1 | //===- ARMFastISel.cpp - ARM 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 | // This file defines the ARM-specific support for the FastISel class. Some |
10 | // of the target-specific code is generated by tablegen in the file |
11 | // ARMGenFastISel.inc, which is #included here. |
12 | // |
13 | //===----------------------------------------------------------------------===// |
14 | |
15 | #include "ARM.h" |
16 | #include "ARMBaseInstrInfo.h" |
17 | #include "ARMBaseRegisterInfo.h" |
18 | #include "ARMCallingConv.h" |
19 | #include "ARMConstantPoolValue.h" |
20 | #include "ARMISelLowering.h" |
21 | #include "ARMMachineFunctionInfo.h" |
22 | #include "ARMSubtarget.h" |
23 | #include "MCTargetDesc/ARMAddressingModes.h" |
24 | #include "MCTargetDesc/ARMBaseInfo.h" |
25 | #include "Utils/ARMBaseInfo.h" |
26 | #include "llvm/ADT/APFloat.h" |
27 | #include "llvm/ADT/APInt.h" |
28 | #include "llvm/ADT/DenseMap.h" |
29 | #include "llvm/ADT/SmallVector.h" |
30 | #include "llvm/CodeGen/CallingConvLower.h" |
31 | #include "llvm/CodeGen/FastISel.h" |
32 | #include "llvm/CodeGen/FunctionLoweringInfo.h" |
33 | #include "llvm/CodeGen/ISDOpcodes.h" |
34 | #include "llvm/CodeGen/MachineBasicBlock.h" |
35 | #include "llvm/CodeGen/MachineConstantPool.h" |
36 | #include "llvm/CodeGen/MachineFrameInfo.h" |
37 | #include "llvm/CodeGen/MachineFunction.h" |
38 | #include "llvm/CodeGen/MachineInstr.h" |
39 | #include "llvm/CodeGen/MachineInstrBuilder.h" |
40 | #include "llvm/CodeGen/MachineMemOperand.h" |
41 | #include "llvm/CodeGen/MachineOperand.h" |
42 | #include "llvm/CodeGen/MachineRegisterInfo.h" |
43 | #include "llvm/CodeGen/TargetInstrInfo.h" |
44 | #include "llvm/CodeGen/TargetLowering.h" |
45 | #include "llvm/CodeGen/TargetOpcodes.h" |
46 | #include "llvm/CodeGen/TargetRegisterInfo.h" |
47 | #include "llvm/CodeGen/ValueTypes.h" |
48 | #include "llvm/CodeGenTypes/MachineValueType.h" |
49 | #include "llvm/IR/Argument.h" |
50 | #include "llvm/IR/Attributes.h" |
51 | #include "llvm/IR/CallingConv.h" |
52 | #include "llvm/IR/Constant.h" |
53 | #include "llvm/IR/Constants.h" |
54 | #include "llvm/IR/DataLayout.h" |
55 | #include "llvm/IR/DerivedTypes.h" |
56 | #include "llvm/IR/Function.h" |
57 | #include "llvm/IR/GetElementPtrTypeIterator.h" |
58 | #include "llvm/IR/GlobalValue.h" |
59 | #include "llvm/IR/GlobalVariable.h" |
60 | #include "llvm/IR/InstrTypes.h" |
61 | #include "llvm/IR/Instruction.h" |
62 | #include "llvm/IR/Instructions.h" |
63 | #include "llvm/IR/IntrinsicInst.h" |
64 | #include "llvm/IR/Intrinsics.h" |
65 | #include "llvm/IR/Module.h" |
66 | #include "llvm/IR/Operator.h" |
67 | #include "llvm/IR/Type.h" |
68 | #include "llvm/IR/User.h" |
69 | #include "llvm/IR/Value.h" |
70 | #include "llvm/MC/MCInstrDesc.h" |
71 | #include "llvm/Support/Casting.h" |
72 | #include "llvm/Support/Compiler.h" |
73 | #include "llvm/Support/ErrorHandling.h" |
74 | #include "llvm/Support/MathExtras.h" |
75 | #include "llvm/Target/TargetMachine.h" |
76 | #include "llvm/Target/TargetOptions.h" |
77 | #include <cassert> |
78 | #include <cstdint> |
79 | #include <utility> |
80 | |
81 | using namespace llvm; |
82 | |
83 | namespace { |
84 | |
85 | // All possible address modes, plus some. |
86 | class Address { |
87 | public: |
88 | using BaseKind = enum { RegBase, FrameIndexBase }; |
89 | |
90 | private: |
91 | BaseKind Kind = RegBase; |
92 | union { |
93 | unsigned Reg; |
94 | int FI; |
95 | } Base; |
96 | |
97 | int Offset = 0; |
98 | |
99 | public: |
100 | // Innocuous defaults for our address. |
101 | Address() { Base.Reg = 0; } |
102 | |
103 | void setKind(BaseKind K) { Kind = K; } |
104 | BaseKind getKind() const { return Kind; } |
105 | bool isRegBase() const { return Kind == RegBase; } |
106 | bool isFIBase() const { return Kind == FrameIndexBase; } |
107 | |
108 | void setReg(Register Reg) { |
109 | assert(isRegBase() && "Invalid base register access!"); |
110 | Base.Reg = Reg.id(); |
111 | } |
112 | |
113 | Register getReg() const { |
114 | assert(isRegBase() && "Invalid base register access!"); |
115 | return Base.Reg; |
116 | } |
117 | |
118 | void setFI(int FI) { |
119 | assert(isFIBase() && "Invalid base frame index access!"); |
120 | Base.FI = FI; |
121 | } |
122 | |
123 | int getFI() const { |
124 | assert(isFIBase() && "Invalid base frame index access!"); |
125 | return Base.FI; |
126 | } |
127 | |
128 | void setOffset(int O) { Offset = O; } |
129 | int getOffset() { return Offset; } |
130 | }; |
131 | |
132 | class ARMFastISel final : public FastISel { |
133 | /// Subtarget - Keep a pointer to the ARMSubtarget around so that we can |
134 | /// make the right decision when generating code for different targets. |
135 | const ARMSubtarget *Subtarget; |
136 | Module &M; |
137 | const TargetMachine &TM; |
138 | const TargetInstrInfo &TII; |
139 | const TargetLowering &TLI; |
140 | ARMFunctionInfo *AFI; |
141 | |
142 | // Convenience variables to avoid some queries. |
143 | bool isThumb2; |
144 | LLVMContext *Context; |
145 | |
146 | public: |
147 | explicit ARMFastISel(FunctionLoweringInfo &funcInfo, |
148 | const TargetLibraryInfo *libInfo) |
149 | : FastISel(funcInfo, libInfo), |
150 | Subtarget(&funcInfo.MF->getSubtarget<ARMSubtarget>()), |
151 | M(const_cast<Module &>(*funcInfo.Fn->getParent())), |
152 | TM(funcInfo.MF->getTarget()), TII(*Subtarget->getInstrInfo()), |
153 | TLI(*Subtarget->getTargetLowering()) { |
154 | AFI = funcInfo.MF->getInfo<ARMFunctionInfo>(); |
155 | isThumb2 = AFI->isThumbFunction(); |
156 | Context = &funcInfo.Fn->getContext(); |
157 | } |
158 | |
159 | private: |
160 | // Code from FastISel.cpp. |
161 | |
162 | Register fastEmitInst_r(unsigned MachineInstOpcode, |
163 | const TargetRegisterClass *RC, Register Op0); |
164 | Register fastEmitInst_rr(unsigned MachineInstOpcode, |
165 | const TargetRegisterClass *RC, Register Op0, |
166 | Register Op1); |
167 | Register fastEmitInst_ri(unsigned MachineInstOpcode, |
168 | const TargetRegisterClass *RC, Register Op0, |
169 | uint64_t Imm); |
170 | Register fastEmitInst_i(unsigned MachineInstOpcode, |
171 | const TargetRegisterClass *RC, uint64_t Imm); |
172 | |
173 | // Backend specific FastISel code. |
174 | |
175 | bool fastSelectInstruction(const Instruction *I) override; |
176 | Register fastMaterializeConstant(const Constant *C) override; |
177 | Register fastMaterializeAlloca(const AllocaInst *AI) override; |
178 | bool tryToFoldLoadIntoMI(MachineInstr *MI, unsigned OpNo, |
179 | const LoadInst *LI) override; |
180 | bool fastLowerArguments() override; |
181 | |
182 | #include "ARMGenFastISel.inc" |
183 | |
184 | // Instruction selection routines. |
185 | |
186 | bool SelectLoad(const Instruction *I); |
187 | bool SelectStore(const Instruction *I); |
188 | bool SelectBranch(const Instruction *I); |
189 | bool SelectIndirectBr(const Instruction *I); |
190 | bool SelectCmp(const Instruction *I); |
191 | bool SelectFPExt(const Instruction *I); |
192 | bool SelectFPTrunc(const Instruction *I); |
193 | bool SelectBinaryIntOp(const Instruction *I, unsigned ISDOpcode); |
194 | bool SelectBinaryFPOp(const Instruction *I, unsigned ISDOpcode); |
195 | bool SelectIToFP(const Instruction *I, bool isSigned); |
196 | bool SelectFPToI(const Instruction *I, bool isSigned); |
197 | bool SelectDiv(const Instruction *I, bool isSigned); |
198 | bool SelectRem(const Instruction *I, bool isSigned); |
199 | bool SelectCall(const Instruction *I, const char *IntrMemName); |
200 | bool SelectIntrinsicCall(const IntrinsicInst &I); |
201 | bool SelectSelect(const Instruction *I); |
202 | bool SelectRet(const Instruction *I); |
203 | bool SelectTrunc(const Instruction *I); |
204 | bool SelectIntExt(const Instruction *I); |
205 | bool SelectShift(const Instruction *I, ARM_AM::ShiftOpc ShiftTy); |
206 | |
207 | // Utility routines. |
208 | |
209 | bool isPositionIndependent() const; |
210 | bool isTypeLegal(Type *Ty, MVT &VT); |
211 | bool isLoadTypeLegal(Type *Ty, MVT &VT); |
212 | bool ARMEmitCmp(const Value *Src1Value, const Value *Src2Value, |
213 | bool isZExt); |
214 | bool ARMEmitLoad(MVT VT, Register &ResultReg, Address &Addr, |
215 | MaybeAlign Alignment = std::nullopt, bool isZExt = true, |
216 | bool allocReg = true); |
217 | bool ARMEmitStore(MVT VT, Register SrcReg, Address &Addr, |
218 | MaybeAlign Alignment = std::nullopt); |
219 | bool ARMComputeAddress(const Value *Obj, Address &Addr); |
220 | void ARMSimplifyAddress(Address &Addr, MVT VT, bool useAM3); |
221 | bool ARMIsMemCpySmall(uint64_t Len); |
222 | bool ARMTryEmitSmallMemCpy(Address Dest, Address Src, uint64_t Len, |
223 | MaybeAlign Alignment); |
224 | Register ARMEmitIntExt(MVT SrcVT, Register SrcReg, MVT DestVT, bool isZExt); |
225 | Register ARMMaterializeFP(const ConstantFP *CFP, MVT VT); |
226 | Register ARMMaterializeInt(const Constant *C, MVT VT); |
227 | Register ARMMaterializeGV(const GlobalValue *GV, MVT VT); |
228 | Register ARMMoveToFPReg(MVT VT, Register SrcReg); |
229 | Register ARMMoveToIntReg(MVT VT, Register SrcReg); |
230 | unsigned ARMSelectCallOp(bool UseReg); |
231 | Register ARMLowerPICELF(const GlobalValue *GV, MVT VT); |
232 | |
233 | const TargetLowering *getTargetLowering() { return &TLI; } |
234 | |
235 | // Call handling routines. |
236 | |
237 | CCAssignFn *CCAssignFnForCall(CallingConv::ID CC, |
238 | bool Return, |
239 | bool isVarArg); |
240 | bool ProcessCallArgs(SmallVectorImpl<Value*> &Args, |
241 | SmallVectorImpl<Register> &ArgRegs, |
242 | SmallVectorImpl<MVT> &ArgVTs, |
243 | SmallVectorImpl<ISD::ArgFlagsTy> &ArgFlags, |
244 | SmallVectorImpl<Register> &RegArgs, |
245 | CallingConv::ID CC, |
246 | unsigned &NumBytes, |
247 | bool isVarArg); |
248 | Register getLibcallReg(const Twine &Name); |
249 | bool FinishCall(MVT RetVT, SmallVectorImpl<Register> &UsedRegs, |
250 | const Instruction *I, CallingConv::ID CC, |
251 | unsigned &NumBytes, bool isVarArg); |
252 | bool ARMEmitLibcall(const Instruction *I, RTLIB::Libcall Call); |
253 | |
254 | // OptionalDef handling routines. |
255 | |
256 | bool isARMNEONPred(const MachineInstr *MI); |
257 | bool DefinesOptionalPredicate(MachineInstr *MI, bool *CPSR); |
258 | const MachineInstrBuilder &AddOptionalDefs(const MachineInstrBuilder &MIB); |
259 | void AddLoadStoreOperands(MVT VT, Address &Addr, |
260 | const MachineInstrBuilder &MIB, |
261 | MachineMemOperand::Flags Flags, bool useAM3); |
262 | }; |
263 | |
264 | } // end anonymous namespace |
265 | |
266 | // DefinesOptionalPredicate - This is different from DefinesPredicate in that |
267 | // we don't care about implicit defs here, just places we'll need to add a |
268 | // default CCReg argument. Sets CPSR if we're setting CPSR instead of CCR. |
269 | bool ARMFastISel::DefinesOptionalPredicate(MachineInstr *MI, bool *CPSR) { |
270 | if (!MI->hasOptionalDef()) |
271 | return false; |
272 | |
273 | // Look to see if our OptionalDef is defining CPSR or CCR. |
274 | for (const MachineOperand &MO : MI->operands()) { |
275 | if (!MO.isReg() || !MO.isDef()) continue; |
276 | if (MO.getReg() == ARM::CPSR) |
277 | *CPSR = true; |
278 | } |
279 | return true; |
280 | } |
281 | |
282 | bool ARMFastISel::isARMNEONPred(const MachineInstr *MI) { |
283 | const MCInstrDesc &MCID = MI->getDesc(); |
284 | |
285 | // If we're a thumb2 or not NEON function we'll be handled via isPredicable. |
286 | if ((MCID.TSFlags & ARMII::DomainMask) != ARMII::DomainNEON || |
287 | AFI->isThumb2Function()) |
288 | return MI->isPredicable(); |
289 | |
290 | for (const MCOperandInfo &opInfo : MCID.operands()) |
291 | if (opInfo.isPredicate()) |
292 | return true; |
293 | |
294 | return false; |
295 | } |
296 | |
297 | // If the machine is predicable go ahead and add the predicate operands, if |
298 | // it needs default CC operands add those. |
299 | // TODO: If we want to support thumb1 then we'll need to deal with optional |
300 | // CPSR defs that need to be added before the remaining operands. See s_cc_out |
301 | // for descriptions why. |
302 | const MachineInstrBuilder & |
303 | ARMFastISel::AddOptionalDefs(const MachineInstrBuilder &MIB) { |
304 | MachineInstr *MI = &*MIB; |
305 | |
306 | // Do we use a predicate? or... |
307 | // Are we NEON in ARM mode and have a predicate operand? If so, I know |
308 | // we're not predicable but add it anyways. |
309 | if (isARMNEONPred(MI)) |
310 | MIB.add(MOs: predOps(Pred: ARMCC::AL)); |
311 | |
312 | // Do we optionally set a predicate? Preds is size > 0 iff the predicate |
313 | // defines CPSR. All other OptionalDefines in ARM are the CCR register. |
314 | bool CPSR = false; |
315 | if (DefinesOptionalPredicate(MI, CPSR: &CPSR)) |
316 | MIB.add(MO: CPSR ? t1CondCodeOp() : condCodeOp()); |
317 | return MIB; |
318 | } |
319 | |
320 | Register ARMFastISel::fastEmitInst_r(unsigned MachineInstOpcode, |
321 | const TargetRegisterClass *RC, |
322 | Register Op0) { |
323 | Register ResultReg = createResultReg(RC); |
324 | const MCInstrDesc &II = TII.get(Opcode: MachineInstOpcode); |
325 | |
326 | // Make sure the input operand is sufficiently constrained to be legal |
327 | // for this instruction. |
328 | Op0 = constrainOperandRegClass(II, Op: Op0, OpNum: 1); |
329 | if (II.getNumDefs() >= 1) { |
330 | AddOptionalDefs(MIB: BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, MCID: II, |
331 | DestReg: ResultReg).addReg(RegNo: Op0)); |
332 | } else { |
333 | AddOptionalDefs(MIB: BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, MCID: II) |
334 | .addReg(RegNo: Op0)); |
335 | AddOptionalDefs(MIB: BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, |
336 | MCID: TII.get(Opcode: TargetOpcode::COPY), DestReg: ResultReg) |
337 | .addReg(RegNo: II.implicit_defs()[0])); |
338 | } |
339 | return ResultReg; |
340 | } |
341 | |
342 | Register ARMFastISel::fastEmitInst_rr(unsigned MachineInstOpcode, |
343 | const TargetRegisterClass *RC, |
344 | Register Op0, Register Op1) { |
345 | Register ResultReg = createResultReg(RC); |
346 | const MCInstrDesc &II = TII.get(Opcode: MachineInstOpcode); |
347 | |
348 | // Make sure the input operands are sufficiently constrained to be legal |
349 | // for this instruction. |
350 | Op0 = constrainOperandRegClass(II, Op: Op0, OpNum: 1); |
351 | Op1 = constrainOperandRegClass(II, Op: Op1, OpNum: 2); |
352 | |
353 | if (II.getNumDefs() >= 1) { |
354 | AddOptionalDefs( |
355 | MIB: BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, MCID: II, DestReg: ResultReg) |
356 | .addReg(RegNo: Op0) |
357 | .addReg(RegNo: Op1)); |
358 | } else { |
359 | AddOptionalDefs(MIB: BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, MCID: II) |
360 | .addReg(RegNo: Op0) |
361 | .addReg(RegNo: Op1)); |
362 | AddOptionalDefs(MIB: BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, |
363 | MCID: TII.get(Opcode: TargetOpcode::COPY), DestReg: ResultReg) |
364 | .addReg(RegNo: II.implicit_defs()[0])); |
365 | } |
366 | return ResultReg; |
367 | } |
368 | |
369 | Register ARMFastISel::fastEmitInst_ri(unsigned MachineInstOpcode, |
370 | const TargetRegisterClass *RC, |
371 | Register Op0, uint64_t Imm) { |
372 | Register ResultReg = createResultReg(RC); |
373 | const MCInstrDesc &II = TII.get(Opcode: MachineInstOpcode); |
374 | |
375 | // Make sure the input operand is sufficiently constrained to be legal |
376 | // for this instruction. |
377 | Op0 = constrainOperandRegClass(II, Op: Op0, OpNum: 1); |
378 | if (II.getNumDefs() >= 1) { |
379 | AddOptionalDefs( |
380 | MIB: BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, MCID: II, DestReg: ResultReg) |
381 | .addReg(RegNo: Op0) |
382 | .addImm(Val: Imm)); |
383 | } else { |
384 | AddOptionalDefs(MIB: BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, MCID: II) |
385 | .addReg(RegNo: Op0) |
386 | .addImm(Val: Imm)); |
387 | AddOptionalDefs(MIB: BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, |
388 | MCID: TII.get(Opcode: TargetOpcode::COPY), DestReg: ResultReg) |
389 | .addReg(RegNo: II.implicit_defs()[0])); |
390 | } |
391 | return ResultReg; |
392 | } |
393 | |
394 | Register ARMFastISel::fastEmitInst_i(unsigned MachineInstOpcode, |
395 | const TargetRegisterClass *RC, |
396 | uint64_t Imm) { |
397 | Register ResultReg = createResultReg(RC); |
398 | const MCInstrDesc &II = TII.get(Opcode: MachineInstOpcode); |
399 | |
400 | if (II.getNumDefs() >= 1) { |
401 | AddOptionalDefs(MIB: BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, MCID: II, |
402 | DestReg: ResultReg).addImm(Val: Imm)); |
403 | } else { |
404 | AddOptionalDefs(MIB: BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, MCID: II) |
405 | .addImm(Val: Imm)); |
406 | AddOptionalDefs(MIB: BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, |
407 | MCID: TII.get(Opcode: TargetOpcode::COPY), DestReg: ResultReg) |
408 | .addReg(RegNo: II.implicit_defs()[0])); |
409 | } |
410 | return ResultReg; |
411 | } |
412 | |
413 | // TODO: Don't worry about 64-bit now, but when this is fixed remove the |
414 | // checks from the various callers. |
415 | Register ARMFastISel::ARMMoveToFPReg(MVT VT, Register SrcReg) { |
416 | if (VT == MVT::f64) |
417 | return Register(); |
418 | |
419 | Register MoveReg = createResultReg(RC: TLI.getRegClassFor(VT)); |
420 | AddOptionalDefs(MIB: BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, |
421 | MCID: TII.get(Opcode: ARM::VMOVSR), DestReg: MoveReg) |
422 | .addReg(RegNo: SrcReg)); |
423 | return MoveReg; |
424 | } |
425 | |
426 | Register ARMFastISel::ARMMoveToIntReg(MVT VT, Register SrcReg) { |
427 | if (VT == MVT::i64) |
428 | return Register(); |
429 | |
430 | Register MoveReg = createResultReg(RC: TLI.getRegClassFor(VT)); |
431 | AddOptionalDefs(MIB: BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, MIMD, |
432 | TII.get(ARM::Opcode: VMOVRS), MoveReg) |
433 | .addReg(SrcReg)); |
434 | return MoveReg; |
435 | } |
436 | |
437 | // For double width floating point we need to materialize two constants |
438 | // (the high and the low) into integer registers then use a move to get |
439 | // the combined constant into an FP reg. |
440 | Register ARMFastISel::ARMMaterializeFP(const ConstantFP *CFP, MVT VT) { |
441 | const APFloat Val = CFP->getValueAPF(); |
442 | bool is64bit = VT == MVT::f64; |
443 | |
444 | // This checks to see if we can use VFP3 instructions to materialize |
445 | // a constant, otherwise we have to go through the constant pool. |
446 | if (TLI.isFPImmLegal(Val, VT)) { |
447 | int Imm; |
448 | unsigned Opc; |
449 | if (is64bit) { |
450 | Imm = ARM_AM::getFP64Imm(FPImm: Val); |
451 | Opc = ARM::FCONSTD; |
452 | } else { |
453 | Imm = ARM_AM::getFP32Imm(FPImm: Val); |
454 | Opc = ARM::FCONSTS; |
455 | } |
456 | Register DestReg = createResultReg(RC: TLI.getRegClassFor(VT)); |
457 | AddOptionalDefs(MIB: BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, |
458 | MCID: TII.get(Opcode: Opc), DestReg).addImm(Val: Imm)); |
459 | return DestReg; |
460 | } |
461 | |
462 | // Require VFP2 for loading fp constants. |
463 | if (!Subtarget->hasVFP2Base()) return false; |
464 | |
465 | // MachineConstantPool wants an explicit alignment. |
466 | Align Alignment = DL.getPrefTypeAlign(Ty: CFP->getType()); |
467 | unsigned Idx = MCP.getConstantPoolIndex(C: cast<Constant>(Val: CFP), Alignment); |
468 | Register DestReg = createResultReg(RC: TLI.getRegClassFor(VT)); |
469 | unsigned Opc = is64bit ? ARM::VLDRD : ARM::VLDRS; |
470 | |
471 | // The extra reg is for addrmode5. |
472 | AddOptionalDefs( |
473 | MIB: BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, MCID: TII.get(Opcode: Opc), DestReg) |
474 | .addConstantPoolIndex(Idx) |
475 | .addReg(RegNo: 0)); |
476 | return DestReg; |
477 | } |
478 | |
479 | Register ARMFastISel::ARMMaterializeInt(const Constant *C, MVT VT) { |
480 | if (VT != MVT::i32 && VT != MVT::i16 && VT != MVT::i8 && VT != MVT::i1) |
481 | return Register(); |
482 | |
483 | // If we can do this in a single instruction without a constant pool entry |
484 | // do so now. |
485 | const ConstantInt *CI = cast<ConstantInt>(Val: C); |
486 | if (Subtarget->hasV6T2Ops() && isUInt<16>(x: CI->getZExtValue())) { |
487 | unsigned Opc = isThumb2 ? ARM::t2MOVi16 : ARM::MOVi16; |
488 | const TargetRegisterClass *RC = isThumb2 ? &ARM::rGPRRegClass : |
489 | &ARM::GPRRegClass; |
490 | Register ImmReg = createResultReg(RC); |
491 | AddOptionalDefs(MIB: BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, |
492 | MCID: TII.get(Opcode: Opc), DestReg: ImmReg) |
493 | .addImm(Val: CI->getZExtValue())); |
494 | return ImmReg; |
495 | } |
496 | |
497 | // Use MVN to emit negative constants. |
498 | if (VT == MVT::i32 && Subtarget->hasV6T2Ops() && CI->isNegative()) { |
499 | unsigned Imm = (unsigned)~(CI->getSExtValue()); |
500 | bool UseImm = isThumb2 ? (ARM_AM::getT2SOImmVal(Arg: Imm) != -1) : |
501 | (ARM_AM::getSOImmVal(Arg: Imm) != -1); |
502 | if (UseImm) { |
503 | unsigned Opc = isThumb2 ? ARM::t2MVNi : ARM::MVNi; |
504 | const TargetRegisterClass *RC = isThumb2 ? &ARM::rGPRRegClass : |
505 | &ARM::GPRRegClass; |
506 | Register ImmReg = createResultReg(RC); |
507 | AddOptionalDefs(MIB: BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, |
508 | MCID: TII.get(Opcode: Opc), DestReg: ImmReg) |
509 | .addImm(Val: Imm)); |
510 | return ImmReg; |
511 | } |
512 | } |
513 | |
514 | Register ResultReg; |
515 | if (Subtarget->useMovt()) |
516 | ResultReg = fastEmit_i(VT, RetVT: VT, Opcode: ISD::Constant, Imm: CI->getZExtValue()); |
517 | |
518 | if (ResultReg) |
519 | return ResultReg; |
520 | |
521 | // Load from constant pool. For now 32-bit only. |
522 | if (VT != MVT::i32) |
523 | return Register(); |
524 | |
525 | // MachineConstantPool wants an explicit alignment. |
526 | Align Alignment = DL.getPrefTypeAlign(Ty: C->getType()); |
527 | unsigned Idx = MCP.getConstantPoolIndex(C, Alignment); |
528 | ResultReg = createResultReg(RC: TLI.getRegClassFor(VT)); |
529 | if (isThumb2) |
530 | AddOptionalDefs(MIB: BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, MIMD, |
531 | TII.get(ARM::Opcode: t2LDRpci), ResultReg) |
532 | .addConstantPoolIndex(Idx)); |
533 | else { |
534 | // The extra immediate is for addrmode2. |
535 | ResultReg = constrainOperandRegClass(II: TII.get(ARM::Opcode: LDRcp), Op: ResultReg, OpNum: 0); |
536 | AddOptionalDefs(MIB: BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, MIMD, |
537 | TII.get(ARM::Opcode: LDRcp), ResultReg) |
538 | .addConstantPoolIndex(Idx) |
539 | .addImm(0)); |
540 | } |
541 | return ResultReg; |
542 | } |
543 | |
544 | bool ARMFastISel::isPositionIndependent() const { |
545 | return TLI.isPositionIndependent(); |
546 | } |
547 | |
548 | Register ARMFastISel::ARMMaterializeGV(const GlobalValue *GV, MVT VT) { |
549 | // For now 32-bit only. |
550 | if (VT != MVT::i32 || GV->isThreadLocal()) |
551 | return Register(); |
552 | |
553 | // ROPI/RWPI not currently supported. |
554 | if (Subtarget->isROPI() || Subtarget->isRWPI()) |
555 | return Register(); |
556 | |
557 | bool IsIndirect = Subtarget->isGVIndirectSymbol(GV); |
558 | const TargetRegisterClass *RC = isThumb2 ? &ARM::rGPRRegClass |
559 | : &ARM::GPRRegClass; |
560 | Register DestReg = createResultReg(RC); |
561 | |
562 | // FastISel TLS support on non-MachO is broken, punt to SelectionDAG. |
563 | const GlobalVariable *GVar = dyn_cast<GlobalVariable>(Val: GV); |
564 | bool IsThreadLocal = GVar && GVar->isThreadLocal(); |
565 | if (!Subtarget->isTargetMachO() && IsThreadLocal) |
566 | return Register(); |
567 | |
568 | bool IsPositionIndependent = isPositionIndependent(); |
569 | // Use movw+movt when possible, it avoids constant pool entries. |
570 | // Non-darwin targets only support static movt relocations in FastISel. |
571 | if (Subtarget->useMovt() && |
572 | (Subtarget->isTargetMachO() || !IsPositionIndependent)) { |
573 | unsigned Opc; |
574 | unsigned char TF = 0; |
575 | if (Subtarget->isTargetMachO()) |
576 | TF = ARMII::MO_NONLAZY; |
577 | |
578 | if (IsPositionIndependent) |
579 | Opc = isThumb2 ? ARM::t2MOV_ga_pcrel : ARM::MOV_ga_pcrel; |
580 | else |
581 | Opc = isThumb2 ? ARM::t2MOVi32imm : ARM::MOVi32imm; |
582 | AddOptionalDefs(MIB: BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, |
583 | MCID: TII.get(Opcode: Opc), DestReg).addGlobalAddress(GV, Offset: 0, TargetFlags: TF)); |
584 | } else { |
585 | // MachineConstantPool wants an explicit alignment. |
586 | Align Alignment = DL.getPrefTypeAlign(Ty: GV->getType()); |
587 | |
588 | if (Subtarget->isTargetELF() && IsPositionIndependent) |
589 | return ARMLowerPICELF(GV, VT); |
590 | |
591 | // Grab index. |
592 | unsigned PCAdj = IsPositionIndependent ? (Subtarget->isThumb() ? 4 : 8) : 0; |
593 | unsigned Id = AFI->createPICLabelUId(); |
594 | ARMConstantPoolValue *CPV = ARMConstantPoolConstant::Create(C: GV, ID: Id, |
595 | Kind: ARMCP::CPValue, |
596 | PCAdj); |
597 | unsigned Idx = MCP.getConstantPoolIndex(V: CPV, Alignment); |
598 | |
599 | // Load value. |
600 | MachineInstrBuilder MIB; |
601 | if (isThumb2) { |
602 | unsigned Opc = IsPositionIndependent ? ARM::t2LDRpci_pic : ARM::t2LDRpci; |
603 | MIB = BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, MCID: TII.get(Opcode: Opc), |
604 | DestReg).addConstantPoolIndex(Idx); |
605 | if (IsPositionIndependent) |
606 | MIB.addImm(Val: Id); |
607 | AddOptionalDefs(MIB); |
608 | } else { |
609 | // The extra immediate is for addrmode2. |
610 | DestReg = constrainOperandRegClass(TII.get(ARM::LDRcp), DestReg, 0); |
611 | MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, MIMD, |
612 | TII.get(ARM::LDRcp), DestReg) |
613 | .addConstantPoolIndex(Idx) |
614 | .addImm(0); |
615 | AddOptionalDefs(MIB); |
616 | |
617 | if (IsPositionIndependent) { |
618 | unsigned Opc = IsIndirect ? ARM::PICLDR : ARM::PICADD; |
619 | Register NewDestReg = createResultReg(RC: TLI.getRegClassFor(VT)); |
620 | |
621 | MachineInstrBuilder MIB = BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, |
622 | MIMD, MCID: TII.get(Opcode: Opc), DestReg: NewDestReg) |
623 | .addReg(RegNo: DestReg) |
624 | .addImm(Val: Id); |
625 | AddOptionalDefs(MIB); |
626 | return NewDestReg; |
627 | } |
628 | } |
629 | } |
630 | |
631 | if ((Subtarget->isTargetELF() && Subtarget->isGVInGOT(GV)) || |
632 | (Subtarget->isTargetMachO() && IsIndirect)) { |
633 | MachineInstrBuilder MIB; |
634 | Register NewDestReg = createResultReg(RC: TLI.getRegClassFor(VT)); |
635 | if (isThumb2) |
636 | MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, MIMD, |
637 | TII.get(ARM::t2LDRi12), NewDestReg) |
638 | .addReg(DestReg) |
639 | .addImm(0); |
640 | else |
641 | MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, MIMD, |
642 | TII.get(ARM::LDRi12), NewDestReg) |
643 | .addReg(DestReg) |
644 | .addImm(0); |
645 | DestReg = NewDestReg; |
646 | AddOptionalDefs(MIB); |
647 | } |
648 | |
649 | return DestReg; |
650 | } |
651 | |
652 | Register ARMFastISel::fastMaterializeConstant(const Constant *C) { |
653 | EVT CEVT = TLI.getValueType(DL, Ty: C->getType(), AllowUnknown: true); |
654 | |
655 | // Only handle simple types. |
656 | if (!CEVT.isSimple()) |
657 | return Register(); |
658 | MVT VT = CEVT.getSimpleVT(); |
659 | |
660 | if (const ConstantFP *CFP = dyn_cast<ConstantFP>(Val: C)) |
661 | return ARMMaterializeFP(CFP, VT); |
662 | else if (const GlobalValue *GV = dyn_cast<GlobalValue>(Val: C)) |
663 | return ARMMaterializeGV(GV, VT); |
664 | else if (isa<ConstantInt>(Val: C)) |
665 | return ARMMaterializeInt(C, VT); |
666 | |
667 | return Register(); |
668 | } |
669 | |
670 | // TODO: Register ARMFastISel::TargetMaterializeFloatZero(const ConstantFP *CF); |
671 | |
672 | Register ARMFastISel::fastMaterializeAlloca(const AllocaInst *AI) { |
673 | // Don't handle dynamic allocas. |
674 | if (!FuncInfo.StaticAllocaMap.count(Val: AI)) |
675 | return Register(); |
676 | |
677 | MVT VT; |
678 | if (!isLoadTypeLegal(Ty: AI->getType(), VT)) |
679 | return Register(); |
680 | |
681 | DenseMap<const AllocaInst*, int>::iterator SI = |
682 | FuncInfo.StaticAllocaMap.find(Val: AI); |
683 | |
684 | // This will get lowered later into the correct offsets and registers |
685 | // via rewriteXFrameIndex. |
686 | if (SI != FuncInfo.StaticAllocaMap.end()) { |
687 | unsigned Opc = isThumb2 ? ARM::t2ADDri : ARM::ADDri; |
688 | const TargetRegisterClass* RC = TLI.getRegClassFor(VT); |
689 | Register ResultReg = createResultReg(RC); |
690 | ResultReg = constrainOperandRegClass(II: TII.get(Opcode: Opc), Op: ResultReg, OpNum: 0); |
691 | |
692 | AddOptionalDefs(MIB: BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, |
693 | MCID: TII.get(Opcode: Opc), DestReg: ResultReg) |
694 | .addFrameIndex(Idx: SI->second) |
695 | .addImm(Val: 0)); |
696 | return ResultReg; |
697 | } |
698 | |
699 | return Register(); |
700 | } |
701 | |
702 | bool ARMFastISel::isTypeLegal(Type *Ty, MVT &VT) { |
703 | EVT evt = TLI.getValueType(DL, Ty, AllowUnknown: true); |
704 | |
705 | // Only handle simple types. |
706 | if (evt == MVT::Other || !evt.isSimple()) return false; |
707 | VT = evt.getSimpleVT(); |
708 | |
709 | // Handle all legal types, i.e. a register that will directly hold this |
710 | // value. |
711 | return TLI.isTypeLegal(VT); |
712 | } |
713 | |
714 | bool ARMFastISel::isLoadTypeLegal(Type *Ty, MVT &VT) { |
715 | if (isTypeLegal(Ty, VT)) return true; |
716 | |
717 | // If this is a type than can be sign or zero-extended to a basic operation |
718 | // go ahead and accept it now. |
719 | if (VT == MVT::i1 || VT == MVT::i8 || VT == MVT::i16) |
720 | return true; |
721 | |
722 | return false; |
723 | } |
724 | |
725 | // Computes the address to get to an object. |
726 | bool ARMFastISel::ARMComputeAddress(const Value *Obj, Address &Addr) { |
727 | // Some boilerplate from the X86 FastISel. |
728 | const User *U = nullptr; |
729 | unsigned Opcode = Instruction::UserOp1; |
730 | if (const Instruction *I = dyn_cast<Instruction>(Val: Obj)) { |
731 | // Don't walk into other basic blocks unless the object is an alloca from |
732 | // another block, otherwise it may not have a virtual register assigned. |
733 | if (FuncInfo.StaticAllocaMap.count(Val: static_cast<const AllocaInst *>(Obj)) || |
734 | FuncInfo.getMBB(BB: I->getParent()) == FuncInfo.MBB) { |
735 | Opcode = I->getOpcode(); |
736 | U = I; |
737 | } |
738 | } else if (const ConstantExpr *C = dyn_cast<ConstantExpr>(Val: Obj)) { |
739 | Opcode = C->getOpcode(); |
740 | U = C; |
741 | } |
742 | |
743 | if (PointerType *Ty = dyn_cast<PointerType>(Val: Obj->getType())) |
744 | if (Ty->getAddressSpace() > 255) |
745 | // Fast instruction selection doesn't support the special |
746 | // address spaces. |
747 | return false; |
748 | |
749 | switch (Opcode) { |
750 | default: |
751 | break; |
752 | case Instruction::BitCast: |
753 | // Look through bitcasts. |
754 | return ARMComputeAddress(Obj: U->getOperand(i: 0), Addr); |
755 | case Instruction::IntToPtr: |
756 | // Look past no-op inttoptrs. |
757 | if (TLI.getValueType(DL, Ty: U->getOperand(i: 0)->getType()) == |
758 | TLI.getPointerTy(DL)) |
759 | return ARMComputeAddress(Obj: U->getOperand(i: 0), Addr); |
760 | break; |
761 | case Instruction::PtrToInt: |
762 | // Look past no-op ptrtoints. |
763 | if (TLI.getValueType(DL, Ty: U->getType()) == TLI.getPointerTy(DL)) |
764 | return ARMComputeAddress(Obj: U->getOperand(i: 0), Addr); |
765 | break; |
766 | case Instruction::GetElementPtr: { |
767 | Address SavedAddr = Addr; |
768 | int TmpOffset = Addr.getOffset(); |
769 | |
770 | // Iterate through the GEP folding the constants into offsets where |
771 | // we can. |
772 | gep_type_iterator GTI = gep_type_begin(GEP: U); |
773 | for (User::const_op_iterator i = U->op_begin() + 1, e = U->op_end(); |
774 | i != e; ++i, ++GTI) { |
775 | const Value *Op = *i; |
776 | if (StructType *STy = GTI.getStructTypeOrNull()) { |
777 | const StructLayout *SL = DL.getStructLayout(Ty: STy); |
778 | unsigned Idx = cast<ConstantInt>(Val: Op)->getZExtValue(); |
779 | TmpOffset += SL->getElementOffset(Idx); |
780 | } else { |
781 | uint64_t S = GTI.getSequentialElementStride(DL); |
782 | while (true) { |
783 | if (const ConstantInt *CI = dyn_cast<ConstantInt>(Val: Op)) { |
784 | // Constant-offset addressing. |
785 | TmpOffset += CI->getSExtValue() * S; |
786 | break; |
787 | } |
788 | if (canFoldAddIntoGEP(GEP: U, Add: Op)) { |
789 | // A compatible add with a constant operand. Fold the constant. |
790 | ConstantInt *CI = |
791 | cast<ConstantInt>(Val: cast<AddOperator>(Val: Op)->getOperand(i_nocapture: 1)); |
792 | TmpOffset += CI->getSExtValue() * S; |
793 | // Iterate on the other operand. |
794 | Op = cast<AddOperator>(Val: Op)->getOperand(i_nocapture: 0); |
795 | continue; |
796 | } |
797 | // Unsupported |
798 | goto unsupported_gep; |
799 | } |
800 | } |
801 | } |
802 | |
803 | // Try to grab the base operand now. |
804 | Addr.setOffset(TmpOffset); |
805 | if (ARMComputeAddress(Obj: U->getOperand(i: 0), Addr)) return true; |
806 | |
807 | // We failed, restore everything and try the other options. |
808 | Addr = SavedAddr; |
809 | |
810 | unsupported_gep: |
811 | break; |
812 | } |
813 | case Instruction::Alloca: { |
814 | const AllocaInst *AI = cast<AllocaInst>(Val: Obj); |
815 | DenseMap<const AllocaInst*, int>::iterator SI = |
816 | FuncInfo.StaticAllocaMap.find(Val: AI); |
817 | if (SI != FuncInfo.StaticAllocaMap.end()) { |
818 | Addr.setKind(Address::FrameIndexBase); |
819 | Addr.setFI(SI->second); |
820 | return true; |
821 | } |
822 | break; |
823 | } |
824 | } |
825 | |
826 | // Try to get this in a register if nothing else has worked. |
827 | if (!Addr.getReg()) |
828 | Addr.setReg(getRegForValue(V: Obj)); |
829 | return Addr.getReg(); |
830 | } |
831 | |
832 | void ARMFastISel::ARMSimplifyAddress(Address &Addr, MVT VT, bool useAM3) { |
833 | bool needsLowering = false; |
834 | switch (VT.SimpleTy) { |
835 | default: llvm_unreachable("Unhandled load/store type!"); |
836 | case MVT::i1: |
837 | case MVT::i8: |
838 | case MVT::i16: |
839 | case MVT::i32: |
840 | if (!useAM3) { |
841 | // Integer loads/stores handle 12-bit offsets. |
842 | needsLowering = ((Addr.getOffset() & 0xfff) != Addr.getOffset()); |
843 | // Handle negative offsets. |
844 | if (needsLowering && isThumb2) |
845 | needsLowering = !(Subtarget->hasV6T2Ops() && Addr.getOffset() < 0 && |
846 | Addr.getOffset() > -256); |
847 | } else { |
848 | // ARM halfword load/stores and signed byte loads use +/-imm8 offsets. |
849 | needsLowering = (Addr.getOffset() > 255 || Addr.getOffset() < -255); |
850 | } |
851 | break; |
852 | case MVT::f32: |
853 | case MVT::f64: |
854 | // Floating point operands handle 8-bit offsets. |
855 | needsLowering = ((Addr.getOffset() & 0xff) != Addr.getOffset()); |
856 | break; |
857 | } |
858 | |
859 | // If this is a stack pointer and the offset needs to be simplified then |
860 | // put the alloca address into a register, set the base type back to |
861 | // register and continue. This should almost never happen. |
862 | if (needsLowering && Addr.isFIBase()) { |
863 | const TargetRegisterClass *RC = isThumb2 ? &ARM::tGPRRegClass |
864 | : &ARM::GPRRegClass; |
865 | Register ResultReg = createResultReg(RC); |
866 | unsigned Opc = isThumb2 ? ARM::t2ADDri : ARM::ADDri; |
867 | AddOptionalDefs( |
868 | MIB: BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, MCID: TII.get(Opcode: Opc), DestReg: ResultReg) |
869 | .addFrameIndex(Idx: Addr.getFI()) |
870 | .addImm(Val: 0)); |
871 | Addr.setKind(Address::RegBase); |
872 | Addr.setReg(ResultReg); |
873 | } |
874 | |
875 | // Since the offset is too large for the load/store instruction |
876 | // get the reg+offset into a register. |
877 | if (needsLowering) { |
878 | Addr.setReg(fastEmit_ri_(MVT::i32, ISD::ADD, Addr.getReg(), |
879 | Addr.getOffset(), MVT::i32)); |
880 | Addr.setOffset(0); |
881 | } |
882 | } |
883 | |
884 | void ARMFastISel::AddLoadStoreOperands(MVT VT, Address &Addr, |
885 | const MachineInstrBuilder &MIB, |
886 | MachineMemOperand::Flags Flags, |
887 | bool useAM3) { |
888 | // addrmode5 output depends on the selection dag addressing dividing the |
889 | // offset by 4 that it then later multiplies. Do this here as well. |
890 | if (VT.SimpleTy == MVT::f32 || VT.SimpleTy == MVT::f64) |
891 | Addr.setOffset(Addr.getOffset() / 4); |
892 | |
893 | // Frame base works a bit differently. Handle it separately. |
894 | if (Addr.isFIBase()) { |
895 | int FI = Addr.getFI(); |
896 | int Offset = Addr.getOffset(); |
897 | MachineMemOperand *MMO = FuncInfo.MF->getMachineMemOperand( |
898 | PtrInfo: MachinePointerInfo::getFixedStack(MF&: *FuncInfo.MF, FI, Offset), F: Flags, |
899 | Size: MFI.getObjectSize(ObjectIdx: FI), BaseAlignment: MFI.getObjectAlign(ObjectIdx: FI)); |
900 | // Now add the rest of the operands. |
901 | MIB.addFrameIndex(Idx: FI); |
902 | |
903 | // ARM halfword load/stores and signed byte loads need an additional |
904 | // operand. |
905 | if (useAM3) { |
906 | int Imm = (Addr.getOffset() < 0) ? (0x100 | -Addr.getOffset()) |
907 | : Addr.getOffset(); |
908 | MIB.addReg(RegNo: 0); |
909 | MIB.addImm(Val: Imm); |
910 | } else { |
911 | MIB.addImm(Val: Addr.getOffset()); |
912 | } |
913 | MIB.addMemOperand(MMO); |
914 | } else { |
915 | // Now add the rest of the operands. |
916 | MIB.addReg(RegNo: Addr.getReg()); |
917 | |
918 | // ARM halfword load/stores and signed byte loads need an additional |
919 | // operand. |
920 | if (useAM3) { |
921 | int Imm = (Addr.getOffset() < 0) ? (0x100 | -Addr.getOffset()) |
922 | : Addr.getOffset(); |
923 | MIB.addReg(RegNo: 0); |
924 | MIB.addImm(Val: Imm); |
925 | } else { |
926 | MIB.addImm(Val: Addr.getOffset()); |
927 | } |
928 | } |
929 | AddOptionalDefs(MIB); |
930 | } |
931 | |
932 | bool ARMFastISel::ARMEmitLoad(MVT VT, Register &ResultReg, Address &Addr, |
933 | MaybeAlign Alignment, bool isZExt, |
934 | bool allocReg) { |
935 | unsigned Opc; |
936 | bool useAM3 = false; |
937 | bool needVMOV = false; |
938 | const TargetRegisterClass *RC; |
939 | switch (VT.SimpleTy) { |
940 | // This is mostly going to be Neon/vector support. |
941 | default: return false; |
942 | case MVT::i1: |
943 | case MVT::i8: |
944 | if (isThumb2) { |
945 | if (Addr.getOffset() < 0 && Addr.getOffset() > -256 && |
946 | Subtarget->hasV6T2Ops()) |
947 | Opc = isZExt ? ARM::t2LDRBi8 : ARM::t2LDRSBi8; |
948 | else |
949 | Opc = isZExt ? ARM::t2LDRBi12 : ARM::t2LDRSBi12; |
950 | } else { |
951 | if (isZExt) { |
952 | Opc = ARM::LDRBi12; |
953 | } else { |
954 | Opc = ARM::LDRSB; |
955 | useAM3 = true; |
956 | } |
957 | } |
958 | RC = isThumb2 ? &ARM::rGPRRegClass : &ARM::GPRnopcRegClass; |
959 | break; |
960 | case MVT::i16: |
961 | if (Alignment && *Alignment < Align(2) && |
962 | !Subtarget->allowsUnalignedMem()) |
963 | return false; |
964 | |
965 | if (isThumb2) { |
966 | if (Addr.getOffset() < 0 && Addr.getOffset() > -256 && |
967 | Subtarget->hasV6T2Ops()) |
968 | Opc = isZExt ? ARM::t2LDRHi8 : ARM::t2LDRSHi8; |
969 | else |
970 | Opc = isZExt ? ARM::t2LDRHi12 : ARM::t2LDRSHi12; |
971 | } else { |
972 | Opc = isZExt ? ARM::LDRH : ARM::LDRSH; |
973 | useAM3 = true; |
974 | } |
975 | RC = isThumb2 ? &ARM::rGPRRegClass : &ARM::GPRnopcRegClass; |
976 | break; |
977 | case MVT::i32: |
978 | if (Alignment && *Alignment < Align(4) && |
979 | !Subtarget->allowsUnalignedMem()) |
980 | return false; |
981 | |
982 | if (isThumb2) { |
983 | if (Addr.getOffset() < 0 && Addr.getOffset() > -256 && |
984 | Subtarget->hasV6T2Ops()) |
985 | Opc = ARM::t2LDRi8; |
986 | else |
987 | Opc = ARM::t2LDRi12; |
988 | } else { |
989 | Opc = ARM::LDRi12; |
990 | } |
991 | RC = isThumb2 ? &ARM::rGPRRegClass : &ARM::GPRnopcRegClass; |
992 | break; |
993 | case MVT::f32: |
994 | if (!Subtarget->hasVFP2Base()) return false; |
995 | // Unaligned loads need special handling. Floats require word-alignment. |
996 | if (Alignment && *Alignment < Align(4)) { |
997 | needVMOV = true; |
998 | VT = MVT::i32; |
999 | Opc = isThumb2 ? ARM::t2LDRi12 : ARM::LDRi12; |
1000 | RC = isThumb2 ? &ARM::rGPRRegClass : &ARM::GPRnopcRegClass; |
1001 | } else { |
1002 | Opc = ARM::VLDRS; |
1003 | RC = TLI.getRegClassFor(VT); |
1004 | } |
1005 | break; |
1006 | case MVT::f64: |
1007 | // Can load and store double precision even without FeatureFP64 |
1008 | if (!Subtarget->hasVFP2Base()) return false; |
1009 | // FIXME: Unaligned loads need special handling. Doublewords require |
1010 | // word-alignment. |
1011 | if (Alignment && *Alignment < Align(4)) |
1012 | return false; |
1013 | |
1014 | Opc = ARM::VLDRD; |
1015 | RC = TLI.getRegClassFor(VT); |
1016 | break; |
1017 | } |
1018 | // Simplify this down to something we can handle. |
1019 | ARMSimplifyAddress(Addr, VT, useAM3); |
1020 | |
1021 | // Create the base instruction, then add the operands. |
1022 | if (allocReg) |
1023 | ResultReg = createResultReg(RC); |
1024 | assert(ResultReg.isVirtual() && "Expected an allocated virtual register."); |
1025 | MachineInstrBuilder MIB = BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, |
1026 | MCID: TII.get(Opcode: Opc), DestReg: ResultReg); |
1027 | AddLoadStoreOperands(VT, Addr, MIB, Flags: MachineMemOperand::MOLoad, useAM3); |
1028 | |
1029 | // If we had an unaligned load of a float we've converted it to an regular |
1030 | // load. Now we must move from the GRP to the FP register. |
1031 | if (needVMOV) { |
1032 | Register MoveReg = createResultReg(TLI.getRegClassFor(MVT::f32)); |
1033 | AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, MIMD, |
1034 | TII.get(ARM::VMOVSR), MoveReg) |
1035 | .addReg(ResultReg)); |
1036 | ResultReg = MoveReg; |
1037 | } |
1038 | return true; |
1039 | } |
1040 | |
1041 | bool ARMFastISel::SelectLoad(const Instruction *I) { |
1042 | // Atomic loads need special handling. |
1043 | if (cast<LoadInst>(Val: I)->isAtomic()) |
1044 | return false; |
1045 | |
1046 | const Value *SV = I->getOperand(i: 0); |
1047 | if (TLI.supportSwiftError()) { |
1048 | // Swifterror values can come from either a function parameter with |
1049 | // swifterror attribute or an alloca with swifterror attribute. |
1050 | if (const Argument *Arg = dyn_cast<Argument>(Val: SV)) { |
1051 | if (Arg->hasSwiftErrorAttr()) |
1052 | return false; |
1053 | } |
1054 | |
1055 | if (const AllocaInst *Alloca = dyn_cast<AllocaInst>(Val: SV)) { |
1056 | if (Alloca->isSwiftError()) |
1057 | return false; |
1058 | } |
1059 | } |
1060 | |
1061 | // Verify we have a legal type before going any further. |
1062 | MVT VT; |
1063 | if (!isLoadTypeLegal(Ty: I->getType(), VT)) |
1064 | return false; |
1065 | |
1066 | // See if we can handle this address. |
1067 | Address Addr; |
1068 | if (!ARMComputeAddress(Obj: I->getOperand(i: 0), Addr)) return false; |
1069 | |
1070 | Register ResultReg; |
1071 | if (!ARMEmitLoad(VT, ResultReg, Addr, Alignment: cast<LoadInst>(Val: I)->getAlign())) |
1072 | return false; |
1073 | updateValueMap(I, Reg: ResultReg); |
1074 | return true; |
1075 | } |
1076 | |
1077 | bool ARMFastISel::ARMEmitStore(MVT VT, Register SrcReg, Address &Addr, |
1078 | MaybeAlign Alignment) { |
1079 | unsigned StrOpc; |
1080 | bool useAM3 = false; |
1081 | switch (VT.SimpleTy) { |
1082 | // This is mostly going to be Neon/vector support. |
1083 | default: return false; |
1084 | case MVT::i1: { |
1085 | Register Res = createResultReg(isThumb2 ? &ARM::tGPRRegClass |
1086 | : &ARM::GPRRegClass); |
1087 | unsigned Opc = isThumb2 ? ARM::t2ANDri : ARM::ANDri; |
1088 | SrcReg = constrainOperandRegClass(II: TII.get(Opcode: Opc), Op: SrcReg, OpNum: 1); |
1089 | AddOptionalDefs(MIB: BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, |
1090 | MCID: TII.get(Opcode: Opc), DestReg: Res) |
1091 | .addReg(RegNo: SrcReg).addImm(Val: 1)); |
1092 | SrcReg = Res; |
1093 | [[fallthrough]]; |
1094 | } |
1095 | case MVT::i8: |
1096 | if (isThumb2) { |
1097 | if (Addr.getOffset() < 0 && Addr.getOffset() > -256 && |
1098 | Subtarget->hasV6T2Ops()) |
1099 | StrOpc = ARM::t2STRBi8; |
1100 | else |
1101 | StrOpc = ARM::t2STRBi12; |
1102 | } else { |
1103 | StrOpc = ARM::STRBi12; |
1104 | } |
1105 | break; |
1106 | case MVT::i16: |
1107 | if (Alignment && *Alignment < Align(2) && |
1108 | !Subtarget->allowsUnalignedMem()) |
1109 | return false; |
1110 | |
1111 | if (isThumb2) { |
1112 | if (Addr.getOffset() < 0 && Addr.getOffset() > -256 && |
1113 | Subtarget->hasV6T2Ops()) |
1114 | StrOpc = ARM::t2STRHi8; |
1115 | else |
1116 | StrOpc = ARM::t2STRHi12; |
1117 | } else { |
1118 | StrOpc = ARM::STRH; |
1119 | useAM3 = true; |
1120 | } |
1121 | break; |
1122 | case MVT::i32: |
1123 | if (Alignment && *Alignment < Align(4) && |
1124 | !Subtarget->allowsUnalignedMem()) |
1125 | return false; |
1126 | |
1127 | if (isThumb2) { |
1128 | if (Addr.getOffset() < 0 && Addr.getOffset() > -256 && |
1129 | Subtarget->hasV6T2Ops()) |
1130 | StrOpc = ARM::t2STRi8; |
1131 | else |
1132 | StrOpc = ARM::t2STRi12; |
1133 | } else { |
1134 | StrOpc = ARM::STRi12; |
1135 | } |
1136 | break; |
1137 | case MVT::f32: |
1138 | if (!Subtarget->hasVFP2Base()) return false; |
1139 | // Unaligned stores need special handling. Floats require word-alignment. |
1140 | if (Alignment && *Alignment < Align(4)) { |
1141 | Register MoveReg = createResultReg(TLI.getRegClassFor(MVT::i32)); |
1142 | AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, MIMD, |
1143 | TII.get(ARM::VMOVRS), MoveReg) |
1144 | .addReg(SrcReg)); |
1145 | SrcReg = MoveReg; |
1146 | VT = MVT::i32; |
1147 | StrOpc = isThumb2 ? ARM::t2STRi12 : ARM::STRi12; |
1148 | } else { |
1149 | StrOpc = ARM::VSTRS; |
1150 | } |
1151 | break; |
1152 | case MVT::f64: |
1153 | // Can load and store double precision even without FeatureFP64 |
1154 | if (!Subtarget->hasVFP2Base()) return false; |
1155 | // FIXME: Unaligned stores need special handling. Doublewords require |
1156 | // word-alignment. |
1157 | if (Alignment && *Alignment < Align(4)) |
1158 | return false; |
1159 | |
1160 | StrOpc = ARM::VSTRD; |
1161 | break; |
1162 | } |
1163 | // Simplify this down to something we can handle. |
1164 | ARMSimplifyAddress(Addr, VT, useAM3); |
1165 | |
1166 | // Create the base instruction, then add the operands. |
1167 | SrcReg = constrainOperandRegClass(II: TII.get(Opcode: StrOpc), Op: SrcReg, OpNum: 0); |
1168 | MachineInstrBuilder MIB = BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, |
1169 | MCID: TII.get(Opcode: StrOpc)) |
1170 | .addReg(RegNo: SrcReg); |
1171 | AddLoadStoreOperands(VT, Addr, MIB, Flags: MachineMemOperand::MOStore, useAM3); |
1172 | return true; |
1173 | } |
1174 | |
1175 | bool ARMFastISel::SelectStore(const Instruction *I) { |
1176 | Value *Op0 = I->getOperand(i: 0); |
1177 | Register SrcReg; |
1178 | |
1179 | // Atomic stores need special handling. |
1180 | if (cast<StoreInst>(Val: I)->isAtomic()) |
1181 | return false; |
1182 | |
1183 | const Value *PtrV = I->getOperand(i: 1); |
1184 | if (TLI.supportSwiftError()) { |
1185 | // Swifterror values can come from either a function parameter with |
1186 | // swifterror attribute or an alloca with swifterror attribute. |
1187 | if (const Argument *Arg = dyn_cast<Argument>(Val: PtrV)) { |
1188 | if (Arg->hasSwiftErrorAttr()) |
1189 | return false; |
1190 | } |
1191 | |
1192 | if (const AllocaInst *Alloca = dyn_cast<AllocaInst>(Val: PtrV)) { |
1193 | if (Alloca->isSwiftError()) |
1194 | return false; |
1195 | } |
1196 | } |
1197 | |
1198 | // Verify we have a legal type before going any further. |
1199 | MVT VT; |
1200 | if (!isLoadTypeLegal(Ty: I->getOperand(i: 0)->getType(), VT)) |
1201 | return false; |
1202 | |
1203 | // Get the value to be stored into a register. |
1204 | SrcReg = getRegForValue(V: Op0); |
1205 | if (!SrcReg) |
1206 | return false; |
1207 | |
1208 | // See if we can handle this address. |
1209 | Address Addr; |
1210 | if (!ARMComputeAddress(Obj: I->getOperand(i: 1), Addr)) |
1211 | return false; |
1212 | |
1213 | if (!ARMEmitStore(VT, SrcReg, Addr, Alignment: cast<StoreInst>(Val: I)->getAlign())) |
1214 | return false; |
1215 | return true; |
1216 | } |
1217 | |
1218 | static ARMCC::CondCodes getComparePred(CmpInst::Predicate Pred) { |
1219 | switch (Pred) { |
1220 | // Needs two compares... |
1221 | case CmpInst::FCMP_ONE: |
1222 | case CmpInst::FCMP_UEQ: |
1223 | default: |
1224 | // AL is our "false" for now. The other two need more compares. |
1225 | return ARMCC::AL; |
1226 | case CmpInst::ICMP_EQ: |
1227 | case CmpInst::FCMP_OEQ: |
1228 | return ARMCC::EQ; |
1229 | case CmpInst::ICMP_SGT: |
1230 | case CmpInst::FCMP_OGT: |
1231 | return ARMCC::GT; |
1232 | case CmpInst::ICMP_SGE: |
1233 | case CmpInst::FCMP_OGE: |
1234 | return ARMCC::GE; |
1235 | case CmpInst::ICMP_UGT: |
1236 | case CmpInst::FCMP_UGT: |
1237 | return ARMCC::HI; |
1238 | case CmpInst::FCMP_OLT: |
1239 | return ARMCC::MI; |
1240 | case CmpInst::ICMP_ULE: |
1241 | case CmpInst::FCMP_OLE: |
1242 | return ARMCC::LS; |
1243 | case CmpInst::FCMP_ORD: |
1244 | return ARMCC::VC; |
1245 | case CmpInst::FCMP_UNO: |
1246 | return ARMCC::VS; |
1247 | case CmpInst::FCMP_UGE: |
1248 | return ARMCC::PL; |
1249 | case CmpInst::ICMP_SLT: |
1250 | case CmpInst::FCMP_ULT: |
1251 | return ARMCC::LT; |
1252 | case CmpInst::ICMP_SLE: |
1253 | case CmpInst::FCMP_ULE: |
1254 | return ARMCC::LE; |
1255 | case CmpInst::FCMP_UNE: |
1256 | case CmpInst::ICMP_NE: |
1257 | return ARMCC::NE; |
1258 | case CmpInst::ICMP_UGE: |
1259 | return ARMCC::HS; |
1260 | case CmpInst::ICMP_ULT: |
1261 | return ARMCC::LO; |
1262 | } |
1263 | } |
1264 | |
1265 | bool ARMFastISel::SelectBranch(const Instruction *I) { |
1266 | const BranchInst *BI = cast<BranchInst>(Val: I); |
1267 | MachineBasicBlock *TBB = FuncInfo.getMBB(BB: BI->getSuccessor(i: 0)); |
1268 | MachineBasicBlock *FBB = FuncInfo.getMBB(BB: BI->getSuccessor(i: 1)); |
1269 | |
1270 | // Simple branch support. |
1271 | |
1272 | // If we can, avoid recomputing the compare - redoing it could lead to wonky |
1273 | // behavior. |
1274 | if (const CmpInst *CI = dyn_cast<CmpInst>(Val: BI->getCondition())) { |
1275 | if (CI->hasOneUse() && (CI->getParent() == I->getParent())) { |
1276 | // Get the compare predicate. |
1277 | // Try to take advantage of fallthrough opportunities. |
1278 | CmpInst::Predicate Predicate = CI->getPredicate(); |
1279 | if (FuncInfo.MBB->isLayoutSuccessor(MBB: TBB)) { |
1280 | std::swap(a&: TBB, b&: FBB); |
1281 | Predicate = CmpInst::getInversePredicate(pred: Predicate); |
1282 | } |
1283 | |
1284 | ARMCC::CondCodes ARMPred = getComparePred(Pred: Predicate); |
1285 | |
1286 | // We may not handle every CC for now. |
1287 | if (ARMPred == ARMCC::AL) return false; |
1288 | |
1289 | // Emit the compare. |
1290 | if (!ARMEmitCmp(Src1Value: CI->getOperand(i_nocapture: 0), Src2Value: CI->getOperand(i_nocapture: 1), isZExt: CI->isUnsigned())) |
1291 | return false; |
1292 | |
1293 | unsigned BrOpc = isThumb2 ? ARM::t2Bcc : ARM::Bcc; |
1294 | BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, MIMD, TII.get(BrOpc)) |
1295 | .addMBB(TBB).addImm(ARMPred).addReg(ARM::CPSR); |
1296 | finishCondBranch(BranchBB: BI->getParent(), TrueMBB: TBB, FalseMBB: FBB); |
1297 | return true; |
1298 | } |
1299 | } else if (TruncInst *TI = dyn_cast<TruncInst>(Val: BI->getCondition())) { |
1300 | MVT SourceVT; |
1301 | if (TI->hasOneUse() && TI->getParent() == I->getParent() && |
1302 | (isLoadTypeLegal(Ty: TI->getOperand(i_nocapture: 0)->getType(), VT&: SourceVT))) { |
1303 | unsigned TstOpc = isThumb2 ? ARM::t2TSTri : ARM::TSTri; |
1304 | Register OpReg = getRegForValue(V: TI->getOperand(i_nocapture: 0)); |
1305 | OpReg = constrainOperandRegClass(II: TII.get(Opcode: TstOpc), Op: OpReg, OpNum: 0); |
1306 | AddOptionalDefs(MIB: BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, |
1307 | MCID: TII.get(Opcode: TstOpc)) |
1308 | .addReg(RegNo: OpReg).addImm(Val: 1)); |
1309 | |
1310 | unsigned CCMode = ARMCC::NE; |
1311 | if (FuncInfo.MBB->isLayoutSuccessor(MBB: TBB)) { |
1312 | std::swap(a&: TBB, b&: FBB); |
1313 | CCMode = ARMCC::EQ; |
1314 | } |
1315 | |
1316 | unsigned BrOpc = isThumb2 ? ARM::t2Bcc : ARM::Bcc; |
1317 | BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, MIMD, TII.get(BrOpc)) |
1318 | .addMBB(TBB).addImm(CCMode).addReg(ARM::CPSR); |
1319 | |
1320 | finishCondBranch(BranchBB: BI->getParent(), TrueMBB: TBB, FalseMBB: FBB); |
1321 | return true; |
1322 | } |
1323 | } else if (const ConstantInt *CI = |
1324 | dyn_cast<ConstantInt>(Val: BI->getCondition())) { |
1325 | uint64_t Imm = CI->getZExtValue(); |
1326 | MachineBasicBlock *Target = (Imm == 0) ? FBB : TBB; |
1327 | fastEmitBranch(MSucc: Target, DbgLoc: MIMD.getDL()); |
1328 | return true; |
1329 | } |
1330 | |
1331 | Register CmpReg = getRegForValue(V: BI->getCondition()); |
1332 | if (!CmpReg) |
1333 | return false; |
1334 | |
1335 | // We've been divorced from our compare! Our block was split, and |
1336 | // now our compare lives in a predecessor block. We musn't |
1337 | // re-compare here, as the children of the compare aren't guaranteed |
1338 | // live across the block boundary (we *could* check for this). |
1339 | // Regardless, the compare has been done in the predecessor block, |
1340 | // and it left a value for us in a virtual register. Ergo, we test |
1341 | // the one-bit value left in the virtual register. |
1342 | unsigned TstOpc = isThumb2 ? ARM::t2TSTri : ARM::TSTri; |
1343 | CmpReg = constrainOperandRegClass(II: TII.get(Opcode: TstOpc), Op: CmpReg, OpNum: 0); |
1344 | AddOptionalDefs( |
1345 | MIB: BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, MCID: TII.get(Opcode: TstOpc)) |
1346 | .addReg(RegNo: CmpReg) |
1347 | .addImm(Val: 1)); |
1348 | |
1349 | unsigned CCMode = ARMCC::NE; |
1350 | if (FuncInfo.MBB->isLayoutSuccessor(MBB: TBB)) { |
1351 | std::swap(a&: TBB, b&: FBB); |
1352 | CCMode = ARMCC::EQ; |
1353 | } |
1354 | |
1355 | unsigned BrOpc = isThumb2 ? ARM::t2Bcc : ARM::Bcc; |
1356 | BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, MIMD, TII.get(BrOpc)) |
1357 | .addMBB(TBB).addImm(CCMode).addReg(ARM::CPSR); |
1358 | finishCondBranch(BranchBB: BI->getParent(), TrueMBB: TBB, FalseMBB: FBB); |
1359 | return true; |
1360 | } |
1361 | |
1362 | bool ARMFastISel::SelectIndirectBr(const Instruction *I) { |
1363 | Register AddrReg = getRegForValue(V: I->getOperand(i: 0)); |
1364 | if (!AddrReg) |
1365 | return false; |
1366 | |
1367 | unsigned Opc = isThumb2 ? ARM::tBRIND : ARM::BX; |
1368 | assert(isThumb2 || Subtarget->hasV4TOps()); |
1369 | |
1370 | AddOptionalDefs(MIB: BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, |
1371 | MCID: TII.get(Opcode: Opc)).addReg(RegNo: AddrReg)); |
1372 | |
1373 | const IndirectBrInst *IB = cast<IndirectBrInst>(Val: I); |
1374 | for (const BasicBlock *SuccBB : IB->successors()) |
1375 | FuncInfo.MBB->addSuccessor(Succ: FuncInfo.getMBB(BB: SuccBB)); |
1376 | |
1377 | return true; |
1378 | } |
1379 | |
1380 | bool ARMFastISel::ARMEmitCmp(const Value *Src1Value, const Value *Src2Value, |
1381 | bool isZExt) { |
1382 | Type *Ty = Src1Value->getType(); |
1383 | EVT SrcEVT = TLI.getValueType(DL, Ty, AllowUnknown: true); |
1384 | if (!SrcEVT.isSimple()) return false; |
1385 | MVT SrcVT = SrcEVT.getSimpleVT(); |
1386 | |
1387 | if (Ty->isFloatTy() && !Subtarget->hasVFP2Base()) |
1388 | return false; |
1389 | |
1390 | if (Ty->isDoubleTy() && (!Subtarget->hasVFP2Base() || !Subtarget->hasFP64())) |
1391 | return false; |
1392 | |
1393 | // Check to see if the 2nd operand is a constant that we can encode directly |
1394 | // in the compare. |
1395 | int Imm = 0; |
1396 | bool UseImm = false; |
1397 | bool isNegativeImm = false; |
1398 | // FIXME: At -O0 we don't have anything that canonicalizes operand order. |
1399 | // Thus, Src1Value may be a ConstantInt, but we're missing it. |
1400 | if (const ConstantInt *ConstInt = dyn_cast<ConstantInt>(Val: Src2Value)) { |
1401 | if (SrcVT == MVT::i32 || SrcVT == MVT::i16 || SrcVT == MVT::i8 || |
1402 | SrcVT == MVT::i1) { |
1403 | const APInt &CIVal = ConstInt->getValue(); |
1404 | Imm = (isZExt) ? (int)CIVal.getZExtValue() : (int)CIVal.getSExtValue(); |
1405 | // For INT_MIN/LONG_MIN (i.e., 0x80000000) we need to use a cmp, rather |
1406 | // then a cmn, because there is no way to represent 2147483648 as a |
1407 | // signed 32-bit int. |
1408 | if (Imm < 0 && Imm != (int)0x80000000) { |
1409 | isNegativeImm = true; |
1410 | Imm = -Imm; |
1411 | } |
1412 | UseImm = isThumb2 ? (ARM_AM::getT2SOImmVal(Arg: Imm) != -1) : |
1413 | (ARM_AM::getSOImmVal(Arg: Imm) != -1); |
1414 | } |
1415 | } else if (const ConstantFP *ConstFP = dyn_cast<ConstantFP>(Val: Src2Value)) { |
1416 | if (SrcVT == MVT::f32 || SrcVT == MVT::f64) |
1417 | if (ConstFP->isZero() && !ConstFP->isNegative()) |
1418 | UseImm = true; |
1419 | } |
1420 | |
1421 | unsigned CmpOpc; |
1422 | bool isICmp = true; |
1423 | bool needsExt = false; |
1424 | switch (SrcVT.SimpleTy) { |
1425 | default: return false; |
1426 | // TODO: Verify compares. |
1427 | case MVT::f32: |
1428 | isICmp = false; |
1429 | CmpOpc = UseImm ? ARM::VCMPZS : ARM::VCMPS; |
1430 | break; |
1431 | case MVT::f64: |
1432 | isICmp = false; |
1433 | CmpOpc = UseImm ? ARM::VCMPZD : ARM::VCMPD; |
1434 | break; |
1435 | case MVT::i1: |
1436 | case MVT::i8: |
1437 | case MVT::i16: |
1438 | needsExt = true; |
1439 | [[fallthrough]]; |
1440 | case MVT::i32: |
1441 | if (isThumb2) { |
1442 | if (!UseImm) |
1443 | CmpOpc = ARM::t2CMPrr; |
1444 | else |
1445 | CmpOpc = isNegativeImm ? ARM::t2CMNri : ARM::t2CMPri; |
1446 | } else { |
1447 | if (!UseImm) |
1448 | CmpOpc = ARM::CMPrr; |
1449 | else |
1450 | CmpOpc = isNegativeImm ? ARM::CMNri : ARM::CMPri; |
1451 | } |
1452 | break; |
1453 | } |
1454 | |
1455 | Register SrcReg1 = getRegForValue(V: Src1Value); |
1456 | if (!SrcReg1) |
1457 | return false; |
1458 | |
1459 | Register SrcReg2; |
1460 | if (!UseImm) { |
1461 | SrcReg2 = getRegForValue(V: Src2Value); |
1462 | if (!SrcReg2) |
1463 | return false; |
1464 | } |
1465 | |
1466 | // We have i1, i8, or i16, we need to either zero extend or sign extend. |
1467 | if (needsExt) { |
1468 | SrcReg1 = ARMEmitIntExt(SrcVT, SrcReg1, MVT::i32, isZExt); |
1469 | if (!SrcReg1) |
1470 | return false; |
1471 | if (!UseImm) { |
1472 | SrcReg2 = ARMEmitIntExt(SrcVT, SrcReg2, MVT::i32, isZExt); |
1473 | if (!SrcReg2) |
1474 | return false; |
1475 | } |
1476 | } |
1477 | |
1478 | const MCInstrDesc &II = TII.get(Opcode: CmpOpc); |
1479 | SrcReg1 = constrainOperandRegClass(II, Op: SrcReg1, OpNum: 0); |
1480 | if (!UseImm) { |
1481 | SrcReg2 = constrainOperandRegClass(II, Op: SrcReg2, OpNum: 1); |
1482 | AddOptionalDefs(MIB: BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, MCID: II) |
1483 | .addReg(RegNo: SrcReg1).addReg(RegNo: SrcReg2)); |
1484 | } else { |
1485 | MachineInstrBuilder MIB; |
1486 | MIB = BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, MCID: II) |
1487 | .addReg(RegNo: SrcReg1); |
1488 | |
1489 | // Only add immediate for icmp as the immediate for fcmp is an implicit 0.0. |
1490 | if (isICmp) |
1491 | MIB.addImm(Val: Imm); |
1492 | AddOptionalDefs(MIB); |
1493 | } |
1494 | |
1495 | // For floating point we need to move the result to a comparison register |
1496 | // that we can then use for branches. |
1497 | if (Ty->isFloatTy() || Ty->isDoubleTy()) |
1498 | AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, MIMD, |
1499 | TII.get(ARM::FMSTAT))); |
1500 | return true; |
1501 | } |
1502 | |
1503 | bool ARMFastISel::SelectCmp(const Instruction *I) { |
1504 | const CmpInst *CI = cast<CmpInst>(Val: I); |
1505 | |
1506 | // Get the compare predicate. |
1507 | ARMCC::CondCodes ARMPred = getComparePred(Pred: CI->getPredicate()); |
1508 | |
1509 | // We may not handle every CC for now. |
1510 | if (ARMPred == ARMCC::AL) return false; |
1511 | |
1512 | // Emit the compare. |
1513 | if (!ARMEmitCmp(Src1Value: CI->getOperand(i_nocapture: 0), Src2Value: CI->getOperand(i_nocapture: 1), isZExt: CI->isUnsigned())) |
1514 | return false; |
1515 | |
1516 | // Now set a register based on the comparison. Explicitly set the predicates |
1517 | // here. |
1518 | unsigned MovCCOpc = isThumb2 ? ARM::t2MOVCCi : ARM::MOVCCi; |
1519 | const TargetRegisterClass *RC = isThumb2 ? &ARM::rGPRRegClass |
1520 | : &ARM::GPRRegClass; |
1521 | Register DestReg = createResultReg(RC); |
1522 | Constant *Zero = ConstantInt::get(Ty: Type::getInt32Ty(C&: *Context), V: 0); |
1523 | Register ZeroReg = fastMaterializeConstant(C: Zero); |
1524 | // ARMEmitCmp emits a FMSTAT when necessary, so it's always safe to use CPSR. |
1525 | BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, MIMD, TII.get(MovCCOpc), DestReg) |
1526 | .addReg(ZeroReg).addImm(1) |
1527 | .addImm(ARMPred).addReg(ARM::CPSR); |
1528 | |
1529 | updateValueMap(I, Reg: DestReg); |
1530 | return true; |
1531 | } |
1532 | |
1533 | bool ARMFastISel::SelectFPExt(const Instruction *I) { |
1534 | // Make sure we have VFP and that we're extending float to double. |
1535 | if (!Subtarget->hasVFP2Base() || !Subtarget->hasFP64()) return false; |
1536 | |
1537 | Value *V = I->getOperand(i: 0); |
1538 | if (!I->getType()->isDoubleTy() || |
1539 | !V->getType()->isFloatTy()) return false; |
1540 | |
1541 | Register Op = getRegForValue(V); |
1542 | if (!Op) |
1543 | return false; |
1544 | |
1545 | Register Result = createResultReg(&ARM::DPRRegClass); |
1546 | AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, MIMD, |
1547 | TII.get(ARM::VCVTDS), Result) |
1548 | .addReg(Op)); |
1549 | updateValueMap(I, Reg: Result); |
1550 | return true; |
1551 | } |
1552 | |
1553 | bool ARMFastISel::SelectFPTrunc(const Instruction *I) { |
1554 | // Make sure we have VFP and that we're truncating double to float. |
1555 | if (!Subtarget->hasVFP2Base() || !Subtarget->hasFP64()) return false; |
1556 | |
1557 | Value *V = I->getOperand(i: 0); |
1558 | if (!(I->getType()->isFloatTy() && |
1559 | V->getType()->isDoubleTy())) return false; |
1560 | |
1561 | Register Op = getRegForValue(V); |
1562 | if (!Op) |
1563 | return false; |
1564 | |
1565 | Register Result = createResultReg(&ARM::SPRRegClass); |
1566 | AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, MIMD, |
1567 | TII.get(ARM::VCVTSD), Result) |
1568 | .addReg(Op)); |
1569 | updateValueMap(I, Reg: Result); |
1570 | return true; |
1571 | } |
1572 | |
1573 | bool ARMFastISel::SelectIToFP(const Instruction *I, bool isSigned) { |
1574 | // Make sure we have VFP. |
1575 | if (!Subtarget->hasVFP2Base()) return false; |
1576 | |
1577 | MVT DstVT; |
1578 | Type *Ty = I->getType(); |
1579 | if (!isTypeLegal(Ty, VT&: DstVT)) |
1580 | return false; |
1581 | |
1582 | Value *Src = I->getOperand(i: 0); |
1583 | EVT SrcEVT = TLI.getValueType(DL, Ty: Src->getType(), AllowUnknown: true); |
1584 | if (!SrcEVT.isSimple()) |
1585 | return false; |
1586 | MVT SrcVT = SrcEVT.getSimpleVT(); |
1587 | if (SrcVT != MVT::i32 && SrcVT != MVT::i16 && SrcVT != MVT::i8) |
1588 | return false; |
1589 | |
1590 | Register SrcReg = getRegForValue(V: Src); |
1591 | if (!SrcReg) |
1592 | return false; |
1593 | |
1594 | // Handle sign-extension. |
1595 | if (SrcVT == MVT::i16 || SrcVT == MVT::i8) { |
1596 | SrcReg = ARMEmitIntExt(SrcVT, SrcReg, MVT::i32, |
1597 | /*isZExt*/!isSigned); |
1598 | if (!SrcReg) |
1599 | return false; |
1600 | } |
1601 | |
1602 | // The conversion routine works on fp-reg to fp-reg and the operand above |
1603 | // was an integer, move it to the fp registers if possible. |
1604 | Register FP = ARMMoveToFPReg(MVT::f32, SrcReg); |
1605 | if (!FP) |
1606 | return false; |
1607 | |
1608 | unsigned Opc; |
1609 | if (Ty->isFloatTy()) Opc = isSigned ? ARM::VSITOS : ARM::VUITOS; |
1610 | else if (Ty->isDoubleTy() && Subtarget->hasFP64()) |
1611 | Opc = isSigned ? ARM::VSITOD : ARM::VUITOD; |
1612 | else return false; |
1613 | |
1614 | Register ResultReg = createResultReg(RC: TLI.getRegClassFor(VT: DstVT)); |
1615 | AddOptionalDefs(MIB: BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, |
1616 | MCID: TII.get(Opcode: Opc), DestReg: ResultReg).addReg(RegNo: FP)); |
1617 | updateValueMap(I, Reg: ResultReg); |
1618 | return true; |
1619 | } |
1620 | |
1621 | bool ARMFastISel::SelectFPToI(const Instruction *I, bool isSigned) { |
1622 | // Make sure we have VFP. |
1623 | if (!Subtarget->hasVFP2Base()) return false; |
1624 | |
1625 | MVT DstVT; |
1626 | Type *RetTy = I->getType(); |
1627 | if (!isTypeLegal(Ty: RetTy, VT&: DstVT)) |
1628 | return false; |
1629 | |
1630 | Register Op = getRegForValue(V: I->getOperand(i: 0)); |
1631 | if (!Op) |
1632 | return false; |
1633 | |
1634 | unsigned Opc; |
1635 | Type *OpTy = I->getOperand(i: 0)->getType(); |
1636 | if (OpTy->isFloatTy()) Opc = isSigned ? ARM::VTOSIZS : ARM::VTOUIZS; |
1637 | else if (OpTy->isDoubleTy() && Subtarget->hasFP64()) |
1638 | Opc = isSigned ? ARM::VTOSIZD : ARM::VTOUIZD; |
1639 | else return false; |
1640 | |
1641 | // f64->s32/u32 or f32->s32/u32 both need an intermediate f32 reg. |
1642 | Register ResultReg = createResultReg(TLI.getRegClassFor(MVT::f32)); |
1643 | AddOptionalDefs(MIB: BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, |
1644 | MCID: TII.get(Opcode: Opc), DestReg: ResultReg).addReg(RegNo: Op)); |
1645 | |
1646 | // This result needs to be in an integer register, but the conversion only |
1647 | // takes place in fp-regs. |
1648 | Register IntReg = ARMMoveToIntReg(VT: DstVT, SrcReg: ResultReg); |
1649 | if (!IntReg) |
1650 | return false; |
1651 | |
1652 | updateValueMap(I, Reg: IntReg); |
1653 | return true; |
1654 | } |
1655 | |
1656 | bool ARMFastISel::SelectSelect(const Instruction *I) { |
1657 | MVT VT; |
1658 | if (!isTypeLegal(Ty: I->getType(), VT)) |
1659 | return false; |
1660 | |
1661 | // Things need to be register sized for register moves. |
1662 | if (VT != MVT::i32) return false; |
1663 | |
1664 | Register CondReg = getRegForValue(V: I->getOperand(i: 0)); |
1665 | if (!CondReg) |
1666 | return false; |
1667 | Register Op1Reg = getRegForValue(V: I->getOperand(i: 1)); |
1668 | if (!Op1Reg) |
1669 | return false; |
1670 | |
1671 | // Check to see if we can use an immediate in the conditional move. |
1672 | int Imm = 0; |
1673 | bool UseImm = false; |
1674 | bool isNegativeImm = false; |
1675 | if (const ConstantInt *ConstInt = dyn_cast<ConstantInt>(Val: I->getOperand(i: 2))) { |
1676 | assert(VT == MVT::i32 && "Expecting an i32."); |
1677 | Imm = (int)ConstInt->getValue().getZExtValue(); |
1678 | if (Imm < 0) { |
1679 | isNegativeImm = true; |
1680 | Imm = ~Imm; |
1681 | } |
1682 | UseImm = isThumb2 ? (ARM_AM::getT2SOImmVal(Arg: Imm) != -1) : |
1683 | (ARM_AM::getSOImmVal(Arg: Imm) != -1); |
1684 | } |
1685 | |
1686 | Register Op2Reg; |
1687 | if (!UseImm) { |
1688 | Op2Reg = getRegForValue(V: I->getOperand(i: 2)); |
1689 | if (!Op2Reg) |
1690 | return false; |
1691 | } |
1692 | |
1693 | unsigned TstOpc = isThumb2 ? ARM::t2TSTri : ARM::TSTri; |
1694 | CondReg = constrainOperandRegClass(II: TII.get(Opcode: TstOpc), Op: CondReg, OpNum: 0); |
1695 | AddOptionalDefs( |
1696 | MIB: BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, MCID: TII.get(Opcode: TstOpc)) |
1697 | .addReg(RegNo: CondReg) |
1698 | .addImm(Val: 1)); |
1699 | |
1700 | unsigned MovCCOpc; |
1701 | const TargetRegisterClass *RC; |
1702 | if (!UseImm) { |
1703 | RC = isThumb2 ? &ARM::tGPRRegClass : &ARM::GPRRegClass; |
1704 | MovCCOpc = isThumb2 ? ARM::t2MOVCCr : ARM::MOVCCr; |
1705 | } else { |
1706 | RC = isThumb2 ? &ARM::rGPRRegClass : &ARM::GPRRegClass; |
1707 | if (!isNegativeImm) |
1708 | MovCCOpc = isThumb2 ? ARM::t2MOVCCi : ARM::MOVCCi; |
1709 | else |
1710 | MovCCOpc = isThumb2 ? ARM::t2MVNCCi : ARM::MVNCCi; |
1711 | } |
1712 | Register ResultReg = createResultReg(RC); |
1713 | if (!UseImm) { |
1714 | Op2Reg = constrainOperandRegClass(II: TII.get(Opcode: MovCCOpc), Op: Op2Reg, OpNum: 1); |
1715 | Op1Reg = constrainOperandRegClass(II: TII.get(Opcode: MovCCOpc), Op: Op1Reg, OpNum: 2); |
1716 | BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, MIMD, TII.get(MovCCOpc), |
1717 | ResultReg) |
1718 | .addReg(Op2Reg) |
1719 | .addReg(Op1Reg) |
1720 | .addImm(ARMCC::NE) |
1721 | .addReg(ARM::CPSR); |
1722 | } else { |
1723 | Op1Reg = constrainOperandRegClass(II: TII.get(Opcode: MovCCOpc), Op: Op1Reg, OpNum: 1); |
1724 | BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, MIMD, TII.get(MovCCOpc), |
1725 | ResultReg) |
1726 | .addReg(Op1Reg) |
1727 | .addImm(Imm) |
1728 | .addImm(ARMCC::EQ) |
1729 | .addReg(ARM::CPSR); |
1730 | } |
1731 | updateValueMap(I, Reg: ResultReg); |
1732 | return true; |
1733 | } |
1734 | |
1735 | bool ARMFastISel::SelectDiv(const Instruction *I, bool isSigned) { |
1736 | MVT VT; |
1737 | Type *Ty = I->getType(); |
1738 | if (!isTypeLegal(Ty, VT)) |
1739 | return false; |
1740 | |
1741 | // If we have integer div support we should have selected this automagically. |
1742 | // In case we have a real miss go ahead and return false and we'll pick |
1743 | // it up later. |
1744 | if (Subtarget->hasDivideInThumbMode()) |
1745 | return false; |
1746 | |
1747 | // Otherwise emit a libcall. |
1748 | RTLIB::Libcall LC = RTLIB::UNKNOWN_LIBCALL; |
1749 | if (VT == MVT::i8) |
1750 | LC = isSigned ? RTLIB::SDIV_I8 : RTLIB::UDIV_I8; |
1751 | else if (VT == MVT::i16) |
1752 | LC = isSigned ? RTLIB::SDIV_I16 : RTLIB::UDIV_I16; |
1753 | else if (VT == MVT::i32) |
1754 | LC = isSigned ? RTLIB::SDIV_I32 : RTLIB::UDIV_I32; |
1755 | else if (VT == MVT::i64) |
1756 | LC = isSigned ? RTLIB::SDIV_I64 : RTLIB::UDIV_I64; |
1757 | else if (VT == MVT::i128) |
1758 | LC = isSigned ? RTLIB::SDIV_I128 : RTLIB::UDIV_I128; |
1759 | assert(LC != RTLIB::UNKNOWN_LIBCALL && "Unsupported SDIV!"); |
1760 | |
1761 | return ARMEmitLibcall(I, Call: LC); |
1762 | } |
1763 | |
1764 | bool ARMFastISel::SelectRem(const Instruction *I, bool isSigned) { |
1765 | MVT VT; |
1766 | Type *Ty = I->getType(); |
1767 | if (!isTypeLegal(Ty, VT)) |
1768 | return false; |
1769 | |
1770 | // Many ABIs do not provide a libcall for standalone remainder, so we need to |
1771 | // use divrem (see the RTABI 4.3.1). Since FastISel can't handle non-double |
1772 | // multi-reg returns, we'll have to bail out. |
1773 | if (!TLI.hasStandaloneRem(VT)) { |
1774 | return false; |
1775 | } |
1776 | |
1777 | RTLIB::Libcall LC = RTLIB::UNKNOWN_LIBCALL; |
1778 | if (VT == MVT::i8) |
1779 | LC = isSigned ? RTLIB::SREM_I8 : RTLIB::UREM_I8; |
1780 | else if (VT == MVT::i16) |
1781 | LC = isSigned ? RTLIB::SREM_I16 : RTLIB::UREM_I16; |
1782 | else if (VT == MVT::i32) |
1783 | LC = isSigned ? RTLIB::SREM_I32 : RTLIB::UREM_I32; |
1784 | else if (VT == MVT::i64) |
1785 | LC = isSigned ? RTLIB::SREM_I64 : RTLIB::UREM_I64; |
1786 | else if (VT == MVT::i128) |
1787 | LC = isSigned ? RTLIB::SREM_I128 : RTLIB::UREM_I128; |
1788 | assert(LC != RTLIB::UNKNOWN_LIBCALL && "Unsupported SREM!"); |
1789 | |
1790 | return ARMEmitLibcall(I, Call: LC); |
1791 | } |
1792 | |
1793 | bool ARMFastISel::SelectBinaryIntOp(const Instruction *I, unsigned ISDOpcode) { |
1794 | EVT DestVT = TLI.getValueType(DL, Ty: I->getType(), AllowUnknown: true); |
1795 | |
1796 | // We can get here in the case when we have a binary operation on a non-legal |
1797 | // type and the target independent selector doesn't know how to handle it. |
1798 | if (DestVT != MVT::i16 && DestVT != MVT::i8 && DestVT != MVT::i1) |
1799 | return false; |
1800 | |
1801 | unsigned Opc; |
1802 | switch (ISDOpcode) { |
1803 | default: return false; |
1804 | case ISD::ADD: |
1805 | Opc = isThumb2 ? ARM::t2ADDrr : ARM::ADDrr; |
1806 | break; |
1807 | case ISD::OR: |
1808 | Opc = isThumb2 ? ARM::t2ORRrr : ARM::ORRrr; |
1809 | break; |
1810 | case ISD::SUB: |
1811 | Opc = isThumb2 ? ARM::t2SUBrr : ARM::SUBrr; |
1812 | break; |
1813 | } |
1814 | |
1815 | Register SrcReg1 = getRegForValue(V: I->getOperand(i: 0)); |
1816 | if (!SrcReg1) |
1817 | return false; |
1818 | |
1819 | // TODO: Often the 2nd operand is an immediate, which can be encoded directly |
1820 | // in the instruction, rather then materializing the value in a register. |
1821 | Register SrcReg2 = getRegForValue(V: I->getOperand(i: 1)); |
1822 | if (!SrcReg2) |
1823 | return false; |
1824 | |
1825 | Register ResultReg = createResultReg(&ARM::GPRnopcRegClass); |
1826 | SrcReg1 = constrainOperandRegClass(II: TII.get(Opcode: Opc), Op: SrcReg1, OpNum: 1); |
1827 | SrcReg2 = constrainOperandRegClass(II: TII.get(Opcode: Opc), Op: SrcReg2, OpNum: 2); |
1828 | AddOptionalDefs(MIB: BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, |
1829 | MCID: TII.get(Opcode: Opc), DestReg: ResultReg) |
1830 | .addReg(RegNo: SrcReg1).addReg(RegNo: SrcReg2)); |
1831 | updateValueMap(I, Reg: ResultReg); |
1832 | return true; |
1833 | } |
1834 | |
1835 | bool ARMFastISel::SelectBinaryFPOp(const Instruction *I, unsigned ISDOpcode) { |
1836 | EVT FPVT = TLI.getValueType(DL, Ty: I->getType(), AllowUnknown: true); |
1837 | if (!FPVT.isSimple()) return false; |
1838 | MVT VT = FPVT.getSimpleVT(); |
1839 | |
1840 | // FIXME: Support vector types where possible. |
1841 | if (VT.isVector()) |
1842 | return false; |
1843 | |
1844 | // We can get here in the case when we want to use NEON for our fp |
1845 | // operations, but can't figure out how to. Just use the vfp instructions |
1846 | // if we have them. |
1847 | // FIXME: It'd be nice to use NEON instructions. |
1848 | Type *Ty = I->getType(); |
1849 | if (Ty->isFloatTy() && !Subtarget->hasVFP2Base()) |
1850 | return false; |
1851 | if (Ty->isDoubleTy() && (!Subtarget->hasVFP2Base() || !Subtarget->hasFP64())) |
1852 | return false; |
1853 | |
1854 | unsigned Opc; |
1855 | bool is64bit = VT == MVT::f64 || VT == MVT::i64; |
1856 | switch (ISDOpcode) { |
1857 | default: return false; |
1858 | case ISD::FADD: |
1859 | Opc = is64bit ? ARM::VADDD : ARM::VADDS; |
1860 | break; |
1861 | case ISD::FSUB: |
1862 | Opc = is64bit ? ARM::VSUBD : ARM::VSUBS; |
1863 | break; |
1864 | case ISD::FMUL: |
1865 | Opc = is64bit ? ARM::VMULD : ARM::VMULS; |
1866 | break; |
1867 | } |
1868 | Register Op1 = getRegForValue(V: I->getOperand(i: 0)); |
1869 | if (!Op1) |
1870 | return false; |
1871 | |
1872 | Register Op2 = getRegForValue(V: I->getOperand(i: 1)); |
1873 | if (!Op2) |
1874 | return false; |
1875 | |
1876 | Register ResultReg = createResultReg(RC: TLI.getRegClassFor(VT: VT.SimpleTy)); |
1877 | AddOptionalDefs(MIB: BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, |
1878 | MCID: TII.get(Opcode: Opc), DestReg: ResultReg) |
1879 | .addReg(RegNo: Op1).addReg(RegNo: Op2)); |
1880 | updateValueMap(I, Reg: ResultReg); |
1881 | return true; |
1882 | } |
1883 | |
1884 | // Call Handling Code |
1885 | |
1886 | // This is largely taken directly from CCAssignFnForNode |
1887 | // TODO: We may not support all of this. |
1888 | CCAssignFn *ARMFastISel::CCAssignFnForCall(CallingConv::ID CC, |
1889 | bool Return, |
1890 | bool isVarArg) { |
1891 | switch (CC) { |
1892 | default: |
1893 | report_fatal_error(reason: "Unsupported calling convention"); |
1894 | case CallingConv::Fast: |
1895 | if (Subtarget->hasVFP2Base() && !isVarArg) { |
1896 | if (!Subtarget->isAAPCS_ABI()) |
1897 | return (Return ? RetFastCC_ARM_APCS : FastCC_ARM_APCS); |
1898 | // For AAPCS ABI targets, just use VFP variant of the calling convention. |
1899 | return (Return ? RetCC_ARM_AAPCS_VFP : CC_ARM_AAPCS_VFP); |
1900 | } |
1901 | [[fallthrough]]; |
1902 | case CallingConv::C: |
1903 | case CallingConv::CXX_FAST_TLS: |
1904 | // Use target triple & subtarget features to do actual dispatch. |
1905 | if (Subtarget->isAAPCS_ABI()) { |
1906 | if (Subtarget->hasFPRegs() && |
1907 | TM.Options.FloatABIType == FloatABI::Hard && !isVarArg) |
1908 | return (Return ? RetCC_ARM_AAPCS_VFP: CC_ARM_AAPCS_VFP); |
1909 | else |
1910 | return (Return ? RetCC_ARM_AAPCS: CC_ARM_AAPCS); |
1911 | } else { |
1912 | return (Return ? RetCC_ARM_APCS: CC_ARM_APCS); |
1913 | } |
1914 | case CallingConv::ARM_AAPCS_VFP: |
1915 | case CallingConv::Swift: |
1916 | case CallingConv::SwiftTail: |
1917 | if (!isVarArg) |
1918 | return (Return ? RetCC_ARM_AAPCS_VFP: CC_ARM_AAPCS_VFP); |
1919 | // Fall through to soft float variant, variadic functions don't |
1920 | // use hard floating point ABI. |
1921 | [[fallthrough]]; |
1922 | case CallingConv::ARM_AAPCS: |
1923 | return (Return ? RetCC_ARM_AAPCS: CC_ARM_AAPCS); |
1924 | case CallingConv::ARM_APCS: |
1925 | return (Return ? RetCC_ARM_APCS: CC_ARM_APCS); |
1926 | case CallingConv::GHC: |
1927 | if (Return) |
1928 | report_fatal_error(reason: "Can't return in GHC call convention"); |
1929 | else |
1930 | return CC_ARM_APCS_GHC; |
1931 | case CallingConv::CFGuard_Check: |
1932 | return (Return ? RetCC_ARM_AAPCS : CC_ARM_Win32_CFGuard_Check); |
1933 | } |
1934 | } |
1935 | |
1936 | bool ARMFastISel::ProcessCallArgs(SmallVectorImpl<Value*> &Args, |
1937 | SmallVectorImpl<Register> &ArgRegs, |
1938 | SmallVectorImpl<MVT> &ArgVTs, |
1939 | SmallVectorImpl<ISD::ArgFlagsTy> &ArgFlags, |
1940 | SmallVectorImpl<Register> &RegArgs, |
1941 | CallingConv::ID CC, |
1942 | unsigned &NumBytes, |
1943 | bool isVarArg) { |
1944 | SmallVector<CCValAssign, 16> ArgLocs; |
1945 | CCState CCInfo(CC, isVarArg, *FuncInfo.MF, ArgLocs, *Context); |
1946 | CCInfo.AnalyzeCallOperands(ArgVTs, Flags&: ArgFlags, |
1947 | Fn: CCAssignFnForCall(CC, Return: false, isVarArg)); |
1948 | |
1949 | // Check that we can handle all of the arguments. If we can't, then bail out |
1950 | // now before we add code to the MBB. |
1951 | for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) { |
1952 | CCValAssign &VA = ArgLocs[i]; |
1953 | MVT ArgVT = ArgVTs[VA.getValNo()]; |
1954 | |
1955 | // We don't handle NEON/vector parameters yet. |
1956 | if (ArgVT.isVector() || ArgVT.getSizeInBits() > 64) |
1957 | return false; |
1958 | |
1959 | // Now copy/store arg to correct locations. |
1960 | if (VA.isRegLoc() && !VA.needsCustom()) { |
1961 | continue; |
1962 | } else if (VA.needsCustom()) { |
1963 | // TODO: We need custom lowering for vector (v2f64) args. |
1964 | if (VA.getLocVT() != MVT::f64 || |
1965 | // TODO: Only handle register args for now. |
1966 | !VA.isRegLoc() || !ArgLocs[++i].isRegLoc()) |
1967 | return false; |
1968 | } else { |
1969 | switch (ArgVT.SimpleTy) { |
1970 | default: |
1971 | return false; |
1972 | case MVT::i1: |
1973 | case MVT::i8: |
1974 | case MVT::i16: |
1975 | case MVT::i32: |
1976 | break; |
1977 | case MVT::f32: |
1978 | if (!Subtarget->hasVFP2Base()) |
1979 | return false; |
1980 | break; |
1981 | case MVT::f64: |
1982 | if (!Subtarget->hasVFP2Base()) |
1983 | return false; |
1984 | break; |
1985 | } |
1986 | } |
1987 | } |
1988 | |
1989 | // At the point, we are able to handle the call's arguments in fast isel. |
1990 | |
1991 | // Get a count of how many bytes are to be pushed on the stack. |
1992 | NumBytes = CCInfo.getStackSize(); |
1993 | |
1994 | // Issue CALLSEQ_START |
1995 | unsigned AdjStackDown = TII.getCallFrameSetupOpcode(); |
1996 | AddOptionalDefs(MIB: BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, |
1997 | MCID: TII.get(Opcode: AdjStackDown)) |
1998 | .addImm(Val: NumBytes).addImm(Val: 0)); |
1999 | |
2000 | // Process the args. |
2001 | for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) { |
2002 | CCValAssign &VA = ArgLocs[i]; |
2003 | const Value *ArgVal = Args[VA.getValNo()]; |
2004 | Register Arg = ArgRegs[VA.getValNo()]; |
2005 | MVT ArgVT = ArgVTs[VA.getValNo()]; |
2006 | |
2007 | assert((!ArgVT.isVector() && ArgVT.getSizeInBits() <= 64) && |
2008 | "We don't handle NEON/vector parameters yet."); |
2009 | |
2010 | // Handle arg promotion, etc. |
2011 | switch (VA.getLocInfo()) { |
2012 | case CCValAssign::Full: break; |
2013 | case CCValAssign::SExt: { |
2014 | MVT DestVT = VA.getLocVT(); |
2015 | Arg = ARMEmitIntExt(SrcVT: ArgVT, SrcReg: Arg, DestVT, /*isZExt*/false); |
2016 | assert(Arg && "Failed to emit a sext"); |
2017 | ArgVT = DestVT; |
2018 | break; |
2019 | } |
2020 | case CCValAssign::AExt: |
2021 | // Intentional fall-through. Handle AExt and ZExt. |
2022 | case CCValAssign::ZExt: { |
2023 | MVT DestVT = VA.getLocVT(); |
2024 | Arg = ARMEmitIntExt(SrcVT: ArgVT, SrcReg: Arg, DestVT, /*isZExt*/true); |
2025 | assert(Arg && "Failed to emit a zext"); |
2026 | ArgVT = DestVT; |
2027 | break; |
2028 | } |
2029 | case CCValAssign::BCvt: { |
2030 | Register BC = fastEmit_r(VT: ArgVT, RetVT: VA.getLocVT(), Opcode: ISD::BITCAST, Op0: Arg); |
2031 | assert(BC && "Failed to emit a bitcast!"); |
2032 | Arg = BC; |
2033 | ArgVT = VA.getLocVT(); |
2034 | break; |
2035 | } |
2036 | default: llvm_unreachable("Unknown arg promotion!"); |
2037 | } |
2038 | |
2039 | // Now copy/store arg to correct locations. |
2040 | if (VA.isRegLoc() && !VA.needsCustom()) { |
2041 | BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, |
2042 | MCID: TII.get(Opcode: TargetOpcode::COPY), DestReg: VA.getLocReg()).addReg(RegNo: Arg); |
2043 | RegArgs.push_back(Elt: VA.getLocReg()); |
2044 | } else if (VA.needsCustom()) { |
2045 | // TODO: We need custom lowering for vector (v2f64) args. |
2046 | assert(VA.getLocVT() == MVT::f64 && |
2047 | "Custom lowering for v2f64 args not available"); |
2048 | |
2049 | // FIXME: ArgLocs[++i] may extend beyond ArgLocs.size() |
2050 | CCValAssign &NextVA = ArgLocs[++i]; |
2051 | |
2052 | assert(VA.isRegLoc() && NextVA.isRegLoc() && |
2053 | "We only handle register args!"); |
2054 | |
2055 | AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, MIMD, |
2056 | TII.get(ARM::VMOVRRD), VA.getLocReg()) |
2057 | .addReg(NextVA.getLocReg(), RegState::Define) |
2058 | .addReg(Arg)); |
2059 | RegArgs.push_back(Elt: VA.getLocReg()); |
2060 | RegArgs.push_back(Elt: NextVA.getLocReg()); |
2061 | } else { |
2062 | assert(VA.isMemLoc()); |
2063 | // Need to store on the stack. |
2064 | |
2065 | // Don't emit stores for undef values. |
2066 | if (isa<UndefValue>(Val: ArgVal)) |
2067 | continue; |
2068 | |
2069 | Address Addr; |
2070 | Addr.setKind(Address::RegBase); |
2071 | Addr.setReg(ARM::SP); |
2072 | Addr.setOffset(VA.getLocMemOffset()); |
2073 | |
2074 | bool EmitRet = ARMEmitStore(VT: ArgVT, SrcReg: Arg, Addr); (void)EmitRet; |
2075 | assert(EmitRet && "Could not emit a store for argument!"); |
2076 | } |
2077 | } |
2078 | |
2079 | return true; |
2080 | } |
2081 | |
2082 | bool ARMFastISel::FinishCall(MVT RetVT, SmallVectorImpl<Register> &UsedRegs, |
2083 | const Instruction *I, CallingConv::ID CC, |
2084 | unsigned &NumBytes, bool isVarArg) { |
2085 | // Issue CALLSEQ_END |
2086 | unsigned AdjStackUp = TII.getCallFrameDestroyOpcode(); |
2087 | AddOptionalDefs(MIB: BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, |
2088 | MCID: TII.get(Opcode: AdjStackUp)) |
2089 | .addImm(Val: NumBytes).addImm(Val: -1ULL)); |
2090 | |
2091 | // Now the return value. |
2092 | if (RetVT != MVT::isVoid) { |
2093 | SmallVector<CCValAssign, 16> RVLocs; |
2094 | CCState CCInfo(CC, isVarArg, *FuncInfo.MF, RVLocs, *Context); |
2095 | CCInfo.AnalyzeCallResult(VT: RetVT, Fn: CCAssignFnForCall(CC, Return: true, isVarArg)); |
2096 | |
2097 | // Copy all of the result registers out of their specified physreg. |
2098 | if (RVLocs.size() == 2 && RetVT == MVT::f64) { |
2099 | // For this move we copy into two registers and then move into the |
2100 | // double fp reg we want. |
2101 | MVT DestVT = RVLocs[0].getValVT(); |
2102 | const TargetRegisterClass* DstRC = TLI.getRegClassFor(VT: DestVT); |
2103 | Register ResultReg = createResultReg(RC: DstRC); |
2104 | AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, MIMD, |
2105 | TII.get(ARM::VMOVDRR), ResultReg) |
2106 | .addReg(RVLocs[0].getLocReg()) |
2107 | .addReg(RVLocs[1].getLocReg())); |
2108 | |
2109 | UsedRegs.push_back(Elt: RVLocs[0].getLocReg()); |
2110 | UsedRegs.push_back(Elt: RVLocs[1].getLocReg()); |
2111 | |
2112 | // Finally update the result. |
2113 | updateValueMap(I, Reg: ResultReg); |
2114 | } else { |
2115 | assert(RVLocs.size() == 1 &&"Can't handle non-double multi-reg retvals!"); |
2116 | MVT CopyVT = RVLocs[0].getValVT(); |
2117 | |
2118 | // Special handling for extended integers. |
2119 | if (RetVT == MVT::i1 || RetVT == MVT::i8 || RetVT == MVT::i16) |
2120 | CopyVT = MVT::i32; |
2121 | |
2122 | const TargetRegisterClass* DstRC = TLI.getRegClassFor(VT: CopyVT); |
2123 | |
2124 | Register ResultReg = createResultReg(RC: DstRC); |
2125 | BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, |
2126 | MCID: TII.get(Opcode: TargetOpcode::COPY), |
2127 | DestReg: ResultReg).addReg(RegNo: RVLocs[0].getLocReg()); |
2128 | UsedRegs.push_back(Elt: RVLocs[0].getLocReg()); |
2129 | |
2130 | // Finally update the result. |
2131 | updateValueMap(I, Reg: ResultReg); |
2132 | } |
2133 | } |
2134 | |
2135 | return true; |
2136 | } |
2137 | |
2138 | bool ARMFastISel::SelectRet(const Instruction *I) { |
2139 | const ReturnInst *Ret = cast<ReturnInst>(Val: I); |
2140 | const Function &F = *I->getParent()->getParent(); |
2141 | const bool IsCmseNSEntry = F.hasFnAttribute(Kind: "cmse_nonsecure_entry"); |
2142 | |
2143 | if (!FuncInfo.CanLowerReturn) |
2144 | return false; |
2145 | |
2146 | if (TLI.supportSwiftError() && |
2147 | F.getAttributes().hasAttrSomewhere(Attribute::SwiftError)) |
2148 | return false; |
2149 | |
2150 | if (TLI.supportSplitCSR(MF: FuncInfo.MF)) |
2151 | return false; |
2152 | |
2153 | // Build a list of return value registers. |
2154 | SmallVector<Register, 4> RetRegs; |
2155 | |
2156 | CallingConv::ID CC = F.getCallingConv(); |
2157 | if (Ret->getNumOperands() > 0) { |
2158 | SmallVector<ISD::OutputArg, 4> Outs; |
2159 | GetReturnInfo(CC, ReturnType: F.getReturnType(), attr: F.getAttributes(), Outs, TLI, DL); |
2160 | |
2161 | // Analyze operands of the call, assigning locations to each operand. |
2162 | SmallVector<CCValAssign, 16> ValLocs; |
2163 | CCState CCInfo(CC, F.isVarArg(), *FuncInfo.MF, ValLocs, I->getContext()); |
2164 | CCInfo.AnalyzeReturn(Outs, Fn: CCAssignFnForCall(CC, Return: true /* is Ret */, |
2165 | isVarArg: F.isVarArg())); |
2166 | |
2167 | const Value *RV = Ret->getOperand(i_nocapture: 0); |
2168 | Register Reg = getRegForValue(V: RV); |
2169 | if (!Reg) |
2170 | return false; |
2171 | |
2172 | // Only handle a single return value for now. |
2173 | if (ValLocs.size() != 1) |
2174 | return false; |
2175 | |
2176 | CCValAssign &VA = ValLocs[0]; |
2177 | |
2178 | // Don't bother handling odd stuff for now. |
2179 | if (VA.getLocInfo() != CCValAssign::Full) |
2180 | return false; |
2181 | // Only handle register returns for now. |
2182 | if (!VA.isRegLoc()) |
2183 | return false; |
2184 | |
2185 | Register SrcReg = Reg + VA.getValNo(); |
2186 | EVT RVEVT = TLI.getValueType(DL, Ty: RV->getType()); |
2187 | if (!RVEVT.isSimple()) return false; |
2188 | MVT RVVT = RVEVT.getSimpleVT(); |
2189 | MVT DestVT = VA.getValVT(); |
2190 | // Special handling for extended integers. |
2191 | if (RVVT != DestVT) { |
2192 | if (RVVT != MVT::i1 && RVVT != MVT::i8 && RVVT != MVT::i16) |
2193 | return false; |
2194 | |
2195 | assert(DestVT == MVT::i32 && "ARM should always ext to i32"); |
2196 | |
2197 | // Perform extension if flagged as either zext or sext. Otherwise, do |
2198 | // nothing. |
2199 | if (Outs[0].Flags.isZExt() || Outs[0].Flags.isSExt()) { |
2200 | SrcReg = ARMEmitIntExt(SrcVT: RVVT, SrcReg, DestVT, isZExt: Outs[0].Flags.isZExt()); |
2201 | if (!SrcReg) |
2202 | return false; |
2203 | } |
2204 | } |
2205 | |
2206 | // Make the copy. |
2207 | Register DstReg = VA.getLocReg(); |
2208 | const TargetRegisterClass* SrcRC = MRI.getRegClass(Reg: SrcReg); |
2209 | // Avoid a cross-class copy. This is very unlikely. |
2210 | if (!SrcRC->contains(Reg: DstReg)) |
2211 | return false; |
2212 | BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, |
2213 | MCID: TII.get(Opcode: TargetOpcode::COPY), DestReg: DstReg).addReg(RegNo: SrcReg); |
2214 | |
2215 | // Add register to return instruction. |
2216 | RetRegs.push_back(Elt: VA.getLocReg()); |
2217 | } |
2218 | |
2219 | unsigned RetOpc; |
2220 | if (IsCmseNSEntry) |
2221 | if (isThumb2) |
2222 | RetOpc = ARM::tBXNS_RET; |
2223 | else |
2224 | llvm_unreachable("CMSE not valid for non-Thumb targets"); |
2225 | else |
2226 | RetOpc = Subtarget->getReturnOpcode(); |
2227 | |
2228 | MachineInstrBuilder MIB = BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, |
2229 | MCID: TII.get(Opcode: RetOpc)); |
2230 | AddOptionalDefs(MIB); |
2231 | for (Register R : RetRegs) |
2232 | MIB.addReg(RegNo: R, flags: RegState::Implicit); |
2233 | return true; |
2234 | } |
2235 | |
2236 | unsigned ARMFastISel::ARMSelectCallOp(bool UseReg) { |
2237 | if (UseReg) |
2238 | return isThumb2 ? gettBLXrOpcode(MF: *MF) : getBLXOpcode(MF: *MF); |
2239 | else |
2240 | return isThumb2 ? ARM::tBL : ARM::BL; |
2241 | } |
2242 | |
2243 | Register ARMFastISel::getLibcallReg(const Twine &Name) { |
2244 | // Manually compute the global's type to avoid building it when unnecessary. |
2245 | Type *GVTy = PointerType::get(C&: *Context, /*AS=*/AddressSpace: 0); |
2246 | EVT LCREVT = TLI.getValueType(DL, Ty: GVTy); |
2247 | if (!LCREVT.isSimple()) |
2248 | return Register(); |
2249 | |
2250 | GlobalValue *GV = M.getNamedGlobal(Name: Name.str()); |
2251 | if (!GV) |
2252 | GV = new GlobalVariable(M, Type::getInt32Ty(C&: *Context), false, |
2253 | GlobalValue::ExternalLinkage, nullptr, Name); |
2254 | |
2255 | return ARMMaterializeGV(GV, VT: LCREVT.getSimpleVT()); |
2256 | } |
2257 | |
2258 | // A quick function that will emit a call for a named libcall in F with the |
2259 | // vector of passed arguments for the Instruction in I. We can assume that we |
2260 | // can emit a call for any libcall we can produce. This is an abridged version |
2261 | // of the full call infrastructure since we won't need to worry about things |
2262 | // like computed function pointers or strange arguments at call sites. |
2263 | // TODO: Try to unify this and the normal call bits for ARM, then try to unify |
2264 | // with X86. |
2265 | bool ARMFastISel::ARMEmitLibcall(const Instruction *I, RTLIB::Libcall Call) { |
2266 | CallingConv::ID CC = TLI.getLibcallCallingConv(Call); |
2267 | |
2268 | // Handle *simple* calls for now. |
2269 | Type *RetTy = I->getType(); |
2270 | MVT RetVT; |
2271 | if (RetTy->isVoidTy()) |
2272 | RetVT = MVT::isVoid; |
2273 | else if (!isTypeLegal(Ty: RetTy, VT&: RetVT)) |
2274 | return false; |
2275 | |
2276 | // Can't handle non-double multi-reg retvals. |
2277 | if (RetVT != MVT::isVoid && RetVT != MVT::i32) { |
2278 | SmallVector<CCValAssign, 16> RVLocs; |
2279 | CCState CCInfo(CC, false, *FuncInfo.MF, RVLocs, *Context); |
2280 | CCInfo.AnalyzeCallResult(VT: RetVT, Fn: CCAssignFnForCall(CC, Return: true, isVarArg: false)); |
2281 | if (RVLocs.size() >= 2 && RetVT != MVT::f64) |
2282 | return false; |
2283 | } |
2284 | |
2285 | // Set up the argument vectors. |
2286 | SmallVector<Value*, 8> Args; |
2287 | SmallVector<Register, 8> ArgRegs; |
2288 | SmallVector<MVT, 8> ArgVTs; |
2289 | SmallVector<ISD::ArgFlagsTy, 8> ArgFlags; |
2290 | Args.reserve(N: I->getNumOperands()); |
2291 | ArgRegs.reserve(N: I->getNumOperands()); |
2292 | ArgVTs.reserve(N: I->getNumOperands()); |
2293 | ArgFlags.reserve(N: I->getNumOperands()); |
2294 | for (Value *Op : I->operands()) { |
2295 | Register Arg = getRegForValue(V: Op); |
2296 | if (!Arg) |
2297 | return false; |
2298 | |
2299 | Type *ArgTy = Op->getType(); |
2300 | MVT ArgVT; |
2301 | if (!isTypeLegal(Ty: ArgTy, VT&: ArgVT)) return false; |
2302 | |
2303 | ISD::ArgFlagsTy Flags; |
2304 | Flags.setOrigAlign(DL.getABITypeAlign(Ty: ArgTy)); |
2305 | |
2306 | Args.push_back(Elt: Op); |
2307 | ArgRegs.push_back(Elt: Arg); |
2308 | ArgVTs.push_back(Elt: ArgVT); |
2309 | ArgFlags.push_back(Elt: Flags); |
2310 | } |
2311 | |
2312 | // Handle the arguments now that we've gotten them. |
2313 | SmallVector<Register, 4> RegArgs; |
2314 | unsigned NumBytes; |
2315 | if (!ProcessCallArgs(Args, ArgRegs, ArgVTs, ArgFlags, |
2316 | RegArgs, CC, NumBytes, isVarArg: false)) |
2317 | return false; |
2318 | |
2319 | Register CalleeReg; |
2320 | if (Subtarget->genLongCalls()) { |
2321 | CalleeReg = getLibcallReg(Name: TLI.getLibcallName(Call)); |
2322 | if (!CalleeReg) |
2323 | return false; |
2324 | } |
2325 | |
2326 | // Issue the call. |
2327 | unsigned CallOpc = ARMSelectCallOp(UseReg: Subtarget->genLongCalls()); |
2328 | MachineInstrBuilder MIB = BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, |
2329 | MIMD, MCID: TII.get(Opcode: CallOpc)); |
2330 | // BL / BLX don't take a predicate, but tBL / tBLX do. |
2331 | if (isThumb2) |
2332 | MIB.add(MOs: predOps(Pred: ARMCC::AL)); |
2333 | if (Subtarget->genLongCalls()) { |
2334 | CalleeReg = |
2335 | constrainOperandRegClass(II: TII.get(Opcode: CallOpc), Op: CalleeReg, OpNum: isThumb2 ? 2 : 0); |
2336 | MIB.addReg(RegNo: CalleeReg); |
2337 | } else |
2338 | MIB.addExternalSymbol(FnName: TLI.getLibcallName(Call)); |
2339 | |
2340 | // Add implicit physical register uses to the call. |
2341 | for (Register R : RegArgs) |
2342 | MIB.addReg(RegNo: R, flags: RegState::Implicit); |
2343 | |
2344 | // Add a register mask with the call-preserved registers. |
2345 | // Proper defs for return values will be added by setPhysRegsDeadExcept(). |
2346 | MIB.addRegMask(Mask: TRI.getCallPreservedMask(MF: *FuncInfo.MF, CC)); |
2347 | |
2348 | // Finish off the call including any return values. |
2349 | SmallVector<Register, 4> UsedRegs; |
2350 | if (!FinishCall(RetVT, UsedRegs, I, CC, NumBytes, isVarArg: false)) return false; |
2351 | |
2352 | // Set all unused physreg defs as dead. |
2353 | static_cast<MachineInstr *>(MIB)->setPhysRegsDeadExcept(UsedRegs, TRI); |
2354 | |
2355 | return true; |
2356 | } |
2357 | |
2358 | bool ARMFastISel::SelectCall(const Instruction *I, |
2359 | const char *IntrMemName = nullptr) { |
2360 | const CallInst *CI = cast<CallInst>(Val: I); |
2361 | const Value *Callee = CI->getCalledOperand(); |
2362 | |
2363 | // Can't handle inline asm. |
2364 | if (isa<InlineAsm>(Val: Callee)) return false; |
2365 | |
2366 | // Allow SelectionDAG isel to handle tail calls. |
2367 | if (CI->isTailCall()) return false; |
2368 | |
2369 | // Check the calling convention. |
2370 | CallingConv::ID CC = CI->getCallingConv(); |
2371 | |
2372 | // TODO: Avoid some calling conventions? |
2373 | |
2374 | FunctionType *FTy = CI->getFunctionType(); |
2375 | bool isVarArg = FTy->isVarArg(); |
2376 | |
2377 | // Handle *simple* calls for now. |
2378 | Type *RetTy = I->getType(); |
2379 | MVT RetVT; |
2380 | if (RetTy->isVoidTy()) |
2381 | RetVT = MVT::isVoid; |
2382 | else if (!isTypeLegal(RetTy, RetVT) && RetVT != MVT::i16 && |
2383 | RetVT != MVT::i8 && RetVT != MVT::i1) |
2384 | return false; |
2385 | |
2386 | // Can't handle non-double multi-reg retvals. |
2387 | if (RetVT != MVT::isVoid && RetVT != MVT::i1 && RetVT != MVT::i8 && |
2388 | RetVT != MVT::i16 && RetVT != MVT::i32) { |
2389 | SmallVector<CCValAssign, 16> RVLocs; |
2390 | CCState CCInfo(CC, isVarArg, *FuncInfo.MF, RVLocs, *Context); |
2391 | CCInfo.AnalyzeCallResult(VT: RetVT, Fn: CCAssignFnForCall(CC, Return: true, isVarArg)); |
2392 | if (RVLocs.size() >= 2 && RetVT != MVT::f64) |
2393 | return false; |
2394 | } |
2395 | |
2396 | // Set up the argument vectors. |
2397 | SmallVector<Value*, 8> Args; |
2398 | SmallVector<Register, 8> ArgRegs; |
2399 | SmallVector<MVT, 8> ArgVTs; |
2400 | SmallVector<ISD::ArgFlagsTy, 8> ArgFlags; |
2401 | unsigned arg_size = CI->arg_size(); |
2402 | Args.reserve(N: arg_size); |
2403 | ArgRegs.reserve(N: arg_size); |
2404 | ArgVTs.reserve(N: arg_size); |
2405 | ArgFlags.reserve(N: arg_size); |
2406 | for (auto ArgI = CI->arg_begin(), ArgE = CI->arg_end(); ArgI != ArgE; ++ArgI) { |
2407 | // If we're lowering a memory intrinsic instead of a regular call, skip the |
2408 | // last argument, which shouldn't be passed to the underlying function. |
2409 | if (IntrMemName && ArgE - ArgI <= 1) |
2410 | break; |
2411 | |
2412 | ISD::ArgFlagsTy Flags; |
2413 | unsigned ArgIdx = ArgI - CI->arg_begin(); |
2414 | if (CI->paramHasAttr(ArgIdx, Attribute::SExt)) |
2415 | Flags.setSExt(); |
2416 | if (CI->paramHasAttr(ArgIdx, Attribute::ZExt)) |
2417 | Flags.setZExt(); |
2418 | |
2419 | // FIXME: Only handle *easy* calls for now. |
2420 | if (CI->paramHasAttr(ArgIdx, Attribute::InReg) || |
2421 | CI->paramHasAttr(ArgIdx, Attribute::StructRet) || |
2422 | CI->paramHasAttr(ArgIdx, Attribute::SwiftSelf) || |
2423 | CI->paramHasAttr(ArgIdx, Attribute::SwiftError) || |
2424 | CI->paramHasAttr(ArgIdx, Attribute::Nest) || |
2425 | CI->paramHasAttr(ArgIdx, Attribute::ByVal)) |
2426 | return false; |
2427 | |
2428 | Type *ArgTy = (*ArgI)->getType(); |
2429 | MVT ArgVT; |
2430 | if (!isTypeLegal(ArgTy, ArgVT) && ArgVT != MVT::i16 && ArgVT != MVT::i8 && |
2431 | ArgVT != MVT::i1) |
2432 | return false; |
2433 | |
2434 | Register Arg = getRegForValue(V: *ArgI); |
2435 | if (!Arg.isValid()) |
2436 | return false; |
2437 | |
2438 | Flags.setOrigAlign(DL.getABITypeAlign(Ty: ArgTy)); |
2439 | |
2440 | Args.push_back(Elt: *ArgI); |
2441 | ArgRegs.push_back(Elt: Arg); |
2442 | ArgVTs.push_back(Elt: ArgVT); |
2443 | ArgFlags.push_back(Elt: Flags); |
2444 | } |
2445 | |
2446 | // Handle the arguments now that we've gotten them. |
2447 | SmallVector<Register, 4> RegArgs; |
2448 | unsigned NumBytes; |
2449 | if (!ProcessCallArgs(Args, ArgRegs, ArgVTs, ArgFlags, |
2450 | RegArgs, CC, NumBytes, isVarArg)) |
2451 | return false; |
2452 | |
2453 | bool UseReg = false; |
2454 | const GlobalValue *GV = dyn_cast<GlobalValue>(Val: Callee); |
2455 | if (!GV || Subtarget->genLongCalls()) UseReg = true; |
2456 | |
2457 | Register CalleeReg; |
2458 | if (UseReg) { |
2459 | if (IntrMemName) |
2460 | CalleeReg = getLibcallReg(Name: IntrMemName); |
2461 | else |
2462 | CalleeReg = getRegForValue(V: Callee); |
2463 | |
2464 | if (!CalleeReg) |
2465 | return false; |
2466 | } |
2467 | |
2468 | // Issue the call. |
2469 | unsigned CallOpc = ARMSelectCallOp(UseReg); |
2470 | MachineInstrBuilder MIB = BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, |
2471 | MIMD, MCID: TII.get(Opcode: CallOpc)); |
2472 | |
2473 | // ARM calls don't take a predicate, but tBL / tBLX do. |
2474 | if(isThumb2) |
2475 | MIB.add(MOs: predOps(Pred: ARMCC::AL)); |
2476 | if (UseReg) { |
2477 | CalleeReg = |
2478 | constrainOperandRegClass(II: TII.get(Opcode: CallOpc), Op: CalleeReg, OpNum: isThumb2 ? 2 : 0); |
2479 | MIB.addReg(RegNo: CalleeReg); |
2480 | } else if (!IntrMemName) |
2481 | MIB.addGlobalAddress(GV, Offset: 0, TargetFlags: 0); |
2482 | else |
2483 | MIB.addExternalSymbol(FnName: IntrMemName, TargetFlags: 0); |
2484 | |
2485 | // Add implicit physical register uses to the call. |
2486 | for (Register R : RegArgs) |
2487 | MIB.addReg(RegNo: R, flags: RegState::Implicit); |
2488 | |
2489 | // Add a register mask with the call-preserved registers. |
2490 | // Proper defs for return values will be added by setPhysRegsDeadExcept(). |
2491 | MIB.addRegMask(Mask: TRI.getCallPreservedMask(MF: *FuncInfo.MF, CC)); |
2492 | |
2493 | // Finish off the call including any return values. |
2494 | SmallVector<Register, 4> UsedRegs; |
2495 | if (!FinishCall(RetVT, UsedRegs, I, CC, NumBytes, isVarArg)) |
2496 | return false; |
2497 | |
2498 | // Set all unused physreg defs as dead. |
2499 | static_cast<MachineInstr *>(MIB)->setPhysRegsDeadExcept(UsedRegs, TRI); |
2500 | |
2501 | return true; |
2502 | } |
2503 | |
2504 | bool ARMFastISel::ARMIsMemCpySmall(uint64_t Len) { |
2505 | return Len <= 16; |
2506 | } |
2507 | |
2508 | bool ARMFastISel::ARMTryEmitSmallMemCpy(Address Dest, Address Src, uint64_t Len, |
2509 | MaybeAlign Alignment) { |
2510 | // Make sure we don't bloat code by inlining very large memcpy's. |
2511 | if (!ARMIsMemCpySmall(Len)) |
2512 | return false; |
2513 | |
2514 | while (Len) { |
2515 | MVT VT; |
2516 | if (!Alignment || *Alignment >= 4) { |
2517 | if (Len >= 4) |
2518 | VT = MVT::i32; |
2519 | else if (Len >= 2) |
2520 | VT = MVT::i16; |
2521 | else { |
2522 | assert(Len == 1 && "Expected a length of 1!"); |
2523 | VT = MVT::i8; |
2524 | } |
2525 | } else { |
2526 | assert(Alignment && "Alignment is set in this branch"); |
2527 | // Bound based on alignment. |
2528 | if (Len >= 2 && *Alignment == 2) |
2529 | VT = MVT::i16; |
2530 | else { |
2531 | VT = MVT::i8; |
2532 | } |
2533 | } |
2534 | |
2535 | bool RV; |
2536 | Register ResultReg; |
2537 | RV = ARMEmitLoad(VT, ResultReg, Addr&: Src); |
2538 | assert(RV && "Should be able to handle this load."); |
2539 | RV = ARMEmitStore(VT, SrcReg: ResultReg, Addr&: Dest); |
2540 | assert(RV && "Should be able to handle this store."); |
2541 | (void)RV; |
2542 | |
2543 | unsigned Size = VT.getSizeInBits()/8; |
2544 | Len -= Size; |
2545 | Dest.setOffset(Dest.getOffset() + Size); |
2546 | Src.setOffset(Src.getOffset() + Size); |
2547 | } |
2548 | |
2549 | return true; |
2550 | } |
2551 | |
2552 | bool ARMFastISel::SelectIntrinsicCall(const IntrinsicInst &I) { |
2553 | // FIXME: Handle more intrinsics. |
2554 | switch (I.getIntrinsicID()) { |
2555 | default: return false; |
2556 | case Intrinsic::frameaddress: { |
2557 | MachineFrameInfo &MFI = FuncInfo.MF->getFrameInfo(); |
2558 | MFI.setFrameAddressIsTaken(true); |
2559 | |
2560 | unsigned LdrOpc = isThumb2 ? ARM::t2LDRi12 : ARM::LDRi12; |
2561 | const TargetRegisterClass *RC = isThumb2 ? &ARM::tGPRRegClass |
2562 | : &ARM::GPRRegClass; |
2563 | |
2564 | const ARMBaseRegisterInfo *RegInfo = |
2565 | static_cast<const ARMBaseRegisterInfo *>(Subtarget->getRegisterInfo()); |
2566 | Register FramePtr = RegInfo->getFrameRegister(MF: *(FuncInfo.MF)); |
2567 | Register SrcReg = FramePtr; |
2568 | |
2569 | // Recursively load frame address |
2570 | // ldr r0 [fp] |
2571 | // ldr r0 [r0] |
2572 | // ldr r0 [r0] |
2573 | // ... |
2574 | Register DestReg; |
2575 | unsigned Depth = cast<ConstantInt>(Val: I.getOperand(i_nocapture: 0))->getZExtValue(); |
2576 | while (Depth--) { |
2577 | DestReg = createResultReg(RC); |
2578 | AddOptionalDefs(MIB: BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, |
2579 | MCID: TII.get(Opcode: LdrOpc), DestReg) |
2580 | .addReg(RegNo: SrcReg).addImm(Val: 0)); |
2581 | SrcReg = DestReg; |
2582 | } |
2583 | updateValueMap(I: &I, Reg: SrcReg); |
2584 | return true; |
2585 | } |
2586 | case Intrinsic::memcpy: |
2587 | case Intrinsic::memmove: { |
2588 | const MemTransferInst &MTI = cast<MemTransferInst>(Val: I); |
2589 | // Don't handle volatile. |
2590 | if (MTI.isVolatile()) |
2591 | return false; |
2592 | |
2593 | // Disable inlining for memmove before calls to ComputeAddress. Otherwise, |
2594 | // we would emit dead code because we don't currently handle memmoves. |
2595 | bool isMemCpy = (I.getIntrinsicID() == Intrinsic::memcpy); |
2596 | if (isa<ConstantInt>(Val: MTI.getLength()) && isMemCpy) { |
2597 | // Small memcpy's are common enough that we want to do them without a call |
2598 | // if possible. |
2599 | uint64_t Len = cast<ConstantInt>(Val: MTI.getLength())->getZExtValue(); |
2600 | if (ARMIsMemCpySmall(Len)) { |
2601 | Address Dest, Src; |
2602 | if (!ARMComputeAddress(Obj: MTI.getRawDest(), Addr&: Dest) || |
2603 | !ARMComputeAddress(Obj: MTI.getRawSource(), Addr&: Src)) |
2604 | return false; |
2605 | MaybeAlign Alignment; |
2606 | if (MTI.getDestAlign() || MTI.getSourceAlign()) |
2607 | Alignment = std::min(a: MTI.getDestAlign().valueOrOne(), |
2608 | b: MTI.getSourceAlign().valueOrOne()); |
2609 | if (ARMTryEmitSmallMemCpy(Dest, Src, Len, Alignment)) |
2610 | return true; |
2611 | } |
2612 | } |
2613 | |
2614 | if (!MTI.getLength()->getType()->isIntegerTy(Bitwidth: 32)) |
2615 | return false; |
2616 | |
2617 | if (MTI.getSourceAddressSpace() > 255 || MTI.getDestAddressSpace() > 255) |
2618 | return false; |
2619 | |
2620 | const char *IntrMemName = isa<MemCpyInst>(Val: I) ? "memcpy": "memmove"; |
2621 | return SelectCall(I: &I, IntrMemName); |
2622 | } |
2623 | case Intrinsic::memset: { |
2624 | const MemSetInst &MSI = cast<MemSetInst>(Val: I); |
2625 | // Don't handle volatile. |
2626 | if (MSI.isVolatile()) |
2627 | return false; |
2628 | |
2629 | if (!MSI.getLength()->getType()->isIntegerTy(Bitwidth: 32)) |
2630 | return false; |
2631 | |
2632 | if (MSI.getDestAddressSpace() > 255) |
2633 | return false; |
2634 | |
2635 | return SelectCall(I: &I, IntrMemName: "memset"); |
2636 | } |
2637 | case Intrinsic::trap: { |
2638 | unsigned Opcode; |
2639 | if (Subtarget->isThumb()) |
2640 | Opcode = ARM::tTRAP; |
2641 | else |
2642 | Opcode = Subtarget->useNaClTrap() ? ARM::TRAPNaCl : ARM::TRAP; |
2643 | BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, MCID: TII.get(Opcode)); |
2644 | return true; |
2645 | } |
2646 | } |
2647 | } |
2648 | |
2649 | bool ARMFastISel::SelectTrunc(const Instruction *I) { |
2650 | // The high bits for a type smaller than the register size are assumed to be |
2651 | // undefined. |
2652 | Value *Op = I->getOperand(i: 0); |
2653 | |
2654 | EVT SrcVT, DestVT; |
2655 | SrcVT = TLI.getValueType(DL, Ty: Op->getType(), AllowUnknown: true); |
2656 | DestVT = TLI.getValueType(DL, Ty: I->getType(), AllowUnknown: true); |
2657 | |
2658 | if (SrcVT != MVT::i32 && SrcVT != MVT::i16 && SrcVT != MVT::i8) |
2659 | return false; |
2660 | if (DestVT != MVT::i16 && DestVT != MVT::i8 && DestVT != MVT::i1) |
2661 | return false; |
2662 | |
2663 | Register SrcReg = getRegForValue(V: Op); |
2664 | if (!SrcReg) return false; |
2665 | |
2666 | // Because the high bits are undefined, a truncate doesn't generate |
2667 | // any code. |
2668 | updateValueMap(I, Reg: SrcReg); |
2669 | return true; |
2670 | } |
2671 | |
2672 | Register ARMFastISel::ARMEmitIntExt(MVT SrcVT, Register SrcReg, MVT DestVT, |
2673 | bool isZExt) { |
2674 | if (DestVT != MVT::i32 && DestVT != MVT::i16 && DestVT != MVT::i8) |
2675 | return Register(); |
2676 | if (SrcVT != MVT::i16 && SrcVT != MVT::i8 && SrcVT != MVT::i1) |
2677 | return Register(); |
2678 | |
2679 | // Table of which combinations can be emitted as a single instruction, |
2680 | // and which will require two. |
2681 | static const uint8_t isSingleInstrTbl[3][2][2][2] = { |
2682 | // ARM Thumb |
2683 | // !hasV6Ops hasV6Ops !hasV6Ops hasV6Ops |
2684 | // ext: s z s z s z s z |
2685 | /* 1 */ { { { 0, 1 }, { 0, 1 } }, { { 0, 0 }, { 0, 1 } } }, |
2686 | /* 8 */ { { { 0, 1 }, { 1, 1 } }, { { 0, 0 }, { 1, 1 } } }, |
2687 | /* 16 */ { { { 0, 0 }, { 1, 1 } }, { { 0, 0 }, { 1, 1 } } } |
2688 | }; |
2689 | |
2690 | // Target registers for: |
2691 | // - For ARM can never be PC. |
2692 | // - For 16-bit Thumb are restricted to lower 8 registers. |
2693 | // - For 32-bit Thumb are restricted to non-SP and non-PC. |
2694 | static const TargetRegisterClass *RCTbl[2][2] = { |
2695 | // Instructions: Two Single |
2696 | /* ARM */ { &ARM::GPRnopcRegClass, &ARM::GPRnopcRegClass }, |
2697 | /* Thumb */ { &ARM::tGPRRegClass, &ARM::rGPRRegClass } |
2698 | }; |
2699 | |
2700 | // Table governing the instruction(s) to be emitted. |
2701 | static const struct InstructionTable { |
2702 | uint32_t Opc : 16; |
2703 | uint32_t hasS : 1; // Some instructions have an S bit, always set it to 0. |
2704 | uint32_t Shift : 7; // For shift operand addressing mode, used by MOVsi. |
2705 | uint32_t Imm : 8; // All instructions have either a shift or a mask. |
2706 | } IT[2][2][3][2] = { |
2707 | { // Two instructions (first is left shift, second is in this table). |
2708 | { // ARM Opc S Shift Imm |
2709 | /* 1 bit sext */ { { ARM::MOVsi , 1, ARM_AM::asr , 31 }, |
2710 | /* 1 bit zext */ { ARM::MOVsi , 1, ARM_AM::lsr , 31 } }, |
2711 | /* 8 bit sext */ { { ARM::MOVsi , 1, ARM_AM::asr , 24 }, |
2712 | /* 8 bit zext */ { ARM::MOVsi , 1, ARM_AM::lsr , 24 } }, |
2713 | /* 16 bit sext */ { { ARM::MOVsi , 1, ARM_AM::asr , 16 }, |
2714 | /* 16 bit zext */ { ARM::MOVsi , 1, ARM_AM::lsr , 16 } } |
2715 | }, |
2716 | { // Thumb Opc S Shift Imm |
2717 | /* 1 bit sext */ { { ARM::tASRri , 0, ARM_AM::no_shift, 31 }, |
2718 | /* 1 bit zext */ { ARM::tLSRri , 0, ARM_AM::no_shift, 31 } }, |
2719 | /* 8 bit sext */ { { ARM::tASRri , 0, ARM_AM::no_shift, 24 }, |
2720 | /* 8 bit zext */ { ARM::tLSRri , 0, ARM_AM::no_shift, 24 } }, |
2721 | /* 16 bit sext */ { { ARM::tASRri , 0, ARM_AM::no_shift, 16 }, |
2722 | /* 16 bit zext */ { ARM::tLSRri , 0, ARM_AM::no_shift, 16 } } |
2723 | } |
2724 | }, |
2725 | { // Single instruction. |
2726 | { // ARM Opc S Shift Imm |
2727 | /* 1 bit sext */ { { ARM::KILL , 0, ARM_AM::no_shift, 0 }, |
2728 | /* 1 bit zext */ { ARM::ANDri , 1, ARM_AM::no_shift, 1 } }, |
2729 | /* 8 bit sext */ { { ARM::SXTB , 0, ARM_AM::no_shift, 0 }, |
2730 | /* 8 bit zext */ { ARM::ANDri , 1, ARM_AM::no_shift, 255 } }, |
2731 | /* 16 bit sext */ { { ARM::SXTH , 0, ARM_AM::no_shift, 0 }, |
2732 | /* 16 bit zext */ { ARM::UXTH , 0, ARM_AM::no_shift, 0 } } |
2733 | }, |
2734 | { // Thumb Opc S Shift Imm |
2735 | /* 1 bit sext */ { { ARM::KILL , 0, ARM_AM::no_shift, 0 }, |
2736 | /* 1 bit zext */ { ARM::t2ANDri, 1, ARM_AM::no_shift, 1 } }, |
2737 | /* 8 bit sext */ { { ARM::t2SXTB , 0, ARM_AM::no_shift, 0 }, |
2738 | /* 8 bit zext */ { ARM::t2ANDri, 1, ARM_AM::no_shift, 255 } }, |
2739 | /* 16 bit sext */ { { ARM::t2SXTH , 0, ARM_AM::no_shift, 0 }, |
2740 | /* 16 bit zext */ { ARM::t2UXTH , 0, ARM_AM::no_shift, 0 } } |
2741 | } |
2742 | } |
2743 | }; |
2744 | |
2745 | unsigned SrcBits = SrcVT.getSizeInBits(); |
2746 | unsigned DestBits = DestVT.getSizeInBits(); |
2747 | (void) DestBits; |
2748 | assert((SrcBits < DestBits) && "can only extend to larger types"); |
2749 | assert((DestBits == 32 || DestBits == 16 || DestBits == 8) && |
2750 | "other sizes unimplemented"); |
2751 | assert((SrcBits == 16 || SrcBits == 8 || SrcBits == 1) && |
2752 | "other sizes unimplemented"); |
2753 | |
2754 | bool hasV6Ops = Subtarget->hasV6Ops(); |
2755 | unsigned Bitness = SrcBits / 8; // {1,8,16}=>{0,1,2} |
2756 | assert((Bitness < 3) && "sanity-check table bounds"); |
2757 | |
2758 | bool isSingleInstr = isSingleInstrTbl[Bitness][isThumb2][hasV6Ops][isZExt]; |
2759 | const TargetRegisterClass *RC = RCTbl[isThumb2][isSingleInstr]; |
2760 | const InstructionTable *ITP = &IT[isSingleInstr][isThumb2][Bitness][isZExt]; |
2761 | unsigned Opc = ITP->Opc; |
2762 | assert(ARM::KILL != Opc && "Invalid table entry"); |
2763 | unsigned hasS = ITP->hasS; |
2764 | ARM_AM::ShiftOpc Shift = (ARM_AM::ShiftOpc) ITP->Shift; |
2765 | assert(((Shift == ARM_AM::no_shift) == (Opc != ARM::MOVsi)) && |
2766 | "only MOVsi has shift operand addressing mode"); |
2767 | unsigned Imm = ITP->Imm; |
2768 | |
2769 | // 16-bit Thumb instructions always set CPSR (unless they're in an IT block). |
2770 | bool setsCPSR = &ARM::tGPRRegClass == RC; |
2771 | unsigned LSLOpc = isThumb2 ? ARM::tLSLri : ARM::MOVsi; |
2772 | Register ResultReg; |
2773 | // MOVsi encodes shift and immediate in shift operand addressing mode. |
2774 | // The following condition has the same value when emitting two |
2775 | // instruction sequences: both are shifts. |
2776 | bool ImmIsSO = (Shift != ARM_AM::no_shift); |
2777 | |
2778 | // Either one or two instructions are emitted. |
2779 | // They're always of the form: |
2780 | // dst = in OP imm |
2781 | // CPSR is set only by 16-bit Thumb instructions. |
2782 | // Predicate, if any, is AL. |
2783 | // S bit, if available, is always 0. |
2784 | // When two are emitted the first's result will feed as the second's input, |
2785 | // that value is then dead. |
2786 | unsigned NumInstrsEmitted = isSingleInstr ? 1 : 2; |
2787 | for (unsigned Instr = 0; Instr != NumInstrsEmitted; ++Instr) { |
2788 | ResultReg = createResultReg(RC); |
2789 | bool isLsl = (0 == Instr) && !isSingleInstr; |
2790 | unsigned Opcode = isLsl ? LSLOpc : Opc; |
2791 | ARM_AM::ShiftOpc ShiftAM = isLsl ? ARM_AM::lsl : Shift; |
2792 | unsigned ImmEnc = ImmIsSO ? ARM_AM::getSORegOpc(ShOp: ShiftAM, Imm) : Imm; |
2793 | bool isKill = 1 == Instr; |
2794 | MachineInstrBuilder MIB = BuildMI( |
2795 | BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, MCID: TII.get(Opcode), DestReg: ResultReg); |
2796 | if (setsCPSR) |
2797 | MIB.addReg(ARM::CPSR, RegState::Define); |
2798 | SrcReg = constrainOperandRegClass(II: TII.get(Opcode), Op: SrcReg, OpNum: 1 + setsCPSR); |
2799 | MIB.addReg(RegNo: SrcReg, flags: isKill * RegState::Kill) |
2800 | .addImm(Val: ImmEnc) |
2801 | .add(MOs: predOps(Pred: ARMCC::AL)); |
2802 | if (hasS) |
2803 | MIB.add(MO: condCodeOp()); |
2804 | // Second instruction consumes the first's result. |
2805 | SrcReg = ResultReg; |
2806 | } |
2807 | |
2808 | return ResultReg; |
2809 | } |
2810 | |
2811 | bool ARMFastISel::SelectIntExt(const Instruction *I) { |
2812 | // On ARM, in general, integer casts don't involve legal types; this code |
2813 | // handles promotable integers. |
2814 | Type *DestTy = I->getType(); |
2815 | Value *Src = I->getOperand(i: 0); |
2816 | Type *SrcTy = Src->getType(); |
2817 | |
2818 | bool isZExt = isa<ZExtInst>(Val: I); |
2819 | Register SrcReg = getRegForValue(V: Src); |
2820 | if (!SrcReg) return false; |
2821 | |
2822 | EVT SrcEVT, DestEVT; |
2823 | SrcEVT = TLI.getValueType(DL, Ty: SrcTy, AllowUnknown: true); |
2824 | DestEVT = TLI.getValueType(DL, Ty: DestTy, AllowUnknown: true); |
2825 | if (!SrcEVT.isSimple()) return false; |
2826 | if (!DestEVT.isSimple()) return false; |
2827 | |
2828 | MVT SrcVT = SrcEVT.getSimpleVT(); |
2829 | MVT DestVT = DestEVT.getSimpleVT(); |
2830 | Register ResultReg = ARMEmitIntExt(SrcVT, SrcReg, DestVT, isZExt); |
2831 | if (!ResultReg) |
2832 | return false; |
2833 | updateValueMap(I, Reg: ResultReg); |
2834 | return true; |
2835 | } |
2836 | |
2837 | bool ARMFastISel::SelectShift(const Instruction *I, |
2838 | ARM_AM::ShiftOpc ShiftTy) { |
2839 | // We handle thumb2 mode by target independent selector |
2840 | // or SelectionDAG ISel. |
2841 | if (isThumb2) |
2842 | return false; |
2843 | |
2844 | // Only handle i32 now. |
2845 | EVT DestVT = TLI.getValueType(DL, Ty: I->getType(), AllowUnknown: true); |
2846 | if (DestVT != MVT::i32) |
2847 | return false; |
2848 | |
2849 | unsigned Opc = ARM::MOVsr; |
2850 | unsigned ShiftImm; |
2851 | Value *Src2Value = I->getOperand(i: 1); |
2852 | if (const ConstantInt *CI = dyn_cast<ConstantInt>(Val: Src2Value)) { |
2853 | ShiftImm = CI->getZExtValue(); |
2854 | |
2855 | // Fall back to selection DAG isel if the shift amount |
2856 | // is zero or greater than the width of the value type. |
2857 | if (ShiftImm == 0 || ShiftImm >=32) |
2858 | return false; |
2859 | |
2860 | Opc = ARM::MOVsi; |
2861 | } |
2862 | |
2863 | Value *Src1Value = I->getOperand(i: 0); |
2864 | Register Reg1 = getRegForValue(V: Src1Value); |
2865 | if (!Reg1) |
2866 | return false; |
2867 | |
2868 | Register Reg2; |
2869 | if (Opc == ARM::MOVsr) { |
2870 | Reg2 = getRegForValue(V: Src2Value); |
2871 | if (!Reg2) |
2872 | return false; |
2873 | } |
2874 | |
2875 | Register ResultReg = createResultReg(&ARM::GPRnopcRegClass); |
2876 | if (!ResultReg) |
2877 | return false; |
2878 | |
2879 | MachineInstrBuilder MIB = BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, |
2880 | MCID: TII.get(Opcode: Opc), DestReg: ResultReg) |
2881 | .addReg(RegNo: Reg1); |
2882 | |
2883 | if (Opc == ARM::MOVsi) |
2884 | MIB.addImm(Val: ARM_AM::getSORegOpc(ShOp: ShiftTy, Imm: ShiftImm)); |
2885 | else if (Opc == ARM::MOVsr) { |
2886 | MIB.addReg(RegNo: Reg2); |
2887 | MIB.addImm(Val: ARM_AM::getSORegOpc(ShOp: ShiftTy, Imm: 0)); |
2888 | } |
2889 | |
2890 | AddOptionalDefs(MIB); |
2891 | updateValueMap(I, Reg: ResultReg); |
2892 | return true; |
2893 | } |
2894 | |
2895 | // TODO: SoftFP support. |
2896 | bool ARMFastISel::fastSelectInstruction(const Instruction *I) { |
2897 | switch (I->getOpcode()) { |
2898 | case Instruction::Load: |
2899 | return SelectLoad(I); |
2900 | case Instruction::Store: |
2901 | return SelectStore(I); |
2902 | case Instruction::Br: |
2903 | return SelectBranch(I); |
2904 | case Instruction::IndirectBr: |
2905 | return SelectIndirectBr(I); |
2906 | case Instruction::ICmp: |
2907 | case Instruction::FCmp: |
2908 | return SelectCmp(I); |
2909 | case Instruction::FPExt: |
2910 | return SelectFPExt(I); |
2911 | case Instruction::FPTrunc: |
2912 | return SelectFPTrunc(I); |
2913 | case Instruction::SIToFP: |
2914 | return SelectIToFP(I, /*isSigned*/ true); |
2915 | case Instruction::UIToFP: |
2916 | return SelectIToFP(I, /*isSigned*/ false); |
2917 | case Instruction::FPToSI: |
2918 | return SelectFPToI(I, /*isSigned*/ true); |
2919 | case Instruction::FPToUI: |
2920 | return SelectFPToI(I, /*isSigned*/ false); |
2921 | case Instruction::Add: |
2922 | return SelectBinaryIntOp(I, ISDOpcode: ISD::ADD); |
2923 | case Instruction::Or: |
2924 | return SelectBinaryIntOp(I, ISDOpcode: ISD::OR); |
2925 | case Instruction::Sub: |
2926 | return SelectBinaryIntOp(I, ISDOpcode: ISD::SUB); |
2927 | case Instruction::FAdd: |
2928 | return SelectBinaryFPOp(I, ISDOpcode: ISD::FADD); |
2929 | case Instruction::FSub: |
2930 | return SelectBinaryFPOp(I, ISDOpcode: ISD::FSUB); |
2931 | case Instruction::FMul: |
2932 | return SelectBinaryFPOp(I, ISDOpcode: ISD::FMUL); |
2933 | case Instruction::SDiv: |
2934 | return SelectDiv(I, /*isSigned*/ true); |
2935 | case Instruction::UDiv: |
2936 | return SelectDiv(I, /*isSigned*/ false); |
2937 | case Instruction::SRem: |
2938 | return SelectRem(I, /*isSigned*/ true); |
2939 | case Instruction::URem: |
2940 | return SelectRem(I, /*isSigned*/ false); |
2941 | case Instruction::Call: |
2942 | if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(Val: I)) |
2943 | return SelectIntrinsicCall(I: *II); |
2944 | return SelectCall(I); |
2945 | case Instruction::Select: |
2946 | return SelectSelect(I); |
2947 | case Instruction::Ret: |
2948 | return SelectRet(I); |
2949 | case Instruction::Trunc: |
2950 | return SelectTrunc(I); |
2951 | case Instruction::ZExt: |
2952 | case Instruction::SExt: |
2953 | return SelectIntExt(I); |
2954 | case Instruction::Shl: |
2955 | return SelectShift(I, ShiftTy: ARM_AM::lsl); |
2956 | case Instruction::LShr: |
2957 | return SelectShift(I, ShiftTy: ARM_AM::lsr); |
2958 | case Instruction::AShr: |
2959 | return SelectShift(I, ShiftTy: ARM_AM::asr); |
2960 | default: break; |
2961 | } |
2962 | return false; |
2963 | } |
2964 | |
2965 | // This table describes sign- and zero-extend instructions which can be |
2966 | // folded into a preceding load. All of these extends have an immediate |
2967 | // (sometimes a mask and sometimes a shift) that's applied after |
2968 | // extension. |
2969 | static const struct FoldableLoadExtendsStruct { |
2970 | uint16_t Opc[2]; // ARM, Thumb. |
2971 | uint8_t ExpectedImm; |
2972 | uint8_t isZExt : 1; |
2973 | uint8_t ExpectedVT : 7; |
2974 | } FoldableLoadExtends[] = { |
2975 | { { ARM::SXTH, ARM::t2SXTH }, 0, 0, MVT::i16 }, |
2976 | { { ARM::UXTH, ARM::t2UXTH }, 0, 1, MVT::i16 }, |
2977 | { { ARM::ANDri, ARM::t2ANDri }, 255, 1, MVT::i8 }, |
2978 | { { ARM::SXTB, ARM::t2SXTB }, 0, 0, MVT::i8 }, |
2979 | { { ARM::UXTB, ARM::t2UXTB }, 0, 1, MVT::i8 } |
2980 | }; |
2981 | |
2982 | /// The specified machine instr operand is a vreg, and that |
2983 | /// vreg is being provided by the specified load instruction. If possible, |
2984 | /// try to fold the load as an operand to the instruction, returning true if |
2985 | /// successful. |
2986 | bool ARMFastISel::tryToFoldLoadIntoMI(MachineInstr *MI, unsigned OpNo, |
2987 | const LoadInst *LI) { |
2988 | // Verify we have a legal type before going any further. |
2989 | MVT VT; |
2990 | if (!isLoadTypeLegal(Ty: LI->getType(), VT)) |
2991 | return false; |
2992 | |
2993 | // Combine load followed by zero- or sign-extend. |
2994 | // ldrb r1, [r0] ldrb r1, [r0] |
2995 | // uxtb r2, r1 => |
2996 | // mov r3, r2 mov r3, r1 |
2997 | if (MI->getNumOperands() < 3 || !MI->getOperand(i: 2).isImm()) |
2998 | return false; |
2999 | const uint64_t Imm = MI->getOperand(i: 2).getImm(); |
3000 | |
3001 | bool Found = false; |
3002 | bool isZExt; |
3003 | for (const FoldableLoadExtendsStruct &FLE : FoldableLoadExtends) { |
3004 | if (FLE.Opc[isThumb2] == MI->getOpcode() && |
3005 | (uint64_t)FLE.ExpectedImm == Imm && |
3006 | MVT((MVT::SimpleValueType)FLE.ExpectedVT) == VT) { |
3007 | Found = true; |
3008 | isZExt = FLE.isZExt; |
3009 | } |
3010 | } |
3011 | if (!Found) return false; |
3012 | |
3013 | // See if we can handle this address. |
3014 | Address Addr; |
3015 | if (!ARMComputeAddress(Obj: LI->getOperand(i_nocapture: 0), Addr)) return false; |
3016 | |
3017 | Register ResultReg = MI->getOperand(i: 0).getReg(); |
3018 | if (!ARMEmitLoad(VT, ResultReg, Addr, Alignment: LI->getAlign(), isZExt, allocReg: false)) |
3019 | return false; |
3020 | MachineBasicBlock::iterator I(MI); |
3021 | removeDeadCode(I, E: std::next(x: I)); |
3022 | return true; |
3023 | } |
3024 | |
3025 | Register ARMFastISel::ARMLowerPICELF(const GlobalValue *GV, MVT VT) { |
3026 | bool UseGOT_PREL = !GV->isDSOLocal(); |
3027 | LLVMContext *Context = &MF->getFunction().getContext(); |
3028 | unsigned ARMPCLabelIndex = AFI->createPICLabelUId(); |
3029 | unsigned PCAdj = Subtarget->isThumb() ? 4 : 8; |
3030 | ARMConstantPoolValue *CPV = ARMConstantPoolConstant::Create( |
3031 | C: GV, ID: ARMPCLabelIndex, Kind: ARMCP::CPValue, PCAdj, |
3032 | Modifier: UseGOT_PREL ? ARMCP::GOT_PREL : ARMCP::no_modifier, |
3033 | /*AddCurrentAddress=*/UseGOT_PREL); |
3034 | |
3035 | Align ConstAlign = |
3036 | MF->getDataLayout().getPrefTypeAlign(Ty: PointerType::get(C&: *Context, AddressSpace: 0)); |
3037 | unsigned Idx = MF->getConstantPool()->getConstantPoolIndex(V: CPV, Alignment: ConstAlign); |
3038 | MachineMemOperand *CPMMO = |
3039 | MF->getMachineMemOperand(PtrInfo: MachinePointerInfo::getConstantPool(MF&: *MF), |
3040 | F: MachineMemOperand::MOLoad, Size: 4, BaseAlignment: Align(4)); |
3041 | |
3042 | Register TempReg = MF->getRegInfo().createVirtualRegister(&ARM::rGPRRegClass); |
3043 | unsigned Opc = isThumb2 ? ARM::t2LDRpci : ARM::LDRcp; |
3044 | MachineInstrBuilder MIB = |
3045 | BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, MCID: TII.get(Opcode: Opc), DestReg: TempReg) |
3046 | .addConstantPoolIndex(Idx) |
3047 | .addMemOperand(MMO: CPMMO); |
3048 | if (Opc == ARM::LDRcp) |
3049 | MIB.addImm(Val: 0); |
3050 | MIB.add(MOs: predOps(Pred: ARMCC::AL)); |
3051 | |
3052 | // Fix the address by adding pc. |
3053 | Register DestReg = createResultReg(RC: TLI.getRegClassFor(VT)); |
3054 | Opc = Subtarget->isThumb() ? ARM::tPICADD : UseGOT_PREL ? ARM::PICLDR |
3055 | : ARM::PICADD; |
3056 | DestReg = constrainOperandRegClass(II: TII.get(Opcode: Opc), Op: DestReg, OpNum: 0); |
3057 | MIB = BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, MCID: TII.get(Opcode: Opc), DestReg) |
3058 | .addReg(RegNo: TempReg) |
3059 | .addImm(Val: ARMPCLabelIndex); |
3060 | |
3061 | if (!Subtarget->isThumb()) |
3062 | MIB.add(MOs: predOps(Pred: ARMCC::AL)); |
3063 | |
3064 | if (UseGOT_PREL && Subtarget->isThumb()) { |
3065 | Register NewDestReg = createResultReg(RC: TLI.getRegClassFor(VT)); |
3066 | MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, MIMD, |
3067 | TII.get(ARM::t2LDRi12), NewDestReg) |
3068 | .addReg(DestReg) |
3069 | .addImm(0); |
3070 | DestReg = NewDestReg; |
3071 | AddOptionalDefs(MIB); |
3072 | } |
3073 | return DestReg; |
3074 | } |
3075 | |
3076 | bool ARMFastISel::fastLowerArguments() { |
3077 | if (!FuncInfo.CanLowerReturn) |
3078 | return false; |
3079 | |
3080 | const Function *F = FuncInfo.Fn; |
3081 | if (F->isVarArg()) |
3082 | return false; |
3083 | |
3084 | CallingConv::ID CC = F->getCallingConv(); |
3085 | switch (CC) { |
3086 | default: |
3087 | return false; |
3088 | case CallingConv::Fast: |
3089 | case CallingConv::C: |
3090 | case CallingConv::ARM_AAPCS_VFP: |
3091 | case CallingConv::ARM_AAPCS: |
3092 | case CallingConv::ARM_APCS: |
3093 | case CallingConv::Swift: |
3094 | case CallingConv::SwiftTail: |
3095 | break; |
3096 | } |
3097 | |
3098 | // Only handle simple cases. i.e. Up to 4 i8/i16/i32 scalar arguments |
3099 | // which are passed in r0 - r3. |
3100 | for (const Argument &Arg : F->args()) { |
3101 | if (Arg.getArgNo() >= 4) |
3102 | return false; |
3103 | |
3104 | if (Arg.hasAttribute(Attribute::InReg) || |
3105 | Arg.hasAttribute(Attribute::StructRet) || |
3106 | Arg.hasAttribute(Attribute::SwiftSelf) || |
3107 | Arg.hasAttribute(Attribute::SwiftError) || |
3108 | Arg.hasAttribute(Attribute::ByVal)) |
3109 | return false; |
3110 | |
3111 | Type *ArgTy = Arg.getType(); |
3112 | if (ArgTy->isStructTy() || ArgTy->isArrayTy() || ArgTy->isVectorTy()) |
3113 | return false; |
3114 | |
3115 | EVT ArgVT = TLI.getValueType(DL, Ty: ArgTy); |
3116 | if (!ArgVT.isSimple()) return false; |
3117 | switch (ArgVT.getSimpleVT().SimpleTy) { |
3118 | case MVT::i8: |
3119 | case MVT::i16: |
3120 | case MVT::i32: |
3121 | break; |
3122 | default: |
3123 | return false; |
3124 | } |
3125 | } |
3126 | |
3127 | static const MCPhysReg GPRArgRegs[] = { |
3128 | ARM::R0, ARM::R1, ARM::R2, ARM::R3 |
3129 | }; |
3130 | |
3131 | const TargetRegisterClass *RC = &ARM::rGPRRegClass; |
3132 | for (const Argument &Arg : F->args()) { |
3133 | unsigned ArgNo = Arg.getArgNo(); |
3134 | MCRegister SrcReg = GPRArgRegs[ArgNo]; |
3135 | Register DstReg = FuncInfo.MF->addLiveIn(PReg: SrcReg, RC); |
3136 | // FIXME: Unfortunately it's necessary to emit a copy from the livein copy. |
3137 | // Without this, EmitLiveInCopies may eliminate the livein if its only |
3138 | // use is a bitcast (which isn't turned into an instruction). |
3139 | Register ResultReg = createResultReg(RC); |
3140 | BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, |
3141 | MCID: TII.get(Opcode: TargetOpcode::COPY), |
3142 | DestReg: ResultReg).addReg(RegNo: DstReg, flags: getKillRegState(B: true)); |
3143 | updateValueMap(I: &Arg, Reg: ResultReg); |
3144 | } |
3145 | |
3146 | return true; |
3147 | } |
3148 | |
3149 | namespace llvm { |
3150 | |
3151 | FastISel *ARM::createFastISel(FunctionLoweringInfo &funcInfo, |
3152 | const TargetLibraryInfo *libInfo) { |
3153 | if (funcInfo.MF->getSubtarget<ARMSubtarget>().useFastISel()) |
3154 | return new ARMFastISel(funcInfo, libInfo); |
3155 | |
3156 | return nullptr; |
3157 | } |
3158 | |
3159 | } // end namespace llvm |
3160 |
Definitions
- Address
- Address
- setKind
- getKind
- isRegBase
- isFIBase
- setReg
- getReg
- setFI
- getFI
- setOffset
- getOffset
- ARMFastISel
- ARMFastISel
- getTargetLowering
- DefinesOptionalPredicate
- isARMNEONPred
- AddOptionalDefs
- fastEmitInst_r
- fastEmitInst_rr
- fastEmitInst_ri
- fastEmitInst_i
- ARMMoveToFPReg
- ARMMoveToIntReg
- ARMMaterializeFP
- ARMMaterializeInt
- isPositionIndependent
- ARMMaterializeGV
- fastMaterializeConstant
- fastMaterializeAlloca
- isTypeLegal
- isLoadTypeLegal
- ARMComputeAddress
- ARMSimplifyAddress
- AddLoadStoreOperands
- ARMEmitLoad
- SelectLoad
- ARMEmitStore
- SelectStore
- getComparePred
- SelectBranch
- SelectIndirectBr
- ARMEmitCmp
- SelectCmp
- SelectFPExt
- SelectFPTrunc
- SelectIToFP
- SelectFPToI
- SelectSelect
- SelectDiv
- SelectRem
- SelectBinaryIntOp
- SelectBinaryFPOp
- CCAssignFnForCall
- ProcessCallArgs
- FinishCall
- SelectRet
- ARMSelectCallOp
- getLibcallReg
- ARMEmitLibcall
- SelectCall
- ARMIsMemCpySmall
- ARMTryEmitSmallMemCpy
- SelectIntrinsicCall
- SelectTrunc
- ARMEmitIntExt
- SelectIntExt
- SelectShift
- fastSelectInstruction
- FoldableLoadExtendsStruct
- FoldableLoadExtends
- tryToFoldLoadIntoMI
- ARMLowerPICELF
- fastLowerArguments
Learn to use CMake with our Intro Training
Find out more