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 //===----------------------------------------------------------------------===//
11 #include "ARMBaseInstrInfo.h"
12 #include "ARMSubtarget.h"
13 #include "MCTargetDesc/ARMAddressingModes.h"
14 #include "Thumb2InstrInfo.h"
15 #include "llvm/ADT/DenseMap.h"
16 #include "llvm/ADT/PostOrderIterator.h"
17 #include "llvm/ADT/Statistic.h"
18 #include "llvm/CodeGen/MachineFunctionPass.h"
19 #include "llvm/CodeGen/MachineInstr.h"
20 #include "llvm/CodeGen/MachineInstrBuilder.h"
21 #include "llvm/IR/Function.h" // To access Function attributes
22 #include "llvm/Support/CommandLine.h"
23 #include "llvm/Support/Debug.h"
24 #include "llvm/Support/raw_ostream.h"
25 #include "llvm/Target/TargetMachine.h"
28 #define DEBUG_TYPE "t2-reduce-size"
30 STATISTIC(NumNarrows, "Number of 32-bit instrs reduced to 16-bit ones");
31 STATISTIC(Num2Addrs, "Number of 32-bit instrs reduced to 2addr 16-bit ones");
32 STATISTIC(NumLdSts, "Number of 32-bit load / store reduced to 16-bit ones");
34 static cl::opt<int> ReduceLimit("t2-reduce-limit",
35 cl::init(-1), cl::Hidden);
36 static cl::opt<int> ReduceLimit2Addr("t2-reduce-limit2",
37 cl::init(-1), cl::Hidden);
38 static cl::opt<int> ReduceLimitLdSt("t2-reduce-limit3",
39 cl::init(-1), cl::Hidden);
42 /// ReduceTable - A static table with information on mapping from wide
45 uint16_t WideOpc; // Wide opcode
46 uint16_t NarrowOpc1; // Narrow opcode to transform to
47 uint16_t NarrowOpc2; // Narrow opcode when it's two-address
48 uint8_t Imm1Limit; // Limit of immediate field (bits)
49 uint8_t Imm2Limit; // Limit of immediate field when it's two-address
50 unsigned LowRegs1 : 1; // Only possible if low-registers are used
51 unsigned LowRegs2 : 1; // Only possible if low-registers are used (2addr)
52 unsigned PredCC1 : 2; // 0 - If predicated, cc is on and vice versa.
54 // 2 - Always set CPSR.
56 unsigned PartFlag : 1; // 16-bit instruction does partial flag update
57 unsigned Special : 1; // Needs to be dealt with specially
58 unsigned AvoidMovs: 1; // Avoid movs with shifter operand (for Swift)
61 static const ReduceEntry ReduceTable[] = {
62 // Wide, Narrow1, Narrow2, imm1,imm2, lo1, lo2, P/C,PF,S,AM
63 { ARM::t2ADCrr, 0, ARM::tADC, 0, 0, 0, 1, 0,0, 0,0,0 },
64 { ARM::t2ADDri, ARM::tADDi3, ARM::tADDi8, 3, 8, 1, 1, 0,0, 0,1,0 },
65 { ARM::t2ADDrr, ARM::tADDrr, ARM::tADDhirr, 0, 0, 1, 0, 0,1, 0,0,0 },
66 { ARM::t2ADDSri,ARM::tADDi3, ARM::tADDi8, 3, 8, 1, 1, 2,2, 0,1,0 },
67 { ARM::t2ADDSrr,ARM::tADDrr, 0, 0, 0, 1, 0, 2,0, 0,1,0 },
68 { ARM::t2ANDrr, 0, ARM::tAND, 0, 0, 0, 1, 0,0, 1,0,0 },
69 { ARM::t2ASRri, ARM::tASRri, 0, 5, 0, 1, 0, 0,0, 1,0,1 },
70 { ARM::t2ASRrr, 0, ARM::tASRrr, 0, 0, 0, 1, 0,0, 1,0,1 },
71 { ARM::t2BICrr, 0, ARM::tBIC, 0, 0, 0, 1, 0,0, 1,0,0 },
72 //FIXME: Disable CMN, as CCodes are backwards from compare expectations
73 //{ ARM::t2CMNrr, ARM::tCMN, 0, 0, 0, 1, 0, 2,0, 0,0,0 },
74 { ARM::t2CMNzrr, ARM::tCMNz, 0, 0, 0, 1, 0, 2,0, 0,0,0 },
75 { ARM::t2CMPri, ARM::tCMPi8, 0, 8, 0, 1, 0, 2,0, 0,0,0 },
76 { ARM::t2CMPrr, ARM::tCMPhir, 0, 0, 0, 0, 0, 2,0, 0,1,0 },
77 { ARM::t2EORrr, 0, ARM::tEOR, 0, 0, 0, 1, 0,0, 1,0,0 },
78 // FIXME: adr.n immediate offset must be multiple of 4.
79 //{ ARM::t2LEApcrelJT,ARM::tLEApcrelJT, 0, 0, 0, 1, 0, 1,0, 0,0,0 },
80 { ARM::t2LSLri, ARM::tLSLri, 0, 5, 0, 1, 0, 0,0, 1,0,1 },
81 { ARM::t2LSLrr, 0, ARM::tLSLrr, 0, 0, 0, 1, 0,0, 1,0,1 },
82 { ARM::t2LSRri, ARM::tLSRri, 0, 5, 0, 1, 0, 0,0, 1,0,1 },
83 { ARM::t2LSRrr, 0, ARM::tLSRrr, 0, 0, 0, 1, 0,0, 1,0,1 },
84 { ARM::t2MOVi, ARM::tMOVi8, 0, 8, 0, 1, 0, 0,0, 1,0,0 },
85 { ARM::t2MOVi16,ARM::tMOVi8, 0, 8, 0, 1, 0, 0,0, 1,1,0 },
86 // FIXME: Do we need the 16-bit 'S' variant?
87 { ARM::t2MOVr,ARM::tMOVr, 0, 0, 0, 0, 0, 1,0, 0,0,0 },
88 { ARM::t2MUL, 0, ARM::tMUL, 0, 0, 0, 1, 0,0, 1,0,0 },
89 { ARM::t2MVNr, ARM::tMVN, 0, 0, 0, 1, 0, 0,0, 0,0,0 },
90 { ARM::t2ORRrr, 0, ARM::tORR, 0, 0, 0, 1, 0,0, 1,0,0 },
91 { ARM::t2REV, ARM::tREV, 0, 0, 0, 1, 0, 1,0, 0,0,0 },
92 { ARM::t2REV16, ARM::tREV16, 0, 0, 0, 1, 0, 1,0, 0,0,0 },
93 { ARM::t2REVSH, ARM::tREVSH, 0, 0, 0, 1, 0, 1,0, 0,0,0 },
94 { ARM::t2RORrr, 0, ARM::tROR, 0, 0, 0, 1, 0,0, 1,0,0 },
95 { ARM::t2RSBri, ARM::tRSB, 0, 0, 0, 1, 0, 0,0, 0,1,0 },
96 { ARM::t2RSBSri,ARM::tRSB, 0, 0, 0, 1, 0, 2,0, 0,1,0 },
97 { ARM::t2SBCrr, 0, ARM::tSBC, 0, 0, 0, 1, 0,0, 0,0,0 },
98 { ARM::t2SUBri, ARM::tSUBi3, ARM::tSUBi8, 3, 8, 1, 1, 0,0, 0,0,0 },
99 { ARM::t2SUBrr, ARM::tSUBrr, 0, 0, 0, 1, 0, 0,0, 0,0,0 },
100 { ARM::t2SUBSri,ARM::tSUBi3, ARM::tSUBi8, 3, 8, 1, 1, 2,2, 0,0,0 },
101 { ARM::t2SUBSrr,ARM::tSUBrr, 0, 0, 0, 1, 0, 2,0, 0,0,0 },
102 { ARM::t2SXTB, ARM::tSXTB, 0, 0, 0, 1, 0, 1,0, 0,1,0 },
103 { ARM::t2SXTH, ARM::tSXTH, 0, 0, 0, 1, 0, 1,0, 0,1,0 },
104 { ARM::t2TSTrr, ARM::tTST, 0, 0, 0, 1, 0, 2,0, 0,0,0 },
105 { ARM::t2UXTB, ARM::tUXTB, 0, 0, 0, 1, 0, 1,0, 0,1,0 },
106 { ARM::t2UXTH, ARM::tUXTH, 0, 0, 0, 1, 0, 1,0, 0,1,0 },
108 // FIXME: Clean this up after splitting each Thumb load / store opcode
109 // into multiple ones.
110 { ARM::t2LDRi12,ARM::tLDRi, ARM::tLDRspi, 5, 8, 1, 0, 0,0, 0,1,0 },
111 { ARM::t2LDRs, ARM::tLDRr, 0, 0, 0, 1, 0, 0,0, 0,1,0 },
112 { ARM::t2LDRBi12,ARM::tLDRBi, 0, 5, 0, 1, 0, 0,0, 0,1,0 },
113 { ARM::t2LDRBs, ARM::tLDRBr, 0, 0, 0, 1, 0, 0,0, 0,1,0 },
114 { ARM::t2LDRHi12,ARM::tLDRHi, 0, 5, 0, 1, 0, 0,0, 0,1,0 },
115 { ARM::t2LDRHs, ARM::tLDRHr, 0, 0, 0, 1, 0, 0,0, 0,1,0 },
116 { ARM::t2LDRSBs,ARM::tLDRSB, 0, 0, 0, 1, 0, 0,0, 0,1,0 },
117 { ARM::t2LDRSHs,ARM::tLDRSH, 0, 0, 0, 1, 0, 0,0, 0,1,0 },
118 { ARM::t2STRi12,ARM::tSTRi, ARM::tSTRspi, 5, 8, 1, 0, 0,0, 0,1,0 },
119 { ARM::t2STRs, ARM::tSTRr, 0, 0, 0, 1, 0, 0,0, 0,1,0 },
120 { ARM::t2STRBi12,ARM::tSTRBi, 0, 5, 0, 1, 0, 0,0, 0,1,0 },
121 { ARM::t2STRBs, ARM::tSTRBr, 0, 0, 0, 1, 0, 0,0, 0,1,0 },
122 { ARM::t2STRHi12,ARM::tSTRHi, 0, 5, 0, 1, 0, 0,0, 0,1,0 },
123 { ARM::t2STRHs, ARM::tSTRHr, 0, 0, 0, 1, 0, 0,0, 0,1,0 },
125 { ARM::t2LDMIA, ARM::tLDMIA, 0, 0, 0, 1, 1, 1,1, 0,1,0 },
126 { ARM::t2LDMIA_RET,0, ARM::tPOP_RET, 0, 0, 1, 1, 1,1, 0,1,0 },
127 { ARM::t2LDMIA_UPD,ARM::tLDMIA_UPD,ARM::tPOP,0, 0, 1, 1, 1,1, 0,1,0 },
128 // ARM::t2STMIA (with no basereg writeback) has no Thumb1 equivalent.
129 // tSTMIA_UPD is a change in semantics which can only be used if the base
130 // register is killed. This difference is correctly handled elsewhere.
131 { ARM::t2STMIA, ARM::tSTMIA_UPD, 0, 0, 0, 1, 1, 1,1, 0,1,0 },
132 { ARM::t2STMIA_UPD,ARM::tSTMIA_UPD, 0, 0, 0, 1, 1, 1,1, 0,1,0 },
133 { ARM::t2STMDB_UPD, 0, ARM::tPUSH, 0, 0, 1, 1, 1,1, 0,1,0 }
136 class Thumb2SizeReduce : public MachineFunctionPass {
141 const Thumb2InstrInfo *TII;
142 const ARMSubtarget *STI;
144 bool runOnMachineFunction(MachineFunction &MF) override;
146 const char *getPassName() const override {
147 return "Thumb2 instruction size reduction pass";
151 /// ReduceOpcodeMap - Maps wide opcode to index of entry in ReduceTable.
152 DenseMap<unsigned, unsigned> ReduceOpcodeMap;
154 bool canAddPseudoFlagDep(MachineInstr *Use, bool IsSelfLoop);
156 bool VerifyPredAndCC(MachineInstr *MI, const ReduceEntry &Entry,
157 bool is2Addr, ARMCC::CondCodes Pred,
158 bool LiveCPSR, bool &HasCC, bool &CCDead);
160 bool ReduceLoadStore(MachineBasicBlock &MBB, MachineInstr *MI,
161 const ReduceEntry &Entry);
163 bool ReduceSpecial(MachineBasicBlock &MBB, MachineInstr *MI,
164 const ReduceEntry &Entry, bool LiveCPSR, bool IsSelfLoop);
166 /// ReduceTo2Addr - Reduce a 32-bit instruction to a 16-bit two-address
168 bool ReduceTo2Addr(MachineBasicBlock &MBB, MachineInstr *MI,
169 const ReduceEntry &Entry, bool LiveCPSR,
172 /// ReduceToNarrow - Reduce a 32-bit instruction to a 16-bit
173 /// non-two-address instruction.
174 bool ReduceToNarrow(MachineBasicBlock &MBB, MachineInstr *MI,
175 const ReduceEntry &Entry, bool LiveCPSR,
178 /// ReduceMI - Attempt to reduce MI, return true on success.
179 bool ReduceMI(MachineBasicBlock &MBB, MachineInstr *MI,
180 bool LiveCPSR, bool IsSelfLoop);
182 /// ReduceMBB - Reduce width of instructions in the specified basic block.
183 bool ReduceMBB(MachineBasicBlock &MBB);
188 // Last instruction to define CPSR in the current block.
189 MachineInstr *CPSRDef;
190 // Was CPSR last defined by a high latency instruction?
191 // When CPSRDef is null, this refers to CPSR defs in predecessors.
192 bool HighLatencyCPSR;
195 // The flags leaving this block have high latency.
196 bool HighLatencyCPSR;
197 // Has this block been visited yet?
200 MBBInfo() : HighLatencyCPSR(false), Visited(false) {}
203 SmallVector<MBBInfo, 8> BlockInfo;
205 char Thumb2SizeReduce::ID = 0;
208 Thumb2SizeReduce::Thumb2SizeReduce() : MachineFunctionPass(ID) {
209 OptimizeSize = MinimizeSize = false;
210 for (unsigned i = 0, e = array_lengthof(ReduceTable); i != e; ++i) {
211 unsigned FromOpc = ReduceTable[i].WideOpc;
212 if (!ReduceOpcodeMap.insert(std::make_pair(FromOpc, i)).second)
213 assert(false && "Duplicated entries?");
217 static bool HasImplicitCPSRDef(const MCInstrDesc &MCID) {
218 for (const uint16_t *Regs = MCID.getImplicitDefs(); *Regs; ++Regs)
219 if (*Regs == ARM::CPSR)
224 // Check for a likely high-latency flag def.
225 static bool isHighLatencyCPSR(MachineInstr *Def) {
226 switch(Def->getOpcode()) {
234 /// canAddPseudoFlagDep - For A9 (and other out-of-order) implementations,
235 /// the 's' 16-bit instruction partially update CPSR. Abort the
236 /// transformation to avoid adding false dependency on last CPSR setting
237 /// instruction which hurts the ability for out-of-order execution engine
238 /// to do register renaming magic.
239 /// This function checks if there is a read-of-write dependency between the
240 /// last instruction that defines the CPSR and the current instruction. If there
241 /// is, then there is no harm done since the instruction cannot be retired
242 /// before the CPSR setting instruction anyway.
243 /// Note, we are not doing full dependency analysis here for the sake of compile
244 /// time. We're not looking for cases like:
246 /// r1 = add.w r0, ...
249 /// In this case it would have been ok to narrow the mul.w to muls since there
250 /// are indirect RAW dependency between the muls and the mul.w
252 Thumb2SizeReduce::canAddPseudoFlagDep(MachineInstr *Use, bool FirstInSelfLoop) {
253 // Disable the check for -Oz (aka OptimizeForSizeHarder).
254 if (MinimizeSize || !STI->avoidCPSRPartialUpdate())
258 // If this BB loops back to itself, conservatively avoid narrowing the
259 // first instruction that does partial flag update.
260 return HighLatencyCPSR || FirstInSelfLoop;
262 SmallSet<unsigned, 2> Defs;
263 for (const MachineOperand &MO : CPSRDef->operands()) {
264 if (!MO.isReg() || MO.isUndef() || MO.isUse())
266 unsigned Reg = MO.getReg();
267 if (Reg == 0 || Reg == ARM::CPSR)
272 for (const MachineOperand &MO : Use->operands()) {
273 if (!MO.isReg() || MO.isUndef() || MO.isDef())
275 unsigned Reg = MO.getReg();
280 // If the current CPSR has high latency, try to avoid the false dependency.
284 // tMOVi8 usually doesn't start long dependency chains, and there are a lot
285 // of them, so always shrink them when CPSR doesn't have high latency.
286 if (Use->getOpcode() == ARM::t2MOVi ||
287 Use->getOpcode() == ARM::t2MOVi16)
290 // No read-after-write dependency. The narrowing will add false dependency.
295 Thumb2SizeReduce::VerifyPredAndCC(MachineInstr *MI, const ReduceEntry &Entry,
296 bool is2Addr, ARMCC::CondCodes Pred,
297 bool LiveCPSR, bool &HasCC, bool &CCDead) {
298 if ((is2Addr && Entry.PredCC2 == 0) ||
299 (!is2Addr && Entry.PredCC1 == 0)) {
300 if (Pred == ARMCC::AL) {
301 // Not predicated, must set CPSR.
303 // Original instruction was not setting CPSR, but CPSR is not
304 // currently live anyway. It's ok to set it. The CPSR def is
314 // Predicated, must not set CPSR.
318 } else if ((is2Addr && Entry.PredCC2 == 2) ||
319 (!is2Addr && Entry.PredCC1 == 2)) {
320 /// Old opcode has an optional def of CPSR.
323 // If old opcode does not implicitly define CPSR, then it's not ok since
324 // these new opcodes' CPSR def is not meant to be thrown away. e.g. CMP.
325 if (!HasImplicitCPSRDef(MI->getDesc()))
329 // 16-bit instruction does not set CPSR.
337 static bool VerifyLowRegs(MachineInstr *MI) {
338 unsigned Opc = MI->getOpcode();
339 bool isPCOk = (Opc == ARM::t2LDMIA_RET || Opc == ARM::t2LDMIA_UPD);
340 bool isLROk = (Opc == ARM::t2STMDB_UPD);
341 bool isSPOk = isPCOk || isLROk;
342 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
343 const MachineOperand &MO = MI->getOperand(i);
344 if (!MO.isReg() || MO.isImplicit())
346 unsigned Reg = MO.getReg();
347 if (Reg == 0 || Reg == ARM::CPSR)
349 if (isPCOk && Reg == ARM::PC)
351 if (isLROk && Reg == ARM::LR)
353 if (Reg == ARM::SP) {
356 if (i == 1 && (Opc == ARM::t2LDRi12 || Opc == ARM::t2STRi12))
357 // Special case for these ldr / str with sp as base register.
360 if (!isARMLowRegister(Reg))
367 Thumb2SizeReduce::ReduceLoadStore(MachineBasicBlock &MBB, MachineInstr *MI,
368 const ReduceEntry &Entry) {
369 if (ReduceLimitLdSt != -1 && ((int)NumLdSts >= ReduceLimitLdSt))
373 bool HasImmOffset = false;
374 bool HasShift = false;
375 bool HasOffReg = true;
376 bool isLdStMul = false;
377 unsigned Opc = Entry.NarrowOpc1;
378 unsigned OpNum = 3; // First 'rest' of operands.
379 uint8_t ImmLimit = Entry.Imm1Limit;
381 switch (Entry.WideOpc) {
383 llvm_unreachable("Unexpected Thumb2 load / store opcode!");
386 if (MI->getOperand(1).getReg() == ARM::SP) {
387 Opc = Entry.NarrowOpc2;
388 ImmLimit = Entry.Imm2Limit;
418 unsigned BaseReg = MI->getOperand(0).getReg();
419 assert(isARMLowRegister(BaseReg));
421 // For the non-writeback version (this one), the base register must be
422 // one of the registers being loaded.
424 for (unsigned i = 3; i < MI->getNumOperands(); ++i) {
425 if (MI->getOperand(i).getReg() == BaseReg) {
439 // If the base register is killed, we don't care what its value is after the
440 // instruction, so we can use an updating STMIA.
441 if (!MI->getOperand(0).isKill())
446 case ARM::t2LDMIA_RET: {
447 unsigned BaseReg = MI->getOperand(1).getReg();
448 if (BaseReg != ARM::SP)
450 Opc = Entry.NarrowOpc2; // tPOP_RET
455 case ARM::t2LDMIA_UPD:
456 case ARM::t2STMIA_UPD:
457 case ARM::t2STMDB_UPD: {
460 unsigned BaseReg = MI->getOperand(1).getReg();
461 if (BaseReg == ARM::SP &&
462 (Entry.WideOpc == ARM::t2LDMIA_UPD ||
463 Entry.WideOpc == ARM::t2STMDB_UPD)) {
464 Opc = Entry.NarrowOpc2; // tPOP or tPUSH
466 } else if (!isARMLowRegister(BaseReg) ||
467 (Entry.WideOpc != ARM::t2LDMIA_UPD &&
468 Entry.WideOpc != ARM::t2STMIA_UPD)) {
477 unsigned OffsetReg = 0;
478 bool OffsetKill = false;
479 bool OffsetInternal = false;
481 OffsetReg = MI->getOperand(2).getReg();
482 OffsetKill = MI->getOperand(2).isKill();
483 OffsetInternal = MI->getOperand(2).isInternalRead();
485 if (MI->getOperand(3).getImm())
486 // Thumb1 addressing mode doesn't support shift.
490 unsigned OffsetImm = 0;
492 OffsetImm = MI->getOperand(2).getImm();
493 unsigned MaxOffset = ((1 << ImmLimit) - 1) * Scale;
495 if ((OffsetImm & (Scale - 1)) || OffsetImm > MaxOffset)
496 // Make sure the immediate field fits.
500 // Add the 16-bit load / store instruction.
501 DebugLoc dl = MI->getDebugLoc();
502 MachineInstrBuilder MIB = BuildMI(MBB, MI, dl, TII->get(Opc));
504 // tSTMIA_UPD takes a defining register operand. We've already checked that
505 // the register is killed, so mark it as dead here.
506 if (Entry.WideOpc == ARM::t2STMIA)
507 MIB.addReg(MI->getOperand(0).getReg(), RegState::Define | RegState::Dead);
510 MIB.addOperand(MI->getOperand(0));
511 MIB.addOperand(MI->getOperand(1));
514 MIB.addImm(OffsetImm / Scale);
516 assert((!HasShift || OffsetReg) && "Invalid so_reg load / store address!");
519 MIB.addReg(OffsetReg, getKillRegState(OffsetKill) |
520 getInternalReadRegState(OffsetInternal));
523 // Transfer the rest of operands.
524 for (unsigned e = MI->getNumOperands(); OpNum != e; ++OpNum)
525 MIB.addOperand(MI->getOperand(OpNum));
527 // Transfer memoperands.
528 MIB->setMemRefs(MI->memoperands_begin(), MI->memoperands_end());
530 // Transfer MI flags.
531 MIB.setMIFlags(MI->getFlags());
533 DEBUG(errs() << "Converted 32-bit: " << *MI << " to 16-bit: " << *MIB);
541 Thumb2SizeReduce::ReduceSpecial(MachineBasicBlock &MBB, MachineInstr *MI,
542 const ReduceEntry &Entry,
543 bool LiveCPSR, bool IsSelfLoop) {
544 unsigned Opc = MI->getOpcode();
545 if (Opc == ARM::t2ADDri) {
546 // If the source register is SP, try to reduce to tADDrSPi, otherwise
547 // it's a normal reduce.
548 if (MI->getOperand(1).getReg() != ARM::SP) {
549 if (ReduceTo2Addr(MBB, MI, Entry, LiveCPSR, IsSelfLoop))
551 return ReduceToNarrow(MBB, MI, Entry, LiveCPSR, IsSelfLoop);
553 // Try to reduce to tADDrSPi.
554 unsigned Imm = MI->getOperand(2).getImm();
555 // The immediate must be in range, the destination register must be a low
556 // reg, the predicate must be "always" and the condition flags must not
558 if (Imm & 3 || Imm > 1020)
560 if (!isARMLowRegister(MI->getOperand(0).getReg()))
562 if (MI->getOperand(3).getImm() != ARMCC::AL)
564 const MCInstrDesc &MCID = MI->getDesc();
565 if (MCID.hasOptionalDef() &&
566 MI->getOperand(MCID.getNumOperands()-1).getReg() == ARM::CPSR)
569 MachineInstrBuilder MIB = BuildMI(MBB, MI, MI->getDebugLoc(),
570 TII->get(ARM::tADDrSPi))
571 .addOperand(MI->getOperand(0))
572 .addOperand(MI->getOperand(1))
573 .addImm(Imm / 4); // The tADDrSPi has an implied scale by four.
576 // Transfer MI flags.
577 MIB.setMIFlags(MI->getFlags());
579 DEBUG(errs() << "Converted 32-bit: " << *MI << " to 16-bit: " <<*MIB);
586 if (Entry.LowRegs1 && !VerifyLowRegs(MI))
589 if (MI->mayLoadOrStore())
590 return ReduceLoadStore(MBB, MI, Entry);
595 case ARM::t2ADDSrr: {
596 unsigned PredReg = 0;
597 if (getInstrPredicate(MI, PredReg) == ARMCC::AL) {
600 case ARM::t2ADDSri: {
601 if (ReduceTo2Addr(MBB, MI, Entry, LiveCPSR, IsSelfLoop))
606 return ReduceToNarrow(MBB, MI, Entry, LiveCPSR, IsSelfLoop);
617 if (MI->getOperand(2).getImm() == 0)
618 return ReduceToNarrow(MBB, MI, Entry, LiveCPSR, IsSelfLoop);
621 // Can convert only 'pure' immediate operands, not immediates obtained as
622 // globals' addresses.
623 if (MI->getOperand(1).isImm())
624 return ReduceToNarrow(MBB, MI, Entry, LiveCPSR, IsSelfLoop);
627 // Try to reduce to the lo-reg only version first. Why there are two
628 // versions of the instruction is a mystery.
629 // It would be nice to just have two entries in the master table that
630 // are prioritized, but the table assumes a unique entry for each
631 // source insn opcode. So for now, we hack a local entry record to use.
632 static const ReduceEntry NarrowEntry =
633 { ARM::t2CMPrr,ARM::tCMPr, 0, 0, 0, 1, 1,2, 0, 0,1,0 };
634 if (ReduceToNarrow(MBB, MI, NarrowEntry, LiveCPSR, IsSelfLoop))
636 return ReduceToNarrow(MBB, MI, Entry, LiveCPSR, IsSelfLoop);
643 Thumb2SizeReduce::ReduceTo2Addr(MachineBasicBlock &MBB, MachineInstr *MI,
644 const ReduceEntry &Entry,
645 bool LiveCPSR, bool IsSelfLoop) {
647 if (ReduceLimit2Addr != -1 && ((int)Num2Addrs >= ReduceLimit2Addr))
650 if (!MinimizeSize && !OptimizeSize && Entry.AvoidMovs &&
651 STI->avoidMOVsShifterOperand())
652 // Don't issue movs with shifter operand for some CPUs unless we
653 // are optimizing / minimizing for size.
656 unsigned Reg0 = MI->getOperand(0).getReg();
657 unsigned Reg1 = MI->getOperand(1).getReg();
658 // t2MUL is "special". The tied source operand is second, not first.
659 if (MI->getOpcode() == ARM::t2MUL) {
660 unsigned Reg2 = MI->getOperand(2).getReg();
661 // Early exit if the regs aren't all low regs.
662 if (!isARMLowRegister(Reg0) || !isARMLowRegister(Reg1)
663 || !isARMLowRegister(Reg2))
666 // If the other operand also isn't the same as the destination, we
670 // Try to commute the operands to make it a 2-address instruction.
671 MachineInstr *CommutedMI = TII->commuteInstruction(MI);
675 } else if (Reg0 != Reg1) {
676 // Try to commute the operands to make it a 2-address instruction.
677 unsigned CommOpIdx1, CommOpIdx2;
678 if (!TII->findCommutedOpIndices(MI, CommOpIdx1, CommOpIdx2) ||
679 CommOpIdx1 != 1 || MI->getOperand(CommOpIdx2).getReg() != Reg0)
681 MachineInstr *CommutedMI = TII->commuteInstruction(MI);
685 if (Entry.LowRegs2 && !isARMLowRegister(Reg0))
687 if (Entry.Imm2Limit) {
688 unsigned Imm = MI->getOperand(2).getImm();
689 unsigned Limit = (1 << Entry.Imm2Limit) - 1;
693 unsigned Reg2 = MI->getOperand(2).getReg();
694 if (Entry.LowRegs2 && !isARMLowRegister(Reg2))
698 // Check if it's possible / necessary to transfer the predicate.
699 const MCInstrDesc &NewMCID = TII->get(Entry.NarrowOpc2);
700 unsigned PredReg = 0;
701 ARMCC::CondCodes Pred = getInstrPredicate(MI, PredReg);
702 bool SkipPred = false;
703 if (Pred != ARMCC::AL) {
704 if (!NewMCID.isPredicable())
705 // Can't transfer predicate, fail.
708 SkipPred = !NewMCID.isPredicable();
713 const MCInstrDesc &MCID = MI->getDesc();
714 if (MCID.hasOptionalDef()) {
715 unsigned NumOps = MCID.getNumOperands();
716 HasCC = (MI->getOperand(NumOps-1).getReg() == ARM::CPSR);
717 if (HasCC && MI->getOperand(NumOps-1).isDead())
720 if (!VerifyPredAndCC(MI, Entry, true, Pred, LiveCPSR, HasCC, CCDead))
723 // Avoid adding a false dependency on partial flag update by some 16-bit
724 // instructions which has the 's' bit set.
725 if (Entry.PartFlag && NewMCID.hasOptionalDef() && HasCC &&
726 canAddPseudoFlagDep(MI, IsSelfLoop))
729 // Add the 16-bit instruction.
730 DebugLoc dl = MI->getDebugLoc();
731 MachineInstrBuilder MIB = BuildMI(MBB, MI, dl, NewMCID);
732 MIB.addOperand(MI->getOperand(0));
733 if (NewMCID.hasOptionalDef()) {
735 AddDefaultT1CC(MIB, CCDead);
740 // Transfer the rest of operands.
741 unsigned NumOps = MCID.getNumOperands();
742 for (unsigned i = 1, e = MI->getNumOperands(); i != e; ++i) {
743 if (i < NumOps && MCID.OpInfo[i].isOptionalDef())
745 if (SkipPred && MCID.OpInfo[i].isPredicate())
747 MIB.addOperand(MI->getOperand(i));
750 // Transfer MI flags.
751 MIB.setMIFlags(MI->getFlags());
753 DEBUG(errs() << "Converted 32-bit: " << *MI << " to 16-bit: " << *MIB);
761 Thumb2SizeReduce::ReduceToNarrow(MachineBasicBlock &MBB, MachineInstr *MI,
762 const ReduceEntry &Entry,
763 bool LiveCPSR, bool IsSelfLoop) {
764 if (ReduceLimit != -1 && ((int)NumNarrows >= ReduceLimit))
767 if (!MinimizeSize && !OptimizeSize && Entry.AvoidMovs &&
768 STI->avoidMOVsShifterOperand())
769 // Don't issue movs with shifter operand for some CPUs unless we
770 // are optimizing / minimizing for size.
773 unsigned Limit = ~0U;
775 Limit = (1 << Entry.Imm1Limit) - 1;
777 const MCInstrDesc &MCID = MI->getDesc();
778 for (unsigned i = 0, e = MCID.getNumOperands(); i != e; ++i) {
779 if (MCID.OpInfo[i].isPredicate())
781 const MachineOperand &MO = MI->getOperand(i);
783 unsigned Reg = MO.getReg();
784 if (!Reg || Reg == ARM::CPSR)
786 if (Entry.LowRegs1 && !isARMLowRegister(Reg))
788 } else if (MO.isImm() &&
789 !MCID.OpInfo[i].isPredicate()) {
790 if (((unsigned)MO.getImm()) > Limit)
795 // Check if it's possible / necessary to transfer the predicate.
796 const MCInstrDesc &NewMCID = TII->get(Entry.NarrowOpc1);
797 unsigned PredReg = 0;
798 ARMCC::CondCodes Pred = getInstrPredicate(MI, PredReg);
799 bool SkipPred = false;
800 if (Pred != ARMCC::AL) {
801 if (!NewMCID.isPredicable())
802 // Can't transfer predicate, fail.
805 SkipPred = !NewMCID.isPredicable();
810 if (MCID.hasOptionalDef()) {
811 unsigned NumOps = MCID.getNumOperands();
812 HasCC = (MI->getOperand(NumOps-1).getReg() == ARM::CPSR);
813 if (HasCC && MI->getOperand(NumOps-1).isDead())
816 if (!VerifyPredAndCC(MI, Entry, false, Pred, LiveCPSR, HasCC, CCDead))
819 // Avoid adding a false dependency on partial flag update by some 16-bit
820 // instructions which has the 's' bit set.
821 if (Entry.PartFlag && NewMCID.hasOptionalDef() && HasCC &&
822 canAddPseudoFlagDep(MI, IsSelfLoop))
825 // Add the 16-bit instruction.
826 DebugLoc dl = MI->getDebugLoc();
827 MachineInstrBuilder MIB = BuildMI(MBB, MI, dl, NewMCID);
828 MIB.addOperand(MI->getOperand(0));
829 if (NewMCID.hasOptionalDef()) {
831 AddDefaultT1CC(MIB, CCDead);
836 // Transfer the rest of operands.
837 unsigned NumOps = MCID.getNumOperands();
838 for (unsigned i = 1, e = MI->getNumOperands(); i != e; ++i) {
839 if (i < NumOps && MCID.OpInfo[i].isOptionalDef())
841 if ((MCID.getOpcode() == ARM::t2RSBSri ||
842 MCID.getOpcode() == ARM::t2RSBri ||
843 MCID.getOpcode() == ARM::t2SXTB ||
844 MCID.getOpcode() == ARM::t2SXTH ||
845 MCID.getOpcode() == ARM::t2UXTB ||
846 MCID.getOpcode() == ARM::t2UXTH) && i == 2)
847 // Skip the zero immediate operand, it's now implicit.
849 bool isPred = (i < NumOps && MCID.OpInfo[i].isPredicate());
850 if (SkipPred && isPred)
852 const MachineOperand &MO = MI->getOperand(i);
853 if (MO.isReg() && MO.isImplicit() && MO.getReg() == ARM::CPSR)
854 // Skip implicit def of CPSR. Either it's modeled as an optional
855 // def now or it's already an implicit def on the new instruction.
859 if (!MCID.isPredicable() && NewMCID.isPredicable())
862 // Transfer MI flags.
863 MIB.setMIFlags(MI->getFlags());
865 DEBUG(errs() << "Converted 32-bit: " << *MI << " to 16-bit: " << *MIB);
872 static bool UpdateCPSRDef(MachineInstr &MI, bool LiveCPSR, bool &DefCPSR) {
874 for (const MachineOperand &MO : MI.operands()) {
875 if (!MO.isReg() || MO.isUndef() || MO.isUse())
877 if (MO.getReg() != ARM::CPSR)
885 return HasDef || LiveCPSR;
888 static bool UpdateCPSRUse(MachineInstr &MI, bool LiveCPSR) {
889 for (const MachineOperand &MO : MI.operands()) {
890 if (!MO.isReg() || MO.isUndef() || MO.isDef())
892 if (MO.getReg() != ARM::CPSR)
894 assert(LiveCPSR && "CPSR liveness tracking is wrong!");
904 bool Thumb2SizeReduce::ReduceMI(MachineBasicBlock &MBB, MachineInstr *MI,
905 bool LiveCPSR, bool IsSelfLoop) {
906 unsigned Opcode = MI->getOpcode();
907 DenseMap<unsigned, unsigned>::iterator OPI = ReduceOpcodeMap.find(Opcode);
908 if (OPI == ReduceOpcodeMap.end())
910 const ReduceEntry &Entry = ReduceTable[OPI->second];
912 // Don't attempt normal reductions on "special" cases for now.
914 return ReduceSpecial(MBB, MI, Entry, LiveCPSR, IsSelfLoop);
916 // Try to transform to a 16-bit two-address instruction.
917 if (Entry.NarrowOpc2 &&
918 ReduceTo2Addr(MBB, MI, Entry, LiveCPSR, IsSelfLoop))
921 // Try to transform to a 16-bit non-two-address instruction.
922 if (Entry.NarrowOpc1 &&
923 ReduceToNarrow(MBB, MI, Entry, LiveCPSR, IsSelfLoop))
929 bool Thumb2SizeReduce::ReduceMBB(MachineBasicBlock &MBB) {
930 bool Modified = false;
932 // Yes, CPSR could be livein.
933 bool LiveCPSR = MBB.isLiveIn(ARM::CPSR);
934 MachineInstr *BundleMI = nullptr;
937 HighLatencyCPSR = false;
939 // Check predecessors for the latest CPSRDef.
940 for (auto *Pred : MBB.predecessors()) {
941 const MBBInfo &PInfo = BlockInfo[Pred->getNumber()];
942 if (!PInfo.Visited) {
943 // Since blocks are visited in RPO, this must be a back-edge.
946 if (PInfo.HighLatencyCPSR) {
947 HighLatencyCPSR = true;
952 // If this BB loops back to itself, conservatively avoid narrowing the
953 // first instruction that does partial flag update.
954 bool IsSelfLoop = MBB.isSuccessor(&MBB);
955 MachineBasicBlock::instr_iterator MII = MBB.instr_begin(),E = MBB.instr_end();
956 MachineBasicBlock::instr_iterator NextMII;
957 for (; MII != E; MII = NextMII) {
958 NextMII = std::next(MII);
960 MachineInstr *MI = &*MII;
961 if (MI->isBundle()) {
965 if (MI->isDebugValue())
968 LiveCPSR = UpdateCPSRUse(*MI, LiveCPSR);
970 // Does NextMII belong to the same bundle as MI?
971 bool NextInSameBundle = NextMII != E && NextMII->isBundledWithPred();
973 if (ReduceMI(MBB, MI, LiveCPSR, IsSelfLoop)) {
975 MachineBasicBlock::instr_iterator I = std::prev(NextMII);
977 // Removing and reinserting the first instruction in a bundle will break
978 // up the bundle. Fix the bundling if it was broken.
979 if (NextInSameBundle && !NextMII->isBundledWithPred())
980 NextMII->bundleWithPred();
983 if (!NextInSameBundle && MI->isInsideBundle()) {
984 // FIXME: Since post-ra scheduler operates on bundles, the CPSR kill
985 // marker is only on the BUNDLE instruction. Process the BUNDLE
986 // instruction as we finish with the bundled instruction to work around
987 // the inconsistency.
988 if (BundleMI->killsRegister(ARM::CPSR))
990 MachineOperand *MO = BundleMI->findRegisterDefOperand(ARM::CPSR);
991 if (MO && !MO->isDead())
993 MO = BundleMI->findRegisterUseOperand(ARM::CPSR);
994 if (MO && !MO->isKill())
998 bool DefCPSR = false;
999 LiveCPSR = UpdateCPSRDef(*MI, LiveCPSR, DefCPSR);
1001 // Calls don't really set CPSR.
1003 HighLatencyCPSR = false;
1005 } else if (DefCPSR) {
1006 // This is the last CPSR defining instruction.
1008 HighLatencyCPSR = isHighLatencyCPSR(CPSRDef);
1013 MBBInfo &Info = BlockInfo[MBB.getNumber()];
1014 Info.HighLatencyCPSR = HighLatencyCPSR;
1015 Info.Visited = true;
1019 bool Thumb2SizeReduce::runOnMachineFunction(MachineFunction &MF) {
1020 STI = &static_cast<const ARMSubtarget &>(MF.getSubtarget());
1021 if (STI->isThumb1Only() || STI->prefers32BitThumb())
1024 TII = static_cast<const Thumb2InstrInfo *>(STI->getInstrInfo());
1026 // Optimizing / minimizing size?
1027 OptimizeSize = MF.getFunction()->hasFnAttribute(Attribute::OptimizeForSize);
1028 MinimizeSize = MF.getFunction()->hasFnAttribute(Attribute::MinSize);
1031 BlockInfo.resize(MF.getNumBlockIDs());
1033 // Visit blocks in reverse post-order so LastCPSRDef is known for all
1035 ReversePostOrderTraversal<MachineFunction*> RPOT(&MF);
1036 bool Modified = false;
1037 for (ReversePostOrderTraversal<MachineFunction*>::rpo_iterator
1038 I = RPOT.begin(), E = RPOT.end(); I != E; ++I)
1039 Modified |= ReduceMBB(**I);
1043 /// createThumb2SizeReductionPass - Returns an instance of the Thumb2 size
1045 FunctionPass *llvm::createThumb2SizeReductionPass() {
1046 return new Thumb2SizeReduce();