1 | //===- MipsISelLowering.cpp - Mips DAG Lowering 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 interfaces that Mips uses to lower LLVM code into a |
10 | // selection DAG. |
11 | // |
12 | //===----------------------------------------------------------------------===// |
13 | |
14 | #include "MipsISelLowering.h" |
15 | #include "MCTargetDesc/MipsBaseInfo.h" |
16 | #include "MCTargetDesc/MipsInstPrinter.h" |
17 | #include "MCTargetDesc/MipsMCTargetDesc.h" |
18 | #include "MipsCCState.h" |
19 | #include "MipsInstrInfo.h" |
20 | #include "MipsMachineFunction.h" |
21 | #include "MipsRegisterInfo.h" |
22 | #include "MipsSubtarget.h" |
23 | #include "MipsTargetMachine.h" |
24 | #include "MipsTargetObjectFile.h" |
25 | #include "llvm/ADT/APFloat.h" |
26 | #include "llvm/ADT/ArrayRef.h" |
27 | #include "llvm/ADT/SmallVector.h" |
28 | #include "llvm/ADT/Statistic.h" |
29 | #include "llvm/ADT/StringRef.h" |
30 | #include "llvm/ADT/StringSwitch.h" |
31 | #include "llvm/CodeGen/CallingConvLower.h" |
32 | #include "llvm/CodeGen/FunctionLoweringInfo.h" |
33 | #include "llvm/CodeGen/ISDOpcodes.h" |
34 | #include "llvm/CodeGen/MachineBasicBlock.h" |
35 | #include "llvm/CodeGen/MachineFrameInfo.h" |
36 | #include "llvm/CodeGen/MachineFunction.h" |
37 | #include "llvm/CodeGen/MachineInstr.h" |
38 | #include "llvm/CodeGen/MachineInstrBuilder.h" |
39 | #include "llvm/CodeGen/MachineJumpTableInfo.h" |
40 | #include "llvm/CodeGen/MachineMemOperand.h" |
41 | #include "llvm/CodeGen/MachineOperand.h" |
42 | #include "llvm/CodeGen/MachineRegisterInfo.h" |
43 | #include "llvm/CodeGen/SelectionDAG.h" |
44 | #include "llvm/CodeGen/SelectionDAGNodes.h" |
45 | #include "llvm/CodeGen/TargetFrameLowering.h" |
46 | #include "llvm/CodeGen/TargetInstrInfo.h" |
47 | #include "llvm/CodeGen/TargetRegisterInfo.h" |
48 | #include "llvm/CodeGen/ValueTypes.h" |
49 | #include "llvm/CodeGenTypes/MachineValueType.h" |
50 | #include "llvm/IR/CallingConv.h" |
51 | #include "llvm/IR/Constants.h" |
52 | #include "llvm/IR/DataLayout.h" |
53 | #include "llvm/IR/DebugLoc.h" |
54 | #include "llvm/IR/DerivedTypes.h" |
55 | #include "llvm/IR/Function.h" |
56 | #include "llvm/IR/GlobalValue.h" |
57 | #include "llvm/IR/Module.h" |
58 | #include "llvm/IR/Type.h" |
59 | #include "llvm/IR/Value.h" |
60 | #include "llvm/MC/MCContext.h" |
61 | #include "llvm/Support/Casting.h" |
62 | #include "llvm/Support/CodeGen.h" |
63 | #include "llvm/Support/CommandLine.h" |
64 | #include "llvm/Support/Compiler.h" |
65 | #include "llvm/Support/ErrorHandling.h" |
66 | #include "llvm/Support/MathExtras.h" |
67 | #include "llvm/Target/TargetMachine.h" |
68 | #include "llvm/Target/TargetOptions.h" |
69 | #include <algorithm> |
70 | #include <cassert> |
71 | #include <cctype> |
72 | #include <cstdint> |
73 | #include <deque> |
74 | #include <iterator> |
75 | #include <utility> |
76 | #include <vector> |
77 | |
78 | using namespace llvm; |
79 | |
80 | #define DEBUG_TYPE "mips-lower" |
81 | |
82 | STATISTIC(NumTailCalls, "Number of tail calls" ); |
83 | |
84 | static cl::opt<bool> |
85 | NoZeroDivCheck("mno-check-zero-division" , cl::Hidden, |
86 | cl::desc("MIPS: Don't trap on integer division by zero." ), |
87 | cl::init(Val: false)); |
88 | |
89 | extern cl::opt<bool> EmitJalrReloc; |
90 | |
91 | static const MCPhysReg Mips64DPRegs[8] = { |
92 | Mips::D12_64, Mips::D13_64, Mips::D14_64, Mips::D15_64, |
93 | Mips::D16_64, Mips::D17_64, Mips::D18_64, Mips::D19_64 |
94 | }; |
95 | |
96 | // The MIPS MSA ABI passes vector arguments in the integer register set. |
97 | // The number of integer registers used is dependant on the ABI used. |
98 | MVT MipsTargetLowering::getRegisterTypeForCallingConv(LLVMContext &Context, |
99 | CallingConv::ID CC, |
100 | EVT VT) const { |
101 | if (!VT.isVector()) |
102 | return getRegisterType(Context, VT); |
103 | |
104 | if (VT.isPow2VectorType() && VT.getVectorElementType().isRound()) |
105 | return Subtarget.isABI_O32() || VT.getSizeInBits() == 32 ? MVT::i32 |
106 | : MVT::i64; |
107 | return getRegisterType(Context, VT: VT.getVectorElementType()); |
108 | } |
109 | |
110 | unsigned MipsTargetLowering::getNumRegistersForCallingConv(LLVMContext &Context, |
111 | CallingConv::ID CC, |
112 | EVT VT) const { |
113 | if (VT.isVector()) { |
114 | if (VT.isPow2VectorType() && VT.getVectorElementType().isRound()) |
115 | return divideCeil(Numerator: VT.getSizeInBits(), Denominator: Subtarget.isABI_O32() ? 32 : 64); |
116 | return VT.getVectorNumElements() * |
117 | getNumRegisters(Context, VT: VT.getVectorElementType()); |
118 | } |
119 | return MipsTargetLowering::getNumRegisters(Context, VT); |
120 | } |
121 | |
122 | unsigned MipsTargetLowering::getVectorTypeBreakdownForCallingConv( |
123 | LLVMContext &Context, CallingConv::ID CC, EVT VT, EVT &IntermediateVT, |
124 | unsigned &NumIntermediates, MVT &RegisterVT) const { |
125 | if (VT.isPow2VectorType() && VT.getVectorElementType().isRound()) { |
126 | IntermediateVT = getRegisterTypeForCallingConv(Context, CC, VT); |
127 | RegisterVT = IntermediateVT.getSimpleVT(); |
128 | NumIntermediates = getNumRegistersForCallingConv(Context, CC, VT); |
129 | return NumIntermediates; |
130 | } |
131 | IntermediateVT = VT.getVectorElementType(); |
132 | NumIntermediates = VT.getVectorNumElements(); |
133 | RegisterVT = getRegisterType(Context, VT: IntermediateVT); |
134 | return NumIntermediates * getNumRegisters(Context, VT: IntermediateVT); |
135 | } |
136 | |
137 | SDValue MipsTargetLowering::getGlobalReg(SelectionDAG &DAG, EVT Ty) const { |
138 | MachineFunction &MF = DAG.getMachineFunction(); |
139 | MipsFunctionInfo *FI = MF.getInfo<MipsFunctionInfo>(); |
140 | return DAG.getRegister(Reg: FI->getGlobalBaseReg(MF), VT: Ty); |
141 | } |
142 | |
143 | SDValue MipsTargetLowering::getTargetNode(GlobalAddressSDNode *N, EVT Ty, |
144 | SelectionDAG &DAG, |
145 | unsigned Flag) const { |
146 | return DAG.getTargetGlobalAddress(GV: N->getGlobal(), DL: SDLoc(N), VT: Ty, offset: 0, TargetFlags: Flag); |
147 | } |
148 | |
149 | SDValue MipsTargetLowering::getTargetNode(ExternalSymbolSDNode *N, EVT Ty, |
150 | SelectionDAG &DAG, |
151 | unsigned Flag) const { |
152 | return DAG.getTargetExternalSymbol(Sym: N->getSymbol(), VT: Ty, TargetFlags: Flag); |
153 | } |
154 | |
155 | SDValue MipsTargetLowering::getTargetNode(BlockAddressSDNode *N, EVT Ty, |
156 | SelectionDAG &DAG, |
157 | unsigned Flag) const { |
158 | return DAG.getTargetBlockAddress(BA: N->getBlockAddress(), VT: Ty, Offset: 0, TargetFlags: Flag); |
159 | } |
160 | |
161 | SDValue MipsTargetLowering::getTargetNode(JumpTableSDNode *N, EVT Ty, |
162 | SelectionDAG &DAG, |
163 | unsigned Flag) const { |
164 | return DAG.getTargetJumpTable(JTI: N->getIndex(), VT: Ty, TargetFlags: Flag); |
165 | } |
166 | |
167 | SDValue MipsTargetLowering::getTargetNode(ConstantPoolSDNode *N, EVT Ty, |
168 | SelectionDAG &DAG, |
169 | unsigned Flag) const { |
170 | return DAG.getTargetConstantPool(C: N->getConstVal(), VT: Ty, Align: N->getAlign(), |
171 | Offset: N->getOffset(), TargetFlags: Flag); |
172 | } |
173 | |
174 | const char *MipsTargetLowering::getTargetNodeName(unsigned Opcode) const { |
175 | switch ((MipsISD::NodeType)Opcode) { |
176 | case MipsISD::FIRST_NUMBER: break; |
177 | case MipsISD::JmpLink: return "MipsISD::JmpLink" ; |
178 | case MipsISD::TailCall: return "MipsISD::TailCall" ; |
179 | case MipsISD::Highest: return "MipsISD::Highest" ; |
180 | case MipsISD::Higher: return "MipsISD::Higher" ; |
181 | case MipsISD::Hi: return "MipsISD::Hi" ; |
182 | case MipsISD::Lo: return "MipsISD::Lo" ; |
183 | case MipsISD::GotHi: return "MipsISD::GotHi" ; |
184 | case MipsISD::TlsHi: return "MipsISD::TlsHi" ; |
185 | case MipsISD::GPRel: return "MipsISD::GPRel" ; |
186 | case MipsISD::ThreadPointer: return "MipsISD::ThreadPointer" ; |
187 | case MipsISD::Ret: return "MipsISD::Ret" ; |
188 | case MipsISD::ERet: return "MipsISD::ERet" ; |
189 | case MipsISD::EH_RETURN: return "MipsISD::EH_RETURN" ; |
190 | case MipsISD::FAbs: return "MipsISD::FAbs" ; |
191 | case MipsISD::FMS: return "MipsISD::FMS" ; |
192 | case MipsISD::FPBrcond: return "MipsISD::FPBrcond" ; |
193 | case MipsISD::FPCmp: return "MipsISD::FPCmp" ; |
194 | case MipsISD::FSELECT: return "MipsISD::FSELECT" ; |
195 | case MipsISD::MTC1_D64: return "MipsISD::MTC1_D64" ; |
196 | case MipsISD::CMovFP_T: return "MipsISD::CMovFP_T" ; |
197 | case MipsISD::CMovFP_F: return "MipsISD::CMovFP_F" ; |
198 | case MipsISD::TruncIntFP: return "MipsISD::TruncIntFP" ; |
199 | case MipsISD::MFHI: return "MipsISD::MFHI" ; |
200 | case MipsISD::MFLO: return "MipsISD::MFLO" ; |
201 | case MipsISD::MTLOHI: return "MipsISD::MTLOHI" ; |
202 | case MipsISD::Mult: return "MipsISD::Mult" ; |
203 | case MipsISD::Multu: return "MipsISD::Multu" ; |
204 | case MipsISD::MAdd: return "MipsISD::MAdd" ; |
205 | case MipsISD::MAddu: return "MipsISD::MAddu" ; |
206 | case MipsISD::MSub: return "MipsISD::MSub" ; |
207 | case MipsISD::MSubu: return "MipsISD::MSubu" ; |
208 | case MipsISD::DivRem: return "MipsISD::DivRem" ; |
209 | case MipsISD::DivRemU: return "MipsISD::DivRemU" ; |
210 | case MipsISD::DivRem16: return "MipsISD::DivRem16" ; |
211 | case MipsISD::DivRemU16: return "MipsISD::DivRemU16" ; |
212 | case MipsISD::BuildPairF64: return "MipsISD::BuildPairF64" ; |
213 | case MipsISD::ExtractElementF64: return "MipsISD::ExtractElementF64" ; |
214 | case MipsISD::Wrapper: return "MipsISD::Wrapper" ; |
215 | case MipsISD::DynAlloc: return "MipsISD::DynAlloc" ; |
216 | case MipsISD::Sync: return "MipsISD::Sync" ; |
217 | case MipsISD::Ext: return "MipsISD::Ext" ; |
218 | case MipsISD::Ins: return "MipsISD::Ins" ; |
219 | case MipsISD::CIns: return "MipsISD::CIns" ; |
220 | case MipsISD::LWL: return "MipsISD::LWL" ; |
221 | case MipsISD::LWR: return "MipsISD::LWR" ; |
222 | case MipsISD::SWL: return "MipsISD::SWL" ; |
223 | case MipsISD::SWR: return "MipsISD::SWR" ; |
224 | case MipsISD::LDL: return "MipsISD::LDL" ; |
225 | case MipsISD::LDR: return "MipsISD::LDR" ; |
226 | case MipsISD::SDL: return "MipsISD::SDL" ; |
227 | case MipsISD::SDR: return "MipsISD::SDR" ; |
228 | case MipsISD::EXTP: return "MipsISD::EXTP" ; |
229 | case MipsISD::EXTPDP: return "MipsISD::EXTPDP" ; |
230 | case MipsISD::EXTR_S_H: return "MipsISD::EXTR_S_H" ; |
231 | case MipsISD::EXTR_W: return "MipsISD::EXTR_W" ; |
232 | case MipsISD::EXTR_R_W: return "MipsISD::EXTR_R_W" ; |
233 | case MipsISD::EXTR_RS_W: return "MipsISD::EXTR_RS_W" ; |
234 | case MipsISD::SHILO: return "MipsISD::SHILO" ; |
235 | case MipsISD::MTHLIP: return "MipsISD::MTHLIP" ; |
236 | case MipsISD::MULSAQ_S_W_PH: return "MipsISD::MULSAQ_S_W_PH" ; |
237 | case MipsISD::MAQ_S_W_PHL: return "MipsISD::MAQ_S_W_PHL" ; |
238 | case MipsISD::MAQ_S_W_PHR: return "MipsISD::MAQ_S_W_PHR" ; |
239 | case MipsISD::MAQ_SA_W_PHL: return "MipsISD::MAQ_SA_W_PHL" ; |
240 | case MipsISD::MAQ_SA_W_PHR: return "MipsISD::MAQ_SA_W_PHR" ; |
241 | case MipsISD::DOUBLE_SELECT_I: return "MipsISD::DOUBLE_SELECT_I" ; |
242 | case MipsISD::DOUBLE_SELECT_I64: return "MipsISD::DOUBLE_SELECT_I64" ; |
243 | case MipsISD::DPAU_H_QBL: return "MipsISD::DPAU_H_QBL" ; |
244 | case MipsISD::DPAU_H_QBR: return "MipsISD::DPAU_H_QBR" ; |
245 | case MipsISD::DPSU_H_QBL: return "MipsISD::DPSU_H_QBL" ; |
246 | case MipsISD::DPSU_H_QBR: return "MipsISD::DPSU_H_QBR" ; |
247 | case MipsISD::DPAQ_S_W_PH: return "MipsISD::DPAQ_S_W_PH" ; |
248 | case MipsISD::DPSQ_S_W_PH: return "MipsISD::DPSQ_S_W_PH" ; |
249 | case MipsISD::DPAQ_SA_L_W: return "MipsISD::DPAQ_SA_L_W" ; |
250 | case MipsISD::DPSQ_SA_L_W: return "MipsISD::DPSQ_SA_L_W" ; |
251 | case MipsISD::DPA_W_PH: return "MipsISD::DPA_W_PH" ; |
252 | case MipsISD::DPS_W_PH: return "MipsISD::DPS_W_PH" ; |
253 | case MipsISD::DPAQX_S_W_PH: return "MipsISD::DPAQX_S_W_PH" ; |
254 | case MipsISD::DPAQX_SA_W_PH: return "MipsISD::DPAQX_SA_W_PH" ; |
255 | case MipsISD::DPAX_W_PH: return "MipsISD::DPAX_W_PH" ; |
256 | case MipsISD::DPSX_W_PH: return "MipsISD::DPSX_W_PH" ; |
257 | case MipsISD::DPSQX_S_W_PH: return "MipsISD::DPSQX_S_W_PH" ; |
258 | case MipsISD::DPSQX_SA_W_PH: return "MipsISD::DPSQX_SA_W_PH" ; |
259 | case MipsISD::MULSA_W_PH: return "MipsISD::MULSA_W_PH" ; |
260 | case MipsISD::MULT: return "MipsISD::MULT" ; |
261 | case MipsISD::MULTU: return "MipsISD::MULTU" ; |
262 | case MipsISD::MADD_DSP: return "MipsISD::MADD_DSP" ; |
263 | case MipsISD::MADDU_DSP: return "MipsISD::MADDU_DSP" ; |
264 | case MipsISD::MSUB_DSP: return "MipsISD::MSUB_DSP" ; |
265 | case MipsISD::MSUBU_DSP: return "MipsISD::MSUBU_DSP" ; |
266 | case MipsISD::SHLL_DSP: return "MipsISD::SHLL_DSP" ; |
267 | case MipsISD::SHRA_DSP: return "MipsISD::SHRA_DSP" ; |
268 | case MipsISD::SHRL_DSP: return "MipsISD::SHRL_DSP" ; |
269 | case MipsISD::SETCC_DSP: return "MipsISD::SETCC_DSP" ; |
270 | case MipsISD::SELECT_CC_DSP: return "MipsISD::SELECT_CC_DSP" ; |
271 | case MipsISD::VALL_ZERO: return "MipsISD::VALL_ZERO" ; |
272 | case MipsISD::VANY_ZERO: return "MipsISD::VANY_ZERO" ; |
273 | case MipsISD::VALL_NONZERO: return "MipsISD::VALL_NONZERO" ; |
274 | case MipsISD::VANY_NONZERO: return "MipsISD::VANY_NONZERO" ; |
275 | case MipsISD::VCEQ: return "MipsISD::VCEQ" ; |
276 | case MipsISD::VCLE_S: return "MipsISD::VCLE_S" ; |
277 | case MipsISD::VCLE_U: return "MipsISD::VCLE_U" ; |
278 | case MipsISD::VCLT_S: return "MipsISD::VCLT_S" ; |
279 | case MipsISD::VCLT_U: return "MipsISD::VCLT_U" ; |
280 | case MipsISD::VEXTRACT_SEXT_ELT: return "MipsISD::VEXTRACT_SEXT_ELT" ; |
281 | case MipsISD::VEXTRACT_ZEXT_ELT: return "MipsISD::VEXTRACT_ZEXT_ELT" ; |
282 | case MipsISD::VNOR: return "MipsISD::VNOR" ; |
283 | case MipsISD::VSHF: return "MipsISD::VSHF" ; |
284 | case MipsISD::SHF: return "MipsISD::SHF" ; |
285 | case MipsISD::ILVEV: return "MipsISD::ILVEV" ; |
286 | case MipsISD::ILVOD: return "MipsISD::ILVOD" ; |
287 | case MipsISD::ILVL: return "MipsISD::ILVL" ; |
288 | case MipsISD::ILVR: return "MipsISD::ILVR" ; |
289 | case MipsISD::PCKEV: return "MipsISD::PCKEV" ; |
290 | case MipsISD::PCKOD: return "MipsISD::PCKOD" ; |
291 | case MipsISD::INSVE: return "MipsISD::INSVE" ; |
292 | } |
293 | return nullptr; |
294 | } |
295 | |
296 | MipsTargetLowering::MipsTargetLowering(const MipsTargetMachine &TM, |
297 | const MipsSubtarget &STI) |
298 | : TargetLowering(TM), Subtarget(STI), ABI(TM.getABI()) { |
299 | // Mips does not have i1 type, so use i32 for |
300 | // setcc operations results (slt, sgt, ...). |
301 | setBooleanContents(ZeroOrOneBooleanContent); |
302 | setBooleanVectorContents(ZeroOrNegativeOneBooleanContent); |
303 | // The cmp.cond.fmt instruction in MIPS32r6/MIPS64r6 uses 0 and -1 like MSA |
304 | // does. Integer booleans still use 0 and 1. |
305 | if (Subtarget.hasMips32r6()) |
306 | setBooleanContents(IntTy: ZeroOrOneBooleanContent, |
307 | FloatTy: ZeroOrNegativeOneBooleanContent); |
308 | |
309 | // Load extented operations for i1 types must be promoted |
310 | for (MVT VT : MVT::integer_valuetypes()) { |
311 | setLoadExtAction(ExtType: ISD::EXTLOAD, ValVT: VT, MemVT: MVT::i1, Action: Promote); |
312 | setLoadExtAction(ExtType: ISD::ZEXTLOAD, ValVT: VT, MemVT: MVT::i1, Action: Promote); |
313 | setLoadExtAction(ExtType: ISD::SEXTLOAD, ValVT: VT, MemVT: MVT::i1, Action: Promote); |
314 | } |
315 | |
316 | // MIPS doesn't have extending float->double load/store. Set LoadExtAction |
317 | // for f32, f16 |
318 | for (MVT VT : MVT::fp_valuetypes()) { |
319 | setLoadExtAction(ExtType: ISD::EXTLOAD, ValVT: VT, MemVT: MVT::f32, Action: Expand); |
320 | setLoadExtAction(ExtType: ISD::EXTLOAD, ValVT: VT, MemVT: MVT::f16, Action: Expand); |
321 | } |
322 | |
323 | // Set LoadExtAction for f16 vectors to Expand |
324 | for (MVT VT : MVT::fp_fixedlen_vector_valuetypes()) { |
325 | MVT F16VT = MVT::getVectorVT(VT: MVT::f16, NumElements: VT.getVectorNumElements()); |
326 | if (F16VT.isValid()) |
327 | setLoadExtAction(ExtType: ISD::EXTLOAD, ValVT: VT, MemVT: F16VT, Action: Expand); |
328 | } |
329 | |
330 | setTruncStoreAction(ValVT: MVT::f32, MemVT: MVT::f16, Action: Expand); |
331 | setTruncStoreAction(ValVT: MVT::f64, MemVT: MVT::f16, Action: Expand); |
332 | |
333 | setTruncStoreAction(ValVT: MVT::f64, MemVT: MVT::f32, Action: Expand); |
334 | |
335 | // Used by legalize types to correctly generate the setcc result. |
336 | // Without this, every float setcc comes with a AND/OR with the result, |
337 | // we don't want this, since the fpcmp result goes to a flag register, |
338 | // which is used implicitly by brcond and select operations. |
339 | AddPromotedToType(Opc: ISD::SETCC, OrigVT: MVT::i1, DestVT: MVT::i32); |
340 | |
341 | // Mips Custom Operations |
342 | setOperationAction(Op: ISD::BR_JT, VT: MVT::Other, Action: Expand); |
343 | setOperationAction(Op: ISD::GlobalAddress, VT: MVT::i32, Action: Custom); |
344 | setOperationAction(Op: ISD::BlockAddress, VT: MVT::i32, Action: Custom); |
345 | setOperationAction(Op: ISD::GlobalTLSAddress, VT: MVT::i32, Action: Custom); |
346 | setOperationAction(Op: ISD::JumpTable, VT: MVT::i32, Action: Custom); |
347 | setOperationAction(Op: ISD::ConstantPool, VT: MVT::i32, Action: Custom); |
348 | setOperationAction(Op: ISD::SELECT, VT: MVT::f32, Action: Custom); |
349 | setOperationAction(Op: ISD::SELECT, VT: MVT::f64, Action: Custom); |
350 | setOperationAction(Op: ISD::SELECT, VT: MVT::i32, Action: Custom); |
351 | setOperationAction(Op: ISD::SETCC, VT: MVT::f32, Action: Custom); |
352 | setOperationAction(Op: ISD::SETCC, VT: MVT::f64, Action: Custom); |
353 | setOperationAction(Op: ISD::BRCOND, VT: MVT::Other, Action: Custom); |
354 | setOperationAction(Op: ISD::FABS, VT: MVT::f32, Action: Custom); |
355 | setOperationAction(Op: ISD::FABS, VT: MVT::f64, Action: Custom); |
356 | setOperationAction(Op: ISD::FCOPYSIGN, VT: MVT::f32, Action: Custom); |
357 | setOperationAction(Op: ISD::FCOPYSIGN, VT: MVT::f64, Action: Custom); |
358 | setOperationAction(Op: ISD::FP_TO_SINT, VT: MVT::i32, Action: Custom); |
359 | |
360 | if (Subtarget.hasMips32r2() || |
361 | getTargetMachine().getTargetTriple().isOSLinux()) |
362 | setOperationAction(Op: ISD::READCYCLECOUNTER, VT: MVT::i64, Action: Custom); |
363 | |
364 | // Lower fmin/fmax/fclass operations for MIPS R6. |
365 | if (Subtarget.hasMips32r6()) { |
366 | setOperationAction(Op: ISD::FMINNUM_IEEE, VT: MVT::f32, Action: Legal); |
367 | setOperationAction(Op: ISD::FMAXNUM_IEEE, VT: MVT::f32, Action: Legal); |
368 | setOperationAction(Op: ISD::FMINNUM, VT: MVT::f32, Action: Legal); |
369 | setOperationAction(Op: ISD::FMAXNUM, VT: MVT::f32, Action: Legal); |
370 | setOperationAction(Op: ISD::FMINNUM_IEEE, VT: MVT::f64, Action: Legal); |
371 | setOperationAction(Op: ISD::FMAXNUM_IEEE, VT: MVT::f64, Action: Legal); |
372 | setOperationAction(Op: ISD::FMINNUM, VT: MVT::f64, Action: Legal); |
373 | setOperationAction(Op: ISD::FMAXNUM, VT: MVT::f64, Action: Legal); |
374 | setOperationAction(Op: ISD::IS_FPCLASS, VT: MVT::f32, Action: Legal); |
375 | setOperationAction(Op: ISD::IS_FPCLASS, VT: MVT::f64, Action: Legal); |
376 | } else { |
377 | setOperationAction(Op: ISD::FCANONICALIZE, VT: MVT::f32, Action: Custom); |
378 | setOperationAction(Op: ISD::FCANONICALIZE, VT: MVT::f64, Action: Custom); |
379 | } |
380 | |
381 | if (Subtarget.isGP64bit()) { |
382 | setOperationAction(Op: ISD::GlobalAddress, VT: MVT::i64, Action: Custom); |
383 | setOperationAction(Op: ISD::BlockAddress, VT: MVT::i64, Action: Custom); |
384 | setOperationAction(Op: ISD::GlobalTLSAddress, VT: MVT::i64, Action: Custom); |
385 | setOperationAction(Op: ISD::JumpTable, VT: MVT::i64, Action: Custom); |
386 | setOperationAction(Op: ISD::ConstantPool, VT: MVT::i64, Action: Custom); |
387 | setOperationAction(Op: ISD::SELECT, VT: MVT::i64, Action: Custom); |
388 | if (Subtarget.hasMips64r6()) { |
389 | setOperationAction(Op: ISD::LOAD, VT: MVT::i64, Action: Legal); |
390 | setOperationAction(Op: ISD::STORE, VT: MVT::i64, Action: Legal); |
391 | } else { |
392 | setOperationAction(Op: ISD::LOAD, VT: MVT::i64, Action: Custom); |
393 | setOperationAction(Op: ISD::STORE, VT: MVT::i64, Action: Custom); |
394 | } |
395 | setOperationAction(Op: ISD::FP_TO_SINT, VT: MVT::i64, Action: Custom); |
396 | setOperationAction(Op: ISD::SHL_PARTS, VT: MVT::i64, Action: Custom); |
397 | setOperationAction(Op: ISD::SRA_PARTS, VT: MVT::i64, Action: Custom); |
398 | setOperationAction(Op: ISD::SRL_PARTS, VT: MVT::i64, Action: Custom); |
399 | } |
400 | |
401 | if (!Subtarget.isGP64bit()) { |
402 | setOperationAction(Op: ISD::SHL_PARTS, VT: MVT::i32, Action: Custom); |
403 | setOperationAction(Op: ISD::SRA_PARTS, VT: MVT::i32, Action: Custom); |
404 | setOperationAction(Op: ISD::SRL_PARTS, VT: MVT::i32, Action: Custom); |
405 | } |
406 | |
407 | setOperationAction(Op: ISD::EH_DWARF_CFA, VT: MVT::i32, Action: Custom); |
408 | if (Subtarget.isGP64bit()) |
409 | setOperationAction(Op: ISD::EH_DWARF_CFA, VT: MVT::i64, Action: Custom); |
410 | |
411 | setOperationAction(Op: ISD::SDIV, VT: MVT::i32, Action: Expand); |
412 | setOperationAction(Op: ISD::SREM, VT: MVT::i32, Action: Expand); |
413 | setOperationAction(Op: ISD::UDIV, VT: MVT::i32, Action: Expand); |
414 | setOperationAction(Op: ISD::UREM, VT: MVT::i32, Action: Expand); |
415 | setOperationAction(Op: ISD::SDIV, VT: MVT::i64, Action: Expand); |
416 | setOperationAction(Op: ISD::SREM, VT: MVT::i64, Action: Expand); |
417 | setOperationAction(Op: ISD::UDIV, VT: MVT::i64, Action: Expand); |
418 | setOperationAction(Op: ISD::UREM, VT: MVT::i64, Action: Expand); |
419 | |
420 | // Operations not directly supported by Mips. |
421 | setOperationAction(Op: ISD::BR_CC, VT: MVT::f32, Action: Expand); |
422 | setOperationAction(Op: ISD::BR_CC, VT: MVT::f64, Action: Expand); |
423 | setOperationAction(Op: ISD::BR_CC, VT: MVT::i32, Action: Expand); |
424 | setOperationAction(Op: ISD::BR_CC, VT: MVT::i64, Action: Expand); |
425 | setOperationAction(Op: ISD::SELECT_CC, VT: MVT::i32, Action: Expand); |
426 | setOperationAction(Op: ISD::SELECT_CC, VT: MVT::i64, Action: Expand); |
427 | setOperationAction(Op: ISD::SELECT_CC, VT: MVT::f32, Action: Expand); |
428 | setOperationAction(Op: ISD::SELECT_CC, VT: MVT::f64, Action: Expand); |
429 | setOperationAction(Op: ISD::UINT_TO_FP, VT: MVT::i32, Action: Expand); |
430 | setOperationAction(Op: ISD::UINT_TO_FP, VT: MVT::i64, Action: Expand); |
431 | setOperationAction(Op: ISD::FP_TO_UINT, VT: MVT::i32, Action: Expand); |
432 | setOperationAction(Op: ISD::FP_TO_UINT, VT: MVT::i64, Action: Expand); |
433 | setOperationAction(Op: ISD::SIGN_EXTEND_INREG, VT: MVT::i1, Action: Expand); |
434 | if (Subtarget.hasCnMips()) { |
435 | setOperationAction(Op: ISD::CTPOP, VT: MVT::i32, Action: Legal); |
436 | setOperationAction(Op: ISD::CTPOP, VT: MVT::i64, Action: Legal); |
437 | } else { |
438 | setOperationAction(Op: ISD::CTPOP, VT: MVT::i32, Action: Expand); |
439 | setOperationAction(Op: ISD::CTPOP, VT: MVT::i64, Action: Expand); |
440 | } |
441 | setOperationAction(Op: ISD::CTTZ, VT: MVT::i32, Action: Expand); |
442 | setOperationAction(Op: ISD::CTTZ, VT: MVT::i64, Action: Expand); |
443 | setOperationAction(Op: ISD::ROTL, VT: MVT::i32, Action: Expand); |
444 | setOperationAction(Op: ISD::ROTL, VT: MVT::i64, Action: Expand); |
445 | setOperationAction(Op: ISD::DYNAMIC_STACKALLOC, VT: MVT::i32, Action: Expand); |
446 | setOperationAction(Op: ISD::DYNAMIC_STACKALLOC, VT: MVT::i64, Action: Expand); |
447 | |
448 | if (!Subtarget.hasMips32r2()) |
449 | setOperationAction(Op: ISD::ROTR, VT: MVT::i32, Action: Expand); |
450 | |
451 | if (!Subtarget.hasMips64r2()) |
452 | setOperationAction(Op: ISD::ROTR, VT: MVT::i64, Action: Expand); |
453 | |
454 | setOperationAction(Op: ISD::FSIN, VT: MVT::f32, Action: Expand); |
455 | setOperationAction(Op: ISD::FSIN, VT: MVT::f64, Action: Expand); |
456 | setOperationAction(Op: ISD::FCOS, VT: MVT::f32, Action: Expand); |
457 | setOperationAction(Op: ISD::FCOS, VT: MVT::f64, Action: Expand); |
458 | setOperationAction(Op: ISD::FSINCOS, VT: MVT::f32, Action: Expand); |
459 | setOperationAction(Op: ISD::FSINCOS, VT: MVT::f64, Action: Expand); |
460 | setOperationAction(Op: ISD::FPOW, VT: MVT::f32, Action: Expand); |
461 | setOperationAction(Op: ISD::FPOW, VT: MVT::f64, Action: Expand); |
462 | setOperationAction(Op: ISD::FLOG, VT: MVT::f32, Action: Expand); |
463 | setOperationAction(Op: ISD::FLOG2, VT: MVT::f32, Action: Expand); |
464 | setOperationAction(Op: ISD::FLOG10, VT: MVT::f32, Action: Expand); |
465 | setOperationAction(Op: ISD::FEXP, VT: MVT::f32, Action: Expand); |
466 | setOperationAction(Op: ISD::FMA, VT: MVT::f32, Action: Expand); |
467 | setOperationAction(Op: ISD::FMA, VT: MVT::f64, Action: Expand); |
468 | setOperationAction(Op: ISD::FREM, VT: MVT::f32, Action: Expand); |
469 | setOperationAction(Op: ISD::FREM, VT: MVT::f64, Action: Expand); |
470 | |
471 | // Lower f16 conversion operations into library calls |
472 | setOperationAction(Op: ISD::FP16_TO_FP, VT: MVT::f32, Action: Expand); |
473 | setOperationAction(Op: ISD::FP_TO_FP16, VT: MVT::f32, Action: Expand); |
474 | setOperationAction(Op: ISD::FP16_TO_FP, VT: MVT::f64, Action: Expand); |
475 | setOperationAction(Op: ISD::FP_TO_FP16, VT: MVT::f64, Action: Expand); |
476 | |
477 | setOperationAction(Op: ISD::EH_RETURN, VT: MVT::Other, Action: Custom); |
478 | |
479 | setOperationAction(Op: ISD::VASTART, VT: MVT::Other, Action: Custom); |
480 | setOperationAction(Op: ISD::VAARG, VT: MVT::Other, Action: Custom); |
481 | setOperationAction(Op: ISD::VACOPY, VT: MVT::Other, Action: Expand); |
482 | setOperationAction(Op: ISD::VAEND, VT: MVT::Other, Action: Expand); |
483 | |
484 | // Use the default for now |
485 | setOperationAction(Op: ISD::STACKSAVE, VT: MVT::Other, Action: Expand); |
486 | setOperationAction(Op: ISD::STACKRESTORE, VT: MVT::Other, Action: Expand); |
487 | |
488 | if (!Subtarget.isGP64bit()) { |
489 | setOperationAction(Op: ISD::ATOMIC_LOAD, VT: MVT::i64, Action: Expand); |
490 | setOperationAction(Op: ISD::ATOMIC_STORE, VT: MVT::i64, Action: Expand); |
491 | } |
492 | |
493 | if (!Subtarget.hasMips32r2()) { |
494 | setOperationAction(Op: ISD::SIGN_EXTEND_INREG, VT: MVT::i8, Action: Expand); |
495 | setOperationAction(Op: ISD::SIGN_EXTEND_INREG, VT: MVT::i16, Action: Expand); |
496 | } |
497 | |
498 | // MIPS16 lacks MIPS32's clz and clo instructions. |
499 | if (!Subtarget.hasMips32() || Subtarget.inMips16Mode()) |
500 | setOperationAction(Op: ISD::CTLZ, VT: MVT::i32, Action: Expand); |
501 | if (!Subtarget.hasMips64()) |
502 | setOperationAction(Op: ISD::CTLZ, VT: MVT::i64, Action: Expand); |
503 | |
504 | if (!Subtarget.hasMips32r2()) |
505 | setOperationAction(Op: ISD::BSWAP, VT: MVT::i32, Action: Expand); |
506 | if (!Subtarget.hasMips64r2()) |
507 | setOperationAction(Op: ISD::BSWAP, VT: MVT::i64, Action: Expand); |
508 | |
509 | if (Subtarget.isGP64bit() && Subtarget.hasMips64r6()) { |
510 | setLoadExtAction(ExtType: ISD::SEXTLOAD, ValVT: MVT::i64, MemVT: MVT::i32, Action: Legal); |
511 | setLoadExtAction(ExtType: ISD::ZEXTLOAD, ValVT: MVT::i64, MemVT: MVT::i32, Action: Legal); |
512 | setLoadExtAction(ExtType: ISD::EXTLOAD, ValVT: MVT::i64, MemVT: MVT::i32, Action: Legal); |
513 | setTruncStoreAction(ValVT: MVT::i64, MemVT: MVT::i32, Action: Legal); |
514 | } else if (Subtarget.isGP64bit()) { |
515 | setLoadExtAction(ExtType: ISD::SEXTLOAD, ValVT: MVT::i64, MemVT: MVT::i32, Action: Custom); |
516 | setLoadExtAction(ExtType: ISD::ZEXTLOAD, ValVT: MVT::i64, MemVT: MVT::i32, Action: Custom); |
517 | setLoadExtAction(ExtType: ISD::EXTLOAD, ValVT: MVT::i64, MemVT: MVT::i32, Action: Custom); |
518 | setTruncStoreAction(ValVT: MVT::i64, MemVT: MVT::i32, Action: Custom); |
519 | } |
520 | |
521 | setOperationAction(Op: ISD::TRAP, VT: MVT::Other, Action: Legal); |
522 | |
523 | setOperationAction(Op: ISD::ConstantFP, VT: MVT::f32, Action: Custom); |
524 | setOperationAction(Op: ISD::ConstantFP, VT: MVT::f64, Action: Custom); |
525 | |
526 | setTargetDAGCombine({ISD::SDIVREM, ISD::UDIVREM, ISD::SELECT, ISD::AND, |
527 | ISD::OR, ISD::ADD, ISD::SUB, ISD::AssertZext, ISD::SHL, |
528 | ISD::SIGN_EXTEND}); |
529 | |
530 | if (Subtarget.isGP64bit()) |
531 | setMaxAtomicSizeInBitsSupported(64); |
532 | else |
533 | setMaxAtomicSizeInBitsSupported(32); |
534 | |
535 | setMinFunctionAlignment(Subtarget.isGP64bit() ? Align(8) : Align(4)); |
536 | |
537 | // The arguments on the stack are defined in terms of 4-byte slots on O32 |
538 | // and 8-byte slots on N32/N64. |
539 | setMinStackArgumentAlignment((ABI.IsN32() || ABI.IsN64()) ? Align(8) |
540 | : Align(4)); |
541 | |
542 | setStackPointerRegisterToSaveRestore(ABI.IsN64() ? Mips::SP_64 : Mips::SP); |
543 | |
544 | MaxStoresPerMemcpy = 16; |
545 | |
546 | isMicroMips = Subtarget.inMicroMipsMode(); |
547 | } |
548 | |
549 | const MipsTargetLowering * |
550 | MipsTargetLowering::create(const MipsTargetMachine &TM, |
551 | const MipsSubtarget &STI) { |
552 | if (STI.inMips16Mode()) |
553 | return createMips16TargetLowering(TM, STI); |
554 | |
555 | return createMipsSETargetLowering(TM, STI); |
556 | } |
557 | |
558 | // Create a fast isel object. |
559 | FastISel * |
560 | MipsTargetLowering::createFastISel(FunctionLoweringInfo &funcInfo, |
561 | const TargetLibraryInfo *libInfo) const { |
562 | const MipsTargetMachine &TM = |
563 | static_cast<const MipsTargetMachine &>(funcInfo.MF->getTarget()); |
564 | |
565 | // We support only the standard encoding [MIPS32,MIPS32R5] ISAs. |
566 | bool UseFastISel = TM.Options.EnableFastISel && Subtarget.hasMips32() && |
567 | !Subtarget.hasMips32r6() && !Subtarget.inMips16Mode() && |
568 | !Subtarget.inMicroMipsMode(); |
569 | |
570 | // Disable if either of the following is true: |
571 | // We do not generate PIC, the ABI is not O32, XGOT is being used. |
572 | if (!TM.isPositionIndependent() || !TM.getABI().IsO32() || |
573 | Subtarget.useXGOT()) |
574 | UseFastISel = false; |
575 | |
576 | return UseFastISel ? Mips::createFastISel(funcInfo, libInfo) : nullptr; |
577 | } |
578 | |
579 | EVT MipsTargetLowering::getSetCCResultType(const DataLayout &, LLVMContext &, |
580 | EVT VT) const { |
581 | if (!VT.isVector()) |
582 | return MVT::i32; |
583 | return VT.changeVectorElementTypeToInteger(); |
584 | } |
585 | |
586 | static SDValue performDivRemCombine(SDNode *N, SelectionDAG &DAG, |
587 | TargetLowering::DAGCombinerInfo &DCI, |
588 | const MipsSubtarget &Subtarget) { |
589 | if (DCI.isBeforeLegalizeOps()) |
590 | return SDValue(); |
591 | |
592 | EVT Ty = N->getValueType(ResNo: 0); |
593 | unsigned LO = (Ty == MVT::i32) ? Mips::LO0 : Mips::LO0_64; |
594 | unsigned HI = (Ty == MVT::i32) ? Mips::HI0 : Mips::HI0_64; |
595 | unsigned Opc = N->getOpcode() == ISD::SDIVREM ? MipsISD::DivRem16 : |
596 | MipsISD::DivRemU16; |
597 | SDLoc DL(N); |
598 | |
599 | SDValue DivRem = DAG.getNode(Opcode: Opc, DL, VT: MVT::Glue, |
600 | N1: N->getOperand(Num: 0), N2: N->getOperand(Num: 1)); |
601 | SDValue InChain = DAG.getEntryNode(); |
602 | SDValue InGlue = DivRem; |
603 | |
604 | // insert MFLO |
605 | if (N->hasAnyUseOfValue(Value: 0)) { |
606 | SDValue CopyFromLo = DAG.getCopyFromReg(Chain: InChain, dl: DL, Reg: LO, VT: Ty, |
607 | Glue: InGlue); |
608 | DAG.ReplaceAllUsesOfValueWith(From: SDValue(N, 0), To: CopyFromLo); |
609 | InChain = CopyFromLo.getValue(R: 1); |
610 | InGlue = CopyFromLo.getValue(R: 2); |
611 | } |
612 | |
613 | // insert MFHI |
614 | if (N->hasAnyUseOfValue(Value: 1)) { |
615 | SDValue CopyFromHi = DAG.getCopyFromReg(Chain: InChain, dl: DL, |
616 | Reg: HI, VT: Ty, Glue: InGlue); |
617 | DAG.ReplaceAllUsesOfValueWith(From: SDValue(N, 1), To: CopyFromHi); |
618 | } |
619 | |
620 | return SDValue(); |
621 | } |
622 | |
623 | static Mips::CondCode condCodeToFCC(ISD::CondCode CC) { |
624 | switch (CC) { |
625 | default: llvm_unreachable("Unknown fp condition code!" ); |
626 | case ISD::SETEQ: |
627 | case ISD::SETOEQ: return Mips::FCOND_OEQ; |
628 | case ISD::SETUNE: return Mips::FCOND_UNE; |
629 | case ISD::SETLT: |
630 | case ISD::SETOLT: return Mips::FCOND_OLT; |
631 | case ISD::SETGT: |
632 | case ISD::SETOGT: return Mips::FCOND_OGT; |
633 | case ISD::SETLE: |
634 | case ISD::SETOLE: return Mips::FCOND_OLE; |
635 | case ISD::SETGE: |
636 | case ISD::SETOGE: return Mips::FCOND_OGE; |
637 | case ISD::SETULT: return Mips::FCOND_ULT; |
638 | case ISD::SETULE: return Mips::FCOND_ULE; |
639 | case ISD::SETUGT: return Mips::FCOND_UGT; |
640 | case ISD::SETUGE: return Mips::FCOND_UGE; |
641 | case ISD::SETUO: return Mips::FCOND_UN; |
642 | case ISD::SETO: return Mips::FCOND_OR; |
643 | case ISD::SETNE: |
644 | case ISD::SETONE: return Mips::FCOND_ONE; |
645 | case ISD::SETUEQ: return Mips::FCOND_UEQ; |
646 | } |
647 | } |
648 | |
649 | /// This function returns true if the floating point conditional branches and |
650 | /// conditional moves which use condition code CC should be inverted. |
651 | static bool invertFPCondCodeUser(Mips::CondCode CC) { |
652 | if (CC >= Mips::FCOND_F && CC <= Mips::FCOND_NGT) |
653 | return false; |
654 | |
655 | assert((CC >= Mips::FCOND_T && CC <= Mips::FCOND_GT) && |
656 | "Illegal Condition Code" ); |
657 | |
658 | return true; |
659 | } |
660 | |
661 | // Creates and returns an FPCmp node from a setcc node. |
662 | // Returns Op if setcc is not a floating point comparison. |
663 | static SDValue createFPCmp(SelectionDAG &DAG, const SDValue &Op) { |
664 | // must be a SETCC node |
665 | if (Op.getOpcode() != ISD::SETCC) |
666 | return Op; |
667 | |
668 | SDValue LHS = Op.getOperand(i: 0); |
669 | |
670 | if (!LHS.getValueType().isFloatingPoint()) |
671 | return Op; |
672 | |
673 | SDValue RHS = Op.getOperand(i: 1); |
674 | SDLoc DL(Op); |
675 | |
676 | // Assume the 3rd operand is a CondCodeSDNode. Add code to check the type of |
677 | // node if necessary. |
678 | ISD::CondCode CC = cast<CondCodeSDNode>(Val: Op.getOperand(i: 2))->get(); |
679 | |
680 | return DAG.getNode(Opcode: MipsISD::FPCmp, DL, VT: MVT::Glue, N1: LHS, N2: RHS, |
681 | N3: DAG.getConstant(Val: condCodeToFCC(CC), DL, VT: MVT::i32)); |
682 | } |
683 | |
684 | // Creates and returns a CMovFPT/F node. |
685 | static SDValue createCMovFP(SelectionDAG &DAG, SDValue Cond, SDValue True, |
686 | SDValue False, const SDLoc &DL) { |
687 | ConstantSDNode *CC = cast<ConstantSDNode>(Val: Cond.getOperand(i: 2)); |
688 | bool invert = invertFPCondCodeUser(CC: (Mips::CondCode)CC->getSExtValue()); |
689 | SDValue FCC0 = DAG.getRegister(Reg: Mips::FCC0, VT: MVT::i32); |
690 | |
691 | return DAG.getNode(Opcode: (invert ? MipsISD::CMovFP_F : MipsISD::CMovFP_T), DL, |
692 | VT: True.getValueType(), N1: True, N2: FCC0, N3: False, N4: Cond); |
693 | } |
694 | |
695 | static SDValue performSELECTCombine(SDNode *N, SelectionDAG &DAG, |
696 | TargetLowering::DAGCombinerInfo &DCI, |
697 | const MipsSubtarget &Subtarget) { |
698 | if (DCI.isBeforeLegalizeOps()) |
699 | return SDValue(); |
700 | |
701 | SDValue SetCC = N->getOperand(Num: 0); |
702 | |
703 | if ((SetCC.getOpcode() != ISD::SETCC) || |
704 | !SetCC.getOperand(i: 0).getValueType().isInteger()) |
705 | return SDValue(); |
706 | |
707 | SDValue False = N->getOperand(Num: 2); |
708 | EVT FalseTy = False.getValueType(); |
709 | |
710 | if (!FalseTy.isInteger()) |
711 | return SDValue(); |
712 | |
713 | ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(Val&: False); |
714 | |
715 | // If the RHS (False) is 0, we swap the order of the operands |
716 | // of ISD::SELECT (obviously also inverting the condition) so that we can |
717 | // take advantage of conditional moves using the $0 register. |
718 | // Example: |
719 | // return (a != 0) ? x : 0; |
720 | // load $reg, x |
721 | // movz $reg, $0, a |
722 | if (!FalseC) |
723 | return SDValue(); |
724 | |
725 | const SDLoc DL(N); |
726 | |
727 | if (!FalseC->getZExtValue()) { |
728 | ISD::CondCode CC = cast<CondCodeSDNode>(Val: SetCC.getOperand(i: 2))->get(); |
729 | SDValue True = N->getOperand(Num: 1); |
730 | |
731 | SetCC = DAG.getSetCC(DL, VT: SetCC.getValueType(), LHS: SetCC.getOperand(i: 0), |
732 | RHS: SetCC.getOperand(i: 1), |
733 | Cond: ISD::getSetCCInverse(Operation: CC, Type: SetCC.getValueType())); |
734 | |
735 | return DAG.getNode(Opcode: ISD::SELECT, DL, VT: FalseTy, N1: SetCC, N2: False, N3: True); |
736 | } |
737 | |
738 | // If both operands are integer constants there's a possibility that we |
739 | // can do some interesting optimizations. |
740 | SDValue True = N->getOperand(Num: 1); |
741 | ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(Val&: True); |
742 | |
743 | if (!TrueC || !True.getValueType().isInteger()) |
744 | return SDValue(); |
745 | |
746 | // We'll also ignore MVT::i64 operands as this optimizations proves |
747 | // to be ineffective because of the required sign extensions as the result |
748 | // of a SETCC operator is always MVT::i32 for non-vector types. |
749 | if (True.getValueType() == MVT::i64) |
750 | return SDValue(); |
751 | |
752 | int64_t Diff = TrueC->getSExtValue() - FalseC->getSExtValue(); |
753 | |
754 | // 1) (a < x) ? y : y-1 |
755 | // slti $reg1, a, x |
756 | // addiu $reg2, $reg1, y-1 |
757 | if (Diff == 1) |
758 | return DAG.getNode(Opcode: ISD::ADD, DL, VT: SetCC.getValueType(), N1: SetCC, N2: False); |
759 | |
760 | // 2) (a < x) ? y-1 : y |
761 | // slti $reg1, a, x |
762 | // xor $reg1, $reg1, 1 |
763 | // addiu $reg2, $reg1, y-1 |
764 | if (Diff == -1) { |
765 | ISD::CondCode CC = cast<CondCodeSDNode>(Val: SetCC.getOperand(i: 2))->get(); |
766 | SetCC = DAG.getSetCC(DL, VT: SetCC.getValueType(), LHS: SetCC.getOperand(i: 0), |
767 | RHS: SetCC.getOperand(i: 1), |
768 | Cond: ISD::getSetCCInverse(Operation: CC, Type: SetCC.getValueType())); |
769 | return DAG.getNode(Opcode: ISD::ADD, DL, VT: SetCC.getValueType(), N1: SetCC, N2: True); |
770 | } |
771 | |
772 | // Could not optimize. |
773 | return SDValue(); |
774 | } |
775 | |
776 | static SDValue performCMovFPCombine(SDNode *N, SelectionDAG &DAG, |
777 | TargetLowering::DAGCombinerInfo &DCI, |
778 | const MipsSubtarget &Subtarget) { |
779 | if (DCI.isBeforeLegalizeOps()) |
780 | return SDValue(); |
781 | |
782 | SDValue ValueIfTrue = N->getOperand(Num: 0), ValueIfFalse = N->getOperand(Num: 2); |
783 | |
784 | ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(Val&: ValueIfFalse); |
785 | if (!FalseC || FalseC->getZExtValue()) |
786 | return SDValue(); |
787 | |
788 | // Since RHS (False) is 0, we swap the order of the True/False operands |
789 | // (obviously also inverting the condition) so that we can |
790 | // take advantage of conditional moves using the $0 register. |
791 | // Example: |
792 | // return (a != 0) ? x : 0; |
793 | // load $reg, x |
794 | // movz $reg, $0, a |
795 | unsigned Opc = (N->getOpcode() == MipsISD::CMovFP_T) ? MipsISD::CMovFP_F : |
796 | MipsISD::CMovFP_T; |
797 | |
798 | SDValue FCC = N->getOperand(Num: 1), Glue = N->getOperand(Num: 3); |
799 | return DAG.getNode(Opcode: Opc, DL: SDLoc(N), VT: ValueIfFalse.getValueType(), |
800 | N1: ValueIfFalse, N2: FCC, N3: ValueIfTrue, N4: Glue); |
801 | } |
802 | |
803 | static SDValue performANDCombine(SDNode *N, SelectionDAG &DAG, |
804 | TargetLowering::DAGCombinerInfo &DCI, |
805 | const MipsSubtarget &Subtarget) { |
806 | if (DCI.isBeforeLegalizeOps() || !Subtarget.hasExtractInsert()) |
807 | return SDValue(); |
808 | |
809 | SDValue FirstOperand = N->getOperand(Num: 0); |
810 | unsigned FirstOperandOpc = FirstOperand.getOpcode(); |
811 | SDValue Mask = N->getOperand(Num: 1); |
812 | EVT ValTy = N->getValueType(ResNo: 0); |
813 | SDLoc DL(N); |
814 | |
815 | uint64_t Pos = 0; |
816 | unsigned SMPos, SMSize; |
817 | ConstantSDNode *CN; |
818 | SDValue NewOperand; |
819 | unsigned Opc; |
820 | |
821 | // Op's second operand must be a shifted mask. |
822 | if (!(CN = dyn_cast<ConstantSDNode>(Val&: Mask)) || |
823 | !isShiftedMask_64(Value: CN->getZExtValue(), MaskIdx&: SMPos, MaskLen&: SMSize)) |
824 | return SDValue(); |
825 | |
826 | if (FirstOperandOpc == ISD::SRA || FirstOperandOpc == ISD::SRL) { |
827 | // Pattern match EXT. |
828 | // $dst = and ((sra or srl) $src , pos), (2**size - 1) |
829 | // => ext $dst, $src, pos, size |
830 | |
831 | // The second operand of the shift must be an immediate. |
832 | if (!(CN = dyn_cast<ConstantSDNode>(Val: FirstOperand.getOperand(i: 1)))) |
833 | return SDValue(); |
834 | |
835 | Pos = CN->getZExtValue(); |
836 | |
837 | // Return if the shifted mask does not start at bit 0 or the sum of its size |
838 | // and Pos exceeds the word's size. |
839 | if (SMPos != 0 || Pos + SMSize > ValTy.getSizeInBits()) |
840 | return SDValue(); |
841 | |
842 | Opc = MipsISD::Ext; |
843 | NewOperand = FirstOperand.getOperand(i: 0); |
844 | } else if (FirstOperandOpc == ISD::SHL && Subtarget.hasCnMips()) { |
845 | // Pattern match CINS. |
846 | // $dst = and (shl $src , pos), mask |
847 | // => cins $dst, $src, pos, size |
848 | // mask is a shifted mask with consecutive 1's, pos = shift amount, |
849 | // size = population count. |
850 | |
851 | // The second operand of the shift must be an immediate. |
852 | if (!(CN = dyn_cast<ConstantSDNode>(Val: FirstOperand.getOperand(i: 1)))) |
853 | return SDValue(); |
854 | |
855 | Pos = CN->getZExtValue(); |
856 | |
857 | if (SMPos != Pos || Pos >= ValTy.getSizeInBits() || SMSize >= 32 || |
858 | Pos + SMSize > ValTy.getSizeInBits()) |
859 | return SDValue(); |
860 | |
861 | NewOperand = FirstOperand.getOperand(i: 0); |
862 | // SMSize is 'location' (position) in this case, not size. |
863 | SMSize--; |
864 | Opc = MipsISD::CIns; |
865 | } else { |
866 | // Pattern match EXT. |
867 | // $dst = and $src, (2**size - 1) , if size > 16 |
868 | // => ext $dst, $src, pos, size , pos = 0 |
869 | |
870 | // If the mask is <= 0xffff, andi can be used instead. |
871 | if (CN->getZExtValue() <= 0xffff) |
872 | return SDValue(); |
873 | |
874 | // Return if the mask doesn't start at position 0. |
875 | if (SMPos) |
876 | return SDValue(); |
877 | |
878 | Opc = MipsISD::Ext; |
879 | NewOperand = FirstOperand; |
880 | } |
881 | return DAG.getNode(Opcode: Opc, DL, VT: ValTy, N1: NewOperand, |
882 | N2: DAG.getConstant(Val: Pos, DL, VT: MVT::i32), |
883 | N3: DAG.getConstant(Val: SMSize, DL, VT: MVT::i32)); |
884 | } |
885 | |
886 | static SDValue performORCombine(SDNode *N, SelectionDAG &DAG, |
887 | TargetLowering::DAGCombinerInfo &DCI, |
888 | const MipsSubtarget &Subtarget) { |
889 | if (DCI.isBeforeLegalizeOps() || !Subtarget.hasExtractInsert()) |
890 | return SDValue(); |
891 | |
892 | SDValue FirstOperand = N->getOperand(Num: 0), SecondOperand = N->getOperand(Num: 1); |
893 | unsigned SMPos0, SMSize0, SMPos1, SMSize1; |
894 | ConstantSDNode *CN, *CN1; |
895 | |
896 | if ((FirstOperand.getOpcode() == ISD::AND && |
897 | SecondOperand.getOpcode() == ISD::SHL) || |
898 | (FirstOperand.getOpcode() == ISD::SHL && |
899 | SecondOperand.getOpcode() == ISD::AND)) { |
900 | // Pattern match INS. |
901 | // $dst = or (and $src1, (2**size0 - 1)), (shl $src2, size0) |
902 | // ==> ins $src1, $src2, pos, size, pos = size0, size = 32 - pos; |
903 | // Or: |
904 | // $dst = or (shl $src2, size0), (and $src1, (2**size0 - 1)) |
905 | // ==> ins $src1, $src2, pos, size, pos = size0, size = 32 - pos; |
906 | SDValue AndOperand0 = FirstOperand.getOpcode() == ISD::AND |
907 | ? FirstOperand.getOperand(i: 0) |
908 | : SecondOperand.getOperand(i: 0); |
909 | SDValue ShlOperand0 = FirstOperand.getOpcode() == ISD::AND |
910 | ? SecondOperand.getOperand(i: 0) |
911 | : FirstOperand.getOperand(i: 0); |
912 | SDValue AndMask = FirstOperand.getOpcode() == ISD::AND |
913 | ? FirstOperand.getOperand(i: 1) |
914 | : SecondOperand.getOperand(i: 1); |
915 | if (!(CN = dyn_cast<ConstantSDNode>(Val&: AndMask)) || |
916 | !isShiftedMask_64(Value: CN->getZExtValue(), MaskIdx&: SMPos0, MaskLen&: SMSize0)) |
917 | return SDValue(); |
918 | |
919 | SDValue ShlShift = FirstOperand.getOpcode() == ISD::AND |
920 | ? SecondOperand.getOperand(i: 1) |
921 | : FirstOperand.getOperand(i: 1); |
922 | if (!(CN = dyn_cast<ConstantSDNode>(Val&: ShlShift))) |
923 | return SDValue(); |
924 | uint64_t ShlShiftValue = CN->getZExtValue(); |
925 | |
926 | if (SMPos0 != 0 || SMSize0 != ShlShiftValue) |
927 | return SDValue(); |
928 | |
929 | SDLoc DL(N); |
930 | EVT ValTy = N->getValueType(ResNo: 0); |
931 | SMPos1 = ShlShiftValue; |
932 | assert(SMPos1 < ValTy.getSizeInBits()); |
933 | SMSize1 = (ValTy == MVT::i64 ? 64 : 32) - SMPos1; |
934 | return DAG.getNode(Opcode: MipsISD::Ins, DL, VT: ValTy, N1: ShlOperand0, |
935 | N2: DAG.getConstant(Val: SMPos1, DL, VT: MVT::i32), |
936 | N3: DAG.getConstant(Val: SMSize1, DL, VT: MVT::i32), N4: AndOperand0); |
937 | } |
938 | |
939 | // See if Op's first operand matches (and $src1 , mask0). |
940 | if (FirstOperand.getOpcode() != ISD::AND) |
941 | return SDValue(); |
942 | |
943 | // Pattern match INS. |
944 | // $dst = or (and $src1 , mask0), (and (shl $src, pos), mask1), |
945 | // where mask1 = (2**size - 1) << pos, mask0 = ~mask1 |
946 | // => ins $dst, $src, size, pos, $src1 |
947 | if (!(CN = dyn_cast<ConstantSDNode>(Val: FirstOperand.getOperand(i: 1))) || |
948 | !isShiftedMask_64(Value: ~CN->getSExtValue(), MaskIdx&: SMPos0, MaskLen&: SMSize0)) |
949 | return SDValue(); |
950 | |
951 | // See if Op's second operand matches (and (shl $src, pos), mask1). |
952 | if (SecondOperand.getOpcode() == ISD::AND && |
953 | SecondOperand.getOperand(i: 0).getOpcode() == ISD::SHL) { |
954 | |
955 | if (!(CN = dyn_cast<ConstantSDNode>(Val: SecondOperand.getOperand(i: 1))) || |
956 | !isShiftedMask_64(Value: CN->getZExtValue(), MaskIdx&: SMPos1, MaskLen&: SMSize1)) |
957 | return SDValue(); |
958 | |
959 | // The shift masks must have the same position and size. |
960 | if (SMPos0 != SMPos1 || SMSize0 != SMSize1) |
961 | return SDValue(); |
962 | |
963 | SDValue Shl = SecondOperand.getOperand(i: 0); |
964 | |
965 | if (!(CN = dyn_cast<ConstantSDNode>(Val: Shl.getOperand(i: 1)))) |
966 | return SDValue(); |
967 | |
968 | unsigned Shamt = CN->getZExtValue(); |
969 | |
970 | // Return if the shift amount and the first bit position of mask are not the |
971 | // same. |
972 | EVT ValTy = N->getValueType(ResNo: 0); |
973 | if ((Shamt != SMPos0) || (SMPos0 + SMSize0 > ValTy.getSizeInBits())) |
974 | return SDValue(); |
975 | |
976 | SDLoc DL(N); |
977 | return DAG.getNode(Opcode: MipsISD::Ins, DL, VT: ValTy, N1: Shl.getOperand(i: 0), |
978 | N2: DAG.getConstant(Val: SMPos0, DL, VT: MVT::i32), |
979 | N3: DAG.getConstant(Val: SMSize0, DL, VT: MVT::i32), |
980 | N4: FirstOperand.getOperand(i: 0)); |
981 | } else { |
982 | // Pattern match DINS. |
983 | // $dst = or (and $src, mask0), mask1 |
984 | // where mask0 = ((1 << SMSize0) -1) << SMPos0 |
985 | // => dins $dst, $src, pos, size |
986 | if (~CN->getSExtValue() == ((((int64_t)1 << SMSize0) - 1) << SMPos0) && |
987 | ((SMSize0 + SMPos0 <= 64 && Subtarget.hasMips64r2()) || |
988 | (SMSize0 + SMPos0 <= 32))) { |
989 | // Check if AND instruction has constant as argument |
990 | bool isConstCase = SecondOperand.getOpcode() != ISD::AND; |
991 | if (SecondOperand.getOpcode() == ISD::AND) { |
992 | if (!(CN1 = dyn_cast<ConstantSDNode>(Val: SecondOperand->getOperand(Num: 1)))) |
993 | return SDValue(); |
994 | } else { |
995 | if (!(CN1 = dyn_cast<ConstantSDNode>(Val: N->getOperand(Num: 1)))) |
996 | return SDValue(); |
997 | } |
998 | // Don't generate INS if constant OR operand doesn't fit into bits |
999 | // cleared by constant AND operand. |
1000 | if (CN->getSExtValue() & CN1->getSExtValue()) |
1001 | return SDValue(); |
1002 | |
1003 | SDLoc DL(N); |
1004 | EVT ValTy = N->getOperand(Num: 0)->getValueType(ResNo: 0); |
1005 | SDValue Const1; |
1006 | SDValue SrlX; |
1007 | if (!isConstCase) { |
1008 | Const1 = DAG.getConstant(Val: SMPos0, DL, VT: MVT::i32); |
1009 | SrlX = DAG.getNode(Opcode: ISD::SRL, DL, VT: SecondOperand->getValueType(ResNo: 0), |
1010 | N1: SecondOperand, N2: Const1); |
1011 | } |
1012 | return DAG.getNode( |
1013 | Opcode: MipsISD::Ins, DL, VT: N->getValueType(ResNo: 0), |
1014 | N1: isConstCase |
1015 | ? DAG.getSignedConstant(Val: CN1->getSExtValue() >> SMPos0, DL, VT: ValTy) |
1016 | : SrlX, |
1017 | N2: DAG.getConstant(Val: SMPos0, DL, VT: MVT::i32), |
1018 | N3: DAG.getConstant(Val: ValTy.getSizeInBits() / 8 < 8 ? SMSize0 & 31 |
1019 | : SMSize0, |
1020 | DL, VT: MVT::i32), |
1021 | N4: FirstOperand->getOperand(Num: 0)); |
1022 | } |
1023 | return SDValue(); |
1024 | } |
1025 | } |
1026 | |
1027 | static SDValue performMADD_MSUBCombine(SDNode *ROOTNode, SelectionDAG &CurDAG, |
1028 | const MipsSubtarget &Subtarget) { |
1029 | // ROOTNode must have a multiplication as an operand for the match to be |
1030 | // successful. |
1031 | if (ROOTNode->getOperand(Num: 0).getOpcode() != ISD::MUL && |
1032 | ROOTNode->getOperand(Num: 1).getOpcode() != ISD::MUL) |
1033 | return SDValue(); |
1034 | |
1035 | // In the case where we have a multiplication as the left operand of |
1036 | // of a subtraction, we can't combine into a MipsISD::MSub node as the |
1037 | // the instruction definition of msub(u) places the multiplication on |
1038 | // on the right. |
1039 | if (ROOTNode->getOpcode() == ISD::SUB && |
1040 | ROOTNode->getOperand(Num: 0).getOpcode() == ISD::MUL) |
1041 | return SDValue(); |
1042 | |
1043 | // We don't handle vector types here. |
1044 | if (ROOTNode->getValueType(ResNo: 0).isVector()) |
1045 | return SDValue(); |
1046 | |
1047 | // For MIPS64, madd / msub instructions are inefficent to use with 64 bit |
1048 | // arithmetic. E.g. |
1049 | // (add (mul a b) c) => |
1050 | // let res = (madd (mthi (drotr c 32))x(mtlo c) a b) in |
1051 | // MIPS64: (or (dsll (mfhi res) 32) (dsrl (dsll (mflo res) 32) 32) |
1052 | // or |
1053 | // MIPS64R2: (dins (mflo res) (mfhi res) 32 32) |
1054 | // |
1055 | // The overhead of setting up the Hi/Lo registers and reassembling the |
1056 | // result makes this a dubious optimzation for MIPS64. The core of the |
1057 | // problem is that Hi/Lo contain the upper and lower 32 bits of the |
1058 | // operand and result. |
1059 | // |
1060 | // It requires a chain of 4 add/mul for MIPS64R2 to get better code |
1061 | // density than doing it naively, 5 for MIPS64. Additionally, using |
1062 | // madd/msub on MIPS64 requires the operands actually be 32 bit sign |
1063 | // extended operands, not true 64 bit values. |
1064 | // |
1065 | // FIXME: For the moment, disable this completely for MIPS64. |
1066 | if (Subtarget.hasMips64()) |
1067 | return SDValue(); |
1068 | |
1069 | SDValue Mult = ROOTNode->getOperand(Num: 0).getOpcode() == ISD::MUL |
1070 | ? ROOTNode->getOperand(Num: 0) |
1071 | : ROOTNode->getOperand(Num: 1); |
1072 | |
1073 | SDValue AddOperand = ROOTNode->getOperand(Num: 0).getOpcode() == ISD::MUL |
1074 | ? ROOTNode->getOperand(Num: 1) |
1075 | : ROOTNode->getOperand(Num: 0); |
1076 | |
1077 | // Transform this to a MADD only if the user of this node is the add. |
1078 | // If there are other users of the mul, this function returns here. |
1079 | if (!Mult.hasOneUse()) |
1080 | return SDValue(); |
1081 | |
1082 | // maddu and madd are unusual instructions in that on MIPS64 bits 63..31 |
1083 | // must be in canonical form, i.e. sign extended. For MIPS32, the operands |
1084 | // of the multiply must have 32 or more sign bits, otherwise we cannot |
1085 | // perform this optimization. We have to check this here as we're performing |
1086 | // this optimization pre-legalization. |
1087 | SDValue MultLHS = Mult->getOperand(Num: 0); |
1088 | SDValue MultRHS = Mult->getOperand(Num: 1); |
1089 | |
1090 | bool IsSigned = MultLHS->getOpcode() == ISD::SIGN_EXTEND && |
1091 | MultRHS->getOpcode() == ISD::SIGN_EXTEND; |
1092 | bool IsUnsigned = MultLHS->getOpcode() == ISD::ZERO_EXTEND && |
1093 | MultRHS->getOpcode() == ISD::ZERO_EXTEND; |
1094 | |
1095 | if (!IsSigned && !IsUnsigned) |
1096 | return SDValue(); |
1097 | |
1098 | // Initialize accumulator. |
1099 | SDLoc DL(ROOTNode); |
1100 | SDValue BottomHalf, TopHalf; |
1101 | std::tie(args&: BottomHalf, args&: TopHalf) = |
1102 | CurDAG.SplitScalar(N: AddOperand, DL, LoVT: MVT::i32, HiVT: MVT::i32); |
1103 | SDValue ACCIn = |
1104 | CurDAG.getNode(Opcode: MipsISD::MTLOHI, DL, VT: MVT::Untyped, N1: BottomHalf, N2: TopHalf); |
1105 | |
1106 | // Create MipsMAdd(u) / MipsMSub(u) node. |
1107 | bool IsAdd = ROOTNode->getOpcode() == ISD::ADD; |
1108 | unsigned Opcode = IsAdd ? (IsUnsigned ? MipsISD::MAddu : MipsISD::MAdd) |
1109 | : (IsUnsigned ? MipsISD::MSubu : MipsISD::MSub); |
1110 | SDValue MAddOps[3] = { |
1111 | CurDAG.getNode(Opcode: ISD::TRUNCATE, DL, VT: MVT::i32, Operand: Mult->getOperand(Num: 0)), |
1112 | CurDAG.getNode(Opcode: ISD::TRUNCATE, DL, VT: MVT::i32, Operand: Mult->getOperand(Num: 1)), ACCIn}; |
1113 | SDValue MAdd = CurDAG.getNode(Opcode, DL, VT: MVT::Untyped, Ops: MAddOps); |
1114 | |
1115 | SDValue ResLo = CurDAG.getNode(Opcode: MipsISD::MFLO, DL, VT: MVT::i32, Operand: MAdd); |
1116 | SDValue ResHi = CurDAG.getNode(Opcode: MipsISD::MFHI, DL, VT: MVT::i32, Operand: MAdd); |
1117 | SDValue Combined = |
1118 | CurDAG.getNode(Opcode: ISD::BUILD_PAIR, DL, VT: MVT::i64, N1: ResLo, N2: ResHi); |
1119 | return Combined; |
1120 | } |
1121 | |
1122 | static SDValue performSUBCombine(SDNode *N, SelectionDAG &DAG, |
1123 | TargetLowering::DAGCombinerInfo &DCI, |
1124 | const MipsSubtarget &Subtarget) { |
1125 | // (sub v0 (mul v1, v2)) => (msub v1, v2, v0) |
1126 | if (DCI.isBeforeLegalizeOps()) { |
1127 | if (Subtarget.hasMips32() && !Subtarget.hasMips32r6() && |
1128 | !Subtarget.inMips16Mode() && N->getValueType(ResNo: 0) == MVT::i64) |
1129 | return performMADD_MSUBCombine(ROOTNode: N, CurDAG&: DAG, Subtarget); |
1130 | |
1131 | return SDValue(); |
1132 | } |
1133 | |
1134 | return SDValue(); |
1135 | } |
1136 | |
1137 | static SDValue performADDCombine(SDNode *N, SelectionDAG &DAG, |
1138 | TargetLowering::DAGCombinerInfo &DCI, |
1139 | const MipsSubtarget &Subtarget) { |
1140 | // (add v0 (mul v1, v2)) => (madd v1, v2, v0) |
1141 | if (DCI.isBeforeLegalizeOps()) { |
1142 | if (Subtarget.hasMips32() && !Subtarget.hasMips32r6() && |
1143 | !Subtarget.inMips16Mode() && N->getValueType(ResNo: 0) == MVT::i64) |
1144 | return performMADD_MSUBCombine(ROOTNode: N, CurDAG&: DAG, Subtarget); |
1145 | |
1146 | return SDValue(); |
1147 | } |
1148 | |
1149 | // When loading from a jump table, push the Lo node to the position that |
1150 | // allows folding it into a load immediate. |
1151 | // (add v0, (add v1, abs_lo(tjt))) => (add (add v0, v1), abs_lo(tjt)) |
1152 | // (add (add abs_lo(tjt), v1), v0) => (add (add v0, v1), abs_lo(tjt)) |
1153 | SDValue InnerAdd = N->getOperand(Num: 1); |
1154 | SDValue Index = N->getOperand(Num: 0); |
1155 | if (InnerAdd.getOpcode() != ISD::ADD) |
1156 | std::swap(a&: InnerAdd, b&: Index); |
1157 | if (InnerAdd.getOpcode() != ISD::ADD) |
1158 | return SDValue(); |
1159 | |
1160 | SDValue Lo = InnerAdd.getOperand(i: 0); |
1161 | SDValue Other = InnerAdd.getOperand(i: 1); |
1162 | if (Lo.getOpcode() != MipsISD::Lo) |
1163 | std::swap(a&: Lo, b&: Other); |
1164 | |
1165 | if ((Lo.getOpcode() != MipsISD::Lo) || |
1166 | (Lo.getOperand(i: 0).getOpcode() != ISD::TargetJumpTable)) |
1167 | return SDValue(); |
1168 | |
1169 | EVT ValTy = N->getValueType(ResNo: 0); |
1170 | SDLoc DL(N); |
1171 | |
1172 | SDValue Add1 = DAG.getNode(Opcode: ISD::ADD, DL, VT: ValTy, N1: Index, N2: Other); |
1173 | return DAG.getNode(Opcode: ISD::ADD, DL, VT: ValTy, N1: Add1, N2: Lo); |
1174 | } |
1175 | |
1176 | static SDValue performSHLCombine(SDNode *N, SelectionDAG &DAG, |
1177 | TargetLowering::DAGCombinerInfo &DCI, |
1178 | const MipsSubtarget &Subtarget) { |
1179 | // Pattern match CINS. |
1180 | // $dst = shl (and $src , imm), pos |
1181 | // => cins $dst, $src, pos, size |
1182 | |
1183 | if (DCI.isBeforeLegalizeOps() || !Subtarget.hasCnMips()) |
1184 | return SDValue(); |
1185 | |
1186 | SDValue FirstOperand = N->getOperand(Num: 0); |
1187 | unsigned FirstOperandOpc = FirstOperand.getOpcode(); |
1188 | SDValue SecondOperand = N->getOperand(Num: 1); |
1189 | EVT ValTy = N->getValueType(ResNo: 0); |
1190 | SDLoc DL(N); |
1191 | |
1192 | uint64_t Pos = 0; |
1193 | unsigned SMPos, SMSize; |
1194 | ConstantSDNode *CN; |
1195 | SDValue NewOperand; |
1196 | |
1197 | // The second operand of the shift must be an immediate. |
1198 | if (!(CN = dyn_cast<ConstantSDNode>(Val&: SecondOperand))) |
1199 | return SDValue(); |
1200 | |
1201 | Pos = CN->getZExtValue(); |
1202 | |
1203 | if (Pos >= ValTy.getSizeInBits()) |
1204 | return SDValue(); |
1205 | |
1206 | if (FirstOperandOpc != ISD::AND) |
1207 | return SDValue(); |
1208 | |
1209 | // AND's second operand must be a shifted mask. |
1210 | if (!(CN = dyn_cast<ConstantSDNode>(Val: FirstOperand.getOperand(i: 1))) || |
1211 | !isShiftedMask_64(Value: CN->getZExtValue(), MaskIdx&: SMPos, MaskLen&: SMSize)) |
1212 | return SDValue(); |
1213 | |
1214 | // Return if the shifted mask does not start at bit 0 or the sum of its size |
1215 | // and Pos exceeds the word's size. |
1216 | if (SMPos != 0 || SMSize > 32 || Pos + SMSize > ValTy.getSizeInBits()) |
1217 | return SDValue(); |
1218 | |
1219 | NewOperand = FirstOperand.getOperand(i: 0); |
1220 | // SMSize is 'location' (position) in this case, not size. |
1221 | SMSize--; |
1222 | |
1223 | return DAG.getNode(Opcode: MipsISD::CIns, DL, VT: ValTy, N1: NewOperand, |
1224 | N2: DAG.getConstant(Val: Pos, DL, VT: MVT::i32), |
1225 | N3: DAG.getConstant(Val: SMSize, DL, VT: MVT::i32)); |
1226 | } |
1227 | |
1228 | static SDValue performSignExtendCombine(SDNode *N, SelectionDAG &DAG, |
1229 | TargetLowering::DAGCombinerInfo &DCI, |
1230 | const MipsSubtarget &Subtarget) { |
1231 | if (DCI.Level != AfterLegalizeDAG || !Subtarget.isGP64bit()) { |
1232 | return SDValue(); |
1233 | } |
1234 | |
1235 | SDValue N0 = N->getOperand(Num: 0); |
1236 | EVT VT = N->getValueType(ResNo: 0); |
1237 | |
1238 | // Pattern match XOR. |
1239 | // $dst = sign_extend (xor (trunc $src, i32), imm) |
1240 | // => $dst = xor (signext_inreg $src, i32), imm |
1241 | if (N0.getOpcode() == ISD::XOR && |
1242 | N0.getOperand(i: 0).getOpcode() == ISD::TRUNCATE && |
1243 | N0.getOperand(i: 1).getOpcode() == ISD::Constant) { |
1244 | SDValue TruncateSource = N0.getOperand(i: 0).getOperand(i: 0); |
1245 | auto *ConstantOperand = dyn_cast<ConstantSDNode>(Val: N0->getOperand(Num: 1)); |
1246 | |
1247 | SDValue FirstOperand = |
1248 | DAG.getNode(Opcode: ISD::SIGN_EXTEND_INREG, DL: SDLoc(N0), VT, N1: TruncateSource, |
1249 | N2: DAG.getValueType(N0.getOperand(i: 0).getValueType())); |
1250 | |
1251 | int64_t ConstImm = ConstantOperand->getSExtValue(); |
1252 | return DAG.getNode(Opcode: ISD::XOR, DL: SDLoc(N0), VT, N1: FirstOperand, |
1253 | N2: DAG.getConstant(Val: ConstImm, DL: SDLoc(N0), VT)); |
1254 | } |
1255 | |
1256 | return SDValue(); |
1257 | } |
1258 | |
1259 | SDValue MipsTargetLowering::PerformDAGCombine(SDNode *N, DAGCombinerInfo &DCI) |
1260 | const { |
1261 | SelectionDAG &DAG = DCI.DAG; |
1262 | unsigned Opc = N->getOpcode(); |
1263 | |
1264 | switch (Opc) { |
1265 | default: break; |
1266 | case ISD::SDIVREM: |
1267 | case ISD::UDIVREM: |
1268 | return performDivRemCombine(N, DAG, DCI, Subtarget); |
1269 | case ISD::SELECT: |
1270 | return performSELECTCombine(N, DAG, DCI, Subtarget); |
1271 | case MipsISD::CMovFP_F: |
1272 | case MipsISD::CMovFP_T: |
1273 | return performCMovFPCombine(N, DAG, DCI, Subtarget); |
1274 | case ISD::AND: |
1275 | return performANDCombine(N, DAG, DCI, Subtarget); |
1276 | case ISD::OR: |
1277 | return performORCombine(N, DAG, DCI, Subtarget); |
1278 | case ISD::ADD: |
1279 | return performADDCombine(N, DAG, DCI, Subtarget); |
1280 | case ISD::SHL: |
1281 | return performSHLCombine(N, DAG, DCI, Subtarget); |
1282 | case ISD::SUB: |
1283 | return performSUBCombine(N, DAG, DCI, Subtarget); |
1284 | case ISD::SIGN_EXTEND: |
1285 | return performSignExtendCombine(N, DAG, DCI, Subtarget); |
1286 | } |
1287 | |
1288 | return SDValue(); |
1289 | } |
1290 | |
1291 | bool MipsTargetLowering::isCheapToSpeculateCttz(Type *Ty) const { |
1292 | return Subtarget.hasMips32(); |
1293 | } |
1294 | |
1295 | bool MipsTargetLowering::isCheapToSpeculateCtlz(Type *Ty) const { |
1296 | return Subtarget.hasMips32(); |
1297 | } |
1298 | |
1299 | bool MipsTargetLowering::hasBitTest(SDValue X, SDValue Y) const { |
1300 | // We can use ANDI+SLTIU as a bit test. Y contains the bit position. |
1301 | // For MIPSR2 or later, we may be able to use the `ext` instruction or its' |
1302 | // double-word variants. |
1303 | if (auto *C = dyn_cast<ConstantSDNode>(Val&: Y)) |
1304 | return C->getAPIntValue().ule(RHS: 15); |
1305 | |
1306 | return false; |
1307 | } |
1308 | |
1309 | bool MipsTargetLowering::shouldFoldConstantShiftPairToMask( |
1310 | const SDNode *N, CombineLevel Level) const { |
1311 | assert(((N->getOpcode() == ISD::SHL && |
1312 | N->getOperand(0).getOpcode() == ISD::SRL) || |
1313 | (N->getOpcode() == ISD::SRL && |
1314 | N->getOperand(0).getOpcode() == ISD::SHL)) && |
1315 | "Expected shift-shift mask" ); |
1316 | |
1317 | if (N->getOperand(Num: 0).getValueType().isVector()) |
1318 | return false; |
1319 | return true; |
1320 | } |
1321 | |
1322 | void |
1323 | MipsTargetLowering::ReplaceNodeResults(SDNode *N, |
1324 | SmallVectorImpl<SDValue> &Results, |
1325 | SelectionDAG &DAG) const { |
1326 | return LowerOperationWrapper(N, Results, DAG); |
1327 | } |
1328 | |
1329 | SDValue MipsTargetLowering:: |
1330 | LowerOperation(SDValue Op, SelectionDAG &DAG) const |
1331 | { |
1332 | switch (Op.getOpcode()) |
1333 | { |
1334 | case ISD::BRCOND: return lowerBRCOND(Op, DAG); |
1335 | case ISD::ConstantPool: return lowerConstantPool(Op, DAG); |
1336 | case ISD::GlobalAddress: return lowerGlobalAddress(Op, DAG); |
1337 | case ISD::BlockAddress: return lowerBlockAddress(Op, DAG); |
1338 | case ISD::GlobalTLSAddress: return lowerGlobalTLSAddress(Op, DAG); |
1339 | case ISD::JumpTable: return lowerJumpTable(Op, DAG); |
1340 | case ISD::SELECT: return lowerSELECT(Op, DAG); |
1341 | case ISD::SETCC: return lowerSETCC(Op, DAG); |
1342 | case ISD::VASTART: return lowerVASTART(Op, DAG); |
1343 | case ISD::VAARG: return lowerVAARG(Op, DAG); |
1344 | case ISD::FCOPYSIGN: return lowerFCOPYSIGN(Op, DAG); |
1345 | case ISD::FABS: return lowerFABS(Op, DAG); |
1346 | case ISD::FCANONICALIZE: |
1347 | return lowerFCANONICALIZE(Op, DAG); |
1348 | case ISD::FRAMEADDR: return lowerFRAMEADDR(Op, DAG); |
1349 | case ISD::RETURNADDR: return lowerRETURNADDR(Op, DAG); |
1350 | case ISD::EH_RETURN: return lowerEH_RETURN(Op, DAG); |
1351 | case ISD::ATOMIC_FENCE: return lowerATOMIC_FENCE(Op, DAG); |
1352 | case ISD::SHL_PARTS: return lowerShiftLeftParts(Op, DAG); |
1353 | case ISD::SRA_PARTS: return lowerShiftRightParts(Op, DAG, IsSRA: true); |
1354 | case ISD::SRL_PARTS: return lowerShiftRightParts(Op, DAG, IsSRA: false); |
1355 | case ISD::LOAD: return lowerLOAD(Op, DAG); |
1356 | case ISD::STORE: return lowerSTORE(Op, DAG); |
1357 | case ISD::EH_DWARF_CFA: return lowerEH_DWARF_CFA(Op, DAG); |
1358 | case ISD::FP_TO_SINT: return lowerFP_TO_SINT(Op, DAG); |
1359 | case ISD::READCYCLECOUNTER: |
1360 | return lowerREADCYCLECOUNTER(Op, DAG); |
1361 | case ISD::ConstantFP: |
1362 | return lowerConstantFP(Op, DAG); |
1363 | } |
1364 | return SDValue(); |
1365 | } |
1366 | |
1367 | //===----------------------------------------------------------------------===// |
1368 | // Lower helper functions |
1369 | //===----------------------------------------------------------------------===// |
1370 | |
1371 | // addLiveIn - This helper function adds the specified physical register to the |
1372 | // MachineFunction as a live in value. It also creates a corresponding |
1373 | // virtual register for it. |
1374 | static unsigned |
1375 | addLiveIn(MachineFunction &MF, unsigned PReg, const TargetRegisterClass *RC) |
1376 | { |
1377 | Register VReg = MF.getRegInfo().createVirtualRegister(RegClass: RC); |
1378 | MF.getRegInfo().addLiveIn(Reg: PReg, vreg: VReg); |
1379 | return VReg; |
1380 | } |
1381 | |
1382 | static MachineBasicBlock *insertDivByZeroTrap(MachineInstr &MI, |
1383 | MachineBasicBlock &MBB, |
1384 | const TargetInstrInfo &TII, |
1385 | bool Is64Bit, bool IsMicroMips) { |
1386 | if (NoZeroDivCheck) |
1387 | return &MBB; |
1388 | |
1389 | // Insert instruction "teq $divisor_reg, $zero, 7". |
1390 | MachineBasicBlock::iterator I(MI); |
1391 | MachineInstrBuilder MIB; |
1392 | MachineOperand &Divisor = MI.getOperand(i: 2); |
1393 | MIB = BuildMI(BB&: MBB, I: std::next(x: I), MIMD: MI.getDebugLoc(), |
1394 | MCID: TII.get(Opcode: IsMicroMips ? Mips::TEQ_MM : Mips::TEQ)) |
1395 | .addReg(RegNo: Divisor.getReg(), flags: getKillRegState(B: Divisor.isKill())) |
1396 | .addReg(RegNo: Mips::ZERO) |
1397 | .addImm(Val: 7); |
1398 | |
1399 | // Use the 32-bit sub-register if this is a 64-bit division. |
1400 | if (Is64Bit) |
1401 | MIB->getOperand(i: 0).setSubReg(Mips::sub_32); |
1402 | |
1403 | // Clear Divisor's kill flag. |
1404 | Divisor.setIsKill(false); |
1405 | |
1406 | // We would normally delete the original instruction here but in this case |
1407 | // we only needed to inject an additional instruction rather than replace it. |
1408 | |
1409 | return &MBB; |
1410 | } |
1411 | |
1412 | MachineBasicBlock * |
1413 | MipsTargetLowering::EmitInstrWithCustomInserter(MachineInstr &MI, |
1414 | MachineBasicBlock *BB) const { |
1415 | switch (MI.getOpcode()) { |
1416 | default: |
1417 | llvm_unreachable("Unexpected instr type to insert" ); |
1418 | case Mips::ATOMIC_LOAD_ADD_I8: |
1419 | return emitAtomicBinaryPartword(MI, BB, Size: 1); |
1420 | case Mips::ATOMIC_LOAD_ADD_I16: |
1421 | return emitAtomicBinaryPartword(MI, BB, Size: 2); |
1422 | case Mips::ATOMIC_LOAD_ADD_I32: |
1423 | return emitAtomicBinary(MI, BB); |
1424 | case Mips::ATOMIC_LOAD_ADD_I64: |
1425 | return emitAtomicBinary(MI, BB); |
1426 | |
1427 | case Mips::ATOMIC_LOAD_AND_I8: |
1428 | return emitAtomicBinaryPartword(MI, BB, Size: 1); |
1429 | case Mips::ATOMIC_LOAD_AND_I16: |
1430 | return emitAtomicBinaryPartword(MI, BB, Size: 2); |
1431 | case Mips::ATOMIC_LOAD_AND_I32: |
1432 | return emitAtomicBinary(MI, BB); |
1433 | case Mips::ATOMIC_LOAD_AND_I64: |
1434 | return emitAtomicBinary(MI, BB); |
1435 | |
1436 | case Mips::ATOMIC_LOAD_OR_I8: |
1437 | return emitAtomicBinaryPartword(MI, BB, Size: 1); |
1438 | case Mips::ATOMIC_LOAD_OR_I16: |
1439 | return emitAtomicBinaryPartword(MI, BB, Size: 2); |
1440 | case Mips::ATOMIC_LOAD_OR_I32: |
1441 | return emitAtomicBinary(MI, BB); |
1442 | case Mips::ATOMIC_LOAD_OR_I64: |
1443 | return emitAtomicBinary(MI, BB); |
1444 | |
1445 | case Mips::ATOMIC_LOAD_XOR_I8: |
1446 | return emitAtomicBinaryPartword(MI, BB, Size: 1); |
1447 | case Mips::ATOMIC_LOAD_XOR_I16: |
1448 | return emitAtomicBinaryPartword(MI, BB, Size: 2); |
1449 | case Mips::ATOMIC_LOAD_XOR_I32: |
1450 | return emitAtomicBinary(MI, BB); |
1451 | case Mips::ATOMIC_LOAD_XOR_I64: |
1452 | return emitAtomicBinary(MI, BB); |
1453 | |
1454 | case Mips::ATOMIC_LOAD_NAND_I8: |
1455 | return emitAtomicBinaryPartword(MI, BB, Size: 1); |
1456 | case Mips::ATOMIC_LOAD_NAND_I16: |
1457 | return emitAtomicBinaryPartword(MI, BB, Size: 2); |
1458 | case Mips::ATOMIC_LOAD_NAND_I32: |
1459 | return emitAtomicBinary(MI, BB); |
1460 | case Mips::ATOMIC_LOAD_NAND_I64: |
1461 | return emitAtomicBinary(MI, BB); |
1462 | |
1463 | case Mips::ATOMIC_LOAD_SUB_I8: |
1464 | return emitAtomicBinaryPartword(MI, BB, Size: 1); |
1465 | case Mips::ATOMIC_LOAD_SUB_I16: |
1466 | return emitAtomicBinaryPartword(MI, BB, Size: 2); |
1467 | case Mips::ATOMIC_LOAD_SUB_I32: |
1468 | return emitAtomicBinary(MI, BB); |
1469 | case Mips::ATOMIC_LOAD_SUB_I64: |
1470 | return emitAtomicBinary(MI, BB); |
1471 | |
1472 | case Mips::ATOMIC_SWAP_I8: |
1473 | return emitAtomicBinaryPartword(MI, BB, Size: 1); |
1474 | case Mips::ATOMIC_SWAP_I16: |
1475 | return emitAtomicBinaryPartword(MI, BB, Size: 2); |
1476 | case Mips::ATOMIC_SWAP_I32: |
1477 | return emitAtomicBinary(MI, BB); |
1478 | case Mips::ATOMIC_SWAP_I64: |
1479 | return emitAtomicBinary(MI, BB); |
1480 | |
1481 | case Mips::ATOMIC_CMP_SWAP_I8: |
1482 | return emitAtomicCmpSwapPartword(MI, BB, Size: 1); |
1483 | case Mips::ATOMIC_CMP_SWAP_I16: |
1484 | return emitAtomicCmpSwapPartword(MI, BB, Size: 2); |
1485 | case Mips::ATOMIC_CMP_SWAP_I32: |
1486 | return emitAtomicCmpSwap(MI, BB); |
1487 | case Mips::ATOMIC_CMP_SWAP_I64: |
1488 | return emitAtomicCmpSwap(MI, BB); |
1489 | |
1490 | case Mips::ATOMIC_LOAD_MIN_I8: |
1491 | return emitAtomicBinaryPartword(MI, BB, Size: 1); |
1492 | case Mips::ATOMIC_LOAD_MIN_I16: |
1493 | return emitAtomicBinaryPartword(MI, BB, Size: 2); |
1494 | case Mips::ATOMIC_LOAD_MIN_I32: |
1495 | return emitAtomicBinary(MI, BB); |
1496 | case Mips::ATOMIC_LOAD_MIN_I64: |
1497 | return emitAtomicBinary(MI, BB); |
1498 | |
1499 | case Mips::ATOMIC_LOAD_MAX_I8: |
1500 | return emitAtomicBinaryPartword(MI, BB, Size: 1); |
1501 | case Mips::ATOMIC_LOAD_MAX_I16: |
1502 | return emitAtomicBinaryPartword(MI, BB, Size: 2); |
1503 | case Mips::ATOMIC_LOAD_MAX_I32: |
1504 | return emitAtomicBinary(MI, BB); |
1505 | case Mips::ATOMIC_LOAD_MAX_I64: |
1506 | return emitAtomicBinary(MI, BB); |
1507 | |
1508 | case Mips::ATOMIC_LOAD_UMIN_I8: |
1509 | return emitAtomicBinaryPartword(MI, BB, Size: 1); |
1510 | case Mips::ATOMIC_LOAD_UMIN_I16: |
1511 | return emitAtomicBinaryPartword(MI, BB, Size: 2); |
1512 | case Mips::ATOMIC_LOAD_UMIN_I32: |
1513 | return emitAtomicBinary(MI, BB); |
1514 | case Mips::ATOMIC_LOAD_UMIN_I64: |
1515 | return emitAtomicBinary(MI, BB); |
1516 | |
1517 | case Mips::ATOMIC_LOAD_UMAX_I8: |
1518 | return emitAtomicBinaryPartword(MI, BB, Size: 1); |
1519 | case Mips::ATOMIC_LOAD_UMAX_I16: |
1520 | return emitAtomicBinaryPartword(MI, BB, Size: 2); |
1521 | case Mips::ATOMIC_LOAD_UMAX_I32: |
1522 | return emitAtomicBinary(MI, BB); |
1523 | case Mips::ATOMIC_LOAD_UMAX_I64: |
1524 | return emitAtomicBinary(MI, BB); |
1525 | |
1526 | case Mips::PseudoSDIV: |
1527 | case Mips::PseudoUDIV: |
1528 | case Mips::DIV: |
1529 | case Mips::DIVU: |
1530 | case Mips::MOD: |
1531 | case Mips::MODU: |
1532 | return insertDivByZeroTrap(MI, MBB&: *BB, TII: *Subtarget.getInstrInfo(), Is64Bit: false, |
1533 | IsMicroMips: false); |
1534 | case Mips::SDIV_MM_Pseudo: |
1535 | case Mips::UDIV_MM_Pseudo: |
1536 | case Mips::SDIV_MM: |
1537 | case Mips::UDIV_MM: |
1538 | case Mips::DIV_MMR6: |
1539 | case Mips::DIVU_MMR6: |
1540 | case Mips::MOD_MMR6: |
1541 | case Mips::MODU_MMR6: |
1542 | return insertDivByZeroTrap(MI, MBB&: *BB, TII: *Subtarget.getInstrInfo(), Is64Bit: false, IsMicroMips: true); |
1543 | case Mips::PseudoDSDIV: |
1544 | case Mips::PseudoDUDIV: |
1545 | case Mips::DDIV: |
1546 | case Mips::DDIVU: |
1547 | case Mips::DMOD: |
1548 | case Mips::DMODU: |
1549 | return insertDivByZeroTrap(MI, MBB&: *BB, TII: *Subtarget.getInstrInfo(), Is64Bit: true, IsMicroMips: false); |
1550 | |
1551 | case Mips::PseudoSELECT_I: |
1552 | case Mips::PseudoSELECT_I64: |
1553 | case Mips::PseudoSELECT_S: |
1554 | case Mips::PseudoSELECT_D32: |
1555 | case Mips::PseudoSELECT_D64: |
1556 | return emitPseudoSELECT(MI, BB, isFPCmp: false, Opc: Mips::BNE); |
1557 | case Mips::PseudoSELECTFP_F_I: |
1558 | case Mips::PseudoSELECTFP_F_I64: |
1559 | case Mips::PseudoSELECTFP_F_S: |
1560 | case Mips::PseudoSELECTFP_F_D32: |
1561 | case Mips::PseudoSELECTFP_F_D64: |
1562 | return emitPseudoSELECT(MI, BB, isFPCmp: true, Opc: Mips::BC1F); |
1563 | case Mips::PseudoSELECTFP_T_I: |
1564 | case Mips::PseudoSELECTFP_T_I64: |
1565 | case Mips::PseudoSELECTFP_T_S: |
1566 | case Mips::PseudoSELECTFP_T_D32: |
1567 | case Mips::PseudoSELECTFP_T_D64: |
1568 | return emitPseudoSELECT(MI, BB, isFPCmp: true, Opc: Mips::BC1T); |
1569 | case Mips::PseudoD_SELECT_I: |
1570 | case Mips::PseudoD_SELECT_I64: |
1571 | return emitPseudoD_SELECT(MI, BB); |
1572 | case Mips::LDR_W: |
1573 | return emitLDR_W(MI, BB); |
1574 | case Mips::LDR_D: |
1575 | return emitLDR_D(MI, BB); |
1576 | case Mips::STR_W: |
1577 | return emitSTR_W(MI, BB); |
1578 | case Mips::STR_D: |
1579 | return emitSTR_D(MI, BB); |
1580 | } |
1581 | } |
1582 | |
1583 | // This function also handles Mips::ATOMIC_SWAP_I32 (when BinOpcode == 0), and |
1584 | // Mips::ATOMIC_LOAD_NAND_I32 (when Nand == true) |
1585 | MachineBasicBlock * |
1586 | MipsTargetLowering::emitAtomicBinary(MachineInstr &MI, |
1587 | MachineBasicBlock *BB) const { |
1588 | |
1589 | MachineFunction *MF = BB->getParent(); |
1590 | MachineRegisterInfo &RegInfo = MF->getRegInfo(); |
1591 | const TargetInstrInfo *TII = Subtarget.getInstrInfo(); |
1592 | DebugLoc DL = MI.getDebugLoc(); |
1593 | |
1594 | unsigned AtomicOp; |
1595 | bool NeedsAdditionalReg = false; |
1596 | switch (MI.getOpcode()) { |
1597 | case Mips::ATOMIC_LOAD_ADD_I32: |
1598 | AtomicOp = Mips::ATOMIC_LOAD_ADD_I32_POSTRA; |
1599 | break; |
1600 | case Mips::ATOMIC_LOAD_SUB_I32: |
1601 | AtomicOp = Mips::ATOMIC_LOAD_SUB_I32_POSTRA; |
1602 | break; |
1603 | case Mips::ATOMIC_LOAD_AND_I32: |
1604 | AtomicOp = Mips::ATOMIC_LOAD_AND_I32_POSTRA; |
1605 | break; |
1606 | case Mips::ATOMIC_LOAD_OR_I32: |
1607 | AtomicOp = Mips::ATOMIC_LOAD_OR_I32_POSTRA; |
1608 | break; |
1609 | case Mips::ATOMIC_LOAD_XOR_I32: |
1610 | AtomicOp = Mips::ATOMIC_LOAD_XOR_I32_POSTRA; |
1611 | break; |
1612 | case Mips::ATOMIC_LOAD_NAND_I32: |
1613 | AtomicOp = Mips::ATOMIC_LOAD_NAND_I32_POSTRA; |
1614 | break; |
1615 | case Mips::ATOMIC_SWAP_I32: |
1616 | AtomicOp = Mips::ATOMIC_SWAP_I32_POSTRA; |
1617 | break; |
1618 | case Mips::ATOMIC_LOAD_ADD_I64: |
1619 | AtomicOp = Mips::ATOMIC_LOAD_ADD_I64_POSTRA; |
1620 | break; |
1621 | case Mips::ATOMIC_LOAD_SUB_I64: |
1622 | AtomicOp = Mips::ATOMIC_LOAD_SUB_I64_POSTRA; |
1623 | break; |
1624 | case Mips::ATOMIC_LOAD_AND_I64: |
1625 | AtomicOp = Mips::ATOMIC_LOAD_AND_I64_POSTRA; |
1626 | break; |
1627 | case Mips::ATOMIC_LOAD_OR_I64: |
1628 | AtomicOp = Mips::ATOMIC_LOAD_OR_I64_POSTRA; |
1629 | break; |
1630 | case Mips::ATOMIC_LOAD_XOR_I64: |
1631 | AtomicOp = Mips::ATOMIC_LOAD_XOR_I64_POSTRA; |
1632 | break; |
1633 | case Mips::ATOMIC_LOAD_NAND_I64: |
1634 | AtomicOp = Mips::ATOMIC_LOAD_NAND_I64_POSTRA; |
1635 | break; |
1636 | case Mips::ATOMIC_SWAP_I64: |
1637 | AtomicOp = Mips::ATOMIC_SWAP_I64_POSTRA; |
1638 | break; |
1639 | case Mips::ATOMIC_LOAD_MIN_I32: |
1640 | AtomicOp = Mips::ATOMIC_LOAD_MIN_I32_POSTRA; |
1641 | NeedsAdditionalReg = true; |
1642 | break; |
1643 | case Mips::ATOMIC_LOAD_MAX_I32: |
1644 | AtomicOp = Mips::ATOMIC_LOAD_MAX_I32_POSTRA; |
1645 | NeedsAdditionalReg = true; |
1646 | break; |
1647 | case Mips::ATOMIC_LOAD_UMIN_I32: |
1648 | AtomicOp = Mips::ATOMIC_LOAD_UMIN_I32_POSTRA; |
1649 | NeedsAdditionalReg = true; |
1650 | break; |
1651 | case Mips::ATOMIC_LOAD_UMAX_I32: |
1652 | AtomicOp = Mips::ATOMIC_LOAD_UMAX_I32_POSTRA; |
1653 | NeedsAdditionalReg = true; |
1654 | break; |
1655 | case Mips::ATOMIC_LOAD_MIN_I64: |
1656 | AtomicOp = Mips::ATOMIC_LOAD_MIN_I64_POSTRA; |
1657 | NeedsAdditionalReg = true; |
1658 | break; |
1659 | case Mips::ATOMIC_LOAD_MAX_I64: |
1660 | AtomicOp = Mips::ATOMIC_LOAD_MAX_I64_POSTRA; |
1661 | NeedsAdditionalReg = true; |
1662 | break; |
1663 | case Mips::ATOMIC_LOAD_UMIN_I64: |
1664 | AtomicOp = Mips::ATOMIC_LOAD_UMIN_I64_POSTRA; |
1665 | NeedsAdditionalReg = true; |
1666 | break; |
1667 | case Mips::ATOMIC_LOAD_UMAX_I64: |
1668 | AtomicOp = Mips::ATOMIC_LOAD_UMAX_I64_POSTRA; |
1669 | NeedsAdditionalReg = true; |
1670 | break; |
1671 | default: |
1672 | llvm_unreachable("Unknown pseudo atomic for replacement!" ); |
1673 | } |
1674 | |
1675 | Register OldVal = MI.getOperand(i: 0).getReg(); |
1676 | Register Ptr = MI.getOperand(i: 1).getReg(); |
1677 | Register Incr = MI.getOperand(i: 2).getReg(); |
1678 | Register Scratch = RegInfo.createVirtualRegister(RegClass: RegInfo.getRegClass(Reg: OldVal)); |
1679 | |
1680 | MachineBasicBlock::iterator II(MI); |
1681 | |
1682 | // The scratch registers here with the EarlyClobber | Define | Implicit |
1683 | // flags is used to persuade the register allocator and the machine |
1684 | // verifier to accept the usage of this register. This has to be a real |
1685 | // register which has an UNDEF value but is dead after the instruction which |
1686 | // is unique among the registers chosen for the instruction. |
1687 | |
1688 | // The EarlyClobber flag has the semantic properties that the operand it is |
1689 | // attached to is clobbered before the rest of the inputs are read. Hence it |
1690 | // must be unique among the operands to the instruction. |
1691 | // The Define flag is needed to coerce the machine verifier that an Undef |
1692 | // value isn't a problem. |
1693 | // The Dead flag is needed as the value in scratch isn't used by any other |
1694 | // instruction. Kill isn't used as Dead is more precise. |
1695 | // The implicit flag is here due to the interaction between the other flags |
1696 | // and the machine verifier. |
1697 | |
1698 | // For correctness purpose, a new pseudo is introduced here. We need this |
1699 | // new pseudo, so that FastRegisterAllocator does not see an ll/sc sequence |
1700 | // that is spread over >1 basic blocks. A register allocator which |
1701 | // introduces (or any codegen infact) a store, can violate the expectations |
1702 | // of the hardware. |
1703 | // |
1704 | // An atomic read-modify-write sequence starts with a linked load |
1705 | // instruction and ends with a store conditional instruction. The atomic |
1706 | // read-modify-write sequence fails if any of the following conditions |
1707 | // occur between the execution of ll and sc: |
1708 | // * A coherent store is completed by another process or coherent I/O |
1709 | // module into the block of synchronizable physical memory containing |
1710 | // the word. The size and alignment of the block is |
1711 | // implementation-dependent. |
1712 | // * A coherent store is executed between an LL and SC sequence on the |
1713 | // same processor to the block of synchornizable physical memory |
1714 | // containing the word. |
1715 | // |
1716 | |
1717 | Register PtrCopy = RegInfo.createVirtualRegister(RegClass: RegInfo.getRegClass(Reg: Ptr)); |
1718 | Register IncrCopy = RegInfo.createVirtualRegister(RegClass: RegInfo.getRegClass(Reg: Incr)); |
1719 | |
1720 | BuildMI(BB&: *BB, I: II, MIMD: DL, MCID: TII->get(Opcode: Mips::COPY), DestReg: IncrCopy).addReg(RegNo: Incr); |
1721 | BuildMI(BB&: *BB, I: II, MIMD: DL, MCID: TII->get(Opcode: Mips::COPY), DestReg: PtrCopy).addReg(RegNo: Ptr); |
1722 | |
1723 | MachineInstrBuilder MIB = |
1724 | BuildMI(BB&: *BB, I: II, MIMD: DL, MCID: TII->get(Opcode: AtomicOp)) |
1725 | .addReg(RegNo: OldVal, flags: RegState::Define | RegState::EarlyClobber) |
1726 | .addReg(RegNo: PtrCopy) |
1727 | .addReg(RegNo: IncrCopy) |
1728 | .addReg(RegNo: Scratch, flags: RegState::Define | RegState::EarlyClobber | |
1729 | RegState::Implicit | RegState::Dead); |
1730 | if (NeedsAdditionalReg) { |
1731 | Register Scratch2 = |
1732 | RegInfo.createVirtualRegister(RegClass: RegInfo.getRegClass(Reg: OldVal)); |
1733 | MIB.addReg(RegNo: Scratch2, flags: RegState::Define | RegState::EarlyClobber | |
1734 | RegState::Implicit | RegState::Dead); |
1735 | } |
1736 | |
1737 | MI.eraseFromParent(); |
1738 | |
1739 | return BB; |
1740 | } |
1741 | |
1742 | MachineBasicBlock *MipsTargetLowering::emitSignExtendToI32InReg( |
1743 | MachineInstr &MI, MachineBasicBlock *BB, unsigned Size, unsigned DstReg, |
1744 | unsigned SrcReg) const { |
1745 | const TargetInstrInfo *TII = Subtarget.getInstrInfo(); |
1746 | const DebugLoc &DL = MI.getDebugLoc(); |
1747 | |
1748 | if (Subtarget.hasMips32r2() && Size == 1) { |
1749 | BuildMI(BB, MIMD: DL, MCID: TII->get(Opcode: Mips::SEB), DestReg: DstReg).addReg(RegNo: SrcReg); |
1750 | return BB; |
1751 | } |
1752 | |
1753 | if (Subtarget.hasMips32r2() && Size == 2) { |
1754 | BuildMI(BB, MIMD: DL, MCID: TII->get(Opcode: Mips::SEH), DestReg: DstReg).addReg(RegNo: SrcReg); |
1755 | return BB; |
1756 | } |
1757 | |
1758 | MachineFunction *MF = BB->getParent(); |
1759 | MachineRegisterInfo &RegInfo = MF->getRegInfo(); |
1760 | const TargetRegisterClass *RC = getRegClassFor(VT: MVT::i32); |
1761 | Register ScrReg = RegInfo.createVirtualRegister(RegClass: RC); |
1762 | |
1763 | assert(Size < 32); |
1764 | int64_t ShiftImm = 32 - (Size * 8); |
1765 | |
1766 | BuildMI(BB, MIMD: DL, MCID: TII->get(Opcode: Mips::SLL), DestReg: ScrReg).addReg(RegNo: SrcReg).addImm(Val: ShiftImm); |
1767 | BuildMI(BB, MIMD: DL, MCID: TII->get(Opcode: Mips::SRA), DestReg: DstReg).addReg(RegNo: ScrReg).addImm(Val: ShiftImm); |
1768 | |
1769 | return BB; |
1770 | } |
1771 | |
1772 | MachineBasicBlock *MipsTargetLowering::emitAtomicBinaryPartword( |
1773 | MachineInstr &MI, MachineBasicBlock *BB, unsigned Size) const { |
1774 | assert((Size == 1 || Size == 2) && |
1775 | "Unsupported size for EmitAtomicBinaryPartial." ); |
1776 | |
1777 | MachineFunction *MF = BB->getParent(); |
1778 | MachineRegisterInfo &RegInfo = MF->getRegInfo(); |
1779 | const TargetRegisterClass *RC = getRegClassFor(VT: MVT::i32); |
1780 | const bool ArePtrs64bit = ABI.ArePtrs64bit(); |
1781 | const TargetRegisterClass *RCp = |
1782 | getRegClassFor(VT: ArePtrs64bit ? MVT::i64 : MVT::i32); |
1783 | const TargetInstrInfo *TII = Subtarget.getInstrInfo(); |
1784 | DebugLoc DL = MI.getDebugLoc(); |
1785 | |
1786 | Register Dest = MI.getOperand(i: 0).getReg(); |
1787 | Register Ptr = MI.getOperand(i: 1).getReg(); |
1788 | Register Incr = MI.getOperand(i: 2).getReg(); |
1789 | |
1790 | Register AlignedAddr = RegInfo.createVirtualRegister(RegClass: RCp); |
1791 | Register ShiftAmt = RegInfo.createVirtualRegister(RegClass: RC); |
1792 | Register Mask = RegInfo.createVirtualRegister(RegClass: RC); |
1793 | Register Mask2 = RegInfo.createVirtualRegister(RegClass: RC); |
1794 | Register Incr2 = RegInfo.createVirtualRegister(RegClass: RC); |
1795 | Register MaskLSB2 = RegInfo.createVirtualRegister(RegClass: RCp); |
1796 | Register PtrLSB2 = RegInfo.createVirtualRegister(RegClass: RC); |
1797 | Register MaskUpper = RegInfo.createVirtualRegister(RegClass: RC); |
1798 | Register Scratch = RegInfo.createVirtualRegister(RegClass: RC); |
1799 | Register Scratch2 = RegInfo.createVirtualRegister(RegClass: RC); |
1800 | Register Scratch3 = RegInfo.createVirtualRegister(RegClass: RC); |
1801 | |
1802 | unsigned AtomicOp = 0; |
1803 | bool NeedsAdditionalReg = false; |
1804 | switch (MI.getOpcode()) { |
1805 | case Mips::ATOMIC_LOAD_NAND_I8: |
1806 | AtomicOp = Mips::ATOMIC_LOAD_NAND_I8_POSTRA; |
1807 | break; |
1808 | case Mips::ATOMIC_LOAD_NAND_I16: |
1809 | AtomicOp = Mips::ATOMIC_LOAD_NAND_I16_POSTRA; |
1810 | break; |
1811 | case Mips::ATOMIC_SWAP_I8: |
1812 | AtomicOp = Mips::ATOMIC_SWAP_I8_POSTRA; |
1813 | break; |
1814 | case Mips::ATOMIC_SWAP_I16: |
1815 | AtomicOp = Mips::ATOMIC_SWAP_I16_POSTRA; |
1816 | break; |
1817 | case Mips::ATOMIC_LOAD_ADD_I8: |
1818 | AtomicOp = Mips::ATOMIC_LOAD_ADD_I8_POSTRA; |
1819 | break; |
1820 | case Mips::ATOMIC_LOAD_ADD_I16: |
1821 | AtomicOp = Mips::ATOMIC_LOAD_ADD_I16_POSTRA; |
1822 | break; |
1823 | case Mips::ATOMIC_LOAD_SUB_I8: |
1824 | AtomicOp = Mips::ATOMIC_LOAD_SUB_I8_POSTRA; |
1825 | break; |
1826 | case Mips::ATOMIC_LOAD_SUB_I16: |
1827 | AtomicOp = Mips::ATOMIC_LOAD_SUB_I16_POSTRA; |
1828 | break; |
1829 | case Mips::ATOMIC_LOAD_AND_I8: |
1830 | AtomicOp = Mips::ATOMIC_LOAD_AND_I8_POSTRA; |
1831 | break; |
1832 | case Mips::ATOMIC_LOAD_AND_I16: |
1833 | AtomicOp = Mips::ATOMIC_LOAD_AND_I16_POSTRA; |
1834 | break; |
1835 | case Mips::ATOMIC_LOAD_OR_I8: |
1836 | AtomicOp = Mips::ATOMIC_LOAD_OR_I8_POSTRA; |
1837 | break; |
1838 | case Mips::ATOMIC_LOAD_OR_I16: |
1839 | AtomicOp = Mips::ATOMIC_LOAD_OR_I16_POSTRA; |
1840 | break; |
1841 | case Mips::ATOMIC_LOAD_XOR_I8: |
1842 | AtomicOp = Mips::ATOMIC_LOAD_XOR_I8_POSTRA; |
1843 | break; |
1844 | case Mips::ATOMIC_LOAD_XOR_I16: |
1845 | AtomicOp = Mips::ATOMIC_LOAD_XOR_I16_POSTRA; |
1846 | break; |
1847 | case Mips::ATOMIC_LOAD_MIN_I8: |
1848 | AtomicOp = Mips::ATOMIC_LOAD_MIN_I8_POSTRA; |
1849 | NeedsAdditionalReg = true; |
1850 | break; |
1851 | case Mips::ATOMIC_LOAD_MIN_I16: |
1852 | AtomicOp = Mips::ATOMIC_LOAD_MIN_I16_POSTRA; |
1853 | NeedsAdditionalReg = true; |
1854 | break; |
1855 | case Mips::ATOMIC_LOAD_MAX_I8: |
1856 | AtomicOp = Mips::ATOMIC_LOAD_MAX_I8_POSTRA; |
1857 | NeedsAdditionalReg = true; |
1858 | break; |
1859 | case Mips::ATOMIC_LOAD_MAX_I16: |
1860 | AtomicOp = Mips::ATOMIC_LOAD_MAX_I16_POSTRA; |
1861 | NeedsAdditionalReg = true; |
1862 | break; |
1863 | case Mips::ATOMIC_LOAD_UMIN_I8: |
1864 | AtomicOp = Mips::ATOMIC_LOAD_UMIN_I8_POSTRA; |
1865 | NeedsAdditionalReg = true; |
1866 | break; |
1867 | case Mips::ATOMIC_LOAD_UMIN_I16: |
1868 | AtomicOp = Mips::ATOMIC_LOAD_UMIN_I16_POSTRA; |
1869 | NeedsAdditionalReg = true; |
1870 | break; |
1871 | case Mips::ATOMIC_LOAD_UMAX_I8: |
1872 | AtomicOp = Mips::ATOMIC_LOAD_UMAX_I8_POSTRA; |
1873 | NeedsAdditionalReg = true; |
1874 | break; |
1875 | case Mips::ATOMIC_LOAD_UMAX_I16: |
1876 | AtomicOp = Mips::ATOMIC_LOAD_UMAX_I16_POSTRA; |
1877 | NeedsAdditionalReg = true; |
1878 | break; |
1879 | default: |
1880 | llvm_unreachable("Unknown subword atomic pseudo for expansion!" ); |
1881 | } |
1882 | |
1883 | // insert new blocks after the current block |
1884 | const BasicBlock *LLVM_BB = BB->getBasicBlock(); |
1885 | MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(BB: LLVM_BB); |
1886 | MachineFunction::iterator It = ++BB->getIterator(); |
1887 | MF->insert(MBBI: It, MBB: exitMBB); |
1888 | |
1889 | // Transfer the remainder of BB and its successor edges to exitMBB. |
1890 | exitMBB->splice(Where: exitMBB->begin(), Other: BB, |
1891 | From: std::next(x: MachineBasicBlock::iterator(MI)), To: BB->end()); |
1892 | exitMBB->transferSuccessorsAndUpdatePHIs(FromMBB: BB); |
1893 | |
1894 | BB->addSuccessor(Succ: exitMBB, Prob: BranchProbability::getOne()); |
1895 | |
1896 | // thisMBB: |
1897 | // addiu masklsb2,$0,-4 # 0xfffffffc |
1898 | // and alignedaddr,ptr,masklsb2 |
1899 | // andi ptrlsb2,ptr,3 |
1900 | // sll shiftamt,ptrlsb2,3 |
1901 | // ori maskupper,$0,255 # 0xff |
1902 | // sll mask,maskupper,shiftamt |
1903 | // nor mask2,$0,mask |
1904 | // sll incr2,incr,shiftamt |
1905 | |
1906 | int64_t MaskImm = (Size == 1) ? 255 : 65535; |
1907 | BuildMI(BB, MIMD: DL, MCID: TII->get(Opcode: ABI.GetPtrAddiuOp()), DestReg: MaskLSB2) |
1908 | .addReg(RegNo: ABI.GetNullPtr()).addImm(Val: -4); |
1909 | BuildMI(BB, MIMD: DL, MCID: TII->get(Opcode: ABI.GetPtrAndOp()), DestReg: AlignedAddr) |
1910 | .addReg(RegNo: Ptr).addReg(RegNo: MaskLSB2); |
1911 | BuildMI(BB, MIMD: DL, MCID: TII->get(Opcode: Mips::ANDi), DestReg: PtrLSB2) |
1912 | .addReg(RegNo: Ptr, flags: 0, SubReg: ArePtrs64bit ? Mips::sub_32 : 0).addImm(Val: 3); |
1913 | if (Subtarget.isLittle()) { |
1914 | BuildMI(BB, MIMD: DL, MCID: TII->get(Opcode: Mips::SLL), DestReg: ShiftAmt).addReg(RegNo: PtrLSB2).addImm(Val: 3); |
1915 | } else { |
1916 | Register Off = RegInfo.createVirtualRegister(RegClass: RC); |
1917 | BuildMI(BB, MIMD: DL, MCID: TII->get(Opcode: Mips::XORi), DestReg: Off) |
1918 | .addReg(RegNo: PtrLSB2).addImm(Val: (Size == 1) ? 3 : 2); |
1919 | BuildMI(BB, MIMD: DL, MCID: TII->get(Opcode: Mips::SLL), DestReg: ShiftAmt).addReg(RegNo: Off).addImm(Val: 3); |
1920 | } |
1921 | BuildMI(BB, MIMD: DL, MCID: TII->get(Opcode: Mips::ORi), DestReg: MaskUpper) |
1922 | .addReg(RegNo: Mips::ZERO).addImm(Val: MaskImm); |
1923 | BuildMI(BB, MIMD: DL, MCID: TII->get(Opcode: Mips::SLLV), DestReg: Mask) |
1924 | .addReg(RegNo: MaskUpper).addReg(RegNo: ShiftAmt); |
1925 | BuildMI(BB, MIMD: DL, MCID: TII->get(Opcode: Mips::NOR), DestReg: Mask2).addReg(RegNo: Mips::ZERO).addReg(RegNo: Mask); |
1926 | BuildMI(BB, MIMD: DL, MCID: TII->get(Opcode: Mips::SLLV), DestReg: Incr2).addReg(RegNo: Incr).addReg(RegNo: ShiftAmt); |
1927 | |
1928 | |
1929 | // The purposes of the flags on the scratch registers is explained in |
1930 | // emitAtomicBinary. In summary, we need a scratch register which is going to |
1931 | // be undef, that is unique among registers chosen for the instruction. |
1932 | |
1933 | MachineInstrBuilder MIB = |
1934 | BuildMI(BB, MIMD: DL, MCID: TII->get(Opcode: AtomicOp)) |
1935 | .addReg(RegNo: Dest, flags: RegState::Define | RegState::EarlyClobber) |
1936 | .addReg(RegNo: AlignedAddr) |
1937 | .addReg(RegNo: Incr2) |
1938 | .addReg(RegNo: Mask) |
1939 | .addReg(RegNo: Mask2) |
1940 | .addReg(RegNo: ShiftAmt) |
1941 | .addReg(RegNo: Scratch, flags: RegState::EarlyClobber | RegState::Define | |
1942 | RegState::Dead | RegState::Implicit) |
1943 | .addReg(RegNo: Scratch2, flags: RegState::EarlyClobber | RegState::Define | |
1944 | RegState::Dead | RegState::Implicit) |
1945 | .addReg(RegNo: Scratch3, flags: RegState::EarlyClobber | RegState::Define | |
1946 | RegState::Dead | RegState::Implicit); |
1947 | if (NeedsAdditionalReg) { |
1948 | Register Scratch4 = RegInfo.createVirtualRegister(RegClass: RC); |
1949 | MIB.addReg(RegNo: Scratch4, flags: RegState::EarlyClobber | RegState::Define | |
1950 | RegState::Dead | RegState::Implicit); |
1951 | } |
1952 | |
1953 | MI.eraseFromParent(); // The instruction is gone now. |
1954 | |
1955 | return exitMBB; |
1956 | } |
1957 | |
1958 | // Lower atomic compare and swap to a pseudo instruction, taking care to |
1959 | // define a scratch register for the pseudo instruction's expansion. The |
1960 | // instruction is expanded after the register allocator as to prevent |
1961 | // the insertion of stores between the linked load and the store conditional. |
1962 | |
1963 | MachineBasicBlock * |
1964 | MipsTargetLowering::emitAtomicCmpSwap(MachineInstr &MI, |
1965 | MachineBasicBlock *BB) const { |
1966 | |
1967 | assert((MI.getOpcode() == Mips::ATOMIC_CMP_SWAP_I32 || |
1968 | MI.getOpcode() == Mips::ATOMIC_CMP_SWAP_I64) && |
1969 | "Unsupported atomic pseudo for EmitAtomicCmpSwap." ); |
1970 | |
1971 | const unsigned Size = MI.getOpcode() == Mips::ATOMIC_CMP_SWAP_I32 ? 4 : 8; |
1972 | |
1973 | MachineFunction *MF = BB->getParent(); |
1974 | MachineRegisterInfo &MRI = MF->getRegInfo(); |
1975 | const TargetRegisterClass *RC = getRegClassFor(VT: MVT::getIntegerVT(BitWidth: Size * 8)); |
1976 | const TargetInstrInfo *TII = Subtarget.getInstrInfo(); |
1977 | DebugLoc DL = MI.getDebugLoc(); |
1978 | |
1979 | unsigned AtomicOp = MI.getOpcode() == Mips::ATOMIC_CMP_SWAP_I32 |
1980 | ? Mips::ATOMIC_CMP_SWAP_I32_POSTRA |
1981 | : Mips::ATOMIC_CMP_SWAP_I64_POSTRA; |
1982 | Register Dest = MI.getOperand(i: 0).getReg(); |
1983 | Register Ptr = MI.getOperand(i: 1).getReg(); |
1984 | Register OldVal = MI.getOperand(i: 2).getReg(); |
1985 | Register NewVal = MI.getOperand(i: 3).getReg(); |
1986 | |
1987 | Register Scratch = MRI.createVirtualRegister(RegClass: RC); |
1988 | MachineBasicBlock::iterator II(MI); |
1989 | |
1990 | // We need to create copies of the various registers and kill them at the |
1991 | // atomic pseudo. If the copies are not made, when the atomic is expanded |
1992 | // after fast register allocation, the spills will end up outside of the |
1993 | // blocks that their values are defined in, causing livein errors. |
1994 | |
1995 | Register PtrCopy = MRI.createVirtualRegister(RegClass: MRI.getRegClass(Reg: Ptr)); |
1996 | Register OldValCopy = MRI.createVirtualRegister(RegClass: MRI.getRegClass(Reg: OldVal)); |
1997 | Register NewValCopy = MRI.createVirtualRegister(RegClass: MRI.getRegClass(Reg: NewVal)); |
1998 | |
1999 | BuildMI(BB&: *BB, I: II, MIMD: DL, MCID: TII->get(Opcode: Mips::COPY), DestReg: PtrCopy).addReg(RegNo: Ptr); |
2000 | BuildMI(BB&: *BB, I: II, MIMD: DL, MCID: TII->get(Opcode: Mips::COPY), DestReg: OldValCopy).addReg(RegNo: OldVal); |
2001 | BuildMI(BB&: *BB, I: II, MIMD: DL, MCID: TII->get(Opcode: Mips::COPY), DestReg: NewValCopy).addReg(RegNo: NewVal); |
2002 | |
2003 | // The purposes of the flags on the scratch registers is explained in |
2004 | // emitAtomicBinary. In summary, we need a scratch register which is going to |
2005 | // be undef, that is unique among registers chosen for the instruction. |
2006 | |
2007 | BuildMI(BB&: *BB, I: II, MIMD: DL, MCID: TII->get(Opcode: AtomicOp)) |
2008 | .addReg(RegNo: Dest, flags: RegState::Define | RegState::EarlyClobber) |
2009 | .addReg(RegNo: PtrCopy, flags: RegState::Kill) |
2010 | .addReg(RegNo: OldValCopy, flags: RegState::Kill) |
2011 | .addReg(RegNo: NewValCopy, flags: RegState::Kill) |
2012 | .addReg(RegNo: Scratch, flags: RegState::EarlyClobber | RegState::Define | |
2013 | RegState::Dead | RegState::Implicit); |
2014 | |
2015 | MI.eraseFromParent(); // The instruction is gone now. |
2016 | |
2017 | return BB; |
2018 | } |
2019 | |
2020 | MachineBasicBlock *MipsTargetLowering::emitAtomicCmpSwapPartword( |
2021 | MachineInstr &MI, MachineBasicBlock *BB, unsigned Size) const { |
2022 | assert((Size == 1 || Size == 2) && |
2023 | "Unsupported size for EmitAtomicCmpSwapPartial." ); |
2024 | |
2025 | MachineFunction *MF = BB->getParent(); |
2026 | MachineRegisterInfo &RegInfo = MF->getRegInfo(); |
2027 | const TargetRegisterClass *RC = getRegClassFor(VT: MVT::i32); |
2028 | const bool ArePtrs64bit = ABI.ArePtrs64bit(); |
2029 | const TargetRegisterClass *RCp = |
2030 | getRegClassFor(VT: ArePtrs64bit ? MVT::i64 : MVT::i32); |
2031 | const TargetInstrInfo *TII = Subtarget.getInstrInfo(); |
2032 | DebugLoc DL = MI.getDebugLoc(); |
2033 | |
2034 | Register Dest = MI.getOperand(i: 0).getReg(); |
2035 | Register Ptr = MI.getOperand(i: 1).getReg(); |
2036 | Register CmpVal = MI.getOperand(i: 2).getReg(); |
2037 | Register NewVal = MI.getOperand(i: 3).getReg(); |
2038 | |
2039 | Register AlignedAddr = RegInfo.createVirtualRegister(RegClass: RCp); |
2040 | Register ShiftAmt = RegInfo.createVirtualRegister(RegClass: RC); |
2041 | Register Mask = RegInfo.createVirtualRegister(RegClass: RC); |
2042 | Register Mask2 = RegInfo.createVirtualRegister(RegClass: RC); |
2043 | Register ShiftedCmpVal = RegInfo.createVirtualRegister(RegClass: RC); |
2044 | Register ShiftedNewVal = RegInfo.createVirtualRegister(RegClass: RC); |
2045 | Register MaskLSB2 = RegInfo.createVirtualRegister(RegClass: RCp); |
2046 | Register PtrLSB2 = RegInfo.createVirtualRegister(RegClass: RC); |
2047 | Register MaskUpper = RegInfo.createVirtualRegister(RegClass: RC); |
2048 | Register MaskedCmpVal = RegInfo.createVirtualRegister(RegClass: RC); |
2049 | Register MaskedNewVal = RegInfo.createVirtualRegister(RegClass: RC); |
2050 | unsigned AtomicOp = MI.getOpcode() == Mips::ATOMIC_CMP_SWAP_I8 |
2051 | ? Mips::ATOMIC_CMP_SWAP_I8_POSTRA |
2052 | : Mips::ATOMIC_CMP_SWAP_I16_POSTRA; |
2053 | |
2054 | // The scratch registers here with the EarlyClobber | Define | Dead | Implicit |
2055 | // flags are used to coerce the register allocator and the machine verifier to |
2056 | // accept the usage of these registers. |
2057 | // The EarlyClobber flag has the semantic properties that the operand it is |
2058 | // attached to is clobbered before the rest of the inputs are read. Hence it |
2059 | // must be unique among the operands to the instruction. |
2060 | // The Define flag is needed to coerce the machine verifier that an Undef |
2061 | // value isn't a problem. |
2062 | // The Dead flag is needed as the value in scratch isn't used by any other |
2063 | // instruction. Kill isn't used as Dead is more precise. |
2064 | Register Scratch = RegInfo.createVirtualRegister(RegClass: RC); |
2065 | Register Scratch2 = RegInfo.createVirtualRegister(RegClass: RC); |
2066 | |
2067 | // insert new blocks after the current block |
2068 | const BasicBlock *LLVM_BB = BB->getBasicBlock(); |
2069 | MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(BB: LLVM_BB); |
2070 | MachineFunction::iterator It = ++BB->getIterator(); |
2071 | MF->insert(MBBI: It, MBB: exitMBB); |
2072 | |
2073 | // Transfer the remainder of BB and its successor edges to exitMBB. |
2074 | exitMBB->splice(Where: exitMBB->begin(), Other: BB, |
2075 | From: std::next(x: MachineBasicBlock::iterator(MI)), To: BB->end()); |
2076 | exitMBB->transferSuccessorsAndUpdatePHIs(FromMBB: BB); |
2077 | |
2078 | BB->addSuccessor(Succ: exitMBB, Prob: BranchProbability::getOne()); |
2079 | |
2080 | // thisMBB: |
2081 | // addiu masklsb2,$0,-4 # 0xfffffffc |
2082 | // and alignedaddr,ptr,masklsb2 |
2083 | // andi ptrlsb2,ptr,3 |
2084 | // xori ptrlsb2,ptrlsb2,3 # Only for BE |
2085 | // sll shiftamt,ptrlsb2,3 |
2086 | // ori maskupper,$0,255 # 0xff |
2087 | // sll mask,maskupper,shiftamt |
2088 | // nor mask2,$0,mask |
2089 | // andi maskedcmpval,cmpval,255 |
2090 | // sll shiftedcmpval,maskedcmpval,shiftamt |
2091 | // andi maskednewval,newval,255 |
2092 | // sll shiftednewval,maskednewval,shiftamt |
2093 | int64_t MaskImm = (Size == 1) ? 255 : 65535; |
2094 | BuildMI(BB, MIMD: DL, MCID: TII->get(Opcode: ArePtrs64bit ? Mips::DADDiu : Mips::ADDiu), DestReg: MaskLSB2) |
2095 | .addReg(RegNo: ABI.GetNullPtr()).addImm(Val: -4); |
2096 | BuildMI(BB, MIMD: DL, MCID: TII->get(Opcode: ArePtrs64bit ? Mips::AND64 : Mips::AND), DestReg: AlignedAddr) |
2097 | .addReg(RegNo: Ptr).addReg(RegNo: MaskLSB2); |
2098 | BuildMI(BB, MIMD: DL, MCID: TII->get(Opcode: Mips::ANDi), DestReg: PtrLSB2) |
2099 | .addReg(RegNo: Ptr, flags: 0, SubReg: ArePtrs64bit ? Mips::sub_32 : 0).addImm(Val: 3); |
2100 | if (Subtarget.isLittle()) { |
2101 | BuildMI(BB, MIMD: DL, MCID: TII->get(Opcode: Mips::SLL), DestReg: ShiftAmt).addReg(RegNo: PtrLSB2).addImm(Val: 3); |
2102 | } else { |
2103 | Register Off = RegInfo.createVirtualRegister(RegClass: RC); |
2104 | BuildMI(BB, MIMD: DL, MCID: TII->get(Opcode: Mips::XORi), DestReg: Off) |
2105 | .addReg(RegNo: PtrLSB2).addImm(Val: (Size == 1) ? 3 : 2); |
2106 | BuildMI(BB, MIMD: DL, MCID: TII->get(Opcode: Mips::SLL), DestReg: ShiftAmt).addReg(RegNo: Off).addImm(Val: 3); |
2107 | } |
2108 | BuildMI(BB, MIMD: DL, MCID: TII->get(Opcode: Mips::ORi), DestReg: MaskUpper) |
2109 | .addReg(RegNo: Mips::ZERO).addImm(Val: MaskImm); |
2110 | BuildMI(BB, MIMD: DL, MCID: TII->get(Opcode: Mips::SLLV), DestReg: Mask) |
2111 | .addReg(RegNo: MaskUpper).addReg(RegNo: ShiftAmt); |
2112 | BuildMI(BB, MIMD: DL, MCID: TII->get(Opcode: Mips::NOR), DestReg: Mask2).addReg(RegNo: Mips::ZERO).addReg(RegNo: Mask); |
2113 | BuildMI(BB, MIMD: DL, MCID: TII->get(Opcode: Mips::ANDi), DestReg: MaskedCmpVal) |
2114 | .addReg(RegNo: CmpVal).addImm(Val: MaskImm); |
2115 | BuildMI(BB, MIMD: DL, MCID: TII->get(Opcode: Mips::SLLV), DestReg: ShiftedCmpVal) |
2116 | .addReg(RegNo: MaskedCmpVal).addReg(RegNo: ShiftAmt); |
2117 | BuildMI(BB, MIMD: DL, MCID: TII->get(Opcode: Mips::ANDi), DestReg: MaskedNewVal) |
2118 | .addReg(RegNo: NewVal).addImm(Val: MaskImm); |
2119 | BuildMI(BB, MIMD: DL, MCID: TII->get(Opcode: Mips::SLLV), DestReg: ShiftedNewVal) |
2120 | .addReg(RegNo: MaskedNewVal).addReg(RegNo: ShiftAmt); |
2121 | |
2122 | // The purposes of the flags on the scratch registers are explained in |
2123 | // emitAtomicBinary. In summary, we need a scratch register which is going to |
2124 | // be undef, that is unique among the register chosen for the instruction. |
2125 | |
2126 | BuildMI(BB, MIMD: DL, MCID: TII->get(Opcode: AtomicOp)) |
2127 | .addReg(RegNo: Dest, flags: RegState::Define | RegState::EarlyClobber) |
2128 | .addReg(RegNo: AlignedAddr) |
2129 | .addReg(RegNo: Mask) |
2130 | .addReg(RegNo: ShiftedCmpVal) |
2131 | .addReg(RegNo: Mask2) |
2132 | .addReg(RegNo: ShiftedNewVal) |
2133 | .addReg(RegNo: ShiftAmt) |
2134 | .addReg(RegNo: Scratch, flags: RegState::EarlyClobber | RegState::Define | |
2135 | RegState::Dead | RegState::Implicit) |
2136 | .addReg(RegNo: Scratch2, flags: RegState::EarlyClobber | RegState::Define | |
2137 | RegState::Dead | RegState::Implicit); |
2138 | |
2139 | MI.eraseFromParent(); // The instruction is gone now. |
2140 | |
2141 | return exitMBB; |
2142 | } |
2143 | |
2144 | SDValue MipsTargetLowering::lowerREADCYCLECOUNTER(SDValue Op, |
2145 | SelectionDAG &DAG) const { |
2146 | SmallVector<SDValue, 3> Results; |
2147 | SDLoc DL(Op); |
2148 | MachineFunction &MF = DAG.getMachineFunction(); |
2149 | unsigned RdhwrOpc, DestReg; |
2150 | EVT PtrVT = getPointerTy(DL: DAG.getDataLayout()); |
2151 | |
2152 | if (PtrVT == MVT::i64) { |
2153 | RdhwrOpc = Mips::RDHWR64; |
2154 | DestReg = MF.getRegInfo().createVirtualRegister(RegClass: getRegClassFor(VT: MVT::i64)); |
2155 | SDNode *Rdhwr = DAG.getMachineNode(Opcode: RdhwrOpc, dl: DL, VT1: MVT::i64, VT2: MVT::Glue, |
2156 | Op1: DAG.getRegister(Reg: Mips::HWR2, VT: MVT::i32), |
2157 | Op2: DAG.getTargetConstant(Val: 0, DL, VT: MVT::i32)); |
2158 | SDValue Chain = DAG.getCopyToReg(Chain: DAG.getEntryNode(), dl: DL, Reg: DestReg, |
2159 | N: SDValue(Rdhwr, 0), Glue: SDValue(Rdhwr, 1)); |
2160 | SDValue ResNode = |
2161 | DAG.getCopyFromReg(Chain, dl: DL, Reg: DestReg, VT: MVT::i64, Glue: Chain.getValue(R: 1)); |
2162 | Results.push_back(Elt: ResNode); |
2163 | Results.push_back(Elt: ResNode.getValue(R: 1)); |
2164 | } else { |
2165 | RdhwrOpc = Mips::RDHWR; |
2166 | DestReg = MF.getRegInfo().createVirtualRegister(RegClass: getRegClassFor(VT: MVT::i32)); |
2167 | SDNode *Rdhwr = DAG.getMachineNode(Opcode: RdhwrOpc, dl: DL, VT1: MVT::i32, VT2: MVT::Glue, |
2168 | Op1: DAG.getRegister(Reg: Mips::HWR2, VT: MVT::i32), |
2169 | Op2: DAG.getTargetConstant(Val: 0, DL, VT: MVT::i32)); |
2170 | SDValue Chain = DAG.getCopyToReg(Chain: DAG.getEntryNode(), dl: DL, Reg: DestReg, |
2171 | N: SDValue(Rdhwr, 0), Glue: SDValue(Rdhwr, 1)); |
2172 | SDValue ResNode = |
2173 | DAG.getCopyFromReg(Chain, dl: DL, Reg: DestReg, VT: MVT::i32, Glue: Chain.getValue(R: 1)); |
2174 | Results.push_back(Elt: DAG.getNode(Opcode: ISD::BUILD_PAIR, DL, VT: MVT::i64, N1: ResNode, |
2175 | N2: DAG.getConstant(Val: 0, DL, VT: MVT::i32))); |
2176 | Results.push_back(Elt: ResNode.getValue(R: 1)); |
2177 | } |
2178 | |
2179 | return DAG.getMergeValues(Ops: Results, dl: DL); |
2180 | } |
2181 | |
2182 | SDValue MipsTargetLowering::lowerBRCOND(SDValue Op, SelectionDAG &DAG) const { |
2183 | // The first operand is the chain, the second is the condition, the third is |
2184 | // the block to branch to if the condition is true. |
2185 | SDValue Chain = Op.getOperand(i: 0); |
2186 | SDValue Dest = Op.getOperand(i: 2); |
2187 | SDLoc DL(Op); |
2188 | |
2189 | assert(!Subtarget.hasMips32r6() && !Subtarget.hasMips64r6()); |
2190 | SDValue CondRes = createFPCmp(DAG, Op: Op.getOperand(i: 1)); |
2191 | |
2192 | // Return if flag is not set by a floating point comparison. |
2193 | if (CondRes.getOpcode() != MipsISD::FPCmp) |
2194 | return Op; |
2195 | |
2196 | SDValue CCNode = CondRes.getOperand(i: 2); |
2197 | Mips::CondCode CC = (Mips::CondCode)CCNode->getAsZExtVal(); |
2198 | unsigned Opc = invertFPCondCodeUser(CC) ? Mips::BRANCH_F : Mips::BRANCH_T; |
2199 | SDValue BrCode = DAG.getConstant(Val: Opc, DL, VT: MVT::i32); |
2200 | SDValue FCC0 = DAG.getRegister(Reg: Mips::FCC0, VT: MVT::i32); |
2201 | return DAG.getNode(Opcode: MipsISD::FPBrcond, DL, VT: Op.getValueType(), N1: Chain, N2: BrCode, |
2202 | N3: FCC0, N4: Dest, N5: CondRes); |
2203 | } |
2204 | |
2205 | SDValue MipsTargetLowering:: |
2206 | lowerSELECT(SDValue Op, SelectionDAG &DAG) const |
2207 | { |
2208 | assert(!Subtarget.hasMips32r6() && !Subtarget.hasMips64r6()); |
2209 | SDValue Cond = createFPCmp(DAG, Op: Op.getOperand(i: 0)); |
2210 | |
2211 | // Return if flag is not set by a floating point comparison. |
2212 | if (Cond.getOpcode() != MipsISD::FPCmp) |
2213 | return Op; |
2214 | |
2215 | return createCMovFP(DAG, Cond, True: Op.getOperand(i: 1), False: Op.getOperand(i: 2), |
2216 | DL: SDLoc(Op)); |
2217 | } |
2218 | |
2219 | SDValue MipsTargetLowering::lowerSETCC(SDValue Op, SelectionDAG &DAG) const { |
2220 | assert(!Subtarget.hasMips32r6() && !Subtarget.hasMips64r6()); |
2221 | SDValue Cond = createFPCmp(DAG, Op); |
2222 | |
2223 | assert(Cond.getOpcode() == MipsISD::FPCmp && |
2224 | "Floating point operand expected." ); |
2225 | |
2226 | SDLoc DL(Op); |
2227 | SDValue True = DAG.getConstant(Val: 1, DL, VT: MVT::i32); |
2228 | SDValue False = DAG.getConstant(Val: 0, DL, VT: MVT::i32); |
2229 | |
2230 | return createCMovFP(DAG, Cond, True, False, DL); |
2231 | } |
2232 | |
2233 | SDValue MipsTargetLowering::lowerGlobalAddress(SDValue Op, |
2234 | SelectionDAG &DAG) const { |
2235 | EVT Ty = Op.getValueType(); |
2236 | GlobalAddressSDNode *N = cast<GlobalAddressSDNode>(Val&: Op); |
2237 | const GlobalValue *GV = N->getGlobal(); |
2238 | |
2239 | if (GV->hasDLLImportStorageClass()) { |
2240 | assert(Subtarget.isTargetWindows() && |
2241 | "Windows is the only supported COFF target" ); |
2242 | return getDllimportVariable( |
2243 | N, DL: SDLoc(N), Ty, DAG, Chain: DAG.getEntryNode(), |
2244 | PtrInfo: MachinePointerInfo::getGOT(MF&: DAG.getMachineFunction())); |
2245 | } |
2246 | |
2247 | if (!isPositionIndependent()) { |
2248 | const MipsTargetObjectFile *TLOF = |
2249 | static_cast<const MipsTargetObjectFile *>( |
2250 | getTargetMachine().getObjFileLowering()); |
2251 | const GlobalObject *GO = GV->getAliaseeObject(); |
2252 | if (GO && TLOF->IsGlobalInSmallSection(GO, TM: getTargetMachine())) |
2253 | // %gp_rel relocation |
2254 | return getAddrGPRel(N, DL: SDLoc(N), Ty, DAG, IsN64: ABI.IsN64()); |
2255 | |
2256 | // %hi/%lo relocation |
2257 | return Subtarget.hasSym32() ? getAddrNonPIC(N, DL: SDLoc(N), Ty, DAG) |
2258 | // %highest/%higher/%hi/%lo relocation |
2259 | : getAddrNonPICSym64(N, DL: SDLoc(N), Ty, DAG); |
2260 | } |
2261 | |
2262 | // Every other architecture would use shouldAssumeDSOLocal in here, but |
2263 | // mips is special. |
2264 | // * In PIC code mips requires got loads even for local statics! |
2265 | // * To save on got entries, for local statics the got entry contains the |
2266 | // page and an additional add instruction takes care of the low bits. |
2267 | // * It is legal to access a hidden symbol with a non hidden undefined, |
2268 | // so one cannot guarantee that all access to a hidden symbol will know |
2269 | // it is hidden. |
2270 | // * Mips linkers don't support creating a page and a full got entry for |
2271 | // the same symbol. |
2272 | // * Given all that, we have to use a full got entry for hidden symbols :-( |
2273 | if (GV->hasLocalLinkage()) |
2274 | return getAddrLocal(N, DL: SDLoc(N), Ty, DAG, IsN32OrN64: ABI.IsN32() || ABI.IsN64()); |
2275 | |
2276 | if (Subtarget.useXGOT()) |
2277 | return getAddrGlobalLargeGOT( |
2278 | N, DL: SDLoc(N), Ty, DAG, HiFlag: MipsII::MO_GOT_HI16, LoFlag: MipsII::MO_GOT_LO16, |
2279 | Chain: DAG.getEntryNode(), |
2280 | PtrInfo: MachinePointerInfo::getGOT(MF&: DAG.getMachineFunction())); |
2281 | |
2282 | return getAddrGlobal( |
2283 | N, DL: SDLoc(N), Ty, DAG, |
2284 | Flag: (ABI.IsN32() || ABI.IsN64()) ? MipsII::MO_GOT_DISP : MipsII::MO_GOT, |
2285 | Chain: DAG.getEntryNode(), PtrInfo: MachinePointerInfo::getGOT(MF&: DAG.getMachineFunction())); |
2286 | } |
2287 | |
2288 | SDValue MipsTargetLowering::lowerBlockAddress(SDValue Op, |
2289 | SelectionDAG &DAG) const { |
2290 | BlockAddressSDNode *N = cast<BlockAddressSDNode>(Val&: Op); |
2291 | EVT Ty = Op.getValueType(); |
2292 | |
2293 | if (!isPositionIndependent()) |
2294 | return Subtarget.hasSym32() ? getAddrNonPIC(N, DL: SDLoc(N), Ty, DAG) |
2295 | : getAddrNonPICSym64(N, DL: SDLoc(N), Ty, DAG); |
2296 | |
2297 | return getAddrLocal(N, DL: SDLoc(N), Ty, DAG, IsN32OrN64: ABI.IsN32() || ABI.IsN64()); |
2298 | } |
2299 | |
2300 | SDValue MipsTargetLowering:: |
2301 | lowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const |
2302 | { |
2303 | // If the relocation model is PIC, use the General Dynamic TLS Model or |
2304 | // Local Dynamic TLS model, otherwise use the Initial Exec or |
2305 | // Local Exec TLS Model. |
2306 | |
2307 | GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Val&: Op); |
2308 | if (DAG.getTarget().useEmulatedTLS()) |
2309 | return LowerToTLSEmulatedModel(GA, DAG); |
2310 | |
2311 | SDLoc DL(GA); |
2312 | const GlobalValue *GV = GA->getGlobal(); |
2313 | EVT PtrVT = getPointerTy(DL: DAG.getDataLayout()); |
2314 | |
2315 | TLSModel::Model model = getTargetMachine().getTLSModel(GV); |
2316 | |
2317 | if (model == TLSModel::GeneralDynamic || model == TLSModel::LocalDynamic) { |
2318 | // General Dynamic and Local Dynamic TLS Model. |
2319 | unsigned Flag = (model == TLSModel::LocalDynamic) ? MipsII::MO_TLSLDM |
2320 | : MipsII::MO_TLSGD; |
2321 | |
2322 | SDValue TGA = DAG.getTargetGlobalAddress(GV, DL, VT: PtrVT, offset: 0, TargetFlags: Flag); |
2323 | SDValue Argument = DAG.getNode(Opcode: MipsISD::Wrapper, DL, VT: PtrVT, |
2324 | N1: getGlobalReg(DAG, Ty: PtrVT), N2: TGA); |
2325 | unsigned PtrSize = PtrVT.getSizeInBits(); |
2326 | IntegerType *PtrTy = Type::getIntNTy(C&: *DAG.getContext(), N: PtrSize); |
2327 | |
2328 | SDValue TlsGetAddr = DAG.getExternalSymbol(Sym: "__tls_get_addr" , VT: PtrVT); |
2329 | |
2330 | ArgListTy Args; |
2331 | ArgListEntry Entry; |
2332 | Entry.Node = Argument; |
2333 | Entry.Ty = PtrTy; |
2334 | Args.push_back(x: Entry); |
2335 | |
2336 | TargetLowering::CallLoweringInfo CLI(DAG); |
2337 | CLI.setDebugLoc(DL) |
2338 | .setChain(DAG.getEntryNode()) |
2339 | .setLibCallee(CC: CallingConv::C, ResultType: PtrTy, Target: TlsGetAddr, ArgsList: std::move(Args)); |
2340 | std::pair<SDValue, SDValue> CallResult = LowerCallTo(CLI); |
2341 | |
2342 | SDValue Ret = CallResult.first; |
2343 | |
2344 | if (model != TLSModel::LocalDynamic) |
2345 | return Ret; |
2346 | |
2347 | SDValue TGAHi = DAG.getTargetGlobalAddress(GV, DL, VT: PtrVT, offset: 0, |
2348 | TargetFlags: MipsII::MO_DTPREL_HI); |
2349 | SDValue Hi = DAG.getNode(Opcode: MipsISD::TlsHi, DL, VT: PtrVT, Operand: TGAHi); |
2350 | SDValue TGALo = DAG.getTargetGlobalAddress(GV, DL, VT: PtrVT, offset: 0, |
2351 | TargetFlags: MipsII::MO_DTPREL_LO); |
2352 | SDValue Lo = DAG.getNode(Opcode: MipsISD::Lo, DL, VT: PtrVT, Operand: TGALo); |
2353 | SDValue Add = DAG.getNode(Opcode: ISD::ADD, DL, VT: PtrVT, N1: Hi, N2: Ret); |
2354 | return DAG.getNode(Opcode: ISD::ADD, DL, VT: PtrVT, N1: Add, N2: Lo); |
2355 | } |
2356 | |
2357 | SDValue Offset; |
2358 | if (model == TLSModel::InitialExec) { |
2359 | // Initial Exec TLS Model |
2360 | SDValue TGA = DAG.getTargetGlobalAddress(GV, DL, VT: PtrVT, offset: 0, |
2361 | TargetFlags: MipsII::MO_GOTTPREL); |
2362 | TGA = DAG.getNode(Opcode: MipsISD::Wrapper, DL, VT: PtrVT, N1: getGlobalReg(DAG, Ty: PtrVT), |
2363 | N2: TGA); |
2364 | Offset = |
2365 | DAG.getLoad(VT: PtrVT, dl: DL, Chain: DAG.getEntryNode(), Ptr: TGA, PtrInfo: MachinePointerInfo()); |
2366 | } else { |
2367 | // Local Exec TLS Model |
2368 | assert(model == TLSModel::LocalExec); |
2369 | SDValue TGAHi = DAG.getTargetGlobalAddress(GV, DL, VT: PtrVT, offset: 0, |
2370 | TargetFlags: MipsII::MO_TPREL_HI); |
2371 | SDValue TGALo = DAG.getTargetGlobalAddress(GV, DL, VT: PtrVT, offset: 0, |
2372 | TargetFlags: MipsII::MO_TPREL_LO); |
2373 | SDValue Hi = DAG.getNode(Opcode: MipsISD::TlsHi, DL, VT: PtrVT, Operand: TGAHi); |
2374 | SDValue Lo = DAG.getNode(Opcode: MipsISD::Lo, DL, VT: PtrVT, Operand: TGALo); |
2375 | Offset = DAG.getNode(Opcode: ISD::ADD, DL, VT: PtrVT, N1: Hi, N2: Lo); |
2376 | } |
2377 | |
2378 | SDValue ThreadPointer = DAG.getNode(Opcode: MipsISD::ThreadPointer, DL, VT: PtrVT); |
2379 | return DAG.getNode(Opcode: ISD::ADD, DL, VT: PtrVT, N1: ThreadPointer, N2: Offset); |
2380 | } |
2381 | |
2382 | SDValue MipsTargetLowering:: |
2383 | lowerJumpTable(SDValue Op, SelectionDAG &DAG) const |
2384 | { |
2385 | JumpTableSDNode *N = cast<JumpTableSDNode>(Val&: Op); |
2386 | EVT Ty = Op.getValueType(); |
2387 | |
2388 | if (!isPositionIndependent()) |
2389 | return Subtarget.hasSym32() ? getAddrNonPIC(N, DL: SDLoc(N), Ty, DAG) |
2390 | : getAddrNonPICSym64(N, DL: SDLoc(N), Ty, DAG); |
2391 | |
2392 | return getAddrLocal(N, DL: SDLoc(N), Ty, DAG, IsN32OrN64: ABI.IsN32() || ABI.IsN64()); |
2393 | } |
2394 | |
2395 | SDValue MipsTargetLowering:: |
2396 | lowerConstantPool(SDValue Op, SelectionDAG &DAG) const |
2397 | { |
2398 | ConstantPoolSDNode *N = cast<ConstantPoolSDNode>(Val&: Op); |
2399 | EVT Ty = Op.getValueType(); |
2400 | |
2401 | if (!isPositionIndependent()) { |
2402 | const MipsTargetObjectFile *TLOF = |
2403 | static_cast<const MipsTargetObjectFile *>( |
2404 | getTargetMachine().getObjFileLowering()); |
2405 | |
2406 | if (TLOF->IsConstantInSmallSection(DL: DAG.getDataLayout(), CN: N->getConstVal(), |
2407 | TM: getTargetMachine())) |
2408 | // %gp_rel relocation |
2409 | return getAddrGPRel(N, DL: SDLoc(N), Ty, DAG, IsN64: ABI.IsN64()); |
2410 | |
2411 | return Subtarget.hasSym32() ? getAddrNonPIC(N, DL: SDLoc(N), Ty, DAG) |
2412 | : getAddrNonPICSym64(N, DL: SDLoc(N), Ty, DAG); |
2413 | } |
2414 | |
2415 | return getAddrLocal(N, DL: SDLoc(N), Ty, DAG, IsN32OrN64: ABI.IsN32() || ABI.IsN64()); |
2416 | } |
2417 | |
2418 | SDValue MipsTargetLowering::lowerVASTART(SDValue Op, SelectionDAG &DAG) const { |
2419 | MachineFunction &MF = DAG.getMachineFunction(); |
2420 | MipsFunctionInfo *FuncInfo = MF.getInfo<MipsFunctionInfo>(); |
2421 | |
2422 | SDLoc DL(Op); |
2423 | SDValue FI = DAG.getFrameIndex(FI: FuncInfo->getVarArgsFrameIndex(), |
2424 | VT: getPointerTy(DL: MF.getDataLayout())); |
2425 | |
2426 | // vastart just stores the address of the VarArgsFrameIndex slot into the |
2427 | // memory location argument. |
2428 | const Value *SV = cast<SrcValueSDNode>(Val: Op.getOperand(i: 2))->getValue(); |
2429 | return DAG.getStore(Chain: Op.getOperand(i: 0), dl: DL, Val: FI, Ptr: Op.getOperand(i: 1), |
2430 | PtrInfo: MachinePointerInfo(SV)); |
2431 | } |
2432 | |
2433 | SDValue MipsTargetLowering::lowerVAARG(SDValue Op, SelectionDAG &DAG) const { |
2434 | SDNode *Node = Op.getNode(); |
2435 | EVT VT = Node->getValueType(ResNo: 0); |
2436 | SDValue Chain = Node->getOperand(Num: 0); |
2437 | SDValue VAListPtr = Node->getOperand(Num: 1); |
2438 | const Align Align = |
2439 | llvm::MaybeAlign(Node->getConstantOperandVal(Num: 3)).valueOrOne(); |
2440 | const Value *SV = cast<SrcValueSDNode>(Val: Node->getOperand(Num: 2))->getValue(); |
2441 | SDLoc DL(Node); |
2442 | unsigned ArgSlotSizeInBytes = (ABI.IsN32() || ABI.IsN64()) ? 8 : 4; |
2443 | |
2444 | SDValue VAListLoad = DAG.getLoad(VT: getPointerTy(DL: DAG.getDataLayout()), dl: DL, Chain, |
2445 | Ptr: VAListPtr, PtrInfo: MachinePointerInfo(SV)); |
2446 | SDValue VAList = VAListLoad; |
2447 | |
2448 | // Re-align the pointer if necessary. |
2449 | // It should only ever be necessary for 64-bit types on O32 since the minimum |
2450 | // argument alignment is the same as the maximum type alignment for N32/N64. |
2451 | // |
2452 | // FIXME: We currently align too often. The code generator doesn't notice |
2453 | // when the pointer is still aligned from the last va_arg (or pair of |
2454 | // va_args for the i64 on O32 case). |
2455 | if (Align > getMinStackArgumentAlignment()) { |
2456 | VAList = DAG.getNode( |
2457 | Opcode: ISD::ADD, DL, VT: VAList.getValueType(), N1: VAList, |
2458 | N2: DAG.getConstant(Val: Align.value() - 1, DL, VT: VAList.getValueType())); |
2459 | |
2460 | VAList = DAG.getNode(Opcode: ISD::AND, DL, VT: VAList.getValueType(), N1: VAList, |
2461 | N2: DAG.getSignedConstant(Val: -(int64_t)Align.value(), DL, |
2462 | VT: VAList.getValueType())); |
2463 | } |
2464 | |
2465 | // Increment the pointer, VAList, to the next vaarg. |
2466 | auto &TD = DAG.getDataLayout(); |
2467 | unsigned ArgSizeInBytes = |
2468 | TD.getTypeAllocSize(Ty: VT.getTypeForEVT(Context&: *DAG.getContext())); |
2469 | SDValue Tmp3 = |
2470 | DAG.getNode(Opcode: ISD::ADD, DL, VT: VAList.getValueType(), N1: VAList, |
2471 | N2: DAG.getConstant(Val: alignTo(Value: ArgSizeInBytes, Align: ArgSlotSizeInBytes), |
2472 | DL, VT: VAList.getValueType())); |
2473 | // Store the incremented VAList to the legalized pointer |
2474 | Chain = DAG.getStore(Chain: VAListLoad.getValue(R: 1), dl: DL, Val: Tmp3, Ptr: VAListPtr, |
2475 | PtrInfo: MachinePointerInfo(SV)); |
2476 | |
2477 | // In big-endian mode we must adjust the pointer when the load size is smaller |
2478 | // than the argument slot size. We must also reduce the known alignment to |
2479 | // match. For example in the N64 ABI, we must add 4 bytes to the offset to get |
2480 | // the correct half of the slot, and reduce the alignment from 8 (slot |
2481 | // alignment) down to 4 (type alignment). |
2482 | if (!Subtarget.isLittle() && ArgSizeInBytes < ArgSlotSizeInBytes) { |
2483 | unsigned Adjustment = ArgSlotSizeInBytes - ArgSizeInBytes; |
2484 | VAList = DAG.getNode(Opcode: ISD::ADD, DL, VT: VAListPtr.getValueType(), N1: VAList, |
2485 | N2: DAG.getIntPtrConstant(Val: Adjustment, DL)); |
2486 | } |
2487 | // Load the actual argument out of the pointer VAList |
2488 | return DAG.getLoad(VT, dl: DL, Chain, Ptr: VAList, PtrInfo: MachinePointerInfo()); |
2489 | } |
2490 | |
2491 | static SDValue lowerFCOPYSIGN32(SDValue Op, SelectionDAG &DAG, |
2492 | bool ) { |
2493 | EVT TyX = Op.getOperand(i: 0).getValueType(); |
2494 | EVT TyY = Op.getOperand(i: 1).getValueType(); |
2495 | SDLoc DL(Op); |
2496 | SDValue Const1 = DAG.getConstant(Val: 1, DL, VT: MVT::i32); |
2497 | SDValue Const31 = DAG.getConstant(Val: 31, DL, VT: MVT::i32); |
2498 | SDValue Res; |
2499 | |
2500 | // If operand is of type f64, extract the upper 32-bit. Otherwise, bitcast it |
2501 | // to i32. |
2502 | SDValue X = (TyX == MVT::f32) ? |
2503 | DAG.getNode(Opcode: ISD::BITCAST, DL, VT: MVT::i32, Operand: Op.getOperand(i: 0)) : |
2504 | DAG.getNode(Opcode: MipsISD::ExtractElementF64, DL, VT: MVT::i32, N1: Op.getOperand(i: 0), |
2505 | N2: Const1); |
2506 | SDValue Y = (TyY == MVT::f32) ? |
2507 | DAG.getNode(Opcode: ISD::BITCAST, DL, VT: MVT::i32, Operand: Op.getOperand(i: 1)) : |
2508 | DAG.getNode(Opcode: MipsISD::ExtractElementF64, DL, VT: MVT::i32, N1: Op.getOperand(i: 1), |
2509 | N2: Const1); |
2510 | |
2511 | if (HasExtractInsert) { |
2512 | // ext E, Y, 31, 1 ; extract bit31 of Y |
2513 | // ins X, E, 31, 1 ; insert extracted bit at bit31 of X |
2514 | SDValue E = DAG.getNode(Opcode: MipsISD::Ext, DL, VT: MVT::i32, N1: Y, N2: Const31, N3: Const1); |
2515 | Res = DAG.getNode(Opcode: MipsISD::Ins, DL, VT: MVT::i32, N1: E, N2: Const31, N3: Const1, N4: X); |
2516 | } else { |
2517 | // sll SllX, X, 1 |
2518 | // srl SrlX, SllX, 1 |
2519 | // srl SrlY, Y, 31 |
2520 | // sll SllY, SrlX, 31 |
2521 | // or Or, SrlX, SllY |
2522 | SDValue SllX = DAG.getNode(Opcode: ISD::SHL, DL, VT: MVT::i32, N1: X, N2: Const1); |
2523 | SDValue SrlX = DAG.getNode(Opcode: ISD::SRL, DL, VT: MVT::i32, N1: SllX, N2: Const1); |
2524 | SDValue SrlY = DAG.getNode(Opcode: ISD::SRL, DL, VT: MVT::i32, N1: Y, N2: Const31); |
2525 | SDValue SllY = DAG.getNode(Opcode: ISD::SHL, DL, VT: MVT::i32, N1: SrlY, N2: Const31); |
2526 | Res = DAG.getNode(Opcode: ISD::OR, DL, VT: MVT::i32, N1: SrlX, N2: SllY); |
2527 | } |
2528 | |
2529 | if (TyX == MVT::f32) |
2530 | return DAG.getNode(Opcode: ISD::BITCAST, DL, VT: Op.getOperand(i: 0).getValueType(), Operand: Res); |
2531 | |
2532 | SDValue LowX = DAG.getNode(Opcode: MipsISD::ExtractElementF64, DL, VT: MVT::i32, |
2533 | N1: Op.getOperand(i: 0), |
2534 | N2: DAG.getConstant(Val: 0, DL, VT: MVT::i32)); |
2535 | return DAG.getNode(Opcode: MipsISD::BuildPairF64, DL, VT: MVT::f64, N1: LowX, N2: Res); |
2536 | } |
2537 | |
2538 | static SDValue lowerFCOPYSIGN64(SDValue Op, SelectionDAG &DAG, |
2539 | bool ) { |
2540 | unsigned WidthX = Op.getOperand(i: 0).getValueSizeInBits(); |
2541 | unsigned WidthY = Op.getOperand(i: 1).getValueSizeInBits(); |
2542 | EVT TyX = MVT::getIntegerVT(BitWidth: WidthX), TyY = MVT::getIntegerVT(BitWidth: WidthY); |
2543 | SDLoc DL(Op); |
2544 | SDValue Const1 = DAG.getConstant(Val: 1, DL, VT: MVT::i32); |
2545 | |
2546 | // Bitcast to integer nodes. |
2547 | SDValue X = DAG.getNode(Opcode: ISD::BITCAST, DL, VT: TyX, Operand: Op.getOperand(i: 0)); |
2548 | SDValue Y = DAG.getNode(Opcode: ISD::BITCAST, DL, VT: TyY, Operand: Op.getOperand(i: 1)); |
2549 | |
2550 | if (HasExtractInsert) { |
2551 | // ext E, Y, width(Y) - 1, 1 ; extract bit width(Y)-1 of Y |
2552 | // ins X, E, width(X) - 1, 1 ; insert extracted bit at bit width(X)-1 of X |
2553 | SDValue E = DAG.getNode(Opcode: MipsISD::Ext, DL, VT: TyY, N1: Y, |
2554 | N2: DAG.getConstant(Val: WidthY - 1, DL, VT: MVT::i32), N3: Const1); |
2555 | |
2556 | if (WidthX > WidthY) |
2557 | E = DAG.getNode(Opcode: ISD::ZERO_EXTEND, DL, VT: TyX, Operand: E); |
2558 | else if (WidthY > WidthX) |
2559 | E = DAG.getNode(Opcode: ISD::TRUNCATE, DL, VT: TyX, Operand: E); |
2560 | |
2561 | SDValue I = DAG.getNode(Opcode: MipsISD::Ins, DL, VT: TyX, N1: E, |
2562 | N2: DAG.getConstant(Val: WidthX - 1, DL, VT: MVT::i32), N3: Const1, |
2563 | N4: X); |
2564 | return DAG.getNode(Opcode: ISD::BITCAST, DL, VT: Op.getOperand(i: 0).getValueType(), Operand: I); |
2565 | } |
2566 | |
2567 | // (d)sll SllX, X, 1 |
2568 | // (d)srl SrlX, SllX, 1 |
2569 | // (d)srl SrlY, Y, width(Y)-1 |
2570 | // (d)sll SllY, SrlX, width(Y)-1 |
2571 | // or Or, SrlX, SllY |
2572 | SDValue SllX = DAG.getNode(Opcode: ISD::SHL, DL, VT: TyX, N1: X, N2: Const1); |
2573 | SDValue SrlX = DAG.getNode(Opcode: ISD::SRL, DL, VT: TyX, N1: SllX, N2: Const1); |
2574 | SDValue SrlY = DAG.getNode(Opcode: ISD::SRL, DL, VT: TyY, N1: Y, |
2575 | N2: DAG.getConstant(Val: WidthY - 1, DL, VT: MVT::i32)); |
2576 | |
2577 | if (WidthX > WidthY) |
2578 | SrlY = DAG.getNode(Opcode: ISD::ZERO_EXTEND, DL, VT: TyX, Operand: SrlY); |
2579 | else if (WidthY > WidthX) |
2580 | SrlY = DAG.getNode(Opcode: ISD::TRUNCATE, DL, VT: TyX, Operand: SrlY); |
2581 | |
2582 | SDValue SllY = DAG.getNode(Opcode: ISD::SHL, DL, VT: TyX, N1: SrlY, |
2583 | N2: DAG.getConstant(Val: WidthX - 1, DL, VT: MVT::i32)); |
2584 | SDValue Or = DAG.getNode(Opcode: ISD::OR, DL, VT: TyX, N1: SrlX, N2: SllY); |
2585 | return DAG.getNode(Opcode: ISD::BITCAST, DL, VT: Op.getOperand(i: 0).getValueType(), Operand: Or); |
2586 | } |
2587 | |
2588 | SDValue |
2589 | MipsTargetLowering::lowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) const { |
2590 | if (Subtarget.isGP64bit()) |
2591 | return lowerFCOPYSIGN64(Op, DAG, HasExtractInsert: Subtarget.hasExtractInsert()); |
2592 | |
2593 | return lowerFCOPYSIGN32(Op, DAG, HasExtractInsert: Subtarget.hasExtractInsert()); |
2594 | } |
2595 | |
2596 | SDValue MipsTargetLowering::lowerFABS32(SDValue Op, SelectionDAG &DAG, |
2597 | bool ) const { |
2598 | SDLoc DL(Op); |
2599 | SDValue Res, Const1 = DAG.getConstant(Val: 1, DL, VT: MVT::i32); |
2600 | |
2601 | if (DAG.getTarget().Options.NoNaNsFPMath || Subtarget.inAbs2008Mode()) |
2602 | return DAG.getNode(Opcode: MipsISD::FAbs, DL, VT: Op.getValueType(), Operand: Op.getOperand(i: 0)); |
2603 | |
2604 | // If operand is of type f64, extract the upper 32-bit. Otherwise, bitcast it |
2605 | // to i32. |
2606 | SDValue X = (Op.getValueType() == MVT::f32) |
2607 | ? DAG.getNode(Opcode: ISD::BITCAST, DL, VT: MVT::i32, Operand: Op.getOperand(i: 0)) |
2608 | : DAG.getNode(Opcode: MipsISD::ExtractElementF64, DL, VT: MVT::i32, |
2609 | N1: Op.getOperand(i: 0), N2: Const1); |
2610 | |
2611 | // Clear MSB. |
2612 | if (HasExtractInsert) |
2613 | Res = DAG.getNode(Opcode: MipsISD::Ins, DL, VT: MVT::i32, |
2614 | N1: DAG.getRegister(Reg: Mips::ZERO, VT: MVT::i32), |
2615 | N2: DAG.getConstant(Val: 31, DL, VT: MVT::i32), N3: Const1, N4: X); |
2616 | else { |
2617 | // TODO: Provide DAG patterns which transform (and x, cst) |
2618 | // back to a (shl (srl x (clz cst)) (clz cst)) sequence. |
2619 | SDValue SllX = DAG.getNode(Opcode: ISD::SHL, DL, VT: MVT::i32, N1: X, N2: Const1); |
2620 | Res = DAG.getNode(Opcode: ISD::SRL, DL, VT: MVT::i32, N1: SllX, N2: Const1); |
2621 | } |
2622 | |
2623 | if (Op.getValueType() == MVT::f32) |
2624 | return DAG.getNode(Opcode: ISD::BITCAST, DL, VT: MVT::f32, Operand: Res); |
2625 | |
2626 | // FIXME: For mips32r2, the sequence of (BuildPairF64 (ins (ExtractElementF64 |
2627 | // Op 1), $zero, 31 1) (ExtractElementF64 Op 0)) and the Op has one use, we |
2628 | // should be able to drop the usage of mfc1/mtc1 and rewrite the register in |
2629 | // place. |
2630 | SDValue LowX = |
2631 | DAG.getNode(Opcode: MipsISD::ExtractElementF64, DL, VT: MVT::i32, N1: Op.getOperand(i: 0), |
2632 | N2: DAG.getConstant(Val: 0, DL, VT: MVT::i32)); |
2633 | return DAG.getNode(Opcode: MipsISD::BuildPairF64, DL, VT: MVT::f64, N1: LowX, N2: Res); |
2634 | } |
2635 | |
2636 | SDValue MipsTargetLowering::lowerFABS64(SDValue Op, SelectionDAG &DAG, |
2637 | bool ) const { |
2638 | SDLoc DL(Op); |
2639 | SDValue Res, Const1 = DAG.getConstant(Val: 1, DL, VT: MVT::i32); |
2640 | |
2641 | if (DAG.getTarget().Options.NoNaNsFPMath || Subtarget.inAbs2008Mode()) |
2642 | return DAG.getNode(Opcode: MipsISD::FAbs, DL, VT: Op.getValueType(), Operand: Op.getOperand(i: 0)); |
2643 | |
2644 | // Bitcast to integer node. |
2645 | SDValue X = DAG.getNode(Opcode: ISD::BITCAST, DL, VT: MVT::i64, Operand: Op.getOperand(i: 0)); |
2646 | |
2647 | // Clear MSB. |
2648 | if (HasExtractInsert) |
2649 | Res = DAG.getNode(Opcode: MipsISD::Ins, DL, VT: MVT::i64, |
2650 | N1: DAG.getRegister(Reg: Mips::ZERO_64, VT: MVT::i64), |
2651 | N2: DAG.getConstant(Val: 63, DL, VT: MVT::i32), N3: Const1, N4: X); |
2652 | else { |
2653 | SDValue SllX = DAG.getNode(Opcode: ISD::SHL, DL, VT: MVT::i64, N1: X, N2: Const1); |
2654 | Res = DAG.getNode(Opcode: ISD::SRL, DL, VT: MVT::i64, N1: SllX, N2: Const1); |
2655 | } |
2656 | |
2657 | return DAG.getNode(Opcode: ISD::BITCAST, DL, VT: MVT::f64, Operand: Res); |
2658 | } |
2659 | |
2660 | SDValue MipsTargetLowering::lowerFABS(SDValue Op, SelectionDAG &DAG) const { |
2661 | if ((ABI.IsN32() || ABI.IsN64()) && (Op.getValueType() == MVT::f64)) |
2662 | return lowerFABS64(Op, DAG, HasExtractInsert: Subtarget.hasExtractInsert()); |
2663 | |
2664 | return lowerFABS32(Op, DAG, HasExtractInsert: Subtarget.hasExtractInsert()); |
2665 | } |
2666 | |
2667 | SDValue MipsTargetLowering::lowerFCANONICALIZE(SDValue Op, |
2668 | SelectionDAG &DAG) const { |
2669 | SDLoc DL(Op); |
2670 | EVT VT = Op.getValueType(); |
2671 | SDValue Operand = Op.getOperand(i: 0); |
2672 | SDNodeFlags Flags = Op->getFlags(); |
2673 | |
2674 | if (Flags.hasNoNaNs() || DAG.isKnownNeverNaN(Op: Operand)) |
2675 | return Operand; |
2676 | |
2677 | SDValue Quiet = DAG.getNode(Opcode: ISD::FADD, DL, VT, N1: Operand, N2: Operand); |
2678 | return DAG.getSelectCC(DL, LHS: Operand, RHS: Operand, True: Quiet, False: Operand, Cond: ISD::SETUO); |
2679 | } |
2680 | |
2681 | SDValue MipsTargetLowering:: |
2682 | lowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const { |
2683 | // check the depth |
2684 | if (Op.getConstantOperandVal(i: 0) != 0) { |
2685 | DAG.getContext()->emitError( |
2686 | ErrorStr: "return address can be determined only for current frame" ); |
2687 | return SDValue(); |
2688 | } |
2689 | |
2690 | MachineFrameInfo &MFI = DAG.getMachineFunction().getFrameInfo(); |
2691 | MFI.setFrameAddressIsTaken(true); |
2692 | EVT VT = Op.getValueType(); |
2693 | SDLoc DL(Op); |
2694 | SDValue FrameAddr = DAG.getCopyFromReg( |
2695 | Chain: DAG.getEntryNode(), dl: DL, Reg: ABI.IsN64() ? Mips::FP_64 : Mips::FP, VT); |
2696 | return FrameAddr; |
2697 | } |
2698 | |
2699 | SDValue MipsTargetLowering::lowerRETURNADDR(SDValue Op, |
2700 | SelectionDAG &DAG) const { |
2701 | if (verifyReturnAddressArgumentIsConstant(Op, DAG)) |
2702 | return SDValue(); |
2703 | |
2704 | // check the depth |
2705 | if (Op.getConstantOperandVal(i: 0) != 0) { |
2706 | DAG.getContext()->emitError( |
2707 | ErrorStr: "return address can be determined only for current frame" ); |
2708 | return SDValue(); |
2709 | } |
2710 | |
2711 | MachineFunction &MF = DAG.getMachineFunction(); |
2712 | MachineFrameInfo &MFI = MF.getFrameInfo(); |
2713 | MVT VT = Op.getSimpleValueType(); |
2714 | unsigned RA = ABI.IsN64() ? Mips::RA_64 : Mips::RA; |
2715 | MFI.setReturnAddressIsTaken(true); |
2716 | |
2717 | // Return RA, which contains the return address. Mark it an implicit live-in. |
2718 | Register Reg = MF.addLiveIn(PReg: RA, RC: getRegClassFor(VT)); |
2719 | return DAG.getCopyFromReg(Chain: DAG.getEntryNode(), dl: SDLoc(Op), Reg, VT); |
2720 | } |
2721 | |
2722 | // An EH_RETURN is the result of lowering llvm.eh.return which in turn is |
2723 | // generated from __builtin_eh_return (offset, handler) |
2724 | // The effect of this is to adjust the stack pointer by "offset" |
2725 | // and then branch to "handler". |
2726 | SDValue MipsTargetLowering::lowerEH_RETURN(SDValue Op, SelectionDAG &DAG) |
2727 | const { |
2728 | MachineFunction &MF = DAG.getMachineFunction(); |
2729 | MipsFunctionInfo *MipsFI = MF.getInfo<MipsFunctionInfo>(); |
2730 | |
2731 | MipsFI->setCallsEhReturn(); |
2732 | SDValue Chain = Op.getOperand(i: 0); |
2733 | SDValue Offset = Op.getOperand(i: 1); |
2734 | SDValue Handler = Op.getOperand(i: 2); |
2735 | SDLoc DL(Op); |
2736 | EVT Ty = ABI.IsN64() ? MVT::i64 : MVT::i32; |
2737 | |
2738 | // Store stack offset in V1, store jump target in V0. Glue CopyToReg and |
2739 | // EH_RETURN nodes, so that instructions are emitted back-to-back. |
2740 | unsigned OffsetReg = ABI.IsN64() ? Mips::V1_64 : Mips::V1; |
2741 | unsigned AddrReg = ABI.IsN64() ? Mips::V0_64 : Mips::V0; |
2742 | Chain = DAG.getCopyToReg(Chain, dl: DL, Reg: OffsetReg, N: Offset, Glue: SDValue()); |
2743 | Chain = DAG.getCopyToReg(Chain, dl: DL, Reg: AddrReg, N: Handler, Glue: Chain.getValue(R: 1)); |
2744 | return DAG.getNode(Opcode: MipsISD::EH_RETURN, DL, VT: MVT::Other, N1: Chain, |
2745 | N2: DAG.getRegister(Reg: OffsetReg, VT: Ty), |
2746 | N3: DAG.getRegister(Reg: AddrReg, VT: getPointerTy(DL: MF.getDataLayout())), |
2747 | N4: Chain.getValue(R: 1)); |
2748 | } |
2749 | |
2750 | SDValue MipsTargetLowering::lowerATOMIC_FENCE(SDValue Op, |
2751 | SelectionDAG &DAG) const { |
2752 | // FIXME: Need pseudo-fence for 'singlethread' fences |
2753 | // FIXME: Set SType for weaker fences where supported/appropriate. |
2754 | unsigned SType = 0; |
2755 | SDLoc DL(Op); |
2756 | return DAG.getNode(Opcode: MipsISD::Sync, DL, VT: MVT::Other, N1: Op.getOperand(i: 0), |
2757 | N2: DAG.getConstant(Val: SType, DL, VT: MVT::i32)); |
2758 | } |
2759 | |
2760 | SDValue MipsTargetLowering::lowerShiftLeftParts(SDValue Op, |
2761 | SelectionDAG &DAG) const { |
2762 | SDLoc DL(Op); |
2763 | MVT VT = Subtarget.isGP64bit() ? MVT::i64 : MVT::i32; |
2764 | |
2765 | SDValue Lo = Op.getOperand(i: 0), Hi = Op.getOperand(i: 1); |
2766 | SDValue Shamt = Op.getOperand(i: 2); |
2767 | // if shamt < (VT.bits): |
2768 | // lo = (shl lo, shamt) |
2769 | // hi = (or (shl hi, shamt) (srl (srl lo, 1), (xor shamt, (VT.bits-1)))) |
2770 | // else: |
2771 | // lo = 0 |
2772 | // hi = (shl lo, shamt[4:0]) |
2773 | SDValue Not = |
2774 | DAG.getNode(Opcode: ISD::XOR, DL, VT: MVT::i32, N1: Shamt, |
2775 | N2: DAG.getConstant(Val: VT.getSizeInBits() - 1, DL, VT: MVT::i32)); |
2776 | SDValue ShiftRight1Lo = DAG.getNode(Opcode: ISD::SRL, DL, VT, N1: Lo, |
2777 | N2: DAG.getConstant(Val: 1, DL, VT)); |
2778 | SDValue ShiftRightLo = DAG.getNode(Opcode: ISD::SRL, DL, VT, N1: ShiftRight1Lo, N2: Not); |
2779 | SDValue ShiftLeftHi = DAG.getNode(Opcode: ISD::SHL, DL, VT, N1: Hi, N2: Shamt); |
2780 | SDValue Or = DAG.getNode(Opcode: ISD::OR, DL, VT, N1: ShiftLeftHi, N2: ShiftRightLo); |
2781 | SDValue ShiftLeftLo = DAG.getNode(Opcode: ISD::SHL, DL, VT, N1: Lo, N2: Shamt); |
2782 | SDValue Cond = DAG.getNode(Opcode: ISD::AND, DL, VT: MVT::i32, N1: Shamt, |
2783 | N2: DAG.getConstant(Val: VT.getSizeInBits(), DL, VT: MVT::i32)); |
2784 | Lo = DAG.getNode(Opcode: ISD::SELECT, DL, VT, N1: Cond, |
2785 | N2: DAG.getConstant(Val: 0, DL, VT), N3: ShiftLeftLo); |
2786 | Hi = DAG.getNode(Opcode: ISD::SELECT, DL, VT, N1: Cond, N2: ShiftLeftLo, N3: Or); |
2787 | |
2788 | SDValue Ops[2] = {Lo, Hi}; |
2789 | return DAG.getMergeValues(Ops, dl: DL); |
2790 | } |
2791 | |
2792 | SDValue MipsTargetLowering::lowerShiftRightParts(SDValue Op, SelectionDAG &DAG, |
2793 | bool IsSRA) const { |
2794 | SDLoc DL(Op); |
2795 | SDValue Lo = Op.getOperand(i: 0), Hi = Op.getOperand(i: 1); |
2796 | SDValue Shamt = Op.getOperand(i: 2); |
2797 | MVT VT = Subtarget.isGP64bit() ? MVT::i64 : MVT::i32; |
2798 | |
2799 | // if shamt < (VT.bits): |
2800 | // lo = (or (shl (shl hi, 1), (xor shamt, (VT.bits-1))) (srl lo, shamt)) |
2801 | // if isSRA: |
2802 | // hi = (sra hi, shamt) |
2803 | // else: |
2804 | // hi = (srl hi, shamt) |
2805 | // else: |
2806 | // if isSRA: |
2807 | // lo = (sra hi, shamt[4:0]) |
2808 | // hi = (sra hi, 31) |
2809 | // else: |
2810 | // lo = (srl hi, shamt[4:0]) |
2811 | // hi = 0 |
2812 | SDValue Not = |
2813 | DAG.getNode(Opcode: ISD::XOR, DL, VT: MVT::i32, N1: Shamt, |
2814 | N2: DAG.getConstant(Val: VT.getSizeInBits() - 1, DL, VT: MVT::i32)); |
2815 | SDValue ShiftLeft1Hi = DAG.getNode(Opcode: ISD::SHL, DL, VT, N1: Hi, |
2816 | N2: DAG.getConstant(Val: 1, DL, VT)); |
2817 | SDValue ShiftLeftHi = DAG.getNode(Opcode: ISD::SHL, DL, VT, N1: ShiftLeft1Hi, N2: Not); |
2818 | SDValue ShiftRightLo = DAG.getNode(Opcode: ISD::SRL, DL, VT, N1: Lo, N2: Shamt); |
2819 | SDValue Or = DAG.getNode(Opcode: ISD::OR, DL, VT, N1: ShiftLeftHi, N2: ShiftRightLo); |
2820 | SDValue ShiftRightHi = DAG.getNode(Opcode: IsSRA ? ISD::SRA : ISD::SRL, |
2821 | DL, VT, N1: Hi, N2: Shamt); |
2822 | SDValue Cond = DAG.getNode(Opcode: ISD::AND, DL, VT: MVT::i32, N1: Shamt, |
2823 | N2: DAG.getConstant(Val: VT.getSizeInBits(), DL, VT: MVT::i32)); |
2824 | SDValue Ext = DAG.getNode(Opcode: ISD::SRA, DL, VT, N1: Hi, |
2825 | N2: DAG.getConstant(Val: VT.getSizeInBits() - 1, DL, VT)); |
2826 | |
2827 | if (!(Subtarget.hasMips4() || Subtarget.hasMips32())) { |
2828 | SDVTList VTList = DAG.getVTList(VT1: VT, VT2: VT); |
2829 | return DAG.getNode(Opcode: Subtarget.isGP64bit() ? MipsISD::DOUBLE_SELECT_I64 |
2830 | : MipsISD::DOUBLE_SELECT_I, |
2831 | DL, VTList, N1: Cond, N2: ShiftRightHi, |
2832 | N3: IsSRA ? Ext : DAG.getConstant(Val: 0, DL, VT), N4: Or, |
2833 | N5: ShiftRightHi); |
2834 | } |
2835 | |
2836 | Lo = DAG.getNode(Opcode: ISD::SELECT, DL, VT, N1: Cond, N2: ShiftRightHi, N3: Or); |
2837 | Hi = DAG.getNode(Opcode: ISD::SELECT, DL, VT, N1: Cond, |
2838 | N2: IsSRA ? Ext : DAG.getConstant(Val: 0, DL, VT), N3: ShiftRightHi); |
2839 | |
2840 | SDValue Ops[2] = {Lo, Hi}; |
2841 | return DAG.getMergeValues(Ops, dl: DL); |
2842 | } |
2843 | |
2844 | static SDValue createLoadLR(unsigned Opc, SelectionDAG &DAG, LoadSDNode *LD, |
2845 | SDValue Chain, SDValue Src, unsigned Offset) { |
2846 | SDValue Ptr = LD->getBasePtr(); |
2847 | EVT VT = LD->getValueType(ResNo: 0), MemVT = LD->getMemoryVT(); |
2848 | EVT BasePtrVT = Ptr.getValueType(); |
2849 | SDLoc DL(LD); |
2850 | SDVTList VTList = DAG.getVTList(VT1: VT, VT2: MVT::Other); |
2851 | |
2852 | if (Offset) |
2853 | Ptr = DAG.getNode(Opcode: ISD::ADD, DL, VT: BasePtrVT, N1: Ptr, |
2854 | N2: DAG.getConstant(Val: Offset, DL, VT: BasePtrVT)); |
2855 | |
2856 | SDValue Ops[] = { Chain, Ptr, Src }; |
2857 | return DAG.getMemIntrinsicNode(Opcode: Opc, dl: DL, VTList, Ops, MemVT, |
2858 | MMO: LD->getMemOperand()); |
2859 | } |
2860 | |
2861 | // Expand an unaligned 32 or 64-bit integer load node. |
2862 | SDValue MipsTargetLowering::lowerLOAD(SDValue Op, SelectionDAG &DAG) const { |
2863 | LoadSDNode *LD = cast<LoadSDNode>(Val&: Op); |
2864 | EVT MemVT = LD->getMemoryVT(); |
2865 | |
2866 | if (Subtarget.systemSupportsUnalignedAccess()) |
2867 | return Op; |
2868 | |
2869 | // Return if load is aligned or if MemVT is neither i32 nor i64. |
2870 | if ((LD->getAlign().value() >= (MemVT.getSizeInBits() / 8)) || |
2871 | ((MemVT != MVT::i32) && (MemVT != MVT::i64))) |
2872 | return SDValue(); |
2873 | |
2874 | bool IsLittle = Subtarget.isLittle(); |
2875 | EVT VT = Op.getValueType(); |
2876 | ISD::LoadExtType ExtType = LD->getExtensionType(); |
2877 | SDValue Chain = LD->getChain(), Undef = DAG.getUNDEF(VT); |
2878 | |
2879 | assert((VT == MVT::i32) || (VT == MVT::i64)); |
2880 | |
2881 | // Expand |
2882 | // (set dst, (i64 (load baseptr))) |
2883 | // to |
2884 | // (set tmp, (ldl (add baseptr, 7), undef)) |
2885 | // (set dst, (ldr baseptr, tmp)) |
2886 | if ((VT == MVT::i64) && (ExtType == ISD::NON_EXTLOAD)) { |
2887 | SDValue LDL = createLoadLR(Opc: MipsISD::LDL, DAG, LD, Chain, Src: Undef, |
2888 | Offset: IsLittle ? 7 : 0); |
2889 | return createLoadLR(Opc: MipsISD::LDR, DAG, LD, Chain: LDL.getValue(R: 1), Src: LDL, |
2890 | Offset: IsLittle ? 0 : 7); |
2891 | } |
2892 | |
2893 | SDValue LWL = createLoadLR(Opc: MipsISD::LWL, DAG, LD, Chain, Src: Undef, |
2894 | Offset: IsLittle ? 3 : 0); |
2895 | SDValue LWR = createLoadLR(Opc: MipsISD::LWR, DAG, LD, Chain: LWL.getValue(R: 1), Src: LWL, |
2896 | Offset: IsLittle ? 0 : 3); |
2897 | |
2898 | // Expand |
2899 | // (set dst, (i32 (load baseptr))) or |
2900 | // (set dst, (i64 (sextload baseptr))) or |
2901 | // (set dst, (i64 (extload baseptr))) |
2902 | // to |
2903 | // (set tmp, (lwl (add baseptr, 3), undef)) |
2904 | // (set dst, (lwr baseptr, tmp)) |
2905 | if ((VT == MVT::i32) || (ExtType == ISD::SEXTLOAD) || |
2906 | (ExtType == ISD::EXTLOAD)) |
2907 | return LWR; |
2908 | |
2909 | assert((VT == MVT::i64) && (ExtType == ISD::ZEXTLOAD)); |
2910 | |
2911 | // Expand |
2912 | // (set dst, (i64 (zextload baseptr))) |
2913 | // to |
2914 | // (set tmp0, (lwl (add baseptr, 3), undef)) |
2915 | // (set tmp1, (lwr baseptr, tmp0)) |
2916 | // (set tmp2, (shl tmp1, 32)) |
2917 | // (set dst, (srl tmp2, 32)) |
2918 | SDLoc DL(LD); |
2919 | SDValue Const32 = DAG.getConstant(Val: 32, DL, VT: MVT::i32); |
2920 | SDValue SLL = DAG.getNode(Opcode: ISD::SHL, DL, VT: MVT::i64, N1: LWR, N2: Const32); |
2921 | SDValue SRL = DAG.getNode(Opcode: ISD::SRL, DL, VT: MVT::i64, N1: SLL, N2: Const32); |
2922 | SDValue Ops[] = { SRL, LWR.getValue(R: 1) }; |
2923 | return DAG.getMergeValues(Ops, dl: DL); |
2924 | } |
2925 | |
2926 | static SDValue createStoreLR(unsigned Opc, SelectionDAG &DAG, StoreSDNode *SD, |
2927 | SDValue Chain, unsigned Offset) { |
2928 | SDValue Ptr = SD->getBasePtr(), Value = SD->getValue(); |
2929 | EVT MemVT = SD->getMemoryVT(), BasePtrVT = Ptr.getValueType(); |
2930 | SDLoc DL(SD); |
2931 | SDVTList VTList = DAG.getVTList(VT: MVT::Other); |
2932 | |
2933 | if (Offset) |
2934 | Ptr = DAG.getNode(Opcode: ISD::ADD, DL, VT: BasePtrVT, N1: Ptr, |
2935 | N2: DAG.getConstant(Val: Offset, DL, VT: BasePtrVT)); |
2936 | |
2937 | SDValue Ops[] = { Chain, Value, Ptr }; |
2938 | return DAG.getMemIntrinsicNode(Opcode: Opc, dl: DL, VTList, Ops, MemVT, |
2939 | MMO: SD->getMemOperand()); |
2940 | } |
2941 | |
2942 | // Expand an unaligned 32 or 64-bit integer store node. |
2943 | static SDValue lowerUnalignedIntStore(StoreSDNode *SD, SelectionDAG &DAG, |
2944 | bool IsLittle) { |
2945 | SDValue Value = SD->getValue(), Chain = SD->getChain(); |
2946 | EVT VT = Value.getValueType(); |
2947 | |
2948 | // Expand |
2949 | // (store val, baseptr) or |
2950 | // (truncstore val, baseptr) |
2951 | // to |
2952 | // (swl val, (add baseptr, 3)) |
2953 | // (swr val, baseptr) |
2954 | if ((VT == MVT::i32) || SD->isTruncatingStore()) { |
2955 | SDValue SWL = createStoreLR(Opc: MipsISD::SWL, DAG, SD, Chain, |
2956 | Offset: IsLittle ? 3 : 0); |
2957 | return createStoreLR(Opc: MipsISD::SWR, DAG, SD, Chain: SWL, Offset: IsLittle ? 0 : 3); |
2958 | } |
2959 | |
2960 | assert(VT == MVT::i64); |
2961 | |
2962 | // Expand |
2963 | // (store val, baseptr) |
2964 | // to |
2965 | // (sdl val, (add baseptr, 7)) |
2966 | // (sdr val, baseptr) |
2967 | SDValue SDL = createStoreLR(Opc: MipsISD::SDL, DAG, SD, Chain, Offset: IsLittle ? 7 : 0); |
2968 | return createStoreLR(Opc: MipsISD::SDR, DAG, SD, Chain: SDL, Offset: IsLittle ? 0 : 7); |
2969 | } |
2970 | |
2971 | // Lower (store (fp_to_sint $fp) $ptr) to (store (TruncIntFP $fp), $ptr). |
2972 | static SDValue lowerFP_TO_SINT_STORE(StoreSDNode *SD, SelectionDAG &DAG, |
2973 | bool SingleFloat) { |
2974 | SDValue Val = SD->getValue(); |
2975 | |
2976 | if (Val.getOpcode() != ISD::FP_TO_SINT || |
2977 | (Val.getValueSizeInBits() > 32 && SingleFloat)) |
2978 | return SDValue(); |
2979 | |
2980 | EVT FPTy = EVT::getFloatingPointVT(BitWidth: Val.getValueSizeInBits()); |
2981 | SDValue Tr = DAG.getNode(Opcode: MipsISD::TruncIntFP, DL: SDLoc(Val), VT: FPTy, |
2982 | Operand: Val.getOperand(i: 0)); |
2983 | return DAG.getStore(Chain: SD->getChain(), dl: SDLoc(SD), Val: Tr, Ptr: SD->getBasePtr(), |
2984 | PtrInfo: SD->getPointerInfo(), Alignment: SD->getAlign(), |
2985 | MMOFlags: SD->getMemOperand()->getFlags()); |
2986 | } |
2987 | |
2988 | SDValue MipsTargetLowering::lowerSTORE(SDValue Op, SelectionDAG &DAG) const { |
2989 | StoreSDNode *SD = cast<StoreSDNode>(Val&: Op); |
2990 | EVT MemVT = SD->getMemoryVT(); |
2991 | |
2992 | // Lower unaligned integer stores. |
2993 | if (!Subtarget.systemSupportsUnalignedAccess() && |
2994 | (SD->getAlign().value() < (MemVT.getSizeInBits() / 8)) && |
2995 | ((MemVT == MVT::i32) || (MemVT == MVT::i64))) |
2996 | return lowerUnalignedIntStore(SD, DAG, IsLittle: Subtarget.isLittle()); |
2997 | |
2998 | return lowerFP_TO_SINT_STORE(SD, DAG, SingleFloat: Subtarget.isSingleFloat()); |
2999 | } |
3000 | |
3001 | SDValue MipsTargetLowering::lowerEH_DWARF_CFA(SDValue Op, |
3002 | SelectionDAG &DAG) const { |
3003 | |
3004 | // Return a fixed StackObject with offset 0 which points to the old stack |
3005 | // pointer. |
3006 | MachineFrameInfo &MFI = DAG.getMachineFunction().getFrameInfo(); |
3007 | EVT ValTy = Op->getValueType(ResNo: 0); |
3008 | int FI = MFI.CreateFixedObject(Size: Op.getValueSizeInBits() / 8, SPOffset: 0, IsImmutable: false); |
3009 | return DAG.getFrameIndex(FI, VT: ValTy); |
3010 | } |
3011 | |
3012 | SDValue MipsTargetLowering::lowerFP_TO_SINT(SDValue Op, |
3013 | SelectionDAG &DAG) const { |
3014 | if (Op.getValueSizeInBits() > 32 && Subtarget.isSingleFloat()) |
3015 | return SDValue(); |
3016 | |
3017 | EVT FPTy = EVT::getFloatingPointVT(BitWidth: Op.getValueSizeInBits()); |
3018 | SDValue Trunc = DAG.getNode(Opcode: MipsISD::TruncIntFP, DL: SDLoc(Op), VT: FPTy, |
3019 | Operand: Op.getOperand(i: 0)); |
3020 | return DAG.getNode(Opcode: ISD::BITCAST, DL: SDLoc(Op), VT: Op.getValueType(), Operand: Trunc); |
3021 | } |
3022 | |
3023 | SDValue MipsTargetLowering::lowerConstantFP(SDValue Op, |
3024 | SelectionDAG &DAG) const { |
3025 | SDLoc DL(Op); |
3026 | EVT VT = Op.getSimpleValueType(); |
3027 | SDNode *N = Op.getNode(); |
3028 | ConstantFPSDNode *CFP = cast<ConstantFPSDNode>(Val: N); |
3029 | |
3030 | if (!CFP->isNaN() || Subtarget.isNaN2008()) { |
3031 | return SDValue(); |
3032 | } |
3033 | |
3034 | APFloat NaNValue = CFP->getValueAPF(); |
3035 | auto &Sem = NaNValue.getSemantics(); |
3036 | |
3037 | // The MSB of the mantissa should be zero for QNaNs in the MIPS legacy NaN |
3038 | // encodings, and one for sNaNs. Check every NaN constants and make sure |
3039 | // they are correctly encoded for legacy encodings. |
3040 | if (!NaNValue.isSignaling()) { |
3041 | APFloat RealQNaN = NaNValue.getSNaN(Sem); |
3042 | return DAG.getConstantFP(Val: RealQNaN, DL, VT); |
3043 | } |
3044 | return SDValue(); |
3045 | } |
3046 | |
3047 | //===----------------------------------------------------------------------===// |
3048 | // Calling Convention Implementation |
3049 | //===----------------------------------------------------------------------===// |
3050 | |
3051 | //===----------------------------------------------------------------------===// |
3052 | // TODO: Implement a generic logic using tblgen that can support this. |
3053 | // Mips O32 ABI rules: |
3054 | // --- |
3055 | // i32 - Passed in A0, A1, A2, A3 and stack |
3056 | // f32 - Only passed in f32 registers if no int reg has been used yet to hold |
3057 | // an argument. Otherwise, passed in A1, A2, A3 and stack. |
3058 | // f64 - Only passed in two aliased f32 registers if no int reg has been used |
3059 | // yet to hold an argument. Otherwise, use A2, A3 and stack. If A1 is |
3060 | // not used, it must be shadowed. If only A3 is available, shadow it and |
3061 | // go to stack. |
3062 | // vXiX - Received as scalarized i32s, passed in A0 - A3 and the stack. |
3063 | // vXf32 - Passed in either a pair of registers {A0, A1}, {A2, A3} or {A0 - A3} |
3064 | // with the remainder spilled to the stack. |
3065 | // vXf64 - Passed in either {A0, A1, A2, A3} or {A2, A3} and in both cases |
3066 | // spilling the remainder to the stack. |
3067 | // |
3068 | // For vararg functions, all arguments are passed in A0, A1, A2, A3 and stack. |
3069 | //===----------------------------------------------------------------------===// |
3070 | |
3071 | static bool CC_MipsO32(unsigned ValNo, MVT ValVT, MVT LocVT, |
3072 | CCValAssign::LocInfo LocInfo, ISD::ArgFlagsTy ArgFlags, |
3073 | CCState &State, ArrayRef<MCPhysReg> F64Regs) { |
3074 | const MipsSubtarget &Subtarget = static_cast<const MipsSubtarget &>( |
3075 | State.getMachineFunction().getSubtarget()); |
3076 | |
3077 | static const MCPhysReg IntRegs[] = { Mips::A0, Mips::A1, Mips::A2, Mips::A3 }; |
3078 | |
3079 | const MipsCCState * MipsState = static_cast<MipsCCState *>(&State); |
3080 | |
3081 | static const MCPhysReg F32Regs[] = { Mips::F12, Mips::F14 }; |
3082 | |
3083 | static const MCPhysReg FloatVectorIntRegs[] = { Mips::A0, Mips::A2 }; |
3084 | |
3085 | // Do not process byval args here. |
3086 | if (ArgFlags.isByVal()) |
3087 | return true; |
3088 | |
3089 | // Promote i8 and i16 |
3090 | if (ArgFlags.isInReg() && !Subtarget.isLittle()) { |
3091 | if (LocVT == MVT::i8 || LocVT == MVT::i16 || LocVT == MVT::i32) { |
3092 | LocVT = MVT::i32; |
3093 | if (ArgFlags.isSExt()) |
3094 | LocInfo = CCValAssign::SExtUpper; |
3095 | else if (ArgFlags.isZExt()) |
3096 | LocInfo = CCValAssign::ZExtUpper; |
3097 | else |
3098 | LocInfo = CCValAssign::AExtUpper; |
3099 | } |
3100 | } |
3101 | |
3102 | // Promote i8 and i16 |
3103 | if (LocVT == MVT::i8 || LocVT == MVT::i16) { |
3104 | LocVT = MVT::i32; |
3105 | if (ArgFlags.isSExt()) |
3106 | LocInfo = CCValAssign::SExt; |
3107 | else if (ArgFlags.isZExt()) |
3108 | LocInfo = CCValAssign::ZExt; |
3109 | else |
3110 | LocInfo = CCValAssign::AExt; |
3111 | } |
3112 | |
3113 | unsigned Reg; |
3114 | |
3115 | // f32 and f64 are allocated in A0, A1, A2, A3 when either of the following |
3116 | // is true: function is vararg, argument is 3rd or higher, there is previous |
3117 | // argument which is not f32 or f64. |
3118 | bool AllocateFloatsInIntReg = State.isVarArg() || ValNo > 1 || |
3119 | State.getFirstUnallocated(Regs: F32Regs) != ValNo; |
3120 | Align OrigAlign = ArgFlags.getNonZeroOrigAlign(); |
3121 | bool isI64 = (ValVT == MVT::i32 && OrigAlign == Align(8)); |
3122 | bool isVectorFloat = MipsState->WasOriginalArgVectorFloat(ValNo); |
3123 | |
3124 | // The MIPS vector ABI for floats passes them in a pair of registers |
3125 | if (ValVT == MVT::i32 && isVectorFloat) { |
3126 | // This is the start of an vector that was scalarized into an unknown number |
3127 | // of components. It doesn't matter how many there are. Allocate one of the |
3128 | // notional 8 byte aligned registers which map onto the argument stack, and |
3129 | // shadow the register lost to alignment requirements. |
3130 | if (ArgFlags.isSplit()) { |
3131 | Reg = State.AllocateReg(Regs: FloatVectorIntRegs); |
3132 | if (Reg == Mips::A2) |
3133 | State.AllocateReg(Reg: Mips::A1); |
3134 | else if (Reg == 0) |
3135 | State.AllocateReg(Reg: Mips::A3); |
3136 | } else { |
3137 | // If we're an intermediate component of the split, we can just attempt to |
3138 | // allocate a register directly. |
3139 | Reg = State.AllocateReg(Regs: IntRegs); |
3140 | } |
3141 | } else if (ValVT == MVT::i32 || |
3142 | (ValVT == MVT::f32 && AllocateFloatsInIntReg)) { |
3143 | Reg = State.AllocateReg(Regs: IntRegs); |
3144 | // If this is the first part of an i64 arg, |
3145 | // the allocated register must be either A0 or A2. |
3146 | if (isI64 && (Reg == Mips::A1 || Reg == Mips::A3)) |
3147 | Reg = State.AllocateReg(Regs: IntRegs); |
3148 | LocVT = MVT::i32; |
3149 | } else if (ValVT == MVT::f64 && AllocateFloatsInIntReg) { |
3150 | // Allocate int register and shadow next int register. If first |
3151 | // available register is Mips::A1 or Mips::A3, shadow it too. |
3152 | Reg = State.AllocateReg(Regs: IntRegs); |
3153 | if (Reg == Mips::A1 || Reg == Mips::A3) |
3154 | Reg = State.AllocateReg(Regs: IntRegs); |
3155 | |
3156 | if (Reg) { |
3157 | LocVT = MVT::i32; |
3158 | |
3159 | State.addLoc( |
3160 | V: CCValAssign::getCustomReg(ValNo, ValVT, Reg, LocVT, HTP: LocInfo)); |
3161 | MCRegister HiReg = State.AllocateReg(Regs: IntRegs); |
3162 | assert(HiReg); |
3163 | State.addLoc( |
3164 | V: CCValAssign::getCustomReg(ValNo, ValVT, Reg: HiReg, LocVT, HTP: LocInfo)); |
3165 | return false; |
3166 | } |
3167 | } else if (ValVT.isFloatingPoint() && !AllocateFloatsInIntReg) { |
3168 | // we are guaranteed to find an available float register |
3169 | if (ValVT == MVT::f32) { |
3170 | Reg = State.AllocateReg(Regs: F32Regs); |
3171 | // Shadow int register |
3172 | State.AllocateReg(Regs: IntRegs); |
3173 | } else { |
3174 | Reg = State.AllocateReg(Regs: F64Regs); |
3175 | // Shadow int registers |
3176 | MCRegister Reg2 = State.AllocateReg(Regs: IntRegs); |
3177 | if (Reg2 == Mips::A1 || Reg2 == Mips::A3) |
3178 | State.AllocateReg(Regs: IntRegs); |
3179 | State.AllocateReg(Regs: IntRegs); |
3180 | } |
3181 | } else |
3182 | llvm_unreachable("Cannot handle this ValVT." ); |
3183 | |
3184 | if (!Reg) { |
3185 | unsigned Offset = State.AllocateStack(Size: ValVT.getStoreSize(), Alignment: OrigAlign); |
3186 | State.addLoc(V: CCValAssign::getMem(ValNo, ValVT, Offset, LocVT, HTP: LocInfo)); |
3187 | } else |
3188 | State.addLoc(V: CCValAssign::getReg(ValNo, ValVT, Reg, LocVT, HTP: LocInfo)); |
3189 | |
3190 | return false; |
3191 | } |
3192 | |
3193 | static bool CC_MipsO32_FP32(unsigned ValNo, MVT ValVT, |
3194 | MVT LocVT, CCValAssign::LocInfo LocInfo, |
3195 | ISD::ArgFlagsTy ArgFlags, CCState &State) { |
3196 | static const MCPhysReg F64Regs[] = { Mips::D6, Mips::D7 }; |
3197 | |
3198 | return CC_MipsO32(ValNo, ValVT, LocVT, LocInfo, ArgFlags, State, F64Regs); |
3199 | } |
3200 | |
3201 | static bool CC_MipsO32_FP64(unsigned ValNo, MVT ValVT, |
3202 | MVT LocVT, CCValAssign::LocInfo LocInfo, |
3203 | ISD::ArgFlagsTy ArgFlags, CCState &State) { |
3204 | static const MCPhysReg F64Regs[] = { Mips::D12_64, Mips::D14_64 }; |
3205 | |
3206 | return CC_MipsO32(ValNo, ValVT, LocVT, LocInfo, ArgFlags, State, F64Regs); |
3207 | } |
3208 | |
3209 | static bool CC_MipsO32(unsigned ValNo, MVT ValVT, MVT LocVT, |
3210 | CCValAssign::LocInfo LocInfo, ISD::ArgFlagsTy ArgFlags, |
3211 | CCState &State) LLVM_ATTRIBUTE_UNUSED; |
3212 | |
3213 | #include "MipsGenCallingConv.inc" |
3214 | |
3215 | CCAssignFn *MipsTargetLowering::CCAssignFnForCall() const{ |
3216 | return CC_Mips_FixedArg; |
3217 | } |
3218 | |
3219 | CCAssignFn *MipsTargetLowering::CCAssignFnForReturn() const{ |
3220 | return RetCC_Mips; |
3221 | } |
3222 | //===----------------------------------------------------------------------===// |
3223 | // Call Calling Convention Implementation |
3224 | //===----------------------------------------------------------------------===// |
3225 | |
3226 | SDValue MipsTargetLowering::passArgOnStack(SDValue StackPtr, unsigned Offset, |
3227 | SDValue Chain, SDValue Arg, |
3228 | const SDLoc &DL, bool IsTailCall, |
3229 | SelectionDAG &DAG) const { |
3230 | if (!IsTailCall) { |
3231 | SDValue PtrOff = |
3232 | DAG.getNode(Opcode: ISD::ADD, DL, VT: getPointerTy(DL: DAG.getDataLayout()), N1: StackPtr, |
3233 | N2: DAG.getIntPtrConstant(Val: Offset, DL)); |
3234 | return DAG.getStore(Chain, dl: DL, Val: Arg, Ptr: PtrOff, PtrInfo: MachinePointerInfo()); |
3235 | } |
3236 | |
3237 | MachineFrameInfo &MFI = DAG.getMachineFunction().getFrameInfo(); |
3238 | int FI = MFI.CreateFixedObject(Size: Arg.getValueSizeInBits() / 8, SPOffset: Offset, IsImmutable: false); |
3239 | SDValue FIN = DAG.getFrameIndex(FI, VT: getPointerTy(DL: DAG.getDataLayout())); |
3240 | return DAG.getStore(Chain, dl: DL, Val: Arg, Ptr: FIN, PtrInfo: MachinePointerInfo(), Alignment: MaybeAlign(), |
3241 | MMOFlags: MachineMemOperand::MOVolatile); |
3242 | } |
3243 | |
3244 | void MipsTargetLowering:: |
3245 | getOpndList(SmallVectorImpl<SDValue> &Ops, |
3246 | std::deque<std::pair<unsigned, SDValue>> &RegsToPass, |
3247 | bool IsPICCall, bool GlobalOrExternal, bool InternalLinkage, |
3248 | bool IsCallReloc, CallLoweringInfo &CLI, SDValue Callee, |
3249 | SDValue Chain) const { |
3250 | // Insert node "GP copy globalreg" before call to function. |
3251 | // |
3252 | // R_MIPS_CALL* operators (emitted when non-internal functions are called |
3253 | // in PIC mode) allow symbols to be resolved via lazy binding. |
3254 | // The lazy binding stub requires GP to point to the GOT. |
3255 | // Note that we don't need GP to point to the GOT for indirect calls |
3256 | // (when R_MIPS_CALL* is not used for the call) because Mips linker generates |
3257 | // lazy binding stub for a function only when R_MIPS_CALL* are the only relocs |
3258 | // used for the function (that is, Mips linker doesn't generate lazy binding |
3259 | // stub for a function whose address is taken in the program). |
3260 | if (IsPICCall && !InternalLinkage && IsCallReloc) { |
3261 | unsigned GPReg = ABI.IsN64() ? Mips::GP_64 : Mips::GP; |
3262 | EVT Ty = ABI.IsN64() ? MVT::i64 : MVT::i32; |
3263 | RegsToPass.push_back(x: std::make_pair(x&: GPReg, y: getGlobalReg(DAG&: CLI.DAG, Ty))); |
3264 | } |
3265 | |
3266 | // Build a sequence of copy-to-reg nodes chained together with token |
3267 | // chain and flag operands which copy the outgoing args into registers. |
3268 | // The InGlue in necessary since all emitted instructions must be |
3269 | // stuck together. |
3270 | SDValue InGlue; |
3271 | |
3272 | for (auto &R : RegsToPass) { |
3273 | Chain = CLI.DAG.getCopyToReg(Chain, dl: CLI.DL, Reg: R.first, N: R.second, Glue: InGlue); |
3274 | InGlue = Chain.getValue(R: 1); |
3275 | } |
3276 | |
3277 | // Add argument registers to the end of the list so that they are |
3278 | // known live into the call. |
3279 | for (auto &R : RegsToPass) |
3280 | Ops.push_back(Elt: CLI.DAG.getRegister(Reg: R.first, VT: R.second.getValueType())); |
3281 | |
3282 | // Add a register mask operand representing the call-preserved registers. |
3283 | const TargetRegisterInfo *TRI = Subtarget.getRegisterInfo(); |
3284 | const uint32_t *Mask = |
3285 | TRI->getCallPreservedMask(MF: CLI.DAG.getMachineFunction(), CLI.CallConv); |
3286 | assert(Mask && "Missing call preserved mask for calling convention" ); |
3287 | if (Subtarget.inMips16HardFloat()) { |
3288 | if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Val&: CLI.Callee)) { |
3289 | StringRef Sym = G->getGlobal()->getName(); |
3290 | Function *F = G->getGlobal()->getParent()->getFunction(Name: Sym); |
3291 | if (F && F->hasFnAttribute(Kind: "__Mips16RetHelper" )) { |
3292 | Mask = MipsRegisterInfo::getMips16RetHelperMask(); |
3293 | } |
3294 | } |
3295 | } |
3296 | Ops.push_back(Elt: CLI.DAG.getRegisterMask(RegMask: Mask)); |
3297 | |
3298 | if (InGlue.getNode()) |
3299 | Ops.push_back(Elt: InGlue); |
3300 | } |
3301 | |
3302 | void MipsTargetLowering::AdjustInstrPostInstrSelection(MachineInstr &MI, |
3303 | SDNode *Node) const { |
3304 | switch (MI.getOpcode()) { |
3305 | default: |
3306 | return; |
3307 | case Mips::JALR: |
3308 | case Mips::JALRPseudo: |
3309 | case Mips::JALR64: |
3310 | case Mips::JALR64Pseudo: |
3311 | case Mips::JALR16_MM: |
3312 | case Mips::JALRC16_MMR6: |
3313 | case Mips::TAILCALLREG: |
3314 | case Mips::TAILCALLREG64: |
3315 | case Mips::TAILCALLR6REG: |
3316 | case Mips::TAILCALL64R6REG: |
3317 | case Mips::TAILCALLREG_MM: |
3318 | case Mips::TAILCALLREG_MMR6: { |
3319 | if (!EmitJalrReloc || |
3320 | Subtarget.inMips16Mode() || |
3321 | !isPositionIndependent() || |
3322 | Node->getNumOperands() < 1 || |
3323 | Node->getOperand(Num: 0).getNumOperands() < 2) { |
3324 | return; |
3325 | } |
3326 | // We are after the callee address, set by LowerCall(). |
3327 | // If added to MI, asm printer will emit .reloc R_MIPS_JALR for the |
3328 | // symbol. |
3329 | const SDValue TargetAddr = Node->getOperand(Num: 0).getOperand(i: 1); |
3330 | StringRef Sym; |
3331 | if (const GlobalAddressSDNode *G = |
3332 | dyn_cast_or_null<const GlobalAddressSDNode>(Val: TargetAddr)) { |
3333 | // We must not emit the R_MIPS_JALR relocation against data symbols |
3334 | // since this will cause run-time crashes if the linker replaces the |
3335 | // call instruction with a relative branch to the data symbol. |
3336 | if (!isa<Function>(Val: G->getGlobal())) { |
3337 | LLVM_DEBUG(dbgs() << "Not adding R_MIPS_JALR against data symbol " |
3338 | << G->getGlobal()->getName() << "\n" ); |
3339 | return; |
3340 | } |
3341 | Sym = G->getGlobal()->getName(); |
3342 | } |
3343 | else if (const ExternalSymbolSDNode *ES = |
3344 | dyn_cast_or_null<const ExternalSymbolSDNode>(Val: TargetAddr)) { |
3345 | Sym = ES->getSymbol(); |
3346 | } |
3347 | |
3348 | if (Sym.empty()) |
3349 | return; |
3350 | |
3351 | MachineFunction *MF = MI.getParent()->getParent(); |
3352 | MCSymbol *S = MF->getContext().getOrCreateSymbol(Name: Sym); |
3353 | LLVM_DEBUG(dbgs() << "Adding R_MIPS_JALR against " << Sym << "\n" ); |
3354 | MI.addOperand(Op: MachineOperand::CreateMCSymbol(Sym: S, TargetFlags: MipsII::MO_JALR)); |
3355 | } |
3356 | } |
3357 | } |
3358 | |
3359 | /// LowerCall - functions arguments are copied from virtual regs to |
3360 | /// (physical regs)/(stack frame), CALLSEQ_START and CALLSEQ_END are emitted. |
3361 | SDValue |
3362 | MipsTargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI, |
3363 | SmallVectorImpl<SDValue> &InVals) const { |
3364 | SelectionDAG &DAG = CLI.DAG; |
3365 | SDLoc DL = CLI.DL; |
3366 | SmallVectorImpl<ISD::OutputArg> &Outs = CLI.Outs; |
3367 | SmallVectorImpl<SDValue> &OutVals = CLI.OutVals; |
3368 | SmallVectorImpl<ISD::InputArg> &Ins = CLI.Ins; |
3369 | SDValue Chain = CLI.Chain; |
3370 | SDValue Callee = CLI.Callee; |
3371 | bool &IsTailCall = CLI.IsTailCall; |
3372 | CallingConv::ID CallConv = CLI.CallConv; |
3373 | bool IsVarArg = CLI.IsVarArg; |
3374 | |
3375 | MachineFunction &MF = DAG.getMachineFunction(); |
3376 | MachineFrameInfo &MFI = MF.getFrameInfo(); |
3377 | const TargetFrameLowering *TFL = Subtarget.getFrameLowering(); |
3378 | MipsFunctionInfo *FuncInfo = MF.getInfo<MipsFunctionInfo>(); |
3379 | bool IsPIC = isPositionIndependent(); |
3380 | |
3381 | // Analyze operands of the call, assigning locations to each operand. |
3382 | SmallVector<CCValAssign, 16> ArgLocs; |
3383 | MipsCCState CCInfo( |
3384 | CallConv, IsVarArg, DAG.getMachineFunction(), ArgLocs, *DAG.getContext(), |
3385 | MipsCCState::getSpecialCallingConvForCallee(Callee: Callee.getNode(), Subtarget)); |
3386 | |
3387 | const ExternalSymbolSDNode *ES = |
3388 | dyn_cast_or_null<const ExternalSymbolSDNode>(Val: Callee.getNode()); |
3389 | |
3390 | // There is one case where CALLSEQ_START..CALLSEQ_END can be nested, which |
3391 | // is during the lowering of a call with a byval argument which produces |
3392 | // a call to memcpy. For the O32 case, this causes the caller to allocate |
3393 | // stack space for the reserved argument area for the callee, then recursively |
3394 | // again for the memcpy call. In the NEWABI case, this doesn't occur as those |
3395 | // ABIs mandate that the callee allocates the reserved argument area. We do |
3396 | // still produce nested CALLSEQ_START..CALLSEQ_END with zero space though. |
3397 | // |
3398 | // If the callee has a byval argument and memcpy is used, we are mandated |
3399 | // to already have produced a reserved argument area for the callee for O32. |
3400 | // Therefore, the reserved argument area can be reused for both calls. |
3401 | // |
3402 | // Other cases of calling memcpy cannot have a chain with a CALLSEQ_START |
3403 | // present, as we have yet to hook that node onto the chain. |
3404 | // |
3405 | // Hence, the CALLSEQ_START and CALLSEQ_END nodes can be eliminated in this |
3406 | // case. GCC does a similar trick, in that wherever possible, it calculates |
3407 | // the maximum out going argument area (including the reserved area), and |
3408 | // preallocates the stack space on entrance to the caller. |
3409 | // |
3410 | // FIXME: We should do the same for efficiency and space. |
3411 | |
3412 | // Note: The check on the calling convention below must match |
3413 | // MipsABIInfo::GetCalleeAllocdArgSizeInBytes(). |
3414 | bool MemcpyInByVal = ES && StringRef(ES->getSymbol()) == "memcpy" && |
3415 | CallConv != CallingConv::Fast && |
3416 | Chain.getOpcode() == ISD::CALLSEQ_START; |
3417 | |
3418 | // Allocate the reserved argument area. It seems strange to do this from the |
3419 | // caller side but removing it breaks the frame size calculation. |
3420 | unsigned ReservedArgArea = |
3421 | MemcpyInByVal ? 0 : ABI.GetCalleeAllocdArgSizeInBytes(CC: CallConv); |
3422 | CCInfo.AllocateStack(Size: ReservedArgArea, Alignment: Align(1)); |
3423 | |
3424 | CCInfo.AnalyzeCallOperands(Outs, Fn: CC_Mips, FuncArgs&: CLI.getArgs(), |
3425 | Func: ES ? ES->getSymbol() : nullptr); |
3426 | |
3427 | // Get a count of how many bytes are to be pushed on the stack. |
3428 | unsigned StackSize = CCInfo.getStackSize(); |
3429 | |
3430 | // Call site info for function parameters tracking. |
3431 | MachineFunction::CallSiteInfo CSInfo; |
3432 | |
3433 | // Check if it's really possible to do a tail call. Restrict it to functions |
3434 | // that are part of this compilation unit. |
3435 | bool InternalLinkage = false; |
3436 | if (IsTailCall) { |
3437 | IsTailCall = isEligibleForTailCallOptimization( |
3438 | CCInfo, NextStackOffset: StackSize, FI: *MF.getInfo<MipsFunctionInfo>()); |
3439 | if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Val&: Callee)) { |
3440 | InternalLinkage = G->getGlobal()->hasInternalLinkage(); |
3441 | IsTailCall &= (InternalLinkage || G->getGlobal()->hasLocalLinkage() || |
3442 | G->getGlobal()->hasPrivateLinkage() || |
3443 | G->getGlobal()->hasHiddenVisibility() || |
3444 | G->getGlobal()->hasProtectedVisibility()); |
3445 | } |
3446 | } |
3447 | if (!IsTailCall && CLI.CB && CLI.CB->isMustTailCall()) |
3448 | report_fatal_error(reason: "failed to perform tail call elimination on a call " |
3449 | "site marked musttail" ); |
3450 | |
3451 | if (IsTailCall) |
3452 | ++NumTailCalls; |
3453 | |
3454 | // Chain is the output chain of the last Load/Store or CopyToReg node. |
3455 | // ByValChain is the output chain of the last Memcpy node created for copying |
3456 | // byval arguments to the stack. |
3457 | unsigned StackAlignment = TFL->getStackAlignment(); |
3458 | StackSize = alignTo(Value: StackSize, Align: StackAlignment); |
3459 | |
3460 | if (!(IsTailCall || MemcpyInByVal)) |
3461 | Chain = DAG.getCALLSEQ_START(Chain, InSize: StackSize, OutSize: 0, DL); |
3462 | |
3463 | SDValue StackPtr = |
3464 | DAG.getCopyFromReg(Chain, dl: DL, Reg: ABI.IsN64() ? Mips::SP_64 : Mips::SP, |
3465 | VT: getPointerTy(DL: DAG.getDataLayout())); |
3466 | std::deque<std::pair<unsigned, SDValue>> RegsToPass; |
3467 | SmallVector<SDValue, 8> MemOpChains; |
3468 | |
3469 | CCInfo.rewindByValRegsInfo(); |
3470 | |
3471 | // Walk the register/memloc assignments, inserting copies/loads. |
3472 | for (unsigned i = 0, e = ArgLocs.size(), OutIdx = 0; i != e; ++i, ++OutIdx) { |
3473 | SDValue Arg = OutVals[OutIdx]; |
3474 | CCValAssign &VA = ArgLocs[i]; |
3475 | MVT ValVT = VA.getValVT(), LocVT = VA.getLocVT(); |
3476 | ISD::ArgFlagsTy Flags = Outs[OutIdx].Flags; |
3477 | bool UseUpperBits = false; |
3478 | |
3479 | // ByVal Arg. |
3480 | if (Flags.isByVal()) { |
3481 | unsigned FirstByValReg, LastByValReg; |
3482 | unsigned ByValIdx = CCInfo.getInRegsParamsProcessed(); |
3483 | CCInfo.getInRegsParamInfo(InRegsParamRecordIndex: ByValIdx, BeginReg&: FirstByValReg, EndReg&: LastByValReg); |
3484 | |
3485 | assert(Flags.getByValSize() && |
3486 | "ByVal args of size 0 should have been ignored by front-end." ); |
3487 | assert(ByValIdx < CCInfo.getInRegsParamsCount()); |
3488 | assert(!IsTailCall && |
3489 | "Do not tail-call optimize if there is a byval argument." ); |
3490 | passByValArg(Chain, DL, RegsToPass, MemOpChains, StackPtr, MFI, DAG, Arg, |
3491 | FirstReg: FirstByValReg, LastReg: LastByValReg, Flags, isLittle: Subtarget.isLittle(), |
3492 | VA); |
3493 | CCInfo.nextInRegsParam(); |
3494 | continue; |
3495 | } |
3496 | |
3497 | // Promote the value if needed. |
3498 | switch (VA.getLocInfo()) { |
3499 | default: |
3500 | llvm_unreachable("Unknown loc info!" ); |
3501 | case CCValAssign::Full: |
3502 | if (VA.isRegLoc()) { |
3503 | if ((ValVT == MVT::f32 && LocVT == MVT::i32) || |
3504 | (ValVT == MVT::f64 && LocVT == MVT::i64) || |
3505 | (ValVT == MVT::i64 && LocVT == MVT::f64)) |
3506 | Arg = DAG.getNode(Opcode: ISD::BITCAST, DL, VT: LocVT, Operand: Arg); |
3507 | else if (ValVT == MVT::f64 && LocVT == MVT::i32) { |
3508 | SDValue Lo = DAG.getNode(Opcode: MipsISD::ExtractElementF64, DL, VT: MVT::i32, |
3509 | N1: Arg, N2: DAG.getConstant(Val: 0, DL, VT: MVT::i32)); |
3510 | SDValue Hi = DAG.getNode(Opcode: MipsISD::ExtractElementF64, DL, VT: MVT::i32, |
3511 | N1: Arg, N2: DAG.getConstant(Val: 1, DL, VT: MVT::i32)); |
3512 | if (!Subtarget.isLittle()) |
3513 | std::swap(a&: Lo, b&: Hi); |
3514 | |
3515 | assert(VA.needsCustom()); |
3516 | |
3517 | Register LocRegLo = VA.getLocReg(); |
3518 | Register LocRegHigh = ArgLocs[++i].getLocReg(); |
3519 | RegsToPass.push_back(x: std::make_pair(x&: LocRegLo, y&: Lo)); |
3520 | RegsToPass.push_back(x: std::make_pair(x&: LocRegHigh, y&: Hi)); |
3521 | continue; |
3522 | } |
3523 | } |
3524 | break; |
3525 | case CCValAssign::BCvt: |
3526 | Arg = DAG.getNode(Opcode: ISD::BITCAST, DL, VT: LocVT, Operand: Arg); |
3527 | break; |
3528 | case CCValAssign::SExtUpper: |
3529 | UseUpperBits = true; |
3530 | [[fallthrough]]; |
3531 | case CCValAssign::SExt: |
3532 | Arg = DAG.getNode(Opcode: ISD::SIGN_EXTEND, DL, VT: LocVT, Operand: Arg); |
3533 | break; |
3534 | case CCValAssign::ZExtUpper: |
3535 | UseUpperBits = true; |
3536 | [[fallthrough]]; |
3537 | case CCValAssign::ZExt: |
3538 | Arg = DAG.getNode(Opcode: ISD::ZERO_EXTEND, DL, VT: LocVT, Operand: Arg); |
3539 | break; |
3540 | case CCValAssign::AExtUpper: |
3541 | UseUpperBits = true; |
3542 | [[fallthrough]]; |
3543 | case CCValAssign::AExt: |
3544 | Arg = DAG.getNode(Opcode: ISD::ANY_EXTEND, DL, VT: LocVT, Operand: Arg); |
3545 | break; |
3546 | } |
3547 | |
3548 | if (UseUpperBits) { |
3549 | unsigned ValSizeInBits = Outs[OutIdx].ArgVT.getSizeInBits(); |
3550 | unsigned LocSizeInBits = VA.getLocVT().getSizeInBits(); |
3551 | Arg = DAG.getNode( |
3552 | Opcode: ISD::SHL, DL, VT: VA.getLocVT(), N1: Arg, |
3553 | N2: DAG.getConstant(Val: LocSizeInBits - ValSizeInBits, DL, VT: VA.getLocVT())); |
3554 | } |
3555 | |
3556 | // Arguments that can be passed on register must be kept at |
3557 | // RegsToPass vector |
3558 | if (VA.isRegLoc()) { |
3559 | RegsToPass.push_back(x: std::make_pair(x: VA.getLocReg(), y&: Arg)); |
3560 | |
3561 | // If the parameter is passed through reg $D, which splits into |
3562 | // two physical registers, avoid creating call site info. |
3563 | if (Mips::AFGR64RegClass.contains(Reg: VA.getLocReg())) |
3564 | continue; |
3565 | |
3566 | // Collect CSInfo about which register passes which parameter. |
3567 | const TargetOptions &Options = DAG.getTarget().Options; |
3568 | if (Options.EmitCallSiteInfo) |
3569 | CSInfo.ArgRegPairs.emplace_back(Args: VA.getLocReg(), Args&: i); |
3570 | |
3571 | continue; |
3572 | } |
3573 | |
3574 | // Register can't get to this point... |
3575 | assert(VA.isMemLoc()); |
3576 | |
3577 | // emit ISD::STORE whichs stores the |
3578 | // parameter value to a stack Location |
3579 | MemOpChains.push_back(Elt: passArgOnStack(StackPtr, Offset: VA.getLocMemOffset(), |
3580 | Chain, Arg, DL, IsTailCall, DAG)); |
3581 | } |
3582 | |
3583 | // Transform all store nodes into one single node because all store |
3584 | // nodes are independent of each other. |
3585 | if (!MemOpChains.empty()) |
3586 | Chain = DAG.getNode(Opcode: ISD::TokenFactor, DL, VT: MVT::Other, Ops: MemOpChains); |
3587 | |
3588 | // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every |
3589 | // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol |
3590 | // node so that legalize doesn't hack it. |
3591 | |
3592 | EVT Ty = Callee.getValueType(); |
3593 | bool GlobalOrExternal = false, IsCallReloc = false; |
3594 | |
3595 | // The long-calls feature is ignored in case of PIC. |
3596 | // While we do not support -mshared / -mno-shared properly, |
3597 | // ignore long-calls in case of -mabicalls too. |
3598 | if (!Subtarget.isABICalls() && !IsPIC) { |
3599 | // If the function should be called using "long call", |
3600 | // get its address into a register to prevent using |
3601 | // of the `jal` instruction for the direct call. |
3602 | if (auto *N = dyn_cast<ExternalSymbolSDNode>(Val&: Callee)) { |
3603 | if (Subtarget.useLongCalls()) |
3604 | Callee = Subtarget.hasSym32() |
3605 | ? getAddrNonPIC(N, DL: SDLoc(N), Ty, DAG) |
3606 | : getAddrNonPICSym64(N, DL: SDLoc(N), Ty, DAG); |
3607 | } else if (auto *N = dyn_cast<GlobalAddressSDNode>(Val&: Callee)) { |
3608 | bool UseLongCalls = Subtarget.useLongCalls(); |
3609 | // If the function has long-call/far/near attribute |
3610 | // it overrides command line switch pased to the backend. |
3611 | if (auto *F = dyn_cast<Function>(Val: N->getGlobal())) { |
3612 | if (F->hasFnAttribute(Kind: "long-call" )) |
3613 | UseLongCalls = true; |
3614 | else if (F->hasFnAttribute(Kind: "short-call" )) |
3615 | UseLongCalls = false; |
3616 | } |
3617 | if (UseLongCalls) |
3618 | Callee = Subtarget.hasSym32() |
3619 | ? getAddrNonPIC(N, DL: SDLoc(N), Ty, DAG) |
3620 | : getAddrNonPICSym64(N, DL: SDLoc(N), Ty, DAG); |
3621 | } |
3622 | } |
3623 | |
3624 | if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Val&: Callee)) { |
3625 | if (Subtarget.isTargetCOFF() && |
3626 | G->getGlobal()->hasDLLImportStorageClass()) { |
3627 | assert(Subtarget.isTargetWindows() && |
3628 | "Windows is the only supported COFF target" ); |
3629 | auto PtrInfo = MachinePointerInfo(); |
3630 | Callee = DAG.getLoad(VT: Ty, dl: DL, Chain, |
3631 | Ptr: getDllimportSymbol(N: G, DL: SDLoc(G), Ty, DAG), PtrInfo); |
3632 | } else if (IsPIC) { |
3633 | const GlobalValue *Val = G->getGlobal(); |
3634 | InternalLinkage = Val->hasInternalLinkage(); |
3635 | |
3636 | if (InternalLinkage) |
3637 | Callee = getAddrLocal(N: G, DL, Ty, DAG, IsN32OrN64: ABI.IsN32() || ABI.IsN64()); |
3638 | else if (Subtarget.useXGOT()) { |
3639 | Callee = getAddrGlobalLargeGOT(N: G, DL, Ty, DAG, HiFlag: MipsII::MO_CALL_HI16, |
3640 | LoFlag: MipsII::MO_CALL_LO16, Chain, |
3641 | PtrInfo: FuncInfo->callPtrInfo(MF, GV: Val)); |
3642 | IsCallReloc = true; |
3643 | } else { |
3644 | Callee = getAddrGlobal(N: G, DL, Ty, DAG, Flag: MipsII::MO_GOT_CALL, Chain, |
3645 | PtrInfo: FuncInfo->callPtrInfo(MF, GV: Val)); |
3646 | IsCallReloc = true; |
3647 | } |
3648 | } else |
3649 | Callee = DAG.getTargetGlobalAddress(GV: G->getGlobal(), DL, |
3650 | VT: getPointerTy(DL: DAG.getDataLayout()), offset: 0, |
3651 | TargetFlags: MipsII::MO_NO_FLAG); |
3652 | GlobalOrExternal = true; |
3653 | } |
3654 | else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Val&: Callee)) { |
3655 | const char *Sym = S->getSymbol(); |
3656 | |
3657 | if (!IsPIC) // static |
3658 | Callee = DAG.getTargetExternalSymbol( |
3659 | Sym, VT: getPointerTy(DL: DAG.getDataLayout()), TargetFlags: MipsII::MO_NO_FLAG); |
3660 | else if (Subtarget.useXGOT()) { |
3661 | Callee = getAddrGlobalLargeGOT(N: S, DL, Ty, DAG, HiFlag: MipsII::MO_CALL_HI16, |
3662 | LoFlag: MipsII::MO_CALL_LO16, Chain, |
3663 | PtrInfo: FuncInfo->callPtrInfo(MF, ES: Sym)); |
3664 | IsCallReloc = true; |
3665 | } else { // PIC |
3666 | Callee = getAddrGlobal(N: S, DL, Ty, DAG, Flag: MipsII::MO_GOT_CALL, Chain, |
3667 | PtrInfo: FuncInfo->callPtrInfo(MF, ES: Sym)); |
3668 | IsCallReloc = true; |
3669 | } |
3670 | |
3671 | GlobalOrExternal = true; |
3672 | } |
3673 | |
3674 | SmallVector<SDValue, 8> Ops(1, Chain); |
3675 | SDVTList NodeTys = DAG.getVTList(VT1: MVT::Other, VT2: MVT::Glue); |
3676 | |
3677 | getOpndList(Ops, RegsToPass, IsPICCall: IsPIC, GlobalOrExternal, InternalLinkage, |
3678 | IsCallReloc, CLI, Callee, Chain); |
3679 | |
3680 | if (IsTailCall) { |
3681 | MF.getFrameInfo().setHasTailCall(); |
3682 | SDValue Ret = DAG.getNode(Opcode: MipsISD::TailCall, DL, VT: MVT::Other, Ops); |
3683 | DAG.addCallSiteInfo(Node: Ret.getNode(), CallInfo: std::move(CSInfo)); |
3684 | return Ret; |
3685 | } |
3686 | |
3687 | Chain = DAG.getNode(Opcode: MipsISD::JmpLink, DL, VTList: NodeTys, Ops); |
3688 | SDValue InGlue = Chain.getValue(R: 1); |
3689 | |
3690 | DAG.addCallSiteInfo(Node: Chain.getNode(), CallInfo: std::move(CSInfo)); |
3691 | |
3692 | // Create the CALLSEQ_END node in the case of where it is not a call to |
3693 | // memcpy. |
3694 | if (!(MemcpyInByVal)) { |
3695 | Chain = DAG.getCALLSEQ_END(Chain, Size1: StackSize, Size2: 0, Glue: InGlue, DL); |
3696 | InGlue = Chain.getValue(R: 1); |
3697 | } |
3698 | |
3699 | // Handle result values, copying them out of physregs into vregs that we |
3700 | // return. |
3701 | return LowerCallResult(Chain, InGlue, CallConv, isVarArg: IsVarArg, Ins, dl: DL, DAG, |
3702 | InVals, CLI); |
3703 | } |
3704 | |
3705 | /// LowerCallResult - Lower the result values of a call into the |
3706 | /// appropriate copies out of appropriate physical registers. |
3707 | SDValue MipsTargetLowering::LowerCallResult( |
3708 | SDValue Chain, SDValue InGlue, CallingConv::ID CallConv, bool IsVarArg, |
3709 | const SmallVectorImpl<ISD::InputArg> &Ins, const SDLoc &DL, |
3710 | SelectionDAG &DAG, SmallVectorImpl<SDValue> &InVals, |
3711 | TargetLowering::CallLoweringInfo &CLI) const { |
3712 | // Assign locations to each value returned by this call. |
3713 | SmallVector<CCValAssign, 16> RVLocs; |
3714 | MipsCCState CCInfo(CallConv, IsVarArg, DAG.getMachineFunction(), RVLocs, |
3715 | *DAG.getContext()); |
3716 | |
3717 | const ExternalSymbolSDNode *ES = |
3718 | dyn_cast_or_null<const ExternalSymbolSDNode>(Val: CLI.Callee.getNode()); |
3719 | CCInfo.AnalyzeCallResult(Ins, Fn: RetCC_Mips, RetTy: CLI.RetTy, |
3720 | Func: ES ? ES->getSymbol() : nullptr); |
3721 | |
3722 | // Copy all of the result registers out of their specified physreg. |
3723 | for (unsigned i = 0; i != RVLocs.size(); ++i) { |
3724 | CCValAssign &VA = RVLocs[i]; |
3725 | assert(VA.isRegLoc() && "Can only return in registers!" ); |
3726 | |
3727 | SDValue Val = DAG.getCopyFromReg(Chain, dl: DL, Reg: RVLocs[i].getLocReg(), |
3728 | VT: RVLocs[i].getLocVT(), Glue: InGlue); |
3729 | Chain = Val.getValue(R: 1); |
3730 | InGlue = Val.getValue(R: 2); |
3731 | |
3732 | if (VA.isUpperBitsInLoc()) { |
3733 | unsigned ValSizeInBits = Ins[i].ArgVT.getSizeInBits(); |
3734 | unsigned LocSizeInBits = VA.getLocVT().getSizeInBits(); |
3735 | unsigned Shift = |
3736 | VA.getLocInfo() == CCValAssign::ZExtUpper ? ISD::SRL : ISD::SRA; |
3737 | Val = DAG.getNode( |
3738 | Opcode: Shift, DL, VT: VA.getLocVT(), N1: Val, |
3739 | N2: DAG.getConstant(Val: LocSizeInBits - ValSizeInBits, DL, VT: VA.getLocVT())); |
3740 | } |
3741 | |
3742 | switch (VA.getLocInfo()) { |
3743 | default: |
3744 | llvm_unreachable("Unknown loc info!" ); |
3745 | case CCValAssign::Full: |
3746 | break; |
3747 | case CCValAssign::BCvt: |
3748 | Val = DAG.getNode(Opcode: ISD::BITCAST, DL, VT: VA.getValVT(), Operand: Val); |
3749 | break; |
3750 | case CCValAssign::AExt: |
3751 | case CCValAssign::AExtUpper: |
3752 | Val = DAG.getNode(Opcode: ISD::TRUNCATE, DL, VT: VA.getValVT(), Operand: Val); |
3753 | break; |
3754 | case CCValAssign::ZExt: |
3755 | case CCValAssign::ZExtUpper: |
3756 | Val = DAG.getNode(Opcode: ISD::AssertZext, DL, VT: VA.getLocVT(), N1: Val, |
3757 | N2: DAG.getValueType(VA.getValVT())); |
3758 | Val = DAG.getNode(Opcode: ISD::TRUNCATE, DL, VT: VA.getValVT(), Operand: Val); |
3759 | break; |
3760 | case CCValAssign::SExt: |
3761 | case CCValAssign::SExtUpper: |
3762 | Val = DAG.getNode(Opcode: ISD::AssertSext, DL, VT: VA.getLocVT(), N1: Val, |
3763 | N2: DAG.getValueType(VA.getValVT())); |
3764 | Val = DAG.getNode(Opcode: ISD::TRUNCATE, DL, VT: VA.getValVT(), Operand: Val); |
3765 | break; |
3766 | } |
3767 | |
3768 | InVals.push_back(Elt: Val); |
3769 | } |
3770 | |
3771 | return Chain; |
3772 | } |
3773 | |
3774 | static SDValue UnpackFromArgumentSlot(SDValue Val, const CCValAssign &VA, |
3775 | EVT ArgVT, const SDLoc &DL, |
3776 | SelectionDAG &DAG) { |
3777 | MVT LocVT = VA.getLocVT(); |
3778 | EVT ValVT = VA.getValVT(); |
3779 | |
3780 | // Shift into the upper bits if necessary. |
3781 | switch (VA.getLocInfo()) { |
3782 | default: |
3783 | break; |
3784 | case CCValAssign::AExtUpper: |
3785 | case CCValAssign::SExtUpper: |
3786 | case CCValAssign::ZExtUpper: { |
3787 | unsigned ValSizeInBits = ArgVT.getSizeInBits(); |
3788 | unsigned LocSizeInBits = VA.getLocVT().getSizeInBits(); |
3789 | unsigned Opcode = |
3790 | VA.getLocInfo() == CCValAssign::ZExtUpper ? ISD::SRL : ISD::SRA; |
3791 | Val = DAG.getNode( |
3792 | Opcode, DL, VT: VA.getLocVT(), N1: Val, |
3793 | N2: DAG.getConstant(Val: LocSizeInBits - ValSizeInBits, DL, VT: VA.getLocVT())); |
3794 | break; |
3795 | } |
3796 | } |
3797 | |
3798 | // If this is an value smaller than the argument slot size (32-bit for O32, |
3799 | // 64-bit for N32/N64), it has been promoted in some way to the argument slot |
3800 | // size. Extract the value and insert any appropriate assertions regarding |
3801 | // sign/zero extension. |
3802 | switch (VA.getLocInfo()) { |
3803 | default: |
3804 | llvm_unreachable("Unknown loc info!" ); |
3805 | case CCValAssign::Full: |
3806 | break; |
3807 | case CCValAssign::AExtUpper: |
3808 | case CCValAssign::AExt: |
3809 | Val = DAG.getNode(Opcode: ISD::TRUNCATE, DL, VT: ValVT, Operand: Val); |
3810 | break; |
3811 | case CCValAssign::SExtUpper: |
3812 | case CCValAssign::SExt: |
3813 | Val = DAG.getNode(Opcode: ISD::AssertSext, DL, VT: LocVT, N1: Val, N2: DAG.getValueType(ValVT)); |
3814 | Val = DAG.getNode(Opcode: ISD::TRUNCATE, DL, VT: ValVT, Operand: Val); |
3815 | break; |
3816 | case CCValAssign::ZExtUpper: |
3817 | case CCValAssign::ZExt: |
3818 | Val = DAG.getNode(Opcode: ISD::AssertZext, DL, VT: LocVT, N1: Val, N2: DAG.getValueType(ValVT)); |
3819 | Val = DAG.getNode(Opcode: ISD::TRUNCATE, DL, VT: ValVT, Operand: Val); |
3820 | break; |
3821 | case CCValAssign::BCvt: |
3822 | Val = DAG.getNode(Opcode: ISD::BITCAST, DL, VT: ValVT, Operand: Val); |
3823 | break; |
3824 | } |
3825 | |
3826 | return Val; |
3827 | } |
3828 | |
3829 | //===----------------------------------------------------------------------===// |
3830 | // Formal Arguments Calling Convention Implementation |
3831 | //===----------------------------------------------------------------------===// |
3832 | /// LowerFormalArguments - transform physical registers into virtual registers |
3833 | /// and generate load operations for arguments places on the stack. |
3834 | SDValue MipsTargetLowering::LowerFormalArguments( |
3835 | SDValue Chain, CallingConv::ID CallConv, bool IsVarArg, |
3836 | const SmallVectorImpl<ISD::InputArg> &Ins, const SDLoc &DL, |
3837 | SelectionDAG &DAG, SmallVectorImpl<SDValue> &InVals) const { |
3838 | MachineFunction &MF = DAG.getMachineFunction(); |
3839 | MachineFrameInfo &MFI = MF.getFrameInfo(); |
3840 | MipsFunctionInfo *MipsFI = MF.getInfo<MipsFunctionInfo>(); |
3841 | |
3842 | MipsFI->setVarArgsFrameIndex(0); |
3843 | |
3844 | // Used with vargs to acumulate store chains. |
3845 | std::vector<SDValue> OutChains; |
3846 | |
3847 | // Assign locations to all of the incoming arguments. |
3848 | SmallVector<CCValAssign, 16> ArgLocs; |
3849 | MipsCCState CCInfo(CallConv, IsVarArg, DAG.getMachineFunction(), ArgLocs, |
3850 | *DAG.getContext()); |
3851 | CCInfo.AllocateStack(Size: ABI.GetCalleeAllocdArgSizeInBytes(CC: CallConv), Alignment: Align(1)); |
3852 | const Function &Func = DAG.getMachineFunction().getFunction(); |
3853 | Function::const_arg_iterator FuncArg = Func.arg_begin(); |
3854 | |
3855 | if (Func.hasFnAttribute(Kind: "interrupt" ) && !Func.arg_empty()) |
3856 | report_fatal_error( |
3857 | reason: "Functions with the interrupt attribute cannot have arguments!" ); |
3858 | |
3859 | CCInfo.AnalyzeFormalArguments(Ins, Fn: CC_Mips_FixedArg); |
3860 | MipsFI->setFormalArgInfo(Size: CCInfo.getStackSize(), |
3861 | HasByval: CCInfo.getInRegsParamsCount() > 0); |
3862 | |
3863 | unsigned CurArgIdx = 0; |
3864 | CCInfo.rewindByValRegsInfo(); |
3865 | |
3866 | for (unsigned i = 0, e = ArgLocs.size(), InsIdx = 0; i != e; ++i, ++InsIdx) { |
3867 | CCValAssign &VA = ArgLocs[i]; |
3868 | if (Ins[InsIdx].isOrigArg()) { |
3869 | std::advance(i&: FuncArg, n: Ins[InsIdx].getOrigArgIndex() - CurArgIdx); |
3870 | CurArgIdx = Ins[InsIdx].getOrigArgIndex(); |
3871 | } |
3872 | EVT ValVT = VA.getValVT(); |
3873 | ISD::ArgFlagsTy Flags = Ins[InsIdx].Flags; |
3874 | bool IsRegLoc = VA.isRegLoc(); |
3875 | |
3876 | if (Flags.isByVal()) { |
3877 | assert(Ins[InsIdx].isOrigArg() && "Byval arguments cannot be implicit" ); |
3878 | unsigned FirstByValReg, LastByValReg; |
3879 | unsigned ByValIdx = CCInfo.getInRegsParamsProcessed(); |
3880 | CCInfo.getInRegsParamInfo(InRegsParamRecordIndex: ByValIdx, BeginReg&: FirstByValReg, EndReg&: LastByValReg); |
3881 | |
3882 | assert(Flags.getByValSize() && |
3883 | "ByVal args of size 0 should have been ignored by front-end." ); |
3884 | assert(ByValIdx < CCInfo.getInRegsParamsCount()); |
3885 | copyByValRegs(Chain, DL, OutChains, DAG, Flags, InVals, FuncArg: &*FuncArg, |
3886 | FirstReg: FirstByValReg, LastReg: LastByValReg, VA, State&: CCInfo); |
3887 | CCInfo.nextInRegsParam(); |
3888 | continue; |
3889 | } |
3890 | |
3891 | // Arguments stored on registers |
3892 | if (IsRegLoc) { |
3893 | MVT RegVT = VA.getLocVT(); |
3894 | Register ArgReg = VA.getLocReg(); |
3895 | const TargetRegisterClass *RC = getRegClassFor(VT: RegVT); |
3896 | |
3897 | // Transform the arguments stored on |
3898 | // physical registers into virtual ones |
3899 | unsigned Reg = addLiveIn(MF&: DAG.getMachineFunction(), PReg: ArgReg, RC); |
3900 | SDValue ArgValue = DAG.getCopyFromReg(Chain, dl: DL, Reg, VT: RegVT); |
3901 | |
3902 | ArgValue = |
3903 | UnpackFromArgumentSlot(Val: ArgValue, VA, ArgVT: Ins[InsIdx].ArgVT, DL, DAG); |
3904 | |
3905 | // Handle floating point arguments passed in integer registers and |
3906 | // long double arguments passed in floating point registers. |
3907 | if ((RegVT == MVT::i32 && ValVT == MVT::f32) || |
3908 | (RegVT == MVT::i64 && ValVT == MVT::f64) || |
3909 | (RegVT == MVT::f64 && ValVT == MVT::i64)) |
3910 | ArgValue = DAG.getNode(Opcode: ISD::BITCAST, DL, VT: ValVT, Operand: ArgValue); |
3911 | else if (ABI.IsO32() && RegVT == MVT::i32 && |
3912 | ValVT == MVT::f64) { |
3913 | assert(VA.needsCustom() && "Expected custom argument for f64 split" ); |
3914 | CCValAssign &NextVA = ArgLocs[++i]; |
3915 | unsigned Reg2 = |
3916 | addLiveIn(MF&: DAG.getMachineFunction(), PReg: NextVA.getLocReg(), RC); |
3917 | SDValue ArgValue2 = DAG.getCopyFromReg(Chain, dl: DL, Reg: Reg2, VT: RegVT); |
3918 | if (!Subtarget.isLittle()) |
3919 | std::swap(a&: ArgValue, b&: ArgValue2); |
3920 | ArgValue = DAG.getNode(Opcode: MipsISD::BuildPairF64, DL, VT: MVT::f64, |
3921 | N1: ArgValue, N2: ArgValue2); |
3922 | } |
3923 | |
3924 | InVals.push_back(Elt: ArgValue); |
3925 | } else { // VA.isRegLoc() |
3926 | MVT LocVT = VA.getLocVT(); |
3927 | |
3928 | assert(!VA.needsCustom() && "unexpected custom memory argument" ); |
3929 | |
3930 | // Only arguments pased on the stack should make it here. |
3931 | assert(VA.isMemLoc()); |
3932 | |
3933 | // The stack pointer offset is relative to the caller stack frame. |
3934 | int FI = MFI.CreateFixedObject(Size: LocVT.getSizeInBits() / 8, |
3935 | SPOffset: VA.getLocMemOffset(), IsImmutable: true); |
3936 | |
3937 | // Create load nodes to retrieve arguments from the stack |
3938 | SDValue FIN = DAG.getFrameIndex(FI, VT: getPointerTy(DL: DAG.getDataLayout())); |
3939 | SDValue ArgValue = DAG.getLoad( |
3940 | VT: LocVT, dl: DL, Chain, Ptr: FIN, |
3941 | PtrInfo: MachinePointerInfo::getFixedStack(MF&: DAG.getMachineFunction(), FI)); |
3942 | OutChains.push_back(x: ArgValue.getValue(R: 1)); |
3943 | |
3944 | ArgValue = |
3945 | UnpackFromArgumentSlot(Val: ArgValue, VA, ArgVT: Ins[InsIdx].ArgVT, DL, DAG); |
3946 | |
3947 | InVals.push_back(Elt: ArgValue); |
3948 | } |
3949 | } |
3950 | |
3951 | for (unsigned i = 0, e = ArgLocs.size(), InsIdx = 0; i != e; ++i, ++InsIdx) { |
3952 | |
3953 | if (ArgLocs[i].needsCustom()) { |
3954 | ++i; |
3955 | continue; |
3956 | } |
3957 | |
3958 | // The mips ABIs for returning structs by value requires that we copy |
3959 | // the sret argument into $v0 for the return. Save the argument into |
3960 | // a virtual register so that we can access it from the return points. |
3961 | if (Ins[InsIdx].Flags.isSRet()) { |
3962 | unsigned Reg = MipsFI->getSRetReturnReg(); |
3963 | if (!Reg) { |
3964 | Reg = MF.getRegInfo().createVirtualRegister( |
3965 | RegClass: getRegClassFor(VT: ABI.IsN64() ? MVT::i64 : MVT::i32)); |
3966 | MipsFI->setSRetReturnReg(Reg); |
3967 | } |
3968 | SDValue Copy = DAG.getCopyToReg(Chain: DAG.getEntryNode(), dl: DL, Reg, N: InVals[i]); |
3969 | Chain = DAG.getNode(Opcode: ISD::TokenFactor, DL, VT: MVT::Other, N1: Copy, N2: Chain); |
3970 | break; |
3971 | } |
3972 | } |
3973 | |
3974 | if (IsVarArg) |
3975 | writeVarArgRegs(OutChains, Chain, DL, DAG, State&: CCInfo); |
3976 | |
3977 | // All stores are grouped in one node to allow the matching between |
3978 | // the size of Ins and InVals. This only happens when on varg functions |
3979 | if (!OutChains.empty()) { |
3980 | OutChains.push_back(x: Chain); |
3981 | Chain = DAG.getNode(Opcode: ISD::TokenFactor, DL, VT: MVT::Other, Ops: OutChains); |
3982 | } |
3983 | |
3984 | return Chain; |
3985 | } |
3986 | |
3987 | //===----------------------------------------------------------------------===// |
3988 | // Return Value Calling Convention Implementation |
3989 | //===----------------------------------------------------------------------===// |
3990 | |
3991 | bool |
3992 | MipsTargetLowering::CanLowerReturn(CallingConv::ID CallConv, |
3993 | MachineFunction &MF, bool IsVarArg, |
3994 | const SmallVectorImpl<ISD::OutputArg> &Outs, |
3995 | LLVMContext &Context, const Type *RetTy) const { |
3996 | SmallVector<CCValAssign, 16> RVLocs; |
3997 | MipsCCState CCInfo(CallConv, IsVarArg, MF, RVLocs, Context); |
3998 | return CCInfo.CheckCallReturn(ArgsFlags: Outs, Fn: RetCC_Mips, RetTy); |
3999 | } |
4000 | |
4001 | bool MipsTargetLowering::shouldSignExtendTypeInLibCall(Type *Ty, |
4002 | bool IsSigned) const { |
4003 | if ((ABI.IsN32() || ABI.IsN64()) && Ty->isIntegerTy(Bitwidth: 32)) |
4004 | return true; |
4005 | |
4006 | return IsSigned; |
4007 | } |
4008 | |
4009 | SDValue |
4010 | MipsTargetLowering::LowerInterruptReturn(SmallVectorImpl<SDValue> &RetOps, |
4011 | const SDLoc &DL, |
4012 | SelectionDAG &DAG) const { |
4013 | MachineFunction &MF = DAG.getMachineFunction(); |
4014 | MipsFunctionInfo *MipsFI = MF.getInfo<MipsFunctionInfo>(); |
4015 | |
4016 | MipsFI->setISR(); |
4017 | |
4018 | return DAG.getNode(Opcode: MipsISD::ERet, DL, VT: MVT::Other, Ops: RetOps); |
4019 | } |
4020 | |
4021 | SDValue |
4022 | MipsTargetLowering::LowerReturn(SDValue Chain, CallingConv::ID CallConv, |
4023 | bool IsVarArg, |
4024 | const SmallVectorImpl<ISD::OutputArg> &Outs, |
4025 | const SmallVectorImpl<SDValue> &OutVals, |
4026 | const SDLoc &DL, SelectionDAG &DAG) const { |
4027 | // CCValAssign - represent the assignment of |
4028 | // the return value to a location |
4029 | SmallVector<CCValAssign, 16> RVLocs; |
4030 | MachineFunction &MF = DAG.getMachineFunction(); |
4031 | |
4032 | // CCState - Info about the registers and stack slot. |
4033 | MipsCCState CCInfo(CallConv, IsVarArg, MF, RVLocs, *DAG.getContext()); |
4034 | |
4035 | // Analyze return values. |
4036 | CCInfo.AnalyzeReturn(Outs, Fn: RetCC_Mips); |
4037 | |
4038 | SDValue Glue; |
4039 | SmallVector<SDValue, 4> RetOps(1, Chain); |
4040 | |
4041 | // Copy the result values into the output registers. |
4042 | for (unsigned i = 0; i != RVLocs.size(); ++i) { |
4043 | SDValue Val = OutVals[i]; |
4044 | CCValAssign &VA = RVLocs[i]; |
4045 | assert(VA.isRegLoc() && "Can only return in registers!" ); |
4046 | bool UseUpperBits = false; |
4047 | |
4048 | switch (VA.getLocInfo()) { |
4049 | default: |
4050 | llvm_unreachable("Unknown loc info!" ); |
4051 | case CCValAssign::Full: |
4052 | break; |
4053 | case CCValAssign::BCvt: |
4054 | Val = DAG.getNode(Opcode: ISD::BITCAST, DL, VT: VA.getLocVT(), Operand: Val); |
4055 | break; |
4056 | case CCValAssign::AExtUpper: |
4057 | UseUpperBits = true; |
4058 | [[fallthrough]]; |
4059 | case CCValAssign::AExt: |
4060 | Val = DAG.getNode(Opcode: ISD::ANY_EXTEND, DL, VT: VA.getLocVT(), Operand: Val); |
4061 | break; |
4062 | case CCValAssign::ZExtUpper: |
4063 | UseUpperBits = true; |
4064 | [[fallthrough]]; |
4065 | case CCValAssign::ZExt: |
4066 | Val = DAG.getNode(Opcode: ISD::ZERO_EXTEND, DL, VT: VA.getLocVT(), Operand: Val); |
4067 | break; |
4068 | case CCValAssign::SExtUpper: |
4069 | UseUpperBits = true; |
4070 | [[fallthrough]]; |
4071 | case CCValAssign::SExt: |
4072 | Val = DAG.getNode(Opcode: ISD::SIGN_EXTEND, DL, VT: VA.getLocVT(), Operand: Val); |
4073 | break; |
4074 | } |
4075 | |
4076 | if (UseUpperBits) { |
4077 | unsigned ValSizeInBits = Outs[i].ArgVT.getSizeInBits(); |
4078 | unsigned LocSizeInBits = VA.getLocVT().getSizeInBits(); |
4079 | Val = DAG.getNode( |
4080 | Opcode: ISD::SHL, DL, VT: VA.getLocVT(), N1: Val, |
4081 | N2: DAG.getConstant(Val: LocSizeInBits - ValSizeInBits, DL, VT: VA.getLocVT())); |
4082 | } |
4083 | |
4084 | Chain = DAG.getCopyToReg(Chain, dl: DL, Reg: VA.getLocReg(), N: Val, Glue); |
4085 | |
4086 | // Guarantee that all emitted copies are stuck together with flags. |
4087 | Glue = Chain.getValue(R: 1); |
4088 | RetOps.push_back(Elt: DAG.getRegister(Reg: VA.getLocReg(), VT: VA.getLocVT())); |
4089 | } |
4090 | |
4091 | // The mips ABIs for returning structs by value requires that we copy |
4092 | // the sret argument into $v0 for the return. We saved the argument into |
4093 | // a virtual register in the entry block, so now we copy the value out |
4094 | // and into $v0. |
4095 | if (MF.getFunction().hasStructRetAttr()) { |
4096 | MipsFunctionInfo *MipsFI = MF.getInfo<MipsFunctionInfo>(); |
4097 | unsigned Reg = MipsFI->getSRetReturnReg(); |
4098 | |
4099 | if (!Reg) |
4100 | llvm_unreachable("sret virtual register not created in the entry block" ); |
4101 | SDValue Val = |
4102 | DAG.getCopyFromReg(Chain, dl: DL, Reg, VT: getPointerTy(DL: DAG.getDataLayout())); |
4103 | unsigned V0 = ABI.IsN64() ? Mips::V0_64 : Mips::V0; |
4104 | |
4105 | Chain = DAG.getCopyToReg(Chain, dl: DL, Reg: V0, N: Val, Glue); |
4106 | Glue = Chain.getValue(R: 1); |
4107 | RetOps.push_back(Elt: DAG.getRegister(Reg: V0, VT: getPointerTy(DL: DAG.getDataLayout()))); |
4108 | } |
4109 | |
4110 | RetOps[0] = Chain; // Update chain. |
4111 | |
4112 | // Add the glue if we have it. |
4113 | if (Glue.getNode()) |
4114 | RetOps.push_back(Elt: Glue); |
4115 | |
4116 | // ISRs must use "eret". |
4117 | if (DAG.getMachineFunction().getFunction().hasFnAttribute(Kind: "interrupt" )) |
4118 | return LowerInterruptReturn(RetOps, DL, DAG); |
4119 | |
4120 | // Standard return on Mips is a "jr $ra" |
4121 | return DAG.getNode(Opcode: MipsISD::Ret, DL, VT: MVT::Other, Ops: RetOps); |
4122 | } |
4123 | |
4124 | //===----------------------------------------------------------------------===// |
4125 | // Mips Inline Assembly Support |
4126 | //===----------------------------------------------------------------------===// |
4127 | |
4128 | /// getConstraintType - Given a constraint letter, return the type of |
4129 | /// constraint it is for this target. |
4130 | MipsTargetLowering::ConstraintType |
4131 | MipsTargetLowering::getConstraintType(StringRef Constraint) const { |
4132 | // Mips specific constraints |
4133 | // GCC config/mips/constraints.md |
4134 | // |
4135 | // 'd' : An address register. Equivalent to r |
4136 | // unless generating MIPS16 code. |
4137 | // 'y' : Equivalent to r; retained for |
4138 | // backwards compatibility. |
4139 | // 'c' : A register suitable for use in an indirect |
4140 | // jump. This will always be $25 for -mabicalls. |
4141 | // 'l' : The lo register. 1 word storage. |
4142 | // 'x' : The hilo register pair. Double word storage. |
4143 | if (Constraint.size() == 1) { |
4144 | switch (Constraint[0]) { |
4145 | default : break; |
4146 | case 'd': |
4147 | case 'y': |
4148 | case 'f': |
4149 | case 'c': |
4150 | case 'l': |
4151 | case 'x': |
4152 | return C_RegisterClass; |
4153 | case 'R': |
4154 | return C_Memory; |
4155 | } |
4156 | } |
4157 | |
4158 | if (Constraint == "ZC" ) |
4159 | return C_Memory; |
4160 | |
4161 | return TargetLowering::getConstraintType(Constraint); |
4162 | } |
4163 | |
4164 | /// Examine constraint type and operand type and determine a weight value. |
4165 | /// This object must already have been set up with the operand type |
4166 | /// and the current alternative constraint selected. |
4167 | TargetLowering::ConstraintWeight |
4168 | MipsTargetLowering::getSingleConstraintMatchWeight( |
4169 | AsmOperandInfo &info, const char *constraint) const { |
4170 | ConstraintWeight weight = CW_Invalid; |
4171 | Value *CallOperandVal = info.CallOperandVal; |
4172 | // If we don't have a value, we can't do a match, |
4173 | // but allow it at the lowest weight. |
4174 | if (!CallOperandVal) |
4175 | return CW_Default; |
4176 | Type *type = CallOperandVal->getType(); |
4177 | // Look at the constraint type. |
4178 | switch (*constraint) { |
4179 | default: |
4180 | weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint); |
4181 | break; |
4182 | case 'd': |
4183 | case 'y': |
4184 | if (type->isIntegerTy()) |
4185 | weight = CW_Register; |
4186 | break; |
4187 | case 'f': // FPU or MSA register |
4188 | if (Subtarget.hasMSA() && type->isVectorTy() && |
4189 | type->getPrimitiveSizeInBits().getFixedValue() == 128) |
4190 | weight = CW_Register; |
4191 | else if (type->isFloatTy()) |
4192 | weight = CW_Register; |
4193 | break; |
4194 | case 'c': // $25 for indirect jumps |
4195 | case 'l': // lo register |
4196 | case 'x': // hilo register pair |
4197 | if (type->isIntegerTy()) |
4198 | weight = CW_SpecificReg; |
4199 | break; |
4200 | case 'I': // signed 16 bit immediate |
4201 | case 'J': // integer zero |
4202 | case 'K': // unsigned 16 bit immediate |
4203 | case 'L': // signed 32 bit immediate where lower 16 bits are 0 |
4204 | case 'N': // immediate in the range of -65535 to -1 (inclusive) |
4205 | case 'O': // signed 15 bit immediate (+- 16383) |
4206 | case 'P': // immediate in the range of 65535 to 1 (inclusive) |
4207 | if (isa<ConstantInt>(Val: CallOperandVal)) |
4208 | weight = CW_Constant; |
4209 | break; |
4210 | case 'R': |
4211 | weight = CW_Memory; |
4212 | break; |
4213 | } |
4214 | return weight; |
4215 | } |
4216 | |
4217 | /// This is a helper function to parse a physical register string and split it |
4218 | /// into non-numeric and numeric parts (Prefix and Reg). The first boolean flag |
4219 | /// that is returned indicates whether parsing was successful. The second flag |
4220 | /// is true if the numeric part exists. |
4221 | static std::pair<bool, bool> parsePhysicalReg(StringRef C, StringRef &Prefix, |
4222 | unsigned long long &Reg) { |
4223 | if (C.front() != '{' || C.back() != '}') |
4224 | return std::make_pair(x: false, y: false); |
4225 | |
4226 | // Search for the first numeric character. |
4227 | StringRef::const_iterator I, B = C.begin() + 1, E = C.end() - 1; |
4228 | I = std::find_if(first: B, last: E, pred: isdigit); |
4229 | |
4230 | Prefix = StringRef(B, I - B); |
4231 | |
4232 | // The second flag is set to false if no numeric characters were found. |
4233 | if (I == E) |
4234 | return std::make_pair(x: true, y: false); |
4235 | |
4236 | // Parse the numeric characters. |
4237 | return std::make_pair(x: !getAsUnsignedInteger(Str: StringRef(I, E - I), Radix: 10, Result&: Reg), |
4238 | y: true); |
4239 | } |
4240 | |
4241 | EVT MipsTargetLowering::getTypeForExtReturn(LLVMContext &Context, EVT VT, |
4242 | ISD::NodeType) const { |
4243 | bool Cond = !Subtarget.isABI_O32() && VT.getSizeInBits() == 32; |
4244 | EVT MinVT = getRegisterType(VT: Cond ? MVT::i64 : MVT::i32); |
4245 | return VT.bitsLT(VT: MinVT) ? MinVT : VT; |
4246 | } |
4247 | |
4248 | std::pair<unsigned, const TargetRegisterClass *> MipsTargetLowering:: |
4249 | parseRegForInlineAsmConstraint(StringRef C, MVT VT) const { |
4250 | const TargetRegisterInfo *TRI = |
4251 | Subtarget.getRegisterInfo(); |
4252 | const TargetRegisterClass *RC; |
4253 | StringRef Prefix; |
4254 | unsigned long long Reg; |
4255 | |
4256 | std::pair<bool, bool> R = parsePhysicalReg(C, Prefix, Reg); |
4257 | |
4258 | if (!R.first) |
4259 | return std::make_pair(x: 0U, y: nullptr); |
4260 | |
4261 | if ((Prefix == "hi" || Prefix == "lo" )) { // Parse hi/lo. |
4262 | // No numeric characters follow "hi" or "lo". |
4263 | if (R.second) |
4264 | return std::make_pair(x: 0U, y: nullptr); |
4265 | |
4266 | RC = TRI->getRegClass(i: Prefix == "hi" ? |
4267 | Mips::HI32RegClassID : Mips::LO32RegClassID); |
4268 | return std::make_pair(x: *(RC->begin()), y&: RC); |
4269 | } else if (Prefix.starts_with(Prefix: "$msa" )) { |
4270 | // Parse $msa(ir|csr|access|save|modify|request|map|unmap) |
4271 | |
4272 | // No numeric characters follow the name. |
4273 | if (R.second) |
4274 | return std::make_pair(x: 0U, y: nullptr); |
4275 | |
4276 | Reg = StringSwitch<unsigned long long>(Prefix) |
4277 | .Case(S: "$msair" , Value: Mips::MSAIR) |
4278 | .Case(S: "$msacsr" , Value: Mips::MSACSR) |
4279 | .Case(S: "$msaaccess" , Value: Mips::MSAAccess) |
4280 | .Case(S: "$msasave" , Value: Mips::MSASave) |
4281 | .Case(S: "$msamodify" , Value: Mips::MSAModify) |
4282 | .Case(S: "$msarequest" , Value: Mips::MSARequest) |
4283 | .Case(S: "$msamap" , Value: Mips::MSAMap) |
4284 | .Case(S: "$msaunmap" , Value: Mips::MSAUnmap) |
4285 | .Default(Value: 0); |
4286 | |
4287 | if (!Reg) |
4288 | return std::make_pair(x: 0U, y: nullptr); |
4289 | |
4290 | RC = TRI->getRegClass(i: Mips::MSACtrlRegClassID); |
4291 | return std::make_pair(x&: Reg, y&: RC); |
4292 | } |
4293 | |
4294 | if (!R.second) |
4295 | return std::make_pair(x: 0U, y: nullptr); |
4296 | |
4297 | if (Prefix == "$f" ) { // Parse $f0-$f31. |
4298 | // If the size of FP registers is 64-bit or Reg is an even number, select |
4299 | // the 64-bit register class. Otherwise, select the 32-bit register class. |
4300 | if (VT == MVT::Other) |
4301 | VT = (Subtarget.isFP64bit() || !(Reg % 2)) ? MVT::f64 : MVT::f32; |
4302 | |
4303 | RC = getRegClassFor(VT); |
4304 | |
4305 | if (RC == &Mips::AFGR64RegClass) { |
4306 | assert(Reg % 2 == 0); |
4307 | Reg >>= 1; |
4308 | } |
4309 | } else if (Prefix == "$fcc" ) // Parse $fcc0-$fcc7. |
4310 | RC = TRI->getRegClass(i: Mips::FCCRegClassID); |
4311 | else if (Prefix == "$w" ) { // Parse $w0-$w31. |
4312 | RC = getRegClassFor(VT: (VT == MVT::Other) ? MVT::v16i8 : VT); |
4313 | } else { // Parse $0-$31. |
4314 | assert(Prefix == "$" ); |
4315 | RC = getRegClassFor(VT: (VT == MVT::Other) ? MVT::i32 : VT); |
4316 | } |
4317 | |
4318 | assert(Reg < RC->getNumRegs()); |
4319 | return std::make_pair(x: *(RC->begin() + Reg), y&: RC); |
4320 | } |
4321 | |
4322 | /// Given a register class constraint, like 'r', if this corresponds directly |
4323 | /// to an LLVM register class, return a register of 0 and the register class |
4324 | /// pointer. |
4325 | std::pair<unsigned, const TargetRegisterClass *> |
4326 | MipsTargetLowering::getRegForInlineAsmConstraint(const TargetRegisterInfo *TRI, |
4327 | StringRef Constraint, |
4328 | MVT VT) const { |
4329 | if (Constraint.size() == 1) { |
4330 | switch (Constraint[0]) { |
4331 | case 'd': // Address register. Same as 'r' unless generating MIPS16 code. |
4332 | case 'y': // Same as 'r'. Exists for compatibility. |
4333 | case 'r': |
4334 | if ((VT == MVT::i32 || VT == MVT::i16 || VT == MVT::i8 || |
4335 | VT == MVT::i1) || |
4336 | (VT == MVT::f32 && Subtarget.useSoftFloat())) { |
4337 | if (Subtarget.inMips16Mode()) |
4338 | return std::make_pair(x: 0U, y: &Mips::CPU16RegsRegClass); |
4339 | return std::make_pair(x: 0U, y: &Mips::GPR32RegClass); |
4340 | } |
4341 | if ((VT == MVT::i64 || (VT == MVT::f64 && Subtarget.useSoftFloat())) && |
4342 | !Subtarget.isGP64bit()) |
4343 | return std::make_pair(x: 0U, y: &Mips::GPR32RegClass); |
4344 | if ((VT == MVT::i64 || (VT == MVT::f64 && Subtarget.useSoftFloat())) && |
4345 | Subtarget.isGP64bit()) |
4346 | return std::make_pair(x: 0U, y: &Mips::GPR64RegClass); |
4347 | // This will generate an error message |
4348 | return std::make_pair(x: 0U, y: nullptr); |
4349 | case 'f': // FPU or MSA register |
4350 | if (VT == MVT::v16i8) |
4351 | return std::make_pair(x: 0U, y: &Mips::MSA128BRegClass); |
4352 | else if (VT == MVT::v8i16 || VT == MVT::v8f16) |
4353 | return std::make_pair(x: 0U, y: &Mips::MSA128HRegClass); |
4354 | else if (VT == MVT::v4i32 || VT == MVT::v4f32) |
4355 | return std::make_pair(x: 0U, y: &Mips::MSA128WRegClass); |
4356 | else if (VT == MVT::v2i64 || VT == MVT::v2f64) |
4357 | return std::make_pair(x: 0U, y: &Mips::MSA128DRegClass); |
4358 | else if (VT == MVT::f32) |
4359 | return std::make_pair(x: 0U, y: &Mips::FGR32RegClass); |
4360 | else if ((VT == MVT::f64) && (!Subtarget.isSingleFloat())) { |
4361 | if (Subtarget.isFP64bit()) |
4362 | return std::make_pair(x: 0U, y: &Mips::FGR64RegClass); |
4363 | return std::make_pair(x: 0U, y: &Mips::AFGR64RegClass); |
4364 | } |
4365 | break; |
4366 | case 'c': // register suitable for indirect jump |
4367 | if (VT == MVT::i32) |
4368 | return std::make_pair(x: (unsigned)Mips::T9, y: &Mips::GPR32RegClass); |
4369 | if (VT == MVT::i64) |
4370 | return std::make_pair(x: (unsigned)Mips::T9_64, y: &Mips::GPR64RegClass); |
4371 | // This will generate an error message |
4372 | return std::make_pair(x: 0U, y: nullptr); |
4373 | case 'l': // use the `lo` register to store values |
4374 | // that are no bigger than a word |
4375 | if (VT == MVT::i32 || VT == MVT::i16 || VT == MVT::i8) |
4376 | return std::make_pair(x: (unsigned)Mips::LO0, y: &Mips::LO32RegClass); |
4377 | return std::make_pair(x: (unsigned)Mips::LO0_64, y: &Mips::LO64RegClass); |
4378 | case 'x': // use the concatenated `hi` and `lo` registers |
4379 | // to store doubleword values |
4380 | // Fixme: Not triggering the use of both hi and low |
4381 | // This will generate an error message |
4382 | return std::make_pair(x: 0U, y: nullptr); |
4383 | } |
4384 | } |
4385 | |
4386 | if (!Constraint.empty()) { |
4387 | std::pair<unsigned, const TargetRegisterClass *> R; |
4388 | R = parseRegForInlineAsmConstraint(C: Constraint, VT); |
4389 | |
4390 | if (R.second) |
4391 | return R; |
4392 | } |
4393 | |
4394 | return TargetLowering::getRegForInlineAsmConstraint(TRI, Constraint, VT); |
4395 | } |
4396 | |
4397 | /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops |
4398 | /// vector. If it is invalid, don't add anything to Ops. |
4399 | void MipsTargetLowering::LowerAsmOperandForConstraint(SDValue Op, |
4400 | StringRef Constraint, |
4401 | std::vector<SDValue> &Ops, |
4402 | SelectionDAG &DAG) const { |
4403 | SDLoc DL(Op); |
4404 | SDValue Result; |
4405 | |
4406 | // Only support length 1 constraints for now. |
4407 | if (Constraint.size() > 1) |
4408 | return; |
4409 | |
4410 | char ConstraintLetter = Constraint[0]; |
4411 | switch (ConstraintLetter) { |
4412 | default: break; // This will fall through to the generic implementation |
4413 | case 'I': // Signed 16 bit constant |
4414 | // If this fails, the parent routine will give an error |
4415 | if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Val&: Op)) { |
4416 | EVT Type = Op.getValueType(); |
4417 | int64_t Val = C->getSExtValue(); |
4418 | if (isInt<16>(x: Val)) { |
4419 | Result = DAG.getSignedTargetConstant(Val, DL, VT: Type); |
4420 | break; |
4421 | } |
4422 | } |
4423 | return; |
4424 | case 'J': // integer zero |
4425 | if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Val&: Op)) { |
4426 | EVT Type = Op.getValueType(); |
4427 | int64_t Val = C->getZExtValue(); |
4428 | if (Val == 0) { |
4429 | Result = DAG.getTargetConstant(Val: 0, DL, VT: Type); |
4430 | break; |
4431 | } |
4432 | } |
4433 | return; |
4434 | case 'K': // unsigned 16 bit immediate |
4435 | if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Val&: Op)) { |
4436 | EVT Type = Op.getValueType(); |
4437 | uint64_t Val = (uint64_t)C->getZExtValue(); |
4438 | if (isUInt<16>(x: Val)) { |
4439 | Result = DAG.getTargetConstant(Val, DL, VT: Type); |
4440 | break; |
4441 | } |
4442 | } |
4443 | return; |
4444 | case 'L': // signed 32 bit immediate where lower 16 bits are 0 |
4445 | if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Val&: Op)) { |
4446 | EVT Type = Op.getValueType(); |
4447 | int64_t Val = C->getSExtValue(); |
4448 | if ((isInt<32>(x: Val)) && ((Val & 0xffff) == 0)){ |
4449 | Result = DAG.getSignedTargetConstant(Val, DL, VT: Type); |
4450 | break; |
4451 | } |
4452 | } |
4453 | return; |
4454 | case 'N': // immediate in the range of -65535 to -1 (inclusive) |
4455 | if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Val&: Op)) { |
4456 | EVT Type = Op.getValueType(); |
4457 | int64_t Val = C->getSExtValue(); |
4458 | if ((Val >= -65535) && (Val <= -1)) { |
4459 | Result = DAG.getSignedTargetConstant(Val, DL, VT: Type); |
4460 | break; |
4461 | } |
4462 | } |
4463 | return; |
4464 | case 'O': // signed 15 bit immediate |
4465 | if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Val&: Op)) { |
4466 | EVT Type = Op.getValueType(); |
4467 | int64_t Val = C->getSExtValue(); |
4468 | if ((isInt<15>(x: Val))) { |
4469 | Result = DAG.getSignedTargetConstant(Val, DL, VT: Type); |
4470 | break; |
4471 | } |
4472 | } |
4473 | return; |
4474 | case 'P': // immediate in the range of 1 to 65535 (inclusive) |
4475 | if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Val&: Op)) { |
4476 | EVT Type = Op.getValueType(); |
4477 | int64_t Val = C->getSExtValue(); |
4478 | if ((Val <= 65535) && (Val >= 1)) { |
4479 | Result = DAG.getTargetConstant(Val, DL, VT: Type); |
4480 | break; |
4481 | } |
4482 | } |
4483 | return; |
4484 | } |
4485 | |
4486 | if (Result.getNode()) { |
4487 | Ops.push_back(x: Result); |
4488 | return; |
4489 | } |
4490 | |
4491 | TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG); |
4492 | } |
4493 | |
4494 | bool MipsTargetLowering::isLegalAddressingMode(const DataLayout &DL, |
4495 | const AddrMode &AM, Type *Ty, |
4496 | unsigned AS, |
4497 | Instruction *I) const { |
4498 | // No global is ever allowed as a base. |
4499 | if (AM.BaseGV) |
4500 | return false; |
4501 | |
4502 | switch (AM.Scale) { |
4503 | case 0: // "r+i" or just "i", depending on HasBaseReg. |
4504 | break; |
4505 | case 1: |
4506 | if (!AM.HasBaseReg) // allow "r+i". |
4507 | break; |
4508 | return false; // disallow "r+r" or "r+r+i". |
4509 | default: |
4510 | return false; |
4511 | } |
4512 | |
4513 | return true; |
4514 | } |
4515 | |
4516 | bool |
4517 | MipsTargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const { |
4518 | // The Mips target isn't yet aware of offsets. |
4519 | return false; |
4520 | } |
4521 | |
4522 | EVT MipsTargetLowering::getOptimalMemOpType( |
4523 | const MemOp &Op, const AttributeList &FuncAttributes) const { |
4524 | if (Subtarget.hasMips64()) |
4525 | return MVT::i64; |
4526 | |
4527 | return MVT::i32; |
4528 | } |
4529 | |
4530 | bool MipsTargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT, |
4531 | bool ForCodeSize) const { |
4532 | if (VT != MVT::f32 && VT != MVT::f64) |
4533 | return false; |
4534 | if (Imm.isNegZero()) |
4535 | return false; |
4536 | return Imm.isZero(); |
4537 | } |
4538 | |
4539 | bool MipsTargetLowering::isLegalICmpImmediate(int64_t Imm) const { |
4540 | return isInt<16>(x: Imm); |
4541 | } |
4542 | |
4543 | bool MipsTargetLowering::isLegalAddImmediate(int64_t Imm) const { |
4544 | return isInt<16>(x: Imm); |
4545 | } |
4546 | |
4547 | unsigned MipsTargetLowering::getJumpTableEncoding() const { |
4548 | if (!isPositionIndependent()) |
4549 | return MachineJumpTableInfo::EK_BlockAddress; |
4550 | if (ABI.IsN64()) |
4551 | return MachineJumpTableInfo::EK_GPRel64BlockAddress; |
4552 | return MachineJumpTableInfo::EK_GPRel32BlockAddress; |
4553 | } |
4554 | |
4555 | SDValue MipsTargetLowering::getPICJumpTableRelocBase(SDValue Table, |
4556 | SelectionDAG &DAG) const { |
4557 | if (!isPositionIndependent()) |
4558 | return Table; |
4559 | return DAG.getGLOBAL_OFFSET_TABLE(VT: getPointerTy(DL: DAG.getDataLayout())); |
4560 | } |
4561 | |
4562 | bool MipsTargetLowering::useSoftFloat() const { |
4563 | return Subtarget.useSoftFloat(); |
4564 | } |
4565 | |
4566 | void MipsTargetLowering::copyByValRegs( |
4567 | SDValue Chain, const SDLoc &DL, std::vector<SDValue> &OutChains, |
4568 | SelectionDAG &DAG, const ISD::ArgFlagsTy &Flags, |
4569 | SmallVectorImpl<SDValue> &InVals, const Argument *FuncArg, |
4570 | unsigned FirstReg, unsigned LastReg, const CCValAssign &VA, |
4571 | MipsCCState &State) const { |
4572 | MachineFunction &MF = DAG.getMachineFunction(); |
4573 | MachineFrameInfo &MFI = MF.getFrameInfo(); |
4574 | unsigned GPRSizeInBytes = Subtarget.getGPRSizeInBytes(); |
4575 | unsigned NumRegs = LastReg - FirstReg; |
4576 | unsigned RegAreaSize = NumRegs * GPRSizeInBytes; |
4577 | unsigned FrameObjSize = std::max(a: Flags.getByValSize(), b: RegAreaSize); |
4578 | int FrameObjOffset; |
4579 | ArrayRef<MCPhysReg> ByValArgRegs = ABI.GetByValArgRegs(); |
4580 | |
4581 | if (RegAreaSize) |
4582 | FrameObjOffset = |
4583 | (int)ABI.GetCalleeAllocdArgSizeInBytes(CC: State.getCallingConv()) - |
4584 | (int)((ByValArgRegs.size() - FirstReg) * GPRSizeInBytes); |
4585 | else |
4586 | FrameObjOffset = VA.getLocMemOffset(); |
4587 | |
4588 | // Create frame object. |
4589 | EVT PtrTy = getPointerTy(DL: DAG.getDataLayout()); |
4590 | // Make the fixed object stored to mutable so that the load instructions |
4591 | // referencing it have their memory dependencies added. |
4592 | // Set the frame object as isAliased which clears the underlying objects |
4593 | // vector in ScheduleDAGInstrs::buildSchedGraph() resulting in addition of all |
4594 | // stores as dependencies for loads referencing this fixed object. |
4595 | int FI = MFI.CreateFixedObject(Size: FrameObjSize, SPOffset: FrameObjOffset, IsImmutable: false, isAliased: true); |
4596 | SDValue FIN = DAG.getFrameIndex(FI, VT: PtrTy); |
4597 | InVals.push_back(Elt: FIN); |
4598 | |
4599 | if (!NumRegs) |
4600 | return; |
4601 | |
4602 | // Copy arg registers. |
4603 | MVT RegTy = MVT::getIntegerVT(BitWidth: GPRSizeInBytes * 8); |
4604 | const TargetRegisterClass *RC = getRegClassFor(VT: RegTy); |
4605 | |
4606 | for (unsigned I = 0; I < NumRegs; ++I) { |
4607 | unsigned ArgReg = ByValArgRegs[FirstReg + I]; |
4608 | unsigned VReg = addLiveIn(MF, PReg: ArgReg, RC); |
4609 | unsigned Offset = I * GPRSizeInBytes; |
4610 | SDValue StorePtr = DAG.getNode(Opcode: ISD::ADD, DL, VT: PtrTy, N1: FIN, |
4611 | N2: DAG.getConstant(Val: Offset, DL, VT: PtrTy)); |
4612 | SDValue Store = DAG.getStore(Chain, dl: DL, Val: DAG.getRegister(Reg: VReg, VT: RegTy), |
4613 | Ptr: StorePtr, PtrInfo: MachinePointerInfo(FuncArg, Offset)); |
4614 | OutChains.push_back(x: Store); |
4615 | } |
4616 | } |
4617 | |
4618 | // Copy byVal arg to registers and stack. |
4619 | void MipsTargetLowering::passByValArg( |
4620 | SDValue Chain, const SDLoc &DL, |
4621 | std::deque<std::pair<unsigned, SDValue>> &RegsToPass, |
4622 | SmallVectorImpl<SDValue> &MemOpChains, SDValue StackPtr, |
4623 | MachineFrameInfo &MFI, SelectionDAG &DAG, SDValue Arg, unsigned FirstReg, |
4624 | unsigned LastReg, const ISD::ArgFlagsTy &Flags, bool isLittle, |
4625 | const CCValAssign &VA) const { |
4626 | unsigned ByValSizeInBytes = Flags.getByValSize(); |
4627 | unsigned OffsetInBytes = 0; // From beginning of struct |
4628 | unsigned RegSizeInBytes = Subtarget.getGPRSizeInBytes(); |
4629 | Align Alignment = |
4630 | std::min(a: Flags.getNonZeroByValAlign(), b: Align(RegSizeInBytes)); |
4631 | EVT PtrTy = getPointerTy(DL: DAG.getDataLayout()), |
4632 | RegTy = MVT::getIntegerVT(BitWidth: RegSizeInBytes * 8); |
4633 | unsigned NumRegs = LastReg - FirstReg; |
4634 | |
4635 | if (NumRegs) { |
4636 | ArrayRef<MCPhysReg> ArgRegs = ABI.GetByValArgRegs(); |
4637 | bool LeftoverBytes = (NumRegs * RegSizeInBytes > ByValSizeInBytes); |
4638 | unsigned I = 0; |
4639 | |
4640 | // Copy words to registers. |
4641 | for (; I < NumRegs - LeftoverBytes; ++I, OffsetInBytes += RegSizeInBytes) { |
4642 | SDValue LoadPtr = DAG.getNode(Opcode: ISD::ADD, DL, VT: PtrTy, N1: Arg, |
4643 | N2: DAG.getConstant(Val: OffsetInBytes, DL, VT: PtrTy)); |
4644 | SDValue LoadVal = DAG.getLoad(VT: RegTy, dl: DL, Chain, Ptr: LoadPtr, |
4645 | PtrInfo: MachinePointerInfo(), Alignment); |
4646 | MemOpChains.push_back(Elt: LoadVal.getValue(R: 1)); |
4647 | unsigned ArgReg = ArgRegs[FirstReg + I]; |
4648 | RegsToPass.push_back(x: std::make_pair(x&: ArgReg, y&: LoadVal)); |
4649 | } |
4650 | |
4651 | // Return if the struct has been fully copied. |
4652 | if (ByValSizeInBytes == OffsetInBytes) |
4653 | return; |
4654 | |
4655 | // Copy the remainder of the byval argument with sub-word loads and shifts. |
4656 | if (LeftoverBytes) { |
4657 | SDValue Val; |
4658 | |
4659 | for (unsigned LoadSizeInBytes = RegSizeInBytes / 2, TotalBytesLoaded = 0; |
4660 | OffsetInBytes < ByValSizeInBytes; LoadSizeInBytes /= 2) { |
4661 | unsigned RemainingSizeInBytes = ByValSizeInBytes - OffsetInBytes; |
4662 | |
4663 | if (RemainingSizeInBytes < LoadSizeInBytes) |
4664 | continue; |
4665 | |
4666 | // Load subword. |
4667 | SDValue LoadPtr = DAG.getNode(Opcode: ISD::ADD, DL, VT: PtrTy, N1: Arg, |
4668 | N2: DAG.getConstant(Val: OffsetInBytes, DL, |
4669 | VT: PtrTy)); |
4670 | SDValue LoadVal = DAG.getExtLoad( |
4671 | ExtType: ISD::ZEXTLOAD, dl: DL, VT: RegTy, Chain, Ptr: LoadPtr, PtrInfo: MachinePointerInfo(), |
4672 | MemVT: MVT::getIntegerVT(BitWidth: LoadSizeInBytes * 8), Alignment); |
4673 | MemOpChains.push_back(Elt: LoadVal.getValue(R: 1)); |
4674 | |
4675 | // Shift the loaded value. |
4676 | unsigned Shamt; |
4677 | |
4678 | if (isLittle) |
4679 | Shamt = TotalBytesLoaded * 8; |
4680 | else |
4681 | Shamt = (RegSizeInBytes - (TotalBytesLoaded + LoadSizeInBytes)) * 8; |
4682 | |
4683 | SDValue Shift = DAG.getNode(Opcode: ISD::SHL, DL, VT: RegTy, N1: LoadVal, |
4684 | N2: DAG.getConstant(Val: Shamt, DL, VT: MVT::i32)); |
4685 | |
4686 | if (Val.getNode()) |
4687 | Val = DAG.getNode(Opcode: ISD::OR, DL, VT: RegTy, N1: Val, N2: Shift); |
4688 | else |
4689 | Val = Shift; |
4690 | |
4691 | OffsetInBytes += LoadSizeInBytes; |
4692 | TotalBytesLoaded += LoadSizeInBytes; |
4693 | Alignment = std::min(a: Alignment, b: Align(LoadSizeInBytes)); |
4694 | } |
4695 | |
4696 | unsigned ArgReg = ArgRegs[FirstReg + I]; |
4697 | RegsToPass.push_back(x: std::make_pair(x&: ArgReg, y&: Val)); |
4698 | return; |
4699 | } |
4700 | } |
4701 | |
4702 | // Copy remainder of byval arg to it with memcpy. |
4703 | unsigned MemCpySize = ByValSizeInBytes - OffsetInBytes; |
4704 | SDValue Src = DAG.getNode(Opcode: ISD::ADD, DL, VT: PtrTy, N1: Arg, |
4705 | N2: DAG.getConstant(Val: OffsetInBytes, DL, VT: PtrTy)); |
4706 | SDValue Dst = DAG.getNode(Opcode: ISD::ADD, DL, VT: PtrTy, N1: StackPtr, |
4707 | N2: DAG.getIntPtrConstant(Val: VA.getLocMemOffset(), DL)); |
4708 | Chain = DAG.getMemcpy( |
4709 | Chain, dl: DL, Dst, Src, Size: DAG.getConstant(Val: MemCpySize, DL, VT: PtrTy), |
4710 | Alignment: Align(Alignment), /*isVolatile=*/isVol: false, /*AlwaysInline=*/false, |
4711 | /*CI=*/nullptr, OverrideTailCall: std::nullopt, DstPtrInfo: MachinePointerInfo(), SrcPtrInfo: MachinePointerInfo()); |
4712 | MemOpChains.push_back(Elt: Chain); |
4713 | } |
4714 | |
4715 | void MipsTargetLowering::writeVarArgRegs(std::vector<SDValue> &OutChains, |
4716 | SDValue Chain, const SDLoc &DL, |
4717 | SelectionDAG &DAG, |
4718 | CCState &State) const { |
4719 | ArrayRef<MCPhysReg> ArgRegs = ABI.getVarArgRegs(isGP64bit: Subtarget.isGP64bit()); |
4720 | unsigned Idx = State.getFirstUnallocated(Regs: ArgRegs); |
4721 | unsigned RegSizeInBytes = Subtarget.getGPRSizeInBytes(); |
4722 | MVT RegTy = MVT::getIntegerVT(BitWidth: RegSizeInBytes * 8); |
4723 | const TargetRegisterClass *RC = getRegClassFor(VT: RegTy); |
4724 | MachineFunction &MF = DAG.getMachineFunction(); |
4725 | MachineFrameInfo &MFI = MF.getFrameInfo(); |
4726 | MipsFunctionInfo *MipsFI = MF.getInfo<MipsFunctionInfo>(); |
4727 | |
4728 | // Offset of the first variable argument from stack pointer. |
4729 | int VaArgOffset; |
4730 | |
4731 | if (ArgRegs.size() == Idx) |
4732 | VaArgOffset = alignTo(Value: State.getStackSize(), Align: RegSizeInBytes); |
4733 | else { |
4734 | VaArgOffset = |
4735 | (int)ABI.GetCalleeAllocdArgSizeInBytes(CC: State.getCallingConv()) - |
4736 | (int)(RegSizeInBytes * (ArgRegs.size() - Idx)); |
4737 | } |
4738 | |
4739 | // Record the frame index of the first variable argument |
4740 | // which is a value necessary to VASTART. |
4741 | int FI = MFI.CreateFixedObject(Size: RegSizeInBytes, SPOffset: VaArgOffset, IsImmutable: true); |
4742 | MipsFI->setVarArgsFrameIndex(FI); |
4743 | |
4744 | // Copy the integer registers that have not been used for argument passing |
4745 | // to the argument register save area. For O32, the save area is allocated |
4746 | // in the caller's stack frame, while for N32/64, it is allocated in the |
4747 | // callee's stack frame. |
4748 | for (unsigned I = Idx; I < ArgRegs.size(); |
4749 | ++I, VaArgOffset += RegSizeInBytes) { |
4750 | unsigned Reg = addLiveIn(MF, PReg: ArgRegs[I], RC); |
4751 | SDValue ArgValue = DAG.getCopyFromReg(Chain, dl: DL, Reg, VT: RegTy); |
4752 | FI = MFI.CreateFixedObject(Size: RegSizeInBytes, SPOffset: VaArgOffset, IsImmutable: true); |
4753 | SDValue PtrOff = DAG.getFrameIndex(FI, VT: getPointerTy(DL: DAG.getDataLayout())); |
4754 | SDValue Store = |
4755 | DAG.getStore(Chain, dl: DL, Val: ArgValue, Ptr: PtrOff, PtrInfo: MachinePointerInfo()); |
4756 | cast<StoreSDNode>(Val: Store.getNode())->getMemOperand()->setValue( |
4757 | (Value *)nullptr); |
4758 | OutChains.push_back(x: Store); |
4759 | } |
4760 | } |
4761 | |
4762 | void MipsTargetLowering::HandleByVal(CCState *State, unsigned &Size, |
4763 | Align Alignment) const { |
4764 | const TargetFrameLowering *TFL = Subtarget.getFrameLowering(); |
4765 | |
4766 | assert(Size && "Byval argument's size shouldn't be 0." ); |
4767 | |
4768 | Alignment = std::min(a: Alignment, b: TFL->getStackAlign()); |
4769 | |
4770 | unsigned FirstReg = 0; |
4771 | unsigned NumRegs = 0; |
4772 | |
4773 | if (State->getCallingConv() != CallingConv::Fast) { |
4774 | unsigned RegSizeInBytes = Subtarget.getGPRSizeInBytes(); |
4775 | ArrayRef<MCPhysReg> IntArgRegs = ABI.GetByValArgRegs(); |
4776 | // FIXME: The O32 case actually describes no shadow registers. |
4777 | const MCPhysReg *ShadowRegs = |
4778 | ABI.IsO32() ? IntArgRegs.data() : Mips64DPRegs; |
4779 | |
4780 | // We used to check the size as well but we can't do that anymore since |
4781 | // CCState::HandleByVal() rounds up the size after calling this function. |
4782 | assert( |
4783 | Alignment >= Align(RegSizeInBytes) && |
4784 | "Byval argument's alignment should be a multiple of RegSizeInBytes." ); |
4785 | |
4786 | FirstReg = State->getFirstUnallocated(Regs: IntArgRegs); |
4787 | |
4788 | // If Alignment > RegSizeInBytes, the first arg register must be even. |
4789 | // FIXME: This condition happens to do the right thing but it's not the |
4790 | // right way to test it. We want to check that the stack frame offset |
4791 | // of the register is aligned. |
4792 | if ((Alignment > RegSizeInBytes) && (FirstReg % 2)) { |
4793 | State->AllocateReg(Reg: IntArgRegs[FirstReg], ShadowReg: ShadowRegs[FirstReg]); |
4794 | ++FirstReg; |
4795 | } |
4796 | |
4797 | // Mark the registers allocated. |
4798 | Size = alignTo(Value: Size, Align: RegSizeInBytes); |
4799 | for (unsigned I = FirstReg; Size > 0 && (I < IntArgRegs.size()); |
4800 | Size -= RegSizeInBytes, ++I, ++NumRegs) |
4801 | State->AllocateReg(Reg: IntArgRegs[I], ShadowReg: ShadowRegs[I]); |
4802 | } |
4803 | |
4804 | State->addInRegsParamInfo(RegBegin: FirstReg, RegEnd: FirstReg + NumRegs); |
4805 | } |
4806 | |
4807 | MachineBasicBlock *MipsTargetLowering::emitPseudoSELECT(MachineInstr &MI, |
4808 | MachineBasicBlock *BB, |
4809 | bool isFPCmp, |
4810 | unsigned Opc) const { |
4811 | assert(!(Subtarget.hasMips4() || Subtarget.hasMips32()) && |
4812 | "Subtarget already supports SELECT nodes with the use of" |
4813 | "conditional-move instructions." ); |
4814 | |
4815 | const TargetInstrInfo *TII = |
4816 | Subtarget.getInstrInfo(); |
4817 | DebugLoc DL = MI.getDebugLoc(); |
4818 | |
4819 | // To "insert" a SELECT instruction, we actually have to insert the |
4820 | // diamond control-flow pattern. The incoming instruction knows the |
4821 | // destination vreg to set, the condition code register to branch on, the |
4822 | // true/false values to select between, and a branch opcode to use. |
4823 | const BasicBlock *LLVM_BB = BB->getBasicBlock(); |
4824 | MachineFunction::iterator It = ++BB->getIterator(); |
4825 | |
4826 | // thisMBB: |
4827 | // ... |
4828 | // TrueVal = ... |
4829 | // setcc r1, r2, r3 |
4830 | // bNE r1, r0, copy1MBB |
4831 | // fallthrough --> copy0MBB |
4832 | MachineBasicBlock *thisMBB = BB; |
4833 | MachineFunction *F = BB->getParent(); |
4834 | MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(BB: LLVM_BB); |
4835 | MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(BB: LLVM_BB); |
4836 | F->insert(MBBI: It, MBB: copy0MBB); |
4837 | F->insert(MBBI: It, MBB: sinkMBB); |
4838 | |
4839 | // Transfer the remainder of BB and its successor edges to sinkMBB. |
4840 | sinkMBB->splice(Where: sinkMBB->begin(), Other: BB, |
4841 | From: std::next(x: MachineBasicBlock::iterator(MI)), To: BB->end()); |
4842 | sinkMBB->transferSuccessorsAndUpdatePHIs(FromMBB: BB); |
4843 | |
4844 | // Next, add the true and fallthrough blocks as its successors. |
4845 | BB->addSuccessor(Succ: copy0MBB); |
4846 | BB->addSuccessor(Succ: sinkMBB); |
4847 | |
4848 | if (isFPCmp) { |
4849 | // bc1[tf] cc, sinkMBB |
4850 | BuildMI(BB, MIMD: DL, MCID: TII->get(Opcode: Opc)) |
4851 | .addReg(RegNo: MI.getOperand(i: 1).getReg()) |
4852 | .addMBB(MBB: sinkMBB); |
4853 | } else { |
4854 | // bne rs, $0, sinkMBB |
4855 | BuildMI(BB, MIMD: DL, MCID: TII->get(Opcode: Opc)) |
4856 | .addReg(RegNo: MI.getOperand(i: 1).getReg()) |
4857 | .addReg(RegNo: Mips::ZERO) |
4858 | .addMBB(MBB: sinkMBB); |
4859 | } |
4860 | |
4861 | // copy0MBB: |
4862 | // %FalseValue = ... |
4863 | // # fallthrough to sinkMBB |
4864 | BB = copy0MBB; |
4865 | |
4866 | // Update machine-CFG edges |
4867 | BB->addSuccessor(Succ: sinkMBB); |
4868 | |
4869 | // sinkMBB: |
4870 | // %Result = phi [ %TrueValue, thisMBB ], [ %FalseValue, copy0MBB ] |
4871 | // ... |
4872 | BB = sinkMBB; |
4873 | |
4874 | BuildMI(BB&: *BB, I: BB->begin(), MIMD: DL, MCID: TII->get(Opcode: Mips::PHI), DestReg: MI.getOperand(i: 0).getReg()) |
4875 | .addReg(RegNo: MI.getOperand(i: 2).getReg()) |
4876 | .addMBB(MBB: thisMBB) |
4877 | .addReg(RegNo: MI.getOperand(i: 3).getReg()) |
4878 | .addMBB(MBB: copy0MBB); |
4879 | |
4880 | MI.eraseFromParent(); // The pseudo instruction is gone now. |
4881 | |
4882 | return BB; |
4883 | } |
4884 | |
4885 | MachineBasicBlock * |
4886 | MipsTargetLowering::emitPseudoD_SELECT(MachineInstr &MI, |
4887 | MachineBasicBlock *BB) const { |
4888 | assert(!(Subtarget.hasMips4() || Subtarget.hasMips32()) && |
4889 | "Subtarget already supports SELECT nodes with the use of" |
4890 | "conditional-move instructions." ); |
4891 | |
4892 | const TargetInstrInfo *TII = Subtarget.getInstrInfo(); |
4893 | DebugLoc DL = MI.getDebugLoc(); |
4894 | |
4895 | // D_SELECT substitutes two SELECT nodes that goes one after another and |
4896 | // have the same condition operand. On machines which don't have |
4897 | // conditional-move instruction, it reduces unnecessary branch instructions |
4898 | // which are result of using two diamond patterns that are result of two |
4899 | // SELECT pseudo instructions. |
4900 | const BasicBlock *LLVM_BB = BB->getBasicBlock(); |
4901 | MachineFunction::iterator It = ++BB->getIterator(); |
4902 | |
4903 | // thisMBB: |
4904 | // ... |
4905 | // TrueVal = ... |
4906 | // setcc r1, r2, r3 |
4907 | // bNE r1, r0, copy1MBB |
4908 | // fallthrough --> copy0MBB |
4909 | MachineBasicBlock *thisMBB = BB; |
4910 | MachineFunction *F = BB->getParent(); |
4911 | MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(BB: LLVM_BB); |
4912 | MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(BB: LLVM_BB); |
4913 | F->insert(MBBI: It, MBB: copy0MBB); |
4914 | F->insert(MBBI: It, MBB: sinkMBB); |
4915 | |
4916 | // Transfer the remainder of BB and its successor edges to sinkMBB. |
4917 | sinkMBB->splice(Where: sinkMBB->begin(), Other: BB, |
4918 | From: std::next(x: MachineBasicBlock::iterator(MI)), To: BB->end()); |
4919 | sinkMBB->transferSuccessorsAndUpdatePHIs(FromMBB: BB); |
4920 | |
4921 | // Next, add the true and fallthrough blocks as its successors. |
4922 | BB->addSuccessor(Succ: copy0MBB); |
4923 | BB->addSuccessor(Succ: sinkMBB); |
4924 | |
4925 | // bne rs, $0, sinkMBB |
4926 | BuildMI(BB, MIMD: DL, MCID: TII->get(Opcode: Mips::BNE)) |
4927 | .addReg(RegNo: MI.getOperand(i: 2).getReg()) |
4928 | .addReg(RegNo: Mips::ZERO) |
4929 | .addMBB(MBB: sinkMBB); |
4930 | |
4931 | // copy0MBB: |
4932 | // %FalseValue = ... |
4933 | // # fallthrough to sinkMBB |
4934 | BB = copy0MBB; |
4935 | |
4936 | // Update machine-CFG edges |
4937 | BB->addSuccessor(Succ: sinkMBB); |
4938 | |
4939 | // sinkMBB: |
4940 | // %Result = phi [ %TrueValue, thisMBB ], [ %FalseValue, copy0MBB ] |
4941 | // ... |
4942 | BB = sinkMBB; |
4943 | |
4944 | // Use two PHI nodes to select two reults |
4945 | BuildMI(BB&: *BB, I: BB->begin(), MIMD: DL, MCID: TII->get(Opcode: Mips::PHI), DestReg: MI.getOperand(i: 0).getReg()) |
4946 | .addReg(RegNo: MI.getOperand(i: 3).getReg()) |
4947 | .addMBB(MBB: thisMBB) |
4948 | .addReg(RegNo: MI.getOperand(i: 5).getReg()) |
4949 | .addMBB(MBB: copy0MBB); |
4950 | BuildMI(BB&: *BB, I: BB->begin(), MIMD: DL, MCID: TII->get(Opcode: Mips::PHI), DestReg: MI.getOperand(i: 1).getReg()) |
4951 | .addReg(RegNo: MI.getOperand(i: 4).getReg()) |
4952 | .addMBB(MBB: thisMBB) |
4953 | .addReg(RegNo: MI.getOperand(i: 6).getReg()) |
4954 | .addMBB(MBB: copy0MBB); |
4955 | |
4956 | MI.eraseFromParent(); // The pseudo instruction is gone now. |
4957 | |
4958 | return BB; |
4959 | } |
4960 | |
4961 | // FIXME? Maybe this could be a TableGen attribute on some registers and |
4962 | // this table could be generated automatically from RegInfo. |
4963 | Register |
4964 | MipsTargetLowering::getRegisterByName(const char *RegName, LLT VT, |
4965 | const MachineFunction &MF) const { |
4966 | // The Linux kernel uses $28 and sp. |
4967 | if (Subtarget.isGP64bit()) { |
4968 | Register Reg = StringSwitch<Register>(RegName) |
4969 | .Case(S: "$28" , Value: Mips::GP_64) |
4970 | .Case(S: "sp" , Value: Mips::SP_64) |
4971 | .Default(Value: Register()); |
4972 | return Reg; |
4973 | } |
4974 | |
4975 | Register Reg = StringSwitch<Register>(RegName) |
4976 | .Case(S: "$28" , Value: Mips::GP) |
4977 | .Case(S: "sp" , Value: Mips::SP) |
4978 | .Default(Value: Register()); |
4979 | return Reg; |
4980 | } |
4981 | |
4982 | MachineBasicBlock *MipsTargetLowering::emitLDR_W(MachineInstr &MI, |
4983 | MachineBasicBlock *BB) const { |
4984 | MachineFunction *MF = BB->getParent(); |
4985 | MachineRegisterInfo &MRI = MF->getRegInfo(); |
4986 | const TargetInstrInfo *TII = Subtarget.getInstrInfo(); |
4987 | const bool IsLittle = Subtarget.isLittle(); |
4988 | DebugLoc DL = MI.getDebugLoc(); |
4989 | |
4990 | Register Dest = MI.getOperand(i: 0).getReg(); |
4991 | Register Address = MI.getOperand(i: 1).getReg(); |
4992 | unsigned Imm = MI.getOperand(i: 2).getImm(); |
4993 | |
4994 | MachineBasicBlock::iterator I(MI); |
4995 | |
4996 | if (Subtarget.hasMips32r6() || Subtarget.hasMips64r6()) { |
4997 | // Mips release 6 can load from adress that is not naturally-aligned. |
4998 | Register Temp = MRI.createVirtualRegister(RegClass: &Mips::GPR32RegClass); |
4999 | BuildMI(BB&: *BB, I, MIMD: DL, MCID: TII->get(Opcode: Mips::LW)) |
5000 | .addDef(RegNo: Temp) |
5001 | .addUse(RegNo: Address) |
5002 | .addImm(Val: Imm); |
5003 | BuildMI(BB&: *BB, I, MIMD: DL, MCID: TII->get(Opcode: Mips::FILL_W)).addDef(RegNo: Dest).addUse(RegNo: Temp); |
5004 | } else { |
5005 | // Mips release 5 needs to use instructions that can load from an unaligned |
5006 | // memory address. |
5007 | Register LoadHalf = MRI.createVirtualRegister(RegClass: &Mips::GPR32RegClass); |
5008 | Register LoadFull = MRI.createVirtualRegister(RegClass: &Mips::GPR32RegClass); |
5009 | Register Undef = MRI.createVirtualRegister(RegClass: &Mips::GPR32RegClass); |
5010 | BuildMI(BB&: *BB, I, MIMD: DL, MCID: TII->get(Opcode: Mips::IMPLICIT_DEF)).addDef(RegNo: Undef); |
5011 | BuildMI(BB&: *BB, I, MIMD: DL, MCID: TII->get(Opcode: Mips::LWR)) |
5012 | .addDef(RegNo: LoadHalf) |
5013 | .addUse(RegNo: Address) |
5014 | .addImm(Val: Imm + (IsLittle ? 0 : 3)) |
5015 | .addUse(RegNo: Undef); |
5016 | BuildMI(BB&: *BB, I, MIMD: DL, MCID: TII->get(Opcode: Mips::LWL)) |
5017 | .addDef(RegNo: LoadFull) |
5018 | .addUse(RegNo: Address) |
5019 | .addImm(Val: Imm + (IsLittle ? 3 : 0)) |
5020 | .addUse(RegNo: LoadHalf); |
5021 | BuildMI(BB&: *BB, I, MIMD: DL, MCID: TII->get(Opcode: Mips::FILL_W)).addDef(RegNo: Dest).addUse(RegNo: LoadFull); |
5022 | } |
5023 | |
5024 | MI.eraseFromParent(); |
5025 | return BB; |
5026 | } |
5027 | |
5028 | MachineBasicBlock *MipsTargetLowering::emitLDR_D(MachineInstr &MI, |
5029 | MachineBasicBlock *BB) const { |
5030 | MachineFunction *MF = BB->getParent(); |
5031 | MachineRegisterInfo &MRI = MF->getRegInfo(); |
5032 | const TargetInstrInfo *TII = Subtarget.getInstrInfo(); |
5033 | const bool IsLittle = Subtarget.isLittle(); |
5034 | DebugLoc DL = MI.getDebugLoc(); |
5035 | |
5036 | Register Dest = MI.getOperand(i: 0).getReg(); |
5037 | Register Address = MI.getOperand(i: 1).getReg(); |
5038 | unsigned Imm = MI.getOperand(i: 2).getImm(); |
5039 | |
5040 | MachineBasicBlock::iterator I(MI); |
5041 | |
5042 | if (Subtarget.hasMips32r6() || Subtarget.hasMips64r6()) { |
5043 | // Mips release 6 can load from adress that is not naturally-aligned. |
5044 | if (Subtarget.isGP64bit()) { |
5045 | Register Temp = MRI.createVirtualRegister(RegClass: &Mips::GPR64RegClass); |
5046 | BuildMI(BB&: *BB, I, MIMD: DL, MCID: TII->get(Opcode: Mips::LD)) |
5047 | .addDef(RegNo: Temp) |
5048 | .addUse(RegNo: Address) |
5049 | .addImm(Val: Imm); |
5050 | BuildMI(BB&: *BB, I, MIMD: DL, MCID: TII->get(Opcode: Mips::FILL_D)).addDef(RegNo: Dest).addUse(RegNo: Temp); |
5051 | } else { |
5052 | Register Wtemp = MRI.createVirtualRegister(RegClass: &Mips::MSA128WRegClass); |
5053 | Register Lo = MRI.createVirtualRegister(RegClass: &Mips::GPR32RegClass); |
5054 | Register Hi = MRI.createVirtualRegister(RegClass: &Mips::GPR32RegClass); |
5055 | BuildMI(BB&: *BB, I, MIMD: DL, MCID: TII->get(Opcode: Mips::LW)) |
5056 | .addDef(RegNo: Lo) |
5057 | .addUse(RegNo: Address) |
5058 | .addImm(Val: Imm + (IsLittle ? 0 : 4)); |
5059 | BuildMI(BB&: *BB, I, MIMD: DL, MCID: TII->get(Opcode: Mips::LW)) |
5060 | .addDef(RegNo: Hi) |
5061 | .addUse(RegNo: Address) |
5062 | .addImm(Val: Imm + (IsLittle ? 4 : 0)); |
5063 | BuildMI(BB&: *BB, I, MIMD: DL, MCID: TII->get(Opcode: Mips::FILL_W)).addDef(RegNo: Wtemp).addUse(RegNo: Lo); |
5064 | BuildMI(BB&: *BB, I, MIMD: DL, MCID: TII->get(Opcode: Mips::INSERT_W), DestReg: Dest) |
5065 | .addUse(RegNo: Wtemp) |
5066 | .addUse(RegNo: Hi) |
5067 | .addImm(Val: 1); |
5068 | } |
5069 | } else { |
5070 | // Mips release 5 needs to use instructions that can load from an unaligned |
5071 | // memory address. |
5072 | Register LoHalf = MRI.createVirtualRegister(RegClass: &Mips::GPR32RegClass); |
5073 | Register LoFull = MRI.createVirtualRegister(RegClass: &Mips::GPR32RegClass); |
5074 | Register LoUndef = MRI.createVirtualRegister(RegClass: &Mips::GPR32RegClass); |
5075 | Register HiHalf = MRI.createVirtualRegister(RegClass: &Mips::GPR32RegClass); |
5076 | Register HiFull = MRI.createVirtualRegister(RegClass: &Mips::GPR32RegClass); |
5077 | Register HiUndef = MRI.createVirtualRegister(RegClass: &Mips::GPR32RegClass); |
5078 | Register Wtemp = MRI.createVirtualRegister(RegClass: &Mips::MSA128WRegClass); |
5079 | BuildMI(BB&: *BB, I, MIMD: DL, MCID: TII->get(Opcode: Mips::IMPLICIT_DEF)).addDef(RegNo: LoUndef); |
5080 | BuildMI(BB&: *BB, I, MIMD: DL, MCID: TII->get(Opcode: Mips::LWR)) |
5081 | .addDef(RegNo: LoHalf) |
5082 | .addUse(RegNo: Address) |
5083 | .addImm(Val: Imm + (IsLittle ? 0 : 7)) |
5084 | .addUse(RegNo: LoUndef); |
5085 | BuildMI(BB&: *BB, I, MIMD: DL, MCID: TII->get(Opcode: Mips::LWL)) |
5086 | .addDef(RegNo: LoFull) |
5087 | .addUse(RegNo: Address) |
5088 | .addImm(Val: Imm + (IsLittle ? 3 : 4)) |
5089 | .addUse(RegNo: LoHalf); |
5090 | BuildMI(BB&: *BB, I, MIMD: DL, MCID: TII->get(Opcode: Mips::IMPLICIT_DEF)).addDef(RegNo: HiUndef); |
5091 | BuildMI(BB&: *BB, I, MIMD: DL, MCID: TII->get(Opcode: Mips::LWR)) |
5092 | .addDef(RegNo: HiHalf) |
5093 | .addUse(RegNo: Address) |
5094 | .addImm(Val: Imm + (IsLittle ? 4 : 3)) |
5095 | .addUse(RegNo: HiUndef); |
5096 | BuildMI(BB&: *BB, I, MIMD: DL, MCID: TII->get(Opcode: Mips::LWL)) |
5097 | .addDef(RegNo: HiFull) |
5098 | .addUse(RegNo: Address) |
5099 | .addImm(Val: Imm + (IsLittle ? 7 : 0)) |
5100 | .addUse(RegNo: HiHalf); |
5101 | BuildMI(BB&: *BB, I, MIMD: DL, MCID: TII->get(Opcode: Mips::FILL_W)).addDef(RegNo: Wtemp).addUse(RegNo: LoFull); |
5102 | BuildMI(BB&: *BB, I, MIMD: DL, MCID: TII->get(Opcode: Mips::INSERT_W), DestReg: Dest) |
5103 | .addUse(RegNo: Wtemp) |
5104 | .addUse(RegNo: HiFull) |
5105 | .addImm(Val: 1); |
5106 | } |
5107 | |
5108 | MI.eraseFromParent(); |
5109 | return BB; |
5110 | } |
5111 | |
5112 | MachineBasicBlock *MipsTargetLowering::emitSTR_W(MachineInstr &MI, |
5113 | MachineBasicBlock *BB) const { |
5114 | MachineFunction *MF = BB->getParent(); |
5115 | MachineRegisterInfo &MRI = MF->getRegInfo(); |
5116 | const TargetInstrInfo *TII = Subtarget.getInstrInfo(); |
5117 | const bool IsLittle = Subtarget.isLittle(); |
5118 | DebugLoc DL = MI.getDebugLoc(); |
5119 | |
5120 | Register StoreVal = MI.getOperand(i: 0).getReg(); |
5121 | Register Address = MI.getOperand(i: 1).getReg(); |
5122 | unsigned Imm = MI.getOperand(i: 2).getImm(); |
5123 | |
5124 | MachineBasicBlock::iterator I(MI); |
5125 | |
5126 | if (Subtarget.hasMips32r6() || Subtarget.hasMips64r6()) { |
5127 | // Mips release 6 can store to adress that is not naturally-aligned. |
5128 | Register BitcastW = MRI.createVirtualRegister(RegClass: &Mips::MSA128WRegClass); |
5129 | Register Tmp = MRI.createVirtualRegister(RegClass: &Mips::GPR32RegClass); |
5130 | BuildMI(BB&: *BB, I, MIMD: DL, MCID: TII->get(Opcode: Mips::COPY)).addDef(RegNo: BitcastW).addUse(RegNo: StoreVal); |
5131 | BuildMI(BB&: *BB, I, MIMD: DL, MCID: TII->get(Opcode: Mips::COPY_S_W)) |
5132 | .addDef(RegNo: Tmp) |
5133 | .addUse(RegNo: BitcastW) |
5134 | .addImm(Val: 0); |
5135 | BuildMI(BB&: *BB, I, MIMD: DL, MCID: TII->get(Opcode: Mips::SW)) |
5136 | .addUse(RegNo: Tmp) |
5137 | .addUse(RegNo: Address) |
5138 | .addImm(Val: Imm); |
5139 | } else { |
5140 | // Mips release 5 needs to use instructions that can store to an unaligned |
5141 | // memory address. |
5142 | Register Tmp = MRI.createVirtualRegister(RegClass: &Mips::GPR32RegClass); |
5143 | BuildMI(BB&: *BB, I, MIMD: DL, MCID: TII->get(Opcode: Mips::COPY_S_W)) |
5144 | .addDef(RegNo: Tmp) |
5145 | .addUse(RegNo: StoreVal) |
5146 | .addImm(Val: 0); |
5147 | BuildMI(BB&: *BB, I, MIMD: DL, MCID: TII->get(Opcode: Mips::SWR)) |
5148 | .addUse(RegNo: Tmp) |
5149 | .addUse(RegNo: Address) |
5150 | .addImm(Val: Imm + (IsLittle ? 0 : 3)); |
5151 | BuildMI(BB&: *BB, I, MIMD: DL, MCID: TII->get(Opcode: Mips::SWL)) |
5152 | .addUse(RegNo: Tmp) |
5153 | .addUse(RegNo: Address) |
5154 | .addImm(Val: Imm + (IsLittle ? 3 : 0)); |
5155 | } |
5156 | |
5157 | MI.eraseFromParent(); |
5158 | |
5159 | return BB; |
5160 | } |
5161 | |
5162 | MachineBasicBlock *MipsTargetLowering::emitSTR_D(MachineInstr &MI, |
5163 | MachineBasicBlock *BB) const { |
5164 | MachineFunction *MF = BB->getParent(); |
5165 | MachineRegisterInfo &MRI = MF->getRegInfo(); |
5166 | const TargetInstrInfo *TII = Subtarget.getInstrInfo(); |
5167 | const bool IsLittle = Subtarget.isLittle(); |
5168 | DebugLoc DL = MI.getDebugLoc(); |
5169 | |
5170 | Register StoreVal = MI.getOperand(i: 0).getReg(); |
5171 | Register Address = MI.getOperand(i: 1).getReg(); |
5172 | unsigned Imm = MI.getOperand(i: 2).getImm(); |
5173 | |
5174 | MachineBasicBlock::iterator I(MI); |
5175 | |
5176 | if (Subtarget.hasMips32r6() || Subtarget.hasMips64r6()) { |
5177 | // Mips release 6 can store to adress that is not naturally-aligned. |
5178 | if (Subtarget.isGP64bit()) { |
5179 | Register BitcastD = MRI.createVirtualRegister(RegClass: &Mips::MSA128DRegClass); |
5180 | Register Lo = MRI.createVirtualRegister(RegClass: &Mips::GPR64RegClass); |
5181 | BuildMI(BB&: *BB, I, MIMD: DL, MCID: TII->get(Opcode: Mips::COPY)) |
5182 | .addDef(RegNo: BitcastD) |
5183 | .addUse(RegNo: StoreVal); |
5184 | BuildMI(BB&: *BB, I, MIMD: DL, MCID: TII->get(Opcode: Mips::COPY_S_D)) |
5185 | .addDef(RegNo: Lo) |
5186 | .addUse(RegNo: BitcastD) |
5187 | .addImm(Val: 0); |
5188 | BuildMI(BB&: *BB, I, MIMD: DL, MCID: TII->get(Opcode: Mips::SD)) |
5189 | .addUse(RegNo: Lo) |
5190 | .addUse(RegNo: Address) |
5191 | .addImm(Val: Imm); |
5192 | } else { |
5193 | Register BitcastW = MRI.createVirtualRegister(RegClass: &Mips::MSA128WRegClass); |
5194 | Register Lo = MRI.createVirtualRegister(RegClass: &Mips::GPR32RegClass); |
5195 | Register Hi = MRI.createVirtualRegister(RegClass: &Mips::GPR32RegClass); |
5196 | BuildMI(BB&: *BB, I, MIMD: DL, MCID: TII->get(Opcode: Mips::COPY)) |
5197 | .addDef(RegNo: BitcastW) |
5198 | .addUse(RegNo: StoreVal); |
5199 | BuildMI(BB&: *BB, I, MIMD: DL, MCID: TII->get(Opcode: Mips::COPY_S_W)) |
5200 | .addDef(RegNo: Lo) |
5201 | .addUse(RegNo: BitcastW) |
5202 | .addImm(Val: 0); |
5203 | BuildMI(BB&: *BB, I, MIMD: DL, MCID: TII->get(Opcode: Mips::COPY_S_W)) |
5204 | .addDef(RegNo: Hi) |
5205 | .addUse(RegNo: BitcastW) |
5206 | .addImm(Val: 1); |
5207 | BuildMI(BB&: *BB, I, MIMD: DL, MCID: TII->get(Opcode: Mips::SW)) |
5208 | .addUse(RegNo: Lo) |
5209 | .addUse(RegNo: Address) |
5210 | .addImm(Val: Imm + (IsLittle ? 0 : 4)); |
5211 | BuildMI(BB&: *BB, I, MIMD: DL, MCID: TII->get(Opcode: Mips::SW)) |
5212 | .addUse(RegNo: Hi) |
5213 | .addUse(RegNo: Address) |
5214 | .addImm(Val: Imm + (IsLittle ? 4 : 0)); |
5215 | } |
5216 | } else { |
5217 | // Mips release 5 needs to use instructions that can store to an unaligned |
5218 | // memory address. |
5219 | Register Bitcast = MRI.createVirtualRegister(RegClass: &Mips::MSA128WRegClass); |
5220 | Register Lo = MRI.createVirtualRegister(RegClass: &Mips::GPR32RegClass); |
5221 | Register Hi = MRI.createVirtualRegister(RegClass: &Mips::GPR32RegClass); |
5222 | BuildMI(BB&: *BB, I, MIMD: DL, MCID: TII->get(Opcode: Mips::COPY)).addDef(RegNo: Bitcast).addUse(RegNo: StoreVal); |
5223 | BuildMI(BB&: *BB, I, MIMD: DL, MCID: TII->get(Opcode: Mips::COPY_S_W)) |
5224 | .addDef(RegNo: Lo) |
5225 | .addUse(RegNo: Bitcast) |
5226 | .addImm(Val: 0); |
5227 | BuildMI(BB&: *BB, I, MIMD: DL, MCID: TII->get(Opcode: Mips::COPY_S_W)) |
5228 | .addDef(RegNo: Hi) |
5229 | .addUse(RegNo: Bitcast) |
5230 | .addImm(Val: 1); |
5231 | BuildMI(BB&: *BB, I, MIMD: DL, MCID: TII->get(Opcode: Mips::SWR)) |
5232 | .addUse(RegNo: Lo) |
5233 | .addUse(RegNo: Address) |
5234 | .addImm(Val: Imm + (IsLittle ? 0 : 3)); |
5235 | BuildMI(BB&: *BB, I, MIMD: DL, MCID: TII->get(Opcode: Mips::SWL)) |
5236 | .addUse(RegNo: Lo) |
5237 | .addUse(RegNo: Address) |
5238 | .addImm(Val: Imm + (IsLittle ? 3 : 0)); |
5239 | BuildMI(BB&: *BB, I, MIMD: DL, MCID: TII->get(Opcode: Mips::SWR)) |
5240 | .addUse(RegNo: Hi) |
5241 | .addUse(RegNo: Address) |
5242 | .addImm(Val: Imm + (IsLittle ? 4 : 7)); |
5243 | BuildMI(BB&: *BB, I, MIMD: DL, MCID: TII->get(Opcode: Mips::SWL)) |
5244 | .addUse(RegNo: Hi) |
5245 | .addUse(RegNo: Address) |
5246 | .addImm(Val: Imm + (IsLittle ? 7 : 4)); |
5247 | } |
5248 | |
5249 | MI.eraseFromParent(); |
5250 | return BB; |
5251 | } |
5252 | |