/src/llvm-project/llvm/lib/CodeGen/MachineCopyPropagation.cpp
Line | Count | Source (jump to first uncovered line) |
1 | | //===- MachineCopyPropagation.cpp - Machine Copy Propagation Pass ---------===// |
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 is an extremely simple MachineInstr-level copy propagation pass. |
10 | | // |
11 | | // This pass forwards the source of COPYs to the users of their destinations |
12 | | // when doing so is legal. For example: |
13 | | // |
14 | | // %reg1 = COPY %reg0 |
15 | | // ... |
16 | | // ... = OP %reg1 |
17 | | // |
18 | | // If |
19 | | // - %reg0 has not been clobbered by the time of the use of %reg1 |
20 | | // - the register class constraints are satisfied |
21 | | // - the COPY def is the only value that reaches OP |
22 | | // then this pass replaces the above with: |
23 | | // |
24 | | // %reg1 = COPY %reg0 |
25 | | // ... |
26 | | // ... = OP %reg0 |
27 | | // |
28 | | // This pass also removes some redundant COPYs. For example: |
29 | | // |
30 | | // %R1 = COPY %R0 |
31 | | // ... // No clobber of %R1 |
32 | | // %R0 = COPY %R1 <<< Removed |
33 | | // |
34 | | // or |
35 | | // |
36 | | // %R1 = COPY %R0 |
37 | | // ... // No clobber of %R0 |
38 | | // %R1 = COPY %R0 <<< Removed |
39 | | // |
40 | | // or |
41 | | // |
42 | | // $R0 = OP ... |
43 | | // ... // No read/clobber of $R0 and $R1 |
44 | | // $R1 = COPY $R0 // $R0 is killed |
45 | | // Replace $R0 with $R1 and remove the COPY |
46 | | // $R1 = OP ... |
47 | | // ... |
48 | | // |
49 | | //===----------------------------------------------------------------------===// |
50 | | |
51 | | #include "llvm/ADT/DenseMap.h" |
52 | | #include "llvm/ADT/STLExtras.h" |
53 | | #include "llvm/ADT/SetVector.h" |
54 | | #include "llvm/ADT/SmallSet.h" |
55 | | #include "llvm/ADT/SmallVector.h" |
56 | | #include "llvm/ADT/Statistic.h" |
57 | | #include "llvm/ADT/iterator_range.h" |
58 | | #include "llvm/CodeGen/MachineBasicBlock.h" |
59 | | #include "llvm/CodeGen/MachineFunction.h" |
60 | | #include "llvm/CodeGen/MachineFunctionPass.h" |
61 | | #include "llvm/CodeGen/MachineInstr.h" |
62 | | #include "llvm/CodeGen/MachineOperand.h" |
63 | | #include "llvm/CodeGen/MachineRegisterInfo.h" |
64 | | #include "llvm/CodeGen/TargetInstrInfo.h" |
65 | | #include "llvm/CodeGen/TargetRegisterInfo.h" |
66 | | #include "llvm/CodeGen/TargetSubtargetInfo.h" |
67 | | #include "llvm/InitializePasses.h" |
68 | | #include "llvm/MC/MCRegisterInfo.h" |
69 | | #include "llvm/Pass.h" |
70 | | #include "llvm/Support/Debug.h" |
71 | | #include "llvm/Support/DebugCounter.h" |
72 | | #include "llvm/Support/raw_ostream.h" |
73 | | #include <cassert> |
74 | | #include <iterator> |
75 | | |
76 | | using namespace llvm; |
77 | | |
78 | | #define DEBUG_TYPE "machine-cp" |
79 | | |
80 | | STATISTIC(NumDeletes, "Number of dead copies deleted"); |
81 | | STATISTIC(NumCopyForwards, "Number of copy uses forwarded"); |
82 | | STATISTIC(NumCopyBackwardPropagated, "Number of copy defs backward propagated"); |
83 | | STATISTIC(SpillageChainsLength, "Length of spillage chains"); |
84 | | STATISTIC(NumSpillageChains, "Number of spillage chains"); |
85 | | DEBUG_COUNTER(FwdCounter, "machine-cp-fwd", |
86 | | "Controls which register COPYs are forwarded"); |
87 | | |
88 | | static cl::opt<bool> MCPUseCopyInstr("mcp-use-is-copy-instr", cl::init(false), |
89 | | cl::Hidden); |
90 | | static cl::opt<cl::boolOrDefault> |
91 | | EnableSpillageCopyElimination("enable-spill-copy-elim", cl::Hidden); |
92 | | |
93 | | namespace { |
94 | | |
95 | | static std::optional<DestSourcePair> isCopyInstr(const MachineInstr &MI, |
96 | | const TargetInstrInfo &TII, |
97 | 22.7M | bool UseCopyInstr) { |
98 | 22.7M | if (UseCopyInstr) |
99 | 2.17M | return TII.isCopyInstr(MI); |
100 | | |
101 | 20.5M | if (MI.isCopy()) |
102 | 3.18M | return std::optional<DestSourcePair>( |
103 | 3.18M | DestSourcePair{MI.getOperand(0), MI.getOperand(1)}); |
104 | | |
105 | 17.4M | return std::nullopt; |
106 | 20.5M | } |
107 | | |
108 | | class CopyTracker { |
109 | | struct CopyInfo { |
110 | | MachineInstr *MI, *LastSeenUseInCopy; |
111 | | SmallVector<MCRegister, 4> DefRegs; |
112 | | bool Avail; |
113 | | }; |
114 | | |
115 | | DenseMap<MCRegister, CopyInfo> Copies; |
116 | | |
117 | | public: |
118 | | /// Mark all of the given registers and their subregisters as unavailable for |
119 | | /// copying. |
120 | | void markRegsUnavailable(ArrayRef<MCRegister> Regs, |
121 | 1.28M | const TargetRegisterInfo &TRI) { |
122 | 1.28M | for (MCRegister Reg : Regs) { |
123 | | // Source of copy is no longer available for propagation. |
124 | 1.05M | for (MCRegUnit Unit : TRI.regunits(Reg)) { |
125 | 1.05M | auto CI = Copies.find(Unit); |
126 | 1.05M | if (CI != Copies.end()) |
127 | 971k | CI->second.Avail = false; |
128 | 1.05M | } |
129 | 775k | } |
130 | 1.28M | } |
131 | | |
132 | | /// Remove register from copy maps. |
133 | | void invalidateRegister(MCRegister Reg, const TargetRegisterInfo &TRI, |
134 | 18.8M | const TargetInstrInfo &TII, bool UseCopyInstr) { |
135 | | // Since Reg might be a subreg of some registers, only invalidate Reg is not |
136 | | // enough. We have to find the COPY defines Reg or registers defined by Reg |
137 | | // and invalidate all of them. Similarly, we must invalidate all of the |
138 | | // the subregisters used in the source of the COPY. |
139 | 18.8M | SmallSet<MCRegUnit, 8> RegUnitsToInvalidate; |
140 | 18.8M | auto InvalidateCopy = [&](MachineInstr *MI) { |
141 | 135k | std::optional<DestSourcePair> CopyOperands = |
142 | 135k | isCopyInstr(*MI, TII, UseCopyInstr); |
143 | 135k | assert(CopyOperands && "Expect copy"); |
144 | | |
145 | 0 | auto Dest = TRI.regunits(CopyOperands->Destination->getReg().asMCReg()); |
146 | 135k | auto Src = TRI.regunits(CopyOperands->Source->getReg().asMCReg()); |
147 | 135k | RegUnitsToInvalidate.insert(Dest.begin(), Dest.end()); |
148 | 135k | RegUnitsToInvalidate.insert(Src.begin(), Src.end()); |
149 | 135k | }; |
150 | | |
151 | 20.7M | for (MCRegUnit Unit : TRI.regunits(Reg)) { |
152 | 20.7M | auto I = Copies.find(Unit); |
153 | 20.7M | if (I != Copies.end()) { |
154 | 135k | if (MachineInstr *MI = I->second.MI) |
155 | 106k | InvalidateCopy(MI); |
156 | 135k | if (MachineInstr *MI = I->second.LastSeenUseInCopy) |
157 | 28.8k | InvalidateCopy(MI); |
158 | 135k | } |
159 | 20.7M | } |
160 | 18.8M | for (MCRegUnit Unit : RegUnitsToInvalidate) |
161 | 277k | Copies.erase(Unit); |
162 | 18.8M | } |
163 | | |
164 | | /// Clobber a single register, removing it from the tracker's copy maps. |
165 | | void clobberRegister(MCRegister Reg, const TargetRegisterInfo &TRI, |
166 | 7.43M | const TargetInstrInfo &TII, bool UseCopyInstr) { |
167 | 8.32M | for (MCRegUnit Unit : TRI.regunits(Reg)) { |
168 | 8.32M | auto I = Copies.find(Unit); |
169 | 8.32M | if (I != Copies.end()) { |
170 | | // When we clobber the source of a copy, we need to clobber everything |
171 | | // it defined. |
172 | 756k | markRegsUnavailable(I->second.DefRegs, TRI); |
173 | | // When we clobber the destination of a copy, we need to clobber the |
174 | | // whole register it defined. |
175 | 756k | if (MachineInstr *MI = I->second.MI) { |
176 | 527k | std::optional<DestSourcePair> CopyOperands = |
177 | 527k | isCopyInstr(*MI, TII, UseCopyInstr); |
178 | | |
179 | 527k | MCRegister Def = CopyOperands->Destination->getReg().asMCReg(); |
180 | 527k | MCRegister Src = CopyOperands->Source->getReg().asMCReg(); |
181 | | |
182 | 527k | markRegsUnavailable(Def, TRI); |
183 | | |
184 | | // Since we clobber the destination of a copy, the semantic of Src's |
185 | | // "DefRegs" to contain Def is no longer effectual. We will also need |
186 | | // to remove the record from the copy maps that indicates Src defined |
187 | | // Def. Failing to do so might cause the target to miss some |
188 | | // opportunities to further eliminate redundant copy instructions. |
189 | | // Consider the following sequence during the |
190 | | // ForwardCopyPropagateBlock procedure: |
191 | | // L1: r0 = COPY r9 <- TrackMI |
192 | | // L2: r0 = COPY r8 <- TrackMI (Remove r9 defined r0 from tracker) |
193 | | // L3: use r0 <- Remove L2 from MaybeDeadCopies |
194 | | // L4: early-clobber r9 <- Clobber r9 (L2 is still valid in tracker) |
195 | | // L5: r0 = COPY r8 <- Remove NopCopy |
196 | 694k | for (MCRegUnit SrcUnit : TRI.regunits(Src)) { |
197 | 694k | auto SrcCopy = Copies.find(SrcUnit); |
198 | 694k | if (SrcCopy != Copies.end() && SrcCopy->second.LastSeenUseInCopy) { |
199 | | // If SrcCopy defines multiple values, we only need |
200 | | // to erase the record for Def in DefRegs. |
201 | 407k | for (auto itr = SrcCopy->second.DefRegs.begin(); |
202 | 427k | itr != SrcCopy->second.DefRegs.end(); itr++) { |
203 | 405k | if (*itr == Def) { |
204 | 384k | SrcCopy->second.DefRegs.erase(itr); |
205 | | // If DefReg becomes empty after removal, we can remove the |
206 | | // SrcCopy from the tracker's copy maps. We only remove those |
207 | | // entries solely record the Def is defined by Src. If an |
208 | | // entry also contains the definition record of other Def' |
209 | | // registers, it cannot be cleared. |
210 | 384k | if (SrcCopy->second.DefRegs.empty() && !SrcCopy->second.MI) { |
211 | 294k | Copies.erase(SrcCopy); |
212 | 294k | } |
213 | 384k | break; |
214 | 384k | } |
215 | 405k | } |
216 | 407k | } |
217 | 694k | } |
218 | 527k | } |
219 | | // Now we can erase the copy. |
220 | 756k | Copies.erase(I); |
221 | 756k | } |
222 | 8.32M | } |
223 | 7.43M | } |
224 | | |
225 | | /// Add this copy's registers into the tracker's copy maps. |
226 | | void trackCopy(MachineInstr *MI, const TargetRegisterInfo &TRI, |
227 | 752k | const TargetInstrInfo &TII, bool UseCopyInstr) { |
228 | 752k | std::optional<DestSourcePair> CopyOperands = |
229 | 752k | isCopyInstr(*MI, TII, UseCopyInstr); |
230 | 752k | assert(CopyOperands && "Tracking non-copy?"); |
231 | | |
232 | 0 | MCRegister Src = CopyOperands->Source->getReg().asMCReg(); |
233 | 752k | MCRegister Def = CopyOperands->Destination->getReg().asMCReg(); |
234 | | |
235 | | // Remember Def is defined by the copy. |
236 | 752k | for (MCRegUnit Unit : TRI.regunits(Def)) |
237 | 863k | Copies[Unit] = {MI, nullptr, {}, true}; |
238 | | |
239 | | // Remember source that's copied to Def. Once it's clobbered, then |
240 | | // it's no longer available for copy propagation. |
241 | 863k | for (MCRegUnit Unit : TRI.regunits(Src)) { |
242 | 863k | auto I = Copies.insert({Unit, {nullptr, nullptr, {}, false}}); |
243 | 863k | auto &Copy = I.first->second; |
244 | 863k | if (!is_contained(Copy.DefRegs, Def)) |
245 | 862k | Copy.DefRegs.push_back(Def); |
246 | 863k | Copy.LastSeenUseInCopy = MI; |
247 | 863k | } |
248 | 752k | } |
249 | | |
250 | 17.6M | bool hasAnyCopies() { |
251 | 17.6M | return !Copies.empty(); |
252 | 17.6M | } |
253 | | |
254 | | MachineInstr *findCopyForUnit(MCRegister RegUnit, |
255 | | const TargetRegisterInfo &TRI, |
256 | 17.0M | bool MustBeAvailable = false) { |
257 | 17.0M | auto CI = Copies.find(RegUnit); |
258 | 17.0M | if (CI == Copies.end()) |
259 | 15.4M | return nullptr; |
260 | 1.53M | if (MustBeAvailable && !CI->second.Avail) |
261 | 461k | return nullptr; |
262 | 1.07M | return CI->second.MI; |
263 | 1.53M | } |
264 | | |
265 | | MachineInstr *findCopyDefViaUnit(MCRegister RegUnit, |
266 | 208k | const TargetRegisterInfo &TRI) { |
267 | 208k | auto CI = Copies.find(RegUnit); |
268 | 208k | if (CI == Copies.end()) |
269 | 203k | return nullptr; |
270 | 4.60k | if (CI->second.DefRegs.size() != 1) |
271 | 122 | return nullptr; |
272 | 4.48k | MCRegUnit RU = *TRI.regunits(CI->second.DefRegs[0]).begin(); |
273 | 4.48k | return findCopyForUnit(RU, TRI, true); |
274 | 4.60k | } |
275 | | |
276 | | MachineInstr *findAvailBackwardCopy(MachineInstr &I, MCRegister Reg, |
277 | | const TargetRegisterInfo &TRI, |
278 | | const TargetInstrInfo &TII, |
279 | 206k | bool UseCopyInstr) { |
280 | 206k | MCRegUnit RU = *TRI.regunits(Reg).begin(); |
281 | 206k | MachineInstr *AvailCopy = findCopyDefViaUnit(RU, TRI); |
282 | | |
283 | 206k | if (!AvailCopy) |
284 | 201k | return nullptr; |
285 | | |
286 | 4.46k | std::optional<DestSourcePair> CopyOperands = |
287 | 4.46k | isCopyInstr(*AvailCopy, TII, UseCopyInstr); |
288 | 4.46k | Register AvailSrc = CopyOperands->Source->getReg(); |
289 | 4.46k | Register AvailDef = CopyOperands->Destination->getReg(); |
290 | 4.46k | if (!TRI.isSubRegisterEq(AvailSrc, Reg)) |
291 | 27 | return nullptr; |
292 | | |
293 | 4.43k | for (const MachineInstr &MI : |
294 | 4.43k | make_range(AvailCopy->getReverseIterator(), I.getReverseIterator())) |
295 | 10.8k | for (const MachineOperand &MO : MI.operands()) |
296 | 31.2k | if (MO.isRegMask()) |
297 | | // FIXME: Shall we simultaneously invalidate AvailSrc or AvailDef? |
298 | 571 | if (MO.clobbersPhysReg(AvailSrc) || MO.clobbersPhysReg(AvailDef)) |
299 | 532 | return nullptr; |
300 | | |
301 | 3.90k | return AvailCopy; |
302 | 4.43k | } |
303 | | |
304 | | MachineInstr *findAvailCopy(MachineInstr &DestCopy, MCRegister Reg, |
305 | | const TargetRegisterInfo &TRI, |
306 | 4.56M | const TargetInstrInfo &TII, bool UseCopyInstr) { |
307 | | // We check the first RegUnit here, since we'll only be interested in the |
308 | | // copy if it copies the entire register anyway. |
309 | 4.56M | MCRegUnit RU = *TRI.regunits(Reg).begin(); |
310 | 4.56M | MachineInstr *AvailCopy = |
311 | 4.56M | findCopyForUnit(RU, TRI, /*MustBeAvailable=*/true); |
312 | | |
313 | 4.56M | if (!AvailCopy) |
314 | 4.43M | return nullptr; |
315 | | |
316 | 134k | std::optional<DestSourcePair> CopyOperands = |
317 | 134k | isCopyInstr(*AvailCopy, TII, UseCopyInstr); |
318 | 134k | Register AvailSrc = CopyOperands->Source->getReg(); |
319 | 134k | Register AvailDef = CopyOperands->Destination->getReg(); |
320 | 134k | if (!TRI.isSubRegisterEq(AvailDef, Reg)) |
321 | 11.0k | return nullptr; |
322 | | |
323 | | // Check that the available copy isn't clobbered by any regmasks between |
324 | | // itself and the destination. |
325 | 123k | for (const MachineInstr &MI : |
326 | 123k | make_range(AvailCopy->getIterator(), DestCopy.getIterator())) |
327 | 668k | for (const MachineOperand &MO : MI.operands()) |
328 | 1.92M | if (MO.isRegMask()) |
329 | 61.8k | if (MO.clobbersPhysReg(AvailSrc) || MO.clobbersPhysReg(AvailDef)) |
330 | 61.0k | return nullptr; |
331 | | |
332 | 62.3k | return AvailCopy; |
333 | 123k | } |
334 | | |
335 | | // Find last COPY that defines Reg before Current MachineInstr. |
336 | | MachineInstr *findLastSeenDefInCopy(const MachineInstr &Current, |
337 | | MCRegister Reg, |
338 | | const TargetRegisterInfo &TRI, |
339 | | const TargetInstrInfo &TII, |
340 | 5.54M | bool UseCopyInstr) { |
341 | 5.54M | MCRegUnit RU = *TRI.regunits(Reg).begin(); |
342 | 5.54M | auto CI = Copies.find(RU); |
343 | 5.54M | if (CI == Copies.end() || !CI->second.Avail) |
344 | 5.48M | return nullptr; |
345 | | |
346 | 55.5k | MachineInstr *DefCopy = CI->second.MI; |
347 | 55.5k | std::optional<DestSourcePair> CopyOperands = |
348 | 55.5k | isCopyInstr(*DefCopy, TII, UseCopyInstr); |
349 | 55.5k | Register Def = CopyOperands->Destination->getReg(); |
350 | 55.5k | if (!TRI.isSubRegisterEq(Def, Reg)) |
351 | 166 | return nullptr; |
352 | | |
353 | 55.4k | for (const MachineInstr &MI : |
354 | 55.4k | make_range(static_cast<const MachineInstr *>(DefCopy)->getIterator(), |
355 | 55.4k | Current.getIterator())) |
356 | 791k | for (const MachineOperand &MO : MI.operands()) |
357 | 2.50M | if (MO.isRegMask()) |
358 | 10.3k | if (MO.clobbersPhysReg(Def)) { |
359 | 0 | LLVM_DEBUG(dbgs() << "MCP: Removed tracking of " |
360 | 0 | << printReg(Def, &TRI) << "\n"); |
361 | 0 | return nullptr; |
362 | 0 | } |
363 | | |
364 | 55.4k | return DefCopy; |
365 | 55.4k | } |
366 | | |
367 | | // Find last COPY that uses Reg. |
368 | | MachineInstr *findLastSeenUseInCopy(MCRegister Reg, |
369 | 5.49M | const TargetRegisterInfo &TRI) { |
370 | 5.49M | MCRegUnit RU = *TRI.regunits(Reg).begin(); |
371 | 5.49M | auto CI = Copies.find(RU); |
372 | 5.49M | if (CI == Copies.end()) |
373 | 5.36M | return nullptr; |
374 | 132k | return CI->second.LastSeenUseInCopy; |
375 | 5.49M | } |
376 | | |
377 | 888k | void clear() { |
378 | 888k | Copies.clear(); |
379 | 888k | } |
380 | | }; |
381 | | |
382 | | class MachineCopyPropagation : public MachineFunctionPass { |
383 | | const TargetRegisterInfo *TRI = nullptr; |
384 | | const TargetInstrInfo *TII = nullptr; |
385 | | const MachineRegisterInfo *MRI = nullptr; |
386 | | |
387 | | // Return true if this is a copy instruction and false otherwise. |
388 | | bool UseCopyInstr; |
389 | | |
390 | | public: |
391 | | static char ID; // Pass identification, replacement for typeid |
392 | | |
393 | | MachineCopyPropagation(bool CopyInstr = false) |
394 | 72.6k | : MachineFunctionPass(ID), UseCopyInstr(CopyInstr || MCPUseCopyInstr) { |
395 | 72.6k | initializeMachineCopyPropagationPass(*PassRegistry::getPassRegistry()); |
396 | 72.6k | } |
397 | | |
398 | 72.6k | void getAnalysisUsage(AnalysisUsage &AU) const override { |
399 | 72.6k | AU.setPreservesCFG(); |
400 | 72.6k | MachineFunctionPass::getAnalysisUsage(AU); |
401 | 72.6k | } |
402 | | |
403 | | bool runOnMachineFunction(MachineFunction &MF) override; |
404 | | |
405 | 72.6k | MachineFunctionProperties getRequiredProperties() const override { |
406 | 72.6k | return MachineFunctionProperties().set( |
407 | 72.6k | MachineFunctionProperties::Property::NoVRegs); |
408 | 72.6k | } |
409 | | |
410 | | private: |
411 | | typedef enum { DebugUse = false, RegularUse = true } DebugType; |
412 | | |
413 | | void ReadRegister(MCRegister Reg, MachineInstr &Reader, DebugType DT); |
414 | | void ForwardCopyPropagateBlock(MachineBasicBlock &MBB); |
415 | | void BackwardCopyPropagateBlock(MachineBasicBlock &MBB); |
416 | | void EliminateSpillageCopies(MachineBasicBlock &MBB); |
417 | | bool eraseIfRedundant(MachineInstr &Copy, MCRegister Src, MCRegister Def); |
418 | | void forwardUses(MachineInstr &MI); |
419 | | void propagateDefs(MachineInstr &MI); |
420 | | bool isForwardableRegClassCopy(const MachineInstr &Copy, |
421 | | const MachineInstr &UseI, unsigned UseIdx); |
422 | | bool isBackwardPropagatableRegClassCopy(const MachineInstr &Copy, |
423 | | const MachineInstr &UseI, |
424 | | unsigned UseIdx); |
425 | | bool hasImplicitOverlap(const MachineInstr &MI, const MachineOperand &Use); |
426 | | bool hasOverlappingMultipleDef(const MachineInstr &MI, |
427 | | const MachineOperand &MODef, Register Def); |
428 | | |
429 | | /// Candidates for deletion. |
430 | | SmallSetVector<MachineInstr *, 8> MaybeDeadCopies; |
431 | | |
432 | | /// Multimap tracking debug users in current BB |
433 | | DenseMap<MachineInstr *, SmallSet<MachineInstr *, 2>> CopyDbgUsers; |
434 | | |
435 | | CopyTracker Tracker; |
436 | | |
437 | | bool Changed = false; |
438 | | }; |
439 | | |
440 | | } // end anonymous namespace |
441 | | |
442 | | char MachineCopyPropagation::ID = 0; |
443 | | |
444 | | char &llvm::MachineCopyPropagationID = MachineCopyPropagation::ID; |
445 | | |
446 | | INITIALIZE_PASS(MachineCopyPropagation, DEBUG_TYPE, |
447 | | "Machine Copy Propagation Pass", false, false) |
448 | | |
449 | | void MachineCopyPropagation::ReadRegister(MCRegister Reg, MachineInstr &Reader, |
450 | 11.4M | DebugType DT) { |
451 | | // If 'Reg' is defined by a copy, the copy is no longer a candidate |
452 | | // for elimination. If a copy is "read" by a debug user, record the user |
453 | | // for propagation. |
454 | 12.4M | for (MCRegUnit Unit : TRI->regunits(Reg)) { |
455 | 12.4M | if (MachineInstr *Copy = Tracker.findCopyForUnit(Unit, *TRI)) { |
456 | 823k | if (DT == RegularUse) { |
457 | 823k | LLVM_DEBUG(dbgs() << "MCP: Copy is used - not dead: "; Copy->dump()); |
458 | 823k | MaybeDeadCopies.remove(Copy); |
459 | 823k | } else { |
460 | 361 | CopyDbgUsers[Copy].insert(&Reader); |
461 | 361 | } |
462 | 823k | } |
463 | 12.4M | } |
464 | 11.4M | } |
465 | | |
466 | | /// Return true if \p PreviousCopy did copy register \p Src to register \p Def. |
467 | | /// This fact may have been obscured by sub register usage or may not be true at |
468 | | /// all even though Src and Def are subregisters of the registers used in |
469 | | /// PreviousCopy. e.g. |
470 | | /// isNopCopy("ecx = COPY eax", AX, CX) == true |
471 | | /// isNopCopy("ecx = COPY eax", AH, CL) == false |
472 | | static bool isNopCopy(const MachineInstr &PreviousCopy, MCRegister Src, |
473 | | MCRegister Def, const TargetRegisterInfo *TRI, |
474 | 8.82k | const TargetInstrInfo *TII, bool UseCopyInstr) { |
475 | | |
476 | 8.82k | std::optional<DestSourcePair> CopyOperands = |
477 | 8.82k | isCopyInstr(PreviousCopy, *TII, UseCopyInstr); |
478 | 8.82k | MCRegister PreviousSrc = CopyOperands->Source->getReg().asMCReg(); |
479 | 8.82k | MCRegister PreviousDef = CopyOperands->Destination->getReg().asMCReg(); |
480 | 8.82k | if (Src == PreviousSrc && Def == PreviousDef) |
481 | 2.16k | return true; |
482 | 6.66k | if (!TRI->isSubRegister(PreviousSrc, Src)) |
483 | 6.66k | return false; |
484 | 2 | unsigned SubIdx = TRI->getSubRegIndex(PreviousSrc, Src); |
485 | 2 | return SubIdx == TRI->getSubRegIndex(PreviousDef, Def); |
486 | 6.66k | } |
487 | | |
488 | | /// Remove instruction \p Copy if there exists a previous copy that copies the |
489 | | /// register \p Src to the register \p Def; This may happen indirectly by |
490 | | /// copying the super registers. |
491 | | bool MachineCopyPropagation::eraseIfRedundant(MachineInstr &Copy, |
492 | 1.13M | MCRegister Src, MCRegister Def) { |
493 | | // Avoid eliminating a copy from/to a reserved registers as we cannot predict |
494 | | // the value (Example: The sparc zero register is writable but stays zero). |
495 | 1.13M | if (MRI->isReserved(Src) || MRI->isReserved(Def)) |
496 | 160k | return false; |
497 | | |
498 | | // Search for an existing copy. |
499 | 977k | MachineInstr *PrevCopy = |
500 | 977k | Tracker.findAvailCopy(Copy, Def, *TRI, *TII, UseCopyInstr); |
501 | 977k | if (!PrevCopy) |
502 | 968k | return false; |
503 | | |
504 | 8.82k | auto PrevCopyOperands = isCopyInstr(*PrevCopy, *TII, UseCopyInstr); |
505 | | // Check that the existing copy uses the correct sub registers. |
506 | 8.82k | if (PrevCopyOperands->Destination->isDead()) |
507 | 0 | return false; |
508 | 8.82k | if (!isNopCopy(*PrevCopy, Src, Def, TRI, TII, UseCopyInstr)) |
509 | 6.66k | return false; |
510 | | |
511 | 2.16k | LLVM_DEBUG(dbgs() << "MCP: copy is a NOP, removing: "; Copy.dump()); |
512 | | |
513 | | // Copy was redundantly redefining either Src or Def. Remove earlier kill |
514 | | // flags between Copy and PrevCopy because the value will be reused now. |
515 | 2.16k | std::optional<DestSourcePair> CopyOperands = |
516 | 2.16k | isCopyInstr(Copy, *TII, UseCopyInstr); |
517 | 2.16k | assert(CopyOperands); |
518 | | |
519 | 0 | Register CopyDef = CopyOperands->Destination->getReg(); |
520 | 2.16k | assert(CopyDef == Src || CopyDef == Def); |
521 | 0 | for (MachineInstr &MI : |
522 | 2.16k | make_range(PrevCopy->getIterator(), Copy.getIterator())) |
523 | 12.0k | MI.clearRegisterKills(CopyDef, TRI); |
524 | | |
525 | | // Clear undef flag from remaining copy if needed. |
526 | 2.16k | if (!CopyOperands->Source->isUndef()) { |
527 | 2.16k | PrevCopy->getOperand(PrevCopyOperands->Source->getOperandNo()) |
528 | 2.16k | .setIsUndef(false); |
529 | 2.16k | } |
530 | | |
531 | 2.16k | Copy.eraseFromParent(); |
532 | 2.16k | Changed = true; |
533 | 2.16k | ++NumDeletes; |
534 | 2.16k | return true; |
535 | 8.82k | } |
536 | | |
537 | | bool MachineCopyPropagation::isBackwardPropagatableRegClassCopy( |
538 | 3.68k | const MachineInstr &Copy, const MachineInstr &UseI, unsigned UseIdx) { |
539 | 3.68k | std::optional<DestSourcePair> CopyOperands = |
540 | 3.68k | isCopyInstr(Copy, *TII, UseCopyInstr); |
541 | 3.68k | Register Def = CopyOperands->Destination->getReg(); |
542 | | |
543 | 3.68k | if (const TargetRegisterClass *URC = |
544 | 3.68k | UseI.getRegClassConstraint(UseIdx, TII, TRI)) |
545 | 3.61k | return URC->contains(Def); |
546 | | |
547 | | // We don't process further if UseI is a COPY, since forward copy propagation |
548 | | // should handle that. |
549 | 72 | return false; |
550 | 3.68k | } |
551 | | |
552 | | /// Decide whether we should forward the source of \param Copy to its use in |
553 | | /// \param UseI based on the physical register class constraints of the opcode |
554 | | /// and avoiding introducing more cross-class COPYs. |
555 | | bool MachineCopyPropagation::isForwardableRegClassCopy(const MachineInstr &Copy, |
556 | | const MachineInstr &UseI, |
557 | 22.8k | unsigned UseIdx) { |
558 | 22.8k | std::optional<DestSourcePair> CopyOperands = |
559 | 22.8k | isCopyInstr(Copy, *TII, UseCopyInstr); |
560 | 22.8k | Register CopySrcReg = CopyOperands->Source->getReg(); |
561 | | |
562 | | // If the new register meets the opcode register constraints, then allow |
563 | | // forwarding. |
564 | 22.8k | if (const TargetRegisterClass *URC = |
565 | 22.8k | UseI.getRegClassConstraint(UseIdx, TII, TRI)) |
566 | 20.3k | return URC->contains(CopySrcReg); |
567 | | |
568 | 2.43k | auto UseICopyOperands = isCopyInstr(UseI, *TII, UseCopyInstr); |
569 | 2.43k | if (!UseICopyOperands) |
570 | 410 | return false; |
571 | | |
572 | | /// COPYs don't have register class constraints, so if the user instruction |
573 | | /// is a COPY, we just try to avoid introducing additional cross-class |
574 | | /// COPYs. For example: |
575 | | /// |
576 | | /// RegClassA = COPY RegClassB // Copy parameter |
577 | | /// ... |
578 | | /// RegClassB = COPY RegClassA // UseI parameter |
579 | | /// |
580 | | /// which after forwarding becomes |
581 | | /// |
582 | | /// RegClassA = COPY RegClassB |
583 | | /// ... |
584 | | /// RegClassB = COPY RegClassB |
585 | | /// |
586 | | /// so we have reduced the number of cross-class COPYs and potentially |
587 | | /// introduced a nop COPY that can be removed. |
588 | | |
589 | | // Allow forwarding if src and dst belong to any common class, so long as they |
590 | | // don't belong to any (possibly smaller) common class that requires copies to |
591 | | // go via a different class. |
592 | 2.02k | Register UseDstReg = UseICopyOperands->Destination->getReg(); |
593 | 2.02k | bool Found = false; |
594 | 2.02k | bool IsCrossClass = false; |
595 | 214k | for (const TargetRegisterClass *RC : TRI->regclasses()) { |
596 | 214k | if (RC->contains(CopySrcReg) && RC->contains(UseDstReg)) { |
597 | 14.1k | Found = true; |
598 | 14.1k | if (TRI->getCrossCopyRegClass(RC) != RC) { |
599 | 0 | IsCrossClass = true; |
600 | 0 | break; |
601 | 0 | } |
602 | 14.1k | } |
603 | 214k | } |
604 | 2.02k | if (!Found) |
605 | 169 | return false; |
606 | 1.85k | if (!IsCrossClass) |
607 | 1.85k | return true; |
608 | | // The forwarded copy would be cross-class. Only do this if the original copy |
609 | | // was also cross-class. |
610 | 0 | Register CopyDstReg = CopyOperands->Destination->getReg(); |
611 | 0 | for (const TargetRegisterClass *RC : TRI->regclasses()) { |
612 | 0 | if (RC->contains(CopySrcReg) && RC->contains(CopyDstReg) && |
613 | 0 | TRI->getCrossCopyRegClass(RC) != RC) |
614 | 0 | return true; |
615 | 0 | } |
616 | 0 | return false; |
617 | 0 | } |
618 | | |
619 | | /// Check that \p MI does not have implicit uses that overlap with it's \p Use |
620 | | /// operand (the register being replaced), since these can sometimes be |
621 | | /// implicitly tied to other operands. For example, on AMDGPU: |
622 | | /// |
623 | | /// V_MOVRELS_B32_e32 %VGPR2, %M0<imp-use>, %EXEC<imp-use>, %VGPR2_VGPR3_VGPR4_VGPR5<imp-use> |
624 | | /// |
625 | | /// the %VGPR2 is implicitly tied to the larger reg operand, but we have no |
626 | | /// way of knowing we need to update the latter when updating the former. |
627 | | bool MachineCopyPropagation::hasImplicitOverlap(const MachineInstr &MI, |
628 | 17.3k | const MachineOperand &Use) { |
629 | 17.3k | for (const MachineOperand &MIUse : MI.uses()) |
630 | 47.7k | if (&MIUse != &Use && MIUse.isReg() && MIUse.isImplicit() && |
631 | 47.7k | MIUse.isUse() && TRI->regsOverlap(Use.getReg(), MIUse.getReg())) |
632 | 43 | return true; |
633 | | |
634 | 17.2k | return false; |
635 | 17.3k | } |
636 | | |
637 | | /// For an MI that has multiple definitions, check whether \p MI has |
638 | | /// a definition that overlaps with another of its definitions. |
639 | | /// For example, on ARM: umull r9, r9, lr, r0 |
640 | | /// The umull instruction is unpredictable unless RdHi and RdLo are different. |
641 | | bool MachineCopyPropagation::hasOverlappingMultipleDef( |
642 | 163 | const MachineInstr &MI, const MachineOperand &MODef, Register Def) { |
643 | 163 | for (const MachineOperand &MIDef : MI.defs()) { |
644 | 163 | if ((&MIDef != &MODef) && MIDef.isReg() && |
645 | 163 | TRI->regsOverlap(Def, MIDef.getReg())) |
646 | 0 | return true; |
647 | 163 | } |
648 | | |
649 | 163 | return false; |
650 | 163 | } |
651 | | |
652 | | /// Look for available copies whose destination register is used by \p MI and |
653 | | /// replace the use in \p MI with the copy's source register. |
654 | 8.89M | void MachineCopyPropagation::forwardUses(MachineInstr &MI) { |
655 | 8.89M | if (!Tracker.hasAnyCopies()) |
656 | 4.39M | return; |
657 | | |
658 | | // Look for non-tied explicit vreg uses that have an active COPY |
659 | | // instruction that defines the physical register allocated to them. |
660 | | // Replace the vreg with the source of the active COPY. |
661 | 19.9M | for (unsigned OpIdx = 0, OpEnd = MI.getNumOperands(); OpIdx < OpEnd; |
662 | 15.4M | ++OpIdx) { |
663 | 15.4M | MachineOperand &MOUse = MI.getOperand(OpIdx); |
664 | | // Don't forward into undef use operands since doing so can cause problems |
665 | | // with the machine verifier, since it doesn't treat undef reads as reads, |
666 | | // so we can end up with a live range that ends on an undef read, leading to |
667 | | // an error that the live range doesn't end on a read of the live range |
668 | | // register. |
669 | 15.4M | if (!MOUse.isReg() || MOUse.isTied() || MOUse.isUndef() || MOUse.isDef() || |
670 | 15.4M | MOUse.isImplicit()) |
671 | 10.1M | continue; |
672 | | |
673 | 5.24M | if (!MOUse.getReg()) |
674 | 453k | continue; |
675 | | |
676 | | // Check that the register is marked 'renamable' so we know it is safe to |
677 | | // rename it without violating any constraints that aren't expressed in the |
678 | | // IR (e.g. ABI or opcode requirements). |
679 | 4.78M | if (!MOUse.isRenamable()) |
680 | 1.19M | continue; |
681 | | |
682 | 3.58M | MachineInstr *Copy = Tracker.findAvailCopy(MI, MOUse.getReg().asMCReg(), |
683 | 3.58M | *TRI, *TII, UseCopyInstr); |
684 | 3.58M | if (!Copy) |
685 | 3.53M | continue; |
686 | | |
687 | 53.5k | std::optional<DestSourcePair> CopyOperands = |
688 | 53.5k | isCopyInstr(*Copy, *TII, UseCopyInstr); |
689 | 53.5k | Register CopyDstReg = CopyOperands->Destination->getReg(); |
690 | 53.5k | const MachineOperand &CopySrc = *CopyOperands->Source; |
691 | 53.5k | Register CopySrcReg = CopySrc.getReg(); |
692 | | |
693 | 53.5k | Register ForwardedReg = CopySrcReg; |
694 | | // MI might use a sub-register of the Copy destination, in which case the |
695 | | // forwarded register is the matching sub-register of the Copy source. |
696 | 53.5k | if (MOUse.getReg() != CopyDstReg) { |
697 | 2.03k | unsigned SubRegIdx = TRI->getSubRegIndex(CopyDstReg, MOUse.getReg()); |
698 | 2.03k | assert(SubRegIdx && |
699 | 2.03k | "MI source is not a sub-register of Copy destination"); |
700 | 0 | ForwardedReg = TRI->getSubReg(CopySrcReg, SubRegIdx); |
701 | 2.03k | if (!ForwardedReg) { |
702 | 110 | LLVM_DEBUG(dbgs() << "MCP: Copy source does not have sub-register " |
703 | 110 | << TRI->getSubRegIndexName(SubRegIdx) << '\n'); |
704 | 110 | continue; |
705 | 110 | } |
706 | 2.03k | } |
707 | | |
708 | | // Don't forward COPYs of reserved regs unless they are constant. |
709 | 53.4k | if (MRI->isReserved(CopySrcReg) && !MRI->isConstantPhysReg(CopySrcReg)) |
710 | 30.6k | continue; |
711 | | |
712 | 22.8k | if (!isForwardableRegClassCopy(*Copy, MI, OpIdx)) |
713 | 5.62k | continue; |
714 | | |
715 | 17.1k | if (hasImplicitOverlap(MI, MOUse)) |
716 | 43 | continue; |
717 | | |
718 | | // Check that the instruction is not a copy that partially overwrites the |
719 | | // original copy source that we are about to use. The tracker mechanism |
720 | | // cannot cope with that. |
721 | 17.1k | if (isCopyInstr(MI, *TII, UseCopyInstr) && |
722 | 17.1k | MI.modifiesRegister(CopySrcReg, TRI) && |
723 | 17.1k | !MI.definesRegister(CopySrcReg)) { |
724 | 0 | LLVM_DEBUG(dbgs() << "MCP: Copy source overlap with dest in " << MI); |
725 | 0 | continue; |
726 | 0 | } |
727 | | |
728 | 17.1k | if (!DebugCounter::shouldExecute(FwdCounter)) { |
729 | 0 | LLVM_DEBUG(dbgs() << "MCP: Skipping forwarding due to debug counter:\n " |
730 | 0 | << MI); |
731 | 0 | continue; |
732 | 0 | } |
733 | | |
734 | 17.1k | LLVM_DEBUG(dbgs() << "MCP: Replacing " << printReg(MOUse.getReg(), TRI) |
735 | 17.1k | << "\n with " << printReg(ForwardedReg, TRI) |
736 | 17.1k | << "\n in " << MI << " from " << *Copy); |
737 | | |
738 | 17.1k | MOUse.setReg(ForwardedReg); |
739 | | |
740 | 17.1k | if (!CopySrc.isRenamable()) |
741 | 8.43k | MOUse.setIsRenamable(false); |
742 | 17.1k | MOUse.setIsUndef(CopySrc.isUndef()); |
743 | | |
744 | 17.1k | LLVM_DEBUG(dbgs() << "MCP: After replacement: " << MI << "\n"); |
745 | | |
746 | | // Clear kill markers that may have been invalidated. |
747 | 17.1k | for (MachineInstr &KMI : |
748 | 17.1k | make_range(Copy->getIterator(), std::next(MI.getIterator()))) |
749 | 118k | KMI.clearRegisterKills(CopySrcReg, TRI); |
750 | | |
751 | 17.1k | ++NumCopyForwards; |
752 | 17.1k | Changed = true; |
753 | 17.1k | } |
754 | 4.50M | } |
755 | | |
756 | 387k | void MachineCopyPropagation::ForwardCopyPropagateBlock(MachineBasicBlock &MBB) { |
757 | 387k | LLVM_DEBUG(dbgs() << "MCP: ForwardCopyPropagateBlock " << MBB.getName() |
758 | 387k | << "\n"); |
759 | | |
760 | 8.89M | for (MachineInstr &MI : llvm::make_early_inc_range(MBB)) { |
761 | | // Analyze copies (which don't overlap themselves). |
762 | 8.89M | std::optional<DestSourcePair> CopyOperands = |
763 | 8.89M | isCopyInstr(MI, *TII, UseCopyInstr); |
764 | 8.89M | if (CopyOperands) { |
765 | | |
766 | 570k | Register RegSrc = CopyOperands->Source->getReg(); |
767 | 570k | Register RegDef = CopyOperands->Destination->getReg(); |
768 | | |
769 | 570k | if (!TRI->regsOverlap(RegDef, RegSrc)) { |
770 | 569k | assert(RegDef.isPhysical() && RegSrc.isPhysical() && |
771 | 569k | "MachineCopyPropagation should be run after register allocation!"); |
772 | | |
773 | 0 | MCRegister Def = RegDef.asMCReg(); |
774 | 569k | MCRegister Src = RegSrc.asMCReg(); |
775 | | |
776 | | // The two copies cancel out and the source of the first copy |
777 | | // hasn't been overridden, eliminate the second one. e.g. |
778 | | // %ecx = COPY %eax |
779 | | // ... nothing clobbered eax. |
780 | | // %eax = COPY %ecx |
781 | | // => |
782 | | // %ecx = COPY %eax |
783 | | // |
784 | | // or |
785 | | // |
786 | | // %ecx = COPY %eax |
787 | | // ... nothing clobbered eax. |
788 | | // %ecx = COPY %eax |
789 | | // => |
790 | | // %ecx = COPY %eax |
791 | 569k | if (eraseIfRedundant(MI, Def, Src) || eraseIfRedundant(MI, Src, Def)) |
792 | 2.16k | continue; |
793 | | |
794 | 567k | forwardUses(MI); |
795 | | |
796 | | // Src may have been changed by forwardUses() |
797 | 567k | CopyOperands = isCopyInstr(MI, *TII, UseCopyInstr); |
798 | 567k | Src = CopyOperands->Source->getReg().asMCReg(); |
799 | | |
800 | | // If Src is defined by a previous copy, the previous copy cannot be |
801 | | // eliminated. |
802 | 567k | ReadRegister(Src, MI, RegularUse); |
803 | 567k | for (const MachineOperand &MO : MI.implicit_operands()) { |
804 | 5.18k | if (!MO.isReg() || !MO.readsReg()) |
805 | 3.03k | continue; |
806 | 2.15k | MCRegister Reg = MO.getReg().asMCReg(); |
807 | 2.15k | if (!Reg) |
808 | 0 | continue; |
809 | 2.15k | ReadRegister(Reg, MI, RegularUse); |
810 | 2.15k | } |
811 | | |
812 | 567k | LLVM_DEBUG(dbgs() << "MCP: Copy is a deletion candidate: "; MI.dump()); |
813 | | |
814 | | // Copy is now a candidate for deletion. |
815 | 567k | if (!MRI->isReserved(Def)) |
816 | 537k | MaybeDeadCopies.insert(&MI); |
817 | | |
818 | | // If 'Def' is previously source of another copy, then this earlier copy's |
819 | | // source is no longer available. e.g. |
820 | | // %xmm9 = copy %xmm2 |
821 | | // ... |
822 | | // %xmm2 = copy %xmm0 |
823 | | // ... |
824 | | // %xmm2 = copy %xmm9 |
825 | 567k | Tracker.clobberRegister(Def, *TRI, *TII, UseCopyInstr); |
826 | 567k | for (const MachineOperand &MO : MI.implicit_operands()) { |
827 | 5.18k | if (!MO.isReg() || !MO.isDef()) |
828 | 2.15k | continue; |
829 | 3.03k | MCRegister Reg = MO.getReg().asMCReg(); |
830 | 3.03k | if (!Reg) |
831 | 0 | continue; |
832 | 3.03k | Tracker.clobberRegister(Reg, *TRI, *TII, UseCopyInstr); |
833 | 3.03k | } |
834 | | |
835 | 567k | Tracker.trackCopy(&MI, *TRI, *TII, UseCopyInstr); |
836 | | |
837 | 567k | continue; |
838 | 569k | } |
839 | 570k | } |
840 | | |
841 | | // Clobber any earlyclobber regs first. |
842 | 8.32M | for (const MachineOperand &MO : MI.operands()) |
843 | 28.1M | if (MO.isReg() && MO.isEarlyClobber()) { |
844 | 18.9k | MCRegister Reg = MO.getReg().asMCReg(); |
845 | | // If we have a tied earlyclobber, that means it is also read by this |
846 | | // instruction, so we need to make sure we don't remove it as dead |
847 | | // later. |
848 | 18.9k | if (MO.isTied()) |
849 | 15.1k | ReadRegister(Reg, MI, RegularUse); |
850 | 18.9k | Tracker.clobberRegister(Reg, *TRI, *TII, UseCopyInstr); |
851 | 18.9k | } |
852 | | |
853 | 8.32M | forwardUses(MI); |
854 | | |
855 | | // Not a copy. |
856 | 8.32M | SmallVector<Register, 2> Defs; |
857 | 8.32M | const MachineOperand *RegMask = nullptr; |
858 | 28.1M | for (const MachineOperand &MO : MI.operands()) { |
859 | 28.1M | if (MO.isRegMask()) |
860 | 218k | RegMask = &MO; |
861 | 28.1M | if (!MO.isReg()) |
862 | 9.51M | continue; |
863 | 18.6M | Register Reg = MO.getReg(); |
864 | 18.6M | if (!Reg) |
865 | 714k | continue; |
866 | | |
867 | 17.9M | assert(!Reg.isVirtual() && |
868 | 17.9M | "MachineCopyPropagation should be run after register allocation!"); |
869 | | |
870 | 17.9M | if (MO.isDef() && !MO.isEarlyClobber()) { |
871 | 6.79M | Defs.push_back(Reg.asMCReg()); |
872 | 6.79M | continue; |
873 | 11.1M | } else if (MO.readsReg()) |
874 | 10.8M | ReadRegister(Reg.asMCReg(), MI, MO.isDebug() ? DebugUse : RegularUse); |
875 | 17.9M | } |
876 | | |
877 | | // The instruction has a register mask operand which means that it clobbers |
878 | | // a large set of registers. Treat clobbered registers the same way as |
879 | | // defined registers. |
880 | 8.32M | if (RegMask) { |
881 | | // Erase any MaybeDeadCopies whose destination register is clobbered. |
882 | 218k | for (SmallSetVector<MachineInstr *, 8>::iterator DI = |
883 | 218k | MaybeDeadCopies.begin(); |
884 | 392k | DI != MaybeDeadCopies.end();) { |
885 | 173k | MachineInstr *MaybeDead = *DI; |
886 | 173k | std::optional<DestSourcePair> CopyOperands = |
887 | 173k | isCopyInstr(*MaybeDead, *TII, UseCopyInstr); |
888 | 173k | MCRegister Reg = CopyOperands->Destination->getReg().asMCReg(); |
889 | 173k | assert(!MRI->isReserved(Reg)); |
890 | | |
891 | 173k | if (!RegMask->clobbersPhysReg(Reg)) { |
892 | 173k | ++DI; |
893 | 173k | continue; |
894 | 173k | } |
895 | | |
896 | 176 | LLVM_DEBUG(dbgs() << "MCP: Removing copy due to regmask clobbering: "; |
897 | 176 | MaybeDead->dump()); |
898 | | |
899 | | // Make sure we invalidate any entries in the copy maps before erasing |
900 | | // the instruction. |
901 | 176 | Tracker.clobberRegister(Reg, *TRI, *TII, UseCopyInstr); |
902 | | |
903 | | // erase() will return the next valid iterator pointing to the next |
904 | | // element after the erased one. |
905 | 176 | DI = MaybeDeadCopies.erase(DI); |
906 | 176 | MaybeDead->eraseFromParent(); |
907 | 176 | Changed = true; |
908 | 176 | ++NumDeletes; |
909 | 176 | } |
910 | 218k | } |
911 | | |
912 | | // Any previous copy definition or reading the Defs is no longer available. |
913 | 8.32M | for (MCRegister Reg : Defs) |
914 | 6.79M | Tracker.clobberRegister(Reg, *TRI, *TII, UseCopyInstr); |
915 | 8.32M | } |
916 | | |
917 | | // If MBB doesn't have successors, delete the copies whose defs are not used. |
918 | | // If MBB does have successors, then conservative assume the defs are live-out |
919 | | // since we don't want to trust live-in lists. |
920 | 387k | if (MBB.succ_empty()) { |
921 | 219k | for (MachineInstr *MaybeDead : MaybeDeadCopies) { |
922 | 759 | LLVM_DEBUG(dbgs() << "MCP: Removing copy due to no live-out succ: "; |
923 | 759 | MaybeDead->dump()); |
924 | | |
925 | 759 | std::optional<DestSourcePair> CopyOperands = |
926 | 759 | isCopyInstr(*MaybeDead, *TII, UseCopyInstr); |
927 | 759 | assert(CopyOperands); |
928 | | |
929 | 0 | Register SrcReg = CopyOperands->Source->getReg(); |
930 | 759 | Register DestReg = CopyOperands->Destination->getReg(); |
931 | 759 | assert(!MRI->isReserved(DestReg)); |
932 | | |
933 | | // Update matching debug values, if any. |
934 | 0 | SmallVector<MachineInstr *> MaybeDeadDbgUsers( |
935 | 759 | CopyDbgUsers[MaybeDead].begin(), CopyDbgUsers[MaybeDead].end()); |
936 | 759 | MRI->updateDbgUsersToReg(DestReg.asMCReg(), SrcReg.asMCReg(), |
937 | 759 | MaybeDeadDbgUsers); |
938 | | |
939 | 759 | MaybeDead->eraseFromParent(); |
940 | 759 | Changed = true; |
941 | 759 | ++NumDeletes; |
942 | 759 | } |
943 | 219k | } |
944 | | |
945 | 387k | MaybeDeadCopies.clear(); |
946 | 387k | CopyDbgUsers.clear(); |
947 | 387k | Tracker.clear(); |
948 | 387k | } |
949 | | |
950 | | static bool isBackwardPropagatableCopy(const DestSourcePair &CopyOperands, |
951 | | const MachineRegisterInfo &MRI, |
952 | 478k | const TargetInstrInfo &TII) { |
953 | 478k | Register Def = CopyOperands.Destination->getReg(); |
954 | 478k | Register Src = CopyOperands.Source->getReg(); |
955 | | |
956 | 478k | if (!Def || !Src) |
957 | 0 | return false; |
958 | | |
959 | 478k | if (MRI.isReserved(Def) || MRI.isReserved(Src)) |
960 | 72.8k | return false; |
961 | | |
962 | 405k | return CopyOperands.Source->isRenamable() && CopyOperands.Source->isKill(); |
963 | 478k | } |
964 | | |
965 | 8.76M | void MachineCopyPropagation::propagateDefs(MachineInstr &MI) { |
966 | 8.76M | if (!Tracker.hasAnyCopies()) |
967 | 8.32M | return; |
968 | | |
969 | 2.05M | for (unsigned OpIdx = 0, OpEnd = MI.getNumOperands(); OpIdx != OpEnd; |
970 | 1.61M | ++OpIdx) { |
971 | 1.61M | MachineOperand &MODef = MI.getOperand(OpIdx); |
972 | | |
973 | 1.61M | if (!MODef.isReg() || MODef.isUse()) |
974 | 1.21M | continue; |
975 | | |
976 | | // Ignore non-trivial cases. |
977 | 404k | if (MODef.isTied() || MODef.isUndef() || MODef.isImplicit()) |
978 | 157k | continue; |
979 | | |
980 | 246k | if (!MODef.getReg()) |
981 | 0 | continue; |
982 | | |
983 | | // We only handle if the register comes from a vreg. |
984 | 246k | if (!MODef.isRenamable()) |
985 | 40.2k | continue; |
986 | | |
987 | 206k | MachineInstr *Copy = Tracker.findAvailBackwardCopy( |
988 | 206k | MI, MODef.getReg().asMCReg(), *TRI, *TII, UseCopyInstr); |
989 | 206k | if (!Copy) |
990 | 202k | continue; |
991 | | |
992 | 3.90k | std::optional<DestSourcePair> CopyOperands = |
993 | 3.90k | isCopyInstr(*Copy, *TII, UseCopyInstr); |
994 | 3.90k | Register Def = CopyOperands->Destination->getReg(); |
995 | 3.90k | Register Src = CopyOperands->Source->getReg(); |
996 | | |
997 | 3.90k | if (MODef.getReg() != Src) |
998 | 216 | continue; |
999 | | |
1000 | 3.68k | if (!isBackwardPropagatableRegClassCopy(*Copy, MI, OpIdx)) |
1001 | 3.52k | continue; |
1002 | | |
1003 | 163 | if (hasImplicitOverlap(MI, MODef)) |
1004 | 0 | continue; |
1005 | | |
1006 | 163 | if (hasOverlappingMultipleDef(MI, MODef, Def)) |
1007 | 0 | continue; |
1008 | | |
1009 | 163 | LLVM_DEBUG(dbgs() << "MCP: Replacing " << printReg(MODef.getReg(), TRI) |
1010 | 163 | << "\n with " << printReg(Def, TRI) << "\n in " |
1011 | 163 | << MI << " from " << *Copy); |
1012 | | |
1013 | 163 | MODef.setReg(Def); |
1014 | 163 | MODef.setIsRenamable(CopyOperands->Destination->isRenamable()); |
1015 | | |
1016 | 163 | LLVM_DEBUG(dbgs() << "MCP: After replacement: " << MI << "\n"); |
1017 | 163 | MaybeDeadCopies.insert(Copy); |
1018 | 163 | Changed = true; |
1019 | 163 | ++NumCopyBackwardPropagated; |
1020 | 163 | } |
1021 | 440k | } |
1022 | | |
1023 | | void MachineCopyPropagation::BackwardCopyPropagateBlock( |
1024 | 387k | MachineBasicBlock &MBB) { |
1025 | 387k | LLVM_DEBUG(dbgs() << "MCP: BackwardCopyPropagateBlock " << MBB.getName() |
1026 | 387k | << "\n"); |
1027 | | |
1028 | 8.89M | for (MachineInstr &MI : llvm::make_early_inc_range(llvm::reverse(MBB))) { |
1029 | | // Ignore non-trivial COPYs. |
1030 | 8.89M | std::optional<DestSourcePair> CopyOperands = |
1031 | 8.89M | isCopyInstr(MI, *TII, UseCopyInstr); |
1032 | 8.89M | if (CopyOperands && MI.getNumOperands() == 2) { |
1033 | 479k | Register DefReg = CopyOperands->Destination->getReg(); |
1034 | 479k | Register SrcReg = CopyOperands->Source->getReg(); |
1035 | | |
1036 | 479k | if (!TRI->regsOverlap(DefReg, SrcReg)) { |
1037 | | // Unlike forward cp, we don't invoke propagateDefs here, |
1038 | | // just let forward cp do COPY-to-COPY propagation. |
1039 | 478k | if (isBackwardPropagatableCopy(*CopyOperands, *MRI, *TII)) { |
1040 | 134k | Tracker.invalidateRegister(SrcReg.asMCReg(), *TRI, *TII, |
1041 | 134k | UseCopyInstr); |
1042 | 134k | Tracker.invalidateRegister(DefReg.asMCReg(), *TRI, *TII, |
1043 | 134k | UseCopyInstr); |
1044 | 134k | Tracker.trackCopy(&MI, *TRI, *TII, UseCopyInstr); |
1045 | 134k | continue; |
1046 | 134k | } |
1047 | 478k | } |
1048 | 479k | } |
1049 | | |
1050 | | // Invalidate any earlyclobber regs first. |
1051 | 8.76M | for (const MachineOperand &MO : MI.operands()) |
1052 | 29.1M | if (MO.isReg() && MO.isEarlyClobber()) { |
1053 | 18.9k | MCRegister Reg = MO.getReg().asMCReg(); |
1054 | 18.9k | if (!Reg) |
1055 | 0 | continue; |
1056 | 18.9k | Tracker.invalidateRegister(Reg, *TRI, *TII, UseCopyInstr); |
1057 | 18.9k | } |
1058 | | |
1059 | 8.76M | propagateDefs(MI); |
1060 | 29.1M | for (const MachineOperand &MO : MI.operands()) { |
1061 | 29.1M | if (!MO.isReg()) |
1062 | 9.60M | continue; |
1063 | | |
1064 | 19.5M | if (!MO.getReg()) |
1065 | 714k | continue; |
1066 | | |
1067 | 18.8M | if (MO.isDef()) |
1068 | 7.25M | Tracker.invalidateRegister(MO.getReg().asMCReg(), *TRI, *TII, |
1069 | 7.25M | UseCopyInstr); |
1070 | | |
1071 | 18.8M | if (MO.readsReg()) { |
1072 | 11.3M | if (MO.isDebug()) { |
1073 | | // Check if the register in the debug instruction is utilized |
1074 | | // in a copy instruction, so we can update the debug info if the |
1075 | | // register is changed. |
1076 | 2.51k | for (MCRegUnit Unit : TRI->regunits(MO.getReg().asMCReg())) { |
1077 | 2.51k | if (auto *Copy = Tracker.findCopyDefViaUnit(Unit, *TRI)) { |
1078 | 21 | CopyDbgUsers[Copy].insert(&MI); |
1079 | 21 | } |
1080 | 2.51k | } |
1081 | 11.2M | } else { |
1082 | 11.2M | Tracker.invalidateRegister(MO.getReg().asMCReg(), *TRI, *TII, |
1083 | 11.2M | UseCopyInstr); |
1084 | 11.2M | } |
1085 | 11.3M | } |
1086 | 18.8M | } |
1087 | 8.76M | } |
1088 | | |
1089 | 387k | for (auto *Copy : MaybeDeadCopies) { |
1090 | 163 | std::optional<DestSourcePair> CopyOperands = |
1091 | 163 | isCopyInstr(*Copy, *TII, UseCopyInstr); |
1092 | 163 | Register Src = CopyOperands->Source->getReg(); |
1093 | 163 | Register Def = CopyOperands->Destination->getReg(); |
1094 | 163 | SmallVector<MachineInstr *> MaybeDeadDbgUsers(CopyDbgUsers[Copy].begin(), |
1095 | 163 | CopyDbgUsers[Copy].end()); |
1096 | | |
1097 | 163 | MRI->updateDbgUsersToReg(Src.asMCReg(), Def.asMCReg(), MaybeDeadDbgUsers); |
1098 | 163 | Copy->eraseFromParent(); |
1099 | 163 | ++NumDeletes; |
1100 | 163 | } |
1101 | | |
1102 | 387k | MaybeDeadCopies.clear(); |
1103 | 387k | CopyDbgUsers.clear(); |
1104 | 387k | Tracker.clear(); |
1105 | 387k | } |
1106 | | |
1107 | | static void LLVM_ATTRIBUTE_UNUSED printSpillReloadChain( |
1108 | | DenseMap<MachineInstr *, SmallVector<MachineInstr *>> &SpillChain, |
1109 | | DenseMap<MachineInstr *, SmallVector<MachineInstr *>> &ReloadChain, |
1110 | 0 | MachineInstr *Leader) { |
1111 | 0 | auto &SC = SpillChain[Leader]; |
1112 | 0 | auto &RC = ReloadChain[Leader]; |
1113 | 0 | for (auto I = SC.rbegin(), E = SC.rend(); I != E; ++I) |
1114 | 0 | (*I)->dump(); |
1115 | 0 | for (MachineInstr *MI : RC) |
1116 | 0 | MI->dump(); |
1117 | 0 | } |
1118 | | |
1119 | | // Remove spill-reload like copy chains. For example |
1120 | | // r0 = COPY r1 |
1121 | | // r1 = COPY r2 |
1122 | | // r2 = COPY r3 |
1123 | | // r3 = COPY r4 |
1124 | | // <def-use r4> |
1125 | | // r4 = COPY r3 |
1126 | | // r3 = COPY r2 |
1127 | | // r2 = COPY r1 |
1128 | | // r1 = COPY r0 |
1129 | | // will be folded into |
1130 | | // r0 = COPY r1 |
1131 | | // r1 = COPY r4 |
1132 | | // <def-use r4> |
1133 | | // r4 = COPY r1 |
1134 | | // r1 = COPY r0 |
1135 | | // TODO: Currently we don't track usage of r0 outside the chain, so we |
1136 | | // conservatively keep its value as it was before the rewrite. |
1137 | | // |
1138 | | // The algorithm is trying to keep |
1139 | | // property#1: No Def of spill COPY in the chain is used or defined until the |
1140 | | // paired reload COPY in the chain uses the Def. |
1141 | | // |
1142 | | // property#2: NO Source of COPY in the chain is used or defined until the next |
1143 | | // COPY in the chain defines the Source, except the innermost spill-reload |
1144 | | // pair. |
1145 | | // |
1146 | | // The algorithm is conducted by checking every COPY inside the MBB, assuming |
1147 | | // the COPY is a reload COPY, then try to find paired spill COPY by searching |
1148 | | // the COPY defines the Src of the reload COPY backward. If such pair is found, |
1149 | | // it either belongs to an existing chain or a new chain depends on |
1150 | | // last available COPY uses the Def of the reload COPY. |
1151 | | // Implementation notes, we use CopyTracker::findLastDefCopy(Reg, ...) to find |
1152 | | // out last COPY that defines Reg; we use CopyTracker::findLastUseCopy(Reg, ...) |
1153 | | // to find out last COPY that uses Reg. When we are encountered with a Non-COPY |
1154 | | // instruction, we check registers in the operands of this instruction. If this |
1155 | | // Reg is defined by a COPY, we untrack this Reg via |
1156 | | // CopyTracker::clobberRegister(Reg, ...). |
1157 | 112k | void MachineCopyPropagation::EliminateSpillageCopies(MachineBasicBlock &MBB) { |
1158 | | // ChainLeader maps MI inside a spill-reload chain to its innermost reload COPY. |
1159 | | // Thus we can track if a MI belongs to an existing spill-reload chain. |
1160 | 112k | DenseMap<MachineInstr *, MachineInstr *> ChainLeader; |
1161 | | // SpillChain maps innermost reload COPY of a spill-reload chain to a sequence |
1162 | | // of COPYs that forms spills of a spill-reload chain. |
1163 | | // ReloadChain maps innermost reload COPY of a spill-reload chain to a |
1164 | | // sequence of COPYs that forms reloads of a spill-reload chain. |
1165 | 112k | DenseMap<MachineInstr *, SmallVector<MachineInstr *>> SpillChain, ReloadChain; |
1166 | | // If a COPY's Source has use or def until next COPY defines the Source, |
1167 | | // we put the COPY in this set to keep property#2. |
1168 | 112k | DenseSet<const MachineInstr *> CopySourceInvalid; |
1169 | | |
1170 | 112k | auto TryFoldSpillageCopies = |
1171 | 112k | [&, this](const SmallVectorImpl<MachineInstr *> &SC, |
1172 | 112k | const SmallVectorImpl<MachineInstr *> &RC) { |
1173 | 389 | assert(SC.size() == RC.size() && "Spill-reload should be paired"); |
1174 | | |
1175 | | // We need at least 3 pairs of copies for the transformation to apply, |
1176 | | // because the first outermost pair cannot be removed since we don't |
1177 | | // recolor outside of the chain and that we need at least one temporary |
1178 | | // spill slot to shorten the chain. If we only have a chain of two |
1179 | | // pairs, we already have the shortest sequence this code can handle: |
1180 | | // the outermost pair for the temporary spill slot, and the pair that |
1181 | | // use that temporary spill slot for the other end of the chain. |
1182 | | // TODO: We might be able to simplify to one spill-reload pair if collecting |
1183 | | // more infomation about the outermost COPY. |
1184 | 389 | if (SC.size() <= 2) |
1185 | 361 | return; |
1186 | | |
1187 | | // If violate property#2, we don't fold the chain. |
1188 | 28 | for (const MachineInstr *Spill : drop_begin(SC)) |
1189 | 228 | if (CopySourceInvalid.count(Spill)) |
1190 | 0 | return; |
1191 | | |
1192 | 28 | for (const MachineInstr *Reload : drop_end(RC)) |
1193 | 228 | if (CopySourceInvalid.count(Reload)) |
1194 | 0 | return; |
1195 | | |
1196 | 56 | auto CheckCopyConstraint = [this](Register Def, Register Src) { |
1197 | 266 | for (const TargetRegisterClass *RC : TRI->regclasses()) { |
1198 | 266 | if (RC->contains(Def) && RC->contains(Src)) |
1199 | 56 | return true; |
1200 | 266 | } |
1201 | 0 | return false; |
1202 | 56 | }; |
1203 | | |
1204 | 28 | auto UpdateReg = [](MachineInstr *MI, const MachineOperand *Old, |
1205 | 56 | const MachineOperand *New) { |
1206 | 112 | for (MachineOperand &MO : MI->operands()) { |
1207 | 112 | if (&MO == Old) |
1208 | 56 | MO.setReg(New->getReg()); |
1209 | 112 | } |
1210 | 56 | }; |
1211 | | |
1212 | 28 | std::optional<DestSourcePair> InnerMostSpillCopy = |
1213 | 28 | isCopyInstr(*SC[0], *TII, UseCopyInstr); |
1214 | 28 | std::optional<DestSourcePair> OuterMostSpillCopy = |
1215 | 28 | isCopyInstr(*SC.back(), *TII, UseCopyInstr); |
1216 | 28 | std::optional<DestSourcePair> InnerMostReloadCopy = |
1217 | 28 | isCopyInstr(*RC[0], *TII, UseCopyInstr); |
1218 | 28 | std::optional<DestSourcePair> OuterMostReloadCopy = |
1219 | 28 | isCopyInstr(*RC.back(), *TII, UseCopyInstr); |
1220 | 28 | if (!CheckCopyConstraint(OuterMostSpillCopy->Source->getReg(), |
1221 | 28 | InnerMostSpillCopy->Source->getReg()) || |
1222 | 28 | !CheckCopyConstraint(InnerMostReloadCopy->Destination->getReg(), |
1223 | 28 | OuterMostReloadCopy->Destination->getReg())) |
1224 | 0 | return; |
1225 | | |
1226 | 28 | SpillageChainsLength += SC.size() + RC.size(); |
1227 | 28 | NumSpillageChains += 1; |
1228 | 28 | UpdateReg(SC[0], InnerMostSpillCopy->Destination, |
1229 | 28 | OuterMostSpillCopy->Source); |
1230 | 28 | UpdateReg(RC[0], InnerMostReloadCopy->Source, |
1231 | 28 | OuterMostReloadCopy->Destination); |
1232 | | |
1233 | 228 | for (size_t I = 1; I < SC.size() - 1; ++I) { |
1234 | 200 | SC[I]->eraseFromParent(); |
1235 | 200 | RC[I]->eraseFromParent(); |
1236 | 200 | NumDeletes += 2; |
1237 | 200 | } |
1238 | 28 | }; |
1239 | | |
1240 | 112k | auto IsFoldableCopy = [this](const MachineInstr &MaybeCopy) { |
1241 | 5.66k | if (MaybeCopy.getNumImplicitOperands() > 0) |
1242 | 0 | return false; |
1243 | 5.66k | std::optional<DestSourcePair> CopyOperands = |
1244 | 5.66k | isCopyInstr(MaybeCopy, *TII, UseCopyInstr); |
1245 | 5.66k | if (!CopyOperands) |
1246 | 0 | return false; |
1247 | 5.66k | Register Src = CopyOperands->Source->getReg(); |
1248 | 5.66k | Register Def = CopyOperands->Destination->getReg(); |
1249 | 5.66k | return Src && Def && !TRI->regsOverlap(Src, Def) && |
1250 | 5.66k | CopyOperands->Source->isRenamable() && |
1251 | 5.66k | CopyOperands->Destination->isRenamable(); |
1252 | 5.66k | }; |
1253 | | |
1254 | 112k | auto IsSpillReloadPair = [&, this](const MachineInstr &Spill, |
1255 | 112k | const MachineInstr &Reload) { |
1256 | 3.93k | if (!IsFoldableCopy(Spill) || !IsFoldableCopy(Reload)) |
1257 | 2.79k | return false; |
1258 | 1.13k | std::optional<DestSourcePair> SpillCopy = |
1259 | 1.13k | isCopyInstr(Spill, *TII, UseCopyInstr); |
1260 | 1.13k | std::optional<DestSourcePair> ReloadCopy = |
1261 | 1.13k | isCopyInstr(Reload, *TII, UseCopyInstr); |
1262 | 1.13k | if (!SpillCopy || !ReloadCopy) |
1263 | 0 | return false; |
1264 | 1.13k | return SpillCopy->Source->getReg() == ReloadCopy->Destination->getReg() && |
1265 | 1.13k | SpillCopy->Destination->getReg() == ReloadCopy->Source->getReg(); |
1266 | 1.13k | }; |
1267 | | |
1268 | 112k | auto IsChainedCopy = [&, this](const MachineInstr &Prev, |
1269 | 112k | const MachineInstr &Current) { |
1270 | 281 | if (!IsFoldableCopy(Prev) || !IsFoldableCopy(Current)) |
1271 | 0 | return false; |
1272 | 281 | std::optional<DestSourcePair> PrevCopy = |
1273 | 281 | isCopyInstr(Prev, *TII, UseCopyInstr); |
1274 | 281 | std::optional<DestSourcePair> CurrentCopy = |
1275 | 281 | isCopyInstr(Current, *TII, UseCopyInstr); |
1276 | 281 | if (!PrevCopy || !CurrentCopy) |
1277 | 0 | return false; |
1278 | 281 | return PrevCopy->Source->getReg() == CurrentCopy->Destination->getReg(); |
1279 | 281 | }; |
1280 | | |
1281 | 2.47M | for (MachineInstr &MI : llvm::make_early_inc_range(MBB)) { |
1282 | 2.47M | std::optional<DestSourcePair> CopyOperands = |
1283 | 2.47M | isCopyInstr(MI, *TII, UseCopyInstr); |
1284 | | |
1285 | | // Update track information via non-copy instruction. |
1286 | 2.47M | SmallSet<Register, 8> RegsToClobber; |
1287 | 2.47M | if (!CopyOperands) { |
1288 | 7.86M | for (const MachineOperand &MO : MI.operands()) { |
1289 | 7.86M | if (!MO.isReg()) |
1290 | 2.36M | continue; |
1291 | 5.49M | Register Reg = MO.getReg(); |
1292 | 5.49M | if (!Reg) |
1293 | 2.30k | continue; |
1294 | 5.49M | MachineInstr *LastUseCopy = |
1295 | 5.49M | Tracker.findLastSeenUseInCopy(Reg.asMCReg(), *TRI); |
1296 | 5.49M | if (LastUseCopy) { |
1297 | 69.7k | LLVM_DEBUG(dbgs() << "MCP: Copy source of\n"); |
1298 | 69.7k | LLVM_DEBUG(LastUseCopy->dump()); |
1299 | 69.7k | LLVM_DEBUG(dbgs() << "might be invalidated by\n"); |
1300 | 69.7k | LLVM_DEBUG(MI.dump()); |
1301 | 69.7k | CopySourceInvalid.insert(LastUseCopy); |
1302 | 69.7k | } |
1303 | | // Must be noted Tracker.clobberRegister(Reg, ...) removes tracking of |
1304 | | // Reg, i.e, COPY that defines Reg is removed from the mapping as well |
1305 | | // as marking COPYs that uses Reg unavailable. |
1306 | | // We don't invoke CopyTracker::clobberRegister(Reg, ...) if Reg is not |
1307 | | // defined by a previous COPY, since we don't want to make COPYs uses |
1308 | | // Reg unavailable. |
1309 | 5.49M | if (Tracker.findLastSeenDefInCopy(MI, Reg.asMCReg(), *TRI, *TII, |
1310 | 5.49M | UseCopyInstr)) |
1311 | | // Thus we can keep the property#1. |
1312 | 51.4k | RegsToClobber.insert(Reg); |
1313 | 5.49M | } |
1314 | 2.41M | for (Register Reg : RegsToClobber) { |
1315 | 41.5k | Tracker.clobberRegister(Reg, *TRI, *TII, UseCopyInstr); |
1316 | 41.5k | LLVM_DEBUG(dbgs() << "MCP: Removed tracking of " << printReg(Reg, TRI) |
1317 | 41.5k | << "\n"); |
1318 | 41.5k | } |
1319 | 2.41M | continue; |
1320 | 2.41M | } |
1321 | | |
1322 | 50.9k | Register Src = CopyOperands->Source->getReg(); |
1323 | 50.9k | Register Def = CopyOperands->Destination->getReg(); |
1324 | | // Check if we can find a pair spill-reload copy. |
1325 | 50.9k | LLVM_DEBUG(dbgs() << "MCP: Searching paired spill for reload: "); |
1326 | 50.9k | LLVM_DEBUG(MI.dump()); |
1327 | 50.9k | MachineInstr *MaybeSpill = |
1328 | 50.9k | Tracker.findLastSeenDefInCopy(MI, Src.asMCReg(), *TRI, *TII, UseCopyInstr); |
1329 | 50.9k | bool MaybeSpillIsChained = ChainLeader.count(MaybeSpill); |
1330 | 50.9k | if (!MaybeSpillIsChained && MaybeSpill && |
1331 | 50.9k | IsSpillReloadPair(*MaybeSpill, MI)) { |
1332 | | // Check if we already have an existing chain. Now we have a |
1333 | | // spill-reload pair. |
1334 | | // L2: r2 = COPY r3 |
1335 | | // L5: r3 = COPY r2 |
1336 | | // Looking for a valid COPY before L5 which uses r3. |
1337 | | // This can be serverial cases. |
1338 | | // Case #1: |
1339 | | // No COPY is found, which can be r3 is def-use between (L2, L5), we |
1340 | | // create a new chain for L2 and L5. |
1341 | | // Case #2: |
1342 | | // L2: r2 = COPY r3 |
1343 | | // L5: r3 = COPY r2 |
1344 | | // Such COPY is found and is L2, we create a new chain for L2 and L5. |
1345 | | // Case #3: |
1346 | | // L2: r2 = COPY r3 |
1347 | | // L3: r1 = COPY r3 |
1348 | | // L5: r3 = COPY r2 |
1349 | | // we create a new chain for L2 and L5. |
1350 | | // Case #4: |
1351 | | // L2: r2 = COPY r3 |
1352 | | // L3: r1 = COPY r3 |
1353 | | // L4: r3 = COPY r1 |
1354 | | // L5: r3 = COPY r2 |
1355 | | // Such COPY won't be found since L4 defines r3. we create a new chain |
1356 | | // for L2 and L5. |
1357 | | // Case #5: |
1358 | | // L2: r2 = COPY r3 |
1359 | | // L3: r3 = COPY r1 |
1360 | | // L4: r1 = COPY r3 |
1361 | | // L5: r3 = COPY r2 |
1362 | | // COPY is found and is L4 which belongs to an existing chain, we add |
1363 | | // L2 and L5 to this chain. |
1364 | 650 | LLVM_DEBUG(dbgs() << "MCP: Found spill: "); |
1365 | 650 | LLVM_DEBUG(MaybeSpill->dump()); |
1366 | 650 | MachineInstr *MaybePrevReload = |
1367 | 650 | Tracker.findLastSeenUseInCopy(Def.asMCReg(), *TRI); |
1368 | 650 | auto Leader = ChainLeader.find(MaybePrevReload); |
1369 | 650 | MachineInstr *L = nullptr; |
1370 | 650 | if (Leader == ChainLeader.end() || |
1371 | 650 | (MaybePrevReload && !IsChainedCopy(*MaybePrevReload, MI))) { |
1372 | 389 | L = &MI; |
1373 | 389 | assert(!SpillChain.count(L) && |
1374 | 389 | "SpillChain should not have contained newly found chain"); |
1375 | 389 | } else { |
1376 | 261 | assert(MaybePrevReload && |
1377 | 261 | "Found a valid leader through nullptr should not happend"); |
1378 | 0 | L = Leader->second; |
1379 | 261 | assert(SpillChain[L].size() > 0 && |
1380 | 261 | "Existing chain's length should be larger than zero"); |
1381 | 261 | } |
1382 | 0 | assert(!ChainLeader.count(&MI) && !ChainLeader.count(MaybeSpill) && |
1383 | 650 | "Newly found paired spill-reload should not belong to any chain " |
1384 | 650 | "at this point"); |
1385 | 0 | ChainLeader.insert({MaybeSpill, L}); |
1386 | 650 | ChainLeader.insert({&MI, L}); |
1387 | 650 | SpillChain[L].push_back(MaybeSpill); |
1388 | 650 | ReloadChain[L].push_back(&MI); |
1389 | 650 | LLVM_DEBUG(dbgs() << "MCP: Chain " << L << " now is:\n"); |
1390 | 650 | LLVM_DEBUG(printSpillReloadChain(SpillChain, ReloadChain, L)); |
1391 | 50.3k | } else if (MaybeSpill && !MaybeSpillIsChained) { |
1392 | | // MaybeSpill is unable to pair with MI. That's to say adding MI makes |
1393 | | // the chain invalid. |
1394 | | // The COPY defines Src is no longer considered as a candidate of a |
1395 | | // valid chain. Since we expect the Def of a spill copy isn't used by |
1396 | | // any COPY instruction until a reload copy. For example: |
1397 | | // L1: r1 = COPY r2 |
1398 | | // L2: r3 = COPY r1 |
1399 | | // If we later have |
1400 | | // L1: r1 = COPY r2 |
1401 | | // L2: r3 = COPY r1 |
1402 | | // L3: r2 = COPY r1 |
1403 | | // L1 and L3 can't be a valid spill-reload pair. |
1404 | | // Thus we keep the property#1. |
1405 | 3.28k | LLVM_DEBUG(dbgs() << "MCP: Not paired spill-reload:\n"); |
1406 | 3.28k | LLVM_DEBUG(MaybeSpill->dump()); |
1407 | 3.28k | LLVM_DEBUG(MI.dump()); |
1408 | 3.28k | Tracker.clobberRegister(Src.asMCReg(), *TRI, *TII, UseCopyInstr); |
1409 | 3.28k | LLVM_DEBUG(dbgs() << "MCP: Removed tracking of " << printReg(Src, TRI) |
1410 | 3.28k | << "\n"); |
1411 | 3.28k | } |
1412 | 0 | Tracker.trackCopy(&MI, *TRI, *TII, UseCopyInstr); |
1413 | 50.9k | } |
1414 | | |
1415 | 113k | for (auto I = SpillChain.begin(), E = SpillChain.end(); I != E; ++I) { |
1416 | 389 | auto &SC = I->second; |
1417 | 389 | assert(ReloadChain.count(I->first) && |
1418 | 389 | "Reload chain of the same leader should exist"); |
1419 | 0 | auto &RC = ReloadChain[I->first]; |
1420 | 389 | TryFoldSpillageCopies(SC, RC); |
1421 | 389 | } |
1422 | | |
1423 | 112k | MaybeDeadCopies.clear(); |
1424 | 112k | CopyDbgUsers.clear(); |
1425 | 112k | Tracker.clear(); |
1426 | 112k | } |
1427 | | |
1428 | 219k | bool MachineCopyPropagation::runOnMachineFunction(MachineFunction &MF) { |
1429 | 219k | if (skipFunction(MF.getFunction())) |
1430 | 0 | return false; |
1431 | | |
1432 | 219k | bool isSpillageCopyElimEnabled = false; |
1433 | 219k | switch (EnableSpillageCopyElimination) { |
1434 | 219k | case cl::BOU_UNSET: |
1435 | 219k | isSpillageCopyElimEnabled = |
1436 | 219k | MF.getSubtarget().enableSpillageCopyElimination(); |
1437 | 219k | break; |
1438 | 0 | case cl::BOU_TRUE: |
1439 | 0 | isSpillageCopyElimEnabled = true; |
1440 | 0 | break; |
1441 | 0 | case cl::BOU_FALSE: |
1442 | 0 | isSpillageCopyElimEnabled = false; |
1443 | 0 | break; |
1444 | 219k | } |
1445 | | |
1446 | 219k | Changed = false; |
1447 | | |
1448 | 219k | TRI = MF.getSubtarget().getRegisterInfo(); |
1449 | 219k | TII = MF.getSubtarget().getInstrInfo(); |
1450 | 219k | MRI = &MF.getRegInfo(); |
1451 | | |
1452 | 387k | for (MachineBasicBlock &MBB : MF) { |
1453 | 387k | if (isSpillageCopyElimEnabled) |
1454 | 112k | EliminateSpillageCopies(MBB); |
1455 | 387k | BackwardCopyPropagateBlock(MBB); |
1456 | 387k | ForwardCopyPropagateBlock(MBB); |
1457 | 387k | } |
1458 | | |
1459 | 219k | return Changed; |
1460 | 219k | } |
1461 | | |
1462 | | MachineFunctionPass * |
1463 | 6.24k | llvm::createMachineCopyPropagationPass(bool UseCopyInstr = false) { |
1464 | 6.24k | return new MachineCopyPropagation(UseCopyInstr); |
1465 | 6.24k | } |