1 //===-- MipsISelLowering.cpp - Mips DAG Lowering Implementation -----------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file defines the interfaces that Mips uses to lower LLVM code into a
13 //===----------------------------------------------------------------------===//
14 #include "MipsISelLowering.h"
15 #include "InstPrinter/MipsInstPrinter.h"
16 #include "MCTargetDesc/MipsBaseInfo.h"
17 #include "MipsCCState.h"
18 #include "MipsMachineFunction.h"
19 #include "MipsSubtarget.h"
20 #include "MipsTargetMachine.h"
21 #include "MipsTargetObjectFile.h"
22 #include "llvm/ADT/Statistic.h"
23 #include "llvm/ADT/StringSwitch.h"
24 #include "llvm/CodeGen/CallingConvLower.h"
25 #include "llvm/CodeGen/MachineFrameInfo.h"
26 #include "llvm/CodeGen/MachineFunction.h"
27 #include "llvm/CodeGen/MachineInstrBuilder.h"
28 #include "llvm/CodeGen/MachineJumpTableInfo.h"
29 #include "llvm/CodeGen/MachineRegisterInfo.h"
30 #include "llvm/CodeGen/FunctionLoweringInfo.h"
31 #include "llvm/CodeGen/SelectionDAGISel.h"
32 #include "llvm/CodeGen/ValueTypes.h"
33 #include "llvm/IR/CallingConv.h"
34 #include "llvm/IR/DerivedTypes.h"
35 #include "llvm/IR/GlobalVariable.h"
36 #include "llvm/Support/CommandLine.h"
37 #include "llvm/Support/Debug.h"
38 #include "llvm/Support/ErrorHandling.h"
39 #include "llvm/Support/raw_ostream.h"
44 #define DEBUG_TYPE "mips-lower"
46 STATISTIC(NumTailCalls, "Number of tail calls");
49 LargeGOT("mxgot", cl::Hidden,
50 cl::desc("MIPS: Enable GOT larger than 64k."), cl::init(false));
53 NoZeroDivCheck("mno-check-zero-division", cl::Hidden,
54 cl::desc("MIPS: Don't trap on integer division by zero."),
57 static const MCPhysReg Mips64DPRegs[8] = {
58 Mips::D12_64, Mips::D13_64, Mips::D14_64, Mips::D15_64,
59 Mips::D16_64, Mips::D17_64, Mips::D18_64, Mips::D19_64
62 // If I is a shifted mask, set the size (Size) and the first bit of the
63 // mask (Pos), and return true.
64 // For example, if I is 0x003ff800, (Pos, Size) = (11, 11).
65 static bool isShiftedMask(uint64_t I, uint64_t &Pos, uint64_t &Size) {
66 if (!isShiftedMask_64(I))
69 Size = countPopulation(I);
70 Pos = countTrailingZeros(I);
74 SDValue MipsTargetLowering::getGlobalReg(SelectionDAG &DAG, EVT Ty) const {
75 MipsFunctionInfo *FI = DAG.getMachineFunction().getInfo<MipsFunctionInfo>();
76 return DAG.getRegister(FI->getGlobalBaseReg(), Ty);
79 SDValue MipsTargetLowering::getTargetNode(GlobalAddressSDNode *N, EVT Ty,
81 unsigned Flag) const {
82 return DAG.getTargetGlobalAddress(N->getGlobal(), SDLoc(N), Ty, 0, Flag);
85 SDValue MipsTargetLowering::getTargetNode(ExternalSymbolSDNode *N, EVT Ty,
87 unsigned Flag) const {
88 return DAG.getTargetExternalSymbol(N->getSymbol(), Ty, Flag);
91 SDValue MipsTargetLowering::getTargetNode(BlockAddressSDNode *N, EVT Ty,
93 unsigned Flag) const {
94 return DAG.getTargetBlockAddress(N->getBlockAddress(), Ty, 0, Flag);
97 SDValue MipsTargetLowering::getTargetNode(JumpTableSDNode *N, EVT Ty,
99 unsigned Flag) const {
100 return DAG.getTargetJumpTable(N->getIndex(), Ty, Flag);
103 SDValue MipsTargetLowering::getTargetNode(ConstantPoolSDNode *N, EVT Ty,
105 unsigned Flag) const {
106 return DAG.getTargetConstantPool(N->getConstVal(), Ty, N->getAlignment(),
107 N->getOffset(), Flag);
110 const char *MipsTargetLowering::getTargetNodeName(unsigned Opcode) const {
111 switch ((MipsISD::NodeType)Opcode) {
112 case MipsISD::FIRST_NUMBER: break;
113 case MipsISD::JmpLink: return "MipsISD::JmpLink";
114 case MipsISD::TailCall: return "MipsISD::TailCall";
115 case MipsISD::Hi: return "MipsISD::Hi";
116 case MipsISD::Lo: return "MipsISD::Lo";
117 case MipsISD::GPRel: return "MipsISD::GPRel";
118 case MipsISD::ThreadPointer: return "MipsISD::ThreadPointer";
119 case MipsISD::Ret: return "MipsISD::Ret";
120 case MipsISD::ERet: return "MipsISD::ERet";
121 case MipsISD::EH_RETURN: return "MipsISD::EH_RETURN";
122 case MipsISD::FPBrcond: return "MipsISD::FPBrcond";
123 case MipsISD::FPCmp: return "MipsISD::FPCmp";
124 case MipsISD::CMovFP_T: return "MipsISD::CMovFP_T";
125 case MipsISD::CMovFP_F: return "MipsISD::CMovFP_F";
126 case MipsISD::TruncIntFP: return "MipsISD::TruncIntFP";
127 case MipsISD::MFHI: return "MipsISD::MFHI";
128 case MipsISD::MFLO: return "MipsISD::MFLO";
129 case MipsISD::MTLOHI: return "MipsISD::MTLOHI";
130 case MipsISD::Mult: return "MipsISD::Mult";
131 case MipsISD::Multu: return "MipsISD::Multu";
132 case MipsISD::MAdd: return "MipsISD::MAdd";
133 case MipsISD::MAddu: return "MipsISD::MAddu";
134 case MipsISD::MSub: return "MipsISD::MSub";
135 case MipsISD::MSubu: return "MipsISD::MSubu";
136 case MipsISD::DivRem: return "MipsISD::DivRem";
137 case MipsISD::DivRemU: return "MipsISD::DivRemU";
138 case MipsISD::DivRem16: return "MipsISD::DivRem16";
139 case MipsISD::DivRemU16: return "MipsISD::DivRemU16";
140 case MipsISD::BuildPairF64: return "MipsISD::BuildPairF64";
141 case MipsISD::ExtractElementF64: return "MipsISD::ExtractElementF64";
142 case MipsISD::Wrapper: return "MipsISD::Wrapper";
143 case MipsISD::DynAlloc: return "MipsISD::DynAlloc";
144 case MipsISD::Sync: return "MipsISD::Sync";
145 case MipsISD::Ext: return "MipsISD::Ext";
146 case MipsISD::Ins: return "MipsISD::Ins";
147 case MipsISD::LWL: return "MipsISD::LWL";
148 case MipsISD::LWR: return "MipsISD::LWR";
149 case MipsISD::SWL: return "MipsISD::SWL";
150 case MipsISD::SWR: return "MipsISD::SWR";
151 case MipsISD::LDL: return "MipsISD::LDL";
152 case MipsISD::LDR: return "MipsISD::LDR";
153 case MipsISD::SDL: return "MipsISD::SDL";
154 case MipsISD::SDR: return "MipsISD::SDR";
155 case MipsISD::EXTP: return "MipsISD::EXTP";
156 case MipsISD::EXTPDP: return "MipsISD::EXTPDP";
157 case MipsISD::EXTR_S_H: return "MipsISD::EXTR_S_H";
158 case MipsISD::EXTR_W: return "MipsISD::EXTR_W";
159 case MipsISD::EXTR_R_W: return "MipsISD::EXTR_R_W";
160 case MipsISD::EXTR_RS_W: return "MipsISD::EXTR_RS_W";
161 case MipsISD::SHILO: return "MipsISD::SHILO";
162 case MipsISD::MTHLIP: return "MipsISD::MTHLIP";
163 case MipsISD::MULSAQ_S_W_PH: return "MipsISD::MULSAQ_S_W_PH";
164 case MipsISD::MAQ_S_W_PHL: return "MipsISD::MAQ_S_W_PHL";
165 case MipsISD::MAQ_S_W_PHR: return "MipsISD::MAQ_S_W_PHR";
166 case MipsISD::MAQ_SA_W_PHL: return "MipsISD::MAQ_SA_W_PHL";
167 case MipsISD::MAQ_SA_W_PHR: return "MipsISD::MAQ_SA_W_PHR";
168 case MipsISD::DPAU_H_QBL: return "MipsISD::DPAU_H_QBL";
169 case MipsISD::DPAU_H_QBR: return "MipsISD::DPAU_H_QBR";
170 case MipsISD::DPSU_H_QBL: return "MipsISD::DPSU_H_QBL";
171 case MipsISD::DPSU_H_QBR: return "MipsISD::DPSU_H_QBR";
172 case MipsISD::DPAQ_S_W_PH: return "MipsISD::DPAQ_S_W_PH";
173 case MipsISD::DPSQ_S_W_PH: return "MipsISD::DPSQ_S_W_PH";
174 case MipsISD::DPAQ_SA_L_W: return "MipsISD::DPAQ_SA_L_W";
175 case MipsISD::DPSQ_SA_L_W: return "MipsISD::DPSQ_SA_L_W";
176 case MipsISD::DPA_W_PH: return "MipsISD::DPA_W_PH";
177 case MipsISD::DPS_W_PH: return "MipsISD::DPS_W_PH";
178 case MipsISD::DPAQX_S_W_PH: return "MipsISD::DPAQX_S_W_PH";
179 case MipsISD::DPAQX_SA_W_PH: return "MipsISD::DPAQX_SA_W_PH";
180 case MipsISD::DPAX_W_PH: return "MipsISD::DPAX_W_PH";
181 case MipsISD::DPSX_W_PH: return "MipsISD::DPSX_W_PH";
182 case MipsISD::DPSQX_S_W_PH: return "MipsISD::DPSQX_S_W_PH";
183 case MipsISD::DPSQX_SA_W_PH: return "MipsISD::DPSQX_SA_W_PH";
184 case MipsISD::MULSA_W_PH: return "MipsISD::MULSA_W_PH";
185 case MipsISD::MULT: return "MipsISD::MULT";
186 case MipsISD::MULTU: return "MipsISD::MULTU";
187 case MipsISD::MADD_DSP: return "MipsISD::MADD_DSP";
188 case MipsISD::MADDU_DSP: return "MipsISD::MADDU_DSP";
189 case MipsISD::MSUB_DSP: return "MipsISD::MSUB_DSP";
190 case MipsISD::MSUBU_DSP: return "MipsISD::MSUBU_DSP";
191 case MipsISD::SHLL_DSP: return "MipsISD::SHLL_DSP";
192 case MipsISD::SHRA_DSP: return "MipsISD::SHRA_DSP";
193 case MipsISD::SHRL_DSP: return "MipsISD::SHRL_DSP";
194 case MipsISD::SETCC_DSP: return "MipsISD::SETCC_DSP";
195 case MipsISD::SELECT_CC_DSP: return "MipsISD::SELECT_CC_DSP";
196 case MipsISD::VALL_ZERO: return "MipsISD::VALL_ZERO";
197 case MipsISD::VANY_ZERO: return "MipsISD::VANY_ZERO";
198 case MipsISD::VALL_NONZERO: return "MipsISD::VALL_NONZERO";
199 case MipsISD::VANY_NONZERO: return "MipsISD::VANY_NONZERO";
200 case MipsISD::VCEQ: return "MipsISD::VCEQ";
201 case MipsISD::VCLE_S: return "MipsISD::VCLE_S";
202 case MipsISD::VCLE_U: return "MipsISD::VCLE_U";
203 case MipsISD::VCLT_S: return "MipsISD::VCLT_S";
204 case MipsISD::VCLT_U: return "MipsISD::VCLT_U";
205 case MipsISD::VSMAX: return "MipsISD::VSMAX";
206 case MipsISD::VSMIN: return "MipsISD::VSMIN";
207 case MipsISD::VUMAX: return "MipsISD::VUMAX";
208 case MipsISD::VUMIN: return "MipsISD::VUMIN";
209 case MipsISD::VEXTRACT_SEXT_ELT: return "MipsISD::VEXTRACT_SEXT_ELT";
210 case MipsISD::VEXTRACT_ZEXT_ELT: return "MipsISD::VEXTRACT_ZEXT_ELT";
211 case MipsISD::VNOR: return "MipsISD::VNOR";
212 case MipsISD::VSHF: return "MipsISD::VSHF";
213 case MipsISD::SHF: return "MipsISD::SHF";
214 case MipsISD::ILVEV: return "MipsISD::ILVEV";
215 case MipsISD::ILVOD: return "MipsISD::ILVOD";
216 case MipsISD::ILVL: return "MipsISD::ILVL";
217 case MipsISD::ILVR: return "MipsISD::ILVR";
218 case MipsISD::PCKEV: return "MipsISD::PCKEV";
219 case MipsISD::PCKOD: return "MipsISD::PCKOD";
220 case MipsISD::INSVE: return "MipsISD::INSVE";
225 MipsTargetLowering::MipsTargetLowering(const MipsTargetMachine &TM,
226 const MipsSubtarget &STI)
227 : TargetLowering(TM), Subtarget(STI), ABI(TM.getABI()) {
228 // Mips does not have i1 type, so use i32 for
229 // setcc operations results (slt, sgt, ...).
230 setBooleanContents(ZeroOrOneBooleanContent);
231 setBooleanVectorContents(ZeroOrNegativeOneBooleanContent);
232 // The cmp.cond.fmt instruction in MIPS32r6/MIPS64r6 uses 0 and -1 like MSA
233 // does. Integer booleans still use 0 and 1.
234 if (Subtarget.hasMips32r6())
235 setBooleanContents(ZeroOrOneBooleanContent,
236 ZeroOrNegativeOneBooleanContent);
238 // Load extented operations for i1 types must be promoted
239 for (MVT VT : MVT::integer_valuetypes()) {
240 setLoadExtAction(ISD::EXTLOAD, VT, MVT::i1, Promote);
241 setLoadExtAction(ISD::ZEXTLOAD, VT, MVT::i1, Promote);
242 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::i1, Promote);
245 // MIPS doesn't have extending float->double load/store. Set LoadExtAction
247 for (MVT VT : MVT::fp_valuetypes()) {
248 setLoadExtAction(ISD::EXTLOAD, VT, MVT::f32, Expand);
249 setLoadExtAction(ISD::EXTLOAD, VT, MVT::f16, Expand);
252 // Set LoadExtAction for f16 vectors to Expand
253 for (MVT VT : MVT::fp_vector_valuetypes()) {
254 MVT F16VT = MVT::getVectorVT(MVT::f16, VT.getVectorNumElements());
256 setLoadExtAction(ISD::EXTLOAD, VT, F16VT, Expand);
259 setTruncStoreAction(MVT::f32, MVT::f16, Expand);
260 setTruncStoreAction(MVT::f64, MVT::f16, Expand);
262 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
264 // Used by legalize types to correctly generate the setcc result.
265 // Without this, every float setcc comes with a AND/OR with the result,
266 // we don't want this, since the fpcmp result goes to a flag register,
267 // which is used implicitly by brcond and select operations.
268 AddPromotedToType(ISD::SETCC, MVT::i1, MVT::i32);
270 // Mips Custom Operations
271 setOperationAction(ISD::BR_JT, MVT::Other, Custom);
272 setOperationAction(ISD::GlobalAddress, MVT::i32, Custom);
273 setOperationAction(ISD::BlockAddress, MVT::i32, Custom);
274 setOperationAction(ISD::GlobalTLSAddress, MVT::i32, Custom);
275 setOperationAction(ISD::JumpTable, MVT::i32, Custom);
276 setOperationAction(ISD::ConstantPool, MVT::i32, Custom);
277 setOperationAction(ISD::SELECT, MVT::f32, Custom);
278 setOperationAction(ISD::SELECT, MVT::f64, Custom);
279 setOperationAction(ISD::SELECT, MVT::i32, Custom);
280 setOperationAction(ISD::SELECT_CC, MVT::f32, Custom);
281 setOperationAction(ISD::SELECT_CC, MVT::f64, Custom);
282 setOperationAction(ISD::SETCC, MVT::f32, Custom);
283 setOperationAction(ISD::SETCC, MVT::f64, Custom);
284 setOperationAction(ISD::BRCOND, MVT::Other, Custom);
285 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
286 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
287 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom);
289 if (Subtarget.isGP64bit()) {
290 setOperationAction(ISD::GlobalAddress, MVT::i64, Custom);
291 setOperationAction(ISD::BlockAddress, MVT::i64, Custom);
292 setOperationAction(ISD::GlobalTLSAddress, MVT::i64, Custom);
293 setOperationAction(ISD::JumpTable, MVT::i64, Custom);
294 setOperationAction(ISD::ConstantPool, MVT::i64, Custom);
295 setOperationAction(ISD::SELECT, MVT::i64, Custom);
296 setOperationAction(ISD::LOAD, MVT::i64, Custom);
297 setOperationAction(ISD::STORE, MVT::i64, Custom);
298 setOperationAction(ISD::FP_TO_SINT, MVT::i64, Custom);
299 setOperationAction(ISD::SHL_PARTS, MVT::i64, Custom);
300 setOperationAction(ISD::SRA_PARTS, MVT::i64, Custom);
301 setOperationAction(ISD::SRL_PARTS, MVT::i64, Custom);
304 if (!Subtarget.isGP64bit()) {
305 setOperationAction(ISD::SHL_PARTS, MVT::i32, Custom);
306 setOperationAction(ISD::SRA_PARTS, MVT::i32, Custom);
307 setOperationAction(ISD::SRL_PARTS, MVT::i32, Custom);
310 setOperationAction(ISD::ADD, MVT::i32, Custom);
311 if (Subtarget.isGP64bit())
312 setOperationAction(ISD::ADD, MVT::i64, Custom);
314 setOperationAction(ISD::SDIV, MVT::i32, Expand);
315 setOperationAction(ISD::SREM, MVT::i32, Expand);
316 setOperationAction(ISD::UDIV, MVT::i32, Expand);
317 setOperationAction(ISD::UREM, MVT::i32, Expand);
318 setOperationAction(ISD::SDIV, MVT::i64, Expand);
319 setOperationAction(ISD::SREM, MVT::i64, Expand);
320 setOperationAction(ISD::UDIV, MVT::i64, Expand);
321 setOperationAction(ISD::UREM, MVT::i64, Expand);
323 // Operations not directly supported by Mips.
324 setOperationAction(ISD::BR_CC, MVT::f32, Expand);
325 setOperationAction(ISD::BR_CC, MVT::f64, Expand);
326 setOperationAction(ISD::BR_CC, MVT::i32, Expand);
327 setOperationAction(ISD::BR_CC, MVT::i64, Expand);
328 setOperationAction(ISD::SELECT_CC, MVT::i32, Expand);
329 setOperationAction(ISD::SELECT_CC, MVT::i64, Expand);
330 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Expand);
331 setOperationAction(ISD::UINT_TO_FP, MVT::i64, Expand);
332 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Expand);
333 setOperationAction(ISD::FP_TO_UINT, MVT::i64, Expand);
334 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand);
335 if (Subtarget.hasCnMips()) {
336 setOperationAction(ISD::CTPOP, MVT::i32, Legal);
337 setOperationAction(ISD::CTPOP, MVT::i64, Legal);
339 setOperationAction(ISD::CTPOP, MVT::i32, Expand);
340 setOperationAction(ISD::CTPOP, MVT::i64, Expand);
342 setOperationAction(ISD::CTTZ, MVT::i32, Expand);
343 setOperationAction(ISD::CTTZ, MVT::i64, Expand);
344 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i32, Expand);
345 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i64, Expand);
346 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i32, Expand);
347 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i64, Expand);
348 setOperationAction(ISD::ROTL, MVT::i32, Expand);
349 setOperationAction(ISD::ROTL, MVT::i64, Expand);
350 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand);
351 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i64, Expand);
353 if (!Subtarget.hasMips32r2())
354 setOperationAction(ISD::ROTR, MVT::i32, Expand);
356 if (!Subtarget.hasMips64r2())
357 setOperationAction(ISD::ROTR, MVT::i64, Expand);
359 setOperationAction(ISD::FSIN, MVT::f32, Expand);
360 setOperationAction(ISD::FSIN, MVT::f64, Expand);
361 setOperationAction(ISD::FCOS, MVT::f32, Expand);
362 setOperationAction(ISD::FCOS, MVT::f64, Expand);
363 setOperationAction(ISD::FSINCOS, MVT::f32, Expand);
364 setOperationAction(ISD::FSINCOS, MVT::f64, Expand);
365 setOperationAction(ISD::FPOWI, MVT::f32, Expand);
366 setOperationAction(ISD::FPOW, MVT::f32, Expand);
367 setOperationAction(ISD::FPOW, MVT::f64, Expand);
368 setOperationAction(ISD::FLOG, MVT::f32, Expand);
369 setOperationAction(ISD::FLOG2, MVT::f32, Expand);
370 setOperationAction(ISD::FLOG10, MVT::f32, Expand);
371 setOperationAction(ISD::FEXP, MVT::f32, Expand);
372 setOperationAction(ISD::FMA, MVT::f32, Expand);
373 setOperationAction(ISD::FMA, MVT::f64, Expand);
374 setOperationAction(ISD::FREM, MVT::f32, Expand);
375 setOperationAction(ISD::FREM, MVT::f64, Expand);
377 // Lower f16 conversion operations into library calls
378 setOperationAction(ISD::FP16_TO_FP, MVT::f32, Expand);
379 setOperationAction(ISD::FP_TO_FP16, MVT::f32, Expand);
380 setOperationAction(ISD::FP16_TO_FP, MVT::f64, Expand);
381 setOperationAction(ISD::FP_TO_FP16, MVT::f64, Expand);
383 setOperationAction(ISD::EH_RETURN, MVT::Other, Custom);
385 setOperationAction(ISD::VASTART, MVT::Other, Custom);
386 setOperationAction(ISD::VAARG, MVT::Other, Custom);
387 setOperationAction(ISD::VACOPY, MVT::Other, Expand);
388 setOperationAction(ISD::VAEND, MVT::Other, Expand);
390 // Use the default for now
391 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
392 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
394 if (!Subtarget.isGP64bit()) {
395 setOperationAction(ISD::ATOMIC_LOAD, MVT::i64, Expand);
396 setOperationAction(ISD::ATOMIC_STORE, MVT::i64, Expand);
399 setInsertFencesForAtomic(true);
401 if (!Subtarget.hasMips32r2()) {
402 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8, Expand);
403 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16, Expand);
406 // MIPS16 lacks MIPS32's clz and clo instructions.
407 if (!Subtarget.hasMips32() || Subtarget.inMips16Mode())
408 setOperationAction(ISD::CTLZ, MVT::i32, Expand);
409 if (!Subtarget.hasMips64())
410 setOperationAction(ISD::CTLZ, MVT::i64, Expand);
412 if (!Subtarget.hasMips32r2())
413 setOperationAction(ISD::BSWAP, MVT::i32, Expand);
414 if (!Subtarget.hasMips64r2())
415 setOperationAction(ISD::BSWAP, MVT::i64, Expand);
417 if (Subtarget.isGP64bit()) {
418 setLoadExtAction(ISD::SEXTLOAD, MVT::i64, MVT::i32, Custom);
419 setLoadExtAction(ISD::ZEXTLOAD, MVT::i64, MVT::i32, Custom);
420 setLoadExtAction(ISD::EXTLOAD, MVT::i64, MVT::i32, Custom);
421 setTruncStoreAction(MVT::i64, MVT::i32, Custom);
424 setOperationAction(ISD::TRAP, MVT::Other, Legal);
426 setTargetDAGCombine(ISD::SDIVREM);
427 setTargetDAGCombine(ISD::UDIVREM);
428 setTargetDAGCombine(ISD::SELECT);
429 setTargetDAGCombine(ISD::AND);
430 setTargetDAGCombine(ISD::OR);
431 setTargetDAGCombine(ISD::ADD);
433 setMinFunctionAlignment(Subtarget.isGP64bit() ? 3 : 2);
435 // The arguments on the stack are defined in terms of 4-byte slots on O32
436 // and 8-byte slots on N32/N64.
437 setMinStackArgumentAlignment((ABI.IsN32() || ABI.IsN64()) ? 8 : 4);
439 setStackPointerRegisterToSaveRestore(ABI.IsN64() ? Mips::SP_64 : Mips::SP);
441 MaxStoresPerMemcpy = 16;
443 isMicroMips = Subtarget.inMicroMipsMode();
446 const MipsTargetLowering *MipsTargetLowering::create(const MipsTargetMachine &TM,
447 const MipsSubtarget &STI) {
448 if (STI.inMips16Mode())
449 return llvm::createMips16TargetLowering(TM, STI);
451 return llvm::createMipsSETargetLowering(TM, STI);
454 // Create a fast isel object.
456 MipsTargetLowering::createFastISel(FunctionLoweringInfo &funcInfo,
457 const TargetLibraryInfo *libInfo) const {
458 if (!funcInfo.MF->getTarget().Options.EnableFastISel)
459 return TargetLowering::createFastISel(funcInfo, libInfo);
460 return Mips::createFastISel(funcInfo, libInfo);
463 EVT MipsTargetLowering::getSetCCResultType(const DataLayout &, LLVMContext &,
467 return VT.changeVectorElementTypeToInteger();
470 static SDValue performDivRemCombine(SDNode *N, SelectionDAG &DAG,
471 TargetLowering::DAGCombinerInfo &DCI,
472 const MipsSubtarget &Subtarget) {
473 if (DCI.isBeforeLegalizeOps())
476 EVT Ty = N->getValueType(0);
477 unsigned LO = (Ty == MVT::i32) ? Mips::LO0 : Mips::LO0_64;
478 unsigned HI = (Ty == MVT::i32) ? Mips::HI0 : Mips::HI0_64;
479 unsigned Opc = N->getOpcode() == ISD::SDIVREM ? MipsISD::DivRem16 :
483 SDValue DivRem = DAG.getNode(Opc, DL, MVT::Glue,
484 N->getOperand(0), N->getOperand(1));
485 SDValue InChain = DAG.getEntryNode();
486 SDValue InGlue = DivRem;
489 if (N->hasAnyUseOfValue(0)) {
490 SDValue CopyFromLo = DAG.getCopyFromReg(InChain, DL, LO, Ty,
492 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), CopyFromLo);
493 InChain = CopyFromLo.getValue(1);
494 InGlue = CopyFromLo.getValue(2);
498 if (N->hasAnyUseOfValue(1)) {
499 SDValue CopyFromHi = DAG.getCopyFromReg(InChain, DL,
501 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), CopyFromHi);
507 static Mips::CondCode condCodeToFCC(ISD::CondCode CC) {
509 default: llvm_unreachable("Unknown fp condition code!");
511 case ISD::SETOEQ: return Mips::FCOND_OEQ;
512 case ISD::SETUNE: return Mips::FCOND_UNE;
514 case ISD::SETOLT: return Mips::FCOND_OLT;
516 case ISD::SETOGT: return Mips::FCOND_OGT;
518 case ISD::SETOLE: return Mips::FCOND_OLE;
520 case ISD::SETOGE: return Mips::FCOND_OGE;
521 case ISD::SETULT: return Mips::FCOND_ULT;
522 case ISD::SETULE: return Mips::FCOND_ULE;
523 case ISD::SETUGT: return Mips::FCOND_UGT;
524 case ISD::SETUGE: return Mips::FCOND_UGE;
525 case ISD::SETUO: return Mips::FCOND_UN;
526 case ISD::SETO: return Mips::FCOND_OR;
528 case ISD::SETONE: return Mips::FCOND_ONE;
529 case ISD::SETUEQ: return Mips::FCOND_UEQ;
534 /// This function returns true if the floating point conditional branches and
535 /// conditional moves which use condition code CC should be inverted.
536 static bool invertFPCondCodeUser(Mips::CondCode CC) {
537 if (CC >= Mips::FCOND_F && CC <= Mips::FCOND_NGT)
540 assert((CC >= Mips::FCOND_T && CC <= Mips::FCOND_GT) &&
541 "Illegal Condition Code");
546 // Creates and returns an FPCmp node from a setcc node.
547 // Returns Op if setcc is not a floating point comparison.
548 static SDValue createFPCmp(SelectionDAG &DAG, const SDValue &Op) {
549 // must be a SETCC node
550 if (Op.getOpcode() != ISD::SETCC)
553 SDValue LHS = Op.getOperand(0);
555 if (!LHS.getValueType().isFloatingPoint())
558 SDValue RHS = Op.getOperand(1);
561 // Assume the 3rd operand is a CondCodeSDNode. Add code to check the type of
562 // node if necessary.
563 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
565 return DAG.getNode(MipsISD::FPCmp, DL, MVT::Glue, LHS, RHS,
566 DAG.getConstant(condCodeToFCC(CC), DL, MVT::i32));
569 // Creates and returns a CMovFPT/F node.
570 static SDValue createCMovFP(SelectionDAG &DAG, SDValue Cond, SDValue True,
571 SDValue False, SDLoc DL) {
572 ConstantSDNode *CC = cast<ConstantSDNode>(Cond.getOperand(2));
573 bool invert = invertFPCondCodeUser((Mips::CondCode)CC->getSExtValue());
574 SDValue FCC0 = DAG.getRegister(Mips::FCC0, MVT::i32);
576 return DAG.getNode((invert ? MipsISD::CMovFP_F : MipsISD::CMovFP_T), DL,
577 True.getValueType(), True, FCC0, False, Cond);
580 static SDValue performSELECTCombine(SDNode *N, SelectionDAG &DAG,
581 TargetLowering::DAGCombinerInfo &DCI,
582 const MipsSubtarget &Subtarget) {
583 if (DCI.isBeforeLegalizeOps())
586 SDValue SetCC = N->getOperand(0);
588 if ((SetCC.getOpcode() != ISD::SETCC) ||
589 !SetCC.getOperand(0).getValueType().isInteger())
592 SDValue False = N->getOperand(2);
593 EVT FalseTy = False.getValueType();
595 if (!FalseTy.isInteger())
598 ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(False);
600 // If the RHS (False) is 0, we swap the order of the operands
601 // of ISD::SELECT (obviously also inverting the condition) so that we can
602 // take advantage of conditional moves using the $0 register.
604 // return (a != 0) ? x : 0;
612 if (!FalseC->getZExtValue()) {
613 ISD::CondCode CC = cast<CondCodeSDNode>(SetCC.getOperand(2))->get();
614 SDValue True = N->getOperand(1);
616 SetCC = DAG.getSetCC(DL, SetCC.getValueType(), SetCC.getOperand(0),
617 SetCC.getOperand(1), ISD::getSetCCInverse(CC, true));
619 return DAG.getNode(ISD::SELECT, DL, FalseTy, SetCC, False, True);
622 // If both operands are integer constants there's a possibility that we
623 // can do some interesting optimizations.
624 SDValue True = N->getOperand(1);
625 ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(True);
627 if (!TrueC || !True.getValueType().isInteger())
630 // We'll also ignore MVT::i64 operands as this optimizations proves
631 // to be ineffective because of the required sign extensions as the result
632 // of a SETCC operator is always MVT::i32 for non-vector types.
633 if (True.getValueType() == MVT::i64)
636 int64_t Diff = TrueC->getSExtValue() - FalseC->getSExtValue();
638 // 1) (a < x) ? y : y-1
640 // addiu $reg2, $reg1, y-1
642 return DAG.getNode(ISD::ADD, DL, SetCC.getValueType(), SetCC, False);
644 // 2) (a < x) ? y-1 : y
646 // xor $reg1, $reg1, 1
647 // addiu $reg2, $reg1, y-1
649 ISD::CondCode CC = cast<CondCodeSDNode>(SetCC.getOperand(2))->get();
650 SetCC = DAG.getSetCC(DL, SetCC.getValueType(), SetCC.getOperand(0),
651 SetCC.getOperand(1), ISD::getSetCCInverse(CC, true));
652 return DAG.getNode(ISD::ADD, DL, SetCC.getValueType(), SetCC, True);
655 // Couldn't optimize.
659 static SDValue performCMovFPCombine(SDNode *N, SelectionDAG &DAG,
660 TargetLowering::DAGCombinerInfo &DCI,
661 const MipsSubtarget &Subtarget) {
662 if (DCI.isBeforeLegalizeOps())
665 SDValue ValueIfTrue = N->getOperand(0), ValueIfFalse = N->getOperand(2);
667 ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(ValueIfFalse);
668 if (!FalseC || FalseC->getZExtValue())
671 // Since RHS (False) is 0, we swap the order of the True/False operands
672 // (obviously also inverting the condition) so that we can
673 // take advantage of conditional moves using the $0 register.
675 // return (a != 0) ? x : 0;
678 unsigned Opc = (N->getOpcode() == MipsISD::CMovFP_T) ? MipsISD::CMovFP_F :
681 SDValue FCC = N->getOperand(1), Glue = N->getOperand(3);
682 return DAG.getNode(Opc, SDLoc(N), ValueIfFalse.getValueType(),
683 ValueIfFalse, FCC, ValueIfTrue, Glue);
686 static SDValue performANDCombine(SDNode *N, SelectionDAG &DAG,
687 TargetLowering::DAGCombinerInfo &DCI,
688 const MipsSubtarget &Subtarget) {
689 // Pattern match EXT.
690 // $dst = and ((sra or srl) $src , pos), (2**size - 1)
691 // => ext $dst, $src, size, pos
692 if (DCI.isBeforeLegalizeOps() || !Subtarget.hasExtractInsert())
695 SDValue ShiftRight = N->getOperand(0), Mask = N->getOperand(1);
696 unsigned ShiftRightOpc = ShiftRight.getOpcode();
698 // Op's first operand must be a shift right.
699 if (ShiftRightOpc != ISD::SRA && ShiftRightOpc != ISD::SRL)
702 // The second operand of the shift must be an immediate.
704 if (!(CN = dyn_cast<ConstantSDNode>(ShiftRight.getOperand(1))))
707 uint64_t Pos = CN->getZExtValue();
708 uint64_t SMPos, SMSize;
710 // Op's second operand must be a shifted mask.
711 if (!(CN = dyn_cast<ConstantSDNode>(Mask)) ||
712 !isShiftedMask(CN->getZExtValue(), SMPos, SMSize))
715 // Return if the shifted mask does not start at bit 0 or the sum of its size
716 // and Pos exceeds the word's size.
717 EVT ValTy = N->getValueType(0);
718 if (SMPos != 0 || Pos + SMSize > ValTy.getSizeInBits())
722 return DAG.getNode(MipsISD::Ext, DL, ValTy,
723 ShiftRight.getOperand(0),
724 DAG.getConstant(Pos, DL, MVT::i32),
725 DAG.getConstant(SMSize, DL, MVT::i32));
728 static SDValue performORCombine(SDNode *N, SelectionDAG &DAG,
729 TargetLowering::DAGCombinerInfo &DCI,
730 const MipsSubtarget &Subtarget) {
731 // Pattern match INS.
732 // $dst = or (and $src1 , mask0), (and (shl $src, pos), mask1),
733 // where mask1 = (2**size - 1) << pos, mask0 = ~mask1
734 // => ins $dst, $src, size, pos, $src1
735 if (DCI.isBeforeLegalizeOps() || !Subtarget.hasExtractInsert())
738 SDValue And0 = N->getOperand(0), And1 = N->getOperand(1);
739 uint64_t SMPos0, SMSize0, SMPos1, SMSize1;
742 // See if Op's first operand matches (and $src1 , mask0).
743 if (And0.getOpcode() != ISD::AND)
746 if (!(CN = dyn_cast<ConstantSDNode>(And0.getOperand(1))) ||
747 !isShiftedMask(~CN->getSExtValue(), SMPos0, SMSize0))
750 // See if Op's second operand matches (and (shl $src, pos), mask1).
751 if (And1.getOpcode() != ISD::AND)
754 if (!(CN = dyn_cast<ConstantSDNode>(And1.getOperand(1))) ||
755 !isShiftedMask(CN->getZExtValue(), SMPos1, SMSize1))
758 // The shift masks must have the same position and size.
759 if (SMPos0 != SMPos1 || SMSize0 != SMSize1)
762 SDValue Shl = And1.getOperand(0);
763 if (Shl.getOpcode() != ISD::SHL)
766 if (!(CN = dyn_cast<ConstantSDNode>(Shl.getOperand(1))))
769 unsigned Shamt = CN->getZExtValue();
771 // Return if the shift amount and the first bit position of mask are not the
773 EVT ValTy = N->getValueType(0);
774 if ((Shamt != SMPos0) || (SMPos0 + SMSize0 > ValTy.getSizeInBits()))
778 return DAG.getNode(MipsISD::Ins, DL, ValTy, Shl.getOperand(0),
779 DAG.getConstant(SMPos0, DL, MVT::i32),
780 DAG.getConstant(SMSize0, DL, MVT::i32),
784 static SDValue performADDCombine(SDNode *N, SelectionDAG &DAG,
785 TargetLowering::DAGCombinerInfo &DCI,
786 const MipsSubtarget &Subtarget) {
787 // (add v0, (add v1, abs_lo(tjt))) => (add (add v0, v1), abs_lo(tjt))
789 if (DCI.isBeforeLegalizeOps())
792 SDValue Add = N->getOperand(1);
794 if (Add.getOpcode() != ISD::ADD)
797 SDValue Lo = Add.getOperand(1);
799 if ((Lo.getOpcode() != MipsISD::Lo) ||
800 (Lo.getOperand(0).getOpcode() != ISD::TargetJumpTable))
803 EVT ValTy = N->getValueType(0);
806 SDValue Add1 = DAG.getNode(ISD::ADD, DL, ValTy, N->getOperand(0),
808 return DAG.getNode(ISD::ADD, DL, ValTy, Add1, Lo);
811 SDValue MipsTargetLowering::PerformDAGCombine(SDNode *N, DAGCombinerInfo &DCI)
813 SelectionDAG &DAG = DCI.DAG;
814 unsigned Opc = N->getOpcode();
820 return performDivRemCombine(N, DAG, DCI, Subtarget);
822 return performSELECTCombine(N, DAG, DCI, Subtarget);
823 case MipsISD::CMovFP_F:
824 case MipsISD::CMovFP_T:
825 return performCMovFPCombine(N, DAG, DCI, Subtarget);
827 return performANDCombine(N, DAG, DCI, Subtarget);
829 return performORCombine(N, DAG, DCI, Subtarget);
831 return performADDCombine(N, DAG, DCI, Subtarget);
838 MipsTargetLowering::LowerOperationWrapper(SDNode *N,
839 SmallVectorImpl<SDValue> &Results,
840 SelectionDAG &DAG) const {
841 SDValue Res = LowerOperation(SDValue(N, 0), DAG);
843 for (unsigned I = 0, E = Res->getNumValues(); I != E; ++I)
844 Results.push_back(Res.getValue(I));
848 MipsTargetLowering::ReplaceNodeResults(SDNode *N,
849 SmallVectorImpl<SDValue> &Results,
850 SelectionDAG &DAG) const {
851 return LowerOperationWrapper(N, Results, DAG);
854 SDValue MipsTargetLowering::
855 LowerOperation(SDValue Op, SelectionDAG &DAG) const
857 switch (Op.getOpcode())
859 case ISD::BR_JT: return lowerBR_JT(Op, DAG);
860 case ISD::BRCOND: return lowerBRCOND(Op, DAG);
861 case ISD::ConstantPool: return lowerConstantPool(Op, DAG);
862 case ISD::GlobalAddress: return lowerGlobalAddress(Op, DAG);
863 case ISD::BlockAddress: return lowerBlockAddress(Op, DAG);
864 case ISD::GlobalTLSAddress: return lowerGlobalTLSAddress(Op, DAG);
865 case ISD::JumpTable: return lowerJumpTable(Op, DAG);
866 case ISD::SELECT: return lowerSELECT(Op, DAG);
867 case ISD::SELECT_CC: return lowerSELECT_CC(Op, DAG);
868 case ISD::SETCC: return lowerSETCC(Op, DAG);
869 case ISD::VASTART: return lowerVASTART(Op, DAG);
870 case ISD::VAARG: return lowerVAARG(Op, DAG);
871 case ISD::FCOPYSIGN: return lowerFCOPYSIGN(Op, DAG);
872 case ISD::FRAMEADDR: return lowerFRAMEADDR(Op, DAG);
873 case ISD::RETURNADDR: return lowerRETURNADDR(Op, DAG);
874 case ISD::EH_RETURN: return lowerEH_RETURN(Op, DAG);
875 case ISD::ATOMIC_FENCE: return lowerATOMIC_FENCE(Op, DAG);
876 case ISD::SHL_PARTS: return lowerShiftLeftParts(Op, DAG);
877 case ISD::SRA_PARTS: return lowerShiftRightParts(Op, DAG, true);
878 case ISD::SRL_PARTS: return lowerShiftRightParts(Op, DAG, false);
879 case ISD::LOAD: return lowerLOAD(Op, DAG);
880 case ISD::STORE: return lowerSTORE(Op, DAG);
881 case ISD::ADD: return lowerADD(Op, DAG);
882 case ISD::FP_TO_SINT: return lowerFP_TO_SINT(Op, DAG);
887 //===----------------------------------------------------------------------===//
888 // Lower helper functions
889 //===----------------------------------------------------------------------===//
891 // addLiveIn - This helper function adds the specified physical register to the
892 // MachineFunction as a live in value. It also creates a corresponding
893 // virtual register for it.
895 addLiveIn(MachineFunction &MF, unsigned PReg, const TargetRegisterClass *RC)
897 unsigned VReg = MF.getRegInfo().createVirtualRegister(RC);
898 MF.getRegInfo().addLiveIn(PReg, VReg);
902 static MachineBasicBlock *insertDivByZeroTrap(MachineInstr *MI,
903 MachineBasicBlock &MBB,
904 const TargetInstrInfo &TII,
909 // Insert instruction "teq $divisor_reg, $zero, 7".
910 MachineBasicBlock::iterator I(MI);
911 MachineInstrBuilder MIB;
912 MachineOperand &Divisor = MI->getOperand(2);
913 MIB = BuildMI(MBB, std::next(I), MI->getDebugLoc(), TII.get(Mips::TEQ))
914 .addReg(Divisor.getReg(), getKillRegState(Divisor.isKill()))
915 .addReg(Mips::ZERO).addImm(7);
917 // Use the 32-bit sub-register if this is a 64-bit division.
919 MIB->getOperand(0).setSubReg(Mips::sub_32);
921 // Clear Divisor's kill flag.
922 Divisor.setIsKill(false);
924 // We would normally delete the original instruction here but in this case
925 // we only needed to inject an additional instruction rather than replace it.
931 MipsTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
932 MachineBasicBlock *BB) const {
933 switch (MI->getOpcode()) {
935 llvm_unreachable("Unexpected instr type to insert");
936 case Mips::ATOMIC_LOAD_ADD_I8:
937 return emitAtomicBinaryPartword(MI, BB, 1, Mips::ADDu);
938 case Mips::ATOMIC_LOAD_ADD_I16:
939 return emitAtomicBinaryPartword(MI, BB, 2, Mips::ADDu);
940 case Mips::ATOMIC_LOAD_ADD_I32:
941 return emitAtomicBinary(MI, BB, 4, Mips::ADDu);
942 case Mips::ATOMIC_LOAD_ADD_I64:
943 return emitAtomicBinary(MI, BB, 8, Mips::DADDu);
945 case Mips::ATOMIC_LOAD_AND_I8:
946 return emitAtomicBinaryPartword(MI, BB, 1, Mips::AND);
947 case Mips::ATOMIC_LOAD_AND_I16:
948 return emitAtomicBinaryPartword(MI, BB, 2, Mips::AND);
949 case Mips::ATOMIC_LOAD_AND_I32:
950 return emitAtomicBinary(MI, BB, 4, Mips::AND);
951 case Mips::ATOMIC_LOAD_AND_I64:
952 return emitAtomicBinary(MI, BB, 8, Mips::AND64);
954 case Mips::ATOMIC_LOAD_OR_I8:
955 return emitAtomicBinaryPartword(MI, BB, 1, Mips::OR);
956 case Mips::ATOMIC_LOAD_OR_I16:
957 return emitAtomicBinaryPartword(MI, BB, 2, Mips::OR);
958 case Mips::ATOMIC_LOAD_OR_I32:
959 return emitAtomicBinary(MI, BB, 4, Mips::OR);
960 case Mips::ATOMIC_LOAD_OR_I64:
961 return emitAtomicBinary(MI, BB, 8, Mips::OR64);
963 case Mips::ATOMIC_LOAD_XOR_I8:
964 return emitAtomicBinaryPartword(MI, BB, 1, Mips::XOR);
965 case Mips::ATOMIC_LOAD_XOR_I16:
966 return emitAtomicBinaryPartword(MI, BB, 2, Mips::XOR);
967 case Mips::ATOMIC_LOAD_XOR_I32:
968 return emitAtomicBinary(MI, BB, 4, Mips::XOR);
969 case Mips::ATOMIC_LOAD_XOR_I64:
970 return emitAtomicBinary(MI, BB, 8, Mips::XOR64);
972 case Mips::ATOMIC_LOAD_NAND_I8:
973 return emitAtomicBinaryPartword(MI, BB, 1, 0, true);
974 case Mips::ATOMIC_LOAD_NAND_I16:
975 return emitAtomicBinaryPartword(MI, BB, 2, 0, true);
976 case Mips::ATOMIC_LOAD_NAND_I32:
977 return emitAtomicBinary(MI, BB, 4, 0, true);
978 case Mips::ATOMIC_LOAD_NAND_I64:
979 return emitAtomicBinary(MI, BB, 8, 0, true);
981 case Mips::ATOMIC_LOAD_SUB_I8:
982 return emitAtomicBinaryPartword(MI, BB, 1, Mips::SUBu);
983 case Mips::ATOMIC_LOAD_SUB_I16:
984 return emitAtomicBinaryPartword(MI, BB, 2, Mips::SUBu);
985 case Mips::ATOMIC_LOAD_SUB_I32:
986 return emitAtomicBinary(MI, BB, 4, Mips::SUBu);
987 case Mips::ATOMIC_LOAD_SUB_I64:
988 return emitAtomicBinary(MI, BB, 8, Mips::DSUBu);
990 case Mips::ATOMIC_SWAP_I8:
991 return emitAtomicBinaryPartword(MI, BB, 1, 0);
992 case Mips::ATOMIC_SWAP_I16:
993 return emitAtomicBinaryPartword(MI, BB, 2, 0);
994 case Mips::ATOMIC_SWAP_I32:
995 return emitAtomicBinary(MI, BB, 4, 0);
996 case Mips::ATOMIC_SWAP_I64:
997 return emitAtomicBinary(MI, BB, 8, 0);
999 case Mips::ATOMIC_CMP_SWAP_I8:
1000 return emitAtomicCmpSwapPartword(MI, BB, 1);
1001 case Mips::ATOMIC_CMP_SWAP_I16:
1002 return emitAtomicCmpSwapPartword(MI, BB, 2);
1003 case Mips::ATOMIC_CMP_SWAP_I32:
1004 return emitAtomicCmpSwap(MI, BB, 4);
1005 case Mips::ATOMIC_CMP_SWAP_I64:
1006 return emitAtomicCmpSwap(MI, BB, 8);
1007 case Mips::PseudoSDIV:
1008 case Mips::PseudoUDIV:
1013 return insertDivByZeroTrap(MI, *BB, *Subtarget.getInstrInfo(), false);
1014 case Mips::PseudoDSDIV:
1015 case Mips::PseudoDUDIV:
1020 return insertDivByZeroTrap(MI, *BB, *Subtarget.getInstrInfo(), true);
1022 return emitSEL_D(MI, BB);
1024 case Mips::PseudoSELECT_I:
1025 case Mips::PseudoSELECT_I64:
1026 case Mips::PseudoSELECT_S:
1027 case Mips::PseudoSELECT_D32:
1028 case Mips::PseudoSELECT_D64:
1029 return emitPseudoSELECT(MI, BB, false, Mips::BNE);
1030 case Mips::PseudoSELECTFP_F_I:
1031 case Mips::PseudoSELECTFP_F_I64:
1032 case Mips::PseudoSELECTFP_F_S:
1033 case Mips::PseudoSELECTFP_F_D32:
1034 case Mips::PseudoSELECTFP_F_D64:
1035 return emitPseudoSELECT(MI, BB, true, Mips::BC1F);
1036 case Mips::PseudoSELECTFP_T_I:
1037 case Mips::PseudoSELECTFP_T_I64:
1038 case Mips::PseudoSELECTFP_T_S:
1039 case Mips::PseudoSELECTFP_T_D32:
1040 case Mips::PseudoSELECTFP_T_D64:
1041 return emitPseudoSELECT(MI, BB, true, Mips::BC1T);
1045 // This function also handles Mips::ATOMIC_SWAP_I32 (when BinOpcode == 0), and
1046 // Mips::ATOMIC_LOAD_NAND_I32 (when Nand == true)
1048 MipsTargetLowering::emitAtomicBinary(MachineInstr *MI, MachineBasicBlock *BB,
1049 unsigned Size, unsigned BinOpcode,
1051 assert((Size == 4 || Size == 8) && "Unsupported size for EmitAtomicBinary.");
1053 MachineFunction *MF = BB->getParent();
1054 MachineRegisterInfo &RegInfo = MF->getRegInfo();
1055 const TargetRegisterClass *RC = getRegClassFor(MVT::getIntegerVT(Size * 8));
1056 const TargetInstrInfo *TII = Subtarget.getInstrInfo();
1057 DebugLoc DL = MI->getDebugLoc();
1058 unsigned LL, SC, AND, NOR, ZERO, BEQ;
1065 LL = Subtarget.hasMips32r6() ? Mips::LL_R6 : Mips::LL;
1066 SC = Subtarget.hasMips32r6() ? Mips::SC_R6 : Mips::SC;
1073 LL = Subtarget.hasMips64r6() ? Mips::LLD_R6 : Mips::LLD;
1074 SC = Subtarget.hasMips64r6() ? Mips::SCD_R6 : Mips::SCD;
1077 ZERO = Mips::ZERO_64;
1081 unsigned OldVal = MI->getOperand(0).getReg();
1082 unsigned Ptr = MI->getOperand(1).getReg();
1083 unsigned Incr = MI->getOperand(2).getReg();
1085 unsigned StoreVal = RegInfo.createVirtualRegister(RC);
1086 unsigned AndRes = RegInfo.createVirtualRegister(RC);
1087 unsigned Success = RegInfo.createVirtualRegister(RC);
1089 // insert new blocks after the current block
1090 const BasicBlock *LLVM_BB = BB->getBasicBlock();
1091 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
1092 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
1093 MachineFunction::iterator It = ++BB->getIterator();
1094 MF->insert(It, loopMBB);
1095 MF->insert(It, exitMBB);
1097 // Transfer the remainder of BB and its successor edges to exitMBB.
1098 exitMBB->splice(exitMBB->begin(), BB,
1099 std::next(MachineBasicBlock::iterator(MI)), BB->end());
1100 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
1104 // fallthrough --> loopMBB
1105 BB->addSuccessor(loopMBB);
1106 loopMBB->addSuccessor(loopMBB);
1107 loopMBB->addSuccessor(exitMBB);
1110 // ll oldval, 0(ptr)
1111 // <binop> storeval, oldval, incr
1112 // sc success, storeval, 0(ptr)
1113 // beq success, $0, loopMBB
1115 BuildMI(BB, DL, TII->get(LL), OldVal).addReg(Ptr).addImm(0);
1117 // and andres, oldval, incr
1118 // nor storeval, $0, andres
1119 BuildMI(BB, DL, TII->get(AND), AndRes).addReg(OldVal).addReg(Incr);
1120 BuildMI(BB, DL, TII->get(NOR), StoreVal).addReg(ZERO).addReg(AndRes);
1121 } else if (BinOpcode) {
1122 // <binop> storeval, oldval, incr
1123 BuildMI(BB, DL, TII->get(BinOpcode), StoreVal).addReg(OldVal).addReg(Incr);
1127 BuildMI(BB, DL, TII->get(SC), Success).addReg(StoreVal).addReg(Ptr).addImm(0);
1128 BuildMI(BB, DL, TII->get(BEQ)).addReg(Success).addReg(ZERO).addMBB(loopMBB);
1130 MI->eraseFromParent(); // The instruction is gone now.
1135 MachineBasicBlock *MipsTargetLowering::emitSignExtendToI32InReg(
1136 MachineInstr *MI, MachineBasicBlock *BB, unsigned Size, unsigned DstReg,
1137 unsigned SrcReg) const {
1138 const TargetInstrInfo *TII = Subtarget.getInstrInfo();
1139 DebugLoc DL = MI->getDebugLoc();
1141 if (Subtarget.hasMips32r2() && Size == 1) {
1142 BuildMI(BB, DL, TII->get(Mips::SEB), DstReg).addReg(SrcReg);
1146 if (Subtarget.hasMips32r2() && Size == 2) {
1147 BuildMI(BB, DL, TII->get(Mips::SEH), DstReg).addReg(SrcReg);
1151 MachineFunction *MF = BB->getParent();
1152 MachineRegisterInfo &RegInfo = MF->getRegInfo();
1153 const TargetRegisterClass *RC = getRegClassFor(MVT::i32);
1154 unsigned ScrReg = RegInfo.createVirtualRegister(RC);
1157 int64_t ShiftImm = 32 - (Size * 8);
1159 BuildMI(BB, DL, TII->get(Mips::SLL), ScrReg).addReg(SrcReg).addImm(ShiftImm);
1160 BuildMI(BB, DL, TII->get(Mips::SRA), DstReg).addReg(ScrReg).addImm(ShiftImm);
1165 MachineBasicBlock *MipsTargetLowering::emitAtomicBinaryPartword(
1166 MachineInstr *MI, MachineBasicBlock *BB, unsigned Size, unsigned BinOpcode,
1168 assert((Size == 1 || Size == 2) &&
1169 "Unsupported size for EmitAtomicBinaryPartial.");
1171 MachineFunction *MF = BB->getParent();
1172 MachineRegisterInfo &RegInfo = MF->getRegInfo();
1173 const TargetRegisterClass *RC = getRegClassFor(MVT::i32);
1174 const TargetInstrInfo *TII = Subtarget.getInstrInfo();
1175 DebugLoc DL = MI->getDebugLoc();
1177 unsigned Dest = MI->getOperand(0).getReg();
1178 unsigned Ptr = MI->getOperand(1).getReg();
1179 unsigned Incr = MI->getOperand(2).getReg();
1181 unsigned AlignedAddr = RegInfo.createVirtualRegister(RC);
1182 unsigned ShiftAmt = RegInfo.createVirtualRegister(RC);
1183 unsigned Mask = RegInfo.createVirtualRegister(RC);
1184 unsigned Mask2 = RegInfo.createVirtualRegister(RC);
1185 unsigned NewVal = RegInfo.createVirtualRegister(RC);
1186 unsigned OldVal = RegInfo.createVirtualRegister(RC);
1187 unsigned Incr2 = RegInfo.createVirtualRegister(RC);
1188 unsigned MaskLSB2 = RegInfo.createVirtualRegister(RC);
1189 unsigned PtrLSB2 = RegInfo.createVirtualRegister(RC);
1190 unsigned MaskUpper = RegInfo.createVirtualRegister(RC);
1191 unsigned AndRes = RegInfo.createVirtualRegister(RC);
1192 unsigned BinOpRes = RegInfo.createVirtualRegister(RC);
1193 unsigned MaskedOldVal0 = RegInfo.createVirtualRegister(RC);
1194 unsigned StoreVal = RegInfo.createVirtualRegister(RC);
1195 unsigned MaskedOldVal1 = RegInfo.createVirtualRegister(RC);
1196 unsigned SrlRes = RegInfo.createVirtualRegister(RC);
1197 unsigned Success = RegInfo.createVirtualRegister(RC);
1199 // insert new blocks after the current block
1200 const BasicBlock *LLVM_BB = BB->getBasicBlock();
1201 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
1202 MachineBasicBlock *sinkMBB = MF->CreateMachineBasicBlock(LLVM_BB);
1203 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
1204 MachineFunction::iterator It = ++BB->getIterator();
1205 MF->insert(It, loopMBB);
1206 MF->insert(It, sinkMBB);
1207 MF->insert(It, exitMBB);
1209 // Transfer the remainder of BB and its successor edges to exitMBB.
1210 exitMBB->splice(exitMBB->begin(), BB,
1211 std::next(MachineBasicBlock::iterator(MI)), BB->end());
1212 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
1214 BB->addSuccessor(loopMBB);
1215 loopMBB->addSuccessor(loopMBB);
1216 loopMBB->addSuccessor(sinkMBB);
1217 sinkMBB->addSuccessor(exitMBB);
1220 // addiu masklsb2,$0,-4 # 0xfffffffc
1221 // and alignedaddr,ptr,masklsb2
1222 // andi ptrlsb2,ptr,3
1223 // sll shiftamt,ptrlsb2,3
1224 // ori maskupper,$0,255 # 0xff
1225 // sll mask,maskupper,shiftamt
1226 // nor mask2,$0,mask
1227 // sll incr2,incr,shiftamt
1229 int64_t MaskImm = (Size == 1) ? 255 : 65535;
1230 BuildMI(BB, DL, TII->get(Mips::ADDiu), MaskLSB2)
1231 .addReg(Mips::ZERO).addImm(-4);
1232 BuildMI(BB, DL, TII->get(Mips::AND), AlignedAddr)
1233 .addReg(Ptr).addReg(MaskLSB2);
1234 BuildMI(BB, DL, TII->get(Mips::ANDi), PtrLSB2).addReg(Ptr).addImm(3);
1235 if (Subtarget.isLittle()) {
1236 BuildMI(BB, DL, TII->get(Mips::SLL), ShiftAmt).addReg(PtrLSB2).addImm(3);
1238 unsigned Off = RegInfo.createVirtualRegister(RC);
1239 BuildMI(BB, DL, TII->get(Mips::XORi), Off)
1240 .addReg(PtrLSB2).addImm((Size == 1) ? 3 : 2);
1241 BuildMI(BB, DL, TII->get(Mips::SLL), ShiftAmt).addReg(Off).addImm(3);
1243 BuildMI(BB, DL, TII->get(Mips::ORi), MaskUpper)
1244 .addReg(Mips::ZERO).addImm(MaskImm);
1245 BuildMI(BB, DL, TII->get(Mips::SLLV), Mask)
1246 .addReg(MaskUpper).addReg(ShiftAmt);
1247 BuildMI(BB, DL, TII->get(Mips::NOR), Mask2).addReg(Mips::ZERO).addReg(Mask);
1248 BuildMI(BB, DL, TII->get(Mips::SLLV), Incr2).addReg(Incr).addReg(ShiftAmt);
1250 // atomic.load.binop
1252 // ll oldval,0(alignedaddr)
1253 // binop binopres,oldval,incr2
1254 // and newval,binopres,mask
1255 // and maskedoldval0,oldval,mask2
1256 // or storeval,maskedoldval0,newval
1257 // sc success,storeval,0(alignedaddr)
1258 // beq success,$0,loopMBB
1262 // ll oldval,0(alignedaddr)
1263 // and newval,incr2,mask
1264 // and maskedoldval0,oldval,mask2
1265 // or storeval,maskedoldval0,newval
1266 // sc success,storeval,0(alignedaddr)
1267 // beq success,$0,loopMBB
1270 unsigned LL = isMicroMips ? Mips::LL_MM : Mips::LL;
1271 BuildMI(BB, DL, TII->get(LL), OldVal).addReg(AlignedAddr).addImm(0);
1273 // and andres, oldval, incr2
1274 // nor binopres, $0, andres
1275 // and newval, binopres, mask
1276 BuildMI(BB, DL, TII->get(Mips::AND), AndRes).addReg(OldVal).addReg(Incr2);
1277 BuildMI(BB, DL, TII->get(Mips::NOR), BinOpRes)
1278 .addReg(Mips::ZERO).addReg(AndRes);
1279 BuildMI(BB, DL, TII->get(Mips::AND), NewVal).addReg(BinOpRes).addReg(Mask);
1280 } else if (BinOpcode) {
1281 // <binop> binopres, oldval, incr2
1282 // and newval, binopres, mask
1283 BuildMI(BB, DL, TII->get(BinOpcode), BinOpRes).addReg(OldVal).addReg(Incr2);
1284 BuildMI(BB, DL, TII->get(Mips::AND), NewVal).addReg(BinOpRes).addReg(Mask);
1285 } else { // atomic.swap
1286 // and newval, incr2, mask
1287 BuildMI(BB, DL, TII->get(Mips::AND), NewVal).addReg(Incr2).addReg(Mask);
1290 BuildMI(BB, DL, TII->get(Mips::AND), MaskedOldVal0)
1291 .addReg(OldVal).addReg(Mask2);
1292 BuildMI(BB, DL, TII->get(Mips::OR), StoreVal)
1293 .addReg(MaskedOldVal0).addReg(NewVal);
1294 unsigned SC = isMicroMips ? Mips::SC_MM : Mips::SC;
1295 BuildMI(BB, DL, TII->get(SC), Success)
1296 .addReg(StoreVal).addReg(AlignedAddr).addImm(0);
1297 BuildMI(BB, DL, TII->get(Mips::BEQ))
1298 .addReg(Success).addReg(Mips::ZERO).addMBB(loopMBB);
1301 // and maskedoldval1,oldval,mask
1302 // srl srlres,maskedoldval1,shiftamt
1303 // sign_extend dest,srlres
1306 BuildMI(BB, DL, TII->get(Mips::AND), MaskedOldVal1)
1307 .addReg(OldVal).addReg(Mask);
1308 BuildMI(BB, DL, TII->get(Mips::SRLV), SrlRes)
1309 .addReg(MaskedOldVal1).addReg(ShiftAmt);
1310 BB = emitSignExtendToI32InReg(MI, BB, Size, Dest, SrlRes);
1312 MI->eraseFromParent(); // The instruction is gone now.
1317 MachineBasicBlock * MipsTargetLowering::emitAtomicCmpSwap(MachineInstr *MI,
1318 MachineBasicBlock *BB,
1319 unsigned Size) const {
1320 assert((Size == 4 || Size == 8) && "Unsupported size for EmitAtomicCmpSwap.");
1322 MachineFunction *MF = BB->getParent();
1323 MachineRegisterInfo &RegInfo = MF->getRegInfo();
1324 const TargetRegisterClass *RC = getRegClassFor(MVT::getIntegerVT(Size * 8));
1325 const TargetInstrInfo *TII = Subtarget.getInstrInfo();
1326 DebugLoc DL = MI->getDebugLoc();
1327 unsigned LL, SC, ZERO, BNE, BEQ;
1334 LL = Subtarget.hasMips32r6() ? Mips::LL_R6 : Mips::LL;
1335 SC = Subtarget.hasMips32r6() ? Mips::SC_R6 : Mips::SC;
1341 LL = Subtarget.hasMips64r6() ? Mips::LLD_R6 : Mips::LLD;
1342 SC = Subtarget.hasMips64r6() ? Mips::SCD_R6 : Mips::SCD;
1343 ZERO = Mips::ZERO_64;
1348 unsigned Dest = MI->getOperand(0).getReg();
1349 unsigned Ptr = MI->getOperand(1).getReg();
1350 unsigned OldVal = MI->getOperand(2).getReg();
1351 unsigned NewVal = MI->getOperand(3).getReg();
1353 unsigned Success = RegInfo.createVirtualRegister(RC);
1355 // insert new blocks after the current block
1356 const BasicBlock *LLVM_BB = BB->getBasicBlock();
1357 MachineBasicBlock *loop1MBB = MF->CreateMachineBasicBlock(LLVM_BB);
1358 MachineBasicBlock *loop2MBB = MF->CreateMachineBasicBlock(LLVM_BB);
1359 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
1360 MachineFunction::iterator It = ++BB->getIterator();
1361 MF->insert(It, loop1MBB);
1362 MF->insert(It, loop2MBB);
1363 MF->insert(It, exitMBB);
1365 // Transfer the remainder of BB and its successor edges to exitMBB.
1366 exitMBB->splice(exitMBB->begin(), BB,
1367 std::next(MachineBasicBlock::iterator(MI)), BB->end());
1368 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
1372 // fallthrough --> loop1MBB
1373 BB->addSuccessor(loop1MBB);
1374 loop1MBB->addSuccessor(exitMBB);
1375 loop1MBB->addSuccessor(loop2MBB);
1376 loop2MBB->addSuccessor(loop1MBB);
1377 loop2MBB->addSuccessor(exitMBB);
1381 // bne dest, oldval, exitMBB
1383 BuildMI(BB, DL, TII->get(LL), Dest).addReg(Ptr).addImm(0);
1384 BuildMI(BB, DL, TII->get(BNE))
1385 .addReg(Dest).addReg(OldVal).addMBB(exitMBB);
1388 // sc success, newval, 0(ptr)
1389 // beq success, $0, loop1MBB
1391 BuildMI(BB, DL, TII->get(SC), Success)
1392 .addReg(NewVal).addReg(Ptr).addImm(0);
1393 BuildMI(BB, DL, TII->get(BEQ))
1394 .addReg(Success).addReg(ZERO).addMBB(loop1MBB);
1396 MI->eraseFromParent(); // The instruction is gone now.
1402 MipsTargetLowering::emitAtomicCmpSwapPartword(MachineInstr *MI,
1403 MachineBasicBlock *BB,
1404 unsigned Size) const {
1405 assert((Size == 1 || Size == 2) &&
1406 "Unsupported size for EmitAtomicCmpSwapPartial.");
1408 MachineFunction *MF = BB->getParent();
1409 MachineRegisterInfo &RegInfo = MF->getRegInfo();
1410 const TargetRegisterClass *RC = getRegClassFor(MVT::i32);
1411 const TargetInstrInfo *TII = Subtarget.getInstrInfo();
1412 DebugLoc DL = MI->getDebugLoc();
1414 unsigned Dest = MI->getOperand(0).getReg();
1415 unsigned Ptr = MI->getOperand(1).getReg();
1416 unsigned CmpVal = MI->getOperand(2).getReg();
1417 unsigned NewVal = MI->getOperand(3).getReg();
1419 unsigned AlignedAddr = RegInfo.createVirtualRegister(RC);
1420 unsigned ShiftAmt = RegInfo.createVirtualRegister(RC);
1421 unsigned Mask = RegInfo.createVirtualRegister(RC);
1422 unsigned Mask2 = RegInfo.createVirtualRegister(RC);
1423 unsigned ShiftedCmpVal = RegInfo.createVirtualRegister(RC);
1424 unsigned OldVal = RegInfo.createVirtualRegister(RC);
1425 unsigned MaskedOldVal0 = RegInfo.createVirtualRegister(RC);
1426 unsigned ShiftedNewVal = RegInfo.createVirtualRegister(RC);
1427 unsigned MaskLSB2 = RegInfo.createVirtualRegister(RC);
1428 unsigned PtrLSB2 = RegInfo.createVirtualRegister(RC);
1429 unsigned MaskUpper = RegInfo.createVirtualRegister(RC);
1430 unsigned MaskedCmpVal = RegInfo.createVirtualRegister(RC);
1431 unsigned MaskedNewVal = RegInfo.createVirtualRegister(RC);
1432 unsigned MaskedOldVal1 = RegInfo.createVirtualRegister(RC);
1433 unsigned StoreVal = RegInfo.createVirtualRegister(RC);
1434 unsigned SrlRes = RegInfo.createVirtualRegister(RC);
1435 unsigned Success = RegInfo.createVirtualRegister(RC);
1437 // insert new blocks after the current block
1438 const BasicBlock *LLVM_BB = BB->getBasicBlock();
1439 MachineBasicBlock *loop1MBB = MF->CreateMachineBasicBlock(LLVM_BB);
1440 MachineBasicBlock *loop2MBB = MF->CreateMachineBasicBlock(LLVM_BB);
1441 MachineBasicBlock *sinkMBB = MF->CreateMachineBasicBlock(LLVM_BB);
1442 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
1443 MachineFunction::iterator It = ++BB->getIterator();
1444 MF->insert(It, loop1MBB);
1445 MF->insert(It, loop2MBB);
1446 MF->insert(It, sinkMBB);
1447 MF->insert(It, exitMBB);
1449 // Transfer the remainder of BB and its successor edges to exitMBB.
1450 exitMBB->splice(exitMBB->begin(), BB,
1451 std::next(MachineBasicBlock::iterator(MI)), BB->end());
1452 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
1454 BB->addSuccessor(loop1MBB);
1455 loop1MBB->addSuccessor(sinkMBB);
1456 loop1MBB->addSuccessor(loop2MBB);
1457 loop2MBB->addSuccessor(loop1MBB);
1458 loop2MBB->addSuccessor(sinkMBB);
1459 sinkMBB->addSuccessor(exitMBB);
1461 // FIXME: computation of newval2 can be moved to loop2MBB.
1463 // addiu masklsb2,$0,-4 # 0xfffffffc
1464 // and alignedaddr,ptr,masklsb2
1465 // andi ptrlsb2,ptr,3
1466 // sll shiftamt,ptrlsb2,3
1467 // ori maskupper,$0,255 # 0xff
1468 // sll mask,maskupper,shiftamt
1469 // nor mask2,$0,mask
1470 // andi maskedcmpval,cmpval,255
1471 // sll shiftedcmpval,maskedcmpval,shiftamt
1472 // andi maskednewval,newval,255
1473 // sll shiftednewval,maskednewval,shiftamt
1474 int64_t MaskImm = (Size == 1) ? 255 : 65535;
1475 BuildMI(BB, DL, TII->get(Mips::ADDiu), MaskLSB2)
1476 .addReg(Mips::ZERO).addImm(-4);
1477 BuildMI(BB, DL, TII->get(Mips::AND), AlignedAddr)
1478 .addReg(Ptr).addReg(MaskLSB2);
1479 BuildMI(BB, DL, TII->get(Mips::ANDi), PtrLSB2).addReg(Ptr).addImm(3);
1480 if (Subtarget.isLittle()) {
1481 BuildMI(BB, DL, TII->get(Mips::SLL), ShiftAmt).addReg(PtrLSB2).addImm(3);
1483 unsigned Off = RegInfo.createVirtualRegister(RC);
1484 BuildMI(BB, DL, TII->get(Mips::XORi), Off)
1485 .addReg(PtrLSB2).addImm((Size == 1) ? 3 : 2);
1486 BuildMI(BB, DL, TII->get(Mips::SLL), ShiftAmt).addReg(Off).addImm(3);
1488 BuildMI(BB, DL, TII->get(Mips::ORi), MaskUpper)
1489 .addReg(Mips::ZERO).addImm(MaskImm);
1490 BuildMI(BB, DL, TII->get(Mips::SLLV), Mask)
1491 .addReg(MaskUpper).addReg(ShiftAmt);
1492 BuildMI(BB, DL, TII->get(Mips::NOR), Mask2).addReg(Mips::ZERO).addReg(Mask);
1493 BuildMI(BB, DL, TII->get(Mips::ANDi), MaskedCmpVal)
1494 .addReg(CmpVal).addImm(MaskImm);
1495 BuildMI(BB, DL, TII->get(Mips::SLLV), ShiftedCmpVal)
1496 .addReg(MaskedCmpVal).addReg(ShiftAmt);
1497 BuildMI(BB, DL, TII->get(Mips::ANDi), MaskedNewVal)
1498 .addReg(NewVal).addImm(MaskImm);
1499 BuildMI(BB, DL, TII->get(Mips::SLLV), ShiftedNewVal)
1500 .addReg(MaskedNewVal).addReg(ShiftAmt);
1503 // ll oldval,0(alginedaddr)
1504 // and maskedoldval0,oldval,mask
1505 // bne maskedoldval0,shiftedcmpval,sinkMBB
1507 unsigned LL = isMicroMips ? Mips::LL_MM : Mips::LL;
1508 BuildMI(BB, DL, TII->get(LL), OldVal).addReg(AlignedAddr).addImm(0);
1509 BuildMI(BB, DL, TII->get(Mips::AND), MaskedOldVal0)
1510 .addReg(OldVal).addReg(Mask);
1511 BuildMI(BB, DL, TII->get(Mips::BNE))
1512 .addReg(MaskedOldVal0).addReg(ShiftedCmpVal).addMBB(sinkMBB);
1515 // and maskedoldval1,oldval,mask2
1516 // or storeval,maskedoldval1,shiftednewval
1517 // sc success,storeval,0(alignedaddr)
1518 // beq success,$0,loop1MBB
1520 BuildMI(BB, DL, TII->get(Mips::AND), MaskedOldVal1)
1521 .addReg(OldVal).addReg(Mask2);
1522 BuildMI(BB, DL, TII->get(Mips::OR), StoreVal)
1523 .addReg(MaskedOldVal1).addReg(ShiftedNewVal);
1524 unsigned SC = isMicroMips ? Mips::SC_MM : Mips::SC;
1525 BuildMI(BB, DL, TII->get(SC), Success)
1526 .addReg(StoreVal).addReg(AlignedAddr).addImm(0);
1527 BuildMI(BB, DL, TII->get(Mips::BEQ))
1528 .addReg(Success).addReg(Mips::ZERO).addMBB(loop1MBB);
1531 // srl srlres,maskedoldval0,shiftamt
1532 // sign_extend dest,srlres
1535 BuildMI(BB, DL, TII->get(Mips::SRLV), SrlRes)
1536 .addReg(MaskedOldVal0).addReg(ShiftAmt);
1537 BB = emitSignExtendToI32InReg(MI, BB, Size, Dest, SrlRes);
1539 MI->eraseFromParent(); // The instruction is gone now.
1544 MachineBasicBlock *MipsTargetLowering::emitSEL_D(MachineInstr *MI,
1545 MachineBasicBlock *BB) const {
1546 MachineFunction *MF = BB->getParent();
1547 const TargetRegisterInfo *TRI = Subtarget.getRegisterInfo();
1548 const TargetInstrInfo *TII = Subtarget.getInstrInfo();
1549 MachineRegisterInfo &RegInfo = MF->getRegInfo();
1550 DebugLoc DL = MI->getDebugLoc();
1551 MachineBasicBlock::iterator II(MI);
1553 unsigned Fc = MI->getOperand(1).getReg();
1554 const auto &FGR64RegClass = TRI->getRegClass(Mips::FGR64RegClassID);
1556 unsigned Fc2 = RegInfo.createVirtualRegister(FGR64RegClass);
1558 BuildMI(*BB, II, DL, TII->get(Mips::SUBREG_TO_REG), Fc2)
1561 .addImm(Mips::sub_lo);
1563 // We don't erase the original instruction, we just replace the condition
1564 // register with the 64-bit super-register.
1565 MI->getOperand(1).setReg(Fc2);
1570 //===----------------------------------------------------------------------===//
1571 // Misc Lower Operation implementation
1572 //===----------------------------------------------------------------------===//
1573 SDValue MipsTargetLowering::lowerBR_JT(SDValue Op, SelectionDAG &DAG) const {
1574 SDValue Chain = Op.getOperand(0);
1575 SDValue Table = Op.getOperand(1);
1576 SDValue Index = Op.getOperand(2);
1578 auto &TD = DAG.getDataLayout();
1579 EVT PTy = getPointerTy(TD);
1580 unsigned EntrySize =
1581 DAG.getMachineFunction().getJumpTableInfo()->getEntrySize(TD);
1583 Index = DAG.getNode(ISD::MUL, DL, PTy, Index,
1584 DAG.getConstant(EntrySize, DL, PTy));
1585 SDValue Addr = DAG.getNode(ISD::ADD, DL, PTy, Index, Table);
1587 EVT MemVT = EVT::getIntegerVT(*DAG.getContext(), EntrySize * 8);
1589 DAG.getExtLoad(ISD::SEXTLOAD, DL, PTy, Chain, Addr,
1590 MachinePointerInfo::getJumpTable(DAG.getMachineFunction()),
1591 MemVT, false, false, false, 0);
1592 Chain = Addr.getValue(1);
1594 if ((getTargetMachine().getRelocationModel() == Reloc::PIC_) || ABI.IsN64()) {
1595 // For PIC, the sequence is:
1596 // BRIND(load(Jumptable + index) + RelocBase)
1597 // RelocBase can be JumpTable, GOT or some sort of global base.
1598 Addr = DAG.getNode(ISD::ADD, DL, PTy, Addr,
1599 getPICJumpTableRelocBase(Table, DAG));
1602 return DAG.getNode(ISD::BRIND, DL, MVT::Other, Chain, Addr);
1605 SDValue MipsTargetLowering::lowerBRCOND(SDValue Op, SelectionDAG &DAG) const {
1606 // The first operand is the chain, the second is the condition, the third is
1607 // the block to branch to if the condition is true.
1608 SDValue Chain = Op.getOperand(0);
1609 SDValue Dest = Op.getOperand(2);
1612 assert(!Subtarget.hasMips32r6() && !Subtarget.hasMips64r6());
1613 SDValue CondRes = createFPCmp(DAG, Op.getOperand(1));
1615 // Return if flag is not set by a floating point comparison.
1616 if (CondRes.getOpcode() != MipsISD::FPCmp)
1619 SDValue CCNode = CondRes.getOperand(2);
1621 (Mips::CondCode)cast<ConstantSDNode>(CCNode)->getZExtValue();
1622 unsigned Opc = invertFPCondCodeUser(CC) ? Mips::BRANCH_F : Mips::BRANCH_T;
1623 SDValue BrCode = DAG.getConstant(Opc, DL, MVT::i32);
1624 SDValue FCC0 = DAG.getRegister(Mips::FCC0, MVT::i32);
1625 return DAG.getNode(MipsISD::FPBrcond, DL, Op.getValueType(), Chain, BrCode,
1626 FCC0, Dest, CondRes);
1629 SDValue MipsTargetLowering::
1630 lowerSELECT(SDValue Op, SelectionDAG &DAG) const
1632 assert(!Subtarget.hasMips32r6() && !Subtarget.hasMips64r6());
1633 SDValue Cond = createFPCmp(DAG, Op.getOperand(0));
1635 // Return if flag is not set by a floating point comparison.
1636 if (Cond.getOpcode() != MipsISD::FPCmp)
1639 return createCMovFP(DAG, Cond, Op.getOperand(1), Op.getOperand(2),
1643 SDValue MipsTargetLowering::
1644 lowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const
1647 EVT Ty = Op.getOperand(0).getValueType();
1649 DAG.getNode(ISD::SETCC, DL, getSetCCResultType(DAG.getDataLayout(),
1650 *DAG.getContext(), Ty),
1651 Op.getOperand(0), Op.getOperand(1), Op.getOperand(4));
1653 return DAG.getNode(ISD::SELECT, DL, Op.getValueType(), Cond, Op.getOperand(2),
1657 SDValue MipsTargetLowering::lowerSETCC(SDValue Op, SelectionDAG &DAG) const {
1658 assert(!Subtarget.hasMips32r6() && !Subtarget.hasMips64r6());
1659 SDValue Cond = createFPCmp(DAG, Op);
1661 assert(Cond.getOpcode() == MipsISD::FPCmp &&
1662 "Floating point operand expected.");
1665 SDValue True = DAG.getConstant(1, DL, MVT::i32);
1666 SDValue False = DAG.getConstant(0, DL, MVT::i32);
1668 return createCMovFP(DAG, Cond, True, False, DL);
1671 SDValue MipsTargetLowering::lowerGlobalAddress(SDValue Op,
1672 SelectionDAG &DAG) const {
1673 EVT Ty = Op.getValueType();
1674 GlobalAddressSDNode *N = cast<GlobalAddressSDNode>(Op);
1675 const GlobalValue *GV = N->getGlobal();
1677 if (getTargetMachine().getRelocationModel() != Reloc::PIC_ && !ABI.IsN64()) {
1678 const MipsTargetObjectFile *TLOF =
1679 static_cast<const MipsTargetObjectFile *>(
1680 getTargetMachine().getObjFileLowering());
1681 if (TLOF->IsGlobalInSmallSection(GV, getTargetMachine()))
1682 // %gp_rel relocation
1683 return getAddrGPRel(N, SDLoc(N), Ty, DAG);
1685 // %hi/%lo relocation
1686 return getAddrNonPIC(N, SDLoc(N), Ty, DAG);
1689 if (GV->hasInternalLinkage() || (GV->hasLocalLinkage() && !isa<Function>(GV)))
1690 return getAddrLocal(N, SDLoc(N), Ty, DAG, ABI.IsN32() || ABI.IsN64());
1693 return getAddrGlobalLargeGOT(
1694 N, SDLoc(N), Ty, DAG, MipsII::MO_GOT_HI16, MipsII::MO_GOT_LO16,
1696 MachinePointerInfo::getGOT(DAG.getMachineFunction()));
1698 return getAddrGlobal(
1699 N, SDLoc(N), Ty, DAG,
1700 (ABI.IsN32() || ABI.IsN64()) ? MipsII::MO_GOT_DISP : MipsII::MO_GOT16,
1701 DAG.getEntryNode(), MachinePointerInfo::getGOT(DAG.getMachineFunction()));
1704 SDValue MipsTargetLowering::lowerBlockAddress(SDValue Op,
1705 SelectionDAG &DAG) const {
1706 BlockAddressSDNode *N = cast<BlockAddressSDNode>(Op);
1707 EVT Ty = Op.getValueType();
1709 if (getTargetMachine().getRelocationModel() != Reloc::PIC_ && !ABI.IsN64())
1710 return getAddrNonPIC(N, SDLoc(N), Ty, DAG);
1712 return getAddrLocal(N, SDLoc(N), Ty, DAG, ABI.IsN32() || ABI.IsN64());
1715 SDValue MipsTargetLowering::
1716 lowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const
1718 // If the relocation model is PIC, use the General Dynamic TLS Model or
1719 // Local Dynamic TLS model, otherwise use the Initial Exec or
1720 // Local Exec TLS Model.
1722 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
1723 if (DAG.getTarget().Options.EmulatedTLS)
1724 return LowerToTLSEmulatedModel(GA, DAG);
1727 const GlobalValue *GV = GA->getGlobal();
1728 EVT PtrVT = getPointerTy(DAG.getDataLayout());
1730 TLSModel::Model model = getTargetMachine().getTLSModel(GV);
1732 if (model == TLSModel::GeneralDynamic || model == TLSModel::LocalDynamic) {
1733 // General Dynamic and Local Dynamic TLS Model.
1734 unsigned Flag = (model == TLSModel::LocalDynamic) ? MipsII::MO_TLSLDM
1737 SDValue TGA = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, Flag);
1738 SDValue Argument = DAG.getNode(MipsISD::Wrapper, DL, PtrVT,
1739 getGlobalReg(DAG, PtrVT), TGA);
1740 unsigned PtrSize = PtrVT.getSizeInBits();
1741 IntegerType *PtrTy = Type::getIntNTy(*DAG.getContext(), PtrSize);
1743 SDValue TlsGetAddr = DAG.getExternalSymbol("__tls_get_addr", PtrVT);
1747 Entry.Node = Argument;
1749 Args.push_back(Entry);
1751 TargetLowering::CallLoweringInfo CLI(DAG);
1752 CLI.setDebugLoc(DL).setChain(DAG.getEntryNode())
1753 .setCallee(CallingConv::C, PtrTy, TlsGetAddr, std::move(Args), 0);
1754 std::pair<SDValue, SDValue> CallResult = LowerCallTo(CLI);
1756 SDValue Ret = CallResult.first;
1758 if (model != TLSModel::LocalDynamic)
1761 SDValue TGAHi = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0,
1762 MipsII::MO_DTPREL_HI);
1763 SDValue Hi = DAG.getNode(MipsISD::Hi, DL, PtrVT, TGAHi);
1764 SDValue TGALo = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0,
1765 MipsII::MO_DTPREL_LO);
1766 SDValue Lo = DAG.getNode(MipsISD::Lo, DL, PtrVT, TGALo);
1767 SDValue Add = DAG.getNode(ISD::ADD, DL, PtrVT, Hi, Ret);
1768 return DAG.getNode(ISD::ADD, DL, PtrVT, Add, Lo);
1772 if (model == TLSModel::InitialExec) {
1773 // Initial Exec TLS Model
1774 SDValue TGA = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0,
1775 MipsII::MO_GOTTPREL);
1776 TGA = DAG.getNode(MipsISD::Wrapper, DL, PtrVT, getGlobalReg(DAG, PtrVT),
1778 Offset = DAG.getLoad(PtrVT, DL,
1779 DAG.getEntryNode(), TGA, MachinePointerInfo(),
1780 false, false, false, 0);
1782 // Local Exec TLS Model
1783 assert(model == TLSModel::LocalExec);
1784 SDValue TGAHi = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0,
1785 MipsII::MO_TPREL_HI);
1786 SDValue TGALo = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0,
1787 MipsII::MO_TPREL_LO);
1788 SDValue Hi = DAG.getNode(MipsISD::Hi, DL, PtrVT, TGAHi);
1789 SDValue Lo = DAG.getNode(MipsISD::Lo, DL, PtrVT, TGALo);
1790 Offset = DAG.getNode(ISD::ADD, DL, PtrVT, Hi, Lo);
1793 SDValue ThreadPointer = DAG.getNode(MipsISD::ThreadPointer, DL, PtrVT);
1794 return DAG.getNode(ISD::ADD, DL, PtrVT, ThreadPointer, Offset);
1797 SDValue MipsTargetLowering::
1798 lowerJumpTable(SDValue Op, SelectionDAG &DAG) const
1800 JumpTableSDNode *N = cast<JumpTableSDNode>(Op);
1801 EVT Ty = Op.getValueType();
1803 if (getTargetMachine().getRelocationModel() != Reloc::PIC_ && !ABI.IsN64())
1804 return getAddrNonPIC(N, SDLoc(N), Ty, DAG);
1806 return getAddrLocal(N, SDLoc(N), Ty, DAG, ABI.IsN32() || ABI.IsN64());
1809 SDValue MipsTargetLowering::
1810 lowerConstantPool(SDValue Op, SelectionDAG &DAG) const
1812 ConstantPoolSDNode *N = cast<ConstantPoolSDNode>(Op);
1813 EVT Ty = Op.getValueType();
1815 if (getTargetMachine().getRelocationModel() != Reloc::PIC_ && !ABI.IsN64()) {
1816 const MipsTargetObjectFile *TLOF =
1817 static_cast<const MipsTargetObjectFile *>(
1818 getTargetMachine().getObjFileLowering());
1820 if (TLOF->IsConstantInSmallSection(DAG.getDataLayout(), N->getConstVal(),
1821 getTargetMachine()))
1822 // %gp_rel relocation
1823 return getAddrGPRel(N, SDLoc(N), Ty, DAG);
1825 return getAddrNonPIC(N, SDLoc(N), Ty, DAG);
1828 return getAddrLocal(N, SDLoc(N), Ty, DAG, ABI.IsN32() || ABI.IsN64());
1831 SDValue MipsTargetLowering::lowerVASTART(SDValue Op, SelectionDAG &DAG) const {
1832 MachineFunction &MF = DAG.getMachineFunction();
1833 MipsFunctionInfo *FuncInfo = MF.getInfo<MipsFunctionInfo>();
1836 SDValue FI = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(),
1837 getPointerTy(MF.getDataLayout()));
1839 // vastart just stores the address of the VarArgsFrameIndex slot into the
1840 // memory location argument.
1841 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
1842 return DAG.getStore(Op.getOperand(0), DL, FI, Op.getOperand(1),
1843 MachinePointerInfo(SV), false, false, 0);
1846 SDValue MipsTargetLowering::lowerVAARG(SDValue Op, SelectionDAG &DAG) const {
1847 SDNode *Node = Op.getNode();
1848 EVT VT = Node->getValueType(0);
1849 SDValue Chain = Node->getOperand(0);
1850 SDValue VAListPtr = Node->getOperand(1);
1851 unsigned Align = Node->getConstantOperandVal(3);
1852 const Value *SV = cast<SrcValueSDNode>(Node->getOperand(2))->getValue();
1854 unsigned ArgSlotSizeInBytes = (ABI.IsN32() || ABI.IsN64()) ? 8 : 4;
1856 SDValue VAListLoad =
1857 DAG.getLoad(getPointerTy(DAG.getDataLayout()), DL, Chain, VAListPtr,
1858 MachinePointerInfo(SV), false, false, false, 0);
1859 SDValue VAList = VAListLoad;
1861 // Re-align the pointer if necessary.
1862 // It should only ever be necessary for 64-bit types on O32 since the minimum
1863 // argument alignment is the same as the maximum type alignment for N32/N64.
1865 // FIXME: We currently align too often. The code generator doesn't notice
1866 // when the pointer is still aligned from the last va_arg (or pair of
1867 // va_args for the i64 on O32 case).
1868 if (Align > getMinStackArgumentAlignment()) {
1869 assert(((Align & (Align-1)) == 0) && "Expected Align to be a power of 2");
1871 VAList = DAG.getNode(ISD::ADD, DL, VAList.getValueType(), VAList,
1872 DAG.getConstant(Align - 1, DL, VAList.getValueType()));
1874 VAList = DAG.getNode(ISD::AND, DL, VAList.getValueType(), VAList,
1875 DAG.getConstant(-(int64_t)Align, DL,
1876 VAList.getValueType()));
1879 // Increment the pointer, VAList, to the next vaarg.
1880 auto &TD = DAG.getDataLayout();
1881 unsigned ArgSizeInBytes =
1882 TD.getTypeAllocSize(VT.getTypeForEVT(*DAG.getContext()));
1883 SDValue Tmp3 = DAG.getNode(ISD::ADD, DL, VAList.getValueType(), VAList,
1884 DAG.getConstant(RoundUpToAlignment(ArgSizeInBytes,
1885 ArgSlotSizeInBytes),
1886 DL, VAList.getValueType()));
1887 // Store the incremented VAList to the legalized pointer
1888 Chain = DAG.getStore(VAListLoad.getValue(1), DL, Tmp3, VAListPtr,
1889 MachinePointerInfo(SV), false, false, 0);
1891 // In big-endian mode we must adjust the pointer when the load size is smaller
1892 // than the argument slot size. We must also reduce the known alignment to
1893 // match. For example in the N64 ABI, we must add 4 bytes to the offset to get
1894 // the correct half of the slot, and reduce the alignment from 8 (slot
1895 // alignment) down to 4 (type alignment).
1896 if (!Subtarget.isLittle() && ArgSizeInBytes < ArgSlotSizeInBytes) {
1897 unsigned Adjustment = ArgSlotSizeInBytes - ArgSizeInBytes;
1898 VAList = DAG.getNode(ISD::ADD, DL, VAListPtr.getValueType(), VAList,
1899 DAG.getIntPtrConstant(Adjustment, DL));
1901 // Load the actual argument out of the pointer VAList
1902 return DAG.getLoad(VT, DL, Chain, VAList, MachinePointerInfo(), false, false,
1906 static SDValue lowerFCOPYSIGN32(SDValue Op, SelectionDAG &DAG,
1907 bool HasExtractInsert) {
1908 EVT TyX = Op.getOperand(0).getValueType();
1909 EVT TyY = Op.getOperand(1).getValueType();
1911 SDValue Const1 = DAG.getConstant(1, DL, MVT::i32);
1912 SDValue Const31 = DAG.getConstant(31, DL, MVT::i32);
1915 // If operand is of type f64, extract the upper 32-bit. Otherwise, bitcast it
1917 SDValue X = (TyX == MVT::f32) ?
1918 DAG.getNode(ISD::BITCAST, DL, MVT::i32, Op.getOperand(0)) :
1919 DAG.getNode(MipsISD::ExtractElementF64, DL, MVT::i32, Op.getOperand(0),
1921 SDValue Y = (TyY == MVT::f32) ?
1922 DAG.getNode(ISD::BITCAST, DL, MVT::i32, Op.getOperand(1)) :
1923 DAG.getNode(MipsISD::ExtractElementF64, DL, MVT::i32, Op.getOperand(1),
1926 if (HasExtractInsert) {
1927 // ext E, Y, 31, 1 ; extract bit31 of Y
1928 // ins X, E, 31, 1 ; insert extracted bit at bit31 of X
1929 SDValue E = DAG.getNode(MipsISD::Ext, DL, MVT::i32, Y, Const31, Const1);
1930 Res = DAG.getNode(MipsISD::Ins, DL, MVT::i32, E, Const31, Const1, X);
1933 // srl SrlX, SllX, 1
1935 // sll SllY, SrlX, 31
1936 // or Or, SrlX, SllY
1937 SDValue SllX = DAG.getNode(ISD::SHL, DL, MVT::i32, X, Const1);
1938 SDValue SrlX = DAG.getNode(ISD::SRL, DL, MVT::i32, SllX, Const1);
1939 SDValue SrlY = DAG.getNode(ISD::SRL, DL, MVT::i32, Y, Const31);
1940 SDValue SllY = DAG.getNode(ISD::SHL, DL, MVT::i32, SrlY, Const31);
1941 Res = DAG.getNode(ISD::OR, DL, MVT::i32, SrlX, SllY);
1944 if (TyX == MVT::f32)
1945 return DAG.getNode(ISD::BITCAST, DL, Op.getOperand(0).getValueType(), Res);
1947 SDValue LowX = DAG.getNode(MipsISD::ExtractElementF64, DL, MVT::i32,
1949 DAG.getConstant(0, DL, MVT::i32));
1950 return DAG.getNode(MipsISD::BuildPairF64, DL, MVT::f64, LowX, Res);
1953 static SDValue lowerFCOPYSIGN64(SDValue Op, SelectionDAG &DAG,
1954 bool HasExtractInsert) {
1955 unsigned WidthX = Op.getOperand(0).getValueSizeInBits();
1956 unsigned WidthY = Op.getOperand(1).getValueSizeInBits();
1957 EVT TyX = MVT::getIntegerVT(WidthX), TyY = MVT::getIntegerVT(WidthY);
1959 SDValue Const1 = DAG.getConstant(1, DL, MVT::i32);
1961 // Bitcast to integer nodes.
1962 SDValue X = DAG.getNode(ISD::BITCAST, DL, TyX, Op.getOperand(0));
1963 SDValue Y = DAG.getNode(ISD::BITCAST, DL, TyY, Op.getOperand(1));
1965 if (HasExtractInsert) {
1966 // ext E, Y, width(Y) - 1, 1 ; extract bit width(Y)-1 of Y
1967 // ins X, E, width(X) - 1, 1 ; insert extracted bit at bit width(X)-1 of X
1968 SDValue E = DAG.getNode(MipsISD::Ext, DL, TyY, Y,
1969 DAG.getConstant(WidthY - 1, DL, MVT::i32), Const1);
1971 if (WidthX > WidthY)
1972 E = DAG.getNode(ISD::ZERO_EXTEND, DL, TyX, E);
1973 else if (WidthY > WidthX)
1974 E = DAG.getNode(ISD::TRUNCATE, DL, TyX, E);
1976 SDValue I = DAG.getNode(MipsISD::Ins, DL, TyX, E,
1977 DAG.getConstant(WidthX - 1, DL, MVT::i32), Const1,
1979 return DAG.getNode(ISD::BITCAST, DL, Op.getOperand(0).getValueType(), I);
1982 // (d)sll SllX, X, 1
1983 // (d)srl SrlX, SllX, 1
1984 // (d)srl SrlY, Y, width(Y)-1
1985 // (d)sll SllY, SrlX, width(Y)-1
1986 // or Or, SrlX, SllY
1987 SDValue SllX = DAG.getNode(ISD::SHL, DL, TyX, X, Const1);
1988 SDValue SrlX = DAG.getNode(ISD::SRL, DL, TyX, SllX, Const1);
1989 SDValue SrlY = DAG.getNode(ISD::SRL, DL, TyY, Y,
1990 DAG.getConstant(WidthY - 1, DL, MVT::i32));
1992 if (WidthX > WidthY)
1993 SrlY = DAG.getNode(ISD::ZERO_EXTEND, DL, TyX, SrlY);
1994 else if (WidthY > WidthX)
1995 SrlY = DAG.getNode(ISD::TRUNCATE, DL, TyX, SrlY);
1997 SDValue SllY = DAG.getNode(ISD::SHL, DL, TyX, SrlY,
1998 DAG.getConstant(WidthX - 1, DL, MVT::i32));
1999 SDValue Or = DAG.getNode(ISD::OR, DL, TyX, SrlX, SllY);
2000 return DAG.getNode(ISD::BITCAST, DL, Op.getOperand(0).getValueType(), Or);
2004 MipsTargetLowering::lowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) const {
2005 if (Subtarget.isGP64bit())
2006 return lowerFCOPYSIGN64(Op, DAG, Subtarget.hasExtractInsert());
2008 return lowerFCOPYSIGN32(Op, DAG, Subtarget.hasExtractInsert());
2011 SDValue MipsTargetLowering::
2012 lowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
2014 assert((cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue() == 0) &&
2015 "Frame address can only be determined for current frame.");
2017 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
2018 MFI->setFrameAddressIsTaken(true);
2019 EVT VT = Op.getValueType();
2021 SDValue FrameAddr = DAG.getCopyFromReg(
2022 DAG.getEntryNode(), DL, ABI.IsN64() ? Mips::FP_64 : Mips::FP, VT);
2026 SDValue MipsTargetLowering::lowerRETURNADDR(SDValue Op,
2027 SelectionDAG &DAG) const {
2028 if (verifyReturnAddressArgumentIsConstant(Op, DAG))
2032 assert((cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue() == 0) &&
2033 "Return address can be determined only for current frame.");
2035 MachineFunction &MF = DAG.getMachineFunction();
2036 MachineFrameInfo *MFI = MF.getFrameInfo();
2037 MVT VT = Op.getSimpleValueType();
2038 unsigned RA = ABI.IsN64() ? Mips::RA_64 : Mips::RA;
2039 MFI->setReturnAddressIsTaken(true);
2041 // Return RA, which contains the return address. Mark it an implicit live-in.
2042 unsigned Reg = MF.addLiveIn(RA, getRegClassFor(VT));
2043 return DAG.getCopyFromReg(DAG.getEntryNode(), SDLoc(Op), Reg, VT);
2046 // An EH_RETURN is the result of lowering llvm.eh.return which in turn is
2047 // generated from __builtin_eh_return (offset, handler)
2048 // The effect of this is to adjust the stack pointer by "offset"
2049 // and then branch to "handler".
2050 SDValue MipsTargetLowering::lowerEH_RETURN(SDValue Op, SelectionDAG &DAG)
2052 MachineFunction &MF = DAG.getMachineFunction();
2053 MipsFunctionInfo *MipsFI = MF.getInfo<MipsFunctionInfo>();
2055 MipsFI->setCallsEhReturn();
2056 SDValue Chain = Op.getOperand(0);
2057 SDValue Offset = Op.getOperand(1);
2058 SDValue Handler = Op.getOperand(2);
2060 EVT Ty = ABI.IsN64() ? MVT::i64 : MVT::i32;
2062 // Store stack offset in V1, store jump target in V0. Glue CopyToReg and
2063 // EH_RETURN nodes, so that instructions are emitted back-to-back.
2064 unsigned OffsetReg = ABI.IsN64() ? Mips::V1_64 : Mips::V1;
2065 unsigned AddrReg = ABI.IsN64() ? Mips::V0_64 : Mips::V0;
2066 Chain = DAG.getCopyToReg(Chain, DL, OffsetReg, Offset, SDValue());
2067 Chain = DAG.getCopyToReg(Chain, DL, AddrReg, Handler, Chain.getValue(1));
2068 return DAG.getNode(MipsISD::EH_RETURN, DL, MVT::Other, Chain,
2069 DAG.getRegister(OffsetReg, Ty),
2070 DAG.getRegister(AddrReg, getPointerTy(MF.getDataLayout())),
2074 SDValue MipsTargetLowering::lowerATOMIC_FENCE(SDValue Op,
2075 SelectionDAG &DAG) const {
2076 // FIXME: Need pseudo-fence for 'singlethread' fences
2077 // FIXME: Set SType for weaker fences where supported/appropriate.
2080 return DAG.getNode(MipsISD::Sync, DL, MVT::Other, Op.getOperand(0),
2081 DAG.getConstant(SType, DL, MVT::i32));
2084 SDValue MipsTargetLowering::lowerShiftLeftParts(SDValue Op,
2085 SelectionDAG &DAG) const {
2087 MVT VT = Subtarget.isGP64bit() ? MVT::i64 : MVT::i32;
2089 SDValue Lo = Op.getOperand(0), Hi = Op.getOperand(1);
2090 SDValue Shamt = Op.getOperand(2);
2091 // if shamt < (VT.bits):
2092 // lo = (shl lo, shamt)
2093 // hi = (or (shl hi, shamt) (srl (srl lo, 1), ~shamt))
2096 // hi = (shl lo, shamt[4:0])
2097 SDValue Not = DAG.getNode(ISD::XOR, DL, MVT::i32, Shamt,
2098 DAG.getConstant(-1, DL, MVT::i32));
2099 SDValue ShiftRight1Lo = DAG.getNode(ISD::SRL, DL, VT, Lo,
2100 DAG.getConstant(1, DL, VT));
2101 SDValue ShiftRightLo = DAG.getNode(ISD::SRL, DL, VT, ShiftRight1Lo, Not);
2102 SDValue ShiftLeftHi = DAG.getNode(ISD::SHL, DL, VT, Hi, Shamt);
2103 SDValue Or = DAG.getNode(ISD::OR, DL, VT, ShiftLeftHi, ShiftRightLo);
2104 SDValue ShiftLeftLo = DAG.getNode(ISD::SHL, DL, VT, Lo, Shamt);
2105 SDValue Cond = DAG.getNode(ISD::AND, DL, MVT::i32, Shamt,
2106 DAG.getConstant(VT.getSizeInBits(), DL, MVT::i32));
2107 Lo = DAG.getNode(ISD::SELECT, DL, VT, Cond,
2108 DAG.getConstant(0, DL, VT), ShiftLeftLo);
2109 Hi = DAG.getNode(ISD::SELECT, DL, VT, Cond, ShiftLeftLo, Or);
2111 SDValue Ops[2] = {Lo, Hi};
2112 return DAG.getMergeValues(Ops, DL);
2115 SDValue MipsTargetLowering::lowerShiftRightParts(SDValue Op, SelectionDAG &DAG,
2118 SDValue Lo = Op.getOperand(0), Hi = Op.getOperand(1);
2119 SDValue Shamt = Op.getOperand(2);
2120 MVT VT = Subtarget.isGP64bit() ? MVT::i64 : MVT::i32;
2122 // if shamt < (VT.bits):
2123 // lo = (or (shl (shl hi, 1), ~shamt) (srl lo, shamt))
2125 // hi = (sra hi, shamt)
2127 // hi = (srl hi, shamt)
2130 // lo = (sra hi, shamt[4:0])
2131 // hi = (sra hi, 31)
2133 // lo = (srl hi, shamt[4:0])
2135 SDValue Not = DAG.getNode(ISD::XOR, DL, MVT::i32, Shamt,
2136 DAG.getConstant(-1, DL, MVT::i32));
2137 SDValue ShiftLeft1Hi = DAG.getNode(ISD::SHL, DL, VT, Hi,
2138 DAG.getConstant(1, DL, VT));
2139 SDValue ShiftLeftHi = DAG.getNode(ISD::SHL, DL, VT, ShiftLeft1Hi, Not);
2140 SDValue ShiftRightLo = DAG.getNode(ISD::SRL, DL, VT, Lo, Shamt);
2141 SDValue Or = DAG.getNode(ISD::OR, DL, VT, ShiftLeftHi, ShiftRightLo);
2142 SDValue ShiftRightHi = DAG.getNode(IsSRA ? ISD::SRA : ISD::SRL,
2144 SDValue Cond = DAG.getNode(ISD::AND, DL, MVT::i32, Shamt,
2145 DAG.getConstant(VT.getSizeInBits(), DL, MVT::i32));
2146 SDValue Ext = DAG.getNode(ISD::SRA, DL, VT, Hi,
2147 DAG.getConstant(VT.getSizeInBits() - 1, DL, VT));
2148 Lo = DAG.getNode(ISD::SELECT, DL, VT, Cond, ShiftRightHi, Or);
2149 Hi = DAG.getNode(ISD::SELECT, DL, VT, Cond,
2150 IsSRA ? Ext : DAG.getConstant(0, DL, VT), ShiftRightHi);
2152 SDValue Ops[2] = {Lo, Hi};
2153 return DAG.getMergeValues(Ops, DL);
2156 static SDValue createLoadLR(unsigned Opc, SelectionDAG &DAG, LoadSDNode *LD,
2157 SDValue Chain, SDValue Src, unsigned Offset) {
2158 SDValue Ptr = LD->getBasePtr();
2159 EVT VT = LD->getValueType(0), MemVT = LD->getMemoryVT();
2160 EVT BasePtrVT = Ptr.getValueType();
2162 SDVTList VTList = DAG.getVTList(VT, MVT::Other);
2165 Ptr = DAG.getNode(ISD::ADD, DL, BasePtrVT, Ptr,
2166 DAG.getConstant(Offset, DL, BasePtrVT));
2168 SDValue Ops[] = { Chain, Ptr, Src };
2169 return DAG.getMemIntrinsicNode(Opc, DL, VTList, Ops, MemVT,
2170 LD->getMemOperand());
2173 // Expand an unaligned 32 or 64-bit integer load node.
2174 SDValue MipsTargetLowering::lowerLOAD(SDValue Op, SelectionDAG &DAG) const {
2175 LoadSDNode *LD = cast<LoadSDNode>(Op);
2176 EVT MemVT = LD->getMemoryVT();
2178 if (Subtarget.systemSupportsUnalignedAccess())
2181 // Return if load is aligned or if MemVT is neither i32 nor i64.
2182 if ((LD->getAlignment() >= MemVT.getSizeInBits() / 8) ||
2183 ((MemVT != MVT::i32) && (MemVT != MVT::i64)))
2186 bool IsLittle = Subtarget.isLittle();
2187 EVT VT = Op.getValueType();
2188 ISD::LoadExtType ExtType = LD->getExtensionType();
2189 SDValue Chain = LD->getChain(), Undef = DAG.getUNDEF(VT);
2191 assert((VT == MVT::i32) || (VT == MVT::i64));
2194 // (set dst, (i64 (load baseptr)))
2196 // (set tmp, (ldl (add baseptr, 7), undef))
2197 // (set dst, (ldr baseptr, tmp))
2198 if ((VT == MVT::i64) && (ExtType == ISD::NON_EXTLOAD)) {
2199 SDValue LDL = createLoadLR(MipsISD::LDL, DAG, LD, Chain, Undef,
2201 return createLoadLR(MipsISD::LDR, DAG, LD, LDL.getValue(1), LDL,
2205 SDValue LWL = createLoadLR(MipsISD::LWL, DAG, LD, Chain, Undef,
2207 SDValue LWR = createLoadLR(MipsISD::LWR, DAG, LD, LWL.getValue(1), LWL,
2211 // (set dst, (i32 (load baseptr))) or
2212 // (set dst, (i64 (sextload baseptr))) or
2213 // (set dst, (i64 (extload baseptr)))
2215 // (set tmp, (lwl (add baseptr, 3), undef))
2216 // (set dst, (lwr baseptr, tmp))
2217 if ((VT == MVT::i32) || (ExtType == ISD::SEXTLOAD) ||
2218 (ExtType == ISD::EXTLOAD))
2221 assert((VT == MVT::i64) && (ExtType == ISD::ZEXTLOAD));
2224 // (set dst, (i64 (zextload baseptr)))
2226 // (set tmp0, (lwl (add baseptr, 3), undef))
2227 // (set tmp1, (lwr baseptr, tmp0))
2228 // (set tmp2, (shl tmp1, 32))
2229 // (set dst, (srl tmp2, 32))
2231 SDValue Const32 = DAG.getConstant(32, DL, MVT::i32);
2232 SDValue SLL = DAG.getNode(ISD::SHL, DL, MVT::i64, LWR, Const32);
2233 SDValue SRL = DAG.getNode(ISD::SRL, DL, MVT::i64, SLL, Const32);
2234 SDValue Ops[] = { SRL, LWR.getValue(1) };
2235 return DAG.getMergeValues(Ops, DL);
2238 static SDValue createStoreLR(unsigned Opc, SelectionDAG &DAG, StoreSDNode *SD,
2239 SDValue Chain, unsigned Offset) {
2240 SDValue Ptr = SD->getBasePtr(), Value = SD->getValue();
2241 EVT MemVT = SD->getMemoryVT(), BasePtrVT = Ptr.getValueType();
2243 SDVTList VTList = DAG.getVTList(MVT::Other);
2246 Ptr = DAG.getNode(ISD::ADD, DL, BasePtrVT, Ptr,
2247 DAG.getConstant(Offset, DL, BasePtrVT));
2249 SDValue Ops[] = { Chain, Value, Ptr };
2250 return DAG.getMemIntrinsicNode(Opc, DL, VTList, Ops, MemVT,
2251 SD->getMemOperand());
2254 // Expand an unaligned 32 or 64-bit integer store node.
2255 static SDValue lowerUnalignedIntStore(StoreSDNode *SD, SelectionDAG &DAG,
2257 SDValue Value = SD->getValue(), Chain = SD->getChain();
2258 EVT VT = Value.getValueType();
2261 // (store val, baseptr) or
2262 // (truncstore val, baseptr)
2264 // (swl val, (add baseptr, 3))
2265 // (swr val, baseptr)
2266 if ((VT == MVT::i32) || SD->isTruncatingStore()) {
2267 SDValue SWL = createStoreLR(MipsISD::SWL, DAG, SD, Chain,
2269 return createStoreLR(MipsISD::SWR, DAG, SD, SWL, IsLittle ? 0 : 3);
2272 assert(VT == MVT::i64);
2275 // (store val, baseptr)
2277 // (sdl val, (add baseptr, 7))
2278 // (sdr val, baseptr)
2279 SDValue SDL = createStoreLR(MipsISD::SDL, DAG, SD, Chain, IsLittle ? 7 : 0);
2280 return createStoreLR(MipsISD::SDR, DAG, SD, SDL, IsLittle ? 0 : 7);
2283 // Lower (store (fp_to_sint $fp) $ptr) to (store (TruncIntFP $fp), $ptr).
2284 static SDValue lowerFP_TO_SINT_STORE(StoreSDNode *SD, SelectionDAG &DAG) {
2285 SDValue Val = SD->getValue();
2287 if (Val.getOpcode() != ISD::FP_TO_SINT)
2290 EVT FPTy = EVT::getFloatingPointVT(Val.getValueSizeInBits());
2291 SDValue Tr = DAG.getNode(MipsISD::TruncIntFP, SDLoc(Val), FPTy,
2294 return DAG.getStore(SD->getChain(), SDLoc(SD), Tr, SD->getBasePtr(),
2295 SD->getPointerInfo(), SD->isVolatile(),
2296 SD->isNonTemporal(), SD->getAlignment());
2299 SDValue MipsTargetLowering::lowerSTORE(SDValue Op, SelectionDAG &DAG) const {
2300 StoreSDNode *SD = cast<StoreSDNode>(Op);
2301 EVT MemVT = SD->getMemoryVT();
2303 // Lower unaligned integer stores.
2304 if (!Subtarget.systemSupportsUnalignedAccess() &&
2305 (SD->getAlignment() < MemVT.getSizeInBits() / 8) &&
2306 ((MemVT == MVT::i32) || (MemVT == MVT::i64)))
2307 return lowerUnalignedIntStore(SD, DAG, Subtarget.isLittle());
2309 return lowerFP_TO_SINT_STORE(SD, DAG);
2312 SDValue MipsTargetLowering::lowerADD(SDValue Op, SelectionDAG &DAG) const {
2313 if (Op->getOperand(0).getOpcode() != ISD::FRAMEADDR
2314 || cast<ConstantSDNode>
2315 (Op->getOperand(0).getOperand(0))->getZExtValue() != 0
2316 || Op->getOperand(1).getOpcode() != ISD::FRAME_TO_ARGS_OFFSET)
2320 // (add (frameaddr 0), (frame_to_args_offset))
2321 // results from lowering llvm.eh.dwarf.cfa intrinsic. Transform it to
2322 // (add FrameObject, 0)
2323 // where FrameObject is a fixed StackObject with offset 0 which points to
2324 // the old stack pointer.
2325 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
2326 EVT ValTy = Op->getValueType(0);
2327 int FI = MFI->CreateFixedObject(Op.getValueSizeInBits() / 8, 0, false);
2328 SDValue InArgsAddr = DAG.getFrameIndex(FI, ValTy);
2330 return DAG.getNode(ISD::ADD, DL, ValTy, InArgsAddr,
2331 DAG.getConstant(0, DL, ValTy));
2334 SDValue MipsTargetLowering::lowerFP_TO_SINT(SDValue Op,
2335 SelectionDAG &DAG) const {
2336 EVT FPTy = EVT::getFloatingPointVT(Op.getValueSizeInBits());
2337 SDValue Trunc = DAG.getNode(MipsISD::TruncIntFP, SDLoc(Op), FPTy,
2339 return DAG.getNode(ISD::BITCAST, SDLoc(Op), Op.getValueType(), Trunc);
2342 //===----------------------------------------------------------------------===//
2343 // Calling Convention Implementation
2344 //===----------------------------------------------------------------------===//
2346 //===----------------------------------------------------------------------===//
2347 // TODO: Implement a generic logic using tblgen that can support this.
2348 // Mips O32 ABI rules:
2350 // i32 - Passed in A0, A1, A2, A3 and stack
2351 // f32 - Only passed in f32 registers if no int reg has been used yet to hold
2352 // an argument. Otherwise, passed in A1, A2, A3 and stack.
2353 // f64 - Only passed in two aliased f32 registers if no int reg has been used
2354 // yet to hold an argument. Otherwise, use A2, A3 and stack. If A1 is
2355 // not used, it must be shadowed. If only A3 is available, shadow it and
2358 // For vararg functions, all arguments are passed in A0, A1, A2, A3 and stack.
2359 //===----------------------------------------------------------------------===//
2361 static bool CC_MipsO32(unsigned ValNo, MVT ValVT, MVT LocVT,
2362 CCValAssign::LocInfo LocInfo, ISD::ArgFlagsTy ArgFlags,
2363 CCState &State, ArrayRef<MCPhysReg> F64Regs) {
2364 const MipsSubtarget &Subtarget = static_cast<const MipsSubtarget &>(
2365 State.getMachineFunction().getSubtarget());
2367 static const MCPhysReg IntRegs[] = { Mips::A0, Mips::A1, Mips::A2, Mips::A3 };
2368 static const MCPhysReg F32Regs[] = { Mips::F12, Mips::F14 };
2370 // Do not process byval args here.
2371 if (ArgFlags.isByVal())
2374 // Promote i8 and i16
2375 if (ArgFlags.isInReg() && !Subtarget.isLittle()) {
2376 if (LocVT == MVT::i8 || LocVT == MVT::i16 || LocVT == MVT::i32) {
2378 if (ArgFlags.isSExt())
2379 LocInfo = CCValAssign::SExtUpper;
2380 else if (ArgFlags.isZExt())
2381 LocInfo = CCValAssign::ZExtUpper;
2383 LocInfo = CCValAssign::AExtUpper;
2387 // Promote i8 and i16
2388 if (LocVT == MVT::i8 || LocVT == MVT::i16) {
2390 if (ArgFlags.isSExt())
2391 LocInfo = CCValAssign::SExt;
2392 else if (ArgFlags.isZExt())
2393 LocInfo = CCValAssign::ZExt;
2395 LocInfo = CCValAssign::AExt;
2400 // f32 and f64 are allocated in A0, A1, A2, A3 when either of the following
2401 // is true: function is vararg, argument is 3rd or higher, there is previous
2402 // argument which is not f32 or f64.
2403 bool AllocateFloatsInIntReg = State.isVarArg() || ValNo > 1 ||
2404 State.getFirstUnallocated(F32Regs) != ValNo;
2405 unsigned OrigAlign = ArgFlags.getOrigAlign();
2406 bool isI64 = (ValVT == MVT::i32 && OrigAlign == 8);
2408 if (ValVT == MVT::i32 || (ValVT == MVT::f32 && AllocateFloatsInIntReg)) {
2409 Reg = State.AllocateReg(IntRegs);
2410 // If this is the first part of an i64 arg,
2411 // the allocated register must be either A0 or A2.
2412 if (isI64 && (Reg == Mips::A1 || Reg == Mips::A3))
2413 Reg = State.AllocateReg(IntRegs);
2415 } else if (ValVT == MVT::f64 && AllocateFloatsInIntReg) {
2416 // Allocate int register and shadow next int register. If first
2417 // available register is Mips::A1 or Mips::A3, shadow it too.
2418 Reg = State.AllocateReg(IntRegs);
2419 if (Reg == Mips::A1 || Reg == Mips::A3)
2420 Reg = State.AllocateReg(IntRegs);
2421 State.AllocateReg(IntRegs);
2423 } else if (ValVT.isFloatingPoint() && !AllocateFloatsInIntReg) {
2424 // we are guaranteed to find an available float register
2425 if (ValVT == MVT::f32) {
2426 Reg = State.AllocateReg(F32Regs);
2427 // Shadow int register
2428 State.AllocateReg(IntRegs);
2430 Reg = State.AllocateReg(F64Regs);
2431 // Shadow int registers
2432 unsigned Reg2 = State.AllocateReg(IntRegs);
2433 if (Reg2 == Mips::A1 || Reg2 == Mips::A3)
2434 State.AllocateReg(IntRegs);
2435 State.AllocateReg(IntRegs);
2438 llvm_unreachable("Cannot handle this ValVT.");
2441 unsigned Offset = State.AllocateStack(ValVT.getSizeInBits() >> 3,
2443 State.addLoc(CCValAssign::getMem(ValNo, ValVT, Offset, LocVT, LocInfo));
2445 State.addLoc(CCValAssign::getReg(ValNo, ValVT, Reg, LocVT, LocInfo));
2450 static bool CC_MipsO32_FP32(unsigned ValNo, MVT ValVT,
2451 MVT LocVT, CCValAssign::LocInfo LocInfo,
2452 ISD::ArgFlagsTy ArgFlags, CCState &State) {
2453 static const MCPhysReg F64Regs[] = { Mips::D6, Mips::D7 };
2455 return CC_MipsO32(ValNo, ValVT, LocVT, LocInfo, ArgFlags, State, F64Regs);
2458 static bool CC_MipsO32_FP64(unsigned ValNo, MVT ValVT,
2459 MVT LocVT, CCValAssign::LocInfo LocInfo,
2460 ISD::ArgFlagsTy ArgFlags, CCState &State) {
2461 static const MCPhysReg F64Regs[] = { Mips::D12_64, Mips::D14_64 };
2463 return CC_MipsO32(ValNo, ValVT, LocVT, LocInfo, ArgFlags, State, F64Regs);
2466 static bool CC_MipsO32(unsigned ValNo, MVT ValVT, MVT LocVT,
2467 CCValAssign::LocInfo LocInfo, ISD::ArgFlagsTy ArgFlags,
2468 CCState &State) LLVM_ATTRIBUTE_UNUSED;
2470 #include "MipsGenCallingConv.inc"
2472 //===----------------------------------------------------------------------===//
2473 // Call Calling Convention Implementation
2474 //===----------------------------------------------------------------------===//
2476 // Return next O32 integer argument register.
2477 static unsigned getNextIntArgReg(unsigned Reg) {
2478 assert((Reg == Mips::A0) || (Reg == Mips::A2));
2479 return (Reg == Mips::A0) ? Mips::A1 : Mips::A3;
2483 MipsTargetLowering::passArgOnStack(SDValue StackPtr, unsigned Offset,
2484 SDValue Chain, SDValue Arg, SDLoc DL,
2485 bool IsTailCall, SelectionDAG &DAG) const {
2488 DAG.getNode(ISD::ADD, DL, getPointerTy(DAG.getDataLayout()), StackPtr,
2489 DAG.getIntPtrConstant(Offset, DL));
2490 return DAG.getStore(Chain, DL, Arg, PtrOff, MachinePointerInfo(), false,
2494 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
2495 int FI = MFI->CreateFixedObject(Arg.getValueSizeInBits() / 8, Offset, false);
2496 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy(DAG.getDataLayout()));
2497 return DAG.getStore(Chain, DL, Arg, FIN, MachinePointerInfo(),
2498 /*isVolatile=*/ true, false, 0);
2501 void MipsTargetLowering::
2502 getOpndList(SmallVectorImpl<SDValue> &Ops,
2503 std::deque< std::pair<unsigned, SDValue> > &RegsToPass,
2504 bool IsPICCall, bool GlobalOrExternal, bool InternalLinkage,
2505 bool IsCallReloc, CallLoweringInfo &CLI, SDValue Callee,
2506 SDValue Chain) const {
2507 // Insert node "GP copy globalreg" before call to function.
2509 // R_MIPS_CALL* operators (emitted when non-internal functions are called
2510 // in PIC mode) allow symbols to be resolved via lazy binding.
2511 // The lazy binding stub requires GP to point to the GOT.
2512 // Note that we don't need GP to point to the GOT for indirect calls
2513 // (when R_MIPS_CALL* is not used for the call) because Mips linker generates
2514 // lazy binding stub for a function only when R_MIPS_CALL* are the only relocs
2515 // used for the function (that is, Mips linker doesn't generate lazy binding
2516 // stub for a function whose address is taken in the program).
2517 if (IsPICCall && !InternalLinkage && IsCallReloc) {
2518 unsigned GPReg = ABI.IsN64() ? Mips::GP_64 : Mips::GP;
2519 EVT Ty = ABI.IsN64() ? MVT::i64 : MVT::i32;
2520 RegsToPass.push_back(std::make_pair(GPReg, getGlobalReg(CLI.DAG, Ty)));
2523 // Build a sequence of copy-to-reg nodes chained together with token
2524 // chain and flag operands which copy the outgoing args into registers.
2525 // The InFlag in necessary since all emitted instructions must be
2529 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
2530 Chain = CLI.DAG.getCopyToReg(Chain, CLI.DL, RegsToPass[i].first,
2531 RegsToPass[i].second, InFlag);
2532 InFlag = Chain.getValue(1);
2535 // Add argument registers to the end of the list so that they are
2536 // known live into the call.
2537 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
2538 Ops.push_back(CLI.DAG.getRegister(RegsToPass[i].first,
2539 RegsToPass[i].second.getValueType()));
2541 // Add a register mask operand representing the call-preserved registers.
2542 const TargetRegisterInfo *TRI = Subtarget.getRegisterInfo();
2543 const uint32_t *Mask =
2544 TRI->getCallPreservedMask(CLI.DAG.getMachineFunction(), CLI.CallConv);
2545 assert(Mask && "Missing call preserved mask for calling convention");
2546 if (Subtarget.inMips16HardFloat()) {
2547 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(CLI.Callee)) {
2548 llvm::StringRef Sym = G->getGlobal()->getName();
2549 Function *F = G->getGlobal()->getParent()->getFunction(Sym);
2550 if (F && F->hasFnAttribute("__Mips16RetHelper")) {
2551 Mask = MipsRegisterInfo::getMips16RetHelperMask();
2555 Ops.push_back(CLI.DAG.getRegisterMask(Mask));
2557 if (InFlag.getNode())
2558 Ops.push_back(InFlag);
2561 /// LowerCall - functions arguments are copied from virtual regs to
2562 /// (physical regs)/(stack frame), CALLSEQ_START and CALLSEQ_END are emitted.
2564 MipsTargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI,
2565 SmallVectorImpl<SDValue> &InVals) const {
2566 SelectionDAG &DAG = CLI.DAG;
2568 SmallVectorImpl<ISD::OutputArg> &Outs = CLI.Outs;
2569 SmallVectorImpl<SDValue> &OutVals = CLI.OutVals;
2570 SmallVectorImpl<ISD::InputArg> &Ins = CLI.Ins;
2571 SDValue Chain = CLI.Chain;
2572 SDValue Callee = CLI.Callee;
2573 bool &IsTailCall = CLI.IsTailCall;
2574 CallingConv::ID CallConv = CLI.CallConv;
2575 bool IsVarArg = CLI.IsVarArg;
2577 MachineFunction &MF = DAG.getMachineFunction();
2578 MachineFrameInfo *MFI = MF.getFrameInfo();
2579 const TargetFrameLowering *TFL = Subtarget.getFrameLowering();
2580 MipsFunctionInfo *FuncInfo = MF.getInfo<MipsFunctionInfo>();
2581 bool IsPIC = getTargetMachine().getRelocationModel() == Reloc::PIC_;
2583 // Analyze operands of the call, assigning locations to each operand.
2584 SmallVector<CCValAssign, 16> ArgLocs;
2586 CallConv, IsVarArg, DAG.getMachineFunction(), ArgLocs, *DAG.getContext(),
2587 MipsCCState::getSpecialCallingConvForCallee(Callee.getNode(), Subtarget));
2589 // Allocate the reserved argument area. It seems strange to do this from the
2590 // caller side but removing it breaks the frame size calculation.
2591 CCInfo.AllocateStack(ABI.GetCalleeAllocdArgSizeInBytes(CallConv), 1);
2593 CCInfo.AnalyzeCallOperands(Outs, CC_Mips, CLI.getArgs(), Callee.getNode());
2595 // Get a count of how many bytes are to be pushed on the stack.
2596 unsigned NextStackOffset = CCInfo.getNextStackOffset();
2598 // Check if it's really possible to do a tail call.
2600 IsTailCall = isEligibleForTailCallOptimization(
2601 CCInfo, NextStackOffset, *MF.getInfo<MipsFunctionInfo>());
2603 if (!IsTailCall && CLI.CS && CLI.CS->isMustTailCall())
2604 report_fatal_error("failed to perform tail call elimination on a call "
2605 "site marked musttail");
2610 // Chain is the output chain of the last Load/Store or CopyToReg node.
2611 // ByValChain is the output chain of the last Memcpy node created for copying
2612 // byval arguments to the stack.
2613 unsigned StackAlignment = TFL->getStackAlignment();
2614 NextStackOffset = RoundUpToAlignment(NextStackOffset, StackAlignment);
2615 SDValue NextStackOffsetVal = DAG.getIntPtrConstant(NextStackOffset, DL, true);
2618 Chain = DAG.getCALLSEQ_START(Chain, NextStackOffsetVal, DL);
2621 DAG.getCopyFromReg(Chain, DL, ABI.IsN64() ? Mips::SP_64 : Mips::SP,
2622 getPointerTy(DAG.getDataLayout()));
2624 // With EABI is it possible to have 16 args on registers.
2625 std::deque< std::pair<unsigned, SDValue> > RegsToPass;
2626 SmallVector<SDValue, 8> MemOpChains;
2628 CCInfo.rewindByValRegsInfo();
2630 // Walk the register/memloc assignments, inserting copies/loads.
2631 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2632 SDValue Arg = OutVals[i];
2633 CCValAssign &VA = ArgLocs[i];
2634 MVT ValVT = VA.getValVT(), LocVT = VA.getLocVT();
2635 ISD::ArgFlagsTy Flags = Outs[i].Flags;
2636 bool UseUpperBits = false;
2639 if (Flags.isByVal()) {
2640 unsigned FirstByValReg, LastByValReg;
2641 unsigned ByValIdx = CCInfo.getInRegsParamsProcessed();
2642 CCInfo.getInRegsParamInfo(ByValIdx, FirstByValReg, LastByValReg);
2644 assert(Flags.getByValSize() &&
2645 "ByVal args of size 0 should have been ignored by front-end.");
2646 assert(ByValIdx < CCInfo.getInRegsParamsCount());
2647 assert(!IsTailCall &&
2648 "Do not tail-call optimize if there is a byval argument.");
2649 passByValArg(Chain, DL, RegsToPass, MemOpChains, StackPtr, MFI, DAG, Arg,
2650 FirstByValReg, LastByValReg, Flags, Subtarget.isLittle(),
2652 CCInfo.nextInRegsParam();
2656 // Promote the value if needed.
2657 switch (VA.getLocInfo()) {
2659 llvm_unreachable("Unknown loc info!");
2660 case CCValAssign::Full:
2661 if (VA.isRegLoc()) {
2662 if ((ValVT == MVT::f32 && LocVT == MVT::i32) ||
2663 (ValVT == MVT::f64 && LocVT == MVT::i64) ||
2664 (ValVT == MVT::i64 && LocVT == MVT::f64))
2665 Arg = DAG.getNode(ISD::BITCAST, DL, LocVT, Arg);
2666 else if (ValVT == MVT::f64 && LocVT == MVT::i32) {
2667 SDValue Lo = DAG.getNode(MipsISD::ExtractElementF64, DL, MVT::i32,
2668 Arg, DAG.getConstant(0, DL, MVT::i32));
2669 SDValue Hi = DAG.getNode(MipsISD::ExtractElementF64, DL, MVT::i32,
2670 Arg, DAG.getConstant(1, DL, MVT::i32));
2671 if (!Subtarget.isLittle())
2673 unsigned LocRegLo = VA.getLocReg();
2674 unsigned LocRegHigh = getNextIntArgReg(LocRegLo);
2675 RegsToPass.push_back(std::make_pair(LocRegLo, Lo));
2676 RegsToPass.push_back(std::make_pair(LocRegHigh, Hi));
2681 case CCValAssign::BCvt:
2682 Arg = DAG.getNode(ISD::BITCAST, DL, LocVT, Arg);
2684 case CCValAssign::SExtUpper:
2685 UseUpperBits = true;
2687 case CCValAssign::SExt:
2688 Arg = DAG.getNode(ISD::SIGN_EXTEND, DL, LocVT, Arg);
2690 case CCValAssign::ZExtUpper:
2691 UseUpperBits = true;
2693 case CCValAssign::ZExt:
2694 Arg = DAG.getNode(ISD::ZERO_EXTEND, DL, LocVT, Arg);
2696 case CCValAssign::AExtUpper:
2697 UseUpperBits = true;
2699 case CCValAssign::AExt:
2700 Arg = DAG.getNode(ISD::ANY_EXTEND, DL, LocVT, Arg);
2705 unsigned ValSizeInBits = Outs[i].ArgVT.getSizeInBits();
2706 unsigned LocSizeInBits = VA.getLocVT().getSizeInBits();
2708 ISD::SHL, DL, VA.getLocVT(), Arg,
2709 DAG.getConstant(LocSizeInBits - ValSizeInBits, DL, VA.getLocVT()));
2712 // Arguments that can be passed on register must be kept at
2713 // RegsToPass vector
2714 if (VA.isRegLoc()) {
2715 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
2719 // Register can't get to this point...
2720 assert(VA.isMemLoc());
2722 // emit ISD::STORE whichs stores the
2723 // parameter value to a stack Location
2724 MemOpChains.push_back(passArgOnStack(StackPtr, VA.getLocMemOffset(),
2725 Chain, Arg, DL, IsTailCall, DAG));
2728 // Transform all store nodes into one single node because all store
2729 // nodes are independent of each other.
2730 if (!MemOpChains.empty())
2731 Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other, MemOpChains);
2733 // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every
2734 // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol
2735 // node so that legalize doesn't hack it.
2736 bool IsPICCall = (ABI.IsN64() || IsPIC); // true if calls are translated to
2738 bool GlobalOrExternal = false, InternalLinkage = false, IsCallReloc = false;
2740 EVT Ty = Callee.getValueType();
2742 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
2744 const GlobalValue *Val = G->getGlobal();
2745 InternalLinkage = Val->hasInternalLinkage();
2747 if (InternalLinkage)
2748 Callee = getAddrLocal(G, DL, Ty, DAG, ABI.IsN32() || ABI.IsN64());
2749 else if (LargeGOT) {
2750 Callee = getAddrGlobalLargeGOT(G, DL, Ty, DAG, MipsII::MO_CALL_HI16,
2751 MipsII::MO_CALL_LO16, Chain,
2752 FuncInfo->callPtrInfo(Val));
2755 Callee = getAddrGlobal(G, DL, Ty, DAG, MipsII::MO_GOT_CALL, Chain,
2756 FuncInfo->callPtrInfo(Val));
2760 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), DL,
2761 getPointerTy(DAG.getDataLayout()), 0,
2762 MipsII::MO_NO_FLAG);
2763 GlobalOrExternal = true;
2765 else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
2766 const char *Sym = S->getSymbol();
2768 if (!ABI.IsN64() && !IsPIC) // !N64 && static
2769 Callee = DAG.getTargetExternalSymbol(
2770 Sym, getPointerTy(DAG.getDataLayout()), MipsII::MO_NO_FLAG);
2771 else if (LargeGOT) {
2772 Callee = getAddrGlobalLargeGOT(S, DL, Ty, DAG, MipsII::MO_CALL_HI16,
2773 MipsII::MO_CALL_LO16, Chain,
2774 FuncInfo->callPtrInfo(Sym));
2776 } else { // N64 || PIC
2777 Callee = getAddrGlobal(S, DL, Ty, DAG, MipsII::MO_GOT_CALL, Chain,
2778 FuncInfo->callPtrInfo(Sym));
2782 GlobalOrExternal = true;
2785 SmallVector<SDValue, 8> Ops(1, Chain);
2786 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
2788 getOpndList(Ops, RegsToPass, IsPICCall, GlobalOrExternal, InternalLinkage,
2789 IsCallReloc, CLI, Callee, Chain);
2792 return DAG.getNode(MipsISD::TailCall, DL, MVT::Other, Ops);
2794 Chain = DAG.getNode(MipsISD::JmpLink, DL, NodeTys, Ops);
2795 SDValue InFlag = Chain.getValue(1);
2797 // Create the CALLSEQ_END node.
2798 Chain = DAG.getCALLSEQ_END(Chain, NextStackOffsetVal,
2799 DAG.getIntPtrConstant(0, DL, true), InFlag, DL);
2800 InFlag = Chain.getValue(1);
2802 // Handle result values, copying them out of physregs into vregs that we
2804 return LowerCallResult(Chain, InFlag, CallConv, IsVarArg, Ins, DL, DAG,
2808 /// LowerCallResult - Lower the result values of a call into the
2809 /// appropriate copies out of appropriate physical registers.
2810 SDValue MipsTargetLowering::LowerCallResult(
2811 SDValue Chain, SDValue InFlag, CallingConv::ID CallConv, bool IsVarArg,
2812 const SmallVectorImpl<ISD::InputArg> &Ins, SDLoc DL, SelectionDAG &DAG,
2813 SmallVectorImpl<SDValue> &InVals,
2814 TargetLowering::CallLoweringInfo &CLI) const {
2815 // Assign locations to each value returned by this call.
2816 SmallVector<CCValAssign, 16> RVLocs;
2817 MipsCCState CCInfo(CallConv, IsVarArg, DAG.getMachineFunction(), RVLocs,
2819 CCInfo.AnalyzeCallResult(Ins, RetCC_Mips, CLI);
2821 // Copy all of the result registers out of their specified physreg.
2822 for (unsigned i = 0; i != RVLocs.size(); ++i) {
2823 CCValAssign &VA = RVLocs[i];
2824 assert(VA.isRegLoc() && "Can only return in registers!");
2826 SDValue Val = DAG.getCopyFromReg(Chain, DL, RVLocs[i].getLocReg(),
2827 RVLocs[i].getLocVT(), InFlag);
2828 Chain = Val.getValue(1);
2829 InFlag = Val.getValue(2);
2831 if (VA.isUpperBitsInLoc()) {
2832 unsigned ValSizeInBits = Ins[i].ArgVT.getSizeInBits();
2833 unsigned LocSizeInBits = VA.getLocVT().getSizeInBits();
2835 VA.getLocInfo() == CCValAssign::ZExtUpper ? ISD::SRL : ISD::SRA;
2837 Shift, DL, VA.getLocVT(), Val,
2838 DAG.getConstant(LocSizeInBits - ValSizeInBits, DL, VA.getLocVT()));
2841 switch (VA.getLocInfo()) {
2843 llvm_unreachable("Unknown loc info!");
2844 case CCValAssign::Full:
2846 case CCValAssign::BCvt:
2847 Val = DAG.getNode(ISD::BITCAST, DL, VA.getValVT(), Val);
2849 case CCValAssign::AExt:
2850 case CCValAssign::AExtUpper:
2851 Val = DAG.getNode(ISD::TRUNCATE, DL, VA.getValVT(), Val);
2853 case CCValAssign::ZExt:
2854 case CCValAssign::ZExtUpper:
2855 Val = DAG.getNode(ISD::AssertZext, DL, VA.getLocVT(), Val,
2856 DAG.getValueType(VA.getValVT()));
2857 Val = DAG.getNode(ISD::TRUNCATE, DL, VA.getValVT(), Val);
2859 case CCValAssign::SExt:
2860 case CCValAssign::SExtUpper:
2861 Val = DAG.getNode(ISD::AssertSext, DL, VA.getLocVT(), Val,
2862 DAG.getValueType(VA.getValVT()));
2863 Val = DAG.getNode(ISD::TRUNCATE, DL, VA.getValVT(), Val);
2867 InVals.push_back(Val);
2873 static SDValue UnpackFromArgumentSlot(SDValue Val, const CCValAssign &VA,
2874 EVT ArgVT, SDLoc DL, SelectionDAG &DAG) {
2875 MVT LocVT = VA.getLocVT();
2876 EVT ValVT = VA.getValVT();
2878 // Shift into the upper bits if necessary.
2879 switch (VA.getLocInfo()) {
2882 case CCValAssign::AExtUpper:
2883 case CCValAssign::SExtUpper:
2884 case CCValAssign::ZExtUpper: {
2885 unsigned ValSizeInBits = ArgVT.getSizeInBits();
2886 unsigned LocSizeInBits = VA.getLocVT().getSizeInBits();
2888 VA.getLocInfo() == CCValAssign::ZExtUpper ? ISD::SRL : ISD::SRA;
2890 Opcode, DL, VA.getLocVT(), Val,
2891 DAG.getConstant(LocSizeInBits - ValSizeInBits, DL, VA.getLocVT()));
2896 // If this is an value smaller than the argument slot size (32-bit for O32,
2897 // 64-bit for N32/N64), it has been promoted in some way to the argument slot
2898 // size. Extract the value and insert any appropriate assertions regarding
2899 // sign/zero extension.
2900 switch (VA.getLocInfo()) {
2902 llvm_unreachable("Unknown loc info!");
2903 case CCValAssign::Full:
2905 case CCValAssign::AExtUpper:
2906 case CCValAssign::AExt:
2907 Val = DAG.getNode(ISD::TRUNCATE, DL, ValVT, Val);
2909 case CCValAssign::SExtUpper:
2910 case CCValAssign::SExt:
2911 Val = DAG.getNode(ISD::AssertSext, DL, LocVT, Val, DAG.getValueType(ValVT));
2912 Val = DAG.getNode(ISD::TRUNCATE, DL, ValVT, Val);
2914 case CCValAssign::ZExtUpper:
2915 case CCValAssign::ZExt:
2916 Val = DAG.getNode(ISD::AssertZext, DL, LocVT, Val, DAG.getValueType(ValVT));
2917 Val = DAG.getNode(ISD::TRUNCATE, DL, ValVT, Val);
2919 case CCValAssign::BCvt:
2920 Val = DAG.getNode(ISD::BITCAST, DL, ValVT, Val);
2927 //===----------------------------------------------------------------------===//
2928 // Formal Arguments Calling Convention Implementation
2929 //===----------------------------------------------------------------------===//
2930 /// LowerFormalArguments - transform physical registers into virtual registers
2931 /// and generate load operations for arguments places on the stack.
2933 MipsTargetLowering::LowerFormalArguments(SDValue Chain,
2934 CallingConv::ID CallConv,
2936 const SmallVectorImpl<ISD::InputArg> &Ins,
2937 SDLoc DL, SelectionDAG &DAG,
2938 SmallVectorImpl<SDValue> &InVals)
2940 MachineFunction &MF = DAG.getMachineFunction();
2941 MachineFrameInfo *MFI = MF.getFrameInfo();
2942 MipsFunctionInfo *MipsFI = MF.getInfo<MipsFunctionInfo>();
2944 MipsFI->setVarArgsFrameIndex(0);
2946 // Used with vargs to acumulate store chains.
2947 std::vector<SDValue> OutChains;
2949 // Assign locations to all of the incoming arguments.
2950 SmallVector<CCValAssign, 16> ArgLocs;
2951 MipsCCState CCInfo(CallConv, IsVarArg, DAG.getMachineFunction(), ArgLocs,
2953 CCInfo.AllocateStack(ABI.GetCalleeAllocdArgSizeInBytes(CallConv), 1);
2954 const Function *Func = DAG.getMachineFunction().getFunction();
2955 Function::const_arg_iterator FuncArg = Func->arg_begin();
2957 if (Func->hasFnAttribute("interrupt") && !Func->arg_empty())
2959 "Functions with the interrupt attribute cannot have arguments!");
2961 CCInfo.AnalyzeFormalArguments(Ins, CC_Mips_FixedArg);
2962 MipsFI->setFormalArgInfo(CCInfo.getNextStackOffset(),
2963 CCInfo.getInRegsParamsCount() > 0);
2965 unsigned CurArgIdx = 0;
2966 CCInfo.rewindByValRegsInfo();
2968 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2969 CCValAssign &VA = ArgLocs[i];
2970 if (Ins[i].isOrigArg()) {
2971 std::advance(FuncArg, Ins[i].getOrigArgIndex() - CurArgIdx);
2972 CurArgIdx = Ins[i].getOrigArgIndex();
2974 EVT ValVT = VA.getValVT();
2975 ISD::ArgFlagsTy Flags = Ins[i].Flags;
2976 bool IsRegLoc = VA.isRegLoc();
2978 if (Flags.isByVal()) {
2979 assert(Ins[i].isOrigArg() && "Byval arguments cannot be implicit");
2980 unsigned FirstByValReg, LastByValReg;
2981 unsigned ByValIdx = CCInfo.getInRegsParamsProcessed();
2982 CCInfo.getInRegsParamInfo(ByValIdx, FirstByValReg, LastByValReg);
2984 assert(Flags.getByValSize() &&
2985 "ByVal args of size 0 should have been ignored by front-end.");
2986 assert(ByValIdx < CCInfo.getInRegsParamsCount());
2987 copyByValRegs(Chain, DL, OutChains, DAG, Flags, InVals, &*FuncArg,
2988 FirstByValReg, LastByValReg, VA, CCInfo);
2989 CCInfo.nextInRegsParam();
2993 // Arguments stored on registers
2995 MVT RegVT = VA.getLocVT();
2996 unsigned ArgReg = VA.getLocReg();
2997 const TargetRegisterClass *RC = getRegClassFor(RegVT);
2999 // Transform the arguments stored on
3000 // physical registers into virtual ones
3001 unsigned Reg = addLiveIn(DAG.getMachineFunction(), ArgReg, RC);
3002 SDValue ArgValue = DAG.getCopyFromReg(Chain, DL, Reg, RegVT);
3004 ArgValue = UnpackFromArgumentSlot(ArgValue, VA, Ins[i].ArgVT, DL, DAG);
3006 // Handle floating point arguments passed in integer registers and
3007 // long double arguments passed in floating point registers.
3008 if ((RegVT == MVT::i32 && ValVT == MVT::f32) ||
3009 (RegVT == MVT::i64 && ValVT == MVT::f64) ||
3010 (RegVT == MVT::f64 && ValVT == MVT::i64))
3011 ArgValue = DAG.getNode(ISD::BITCAST, DL, ValVT, ArgValue);
3012 else if (ABI.IsO32() && RegVT == MVT::i32 &&
3013 ValVT == MVT::f64) {
3014 unsigned Reg2 = addLiveIn(DAG.getMachineFunction(),
3015 getNextIntArgReg(ArgReg), RC);
3016 SDValue ArgValue2 = DAG.getCopyFromReg(Chain, DL, Reg2, RegVT);
3017 if (!Subtarget.isLittle())
3018 std::swap(ArgValue, ArgValue2);
3019 ArgValue = DAG.getNode(MipsISD::BuildPairF64, DL, MVT::f64,
3020 ArgValue, ArgValue2);
3023 InVals.push_back(ArgValue);
3024 } else { // VA.isRegLoc()
3025 MVT LocVT = VA.getLocVT();
3028 // We ought to be able to use LocVT directly but O32 sets it to i32
3029 // when allocating floating point values to integer registers.
3030 // This shouldn't influence how we load the value into registers unless
3031 // we are targeting softfloat.
3032 if (VA.getValVT().isFloatingPoint() && !Subtarget.useSoftFloat())
3033 LocVT = VA.getValVT();
3037 assert(VA.isMemLoc());
3039 // The stack pointer offset is relative to the caller stack frame.
3040 int FI = MFI->CreateFixedObject(LocVT.getSizeInBits() / 8,
3041 VA.getLocMemOffset(), true);
3043 // Create load nodes to retrieve arguments from the stack
3044 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy(DAG.getDataLayout()));
3045 SDValue ArgValue = DAG.getLoad(
3046 LocVT, DL, Chain, FIN,
3047 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), FI),
3048 false, false, false, 0);
3049 OutChains.push_back(ArgValue.getValue(1));
3051 ArgValue = UnpackFromArgumentSlot(ArgValue, VA, Ins[i].ArgVT, DL, DAG);
3053 InVals.push_back(ArgValue);
3057 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
3058 // The mips ABIs for returning structs by value requires that we copy
3059 // the sret argument into $v0 for the return. Save the argument into
3060 // a virtual register so that we can access it from the return points.
3061 if (Ins[i].Flags.isSRet()) {
3062 unsigned Reg = MipsFI->getSRetReturnReg();
3064 Reg = MF.getRegInfo().createVirtualRegister(
3065 getRegClassFor(ABI.IsN64() ? MVT::i64 : MVT::i32));
3066 MipsFI->setSRetReturnReg(Reg);
3068 SDValue Copy = DAG.getCopyToReg(DAG.getEntryNode(), DL, Reg, InVals[i]);
3069 Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other, Copy, Chain);
3075 writeVarArgRegs(OutChains, Chain, DL, DAG, CCInfo);
3077 // All stores are grouped in one node to allow the matching between
3078 // the size of Ins and InVals. This only happens when on varg functions
3079 if (!OutChains.empty()) {
3080 OutChains.push_back(Chain);
3081 Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other, OutChains);
3087 //===----------------------------------------------------------------------===//
3088 // Return Value Calling Convention Implementation
3089 //===----------------------------------------------------------------------===//
3092 MipsTargetLowering::CanLowerReturn(CallingConv::ID CallConv,
3093 MachineFunction &MF, bool IsVarArg,
3094 const SmallVectorImpl<ISD::OutputArg> &Outs,
3095 LLVMContext &Context) const {
3096 SmallVector<CCValAssign, 16> RVLocs;
3097 MipsCCState CCInfo(CallConv, IsVarArg, MF, RVLocs, Context);
3098 return CCInfo.CheckReturn(Outs, RetCC_Mips);
3102 MipsTargetLowering::shouldSignExtendTypeInLibCall(EVT Type, bool IsSigned) const {
3103 if (Subtarget.hasMips3() && Subtarget.useSoftFloat()) {
3104 if (Type == MVT::i32)
3111 MipsTargetLowering::LowerInterruptReturn(SmallVectorImpl<SDValue> &RetOps,
3112 SDLoc DL, SelectionDAG &DAG) const {
3114 MachineFunction &MF = DAG.getMachineFunction();
3115 MipsFunctionInfo *MipsFI = MF.getInfo<MipsFunctionInfo>();
3119 return DAG.getNode(MipsISD::ERet, DL, MVT::Other, RetOps);
3123 MipsTargetLowering::LowerReturn(SDValue Chain, CallingConv::ID CallConv,
3125 const SmallVectorImpl<ISD::OutputArg> &Outs,
3126 const SmallVectorImpl<SDValue> &OutVals,
3127 SDLoc DL, SelectionDAG &DAG) const {
3128 // CCValAssign - represent the assignment of
3129 // the return value to a location
3130 SmallVector<CCValAssign, 16> RVLocs;
3131 MachineFunction &MF = DAG.getMachineFunction();
3133 // CCState - Info about the registers and stack slot.
3134 MipsCCState CCInfo(CallConv, IsVarArg, MF, RVLocs, *DAG.getContext());
3136 // Analyze return values.
3137 CCInfo.AnalyzeReturn(Outs, RetCC_Mips);
3140 SmallVector<SDValue, 4> RetOps(1, Chain);
3142 // Copy the result values into the output registers.
3143 for (unsigned i = 0; i != RVLocs.size(); ++i) {
3144 SDValue Val = OutVals[i];
3145 CCValAssign &VA = RVLocs[i];
3146 assert(VA.isRegLoc() && "Can only return in registers!");
3147 bool UseUpperBits = false;
3149 switch (VA.getLocInfo()) {
3151 llvm_unreachable("Unknown loc info!");
3152 case CCValAssign::Full:
3154 case CCValAssign::BCvt:
3155 Val = DAG.getNode(ISD::BITCAST, DL, VA.getLocVT(), Val);
3157 case CCValAssign::AExtUpper:
3158 UseUpperBits = true;
3160 case CCValAssign::AExt:
3161 Val = DAG.getNode(ISD::ANY_EXTEND, DL, VA.getLocVT(), Val);
3163 case CCValAssign::ZExtUpper:
3164 UseUpperBits = true;
3166 case CCValAssign::ZExt:
3167 Val = DAG.getNode(ISD::ZERO_EXTEND, DL, VA.getLocVT(), Val);
3169 case CCValAssign::SExtUpper:
3170 UseUpperBits = true;
3172 case CCValAssign::SExt:
3173 Val = DAG.getNode(ISD::SIGN_EXTEND, DL, VA.getLocVT(), Val);
3178 unsigned ValSizeInBits = Outs[i].ArgVT.getSizeInBits();
3179 unsigned LocSizeInBits = VA.getLocVT().getSizeInBits();
3181 ISD::SHL, DL, VA.getLocVT(), Val,
3182 DAG.getConstant(LocSizeInBits - ValSizeInBits, DL, VA.getLocVT()));
3185 Chain = DAG.getCopyToReg(Chain, DL, VA.getLocReg(), Val, Flag);
3187 // Guarantee that all emitted copies are stuck together with flags.
3188 Flag = Chain.getValue(1);
3189 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
3192 // The mips ABIs for returning structs by value requires that we copy
3193 // the sret argument into $v0 for the return. We saved the argument into
3194 // a virtual register in the entry block, so now we copy the value out
3196 if (MF.getFunction()->hasStructRetAttr()) {
3197 MipsFunctionInfo *MipsFI = MF.getInfo<MipsFunctionInfo>();
3198 unsigned Reg = MipsFI->getSRetReturnReg();
3201 llvm_unreachable("sret virtual register not created in the entry block");
3203 DAG.getCopyFromReg(Chain, DL, Reg, getPointerTy(DAG.getDataLayout()));
3204 unsigned V0 = ABI.IsN64() ? Mips::V0_64 : Mips::V0;
3206 Chain = DAG.getCopyToReg(Chain, DL, V0, Val, Flag);
3207 Flag = Chain.getValue(1);
3208 RetOps.push_back(DAG.getRegister(V0, getPointerTy(DAG.getDataLayout())));
3211 RetOps[0] = Chain; // Update chain.
3213 // Add the flag if we have it.
3215 RetOps.push_back(Flag);
3217 // ISRs must use "eret".
3218 if (DAG.getMachineFunction().getFunction()->hasFnAttribute("interrupt"))
3219 return LowerInterruptReturn(RetOps, DL, DAG);
3221 // Standard return on Mips is a "jr $ra"
3222 return DAG.getNode(MipsISD::Ret, DL, MVT::Other, RetOps);
3225 //===----------------------------------------------------------------------===//
3226 // Mips Inline Assembly Support
3227 //===----------------------------------------------------------------------===//
3229 /// getConstraintType - Given a constraint letter, return the type of
3230 /// constraint it is for this target.
3231 MipsTargetLowering::ConstraintType
3232 MipsTargetLowering::getConstraintType(StringRef Constraint) const {
3233 // Mips specific constraints
3234 // GCC config/mips/constraints.md
3236 // 'd' : An address register. Equivalent to r
3237 // unless generating MIPS16 code.
3238 // 'y' : Equivalent to r; retained for
3239 // backwards compatibility.
3240 // 'c' : A register suitable for use in an indirect
3241 // jump. This will always be $25 for -mabicalls.
3242 // 'l' : The lo register. 1 word storage.
3243 // 'x' : The hilo register pair. Double word storage.
3244 if (Constraint.size() == 1) {
3245 switch (Constraint[0]) {
3253 return C_RegisterClass;
3259 if (Constraint == "ZC")
3262 return TargetLowering::getConstraintType(Constraint);
3265 /// Examine constraint type and operand type and determine a weight value.
3266 /// This object must already have been set up with the operand type
3267 /// and the current alternative constraint selected.
3268 TargetLowering::ConstraintWeight
3269 MipsTargetLowering::getSingleConstraintMatchWeight(
3270 AsmOperandInfo &info, const char *constraint) const {
3271 ConstraintWeight weight = CW_Invalid;
3272 Value *CallOperandVal = info.CallOperandVal;
3273 // If we don't have a value, we can't do a match,
3274 // but allow it at the lowest weight.
3275 if (!CallOperandVal)
3277 Type *type = CallOperandVal->getType();
3278 // Look at the constraint type.
3279 switch (*constraint) {
3281 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
3285 if (type->isIntegerTy())
3286 weight = CW_Register;
3288 case 'f': // FPU or MSA register
3289 if (Subtarget.hasMSA() && type->isVectorTy() &&
3290 cast<VectorType>(type)->getBitWidth() == 128)
3291 weight = CW_Register;
3292 else if (type->isFloatTy())
3293 weight = CW_Register;
3295 case 'c': // $25 for indirect jumps
3296 case 'l': // lo register
3297 case 'x': // hilo register pair
3298 if (type->isIntegerTy())
3299 weight = CW_SpecificReg;
3301 case 'I': // signed 16 bit immediate
3302 case 'J': // integer zero
3303 case 'K': // unsigned 16 bit immediate
3304 case 'L': // signed 32 bit immediate where lower 16 bits are 0
3305 case 'N': // immediate in the range of -65535 to -1 (inclusive)
3306 case 'O': // signed 15 bit immediate (+- 16383)
3307 case 'P': // immediate in the range of 65535 to 1 (inclusive)
3308 if (isa<ConstantInt>(CallOperandVal))
3309 weight = CW_Constant;
3318 /// This is a helper function to parse a physical register string and split it
3319 /// into non-numeric and numeric parts (Prefix and Reg). The first boolean flag
3320 /// that is returned indicates whether parsing was successful. The second flag
3321 /// is true if the numeric part exists.
3322 static std::pair<bool, bool> parsePhysicalReg(StringRef C, StringRef &Prefix,
3323 unsigned long long &Reg) {
3324 if (C.front() != '{' || C.back() != '}')
3325 return std::make_pair(false, false);
3327 // Search for the first numeric character.
3328 StringRef::const_iterator I, B = C.begin() + 1, E = C.end() - 1;
3329 I = std::find_if(B, E, std::ptr_fun(isdigit));
3331 Prefix = StringRef(B, I - B);
3333 // The second flag is set to false if no numeric characters were found.
3335 return std::make_pair(true, false);
3337 // Parse the numeric characters.
3338 return std::make_pair(!getAsUnsignedInteger(StringRef(I, E - I), 10, Reg),
3342 std::pair<unsigned, const TargetRegisterClass *> MipsTargetLowering::
3343 parseRegForInlineAsmConstraint(StringRef C, MVT VT) const {
3344 const TargetRegisterInfo *TRI =
3345 Subtarget.getRegisterInfo();
3346 const TargetRegisterClass *RC;
3348 unsigned long long Reg;
3350 std::pair<bool, bool> R = parsePhysicalReg(C, Prefix, Reg);
3353 return std::make_pair(0U, nullptr);
3355 if ((Prefix == "hi" || Prefix == "lo")) { // Parse hi/lo.
3356 // No numeric characters follow "hi" or "lo".
3358 return std::make_pair(0U, nullptr);
3360 RC = TRI->getRegClass(Prefix == "hi" ?
3361 Mips::HI32RegClassID : Mips::LO32RegClassID);
3362 return std::make_pair(*(RC->begin()), RC);
3363 } else if (Prefix.startswith("$msa")) {
3364 // Parse $msa(ir|csr|access|save|modify|request|map|unmap)
3366 // No numeric characters follow the name.
3368 return std::make_pair(0U, nullptr);
3370 Reg = StringSwitch<unsigned long long>(Prefix)
3371 .Case("$msair", Mips::MSAIR)
3372 .Case("$msacsr", Mips::MSACSR)
3373 .Case("$msaaccess", Mips::MSAAccess)
3374 .Case("$msasave", Mips::MSASave)
3375 .Case("$msamodify", Mips::MSAModify)
3376 .Case("$msarequest", Mips::MSARequest)
3377 .Case("$msamap", Mips::MSAMap)
3378 .Case("$msaunmap", Mips::MSAUnmap)
3382 return std::make_pair(0U, nullptr);
3384 RC = TRI->getRegClass(Mips::MSACtrlRegClassID);
3385 return std::make_pair(Reg, RC);
3389 return std::make_pair(0U, nullptr);
3391 if (Prefix == "$f") { // Parse $f0-$f31.
3392 // If the size of FP registers is 64-bit or Reg is an even number, select
3393 // the 64-bit register class. Otherwise, select the 32-bit register class.
3394 if (VT == MVT::Other)
3395 VT = (Subtarget.isFP64bit() || !(Reg % 2)) ? MVT::f64 : MVT::f32;
3397 RC = getRegClassFor(VT);
3399 if (RC == &Mips::AFGR64RegClass) {
3400 assert(Reg % 2 == 0);
3403 } else if (Prefix == "$fcc") // Parse $fcc0-$fcc7.
3404 RC = TRI->getRegClass(Mips::FCCRegClassID);
3405 else if (Prefix == "$w") { // Parse $w0-$w31.
3406 RC = getRegClassFor((VT == MVT::Other) ? MVT::v16i8 : VT);
3407 } else { // Parse $0-$31.
3408 assert(Prefix == "$");
3409 RC = getRegClassFor((VT == MVT::Other) ? MVT::i32 : VT);
3412 assert(Reg < RC->getNumRegs());
3413 return std::make_pair(*(RC->begin() + Reg), RC);
3416 /// Given a register class constraint, like 'r', if this corresponds directly
3417 /// to an LLVM register class, return a register of 0 and the register class
3419 std::pair<unsigned, const TargetRegisterClass *>
3420 MipsTargetLowering::getRegForInlineAsmConstraint(const TargetRegisterInfo *TRI,
3421 StringRef Constraint,
3423 if (Constraint.size() == 1) {
3424 switch (Constraint[0]) {
3425 case 'd': // Address register. Same as 'r' unless generating MIPS16 code.
3426 case 'y': // Same as 'r'. Exists for compatibility.
3428 if (VT == MVT::i32 || VT == MVT::i16 || VT == MVT::i8) {
3429 if (Subtarget.inMips16Mode())
3430 return std::make_pair(0U, &Mips::CPU16RegsRegClass);
3431 return std::make_pair(0U, &Mips::GPR32RegClass);
3433 if (VT == MVT::i64 && !Subtarget.isGP64bit())
3434 return std::make_pair(0U, &Mips::GPR32RegClass);
3435 if (VT == MVT::i64 && Subtarget.isGP64bit())
3436 return std::make_pair(0U, &Mips::GPR64RegClass);
3437 // This will generate an error message
3438 return std::make_pair(0U, nullptr);
3439 case 'f': // FPU or MSA register
3440 if (VT == MVT::v16i8)
3441 return std::make_pair(0U, &Mips::MSA128BRegClass);
3442 else if (VT == MVT::v8i16 || VT == MVT::v8f16)
3443 return std::make_pair(0U, &Mips::MSA128HRegClass);
3444 else if (VT == MVT::v4i32 || VT == MVT::v4f32)
3445 return std::make_pair(0U, &Mips::MSA128WRegClass);
3446 else if (VT == MVT::v2i64 || VT == MVT::v2f64)
3447 return std::make_pair(0U, &Mips::MSA128DRegClass);
3448 else if (VT == MVT::f32)
3449 return std::make_pair(0U, &Mips::FGR32RegClass);
3450 else if ((VT == MVT::f64) && (!Subtarget.isSingleFloat())) {
3451 if (Subtarget.isFP64bit())
3452 return std::make_pair(0U, &Mips::FGR64RegClass);
3453 return std::make_pair(0U, &Mips::AFGR64RegClass);
3456 case 'c': // register suitable for indirect jump
3458 return std::make_pair((unsigned)Mips::T9, &Mips::GPR32RegClass);
3459 assert(VT == MVT::i64 && "Unexpected type.");
3460 return std::make_pair((unsigned)Mips::T9_64, &Mips::GPR64RegClass);
3461 case 'l': // register suitable for indirect jump
3463 return std::make_pair((unsigned)Mips::LO0, &Mips::LO32RegClass);
3464 return std::make_pair((unsigned)Mips::LO0_64, &Mips::LO64RegClass);
3465 case 'x': // register suitable for indirect jump
3466 // Fixme: Not triggering the use of both hi and low
3467 // This will generate an error message
3468 return std::make_pair(0U, nullptr);
3472 std::pair<unsigned, const TargetRegisterClass *> R;
3473 R = parseRegForInlineAsmConstraint(Constraint, VT);
3478 return TargetLowering::getRegForInlineAsmConstraint(TRI, Constraint, VT);
3481 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
3482 /// vector. If it is invalid, don't add anything to Ops.
3483 void MipsTargetLowering::LowerAsmOperandForConstraint(SDValue Op,
3484 std::string &Constraint,
3485 std::vector<SDValue>&Ops,
3486 SelectionDAG &DAG) const {
3490 // Only support length 1 constraints for now.
3491 if (Constraint.length() > 1) return;
3493 char ConstraintLetter = Constraint[0];
3494 switch (ConstraintLetter) {
3495 default: break; // This will fall through to the generic implementation
3496 case 'I': // Signed 16 bit constant
3497 // If this fails, the parent routine will give an error
3498 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
3499 EVT Type = Op.getValueType();
3500 int64_t Val = C->getSExtValue();
3501 if (isInt<16>(Val)) {
3502 Result = DAG.getTargetConstant(Val, DL, Type);
3507 case 'J': // integer zero
3508 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
3509 EVT Type = Op.getValueType();
3510 int64_t Val = C->getZExtValue();
3512 Result = DAG.getTargetConstant(0, DL, Type);
3517 case 'K': // unsigned 16 bit immediate
3518 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
3519 EVT Type = Op.getValueType();
3520 uint64_t Val = (uint64_t)C->getZExtValue();
3521 if (isUInt<16>(Val)) {
3522 Result = DAG.getTargetConstant(Val, DL, Type);
3527 case 'L': // signed 32 bit immediate where lower 16 bits are 0
3528 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
3529 EVT Type = Op.getValueType();
3530 int64_t Val = C->getSExtValue();
3531 if ((isInt<32>(Val)) && ((Val & 0xffff) == 0)){
3532 Result = DAG.getTargetConstant(Val, DL, Type);
3537 case 'N': // immediate in the range of -65535 to -1 (inclusive)
3538 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
3539 EVT Type = Op.getValueType();
3540 int64_t Val = C->getSExtValue();
3541 if ((Val >= -65535) && (Val <= -1)) {
3542 Result = DAG.getTargetConstant(Val, DL, Type);
3547 case 'O': // signed 15 bit immediate
3548 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
3549 EVT Type = Op.getValueType();
3550 int64_t Val = C->getSExtValue();
3551 if ((isInt<15>(Val))) {
3552 Result = DAG.getTargetConstant(Val, DL, Type);
3557 case 'P': // immediate in the range of 1 to 65535 (inclusive)
3558 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
3559 EVT Type = Op.getValueType();
3560 int64_t Val = C->getSExtValue();
3561 if ((Val <= 65535) && (Val >= 1)) {
3562 Result = DAG.getTargetConstant(Val, DL, Type);
3569 if (Result.getNode()) {
3570 Ops.push_back(Result);
3574 TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
3577 bool MipsTargetLowering::isLegalAddressingMode(const DataLayout &DL,
3578 const AddrMode &AM, Type *Ty,
3579 unsigned AS) const {
3580 // No global is ever allowed as a base.
3585 case 0: // "r+i" or just "i", depending on HasBaseReg.
3588 if (!AM.HasBaseReg) // allow "r+i".
3590 return false; // disallow "r+r" or "r+r+i".
3599 MipsTargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
3600 // The Mips target isn't yet aware of offsets.
3604 EVT MipsTargetLowering::getOptimalMemOpType(uint64_t Size, unsigned DstAlign,
3606 bool IsMemset, bool ZeroMemset,
3608 MachineFunction &MF) const {
3609 if (Subtarget.hasMips64())
3615 bool MipsTargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
3616 if (VT != MVT::f32 && VT != MVT::f64)
3618 if (Imm.isNegZero())
3620 return Imm.isZero();
3623 unsigned MipsTargetLowering::getJumpTableEncoding() const {
3625 return MachineJumpTableInfo::EK_GPRel64BlockAddress;
3627 return TargetLowering::getJumpTableEncoding();
3630 bool MipsTargetLowering::useSoftFloat() const {
3631 return Subtarget.useSoftFloat();
3634 void MipsTargetLowering::copyByValRegs(
3635 SDValue Chain, SDLoc DL, std::vector<SDValue> &OutChains, SelectionDAG &DAG,
3636 const ISD::ArgFlagsTy &Flags, SmallVectorImpl<SDValue> &InVals,
3637 const Argument *FuncArg, unsigned FirstReg, unsigned LastReg,
3638 const CCValAssign &VA, MipsCCState &State) const {
3639 MachineFunction &MF = DAG.getMachineFunction();
3640 MachineFrameInfo *MFI = MF.getFrameInfo();
3641 unsigned GPRSizeInBytes = Subtarget.getGPRSizeInBytes();
3642 unsigned NumRegs = LastReg - FirstReg;
3643 unsigned RegAreaSize = NumRegs * GPRSizeInBytes;
3644 unsigned FrameObjSize = std::max(Flags.getByValSize(), RegAreaSize);
3646 ArrayRef<MCPhysReg> ByValArgRegs = ABI.GetByValArgRegs();
3650 (int)ABI.GetCalleeAllocdArgSizeInBytes(State.getCallingConv()) -
3651 (int)((ByValArgRegs.size() - FirstReg) * GPRSizeInBytes);
3653 FrameObjOffset = VA.getLocMemOffset();
3655 // Create frame object.
3656 EVT PtrTy = getPointerTy(DAG.getDataLayout());
3657 int FI = MFI->CreateFixedObject(FrameObjSize, FrameObjOffset, true);
3658 SDValue FIN = DAG.getFrameIndex(FI, PtrTy);
3659 InVals.push_back(FIN);
3664 // Copy arg registers.
3665 MVT RegTy = MVT::getIntegerVT(GPRSizeInBytes * 8);
3666 const TargetRegisterClass *RC = getRegClassFor(RegTy);
3668 for (unsigned I = 0; I < NumRegs; ++I) {
3669 unsigned ArgReg = ByValArgRegs[FirstReg + I];
3670 unsigned VReg = addLiveIn(MF, ArgReg, RC);
3671 unsigned Offset = I * GPRSizeInBytes;
3672 SDValue StorePtr = DAG.getNode(ISD::ADD, DL, PtrTy, FIN,
3673 DAG.getConstant(Offset, DL, PtrTy));
3674 SDValue Store = DAG.getStore(Chain, DL, DAG.getRegister(VReg, RegTy),
3675 StorePtr, MachinePointerInfo(FuncArg, Offset),
3677 OutChains.push_back(Store);
3681 // Copy byVal arg to registers and stack.
3682 void MipsTargetLowering::passByValArg(
3683 SDValue Chain, SDLoc DL,
3684 std::deque<std::pair<unsigned, SDValue>> &RegsToPass,
3685 SmallVectorImpl<SDValue> &MemOpChains, SDValue StackPtr,
3686 MachineFrameInfo *MFI, SelectionDAG &DAG, SDValue Arg, unsigned FirstReg,
3687 unsigned LastReg, const ISD::ArgFlagsTy &Flags, bool isLittle,
3688 const CCValAssign &VA) const {
3689 unsigned ByValSizeInBytes = Flags.getByValSize();
3690 unsigned OffsetInBytes = 0; // From beginning of struct
3691 unsigned RegSizeInBytes = Subtarget.getGPRSizeInBytes();
3692 unsigned Alignment = std::min(Flags.getByValAlign(), RegSizeInBytes);
3693 EVT PtrTy = getPointerTy(DAG.getDataLayout()),
3694 RegTy = MVT::getIntegerVT(RegSizeInBytes * 8);
3695 unsigned NumRegs = LastReg - FirstReg;
3698 ArrayRef<MCPhysReg> ArgRegs = ABI.GetByValArgRegs();
3699 bool LeftoverBytes = (NumRegs * RegSizeInBytes > ByValSizeInBytes);
3702 // Copy words to registers.
3703 for (; I < NumRegs - LeftoverBytes; ++I, OffsetInBytes += RegSizeInBytes) {
3704 SDValue LoadPtr = DAG.getNode(ISD::ADD, DL, PtrTy, Arg,
3705 DAG.getConstant(OffsetInBytes, DL, PtrTy));
3706 SDValue LoadVal = DAG.getLoad(RegTy, DL, Chain, LoadPtr,
3707 MachinePointerInfo(), false, false, false,
3709 MemOpChains.push_back(LoadVal.getValue(1));
3710 unsigned ArgReg = ArgRegs[FirstReg + I];
3711 RegsToPass.push_back(std::make_pair(ArgReg, LoadVal));
3714 // Return if the struct has been fully copied.
3715 if (ByValSizeInBytes == OffsetInBytes)
3718 // Copy the remainder of the byval argument with sub-word loads and shifts.
3719 if (LeftoverBytes) {
3722 for (unsigned LoadSizeInBytes = RegSizeInBytes / 2, TotalBytesLoaded = 0;
3723 OffsetInBytes < ByValSizeInBytes; LoadSizeInBytes /= 2) {
3724 unsigned RemainingSizeInBytes = ByValSizeInBytes - OffsetInBytes;
3726 if (RemainingSizeInBytes < LoadSizeInBytes)
3730 SDValue LoadPtr = DAG.getNode(ISD::ADD, DL, PtrTy, Arg,
3731 DAG.getConstant(OffsetInBytes, DL,
3733 SDValue LoadVal = DAG.getExtLoad(
3734 ISD::ZEXTLOAD, DL, RegTy, Chain, LoadPtr, MachinePointerInfo(),
3735 MVT::getIntegerVT(LoadSizeInBytes * 8), false, false, false,
3737 MemOpChains.push_back(LoadVal.getValue(1));
3739 // Shift the loaded value.
3743 Shamt = TotalBytesLoaded * 8;
3745 Shamt = (RegSizeInBytes - (TotalBytesLoaded + LoadSizeInBytes)) * 8;
3747 SDValue Shift = DAG.getNode(ISD::SHL, DL, RegTy, LoadVal,
3748 DAG.getConstant(Shamt, DL, MVT::i32));
3751 Val = DAG.getNode(ISD::OR, DL, RegTy, Val, Shift);
3755 OffsetInBytes += LoadSizeInBytes;
3756 TotalBytesLoaded += LoadSizeInBytes;
3757 Alignment = std::min(Alignment, LoadSizeInBytes);
3760 unsigned ArgReg = ArgRegs[FirstReg + I];
3761 RegsToPass.push_back(std::make_pair(ArgReg, Val));
3766 // Copy remainder of byval arg to it with memcpy.
3767 unsigned MemCpySize = ByValSizeInBytes - OffsetInBytes;
3768 SDValue Src = DAG.getNode(ISD::ADD, DL, PtrTy, Arg,
3769 DAG.getConstant(OffsetInBytes, DL, PtrTy));
3770 SDValue Dst = DAG.getNode(ISD::ADD, DL, PtrTy, StackPtr,
3771 DAG.getIntPtrConstant(VA.getLocMemOffset(), DL));
3772 Chain = DAG.getMemcpy(Chain, DL, Dst, Src,
3773 DAG.getConstant(MemCpySize, DL, PtrTy),
3774 Alignment, /*isVolatile=*/false, /*AlwaysInline=*/false,
3775 /*isTailCall=*/false,
3776 MachinePointerInfo(), MachinePointerInfo());
3777 MemOpChains.push_back(Chain);
3780 void MipsTargetLowering::writeVarArgRegs(std::vector<SDValue> &OutChains,
3781 SDValue Chain, SDLoc DL,
3783 CCState &State) const {
3784 ArrayRef<MCPhysReg> ArgRegs = ABI.GetVarArgRegs();
3785 unsigned Idx = State.getFirstUnallocated(ArgRegs);
3786 unsigned RegSizeInBytes = Subtarget.getGPRSizeInBytes();
3787 MVT RegTy = MVT::getIntegerVT(RegSizeInBytes * 8);
3788 const TargetRegisterClass *RC = getRegClassFor(RegTy);
3789 MachineFunction &MF = DAG.getMachineFunction();
3790 MachineFrameInfo *MFI = MF.getFrameInfo();
3791 MipsFunctionInfo *MipsFI = MF.getInfo<MipsFunctionInfo>();
3793 // Offset of the first variable argument from stack pointer.
3796 if (ArgRegs.size() == Idx)
3798 RoundUpToAlignment(State.getNextStackOffset(), RegSizeInBytes);
3801 (int)ABI.GetCalleeAllocdArgSizeInBytes(State.getCallingConv()) -
3802 (int)(RegSizeInBytes * (ArgRegs.size() - Idx));
3805 // Record the frame index of the first variable argument
3806 // which is a value necessary to VASTART.
3807 int FI = MFI->CreateFixedObject(RegSizeInBytes, VaArgOffset, true);
3808 MipsFI->setVarArgsFrameIndex(FI);
3810 // Copy the integer registers that have not been used for argument passing
3811 // to the argument register save area. For O32, the save area is allocated
3812 // in the caller's stack frame, while for N32/64, it is allocated in the
3813 // callee's stack frame.
3814 for (unsigned I = Idx; I < ArgRegs.size();
3815 ++I, VaArgOffset += RegSizeInBytes) {
3816 unsigned Reg = addLiveIn(MF, ArgRegs[I], RC);
3817 SDValue ArgValue = DAG.getCopyFromReg(Chain, DL, Reg, RegTy);
3818 FI = MFI->CreateFixedObject(RegSizeInBytes, VaArgOffset, true);
3819 SDValue PtrOff = DAG.getFrameIndex(FI, getPointerTy(DAG.getDataLayout()));
3820 SDValue Store = DAG.getStore(Chain, DL, ArgValue, PtrOff,
3821 MachinePointerInfo(), false, false, 0);
3822 cast<StoreSDNode>(Store.getNode())->getMemOperand()->setValue(
3824 OutChains.push_back(Store);
3828 void MipsTargetLowering::HandleByVal(CCState *State, unsigned &Size,
3829 unsigned Align) const {
3830 const TargetFrameLowering *TFL = Subtarget.getFrameLowering();
3832 assert(Size && "Byval argument's size shouldn't be 0.");
3834 Align = std::min(Align, TFL->getStackAlignment());
3836 unsigned FirstReg = 0;
3837 unsigned NumRegs = 0;
3839 if (State->getCallingConv() != CallingConv::Fast) {
3840 unsigned RegSizeInBytes = Subtarget.getGPRSizeInBytes();
3841 ArrayRef<MCPhysReg> IntArgRegs = ABI.GetByValArgRegs();
3842 // FIXME: The O32 case actually describes no shadow registers.
3843 const MCPhysReg *ShadowRegs =
3844 ABI.IsO32() ? IntArgRegs.data() : Mips64DPRegs;
3846 // We used to check the size as well but we can't do that anymore since
3847 // CCState::HandleByVal() rounds up the size after calling this function.
3848 assert(!(Align % RegSizeInBytes) &&
3849 "Byval argument's alignment should be a multiple of"
3852 FirstReg = State->getFirstUnallocated(IntArgRegs);
3854 // If Align > RegSizeInBytes, the first arg register must be even.
3855 // FIXME: This condition happens to do the right thing but it's not the
3856 // right way to test it. We want to check that the stack frame offset
3857 // of the register is aligned.
3858 if ((Align > RegSizeInBytes) && (FirstReg % 2)) {
3859 State->AllocateReg(IntArgRegs[FirstReg], ShadowRegs[FirstReg]);
3863 // Mark the registers allocated.
3864 Size = RoundUpToAlignment(Size, RegSizeInBytes);
3865 for (unsigned I = FirstReg; Size > 0 && (I < IntArgRegs.size());
3866 Size -= RegSizeInBytes, ++I, ++NumRegs)
3867 State->AllocateReg(IntArgRegs[I], ShadowRegs[I]);
3870 State->addInRegsParamInfo(FirstReg, FirstReg + NumRegs);
3874 MipsTargetLowering::emitPseudoSELECT(MachineInstr *MI, MachineBasicBlock *BB,
3875 bool isFPCmp, unsigned Opc) const {
3876 assert(!(Subtarget.hasMips4() || Subtarget.hasMips32()) &&
3877 "Subtarget already supports SELECT nodes with the use of"
3878 "conditional-move instructions.");
3880 const TargetInstrInfo *TII =
3881 Subtarget.getInstrInfo();
3882 DebugLoc DL = MI->getDebugLoc();
3884 // To "insert" a SELECT instruction, we actually have to insert the
3885 // diamond control-flow pattern. The incoming instruction knows the
3886 // destination vreg to set, the condition code register to branch on, the
3887 // true/false values to select between, and a branch opcode to use.
3888 const BasicBlock *LLVM_BB = BB->getBasicBlock();
3889 MachineFunction::iterator It = ++BB->getIterator();
3895 // bNE r1, r0, copy1MBB
3896 // fallthrough --> copy0MBB
3897 MachineBasicBlock *thisMBB = BB;
3898 MachineFunction *F = BB->getParent();
3899 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
3900 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
3901 F->insert(It, copy0MBB);
3902 F->insert(It, sinkMBB);
3904 // Transfer the remainder of BB and its successor edges to sinkMBB.
3905 sinkMBB->splice(sinkMBB->begin(), BB,
3906 std::next(MachineBasicBlock::iterator(MI)), BB->end());
3907 sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
3909 // Next, add the true and fallthrough blocks as its successors.
3910 BB->addSuccessor(copy0MBB);
3911 BB->addSuccessor(sinkMBB);
3914 // bc1[tf] cc, sinkMBB
3915 BuildMI(BB, DL, TII->get(Opc))
3916 .addReg(MI->getOperand(1).getReg())
3919 // bne rs, $0, sinkMBB
3920 BuildMI(BB, DL, TII->get(Opc))
3921 .addReg(MI->getOperand(1).getReg())
3927 // %FalseValue = ...
3928 // # fallthrough to sinkMBB
3931 // Update machine-CFG edges
3932 BB->addSuccessor(sinkMBB);
3935 // %Result = phi [ %TrueValue, thisMBB ], [ %FalseValue, copy0MBB ]
3939 BuildMI(*BB, BB->begin(), DL,
3940 TII->get(Mips::PHI), MI->getOperand(0).getReg())
3941 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB)
3942 .addReg(MI->getOperand(3).getReg()).addMBB(copy0MBB);
3944 MI->eraseFromParent(); // The pseudo instruction is gone now.
3949 // FIXME? Maybe this could be a TableGen attribute on some registers and
3950 // this table could be generated automatically from RegInfo.
3951 unsigned MipsTargetLowering::getRegisterByName(const char* RegName, EVT VT,
3952 SelectionDAG &DAG) const {
3953 // Named registers is expected to be fairly rare. For now, just support $28
3954 // since the linux kernel uses it.
3955 if (Subtarget.isGP64bit()) {
3956 unsigned Reg = StringSwitch<unsigned>(RegName)
3957 .Case("$28", Mips::GP_64)
3962 unsigned Reg = StringSwitch<unsigned>(RegName)
3963 .Case("$28", Mips::GP)
3968 report_fatal_error("Invalid register name global variable");