1 | //===- MipsFastISel.cpp - Mips FastISel implementation --------------------===// |
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 | /// \file |
10 | /// This file defines the MIPS-specific support for the FastISel class. |
11 | /// Some of the target-specific code is generated by tablegen in the file |
12 | /// MipsGenFastISel.inc, which is #included here. |
13 | /// |
14 | //===----------------------------------------------------------------------===// |
15 | |
16 | #include "MCTargetDesc/MipsABIInfo.h" |
17 | #include "MCTargetDesc/MipsBaseInfo.h" |
18 | #include "MipsCCState.h" |
19 | #include "MipsISelLowering.h" |
20 | #include "MipsInstrInfo.h" |
21 | #include "MipsMachineFunction.h" |
22 | #include "MipsSubtarget.h" |
23 | #include "MipsTargetMachine.h" |
24 | #include "llvm/ADT/APInt.h" |
25 | #include "llvm/ADT/ArrayRef.h" |
26 | #include "llvm/ADT/DenseMap.h" |
27 | #include "llvm/ADT/SmallVector.h" |
28 | #include "llvm/Analysis/TargetLibraryInfo.h" |
29 | #include "llvm/CodeGen/CallingConvLower.h" |
30 | #include "llvm/CodeGen/FastISel.h" |
31 | #include "llvm/CodeGen/FunctionLoweringInfo.h" |
32 | #include "llvm/CodeGen/ISDOpcodes.h" |
33 | #include "llvm/CodeGen/MachineBasicBlock.h" |
34 | #include "llvm/CodeGen/MachineFrameInfo.h" |
35 | #include "llvm/CodeGen/MachineInstrBuilder.h" |
36 | #include "llvm/CodeGen/MachineMemOperand.h" |
37 | #include "llvm/CodeGen/MachineRegisterInfo.h" |
38 | #include "llvm/CodeGen/TargetInstrInfo.h" |
39 | #include "llvm/CodeGen/TargetLowering.h" |
40 | #include "llvm/CodeGen/ValueTypes.h" |
41 | #include "llvm/CodeGenTypes/MachineValueType.h" |
42 | #include "llvm/IR/Attributes.h" |
43 | #include "llvm/IR/CallingConv.h" |
44 | #include "llvm/IR/Constant.h" |
45 | #include "llvm/IR/Constants.h" |
46 | #include "llvm/IR/DataLayout.h" |
47 | #include "llvm/IR/Function.h" |
48 | #include "llvm/IR/GetElementPtrTypeIterator.h" |
49 | #include "llvm/IR/GlobalValue.h" |
50 | #include "llvm/IR/GlobalVariable.h" |
51 | #include "llvm/IR/InstrTypes.h" |
52 | #include "llvm/IR/Instruction.h" |
53 | #include "llvm/IR/Instructions.h" |
54 | #include "llvm/IR/IntrinsicInst.h" |
55 | #include "llvm/IR/Operator.h" |
56 | #include "llvm/IR/Type.h" |
57 | #include "llvm/IR/User.h" |
58 | #include "llvm/IR/Value.h" |
59 | #include "llvm/MC/MCContext.h" |
60 | #include "llvm/MC/MCInstrDesc.h" |
61 | #include "llvm/MC/MCRegisterInfo.h" |
62 | #include "llvm/MC/MCSymbol.h" |
63 | #include "llvm/Support/Casting.h" |
64 | #include "llvm/Support/Compiler.h" |
65 | #include "llvm/Support/Debug.h" |
66 | #include "llvm/Support/ErrorHandling.h" |
67 | #include "llvm/Support/MathExtras.h" |
68 | #include "llvm/Support/raw_ostream.h" |
69 | #include <algorithm> |
70 | #include <array> |
71 | #include <cassert> |
72 | #include <cstdint> |
73 | |
74 | #define DEBUG_TYPE "mips-fastisel" |
75 | |
76 | using namespace llvm; |
77 | |
78 | extern cl::opt<bool> EmitJalrReloc; |
79 | |
80 | namespace { |
81 | |
82 | class MipsFastISel final : public FastISel { |
83 | |
84 | // All possible address modes. |
85 | class Address { |
86 | public: |
87 | using BaseKind = enum { RegBase, FrameIndexBase }; |
88 | |
89 | private: |
90 | BaseKind Kind = RegBase; |
91 | union { |
92 | unsigned Reg; |
93 | int FI; |
94 | } Base; |
95 | |
96 | int64_t Offset = 0; |
97 | |
98 | const GlobalValue *GV = nullptr; |
99 | |
100 | public: |
101 | // Innocuous defaults for our address. |
102 | Address() { Base.Reg = 0; } |
103 | |
104 | void setKind(BaseKind K) { Kind = K; } |
105 | BaseKind getKind() const { return Kind; } |
106 | bool isRegBase() const { return Kind == RegBase; } |
107 | bool isFIBase() const { return Kind == FrameIndexBase; } |
108 | |
109 | void setReg(unsigned Reg) { |
110 | assert(isRegBase() && "Invalid base register access!" ); |
111 | Base.Reg = Reg; |
112 | } |
113 | |
114 | unsigned getReg() const { |
115 | assert(isRegBase() && "Invalid base register access!" ); |
116 | return Base.Reg; |
117 | } |
118 | |
119 | void setFI(unsigned FI) { |
120 | assert(isFIBase() && "Invalid base frame index access!" ); |
121 | Base.FI = FI; |
122 | } |
123 | |
124 | unsigned getFI() const { |
125 | assert(isFIBase() && "Invalid base frame index access!" ); |
126 | return Base.FI; |
127 | } |
128 | |
129 | void setOffset(int64_t Offset_) { Offset = Offset_; } |
130 | int64_t getOffset() const { return Offset; } |
131 | void setGlobalValue(const GlobalValue *G) { GV = G; } |
132 | const GlobalValue *getGlobalValue() { return GV; } |
133 | }; |
134 | |
135 | /// Subtarget - Keep a pointer to the MipsSubtarget around so that we can |
136 | /// make the right decision when generating code for different targets. |
137 | const TargetMachine &TM; |
138 | const MipsSubtarget *Subtarget; |
139 | const TargetInstrInfo &TII; |
140 | const TargetLowering &TLI; |
141 | MipsFunctionInfo *MFI; |
142 | |
143 | // Convenience variables to avoid some queries. |
144 | LLVMContext *Context; |
145 | |
146 | bool fastLowerArguments() override; |
147 | bool fastLowerCall(CallLoweringInfo &CLI) override; |
148 | bool fastLowerIntrinsicCall(const IntrinsicInst *II) override; |
149 | |
150 | bool UnsupportedFPMode; // To allow fast-isel to proceed and just not handle |
151 | // floating point but not reject doing fast-isel in other |
152 | // situations |
153 | |
154 | private: |
155 | // Selection routines. |
156 | bool selectLogicalOp(const Instruction *I); |
157 | bool selectLoad(const Instruction *I); |
158 | bool selectStore(const Instruction *I); |
159 | bool selectBranch(const Instruction *I); |
160 | bool selectSelect(const Instruction *I); |
161 | bool selectCmp(const Instruction *I); |
162 | bool selectFPExt(const Instruction *I); |
163 | bool selectFPTrunc(const Instruction *I); |
164 | bool selectFPToInt(const Instruction *I, bool IsSigned); |
165 | bool selectRet(const Instruction *I); |
166 | bool selectTrunc(const Instruction *I); |
167 | bool selectIntExt(const Instruction *I); |
168 | bool selectShift(const Instruction *I); |
169 | bool selectDivRem(const Instruction *I, unsigned ISDOpcode); |
170 | |
171 | // Utility helper routines. |
172 | bool isTypeLegal(Type *Ty, MVT &VT); |
173 | bool isTypeSupported(Type *Ty, MVT &VT); |
174 | bool isLoadTypeLegal(Type *Ty, MVT &VT); |
175 | bool computeAddress(const Value *Obj, Address &Addr); |
176 | bool computeCallAddress(const Value *V, Address &Addr); |
177 | void simplifyAddress(Address &Addr); |
178 | |
179 | // Emit helper routines. |
180 | bool emitCmp(unsigned DestReg, const CmpInst *CI); |
181 | bool emitLoad(MVT VT, unsigned &ResultReg, Address &Addr); |
182 | bool emitStore(MVT VT, unsigned SrcReg, Address &Addr); |
183 | unsigned emitIntExt(MVT SrcVT, unsigned SrcReg, MVT DestVT, bool isZExt); |
184 | bool emitIntExt(MVT SrcVT, unsigned SrcReg, MVT DestVT, unsigned DestReg, |
185 | |
186 | bool IsZExt); |
187 | bool emitIntZExt(MVT SrcVT, unsigned SrcReg, MVT DestVT, unsigned DestReg); |
188 | |
189 | bool emitIntSExt(MVT SrcVT, unsigned SrcReg, MVT DestVT, unsigned DestReg); |
190 | bool emitIntSExt32r1(MVT SrcVT, unsigned SrcReg, MVT DestVT, |
191 | unsigned DestReg); |
192 | bool emitIntSExt32r2(MVT SrcVT, unsigned SrcReg, MVT DestVT, |
193 | unsigned DestReg); |
194 | |
195 | unsigned getRegEnsuringSimpleIntegerWidening(const Value *, bool IsUnsigned); |
196 | |
197 | unsigned emitLogicalOp(unsigned ISDOpc, MVT RetVT, const Value *LHS, |
198 | const Value *RHS); |
199 | |
200 | unsigned materializeFP(const ConstantFP *CFP, MVT VT); |
201 | unsigned materializeGV(const GlobalValue *GV, MVT VT); |
202 | unsigned materializeInt(const Constant *C, MVT VT); |
203 | unsigned materialize32BitInt(int64_t Imm, const TargetRegisterClass *RC); |
204 | unsigned materializeExternalCallSym(MCSymbol *Syn); |
205 | |
206 | MachineInstrBuilder emitInst(unsigned Opc) { |
207 | return BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, MCID: TII.get(Opcode: Opc)); |
208 | } |
209 | |
210 | MachineInstrBuilder emitInst(unsigned Opc, unsigned DstReg) { |
211 | return BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, MCID: TII.get(Opcode: Opc), |
212 | DestReg: DstReg); |
213 | } |
214 | |
215 | MachineInstrBuilder emitInstStore(unsigned Opc, unsigned SrcReg, |
216 | unsigned MemReg, int64_t MemOffset) { |
217 | return emitInst(Opc).addReg(RegNo: SrcReg).addReg(RegNo: MemReg).addImm(Val: MemOffset); |
218 | } |
219 | |
220 | MachineInstrBuilder emitInstLoad(unsigned Opc, unsigned DstReg, |
221 | unsigned MemReg, int64_t MemOffset) { |
222 | return emitInst(Opc, DstReg).addReg(RegNo: MemReg).addImm(Val: MemOffset); |
223 | } |
224 | |
225 | unsigned fastEmitInst_rr(unsigned MachineInstOpcode, |
226 | const TargetRegisterClass *RC, |
227 | unsigned Op0, unsigned Op1); |
228 | |
229 | // for some reason, this default is not generated by tablegen |
230 | // so we explicitly generate it here. |
231 | unsigned fastEmitInst_riir(uint64_t inst, const TargetRegisterClass *RC, |
232 | unsigned Op0, uint64_t imm1, uint64_t imm2, |
233 | unsigned Op3) { |
234 | return 0; |
235 | } |
236 | |
237 | // Call handling routines. |
238 | private: |
239 | CCAssignFn *CCAssignFnForCall(CallingConv::ID CC) const; |
240 | bool processCallArgs(CallLoweringInfo &CLI, SmallVectorImpl<MVT> &ArgVTs, |
241 | unsigned &NumBytes); |
242 | bool finishCall(CallLoweringInfo &CLI, MVT RetVT, unsigned NumBytes); |
243 | |
244 | const MipsABIInfo &getABI() const { |
245 | return static_cast<const MipsTargetMachine &>(TM).getABI(); |
246 | } |
247 | |
248 | public: |
249 | // Backend specific FastISel code. |
250 | explicit MipsFastISel(FunctionLoweringInfo &funcInfo, |
251 | const TargetLibraryInfo *libInfo) |
252 | : FastISel(funcInfo, libInfo), TM(funcInfo.MF->getTarget()), |
253 | Subtarget(&funcInfo.MF->getSubtarget<MipsSubtarget>()), |
254 | TII(*Subtarget->getInstrInfo()), TLI(*Subtarget->getTargetLowering()) { |
255 | MFI = funcInfo.MF->getInfo<MipsFunctionInfo>(); |
256 | Context = &funcInfo.Fn->getContext(); |
257 | UnsupportedFPMode = Subtarget->isFP64bit() || Subtarget->useSoftFloat(); |
258 | } |
259 | |
260 | unsigned fastMaterializeAlloca(const AllocaInst *AI) override; |
261 | unsigned fastMaterializeConstant(const Constant *C) override; |
262 | bool fastSelectInstruction(const Instruction *I) override; |
263 | |
264 | #include "MipsGenFastISel.inc" |
265 | }; |
266 | |
267 | } // end anonymous namespace |
268 | |
269 | static bool CC_Mips(unsigned ValNo, MVT ValVT, MVT LocVT, |
270 | CCValAssign::LocInfo LocInfo, ISD::ArgFlagsTy ArgFlags, |
271 | CCState &State) LLVM_ATTRIBUTE_UNUSED; |
272 | |
273 | static bool CC_MipsO32_FP32(unsigned ValNo, MVT ValVT, MVT LocVT, |
274 | CCValAssign::LocInfo LocInfo, |
275 | ISD::ArgFlagsTy ArgFlags, CCState &State) { |
276 | llvm_unreachable("should not be called" ); |
277 | } |
278 | |
279 | static bool CC_MipsO32_FP64(unsigned ValNo, MVT ValVT, MVT LocVT, |
280 | CCValAssign::LocInfo LocInfo, |
281 | ISD::ArgFlagsTy ArgFlags, CCState &State) { |
282 | llvm_unreachable("should not be called" ); |
283 | } |
284 | |
285 | #include "MipsGenCallingConv.inc" |
286 | |
287 | CCAssignFn *MipsFastISel::CCAssignFnForCall(CallingConv::ID CC) const { |
288 | return CC_MipsO32; |
289 | } |
290 | |
291 | unsigned MipsFastISel::emitLogicalOp(unsigned ISDOpc, MVT RetVT, |
292 | const Value *LHS, const Value *RHS) { |
293 | // Canonicalize immediates to the RHS first. |
294 | if (isa<ConstantInt>(Val: LHS) && !isa<ConstantInt>(Val: RHS)) |
295 | std::swap(a&: LHS, b&: RHS); |
296 | |
297 | unsigned Opc; |
298 | switch (ISDOpc) { |
299 | case ISD::AND: |
300 | Opc = Mips::AND; |
301 | break; |
302 | case ISD::OR: |
303 | Opc = Mips::OR; |
304 | break; |
305 | case ISD::XOR: |
306 | Opc = Mips::XOR; |
307 | break; |
308 | default: |
309 | llvm_unreachable("unexpected opcode" ); |
310 | } |
311 | |
312 | Register LHSReg = getRegForValue(V: LHS); |
313 | if (!LHSReg) |
314 | return 0; |
315 | |
316 | unsigned RHSReg; |
317 | if (const auto *C = dyn_cast<ConstantInt>(Val: RHS)) |
318 | RHSReg = materializeInt(C, VT: MVT::i32); |
319 | else |
320 | RHSReg = getRegForValue(V: RHS); |
321 | if (!RHSReg) |
322 | return 0; |
323 | |
324 | Register ResultReg = createResultReg(RC: &Mips::GPR32RegClass); |
325 | if (!ResultReg) |
326 | return 0; |
327 | |
328 | emitInst(Opc, DstReg: ResultReg).addReg(RegNo: LHSReg).addReg(RegNo: RHSReg); |
329 | return ResultReg; |
330 | } |
331 | |
332 | unsigned MipsFastISel::fastMaterializeAlloca(const AllocaInst *AI) { |
333 | assert(TLI.getValueType(DL, AI->getType(), true) == MVT::i32 && |
334 | "Alloca should always return a pointer." ); |
335 | |
336 | DenseMap<const AllocaInst *, int>::iterator SI = |
337 | FuncInfo.StaticAllocaMap.find(Val: AI); |
338 | |
339 | if (SI != FuncInfo.StaticAllocaMap.end()) { |
340 | Register ResultReg = createResultReg(RC: &Mips::GPR32RegClass); |
341 | BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, MCID: TII.get(Opcode: Mips::LEA_ADDiu), |
342 | DestReg: ResultReg) |
343 | .addFrameIndex(Idx: SI->second) |
344 | .addImm(Val: 0); |
345 | return ResultReg; |
346 | } |
347 | |
348 | return 0; |
349 | } |
350 | |
351 | unsigned MipsFastISel::materializeInt(const Constant *C, MVT VT) { |
352 | if (VT != MVT::i32 && VT != MVT::i16 && VT != MVT::i8 && VT != MVT::i1) |
353 | return 0; |
354 | const TargetRegisterClass *RC = &Mips::GPR32RegClass; |
355 | const ConstantInt *CI = cast<ConstantInt>(Val: C); |
356 | return materialize32BitInt(Imm: CI->getZExtValue(), RC); |
357 | } |
358 | |
359 | unsigned MipsFastISel::materialize32BitInt(int64_t Imm, |
360 | const TargetRegisterClass *RC) { |
361 | Register ResultReg = createResultReg(RC); |
362 | |
363 | if (isInt<16>(x: Imm)) { |
364 | unsigned Opc = Mips::ADDiu; |
365 | emitInst(Opc, DstReg: ResultReg).addReg(RegNo: Mips::ZERO).addImm(Val: Imm); |
366 | return ResultReg; |
367 | } else if (isUInt<16>(x: Imm)) { |
368 | emitInst(Opc: Mips::ORi, DstReg: ResultReg).addReg(RegNo: Mips::ZERO).addImm(Val: Imm); |
369 | return ResultReg; |
370 | } |
371 | unsigned Lo = Imm & 0xFFFF; |
372 | unsigned Hi = (Imm >> 16) & 0xFFFF; |
373 | if (Lo) { |
374 | // Both Lo and Hi have nonzero bits. |
375 | Register TmpReg = createResultReg(RC); |
376 | emitInst(Opc: Mips::LUi, DstReg: TmpReg).addImm(Val: Hi); |
377 | emitInst(Opc: Mips::ORi, DstReg: ResultReg).addReg(RegNo: TmpReg).addImm(Val: Lo); |
378 | } else { |
379 | emitInst(Opc: Mips::LUi, DstReg: ResultReg).addImm(Val: Hi); |
380 | } |
381 | return ResultReg; |
382 | } |
383 | |
384 | unsigned MipsFastISel::materializeFP(const ConstantFP *CFP, MVT VT) { |
385 | if (UnsupportedFPMode) |
386 | return 0; |
387 | int64_t Imm = CFP->getValueAPF().bitcastToAPInt().getZExtValue(); |
388 | if (VT == MVT::f32) { |
389 | const TargetRegisterClass *RC = &Mips::FGR32RegClass; |
390 | Register DestReg = createResultReg(RC); |
391 | unsigned TempReg = materialize32BitInt(Imm, RC: &Mips::GPR32RegClass); |
392 | emitInst(Opc: Mips::MTC1, DstReg: DestReg).addReg(RegNo: TempReg); |
393 | return DestReg; |
394 | } else if (VT == MVT::f64) { |
395 | const TargetRegisterClass *RC = &Mips::AFGR64RegClass; |
396 | Register DestReg = createResultReg(RC); |
397 | unsigned TempReg1 = materialize32BitInt(Imm: Imm >> 32, RC: &Mips::GPR32RegClass); |
398 | unsigned TempReg2 = |
399 | materialize32BitInt(Imm: Imm & 0xFFFFFFFF, RC: &Mips::GPR32RegClass); |
400 | emitInst(Opc: Mips::BuildPairF64, DstReg: DestReg).addReg(RegNo: TempReg2).addReg(RegNo: TempReg1); |
401 | return DestReg; |
402 | } |
403 | return 0; |
404 | } |
405 | |
406 | unsigned MipsFastISel::materializeGV(const GlobalValue *GV, MVT VT) { |
407 | // For now 32-bit only. |
408 | if (VT != MVT::i32) |
409 | return 0; |
410 | const TargetRegisterClass *RC = &Mips::GPR32RegClass; |
411 | Register DestReg = createResultReg(RC); |
412 | const GlobalVariable *GVar = dyn_cast<GlobalVariable>(Val: GV); |
413 | bool IsThreadLocal = GVar && GVar->isThreadLocal(); |
414 | // TLS not supported at this time. |
415 | if (IsThreadLocal) |
416 | return 0; |
417 | emitInst(Opc: Mips::LW, DstReg: DestReg) |
418 | .addReg(RegNo: MFI->getGlobalBaseReg(MF&: *MF)) |
419 | .addGlobalAddress(GV, Offset: 0, TargetFlags: MipsII::MO_GOT); |
420 | if ((GV->hasInternalLinkage() || |
421 | (GV->hasLocalLinkage() && !isa<Function>(Val: GV)))) { |
422 | Register TempReg = createResultReg(RC); |
423 | emitInst(Opc: Mips::ADDiu, DstReg: TempReg) |
424 | .addReg(RegNo: DestReg) |
425 | .addGlobalAddress(GV, Offset: 0, TargetFlags: MipsII::MO_ABS_LO); |
426 | DestReg = TempReg; |
427 | } |
428 | return DestReg; |
429 | } |
430 | |
431 | unsigned MipsFastISel::materializeExternalCallSym(MCSymbol *Sym) { |
432 | const TargetRegisterClass *RC = &Mips::GPR32RegClass; |
433 | Register DestReg = createResultReg(RC); |
434 | emitInst(Opc: Mips::LW, DstReg: DestReg) |
435 | .addReg(RegNo: MFI->getGlobalBaseReg(MF&: *MF)) |
436 | .addSym(Sym, TargetFlags: MipsII::MO_GOT); |
437 | return DestReg; |
438 | } |
439 | |
440 | // Materialize a constant into a register, and return the register |
441 | // number (or zero if we failed to handle it). |
442 | unsigned MipsFastISel::fastMaterializeConstant(const Constant *C) { |
443 | EVT CEVT = TLI.getValueType(DL, Ty: C->getType(), AllowUnknown: true); |
444 | |
445 | // Only handle simple types. |
446 | if (!CEVT.isSimple()) |
447 | return 0; |
448 | MVT VT = CEVT.getSimpleVT(); |
449 | |
450 | if (const ConstantFP *CFP = dyn_cast<ConstantFP>(Val: C)) |
451 | return (UnsupportedFPMode) ? 0 : materializeFP(CFP, VT); |
452 | else if (const GlobalValue *GV = dyn_cast<GlobalValue>(Val: C)) |
453 | return materializeGV(GV, VT); |
454 | else if (isa<ConstantInt>(Val: C)) |
455 | return materializeInt(C, VT); |
456 | |
457 | return 0; |
458 | } |
459 | |
460 | bool MipsFastISel::computeAddress(const Value *Obj, Address &Addr) { |
461 | const User *U = nullptr; |
462 | unsigned Opcode = Instruction::UserOp1; |
463 | if (const Instruction *I = dyn_cast<Instruction>(Val: Obj)) { |
464 | // Don't walk into other basic blocks unless the object is an alloca from |
465 | // another block, otherwise it may not have a virtual register assigned. |
466 | if (FuncInfo.StaticAllocaMap.count(Val: static_cast<const AllocaInst *>(Obj)) || |
467 | FuncInfo.MBBMap[I->getParent()] == FuncInfo.MBB) { |
468 | Opcode = I->getOpcode(); |
469 | U = I; |
470 | } |
471 | } else if (const ConstantExpr *C = dyn_cast<ConstantExpr>(Val: Obj)) { |
472 | Opcode = C->getOpcode(); |
473 | U = C; |
474 | } |
475 | switch (Opcode) { |
476 | default: |
477 | break; |
478 | case Instruction::BitCast: |
479 | // Look through bitcasts. |
480 | return computeAddress(Obj: U->getOperand(i: 0), Addr); |
481 | case Instruction::GetElementPtr: { |
482 | Address SavedAddr = Addr; |
483 | int64_t TmpOffset = Addr.getOffset(); |
484 | // Iterate through the GEP folding the constants into offsets where |
485 | // we can. |
486 | gep_type_iterator GTI = gep_type_begin(GEP: U); |
487 | for (User::const_op_iterator i = U->op_begin() + 1, e = U->op_end(); i != e; |
488 | ++i, ++GTI) { |
489 | const Value *Op = *i; |
490 | if (StructType *STy = GTI.getStructTypeOrNull()) { |
491 | const StructLayout *SL = DL.getStructLayout(Ty: STy); |
492 | unsigned Idx = cast<ConstantInt>(Val: Op)->getZExtValue(); |
493 | TmpOffset += SL->getElementOffset(Idx); |
494 | } else { |
495 | uint64_t S = GTI.getSequentialElementStride(DL); |
496 | while (true) { |
497 | if (const ConstantInt *CI = dyn_cast<ConstantInt>(Val: Op)) { |
498 | // Constant-offset addressing. |
499 | TmpOffset += CI->getSExtValue() * S; |
500 | break; |
501 | } |
502 | if (canFoldAddIntoGEP(GEP: U, Add: Op)) { |
503 | // A compatible add with a constant operand. Fold the constant. |
504 | ConstantInt *CI = |
505 | cast<ConstantInt>(Val: cast<AddOperator>(Val: Op)->getOperand(i_nocapture: 1)); |
506 | TmpOffset += CI->getSExtValue() * S; |
507 | // Iterate on the other operand. |
508 | Op = cast<AddOperator>(Val: Op)->getOperand(i_nocapture: 0); |
509 | continue; |
510 | } |
511 | // Unsupported |
512 | goto unsupported_gep; |
513 | } |
514 | } |
515 | } |
516 | // Try to grab the base operand now. |
517 | Addr.setOffset(TmpOffset); |
518 | if (computeAddress(Obj: U->getOperand(i: 0), Addr)) |
519 | return true; |
520 | // We failed, restore everything and try the other options. |
521 | Addr = SavedAddr; |
522 | unsupported_gep: |
523 | break; |
524 | } |
525 | case Instruction::Alloca: { |
526 | const AllocaInst *AI = cast<AllocaInst>(Val: Obj); |
527 | DenseMap<const AllocaInst *, int>::iterator SI = |
528 | FuncInfo.StaticAllocaMap.find(Val: AI); |
529 | if (SI != FuncInfo.StaticAllocaMap.end()) { |
530 | Addr.setKind(Address::FrameIndexBase); |
531 | Addr.setFI(SI->second); |
532 | return true; |
533 | } |
534 | break; |
535 | } |
536 | } |
537 | Addr.setReg(getRegForValue(V: Obj)); |
538 | return Addr.getReg() != 0; |
539 | } |
540 | |
541 | bool MipsFastISel::computeCallAddress(const Value *V, Address &Addr) { |
542 | const User *U = nullptr; |
543 | unsigned Opcode = Instruction::UserOp1; |
544 | |
545 | if (const auto *I = dyn_cast<Instruction>(Val: V)) { |
546 | // Check if the value is defined in the same basic block. This information |
547 | // is crucial to know whether or not folding an operand is valid. |
548 | if (I->getParent() == FuncInfo.MBB->getBasicBlock()) { |
549 | Opcode = I->getOpcode(); |
550 | U = I; |
551 | } |
552 | } else if (const auto *C = dyn_cast<ConstantExpr>(Val: V)) { |
553 | Opcode = C->getOpcode(); |
554 | U = C; |
555 | } |
556 | |
557 | switch (Opcode) { |
558 | default: |
559 | break; |
560 | case Instruction::BitCast: |
561 | // Look past bitcasts if its operand is in the same BB. |
562 | return computeCallAddress(V: U->getOperand(i: 0), Addr); |
563 | break; |
564 | case Instruction::IntToPtr: |
565 | // Look past no-op inttoptrs if its operand is in the same BB. |
566 | if (TLI.getValueType(DL, Ty: U->getOperand(i: 0)->getType()) == |
567 | TLI.getPointerTy(DL)) |
568 | return computeCallAddress(V: U->getOperand(i: 0), Addr); |
569 | break; |
570 | case Instruction::PtrToInt: |
571 | // Look past no-op ptrtoints if its operand is in the same BB. |
572 | if (TLI.getValueType(DL, Ty: U->getType()) == TLI.getPointerTy(DL)) |
573 | return computeCallAddress(V: U->getOperand(i: 0), Addr); |
574 | break; |
575 | } |
576 | |
577 | if (const GlobalValue *GV = dyn_cast<GlobalValue>(Val: V)) { |
578 | Addr.setGlobalValue(GV); |
579 | return true; |
580 | } |
581 | |
582 | // If all else fails, try to materialize the value in a register. |
583 | if (!Addr.getGlobalValue()) { |
584 | Addr.setReg(getRegForValue(V)); |
585 | return Addr.getReg() != 0; |
586 | } |
587 | |
588 | return false; |
589 | } |
590 | |
591 | bool MipsFastISel::isTypeLegal(Type *Ty, MVT &VT) { |
592 | EVT evt = TLI.getValueType(DL, Ty, AllowUnknown: true); |
593 | // Only handle simple types. |
594 | if (evt == MVT::Other || !evt.isSimple()) |
595 | return false; |
596 | VT = evt.getSimpleVT(); |
597 | |
598 | // Handle all legal types, i.e. a register that will directly hold this |
599 | // value. |
600 | return TLI.isTypeLegal(VT); |
601 | } |
602 | |
603 | bool MipsFastISel::isTypeSupported(Type *Ty, MVT &VT) { |
604 | if (Ty->isVectorTy()) |
605 | return false; |
606 | |
607 | if (isTypeLegal(Ty, VT)) |
608 | return true; |
609 | |
610 | // If this is a type than can be sign or zero-extended to a basic operation |
611 | // go ahead and accept it now. |
612 | if (VT == MVT::i1 || VT == MVT::i8 || VT == MVT::i16) |
613 | return true; |
614 | |
615 | return false; |
616 | } |
617 | |
618 | bool MipsFastISel::isLoadTypeLegal(Type *Ty, MVT &VT) { |
619 | if (isTypeLegal(Ty, VT)) |
620 | return true; |
621 | // We will extend this in a later patch: |
622 | // If this is a type than can be sign or zero-extended to a basic operation |
623 | // go ahead and accept it now. |
624 | if (VT == MVT::i8 || VT == MVT::i16) |
625 | return true; |
626 | return false; |
627 | } |
628 | |
629 | // Because of how EmitCmp is called with fast-isel, you can |
630 | // end up with redundant "andi" instructions after the sequences emitted below. |
631 | // We should try and solve this issue in the future. |
632 | // |
633 | bool MipsFastISel::emitCmp(unsigned ResultReg, const CmpInst *CI) { |
634 | const Value *Left = CI->getOperand(i_nocapture: 0), *Right = CI->getOperand(i_nocapture: 1); |
635 | bool IsUnsigned = CI->isUnsigned(); |
636 | unsigned LeftReg = getRegEnsuringSimpleIntegerWidening(Left, IsUnsigned); |
637 | if (LeftReg == 0) |
638 | return false; |
639 | unsigned RightReg = getRegEnsuringSimpleIntegerWidening(Right, IsUnsigned); |
640 | if (RightReg == 0) |
641 | return false; |
642 | CmpInst::Predicate P = CI->getPredicate(); |
643 | |
644 | switch (P) { |
645 | default: |
646 | return false; |
647 | case CmpInst::ICMP_EQ: { |
648 | Register TempReg = createResultReg(RC: &Mips::GPR32RegClass); |
649 | emitInst(Opc: Mips::XOR, DstReg: TempReg).addReg(RegNo: LeftReg).addReg(RegNo: RightReg); |
650 | emitInst(Opc: Mips::SLTiu, DstReg: ResultReg).addReg(RegNo: TempReg).addImm(Val: 1); |
651 | break; |
652 | } |
653 | case CmpInst::ICMP_NE: { |
654 | Register TempReg = createResultReg(RC: &Mips::GPR32RegClass); |
655 | emitInst(Opc: Mips::XOR, DstReg: TempReg).addReg(RegNo: LeftReg).addReg(RegNo: RightReg); |
656 | emitInst(Opc: Mips::SLTu, DstReg: ResultReg).addReg(RegNo: Mips::ZERO).addReg(RegNo: TempReg); |
657 | break; |
658 | } |
659 | case CmpInst::ICMP_UGT: |
660 | emitInst(Opc: Mips::SLTu, DstReg: ResultReg).addReg(RegNo: RightReg).addReg(RegNo: LeftReg); |
661 | break; |
662 | case CmpInst::ICMP_ULT: |
663 | emitInst(Opc: Mips::SLTu, DstReg: ResultReg).addReg(RegNo: LeftReg).addReg(RegNo: RightReg); |
664 | break; |
665 | case CmpInst::ICMP_UGE: { |
666 | Register TempReg = createResultReg(RC: &Mips::GPR32RegClass); |
667 | emitInst(Opc: Mips::SLTu, DstReg: TempReg).addReg(RegNo: LeftReg).addReg(RegNo: RightReg); |
668 | emitInst(Opc: Mips::XORi, DstReg: ResultReg).addReg(RegNo: TempReg).addImm(Val: 1); |
669 | break; |
670 | } |
671 | case CmpInst::ICMP_ULE: { |
672 | Register TempReg = createResultReg(RC: &Mips::GPR32RegClass); |
673 | emitInst(Opc: Mips::SLTu, DstReg: TempReg).addReg(RegNo: RightReg).addReg(RegNo: LeftReg); |
674 | emitInst(Opc: Mips::XORi, DstReg: ResultReg).addReg(RegNo: TempReg).addImm(Val: 1); |
675 | break; |
676 | } |
677 | case CmpInst::ICMP_SGT: |
678 | emitInst(Opc: Mips::SLT, DstReg: ResultReg).addReg(RegNo: RightReg).addReg(RegNo: LeftReg); |
679 | break; |
680 | case CmpInst::ICMP_SLT: |
681 | emitInst(Opc: Mips::SLT, DstReg: ResultReg).addReg(RegNo: LeftReg).addReg(RegNo: RightReg); |
682 | break; |
683 | case CmpInst::ICMP_SGE: { |
684 | Register TempReg = createResultReg(RC: &Mips::GPR32RegClass); |
685 | emitInst(Opc: Mips::SLT, DstReg: TempReg).addReg(RegNo: LeftReg).addReg(RegNo: RightReg); |
686 | emitInst(Opc: Mips::XORi, DstReg: ResultReg).addReg(RegNo: TempReg).addImm(Val: 1); |
687 | break; |
688 | } |
689 | case CmpInst::ICMP_SLE: { |
690 | Register TempReg = createResultReg(RC: &Mips::GPR32RegClass); |
691 | emitInst(Opc: Mips::SLT, DstReg: TempReg).addReg(RegNo: RightReg).addReg(RegNo: LeftReg); |
692 | emitInst(Opc: Mips::XORi, DstReg: ResultReg).addReg(RegNo: TempReg).addImm(Val: 1); |
693 | break; |
694 | } |
695 | case CmpInst::FCMP_OEQ: |
696 | case CmpInst::FCMP_UNE: |
697 | case CmpInst::FCMP_OLT: |
698 | case CmpInst::FCMP_OLE: |
699 | case CmpInst::FCMP_OGT: |
700 | case CmpInst::FCMP_OGE: { |
701 | if (UnsupportedFPMode) |
702 | return false; |
703 | bool IsFloat = Left->getType()->isFloatTy(); |
704 | bool IsDouble = Left->getType()->isDoubleTy(); |
705 | if (!IsFloat && !IsDouble) |
706 | return false; |
707 | unsigned Opc, CondMovOpc; |
708 | switch (P) { |
709 | case CmpInst::FCMP_OEQ: |
710 | Opc = IsFloat ? Mips::C_EQ_S : Mips::C_EQ_D32; |
711 | CondMovOpc = Mips::MOVT_I; |
712 | break; |
713 | case CmpInst::FCMP_UNE: |
714 | Opc = IsFloat ? Mips::C_EQ_S : Mips::C_EQ_D32; |
715 | CondMovOpc = Mips::MOVF_I; |
716 | break; |
717 | case CmpInst::FCMP_OLT: |
718 | Opc = IsFloat ? Mips::C_OLT_S : Mips::C_OLT_D32; |
719 | CondMovOpc = Mips::MOVT_I; |
720 | break; |
721 | case CmpInst::FCMP_OLE: |
722 | Opc = IsFloat ? Mips::C_OLE_S : Mips::C_OLE_D32; |
723 | CondMovOpc = Mips::MOVT_I; |
724 | break; |
725 | case CmpInst::FCMP_OGT: |
726 | Opc = IsFloat ? Mips::C_ULE_S : Mips::C_ULE_D32; |
727 | CondMovOpc = Mips::MOVF_I; |
728 | break; |
729 | case CmpInst::FCMP_OGE: |
730 | Opc = IsFloat ? Mips::C_ULT_S : Mips::C_ULT_D32; |
731 | CondMovOpc = Mips::MOVF_I; |
732 | break; |
733 | default: |
734 | llvm_unreachable("Only switching of a subset of CCs." ); |
735 | } |
736 | Register RegWithZero = createResultReg(RC: &Mips::GPR32RegClass); |
737 | Register RegWithOne = createResultReg(RC: &Mips::GPR32RegClass); |
738 | emitInst(Opc: Mips::ADDiu, DstReg: RegWithZero).addReg(RegNo: Mips::ZERO).addImm(Val: 0); |
739 | emitInst(Opc: Mips::ADDiu, DstReg: RegWithOne).addReg(RegNo: Mips::ZERO).addImm(Val: 1); |
740 | emitInst(Opc).addReg(RegNo: Mips::FCC0, flags: RegState::Define).addReg(RegNo: LeftReg) |
741 | .addReg(RegNo: RightReg); |
742 | emitInst(Opc: CondMovOpc, DstReg: ResultReg) |
743 | .addReg(RegNo: RegWithOne) |
744 | .addReg(RegNo: Mips::FCC0) |
745 | .addReg(RegNo: RegWithZero); |
746 | break; |
747 | } |
748 | } |
749 | return true; |
750 | } |
751 | |
752 | bool MipsFastISel::emitLoad(MVT VT, unsigned &ResultReg, Address &Addr) { |
753 | // |
754 | // more cases will be handled here in following patches. |
755 | // |
756 | unsigned Opc; |
757 | switch (VT.SimpleTy) { |
758 | case MVT::i32: |
759 | ResultReg = createResultReg(RC: &Mips::GPR32RegClass); |
760 | Opc = Mips::LW; |
761 | break; |
762 | case MVT::i16: |
763 | ResultReg = createResultReg(RC: &Mips::GPR32RegClass); |
764 | Opc = Mips::LHu; |
765 | break; |
766 | case MVT::i8: |
767 | ResultReg = createResultReg(RC: &Mips::GPR32RegClass); |
768 | Opc = Mips::LBu; |
769 | break; |
770 | case MVT::f32: |
771 | if (UnsupportedFPMode) |
772 | return false; |
773 | ResultReg = createResultReg(RC: &Mips::FGR32RegClass); |
774 | Opc = Mips::LWC1; |
775 | break; |
776 | case MVT::f64: |
777 | if (UnsupportedFPMode) |
778 | return false; |
779 | ResultReg = createResultReg(RC: &Mips::AFGR64RegClass); |
780 | Opc = Mips::LDC1; |
781 | break; |
782 | default: |
783 | return false; |
784 | } |
785 | if (Addr.isRegBase()) { |
786 | simplifyAddress(Addr); |
787 | emitInstLoad(Opc, DstReg: ResultReg, MemReg: Addr.getReg(), MemOffset: Addr.getOffset()); |
788 | return true; |
789 | } |
790 | if (Addr.isFIBase()) { |
791 | unsigned FI = Addr.getFI(); |
792 | int64_t Offset = Addr.getOffset(); |
793 | MachineFrameInfo &MFI = MF->getFrameInfo(); |
794 | MachineMemOperand *MMO = MF->getMachineMemOperand( |
795 | PtrInfo: MachinePointerInfo::getFixedStack(MF&: *MF, FI), F: MachineMemOperand::MOLoad, |
796 | Size: MFI.getObjectSize(ObjectIdx: FI), BaseAlignment: Align(4)); |
797 | BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, MCID: TII.get(Opcode: Opc), DestReg: ResultReg) |
798 | .addFrameIndex(Idx: FI) |
799 | .addImm(Val: Offset) |
800 | .addMemOperand(MMO); |
801 | return true; |
802 | } |
803 | return false; |
804 | } |
805 | |
806 | bool MipsFastISel::emitStore(MVT VT, unsigned SrcReg, Address &Addr) { |
807 | // |
808 | // more cases will be handled here in following patches. |
809 | // |
810 | unsigned Opc; |
811 | switch (VT.SimpleTy) { |
812 | case MVT::i8: |
813 | Opc = Mips::SB; |
814 | break; |
815 | case MVT::i16: |
816 | Opc = Mips::SH; |
817 | break; |
818 | case MVT::i32: |
819 | Opc = Mips::SW; |
820 | break; |
821 | case MVT::f32: |
822 | if (UnsupportedFPMode) |
823 | return false; |
824 | Opc = Mips::SWC1; |
825 | break; |
826 | case MVT::f64: |
827 | if (UnsupportedFPMode) |
828 | return false; |
829 | Opc = Mips::SDC1; |
830 | break; |
831 | default: |
832 | return false; |
833 | } |
834 | if (Addr.isRegBase()) { |
835 | simplifyAddress(Addr); |
836 | emitInstStore(Opc, SrcReg, MemReg: Addr.getReg(), MemOffset: Addr.getOffset()); |
837 | return true; |
838 | } |
839 | if (Addr.isFIBase()) { |
840 | unsigned FI = Addr.getFI(); |
841 | int64_t Offset = Addr.getOffset(); |
842 | MachineFrameInfo &MFI = MF->getFrameInfo(); |
843 | MachineMemOperand *MMO = MF->getMachineMemOperand( |
844 | PtrInfo: MachinePointerInfo::getFixedStack(MF&: *MF, FI), F: MachineMemOperand::MOStore, |
845 | Size: MFI.getObjectSize(ObjectIdx: FI), BaseAlignment: Align(4)); |
846 | BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, MCID: TII.get(Opcode: Opc)) |
847 | .addReg(RegNo: SrcReg) |
848 | .addFrameIndex(Idx: FI) |
849 | .addImm(Val: Offset) |
850 | .addMemOperand(MMO); |
851 | return true; |
852 | } |
853 | return false; |
854 | } |
855 | |
856 | bool MipsFastISel::selectLogicalOp(const Instruction *I) { |
857 | MVT VT; |
858 | if (!isTypeSupported(Ty: I->getType(), VT)) |
859 | return false; |
860 | |
861 | unsigned ResultReg; |
862 | switch (I->getOpcode()) { |
863 | default: |
864 | llvm_unreachable("Unexpected instruction." ); |
865 | case Instruction::And: |
866 | ResultReg = emitLogicalOp(ISDOpc: ISD::AND, RetVT: VT, LHS: I->getOperand(i: 0), RHS: I->getOperand(i: 1)); |
867 | break; |
868 | case Instruction::Or: |
869 | ResultReg = emitLogicalOp(ISDOpc: ISD::OR, RetVT: VT, LHS: I->getOperand(i: 0), RHS: I->getOperand(i: 1)); |
870 | break; |
871 | case Instruction::Xor: |
872 | ResultReg = emitLogicalOp(ISDOpc: ISD::XOR, RetVT: VT, LHS: I->getOperand(i: 0), RHS: I->getOperand(i: 1)); |
873 | break; |
874 | } |
875 | |
876 | if (!ResultReg) |
877 | return false; |
878 | |
879 | updateValueMap(I, Reg: ResultReg); |
880 | return true; |
881 | } |
882 | |
883 | bool MipsFastISel::selectLoad(const Instruction *I) { |
884 | // Atomic loads need special handling. |
885 | if (cast<LoadInst>(Val: I)->isAtomic()) |
886 | return false; |
887 | |
888 | // Verify we have a legal type before going any further. |
889 | MVT VT; |
890 | if (!isLoadTypeLegal(Ty: I->getType(), VT)) |
891 | return false; |
892 | |
893 | // See if we can handle this address. |
894 | Address Addr; |
895 | if (!computeAddress(Obj: I->getOperand(i: 0), Addr)) |
896 | return false; |
897 | |
898 | unsigned ResultReg; |
899 | if (!emitLoad(VT, ResultReg, Addr)) |
900 | return false; |
901 | updateValueMap(I, Reg: ResultReg); |
902 | return true; |
903 | } |
904 | |
905 | bool MipsFastISel::selectStore(const Instruction *I) { |
906 | Value *Op0 = I->getOperand(i: 0); |
907 | unsigned SrcReg = 0; |
908 | |
909 | // Atomic stores need special handling. |
910 | if (cast<StoreInst>(Val: I)->isAtomic()) |
911 | return false; |
912 | |
913 | // Verify we have a legal type before going any further. |
914 | MVT VT; |
915 | if (!isLoadTypeLegal(Ty: I->getOperand(i: 0)->getType(), VT)) |
916 | return false; |
917 | |
918 | // Get the value to be stored into a register. |
919 | SrcReg = getRegForValue(V: Op0); |
920 | if (SrcReg == 0) |
921 | return false; |
922 | |
923 | // See if we can handle this address. |
924 | Address Addr; |
925 | if (!computeAddress(Obj: I->getOperand(i: 1), Addr)) |
926 | return false; |
927 | |
928 | if (!emitStore(VT, SrcReg, Addr)) |
929 | return false; |
930 | return true; |
931 | } |
932 | |
933 | // This can cause a redundant sltiu to be generated. |
934 | // FIXME: try and eliminate this in a future patch. |
935 | bool MipsFastISel::selectBranch(const Instruction *I) { |
936 | const BranchInst *BI = cast<BranchInst>(Val: I); |
937 | MachineBasicBlock *BrBB = FuncInfo.MBB; |
938 | // |
939 | // TBB is the basic block for the case where the comparison is true. |
940 | // FBB is the basic block for the case where the comparison is false. |
941 | // if (cond) goto TBB |
942 | // goto FBB |
943 | // TBB: |
944 | // |
945 | MachineBasicBlock *TBB = FuncInfo.MBBMap[BI->getSuccessor(i: 0)]; |
946 | MachineBasicBlock *FBB = FuncInfo.MBBMap[BI->getSuccessor(i: 1)]; |
947 | |
948 | // Fold the common case of a conditional branch with a comparison |
949 | // in the same block. |
950 | unsigned ZExtCondReg = 0; |
951 | if (const CmpInst *CI = dyn_cast<CmpInst>(Val: BI->getCondition())) { |
952 | if (CI->hasOneUse() && CI->getParent() == I->getParent()) { |
953 | ZExtCondReg = createResultReg(RC: &Mips::GPR32RegClass); |
954 | if (!emitCmp(ResultReg: ZExtCondReg, CI)) |
955 | return false; |
956 | } |
957 | } |
958 | |
959 | // For the general case, we need to mask with 1. |
960 | if (ZExtCondReg == 0) { |
961 | Register CondReg = getRegForValue(V: BI->getCondition()); |
962 | if (CondReg == 0) |
963 | return false; |
964 | |
965 | ZExtCondReg = emitIntExt(SrcVT: MVT::i1, SrcReg: CondReg, DestVT: MVT::i32, isZExt: true); |
966 | if (ZExtCondReg == 0) |
967 | return false; |
968 | } |
969 | |
970 | BuildMI(BB&: *BrBB, I: FuncInfo.InsertPt, MIMD, MCID: TII.get(Opcode: Mips::BGTZ)) |
971 | .addReg(RegNo: ZExtCondReg) |
972 | .addMBB(MBB: TBB); |
973 | finishCondBranch(BranchBB: BI->getParent(), TrueMBB: TBB, FalseMBB: FBB); |
974 | return true; |
975 | } |
976 | |
977 | bool MipsFastISel::selectCmp(const Instruction *I) { |
978 | const CmpInst *CI = cast<CmpInst>(Val: I); |
979 | Register ResultReg = createResultReg(RC: &Mips::GPR32RegClass); |
980 | if (!emitCmp(ResultReg, CI)) |
981 | return false; |
982 | updateValueMap(I, Reg: ResultReg); |
983 | return true; |
984 | } |
985 | |
986 | // Attempt to fast-select a floating-point extend instruction. |
987 | bool MipsFastISel::selectFPExt(const Instruction *I) { |
988 | if (UnsupportedFPMode) |
989 | return false; |
990 | Value *Src = I->getOperand(i: 0); |
991 | EVT SrcVT = TLI.getValueType(DL, Ty: Src->getType(), AllowUnknown: true); |
992 | EVT DestVT = TLI.getValueType(DL, Ty: I->getType(), AllowUnknown: true); |
993 | |
994 | if (SrcVT != MVT::f32 || DestVT != MVT::f64) |
995 | return false; |
996 | |
997 | Register SrcReg = |
998 | getRegForValue(V: Src); // this must be a 32bit floating point register class |
999 | // maybe we should handle this differently |
1000 | if (!SrcReg) |
1001 | return false; |
1002 | |
1003 | Register DestReg = createResultReg(RC: &Mips::AFGR64RegClass); |
1004 | emitInst(Opc: Mips::CVT_D32_S, DstReg: DestReg).addReg(RegNo: SrcReg); |
1005 | updateValueMap(I, Reg: DestReg); |
1006 | return true; |
1007 | } |
1008 | |
1009 | bool MipsFastISel::selectSelect(const Instruction *I) { |
1010 | assert(isa<SelectInst>(I) && "Expected a select instruction." ); |
1011 | |
1012 | LLVM_DEBUG(dbgs() << "selectSelect\n" ); |
1013 | |
1014 | MVT VT; |
1015 | if (!isTypeSupported(Ty: I->getType(), VT) || UnsupportedFPMode) { |
1016 | LLVM_DEBUG( |
1017 | dbgs() << ".. .. gave up (!isTypeSupported || UnsupportedFPMode)\n" ); |
1018 | return false; |
1019 | } |
1020 | |
1021 | unsigned CondMovOpc; |
1022 | const TargetRegisterClass *RC; |
1023 | |
1024 | if (VT.isInteger() && !VT.isVector() && VT.getSizeInBits() <= 32) { |
1025 | CondMovOpc = Mips::MOVN_I_I; |
1026 | RC = &Mips::GPR32RegClass; |
1027 | } else if (VT == MVT::f32) { |
1028 | CondMovOpc = Mips::MOVN_I_S; |
1029 | RC = &Mips::FGR32RegClass; |
1030 | } else if (VT == MVT::f64) { |
1031 | CondMovOpc = Mips::MOVN_I_D32; |
1032 | RC = &Mips::AFGR64RegClass; |
1033 | } else |
1034 | return false; |
1035 | |
1036 | const SelectInst *SI = cast<SelectInst>(Val: I); |
1037 | const Value *Cond = SI->getCondition(); |
1038 | Register Src1Reg = getRegForValue(V: SI->getTrueValue()); |
1039 | Register Src2Reg = getRegForValue(V: SI->getFalseValue()); |
1040 | Register CondReg = getRegForValue(V: Cond); |
1041 | |
1042 | if (!Src1Reg || !Src2Reg || !CondReg) |
1043 | return false; |
1044 | |
1045 | Register ZExtCondReg = createResultReg(RC: &Mips::GPR32RegClass); |
1046 | if (!ZExtCondReg) |
1047 | return false; |
1048 | |
1049 | if (!emitIntExt(SrcVT: MVT::i1, SrcReg: CondReg, DestVT: MVT::i32, DestReg: ZExtCondReg, IsZExt: true)) |
1050 | return false; |
1051 | |
1052 | Register ResultReg = createResultReg(RC); |
1053 | Register TempReg = createResultReg(RC); |
1054 | |
1055 | if (!ResultReg || !TempReg) |
1056 | return false; |
1057 | |
1058 | emitInst(Opc: TargetOpcode::COPY, DstReg: TempReg).addReg(RegNo: Src2Reg); |
1059 | emitInst(Opc: CondMovOpc, DstReg: ResultReg) |
1060 | .addReg(RegNo: Src1Reg).addReg(RegNo: ZExtCondReg).addReg(RegNo: TempReg); |
1061 | updateValueMap(I, Reg: ResultReg); |
1062 | return true; |
1063 | } |
1064 | |
1065 | // Attempt to fast-select a floating-point truncate instruction. |
1066 | bool MipsFastISel::selectFPTrunc(const Instruction *I) { |
1067 | if (UnsupportedFPMode) |
1068 | return false; |
1069 | Value *Src = I->getOperand(i: 0); |
1070 | EVT SrcVT = TLI.getValueType(DL, Ty: Src->getType(), AllowUnknown: true); |
1071 | EVT DestVT = TLI.getValueType(DL, Ty: I->getType(), AllowUnknown: true); |
1072 | |
1073 | if (SrcVT != MVT::f64 || DestVT != MVT::f32) |
1074 | return false; |
1075 | |
1076 | Register SrcReg = getRegForValue(V: Src); |
1077 | if (!SrcReg) |
1078 | return false; |
1079 | |
1080 | Register DestReg = createResultReg(RC: &Mips::FGR32RegClass); |
1081 | if (!DestReg) |
1082 | return false; |
1083 | |
1084 | emitInst(Opc: Mips::CVT_S_D32, DstReg: DestReg).addReg(RegNo: SrcReg); |
1085 | updateValueMap(I, Reg: DestReg); |
1086 | return true; |
1087 | } |
1088 | |
1089 | // Attempt to fast-select a floating-point-to-integer conversion. |
1090 | bool MipsFastISel::selectFPToInt(const Instruction *I, bool IsSigned) { |
1091 | if (UnsupportedFPMode) |
1092 | return false; |
1093 | MVT DstVT, SrcVT; |
1094 | if (!IsSigned) |
1095 | return false; // We don't handle this case yet. There is no native |
1096 | // instruction for this but it can be synthesized. |
1097 | Type *DstTy = I->getType(); |
1098 | if (!isTypeLegal(Ty: DstTy, VT&: DstVT)) |
1099 | return false; |
1100 | |
1101 | if (DstVT != MVT::i32) |
1102 | return false; |
1103 | |
1104 | Value *Src = I->getOperand(i: 0); |
1105 | Type *SrcTy = Src->getType(); |
1106 | if (!isTypeLegal(Ty: SrcTy, VT&: SrcVT)) |
1107 | return false; |
1108 | |
1109 | if (SrcVT != MVT::f32 && SrcVT != MVT::f64) |
1110 | return false; |
1111 | |
1112 | Register SrcReg = getRegForValue(V: Src); |
1113 | if (SrcReg == 0) |
1114 | return false; |
1115 | |
1116 | // Determine the opcode for the conversion, which takes place |
1117 | // entirely within FPRs. |
1118 | Register DestReg = createResultReg(RC: &Mips::GPR32RegClass); |
1119 | Register TempReg = createResultReg(RC: &Mips::FGR32RegClass); |
1120 | unsigned Opc = (SrcVT == MVT::f32) ? Mips::TRUNC_W_S : Mips::TRUNC_W_D32; |
1121 | |
1122 | // Generate the convert. |
1123 | emitInst(Opc, DstReg: TempReg).addReg(RegNo: SrcReg); |
1124 | emitInst(Opc: Mips::MFC1, DstReg: DestReg).addReg(RegNo: TempReg); |
1125 | |
1126 | updateValueMap(I, Reg: DestReg); |
1127 | return true; |
1128 | } |
1129 | |
1130 | bool MipsFastISel::processCallArgs(CallLoweringInfo &CLI, |
1131 | SmallVectorImpl<MVT> &OutVTs, |
1132 | unsigned &NumBytes) { |
1133 | CallingConv::ID CC = CLI.CallConv; |
1134 | SmallVector<CCValAssign, 16> ArgLocs; |
1135 | CCState CCInfo(CC, false, *FuncInfo.MF, ArgLocs, *Context); |
1136 | CCInfo.AnalyzeCallOperands(ArgVTs&: OutVTs, Flags&: CLI.OutFlags, Fn: CCAssignFnForCall(CC)); |
1137 | // Get a count of how many bytes are to be pushed on the stack. |
1138 | NumBytes = CCInfo.getStackSize(); |
1139 | // This is the minimum argument area used for A0-A3. |
1140 | if (NumBytes < 16) |
1141 | NumBytes = 16; |
1142 | |
1143 | emitInst(Opc: Mips::ADJCALLSTACKDOWN).addImm(Val: 16).addImm(Val: 0); |
1144 | // Process the args. |
1145 | MVT firstMVT; |
1146 | for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) { |
1147 | CCValAssign &VA = ArgLocs[i]; |
1148 | const Value *ArgVal = CLI.OutVals[VA.getValNo()]; |
1149 | MVT ArgVT = OutVTs[VA.getValNo()]; |
1150 | |
1151 | if (i == 0) { |
1152 | firstMVT = ArgVT; |
1153 | if (ArgVT == MVT::f32) { |
1154 | VA.convertToReg(RegNo: Mips::F12); |
1155 | } else if (ArgVT == MVT::f64) { |
1156 | if (Subtarget->isFP64bit()) |
1157 | VA.convertToReg(RegNo: Mips::D6_64); |
1158 | else |
1159 | VA.convertToReg(RegNo: Mips::D6); |
1160 | } |
1161 | } else if (i == 1) { |
1162 | if ((firstMVT == MVT::f32) || (firstMVT == MVT::f64)) { |
1163 | if (ArgVT == MVT::f32) { |
1164 | VA.convertToReg(RegNo: Mips::F14); |
1165 | } else if (ArgVT == MVT::f64) { |
1166 | if (Subtarget->isFP64bit()) |
1167 | VA.convertToReg(RegNo: Mips::D7_64); |
1168 | else |
1169 | VA.convertToReg(RegNo: Mips::D7); |
1170 | } |
1171 | } |
1172 | } |
1173 | if (((ArgVT == MVT::i32) || (ArgVT == MVT::f32) || (ArgVT == MVT::i16) || |
1174 | (ArgVT == MVT::i8)) && |
1175 | VA.isMemLoc()) { |
1176 | switch (VA.getLocMemOffset()) { |
1177 | case 0: |
1178 | VA.convertToReg(RegNo: Mips::A0); |
1179 | break; |
1180 | case 4: |
1181 | VA.convertToReg(RegNo: Mips::A1); |
1182 | break; |
1183 | case 8: |
1184 | VA.convertToReg(RegNo: Mips::A2); |
1185 | break; |
1186 | case 12: |
1187 | VA.convertToReg(RegNo: Mips::A3); |
1188 | break; |
1189 | default: |
1190 | break; |
1191 | } |
1192 | } |
1193 | Register ArgReg = getRegForValue(V: ArgVal); |
1194 | if (!ArgReg) |
1195 | return false; |
1196 | |
1197 | // Handle arg promotion: SExt, ZExt, AExt. |
1198 | switch (VA.getLocInfo()) { |
1199 | case CCValAssign::Full: |
1200 | break; |
1201 | case CCValAssign::AExt: |
1202 | case CCValAssign::SExt: { |
1203 | MVT DestVT = VA.getLocVT(); |
1204 | MVT SrcVT = ArgVT; |
1205 | ArgReg = emitIntExt(SrcVT, SrcReg: ArgReg, DestVT, /*isZExt=*/false); |
1206 | if (!ArgReg) |
1207 | return false; |
1208 | break; |
1209 | } |
1210 | case CCValAssign::ZExt: { |
1211 | MVT DestVT = VA.getLocVT(); |
1212 | MVT SrcVT = ArgVT; |
1213 | ArgReg = emitIntExt(SrcVT, SrcReg: ArgReg, DestVT, /*isZExt=*/true); |
1214 | if (!ArgReg) |
1215 | return false; |
1216 | break; |
1217 | } |
1218 | default: |
1219 | llvm_unreachable("Unknown arg promotion!" ); |
1220 | } |
1221 | |
1222 | // Now copy/store arg to correct locations. |
1223 | if (VA.isRegLoc() && !VA.needsCustom()) { |
1224 | BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, |
1225 | MCID: TII.get(Opcode: TargetOpcode::COPY), DestReg: VA.getLocReg()).addReg(RegNo: ArgReg); |
1226 | CLI.OutRegs.push_back(Elt: VA.getLocReg()); |
1227 | } else if (VA.needsCustom()) { |
1228 | llvm_unreachable("Mips does not use custom args." ); |
1229 | return false; |
1230 | } else { |
1231 | // |
1232 | // FIXME: This path will currently return false. It was copied |
1233 | // from the AArch64 port and should be essentially fine for Mips too. |
1234 | // The work to finish up this path will be done in a follow-on patch. |
1235 | // |
1236 | assert(VA.isMemLoc() && "Assuming store on stack." ); |
1237 | // Don't emit stores for undef values. |
1238 | if (isa<UndefValue>(Val: ArgVal)) |
1239 | continue; |
1240 | |
1241 | // Need to store on the stack. |
1242 | // FIXME: This alignment is incorrect but this path is disabled |
1243 | // for now (will return false). We need to determine the right alignment |
1244 | // based on the normal alignment for the underlying machine type. |
1245 | // |
1246 | unsigned ArgSize = alignTo(Size: ArgVT.getSizeInBits(), Align: 4); |
1247 | |
1248 | unsigned BEAlign = 0; |
1249 | if (ArgSize < 8 && !Subtarget->isLittle()) |
1250 | BEAlign = 8 - ArgSize; |
1251 | |
1252 | Address Addr; |
1253 | Addr.setKind(Address::RegBase); |
1254 | Addr.setReg(Mips::SP); |
1255 | Addr.setOffset(VA.getLocMemOffset() + BEAlign); |
1256 | |
1257 | Align Alignment = DL.getABITypeAlign(Ty: ArgVal->getType()); |
1258 | MachineMemOperand *MMO = FuncInfo.MF->getMachineMemOperand( |
1259 | PtrInfo: MachinePointerInfo::getStack(MF&: *FuncInfo.MF, Offset: Addr.getOffset()), |
1260 | F: MachineMemOperand::MOStore, Size: ArgVT.getStoreSize(), BaseAlignment: Alignment); |
1261 | (void)(MMO); |
1262 | // if (!emitStore(ArgVT, ArgReg, Addr, MMO)) |
1263 | return false; // can't store on the stack yet. |
1264 | } |
1265 | } |
1266 | |
1267 | return true; |
1268 | } |
1269 | |
1270 | bool MipsFastISel::finishCall(CallLoweringInfo &CLI, MVT RetVT, |
1271 | unsigned NumBytes) { |
1272 | CallingConv::ID CC = CLI.CallConv; |
1273 | emitInst(Opc: Mips::ADJCALLSTACKUP).addImm(Val: 16).addImm(Val: 0); |
1274 | if (RetVT != MVT::isVoid) { |
1275 | SmallVector<CCValAssign, 16> RVLocs; |
1276 | MipsCCState CCInfo(CC, false, *FuncInfo.MF, RVLocs, *Context); |
1277 | |
1278 | CCInfo.AnalyzeCallResult(Ins: CLI.Ins, Fn: RetCC_Mips, RetTy: CLI.RetTy, |
1279 | Func: CLI.Symbol ? CLI.Symbol->getName().data() |
1280 | : nullptr); |
1281 | |
1282 | // Only handle a single return value. |
1283 | if (RVLocs.size() != 1) |
1284 | return false; |
1285 | // Copy all of the result registers out of their specified physreg. |
1286 | MVT CopyVT = RVLocs[0].getValVT(); |
1287 | // Special handling for extended integers. |
1288 | if (RetVT == MVT::i1 || RetVT == MVT::i8 || RetVT == MVT::i16) |
1289 | CopyVT = MVT::i32; |
1290 | |
1291 | Register ResultReg = createResultReg(RC: TLI.getRegClassFor(VT: CopyVT)); |
1292 | if (!ResultReg) |
1293 | return false; |
1294 | BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, |
1295 | MCID: TII.get(Opcode: TargetOpcode::COPY), |
1296 | DestReg: ResultReg).addReg(RegNo: RVLocs[0].getLocReg()); |
1297 | CLI.InRegs.push_back(Elt: RVLocs[0].getLocReg()); |
1298 | |
1299 | CLI.ResultReg = ResultReg; |
1300 | CLI.NumResultRegs = 1; |
1301 | } |
1302 | return true; |
1303 | } |
1304 | |
1305 | bool MipsFastISel::fastLowerArguments() { |
1306 | LLVM_DEBUG(dbgs() << "fastLowerArguments\n" ); |
1307 | |
1308 | if (!FuncInfo.CanLowerReturn) { |
1309 | LLVM_DEBUG(dbgs() << ".. gave up (!CanLowerReturn)\n" ); |
1310 | return false; |
1311 | } |
1312 | |
1313 | const Function *F = FuncInfo.Fn; |
1314 | if (F->isVarArg()) { |
1315 | LLVM_DEBUG(dbgs() << ".. gave up (varargs)\n" ); |
1316 | return false; |
1317 | } |
1318 | |
1319 | CallingConv::ID CC = F->getCallingConv(); |
1320 | if (CC != CallingConv::C) { |
1321 | LLVM_DEBUG(dbgs() << ".. gave up (calling convention is not C)\n" ); |
1322 | return false; |
1323 | } |
1324 | |
1325 | std::array<MCPhysReg, 4> GPR32ArgRegs = {._M_elems: {Mips::A0, Mips::A1, Mips::A2, |
1326 | Mips::A3}}; |
1327 | std::array<MCPhysReg, 2> FGR32ArgRegs = {._M_elems: {Mips::F12, Mips::F14}}; |
1328 | std::array<MCPhysReg, 2> AFGR64ArgRegs = {._M_elems: {Mips::D6, Mips::D7}}; |
1329 | auto NextGPR32 = GPR32ArgRegs.begin(); |
1330 | auto NextFGR32 = FGR32ArgRegs.begin(); |
1331 | auto NextAFGR64 = AFGR64ArgRegs.begin(); |
1332 | |
1333 | struct AllocatedReg { |
1334 | const TargetRegisterClass *RC; |
1335 | unsigned Reg; |
1336 | AllocatedReg(const TargetRegisterClass *RC, unsigned Reg) |
1337 | : RC(RC), Reg(Reg) {} |
1338 | }; |
1339 | |
1340 | // Only handle simple cases. i.e. All arguments are directly mapped to |
1341 | // registers of the appropriate type. |
1342 | SmallVector<AllocatedReg, 4> Allocation; |
1343 | for (const auto &FormalArg : F->args()) { |
1344 | if (FormalArg.hasAttribute(Kind: Attribute::InReg) || |
1345 | FormalArg.hasAttribute(Kind: Attribute::StructRet) || |
1346 | FormalArg.hasAttribute(Kind: Attribute::ByVal)) { |
1347 | LLVM_DEBUG(dbgs() << ".. gave up (inreg, structret, byval)\n" ); |
1348 | return false; |
1349 | } |
1350 | |
1351 | Type *ArgTy = FormalArg.getType(); |
1352 | if (ArgTy->isStructTy() || ArgTy->isArrayTy() || ArgTy->isVectorTy()) { |
1353 | LLVM_DEBUG(dbgs() << ".. gave up (struct, array, or vector)\n" ); |
1354 | return false; |
1355 | } |
1356 | |
1357 | EVT ArgVT = TLI.getValueType(DL, Ty: ArgTy); |
1358 | LLVM_DEBUG(dbgs() << ".. " << FormalArg.getArgNo() << ": " |
1359 | << ArgVT << "\n" ); |
1360 | if (!ArgVT.isSimple()) { |
1361 | LLVM_DEBUG(dbgs() << ".. .. gave up (not a simple type)\n" ); |
1362 | return false; |
1363 | } |
1364 | |
1365 | switch (ArgVT.getSimpleVT().SimpleTy) { |
1366 | case MVT::i1: |
1367 | case MVT::i8: |
1368 | case MVT::i16: |
1369 | if (!FormalArg.hasAttribute(Kind: Attribute::SExt) && |
1370 | !FormalArg.hasAttribute(Kind: Attribute::ZExt)) { |
1371 | // It must be any extend, this shouldn't happen for clang-generated IR |
1372 | // so just fall back on SelectionDAG. |
1373 | LLVM_DEBUG(dbgs() << ".. .. gave up (i8/i16 arg is not extended)\n" ); |
1374 | return false; |
1375 | } |
1376 | |
1377 | if (NextGPR32 == GPR32ArgRegs.end()) { |
1378 | LLVM_DEBUG(dbgs() << ".. .. gave up (ran out of GPR32 arguments)\n" ); |
1379 | return false; |
1380 | } |
1381 | |
1382 | LLVM_DEBUG(dbgs() << ".. .. GPR32(" << *NextGPR32 << ")\n" ); |
1383 | Allocation.emplace_back(Args: &Mips::GPR32RegClass, Args&: *NextGPR32++); |
1384 | |
1385 | // Allocating any GPR32 prohibits further use of floating point arguments. |
1386 | NextFGR32 = FGR32ArgRegs.end(); |
1387 | NextAFGR64 = AFGR64ArgRegs.end(); |
1388 | break; |
1389 | |
1390 | case MVT::i32: |
1391 | if (FormalArg.hasAttribute(Kind: Attribute::ZExt)) { |
1392 | // The O32 ABI does not permit a zero-extended i32. |
1393 | LLVM_DEBUG(dbgs() << ".. .. gave up (i32 arg is zero extended)\n" ); |
1394 | return false; |
1395 | } |
1396 | |
1397 | if (NextGPR32 == GPR32ArgRegs.end()) { |
1398 | LLVM_DEBUG(dbgs() << ".. .. gave up (ran out of GPR32 arguments)\n" ); |
1399 | return false; |
1400 | } |
1401 | |
1402 | LLVM_DEBUG(dbgs() << ".. .. GPR32(" << *NextGPR32 << ")\n" ); |
1403 | Allocation.emplace_back(Args: &Mips::GPR32RegClass, Args&: *NextGPR32++); |
1404 | |
1405 | // Allocating any GPR32 prohibits further use of floating point arguments. |
1406 | NextFGR32 = FGR32ArgRegs.end(); |
1407 | NextAFGR64 = AFGR64ArgRegs.end(); |
1408 | break; |
1409 | |
1410 | case MVT::f32: |
1411 | if (UnsupportedFPMode) { |
1412 | LLVM_DEBUG(dbgs() << ".. .. gave up (UnsupportedFPMode)\n" ); |
1413 | return false; |
1414 | } |
1415 | if (NextFGR32 == FGR32ArgRegs.end()) { |
1416 | LLVM_DEBUG(dbgs() << ".. .. gave up (ran out of FGR32 arguments)\n" ); |
1417 | return false; |
1418 | } |
1419 | LLVM_DEBUG(dbgs() << ".. .. FGR32(" << *NextFGR32 << ")\n" ); |
1420 | Allocation.emplace_back(Args: &Mips::FGR32RegClass, Args&: *NextFGR32++); |
1421 | // Allocating an FGR32 also allocates the super-register AFGR64, and |
1422 | // ABI rules require us to skip the corresponding GPR32. |
1423 | if (NextGPR32 != GPR32ArgRegs.end()) |
1424 | NextGPR32++; |
1425 | if (NextAFGR64 != AFGR64ArgRegs.end()) |
1426 | NextAFGR64++; |
1427 | break; |
1428 | |
1429 | case MVT::f64: |
1430 | if (UnsupportedFPMode) { |
1431 | LLVM_DEBUG(dbgs() << ".. .. gave up (UnsupportedFPMode)\n" ); |
1432 | return false; |
1433 | } |
1434 | if (NextAFGR64 == AFGR64ArgRegs.end()) { |
1435 | LLVM_DEBUG(dbgs() << ".. .. gave up (ran out of AFGR64 arguments)\n" ); |
1436 | return false; |
1437 | } |
1438 | LLVM_DEBUG(dbgs() << ".. .. AFGR64(" << *NextAFGR64 << ")\n" ); |
1439 | Allocation.emplace_back(Args: &Mips::AFGR64RegClass, Args&: *NextAFGR64++); |
1440 | // Allocating an FGR32 also allocates the super-register AFGR64, and |
1441 | // ABI rules require us to skip the corresponding GPR32 pair. |
1442 | if (NextGPR32 != GPR32ArgRegs.end()) |
1443 | NextGPR32++; |
1444 | if (NextGPR32 != GPR32ArgRegs.end()) |
1445 | NextGPR32++; |
1446 | if (NextFGR32 != FGR32ArgRegs.end()) |
1447 | NextFGR32++; |
1448 | break; |
1449 | |
1450 | default: |
1451 | LLVM_DEBUG(dbgs() << ".. .. gave up (unknown type)\n" ); |
1452 | return false; |
1453 | } |
1454 | } |
1455 | |
1456 | for (const auto &FormalArg : F->args()) { |
1457 | unsigned ArgNo = FormalArg.getArgNo(); |
1458 | unsigned SrcReg = Allocation[ArgNo].Reg; |
1459 | Register DstReg = FuncInfo.MF->addLiveIn(PReg: SrcReg, RC: Allocation[ArgNo].RC); |
1460 | // FIXME: Unfortunately it's necessary to emit a copy from the livein copy. |
1461 | // Without this, EmitLiveInCopies may eliminate the livein if its only |
1462 | // use is a bitcast (which isn't turned into an instruction). |
1463 | Register ResultReg = createResultReg(RC: Allocation[ArgNo].RC); |
1464 | BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, |
1465 | MCID: TII.get(Opcode: TargetOpcode::COPY), DestReg: ResultReg) |
1466 | .addReg(RegNo: DstReg, flags: getKillRegState(B: true)); |
1467 | updateValueMap(I: &FormalArg, Reg: ResultReg); |
1468 | } |
1469 | |
1470 | // Calculate the size of the incoming arguments area. |
1471 | // We currently reject all the cases where this would be non-zero. |
1472 | unsigned IncomingArgSizeInBytes = 0; |
1473 | |
1474 | // Account for the reserved argument area on ABI's that have one (O32). |
1475 | // It seems strange to do this on the caller side but it's necessary in |
1476 | // SelectionDAG's implementation. |
1477 | IncomingArgSizeInBytes = std::min(a: getABI().GetCalleeAllocdArgSizeInBytes(CC), |
1478 | b: IncomingArgSizeInBytes); |
1479 | |
1480 | MF->getInfo<MipsFunctionInfo>()->setFormalArgInfo(Size: IncomingArgSizeInBytes, |
1481 | HasByval: false); |
1482 | |
1483 | return true; |
1484 | } |
1485 | |
1486 | bool MipsFastISel::fastLowerCall(CallLoweringInfo &CLI) { |
1487 | CallingConv::ID CC = CLI.CallConv; |
1488 | bool IsTailCall = CLI.IsTailCall; |
1489 | bool IsVarArg = CLI.IsVarArg; |
1490 | const Value *Callee = CLI.Callee; |
1491 | MCSymbol *Symbol = CLI.Symbol; |
1492 | |
1493 | // Do not handle FastCC. |
1494 | if (CC == CallingConv::Fast) |
1495 | return false; |
1496 | |
1497 | // Allow SelectionDAG isel to handle tail calls. |
1498 | if (IsTailCall) |
1499 | return false; |
1500 | |
1501 | // Let SDISel handle vararg functions. |
1502 | if (IsVarArg) |
1503 | return false; |
1504 | |
1505 | // FIXME: Only handle *simple* calls for now. |
1506 | MVT RetVT; |
1507 | if (CLI.RetTy->isVoidTy()) |
1508 | RetVT = MVT::isVoid; |
1509 | else if (!isTypeSupported(Ty: CLI.RetTy, VT&: RetVT)) |
1510 | return false; |
1511 | |
1512 | for (auto Flag : CLI.OutFlags) |
1513 | if (Flag.isInReg() || Flag.isSRet() || Flag.isNest() || Flag.isByVal()) |
1514 | return false; |
1515 | |
1516 | // Set up the argument vectors. |
1517 | SmallVector<MVT, 16> OutVTs; |
1518 | OutVTs.reserve(N: CLI.OutVals.size()); |
1519 | |
1520 | for (auto *Val : CLI.OutVals) { |
1521 | MVT VT; |
1522 | if (!isTypeLegal(Ty: Val->getType(), VT) && |
1523 | !(VT == MVT::i1 || VT == MVT::i8 || VT == MVT::i16)) |
1524 | return false; |
1525 | |
1526 | // We don't handle vector parameters yet. |
1527 | if (VT.isVector() || VT.getSizeInBits() > 64) |
1528 | return false; |
1529 | |
1530 | OutVTs.push_back(Elt: VT); |
1531 | } |
1532 | |
1533 | Address Addr; |
1534 | if (!computeCallAddress(V: Callee, Addr)) |
1535 | return false; |
1536 | |
1537 | // Handle the arguments now that we've gotten them. |
1538 | unsigned NumBytes; |
1539 | if (!processCallArgs(CLI, OutVTs, NumBytes)) |
1540 | return false; |
1541 | |
1542 | if (!Addr.getGlobalValue()) |
1543 | return false; |
1544 | |
1545 | // Issue the call. |
1546 | unsigned DestAddress; |
1547 | if (Symbol) |
1548 | DestAddress = materializeExternalCallSym(Sym: Symbol); |
1549 | else |
1550 | DestAddress = materializeGV(GV: Addr.getGlobalValue(), VT: MVT::i32); |
1551 | emitInst(Opc: TargetOpcode::COPY, DstReg: Mips::T9).addReg(RegNo: DestAddress); |
1552 | MachineInstrBuilder MIB = |
1553 | BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, MCID: TII.get(Opcode: Mips::JALR), |
1554 | DestReg: Mips::RA).addReg(RegNo: Mips::T9); |
1555 | |
1556 | // Add implicit physical register uses to the call. |
1557 | for (auto Reg : CLI.OutRegs) |
1558 | MIB.addReg(RegNo: Reg, flags: RegState::Implicit); |
1559 | |
1560 | // Add a register mask with the call-preserved registers. |
1561 | // Proper defs for return values will be added by setPhysRegsDeadExcept(). |
1562 | MIB.addRegMask(Mask: TRI.getCallPreservedMask(MF: *FuncInfo.MF, CC)); |
1563 | |
1564 | CLI.Call = MIB; |
1565 | |
1566 | if (EmitJalrReloc && !Subtarget->inMips16Mode()) { |
1567 | // Attach callee address to the instruction, let asm printer emit |
1568 | // .reloc R_MIPS_JALR. |
1569 | if (Symbol) |
1570 | MIB.addSym(Sym: Symbol, TargetFlags: MipsII::MO_JALR); |
1571 | else |
1572 | MIB.addSym(Sym: FuncInfo.MF->getContext().getOrCreateSymbol( |
1573 | Name: Addr.getGlobalValue()->getName()), TargetFlags: MipsII::MO_JALR); |
1574 | } |
1575 | |
1576 | // Finish off the call including any return values. |
1577 | return finishCall(CLI, RetVT, NumBytes); |
1578 | } |
1579 | |
1580 | bool MipsFastISel::fastLowerIntrinsicCall(const IntrinsicInst *II) { |
1581 | switch (II->getIntrinsicID()) { |
1582 | default: |
1583 | return false; |
1584 | case Intrinsic::bswap: { |
1585 | Type *RetTy = II->getCalledFunction()->getReturnType(); |
1586 | |
1587 | MVT VT; |
1588 | if (!isTypeSupported(Ty: RetTy, VT)) |
1589 | return false; |
1590 | |
1591 | Register SrcReg = getRegForValue(V: II->getOperand(i_nocapture: 0)); |
1592 | if (SrcReg == 0) |
1593 | return false; |
1594 | Register DestReg = createResultReg(RC: &Mips::GPR32RegClass); |
1595 | if (DestReg == 0) |
1596 | return false; |
1597 | if (VT == MVT::i16) { |
1598 | if (Subtarget->hasMips32r2()) { |
1599 | emitInst(Opc: Mips::WSBH, DstReg: DestReg).addReg(RegNo: SrcReg); |
1600 | updateValueMap(I: II, Reg: DestReg); |
1601 | return true; |
1602 | } else { |
1603 | unsigned TempReg[3]; |
1604 | for (unsigned &R : TempReg) { |
1605 | R = createResultReg(RC: &Mips::GPR32RegClass); |
1606 | if (R == 0) |
1607 | return false; |
1608 | } |
1609 | emitInst(Opc: Mips::SLL, DstReg: TempReg[0]).addReg(RegNo: SrcReg).addImm(Val: 8); |
1610 | emitInst(Opc: Mips::SRL, DstReg: TempReg[1]).addReg(RegNo: SrcReg).addImm(Val: 8); |
1611 | emitInst(Opc: Mips::ANDi, DstReg: TempReg[2]).addReg(RegNo: TempReg[1]).addImm(Val: 0xFF); |
1612 | emitInst(Opc: Mips::OR, DstReg: DestReg).addReg(RegNo: TempReg[0]).addReg(RegNo: TempReg[2]); |
1613 | updateValueMap(I: II, Reg: DestReg); |
1614 | return true; |
1615 | } |
1616 | } else if (VT == MVT::i32) { |
1617 | if (Subtarget->hasMips32r2()) { |
1618 | Register TempReg = createResultReg(RC: &Mips::GPR32RegClass); |
1619 | emitInst(Opc: Mips::WSBH, DstReg: TempReg).addReg(RegNo: SrcReg); |
1620 | emitInst(Opc: Mips::ROTR, DstReg: DestReg).addReg(RegNo: TempReg).addImm(Val: 16); |
1621 | updateValueMap(I: II, Reg: DestReg); |
1622 | return true; |
1623 | } else { |
1624 | unsigned TempReg[8]; |
1625 | for (unsigned &R : TempReg) { |
1626 | R = createResultReg(RC: &Mips::GPR32RegClass); |
1627 | if (R == 0) |
1628 | return false; |
1629 | } |
1630 | |
1631 | emitInst(Opc: Mips::SRL, DstReg: TempReg[0]).addReg(RegNo: SrcReg).addImm(Val: 8); |
1632 | emitInst(Opc: Mips::SRL, DstReg: TempReg[1]).addReg(RegNo: SrcReg).addImm(Val: 24); |
1633 | emitInst(Opc: Mips::ANDi, DstReg: TempReg[2]).addReg(RegNo: TempReg[0]).addImm(Val: 0xFF00); |
1634 | emitInst(Opc: Mips::OR, DstReg: TempReg[3]).addReg(RegNo: TempReg[1]).addReg(RegNo: TempReg[2]); |
1635 | |
1636 | emitInst(Opc: Mips::ANDi, DstReg: TempReg[4]).addReg(RegNo: SrcReg).addImm(Val: 0xFF00); |
1637 | emitInst(Opc: Mips::SLL, DstReg: TempReg[5]).addReg(RegNo: TempReg[4]).addImm(Val: 8); |
1638 | |
1639 | emitInst(Opc: Mips::SLL, DstReg: TempReg[6]).addReg(RegNo: SrcReg).addImm(Val: 24); |
1640 | emitInst(Opc: Mips::OR, DstReg: TempReg[7]).addReg(RegNo: TempReg[3]).addReg(RegNo: TempReg[5]); |
1641 | emitInst(Opc: Mips::OR, DstReg: DestReg).addReg(RegNo: TempReg[6]).addReg(RegNo: TempReg[7]); |
1642 | updateValueMap(I: II, Reg: DestReg); |
1643 | return true; |
1644 | } |
1645 | } |
1646 | return false; |
1647 | } |
1648 | case Intrinsic::memcpy: |
1649 | case Intrinsic::memmove: { |
1650 | const auto *MTI = cast<MemTransferInst>(Val: II); |
1651 | // Don't handle volatile. |
1652 | if (MTI->isVolatile()) |
1653 | return false; |
1654 | if (!MTI->getLength()->getType()->isIntegerTy(Bitwidth: 32)) |
1655 | return false; |
1656 | const char *IntrMemName = isa<MemCpyInst>(Val: II) ? "memcpy" : "memmove" ; |
1657 | return lowerCallTo(CI: II, SymName: IntrMemName, NumArgs: II->arg_size() - 1); |
1658 | } |
1659 | case Intrinsic::memset: { |
1660 | const MemSetInst *MSI = cast<MemSetInst>(Val: II); |
1661 | // Don't handle volatile. |
1662 | if (MSI->isVolatile()) |
1663 | return false; |
1664 | if (!MSI->getLength()->getType()->isIntegerTy(Bitwidth: 32)) |
1665 | return false; |
1666 | return lowerCallTo(CI: II, SymName: "memset" , NumArgs: II->arg_size() - 1); |
1667 | } |
1668 | } |
1669 | return false; |
1670 | } |
1671 | |
1672 | bool MipsFastISel::selectRet(const Instruction *I) { |
1673 | const Function &F = *I->getParent()->getParent(); |
1674 | const ReturnInst *Ret = cast<ReturnInst>(Val: I); |
1675 | |
1676 | LLVM_DEBUG(dbgs() << "selectRet\n" ); |
1677 | |
1678 | if (!FuncInfo.CanLowerReturn) |
1679 | return false; |
1680 | |
1681 | // Build a list of return value registers. |
1682 | SmallVector<unsigned, 4> RetRegs; |
1683 | |
1684 | if (Ret->getNumOperands() > 0) { |
1685 | CallingConv::ID CC = F.getCallingConv(); |
1686 | |
1687 | // Do not handle FastCC. |
1688 | if (CC == CallingConv::Fast) |
1689 | return false; |
1690 | |
1691 | SmallVector<ISD::OutputArg, 4> Outs; |
1692 | GetReturnInfo(CC, ReturnType: F.getReturnType(), attr: F.getAttributes(), Outs, TLI, DL); |
1693 | |
1694 | // Analyze operands of the call, assigning locations to each operand. |
1695 | SmallVector<CCValAssign, 16> ValLocs; |
1696 | MipsCCState CCInfo(CC, F.isVarArg(), *FuncInfo.MF, ValLocs, |
1697 | I->getContext()); |
1698 | CCAssignFn *RetCC = RetCC_Mips; |
1699 | CCInfo.AnalyzeReturn(Outs, Fn: RetCC); |
1700 | |
1701 | // Only handle a single return value for now. |
1702 | if (ValLocs.size() != 1) |
1703 | return false; |
1704 | |
1705 | CCValAssign &VA = ValLocs[0]; |
1706 | const Value *RV = Ret->getOperand(i_nocapture: 0); |
1707 | |
1708 | // Don't bother handling odd stuff for now. |
1709 | if ((VA.getLocInfo() != CCValAssign::Full) && |
1710 | (VA.getLocInfo() != CCValAssign::BCvt)) |
1711 | return false; |
1712 | |
1713 | // Only handle register returns for now. |
1714 | if (!VA.isRegLoc()) |
1715 | return false; |
1716 | |
1717 | Register Reg = getRegForValue(V: RV); |
1718 | if (Reg == 0) |
1719 | return false; |
1720 | |
1721 | unsigned SrcReg = Reg + VA.getValNo(); |
1722 | Register DestReg = VA.getLocReg(); |
1723 | // Avoid a cross-class copy. This is very unlikely. |
1724 | if (!MRI.getRegClass(Reg: SrcReg)->contains(Reg: DestReg)) |
1725 | return false; |
1726 | |
1727 | EVT RVEVT = TLI.getValueType(DL, Ty: RV->getType()); |
1728 | if (!RVEVT.isSimple()) |
1729 | return false; |
1730 | |
1731 | if (RVEVT.isVector()) |
1732 | return false; |
1733 | |
1734 | MVT RVVT = RVEVT.getSimpleVT(); |
1735 | if (RVVT == MVT::f128) |
1736 | return false; |
1737 | |
1738 | // Do not handle FGR64 returns for now. |
1739 | if (RVVT == MVT::f64 && UnsupportedFPMode) { |
1740 | LLVM_DEBUG(dbgs() << ".. .. gave up (UnsupportedFPMode\n" ); |
1741 | return false; |
1742 | } |
1743 | |
1744 | MVT DestVT = VA.getValVT(); |
1745 | // Special handling for extended integers. |
1746 | if (RVVT != DestVT) { |
1747 | if (RVVT != MVT::i1 && RVVT != MVT::i8 && RVVT != MVT::i16) |
1748 | return false; |
1749 | |
1750 | if (Outs[0].Flags.isZExt() || Outs[0].Flags.isSExt()) { |
1751 | bool IsZExt = Outs[0].Flags.isZExt(); |
1752 | SrcReg = emitIntExt(SrcVT: RVVT, SrcReg, DestVT, isZExt: IsZExt); |
1753 | if (SrcReg == 0) |
1754 | return false; |
1755 | } |
1756 | } |
1757 | |
1758 | // Make the copy. |
1759 | BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, |
1760 | MCID: TII.get(Opcode: TargetOpcode::COPY), DestReg).addReg(RegNo: SrcReg); |
1761 | |
1762 | // Add register to return instruction. |
1763 | RetRegs.push_back(Elt: VA.getLocReg()); |
1764 | } |
1765 | MachineInstrBuilder MIB = emitInst(Opc: Mips::RetRA); |
1766 | for (unsigned Reg : RetRegs) |
1767 | MIB.addReg(RegNo: Reg, flags: RegState::Implicit); |
1768 | return true; |
1769 | } |
1770 | |
1771 | bool MipsFastISel::selectTrunc(const Instruction *I) { |
1772 | // The high bits for a type smaller than the register size are assumed to be |
1773 | // undefined. |
1774 | Value *Op = I->getOperand(i: 0); |
1775 | |
1776 | EVT SrcVT, DestVT; |
1777 | SrcVT = TLI.getValueType(DL, Ty: Op->getType(), AllowUnknown: true); |
1778 | DestVT = TLI.getValueType(DL, Ty: I->getType(), AllowUnknown: true); |
1779 | |
1780 | if (SrcVT != MVT::i32 && SrcVT != MVT::i16 && SrcVT != MVT::i8) |
1781 | return false; |
1782 | if (DestVT != MVT::i16 && DestVT != MVT::i8 && DestVT != MVT::i1) |
1783 | return false; |
1784 | |
1785 | Register SrcReg = getRegForValue(V: Op); |
1786 | if (!SrcReg) |
1787 | return false; |
1788 | |
1789 | // Because the high bits are undefined, a truncate doesn't generate |
1790 | // any code. |
1791 | updateValueMap(I, Reg: SrcReg); |
1792 | return true; |
1793 | } |
1794 | |
1795 | bool MipsFastISel::selectIntExt(const Instruction *I) { |
1796 | Type *DestTy = I->getType(); |
1797 | Value *Src = I->getOperand(i: 0); |
1798 | Type *SrcTy = Src->getType(); |
1799 | |
1800 | bool isZExt = isa<ZExtInst>(Val: I); |
1801 | Register SrcReg = getRegForValue(V: Src); |
1802 | if (!SrcReg) |
1803 | return false; |
1804 | |
1805 | EVT SrcEVT, DestEVT; |
1806 | SrcEVT = TLI.getValueType(DL, Ty: SrcTy, AllowUnknown: true); |
1807 | DestEVT = TLI.getValueType(DL, Ty: DestTy, AllowUnknown: true); |
1808 | if (!SrcEVT.isSimple()) |
1809 | return false; |
1810 | if (!DestEVT.isSimple()) |
1811 | return false; |
1812 | |
1813 | MVT SrcVT = SrcEVT.getSimpleVT(); |
1814 | MVT DestVT = DestEVT.getSimpleVT(); |
1815 | Register ResultReg = createResultReg(RC: &Mips::GPR32RegClass); |
1816 | |
1817 | if (!emitIntExt(SrcVT, SrcReg, DestVT, DestReg: ResultReg, IsZExt: isZExt)) |
1818 | return false; |
1819 | updateValueMap(I, Reg: ResultReg); |
1820 | return true; |
1821 | } |
1822 | |
1823 | bool MipsFastISel::emitIntSExt32r1(MVT SrcVT, unsigned SrcReg, MVT DestVT, |
1824 | unsigned DestReg) { |
1825 | unsigned ShiftAmt; |
1826 | switch (SrcVT.SimpleTy) { |
1827 | default: |
1828 | return false; |
1829 | case MVT::i8: |
1830 | ShiftAmt = 24; |
1831 | break; |
1832 | case MVT::i16: |
1833 | ShiftAmt = 16; |
1834 | break; |
1835 | } |
1836 | Register TempReg = createResultReg(RC: &Mips::GPR32RegClass); |
1837 | emitInst(Opc: Mips::SLL, DstReg: TempReg).addReg(RegNo: SrcReg).addImm(Val: ShiftAmt); |
1838 | emitInst(Opc: Mips::SRA, DstReg: DestReg).addReg(RegNo: TempReg).addImm(Val: ShiftAmt); |
1839 | return true; |
1840 | } |
1841 | |
1842 | bool MipsFastISel::emitIntSExt32r2(MVT SrcVT, unsigned SrcReg, MVT DestVT, |
1843 | unsigned DestReg) { |
1844 | switch (SrcVT.SimpleTy) { |
1845 | default: |
1846 | return false; |
1847 | case MVT::i8: |
1848 | emitInst(Opc: Mips::SEB, DstReg: DestReg).addReg(RegNo: SrcReg); |
1849 | break; |
1850 | case MVT::i16: |
1851 | emitInst(Opc: Mips::SEH, DstReg: DestReg).addReg(RegNo: SrcReg); |
1852 | break; |
1853 | } |
1854 | return true; |
1855 | } |
1856 | |
1857 | bool MipsFastISel::emitIntSExt(MVT SrcVT, unsigned SrcReg, MVT DestVT, |
1858 | unsigned DestReg) { |
1859 | if ((DestVT != MVT::i32) && (DestVT != MVT::i16)) |
1860 | return false; |
1861 | if (Subtarget->hasMips32r2()) |
1862 | return emitIntSExt32r2(SrcVT, SrcReg, DestVT, DestReg); |
1863 | return emitIntSExt32r1(SrcVT, SrcReg, DestVT, DestReg); |
1864 | } |
1865 | |
1866 | bool MipsFastISel::emitIntZExt(MVT SrcVT, unsigned SrcReg, MVT DestVT, |
1867 | unsigned DestReg) { |
1868 | int64_t Imm; |
1869 | |
1870 | switch (SrcVT.SimpleTy) { |
1871 | default: |
1872 | return false; |
1873 | case MVT::i1: |
1874 | Imm = 1; |
1875 | break; |
1876 | case MVT::i8: |
1877 | Imm = 0xff; |
1878 | break; |
1879 | case MVT::i16: |
1880 | Imm = 0xffff; |
1881 | break; |
1882 | } |
1883 | |
1884 | emitInst(Opc: Mips::ANDi, DstReg: DestReg).addReg(RegNo: SrcReg).addImm(Val: Imm); |
1885 | return true; |
1886 | } |
1887 | |
1888 | bool MipsFastISel::emitIntExt(MVT SrcVT, unsigned SrcReg, MVT DestVT, |
1889 | unsigned DestReg, bool IsZExt) { |
1890 | // FastISel does not have plumbing to deal with extensions where the SrcVT or |
1891 | // DestVT are odd things, so test to make sure that they are both types we can |
1892 | // handle (i1/i8/i16/i32 for SrcVT and i8/i16/i32/i64 for DestVT), otherwise |
1893 | // bail out to SelectionDAG. |
1894 | if (((DestVT != MVT::i8) && (DestVT != MVT::i16) && (DestVT != MVT::i32)) || |
1895 | ((SrcVT != MVT::i1) && (SrcVT != MVT::i8) && (SrcVT != MVT::i16))) |
1896 | return false; |
1897 | if (IsZExt) |
1898 | return emitIntZExt(SrcVT, SrcReg, DestVT, DestReg); |
1899 | return emitIntSExt(SrcVT, SrcReg, DestVT, DestReg); |
1900 | } |
1901 | |
1902 | unsigned MipsFastISel::emitIntExt(MVT SrcVT, unsigned SrcReg, MVT DestVT, |
1903 | bool isZExt) { |
1904 | unsigned DestReg = createResultReg(RC: &Mips::GPR32RegClass); |
1905 | bool Success = emitIntExt(SrcVT, SrcReg, DestVT, DestReg, IsZExt: isZExt); |
1906 | return Success ? DestReg : 0; |
1907 | } |
1908 | |
1909 | bool MipsFastISel::selectDivRem(const Instruction *I, unsigned ISDOpcode) { |
1910 | EVT DestEVT = TLI.getValueType(DL, Ty: I->getType(), AllowUnknown: true); |
1911 | if (!DestEVT.isSimple()) |
1912 | return false; |
1913 | |
1914 | MVT DestVT = DestEVT.getSimpleVT(); |
1915 | if (DestVT != MVT::i32) |
1916 | return false; |
1917 | |
1918 | unsigned DivOpc; |
1919 | switch (ISDOpcode) { |
1920 | default: |
1921 | return false; |
1922 | case ISD::SDIV: |
1923 | case ISD::SREM: |
1924 | DivOpc = Mips::SDIV; |
1925 | break; |
1926 | case ISD::UDIV: |
1927 | case ISD::UREM: |
1928 | DivOpc = Mips::UDIV; |
1929 | break; |
1930 | } |
1931 | |
1932 | Register Src0Reg = getRegForValue(V: I->getOperand(i: 0)); |
1933 | Register Src1Reg = getRegForValue(V: I->getOperand(i: 1)); |
1934 | if (!Src0Reg || !Src1Reg) |
1935 | return false; |
1936 | |
1937 | emitInst(Opc: DivOpc).addReg(RegNo: Src0Reg).addReg(RegNo: Src1Reg); |
1938 | emitInst(Opc: Mips::TEQ).addReg(RegNo: Src1Reg).addReg(RegNo: Mips::ZERO).addImm(Val: 7); |
1939 | |
1940 | Register ResultReg = createResultReg(RC: &Mips::GPR32RegClass); |
1941 | if (!ResultReg) |
1942 | return false; |
1943 | |
1944 | unsigned MFOpc = (ISDOpcode == ISD::SREM || ISDOpcode == ISD::UREM) |
1945 | ? Mips::MFHI |
1946 | : Mips::MFLO; |
1947 | emitInst(Opc: MFOpc, DstReg: ResultReg); |
1948 | |
1949 | updateValueMap(I, Reg: ResultReg); |
1950 | return true; |
1951 | } |
1952 | |
1953 | bool MipsFastISel::selectShift(const Instruction *I) { |
1954 | MVT RetVT; |
1955 | |
1956 | if (!isTypeSupported(Ty: I->getType(), VT&: RetVT)) |
1957 | return false; |
1958 | |
1959 | Register ResultReg = createResultReg(RC: &Mips::GPR32RegClass); |
1960 | if (!ResultReg) |
1961 | return false; |
1962 | |
1963 | unsigned Opcode = I->getOpcode(); |
1964 | const Value *Op0 = I->getOperand(i: 0); |
1965 | Register Op0Reg = getRegForValue(V: Op0); |
1966 | if (!Op0Reg) |
1967 | return false; |
1968 | |
1969 | // If AShr or LShr, then we need to make sure the operand0 is sign extended. |
1970 | if (Opcode == Instruction::AShr || Opcode == Instruction::LShr) { |
1971 | Register TempReg = createResultReg(RC: &Mips::GPR32RegClass); |
1972 | if (!TempReg) |
1973 | return false; |
1974 | |
1975 | MVT Op0MVT = TLI.getValueType(DL, Ty: Op0->getType(), AllowUnknown: true).getSimpleVT(); |
1976 | bool IsZExt = Opcode == Instruction::LShr; |
1977 | if (!emitIntExt(SrcVT: Op0MVT, SrcReg: Op0Reg, DestVT: MVT::i32, DestReg: TempReg, IsZExt)) |
1978 | return false; |
1979 | |
1980 | Op0Reg = TempReg; |
1981 | } |
1982 | |
1983 | if (const auto *C = dyn_cast<ConstantInt>(Val: I->getOperand(i: 1))) { |
1984 | uint64_t ShiftVal = C->getZExtValue(); |
1985 | |
1986 | switch (Opcode) { |
1987 | default: |
1988 | llvm_unreachable("Unexpected instruction." ); |
1989 | case Instruction::Shl: |
1990 | Opcode = Mips::SLL; |
1991 | break; |
1992 | case Instruction::AShr: |
1993 | Opcode = Mips::SRA; |
1994 | break; |
1995 | case Instruction::LShr: |
1996 | Opcode = Mips::SRL; |
1997 | break; |
1998 | } |
1999 | |
2000 | emitInst(Opc: Opcode, DstReg: ResultReg).addReg(RegNo: Op0Reg).addImm(Val: ShiftVal); |
2001 | updateValueMap(I, Reg: ResultReg); |
2002 | return true; |
2003 | } |
2004 | |
2005 | Register Op1Reg = getRegForValue(V: I->getOperand(i: 1)); |
2006 | if (!Op1Reg) |
2007 | return false; |
2008 | |
2009 | switch (Opcode) { |
2010 | default: |
2011 | llvm_unreachable("Unexpected instruction." ); |
2012 | case Instruction::Shl: |
2013 | Opcode = Mips::SLLV; |
2014 | break; |
2015 | case Instruction::AShr: |
2016 | Opcode = Mips::SRAV; |
2017 | break; |
2018 | case Instruction::LShr: |
2019 | Opcode = Mips::SRLV; |
2020 | break; |
2021 | } |
2022 | |
2023 | emitInst(Opc: Opcode, DstReg: ResultReg).addReg(RegNo: Op0Reg).addReg(RegNo: Op1Reg); |
2024 | updateValueMap(I, Reg: ResultReg); |
2025 | return true; |
2026 | } |
2027 | |
2028 | bool MipsFastISel::fastSelectInstruction(const Instruction *I) { |
2029 | switch (I->getOpcode()) { |
2030 | default: |
2031 | break; |
2032 | case Instruction::Load: |
2033 | return selectLoad(I); |
2034 | case Instruction::Store: |
2035 | return selectStore(I); |
2036 | case Instruction::SDiv: |
2037 | if (!selectBinaryOp(I, ISDOpcode: ISD::SDIV)) |
2038 | return selectDivRem(I, ISDOpcode: ISD::SDIV); |
2039 | return true; |
2040 | case Instruction::UDiv: |
2041 | if (!selectBinaryOp(I, ISDOpcode: ISD::UDIV)) |
2042 | return selectDivRem(I, ISDOpcode: ISD::UDIV); |
2043 | return true; |
2044 | case Instruction::SRem: |
2045 | if (!selectBinaryOp(I, ISDOpcode: ISD::SREM)) |
2046 | return selectDivRem(I, ISDOpcode: ISD::SREM); |
2047 | return true; |
2048 | case Instruction::URem: |
2049 | if (!selectBinaryOp(I, ISDOpcode: ISD::UREM)) |
2050 | return selectDivRem(I, ISDOpcode: ISD::UREM); |
2051 | return true; |
2052 | case Instruction::Shl: |
2053 | case Instruction::LShr: |
2054 | case Instruction::AShr: |
2055 | return selectShift(I); |
2056 | case Instruction::And: |
2057 | case Instruction::Or: |
2058 | case Instruction::Xor: |
2059 | return selectLogicalOp(I); |
2060 | case Instruction::Br: |
2061 | return selectBranch(I); |
2062 | case Instruction::Ret: |
2063 | return selectRet(I); |
2064 | case Instruction::Trunc: |
2065 | return selectTrunc(I); |
2066 | case Instruction::ZExt: |
2067 | case Instruction::SExt: |
2068 | return selectIntExt(I); |
2069 | case Instruction::FPTrunc: |
2070 | return selectFPTrunc(I); |
2071 | case Instruction::FPExt: |
2072 | return selectFPExt(I); |
2073 | case Instruction::FPToSI: |
2074 | return selectFPToInt(I, /*isSigned*/ IsSigned: true); |
2075 | case Instruction::FPToUI: |
2076 | return selectFPToInt(I, /*isSigned*/ IsSigned: false); |
2077 | case Instruction::ICmp: |
2078 | case Instruction::FCmp: |
2079 | return selectCmp(I); |
2080 | case Instruction::Select: |
2081 | return selectSelect(I); |
2082 | } |
2083 | return false; |
2084 | } |
2085 | |
2086 | unsigned MipsFastISel::getRegEnsuringSimpleIntegerWidening(const Value *V, |
2087 | bool IsUnsigned) { |
2088 | Register VReg = getRegForValue(V); |
2089 | if (VReg == 0) |
2090 | return 0; |
2091 | MVT VMVT = TLI.getValueType(DL, Ty: V->getType(), AllowUnknown: true).getSimpleVT(); |
2092 | |
2093 | if (VMVT == MVT::i1) |
2094 | return 0; |
2095 | |
2096 | if ((VMVT == MVT::i8) || (VMVT == MVT::i16)) { |
2097 | Register TempReg = createResultReg(RC: &Mips::GPR32RegClass); |
2098 | if (!emitIntExt(SrcVT: VMVT, SrcReg: VReg, DestVT: MVT::i32, DestReg: TempReg, IsZExt: IsUnsigned)) |
2099 | return 0; |
2100 | VReg = TempReg; |
2101 | } |
2102 | return VReg; |
2103 | } |
2104 | |
2105 | void MipsFastISel::simplifyAddress(Address &Addr) { |
2106 | if (!isInt<16>(x: Addr.getOffset())) { |
2107 | unsigned TempReg = |
2108 | materialize32BitInt(Imm: Addr.getOffset(), RC: &Mips::GPR32RegClass); |
2109 | Register DestReg = createResultReg(RC: &Mips::GPR32RegClass); |
2110 | emitInst(Opc: Mips::ADDu, DstReg: DestReg).addReg(RegNo: TempReg).addReg(RegNo: Addr.getReg()); |
2111 | Addr.setReg(DestReg); |
2112 | Addr.setOffset(0); |
2113 | } |
2114 | } |
2115 | |
2116 | unsigned MipsFastISel::fastEmitInst_rr(unsigned MachineInstOpcode, |
2117 | const TargetRegisterClass *RC, |
2118 | unsigned Op0, unsigned Op1) { |
2119 | // We treat the MUL instruction in a special way because it clobbers |
2120 | // the HI0 & LO0 registers. The TableGen definition of this instruction can |
2121 | // mark these registers only as implicitly defined. As a result, the |
2122 | // register allocator runs out of registers when this instruction is |
2123 | // followed by another instruction that defines the same registers too. |
2124 | // We can fix this by explicitly marking those registers as dead. |
2125 | if (MachineInstOpcode == Mips::MUL) { |
2126 | Register ResultReg = createResultReg(RC); |
2127 | const MCInstrDesc &II = TII.get(Opcode: MachineInstOpcode); |
2128 | Op0 = constrainOperandRegClass(II, Op: Op0, OpNum: II.getNumDefs()); |
2129 | Op1 = constrainOperandRegClass(II, Op: Op1, OpNum: II.getNumDefs() + 1); |
2130 | BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, MCID: II, DestReg: ResultReg) |
2131 | .addReg(RegNo: Op0) |
2132 | .addReg(RegNo: Op1) |
2133 | .addReg(RegNo: Mips::HI0, flags: RegState::ImplicitDefine | RegState::Dead) |
2134 | .addReg(RegNo: Mips::LO0, flags: RegState::ImplicitDefine | RegState::Dead); |
2135 | return ResultReg; |
2136 | } |
2137 | |
2138 | return FastISel::fastEmitInst_rr(MachineInstOpcode, RC, Op0, Op1); |
2139 | } |
2140 | |
2141 | namespace llvm { |
2142 | |
2143 | FastISel *Mips::createFastISel(FunctionLoweringInfo &funcInfo, |
2144 | const TargetLibraryInfo *libInfo) { |
2145 | return new MipsFastISel(funcInfo, libInfo); |
2146 | } |
2147 | |
2148 | } // end namespace llvm |
2149 | |