[InstCombine] Signed saturation patterns
[llvm-complete.git] / lib / Target / ARM / Thumb2InstrInfo.cpp
blobaf1f0aeb27ba572ef36e1c897974a3c6697cebfd
1 //===- Thumb2InstrInfo.cpp - Thumb-2 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 Thumb-2 implementation of the TargetInstrInfo class.
11 //===----------------------------------------------------------------------===//
13 #include "Thumb2InstrInfo.h"
14 #include "ARMMachineFunctionInfo.h"
15 #include "MCTargetDesc/ARMAddressingModes.h"
16 #include "llvm/CodeGen/MachineBasicBlock.h"
17 #include "llvm/CodeGen/MachineFrameInfo.h"
18 #include "llvm/CodeGen/MachineFunction.h"
19 #include "llvm/CodeGen/MachineInstr.h"
20 #include "llvm/CodeGen/MachineInstrBuilder.h"
21 #include "llvm/CodeGen/MachineMemOperand.h"
22 #include "llvm/CodeGen/MachineOperand.h"
23 #include "llvm/CodeGen/MachineRegisterInfo.h"
24 #include "llvm/CodeGen/TargetRegisterInfo.h"
25 #include "llvm/IR/DebugLoc.h"
26 #include "llvm/MC/MCInst.h"
27 #include "llvm/MC/MCInstrDesc.h"
28 #include "llvm/Support/CommandLine.h"
29 #include "llvm/Support/ErrorHandling.h"
30 #include "llvm/Support/MathExtras.h"
31 #include "llvm/Target/TargetMachine.h"
32 #include <cassert>
34 using namespace llvm;
36 static cl::opt<bool>
37 OldT2IfCvt("old-thumb2-ifcvt", cl::Hidden,
38 cl::desc("Use old-style Thumb2 if-conversion heuristics"),
39 cl::init(false));
41 Thumb2InstrInfo::Thumb2InstrInfo(const ARMSubtarget &STI)
42 : ARMBaseInstrInfo(STI) {}
44 /// Return the noop instruction to use for a noop.
45 void Thumb2InstrInfo::getNoop(MCInst &NopInst) const {
46 NopInst.setOpcode(ARM::tHINT);
47 NopInst.addOperand(MCOperand::createImm(0));
48 NopInst.addOperand(MCOperand::createImm(ARMCC::AL));
49 NopInst.addOperand(MCOperand::createReg(0));
52 unsigned Thumb2InstrInfo::getUnindexedOpcode(unsigned Opc) const {
53 // FIXME
54 return 0;
57 void
58 Thumb2InstrInfo::ReplaceTailWithBranchTo(MachineBasicBlock::iterator Tail,
59 MachineBasicBlock *NewDest) const {
60 MachineBasicBlock *MBB = Tail->getParent();
61 ARMFunctionInfo *AFI = MBB->getParent()->getInfo<ARMFunctionInfo>();
62 if (!AFI->hasITBlocks() || Tail->isBranch()) {
63 TargetInstrInfo::ReplaceTailWithBranchTo(Tail, NewDest);
64 return;
67 // If the first instruction of Tail is predicated, we may have to update
68 // the IT instruction.
69 unsigned PredReg = 0;
70 ARMCC::CondCodes CC = getInstrPredicate(*Tail, PredReg);
71 MachineBasicBlock::iterator MBBI = Tail;
72 if (CC != ARMCC::AL)
73 // Expecting at least the t2IT instruction before it.
74 --MBBI;
76 // Actually replace the tail.
77 TargetInstrInfo::ReplaceTailWithBranchTo(Tail, NewDest);
79 // Fix up IT.
80 if (CC != ARMCC::AL) {
81 MachineBasicBlock::iterator E = MBB->begin();
82 unsigned Count = 4; // At most 4 instructions in an IT block.
83 while (Count && MBBI != E) {
84 if (MBBI->isDebugInstr()) {
85 --MBBI;
86 continue;
88 if (MBBI->getOpcode() == ARM::t2IT) {
89 unsigned Mask = MBBI->getOperand(1).getImm();
90 if (Count == 4)
91 MBBI->eraseFromParent();
92 else {
93 unsigned MaskOn = 1 << Count;
94 unsigned MaskOff = ~(MaskOn - 1);
95 MBBI->getOperand(1).setImm((Mask & MaskOff) | MaskOn);
97 return;
99 --MBBI;
100 --Count;
103 // Ctrl flow can reach here if branch folding is run before IT block
104 // formation pass.
108 bool
109 Thumb2InstrInfo::isLegalToSplitMBBAt(MachineBasicBlock &MBB,
110 MachineBasicBlock::iterator MBBI) const {
111 while (MBBI->isDebugInstr()) {
112 ++MBBI;
113 if (MBBI == MBB.end())
114 return false;
117 unsigned PredReg = 0;
118 return getITInstrPredicate(*MBBI, PredReg) == ARMCC::AL;
121 void Thumb2InstrInfo::copyPhysReg(MachineBasicBlock &MBB,
122 MachineBasicBlock::iterator I,
123 const DebugLoc &DL, unsigned DestReg,
124 unsigned SrcReg, bool KillSrc) const {
125 // Handle SPR, DPR, and QPR copies.
126 if (!ARM::GPRRegClass.contains(DestReg, SrcReg))
127 return ARMBaseInstrInfo::copyPhysReg(MBB, I, DL, DestReg, SrcReg, KillSrc);
129 BuildMI(MBB, I, DL, get(ARM::tMOVr), DestReg)
130 .addReg(SrcReg, getKillRegState(KillSrc))
131 .add(predOps(ARMCC::AL));
134 void Thumb2InstrInfo::
135 storeRegToStackSlot(MachineBasicBlock &MBB, MachineBasicBlock::iterator I,
136 unsigned SrcReg, bool isKill, int FI,
137 const TargetRegisterClass *RC,
138 const TargetRegisterInfo *TRI) const {
139 DebugLoc DL;
140 if (I != MBB.end()) DL = I->getDebugLoc();
142 MachineFunction &MF = *MBB.getParent();
143 MachineFrameInfo &MFI = MF.getFrameInfo();
144 MachineMemOperand *MMO = MF.getMachineMemOperand(
145 MachinePointerInfo::getFixedStack(MF, FI), MachineMemOperand::MOStore,
146 MFI.getObjectSize(FI), MFI.getObjectAlignment(FI));
148 if (ARM::GPRRegClass.hasSubClassEq(RC)) {
149 BuildMI(MBB, I, DL, get(ARM::t2STRi12))
150 .addReg(SrcReg, getKillRegState(isKill))
151 .addFrameIndex(FI)
152 .addImm(0)
153 .addMemOperand(MMO)
154 .add(predOps(ARMCC::AL));
155 return;
158 if (ARM::GPRPairRegClass.hasSubClassEq(RC)) {
159 // Thumb2 STRD expects its dest-registers to be in rGPR. Not a problem for
160 // gsub_0, but needs an extra constraint for gsub_1 (which could be sp
161 // otherwise).
162 if (Register::isVirtualRegister(SrcReg)) {
163 MachineRegisterInfo *MRI = &MF.getRegInfo();
164 MRI->constrainRegClass(SrcReg, &ARM::GPRPairnospRegClass);
167 MachineInstrBuilder MIB = BuildMI(MBB, I, DL, get(ARM::t2STRDi8));
168 AddDReg(MIB, SrcReg, ARM::gsub_0, getKillRegState(isKill), TRI);
169 AddDReg(MIB, SrcReg, ARM::gsub_1, 0, TRI);
170 MIB.addFrameIndex(FI).addImm(0).addMemOperand(MMO).add(predOps(ARMCC::AL));
171 return;
174 ARMBaseInstrInfo::storeRegToStackSlot(MBB, I, SrcReg, isKill, FI, RC, TRI);
177 void Thumb2InstrInfo::
178 loadRegFromStackSlot(MachineBasicBlock &MBB, MachineBasicBlock::iterator I,
179 unsigned DestReg, int FI,
180 const TargetRegisterClass *RC,
181 const TargetRegisterInfo *TRI) const {
182 MachineFunction &MF = *MBB.getParent();
183 MachineFrameInfo &MFI = MF.getFrameInfo();
184 MachineMemOperand *MMO = MF.getMachineMemOperand(
185 MachinePointerInfo::getFixedStack(MF, FI), MachineMemOperand::MOLoad,
186 MFI.getObjectSize(FI), MFI.getObjectAlignment(FI));
187 DebugLoc DL;
188 if (I != MBB.end()) DL = I->getDebugLoc();
190 if (ARM::GPRRegClass.hasSubClassEq(RC)) {
191 BuildMI(MBB, I, DL, get(ARM::t2LDRi12), DestReg)
192 .addFrameIndex(FI)
193 .addImm(0)
194 .addMemOperand(MMO)
195 .add(predOps(ARMCC::AL));
196 return;
199 if (ARM::GPRPairRegClass.hasSubClassEq(RC)) {
200 // Thumb2 LDRD expects its dest-registers to be in rGPR. Not a problem for
201 // gsub_0, but needs an extra constraint for gsub_1 (which could be sp
202 // otherwise).
203 if (Register::isVirtualRegister(DestReg)) {
204 MachineRegisterInfo *MRI = &MF.getRegInfo();
205 MRI->constrainRegClass(DestReg, &ARM::GPRPairnospRegClass);
208 MachineInstrBuilder MIB = BuildMI(MBB, I, DL, get(ARM::t2LDRDi8));
209 AddDReg(MIB, DestReg, ARM::gsub_0, RegState::DefineNoRead, TRI);
210 AddDReg(MIB, DestReg, ARM::gsub_1, RegState::DefineNoRead, TRI);
211 MIB.addFrameIndex(FI).addImm(0).addMemOperand(MMO).add(predOps(ARMCC::AL));
213 if (Register::isPhysicalRegister(DestReg))
214 MIB.addReg(DestReg, RegState::ImplicitDefine);
215 return;
218 ARMBaseInstrInfo::loadRegFromStackSlot(MBB, I, DestReg, FI, RC, TRI);
221 void Thumb2InstrInfo::expandLoadStackGuard(
222 MachineBasicBlock::iterator MI) const {
223 MachineFunction &MF = *MI->getParent()->getParent();
224 if (MF.getTarget().isPositionIndependent())
225 expandLoadStackGuardBase(MI, ARM::t2MOV_ga_pcrel, ARM::t2LDRi12);
226 else
227 expandLoadStackGuardBase(MI, ARM::t2MOVi32imm, ARM::t2LDRi12);
230 void llvm::emitT2RegPlusImmediate(MachineBasicBlock &MBB,
231 MachineBasicBlock::iterator &MBBI,
232 const DebugLoc &dl, unsigned DestReg,
233 unsigned BaseReg, int NumBytes,
234 ARMCC::CondCodes Pred, unsigned PredReg,
235 const ARMBaseInstrInfo &TII,
236 unsigned MIFlags) {
237 if (NumBytes == 0 && DestReg != BaseReg) {
238 BuildMI(MBB, MBBI, dl, TII.get(ARM::tMOVr), DestReg)
239 .addReg(BaseReg, RegState::Kill)
240 .addImm((unsigned)Pred).addReg(PredReg).setMIFlags(MIFlags);
241 return;
244 bool isSub = NumBytes < 0;
245 if (isSub) NumBytes = -NumBytes;
247 // If profitable, use a movw or movt to materialize the offset.
248 // FIXME: Use the scavenger to grab a scratch register.
249 if (DestReg != ARM::SP && DestReg != BaseReg &&
250 NumBytes >= 4096 &&
251 ARM_AM::getT2SOImmVal(NumBytes) == -1) {
252 bool Fits = false;
253 if (NumBytes < 65536) {
254 // Use a movw to materialize the 16-bit constant.
255 BuildMI(MBB, MBBI, dl, TII.get(ARM::t2MOVi16), DestReg)
256 .addImm(NumBytes)
257 .addImm((unsigned)Pred).addReg(PredReg).setMIFlags(MIFlags);
258 Fits = true;
259 } else if ((NumBytes & 0xffff) == 0) {
260 // Use a movt to materialize the 32-bit constant.
261 BuildMI(MBB, MBBI, dl, TII.get(ARM::t2MOVTi16), DestReg)
262 .addReg(DestReg)
263 .addImm(NumBytes >> 16)
264 .addImm((unsigned)Pred).addReg(PredReg).setMIFlags(MIFlags);
265 Fits = true;
268 if (Fits) {
269 if (isSub) {
270 BuildMI(MBB, MBBI, dl, TII.get(ARM::t2SUBrr), DestReg)
271 .addReg(BaseReg)
272 .addReg(DestReg, RegState::Kill)
273 .add(predOps(Pred, PredReg))
274 .add(condCodeOp())
275 .setMIFlags(MIFlags);
276 } else {
277 // Here we know that DestReg is not SP but we do not
278 // know anything about BaseReg. t2ADDrr is an invalid
279 // instruction is SP is used as the second argument, but
280 // is fine if SP is the first argument. To be sure we
281 // do not generate invalid encoding, put BaseReg first.
282 BuildMI(MBB, MBBI, dl, TII.get(ARM::t2ADDrr), DestReg)
283 .addReg(BaseReg)
284 .addReg(DestReg, RegState::Kill)
285 .add(predOps(Pred, PredReg))
286 .add(condCodeOp())
287 .setMIFlags(MIFlags);
289 return;
293 while (NumBytes) {
294 unsigned ThisVal = NumBytes;
295 unsigned Opc = 0;
296 if (DestReg == ARM::SP && BaseReg != ARM::SP) {
297 // mov sp, rn. Note t2MOVr cannot be used.
298 BuildMI(MBB, MBBI, dl, TII.get(ARM::tMOVr), DestReg)
299 .addReg(BaseReg)
300 .setMIFlags(MIFlags)
301 .add(predOps(ARMCC::AL));
302 BaseReg = ARM::SP;
303 continue;
306 bool HasCCOut = true;
307 if (BaseReg == ARM::SP) {
308 // sub sp, sp, #imm7
309 if (DestReg == ARM::SP && (ThisVal < ((1 << 7)-1) * 4)) {
310 assert((ThisVal & 3) == 0 && "Stack update is not multiple of 4?");
311 Opc = isSub ? ARM::tSUBspi : ARM::tADDspi;
312 BuildMI(MBB, MBBI, dl, TII.get(Opc), DestReg)
313 .addReg(BaseReg)
314 .addImm(ThisVal / 4)
315 .setMIFlags(MIFlags)
316 .add(predOps(ARMCC::AL));
317 NumBytes = 0;
318 continue;
321 // sub rd, sp, so_imm
322 Opc = isSub ? ARM::t2SUBri : ARM::t2ADDri;
323 if (ARM_AM::getT2SOImmVal(NumBytes) != -1) {
324 NumBytes = 0;
325 } else {
326 // FIXME: Move this to ARMAddressingModes.h?
327 unsigned RotAmt = countLeadingZeros(ThisVal);
328 ThisVal = ThisVal & ARM_AM::rotr32(0xff000000U, RotAmt);
329 NumBytes &= ~ThisVal;
330 assert(ARM_AM::getT2SOImmVal(ThisVal) != -1 &&
331 "Bit extraction didn't work?");
333 } else {
334 assert(DestReg != ARM::SP && BaseReg != ARM::SP);
335 Opc = isSub ? ARM::t2SUBri : ARM::t2ADDri;
336 if (ARM_AM::getT2SOImmVal(NumBytes) != -1) {
337 NumBytes = 0;
338 } else if (ThisVal < 4096) {
339 Opc = isSub ? ARM::t2SUBri12 : ARM::t2ADDri12;
340 HasCCOut = false;
341 NumBytes = 0;
342 } else {
343 // FIXME: Move this to ARMAddressingModes.h?
344 unsigned RotAmt = countLeadingZeros(ThisVal);
345 ThisVal = ThisVal & ARM_AM::rotr32(0xff000000U, RotAmt);
346 NumBytes &= ~ThisVal;
347 assert(ARM_AM::getT2SOImmVal(ThisVal) != -1 &&
348 "Bit extraction didn't work?");
352 // Build the new ADD / SUB.
353 MachineInstrBuilder MIB = BuildMI(MBB, MBBI, dl, TII.get(Opc), DestReg)
354 .addReg(BaseReg, RegState::Kill)
355 .addImm(ThisVal)
356 .add(predOps(ARMCC::AL))
357 .setMIFlags(MIFlags);
358 if (HasCCOut)
359 MIB.add(condCodeOp());
361 BaseReg = DestReg;
365 static unsigned
366 negativeOffsetOpcode(unsigned opcode)
368 switch (opcode) {
369 case ARM::t2LDRi12: return ARM::t2LDRi8;
370 case ARM::t2LDRHi12: return ARM::t2LDRHi8;
371 case ARM::t2LDRBi12: return ARM::t2LDRBi8;
372 case ARM::t2LDRSHi12: return ARM::t2LDRSHi8;
373 case ARM::t2LDRSBi12: return ARM::t2LDRSBi8;
374 case ARM::t2STRi12: return ARM::t2STRi8;
375 case ARM::t2STRBi12: return ARM::t2STRBi8;
376 case ARM::t2STRHi12: return ARM::t2STRHi8;
377 case ARM::t2PLDi12: return ARM::t2PLDi8;
379 case ARM::t2LDRi8:
380 case ARM::t2LDRHi8:
381 case ARM::t2LDRBi8:
382 case ARM::t2LDRSHi8:
383 case ARM::t2LDRSBi8:
384 case ARM::t2STRi8:
385 case ARM::t2STRBi8:
386 case ARM::t2STRHi8:
387 case ARM::t2PLDi8:
388 return opcode;
390 default:
391 break;
394 return 0;
397 static unsigned
398 positiveOffsetOpcode(unsigned opcode)
400 switch (opcode) {
401 case ARM::t2LDRi8: return ARM::t2LDRi12;
402 case ARM::t2LDRHi8: return ARM::t2LDRHi12;
403 case ARM::t2LDRBi8: return ARM::t2LDRBi12;
404 case ARM::t2LDRSHi8: return ARM::t2LDRSHi12;
405 case ARM::t2LDRSBi8: return ARM::t2LDRSBi12;
406 case ARM::t2STRi8: return ARM::t2STRi12;
407 case ARM::t2STRBi8: return ARM::t2STRBi12;
408 case ARM::t2STRHi8: return ARM::t2STRHi12;
409 case ARM::t2PLDi8: return ARM::t2PLDi12;
411 case ARM::t2LDRi12:
412 case ARM::t2LDRHi12:
413 case ARM::t2LDRBi12:
414 case ARM::t2LDRSHi12:
415 case ARM::t2LDRSBi12:
416 case ARM::t2STRi12:
417 case ARM::t2STRBi12:
418 case ARM::t2STRHi12:
419 case ARM::t2PLDi12:
420 return opcode;
422 default:
423 break;
426 return 0;
429 static unsigned
430 immediateOffsetOpcode(unsigned opcode)
432 switch (opcode) {
433 case ARM::t2LDRs: return ARM::t2LDRi12;
434 case ARM::t2LDRHs: return ARM::t2LDRHi12;
435 case ARM::t2LDRBs: return ARM::t2LDRBi12;
436 case ARM::t2LDRSHs: return ARM::t2LDRSHi12;
437 case ARM::t2LDRSBs: return ARM::t2LDRSBi12;
438 case ARM::t2STRs: return ARM::t2STRi12;
439 case ARM::t2STRBs: return ARM::t2STRBi12;
440 case ARM::t2STRHs: return ARM::t2STRHi12;
441 case ARM::t2PLDs: return ARM::t2PLDi12;
443 case ARM::t2LDRi12:
444 case ARM::t2LDRHi12:
445 case ARM::t2LDRBi12:
446 case ARM::t2LDRSHi12:
447 case ARM::t2LDRSBi12:
448 case ARM::t2STRi12:
449 case ARM::t2STRBi12:
450 case ARM::t2STRHi12:
451 case ARM::t2PLDi12:
452 case ARM::t2LDRi8:
453 case ARM::t2LDRHi8:
454 case ARM::t2LDRBi8:
455 case ARM::t2LDRSHi8:
456 case ARM::t2LDRSBi8:
457 case ARM::t2STRi8:
458 case ARM::t2STRBi8:
459 case ARM::t2STRHi8:
460 case ARM::t2PLDi8:
461 return opcode;
463 default:
464 break;
467 return 0;
470 bool llvm::rewriteT2FrameIndex(MachineInstr &MI, unsigned FrameRegIdx,
471 unsigned FrameReg, int &Offset,
472 const ARMBaseInstrInfo &TII,
473 const TargetRegisterInfo *TRI) {
474 unsigned Opcode = MI.getOpcode();
475 const MCInstrDesc &Desc = MI.getDesc();
476 unsigned AddrMode = (Desc.TSFlags & ARMII::AddrModeMask);
477 bool isSub = false;
479 MachineFunction &MF = *MI.getParent()->getParent();
480 const TargetRegisterClass *RegClass =
481 TII.getRegClass(Desc, FrameRegIdx, TRI, MF);
483 // Memory operands in inline assembly always use AddrModeT2_i12.
484 if (Opcode == ARM::INLINEASM || Opcode == ARM::INLINEASM_BR)
485 AddrMode = ARMII::AddrModeT2_i12; // FIXME. mode for thumb2?
487 if (Opcode == ARM::t2ADDri || Opcode == ARM::t2ADDri12) {
488 Offset += MI.getOperand(FrameRegIdx+1).getImm();
490 unsigned PredReg;
491 if (Offset == 0 && getInstrPredicate(MI, PredReg) == ARMCC::AL &&
492 !MI.definesRegister(ARM::CPSR)) {
493 // Turn it into a move.
494 MI.setDesc(TII.get(ARM::tMOVr));
495 MI.getOperand(FrameRegIdx).ChangeToRegister(FrameReg, false);
496 // Remove offset and remaining explicit predicate operands.
497 do MI.RemoveOperand(FrameRegIdx+1);
498 while (MI.getNumOperands() > FrameRegIdx+1);
499 MachineInstrBuilder MIB(*MI.getParent()->getParent(), &MI);
500 MIB.add(predOps(ARMCC::AL));
501 return true;
504 bool HasCCOut = Opcode != ARM::t2ADDri12;
506 if (Offset < 0) {
507 Offset = -Offset;
508 isSub = true;
509 MI.setDesc(TII.get(ARM::t2SUBri));
510 } else {
511 MI.setDesc(TII.get(ARM::t2ADDri));
514 // Common case: small offset, fits into instruction.
515 if (ARM_AM::getT2SOImmVal(Offset) != -1) {
516 MI.getOperand(FrameRegIdx).ChangeToRegister(FrameReg, false);
517 MI.getOperand(FrameRegIdx+1).ChangeToImmediate(Offset);
518 // Add cc_out operand if the original instruction did not have one.
519 if (!HasCCOut)
520 MI.addOperand(MachineOperand::CreateReg(0, false));
521 Offset = 0;
522 return true;
524 // Another common case: imm12.
525 if (Offset < 4096 &&
526 (!HasCCOut || MI.getOperand(MI.getNumOperands()-1).getReg() == 0)) {
527 unsigned NewOpc = isSub ? ARM::t2SUBri12 : ARM::t2ADDri12;
528 MI.setDesc(TII.get(NewOpc));
529 MI.getOperand(FrameRegIdx).ChangeToRegister(FrameReg, false);
530 MI.getOperand(FrameRegIdx+1).ChangeToImmediate(Offset);
531 // Remove the cc_out operand.
532 if (HasCCOut)
533 MI.RemoveOperand(MI.getNumOperands()-1);
534 Offset = 0;
535 return true;
538 // Otherwise, extract 8 adjacent bits from the immediate into this
539 // t2ADDri/t2SUBri.
540 unsigned RotAmt = countLeadingZeros<unsigned>(Offset);
541 unsigned ThisImmVal = Offset & ARM_AM::rotr32(0xff000000U, RotAmt);
543 // We will handle these bits from offset, clear them.
544 Offset &= ~ThisImmVal;
546 assert(ARM_AM::getT2SOImmVal(ThisImmVal) != -1 &&
547 "Bit extraction didn't work?");
548 MI.getOperand(FrameRegIdx+1).ChangeToImmediate(ThisImmVal);
549 // Add cc_out operand if the original instruction did not have one.
550 if (!HasCCOut)
551 MI.addOperand(MachineOperand::CreateReg(0, false));
552 } else {
553 // AddrMode4 and AddrMode6 cannot handle any offset.
554 if (AddrMode == ARMII::AddrMode4 || AddrMode == ARMII::AddrMode6)
555 return false;
557 // AddrModeT2_so cannot handle any offset. If there is no offset
558 // register then we change to an immediate version.
559 unsigned NewOpc = Opcode;
560 if (AddrMode == ARMII::AddrModeT2_so) {
561 Register OffsetReg = MI.getOperand(FrameRegIdx + 1).getReg();
562 if (OffsetReg != 0) {
563 MI.getOperand(FrameRegIdx).ChangeToRegister(FrameReg, false);
564 return Offset == 0;
567 MI.RemoveOperand(FrameRegIdx+1);
568 MI.getOperand(FrameRegIdx+1).ChangeToImmediate(0);
569 NewOpc = immediateOffsetOpcode(Opcode);
570 AddrMode = ARMII::AddrModeT2_i12;
573 unsigned NumBits = 0;
574 unsigned Scale = 1;
575 if (AddrMode == ARMII::AddrModeT2_i8 || AddrMode == ARMII::AddrModeT2_i12) {
576 // i8 supports only negative, and i12 supports only positive, so
577 // based on Offset sign convert Opcode to the appropriate
578 // instruction
579 Offset += MI.getOperand(FrameRegIdx+1).getImm();
580 if (Offset < 0) {
581 NewOpc = negativeOffsetOpcode(Opcode);
582 NumBits = 8;
583 isSub = true;
584 Offset = -Offset;
585 } else {
586 NewOpc = positiveOffsetOpcode(Opcode);
587 NumBits = 12;
589 } else if (AddrMode == ARMII::AddrMode5) {
590 // VFP address mode.
591 const MachineOperand &OffOp = MI.getOperand(FrameRegIdx+1);
592 int InstrOffs = ARM_AM::getAM5Offset(OffOp.getImm());
593 if (ARM_AM::getAM5Op(OffOp.getImm()) == ARM_AM::sub)
594 InstrOffs *= -1;
595 NumBits = 8;
596 Scale = 4;
597 Offset += InstrOffs * 4;
598 assert((Offset & (Scale-1)) == 0 && "Can't encode this offset!");
599 if (Offset < 0) {
600 Offset = -Offset;
601 isSub = true;
603 } else if (AddrMode == ARMII::AddrMode5FP16) {
604 // VFP address mode.
605 const MachineOperand &OffOp = MI.getOperand(FrameRegIdx+1);
606 int InstrOffs = ARM_AM::getAM5FP16Offset(OffOp.getImm());
607 if (ARM_AM::getAM5FP16Op(OffOp.getImm()) == ARM_AM::sub)
608 InstrOffs *= -1;
609 NumBits = 8;
610 Scale = 2;
611 Offset += InstrOffs * 2;
612 assert((Offset & (Scale-1)) == 0 && "Can't encode this offset!");
613 if (Offset < 0) {
614 Offset = -Offset;
615 isSub = true;
617 } else if (AddrMode == ARMII::AddrModeT2_i7s4 ||
618 AddrMode == ARMII::AddrModeT2_i7s2 ||
619 AddrMode == ARMII::AddrModeT2_i7) {
620 Offset += MI.getOperand(FrameRegIdx + 1).getImm();
621 unsigned OffsetMask;
622 switch (AddrMode) {
623 case ARMII::AddrModeT2_i7s4: NumBits = 9; OffsetMask = 0x3; break;
624 case ARMII::AddrModeT2_i7s2: NumBits = 8; OffsetMask = 0x1; break;
625 default: NumBits = 7; OffsetMask = 0x0; break;
627 // MCInst operand expects already scaled value.
628 Scale = 1;
629 assert((Offset & OffsetMask) == 0 && "Can't encode this offset!");
630 (void)OffsetMask; // squash unused-variable warning at -NDEBUG
631 } else if (AddrMode == ARMII::AddrModeT2_i8s4) {
632 Offset += MI.getOperand(FrameRegIdx + 1).getImm() * 4;
633 NumBits = 8 + 2;
634 // MCInst operand expects already scaled value.
635 Scale = 1;
636 assert((Offset & 3) == 0 && "Can't encode this offset!");
637 } else if (AddrMode == ARMII::AddrModeT2_ldrex) {
638 Offset += MI.getOperand(FrameRegIdx + 1).getImm() * 4;
639 NumBits = 8; // 8 bits scaled by 4
640 Scale = 4;
641 assert((Offset & 3) == 0 && "Can't encode this offset!");
642 } else {
643 llvm_unreachable("Unsupported addressing mode!");
646 if (NewOpc != Opcode)
647 MI.setDesc(TII.get(NewOpc));
649 MachineOperand &ImmOp = MI.getOperand(FrameRegIdx+1);
651 // Attempt to fold address computation
652 // Common case: small offset, fits into instruction. We need to make sure
653 // the register class is correct too, for instructions like the MVE
654 // VLDRH.32, which only accepts low tGPR registers.
655 int ImmedOffset = Offset / Scale;
656 unsigned Mask = (1 << NumBits) - 1;
657 if ((unsigned)Offset <= Mask * Scale &&
658 (Register::isVirtualRegister(FrameReg) ||
659 RegClass->contains(FrameReg))) {
660 if (Register::isVirtualRegister(FrameReg)) {
661 // Make sure the register class for the virtual register is correct
662 MachineRegisterInfo *MRI = &MF.getRegInfo();
663 if (!MRI->constrainRegClass(FrameReg, RegClass))
664 llvm_unreachable("Unable to constrain virtual register class.");
667 // Replace the FrameIndex with fp/sp
668 MI.getOperand(FrameRegIdx).ChangeToRegister(FrameReg, false);
669 if (isSub) {
670 if (AddrMode == ARMII::AddrMode5 || AddrMode == ARMII::AddrMode5FP16)
671 // FIXME: Not consistent.
672 ImmedOffset |= 1 << NumBits;
673 else
674 ImmedOffset = -ImmedOffset;
676 ImmOp.ChangeToImmediate(ImmedOffset);
677 Offset = 0;
678 return true;
681 // Otherwise, offset doesn't fit. Pull in what we can to simplify
682 ImmedOffset = ImmedOffset & Mask;
683 if (isSub) {
684 if (AddrMode == ARMII::AddrMode5 || AddrMode == ARMII::AddrMode5FP16)
685 // FIXME: Not consistent.
686 ImmedOffset |= 1 << NumBits;
687 else {
688 ImmedOffset = -ImmedOffset;
689 if (ImmedOffset == 0)
690 // Change the opcode back if the encoded offset is zero.
691 MI.setDesc(TII.get(positiveOffsetOpcode(NewOpc)));
694 ImmOp.ChangeToImmediate(ImmedOffset);
695 Offset &= ~(Mask*Scale);
698 Offset = (isSub) ? -Offset : Offset;
699 return Offset == 0 && (Register::isVirtualRegister(FrameReg) ||
700 RegClass->contains(FrameReg));
703 ARMCC::CondCodes llvm::getITInstrPredicate(const MachineInstr &MI,
704 unsigned &PredReg) {
705 unsigned Opc = MI.getOpcode();
706 if (Opc == ARM::tBcc || Opc == ARM::t2Bcc)
707 return ARMCC::AL;
708 return getInstrPredicate(MI, PredReg);
711 int llvm::findFirstVPTPredOperandIdx(const MachineInstr &MI) {
712 const MCInstrDesc &MCID = MI.getDesc();
714 if (!MCID.OpInfo)
715 return -1;
717 for (unsigned i = 0, e = MCID.getNumOperands(); i != e; ++i)
718 if (ARM::isVpred(MCID.OpInfo[i].OperandType))
719 return i;
721 return -1;
724 ARMVCC::VPTCodes llvm::getVPTInstrPredicate(const MachineInstr &MI,
725 unsigned &PredReg) {
726 int PIdx = findFirstVPTPredOperandIdx(MI);
727 if (PIdx == -1) {
728 PredReg = 0;
729 return ARMVCC::None;
732 PredReg = MI.getOperand(PIdx+1).getReg();
733 return (ARMVCC::VPTCodes)MI.getOperand(PIdx).getImm();