before: I) != |
562 | MachineBasicBlock::LQR_Dead) |
563 | return false; |
564 | |
565 | Register DestReg = MI.getOperand(i: 0).getReg(); |
566 | Register BaseReg = Base.getReg(); |
567 | Register IndexReg = Index.getReg(); |
568 | |
569 | // Don't change stack adjustment LEAs. |
570 | if (UseLEAForSP && (DestReg == X86::ESP || DestReg == X86::RSP)) |
571 | return false; |
572 | |
573 | // LEA64_32 has 64-bit operands but 32-bit result. |
574 | if (MI.getOpcode() == X86::LEA64_32r) { |
575 | if (BaseReg != 0) |
576 | BaseReg = TRI->getSubReg(Reg: BaseReg, Idx: X86::sub_32bit); |
577 | if (IndexReg != 0) |
578 | IndexReg = TRI->getSubReg(Reg: IndexReg, Idx: X86::sub_32bit); |
579 | } |
580 | |
581 | MachineInstr *NewMI = nullptr; |
582 | |
583 | // Case 1. |
584 | // Look for lea(%reg1, %reg2), %reg1 or lea(%reg2, %reg1), %reg1 |
585 | // which can be turned into add %reg2, %reg1 |
586 | if (BaseReg != 0 && IndexReg != 0 && Disp.getImm() == 0 && |
587 | (DestReg == BaseReg || DestReg == IndexReg)) { |
588 | unsigned NewOpcode = getADDrrFromLEA(LEAOpcode: MI.getOpcode()); |
589 | if (DestReg != BaseReg) |
590 | std::swap(a&: BaseReg, b&: IndexReg); |
591 | |
592 | if (MI.getOpcode() == X86::LEA64_32r) { |
593 | // TODO: Do we need the super register implicit use? |
594 | NewMI = BuildMI(BB&: MBB, I, MIMD: MI.getDebugLoc(), MCID: TII->get(Opcode: NewOpcode), DestReg) |
595 | .addReg(RegNo: BaseReg).addReg(RegNo: IndexReg) |
596 | .addReg(RegNo: Base.getReg(), flags: RegState::Implicit) |
597 | .addReg(RegNo: Index.getReg(), flags: RegState::Implicit); |
598 | } else { |
599 | NewMI = BuildMI(BB&: MBB, I, MIMD: MI.getDebugLoc(), MCID: TII->get(Opcode: NewOpcode), DestReg) |
600 | .addReg(RegNo: BaseReg).addReg(RegNo: IndexReg); |
601 | } |
602 | } else if (DestReg == BaseReg && IndexReg == 0) { |
603 | // Case 2. |
604 | // This is an LEA with only a base register and a displacement, |
605 | // We can use ADDri or INC/DEC. |
606 | |
607 | // Does this LEA have one these forms: |
608 | // lea %reg, 1(%reg) |
609 | // lea %reg, -1(%reg) |
610 | if (OptIncDec && (Disp.getImm() == 1 || Disp.getImm() == -1)) { |
611 | bool IsINC = Disp.getImm() == 1; |
612 | unsigned NewOpcode = getINCDECFromLEA(LEAOpcode: MI.getOpcode(), IsINC); |
613 | |
614 | if (MI.getOpcode() == X86::LEA64_32r) { |
615 | // TODO: Do we need the super register implicit use? |
616 | NewMI = BuildMI(BB&: MBB, I, MIMD: MI.getDebugLoc(), MCID: TII->get(Opcode: NewOpcode), DestReg) |
617 | .addReg(RegNo: BaseReg).addReg(RegNo: Base.getReg(), flags: RegState::Implicit); |
618 | } else { |
619 | NewMI = BuildMI(BB&: MBB, I, MIMD: MI.getDebugLoc(), MCID: TII->get(Opcode: NewOpcode), DestReg) |
620 | .addReg(RegNo: BaseReg); |
621 | } |
622 | } else { |
623 | unsigned NewOpcode = getADDriFromLEA(LEAOpcode: MI.getOpcode(), Offset: Disp); |
624 | if (MI.getOpcode() == X86::LEA64_32r) { |
625 | // TODO: Do we need the super register implicit use? |
626 | NewMI = BuildMI(BB&: MBB, I, MIMD: MI.getDebugLoc(), MCID: TII->get(Opcode: NewOpcode), DestReg) |
627 | .addReg(RegNo: BaseReg).addImm(Val: Disp.getImm()) |
628 | .addReg(RegNo: Base.getReg(), flags: RegState::Implicit); |
629 | } else { |
630 | NewMI = BuildMI(BB&: MBB, I, MIMD: MI.getDebugLoc(), MCID: TII->get(Opcode: NewOpcode), DestReg) |
631 | .addReg(RegNo: BaseReg).addImm(Val: Disp.getImm()); |
632 | } |
633 | } |
634 | } else if (BaseReg != 0 && IndexReg != 0 && Disp.getImm() == 0) { |
635 | // Case 3. |
636 | // Look for and transform the sequence |
637 | // lea (reg1, reg2), reg3 |
638 | // sub reg3, reg4 |
639 | return optLEAALU(I, MBB); |
640 | } else |
641 | return false; |
642 | |
643 | MBB.getParent()->substituteDebugValuesForInst(Old: *I, New&: *NewMI, MaxOperand: 1); |
644 | MBB.erase(I); |
645 | I = NewMI; |
646 | return true; |
647 | } |
648 | |
649 | void FixupLEAPass::processInstruction(MachineBasicBlock::iterator &I, |
650 | MachineBasicBlock &MBB) { |
651 | // Process a load, store, or LEA instruction. |
652 | MachineInstr &MI = *I; |
653 | const MCInstrDesc &Desc = MI.getDesc(); |
654 | int AddrOffset = X86II::getMemoryOperandNo(TSFlags: Desc.TSFlags); |
655 | if (AddrOffset >= 0) { |
656 | AddrOffset += X86II::getOperandBias(Desc); |
657 | MachineOperand &p = MI.getOperand(i: AddrOffset + X86::AddrBaseReg); |
658 | if (p.isReg() && p.getReg() != X86::ESP) { |
659 | seekLEAFixup(p, I, MBB); |
660 | } |
661 | MachineOperand &q = MI.getOperand(i: AddrOffset + X86::AddrIndexReg); |
662 | if (q.isReg() && q.getReg() != X86::ESP) { |
663 | seekLEAFixup(p&: q, I, MBB); |
664 | } |
665 | } |
666 | } |
667 | |
668 | void FixupLEAPass::seekLEAFixup(MachineOperand &p, |
669 | MachineBasicBlock::iterator &I, |
670 | MachineBasicBlock &MBB) { |
671 | MachineBasicBlock::iterator MBI = searchBackwards(p, I, MBB); |
672 | if (MBI != MachineBasicBlock::iterator()) { |
673 | MachineInstr *NewMI = postRAConvertToLEA(MBB, MBBI&: MBI); |
674 | if (NewMI) { |
675 | ++NumLEAs; |
676 | LLVM_DEBUG(dbgs() << "FixLEA: Candidate to replace:" ; MBI->dump();); |
677 | // now to replace with an equivalent LEA... |
678 | LLVM_DEBUG(dbgs() << "FixLEA: Replaced by: " ; NewMI->dump();); |
679 | MBB.getParent()->substituteDebugValuesForInst(Old: *MBI, New&: *NewMI, MaxOperand: 1); |
680 | MBB.erase(I: MBI); |
681 | MachineBasicBlock::iterator J = |
682 | static_cast<MachineBasicBlock::iterator>(NewMI); |
683 | processInstruction(I&: J, MBB); |
684 | } |
685 | } |
686 | } |
687 | |
688 | void FixupLEAPass::processInstructionForSlowLEA(MachineBasicBlock::iterator &I, |
689 | MachineBasicBlock &MBB) { |
690 | MachineInstr &MI = *I; |
691 | const unsigned Opcode = MI.getOpcode(); |
692 | |
693 | const MachineOperand &Dst = MI.getOperand(i: 0); |
694 | const MachineOperand &Base = MI.getOperand(i: 1 + X86::AddrBaseReg); |
695 | const MachineOperand &Scale = MI.getOperand(i: 1 + X86::AddrScaleAmt); |
696 | const MachineOperand &Index = MI.getOperand(i: 1 + X86::AddrIndexReg); |
697 | const MachineOperand &Offset = MI.getOperand(i: 1 + X86::AddrDisp); |
698 | const MachineOperand &Segment = MI.getOperand(i: 1 + X86::AddrSegmentReg); |
699 | |
700 | if (Segment.getReg() != 0 || !Offset.isImm() || |
701 | MBB.computeRegisterLiveness(TRI, Reg: X86::EFLAGS, Before: I, Neighborhood: 4) != |
702 | MachineBasicBlock::LQR_Dead) |
703 | return; |
704 | const Register DstR = Dst.getReg(); |
705 | const Register SrcR1 = Base.getReg(); |
706 | const Register SrcR2 = Index.getReg(); |
707 | if ((SrcR1 == 0 || SrcR1 != DstR) && (SrcR2 == 0 || SrcR2 != DstR)) |
708 | return; |
709 | if (Scale.getImm() > 1) |
710 | return; |
711 | LLVM_DEBUG(dbgs() << "FixLEA: Candidate to replace:" ; I->dump();); |
712 | LLVM_DEBUG(dbgs() << "FixLEA: Replaced by: " ;); |
713 | MachineInstr *NewMI = nullptr; |
714 | // Make ADD instruction for two registers writing to LEA's destination |
715 | if (SrcR1 != 0 && SrcR2 != 0) { |
716 | const MCInstrDesc &ADDrr = TII->get(Opcode: getADDrrFromLEA(LEAOpcode: Opcode)); |
717 | const MachineOperand &Src = SrcR1 == DstR ? Index : Base; |
718 | NewMI = |
719 | BuildMI(BB&: MBB, I, MIMD: MI.getDebugLoc(), MCID: ADDrr, DestReg: DstR).addReg(RegNo: DstR).add(MO: Src); |
720 | LLVM_DEBUG(NewMI->dump();); |
721 | } |
722 | // Make ADD instruction for immediate |
723 | if (Offset.getImm() != 0) { |
724 | const MCInstrDesc &ADDri = |
725 | TII->get(Opcode: getADDriFromLEA(LEAOpcode: Opcode, Offset)); |
726 | const MachineOperand &SrcR = SrcR1 == DstR ? Base : Index; |
727 | NewMI = BuildMI(BB&: MBB, I, MIMD: MI.getDebugLoc(), MCID: ADDri, DestReg: DstR) |
728 | .add(MO: SrcR) |
729 | .addImm(Val: Offset.getImm()); |
730 | LLVM_DEBUG(NewMI->dump();); |
731 | } |
732 | if (NewMI) { |
733 | MBB.getParent()->substituteDebugValuesForInst(Old: *I, New&: *NewMI, MaxOperand: 1); |
734 | MBB.erase(I); |
735 | I = NewMI; |
736 | } |
737 | } |
738 | |
739 | void FixupLEAPass::processInstrForSlow3OpLEA(MachineBasicBlock::iterator &I, |
740 | MachineBasicBlock &MBB, |
741 | bool OptIncDec) { |
742 | MachineInstr &MI = *I; |
743 | const unsigned LEAOpcode = MI.getOpcode(); |
744 | |
745 | const MachineOperand &Dest = MI.getOperand(i: 0); |
746 | const MachineOperand &Base = MI.getOperand(i: 1 + X86::AddrBaseReg); |
747 | const MachineOperand &Scale = MI.getOperand(i: 1 + X86::AddrScaleAmt); |
748 | const MachineOperand &Index = MI.getOperand(i: 1 + X86::AddrIndexReg); |
749 | const MachineOperand &Offset = MI.getOperand(i: 1 + X86::AddrDisp); |
750 | const MachineOperand &Segment = MI.getOperand(i: 1 + X86::AddrSegmentReg); |
751 | |
752 | if (!(TII->isThreeOperandsLEA(MI) || hasInefficientLEABaseReg(Base, Index)) || |
753 | MBB.computeRegisterLiveness(TRI, Reg: X86::EFLAGS, Before: I, Neighborhood: 4) != |
754 | MachineBasicBlock::LQR_Dead || |
755 | Segment.getReg() != X86::NoRegister) |
756 | return; |
757 | |
758 | Register DestReg = Dest.getReg(); |
759 | Register BaseReg = Base.getReg(); |
760 | Register IndexReg = Index.getReg(); |
761 | |
762 | if (MI.getOpcode() == X86::LEA64_32r) { |
763 | if (BaseReg != 0) |
764 | BaseReg = TRI->getSubReg(Reg: BaseReg, Idx: X86::sub_32bit); |
765 | if (IndexReg != 0) |
766 | IndexReg = TRI->getSubReg(Reg: IndexReg, Idx: X86::sub_32bit); |
767 | } |
768 | |
769 | bool IsScale1 = Scale.getImm() == 1; |
770 | bool IsInefficientBase = isInefficientLEAReg(Reg: BaseReg); |
771 | bool IsInefficientIndex = isInefficientLEAReg(Reg: IndexReg); |
772 | |
773 | // Skip these cases since it takes more than 2 instructions |
774 | // to replace the LEA instruction. |
775 | if (IsInefficientBase && DestReg == BaseReg && !IsScale1) |
776 | return; |
777 | |
778 | LLVM_DEBUG(dbgs() << "FixLEA: Candidate to replace:" ; MI.dump();); |
779 | LLVM_DEBUG(dbgs() << "FixLEA: Replaced by: " ;); |
780 | |
781 | MachineInstr *NewMI = nullptr; |
782 | bool BaseOrIndexIsDst = DestReg == BaseReg || DestReg == IndexReg; |
783 | // First try and remove the base while sticking with LEA iff base == index and |
784 | // scale == 1. We can handle: |
785 | // 1. lea D(%base,%index,1) -> lea D(,%index,2) |
786 | // 2. lea D(%r13/%rbp,%index) -> lea D(,%index,2) |
787 | // Only do this if the LEA would otherwise be split into 2-instruction |
788 | // (either it has a an Offset or neither base nor index are dst) |
789 | if (IsScale1 && BaseReg == IndexReg && |
790 | (hasLEAOffset(Offset) || (IsInefficientBase && !BaseOrIndexIsDst))) { |
791 | NewMI = BuildMI(BB&: MBB, I&: MI, MIMD: MI.getDebugLoc(), MCID: TII->get(Opcode: LEAOpcode)) |
792 | .add(MO: Dest) |
793 | .addReg(RegNo: 0) |
794 | .addImm(Val: 2) |
795 | .add(MO: Index) |
796 | .add(MO: Offset) |
797 | .add(MO: Segment); |
798 | LLVM_DEBUG(NewMI->dump();); |
799 | |
800 | MBB.getParent()->substituteDebugValuesForInst(Old: *I, New&: *NewMI, MaxOperand: 1); |
801 | MBB.erase(I); |
802 | I = NewMI; |
803 | return; |
804 | } else if (IsScale1 && BaseOrIndexIsDst) { |
805 | // Try to replace LEA with one or two (for the 3-op LEA case) |
806 | // add instructions: |
807 | // 1.lea (%base,%index,1), %base => add %index,%base |
808 | // 2.lea (%base,%index,1), %index => add %base,%index |
809 | |
810 | unsigned NewOpc = getADDrrFromLEA(LEAOpcode: MI.getOpcode()); |
811 | if (DestReg != BaseReg) |
812 | std::swap(a&: BaseReg, b&: IndexReg); |
813 | |
814 | if (MI.getOpcode() == X86::LEA64_32r) { |
815 | // TODO: Do we need the super register implicit use? |
816 | NewMI = BuildMI(BB&: MBB, I, MIMD: MI.getDebugLoc(), MCID: TII->get(Opcode: NewOpc), DestReg) |
817 | .addReg(RegNo: BaseReg) |
818 | .addReg(RegNo: IndexReg) |
819 | .addReg(RegNo: Base.getReg(), flags: RegState::Implicit) |
820 | .addReg(RegNo: Index.getReg(), flags: RegState::Implicit); |
821 | } else { |
822 | NewMI = BuildMI(BB&: MBB, I, MIMD: MI.getDebugLoc(), MCID: TII->get(Opcode: NewOpc), DestReg) |
823 | .addReg(RegNo: BaseReg) |
824 | .addReg(RegNo: IndexReg); |
825 | } |
826 | } else if (!IsInefficientBase || (!IsInefficientIndex && IsScale1)) { |
827 | // If the base is inefficient try switching the index and base operands, |
828 | // otherwise just break the 3-Ops LEA inst into 2-Ops LEA + ADD instruction: |
829 | // lea offset(%base,%index,scale),%dst => |
830 | // lea (%base,%index,scale); add offset,%dst |
831 | NewMI = BuildMI(BB&: MBB, I&: MI, MIMD: MI.getDebugLoc(), MCID: TII->get(Opcode: LEAOpcode)) |
832 | .add(MO: Dest) |
833 | .add(MO: IsInefficientBase ? Index : Base) |
834 | .add(MO: Scale) |
835 | .add(MO: IsInefficientBase ? Base : Index) |
836 | .addImm(Val: 0) |
837 | .add(MO: Segment); |
838 | LLVM_DEBUG(NewMI->dump();); |
839 | } |
840 | |
841 | // If either replacement succeeded above, add the offset if needed, then |
842 | // replace the instruction. |
843 | if (NewMI) { |
844 | // Create ADD instruction for the Offset in case of 3-Ops LEA. |
845 | if (hasLEAOffset(Offset)) { |
846 | if (OptIncDec && Offset.isImm() && |
847 | (Offset.getImm() == 1 || Offset.getImm() == -1)) { |
848 | unsigned NewOpc = |
849 | getINCDECFromLEA(LEAOpcode: MI.getOpcode(), IsINC: Offset.getImm() == 1); |
850 | NewMI = BuildMI(BB&: MBB, I, MIMD: MI.getDebugLoc(), MCID: TII->get(Opcode: NewOpc), DestReg) |
851 | .addReg(RegNo: DestReg); |
852 | LLVM_DEBUG(NewMI->dump();); |
853 | } else { |
854 | unsigned NewOpc = getADDriFromLEA(LEAOpcode: MI.getOpcode(), Offset); |
855 | NewMI = BuildMI(BB&: MBB, I, MIMD: MI.getDebugLoc(), MCID: TII->get(Opcode: NewOpc), DestReg) |
856 | .addReg(RegNo: DestReg) |
857 | .add(MO: Offset); |
858 | LLVM_DEBUG(NewMI->dump();); |
859 | } |
860 | } |
861 | |
862 | MBB.getParent()->substituteDebugValuesForInst(Old: *I, New&: *NewMI, MaxOperand: 1); |
863 | MBB.erase(I); |
864 | I = NewMI; |
865 | return; |
866 | } |
867 | |
868 | // Handle the rest of the cases with inefficient base register: |
869 | assert(DestReg != BaseReg && "DestReg == BaseReg should be handled already!" ); |
870 | assert(IsInefficientBase && "efficient base should be handled already!" ); |
871 | |
872 | // FIXME: Handle LEA64_32r. |
873 | if (LEAOpcode == X86::LEA64_32r) |
874 | return; |
875 | |
876 | // lea (%base,%index,1), %dst => mov %base,%dst; add %index,%dst |
877 | if (IsScale1 && !hasLEAOffset(Offset)) { |
878 | bool BIK = Base.isKill() && BaseReg != IndexReg; |
879 | TII->copyPhysReg(MBB, MI, DL: MI.getDebugLoc(), DestReg, SrcReg: BaseReg, KillSrc: BIK); |
880 | LLVM_DEBUG(MI.getPrevNode()->dump();); |
881 | |
882 | unsigned NewOpc = getADDrrFromLEA(LEAOpcode: MI.getOpcode()); |
883 | NewMI = BuildMI(BB&: MBB, I&: MI, MIMD: MI.getDebugLoc(), MCID: TII->get(Opcode: NewOpc), DestReg) |
884 | .addReg(RegNo: DestReg) |
885 | .add(MO: Index); |
886 | LLVM_DEBUG(NewMI->dump();); |
887 | |
888 | MBB.getParent()->substituteDebugValuesForInst(Old: *I, New&: *NewMI, MaxOperand: 1); |
889 | MBB.erase(I); |
890 | I = NewMI; |
891 | return; |
892 | } |
893 | |
894 | // lea offset(%base,%index,scale), %dst => |
895 | // lea offset( ,%index,scale), %dst; add %base,%dst |
896 | NewMI = BuildMI(BB&: MBB, I&: MI, MIMD: MI.getDebugLoc(), MCID: TII->get(Opcode: LEAOpcode)) |
897 | .add(MO: Dest) |
898 | .addReg(RegNo: 0) |
899 | .add(MO: Scale) |
900 | .add(MO: Index) |
901 | .add(MO: Offset) |
902 | .add(MO: Segment); |
903 | LLVM_DEBUG(NewMI->dump();); |
904 | |
905 | unsigned NewOpc = getADDrrFromLEA(LEAOpcode: MI.getOpcode()); |
906 | NewMI = BuildMI(BB&: MBB, I&: MI, MIMD: MI.getDebugLoc(), MCID: TII->get(Opcode: NewOpc), DestReg) |
907 | .addReg(RegNo: DestReg) |
908 | .add(MO: Base); |
909 | LLVM_DEBUG(NewMI->dump();); |
910 | |
911 | MBB.getParent()->substituteDebugValuesForInst(Old: *I, New&: *NewMI, MaxOperand: 1); |
912 | MBB.erase(I); |
913 | I = NewMI; |
914 | } |
915 | |