1 | //===- AArch64InstrInfo.cpp - AArch64 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 contains the AArch64 implementation of the TargetInstrInfo class. |
10 | // |
11 | //===----------------------------------------------------------------------===// |
12 | |
13 | #include "AArch64InstrInfo.h" |
14 | #include "AArch64ExpandImm.h" |
15 | #include "AArch64FrameLowering.h" |
16 | #include "AArch64MachineFunctionInfo.h" |
17 | #include "AArch64PointerAuth.h" |
18 | #include "AArch64Subtarget.h" |
19 | #include "MCTargetDesc/AArch64AddressingModes.h" |
20 | #include "MCTargetDesc/AArch64MCTargetDesc.h" |
21 | #include "Utils/AArch64BaseInfo.h" |
22 | #include "llvm/ADT/ArrayRef.h" |
23 | #include "llvm/ADT/STLExtras.h" |
24 | #include "llvm/ADT/SmallVector.h" |
25 | #include "llvm/CodeGen/LivePhysRegs.h" |
26 | #include "llvm/CodeGen/MachineBasicBlock.h" |
27 | #include "llvm/CodeGen/MachineCombinerPattern.h" |
28 | #include "llvm/CodeGen/MachineFrameInfo.h" |
29 | #include "llvm/CodeGen/MachineFunction.h" |
30 | #include "llvm/CodeGen/MachineInstr.h" |
31 | #include "llvm/CodeGen/MachineInstrBuilder.h" |
32 | #include "llvm/CodeGen/MachineMemOperand.h" |
33 | #include "llvm/CodeGen/MachineModuleInfo.h" |
34 | #include "llvm/CodeGen/MachineOperand.h" |
35 | #include "llvm/CodeGen/MachineRegisterInfo.h" |
36 | #include "llvm/CodeGen/RegisterScavenging.h" |
37 | #include "llvm/CodeGen/StackMaps.h" |
38 | #include "llvm/CodeGen/TargetRegisterInfo.h" |
39 | #include "llvm/CodeGen/TargetSubtargetInfo.h" |
40 | #include "llvm/IR/DebugInfoMetadata.h" |
41 | #include "llvm/IR/DebugLoc.h" |
42 | #include "llvm/IR/GlobalValue.h" |
43 | #include "llvm/IR/Module.h" |
44 | #include "llvm/MC/MCAsmInfo.h" |
45 | #include "llvm/MC/MCInst.h" |
46 | #include "llvm/MC/MCInstBuilder.h" |
47 | #include "llvm/MC/MCInstrDesc.h" |
48 | #include "llvm/Support/Casting.h" |
49 | #include "llvm/Support/CodeGen.h" |
50 | #include "llvm/Support/CommandLine.h" |
51 | #include "llvm/Support/ErrorHandling.h" |
52 | #include "llvm/Support/LEB128.h" |
53 | #include "llvm/Support/MathExtras.h" |
54 | #include "llvm/Target/TargetMachine.h" |
55 | #include "llvm/Target/TargetOptions.h" |
56 | #include <cassert> |
57 | #include <cstdint> |
58 | #include <iterator> |
59 | #include <utility> |
60 | |
61 | using namespace llvm; |
62 | |
63 | #define GET_INSTRINFO_CTOR_DTOR |
64 | #include "AArch64GenInstrInfo.inc" |
65 | |
66 | static cl::opt<unsigned> TBZDisplacementBits( |
67 | "aarch64-tbz-offset-bits" , cl::Hidden, cl::init(Val: 14), |
68 | cl::desc("Restrict range of TB[N]Z instructions (DEBUG)" )); |
69 | |
70 | static cl::opt<unsigned> CBZDisplacementBits( |
71 | "aarch64-cbz-offset-bits" , cl::Hidden, cl::init(Val: 19), |
72 | cl::desc("Restrict range of CB[N]Z instructions (DEBUG)" )); |
73 | |
74 | static cl::opt<unsigned> |
75 | BCCDisplacementBits("aarch64-bcc-offset-bits" , cl::Hidden, cl::init(Val: 19), |
76 | cl::desc("Restrict range of Bcc instructions (DEBUG)" )); |
77 | |
78 | static cl::opt<unsigned> |
79 | BDisplacementBits("aarch64-b-offset-bits" , cl::Hidden, cl::init(Val: 26), |
80 | cl::desc("Restrict range of B instructions (DEBUG)" )); |
81 | |
82 | AArch64InstrInfo::AArch64InstrInfo(const AArch64Subtarget &STI) |
83 | : AArch64GenInstrInfo(AArch64::ADJCALLSTACKDOWN, AArch64::ADJCALLSTACKUP, |
84 | AArch64::CATCHRET), |
85 | RI(STI.getTargetTriple()), Subtarget(STI) {} |
86 | |
87 | /// GetInstSize - Return the number of bytes of code the specified |
88 | /// instruction may be. This returns the maximum number of bytes. |
89 | unsigned AArch64InstrInfo::getInstSizeInBytes(const MachineInstr &MI) const { |
90 | const MachineBasicBlock &MBB = *MI.getParent(); |
91 | const MachineFunction *MF = MBB.getParent(); |
92 | const Function &F = MF->getFunction(); |
93 | const MCAsmInfo *MAI = MF->getTarget().getMCAsmInfo(); |
94 | |
95 | { |
96 | auto Op = MI.getOpcode(); |
97 | if (Op == AArch64::INLINEASM || Op == AArch64::INLINEASM_BR) |
98 | return getInlineAsmLength(Str: MI.getOperand(i: 0).getSymbolName(), MAI: *MAI); |
99 | } |
100 | |
101 | // Meta-instructions emit no code. |
102 | if (MI.isMetaInstruction()) |
103 | return 0; |
104 | |
105 | // FIXME: We currently only handle pseudoinstructions that don't get expanded |
106 | // before the assembly printer. |
107 | unsigned NumBytes = 0; |
108 | const MCInstrDesc &Desc = MI.getDesc(); |
109 | |
110 | // Size should be preferably set in |
111 | // llvm/lib/Target/AArch64/AArch64InstrInfo.td (default case). |
112 | // Specific cases handle instructions of variable sizes |
113 | switch (Desc.getOpcode()) { |
114 | default: |
115 | if (Desc.getSize()) |
116 | return Desc.getSize(); |
117 | |
118 | // Anything not explicitly designated otherwise (i.e. pseudo-instructions |
119 | // with fixed constant size but not specified in .td file) is a normal |
120 | // 4-byte insn. |
121 | NumBytes = 4; |
122 | break; |
123 | case TargetOpcode::STACKMAP: |
124 | // The upper bound for a stackmap intrinsic is the full length of its shadow |
125 | NumBytes = StackMapOpers(&MI).getNumPatchBytes(); |
126 | assert(NumBytes % 4 == 0 && "Invalid number of NOP bytes requested!" ); |
127 | break; |
128 | case TargetOpcode::PATCHPOINT: |
129 | // The size of the patchpoint intrinsic is the number of bytes requested |
130 | NumBytes = PatchPointOpers(&MI).getNumPatchBytes(); |
131 | assert(NumBytes % 4 == 0 && "Invalid number of NOP bytes requested!" ); |
132 | break; |
133 | case TargetOpcode::STATEPOINT: |
134 | NumBytes = StatepointOpers(&MI).getNumPatchBytes(); |
135 | assert(NumBytes % 4 == 0 && "Invalid number of NOP bytes requested!" ); |
136 | // No patch bytes means a normal call inst is emitted |
137 | if (NumBytes == 0) |
138 | NumBytes = 4; |
139 | break; |
140 | case TargetOpcode::PATCHABLE_FUNCTION_ENTER: |
141 | // If `patchable-function-entry` is set, PATCHABLE_FUNCTION_ENTER |
142 | // instructions are expanded to the specified number of NOPs. Otherwise, |
143 | // they are expanded to 36-byte XRay sleds. |
144 | NumBytes = |
145 | F.getFnAttributeAsParsedInteger(Kind: "patchable-function-entry" , Default: 9) * 4; |
146 | break; |
147 | case TargetOpcode::PATCHABLE_FUNCTION_EXIT: |
148 | case TargetOpcode::PATCHABLE_TYPED_EVENT_CALL: |
149 | // An XRay sled can be 4 bytes of alignment plus a 32-byte block. |
150 | NumBytes = 36; |
151 | break; |
152 | case TargetOpcode::PATCHABLE_EVENT_CALL: |
153 | // EVENT_CALL XRay sleds are exactly 6 instructions long (no alignment). |
154 | NumBytes = 24; |
155 | break; |
156 | |
157 | case AArch64::SPACE: |
158 | NumBytes = MI.getOperand(i: 1).getImm(); |
159 | break; |
160 | case TargetOpcode::BUNDLE: |
161 | NumBytes = getInstBundleLength(MI); |
162 | break; |
163 | } |
164 | |
165 | return NumBytes; |
166 | } |
167 | |
168 | unsigned AArch64InstrInfo::getInstBundleLength(const MachineInstr &MI) const { |
169 | unsigned Size = 0; |
170 | MachineBasicBlock::const_instr_iterator I = MI.getIterator(); |
171 | MachineBasicBlock::const_instr_iterator E = MI.getParent()->instr_end(); |
172 | while (++I != E && I->isInsideBundle()) { |
173 | assert(!I->isBundle() && "No nested bundle!" ); |
174 | Size += getInstSizeInBytes(MI: *I); |
175 | } |
176 | return Size; |
177 | } |
178 | |
179 | static void parseCondBranch(MachineInstr *LastInst, MachineBasicBlock *&Target, |
180 | SmallVectorImpl<MachineOperand> &Cond) { |
181 | // Block ends with fall-through condbranch. |
182 | switch (LastInst->getOpcode()) { |
183 | default: |
184 | llvm_unreachable("Unknown branch instruction?" ); |
185 | case AArch64::Bcc: |
186 | Target = LastInst->getOperand(i: 1).getMBB(); |
187 | Cond.push_back(Elt: LastInst->getOperand(i: 0)); |
188 | break; |
189 | case AArch64::CBZW: |
190 | case AArch64::CBZX: |
191 | case AArch64::CBNZW: |
192 | case AArch64::CBNZX: |
193 | Target = LastInst->getOperand(i: 1).getMBB(); |
194 | Cond.push_back(Elt: MachineOperand::CreateImm(Val: -1)); |
195 | Cond.push_back(Elt: MachineOperand::CreateImm(Val: LastInst->getOpcode())); |
196 | Cond.push_back(Elt: LastInst->getOperand(i: 0)); |
197 | break; |
198 | case AArch64::TBZW: |
199 | case AArch64::TBZX: |
200 | case AArch64::TBNZW: |
201 | case AArch64::TBNZX: |
202 | Target = LastInst->getOperand(i: 2).getMBB(); |
203 | Cond.push_back(Elt: MachineOperand::CreateImm(Val: -1)); |
204 | Cond.push_back(Elt: MachineOperand::CreateImm(Val: LastInst->getOpcode())); |
205 | Cond.push_back(Elt: LastInst->getOperand(i: 0)); |
206 | Cond.push_back(Elt: LastInst->getOperand(i: 1)); |
207 | } |
208 | } |
209 | |
210 | static unsigned getBranchDisplacementBits(unsigned Opc) { |
211 | switch (Opc) { |
212 | default: |
213 | llvm_unreachable("unexpected opcode!" ); |
214 | case AArch64::B: |
215 | return BDisplacementBits; |
216 | case AArch64::TBNZW: |
217 | case AArch64::TBZW: |
218 | case AArch64::TBNZX: |
219 | case AArch64::TBZX: |
220 | return TBZDisplacementBits; |
221 | case AArch64::CBNZW: |
222 | case AArch64::CBZW: |
223 | case AArch64::CBNZX: |
224 | case AArch64::CBZX: |
225 | return CBZDisplacementBits; |
226 | case AArch64::Bcc: |
227 | return BCCDisplacementBits; |
228 | } |
229 | } |
230 | |
231 | bool AArch64InstrInfo::isBranchOffsetInRange(unsigned BranchOp, |
232 | int64_t BrOffset) const { |
233 | unsigned Bits = getBranchDisplacementBits(Opc: BranchOp); |
234 | assert(Bits >= 3 && "max branch displacement must be enough to jump" |
235 | "over conditional branch expansion" ); |
236 | return isIntN(N: Bits, x: BrOffset / 4); |
237 | } |
238 | |
239 | MachineBasicBlock * |
240 | AArch64InstrInfo::getBranchDestBlock(const MachineInstr &MI) const { |
241 | switch (MI.getOpcode()) { |
242 | default: |
243 | llvm_unreachable("unexpected opcode!" ); |
244 | case AArch64::B: |
245 | return MI.getOperand(i: 0).getMBB(); |
246 | case AArch64::TBZW: |
247 | case AArch64::TBNZW: |
248 | case AArch64::TBZX: |
249 | case AArch64::TBNZX: |
250 | return MI.getOperand(i: 2).getMBB(); |
251 | case AArch64::CBZW: |
252 | case AArch64::CBNZW: |
253 | case AArch64::CBZX: |
254 | case AArch64::CBNZX: |
255 | case AArch64::Bcc: |
256 | return MI.getOperand(i: 1).getMBB(); |
257 | } |
258 | } |
259 | |
260 | void AArch64InstrInfo::insertIndirectBranch(MachineBasicBlock &MBB, |
261 | MachineBasicBlock &NewDestBB, |
262 | MachineBasicBlock &RestoreBB, |
263 | const DebugLoc &DL, |
264 | int64_t BrOffset, |
265 | RegScavenger *RS) const { |
266 | assert(RS && "RegScavenger required for long branching" ); |
267 | assert(MBB.empty() && |
268 | "new block should be inserted for expanding unconditional branch" ); |
269 | assert(MBB.pred_size() == 1); |
270 | assert(RestoreBB.empty() && |
271 | "restore block should be inserted for restoring clobbered registers" ); |
272 | |
273 | auto buildIndirectBranch = [&](Register Reg, MachineBasicBlock &DestBB) { |
274 | // Offsets outside of the signed 33-bit range are not supported for ADRP + |
275 | // ADD. |
276 | if (!isInt<33>(x: BrOffset)) |
277 | report_fatal_error( |
278 | reason: "Branch offsets outside of the signed 33-bit range not supported" ); |
279 | |
280 | BuildMI(BB&: MBB, I: MBB.end(), MIMD: DL, MCID: get(Opcode: AArch64::ADRP), DestReg: Reg) |
281 | .addSym(Sym: DestBB.getSymbol(), TargetFlags: AArch64II::MO_PAGE); |
282 | BuildMI(BB&: MBB, I: MBB.end(), MIMD: DL, MCID: get(Opcode: AArch64::ADDXri), DestReg: Reg) |
283 | .addReg(RegNo: Reg) |
284 | .addSym(Sym: DestBB.getSymbol(), TargetFlags: AArch64II::MO_PAGEOFF | AArch64II::MO_NC) |
285 | .addImm(Val: 0); |
286 | BuildMI(BB&: MBB, I: MBB.end(), MIMD: DL, MCID: get(Opcode: AArch64::BR)).addReg(RegNo: Reg); |
287 | }; |
288 | |
289 | RS->enterBasicBlockEnd(MBB); |
290 | // If X16 is unused, we can rely on the linker to insert a range extension |
291 | // thunk if NewDestBB is out of range of a single B instruction. |
292 | constexpr Register Reg = AArch64::X16; |
293 | if (!RS->isRegUsed(Reg)) { |
294 | insertUnconditionalBranch(MBB, DestBB: &NewDestBB, DL); |
295 | RS->setRegUsed(Reg); |
296 | return; |
297 | } |
298 | |
299 | // If there's a free register and it's worth inflating the code size, |
300 | // manually insert the indirect branch. |
301 | Register Scavenged = RS->FindUnusedReg(RC: &AArch64::GPR64RegClass); |
302 | if (Scavenged != AArch64::NoRegister && |
303 | MBB.getSectionID() == MBBSectionID::ColdSectionID) { |
304 | buildIndirectBranch(Scavenged, NewDestBB); |
305 | RS->setRegUsed(Reg: Scavenged); |
306 | return; |
307 | } |
308 | |
309 | // Note: Spilling X16 briefly moves the stack pointer, making it incompatible |
310 | // with red zones. |
311 | AArch64FunctionInfo *AFI = MBB.getParent()->getInfo<AArch64FunctionInfo>(); |
312 | if (!AFI || AFI->hasRedZone().value_or(u: true)) |
313 | report_fatal_error( |
314 | reason: "Unable to insert indirect branch inside function that has red zone" ); |
315 | |
316 | // Otherwise, spill X16 and defer range extension to the linker. |
317 | BuildMI(BB&: MBB, I: MBB.end(), MIMD: DL, MCID: get(Opcode: AArch64::STRXpre)) |
318 | .addReg(RegNo: AArch64::SP, flags: RegState::Define) |
319 | .addReg(RegNo: Reg) |
320 | .addReg(RegNo: AArch64::SP) |
321 | .addImm(Val: -16); |
322 | |
323 | BuildMI(BB&: MBB, I: MBB.end(), MIMD: DL, MCID: get(Opcode: AArch64::B)).addMBB(MBB: &RestoreBB); |
324 | |
325 | BuildMI(BB&: RestoreBB, I: RestoreBB.end(), MIMD: DL, MCID: get(Opcode: AArch64::LDRXpost)) |
326 | .addReg(RegNo: AArch64::SP, flags: RegState::Define) |
327 | .addReg(RegNo: Reg, flags: RegState::Define) |
328 | .addReg(RegNo: AArch64::SP) |
329 | .addImm(Val: 16); |
330 | } |
331 | |
332 | // Branch analysis. |
333 | bool AArch64InstrInfo::analyzeBranch(MachineBasicBlock &MBB, |
334 | MachineBasicBlock *&TBB, |
335 | MachineBasicBlock *&FBB, |
336 | SmallVectorImpl<MachineOperand> &Cond, |
337 | bool AllowModify) const { |
338 | // If the block has no terminators, it just falls into the block after it. |
339 | MachineBasicBlock::iterator I = MBB.getLastNonDebugInstr(); |
340 | if (I == MBB.end()) |
341 | return false; |
342 | |
343 | // Skip over SpeculationBarrierEndBB terminators |
344 | if (I->getOpcode() == AArch64::SpeculationBarrierISBDSBEndBB || |
345 | I->getOpcode() == AArch64::SpeculationBarrierSBEndBB) { |
346 | --I; |
347 | } |
348 | |
349 | if (!isUnpredicatedTerminator(MI: *I)) |
350 | return false; |
351 | |
352 | // Get the last instruction in the block. |
353 | MachineInstr *LastInst = &*I; |
354 | |
355 | // If there is only one terminator instruction, process it. |
356 | unsigned LastOpc = LastInst->getOpcode(); |
357 | if (I == MBB.begin() || !isUnpredicatedTerminator(MI: *--I)) { |
358 | if (isUncondBranchOpcode(Opc: LastOpc)) { |
359 | TBB = LastInst->getOperand(i: 0).getMBB(); |
360 | return false; |
361 | } |
362 | if (isCondBranchOpcode(Opc: LastOpc)) { |
363 | // Block ends with fall-through condbranch. |
364 | parseCondBranch(LastInst, Target&: TBB, Cond); |
365 | return false; |
366 | } |
367 | return true; // Can't handle indirect branch. |
368 | } |
369 | |
370 | // Get the instruction before it if it is a terminator. |
371 | MachineInstr *SecondLastInst = &*I; |
372 | unsigned SecondLastOpc = SecondLastInst->getOpcode(); |
373 | |
374 | // If AllowModify is true and the block ends with two or more unconditional |
375 | // branches, delete all but the first unconditional branch. |
376 | if (AllowModify && isUncondBranchOpcode(Opc: LastOpc)) { |
377 | while (isUncondBranchOpcode(Opc: SecondLastOpc)) { |
378 | LastInst->eraseFromParent(); |
379 | LastInst = SecondLastInst; |
380 | LastOpc = LastInst->getOpcode(); |
381 | if (I == MBB.begin() || !isUnpredicatedTerminator(MI: *--I)) { |
382 | // Return now the only terminator is an unconditional branch. |
383 | TBB = LastInst->getOperand(i: 0).getMBB(); |
384 | return false; |
385 | } |
386 | SecondLastInst = &*I; |
387 | SecondLastOpc = SecondLastInst->getOpcode(); |
388 | } |
389 | } |
390 | |
391 | // If we're allowed to modify and the block ends in a unconditional branch |
392 | // which could simply fallthrough, remove the branch. (Note: This case only |
393 | // matters when we can't understand the whole sequence, otherwise it's also |
394 | // handled by BranchFolding.cpp.) |
395 | if (AllowModify && isUncondBranchOpcode(Opc: LastOpc) && |
396 | MBB.isLayoutSuccessor(MBB: getBranchDestBlock(MI: *LastInst))) { |
397 | LastInst->eraseFromParent(); |
398 | LastInst = SecondLastInst; |
399 | LastOpc = LastInst->getOpcode(); |
400 | if (I == MBB.begin() || !isUnpredicatedTerminator(MI: *--I)) { |
401 | assert(!isUncondBranchOpcode(LastOpc) && |
402 | "unreachable unconditional branches removed above" ); |
403 | |
404 | if (isCondBranchOpcode(Opc: LastOpc)) { |
405 | // Block ends with fall-through condbranch. |
406 | parseCondBranch(LastInst, Target&: TBB, Cond); |
407 | return false; |
408 | } |
409 | return true; // Can't handle indirect branch. |
410 | } |
411 | SecondLastInst = &*I; |
412 | SecondLastOpc = SecondLastInst->getOpcode(); |
413 | } |
414 | |
415 | // If there are three terminators, we don't know what sort of block this is. |
416 | if (SecondLastInst && I != MBB.begin() && isUnpredicatedTerminator(MI: *--I)) |
417 | return true; |
418 | |
419 | // If the block ends with a B and a Bcc, handle it. |
420 | if (isCondBranchOpcode(Opc: SecondLastOpc) && isUncondBranchOpcode(Opc: LastOpc)) { |
421 | parseCondBranch(LastInst: SecondLastInst, Target&: TBB, Cond); |
422 | FBB = LastInst->getOperand(i: 0).getMBB(); |
423 | return false; |
424 | } |
425 | |
426 | // If the block ends with two unconditional branches, handle it. The second |
427 | // one is not executed, so remove it. |
428 | if (isUncondBranchOpcode(Opc: SecondLastOpc) && isUncondBranchOpcode(Opc: LastOpc)) { |
429 | TBB = SecondLastInst->getOperand(i: 0).getMBB(); |
430 | I = LastInst; |
431 | if (AllowModify) |
432 | I->eraseFromParent(); |
433 | return false; |
434 | } |
435 | |
436 | // ...likewise if it ends with an indirect branch followed by an unconditional |
437 | // branch. |
438 | if (isIndirectBranchOpcode(Opc: SecondLastOpc) && isUncondBranchOpcode(Opc: LastOpc)) { |
439 | I = LastInst; |
440 | if (AllowModify) |
441 | I->eraseFromParent(); |
442 | return true; |
443 | } |
444 | |
445 | // Otherwise, can't handle this. |
446 | return true; |
447 | } |
448 | |
449 | bool AArch64InstrInfo::analyzeBranchPredicate(MachineBasicBlock &MBB, |
450 | MachineBranchPredicate &MBP, |
451 | bool AllowModify) const { |
452 | // For the moment, handle only a block which ends with a cb(n)zx followed by |
453 | // a fallthrough. Why this? Because it is a common form. |
454 | // TODO: Should we handle b.cc? |
455 | |
456 | MachineBasicBlock::iterator I = MBB.getLastNonDebugInstr(); |
457 | if (I == MBB.end()) |
458 | return true; |
459 | |
460 | // Skip over SpeculationBarrierEndBB terminators |
461 | if (I->getOpcode() == AArch64::SpeculationBarrierISBDSBEndBB || |
462 | I->getOpcode() == AArch64::SpeculationBarrierSBEndBB) { |
463 | --I; |
464 | } |
465 | |
466 | if (!isUnpredicatedTerminator(MI: *I)) |
467 | return true; |
468 | |
469 | // Get the last instruction in the block. |
470 | MachineInstr *LastInst = &*I; |
471 | unsigned LastOpc = LastInst->getOpcode(); |
472 | if (!isCondBranchOpcode(Opc: LastOpc)) |
473 | return true; |
474 | |
475 | switch (LastOpc) { |
476 | default: |
477 | return true; |
478 | case AArch64::CBZW: |
479 | case AArch64::CBZX: |
480 | case AArch64::CBNZW: |
481 | case AArch64::CBNZX: |
482 | break; |
483 | }; |
484 | |
485 | MBP.TrueDest = LastInst->getOperand(i: 1).getMBB(); |
486 | assert(MBP.TrueDest && "expected!" ); |
487 | MBP.FalseDest = MBB.getNextNode(); |
488 | |
489 | MBP.ConditionDef = nullptr; |
490 | MBP.SingleUseCondition = false; |
491 | |
492 | MBP.LHS = LastInst->getOperand(i: 0); |
493 | MBP.RHS = MachineOperand::CreateImm(Val: 0); |
494 | MBP.Predicate = LastOpc == AArch64::CBNZX ? MachineBranchPredicate::PRED_NE |
495 | : MachineBranchPredicate::PRED_EQ; |
496 | return false; |
497 | } |
498 | |
499 | bool AArch64InstrInfo::reverseBranchCondition( |
500 | SmallVectorImpl<MachineOperand> &Cond) const { |
501 | if (Cond[0].getImm() != -1) { |
502 | // Regular Bcc |
503 | AArch64CC::CondCode CC = (AArch64CC::CondCode)(int)Cond[0].getImm(); |
504 | Cond[0].setImm(AArch64CC::getInvertedCondCode(Code: CC)); |
505 | } else { |
506 | // Folded compare-and-branch |
507 | switch (Cond[1].getImm()) { |
508 | default: |
509 | llvm_unreachable("Unknown conditional branch!" ); |
510 | case AArch64::CBZW: |
511 | Cond[1].setImm(AArch64::CBNZW); |
512 | break; |
513 | case AArch64::CBNZW: |
514 | Cond[1].setImm(AArch64::CBZW); |
515 | break; |
516 | case AArch64::CBZX: |
517 | Cond[1].setImm(AArch64::CBNZX); |
518 | break; |
519 | case AArch64::CBNZX: |
520 | Cond[1].setImm(AArch64::CBZX); |
521 | break; |
522 | case AArch64::TBZW: |
523 | Cond[1].setImm(AArch64::TBNZW); |
524 | break; |
525 | case AArch64::TBNZW: |
526 | Cond[1].setImm(AArch64::TBZW); |
527 | break; |
528 | case AArch64::TBZX: |
529 | Cond[1].setImm(AArch64::TBNZX); |
530 | break; |
531 | case AArch64::TBNZX: |
532 | Cond[1].setImm(AArch64::TBZX); |
533 | break; |
534 | } |
535 | } |
536 | |
537 | return false; |
538 | } |
539 | |
540 | unsigned AArch64InstrInfo::removeBranch(MachineBasicBlock &MBB, |
541 | int *BytesRemoved) const { |
542 | MachineBasicBlock::iterator I = MBB.getLastNonDebugInstr(); |
543 | if (I == MBB.end()) |
544 | return 0; |
545 | |
546 | if (!isUncondBranchOpcode(Opc: I->getOpcode()) && |
547 | !isCondBranchOpcode(Opc: I->getOpcode())) |
548 | return 0; |
549 | |
550 | // Remove the branch. |
551 | I->eraseFromParent(); |
552 | |
553 | I = MBB.end(); |
554 | |
555 | if (I == MBB.begin()) { |
556 | if (BytesRemoved) |
557 | *BytesRemoved = 4; |
558 | return 1; |
559 | } |
560 | --I; |
561 | if (!isCondBranchOpcode(Opc: I->getOpcode())) { |
562 | if (BytesRemoved) |
563 | *BytesRemoved = 4; |
564 | return 1; |
565 | } |
566 | |
567 | // Remove the branch. |
568 | I->eraseFromParent(); |
569 | if (BytesRemoved) |
570 | *BytesRemoved = 8; |
571 | |
572 | return 2; |
573 | } |
574 | |
575 | void AArch64InstrInfo::instantiateCondBranch( |
576 | MachineBasicBlock &MBB, const DebugLoc &DL, MachineBasicBlock *TBB, |
577 | ArrayRef<MachineOperand> Cond) const { |
578 | if (Cond[0].getImm() != -1) { |
579 | // Regular Bcc |
580 | BuildMI(BB: &MBB, MIMD: DL, MCID: get(Opcode: AArch64::Bcc)).addImm(Val: Cond[0].getImm()).addMBB(MBB: TBB); |
581 | } else { |
582 | // Folded compare-and-branch |
583 | // Note that we use addOperand instead of addReg to keep the flags. |
584 | const MachineInstrBuilder MIB = |
585 | BuildMI(BB: &MBB, MIMD: DL, MCID: get(Opcode: Cond[1].getImm())).add(MO: Cond[2]); |
586 | if (Cond.size() > 3) |
587 | MIB.addImm(Val: Cond[3].getImm()); |
588 | MIB.addMBB(MBB: TBB); |
589 | } |
590 | } |
591 | |
592 | unsigned AArch64InstrInfo::insertBranch( |
593 | MachineBasicBlock &MBB, MachineBasicBlock *TBB, MachineBasicBlock *FBB, |
594 | ArrayRef<MachineOperand> Cond, const DebugLoc &DL, int *BytesAdded) const { |
595 | // Shouldn't be a fall through. |
596 | assert(TBB && "insertBranch must not be told to insert a fallthrough" ); |
597 | |
598 | if (!FBB) { |
599 | if (Cond.empty()) // Unconditional branch? |
600 | BuildMI(BB: &MBB, MIMD: DL, MCID: get(Opcode: AArch64::B)).addMBB(MBB: TBB); |
601 | else |
602 | instantiateCondBranch(MBB, DL, TBB, Cond); |
603 | |
604 | if (BytesAdded) |
605 | *BytesAdded = 4; |
606 | |
607 | return 1; |
608 | } |
609 | |
610 | // Two-way conditional branch. |
611 | instantiateCondBranch(MBB, DL, TBB, Cond); |
612 | BuildMI(BB: &MBB, MIMD: DL, MCID: get(Opcode: AArch64::B)).addMBB(MBB: FBB); |
613 | |
614 | if (BytesAdded) |
615 | *BytesAdded = 8; |
616 | |
617 | return 2; |
618 | } |
619 | |
620 | // Find the original register that VReg is copied from. |
621 | static unsigned removeCopies(const MachineRegisterInfo &MRI, unsigned VReg) { |
622 | while (Register::isVirtualRegister(Reg: VReg)) { |
623 | const MachineInstr *DefMI = MRI.getVRegDef(Reg: VReg); |
624 | if (!DefMI->isFullCopy()) |
625 | return VReg; |
626 | VReg = DefMI->getOperand(i: 1).getReg(); |
627 | } |
628 | return VReg; |
629 | } |
630 | |
631 | // Determine if VReg is defined by an instruction that can be folded into a |
632 | // csel instruction. If so, return the folded opcode, and the replacement |
633 | // register. |
634 | static unsigned canFoldIntoCSel(const MachineRegisterInfo &MRI, unsigned VReg, |
635 | unsigned *NewVReg = nullptr) { |
636 | VReg = removeCopies(MRI, VReg); |
637 | if (!Register::isVirtualRegister(Reg: VReg)) |
638 | return 0; |
639 | |
640 | bool Is64Bit = AArch64::GPR64allRegClass.hasSubClassEq(RC: MRI.getRegClass(Reg: VReg)); |
641 | const MachineInstr *DefMI = MRI.getVRegDef(Reg: VReg); |
642 | unsigned Opc = 0; |
643 | unsigned SrcOpNum = 0; |
644 | switch (DefMI->getOpcode()) { |
645 | case AArch64::ADDSXri: |
646 | case AArch64::ADDSWri: |
647 | // if NZCV is used, do not fold. |
648 | if (DefMI->findRegisterDefOperandIdx(Reg: AArch64::NZCV, /*TRI=*/nullptr, |
649 | isDead: true) == -1) |
650 | return 0; |
651 | // fall-through to ADDXri and ADDWri. |
652 | [[fallthrough]]; |
653 | case AArch64::ADDXri: |
654 | case AArch64::ADDWri: |
655 | // add x, 1 -> csinc. |
656 | if (!DefMI->getOperand(i: 2).isImm() || DefMI->getOperand(i: 2).getImm() != 1 || |
657 | DefMI->getOperand(i: 3).getImm() != 0) |
658 | return 0; |
659 | SrcOpNum = 1; |
660 | Opc = Is64Bit ? AArch64::CSINCXr : AArch64::CSINCWr; |
661 | break; |
662 | |
663 | case AArch64::ORNXrr: |
664 | case AArch64::ORNWrr: { |
665 | // not x -> csinv, represented as orn dst, xzr, src. |
666 | unsigned ZReg = removeCopies(MRI, VReg: DefMI->getOperand(i: 1).getReg()); |
667 | if (ZReg != AArch64::XZR && ZReg != AArch64::WZR) |
668 | return 0; |
669 | SrcOpNum = 2; |
670 | Opc = Is64Bit ? AArch64::CSINVXr : AArch64::CSINVWr; |
671 | break; |
672 | } |
673 | |
674 | case AArch64::SUBSXrr: |
675 | case AArch64::SUBSWrr: |
676 | // if NZCV is used, do not fold. |
677 | if (DefMI->findRegisterDefOperandIdx(Reg: AArch64::NZCV, /*TRI=*/nullptr, |
678 | isDead: true) == -1) |
679 | return 0; |
680 | // fall-through to SUBXrr and SUBWrr. |
681 | [[fallthrough]]; |
682 | case AArch64::SUBXrr: |
683 | case AArch64::SUBWrr: { |
684 | // neg x -> csneg, represented as sub dst, xzr, src. |
685 | unsigned ZReg = removeCopies(MRI, VReg: DefMI->getOperand(i: 1).getReg()); |
686 | if (ZReg != AArch64::XZR && ZReg != AArch64::WZR) |
687 | return 0; |
688 | SrcOpNum = 2; |
689 | Opc = Is64Bit ? AArch64::CSNEGXr : AArch64::CSNEGWr; |
690 | break; |
691 | } |
692 | default: |
693 | return 0; |
694 | } |
695 | assert(Opc && SrcOpNum && "Missing parameters" ); |
696 | |
697 | if (NewVReg) |
698 | *NewVReg = DefMI->getOperand(i: SrcOpNum).getReg(); |
699 | return Opc; |
700 | } |
701 | |
702 | bool AArch64InstrInfo::canInsertSelect(const MachineBasicBlock &MBB, |
703 | ArrayRef<MachineOperand> Cond, |
704 | Register DstReg, Register TrueReg, |
705 | Register FalseReg, int &CondCycles, |
706 | int &TrueCycles, |
707 | int &FalseCycles) const { |
708 | // Check register classes. |
709 | const MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo(); |
710 | const TargetRegisterClass *RC = |
711 | RI.getCommonSubClass(A: MRI.getRegClass(Reg: TrueReg), B: MRI.getRegClass(Reg: FalseReg)); |
712 | if (!RC) |
713 | return false; |
714 | |
715 | // Also need to check the dest regclass, in case we're trying to optimize |
716 | // something like: |
717 | // %1(gpr) = PHI %2(fpr), bb1, %(fpr), bb2 |
718 | if (!RI.getCommonSubClass(A: RC, B: MRI.getRegClass(Reg: DstReg))) |
719 | return false; |
720 | |
721 | // Expanding cbz/tbz requires an extra cycle of latency on the condition. |
722 | unsigned = Cond.size() != 1; |
723 | |
724 | // GPRs are handled by csel. |
725 | // FIXME: Fold in x+1, -x, and ~x when applicable. |
726 | if (AArch64::GPR64allRegClass.hasSubClassEq(RC) || |
727 | AArch64::GPR32allRegClass.hasSubClassEq(RC)) { |
728 | // Single-cycle csel, csinc, csinv, and csneg. |
729 | CondCycles = 1 + ExtraCondLat; |
730 | TrueCycles = FalseCycles = 1; |
731 | if (canFoldIntoCSel(MRI, VReg: TrueReg)) |
732 | TrueCycles = 0; |
733 | else if (canFoldIntoCSel(MRI, VReg: FalseReg)) |
734 | FalseCycles = 0; |
735 | return true; |
736 | } |
737 | |
738 | // Scalar floating point is handled by fcsel. |
739 | // FIXME: Form fabs, fmin, and fmax when applicable. |
740 | if (AArch64::FPR64RegClass.hasSubClassEq(RC) || |
741 | AArch64::FPR32RegClass.hasSubClassEq(RC)) { |
742 | CondCycles = 5 + ExtraCondLat; |
743 | TrueCycles = FalseCycles = 2; |
744 | return true; |
745 | } |
746 | |
747 | // Can't do vectors. |
748 | return false; |
749 | } |
750 | |
751 | void AArch64InstrInfo::insertSelect(MachineBasicBlock &MBB, |
752 | MachineBasicBlock::iterator I, |
753 | const DebugLoc &DL, Register DstReg, |
754 | ArrayRef<MachineOperand> Cond, |
755 | Register TrueReg, Register FalseReg) const { |
756 | MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo(); |
757 | |
758 | // Parse the condition code, see parseCondBranch() above. |
759 | AArch64CC::CondCode CC; |
760 | switch (Cond.size()) { |
761 | default: |
762 | llvm_unreachable("Unknown condition opcode in Cond" ); |
763 | case 1: // b.cc |
764 | CC = AArch64CC::CondCode(Cond[0].getImm()); |
765 | break; |
766 | case 3: { // cbz/cbnz |
767 | // We must insert a compare against 0. |
768 | bool Is64Bit; |
769 | switch (Cond[1].getImm()) { |
770 | default: |
771 | llvm_unreachable("Unknown branch opcode in Cond" ); |
772 | case AArch64::CBZW: |
773 | Is64Bit = false; |
774 | CC = AArch64CC::EQ; |
775 | break; |
776 | case AArch64::CBZX: |
777 | Is64Bit = true; |
778 | CC = AArch64CC::EQ; |
779 | break; |
780 | case AArch64::CBNZW: |
781 | Is64Bit = false; |
782 | CC = AArch64CC::NE; |
783 | break; |
784 | case AArch64::CBNZX: |
785 | Is64Bit = true; |
786 | CC = AArch64CC::NE; |
787 | break; |
788 | } |
789 | Register SrcReg = Cond[2].getReg(); |
790 | if (Is64Bit) { |
791 | // cmp reg, #0 is actually subs xzr, reg, #0. |
792 | MRI.constrainRegClass(Reg: SrcReg, RC: &AArch64::GPR64spRegClass); |
793 | BuildMI(BB&: MBB, I, MIMD: DL, MCID: get(Opcode: AArch64::SUBSXri), DestReg: AArch64::XZR) |
794 | .addReg(RegNo: SrcReg) |
795 | .addImm(Val: 0) |
796 | .addImm(Val: 0); |
797 | } else { |
798 | MRI.constrainRegClass(Reg: SrcReg, RC: &AArch64::GPR32spRegClass); |
799 | BuildMI(BB&: MBB, I, MIMD: DL, MCID: get(Opcode: AArch64::SUBSWri), DestReg: AArch64::WZR) |
800 | .addReg(RegNo: SrcReg) |
801 | .addImm(Val: 0) |
802 | .addImm(Val: 0); |
803 | } |
804 | break; |
805 | } |
806 | case 4: { // tbz/tbnz |
807 | // We must insert a tst instruction. |
808 | switch (Cond[1].getImm()) { |
809 | default: |
810 | llvm_unreachable("Unknown branch opcode in Cond" ); |
811 | case AArch64::TBZW: |
812 | case AArch64::TBZX: |
813 | CC = AArch64CC::EQ; |
814 | break; |
815 | case AArch64::TBNZW: |
816 | case AArch64::TBNZX: |
817 | CC = AArch64CC::NE; |
818 | break; |
819 | } |
820 | // cmp reg, #foo is actually ands xzr, reg, #1<<foo. |
821 | if (Cond[1].getImm() == AArch64::TBZW || Cond[1].getImm() == AArch64::TBNZW) |
822 | BuildMI(BB&: MBB, I, MIMD: DL, MCID: get(Opcode: AArch64::ANDSWri), DestReg: AArch64::WZR) |
823 | .addReg(RegNo: Cond[2].getReg()) |
824 | .addImm( |
825 | Val: AArch64_AM::encodeLogicalImmediate(imm: 1ull << Cond[3].getImm(), regSize: 32)); |
826 | else |
827 | BuildMI(BB&: MBB, I, MIMD: DL, MCID: get(Opcode: AArch64::ANDSXri), DestReg: AArch64::XZR) |
828 | .addReg(RegNo: Cond[2].getReg()) |
829 | .addImm( |
830 | Val: AArch64_AM::encodeLogicalImmediate(imm: 1ull << Cond[3].getImm(), regSize: 64)); |
831 | break; |
832 | } |
833 | } |
834 | |
835 | unsigned Opc = 0; |
836 | const TargetRegisterClass *RC = nullptr; |
837 | bool TryFold = false; |
838 | if (MRI.constrainRegClass(Reg: DstReg, RC: &AArch64::GPR64RegClass)) { |
839 | RC = &AArch64::GPR64RegClass; |
840 | Opc = AArch64::CSELXr; |
841 | TryFold = true; |
842 | } else if (MRI.constrainRegClass(Reg: DstReg, RC: &AArch64::GPR32RegClass)) { |
843 | RC = &AArch64::GPR32RegClass; |
844 | Opc = AArch64::CSELWr; |
845 | TryFold = true; |
846 | } else if (MRI.constrainRegClass(Reg: DstReg, RC: &AArch64::FPR64RegClass)) { |
847 | RC = &AArch64::FPR64RegClass; |
848 | Opc = AArch64::FCSELDrrr; |
849 | } else if (MRI.constrainRegClass(Reg: DstReg, RC: &AArch64::FPR32RegClass)) { |
850 | RC = &AArch64::FPR32RegClass; |
851 | Opc = AArch64::FCSELSrrr; |
852 | } |
853 | assert(RC && "Unsupported regclass" ); |
854 | |
855 | // Try folding simple instructions into the csel. |
856 | if (TryFold) { |
857 | unsigned NewVReg = 0; |
858 | unsigned FoldedOpc = canFoldIntoCSel(MRI, VReg: TrueReg, NewVReg: &NewVReg); |
859 | if (FoldedOpc) { |
860 | // The folded opcodes csinc, csinc and csneg apply the operation to |
861 | // FalseReg, so we need to invert the condition. |
862 | CC = AArch64CC::getInvertedCondCode(Code: CC); |
863 | TrueReg = FalseReg; |
864 | } else |
865 | FoldedOpc = canFoldIntoCSel(MRI, VReg: FalseReg, NewVReg: &NewVReg); |
866 | |
867 | // Fold the operation. Leave any dead instructions for DCE to clean up. |
868 | if (FoldedOpc) { |
869 | FalseReg = NewVReg; |
870 | Opc = FoldedOpc; |
871 | // The extends the live range of NewVReg. |
872 | MRI.clearKillFlags(Reg: NewVReg); |
873 | } |
874 | } |
875 | |
876 | // Pull all virtual register into the appropriate class. |
877 | MRI.constrainRegClass(Reg: TrueReg, RC); |
878 | MRI.constrainRegClass(Reg: FalseReg, RC); |
879 | |
880 | // Insert the csel. |
881 | BuildMI(BB&: MBB, I, MIMD: DL, MCID: get(Opcode: Opc), DestReg: DstReg) |
882 | .addReg(RegNo: TrueReg) |
883 | .addReg(RegNo: FalseReg) |
884 | .addImm(Val: CC); |
885 | } |
886 | |
887 | // Return true if Imm can be loaded into a register by a "cheap" sequence of |
888 | // instructions. For now, "cheap" means at most two instructions. |
889 | static bool isCheapImmediate(const MachineInstr &MI, unsigned BitSize) { |
890 | if (BitSize == 32) |
891 | return true; |
892 | |
893 | assert(BitSize == 64 && "Only bit sizes of 32 or 64 allowed" ); |
894 | uint64_t Imm = static_cast<uint64_t>(MI.getOperand(i: 1).getImm()); |
895 | SmallVector<AArch64_IMM::ImmInsnModel, 4> Is; |
896 | AArch64_IMM::expandMOVImm(Imm, BitSize, Insn&: Is); |
897 | |
898 | return Is.size() <= 2; |
899 | } |
900 | |
901 | // FIXME: this implementation should be micro-architecture dependent, so a |
902 | // micro-architecture target hook should be introduced here in future. |
903 | bool AArch64InstrInfo::isAsCheapAsAMove(const MachineInstr &MI) const { |
904 | if (Subtarget.hasExynosCheapAsMoveHandling()) { |
905 | if (isExynosCheapAsMove(MI)) |
906 | return true; |
907 | return MI.isAsCheapAsAMove(); |
908 | } |
909 | |
910 | switch (MI.getOpcode()) { |
911 | default: |
912 | return MI.isAsCheapAsAMove(); |
913 | |
914 | case AArch64::ADDWrs: |
915 | case AArch64::ADDXrs: |
916 | case AArch64::SUBWrs: |
917 | case AArch64::SUBXrs: |
918 | return Subtarget.hasALULSLFast() && MI.getOperand(i: 3).getImm() <= 4; |
919 | |
920 | // If MOVi32imm or MOVi64imm can be expanded into ORRWri or |
921 | // ORRXri, it is as cheap as MOV. |
922 | // Likewise if it can be expanded to MOVZ/MOVN/MOVK. |
923 | case AArch64::MOVi32imm: |
924 | return isCheapImmediate(MI, BitSize: 32); |
925 | case AArch64::MOVi64imm: |
926 | return isCheapImmediate(MI, BitSize: 64); |
927 | } |
928 | } |
929 | |
930 | bool AArch64InstrInfo::isFalkorShiftExtFast(const MachineInstr &MI) { |
931 | switch (MI.getOpcode()) { |
932 | default: |
933 | return false; |
934 | |
935 | case AArch64::ADDWrs: |
936 | case AArch64::ADDXrs: |
937 | case AArch64::ADDSWrs: |
938 | case AArch64::ADDSXrs: { |
939 | unsigned Imm = MI.getOperand(i: 3).getImm(); |
940 | unsigned ShiftVal = AArch64_AM::getShiftValue(Imm); |
941 | if (ShiftVal == 0) |
942 | return true; |
943 | return AArch64_AM::getShiftType(Imm) == AArch64_AM::LSL && ShiftVal <= 5; |
944 | } |
945 | |
946 | case AArch64::ADDWrx: |
947 | case AArch64::ADDXrx: |
948 | case AArch64::ADDXrx64: |
949 | case AArch64::ADDSWrx: |
950 | case AArch64::ADDSXrx: |
951 | case AArch64::ADDSXrx64: { |
952 | unsigned Imm = MI.getOperand(i: 3).getImm(); |
953 | switch (AArch64_AM::getArithExtendType(Imm)) { |
954 | default: |
955 | return false; |
956 | case AArch64_AM::UXTB: |
957 | case AArch64_AM::UXTH: |
958 | case AArch64_AM::UXTW: |
959 | case AArch64_AM::UXTX: |
960 | return AArch64_AM::getArithShiftValue(Imm) <= 4; |
961 | } |
962 | } |
963 | |
964 | case AArch64::SUBWrs: |
965 | case AArch64::SUBSWrs: { |
966 | unsigned Imm = MI.getOperand(i: 3).getImm(); |
967 | unsigned ShiftVal = AArch64_AM::getShiftValue(Imm); |
968 | return ShiftVal == 0 || |
969 | (AArch64_AM::getShiftType(Imm) == AArch64_AM::ASR && ShiftVal == 31); |
970 | } |
971 | |
972 | case AArch64::SUBXrs: |
973 | case AArch64::SUBSXrs: { |
974 | unsigned Imm = MI.getOperand(i: 3).getImm(); |
975 | unsigned ShiftVal = AArch64_AM::getShiftValue(Imm); |
976 | return ShiftVal == 0 || |
977 | (AArch64_AM::getShiftType(Imm) == AArch64_AM::ASR && ShiftVal == 63); |
978 | } |
979 | |
980 | case AArch64::SUBWrx: |
981 | case AArch64::SUBXrx: |
982 | case AArch64::SUBXrx64: |
983 | case AArch64::SUBSWrx: |
984 | case AArch64::SUBSXrx: |
985 | case AArch64::SUBSXrx64: { |
986 | unsigned Imm = MI.getOperand(i: 3).getImm(); |
987 | switch (AArch64_AM::getArithExtendType(Imm)) { |
988 | default: |
989 | return false; |
990 | case AArch64_AM::UXTB: |
991 | case AArch64_AM::UXTH: |
992 | case AArch64_AM::UXTW: |
993 | case AArch64_AM::UXTX: |
994 | return AArch64_AM::getArithShiftValue(Imm) == 0; |
995 | } |
996 | } |
997 | |
998 | case AArch64::LDRBBroW: |
999 | case AArch64::LDRBBroX: |
1000 | case AArch64::LDRBroW: |
1001 | case AArch64::LDRBroX: |
1002 | case AArch64::LDRDroW: |
1003 | case AArch64::LDRDroX: |
1004 | case AArch64::LDRHHroW: |
1005 | case AArch64::LDRHHroX: |
1006 | case AArch64::LDRHroW: |
1007 | case AArch64::LDRHroX: |
1008 | case AArch64::LDRQroW: |
1009 | case AArch64::LDRQroX: |
1010 | case AArch64::LDRSBWroW: |
1011 | case AArch64::LDRSBWroX: |
1012 | case AArch64::LDRSBXroW: |
1013 | case AArch64::LDRSBXroX: |
1014 | case AArch64::LDRSHWroW: |
1015 | case AArch64::LDRSHWroX: |
1016 | case AArch64::LDRSHXroW: |
1017 | case AArch64::LDRSHXroX: |
1018 | case AArch64::LDRSWroW: |
1019 | case AArch64::LDRSWroX: |
1020 | case AArch64::LDRSroW: |
1021 | case AArch64::LDRSroX: |
1022 | case AArch64::LDRWroW: |
1023 | case AArch64::LDRWroX: |
1024 | case AArch64::LDRXroW: |
1025 | case AArch64::LDRXroX: |
1026 | case AArch64::PRFMroW: |
1027 | case AArch64::PRFMroX: |
1028 | case AArch64::STRBBroW: |
1029 | case AArch64::STRBBroX: |
1030 | case AArch64::STRBroW: |
1031 | case AArch64::STRBroX: |
1032 | case AArch64::STRDroW: |
1033 | case AArch64::STRDroX: |
1034 | case AArch64::STRHHroW: |
1035 | case AArch64::STRHHroX: |
1036 | case AArch64::STRHroW: |
1037 | case AArch64::STRHroX: |
1038 | case AArch64::STRQroW: |
1039 | case AArch64::STRQroX: |
1040 | case AArch64::STRSroW: |
1041 | case AArch64::STRSroX: |
1042 | case AArch64::STRWroW: |
1043 | case AArch64::STRWroX: |
1044 | case AArch64::STRXroW: |
1045 | case AArch64::STRXroX: { |
1046 | unsigned IsSigned = MI.getOperand(i: 3).getImm(); |
1047 | return !IsSigned; |
1048 | } |
1049 | } |
1050 | } |
1051 | |
1052 | bool AArch64InstrInfo::isSEHInstruction(const MachineInstr &MI) { |
1053 | unsigned Opc = MI.getOpcode(); |
1054 | switch (Opc) { |
1055 | default: |
1056 | return false; |
1057 | case AArch64::SEH_StackAlloc: |
1058 | case AArch64::SEH_SaveFPLR: |
1059 | case AArch64::SEH_SaveFPLR_X: |
1060 | case AArch64::SEH_SaveReg: |
1061 | case AArch64::SEH_SaveReg_X: |
1062 | case AArch64::SEH_SaveRegP: |
1063 | case AArch64::SEH_SaveRegP_X: |
1064 | case AArch64::SEH_SaveFReg: |
1065 | case AArch64::SEH_SaveFReg_X: |
1066 | case AArch64::SEH_SaveFRegP: |
1067 | case AArch64::SEH_SaveFRegP_X: |
1068 | case AArch64::SEH_SetFP: |
1069 | case AArch64::SEH_AddFP: |
1070 | case AArch64::SEH_Nop: |
1071 | case AArch64::SEH_PrologEnd: |
1072 | case AArch64::SEH_EpilogStart: |
1073 | case AArch64::SEH_EpilogEnd: |
1074 | case AArch64::SEH_PACSignLR: |
1075 | case AArch64::SEH_SaveAnyRegQP: |
1076 | case AArch64::SEH_SaveAnyRegQPX: |
1077 | return true; |
1078 | } |
1079 | } |
1080 | |
1081 | bool AArch64InstrInfo::isCoalescableExtInstr(const MachineInstr &MI, |
1082 | Register &SrcReg, Register &DstReg, |
1083 | unsigned &SubIdx) const { |
1084 | switch (MI.getOpcode()) { |
1085 | default: |
1086 | return false; |
1087 | case AArch64::SBFMXri: // aka sxtw |
1088 | case AArch64::UBFMXri: // aka uxtw |
1089 | // Check for the 32 -> 64 bit extension case, these instructions can do |
1090 | // much more. |
1091 | if (MI.getOperand(i: 2).getImm() != 0 || MI.getOperand(i: 3).getImm() != 31) |
1092 | return false; |
1093 | // This is a signed or unsigned 32 -> 64 bit extension. |
1094 | SrcReg = MI.getOperand(i: 1).getReg(); |
1095 | DstReg = MI.getOperand(i: 0).getReg(); |
1096 | SubIdx = AArch64::sub_32; |
1097 | return true; |
1098 | } |
1099 | } |
1100 | |
1101 | bool AArch64InstrInfo::areMemAccessesTriviallyDisjoint( |
1102 | const MachineInstr &MIa, const MachineInstr &MIb) const { |
1103 | const TargetRegisterInfo *TRI = &getRegisterInfo(); |
1104 | const MachineOperand *BaseOpA = nullptr, *BaseOpB = nullptr; |
1105 | int64_t OffsetA = 0, OffsetB = 0; |
1106 | TypeSize WidthA(0, false), WidthB(0, false); |
1107 | bool OffsetAIsScalable = false, OffsetBIsScalable = false; |
1108 | |
1109 | assert(MIa.mayLoadOrStore() && "MIa must be a load or store." ); |
1110 | assert(MIb.mayLoadOrStore() && "MIb must be a load or store." ); |
1111 | |
1112 | if (MIa.hasUnmodeledSideEffects() || MIb.hasUnmodeledSideEffects() || |
1113 | MIa.hasOrderedMemoryRef() || MIb.hasOrderedMemoryRef()) |
1114 | return false; |
1115 | |
1116 | // Retrieve the base, offset from the base and width. Width |
1117 | // is the size of memory that is being loaded/stored (e.g. 1, 2, 4, 8). If |
1118 | // base are identical, and the offset of a lower memory access + |
1119 | // the width doesn't overlap the offset of a higher memory access, |
1120 | // then the memory accesses are different. |
1121 | // If OffsetAIsScalable and OffsetBIsScalable are both true, they |
1122 | // are assumed to have the same scale (vscale). |
1123 | if (getMemOperandWithOffsetWidth(MI: MIa, BaseOp&: BaseOpA, Offset&: OffsetA, OffsetIsScalable&: OffsetAIsScalable, |
1124 | Width&: WidthA, TRI) && |
1125 | getMemOperandWithOffsetWidth(MI: MIb, BaseOp&: BaseOpB, Offset&: OffsetB, OffsetIsScalable&: OffsetBIsScalable, |
1126 | Width&: WidthB, TRI)) { |
1127 | if (BaseOpA->isIdenticalTo(Other: *BaseOpB) && |
1128 | OffsetAIsScalable == OffsetBIsScalable) { |
1129 | int LowOffset = OffsetA < OffsetB ? OffsetA : OffsetB; |
1130 | int HighOffset = OffsetA < OffsetB ? OffsetB : OffsetA; |
1131 | TypeSize LowWidth = (LowOffset == OffsetA) ? WidthA : WidthB; |
1132 | if (LowWidth.isScalable() == OffsetAIsScalable && |
1133 | LowOffset + (int)LowWidth.getKnownMinValue() <= HighOffset) |
1134 | return true; |
1135 | } |
1136 | } |
1137 | return false; |
1138 | } |
1139 | |
1140 | bool AArch64InstrInfo::isSchedulingBoundary(const MachineInstr &MI, |
1141 | const MachineBasicBlock *MBB, |
1142 | const MachineFunction &MF) const { |
1143 | if (TargetInstrInfo::isSchedulingBoundary(MI, MBB, MF)) |
1144 | return true; |
1145 | |
1146 | // Do not move an instruction that can be recognized as a branch target. |
1147 | if (hasBTISemantics(MI)) |
1148 | return true; |
1149 | |
1150 | switch (MI.getOpcode()) { |
1151 | case AArch64::HINT: |
1152 | // CSDB hints are scheduling barriers. |
1153 | if (MI.getOperand(i: 0).getImm() == 0x14) |
1154 | return true; |
1155 | break; |
1156 | case AArch64::DSB: |
1157 | case AArch64::ISB: |
1158 | // DSB and ISB also are scheduling barriers. |
1159 | return true; |
1160 | case AArch64::MSRpstatesvcrImm1: |
1161 | // SMSTART and SMSTOP are also scheduling barriers. |
1162 | return true; |
1163 | default:; |
1164 | } |
1165 | if (isSEHInstruction(MI)) |
1166 | return true; |
1167 | auto Next = std::next(x: MI.getIterator()); |
1168 | return Next != MBB->end() && Next->isCFIInstruction(); |
1169 | } |
1170 | |
1171 | /// analyzeCompare - For a comparison instruction, return the source registers |
1172 | /// in SrcReg and SrcReg2, and the value it compares against in CmpValue. |
1173 | /// Return true if the comparison instruction can be analyzed. |
1174 | bool AArch64InstrInfo::analyzeCompare(const MachineInstr &MI, Register &SrcReg, |
1175 | Register &SrcReg2, int64_t &CmpMask, |
1176 | int64_t &CmpValue) const { |
1177 | // The first operand can be a frame index where we'd normally expect a |
1178 | // register. |
1179 | assert(MI.getNumOperands() >= 2 && "All AArch64 cmps should have 2 operands" ); |
1180 | if (!MI.getOperand(i: 1).isReg()) |
1181 | return false; |
1182 | |
1183 | switch (MI.getOpcode()) { |
1184 | default: |
1185 | break; |
1186 | case AArch64::PTEST_PP: |
1187 | case AArch64::PTEST_PP_ANY: |
1188 | SrcReg = MI.getOperand(i: 0).getReg(); |
1189 | SrcReg2 = MI.getOperand(i: 1).getReg(); |
1190 | // Not sure about the mask and value for now... |
1191 | CmpMask = ~0; |
1192 | CmpValue = 0; |
1193 | return true; |
1194 | case AArch64::SUBSWrr: |
1195 | case AArch64::SUBSWrs: |
1196 | case AArch64::SUBSWrx: |
1197 | case AArch64::SUBSXrr: |
1198 | case AArch64::SUBSXrs: |
1199 | case AArch64::SUBSXrx: |
1200 | case AArch64::ADDSWrr: |
1201 | case AArch64::ADDSWrs: |
1202 | case AArch64::ADDSWrx: |
1203 | case AArch64::ADDSXrr: |
1204 | case AArch64::ADDSXrs: |
1205 | case AArch64::ADDSXrx: |
1206 | // Replace SUBSWrr with SUBWrr if NZCV is not used. |
1207 | SrcReg = MI.getOperand(i: 1).getReg(); |
1208 | SrcReg2 = MI.getOperand(i: 2).getReg(); |
1209 | CmpMask = ~0; |
1210 | CmpValue = 0; |
1211 | return true; |
1212 | case AArch64::SUBSWri: |
1213 | case AArch64::ADDSWri: |
1214 | case AArch64::SUBSXri: |
1215 | case AArch64::ADDSXri: |
1216 | SrcReg = MI.getOperand(i: 1).getReg(); |
1217 | SrcReg2 = 0; |
1218 | CmpMask = ~0; |
1219 | CmpValue = MI.getOperand(i: 2).getImm(); |
1220 | return true; |
1221 | case AArch64::ANDSWri: |
1222 | case AArch64::ANDSXri: |
1223 | // ANDS does not use the same encoding scheme as the others xxxS |
1224 | // instructions. |
1225 | SrcReg = MI.getOperand(i: 1).getReg(); |
1226 | SrcReg2 = 0; |
1227 | CmpMask = ~0; |
1228 | CmpValue = AArch64_AM::decodeLogicalImmediate( |
1229 | val: MI.getOperand(i: 2).getImm(), |
1230 | regSize: MI.getOpcode() == AArch64::ANDSWri ? 32 : 64); |
1231 | return true; |
1232 | } |
1233 | |
1234 | return false; |
1235 | } |
1236 | |
1237 | static bool UpdateOperandRegClass(MachineInstr &Instr) { |
1238 | MachineBasicBlock *MBB = Instr.getParent(); |
1239 | assert(MBB && "Can't get MachineBasicBlock here" ); |
1240 | MachineFunction *MF = MBB->getParent(); |
1241 | assert(MF && "Can't get MachineFunction here" ); |
1242 | const TargetInstrInfo *TII = MF->getSubtarget().getInstrInfo(); |
1243 | const TargetRegisterInfo *TRI = MF->getSubtarget().getRegisterInfo(); |
1244 | MachineRegisterInfo *MRI = &MF->getRegInfo(); |
1245 | |
1246 | for (unsigned OpIdx = 0, EndIdx = Instr.getNumOperands(); OpIdx < EndIdx; |
1247 | ++OpIdx) { |
1248 | MachineOperand &MO = Instr.getOperand(i: OpIdx); |
1249 | const TargetRegisterClass *OpRegCstraints = |
1250 | Instr.getRegClassConstraint(OpIdx, TII, TRI); |
1251 | |
1252 | // If there's no constraint, there's nothing to do. |
1253 | if (!OpRegCstraints) |
1254 | continue; |
1255 | // If the operand is a frame index, there's nothing to do here. |
1256 | // A frame index operand will resolve correctly during PEI. |
1257 | if (MO.isFI()) |
1258 | continue; |
1259 | |
1260 | assert(MO.isReg() && |
1261 | "Operand has register constraints without being a register!" ); |
1262 | |
1263 | Register Reg = MO.getReg(); |
1264 | if (Reg.isPhysical()) { |
1265 | if (!OpRegCstraints->contains(Reg)) |
1266 | return false; |
1267 | } else if (!OpRegCstraints->hasSubClassEq(RC: MRI->getRegClass(Reg)) && |
1268 | !MRI->constrainRegClass(Reg, RC: OpRegCstraints)) |
1269 | return false; |
1270 | } |
1271 | |
1272 | return true; |
1273 | } |
1274 | |
1275 | /// Return the opcode that does not set flags when possible - otherwise |
1276 | /// return the original opcode. The caller is responsible to do the actual |
1277 | /// substitution and legality checking. |
1278 | static unsigned convertToNonFlagSettingOpc(const MachineInstr &MI) { |
1279 | // Don't convert all compare instructions, because for some the zero register |
1280 | // encoding becomes the sp register. |
1281 | bool MIDefinesZeroReg = false; |
1282 | if (MI.definesRegister(Reg: AArch64::WZR, /*TRI=*/nullptr) || |
1283 | MI.definesRegister(Reg: AArch64::XZR, /*TRI=*/nullptr)) |
1284 | MIDefinesZeroReg = true; |
1285 | |
1286 | switch (MI.getOpcode()) { |
1287 | default: |
1288 | return MI.getOpcode(); |
1289 | case AArch64::ADDSWrr: |
1290 | return AArch64::ADDWrr; |
1291 | case AArch64::ADDSWri: |
1292 | return MIDefinesZeroReg ? AArch64::ADDSWri : AArch64::ADDWri; |
1293 | case AArch64::ADDSWrs: |
1294 | return MIDefinesZeroReg ? AArch64::ADDSWrs : AArch64::ADDWrs; |
1295 | case AArch64::ADDSWrx: |
1296 | return AArch64::ADDWrx; |
1297 | case AArch64::ADDSXrr: |
1298 | return AArch64::ADDXrr; |
1299 | case AArch64::ADDSXri: |
1300 | return MIDefinesZeroReg ? AArch64::ADDSXri : AArch64::ADDXri; |
1301 | case AArch64::ADDSXrs: |
1302 | return MIDefinesZeroReg ? AArch64::ADDSXrs : AArch64::ADDXrs; |
1303 | case AArch64::ADDSXrx: |
1304 | return AArch64::ADDXrx; |
1305 | case AArch64::SUBSWrr: |
1306 | return AArch64::SUBWrr; |
1307 | case AArch64::SUBSWri: |
1308 | return MIDefinesZeroReg ? AArch64::SUBSWri : AArch64::SUBWri; |
1309 | case AArch64::SUBSWrs: |
1310 | return MIDefinesZeroReg ? AArch64::SUBSWrs : AArch64::SUBWrs; |
1311 | case AArch64::SUBSWrx: |
1312 | return AArch64::SUBWrx; |
1313 | case AArch64::SUBSXrr: |
1314 | return AArch64::SUBXrr; |
1315 | case AArch64::SUBSXri: |
1316 | return MIDefinesZeroReg ? AArch64::SUBSXri : AArch64::SUBXri; |
1317 | case AArch64::SUBSXrs: |
1318 | return MIDefinesZeroReg ? AArch64::SUBSXrs : AArch64::SUBXrs; |
1319 | case AArch64::SUBSXrx: |
1320 | return AArch64::SUBXrx; |
1321 | } |
1322 | } |
1323 | |
1324 | enum AccessKind { AK_Write = 0x01, AK_Read = 0x10, AK_All = 0x11 }; |
1325 | |
1326 | /// True when condition flags are accessed (either by writing or reading) |
1327 | /// on the instruction trace starting at From and ending at To. |
1328 | /// |
1329 | /// Note: If From and To are from different blocks it's assumed CC are accessed |
1330 | /// on the path. |
1331 | static bool areCFlagsAccessedBetweenInstrs( |
1332 | MachineBasicBlock::iterator From, MachineBasicBlock::iterator To, |
1333 | const TargetRegisterInfo *TRI, const AccessKind AccessToCheck = AK_All) { |
1334 | // Early exit if To is at the beginning of the BB. |
1335 | if (To == To->getParent()->begin()) |
1336 | return true; |
1337 | |
1338 | // Check whether the instructions are in the same basic block |
1339 | // If not, assume the condition flags might get modified somewhere. |
1340 | if (To->getParent() != From->getParent()) |
1341 | return true; |
1342 | |
1343 | // From must be above To. |
1344 | assert(std::any_of( |
1345 | ++To.getReverse(), To->getParent()->rend(), |
1346 | [From](MachineInstr &MI) { return MI.getIterator() == From; })); |
1347 | |
1348 | // We iterate backward starting at \p To until we hit \p From. |
1349 | for (const MachineInstr &Instr : |
1350 | instructionsWithoutDebug(It: ++To.getReverse(), End: From.getReverse())) { |
1351 | if (((AccessToCheck & AK_Write) && |
1352 | Instr.modifiesRegister(Reg: AArch64::NZCV, TRI)) || |
1353 | ((AccessToCheck & AK_Read) && Instr.readsRegister(Reg: AArch64::NZCV, TRI))) |
1354 | return true; |
1355 | } |
1356 | return false; |
1357 | } |
1358 | |
1359 | std::optional<unsigned> |
1360 | AArch64InstrInfo::canRemovePTestInstr(MachineInstr *PTest, MachineInstr *Mask, |
1361 | MachineInstr *Pred, |
1362 | const MachineRegisterInfo *MRI) const { |
1363 | unsigned MaskOpcode = Mask->getOpcode(); |
1364 | unsigned PredOpcode = Pred->getOpcode(); |
1365 | bool PredIsPTestLike = isPTestLikeOpcode(Opc: PredOpcode); |
1366 | bool PredIsWhileLike = isWhileOpcode(Opc: PredOpcode); |
1367 | |
1368 | if (PredIsWhileLike) { |
1369 | // For PTEST(PG, PG), PTEST is redundant when PG is the result of a WHILEcc |
1370 | // instruction and the condition is "any" since WHILcc does an implicit |
1371 | // PTEST(ALL, PG) check and PG is always a subset of ALL. |
1372 | if ((Mask == Pred) && PTest->getOpcode() == AArch64::PTEST_PP_ANY) |
1373 | return PredOpcode; |
1374 | |
1375 | // For PTEST(PTRUE_ALL, WHILE), if the element size matches, the PTEST is |
1376 | // redundant since WHILE performs an implicit PTEST with an all active |
1377 | // mask. |
1378 | if (isPTrueOpcode(Opc: MaskOpcode) && Mask->getOperand(i: 1).getImm() == 31 && |
1379 | getElementSizeForOpcode(Opc: MaskOpcode) == |
1380 | getElementSizeForOpcode(Opc: PredOpcode)) |
1381 | return PredOpcode; |
1382 | |
1383 | return {}; |
1384 | } |
1385 | |
1386 | if (PredIsPTestLike) { |
1387 | // For PTEST(PG, PG), PTEST is redundant when PG is the result of an |
1388 | // instruction that sets the flags as PTEST would and the condition is |
1389 | // "any" since PG is always a subset of the governing predicate of the |
1390 | // ptest-like instruction. |
1391 | if ((Mask == Pred) && PTest->getOpcode() == AArch64::PTEST_PP_ANY) |
1392 | return PredOpcode; |
1393 | |
1394 | // For PTEST(PTRUE_ALL, PTEST_LIKE), the PTEST is redundant if the |
1395 | // the element size matches and either the PTEST_LIKE instruction uses |
1396 | // the same all active mask or the condition is "any". |
1397 | if (isPTrueOpcode(Opc: MaskOpcode) && Mask->getOperand(i: 1).getImm() == 31 && |
1398 | getElementSizeForOpcode(Opc: MaskOpcode) == |
1399 | getElementSizeForOpcode(Opc: PredOpcode)) { |
1400 | auto PTestLikeMask = MRI->getUniqueVRegDef(Reg: Pred->getOperand(i: 1).getReg()); |
1401 | if (Mask == PTestLikeMask || PTest->getOpcode() == AArch64::PTEST_PP_ANY) |
1402 | return PredOpcode; |
1403 | } |
1404 | |
1405 | // For PTEST(PG, PTEST_LIKE(PG, ...)), the PTEST is redundant since the |
1406 | // flags are set based on the same mask 'PG', but PTEST_LIKE must operate |
1407 | // on 8-bit predicates like the PTEST. Otherwise, for instructions like |
1408 | // compare that also support 16/32/64-bit predicates, the implicit PTEST |
1409 | // performed by the compare could consider fewer lanes for these element |
1410 | // sizes. |
1411 | // |
1412 | // For example, consider |
1413 | // |
1414 | // ptrue p0.b ; P0=1111-1111-1111-1111 |
1415 | // index z0.s, #0, #1 ; Z0=<0,1,2,3> |
1416 | // index z1.s, #1, #1 ; Z1=<1,2,3,4> |
1417 | // cmphi p1.s, p0/z, z1.s, z0.s ; P1=0001-0001-0001-0001 |
1418 | // ; ^ last active |
1419 | // ptest p0, p1.b ; P1=0001-0001-0001-0001 |
1420 | // ; ^ last active |
1421 | // |
1422 | // where the compare generates a canonical all active 32-bit predicate |
1423 | // (equivalent to 'ptrue p1.s, all'). The implicit PTEST sets the last |
1424 | // active flag, whereas the PTEST instruction with the same mask doesn't. |
1425 | // For PTEST_ANY this doesn't apply as the flags in this case would be |
1426 | // identical regardless of element size. |
1427 | auto PTestLikeMask = MRI->getUniqueVRegDef(Reg: Pred->getOperand(i: 1).getReg()); |
1428 | uint64_t PredElementSize = getElementSizeForOpcode(Opc: PredOpcode); |
1429 | if (Mask == PTestLikeMask && (PredElementSize == AArch64::ElementSizeB || |
1430 | PTest->getOpcode() == AArch64::PTEST_PP_ANY)) |
1431 | return PredOpcode; |
1432 | |
1433 | return {}; |
1434 | } |
1435 | |
1436 | // If OP in PTEST(PG, OP(PG, ...)) has a flag-setting variant change the |
1437 | // opcode so the PTEST becomes redundant. |
1438 | switch (PredOpcode) { |
1439 | case AArch64::AND_PPzPP: |
1440 | case AArch64::BIC_PPzPP: |
1441 | case AArch64::EOR_PPzPP: |
1442 | case AArch64::NAND_PPzPP: |
1443 | case AArch64::NOR_PPzPP: |
1444 | case AArch64::ORN_PPzPP: |
1445 | case AArch64::ORR_PPzPP: |
1446 | case AArch64::BRKA_PPzP: |
1447 | case AArch64::BRKPA_PPzPP: |
1448 | case AArch64::BRKB_PPzP: |
1449 | case AArch64::BRKPB_PPzPP: |
1450 | case AArch64::RDFFR_PPz: { |
1451 | // Check to see if our mask is the same. If not the resulting flag bits |
1452 | // may be different and we can't remove the ptest. |
1453 | auto *PredMask = MRI->getUniqueVRegDef(Reg: Pred->getOperand(i: 1).getReg()); |
1454 | if (Mask != PredMask) |
1455 | return {}; |
1456 | break; |
1457 | } |
1458 | case AArch64::BRKN_PPzP: { |
1459 | // BRKN uses an all active implicit mask to set flags unlike the other |
1460 | // flag-setting instructions. |
1461 | // PTEST(PTRUE_B(31), BRKN(PG, A, B)) -> BRKNS(PG, A, B). |
1462 | if ((MaskOpcode != AArch64::PTRUE_B) || |
1463 | (Mask->getOperand(i: 1).getImm() != 31)) |
1464 | return {}; |
1465 | break; |
1466 | } |
1467 | case AArch64::PTRUE_B: |
1468 | // PTEST(OP=PTRUE_B(A), OP) -> PTRUES_B(A) |
1469 | break; |
1470 | default: |
1471 | // Bail out if we don't recognize the input |
1472 | return {}; |
1473 | } |
1474 | |
1475 | return convertToFlagSettingOpc(Opc: PredOpcode); |
1476 | } |
1477 | |
1478 | /// optimizePTestInstr - Attempt to remove a ptest of a predicate-generating |
1479 | /// operation which could set the flags in an identical manner |
1480 | bool AArch64InstrInfo::optimizePTestInstr( |
1481 | MachineInstr *PTest, unsigned MaskReg, unsigned PredReg, |
1482 | const MachineRegisterInfo *MRI) const { |
1483 | auto *Mask = MRI->getUniqueVRegDef(Reg: MaskReg); |
1484 | auto *Pred = MRI->getUniqueVRegDef(Reg: PredReg); |
1485 | unsigned PredOpcode = Pred->getOpcode(); |
1486 | auto NewOp = canRemovePTestInstr(PTest, Mask, Pred, MRI); |
1487 | if (!NewOp) |
1488 | return false; |
1489 | |
1490 | const TargetRegisterInfo *TRI = &getRegisterInfo(); |
1491 | |
1492 | // If another instruction between Pred and PTest accesses flags, don't remove |
1493 | // the ptest or update the earlier instruction to modify them. |
1494 | if (areCFlagsAccessedBetweenInstrs(From: Pred, To: PTest, TRI)) |
1495 | return false; |
1496 | |
1497 | // If we pass all the checks, it's safe to remove the PTEST and use the flags |
1498 | // as they are prior to PTEST. Sometimes this requires the tested PTEST |
1499 | // operand to be replaced with an equivalent instruction that also sets the |
1500 | // flags. |
1501 | PTest->eraseFromParent(); |
1502 | if (*NewOp != PredOpcode) { |
1503 | Pred->setDesc(get(Opcode: *NewOp)); |
1504 | bool succeeded = UpdateOperandRegClass(Instr&: *Pred); |
1505 | (void)succeeded; |
1506 | assert(succeeded && "Operands have incompatible register classes!" ); |
1507 | Pred->addRegisterDefined(Reg: AArch64::NZCV, RegInfo: TRI); |
1508 | } |
1509 | |
1510 | // Ensure that the flags def is live. |
1511 | if (Pred->registerDefIsDead(Reg: AArch64::NZCV, TRI)) { |
1512 | unsigned i = 0, e = Pred->getNumOperands(); |
1513 | for (; i != e; ++i) { |
1514 | MachineOperand &MO = Pred->getOperand(i); |
1515 | if (MO.isReg() && MO.isDef() && MO.getReg() == AArch64::NZCV) { |
1516 | MO.setIsDead(false); |
1517 | break; |
1518 | } |
1519 | } |
1520 | } |
1521 | return true; |
1522 | } |
1523 | |
1524 | /// Try to optimize a compare instruction. A compare instruction is an |
1525 | /// instruction which produces AArch64::NZCV. It can be truly compare |
1526 | /// instruction |
1527 | /// when there are no uses of its destination register. |
1528 | /// |
1529 | /// The following steps are tried in order: |
1530 | /// 1. Convert CmpInstr into an unconditional version. |
1531 | /// 2. Remove CmpInstr if above there is an instruction producing a needed |
1532 | /// condition code or an instruction which can be converted into such an |
1533 | /// instruction. |
1534 | /// Only comparison with zero is supported. |
1535 | bool AArch64InstrInfo::optimizeCompareInstr( |
1536 | MachineInstr &CmpInstr, Register SrcReg, Register SrcReg2, int64_t CmpMask, |
1537 | int64_t CmpValue, const MachineRegisterInfo *MRI) const { |
1538 | assert(CmpInstr.getParent()); |
1539 | assert(MRI); |
1540 | |
1541 | // Replace SUBSWrr with SUBWrr if NZCV is not used. |
1542 | int DeadNZCVIdx = |
1543 | CmpInstr.findRegisterDefOperandIdx(Reg: AArch64::NZCV, /*TRI=*/nullptr, isDead: true); |
1544 | if (DeadNZCVIdx != -1) { |
1545 | if (CmpInstr.definesRegister(Reg: AArch64::WZR, /*TRI=*/nullptr) || |
1546 | CmpInstr.definesRegister(Reg: AArch64::XZR, /*TRI=*/nullptr)) { |
1547 | CmpInstr.eraseFromParent(); |
1548 | return true; |
1549 | } |
1550 | unsigned Opc = CmpInstr.getOpcode(); |
1551 | unsigned NewOpc = convertToNonFlagSettingOpc(MI: CmpInstr); |
1552 | if (NewOpc == Opc) |
1553 | return false; |
1554 | const MCInstrDesc &MCID = get(Opcode: NewOpc); |
1555 | CmpInstr.setDesc(MCID); |
1556 | CmpInstr.removeOperand(OpNo: DeadNZCVIdx); |
1557 | bool succeeded = UpdateOperandRegClass(Instr&: CmpInstr); |
1558 | (void)succeeded; |
1559 | assert(succeeded && "Some operands reg class are incompatible!" ); |
1560 | return true; |
1561 | } |
1562 | |
1563 | if (CmpInstr.getOpcode() == AArch64::PTEST_PP || |
1564 | CmpInstr.getOpcode() == AArch64::PTEST_PP_ANY) |
1565 | return optimizePTestInstr(PTest: &CmpInstr, MaskReg: SrcReg, PredReg: SrcReg2, MRI); |
1566 | |
1567 | if (SrcReg2 != 0) |
1568 | return false; |
1569 | |
1570 | // CmpInstr is a Compare instruction if destination register is not used. |
1571 | if (!MRI->use_nodbg_empty(RegNo: CmpInstr.getOperand(i: 0).getReg())) |
1572 | return false; |
1573 | |
1574 | if (CmpValue == 0 && substituteCmpToZero(CmpInstr, SrcReg, MRI: *MRI)) |
1575 | return true; |
1576 | return (CmpValue == 0 || CmpValue == 1) && |
1577 | removeCmpToZeroOrOne(CmpInstr, SrcReg, CmpValue, MRI: *MRI); |
1578 | } |
1579 | |
1580 | /// Get opcode of S version of Instr. |
1581 | /// If Instr is S version its opcode is returned. |
1582 | /// AArch64::INSTRUCTION_LIST_END is returned if Instr does not have S version |
1583 | /// or we are not interested in it. |
1584 | static unsigned sForm(MachineInstr &Instr) { |
1585 | switch (Instr.getOpcode()) { |
1586 | default: |
1587 | return AArch64::INSTRUCTION_LIST_END; |
1588 | |
1589 | case AArch64::ADDSWrr: |
1590 | case AArch64::ADDSWri: |
1591 | case AArch64::ADDSXrr: |
1592 | case AArch64::ADDSXri: |
1593 | case AArch64::SUBSWrr: |
1594 | case AArch64::SUBSWri: |
1595 | case AArch64::SUBSXrr: |
1596 | case AArch64::SUBSXri: |
1597 | return Instr.getOpcode(); |
1598 | |
1599 | case AArch64::ADDWrr: |
1600 | return AArch64::ADDSWrr; |
1601 | case AArch64::ADDWri: |
1602 | return AArch64::ADDSWri; |
1603 | case AArch64::ADDXrr: |
1604 | return AArch64::ADDSXrr; |
1605 | case AArch64::ADDXri: |
1606 | return AArch64::ADDSXri; |
1607 | case AArch64::ADCWr: |
1608 | return AArch64::ADCSWr; |
1609 | case AArch64::ADCXr: |
1610 | return AArch64::ADCSXr; |
1611 | case AArch64::SUBWrr: |
1612 | return AArch64::SUBSWrr; |
1613 | case AArch64::SUBWri: |
1614 | return AArch64::SUBSWri; |
1615 | case AArch64::SUBXrr: |
1616 | return AArch64::SUBSXrr; |
1617 | case AArch64::SUBXri: |
1618 | return AArch64::SUBSXri; |
1619 | case AArch64::SBCWr: |
1620 | return AArch64::SBCSWr; |
1621 | case AArch64::SBCXr: |
1622 | return AArch64::SBCSXr; |
1623 | case AArch64::ANDWri: |
1624 | return AArch64::ANDSWri; |
1625 | case AArch64::ANDXri: |
1626 | return AArch64::ANDSXri; |
1627 | } |
1628 | } |
1629 | |
1630 | /// Check if AArch64::NZCV should be alive in successors of MBB. |
1631 | static bool areCFlagsAliveInSuccessors(const MachineBasicBlock *MBB) { |
1632 | for (auto *BB : MBB->successors()) |
1633 | if (BB->isLiveIn(Reg: AArch64::NZCV)) |
1634 | return true; |
1635 | return false; |
1636 | } |
1637 | |
1638 | /// \returns The condition code operand index for \p Instr if it is a branch |
1639 | /// or select and -1 otherwise. |
1640 | static int |
1641 | findCondCodeUseOperandIdxForBranchOrSelect(const MachineInstr &Instr) { |
1642 | switch (Instr.getOpcode()) { |
1643 | default: |
1644 | return -1; |
1645 | |
1646 | case AArch64::Bcc: { |
1647 | int Idx = Instr.findRegisterUseOperandIdx(Reg: AArch64::NZCV, /*TRI=*/nullptr); |
1648 | assert(Idx >= 2); |
1649 | return Idx - 2; |
1650 | } |
1651 | |
1652 | case AArch64::CSINVWr: |
1653 | case AArch64::CSINVXr: |
1654 | case AArch64::CSINCWr: |
1655 | case AArch64::CSINCXr: |
1656 | case AArch64::CSELWr: |
1657 | case AArch64::CSELXr: |
1658 | case AArch64::CSNEGWr: |
1659 | case AArch64::CSNEGXr: |
1660 | case AArch64::FCSELSrrr: |
1661 | case AArch64::FCSELDrrr: { |
1662 | int Idx = Instr.findRegisterUseOperandIdx(Reg: AArch64::NZCV, /*TRI=*/nullptr); |
1663 | assert(Idx >= 1); |
1664 | return Idx - 1; |
1665 | } |
1666 | } |
1667 | } |
1668 | |
1669 | /// Find a condition code used by the instruction. |
1670 | /// Returns AArch64CC::Invalid if either the instruction does not use condition |
1671 | /// codes or we don't optimize CmpInstr in the presence of such instructions. |
1672 | static AArch64CC::CondCode findCondCodeUsedByInstr(const MachineInstr &Instr) { |
1673 | int CCIdx = findCondCodeUseOperandIdxForBranchOrSelect(Instr); |
1674 | return CCIdx >= 0 ? static_cast<AArch64CC::CondCode>( |
1675 | Instr.getOperand(i: CCIdx).getImm()) |
1676 | : AArch64CC::Invalid; |
1677 | } |
1678 | |
1679 | static UsedNZCV getUsedNZCV(AArch64CC::CondCode CC) { |
1680 | assert(CC != AArch64CC::Invalid); |
1681 | UsedNZCV UsedFlags; |
1682 | switch (CC) { |
1683 | default: |
1684 | break; |
1685 | |
1686 | case AArch64CC::EQ: // Z set |
1687 | case AArch64CC::NE: // Z clear |
1688 | UsedFlags.Z = true; |
1689 | break; |
1690 | |
1691 | case AArch64CC::HI: // Z clear and C set |
1692 | case AArch64CC::LS: // Z set or C clear |
1693 | UsedFlags.Z = true; |
1694 | [[fallthrough]]; |
1695 | case AArch64CC::HS: // C set |
1696 | case AArch64CC::LO: // C clear |
1697 | UsedFlags.C = true; |
1698 | break; |
1699 | |
1700 | case AArch64CC::MI: // N set |
1701 | case AArch64CC::PL: // N clear |
1702 | UsedFlags.N = true; |
1703 | break; |
1704 | |
1705 | case AArch64CC::VS: // V set |
1706 | case AArch64CC::VC: // V clear |
1707 | UsedFlags.V = true; |
1708 | break; |
1709 | |
1710 | case AArch64CC::GT: // Z clear, N and V the same |
1711 | case AArch64CC::LE: // Z set, N and V differ |
1712 | UsedFlags.Z = true; |
1713 | [[fallthrough]]; |
1714 | case AArch64CC::GE: // N and V the same |
1715 | case AArch64CC::LT: // N and V differ |
1716 | UsedFlags.N = true; |
1717 | UsedFlags.V = true; |
1718 | break; |
1719 | } |
1720 | return UsedFlags; |
1721 | } |
1722 | |
1723 | /// \returns Conditions flags used after \p CmpInstr in its MachineBB if NZCV |
1724 | /// flags are not alive in successors of the same \p CmpInstr and \p MI parent. |
1725 | /// \returns std::nullopt otherwise. |
1726 | /// |
1727 | /// Collect instructions using that flags in \p CCUseInstrs if provided. |
1728 | std::optional<UsedNZCV> |
1729 | llvm::examineCFlagsUse(MachineInstr &MI, MachineInstr &CmpInstr, |
1730 | const TargetRegisterInfo &TRI, |
1731 | SmallVectorImpl<MachineInstr *> *CCUseInstrs) { |
1732 | MachineBasicBlock *CmpParent = CmpInstr.getParent(); |
1733 | if (MI.getParent() != CmpParent) |
1734 | return std::nullopt; |
1735 | |
1736 | if (areCFlagsAliveInSuccessors(MBB: CmpParent)) |
1737 | return std::nullopt; |
1738 | |
1739 | UsedNZCV NZCVUsedAfterCmp; |
1740 | for (MachineInstr &Instr : instructionsWithoutDebug( |
1741 | It: std::next(x: CmpInstr.getIterator()), End: CmpParent->instr_end())) { |
1742 | if (Instr.readsRegister(Reg: AArch64::NZCV, TRI: &TRI)) { |
1743 | AArch64CC::CondCode CC = findCondCodeUsedByInstr(Instr); |
1744 | if (CC == AArch64CC::Invalid) // Unsupported conditional instruction |
1745 | return std::nullopt; |
1746 | NZCVUsedAfterCmp |= getUsedNZCV(CC); |
1747 | if (CCUseInstrs) |
1748 | CCUseInstrs->push_back(Elt: &Instr); |
1749 | } |
1750 | if (Instr.modifiesRegister(Reg: AArch64::NZCV, TRI: &TRI)) |
1751 | break; |
1752 | } |
1753 | return NZCVUsedAfterCmp; |
1754 | } |
1755 | |
1756 | static bool isADDSRegImm(unsigned Opcode) { |
1757 | return Opcode == AArch64::ADDSWri || Opcode == AArch64::ADDSXri; |
1758 | } |
1759 | |
1760 | static bool isSUBSRegImm(unsigned Opcode) { |
1761 | return Opcode == AArch64::SUBSWri || Opcode == AArch64::SUBSXri; |
1762 | } |
1763 | |
1764 | /// Check if CmpInstr can be substituted by MI. |
1765 | /// |
1766 | /// CmpInstr can be substituted: |
1767 | /// - CmpInstr is either 'ADDS %vreg, 0' or 'SUBS %vreg, 0' |
1768 | /// - and, MI and CmpInstr are from the same MachineBB |
1769 | /// - and, condition flags are not alive in successors of the CmpInstr parent |
1770 | /// - and, if MI opcode is the S form there must be no defs of flags between |
1771 | /// MI and CmpInstr |
1772 | /// or if MI opcode is not the S form there must be neither defs of flags |
1773 | /// nor uses of flags between MI and CmpInstr. |
1774 | /// - and, if C/V flags are not used after CmpInstr |
1775 | /// or if N flag is used but MI produces poison value if signed overflow |
1776 | /// occurs. |
1777 | static bool canInstrSubstituteCmpInstr(MachineInstr &MI, MachineInstr &CmpInstr, |
1778 | const TargetRegisterInfo &TRI) { |
1779 | // NOTE this assertion guarantees that MI.getOpcode() is add or subtraction |
1780 | // that may or may not set flags. |
1781 | assert(sForm(MI) != AArch64::INSTRUCTION_LIST_END); |
1782 | |
1783 | const unsigned CmpOpcode = CmpInstr.getOpcode(); |
1784 | if (!isADDSRegImm(Opcode: CmpOpcode) && !isSUBSRegImm(Opcode: CmpOpcode)) |
1785 | return false; |
1786 | |
1787 | assert((CmpInstr.getOperand(2).isImm() && |
1788 | CmpInstr.getOperand(2).getImm() == 0) && |
1789 | "Caller guarantees that CmpInstr compares with constant 0" ); |
1790 | |
1791 | std::optional<UsedNZCV> NZVCUsed = examineCFlagsUse(MI, CmpInstr, TRI); |
1792 | if (!NZVCUsed || NZVCUsed->C) |
1793 | return false; |
1794 | |
1795 | // CmpInstr is either 'ADDS %vreg, 0' or 'SUBS %vreg, 0', and MI is either |
1796 | // '%vreg = add ...' or '%vreg = sub ...'. |
1797 | // Condition flag V is used to indicate signed overflow. |
1798 | // 1) MI and CmpInstr set N and V to the same value. |
1799 | // 2) If MI is add/sub with no-signed-wrap, it produces a poison value when |
1800 | // signed overflow occurs, so CmpInstr could still be simplified away. |
1801 | if (NZVCUsed->V && !MI.getFlag(Flag: MachineInstr::NoSWrap)) |
1802 | return false; |
1803 | |
1804 | AccessKind AccessToCheck = AK_Write; |
1805 | if (sForm(Instr&: MI) != MI.getOpcode()) |
1806 | AccessToCheck = AK_All; |
1807 | return !areCFlagsAccessedBetweenInstrs(From: &MI, To: &CmpInstr, TRI: &TRI, AccessToCheck); |
1808 | } |
1809 | |
1810 | /// Substitute an instruction comparing to zero with another instruction |
1811 | /// which produces needed condition flags. |
1812 | /// |
1813 | /// Return true on success. |
1814 | bool AArch64InstrInfo::substituteCmpToZero( |
1815 | MachineInstr &CmpInstr, unsigned SrcReg, |
1816 | const MachineRegisterInfo &MRI) const { |
1817 | // Get the unique definition of SrcReg. |
1818 | MachineInstr *MI = MRI.getUniqueVRegDef(Reg: SrcReg); |
1819 | if (!MI) |
1820 | return false; |
1821 | |
1822 | const TargetRegisterInfo &TRI = getRegisterInfo(); |
1823 | |
1824 | unsigned NewOpc = sForm(Instr&: *MI); |
1825 | if (NewOpc == AArch64::INSTRUCTION_LIST_END) |
1826 | return false; |
1827 | |
1828 | if (!canInstrSubstituteCmpInstr(MI&: *MI, CmpInstr, TRI)) |
1829 | return false; |
1830 | |
1831 | // Update the instruction to set NZCV. |
1832 | MI->setDesc(get(Opcode: NewOpc)); |
1833 | CmpInstr.eraseFromParent(); |
1834 | bool succeeded = UpdateOperandRegClass(Instr&: *MI); |
1835 | (void)succeeded; |
1836 | assert(succeeded && "Some operands reg class are incompatible!" ); |
1837 | MI->addRegisterDefined(Reg: AArch64::NZCV, RegInfo: &TRI); |
1838 | return true; |
1839 | } |
1840 | |
1841 | /// \returns True if \p CmpInstr can be removed. |
1842 | /// |
1843 | /// \p IsInvertCC is true if, after removing \p CmpInstr, condition |
1844 | /// codes used in \p CCUseInstrs must be inverted. |
1845 | static bool canCmpInstrBeRemoved(MachineInstr &MI, MachineInstr &CmpInstr, |
1846 | int CmpValue, const TargetRegisterInfo &TRI, |
1847 | SmallVectorImpl<MachineInstr *> &CCUseInstrs, |
1848 | bool &IsInvertCC) { |
1849 | assert((CmpValue == 0 || CmpValue == 1) && |
1850 | "Only comparisons to 0 or 1 considered for removal!" ); |
1851 | |
1852 | // MI is 'CSINCWr %vreg, wzr, wzr, <cc>' or 'CSINCXr %vreg, xzr, xzr, <cc>' |
1853 | unsigned MIOpc = MI.getOpcode(); |
1854 | if (MIOpc == AArch64::CSINCWr) { |
1855 | if (MI.getOperand(i: 1).getReg() != AArch64::WZR || |
1856 | MI.getOperand(i: 2).getReg() != AArch64::WZR) |
1857 | return false; |
1858 | } else if (MIOpc == AArch64::CSINCXr) { |
1859 | if (MI.getOperand(i: 1).getReg() != AArch64::XZR || |
1860 | MI.getOperand(i: 2).getReg() != AArch64::XZR) |
1861 | return false; |
1862 | } else { |
1863 | return false; |
1864 | } |
1865 | AArch64CC::CondCode MICC = findCondCodeUsedByInstr(Instr: MI); |
1866 | if (MICC == AArch64CC::Invalid) |
1867 | return false; |
1868 | |
1869 | // NZCV needs to be defined |
1870 | if (MI.findRegisterDefOperandIdx(Reg: AArch64::NZCV, /*TRI=*/nullptr, isDead: true) != -1) |
1871 | return false; |
1872 | |
1873 | // CmpInstr is 'ADDS %vreg, 0' or 'SUBS %vreg, 0' or 'SUBS %vreg, 1' |
1874 | const unsigned CmpOpcode = CmpInstr.getOpcode(); |
1875 | bool IsSubsRegImm = isSUBSRegImm(Opcode: CmpOpcode); |
1876 | if (CmpValue && !IsSubsRegImm) |
1877 | return false; |
1878 | if (!CmpValue && !IsSubsRegImm && !isADDSRegImm(Opcode: CmpOpcode)) |
1879 | return false; |
1880 | |
1881 | // MI conditions allowed: eq, ne, mi, pl |
1882 | UsedNZCV MIUsedNZCV = getUsedNZCV(CC: MICC); |
1883 | if (MIUsedNZCV.C || MIUsedNZCV.V) |
1884 | return false; |
1885 | |
1886 | std::optional<UsedNZCV> NZCVUsedAfterCmp = |
1887 | examineCFlagsUse(MI, CmpInstr, TRI, CCUseInstrs: &CCUseInstrs); |
1888 | // Condition flags are not used in CmpInstr basic block successors and only |
1889 | // Z or N flags allowed to be used after CmpInstr within its basic block |
1890 | if (!NZCVUsedAfterCmp || NZCVUsedAfterCmp->C || NZCVUsedAfterCmp->V) |
1891 | return false; |
1892 | // Z or N flag used after CmpInstr must correspond to the flag used in MI |
1893 | if ((MIUsedNZCV.Z && NZCVUsedAfterCmp->N) || |
1894 | (MIUsedNZCV.N && NZCVUsedAfterCmp->Z)) |
1895 | return false; |
1896 | // If CmpInstr is comparison to zero MI conditions are limited to eq, ne |
1897 | if (MIUsedNZCV.N && !CmpValue) |
1898 | return false; |
1899 | |
1900 | // There must be no defs of flags between MI and CmpInstr |
1901 | if (areCFlagsAccessedBetweenInstrs(From: &MI, To: &CmpInstr, TRI: &TRI, AccessToCheck: AK_Write)) |
1902 | return false; |
1903 | |
1904 | // Condition code is inverted in the following cases: |
1905 | // 1. MI condition is ne; CmpInstr is 'ADDS %vreg, 0' or 'SUBS %vreg, 0' |
1906 | // 2. MI condition is eq, pl; CmpInstr is 'SUBS %vreg, 1' |
1907 | IsInvertCC = (CmpValue && (MICC == AArch64CC::EQ || MICC == AArch64CC::PL)) || |
1908 | (!CmpValue && MICC == AArch64CC::NE); |
1909 | return true; |
1910 | } |
1911 | |
1912 | /// Remove comparison in csinc-cmp sequence |
1913 | /// |
1914 | /// Examples: |
1915 | /// 1. \code |
1916 | /// csinc w9, wzr, wzr, ne |
1917 | /// cmp w9, #0 |
1918 | /// b.eq |
1919 | /// \endcode |
1920 | /// to |
1921 | /// \code |
1922 | /// csinc w9, wzr, wzr, ne |
1923 | /// b.ne |
1924 | /// \endcode |
1925 | /// |
1926 | /// 2. \code |
1927 | /// csinc x2, xzr, xzr, mi |
1928 | /// cmp x2, #1 |
1929 | /// b.pl |
1930 | /// \endcode |
1931 | /// to |
1932 | /// \code |
1933 | /// csinc x2, xzr, xzr, mi |
1934 | /// b.pl |
1935 | /// \endcode |
1936 | /// |
1937 | /// \param CmpInstr comparison instruction |
1938 | /// \return True when comparison removed |
1939 | bool AArch64InstrInfo::removeCmpToZeroOrOne( |
1940 | MachineInstr &CmpInstr, unsigned SrcReg, int CmpValue, |
1941 | const MachineRegisterInfo &MRI) const { |
1942 | MachineInstr *MI = MRI.getUniqueVRegDef(Reg: SrcReg); |
1943 | if (!MI) |
1944 | return false; |
1945 | const TargetRegisterInfo &TRI = getRegisterInfo(); |
1946 | SmallVector<MachineInstr *, 4> CCUseInstrs; |
1947 | bool IsInvertCC = false; |
1948 | if (!canCmpInstrBeRemoved(MI&: *MI, CmpInstr, CmpValue, TRI, CCUseInstrs, |
1949 | IsInvertCC)) |
1950 | return false; |
1951 | // Make transformation |
1952 | CmpInstr.eraseFromParent(); |
1953 | if (IsInvertCC) { |
1954 | // Invert condition codes in CmpInstr CC users |
1955 | for (MachineInstr *CCUseInstr : CCUseInstrs) { |
1956 | int Idx = findCondCodeUseOperandIdxForBranchOrSelect(Instr: *CCUseInstr); |
1957 | assert(Idx >= 0 && "Unexpected instruction using CC." ); |
1958 | MachineOperand &CCOperand = CCUseInstr->getOperand(i: Idx); |
1959 | AArch64CC::CondCode CCUse = AArch64CC::getInvertedCondCode( |
1960 | Code: static_cast<AArch64CC::CondCode>(CCOperand.getImm())); |
1961 | CCOperand.setImm(CCUse); |
1962 | } |
1963 | } |
1964 | return true; |
1965 | } |
1966 | |
1967 | bool AArch64InstrInfo::expandPostRAPseudo(MachineInstr &MI) const { |
1968 | if (MI.getOpcode() != TargetOpcode::LOAD_STACK_GUARD && |
1969 | MI.getOpcode() != AArch64::CATCHRET) |
1970 | return false; |
1971 | |
1972 | MachineBasicBlock &MBB = *MI.getParent(); |
1973 | auto &Subtarget = MBB.getParent()->getSubtarget<AArch64Subtarget>(); |
1974 | auto TRI = Subtarget.getRegisterInfo(); |
1975 | DebugLoc DL = MI.getDebugLoc(); |
1976 | |
1977 | if (MI.getOpcode() == AArch64::CATCHRET) { |
1978 | // Skip to the first instruction before the epilog. |
1979 | const TargetInstrInfo *TII = |
1980 | MBB.getParent()->getSubtarget().getInstrInfo(); |
1981 | MachineBasicBlock *TargetMBB = MI.getOperand(i: 0).getMBB(); |
1982 | auto MBBI = MachineBasicBlock::iterator(MI); |
1983 | MachineBasicBlock::iterator FirstEpilogSEH = std::prev(x: MBBI); |
1984 | while (FirstEpilogSEH->getFlag(Flag: MachineInstr::FrameDestroy) && |
1985 | FirstEpilogSEH != MBB.begin()) |
1986 | FirstEpilogSEH = std::prev(x: FirstEpilogSEH); |
1987 | if (FirstEpilogSEH != MBB.begin()) |
1988 | FirstEpilogSEH = std::next(x: FirstEpilogSEH); |
1989 | BuildMI(BB&: MBB, I: FirstEpilogSEH, MIMD: DL, MCID: TII->get(Opcode: AArch64::ADRP)) |
1990 | .addReg(RegNo: AArch64::X0, flags: RegState::Define) |
1991 | .addMBB(MBB: TargetMBB); |
1992 | BuildMI(BB&: MBB, I: FirstEpilogSEH, MIMD: DL, MCID: TII->get(Opcode: AArch64::ADDXri)) |
1993 | .addReg(RegNo: AArch64::X0, flags: RegState::Define) |
1994 | .addReg(RegNo: AArch64::X0) |
1995 | .addMBB(MBB: TargetMBB) |
1996 | .addImm(Val: 0); |
1997 | return true; |
1998 | } |
1999 | |
2000 | Register Reg = MI.getOperand(i: 0).getReg(); |
2001 | Module &M = *MBB.getParent()->getFunction().getParent(); |
2002 | if (M.getStackProtectorGuard() == "sysreg" ) { |
2003 | const AArch64SysReg::SysReg *SrcReg = |
2004 | AArch64SysReg::lookupSysRegByName(M.getStackProtectorGuardReg()); |
2005 | if (!SrcReg) |
2006 | report_fatal_error(reason: "Unknown SysReg for Stack Protector Guard Register" ); |
2007 | |
2008 | // mrs xN, sysreg |
2009 | BuildMI(BB&: MBB, I&: MI, MIMD: DL, MCID: get(Opcode: AArch64::MRS)) |
2010 | .addDef(RegNo: Reg, Flags: RegState::Renamable) |
2011 | .addImm(Val: SrcReg->Encoding); |
2012 | int Offset = M.getStackProtectorGuardOffset(); |
2013 | if (Offset >= 0 && Offset <= 32760 && Offset % 8 == 0) { |
2014 | // ldr xN, [xN, #offset] |
2015 | BuildMI(BB&: MBB, I&: MI, MIMD: DL, MCID: get(Opcode: AArch64::LDRXui)) |
2016 | .addDef(RegNo: Reg) |
2017 | .addUse(RegNo: Reg, Flags: RegState::Kill) |
2018 | .addImm(Val: Offset / 8); |
2019 | } else if (Offset >= -256 && Offset <= 255) { |
2020 | // ldur xN, [xN, #offset] |
2021 | BuildMI(BB&: MBB, I&: MI, MIMD: DL, MCID: get(Opcode: AArch64::LDURXi)) |
2022 | .addDef(RegNo: Reg) |
2023 | .addUse(RegNo: Reg, Flags: RegState::Kill) |
2024 | .addImm(Val: Offset); |
2025 | } else if (Offset >= -4095 && Offset <= 4095) { |
2026 | if (Offset > 0) { |
2027 | // add xN, xN, #offset |
2028 | BuildMI(BB&: MBB, I&: MI, MIMD: DL, MCID: get(Opcode: AArch64::ADDXri)) |
2029 | .addDef(RegNo: Reg) |
2030 | .addUse(RegNo: Reg, Flags: RegState::Kill) |
2031 | .addImm(Val: Offset) |
2032 | .addImm(Val: 0); |
2033 | } else { |
2034 | // sub xN, xN, #offset |
2035 | BuildMI(BB&: MBB, I&: MI, MIMD: DL, MCID: get(Opcode: AArch64::SUBXri)) |
2036 | .addDef(RegNo: Reg) |
2037 | .addUse(RegNo: Reg, Flags: RegState::Kill) |
2038 | .addImm(Val: -Offset) |
2039 | .addImm(Val: 0); |
2040 | } |
2041 | // ldr xN, [xN] |
2042 | BuildMI(BB&: MBB, I&: MI, MIMD: DL, MCID: get(Opcode: AArch64::LDRXui)) |
2043 | .addDef(RegNo: Reg) |
2044 | .addUse(RegNo: Reg, Flags: RegState::Kill) |
2045 | .addImm(Val: 0); |
2046 | } else { |
2047 | // Cases that are larger than +/- 4095 and not a multiple of 8, or larger |
2048 | // than 23760. |
2049 | // It might be nice to use AArch64::MOVi32imm here, which would get |
2050 | // expanded in PreSched2 after PostRA, but our lone scratch Reg already |
2051 | // contains the MRS result. findScratchNonCalleeSaveRegister() in |
2052 | // AArch64FrameLowering might help us find such a scratch register |
2053 | // though. If we failed to find a scratch register, we could emit a |
2054 | // stream of add instructions to build up the immediate. Or, we could try |
2055 | // to insert a AArch64::MOVi32imm before register allocation so that we |
2056 | // didn't need to scavenge for a scratch register. |
2057 | report_fatal_error(reason: "Unable to encode Stack Protector Guard Offset" ); |
2058 | } |
2059 | MBB.erase(I: MI); |
2060 | return true; |
2061 | } |
2062 | |
2063 | const GlobalValue *GV = |
2064 | cast<GlobalValue>(Val: (*MI.memoperands_begin())->getValue()); |
2065 | const TargetMachine &TM = MBB.getParent()->getTarget(); |
2066 | unsigned OpFlags = Subtarget.ClassifyGlobalReference(GV, TM); |
2067 | const unsigned char MO_NC = AArch64II::MO_NC; |
2068 | |
2069 | if ((OpFlags & AArch64II::MO_GOT) != 0) { |
2070 | BuildMI(BB&: MBB, I&: MI, MIMD: DL, MCID: get(Opcode: AArch64::LOADgot), DestReg: Reg) |
2071 | .addGlobalAddress(GV, Offset: 0, TargetFlags: OpFlags); |
2072 | if (Subtarget.isTargetILP32()) { |
2073 | unsigned Reg32 = TRI->getSubReg(Reg, Idx: AArch64::sub_32); |
2074 | BuildMI(BB&: MBB, I&: MI, MIMD: DL, MCID: get(Opcode: AArch64::LDRWui)) |
2075 | .addDef(RegNo: Reg32, Flags: RegState::Dead) |
2076 | .addUse(RegNo: Reg, Flags: RegState::Kill) |
2077 | .addImm(Val: 0) |
2078 | .addMemOperand(MMO: *MI.memoperands_begin()) |
2079 | .addDef(RegNo: Reg, Flags: RegState::Implicit); |
2080 | } else { |
2081 | BuildMI(BB&: MBB, I&: MI, MIMD: DL, MCID: get(Opcode: AArch64::LDRXui), DestReg: Reg) |
2082 | .addReg(RegNo: Reg, flags: RegState::Kill) |
2083 | .addImm(Val: 0) |
2084 | .addMemOperand(MMO: *MI.memoperands_begin()); |
2085 | } |
2086 | } else if (TM.getCodeModel() == CodeModel::Large) { |
2087 | assert(!Subtarget.isTargetILP32() && "how can large exist in ILP32?" ); |
2088 | BuildMI(BB&: MBB, I&: MI, MIMD: DL, MCID: get(Opcode: AArch64::MOVZXi), DestReg: Reg) |
2089 | .addGlobalAddress(GV, Offset: 0, TargetFlags: AArch64II::MO_G0 | MO_NC) |
2090 | .addImm(Val: 0); |
2091 | BuildMI(BB&: MBB, I&: MI, MIMD: DL, MCID: get(Opcode: AArch64::MOVKXi), DestReg: Reg) |
2092 | .addReg(RegNo: Reg, flags: RegState::Kill) |
2093 | .addGlobalAddress(GV, Offset: 0, TargetFlags: AArch64II::MO_G1 | MO_NC) |
2094 | .addImm(Val: 16); |
2095 | BuildMI(BB&: MBB, I&: MI, MIMD: DL, MCID: get(Opcode: AArch64::MOVKXi), DestReg: Reg) |
2096 | .addReg(RegNo: Reg, flags: RegState::Kill) |
2097 | .addGlobalAddress(GV, Offset: 0, TargetFlags: AArch64II::MO_G2 | MO_NC) |
2098 | .addImm(Val: 32); |
2099 | BuildMI(BB&: MBB, I&: MI, MIMD: DL, MCID: get(Opcode: AArch64::MOVKXi), DestReg: Reg) |
2100 | .addReg(RegNo: Reg, flags: RegState::Kill) |
2101 | .addGlobalAddress(GV, Offset: 0, TargetFlags: AArch64II::MO_G3) |
2102 | .addImm(Val: 48); |
2103 | BuildMI(BB&: MBB, I&: MI, MIMD: DL, MCID: get(Opcode: AArch64::LDRXui), DestReg: Reg) |
2104 | .addReg(RegNo: Reg, flags: RegState::Kill) |
2105 | .addImm(Val: 0) |
2106 | .addMemOperand(MMO: *MI.memoperands_begin()); |
2107 | } else if (TM.getCodeModel() == CodeModel::Tiny) { |
2108 | BuildMI(BB&: MBB, I&: MI, MIMD: DL, MCID: get(Opcode: AArch64::ADR), DestReg: Reg) |
2109 | .addGlobalAddress(GV, Offset: 0, TargetFlags: OpFlags); |
2110 | } else { |
2111 | BuildMI(BB&: MBB, I&: MI, MIMD: DL, MCID: get(Opcode: AArch64::ADRP), DestReg: Reg) |
2112 | .addGlobalAddress(GV, Offset: 0, TargetFlags: OpFlags | AArch64II::MO_PAGE); |
2113 | unsigned char LoFlags = OpFlags | AArch64II::MO_PAGEOFF | MO_NC; |
2114 | if (Subtarget.isTargetILP32()) { |
2115 | unsigned Reg32 = TRI->getSubReg(Reg, Idx: AArch64::sub_32); |
2116 | BuildMI(BB&: MBB, I&: MI, MIMD: DL, MCID: get(Opcode: AArch64::LDRWui)) |
2117 | .addDef(RegNo: Reg32, Flags: RegState::Dead) |
2118 | .addUse(RegNo: Reg, Flags: RegState::Kill) |
2119 | .addGlobalAddress(GV, Offset: 0, TargetFlags: LoFlags) |
2120 | .addMemOperand(MMO: *MI.memoperands_begin()) |
2121 | .addDef(RegNo: Reg, Flags: RegState::Implicit); |
2122 | } else { |
2123 | BuildMI(BB&: MBB, I&: MI, MIMD: DL, MCID: get(Opcode: AArch64::LDRXui), DestReg: Reg) |
2124 | .addReg(RegNo: Reg, flags: RegState::Kill) |
2125 | .addGlobalAddress(GV, Offset: 0, TargetFlags: LoFlags) |
2126 | .addMemOperand(MMO: *MI.memoperands_begin()); |
2127 | } |
2128 | } |
2129 | |
2130 | MBB.erase(I: MI); |
2131 | |
2132 | return true; |
2133 | } |
2134 | |
2135 | // Return true if this instruction simply sets its single destination register |
2136 | // to zero. This is equivalent to a register rename of the zero-register. |
2137 | bool AArch64InstrInfo::isGPRZero(const MachineInstr &MI) { |
2138 | switch (MI.getOpcode()) { |
2139 | default: |
2140 | break; |
2141 | case AArch64::MOVZWi: |
2142 | case AArch64::MOVZXi: // movz Rd, #0 (LSL #0) |
2143 | if (MI.getOperand(i: 1).isImm() && MI.getOperand(i: 1).getImm() == 0) { |
2144 | assert(MI.getDesc().getNumOperands() == 3 && |
2145 | MI.getOperand(2).getImm() == 0 && "invalid MOVZi operands" ); |
2146 | return true; |
2147 | } |
2148 | break; |
2149 | case AArch64::ANDWri: // and Rd, Rzr, #imm |
2150 | return MI.getOperand(i: 1).getReg() == AArch64::WZR; |
2151 | case AArch64::ANDXri: |
2152 | return MI.getOperand(i: 1).getReg() == AArch64::XZR; |
2153 | case TargetOpcode::COPY: |
2154 | return MI.getOperand(i: 1).getReg() == AArch64::WZR; |
2155 | } |
2156 | return false; |
2157 | } |
2158 | |
2159 | // Return true if this instruction simply renames a general register without |
2160 | // modifying bits. |
2161 | bool AArch64InstrInfo::isGPRCopy(const MachineInstr &MI) { |
2162 | switch (MI.getOpcode()) { |
2163 | default: |
2164 | break; |
2165 | case TargetOpcode::COPY: { |
2166 | // GPR32 copies will by lowered to ORRXrs |
2167 | Register DstReg = MI.getOperand(i: 0).getReg(); |
2168 | return (AArch64::GPR32RegClass.contains(Reg: DstReg) || |
2169 | AArch64::GPR64RegClass.contains(Reg: DstReg)); |
2170 | } |
2171 | case AArch64::ORRXrs: // orr Xd, Xzr, Xm (LSL #0) |
2172 | if (MI.getOperand(i: 1).getReg() == AArch64::XZR) { |
2173 | assert(MI.getDesc().getNumOperands() == 4 && |
2174 | MI.getOperand(3).getImm() == 0 && "invalid ORRrs operands" ); |
2175 | return true; |
2176 | } |
2177 | break; |
2178 | case AArch64::ADDXri: // add Xd, Xn, #0 (LSL #0) |
2179 | if (MI.getOperand(i: 2).getImm() == 0) { |
2180 | assert(MI.getDesc().getNumOperands() == 4 && |
2181 | MI.getOperand(3).getImm() == 0 && "invalid ADDXri operands" ); |
2182 | return true; |
2183 | } |
2184 | break; |
2185 | } |
2186 | return false; |
2187 | } |
2188 | |
2189 | // Return true if this instruction simply renames a general register without |
2190 | // modifying bits. |
2191 | bool AArch64InstrInfo::isFPRCopy(const MachineInstr &MI) { |
2192 | switch (MI.getOpcode()) { |
2193 | default: |
2194 | break; |
2195 | case TargetOpcode::COPY: { |
2196 | Register DstReg = MI.getOperand(i: 0).getReg(); |
2197 | return AArch64::FPR128RegClass.contains(Reg: DstReg); |
2198 | } |
2199 | case AArch64::ORRv16i8: |
2200 | if (MI.getOperand(i: 1).getReg() == MI.getOperand(i: 2).getReg()) { |
2201 | assert(MI.getDesc().getNumOperands() == 3 && MI.getOperand(0).isReg() && |
2202 | "invalid ORRv16i8 operands" ); |
2203 | return true; |
2204 | } |
2205 | break; |
2206 | } |
2207 | return false; |
2208 | } |
2209 | |
2210 | Register AArch64InstrInfo::isLoadFromStackSlot(const MachineInstr &MI, |
2211 | int &FrameIndex) const { |
2212 | switch (MI.getOpcode()) { |
2213 | default: |
2214 | break; |
2215 | case AArch64::LDRWui: |
2216 | case AArch64::LDRXui: |
2217 | case AArch64::LDRBui: |
2218 | case AArch64::LDRHui: |
2219 | case AArch64::LDRSui: |
2220 | case AArch64::LDRDui: |
2221 | case AArch64::LDRQui: |
2222 | case AArch64::LDR_PXI: |
2223 | if (MI.getOperand(i: 0).getSubReg() == 0 && MI.getOperand(i: 1).isFI() && |
2224 | MI.getOperand(i: 2).isImm() && MI.getOperand(i: 2).getImm() == 0) { |
2225 | FrameIndex = MI.getOperand(i: 1).getIndex(); |
2226 | return MI.getOperand(i: 0).getReg(); |
2227 | } |
2228 | break; |
2229 | } |
2230 | |
2231 | return 0; |
2232 | } |
2233 | |
2234 | Register AArch64InstrInfo::isStoreToStackSlot(const MachineInstr &MI, |
2235 | int &FrameIndex) const { |
2236 | switch (MI.getOpcode()) { |
2237 | default: |
2238 | break; |
2239 | case AArch64::STRWui: |
2240 | case AArch64::STRXui: |
2241 | case AArch64::STRBui: |
2242 | case AArch64::STRHui: |
2243 | case AArch64::STRSui: |
2244 | case AArch64::STRDui: |
2245 | case AArch64::STRQui: |
2246 | case AArch64::STR_PXI: |
2247 | if (MI.getOperand(i: 0).getSubReg() == 0 && MI.getOperand(i: 1).isFI() && |
2248 | MI.getOperand(i: 2).isImm() && MI.getOperand(i: 2).getImm() == 0) { |
2249 | FrameIndex = MI.getOperand(i: 1).getIndex(); |
2250 | return MI.getOperand(i: 0).getReg(); |
2251 | } |
2252 | break; |
2253 | } |
2254 | return 0; |
2255 | } |
2256 | |
2257 | /// Check all MachineMemOperands for a hint to suppress pairing. |
2258 | bool AArch64InstrInfo::isLdStPairSuppressed(const MachineInstr &MI) { |
2259 | return llvm::any_of(Range: MI.memoperands(), P: [](MachineMemOperand *MMO) { |
2260 | return MMO->getFlags() & MOSuppressPair; |
2261 | }); |
2262 | } |
2263 | |
2264 | /// Set a flag on the first MachineMemOperand to suppress pairing. |
2265 | void AArch64InstrInfo::suppressLdStPair(MachineInstr &MI) { |
2266 | if (MI.memoperands_empty()) |
2267 | return; |
2268 | (*MI.memoperands_begin())->setFlags(MOSuppressPair); |
2269 | } |
2270 | |
2271 | /// Check all MachineMemOperands for a hint that the load/store is strided. |
2272 | bool AArch64InstrInfo::isStridedAccess(const MachineInstr &MI) { |
2273 | return llvm::any_of(Range: MI.memoperands(), P: [](MachineMemOperand *MMO) { |
2274 | return MMO->getFlags() & MOStridedAccess; |
2275 | }); |
2276 | } |
2277 | |
2278 | bool AArch64InstrInfo::hasUnscaledLdStOffset(unsigned Opc) { |
2279 | switch (Opc) { |
2280 | default: |
2281 | return false; |
2282 | case AArch64::STURSi: |
2283 | case AArch64::STRSpre: |
2284 | case AArch64::STURDi: |
2285 | case AArch64::STRDpre: |
2286 | case AArch64::STURQi: |
2287 | case AArch64::STRQpre: |
2288 | case AArch64::STURBBi: |
2289 | case AArch64::STURHHi: |
2290 | case AArch64::STURWi: |
2291 | case AArch64::STRWpre: |
2292 | case AArch64::STURXi: |
2293 | case AArch64::STRXpre: |
2294 | case AArch64::LDURSi: |
2295 | case AArch64::LDRSpre: |
2296 | case AArch64::LDURDi: |
2297 | case AArch64::LDRDpre: |
2298 | case AArch64::LDURQi: |
2299 | case AArch64::LDRQpre: |
2300 | case AArch64::LDURWi: |
2301 | case AArch64::LDRWpre: |
2302 | case AArch64::LDURXi: |
2303 | case AArch64::LDRXpre: |
2304 | case AArch64::LDRSWpre: |
2305 | case AArch64::LDURSWi: |
2306 | case AArch64::LDURHHi: |
2307 | case AArch64::LDURBBi: |
2308 | case AArch64::LDURSBWi: |
2309 | case AArch64::LDURSHWi: |
2310 | return true; |
2311 | } |
2312 | } |
2313 | |
2314 | std::optional<unsigned> AArch64InstrInfo::getUnscaledLdSt(unsigned Opc) { |
2315 | switch (Opc) { |
2316 | default: return {}; |
2317 | case AArch64::PRFMui: return AArch64::PRFUMi; |
2318 | case AArch64::LDRXui: return AArch64::LDURXi; |
2319 | case AArch64::LDRWui: return AArch64::LDURWi; |
2320 | case AArch64::LDRBui: return AArch64::LDURBi; |
2321 | case AArch64::LDRHui: return AArch64::LDURHi; |
2322 | case AArch64::LDRSui: return AArch64::LDURSi; |
2323 | case AArch64::LDRDui: return AArch64::LDURDi; |
2324 | case AArch64::LDRQui: return AArch64::LDURQi; |
2325 | case AArch64::LDRBBui: return AArch64::LDURBBi; |
2326 | case AArch64::LDRHHui: return AArch64::LDURHHi; |
2327 | case AArch64::LDRSBXui: return AArch64::LDURSBXi; |
2328 | case AArch64::LDRSBWui: return AArch64::LDURSBWi; |
2329 | case AArch64::LDRSHXui: return AArch64::LDURSHXi; |
2330 | case AArch64::LDRSHWui: return AArch64::LDURSHWi; |
2331 | case AArch64::LDRSWui: return AArch64::LDURSWi; |
2332 | case AArch64::STRXui: return AArch64::STURXi; |
2333 | case AArch64::STRWui: return AArch64::STURWi; |
2334 | case AArch64::STRBui: return AArch64::STURBi; |
2335 | case AArch64::STRHui: return AArch64::STURHi; |
2336 | case AArch64::STRSui: return AArch64::STURSi; |
2337 | case AArch64::STRDui: return AArch64::STURDi; |
2338 | case AArch64::STRQui: return AArch64::STURQi; |
2339 | case AArch64::STRBBui: return AArch64::STURBBi; |
2340 | case AArch64::STRHHui: return AArch64::STURHHi; |
2341 | } |
2342 | } |
2343 | |
2344 | unsigned AArch64InstrInfo::getLoadStoreImmIdx(unsigned Opc) { |
2345 | switch (Opc) { |
2346 | default: |
2347 | return 2; |
2348 | case AArch64::LDPXi: |
2349 | case AArch64::LDPDi: |
2350 | case AArch64::STPXi: |
2351 | case AArch64::STPDi: |
2352 | case AArch64::LDNPXi: |
2353 | case AArch64::LDNPDi: |
2354 | case AArch64::STNPXi: |
2355 | case AArch64::STNPDi: |
2356 | case AArch64::LDPQi: |
2357 | case AArch64::STPQi: |
2358 | case AArch64::LDNPQi: |
2359 | case AArch64::STNPQi: |
2360 | case AArch64::LDPWi: |
2361 | case AArch64::LDPSi: |
2362 | case AArch64::STPWi: |
2363 | case AArch64::STPSi: |
2364 | case AArch64::LDNPWi: |
2365 | case AArch64::LDNPSi: |
2366 | case AArch64::STNPWi: |
2367 | case AArch64::STNPSi: |
2368 | case AArch64::LDG: |
2369 | case AArch64::STGPi: |
2370 | |
2371 | case AArch64::LD1B_IMM: |
2372 | case AArch64::LD1B_H_IMM: |
2373 | case AArch64::LD1B_S_IMM: |
2374 | case AArch64::LD1B_D_IMM: |
2375 | case AArch64::LD1SB_H_IMM: |
2376 | case AArch64::LD1SB_S_IMM: |
2377 | case AArch64::LD1SB_D_IMM: |
2378 | case AArch64::LD1H_IMM: |
2379 | case AArch64::LD1H_S_IMM: |
2380 | case AArch64::LD1H_D_IMM: |
2381 | case AArch64::LD1SH_S_IMM: |
2382 | case AArch64::LD1SH_D_IMM: |
2383 | case AArch64::LD1W_IMM: |
2384 | case AArch64::LD1W_D_IMM: |
2385 | case AArch64::LD1SW_D_IMM: |
2386 | case AArch64::LD1D_IMM: |
2387 | |
2388 | case AArch64::LD2B_IMM: |
2389 | case AArch64::LD2H_IMM: |
2390 | case AArch64::LD2W_IMM: |
2391 | case AArch64::LD2D_IMM: |
2392 | case AArch64::LD3B_IMM: |
2393 | case AArch64::LD3H_IMM: |
2394 | case AArch64::LD3W_IMM: |
2395 | case AArch64::LD3D_IMM: |
2396 | case AArch64::LD4B_IMM: |
2397 | case AArch64::LD4H_IMM: |
2398 | case AArch64::LD4W_IMM: |
2399 | case AArch64::LD4D_IMM: |
2400 | |
2401 | case AArch64::ST1B_IMM: |
2402 | case AArch64::ST1B_H_IMM: |
2403 | case AArch64::ST1B_S_IMM: |
2404 | case AArch64::ST1B_D_IMM: |
2405 | case AArch64::ST1H_IMM: |
2406 | case AArch64::ST1H_S_IMM: |
2407 | case AArch64::ST1H_D_IMM: |
2408 | case AArch64::ST1W_IMM: |
2409 | case AArch64::ST1W_D_IMM: |
2410 | case AArch64::ST1D_IMM: |
2411 | |
2412 | case AArch64::ST2B_IMM: |
2413 | case AArch64::ST2H_IMM: |
2414 | case AArch64::ST2W_IMM: |
2415 | case AArch64::ST2D_IMM: |
2416 | case AArch64::ST3B_IMM: |
2417 | case AArch64::ST3H_IMM: |
2418 | case AArch64::ST3W_IMM: |
2419 | case AArch64::ST3D_IMM: |
2420 | case AArch64::ST4B_IMM: |
2421 | case AArch64::ST4H_IMM: |
2422 | case AArch64::ST4W_IMM: |
2423 | case AArch64::ST4D_IMM: |
2424 | |
2425 | case AArch64::LD1RB_IMM: |
2426 | case AArch64::LD1RB_H_IMM: |
2427 | case AArch64::LD1RB_S_IMM: |
2428 | case AArch64::LD1RB_D_IMM: |
2429 | case AArch64::LD1RSB_H_IMM: |
2430 | case AArch64::LD1RSB_S_IMM: |
2431 | case AArch64::LD1RSB_D_IMM: |
2432 | case AArch64::LD1RH_IMM: |
2433 | case AArch64::LD1RH_S_IMM: |
2434 | case AArch64::LD1RH_D_IMM: |
2435 | case AArch64::LD1RSH_S_IMM: |
2436 | case AArch64::LD1RSH_D_IMM: |
2437 | case AArch64::LD1RW_IMM: |
2438 | case AArch64::LD1RW_D_IMM: |
2439 | case AArch64::LD1RSW_IMM: |
2440 | case AArch64::LD1RD_IMM: |
2441 | |
2442 | case AArch64::LDNT1B_ZRI: |
2443 | case AArch64::LDNT1H_ZRI: |
2444 | case AArch64::LDNT1W_ZRI: |
2445 | case AArch64::LDNT1D_ZRI: |
2446 | case AArch64::STNT1B_ZRI: |
2447 | case AArch64::STNT1H_ZRI: |
2448 | case AArch64::STNT1W_ZRI: |
2449 | case AArch64::STNT1D_ZRI: |
2450 | |
2451 | case AArch64::LDNF1B_IMM: |
2452 | case AArch64::LDNF1B_H_IMM: |
2453 | case AArch64::LDNF1B_S_IMM: |
2454 | case AArch64::LDNF1B_D_IMM: |
2455 | case AArch64::LDNF1SB_H_IMM: |
2456 | case AArch64::LDNF1SB_S_IMM: |
2457 | case AArch64::LDNF1SB_D_IMM: |
2458 | case AArch64::LDNF1H_IMM: |
2459 | case AArch64::LDNF1H_S_IMM: |
2460 | case AArch64::LDNF1H_D_IMM: |
2461 | case AArch64::LDNF1SH_S_IMM: |
2462 | case AArch64::LDNF1SH_D_IMM: |
2463 | case AArch64::LDNF1W_IMM: |
2464 | case AArch64::LDNF1W_D_IMM: |
2465 | case AArch64::LDNF1SW_D_IMM: |
2466 | case AArch64::LDNF1D_IMM: |
2467 | return 3; |
2468 | case AArch64::ADDG: |
2469 | case AArch64::STGi: |
2470 | case AArch64::LDR_PXI: |
2471 | case AArch64::STR_PXI: |
2472 | return 2; |
2473 | } |
2474 | } |
2475 | |
2476 | bool AArch64InstrInfo::isPairableLdStInst(const MachineInstr &MI) { |
2477 | switch (MI.getOpcode()) { |
2478 | default: |
2479 | return false; |
2480 | // Scaled instructions. |
2481 | case AArch64::STRSui: |
2482 | case AArch64::STRDui: |
2483 | case AArch64::STRQui: |
2484 | case AArch64::STRXui: |
2485 | case AArch64::STRWui: |
2486 | case AArch64::LDRSui: |
2487 | case AArch64::LDRDui: |
2488 | case AArch64::LDRQui: |
2489 | case AArch64::LDRXui: |
2490 | case AArch64::LDRWui: |
2491 | case AArch64::LDRSWui: |
2492 | // Unscaled instructions. |
2493 | case AArch64::STURSi: |
2494 | case AArch64::STRSpre: |
2495 | case AArch64::STURDi: |
2496 | case AArch64::STRDpre: |
2497 | case AArch64::STURQi: |
2498 | case AArch64::STRQpre: |
2499 | case AArch64::STURWi: |
2500 | case AArch64::STRWpre: |
2501 | case AArch64::STURXi: |
2502 | case AArch64::STRXpre: |
2503 | case AArch64::LDURSi: |
2504 | case AArch64::LDRSpre: |
2505 | case AArch64::LDURDi: |
2506 | case AArch64::LDRDpre: |
2507 | case AArch64::LDURQi: |
2508 | case AArch64::LDRQpre: |
2509 | case AArch64::LDURWi: |
2510 | case AArch64::LDRWpre: |
2511 | case AArch64::LDURXi: |
2512 | case AArch64::LDRXpre: |
2513 | case AArch64::LDURSWi: |
2514 | case AArch64::LDRSWpre: |
2515 | return true; |
2516 | } |
2517 | } |
2518 | |
2519 | bool AArch64InstrInfo::isTailCallReturnInst(const MachineInstr &MI) { |
2520 | switch (MI.getOpcode()) { |
2521 | default: |
2522 | assert((!MI.isCall() || !MI.isReturn()) && |
2523 | "Unexpected instruction - was a new tail call opcode introduced?" ); |
2524 | return false; |
2525 | case AArch64::TCRETURNdi: |
2526 | case AArch64::TCRETURNri: |
2527 | case AArch64::TCRETURNrix16x17: |
2528 | case AArch64::TCRETURNrix17: |
2529 | case AArch64::TCRETURNrinotx16: |
2530 | case AArch64::TCRETURNriALL: |
2531 | case AArch64::AUTH_TCRETURN: |
2532 | case AArch64::AUTH_TCRETURN_BTI: |
2533 | return true; |
2534 | } |
2535 | } |
2536 | |
2537 | unsigned AArch64InstrInfo::convertToFlagSettingOpc(unsigned Opc) { |
2538 | switch (Opc) { |
2539 | default: |
2540 | llvm_unreachable("Opcode has no flag setting equivalent!" ); |
2541 | // 32-bit cases: |
2542 | case AArch64::ADDWri: |
2543 | return AArch64::ADDSWri; |
2544 | case AArch64::ADDWrr: |
2545 | return AArch64::ADDSWrr; |
2546 | case AArch64::ADDWrs: |
2547 | return AArch64::ADDSWrs; |
2548 | case AArch64::ADDWrx: |
2549 | return AArch64::ADDSWrx; |
2550 | case AArch64::ANDWri: |
2551 | return AArch64::ANDSWri; |
2552 | case AArch64::ANDWrr: |
2553 | return AArch64::ANDSWrr; |
2554 | case AArch64::ANDWrs: |
2555 | return AArch64::ANDSWrs; |
2556 | case AArch64::BICWrr: |
2557 | return AArch64::BICSWrr; |
2558 | case AArch64::BICWrs: |
2559 | return AArch64::BICSWrs; |
2560 | case AArch64::SUBWri: |
2561 | return AArch64::SUBSWri; |
2562 | case AArch64::SUBWrr: |
2563 | return AArch64::SUBSWrr; |
2564 | case AArch64::SUBWrs: |
2565 | return AArch64::SUBSWrs; |
2566 | case AArch64::SUBWrx: |
2567 | return AArch64::SUBSWrx; |
2568 | // 64-bit cases: |
2569 | case AArch64::ADDXri: |
2570 | return AArch64::ADDSXri; |
2571 | case AArch64::ADDXrr: |
2572 | return AArch64::ADDSXrr; |
2573 | case AArch64::ADDXrs: |
2574 | return AArch64::ADDSXrs; |
2575 | case AArch64::ADDXrx: |
2576 | return AArch64::ADDSXrx; |
2577 | case AArch64::ANDXri: |
2578 | return AArch64::ANDSXri; |
2579 | case AArch64::ANDXrr: |
2580 | return AArch64::ANDSXrr; |
2581 | case AArch64::ANDXrs: |
2582 | return AArch64::ANDSXrs; |
2583 | case AArch64::BICXrr: |
2584 | return AArch64::BICSXrr; |
2585 | case AArch64::BICXrs: |
2586 | return AArch64::BICSXrs; |
2587 | case AArch64::SUBXri: |
2588 | return AArch64::SUBSXri; |
2589 | case AArch64::SUBXrr: |
2590 | return AArch64::SUBSXrr; |
2591 | case AArch64::SUBXrs: |
2592 | return AArch64::SUBSXrs; |
2593 | case AArch64::SUBXrx: |
2594 | return AArch64::SUBSXrx; |
2595 | // SVE instructions: |
2596 | case AArch64::AND_PPzPP: |
2597 | return AArch64::ANDS_PPzPP; |
2598 | case AArch64::BIC_PPzPP: |
2599 | return AArch64::BICS_PPzPP; |
2600 | case AArch64::EOR_PPzPP: |
2601 | return AArch64::EORS_PPzPP; |
2602 | case AArch64::NAND_PPzPP: |
2603 | return AArch64::NANDS_PPzPP; |
2604 | case AArch64::NOR_PPzPP: |
2605 | return AArch64::NORS_PPzPP; |
2606 | case AArch64::ORN_PPzPP: |
2607 | return AArch64::ORNS_PPzPP; |
2608 | case AArch64::ORR_PPzPP: |
2609 | return AArch64::ORRS_PPzPP; |
2610 | case AArch64::BRKA_PPzP: |
2611 | return AArch64::BRKAS_PPzP; |
2612 | case AArch64::BRKPA_PPzPP: |
2613 | return AArch64::BRKPAS_PPzPP; |
2614 | case AArch64::BRKB_PPzP: |
2615 | return AArch64::BRKBS_PPzP; |
2616 | case AArch64::BRKPB_PPzPP: |
2617 | return AArch64::BRKPBS_PPzPP; |
2618 | case AArch64::BRKN_PPzP: |
2619 | return AArch64::BRKNS_PPzP; |
2620 | case AArch64::RDFFR_PPz: |
2621 | return AArch64::RDFFRS_PPz; |
2622 | case AArch64::PTRUE_B: |
2623 | return AArch64::PTRUES_B; |
2624 | } |
2625 | } |
2626 | |
2627 | // Is this a candidate for ld/st merging or pairing? For example, we don't |
2628 | // touch volatiles or load/stores that have a hint to avoid pair formation. |
2629 | bool AArch64InstrInfo::isCandidateToMergeOrPair(const MachineInstr &MI) const { |
2630 | |
2631 | bool IsPreLdSt = isPreLdSt(MI); |
2632 | |
2633 | // If this is a volatile load/store, don't mess with it. |
2634 | if (MI.hasOrderedMemoryRef()) |
2635 | return false; |
2636 | |
2637 | // Make sure this is a reg/fi+imm (as opposed to an address reloc). |
2638 | // For Pre-inc LD/ST, the operand is shifted by one. |
2639 | assert((MI.getOperand(IsPreLdSt ? 2 : 1).isReg() || |
2640 | MI.getOperand(IsPreLdSt ? 2 : 1).isFI()) && |
2641 | "Expected a reg or frame index operand." ); |
2642 | |
2643 | // For Pre-indexed addressing quadword instructions, the third operand is the |
2644 | // immediate value. |
2645 | bool IsImmPreLdSt = IsPreLdSt && MI.getOperand(i: 3).isImm(); |
2646 | |
2647 | if (!MI.getOperand(i: 2).isImm() && !IsImmPreLdSt) |
2648 | return false; |
2649 | |
2650 | // Can't merge/pair if the instruction modifies the base register. |
2651 | // e.g., ldr x0, [x0] |
2652 | // This case will never occur with an FI base. |
2653 | // However, if the instruction is an LDR<S,D,Q,W,X,SW>pre or |
2654 | // STR<S,D,Q,W,X>pre, it can be merged. |
2655 | // For example: |
2656 | // ldr q0, [x11, #32]! |
2657 | // ldr q1, [x11, #16] |
2658 | // to |
2659 | // ldp q0, q1, [x11, #32]! |
2660 | if (MI.getOperand(i: 1).isReg() && !IsPreLdSt) { |
2661 | Register BaseReg = MI.getOperand(i: 1).getReg(); |
2662 | const TargetRegisterInfo *TRI = &getRegisterInfo(); |
2663 | if (MI.modifiesRegister(Reg: BaseReg, TRI)) |
2664 | return false; |
2665 | } |
2666 | |
2667 | // Check if this load/store has a hint to avoid pair formation. |
2668 | // MachineMemOperands hints are set by the AArch64StorePairSuppress pass. |
2669 | if (isLdStPairSuppressed(MI)) |
2670 | return false; |
2671 | |
2672 | // Do not pair any callee-save store/reload instructions in the |
2673 | // prologue/epilogue if the CFI information encoded the operations as separate |
2674 | // instructions, as that will cause the size of the actual prologue to mismatch |
2675 | // with the prologue size recorded in the Windows CFI. |
2676 | const MCAsmInfo *MAI = MI.getMF()->getTarget().getMCAsmInfo(); |
2677 | bool NeedsWinCFI = MAI->usesWindowsCFI() && |
2678 | MI.getMF()->getFunction().needsUnwindTableEntry(); |
2679 | if (NeedsWinCFI && (MI.getFlag(Flag: MachineInstr::FrameSetup) || |
2680 | MI.getFlag(Flag: MachineInstr::FrameDestroy))) |
2681 | return false; |
2682 | |
2683 | // On some CPUs quad load/store pairs are slower than two single load/stores. |
2684 | if (Subtarget.isPaired128Slow()) { |
2685 | switch (MI.getOpcode()) { |
2686 | default: |
2687 | break; |
2688 | case AArch64::LDURQi: |
2689 | case AArch64::STURQi: |
2690 | case AArch64::LDRQui: |
2691 | case AArch64::STRQui: |
2692 | return false; |
2693 | } |
2694 | } |
2695 | |
2696 | return true; |
2697 | } |
2698 | |
2699 | bool AArch64InstrInfo::getMemOperandsWithOffsetWidth( |
2700 | const MachineInstr &LdSt, SmallVectorImpl<const MachineOperand *> &BaseOps, |
2701 | int64_t &Offset, bool &OffsetIsScalable, LocationSize &Width, |
2702 | const TargetRegisterInfo *TRI) const { |
2703 | if (!LdSt.mayLoadOrStore()) |
2704 | return false; |
2705 | |
2706 | const MachineOperand *BaseOp; |
2707 | TypeSize WidthN(0, false); |
2708 | if (!getMemOperandWithOffsetWidth(MI: LdSt, BaseOp, Offset, OffsetIsScalable, |
2709 | Width&: WidthN, TRI)) |
2710 | return false; |
2711 | // The maximum vscale is 16 under AArch64, return the maximal extent for the |
2712 | // vector. |
2713 | Width = LocationSize::precise(Value: WidthN); |
2714 | BaseOps.push_back(Elt: BaseOp); |
2715 | return true; |
2716 | } |
2717 | |
2718 | std::optional<ExtAddrMode> |
2719 | AArch64InstrInfo::getAddrModeFromMemoryOp(const MachineInstr &MemI, |
2720 | const TargetRegisterInfo *TRI) const { |
2721 | const MachineOperand *Base; // Filled with the base operand of MI. |
2722 | int64_t Offset; // Filled with the offset of MI. |
2723 | bool OffsetIsScalable; |
2724 | if (!getMemOperandWithOffset(MI: MemI, BaseOp&: Base, Offset, OffsetIsScalable, TRI)) |
2725 | return std::nullopt; |
2726 | |
2727 | if (!Base->isReg()) |
2728 | return std::nullopt; |
2729 | ExtAddrMode AM; |
2730 | AM.BaseReg = Base->getReg(); |
2731 | AM.Displacement = Offset; |
2732 | AM.ScaledReg = 0; |
2733 | AM.Scale = 0; |
2734 | return AM; |
2735 | } |
2736 | |
2737 | bool AArch64InstrInfo::canFoldIntoAddrMode(const MachineInstr &MemI, |
2738 | Register Reg, |
2739 | const MachineInstr &AddrI, |
2740 | ExtAddrMode &AM) const { |
2741 | // Filter out instructions into which we cannot fold. |
2742 | unsigned NumBytes; |
2743 | int64_t OffsetScale = 1; |
2744 | switch (MemI.getOpcode()) { |
2745 | default: |
2746 | return false; |
2747 | |
2748 | case AArch64::LDURQi: |
2749 | case AArch64::STURQi: |
2750 | NumBytes = 16; |
2751 | break; |
2752 | |
2753 | case AArch64::LDURDi: |
2754 | case AArch64::STURDi: |
2755 | case AArch64::LDURXi: |
2756 | case AArch64::STURXi: |
2757 | NumBytes = 8; |
2758 | break; |
2759 | |
2760 | case AArch64::LDURWi: |
2761 | case AArch64::LDURSWi: |
2762 | case AArch64::STURWi: |
2763 | NumBytes = 4; |
2764 | break; |
2765 | |
2766 | case AArch64::LDURHi: |
2767 | case AArch64::STURHi: |
2768 | case AArch64::LDURHHi: |
2769 | case AArch64::STURHHi: |
2770 | case AArch64::LDURSHXi: |
2771 | case AArch64::LDURSHWi: |
2772 | NumBytes = 2; |
2773 | break; |
2774 | |
2775 | case AArch64::LDRBroX: |
2776 | case AArch64::LDRBBroX: |
2777 | case AArch64::LDRSBXroX: |
2778 | case AArch64::LDRSBWroX: |
2779 | case AArch64::STRBroX: |
2780 | case AArch64::STRBBroX: |
2781 | case AArch64::LDURBi: |
2782 | case AArch64::LDURBBi: |
2783 | case AArch64::LDURSBXi: |
2784 | case AArch64::LDURSBWi: |
2785 | case AArch64::STURBi: |
2786 | case AArch64::STURBBi: |
2787 | case AArch64::LDRBui: |
2788 | case AArch64::LDRBBui: |
2789 | case AArch64::LDRSBXui: |
2790 | case AArch64::LDRSBWui: |
2791 | case AArch64::STRBui: |
2792 | case AArch64::STRBBui: |
2793 | NumBytes = 1; |
2794 | break; |
2795 | |
2796 | case AArch64::LDRQroX: |
2797 | case AArch64::STRQroX: |
2798 | case AArch64::LDRQui: |
2799 | case AArch64::STRQui: |
2800 | NumBytes = 16; |
2801 | OffsetScale = 16; |
2802 | break; |
2803 | |
2804 | case AArch64::LDRDroX: |
2805 | case AArch64::STRDroX: |
2806 | case AArch64::LDRXroX: |
2807 | case AArch64::STRXroX: |
2808 | case AArch64::LDRDui: |
2809 | case AArch64::STRDui: |
2810 | case AArch64::LDRXui: |
2811 | case AArch64::STRXui: |
2812 | NumBytes = 8; |
2813 | OffsetScale = 8; |
2814 | break; |
2815 | |
2816 | case AArch64::LDRWroX: |
2817 | case AArch64::LDRSWroX: |
2818 | case AArch64::STRWroX: |
2819 | case AArch64::LDRWui: |
2820 | case AArch64::LDRSWui: |
2821 | case AArch64::STRWui: |
2822 | NumBytes = 4; |
2823 | OffsetScale = 4; |
2824 | break; |
2825 | |
2826 | case AArch64::LDRHroX: |
2827 | case AArch64::STRHroX: |
2828 | case AArch64::LDRHHroX: |
2829 | case AArch64::STRHHroX: |
2830 | case AArch64::LDRSHXroX: |
2831 | case AArch64::LDRSHWroX: |
2832 | case AArch64::LDRHui: |
2833 | case AArch64::STRHui: |
2834 | case AArch64::LDRHHui: |
2835 | case AArch64::STRHHui: |
2836 | case AArch64::LDRSHXui: |
2837 | case AArch64::LDRSHWui: |
2838 | NumBytes = 2; |
2839 | OffsetScale = 2; |
2840 | break; |
2841 | } |
2842 | |
2843 | // Check the fold operand is not the loaded/stored value. |
2844 | const MachineOperand &BaseRegOp = MemI.getOperand(i: 0); |
2845 | if (BaseRegOp.isReg() && BaseRegOp.getReg() == Reg) |
2846 | return false; |
2847 | |
2848 | // Handle memory instructions with a [Reg, Reg] addressing mode. |
2849 | if (MemI.getOperand(i: 2).isReg()) { |
2850 | // Bail if the addressing mode already includes extension of the offset |
2851 | // register. |
2852 | if (MemI.getOperand(i: 3).getImm()) |
2853 | return false; |
2854 | |
2855 | // Check if we actually have a scaled offset. |
2856 | if (MemI.getOperand(i: 4).getImm() == 0) |
2857 | OffsetScale = 1; |
2858 | |
2859 | // If the address instructions is folded into the base register, then the |
2860 | // addressing mode must not have a scale. Then we can swap the base and the |
2861 | // scaled registers. |
2862 | if (MemI.getOperand(i: 1).getReg() == Reg && OffsetScale != 1) |
2863 | return false; |
2864 | |
2865 | switch (AddrI.getOpcode()) { |
2866 | default: |
2867 | return false; |
2868 | |
2869 | case AArch64::SBFMXri: |
2870 | // sxtw Xa, Wm |
2871 | // ldr Xd, [Xn, Xa, lsl #N] |
2872 | // -> |
2873 | // ldr Xd, [Xn, Wm, sxtw #N] |
2874 | if (AddrI.getOperand(i: 2).getImm() != 0 || |
2875 | AddrI.getOperand(i: 3).getImm() != 31) |
2876 | return false; |
2877 | |
2878 | AM.BaseReg = MemI.getOperand(i: 1).getReg(); |
2879 | if (AM.BaseReg == Reg) |
2880 | AM.BaseReg = MemI.getOperand(i: 2).getReg(); |
2881 | AM.ScaledReg = AddrI.getOperand(i: 1).getReg(); |
2882 | AM.Scale = OffsetScale; |
2883 | AM.Displacement = 0; |
2884 | AM.Form = ExtAddrMode::Formula::SExtScaledReg; |
2885 | return true; |
2886 | |
2887 | case TargetOpcode::SUBREG_TO_REG: { |
2888 | // mov Wa, Wm |
2889 | // ldr Xd, [Xn, Xa, lsl #N] |
2890 | // -> |
2891 | // ldr Xd, [Xn, Wm, uxtw #N] |
2892 | |
2893 | // Zero-extension looks like an ORRWrs followed by a SUBREG_TO_REG. |
2894 | if (AddrI.getOperand(i: 1).getImm() != 0 || |
2895 | AddrI.getOperand(i: 3).getImm() != AArch64::sub_32) |
2896 | return false; |
2897 | |
2898 | const MachineRegisterInfo &MRI = AddrI.getMF()->getRegInfo(); |
2899 | Register OffsetReg = AddrI.getOperand(i: 2).getReg(); |
2900 | if (!OffsetReg.isVirtual() || !MRI.hasOneNonDBGUse(RegNo: OffsetReg)) |
2901 | return false; |
2902 | |
2903 | const MachineInstr &DefMI = *MRI.getVRegDef(Reg: OffsetReg); |
2904 | if (DefMI.getOpcode() != AArch64::ORRWrs || |
2905 | DefMI.getOperand(i: 1).getReg() != AArch64::WZR || |
2906 | DefMI.getOperand(i: 3).getImm() != 0) |
2907 | return false; |
2908 | |
2909 | AM.BaseReg = MemI.getOperand(i: 1).getReg(); |
2910 | if (AM.BaseReg == Reg) |
2911 | AM.BaseReg = MemI.getOperand(i: 2).getReg(); |
2912 | AM.ScaledReg = DefMI.getOperand(i: 2).getReg(); |
2913 | AM.Scale = OffsetScale; |
2914 | AM.Displacement = 0; |
2915 | AM.Form = ExtAddrMode::Formula::ZExtScaledReg; |
2916 | return true; |
2917 | } |
2918 | } |
2919 | } |
2920 | |
2921 | // Handle memory instructions with a [Reg, #Imm] addressing mode. |
2922 | |
2923 | // Check we are not breaking a potential conversion to an LDP. |
2924 | auto validateOffsetForLDP = [](unsigned NumBytes, int64_t OldOffset, |
2925 | int64_t NewOffset) -> bool { |
2926 | int64_t MinOffset, MaxOffset; |
2927 | switch (NumBytes) { |
2928 | default: |
2929 | return true; |
2930 | case 4: |
2931 | MinOffset = -256; |
2932 | MaxOffset = 252; |
2933 | break; |
2934 | case 8: |
2935 | MinOffset = -512; |
2936 | MaxOffset = 504; |
2937 | break; |
2938 | case 16: |
2939 | MinOffset = -1024; |
2940 | MaxOffset = 1008; |
2941 | break; |
2942 | } |
2943 | return OldOffset < MinOffset || OldOffset > MaxOffset || |
2944 | (NewOffset >= MinOffset && NewOffset <= MaxOffset); |
2945 | }; |
2946 | auto canFoldAddSubImmIntoAddrMode = [&](int64_t Disp) -> bool { |
2947 | int64_t OldOffset = MemI.getOperand(i: 2).getImm() * OffsetScale; |
2948 | int64_t NewOffset = OldOffset + Disp; |
2949 | if (!isLegalAddressingMode(NumBytes, Offset: NewOffset, /* Scale */ 0)) |
2950 | return false; |
2951 | // If the old offset would fit into an LDP, but the new offset wouldn't, |
2952 | // bail out. |
2953 | if (!validateOffsetForLDP(NumBytes, OldOffset, NewOffset)) |
2954 | return false; |
2955 | AM.BaseReg = AddrI.getOperand(i: 1).getReg(); |
2956 | AM.ScaledReg = 0; |
2957 | AM.Scale = 0; |
2958 | AM.Displacement = NewOffset; |
2959 | AM.Form = ExtAddrMode::Formula::Basic; |
2960 | return true; |
2961 | }; |
2962 | |
2963 | auto canFoldAddRegIntoAddrMode = |
2964 | [&](int64_t Scale, |
2965 | ExtAddrMode::Formula Form = ExtAddrMode::Formula::Basic) -> bool { |
2966 | if (MemI.getOperand(i: 2).getImm() != 0) |
2967 | return false; |
2968 | if (!isLegalAddressingMode(NumBytes, /* Offset */ 0, Scale)) |
2969 | return false; |
2970 | AM.BaseReg = AddrI.getOperand(i: 1).getReg(); |
2971 | AM.ScaledReg = AddrI.getOperand(i: 2).getReg(); |
2972 | AM.Scale = Scale; |
2973 | AM.Displacement = 0; |
2974 | AM.Form = Form; |
2975 | return true; |
2976 | }; |
2977 | |
2978 | auto avoidSlowSTRQ = [&](const MachineInstr &MemI) { |
2979 | unsigned Opcode = MemI.getOpcode(); |
2980 | return (Opcode == AArch64::STURQi || Opcode == AArch64::STRQui) && |
2981 | Subtarget.isSTRQroSlow(); |
2982 | }; |
2983 | |
2984 | int64_t Disp = 0; |
2985 | const bool OptSize = MemI.getMF()->getFunction().hasOptSize(); |
2986 | switch (AddrI.getOpcode()) { |
2987 | default: |
2988 | return false; |
2989 | |
2990 | case AArch64::ADDXri: |
2991 | // add Xa, Xn, #N |
2992 | // ldr Xd, [Xa, #M] |
2993 | // -> |
2994 | // ldr Xd, [Xn, #N'+M] |
2995 | Disp = AddrI.getOperand(i: 2).getImm() << AddrI.getOperand(i: 3).getImm(); |
2996 | return canFoldAddSubImmIntoAddrMode(Disp); |
2997 | |
2998 | case AArch64::SUBXri: |
2999 | // sub Xa, Xn, #N |
3000 | // ldr Xd, [Xa, #M] |
3001 | // -> |
3002 | // ldr Xd, [Xn, #N'+M] |
3003 | Disp = AddrI.getOperand(i: 2).getImm() << AddrI.getOperand(i: 3).getImm(); |
3004 | return canFoldAddSubImmIntoAddrMode(-Disp); |
3005 | |
3006 | case AArch64::ADDXrs: { |
3007 | // add Xa, Xn, Xm, lsl #N |
3008 | // ldr Xd, [Xa] |
3009 | // -> |
3010 | // ldr Xd, [Xn, Xm, lsl #N] |
3011 | |
3012 | // Don't fold the add if the result would be slower, unless optimising for |
3013 | // size. |
3014 | unsigned Shift = static_cast<unsigned>(AddrI.getOperand(i: 3).getImm()); |
3015 | if (AArch64_AM::getShiftType(Imm: Shift) != AArch64_AM::ShiftExtendType::LSL) |
3016 | return false; |
3017 | Shift = AArch64_AM::getShiftValue(Imm: Shift); |
3018 | if (!OptSize) { |
3019 | if (Shift != 2 && Shift != 3 && Subtarget.hasAddrLSLSlow14()) |
3020 | return false; |
3021 | if (avoidSlowSTRQ(MemI)) |
3022 | return false; |
3023 | } |
3024 | return canFoldAddRegIntoAddrMode(1ULL << Shift); |
3025 | } |
3026 | |
3027 | case AArch64::ADDXrr: |
3028 | // add Xa, Xn, Xm |
3029 | // ldr Xd, [Xa] |
3030 | // -> |
3031 | // ldr Xd, [Xn, Xm, lsl #0] |
3032 | |
3033 | // Don't fold the add if the result would be slower, unless optimising for |
3034 | // size. |
3035 | if (!OptSize && avoidSlowSTRQ(MemI)) |
3036 | return false; |
3037 | return canFoldAddRegIntoAddrMode(1); |
3038 | |
3039 | case AArch64::ADDXrx: |
3040 | // add Xa, Xn, Wm, {s,u}xtw #N |
3041 | // ldr Xd, [Xa] |
3042 | // -> |
3043 | // ldr Xd, [Xn, Wm, {s,u}xtw #N] |
3044 | |
3045 | // Don't fold the add if the result would be slower, unless optimising for |
3046 | // size. |
3047 | if (!OptSize && avoidSlowSTRQ(MemI)) |
3048 | return false; |
3049 | |
3050 | // Can fold only sign-/zero-extend of a word. |
3051 | unsigned Imm = static_cast<unsigned>(AddrI.getOperand(i: 3).getImm()); |
3052 | AArch64_AM::ShiftExtendType Extend = AArch64_AM::getArithExtendType(Imm); |
3053 | if (Extend != AArch64_AM::UXTW && Extend != AArch64_AM::SXTW) |
3054 | return false; |
3055 | |
3056 | return canFoldAddRegIntoAddrMode( |
3057 | 1ULL << AArch64_AM::getArithShiftValue(Imm), |
3058 | (Extend == AArch64_AM::SXTW) ? ExtAddrMode::Formula::SExtScaledReg |
3059 | : ExtAddrMode::Formula::ZExtScaledReg); |
3060 | } |
3061 | } |
3062 | |
3063 | // Given an opcode for an instruction with a [Reg, #Imm] addressing mode, |
3064 | // return the opcode of an instruction performing the same operation, but using |
3065 | // the [Reg, Reg] addressing mode. |
3066 | static unsigned regOffsetOpcode(unsigned Opcode) { |
3067 | switch (Opcode) { |
3068 | default: |
3069 | llvm_unreachable("Address folding not implemented for instruction" ); |
3070 | |
3071 | case AArch64::LDURQi: |
3072 | case AArch64::LDRQui: |
3073 | return AArch64::LDRQroX; |
3074 | case AArch64::STURQi: |
3075 | case AArch64::STRQui: |
3076 | return AArch64::STRQroX; |
3077 | case AArch64::LDURDi: |
3078 | case AArch64::LDRDui: |
3079 | return AArch64::LDRDroX; |
3080 | case AArch64::STURDi: |
3081 | case AArch64::STRDui: |
3082 | return AArch64::STRDroX; |
3083 | case AArch64::LDURXi: |
3084 | case AArch64::LDRXui: |
3085 | return AArch64::LDRXroX; |
3086 | case AArch64::STURXi: |
3087 | case AArch64::STRXui: |
3088 | return AArch64::STRXroX; |
3089 | case AArch64::LDURWi: |
3090 | case AArch64::LDRWui: |
3091 | return AArch64::LDRWroX; |
3092 | case AArch64::LDURSWi: |
3093 | case AArch64::LDRSWui: |
3094 | return AArch64::LDRSWroX; |
3095 | case AArch64::STURWi: |
3096 | case AArch64::STRWui: |
3097 | return AArch64::STRWroX; |
3098 | case AArch64::LDURHi: |
3099 | case AArch64::LDRHui: |
3100 | return AArch64::LDRHroX; |
3101 | case AArch64::STURHi: |
3102 | case AArch64::STRHui: |
3103 | return AArch64::STRHroX; |
3104 | case AArch64::LDURHHi: |
3105 | case AArch64::LDRHHui: |
3106 | return AArch64::LDRHHroX; |
3107 | case AArch64::STURHHi: |
3108 | case AArch64::STRHHui: |
3109 | return AArch64::STRHHroX; |
3110 | case AArch64::LDURSHXi: |
3111 | case AArch64::LDRSHXui: |
3112 | return AArch64::LDRSHXroX; |
3113 | case AArch64::LDURSHWi: |
3114 | case AArch64::LDRSHWui: |
3115 | return AArch64::LDRSHWroX; |
3116 | case AArch64::LDURBi: |
3117 | case AArch64::LDRBui: |
3118 | return AArch64::LDRBroX; |
3119 | case AArch64::LDURBBi: |
3120 | case AArch64::LDRBBui: |
3121 | return AArch64::LDRBBroX; |
3122 | case AArch64::LDURSBXi: |
3123 | case AArch64::LDRSBXui: |
3124 | return AArch64::LDRSBXroX; |
3125 | case AArch64::LDURSBWi: |
3126 | case AArch64::LDRSBWui: |
3127 | return AArch64::LDRSBWroX; |
3128 | case AArch64::STURBi: |
3129 | case AArch64::STRBui: |
3130 | return AArch64::STRBroX; |
3131 | case AArch64::STURBBi: |
3132 | case AArch64::STRBBui: |
3133 | return AArch64::STRBBroX; |
3134 | } |
3135 | } |
3136 | |
3137 | // Given an opcode for an instruction with a [Reg, #Imm] addressing mode, return |
3138 | // the opcode of an instruction performing the same operation, but using the |
3139 | // [Reg, #Imm] addressing mode with scaled offset. |
3140 | unsigned scaledOffsetOpcode(unsigned Opcode, unsigned &Scale) { |
3141 | switch (Opcode) { |
3142 | default: |
3143 | llvm_unreachable("Address folding not implemented for instruction" ); |
3144 | |
3145 | case AArch64::LDURQi: |
3146 | Scale = 16; |
3147 | return AArch64::LDRQui; |
3148 | case AArch64::STURQi: |
3149 | Scale = 16; |
3150 | return AArch64::STRQui; |
3151 | case AArch64::LDURDi: |
3152 | Scale = 8; |
3153 | return AArch64::LDRDui; |
3154 | case AArch64::STURDi: |
3155 | Scale = 8; |
3156 | return AArch64::STRDui; |
3157 | case AArch64::LDURXi: |
3158 | Scale = 8; |
3159 | return AArch64::LDRXui; |
3160 | case AArch64::STURXi: |
3161 | Scale = 8; |
3162 | return AArch64::STRXui; |
3163 | case AArch64::LDURWi: |
3164 | Scale = 4; |
3165 | return AArch64::LDRWui; |
3166 | case AArch64::LDURSWi: |
3167 | Scale = 4; |
3168 | return AArch64::LDRSWui; |
3169 | case AArch64::STURWi: |
3170 | Scale = 4; |
3171 | return AArch64::STRWui; |
3172 | case AArch64::LDURHi: |
3173 | Scale = 2; |
3174 | return AArch64::LDRHui; |
3175 | case AArch64::STURHi: |
3176 | Scale = 2; |
3177 | return AArch64::STRHui; |
3178 | case AArch64::LDURHHi: |
3179 | Scale = 2; |
3180 | return AArch64::LDRHHui; |
3181 | case AArch64::STURHHi: |
3182 | Scale = 2; |
3183 | return AArch64::STRHHui; |
3184 | case AArch64::LDURSHXi: |
3185 | Scale = 2; |
3186 | return AArch64::LDRSHXui; |
3187 | case AArch64::LDURSHWi: |
3188 | Scale = 2; |
3189 | return AArch64::LDRSHWui; |
3190 | case AArch64::LDURBi: |
3191 | Scale = 1; |
3192 | return AArch64::LDRBui; |
3193 | case AArch64::LDURBBi: |
3194 | Scale = 1; |
3195 | return AArch64::LDRBBui; |
3196 | case AArch64::LDURSBXi: |
3197 | Scale = 1; |
3198 | return AArch64::LDRSBXui; |
3199 | case AArch64::LDURSBWi: |
3200 | Scale = 1; |
3201 | return AArch64::LDRSBWui; |
3202 | case AArch64::STURBi: |
3203 | Scale = 1; |
3204 | return AArch64::STRBui; |
3205 | case AArch64::STURBBi: |
3206 | Scale = 1; |
3207 | return AArch64::STRBBui; |
3208 | case AArch64::LDRQui: |
3209 | case AArch64::STRQui: |
3210 | Scale = 16; |
3211 | return Opcode; |
3212 | case AArch64::LDRDui: |
3213 | case AArch64::STRDui: |
3214 | case AArch64::LDRXui: |
3215 | case AArch64::STRXui: |
3216 | Scale = 8; |
3217 | return Opcode; |
3218 | case AArch64::LDRWui: |
3219 | case AArch64::LDRSWui: |
3220 | case AArch64::STRWui: |
3221 | Scale = 4; |
3222 | return Opcode; |
3223 | case AArch64::LDRHui: |
3224 | case AArch64::STRHui: |
3225 | case AArch64::LDRHHui: |
3226 | case AArch64::STRHHui: |
3227 | case AArch64::LDRSHXui: |
3228 | case AArch64::LDRSHWui: |
3229 | Scale = 2; |
3230 | return Opcode; |
3231 | case AArch64::LDRBui: |
3232 | case AArch64::LDRBBui: |
3233 | case AArch64::LDRSBXui: |
3234 | case AArch64::LDRSBWui: |
3235 | case AArch64::STRBui: |
3236 | case AArch64::STRBBui: |
3237 | Scale = 1; |
3238 | return Opcode; |
3239 | } |
3240 | } |
3241 | |
3242 | // Given an opcode for an instruction with a [Reg, #Imm] addressing mode, return |
3243 | // the opcode of an instruction performing the same operation, but using the |
3244 | // [Reg, #Imm] addressing mode with unscaled offset. |
3245 | unsigned unscaledOffsetOpcode(unsigned Opcode) { |
3246 | switch (Opcode) { |
3247 | default: |
3248 | llvm_unreachable("Address folding not implemented for instruction" ); |
3249 | |
3250 | case AArch64::LDURQi: |
3251 | case AArch64::STURQi: |
3252 | case AArch64::LDURDi: |
3253 | case AArch64::STURDi: |
3254 | case AArch64::LDURXi: |
3255 | case AArch64::STURXi: |
3256 | case AArch64::LDURWi: |
3257 | case AArch64::LDURSWi: |
3258 | case AArch64::STURWi: |
3259 | case AArch64::LDURHi: |
3260 | case AArch64::STURHi: |
3261 | case AArch64::LDURHHi: |
3262 | case AArch64::STURHHi: |
3263 | case AArch64::LDURSHXi: |
3264 | case AArch64::LDURSHWi: |
3265 | case AArch64::LDURBi: |
3266 | case AArch64::STURBi: |
3267 | case AArch64::LDURBBi: |
3268 | case AArch64::STURBBi: |
3269 | case AArch64::LDURSBWi: |
3270 | case AArch64::LDURSBXi: |
3271 | return Opcode; |
3272 | case AArch64::LDRQui: |
3273 | return AArch64::LDURQi; |
3274 | case AArch64::STRQui: |
3275 | return AArch64::STURQi; |
3276 | case AArch64::LDRDui: |
3277 | return AArch64::LDURDi; |
3278 | case AArch64::STRDui: |
3279 | return AArch64::STURDi; |
3280 | case AArch64::LDRXui: |
3281 | return AArch64::LDURXi; |
3282 | case AArch64::STRXui: |
3283 | return AArch64::STURXi; |
3284 | case AArch64::LDRWui: |
3285 | return AArch64::LDURWi; |
3286 | case AArch64::LDRSWui: |
3287 | return AArch64::LDURSWi; |
3288 | case AArch64::STRWui: |
3289 | return AArch64::STURWi; |
3290 | case AArch64::LDRHui: |
3291 | return AArch64::LDURHi; |
3292 | case AArch64::STRHui: |
3293 | return AArch64::STURHi; |
3294 | case AArch64::LDRHHui: |
3295 | return AArch64::LDURHHi; |
3296 | case AArch64::STRHHui: |
3297 | return AArch64::STURHHi; |
3298 | case AArch64::LDRSHXui: |
3299 | return AArch64::LDURSHXi; |
3300 | case AArch64::LDRSHWui: |
3301 | return AArch64::LDURSHWi; |
3302 | case AArch64::LDRBBui: |
3303 | return AArch64::LDURBBi; |
3304 | case AArch64::LDRBui: |
3305 | return AArch64::LDURBi; |
3306 | case AArch64::STRBBui: |
3307 | return AArch64::STURBBi; |
3308 | case AArch64::STRBui: |
3309 | return AArch64::STURBi; |
3310 | case AArch64::LDRSBWui: |
3311 | return AArch64::LDURSBWi; |
3312 | case AArch64::LDRSBXui: |
3313 | return AArch64::LDURSBXi; |
3314 | } |
3315 | } |
3316 | |
3317 | // Given the opcode of a memory load/store instruction, return the opcode of an |
3318 | // instruction performing the same operation, but using |
3319 | // the [Reg, Reg, {s,u}xtw #N] addressing mode with sign-/zero-extend of the |
3320 | // offset register. |
3321 | static unsigned offsetExtendOpcode(unsigned Opcode) { |
3322 | switch (Opcode) { |
3323 | default: |
3324 | llvm_unreachable("Address folding not implemented for instruction" ); |
3325 | |
3326 | case AArch64::LDRQroX: |
3327 | case AArch64::LDURQi: |
3328 | case AArch64::LDRQui: |
3329 | return AArch64::LDRQroW; |
3330 | case AArch64::STRQroX: |
3331 | case AArch64::STURQi: |
3332 | case AArch64::STRQui: |
3333 | return AArch64::STRQroW; |
3334 | case AArch64::LDRDroX: |
3335 | case AArch64::LDURDi: |
3336 | case AArch64::LDRDui: |
3337 | return AArch64::LDRDroW; |
3338 | case AArch64::STRDroX: |
3339 | case AArch64::STURDi: |
3340 | case AArch64::STRDui: |
3341 | return AArch64::STRDroW; |
3342 | case AArch64::LDRXroX: |
3343 | case AArch64::LDURXi: |
3344 | case AArch64::LDRXui: |
3345 | return AArch64::LDRXroW; |
3346 | case AArch64::STRXroX: |
3347 | case AArch64::STURXi: |
3348 | case AArch64::STRXui: |
3349 | return AArch64::STRXroW; |
3350 | case AArch64::LDRWroX: |
3351 | case AArch64::LDURWi: |
3352 | case AArch64::LDRWui: |
3353 | return AArch64::LDRWroW; |
3354 | case AArch64::LDRSWroX: |
3355 | case AArch64::LDURSWi: |
3356 | case AArch64::LDRSWui: |
3357 | return AArch64::LDRSWroW; |
3358 | case AArch64::STRWroX: |
3359 | case AArch64::STURWi: |
3360 | case AArch64::STRWui: |
3361 | return AArch64::STRWroW; |
3362 | case AArch64::LDRHroX: |
3363 | case AArch64::LDURHi: |
3364 | case AArch64::LDRHui: |
3365 | return AArch64::LDRHroW; |
3366 | case AArch64::STRHroX: |
3367 | case AArch64::STURHi: |
3368 | case AArch64::STRHui: |
3369 | return AArch64::STRHroW; |
3370 | case AArch64::LDRHHroX: |
3371 | case AArch64::LDURHHi: |
3372 | case AArch64::LDRHHui: |
3373 | return AArch64::LDRHHroW; |
3374 | case AArch64::STRHHroX: |
3375 | case AArch64::STURHHi: |
3376 | case AArch64::STRHHui: |
3377 | return AArch64::STRHHroW; |
3378 | case AArch64::LDRSHXroX: |
3379 | case AArch64::LDURSHXi: |
3380 | case AArch64::LDRSHXui: |
3381 | return AArch64::LDRSHXroW; |
3382 | case AArch64::LDRSHWroX: |
3383 | case AArch64::LDURSHWi: |
3384 | case AArch64::LDRSHWui: |
3385 | return AArch64::LDRSHWroW; |
3386 | case AArch64::LDRBroX: |
3387 | case AArch64::LDURBi: |
3388 | case AArch64::LDRBui: |
3389 | return AArch64::LDRBroW; |
3390 | case AArch64::LDRBBroX: |
3391 | case AArch64::LDURBBi: |
3392 | case AArch64::LDRBBui: |
3393 | return AArch64::LDRBBroW; |
3394 | case AArch64::LDRSBXroX: |
3395 | case AArch64::LDURSBXi: |
3396 | case AArch64::LDRSBXui: |
3397 | return AArch64::LDRSBXroW; |
3398 | case AArch64::LDRSBWroX: |
3399 | case AArch64::LDURSBWi: |
3400 | case AArch64::LDRSBWui: |
3401 | return AArch64::LDRSBWroW; |
3402 | case AArch64::STRBroX: |
3403 | case AArch64::STURBi: |
3404 | case AArch64::STRBui: |
3405 | return AArch64::STRBroW; |
3406 | case AArch64::STRBBroX: |
3407 | case AArch64::STURBBi: |
3408 | case AArch64::STRBBui: |
3409 | return AArch64::STRBBroW; |
3410 | } |
3411 | } |
3412 | |
3413 | MachineInstr *AArch64InstrInfo::emitLdStWithAddr(MachineInstr &MemI, |
3414 | const ExtAddrMode &AM) const { |
3415 | |
3416 | const DebugLoc &DL = MemI.getDebugLoc(); |
3417 | MachineBasicBlock &MBB = *MemI.getParent(); |
3418 | MachineRegisterInfo &MRI = MemI.getMF()->getRegInfo(); |
3419 | |
3420 | if (AM.Form == ExtAddrMode::Formula::Basic) { |
3421 | if (AM.ScaledReg) { |
3422 | // The new instruction will be in the form `ldr Rt, [Xn, Xm, lsl #imm]`. |
3423 | unsigned Opcode = regOffsetOpcode(Opcode: MemI.getOpcode()); |
3424 | MRI.constrainRegClass(Reg: AM.BaseReg, RC: &AArch64::GPR64spRegClass); |
3425 | auto B = BuildMI(BB&: MBB, I&: MemI, MIMD: DL, MCID: get(Opcode)) |
3426 | .addReg(RegNo: MemI.getOperand(i: 0).getReg(), |
3427 | flags: MemI.mayLoad() ? RegState::Define : 0) |
3428 | .addReg(RegNo: AM.BaseReg) |
3429 | .addReg(RegNo: AM.ScaledReg) |
3430 | .addImm(Val: 0) |
3431 | .addImm(Val: AM.Scale > 1) |
3432 | .setMemRefs(MemI.memoperands()) |
3433 | .setMIFlags(MemI.getFlags()); |
3434 | return B.getInstr(); |
3435 | } |
3436 | |
3437 | assert(AM.ScaledReg == 0 && AM.Scale == 0 && |
3438 | "Addressing mode not supported for folding" ); |
3439 | |
3440 | // The new instruction will be in the form `ld[u]r Rt, [Xn, #imm]`. |
3441 | unsigned Scale = 1; |
3442 | unsigned Opcode = MemI.getOpcode(); |
3443 | if (isInt<9>(x: AM.Displacement)) |
3444 | Opcode = unscaledOffsetOpcode(Opcode); |
3445 | else |
3446 | Opcode = scaledOffsetOpcode(Opcode, Scale); |
3447 | |
3448 | auto B = BuildMI(BB&: MBB, I&: MemI, MIMD: DL, MCID: get(Opcode)) |
3449 | .addReg(RegNo: MemI.getOperand(i: 0).getReg(), |
3450 | flags: MemI.mayLoad() ? RegState::Define : 0) |
3451 | .addReg(RegNo: AM.BaseReg) |
3452 | .addImm(Val: AM.Displacement / Scale) |
3453 | .setMemRefs(MemI.memoperands()) |
3454 | .setMIFlags(MemI.getFlags()); |
3455 | return B.getInstr(); |
3456 | } |
3457 | |
3458 | if (AM.Form == ExtAddrMode::Formula::SExtScaledReg || |
3459 | AM.Form == ExtAddrMode::Formula::ZExtScaledReg) { |
3460 | // The new instruction will be in the form `ldr Rt, [Xn, Wm, {s,u}xtw #N]`. |
3461 | assert(AM.ScaledReg && !AM.Displacement && |
3462 | "Address offset can be a register or an immediate, but not both" ); |
3463 | unsigned Opcode = offsetExtendOpcode(Opcode: MemI.getOpcode()); |
3464 | MRI.constrainRegClass(Reg: AM.BaseReg, RC: &AArch64::GPR64spRegClass); |
3465 | // Make sure the offset register is in the correct register class. |
3466 | Register OffsetReg = AM.ScaledReg; |
3467 | const TargetRegisterClass *RC = MRI.getRegClass(Reg: OffsetReg); |
3468 | if (RC->hasSuperClassEq(RC: &AArch64::GPR64RegClass)) { |
3469 | OffsetReg = MRI.createVirtualRegister(RegClass: &AArch64::GPR32RegClass); |
3470 | BuildMI(BB&: MBB, I&: MemI, MIMD: DL, MCID: get(Opcode: TargetOpcode::COPY), DestReg: OffsetReg) |
3471 | .addReg(RegNo: AM.ScaledReg, flags: 0, SubReg: AArch64::sub_32); |
3472 | } |
3473 | auto B = BuildMI(BB&: MBB, I&: MemI, MIMD: DL, MCID: get(Opcode)) |
3474 | .addReg(RegNo: MemI.getOperand(i: 0).getReg(), |
3475 | flags: MemI.mayLoad() ? RegState::Define : 0) |
3476 | .addReg(RegNo: AM.BaseReg) |
3477 | .addReg(RegNo: OffsetReg) |
3478 | .addImm(Val: AM.Form == ExtAddrMode::Formula::SExtScaledReg) |
3479 | .addImm(Val: AM.Scale != 1) |
3480 | .setMemRefs(MemI.memoperands()) |
3481 | .setMIFlags(MemI.getFlags()); |
3482 | |
3483 | return B.getInstr(); |
3484 | } |
3485 | |
3486 | llvm_unreachable( |
3487 | "Function must not be called with an addressing mode it can't handle" ); |
3488 | } |
3489 | |
3490 | bool AArch64InstrInfo::getMemOperandWithOffsetWidth( |
3491 | const MachineInstr &LdSt, const MachineOperand *&BaseOp, int64_t &Offset, |
3492 | bool &OffsetIsScalable, TypeSize &Width, |
3493 | const TargetRegisterInfo *TRI) const { |
3494 | assert(LdSt.mayLoadOrStore() && "Expected a memory operation." ); |
3495 | // Handle only loads/stores with base register followed by immediate offset. |
3496 | if (LdSt.getNumExplicitOperands() == 3) { |
3497 | // Non-paired instruction (e.g., ldr x1, [x0, #8]). |
3498 | if ((!LdSt.getOperand(i: 1).isReg() && !LdSt.getOperand(i: 1).isFI()) || |
3499 | !LdSt.getOperand(i: 2).isImm()) |
3500 | return false; |
3501 | } else if (LdSt.getNumExplicitOperands() == 4) { |
3502 | // Paired instruction (e.g., ldp x1, x2, [x0, #8]). |
3503 | if (!LdSt.getOperand(i: 1).isReg() || |
3504 | (!LdSt.getOperand(i: 2).isReg() && !LdSt.getOperand(i: 2).isFI()) || |
3505 | !LdSt.getOperand(i: 3).isImm()) |
3506 | return false; |
3507 | } else |
3508 | return false; |
3509 | |
3510 | // Get the scaling factor for the instruction and set the width for the |
3511 | // instruction. |
3512 | TypeSize Scale(0U, false); |
3513 | int64_t Dummy1, Dummy2; |
3514 | |
3515 | // If this returns false, then it's an instruction we don't want to handle. |
3516 | if (!getMemOpInfo(Opcode: LdSt.getOpcode(), Scale, Width, MinOffset&: Dummy1, MaxOffset&: Dummy2)) |
3517 | return false; |
3518 | |
3519 | // Compute the offset. Offset is calculated as the immediate operand |
3520 | // multiplied by the scaling factor. Unscaled instructions have scaling factor |
3521 | // set to 1. |
3522 | if (LdSt.getNumExplicitOperands() == 3) { |
3523 | BaseOp = &LdSt.getOperand(i: 1); |
3524 | Offset = LdSt.getOperand(i: 2).getImm() * Scale.getKnownMinValue(); |
3525 | } else { |
3526 | assert(LdSt.getNumExplicitOperands() == 4 && "invalid number of operands" ); |
3527 | BaseOp = &LdSt.getOperand(i: 2); |
3528 | Offset = LdSt.getOperand(i: 3).getImm() * Scale.getKnownMinValue(); |
3529 | } |
3530 | OffsetIsScalable = Scale.isScalable(); |
3531 | |
3532 | if (!BaseOp->isReg() && !BaseOp->isFI()) |
3533 | return false; |
3534 | |
3535 | return true; |
3536 | } |
3537 | |
3538 | MachineOperand & |
3539 | AArch64InstrInfo::getMemOpBaseRegImmOfsOffsetOperand(MachineInstr &LdSt) const { |
3540 | assert(LdSt.mayLoadOrStore() && "Expected a memory operation." ); |
3541 | MachineOperand &OfsOp = LdSt.getOperand(i: LdSt.getNumExplicitOperands() - 1); |
3542 | assert(OfsOp.isImm() && "Offset operand wasn't immediate." ); |
3543 | return OfsOp; |
3544 | } |
3545 | |
3546 | bool AArch64InstrInfo::getMemOpInfo(unsigned Opcode, TypeSize &Scale, |
3547 | TypeSize &Width, int64_t &MinOffset, |
3548 | int64_t &MaxOffset) { |
3549 | switch (Opcode) { |
3550 | // Not a memory operation or something we want to handle. |
3551 | default: |
3552 | Scale = TypeSize::getFixed(ExactSize: 0); |
3553 | Width = TypeSize::getFixed(ExactSize: 0); |
3554 | MinOffset = MaxOffset = 0; |
3555 | return false; |
3556 | // LDR / STR |
3557 | case AArch64::LDRQui: |
3558 | case AArch64::STRQui: |
3559 | Scale = TypeSize::getFixed(ExactSize: 16); |
3560 | Width = TypeSize::getFixed(ExactSize: 16); |
3561 | MinOffset = 0; |
3562 | MaxOffset = 4095; |
3563 | break; |
3564 | case AArch64::LDRXui: |
3565 | case AArch64::LDRDui: |
3566 | case AArch64::STRXui: |
3567 | case AArch64::STRDui: |
3568 | case AArch64::PRFMui: |
3569 | Scale = TypeSize::getFixed(ExactSize: 8); |
3570 | Width = TypeSize::getFixed(ExactSize: 8); |
3571 | MinOffset = 0; |
3572 | MaxOffset = 4095; |
3573 | break; |
3574 | case AArch64::LDRWui: |
3575 | case AArch64::LDRSui: |
3576 | case AArch64::LDRSWui: |
3577 | case AArch64::STRWui: |
3578 | case AArch64::STRSui: |
3579 | Scale = TypeSize::getFixed(ExactSize: 4); |
3580 | Width = TypeSize::getFixed(ExactSize: 4); |
3581 | MinOffset = 0; |
3582 | MaxOffset = 4095; |
3583 | break; |
3584 | case AArch64::LDRHui: |
3585 | case AArch64::LDRHHui: |
3586 | case AArch64::LDRSHWui: |
3587 | case AArch64::LDRSHXui: |
3588 | case AArch64::STRHui: |
3589 | case AArch64::STRHHui: |
3590 | Scale = TypeSize::getFixed(ExactSize: 2); |
3591 | Width = TypeSize::getFixed(ExactSize: 2); |
3592 | MinOffset = 0; |
3593 | MaxOffset = 4095; |
3594 | break; |
3595 | case AArch64::LDRBui: |
3596 | case AArch64::LDRBBui: |
3597 | case AArch64::LDRSBWui: |
3598 | case AArch64::LDRSBXui: |
3599 | case AArch64::STRBui: |
3600 | case AArch64::STRBBui: |
3601 | Scale = TypeSize::getFixed(ExactSize: 1); |
3602 | Width = TypeSize::getFixed(ExactSize: 1); |
3603 | MinOffset = 0; |
3604 | MaxOffset = 4095; |
3605 | break; |
3606 | // post/pre inc |
3607 | case AArch64::STRQpre: |
3608 | case AArch64::LDRQpost: |
3609 | Scale = TypeSize::getFixed(ExactSize: 1); |
3610 | Width = TypeSize::getFixed(ExactSize: 16); |
3611 | MinOffset = -256; |
3612 | MaxOffset = 255; |
3613 | break; |
3614 | case AArch64::STRXpre: |
3615 | case AArch64::STRDpre: |
3616 | case AArch64::LDRXpost: |
3617 | case AArch64::LDRDpost: |
3618 | Scale = TypeSize::getFixed(ExactSize: 1); |
3619 | Width = TypeSize::getFixed(ExactSize: 8); |
3620 | MinOffset = -256; |
3621 | MaxOffset = 255; |
3622 | break; |
3623 | case AArch64::STRWpost: |
3624 | case AArch64::LDRWpost: |
3625 | Scale = TypeSize::getFixed(ExactSize: 4); |
3626 | Width = TypeSize::getFixed(ExactSize: 32); |
3627 | MinOffset = -256; |
3628 | MaxOffset = 255; |
3629 | break; |
3630 | // Unscaled |
3631 | case AArch64::LDURQi: |
3632 | case AArch64::STURQi: |
3633 | Scale = TypeSize::getFixed(ExactSize: 1); |
3634 | Width = TypeSize::getFixed(ExactSize: 16); |
3635 | MinOffset = -256; |
3636 | MaxOffset = 255; |
3637 | break; |
3638 | case AArch64::LDURXi: |
3639 | case AArch64::LDURDi: |
3640 | case AArch64::LDAPURXi: |
3641 | case AArch64::STURXi: |
3642 | case AArch64::STURDi: |
3643 | case AArch64::STLURXi: |
3644 | case AArch64::PRFUMi: |
3645 | Scale = TypeSize::getFixed(ExactSize: 1); |
3646 | Width = TypeSize::getFixed(ExactSize: 8); |
3647 | MinOffset = -256; |
3648 | MaxOffset = 255; |
3649 | break; |
3650 | case AArch64::LDURWi: |
3651 | case AArch64::LDURSi: |
3652 | case AArch64::LDURSWi: |
3653 | case AArch64::LDAPURi: |
3654 | case AArch64::LDAPURSWi: |
3655 | case AArch64::STURWi: |
3656 | case AArch64::STURSi: |
3657 | case AArch64::STLURWi: |
3658 | Scale = TypeSize::getFixed(ExactSize: 1); |
3659 | Width = TypeSize::getFixed(ExactSize: 4); |
3660 | MinOffset = -256; |
3661 | MaxOffset = 255; |
3662 | break; |
3663 | case AArch64::LDURHi: |
3664 | case AArch64::LDURHHi: |
3665 | case AArch64::LDURSHXi: |
3666 | case AArch64::LDURSHWi: |
3667 | case AArch64::LDAPURHi: |
3668 | case AArch64::LDAPURSHWi: |
3669 | case AArch64::LDAPURSHXi: |
3670 | case AArch64::STURHi: |
3671 | case AArch64::STURHHi: |
3672 | case AArch64::STLURHi: |
3673 | Scale = TypeSize::getFixed(ExactSize: 1); |
3674 | Width = TypeSize::getFixed(ExactSize: 2); |
3675 | MinOffset = -256; |
3676 | MaxOffset = 255; |
3677 | break; |
3678 | case AArch64::LDURBi: |
3679 | case AArch64::LDURBBi: |
3680 | case AArch64::LDURSBXi: |
3681 | case AArch64::LDURSBWi: |
3682 | case AArch64::LDAPURBi: |
3683 | case AArch64::LDAPURSBWi: |
3684 | case AArch64::LDAPURSBXi: |
3685 | case AArch64::STURBi: |
3686 | case AArch64::STURBBi: |
3687 | case AArch64::STLURBi: |
3688 | Scale = TypeSize::getFixed(ExactSize: 1); |
3689 | Width = TypeSize::getFixed(ExactSize: 1); |
3690 | MinOffset = -256; |
3691 | MaxOffset = 255; |
3692 | break; |
3693 | // LDP / STP |
3694 | case AArch64::LDPQi: |
3695 | case AArch64::LDNPQi: |
3696 | case AArch64::STPQi: |
3697 | case AArch64::STNPQi: |
3698 | Scale = TypeSize::getFixed(ExactSize: 16); |
3699 | Width = TypeSize::getFixed(ExactSize: 32); |
3700 | MinOffset = -64; |
3701 | MaxOffset = 63; |
3702 | break; |
3703 | case AArch64::LDPXi: |
3704 | case AArch64::LDPDi: |
3705 | case AArch64::LDNPXi: |
3706 | case AArch64::LDNPDi: |
3707 | case AArch64::STPXi: |
3708 | case AArch64::STPDi: |
3709 | case AArch64::STNPXi: |
3710 | case AArch64::STNPDi: |
3711 | Scale = TypeSize::getFixed(ExactSize: 8); |
3712 | Width = TypeSize::getFixed(ExactSize: 16); |
3713 | MinOffset = -64; |
3714 | MaxOffset = 63; |
3715 | break; |
3716 | case AArch64::LDPWi: |
3717 | case AArch64::LDPSi: |
3718 | case AArch64::LDNPWi: |
3719 | case AArch64::LDNPSi: |
3720 | case AArch64::STPWi: |
3721 | case AArch64::STPSi: |
3722 | case AArch64::STNPWi: |
3723 | case AArch64::STNPSi: |
3724 | Scale = TypeSize::getFixed(ExactSize: 4); |
3725 | Width = TypeSize::getFixed(ExactSize: 8); |
3726 | MinOffset = -64; |
3727 | MaxOffset = 63; |
3728 | break; |
3729 | // pre/post inc |
3730 | case AArch64::STPQpre: |
3731 | case AArch64::LDPQpost: |
3732 | Scale = TypeSize::getFixed(ExactSize: 16); |
3733 | Width = TypeSize::getFixed(ExactSize: 16); |
3734 | MinOffset = -1024; |
3735 | MaxOffset = 1008; |
3736 | break; |
3737 | case AArch64::STPXpre: |
3738 | case AArch64::LDPXpost: |
3739 | case AArch64::STPDpre: |
3740 | case AArch64::LDPDpost: |
3741 | Scale = TypeSize::getFixed(ExactSize: 8); |
3742 | Width = TypeSize::getFixed(ExactSize: 8); |
3743 | MinOffset = -512; |
3744 | MaxOffset = 504; |
3745 | break; |
3746 | case AArch64::StoreSwiftAsyncContext: |
3747 | // Store is an STRXui, but there might be an ADDXri in the expansion too. |
3748 | Scale = TypeSize::getFixed(ExactSize: 1); |
3749 | Width = TypeSize::getFixed(ExactSize: 8); |
3750 | MinOffset = 0; |
3751 | MaxOffset = 4095; |
3752 | break; |
3753 | case AArch64::ADDG: |
3754 | Scale = TypeSize::getFixed(ExactSize: 16); |
3755 | Width = TypeSize::getFixed(ExactSize: 0); |
3756 | MinOffset = 0; |
3757 | MaxOffset = 63; |
3758 | break; |
3759 | case AArch64::TAGPstack: |
3760 | Scale = TypeSize::getFixed(ExactSize: 16); |
3761 | Width = TypeSize::getFixed(ExactSize: 0); |
3762 | // TAGP with a negative offset turns into SUBP, which has a maximum offset |
3763 | // of 63 (not 64!). |
3764 | MinOffset = -63; |
3765 | MaxOffset = 63; |
3766 | break; |
3767 | case AArch64::LDG: |
3768 | case AArch64::STGi: |
3769 | case AArch64::STZGi: |
3770 | Scale = TypeSize::getFixed(ExactSize: 16); |
3771 | Width = TypeSize::getFixed(ExactSize: 16); |
3772 | MinOffset = -256; |
3773 | MaxOffset = 255; |
3774 | break; |
3775 | // SVE |
3776 | case AArch64::STR_ZZZZXI: |
3777 | case AArch64::LDR_ZZZZXI: |
3778 | Scale = TypeSize::getScalable(MinimumSize: 16); |
3779 | Width = TypeSize::getScalable(MinimumSize: 16 * 4); |
3780 | MinOffset = -256; |
3781 | MaxOffset = 252; |
3782 | break; |
3783 | case AArch64::STR_ZZZXI: |
3784 | case AArch64::LDR_ZZZXI: |
3785 | Scale = TypeSize::getScalable(MinimumSize: 16); |
3786 | Width = TypeSize::getScalable(MinimumSize: 16 * 3); |
3787 | MinOffset = -256; |
3788 | MaxOffset = 253; |
3789 | break; |
3790 | case AArch64::STR_ZZXI: |
3791 | case AArch64::LDR_ZZXI: |
3792 | Scale = TypeSize::getScalable(MinimumSize: 16); |
3793 | Width = TypeSize::getScalable(MinimumSize: 16 * 2); |
3794 | MinOffset = -256; |
3795 | MaxOffset = 254; |
3796 | break; |
3797 | case AArch64::LDR_PXI: |
3798 | case AArch64::STR_PXI: |
3799 | Scale = TypeSize::getScalable(MinimumSize: 2); |
3800 | Width = TypeSize::getScalable(MinimumSize: 2); |
3801 | MinOffset = -256; |
3802 | MaxOffset = 255; |
3803 | break; |
3804 | case AArch64::LDR_PPXI: |
3805 | case AArch64::STR_PPXI: |
3806 | Scale = TypeSize::getScalable(MinimumSize: 2); |
3807 | Width = TypeSize::getScalable(MinimumSize: 2 * 2); |
3808 | MinOffset = -256; |
3809 | MaxOffset = 254; |
3810 | break; |
3811 | case AArch64::LDR_ZXI: |
3812 | case AArch64::STR_ZXI: |
3813 | Scale = TypeSize::getScalable(MinimumSize: 16); |
3814 | Width = TypeSize::getScalable(MinimumSize: 16); |
3815 | MinOffset = -256; |
3816 | MaxOffset = 255; |
3817 | break; |
3818 | case AArch64::LD1B_IMM: |
3819 | case AArch64::LD1H_IMM: |
3820 | case AArch64::LD1W_IMM: |
3821 | case AArch64::LD1D_IMM: |
3822 | case AArch64::LDNT1B_ZRI: |
3823 | case AArch64::LDNT1H_ZRI: |
3824 | case AArch64::LDNT1W_ZRI: |
3825 | case AArch64::LDNT1D_ZRI: |
3826 | case AArch64::ST1B_IMM: |
3827 | case AArch64::ST1H_IMM: |
3828 | case AArch64::ST1W_IMM: |
3829 | case AArch64::ST1D_IMM: |
3830 | case AArch64::STNT1B_ZRI: |
3831 | case AArch64::STNT1H_ZRI: |
3832 | case AArch64::STNT1W_ZRI: |
3833 | case AArch64::STNT1D_ZRI: |
3834 | case AArch64::LDNF1B_IMM: |
3835 | case AArch64::LDNF1H_IMM: |
3836 | case AArch64::LDNF1W_IMM: |
3837 | case AArch64::LDNF1D_IMM: |
3838 | // A full vectors worth of data |
3839 | // Width = mbytes * elements |
3840 | Scale = TypeSize::getScalable(MinimumSize: 16); |
3841 | Width = TypeSize::getScalable(MinimumSize: 16); |
3842 | MinOffset = -8; |
3843 | MaxOffset = 7; |
3844 | break; |
3845 | case AArch64::LD2B_IMM: |
3846 | case AArch64::LD2H_IMM: |
3847 | case AArch64::LD2W_IMM: |
3848 | case AArch64::LD2D_IMM: |
3849 | case AArch64::ST2B_IMM: |
3850 | case AArch64::ST2H_IMM: |
3851 | case AArch64::ST2W_IMM: |
3852 | case AArch64::ST2D_IMM: |
3853 | Scale = TypeSize::getScalable(MinimumSize: 32); |
3854 | Width = TypeSize::getScalable(MinimumSize: 16 * 2); |
3855 | MinOffset = -8; |
3856 | MaxOffset = 7; |
3857 | break; |
3858 | case AArch64::LD3B_IMM: |
3859 | case AArch64::LD3H_IMM: |
3860 | case AArch64::LD3W_IMM: |
3861 | case AArch64::LD3D_IMM: |
3862 | case AArch64::ST3B_IMM: |
3863 | case AArch64::ST3H_IMM: |
3864 | case AArch64::ST3W_IMM: |
3865 | case AArch64::ST3D_IMM: |
3866 | Scale = TypeSize::getScalable(MinimumSize: 48); |
3867 | Width = TypeSize::getScalable(MinimumSize: 16 * 3); |
3868 | MinOffset = -8; |
3869 | MaxOffset = 7; |
3870 | break; |
3871 | case AArch64::LD4B_IMM: |
3872 | case AArch64::LD4H_IMM: |
3873 | case AArch64::LD4W_IMM: |
3874 | case AArch64::LD4D_IMM: |
3875 | case AArch64::ST4B_IMM: |
3876 | case AArch64::ST4H_IMM: |
3877 | case AArch64::ST4W_IMM: |
3878 | case AArch64::ST4D_IMM: |
3879 | Scale = TypeSize::getScalable(MinimumSize: 64); |
3880 | Width = TypeSize::getScalable(MinimumSize: 16 * 4); |
3881 | MinOffset = -8; |
3882 | MaxOffset = 7; |
3883 | break; |
3884 | case AArch64::LD1B_H_IMM: |
3885 | case AArch64::LD1SB_H_IMM: |
3886 | case AArch64::LD1H_S_IMM: |
3887 | case AArch64::LD1SH_S_IMM: |
3888 | case AArch64::LD1W_D_IMM: |
3889 | case AArch64::LD1SW_D_IMM: |
3890 | case AArch64::ST1B_H_IMM: |
3891 | case AArch64::ST1H_S_IMM: |
3892 | case AArch64::ST1W_D_IMM: |
3893 | case AArch64::LDNF1B_H_IMM: |
3894 | case AArch64::LDNF1SB_H_IMM: |
3895 | case AArch64::LDNF1H_S_IMM: |
3896 | case AArch64::LDNF1SH_S_IMM: |
3897 | case AArch64::LDNF1W_D_IMM: |
3898 | case AArch64::LDNF1SW_D_IMM: |
3899 | // A half vector worth of data |
3900 | // Width = mbytes * elements |
3901 | Scale = TypeSize::getScalable(MinimumSize: 8); |
3902 | Width = TypeSize::getScalable(MinimumSize: 8); |
3903 | MinOffset = -8; |
3904 | MaxOffset = 7; |
3905 | break; |
3906 | case AArch64::LD1B_S_IMM: |
3907 | case AArch64::LD1SB_S_IMM: |
3908 | case AArch64::LD1H_D_IMM: |
3909 | case AArch64::LD1SH_D_IMM: |
3910 | case AArch64::ST1B_S_IMM: |
3911 | case AArch64::ST1H_D_IMM: |
3912 | case AArch64::LDNF1B_S_IMM: |
3913 | case AArch64::LDNF1SB_S_IMM: |
3914 | case AArch64::LDNF1H_D_IMM: |
3915 | case AArch64::LDNF1SH_D_IMM: |
3916 | // A quarter vector worth of data |
3917 | // Width = mbytes * elements |
3918 | Scale = TypeSize::getScalable(MinimumSize: 4); |
3919 | Width = TypeSize::getScalable(MinimumSize: 4); |
3920 | MinOffset = -8; |
3921 | MaxOffset = 7; |
3922 | break; |
3923 | case AArch64::LD1B_D_IMM: |
3924 | case AArch64::LD1SB_D_IMM: |
3925 | case AArch64::ST1B_D_IMM: |
3926 | case AArch64::LDNF1B_D_IMM: |
3927 | case AArch64::LDNF1SB_D_IMM: |
3928 | // A eighth vector worth of data |
3929 | // Width = mbytes * elements |
3930 | Scale = TypeSize::getScalable(MinimumSize: 2); |
3931 | Width = TypeSize::getScalable(MinimumSize: 2); |
3932 | MinOffset = -8; |
3933 | MaxOffset = 7; |
3934 | break; |
3935 | case AArch64::ST2Gi: |
3936 | case AArch64::STZ2Gi: |
3937 | Scale = TypeSize::getFixed(ExactSize: 16); |
3938 | Width = TypeSize::getFixed(ExactSize: 32); |
3939 | MinOffset = -256; |
3940 | MaxOffset = 255; |
3941 | break; |
3942 | case AArch64::STGPi: |
3943 | Scale = TypeSize::getFixed(ExactSize: 16); |
3944 | Width = TypeSize::getFixed(ExactSize: 16); |
3945 | MinOffset = -64; |
3946 | MaxOffset = 63; |
3947 | break; |
3948 | case AArch64::LD1RB_IMM: |
3949 | case AArch64::LD1RB_H_IMM: |
3950 | case AArch64::LD1RB_S_IMM: |
3951 | case AArch64::LD1RB_D_IMM: |
3952 | case AArch64::LD1RSB_H_IMM: |
3953 | case AArch64::LD1RSB_S_IMM: |
3954 | case AArch64::LD1RSB_D_IMM: |
3955 | Scale = TypeSize::getFixed(ExactSize: 1); |
3956 | Width = TypeSize::getFixed(ExactSize: 1); |
3957 | MinOffset = 0; |
3958 | MaxOffset = 63; |
3959 | break; |
3960 | case AArch64::LD1RH_IMM: |
3961 | case AArch64::LD1RH_S_IMM: |
3962 | case AArch64::LD1RH_D_IMM: |
3963 | case AArch64::LD1RSH_S_IMM: |
3964 | case AArch64::LD1RSH_D_IMM: |
3965 | Scale = TypeSize::getFixed(ExactSize: 2); |
3966 | Width = TypeSize::getFixed(ExactSize: 2); |
3967 | MinOffset = 0; |
3968 | MaxOffset = 63; |
3969 | break; |
3970 | case AArch64::LD1RW_IMM: |
3971 | case AArch64::LD1RW_D_IMM: |
3972 | case AArch64::LD1RSW_IMM: |
3973 | Scale = TypeSize::getFixed(ExactSize: 4); |
3974 | Width = TypeSize::getFixed(ExactSize: 4); |
3975 | MinOffset = 0; |
3976 | MaxOffset = 63; |
3977 | break; |
3978 | case AArch64::LD1RD_IMM: |
3979 | Scale = TypeSize::getFixed(ExactSize: 8); |
3980 | Width = TypeSize::getFixed(ExactSize: 8); |
3981 | MinOffset = 0; |
3982 | MaxOffset = 63; |
3983 | break; |
3984 | } |
3985 | |
3986 | return true; |
3987 | } |
3988 | |
3989 | // Scaling factor for unscaled load or store. |
3990 | int AArch64InstrInfo::getMemScale(unsigned Opc) { |
3991 | switch (Opc) { |
3992 | default: |
3993 | llvm_unreachable("Opcode has unknown scale!" ); |
3994 | case AArch64::LDRBBui: |
3995 | case AArch64::LDURBBi: |
3996 | case AArch64::LDRSBWui: |
3997 | case AArch64::LDURSBWi: |
3998 | case AArch64::STRBBui: |
3999 | case AArch64::STURBBi: |
4000 | return 1; |
4001 | case AArch64::LDRHHui: |
4002 | case AArch64::LDURHHi: |
4003 | case AArch64::LDRSHWui: |
4004 | case AArch64::LDURSHWi: |
4005 | case AArch64::STRHHui: |
4006 | case AArch64::STURHHi: |
4007 | return 2; |
4008 | case AArch64::LDRSui: |
4009 | case AArch64::LDURSi: |
4010 | case AArch64::LDRSpre: |
4011 | case AArch64::LDRSWui: |
4012 | case AArch64::LDURSWi: |
4013 | case AArch64::LDRSWpre: |
4014 | case AArch64::LDRWpre: |
4015 | case AArch64::LDRWui: |
4016 | case AArch64::LDURWi: |
4017 | case AArch64::STRSui: |
4018 | case AArch64::STURSi: |
4019 | case AArch64::STRSpre: |
4020 | case AArch64::STRWui: |
4021 | case AArch64::STURWi: |
4022 | case AArch64::STRWpre: |
4023 | case AArch64::LDPSi: |
4024 | case AArch64::LDPSWi: |
4025 | case AArch64::LDPWi: |
4026 | case AArch64::STPSi: |
4027 | case AArch64::STPWi: |
4028 | return 4; |
4029 | case AArch64::LDRDui: |
4030 | case AArch64::LDURDi: |
4031 | case AArch64::LDRDpre: |
4032 | case AArch64::LDRXui: |
4033 | case AArch64::LDURXi: |
4034 | case AArch64::LDRXpre: |
4035 | case AArch64::STRDui: |
4036 | case AArch64::STURDi: |
4037 | case AArch64::STRDpre: |
4038 | case AArch64::STRXui: |
4039 | case AArch64::STURXi: |
4040 | case AArch64::STRXpre: |
4041 | case AArch64::LDPDi: |
4042 | case AArch64::LDPXi: |
4043 | case AArch64::STPDi: |
4044 | case AArch64::STPXi: |
4045 | return 8; |
4046 | case AArch64::LDRQui: |
4047 | case AArch64::LDURQi: |
4048 | case AArch64::STRQui: |
4049 | case AArch64::STURQi: |
4050 | case AArch64::STRQpre: |
4051 | case AArch64::LDPQi: |
4052 | case AArch64::LDRQpre: |
4053 | case AArch64::STPQi: |
4054 | case AArch64::STGi: |
4055 | case AArch64::STZGi: |
4056 | case AArch64::ST2Gi: |
4057 | case AArch64::STZ2Gi: |
4058 | case AArch64::STGPi: |
4059 | return 16; |
4060 | } |
4061 | } |
4062 | |
4063 | bool AArch64InstrInfo::isPreLd(const MachineInstr &MI) { |
4064 | switch (MI.getOpcode()) { |
4065 | default: |
4066 | return false; |
4067 | case AArch64::LDRWpre: |
4068 | case AArch64::LDRXpre: |
4069 | case AArch64::LDRSWpre: |
4070 | case AArch64::LDRSpre: |
4071 | case AArch64::LDRDpre: |
4072 | case AArch64::LDRQpre: |
4073 | return true; |
4074 | } |
4075 | } |
4076 | |
4077 | bool AArch64InstrInfo::isPreSt(const MachineInstr &MI) { |
4078 | switch (MI.getOpcode()) { |
4079 | default: |
4080 | return false; |
4081 | case AArch64::STRWpre: |
4082 | case AArch64::STRXpre: |
4083 | case AArch64::STRSpre: |
4084 | case AArch64::STRDpre: |
4085 | case AArch64::STRQpre: |
4086 | return true; |
4087 | } |
4088 | } |
4089 | |
4090 | bool AArch64InstrInfo::isPreLdSt(const MachineInstr &MI) { |
4091 | return isPreLd(MI) || isPreSt(MI); |
4092 | } |
4093 | |
4094 | bool AArch64InstrInfo::isPairedLdSt(const MachineInstr &MI) { |
4095 | switch (MI.getOpcode()) { |
4096 | default: |
4097 | return false; |
4098 | case AArch64::LDPSi: |
4099 | case AArch64::LDPSWi: |
4100 | case AArch64::LDPDi: |
4101 | case AArch64::LDPQi: |
4102 | case AArch64::LDPWi: |
4103 | case AArch64::LDPXi: |
4104 | case AArch64::STPSi: |
4105 | case AArch64::STPDi: |
4106 | case AArch64::STPQi: |
4107 | case AArch64::STPWi: |
4108 | case AArch64::STPXi: |
4109 | case AArch64::STGPi: |
4110 | return true; |
4111 | } |
4112 | } |
4113 | |
4114 | const MachineOperand &AArch64InstrInfo::getLdStBaseOp(const MachineInstr &MI) { |
4115 | unsigned Idx = |
4116 | AArch64InstrInfo::isPairedLdSt(MI) || AArch64InstrInfo::isPreLdSt(MI) ? 2 |
4117 | : 1; |
4118 | return MI.getOperand(i: Idx); |
4119 | } |
4120 | |
4121 | const MachineOperand & |
4122 | AArch64InstrInfo::getLdStOffsetOp(const MachineInstr &MI) { |
4123 | unsigned Idx = |
4124 | AArch64InstrInfo::isPairedLdSt(MI) || AArch64InstrInfo::isPreLdSt(MI) ? 3 |
4125 | : 2; |
4126 | return MI.getOperand(i: Idx); |
4127 | } |
4128 | |
4129 | static const TargetRegisterClass *getRegClass(const MachineInstr &MI, |
4130 | Register Reg) { |
4131 | if (MI.getParent() == nullptr) |
4132 | return nullptr; |
4133 | const MachineFunction *MF = MI.getParent()->getParent(); |
4134 | return MF ? MF->getRegInfo().getRegClassOrNull(Reg) : nullptr; |
4135 | } |
4136 | |
4137 | bool AArch64InstrInfo::isHForm(const MachineInstr &MI) { |
4138 | auto IsHFPR = [&](const MachineOperand &Op) { |
4139 | if (!Op.isReg()) |
4140 | return false; |
4141 | auto Reg = Op.getReg(); |
4142 | if (Reg.isPhysical()) |
4143 | return AArch64::FPR16RegClass.contains(Reg); |
4144 | const TargetRegisterClass *TRC = ::getRegClass(MI, Reg); |
4145 | return TRC == &AArch64::FPR16RegClass || |
4146 | TRC == &AArch64::FPR16_loRegClass; |
4147 | }; |
4148 | return llvm::any_of(Range: MI.operands(), P: IsHFPR); |
4149 | } |
4150 | |
4151 | bool AArch64InstrInfo::isQForm(const MachineInstr &MI) { |
4152 | auto IsQFPR = [&](const MachineOperand &Op) { |
4153 | if (!Op.isReg()) |
4154 | return false; |
4155 | auto Reg = Op.getReg(); |
4156 | if (Reg.isPhysical()) |
4157 | return AArch64::FPR128RegClass.contains(Reg); |
4158 | const TargetRegisterClass *TRC = ::getRegClass(MI, Reg); |
4159 | return TRC == &AArch64::FPR128RegClass || |
4160 | TRC == &AArch64::FPR128_loRegClass; |
4161 | }; |
4162 | return llvm::any_of(Range: MI.operands(), P: IsQFPR); |
4163 | } |
4164 | |
4165 | bool AArch64InstrInfo::hasBTISemantics(const MachineInstr &MI) { |
4166 | switch (MI.getOpcode()) { |
4167 | case AArch64::BRK: |
4168 | case AArch64::HLT: |
4169 | case AArch64::PACIASP: |
4170 | case AArch64::PACIBSP: |
4171 | // Implicit BTI behavior. |
4172 | return true; |
4173 | case AArch64::PAUTH_PROLOGUE: |
4174 | // PAUTH_PROLOGUE expands to PACI(A|B)SP. |
4175 | return true; |
4176 | case AArch64::HINT: { |
4177 | unsigned Imm = MI.getOperand(i: 0).getImm(); |
4178 | // Explicit BTI instruction. |
4179 | if (Imm == 32 || Imm == 34 || Imm == 36 || Imm == 38) |
4180 | return true; |
4181 | // PACI(A|B)SP instructions. |
4182 | if (Imm == 25 || Imm == 27) |
4183 | return true; |
4184 | return false; |
4185 | } |
4186 | default: |
4187 | return false; |
4188 | } |
4189 | } |
4190 | |
4191 | bool AArch64InstrInfo::isFpOrNEON(Register Reg) { |
4192 | if (Reg == 0) |
4193 | return false; |
4194 | assert(Reg.isPhysical() && "Expected physical register in isFpOrNEON" ); |
4195 | return AArch64::FPR128RegClass.contains(Reg) || |
4196 | AArch64::FPR64RegClass.contains(Reg) || |
4197 | AArch64::FPR32RegClass.contains(Reg) || |
4198 | AArch64::FPR16RegClass.contains(Reg) || |
4199 | AArch64::FPR8RegClass.contains(Reg); |
4200 | } |
4201 | |
4202 | bool AArch64InstrInfo::isFpOrNEON(const MachineInstr &MI) { |
4203 | auto IsFPR = [&](const MachineOperand &Op) { |
4204 | if (!Op.isReg()) |
4205 | return false; |
4206 | auto Reg = Op.getReg(); |
4207 | if (Reg.isPhysical()) |
4208 | return isFpOrNEON(Reg); |
4209 | |
4210 | const TargetRegisterClass *TRC = ::getRegClass(MI, Reg); |
4211 | return TRC == &AArch64::FPR128RegClass || |
4212 | TRC == &AArch64::FPR128_loRegClass || |
4213 | TRC == &AArch64::FPR64RegClass || |
4214 | TRC == &AArch64::FPR64_loRegClass || |
4215 | TRC == &AArch64::FPR32RegClass || TRC == &AArch64::FPR16RegClass || |
4216 | TRC == &AArch64::FPR8RegClass; |
4217 | }; |
4218 | return llvm::any_of(Range: MI.operands(), P: IsFPR); |
4219 | } |
4220 | |
4221 | // Scale the unscaled offsets. Returns false if the unscaled offset can't be |
4222 | // scaled. |
4223 | static bool scaleOffset(unsigned Opc, int64_t &Offset) { |
4224 | int Scale = AArch64InstrInfo::getMemScale(Opc); |
4225 | |
4226 | // If the byte-offset isn't a multiple of the stride, we can't scale this |
4227 | // offset. |
4228 | if (Offset % Scale != 0) |
4229 | return false; |
4230 | |
4231 | // Convert the byte-offset used by unscaled into an "element" offset used |
4232 | // by the scaled pair load/store instructions. |
4233 | Offset /= Scale; |
4234 | return true; |
4235 | } |
4236 | |
4237 | static bool canPairLdStOpc(unsigned FirstOpc, unsigned SecondOpc) { |
4238 | if (FirstOpc == SecondOpc) |
4239 | return true; |
4240 | // We can also pair sign-ext and zero-ext instructions. |
4241 | switch (FirstOpc) { |
4242 | default: |
4243 | return false; |
4244 | case AArch64::STRSui: |
4245 | case AArch64::STURSi: |
4246 | return SecondOpc == AArch64::STRSui || SecondOpc == AArch64::STURSi; |
4247 | case AArch64::STRDui: |
4248 | case AArch64::STURDi: |
4249 | return SecondOpc == AArch64::STRDui || SecondOpc == AArch64::STURDi; |
4250 | case AArch64::STRQui: |
4251 | case AArch64::STURQi: |
4252 | return SecondOpc == AArch64::STRQui || SecondOpc == AArch64::STURQi; |
4253 | case AArch64::STRWui: |
4254 | case AArch64::STURWi: |
4255 | return SecondOpc == AArch64::STRWui || SecondOpc == AArch64::STURWi; |
4256 | case AArch64::STRXui: |
4257 | case AArch64::STURXi: |
4258 | return SecondOpc == AArch64::STRXui || SecondOpc == AArch64::STURXi; |
4259 | case AArch64::LDRSui: |
4260 | case AArch64::LDURSi: |
4261 | return SecondOpc == AArch64::LDRSui || SecondOpc == AArch64::LDURSi; |
4262 | case AArch64::LDRDui: |
4263 | case AArch64::LDURDi: |
4264 | return SecondOpc == AArch64::LDRDui || SecondOpc == AArch64::LDURDi; |
4265 | case AArch64::LDRQui: |
4266 | case AArch64::LDURQi: |
4267 | return SecondOpc == AArch64::LDRQui || SecondOpc == AArch64::LDURQi; |
4268 | case AArch64::LDRWui: |
4269 | case AArch64::LDURWi: |
4270 | return SecondOpc == AArch64::LDRSWui || SecondOpc == AArch64::LDURSWi; |
4271 | case AArch64::LDRSWui: |
4272 | case AArch64::LDURSWi: |
4273 | return SecondOpc == AArch64::LDRWui || SecondOpc == AArch64::LDURWi; |
4274 | case AArch64::LDRXui: |
4275 | case AArch64::LDURXi: |
4276 | return SecondOpc == AArch64::LDRXui || SecondOpc == AArch64::LDURXi; |
4277 | } |
4278 | // These instructions can't be paired based on their opcodes. |
4279 | return false; |
4280 | } |
4281 | |
4282 | static bool shouldClusterFI(const MachineFrameInfo &MFI, int FI1, |
4283 | int64_t Offset1, unsigned Opcode1, int FI2, |
4284 | int64_t Offset2, unsigned Opcode2) { |
4285 | // Accesses through fixed stack object frame indices may access a different |
4286 | // fixed stack slot. Check that the object offsets + offsets match. |
4287 | if (MFI.isFixedObjectIndex(ObjectIdx: FI1) && MFI.isFixedObjectIndex(ObjectIdx: FI2)) { |
4288 | int64_t ObjectOffset1 = MFI.getObjectOffset(ObjectIdx: FI1); |
4289 | int64_t ObjectOffset2 = MFI.getObjectOffset(ObjectIdx: FI2); |
4290 | assert(ObjectOffset1 <= ObjectOffset2 && "Object offsets are not ordered." ); |
4291 | // Convert to scaled object offsets. |
4292 | int Scale1 = AArch64InstrInfo::getMemScale(Opc: Opcode1); |
4293 | if (ObjectOffset1 % Scale1 != 0) |
4294 | return false; |
4295 | ObjectOffset1 /= Scale1; |
4296 | int Scale2 = AArch64InstrInfo::getMemScale(Opc: Opcode2); |
4297 | if (ObjectOffset2 % Scale2 != 0) |
4298 | return false; |
4299 | ObjectOffset2 /= Scale2; |
4300 | ObjectOffset1 += Offset1; |
4301 | ObjectOffset2 += Offset2; |
4302 | return ObjectOffset1 + 1 == ObjectOffset2; |
4303 | } |
4304 | |
4305 | return FI1 == FI2; |
4306 | } |
4307 | |
4308 | /// Detect opportunities for ldp/stp formation. |
4309 | /// |
4310 | /// Only called for LdSt for which getMemOperandWithOffset returns true. |
4311 | bool AArch64InstrInfo::shouldClusterMemOps( |
4312 | ArrayRef<const MachineOperand *> BaseOps1, int64_t OpOffset1, |
4313 | bool OffsetIsScalable1, ArrayRef<const MachineOperand *> BaseOps2, |
4314 | int64_t OpOffset2, bool OffsetIsScalable2, unsigned ClusterSize, |
4315 | unsigned NumBytes) const { |
4316 | assert(BaseOps1.size() == 1 && BaseOps2.size() == 1); |
4317 | const MachineOperand &BaseOp1 = *BaseOps1.front(); |
4318 | const MachineOperand &BaseOp2 = *BaseOps2.front(); |
4319 | const MachineInstr &FirstLdSt = *BaseOp1.getParent(); |
4320 | const MachineInstr &SecondLdSt = *BaseOp2.getParent(); |
4321 | if (BaseOp1.getType() != BaseOp2.getType()) |
4322 | return false; |
4323 | |
4324 | assert((BaseOp1.isReg() || BaseOp1.isFI()) && |
4325 | "Only base registers and frame indices are supported." ); |
4326 | |
4327 | // Check for both base regs and base FI. |
4328 | if (BaseOp1.isReg() && BaseOp1.getReg() != BaseOp2.getReg()) |
4329 | return false; |
4330 | |
4331 | // Only cluster up to a single pair. |
4332 | if (ClusterSize > 2) |
4333 | return false; |
4334 | |
4335 | if (!isPairableLdStInst(MI: FirstLdSt) || !isPairableLdStInst(MI: SecondLdSt)) |
4336 | return false; |
4337 | |
4338 | // Can we pair these instructions based on their opcodes? |
4339 | unsigned FirstOpc = FirstLdSt.getOpcode(); |
4340 | unsigned SecondOpc = SecondLdSt.getOpcode(); |
4341 | if (!canPairLdStOpc(FirstOpc, SecondOpc)) |
4342 | return false; |
4343 | |
4344 | // Can't merge volatiles or load/stores that have a hint to avoid pair |
4345 | // formation, for example. |
4346 | if (!isCandidateToMergeOrPair(MI: FirstLdSt) || |
4347 | !isCandidateToMergeOrPair(MI: SecondLdSt)) |
4348 | return false; |
4349 | |
4350 | // isCandidateToMergeOrPair guarantees that operand 2 is an immediate. |
4351 | int64_t Offset1 = FirstLdSt.getOperand(i: 2).getImm(); |
4352 | if (hasUnscaledLdStOffset(Opc: FirstOpc) && !scaleOffset(Opc: FirstOpc, Offset&: Offset1)) |
4353 | return false; |
4354 | |
4355 | int64_t Offset2 = SecondLdSt.getOperand(i: 2).getImm(); |
4356 | if (hasUnscaledLdStOffset(Opc: SecondOpc) && !scaleOffset(Opc: SecondOpc, Offset&: Offset2)) |
4357 | return false; |
4358 | |
4359 | // Pairwise instructions have a 7-bit signed offset field. |
4360 | if (Offset1 > 63 || Offset1 < -64) |
4361 | return false; |
4362 | |
4363 | // The caller should already have ordered First/SecondLdSt by offset. |
4364 | // Note: except for non-equal frame index bases |
4365 | if (BaseOp1.isFI()) { |
4366 | assert((!BaseOp1.isIdenticalTo(BaseOp2) || Offset1 <= Offset2) && |
4367 | "Caller should have ordered offsets." ); |
4368 | |
4369 | const MachineFrameInfo &MFI = |
4370 | FirstLdSt.getParent()->getParent()->getFrameInfo(); |
4371 | return shouldClusterFI(MFI, FI1: BaseOp1.getIndex(), Offset1, Opcode1: FirstOpc, |
4372 | FI2: BaseOp2.getIndex(), Offset2, Opcode2: SecondOpc); |
4373 | } |
4374 | |
4375 | assert(Offset1 <= Offset2 && "Caller should have ordered offsets." ); |
4376 | |
4377 | return Offset1 + 1 == Offset2; |
4378 | } |
4379 | |
4380 | static const MachineInstrBuilder &AddSubReg(const MachineInstrBuilder &MIB, |
4381 | unsigned Reg, unsigned SubIdx, |
4382 | unsigned State, |
4383 | const TargetRegisterInfo *TRI) { |
4384 | if (!SubIdx) |
4385 | return MIB.addReg(RegNo: Reg, flags: State); |
4386 | |
4387 | if (Register::isPhysicalRegister(Reg)) |
4388 | return MIB.addReg(RegNo: TRI->getSubReg(Reg, Idx: SubIdx), flags: State); |
4389 | return MIB.addReg(RegNo: Reg, flags: State, SubReg: SubIdx); |
4390 | } |
4391 | |
4392 | static bool forwardCopyWillClobberTuple(unsigned DestReg, unsigned SrcReg, |
4393 | unsigned NumRegs) { |
4394 | // We really want the positive remainder mod 32 here, that happens to be |
4395 | // easily obtainable with a mask. |
4396 | return ((DestReg - SrcReg) & 0x1f) < NumRegs; |
4397 | } |
4398 | |
4399 | void AArch64InstrInfo::copyPhysRegTuple(MachineBasicBlock &MBB, |
4400 | MachineBasicBlock::iterator I, |
4401 | const DebugLoc &DL, MCRegister DestReg, |
4402 | MCRegister SrcReg, bool KillSrc, |
4403 | unsigned Opcode, |
4404 | ArrayRef<unsigned> Indices) const { |
4405 | assert(Subtarget.hasNEON() && "Unexpected register copy without NEON" ); |
4406 | const TargetRegisterInfo *TRI = &getRegisterInfo(); |
4407 | uint16_t DestEncoding = TRI->getEncodingValue(RegNo: DestReg); |
4408 | uint16_t SrcEncoding = TRI->getEncodingValue(RegNo: SrcReg); |
4409 | unsigned NumRegs = Indices.size(); |
4410 | |
4411 | int SubReg = 0, End = NumRegs, Incr = 1; |
4412 | if (forwardCopyWillClobberTuple(DestReg: DestEncoding, SrcReg: SrcEncoding, NumRegs)) { |
4413 | SubReg = NumRegs - 1; |
4414 | End = -1; |
4415 | Incr = -1; |
4416 | } |
4417 | |
4418 | for (; SubReg != End; SubReg += Incr) { |
4419 | const MachineInstrBuilder MIB = BuildMI(BB&: MBB, I, MIMD: DL, MCID: get(Opcode)); |
4420 | AddSubReg(MIB, Reg: DestReg, SubIdx: Indices[SubReg], State: RegState::Define, TRI); |
4421 | AddSubReg(MIB, Reg: SrcReg, SubIdx: Indices[SubReg], State: 0, TRI); |
4422 | AddSubReg(MIB, Reg: SrcReg, SubIdx: Indices[SubReg], State: getKillRegState(B: KillSrc), TRI); |
4423 | } |
4424 | } |
4425 | |
4426 | void AArch64InstrInfo::copyGPRRegTuple(MachineBasicBlock &MBB, |
4427 | MachineBasicBlock::iterator I, |
4428 | DebugLoc DL, unsigned DestReg, |
4429 | unsigned SrcReg, bool KillSrc, |
4430 | unsigned Opcode, unsigned ZeroReg, |
4431 | llvm::ArrayRef<unsigned> Indices) const { |
4432 | const TargetRegisterInfo *TRI = &getRegisterInfo(); |
4433 | unsigned NumRegs = Indices.size(); |
4434 | |
4435 | #ifndef NDEBUG |
4436 | uint16_t DestEncoding = TRI->getEncodingValue(DestReg); |
4437 | uint16_t SrcEncoding = TRI->getEncodingValue(SrcReg); |
4438 | assert(DestEncoding % NumRegs == 0 && SrcEncoding % NumRegs == 0 && |
4439 | "GPR reg sequences should not be able to overlap" ); |
4440 | #endif |
4441 | |
4442 | for (unsigned SubReg = 0; SubReg != NumRegs; ++SubReg) { |
4443 | const MachineInstrBuilder MIB = BuildMI(BB&: MBB, I, MIMD: DL, MCID: get(Opcode)); |
4444 | AddSubReg(MIB, Reg: DestReg, SubIdx: Indices[SubReg], State: RegState::Define, TRI); |
4445 | MIB.addReg(RegNo: ZeroReg); |
4446 | AddSubReg(MIB, Reg: SrcReg, SubIdx: Indices[SubReg], State: getKillRegState(B: KillSrc), TRI); |
4447 | MIB.addImm(Val: 0); |
4448 | } |
4449 | } |
4450 | |
4451 | void AArch64InstrInfo::copyPhysReg(MachineBasicBlock &MBB, |
4452 | MachineBasicBlock::iterator I, |
4453 | const DebugLoc &DL, MCRegister DestReg, |
4454 | MCRegister SrcReg, bool KillSrc) const { |
4455 | if (AArch64::GPR32spRegClass.contains(Reg: DestReg) && |
4456 | (AArch64::GPR32spRegClass.contains(Reg: SrcReg) || SrcReg == AArch64::WZR)) { |
4457 | const TargetRegisterInfo *TRI = &getRegisterInfo(); |
4458 | |
4459 | if (DestReg == AArch64::WSP || SrcReg == AArch64::WSP) { |
4460 | // If either operand is WSP, expand to ADD #0. |
4461 | if (Subtarget.hasZeroCycleRegMove()) { |
4462 | // Cyclone recognizes "ADD Xd, Xn, #0" as a zero-cycle register move. |
4463 | MCRegister DestRegX = TRI->getMatchingSuperReg( |
4464 | Reg: DestReg, SubIdx: AArch64::sub_32, RC: &AArch64::GPR64spRegClass); |
4465 | MCRegister SrcRegX = TRI->getMatchingSuperReg( |
4466 | Reg: SrcReg, SubIdx: AArch64::sub_32, RC: &AArch64::GPR64spRegClass); |
4467 | // This instruction is reading and writing X registers. This may upset |
4468 | // the register scavenger and machine verifier, so we need to indicate |
4469 | // that we are reading an undefined value from SrcRegX, but a proper |
4470 | // value from SrcReg. |
4471 | BuildMI(BB&: MBB, I, MIMD: DL, MCID: get(Opcode: AArch64::ADDXri), DestReg: DestRegX) |
4472 | .addReg(RegNo: SrcRegX, flags: RegState::Undef) |
4473 | .addImm(Val: 0) |
4474 | .addImm(Val: AArch64_AM::getShifterImm(ST: AArch64_AM::LSL, Imm: 0)) |
4475 | .addReg(RegNo: SrcReg, flags: RegState::Implicit | getKillRegState(B: KillSrc)); |
4476 | } else { |
4477 | BuildMI(BB&: MBB, I, MIMD: DL, MCID: get(Opcode: AArch64::ADDWri), DestReg) |
4478 | .addReg(RegNo: SrcReg, flags: getKillRegState(B: KillSrc)) |
4479 | .addImm(Val: 0) |
4480 | .addImm(Val: AArch64_AM::getShifterImm(ST: AArch64_AM::LSL, Imm: 0)); |
4481 | } |
4482 | } else if (SrcReg == AArch64::WZR && Subtarget.hasZeroCycleZeroingGP()) { |
4483 | BuildMI(BB&: MBB, I, MIMD: DL, MCID: get(Opcode: AArch64::MOVZWi), DestReg) |
4484 | .addImm(Val: 0) |
4485 | .addImm(Val: AArch64_AM::getShifterImm(ST: AArch64_AM::LSL, Imm: 0)); |
4486 | } else { |
4487 | if (Subtarget.hasZeroCycleRegMove()) { |
4488 | // Cyclone recognizes "ORR Xd, XZR, Xm" as a zero-cycle register move. |
4489 | MCRegister DestRegX = TRI->getMatchingSuperReg( |
4490 | Reg: DestReg, SubIdx: AArch64::sub_32, RC: &AArch64::GPR64spRegClass); |
4491 | MCRegister SrcRegX = TRI->getMatchingSuperReg( |
4492 | Reg: SrcReg, SubIdx: AArch64::sub_32, RC: &AArch64::GPR64spRegClass); |
4493 | // This instruction is reading and writing X registers. This may upset |
4494 | // the register scavenger and machine verifier, so we need to indicate |
4495 | // that we are reading an undefined value from SrcRegX, but a proper |
4496 | // value from SrcReg. |
4497 | BuildMI(BB&: MBB, I, MIMD: DL, MCID: get(Opcode: AArch64::ORRXrr), DestReg: DestRegX) |
4498 | .addReg(RegNo: AArch64::XZR) |
4499 | .addReg(RegNo: SrcRegX, flags: RegState::Undef) |
4500 | .addReg(RegNo: SrcReg, flags: RegState::Implicit | getKillRegState(B: KillSrc)); |
4501 | } else { |
4502 | // Otherwise, expand to ORR WZR. |
4503 | BuildMI(BB&: MBB, I, MIMD: DL, MCID: get(Opcode: AArch64::ORRWrr), DestReg) |
4504 | .addReg(RegNo: AArch64::WZR) |
4505 | .addReg(RegNo: SrcReg, flags: getKillRegState(B: KillSrc)); |
4506 | } |
4507 | } |
4508 | return; |
4509 | } |
4510 | |
4511 | // Copy a Predicate register by ORRing with itself. |
4512 | if (AArch64::PPRRegClass.contains(Reg: DestReg) && |
4513 | AArch64::PPRRegClass.contains(Reg: SrcReg)) { |
4514 | assert(Subtarget.isSVEorStreamingSVEAvailable() && |
4515 | "Unexpected SVE register." ); |
4516 | BuildMI(BB&: MBB, I, MIMD: DL, MCID: get(Opcode: AArch64::ORR_PPzPP), DestReg) |
4517 | .addReg(RegNo: SrcReg) // Pg |
4518 | .addReg(RegNo: SrcReg) |
4519 | .addReg(RegNo: SrcReg, flags: getKillRegState(B: KillSrc)); |
4520 | return; |
4521 | } |
4522 | |
4523 | // Copy a predicate-as-counter register by ORRing with itself as if it |
4524 | // were a regular predicate (mask) register. |
4525 | bool DestIsPNR = AArch64::PNRRegClass.contains(Reg: DestReg); |
4526 | bool SrcIsPNR = AArch64::PNRRegClass.contains(Reg: SrcReg); |
4527 | if (DestIsPNR || SrcIsPNR) { |
4528 | auto ToPPR = [](MCRegister R) -> MCRegister { |
4529 | return (R - AArch64::PN0) + AArch64::P0; |
4530 | }; |
4531 | MCRegister PPRSrcReg = SrcIsPNR ? ToPPR(SrcReg) : SrcReg; |
4532 | MCRegister PPRDestReg = DestIsPNR ? ToPPR(DestReg) : DestReg; |
4533 | |
4534 | if (PPRSrcReg != PPRDestReg) { |
4535 | auto NewMI = BuildMI(BB&: MBB, I, MIMD: DL, MCID: get(Opcode: AArch64::ORR_PPzPP), DestReg: PPRDestReg) |
4536 | .addReg(RegNo: PPRSrcReg) // Pg |
4537 | .addReg(RegNo: PPRSrcReg) |
4538 | .addReg(RegNo: PPRSrcReg, flags: getKillRegState(B: KillSrc)); |
4539 | if (DestIsPNR) |
4540 | NewMI.addDef(RegNo: DestReg, Flags: RegState::Implicit); |
4541 | } |
4542 | return; |
4543 | } |
4544 | |
4545 | // Copy a Z register by ORRing with itself. |
4546 | if (AArch64::ZPRRegClass.contains(Reg: DestReg) && |
4547 | AArch64::ZPRRegClass.contains(Reg: SrcReg)) { |
4548 | assert(Subtarget.isSVEorStreamingSVEAvailable() && |
4549 | "Unexpected SVE register." ); |
4550 | BuildMI(BB&: MBB, I, MIMD: DL, MCID: get(Opcode: AArch64::ORR_ZZZ), DestReg) |
4551 | .addReg(RegNo: SrcReg) |
4552 | .addReg(RegNo: SrcReg, flags: getKillRegState(B: KillSrc)); |
4553 | return; |
4554 | } |
4555 | |
4556 | // Copy a Z register pair by copying the individual sub-registers. |
4557 | if ((AArch64::ZPR2RegClass.contains(Reg: DestReg) || |
4558 | AArch64::ZPR2StridedOrContiguousRegClass.contains(Reg: DestReg)) && |
4559 | (AArch64::ZPR2RegClass.contains(Reg: SrcReg) || |
4560 | AArch64::ZPR2StridedOrContiguousRegClass.contains(Reg: SrcReg))) { |
4561 | assert(Subtarget.isSVEorStreamingSVEAvailable() && |
4562 | "Unexpected SVE register." ); |
4563 | static const unsigned Indices[] = {AArch64::zsub0, AArch64::zsub1}; |
4564 | copyPhysRegTuple(MBB, I, DL, DestReg, SrcReg, KillSrc, Opcode: AArch64::ORR_ZZZ, |
4565 | Indices); |
4566 | return; |
4567 | } |
4568 | |
4569 | // Copy a Z register triple by copying the individual sub-registers. |
4570 | if (AArch64::ZPR3RegClass.contains(Reg: DestReg) && |
4571 | AArch64::ZPR3RegClass.contains(Reg: SrcReg)) { |
4572 | assert(Subtarget.isSVEorStreamingSVEAvailable() && |
4573 | "Unexpected SVE register." ); |
4574 | static const unsigned Indices[] = {AArch64::zsub0, AArch64::zsub1, |
4575 | AArch64::zsub2}; |
4576 | copyPhysRegTuple(MBB, I, DL, DestReg, SrcReg, KillSrc, Opcode: AArch64::ORR_ZZZ, |
4577 | Indices); |
4578 | return; |
4579 | } |
4580 | |
4581 | // Copy a Z register quad by copying the individual sub-registers. |
4582 | if ((AArch64::ZPR4RegClass.contains(Reg: DestReg) || |
4583 | AArch64::ZPR4StridedOrContiguousRegClass.contains(Reg: DestReg)) && |
4584 | (AArch64::ZPR4RegClass.contains(Reg: SrcReg) || |
4585 | AArch64::ZPR4StridedOrContiguousRegClass.contains(Reg: SrcReg))) { |
4586 | assert(Subtarget.isSVEorStreamingSVEAvailable() && |
4587 | "Unexpected SVE register." ); |
4588 | static const unsigned Indices[] = {AArch64::zsub0, AArch64::zsub1, |
4589 | AArch64::zsub2, AArch64::zsub3}; |
4590 | copyPhysRegTuple(MBB, I, DL, DestReg, SrcReg, KillSrc, Opcode: AArch64::ORR_ZZZ, |
4591 | Indices); |
4592 | return; |
4593 | } |
4594 | |
4595 | if (AArch64::GPR64spRegClass.contains(Reg: DestReg) && |
4596 | (AArch64::GPR64spRegClass.contains(Reg: SrcReg) || SrcReg == AArch64::XZR)) { |
4597 | if (DestReg == AArch64::SP || SrcReg == AArch64::SP) { |
4598 | // If either operand is SP, expand to ADD #0. |
4599 | BuildMI(BB&: MBB, I, MIMD: DL, MCID: get(Opcode: AArch64::ADDXri), DestReg) |
4600 | .addReg(RegNo: SrcReg, flags: getKillRegState(B: KillSrc)) |
4601 | .addImm(Val: 0) |
4602 | .addImm(Val: AArch64_AM::getShifterImm(ST: AArch64_AM::LSL, Imm: 0)); |
4603 | } else if (SrcReg == AArch64::XZR && Subtarget.hasZeroCycleZeroingGP()) { |
4604 | BuildMI(BB&: MBB, I, MIMD: DL, MCID: get(Opcode: AArch64::MOVZXi), DestReg) |
4605 | .addImm(Val: 0) |
4606 | .addImm(Val: AArch64_AM::getShifterImm(ST: AArch64_AM::LSL, Imm: 0)); |
4607 | } else { |
4608 | // Otherwise, expand to ORR XZR. |
4609 | BuildMI(BB&: MBB, I, MIMD: DL, MCID: get(Opcode: AArch64::ORRXrr), DestReg) |
4610 | .addReg(RegNo: AArch64::XZR) |
4611 | .addReg(RegNo: SrcReg, flags: getKillRegState(B: KillSrc)); |
4612 | } |
4613 | return; |
4614 | } |
4615 | |
4616 | // Copy a DDDD register quad by copying the individual sub-registers. |
4617 | if (AArch64::DDDDRegClass.contains(Reg: DestReg) && |
4618 | AArch64::DDDDRegClass.contains(Reg: SrcReg)) { |
4619 | static const unsigned Indices[] = {AArch64::dsub0, AArch64::dsub1, |
4620 | AArch64::dsub2, AArch64::dsub3}; |
4621 | copyPhysRegTuple(MBB, I, DL, DestReg, SrcReg, KillSrc, Opcode: AArch64::ORRv8i8, |
4622 | Indices); |
4623 | return; |
4624 | } |
4625 | |
4626 | // Copy a DDD register triple by copying the individual sub-registers. |
4627 | if (AArch64::DDDRegClass.contains(Reg: DestReg) && |
4628 | AArch64::DDDRegClass.contains(Reg: SrcReg)) { |
4629 | static const unsigned Indices[] = {AArch64::dsub0, AArch64::dsub1, |
4630 | AArch64::dsub2}; |
4631 | copyPhysRegTuple(MBB, I, DL, DestReg, SrcReg, KillSrc, Opcode: AArch64::ORRv8i8, |
4632 | Indices); |
4633 | return; |
4634 | } |
4635 | |
4636 | // Copy a DD register pair by copying the individual sub-registers. |
4637 | if (AArch64::DDRegClass.contains(Reg: DestReg) && |
4638 | AArch64::DDRegClass.contains(Reg: SrcReg)) { |
4639 | static const unsigned Indices[] = {AArch64::dsub0, AArch64::dsub1}; |
4640 | copyPhysRegTuple(MBB, I, DL, DestReg, SrcReg, KillSrc, Opcode: AArch64::ORRv8i8, |
4641 | Indices); |
4642 | return; |
4643 | } |
4644 | |
4645 | // Copy a QQQQ register quad by copying the individual sub-registers. |
4646 | if (AArch64::QQQQRegClass.contains(Reg: DestReg) && |
4647 | AArch64::QQQQRegClass.contains(Reg: SrcReg)) { |
4648 | static const unsigned Indices[] = {AArch64::qsub0, AArch64::qsub1, |
4649 | AArch64::qsub2, AArch64::qsub3}; |
4650 | copyPhysRegTuple(MBB, I, DL, DestReg, SrcReg, KillSrc, Opcode: AArch64::ORRv16i8, |
4651 | Indices); |
4652 | return; |
4653 | } |
4654 | |
4655 | // Copy a QQQ register triple by copying the individual sub-registers. |
4656 | if (AArch64::QQQRegClass.contains(Reg: DestReg) && |
4657 | AArch64::QQQRegClass.contains(Reg: SrcReg)) { |
4658 | static const unsigned Indices[] = {AArch64::qsub0, AArch64::qsub1, |
4659 | AArch64::qsub2}; |
4660 | copyPhysRegTuple(MBB, I, DL, DestReg, SrcReg, KillSrc, Opcode: AArch64::ORRv16i8, |
4661 | Indices); |
4662 | return; |
4663 | } |
4664 | |
4665 | // Copy a QQ register pair by copying the individual sub-registers. |
4666 | if (AArch64::QQRegClass.contains(Reg: DestReg) && |
4667 | AArch64::QQRegClass.contains(Reg: SrcReg)) { |
4668 | static const unsigned Indices[] = {AArch64::qsub0, AArch64::qsub1}; |
4669 | copyPhysRegTuple(MBB, I, DL, DestReg, SrcReg, KillSrc, Opcode: AArch64::ORRv16i8, |
4670 | Indices); |
4671 | return; |
4672 | } |
4673 | |
4674 | if (AArch64::XSeqPairsClassRegClass.contains(Reg: DestReg) && |
4675 | AArch64::XSeqPairsClassRegClass.contains(Reg: SrcReg)) { |
4676 | static const unsigned Indices[] = {AArch64::sube64, AArch64::subo64}; |
4677 | copyGPRRegTuple(MBB, I, DL, DestReg, SrcReg, KillSrc, Opcode: AArch64::ORRXrs, |
4678 | ZeroReg: AArch64::XZR, Indices); |
4679 | return; |
4680 | } |
4681 | |
4682 | if (AArch64::WSeqPairsClassRegClass.contains(Reg: DestReg) && |
4683 | AArch64::WSeqPairsClassRegClass.contains(Reg: SrcReg)) { |
4684 | static const unsigned Indices[] = {AArch64::sube32, AArch64::subo32}; |
4685 | copyGPRRegTuple(MBB, I, DL, DestReg, SrcReg, KillSrc, Opcode: AArch64::ORRWrs, |
4686 | ZeroReg: AArch64::WZR, Indices); |
4687 | return; |
4688 | } |
4689 | |
4690 | if (AArch64::FPR128RegClass.contains(Reg: DestReg) && |
4691 | AArch64::FPR128RegClass.contains(Reg: SrcReg)) { |
4692 | if (Subtarget.isSVEorStreamingSVEAvailable() && |
4693 | !Subtarget.isNeonAvailable()) |
4694 | BuildMI(BB&: MBB, I, MIMD: DL, MCID: get(Opcode: AArch64::ORR_ZZZ)) |
4695 | .addReg(RegNo: AArch64::Z0 + (DestReg - AArch64::Q0), flags: RegState::Define) |
4696 | .addReg(RegNo: AArch64::Z0 + (SrcReg - AArch64::Q0)) |
4697 | .addReg(RegNo: AArch64::Z0 + (SrcReg - AArch64::Q0)); |
4698 | else if (Subtarget.isNeonAvailable()) |
4699 | BuildMI(BB&: MBB, I, MIMD: DL, MCID: get(Opcode: AArch64::ORRv16i8), DestReg) |
4700 | .addReg(RegNo: SrcReg) |
4701 | .addReg(RegNo: SrcReg, flags: getKillRegState(B: KillSrc)); |
4702 | else { |
4703 | BuildMI(BB&: MBB, I, MIMD: DL, MCID: get(Opcode: AArch64::STRQpre)) |
4704 | .addReg(RegNo: AArch64::SP, flags: RegState::Define) |
4705 | .addReg(RegNo: SrcReg, flags: getKillRegState(B: KillSrc)) |
4706 | .addReg(RegNo: AArch64::SP) |
4707 | .addImm(Val: -16); |
4708 | BuildMI(BB&: MBB, I, MIMD: DL, MCID: get(Opcode: AArch64::LDRQpost)) |
4709 | .addReg(RegNo: AArch64::SP, flags: RegState::Define) |
4710 | .addReg(RegNo: DestReg, flags: RegState::Define) |
4711 | .addReg(RegNo: AArch64::SP) |
4712 | .addImm(Val: 16); |
4713 | } |
4714 | return; |
4715 | } |
4716 | |
4717 | if (AArch64::FPR64RegClass.contains(Reg: DestReg) && |
4718 | AArch64::FPR64RegClass.contains(Reg: SrcReg)) { |
4719 | BuildMI(BB&: MBB, I, MIMD: DL, MCID: get(Opcode: AArch64::FMOVDr), DestReg) |
4720 | .addReg(RegNo: SrcReg, flags: getKillRegState(B: KillSrc)); |
4721 | return; |
4722 | } |
4723 | |
4724 | if (AArch64::FPR32RegClass.contains(Reg: DestReg) && |
4725 | AArch64::FPR32RegClass.contains(Reg: SrcReg)) { |
4726 | BuildMI(BB&: MBB, I, MIMD: DL, MCID: get(Opcode: AArch64::FMOVSr), DestReg) |
4727 | .addReg(RegNo: SrcReg, flags: getKillRegState(B: KillSrc)); |
4728 | return; |
4729 | } |
4730 | |
4731 | if (AArch64::FPR16RegClass.contains(Reg: DestReg) && |
4732 | AArch64::FPR16RegClass.contains(Reg: SrcReg)) { |
4733 | DestReg = |
4734 | RI.getMatchingSuperReg(Reg: DestReg, SubIdx: AArch64::hsub, RC: &AArch64::FPR32RegClass); |
4735 | SrcReg = |
4736 | RI.getMatchingSuperReg(Reg: SrcReg, SubIdx: AArch64::hsub, RC: &AArch64::FPR32RegClass); |
4737 | BuildMI(BB&: MBB, I, MIMD: DL, MCID: get(Opcode: AArch64::FMOVSr), DestReg) |
4738 | .addReg(RegNo: SrcReg, flags: getKillRegState(B: KillSrc)); |
4739 | return; |
4740 | } |
4741 | |
4742 | if (AArch64::FPR8RegClass.contains(Reg: DestReg) && |
4743 | AArch64::FPR8RegClass.contains(Reg: SrcReg)) { |
4744 | DestReg = |
4745 | RI.getMatchingSuperReg(Reg: DestReg, SubIdx: AArch64::bsub, RC: &AArch64::FPR32RegClass); |
4746 | SrcReg = |
4747 | RI.getMatchingSuperReg(Reg: SrcReg, SubIdx: AArch64::bsub, RC: &AArch64::FPR32RegClass); |
4748 | BuildMI(BB&: MBB, I, MIMD: DL, MCID: get(Opcode: AArch64::FMOVSr), DestReg) |
4749 | .addReg(RegNo: SrcReg, flags: getKillRegState(B: KillSrc)); |
4750 | return; |
4751 | } |
4752 | |
4753 | // Copies between GPR64 and FPR64. |
4754 | if (AArch64::FPR64RegClass.contains(Reg: DestReg) && |
4755 | AArch64::GPR64RegClass.contains(Reg: SrcReg)) { |
4756 | BuildMI(BB&: MBB, I, MIMD: DL, MCID: get(Opcode: AArch64::FMOVXDr), DestReg) |
4757 | .addReg(RegNo: SrcReg, flags: getKillRegState(B: KillSrc)); |
4758 | return; |
4759 | } |
4760 | if (AArch64::GPR64RegClass.contains(Reg: DestReg) && |
4761 | AArch64::FPR64RegClass.contains(Reg: SrcReg)) { |
4762 | BuildMI(BB&: MBB, I, MIMD: DL, MCID: get(Opcode: AArch64::FMOVDXr), DestReg) |
4763 | .addReg(RegNo: SrcReg, flags: getKillRegState(B: KillSrc)); |
4764 | return; |
4765 | } |
4766 | // Copies between GPR32 and FPR32. |
4767 | if (AArch64::FPR32RegClass.contains(Reg: DestReg) && |
4768 | AArch64::GPR32RegClass.contains(Reg: SrcReg)) { |
4769 | BuildMI(BB&: MBB, I, MIMD: DL, MCID: get(Opcode: AArch64::FMOVWSr), DestReg) |
4770 | .addReg(RegNo: SrcReg, flags: getKillRegState(B: KillSrc)); |
4771 | return; |
4772 | } |
4773 | if (AArch64::GPR32RegClass.contains(Reg: DestReg) && |
4774 | AArch64::FPR32RegClass.contains(Reg: SrcReg)) { |
4775 | BuildMI(BB&: MBB, I, MIMD: DL, MCID: get(Opcode: AArch64::FMOVSWr), DestReg) |
4776 | .addReg(RegNo: SrcReg, flags: getKillRegState(B: KillSrc)); |
4777 | return; |
4778 | } |
4779 | |
4780 | if (DestReg == AArch64::NZCV) { |
4781 | assert(AArch64::GPR64RegClass.contains(SrcReg) && "Invalid NZCV copy" ); |
4782 | BuildMI(BB&: MBB, I, MIMD: DL, MCID: get(Opcode: AArch64::MSR)) |
4783 | .addImm(Val: AArch64SysReg::NZCV) |
4784 | .addReg(RegNo: SrcReg, flags: getKillRegState(B: KillSrc)) |
4785 | .addReg(RegNo: AArch64::NZCV, flags: RegState::Implicit | RegState::Define); |
4786 | return; |
4787 | } |
4788 | |
4789 | if (SrcReg == AArch64::NZCV) { |
4790 | assert(AArch64::GPR64RegClass.contains(DestReg) && "Invalid NZCV copy" ); |
4791 | BuildMI(BB&: MBB, I, MIMD: DL, MCID: get(Opcode: AArch64::MRS), DestReg) |
4792 | .addImm(Val: AArch64SysReg::NZCV) |
4793 | .addReg(RegNo: AArch64::NZCV, flags: RegState::Implicit | getKillRegState(B: KillSrc)); |
4794 | return; |
4795 | } |
4796 | |
4797 | #ifndef NDEBUG |
4798 | const TargetRegisterInfo &TRI = getRegisterInfo(); |
4799 | errs() << TRI.getRegAsmName(DestReg) << " = COPY " |
4800 | << TRI.getRegAsmName(SrcReg) << "\n" ; |
4801 | #endif |
4802 | llvm_unreachable("unimplemented reg-to-reg copy" ); |
4803 | } |
4804 | |
4805 | static void storeRegPairToStackSlot(const TargetRegisterInfo &TRI, |
4806 | MachineBasicBlock &MBB, |
4807 | MachineBasicBlock::iterator InsertBefore, |
4808 | const MCInstrDesc &MCID, |
4809 | Register SrcReg, bool IsKill, |
4810 | unsigned SubIdx0, unsigned SubIdx1, int FI, |
4811 | MachineMemOperand *MMO) { |
4812 | Register SrcReg0 = SrcReg; |
4813 | Register SrcReg1 = SrcReg; |
4814 | if (SrcReg.isPhysical()) { |
4815 | SrcReg0 = TRI.getSubReg(Reg: SrcReg, Idx: SubIdx0); |
4816 | SubIdx0 = 0; |
4817 | SrcReg1 = TRI.getSubReg(Reg: SrcReg, Idx: SubIdx1); |
4818 | SubIdx1 = 0; |
4819 | } |
4820 | BuildMI(BB&: MBB, I: InsertBefore, MIMD: DebugLoc(), MCID) |
4821 | .addReg(RegNo: SrcReg0, flags: getKillRegState(B: IsKill), SubReg: SubIdx0) |
4822 | .addReg(RegNo: SrcReg1, flags: getKillRegState(B: IsKill), SubReg: SubIdx1) |
4823 | .addFrameIndex(Idx: FI) |
4824 | .addImm(Val: 0) |
4825 | .addMemOperand(MMO); |
4826 | } |
4827 | |
4828 | void AArch64InstrInfo::storeRegToStackSlot(MachineBasicBlock &MBB, |
4829 | MachineBasicBlock::iterator MBBI, |
4830 | Register SrcReg, bool isKill, int FI, |
4831 | const TargetRegisterClass *RC, |
4832 | const TargetRegisterInfo *TRI, |
4833 | Register VReg) const { |
4834 | MachineFunction &MF = *MBB.getParent(); |
4835 | MachineFrameInfo &MFI = MF.getFrameInfo(); |
4836 | |
4837 | MachinePointerInfo PtrInfo = MachinePointerInfo::getFixedStack(MF, FI); |
4838 | MachineMemOperand *MMO = |
4839 | MF.getMachineMemOperand(PtrInfo, F: MachineMemOperand::MOStore, |
4840 | Size: MFI.getObjectSize(ObjectIdx: FI), BaseAlignment: MFI.getObjectAlign(ObjectIdx: FI)); |
4841 | unsigned Opc = 0; |
4842 | bool Offset = true; |
4843 | MCRegister PNRReg = MCRegister::NoRegister; |
4844 | unsigned StackID = TargetStackID::Default; |
4845 | switch (TRI->getSpillSize(RC: *RC)) { |
4846 | case 1: |
4847 | if (AArch64::FPR8RegClass.hasSubClassEq(RC)) |
4848 | Opc = AArch64::STRBui; |
4849 | break; |
4850 | case 2: { |
4851 | if (AArch64::FPR16RegClass.hasSubClassEq(RC)) |
4852 | Opc = AArch64::STRHui; |
4853 | else if (AArch64::PNRRegClass.hasSubClassEq(RC) || |
4854 | AArch64::PPRRegClass.hasSubClassEq(RC)) { |
4855 | assert(Subtarget.isSVEorStreamingSVEAvailable() && |
4856 | "Unexpected register store without SVE store instructions" ); |
4857 | Opc = AArch64::STR_PXI; |
4858 | StackID = TargetStackID::ScalableVector; |
4859 | } |
4860 | break; |
4861 | } |
4862 | case 4: |
4863 | if (AArch64::GPR32allRegClass.hasSubClassEq(RC)) { |
4864 | Opc = AArch64::STRWui; |
4865 | if (SrcReg.isVirtual()) |
4866 | MF.getRegInfo().constrainRegClass(Reg: SrcReg, RC: &AArch64::GPR32RegClass); |
4867 | else |
4868 | assert(SrcReg != AArch64::WSP); |
4869 | } else if (AArch64::FPR32RegClass.hasSubClassEq(RC)) |
4870 | Opc = AArch64::STRSui; |
4871 | else if (AArch64::PPR2RegClass.hasSubClassEq(RC)) { |
4872 | Opc = AArch64::STR_PPXI; |
4873 | StackID = TargetStackID::ScalableVector; |
4874 | } |
4875 | break; |
4876 | case 8: |
4877 | if (AArch64::GPR64allRegClass.hasSubClassEq(RC)) { |
4878 | Opc = AArch64::STRXui; |
4879 | if (SrcReg.isVirtual()) |
4880 | MF.getRegInfo().constrainRegClass(Reg: SrcReg, RC: &AArch64::GPR64RegClass); |
4881 | else |
4882 | assert(SrcReg != AArch64::SP); |
4883 | } else if (AArch64::FPR64RegClass.hasSubClassEq(RC)) { |
4884 | Opc = AArch64::STRDui; |
4885 | } else if (AArch64::WSeqPairsClassRegClass.hasSubClassEq(RC)) { |
4886 | storeRegPairToStackSlot(TRI: getRegisterInfo(), MBB, InsertBefore: MBBI, |
4887 | MCID: get(Opcode: AArch64::STPWi), SrcReg, IsKill: isKill, |
4888 | SubIdx0: AArch64::sube32, SubIdx1: AArch64::subo32, FI, MMO); |
4889 | return; |
4890 | } |
4891 | break; |
4892 | case 16: |
4893 | if (AArch64::FPR128RegClass.hasSubClassEq(RC)) |
4894 | Opc = AArch64::STRQui; |
4895 | else if (AArch64::DDRegClass.hasSubClassEq(RC)) { |
4896 | assert(Subtarget.hasNEON() && "Unexpected register store without NEON" ); |
4897 | Opc = AArch64::ST1Twov1d; |
4898 | Offset = false; |
4899 | } else if (AArch64::XSeqPairsClassRegClass.hasSubClassEq(RC)) { |
4900 | storeRegPairToStackSlot(TRI: getRegisterInfo(), MBB, InsertBefore: MBBI, |
4901 | MCID: get(Opcode: AArch64::STPXi), SrcReg, IsKill: isKill, |
4902 | SubIdx0: AArch64::sube64, SubIdx1: AArch64::subo64, FI, MMO); |
4903 | return; |
4904 | } else if (AArch64::ZPRRegClass.hasSubClassEq(RC)) { |
4905 | assert(Subtarget.isSVEorStreamingSVEAvailable() && |
4906 | "Unexpected register store without SVE store instructions" ); |
4907 | Opc = AArch64::STR_ZXI; |
4908 | StackID = TargetStackID::ScalableVector; |
4909 | } |
4910 | break; |
4911 | case 24: |
4912 | if (AArch64::DDDRegClass.hasSubClassEq(RC)) { |
4913 | assert(Subtarget.hasNEON() && "Unexpected register store without NEON" ); |
4914 | Opc = AArch64::ST1Threev1d; |
4915 | Offset = false; |
4916 | } |
4917 | break; |
4918 | case 32: |
4919 | if (AArch64::DDDDRegClass.hasSubClassEq(RC)) { |
4920 | assert(Subtarget.hasNEON() && "Unexpected register store without NEON" ); |
4921 | Opc = AArch64::ST1Fourv1d; |
4922 | Offset = false; |
4923 | } else if (AArch64::QQRegClass.hasSubClassEq(RC)) { |
4924 | assert(Subtarget.hasNEON() && "Unexpected register store without NEON" ); |
4925 | Opc = AArch64::ST1Twov2d; |
4926 | Offset = false; |
4927 | } else if (AArch64::ZPR2RegClass.hasSubClassEq(RC) || |
4928 | AArch64::ZPR2StridedOrContiguousRegClass.hasSubClassEq(RC)) { |
4929 | assert(Subtarget.isSVEorStreamingSVEAvailable() && |
4930 | "Unexpected register store without SVE store instructions" ); |
4931 | Opc = AArch64::STR_ZZXI; |
4932 | StackID = TargetStackID::ScalableVector; |
4933 | } |
4934 | break; |
4935 | case 48: |
4936 | if (AArch64::QQQRegClass.hasSubClassEq(RC)) { |
4937 | assert(Subtarget.hasNEON() && "Unexpected register store without NEON" ); |
4938 | Opc = AArch64::ST1Threev2d; |
4939 | Offset = false; |
4940 | } else if (AArch64::ZPR3RegClass.hasSubClassEq(RC)) { |
4941 | assert(Subtarget.isSVEorStreamingSVEAvailable() && |
4942 | "Unexpected register store without SVE store instructions" ); |
4943 | Opc = AArch64::STR_ZZZXI; |
4944 | StackID = TargetStackID::ScalableVector; |
4945 | } |
4946 | break; |
4947 | case 64: |
4948 | if (AArch64::QQQQRegClass.hasSubClassEq(RC)) { |
4949 | assert(Subtarget.hasNEON() && "Unexpected register store without NEON" ); |
4950 | Opc = AArch64::ST1Fourv2d; |
4951 | Offset = false; |
4952 | } else if (AArch64::ZPR4RegClass.hasSubClassEq(RC) || |
4953 | AArch64::ZPR4StridedOrContiguousRegClass.hasSubClassEq(RC)) { |
4954 | assert(Subtarget.isSVEorStreamingSVEAvailable() && |
4955 | "Unexpected register store without SVE store instructions" ); |
4956 | Opc = AArch64::STR_ZZZZXI; |
4957 | StackID = TargetStackID::ScalableVector; |
4958 | } |
4959 | break; |
4960 | } |
4961 | assert(Opc && "Unknown register class" ); |
4962 | MFI.setStackID(ObjectIdx: FI, ID: StackID); |
4963 | |
4964 | const MachineInstrBuilder MI = BuildMI(BB&: MBB, I: MBBI, MIMD: DebugLoc(), MCID: get(Opcode: Opc)) |
4965 | .addReg(RegNo: SrcReg, flags: getKillRegState(B: isKill)) |
4966 | .addFrameIndex(Idx: FI); |
4967 | |
4968 | if (Offset) |
4969 | MI.addImm(Val: 0); |
4970 | if (PNRReg.isValid()) |
4971 | MI.addDef(RegNo: PNRReg, Flags: RegState::Implicit); |
4972 | MI.addMemOperand(MMO); |
4973 | } |
4974 | |
4975 | static void loadRegPairFromStackSlot(const TargetRegisterInfo &TRI, |
4976 | MachineBasicBlock &MBB, |
4977 | MachineBasicBlock::iterator InsertBefore, |
4978 | const MCInstrDesc &MCID, |
4979 | Register DestReg, unsigned SubIdx0, |
4980 | unsigned SubIdx1, int FI, |
4981 | MachineMemOperand *MMO) { |
4982 | Register DestReg0 = DestReg; |
4983 | Register DestReg1 = DestReg; |
4984 | bool IsUndef = true; |
4985 | if (DestReg.isPhysical()) { |
4986 | DestReg0 = TRI.getSubReg(Reg: DestReg, Idx: SubIdx0); |
4987 | SubIdx0 = 0; |
4988 | DestReg1 = TRI.getSubReg(Reg: DestReg, Idx: SubIdx1); |
4989 | SubIdx1 = 0; |
4990 | IsUndef = false; |
4991 | } |
4992 | BuildMI(BB&: MBB, I: InsertBefore, MIMD: DebugLoc(), MCID) |
4993 | .addReg(RegNo: DestReg0, flags: RegState::Define | getUndefRegState(B: IsUndef), SubReg: SubIdx0) |
4994 | .addReg(RegNo: DestReg1, flags: RegState::Define | getUndefRegState(B: IsUndef), SubReg: SubIdx1) |
4995 | .addFrameIndex(Idx: FI) |
4996 | .addImm(Val: 0) |
4997 | .addMemOperand(MMO); |
4998 | } |
4999 | |
5000 | void AArch64InstrInfo::loadRegFromStackSlot(MachineBasicBlock &MBB, |
5001 | MachineBasicBlock::iterator MBBI, |
5002 | Register DestReg, int FI, |
5003 | const TargetRegisterClass *RC, |
5004 | const TargetRegisterInfo *TRI, |
5005 | Register VReg) const { |
5006 | MachineFunction &MF = *MBB.getParent(); |
5007 | MachineFrameInfo &MFI = MF.getFrameInfo(); |
5008 | MachinePointerInfo PtrInfo = MachinePointerInfo::getFixedStack(MF, FI); |
5009 | MachineMemOperand *MMO = |
5010 | MF.getMachineMemOperand(PtrInfo, F: MachineMemOperand::MOLoad, |
5011 | Size: MFI.getObjectSize(ObjectIdx: FI), BaseAlignment: MFI.getObjectAlign(ObjectIdx: FI)); |
5012 | |
5013 | unsigned Opc = 0; |
5014 | bool Offset = true; |
5015 | unsigned StackID = TargetStackID::Default; |
5016 | Register PNRReg = MCRegister::NoRegister; |
5017 | switch (TRI->getSpillSize(RC: *RC)) { |
5018 | case 1: |
5019 | if (AArch64::FPR8RegClass.hasSubClassEq(RC)) |
5020 | Opc = AArch64::LDRBui; |
5021 | break; |
5022 | case 2: { |
5023 | bool IsPNR = AArch64::PNRRegClass.hasSubClassEq(RC); |
5024 | if (AArch64::FPR16RegClass.hasSubClassEq(RC)) |
5025 | Opc = AArch64::LDRHui; |
5026 | else if (IsPNR || AArch64::PPRRegClass.hasSubClassEq(RC)) { |
5027 | assert(Subtarget.isSVEorStreamingSVEAvailable() && |
5028 | "Unexpected register load without SVE load instructions" ); |
5029 | if (IsPNR) |
5030 | PNRReg = DestReg; |
5031 | Opc = AArch64::LDR_PXI; |
5032 | StackID = TargetStackID::ScalableVector; |
5033 | } |
5034 | break; |
5035 | } |
5036 | case 4: |
5037 | if (AArch64::GPR32allRegClass.hasSubClassEq(RC)) { |
5038 | Opc = AArch64::LDRWui; |
5039 | if (DestReg.isVirtual()) |
5040 | MF.getRegInfo().constrainRegClass(Reg: DestReg, RC: &AArch64::GPR32RegClass); |
5041 | else |
5042 | assert(DestReg != AArch64::WSP); |
5043 | } else if (AArch64::FPR32RegClass.hasSubClassEq(RC)) |
5044 | Opc = AArch64::LDRSui; |
5045 | else if (AArch64::PPR2RegClass.hasSubClassEq(RC)) { |
5046 | Opc = AArch64::LDR_PPXI; |
5047 | StackID = TargetStackID::ScalableVector; |
5048 | } |
5049 | break; |
5050 | case 8: |
5051 | if (AArch64::GPR64allRegClass.hasSubClassEq(RC)) { |
5052 | Opc = AArch64::LDRXui; |
5053 | if (DestReg.isVirtual()) |
5054 | MF.getRegInfo().constrainRegClass(Reg: DestReg, RC: &AArch64::GPR64RegClass); |
5055 | else |
5056 | assert(DestReg != AArch64::SP); |
5057 | } else if (AArch64::FPR64RegClass.hasSubClassEq(RC)) { |
5058 | Opc = AArch64::LDRDui; |
5059 | } else if (AArch64::WSeqPairsClassRegClass.hasSubClassEq(RC)) { |
5060 | loadRegPairFromStackSlot(TRI: getRegisterInfo(), MBB, InsertBefore: MBBI, |
5061 | MCID: get(Opcode: AArch64::LDPWi), DestReg, SubIdx0: AArch64::sube32, |
5062 | SubIdx1: AArch64::subo32, FI, MMO); |
5063 | return; |
5064 | } |
5065 | break; |
5066 | case 16: |
5067 | if (AArch64::FPR128RegClass.hasSubClassEq(RC)) |
5068 | Opc = AArch64::LDRQui; |
5069 | else if (AArch64::DDRegClass.hasSubClassEq(RC)) { |
5070 | assert(Subtarget.hasNEON() && "Unexpected register load without NEON" ); |
5071 | Opc = AArch64::LD1Twov1d; |
5072 | Offset = false; |
5073 | } else if (AArch64::XSeqPairsClassRegClass.hasSubClassEq(RC)) { |
5074 | loadRegPairFromStackSlot(TRI: getRegisterInfo(), MBB, InsertBefore: MBBI, |
5075 | MCID: get(Opcode: AArch64::LDPXi), DestReg, SubIdx0: AArch64::sube64, |
5076 | SubIdx1: AArch64::subo64, FI, MMO); |
5077 | return; |
5078 | } else if (AArch64::ZPRRegClass.hasSubClassEq(RC)) { |
5079 | assert(Subtarget.isSVEorStreamingSVEAvailable() && |
5080 | "Unexpected register load without SVE load instructions" ); |
5081 | Opc = AArch64::LDR_ZXI; |
5082 | StackID = TargetStackID::ScalableVector; |
5083 | } |
5084 | break; |
5085 | case 24: |
5086 | if (AArch64::DDDRegClass.hasSubClassEq(RC)) { |
5087 | assert(Subtarget.hasNEON() && "Unexpected register load without NEON" ); |
5088 | Opc = AArch64::LD1Threev1d; |
5089 | Offset = false; |
5090 | } |
5091 | break; |
5092 | case 32: |
5093 | if (AArch64::DDDDRegClass.hasSubClassEq(RC)) { |
5094 | assert(Subtarget.hasNEON() && "Unexpected register load without NEON" ); |
5095 | Opc = AArch64::LD1Fourv1d; |
5096 | Offset = false; |
5097 | } else if (AArch64::QQRegClass.hasSubClassEq(RC)) { |
5098 | assert(Subtarget.hasNEON() && "Unexpected register load without NEON" ); |
5099 | Opc = AArch64::LD1Twov2d; |
5100 | Offset = false; |
5101 | } else if (AArch64::ZPR2RegClass.hasSubClassEq(RC) || |
5102 | AArch64::ZPR2StridedOrContiguousRegClass.hasSubClassEq(RC)) { |
5103 | assert(Subtarget.isSVEorStreamingSVEAvailable() && |
5104 | "Unexpected register load without SVE load instructions" ); |
5105 | Opc = AArch64::LDR_ZZXI; |
5106 | StackID = TargetStackID::ScalableVector; |
5107 | } |
5108 | break; |
5109 | case 48: |
5110 | if (AArch64::QQQRegClass.hasSubClassEq(RC)) { |
5111 | assert(Subtarget.hasNEON() && "Unexpected register load without NEON" ); |
5112 | Opc = AArch64::LD1Threev2d; |
5113 | Offset = false; |
5114 | } else if (AArch64::ZPR3RegClass.hasSubClassEq(RC)) { |
5115 | assert(Subtarget.isSVEorStreamingSVEAvailable() && |
5116 | "Unexpected register load without SVE load instructions" ); |
5117 | Opc = AArch64::LDR_ZZZXI; |
5118 | StackID = TargetStackID::ScalableVector; |
5119 | } |
5120 | break; |
5121 | case 64: |
5122 | if (AArch64::QQQQRegClass.hasSubClassEq(RC)) { |
5123 | assert(Subtarget.hasNEON() && "Unexpected register load without NEON" ); |
5124 | Opc = AArch64::LD1Fourv2d; |
5125 | Offset = false; |
5126 | } else if (AArch64::ZPR4RegClass.hasSubClassEq(RC) || |
5127 | AArch64::ZPR4StridedOrContiguousRegClass.hasSubClassEq(RC)) { |
5128 | assert(Subtarget.isSVEorStreamingSVEAvailable() && |
5129 | "Unexpected register load without SVE load instructions" ); |
5130 | Opc = AArch64::LDR_ZZZZXI; |
5131 | StackID = TargetStackID::ScalableVector; |
5132 | } |
5133 | break; |
5134 | } |
5135 | |
5136 | assert(Opc && "Unknown register class" ); |
5137 | MFI.setStackID(ObjectIdx: FI, ID: StackID); |
5138 | |
5139 | const MachineInstrBuilder MI = BuildMI(BB&: MBB, I: MBBI, MIMD: DebugLoc(), MCID: get(Opcode: Opc)) |
5140 | .addReg(RegNo: DestReg, flags: getDefRegState(B: true)) |
5141 | .addFrameIndex(Idx: FI); |
5142 | if (Offset) |
5143 | MI.addImm(Val: 0); |
5144 | if (PNRReg.isValid() && !PNRReg.isVirtual()) |
5145 | MI.addDef(RegNo: PNRReg, Flags: RegState::Implicit); |
5146 | MI.addMemOperand(MMO); |
5147 | } |
5148 | |
5149 | bool llvm::isNZCVTouchedInInstructionRange(const MachineInstr &DefMI, |
5150 | const MachineInstr &UseMI, |
5151 | const TargetRegisterInfo *TRI) { |
5152 | return any_of(Range: instructionsWithoutDebug(It: std::next(x: DefMI.getIterator()), |
5153 | End: UseMI.getIterator()), |
5154 | P: [TRI](const MachineInstr &I) { |
5155 | return I.modifiesRegister(Reg: AArch64::NZCV, TRI) || |
5156 | I.readsRegister(Reg: AArch64::NZCV, TRI); |
5157 | }); |
5158 | } |
5159 | |
5160 | void AArch64InstrInfo::decomposeStackOffsetForDwarfOffsets( |
5161 | const StackOffset &Offset, int64_t &ByteSized, int64_t &VGSized) { |
5162 | // The smallest scalable element supported by scaled SVE addressing |
5163 | // modes are predicates, which are 2 scalable bytes in size. So the scalable |
5164 | // byte offset must always be a multiple of 2. |
5165 | assert(Offset.getScalable() % 2 == 0 && "Invalid frame offset" ); |
5166 | |
5167 | // VGSized offsets are divided by '2', because the VG register is the |
5168 | // the number of 64bit granules as opposed to 128bit vector chunks, |
5169 | // which is how the 'n' in e.g. MVT::nxv1i8 is modelled. |
5170 | // So, for a stack offset of 16 MVT::nxv1i8's, the size is n x 16 bytes. |
5171 | // VG = n * 2 and the dwarf offset must be VG * 8 bytes. |
5172 | ByteSized = Offset.getFixed(); |
5173 | VGSized = Offset.getScalable() / 2; |
5174 | } |
5175 | |
5176 | /// Returns the offset in parts to which this frame offset can be |
5177 | /// decomposed for the purpose of describing a frame offset. |
5178 | /// For non-scalable offsets this is simply its byte size. |
5179 | void AArch64InstrInfo::decomposeStackOffsetForFrameOffsets( |
5180 | const StackOffset &Offset, int64_t &NumBytes, int64_t &NumPredicateVectors, |
5181 | int64_t &NumDataVectors) { |
5182 | // The smallest scalable element supported by scaled SVE addressing |
5183 | // modes are predicates, which are 2 scalable bytes in size. So the scalable |
5184 | // byte offset must always be a multiple of 2. |
5185 | assert(Offset.getScalable() % 2 == 0 && "Invalid frame offset" ); |
5186 | |
5187 | NumBytes = Offset.getFixed(); |
5188 | NumDataVectors = 0; |
5189 | NumPredicateVectors = Offset.getScalable() / 2; |
5190 | // This method is used to get the offsets to adjust the frame offset. |
5191 | // If the function requires ADDPL to be used and needs more than two ADDPL |
5192 | // instructions, part of the offset is folded into NumDataVectors so that it |
5193 | // uses ADDVL for part of it, reducing the number of ADDPL instructions. |
5194 | if (NumPredicateVectors % 8 == 0 || NumPredicateVectors < -64 || |
5195 | NumPredicateVectors > 62) { |
5196 | NumDataVectors = NumPredicateVectors / 8; |
5197 | NumPredicateVectors -= NumDataVectors * 8; |
5198 | } |
5199 | } |
5200 | |
5201 | // Convenience function to create a DWARF expression for |
5202 | // Expr + NumBytes + NumVGScaledBytes * AArch64::VG |
5203 | static void appendVGScaledOffsetExpr(SmallVectorImpl<char> &Expr, int NumBytes, |
5204 | int NumVGScaledBytes, unsigned VG, |
5205 | llvm::raw_string_ostream &) { |
5206 | uint8_t buffer[16]; |
5207 | |
5208 | if (NumBytes) { |
5209 | Expr.push_back(Elt: dwarf::DW_OP_consts); |
5210 | Expr.append(in_start: buffer, in_end: buffer + encodeSLEB128(Value: NumBytes, p: buffer)); |
5211 | Expr.push_back(Elt: (uint8_t)dwarf::DW_OP_plus); |
5212 | Comment << (NumBytes < 0 ? " - " : " + " ) << std::abs(x: NumBytes); |
5213 | } |
5214 | |
5215 | if (NumVGScaledBytes) { |
5216 | Expr.push_back(Elt: (uint8_t)dwarf::DW_OP_consts); |
5217 | Expr.append(in_start: buffer, in_end: buffer + encodeSLEB128(Value: NumVGScaledBytes, p: buffer)); |
5218 | |
5219 | Expr.push_back(Elt: (uint8_t)dwarf::DW_OP_bregx); |
5220 | Expr.append(in_start: buffer, in_end: buffer + encodeULEB128(Value: VG, p: buffer)); |
5221 | Expr.push_back(Elt: 0); |
5222 | |
5223 | Expr.push_back(Elt: (uint8_t)dwarf::DW_OP_mul); |
5224 | Expr.push_back(Elt: (uint8_t)dwarf::DW_OP_plus); |
5225 | |
5226 | Comment << (NumVGScaledBytes < 0 ? " - " : " + " ) |
5227 | << std::abs(x: NumVGScaledBytes) << " * VG" ; |
5228 | } |
5229 | } |
5230 | |
5231 | // Creates an MCCFIInstruction: |
5232 | // { DW_CFA_def_cfa_expression, ULEB128 (sizeof expr), expr } |
5233 | static MCCFIInstruction createDefCFAExpression(const TargetRegisterInfo &TRI, |
5234 | unsigned Reg, |
5235 | const StackOffset &Offset) { |
5236 | int64_t NumBytes, NumVGScaledBytes; |
5237 | AArch64InstrInfo::decomposeStackOffsetForDwarfOffsets(Offset, ByteSized&: NumBytes, |
5238 | VGSized&: NumVGScaledBytes); |
5239 | std::string ; |
5240 | llvm::raw_string_ostream (CommentBuffer); |
5241 | |
5242 | if (Reg == AArch64::SP) |
5243 | Comment << "sp" ; |
5244 | else if (Reg == AArch64::FP) |
5245 | Comment << "fp" ; |
5246 | else |
5247 | Comment << printReg(Reg, TRI: &TRI); |
5248 | |
5249 | // Build up the expression (Reg + NumBytes + NumVGScaledBytes * AArch64::VG) |
5250 | SmallString<64> Expr; |
5251 | unsigned DwarfReg = TRI.getDwarfRegNum(RegNum: Reg, isEH: true); |
5252 | Expr.push_back(Elt: (uint8_t)(dwarf::DW_OP_breg0 + DwarfReg)); |
5253 | Expr.push_back(Elt: 0); |
5254 | appendVGScaledOffsetExpr(Expr, NumBytes, NumVGScaledBytes, |
5255 | VG: TRI.getDwarfRegNum(RegNum: AArch64::VG, isEH: true), Comment); |
5256 | |
5257 | // Wrap this into DW_CFA_def_cfa. |
5258 | SmallString<64> DefCfaExpr; |
5259 | DefCfaExpr.push_back(Elt: dwarf::DW_CFA_def_cfa_expression); |
5260 | uint8_t buffer[16]; |
5261 | DefCfaExpr.append(in_start: buffer, in_end: buffer + encodeULEB128(Value: Expr.size(), p: buffer)); |
5262 | DefCfaExpr.append(RHS: Expr.str()); |
5263 | return MCCFIInstruction::createEscape(L: nullptr, Vals: DefCfaExpr.str(), Loc: SMLoc(), |
5264 | Comment: Comment.str()); |
5265 | } |
5266 | |
5267 | MCCFIInstruction llvm::createDefCFA(const TargetRegisterInfo &TRI, |
5268 | unsigned FrameReg, unsigned Reg, |
5269 | const StackOffset &Offset, |
5270 | bool LastAdjustmentWasScalable) { |
5271 | if (Offset.getScalable()) |
5272 | return createDefCFAExpression(TRI, Reg, Offset); |
5273 | |
5274 | if (FrameReg == Reg && !LastAdjustmentWasScalable) |
5275 | return MCCFIInstruction::cfiDefCfaOffset(L: nullptr, Offset: int(Offset.getFixed())); |
5276 | |
5277 | unsigned DwarfReg = TRI.getDwarfRegNum(RegNum: Reg, isEH: true); |
5278 | return MCCFIInstruction::cfiDefCfa(L: nullptr, Register: DwarfReg, Offset: (int)Offset.getFixed()); |
5279 | } |
5280 | |
5281 | MCCFIInstruction llvm::createCFAOffset(const TargetRegisterInfo &TRI, |
5282 | unsigned Reg, |
5283 | const StackOffset &OffsetFromDefCFA) { |
5284 | int64_t NumBytes, NumVGScaledBytes; |
5285 | AArch64InstrInfo::decomposeStackOffsetForDwarfOffsets( |
5286 | Offset: OffsetFromDefCFA, ByteSized&: NumBytes, VGSized&: NumVGScaledBytes); |
5287 | |
5288 | unsigned DwarfReg = TRI.getDwarfRegNum(RegNum: Reg, isEH: true); |
5289 | |
5290 | // Non-scalable offsets can use DW_CFA_offset directly. |
5291 | if (!NumVGScaledBytes) |
5292 | return MCCFIInstruction::createOffset(L: nullptr, Register: DwarfReg, Offset: NumBytes); |
5293 | |
5294 | std::string ; |
5295 | llvm::raw_string_ostream (CommentBuffer); |
5296 | Comment << printReg(Reg, TRI: &TRI) << " @ cfa" ; |
5297 | |
5298 | // Build up expression (NumBytes + NumVGScaledBytes * AArch64::VG) |
5299 | SmallString<64> OffsetExpr; |
5300 | appendVGScaledOffsetExpr(Expr&: OffsetExpr, NumBytes, NumVGScaledBytes, |
5301 | VG: TRI.getDwarfRegNum(RegNum: AArch64::VG, isEH: true), Comment); |
5302 | |
5303 | // Wrap this into DW_CFA_expression |
5304 | SmallString<64> CfaExpr; |
5305 | CfaExpr.push_back(Elt: dwarf::DW_CFA_expression); |
5306 | uint8_t buffer[16]; |
5307 | CfaExpr.append(in_start: buffer, in_end: buffer + encodeULEB128(Value: DwarfReg, p: buffer)); |
5308 | CfaExpr.append(in_start: buffer, in_end: buffer + encodeULEB128(Value: OffsetExpr.size(), p: buffer)); |
5309 | CfaExpr.append(RHS: OffsetExpr.str()); |
5310 | |
5311 | return MCCFIInstruction::createEscape(L: nullptr, Vals: CfaExpr.str(), Loc: SMLoc(), |
5312 | Comment: Comment.str()); |
5313 | } |
5314 | |
5315 | // Helper function to emit a frame offset adjustment from a given |
5316 | // pointer (SrcReg), stored into DestReg. This function is explicit |
5317 | // in that it requires the opcode. |
5318 | static void emitFrameOffsetAdj(MachineBasicBlock &MBB, |
5319 | MachineBasicBlock::iterator MBBI, |
5320 | const DebugLoc &DL, unsigned DestReg, |
5321 | unsigned SrcReg, int64_t Offset, unsigned Opc, |
5322 | const TargetInstrInfo *TII, |
5323 | MachineInstr::MIFlag Flag, bool NeedsWinCFI, |
5324 | bool *HasWinCFI, bool EmitCFAOffset, |
5325 | StackOffset CFAOffset, unsigned FrameReg) { |
5326 | int Sign = 1; |
5327 | unsigned MaxEncoding, ShiftSize; |
5328 | switch (Opc) { |
5329 | case AArch64::ADDXri: |
5330 | case AArch64::ADDSXri: |
5331 | case AArch64::SUBXri: |
5332 | case AArch64::SUBSXri: |
5333 | MaxEncoding = 0xfff; |
5334 | ShiftSize = 12; |
5335 | break; |
5336 | case AArch64::ADDVL_XXI: |
5337 | case AArch64::ADDPL_XXI: |
5338 | case AArch64::ADDSVL_XXI: |
5339 | case AArch64::ADDSPL_XXI: |
5340 | MaxEncoding = 31; |
5341 | ShiftSize = 0; |
5342 | if (Offset < 0) { |
5343 | MaxEncoding = 32; |
5344 | Sign = -1; |
5345 | Offset = -Offset; |
5346 | } |
5347 | break; |
5348 | default: |
5349 | llvm_unreachable("Unsupported opcode" ); |
5350 | } |
5351 | |
5352 | // `Offset` can be in bytes or in "scalable bytes". |
5353 | int VScale = 1; |
5354 | if (Opc == AArch64::ADDVL_XXI || Opc == AArch64::ADDSVL_XXI) |
5355 | VScale = 16; |
5356 | else if (Opc == AArch64::ADDPL_XXI || Opc == AArch64::ADDSPL_XXI) |
5357 | VScale = 2; |
5358 | |
5359 | // FIXME: If the offset won't fit in 24-bits, compute the offset into a |
5360 | // scratch register. If DestReg is a virtual register, use it as the |
5361 | // scratch register; otherwise, create a new virtual register (to be |
5362 | // replaced by the scavenger at the end of PEI). That case can be optimized |
5363 | // slightly if DestReg is SP which is always 16-byte aligned, so the scratch |
5364 | // register can be loaded with offset%8 and the add/sub can use an extending |
5365 | // instruction with LSL#3. |
5366 | // Currently the function handles any offsets but generates a poor sequence |
5367 | // of code. |
5368 | // assert(Offset < (1 << 24) && "unimplemented reg plus immediate"); |
5369 | |
5370 | const unsigned MaxEncodableValue = MaxEncoding << ShiftSize; |
5371 | Register TmpReg = DestReg; |
5372 | if (TmpReg == AArch64::XZR) |
5373 | TmpReg = MBB.getParent()->getRegInfo().createVirtualRegister( |
5374 | RegClass: &AArch64::GPR64RegClass); |
5375 | do { |
5376 | uint64_t ThisVal = std::min<uint64_t>(a: Offset, b: MaxEncodableValue); |
5377 | unsigned LocalShiftSize = 0; |
5378 | if (ThisVal > MaxEncoding) { |
5379 | ThisVal = ThisVal >> ShiftSize; |
5380 | LocalShiftSize = ShiftSize; |
5381 | } |
5382 | assert((ThisVal >> ShiftSize) <= MaxEncoding && |
5383 | "Encoding cannot handle value that big" ); |
5384 | |
5385 | Offset -= ThisVal << LocalShiftSize; |
5386 | if (Offset == 0) |
5387 | TmpReg = DestReg; |
5388 | auto MBI = BuildMI(BB&: MBB, I: MBBI, MIMD: DL, MCID: TII->get(Opcode: Opc), DestReg: TmpReg) |
5389 | .addReg(RegNo: SrcReg) |
5390 | .addImm(Val: Sign * (int)ThisVal); |
5391 | if (ShiftSize) |
5392 | MBI = MBI.addImm( |
5393 | Val: AArch64_AM::getShifterImm(ST: AArch64_AM::LSL, Imm: LocalShiftSize)); |
5394 | MBI = MBI.setMIFlag(Flag); |
5395 | |
5396 | auto Change = |
5397 | VScale == 1 |
5398 | ? StackOffset::getFixed(Fixed: ThisVal << LocalShiftSize) |
5399 | : StackOffset::getScalable(Scalable: VScale * (ThisVal << LocalShiftSize)); |
5400 | if (Sign == -1 || Opc == AArch64::SUBXri || Opc == AArch64::SUBSXri) |
5401 | CFAOffset += Change; |
5402 | else |
5403 | CFAOffset -= Change; |
5404 | if (EmitCFAOffset && DestReg == TmpReg) { |
5405 | MachineFunction &MF = *MBB.getParent(); |
5406 | const TargetSubtargetInfo &STI = MF.getSubtarget(); |
5407 | const TargetRegisterInfo &TRI = *STI.getRegisterInfo(); |
5408 | |
5409 | unsigned CFIIndex = MF.addFrameInst( |
5410 | Inst: createDefCFA(TRI, FrameReg, Reg: DestReg, Offset: CFAOffset, LastAdjustmentWasScalable: VScale != 1)); |
5411 | BuildMI(BB&: MBB, I: MBBI, MIMD: DL, MCID: TII->get(Opcode: TargetOpcode::CFI_INSTRUCTION)) |
5412 | .addCFIIndex(CFIIndex) |
5413 | .setMIFlags(Flag); |
5414 | } |
5415 | |
5416 | if (NeedsWinCFI) { |
5417 | assert(Sign == 1 && "SEH directives should always have a positive sign" ); |
5418 | int Imm = (int)(ThisVal << LocalShiftSize); |
5419 | if ((DestReg == AArch64::FP && SrcReg == AArch64::SP) || |
5420 | (SrcReg == AArch64::FP && DestReg == AArch64::SP)) { |
5421 | if (HasWinCFI) |
5422 | *HasWinCFI = true; |
5423 | if (Imm == 0) |
5424 | BuildMI(BB&: MBB, I: MBBI, MIMD: DL, MCID: TII->get(Opcode: AArch64::SEH_SetFP)).setMIFlag(Flag); |
5425 | else |
5426 | BuildMI(BB&: MBB, I: MBBI, MIMD: DL, MCID: TII->get(Opcode: AArch64::SEH_AddFP)) |
5427 | .addImm(Val: Imm) |
5428 | .setMIFlag(Flag); |
5429 | assert(Offset == 0 && "Expected remaining offset to be zero to " |
5430 | "emit a single SEH directive" ); |
5431 | } else if (DestReg == AArch64::SP) { |
5432 | if (HasWinCFI) |
5433 | *HasWinCFI = true; |
5434 | assert(SrcReg == AArch64::SP && "Unexpected SrcReg for SEH_StackAlloc" ); |
5435 | BuildMI(BB&: MBB, I: MBBI, MIMD: DL, MCID: TII->get(Opcode: AArch64::SEH_StackAlloc)) |
5436 | .addImm(Val: Imm) |
5437 | .setMIFlag(Flag); |
5438 | } |
5439 | } |
5440 | |
5441 | SrcReg = TmpReg; |
5442 | } while (Offset); |
5443 | } |
5444 | |
5445 | void llvm::emitFrameOffset(MachineBasicBlock &MBB, |
5446 | MachineBasicBlock::iterator MBBI, const DebugLoc &DL, |
5447 | unsigned DestReg, unsigned SrcReg, |
5448 | StackOffset Offset, const TargetInstrInfo *TII, |
5449 | MachineInstr::MIFlag Flag, bool SetNZCV, |
5450 | bool NeedsWinCFI, bool *HasWinCFI, |
5451 | bool EmitCFAOffset, StackOffset CFAOffset, |
5452 | unsigned FrameReg) { |
5453 | // If a function is marked as arm_locally_streaming, then the runtime value of |
5454 | // vscale in the prologue/epilogue is different the runtime value of vscale |
5455 | // in the function's body. To avoid having to consider multiple vscales, |
5456 | // we can use `addsvl` to allocate any scalable stack-slots, which under |
5457 | // most circumstances will be only locals, not callee-save slots. |
5458 | const Function &F = MBB.getParent()->getFunction(); |
5459 | bool UseSVL = F.hasFnAttribute(Kind: "aarch64_pstate_sm_body" ); |
5460 | |
5461 | int64_t Bytes, NumPredicateVectors, NumDataVectors; |
5462 | AArch64InstrInfo::decomposeStackOffsetForFrameOffsets( |
5463 | Offset, NumBytes&: Bytes, NumPredicateVectors, NumDataVectors); |
5464 | |
5465 | // First emit non-scalable frame offsets, or a simple 'mov'. |
5466 | if (Bytes || (!Offset && SrcReg != DestReg)) { |
5467 | assert((DestReg != AArch64::SP || Bytes % 8 == 0) && |
5468 | "SP increment/decrement not 8-byte aligned" ); |
5469 | unsigned Opc = SetNZCV ? AArch64::ADDSXri : AArch64::ADDXri; |
5470 | if (Bytes < 0) { |
5471 | Bytes = -Bytes; |
5472 | Opc = SetNZCV ? AArch64::SUBSXri : AArch64::SUBXri; |
5473 | } |
5474 | emitFrameOffsetAdj(MBB, MBBI, DL, DestReg, SrcReg, Offset: Bytes, Opc, TII, Flag, |
5475 | NeedsWinCFI, HasWinCFI, EmitCFAOffset, CFAOffset, |
5476 | FrameReg); |
5477 | CFAOffset += (Opc == AArch64::ADDXri || Opc == AArch64::ADDSXri) |
5478 | ? StackOffset::getFixed(Fixed: -Bytes) |
5479 | : StackOffset::getFixed(Fixed: Bytes); |
5480 | SrcReg = DestReg; |
5481 | FrameReg = DestReg; |
5482 | } |
5483 | |
5484 | assert(!(SetNZCV && (NumPredicateVectors || NumDataVectors)) && |
5485 | "SetNZCV not supported with SVE vectors" ); |
5486 | assert(!(NeedsWinCFI && (NumPredicateVectors || NumDataVectors)) && |
5487 | "WinCFI not supported with SVE vectors" ); |
5488 | |
5489 | if (NumDataVectors) { |
5490 | emitFrameOffsetAdj(MBB, MBBI, DL, DestReg, SrcReg, Offset: NumDataVectors, |
5491 | Opc: UseSVL ? AArch64::ADDSVL_XXI : AArch64::ADDVL_XXI, |
5492 | TII, Flag, NeedsWinCFI, HasWinCFI: nullptr, EmitCFAOffset, |
5493 | CFAOffset, FrameReg); |
5494 | CFAOffset += StackOffset::getScalable(Scalable: -NumDataVectors * 16); |
5495 | SrcReg = DestReg; |
5496 | } |
5497 | |
5498 | if (NumPredicateVectors) { |
5499 | assert(DestReg != AArch64::SP && "Unaligned access to SP" ); |
5500 | emitFrameOffsetAdj(MBB, MBBI, DL, DestReg, SrcReg, Offset: NumPredicateVectors, |
5501 | Opc: UseSVL ? AArch64::ADDSPL_XXI : AArch64::ADDPL_XXI, |
5502 | TII, Flag, NeedsWinCFI, HasWinCFI: nullptr, EmitCFAOffset, |
5503 | CFAOffset, FrameReg); |
5504 | } |
5505 | } |
5506 | |
5507 | MachineInstr *AArch64InstrInfo::foldMemoryOperandImpl( |
5508 | MachineFunction &MF, MachineInstr &MI, ArrayRef<unsigned> Ops, |
5509 | MachineBasicBlock::iterator InsertPt, int FrameIndex, |
5510 | LiveIntervals *LIS, VirtRegMap *VRM) const { |
5511 | // This is a bit of a hack. Consider this instruction: |
5512 | // |
5513 | // %0 = COPY %sp; GPR64all:%0 |
5514 | // |
5515 | // We explicitly chose GPR64all for the virtual register so such a copy might |
5516 | // be eliminated by RegisterCoalescer. However, that may not be possible, and |
5517 | // %0 may even spill. We can't spill %sp, and since it is in the GPR64all |
5518 | // register class, TargetInstrInfo::foldMemoryOperand() is going to try. |
5519 | // |
5520 | // To prevent that, we are going to constrain the %0 register class here. |
5521 | if (MI.isFullCopy()) { |
5522 | Register DstReg = MI.getOperand(i: 0).getReg(); |
5523 | Register SrcReg = MI.getOperand(i: 1).getReg(); |
5524 | if (SrcReg == AArch64::SP && DstReg.isVirtual()) { |
5525 | MF.getRegInfo().constrainRegClass(Reg: DstReg, RC: &AArch64::GPR64RegClass); |
5526 | return nullptr; |
5527 | } |
5528 | if (DstReg == AArch64::SP && SrcReg.isVirtual()) { |
5529 | MF.getRegInfo().constrainRegClass(Reg: SrcReg, RC: &AArch64::GPR64RegClass); |
5530 | return nullptr; |
5531 | } |
5532 | // Nothing can folded with copy from/to NZCV. |
5533 | if (SrcReg == AArch64::NZCV || DstReg == AArch64::NZCV) |
5534 | return nullptr; |
5535 | } |
5536 | |
5537 | // Handle the case where a copy is being spilled or filled but the source |
5538 | // and destination register class don't match. For example: |
5539 | // |
5540 | // %0 = COPY %xzr; GPR64common:%0 |
5541 | // |
5542 | // In this case we can still safely fold away the COPY and generate the |
5543 | // following spill code: |
5544 | // |
5545 | // STRXui %xzr, %stack.0 |
5546 | // |
5547 | // This also eliminates spilled cross register class COPYs (e.g. between x and |
5548 | // d regs) of the same size. For example: |
5549 | // |
5550 | // %0 = COPY %1; GPR64:%0, FPR64:%1 |
5551 | // |
5552 | // will be filled as |
5553 | // |
5554 | // LDRDui %0, fi<#0> |
5555 | // |
5556 | // instead of |
5557 | // |
5558 | // LDRXui %Temp, fi<#0> |
5559 | // %0 = FMOV %Temp |
5560 | // |
5561 | if (MI.isCopy() && Ops.size() == 1 && |
5562 | // Make sure we're only folding the explicit COPY defs/uses. |
5563 | (Ops[0] == 0 || Ops[0] == 1)) { |
5564 | bool IsSpill = Ops[0] == 0; |
5565 | bool IsFill = !IsSpill; |
5566 | const TargetRegisterInfo &TRI = *MF.getSubtarget().getRegisterInfo(); |
5567 | const MachineRegisterInfo &MRI = MF.getRegInfo(); |
5568 | MachineBasicBlock &MBB = *MI.getParent(); |
5569 | const MachineOperand &DstMO = MI.getOperand(i: 0); |
5570 | const MachineOperand &SrcMO = MI.getOperand(i: 1); |
5571 | Register DstReg = DstMO.getReg(); |
5572 | Register SrcReg = SrcMO.getReg(); |
5573 | // This is slightly expensive to compute for physical regs since |
5574 | // getMinimalPhysRegClass is slow. |
5575 | auto getRegClass = [&](unsigned Reg) { |
5576 | return Register::isVirtualRegister(Reg) ? MRI.getRegClass(Reg) |
5577 | : TRI.getMinimalPhysRegClass(Reg); |
5578 | }; |
5579 | |
5580 | if (DstMO.getSubReg() == 0 && SrcMO.getSubReg() == 0) { |
5581 | assert(TRI.getRegSizeInBits(*getRegClass(DstReg)) == |
5582 | TRI.getRegSizeInBits(*getRegClass(SrcReg)) && |
5583 | "Mismatched register size in non subreg COPY" ); |
5584 | if (IsSpill) |
5585 | storeRegToStackSlot(MBB, MBBI: InsertPt, SrcReg, isKill: SrcMO.isKill(), FI: FrameIndex, |
5586 | RC: getRegClass(SrcReg), TRI: &TRI, VReg: Register()); |
5587 | else |
5588 | loadRegFromStackSlot(MBB, MBBI: InsertPt, DestReg: DstReg, FI: FrameIndex, |
5589 | RC: getRegClass(DstReg), TRI: &TRI, VReg: Register()); |
5590 | return &*--InsertPt; |
5591 | } |
5592 | |
5593 | // Handle cases like spilling def of: |
5594 | // |
5595 | // %0:sub_32<def,read-undef> = COPY %wzr; GPR64common:%0 |
5596 | // |
5597 | // where the physical register source can be widened and stored to the full |
5598 | // virtual reg destination stack slot, in this case producing: |
5599 | // |
5600 | // STRXui %xzr, %stack.0 |
5601 | // |
5602 | if (IsSpill && DstMO.isUndef() && SrcReg == AArch64::WZR && |
5603 | TRI.getRegSizeInBits(RC: *getRegClass(DstReg)) == 64) { |
5604 | assert(SrcMO.getSubReg() == 0 && |
5605 | "Unexpected subreg on physical register" ); |
5606 | storeRegToStackSlot(MBB, MBBI: InsertPt, SrcReg: AArch64::XZR, isKill: SrcMO.isKill(), |
5607 | FI: FrameIndex, RC: &AArch64::GPR64RegClass, TRI: &TRI, |
5608 | VReg: Register()); |
5609 | return &*--InsertPt; |
5610 | } |
5611 | |
5612 | // Handle cases like filling use of: |
5613 | // |
5614 | // %0:sub_32<def,read-undef> = COPY %1; GPR64:%0, GPR32:%1 |
5615 | // |
5616 | // where we can load the full virtual reg source stack slot, into the subreg |
5617 | // destination, in this case producing: |
5618 | // |
5619 | // LDRWui %0:sub_32<def,read-undef>, %stack.0 |
5620 | // |
5621 | if (IsFill && SrcMO.getSubReg() == 0 && DstMO.isUndef()) { |
5622 | const TargetRegisterClass *FillRC; |
5623 | switch (DstMO.getSubReg()) { |
5624 | default: |
5625 | FillRC = nullptr; |
5626 | break; |
5627 | case AArch64::sub_32: |
5628 | FillRC = &AArch64::GPR32RegClass; |
5629 | break; |
5630 | case AArch64::ssub: |
5631 | FillRC = &AArch64::FPR32RegClass; |
5632 | break; |
5633 | case AArch64::dsub: |
5634 | FillRC = &AArch64::FPR64RegClass; |
5635 | break; |
5636 | } |
5637 | |
5638 | if (FillRC) { |
5639 | assert(TRI.getRegSizeInBits(*getRegClass(SrcReg)) == |
5640 | TRI.getRegSizeInBits(*FillRC) && |
5641 | "Mismatched regclass size on folded subreg COPY" ); |
5642 | loadRegFromStackSlot(MBB, MBBI: InsertPt, DestReg: DstReg, FI: FrameIndex, RC: FillRC, TRI: &TRI, |
5643 | VReg: Register()); |
5644 | MachineInstr &LoadMI = *--InsertPt; |
5645 | MachineOperand &LoadDst = LoadMI.getOperand(i: 0); |
5646 | assert(LoadDst.getSubReg() == 0 && "unexpected subreg on fill load" ); |
5647 | LoadDst.setSubReg(DstMO.getSubReg()); |
5648 | LoadDst.setIsUndef(); |
5649 | return &LoadMI; |
5650 | } |
5651 | } |
5652 | } |
5653 | |
5654 | // Cannot fold. |
5655 | return nullptr; |
5656 | } |
5657 | |
5658 | int llvm::isAArch64FrameOffsetLegal(const MachineInstr &MI, |
5659 | StackOffset &SOffset, |
5660 | bool *OutUseUnscaledOp, |
5661 | unsigned *OutUnscaledOp, |
5662 | int64_t *EmittableOffset) { |
5663 | // Set output values in case of early exit. |
5664 | if (EmittableOffset) |
5665 | *EmittableOffset = 0; |
5666 | if (OutUseUnscaledOp) |
5667 | *OutUseUnscaledOp = false; |
5668 | if (OutUnscaledOp) |
5669 | *OutUnscaledOp = 0; |
5670 | |
5671 | // Exit early for structured vector spills/fills as they can't take an |
5672 | // immediate offset. |
5673 | switch (MI.getOpcode()) { |
5674 | default: |
5675 | break; |
5676 | case AArch64::LD1Rv1d: |
5677 | case AArch64::LD1Rv2s: |
5678 | case AArch64::LD1Rv2d: |
5679 | case AArch64::LD1Rv4h: |
5680 | case AArch64::LD1Rv4s: |
5681 | case AArch64::LD1Rv8b: |
5682 | case AArch64::LD1Rv8h: |
5683 | case AArch64::LD1Rv16b: |
5684 | case AArch64::LD1Twov2d: |
5685 | case AArch64::LD1Threev2d: |
5686 | case AArch64::LD1Fourv2d: |
5687 | case AArch64::LD1Twov1d: |
5688 | case AArch64::LD1Threev1d: |
5689 | case AArch64::LD1Fourv1d: |
5690 | case AArch64::ST1Twov2d: |
5691 | case AArch64::ST1Threev2d: |
5692 | case AArch64::ST1Fourv2d: |
5693 | case AArch64::ST1Twov1d: |
5694 | case AArch64::ST1Threev1d: |
5695 | case AArch64::ST1Fourv1d: |
5696 | case AArch64::ST1i8: |
5697 | case AArch64::ST1i16: |
5698 | case AArch64::ST1i32: |
5699 | case AArch64::ST1i64: |
5700 | case AArch64::IRG: |
5701 | case AArch64::IRGstack: |
5702 | case AArch64::STGloop: |
5703 | case AArch64::STZGloop: |
5704 | return AArch64FrameOffsetCannotUpdate; |
5705 | } |
5706 | |
5707 | // Get the min/max offset and the scale. |
5708 | TypeSize ScaleValue(0U, false), Width(0U, false); |
5709 | int64_t MinOff, MaxOff; |
5710 | if (!AArch64InstrInfo::getMemOpInfo(Opcode: MI.getOpcode(), Scale&: ScaleValue, Width, MinOffset&: MinOff, |
5711 | MaxOffset&: MaxOff)) |
5712 | llvm_unreachable("unhandled opcode in isAArch64FrameOffsetLegal" ); |
5713 | |
5714 | // Construct the complete offset. |
5715 | bool IsMulVL = ScaleValue.isScalable(); |
5716 | unsigned Scale = ScaleValue.getKnownMinValue(); |
5717 | int64_t Offset = IsMulVL ? SOffset.getScalable() : SOffset.getFixed(); |
5718 | |
5719 | const MachineOperand &ImmOpnd = |
5720 | MI.getOperand(i: AArch64InstrInfo::getLoadStoreImmIdx(Opc: MI.getOpcode())); |
5721 | Offset += ImmOpnd.getImm() * Scale; |
5722 | |
5723 | // If the offset doesn't match the scale, we rewrite the instruction to |
5724 | // use the unscaled instruction instead. Likewise, if we have a negative |
5725 | // offset and there is an unscaled op to use. |
5726 | std::optional<unsigned> UnscaledOp = |
5727 | AArch64InstrInfo::getUnscaledLdSt(Opc: MI.getOpcode()); |
5728 | bool useUnscaledOp = UnscaledOp && (Offset % Scale || Offset < 0); |
5729 | if (useUnscaledOp && |
5730 | !AArch64InstrInfo::getMemOpInfo(Opcode: *UnscaledOp, Scale&: ScaleValue, Width, MinOffset&: MinOff, |
5731 | MaxOffset&: MaxOff)) |
5732 | llvm_unreachable("unhandled opcode in isAArch64FrameOffsetLegal" ); |
5733 | |
5734 | Scale = ScaleValue.getKnownMinValue(); |
5735 | assert(IsMulVL == ScaleValue.isScalable() && |
5736 | "Unscaled opcode has different value for scalable" ); |
5737 | |
5738 | int64_t Remainder = Offset % Scale; |
5739 | assert(!(Remainder && useUnscaledOp) && |
5740 | "Cannot have remainder when using unscaled op" ); |
5741 | |
5742 | assert(MinOff < MaxOff && "Unexpected Min/Max offsets" ); |
5743 | int64_t NewOffset = Offset / Scale; |
5744 | if (MinOff <= NewOffset && NewOffset <= MaxOff) |
5745 | Offset = Remainder; |
5746 | else { |
5747 | NewOffset = NewOffset < 0 ? MinOff : MaxOff; |
5748 | Offset = Offset - (NewOffset * Scale); |
5749 | } |
5750 | |
5751 | if (EmittableOffset) |
5752 | *EmittableOffset = NewOffset; |
5753 | if (OutUseUnscaledOp) |
5754 | *OutUseUnscaledOp = useUnscaledOp; |
5755 | if (OutUnscaledOp && UnscaledOp) |
5756 | *OutUnscaledOp = *UnscaledOp; |
5757 | |
5758 | if (IsMulVL) |
5759 | SOffset = StackOffset::get(Fixed: SOffset.getFixed(), Scalable: Offset); |
5760 | else |
5761 | SOffset = StackOffset::get(Fixed: Offset, Scalable: SOffset.getScalable()); |
5762 | return AArch64FrameOffsetCanUpdate | |
5763 | (SOffset ? 0 : AArch64FrameOffsetIsLegal); |
5764 | } |
5765 | |
5766 | bool llvm::rewriteAArch64FrameIndex(MachineInstr &MI, unsigned FrameRegIdx, |
5767 | unsigned FrameReg, StackOffset &Offset, |
5768 | const AArch64InstrInfo *TII) { |
5769 | unsigned Opcode = MI.getOpcode(); |
5770 | unsigned ImmIdx = FrameRegIdx + 1; |
5771 | |
5772 | if (Opcode == AArch64::ADDSXri || Opcode == AArch64::ADDXri) { |
5773 | Offset += StackOffset::getFixed(Fixed: MI.getOperand(i: ImmIdx).getImm()); |
5774 | emitFrameOffset(MBB&: *MI.getParent(), MBBI: MI, DL: MI.getDebugLoc(), |
5775 | DestReg: MI.getOperand(i: 0).getReg(), SrcReg: FrameReg, Offset, TII, |
5776 | Flag: MachineInstr::NoFlags, SetNZCV: (Opcode == AArch64::ADDSXri)); |
5777 | MI.eraseFromParent(); |
5778 | Offset = StackOffset(); |
5779 | return true; |
5780 | } |
5781 | |
5782 | int64_t NewOffset; |
5783 | unsigned UnscaledOp; |
5784 | bool UseUnscaledOp; |
5785 | int Status = isAArch64FrameOffsetLegal(MI, SOffset&: Offset, OutUseUnscaledOp: &UseUnscaledOp, |
5786 | OutUnscaledOp: &UnscaledOp, EmittableOffset: &NewOffset); |
5787 | if (Status & AArch64FrameOffsetCanUpdate) { |
5788 | if (Status & AArch64FrameOffsetIsLegal) |
5789 | // Replace the FrameIndex with FrameReg. |
5790 | MI.getOperand(i: FrameRegIdx).ChangeToRegister(Reg: FrameReg, isDef: false); |
5791 | if (UseUnscaledOp) |
5792 | MI.setDesc(TII->get(Opcode: UnscaledOp)); |
5793 | |
5794 | MI.getOperand(i: ImmIdx).ChangeToImmediate(ImmVal: NewOffset); |
5795 | return !Offset; |
5796 | } |
5797 | |
5798 | return false; |
5799 | } |
5800 | |
5801 | void AArch64InstrInfo::insertNoop(MachineBasicBlock &MBB, |
5802 | MachineBasicBlock::iterator MI) const { |
5803 | DebugLoc DL; |
5804 | BuildMI(BB&: MBB, I: MI, MIMD: DL, MCID: get(Opcode: AArch64::HINT)).addImm(Val: 0); |
5805 | } |
5806 | |
5807 | MCInst AArch64InstrInfo::getNop() const { |
5808 | return MCInstBuilder(AArch64::HINT).addImm(Val: 0); |
5809 | } |
5810 | |
5811 | // AArch64 supports MachineCombiner. |
5812 | bool AArch64InstrInfo::useMachineCombiner() const { return true; } |
5813 | |
5814 | // True when Opc sets flag |
5815 | static bool isCombineInstrSettingFlag(unsigned Opc) { |
5816 | switch (Opc) { |
5817 | case AArch64::ADDSWrr: |
5818 | case AArch64::ADDSWri: |
5819 | case AArch64::ADDSXrr: |
5820 | case AArch64::ADDSXri: |
5821 | case AArch64::SUBSWrr: |
5822 | case AArch64::SUBSXrr: |
5823 | // Note: MSUB Wd,Wn,Wm,Wi -> Wd = Wi - WnxWm, not Wd=WnxWm - Wi. |
5824 | case AArch64::SUBSWri: |
5825 | case AArch64::SUBSXri: |
5826 | return true; |
5827 | default: |
5828 | break; |
5829 | } |
5830 | return false; |
5831 | } |
5832 | |
5833 | // 32b Opcodes that can be combined with a MUL |
5834 | static bool isCombineInstrCandidate32(unsigned Opc) { |
5835 | switch (Opc) { |
5836 | case AArch64::ADDWrr: |
5837 | case AArch64::ADDWri: |
5838 | case AArch64::SUBWrr: |
5839 | case AArch64::ADDSWrr: |
5840 | case AArch64::ADDSWri: |
5841 | case AArch64::SUBSWrr: |
5842 | // Note: MSUB Wd,Wn,Wm,Wi -> Wd = Wi - WnxWm, not Wd=WnxWm - Wi. |
5843 | case AArch64::SUBWri: |
5844 | case AArch64::SUBSWri: |
5845 | return true; |
5846 | default: |
5847 | break; |
5848 | } |
5849 | return false; |
5850 | } |
5851 | |
5852 | // 64b Opcodes that can be combined with a MUL |
5853 | static bool isCombineInstrCandidate64(unsigned Opc) { |
5854 | switch (Opc) { |
5855 | case AArch64::ADDXrr: |
5856 | case AArch64::ADDXri: |
5857 | case AArch64::SUBXrr: |
5858 | case AArch64::ADDSXrr: |
5859 | case AArch64::ADDSXri: |
5860 | case AArch64::SUBSXrr: |
5861 | // Note: MSUB Wd,Wn,Wm,Wi -> Wd = Wi - WnxWm, not Wd=WnxWm - Wi. |
5862 | case AArch64::SUBXri: |
5863 | case AArch64::SUBSXri: |
5864 | case AArch64::ADDv8i8: |
5865 | case AArch64::ADDv16i8: |
5866 | case AArch64::ADDv4i16: |
5867 | case AArch64::ADDv8i16: |
5868 | case AArch64::ADDv2i32: |
5869 | case AArch64::ADDv4i32: |
5870 | case AArch64::SUBv8i8: |
5871 | case AArch64::SUBv16i8: |
5872 | case AArch64::SUBv4i16: |
5873 | case AArch64::SUBv8i16: |
5874 | case AArch64::SUBv2i32: |
5875 | case AArch64::SUBv4i32: |
5876 | return true; |
5877 | default: |
5878 | break; |
5879 | } |
5880 | return false; |
5881 | } |
5882 | |
5883 | // FP Opcodes that can be combined with a FMUL. |
5884 | static bool isCombineInstrCandidateFP(const MachineInstr &Inst) { |
5885 | switch (Inst.getOpcode()) { |
5886 | default: |
5887 | break; |
5888 | case AArch64::FADDHrr: |
5889 | case AArch64::FADDSrr: |
5890 | case AArch64::FADDDrr: |
5891 | case AArch64::FADDv4f16: |
5892 | case AArch64::FADDv8f16: |
5893 | case AArch64::FADDv2f32: |
5894 | case AArch64::FADDv2f64: |
5895 | case AArch64::FADDv4f32: |
5896 | case AArch64::FSUBHrr: |
5897 | case AArch64::FSUBSrr: |
5898 | case AArch64::FSUBDrr: |
5899 | case AArch64::FSUBv4f16: |
5900 | case AArch64::FSUBv8f16: |
5901 | case AArch64::FSUBv2f32: |
5902 | case AArch64::FSUBv2f64: |
5903 | case AArch64::FSUBv4f32: |
5904 | TargetOptions Options = Inst.getParent()->getParent()->getTarget().Options; |
5905 | // We can fuse FADD/FSUB with FMUL, if fusion is either allowed globally by |
5906 | // the target options or if FADD/FSUB has the contract fast-math flag. |
5907 | return Options.UnsafeFPMath || |
5908 | Options.AllowFPOpFusion == FPOpFusion::Fast || |
5909 | Inst.getFlag(Flag: MachineInstr::FmContract); |
5910 | return true; |
5911 | } |
5912 | return false; |
5913 | } |
5914 | |
5915 | // Opcodes that can be combined with a MUL |
5916 | static bool isCombineInstrCandidate(unsigned Opc) { |
5917 | return (isCombineInstrCandidate32(Opc) || isCombineInstrCandidate64(Opc)); |
5918 | } |
5919 | |
5920 | // |
5921 | // Utility routine that checks if \param MO is defined by an |
5922 | // \param CombineOpc instruction in the basic block \param MBB |
5923 | static bool canCombine(MachineBasicBlock &MBB, MachineOperand &MO, |
5924 | unsigned CombineOpc, unsigned ZeroReg = 0, |
5925 | bool CheckZeroReg = false) { |
5926 | MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo(); |
5927 | MachineInstr *MI = nullptr; |
5928 | |
5929 | if (MO.isReg() && MO.getReg().isVirtual()) |
5930 | MI = MRI.getUniqueVRegDef(Reg: MO.getReg()); |
5931 | // And it needs to be in the trace (otherwise, it won't have a depth). |
5932 | if (!MI || MI->getParent() != &MBB || (unsigned)MI->getOpcode() != CombineOpc) |
5933 | return false; |
5934 | // Must only used by the user we combine with. |
5935 | if (!MRI.hasOneNonDBGUse(RegNo: MI->getOperand(i: 0).getReg())) |
5936 | return false; |
5937 | |
5938 | if (CheckZeroReg) { |
5939 | assert(MI->getNumOperands() >= 4 && MI->getOperand(0).isReg() && |
5940 | MI->getOperand(1).isReg() && MI->getOperand(2).isReg() && |
5941 | MI->getOperand(3).isReg() && "MAdd/MSub must have a least 4 regs" ); |
5942 | // The third input reg must be zero. |
5943 | if (MI->getOperand(i: 3).getReg() != ZeroReg) |
5944 | return false; |
5945 | } |
5946 | |
5947 | if (isCombineInstrSettingFlag(Opc: CombineOpc) && |
5948 | MI->findRegisterDefOperandIdx(Reg: AArch64::NZCV, /*TRI=*/nullptr, isDead: true) == -1) |
5949 | return false; |
5950 | |
5951 | return true; |
5952 | } |
5953 | |
5954 | // |
5955 | // Is \param MO defined by an integer multiply and can be combined? |
5956 | static bool canCombineWithMUL(MachineBasicBlock &MBB, MachineOperand &MO, |
5957 | unsigned MulOpc, unsigned ZeroReg) { |
5958 | return canCombine(MBB, MO, CombineOpc: MulOpc, ZeroReg, CheckZeroReg: true); |
5959 | } |
5960 | |
5961 | // |
5962 | // Is \param MO defined by a floating-point multiply and can be combined? |
5963 | static bool canCombineWithFMUL(MachineBasicBlock &MBB, MachineOperand &MO, |
5964 | unsigned MulOpc) { |
5965 | return canCombine(MBB, MO, CombineOpc: MulOpc); |
5966 | } |
5967 | |
5968 | // TODO: There are many more machine instruction opcodes to match: |
5969 | // 1. Other data types (integer, vectors) |
5970 | // 2. Other math / logic operations (xor, or) |
5971 | // 3. Other forms of the same operation (intrinsics and other variants) |
5972 | bool AArch64InstrInfo::isAssociativeAndCommutative(const MachineInstr &Inst, |
5973 | bool Invert) const { |
5974 | if (Invert) |
5975 | return false; |
5976 | switch (Inst.getOpcode()) { |
5977 | // == Floating-point types == |
5978 | // -- Floating-point instructions -- |
5979 | case AArch64::FADDHrr: |
5980 | case AArch64::FADDSrr: |
5981 | case AArch64::FADDDrr: |
5982 | case AArch64::FMULHrr: |
5983 | case AArch64::FMULSrr: |
5984 | case AArch64::FMULDrr: |
5985 | case AArch64::FMULX16: |
5986 | case AArch64::FMULX32: |
5987 | case AArch64::FMULX64: |
5988 | // -- Advanced SIMD instructions -- |
5989 | case AArch64::FADDv4f16: |
5990 | case AArch64::FADDv8f16: |
5991 | case AArch64::FADDv2f32: |
5992 | case AArch64::FADDv4f32: |
5993 | case AArch64::FADDv2f64: |
5994 | case AArch64::FMULv4f16: |
5995 | case AArch64::FMULv8f16: |
5996 | case AArch64::FMULv2f32: |
5997 | case AArch64::FMULv4f32: |
5998 | case AArch64::FMULv2f64: |
5999 | case AArch64::FMULXv4f16: |
6000 | case AArch64::FMULXv8f16: |
6001 | case AArch64::FMULXv2f32: |
6002 | case AArch64::FMULXv4f32: |
6003 | case AArch64::FMULXv2f64: |
6004 | // -- SVE instructions -- |
6005 | // Opcodes FMULX_ZZZ_? don't exist because there is no unpredicated FMULX |
6006 | // in the SVE instruction set (though there are predicated ones). |
6007 | case AArch64::FADD_ZZZ_H: |
6008 | case AArch64::FADD_ZZZ_S: |
6009 | case AArch64::FADD_ZZZ_D: |
6010 | case AArch64::FMUL_ZZZ_H: |
6011 | case AArch64::FMUL_ZZZ_S: |
6012 | case AArch64::FMUL_ZZZ_D: |
6013 | return Inst.getParent()->getParent()->getTarget().Options.UnsafeFPMath || |
6014 | (Inst.getFlag(Flag: MachineInstr::MIFlag::FmReassoc) && |
6015 | Inst.getFlag(Flag: MachineInstr::MIFlag::FmNsz)); |
6016 | |
6017 | // == Integer types == |
6018 | // -- Base instructions -- |
6019 | // Opcodes MULWrr and MULXrr don't exist because |
6020 | // `MUL <Wd>, <Wn>, <Wm>` and `MUL <Xd>, <Xn>, <Xm>` are aliases of |
6021 | // `MADD <Wd>, <Wn>, <Wm>, WZR` and `MADD <Xd>, <Xn>, <Xm>, XZR` respectively. |
6022 | // The machine-combiner does not support three-source-operands machine |
6023 | // instruction. So we cannot reassociate MULs. |
6024 | case AArch64::ADDWrr: |
6025 | case AArch64::ADDXrr: |
6026 | case AArch64::ANDWrr: |
6027 | case AArch64::ANDXrr: |
6028 | case AArch64::ORRWrr: |
6029 | case AArch64::ORRXrr: |
6030 | case AArch64::EORWrr: |
6031 | case AArch64::EORXrr: |
6032 | case AArch64::EONWrr: |
6033 | case AArch64::EONXrr: |
6034 | // -- Advanced SIMD instructions -- |
6035 | // Opcodes MULv1i64 and MULv2i64 don't exist because there is no 64-bit MUL |
6036 | // in the Advanced SIMD instruction set. |
6037 | case AArch64::ADDv8i8: |
6038 | case AArch64::ADDv16i8: |
6039 | case AArch64::ADDv4i16: |
6040 | case AArch64::ADDv8i16: |
6041 | case AArch64::ADDv2i32: |
6042 | case AArch64::ADDv4i32: |
6043 | case AArch64::ADDv1i64: |
6044 | case AArch64::ADDv2i64: |
6045 | case AArch64::MULv8i8: |
6046 | case AArch64::MULv16i8: |
6047 | case AArch64::MULv4i16: |
6048 | case AArch64::MULv8i16: |
6049 | case AArch64::MULv2i32: |
6050 | case AArch64::MULv4i32: |
6051 | case AArch64::ANDv8i8: |
6052 | case AArch64::ANDv16i8: |
6053 | case AArch64::ORRv8i8: |
6054 | case AArch64::ORRv16i8: |
6055 | case AArch64::EORv8i8: |
6056 | case AArch64::EORv16i8: |
6057 | // -- SVE instructions -- |
6058 | case AArch64::ADD_ZZZ_B: |
6059 | case AArch64::ADD_ZZZ_H: |
6060 | case AArch64::ADD_ZZZ_S: |
6061 | case AArch64::ADD_ZZZ_D: |
6062 | case AArch64::MUL_ZZZ_B: |
6063 | case AArch64::MUL_ZZZ_H: |
6064 | case AArch64::MUL_ZZZ_S: |
6065 | case AArch64::MUL_ZZZ_D: |
6066 | case AArch64::AND_ZZZ: |
6067 | case AArch64::ORR_ZZZ: |
6068 | case AArch64::EOR_ZZZ: |
6069 | return true; |
6070 | |
6071 | default: |
6072 | return false; |
6073 | } |
6074 | } |
6075 | |
6076 | /// Find instructions that can be turned into madd. |
6077 | static bool getMaddPatterns(MachineInstr &Root, |
6078 | SmallVectorImpl<unsigned> &Patterns) { |
6079 | unsigned Opc = Root.getOpcode(); |
6080 | MachineBasicBlock &MBB = *Root.getParent(); |
6081 | bool Found = false; |
6082 | |
6083 | if (!isCombineInstrCandidate(Opc)) |
6084 | return false; |
6085 | if (isCombineInstrSettingFlag(Opc)) { |
6086 | int Cmp_NZCV = |
6087 | Root.findRegisterDefOperandIdx(Reg: AArch64::NZCV, /*TRI=*/nullptr, isDead: true); |
6088 | // When NZCV is live bail out. |
6089 | if (Cmp_NZCV == -1) |
6090 | return false; |
6091 | unsigned NewOpc = convertToNonFlagSettingOpc(MI: Root); |
6092 | // When opcode can't change bail out. |
6093 | // CHECKME: do we miss any cases for opcode conversion? |
6094 | if (NewOpc == Opc) |
6095 | return false; |
6096 | Opc = NewOpc; |
6097 | } |
6098 | |
6099 | auto setFound = [&](int Opcode, int Operand, unsigned ZeroReg, |
6100 | unsigned Pattern) { |
6101 | if (canCombineWithMUL(MBB, MO&: Root.getOperand(i: Operand), MulOpc: Opcode, ZeroReg)) { |
6102 | Patterns.push_back(Elt: Pattern); |
6103 | Found = true; |
6104 | } |
6105 | }; |
6106 | |
6107 | auto setVFound = [&](int Opcode, int Operand, unsigned Pattern) { |
6108 | if (canCombine(MBB, MO&: Root.getOperand(i: Operand), CombineOpc: Opcode)) { |
6109 | Patterns.push_back(Elt: Pattern); |
6110 | Found = true; |
6111 | } |
6112 | }; |
6113 | |
6114 | typedef AArch64MachineCombinerPattern MCP; |
6115 | |
6116 | switch (Opc) { |
6117 | default: |
6118 | break; |
6119 | case AArch64::ADDWrr: |
6120 | assert(Root.getOperand(1).isReg() && Root.getOperand(2).isReg() && |
6121 | "ADDWrr does not have register operands" ); |
6122 | setFound(AArch64::MADDWrrr, 1, AArch64::WZR, MCP::MULADDW_OP1); |
6123 | setFound(AArch64::MADDWrrr, 2, AArch64::WZR, MCP::MULADDW_OP2); |
6124 | break; |
6125 | case AArch64::ADDXrr: |
6126 | setFound(AArch64::MADDXrrr, 1, AArch64::XZR, MCP::MULADDX_OP1); |
6127 | setFound(AArch64::MADDXrrr, 2, AArch64::XZR, MCP::MULADDX_OP2); |
6128 | break; |
6129 | case AArch64::SUBWrr: |
6130 | setFound(AArch64::MADDWrrr, 2, AArch64::WZR, MCP::MULSUBW_OP2); |
6131 | setFound(AArch64::MADDWrrr, 1, AArch64::WZR, MCP::MULSUBW_OP1); |
6132 | break; |
6133 | case AArch64::SUBXrr: |
6134 | setFound(AArch64::MADDXrrr, 2, AArch64::XZR, MCP::MULSUBX_OP2); |
6135 | setFound(AArch64::MADDXrrr, 1, AArch64::XZR, MCP::MULSUBX_OP1); |
6136 | break; |
6137 | case AArch64::ADDWri: |
6138 | setFound(AArch64::MADDWrrr, 1, AArch64::WZR, MCP::MULADDWI_OP1); |
6139 | break; |
6140 | case AArch64::ADDXri: |
6141 | setFound(AArch64::MADDXrrr, 1, AArch64::XZR, MCP::MULADDXI_OP1); |
6142 | break; |
6143 | case AArch64::SUBWri: |
6144 | setFound(AArch64::MADDWrrr, 1, AArch64::WZR, MCP::MULSUBWI_OP1); |
6145 | break; |
6146 | case AArch64::SUBXri: |
6147 | setFound(AArch64::MADDXrrr, 1, AArch64::XZR, MCP::MULSUBXI_OP1); |
6148 | break; |
6149 | case AArch64::ADDv8i8: |
6150 | setVFound(AArch64::MULv8i8, 1, MCP::MULADDv8i8_OP1); |
6151 | setVFound(AArch64::MULv8i8, 2, MCP::MULADDv8i8_OP2); |
6152 | break; |
6153 | case AArch64::ADDv16i8: |
6154 | setVFound(AArch64::MULv16i8, 1, MCP::MULADDv16i8_OP1); |
6155 | setVFound(AArch64::MULv16i8, 2, MCP::MULADDv16i8_OP2); |
6156 | break; |
6157 | case AArch64::ADDv4i16: |
6158 | setVFound(AArch64::MULv4i16, 1, MCP::MULADDv4i16_OP1); |
6159 | setVFound(AArch64::MULv4i16, 2, MCP::MULADDv4i16_OP2); |
6160 | setVFound(AArch64::MULv4i16_indexed, 1, MCP::MULADDv4i16_indexed_OP1); |
6161 | setVFound(AArch64::MULv4i16_indexed, 2, MCP::MULADDv4i16_indexed_OP2); |
6162 | break; |
6163 | case AArch64::ADDv8i16: |
6164 | setVFound(AArch64::MULv8i16, 1, MCP::MULADDv8i16_OP1); |
6165 | setVFound(AArch64::MULv8i16, 2, MCP::MULADDv8i16_OP2); |
6166 | setVFound(AArch64::MULv8i16_indexed, 1, MCP::MULADDv8i16_indexed_OP1); |
6167 | setVFound(AArch64::MULv8i16_indexed, 2, MCP::MULADDv8i16_indexed_OP2); |
6168 | break; |
6169 | case AArch64::ADDv2i32: |
6170 | setVFound(AArch64::MULv2i32, 1, MCP::MULADDv2i32_OP1); |
6171 | setVFound(AArch64::MULv2i32, 2, MCP::MULADDv2i32_OP2); |
6172 | setVFound(AArch64::MULv2i32_indexed, 1, MCP::MULADDv2i32_indexed_OP1); |
6173 | setVFound(AArch64::MULv2i32_indexed, 2, MCP::MULADDv2i32_indexed_OP2); |
6174 | break; |
6175 | case AArch64::ADDv4i32: |
6176 | setVFound(AArch64::MULv4i32, 1, MCP::MULADDv4i32_OP1); |
6177 | setVFound(AArch64::MULv4i32, 2, MCP::MULADDv4i32_OP2); |
6178 | setVFound(AArch64::MULv4i32_indexed, 1, MCP::MULADDv4i32_indexed_OP1); |
6179 | setVFound(AArch64::MULv4i32_indexed, 2, MCP::MULADDv4i32_indexed_OP2); |
6180 | break; |
6181 | case AArch64::SUBv8i8: |
6182 | setVFound(AArch64::MULv8i8, 1, MCP::MULSUBv8i8_OP1); |
6183 | setVFound(AArch64::MULv8i8, 2, MCP::MULSUBv8i8_OP2); |
6184 | break; |
6185 | case AArch64::SUBv16i8: |
6186 | setVFound(AArch64::MULv16i8, 1, MCP::MULSUBv16i8_OP1); |
6187 | setVFound(AArch64::MULv16i8, 2, MCP::MULSUBv16i8_OP2); |
6188 | break; |
6189 | case AArch64::SUBv4i16: |
6190 | setVFound(AArch64::MULv4i16, 1, MCP::MULSUBv4i16_OP1); |
6191 | setVFound(AArch64::MULv4i16, 2, MCP::MULSUBv4i16_OP2); |
6192 | setVFound(AArch64::MULv4i16_indexed, 1, MCP::MULSUBv4i16_indexed_OP1); |
6193 | setVFound(AArch64::MULv4i16_indexed, 2, MCP::MULSUBv4i16_indexed_OP2); |
6194 | break; |
6195 | case AArch64::SUBv8i16: |
6196 | setVFound(AArch64::MULv8i16, 1, MCP::MULSUBv8i16_OP1); |
6197 | setVFound(AArch64::MULv8i16, 2, MCP::MULSUBv8i16_OP2); |
6198 | setVFound(AArch64::MULv8i16_indexed, 1, MCP::MULSUBv8i16_indexed_OP1); |
6199 | setVFound(AArch64::MULv8i16_indexed, 2, MCP::MULSUBv8i16_indexed_OP2); |
6200 | break; |
6201 | case AArch64::SUBv2i32: |
6202 | setVFound(AArch64::MULv2i32, 1, MCP::MULSUBv2i32_OP1); |
6203 | setVFound(AArch64::MULv2i32, 2, MCP::MULSUBv2i32_OP2); |
6204 | setVFound(AArch64::MULv2i32_indexed, 1, MCP::MULSUBv2i32_indexed_OP1); |
6205 | setVFound(AArch64::MULv2i32_indexed, 2, MCP::MULSUBv2i32_indexed_OP2); |
6206 | break; |
6207 | case AArch64::SUBv4i32: |
6208 | setVFound(AArch64::MULv4i32, 1, MCP::MULSUBv4i32_OP1); |
6209 | setVFound(AArch64::MULv4i32, 2, MCP::MULSUBv4i32_OP2); |
6210 | setVFound(AArch64::MULv4i32_indexed, 1, MCP::MULSUBv4i32_indexed_OP1); |
6211 | setVFound(AArch64::MULv4i32_indexed, 2, MCP::MULSUBv4i32_indexed_OP2); |
6212 | break; |
6213 | } |
6214 | return Found; |
6215 | } |
6216 | /// Floating-Point Support |
6217 | |
6218 | /// Find instructions that can be turned into madd. |
6219 | static bool getFMAPatterns(MachineInstr &Root, |
6220 | SmallVectorImpl<unsigned> &Patterns) { |
6221 | |
6222 | if (!isCombineInstrCandidateFP(Inst: Root)) |
6223 | return false; |
6224 | |
6225 | MachineBasicBlock &MBB = *Root.getParent(); |
6226 | bool Found = false; |
6227 | |
6228 | auto Match = [&](int Opcode, int Operand, unsigned Pattern) -> bool { |
6229 | if (canCombineWithFMUL(MBB, MO&: Root.getOperand(i: Operand), MulOpc: Opcode)) { |
6230 | Patterns.push_back(Elt: Pattern); |
6231 | return true; |
6232 | } |
6233 | return false; |
6234 | }; |
6235 | |
6236 | typedef AArch64MachineCombinerPattern MCP; |
6237 | |
6238 | switch (Root.getOpcode()) { |
6239 | default: |
6240 | assert(false && "Unsupported FP instruction in combiner\n" ); |
6241 | break; |
6242 | case AArch64::FADDHrr: |
6243 | assert(Root.getOperand(1).isReg() && Root.getOperand(2).isReg() && |
6244 | "FADDHrr does not have register operands" ); |
6245 | |
6246 | Found = Match(AArch64::FMULHrr, 1, MCP::FMULADDH_OP1); |
6247 | Found |= Match(AArch64::FMULHrr, 2, MCP::FMULADDH_OP2); |
6248 | break; |
6249 | case AArch64::FADDSrr: |
6250 | assert(Root.getOperand(1).isReg() && Root.getOperand(2).isReg() && |
6251 | "FADDSrr does not have register operands" ); |
6252 | |
6253 | Found |= Match(AArch64::FMULSrr, 1, MCP::FMULADDS_OP1) || |
6254 | Match(AArch64::FMULv1i32_indexed, 1, MCP::FMLAv1i32_indexed_OP1); |
6255 | |
6256 | Found |= Match(AArch64::FMULSrr, 2, MCP::FMULADDS_OP2) || |
6257 | Match(AArch64::FMULv1i32_indexed, 2, MCP::FMLAv1i32_indexed_OP2); |
6258 | break; |
6259 | case AArch64::FADDDrr: |
6260 | Found |= Match(AArch64::FMULDrr, 1, MCP::FMULADDD_OP1) || |
6261 | Match(AArch64::FMULv1i64_indexed, 1, MCP::FMLAv1i64_indexed_OP1); |
6262 | |
6263 | Found |= Match(AArch64::FMULDrr, 2, MCP::FMULADDD_OP2) || |
6264 | Match(AArch64::FMULv1i64_indexed, 2, MCP::FMLAv1i64_indexed_OP2); |
6265 | break; |
6266 | case AArch64::FADDv4f16: |
6267 | Found |= Match(AArch64::FMULv4i16_indexed, 1, MCP::FMLAv4i16_indexed_OP1) || |
6268 | Match(AArch64::FMULv4f16, 1, MCP::FMLAv4f16_OP1); |
6269 | |
6270 | Found |= Match(AArch64::FMULv4i16_indexed, 2, MCP::FMLAv4i16_indexed_OP2) || |
6271 | Match(AArch64::FMULv4f16, 2, MCP::FMLAv4f16_OP2); |
6272 | break; |
6273 | case AArch64::FADDv8f16: |
6274 | Found |= Match(AArch64::FMULv8i16_indexed, 1, MCP::FMLAv8i16_indexed_OP1) || |
6275 | Match(AArch64::FMULv8f16, 1, MCP::FMLAv8f16_OP1); |
6276 | |
6277 | Found |= Match(AArch64::FMULv8i16_indexed, 2, MCP::FMLAv8i16_indexed_OP2) || |
6278 | Match(AArch64::FMULv8f16, 2, MCP::FMLAv8f16_OP2); |
6279 | break; |
6280 | case AArch64::FADDv2f32: |
6281 | Found |= Match(AArch64::FMULv2i32_indexed, 1, MCP::FMLAv2i32_indexed_OP1) || |
6282 | Match(AArch64::FMULv2f32, 1, MCP::FMLAv2f32_OP1); |
6283 | |
6284 | Found |= Match(AArch64::FMULv2i32_indexed, 2, MCP::FMLAv2i32_indexed_OP2) || |
6285 | Match(AArch64::FMULv2f32, 2, MCP::FMLAv2f32_OP2); |
6286 | break; |
6287 | case AArch64::FADDv2f64: |
6288 | Found |= Match(AArch64::FMULv2i64_indexed, 1, MCP::FMLAv2i64_indexed_OP1) || |
6289 | Match(AArch64::FMULv2f64, 1, MCP::FMLAv2f64_OP1); |
6290 | |
6291 | Found |= Match(AArch64::FMULv2i64_indexed, 2, MCP::FMLAv2i64_indexed_OP2) || |
6292 | Match(AArch64::FMULv2f64, 2, MCP::FMLAv2f64_OP2); |
6293 | break; |
6294 | case AArch64::FADDv4f32: |
6295 | Found |= Match(AArch64::FMULv4i32_indexed, 1, MCP::FMLAv4i32_indexed_OP1) || |
6296 | Match(AArch64::FMULv4f32, 1, MCP::FMLAv4f32_OP1); |
6297 | |
6298 | Found |= Match(AArch64::FMULv4i32_indexed, 2, MCP::FMLAv4i32_indexed_OP2) || |
6299 | Match(AArch64::FMULv4f32, 2, MCP::FMLAv4f32_OP2); |
6300 | break; |
6301 | case AArch64::FSUBHrr: |
6302 | Found = Match(AArch64::FMULHrr, 1, MCP::FMULSUBH_OP1); |
6303 | Found |= Match(AArch64::FMULHrr, 2, MCP::FMULSUBH_OP2); |
6304 | Found |= Match(AArch64::FNMULHrr, 1, MCP::FNMULSUBH_OP1); |
6305 | break; |
6306 | case AArch64::FSUBSrr: |
6307 | Found = Match(AArch64::FMULSrr, 1, MCP::FMULSUBS_OP1); |
6308 | |
6309 | Found |= Match(AArch64::FMULSrr, 2, MCP::FMULSUBS_OP2) || |
6310 | Match(AArch64::FMULv1i32_indexed, 2, MCP::FMLSv1i32_indexed_OP2); |
6311 | |
6312 | Found |= Match(AArch64::FNMULSrr, 1, MCP::FNMULSUBS_OP1); |
6313 | break; |
6314 | case AArch64::FSUBDrr: |
6315 | Found = Match(AArch64::FMULDrr, 1, MCP::FMULSUBD_OP1); |
6316 | |
6317 | Found |= Match(AArch64::FMULDrr, 2, MCP::FMULSUBD_OP2) || |
6318 | Match(AArch64::FMULv1i64_indexed, 2, MCP::FMLSv1i64_indexed_OP2); |
6319 | |
6320 | Found |= Match(AArch64::FNMULDrr, 1, MCP::FNMULSUBD_OP1); |
6321 | break; |
6322 | case AArch64::FSUBv4f16: |
6323 | Found |= Match(AArch64::FMULv4i16_indexed, 2, MCP::FMLSv4i16_indexed_OP2) || |
6324 | Match(AArch64::FMULv4f16, 2, MCP::FMLSv4f16_OP2); |
6325 | |
6326 | Found |= Match(AArch64::FMULv4i16_indexed, 1, MCP::FMLSv4i16_indexed_OP1) || |
6327 | Match(AArch64::FMULv4f16, 1, MCP::FMLSv4f16_OP1); |
6328 | break; |
6329 | case AArch64::FSUBv8f16: |
6330 | Found |= Match(AArch64::FMULv8i16_indexed, 2, MCP::FMLSv8i16_indexed_OP2) || |
6331 | Match(AArch64::FMULv8f16, 2, MCP::FMLSv8f16_OP2); |
6332 | |
6333 | Found |= Match(AArch64::FMULv8i16_indexed, 1, MCP::FMLSv8i16_indexed_OP1) || |
6334 | Match(AArch64::FMULv8f16, 1, MCP::FMLSv8f16_OP1); |
6335 | break; |
6336 | case AArch64::FSUBv2f32: |
6337 | Found |= Match(AArch64::FMULv2i32_indexed, 2, MCP::FMLSv2i32_indexed_OP2) || |
6338 | Match(AArch64::FMULv2f32, 2, MCP::FMLSv2f32_OP2); |
6339 | |
6340 | Found |= Match(AArch64::FMULv2i32_indexed, 1, MCP::FMLSv2i32_indexed_OP1) || |
6341 | Match(AArch64::FMULv2f32, 1, MCP::FMLSv2f32_OP1); |
6342 | break; |
6343 | case AArch64::FSUBv2f64: |
6344 | Found |= Match(AArch64::FMULv2i64_indexed, 2, MCP::FMLSv2i64_indexed_OP2) || |
6345 | Match(AArch64::FMULv2f64, 2, MCP::FMLSv2f64_OP2); |
6346 | |
6347 | Found |= Match(AArch64::FMULv2i64_indexed, 1, MCP::FMLSv2i64_indexed_OP1) || |
6348 | Match(AArch64::FMULv2f64, 1, MCP::FMLSv2f64_OP1); |
6349 | break; |
6350 | case AArch64::FSUBv4f32: |
6351 | Found |= Match(AArch64::FMULv4i32_indexed, 2, MCP::FMLSv4i32_indexed_OP2) || |
6352 | Match(AArch64::FMULv4f32, 2, MCP::FMLSv4f32_OP2); |
6353 | |
6354 | Found |= Match(AArch64::FMULv4i32_indexed, 1, MCP::FMLSv4i32_indexed_OP1) || |
6355 | Match(AArch64::FMULv4f32, 1, MCP::FMLSv4f32_OP1); |
6356 | break; |
6357 | } |
6358 | return Found; |
6359 | } |
6360 | |
6361 | static bool getFMULPatterns(MachineInstr &Root, |
6362 | SmallVectorImpl<unsigned> &Patterns) { |
6363 | MachineBasicBlock &MBB = *Root.getParent(); |
6364 | bool Found = false; |
6365 | |
6366 | auto Match = [&](unsigned Opcode, int Operand, unsigned Pattern) -> bool { |
6367 | MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo(); |
6368 | MachineOperand &MO = Root.getOperand(i: Operand); |
6369 | MachineInstr *MI = nullptr; |
6370 | if (MO.isReg() && MO.getReg().isVirtual()) |
6371 | MI = MRI.getUniqueVRegDef(Reg: MO.getReg()); |
6372 | // Ignore No-op COPYs in FMUL(COPY(DUP(..))) |
6373 | if (MI && MI->getOpcode() == TargetOpcode::COPY && |
6374 | MI->getOperand(i: 1).getReg().isVirtual()) |
6375 | MI = MRI.getUniqueVRegDef(Reg: MI->getOperand(i: 1).getReg()); |
6376 | if (MI && MI->getOpcode() == Opcode) { |
6377 | Patterns.push_back(Elt: Pattern); |
6378 | return true; |
6379 | } |
6380 | return false; |
6381 | }; |
6382 | |
6383 | typedef AArch64MachineCombinerPattern MCP; |
6384 | |
6385 | switch (Root.getOpcode()) { |
6386 | default: |
6387 | return false; |
6388 | case AArch64::FMULv2f32: |
6389 | Found = Match(AArch64::DUPv2i32lane, 1, MCP::FMULv2i32_indexed_OP1); |
6390 | Found |= Match(AArch64::DUPv2i32lane, 2, MCP::FMULv2i32_indexed_OP2); |
6391 | break; |
6392 | case AArch64::FMULv2f64: |
6393 | Found = Match(AArch64::DUPv2i64lane, 1, MCP::FMULv2i64_indexed_OP1); |
6394 | Found |= Match(AArch64::DUPv2i64lane, 2, MCP::FMULv2i64_indexed_OP2); |
6395 | break; |
6396 | case AArch64::FMULv4f16: |
6397 | Found = Match(AArch64::DUPv4i16lane, 1, MCP::FMULv4i16_indexed_OP1); |
6398 | Found |= Match(AArch64::DUPv4i16lane, 2, MCP::FMULv4i16_indexed_OP2); |
6399 | break; |
6400 | case AArch64::FMULv4f32: |
6401 | Found = Match(AArch64::DUPv4i32lane, 1, MCP::FMULv4i32_indexed_OP1); |
6402 | Found |= Match(AArch64::DUPv4i32lane, 2, MCP::FMULv4i32_indexed_OP2); |
6403 | break; |
6404 | case AArch64::FMULv8f16: |
6405 | Found = Match(AArch64::DUPv8i16lane, 1, MCP::FMULv8i16_indexed_OP1); |
6406 | Found |= Match(AArch64::DUPv8i16lane, 2, MCP::FMULv8i16_indexed_OP2); |
6407 | break; |
6408 | } |
6409 | |
6410 | return Found; |
6411 | } |
6412 | |
6413 | static bool getFNEGPatterns(MachineInstr &Root, |
6414 | SmallVectorImpl<unsigned> &Patterns) { |
6415 | unsigned Opc = Root.getOpcode(); |
6416 | MachineBasicBlock &MBB = *Root.getParent(); |
6417 | MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo(); |
6418 | |
6419 | auto Match = [&](unsigned Opcode, unsigned Pattern) -> bool { |
6420 | MachineOperand &MO = Root.getOperand(i: 1); |
6421 | MachineInstr *MI = MRI.getUniqueVRegDef(Reg: MO.getReg()); |
6422 | if (MI != nullptr && (MI->getOpcode() == Opcode) && |
6423 | MRI.hasOneNonDBGUse(RegNo: MI->getOperand(i: 0).getReg()) && |
6424 | Root.getFlag(Flag: MachineInstr::MIFlag::FmContract) && |
6425 | Root.getFlag(Flag: MachineInstr::MIFlag::FmNsz) && |
6426 | MI->getFlag(Flag: MachineInstr::MIFlag::FmContract) && |
6427 | MI->getFlag(Flag: MachineInstr::MIFlag::FmNsz)) { |
6428 | Patterns.push_back(Elt: Pattern); |
6429 | return true; |
6430 | } |
6431 | return false; |
6432 | }; |
6433 | |
6434 | switch (Opc) { |
6435 | default: |
6436 | break; |
6437 | case AArch64::FNEGDr: |
6438 | return Match(AArch64::FMADDDrrr, AArch64MachineCombinerPattern::FNMADD); |
6439 | case AArch64::FNEGSr: |
6440 | return Match(AArch64::FMADDSrrr, AArch64MachineCombinerPattern::FNMADD); |
6441 | } |
6442 | |
6443 | return false; |
6444 | } |
6445 | |
6446 | /// Return true when a code sequence can improve throughput. It |
6447 | /// should be called only for instructions in loops. |
6448 | /// \param Pattern - combiner pattern |
6449 | bool AArch64InstrInfo::isThroughputPattern(unsigned Pattern) const { |
6450 | switch (Pattern) { |
6451 | default: |
6452 | break; |
6453 | case AArch64MachineCombinerPattern::FMULADDH_OP1: |
6454 | case AArch64MachineCombinerPattern::FMULADDH_OP2: |
6455 | case AArch64MachineCombinerPattern::FMULSUBH_OP1: |
6456 | case AArch64MachineCombinerPattern::FMULSUBH_OP2: |
6457 | case AArch64MachineCombinerPattern::FMULADDS_OP1: |
6458 | case AArch64MachineCombinerPattern::FMULADDS_OP2: |
6459 | case AArch64MachineCombinerPattern::FMULSUBS_OP1: |
6460 | case AArch64MachineCombinerPattern::FMULSUBS_OP2: |
6461 | case AArch64MachineCombinerPattern::FMULADDD_OP1: |
6462 | case AArch64MachineCombinerPattern::FMULADDD_OP2: |
6463 | case AArch64MachineCombinerPattern::FMULSUBD_OP1: |
6464 | case AArch64MachineCombinerPattern::FMULSUBD_OP2: |
6465 | case AArch64MachineCombinerPattern::FNMULSUBH_OP1: |
6466 | case AArch64MachineCombinerPattern::FNMULSUBS_OP1: |
6467 | case AArch64MachineCombinerPattern::FNMULSUBD_OP1: |
6468 | case AArch64MachineCombinerPattern::FMLAv4i16_indexed_OP1: |
6469 | case AArch64MachineCombinerPattern::FMLAv4i16_indexed_OP2: |
6470 | case AArch64MachineCombinerPattern::FMLAv8i16_indexed_OP1: |
6471 | case AArch64MachineCombinerPattern::FMLAv8i16_indexed_OP2: |
6472 | case AArch64MachineCombinerPattern::FMLAv1i32_indexed_OP1: |
6473 | case AArch64MachineCombinerPattern::FMLAv1i32_indexed_OP2: |
6474 | case AArch64MachineCombinerPattern::FMLAv1i64_indexed_OP1: |
6475 | case AArch64MachineCombinerPattern::FMLAv1i64_indexed_OP2: |
6476 | case AArch64MachineCombinerPattern::FMLAv4f16_OP2: |
6477 | case AArch64MachineCombinerPattern::FMLAv4f16_OP1: |
6478 | case AArch64MachineCombinerPattern::FMLAv8f16_OP1: |
6479 | case AArch64MachineCombinerPattern::FMLAv8f16_OP2: |
6480 | case AArch64MachineCombinerPattern::FMLAv2f32_OP2: |
6481 | case AArch64MachineCombinerPattern::FMLAv2f32_OP1: |
6482 | case AArch64MachineCombinerPattern::FMLAv2f64_OP1: |
6483 | case AArch64MachineCombinerPattern::FMLAv2f64_OP2: |
6484 | case AArch64MachineCombinerPattern::FMLAv2i32_indexed_OP1: |
6485 | case AArch64MachineCombinerPattern::FMLAv2i32_indexed_OP2: |
6486 | case AArch64MachineCombinerPattern::FMLAv2i64_indexed_OP1: |
6487 | case AArch64MachineCombinerPattern::FMLAv2i64_indexed_OP2: |
6488 | case AArch64MachineCombinerPattern::FMLAv4f32_OP1: |
6489 | case AArch64MachineCombinerPattern::FMLAv4f32_OP2: |
6490 | case AArch64MachineCombinerPattern::FMLAv4i32_indexed_OP1: |
6491 | case AArch64MachineCombinerPattern::FMLAv4i32_indexed_OP2: |
6492 | case AArch64MachineCombinerPattern::FMLSv4i16_indexed_OP1: |
6493 | case AArch64MachineCombinerPattern::FMLSv4i16_indexed_OP2: |
6494 | case AArch64MachineCombinerPattern::FMLSv8i16_indexed_OP1: |
6495 | case AArch64MachineCombinerPattern::FMLSv8i16_indexed_OP2: |
6496 | case AArch64MachineCombinerPattern::FMLSv1i32_indexed_OP2: |
6497 | case AArch64MachineCombinerPattern::FMLSv1i64_indexed_OP2: |
6498 | case AArch64MachineCombinerPattern::FMLSv2i32_indexed_OP2: |
6499 | case AArch64MachineCombinerPattern::FMLSv2i64_indexed_OP2: |
6500 | case AArch64MachineCombinerPattern::FMLSv4f16_OP1: |
6501 | case AArch64MachineCombinerPattern::FMLSv4f16_OP2: |
6502 | case AArch64MachineCombinerPattern::FMLSv8f16_OP1: |
6503 | case AArch64MachineCombinerPattern::FMLSv8f16_OP2: |
6504 | case AArch64MachineCombinerPattern::FMLSv2f32_OP2: |
6505 | case AArch64MachineCombinerPattern::FMLSv2f64_OP2: |
6506 | case AArch64MachineCombinerPattern::FMLSv4i32_indexed_OP2: |
6507 | case AArch64MachineCombinerPattern::FMLSv4f32_OP2: |
6508 | case AArch64MachineCombinerPattern::FMULv2i32_indexed_OP1: |
6509 | case AArch64MachineCombinerPattern::FMULv2i32_indexed_OP2: |
6510 | case AArch64MachineCombinerPattern::FMULv2i64_indexed_OP1: |
6511 | case AArch64MachineCombinerPattern::FMULv2i64_indexed_OP2: |
6512 | case AArch64MachineCombinerPattern::FMULv4i16_indexed_OP1: |
6513 | case AArch64MachineCombinerPattern::FMULv4i16_indexed_OP2: |
6514 | case AArch64MachineCombinerPattern::FMULv4i32_indexed_OP1: |
6515 | case AArch64MachineCombinerPattern::FMULv4i32_indexed_OP2: |
6516 | case AArch64MachineCombinerPattern::FMULv8i16_indexed_OP1: |
6517 | case AArch64MachineCombinerPattern::FMULv8i16_indexed_OP2: |
6518 | case AArch64MachineCombinerPattern::MULADDv8i8_OP1: |
6519 | case AArch64MachineCombinerPattern::MULADDv8i8_OP2: |
6520 | case AArch64MachineCombinerPattern::MULADDv16i8_OP1: |
6521 | case AArch64MachineCombinerPattern::MULADDv16i8_OP2: |
6522 | case AArch64MachineCombinerPattern::MULADDv4i16_OP1: |
6523 | case AArch64MachineCombinerPattern::MULADDv4i16_OP2: |
6524 | case AArch64MachineCombinerPattern::MULADDv8i16_OP1: |
6525 | case AArch64MachineCombinerPattern::MULADDv8i16_OP2: |
6526 | case AArch64MachineCombinerPattern::MULADDv2i32_OP1: |
6527 | case AArch64MachineCombinerPattern::MULADDv2i32_OP2: |
6528 | case AArch64MachineCombinerPattern::MULADDv4i32_OP1: |
6529 | case AArch64MachineCombinerPattern::MULADDv4i32_OP2: |
6530 | case AArch64MachineCombinerPattern::MULSUBv8i8_OP1: |
6531 | case AArch64MachineCombinerPattern::MULSUBv8i8_OP2: |
6532 | case AArch64MachineCombinerPattern::MULSUBv16i8_OP1: |
6533 | case AArch64MachineCombinerPattern::MULSUBv16i8_OP2: |
6534 | case AArch64MachineCombinerPattern::MULSUBv4i16_OP1: |
6535 | case AArch64MachineCombinerPattern::MULSUBv4i16_OP2: |
6536 | case AArch64MachineCombinerPattern::MULSUBv8i16_OP1: |
6537 | case AArch64MachineCombinerPattern::MULSUBv8i16_OP2: |
6538 | case AArch64MachineCombinerPattern::MULSUBv2i32_OP1: |
6539 | case AArch64MachineCombinerPattern::MULSUBv2i32_OP2: |
6540 | case AArch64MachineCombinerPattern::MULSUBv4i32_OP1: |
6541 | case AArch64MachineCombinerPattern::MULSUBv4i32_OP2: |
6542 | case AArch64MachineCombinerPattern::MULADDv4i16_indexed_OP1: |
6543 | case AArch64MachineCombinerPattern::MULADDv4i16_indexed_OP2: |
6544 | case AArch64MachineCombinerPattern::MULADDv8i16_indexed_OP1: |
6545 | case AArch64MachineCombinerPattern::MULADDv8i16_indexed_OP2: |
6546 | case AArch64MachineCombinerPattern::MULADDv2i32_indexed_OP1: |
6547 | case AArch64MachineCombinerPattern::MULADDv2i32_indexed_OP2: |
6548 | case AArch64MachineCombinerPattern::MULADDv4i32_indexed_OP1: |
6549 | case AArch64MachineCombinerPattern::MULADDv4i32_indexed_OP2: |
6550 | case AArch64MachineCombinerPattern::MULSUBv4i16_indexed_OP1: |
6551 | case AArch64MachineCombinerPattern::MULSUBv4i16_indexed_OP2: |
6552 | case AArch64MachineCombinerPattern::MULSUBv8i16_indexed_OP1: |
6553 | case AArch64MachineCombinerPattern::MULSUBv8i16_indexed_OP2: |
6554 | case AArch64MachineCombinerPattern::MULSUBv2i32_indexed_OP1: |
6555 | case AArch64MachineCombinerPattern::MULSUBv2i32_indexed_OP2: |
6556 | case AArch64MachineCombinerPattern::MULSUBv4i32_indexed_OP1: |
6557 | case AArch64MachineCombinerPattern::MULSUBv4i32_indexed_OP2: |
6558 | return true; |
6559 | } // end switch (Pattern) |
6560 | return false; |
6561 | } |
6562 | |
6563 | /// Find other MI combine patterns. |
6564 | static bool getMiscPatterns(MachineInstr &Root, |
6565 | SmallVectorImpl<unsigned> &Patterns) { |
6566 | // A - (B + C) ==> (A - B) - C or (A - C) - B |
6567 | unsigned Opc = Root.getOpcode(); |
6568 | MachineBasicBlock &MBB = *Root.getParent(); |
6569 | |
6570 | switch (Opc) { |
6571 | case AArch64::SUBWrr: |
6572 | case AArch64::SUBSWrr: |
6573 | case AArch64::SUBXrr: |
6574 | case AArch64::SUBSXrr: |
6575 | // Found candidate root. |
6576 | break; |
6577 | default: |
6578 | return false; |
6579 | } |
6580 | |
6581 | if (isCombineInstrSettingFlag(Opc) && |
6582 | Root.findRegisterDefOperandIdx(Reg: AArch64::NZCV, /*TRI=*/nullptr, isDead: true) == |
6583 | -1) |
6584 | return false; |
6585 | |
6586 | if (canCombine(MBB, MO&: Root.getOperand(i: 2), CombineOpc: AArch64::ADDWrr) || |
6587 | canCombine(MBB, MO&: Root.getOperand(i: 2), CombineOpc: AArch64::ADDSWrr) || |
6588 | canCombine(MBB, MO&: Root.getOperand(i: 2), CombineOpc: AArch64::ADDXrr) || |
6589 | canCombine(MBB, MO&: Root.getOperand(i: 2), CombineOpc: AArch64::ADDSXrr)) { |
6590 | Patterns.push_back(Elt: AArch64MachineCombinerPattern::SUBADD_OP1); |
6591 | Patterns.push_back(Elt: AArch64MachineCombinerPattern::SUBADD_OP2); |
6592 | return true; |
6593 | } |
6594 | |
6595 | return false; |
6596 | } |
6597 | |
6598 | CombinerObjective |
6599 | AArch64InstrInfo::getCombinerObjective(unsigned Pattern) const { |
6600 | switch (Pattern) { |
6601 | case AArch64MachineCombinerPattern::SUBADD_OP1: |
6602 | case AArch64MachineCombinerPattern::SUBADD_OP2: |
6603 | return CombinerObjective::MustReduceDepth; |
6604 | default: |
6605 | return TargetInstrInfo::getCombinerObjective(Pattern); |
6606 | } |
6607 | } |
6608 | |
6609 | /// Return true when there is potentially a faster code sequence for an |
6610 | /// instruction chain ending in \p Root. All potential patterns are listed in |
6611 | /// the \p Pattern vector. Pattern should be sorted in priority order since the |
6612 | /// pattern evaluator stops checking as soon as it finds a faster sequence. |
6613 | |
6614 | bool AArch64InstrInfo::getMachineCombinerPatterns( |
6615 | MachineInstr &Root, SmallVectorImpl<unsigned> &Patterns, |
6616 | bool DoRegPressureReduce) const { |
6617 | // Integer patterns |
6618 | if (getMaddPatterns(Root, Patterns)) |
6619 | return true; |
6620 | // Floating point patterns |
6621 | if (getFMULPatterns(Root, Patterns)) |
6622 | return true; |
6623 | if (getFMAPatterns(Root, Patterns)) |
6624 | return true; |
6625 | if (getFNEGPatterns(Root, Patterns)) |
6626 | return true; |
6627 | |
6628 | // Other patterns |
6629 | if (getMiscPatterns(Root, Patterns)) |
6630 | return true; |
6631 | |
6632 | return TargetInstrInfo::getMachineCombinerPatterns(Root, Patterns, |
6633 | DoRegPressureReduce); |
6634 | } |
6635 | |
6636 | enum class FMAInstKind { Default, Indexed, Accumulator }; |
6637 | /// genFusedMultiply - Generate fused multiply instructions. |
6638 | /// This function supports both integer and floating point instructions. |
6639 | /// A typical example: |
6640 | /// F|MUL I=A,B,0 |
6641 | /// F|ADD R,I,C |
6642 | /// ==> F|MADD R,A,B,C |
6643 | /// \param MF Containing MachineFunction |
6644 | /// \param MRI Register information |
6645 | /// \param TII Target information |
6646 | /// \param Root is the F|ADD instruction |
6647 | /// \param [out] InsInstrs is a vector of machine instructions and will |
6648 | /// contain the generated madd instruction |
6649 | /// \param IdxMulOpd is index of operand in Root that is the result of |
6650 | /// the F|MUL. In the example above IdxMulOpd is 1. |
6651 | /// \param MaddOpc the opcode fo the f|madd instruction |
6652 | /// \param RC Register class of operands |
6653 | /// \param kind of fma instruction (addressing mode) to be generated |
6654 | /// \param ReplacedAddend is the result register from the instruction |
6655 | /// replacing the non-combined operand, if any. |
6656 | static MachineInstr * |
6657 | genFusedMultiply(MachineFunction &MF, MachineRegisterInfo &MRI, |
6658 | const TargetInstrInfo *TII, MachineInstr &Root, |
6659 | SmallVectorImpl<MachineInstr *> &InsInstrs, unsigned IdxMulOpd, |
6660 | unsigned MaddOpc, const TargetRegisterClass *RC, |
6661 | FMAInstKind kind = FMAInstKind::Default, |
6662 | const Register *ReplacedAddend = nullptr) { |
6663 | assert(IdxMulOpd == 1 || IdxMulOpd == 2); |
6664 | |
6665 | unsigned IdxOtherOpd = IdxMulOpd == 1 ? 2 : 1; |
6666 | MachineInstr *MUL = MRI.getUniqueVRegDef(Reg: Root.getOperand(i: IdxMulOpd).getReg()); |
6667 | Register ResultReg = Root.getOperand(i: 0).getReg(); |
6668 | Register SrcReg0 = MUL->getOperand(i: 1).getReg(); |
6669 | bool Src0IsKill = MUL->getOperand(i: 1).isKill(); |
6670 | Register SrcReg1 = MUL->getOperand(i: 2).getReg(); |
6671 | bool Src1IsKill = MUL->getOperand(i: 2).isKill(); |
6672 | |
6673 | Register SrcReg2; |
6674 | bool Src2IsKill; |
6675 | if (ReplacedAddend) { |
6676 | // If we just generated a new addend, we must be it's only use. |
6677 | SrcReg2 = *ReplacedAddend; |
6678 | Src2IsKill = true; |
6679 | } else { |
6680 | SrcReg2 = Root.getOperand(i: IdxOtherOpd).getReg(); |
6681 | Src2IsKill = Root.getOperand(i: IdxOtherOpd).isKill(); |
6682 | } |
6683 | |
6684 | if (ResultReg.isVirtual()) |
6685 | MRI.constrainRegClass(Reg: ResultReg, RC); |
6686 | if (SrcReg0.isVirtual()) |
6687 | MRI.constrainRegClass(Reg: SrcReg0, RC); |
6688 | if (SrcReg1.isVirtual()) |
6689 | MRI.constrainRegClass(Reg: SrcReg1, RC); |
6690 | if (SrcReg2.isVirtual()) |
6691 | MRI.constrainRegClass(Reg: SrcReg2, RC); |
6692 | |
6693 | MachineInstrBuilder MIB; |
6694 | if (kind == FMAInstKind::Default) |
6695 | MIB = BuildMI(MF, MIMD: MIMetadata(Root), MCID: TII->get(Opcode: MaddOpc), DestReg: ResultReg) |
6696 | .addReg(RegNo: SrcReg0, flags: getKillRegState(B: Src0IsKill)) |
6697 | .addReg(RegNo: SrcReg1, flags: getKillRegState(B: Src1IsKill)) |
6698 | .addReg(RegNo: SrcReg2, flags: getKillRegState(B: Src2IsKill)); |
6699 | else if (kind == FMAInstKind::Indexed) |
6700 | MIB = BuildMI(MF, MIMD: MIMetadata(Root), MCID: TII->get(Opcode: MaddOpc), DestReg: ResultReg) |
6701 | .addReg(RegNo: SrcReg2, flags: getKillRegState(B: Src2IsKill)) |
6702 | .addReg(RegNo: SrcReg0, flags: getKillRegState(B: Src0IsKill)) |
6703 | .addReg(RegNo: SrcReg1, flags: getKillRegState(B: Src1IsKill)) |
6704 | .addImm(Val: MUL->getOperand(i: 3).getImm()); |
6705 | else if (kind == FMAInstKind::Accumulator) |
6706 | MIB = BuildMI(MF, MIMD: MIMetadata(Root), MCID: TII->get(Opcode: MaddOpc), DestReg: ResultReg) |
6707 | .addReg(RegNo: SrcReg2, flags: getKillRegState(B: Src2IsKill)) |
6708 | .addReg(RegNo: SrcReg0, flags: getKillRegState(B: Src0IsKill)) |
6709 | .addReg(RegNo: SrcReg1, flags: getKillRegState(B: Src1IsKill)); |
6710 | else |
6711 | assert(false && "Invalid FMA instruction kind \n" ); |
6712 | // Insert the MADD (MADD, FMA, FMS, FMLA, FMSL) |
6713 | InsInstrs.push_back(Elt: MIB); |
6714 | return MUL; |
6715 | } |
6716 | |
6717 | static MachineInstr * |
6718 | genFNegatedMAD(MachineFunction &MF, MachineRegisterInfo &MRI, |
6719 | const TargetInstrInfo *TII, MachineInstr &Root, |
6720 | SmallVectorImpl<MachineInstr *> &InsInstrs) { |
6721 | MachineInstr *MAD = MRI.getUniqueVRegDef(Reg: Root.getOperand(i: 1).getReg()); |
6722 | |
6723 | unsigned Opc = 0; |
6724 | const TargetRegisterClass *RC = MRI.getRegClass(Reg: MAD->getOperand(i: 0).getReg()); |
6725 | if (AArch64::FPR32RegClass.hasSubClassEq(RC)) |
6726 | Opc = AArch64::FNMADDSrrr; |
6727 | else if (AArch64::FPR64RegClass.hasSubClassEq(RC)) |
6728 | Opc = AArch64::FNMADDDrrr; |
6729 | else |
6730 | return nullptr; |
6731 | |
6732 | Register ResultReg = Root.getOperand(i: 0).getReg(); |
6733 | Register SrcReg0 = MAD->getOperand(i: 1).getReg(); |
6734 | Register SrcReg1 = MAD->getOperand(i: 2).getReg(); |
6735 | Register SrcReg2 = MAD->getOperand(i: 3).getReg(); |
6736 | bool Src0IsKill = MAD->getOperand(i: 1).isKill(); |
6737 | bool Src1IsKill = MAD->getOperand(i: 2).isKill(); |
6738 | bool Src2IsKill = MAD->getOperand(i: 3).isKill(); |
6739 | if (ResultReg.isVirtual()) |
6740 | MRI.constrainRegClass(Reg: ResultReg, RC); |
6741 | if (SrcReg0.isVirtual()) |
6742 | MRI.constrainRegClass(Reg: SrcReg0, RC); |
6743 | if (SrcReg1.isVirtual()) |
6744 | MRI.constrainRegClass(Reg: SrcReg1, RC); |
6745 | if (SrcReg2.isVirtual()) |
6746 | MRI.constrainRegClass(Reg: SrcReg2, RC); |
6747 | |
6748 | MachineInstrBuilder MIB = |
6749 | BuildMI(MF, MIMD: MIMetadata(Root), MCID: TII->get(Opcode: Opc), DestReg: ResultReg) |
6750 | .addReg(RegNo: SrcReg0, flags: getKillRegState(B: Src0IsKill)) |
6751 | .addReg(RegNo: SrcReg1, flags: getKillRegState(B: Src1IsKill)) |
6752 | .addReg(RegNo: SrcReg2, flags: getKillRegState(B: Src2IsKill)); |
6753 | InsInstrs.push_back(Elt: MIB); |
6754 | |
6755 | return MAD; |
6756 | } |
6757 | |
6758 | /// Fold (FMUL x (DUP y lane)) into (FMUL_indexed x y lane) |
6759 | static MachineInstr * |
6760 | genIndexedMultiply(MachineInstr &Root, |
6761 | SmallVectorImpl<MachineInstr *> &InsInstrs, |
6762 | unsigned IdxDupOp, unsigned MulOpc, |
6763 | const TargetRegisterClass *RC, MachineRegisterInfo &MRI) { |
6764 | assert(((IdxDupOp == 1) || (IdxDupOp == 2)) && |
6765 | "Invalid index of FMUL operand" ); |
6766 | |
6767 | MachineFunction &MF = *Root.getMF(); |
6768 | const TargetInstrInfo *TII = MF.getSubtarget().getInstrInfo(); |
6769 | |
6770 | MachineInstr *Dup = |
6771 | MF.getRegInfo().getUniqueVRegDef(Reg: Root.getOperand(i: IdxDupOp).getReg()); |
6772 | |
6773 | if (Dup->getOpcode() == TargetOpcode::COPY) |
6774 | Dup = MRI.getUniqueVRegDef(Reg: Dup->getOperand(i: 1).getReg()); |
6775 | |
6776 | Register DupSrcReg = Dup->getOperand(i: 1).getReg(); |
6777 | MRI.clearKillFlags(Reg: DupSrcReg); |
6778 | MRI.constrainRegClass(Reg: DupSrcReg, RC); |
6779 | |
6780 | unsigned DupSrcLane = Dup->getOperand(i: 2).getImm(); |
6781 | |
6782 | unsigned IdxMulOp = IdxDupOp == 1 ? 2 : 1; |
6783 | MachineOperand &MulOp = Root.getOperand(i: IdxMulOp); |
6784 | |
6785 | Register ResultReg = Root.getOperand(i: 0).getReg(); |
6786 | |
6787 | MachineInstrBuilder MIB; |
6788 | MIB = BuildMI(MF, MIMD: MIMetadata(Root), MCID: TII->get(Opcode: MulOpc), DestReg: ResultReg) |
6789 | .add(MO: MulOp) |
6790 | .addReg(RegNo: DupSrcReg) |
6791 | .addImm(Val: DupSrcLane); |
6792 | |
6793 | InsInstrs.push_back(Elt: MIB); |
6794 | return &Root; |
6795 | } |
6796 | |
6797 | /// genFusedMultiplyAcc - Helper to generate fused multiply accumulate |
6798 | /// instructions. |
6799 | /// |
6800 | /// \see genFusedMultiply |
6801 | static MachineInstr *genFusedMultiplyAcc( |
6802 | MachineFunction &MF, MachineRegisterInfo &MRI, const TargetInstrInfo *TII, |
6803 | MachineInstr &Root, SmallVectorImpl<MachineInstr *> &InsInstrs, |
6804 | unsigned IdxMulOpd, unsigned MaddOpc, const TargetRegisterClass *RC) { |
6805 | return genFusedMultiply(MF, MRI, TII, Root, InsInstrs, IdxMulOpd, MaddOpc, RC, |
6806 | kind: FMAInstKind::Accumulator); |
6807 | } |
6808 | |
6809 | /// genNeg - Helper to generate an intermediate negation of the second operand |
6810 | /// of Root |
6811 | static Register genNeg(MachineFunction &MF, MachineRegisterInfo &MRI, |
6812 | const TargetInstrInfo *TII, MachineInstr &Root, |
6813 | SmallVectorImpl<MachineInstr *> &InsInstrs, |
6814 | DenseMap<unsigned, unsigned> &InstrIdxForVirtReg, |
6815 | unsigned MnegOpc, const TargetRegisterClass *RC) { |
6816 | Register NewVR = MRI.createVirtualRegister(RegClass: RC); |
6817 | MachineInstrBuilder MIB = |
6818 | BuildMI(MF, MIMD: MIMetadata(Root), MCID: TII->get(Opcode: MnegOpc), DestReg: NewVR) |
6819 | .add(MO: Root.getOperand(i: 2)); |
6820 | InsInstrs.push_back(Elt: MIB); |
6821 | |
6822 | assert(InstrIdxForVirtReg.empty()); |
6823 | InstrIdxForVirtReg.insert(KV: std::make_pair(x&: NewVR, y: 0)); |
6824 | |
6825 | return NewVR; |
6826 | } |
6827 | |
6828 | /// genFusedMultiplyAccNeg - Helper to generate fused multiply accumulate |
6829 | /// instructions with an additional negation of the accumulator |
6830 | static MachineInstr *genFusedMultiplyAccNeg( |
6831 | MachineFunction &MF, MachineRegisterInfo &MRI, const TargetInstrInfo *TII, |
6832 | MachineInstr &Root, SmallVectorImpl<MachineInstr *> &InsInstrs, |
6833 | DenseMap<unsigned, unsigned> &InstrIdxForVirtReg, unsigned IdxMulOpd, |
6834 | unsigned MaddOpc, unsigned MnegOpc, const TargetRegisterClass *RC) { |
6835 | assert(IdxMulOpd == 1); |
6836 | |
6837 | Register NewVR = |
6838 | genNeg(MF, MRI, TII, Root, InsInstrs, InstrIdxForVirtReg, MnegOpc, RC); |
6839 | return genFusedMultiply(MF, MRI, TII, Root, InsInstrs, IdxMulOpd, MaddOpc, RC, |
6840 | kind: FMAInstKind::Accumulator, ReplacedAddend: &NewVR); |
6841 | } |
6842 | |
6843 | /// genFusedMultiplyIdx - Helper to generate fused multiply accumulate |
6844 | /// instructions. |
6845 | /// |
6846 | /// \see genFusedMultiply |
6847 | static MachineInstr *genFusedMultiplyIdx( |
6848 | MachineFunction &MF, MachineRegisterInfo &MRI, const TargetInstrInfo *TII, |
6849 | MachineInstr &Root, SmallVectorImpl<MachineInstr *> &InsInstrs, |
6850 | unsigned IdxMulOpd, unsigned MaddOpc, const TargetRegisterClass *RC) { |
6851 | return genFusedMultiply(MF, MRI, TII, Root, InsInstrs, IdxMulOpd, MaddOpc, RC, |
6852 | kind: FMAInstKind::Indexed); |
6853 | } |
6854 | |
6855 | /// genFusedMultiplyAccNeg - Helper to generate fused multiply accumulate |
6856 | /// instructions with an additional negation of the accumulator |
6857 | static MachineInstr *genFusedMultiplyIdxNeg( |
6858 | MachineFunction &MF, MachineRegisterInfo &MRI, const TargetInstrInfo *TII, |
6859 | MachineInstr &Root, SmallVectorImpl<MachineInstr *> &InsInstrs, |
6860 | DenseMap<unsigned, unsigned> &InstrIdxForVirtReg, unsigned IdxMulOpd, |
6861 | unsigned MaddOpc, unsigned MnegOpc, const TargetRegisterClass *RC) { |
6862 | assert(IdxMulOpd == 1); |
6863 | |
6864 | Register NewVR = |
6865 | genNeg(MF, MRI, TII, Root, InsInstrs, InstrIdxForVirtReg, MnegOpc, RC); |
6866 | |
6867 | return genFusedMultiply(MF, MRI, TII, Root, InsInstrs, IdxMulOpd, MaddOpc, RC, |
6868 | kind: FMAInstKind::Indexed, ReplacedAddend: &NewVR); |
6869 | } |
6870 | |
6871 | /// genMaddR - Generate madd instruction and combine mul and add using |
6872 | /// an extra virtual register |
6873 | /// Example - an ADD intermediate needs to be stored in a register: |
6874 | /// MUL I=A,B,0 |
6875 | /// ADD R,I,Imm |
6876 | /// ==> ORR V, ZR, Imm |
6877 | /// ==> MADD R,A,B,V |
6878 | /// \param MF Containing MachineFunction |
6879 | /// \param MRI Register information |
6880 | /// \param TII Target information |
6881 | /// \param Root is the ADD instruction |
6882 | /// \param [out] InsInstrs is a vector of machine instructions and will |
6883 | /// contain the generated madd instruction |
6884 | /// \param IdxMulOpd is index of operand in Root that is the result of |
6885 | /// the MUL. In the example above IdxMulOpd is 1. |
6886 | /// \param MaddOpc the opcode fo the madd instruction |
6887 | /// \param VR is a virtual register that holds the value of an ADD operand |
6888 | /// (V in the example above). |
6889 | /// \param RC Register class of operands |
6890 | static MachineInstr *genMaddR(MachineFunction &MF, MachineRegisterInfo &MRI, |
6891 | const TargetInstrInfo *TII, MachineInstr &Root, |
6892 | SmallVectorImpl<MachineInstr *> &InsInstrs, |
6893 | unsigned IdxMulOpd, unsigned MaddOpc, unsigned VR, |
6894 | const TargetRegisterClass *RC) { |
6895 | assert(IdxMulOpd == 1 || IdxMulOpd == 2); |
6896 | |
6897 | MachineInstr *MUL = MRI.getUniqueVRegDef(Reg: Root.getOperand(i: IdxMulOpd).getReg()); |
6898 | Register ResultReg = Root.getOperand(i: 0).getReg(); |
6899 | Register SrcReg0 = MUL->getOperand(i: 1).getReg(); |
6900 | bool Src0IsKill = MUL->getOperand(i: 1).isKill(); |
6901 | Register SrcReg1 = MUL->getOperand(i: 2).getReg(); |
6902 | bool Src1IsKill = MUL->getOperand(i: 2).isKill(); |
6903 | |
6904 | if (ResultReg.isVirtual()) |
6905 | MRI.constrainRegClass(Reg: ResultReg, RC); |
6906 | if (SrcReg0.isVirtual()) |
6907 | MRI.constrainRegClass(Reg: SrcReg0, RC); |
6908 | if (SrcReg1.isVirtual()) |
6909 | MRI.constrainRegClass(Reg: SrcReg1, RC); |
6910 | if (Register::isVirtualRegister(Reg: VR)) |
6911 | MRI.constrainRegClass(Reg: VR, RC); |
6912 | |
6913 | MachineInstrBuilder MIB = |
6914 | BuildMI(MF, MIMD: MIMetadata(Root), MCID: TII->get(Opcode: MaddOpc), DestReg: ResultReg) |
6915 | .addReg(RegNo: SrcReg0, flags: getKillRegState(B: Src0IsKill)) |
6916 | .addReg(RegNo: SrcReg1, flags: getKillRegState(B: Src1IsKill)) |
6917 | .addReg(RegNo: VR); |
6918 | // Insert the MADD |
6919 | InsInstrs.push_back(Elt: MIB); |
6920 | return MUL; |
6921 | } |
6922 | |
6923 | /// Do the following transformation |
6924 | /// A - (B + C) ==> (A - B) - C |
6925 | /// A - (B + C) ==> (A - C) - B |
6926 | static void |
6927 | genSubAdd2SubSub(MachineFunction &MF, MachineRegisterInfo &MRI, |
6928 | const TargetInstrInfo *TII, MachineInstr &Root, |
6929 | SmallVectorImpl<MachineInstr *> &InsInstrs, |
6930 | SmallVectorImpl<MachineInstr *> &DelInstrs, |
6931 | unsigned IdxOpd1, |
6932 | DenseMap<unsigned, unsigned> &InstrIdxForVirtReg) { |
6933 | assert(IdxOpd1 == 1 || IdxOpd1 == 2); |
6934 | unsigned IdxOtherOpd = IdxOpd1 == 1 ? 2 : 1; |
6935 | MachineInstr *AddMI = MRI.getUniqueVRegDef(Reg: Root.getOperand(i: 2).getReg()); |
6936 | |
6937 | Register ResultReg = Root.getOperand(i: 0).getReg(); |
6938 | Register RegA = Root.getOperand(i: 1).getReg(); |
6939 | bool RegAIsKill = Root.getOperand(i: 1).isKill(); |
6940 | Register RegB = AddMI->getOperand(i: IdxOpd1).getReg(); |
6941 | bool RegBIsKill = AddMI->getOperand(i: IdxOpd1).isKill(); |
6942 | Register RegC = AddMI->getOperand(i: IdxOtherOpd).getReg(); |
6943 | bool RegCIsKill = AddMI->getOperand(i: IdxOtherOpd).isKill(); |
6944 | Register NewVR = MRI.createVirtualRegister(RegClass: MRI.getRegClass(Reg: RegA)); |
6945 | |
6946 | unsigned Opcode = Root.getOpcode(); |
6947 | if (Opcode == AArch64::SUBSWrr) |
6948 | Opcode = AArch64::SUBWrr; |
6949 | else if (Opcode == AArch64::SUBSXrr) |
6950 | Opcode = AArch64::SUBXrr; |
6951 | else |
6952 | assert((Opcode == AArch64::SUBWrr || Opcode == AArch64::SUBXrr) && |
6953 | "Unexpected instruction opcode." ); |
6954 | |
6955 | uint32_t Flags = Root.mergeFlagsWith(Other: *AddMI); |
6956 | Flags &= ~MachineInstr::NoSWrap; |
6957 | Flags &= ~MachineInstr::NoUWrap; |
6958 | |
6959 | MachineInstrBuilder MIB1 = |
6960 | BuildMI(MF, MIMD: MIMetadata(Root), MCID: TII->get(Opcode), DestReg: NewVR) |
6961 | .addReg(RegNo: RegA, flags: getKillRegState(B: RegAIsKill)) |
6962 | .addReg(RegNo: RegB, flags: getKillRegState(B: RegBIsKill)) |
6963 | .setMIFlags(Flags); |
6964 | MachineInstrBuilder MIB2 = |
6965 | BuildMI(MF, MIMD: MIMetadata(Root), MCID: TII->get(Opcode), DestReg: ResultReg) |
6966 | .addReg(RegNo: NewVR, flags: getKillRegState(B: true)) |
6967 | .addReg(RegNo: RegC, flags: getKillRegState(B: RegCIsKill)) |
6968 | .setMIFlags(Flags); |
6969 | |
6970 | InstrIdxForVirtReg.insert(KV: std::make_pair(x&: NewVR, y: 0)); |
6971 | InsInstrs.push_back(Elt: MIB1); |
6972 | InsInstrs.push_back(Elt: MIB2); |
6973 | DelInstrs.push_back(Elt: AddMI); |
6974 | DelInstrs.push_back(Elt: &Root); |
6975 | } |
6976 | |
6977 | /// When getMachineCombinerPatterns() finds potential patterns, |
6978 | /// this function generates the instructions that could replace the |
6979 | /// original code sequence |
6980 | void AArch64InstrInfo::genAlternativeCodeSequence( |
6981 | MachineInstr &Root, unsigned Pattern, |
6982 | SmallVectorImpl<MachineInstr *> &InsInstrs, |
6983 | SmallVectorImpl<MachineInstr *> &DelInstrs, |
6984 | DenseMap<unsigned, unsigned> &InstrIdxForVirtReg) const { |
6985 | MachineBasicBlock &MBB = *Root.getParent(); |
6986 | MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo(); |
6987 | MachineFunction &MF = *MBB.getParent(); |
6988 | const TargetInstrInfo *TII = MF.getSubtarget().getInstrInfo(); |
6989 | |
6990 | MachineInstr *MUL = nullptr; |
6991 | const TargetRegisterClass *RC; |
6992 | unsigned Opc; |
6993 | switch (Pattern) { |
6994 | default: |
6995 | // Reassociate instructions. |
6996 | TargetInstrInfo::genAlternativeCodeSequence(Root, Pattern, InsInstrs, |
6997 | DelInstrs, InstIdxForVirtReg&: InstrIdxForVirtReg); |
6998 | return; |
6999 | case AArch64MachineCombinerPattern::SUBADD_OP1: |
7000 | // A - (B + C) |
7001 | // ==> (A - B) - C |
7002 | genSubAdd2SubSub(MF, MRI, TII, Root, InsInstrs, DelInstrs, IdxOpd1: 1, |
7003 | InstrIdxForVirtReg); |
7004 | return; |
7005 | case AArch64MachineCombinerPattern::SUBADD_OP2: |
7006 | // A - (B + C) |
7007 | // ==> (A - C) - B |
7008 | genSubAdd2SubSub(MF, MRI, TII, Root, InsInstrs, DelInstrs, IdxOpd1: 2, |
7009 | InstrIdxForVirtReg); |
7010 | return; |
7011 | case AArch64MachineCombinerPattern::MULADDW_OP1: |
7012 | case AArch64MachineCombinerPattern::MULADDX_OP1: |
7013 | // MUL I=A,B,0 |
7014 | // ADD R,I,C |
7015 | // ==> MADD R,A,B,C |
7016 | // --- Create(MADD); |
7017 | if (Pattern == AArch64MachineCombinerPattern::MULADDW_OP1) { |
7018 | Opc = AArch64::MADDWrrr; |
7019 | RC = &AArch64::GPR32RegClass; |
7020 | } else { |
7021 | Opc = AArch64::MADDXrrr; |
7022 | RC = &AArch64::GPR64RegClass; |
7023 | } |
7024 | MUL = genFusedMultiply(MF, MRI, TII, Root, InsInstrs, IdxMulOpd: 1, MaddOpc: Opc, RC); |
7025 | break; |
7026 | case AArch64MachineCombinerPattern::MULADDW_OP2: |
7027 | case AArch64MachineCombinerPattern::MULADDX_OP2: |
7028 | // MUL I=A,B,0 |
7029 | // ADD R,C,I |
7030 | // ==> MADD R,A,B,C |
7031 | // --- Create(MADD); |
7032 | if (Pattern == AArch64MachineCombinerPattern::MULADDW_OP2) { |
7033 | Opc = AArch64::MADDWrrr; |
7034 | RC = &AArch64::GPR32RegClass; |
7035 | } else { |
7036 | Opc = AArch64::MADDXrrr; |
7037 | RC = &AArch64::GPR64RegClass; |
7038 | } |
7039 | MUL = genFusedMultiply(MF, MRI, TII, Root, InsInstrs, IdxMulOpd: 2, MaddOpc: Opc, RC); |
7040 | break; |
7041 | case AArch64MachineCombinerPattern::MULADDWI_OP1: |
7042 | case AArch64MachineCombinerPattern::MULADDXI_OP1: { |
7043 | // MUL I=A,B,0 |
7044 | // ADD R,I,Imm |
7045 | // ==> MOV V, Imm |
7046 | // ==> MADD R,A,B,V |
7047 | // --- Create(MADD); |
7048 | const TargetRegisterClass *OrrRC; |
7049 | unsigned BitSize, OrrOpc, ZeroReg; |
7050 | if (Pattern == AArch64MachineCombinerPattern::MULADDWI_OP1) { |
7051 | OrrOpc = AArch64::ORRWri; |
7052 | OrrRC = &AArch64::GPR32spRegClass; |
7053 | BitSize = 32; |
7054 | ZeroReg = AArch64::WZR; |
7055 | Opc = AArch64::MADDWrrr; |
7056 | RC = &AArch64::GPR32RegClass; |
7057 | } else { |
7058 | OrrOpc = AArch64::ORRXri; |
7059 | OrrRC = &AArch64::GPR64spRegClass; |
7060 | BitSize = 64; |
7061 | ZeroReg = AArch64::XZR; |
7062 | Opc = AArch64::MADDXrrr; |
7063 | RC = &AArch64::GPR64RegClass; |
7064 | } |
7065 | Register NewVR = MRI.createVirtualRegister(RegClass: OrrRC); |
7066 | uint64_t Imm = Root.getOperand(i: 2).getImm(); |
7067 | |
7068 | if (Root.getOperand(i: 3).isImm()) { |
7069 | unsigned Val = Root.getOperand(i: 3).getImm(); |
7070 | Imm = Imm << Val; |
7071 | } |
7072 | uint64_t UImm = SignExtend64(X: Imm, B: BitSize); |
7073 | // The immediate can be composed via a single instruction. |
7074 | SmallVector<AArch64_IMM::ImmInsnModel, 4> Insn; |
7075 | AArch64_IMM::expandMOVImm(Imm: UImm, BitSize, Insn); |
7076 | if (Insn.size() != 1) |
7077 | return; |
7078 | auto MovI = Insn.begin(); |
7079 | MachineInstrBuilder MIB1; |
7080 | // MOV is an alias for one of three instructions: movz, movn, and orr. |
7081 | if (MovI->Opcode == OrrOpc) |
7082 | MIB1 = BuildMI(MF, MIMD: MIMetadata(Root), MCID: TII->get(Opcode: OrrOpc), DestReg: NewVR) |
7083 | .addReg(RegNo: ZeroReg) |
7084 | .addImm(Val: MovI->Op2); |
7085 | else { |
7086 | if (BitSize == 32) |
7087 | assert((MovI->Opcode == AArch64::MOVNWi || |
7088 | MovI->Opcode == AArch64::MOVZWi) && |
7089 | "Expected opcode" ); |
7090 | else |
7091 | assert((MovI->Opcode == AArch64::MOVNXi || |
7092 | MovI->Opcode == AArch64::MOVZXi) && |
7093 | "Expected opcode" ); |
7094 | MIB1 = BuildMI(MF, MIMD: MIMetadata(Root), MCID: TII->get(Opcode: MovI->Opcode), DestReg: NewVR) |
7095 | .addImm(Val: MovI->Op1) |
7096 | .addImm(Val: MovI->Op2); |
7097 | } |
7098 | InsInstrs.push_back(Elt: MIB1); |
7099 | InstrIdxForVirtReg.insert(KV: std::make_pair(x&: NewVR, y: 0)); |
7100 | MUL = genMaddR(MF, MRI, TII, Root, InsInstrs, IdxMulOpd: 1, MaddOpc: Opc, VR: NewVR, RC); |
7101 | break; |
7102 | } |
7103 | case AArch64MachineCombinerPattern::MULSUBW_OP1: |
7104 | case AArch64MachineCombinerPattern::MULSUBX_OP1: { |
7105 | // MUL I=A,B,0 |
7106 | // SUB R,I, C |
7107 | // ==> SUB V, 0, C |
7108 | // ==> MADD R,A,B,V // = -C + A*B |
7109 | // --- Create(MADD); |
7110 | const TargetRegisterClass *SubRC; |
7111 | unsigned SubOpc, ZeroReg; |
7112 | if (Pattern == AArch64MachineCombinerPattern::MULSUBW_OP1) { |
7113 | SubOpc = AArch64::SUBWrr; |
7114 | SubRC = &AArch64::GPR32spRegClass; |
7115 | ZeroReg = AArch64::WZR; |
7116 | Opc = AArch64::MADDWrrr; |
7117 | RC = &AArch64::GPR32RegClass; |
7118 | } else { |
7119 | SubOpc = AArch64::SUBXrr; |
7120 | SubRC = &AArch64::GPR64spRegClass; |
7121 | ZeroReg = AArch64::XZR; |
7122 | Opc = AArch64::MADDXrrr; |
7123 | RC = &AArch64::GPR64RegClass; |
7124 | } |
7125 | Register NewVR = MRI.createVirtualRegister(RegClass: SubRC); |
7126 | // SUB NewVR, 0, C |
7127 | MachineInstrBuilder MIB1 = |
7128 | BuildMI(MF, MIMD: MIMetadata(Root), MCID: TII->get(Opcode: SubOpc), DestReg: NewVR) |
7129 | .addReg(RegNo: ZeroReg) |
7130 | .add(MO: Root.getOperand(i: 2)); |
7131 | InsInstrs.push_back(Elt: MIB1); |
7132 | InstrIdxForVirtReg.insert(KV: std::make_pair(x&: NewVR, y: 0)); |
7133 | MUL = genMaddR(MF, MRI, TII, Root, InsInstrs, IdxMulOpd: 1, MaddOpc: Opc, VR: NewVR, RC); |
7134 | break; |
7135 | } |
7136 | case AArch64MachineCombinerPattern::MULSUBW_OP2: |
7137 | case AArch64MachineCombinerPattern::MULSUBX_OP2: |
7138 | // MUL I=A,B,0 |
7139 | // SUB R,C,I |
7140 | // ==> MSUB R,A,B,C (computes C - A*B) |
7141 | // --- Create(MSUB); |
7142 | if (Pattern == AArch64MachineCombinerPattern::MULSUBW_OP2) { |
7143 | Opc = AArch64::MSUBWrrr; |
7144 | RC = &AArch64::GPR32RegClass; |
7145 | } else { |
7146 | Opc = AArch64::MSUBXrrr; |
7147 | RC = &AArch64::GPR64RegClass; |
7148 | } |
7149 | MUL = genFusedMultiply(MF, MRI, TII, Root, InsInstrs, IdxMulOpd: 2, MaddOpc: Opc, RC); |
7150 | break; |
7151 | case AArch64MachineCombinerPattern::MULSUBWI_OP1: |
7152 | case AArch64MachineCombinerPattern::MULSUBXI_OP1: { |
7153 | // MUL I=A,B,0 |
7154 | // SUB R,I, Imm |
7155 | // ==> MOV V, -Imm |
7156 | // ==> MADD R,A,B,V // = -Imm + A*B |
7157 | // --- Create(MADD); |
7158 | const TargetRegisterClass *OrrRC; |
7159 | unsigned BitSize, OrrOpc, ZeroReg; |
7160 | if (Pattern == AArch64MachineCombinerPattern::MULSUBWI_OP1) { |
7161 | OrrOpc = AArch64::ORRWri; |
7162 | OrrRC = &AArch64::GPR32spRegClass; |
7163 | BitSize = 32; |
7164 | ZeroReg = AArch64::WZR; |
7165 | Opc = AArch64::MADDWrrr; |
7166 | RC = &AArch64::GPR32RegClass; |
7167 | } else { |
7168 | OrrOpc = AArch64::ORRXri; |
7169 | OrrRC = &AArch64::GPR64spRegClass; |
7170 | BitSize = 64; |
7171 | ZeroReg = AArch64::XZR; |
7172 | Opc = AArch64::MADDXrrr; |
7173 | RC = &AArch64::GPR64RegClass; |
7174 | } |
7175 | Register NewVR = MRI.createVirtualRegister(RegClass: OrrRC); |
7176 | uint64_t Imm = Root.getOperand(i: 2).getImm(); |
7177 | if (Root.getOperand(i: 3).isImm()) { |
7178 | unsigned Val = Root.getOperand(i: 3).getImm(); |
7179 | Imm = Imm << Val; |
7180 | } |
7181 | uint64_t UImm = SignExtend64(X: -Imm, B: BitSize); |
7182 | // The immediate can be composed via a single instruction. |
7183 | SmallVector<AArch64_IMM::ImmInsnModel, 4> Insn; |
7184 | AArch64_IMM::expandMOVImm(Imm: UImm, BitSize, Insn); |
7185 | if (Insn.size() != 1) |
7186 | return; |
7187 | auto MovI = Insn.begin(); |
7188 | MachineInstrBuilder MIB1; |
7189 | // MOV is an alias for one of three instructions: movz, movn, and orr. |
7190 | if (MovI->Opcode == OrrOpc) |
7191 | MIB1 = BuildMI(MF, MIMD: MIMetadata(Root), MCID: TII->get(Opcode: OrrOpc), DestReg: NewVR) |
7192 | .addReg(RegNo: ZeroReg) |
7193 | .addImm(Val: MovI->Op2); |
7194 | else { |
7195 | if (BitSize == 32) |
7196 | assert((MovI->Opcode == AArch64::MOVNWi || |
7197 | MovI->Opcode == AArch64::MOVZWi) && |
7198 | "Expected opcode" ); |
7199 | else |
7200 | assert((MovI->Opcode == AArch64::MOVNXi || |
7201 | MovI->Opcode == AArch64::MOVZXi) && |
7202 | "Expected opcode" ); |
7203 | MIB1 = BuildMI(MF, MIMD: MIMetadata(Root), MCID: TII->get(Opcode: MovI->Opcode), DestReg: NewVR) |
7204 | .addImm(Val: MovI->Op1) |
7205 | .addImm(Val: MovI->Op2); |
7206 | } |
7207 | InsInstrs.push_back(Elt: MIB1); |
7208 | InstrIdxForVirtReg.insert(KV: std::make_pair(x&: NewVR, y: 0)); |
7209 | MUL = genMaddR(MF, MRI, TII, Root, InsInstrs, IdxMulOpd: 1, MaddOpc: Opc, VR: NewVR, RC); |
7210 | break; |
7211 | } |
7212 | |
7213 | case AArch64MachineCombinerPattern::MULADDv8i8_OP1: |
7214 | Opc = AArch64::MLAv8i8; |
7215 | RC = &AArch64::FPR64RegClass; |
7216 | MUL = genFusedMultiplyAcc(MF, MRI, TII, Root, InsInstrs, IdxMulOpd: 1, MaddOpc: Opc, RC); |
7217 | break; |
7218 | case AArch64MachineCombinerPattern::MULADDv8i8_OP2: |
7219 | Opc = AArch64::MLAv8i8; |
7220 | RC = &AArch64::FPR64RegClass; |
7221 | MUL = genFusedMultiplyAcc(MF, MRI, TII, Root, InsInstrs, IdxMulOpd: 2, MaddOpc: Opc, RC); |
7222 | break; |
7223 | case AArch64MachineCombinerPattern::MULADDv16i8_OP1: |
7224 | Opc = AArch64::MLAv16i8; |
7225 | RC = &AArch64::FPR128RegClass; |
7226 | MUL = genFusedMultiplyAcc(MF, MRI, TII, Root, InsInstrs, IdxMulOpd: 1, MaddOpc: Opc, RC); |
7227 | break; |
7228 | case AArch64MachineCombinerPattern::MULADDv16i8_OP2: |
7229 | Opc = AArch64::MLAv16i8; |
7230 | RC = &AArch64::FPR128RegClass; |
7231 | MUL = genFusedMultiplyAcc(MF, MRI, TII, Root, InsInstrs, IdxMulOpd: 2, MaddOpc: Opc, RC); |
7232 | break; |
7233 | case AArch64MachineCombinerPattern::MULADDv4i16_OP1: |
7234 | Opc = AArch64::MLAv4i16; |
7235 | RC = &AArch64::FPR64RegClass; |
7236 | MUL = genFusedMultiplyAcc(MF, MRI, TII, Root, InsInstrs, IdxMulOpd: 1, MaddOpc: Opc, RC); |
7237 | break; |
7238 | case AArch64MachineCombinerPattern::MULADDv4i16_OP2: |
7239 | Opc = AArch64::MLAv4i16; |
7240 | RC = &AArch64::FPR64RegClass; |
7241 | MUL = genFusedMultiplyAcc(MF, MRI, TII, Root, InsInstrs, IdxMulOpd: 2, MaddOpc: Opc, RC); |
7242 | break; |
7243 | case AArch64MachineCombinerPattern::MULADDv8i16_OP1: |
7244 | Opc = AArch64::MLAv8i16; |
7245 | RC = &AArch64::FPR128RegClass; |
7246 | MUL = genFusedMultiplyAcc(MF, MRI, TII, Root, InsInstrs, IdxMulOpd: 1, MaddOpc: Opc, RC); |
7247 | break; |
7248 | case AArch64MachineCombinerPattern::MULADDv8i16_OP2: |
7249 | Opc = AArch64::MLAv8i16; |
7250 | RC = &AArch64::FPR128RegClass; |
7251 | MUL = genFusedMultiplyAcc(MF, MRI, TII, Root, InsInstrs, IdxMulOpd: 2, MaddOpc: Opc, RC); |
7252 | break; |
7253 | case AArch64MachineCombinerPattern::MULADDv2i32_OP1: |
7254 | Opc = AArch64::MLAv2i32; |
7255 | RC = &AArch64::FPR64RegClass; |
7256 | MUL = genFusedMultiplyAcc(MF, MRI, TII, Root, InsInstrs, IdxMulOpd: 1, MaddOpc: Opc, RC); |
7257 | break; |
7258 | case AArch64MachineCombinerPattern::MULADDv2i32_OP2: |
7259 | Opc = AArch64::MLAv2i32; |
7260 | RC = &AArch64::FPR64RegClass; |
7261 | MUL = genFusedMultiplyAcc(MF, MRI, TII, Root, InsInstrs, IdxMulOpd: 2, MaddOpc: Opc, RC); |
7262 | break; |
7263 | case AArch64MachineCombinerPattern::MULADDv4i32_OP1: |
7264 | Opc = AArch64::MLAv4i32; |
7265 | RC = &AArch64::FPR128RegClass; |
7266 | MUL = genFusedMultiplyAcc(MF, MRI, TII, Root, InsInstrs, IdxMulOpd: 1, MaddOpc: Opc, RC); |
7267 | break; |
7268 | case AArch64MachineCombinerPattern::MULADDv4i32_OP2: |
7269 | Opc = AArch64::MLAv4i32; |
7270 | RC = &AArch64::FPR128RegClass; |
7271 | MUL = genFusedMultiplyAcc(MF, MRI, TII, Root, InsInstrs, IdxMulOpd: 2, MaddOpc: Opc, RC); |
7272 | break; |
7273 | |
7274 | case AArch64MachineCombinerPattern::MULSUBv8i8_OP1: |
7275 | Opc = AArch64::MLAv8i8; |
7276 | RC = &AArch64::FPR64RegClass; |
7277 | MUL = genFusedMultiplyAccNeg(MF, MRI, TII, Root, InsInstrs, |
7278 | InstrIdxForVirtReg, IdxMulOpd: 1, MaddOpc: Opc, MnegOpc: AArch64::NEGv8i8, |
7279 | RC); |
7280 | break; |
7281 | case AArch64MachineCombinerPattern::MULSUBv8i8_OP2: |
7282 | Opc = AArch64::MLSv8i8; |
7283 | RC = &AArch64::FPR64RegClass; |
7284 | MUL = genFusedMultiplyAcc(MF, MRI, TII, Root, InsInstrs, IdxMulOpd: 2, MaddOpc: Opc, RC); |
7285 | break; |
7286 | case AArch64MachineCombinerPattern::MULSUBv16i8_OP1: |
7287 | Opc = AArch64::MLAv16i8; |
7288 | RC = &AArch64::FPR128RegClass; |
7289 | MUL = genFusedMultiplyAccNeg(MF, MRI, TII, Root, InsInstrs, |
7290 | InstrIdxForVirtReg, IdxMulOpd: 1, MaddOpc: Opc, MnegOpc: AArch64::NEGv16i8, |
7291 | RC); |
7292 | break; |
7293 | case AArch64MachineCombinerPattern::MULSUBv16i8_OP2: |
7294 | Opc = AArch64::MLSv16i8; |
7295 | RC = &AArch64::FPR128RegClass; |
7296 | MUL = genFusedMultiplyAcc(MF, MRI, TII, Root, InsInstrs, IdxMulOpd: 2, MaddOpc: Opc, RC); |
7297 | break; |
7298 | case AArch64MachineCombinerPattern::MULSUBv4i16_OP1: |
7299 | Opc = AArch64::MLAv4i16; |
7300 | RC = &AArch64::FPR64RegClass; |
7301 | MUL = genFusedMultiplyAccNeg(MF, MRI, TII, Root, InsInstrs, |
7302 | InstrIdxForVirtReg, IdxMulOpd: 1, MaddOpc: Opc, MnegOpc: AArch64::NEGv4i16, |
7303 | RC); |
7304 | break; |
7305 | case AArch64MachineCombinerPattern::MULSUBv4i16_OP2: |
7306 | Opc = AArch64::MLSv4i16; |
7307 | RC = &AArch64::FPR64RegClass; |
7308 | MUL = genFusedMultiplyAcc(MF, MRI, TII, Root, InsInstrs, IdxMulOpd: 2, MaddOpc: Opc, RC); |
7309 | break; |
7310 | case AArch64MachineCombinerPattern::MULSUBv8i16_OP1: |
7311 | Opc = AArch64::MLAv8i16; |
7312 | RC = &AArch64::FPR128RegClass; |
7313 | MUL = genFusedMultiplyAccNeg(MF, MRI, TII, Root, InsInstrs, |
7314 | InstrIdxForVirtReg, IdxMulOpd: 1, MaddOpc: Opc, MnegOpc: AArch64::NEGv8i16, |
7315 | RC); |
7316 | break; |
7317 | case AArch64MachineCombinerPattern::MULSUBv8i16_OP2: |
7318 | Opc = AArch64::MLSv8i16; |
7319 | RC = &AArch64::FPR128RegClass; |
7320 | MUL = genFusedMultiplyAcc(MF, MRI, TII, Root, InsInstrs, IdxMulOpd: 2, MaddOpc: Opc, RC); |
7321 | break; |
7322 | case AArch64MachineCombinerPattern::MULSUBv2i32_OP1: |
7323 | Opc = AArch64::MLAv2i32; |
7324 | RC = &AArch64::FPR64RegClass; |
7325 | MUL = genFusedMultiplyAccNeg(MF, MRI, TII, Root, InsInstrs, |
7326 | InstrIdxForVirtReg, IdxMulOpd: 1, MaddOpc: Opc, MnegOpc: AArch64::NEGv2i32, |
7327 | RC); |
7328 | break; |
7329 | case AArch64MachineCombinerPattern::MULSUBv2i32_OP2: |
7330 | Opc = AArch64::MLSv2i32; |
7331 | RC = &AArch64::FPR64RegClass; |
7332 | MUL = genFusedMultiplyAcc(MF, MRI, TII, Root, InsInstrs, IdxMulOpd: 2, MaddOpc: Opc, RC); |
7333 | break; |
7334 | case AArch64MachineCombinerPattern::MULSUBv4i32_OP1: |
7335 | Opc = AArch64::MLAv4i32; |
7336 | RC = &AArch64::FPR128RegClass; |
7337 | MUL = genFusedMultiplyAccNeg(MF, MRI, TII, Root, InsInstrs, |
7338 | InstrIdxForVirtReg, IdxMulOpd: 1, MaddOpc: Opc, MnegOpc: AArch64::NEGv4i32, |
7339 | RC); |
7340 | break; |
7341 | case AArch64MachineCombinerPattern::MULSUBv4i32_OP2: |
7342 | Opc = AArch64::MLSv4i32; |
7343 | RC = &AArch64::FPR128RegClass; |
7344 | MUL = genFusedMultiplyAcc(MF, MRI, TII, Root, InsInstrs, IdxMulOpd: 2, MaddOpc: Opc, RC); |
7345 | break; |
7346 | |
7347 | case AArch64MachineCombinerPattern::MULADDv4i16_indexed_OP1: |
7348 | Opc = AArch64::MLAv4i16_indexed; |
7349 | RC = &AArch64::FPR64RegClass; |
7350 | MUL = genFusedMultiplyIdx(MF, MRI, TII, Root, InsInstrs, IdxMulOpd: 1, MaddOpc: Opc, RC); |
7351 | break; |
7352 | case AArch64MachineCombinerPattern::MULADDv4i16_indexed_OP2: |
7353 | Opc = AArch64::MLAv4i16_indexed; |
7354 | RC = &AArch64::FPR64RegClass; |
7355 | MUL = genFusedMultiplyIdx(MF, MRI, TII, Root, InsInstrs, IdxMulOpd: 2, MaddOpc: Opc, RC); |
7356 | break; |
7357 | case AArch64MachineCombinerPattern::MULADDv8i16_indexed_OP1: |
7358 | Opc = AArch64::MLAv8i16_indexed; |
7359 | RC = &AArch64::FPR128RegClass; |
7360 | MUL = genFusedMultiplyIdx(MF, MRI, TII, Root, InsInstrs, IdxMulOpd: 1, MaddOpc: Opc, RC); |
7361 | break; |
7362 | case AArch64MachineCombinerPattern::MULADDv8i16_indexed_OP2: |
7363 | Opc = AArch64::MLAv8i16_indexed; |
7364 | RC = &AArch64::FPR128RegClass; |
7365 | MUL = genFusedMultiplyIdx(MF, MRI, TII, Root, InsInstrs, IdxMulOpd: 2, MaddOpc: Opc, RC); |
7366 | break; |
7367 | case AArch64MachineCombinerPattern::MULADDv2i32_indexed_OP1: |
7368 | Opc = AArch64::MLAv2i32_indexed; |
7369 | RC = &AArch64::FPR64RegClass; |
7370 | MUL = genFusedMultiplyIdx(MF, MRI, TII, Root, InsInstrs, IdxMulOpd: 1, MaddOpc: Opc, RC); |
7371 | break; |
7372 | case AArch64MachineCombinerPattern::MULADDv2i32_indexed_OP2: |
7373 | Opc = AArch64::MLAv2i32_indexed; |
7374 | RC = &AArch64::FPR64RegClass; |
7375 | MUL = genFusedMultiplyIdx(MF, MRI, TII, Root, InsInstrs, IdxMulOpd: 2, MaddOpc: Opc, RC); |
7376 | break; |
7377 | case AArch64MachineCombinerPattern::MULADDv4i32_indexed_OP1: |
7378 | Opc = AArch64::MLAv4i32_indexed; |
7379 | RC = &AArch64::FPR128RegClass; |
7380 | MUL = genFusedMultiplyIdx(MF, MRI, TII, Root, InsInstrs, IdxMulOpd: 1, MaddOpc: Opc, RC); |
7381 | break; |
7382 | case AArch64MachineCombinerPattern::MULADDv4i32_indexed_OP2: |
7383 | Opc = AArch64::MLAv4i32_indexed; |
7384 | RC = &AArch64::FPR128RegClass; |
7385 | MUL = genFusedMultiplyIdx(MF, MRI, TII, Root, InsInstrs, IdxMulOpd: 2, MaddOpc: Opc, RC); |
7386 | break; |
7387 | |
7388 | case AArch64MachineCombinerPattern::MULSUBv4i16_indexed_OP1: |
7389 | Opc = AArch64::MLAv4i16_indexed; |
7390 | RC = &AArch64::FPR64RegClass; |
7391 | MUL = genFusedMultiplyIdxNeg(MF, MRI, TII, Root, InsInstrs, |
7392 | InstrIdxForVirtReg, IdxMulOpd: 1, MaddOpc: Opc, MnegOpc: AArch64::NEGv4i16, |
7393 | RC); |
7394 | break; |
7395 | case AArch64MachineCombinerPattern::MULSUBv4i16_indexed_OP2: |
7396 | Opc = AArch64::MLSv4i16_indexed; |
7397 | RC = &AArch64::FPR64RegClass; |
7398 | MUL = genFusedMultiplyIdx(MF, MRI, TII, Root, InsInstrs, IdxMulOpd: 2, MaddOpc: Opc, RC); |
7399 | break; |
7400 | case AArch64MachineCombinerPattern::MULSUBv8i16_indexed_OP1: |
7401 | Opc = AArch64::MLAv8i16_indexed; |
7402 | RC = &AArch64::FPR128RegClass; |
7403 | MUL = genFusedMultiplyIdxNeg(MF, MRI, TII, Root, InsInstrs, |
7404 | InstrIdxForVirtReg, IdxMulOpd: 1, MaddOpc: Opc, MnegOpc: AArch64::NEGv8i16, |
7405 | RC); |
7406 | break; |
7407 | case AArch64MachineCombinerPattern::MULSUBv8i16_indexed_OP2: |
7408 | Opc = AArch64::MLSv8i16_indexed; |
7409 | RC = &AArch64::FPR128RegClass; |
7410 | MUL = genFusedMultiplyIdx(MF, MRI, TII, Root, InsInstrs, IdxMulOpd: 2, MaddOpc: Opc, RC); |
7411 | break; |
7412 | case AArch64MachineCombinerPattern::MULSUBv2i32_indexed_OP1: |
7413 | Opc = AArch64::MLAv2i32_indexed; |
7414 | RC = &AArch64::FPR64RegClass; |
7415 | MUL = genFusedMultiplyIdxNeg(MF, MRI, TII, Root, InsInstrs, |
7416 | InstrIdxForVirtReg, IdxMulOpd: 1, MaddOpc: Opc, MnegOpc: AArch64::NEGv2i32, |
7417 | RC); |
7418 | break; |
7419 | case AArch64MachineCombinerPattern::MULSUBv2i32_indexed_OP2: |
7420 | Opc = AArch64::MLSv2i32_indexed; |
7421 | RC = &AArch64::FPR64RegClass; |
7422 | MUL = genFusedMultiplyIdx(MF, MRI, TII, Root, InsInstrs, IdxMulOpd: 2, MaddOpc: Opc, RC); |
7423 | break; |
7424 | case AArch64MachineCombinerPattern::MULSUBv4i32_indexed_OP1: |
7425 | Opc = AArch64::MLAv4i32_indexed; |
7426 | RC = &AArch64::FPR128RegClass; |
7427 | MUL = genFusedMultiplyIdxNeg(MF, MRI, TII, Root, InsInstrs, |
7428 | InstrIdxForVirtReg, IdxMulOpd: 1, MaddOpc: Opc, MnegOpc: AArch64::NEGv4i32, |
7429 | RC); |
7430 | break; |
7431 | case AArch64MachineCombinerPattern::MULSUBv4i32_indexed_OP2: |
7432 | Opc = AArch64::MLSv4i32_indexed; |
7433 | RC = &AArch64::FPR128RegClass; |
7434 | MUL = genFusedMultiplyIdx(MF, MRI, TII, Root, InsInstrs, IdxMulOpd: 2, MaddOpc: Opc, RC); |
7435 | break; |
7436 | |
7437 | // Floating Point Support |
7438 | case AArch64MachineCombinerPattern::FMULADDH_OP1: |
7439 | Opc = AArch64::FMADDHrrr; |
7440 | RC = &AArch64::FPR16RegClass; |
7441 | MUL = genFusedMultiply(MF, MRI, TII, Root, InsInstrs, IdxMulOpd: 1, MaddOpc: Opc, RC); |
7442 | break; |
7443 | case AArch64MachineCombinerPattern::FMULADDS_OP1: |
7444 | Opc = AArch64::FMADDSrrr; |
7445 | RC = &AArch64::FPR32RegClass; |
7446 | MUL = genFusedMultiply(MF, MRI, TII, Root, InsInstrs, IdxMulOpd: 1, MaddOpc: Opc, RC); |
7447 | break; |
7448 | case AArch64MachineCombinerPattern::FMULADDD_OP1: |
7449 | Opc = AArch64::FMADDDrrr; |
7450 | RC = &AArch64::FPR64RegClass; |
7451 | MUL = genFusedMultiply(MF, MRI, TII, Root, InsInstrs, IdxMulOpd: 1, MaddOpc: Opc, RC); |
7452 | break; |
7453 | |
7454 | case AArch64MachineCombinerPattern::FMULADDH_OP2: |
7455 | Opc = AArch64::FMADDHrrr; |
7456 | RC = &AArch64::FPR16RegClass; |
7457 | MUL = genFusedMultiply(MF, MRI, TII, Root, InsInstrs, IdxMulOpd: 2, MaddOpc: Opc, RC); |
7458 | break; |
7459 | case AArch64MachineCombinerPattern::FMULADDS_OP2: |
7460 | Opc = AArch64::FMADDSrrr; |
7461 | RC = &AArch64::FPR32RegClass; |
7462 | MUL = genFusedMultiply(MF, MRI, TII, Root, InsInstrs, IdxMulOpd: 2, MaddOpc: Opc, RC); |
7463 | break; |
7464 | case AArch64MachineCombinerPattern::FMULADDD_OP2: |
7465 | Opc = AArch64::FMADDDrrr; |
7466 | RC = &AArch64::FPR64RegClass; |
7467 | MUL = genFusedMultiply(MF, MRI, TII, Root, InsInstrs, IdxMulOpd: 2, MaddOpc: Opc, RC); |
7468 | break; |
7469 | |
7470 | case AArch64MachineCombinerPattern::FMLAv1i32_indexed_OP1: |
7471 | Opc = AArch64::FMLAv1i32_indexed; |
7472 | RC = &AArch64::FPR32RegClass; |
7473 | MUL = genFusedMultiply(MF, MRI, TII, Root, InsInstrs, IdxMulOpd: 1, MaddOpc: Opc, RC, |
7474 | kind: FMAInstKind::Indexed); |
7475 | break; |
7476 | case AArch64MachineCombinerPattern::FMLAv1i32_indexed_OP2: |
7477 | Opc = AArch64::FMLAv1i32_indexed; |
7478 | RC = &AArch64::FPR32RegClass; |
7479 | MUL = genFusedMultiply(MF, MRI, TII, Root, InsInstrs, IdxMulOpd: 2, MaddOpc: Opc, RC, |
7480 | kind: FMAInstKind::Indexed); |
7481 | break; |
7482 | |
7483 | case AArch64MachineCombinerPattern::FMLAv1i64_indexed_OP1: |
7484 | Opc = AArch64::FMLAv1i64_indexed; |
7485 | RC = &AArch64::FPR64RegClass; |
7486 | MUL = genFusedMultiply(MF, MRI, TII, Root, InsInstrs, IdxMulOpd: 1, MaddOpc: Opc, RC, |
7487 | kind: FMAInstKind::Indexed); |
7488 | break; |
7489 | case AArch64MachineCombinerPattern::FMLAv1i64_indexed_OP2: |
7490 | Opc = AArch64::FMLAv1i64_indexed; |
7491 | RC = &AArch64::FPR64RegClass; |
7492 | MUL = genFusedMultiply(MF, MRI, TII, Root, InsInstrs, IdxMulOpd: 2, MaddOpc: Opc, RC, |
7493 | kind: FMAInstKind::Indexed); |
7494 | break; |
7495 | |
7496 | case AArch64MachineCombinerPattern::FMLAv4i16_indexed_OP1: |
7497 | RC = &AArch64::FPR64RegClass; |
7498 | Opc = AArch64::FMLAv4i16_indexed; |
7499 | MUL = genFusedMultiply(MF, MRI, TII, Root, InsInstrs, IdxMulOpd: 1, MaddOpc: Opc, RC, |
7500 | kind: FMAInstKind::Indexed); |
7501 | break; |
7502 | case AArch64MachineCombinerPattern::FMLAv4f16_OP1: |
7503 | RC = &AArch64::FPR64RegClass; |
7504 | Opc = AArch64::FMLAv4f16; |
7505 | MUL = genFusedMultiply(MF, MRI, TII, Root, InsInstrs, IdxMulOpd: 1, MaddOpc: Opc, RC, |
7506 | kind: FMAInstKind::Accumulator); |
7507 | break; |
7508 | case AArch64MachineCombinerPattern::FMLAv4i16_indexed_OP2: |
7509 | RC = &AArch64::FPR64RegClass; |
7510 | Opc = AArch64::FMLAv4i16_indexed; |
7511 | MUL = genFusedMultiply(MF, MRI, TII, Root, InsInstrs, IdxMulOpd: 2, MaddOpc: Opc, RC, |
7512 | kind: FMAInstKind::Indexed); |
7513 | break; |
7514 | case AArch64MachineCombinerPattern::FMLAv4f16_OP2: |
7515 | RC = &AArch64::FPR64RegClass; |
7516 | Opc = AArch64::FMLAv4f16; |
7517 | MUL = genFusedMultiply(MF, MRI, TII, Root, InsInstrs, IdxMulOpd: 2, MaddOpc: Opc, RC, |
7518 | kind: FMAInstKind::Accumulator); |
7519 | break; |
7520 | |
7521 | case AArch64MachineCombinerPattern::FMLAv2i32_indexed_OP1: |
7522 | case AArch64MachineCombinerPattern::FMLAv2f32_OP1: |
7523 | RC = &AArch64::FPR64RegClass; |
7524 | if (Pattern == AArch64MachineCombinerPattern::FMLAv2i32_indexed_OP1) { |
7525 | Opc = AArch64::FMLAv2i32_indexed; |
7526 | MUL = genFusedMultiply(MF, MRI, TII, Root, InsInstrs, IdxMulOpd: 1, MaddOpc: Opc, RC, |
7527 | kind: FMAInstKind::Indexed); |
7528 | } else { |
7529 | Opc = AArch64::FMLAv2f32; |
7530 | MUL = genFusedMultiply(MF, MRI, TII, Root, InsInstrs, IdxMulOpd: 1, MaddOpc: Opc, RC, |
7531 | kind: FMAInstKind::Accumulator); |
7532 | } |
7533 | break; |
7534 | case AArch64MachineCombinerPattern::FMLAv2i32_indexed_OP2: |
7535 | case AArch64MachineCombinerPattern::FMLAv2f32_OP2: |
7536 | RC = &AArch64::FPR64RegClass; |
7537 | if (Pattern == AArch64MachineCombinerPattern::FMLAv2i32_indexed_OP2) { |
7538 | Opc = AArch64::FMLAv2i32_indexed; |
7539 | MUL = genFusedMultiply(MF, MRI, TII, Root, InsInstrs, IdxMulOpd: 2, MaddOpc: Opc, RC, |
7540 | kind: FMAInstKind::Indexed); |
7541 | } else { |
7542 | Opc = AArch64::FMLAv2f32; |
7543 | MUL = genFusedMultiply(MF, MRI, TII, Root, InsInstrs, IdxMulOpd: 2, MaddOpc: Opc, RC, |
7544 | kind: FMAInstKind::Accumulator); |
7545 | } |
7546 | break; |
7547 | |
7548 | case AArch64MachineCombinerPattern::FMLAv8i16_indexed_OP1: |
7549 | RC = &AArch64::FPR128RegClass; |
7550 | Opc = AArch64::FMLAv8i16_indexed; |
7551 | MUL = genFusedMultiply(MF, MRI, TII, Root, InsInstrs, IdxMulOpd: 1, MaddOpc: Opc, RC, |
7552 | kind: FMAInstKind::Indexed); |
7553 | break; |
7554 | case AArch64MachineCombinerPattern::FMLAv8f16_OP1: |
7555 | RC = &AArch64::FPR128RegClass; |
7556 | Opc = AArch64::FMLAv8f16; |
7557 | MUL = genFusedMultiply(MF, MRI, TII, Root, InsInstrs, IdxMulOpd: 1, MaddOpc: Opc, RC, |
7558 | kind: FMAInstKind::Accumulator); |
7559 | break; |
7560 | case AArch64MachineCombinerPattern::FMLAv8i16_indexed_OP2: |
7561 | RC = &AArch64::FPR128RegClass; |
7562 | Opc = AArch64::FMLAv8i16_indexed; |
7563 | MUL = genFusedMultiply(MF, MRI, TII, Root, InsInstrs, IdxMulOpd: 2, MaddOpc: Opc, RC, |
7564 | kind: FMAInstKind::Indexed); |
7565 | break; |
7566 | case AArch64MachineCombinerPattern::FMLAv8f16_OP2: |
7567 | RC = &AArch64::FPR128RegClass; |
7568 | Opc = AArch64::FMLAv8f16; |
7569 | MUL = genFusedMultiply(MF, MRI, TII, Root, InsInstrs, IdxMulOpd: 2, MaddOpc: Opc, RC, |
7570 | kind: FMAInstKind::Accumulator); |
7571 | break; |
7572 | |
7573 | case AArch64MachineCombinerPattern::FMLAv2i64_indexed_OP1: |
7574 | case AArch64MachineCombinerPattern::FMLAv2f64_OP1: |
7575 | RC = &AArch64::FPR128RegClass; |
7576 | if (Pattern == AArch64MachineCombinerPattern::FMLAv2i64_indexed_OP1) { |
7577 | Opc = AArch64::FMLAv2i64_indexed; |
7578 | MUL = genFusedMultiply(MF, MRI, TII, Root, InsInstrs, IdxMulOpd: 1, MaddOpc: Opc, RC, |
7579 | kind: FMAInstKind::Indexed); |
7580 | } else { |
7581 | Opc = AArch64::FMLAv2f64; |
7582 | MUL = genFusedMultiply(MF, MRI, TII, Root, InsInstrs, IdxMulOpd: 1, MaddOpc: Opc, RC, |
7583 | kind: FMAInstKind::Accumulator); |
7584 | } |
7585 | break; |
7586 | case AArch64MachineCombinerPattern::FMLAv2i64_indexed_OP2: |
7587 | case AArch64MachineCombinerPattern::FMLAv2f64_OP2: |
7588 | RC = &AArch64::FPR128RegClass; |
7589 | if (Pattern == AArch64MachineCombinerPattern::FMLAv2i64_indexed_OP2) { |
7590 | Opc = AArch64::FMLAv2i64_indexed; |
7591 | MUL = genFusedMultiply(MF, MRI, TII, Root, InsInstrs, IdxMulOpd: 2, MaddOpc: Opc, RC, |
7592 | kind: FMAInstKind::Indexed); |
7593 | } else { |
7594 | Opc = AArch64::FMLAv2f64; |
7595 | MUL = genFusedMultiply(MF, MRI, TII, Root, InsInstrs, IdxMulOpd: 2, MaddOpc: Opc, RC, |
7596 | kind: FMAInstKind::Accumulator); |
7597 | } |
7598 | break; |
7599 | |
7600 | case AArch64MachineCombinerPattern::FMLAv4i32_indexed_OP1: |
7601 | case AArch64MachineCombinerPattern::FMLAv4f32_OP1: |
7602 | RC = &AArch64::FPR128RegClass; |
7603 | if (Pattern == AArch64MachineCombinerPattern::FMLAv4i32_indexed_OP1) { |
7604 | Opc = AArch64::FMLAv4i32_indexed; |
7605 | MUL = genFusedMultiply(MF, MRI, TII, Root, InsInstrs, IdxMulOpd: 1, MaddOpc: Opc, RC, |
7606 | kind: FMAInstKind::Indexed); |
7607 | } else { |
7608 | Opc = AArch64::FMLAv4f32; |
7609 | MUL = genFusedMultiply(MF, MRI, TII, Root, InsInstrs, IdxMulOpd: 1, MaddOpc: Opc, RC, |
7610 | kind: FMAInstKind::Accumulator); |
7611 | } |
7612 | break; |
7613 | |
7614 | case AArch64MachineCombinerPattern::FMLAv4i32_indexed_OP2: |
7615 | case AArch64MachineCombinerPattern::FMLAv4f32_OP2: |
7616 | RC = &AArch64::FPR128RegClass; |
7617 | if (Pattern == AArch64MachineCombinerPattern::FMLAv4i32_indexed_OP2) { |
7618 | Opc = AArch64::FMLAv4i32_indexed; |
7619 | MUL = genFusedMultiply(MF, MRI, TII, Root, InsInstrs, IdxMulOpd: 2, MaddOpc: Opc, RC, |
7620 | kind: FMAInstKind::Indexed); |
7621 | } else { |
7622 | Opc = AArch64::FMLAv4f32; |
7623 | MUL = genFusedMultiply(MF, MRI, TII, Root, InsInstrs, IdxMulOpd: 2, MaddOpc: Opc, RC, |
7624 | kind: FMAInstKind::Accumulator); |
7625 | } |
7626 | break; |
7627 | |
7628 | case AArch64MachineCombinerPattern::FMULSUBH_OP1: |
7629 | Opc = AArch64::FNMSUBHrrr; |
7630 | RC = &AArch64::FPR16RegClass; |
7631 | MUL = genFusedMultiply(MF, MRI, TII, Root, InsInstrs, IdxMulOpd: 1, MaddOpc: Opc, RC); |
7632 | break; |
7633 | case AArch64MachineCombinerPattern::FMULSUBS_OP1: |
7634 | Opc = AArch64::FNMSUBSrrr; |
7635 | RC = &AArch64::FPR32RegClass; |
7636 | MUL = genFusedMultiply(MF, MRI, TII, Root, InsInstrs, IdxMulOpd: 1, MaddOpc: Opc, RC); |
7637 | break; |
7638 | case AArch64MachineCombinerPattern::FMULSUBD_OP1: |
7639 | Opc = AArch64::FNMSUBDrrr; |
7640 | RC = &AArch64::FPR64RegClass; |
7641 | MUL = genFusedMultiply(MF, MRI, TII, Root, InsInstrs, IdxMulOpd: 1, MaddOpc: Opc, RC); |
7642 | break; |
7643 | |
7644 | case AArch64MachineCombinerPattern::FNMULSUBH_OP1: |
7645 | Opc = AArch64::FNMADDHrrr; |
7646 | RC = &AArch64::FPR16RegClass; |
7647 | MUL = genFusedMultiply(MF, MRI, TII, Root, InsInstrs, IdxMulOpd: 1, MaddOpc: Opc, RC); |
7648 | break; |
7649 | case AArch64MachineCombinerPattern::FNMULSUBS_OP1: |
7650 | Opc = AArch64::FNMADDSrrr; |
7651 | RC = &AArch64::FPR32RegClass; |
7652 | MUL = genFusedMultiply(MF, MRI, TII, Root, InsInstrs, IdxMulOpd: 1, MaddOpc: Opc, RC); |
7653 | break; |
7654 | case AArch64MachineCombinerPattern::FNMULSUBD_OP1: |
7655 | Opc = AArch64::FNMADDDrrr; |
7656 | RC = &AArch64::FPR64RegClass; |
7657 | MUL = genFusedMultiply(MF, MRI, TII, Root, InsInstrs, IdxMulOpd: 1, MaddOpc: Opc, RC); |
7658 | break; |
7659 | |
7660 | case AArch64MachineCombinerPattern::FMULSUBH_OP2: |
7661 | Opc = AArch64::FMSUBHrrr; |
7662 | RC = &AArch64::FPR16RegClass; |
7663 | MUL = genFusedMultiply(MF, MRI, TII, Root, InsInstrs, IdxMulOpd: 2, MaddOpc: Opc, RC); |
7664 | break; |
7665 | case AArch64MachineCombinerPattern::FMULSUBS_OP2: |
7666 | Opc = AArch64::FMSUBSrrr; |
7667 | RC = &AArch64::FPR32RegClass; |
7668 | MUL = genFusedMultiply(MF, MRI, TII, Root, InsInstrs, IdxMulOpd: 2, MaddOpc: Opc, RC); |
7669 | break; |
7670 | case AArch64MachineCombinerPattern::FMULSUBD_OP2: |
7671 | Opc = AArch64::FMSUBDrrr; |
7672 | RC = &AArch64::FPR64RegClass; |
7673 | MUL = genFusedMultiply(MF, MRI, TII, Root, InsInstrs, IdxMulOpd: 2, MaddOpc: Opc, RC); |
7674 | break; |
7675 | |
7676 | case AArch64MachineCombinerPattern::FMLSv1i32_indexed_OP2: |
7677 | Opc = AArch64::FMLSv1i32_indexed; |
7678 | RC = &AArch64::FPR32RegClass; |
7679 | MUL = genFusedMultiply(MF, MRI, TII, Root, InsInstrs, IdxMulOpd: 2, MaddOpc: Opc, RC, |
7680 | kind: FMAInstKind::Indexed); |
7681 | break; |
7682 | |
7683 | case AArch64MachineCombinerPattern::FMLSv1i64_indexed_OP2: |
7684 | Opc = AArch64::FMLSv1i64_indexed; |
7685 | RC = &AArch64::FPR64RegClass; |
7686 | MUL = genFusedMultiply(MF, MRI, TII, Root, InsInstrs, IdxMulOpd: 2, MaddOpc: Opc, RC, |
7687 | kind: FMAInstKind::Indexed); |
7688 | break; |
7689 | |
7690 | case AArch64MachineCombinerPattern::FMLSv4f16_OP1: |
7691 | case AArch64MachineCombinerPattern::FMLSv4i16_indexed_OP1: { |
7692 | RC = &AArch64::FPR64RegClass; |
7693 | Register NewVR = MRI.createVirtualRegister(RegClass: RC); |
7694 | MachineInstrBuilder MIB1 = |
7695 | BuildMI(MF, MIMD: MIMetadata(Root), MCID: TII->get(Opcode: AArch64::FNEGv4f16), DestReg: NewVR) |
7696 | .add(MO: Root.getOperand(i: 2)); |
7697 | InsInstrs.push_back(Elt: MIB1); |
7698 | InstrIdxForVirtReg.insert(KV: std::make_pair(x&: NewVR, y: 0)); |
7699 | if (Pattern == AArch64MachineCombinerPattern::FMLSv4f16_OP1) { |
7700 | Opc = AArch64::FMLAv4f16; |
7701 | MUL = genFusedMultiply(MF, MRI, TII, Root, InsInstrs, IdxMulOpd: 1, MaddOpc: Opc, RC, |
7702 | kind: FMAInstKind::Accumulator, ReplacedAddend: &NewVR); |
7703 | } else { |
7704 | Opc = AArch64::FMLAv4i16_indexed; |
7705 | MUL = genFusedMultiply(MF, MRI, TII, Root, InsInstrs, IdxMulOpd: 1, MaddOpc: Opc, RC, |
7706 | kind: FMAInstKind::Indexed, ReplacedAddend: &NewVR); |
7707 | } |
7708 | break; |
7709 | } |
7710 | case AArch64MachineCombinerPattern::FMLSv4f16_OP2: |
7711 | RC = &AArch64::FPR64RegClass; |
7712 | Opc = AArch64::FMLSv4f16; |
7713 | MUL = genFusedMultiply(MF, MRI, TII, Root, InsInstrs, IdxMulOpd: 2, MaddOpc: Opc, RC, |
7714 | kind: FMAInstKind::Accumulator); |
7715 | break; |
7716 | case AArch64MachineCombinerPattern::FMLSv4i16_indexed_OP2: |
7717 | RC = &AArch64::FPR64RegClass; |
7718 | Opc = AArch64::FMLSv4i16_indexed; |
7719 | MUL = genFusedMultiply(MF, MRI, TII, Root, InsInstrs, IdxMulOpd: 2, MaddOpc: Opc, RC, |
7720 | kind: FMAInstKind::Indexed); |
7721 | break; |
7722 | |
7723 | case AArch64MachineCombinerPattern::FMLSv2f32_OP2: |
7724 | case AArch64MachineCombinerPattern::FMLSv2i32_indexed_OP2: |
7725 | RC = &AArch64::FPR64RegClass; |
7726 | if (Pattern == AArch64MachineCombinerPattern::FMLSv2i32_indexed_OP2) { |
7727 | Opc = AArch64::FMLSv2i32_indexed; |
7728 | MUL = genFusedMultiply(MF, MRI, TII, Root, InsInstrs, IdxMulOpd: 2, MaddOpc: Opc, RC, |
7729 | kind: FMAInstKind::Indexed); |
7730 | } else { |
7731 | Opc = AArch64::FMLSv2f32; |
7732 | MUL = genFusedMultiply(MF, MRI, TII, Root, InsInstrs, IdxMulOpd: 2, MaddOpc: Opc, RC, |
7733 | kind: FMAInstKind::Accumulator); |
7734 | } |
7735 | break; |
7736 | |
7737 | case AArch64MachineCombinerPattern::FMLSv8f16_OP1: |
7738 | case AArch64MachineCombinerPattern::FMLSv8i16_indexed_OP1: { |
7739 | RC = &AArch64::FPR128RegClass; |
7740 | Register NewVR = MRI.createVirtualRegister(RegClass: RC); |
7741 | MachineInstrBuilder MIB1 = |
7742 | BuildMI(MF, MIMD: MIMetadata(Root), MCID: TII->get(Opcode: AArch64::FNEGv8f16), DestReg: NewVR) |
7743 | .add(MO: Root.getOperand(i: 2)); |
7744 | InsInstrs.push_back(Elt: MIB1); |
7745 | InstrIdxForVirtReg.insert(KV: std::make_pair(x&: NewVR, y: 0)); |
7746 | if (Pattern == AArch64MachineCombinerPattern::FMLSv8f16_OP1) { |
7747 | Opc = AArch64::FMLAv8f16; |
7748 | MUL = genFusedMultiply(MF, MRI, TII, Root, InsInstrs, IdxMulOpd: 1, MaddOpc: Opc, RC, |
7749 | kind: FMAInstKind::Accumulator, ReplacedAddend: &NewVR); |
7750 | } else { |
7751 | Opc = AArch64::FMLAv8i16_indexed; |
7752 | MUL = genFusedMultiply(MF, MRI, TII, Root, InsInstrs, IdxMulOpd: 1, MaddOpc: Opc, RC, |
7753 | kind: FMAInstKind::Indexed, ReplacedAddend: &NewVR); |
7754 | } |
7755 | break; |
7756 | } |
7757 | case AArch64MachineCombinerPattern::FMLSv8f16_OP2: |
7758 | RC = &AArch64::FPR128RegClass; |
7759 | Opc = AArch64::FMLSv8f16; |
7760 | MUL = genFusedMultiply(MF, MRI, TII, Root, InsInstrs, IdxMulOpd: 2, MaddOpc: Opc, RC, |
7761 | kind: FMAInstKind::Accumulator); |
7762 | break; |
7763 | case AArch64MachineCombinerPattern::FMLSv8i16_indexed_OP2: |
7764 | RC = &AArch64::FPR128RegClass; |
7765 | Opc = AArch64::FMLSv8i16_indexed; |
7766 | MUL = genFusedMultiply(MF, MRI, TII, Root, InsInstrs, IdxMulOpd: 2, MaddOpc: Opc, RC, |
7767 | kind: FMAInstKind::Indexed); |
7768 | break; |
7769 | |
7770 | case AArch64MachineCombinerPattern::FMLSv2f64_OP2: |
7771 | case AArch64MachineCombinerPattern::FMLSv2i64_indexed_OP2: |
7772 | RC = &AArch64::FPR128RegClass; |
7773 | if (Pattern == AArch64MachineCombinerPattern::FMLSv2i64_indexed_OP2) { |
7774 | Opc = AArch64::FMLSv2i64_indexed; |
7775 | MUL = genFusedMultiply(MF, MRI, TII, Root, InsInstrs, IdxMulOpd: 2, MaddOpc: Opc, RC, |
7776 | kind: FMAInstKind::Indexed); |
7777 | } else { |
7778 | Opc = AArch64::FMLSv2f64; |
7779 | MUL = genFusedMultiply(MF, MRI, TII, Root, InsInstrs, IdxMulOpd: 2, MaddOpc: Opc, RC, |
7780 | kind: FMAInstKind::Accumulator); |
7781 | } |
7782 | break; |
7783 | |
7784 | case AArch64MachineCombinerPattern::FMLSv4f32_OP2: |
7785 | case AArch64MachineCombinerPattern::FMLSv4i32_indexed_OP2: |
7786 | RC = &AArch64::FPR128RegClass; |
7787 | if (Pattern == AArch64MachineCombinerPattern::FMLSv4i32_indexed_OP2) { |
7788 | Opc = AArch64::FMLSv4i32_indexed; |
7789 | MUL = genFusedMultiply(MF, MRI, TII, Root, InsInstrs, IdxMulOpd: 2, MaddOpc: Opc, RC, |
7790 | kind: FMAInstKind::Indexed); |
7791 | } else { |
7792 | Opc = AArch64::FMLSv4f32; |
7793 | MUL = genFusedMultiply(MF, MRI, TII, Root, InsInstrs, IdxMulOpd: 2, MaddOpc: Opc, RC, |
7794 | kind: FMAInstKind::Accumulator); |
7795 | } |
7796 | break; |
7797 | case AArch64MachineCombinerPattern::FMLSv2f32_OP1: |
7798 | case AArch64MachineCombinerPattern::FMLSv2i32_indexed_OP1: { |
7799 | RC = &AArch64::FPR64RegClass; |
7800 | Register NewVR = MRI.createVirtualRegister(RegClass: RC); |
7801 | MachineInstrBuilder MIB1 = |
7802 | BuildMI(MF, MIMD: MIMetadata(Root), MCID: TII->get(Opcode: AArch64::FNEGv2f32), DestReg: NewVR) |
7803 | .add(MO: Root.getOperand(i: 2)); |
7804 | InsInstrs.push_back(Elt: MIB1); |
7805 | InstrIdxForVirtReg.insert(KV: std::make_pair(x&: NewVR, y: 0)); |
7806 | if (Pattern == AArch64MachineCombinerPattern::FMLSv2i32_indexed_OP1) { |
7807 | Opc = AArch64::FMLAv2i32_indexed; |
7808 | MUL = genFusedMultiply(MF, MRI, TII, Root, InsInstrs, IdxMulOpd: 1, MaddOpc: Opc, RC, |
7809 | kind: FMAInstKind::Indexed, ReplacedAddend: &NewVR); |
7810 | } else { |
7811 | Opc = AArch64::FMLAv2f32; |
7812 | MUL = genFusedMultiply(MF, MRI, TII, Root, InsInstrs, IdxMulOpd: 1, MaddOpc: Opc, RC, |
7813 | kind: FMAInstKind::Accumulator, ReplacedAddend: &NewVR); |
7814 | } |
7815 | break; |
7816 | } |
7817 | case AArch64MachineCombinerPattern::FMLSv4f32_OP1: |
7818 | case AArch64MachineCombinerPattern::FMLSv4i32_indexed_OP1: { |
7819 | RC = &AArch64::FPR128RegClass; |
7820 | Register NewVR = MRI.createVirtualRegister(RegClass: RC); |
7821 | MachineInstrBuilder MIB1 = |
7822 | BuildMI(MF, MIMD: MIMetadata(Root), MCID: TII->get(Opcode: AArch64::FNEGv4f32), DestReg: NewVR) |
7823 | .add(MO: Root.getOperand(i: 2)); |
7824 | InsInstrs.push_back(Elt: MIB1); |
7825 | InstrIdxForVirtReg.insert(KV: std::make_pair(x&: NewVR, y: 0)); |
7826 | if (Pattern == AArch64MachineCombinerPattern::FMLSv4i32_indexed_OP1) { |
7827 | Opc = AArch64::FMLAv4i32_indexed; |
7828 | MUL = genFusedMultiply(MF, MRI, TII, Root, InsInstrs, IdxMulOpd: 1, MaddOpc: Opc, RC, |
7829 | kind: FMAInstKind::Indexed, ReplacedAddend: &NewVR); |
7830 | } else { |
7831 | Opc = AArch64::FMLAv4f32; |
7832 | MUL = genFusedMultiply(MF, MRI, TII, Root, InsInstrs, IdxMulOpd: 1, MaddOpc: Opc, RC, |
7833 | kind: FMAInstKind::Accumulator, ReplacedAddend: &NewVR); |
7834 | } |
7835 | break; |
7836 | } |
7837 | case AArch64MachineCombinerPattern::FMLSv2f64_OP1: |
7838 | case AArch64MachineCombinerPattern::FMLSv2i64_indexed_OP1: { |
7839 | RC = &AArch64::FPR128RegClass; |
7840 | Register NewVR = MRI.createVirtualRegister(RegClass: RC); |
7841 | MachineInstrBuilder MIB1 = |
7842 | BuildMI(MF, MIMD: MIMetadata(Root), MCID: TII->get(Opcode: AArch64::FNEGv2f64), DestReg: NewVR) |
7843 | .add(MO: Root.getOperand(i: 2)); |
7844 | InsInstrs.push_back(Elt: MIB1); |
7845 | InstrIdxForVirtReg.insert(KV: std::make_pair(x&: NewVR, y: 0)); |
7846 | if (Pattern == AArch64MachineCombinerPattern::FMLSv2i64_indexed_OP1) { |
7847 | Opc = AArch64::FMLAv2i64_indexed; |
7848 | MUL = genFusedMultiply(MF, MRI, TII, Root, InsInstrs, IdxMulOpd: 1, MaddOpc: Opc, RC, |
7849 | kind: FMAInstKind::Indexed, ReplacedAddend: &NewVR); |
7850 | } else { |
7851 | Opc = AArch64::FMLAv2f64; |
7852 | MUL = genFusedMultiply(MF, MRI, TII, Root, InsInstrs, IdxMulOpd: 1, MaddOpc: Opc, RC, |
7853 | kind: FMAInstKind::Accumulator, ReplacedAddend: &NewVR); |
7854 | } |
7855 | break; |
7856 | } |
7857 | case AArch64MachineCombinerPattern::FMULv2i32_indexed_OP1: |
7858 | case AArch64MachineCombinerPattern::FMULv2i32_indexed_OP2: { |
7859 | unsigned IdxDupOp = |
7860 | (Pattern == AArch64MachineCombinerPattern::FMULv2i32_indexed_OP1) ? 1 |
7861 | : 2; |
7862 | genIndexedMultiply(Root, InsInstrs, IdxDupOp, MulOpc: AArch64::FMULv2i32_indexed, |
7863 | RC: &AArch64::FPR128RegClass, MRI); |
7864 | break; |
7865 | } |
7866 | case AArch64MachineCombinerPattern::FMULv2i64_indexed_OP1: |
7867 | case AArch64MachineCombinerPattern::FMULv2i64_indexed_OP2: { |
7868 | unsigned IdxDupOp = |
7869 | (Pattern == AArch64MachineCombinerPattern::FMULv2i64_indexed_OP1) ? 1 |
7870 | : 2; |
7871 | genIndexedMultiply(Root, InsInstrs, IdxDupOp, MulOpc: AArch64::FMULv2i64_indexed, |
7872 | RC: &AArch64::FPR128RegClass, MRI); |
7873 | break; |
7874 | } |
7875 | case AArch64MachineCombinerPattern::FMULv4i16_indexed_OP1: |
7876 | case AArch64MachineCombinerPattern::FMULv4i16_indexed_OP2: { |
7877 | unsigned IdxDupOp = |
7878 | (Pattern == AArch64MachineCombinerPattern::FMULv4i16_indexed_OP1) ? 1 |
7879 | : 2; |
7880 | genIndexedMultiply(Root, InsInstrs, IdxDupOp, MulOpc: AArch64::FMULv4i16_indexed, |
7881 | RC: &AArch64::FPR128_loRegClass, MRI); |
7882 | break; |
7883 | } |
7884 | case AArch64MachineCombinerPattern::FMULv4i32_indexed_OP1: |
7885 | case AArch64MachineCombinerPattern::FMULv4i32_indexed_OP2: { |
7886 | unsigned IdxDupOp = |
7887 | (Pattern == AArch64MachineCombinerPattern::FMULv4i32_indexed_OP1) ? 1 |
7888 | : 2; |
7889 | genIndexedMultiply(Root, InsInstrs, IdxDupOp, MulOpc: AArch64::FMULv4i32_indexed, |
7890 | RC: &AArch64::FPR128RegClass, MRI); |
7891 | break; |
7892 | } |
7893 | case AArch64MachineCombinerPattern::FMULv8i16_indexed_OP1: |
7894 | case AArch64MachineCombinerPattern::FMULv8i16_indexed_OP2: { |
7895 | unsigned IdxDupOp = |
7896 | (Pattern == AArch64MachineCombinerPattern::FMULv8i16_indexed_OP1) ? 1 |
7897 | : 2; |
7898 | genIndexedMultiply(Root, InsInstrs, IdxDupOp, MulOpc: AArch64::FMULv8i16_indexed, |
7899 | RC: &AArch64::FPR128_loRegClass, MRI); |
7900 | break; |
7901 | } |
7902 | case AArch64MachineCombinerPattern::FNMADD: { |
7903 | MUL = genFNegatedMAD(MF, MRI, TII, Root, InsInstrs); |
7904 | break; |
7905 | } |
7906 | |
7907 | } // end switch (Pattern) |
7908 | // Record MUL and ADD/SUB for deletion |
7909 | if (MUL) |
7910 | DelInstrs.push_back(Elt: MUL); |
7911 | DelInstrs.push_back(Elt: &Root); |
7912 | |
7913 | // Set the flags on the inserted instructions to be the merged flags of the |
7914 | // instructions that we have combined. |
7915 | uint32_t Flags = Root.getFlags(); |
7916 | if (MUL) |
7917 | Flags = Root.mergeFlagsWith(Other: *MUL); |
7918 | for (auto *MI : InsInstrs) |
7919 | MI->setFlags(Flags); |
7920 | } |
7921 | |
7922 | /// Replace csincr-branch sequence by simple conditional branch |
7923 | /// |
7924 | /// Examples: |
7925 | /// 1. \code |
7926 | /// csinc w9, wzr, wzr, <condition code> |
7927 | /// tbnz w9, #0, 0x44 |
7928 | /// \endcode |
7929 | /// to |
7930 | /// \code |
7931 | /// b.<inverted condition code> |
7932 | /// \endcode |
7933 | /// |
7934 | /// 2. \code |
7935 | /// csinc w9, wzr, wzr, <condition code> |
7936 | /// tbz w9, #0, 0x44 |
7937 | /// \endcode |
7938 | /// to |
7939 | /// \code |
7940 | /// b.<condition code> |
7941 | /// \endcode |
7942 | /// |
7943 | /// Replace compare and branch sequence by TBZ/TBNZ instruction when the |
7944 | /// compare's constant operand is power of 2. |
7945 | /// |
7946 | /// Examples: |
7947 | /// \code |
7948 | /// and w8, w8, #0x400 |
7949 | /// cbnz w8, L1 |
7950 | /// \endcode |
7951 | /// to |
7952 | /// \code |
7953 | /// tbnz w8, #10, L1 |
7954 | /// \endcode |
7955 | /// |
7956 | /// \param MI Conditional Branch |
7957 | /// \return True when the simple conditional branch is generated |
7958 | /// |
7959 | bool AArch64InstrInfo::optimizeCondBranch(MachineInstr &MI) const { |
7960 | bool IsNegativeBranch = false; |
7961 | bool IsTestAndBranch = false; |
7962 | unsigned TargetBBInMI = 0; |
7963 | switch (MI.getOpcode()) { |
7964 | default: |
7965 | llvm_unreachable("Unknown branch instruction?" ); |
7966 | case AArch64::Bcc: |
7967 | return false; |
7968 | case AArch64::CBZW: |
7969 | case AArch64::CBZX: |
7970 | TargetBBInMI = 1; |
7971 | break; |
7972 | case AArch64::CBNZW: |
7973 | case AArch64::CBNZX: |
7974 | TargetBBInMI = 1; |
7975 | IsNegativeBranch = true; |
7976 | break; |
7977 | case AArch64::TBZW: |
7978 | case AArch64::TBZX: |
7979 | TargetBBInMI = 2; |
7980 | IsTestAndBranch = true; |
7981 | break; |
7982 | case AArch64::TBNZW: |
7983 | case AArch64::TBNZX: |
7984 | TargetBBInMI = 2; |
7985 | IsNegativeBranch = true; |
7986 | IsTestAndBranch = true; |
7987 | break; |
7988 | } |
7989 | // So we increment a zero register and test for bits other |
7990 | // than bit 0? Conservatively bail out in case the verifier |
7991 | // missed this case. |
7992 | if (IsTestAndBranch && MI.getOperand(i: 1).getImm()) |
7993 | return false; |
7994 | |
7995 | // Find Definition. |
7996 | assert(MI.getParent() && "Incomplete machine instruciton\n" ); |
7997 | MachineBasicBlock *MBB = MI.getParent(); |
7998 | MachineFunction *MF = MBB->getParent(); |
7999 | MachineRegisterInfo *MRI = &MF->getRegInfo(); |
8000 | Register VReg = MI.getOperand(i: 0).getReg(); |
8001 | if (!VReg.isVirtual()) |
8002 | return false; |
8003 | |
8004 | MachineInstr *DefMI = MRI->getVRegDef(Reg: VReg); |
8005 | |
8006 | // Look through COPY instructions to find definition. |
8007 | while (DefMI->isCopy()) { |
8008 | Register CopyVReg = DefMI->getOperand(i: 1).getReg(); |
8009 | if (!MRI->hasOneNonDBGUse(RegNo: CopyVReg)) |
8010 | return false; |
8011 | if (!MRI->hasOneDef(RegNo: CopyVReg)) |
8012 | return false; |
8013 | DefMI = MRI->getVRegDef(Reg: CopyVReg); |
8014 | } |
8015 | |
8016 | switch (DefMI->getOpcode()) { |
8017 | default: |
8018 | return false; |
8019 | // Fold AND into a TBZ/TBNZ if constant operand is power of 2. |
8020 | case AArch64::ANDWri: |
8021 | case AArch64::ANDXri: { |
8022 | if (IsTestAndBranch) |
8023 | return false; |
8024 | if (DefMI->getParent() != MBB) |
8025 | return false; |
8026 | if (!MRI->hasOneNonDBGUse(RegNo: VReg)) |
8027 | return false; |
8028 | |
8029 | bool Is32Bit = (DefMI->getOpcode() == AArch64::ANDWri); |
8030 | uint64_t Mask = AArch64_AM::decodeLogicalImmediate( |
8031 | val: DefMI->getOperand(i: 2).getImm(), regSize: Is32Bit ? 32 : 64); |
8032 | if (!isPowerOf2_64(Value: Mask)) |
8033 | return false; |
8034 | |
8035 | MachineOperand &MO = DefMI->getOperand(i: 1); |
8036 | Register NewReg = MO.getReg(); |
8037 | if (!NewReg.isVirtual()) |
8038 | return false; |
8039 | |
8040 | assert(!MRI->def_empty(NewReg) && "Register must be defined." ); |
8041 | |
8042 | MachineBasicBlock &RefToMBB = *MBB; |
8043 | MachineBasicBlock *TBB = MI.getOperand(i: 1).getMBB(); |
8044 | DebugLoc DL = MI.getDebugLoc(); |
8045 | unsigned Imm = Log2_64(Value: Mask); |
8046 | unsigned Opc = (Imm < 32) |
8047 | ? (IsNegativeBranch ? AArch64::TBNZW : AArch64::TBZW) |
8048 | : (IsNegativeBranch ? AArch64::TBNZX : AArch64::TBZX); |
8049 | MachineInstr *NewMI = BuildMI(BB&: RefToMBB, I&: MI, MIMD: DL, MCID: get(Opcode: Opc)) |
8050 | .addReg(RegNo: NewReg) |
8051 | .addImm(Val: Imm) |
8052 | .addMBB(MBB: TBB); |
8053 | // Register lives on to the CBZ now. |
8054 | MO.setIsKill(false); |
8055 | |
8056 | // For immediate smaller than 32, we need to use the 32-bit |
8057 | // variant (W) in all cases. Indeed the 64-bit variant does not |
8058 | // allow to encode them. |
8059 | // Therefore, if the input register is 64-bit, we need to take the |
8060 | // 32-bit sub-part. |
8061 | if (!Is32Bit && Imm < 32) |
8062 | NewMI->getOperand(i: 0).setSubReg(AArch64::sub_32); |
8063 | MI.eraseFromParent(); |
8064 | return true; |
8065 | } |
8066 | // Look for CSINC |
8067 | case AArch64::CSINCWr: |
8068 | case AArch64::CSINCXr: { |
8069 | if (!(DefMI->getOperand(i: 1).getReg() == AArch64::WZR && |
8070 | DefMI->getOperand(i: 2).getReg() == AArch64::WZR) && |
8071 | !(DefMI->getOperand(i: 1).getReg() == AArch64::XZR && |
8072 | DefMI->getOperand(i: 2).getReg() == AArch64::XZR)) |
8073 | return false; |
8074 | |
8075 | if (DefMI->findRegisterDefOperandIdx(Reg: AArch64::NZCV, /*TRI=*/nullptr, |
8076 | isDead: true) != -1) |
8077 | return false; |
8078 | |
8079 | AArch64CC::CondCode CC = (AArch64CC::CondCode)DefMI->getOperand(i: 3).getImm(); |
8080 | // Convert only when the condition code is not modified between |
8081 | // the CSINC and the branch. The CC may be used by other |
8082 | // instructions in between. |
8083 | if (areCFlagsAccessedBetweenInstrs(From: DefMI, To: MI, TRI: &getRegisterInfo(), AccessToCheck: AK_Write)) |
8084 | return false; |
8085 | MachineBasicBlock &RefToMBB = *MBB; |
8086 | MachineBasicBlock *TBB = MI.getOperand(i: TargetBBInMI).getMBB(); |
8087 | DebugLoc DL = MI.getDebugLoc(); |
8088 | if (IsNegativeBranch) |
8089 | CC = AArch64CC::getInvertedCondCode(Code: CC); |
8090 | BuildMI(BB&: RefToMBB, I&: MI, MIMD: DL, MCID: get(Opcode: AArch64::Bcc)).addImm(Val: CC).addMBB(MBB: TBB); |
8091 | MI.eraseFromParent(); |
8092 | return true; |
8093 | } |
8094 | } |
8095 | } |
8096 | |
8097 | std::pair<unsigned, unsigned> |
8098 | AArch64InstrInfo::decomposeMachineOperandsTargetFlags(unsigned TF) const { |
8099 | const unsigned Mask = AArch64II::MO_FRAGMENT; |
8100 | return std::make_pair(x: TF & Mask, y: TF & ~Mask); |
8101 | } |
8102 | |
8103 | ArrayRef<std::pair<unsigned, const char *>> |
8104 | AArch64InstrInfo::getSerializableDirectMachineOperandTargetFlags() const { |
8105 | using namespace AArch64II; |
8106 | |
8107 | static const std::pair<unsigned, const char *> TargetFlags[] = { |
8108 | {MO_PAGE, "aarch64-page" }, {MO_PAGEOFF, "aarch64-pageoff" }, |
8109 | {MO_G3, "aarch64-g3" }, {MO_G2, "aarch64-g2" }, |
8110 | {MO_G1, "aarch64-g1" }, {MO_G0, "aarch64-g0" }, |
8111 | {MO_HI12, "aarch64-hi12" }}; |
8112 | return ArrayRef(TargetFlags); |
8113 | } |
8114 | |
8115 | ArrayRef<std::pair<unsigned, const char *>> |
8116 | AArch64InstrInfo::getSerializableBitmaskMachineOperandTargetFlags() const { |
8117 | using namespace AArch64II; |
8118 | |
8119 | static const std::pair<unsigned, const char *> TargetFlags[] = { |
8120 | {MO_COFFSTUB, "aarch64-coffstub" }, |
8121 | {MO_GOT, "aarch64-got" }, |
8122 | {MO_NC, "aarch64-nc" }, |
8123 | {MO_S, "aarch64-s" }, |
8124 | {MO_TLS, "aarch64-tls" }, |
8125 | {MO_DLLIMPORT, "aarch64-dllimport" }, |
8126 | {MO_PREL, "aarch64-prel" }, |
8127 | {MO_TAGGED, "aarch64-tagged" }, |
8128 | {MO_ARM64EC_CALLMANGLE, "aarch64-arm64ec-callmangle" }, |
8129 | }; |
8130 | return ArrayRef(TargetFlags); |
8131 | } |
8132 | |
8133 | ArrayRef<std::pair<MachineMemOperand::Flags, const char *>> |
8134 | AArch64InstrInfo::getSerializableMachineMemOperandTargetFlags() const { |
8135 | static const std::pair<MachineMemOperand::Flags, const char *> TargetFlags[] = |
8136 | {{MOSuppressPair, "aarch64-suppress-pair" }, |
8137 | {MOStridedAccess, "aarch64-strided-access" }}; |
8138 | return ArrayRef(TargetFlags); |
8139 | } |
8140 | |
8141 | /// Constants defining how certain sequences should be outlined. |
8142 | /// This encompasses how an outlined function should be called, and what kind of |
8143 | /// frame should be emitted for that outlined function. |
8144 | /// |
8145 | /// \p MachineOutlinerDefault implies that the function should be called with |
8146 | /// a save and restore of LR to the stack. |
8147 | /// |
8148 | /// That is, |
8149 | /// |
8150 | /// I1 Save LR OUTLINED_FUNCTION: |
8151 | /// I2 --> BL OUTLINED_FUNCTION I1 |
8152 | /// I3 Restore LR I2 |
8153 | /// I3 |
8154 | /// RET |
8155 | /// |
8156 | /// * Call construction overhead: 3 (save + BL + restore) |
8157 | /// * Frame construction overhead: 1 (ret) |
8158 | /// * Requires stack fixups? Yes |
8159 | /// |
8160 | /// \p MachineOutlinerTailCall implies that the function is being created from |
8161 | /// a sequence of instructions ending in a return. |
8162 | /// |
8163 | /// That is, |
8164 | /// |
8165 | /// I1 OUTLINED_FUNCTION: |
8166 | /// I2 --> B OUTLINED_FUNCTION I1 |
8167 | /// RET I2 |
8168 | /// RET |
8169 | /// |
8170 | /// * Call construction overhead: 1 (B) |
8171 | /// * Frame construction overhead: 0 (Return included in sequence) |
8172 | /// * Requires stack fixups? No |
8173 | /// |
8174 | /// \p MachineOutlinerNoLRSave implies that the function should be called using |
8175 | /// a BL instruction, but doesn't require LR to be saved and restored. This |
8176 | /// happens when LR is known to be dead. |
8177 | /// |
8178 | /// That is, |
8179 | /// |
8180 | /// I1 OUTLINED_FUNCTION: |
8181 | /// I2 --> BL OUTLINED_FUNCTION I1 |
8182 | /// I3 I2 |
8183 | /// I3 |
8184 | /// RET |
8185 | /// |
8186 | /// * Call construction overhead: 1 (BL) |
8187 | /// * Frame construction overhead: 1 (RET) |
8188 | /// * Requires stack fixups? No |
8189 | /// |
8190 | /// \p MachineOutlinerThunk implies that the function is being created from |
8191 | /// a sequence of instructions ending in a call. The outlined function is |
8192 | /// called with a BL instruction, and the outlined function tail-calls the |
8193 | /// original call destination. |
8194 | /// |
8195 | /// That is, |
8196 | /// |
8197 | /// I1 OUTLINED_FUNCTION: |
8198 | /// I2 --> BL OUTLINED_FUNCTION I1 |
8199 | /// BL f I2 |
8200 | /// B f |
8201 | /// * Call construction overhead: 1 (BL) |
8202 | /// * Frame construction overhead: 0 |
8203 | /// * Requires stack fixups? No |
8204 | /// |
8205 | /// \p MachineOutlinerRegSave implies that the function should be called with a |
8206 | /// save and restore of LR to an available register. This allows us to avoid |
8207 | /// stack fixups. Note that this outlining variant is compatible with the |
8208 | /// NoLRSave case. |
8209 | /// |
8210 | /// That is, |
8211 | /// |
8212 | /// I1 Save LR OUTLINED_FUNCTION: |
8213 | /// I2 --> BL OUTLINED_FUNCTION I1 |
8214 | /// I3 Restore LR I2 |
8215 | /// I3 |
8216 | /// RET |
8217 | /// |
8218 | /// * Call construction overhead: 3 (save + BL + restore) |
8219 | /// * Frame construction overhead: 1 (ret) |
8220 | /// * Requires stack fixups? No |
8221 | enum MachineOutlinerClass { |
8222 | MachineOutlinerDefault, /// Emit a save, restore, call, and return. |
8223 | MachineOutlinerTailCall, /// Only emit a branch. |
8224 | MachineOutlinerNoLRSave, /// Emit a call and return. |
8225 | MachineOutlinerThunk, /// Emit a call and tail-call. |
8226 | MachineOutlinerRegSave /// Same as default, but save to a register. |
8227 | }; |
8228 | |
8229 | enum MachineOutlinerMBBFlags { |
8230 | LRUnavailableSomewhere = 0x2, |
8231 | HasCalls = 0x4, |
8232 | UnsafeRegsDead = 0x8 |
8233 | }; |
8234 | |
8235 | Register |
8236 | AArch64InstrInfo::findRegisterToSaveLRTo(outliner::Candidate &C) const { |
8237 | MachineFunction *MF = C.getMF(); |
8238 | const TargetRegisterInfo &TRI = *MF->getSubtarget().getRegisterInfo(); |
8239 | const AArch64RegisterInfo *ARI = |
8240 | static_cast<const AArch64RegisterInfo *>(&TRI); |
8241 | // Check if there is an available register across the sequence that we can |
8242 | // use. |
8243 | for (unsigned Reg : AArch64::GPR64RegClass) { |
8244 | if (!ARI->isReservedReg(MF: *MF, Reg) && |
8245 | Reg != AArch64::LR && // LR is not reserved, but don't use it. |
8246 | Reg != AArch64::X16 && // X16 is not guaranteed to be preserved. |
8247 | Reg != AArch64::X17 && // Ditto for X17. |
8248 | C.isAvailableAcrossAndOutOfSeq(Reg, TRI) && |
8249 | C.isAvailableInsideSeq(Reg, TRI)) |
8250 | return Reg; |
8251 | } |
8252 | return Register(); |
8253 | } |
8254 | |
8255 | static bool |
8256 | outliningCandidatesSigningScopeConsensus(const outliner::Candidate &a, |
8257 | const outliner::Candidate &b) { |
8258 | const auto &MFIa = a.getMF()->getInfo<AArch64FunctionInfo>(); |
8259 | const auto &MFIb = b.getMF()->getInfo<AArch64FunctionInfo>(); |
8260 | |
8261 | return MFIa->shouldSignReturnAddress(SpillsLR: false) == MFIb->shouldSignReturnAddress(SpillsLR: false) && |
8262 | MFIa->shouldSignReturnAddress(SpillsLR: true) == MFIb->shouldSignReturnAddress(SpillsLR: true); |
8263 | } |
8264 | |
8265 | static bool |
8266 | outliningCandidatesSigningKeyConsensus(const outliner::Candidate &a, |
8267 | const outliner::Candidate &b) { |
8268 | const auto &MFIa = a.getMF()->getInfo<AArch64FunctionInfo>(); |
8269 | const auto &MFIb = b.getMF()->getInfo<AArch64FunctionInfo>(); |
8270 | |
8271 | return MFIa->shouldSignWithBKey() == MFIb->shouldSignWithBKey(); |
8272 | } |
8273 | |
8274 | static bool outliningCandidatesV8_3OpsConsensus(const outliner::Candidate &a, |
8275 | const outliner::Candidate &b) { |
8276 | const AArch64Subtarget &SubtargetA = |
8277 | a.getMF()->getSubtarget<AArch64Subtarget>(); |
8278 | const AArch64Subtarget &SubtargetB = |
8279 | b.getMF()->getSubtarget<AArch64Subtarget>(); |
8280 | return SubtargetA.hasV8_3aOps() == SubtargetB.hasV8_3aOps(); |
8281 | } |
8282 | |
8283 | std::optional<outliner::OutlinedFunction> |
8284 | AArch64InstrInfo::getOutliningCandidateInfo( |
8285 | std::vector<outliner::Candidate> &RepeatedSequenceLocs) const { |
8286 | unsigned SequenceSize = 0; |
8287 | for (auto &MI : RepeatedSequenceLocs[0]) |
8288 | SequenceSize += getInstSizeInBytes(MI); |
8289 | |
8290 | unsigned NumBytesToCreateFrame = 0; |
8291 | |
8292 | // We only allow outlining for functions having exactly matching return |
8293 | // address signing attributes, i.e., all share the same value for the |
8294 | // attribute "sign-return-address" and all share the same type of key they |
8295 | // are signed with. |
8296 | // Additionally we require all functions to simultaniously either support |
8297 | // v8.3a features or not. Otherwise an outlined function could get signed |
8298 | // using dedicated v8.3 instructions and a call from a function that doesn't |
8299 | // support v8.3 instructions would therefore be invalid. |
8300 | if (std::adjacent_find( |
8301 | first: RepeatedSequenceLocs.begin(), last: RepeatedSequenceLocs.end(), |
8302 | binary_pred: [](const outliner::Candidate &a, const outliner::Candidate &b) { |
8303 | // Return true if a and b are non-equal w.r.t. return address |
8304 | // signing or support of v8.3a features |
8305 | if (outliningCandidatesSigningScopeConsensus(a, b) && |
8306 | outliningCandidatesSigningKeyConsensus(a, b) && |
8307 | outliningCandidatesV8_3OpsConsensus(a, b)) { |
8308 | return false; |
8309 | } |
8310 | return true; |
8311 | }) != RepeatedSequenceLocs.end()) { |
8312 | return std::nullopt; |
8313 | } |
8314 | |
8315 | // Since at this point all candidates agree on their return address signing |
8316 | // picking just one is fine. If the candidate functions potentially sign their |
8317 | // return addresses, the outlined function should do the same. Note that in |
8318 | // the case of "sign-return-address"="non-leaf" this is an assumption: It is |
8319 | // not certainly true that the outlined function will have to sign its return |
8320 | // address but this decision is made later, when the decision to outline |
8321 | // has already been made. |
8322 | // The same holds for the number of additional instructions we need: On |
8323 | // v8.3a RET can be replaced by RETAA/RETAB and no AUT instruction is |
8324 | // necessary. However, at this point we don't know if the outlined function |
8325 | // will have a RET instruction so we assume the worst. |
8326 | const TargetRegisterInfo &TRI = getRegisterInfo(); |
8327 | // Performing a tail call may require extra checks when PAuth is enabled. |
8328 | // If PAuth is disabled, set it to zero for uniformity. |
8329 | unsigned NumBytesToCheckLRInTCEpilogue = 0; |
8330 | if (RepeatedSequenceLocs[0] |
8331 | .getMF() |
8332 | ->getInfo<AArch64FunctionInfo>() |
8333 | ->shouldSignReturnAddress(SpillsLR: true)) { |
8334 | // One PAC and one AUT instructions |
8335 | NumBytesToCreateFrame += 8; |
8336 | |
8337 | // PAuth is enabled - set extra tail call cost, if any. |
8338 | auto LRCheckMethod = Subtarget.getAuthenticatedLRCheckMethod( |
8339 | MF: *RepeatedSequenceLocs[0].getMF()); |
8340 | NumBytesToCheckLRInTCEpilogue = |
8341 | AArch64PAuth::getCheckerSizeInBytes(Method: LRCheckMethod); |
8342 | // Checking the authenticated LR value may significantly impact |
8343 | // SequenceSize, so account for it for more precise results. |
8344 | if (isTailCallReturnInst(MI: RepeatedSequenceLocs[0].back())) |
8345 | SequenceSize += NumBytesToCheckLRInTCEpilogue; |
8346 | |
8347 | // We have to check if sp modifying instructions would get outlined. |
8348 | // If so we only allow outlining if sp is unchanged overall, so matching |
8349 | // sub and add instructions are okay to outline, all other sp modifications |
8350 | // are not |
8351 | auto hasIllegalSPModification = [&TRI](outliner::Candidate &C) { |
8352 | int SPValue = 0; |
8353 | for (auto &MI : C) { |
8354 | if (MI.modifiesRegister(Reg: AArch64::SP, TRI: &TRI)) { |
8355 | switch (MI.getOpcode()) { |
8356 | case AArch64::ADDXri: |
8357 | case AArch64::ADDWri: |
8358 | assert(MI.getNumOperands() == 4 && "Wrong number of operands" ); |
8359 | assert(MI.getOperand(2).isImm() && |
8360 | "Expected operand to be immediate" ); |
8361 | assert(MI.getOperand(1).isReg() && |
8362 | "Expected operand to be a register" ); |
8363 | // Check if the add just increments sp. If so, we search for |
8364 | // matching sub instructions that decrement sp. If not, the |
8365 | // modification is illegal |
8366 | if (MI.getOperand(i: 1).getReg() == AArch64::SP) |
8367 | SPValue += MI.getOperand(i: 2).getImm(); |
8368 | else |
8369 | return true; |
8370 | break; |
8371 | case AArch64::SUBXri: |
8372 | case AArch64::SUBWri: |
8373 | assert(MI.getNumOperands() == 4 && "Wrong number of operands" ); |
8374 | assert(MI.getOperand(2).isImm() && |
8375 | "Expected operand to be immediate" ); |
8376 | assert(MI.getOperand(1).isReg() && |
8377 | "Expected operand to be a register" ); |
8378 | // Check if the sub just decrements sp. If so, we search for |
8379 | // matching add instructions that increment sp. If not, the |
8380 | // modification is illegal |
8381 | if (MI.getOperand(i: 1).getReg() == AArch64::SP) |
8382 | SPValue -= MI.getOperand(i: 2).getImm(); |
8383 | else |
8384 | return true; |
8385 | break; |
8386 | default: |
8387 | return true; |
8388 | } |
8389 | } |
8390 | } |
8391 | if (SPValue) |
8392 | return true; |
8393 | return false; |
8394 | }; |
8395 | // Remove candidates with illegal stack modifying instructions |
8396 | llvm::erase_if(C&: RepeatedSequenceLocs, P: hasIllegalSPModification); |
8397 | |
8398 | // If the sequence doesn't have enough candidates left, then we're done. |
8399 | if (RepeatedSequenceLocs.size() < 2) |
8400 | return std::nullopt; |
8401 | } |
8402 | |
8403 | // Properties about candidate MBBs that hold for all of them. |
8404 | unsigned FlagsSetInAll = 0xF; |
8405 | |
8406 | // Compute liveness information for each candidate, and set FlagsSetInAll. |
8407 | for (outliner::Candidate &C : RepeatedSequenceLocs) |
8408 | FlagsSetInAll &= C.Flags; |
8409 | |
8410 | unsigned LastInstrOpcode = RepeatedSequenceLocs[0].back().getOpcode(); |
8411 | |
8412 | // Helper lambda which sets call information for every candidate. |
8413 | auto SetCandidateCallInfo = |
8414 | [&RepeatedSequenceLocs](unsigned CallID, unsigned NumBytesForCall) { |
8415 | for (outliner::Candidate &C : RepeatedSequenceLocs) |
8416 | C.setCallInfo(CID: CallID, CO: NumBytesForCall); |
8417 | }; |
8418 | |
8419 | unsigned FrameID = MachineOutlinerDefault; |
8420 | NumBytesToCreateFrame += 4; |
8421 | |
8422 | bool HasBTI = any_of(Range&: RepeatedSequenceLocs, P: [](outliner::Candidate &C) { |
8423 | return C.getMF()->getInfo<AArch64FunctionInfo>()->branchTargetEnforcement(); |
8424 | }); |
8425 | |
8426 | // We check to see if CFI Instructions are present, and if they are |
8427 | // we find the number of CFI Instructions in the candidates. |
8428 | unsigned CFICount = 0; |
8429 | for (auto &I : RepeatedSequenceLocs[0]) { |
8430 | if (I.isCFIInstruction()) |
8431 | CFICount++; |
8432 | } |
8433 | |
8434 | // We compare the number of found CFI Instructions to the number of CFI |
8435 | // instructions in the parent function for each candidate. We must check this |
8436 | // since if we outline one of the CFI instructions in a function, we have to |
8437 | // outline them all for correctness. If we do not, the address offsets will be |
8438 | // incorrect between the two sections of the program. |
8439 | for (outliner::Candidate &C : RepeatedSequenceLocs) { |
8440 | std::vector<MCCFIInstruction> CFIInstructions = |
8441 | C.getMF()->getFrameInstructions(); |
8442 | |
8443 | if (CFICount > 0 && CFICount != CFIInstructions.size()) |
8444 | return std::nullopt; |
8445 | } |
8446 | |
8447 | // Returns true if an instructions is safe to fix up, false otherwise. |
8448 | auto IsSafeToFixup = [this, &TRI](MachineInstr &MI) { |
8449 | if (MI.isCall()) |
8450 | return true; |
8451 | |
8452 | if (!MI.modifiesRegister(Reg: AArch64::SP, TRI: &TRI) && |
8453 | !MI.readsRegister(Reg: AArch64::SP, TRI: &TRI)) |
8454 | return true; |
8455 | |
8456 | // Any modification of SP will break our code to save/restore LR. |
8457 | // FIXME: We could handle some instructions which add a constant |
8458 | // offset to SP, with a bit more work. |
8459 | if (MI.modifiesRegister(Reg: AArch64::SP, TRI: &TRI)) |
8460 | return false; |
8461 | |
8462 | // At this point, we have a stack instruction that we might need to |
8463 | // fix up. We'll handle it if it's a load or store. |
8464 | if (MI.mayLoadOrStore()) { |
8465 | const MachineOperand *Base; // Filled with the base operand of MI. |
8466 | int64_t Offset; // Filled with the offset of MI. |
8467 | bool OffsetIsScalable; |
8468 | |
8469 | // Does it allow us to offset the base operand and is the base the |
8470 | // register SP? |
8471 | if (!getMemOperandWithOffset(MI, BaseOp&: Base, Offset, OffsetIsScalable, TRI: &TRI) || |
8472 | !Base->isReg() || Base->getReg() != AArch64::SP) |
8473 | return false; |
8474 | |
8475 | // Fixe-up code below assumes bytes. |
8476 | if (OffsetIsScalable) |
8477 | return false; |
8478 | |
8479 | // Find the minimum/maximum offset for this instruction and check |
8480 | // if fixing it up would be in range. |
8481 | int64_t MinOffset, |
8482 | MaxOffset; // Unscaled offsets for the instruction. |
8483 | // The scale to multiply the offsets by. |
8484 | TypeSize Scale(0U, false), DummyWidth(0U, false); |
8485 | getMemOpInfo(Opcode: MI.getOpcode(), Scale, Width&: DummyWidth, MinOffset, MaxOffset); |
8486 | |
8487 | Offset += 16; // Update the offset to what it would be if we outlined. |
8488 | if (Offset < MinOffset * (int64_t)Scale.getFixedValue() || |
8489 | Offset > MaxOffset * (int64_t)Scale.getFixedValue()) |
8490 | return false; |
8491 | |
8492 | // It's in range, so we can outline it. |
8493 | return true; |
8494 | } |
8495 | |
8496 | // FIXME: Add handling for instructions like "add x0, sp, #8". |
8497 | |
8498 | // We can't fix it up, so don't outline it. |
8499 | return false; |
8500 | }; |
8501 | |
8502 | // True if it's possible to fix up each stack instruction in this sequence. |
8503 | // Important for frames/call variants that modify the stack. |
8504 | bool = |
8505 | llvm::all_of(Range&: RepeatedSequenceLocs[0], P: IsSafeToFixup); |
8506 | |
8507 | // If the last instruction in any candidate is a terminator, then we should |
8508 | // tail call all of the candidates. |
8509 | if (RepeatedSequenceLocs[0].back().isTerminator()) { |
8510 | FrameID = MachineOutlinerTailCall; |
8511 | NumBytesToCreateFrame = 0; |
8512 | unsigned NumBytesForCall = 4 + NumBytesToCheckLRInTCEpilogue; |
8513 | SetCandidateCallInfo(MachineOutlinerTailCall, NumBytesForCall); |
8514 | } |
8515 | |
8516 | else if (LastInstrOpcode == AArch64::BL || |
8517 | ((LastInstrOpcode == AArch64::BLR || |
8518 | LastInstrOpcode == AArch64::BLRNoIP) && |
8519 | !HasBTI)) { |
8520 | // FIXME: Do we need to check if the code after this uses the value of LR? |
8521 | FrameID = MachineOutlinerThunk; |
8522 | NumBytesToCreateFrame = NumBytesToCheckLRInTCEpilogue; |
8523 | SetCandidateCallInfo(MachineOutlinerThunk, 4); |
8524 | } |
8525 | |
8526 | else { |
8527 | // We need to decide how to emit calls + frames. We can always emit the same |
8528 | // frame if we don't need to save to the stack. If we have to save to the |
8529 | // stack, then we need a different frame. |
8530 | unsigned NumBytesNoStackCalls = 0; |
8531 | std::vector<outliner::Candidate> CandidatesWithoutStackFixups; |
8532 | |
8533 | // Check if we have to save LR. |
8534 | for (outliner::Candidate &C : RepeatedSequenceLocs) { |
8535 | bool LRAvailable = |
8536 | (C.Flags & MachineOutlinerMBBFlags::LRUnavailableSomewhere) |
8537 | ? C.isAvailableAcrossAndOutOfSeq(Reg: AArch64::LR, TRI) |
8538 | : true; |
8539 | // If we have a noreturn caller, then we're going to be conservative and |
8540 | // say that we have to save LR. If we don't have a ret at the end of the |
8541 | // block, then we can't reason about liveness accurately. |
8542 | // |
8543 | // FIXME: We can probably do better than always disabling this in |
8544 | // noreturn functions by fixing up the liveness info. |
8545 | bool IsNoReturn = |
8546 | C.getMF()->getFunction().hasFnAttribute(Kind: Attribute::NoReturn); |
8547 | |
8548 | // Is LR available? If so, we don't need a save. |
8549 | if (LRAvailable && !IsNoReturn) { |
8550 | NumBytesNoStackCalls += 4; |
8551 | C.setCallInfo(CID: MachineOutlinerNoLRSave, CO: 4); |
8552 | CandidatesWithoutStackFixups.push_back(x: C); |
8553 | } |
8554 | |
8555 | // Is an unused register available? If so, we won't modify the stack, so |
8556 | // we can outline with the same frame type as those that don't save LR. |
8557 | else if (findRegisterToSaveLRTo(C)) { |
8558 | NumBytesNoStackCalls += 12; |
8559 | C.setCallInfo(CID: MachineOutlinerRegSave, CO: 12); |
8560 | CandidatesWithoutStackFixups.push_back(x: C); |
8561 | } |
8562 | |
8563 | // Is SP used in the sequence at all? If not, we don't have to modify |
8564 | // the stack, so we are guaranteed to get the same frame. |
8565 | else if (C.isAvailableInsideSeq(Reg: AArch64::SP, TRI)) { |
8566 | NumBytesNoStackCalls += 12; |
8567 | C.setCallInfo(CID: MachineOutlinerDefault, CO: 12); |
8568 | CandidatesWithoutStackFixups.push_back(x: C); |
8569 | } |
8570 | |
8571 | // If we outline this, we need to modify the stack. Pretend we don't |
8572 | // outline this by saving all of its bytes. |
8573 | else { |
8574 | NumBytesNoStackCalls += SequenceSize; |
8575 | } |
8576 | } |
8577 | |
8578 | // If there are no places where we have to save LR, then note that we |
8579 | // don't have to update the stack. Otherwise, give every candidate the |
8580 | // default call type, as long as it's safe to do so. |
8581 | if (!AllStackInstrsSafe || |
8582 | NumBytesNoStackCalls <= RepeatedSequenceLocs.size() * 12) { |
8583 | RepeatedSequenceLocs = CandidatesWithoutStackFixups; |
8584 | FrameID = MachineOutlinerNoLRSave; |
8585 | if (RepeatedSequenceLocs.size() < 2) |
8586 | return std::nullopt; |
8587 | } else { |
8588 | SetCandidateCallInfo(MachineOutlinerDefault, 12); |
8589 | |
8590 | // Bugzilla ID: 46767 |
8591 | // TODO: Check if fixing up the stack more than once is safe so we can |
8592 | // outline these. |
8593 | // |
8594 | // An outline resulting in a caller that requires stack fixups at the |
8595 | // callsite to a callee that also requires stack fixups can happen when |
8596 | // there are no available registers at the candidate callsite for a |
8597 | // candidate that itself also has calls. |
8598 | // |
8599 | // In other words if function_containing_sequence in the following pseudo |
8600 | // assembly requires that we save LR at the point of the call, but there |
8601 | // are no available registers: in this case we save using SP and as a |
8602 | // result the SP offsets requires stack fixups by multiples of 16. |
8603 | // |
8604 | // function_containing_sequence: |
8605 | // ... |
8606 | // save LR to SP <- Requires stack instr fixups in OUTLINED_FUNCTION_N |
8607 | // call OUTLINED_FUNCTION_N |
8608 | // restore LR from SP |
8609 | // ... |
8610 | // |
8611 | // OUTLINED_FUNCTION_N: |
8612 | // save LR to SP <- Requires stack instr fixups in OUTLINED_FUNCTION_N |
8613 | // ... |
8614 | // bl foo |
8615 | // restore LR from SP |
8616 | // ret |
8617 | // |
8618 | // Because the code to handle more than one stack fixup does not |
8619 | // currently have the proper checks for legality, these cases will assert |
8620 | // in the AArch64 MachineOutliner. This is because the code to do this |
8621 | // needs more hardening, testing, better checks that generated code is |
8622 | // legal, etc and because it is only verified to handle a single pass of |
8623 | // stack fixup. |
8624 | // |
8625 | // The assert happens in AArch64InstrInfo::buildOutlinedFrame to catch |
8626 | // these cases until they are known to be handled. Bugzilla 46767 is |
8627 | // referenced in comments at the assert site. |
8628 | // |
8629 | // To avoid asserting (or generating non-legal code on noassert builds) |
8630 | // we remove all candidates which would need more than one stack fixup by |
8631 | // pruning the cases where the candidate has calls while also having no |
8632 | // available LR and having no available general purpose registers to copy |
8633 | // LR to (ie one extra stack save/restore). |
8634 | // |
8635 | if (FlagsSetInAll & MachineOutlinerMBBFlags::HasCalls) { |
8636 | erase_if(C&: RepeatedSequenceLocs, P: [this, &TRI](outliner::Candidate &C) { |
8637 | auto IsCall = [](const MachineInstr &MI) { return MI.isCall(); }; |
8638 | return (llvm::any_of(Range&: C, P: IsCall)) && |
8639 | (!C.isAvailableAcrossAndOutOfSeq(Reg: AArch64::LR, TRI) || |
8640 | !findRegisterToSaveLRTo(C)); |
8641 | }); |
8642 | } |
8643 | } |
8644 | |
8645 | // If we dropped all of the candidates, bail out here. |
8646 | if (RepeatedSequenceLocs.size() < 2) { |
8647 | RepeatedSequenceLocs.clear(); |
8648 | return std::nullopt; |
8649 | } |
8650 | } |
8651 | |
8652 | // Does every candidate's MBB contain a call? If so, then we might have a call |
8653 | // in the range. |
8654 | if (FlagsSetInAll & MachineOutlinerMBBFlags::HasCalls) { |
8655 | // Check if the range contains a call. These require a save + restore of the |
8656 | // link register. |
8657 | outliner::Candidate &FirstCand = RepeatedSequenceLocs[0]; |
8658 | bool ModStackToSaveLR = false; |
8659 | if (std::any_of(first: FirstCand.begin(), last: std::prev(x: FirstCand.end()), |
8660 | pred: [](const MachineInstr &MI) { return MI.isCall(); })) |
8661 | ModStackToSaveLR = true; |
8662 | |
8663 | // Handle the last instruction separately. If this is a tail call, then the |
8664 | // last instruction is a call. We don't want to save + restore in this case. |
8665 | // However, it could be possible that the last instruction is a call without |
8666 | // it being valid to tail call this sequence. We should consider this as |
8667 | // well. |
8668 | else if (FrameID != MachineOutlinerThunk && |
8669 | FrameID != MachineOutlinerTailCall && FirstCand.back().isCall()) |
8670 | ModStackToSaveLR = true; |
8671 | |
8672 | if (ModStackToSaveLR) { |
8673 | // We can't fix up the stack. Bail out. |
8674 | if (!AllStackInstrsSafe) { |
8675 | RepeatedSequenceLocs.clear(); |
8676 | return std::nullopt; |
8677 | } |
8678 | |
8679 | // Save + restore LR. |
8680 | NumBytesToCreateFrame += 8; |
8681 | } |
8682 | } |
8683 | |
8684 | // If we have CFI instructions, we can only outline if the outlined section |
8685 | // can be a tail call |
8686 | if (FrameID != MachineOutlinerTailCall && CFICount > 0) |
8687 | return std::nullopt; |
8688 | |
8689 | return outliner::OutlinedFunction(RepeatedSequenceLocs, SequenceSize, |
8690 | NumBytesToCreateFrame, FrameID); |
8691 | } |
8692 | |
8693 | void AArch64InstrInfo::mergeOutliningCandidateAttributes( |
8694 | Function &F, std::vector<outliner::Candidate> &Candidates) const { |
8695 | // If a bunch of candidates reach this point they must agree on their return |
8696 | // address signing. It is therefore enough to just consider the signing |
8697 | // behaviour of one of them |
8698 | const auto &CFn = Candidates.front().getMF()->getFunction(); |
8699 | |
8700 | if (CFn.hasFnAttribute(Kind: "ptrauth-returns" )) |
8701 | F.addFnAttr(Attr: CFn.getFnAttribute(Kind: "ptrauth-returns" )); |
8702 | if (CFn.hasFnAttribute(Kind: "ptrauth-auth-traps" )) |
8703 | F.addFnAttr(Attr: CFn.getFnAttribute(Kind: "ptrauth-auth-traps" )); |
8704 | // Since all candidates belong to the same module, just copy the |
8705 | // function-level attributes of an arbitrary function. |
8706 | if (CFn.hasFnAttribute(Kind: "sign-return-address" )) |
8707 | F.addFnAttr(Attr: CFn.getFnAttribute(Kind: "sign-return-address" )); |
8708 | if (CFn.hasFnAttribute(Kind: "sign-return-address-key" )) |
8709 | F.addFnAttr(Attr: CFn.getFnAttribute(Kind: "sign-return-address-key" )); |
8710 | |
8711 | AArch64GenInstrInfo::mergeOutliningCandidateAttributes(F, Candidates); |
8712 | } |
8713 | |
8714 | bool AArch64InstrInfo::isFunctionSafeToOutlineFrom( |
8715 | MachineFunction &MF, bool OutlineFromLinkOnceODRs) const { |
8716 | const Function &F = MF.getFunction(); |
8717 | |
8718 | // Can F be deduplicated by the linker? If it can, don't outline from it. |
8719 | if (!OutlineFromLinkOnceODRs && F.hasLinkOnceODRLinkage()) |
8720 | return false; |
8721 | |
8722 | // Don't outline from functions with section markings; the program could |
8723 | // expect that all the code is in the named section. |
8724 | // FIXME: Allow outlining from multiple functions with the same section |
8725 | // marking. |
8726 | if (F.hasSection()) |
8727 | return false; |
8728 | |
8729 | // Outlining from functions with redzones is unsafe since the outliner may |
8730 | // modify the stack. Check if hasRedZone is true or unknown; if yes, don't |
8731 | // outline from it. |
8732 | AArch64FunctionInfo *AFI = MF.getInfo<AArch64FunctionInfo>(); |
8733 | if (!AFI || AFI->hasRedZone().value_or(u: true)) |
8734 | return false; |
8735 | |
8736 | // FIXME: Determine whether it is safe to outline from functions which contain |
8737 | // streaming-mode changes. We may need to ensure any smstart/smstop pairs are |
8738 | // outlined together and ensure it is safe to outline with async unwind info, |
8739 | // required for saving & restoring VG around calls. |
8740 | if (AFI->hasStreamingModeChanges()) |
8741 | return false; |
8742 | |
8743 | // FIXME: Teach the outliner to generate/handle Windows unwind info. |
8744 | if (MF.getTarget().getMCAsmInfo()->usesWindowsCFI()) |
8745 | return false; |
8746 | |
8747 | // It's safe to outline from MF. |
8748 | return true; |
8749 | } |
8750 | |
8751 | SmallVector<std::pair<MachineBasicBlock::iterator, MachineBasicBlock::iterator>> |
8752 | AArch64InstrInfo::getOutlinableRanges(MachineBasicBlock &MBB, |
8753 | unsigned &Flags) const { |
8754 | assert(MBB.getParent()->getRegInfo().tracksLiveness() && |
8755 | "Must track liveness!" ); |
8756 | SmallVector< |
8757 | std::pair<MachineBasicBlock::iterator, MachineBasicBlock::iterator>> |
8758 | Ranges; |
8759 | // According to the AArch64 Procedure Call Standard, the following are |
8760 | // undefined on entry/exit from a function call: |
8761 | // |
8762 | // * Registers x16, x17, (and thus w16, w17) |
8763 | // * Condition codes (and thus the NZCV register) |
8764 | // |
8765 | // If any of these registers are used inside or live across an outlined |
8766 | // function, then they may be modified later, either by the compiler or |
8767 | // some other tool (like the linker). |
8768 | // |
8769 | // To avoid outlining in these situations, partition each block into ranges |
8770 | // where these registers are dead. We will only outline from those ranges. |
8771 | LiveRegUnits LRU(getRegisterInfo()); |
8772 | auto AreAllUnsafeRegsDead = [&LRU]() { |
8773 | return LRU.available(Reg: AArch64::W16) && LRU.available(Reg: AArch64::W17) && |
8774 | LRU.available(Reg: AArch64::NZCV); |
8775 | }; |
8776 | |
8777 | // We need to know if LR is live across an outlining boundary later on in |
8778 | // order to decide how we'll create the outlined call, frame, etc. |
8779 | // |
8780 | // It's pretty expensive to check this for *every candidate* within a block. |
8781 | // That's some potentially n^2 behaviour, since in the worst case, we'd need |
8782 | // to compute liveness from the end of the block for O(n) candidates within |
8783 | // the block. |
8784 | // |
8785 | // So, to improve the average case, let's keep track of liveness from the end |
8786 | // of the block to the beginning of *every outlinable range*. If we know that |
8787 | // LR is available in every range we could outline from, then we know that |
8788 | // we don't need to check liveness for any candidate within that range. |
8789 | bool LRAvailableEverywhere = true; |
8790 | // Compute liveness bottom-up. |
8791 | LRU.addLiveOuts(MBB); |
8792 | // Update flags that require info about the entire MBB. |
8793 | auto UpdateWholeMBBFlags = [&Flags](const MachineInstr &MI) { |
8794 | if (MI.isCall() && !MI.isTerminator()) |
8795 | Flags |= MachineOutlinerMBBFlags::HasCalls; |
8796 | }; |
8797 | // Range: [RangeBegin, RangeEnd) |
8798 | MachineBasicBlock::instr_iterator RangeBegin, RangeEnd; |
8799 | unsigned RangeLen; |
8800 | auto CreateNewRangeStartingAt = |
8801 | [&RangeBegin, &RangeEnd, |
8802 | &RangeLen](MachineBasicBlock::instr_iterator NewBegin) { |
8803 | RangeBegin = NewBegin; |
8804 | RangeEnd = std::next(x: RangeBegin); |
8805 | RangeLen = 0; |
8806 | }; |
8807 | auto SaveRangeIfNonEmpty = [&RangeLen, &Ranges, &RangeBegin, &RangeEnd]() { |
8808 | // At least one unsafe register is not dead. We do not want to outline at |
8809 | // this point. If it is long enough to outline from, save the range |
8810 | // [RangeBegin, RangeEnd). |
8811 | if (RangeLen > 1) |
8812 | Ranges.push_back(Elt: std::make_pair(x&: RangeBegin, y&: RangeEnd)); |
8813 | }; |
8814 | // Find the first point where all unsafe registers are dead. |
8815 | // FIND: <safe instr> <-- end of first potential range |
8816 | // SKIP: <unsafe def> |
8817 | // SKIP: ... everything between ... |
8818 | // SKIP: <unsafe use> |
8819 | auto FirstPossibleEndPt = MBB.instr_rbegin(); |
8820 | for (; FirstPossibleEndPt != MBB.instr_rend(); ++FirstPossibleEndPt) { |
8821 | LRU.stepBackward(MI: *FirstPossibleEndPt); |
8822 | // Update flags that impact how we outline across the entire block, |
8823 | // regardless of safety. |
8824 | UpdateWholeMBBFlags(*FirstPossibleEndPt); |
8825 | if (AreAllUnsafeRegsDead()) |
8826 | break; |
8827 | } |
8828 | // If we exhausted the entire block, we have no safe ranges to outline. |
8829 | if (FirstPossibleEndPt == MBB.instr_rend()) |
8830 | return Ranges; |
8831 | // Current range. |
8832 | CreateNewRangeStartingAt(FirstPossibleEndPt->getIterator()); |
8833 | // StartPt points to the first place where all unsafe registers |
8834 | // are dead (if there is any such point). Begin partitioning the MBB into |
8835 | // ranges. |
8836 | for (auto &MI : make_range(x: FirstPossibleEndPt, y: MBB.instr_rend())) { |
8837 | LRU.stepBackward(MI); |
8838 | UpdateWholeMBBFlags(MI); |
8839 | if (!AreAllUnsafeRegsDead()) { |
8840 | SaveRangeIfNonEmpty(); |
8841 | CreateNewRangeStartingAt(MI.getIterator()); |
8842 | continue; |
8843 | } |
8844 | LRAvailableEverywhere &= LRU.available(Reg: AArch64::LR); |
8845 | RangeBegin = MI.getIterator(); |
8846 | ++RangeLen; |
8847 | } |
8848 | // Above loop misses the last (or only) range. If we are still safe, then |
8849 | // let's save the range. |
8850 | if (AreAllUnsafeRegsDead()) |
8851 | SaveRangeIfNonEmpty(); |
8852 | if (Ranges.empty()) |
8853 | return Ranges; |
8854 | // We found the ranges bottom-up. Mapping expects the top-down. Reverse |
8855 | // the order. |
8856 | std::reverse(first: Ranges.begin(), last: Ranges.end()); |
8857 | // If there is at least one outlinable range where LR is unavailable |
8858 | // somewhere, remember that. |
8859 | if (!LRAvailableEverywhere) |
8860 | Flags |= MachineOutlinerMBBFlags::LRUnavailableSomewhere; |
8861 | return Ranges; |
8862 | } |
8863 | |
8864 | outliner::InstrType |
8865 | AArch64InstrInfo::getOutliningTypeImpl(MachineBasicBlock::iterator &MIT, |
8866 | unsigned Flags) const { |
8867 | MachineInstr &MI = *MIT; |
8868 | MachineBasicBlock *MBB = MI.getParent(); |
8869 | MachineFunction *MF = MBB->getParent(); |
8870 | AArch64FunctionInfo *FuncInfo = MF->getInfo<AArch64FunctionInfo>(); |
8871 | |
8872 | // Don't outline anything used for return address signing. The outlined |
8873 | // function will get signed later if needed |
8874 | switch (MI.getOpcode()) { |
8875 | case AArch64::PACM: |
8876 | case AArch64::PACIASP: |
8877 | case AArch64::PACIBSP: |
8878 | case AArch64::PACIASPPC: |
8879 | case AArch64::PACIBSPPC: |
8880 | case AArch64::AUTIASP: |
8881 | case AArch64::AUTIBSP: |
8882 | case AArch64::AUTIASPPCi: |
8883 | case AArch64::AUTIASPPCr: |
8884 | case AArch64::AUTIBSPPCi: |
8885 | case AArch64::AUTIBSPPCr: |
8886 | case AArch64::RETAA: |
8887 | case AArch64::RETAB: |
8888 | case AArch64::RETAASPPCi: |
8889 | case AArch64::RETAASPPCr: |
8890 | case AArch64::RETABSPPCi: |
8891 | case AArch64::RETABSPPCr: |
8892 | case AArch64::EMITBKEY: |
8893 | case AArch64::PAUTH_PROLOGUE: |
8894 | case AArch64::PAUTH_EPILOGUE: |
8895 | return outliner::InstrType::Illegal; |
8896 | } |
8897 | |
8898 | // Don't outline LOHs. |
8899 | if (FuncInfo->getLOHRelated().count(Ptr: &MI)) |
8900 | return outliner::InstrType::Illegal; |
8901 | |
8902 | // We can only outline these if we will tail call the outlined function, or |
8903 | // fix up the CFI offsets. Currently, CFI instructions are outlined only if |
8904 | // in a tail call. |
8905 | // |
8906 | // FIXME: If the proper fixups for the offset are implemented, this should be |
8907 | // possible. |
8908 | if (MI.isCFIInstruction()) |
8909 | return outliner::InstrType::Legal; |
8910 | |
8911 | // Is this a terminator for a basic block? |
8912 | if (MI.isTerminator()) |
8913 | // TargetInstrInfo::getOutliningType has already filtered out anything |
8914 | // that would break this, so we can allow it here. |
8915 | return outliner::InstrType::Legal; |
8916 | |
8917 | // Make sure none of the operands are un-outlinable. |
8918 | for (const MachineOperand &MOP : MI.operands()) { |
8919 | // A check preventing CFI indices was here before, but only CFI |
8920 | // instructions should have those. |
8921 | assert(!MOP.isCFIIndex()); |
8922 | |
8923 | // If it uses LR or W30 explicitly, then don't touch it. |
8924 | if (MOP.isReg() && !MOP.isImplicit() && |
8925 | (MOP.getReg() == AArch64::LR || MOP.getReg() == AArch64::W30)) |
8926 | return outliner::InstrType::Illegal; |
8927 | } |
8928 | |
8929 | // Special cases for instructions that can always be outlined, but will fail |
8930 | // the later tests. e.g, ADRPs, which are PC-relative use LR, but can always |
8931 | // be outlined because they don't require a *specific* value to be in LR. |
8932 | if (MI.getOpcode() == AArch64::ADRP) |
8933 | return outliner::InstrType::Legal; |
8934 | |
8935 | // If MI is a call we might be able to outline it. We don't want to outline |
8936 | // any calls that rely on the position of items on the stack. When we outline |
8937 | // something containing a call, we have to emit a save and restore of LR in |
8938 | // the outlined function. Currently, this always happens by saving LR to the |
8939 | // stack. Thus, if we outline, say, half the parameters for a function call |
8940 | // plus the call, then we'll break the callee's expectations for the layout |
8941 | // of the stack. |
8942 | // |
8943 | // FIXME: Allow calls to functions which construct a stack frame, as long |
8944 | // as they don't access arguments on the stack. |
8945 | // FIXME: Figure out some way to analyze functions defined in other modules. |
8946 | // We should be able to compute the memory usage based on the IR calling |
8947 | // convention, even if we can't see the definition. |
8948 | if (MI.isCall()) { |
8949 | // Get the function associated with the call. Look at each operand and find |
8950 | // the one that represents the callee and get its name. |
8951 | const Function *Callee = nullptr; |
8952 | for (const MachineOperand &MOP : MI.operands()) { |
8953 | if (MOP.isGlobal()) { |
8954 | Callee = dyn_cast<Function>(Val: MOP.getGlobal()); |
8955 | break; |
8956 | } |
8957 | } |
8958 | |
8959 | // Never outline calls to mcount. There isn't any rule that would require |
8960 | // this, but the Linux kernel's "ftrace" feature depends on it. |
8961 | if (Callee && Callee->getName() == "\01_mcount" ) |
8962 | return outliner::InstrType::Illegal; |
8963 | |
8964 | // If we don't know anything about the callee, assume it depends on the |
8965 | // stack layout of the caller. In that case, it's only legal to outline |
8966 | // as a tail-call. Explicitly list the call instructions we know about so we |
8967 | // don't get unexpected results with call pseudo-instructions. |
8968 | auto UnknownCallOutlineType = outliner::InstrType::Illegal; |
8969 | if (MI.getOpcode() == AArch64::BLR || |
8970 | MI.getOpcode() == AArch64::BLRNoIP || MI.getOpcode() == AArch64::BL) |
8971 | UnknownCallOutlineType = outliner::InstrType::LegalTerminator; |
8972 | |
8973 | if (!Callee) |
8974 | return UnknownCallOutlineType; |
8975 | |
8976 | // We have a function we have information about. Check it if it's something |
8977 | // can safely outline. |
8978 | MachineFunction *CalleeMF = MF->getMMI().getMachineFunction(F: *Callee); |
8979 | |
8980 | // We don't know what's going on with the callee at all. Don't touch it. |
8981 | if (!CalleeMF) |
8982 | return UnknownCallOutlineType; |
8983 | |
8984 | // Check if we know anything about the callee saves on the function. If we |
8985 | // don't, then don't touch it, since that implies that we haven't |
8986 | // computed anything about its stack frame yet. |
8987 | MachineFrameInfo &MFI = CalleeMF->getFrameInfo(); |
8988 | if (!MFI.isCalleeSavedInfoValid() || MFI.getStackSize() > 0 || |
8989 | MFI.getNumObjects() > 0) |
8990 | return UnknownCallOutlineType; |
8991 | |
8992 | // At this point, we can say that CalleeMF ought to not pass anything on the |
8993 | // stack. Therefore, we can outline it. |
8994 | return outliner::InstrType::Legal; |
8995 | } |
8996 | |
8997 | // Don't touch the link register or W30. |
8998 | if (MI.readsRegister(Reg: AArch64::W30, TRI: &getRegisterInfo()) || |
8999 | MI.modifiesRegister(Reg: AArch64::W30, TRI: &getRegisterInfo())) |
9000 | return outliner::InstrType::Illegal; |
9001 | |
9002 | // Don't outline BTI instructions, because that will prevent the outlining |
9003 | // site from being indirectly callable. |
9004 | if (hasBTISemantics(MI)) |
9005 | return outliner::InstrType::Illegal; |
9006 | |
9007 | return outliner::InstrType::Legal; |
9008 | } |
9009 | |
9010 | void AArch64InstrInfo::fixupPostOutline(MachineBasicBlock &MBB) const { |
9011 | for (MachineInstr &MI : MBB) { |
9012 | const MachineOperand *Base; |
9013 | TypeSize Width(0, false); |
9014 | int64_t Offset; |
9015 | bool OffsetIsScalable; |
9016 | |
9017 | // Is this a load or store with an immediate offset with SP as the base? |
9018 | if (!MI.mayLoadOrStore() || |
9019 | !getMemOperandWithOffsetWidth(LdSt: MI, BaseOp&: Base, Offset, OffsetIsScalable, Width, |
9020 | TRI: &RI) || |
9021 | (Base->isReg() && Base->getReg() != AArch64::SP)) |
9022 | continue; |
9023 | |
9024 | // It is, so we have to fix it up. |
9025 | TypeSize Scale(0U, false); |
9026 | int64_t Dummy1, Dummy2; |
9027 | |
9028 | MachineOperand &StackOffsetOperand = getMemOpBaseRegImmOfsOffsetOperand(LdSt&: MI); |
9029 | assert(StackOffsetOperand.isImm() && "Stack offset wasn't immediate!" ); |
9030 | getMemOpInfo(Opcode: MI.getOpcode(), Scale, Width, MinOffset&: Dummy1, MaxOffset&: Dummy2); |
9031 | assert(Scale != 0 && "Unexpected opcode!" ); |
9032 | assert(!OffsetIsScalable && "Expected offset to be a byte offset" ); |
9033 | |
9034 | // We've pushed the return address to the stack, so add 16 to the offset. |
9035 | // This is safe, since we already checked if it would overflow when we |
9036 | // checked if this instruction was legal to outline. |
9037 | int64_t NewImm = (Offset + 16) / (int64_t)Scale.getFixedValue(); |
9038 | StackOffsetOperand.setImm(NewImm); |
9039 | } |
9040 | } |
9041 | |
9042 | static void signOutlinedFunction(MachineFunction &MF, MachineBasicBlock &MBB, |
9043 | const AArch64InstrInfo *TII, |
9044 | bool ShouldSignReturnAddr) { |
9045 | if (!ShouldSignReturnAddr) |
9046 | return; |
9047 | |
9048 | BuildMI(BB&: MBB, I: MBB.begin(), MIMD: DebugLoc(), MCID: TII->get(Opcode: AArch64::PAUTH_PROLOGUE)) |
9049 | .setMIFlag(MachineInstr::FrameSetup); |
9050 | BuildMI(BB&: MBB, I: MBB.getFirstInstrTerminator(), MIMD: DebugLoc(), |
9051 | MCID: TII->get(Opcode: AArch64::PAUTH_EPILOGUE)) |
9052 | .setMIFlag(MachineInstr::FrameDestroy); |
9053 | } |
9054 | |
9055 | void AArch64InstrInfo::buildOutlinedFrame( |
9056 | MachineBasicBlock &MBB, MachineFunction &MF, |
9057 | const outliner::OutlinedFunction &OF) const { |
9058 | |
9059 | AArch64FunctionInfo *FI = MF.getInfo<AArch64FunctionInfo>(); |
9060 | |
9061 | if (OF.FrameConstructionID == MachineOutlinerTailCall) |
9062 | FI->setOutliningStyle("Tail Call" ); |
9063 | else if (OF.FrameConstructionID == MachineOutlinerThunk) { |
9064 | // For thunk outlining, rewrite the last instruction from a call to a |
9065 | // tail-call. |
9066 | MachineInstr *Call = &*--MBB.instr_end(); |
9067 | unsigned TailOpcode; |
9068 | if (Call->getOpcode() == AArch64::BL) { |
9069 | TailOpcode = AArch64::TCRETURNdi; |
9070 | } else { |
9071 | assert(Call->getOpcode() == AArch64::BLR || |
9072 | Call->getOpcode() == AArch64::BLRNoIP); |
9073 | TailOpcode = AArch64::TCRETURNriALL; |
9074 | } |
9075 | MachineInstr *TC = BuildMI(MF, MIMD: DebugLoc(), MCID: get(Opcode: TailOpcode)) |
9076 | .add(MO: Call->getOperand(i: 0)) |
9077 | .addImm(Val: 0); |
9078 | MBB.insert(I: MBB.end(), MI: TC); |
9079 | Call->eraseFromParent(); |
9080 | |
9081 | FI->setOutliningStyle("Thunk" ); |
9082 | } |
9083 | |
9084 | bool IsLeafFunction = true; |
9085 | |
9086 | // Is there a call in the outlined range? |
9087 | auto IsNonTailCall = [](const MachineInstr &MI) { |
9088 | return MI.isCall() && !MI.isReturn(); |
9089 | }; |
9090 | |
9091 | if (llvm::any_of(Range: MBB.instrs(), P: IsNonTailCall)) { |
9092 | // Fix up the instructions in the range, since we're going to modify the |
9093 | // stack. |
9094 | |
9095 | // Bugzilla ID: 46767 |
9096 | // TODO: Check if fixing up twice is safe so we can outline these. |
9097 | assert(OF.FrameConstructionID != MachineOutlinerDefault && |
9098 | "Can only fix up stack references once" ); |
9099 | fixupPostOutline(MBB); |
9100 | |
9101 | IsLeafFunction = false; |
9102 | |
9103 | // LR has to be a live in so that we can save it. |
9104 | if (!MBB.isLiveIn(Reg: AArch64::LR)) |
9105 | MBB.addLiveIn(PhysReg: AArch64::LR); |
9106 | |
9107 | MachineBasicBlock::iterator It = MBB.begin(); |
9108 | MachineBasicBlock::iterator Et = MBB.end(); |
9109 | |
9110 | if (OF.FrameConstructionID == MachineOutlinerTailCall || |
9111 | OF.FrameConstructionID == MachineOutlinerThunk) |
9112 | Et = std::prev(x: MBB.end()); |
9113 | |
9114 | // Insert a save before the outlined region |
9115 | MachineInstr *STRXpre = BuildMI(MF, MIMD: DebugLoc(), MCID: get(Opcode: AArch64::STRXpre)) |
9116 | .addReg(RegNo: AArch64::SP, flags: RegState::Define) |
9117 | .addReg(RegNo: AArch64::LR) |
9118 | .addReg(RegNo: AArch64::SP) |
9119 | .addImm(Val: -16); |
9120 | It = MBB.insert(I: It, MI: STRXpre); |
9121 | |
9122 | if (MF.getInfo<AArch64FunctionInfo>()->needsDwarfUnwindInfo(MF)) { |
9123 | const TargetSubtargetInfo &STI = MF.getSubtarget(); |
9124 | const MCRegisterInfo *MRI = STI.getRegisterInfo(); |
9125 | unsigned DwarfReg = MRI->getDwarfRegNum(RegNum: AArch64::LR, isEH: true); |
9126 | |
9127 | // Add a CFI saying the stack was moved 16 B down. |
9128 | int64_t StackPosEntry = |
9129 | MF.addFrameInst(Inst: MCCFIInstruction::cfiDefCfaOffset(L: nullptr, Offset: 16)); |
9130 | BuildMI(BB&: MBB, I: It, MIMD: DebugLoc(), MCID: get(Opcode: AArch64::CFI_INSTRUCTION)) |
9131 | .addCFIIndex(CFIIndex: StackPosEntry) |
9132 | .setMIFlags(MachineInstr::FrameSetup); |
9133 | |
9134 | // Add a CFI saying that the LR that we want to find is now 16 B higher |
9135 | // than before. |
9136 | int64_t LRPosEntry = MF.addFrameInst( |
9137 | Inst: MCCFIInstruction::createOffset(L: nullptr, Register: DwarfReg, Offset: -16)); |
9138 | BuildMI(BB&: MBB, I: It, MIMD: DebugLoc(), MCID: get(Opcode: AArch64::CFI_INSTRUCTION)) |
9139 | .addCFIIndex(CFIIndex: LRPosEntry) |
9140 | .setMIFlags(MachineInstr::FrameSetup); |
9141 | } |
9142 | |
9143 | // Insert a restore before the terminator for the function. |
9144 | MachineInstr *LDRXpost = BuildMI(MF, MIMD: DebugLoc(), MCID: get(Opcode: AArch64::LDRXpost)) |
9145 | .addReg(RegNo: AArch64::SP, flags: RegState::Define) |
9146 | .addReg(RegNo: AArch64::LR, flags: RegState::Define) |
9147 | .addReg(RegNo: AArch64::SP) |
9148 | .addImm(Val: 16); |
9149 | Et = MBB.insert(I: Et, MI: LDRXpost); |
9150 | } |
9151 | |
9152 | bool ShouldSignReturnAddr = FI->shouldSignReturnAddress(SpillsLR: !IsLeafFunction); |
9153 | |
9154 | // If this is a tail call outlined function, then there's already a return. |
9155 | if (OF.FrameConstructionID == MachineOutlinerTailCall || |
9156 | OF.FrameConstructionID == MachineOutlinerThunk) { |
9157 | signOutlinedFunction(MF, MBB, TII: this, ShouldSignReturnAddr); |
9158 | return; |
9159 | } |
9160 | |
9161 | // It's not a tail call, so we have to insert the return ourselves. |
9162 | |
9163 | // LR has to be a live in so that we can return to it. |
9164 | if (!MBB.isLiveIn(Reg: AArch64::LR)) |
9165 | MBB.addLiveIn(PhysReg: AArch64::LR); |
9166 | |
9167 | MachineInstr *ret = BuildMI(MF, MIMD: DebugLoc(), MCID: get(Opcode: AArch64::RET)) |
9168 | .addReg(RegNo: AArch64::LR); |
9169 | MBB.insert(I: MBB.end(), MI: ret); |
9170 | |
9171 | signOutlinedFunction(MF, MBB, TII: this, ShouldSignReturnAddr); |
9172 | |
9173 | FI->setOutliningStyle("Function" ); |
9174 | |
9175 | // Did we have to modify the stack by saving the link register? |
9176 | if (OF.FrameConstructionID != MachineOutlinerDefault) |
9177 | return; |
9178 | |
9179 | // We modified the stack. |
9180 | // Walk over the basic block and fix up all the stack accesses. |
9181 | fixupPostOutline(MBB); |
9182 | } |
9183 | |
9184 | MachineBasicBlock::iterator AArch64InstrInfo::insertOutlinedCall( |
9185 | Module &M, MachineBasicBlock &MBB, MachineBasicBlock::iterator &It, |
9186 | MachineFunction &MF, outliner::Candidate &C) const { |
9187 | |
9188 | // Are we tail calling? |
9189 | if (C.CallConstructionID == MachineOutlinerTailCall) { |
9190 | // If yes, then we can just branch to the label. |
9191 | It = MBB.insert(I: It, MI: BuildMI(MF, MIMD: DebugLoc(), MCID: get(Opcode: AArch64::TCRETURNdi)) |
9192 | .addGlobalAddress(GV: M.getNamedValue(Name: MF.getName())) |
9193 | .addImm(Val: 0)); |
9194 | return It; |
9195 | } |
9196 | |
9197 | // Are we saving the link register? |
9198 | if (C.CallConstructionID == MachineOutlinerNoLRSave || |
9199 | C.CallConstructionID == MachineOutlinerThunk) { |
9200 | // No, so just insert the call. |
9201 | It = MBB.insert(I: It, MI: BuildMI(MF, MIMD: DebugLoc(), MCID: get(Opcode: AArch64::BL)) |
9202 | .addGlobalAddress(GV: M.getNamedValue(Name: MF.getName()))); |
9203 | return It; |
9204 | } |
9205 | |
9206 | // We want to return the spot where we inserted the call. |
9207 | MachineBasicBlock::iterator CallPt; |
9208 | |
9209 | // Instructions for saving and restoring LR around the call instruction we're |
9210 | // going to insert. |
9211 | MachineInstr *Save; |
9212 | MachineInstr *Restore; |
9213 | // Can we save to a register? |
9214 | if (C.CallConstructionID == MachineOutlinerRegSave) { |
9215 | // FIXME: This logic should be sunk into a target-specific interface so that |
9216 | // we don't have to recompute the register. |
9217 | Register Reg = findRegisterToSaveLRTo(C); |
9218 | assert(Reg && "No callee-saved register available?" ); |
9219 | |
9220 | // LR has to be a live in so that we can save it. |
9221 | if (!MBB.isLiveIn(Reg: AArch64::LR)) |
9222 | MBB.addLiveIn(PhysReg: AArch64::LR); |
9223 | |
9224 | // Save and restore LR from Reg. |
9225 | Save = BuildMI(MF, MIMD: DebugLoc(), MCID: get(Opcode: AArch64::ORRXrs), DestReg: Reg) |
9226 | .addReg(RegNo: AArch64::XZR) |
9227 | .addReg(RegNo: AArch64::LR) |
9228 | .addImm(Val: 0); |
9229 | Restore = BuildMI(MF, MIMD: DebugLoc(), MCID: get(Opcode: AArch64::ORRXrs), DestReg: AArch64::LR) |
9230 | .addReg(RegNo: AArch64::XZR) |
9231 | .addReg(RegNo: Reg) |
9232 | .addImm(Val: 0); |
9233 | } else { |
9234 | // We have the default case. Save and restore from SP. |
9235 | Save = BuildMI(MF, MIMD: DebugLoc(), MCID: get(Opcode: AArch64::STRXpre)) |
9236 | .addReg(RegNo: AArch64::SP, flags: RegState::Define) |
9237 | .addReg(RegNo: AArch64::LR) |
9238 | .addReg(RegNo: AArch64::SP) |
9239 | .addImm(Val: -16); |
9240 | Restore = BuildMI(MF, MIMD: DebugLoc(), MCID: get(Opcode: AArch64::LDRXpost)) |
9241 | .addReg(RegNo: AArch64::SP, flags: RegState::Define) |
9242 | .addReg(RegNo: AArch64::LR, flags: RegState::Define) |
9243 | .addReg(RegNo: AArch64::SP) |
9244 | .addImm(Val: 16); |
9245 | } |
9246 | |
9247 | It = MBB.insert(I: It, MI: Save); |
9248 | It++; |
9249 | |
9250 | // Insert the call. |
9251 | It = MBB.insert(I: It, MI: BuildMI(MF, MIMD: DebugLoc(), MCID: get(Opcode: AArch64::BL)) |
9252 | .addGlobalAddress(GV: M.getNamedValue(Name: MF.getName()))); |
9253 | CallPt = It; |
9254 | It++; |
9255 | |
9256 | It = MBB.insert(I: It, MI: Restore); |
9257 | return CallPt; |
9258 | } |
9259 | |
9260 | bool AArch64InstrInfo::shouldOutlineFromFunctionByDefault( |
9261 | MachineFunction &MF) const { |
9262 | return MF.getFunction().hasMinSize(); |
9263 | } |
9264 | |
9265 | void AArch64InstrInfo::buildClearRegister(Register Reg, MachineBasicBlock &MBB, |
9266 | MachineBasicBlock::iterator Iter, |
9267 | DebugLoc &DL, |
9268 | bool AllowSideEffects) const { |
9269 | const MachineFunction &MF = *MBB.getParent(); |
9270 | const AArch64Subtarget &STI = MF.getSubtarget<AArch64Subtarget>(); |
9271 | const AArch64RegisterInfo &TRI = *STI.getRegisterInfo(); |
9272 | |
9273 | if (TRI.isGeneralPurposeRegister(MF, Reg)) { |
9274 | BuildMI(BB&: MBB, I: Iter, MIMD: DL, MCID: get(Opcode: AArch64::MOVZXi), DestReg: Reg).addImm(Val: 0).addImm(Val: 0); |
9275 | } else if (STI.hasSVE()) { |
9276 | BuildMI(BB&: MBB, I: Iter, MIMD: DL, MCID: get(Opcode: AArch64::DUP_ZI_D), DestReg: Reg) |
9277 | .addImm(Val: 0) |
9278 | .addImm(Val: 0); |
9279 | } else { |
9280 | BuildMI(BB&: MBB, I: Iter, MIMD: DL, MCID: get(Opcode: AArch64::MOVIv2d_ns), DestReg: Reg) |
9281 | .addImm(Val: 0); |
9282 | } |
9283 | } |
9284 | |
9285 | std::optional<DestSourcePair> |
9286 | AArch64InstrInfo::isCopyInstrImpl(const MachineInstr &MI) const { |
9287 | |
9288 | // AArch64::ORRWrs and AArch64::ORRXrs with WZR/XZR reg |
9289 | // and zero immediate operands used as an alias for mov instruction. |
9290 | if (MI.getOpcode() == AArch64::ORRWrs && |
9291 | MI.getOperand(i: 1).getReg() == AArch64::WZR && |
9292 | MI.getOperand(i: 3).getImm() == 0x0 && |
9293 | // Check that the w->w move is not a zero-extending w->x mov. |
9294 | (!MI.getOperand(i: 0).getReg().isVirtual() || |
9295 | MI.getOperand(i: 0).getSubReg() == 0) && |
9296 | (!MI.getOperand(i: 0).getReg().isPhysical() || |
9297 | MI.findRegisterDefOperandIdx(Reg: MI.getOperand(i: 0).getReg() - AArch64::W0 + |
9298 | AArch64::X0, |
9299 | /*TRI=*/nullptr) == -1)) |
9300 | return DestSourcePair{MI.getOperand(i: 0), MI.getOperand(i: 2)}; |
9301 | |
9302 | if (MI.getOpcode() == AArch64::ORRXrs && |
9303 | MI.getOperand(i: 1).getReg() == AArch64::XZR && |
9304 | MI.getOperand(i: 3).getImm() == 0x0) |
9305 | return DestSourcePair{MI.getOperand(i: 0), MI.getOperand(i: 2)}; |
9306 | |
9307 | return std::nullopt; |
9308 | } |
9309 | |
9310 | std::optional<DestSourcePair> |
9311 | AArch64InstrInfo::isCopyLikeInstrImpl(const MachineInstr &MI) const { |
9312 | if (MI.getOpcode() == AArch64::ORRWrs && |
9313 | MI.getOperand(i: 1).getReg() == AArch64::WZR && |
9314 | MI.getOperand(i: 3).getImm() == 0x0) |
9315 | return DestSourcePair{MI.getOperand(i: 0), MI.getOperand(i: 2)}; |
9316 | return std::nullopt; |
9317 | } |
9318 | |
9319 | std::optional<RegImmPair> |
9320 | AArch64InstrInfo::isAddImmediate(const MachineInstr &MI, Register Reg) const { |
9321 | int Sign = 1; |
9322 | int64_t Offset = 0; |
9323 | |
9324 | // TODO: Handle cases where Reg is a super- or sub-register of the |
9325 | // destination register. |
9326 | const MachineOperand &Op0 = MI.getOperand(i: 0); |
9327 | if (!Op0.isReg() || Reg != Op0.getReg()) |
9328 | return std::nullopt; |
9329 | |
9330 | switch (MI.getOpcode()) { |
9331 | default: |
9332 | return std::nullopt; |
9333 | case AArch64::SUBWri: |
9334 | case AArch64::SUBXri: |
9335 | case AArch64::SUBSWri: |
9336 | case AArch64::SUBSXri: |
9337 | Sign *= -1; |
9338 | [[fallthrough]]; |
9339 | case AArch64::ADDSWri: |
9340 | case AArch64::ADDSXri: |
9341 | case AArch64::ADDWri: |
9342 | case AArch64::ADDXri: { |
9343 | // TODO: Third operand can be global address (usually some string). |
9344 | if (!MI.getOperand(i: 0).isReg() || !MI.getOperand(i: 1).isReg() || |
9345 | !MI.getOperand(i: 2).isImm()) |
9346 | return std::nullopt; |
9347 | int Shift = MI.getOperand(i: 3).getImm(); |
9348 | assert((Shift == 0 || Shift == 12) && "Shift can be either 0 or 12" ); |
9349 | Offset = Sign * (MI.getOperand(i: 2).getImm() << Shift); |
9350 | } |
9351 | } |
9352 | return RegImmPair{MI.getOperand(i: 1).getReg(), Offset}; |
9353 | } |
9354 | |
9355 | /// If the given ORR instruction is a copy, and \p DescribedReg overlaps with |
9356 | /// the destination register then, if possible, describe the value in terms of |
9357 | /// the source register. |
9358 | static std::optional<ParamLoadedValue> |
9359 | describeORRLoadedValue(const MachineInstr &MI, Register DescribedReg, |
9360 | const TargetInstrInfo *TII, |
9361 | const TargetRegisterInfo *TRI) { |
9362 | auto DestSrc = TII->isCopyLikeInstr(MI); |
9363 | if (!DestSrc) |
9364 | return std::nullopt; |
9365 | |
9366 | Register DestReg = DestSrc->Destination->getReg(); |
9367 | Register SrcReg = DestSrc->Source->getReg(); |
9368 | |
9369 | auto Expr = DIExpression::get(Context&: MI.getMF()->getFunction().getContext(), Elements: {}); |
9370 | |
9371 | // If the described register is the destination, just return the source. |
9372 | if (DestReg == DescribedReg) |
9373 | return ParamLoadedValue(MachineOperand::CreateReg(Reg: SrcReg, isDef: false), Expr); |
9374 | |
9375 | // ORRWrs zero-extends to 64-bits, so we need to consider such cases. |
9376 | if (MI.getOpcode() == AArch64::ORRWrs && |
9377 | TRI->isSuperRegister(RegA: DestReg, RegB: DescribedReg)) |
9378 | return ParamLoadedValue(MachineOperand::CreateReg(Reg: SrcReg, isDef: false), Expr); |
9379 | |
9380 | // We may need to describe the lower part of a ORRXrs move. |
9381 | if (MI.getOpcode() == AArch64::ORRXrs && |
9382 | TRI->isSubRegister(RegA: DestReg, RegB: DescribedReg)) { |
9383 | Register SrcSubReg = TRI->getSubReg(Reg: SrcReg, Idx: AArch64::sub_32); |
9384 | return ParamLoadedValue(MachineOperand::CreateReg(Reg: SrcSubReg, isDef: false), Expr); |
9385 | } |
9386 | |
9387 | assert(!TRI->isSuperOrSubRegisterEq(DestReg, DescribedReg) && |
9388 | "Unhandled ORR[XW]rs copy case" ); |
9389 | |
9390 | return std::nullopt; |
9391 | } |
9392 | |
9393 | bool AArch64InstrInfo::isFunctionSafeToSplit(const MachineFunction &MF) const { |
9394 | // Functions cannot be split to different sections on AArch64 if they have |
9395 | // a red zone. This is because relaxing a cross-section branch may require |
9396 | // incrementing the stack pointer to spill a register, which would overwrite |
9397 | // the red zone. |
9398 | if (MF.getInfo<AArch64FunctionInfo>()->hasRedZone().value_or(u: true)) |
9399 | return false; |
9400 | |
9401 | return TargetInstrInfo::isFunctionSafeToSplit(MF); |
9402 | } |
9403 | |
9404 | bool AArch64InstrInfo::isMBBSafeToSplitToCold( |
9405 | const MachineBasicBlock &MBB) const { |
9406 | // Asm Goto blocks can contain conditional branches to goto labels, which can |
9407 | // get moved out of range of the branch instruction. |
9408 | auto isAsmGoto = [](const MachineInstr &MI) { |
9409 | return MI.getOpcode() == AArch64::INLINEASM_BR; |
9410 | }; |
9411 | if (llvm::any_of(Range: MBB, P: isAsmGoto) || MBB.isInlineAsmBrIndirectTarget()) |
9412 | return false; |
9413 | |
9414 | // Because jump tables are label-relative instead of table-relative, they all |
9415 | // must be in the same section or relocation fixup handling will fail. |
9416 | |
9417 | // Check if MBB is a jump table target |
9418 | const MachineJumpTableInfo *MJTI = MBB.getParent()->getJumpTableInfo(); |
9419 | auto containsMBB = [&MBB](const MachineJumpTableEntry &JTE) { |
9420 | return llvm::is_contained(Range: JTE.MBBs, Element: &MBB); |
9421 | }; |
9422 | if (MJTI != nullptr && llvm::any_of(Range: MJTI->getJumpTables(), P: containsMBB)) |
9423 | return false; |
9424 | |
9425 | // Check if MBB contains a jump table lookup |
9426 | for (const MachineInstr &MI : MBB) { |
9427 | switch (MI.getOpcode()) { |
9428 | case TargetOpcode::G_BRJT: |
9429 | case AArch64::JumpTableDest32: |
9430 | case AArch64::JumpTableDest16: |
9431 | case AArch64::JumpTableDest8: |
9432 | return false; |
9433 | default: |
9434 | continue; |
9435 | } |
9436 | } |
9437 | |
9438 | // MBB isn't a special case, so it's safe to be split to the cold section. |
9439 | return true; |
9440 | } |
9441 | |
9442 | std::optional<ParamLoadedValue> |
9443 | AArch64InstrInfo::describeLoadedValue(const MachineInstr &MI, |
9444 | Register Reg) const { |
9445 | const MachineFunction *MF = MI.getMF(); |
9446 | const TargetRegisterInfo *TRI = MF->getSubtarget().getRegisterInfo(); |
9447 | switch (MI.getOpcode()) { |
9448 | case AArch64::MOVZWi: |
9449 | case AArch64::MOVZXi: { |
9450 | // MOVZWi may be used for producing zero-extended 32-bit immediates in |
9451 | // 64-bit parameters, so we need to consider super-registers. |
9452 | if (!TRI->isSuperRegisterEq(RegA: MI.getOperand(i: 0).getReg(), RegB: Reg)) |
9453 | return std::nullopt; |
9454 | |
9455 | if (!MI.getOperand(i: 1).isImm()) |
9456 | return std::nullopt; |
9457 | int64_t Immediate = MI.getOperand(i: 1).getImm(); |
9458 | int Shift = MI.getOperand(i: 2).getImm(); |
9459 | return ParamLoadedValue(MachineOperand::CreateImm(Val: Immediate << Shift), |
9460 | nullptr); |
9461 | } |
9462 | case AArch64::ORRWrs: |
9463 | case AArch64::ORRXrs: |
9464 | return describeORRLoadedValue(MI, DescribedReg: Reg, TII: this, TRI); |
9465 | } |
9466 | |
9467 | return TargetInstrInfo::describeLoadedValue(MI, Reg); |
9468 | } |
9469 | |
9470 | bool AArch64InstrInfo::isExtendLikelyToBeFolded( |
9471 | MachineInstr &ExtMI, MachineRegisterInfo &MRI) const { |
9472 | assert(ExtMI.getOpcode() == TargetOpcode::G_SEXT || |
9473 | ExtMI.getOpcode() == TargetOpcode::G_ZEXT || |
9474 | ExtMI.getOpcode() == TargetOpcode::G_ANYEXT); |
9475 | |
9476 | // Anyexts are nops. |
9477 | if (ExtMI.getOpcode() == TargetOpcode::G_ANYEXT) |
9478 | return true; |
9479 | |
9480 | Register DefReg = ExtMI.getOperand(i: 0).getReg(); |
9481 | if (!MRI.hasOneNonDBGUse(RegNo: DefReg)) |
9482 | return false; |
9483 | |
9484 | // It's likely that a sext/zext as a G_PTR_ADD offset will be folded into an |
9485 | // addressing mode. |
9486 | auto *UserMI = &*MRI.use_instr_nodbg_begin(RegNo: DefReg); |
9487 | return UserMI->getOpcode() == TargetOpcode::G_PTR_ADD; |
9488 | } |
9489 | |
9490 | uint64_t AArch64InstrInfo::getElementSizeForOpcode(unsigned Opc) const { |
9491 | return get(Opcode: Opc).TSFlags & AArch64::ElementSizeMask; |
9492 | } |
9493 | |
9494 | bool AArch64InstrInfo::isPTestLikeOpcode(unsigned Opc) const { |
9495 | return get(Opcode: Opc).TSFlags & AArch64::InstrFlagIsPTestLike; |
9496 | } |
9497 | |
9498 | bool AArch64InstrInfo::isWhileOpcode(unsigned Opc) const { |
9499 | return get(Opcode: Opc).TSFlags & AArch64::InstrFlagIsWhile; |
9500 | } |
9501 | |
9502 | unsigned int |
9503 | AArch64InstrInfo::getTailDuplicateSize(CodeGenOptLevel OptLevel) const { |
9504 | return OptLevel >= CodeGenOptLevel::Aggressive ? 6 : 2; |
9505 | } |
9506 | |
9507 | bool AArch64InstrInfo::isLegalAddressingMode(unsigned NumBytes, int64_t Offset, |
9508 | unsigned Scale) const { |
9509 | if (Offset && Scale) |
9510 | return false; |
9511 | |
9512 | // Check Reg + Imm |
9513 | if (!Scale) { |
9514 | // 9-bit signed offset |
9515 | if (isInt<9>(x: Offset)) |
9516 | return true; |
9517 | |
9518 | // 12-bit unsigned offset |
9519 | unsigned Shift = Log2_64(Value: NumBytes); |
9520 | if (NumBytes && Offset > 0 && (Offset / NumBytes) <= (1LL << 12) - 1 && |
9521 | // Must be a multiple of NumBytes (NumBytes is a power of 2) |
9522 | (Offset >> Shift) << Shift == Offset) |
9523 | return true; |
9524 | return false; |
9525 | } |
9526 | |
9527 | // Check reg1 + SIZE_IN_BYTES * reg2 and reg1 + reg2 |
9528 | return Scale == 1 || (Scale > 0 && Scale == NumBytes); |
9529 | } |
9530 | |
9531 | unsigned llvm::getBLRCallOpcode(const MachineFunction &MF) { |
9532 | if (MF.getSubtarget<AArch64Subtarget>().hardenSlsBlr()) |
9533 | return AArch64::BLRNoIP; |
9534 | else |
9535 | return AArch64::BLR; |
9536 | } |
9537 | |
9538 | MachineBasicBlock::iterator |
9539 | AArch64InstrInfo::probedStackAlloc(MachineBasicBlock::iterator MBBI, |
9540 | Register TargetReg, bool FrameSetup) const { |
9541 | assert(TargetReg != AArch64::SP && "New top of stack cannot aleady be in SP" ); |
9542 | |
9543 | MachineBasicBlock &MBB = *MBBI->getParent(); |
9544 | MachineFunction &MF = *MBB.getParent(); |
9545 | const AArch64InstrInfo *TII = |
9546 | MF.getSubtarget<AArch64Subtarget>().getInstrInfo(); |
9547 | int64_t ProbeSize = MF.getInfo<AArch64FunctionInfo>()->getStackProbeSize(); |
9548 | DebugLoc DL = MBB.findDebugLoc(MBBI); |
9549 | |
9550 | MachineFunction::iterator MBBInsertPoint = std::next(x: MBB.getIterator()); |
9551 | MachineBasicBlock *LoopTestMBB = |
9552 | MF.CreateMachineBasicBlock(BB: MBB.getBasicBlock()); |
9553 | MF.insert(MBBI: MBBInsertPoint, MBB: LoopTestMBB); |
9554 | MachineBasicBlock *LoopBodyMBB = |
9555 | MF.CreateMachineBasicBlock(BB: MBB.getBasicBlock()); |
9556 | MF.insert(MBBI: MBBInsertPoint, MBB: LoopBodyMBB); |
9557 | MachineBasicBlock *ExitMBB = MF.CreateMachineBasicBlock(BB: MBB.getBasicBlock()); |
9558 | MF.insert(MBBI: MBBInsertPoint, MBB: ExitMBB); |
9559 | MachineInstr::MIFlag Flags = |
9560 | FrameSetup ? MachineInstr::FrameSetup : MachineInstr::NoFlags; |
9561 | |
9562 | // LoopTest: |
9563 | // SUB SP, SP, #ProbeSize |
9564 | emitFrameOffset(MBB&: *LoopTestMBB, MBBI: LoopTestMBB->end(), DL, DestReg: AArch64::SP, |
9565 | SrcReg: AArch64::SP, Offset: StackOffset::getFixed(Fixed: -ProbeSize), TII, Flag: Flags); |
9566 | |
9567 | // CMP SP, TargetReg |
9568 | BuildMI(BB&: *LoopTestMBB, I: LoopTestMBB->end(), MIMD: DL, MCID: TII->get(Opcode: AArch64::SUBSXrx64), |
9569 | DestReg: AArch64::XZR) |
9570 | .addReg(RegNo: AArch64::SP) |
9571 | .addReg(RegNo: TargetReg) |
9572 | .addImm(Val: AArch64_AM::getArithExtendImm(ET: AArch64_AM::UXTX, Imm: 0)) |
9573 | .setMIFlags(Flags); |
9574 | |
9575 | // B.<Cond> LoopExit |
9576 | BuildMI(BB&: *LoopTestMBB, I: LoopTestMBB->end(), MIMD: DL, MCID: TII->get(Opcode: AArch64::Bcc)) |
9577 | .addImm(Val: AArch64CC::LE) |
9578 | .addMBB(MBB: ExitMBB) |
9579 | .setMIFlags(Flags); |
9580 | |
9581 | // STR XZR, [SP] |
9582 | BuildMI(BB&: *LoopBodyMBB, I: LoopBodyMBB->end(), MIMD: DL, MCID: TII->get(Opcode: AArch64::STRXui)) |
9583 | .addReg(RegNo: AArch64::XZR) |
9584 | .addReg(RegNo: AArch64::SP) |
9585 | .addImm(Val: 0) |
9586 | .setMIFlags(Flags); |
9587 | |
9588 | // B loop |
9589 | BuildMI(BB&: *LoopBodyMBB, I: LoopBodyMBB->end(), MIMD: DL, MCID: TII->get(Opcode: AArch64::B)) |
9590 | .addMBB(MBB: LoopTestMBB) |
9591 | .setMIFlags(Flags); |
9592 | |
9593 | // LoopExit: |
9594 | // MOV SP, TargetReg |
9595 | BuildMI(BB&: *ExitMBB, I: ExitMBB->end(), MIMD: DL, MCID: TII->get(Opcode: AArch64::ADDXri), DestReg: AArch64::SP) |
9596 | .addReg(RegNo: TargetReg) |
9597 | .addImm(Val: 0) |
9598 | .addImm(Val: AArch64_AM::getShifterImm(ST: AArch64_AM::LSL, Imm: 0)) |
9599 | .setMIFlags(Flags); |
9600 | |
9601 | // LDR XZR, [SP] |
9602 | BuildMI(BB&: *ExitMBB, I: ExitMBB->end(), MIMD: DL, MCID: TII->get(Opcode: AArch64::LDRXui)) |
9603 | .addReg(RegNo: AArch64::XZR, flags: RegState::Define) |
9604 | .addReg(RegNo: AArch64::SP) |
9605 | .addImm(Val: 0) |
9606 | .setMIFlags(Flags); |
9607 | |
9608 | ExitMBB->splice(Where: ExitMBB->end(), Other: &MBB, From: std::next(x: MBBI), To: MBB.end()); |
9609 | ExitMBB->transferSuccessorsAndUpdatePHIs(FromMBB: &MBB); |
9610 | |
9611 | LoopTestMBB->addSuccessor(Succ: ExitMBB); |
9612 | LoopTestMBB->addSuccessor(Succ: LoopBodyMBB); |
9613 | LoopBodyMBB->addSuccessor(Succ: LoopTestMBB); |
9614 | MBB.addSuccessor(Succ: LoopTestMBB); |
9615 | |
9616 | // Update liveins. |
9617 | if (MF.getRegInfo().reservedRegsFrozen()) |
9618 | fullyRecomputeLiveIns(MBBs: {ExitMBB, LoopBodyMBB, LoopTestMBB}); |
9619 | |
9620 | return ExitMBB->begin(); |
9621 | } |
9622 | |
9623 | namespace { |
9624 | class AArch64PipelinerLoopInfo : public TargetInstrInfo::PipelinerLoopInfo { |
9625 | MachineFunction *MF; |
9626 | const TargetInstrInfo *TII; |
9627 | const TargetRegisterInfo *TRI; |
9628 | MachineRegisterInfo &MRI; |
9629 | |
9630 | /// The block of the loop |
9631 | MachineBasicBlock *LoopBB; |
9632 | /// The conditional branch of the loop |
9633 | MachineInstr *CondBranch; |
9634 | /// The compare instruction for loop control |
9635 | MachineInstr *Comp; |
9636 | /// The number of the operand of the loop counter value in Comp |
9637 | unsigned CompCounterOprNum; |
9638 | /// The instruction that updates the loop counter value |
9639 | MachineInstr *Update; |
9640 | /// The number of the operand of the loop counter value in Update |
9641 | unsigned UpdateCounterOprNum; |
9642 | /// The initial value of the loop counter |
9643 | Register Init; |
9644 | /// True iff Update is a predecessor of Comp |
9645 | bool IsUpdatePriorComp; |
9646 | |
9647 | /// The normalized condition used by createTripCountGreaterCondition() |
9648 | SmallVector<MachineOperand, 4> Cond; |
9649 | |
9650 | public: |
9651 | AArch64PipelinerLoopInfo(MachineBasicBlock *LoopBB, MachineInstr *CondBranch, |
9652 | MachineInstr *Comp, unsigned CompCounterOprNum, |
9653 | MachineInstr *Update, unsigned UpdateCounterOprNum, |
9654 | Register Init, bool IsUpdatePriorComp, |
9655 | const SmallVectorImpl<MachineOperand> &Cond) |
9656 | : MF(Comp->getParent()->getParent()), |
9657 | TII(MF->getSubtarget().getInstrInfo()), |
9658 | TRI(MF->getSubtarget().getRegisterInfo()), MRI(MF->getRegInfo()), |
9659 | LoopBB(LoopBB), CondBranch(CondBranch), Comp(Comp), |
9660 | CompCounterOprNum(CompCounterOprNum), Update(Update), |
9661 | UpdateCounterOprNum(UpdateCounterOprNum), Init(Init), |
9662 | IsUpdatePriorComp(IsUpdatePriorComp), Cond(Cond.begin(), Cond.end()) {} |
9663 | |
9664 | bool shouldIgnoreForPipelining(const MachineInstr *MI) const override { |
9665 | // Make the instructions for loop control be placed in stage 0. |
9666 | // The predecessors of Comp are considered by the caller. |
9667 | return MI == Comp; |
9668 | } |
9669 | |
9670 | std::optional<bool> createTripCountGreaterCondition( |
9671 | int TC, MachineBasicBlock &MBB, |
9672 | SmallVectorImpl<MachineOperand> &CondParam) override { |
9673 | // A branch instruction will be inserted as "if (Cond) goto epilogue". |
9674 | // Cond is normalized for such use. |
9675 | // The predecessors of the branch are assumed to have already been inserted. |
9676 | CondParam = Cond; |
9677 | return {}; |
9678 | } |
9679 | |
9680 | void createRemainingIterationsGreaterCondition( |
9681 | int TC, MachineBasicBlock &MBB, SmallVectorImpl<MachineOperand> &Cond, |
9682 | DenseMap<MachineInstr *, MachineInstr *> &LastStage0Insts) override; |
9683 | |
9684 | void (MachineBasicBlock *) override {} |
9685 | |
9686 | void adjustTripCount(int TripCountAdjust) override {} |
9687 | |
9688 | void disposed() override {} |
9689 | bool isMVEExpanderSupported() override { return true; } |
9690 | }; |
9691 | } // namespace |
9692 | |
9693 | /// Clone an instruction from MI. The register of ReplaceOprNum-th operand |
9694 | /// is replaced by ReplaceReg. The output register is newly created. |
9695 | /// The other operands are unchanged from MI. |
9696 | static Register cloneInstr(const MachineInstr *MI, unsigned ReplaceOprNum, |
9697 | Register ReplaceReg, MachineBasicBlock &MBB, |
9698 | MachineBasicBlock::iterator InsertTo) { |
9699 | MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo(); |
9700 | const TargetInstrInfo *TII = MBB.getParent()->getSubtarget().getInstrInfo(); |
9701 | const TargetRegisterInfo *TRI = |
9702 | MBB.getParent()->getSubtarget().getRegisterInfo(); |
9703 | MachineInstr *NewMI = MBB.getParent()->CloneMachineInstr(Orig: MI); |
9704 | Register Result = 0; |
9705 | for (unsigned I = 0; I < NewMI->getNumOperands(); ++I) { |
9706 | if (I == 0 && NewMI->getOperand(i: 0).getReg().isVirtual()) { |
9707 | Result = MRI.createVirtualRegister( |
9708 | RegClass: MRI.getRegClass(Reg: NewMI->getOperand(i: 0).getReg())); |
9709 | NewMI->getOperand(i: I).setReg(Result); |
9710 | } else if (I == ReplaceOprNum) { |
9711 | MRI.constrainRegClass( |
9712 | Reg: ReplaceReg, |
9713 | RC: TII->getRegClass(MCID: NewMI->getDesc(), OpNum: I, TRI, MF: *MBB.getParent())); |
9714 | NewMI->getOperand(i: I).setReg(ReplaceReg); |
9715 | } |
9716 | } |
9717 | MBB.insert(I: InsertTo, MI: NewMI); |
9718 | return Result; |
9719 | } |
9720 | |
9721 | void AArch64PipelinerLoopInfo::createRemainingIterationsGreaterCondition( |
9722 | int TC, MachineBasicBlock &MBB, SmallVectorImpl<MachineOperand> &Cond, |
9723 | DenseMap<MachineInstr *, MachineInstr *> &LastStage0Insts) { |
9724 | // Create and accumulate conditions for next TC iterations. |
9725 | // Example: |
9726 | // SUBSXrr N, counter, implicit-def $nzcv # compare instruction for the last |
9727 | // # iteration of the kernel |
9728 | // |
9729 | // # insert the following instructions |
9730 | // cond = CSINCXr 0, 0, C, implicit $nzcv |
9731 | // counter = ADDXri counter, 1 # clone from this->Update |
9732 | // SUBSXrr n, counter, implicit-def $nzcv # clone from this->Comp |
9733 | // cond = CSINCXr cond, cond, C, implicit $nzcv |
9734 | // ... (repeat TC times) |
9735 | // SUBSXri cond, 0, implicit-def $nzcv |
9736 | |
9737 | assert(CondBranch->getOpcode() == AArch64::Bcc); |
9738 | // CondCode to exit the loop |
9739 | AArch64CC::CondCode CC = |
9740 | (AArch64CC::CondCode)CondBranch->getOperand(i: 0).getImm(); |
9741 | if (CondBranch->getOperand(i: 1).getMBB() == LoopBB) |
9742 | CC = AArch64CC::getInvertedCondCode(Code: CC); |
9743 | |
9744 | // Accumulate conditions to exit the loop |
9745 | Register AccCond = AArch64::XZR; |
9746 | |
9747 | // If CC holds, CurCond+1 is returned; otherwise CurCond is returned. |
9748 | auto AccumulateCond = [&](Register CurCond, |
9749 | AArch64CC::CondCode CC) -> Register { |
9750 | Register NewCond = MRI.createVirtualRegister(RegClass: &AArch64::GPR64commonRegClass); |
9751 | BuildMI(BB&: MBB, I: MBB.end(), MIMD: Comp->getDebugLoc(), MCID: TII->get(Opcode: AArch64::CSINCXr)) |
9752 | .addReg(RegNo: NewCond, flags: RegState::Define) |
9753 | .addReg(RegNo: CurCond) |
9754 | .addReg(RegNo: CurCond) |
9755 | .addImm(Val: AArch64CC::getInvertedCondCode(Code: CC)); |
9756 | return NewCond; |
9757 | }; |
9758 | |
9759 | if (!LastStage0Insts.empty() && LastStage0Insts[Comp]->getParent() == &MBB) { |
9760 | // Update and Comp for I==0 are already exists in MBB |
9761 | // (MBB is an unrolled kernel) |
9762 | Register Counter; |
9763 | for (int I = 0; I <= TC; ++I) { |
9764 | Register NextCounter; |
9765 | if (I != 0) |
9766 | NextCounter = |
9767 | cloneInstr(MI: Comp, ReplaceOprNum: CompCounterOprNum, ReplaceReg: Counter, MBB, InsertTo: MBB.end()); |
9768 | |
9769 | AccCond = AccumulateCond(AccCond, CC); |
9770 | |
9771 | if (I != TC) { |
9772 | if (I == 0) { |
9773 | if (Update != Comp && IsUpdatePriorComp) { |
9774 | Counter = |
9775 | LastStage0Insts[Comp]->getOperand(i: CompCounterOprNum).getReg(); |
9776 | NextCounter = cloneInstr(MI: Update, ReplaceOprNum: UpdateCounterOprNum, ReplaceReg: Counter, MBB, |
9777 | InsertTo: MBB.end()); |
9778 | } else { |
9779 | // can use already calculated value |
9780 | NextCounter = LastStage0Insts[Update]->getOperand(i: 0).getReg(); |
9781 | } |
9782 | } else if (Update != Comp) { |
9783 | NextCounter = |
9784 | cloneInstr(MI: Update, ReplaceOprNum: UpdateCounterOprNum, ReplaceReg: Counter, MBB, InsertTo: MBB.end()); |
9785 | } |
9786 | } |
9787 | Counter = NextCounter; |
9788 | } |
9789 | } else { |
9790 | Register Counter; |
9791 | if (LastStage0Insts.empty()) { |
9792 | // use initial counter value (testing if the trip count is sufficient to |
9793 | // be executed by pipelined code) |
9794 | Counter = Init; |
9795 | if (IsUpdatePriorComp) |
9796 | Counter = |
9797 | cloneInstr(MI: Update, ReplaceOprNum: UpdateCounterOprNum, ReplaceReg: Counter, MBB, InsertTo: MBB.end()); |
9798 | } else { |
9799 | // MBB is an epilogue block. LastStage0Insts[Comp] is in the kernel block. |
9800 | Counter = LastStage0Insts[Comp]->getOperand(i: CompCounterOprNum).getReg(); |
9801 | } |
9802 | |
9803 | for (int I = 0; I <= TC; ++I) { |
9804 | Register NextCounter; |
9805 | NextCounter = |
9806 | cloneInstr(MI: Comp, ReplaceOprNum: CompCounterOprNum, ReplaceReg: Counter, MBB, InsertTo: MBB.end()); |
9807 | AccCond = AccumulateCond(AccCond, CC); |
9808 | if (I != TC && Update != Comp) |
9809 | NextCounter = |
9810 | cloneInstr(MI: Update, ReplaceOprNum: UpdateCounterOprNum, ReplaceReg: Counter, MBB, InsertTo: MBB.end()); |
9811 | Counter = NextCounter; |
9812 | } |
9813 | } |
9814 | |
9815 | // If AccCond == 0, the remainder is greater than TC. |
9816 | BuildMI(BB&: MBB, I: MBB.end(), MIMD: Comp->getDebugLoc(), MCID: TII->get(Opcode: AArch64::SUBSXri)) |
9817 | .addReg(RegNo: AArch64::XZR, flags: RegState::Define | RegState::Dead) |
9818 | .addReg(RegNo: AccCond) |
9819 | .addImm(Val: 0) |
9820 | .addImm(Val: 0); |
9821 | Cond.clear(); |
9822 | Cond.push_back(Elt: MachineOperand::CreateImm(Val: AArch64CC::EQ)); |
9823 | } |
9824 | |
9825 | static void (const MachineInstr &Phi, const MachineBasicBlock *MBB, |
9826 | Register &RegMBB, Register &RegOther) { |
9827 | assert(Phi.getNumOperands() == 5); |
9828 | if (Phi.getOperand(i: 2).getMBB() == MBB) { |
9829 | RegMBB = Phi.getOperand(i: 1).getReg(); |
9830 | RegOther = Phi.getOperand(i: 3).getReg(); |
9831 | } else { |
9832 | assert(Phi.getOperand(4).getMBB() == MBB); |
9833 | RegMBB = Phi.getOperand(i: 3).getReg(); |
9834 | RegOther = Phi.getOperand(i: 1).getReg(); |
9835 | } |
9836 | } |
9837 | |
9838 | static bool isDefinedOutside(Register Reg, const MachineBasicBlock *BB) { |
9839 | if (!Reg.isVirtual()) |
9840 | return false; |
9841 | const MachineRegisterInfo &MRI = BB->getParent()->getRegInfo(); |
9842 | return MRI.getVRegDef(Reg)->getParent() != BB; |
9843 | } |
9844 | |
9845 | /// If Reg is an induction variable, return true and set some parameters |
9846 | static bool getIndVarInfo(Register Reg, const MachineBasicBlock *LoopBB, |
9847 | MachineInstr *&UpdateInst, |
9848 | unsigned &UpdateCounterOprNum, Register &InitReg, |
9849 | bool &IsUpdatePriorComp) { |
9850 | // Example: |
9851 | // |
9852 | // Preheader: |
9853 | // InitReg = ... |
9854 | // LoopBB: |
9855 | // Reg0 = PHI (InitReg, Preheader), (Reg1, LoopBB) |
9856 | // Reg = COPY Reg0 ; COPY is ignored. |
9857 | // Reg1 = ADD Reg, #1; UpdateInst. Incremented by a loop invariant value. |
9858 | // ; Reg is the value calculated in the previous |
9859 | // ; iteration, so IsUpdatePriorComp == false. |
9860 | |
9861 | if (LoopBB->pred_size() != 2) |
9862 | return false; |
9863 | if (!Reg.isVirtual()) |
9864 | return false; |
9865 | const MachineRegisterInfo &MRI = LoopBB->getParent()->getRegInfo(); |
9866 | UpdateInst = nullptr; |
9867 | UpdateCounterOprNum = 0; |
9868 | InitReg = 0; |
9869 | IsUpdatePriorComp = true; |
9870 | Register CurReg = Reg; |
9871 | while (true) { |
9872 | MachineInstr *Def = MRI.getVRegDef(Reg: CurReg); |
9873 | if (Def->getParent() != LoopBB) |
9874 | return false; |
9875 | if (Def->isCopy()) { |
9876 | // Ignore copy instructions unless they contain subregisters |
9877 | if (Def->getOperand(i: 0).getSubReg() || Def->getOperand(i: 1).getSubReg()) |
9878 | return false; |
9879 | CurReg = Def->getOperand(i: 1).getReg(); |
9880 | } else if (Def->isPHI()) { |
9881 | if (InitReg != 0) |
9882 | return false; |
9883 | if (!UpdateInst) |
9884 | IsUpdatePriorComp = false; |
9885 | extractPhiReg(Phi: *Def, MBB: LoopBB, RegMBB&: CurReg, RegOther&: InitReg); |
9886 | } else { |
9887 | if (UpdateInst) |
9888 | return false; |
9889 | switch (Def->getOpcode()) { |
9890 | case AArch64::ADDSXri: |
9891 | case AArch64::ADDSWri: |
9892 | case AArch64::SUBSXri: |
9893 | case AArch64::SUBSWri: |
9894 | case AArch64::ADDXri: |
9895 | case AArch64::ADDWri: |
9896 | case AArch64::SUBXri: |
9897 | case AArch64::SUBWri: |
9898 | UpdateInst = Def; |
9899 | UpdateCounterOprNum = 1; |
9900 | break; |
9901 | case AArch64::ADDSXrr: |
9902 | case AArch64::ADDSWrr: |
9903 | case AArch64::SUBSXrr: |
9904 | case AArch64::SUBSWrr: |
9905 | case AArch64::ADDXrr: |
9906 | case AArch64::ADDWrr: |
9907 | case AArch64::SUBXrr: |
9908 | case AArch64::SUBWrr: |
9909 | UpdateInst = Def; |
9910 | if (isDefinedOutside(Reg: Def->getOperand(i: 2).getReg(), BB: LoopBB)) |
9911 | UpdateCounterOprNum = 1; |
9912 | else if (isDefinedOutside(Reg: Def->getOperand(i: 1).getReg(), BB: LoopBB)) |
9913 | UpdateCounterOprNum = 2; |
9914 | else |
9915 | return false; |
9916 | break; |
9917 | default: |
9918 | return false; |
9919 | } |
9920 | CurReg = Def->getOperand(i: UpdateCounterOprNum).getReg(); |
9921 | } |
9922 | |
9923 | if (!CurReg.isVirtual()) |
9924 | return false; |
9925 | if (Reg == CurReg) |
9926 | break; |
9927 | } |
9928 | |
9929 | if (!UpdateInst) |
9930 | return false; |
9931 | |
9932 | return true; |
9933 | } |
9934 | |
9935 | std::unique_ptr<TargetInstrInfo::PipelinerLoopInfo> |
9936 | AArch64InstrInfo::analyzeLoopForPipelining(MachineBasicBlock *LoopBB) const { |
9937 | // Accept loops that meet the following conditions |
9938 | // * The conditional branch is BCC |
9939 | // * The compare instruction is ADDS/SUBS/WHILEXX |
9940 | // * One operand of the compare is an induction variable and the other is a |
9941 | // loop invariant value |
9942 | // * The induction variable is incremented/decremented by a single instruction |
9943 | // * Does not contain CALL or instructions which have unmodeled side effects |
9944 | |
9945 | for (MachineInstr &MI : *LoopBB) |
9946 | if (MI.isCall() || MI.hasUnmodeledSideEffects()) |
9947 | // This instruction may use NZCV, which interferes with the instruction to |
9948 | // be inserted for loop control. |
9949 | return nullptr; |
9950 | |
9951 | MachineBasicBlock *TBB = nullptr, *FBB = nullptr; |
9952 | SmallVector<MachineOperand, 4> Cond; |
9953 | if (analyzeBranch(MBB&: *LoopBB, TBB, FBB, Cond)) |
9954 | return nullptr; |
9955 | |
9956 | // Infinite loops are not supported |
9957 | if (TBB == LoopBB && FBB == LoopBB) |
9958 | return nullptr; |
9959 | |
9960 | // Must be conditional branch |
9961 | if (TBB != LoopBB && FBB == nullptr) |
9962 | return nullptr; |
9963 | |
9964 | assert((TBB == LoopBB || FBB == LoopBB) && |
9965 | "The Loop must be a single-basic-block loop" ); |
9966 | |
9967 | MachineInstr *CondBranch = &*LoopBB->getFirstTerminator(); |
9968 | const TargetRegisterInfo &TRI = getRegisterInfo(); |
9969 | |
9970 | if (CondBranch->getOpcode() != AArch64::Bcc) |
9971 | return nullptr; |
9972 | |
9973 | // Normalization for createTripCountGreaterCondition() |
9974 | if (TBB == LoopBB) |
9975 | reverseBranchCondition(Cond); |
9976 | |
9977 | MachineInstr *Comp = nullptr; |
9978 | unsigned CompCounterOprNum = 0; |
9979 | for (MachineInstr &MI : reverse(C&: *LoopBB)) { |
9980 | if (MI.modifiesRegister(Reg: AArch64::NZCV, TRI: &TRI)) { |
9981 | // Guarantee that the compare is SUBS/ADDS/WHILEXX and that one of the |
9982 | // operands is a loop invariant value |
9983 | |
9984 | switch (MI.getOpcode()) { |
9985 | case AArch64::SUBSXri: |
9986 | case AArch64::SUBSWri: |
9987 | case AArch64::ADDSXri: |
9988 | case AArch64::ADDSWri: |
9989 | Comp = &MI; |
9990 | CompCounterOprNum = 1; |
9991 | break; |
9992 | case AArch64::ADDSWrr: |
9993 | case AArch64::ADDSXrr: |
9994 | case AArch64::SUBSWrr: |
9995 | case AArch64::SUBSXrr: |
9996 | Comp = &MI; |
9997 | break; |
9998 | default: |
9999 | if (isWhileOpcode(Opc: MI.getOpcode())) { |
10000 | Comp = &MI; |
10001 | break; |
10002 | } |
10003 | return nullptr; |
10004 | } |
10005 | |
10006 | if (CompCounterOprNum == 0) { |
10007 | if (isDefinedOutside(Reg: Comp->getOperand(i: 1).getReg(), BB: LoopBB)) |
10008 | CompCounterOprNum = 2; |
10009 | else if (isDefinedOutside(Reg: Comp->getOperand(i: 2).getReg(), BB: LoopBB)) |
10010 | CompCounterOprNum = 1; |
10011 | else |
10012 | return nullptr; |
10013 | } |
10014 | break; |
10015 | } |
10016 | } |
10017 | if (!Comp) |
10018 | return nullptr; |
10019 | |
10020 | MachineInstr *Update = nullptr; |
10021 | Register Init; |
10022 | bool IsUpdatePriorComp; |
10023 | unsigned UpdateCounterOprNum; |
10024 | if (!getIndVarInfo(Reg: Comp->getOperand(i: CompCounterOprNum).getReg(), LoopBB, |
10025 | UpdateInst&: Update, UpdateCounterOprNum, InitReg&: Init, IsUpdatePriorComp)) |
10026 | return nullptr; |
10027 | |
10028 | return std::make_unique<AArch64PipelinerLoopInfo>( |
10029 | args&: LoopBB, args&: CondBranch, args&: Comp, args&: CompCounterOprNum, args&: Update, args&: UpdateCounterOprNum, |
10030 | args&: Init, args&: IsUpdatePriorComp, args&: Cond); |
10031 | } |
10032 | |
10033 | #define GET_INSTRINFO_HELPERS |
10034 | #define GET_INSTRMAP_INFO |
10035 | #include "AArch64GenInstrInfo.inc" |
10036 | |