1 | //===-- TargetInstrInfo.cpp - Target Instruction Information --------------===// |
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 implements the TargetInstrInfo class. |
10 | // |
11 | //===----------------------------------------------------------------------===// |
12 | |
13 | #include "llvm/CodeGen/TargetInstrInfo.h" |
14 | #include "llvm/ADT/SmallSet.h" |
15 | #include "llvm/ADT/StringExtras.h" |
16 | #include "llvm/BinaryFormat/Dwarf.h" |
17 | #include "llvm/CodeGen/MachineCombinerPattern.h" |
18 | #include "llvm/CodeGen/MachineFrameInfo.h" |
19 | #include "llvm/CodeGen/MachineInstrBuilder.h" |
20 | #include "llvm/CodeGen/MachineMemOperand.h" |
21 | #include "llvm/CodeGen/MachineRegisterInfo.h" |
22 | #include "llvm/CodeGen/MachineScheduler.h" |
23 | #include "llvm/CodeGen/MachineTraceMetrics.h" |
24 | #include "llvm/CodeGen/PseudoSourceValue.h" |
25 | #include "llvm/CodeGen/ScoreboardHazardRecognizer.h" |
26 | #include "llvm/CodeGen/StackMaps.h" |
27 | #include "llvm/CodeGen/TargetFrameLowering.h" |
28 | #include "llvm/CodeGen/TargetLowering.h" |
29 | #include "llvm/CodeGen/TargetRegisterInfo.h" |
30 | #include "llvm/CodeGen/TargetSchedule.h" |
31 | #include "llvm/IR/DataLayout.h" |
32 | #include "llvm/IR/DebugInfoMetadata.h" |
33 | #include "llvm/MC/MCAsmInfo.h" |
34 | #include "llvm/MC/MCInstrItineraries.h" |
35 | #include "llvm/Support/CommandLine.h" |
36 | #include "llvm/Support/ErrorHandling.h" |
37 | #include "llvm/Support/raw_ostream.h" |
38 | #include "llvm/Target/TargetMachine.h" |
39 | |
40 | using namespace llvm; |
41 | |
42 | static cl::opt<bool> DisableHazardRecognizer( |
43 | "disable-sched-hazard" , cl::Hidden, cl::init(Val: false), |
44 | cl::desc("Disable hazard detection during preRA scheduling" )); |
45 | |
46 | static cl::opt<bool> EnableAccReassociation( |
47 | "acc-reassoc" , cl::Hidden, cl::init(Val: true), |
48 | cl::desc("Enable reassociation of accumulation chains" )); |
49 | |
50 | static cl::opt<unsigned int> |
51 | MinAccumulatorDepth("acc-min-depth" , cl::Hidden, cl::init(Val: 8), |
52 | cl::desc("Minimum length of accumulator chains " |
53 | "required for the optimization to kick in" )); |
54 | |
55 | static cl::opt<unsigned int> MaxAccumulatorWidth( |
56 | "acc-max-width" , cl::Hidden, cl::init(Val: 3), |
57 | cl::desc("Maximum number of branches in the accumulator tree" )); |
58 | |
59 | TargetInstrInfo::~TargetInstrInfo() = default; |
60 | |
61 | const TargetRegisterClass* |
62 | TargetInstrInfo::getRegClass(const MCInstrDesc &MCID, unsigned OpNum, |
63 | const TargetRegisterInfo *TRI, |
64 | const MachineFunction &MF) const { |
65 | if (OpNum >= MCID.getNumOperands()) |
66 | return nullptr; |
67 | |
68 | short RegClass = MCID.operands()[OpNum].RegClass; |
69 | if (MCID.operands()[OpNum].isLookupPtrRegClass()) |
70 | return TRI->getPointerRegClass(MF, Kind: RegClass); |
71 | |
72 | // Instructions like INSERT_SUBREG do not have fixed register classes. |
73 | if (RegClass < 0) |
74 | return nullptr; |
75 | |
76 | // Otherwise just look it up normally. |
77 | return TRI->getRegClass(i: RegClass); |
78 | } |
79 | |
80 | /// insertNoop - Insert a noop into the instruction stream at the specified |
81 | /// point. |
82 | void TargetInstrInfo::insertNoop(MachineBasicBlock &MBB, |
83 | MachineBasicBlock::iterator MI) const { |
84 | llvm_unreachable("Target didn't implement insertNoop!" ); |
85 | } |
86 | |
87 | /// insertNoops - Insert noops into the instruction stream at the specified |
88 | /// point. |
89 | void TargetInstrInfo::insertNoops(MachineBasicBlock &MBB, |
90 | MachineBasicBlock::iterator MI, |
91 | unsigned Quantity) const { |
92 | for (unsigned i = 0; i < Quantity; ++i) |
93 | insertNoop(MBB, MI); |
94 | } |
95 | |
96 | static bool (const char *Str, const MCAsmInfo &MAI) { |
97 | return strncmp(s1: Str, s2: MAI.getCommentString().data(), |
98 | n: MAI.getCommentString().size()) == 0; |
99 | } |
100 | |
101 | /// Measure the specified inline asm to determine an approximation of its |
102 | /// length. |
103 | /// Comments (which run till the next SeparatorString or newline) do not |
104 | /// count as an instruction. |
105 | /// Any other non-whitespace text is considered an instruction, with |
106 | /// multiple instructions separated by SeparatorString or newlines. |
107 | /// Variable-length instructions are not handled here; this function |
108 | /// may be overloaded in the target code to do that. |
109 | /// We implement a special case of the .space directive which takes only a |
110 | /// single integer argument in base 10 that is the size in bytes. This is a |
111 | /// restricted form of the GAS directive in that we only interpret |
112 | /// simple--i.e. not a logical or arithmetic expression--size values without |
113 | /// the optional fill value. This is primarily used for creating arbitrary |
114 | /// sized inline asm blocks for testing purposes. |
115 | unsigned TargetInstrInfo::getInlineAsmLength( |
116 | const char *Str, |
117 | const MCAsmInfo &MAI, const TargetSubtargetInfo *STI) const { |
118 | // Count the number of instructions in the asm. |
119 | bool AtInsnStart = true; |
120 | unsigned Length = 0; |
121 | const unsigned MaxInstLength = MAI.getMaxInstLength(STI); |
122 | for (; *Str; ++Str) { |
123 | if (*Str == '\n' || strncmp(s1: Str, s2: MAI.getSeparatorString(), |
124 | n: strlen(s: MAI.getSeparatorString())) == 0) { |
125 | AtInsnStart = true; |
126 | } else if (isAsmComment(Str, MAI)) { |
127 | // Stop counting as an instruction after a comment until the next |
128 | // separator. |
129 | AtInsnStart = false; |
130 | } |
131 | |
132 | if (AtInsnStart && !isSpace(C: static_cast<unsigned char>(*Str))) { |
133 | unsigned AddLength = MaxInstLength; |
134 | if (strncmp(s1: Str, s2: ".space" , n: 6) == 0) { |
135 | char *EStr; |
136 | int SpaceSize; |
137 | SpaceSize = strtol(nptr: Str + 6, endptr: &EStr, base: 10); |
138 | SpaceSize = SpaceSize < 0 ? 0 : SpaceSize; |
139 | while (*EStr != '\n' && isSpace(C: static_cast<unsigned char>(*EStr))) |
140 | ++EStr; |
141 | if (*EStr == '\0' || *EStr == '\n' || |
142 | isAsmComment(Str: EStr, MAI)) // Successfully parsed .space argument |
143 | AddLength = SpaceSize; |
144 | } |
145 | Length += AddLength; |
146 | AtInsnStart = false; |
147 | } |
148 | } |
149 | |
150 | return Length; |
151 | } |
152 | |
153 | /// ReplaceTailWithBranchTo - Delete the instruction OldInst and everything |
154 | /// after it, replacing it with an unconditional branch to NewDest. |
155 | void |
156 | TargetInstrInfo::ReplaceTailWithBranchTo(MachineBasicBlock::iterator Tail, |
157 | MachineBasicBlock *NewDest) const { |
158 | MachineBasicBlock *MBB = Tail->getParent(); |
159 | |
160 | // Remove all the old successors of MBB from the CFG. |
161 | while (!MBB->succ_empty()) |
162 | MBB->removeSuccessor(I: MBB->succ_begin()); |
163 | |
164 | // Save off the debug loc before erasing the instruction. |
165 | DebugLoc DL = Tail->getDebugLoc(); |
166 | |
167 | // Update call info and remove all the dead instructions |
168 | // from the end of MBB. |
169 | while (Tail != MBB->end()) { |
170 | auto MI = Tail++; |
171 | if (MI->shouldUpdateAdditionalCallInfo()) |
172 | MBB->getParent()->eraseAdditionalCallInfo(MI: &*MI); |
173 | MBB->erase(I: MI); |
174 | } |
175 | |
176 | // If MBB isn't immediately before MBB, insert a branch to it. |
177 | if (++MachineFunction::iterator(MBB) != MachineFunction::iterator(NewDest)) |
178 | insertBranch(MBB&: *MBB, TBB: NewDest, FBB: nullptr, Cond: SmallVector<MachineOperand, 0>(), DL); |
179 | MBB->addSuccessor(Succ: NewDest); |
180 | } |
181 | |
182 | MachineInstr *TargetInstrInfo::commuteInstructionImpl(MachineInstr &MI, |
183 | bool NewMI, unsigned Idx1, |
184 | unsigned Idx2) const { |
185 | const MCInstrDesc &MCID = MI.getDesc(); |
186 | bool HasDef = MCID.getNumDefs(); |
187 | if (HasDef && !MI.getOperand(i: 0).isReg()) |
188 | // No idea how to commute this instruction. Target should implement its own. |
189 | return nullptr; |
190 | |
191 | unsigned CommutableOpIdx1 = Idx1; (void)CommutableOpIdx1; |
192 | unsigned CommutableOpIdx2 = Idx2; (void)CommutableOpIdx2; |
193 | assert(findCommutedOpIndices(MI, CommutableOpIdx1, CommutableOpIdx2) && |
194 | CommutableOpIdx1 == Idx1 && CommutableOpIdx2 == Idx2 && |
195 | "TargetInstrInfo::CommuteInstructionImpl(): not commutable operands." ); |
196 | assert(MI.getOperand(Idx1).isReg() && MI.getOperand(Idx2).isReg() && |
197 | "This only knows how to commute register operands so far" ); |
198 | |
199 | Register Reg0 = HasDef ? MI.getOperand(i: 0).getReg() : Register(); |
200 | Register Reg1 = MI.getOperand(i: Idx1).getReg(); |
201 | Register Reg2 = MI.getOperand(i: Idx2).getReg(); |
202 | unsigned SubReg0 = HasDef ? MI.getOperand(i: 0).getSubReg() : 0; |
203 | unsigned SubReg1 = MI.getOperand(i: Idx1).getSubReg(); |
204 | unsigned SubReg2 = MI.getOperand(i: Idx2).getSubReg(); |
205 | bool Reg1IsKill = MI.getOperand(i: Idx1).isKill(); |
206 | bool Reg2IsKill = MI.getOperand(i: Idx2).isKill(); |
207 | bool Reg1IsUndef = MI.getOperand(i: Idx1).isUndef(); |
208 | bool Reg2IsUndef = MI.getOperand(i: Idx2).isUndef(); |
209 | bool Reg1IsInternal = MI.getOperand(i: Idx1).isInternalRead(); |
210 | bool Reg2IsInternal = MI.getOperand(i: Idx2).isInternalRead(); |
211 | // Avoid calling isRenamable for virtual registers since we assert that |
212 | // renamable property is only queried/set for physical registers. |
213 | bool Reg1IsRenamable = |
214 | Reg1.isPhysical() ? MI.getOperand(i: Idx1).isRenamable() : false; |
215 | bool Reg2IsRenamable = |
216 | Reg2.isPhysical() ? MI.getOperand(i: Idx2).isRenamable() : false; |
217 | // If destination is tied to either of the commuted source register, then |
218 | // it must be updated. |
219 | if (HasDef && Reg0 == Reg1 && |
220 | MI.getDesc().getOperandConstraint(OpNum: Idx1, Constraint: MCOI::TIED_TO) == 0) { |
221 | Reg2IsKill = false; |
222 | Reg0 = Reg2; |
223 | SubReg0 = SubReg2; |
224 | } else if (HasDef && Reg0 == Reg2 && |
225 | MI.getDesc().getOperandConstraint(OpNum: Idx2, Constraint: MCOI::TIED_TO) == 0) { |
226 | Reg1IsKill = false; |
227 | Reg0 = Reg1; |
228 | SubReg0 = SubReg1; |
229 | } |
230 | |
231 | MachineInstr *CommutedMI = nullptr; |
232 | if (NewMI) { |
233 | // Create a new instruction. |
234 | MachineFunction &MF = *MI.getMF(); |
235 | CommutedMI = MF.CloneMachineInstr(Orig: &MI); |
236 | } else { |
237 | CommutedMI = &MI; |
238 | } |
239 | |
240 | if (HasDef) { |
241 | CommutedMI->getOperand(i: 0).setReg(Reg0); |
242 | CommutedMI->getOperand(i: 0).setSubReg(SubReg0); |
243 | } |
244 | CommutedMI->getOperand(i: Idx2).setReg(Reg1); |
245 | CommutedMI->getOperand(i: Idx1).setReg(Reg2); |
246 | CommutedMI->getOperand(i: Idx2).setSubReg(SubReg1); |
247 | CommutedMI->getOperand(i: Idx1).setSubReg(SubReg2); |
248 | CommutedMI->getOperand(i: Idx2).setIsKill(Reg1IsKill); |
249 | CommutedMI->getOperand(i: Idx1).setIsKill(Reg2IsKill); |
250 | CommutedMI->getOperand(i: Idx2).setIsUndef(Reg1IsUndef); |
251 | CommutedMI->getOperand(i: Idx1).setIsUndef(Reg2IsUndef); |
252 | CommutedMI->getOperand(i: Idx2).setIsInternalRead(Reg1IsInternal); |
253 | CommutedMI->getOperand(i: Idx1).setIsInternalRead(Reg2IsInternal); |
254 | // Avoid calling setIsRenamable for virtual registers since we assert that |
255 | // renamable property is only queried/set for physical registers. |
256 | if (Reg1.isPhysical()) |
257 | CommutedMI->getOperand(i: Idx2).setIsRenamable(Reg1IsRenamable); |
258 | if (Reg2.isPhysical()) |
259 | CommutedMI->getOperand(i: Idx1).setIsRenamable(Reg2IsRenamable); |
260 | return CommutedMI; |
261 | } |
262 | |
263 | MachineInstr *TargetInstrInfo::commuteInstruction(MachineInstr &MI, bool NewMI, |
264 | unsigned OpIdx1, |
265 | unsigned OpIdx2) const { |
266 | // If OpIdx1 or OpIdx2 is not specified, then this method is free to choose |
267 | // any commutable operand, which is done in findCommutedOpIndices() method |
268 | // called below. |
269 | if ((OpIdx1 == CommuteAnyOperandIndex || OpIdx2 == CommuteAnyOperandIndex) && |
270 | !findCommutedOpIndices(MI, SrcOpIdx1&: OpIdx1, SrcOpIdx2&: OpIdx2)) { |
271 | assert(MI.isCommutable() && |
272 | "Precondition violation: MI must be commutable." ); |
273 | return nullptr; |
274 | } |
275 | return commuteInstructionImpl(MI, NewMI, Idx1: OpIdx1, Idx2: OpIdx2); |
276 | } |
277 | |
278 | bool TargetInstrInfo::fixCommutedOpIndices(unsigned &ResultIdx1, |
279 | unsigned &ResultIdx2, |
280 | unsigned CommutableOpIdx1, |
281 | unsigned CommutableOpIdx2) { |
282 | if (ResultIdx1 == CommuteAnyOperandIndex && |
283 | ResultIdx2 == CommuteAnyOperandIndex) { |
284 | ResultIdx1 = CommutableOpIdx1; |
285 | ResultIdx2 = CommutableOpIdx2; |
286 | } else if (ResultIdx1 == CommuteAnyOperandIndex) { |
287 | if (ResultIdx2 == CommutableOpIdx1) |
288 | ResultIdx1 = CommutableOpIdx2; |
289 | else if (ResultIdx2 == CommutableOpIdx2) |
290 | ResultIdx1 = CommutableOpIdx1; |
291 | else |
292 | return false; |
293 | } else if (ResultIdx2 == CommuteAnyOperandIndex) { |
294 | if (ResultIdx1 == CommutableOpIdx1) |
295 | ResultIdx2 = CommutableOpIdx2; |
296 | else if (ResultIdx1 == CommutableOpIdx2) |
297 | ResultIdx2 = CommutableOpIdx1; |
298 | else |
299 | return false; |
300 | } else |
301 | // Check that the result operand indices match the given commutable |
302 | // operand indices. |
303 | return (ResultIdx1 == CommutableOpIdx1 && ResultIdx2 == CommutableOpIdx2) || |
304 | (ResultIdx1 == CommutableOpIdx2 && ResultIdx2 == CommutableOpIdx1); |
305 | |
306 | return true; |
307 | } |
308 | |
309 | bool TargetInstrInfo::findCommutedOpIndices(const MachineInstr &MI, |
310 | unsigned &SrcOpIdx1, |
311 | unsigned &SrcOpIdx2) const { |
312 | assert(!MI.isBundle() && |
313 | "TargetInstrInfo::findCommutedOpIndices() can't handle bundles" ); |
314 | |
315 | const MCInstrDesc &MCID = MI.getDesc(); |
316 | if (!MCID.isCommutable()) |
317 | return false; |
318 | |
319 | // This assumes v0 = op v1, v2 and commuting would swap v1 and v2. If this |
320 | // is not true, then the target must implement this. |
321 | unsigned CommutableOpIdx1 = MCID.getNumDefs(); |
322 | unsigned CommutableOpIdx2 = CommutableOpIdx1 + 1; |
323 | if (!fixCommutedOpIndices(ResultIdx1&: SrcOpIdx1, ResultIdx2&: SrcOpIdx2, |
324 | CommutableOpIdx1, CommutableOpIdx2)) |
325 | return false; |
326 | |
327 | if (!MI.getOperand(i: SrcOpIdx1).isReg() || !MI.getOperand(i: SrcOpIdx2).isReg()) |
328 | // No idea. |
329 | return false; |
330 | return true; |
331 | } |
332 | |
333 | bool TargetInstrInfo::isUnpredicatedTerminator(const MachineInstr &MI) const { |
334 | if (!MI.isTerminator()) return false; |
335 | |
336 | // Conditional branch is a special case. |
337 | if (MI.isBranch() && !MI.isBarrier()) |
338 | return true; |
339 | if (!MI.isPredicable()) |
340 | return true; |
341 | return !isPredicated(MI); |
342 | } |
343 | |
344 | bool TargetInstrInfo::PredicateInstruction( |
345 | MachineInstr &MI, ArrayRef<MachineOperand> Pred) const { |
346 | bool MadeChange = false; |
347 | |
348 | assert(!MI.isBundle() && |
349 | "TargetInstrInfo::PredicateInstruction() can't handle bundles" ); |
350 | |
351 | const MCInstrDesc &MCID = MI.getDesc(); |
352 | if (!MI.isPredicable()) |
353 | return false; |
354 | |
355 | for (unsigned j = 0, i = 0, e = MI.getNumOperands(); i != e; ++i) { |
356 | if (MCID.operands()[i].isPredicate()) { |
357 | MachineOperand &MO = MI.getOperand(i); |
358 | if (MO.isReg()) { |
359 | MO.setReg(Pred[j].getReg()); |
360 | MadeChange = true; |
361 | } else if (MO.isImm()) { |
362 | MO.setImm(Pred[j].getImm()); |
363 | MadeChange = true; |
364 | } else if (MO.isMBB()) { |
365 | MO.setMBB(Pred[j].getMBB()); |
366 | MadeChange = true; |
367 | } |
368 | ++j; |
369 | } |
370 | } |
371 | return MadeChange; |
372 | } |
373 | |
374 | bool TargetInstrInfo::hasLoadFromStackSlot( |
375 | const MachineInstr &MI, |
376 | SmallVectorImpl<const MachineMemOperand *> &Accesses) const { |
377 | size_t StartSize = Accesses.size(); |
378 | for (MachineInstr::mmo_iterator o = MI.memoperands_begin(), |
379 | oe = MI.memoperands_end(); |
380 | o != oe; ++o) { |
381 | if ((*o)->isLoad() && |
382 | isa_and_nonnull<FixedStackPseudoSourceValue>(Val: (*o)->getPseudoValue())) |
383 | Accesses.push_back(Elt: *o); |
384 | } |
385 | return Accesses.size() != StartSize; |
386 | } |
387 | |
388 | bool TargetInstrInfo::hasStoreToStackSlot( |
389 | const MachineInstr &MI, |
390 | SmallVectorImpl<const MachineMemOperand *> &Accesses) const { |
391 | size_t StartSize = Accesses.size(); |
392 | for (MachineInstr::mmo_iterator o = MI.memoperands_begin(), |
393 | oe = MI.memoperands_end(); |
394 | o != oe; ++o) { |
395 | if ((*o)->isStore() && |
396 | isa_and_nonnull<FixedStackPseudoSourceValue>(Val: (*o)->getPseudoValue())) |
397 | Accesses.push_back(Elt: *o); |
398 | } |
399 | return Accesses.size() != StartSize; |
400 | } |
401 | |
402 | bool TargetInstrInfo::getStackSlotRange(const TargetRegisterClass *RC, |
403 | unsigned SubIdx, unsigned &Size, |
404 | unsigned &Offset, |
405 | const MachineFunction &MF) const { |
406 | const TargetRegisterInfo *TRI = MF.getSubtarget().getRegisterInfo(); |
407 | if (!SubIdx) { |
408 | Size = TRI->getSpillSize(RC: *RC); |
409 | Offset = 0; |
410 | return true; |
411 | } |
412 | unsigned BitSize = TRI->getSubRegIdxSize(Idx: SubIdx); |
413 | // Convert bit size to byte size. |
414 | if (BitSize % 8) |
415 | return false; |
416 | |
417 | int BitOffset = TRI->getSubRegIdxOffset(Idx: SubIdx); |
418 | if (BitOffset < 0 || BitOffset % 8) |
419 | return false; |
420 | |
421 | Size = BitSize / 8; |
422 | Offset = (unsigned)BitOffset / 8; |
423 | |
424 | assert(TRI->getSpillSize(*RC) >= (Offset + Size) && "bad subregister range" ); |
425 | |
426 | if (!MF.getDataLayout().isLittleEndian()) { |
427 | Offset = TRI->getSpillSize(RC: *RC) - (Offset + Size); |
428 | } |
429 | return true; |
430 | } |
431 | |
432 | void TargetInstrInfo::reMaterialize(MachineBasicBlock &MBB, |
433 | MachineBasicBlock::iterator I, |
434 | Register DestReg, unsigned SubIdx, |
435 | const MachineInstr &Orig, |
436 | const TargetRegisterInfo &TRI) const { |
437 | MachineInstr *MI = MBB.getParent()->CloneMachineInstr(Orig: &Orig); |
438 | MI->substituteRegister(FromReg: MI->getOperand(i: 0).getReg(), ToReg: DestReg, SubIdx, RegInfo: TRI); |
439 | MBB.insert(I, MI); |
440 | } |
441 | |
442 | bool TargetInstrInfo::produceSameValue(const MachineInstr &MI0, |
443 | const MachineInstr &MI1, |
444 | const MachineRegisterInfo *MRI) const { |
445 | return MI0.isIdenticalTo(Other: MI1, Check: MachineInstr::IgnoreVRegDefs); |
446 | } |
447 | |
448 | MachineInstr & |
449 | TargetInstrInfo::duplicate(MachineBasicBlock &MBB, |
450 | MachineBasicBlock::iterator InsertBefore, |
451 | const MachineInstr &Orig) const { |
452 | MachineFunction &MF = *MBB.getParent(); |
453 | // CFI instructions are marked as non-duplicable, because Darwin compact |
454 | // unwind info emission can't handle multiple prologue setups. |
455 | assert((!Orig.isNotDuplicable() || |
456 | (!MF.getTarget().getTargetTriple().isOSDarwin() && |
457 | Orig.isCFIInstruction())) && |
458 | "Instruction cannot be duplicated" ); |
459 | |
460 | return MF.cloneMachineInstrBundle(MBB, InsertBefore, Orig); |
461 | } |
462 | |
463 | // If the COPY instruction in MI can be folded to a stack operation, return |
464 | // the register class to use. |
465 | static const TargetRegisterClass *canFoldCopy(const MachineInstr &MI, |
466 | const TargetInstrInfo &TII, |
467 | unsigned FoldIdx) { |
468 | assert(TII.isCopyInstr(MI) && "MI must be a COPY instruction" ); |
469 | if (MI.getNumOperands() != 2) |
470 | return nullptr; |
471 | assert(FoldIdx<2 && "FoldIdx refers no nonexistent operand" ); |
472 | |
473 | const MachineOperand &FoldOp = MI.getOperand(i: FoldIdx); |
474 | const MachineOperand &LiveOp = MI.getOperand(i: 1 - FoldIdx); |
475 | |
476 | if (FoldOp.getSubReg() || LiveOp.getSubReg()) |
477 | return nullptr; |
478 | |
479 | Register FoldReg = FoldOp.getReg(); |
480 | Register LiveReg = LiveOp.getReg(); |
481 | |
482 | assert(FoldReg.isVirtual() && "Cannot fold physregs" ); |
483 | |
484 | const MachineRegisterInfo &MRI = MI.getMF()->getRegInfo(); |
485 | const TargetRegisterClass *RC = MRI.getRegClass(Reg: FoldReg); |
486 | |
487 | if (LiveOp.getReg().isPhysical()) |
488 | return RC->contains(Reg: LiveOp.getReg()) ? RC : nullptr; |
489 | |
490 | if (RC->hasSubClassEq(RC: MRI.getRegClass(Reg: LiveReg))) |
491 | return RC; |
492 | |
493 | // FIXME: Allow folding when register classes are memory compatible. |
494 | return nullptr; |
495 | } |
496 | |
497 | MCInst TargetInstrInfo::getNop() const { llvm_unreachable("Not implemented" ); } |
498 | |
499 | /// Try to remove the load by folding it to a register |
500 | /// operand at the use. We fold the load instructions if load defines a virtual |
501 | /// register, the virtual register is used once in the same BB, and the |
502 | /// instructions in-between do not load or store, and have no side effects. |
503 | MachineInstr *TargetInstrInfo::optimizeLoadInstr(MachineInstr &MI, |
504 | const MachineRegisterInfo *MRI, |
505 | Register &FoldAsLoadDefReg, |
506 | MachineInstr *&DefMI) const { |
507 | // Check whether we can move DefMI here. |
508 | DefMI = MRI->getVRegDef(Reg: FoldAsLoadDefReg); |
509 | assert(DefMI); |
510 | bool SawStore = false; |
511 | if (!DefMI->isSafeToMove(SawStore)) |
512 | return nullptr; |
513 | |
514 | // Collect information about virtual register operands of MI. |
515 | SmallVector<unsigned, 1> SrcOperandIds; |
516 | for (unsigned i = 0, e = MI.getNumOperands(); i != e; ++i) { |
517 | MachineOperand &MO = MI.getOperand(i); |
518 | if (!MO.isReg()) |
519 | continue; |
520 | Register Reg = MO.getReg(); |
521 | if (Reg != FoldAsLoadDefReg) |
522 | continue; |
523 | // Do not fold if we have a subreg use or a def. |
524 | if (MO.getSubReg() || MO.isDef()) |
525 | return nullptr; |
526 | SrcOperandIds.push_back(Elt: i); |
527 | } |
528 | if (SrcOperandIds.empty()) |
529 | return nullptr; |
530 | |
531 | // Check whether we can fold the def into SrcOperandId. |
532 | if (MachineInstr *FoldMI = foldMemoryOperand(MI, Ops: SrcOperandIds, LoadMI&: *DefMI)) { |
533 | FoldAsLoadDefReg = 0; |
534 | return FoldMI; |
535 | } |
536 | |
537 | return nullptr; |
538 | } |
539 | |
540 | std::pair<unsigned, unsigned> |
541 | TargetInstrInfo::getPatchpointUnfoldableRange(const MachineInstr &MI) const { |
542 | switch (MI.getOpcode()) { |
543 | case TargetOpcode::STACKMAP: |
544 | // StackMapLiveValues are foldable |
545 | return std::make_pair(x: 0, y: StackMapOpers(&MI).getVarIdx()); |
546 | case TargetOpcode::PATCHPOINT: |
547 | // For PatchPoint, the call args are not foldable (even if reported in the |
548 | // stackmap e.g. via anyregcc). |
549 | return std::make_pair(x: 0, y: PatchPointOpers(&MI).getVarIdx()); |
550 | case TargetOpcode::STATEPOINT: |
551 | // For statepoints, fold deopt and gc arguments, but not call arguments. |
552 | return std::make_pair(x: MI.getNumDefs(), y: StatepointOpers(&MI).getVarIdx()); |
553 | default: |
554 | llvm_unreachable("unexpected stackmap opcode" ); |
555 | } |
556 | } |
557 | |
558 | static MachineInstr *foldPatchpoint(MachineFunction &MF, MachineInstr &MI, |
559 | ArrayRef<unsigned> Ops, int FrameIndex, |
560 | const TargetInstrInfo &TII) { |
561 | unsigned StartIdx = 0; |
562 | unsigned NumDefs = 0; |
563 | // getPatchpointUnfoldableRange throws guarantee if MI is not a patchpoint. |
564 | std::tie(args&: NumDefs, args&: StartIdx) = TII.getPatchpointUnfoldableRange(MI); |
565 | |
566 | unsigned DefToFoldIdx = MI.getNumOperands(); |
567 | |
568 | // Return false if any operands requested for folding are not foldable (not |
569 | // part of the stackmap's live values). |
570 | for (unsigned Op : Ops) { |
571 | if (Op < NumDefs) { |
572 | assert(DefToFoldIdx == MI.getNumOperands() && "Folding multiple defs" ); |
573 | DefToFoldIdx = Op; |
574 | } else if (Op < StartIdx) { |
575 | return nullptr; |
576 | } |
577 | if (MI.getOperand(i: Op).isTied()) |
578 | return nullptr; |
579 | } |
580 | |
581 | MachineInstr *NewMI = |
582 | MF.CreateMachineInstr(MCID: TII.get(Opcode: MI.getOpcode()), DL: MI.getDebugLoc(), NoImplicit: true); |
583 | MachineInstrBuilder MIB(MF, NewMI); |
584 | |
585 | // No need to fold return, the meta data, and function arguments |
586 | for (unsigned i = 0; i < StartIdx; ++i) |
587 | if (i != DefToFoldIdx) |
588 | MIB.add(MO: MI.getOperand(i)); |
589 | |
590 | for (unsigned i = StartIdx, e = MI.getNumOperands(); i < e; ++i) { |
591 | MachineOperand &MO = MI.getOperand(i); |
592 | unsigned TiedTo = e; |
593 | (void)MI.isRegTiedToDefOperand(UseOpIdx: i, DefOpIdx: &TiedTo); |
594 | |
595 | if (is_contained(Range&: Ops, Element: i)) { |
596 | assert(TiedTo == e && "Cannot fold tied operands" ); |
597 | unsigned SpillSize; |
598 | unsigned SpillOffset; |
599 | // Compute the spill slot size and offset. |
600 | const TargetRegisterClass *RC = |
601 | MF.getRegInfo().getRegClass(Reg: MO.getReg()); |
602 | bool Valid = |
603 | TII.getStackSlotRange(RC, SubIdx: MO.getSubReg(), Size&: SpillSize, Offset&: SpillOffset, MF); |
604 | if (!Valid) |
605 | report_fatal_error(reason: "cannot spill patchpoint subregister operand" ); |
606 | MIB.addImm(Val: StackMaps::IndirectMemRefOp); |
607 | MIB.addImm(Val: SpillSize); |
608 | MIB.addFrameIndex(Idx: FrameIndex); |
609 | MIB.addImm(Val: SpillOffset); |
610 | } else { |
611 | MIB.add(MO); |
612 | if (TiedTo < e) { |
613 | assert(TiedTo < NumDefs && "Bad tied operand" ); |
614 | if (TiedTo > DefToFoldIdx) |
615 | --TiedTo; |
616 | NewMI->tieOperands(DefIdx: TiedTo, UseIdx: NewMI->getNumOperands() - 1); |
617 | } |
618 | } |
619 | } |
620 | return NewMI; |
621 | } |
622 | |
623 | static void foldInlineAsmMemOperand(MachineInstr *MI, unsigned OpNo, int FI, |
624 | const TargetInstrInfo &TII) { |
625 | // If the machine operand is tied, untie it first. |
626 | if (MI->getOperand(i: OpNo).isTied()) { |
627 | unsigned TiedTo = MI->findTiedOperandIdx(OpIdx: OpNo); |
628 | MI->untieRegOperand(OpIdx: OpNo); |
629 | // Intentional recursion! |
630 | foldInlineAsmMemOperand(MI, OpNo: TiedTo, FI, TII); |
631 | } |
632 | |
633 | SmallVector<MachineOperand, 5> NewOps; |
634 | TII.getFrameIndexOperands(Ops&: NewOps, FI); |
635 | assert(!NewOps.empty() && "getFrameIndexOperands didn't create any operands" ); |
636 | MI->removeOperand(OpNo); |
637 | MI->insert(InsertBefore: MI->operands_begin() + OpNo, Ops: NewOps); |
638 | |
639 | // Change the previous operand to a MemKind InlineAsm::Flag. The second param |
640 | // is the per-target number of operands that represent the memory operand |
641 | // excluding this one (MD). This includes MO. |
642 | InlineAsm::Flag F(InlineAsm::Kind::Mem, NewOps.size()); |
643 | F.setMemConstraint(InlineAsm::ConstraintCode::m); |
644 | MachineOperand &MD = MI->getOperand(i: OpNo - 1); |
645 | MD.setImm(F); |
646 | } |
647 | |
648 | // Returns nullptr if not possible to fold. |
649 | static MachineInstr *foldInlineAsmMemOperand(MachineInstr &MI, |
650 | ArrayRef<unsigned> Ops, int FI, |
651 | const TargetInstrInfo &TII) { |
652 | assert(MI.isInlineAsm() && "wrong opcode" ); |
653 | if (Ops.size() > 1) |
654 | return nullptr; |
655 | unsigned Op = Ops[0]; |
656 | assert(Op && "should never be first operand" ); |
657 | assert(MI.getOperand(Op).isReg() && "shouldn't be folding non-reg operands" ); |
658 | |
659 | if (!MI.mayFoldInlineAsmRegOp(OpId: Op)) |
660 | return nullptr; |
661 | |
662 | MachineInstr &NewMI = TII.duplicate(MBB&: *MI.getParent(), InsertBefore: MI.getIterator(), Orig: MI); |
663 | |
664 | foldInlineAsmMemOperand(MI: &NewMI, OpNo: Op, FI, TII); |
665 | |
666 | // Update mayload/maystore metadata, and memoperands. |
667 | const VirtRegInfo &RI = |
668 | AnalyzeVirtRegInBundle(MI, Reg: MI.getOperand(i: Op).getReg()); |
669 | MachineOperand & = NewMI.getOperand(i: InlineAsm::MIOp_ExtraInfo); |
670 | MachineMemOperand::Flags Flags = MachineMemOperand::MONone; |
671 | if (RI.Reads) { |
672 | ExtraMO.setImm(ExtraMO.getImm() | InlineAsm::Extra_MayLoad); |
673 | Flags |= MachineMemOperand::MOLoad; |
674 | } |
675 | if (RI.Writes) { |
676 | ExtraMO.setImm(ExtraMO.getImm() | InlineAsm::Extra_MayStore); |
677 | Flags |= MachineMemOperand::MOStore; |
678 | } |
679 | MachineFunction *MF = NewMI.getMF(); |
680 | const MachineFrameInfo &MFI = MF->getFrameInfo(); |
681 | MachineMemOperand *MMO = MF->getMachineMemOperand( |
682 | PtrInfo: MachinePointerInfo::getFixedStack(MF&: *MF, FI), F: Flags, Size: MFI.getObjectSize(ObjectIdx: FI), |
683 | BaseAlignment: MFI.getObjectAlign(ObjectIdx: FI)); |
684 | NewMI.addMemOperand(MF&: *MF, MO: MMO); |
685 | |
686 | return &NewMI; |
687 | } |
688 | |
689 | MachineInstr *TargetInstrInfo::foldMemoryOperand(MachineInstr &MI, |
690 | ArrayRef<unsigned> Ops, int FI, |
691 | LiveIntervals *LIS, |
692 | VirtRegMap *VRM) const { |
693 | auto Flags = MachineMemOperand::MONone; |
694 | for (unsigned OpIdx : Ops) |
695 | Flags |= MI.getOperand(i: OpIdx).isDef() ? MachineMemOperand::MOStore |
696 | : MachineMemOperand::MOLoad; |
697 | |
698 | MachineBasicBlock *MBB = MI.getParent(); |
699 | assert(MBB && "foldMemoryOperand needs an inserted instruction" ); |
700 | MachineFunction &MF = *MBB->getParent(); |
701 | |
702 | // If we're not folding a load into a subreg, the size of the load is the |
703 | // size of the spill slot. But if we are, we need to figure out what the |
704 | // actual load size is. |
705 | int64_t MemSize = 0; |
706 | const MachineFrameInfo &MFI = MF.getFrameInfo(); |
707 | const TargetRegisterInfo *TRI = MF.getSubtarget().getRegisterInfo(); |
708 | |
709 | if (Flags & MachineMemOperand::MOStore) { |
710 | MemSize = MFI.getObjectSize(ObjectIdx: FI); |
711 | } else { |
712 | for (unsigned OpIdx : Ops) { |
713 | int64_t OpSize = MFI.getObjectSize(ObjectIdx: FI); |
714 | |
715 | if (auto SubReg = MI.getOperand(i: OpIdx).getSubReg()) { |
716 | unsigned SubRegSize = TRI->getSubRegIdxSize(Idx: SubReg); |
717 | if (SubRegSize > 0 && !(SubRegSize % 8)) |
718 | OpSize = SubRegSize / 8; |
719 | } |
720 | |
721 | MemSize = std::max(a: MemSize, b: OpSize); |
722 | } |
723 | } |
724 | |
725 | assert(MemSize && "Did not expect a zero-sized stack slot" ); |
726 | |
727 | MachineInstr *NewMI = nullptr; |
728 | |
729 | if (MI.getOpcode() == TargetOpcode::STACKMAP || |
730 | MI.getOpcode() == TargetOpcode::PATCHPOINT || |
731 | MI.getOpcode() == TargetOpcode::STATEPOINT) { |
732 | // Fold stackmap/patchpoint. |
733 | NewMI = foldPatchpoint(MF, MI, Ops, FrameIndex: FI, TII: *this); |
734 | if (NewMI) |
735 | MBB->insert(I: MI, MI: NewMI); |
736 | } else if (MI.isInlineAsm()) { |
737 | return foldInlineAsmMemOperand(MI, Ops, FI, TII: *this); |
738 | } else { |
739 | // Ask the target to do the actual folding. |
740 | NewMI = foldMemoryOperandImpl(MF, MI, Ops, InsertPt: MI, FrameIndex: FI, LIS, VRM); |
741 | } |
742 | |
743 | if (NewMI) { |
744 | NewMI->setMemRefs(MF, MemRefs: MI.memoperands()); |
745 | // Add a memory operand, foldMemoryOperandImpl doesn't do that. |
746 | assert((!(Flags & MachineMemOperand::MOStore) || |
747 | NewMI->mayStore()) && |
748 | "Folded a def to a non-store!" ); |
749 | assert((!(Flags & MachineMemOperand::MOLoad) || |
750 | NewMI->mayLoad()) && |
751 | "Folded a use to a non-load!" ); |
752 | assert(MFI.getObjectOffset(FI) != -1); |
753 | MachineMemOperand *MMO = |
754 | MF.getMachineMemOperand(PtrInfo: MachinePointerInfo::getFixedStack(MF, FI), |
755 | F: Flags, Size: MemSize, BaseAlignment: MFI.getObjectAlign(ObjectIdx: FI)); |
756 | NewMI->addMemOperand(MF, MO: MMO); |
757 | |
758 | // The pass "x86 speculative load hardening" always attaches symbols to |
759 | // call instructions. We need copy it form old instruction. |
760 | NewMI->cloneInstrSymbols(MF, MI); |
761 | |
762 | return NewMI; |
763 | } |
764 | |
765 | // Straight COPY may fold as load/store. |
766 | if (!isCopyInstr(MI) || Ops.size() != 1) |
767 | return nullptr; |
768 | |
769 | const TargetRegisterClass *RC = canFoldCopy(MI, TII: *this, FoldIdx: Ops[0]); |
770 | if (!RC) |
771 | return nullptr; |
772 | |
773 | const MachineOperand &MO = MI.getOperand(i: 1 - Ops[0]); |
774 | MachineBasicBlock::iterator Pos = MI; |
775 | |
776 | if (Flags == MachineMemOperand::MOStore) |
777 | storeRegToStackSlot(MBB&: *MBB, MI: Pos, SrcReg: MO.getReg(), isKill: MO.isKill(), FrameIndex: FI, RC, TRI, |
778 | VReg: Register()); |
779 | else |
780 | loadRegFromStackSlot(MBB&: *MBB, MI: Pos, DestReg: MO.getReg(), FrameIndex: FI, RC, TRI, VReg: Register()); |
781 | return &*--Pos; |
782 | } |
783 | |
784 | MachineInstr *TargetInstrInfo::foldMemoryOperand(MachineInstr &MI, |
785 | ArrayRef<unsigned> Ops, |
786 | MachineInstr &LoadMI, |
787 | LiveIntervals *LIS) const { |
788 | assert(LoadMI.canFoldAsLoad() && "LoadMI isn't foldable!" ); |
789 | #ifndef NDEBUG |
790 | for (unsigned OpIdx : Ops) |
791 | assert(MI.getOperand(OpIdx).isUse() && "Folding load into def!" ); |
792 | #endif |
793 | |
794 | MachineBasicBlock &MBB = *MI.getParent(); |
795 | MachineFunction &MF = *MBB.getParent(); |
796 | |
797 | // Ask the target to do the actual folding. |
798 | MachineInstr *NewMI = nullptr; |
799 | int FrameIndex = 0; |
800 | |
801 | if ((MI.getOpcode() == TargetOpcode::STACKMAP || |
802 | MI.getOpcode() == TargetOpcode::PATCHPOINT || |
803 | MI.getOpcode() == TargetOpcode::STATEPOINT) && |
804 | isLoadFromStackSlot(MI: LoadMI, FrameIndex)) { |
805 | // Fold stackmap/patchpoint. |
806 | NewMI = foldPatchpoint(MF, MI, Ops, FrameIndex, TII: *this); |
807 | if (NewMI) |
808 | NewMI = &*MBB.insert(I: MI, MI: NewMI); |
809 | } else if (MI.isInlineAsm() && isLoadFromStackSlot(MI: LoadMI, FrameIndex)) { |
810 | return foldInlineAsmMemOperand(MI, Ops, FI: FrameIndex, TII: *this); |
811 | } else { |
812 | // Ask the target to do the actual folding. |
813 | NewMI = foldMemoryOperandImpl(MF, MI, Ops, InsertPt: MI, LoadMI, LIS); |
814 | } |
815 | |
816 | if (!NewMI) |
817 | return nullptr; |
818 | |
819 | // Copy the memoperands from the load to the folded instruction. |
820 | if (MI.memoperands_empty()) { |
821 | NewMI->setMemRefs(MF, MemRefs: LoadMI.memoperands()); |
822 | } else { |
823 | // Handle the rare case of folding multiple loads. |
824 | NewMI->setMemRefs(MF, MemRefs: MI.memoperands()); |
825 | for (MachineInstr::mmo_iterator I = LoadMI.memoperands_begin(), |
826 | E = LoadMI.memoperands_end(); |
827 | I != E; ++I) { |
828 | NewMI->addMemOperand(MF, MO: *I); |
829 | } |
830 | } |
831 | return NewMI; |
832 | } |
833 | |
834 | /// transferImplicitOperands - MI is a pseudo-instruction, and the lowered |
835 | /// replacement instructions immediately precede it. Copy any implicit |
836 | /// operands from MI to the replacement instruction. |
837 | static void transferImplicitOperands(MachineInstr *MI, |
838 | const TargetRegisterInfo *TRI) { |
839 | MachineBasicBlock::iterator CopyMI = MI; |
840 | --CopyMI; |
841 | |
842 | Register DstReg = MI->getOperand(i: 0).getReg(); |
843 | for (const MachineOperand &MO : MI->implicit_operands()) { |
844 | CopyMI->addOperand(Op: MO); |
845 | |
846 | // Be conservative about preserving kills when subregister defs are |
847 | // involved. If there was implicit kill of a super-register overlapping the |
848 | // copy result, we would kill the subregisters previous copies defined. |
849 | |
850 | if (MO.isKill() && TRI->regsOverlap(RegA: DstReg, RegB: MO.getReg())) |
851 | CopyMI->getOperand(i: CopyMI->getNumOperands() - 1).setIsKill(false); |
852 | } |
853 | } |
854 | |
855 | void TargetInstrInfo::lowerCopy(MachineInstr *MI, |
856 | const TargetRegisterInfo *TRI) const { |
857 | if (MI->allDefsAreDead()) { |
858 | MI->setDesc(get(Opcode: TargetOpcode::KILL)); |
859 | return; |
860 | } |
861 | |
862 | MachineOperand &DstMO = MI->getOperand(i: 0); |
863 | MachineOperand &SrcMO = MI->getOperand(i: 1); |
864 | |
865 | bool IdentityCopy = (SrcMO.getReg() == DstMO.getReg()); |
866 | if (IdentityCopy || SrcMO.isUndef()) { |
867 | // No need to insert an identity copy instruction, but replace with a KILL |
868 | // if liveness is changed. |
869 | if (SrcMO.isUndef() || MI->getNumOperands() > 2) { |
870 | // We must make sure the super-register gets killed. Replace the |
871 | // instruction with KILL. |
872 | MI->setDesc(get(Opcode: TargetOpcode::KILL)); |
873 | return; |
874 | } |
875 | // Vanilla identity copy. |
876 | MI->eraseFromParent(); |
877 | return; |
878 | } |
879 | |
880 | copyPhysReg(MBB&: *MI->getParent(), MI, DL: MI->getDebugLoc(), DestReg: DstMO.getReg(), |
881 | SrcReg: SrcMO.getReg(), KillSrc: SrcMO.isKill(), |
882 | RenamableDest: DstMO.getReg().isPhysical() ? DstMO.isRenamable() : false, |
883 | RenamableSrc: SrcMO.getReg().isPhysical() ? SrcMO.isRenamable() : false); |
884 | |
885 | if (MI->getNumOperands() > 2) |
886 | transferImplicitOperands(MI, TRI); |
887 | MI->eraseFromParent(); |
888 | } |
889 | |
890 | bool TargetInstrInfo::hasReassociableOperands( |
891 | const MachineInstr &Inst, const MachineBasicBlock *MBB) const { |
892 | const MachineOperand &Op1 = Inst.getOperand(i: 1); |
893 | const MachineOperand &Op2 = Inst.getOperand(i: 2); |
894 | const MachineRegisterInfo &MRI = MBB->getParent()->getRegInfo(); |
895 | |
896 | // We need virtual register definitions for the operands that we will |
897 | // reassociate. |
898 | MachineInstr *MI1 = nullptr; |
899 | MachineInstr *MI2 = nullptr; |
900 | if (Op1.isReg() && Op1.getReg().isVirtual()) |
901 | MI1 = MRI.getUniqueVRegDef(Reg: Op1.getReg()); |
902 | if (Op2.isReg() && Op2.getReg().isVirtual()) |
903 | MI2 = MRI.getUniqueVRegDef(Reg: Op2.getReg()); |
904 | |
905 | // And at least one operand must be defined in MBB. |
906 | return MI1 && MI2 && (MI1->getParent() == MBB || MI2->getParent() == MBB); |
907 | } |
908 | |
909 | bool TargetInstrInfo::areOpcodesEqualOrInverse(unsigned Opcode1, |
910 | unsigned Opcode2) const { |
911 | return Opcode1 == Opcode2 || getInverseOpcode(Opcode: Opcode1) == Opcode2; |
912 | } |
913 | |
914 | bool TargetInstrInfo::hasReassociableSibling(const MachineInstr &Inst, |
915 | bool &Commuted) const { |
916 | const MachineBasicBlock *MBB = Inst.getParent(); |
917 | const MachineRegisterInfo &MRI = MBB->getParent()->getRegInfo(); |
918 | MachineInstr *MI1 = MRI.getUniqueVRegDef(Reg: Inst.getOperand(i: 1).getReg()); |
919 | MachineInstr *MI2 = MRI.getUniqueVRegDef(Reg: Inst.getOperand(i: 2).getReg()); |
920 | unsigned Opcode = Inst.getOpcode(); |
921 | |
922 | // If only one operand has the same or inverse opcode and it's the second |
923 | // source operand, the operands must be commuted. |
924 | Commuted = !areOpcodesEqualOrInverse(Opcode1: Opcode, Opcode2: MI1->getOpcode()) && |
925 | areOpcodesEqualOrInverse(Opcode1: Opcode, Opcode2: MI2->getOpcode()); |
926 | if (Commuted) |
927 | std::swap(a&: MI1, b&: MI2); |
928 | |
929 | // 1. The previous instruction must be the same type as Inst. |
930 | // 2. The previous instruction must also be associative/commutative or be the |
931 | // inverse of such an operation (this can be different even for |
932 | // instructions with the same opcode if traits like fast-math-flags are |
933 | // included). |
934 | // 3. The previous instruction must have virtual register definitions for its |
935 | // operands in the same basic block as Inst. |
936 | // 4. The previous instruction's result must only be used by Inst. |
937 | return areOpcodesEqualOrInverse(Opcode1: Opcode, Opcode2: MI1->getOpcode()) && |
938 | (isAssociativeAndCommutative(Inst: *MI1) || |
939 | isAssociativeAndCommutative(Inst: *MI1, /* Invert */ true)) && |
940 | hasReassociableOperands(Inst: *MI1, MBB) && |
941 | MRI.hasOneNonDBGUse(RegNo: MI1->getOperand(i: 0).getReg()); |
942 | } |
943 | |
944 | // 1. The operation must be associative and commutative or be the inverse of |
945 | // such an operation. |
946 | // 2. The instruction must have virtual register definitions for its |
947 | // operands in the same basic block. |
948 | // 3. The instruction must have a reassociable sibling. |
949 | bool TargetInstrInfo::isReassociationCandidate(const MachineInstr &Inst, |
950 | bool &Commuted) const { |
951 | return (isAssociativeAndCommutative(Inst) || |
952 | isAssociativeAndCommutative(Inst, /* Invert */ true)) && |
953 | hasReassociableOperands(Inst, MBB: Inst.getParent()) && |
954 | hasReassociableSibling(Inst, Commuted); |
955 | } |
956 | |
957 | // Utility routine that checks if \param MO is defined by an |
958 | // \param CombineOpc instruction in the basic block \param MBB. |
959 | // If \param CombineOpc is not provided, the OpCode check will |
960 | // be skipped. |
961 | static bool canCombine(MachineBasicBlock &MBB, MachineOperand &MO, |
962 | unsigned CombineOpc = 0) { |
963 | MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo(); |
964 | MachineInstr *MI = nullptr; |
965 | |
966 | if (MO.isReg() && MO.getReg().isVirtual()) |
967 | MI = MRI.getUniqueVRegDef(Reg: MO.getReg()); |
968 | // And it needs to be in the trace (otherwise, it won't have a depth). |
969 | if (!MI || MI->getParent() != &MBB || |
970 | ((unsigned)MI->getOpcode() != CombineOpc && CombineOpc != 0)) |
971 | return false; |
972 | // Must only used by the user we combine with. |
973 | if (!MRI.hasOneNonDBGUse(RegNo: MI->getOperand(i: 0).getReg())) |
974 | return false; |
975 | |
976 | return true; |
977 | } |
978 | |
979 | // A chain of accumulation instructions will be selected IFF: |
980 | // 1. All the accumulation instructions in the chain have the same opcode, |
981 | // besides the first that has a slightly different opcode because it does |
982 | // not accumulate into a register. |
983 | // 2. All the instructions in the chain are combinable (have a single use |
984 | // which itself is part of the chain). |
985 | // 3. Meets the required minimum length. |
986 | void TargetInstrInfo::getAccumulatorChain( |
987 | MachineInstr *CurrentInstr, SmallVectorImpl<Register> &Chain) const { |
988 | // Walk up the chain of accumulation instructions and collect them in the |
989 | // vector. |
990 | MachineBasicBlock &MBB = *CurrentInstr->getParent(); |
991 | const MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo(); |
992 | unsigned AccumulatorOpcode = CurrentInstr->getOpcode(); |
993 | std::optional<unsigned> ChainStartOpCode = |
994 | getAccumulationStartOpcode(Opcode: AccumulatorOpcode); |
995 | |
996 | if (!ChainStartOpCode.has_value()) |
997 | return; |
998 | |
999 | // Push the first accumulator result to the start of the chain. |
1000 | Chain.push_back(Elt: CurrentInstr->getOperand(i: 0).getReg()); |
1001 | |
1002 | // Collect the accumulator input register from all instructions in the chain. |
1003 | while (CurrentInstr && |
1004 | canCombine(MBB, MO&: CurrentInstr->getOperand(i: 1), CombineOpc: AccumulatorOpcode)) { |
1005 | Chain.push_back(Elt: CurrentInstr->getOperand(i: 1).getReg()); |
1006 | CurrentInstr = MRI.getUniqueVRegDef(Reg: CurrentInstr->getOperand(i: 1).getReg()); |
1007 | } |
1008 | |
1009 | // Add the instruction at the top of the chain. |
1010 | if (CurrentInstr->getOpcode() == AccumulatorOpcode && |
1011 | canCombine(MBB, MO&: CurrentInstr->getOperand(i: 1))) |
1012 | Chain.push_back(Elt: CurrentInstr->getOperand(i: 1).getReg()); |
1013 | } |
1014 | |
1015 | /// Find chains of accumulations that can be rewritten as a tree for increased |
1016 | /// ILP. |
1017 | bool TargetInstrInfo::getAccumulatorReassociationPatterns( |
1018 | MachineInstr &Root, SmallVectorImpl<unsigned> &Patterns) const { |
1019 | if (!EnableAccReassociation) |
1020 | return false; |
1021 | |
1022 | unsigned Opc = Root.getOpcode(); |
1023 | if (!isAccumulationOpcode(Opcode: Opc)) |
1024 | return false; |
1025 | |
1026 | // Verify that this is the end of the chain. |
1027 | MachineBasicBlock &MBB = *Root.getParent(); |
1028 | MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo(); |
1029 | if (!MRI.hasOneNonDBGUser(RegNo: Root.getOperand(i: 0).getReg())) |
1030 | return false; |
1031 | |
1032 | auto User = MRI.use_instr_begin(RegNo: Root.getOperand(i: 0).getReg()); |
1033 | if (User->getOpcode() == Opc) |
1034 | return false; |
1035 | |
1036 | // Walk up the use chain and collect the reduction chain. |
1037 | SmallVector<Register, 32> Chain; |
1038 | getAccumulatorChain(CurrentInstr: &Root, Chain); |
1039 | |
1040 | // Reject chains which are too short to be worth modifying. |
1041 | if (Chain.size() < MinAccumulatorDepth) |
1042 | return false; |
1043 | |
1044 | // Check if the MBB this instruction is a part of contains any other chains. |
1045 | // If so, don't apply it. |
1046 | SmallSet<Register, 32> ReductionChain(llvm::from_range, Chain); |
1047 | for (const auto &I : MBB) { |
1048 | if (I.getOpcode() == Opc && |
1049 | !ReductionChain.contains(V: I.getOperand(i: 0).getReg())) |
1050 | return false; |
1051 | } |
1052 | |
1053 | Patterns.push_back(Elt: MachineCombinerPattern::ACC_CHAIN); |
1054 | return true; |
1055 | } |
1056 | |
1057 | // Reduce branches of the accumulator tree by adding them together. |
1058 | void TargetInstrInfo::reduceAccumulatorTree( |
1059 | SmallVectorImpl<Register> &RegistersToReduce, |
1060 | SmallVectorImpl<MachineInstr *> &InsInstrs, MachineFunction &MF, |
1061 | MachineInstr &Root, MachineRegisterInfo &MRI, |
1062 | DenseMap<Register, unsigned> &InstrIdxForVirtReg, |
1063 | Register ResultReg) const { |
1064 | const TargetInstrInfo *TII = MF.getSubtarget().getInstrInfo(); |
1065 | SmallVector<Register, 8> NewRegs; |
1066 | |
1067 | // Get the opcode for the reduction instruction we will need to build. |
1068 | // If for some reason it is not defined, early exit and don't apply this. |
1069 | unsigned ReduceOpCode = getReduceOpcodeForAccumulator(AccumulatorOpCode: Root.getOpcode()); |
1070 | |
1071 | for (unsigned int i = 1; i <= (RegistersToReduce.size() / 2); i += 2) { |
1072 | auto RHS = RegistersToReduce[i - 1]; |
1073 | auto LHS = RegistersToReduce[i]; |
1074 | Register Dest; |
1075 | // If we are reducing 2 registers, reuse the original result register. |
1076 | if (RegistersToReduce.size() == 2) |
1077 | Dest = ResultReg; |
1078 | // Otherwise, create a new virtual register to hold the partial sum. |
1079 | else { |
1080 | auto NewVR = MRI.createVirtualRegister( |
1081 | RegClass: MRI.getRegClass(Reg: Root.getOperand(i: 0).getReg())); |
1082 | Dest = NewVR; |
1083 | NewRegs.push_back(Elt: Dest); |
1084 | InstrIdxForVirtReg.insert(KV: std::make_pair(x&: Dest, y: InsInstrs.size())); |
1085 | } |
1086 | |
1087 | // Create the new reduction instruction. |
1088 | MachineInstrBuilder MIB = |
1089 | BuildMI(MF, MIMD: MIMetadata(Root), MCID: TII->get(Opcode: ReduceOpCode), DestReg: Dest) |
1090 | .addReg(RegNo: RHS, flags: getKillRegState(B: true)) |
1091 | .addReg(RegNo: LHS, flags: getKillRegState(B: true)); |
1092 | // Copy any flags needed from the original instruction. |
1093 | MIB->setFlags(Root.getFlags()); |
1094 | InsInstrs.push_back(Elt: MIB); |
1095 | } |
1096 | |
1097 | // If the number of registers to reduce is odd, add the remaining register to |
1098 | // the vector of registers to reduce. |
1099 | if (RegistersToReduce.size() % 2 != 0) |
1100 | NewRegs.push_back(Elt: RegistersToReduce[RegistersToReduce.size() - 1]); |
1101 | |
1102 | RegistersToReduce = NewRegs; |
1103 | } |
1104 | |
1105 | // The concept of the reassociation pass is that these operations can benefit |
1106 | // from this kind of transformation: |
1107 | // |
1108 | // A = ? op ? |
1109 | // B = A op X (Prev) |
1110 | // C = B op Y (Root) |
1111 | // --> |
1112 | // A = ? op ? |
1113 | // B = X op Y |
1114 | // C = A op B |
1115 | // |
1116 | // breaking the dependency between A and B, allowing them to be executed in |
1117 | // parallel (or back-to-back in a pipeline) instead of depending on each other. |
1118 | |
1119 | // FIXME: This has the potential to be expensive (compile time) while not |
1120 | // improving the code at all. Some ways to limit the overhead: |
1121 | // 1. Track successful transforms; bail out if hit rate gets too low. |
1122 | // 2. Only enable at -O3 or some other non-default optimization level. |
1123 | // 3. Pre-screen pattern candidates here: if an operand of the previous |
1124 | // instruction is known to not increase the critical path, then don't match |
1125 | // that pattern. |
1126 | bool TargetInstrInfo::getMachineCombinerPatterns( |
1127 | MachineInstr &Root, SmallVectorImpl<unsigned> &Patterns, |
1128 | bool DoRegPressureReduce) const { |
1129 | bool Commute; |
1130 | if (isReassociationCandidate(Inst: Root, Commuted&: Commute)) { |
1131 | // We found a sequence of instructions that may be suitable for a |
1132 | // reassociation of operands to increase ILP. Specify each commutation |
1133 | // possibility for the Prev instruction in the sequence and let the |
1134 | // machine combiner decide if changing the operands is worthwhile. |
1135 | if (Commute) { |
1136 | Patterns.push_back(Elt: MachineCombinerPattern::REASSOC_AX_YB); |
1137 | Patterns.push_back(Elt: MachineCombinerPattern::REASSOC_XA_YB); |
1138 | } else { |
1139 | Patterns.push_back(Elt: MachineCombinerPattern::REASSOC_AX_BY); |
1140 | Patterns.push_back(Elt: MachineCombinerPattern::REASSOC_XA_BY); |
1141 | } |
1142 | return true; |
1143 | } |
1144 | if (getAccumulatorReassociationPatterns(Root, Patterns)) |
1145 | return true; |
1146 | |
1147 | return false; |
1148 | } |
1149 | |
1150 | /// Return true when a code sequence can improve loop throughput. |
1151 | bool TargetInstrInfo::isThroughputPattern(unsigned Pattern) const { |
1152 | return false; |
1153 | } |
1154 | |
1155 | CombinerObjective |
1156 | TargetInstrInfo::getCombinerObjective(unsigned Pattern) const { |
1157 | switch (Pattern) { |
1158 | case MachineCombinerPattern::ACC_CHAIN: |
1159 | return CombinerObjective::MustReduceDepth; |
1160 | default: |
1161 | return CombinerObjective::Default; |
1162 | } |
1163 | } |
1164 | |
1165 | std::pair<unsigned, unsigned> |
1166 | TargetInstrInfo::getReassociationOpcodes(unsigned Pattern, |
1167 | const MachineInstr &Root, |
1168 | const MachineInstr &Prev) const { |
1169 | bool AssocCommutRoot = isAssociativeAndCommutative(Inst: Root); |
1170 | bool AssocCommutPrev = isAssociativeAndCommutative(Inst: Prev); |
1171 | |
1172 | // Early exit if both opcodes are associative and commutative. It's a trivial |
1173 | // reassociation when we only change operands order. In this case opcodes are |
1174 | // not required to have inverse versions. |
1175 | if (AssocCommutRoot && AssocCommutPrev) { |
1176 | assert(Root.getOpcode() == Prev.getOpcode() && "Expected to be equal" ); |
1177 | return std::make_pair(x: Root.getOpcode(), y: Root.getOpcode()); |
1178 | } |
1179 | |
1180 | // At least one instruction is not associative or commutative. |
1181 | // Since we have matched one of the reassociation patterns, we expect that the |
1182 | // instructions' opcodes are equal or one of them is the inversion of the |
1183 | // other. |
1184 | assert(areOpcodesEqualOrInverse(Root.getOpcode(), Prev.getOpcode()) && |
1185 | "Incorrectly matched pattern" ); |
1186 | unsigned AssocCommutOpcode = Root.getOpcode(); |
1187 | unsigned InverseOpcode = *getInverseOpcode(Opcode: Root.getOpcode()); |
1188 | if (!AssocCommutRoot) |
1189 | std::swap(a&: AssocCommutOpcode, b&: InverseOpcode); |
1190 | |
1191 | // The transformation rule (`+` is any associative and commutative binary |
1192 | // operation, `-` is the inverse): |
1193 | // REASSOC_AX_BY: |
1194 | // (A + X) + Y => A + (X + Y) |
1195 | // (A + X) - Y => A + (X - Y) |
1196 | // (A - X) + Y => A - (X - Y) |
1197 | // (A - X) - Y => A - (X + Y) |
1198 | // REASSOC_XA_BY: |
1199 | // (X + A) + Y => (X + Y) + A |
1200 | // (X + A) - Y => (X - Y) + A |
1201 | // (X - A) + Y => (X + Y) - A |
1202 | // (X - A) - Y => (X - Y) - A |
1203 | // REASSOC_AX_YB: |
1204 | // Y + (A + X) => (Y + X) + A |
1205 | // Y - (A + X) => (Y - X) - A |
1206 | // Y + (A - X) => (Y - X) + A |
1207 | // Y - (A - X) => (Y + X) - A |
1208 | // REASSOC_XA_YB: |
1209 | // Y + (X + A) => (Y + X) + A |
1210 | // Y - (X + A) => (Y - X) - A |
1211 | // Y + (X - A) => (Y + X) - A |
1212 | // Y - (X - A) => (Y - X) + A |
1213 | switch (Pattern) { |
1214 | default: |
1215 | llvm_unreachable("Unexpected pattern" ); |
1216 | case MachineCombinerPattern::REASSOC_AX_BY: |
1217 | if (!AssocCommutRoot && AssocCommutPrev) |
1218 | return {AssocCommutOpcode, InverseOpcode}; |
1219 | if (AssocCommutRoot && !AssocCommutPrev) |
1220 | return {InverseOpcode, InverseOpcode}; |
1221 | if (!AssocCommutRoot && !AssocCommutPrev) |
1222 | return {InverseOpcode, AssocCommutOpcode}; |
1223 | break; |
1224 | case MachineCombinerPattern::REASSOC_XA_BY: |
1225 | if (!AssocCommutRoot && AssocCommutPrev) |
1226 | return {AssocCommutOpcode, InverseOpcode}; |
1227 | if (AssocCommutRoot && !AssocCommutPrev) |
1228 | return {InverseOpcode, AssocCommutOpcode}; |
1229 | if (!AssocCommutRoot && !AssocCommutPrev) |
1230 | return {InverseOpcode, InverseOpcode}; |
1231 | break; |
1232 | case MachineCombinerPattern::REASSOC_AX_YB: |
1233 | if (!AssocCommutRoot && AssocCommutPrev) |
1234 | return {InverseOpcode, InverseOpcode}; |
1235 | if (AssocCommutRoot && !AssocCommutPrev) |
1236 | return {AssocCommutOpcode, InverseOpcode}; |
1237 | if (!AssocCommutRoot && !AssocCommutPrev) |
1238 | return {InverseOpcode, AssocCommutOpcode}; |
1239 | break; |
1240 | case MachineCombinerPattern::REASSOC_XA_YB: |
1241 | if (!AssocCommutRoot && AssocCommutPrev) |
1242 | return {InverseOpcode, InverseOpcode}; |
1243 | if (AssocCommutRoot && !AssocCommutPrev) |
1244 | return {InverseOpcode, AssocCommutOpcode}; |
1245 | if (!AssocCommutRoot && !AssocCommutPrev) |
1246 | return {AssocCommutOpcode, InverseOpcode}; |
1247 | break; |
1248 | } |
1249 | llvm_unreachable("Unhandled combination" ); |
1250 | } |
1251 | |
1252 | // Return a pair of boolean flags showing if the new root and new prev operands |
1253 | // must be swapped. See visual example of the rule in |
1254 | // TargetInstrInfo::getReassociationOpcodes. |
1255 | static std::pair<bool, bool> mustSwapOperands(unsigned Pattern) { |
1256 | switch (Pattern) { |
1257 | default: |
1258 | llvm_unreachable("Unexpected pattern" ); |
1259 | case MachineCombinerPattern::REASSOC_AX_BY: |
1260 | return {false, false}; |
1261 | case MachineCombinerPattern::REASSOC_XA_BY: |
1262 | return {true, false}; |
1263 | case MachineCombinerPattern::REASSOC_AX_YB: |
1264 | return {true, true}; |
1265 | case MachineCombinerPattern::REASSOC_XA_YB: |
1266 | return {true, true}; |
1267 | } |
1268 | } |
1269 | |
1270 | void TargetInstrInfo::getReassociateOperandIndices( |
1271 | const MachineInstr &Root, unsigned Pattern, |
1272 | std::array<unsigned, 5> &OperandIndices) const { |
1273 | switch (Pattern) { |
1274 | case MachineCombinerPattern::REASSOC_AX_BY: |
1275 | OperandIndices = {1, 1, 1, 2, 2}; |
1276 | break; |
1277 | case MachineCombinerPattern::REASSOC_AX_YB: |
1278 | OperandIndices = {2, 1, 2, 2, 1}; |
1279 | break; |
1280 | case MachineCombinerPattern::REASSOC_XA_BY: |
1281 | OperandIndices = {1, 2, 1, 1, 2}; |
1282 | break; |
1283 | case MachineCombinerPattern::REASSOC_XA_YB: |
1284 | OperandIndices = {2, 2, 2, 1, 1}; |
1285 | break; |
1286 | default: |
1287 | llvm_unreachable("unexpected MachineCombinerPattern" ); |
1288 | } |
1289 | } |
1290 | |
1291 | /// Attempt the reassociation transformation to reduce critical path length. |
1292 | /// See the above comments before getMachineCombinerPatterns(). |
1293 | void TargetInstrInfo::reassociateOps( |
1294 | MachineInstr &Root, MachineInstr &Prev, unsigned Pattern, |
1295 | SmallVectorImpl<MachineInstr *> &InsInstrs, |
1296 | SmallVectorImpl<MachineInstr *> &DelInstrs, |
1297 | ArrayRef<unsigned> OperandIndices, |
1298 | DenseMap<Register, unsigned> &InstrIdxForVirtReg) const { |
1299 | MachineFunction *MF = Root.getMF(); |
1300 | MachineRegisterInfo &MRI = MF->getRegInfo(); |
1301 | const TargetInstrInfo *TII = MF->getSubtarget().getInstrInfo(); |
1302 | const TargetRegisterInfo *TRI = MF->getSubtarget().getRegisterInfo(); |
1303 | const TargetRegisterClass *RC = Root.getRegClassConstraint(OpIdx: 0, TII, TRI); |
1304 | |
1305 | MachineOperand &OpA = Prev.getOperand(i: OperandIndices[1]); |
1306 | MachineOperand &OpB = Root.getOperand(i: OperandIndices[2]); |
1307 | MachineOperand &OpX = Prev.getOperand(i: OperandIndices[3]); |
1308 | MachineOperand &OpY = Root.getOperand(i: OperandIndices[4]); |
1309 | MachineOperand &OpC = Root.getOperand(i: 0); |
1310 | |
1311 | Register RegA = OpA.getReg(); |
1312 | Register RegB = OpB.getReg(); |
1313 | Register RegX = OpX.getReg(); |
1314 | Register RegY = OpY.getReg(); |
1315 | Register RegC = OpC.getReg(); |
1316 | |
1317 | if (RegA.isVirtual()) |
1318 | MRI.constrainRegClass(Reg: RegA, RC); |
1319 | if (RegB.isVirtual()) |
1320 | MRI.constrainRegClass(Reg: RegB, RC); |
1321 | if (RegX.isVirtual()) |
1322 | MRI.constrainRegClass(Reg: RegX, RC); |
1323 | if (RegY.isVirtual()) |
1324 | MRI.constrainRegClass(Reg: RegY, RC); |
1325 | if (RegC.isVirtual()) |
1326 | MRI.constrainRegClass(Reg: RegC, RC); |
1327 | |
1328 | // Create a new virtual register for the result of (X op Y) instead of |
1329 | // recycling RegB because the MachineCombiner's computation of the critical |
1330 | // path requires a new register definition rather than an existing one. |
1331 | Register NewVR = MRI.createVirtualRegister(RegClass: RC); |
1332 | InstrIdxForVirtReg.insert(KV: std::make_pair(x&: NewVR, y: 0)); |
1333 | |
1334 | auto [NewRootOpc, NewPrevOpc] = getReassociationOpcodes(Pattern, Root, Prev); |
1335 | bool KillA = OpA.isKill(); |
1336 | bool KillX = OpX.isKill(); |
1337 | bool KillY = OpY.isKill(); |
1338 | bool KillNewVR = true; |
1339 | |
1340 | auto [SwapRootOperands, SwapPrevOperands] = mustSwapOperands(Pattern); |
1341 | |
1342 | if (SwapPrevOperands) { |
1343 | std::swap(a&: RegX, b&: RegY); |
1344 | std::swap(a&: KillX, b&: KillY); |
1345 | } |
1346 | |
1347 | unsigned PrevFirstOpIdx, PrevSecondOpIdx; |
1348 | unsigned RootFirstOpIdx, RootSecondOpIdx; |
1349 | switch (Pattern) { |
1350 | case MachineCombinerPattern::REASSOC_AX_BY: |
1351 | PrevFirstOpIdx = OperandIndices[1]; |
1352 | PrevSecondOpIdx = OperandIndices[3]; |
1353 | RootFirstOpIdx = OperandIndices[2]; |
1354 | RootSecondOpIdx = OperandIndices[4]; |
1355 | break; |
1356 | case MachineCombinerPattern::REASSOC_AX_YB: |
1357 | PrevFirstOpIdx = OperandIndices[1]; |
1358 | PrevSecondOpIdx = OperandIndices[3]; |
1359 | RootFirstOpIdx = OperandIndices[4]; |
1360 | RootSecondOpIdx = OperandIndices[2]; |
1361 | break; |
1362 | case MachineCombinerPattern::REASSOC_XA_BY: |
1363 | PrevFirstOpIdx = OperandIndices[3]; |
1364 | PrevSecondOpIdx = OperandIndices[1]; |
1365 | RootFirstOpIdx = OperandIndices[2]; |
1366 | RootSecondOpIdx = OperandIndices[4]; |
1367 | break; |
1368 | case MachineCombinerPattern::REASSOC_XA_YB: |
1369 | PrevFirstOpIdx = OperandIndices[3]; |
1370 | PrevSecondOpIdx = OperandIndices[1]; |
1371 | RootFirstOpIdx = OperandIndices[4]; |
1372 | RootSecondOpIdx = OperandIndices[2]; |
1373 | break; |
1374 | default: |
1375 | llvm_unreachable("unexpected MachineCombinerPattern" ); |
1376 | } |
1377 | |
1378 | // Basically BuildMI but doesn't add implicit operands by default. |
1379 | auto buildMINoImplicit = [](MachineFunction &MF, const MIMetadata &MIMD, |
1380 | const MCInstrDesc &MCID, Register DestReg) { |
1381 | return MachineInstrBuilder( |
1382 | MF, MF.CreateMachineInstr(MCID, DL: MIMD.getDL(), /*NoImpl=*/NoImplicit: true)) |
1383 | .setPCSections(MIMD.getPCSections()) |
1384 | .addReg(RegNo: DestReg, flags: RegState::Define); |
1385 | }; |
1386 | |
1387 | // Create new instructions for insertion. |
1388 | MachineInstrBuilder MIB1 = |
1389 | buildMINoImplicit(*MF, MIMetadata(Prev), TII->get(Opcode: NewPrevOpc), NewVR); |
1390 | for (const auto &MO : Prev.explicit_operands()) { |
1391 | unsigned Idx = MO.getOperandNo(); |
1392 | // Skip the result operand we'd already added. |
1393 | if (Idx == 0) |
1394 | continue; |
1395 | if (Idx == PrevFirstOpIdx) |
1396 | MIB1.addReg(RegNo: RegX, flags: getKillRegState(B: KillX)); |
1397 | else if (Idx == PrevSecondOpIdx) |
1398 | MIB1.addReg(RegNo: RegY, flags: getKillRegState(B: KillY)); |
1399 | else |
1400 | MIB1.add(MO); |
1401 | } |
1402 | MIB1.copyImplicitOps(OtherMI: Prev); |
1403 | |
1404 | if (SwapRootOperands) { |
1405 | std::swap(a&: RegA, b&: NewVR); |
1406 | std::swap(a&: KillA, b&: KillNewVR); |
1407 | } |
1408 | |
1409 | MachineInstrBuilder MIB2 = |
1410 | buildMINoImplicit(*MF, MIMetadata(Root), TII->get(Opcode: NewRootOpc), RegC); |
1411 | for (const auto &MO : Root.explicit_operands()) { |
1412 | unsigned Idx = MO.getOperandNo(); |
1413 | // Skip the result operand. |
1414 | if (Idx == 0) |
1415 | continue; |
1416 | if (Idx == RootFirstOpIdx) |
1417 | MIB2 = MIB2.addReg(RegNo: RegA, flags: getKillRegState(B: KillA)); |
1418 | else if (Idx == RootSecondOpIdx) |
1419 | MIB2 = MIB2.addReg(RegNo: NewVR, flags: getKillRegState(B: KillNewVR)); |
1420 | else |
1421 | MIB2 = MIB2.add(MO); |
1422 | } |
1423 | MIB2.copyImplicitOps(OtherMI: Root); |
1424 | |
1425 | // Propagate FP flags from the original instructions. |
1426 | // But clear poison-generating flags because those may not be valid now. |
1427 | // TODO: There should be a helper function for copying only fast-math-flags. |
1428 | uint32_t IntersectedFlags = Root.getFlags() & Prev.getFlags(); |
1429 | MIB1->setFlags(IntersectedFlags); |
1430 | MIB1->clearFlag(Flag: MachineInstr::MIFlag::NoSWrap); |
1431 | MIB1->clearFlag(Flag: MachineInstr::MIFlag::NoUWrap); |
1432 | MIB1->clearFlag(Flag: MachineInstr::MIFlag::IsExact); |
1433 | |
1434 | MIB2->setFlags(IntersectedFlags); |
1435 | MIB2->clearFlag(Flag: MachineInstr::MIFlag::NoSWrap); |
1436 | MIB2->clearFlag(Flag: MachineInstr::MIFlag::NoUWrap); |
1437 | MIB2->clearFlag(Flag: MachineInstr::MIFlag::IsExact); |
1438 | |
1439 | setSpecialOperandAttr(OldMI1&: Root, OldMI2&: Prev, NewMI1&: *MIB1, NewMI2&: *MIB2); |
1440 | |
1441 | // Record new instructions for insertion and old instructions for deletion. |
1442 | InsInstrs.push_back(Elt: MIB1); |
1443 | InsInstrs.push_back(Elt: MIB2); |
1444 | DelInstrs.push_back(Elt: &Prev); |
1445 | DelInstrs.push_back(Elt: &Root); |
1446 | |
1447 | // We transformed: |
1448 | // B = A op X (Prev) |
1449 | // C = B op Y (Root) |
1450 | // Into: |
1451 | // B = X op Y (MIB1) |
1452 | // C = A op B (MIB2) |
1453 | // C has the same value as before, B doesn't; as such, keep the debug number |
1454 | // of C but not of B. |
1455 | if (unsigned OldRootNum = Root.peekDebugInstrNum()) |
1456 | MIB2.getInstr()->setDebugInstrNum(OldRootNum); |
1457 | } |
1458 | |
1459 | void TargetInstrInfo::genAlternativeCodeSequence( |
1460 | MachineInstr &Root, unsigned Pattern, |
1461 | SmallVectorImpl<MachineInstr *> &InsInstrs, |
1462 | SmallVectorImpl<MachineInstr *> &DelInstrs, |
1463 | DenseMap<Register, unsigned> &InstIdxForVirtReg) const { |
1464 | MachineRegisterInfo &MRI = Root.getMF()->getRegInfo(); |
1465 | MachineBasicBlock &MBB = *Root.getParent(); |
1466 | MachineFunction &MF = *MBB.getParent(); |
1467 | const TargetInstrInfo *TII = MF.getSubtarget().getInstrInfo(); |
1468 | |
1469 | switch (Pattern) { |
1470 | case MachineCombinerPattern::REASSOC_AX_BY: |
1471 | case MachineCombinerPattern::REASSOC_AX_YB: |
1472 | case MachineCombinerPattern::REASSOC_XA_BY: |
1473 | case MachineCombinerPattern::REASSOC_XA_YB: { |
1474 | // Select the previous instruction in the sequence based on the input |
1475 | // pattern. |
1476 | std::array<unsigned, 5> OperandIndices; |
1477 | getReassociateOperandIndices(Root, Pattern, OperandIndices); |
1478 | MachineInstr *Prev = |
1479 | MRI.getUniqueVRegDef(Reg: Root.getOperand(i: OperandIndices[0]).getReg()); |
1480 | |
1481 | // Don't reassociate if Prev and Root are in different blocks. |
1482 | if (Prev->getParent() != Root.getParent()) |
1483 | return; |
1484 | |
1485 | reassociateOps(Root, Prev&: *Prev, Pattern, InsInstrs, DelInstrs, OperandIndices, |
1486 | InstrIdxForVirtReg&: InstIdxForVirtReg); |
1487 | break; |
1488 | } |
1489 | case MachineCombinerPattern::ACC_CHAIN: { |
1490 | SmallVector<Register, 32> ChainRegs; |
1491 | getAccumulatorChain(CurrentInstr: &Root, Chain&: ChainRegs); |
1492 | unsigned int Depth = ChainRegs.size(); |
1493 | assert(MaxAccumulatorWidth > 1 && |
1494 | "Max accumulator width set to illegal value" ); |
1495 | unsigned int MaxWidth = Log2_32(Value: Depth) < MaxAccumulatorWidth |
1496 | ? Log2_32(Value: Depth) |
1497 | : MaxAccumulatorWidth; |
1498 | |
1499 | // Walk down the chain and rewrite it as a tree. |
1500 | for (auto IndexedReg : llvm::enumerate(First: llvm::reverse(C&: ChainRegs))) { |
1501 | // No need to rewrite the first node, it is already perfect as it is. |
1502 | if (IndexedReg.index() == 0) |
1503 | continue; |
1504 | |
1505 | MachineInstr *Instr = MRI.getUniqueVRegDef(Reg: IndexedReg.value()); |
1506 | MachineInstrBuilder MIB; |
1507 | Register AccReg; |
1508 | if (IndexedReg.index() < MaxWidth) { |
1509 | // Now we need to create new instructions for the first row. |
1510 | AccReg = Instr->getOperand(i: 0).getReg(); |
1511 | unsigned OpCode = getAccumulationStartOpcode(Opcode: Root.getOpcode()); |
1512 | |
1513 | MIB = BuildMI(MF, MIMD: MIMetadata(*Instr), MCID: TII->get(Opcode: OpCode), DestReg: AccReg) |
1514 | .addReg(RegNo: Instr->getOperand(i: 2).getReg(), |
1515 | flags: getKillRegState(B: Instr->getOperand(i: 2).isKill())) |
1516 | .addReg(RegNo: Instr->getOperand(i: 3).getReg(), |
1517 | flags: getKillRegState(B: Instr->getOperand(i: 3).isKill())); |
1518 | } else { |
1519 | // For the remaining cases, we need to use an output register of one of |
1520 | // the newly inserted instuctions as operand 1 |
1521 | AccReg = Instr->getOperand(i: 0).getReg() == Root.getOperand(i: 0).getReg() |
1522 | ? MRI.createVirtualRegister( |
1523 | RegClass: MRI.getRegClass(Reg: Root.getOperand(i: 0).getReg())) |
1524 | : Instr->getOperand(i: 0).getReg(); |
1525 | assert(IndexedReg.index() >= MaxWidth); |
1526 | auto AccumulatorInput = |
1527 | ChainRegs[Depth - (IndexedReg.index() - MaxWidth) - 1]; |
1528 | MIB = BuildMI(MF, MIMD: MIMetadata(*Instr), MCID: TII->get(Opcode: Instr->getOpcode()), |
1529 | DestReg: AccReg) |
1530 | .addReg(RegNo: AccumulatorInput, flags: getKillRegState(B: true)) |
1531 | .addReg(RegNo: Instr->getOperand(i: 2).getReg(), |
1532 | flags: getKillRegState(B: Instr->getOperand(i: 2).isKill())) |
1533 | .addReg(RegNo: Instr->getOperand(i: 3).getReg(), |
1534 | flags: getKillRegState(B: Instr->getOperand(i: 3).isKill())); |
1535 | } |
1536 | |
1537 | MIB->setFlags(Instr->getFlags()); |
1538 | InstIdxForVirtReg.insert(KV: std::make_pair(x&: AccReg, y: InsInstrs.size())); |
1539 | InsInstrs.push_back(Elt: MIB); |
1540 | DelInstrs.push_back(Elt: Instr); |
1541 | } |
1542 | |
1543 | SmallVector<Register, 8> RegistersToReduce; |
1544 | for (unsigned i = (InsInstrs.size() - MaxWidth); i < InsInstrs.size(); |
1545 | ++i) { |
1546 | auto Reg = InsInstrs[i]->getOperand(i: 0).getReg(); |
1547 | RegistersToReduce.push_back(Elt: Reg); |
1548 | } |
1549 | |
1550 | while (RegistersToReduce.size() > 1) |
1551 | reduceAccumulatorTree(RegistersToReduce, InsInstrs, MF, Root, MRI, |
1552 | InstrIdxForVirtReg&: InstIdxForVirtReg, ResultReg: Root.getOperand(i: 0).getReg()); |
1553 | |
1554 | break; |
1555 | } |
1556 | } |
1557 | } |
1558 | |
1559 | MachineTraceStrategy TargetInstrInfo::getMachineCombinerTraceStrategy() const { |
1560 | return MachineTraceStrategy::TS_MinInstrCount; |
1561 | } |
1562 | |
1563 | bool TargetInstrInfo::isReallyTriviallyReMaterializable( |
1564 | const MachineInstr &MI) const { |
1565 | const MachineFunction &MF = *MI.getMF(); |
1566 | const MachineRegisterInfo &MRI = MF.getRegInfo(); |
1567 | |
1568 | // Remat clients assume operand 0 is the defined register. |
1569 | if (!MI.getNumOperands() || !MI.getOperand(i: 0).isReg()) |
1570 | return false; |
1571 | Register DefReg = MI.getOperand(i: 0).getReg(); |
1572 | |
1573 | // A sub-register definition can only be rematerialized if the instruction |
1574 | // doesn't read the other parts of the register. Otherwise it is really a |
1575 | // read-modify-write operation on the full virtual register which cannot be |
1576 | // moved safely. |
1577 | if (DefReg.isVirtual() && MI.getOperand(i: 0).getSubReg() && |
1578 | MI.readsVirtualRegister(Reg: DefReg)) |
1579 | return false; |
1580 | |
1581 | // A load from a fixed stack slot can be rematerialized. This may be |
1582 | // redundant with subsequent checks, but it's target-independent, |
1583 | // simple, and a common case. |
1584 | int FrameIdx = 0; |
1585 | if (isLoadFromStackSlot(MI, FrameIndex&: FrameIdx) && |
1586 | MF.getFrameInfo().isImmutableObjectIndex(ObjectIdx: FrameIdx)) |
1587 | return true; |
1588 | |
1589 | // Avoid instructions obviously unsafe for remat. |
1590 | if (MI.isNotDuplicable() || MI.mayStore() || MI.mayRaiseFPException() || |
1591 | MI.hasUnmodeledSideEffects()) |
1592 | return false; |
1593 | |
1594 | // Don't remat inline asm. We have no idea how expensive it is |
1595 | // even if it's side effect free. |
1596 | if (MI.isInlineAsm()) |
1597 | return false; |
1598 | |
1599 | // Avoid instructions which load from potentially varying memory. |
1600 | if (MI.mayLoad() && !MI.isDereferenceableInvariantLoad()) |
1601 | return false; |
1602 | |
1603 | // If any of the registers accessed are non-constant, conservatively assume |
1604 | // the instruction is not rematerializable. |
1605 | for (const MachineOperand &MO : MI.operands()) { |
1606 | if (!MO.isReg()) continue; |
1607 | Register Reg = MO.getReg(); |
1608 | if (Reg == 0) |
1609 | continue; |
1610 | |
1611 | // Check for a well-behaved physical register. |
1612 | if (Reg.isPhysical()) { |
1613 | if (MO.isUse()) { |
1614 | // If the physreg has no defs anywhere, it's just an ambient register |
1615 | // and we can freely move its uses. Alternatively, if it's allocatable, |
1616 | // it could get allocated to something with a def during allocation. |
1617 | if (!MRI.isConstantPhysReg(PhysReg: Reg)) |
1618 | return false; |
1619 | } else { |
1620 | // A physreg def. We can't remat it. |
1621 | return false; |
1622 | } |
1623 | continue; |
1624 | } |
1625 | |
1626 | // Only allow one virtual-register def. There may be multiple defs of the |
1627 | // same virtual register, though. |
1628 | if (MO.isDef() && Reg != DefReg) |
1629 | return false; |
1630 | |
1631 | // Don't allow any virtual-register uses. Rematting an instruction with |
1632 | // virtual register uses would length the live ranges of the uses, which |
1633 | // is not necessarily a good idea, certainly not "trivial". |
1634 | if (MO.isUse()) |
1635 | return false; |
1636 | } |
1637 | |
1638 | // Everything checked out. |
1639 | return true; |
1640 | } |
1641 | |
1642 | int TargetInstrInfo::getSPAdjust(const MachineInstr &MI) const { |
1643 | const MachineFunction *MF = MI.getMF(); |
1644 | const TargetFrameLowering *TFI = MF->getSubtarget().getFrameLowering(); |
1645 | bool StackGrowsDown = |
1646 | TFI->getStackGrowthDirection() == TargetFrameLowering::StackGrowsDown; |
1647 | |
1648 | unsigned FrameSetupOpcode = getCallFrameSetupOpcode(); |
1649 | unsigned FrameDestroyOpcode = getCallFrameDestroyOpcode(); |
1650 | |
1651 | if (!isFrameInstr(I: MI)) |
1652 | return 0; |
1653 | |
1654 | int SPAdj = TFI->alignSPAdjust(SPAdj: getFrameSize(I: MI)); |
1655 | |
1656 | if ((!StackGrowsDown && MI.getOpcode() == FrameSetupOpcode) || |
1657 | (StackGrowsDown && MI.getOpcode() == FrameDestroyOpcode)) |
1658 | SPAdj = -SPAdj; |
1659 | |
1660 | return SPAdj; |
1661 | } |
1662 | |
1663 | /// isSchedulingBoundary - Test if the given instruction should be |
1664 | /// considered a scheduling boundary. This primarily includes labels |
1665 | /// and terminators. |
1666 | bool TargetInstrInfo::isSchedulingBoundary(const MachineInstr &MI, |
1667 | const MachineBasicBlock *MBB, |
1668 | const MachineFunction &MF) const { |
1669 | // Terminators and labels can't be scheduled around. |
1670 | if (MI.isTerminator() || MI.isPosition()) |
1671 | return true; |
1672 | |
1673 | // INLINEASM_BR can jump to another block |
1674 | if (MI.getOpcode() == TargetOpcode::INLINEASM_BR) |
1675 | return true; |
1676 | |
1677 | // Don't attempt to schedule around any instruction that defines |
1678 | // a stack-oriented pointer, as it's unlikely to be profitable. This |
1679 | // saves compile time, because it doesn't require every single |
1680 | // stack slot reference to depend on the instruction that does the |
1681 | // modification. |
1682 | const TargetLowering &TLI = *MF.getSubtarget().getTargetLowering(); |
1683 | const TargetRegisterInfo *TRI = MF.getSubtarget().getRegisterInfo(); |
1684 | return MI.modifiesRegister(Reg: TLI.getStackPointerRegisterToSaveRestore(), TRI); |
1685 | } |
1686 | |
1687 | // Provide a global flag for disabling the PreRA hazard recognizer that targets |
1688 | // may choose to honor. |
1689 | bool TargetInstrInfo::usePreRAHazardRecognizer() const { |
1690 | return !DisableHazardRecognizer; |
1691 | } |
1692 | |
1693 | // Default implementation of CreateTargetRAHazardRecognizer. |
1694 | ScheduleHazardRecognizer *TargetInstrInfo:: |
1695 | CreateTargetHazardRecognizer(const TargetSubtargetInfo *STI, |
1696 | const ScheduleDAG *DAG) const { |
1697 | // Dummy hazard recognizer allows all instructions to issue. |
1698 | return new ScheduleHazardRecognizer(); |
1699 | } |
1700 | |
1701 | // Default implementation of CreateTargetMIHazardRecognizer. |
1702 | ScheduleHazardRecognizer *TargetInstrInfo::CreateTargetMIHazardRecognizer( |
1703 | const InstrItineraryData *II, const ScheduleDAGMI *DAG) const { |
1704 | return new ScoreboardHazardRecognizer(II, DAG, "machine-scheduler" ); |
1705 | } |
1706 | |
1707 | // Default implementation of CreateTargetPostRAHazardRecognizer. |
1708 | ScheduleHazardRecognizer *TargetInstrInfo:: |
1709 | CreateTargetPostRAHazardRecognizer(const InstrItineraryData *II, |
1710 | const ScheduleDAG *DAG) const { |
1711 | return new ScoreboardHazardRecognizer(II, DAG, "post-RA-sched" ); |
1712 | } |
1713 | |
1714 | // Default implementation of getMemOperandWithOffset. |
1715 | bool TargetInstrInfo::getMemOperandWithOffset( |
1716 | const MachineInstr &MI, const MachineOperand *&BaseOp, int64_t &Offset, |
1717 | bool &OffsetIsScalable, const TargetRegisterInfo *TRI) const { |
1718 | SmallVector<const MachineOperand *, 4> BaseOps; |
1719 | LocationSize Width = LocationSize::precise(Value: 0); |
1720 | if (!getMemOperandsWithOffsetWidth(MI, BaseOps, Offset, OffsetIsScalable, |
1721 | Width, TRI) || |
1722 | BaseOps.size() != 1) |
1723 | return false; |
1724 | BaseOp = BaseOps.front(); |
1725 | return true; |
1726 | } |
1727 | |
1728 | //===----------------------------------------------------------------------===// |
1729 | // SelectionDAG latency interface. |
1730 | //===----------------------------------------------------------------------===// |
1731 | |
1732 | std::optional<unsigned> |
1733 | TargetInstrInfo::getOperandLatency(const InstrItineraryData *ItinData, |
1734 | SDNode *DefNode, unsigned DefIdx, |
1735 | SDNode *UseNode, unsigned UseIdx) const { |
1736 | if (!ItinData || ItinData->isEmpty()) |
1737 | return std::nullopt; |
1738 | |
1739 | if (!DefNode->isMachineOpcode()) |
1740 | return std::nullopt; |
1741 | |
1742 | unsigned DefClass = get(Opcode: DefNode->getMachineOpcode()).getSchedClass(); |
1743 | if (!UseNode->isMachineOpcode()) |
1744 | return ItinData->getOperandCycle(ItinClassIndx: DefClass, OperandIdx: DefIdx); |
1745 | unsigned UseClass = get(Opcode: UseNode->getMachineOpcode()).getSchedClass(); |
1746 | return ItinData->getOperandLatency(DefClass, DefIdx, UseClass, UseIdx); |
1747 | } |
1748 | |
1749 | unsigned TargetInstrInfo::getInstrLatency(const InstrItineraryData *ItinData, |
1750 | SDNode *N) const { |
1751 | if (!ItinData || ItinData->isEmpty()) |
1752 | return 1; |
1753 | |
1754 | if (!N->isMachineOpcode()) |
1755 | return 1; |
1756 | |
1757 | return ItinData->getStageLatency(ItinClassIndx: get(Opcode: N->getMachineOpcode()).getSchedClass()); |
1758 | } |
1759 | |
1760 | //===----------------------------------------------------------------------===// |
1761 | // MachineInstr latency interface. |
1762 | //===----------------------------------------------------------------------===// |
1763 | |
1764 | unsigned TargetInstrInfo::getNumMicroOps(const InstrItineraryData *ItinData, |
1765 | const MachineInstr &MI) const { |
1766 | if (!ItinData || ItinData->isEmpty()) |
1767 | return 1; |
1768 | |
1769 | unsigned Class = MI.getDesc().getSchedClass(); |
1770 | int UOps = ItinData->Itineraries[Class].NumMicroOps; |
1771 | if (UOps >= 0) |
1772 | return UOps; |
1773 | |
1774 | // The # of u-ops is dynamically determined. The specific target should |
1775 | // override this function to return the right number. |
1776 | return 1; |
1777 | } |
1778 | |
1779 | /// Return the default expected latency for a def based on it's opcode. |
1780 | unsigned TargetInstrInfo::defaultDefLatency(const MCSchedModel &SchedModel, |
1781 | const MachineInstr &DefMI) const { |
1782 | if (DefMI.isTransient()) |
1783 | return 0; |
1784 | if (DefMI.mayLoad()) |
1785 | return SchedModel.LoadLatency; |
1786 | if (isHighLatencyDef(opc: DefMI.getOpcode())) |
1787 | return SchedModel.HighLatency; |
1788 | return 1; |
1789 | } |
1790 | |
1791 | unsigned TargetInstrInfo::getPredicationCost(const MachineInstr &) const { |
1792 | return 0; |
1793 | } |
1794 | |
1795 | unsigned TargetInstrInfo::getInstrLatency(const InstrItineraryData *ItinData, |
1796 | const MachineInstr &MI, |
1797 | unsigned *PredCost) const { |
1798 | // Default to one cycle for no itinerary. However, an "empty" itinerary may |
1799 | // still have a MinLatency property, which getStageLatency checks. |
1800 | if (!ItinData) |
1801 | return MI.mayLoad() ? 2 : 1; |
1802 | |
1803 | return ItinData->getStageLatency(ItinClassIndx: MI.getDesc().getSchedClass()); |
1804 | } |
1805 | |
1806 | bool TargetInstrInfo::hasLowDefLatency(const TargetSchedModel &SchedModel, |
1807 | const MachineInstr &DefMI, |
1808 | unsigned DefIdx) const { |
1809 | const InstrItineraryData *ItinData = SchedModel.getInstrItineraries(); |
1810 | if (!ItinData || ItinData->isEmpty()) |
1811 | return false; |
1812 | |
1813 | unsigned DefClass = DefMI.getDesc().getSchedClass(); |
1814 | std::optional<unsigned> DefCycle = |
1815 | ItinData->getOperandCycle(ItinClassIndx: DefClass, OperandIdx: DefIdx); |
1816 | return DefCycle && DefCycle <= 1U; |
1817 | } |
1818 | |
1819 | bool TargetInstrInfo::isFunctionSafeToSplit(const MachineFunction &MF) const { |
1820 | // TODO: We don't split functions where a section attribute has been set |
1821 | // since the split part may not be placed in a contiguous region. It may also |
1822 | // be more beneficial to augment the linker to ensure contiguous layout of |
1823 | // split functions within the same section as specified by the attribute. |
1824 | if (MF.getFunction().hasSection()) |
1825 | return false; |
1826 | |
1827 | // We don't want to proceed further for cold functions |
1828 | // or functions of unknown hotness. Lukewarm functions have no prefix. |
1829 | std::optional<StringRef> SectionPrefix = MF.getFunction().getSectionPrefix(); |
1830 | if (SectionPrefix && |
1831 | (*SectionPrefix == "unlikely" || *SectionPrefix == "unknown" )) { |
1832 | return false; |
1833 | } |
1834 | |
1835 | return true; |
1836 | } |
1837 | |
1838 | std::optional<ParamLoadedValue> |
1839 | TargetInstrInfo::describeLoadedValue(const MachineInstr &MI, |
1840 | Register Reg) const { |
1841 | const MachineFunction *MF = MI.getMF(); |
1842 | const TargetRegisterInfo *TRI = MF->getSubtarget().getRegisterInfo(); |
1843 | DIExpression *Expr = DIExpression::get(Context&: MF->getFunction().getContext(), Elements: {}); |
1844 | int64_t Offset; |
1845 | bool OffsetIsScalable; |
1846 | |
1847 | // To simplify the sub-register handling, verify that we only need to |
1848 | // consider physical registers. |
1849 | assert(MF->getProperties().hasNoVRegs()); |
1850 | |
1851 | if (auto DestSrc = isCopyInstr(MI)) { |
1852 | Register DestReg = DestSrc->Destination->getReg(); |
1853 | |
1854 | // If the copy destination is the forwarding reg, describe the forwarding |
1855 | // reg using the copy source as the backup location. Example: |
1856 | // |
1857 | // x0 = MOV x7 |
1858 | // call callee(x0) ; x0 described as x7 |
1859 | if (Reg == DestReg) |
1860 | return ParamLoadedValue(*DestSrc->Source, Expr); |
1861 | |
1862 | // If the target's hook couldn't describe this copy, give up. |
1863 | return std::nullopt; |
1864 | } else if (auto RegImm = isAddImmediate(MI, Reg)) { |
1865 | Register SrcReg = RegImm->Reg; |
1866 | Offset = RegImm->Imm; |
1867 | Expr = DIExpression::prepend(Expr, Flags: DIExpression::ApplyOffset, Offset); |
1868 | return ParamLoadedValue(MachineOperand::CreateReg(Reg: SrcReg, isDef: false), Expr); |
1869 | } else if (MI.hasOneMemOperand()) { |
1870 | // Only describe memory which provably does not escape the function. As |
1871 | // described in llvm.org/PR43343, escaped memory may be clobbered by the |
1872 | // callee (or by another thread). |
1873 | const auto &TII = MF->getSubtarget().getInstrInfo(); |
1874 | const MachineFrameInfo &MFI = MF->getFrameInfo(); |
1875 | const MachineMemOperand *MMO = MI.memoperands()[0]; |
1876 | const PseudoSourceValue *PSV = MMO->getPseudoValue(); |
1877 | |
1878 | // If the address points to "special" memory (e.g. a spill slot), it's |
1879 | // sufficient to check that it isn't aliased by any high-level IR value. |
1880 | if (!PSV || PSV->mayAlias(&MFI)) |
1881 | return std::nullopt; |
1882 | |
1883 | const MachineOperand *BaseOp; |
1884 | if (!TII->getMemOperandWithOffset(MI, BaseOp, Offset, OffsetIsScalable, |
1885 | TRI)) |
1886 | return std::nullopt; |
1887 | |
1888 | // FIXME: Scalable offsets are not yet handled in the offset code below. |
1889 | if (OffsetIsScalable) |
1890 | return std::nullopt; |
1891 | |
1892 | // TODO: Can currently only handle mem instructions with a single define. |
1893 | // An example from the x86 target: |
1894 | // ... |
1895 | // DIV64m $rsp, 1, $noreg, 24, $noreg, implicit-def dead $rax, implicit-def $rdx |
1896 | // ... |
1897 | // |
1898 | if (MI.getNumExplicitDefs() != 1) |
1899 | return std::nullopt; |
1900 | |
1901 | // TODO: In what way do we need to take Reg into consideration here? |
1902 | |
1903 | SmallVector<uint64_t, 8> Ops; |
1904 | DIExpression::appendOffset(Ops, Offset); |
1905 | Ops.push_back(Elt: dwarf::DW_OP_deref_size); |
1906 | Ops.push_back(Elt: MMO->getSize().hasValue() ? MMO->getSize().getValue() |
1907 | : ~UINT64_C(0)); |
1908 | Expr = DIExpression::prependOpcodes(Expr, Ops); |
1909 | return ParamLoadedValue(*BaseOp, Expr); |
1910 | } |
1911 | |
1912 | return std::nullopt; |
1913 | } |
1914 | |
1915 | // Get the call frame size just before MI. |
1916 | unsigned TargetInstrInfo::getCallFrameSizeAt(MachineInstr &MI) const { |
1917 | // Search backwards from MI for the most recent call frame instruction. |
1918 | MachineBasicBlock *MBB = MI.getParent(); |
1919 | for (auto &AdjI : reverse(C: make_range(x: MBB->instr_begin(), y: MI.getIterator()))) { |
1920 | if (AdjI.getOpcode() == getCallFrameSetupOpcode()) |
1921 | return getFrameTotalSize(I: AdjI); |
1922 | if (AdjI.getOpcode() == getCallFrameDestroyOpcode()) |
1923 | return 0; |
1924 | } |
1925 | |
1926 | // If none was found, use the call frame size from the start of the basic |
1927 | // block. |
1928 | return MBB->getCallFrameSize(); |
1929 | } |
1930 | |
1931 | /// Both DefMI and UseMI must be valid. By default, call directly to the |
1932 | /// itinerary. This may be overriden by the target. |
1933 | std::optional<unsigned> TargetInstrInfo::getOperandLatency( |
1934 | const InstrItineraryData *ItinData, const MachineInstr &DefMI, |
1935 | unsigned DefIdx, const MachineInstr &UseMI, unsigned UseIdx) const { |
1936 | unsigned DefClass = DefMI.getDesc().getSchedClass(); |
1937 | unsigned UseClass = UseMI.getDesc().getSchedClass(); |
1938 | return ItinData->getOperandLatency(DefClass, DefIdx, UseClass, UseIdx); |
1939 | } |
1940 | |
1941 | bool TargetInstrInfo::getRegSequenceInputs( |
1942 | const MachineInstr &MI, unsigned DefIdx, |
1943 | SmallVectorImpl<RegSubRegPairAndIdx> &InputRegs) const { |
1944 | assert((MI.isRegSequence() || |
1945 | MI.isRegSequenceLike()) && "Instruction do not have the proper type" ); |
1946 | |
1947 | if (!MI.isRegSequence()) |
1948 | return getRegSequenceLikeInputs(MI, DefIdx, InputRegs); |
1949 | |
1950 | // We are looking at: |
1951 | // Def = REG_SEQUENCE v0, sub0, v1, sub1, ... |
1952 | assert(DefIdx == 0 && "REG_SEQUENCE only has one def" ); |
1953 | for (unsigned OpIdx = 1, EndOpIdx = MI.getNumOperands(); OpIdx != EndOpIdx; |
1954 | OpIdx += 2) { |
1955 | const MachineOperand &MOReg = MI.getOperand(i: OpIdx); |
1956 | if (MOReg.isUndef()) |
1957 | continue; |
1958 | const MachineOperand &MOSubIdx = MI.getOperand(i: OpIdx + 1); |
1959 | assert(MOSubIdx.isImm() && |
1960 | "One of the subindex of the reg_sequence is not an immediate" ); |
1961 | // Record Reg:SubReg, SubIdx. |
1962 | InputRegs.push_back(Elt: RegSubRegPairAndIdx(MOReg.getReg(), MOReg.getSubReg(), |
1963 | (unsigned)MOSubIdx.getImm())); |
1964 | } |
1965 | return true; |
1966 | } |
1967 | |
1968 | bool TargetInstrInfo::getExtractSubregInputs( |
1969 | const MachineInstr &MI, unsigned DefIdx, |
1970 | RegSubRegPairAndIdx &InputReg) const { |
1971 | assert((MI.isExtractSubreg() || |
1972 | MI.isExtractSubregLike()) && "Instruction do not have the proper type" ); |
1973 | |
1974 | if (!MI.isExtractSubreg()) |
1975 | return getExtractSubregLikeInputs(MI, DefIdx, InputReg); |
1976 | |
1977 | // We are looking at: |
1978 | // Def = EXTRACT_SUBREG v0.sub1, sub0. |
1979 | assert(DefIdx == 0 && "EXTRACT_SUBREG only has one def" ); |
1980 | const MachineOperand &MOReg = MI.getOperand(i: 1); |
1981 | if (MOReg.isUndef()) |
1982 | return false; |
1983 | const MachineOperand &MOSubIdx = MI.getOperand(i: 2); |
1984 | assert(MOSubIdx.isImm() && |
1985 | "The subindex of the extract_subreg is not an immediate" ); |
1986 | |
1987 | InputReg.Reg = MOReg.getReg(); |
1988 | InputReg.SubReg = MOReg.getSubReg(); |
1989 | InputReg.SubIdx = (unsigned)MOSubIdx.getImm(); |
1990 | return true; |
1991 | } |
1992 | |
1993 | bool TargetInstrInfo::getInsertSubregInputs( |
1994 | const MachineInstr &MI, unsigned DefIdx, |
1995 | RegSubRegPair &BaseReg, RegSubRegPairAndIdx &InsertedReg) const { |
1996 | assert((MI.isInsertSubreg() || |
1997 | MI.isInsertSubregLike()) && "Instruction do not have the proper type" ); |
1998 | |
1999 | if (!MI.isInsertSubreg()) |
2000 | return getInsertSubregLikeInputs(MI, DefIdx, BaseReg, InsertedReg); |
2001 | |
2002 | // We are looking at: |
2003 | // Def = INSERT_SEQUENCE v0, v1, sub0. |
2004 | assert(DefIdx == 0 && "INSERT_SUBREG only has one def" ); |
2005 | const MachineOperand &MOBaseReg = MI.getOperand(i: 1); |
2006 | const MachineOperand &MOInsertedReg = MI.getOperand(i: 2); |
2007 | if (MOInsertedReg.isUndef()) |
2008 | return false; |
2009 | const MachineOperand &MOSubIdx = MI.getOperand(i: 3); |
2010 | assert(MOSubIdx.isImm() && |
2011 | "One of the subindex of the reg_sequence is not an immediate" ); |
2012 | BaseReg.Reg = MOBaseReg.getReg(); |
2013 | BaseReg.SubReg = MOBaseReg.getSubReg(); |
2014 | |
2015 | InsertedReg.Reg = MOInsertedReg.getReg(); |
2016 | InsertedReg.SubReg = MOInsertedReg.getSubReg(); |
2017 | InsertedReg.SubIdx = (unsigned)MOSubIdx.getImm(); |
2018 | return true; |
2019 | } |
2020 | |
2021 | // Returns a MIRPrinter comment for this machine operand. |
2022 | std::string TargetInstrInfo::createMIROperandComment( |
2023 | const MachineInstr &MI, const MachineOperand &Op, unsigned OpIdx, |
2024 | const TargetRegisterInfo *TRI) const { |
2025 | |
2026 | if (!MI.isInlineAsm()) |
2027 | return "" ; |
2028 | |
2029 | std::string Flags; |
2030 | raw_string_ostream OS(Flags); |
2031 | |
2032 | if (OpIdx == InlineAsm::MIOp_ExtraInfo) { |
2033 | // Print HasSideEffects, MayLoad, MayStore, IsAlignStack |
2034 | unsigned = Op.getImm(); |
2035 | bool First = true; |
2036 | for (StringRef Info : InlineAsm::getExtraInfoNames(ExtraInfo)) { |
2037 | if (!First) |
2038 | OS << " " ; |
2039 | First = false; |
2040 | OS << Info; |
2041 | } |
2042 | |
2043 | return Flags; |
2044 | } |
2045 | |
2046 | int FlagIdx = MI.findInlineAsmFlagIdx(OpIdx); |
2047 | if (FlagIdx < 0 || (unsigned)FlagIdx != OpIdx) |
2048 | return "" ; |
2049 | |
2050 | assert(Op.isImm() && "Expected flag operand to be an immediate" ); |
2051 | // Pretty print the inline asm operand descriptor. |
2052 | unsigned Flag = Op.getImm(); |
2053 | const InlineAsm::Flag F(Flag); |
2054 | OS << F.getKindName(); |
2055 | |
2056 | unsigned RCID; |
2057 | if (!F.isImmKind() && !F.isMemKind() && F.hasRegClassConstraint(RC&: RCID)) { |
2058 | if (TRI) { |
2059 | OS << ':' << TRI->getRegClassName(Class: TRI->getRegClass(i: RCID)); |
2060 | } else |
2061 | OS << ":RC" << RCID; |
2062 | } |
2063 | |
2064 | if (F.isMemKind()) { |
2065 | InlineAsm::ConstraintCode MCID = F.getMemoryConstraintID(); |
2066 | OS << ":" << InlineAsm::getMemConstraintName(C: MCID); |
2067 | } |
2068 | |
2069 | unsigned TiedTo; |
2070 | if (F.isUseOperandTiedToDef(Idx&: TiedTo)) |
2071 | OS << " tiedto:$" << TiedTo; |
2072 | |
2073 | if ((F.isRegDefKind() || F.isRegDefEarlyClobberKind() || F.isRegUseKind()) && |
2074 | F.getRegMayBeFolded()) |
2075 | OS << " foldable" ; |
2076 | |
2077 | return Flags; |
2078 | } |
2079 | |
2080 | TargetInstrInfo::PipelinerLoopInfo::~PipelinerLoopInfo() = default; |
2081 | |
2082 | void TargetInstrInfo::mergeOutliningCandidateAttributes( |
2083 | Function &F, std::vector<outliner::Candidate> &Candidates) const { |
2084 | // Include target features from an arbitrary candidate for the outlined |
2085 | // function. This makes sure the outlined function knows what kinds of |
2086 | // instructions are going into it. This is fine, since all parent functions |
2087 | // must necessarily support the instructions that are in the outlined region. |
2088 | outliner::Candidate &FirstCand = Candidates.front(); |
2089 | const Function &ParentFn = FirstCand.getMF()->getFunction(); |
2090 | if (ParentFn.hasFnAttribute(Kind: "target-features" )) |
2091 | F.addFnAttr(Attr: ParentFn.getFnAttribute(Kind: "target-features" )); |
2092 | if (ParentFn.hasFnAttribute(Kind: "target-cpu" )) |
2093 | F.addFnAttr(Attr: ParentFn.getFnAttribute(Kind: "target-cpu" )); |
2094 | |
2095 | // Set nounwind, so we don't generate eh_frame. |
2096 | if (llvm::all_of(Range&: Candidates, P: [](const outliner::Candidate &C) { |
2097 | return C.getMF()->getFunction().hasFnAttribute(Kind: Attribute::NoUnwind); |
2098 | })) |
2099 | F.addFnAttr(Kind: Attribute::NoUnwind); |
2100 | } |
2101 | |
2102 | outliner::InstrType |
2103 | TargetInstrInfo::getOutliningType(const MachineModuleInfo &MMI, |
2104 | MachineBasicBlock::iterator &MIT, |
2105 | unsigned Flags) const { |
2106 | MachineInstr &MI = *MIT; |
2107 | |
2108 | // NOTE: MI.isMetaInstruction() will match CFI_INSTRUCTION, but some targets |
2109 | // have support for outlining those. Special-case that here. |
2110 | if (MI.isCFIInstruction()) |
2111 | // Just go right to the target implementation. |
2112 | return getOutliningTypeImpl(MMI, MIT, Flags); |
2113 | |
2114 | // Be conservative about inline assembly. |
2115 | if (MI.isInlineAsm()) |
2116 | return outliner::InstrType::Illegal; |
2117 | |
2118 | // Labels generally can't safely be outlined. |
2119 | if (MI.isLabel()) |
2120 | return outliner::InstrType::Illegal; |
2121 | |
2122 | // Don't let debug instructions impact analysis. |
2123 | if (MI.isDebugInstr()) |
2124 | return outliner::InstrType::Invisible; |
2125 | |
2126 | // Some other special cases. |
2127 | switch (MI.getOpcode()) { |
2128 | case TargetOpcode::IMPLICIT_DEF: |
2129 | case TargetOpcode::KILL: |
2130 | case TargetOpcode::LIFETIME_START: |
2131 | case TargetOpcode::LIFETIME_END: |
2132 | return outliner::InstrType::Invisible; |
2133 | default: |
2134 | break; |
2135 | } |
2136 | |
2137 | // Is this a terminator for a basic block? |
2138 | if (MI.isTerminator()) { |
2139 | // If this is a branch to another block, we can't outline it. |
2140 | if (!MI.getParent()->succ_empty()) |
2141 | return outliner::InstrType::Illegal; |
2142 | |
2143 | // Don't outline if the branch is not unconditional. |
2144 | if (isPredicated(MI)) |
2145 | return outliner::InstrType::Illegal; |
2146 | } |
2147 | |
2148 | // Make sure none of the operands of this instruction do anything that |
2149 | // might break if they're moved outside their current function. |
2150 | // This includes MachineBasicBlock references, BlockAddressses, |
2151 | // Constant pool indices and jump table indices. |
2152 | // |
2153 | // A quick note on MO_TargetIndex: |
2154 | // This doesn't seem to be used in any of the architectures that the |
2155 | // MachineOutliner supports, but it was still filtered out in all of them. |
2156 | // There was one exception (RISC-V), but MO_TargetIndex also isn't used there. |
2157 | // As such, this check is removed both here and in the target-specific |
2158 | // implementations. Instead, we assert to make sure this doesn't |
2159 | // catch anyone off-guard somewhere down the line. |
2160 | for (const MachineOperand &MOP : MI.operands()) { |
2161 | // If you hit this assertion, please remove it and adjust |
2162 | // `getOutliningTypeImpl` for your target appropriately if necessary. |
2163 | // Adding the assertion back to other supported architectures |
2164 | // would be nice too :) |
2165 | assert(!MOP.isTargetIndex() && "This isn't used quite yet!" ); |
2166 | |
2167 | // CFI instructions should already have been filtered out at this point. |
2168 | assert(!MOP.isCFIIndex() && "CFI instructions handled elsewhere!" ); |
2169 | |
2170 | // PrologEpilogInserter should've already run at this point. |
2171 | assert(!MOP.isFI() && "FrameIndex instructions should be gone by now!" ); |
2172 | |
2173 | if (MOP.isMBB() || MOP.isBlockAddress() || MOP.isCPI() || MOP.isJTI()) |
2174 | return outliner::InstrType::Illegal; |
2175 | } |
2176 | |
2177 | // If we don't know, delegate to the target-specific hook. |
2178 | return getOutliningTypeImpl(MMI, MIT, Flags); |
2179 | } |
2180 | |
2181 | bool TargetInstrInfo::isMBBSafeToOutlineFrom(MachineBasicBlock &MBB, |
2182 | unsigned &Flags) const { |
2183 | // Some instrumentations create special TargetOpcode at the start which |
2184 | // expands to special code sequences which must be present. |
2185 | auto First = MBB.getFirstNonDebugInstr(); |
2186 | if (First == MBB.end()) |
2187 | return true; |
2188 | |
2189 | if (First->getOpcode() == TargetOpcode::FENTRY_CALL || |
2190 | First->getOpcode() == TargetOpcode::PATCHABLE_FUNCTION_ENTER) |
2191 | return false; |
2192 | |
2193 | // Some instrumentations create special pseudo-instructions at or just before |
2194 | // the end that must be present. |
2195 | auto Last = MBB.getLastNonDebugInstr(); |
2196 | if (Last->getOpcode() == TargetOpcode::PATCHABLE_RET || |
2197 | Last->getOpcode() == TargetOpcode::PATCHABLE_TAIL_CALL) |
2198 | return false; |
2199 | |
2200 | if (Last != First && Last->isReturn()) { |
2201 | --Last; |
2202 | if (Last->getOpcode() == TargetOpcode::PATCHABLE_FUNCTION_EXIT || |
2203 | Last->getOpcode() == TargetOpcode::PATCHABLE_TAIL_CALL) |
2204 | return false; |
2205 | } |
2206 | return true; |
2207 | } |
2208 | |
2209 | bool TargetInstrInfo::isGlobalMemoryObject(const MachineInstr *MI) const { |
2210 | return MI->isCall() || MI->hasUnmodeledSideEffects() || |
2211 | (MI->hasOrderedMemoryRef() && !MI->isDereferenceableInvariantLoad()); |
2212 | } |
2213 | |