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/ADT/DenseMap.h"
23 #include "llvm/ADT/Statistic.h"
26 STATISTIC(NumNarrows, "Number of 32-bit instrs reduced to 16-bit ones");
27 STATISTIC(Num2Addrs, "Number of 32-bit instrs reduced to 2addr 16-bit ones");
28 STATISTIC(NumLdSts, "Number of 32-bit load / store reduced to 16-bit ones");
30 static cl::opt<int> ReduceLimit("t2-reduce-limit",
31 cl::init(-1), cl::Hidden);
32 static cl::opt<int> ReduceLimit2Addr("t2-reduce-limit2",
33 cl::init(-1), cl::Hidden);
34 static cl::opt<int> ReduceLimitLdSt("t2-reduce-limit3",
35 cl::init(-1), cl::Hidden);
38 /// ReduceTable - A static table with information on mapping from wide
41 unsigned WideOpc; // Wide opcode
42 unsigned NarrowOpc1; // Narrow opcode to transform to
43 unsigned NarrowOpc2; // Narrow opcode when it's two-address
44 uint8_t Imm1Limit; // Limit of immediate field (bits)
45 uint8_t Imm2Limit; // Limit of immediate field when it's two-address
46 unsigned LowRegs1 : 1; // Only possible if low-registers are used
47 unsigned LowRegs2 : 1; // Only possible if low-registers are used (2addr)
48 unsigned PredCC1 : 1; // 0 - If predicated, cc is on and vice versa.
51 unsigned Special : 1; // Needs to be dealt with specially
54 static const ReduceEntry ReduceTable[] = {
55 // Wide, Narrow1, Narrow2, imm1,imm2, lo1, lo2, P/C, S
56 { ARM::t2ADCrr, ARM::tADC, 0, 0, 0, 1, 0, 0,0, 0 },
57 // FIXME: t2ADDS variants.
58 { ARM::t2ADDri, ARM::tADDi3, ARM::tADDi8, 3, 8, 1, 1, 0,0, 0 },
59 { ARM::t2ADDrr, ARM::tADDrr, ARM::tADDhirr, 0, 0, 1, 0, 0,1, 0 },
60 { ARM::t2ANDrr, 0, ARM::tAND, 0, 0, 0, 1, 0,0, 0 },
61 { ARM::t2ASRri, ARM::tASRri, 0, 5, 0, 1, 0, 0,0, 0 },
62 { ARM::t2ASRrr, 0, ARM::tASRrr, 0, 0, 0, 1, 0,0, 0 },
63 { ARM::t2BICrr, 0, ARM::tBIC, 0, 0, 0, 1, 0,0, 0 },
64 { ARM::t2CMNrr, ARM::tCMN, 0, 0, 0, 1, 0, 1,0, 0 },
65 { ARM::t2CMPri, ARM::tCMPi8, 0, 8, 0, 1, 0, 1,0, 0 },
66 { ARM::t2CMPrr, ARM::tCMPhir, 0, 0, 0, 0, 0, 1,0, 0 },
67 { ARM::t2CMPzri,ARM::tCMPzi8, 0, 8, 0, 1, 0, 1,0, 0 },
68 { ARM::t2CMPzrr,ARM::tCMPzhir,0, 0, 0, 0, 0, 1,0, 0 },
69 { ARM::t2EORrr, 0, ARM::tEOR, 0, 0, 0, 1, 0,0, 0 },
70 { ARM::t2LSLri, ARM::tLSLri, 0, 5, 0, 1, 0, 0,0, 0 },
71 { ARM::t2LSLrr, 0, ARM::tLSLrr, 0, 0, 0, 1, 0,0, 0 },
72 { ARM::t2LSRri, ARM::tLSRri, 0, 5, 0, 1, 0, 0,0, 0 },
73 { ARM::t2LSRrr, 0, ARM::tLSRrr, 0, 0, 0, 1, 0,0, 0 },
74 { ARM::t2MOVi, ARM::tMOVi8, 0, 8, 0, 1, 0, 0,0, 0 },
75 // FIXME: Do we need the 16-bit 'S' variant?
77 { ARM::t2MOVr,ARM::tMOVgpr2gpr,0, 0, 0, 0, 0, 1,0, 0 },
78 { ARM::t2MUL, 0, ARM::tMUL, 0, 0, 0, 1, 0,0, 0 },
79 { ARM::t2MVNr, ARM::tMVN, 0, 0, 0, 1, 0, 0,0, 0 },
80 { ARM::t2ORRrr, 0, ARM::tORR, 0, 0, 0, 1, 0,0, 0 },
81 { ARM::t2REV, ARM::tREV, 0, 0, 0, 1, 0, 1,0, 0 },
82 { ARM::t2REV16, ARM::tREV16, 0, 0, 0, 1, 0, 1,0, 0 },
83 { ARM::t2REVSH, ARM::tREVSH, 0, 0, 0, 1, 0, 1,0, 0 },
84 { ARM::t2RORrr, 0, ARM::tROR, 0, 0, 0, 1, 0,0, 0 },
85 // FIXME: T2RSBri immediate must be zero. Also need entry for T2RSBS
86 //{ ARM::t2RSBri, ARM::tRSB, 0, 0, 0, 1, 0, 0,0, 0 },
87 { ARM::t2SUBri, ARM::tSUBi3, ARM::tSUBi8, 3, 8, 1, 1, 0,0, 0 },
88 { ARM::t2SUBrr, ARM::tSUBrr, 0, 0, 0, 1, 0, 0,0, 0 },
89 { ARM::t2SXTBr, ARM::tSXTB, 0, 0, 0, 1, 0, 1,0, 0 },
90 { ARM::t2SXTHr, ARM::tSXTH, 0, 0, 0, 1, 0, 1,0, 0 },
91 { ARM::t2TSTrr, ARM::tTST, 0, 0, 0, 1, 0, 1,0, 0 },
92 { ARM::t2UXTBr, ARM::tUXTB, 0, 0, 0, 1, 0, 1,0, 0 },
93 { ARM::t2UXTHr, ARM::tUXTH, 0, 0, 0, 1, 0, 1,0, 0 },
95 // FIXME: Clean this up after splitting each Thumb load / store opcode
96 // into multiple ones.
97 { ARM::t2LDRi12,ARM::tLDR, 0, 5, 0, 1, 0, 0,0, 1 },
98 { ARM::t2LDRs, ARM::tLDR, 0, 0, 0, 1, 0, 0,0, 1 },
99 { ARM::t2LDRBi12,ARM::tLDRB, 0, 5, 0, 1, 0, 0,0, 1 },
100 { ARM::t2LDRBs, ARM::tLDRB, 0, 0, 0, 1, 0, 0,0, 1 },
101 { ARM::t2LDRHi12,ARM::tLDRH, 0, 5, 0, 1, 0, 0,0, 1 },
102 { ARM::t2LDRHs, ARM::tLDRH, 0, 0, 0, 1, 0, 0,0, 1 },
103 { ARM::t2LDRSBs,ARM::tLDRSB, 0, 0, 0, 1, 0, 0,0, 1 },
104 { ARM::t2LDRSHs,ARM::tLDRSH, 0, 0, 0, 1, 0, 0,0, 1 },
105 { ARM::t2STRi12,ARM::tSTR, 0, 5, 0, 1, 0, 0,0, 1 },
106 { ARM::t2STRs, ARM::tSTR, 0, 0, 0, 1, 0, 0,0, 1 },
107 { ARM::t2STRBi12,ARM::tSTRB, 0, 5, 0, 1, 0, 0,0, 1 },
108 { ARM::t2STRBs, ARM::tSTRB, 0, 0, 0, 1, 0, 0,0, 1 },
109 { ARM::t2STRHi12,ARM::tSTRH, 0, 5, 0, 1, 0, 0,0, 1 },
110 { ARM::t2STRHs, ARM::tSTRH, 0, 0, 0, 1, 0, 0,0, 1 },
112 { ARM::t2LDM_RET,0, ARM::tPOP_RET, 0, 0, 1, 1, 1,1, 1 },
113 { ARM::t2LDM, ARM::tLDM, ARM::tPOP, 0, 0, 1, 1, 1,1, 1 },
114 { ARM::t2STM, ARM::tSTM, ARM::tPUSH, 0, 0, 1, 1, 1,1, 1 },
117 class VISIBILITY_HIDDEN Thumb2SizeReduce : public MachineFunctionPass {
122 const TargetInstrInfo *TII;
124 virtual bool runOnMachineFunction(MachineFunction &MF);
126 virtual const char *getPassName() const {
127 return "Thumb2 instruction size reduction pass";
131 /// ReduceOpcodeMap - Maps wide opcode to index of entry in ReduceTable.
132 DenseMap<unsigned, unsigned> ReduceOpcodeMap;
134 bool ReduceLoadStore(MachineBasicBlock &MBB, MachineInstr *MI,
135 const ReduceEntry &Entry);
137 bool ReduceSpecial(MachineBasicBlock &MBB, MachineInstr *MI,
138 const ReduceEntry &Entry, bool LiveCPSR);
140 /// ReduceTo2Addr - Reduce a 32-bit instruction to a 16-bit two-address
142 bool ReduceTo2Addr(MachineBasicBlock &MBB, MachineInstr *MI,
143 const ReduceEntry &Entry,
146 /// ReduceToNarrow - Reduce a 32-bit instruction to a 16-bit
147 /// non-two-address instruction.
148 bool ReduceToNarrow(MachineBasicBlock &MBB, MachineInstr *MI,
149 const ReduceEntry &Entry,
152 /// ReduceMBB - Reduce width of instructions in the specified basic block.
153 bool ReduceMBB(MachineBasicBlock &MBB);
155 char Thumb2SizeReduce::ID = 0;
158 Thumb2SizeReduce::Thumb2SizeReduce() : MachineFunctionPass(&ID) {
159 for (unsigned i = 0, e = array_lengthof(ReduceTable); i != e; ++i) {
160 unsigned FromOpc = ReduceTable[i].WideOpc;
161 if (!ReduceOpcodeMap.insert(std::make_pair(FromOpc, i)).second)
162 assert(false && "Duplicated entries?");
166 static bool VerifyPredAndCC(MachineInstr *MI, const ReduceEntry &Entry,
167 bool is2Addr, ARMCC::CondCodes Pred,
168 bool LiveCPSR, bool &HasCC, bool &CCDead) {
169 if ((is2Addr && Entry.PredCC2 == 0) ||
170 (!is2Addr && Entry.PredCC1 == 0)) {
171 if (Pred == ARMCC::AL) {
172 // Not predicated, must set CPSR.
174 // Original instruction was not setting CPSR, but CPSR is not
175 // currently live anyway. It's ok to set it. The CPSR def is
185 // Predicated, must not set CPSR.
190 // 16-bit instruction does not set CPSR.
198 static bool VerifyLowRegs(MachineInstr *MI) {
199 unsigned Opc = MI->getOpcode();
200 bool isPCOk = (Opc == ARM::t2LDM_RET) || (Opc == ARM::t2LDM);
201 bool isLROk = (Opc == ARM::t2STM);
202 bool isSPOk = isPCOk || isLROk;
203 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
204 const MachineOperand &MO = MI->getOperand(i);
205 if (!MO.isReg() || MO.isImplicit())
207 unsigned Reg = MO.getReg();
208 if (Reg == 0 || Reg == ARM::CPSR)
210 if (isPCOk && Reg == ARM::PC)
212 if (isLROk && Reg == ARM::LR)
214 if (isSPOk && Reg == ARM::SP)
216 if (!isARMLowRegister(Reg))
223 Thumb2SizeReduce::ReduceLoadStore(MachineBasicBlock &MBB, MachineInstr *MI,
224 const ReduceEntry &Entry) {
225 if (ReduceLimitLdSt != -1 && ((int)NumLdSts >= ReduceLimitLdSt))
229 bool HasImmOffset = false;
230 bool HasShift = false;
231 bool isLdStMul = false;
232 bool isPopPush = false;
233 unsigned Opc = Entry.NarrowOpc1;
234 unsigned OpNum = 3; // First 'rest' of operands.
235 switch (Entry.WideOpc) {
237 llvm_unreachable("Unexpected Thumb2 load / store opcode!");
267 unsigned BaseReg = MI->getOperand(0).getReg();
268 unsigned Mode = MI->getOperand(1).getImm();
269 if (BaseReg == ARM::SP && ARM_AM::getAM4WBFlag(Mode)) {
270 Opc = Entry.NarrowOpc2;
273 } else if (Entry.WideOpc == ARM::t2LDM_RET ||
274 !isARMLowRegister(BaseReg) ||
275 !ARM_AM::getAM4WBFlag(Mode) ||
276 ARM_AM::getAM4SubMode(Mode) != ARM_AM::ia) {
284 unsigned OffsetReg = 0;
285 bool OffsetKill = false;
287 OffsetReg = MI->getOperand(2).getReg();
288 OffsetKill = MI->getOperand(2).isKill();
289 if (MI->getOperand(3).getImm())
290 // Thumb1 addressing mode doesn't support shift.
294 unsigned OffsetImm = 0;
296 OffsetImm = MI->getOperand(2).getImm();
297 unsigned MaxOffset = ((1 << Entry.Imm1Limit) - 1) * Scale;
298 if ((OffsetImm & (Scale-1)) || OffsetImm > MaxOffset)
299 // Make sure the immediate field fits.
303 // Add the 16-bit load / store instruction.
304 // FIXME: Thumb1 addressing mode encode both immediate and register offset.
305 DebugLoc dl = MI->getDebugLoc();
306 MachineInstrBuilder MIB = BuildMI(MBB, *MI, dl, TII->get(Opc));
308 MIB.addOperand(MI->getOperand(0)).addOperand(MI->getOperand(1));
309 if (Entry.NarrowOpc1 != ARM::tLDRSB && Entry.NarrowOpc1 != ARM::tLDRSH) {
310 // tLDRSB and tLDRSH do not have an immediate offset field. On the other
311 // hand, it must have an offset register.
312 // FIXME: Remove this special case.
313 MIB.addImm(OffsetImm/Scale);
315 assert((!HasShift || OffsetReg) && "Invalid so_reg load / store address!");
317 MIB.addReg(OffsetReg, getKillRegState(OffsetKill));
320 // Transfer the rest of operands.
321 for (unsigned e = MI->getNumOperands(); OpNum != e; ++OpNum)
322 MIB.addOperand(MI->getOperand(OpNum));
324 DOUT << "Converted 32-bit: " << *MI << " to 16-bit: " << *MIB;
332 Thumb2SizeReduce::ReduceSpecial(MachineBasicBlock &MBB, MachineInstr *MI,
333 const ReduceEntry &Entry,
335 if (Entry.LowRegs1 && !VerifyLowRegs(MI))
338 const TargetInstrDesc &TID = MI->getDesc();
339 if (TID.mayLoad() || TID.mayStore())
340 return ReduceLoadStore(MBB, MI, Entry);
345 Thumb2SizeReduce::ReduceTo2Addr(MachineBasicBlock &MBB, MachineInstr *MI,
346 const ReduceEntry &Entry,
349 if (ReduceLimit2Addr != -1 && ((int)Num2Addrs >= ReduceLimit2Addr))
352 const TargetInstrDesc &TID = MI->getDesc();
353 unsigned Reg0 = MI->getOperand(0).getReg();
354 unsigned Reg1 = MI->getOperand(1).getReg();
357 if (Entry.LowRegs2 && !isARMLowRegister(Reg0))
359 if (Entry.Imm2Limit) {
360 unsigned Imm = MI->getOperand(2).getImm();
361 unsigned Limit = (1 << Entry.Imm2Limit) - 1;
365 unsigned Reg2 = MI->getOperand(2).getReg();
366 if (Entry.LowRegs2 && !isARMLowRegister(Reg2))
370 // Check if it's possible / necessary to transfer the predicate.
371 const TargetInstrDesc &NewTID = TII->get(Entry.NarrowOpc2);
372 unsigned PredReg = 0;
373 ARMCC::CondCodes Pred = getInstrPredicate(MI, PredReg);
374 bool SkipPred = false;
375 if (Pred != ARMCC::AL) {
376 if (!NewTID.isPredicable())
377 // Can't transfer predicate, fail.
380 SkipPred = !NewTID.isPredicable();
385 if (TID.hasOptionalDef()) {
386 unsigned NumOps = TID.getNumOperands();
387 HasCC = (MI->getOperand(NumOps-1).getReg() == ARM::CPSR);
388 if (HasCC && MI->getOperand(NumOps-1).isDead())
391 if (!VerifyPredAndCC(MI, Entry, true, Pred, LiveCPSR, HasCC, CCDead))
394 // Add the 16-bit instruction.
395 DebugLoc dl = MI->getDebugLoc();
396 MachineInstrBuilder MIB = BuildMI(MBB, *MI, dl, TII->get(Entry.NarrowOpc2));
397 MIB.addOperand(MI->getOperand(0));
399 AddDefaultT1CC(MIB, CCDead);
401 // Transfer the rest of operands.
402 unsigned NumOps = TID.getNumOperands();
403 for (unsigned i = 1, e = MI->getNumOperands(); i != e; ++i) {
404 if (i < NumOps && TID.OpInfo[i].isOptionalDef())
406 if (SkipPred && TID.OpInfo[i].isPredicate())
408 MIB.addOperand(MI->getOperand(i));
411 DOUT << "Converted 32-bit: " << *MI << " to 16-bit: " << *MIB;
419 Thumb2SizeReduce::ReduceToNarrow(MachineBasicBlock &MBB, MachineInstr *MI,
420 const ReduceEntry &Entry,
422 if (ReduceLimit != -1 && ((int)NumNarrows >= ReduceLimit))
425 unsigned Limit = ~0U;
427 Limit = (1 << Entry.Imm1Limit) - 1;
429 const TargetInstrDesc &TID = MI->getDesc();
430 for (unsigned i = 0, e = TID.getNumOperands(); i != e; ++i) {
431 if (TID.OpInfo[i].isPredicate())
433 const MachineOperand &MO = MI->getOperand(i);
435 unsigned Reg = MO.getReg();
436 if (!Reg || Reg == ARM::CPSR)
438 if (Entry.LowRegs1 && !isARMLowRegister(Reg))
440 } else if (MO.isImm()) {
441 if (MO.getImm() > Limit)
446 // Check if it's possible / necessary to transfer the predicate.
447 const TargetInstrDesc &NewTID = TII->get(Entry.NarrowOpc1);
448 unsigned PredReg = 0;
449 ARMCC::CondCodes Pred = getInstrPredicate(MI, PredReg);
450 bool SkipPred = false;
451 if (Pred != ARMCC::AL) {
452 if (!NewTID.isPredicable())
453 // Can't transfer predicate, fail.
456 SkipPred = !NewTID.isPredicable();
461 if (TID.hasOptionalDef()) {
462 unsigned NumOps = TID.getNumOperands();
463 HasCC = (MI->getOperand(NumOps-1).getReg() == ARM::CPSR);
464 if (HasCC && MI->getOperand(NumOps-1).isDead())
467 if (!VerifyPredAndCC(MI, Entry, false, Pred, LiveCPSR, HasCC, CCDead))
470 // Add the 16-bit instruction.
471 DebugLoc dl = MI->getDebugLoc();
472 MachineInstrBuilder MIB = BuildMI(MBB, *MI, dl, TII->get(Entry.NarrowOpc1));
473 MIB.addOperand(MI->getOperand(0));
475 AddDefaultT1CC(MIB, CCDead);
477 // Transfer the rest of operands.
478 unsigned NumOps = TID.getNumOperands();
479 for (unsigned i = 1, e = MI->getNumOperands(); i != e; ++i) {
480 if (i < NumOps && TID.OpInfo[i].isOptionalDef())
482 if (SkipPred && TID.OpInfo[i].isPredicate())
484 MIB.addOperand(MI->getOperand(i));
488 DOUT << "Converted 32-bit: " << *MI << " to 16-bit: " << *MIB;
495 static bool UpdateCPSRLiveness(MachineInstr &MI, bool LiveCPSR) {
497 for (unsigned i = 0, e = MI.getNumOperands(); i != e; ++i) {
498 const MachineOperand &MO = MI.getOperand(i);
499 if (!MO.isReg() || MO.isUndef())
501 if (MO.getReg() != ARM::CPSR)
509 assert(LiveCPSR && "CPSR liveness tracking is wrong!");
516 return HasDef || LiveCPSR;
519 bool Thumb2SizeReduce::ReduceMBB(MachineBasicBlock &MBB) {
520 bool Modified = false;
522 bool LiveCPSR = false;
523 // Yes, CPSR could be livein.
524 for (MachineBasicBlock::const_livein_iterator I = MBB.livein_begin(),
525 E = MBB.livein_end(); I != E; ++I) {
526 if (*I == ARM::CPSR) {
532 MachineBasicBlock::iterator MII = MBB.begin(), E = MBB.end();
533 MachineBasicBlock::iterator NextMII;
534 for (; MII != E; MII = NextMII) {
537 MachineInstr *MI = &*MII;
538 unsigned Opcode = MI->getOpcode();
539 DenseMap<unsigned, unsigned>::iterator OPI = ReduceOpcodeMap.find(Opcode);
540 if (OPI != ReduceOpcodeMap.end()) {
541 const ReduceEntry &Entry = ReduceTable[OPI->second];
542 // Ignore "special" cases for now.
544 if (ReduceSpecial(MBB, MI, Entry, LiveCPSR)) {
546 MachineBasicBlock::iterator I = prior(NextMII);
552 // Try to transform to a 16-bit two-address instruction.
553 if (Entry.NarrowOpc2 && ReduceTo2Addr(MBB, MI, Entry, LiveCPSR)) {
555 MachineBasicBlock::iterator I = prior(NextMII);
560 // Try to transform ro a 16-bit non-two-address instruction.
561 if (Entry.NarrowOpc1 && ReduceToNarrow(MBB, MI, Entry, LiveCPSR))
566 LiveCPSR = UpdateCPSRLiveness(*MI, LiveCPSR);
572 bool Thumb2SizeReduce::runOnMachineFunction(MachineFunction &MF) {
573 const TargetMachine &TM = MF.getTarget();
574 TII = TM.getInstrInfo();
576 bool Modified = false;
577 for (MachineFunction::iterator I = MF.begin(), E = MF.end(); I != E; ++I)
578 Modified |= ReduceMBB(*I);
582 /// createThumb2SizeReductionPass - Returns an instance of the Thumb2 size
584 FunctionPass *llvm::createThumb2SizeReductionPass() {
585 return new Thumb2SizeReduce();