Shrink Thumb2 movcc instructions.
[llvm/avr.git] / lib / Target / ARM / Thumb2SizeReduction.cpp
blobc9f3ad2de2da6d781eb47758e19034e9f5b9c5d0
1 //===-- Thumb2SizeReduction.cpp - Thumb2 code size reduction pass -*- C++ -*-=//
2 //
3 // The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
10 #define DEBUG_TYPE "t2-reduce-size"
11 #include "ARM.h"
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"
24 using namespace llvm;
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);
37 namespace {
38 /// ReduceTable - A static table with information on mapping from wide
39 /// opcodes to narrow
40 struct ReduceEntry {
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 : 2; // 0 - If predicated, cc is on and vice versa.
49 // 1 - No cc field.
50 // 2 - Always set CPSR.
51 unsigned PredCC2 : 1;
52 unsigned Special : 1; // Needs to be dealt with specially
55 static const ReduceEntry ReduceTable[] = {
56 // Wide, Narrow1, Narrow2, imm1,imm2, lo1, lo2, P/C, S
57 { ARM::t2ADCrr, 0, ARM::tADC, 0, 0, 0, 1, 0,0, 0 },
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 // Note: immediate scale is 4.
61 { ARM::t2ADDrSPi,ARM::tADDrSPi,0, 8, 0, 1, 0, 1,0, 0 },
62 { ARM::t2ADDSri,ARM::tADDi3, ARM::tADDi8, 3, 8, 1, 1, 2,2, 1 },
63 { ARM::t2ADDSrr,ARM::tADDrr, 0, 0, 0, 1, 0, 2,0, 1 },
64 { ARM::t2ANDrr, 0, ARM::tAND, 0, 0, 0, 1, 0,0, 0 },
65 { ARM::t2ASRri, ARM::tASRri, 0, 5, 0, 1, 0, 0,0, 0 },
66 { ARM::t2ASRrr, 0, ARM::tASRrr, 0, 0, 0, 1, 0,0, 0 },
67 { ARM::t2BICrr, 0, ARM::tBIC, 0, 0, 0, 1, 0,0, 0 },
68 { ARM::t2CMNrr, ARM::tCMN, 0, 0, 0, 1, 0, 2,0, 0 },
69 { ARM::t2CMPri, ARM::tCMPi8, 0, 8, 0, 1, 0, 2,0, 0 },
70 { ARM::t2CMPrr, ARM::tCMPhir, 0, 0, 0, 0, 0, 2,0, 0 },
71 { ARM::t2CMPzri,ARM::tCMPzi8, 0, 8, 0, 1, 0, 2,0, 0 },
72 { ARM::t2CMPzrr,ARM::tCMPzhir,0, 0, 0, 0, 0, 2,0, 0 },
73 { ARM::t2EORrr, 0, ARM::tEOR, 0, 0, 0, 1, 0,0, 0 },
74 { ARM::t2LSLri, ARM::tLSLri, 0, 5, 0, 1, 0, 0,0, 0 },
75 { ARM::t2LSLrr, 0, ARM::tLSLrr, 0, 0, 0, 1, 0,0, 0 },
76 { ARM::t2LSRri, ARM::tLSRri, 0, 5, 0, 1, 0, 0,0, 0 },
77 { ARM::t2LSRrr, 0, ARM::tLSRrr, 0, 0, 0, 1, 0,0, 0 },
78 { ARM::t2MOVi, ARM::tMOVi8, 0, 8, 0, 1, 0, 0,0, 0 },
79 // FIXME: Do we need the 16-bit 'S' variant?
80 { ARM::t2MOVr,ARM::tMOVgpr2gpr,0, 0, 0, 0, 0, 1,0, 0 },
81 { ARM::t2MOVCCr,0, ARM::tMOVCCr, 0, 0, 0, 0, 0,1, 0 },
82 { ARM::t2MOVCCi,0, ARM::tMOVCCi, 0, 8, 0, 0, 0,1, 0 },
83 { ARM::t2MUL, 0, ARM::tMUL, 0, 0, 0, 1, 0,0, 0 },
84 { ARM::t2MVNr, ARM::tMVN, 0, 0, 0, 1, 0, 0,0, 0 },
85 { ARM::t2ORRrr, 0, ARM::tORR, 0, 0, 0, 1, 0,0, 0 },
86 { ARM::t2REV, ARM::tREV, 0, 0, 0, 1, 0, 1,0, 0 },
87 { ARM::t2REV16, ARM::tREV16, 0, 0, 0, 1, 0, 1,0, 0 },
88 { ARM::t2REVSH, ARM::tREVSH, 0, 0, 0, 1, 0, 1,0, 0 },
89 { ARM::t2RORrr, 0, ARM::tROR, 0, 0, 0, 1, 0,0, 0 },
90 { ARM::t2RSBri, ARM::tRSB, 0, 0, 0, 1, 0, 0,0, 1 },
91 { ARM::t2RSBSri,ARM::tRSB, 0, 0, 0, 1, 0, 2,0, 1 },
92 { ARM::t2SBCrr, 0, ARM::tSBC, 0, 0, 0, 1, 0,0, 0 },
93 { ARM::t2SUBri, ARM::tSUBi3, ARM::tSUBi8, 3, 8, 1, 1, 0,0, 0 },
94 { ARM::t2SUBrr, ARM::tSUBrr, 0, 0, 0, 1, 0, 0,0, 0 },
95 { ARM::t2SUBSri,ARM::tSUBi3, ARM::tSUBi8, 3, 8, 1, 1, 2,2, 0 },
96 { ARM::t2SUBSrr,ARM::tSUBrr, 0, 0, 0, 1, 0, 2,0, 0 },
97 { ARM::t2SXTBr, ARM::tSXTB, 0, 0, 0, 1, 0, 1,0, 0 },
98 { ARM::t2SXTHr, ARM::tSXTH, 0, 0, 0, 1, 0, 1,0, 0 },
99 { ARM::t2TSTrr, ARM::tTST, 0, 0, 0, 1, 0, 2,0, 0 },
100 { ARM::t2UXTBr, ARM::tUXTB, 0, 0, 0, 1, 0, 1,0, 0 },
101 { ARM::t2UXTHr, ARM::tUXTH, 0, 0, 0, 1, 0, 1,0, 0 },
103 // FIXME: Clean this up after splitting each Thumb load / store opcode
104 // into multiple ones.
105 { ARM::t2LDRi12,ARM::tLDR, 0, 5, 0, 1, 0, 0,0, 1 },
106 { ARM::t2LDRs, ARM::tLDR, 0, 0, 0, 1, 0, 0,0, 1 },
107 { ARM::t2LDRBi12,ARM::tLDRB, 0, 5, 0, 1, 0, 0,0, 1 },
108 { ARM::t2LDRBs, ARM::tLDRB, 0, 0, 0, 1, 0, 0,0, 1 },
109 { ARM::t2LDRHi12,ARM::tLDRH, 0, 5, 0, 1, 0, 0,0, 1 },
110 { ARM::t2LDRHs, ARM::tLDRH, 0, 0, 0, 1, 0, 0,0, 1 },
111 { ARM::t2LDRSBs,ARM::tLDRSB, 0, 0, 0, 1, 0, 0,0, 1 },
112 { ARM::t2LDRSHs,ARM::tLDRSH, 0, 0, 0, 1, 0, 0,0, 1 },
113 { ARM::t2STRi12,ARM::tSTR, 0, 5, 0, 1, 0, 0,0, 1 },
114 { ARM::t2STRs, ARM::tSTR, 0, 0, 0, 1, 0, 0,0, 1 },
115 { ARM::t2STRBi12,ARM::tSTRB, 0, 5, 0, 1, 0, 0,0, 1 },
116 { ARM::t2STRBs, ARM::tSTRB, 0, 0, 0, 1, 0, 0,0, 1 },
117 { ARM::t2STRHi12,ARM::tSTRH, 0, 5, 0, 1, 0, 0,0, 1 },
118 { ARM::t2STRHs, ARM::tSTRH, 0, 0, 0, 1, 0, 0,0, 1 },
120 { ARM::t2LDM_RET,0, ARM::tPOP_RET, 0, 0, 1, 1, 1,1, 1 },
121 { ARM::t2LDM, ARM::tLDM, ARM::tPOP, 0, 0, 1, 1, 1,1, 1 },
122 { ARM::t2STM, ARM::tSTM, ARM::tPUSH, 0, 0, 1, 1, 1,1, 1 },
125 class VISIBILITY_HIDDEN Thumb2SizeReduce : public MachineFunctionPass {
126 public:
127 static char ID;
128 Thumb2SizeReduce();
130 const TargetInstrInfo *TII;
132 virtual bool runOnMachineFunction(MachineFunction &MF);
134 virtual const char *getPassName() const {
135 return "Thumb2 instruction size reduction pass";
138 private:
139 /// ReduceOpcodeMap - Maps wide opcode to index of entry in ReduceTable.
140 DenseMap<unsigned, unsigned> ReduceOpcodeMap;
142 bool VerifyPredAndCC(MachineInstr *MI, const ReduceEntry &Entry,
143 bool is2Addr, ARMCC::CondCodes Pred,
144 bool LiveCPSR, bool &HasCC, bool &CCDead);
146 bool ReduceLoadStore(MachineBasicBlock &MBB, MachineInstr *MI,
147 const ReduceEntry &Entry);
149 bool ReduceSpecial(MachineBasicBlock &MBB, MachineInstr *MI,
150 const ReduceEntry &Entry, bool LiveCPSR);
152 /// ReduceTo2Addr - Reduce a 32-bit instruction to a 16-bit two-address
153 /// instruction.
154 bool ReduceTo2Addr(MachineBasicBlock &MBB, MachineInstr *MI,
155 const ReduceEntry &Entry,
156 bool LiveCPSR);
158 /// ReduceToNarrow - Reduce a 32-bit instruction to a 16-bit
159 /// non-two-address instruction.
160 bool ReduceToNarrow(MachineBasicBlock &MBB, MachineInstr *MI,
161 const ReduceEntry &Entry,
162 bool LiveCPSR);
164 /// ReduceMBB - Reduce width of instructions in the specified basic block.
165 bool ReduceMBB(MachineBasicBlock &MBB);
167 char Thumb2SizeReduce::ID = 0;
170 Thumb2SizeReduce::Thumb2SizeReduce() : MachineFunctionPass(&ID) {
171 for (unsigned i = 0, e = array_lengthof(ReduceTable); i != e; ++i) {
172 unsigned FromOpc = ReduceTable[i].WideOpc;
173 if (!ReduceOpcodeMap.insert(std::make_pair(FromOpc, i)).second)
174 assert(false && "Duplicated entries?");
178 static bool HasImplicitCPSRDef(const TargetInstrDesc &TID) {
179 for (const unsigned *Regs = TID.ImplicitDefs; *Regs; ++Regs)
180 if (*Regs == ARM::CPSR)
181 return true;
182 return false;
185 bool
186 Thumb2SizeReduce::VerifyPredAndCC(MachineInstr *MI, const ReduceEntry &Entry,
187 bool is2Addr, ARMCC::CondCodes Pred,
188 bool LiveCPSR, bool &HasCC, bool &CCDead) {
189 if ((is2Addr && Entry.PredCC2 == 0) ||
190 (!is2Addr && Entry.PredCC1 == 0)) {
191 if (Pred == ARMCC::AL) {
192 // Not predicated, must set CPSR.
193 if (!HasCC) {
194 // Original instruction was not setting CPSR, but CPSR is not
195 // currently live anyway. It's ok to set it. The CPSR def is
196 // dead though.
197 if (!LiveCPSR) {
198 HasCC = true;
199 CCDead = true;
200 return true;
202 return false;
204 } else {
205 // Predicated, must not set CPSR.
206 if (HasCC)
207 return false;
209 } else if ((is2Addr && Entry.PredCC2 == 2) ||
210 (!is2Addr && Entry.PredCC1 == 2)) {
211 /// Old opcode has an optional def of CPSR.
212 if (HasCC)
213 return true;
214 // If both old opcode does not implicit CPSR def, then it's not ok since
215 // these new opcodes CPSR def is not meant to be thrown away. e.g. CMP.
216 if (!HasImplicitCPSRDef(MI->getDesc()))
217 return false;
218 HasCC = true;
219 } else {
220 // 16-bit instruction does not set CPSR.
221 if (HasCC)
222 return false;
225 return true;
228 static bool VerifyLowRegs(MachineInstr *MI) {
229 unsigned Opc = MI->getOpcode();
230 bool isPCOk = (Opc == ARM::t2LDM_RET) || (Opc == ARM::t2LDM);
231 bool isLROk = (Opc == ARM::t2STM);
232 bool isSPOk = isPCOk || isLROk || (Opc == ARM::t2ADDrSPi);
233 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
234 const MachineOperand &MO = MI->getOperand(i);
235 if (!MO.isReg() || MO.isImplicit())
236 continue;
237 unsigned Reg = MO.getReg();
238 if (Reg == 0 || Reg == ARM::CPSR)
239 continue;
240 if (isPCOk && Reg == ARM::PC)
241 continue;
242 if (isLROk && Reg == ARM::LR)
243 continue;
244 if (isSPOk && Reg == ARM::SP)
245 continue;
246 if (!isARMLowRegister(Reg))
247 return false;
249 return true;
252 bool
253 Thumb2SizeReduce::ReduceLoadStore(MachineBasicBlock &MBB, MachineInstr *MI,
254 const ReduceEntry &Entry) {
255 if (ReduceLimitLdSt != -1 && ((int)NumLdSts >= ReduceLimitLdSt))
256 return false;
258 unsigned Scale = 1;
259 bool HasImmOffset = false;
260 bool HasShift = false;
261 bool isLdStMul = false;
262 bool isPopPush = false;
263 unsigned Opc = Entry.NarrowOpc1;
264 unsigned OpNum = 3; // First 'rest' of operands.
265 switch (Entry.WideOpc) {
266 default:
267 llvm_unreachable("Unexpected Thumb2 load / store opcode!");
268 case ARM::t2LDRi12:
269 case ARM::t2STRi12:
270 Scale = 4;
271 HasImmOffset = true;
272 break;
273 case ARM::t2LDRBi12:
274 case ARM::t2STRBi12:
275 HasImmOffset = true;
276 break;
277 case ARM::t2LDRHi12:
278 case ARM::t2STRHi12:
279 Scale = 2;
280 HasImmOffset = true;
281 break;
282 case ARM::t2LDRs:
283 case ARM::t2LDRBs:
284 case ARM::t2LDRHs:
285 case ARM::t2LDRSBs:
286 case ARM::t2LDRSHs:
287 case ARM::t2STRs:
288 case ARM::t2STRBs:
289 case ARM::t2STRHs:
290 HasShift = true;
291 OpNum = 4;
292 break;
293 case ARM::t2LDM_RET:
294 case ARM::t2LDM:
295 case ARM::t2STM: {
296 OpNum = 0;
297 unsigned BaseReg = MI->getOperand(0).getReg();
298 unsigned Mode = MI->getOperand(1).getImm();
299 if (BaseReg == ARM::SP && ARM_AM::getAM4WBFlag(Mode)) {
300 Opc = Entry.NarrowOpc2;
301 isPopPush = true;
302 OpNum = 2;
303 } else if (Entry.WideOpc == ARM::t2LDM_RET ||
304 !isARMLowRegister(BaseReg) ||
305 !ARM_AM::getAM4WBFlag(Mode) ||
306 ARM_AM::getAM4SubMode(Mode) != ARM_AM::ia) {
307 return false;
309 isLdStMul = true;
310 break;
314 unsigned OffsetReg = 0;
315 bool OffsetKill = false;
316 if (HasShift) {
317 OffsetReg = MI->getOperand(2).getReg();
318 OffsetKill = MI->getOperand(2).isKill();
319 if (MI->getOperand(3).getImm())
320 // Thumb1 addressing mode doesn't support shift.
321 return false;
324 unsigned OffsetImm = 0;
325 if (HasImmOffset) {
326 OffsetImm = MI->getOperand(2).getImm();
327 unsigned MaxOffset = ((1 << Entry.Imm1Limit) - 1) * Scale;
328 if ((OffsetImm & (Scale-1)) || OffsetImm > MaxOffset)
329 // Make sure the immediate field fits.
330 return false;
333 // Add the 16-bit load / store instruction.
334 // FIXME: Thumb1 addressing mode encode both immediate and register offset.
335 DebugLoc dl = MI->getDebugLoc();
336 MachineInstrBuilder MIB = BuildMI(MBB, *MI, dl, TII->get(Opc));
337 if (!isLdStMul) {
338 MIB.addOperand(MI->getOperand(0)).addOperand(MI->getOperand(1));
339 if (Entry.NarrowOpc1 != ARM::tLDRSB && Entry.NarrowOpc1 != ARM::tLDRSH) {
340 // tLDRSB and tLDRSH do not have an immediate offset field. On the other
341 // hand, it must have an offset register.
342 // FIXME: Remove this special case.
343 MIB.addImm(OffsetImm/Scale);
345 assert((!HasShift || OffsetReg) && "Invalid so_reg load / store address!");
347 MIB.addReg(OffsetReg, getKillRegState(OffsetKill));
350 // Transfer the rest of operands.
351 for (unsigned e = MI->getNumOperands(); OpNum != e; ++OpNum)
352 MIB.addOperand(MI->getOperand(OpNum));
354 DOUT << "Converted 32-bit: " << *MI << " to 16-bit: " << *MIB;
356 MBB.erase(MI);
357 ++NumLdSts;
358 return true;
361 bool
362 Thumb2SizeReduce::ReduceSpecial(MachineBasicBlock &MBB, MachineInstr *MI,
363 const ReduceEntry &Entry,
364 bool LiveCPSR) {
365 if (Entry.LowRegs1 && !VerifyLowRegs(MI))
366 return false;
368 const TargetInstrDesc &TID = MI->getDesc();
369 if (TID.mayLoad() || TID.mayStore())
370 return ReduceLoadStore(MBB, MI, Entry);
372 unsigned Opc = MI->getOpcode();
373 switch (Opc) {
374 default: break;
375 case ARM::t2ADDSri:
376 case ARM::t2ADDSrr: {
377 unsigned PredReg = 0;
378 if (getInstrPredicate(MI, PredReg) == ARMCC::AL) {
379 switch (Opc) {
380 default: break;
381 case ARM::t2ADDSri: {
382 if (ReduceTo2Addr(MBB, MI, Entry, LiveCPSR))
383 return true;
384 // fallthrough
386 case ARM::t2ADDSrr:
387 return ReduceToNarrow(MBB, MI, Entry, LiveCPSR);
390 break;
392 case ARM::t2RSBri:
393 case ARM::t2RSBSri:
394 if (MI->getOperand(2).getImm() == 0)
395 return ReduceToNarrow(MBB, MI, Entry, LiveCPSR);
396 break;
398 return false;
401 bool
402 Thumb2SizeReduce::ReduceTo2Addr(MachineBasicBlock &MBB, MachineInstr *MI,
403 const ReduceEntry &Entry,
404 bool LiveCPSR) {
406 if (ReduceLimit2Addr != -1 && ((int)Num2Addrs >= ReduceLimit2Addr))
407 return false;
409 const TargetInstrDesc &TID = MI->getDesc();
410 unsigned Reg0 = MI->getOperand(0).getReg();
411 unsigned Reg1 = MI->getOperand(1).getReg();
412 if (Reg0 != Reg1)
413 return false;
414 if (Entry.LowRegs2 && !isARMLowRegister(Reg0))
415 return false;
416 if (Entry.Imm2Limit) {
417 unsigned Imm = MI->getOperand(2).getImm();
418 unsigned Limit = (1 << Entry.Imm2Limit) - 1;
419 if (Imm > Limit)
420 return false;
421 } else {
422 unsigned Reg2 = MI->getOperand(2).getReg();
423 if (Entry.LowRegs2 && !isARMLowRegister(Reg2))
424 return false;
427 // Check if it's possible / necessary to transfer the predicate.
428 const TargetInstrDesc &NewTID = TII->get(Entry.NarrowOpc2);
429 unsigned PredReg = 0;
430 ARMCC::CondCodes Pred = getInstrPredicate(MI, PredReg);
431 bool SkipPred = false;
432 if (Pred != ARMCC::AL) {
433 if (!NewTID.isPredicable())
434 // Can't transfer predicate, fail.
435 return false;
436 } else {
437 SkipPred = !NewTID.isPredicable();
440 bool HasCC = false;
441 bool CCDead = false;
442 if (TID.hasOptionalDef()) {
443 unsigned NumOps = TID.getNumOperands();
444 HasCC = (MI->getOperand(NumOps-1).getReg() == ARM::CPSR);
445 if (HasCC && MI->getOperand(NumOps-1).isDead())
446 CCDead = true;
448 if (!VerifyPredAndCC(MI, Entry, true, Pred, LiveCPSR, HasCC, CCDead))
449 return false;
451 // Add the 16-bit instruction.
452 DebugLoc dl = MI->getDebugLoc();
453 MachineInstrBuilder MIB = BuildMI(MBB, *MI, dl, NewTID);
454 MIB.addOperand(MI->getOperand(0));
455 if (HasCC && NewTID.hasOptionalDef())
456 AddDefaultT1CC(MIB, CCDead);
458 // Transfer the rest of operands.
459 unsigned NumOps = TID.getNumOperands();
460 for (unsigned i = 1, e = MI->getNumOperands(); i != e; ++i) {
461 if (i < NumOps && TID.OpInfo[i].isOptionalDef())
462 continue;
463 if (SkipPred && TID.OpInfo[i].isPredicate())
464 continue;
465 MIB.addOperand(MI->getOperand(i));
468 DOUT << "Converted 32-bit: " << *MI << " to 16-bit: " << *MIB;
470 MBB.erase(MI);
471 ++Num2Addrs;
472 return true;
475 bool
476 Thumb2SizeReduce::ReduceToNarrow(MachineBasicBlock &MBB, MachineInstr *MI,
477 const ReduceEntry &Entry,
478 bool LiveCPSR) {
479 if (ReduceLimit != -1 && ((int)NumNarrows >= ReduceLimit))
480 return false;
482 unsigned Limit = ~0U;
483 unsigned Scale = (Entry.WideOpc == ARM::t2ADDrSPi) ? 4 : 1;
484 if (Entry.Imm1Limit)
485 Limit = ((1 << Entry.Imm1Limit) - 1) * Scale;
487 const TargetInstrDesc &TID = MI->getDesc();
488 for (unsigned i = 0, e = TID.getNumOperands(); i != e; ++i) {
489 if (TID.OpInfo[i].isPredicate())
490 continue;
491 const MachineOperand &MO = MI->getOperand(i);
492 if (MO.isReg()) {
493 unsigned Reg = MO.getReg();
494 if (!Reg || Reg == ARM::CPSR)
495 continue;
496 if (Entry.WideOpc == ARM::t2ADDrSPi && Reg == ARM::SP)
497 continue;
498 if (Entry.LowRegs1 && !isARMLowRegister(Reg))
499 return false;
500 } else if (MO.isImm() &&
501 !TID.OpInfo[i].isPredicate()) {
502 if (MO.getImm() > Limit || (MO.getImm() & (Scale-1)) != 0)
503 return false;
507 // Check if it's possible / necessary to transfer the predicate.
508 const TargetInstrDesc &NewTID = TII->get(Entry.NarrowOpc1);
509 unsigned PredReg = 0;
510 ARMCC::CondCodes Pred = getInstrPredicate(MI, PredReg);
511 bool SkipPred = false;
512 if (Pred != ARMCC::AL) {
513 if (!NewTID.isPredicable())
514 // Can't transfer predicate, fail.
515 return false;
516 } else {
517 SkipPred = !NewTID.isPredicable();
520 bool HasCC = false;
521 bool CCDead = false;
522 if (TID.hasOptionalDef()) {
523 unsigned NumOps = TID.getNumOperands();
524 HasCC = (MI->getOperand(NumOps-1).getReg() == ARM::CPSR);
525 if (HasCC && MI->getOperand(NumOps-1).isDead())
526 CCDead = true;
528 if (!VerifyPredAndCC(MI, Entry, false, Pred, LiveCPSR, HasCC, CCDead))
529 return false;
531 // Add the 16-bit instruction.
532 DebugLoc dl = MI->getDebugLoc();
533 MachineInstrBuilder MIB = BuildMI(MBB, *MI, dl, NewTID);
534 MIB.addOperand(MI->getOperand(0));
535 if (HasCC && NewTID.hasOptionalDef())
536 AddDefaultT1CC(MIB, CCDead);
538 // Transfer the rest of operands.
539 unsigned NumOps = TID.getNumOperands();
540 for (unsigned i = 1, e = MI->getNumOperands(); i != e; ++i) {
541 if (i < NumOps && TID.OpInfo[i].isOptionalDef())
542 continue;
543 if ((TID.getOpcode() == ARM::t2RSBSri ||
544 TID.getOpcode() == ARM::t2RSBri) && i == 2)
545 // Skip the zero immediate operand, it's now implicit.
546 continue;
547 bool isPred = (i < NumOps && TID.OpInfo[i].isPredicate());
548 if (SkipPred && isPred)
549 continue;
550 const MachineOperand &MO = MI->getOperand(i);
551 if (Scale > 1 && !isPred && MO.isImm())
552 MIB.addImm(MO.getImm() / Scale);
553 else {
554 if (MO.isReg() && MO.isImplicit() && MO.getReg() == ARM::CPSR)
555 // Skip implicit def of CPSR. Either it's modeled as an optional
556 // def now or it's already an implicit def on the new instruction.
557 continue;
558 MIB.addOperand(MO);
561 if (!TID.isPredicable() && NewTID.isPredicable())
562 AddDefaultPred(MIB);
564 DOUT << "Converted 32-bit: " << *MI << " to 16-bit: " << *MIB;
566 MBB.erase(MI);
567 ++NumNarrows;
568 return true;
571 static bool UpdateCPSRDef(MachineInstr &MI, bool LiveCPSR) {
572 bool HasDef = false;
573 for (unsigned i = 0, e = MI.getNumOperands(); i != e; ++i) {
574 const MachineOperand &MO = MI.getOperand(i);
575 if (!MO.isReg() || MO.isUndef() || MO.isUse())
576 continue;
577 if (MO.getReg() != ARM::CPSR)
578 continue;
579 if (!MO.isDead())
580 HasDef = true;
583 return HasDef || LiveCPSR;
586 static bool UpdateCPSRUse(MachineInstr &MI, bool LiveCPSR) {
587 for (unsigned i = 0, e = MI.getNumOperands(); i != e; ++i) {
588 const MachineOperand &MO = MI.getOperand(i);
589 if (!MO.isReg() || MO.isUndef() || MO.isDef())
590 continue;
591 if (MO.getReg() != ARM::CPSR)
592 continue;
593 assert(LiveCPSR && "CPSR liveness tracking is wrong!");
594 if (MO.isKill()) {
595 LiveCPSR = false;
596 break;
600 return LiveCPSR;
603 bool Thumb2SizeReduce::ReduceMBB(MachineBasicBlock &MBB) {
604 bool Modified = false;
606 bool LiveCPSR = false;
607 // Yes, CPSR could be livein.
608 for (MachineBasicBlock::const_livein_iterator I = MBB.livein_begin(),
609 E = MBB.livein_end(); I != E; ++I) {
610 if (*I == ARM::CPSR) {
611 LiveCPSR = true;
612 break;
616 MachineBasicBlock::iterator MII = MBB.begin(), E = MBB.end();
617 MachineBasicBlock::iterator NextMII;
618 for (; MII != E; MII = NextMII) {
619 NextMII = next(MII);
621 MachineInstr *MI = &*MII;
622 LiveCPSR = UpdateCPSRUse(*MI, LiveCPSR);
624 unsigned Opcode = MI->getOpcode();
625 DenseMap<unsigned, unsigned>::iterator OPI = ReduceOpcodeMap.find(Opcode);
626 if (OPI != ReduceOpcodeMap.end()) {
627 const ReduceEntry &Entry = ReduceTable[OPI->second];
628 // Ignore "special" cases for now.
629 if (Entry.Special) {
630 if (ReduceSpecial(MBB, MI, Entry, LiveCPSR)) {
631 Modified = true;
632 MachineBasicBlock::iterator I = prior(NextMII);
633 MI = &*I;
635 goto ProcessNext;
638 // Try to transform to a 16-bit two-address instruction.
639 if (Entry.NarrowOpc2 && ReduceTo2Addr(MBB, MI, Entry, LiveCPSR)) {
640 Modified = true;
641 MachineBasicBlock::iterator I = prior(NextMII);
642 MI = &*I;
643 goto ProcessNext;
646 // Try to transform ro a 16-bit non-two-address instruction.
647 if (Entry.NarrowOpc1 && ReduceToNarrow(MBB, MI, Entry, LiveCPSR))
648 Modified = true;
651 ProcessNext:
652 LiveCPSR = UpdateCPSRDef(*MI, LiveCPSR);
655 return Modified;
658 bool Thumb2SizeReduce::runOnMachineFunction(MachineFunction &MF) {
659 const TargetMachine &TM = MF.getTarget();
660 TII = TM.getInstrInfo();
662 bool Modified = false;
663 for (MachineFunction::iterator I = MF.begin(), E = MF.end(); I != E; ++I)
664 Modified |= ReduceMBB(*I);
665 return Modified;
668 /// createThumb2SizeReductionPass - Returns an instance of the Thumb2 size
669 /// reduction pass.
670 FunctionPass *llvm::createThumb2SizeReductionPass() {
671 return new Thumb2SizeReduce();