1 //===-- Thumb2SizeReduction.cpp - Thumb2 code size reduction pass -*- C++ -*-=//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 #define DEBUG_TYPE "t2-reduce-size"
12 #include "ARMAddressingModes.h"
13 #include "ARMBaseRegisterInfo.h"
14 #include "ARMBaseInstrInfo.h"
15 #include "Thumb2InstrInfo.h"
16 #include "llvm/CodeGen/MachineInstr.h"
17 #include "llvm/CodeGen/MachineInstrBuilder.h"
18 #include "llvm/CodeGen/MachineFunctionPass.h"
19 #include "llvm/Support/CommandLine.h"
20 #include "llvm/Support/Compiler.h"
21 #include "llvm/Support/Debug.h"
22 #include "llvm/Support/raw_ostream.h"
23 #include "llvm/ADT/DenseMap.h"
24 #include "llvm/ADT/Statistic.h"
27 STATISTIC(NumNarrows, "Number of 32-bit instrs reduced to 16-bit ones");
28 STATISTIC(Num2Addrs, "Number of 32-bit instrs reduced to 2addr 16-bit ones");
29 STATISTIC(NumLdSts, "Number of 32-bit load / store reduced to 16-bit ones");
31 static cl::opt<int> ReduceLimit("t2-reduce-limit",
32 cl::init(-1), cl::Hidden);
33 static cl::opt<int> ReduceLimit2Addr("t2-reduce-limit2",
34 cl::init(-1), cl::Hidden);
35 static cl::opt<int> ReduceLimitLdSt("t2-reduce-limit3",
36 cl::init(-1), cl::Hidden);
39 /// ReduceTable - A static table with information on mapping from wide
42 unsigned WideOpc; // Wide opcode
43 unsigned NarrowOpc1; // Narrow opcode to transform to
44 unsigned NarrowOpc2; // Narrow opcode when it's two-address
45 uint8_t Imm1Limit; // Limit of immediate field (bits)
46 uint8_t Imm2Limit; // Limit of immediate field when it's two-address
47 unsigned LowRegs1 : 1; // Only possible if low-registers are used
48 unsigned LowRegs2 : 1; // Only possible if low-registers are used (2addr)
49 unsigned PredCC1 : 2; // 0 - If predicated, cc is on and vice versa.
51 // 2 - Always set CPSR.
53 unsigned Special : 1; // Needs to be dealt with specially
56 static const ReduceEntry ReduceTable[] = {
57 // Wide, Narrow1, Narrow2, imm1,imm2, lo1, lo2, P/C, S
58 { ARM::t2ADCrr, 0, ARM::tADC, 0, 0, 0, 1, 0,0, 0 },
59 { ARM::t2ADDri, ARM::tADDi3, ARM::tADDi8, 3, 8, 1, 1, 0,0, 0 },
60 { ARM::t2ADDrr, ARM::tADDrr, ARM::tADDhirr, 0, 0, 1, 0, 0,1, 0 },
61 // Note: immediate scale is 4.
62 { ARM::t2ADDrSPi,ARM::tADDrSPi,0, 8, 0, 1, 0, 1,0, 0 },
63 { ARM::t2ADDSri,ARM::tADDi3, ARM::tADDi8, 3, 8, 1, 1, 2,2, 1 },
64 { ARM::t2ADDSrr,ARM::tADDrr, 0, 0, 0, 1, 0, 2,0, 1 },
65 { ARM::t2ANDrr, 0, ARM::tAND, 0, 0, 0, 1, 0,0, 0 },
66 { ARM::t2ASRri, ARM::tASRri, 0, 5, 0, 1, 0, 0,0, 0 },
67 { ARM::t2ASRrr, 0, ARM::tASRrr, 0, 0, 0, 1, 0,0, 0 },
68 { ARM::t2BICrr, 0, ARM::tBIC, 0, 0, 0, 1, 0,0, 0 },
69 { ARM::t2CMNrr, ARM::tCMN, 0, 0, 0, 1, 0, 2,0, 0 },
70 { ARM::t2CMPri, ARM::tCMPi8, 0, 8, 0, 1, 0, 2,0, 0 },
71 { ARM::t2CMPrr, ARM::tCMPhir, 0, 0, 0, 0, 0, 2,0, 0 },
72 { ARM::t2CMPzri,ARM::tCMPzi8, 0, 8, 0, 1, 0, 2,0, 0 },
73 { ARM::t2CMPzrr,ARM::tCMPzhir,0, 0, 0, 0, 0, 2,0, 0 },
74 { ARM::t2EORrr, 0, ARM::tEOR, 0, 0, 0, 1, 0,0, 0 },
75 // FIXME: adr.n immediate offset must be multiple of 4.
76 //{ ARM::t2LEApcrelJT,ARM::tLEApcrelJT, 0, 0, 0, 1, 0, 1,0, 0 },
77 { ARM::t2LSLri, ARM::tLSLri, 0, 5, 0, 1, 0, 0,0, 0 },
78 { ARM::t2LSLrr, 0, ARM::tLSLrr, 0, 0, 0, 1, 0,0, 0 },
79 { ARM::t2LSRri, ARM::tLSRri, 0, 5, 0, 1, 0, 0,0, 0 },
80 { ARM::t2LSRrr, 0, ARM::tLSRrr, 0, 0, 0, 1, 0,0, 0 },
81 { ARM::t2MOVi, ARM::tMOVi8, 0, 8, 0, 1, 0, 0,0, 0 },
82 { ARM::t2MOVi16,ARM::tMOVi8, 0, 8, 0, 1, 0, 0,0, 0 },
83 // FIXME: Do we need the 16-bit 'S' variant?
84 { ARM::t2MOVr,ARM::tMOVgpr2gpr,0, 0, 0, 0, 0, 1,0, 0 },
85 { ARM::t2MOVCCr,0, ARM::tMOVCCr, 0, 0, 0, 0, 0,1, 0 },
86 { ARM::t2MOVCCi,0, ARM::tMOVCCi, 0, 8, 0, 0, 0,1, 0 },
87 { ARM::t2MUL, 0, ARM::tMUL, 0, 0, 0, 1, 0,0, 0 },
88 { ARM::t2MVNr, ARM::tMVN, 0, 0, 0, 1, 0, 0,0, 0 },
89 { ARM::t2ORRrr, 0, ARM::tORR, 0, 0, 0, 1, 0,0, 0 },
90 { ARM::t2REV, ARM::tREV, 0, 0, 0, 1, 0, 1,0, 0 },
91 { ARM::t2REV16, ARM::tREV16, 0, 0, 0, 1, 0, 1,0, 0 },
92 { ARM::t2REVSH, ARM::tREVSH, 0, 0, 0, 1, 0, 1,0, 0 },
93 { ARM::t2RORrr, 0, ARM::tROR, 0, 0, 0, 1, 0,0, 0 },
94 { ARM::t2RSBri, ARM::tRSB, 0, 0, 0, 1, 0, 0,0, 1 },
95 { ARM::t2RSBSri,ARM::tRSB, 0, 0, 0, 1, 0, 2,0, 1 },
96 { ARM::t2SBCrr, 0, ARM::tSBC, 0, 0, 0, 1, 0,0, 0 },
97 { ARM::t2SUBri, ARM::tSUBi3, ARM::tSUBi8, 3, 8, 1, 1, 0,0, 0 },
98 { ARM::t2SUBrr, ARM::tSUBrr, 0, 0, 0, 1, 0, 0,0, 0 },
99 { ARM::t2SUBSri,ARM::tSUBi3, ARM::tSUBi8, 3, 8, 1, 1, 2,2, 0 },
100 { ARM::t2SUBSrr,ARM::tSUBrr, 0, 0, 0, 1, 0, 2,0, 0 },
101 { ARM::t2SXTBr, ARM::tSXTB, 0, 0, 0, 1, 0, 1,0, 0 },
102 { ARM::t2SXTHr, ARM::tSXTH, 0, 0, 0, 1, 0, 1,0, 0 },
103 { ARM::t2TSTrr, ARM::tTST, 0, 0, 0, 1, 0, 2,0, 0 },
104 { ARM::t2UXTBr, ARM::tUXTB, 0, 0, 0, 1, 0, 1,0, 0 },
105 { ARM::t2UXTHr, ARM::tUXTH, 0, 0, 0, 1, 0, 1,0, 0 },
107 // FIXME: Clean this up after splitting each Thumb load / store opcode
108 // into multiple ones.
109 { ARM::t2LDRi12,ARM::tLDR, 0, 5, 0, 1, 0, 0,0, 1 },
110 { ARM::t2LDRs, ARM::tLDR, 0, 0, 0, 1, 0, 0,0, 1 },
111 { ARM::t2LDRBi12,ARM::tLDRB, 0, 5, 0, 1, 0, 0,0, 1 },
112 { ARM::t2LDRBs, ARM::tLDRB, 0, 0, 0, 1, 0, 0,0, 1 },
113 { ARM::t2LDRHi12,ARM::tLDRH, 0, 5, 0, 1, 0, 0,0, 1 },
114 { ARM::t2LDRHs, ARM::tLDRH, 0, 0, 0, 1, 0, 0,0, 1 },
115 { ARM::t2LDRSBs,ARM::tLDRSB, 0, 0, 0, 1, 0, 0,0, 1 },
116 { ARM::t2LDRSHs,ARM::tLDRSH, 0, 0, 0, 1, 0, 0,0, 1 },
117 { ARM::t2STRi12,ARM::tSTR, 0, 5, 0, 1, 0, 0,0, 1 },
118 { ARM::t2STRs, ARM::tSTR, 0, 0, 0, 1, 0, 0,0, 1 },
119 { ARM::t2STRBi12,ARM::tSTRB, 0, 5, 0, 1, 0, 0,0, 1 },
120 { ARM::t2STRBs, ARM::tSTRB, 0, 0, 0, 1, 0, 0,0, 1 },
121 { ARM::t2STRHi12,ARM::tSTRH, 0, 5, 0, 1, 0, 0,0, 1 },
122 { ARM::t2STRHs, ARM::tSTRH, 0, 0, 0, 1, 0, 0,0, 1 },
124 { ARM::t2LDM_RET,0, ARM::tPOP_RET, 0, 0, 1, 1, 1,1, 1 },
125 { ARM::t2LDM, ARM::tLDM, ARM::tPOP, 0, 0, 1, 1, 1,1, 1 },
126 { ARM::t2STM, ARM::tSTM, ARM::tPUSH, 0, 0, 1, 1, 1,1, 1 },
129 class VISIBILITY_HIDDEN Thumb2SizeReduce : public MachineFunctionPass {
134 const Thumb2InstrInfo *TII;
136 virtual bool runOnMachineFunction(MachineFunction &MF);
138 virtual const char *getPassName() const {
139 return "Thumb2 instruction size reduction pass";
143 /// ReduceOpcodeMap - Maps wide opcode to index of entry in ReduceTable.
144 DenseMap<unsigned, unsigned> ReduceOpcodeMap;
146 bool VerifyPredAndCC(MachineInstr *MI, const ReduceEntry &Entry,
147 bool is2Addr, ARMCC::CondCodes Pred,
148 bool LiveCPSR, bool &HasCC, bool &CCDead);
150 bool ReduceLoadStore(MachineBasicBlock &MBB, MachineInstr *MI,
151 const ReduceEntry &Entry);
153 bool ReduceSpecial(MachineBasicBlock &MBB, MachineInstr *MI,
154 const ReduceEntry &Entry, bool LiveCPSR);
156 /// ReduceTo2Addr - Reduce a 32-bit instruction to a 16-bit two-address
158 bool ReduceTo2Addr(MachineBasicBlock &MBB, MachineInstr *MI,
159 const ReduceEntry &Entry,
162 /// ReduceToNarrow - Reduce a 32-bit instruction to a 16-bit
163 /// non-two-address instruction.
164 bool ReduceToNarrow(MachineBasicBlock &MBB, MachineInstr *MI,
165 const ReduceEntry &Entry,
168 /// ReduceMBB - Reduce width of instructions in the specified basic block.
169 bool ReduceMBB(MachineBasicBlock &MBB);
171 char Thumb2SizeReduce::ID = 0;
174 Thumb2SizeReduce::Thumb2SizeReduce() : MachineFunctionPass(&ID) {
175 for (unsigned i = 0, e = array_lengthof(ReduceTable); i != e; ++i) {
176 unsigned FromOpc = ReduceTable[i].WideOpc;
177 if (!ReduceOpcodeMap.insert(std::make_pair(FromOpc, i)).second)
178 assert(false && "Duplicated entries?");
182 static bool HasImplicitCPSRDef(const TargetInstrDesc &TID) {
183 for (const unsigned *Regs = TID.ImplicitDefs; *Regs; ++Regs)
184 if (*Regs == ARM::CPSR)
190 Thumb2SizeReduce::VerifyPredAndCC(MachineInstr *MI, const ReduceEntry &Entry,
191 bool is2Addr, ARMCC::CondCodes Pred,
192 bool LiveCPSR, bool &HasCC, bool &CCDead) {
193 if ((is2Addr && Entry.PredCC2 == 0) ||
194 (!is2Addr && Entry.PredCC1 == 0)) {
195 if (Pred == ARMCC::AL) {
196 // Not predicated, must set CPSR.
198 // Original instruction was not setting CPSR, but CPSR is not
199 // currently live anyway. It's ok to set it. The CPSR def is
209 // Predicated, must not set CPSR.
213 } else if ((is2Addr && Entry.PredCC2 == 2) ||
214 (!is2Addr && Entry.PredCC1 == 2)) {
215 /// Old opcode has an optional def of CPSR.
218 // If both old opcode does not implicit CPSR def, then it's not ok since
219 // these new opcodes CPSR def is not meant to be thrown away. e.g. CMP.
220 if (!HasImplicitCPSRDef(MI->getDesc()))
224 // 16-bit instruction does not set CPSR.
232 static bool VerifyLowRegs(MachineInstr *MI) {
233 unsigned Opc = MI->getOpcode();
234 bool isPCOk = (Opc == ARM::t2LDM_RET) || (Opc == ARM::t2LDM);
235 bool isLROk = (Opc == ARM::t2STM);
236 bool isSPOk = isPCOk || isLROk || (Opc == ARM::t2ADDrSPi);
237 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
238 const MachineOperand &MO = MI->getOperand(i);
239 if (!MO.isReg() || MO.isImplicit())
241 unsigned Reg = MO.getReg();
242 if (Reg == 0 || Reg == ARM::CPSR)
244 if (isPCOk && Reg == ARM::PC)
246 if (isLROk && Reg == ARM::LR)
248 if (isSPOk && Reg == ARM::SP)
250 if (!isARMLowRegister(Reg))
257 Thumb2SizeReduce::ReduceLoadStore(MachineBasicBlock &MBB, MachineInstr *MI,
258 const ReduceEntry &Entry) {
259 if (ReduceLimitLdSt != -1 && ((int)NumLdSts >= ReduceLimitLdSt))
263 bool HasImmOffset = false;
264 bool HasShift = false;
265 bool isLdStMul = false;
266 unsigned Opc = Entry.NarrowOpc1;
267 unsigned OpNum = 3; // First 'rest' of operands.
268 switch (Entry.WideOpc) {
270 llvm_unreachable("Unexpected Thumb2 load / store opcode!");
300 unsigned BaseReg = MI->getOperand(0).getReg();
301 unsigned Mode = MI->getOperand(1).getImm();
302 if (BaseReg == ARM::SP && ARM_AM::getAM4WBFlag(Mode)) {
303 Opc = Entry.NarrowOpc2;
305 } else if (Entry.WideOpc == ARM::t2LDM_RET ||
306 !isARMLowRegister(BaseReg) ||
307 !ARM_AM::getAM4WBFlag(Mode) ||
308 ARM_AM::getAM4SubMode(Mode) != ARM_AM::ia) {
316 unsigned OffsetReg = 0;
317 bool OffsetKill = false;
319 OffsetReg = MI->getOperand(2).getReg();
320 OffsetKill = MI->getOperand(2).isKill();
321 if (MI->getOperand(3).getImm())
322 // Thumb1 addressing mode doesn't support shift.
326 unsigned OffsetImm = 0;
328 OffsetImm = MI->getOperand(2).getImm();
329 unsigned MaxOffset = ((1 << Entry.Imm1Limit) - 1) * Scale;
330 if ((OffsetImm & (Scale-1)) || OffsetImm > MaxOffset)
331 // Make sure the immediate field fits.
335 // Add the 16-bit load / store instruction.
336 // FIXME: Thumb1 addressing mode encode both immediate and register offset.
337 DebugLoc dl = MI->getDebugLoc();
338 MachineInstrBuilder MIB = BuildMI(MBB, *MI, dl, TII->get(Opc));
340 MIB.addOperand(MI->getOperand(0)).addOperand(MI->getOperand(1));
341 if (Entry.NarrowOpc1 != ARM::tLDRSB && Entry.NarrowOpc1 != ARM::tLDRSH) {
342 // tLDRSB and tLDRSH do not have an immediate offset field. On the other
343 // hand, it must have an offset register.
344 // FIXME: Remove this special case.
345 MIB.addImm(OffsetImm/Scale);
347 assert((!HasShift || OffsetReg) && "Invalid so_reg load / store address!");
349 MIB.addReg(OffsetReg, getKillRegState(OffsetKill));
352 // Transfer the rest of operands.
353 for (unsigned e = MI->getNumOperands(); OpNum != e; ++OpNum)
354 MIB.addOperand(MI->getOperand(OpNum));
356 DEBUG(errs() << "Converted 32-bit: " << *MI << " to 16-bit: " << *MIB);
364 Thumb2SizeReduce::ReduceSpecial(MachineBasicBlock &MBB, MachineInstr *MI,
365 const ReduceEntry &Entry,
367 if (Entry.LowRegs1 && !VerifyLowRegs(MI))
370 const TargetInstrDesc &TID = MI->getDesc();
371 if (TID.mayLoad() || TID.mayStore())
372 return ReduceLoadStore(MBB, MI, Entry);
374 unsigned Opc = MI->getOpcode();
378 case ARM::t2ADDSrr: {
379 unsigned PredReg = 0;
380 if (getInstrPredicate(MI, PredReg) == ARMCC::AL) {
383 case ARM::t2ADDSri: {
384 if (ReduceTo2Addr(MBB, MI, Entry, LiveCPSR))
389 return ReduceToNarrow(MBB, MI, Entry, LiveCPSR);
396 if (MI->getOperand(2).getImm() == 0)
397 return ReduceToNarrow(MBB, MI, Entry, LiveCPSR);
404 Thumb2SizeReduce::ReduceTo2Addr(MachineBasicBlock &MBB, MachineInstr *MI,
405 const ReduceEntry &Entry,
408 if (ReduceLimit2Addr != -1 && ((int)Num2Addrs >= ReduceLimit2Addr))
411 const TargetInstrDesc &TID = MI->getDesc();
412 unsigned Reg0 = MI->getOperand(0).getReg();
413 unsigned Reg1 = MI->getOperand(1).getReg();
416 if (Entry.LowRegs2 && !isARMLowRegister(Reg0))
418 if (Entry.Imm2Limit) {
419 unsigned Imm = MI->getOperand(2).getImm();
420 unsigned Limit = (1 << Entry.Imm2Limit) - 1;
424 unsigned Reg2 = MI->getOperand(2).getReg();
425 if (Entry.LowRegs2 && !isARMLowRegister(Reg2))
429 // Check if it's possible / necessary to transfer the predicate.
430 const TargetInstrDesc &NewTID = TII->get(Entry.NarrowOpc2);
431 unsigned PredReg = 0;
432 ARMCC::CondCodes Pred = getInstrPredicate(MI, PredReg);
433 bool SkipPred = false;
434 if (Pred != ARMCC::AL) {
435 if (!NewTID.isPredicable())
436 // Can't transfer predicate, fail.
439 SkipPred = !NewTID.isPredicable();
444 if (TID.hasOptionalDef()) {
445 unsigned NumOps = TID.getNumOperands();
446 HasCC = (MI->getOperand(NumOps-1).getReg() == ARM::CPSR);
447 if (HasCC && MI->getOperand(NumOps-1).isDead())
450 if (!VerifyPredAndCC(MI, Entry, true, Pred, LiveCPSR, HasCC, CCDead))
453 // Add the 16-bit instruction.
454 DebugLoc dl = MI->getDebugLoc();
455 MachineInstrBuilder MIB = BuildMI(MBB, *MI, dl, NewTID);
456 MIB.addOperand(MI->getOperand(0));
457 if (NewTID.hasOptionalDef()) {
459 AddDefaultT1CC(MIB, CCDead);
464 // Transfer the rest of operands.
465 unsigned NumOps = TID.getNumOperands();
466 for (unsigned i = 1, e = MI->getNumOperands(); i != e; ++i) {
467 if (i < NumOps && TID.OpInfo[i].isOptionalDef())
469 if (SkipPred && TID.OpInfo[i].isPredicate())
471 MIB.addOperand(MI->getOperand(i));
474 DEBUG(errs() << "Converted 32-bit: " << *MI << " to 16-bit: " << *MIB);
482 Thumb2SizeReduce::ReduceToNarrow(MachineBasicBlock &MBB, MachineInstr *MI,
483 const ReduceEntry &Entry,
485 if (ReduceLimit != -1 && ((int)NumNarrows >= ReduceLimit))
488 unsigned Limit = ~0U;
489 unsigned Scale = (Entry.WideOpc == ARM::t2ADDrSPi) ? 4 : 1;
491 Limit = ((1 << Entry.Imm1Limit) - 1) * Scale;
493 const TargetInstrDesc &TID = MI->getDesc();
494 for (unsigned i = 0, e = TID.getNumOperands(); i != e; ++i) {
495 if (TID.OpInfo[i].isPredicate())
497 const MachineOperand &MO = MI->getOperand(i);
499 unsigned Reg = MO.getReg();
500 if (!Reg || Reg == ARM::CPSR)
502 if (Entry.WideOpc == ARM::t2ADDrSPi && Reg == ARM::SP)
504 if (Entry.LowRegs1 && !isARMLowRegister(Reg))
506 } else if (MO.isImm() &&
507 !TID.OpInfo[i].isPredicate()) {
508 if (((unsigned)MO.getImm()) > Limit || (MO.getImm() & (Scale-1)) != 0)
513 // Check if it's possible / necessary to transfer the predicate.
514 const TargetInstrDesc &NewTID = TII->get(Entry.NarrowOpc1);
515 unsigned PredReg = 0;
516 ARMCC::CondCodes Pred = getInstrPredicate(MI, PredReg);
517 bool SkipPred = false;
518 if (Pred != ARMCC::AL) {
519 if (!NewTID.isPredicable())
520 // Can't transfer predicate, fail.
523 SkipPred = !NewTID.isPredicable();
528 if (TID.hasOptionalDef()) {
529 unsigned NumOps = TID.getNumOperands();
530 HasCC = (MI->getOperand(NumOps-1).getReg() == ARM::CPSR);
531 if (HasCC && MI->getOperand(NumOps-1).isDead())
534 if (!VerifyPredAndCC(MI, Entry, false, Pred, LiveCPSR, HasCC, CCDead))
537 // Add the 16-bit instruction.
538 DebugLoc dl = MI->getDebugLoc();
539 MachineInstrBuilder MIB = BuildMI(MBB, *MI, dl, NewTID);
540 MIB.addOperand(MI->getOperand(0));
541 if (NewTID.hasOptionalDef()) {
543 AddDefaultT1CC(MIB, CCDead);
548 // Transfer the rest of operands.
549 unsigned NumOps = TID.getNumOperands();
550 for (unsigned i = 1, e = MI->getNumOperands(); i != e; ++i) {
551 if (i < NumOps && TID.OpInfo[i].isOptionalDef())
553 if ((TID.getOpcode() == ARM::t2RSBSri ||
554 TID.getOpcode() == ARM::t2RSBri) && i == 2)
555 // Skip the zero immediate operand, it's now implicit.
557 bool isPred = (i < NumOps && TID.OpInfo[i].isPredicate());
558 if (SkipPred && isPred)
560 const MachineOperand &MO = MI->getOperand(i);
561 if (Scale > 1 && !isPred && MO.isImm())
562 MIB.addImm(MO.getImm() / Scale);
564 if (MO.isReg() && MO.isImplicit() && MO.getReg() == ARM::CPSR)
565 // Skip implicit def of CPSR. Either it's modeled as an optional
566 // def now or it's already an implicit def on the new instruction.
571 if (!TID.isPredicable() && NewTID.isPredicable())
574 DEBUG(errs() << "Converted 32-bit: " << *MI << " to 16-bit: " << *MIB);
581 static bool UpdateCPSRDef(MachineInstr &MI, bool LiveCPSR) {
583 for (unsigned i = 0, e = MI.getNumOperands(); i != e; ++i) {
584 const MachineOperand &MO = MI.getOperand(i);
585 if (!MO.isReg() || MO.isUndef() || MO.isUse())
587 if (MO.getReg() != ARM::CPSR)
593 return HasDef || LiveCPSR;
596 static bool UpdateCPSRUse(MachineInstr &MI, bool LiveCPSR) {
597 for (unsigned i = 0, e = MI.getNumOperands(); i != e; ++i) {
598 const MachineOperand &MO = MI.getOperand(i);
599 if (!MO.isReg() || MO.isUndef() || MO.isDef())
601 if (MO.getReg() != ARM::CPSR)
603 assert(LiveCPSR && "CPSR liveness tracking is wrong!");
613 bool Thumb2SizeReduce::ReduceMBB(MachineBasicBlock &MBB) {
614 bool Modified = false;
616 bool LiveCPSR = false;
617 // Yes, CPSR could be livein.
618 for (MachineBasicBlock::const_livein_iterator I = MBB.livein_begin(),
619 E = MBB.livein_end(); I != E; ++I) {
620 if (*I == ARM::CPSR) {
626 MachineBasicBlock::iterator MII = MBB.begin(), E = MBB.end();
627 MachineBasicBlock::iterator NextMII;
628 for (; MII != E; MII = NextMII) {
631 MachineInstr *MI = &*MII;
632 LiveCPSR = UpdateCPSRUse(*MI, LiveCPSR);
634 unsigned Opcode = MI->getOpcode();
635 DenseMap<unsigned, unsigned>::iterator OPI = ReduceOpcodeMap.find(Opcode);
636 if (OPI != ReduceOpcodeMap.end()) {
637 const ReduceEntry &Entry = ReduceTable[OPI->second];
638 // Ignore "special" cases for now.
640 if (ReduceSpecial(MBB, MI, Entry, LiveCPSR)) {
642 MachineBasicBlock::iterator I = prior(NextMII);
648 // Try to transform to a 16-bit two-address instruction.
649 if (Entry.NarrowOpc2 && ReduceTo2Addr(MBB, MI, Entry, LiveCPSR)) {
651 MachineBasicBlock::iterator I = prior(NextMII);
656 // Try to transform ro a 16-bit non-two-address instruction.
657 if (Entry.NarrowOpc1 && ReduceToNarrow(MBB, MI, Entry, LiveCPSR)) {
659 MachineBasicBlock::iterator I = prior(NextMII);
665 LiveCPSR = UpdateCPSRDef(*MI, LiveCPSR);
671 bool Thumb2SizeReduce::runOnMachineFunction(MachineFunction &MF) {
672 const TargetMachine &TM = MF.getTarget();
673 TII = static_cast<const Thumb2InstrInfo*>(TM.getInstrInfo());
675 bool Modified = false;
676 for (MachineFunction::iterator I = MF.begin(), E = MF.end(); I != E; ++I)
677 Modified |= ReduceMBB(*I);
681 /// createThumb2SizeReductionPass - Returns an instance of the Thumb2 size
683 FunctionPass *llvm::createThumb2SizeReductionPass() {
684 return new Thumb2SizeReduce();