1 | //===-- SystemZInstrInfo.cpp - SystemZ 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 SystemZ implementation of the TargetInstrInfo class. |
10 | // |
11 | //===----------------------------------------------------------------------===// |
12 | |
13 | #include "SystemZInstrInfo.h" |
14 | #include "MCTargetDesc/SystemZMCTargetDesc.h" |
15 | #include "SystemZ.h" |
16 | #include "SystemZInstrBuilder.h" |
17 | #include "SystemZSubtarget.h" |
18 | #include "llvm/ADT/Statistic.h" |
19 | #include "llvm/CodeGen/LiveInterval.h" |
20 | #include "llvm/CodeGen/LiveIntervals.h" |
21 | #include "llvm/CodeGen/LiveRegUnits.h" |
22 | #include "llvm/CodeGen/LiveVariables.h" |
23 | #include "llvm/CodeGen/MachineBasicBlock.h" |
24 | #include "llvm/CodeGen/MachineFrameInfo.h" |
25 | #include "llvm/CodeGen/MachineFunction.h" |
26 | #include "llvm/CodeGen/MachineInstr.h" |
27 | #include "llvm/CodeGen/MachineMemOperand.h" |
28 | #include "llvm/CodeGen/MachineOperand.h" |
29 | #include "llvm/CodeGen/MachineRegisterInfo.h" |
30 | #include "llvm/CodeGen/SlotIndexes.h" |
31 | #include "llvm/CodeGen/StackMaps.h" |
32 | #include "llvm/CodeGen/TargetInstrInfo.h" |
33 | #include "llvm/CodeGen/TargetSubtargetInfo.h" |
34 | #include "llvm/CodeGen/VirtRegMap.h" |
35 | #include "llvm/MC/MCInstrDesc.h" |
36 | #include "llvm/MC/MCRegisterInfo.h" |
37 | #include "llvm/Support/BranchProbability.h" |
38 | #include "llvm/Support/ErrorHandling.h" |
39 | #include "llvm/Support/MathExtras.h" |
40 | #include "llvm/Target/TargetMachine.h" |
41 | #include <cassert> |
42 | #include <cstdint> |
43 | #include <iterator> |
44 | |
45 | using namespace llvm; |
46 | |
47 | #define GET_INSTRINFO_CTOR_DTOR |
48 | #define GET_INSTRMAP_INFO |
49 | #include "SystemZGenInstrInfo.inc" |
50 | |
51 | #define DEBUG_TYPE "systemz-II" |
52 | |
53 | // Return a mask with Count low bits set. |
54 | static uint64_t allOnes(unsigned int Count) { |
55 | return Count == 0 ? 0 : (uint64_t(1) << (Count - 1) << 1) - 1; |
56 | } |
57 | |
58 | // Pin the vtable to this file. |
59 | void SystemZInstrInfo::anchor() {} |
60 | |
61 | SystemZInstrInfo::SystemZInstrInfo(SystemZSubtarget &sti) |
62 | : SystemZGenInstrInfo(-1, -1), |
63 | RI(sti.getSpecialRegisters()->getReturnFunctionAddressRegister()), |
64 | STI(sti) {} |
65 | |
66 | // MI is a 128-bit load or store. Split it into two 64-bit loads or stores, |
67 | // each having the opcode given by NewOpcode. |
68 | void SystemZInstrInfo::splitMove(MachineBasicBlock::iterator MI, |
69 | unsigned NewOpcode) const { |
70 | MachineBasicBlock *MBB = MI->getParent(); |
71 | MachineFunction &MF = *MBB->getParent(); |
72 | |
73 | // Get two load or store instructions. Use the original instruction for |
74 | // one of them and create a clone for the other. |
75 | MachineInstr *HighPartMI = MF.CloneMachineInstr(Orig: &*MI); |
76 | MachineInstr *LowPartMI = &*MI; |
77 | MBB->insert(I: LowPartMI, MI: HighPartMI); |
78 | |
79 | // Set up the two 64-bit registers and remember super reg and its flags. |
80 | MachineOperand &HighRegOp = HighPartMI->getOperand(i: 0); |
81 | MachineOperand &LowRegOp = LowPartMI->getOperand(i: 0); |
82 | Register Reg128 = LowRegOp.getReg(); |
83 | unsigned Reg128Killed = getKillRegState(B: LowRegOp.isKill()); |
84 | unsigned Reg128Undef = getUndefRegState(B: LowRegOp.isUndef()); |
85 | HighRegOp.setReg(RI.getSubReg(Reg: HighRegOp.getReg(), Idx: SystemZ::subreg_h64)); |
86 | LowRegOp.setReg(RI.getSubReg(Reg: LowRegOp.getReg(), Idx: SystemZ::subreg_l64)); |
87 | |
88 | // The address in the first (high) instruction is already correct. |
89 | // Adjust the offset in the second (low) instruction. |
90 | MachineOperand &HighOffsetOp = HighPartMI->getOperand(i: 2); |
91 | MachineOperand &LowOffsetOp = LowPartMI->getOperand(i: 2); |
92 | LowOffsetOp.setImm(LowOffsetOp.getImm() + 8); |
93 | |
94 | // Set the opcodes. |
95 | unsigned HighOpcode = getOpcodeForOffset(Opcode: NewOpcode, Offset: HighOffsetOp.getImm()); |
96 | unsigned LowOpcode = getOpcodeForOffset(Opcode: NewOpcode, Offset: LowOffsetOp.getImm()); |
97 | assert(HighOpcode && LowOpcode && "Both offsets should be in range" ); |
98 | HighPartMI->setDesc(get(Opcode: HighOpcode)); |
99 | LowPartMI->setDesc(get(Opcode: LowOpcode)); |
100 | |
101 | MachineInstr *FirstMI = HighPartMI; |
102 | if (MI->mayStore()) { |
103 | FirstMI->getOperand(i: 0).setIsKill(false); |
104 | // Add implicit uses of the super register in case one of the subregs is |
105 | // undefined. We could track liveness and skip storing an undefined |
106 | // subreg, but this is hopefully rare (discovered with llvm-stress). |
107 | // If Reg128 was killed, set kill flag on MI. |
108 | unsigned Reg128UndefImpl = (Reg128Undef | RegState::Implicit); |
109 | MachineInstrBuilder(MF, HighPartMI).addReg(RegNo: Reg128, flags: Reg128UndefImpl); |
110 | MachineInstrBuilder(MF, LowPartMI).addReg(RegNo: Reg128, flags: (Reg128UndefImpl | Reg128Killed)); |
111 | } else { |
112 | // If HighPartMI clobbers any of the address registers, it needs to come |
113 | // after LowPartMI. |
114 | auto overlapsAddressReg = [&](Register Reg) -> bool { |
115 | return RI.regsOverlap(RegA: Reg, RegB: MI->getOperand(i: 1).getReg()) || |
116 | RI.regsOverlap(RegA: Reg, RegB: MI->getOperand(i: 3).getReg()); |
117 | }; |
118 | if (overlapsAddressReg(HighRegOp.getReg())) { |
119 | assert(!overlapsAddressReg(LowRegOp.getReg()) && |
120 | "Both loads clobber address!" ); |
121 | MBB->splice(Where: HighPartMI, Other: MBB, From: LowPartMI); |
122 | FirstMI = LowPartMI; |
123 | } |
124 | } |
125 | |
126 | // Clear the kill flags on the address registers in the first instruction. |
127 | FirstMI->getOperand(i: 1).setIsKill(false); |
128 | FirstMI->getOperand(i: 3).setIsKill(false); |
129 | } |
130 | |
131 | // Split ADJDYNALLOC instruction MI. |
132 | void SystemZInstrInfo::splitAdjDynAlloc(MachineBasicBlock::iterator MI) const { |
133 | MachineBasicBlock *MBB = MI->getParent(); |
134 | MachineFunction &MF = *MBB->getParent(); |
135 | MachineFrameInfo &MFFrame = MF.getFrameInfo(); |
136 | MachineOperand &OffsetMO = MI->getOperand(i: 2); |
137 | SystemZCallingConventionRegisters *Regs = STI.getSpecialRegisters(); |
138 | |
139 | uint64_t Offset = (MFFrame.getMaxCallFrameSize() + |
140 | Regs->getCallFrameSize() + |
141 | Regs->getStackPointerBias() + |
142 | OffsetMO.getImm()); |
143 | unsigned NewOpcode = getOpcodeForOffset(Opcode: SystemZ::LA, Offset); |
144 | assert(NewOpcode && "No support for huge argument lists yet" ); |
145 | MI->setDesc(get(Opcode: NewOpcode)); |
146 | OffsetMO.setImm(Offset); |
147 | } |
148 | |
149 | // MI is an RI-style pseudo instruction. Replace it with LowOpcode |
150 | // if the first operand is a low GR32 and HighOpcode if the first operand |
151 | // is a high GR32. ConvertHigh is true if LowOpcode takes a signed operand |
152 | // and HighOpcode takes an unsigned 32-bit operand. In those cases, |
153 | // MI has the same kind of operand as LowOpcode, so needs to be converted |
154 | // if HighOpcode is used. |
155 | void SystemZInstrInfo::expandRIPseudo(MachineInstr &MI, unsigned LowOpcode, |
156 | unsigned HighOpcode, |
157 | bool ConvertHigh) const { |
158 | Register Reg = MI.getOperand(i: 0).getReg(); |
159 | bool IsHigh = SystemZ::isHighReg(Reg); |
160 | MI.setDesc(get(Opcode: IsHigh ? HighOpcode : LowOpcode)); |
161 | if (IsHigh && ConvertHigh) |
162 | MI.getOperand(i: 1).setImm(uint32_t(MI.getOperand(i: 1).getImm())); |
163 | } |
164 | |
165 | // MI is a three-operand RIE-style pseudo instruction. Replace it with |
166 | // LowOpcodeK if the registers are both low GR32s, otherwise use a move |
167 | // followed by HighOpcode or LowOpcode, depending on whether the target |
168 | // is a high or low GR32. |
169 | void SystemZInstrInfo::expandRIEPseudo(MachineInstr &MI, unsigned LowOpcode, |
170 | unsigned LowOpcodeK, |
171 | unsigned HighOpcode) const { |
172 | Register DestReg = MI.getOperand(i: 0).getReg(); |
173 | Register SrcReg = MI.getOperand(i: 1).getReg(); |
174 | bool DestIsHigh = SystemZ::isHighReg(Reg: DestReg); |
175 | bool SrcIsHigh = SystemZ::isHighReg(Reg: SrcReg); |
176 | if (!DestIsHigh && !SrcIsHigh) |
177 | MI.setDesc(get(Opcode: LowOpcodeK)); |
178 | else { |
179 | if (DestReg != SrcReg) { |
180 | emitGRX32Move(MBB&: *MI.getParent(), MBBI: MI, DL: MI.getDebugLoc(), DestReg, SrcReg, |
181 | LowLowOpcode: SystemZ::LR, Size: 32, KillSrc: MI.getOperand(i: 1).isKill(), |
182 | UndefSrc: MI.getOperand(i: 1).isUndef()); |
183 | MI.getOperand(i: 1).setReg(DestReg); |
184 | } |
185 | MI.setDesc(get(Opcode: DestIsHigh ? HighOpcode : LowOpcode)); |
186 | MI.tieOperands(DefIdx: 0, UseIdx: 1); |
187 | } |
188 | } |
189 | |
190 | // MI is an RXY-style pseudo instruction. Replace it with LowOpcode |
191 | // if the first operand is a low GR32 and HighOpcode if the first operand |
192 | // is a high GR32. |
193 | void SystemZInstrInfo::expandRXYPseudo(MachineInstr &MI, unsigned LowOpcode, |
194 | unsigned HighOpcode) const { |
195 | Register Reg = MI.getOperand(i: 0).getReg(); |
196 | unsigned Opcode = getOpcodeForOffset( |
197 | Opcode: SystemZ::isHighReg(Reg) ? HighOpcode : LowOpcode, |
198 | Offset: MI.getOperand(i: 2).getImm()); |
199 | MI.setDesc(get(Opcode)); |
200 | } |
201 | |
202 | // MI is a load-on-condition pseudo instruction with a single register |
203 | // (source or destination) operand. Replace it with LowOpcode if the |
204 | // register is a low GR32 and HighOpcode if the register is a high GR32. |
205 | void SystemZInstrInfo::expandLOCPseudo(MachineInstr &MI, unsigned LowOpcode, |
206 | unsigned HighOpcode) const { |
207 | Register Reg = MI.getOperand(i: 0).getReg(); |
208 | unsigned Opcode = SystemZ::isHighReg(Reg) ? HighOpcode : LowOpcode; |
209 | MI.setDesc(get(Opcode)); |
210 | } |
211 | |
212 | // MI is an RR-style pseudo instruction that zero-extends the low Size bits |
213 | // of one GRX32 into another. Replace it with LowOpcode if both operands |
214 | // are low registers, otherwise use RISB[LH]G. |
215 | void SystemZInstrInfo::expandZExtPseudo(MachineInstr &MI, unsigned LowOpcode, |
216 | unsigned Size) const { |
217 | MachineInstrBuilder MIB = |
218 | emitGRX32Move(MBB&: *MI.getParent(), MBBI: MI, DL: MI.getDebugLoc(), |
219 | DestReg: MI.getOperand(i: 0).getReg(), SrcReg: MI.getOperand(i: 1).getReg(), LowLowOpcode: LowOpcode, |
220 | Size, KillSrc: MI.getOperand(i: 1).isKill(), UndefSrc: MI.getOperand(i: 1).isUndef()); |
221 | |
222 | // Keep the remaining operands as-is. |
223 | for (const MachineOperand &MO : llvm::drop_begin(RangeOrContainer: MI.operands(), N: 2)) |
224 | MIB.add(MO); |
225 | |
226 | MI.eraseFromParent(); |
227 | } |
228 | |
229 | void SystemZInstrInfo::expandLoadStackGuard(MachineInstr *MI) const { |
230 | MachineBasicBlock *MBB = MI->getParent(); |
231 | MachineFunction &MF = *MBB->getParent(); |
232 | const Register Reg64 = MI->getOperand(i: 0).getReg(); |
233 | const Register Reg32 = RI.getSubReg(Reg: Reg64, Idx: SystemZ::subreg_l32); |
234 | |
235 | // EAR can only load the low subregister so us a shift for %a0 to produce |
236 | // the GR containing %a0 and %a1. |
237 | |
238 | // ear <reg>, %a0 |
239 | BuildMI(BB&: *MBB, I: MI, MIMD: MI->getDebugLoc(), MCID: get(Opcode: SystemZ::EAR), DestReg: Reg32) |
240 | .addReg(RegNo: SystemZ::A0) |
241 | .addReg(RegNo: Reg64, flags: RegState::ImplicitDefine); |
242 | |
243 | // sllg <reg>, <reg>, 32 |
244 | BuildMI(BB&: *MBB, I: MI, MIMD: MI->getDebugLoc(), MCID: get(Opcode: SystemZ::SLLG), DestReg: Reg64) |
245 | .addReg(RegNo: Reg64) |
246 | .addReg(RegNo: 0) |
247 | .addImm(Val: 32); |
248 | |
249 | // ear <reg>, %a1 |
250 | BuildMI(BB&: *MBB, I: MI, MIMD: MI->getDebugLoc(), MCID: get(Opcode: SystemZ::EAR), DestReg: Reg32) |
251 | .addReg(RegNo: SystemZ::A1); |
252 | |
253 | // lg <reg>, 40(<reg>) |
254 | MI->setDesc(get(Opcode: SystemZ::LG)); |
255 | MachineInstrBuilder(MF, MI).addReg(RegNo: Reg64).addImm(Val: 40).addReg(RegNo: 0); |
256 | } |
257 | |
258 | // Emit a zero-extending move from 32-bit GPR SrcReg to 32-bit GPR |
259 | // DestReg before MBBI in MBB. Use LowLowOpcode when both DestReg and SrcReg |
260 | // are low registers, otherwise use RISB[LH]G. Size is the number of bits |
261 | // taken from the low end of SrcReg (8 for LLCR, 16 for LLHR and 32 for LR). |
262 | // KillSrc is true if this move is the last use of SrcReg. |
263 | MachineInstrBuilder |
264 | SystemZInstrInfo::emitGRX32Move(MachineBasicBlock &MBB, |
265 | MachineBasicBlock::iterator MBBI, |
266 | const DebugLoc &DL, unsigned DestReg, |
267 | unsigned SrcReg, unsigned LowLowOpcode, |
268 | unsigned Size, bool KillSrc, |
269 | bool UndefSrc) const { |
270 | unsigned Opcode; |
271 | bool DestIsHigh = SystemZ::isHighReg(Reg: DestReg); |
272 | bool SrcIsHigh = SystemZ::isHighReg(Reg: SrcReg); |
273 | if (DestIsHigh && SrcIsHigh) |
274 | Opcode = SystemZ::RISBHH; |
275 | else if (DestIsHigh && !SrcIsHigh) |
276 | Opcode = SystemZ::RISBHL; |
277 | else if (!DestIsHigh && SrcIsHigh) |
278 | Opcode = SystemZ::RISBLH; |
279 | else { |
280 | return BuildMI(BB&: MBB, I: MBBI, MIMD: DL, MCID: get(Opcode: LowLowOpcode), DestReg) |
281 | .addReg(RegNo: SrcReg, flags: getKillRegState(B: KillSrc) | getUndefRegState(B: UndefSrc)); |
282 | } |
283 | unsigned Rotate = (DestIsHigh != SrcIsHigh ? 32 : 0); |
284 | return BuildMI(BB&: MBB, I: MBBI, MIMD: DL, MCID: get(Opcode), DestReg) |
285 | .addReg(RegNo: DestReg, flags: RegState::Undef) |
286 | .addReg(RegNo: SrcReg, flags: getKillRegState(B: KillSrc) | getUndefRegState(B: UndefSrc)) |
287 | .addImm(Val: 32 - Size).addImm(Val: 128 + 31).addImm(Val: Rotate); |
288 | } |
289 | |
290 | MachineInstr *SystemZInstrInfo::commuteInstructionImpl(MachineInstr &MI, |
291 | bool NewMI, |
292 | unsigned OpIdx1, |
293 | unsigned OpIdx2) const { |
294 | auto cloneIfNew = [NewMI](MachineInstr &MI) -> MachineInstr & { |
295 | if (NewMI) |
296 | return *MI.getParent()->getParent()->CloneMachineInstr(Orig: &MI); |
297 | return MI; |
298 | }; |
299 | |
300 | switch (MI.getOpcode()) { |
301 | case SystemZ::SELRMux: |
302 | case SystemZ::SELFHR: |
303 | case SystemZ::SELR: |
304 | case SystemZ::SELGR: |
305 | case SystemZ::LOCRMux: |
306 | case SystemZ::LOCFHR: |
307 | case SystemZ::LOCR: |
308 | case SystemZ::LOCGR: { |
309 | auto &WorkingMI = cloneIfNew(MI); |
310 | // Invert condition. |
311 | unsigned CCValid = WorkingMI.getOperand(i: 3).getImm(); |
312 | unsigned CCMask = WorkingMI.getOperand(i: 4).getImm(); |
313 | WorkingMI.getOperand(i: 4).setImm(CCMask ^ CCValid); |
314 | return TargetInstrInfo::commuteInstructionImpl(MI&: WorkingMI, /*NewMI=*/false, |
315 | OpIdx1, OpIdx2); |
316 | } |
317 | default: |
318 | return TargetInstrInfo::commuteInstructionImpl(MI, NewMI, OpIdx1, OpIdx2); |
319 | } |
320 | } |
321 | |
322 | // If MI is a simple load or store for a frame object, return the register |
323 | // it loads or stores and set FrameIndex to the index of the frame object. |
324 | // Return 0 otherwise. |
325 | // |
326 | // Flag is SimpleBDXLoad for loads and SimpleBDXStore for stores. |
327 | static int isSimpleMove(const MachineInstr &MI, int &FrameIndex, |
328 | unsigned Flag) { |
329 | const MCInstrDesc &MCID = MI.getDesc(); |
330 | if ((MCID.TSFlags & Flag) && MI.getOperand(i: 1).isFI() && |
331 | MI.getOperand(i: 2).getImm() == 0 && MI.getOperand(i: 3).getReg() == 0) { |
332 | FrameIndex = MI.getOperand(i: 1).getIndex(); |
333 | return MI.getOperand(i: 0).getReg(); |
334 | } |
335 | return 0; |
336 | } |
337 | |
338 | Register SystemZInstrInfo::isLoadFromStackSlot(const MachineInstr &MI, |
339 | int &FrameIndex) const { |
340 | return isSimpleMove(MI, FrameIndex, Flag: SystemZII::SimpleBDXLoad); |
341 | } |
342 | |
343 | Register SystemZInstrInfo::isStoreToStackSlot(const MachineInstr &MI, |
344 | int &FrameIndex) const { |
345 | return isSimpleMove(MI, FrameIndex, Flag: SystemZII::SimpleBDXStore); |
346 | } |
347 | |
348 | bool SystemZInstrInfo::isStackSlotCopy(const MachineInstr &MI, |
349 | int &DestFrameIndex, |
350 | int &SrcFrameIndex) const { |
351 | // Check for MVC 0(Length,FI1),0(FI2) |
352 | const MachineFrameInfo &MFI = MI.getParent()->getParent()->getFrameInfo(); |
353 | if (MI.getOpcode() != SystemZ::MVC || !MI.getOperand(i: 0).isFI() || |
354 | MI.getOperand(i: 1).getImm() != 0 || !MI.getOperand(i: 3).isFI() || |
355 | MI.getOperand(i: 4).getImm() != 0) |
356 | return false; |
357 | |
358 | // Check that Length covers the full slots. |
359 | int64_t Length = MI.getOperand(i: 2).getImm(); |
360 | unsigned FI1 = MI.getOperand(i: 0).getIndex(); |
361 | unsigned FI2 = MI.getOperand(i: 3).getIndex(); |
362 | if (MFI.getObjectSize(ObjectIdx: FI1) != Length || |
363 | MFI.getObjectSize(ObjectIdx: FI2) != Length) |
364 | return false; |
365 | |
366 | DestFrameIndex = FI1; |
367 | SrcFrameIndex = FI2; |
368 | return true; |
369 | } |
370 | |
371 | bool SystemZInstrInfo::analyzeBranch(MachineBasicBlock &MBB, |
372 | MachineBasicBlock *&TBB, |
373 | MachineBasicBlock *&FBB, |
374 | SmallVectorImpl<MachineOperand> &Cond, |
375 | bool AllowModify) const { |
376 | // Most of the code and comments here are boilerplate. |
377 | |
378 | // Start from the bottom of the block and work up, examining the |
379 | // terminator instructions. |
380 | MachineBasicBlock::iterator I = MBB.end(); |
381 | while (I != MBB.begin()) { |
382 | --I; |
383 | if (I->isDebugInstr()) |
384 | continue; |
385 | |
386 | // Working from the bottom, when we see a non-terminator instruction, we're |
387 | // done. |
388 | if (!isUnpredicatedTerminator(MI: *I)) |
389 | break; |
390 | |
391 | // A terminator that isn't a branch can't easily be handled by this |
392 | // analysis. |
393 | if (!I->isBranch()) |
394 | return true; |
395 | |
396 | // Can't handle indirect branches. |
397 | SystemZII::Branch Branch(getBranchInfo(MI: *I)); |
398 | if (!Branch.hasMBBTarget()) |
399 | return true; |
400 | |
401 | // Punt on compound branches. |
402 | if (Branch.Type != SystemZII::BranchNormal) |
403 | return true; |
404 | |
405 | if (Branch.CCMask == SystemZ::CCMASK_ANY) { |
406 | // Handle unconditional branches. |
407 | if (!AllowModify) { |
408 | TBB = Branch.getMBBTarget(); |
409 | continue; |
410 | } |
411 | |
412 | // If the block has any instructions after a JMP, delete them. |
413 | MBB.erase(I: std::next(x: I), E: MBB.end()); |
414 | |
415 | Cond.clear(); |
416 | FBB = nullptr; |
417 | |
418 | // Delete the JMP if it's equivalent to a fall-through. |
419 | if (MBB.isLayoutSuccessor(MBB: Branch.getMBBTarget())) { |
420 | TBB = nullptr; |
421 | I->eraseFromParent(); |
422 | I = MBB.end(); |
423 | continue; |
424 | } |
425 | |
426 | // TBB is used to indicate the unconditinal destination. |
427 | TBB = Branch.getMBBTarget(); |
428 | continue; |
429 | } |
430 | |
431 | // Working from the bottom, handle the first conditional branch. |
432 | if (Cond.empty()) { |
433 | // FIXME: add X86-style branch swap |
434 | FBB = TBB; |
435 | TBB = Branch.getMBBTarget(); |
436 | Cond.push_back(Elt: MachineOperand::CreateImm(Val: Branch.CCValid)); |
437 | Cond.push_back(Elt: MachineOperand::CreateImm(Val: Branch.CCMask)); |
438 | continue; |
439 | } |
440 | |
441 | // Handle subsequent conditional branches. |
442 | assert(Cond.size() == 2 && TBB && "Should have seen a conditional branch" ); |
443 | |
444 | // Only handle the case where all conditional branches branch to the same |
445 | // destination. |
446 | if (TBB != Branch.getMBBTarget()) |
447 | return true; |
448 | |
449 | // If the conditions are the same, we can leave them alone. |
450 | unsigned OldCCValid = Cond[0].getImm(); |
451 | unsigned OldCCMask = Cond[1].getImm(); |
452 | if (OldCCValid == Branch.CCValid && OldCCMask == Branch.CCMask) |
453 | continue; |
454 | |
455 | // FIXME: Try combining conditions like X86 does. Should be easy on Z! |
456 | return false; |
457 | } |
458 | |
459 | return false; |
460 | } |
461 | |
462 | unsigned SystemZInstrInfo::removeBranch(MachineBasicBlock &MBB, |
463 | int *BytesRemoved) const { |
464 | assert(!BytesRemoved && "code size not handled" ); |
465 | |
466 | // Most of the code and comments here are boilerplate. |
467 | MachineBasicBlock::iterator I = MBB.end(); |
468 | unsigned Count = 0; |
469 | |
470 | while (I != MBB.begin()) { |
471 | --I; |
472 | if (I->isDebugInstr()) |
473 | continue; |
474 | if (!I->isBranch()) |
475 | break; |
476 | if (!getBranchInfo(MI: *I).hasMBBTarget()) |
477 | break; |
478 | // Remove the branch. |
479 | I->eraseFromParent(); |
480 | I = MBB.end(); |
481 | ++Count; |
482 | } |
483 | |
484 | return Count; |
485 | } |
486 | |
487 | bool SystemZInstrInfo:: |
488 | reverseBranchCondition(SmallVectorImpl<MachineOperand> &Cond) const { |
489 | assert(Cond.size() == 2 && "Invalid condition" ); |
490 | Cond[1].setImm(Cond[1].getImm() ^ Cond[0].getImm()); |
491 | return false; |
492 | } |
493 | |
494 | unsigned SystemZInstrInfo::insertBranch(MachineBasicBlock &MBB, |
495 | MachineBasicBlock *TBB, |
496 | MachineBasicBlock *FBB, |
497 | ArrayRef<MachineOperand> Cond, |
498 | const DebugLoc &DL, |
499 | int *BytesAdded) const { |
500 | // In this function we output 32-bit branches, which should always |
501 | // have enough range. They can be shortened and relaxed by later code |
502 | // in the pipeline, if desired. |
503 | |
504 | // Shouldn't be a fall through. |
505 | assert(TBB && "insertBranch must not be told to insert a fallthrough" ); |
506 | assert((Cond.size() == 2 || Cond.size() == 0) && |
507 | "SystemZ branch conditions have one component!" ); |
508 | assert(!BytesAdded && "code size not handled" ); |
509 | |
510 | if (Cond.empty()) { |
511 | // Unconditional branch? |
512 | assert(!FBB && "Unconditional branch with multiple successors!" ); |
513 | BuildMI(BB: &MBB, MIMD: DL, MCID: get(Opcode: SystemZ::J)).addMBB(MBB: TBB); |
514 | return 1; |
515 | } |
516 | |
517 | // Conditional branch. |
518 | unsigned Count = 0; |
519 | unsigned CCValid = Cond[0].getImm(); |
520 | unsigned CCMask = Cond[1].getImm(); |
521 | BuildMI(BB: &MBB, MIMD: DL, MCID: get(Opcode: SystemZ::BRC)) |
522 | .addImm(Val: CCValid).addImm(Val: CCMask).addMBB(MBB: TBB); |
523 | ++Count; |
524 | |
525 | if (FBB) { |
526 | // Two-way Conditional branch. Insert the second branch. |
527 | BuildMI(BB: &MBB, MIMD: DL, MCID: get(Opcode: SystemZ::J)).addMBB(MBB: FBB); |
528 | ++Count; |
529 | } |
530 | return Count; |
531 | } |
532 | |
533 | bool SystemZInstrInfo::analyzeCompare(const MachineInstr &MI, Register &SrcReg, |
534 | Register &SrcReg2, int64_t &Mask, |
535 | int64_t &Value) const { |
536 | assert(MI.isCompare() && "Caller should have checked for a comparison" ); |
537 | |
538 | if (MI.getNumExplicitOperands() == 2 && MI.getOperand(i: 0).isReg() && |
539 | MI.getOperand(i: 1).isImm()) { |
540 | SrcReg = MI.getOperand(i: 0).getReg(); |
541 | SrcReg2 = 0; |
542 | Value = MI.getOperand(i: 1).getImm(); |
543 | Mask = ~0; |
544 | return true; |
545 | } |
546 | |
547 | return false; |
548 | } |
549 | |
550 | bool SystemZInstrInfo::canInsertSelect(const MachineBasicBlock &MBB, |
551 | ArrayRef<MachineOperand> Pred, |
552 | Register DstReg, Register TrueReg, |
553 | Register FalseReg, int &CondCycles, |
554 | int &TrueCycles, |
555 | int &FalseCycles) const { |
556 | // Not all subtargets have LOCR instructions. |
557 | if (!STI.hasLoadStoreOnCond()) |
558 | return false; |
559 | if (Pred.size() != 2) |
560 | return false; |
561 | |
562 | // Check register classes. |
563 | const MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo(); |
564 | const TargetRegisterClass *RC = |
565 | RI.getCommonSubClass(A: MRI.getRegClass(Reg: TrueReg), B: MRI.getRegClass(Reg: FalseReg)); |
566 | if (!RC) |
567 | return false; |
568 | |
569 | // We have LOCR instructions for 32 and 64 bit general purpose registers. |
570 | if ((STI.hasLoadStoreOnCond2() && |
571 | SystemZ::GRX32BitRegClass.hasSubClassEq(RC)) || |
572 | SystemZ::GR32BitRegClass.hasSubClassEq(RC) || |
573 | SystemZ::GR64BitRegClass.hasSubClassEq(RC)) { |
574 | CondCycles = 2; |
575 | TrueCycles = 2; |
576 | FalseCycles = 2; |
577 | return true; |
578 | } |
579 | |
580 | // Can't do anything else. |
581 | return false; |
582 | } |
583 | |
584 | void SystemZInstrInfo::insertSelect(MachineBasicBlock &MBB, |
585 | MachineBasicBlock::iterator I, |
586 | const DebugLoc &DL, Register DstReg, |
587 | ArrayRef<MachineOperand> Pred, |
588 | Register TrueReg, |
589 | Register FalseReg) const { |
590 | MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo(); |
591 | const TargetRegisterClass *RC = MRI.getRegClass(Reg: DstReg); |
592 | |
593 | assert(Pred.size() == 2 && "Invalid condition" ); |
594 | unsigned CCValid = Pred[0].getImm(); |
595 | unsigned CCMask = Pred[1].getImm(); |
596 | |
597 | unsigned Opc; |
598 | if (SystemZ::GRX32BitRegClass.hasSubClassEq(RC)) { |
599 | if (STI.hasMiscellaneousExtensions3()) |
600 | Opc = SystemZ::SELRMux; |
601 | else if (STI.hasLoadStoreOnCond2()) |
602 | Opc = SystemZ::LOCRMux; |
603 | else { |
604 | Opc = SystemZ::LOCR; |
605 | MRI.constrainRegClass(Reg: DstReg, RC: &SystemZ::GR32BitRegClass); |
606 | Register TReg = MRI.createVirtualRegister(RegClass: &SystemZ::GR32BitRegClass); |
607 | Register FReg = MRI.createVirtualRegister(RegClass: &SystemZ::GR32BitRegClass); |
608 | BuildMI(BB&: MBB, I, MIMD: DL, MCID: get(Opcode: TargetOpcode::COPY), DestReg: TReg).addReg(RegNo: TrueReg); |
609 | BuildMI(BB&: MBB, I, MIMD: DL, MCID: get(Opcode: TargetOpcode::COPY), DestReg: FReg).addReg(RegNo: FalseReg); |
610 | TrueReg = TReg; |
611 | FalseReg = FReg; |
612 | } |
613 | } else if (SystemZ::GR64BitRegClass.hasSubClassEq(RC)) { |
614 | if (STI.hasMiscellaneousExtensions3()) |
615 | Opc = SystemZ::SELGR; |
616 | else |
617 | Opc = SystemZ::LOCGR; |
618 | } else |
619 | llvm_unreachable("Invalid register class" ); |
620 | |
621 | BuildMI(BB&: MBB, I, MIMD: DL, MCID: get(Opcode: Opc), DestReg: DstReg) |
622 | .addReg(RegNo: FalseReg).addReg(RegNo: TrueReg) |
623 | .addImm(Val: CCValid).addImm(Val: CCMask); |
624 | } |
625 | |
626 | MachineInstr *SystemZInstrInfo::optimizeLoadInstr(MachineInstr &MI, |
627 | const MachineRegisterInfo *MRI, |
628 | Register &FoldAsLoadDefReg, |
629 | MachineInstr *&DefMI) const { |
630 | // Check whether we can move the DefMI load, and that it only has one use. |
631 | DefMI = MRI->getVRegDef(Reg: FoldAsLoadDefReg); |
632 | assert(DefMI); |
633 | bool SawStore = false; |
634 | if (!DefMI->isSafeToMove(AA: nullptr, SawStore) || |
635 | !MRI->hasOneNonDBGUse(RegNo: FoldAsLoadDefReg)) |
636 | return nullptr; |
637 | |
638 | int UseOpIdx = |
639 | MI.findRegisterUseOperandIdx(Reg: FoldAsLoadDefReg, /*TRI=*/nullptr); |
640 | assert(UseOpIdx != -1 && "Expected FoldAsLoadDefReg to be used by MI." ); |
641 | |
642 | // Check whether we can fold the load. |
643 | if (MachineInstr *FoldMI = |
644 | foldMemoryOperand(MI, Ops: {((unsigned)UseOpIdx)}, LoadMI&: *DefMI)) { |
645 | FoldAsLoadDefReg = 0; |
646 | return FoldMI; |
647 | } |
648 | |
649 | return nullptr; |
650 | } |
651 | |
652 | bool SystemZInstrInfo::foldImmediate(MachineInstr &UseMI, MachineInstr &DefMI, |
653 | Register Reg, |
654 | MachineRegisterInfo *MRI) const { |
655 | unsigned DefOpc = DefMI.getOpcode(); |
656 | |
657 | if (DefOpc == SystemZ::VGBM) { |
658 | int64_t ImmVal = DefMI.getOperand(i: 1).getImm(); |
659 | if (ImmVal != 0) // TODO: Handle other values |
660 | return false; |
661 | |
662 | // Fold gr128 = COPY (vr128 VGBM imm) |
663 | // |
664 | // %tmp:gr64 = LGHI 0 |
665 | // to gr128 = REG_SEQUENCE %tmp, %tmp |
666 | assert(DefMI.getOperand(0).getReg() == Reg); |
667 | |
668 | if (!UseMI.isCopy()) |
669 | return false; |
670 | |
671 | Register CopyDstReg = UseMI.getOperand(i: 0).getReg(); |
672 | if (CopyDstReg.isVirtual() && |
673 | MRI->getRegClass(Reg: CopyDstReg) == &SystemZ::GR128BitRegClass && |
674 | MRI->hasOneNonDBGUse(RegNo: Reg)) { |
675 | // TODO: Handle physical registers |
676 | // TODO: Handle gr64 uses with subregister indexes |
677 | // TODO: Should this multi-use cases? |
678 | Register TmpReg = MRI->createVirtualRegister(RegClass: &SystemZ::GR64BitRegClass); |
679 | MachineBasicBlock &MBB = *UseMI.getParent(); |
680 | |
681 | loadImmediate(MBB, MBBI: UseMI.getIterator(), Reg: TmpReg, Value: ImmVal); |
682 | |
683 | UseMI.setDesc(get(Opcode: SystemZ::REG_SEQUENCE)); |
684 | UseMI.getOperand(i: 1).setReg(TmpReg); |
685 | MachineInstrBuilder(*MBB.getParent(), &UseMI) |
686 | .addImm(Val: SystemZ::subreg_h64) |
687 | .addReg(RegNo: TmpReg) |
688 | .addImm(Val: SystemZ::subreg_l64); |
689 | |
690 | if (MRI->use_nodbg_empty(RegNo: Reg)) |
691 | DefMI.eraseFromParent(); |
692 | return true; |
693 | } |
694 | |
695 | return false; |
696 | } |
697 | |
698 | if (DefOpc != SystemZ::LHIMux && DefOpc != SystemZ::LHI && |
699 | DefOpc != SystemZ::LGHI) |
700 | return false; |
701 | if (DefMI.getOperand(i: 0).getReg() != Reg) |
702 | return false; |
703 | int32_t ImmVal = (int32_t)DefMI.getOperand(i: 1).getImm(); |
704 | |
705 | unsigned UseOpc = UseMI.getOpcode(); |
706 | unsigned NewUseOpc; |
707 | unsigned UseIdx; |
708 | int CommuteIdx = -1; |
709 | bool TieOps = false; |
710 | switch (UseOpc) { |
711 | case SystemZ::SELRMux: |
712 | TieOps = true; |
713 | [[fallthrough]]; |
714 | case SystemZ::LOCRMux: |
715 | if (!STI.hasLoadStoreOnCond2()) |
716 | return false; |
717 | NewUseOpc = SystemZ::LOCHIMux; |
718 | if (UseMI.getOperand(i: 2).getReg() == Reg) |
719 | UseIdx = 2; |
720 | else if (UseMI.getOperand(i: 1).getReg() == Reg) |
721 | UseIdx = 2, CommuteIdx = 1; |
722 | else |
723 | return false; |
724 | break; |
725 | case SystemZ::SELGR: |
726 | TieOps = true; |
727 | [[fallthrough]]; |
728 | case SystemZ::LOCGR: |
729 | if (!STI.hasLoadStoreOnCond2()) |
730 | return false; |
731 | NewUseOpc = SystemZ::LOCGHI; |
732 | if (UseMI.getOperand(i: 2).getReg() == Reg) |
733 | UseIdx = 2; |
734 | else if (UseMI.getOperand(i: 1).getReg() == Reg) |
735 | UseIdx = 2, CommuteIdx = 1; |
736 | else |
737 | return false; |
738 | break; |
739 | default: |
740 | return false; |
741 | } |
742 | |
743 | if (CommuteIdx != -1) |
744 | if (!commuteInstruction(MI&: UseMI, NewMI: false, OpIdx1: CommuteIdx, OpIdx2: UseIdx)) |
745 | return false; |
746 | |
747 | bool DeleteDef = MRI->hasOneNonDBGUse(RegNo: Reg); |
748 | UseMI.setDesc(get(Opcode: NewUseOpc)); |
749 | if (TieOps) |
750 | UseMI.tieOperands(DefIdx: 0, UseIdx: 1); |
751 | UseMI.getOperand(i: UseIdx).ChangeToImmediate(ImmVal); |
752 | if (DeleteDef) |
753 | DefMI.eraseFromParent(); |
754 | |
755 | return true; |
756 | } |
757 | |
758 | bool SystemZInstrInfo::isPredicable(const MachineInstr &MI) const { |
759 | unsigned Opcode = MI.getOpcode(); |
760 | if (Opcode == SystemZ::Return || |
761 | Opcode == SystemZ::Return_XPLINK || |
762 | Opcode == SystemZ::Trap || |
763 | Opcode == SystemZ::CallJG || |
764 | Opcode == SystemZ::CallBR) |
765 | return true; |
766 | return false; |
767 | } |
768 | |
769 | bool SystemZInstrInfo:: |
770 | isProfitableToIfCvt(MachineBasicBlock &MBB, |
771 | unsigned NumCycles, unsigned , |
772 | BranchProbability Probability) const { |
773 | // Avoid using conditional returns at the end of a loop (since then |
774 | // we'd need to emit an unconditional branch to the beginning anyway, |
775 | // making the loop body longer). This doesn't apply for low-probability |
776 | // loops (eg. compare-and-swap retry), so just decide based on branch |
777 | // probability instead of looping structure. |
778 | // However, since Compare and Trap instructions cost the same as a regular |
779 | // Compare instruction, we should allow the if conversion to convert this |
780 | // into a Conditional Compare regardless of the branch probability. |
781 | if (MBB.getLastNonDebugInstr()->getOpcode() != SystemZ::Trap && |
782 | MBB.succ_empty() && Probability < BranchProbability(1, 8)) |
783 | return false; |
784 | // For now only convert single instructions. |
785 | return NumCycles == 1; |
786 | } |
787 | |
788 | bool SystemZInstrInfo:: |
789 | isProfitableToIfCvt(MachineBasicBlock &TMBB, |
790 | unsigned NumCyclesT, unsigned , |
791 | MachineBasicBlock &FMBB, |
792 | unsigned NumCyclesF, unsigned , |
793 | BranchProbability Probability) const { |
794 | // For now avoid converting mutually-exclusive cases. |
795 | return false; |
796 | } |
797 | |
798 | bool SystemZInstrInfo:: |
799 | isProfitableToDupForIfCvt(MachineBasicBlock &MBB, unsigned NumCycles, |
800 | BranchProbability Probability) const { |
801 | // For now only duplicate single instructions. |
802 | return NumCycles == 1; |
803 | } |
804 | |
805 | bool SystemZInstrInfo::PredicateInstruction( |
806 | MachineInstr &MI, ArrayRef<MachineOperand> Pred) const { |
807 | assert(Pred.size() == 2 && "Invalid condition" ); |
808 | unsigned CCValid = Pred[0].getImm(); |
809 | unsigned CCMask = Pred[1].getImm(); |
810 | assert(CCMask > 0 && CCMask < 15 && "Invalid predicate" ); |
811 | unsigned Opcode = MI.getOpcode(); |
812 | if (Opcode == SystemZ::Trap) { |
813 | MI.setDesc(get(Opcode: SystemZ::CondTrap)); |
814 | MachineInstrBuilder(*MI.getParent()->getParent(), MI) |
815 | .addImm(Val: CCValid).addImm(Val: CCMask) |
816 | .addReg(RegNo: SystemZ::CC, flags: RegState::Implicit); |
817 | return true; |
818 | } |
819 | if (Opcode == SystemZ::Return || Opcode == SystemZ::Return_XPLINK) { |
820 | MI.setDesc(get(Opcode: Opcode == SystemZ::Return ? SystemZ::CondReturn |
821 | : SystemZ::CondReturn_XPLINK)); |
822 | MachineInstrBuilder(*MI.getParent()->getParent(), MI) |
823 | .addImm(Val: CCValid) |
824 | .addImm(Val: CCMask) |
825 | .addReg(RegNo: SystemZ::CC, flags: RegState::Implicit); |
826 | return true; |
827 | } |
828 | if (Opcode == SystemZ::CallJG) { |
829 | MachineOperand FirstOp = MI.getOperand(i: 0); |
830 | const uint32_t *RegMask = MI.getOperand(i: 1).getRegMask(); |
831 | MI.removeOperand(OpNo: 1); |
832 | MI.removeOperand(OpNo: 0); |
833 | MI.setDesc(get(Opcode: SystemZ::CallBRCL)); |
834 | MachineInstrBuilder(*MI.getParent()->getParent(), MI) |
835 | .addImm(Val: CCValid) |
836 | .addImm(Val: CCMask) |
837 | .add(MO: FirstOp) |
838 | .addRegMask(Mask: RegMask) |
839 | .addReg(RegNo: SystemZ::CC, flags: RegState::Implicit); |
840 | return true; |
841 | } |
842 | if (Opcode == SystemZ::CallBR) { |
843 | MachineOperand Target = MI.getOperand(i: 0); |
844 | const uint32_t *RegMask = MI.getOperand(i: 1).getRegMask(); |
845 | MI.removeOperand(OpNo: 1); |
846 | MI.removeOperand(OpNo: 0); |
847 | MI.setDesc(get(Opcode: SystemZ::CallBCR)); |
848 | MachineInstrBuilder(*MI.getParent()->getParent(), MI) |
849 | .addImm(Val: CCValid).addImm(Val: CCMask) |
850 | .add(MO: Target) |
851 | .addRegMask(Mask: RegMask) |
852 | .addReg(RegNo: SystemZ::CC, flags: RegState::Implicit); |
853 | return true; |
854 | } |
855 | return false; |
856 | } |
857 | |
858 | void SystemZInstrInfo::copyPhysReg(MachineBasicBlock &MBB, |
859 | MachineBasicBlock::iterator MBBI, |
860 | const DebugLoc &DL, MCRegister DestReg, |
861 | MCRegister SrcReg, bool KillSrc) const { |
862 | // Split 128-bit GPR moves into two 64-bit moves. Add implicit uses of the |
863 | // super register in case one of the subregs is undefined. |
864 | // This handles ADDR128 too. |
865 | if (SystemZ::GR128BitRegClass.contains(Reg1: DestReg, Reg2: SrcReg)) { |
866 | copyPhysReg(MBB, MBBI, DL, DestReg: RI.getSubReg(Reg: DestReg, Idx: SystemZ::subreg_h64), |
867 | SrcReg: RI.getSubReg(Reg: SrcReg, Idx: SystemZ::subreg_h64), KillSrc); |
868 | MachineInstrBuilder(*MBB.getParent(), std::prev(x: MBBI)) |
869 | .addReg(RegNo: SrcReg, flags: RegState::Implicit); |
870 | copyPhysReg(MBB, MBBI, DL, DestReg: RI.getSubReg(Reg: DestReg, Idx: SystemZ::subreg_l64), |
871 | SrcReg: RI.getSubReg(Reg: SrcReg, Idx: SystemZ::subreg_l64), KillSrc); |
872 | MachineInstrBuilder(*MBB.getParent(), std::prev(x: MBBI)) |
873 | .addReg(RegNo: SrcReg, flags: (getKillRegState(B: KillSrc) | RegState::Implicit)); |
874 | return; |
875 | } |
876 | |
877 | if (SystemZ::GRX32BitRegClass.contains(Reg1: DestReg, Reg2: SrcReg)) { |
878 | emitGRX32Move(MBB, MBBI, DL, DestReg, SrcReg, LowLowOpcode: SystemZ::LR, Size: 32, KillSrc, |
879 | UndefSrc: false); |
880 | return; |
881 | } |
882 | |
883 | // Move 128-bit floating-point values between VR128 and FP128. |
884 | if (SystemZ::VR128BitRegClass.contains(Reg: DestReg) && |
885 | SystemZ::FP128BitRegClass.contains(Reg: SrcReg)) { |
886 | MCRegister SrcRegHi = |
887 | RI.getMatchingSuperReg(Reg: RI.getSubReg(Reg: SrcReg, Idx: SystemZ::subreg_h64), |
888 | SubIdx: SystemZ::subreg_h64, RC: &SystemZ::VR128BitRegClass); |
889 | MCRegister SrcRegLo = |
890 | RI.getMatchingSuperReg(Reg: RI.getSubReg(Reg: SrcReg, Idx: SystemZ::subreg_l64), |
891 | SubIdx: SystemZ::subreg_h64, RC: &SystemZ::VR128BitRegClass); |
892 | |
893 | BuildMI(BB&: MBB, I: MBBI, MIMD: DL, MCID: get(Opcode: SystemZ::VMRHG), DestReg) |
894 | .addReg(RegNo: SrcRegHi, flags: getKillRegState(B: KillSrc)) |
895 | .addReg(RegNo: SrcRegLo, flags: getKillRegState(B: KillSrc)); |
896 | return; |
897 | } |
898 | if (SystemZ::FP128BitRegClass.contains(Reg: DestReg) && |
899 | SystemZ::VR128BitRegClass.contains(Reg: SrcReg)) { |
900 | MCRegister DestRegHi = |
901 | RI.getMatchingSuperReg(Reg: RI.getSubReg(Reg: DestReg, Idx: SystemZ::subreg_h64), |
902 | SubIdx: SystemZ::subreg_h64, RC: &SystemZ::VR128BitRegClass); |
903 | MCRegister DestRegLo = |
904 | RI.getMatchingSuperReg(Reg: RI.getSubReg(Reg: DestReg, Idx: SystemZ::subreg_l64), |
905 | SubIdx: SystemZ::subreg_h64, RC: &SystemZ::VR128BitRegClass); |
906 | |
907 | if (DestRegHi != SrcReg) |
908 | copyPhysReg(MBB, MBBI, DL, DestReg: DestRegHi, SrcReg, KillSrc: false); |
909 | BuildMI(BB&: MBB, I: MBBI, MIMD: DL, MCID: get(Opcode: SystemZ::VREPG), DestReg: DestRegLo) |
910 | .addReg(RegNo: SrcReg, flags: getKillRegState(B: KillSrc)).addImm(Val: 1); |
911 | return; |
912 | } |
913 | |
914 | if (SystemZ::FP128BitRegClass.contains(Reg: DestReg) && |
915 | SystemZ::GR128BitRegClass.contains(Reg: SrcReg)) { |
916 | MCRegister DestRegHi = RI.getSubReg(Reg: DestReg, Idx: SystemZ::subreg_h64); |
917 | MCRegister DestRegLo = RI.getSubReg(Reg: DestReg, Idx: SystemZ::subreg_l64); |
918 | MCRegister SrcRegHi = RI.getSubReg(Reg: SrcReg, Idx: SystemZ::subreg_h64); |
919 | MCRegister SrcRegLo = RI.getSubReg(Reg: SrcReg, Idx: SystemZ::subreg_l64); |
920 | |
921 | BuildMI(BB&: MBB, I: MBBI, MIMD: DL, MCID: get(Opcode: SystemZ::LDGR), DestReg: DestRegHi) |
922 | .addReg(RegNo: SrcRegHi) |
923 | .addReg(RegNo: DestReg, flags: RegState::ImplicitDefine); |
924 | |
925 | BuildMI(BB&: MBB, I: MBBI, MIMD: DL, MCID: get(Opcode: SystemZ::LDGR), DestReg: DestRegLo) |
926 | .addReg(RegNo: SrcRegLo, flags: getKillRegState(B: KillSrc)); |
927 | return; |
928 | } |
929 | |
930 | // Move CC value from a GR32. |
931 | if (DestReg == SystemZ::CC) { |
932 | unsigned Opcode = |
933 | SystemZ::GR32BitRegClass.contains(Reg: SrcReg) ? SystemZ::TMLH : SystemZ::TMHH; |
934 | BuildMI(BB&: MBB, I: MBBI, MIMD: DL, MCID: get(Opcode)) |
935 | .addReg(RegNo: SrcReg, flags: getKillRegState(B: KillSrc)) |
936 | .addImm(Val: 3 << (SystemZ::IPM_CC - 16)); |
937 | return; |
938 | } |
939 | |
940 | if (SystemZ::GR128BitRegClass.contains(Reg: DestReg) && |
941 | SystemZ::VR128BitRegClass.contains(Reg: SrcReg)) { |
942 | MCRegister DestH64 = RI.getSubReg(Reg: DestReg, Idx: SystemZ::subreg_h64); |
943 | MCRegister DestL64 = RI.getSubReg(Reg: DestReg, Idx: SystemZ::subreg_l64); |
944 | |
945 | BuildMI(BB&: MBB, I: MBBI, MIMD: DL, MCID: get(Opcode: SystemZ::VLGVG), DestReg: DestH64) |
946 | .addReg(RegNo: SrcReg) |
947 | .addReg(RegNo: SystemZ::NoRegister) |
948 | .addImm(Val: 0) |
949 | .addDef(RegNo: DestReg, Flags: RegState::Implicit); |
950 | BuildMI(BB&: MBB, I: MBBI, MIMD: DL, MCID: get(Opcode: SystemZ::VLGVG), DestReg: DestL64) |
951 | .addReg(RegNo: SrcReg, flags: getKillRegState(B: KillSrc)) |
952 | .addReg(RegNo: SystemZ::NoRegister) |
953 | .addImm(Val: 1); |
954 | return; |
955 | } |
956 | |
957 | if (SystemZ::VR128BitRegClass.contains(Reg: DestReg) && |
958 | SystemZ::GR128BitRegClass.contains(Reg: SrcReg)) { |
959 | BuildMI(BB&: MBB, I: MBBI, MIMD: DL, MCID: get(Opcode: SystemZ::VLVGP), DestReg) |
960 | .addReg(RegNo: RI.getSubReg(Reg: SrcReg, Idx: SystemZ::subreg_h64)) |
961 | .addReg(RegNo: RI.getSubReg(Reg: SrcReg, Idx: SystemZ::subreg_l64)); |
962 | return; |
963 | } |
964 | |
965 | // Everything else needs only one instruction. |
966 | unsigned Opcode; |
967 | if (SystemZ::GR64BitRegClass.contains(Reg1: DestReg, Reg2: SrcReg)) |
968 | Opcode = SystemZ::LGR; |
969 | else if (SystemZ::FP32BitRegClass.contains(Reg1: DestReg, Reg2: SrcReg)) |
970 | // For z13 we prefer LDR over LER to avoid partial register dependencies. |
971 | Opcode = STI.hasVector() ? SystemZ::LDR32 : SystemZ::LER; |
972 | else if (SystemZ::FP64BitRegClass.contains(Reg1: DestReg, Reg2: SrcReg)) |
973 | Opcode = SystemZ::LDR; |
974 | else if (SystemZ::FP128BitRegClass.contains(Reg1: DestReg, Reg2: SrcReg)) |
975 | Opcode = SystemZ::LXR; |
976 | else if (SystemZ::VR32BitRegClass.contains(Reg1: DestReg, Reg2: SrcReg)) |
977 | Opcode = SystemZ::VLR32; |
978 | else if (SystemZ::VR64BitRegClass.contains(Reg1: DestReg, Reg2: SrcReg)) |
979 | Opcode = SystemZ::VLR64; |
980 | else if (SystemZ::VR128BitRegClass.contains(Reg1: DestReg, Reg2: SrcReg)) |
981 | Opcode = SystemZ::VLR; |
982 | else if (SystemZ::AR32BitRegClass.contains(Reg1: DestReg, Reg2: SrcReg)) |
983 | Opcode = SystemZ::CPYA; |
984 | else |
985 | llvm_unreachable("Impossible reg-to-reg copy" ); |
986 | |
987 | BuildMI(BB&: MBB, I: MBBI, MIMD: DL, MCID: get(Opcode), DestReg) |
988 | .addReg(RegNo: SrcReg, flags: getKillRegState(B: KillSrc)); |
989 | } |
990 | |
991 | void SystemZInstrInfo::storeRegToStackSlot( |
992 | MachineBasicBlock &MBB, MachineBasicBlock::iterator MBBI, Register SrcReg, |
993 | bool isKill, int FrameIdx, const TargetRegisterClass *RC, |
994 | const TargetRegisterInfo *TRI, Register VReg) const { |
995 | DebugLoc DL = MBBI != MBB.end() ? MBBI->getDebugLoc() : DebugLoc(); |
996 | |
997 | // Callers may expect a single instruction, so keep 128-bit moves |
998 | // together for now and lower them after register allocation. |
999 | unsigned LoadOpcode, StoreOpcode; |
1000 | getLoadStoreOpcodes(RC, LoadOpcode, StoreOpcode); |
1001 | addFrameReference(MIB: BuildMI(BB&: MBB, I: MBBI, MIMD: DL, MCID: get(Opcode: StoreOpcode)) |
1002 | .addReg(RegNo: SrcReg, flags: getKillRegState(B: isKill)), |
1003 | FI: FrameIdx); |
1004 | } |
1005 | |
1006 | void SystemZInstrInfo::loadRegFromStackSlot(MachineBasicBlock &MBB, |
1007 | MachineBasicBlock::iterator MBBI, |
1008 | Register DestReg, int FrameIdx, |
1009 | const TargetRegisterClass *RC, |
1010 | const TargetRegisterInfo *TRI, |
1011 | Register VReg) const { |
1012 | DebugLoc DL = MBBI != MBB.end() ? MBBI->getDebugLoc() : DebugLoc(); |
1013 | |
1014 | // Callers may expect a single instruction, so keep 128-bit moves |
1015 | // together for now and lower them after register allocation. |
1016 | unsigned LoadOpcode, StoreOpcode; |
1017 | getLoadStoreOpcodes(RC, LoadOpcode, StoreOpcode); |
1018 | addFrameReference(MIB: BuildMI(BB&: MBB, I: MBBI, MIMD: DL, MCID: get(Opcode: LoadOpcode), DestReg), |
1019 | FI: FrameIdx); |
1020 | } |
1021 | |
1022 | // Return true if MI is a simple load or store with a 12-bit displacement |
1023 | // and no index. Flag is SimpleBDXLoad for loads and SimpleBDXStore for stores. |
1024 | static bool isSimpleBD12Move(const MachineInstr *MI, unsigned Flag) { |
1025 | const MCInstrDesc &MCID = MI->getDesc(); |
1026 | return ((MCID.TSFlags & Flag) && |
1027 | isUInt<12>(x: MI->getOperand(i: 2).getImm()) && |
1028 | MI->getOperand(i: 3).getReg() == 0); |
1029 | } |
1030 | |
1031 | namespace { |
1032 | |
1033 | struct LogicOp { |
1034 | LogicOp() = default; |
1035 | LogicOp(unsigned regSize, unsigned immLSB, unsigned immSize) |
1036 | : RegSize(regSize), ImmLSB(immLSB), ImmSize(immSize) {} |
1037 | |
1038 | explicit operator bool() const { return RegSize; } |
1039 | |
1040 | unsigned RegSize = 0; |
1041 | unsigned ImmLSB = 0; |
1042 | unsigned ImmSize = 0; |
1043 | }; |
1044 | |
1045 | } // end anonymous namespace |
1046 | |
1047 | static LogicOp interpretAndImmediate(unsigned Opcode) { |
1048 | switch (Opcode) { |
1049 | case SystemZ::NILMux: return LogicOp(32, 0, 16); |
1050 | case SystemZ::NIHMux: return LogicOp(32, 16, 16); |
1051 | case SystemZ::NILL64: return LogicOp(64, 0, 16); |
1052 | case SystemZ::NILH64: return LogicOp(64, 16, 16); |
1053 | case SystemZ::NIHL64: return LogicOp(64, 32, 16); |
1054 | case SystemZ::NIHH64: return LogicOp(64, 48, 16); |
1055 | case SystemZ::NIFMux: return LogicOp(32, 0, 32); |
1056 | case SystemZ::NILF64: return LogicOp(64, 0, 32); |
1057 | case SystemZ::NIHF64: return LogicOp(64, 32, 32); |
1058 | default: return LogicOp(); |
1059 | } |
1060 | } |
1061 | |
1062 | static void transferDeadCC(MachineInstr *OldMI, MachineInstr *NewMI) { |
1063 | if (OldMI->registerDefIsDead(Reg: SystemZ::CC, /*TRI=*/nullptr)) { |
1064 | MachineOperand *CCDef = |
1065 | NewMI->findRegisterDefOperand(Reg: SystemZ::CC, /*TRI=*/nullptr); |
1066 | if (CCDef != nullptr) |
1067 | CCDef->setIsDead(true); |
1068 | } |
1069 | } |
1070 | |
1071 | static void transferMIFlag(MachineInstr *OldMI, MachineInstr *NewMI, |
1072 | MachineInstr::MIFlag Flag) { |
1073 | if (OldMI->getFlag(Flag)) |
1074 | NewMI->setFlag(Flag); |
1075 | } |
1076 | |
1077 | MachineInstr * |
1078 | SystemZInstrInfo::convertToThreeAddress(MachineInstr &MI, LiveVariables *LV, |
1079 | LiveIntervals *LIS) const { |
1080 | MachineBasicBlock *MBB = MI.getParent(); |
1081 | |
1082 | // Try to convert an AND into an RISBG-type instruction. |
1083 | // TODO: It might be beneficial to select RISBG and shorten to AND instead. |
1084 | if (LogicOp And = interpretAndImmediate(Opcode: MI.getOpcode())) { |
1085 | uint64_t Imm = MI.getOperand(i: 2).getImm() << And.ImmLSB; |
1086 | // AND IMMEDIATE leaves the other bits of the register unchanged. |
1087 | Imm |= allOnes(Count: And.RegSize) & ~(allOnes(Count: And.ImmSize) << And.ImmLSB); |
1088 | unsigned Start, End; |
1089 | if (isRxSBGMask(Mask: Imm, BitSize: And.RegSize, Start, End)) { |
1090 | unsigned NewOpcode; |
1091 | if (And.RegSize == 64) { |
1092 | NewOpcode = SystemZ::RISBG; |
1093 | // Prefer RISBGN if available, since it does not clobber CC. |
1094 | if (STI.hasMiscellaneousExtensions()) |
1095 | NewOpcode = SystemZ::RISBGN; |
1096 | } else { |
1097 | NewOpcode = SystemZ::RISBMux; |
1098 | Start &= 31; |
1099 | End &= 31; |
1100 | } |
1101 | MachineOperand &Dest = MI.getOperand(i: 0); |
1102 | MachineOperand &Src = MI.getOperand(i: 1); |
1103 | MachineInstrBuilder MIB = |
1104 | BuildMI(BB&: *MBB, I&: MI, MIMD: MI.getDebugLoc(), MCID: get(Opcode: NewOpcode)) |
1105 | .add(MO: Dest) |
1106 | .addReg(RegNo: 0) |
1107 | .addReg(RegNo: Src.getReg(), flags: getKillRegState(B: Src.isKill()), |
1108 | SubReg: Src.getSubReg()) |
1109 | .addImm(Val: Start) |
1110 | .addImm(Val: End + 128) |
1111 | .addImm(Val: 0); |
1112 | if (LV) { |
1113 | unsigned NumOps = MI.getNumOperands(); |
1114 | for (unsigned I = 1; I < NumOps; ++I) { |
1115 | MachineOperand &Op = MI.getOperand(i: I); |
1116 | if (Op.isReg() && Op.isKill()) |
1117 | LV->replaceKillInstruction(Reg: Op.getReg(), OldMI&: MI, NewMI&: *MIB); |
1118 | } |
1119 | } |
1120 | if (LIS) |
1121 | LIS->ReplaceMachineInstrInMaps(MI, NewMI&: *MIB); |
1122 | transferDeadCC(OldMI: &MI, NewMI: MIB); |
1123 | return MIB; |
1124 | } |
1125 | } |
1126 | return nullptr; |
1127 | } |
1128 | |
1129 | bool SystemZInstrInfo::isAssociativeAndCommutative(const MachineInstr &Inst, |
1130 | bool Invert) const { |
1131 | unsigned Opc = Inst.getOpcode(); |
1132 | if (Invert) { |
1133 | auto InverseOpcode = getInverseOpcode(Opcode: Opc); |
1134 | if (!InverseOpcode) |
1135 | return false; |
1136 | Opc = *InverseOpcode; |
1137 | } |
1138 | |
1139 | switch (Opc) { |
1140 | default: |
1141 | break; |
1142 | // Adds and multiplications. |
1143 | case SystemZ::WFADB: |
1144 | case SystemZ::WFASB: |
1145 | case SystemZ::WFAXB: |
1146 | case SystemZ::VFADB: |
1147 | case SystemZ::VFASB: |
1148 | case SystemZ::WFMDB: |
1149 | case SystemZ::WFMSB: |
1150 | case SystemZ::WFMXB: |
1151 | case SystemZ::VFMDB: |
1152 | case SystemZ::VFMSB: |
1153 | return (Inst.getFlag(Flag: MachineInstr::MIFlag::FmReassoc) && |
1154 | Inst.getFlag(Flag: MachineInstr::MIFlag::FmNsz)); |
1155 | } |
1156 | |
1157 | return false; |
1158 | } |
1159 | |
1160 | std::optional<unsigned> |
1161 | SystemZInstrInfo::getInverseOpcode(unsigned Opcode) const { |
1162 | // fadd => fsub |
1163 | switch (Opcode) { |
1164 | case SystemZ::WFADB: |
1165 | return SystemZ::WFSDB; |
1166 | case SystemZ::WFASB: |
1167 | return SystemZ::WFSSB; |
1168 | case SystemZ::WFAXB: |
1169 | return SystemZ::WFSXB; |
1170 | case SystemZ::VFADB: |
1171 | return SystemZ::VFSDB; |
1172 | case SystemZ::VFASB: |
1173 | return SystemZ::VFSSB; |
1174 | // fsub => fadd |
1175 | case SystemZ::WFSDB: |
1176 | return SystemZ::WFADB; |
1177 | case SystemZ::WFSSB: |
1178 | return SystemZ::WFASB; |
1179 | case SystemZ::WFSXB: |
1180 | return SystemZ::WFAXB; |
1181 | case SystemZ::VFSDB: |
1182 | return SystemZ::VFADB; |
1183 | case SystemZ::VFSSB: |
1184 | return SystemZ::VFASB; |
1185 | default: |
1186 | return std::nullopt; |
1187 | } |
1188 | } |
1189 | |
1190 | MachineInstr *SystemZInstrInfo::foldMemoryOperandImpl( |
1191 | MachineFunction &MF, MachineInstr &MI, ArrayRef<unsigned> Ops, |
1192 | MachineBasicBlock::iterator InsertPt, int FrameIndex, |
1193 | LiveIntervals *LIS, VirtRegMap *VRM) const { |
1194 | const TargetRegisterInfo *TRI = MF.getSubtarget().getRegisterInfo(); |
1195 | MachineRegisterInfo &MRI = MF.getRegInfo(); |
1196 | const MachineFrameInfo &MFI = MF.getFrameInfo(); |
1197 | unsigned Size = MFI.getObjectSize(ObjectIdx: FrameIndex); |
1198 | unsigned Opcode = MI.getOpcode(); |
1199 | |
1200 | // Check CC liveness if new instruction introduces a dead def of CC. |
1201 | SlotIndex MISlot = SlotIndex(); |
1202 | LiveRange *CCLiveRange = nullptr; |
1203 | bool CCLiveAtMI = true; |
1204 | if (LIS) { |
1205 | MISlot = LIS->getSlotIndexes()->getInstructionIndex(MI).getRegSlot(); |
1206 | auto CCUnits = TRI->regunits(Reg: MCRegister::from(Val: SystemZ::CC)); |
1207 | assert(range_size(CCUnits) == 1 && "CC only has one reg unit." ); |
1208 | CCLiveRange = &LIS->getRegUnit(Unit: *CCUnits.begin()); |
1209 | CCLiveAtMI = CCLiveRange->liveAt(index: MISlot); |
1210 | } |
1211 | |
1212 | if (Ops.size() == 2 && Ops[0] == 0 && Ops[1] == 1) { |
1213 | if (!CCLiveAtMI && (Opcode == SystemZ::LA || Opcode == SystemZ::LAY) && |
1214 | isInt<8>(x: MI.getOperand(i: 2).getImm()) && !MI.getOperand(i: 3).getReg()) { |
1215 | // LA(Y) %reg, CONST(%reg) -> AGSI %mem, CONST |
1216 | MachineInstr *BuiltMI = BuildMI(BB&: *InsertPt->getParent(), I: InsertPt, |
1217 | MIMD: MI.getDebugLoc(), MCID: get(Opcode: SystemZ::AGSI)) |
1218 | .addFrameIndex(Idx: FrameIndex) |
1219 | .addImm(Val: 0) |
1220 | .addImm(Val: MI.getOperand(i: 2).getImm()); |
1221 | BuiltMI->findRegisterDefOperand(Reg: SystemZ::CC, /*TRI=*/nullptr) |
1222 | ->setIsDead(true); |
1223 | CCLiveRange->createDeadDef(Def: MISlot, VNIAlloc&: LIS->getVNInfoAllocator()); |
1224 | return BuiltMI; |
1225 | } |
1226 | return nullptr; |
1227 | } |
1228 | |
1229 | // All other cases require a single operand. |
1230 | if (Ops.size() != 1) |
1231 | return nullptr; |
1232 | |
1233 | unsigned OpNum = Ops[0]; |
1234 | assert(Size * 8 == |
1235 | TRI->getRegSizeInBits(*MF.getRegInfo() |
1236 | .getRegClass(MI.getOperand(OpNum).getReg())) && |
1237 | "Invalid size combination" ); |
1238 | |
1239 | if ((Opcode == SystemZ::AHI || Opcode == SystemZ::AGHI) && OpNum == 0 && |
1240 | isInt<8>(x: MI.getOperand(i: 2).getImm())) { |
1241 | // A(G)HI %reg, CONST -> A(G)SI %mem, CONST |
1242 | Opcode = (Opcode == SystemZ::AHI ? SystemZ::ASI : SystemZ::AGSI); |
1243 | MachineInstr *BuiltMI = |
1244 | BuildMI(BB&: *InsertPt->getParent(), I: InsertPt, MIMD: MI.getDebugLoc(), MCID: get(Opcode)) |
1245 | .addFrameIndex(Idx: FrameIndex) |
1246 | .addImm(Val: 0) |
1247 | .addImm(Val: MI.getOperand(i: 2).getImm()); |
1248 | transferDeadCC(OldMI: &MI, NewMI: BuiltMI); |
1249 | transferMIFlag(OldMI: &MI, NewMI: BuiltMI, Flag: MachineInstr::NoSWrap); |
1250 | return BuiltMI; |
1251 | } |
1252 | |
1253 | if ((Opcode == SystemZ::ALFI && OpNum == 0 && |
1254 | isInt<8>(x: (int32_t)MI.getOperand(i: 2).getImm())) || |
1255 | (Opcode == SystemZ::ALGFI && OpNum == 0 && |
1256 | isInt<8>(x: (int64_t)MI.getOperand(i: 2).getImm()))) { |
1257 | // AL(G)FI %reg, CONST -> AL(G)SI %mem, CONST |
1258 | Opcode = (Opcode == SystemZ::ALFI ? SystemZ::ALSI : SystemZ::ALGSI); |
1259 | MachineInstr *BuiltMI = |
1260 | BuildMI(BB&: *InsertPt->getParent(), I: InsertPt, MIMD: MI.getDebugLoc(), MCID: get(Opcode)) |
1261 | .addFrameIndex(Idx: FrameIndex) |
1262 | .addImm(Val: 0) |
1263 | .addImm(Val: (int8_t)MI.getOperand(i: 2).getImm()); |
1264 | transferDeadCC(OldMI: &MI, NewMI: BuiltMI); |
1265 | return BuiltMI; |
1266 | } |
1267 | |
1268 | if ((Opcode == SystemZ::SLFI && OpNum == 0 && |
1269 | isInt<8>(x: (int32_t)-MI.getOperand(i: 2).getImm())) || |
1270 | (Opcode == SystemZ::SLGFI && OpNum == 0 && |
1271 | isInt<8>(x: (int64_t)-MI.getOperand(i: 2).getImm()))) { |
1272 | // SL(G)FI %reg, CONST -> AL(G)SI %mem, -CONST |
1273 | Opcode = (Opcode == SystemZ::SLFI ? SystemZ::ALSI : SystemZ::ALGSI); |
1274 | MachineInstr *BuiltMI = |
1275 | BuildMI(BB&: *InsertPt->getParent(), I: InsertPt, MIMD: MI.getDebugLoc(), MCID: get(Opcode)) |
1276 | .addFrameIndex(Idx: FrameIndex) |
1277 | .addImm(Val: 0) |
1278 | .addImm(Val: (int8_t)-MI.getOperand(i: 2).getImm()); |
1279 | transferDeadCC(OldMI: &MI, NewMI: BuiltMI); |
1280 | return BuiltMI; |
1281 | } |
1282 | |
1283 | unsigned MemImmOpc = 0; |
1284 | switch (Opcode) { |
1285 | case SystemZ::LHIMux: |
1286 | case SystemZ::LHI: MemImmOpc = SystemZ::MVHI; break; |
1287 | case SystemZ::LGHI: MemImmOpc = SystemZ::MVGHI; break; |
1288 | case SystemZ::CHIMux: |
1289 | case SystemZ::CHI: MemImmOpc = SystemZ::CHSI; break; |
1290 | case SystemZ::CGHI: MemImmOpc = SystemZ::CGHSI; break; |
1291 | case SystemZ::CLFIMux: |
1292 | case SystemZ::CLFI: |
1293 | if (isUInt<16>(x: MI.getOperand(i: 1).getImm())) |
1294 | MemImmOpc = SystemZ::CLFHSI; |
1295 | break; |
1296 | case SystemZ::CLGFI: |
1297 | if (isUInt<16>(x: MI.getOperand(i: 1).getImm())) |
1298 | MemImmOpc = SystemZ::CLGHSI; |
1299 | break; |
1300 | default: break; |
1301 | } |
1302 | if (MemImmOpc) |
1303 | return BuildMI(BB&: *InsertPt->getParent(), I: InsertPt, MIMD: MI.getDebugLoc(), |
1304 | MCID: get(Opcode: MemImmOpc)) |
1305 | .addFrameIndex(Idx: FrameIndex) |
1306 | .addImm(Val: 0) |
1307 | .addImm(Val: MI.getOperand(i: 1).getImm()); |
1308 | |
1309 | if (Opcode == SystemZ::LGDR || Opcode == SystemZ::LDGR) { |
1310 | bool Op0IsGPR = (Opcode == SystemZ::LGDR); |
1311 | bool Op1IsGPR = (Opcode == SystemZ::LDGR); |
1312 | // If we're spilling the destination of an LDGR or LGDR, store the |
1313 | // source register instead. |
1314 | if (OpNum == 0) { |
1315 | unsigned StoreOpcode = Op1IsGPR ? SystemZ::STG : SystemZ::STD; |
1316 | return BuildMI(BB&: *InsertPt->getParent(), I: InsertPt, MIMD: MI.getDebugLoc(), |
1317 | MCID: get(Opcode: StoreOpcode)) |
1318 | .add(MO: MI.getOperand(i: 1)) |
1319 | .addFrameIndex(Idx: FrameIndex) |
1320 | .addImm(Val: 0) |
1321 | .addReg(RegNo: 0); |
1322 | } |
1323 | // If we're spilling the source of an LDGR or LGDR, load the |
1324 | // destination register instead. |
1325 | if (OpNum == 1) { |
1326 | unsigned LoadOpcode = Op0IsGPR ? SystemZ::LG : SystemZ::LD; |
1327 | return BuildMI(BB&: *InsertPt->getParent(), I: InsertPt, MIMD: MI.getDebugLoc(), |
1328 | MCID: get(Opcode: LoadOpcode)) |
1329 | .add(MO: MI.getOperand(i: 0)) |
1330 | .addFrameIndex(Idx: FrameIndex) |
1331 | .addImm(Val: 0) |
1332 | .addReg(RegNo: 0); |
1333 | } |
1334 | } |
1335 | |
1336 | // Look for cases where the source of a simple store or the destination |
1337 | // of a simple load is being spilled. Try to use MVC instead. |
1338 | // |
1339 | // Although MVC is in practice a fast choice in these cases, it is still |
1340 | // logically a bytewise copy. This means that we cannot use it if the |
1341 | // load or store is volatile. We also wouldn't be able to use MVC if |
1342 | // the two memories partially overlap, but that case cannot occur here, |
1343 | // because we know that one of the memories is a full frame index. |
1344 | // |
1345 | // For performance reasons, we also want to avoid using MVC if the addresses |
1346 | // might be equal. We don't worry about that case here, because spill slot |
1347 | // coloring happens later, and because we have special code to remove |
1348 | // MVCs that turn out to be redundant. |
1349 | if (OpNum == 0 && MI.hasOneMemOperand()) { |
1350 | MachineMemOperand *MMO = *MI.memoperands_begin(); |
1351 | if (MMO->getSize() == Size && !MMO->isVolatile() && !MMO->isAtomic()) { |
1352 | // Handle conversion of loads. |
1353 | if (isSimpleBD12Move(MI: &MI, Flag: SystemZII::SimpleBDXLoad)) { |
1354 | return BuildMI(BB&: *InsertPt->getParent(), I: InsertPt, MIMD: MI.getDebugLoc(), |
1355 | MCID: get(Opcode: SystemZ::MVC)) |
1356 | .addFrameIndex(Idx: FrameIndex) |
1357 | .addImm(Val: 0) |
1358 | .addImm(Val: Size) |
1359 | .add(MO: MI.getOperand(i: 1)) |
1360 | .addImm(Val: MI.getOperand(i: 2).getImm()) |
1361 | .addMemOperand(MMO); |
1362 | } |
1363 | // Handle conversion of stores. |
1364 | if (isSimpleBD12Move(MI: &MI, Flag: SystemZII::SimpleBDXStore)) { |
1365 | return BuildMI(BB&: *InsertPt->getParent(), I: InsertPt, MIMD: MI.getDebugLoc(), |
1366 | MCID: get(Opcode: SystemZ::MVC)) |
1367 | .add(MO: MI.getOperand(i: 1)) |
1368 | .addImm(Val: MI.getOperand(i: 2).getImm()) |
1369 | .addImm(Val: Size) |
1370 | .addFrameIndex(Idx: FrameIndex) |
1371 | .addImm(Val: 0) |
1372 | .addMemOperand(MMO); |
1373 | } |
1374 | } |
1375 | } |
1376 | |
1377 | // If the spilled operand is the final one or the instruction is |
1378 | // commutable, try to change <INSN>R into <INSN>. Don't introduce a def of |
1379 | // CC if it is live and MI does not define it. |
1380 | unsigned NumOps = MI.getNumExplicitOperands(); |
1381 | int MemOpcode = SystemZ::getMemOpcode(Opcode); |
1382 | if (MemOpcode == -1 || |
1383 | (CCLiveAtMI && !MI.definesRegister(Reg: SystemZ::CC, /*TRI=*/nullptr) && |
1384 | get(Opcode: MemOpcode).hasImplicitDefOfPhysReg(Reg: SystemZ::CC))) |
1385 | return nullptr; |
1386 | |
1387 | // Check if all other vregs have a usable allocation in the case of vector |
1388 | // to FP conversion. |
1389 | const MCInstrDesc &MCID = MI.getDesc(); |
1390 | for (unsigned I = 0, E = MCID.getNumOperands(); I != E; ++I) { |
1391 | const MCOperandInfo &MCOI = MCID.operands()[I]; |
1392 | if (MCOI.OperandType != MCOI::OPERAND_REGISTER || I == OpNum) |
1393 | continue; |
1394 | const TargetRegisterClass *RC = TRI->getRegClass(i: MCOI.RegClass); |
1395 | if (RC == &SystemZ::VR32BitRegClass || RC == &SystemZ::VR64BitRegClass) { |
1396 | Register Reg = MI.getOperand(i: I).getReg(); |
1397 | Register PhysReg = Reg.isVirtual() |
1398 | ? (VRM ? Register(VRM->getPhys(virtReg: Reg)) : Register()) |
1399 | : Reg; |
1400 | if (!PhysReg || |
1401 | !(SystemZ::FP32BitRegClass.contains(Reg: PhysReg) || |
1402 | SystemZ::FP64BitRegClass.contains(Reg: PhysReg) || |
1403 | SystemZ::VF128BitRegClass.contains(Reg: PhysReg))) |
1404 | return nullptr; |
1405 | } |
1406 | } |
1407 | // Fused multiply and add/sub need to have the same dst and accumulator reg. |
1408 | bool FusedFPOp = (Opcode == SystemZ::WFMADB || Opcode == SystemZ::WFMASB || |
1409 | Opcode == SystemZ::WFMSDB || Opcode == SystemZ::WFMSSB); |
1410 | if (FusedFPOp) { |
1411 | Register DstReg = VRM->getPhys(virtReg: MI.getOperand(i: 0).getReg()); |
1412 | Register AccReg = VRM->getPhys(virtReg: MI.getOperand(i: 3).getReg()); |
1413 | if (OpNum == 0 || OpNum == 3 || DstReg != AccReg) |
1414 | return nullptr; |
1415 | } |
1416 | |
1417 | // Try to swap compare operands if possible. |
1418 | bool NeedsCommute = false; |
1419 | if ((MI.getOpcode() == SystemZ::CR || MI.getOpcode() == SystemZ::CGR || |
1420 | MI.getOpcode() == SystemZ::CLR || MI.getOpcode() == SystemZ::CLGR || |
1421 | MI.getOpcode() == SystemZ::WFCDB || MI.getOpcode() == SystemZ::WFCSB || |
1422 | MI.getOpcode() == SystemZ::WFKDB || MI.getOpcode() == SystemZ::WFKSB) && |
1423 | OpNum == 0 && prepareCompareSwapOperands(MBBI: MI)) |
1424 | NeedsCommute = true; |
1425 | |
1426 | bool CCOperands = false; |
1427 | if (MI.getOpcode() == SystemZ::LOCRMux || MI.getOpcode() == SystemZ::LOCGR || |
1428 | MI.getOpcode() == SystemZ::SELRMux || MI.getOpcode() == SystemZ::SELGR) { |
1429 | assert(MI.getNumOperands() == 6 && NumOps == 5 && |
1430 | "LOCR/SELR instruction operands corrupt?" ); |
1431 | NumOps -= 2; |
1432 | CCOperands = true; |
1433 | } |
1434 | |
1435 | // See if this is a 3-address instruction that is convertible to 2-address |
1436 | // and suitable for folding below. Only try this with virtual registers |
1437 | // and a provided VRM (during regalloc). |
1438 | if (NumOps == 3 && SystemZ::getTargetMemOpcode(Opcode: MemOpcode) != -1) { |
1439 | if (VRM == nullptr) |
1440 | return nullptr; |
1441 | else { |
1442 | Register DstReg = MI.getOperand(i: 0).getReg(); |
1443 | Register DstPhys = |
1444 | (DstReg.isVirtual() ? Register(VRM->getPhys(virtReg: DstReg)) : DstReg); |
1445 | Register SrcReg = (OpNum == 2 ? MI.getOperand(i: 1).getReg() |
1446 | : ((OpNum == 1 && MI.isCommutable()) |
1447 | ? MI.getOperand(i: 2).getReg() |
1448 | : Register())); |
1449 | if (DstPhys && !SystemZ::GRH32BitRegClass.contains(Reg: DstPhys) && SrcReg && |
1450 | SrcReg.isVirtual() && DstPhys == VRM->getPhys(virtReg: SrcReg)) |
1451 | NeedsCommute = (OpNum == 1); |
1452 | else |
1453 | return nullptr; |
1454 | } |
1455 | } |
1456 | |
1457 | if ((OpNum == NumOps - 1) || NeedsCommute || FusedFPOp) { |
1458 | const MCInstrDesc &MemDesc = get(Opcode: MemOpcode); |
1459 | uint64_t AccessBytes = SystemZII::getAccessSize(Flags: MemDesc.TSFlags); |
1460 | assert(AccessBytes != 0 && "Size of access should be known" ); |
1461 | assert(AccessBytes <= Size && "Access outside the frame index" ); |
1462 | uint64_t Offset = Size - AccessBytes; |
1463 | MachineInstrBuilder MIB = BuildMI(BB&: *InsertPt->getParent(), I: InsertPt, |
1464 | MIMD: MI.getDebugLoc(), MCID: get(Opcode: MemOpcode)); |
1465 | if (MI.isCompare()) { |
1466 | assert(NumOps == 2 && "Expected 2 register operands for a compare." ); |
1467 | MIB.add(MO: MI.getOperand(i: NeedsCommute ? 1 : 0)); |
1468 | } |
1469 | else if (FusedFPOp) { |
1470 | MIB.add(MO: MI.getOperand(i: 0)); |
1471 | MIB.add(MO: MI.getOperand(i: 3)); |
1472 | MIB.add(MO: MI.getOperand(i: OpNum == 1 ? 2 : 1)); |
1473 | } |
1474 | else { |
1475 | MIB.add(MO: MI.getOperand(i: 0)); |
1476 | if (NeedsCommute) |
1477 | MIB.add(MO: MI.getOperand(i: 2)); |
1478 | else |
1479 | for (unsigned I = 1; I < OpNum; ++I) |
1480 | MIB.add(MO: MI.getOperand(i: I)); |
1481 | } |
1482 | MIB.addFrameIndex(Idx: FrameIndex).addImm(Val: Offset); |
1483 | if (MemDesc.TSFlags & SystemZII::HasIndex) |
1484 | MIB.addReg(RegNo: 0); |
1485 | if (CCOperands) { |
1486 | unsigned CCValid = MI.getOperand(i: NumOps).getImm(); |
1487 | unsigned CCMask = MI.getOperand(i: NumOps + 1).getImm(); |
1488 | MIB.addImm(Val: CCValid); |
1489 | MIB.addImm(Val: NeedsCommute ? CCMask ^ CCValid : CCMask); |
1490 | } |
1491 | if (MIB->definesRegister(Reg: SystemZ::CC, /*TRI=*/nullptr) && |
1492 | (!MI.definesRegister(Reg: SystemZ::CC, /*TRI=*/nullptr) || |
1493 | MI.registerDefIsDead(Reg: SystemZ::CC, /*TRI=*/nullptr))) { |
1494 | MIB->addRegisterDead(Reg: SystemZ::CC, RegInfo: TRI); |
1495 | if (CCLiveRange) |
1496 | CCLiveRange->createDeadDef(Def: MISlot, VNIAlloc&: LIS->getVNInfoAllocator()); |
1497 | } |
1498 | // Constrain the register classes if converted from a vector opcode. The |
1499 | // allocated regs are in an FP reg-class per previous check above. |
1500 | for (const MachineOperand &MO : MIB->operands()) |
1501 | if (MO.isReg() && MO.getReg().isVirtual()) { |
1502 | Register Reg = MO.getReg(); |
1503 | if (MRI.getRegClass(Reg) == &SystemZ::VR32BitRegClass) |
1504 | MRI.setRegClass(Reg, RC: &SystemZ::FP32BitRegClass); |
1505 | else if (MRI.getRegClass(Reg) == &SystemZ::VR64BitRegClass) |
1506 | MRI.setRegClass(Reg, RC: &SystemZ::FP64BitRegClass); |
1507 | else if (MRI.getRegClass(Reg) == &SystemZ::VR128BitRegClass) |
1508 | MRI.setRegClass(Reg, RC: &SystemZ::VF128BitRegClass); |
1509 | } |
1510 | |
1511 | transferDeadCC(OldMI: &MI, NewMI: MIB); |
1512 | transferMIFlag(OldMI: &MI, NewMI: MIB, Flag: MachineInstr::NoSWrap); |
1513 | transferMIFlag(OldMI: &MI, NewMI: MIB, Flag: MachineInstr::NoFPExcept); |
1514 | return MIB; |
1515 | } |
1516 | |
1517 | return nullptr; |
1518 | } |
1519 | |
1520 | MachineInstr *SystemZInstrInfo::foldMemoryOperandImpl( |
1521 | MachineFunction &MF, MachineInstr &MI, ArrayRef<unsigned> Ops, |
1522 | MachineBasicBlock::iterator InsertPt, MachineInstr &LoadMI, |
1523 | LiveIntervals *LIS) const { |
1524 | MachineRegisterInfo *MRI = &MF.getRegInfo(); |
1525 | MachineBasicBlock *MBB = MI.getParent(); |
1526 | |
1527 | // For reassociable FP operations, any loads have been purposefully left |
1528 | // unfolded so that MachineCombiner can do its work on reg/reg |
1529 | // opcodes. After that, as many loads as possible are now folded. |
1530 | // TODO: This may be beneficial with other opcodes as well as machine-sink |
1531 | // can move loads close to their user in a different MBB, which the isel |
1532 | // matcher did not see. |
1533 | unsigned LoadOpc = 0; |
1534 | unsigned RegMemOpcode = 0; |
1535 | const TargetRegisterClass *FPRC = nullptr; |
1536 | RegMemOpcode = MI.getOpcode() == SystemZ::WFADB ? SystemZ::ADB |
1537 | : MI.getOpcode() == SystemZ::WFSDB ? SystemZ::SDB |
1538 | : MI.getOpcode() == SystemZ::WFMDB ? SystemZ::MDB |
1539 | : 0; |
1540 | if (RegMemOpcode) { |
1541 | LoadOpc = SystemZ::VL64; |
1542 | FPRC = &SystemZ::FP64BitRegClass; |
1543 | } else { |
1544 | RegMemOpcode = MI.getOpcode() == SystemZ::WFASB ? SystemZ::AEB |
1545 | : MI.getOpcode() == SystemZ::WFSSB ? SystemZ::SEB |
1546 | : MI.getOpcode() == SystemZ::WFMSB ? SystemZ::MEEB |
1547 | : 0; |
1548 | if (RegMemOpcode) { |
1549 | LoadOpc = SystemZ::VL32; |
1550 | FPRC = &SystemZ::FP32BitRegClass; |
1551 | } |
1552 | } |
1553 | if (!RegMemOpcode || LoadMI.getOpcode() != LoadOpc) |
1554 | return nullptr; |
1555 | |
1556 | // If RegMemOpcode clobbers CC, first make sure CC is not live at this point. |
1557 | if (get(Opcode: RegMemOpcode).hasImplicitDefOfPhysReg(Reg: SystemZ::CC)) { |
1558 | assert(LoadMI.getParent() == MI.getParent() && "Assuming a local fold." ); |
1559 | assert(LoadMI != InsertPt && "Assuming InsertPt not to be first in MBB." ); |
1560 | for (MachineBasicBlock::iterator MII = std::prev(x: InsertPt);; |
1561 | --MII) { |
1562 | if (MII->definesRegister(Reg: SystemZ::CC, /*TRI=*/nullptr)) { |
1563 | if (!MII->registerDefIsDead(Reg: SystemZ::CC, /*TRI=*/nullptr)) |
1564 | return nullptr; |
1565 | break; |
1566 | } |
1567 | if (MII == MBB->begin()) { |
1568 | if (MBB->isLiveIn(Reg: SystemZ::CC)) |
1569 | return nullptr; |
1570 | break; |
1571 | } |
1572 | } |
1573 | } |
1574 | |
1575 | Register FoldAsLoadDefReg = LoadMI.getOperand(i: 0).getReg(); |
1576 | if (Ops.size() != 1 || FoldAsLoadDefReg != MI.getOperand(i: Ops[0]).getReg()) |
1577 | return nullptr; |
1578 | Register DstReg = MI.getOperand(i: 0).getReg(); |
1579 | MachineOperand LHS = MI.getOperand(i: 1); |
1580 | MachineOperand RHS = MI.getOperand(i: 2); |
1581 | MachineOperand &RegMO = RHS.getReg() == FoldAsLoadDefReg ? LHS : RHS; |
1582 | if ((RegMemOpcode == SystemZ::SDB || RegMemOpcode == SystemZ::SEB) && |
1583 | FoldAsLoadDefReg != RHS.getReg()) |
1584 | return nullptr; |
1585 | |
1586 | MachineOperand &Base = LoadMI.getOperand(i: 1); |
1587 | MachineOperand &Disp = LoadMI.getOperand(i: 2); |
1588 | MachineOperand &Indx = LoadMI.getOperand(i: 3); |
1589 | MachineInstrBuilder MIB = |
1590 | BuildMI(BB&: *MI.getParent(), I: InsertPt, MIMD: MI.getDebugLoc(), MCID: get(Opcode: RegMemOpcode), DestReg: DstReg) |
1591 | .add(MO: RegMO) |
1592 | .add(MO: Base) |
1593 | .add(MO: Disp) |
1594 | .add(MO: Indx); |
1595 | MIB->addRegisterDead(Reg: SystemZ::CC, RegInfo: &RI); |
1596 | MRI->setRegClass(Reg: DstReg, RC: FPRC); |
1597 | MRI->setRegClass(Reg: RegMO.getReg(), RC: FPRC); |
1598 | transferMIFlag(OldMI: &MI, NewMI: MIB, Flag: MachineInstr::NoFPExcept); |
1599 | |
1600 | return MIB; |
1601 | } |
1602 | |
1603 | bool SystemZInstrInfo::expandPostRAPseudo(MachineInstr &MI) const { |
1604 | switch (MI.getOpcode()) { |
1605 | case SystemZ::L128: |
1606 | splitMove(MI, NewOpcode: SystemZ::LG); |
1607 | return true; |
1608 | |
1609 | case SystemZ::ST128: |
1610 | splitMove(MI, NewOpcode: SystemZ::STG); |
1611 | return true; |
1612 | |
1613 | case SystemZ::LX: |
1614 | splitMove(MI, NewOpcode: SystemZ::LD); |
1615 | return true; |
1616 | |
1617 | case SystemZ::STX: |
1618 | splitMove(MI, NewOpcode: SystemZ::STD); |
1619 | return true; |
1620 | |
1621 | case SystemZ::LBMux: |
1622 | expandRXYPseudo(MI, LowOpcode: SystemZ::LB, HighOpcode: SystemZ::LBH); |
1623 | return true; |
1624 | |
1625 | case SystemZ::LHMux: |
1626 | expandRXYPseudo(MI, LowOpcode: SystemZ::LH, HighOpcode: SystemZ::LHH); |
1627 | return true; |
1628 | |
1629 | case SystemZ::LLCRMux: |
1630 | expandZExtPseudo(MI, LowOpcode: SystemZ::LLCR, Size: 8); |
1631 | return true; |
1632 | |
1633 | case SystemZ::LLHRMux: |
1634 | expandZExtPseudo(MI, LowOpcode: SystemZ::LLHR, Size: 16); |
1635 | return true; |
1636 | |
1637 | case SystemZ::LLCMux: |
1638 | expandRXYPseudo(MI, LowOpcode: SystemZ::LLC, HighOpcode: SystemZ::LLCH); |
1639 | return true; |
1640 | |
1641 | case SystemZ::LLHMux: |
1642 | expandRXYPseudo(MI, LowOpcode: SystemZ::LLH, HighOpcode: SystemZ::LLHH); |
1643 | return true; |
1644 | |
1645 | case SystemZ::LMux: |
1646 | expandRXYPseudo(MI, LowOpcode: SystemZ::L, HighOpcode: SystemZ::LFH); |
1647 | return true; |
1648 | |
1649 | case SystemZ::LOCMux: |
1650 | expandLOCPseudo(MI, LowOpcode: SystemZ::LOC, HighOpcode: SystemZ::LOCFH); |
1651 | return true; |
1652 | |
1653 | case SystemZ::LOCHIMux: |
1654 | expandLOCPseudo(MI, LowOpcode: SystemZ::LOCHI, HighOpcode: SystemZ::LOCHHI); |
1655 | return true; |
1656 | |
1657 | case SystemZ::STCMux: |
1658 | expandRXYPseudo(MI, LowOpcode: SystemZ::STC, HighOpcode: SystemZ::STCH); |
1659 | return true; |
1660 | |
1661 | case SystemZ::STHMux: |
1662 | expandRXYPseudo(MI, LowOpcode: SystemZ::STH, HighOpcode: SystemZ::STHH); |
1663 | return true; |
1664 | |
1665 | case SystemZ::STMux: |
1666 | expandRXYPseudo(MI, LowOpcode: SystemZ::ST, HighOpcode: SystemZ::STFH); |
1667 | return true; |
1668 | |
1669 | case SystemZ::STOCMux: |
1670 | expandLOCPseudo(MI, LowOpcode: SystemZ::STOC, HighOpcode: SystemZ::STOCFH); |
1671 | return true; |
1672 | |
1673 | case SystemZ::LHIMux: |
1674 | expandRIPseudo(MI, LowOpcode: SystemZ::LHI, HighOpcode: SystemZ::IIHF, ConvertHigh: true); |
1675 | return true; |
1676 | |
1677 | case SystemZ::IIFMux: |
1678 | expandRIPseudo(MI, LowOpcode: SystemZ::IILF, HighOpcode: SystemZ::IIHF, ConvertHigh: false); |
1679 | return true; |
1680 | |
1681 | case SystemZ::IILMux: |
1682 | expandRIPseudo(MI, LowOpcode: SystemZ::IILL, HighOpcode: SystemZ::IIHL, ConvertHigh: false); |
1683 | return true; |
1684 | |
1685 | case SystemZ::IIHMux: |
1686 | expandRIPseudo(MI, LowOpcode: SystemZ::IILH, HighOpcode: SystemZ::IIHH, ConvertHigh: false); |
1687 | return true; |
1688 | |
1689 | case SystemZ::NIFMux: |
1690 | expandRIPseudo(MI, LowOpcode: SystemZ::NILF, HighOpcode: SystemZ::NIHF, ConvertHigh: false); |
1691 | return true; |
1692 | |
1693 | case SystemZ::NILMux: |
1694 | expandRIPseudo(MI, LowOpcode: SystemZ::NILL, HighOpcode: SystemZ::NIHL, ConvertHigh: false); |
1695 | return true; |
1696 | |
1697 | case SystemZ::NIHMux: |
1698 | expandRIPseudo(MI, LowOpcode: SystemZ::NILH, HighOpcode: SystemZ::NIHH, ConvertHigh: false); |
1699 | return true; |
1700 | |
1701 | case SystemZ::OIFMux: |
1702 | expandRIPseudo(MI, LowOpcode: SystemZ::OILF, HighOpcode: SystemZ::OIHF, ConvertHigh: false); |
1703 | return true; |
1704 | |
1705 | case SystemZ::OILMux: |
1706 | expandRIPseudo(MI, LowOpcode: SystemZ::OILL, HighOpcode: SystemZ::OIHL, ConvertHigh: false); |
1707 | return true; |
1708 | |
1709 | case SystemZ::OIHMux: |
1710 | expandRIPseudo(MI, LowOpcode: SystemZ::OILH, HighOpcode: SystemZ::OIHH, ConvertHigh: false); |
1711 | return true; |
1712 | |
1713 | case SystemZ::XIFMux: |
1714 | expandRIPseudo(MI, LowOpcode: SystemZ::XILF, HighOpcode: SystemZ::XIHF, ConvertHigh: false); |
1715 | return true; |
1716 | |
1717 | case SystemZ::TMLMux: |
1718 | expandRIPseudo(MI, LowOpcode: SystemZ::TMLL, HighOpcode: SystemZ::TMHL, ConvertHigh: false); |
1719 | return true; |
1720 | |
1721 | case SystemZ::TMHMux: |
1722 | expandRIPseudo(MI, LowOpcode: SystemZ::TMLH, HighOpcode: SystemZ::TMHH, ConvertHigh: false); |
1723 | return true; |
1724 | |
1725 | case SystemZ::AHIMux: |
1726 | expandRIPseudo(MI, LowOpcode: SystemZ::AHI, HighOpcode: SystemZ::AIH, ConvertHigh: false); |
1727 | return true; |
1728 | |
1729 | case SystemZ::AHIMuxK: |
1730 | expandRIEPseudo(MI, LowOpcode: SystemZ::AHI, LowOpcodeK: SystemZ::AHIK, HighOpcode: SystemZ::AIH); |
1731 | return true; |
1732 | |
1733 | case SystemZ::AFIMux: |
1734 | expandRIPseudo(MI, LowOpcode: SystemZ::AFI, HighOpcode: SystemZ::AIH, ConvertHigh: false); |
1735 | return true; |
1736 | |
1737 | case SystemZ::CHIMux: |
1738 | expandRIPseudo(MI, LowOpcode: SystemZ::CHI, HighOpcode: SystemZ::CIH, ConvertHigh: false); |
1739 | return true; |
1740 | |
1741 | case SystemZ::CFIMux: |
1742 | expandRIPseudo(MI, LowOpcode: SystemZ::CFI, HighOpcode: SystemZ::CIH, ConvertHigh: false); |
1743 | return true; |
1744 | |
1745 | case SystemZ::CLFIMux: |
1746 | expandRIPseudo(MI, LowOpcode: SystemZ::CLFI, HighOpcode: SystemZ::CLIH, ConvertHigh: false); |
1747 | return true; |
1748 | |
1749 | case SystemZ::CMux: |
1750 | expandRXYPseudo(MI, LowOpcode: SystemZ::C, HighOpcode: SystemZ::CHF); |
1751 | return true; |
1752 | |
1753 | case SystemZ::CLMux: |
1754 | expandRXYPseudo(MI, LowOpcode: SystemZ::CL, HighOpcode: SystemZ::CLHF); |
1755 | return true; |
1756 | |
1757 | case SystemZ::RISBMux: { |
1758 | bool DestIsHigh = SystemZ::isHighReg(Reg: MI.getOperand(i: 0).getReg()); |
1759 | bool SrcIsHigh = SystemZ::isHighReg(Reg: MI.getOperand(i: 2).getReg()); |
1760 | if (SrcIsHigh == DestIsHigh) |
1761 | MI.setDesc(get(Opcode: DestIsHigh ? SystemZ::RISBHH : SystemZ::RISBLL)); |
1762 | else { |
1763 | MI.setDesc(get(Opcode: DestIsHigh ? SystemZ::RISBHL : SystemZ::RISBLH)); |
1764 | MI.getOperand(i: 5).setImm(MI.getOperand(i: 5).getImm() ^ 32); |
1765 | } |
1766 | return true; |
1767 | } |
1768 | |
1769 | case SystemZ::ADJDYNALLOC: |
1770 | splitAdjDynAlloc(MI); |
1771 | return true; |
1772 | |
1773 | case TargetOpcode::LOAD_STACK_GUARD: |
1774 | expandLoadStackGuard(MI: &MI); |
1775 | return true; |
1776 | |
1777 | default: |
1778 | return false; |
1779 | } |
1780 | } |
1781 | |
1782 | unsigned SystemZInstrInfo::getInstSizeInBytes(const MachineInstr &MI) const { |
1783 | if (MI.isInlineAsm()) { |
1784 | const MachineFunction *MF = MI.getParent()->getParent(); |
1785 | const char *AsmStr = MI.getOperand(i: 0).getSymbolName(); |
1786 | return getInlineAsmLength(Str: AsmStr, MAI: *MF->getTarget().getMCAsmInfo()); |
1787 | } |
1788 | else if (MI.getOpcode() == SystemZ::PATCHPOINT) |
1789 | return PatchPointOpers(&MI).getNumPatchBytes(); |
1790 | else if (MI.getOpcode() == SystemZ::STACKMAP) |
1791 | return MI.getOperand(i: 1).getImm(); |
1792 | else if (MI.getOpcode() == SystemZ::FENTRY_CALL) |
1793 | return 6; |
1794 | |
1795 | return MI.getDesc().getSize(); |
1796 | } |
1797 | |
1798 | SystemZII::Branch |
1799 | SystemZInstrInfo::getBranchInfo(const MachineInstr &MI) const { |
1800 | switch (MI.getOpcode()) { |
1801 | case SystemZ::BR: |
1802 | case SystemZ::BI: |
1803 | case SystemZ::J: |
1804 | case SystemZ::JG: |
1805 | return SystemZII::Branch(SystemZII::BranchNormal, SystemZ::CCMASK_ANY, |
1806 | SystemZ::CCMASK_ANY, &MI.getOperand(i: 0)); |
1807 | |
1808 | case SystemZ::BRC: |
1809 | case SystemZ::BRCL: |
1810 | return SystemZII::Branch(SystemZII::BranchNormal, MI.getOperand(i: 0).getImm(), |
1811 | MI.getOperand(i: 1).getImm(), &MI.getOperand(i: 2)); |
1812 | |
1813 | case SystemZ::BRCT: |
1814 | case SystemZ::BRCTH: |
1815 | return SystemZII::Branch(SystemZII::BranchCT, SystemZ::CCMASK_ICMP, |
1816 | SystemZ::CCMASK_CMP_NE, &MI.getOperand(i: 2)); |
1817 | |
1818 | case SystemZ::BRCTG: |
1819 | return SystemZII::Branch(SystemZII::BranchCTG, SystemZ::CCMASK_ICMP, |
1820 | SystemZ::CCMASK_CMP_NE, &MI.getOperand(i: 2)); |
1821 | |
1822 | case SystemZ::CIJ: |
1823 | case SystemZ::CRJ: |
1824 | return SystemZII::Branch(SystemZII::BranchC, SystemZ::CCMASK_ICMP, |
1825 | MI.getOperand(i: 2).getImm(), &MI.getOperand(i: 3)); |
1826 | |
1827 | case SystemZ::CLIJ: |
1828 | case SystemZ::CLRJ: |
1829 | return SystemZII::Branch(SystemZII::BranchCL, SystemZ::CCMASK_ICMP, |
1830 | MI.getOperand(i: 2).getImm(), &MI.getOperand(i: 3)); |
1831 | |
1832 | case SystemZ::CGIJ: |
1833 | case SystemZ::CGRJ: |
1834 | return SystemZII::Branch(SystemZII::BranchCG, SystemZ::CCMASK_ICMP, |
1835 | MI.getOperand(i: 2).getImm(), &MI.getOperand(i: 3)); |
1836 | |
1837 | case SystemZ::CLGIJ: |
1838 | case SystemZ::CLGRJ: |
1839 | return SystemZII::Branch(SystemZII::BranchCLG, SystemZ::CCMASK_ICMP, |
1840 | MI.getOperand(i: 2).getImm(), &MI.getOperand(i: 3)); |
1841 | |
1842 | case SystemZ::INLINEASM_BR: |
1843 | // Don't try to analyze asm goto, so pass nullptr as branch target argument. |
1844 | return SystemZII::Branch(SystemZII::AsmGoto, 0, 0, nullptr); |
1845 | |
1846 | default: |
1847 | llvm_unreachable("Unrecognized branch opcode" ); |
1848 | } |
1849 | } |
1850 | |
1851 | void SystemZInstrInfo::getLoadStoreOpcodes(const TargetRegisterClass *RC, |
1852 | unsigned &LoadOpcode, |
1853 | unsigned &StoreOpcode) const { |
1854 | if (RC == &SystemZ::GR32BitRegClass || RC == &SystemZ::ADDR32BitRegClass) { |
1855 | LoadOpcode = SystemZ::L; |
1856 | StoreOpcode = SystemZ::ST; |
1857 | } else if (RC == &SystemZ::GRH32BitRegClass) { |
1858 | LoadOpcode = SystemZ::LFH; |
1859 | StoreOpcode = SystemZ::STFH; |
1860 | } else if (RC == &SystemZ::GRX32BitRegClass) { |
1861 | LoadOpcode = SystemZ::LMux; |
1862 | StoreOpcode = SystemZ::STMux; |
1863 | } else if (RC == &SystemZ::GR64BitRegClass || |
1864 | RC == &SystemZ::ADDR64BitRegClass) { |
1865 | LoadOpcode = SystemZ::LG; |
1866 | StoreOpcode = SystemZ::STG; |
1867 | } else if (RC == &SystemZ::GR128BitRegClass || |
1868 | RC == &SystemZ::ADDR128BitRegClass) { |
1869 | LoadOpcode = SystemZ::L128; |
1870 | StoreOpcode = SystemZ::ST128; |
1871 | } else if (RC == &SystemZ::FP32BitRegClass) { |
1872 | LoadOpcode = SystemZ::LE; |
1873 | StoreOpcode = SystemZ::STE; |
1874 | } else if (RC == &SystemZ::FP64BitRegClass) { |
1875 | LoadOpcode = SystemZ::LD; |
1876 | StoreOpcode = SystemZ::STD; |
1877 | } else if (RC == &SystemZ::FP128BitRegClass) { |
1878 | LoadOpcode = SystemZ::LX; |
1879 | StoreOpcode = SystemZ::STX; |
1880 | } else if (RC == &SystemZ::VR32BitRegClass) { |
1881 | LoadOpcode = SystemZ::VL32; |
1882 | StoreOpcode = SystemZ::VST32; |
1883 | } else if (RC == &SystemZ::VR64BitRegClass) { |
1884 | LoadOpcode = SystemZ::VL64; |
1885 | StoreOpcode = SystemZ::VST64; |
1886 | } else if (RC == &SystemZ::VF128BitRegClass || |
1887 | RC == &SystemZ::VR128BitRegClass) { |
1888 | LoadOpcode = SystemZ::VL; |
1889 | StoreOpcode = SystemZ::VST; |
1890 | } else |
1891 | llvm_unreachable("Unsupported regclass to load or store" ); |
1892 | } |
1893 | |
1894 | unsigned SystemZInstrInfo::getOpcodeForOffset(unsigned Opcode, |
1895 | int64_t Offset, |
1896 | const MachineInstr *MI) const { |
1897 | const MCInstrDesc &MCID = get(Opcode); |
1898 | int64_t Offset2 = (MCID.TSFlags & SystemZII::Is128Bit ? Offset + 8 : Offset); |
1899 | if (isUInt<12>(x: Offset) && isUInt<12>(x: Offset2)) { |
1900 | // Get the instruction to use for unsigned 12-bit displacements. |
1901 | int Disp12Opcode = SystemZ::getDisp12Opcode(Opcode); |
1902 | if (Disp12Opcode >= 0) |
1903 | return Disp12Opcode; |
1904 | |
1905 | // All address-related instructions can use unsigned 12-bit |
1906 | // displacements. |
1907 | return Opcode; |
1908 | } |
1909 | if (isInt<20>(x: Offset) && isInt<20>(x: Offset2)) { |
1910 | // Get the instruction to use for signed 20-bit displacements. |
1911 | int Disp20Opcode = SystemZ::getDisp20Opcode(Opcode); |
1912 | if (Disp20Opcode >= 0) |
1913 | return Disp20Opcode; |
1914 | |
1915 | // Check whether Opcode allows signed 20-bit displacements. |
1916 | if (MCID.TSFlags & SystemZII::Has20BitOffset) |
1917 | return Opcode; |
1918 | |
1919 | // If a VR32/VR64 reg ended up in an FP register, use the FP opcode. |
1920 | if (MI && MI->getOperand(i: 0).isReg()) { |
1921 | Register Reg = MI->getOperand(i: 0).getReg(); |
1922 | if (Reg.isPhysical() && SystemZMC::getFirstReg(Reg) < 16) { |
1923 | switch (Opcode) { |
1924 | case SystemZ::VL32: |
1925 | return SystemZ::LEY; |
1926 | case SystemZ::VST32: |
1927 | return SystemZ::STEY; |
1928 | case SystemZ::VL64: |
1929 | return SystemZ::LDY; |
1930 | case SystemZ::VST64: |
1931 | return SystemZ::STDY; |
1932 | default: break; |
1933 | } |
1934 | } |
1935 | } |
1936 | } |
1937 | return 0; |
1938 | } |
1939 | |
1940 | bool SystemZInstrInfo::hasDisplacementPairInsn(unsigned Opcode) const { |
1941 | const MCInstrDesc &MCID = get(Opcode); |
1942 | if (MCID.TSFlags & SystemZII::Has20BitOffset) |
1943 | return SystemZ::getDisp12Opcode(Opcode) >= 0; |
1944 | return SystemZ::getDisp20Opcode(Opcode) >= 0; |
1945 | } |
1946 | |
1947 | unsigned SystemZInstrInfo::getLoadAndTest(unsigned Opcode) const { |
1948 | switch (Opcode) { |
1949 | case SystemZ::L: return SystemZ::LT; |
1950 | case SystemZ::LY: return SystemZ::LT; |
1951 | case SystemZ::LG: return SystemZ::LTG; |
1952 | case SystemZ::LGF: return SystemZ::LTGF; |
1953 | case SystemZ::LR: return SystemZ::LTR; |
1954 | case SystemZ::LGFR: return SystemZ::LTGFR; |
1955 | case SystemZ::LGR: return SystemZ::LTGR; |
1956 | case SystemZ::LCDFR: return SystemZ::LCDBR; |
1957 | case SystemZ::LPDFR: return SystemZ::LPDBR; |
1958 | case SystemZ::LNDFR: return SystemZ::LNDBR; |
1959 | case SystemZ::LCDFR_32: return SystemZ::LCEBR; |
1960 | case SystemZ::LPDFR_32: return SystemZ::LPEBR; |
1961 | case SystemZ::LNDFR_32: return SystemZ::LNEBR; |
1962 | // On zEC12 we prefer to use RISBGN. But if there is a chance to |
1963 | // actually use the condition code, we may turn it back into RISGB. |
1964 | // Note that RISBG is not really a "load-and-test" instruction, |
1965 | // but sets the same condition code values, so is OK to use here. |
1966 | case SystemZ::RISBGN: return SystemZ::RISBG; |
1967 | default: return 0; |
1968 | } |
1969 | } |
1970 | |
1971 | bool SystemZInstrInfo::isRxSBGMask(uint64_t Mask, unsigned BitSize, |
1972 | unsigned &Start, unsigned &End) const { |
1973 | // Reject trivial all-zero masks. |
1974 | Mask &= allOnes(Count: BitSize); |
1975 | if (Mask == 0) |
1976 | return false; |
1977 | |
1978 | // Handle the 1+0+ or 0+1+0* cases. Start then specifies the index of |
1979 | // the msb and End specifies the index of the lsb. |
1980 | unsigned LSB, Length; |
1981 | if (isShiftedMask_64(Value: Mask, MaskIdx&: LSB, MaskLen&: Length)) { |
1982 | Start = 63 - (LSB + Length - 1); |
1983 | End = 63 - LSB; |
1984 | return true; |
1985 | } |
1986 | |
1987 | // Handle the wrap-around 1+0+1+ cases. Start then specifies the msb |
1988 | // of the low 1s and End specifies the lsb of the high 1s. |
1989 | if (isShiftedMask_64(Value: Mask ^ allOnes(Count: BitSize), MaskIdx&: LSB, MaskLen&: Length)) { |
1990 | assert(LSB > 0 && "Bottom bit must be set" ); |
1991 | assert(LSB + Length < BitSize && "Top bit must be set" ); |
1992 | Start = 63 - (LSB - 1); |
1993 | End = 63 - (LSB + Length); |
1994 | return true; |
1995 | } |
1996 | |
1997 | return false; |
1998 | } |
1999 | |
2000 | unsigned SystemZInstrInfo::getFusedCompare(unsigned Opcode, |
2001 | SystemZII::FusedCompareType Type, |
2002 | const MachineInstr *MI) const { |
2003 | switch (Opcode) { |
2004 | case SystemZ::CHI: |
2005 | case SystemZ::CGHI: |
2006 | if (!(MI && isInt<8>(x: MI->getOperand(i: 1).getImm()))) |
2007 | return 0; |
2008 | break; |
2009 | case SystemZ::CLFI: |
2010 | case SystemZ::CLGFI: |
2011 | if (!(MI && isUInt<8>(x: MI->getOperand(i: 1).getImm()))) |
2012 | return 0; |
2013 | break; |
2014 | case SystemZ::CL: |
2015 | case SystemZ::CLG: |
2016 | if (!STI.hasMiscellaneousExtensions()) |
2017 | return 0; |
2018 | if (!(MI && MI->getOperand(i: 3).getReg() == 0)) |
2019 | return 0; |
2020 | break; |
2021 | } |
2022 | switch (Type) { |
2023 | case SystemZII::CompareAndBranch: |
2024 | switch (Opcode) { |
2025 | case SystemZ::CR: |
2026 | return SystemZ::CRJ; |
2027 | case SystemZ::CGR: |
2028 | return SystemZ::CGRJ; |
2029 | case SystemZ::CHI: |
2030 | return SystemZ::CIJ; |
2031 | case SystemZ::CGHI: |
2032 | return SystemZ::CGIJ; |
2033 | case SystemZ::CLR: |
2034 | return SystemZ::CLRJ; |
2035 | case SystemZ::CLGR: |
2036 | return SystemZ::CLGRJ; |
2037 | case SystemZ::CLFI: |
2038 | return SystemZ::CLIJ; |
2039 | case SystemZ::CLGFI: |
2040 | return SystemZ::CLGIJ; |
2041 | default: |
2042 | return 0; |
2043 | } |
2044 | case SystemZII::CompareAndReturn: |
2045 | switch (Opcode) { |
2046 | case SystemZ::CR: |
2047 | return SystemZ::CRBReturn; |
2048 | case SystemZ::CGR: |
2049 | return SystemZ::CGRBReturn; |
2050 | case SystemZ::CHI: |
2051 | return SystemZ::CIBReturn; |
2052 | case SystemZ::CGHI: |
2053 | return SystemZ::CGIBReturn; |
2054 | case SystemZ::CLR: |
2055 | return SystemZ::CLRBReturn; |
2056 | case SystemZ::CLGR: |
2057 | return SystemZ::CLGRBReturn; |
2058 | case SystemZ::CLFI: |
2059 | return SystemZ::CLIBReturn; |
2060 | case SystemZ::CLGFI: |
2061 | return SystemZ::CLGIBReturn; |
2062 | default: |
2063 | return 0; |
2064 | } |
2065 | case SystemZII::CompareAndSibcall: |
2066 | switch (Opcode) { |
2067 | case SystemZ::CR: |
2068 | return SystemZ::CRBCall; |
2069 | case SystemZ::CGR: |
2070 | return SystemZ::CGRBCall; |
2071 | case SystemZ::CHI: |
2072 | return SystemZ::CIBCall; |
2073 | case SystemZ::CGHI: |
2074 | return SystemZ::CGIBCall; |
2075 | case SystemZ::CLR: |
2076 | return SystemZ::CLRBCall; |
2077 | case SystemZ::CLGR: |
2078 | return SystemZ::CLGRBCall; |
2079 | case SystemZ::CLFI: |
2080 | return SystemZ::CLIBCall; |
2081 | case SystemZ::CLGFI: |
2082 | return SystemZ::CLGIBCall; |
2083 | default: |
2084 | return 0; |
2085 | } |
2086 | case SystemZII::CompareAndTrap: |
2087 | switch (Opcode) { |
2088 | case SystemZ::CR: |
2089 | return SystemZ::CRT; |
2090 | case SystemZ::CGR: |
2091 | return SystemZ::CGRT; |
2092 | case SystemZ::CHI: |
2093 | return SystemZ::CIT; |
2094 | case SystemZ::CGHI: |
2095 | return SystemZ::CGIT; |
2096 | case SystemZ::CLR: |
2097 | return SystemZ::CLRT; |
2098 | case SystemZ::CLGR: |
2099 | return SystemZ::CLGRT; |
2100 | case SystemZ::CLFI: |
2101 | return SystemZ::CLFIT; |
2102 | case SystemZ::CLGFI: |
2103 | return SystemZ::CLGIT; |
2104 | case SystemZ::CL: |
2105 | return SystemZ::CLT; |
2106 | case SystemZ::CLG: |
2107 | return SystemZ::CLGT; |
2108 | default: |
2109 | return 0; |
2110 | } |
2111 | } |
2112 | return 0; |
2113 | } |
2114 | |
2115 | bool SystemZInstrInfo:: |
2116 | prepareCompareSwapOperands(MachineBasicBlock::iterator const MBBI) const { |
2117 | assert(MBBI->isCompare() && MBBI->getOperand(0).isReg() && |
2118 | MBBI->getOperand(1).isReg() && !MBBI->mayLoad() && |
2119 | "Not a compare reg/reg." ); |
2120 | |
2121 | MachineBasicBlock *MBB = MBBI->getParent(); |
2122 | bool CCLive = true; |
2123 | SmallVector<MachineInstr *, 4> CCUsers; |
2124 | for (MachineInstr &MI : llvm::make_range(x: std::next(x: MBBI), y: MBB->end())) { |
2125 | if (MI.readsRegister(Reg: SystemZ::CC, /*TRI=*/nullptr)) { |
2126 | unsigned Flags = MI.getDesc().TSFlags; |
2127 | if ((Flags & SystemZII::CCMaskFirst) || (Flags & SystemZII::CCMaskLast)) |
2128 | CCUsers.push_back(Elt: &MI); |
2129 | else |
2130 | return false; |
2131 | } |
2132 | if (MI.definesRegister(Reg: SystemZ::CC, /*TRI=*/nullptr)) { |
2133 | CCLive = false; |
2134 | break; |
2135 | } |
2136 | } |
2137 | if (CCLive) { |
2138 | LiveRegUnits LiveRegs(*MBB->getParent()->getSubtarget().getRegisterInfo()); |
2139 | LiveRegs.addLiveOuts(MBB: *MBB); |
2140 | if (!LiveRegs.available(Reg: SystemZ::CC)) |
2141 | return false; |
2142 | } |
2143 | |
2144 | // Update all CC users. |
2145 | for (unsigned Idx = 0; Idx < CCUsers.size(); ++Idx) { |
2146 | unsigned Flags = CCUsers[Idx]->getDesc().TSFlags; |
2147 | unsigned FirstOpNum = ((Flags & SystemZII::CCMaskFirst) ? |
2148 | 0 : CCUsers[Idx]->getNumExplicitOperands() - 2); |
2149 | MachineOperand &CCMaskMO = CCUsers[Idx]->getOperand(i: FirstOpNum + 1); |
2150 | unsigned NewCCMask = SystemZ::reverseCCMask(CCMask: CCMaskMO.getImm()); |
2151 | CCMaskMO.setImm(NewCCMask); |
2152 | } |
2153 | |
2154 | return true; |
2155 | } |
2156 | |
2157 | unsigned SystemZ::reverseCCMask(unsigned CCMask) { |
2158 | return ((CCMask & SystemZ::CCMASK_CMP_EQ) | |
2159 | ((CCMask & SystemZ::CCMASK_CMP_GT) ? SystemZ::CCMASK_CMP_LT : 0) | |
2160 | ((CCMask & SystemZ::CCMASK_CMP_LT) ? SystemZ::CCMASK_CMP_GT : 0) | |
2161 | (CCMask & SystemZ::CCMASK_CMP_UO)); |
2162 | } |
2163 | |
2164 | MachineBasicBlock *SystemZ::emitBlockAfter(MachineBasicBlock *MBB) { |
2165 | MachineFunction &MF = *MBB->getParent(); |
2166 | MachineBasicBlock *NewMBB = MF.CreateMachineBasicBlock(BB: MBB->getBasicBlock()); |
2167 | MF.insert(MBBI: std::next(x: MachineFunction::iterator(MBB)), MBB: NewMBB); |
2168 | return NewMBB; |
2169 | } |
2170 | |
2171 | MachineBasicBlock *SystemZ::splitBlockAfter(MachineBasicBlock::iterator MI, |
2172 | MachineBasicBlock *MBB) { |
2173 | MachineBasicBlock *NewMBB = emitBlockAfter(MBB); |
2174 | NewMBB->splice(Where: NewMBB->begin(), Other: MBB, |
2175 | From: std::next(x: MachineBasicBlock::iterator(MI)), To: MBB->end()); |
2176 | NewMBB->transferSuccessorsAndUpdatePHIs(FromMBB: MBB); |
2177 | return NewMBB; |
2178 | } |
2179 | |
2180 | MachineBasicBlock *SystemZ::splitBlockBefore(MachineBasicBlock::iterator MI, |
2181 | MachineBasicBlock *MBB) { |
2182 | MachineBasicBlock *NewMBB = emitBlockAfter(MBB); |
2183 | NewMBB->splice(Where: NewMBB->begin(), Other: MBB, From: MI, To: MBB->end()); |
2184 | NewMBB->transferSuccessorsAndUpdatePHIs(FromMBB: MBB); |
2185 | return NewMBB; |
2186 | } |
2187 | |
2188 | unsigned SystemZInstrInfo::getLoadAndTrap(unsigned Opcode) const { |
2189 | if (!STI.hasLoadAndTrap()) |
2190 | return 0; |
2191 | switch (Opcode) { |
2192 | case SystemZ::L: |
2193 | case SystemZ::LY: |
2194 | return SystemZ::LAT; |
2195 | case SystemZ::LG: |
2196 | return SystemZ::LGAT; |
2197 | case SystemZ::LFH: |
2198 | return SystemZ::LFHAT; |
2199 | case SystemZ::LLGF: |
2200 | return SystemZ::LLGFAT; |
2201 | case SystemZ::LLGT: |
2202 | return SystemZ::LLGTAT; |
2203 | } |
2204 | return 0; |
2205 | } |
2206 | |
2207 | void SystemZInstrInfo::loadImmediate(MachineBasicBlock &MBB, |
2208 | MachineBasicBlock::iterator MBBI, |
2209 | unsigned Reg, uint64_t Value) const { |
2210 | DebugLoc DL = MBBI != MBB.end() ? MBBI->getDebugLoc() : DebugLoc(); |
2211 | unsigned Opcode = 0; |
2212 | if (isInt<16>(x: Value)) |
2213 | Opcode = SystemZ::LGHI; |
2214 | else if (SystemZ::isImmLL(Val: Value)) |
2215 | Opcode = SystemZ::LLILL; |
2216 | else if (SystemZ::isImmLH(Val: Value)) { |
2217 | Opcode = SystemZ::LLILH; |
2218 | Value >>= 16; |
2219 | } |
2220 | else if (isInt<32>(x: Value)) |
2221 | Opcode = SystemZ::LGFI; |
2222 | if (Opcode) { |
2223 | BuildMI(BB&: MBB, I: MBBI, MIMD: DL, MCID: get(Opcode), DestReg: Reg).addImm(Val: Value); |
2224 | return; |
2225 | } |
2226 | |
2227 | MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo(); |
2228 | assert (MRI.isSSA() && "Huge values only handled before reg-alloc ." ); |
2229 | Register Reg0 = MRI.createVirtualRegister(RegClass: &SystemZ::GR64BitRegClass); |
2230 | Register Reg1 = MRI.createVirtualRegister(RegClass: &SystemZ::GR64BitRegClass); |
2231 | BuildMI(BB&: MBB, I: MBBI, MIMD: DL, MCID: get(Opcode: SystemZ::IMPLICIT_DEF), DestReg: Reg0); |
2232 | BuildMI(BB&: MBB, I: MBBI, MIMD: DL, MCID: get(Opcode: SystemZ::IIHF64), DestReg: Reg1) |
2233 | .addReg(RegNo: Reg0).addImm(Val: Value >> 32); |
2234 | BuildMI(BB&: MBB, I: MBBI, MIMD: DL, MCID: get(Opcode: SystemZ::IILF64), DestReg: Reg) |
2235 | .addReg(RegNo: Reg1).addImm(Val: Value & ((uint64_t(1) << 32) - 1)); |
2236 | } |
2237 | |
2238 | bool SystemZInstrInfo::verifyInstruction(const MachineInstr &MI, |
2239 | StringRef &ErrInfo) const { |
2240 | const MCInstrDesc &MCID = MI.getDesc(); |
2241 | for (unsigned I = 0, E = MI.getNumOperands(); I != E; ++I) { |
2242 | if (I >= MCID.getNumOperands()) |
2243 | break; |
2244 | const MachineOperand &Op = MI.getOperand(i: I); |
2245 | const MCOperandInfo &MCOI = MCID.operands()[I]; |
2246 | // Addressing modes have register and immediate operands. Op should be a |
2247 | // register (or frame index) operand if MCOI.RegClass contains a valid |
2248 | // register class, or an immediate otherwise. |
2249 | if (MCOI.OperandType == MCOI::OPERAND_MEMORY && |
2250 | ((MCOI.RegClass != -1 && !Op.isReg() && !Op.isFI()) || |
2251 | (MCOI.RegClass == -1 && !Op.isImm()))) { |
2252 | ErrInfo = "Addressing mode operands corrupt!" ; |
2253 | return false; |
2254 | } |
2255 | } |
2256 | |
2257 | return true; |
2258 | } |
2259 | |
2260 | bool SystemZInstrInfo:: |
2261 | areMemAccessesTriviallyDisjoint(const MachineInstr &MIa, |
2262 | const MachineInstr &MIb) const { |
2263 | |
2264 | if (!MIa.hasOneMemOperand() || !MIb.hasOneMemOperand()) |
2265 | return false; |
2266 | |
2267 | // If mem-operands show that the same address Value is used by both |
2268 | // instructions, check for non-overlapping offsets and widths. Not |
2269 | // sure if a register based analysis would be an improvement... |
2270 | |
2271 | MachineMemOperand *MMOa = *MIa.memoperands_begin(); |
2272 | MachineMemOperand *MMOb = *MIb.memoperands_begin(); |
2273 | const Value *VALa = MMOa->getValue(); |
2274 | const Value *VALb = MMOb->getValue(); |
2275 | bool SameVal = (VALa && VALb && (VALa == VALb)); |
2276 | if (!SameVal) { |
2277 | const PseudoSourceValue *PSVa = MMOa->getPseudoValue(); |
2278 | const PseudoSourceValue *PSVb = MMOb->getPseudoValue(); |
2279 | if (PSVa && PSVb && (PSVa == PSVb)) |
2280 | SameVal = true; |
2281 | } |
2282 | if (SameVal) { |
2283 | int OffsetA = MMOa->getOffset(), OffsetB = MMOb->getOffset(); |
2284 | LocationSize WidthA = MMOa->getSize(), WidthB = MMOb->getSize(); |
2285 | int LowOffset = OffsetA < OffsetB ? OffsetA : OffsetB; |
2286 | int HighOffset = OffsetA < OffsetB ? OffsetB : OffsetA; |
2287 | LocationSize LowWidth = (LowOffset == OffsetA) ? WidthA : WidthB; |
2288 | if (LowWidth.hasValue() && |
2289 | LowOffset + (int)LowWidth.getValue() <= HighOffset) |
2290 | return true; |
2291 | } |
2292 | |
2293 | return false; |
2294 | } |
2295 | |
2296 | bool SystemZInstrInfo::getConstValDefinedInReg(const MachineInstr &MI, |
2297 | const Register Reg, |
2298 | int64_t &ImmVal) const { |
2299 | |
2300 | if (MI.getOpcode() == SystemZ::VGBM && Reg == MI.getOperand(i: 0).getReg()) { |
2301 | ImmVal = MI.getOperand(i: 1).getImm(); |
2302 | // TODO: Handle non-0 values |
2303 | return ImmVal == 0; |
2304 | } |
2305 | |
2306 | return false; |
2307 | } |
2308 | |