1 | //===-- ARMISelDAGToDAG.cpp - A dag to dag inst selector for ARM ----------===// |
2 | // |
3 | // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. |
4 | // See https://llvm.org/LICENSE.txt for license information. |
5 | // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception |
6 | // |
7 | //===----------------------------------------------------------------------===// |
8 | // |
9 | // This file defines an instruction selector for the ARM target. |
10 | // |
11 | //===----------------------------------------------------------------------===// |
12 | |
13 | #include "ARM.h" |
14 | #include "ARMBaseInstrInfo.h" |
15 | #include "ARMTargetMachine.h" |
16 | #include "MCTargetDesc/ARMAddressingModes.h" |
17 | #include "Utils/ARMBaseInfo.h" |
18 | #include "llvm/ADT/APSInt.h" |
19 | #include "llvm/ADT/StringSwitch.h" |
20 | #include "llvm/CodeGen/MachineFrameInfo.h" |
21 | #include "llvm/CodeGen/MachineFunction.h" |
22 | #include "llvm/CodeGen/MachineInstrBuilder.h" |
23 | #include "llvm/CodeGen/MachineRegisterInfo.h" |
24 | #include "llvm/CodeGen/SelectionDAG.h" |
25 | #include "llvm/CodeGen/SelectionDAGISel.h" |
26 | #include "llvm/CodeGen/TargetLowering.h" |
27 | #include "llvm/IR/Constants.h" |
28 | #include "llvm/IR/DerivedTypes.h" |
29 | #include "llvm/IR/Function.h" |
30 | #include "llvm/IR/Intrinsics.h" |
31 | #include "llvm/IR/IntrinsicsARM.h" |
32 | #include "llvm/IR/LLVMContext.h" |
33 | #include "llvm/Support/CommandLine.h" |
34 | #include "llvm/Support/ErrorHandling.h" |
35 | #include "llvm/Target/TargetOptions.h" |
36 | #include <optional> |
37 | |
38 | using namespace llvm; |
39 | |
40 | #define DEBUG_TYPE "arm-isel" |
41 | #define PASS_NAME "ARM Instruction Selection" |
42 | |
43 | static cl::opt<bool> |
44 | DisableShifterOp("disable-shifter-op" , cl::Hidden, |
45 | cl::desc("Disable isel of shifter-op" ), |
46 | cl::init(Val: false)); |
47 | |
48 | //===--------------------------------------------------------------------===// |
49 | /// ARMDAGToDAGISel - ARM specific code to select ARM machine |
50 | /// instructions for SelectionDAG operations. |
51 | /// |
52 | namespace { |
53 | |
54 | class ARMDAGToDAGISel : public SelectionDAGISel { |
55 | /// Subtarget - Keep a pointer to the ARMSubtarget around so that we can |
56 | /// make the right decision when generating code for different targets. |
57 | const ARMSubtarget *Subtarget; |
58 | |
59 | public: |
60 | ARMDAGToDAGISel() = delete; |
61 | |
62 | explicit ARMDAGToDAGISel(ARMBaseTargetMachine &tm, CodeGenOptLevel OptLevel) |
63 | : SelectionDAGISel(tm, OptLevel) {} |
64 | |
65 | bool runOnMachineFunction(MachineFunction &MF) override { |
66 | // Reset the subtarget each time through. |
67 | Subtarget = &MF.getSubtarget<ARMSubtarget>(); |
68 | SelectionDAGISel::runOnMachineFunction(mf&: MF); |
69 | return true; |
70 | } |
71 | |
72 | void PreprocessISelDAG() override; |
73 | |
74 | /// getI32Imm - Return a target constant of type i32 with the specified |
75 | /// value. |
76 | inline SDValue getI32Imm(unsigned Imm, const SDLoc &dl) { |
77 | return CurDAG->getTargetConstant(Val: Imm, DL: dl, VT: MVT::i32); |
78 | } |
79 | |
80 | void Select(SDNode *N) override; |
81 | |
82 | /// Return true as some complex patterns, like those that call |
83 | /// canExtractShiftFromMul can modify the DAG inplace. |
84 | bool ComplexPatternFuncMutatesDAG() const override { return true; } |
85 | |
86 | bool hasNoVMLxHazardUse(SDNode *N) const; |
87 | bool isShifterOpProfitable(const SDValue &Shift, |
88 | ARM_AM::ShiftOpc ShOpcVal, unsigned ShAmt); |
89 | bool SelectRegShifterOperand(SDValue N, SDValue &A, |
90 | SDValue &B, SDValue &C, |
91 | bool CheckProfitability = true); |
92 | bool SelectImmShifterOperand(SDValue N, SDValue &A, |
93 | SDValue &B, bool CheckProfitability = true); |
94 | bool SelectShiftRegShifterOperand(SDValue N, SDValue &A, SDValue &B, |
95 | SDValue &C) { |
96 | // Don't apply the profitability check |
97 | return SelectRegShifterOperand(N, A, B, C, CheckProfitability: false); |
98 | } |
99 | bool SelectShiftImmShifterOperand(SDValue N, SDValue &A, SDValue &B) { |
100 | // Don't apply the profitability check |
101 | return SelectImmShifterOperand(N, A, B, CheckProfitability: false); |
102 | } |
103 | bool SelectShiftImmShifterOperandOneUse(SDValue N, SDValue &A, SDValue &B) { |
104 | if (!N.hasOneUse()) |
105 | return false; |
106 | return SelectImmShifterOperand(N, A, B, CheckProfitability: false); |
107 | } |
108 | |
109 | bool SelectAddLikeOr(SDNode *Parent, SDValue N, SDValue &Out); |
110 | |
111 | bool SelectAddrModeImm12(SDValue N, SDValue &Base, SDValue &OffImm); |
112 | bool SelectLdStSOReg(SDValue N, SDValue &Base, SDValue &Offset, SDValue &Opc); |
113 | |
114 | bool SelectAddrMode2OffsetReg(SDNode *Op, SDValue N, |
115 | SDValue &Offset, SDValue &Opc); |
116 | bool SelectAddrMode2OffsetImm(SDNode *Op, SDValue N, |
117 | SDValue &Offset, SDValue &Opc); |
118 | bool SelectAddrMode2OffsetImmPre(SDNode *Op, SDValue N, |
119 | SDValue &Offset, SDValue &Opc); |
120 | bool SelectAddrOffsetNone(SDValue N, SDValue &Base); |
121 | bool SelectAddrMode3(SDValue N, SDValue &Base, |
122 | SDValue &Offset, SDValue &Opc); |
123 | bool SelectAddrMode3Offset(SDNode *Op, SDValue N, |
124 | SDValue &Offset, SDValue &Opc); |
125 | bool IsAddressingMode5(SDValue N, SDValue &Base, SDValue &Offset, bool FP16); |
126 | bool SelectAddrMode5(SDValue N, SDValue &Base, SDValue &Offset); |
127 | bool SelectAddrMode5FP16(SDValue N, SDValue &Base, SDValue &Offset); |
128 | bool SelectAddrMode6(SDNode *Parent, SDValue N, SDValue &Addr,SDValue &Align); |
129 | bool SelectAddrMode6Offset(SDNode *Op, SDValue N, SDValue &Offset); |
130 | |
131 | bool SelectAddrModePC(SDValue N, SDValue &Offset, SDValue &Label); |
132 | |
133 | // Thumb Addressing Modes: |
134 | bool SelectThumbAddrModeRR(SDValue N, SDValue &Base, SDValue &Offset); |
135 | bool SelectThumbAddrModeRRSext(SDValue N, SDValue &Base, SDValue &Offset); |
136 | bool SelectThumbAddrModeImm5S(SDValue N, unsigned Scale, SDValue &Base, |
137 | SDValue &OffImm); |
138 | bool SelectThumbAddrModeImm5S1(SDValue N, SDValue &Base, |
139 | SDValue &OffImm); |
140 | bool SelectThumbAddrModeImm5S2(SDValue N, SDValue &Base, |
141 | SDValue &OffImm); |
142 | bool SelectThumbAddrModeImm5S4(SDValue N, SDValue &Base, |
143 | SDValue &OffImm); |
144 | bool SelectThumbAddrModeSP(SDValue N, SDValue &Base, SDValue &OffImm); |
145 | template <unsigned Shift> |
146 | bool SelectTAddrModeImm7(SDValue N, SDValue &Base, SDValue &OffImm); |
147 | |
148 | // Thumb 2 Addressing Modes: |
149 | bool SelectT2AddrModeImm12(SDValue N, SDValue &Base, SDValue &OffImm); |
150 | template <unsigned Shift> |
151 | bool SelectT2AddrModeImm8(SDValue N, SDValue &Base, SDValue &OffImm); |
152 | bool SelectT2AddrModeImm8(SDValue N, SDValue &Base, |
153 | SDValue &OffImm); |
154 | bool SelectT2AddrModeImm8Offset(SDNode *Op, SDValue N, |
155 | SDValue &OffImm); |
156 | template <unsigned Shift> |
157 | bool SelectT2AddrModeImm7Offset(SDNode *Op, SDValue N, SDValue &OffImm); |
158 | bool SelectT2AddrModeImm7Offset(SDNode *Op, SDValue N, SDValue &OffImm, |
159 | unsigned Shift); |
160 | template <unsigned Shift> |
161 | bool SelectT2AddrModeImm7(SDValue N, SDValue &Base, SDValue &OffImm); |
162 | bool SelectT2AddrModeSoReg(SDValue N, SDValue &Base, |
163 | SDValue &OffReg, SDValue &ShImm); |
164 | bool SelectT2AddrModeExclusive(SDValue N, SDValue &Base, SDValue &OffImm); |
165 | |
166 | template<int Min, int Max> |
167 | bool SelectImmediateInRange(SDValue N, SDValue &OffImm); |
168 | |
169 | inline bool is_so_imm(unsigned Imm) const { |
170 | return ARM_AM::getSOImmVal(Arg: Imm) != -1; |
171 | } |
172 | |
173 | inline bool is_so_imm_not(unsigned Imm) const { |
174 | return ARM_AM::getSOImmVal(Arg: ~Imm) != -1; |
175 | } |
176 | |
177 | inline bool is_t2_so_imm(unsigned Imm) const { |
178 | return ARM_AM::getT2SOImmVal(Arg: Imm) != -1; |
179 | } |
180 | |
181 | inline bool is_t2_so_imm_not(unsigned Imm) const { |
182 | return ARM_AM::getT2SOImmVal(Arg: ~Imm) != -1; |
183 | } |
184 | |
185 | // Include the pieces autogenerated from the target description. |
186 | #include "ARMGenDAGISel.inc" |
187 | |
188 | private: |
189 | void transferMemOperands(SDNode *Src, SDNode *Dst); |
190 | |
191 | /// Indexed (pre/post inc/dec) load matching code for ARM. |
192 | bool tryARMIndexedLoad(SDNode *N); |
193 | bool tryT1IndexedLoad(SDNode *N); |
194 | bool tryT2IndexedLoad(SDNode *N); |
195 | bool tryMVEIndexedLoad(SDNode *N); |
196 | bool tryFMULFixed(SDNode *N, SDLoc dl); |
197 | bool tryFP_TO_INT(SDNode *N, SDLoc dl); |
198 | bool transformFixedFloatingPointConversion(SDNode *N, SDNode *FMul, |
199 | bool IsUnsigned, |
200 | bool FixedToFloat); |
201 | |
202 | /// SelectVLD - Select NEON load intrinsics. NumVecs should be |
203 | /// 1, 2, 3 or 4. The opcode arrays specify the instructions used for |
204 | /// loads of D registers and even subregs and odd subregs of Q registers. |
205 | /// For NumVecs <= 2, QOpcodes1 is not used. |
206 | void SelectVLD(SDNode *N, bool isUpdating, unsigned NumVecs, |
207 | const uint16_t *DOpcodes, const uint16_t *QOpcodes0, |
208 | const uint16_t *QOpcodes1); |
209 | |
210 | /// SelectVST - Select NEON store intrinsics. NumVecs should |
211 | /// be 1, 2, 3 or 4. The opcode arrays specify the instructions used for |
212 | /// stores of D registers and even subregs and odd subregs of Q registers. |
213 | /// For NumVecs <= 2, QOpcodes1 is not used. |
214 | void SelectVST(SDNode *N, bool isUpdating, unsigned NumVecs, |
215 | const uint16_t *DOpcodes, const uint16_t *QOpcodes0, |
216 | const uint16_t *QOpcodes1); |
217 | |
218 | /// SelectVLDSTLane - Select NEON load/store lane intrinsics. NumVecs should |
219 | /// be 2, 3 or 4. The opcode arrays specify the instructions used for |
220 | /// load/store of D registers and Q registers. |
221 | void SelectVLDSTLane(SDNode *N, bool IsLoad, bool isUpdating, |
222 | unsigned NumVecs, const uint16_t *DOpcodes, |
223 | const uint16_t *QOpcodes); |
224 | |
225 | /// Helper functions for setting up clusters of MVE predication operands. |
226 | template <typename SDValueVector> |
227 | void AddMVEPredicateToOps(SDValueVector &Ops, SDLoc Loc, |
228 | SDValue PredicateMask); |
229 | template <typename SDValueVector> |
230 | void AddMVEPredicateToOps(SDValueVector &Ops, SDLoc Loc, |
231 | SDValue PredicateMask, SDValue Inactive); |
232 | |
233 | template <typename SDValueVector> |
234 | void AddEmptyMVEPredicateToOps(SDValueVector &Ops, SDLoc Loc); |
235 | template <typename SDValueVector> |
236 | void AddEmptyMVEPredicateToOps(SDValueVector &Ops, SDLoc Loc, EVT InactiveTy); |
237 | |
238 | /// SelectMVE_WB - Select MVE writeback load/store intrinsics. |
239 | void SelectMVE_WB(SDNode *N, const uint16_t *Opcodes, bool Predicated); |
240 | |
241 | /// SelectMVE_LongShift - Select MVE 64-bit scalar shift intrinsics. |
242 | void SelectMVE_LongShift(SDNode *N, uint16_t Opcode, bool Immediate, |
243 | bool HasSaturationOperand); |
244 | |
245 | /// SelectMVE_VADCSBC - Select MVE vector add/sub-with-carry intrinsics. |
246 | void SelectMVE_VADCSBC(SDNode *N, uint16_t OpcodeWithCarry, |
247 | uint16_t OpcodeWithNoCarry, bool Add, bool Predicated); |
248 | |
249 | /// SelectMVE_VSHLC - Select MVE intrinsics for a shift that carries between |
250 | /// vector lanes. |
251 | void SelectMVE_VSHLC(SDNode *N, bool Predicated); |
252 | |
253 | /// Select long MVE vector reductions with two vector operands |
254 | /// Stride is the number of vector element widths the instruction can operate |
255 | /// on: |
256 | /// 2 for long non-rounding variants, vml{a,s}ldav[a][x]: [i16, i32] |
257 | /// 1 for long rounding variants: vrml{a,s}ldavh[a][x]: [i32] |
258 | /// Stride is used when addressing the OpcodesS array which contains multiple |
259 | /// opcodes for each element width. |
260 | /// TySize is the index into the list of element types listed above |
261 | void SelectBaseMVE_VMLLDAV(SDNode *N, bool Predicated, |
262 | const uint16_t *OpcodesS, const uint16_t *OpcodesU, |
263 | size_t Stride, size_t TySize); |
264 | |
265 | /// Select a 64-bit MVE vector reduction with two vector operands |
266 | /// arm_mve_vmlldava_[predicated] |
267 | void SelectMVE_VMLLDAV(SDNode *N, bool Predicated, const uint16_t *OpcodesS, |
268 | const uint16_t *OpcodesU); |
269 | /// Select a 72-bit MVE vector rounding reduction with two vector operands |
270 | /// int_arm_mve_vrmlldavha[_predicated] |
271 | void SelectMVE_VRMLLDAVH(SDNode *N, bool Predicated, const uint16_t *OpcodesS, |
272 | const uint16_t *OpcodesU); |
273 | |
274 | /// SelectMVE_VLD - Select MVE interleaving load intrinsics. NumVecs |
275 | /// should be 2 or 4. The opcode array specifies the instructions |
276 | /// used for 8, 16 and 32-bit lane sizes respectively, and each |
277 | /// pointer points to a set of NumVecs sub-opcodes used for the |
278 | /// different stages (e.g. VLD20 versus VLD21) of each load family. |
279 | void SelectMVE_VLD(SDNode *N, unsigned NumVecs, |
280 | const uint16_t *const *Opcodes, bool HasWriteback); |
281 | |
282 | /// SelectMVE_VxDUP - Select MVE incrementing-dup instructions. Opcodes is an |
283 | /// array of 3 elements for the 8, 16 and 32-bit lane sizes. |
284 | void SelectMVE_VxDUP(SDNode *N, const uint16_t *Opcodes, |
285 | bool Wrapping, bool Predicated); |
286 | |
287 | /// Select SelectCDE_CXxD - Select CDE dual-GPR instruction (one of CX1D, |
288 | /// CX1DA, CX2D, CX2DA, CX3, CX3DA). |
289 | /// \arg \c NumExtraOps number of extra operands besides the coprocossor, |
290 | /// the accumulator and the immediate operand, i.e. 0 |
291 | /// for CX1*, 1 for CX2*, 2 for CX3* |
292 | /// \arg \c HasAccum whether the instruction has an accumulator operand |
293 | void SelectCDE_CXxD(SDNode *N, uint16_t Opcode, size_t , |
294 | bool HasAccum); |
295 | |
296 | /// SelectVLDDup - Select NEON load-duplicate intrinsics. NumVecs |
297 | /// should be 1, 2, 3 or 4. The opcode array specifies the instructions used |
298 | /// for loading D registers. |
299 | void SelectVLDDup(SDNode *N, bool IsIntrinsic, bool isUpdating, |
300 | unsigned NumVecs, const uint16_t *DOpcodes, |
301 | const uint16_t *QOpcodes0 = nullptr, |
302 | const uint16_t *QOpcodes1 = nullptr); |
303 | |
304 | /// Try to select SBFX/UBFX instructions for ARM. |
305 | bool tryV6T2BitfieldExtractOp(SDNode *N, bool isSigned); |
306 | |
307 | bool tryInsertVectorElt(SDNode *N); |
308 | |
309 | // Select special operations if node forms integer ABS pattern |
310 | bool tryABSOp(SDNode *N); |
311 | |
312 | bool tryReadRegister(SDNode *N); |
313 | bool tryWriteRegister(SDNode *N); |
314 | |
315 | bool tryInlineAsm(SDNode *N); |
316 | |
317 | void SelectCMPZ(SDNode *N, bool &SwitchEQNEToPLMI); |
318 | |
319 | void SelectCMP_SWAP(SDNode *N); |
320 | |
321 | /// SelectInlineAsmMemoryOperand - Implement addressing mode selection for |
322 | /// inline asm expressions. |
323 | bool SelectInlineAsmMemoryOperand(const SDValue &Op, |
324 | InlineAsm::ConstraintCode ConstraintID, |
325 | std::vector<SDValue> &OutOps) override; |
326 | |
327 | // Form pairs of consecutive R, S, D, or Q registers. |
328 | SDNode *createGPRPairNode(EVT VT, SDValue V0, SDValue V1); |
329 | SDNode *createSRegPairNode(EVT VT, SDValue V0, SDValue V1); |
330 | SDNode *createDRegPairNode(EVT VT, SDValue V0, SDValue V1); |
331 | SDNode *createQRegPairNode(EVT VT, SDValue V0, SDValue V1); |
332 | |
333 | // Form sequences of 4 consecutive S, D, or Q registers. |
334 | SDNode *createQuadSRegsNode(EVT VT, SDValue V0, SDValue V1, SDValue V2, SDValue V3); |
335 | SDNode *createQuadDRegsNode(EVT VT, SDValue V0, SDValue V1, SDValue V2, SDValue V3); |
336 | SDNode *createQuadQRegsNode(EVT VT, SDValue V0, SDValue V1, SDValue V2, SDValue V3); |
337 | |
338 | // Get the alignment operand for a NEON VLD or VST instruction. |
339 | SDValue GetVLDSTAlign(SDValue Align, const SDLoc &dl, unsigned NumVecs, |
340 | bool is64BitVector); |
341 | |
342 | /// Checks if N is a multiplication by a constant where we can extract out a |
343 | /// power of two from the constant so that it can be used in a shift, but only |
344 | /// if it simplifies the materialization of the constant. Returns true if it |
345 | /// is, and assigns to PowerOfTwo the power of two that should be extracted |
346 | /// out and to NewMulConst the new constant to be multiplied by. |
347 | bool canExtractShiftFromMul(const SDValue &N, unsigned MaxShift, |
348 | unsigned &PowerOfTwo, SDValue &NewMulConst) const; |
349 | |
350 | /// Replace N with M in CurDAG, in a way that also ensures that M gets |
351 | /// selected when N would have been selected. |
352 | void replaceDAGValue(const SDValue &N, SDValue M); |
353 | }; |
354 | |
355 | class ARMDAGToDAGISelLegacy : public SelectionDAGISelLegacy { |
356 | public: |
357 | static char ID; |
358 | ARMDAGToDAGISelLegacy(ARMBaseTargetMachine &tm, CodeGenOptLevel OptLevel) |
359 | : SelectionDAGISelLegacy( |
360 | ID, std::make_unique<ARMDAGToDAGISel>(args&: tm, args&: OptLevel)) {} |
361 | }; |
362 | } |
363 | |
364 | char ARMDAGToDAGISelLegacy::ID = 0; |
365 | |
366 | INITIALIZE_PASS(ARMDAGToDAGISelLegacy, DEBUG_TYPE, PASS_NAME, false, false) |
367 | |
368 | /// isInt32Immediate - This method tests to see if the node is a 32-bit constant |
369 | /// operand. If so Imm will receive the 32-bit value. |
370 | static bool isInt32Immediate(SDNode *N, unsigned &Imm) { |
371 | if (N->getOpcode() == ISD::Constant && N->getValueType(ResNo: 0) == MVT::i32) { |
372 | Imm = N->getAsZExtVal(); |
373 | return true; |
374 | } |
375 | return false; |
376 | } |
377 | |
378 | // isInt32Immediate - This method tests to see if a constant operand. |
379 | // If so Imm will receive the 32 bit value. |
380 | static bool isInt32Immediate(SDValue N, unsigned &Imm) { |
381 | return isInt32Immediate(N: N.getNode(), Imm); |
382 | } |
383 | |
384 | // isOpcWithIntImmediate - This method tests to see if the node is a specific |
385 | // opcode and that it has a immediate integer right operand. |
386 | // If so Imm will receive the 32 bit value. |
387 | static bool isOpcWithIntImmediate(SDNode *N, unsigned Opc, unsigned& Imm) { |
388 | return N->getOpcode() == Opc && |
389 | isInt32Immediate(N: N->getOperand(Num: 1).getNode(), Imm); |
390 | } |
391 | |
392 | /// Check whether a particular node is a constant value representable as |
393 | /// (N * Scale) where (N in [\p RangeMin, \p RangeMax). |
394 | /// |
395 | /// \param ScaledConstant [out] - On success, the pre-scaled constant value. |
396 | static bool isScaledConstantInRange(SDValue Node, int Scale, |
397 | int RangeMin, int RangeMax, |
398 | int &ScaledConstant) { |
399 | assert(Scale > 0 && "Invalid scale!" ); |
400 | |
401 | // Check that this is a constant. |
402 | const ConstantSDNode *C = dyn_cast<ConstantSDNode>(Val&: Node); |
403 | if (!C) |
404 | return false; |
405 | |
406 | ScaledConstant = (int) C->getZExtValue(); |
407 | if ((ScaledConstant % Scale) != 0) |
408 | return false; |
409 | |
410 | ScaledConstant /= Scale; |
411 | return ScaledConstant >= RangeMin && ScaledConstant < RangeMax; |
412 | } |
413 | |
414 | void ARMDAGToDAGISel::PreprocessISelDAG() { |
415 | if (!Subtarget->hasV6T2Ops()) |
416 | return; |
417 | |
418 | bool isThumb2 = Subtarget->isThumb(); |
419 | // We use make_early_inc_range to avoid invalidation issues. |
420 | for (SDNode &N : llvm::make_early_inc_range(Range: CurDAG->allnodes())) { |
421 | if (N.getOpcode() != ISD::ADD) |
422 | continue; |
423 | |
424 | // Look for (add X1, (and (srl X2, c1), c2)) where c2 is constant with |
425 | // leading zeros, followed by consecutive set bits, followed by 1 or 2 |
426 | // trailing zeros, e.g. 1020. |
427 | // Transform the expression to |
428 | // (add X1, (shl (and (srl X2, c1), (c2>>tz)), tz)) where tz is the number |
429 | // of trailing zeros of c2. The left shift would be folded as an shifter |
430 | // operand of 'add' and the 'and' and 'srl' would become a bits extraction |
431 | // node (UBFX). |
432 | |
433 | SDValue N0 = N.getOperand(Num: 0); |
434 | SDValue N1 = N.getOperand(Num: 1); |
435 | unsigned And_imm = 0; |
436 | if (!isOpcWithIntImmediate(N: N1.getNode(), Opc: ISD::AND, Imm&: And_imm)) { |
437 | if (isOpcWithIntImmediate(N: N0.getNode(), Opc: ISD::AND, Imm&: And_imm)) |
438 | std::swap(a&: N0, b&: N1); |
439 | } |
440 | if (!And_imm) |
441 | continue; |
442 | |
443 | // Check if the AND mask is an immediate of the form: 000.....1111111100 |
444 | unsigned TZ = llvm::countr_zero(Val: And_imm); |
445 | if (TZ != 1 && TZ != 2) |
446 | // Be conservative here. Shifter operands aren't always free. e.g. On |
447 | // Swift, left shifter operand of 1 / 2 for free but others are not. |
448 | // e.g. |
449 | // ubfx r3, r1, #16, #8 |
450 | // ldr.w r3, [r0, r3, lsl #2] |
451 | // vs. |
452 | // mov.w r9, #1020 |
453 | // and.w r2, r9, r1, lsr #14 |
454 | // ldr r2, [r0, r2] |
455 | continue; |
456 | And_imm >>= TZ; |
457 | if (And_imm & (And_imm + 1)) |
458 | continue; |
459 | |
460 | // Look for (and (srl X, c1), c2). |
461 | SDValue Srl = N1.getOperand(i: 0); |
462 | unsigned Srl_imm = 0; |
463 | if (!isOpcWithIntImmediate(N: Srl.getNode(), Opc: ISD::SRL, Imm&: Srl_imm) || |
464 | (Srl_imm <= 2)) |
465 | continue; |
466 | |
467 | // Make sure first operand is not a shifter operand which would prevent |
468 | // folding of the left shift. |
469 | SDValue CPTmp0; |
470 | SDValue CPTmp1; |
471 | SDValue CPTmp2; |
472 | if (isThumb2) { |
473 | if (SelectImmShifterOperand(N: N0, A&: CPTmp0, B&: CPTmp1)) |
474 | continue; |
475 | } else { |
476 | if (SelectImmShifterOperand(N: N0, A&: CPTmp0, B&: CPTmp1) || |
477 | SelectRegShifterOperand(N: N0, A&: CPTmp0, B&: CPTmp1, C&: CPTmp2)) |
478 | continue; |
479 | } |
480 | |
481 | // Now make the transformation. |
482 | Srl = CurDAG->getNode(Opcode: ISD::SRL, DL: SDLoc(Srl), VT: MVT::i32, |
483 | N1: Srl.getOperand(i: 0), |
484 | N2: CurDAG->getConstant(Val: Srl_imm + TZ, DL: SDLoc(Srl), |
485 | VT: MVT::i32)); |
486 | N1 = CurDAG->getNode(Opcode: ISD::AND, DL: SDLoc(N1), VT: MVT::i32, |
487 | N1: Srl, |
488 | N2: CurDAG->getConstant(Val: And_imm, DL: SDLoc(Srl), VT: MVT::i32)); |
489 | N1 = CurDAG->getNode(Opcode: ISD::SHL, DL: SDLoc(N1), VT: MVT::i32, |
490 | N1, N2: CurDAG->getConstant(Val: TZ, DL: SDLoc(Srl), VT: MVT::i32)); |
491 | CurDAG->UpdateNodeOperands(N: &N, Op1: N0, Op2: N1); |
492 | } |
493 | } |
494 | |
495 | /// hasNoVMLxHazardUse - Return true if it's desirable to select a FP MLA / MLS |
496 | /// node. VFP / NEON fp VMLA / VMLS instructions have special RAW hazards (at |
497 | /// least on current ARM implementations) which should be avoidded. |
498 | bool ARMDAGToDAGISel::hasNoVMLxHazardUse(SDNode *N) const { |
499 | if (OptLevel == CodeGenOptLevel::None) |
500 | return true; |
501 | |
502 | if (!Subtarget->hasVMLxHazards()) |
503 | return true; |
504 | |
505 | if (!N->hasOneUse()) |
506 | return false; |
507 | |
508 | SDNode *User = *N->user_begin(); |
509 | if (User->getOpcode() == ISD::CopyToReg) |
510 | return true; |
511 | if (User->isMachineOpcode()) { |
512 | const ARMBaseInstrInfo *TII = static_cast<const ARMBaseInstrInfo *>( |
513 | CurDAG->getSubtarget().getInstrInfo()); |
514 | |
515 | const MCInstrDesc &MCID = TII->get(Opcode: User->getMachineOpcode()); |
516 | if (MCID.mayStore()) |
517 | return true; |
518 | unsigned Opcode = MCID.getOpcode(); |
519 | if (Opcode == ARM::VMOVRS || Opcode == ARM::VMOVRRD) |
520 | return true; |
521 | // vmlx feeding into another vmlx. We actually want to unfold |
522 | // the use later in the MLxExpansion pass. e.g. |
523 | // vmla |
524 | // vmla (stall 8 cycles) |
525 | // |
526 | // vmul (5 cycles) |
527 | // vadd (5 cycles) |
528 | // vmla |
529 | // This adds up to about 18 - 19 cycles. |
530 | // |
531 | // vmla |
532 | // vmul (stall 4 cycles) |
533 | // vadd adds up to about 14 cycles. |
534 | return TII->isFpMLxInstruction(Opcode); |
535 | } |
536 | |
537 | return false; |
538 | } |
539 | |
540 | bool ARMDAGToDAGISel::isShifterOpProfitable(const SDValue &Shift, |
541 | ARM_AM::ShiftOpc ShOpcVal, |
542 | unsigned ShAmt) { |
543 | if (!Subtarget->isLikeA9() && !Subtarget->isSwift()) |
544 | return true; |
545 | if (Shift.hasOneUse()) |
546 | return true; |
547 | // R << 2 is free. |
548 | return ShOpcVal == ARM_AM::lsl && |
549 | (ShAmt == 2 || (Subtarget->isSwift() && ShAmt == 1)); |
550 | } |
551 | |
552 | bool ARMDAGToDAGISel::(const SDValue &N, |
553 | unsigned MaxShift, |
554 | unsigned &PowerOfTwo, |
555 | SDValue &NewMulConst) const { |
556 | assert(N.getOpcode() == ISD::MUL); |
557 | assert(MaxShift > 0); |
558 | |
559 | // If the multiply is used in more than one place then changing the constant |
560 | // will make other uses incorrect, so don't. |
561 | if (!N.hasOneUse()) return false; |
562 | // Check if the multiply is by a constant |
563 | ConstantSDNode *MulConst = dyn_cast<ConstantSDNode>(Val: N.getOperand(i: 1)); |
564 | if (!MulConst) return false; |
565 | // If the constant is used in more than one place then modifying it will mean |
566 | // we need to materialize two constants instead of one, which is a bad idea. |
567 | if (!MulConst->hasOneUse()) return false; |
568 | unsigned MulConstVal = MulConst->getZExtValue(); |
569 | if (MulConstVal == 0) return false; |
570 | |
571 | // Find the largest power of 2 that MulConstVal is a multiple of |
572 | PowerOfTwo = MaxShift; |
573 | while ((MulConstVal % (1 << PowerOfTwo)) != 0) { |
574 | --PowerOfTwo; |
575 | if (PowerOfTwo == 0) return false; |
576 | } |
577 | |
578 | // Only optimise if the new cost is better |
579 | unsigned NewMulConstVal = MulConstVal / (1 << PowerOfTwo); |
580 | NewMulConst = CurDAG->getConstant(Val: NewMulConstVal, DL: SDLoc(N), VT: MVT::i32); |
581 | unsigned OldCost = ConstantMaterializationCost(Val: MulConstVal, Subtarget); |
582 | unsigned NewCost = ConstantMaterializationCost(Val: NewMulConstVal, Subtarget); |
583 | return NewCost < OldCost; |
584 | } |
585 | |
586 | void ARMDAGToDAGISel::replaceDAGValue(const SDValue &N, SDValue M) { |
587 | CurDAG->RepositionNode(Position: N.getNode()->getIterator(), N: M.getNode()); |
588 | ReplaceUses(F: N, T: M); |
589 | } |
590 | |
591 | bool ARMDAGToDAGISel::SelectImmShifterOperand(SDValue N, |
592 | SDValue &BaseReg, |
593 | SDValue &Opc, |
594 | bool CheckProfitability) { |
595 | if (DisableShifterOp) |
596 | return false; |
597 | |
598 | // If N is a multiply-by-constant and it's profitable to extract a shift and |
599 | // use it in a shifted operand do so. |
600 | if (N.getOpcode() == ISD::MUL) { |
601 | unsigned PowerOfTwo = 0; |
602 | SDValue NewMulConst; |
603 | if (canExtractShiftFromMul(N, MaxShift: 31, PowerOfTwo, NewMulConst)) { |
604 | HandleSDNode Handle(N); |
605 | SDLoc Loc(N); |
606 | replaceDAGValue(N: N.getOperand(i: 1), M: NewMulConst); |
607 | BaseReg = Handle.getValue(); |
608 | Opc = CurDAG->getTargetConstant( |
609 | Val: ARM_AM::getSORegOpc(ShOp: ARM_AM::lsl, Imm: PowerOfTwo), DL: Loc, VT: MVT::i32); |
610 | return true; |
611 | } |
612 | } |
613 | |
614 | ARM_AM::ShiftOpc ShOpcVal = ARM_AM::getShiftOpcForNode(Opcode: N.getOpcode()); |
615 | |
616 | // Don't match base register only case. That is matched to a separate |
617 | // lower complexity pattern with explicit register operand. |
618 | if (ShOpcVal == ARM_AM::no_shift) return false; |
619 | |
620 | BaseReg = N.getOperand(i: 0); |
621 | unsigned ShImmVal = 0; |
622 | ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Val: N.getOperand(i: 1)); |
623 | if (!RHS) return false; |
624 | ShImmVal = RHS->getZExtValue() & 31; |
625 | Opc = CurDAG->getTargetConstant(Val: ARM_AM::getSORegOpc(ShOp: ShOpcVal, Imm: ShImmVal), |
626 | DL: SDLoc(N), VT: MVT::i32); |
627 | return true; |
628 | } |
629 | |
630 | bool ARMDAGToDAGISel::SelectRegShifterOperand(SDValue N, |
631 | SDValue &BaseReg, |
632 | SDValue &ShReg, |
633 | SDValue &Opc, |
634 | bool CheckProfitability) { |
635 | if (DisableShifterOp) |
636 | return false; |
637 | |
638 | ARM_AM::ShiftOpc ShOpcVal = ARM_AM::getShiftOpcForNode(Opcode: N.getOpcode()); |
639 | |
640 | // Don't match base register only case. That is matched to a separate |
641 | // lower complexity pattern with explicit register operand. |
642 | if (ShOpcVal == ARM_AM::no_shift) return false; |
643 | |
644 | BaseReg = N.getOperand(i: 0); |
645 | unsigned ShImmVal = 0; |
646 | ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Val: N.getOperand(i: 1)); |
647 | if (RHS) return false; |
648 | |
649 | ShReg = N.getOperand(i: 1); |
650 | if (CheckProfitability && !isShifterOpProfitable(Shift: N, ShOpcVal, ShAmt: ShImmVal)) |
651 | return false; |
652 | Opc = CurDAG->getTargetConstant(Val: ARM_AM::getSORegOpc(ShOp: ShOpcVal, Imm: ShImmVal), |
653 | DL: SDLoc(N), VT: MVT::i32); |
654 | return true; |
655 | } |
656 | |
657 | // Determine whether an ISD::OR's operands are suitable to turn the operation |
658 | // into an addition, which often has more compact encodings. |
659 | bool ARMDAGToDAGISel::SelectAddLikeOr(SDNode *Parent, SDValue N, SDValue &Out) { |
660 | assert(Parent->getOpcode() == ISD::OR && "unexpected parent" ); |
661 | Out = N; |
662 | return CurDAG->haveNoCommonBitsSet(A: N, B: Parent->getOperand(Num: 1)); |
663 | } |
664 | |
665 | |
666 | bool ARMDAGToDAGISel::SelectAddrModeImm12(SDValue N, |
667 | SDValue &Base, |
668 | SDValue &OffImm) { |
669 | // Match simple R + imm12 operands. |
670 | |
671 | // Base only. |
672 | if (N.getOpcode() != ISD::ADD && N.getOpcode() != ISD::SUB && |
673 | !CurDAG->isBaseWithConstantOffset(Op: N)) { |
674 | if (N.getOpcode() == ISD::FrameIndex) { |
675 | // Match frame index. |
676 | int FI = cast<FrameIndexSDNode>(Val&: N)->getIndex(); |
677 | Base = CurDAG->getTargetFrameIndex( |
678 | FI, VT: TLI->getPointerTy(DL: CurDAG->getDataLayout())); |
679 | OffImm = CurDAG->getTargetConstant(Val: 0, DL: SDLoc(N), VT: MVT::i32); |
680 | return true; |
681 | } |
682 | |
683 | if (N.getOpcode() == ARMISD::Wrapper && |
684 | N.getOperand(i: 0).getOpcode() != ISD::TargetGlobalAddress && |
685 | N.getOperand(i: 0).getOpcode() != ISD::TargetExternalSymbol && |
686 | N.getOperand(i: 0).getOpcode() != ISD::TargetGlobalTLSAddress) { |
687 | Base = N.getOperand(i: 0); |
688 | } else |
689 | Base = N; |
690 | OffImm = CurDAG->getTargetConstant(Val: 0, DL: SDLoc(N), VT: MVT::i32); |
691 | return true; |
692 | } |
693 | |
694 | if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Val: N.getOperand(i: 1))) { |
695 | int RHSC = (int)RHS->getSExtValue(); |
696 | if (N.getOpcode() == ISD::SUB) |
697 | RHSC = -RHSC; |
698 | |
699 | if (RHSC > -0x1000 && RHSC < 0x1000) { // 12 bits |
700 | Base = N.getOperand(i: 0); |
701 | if (Base.getOpcode() == ISD::FrameIndex) { |
702 | int FI = cast<FrameIndexSDNode>(Val&: Base)->getIndex(); |
703 | Base = CurDAG->getTargetFrameIndex( |
704 | FI, VT: TLI->getPointerTy(DL: CurDAG->getDataLayout())); |
705 | } |
706 | OffImm = CurDAG->getSignedTargetConstant(Val: RHSC, DL: SDLoc(N), VT: MVT::i32); |
707 | return true; |
708 | } |
709 | } |
710 | |
711 | // Base only. |
712 | Base = N; |
713 | OffImm = CurDAG->getTargetConstant(Val: 0, DL: SDLoc(N), VT: MVT::i32); |
714 | return true; |
715 | } |
716 | |
717 | |
718 | |
719 | bool ARMDAGToDAGISel::SelectLdStSOReg(SDValue N, SDValue &Base, SDValue &Offset, |
720 | SDValue &Opc) { |
721 | if (N.getOpcode() == ISD::MUL && |
722 | ((!Subtarget->isLikeA9() && !Subtarget->isSwift()) || N.hasOneUse())) { |
723 | if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Val: N.getOperand(i: 1))) { |
724 | // X * [3,5,9] -> X + X * [2,4,8] etc. |
725 | int RHSC = (int)RHS->getZExtValue(); |
726 | if (RHSC & 1) { |
727 | RHSC = RHSC & ~1; |
728 | ARM_AM::AddrOpc AddSub = ARM_AM::add; |
729 | if (RHSC < 0) { |
730 | AddSub = ARM_AM::sub; |
731 | RHSC = - RHSC; |
732 | } |
733 | if (isPowerOf2_32(Value: RHSC)) { |
734 | unsigned ShAmt = Log2_32(Value: RHSC); |
735 | Base = Offset = N.getOperand(i: 0); |
736 | Opc = CurDAG->getTargetConstant(Val: ARM_AM::getAM2Opc(Opc: AddSub, Imm12: ShAmt, |
737 | SO: ARM_AM::lsl), |
738 | DL: SDLoc(N), VT: MVT::i32); |
739 | return true; |
740 | } |
741 | } |
742 | } |
743 | } |
744 | |
745 | if (N.getOpcode() != ISD::ADD && N.getOpcode() != ISD::SUB && |
746 | // ISD::OR that is equivalent to an ISD::ADD. |
747 | !CurDAG->isBaseWithConstantOffset(Op: N)) |
748 | return false; |
749 | |
750 | // Leave simple R +/- imm12 operands for LDRi12 |
751 | if (N.getOpcode() == ISD::ADD || N.getOpcode() == ISD::OR) { |
752 | int RHSC; |
753 | if (isScaledConstantInRange(Node: N.getOperand(i: 1), /*Scale=*/1, |
754 | RangeMin: -0x1000+1, RangeMax: 0x1000, ScaledConstant&: RHSC)) // 12 bits. |
755 | return false; |
756 | } |
757 | |
758 | // Otherwise this is R +/- [possibly shifted] R. |
759 | ARM_AM::AddrOpc AddSub = N.getOpcode() == ISD::SUB ? ARM_AM::sub:ARM_AM::add; |
760 | ARM_AM::ShiftOpc ShOpcVal = |
761 | ARM_AM::getShiftOpcForNode(Opcode: N.getOperand(i: 1).getOpcode()); |
762 | unsigned ShAmt = 0; |
763 | |
764 | Base = N.getOperand(i: 0); |
765 | Offset = N.getOperand(i: 1); |
766 | |
767 | if (ShOpcVal != ARM_AM::no_shift) { |
768 | // Check to see if the RHS of the shift is a constant, if not, we can't fold |
769 | // it. |
770 | if (ConstantSDNode *Sh = |
771 | dyn_cast<ConstantSDNode>(Val: N.getOperand(i: 1).getOperand(i: 1))) { |
772 | ShAmt = Sh->getZExtValue(); |
773 | if (isShifterOpProfitable(Shift: Offset, ShOpcVal, ShAmt)) |
774 | Offset = N.getOperand(i: 1).getOperand(i: 0); |
775 | else { |
776 | ShAmt = 0; |
777 | ShOpcVal = ARM_AM::no_shift; |
778 | } |
779 | } else { |
780 | ShOpcVal = ARM_AM::no_shift; |
781 | } |
782 | } |
783 | |
784 | // Try matching (R shl C) + (R). |
785 | if (N.getOpcode() != ISD::SUB && ShOpcVal == ARM_AM::no_shift && |
786 | !(Subtarget->isLikeA9() || Subtarget->isSwift() || |
787 | N.getOperand(i: 0).hasOneUse())) { |
788 | ShOpcVal = ARM_AM::getShiftOpcForNode(Opcode: N.getOperand(i: 0).getOpcode()); |
789 | if (ShOpcVal != ARM_AM::no_shift) { |
790 | // Check to see if the RHS of the shift is a constant, if not, we can't |
791 | // fold it. |
792 | if (ConstantSDNode *Sh = |
793 | dyn_cast<ConstantSDNode>(Val: N.getOperand(i: 0).getOperand(i: 1))) { |
794 | ShAmt = Sh->getZExtValue(); |
795 | if (isShifterOpProfitable(Shift: N.getOperand(i: 0), ShOpcVal, ShAmt)) { |
796 | Offset = N.getOperand(i: 0).getOperand(i: 0); |
797 | Base = N.getOperand(i: 1); |
798 | } else { |
799 | ShAmt = 0; |
800 | ShOpcVal = ARM_AM::no_shift; |
801 | } |
802 | } else { |
803 | ShOpcVal = ARM_AM::no_shift; |
804 | } |
805 | } |
806 | } |
807 | |
808 | // If Offset is a multiply-by-constant and it's profitable to extract a shift |
809 | // and use it in a shifted operand do so. |
810 | if (Offset.getOpcode() == ISD::MUL && N.hasOneUse()) { |
811 | unsigned PowerOfTwo = 0; |
812 | SDValue NewMulConst; |
813 | if (canExtractShiftFromMul(N: Offset, MaxShift: 31, PowerOfTwo, NewMulConst)) { |
814 | HandleSDNode Handle(Offset); |
815 | replaceDAGValue(N: Offset.getOperand(i: 1), M: NewMulConst); |
816 | Offset = Handle.getValue(); |
817 | ShAmt = PowerOfTwo; |
818 | ShOpcVal = ARM_AM::lsl; |
819 | } |
820 | } |
821 | |
822 | Opc = CurDAG->getTargetConstant(Val: ARM_AM::getAM2Opc(Opc: AddSub, Imm12: ShAmt, SO: ShOpcVal), |
823 | DL: SDLoc(N), VT: MVT::i32); |
824 | return true; |
825 | } |
826 | |
827 | bool ARMDAGToDAGISel::SelectAddrMode2OffsetReg(SDNode *Op, SDValue N, |
828 | SDValue &Offset, SDValue &Opc) { |
829 | unsigned Opcode = Op->getOpcode(); |
830 | ISD::MemIndexedMode AM = (Opcode == ISD::LOAD) |
831 | ? cast<LoadSDNode>(Val: Op)->getAddressingMode() |
832 | : cast<StoreSDNode>(Val: Op)->getAddressingMode(); |
833 | ARM_AM::AddrOpc AddSub = (AM == ISD::PRE_INC || AM == ISD::POST_INC) |
834 | ? ARM_AM::add : ARM_AM::sub; |
835 | int Val; |
836 | if (isScaledConstantInRange(Node: N, /*Scale=*/1, RangeMin: 0, RangeMax: 0x1000, ScaledConstant&: Val)) |
837 | return false; |
838 | |
839 | Offset = N; |
840 | ARM_AM::ShiftOpc ShOpcVal = ARM_AM::getShiftOpcForNode(Opcode: N.getOpcode()); |
841 | unsigned ShAmt = 0; |
842 | if (ShOpcVal != ARM_AM::no_shift) { |
843 | // Check to see if the RHS of the shift is a constant, if not, we can't fold |
844 | // it. |
845 | if (ConstantSDNode *Sh = dyn_cast<ConstantSDNode>(Val: N.getOperand(i: 1))) { |
846 | ShAmt = Sh->getZExtValue(); |
847 | if (isShifterOpProfitable(Shift: N, ShOpcVal, ShAmt)) |
848 | Offset = N.getOperand(i: 0); |
849 | else { |
850 | ShAmt = 0; |
851 | ShOpcVal = ARM_AM::no_shift; |
852 | } |
853 | } else { |
854 | ShOpcVal = ARM_AM::no_shift; |
855 | } |
856 | } |
857 | |
858 | Opc = CurDAG->getTargetConstant(Val: ARM_AM::getAM2Opc(Opc: AddSub, Imm12: ShAmt, SO: ShOpcVal), |
859 | DL: SDLoc(N), VT: MVT::i32); |
860 | return true; |
861 | } |
862 | |
863 | bool ARMDAGToDAGISel::SelectAddrMode2OffsetImmPre(SDNode *Op, SDValue N, |
864 | SDValue &Offset, SDValue &Opc) { |
865 | unsigned Opcode = Op->getOpcode(); |
866 | ISD::MemIndexedMode AM = (Opcode == ISD::LOAD) |
867 | ? cast<LoadSDNode>(Val: Op)->getAddressingMode() |
868 | : cast<StoreSDNode>(Val: Op)->getAddressingMode(); |
869 | ARM_AM::AddrOpc AddSub = (AM == ISD::PRE_INC || AM == ISD::POST_INC) |
870 | ? ARM_AM::add : ARM_AM::sub; |
871 | int Val; |
872 | if (isScaledConstantInRange(Node: N, /*Scale=*/1, RangeMin: 0, RangeMax: 0x1000, ScaledConstant&: Val)) { // 12 bits. |
873 | if (AddSub == ARM_AM::sub) Val *= -1; |
874 | Offset = CurDAG->getRegister(Reg: 0, VT: MVT::i32); |
875 | Opc = CurDAG->getSignedTargetConstant(Val, DL: SDLoc(Op), VT: MVT::i32); |
876 | return true; |
877 | } |
878 | |
879 | return false; |
880 | } |
881 | |
882 | |
883 | bool ARMDAGToDAGISel::SelectAddrMode2OffsetImm(SDNode *Op, SDValue N, |
884 | SDValue &Offset, SDValue &Opc) { |
885 | unsigned Opcode = Op->getOpcode(); |
886 | ISD::MemIndexedMode AM = (Opcode == ISD::LOAD) |
887 | ? cast<LoadSDNode>(Val: Op)->getAddressingMode() |
888 | : cast<StoreSDNode>(Val: Op)->getAddressingMode(); |
889 | ARM_AM::AddrOpc AddSub = (AM == ISD::PRE_INC || AM == ISD::POST_INC) |
890 | ? ARM_AM::add : ARM_AM::sub; |
891 | int Val; |
892 | if (isScaledConstantInRange(Node: N, /*Scale=*/1, RangeMin: 0, RangeMax: 0x1000, ScaledConstant&: Val)) { // 12 bits. |
893 | Offset = CurDAG->getRegister(Reg: 0, VT: MVT::i32); |
894 | Opc = CurDAG->getTargetConstant(Val: ARM_AM::getAM2Opc(Opc: AddSub, Imm12: Val, |
895 | SO: ARM_AM::no_shift), |
896 | DL: SDLoc(Op), VT: MVT::i32); |
897 | return true; |
898 | } |
899 | |
900 | return false; |
901 | } |
902 | |
903 | bool ARMDAGToDAGISel::SelectAddrOffsetNone(SDValue N, SDValue &Base) { |
904 | Base = N; |
905 | return true; |
906 | } |
907 | |
908 | bool ARMDAGToDAGISel::SelectAddrMode3(SDValue N, |
909 | SDValue &Base, SDValue &Offset, |
910 | SDValue &Opc) { |
911 | if (N.getOpcode() == ISD::SUB) { |
912 | // X - C is canonicalize to X + -C, no need to handle it here. |
913 | Base = N.getOperand(i: 0); |
914 | Offset = N.getOperand(i: 1); |
915 | Opc = CurDAG->getTargetConstant(Val: ARM_AM::getAM3Opc(Opc: ARM_AM::sub, Offset: 0), DL: SDLoc(N), |
916 | VT: MVT::i32); |
917 | return true; |
918 | } |
919 | |
920 | if (!CurDAG->isBaseWithConstantOffset(Op: N)) { |
921 | Base = N; |
922 | if (N.getOpcode() == ISD::FrameIndex) { |
923 | int FI = cast<FrameIndexSDNode>(Val&: N)->getIndex(); |
924 | Base = CurDAG->getTargetFrameIndex( |
925 | FI, VT: TLI->getPointerTy(DL: CurDAG->getDataLayout())); |
926 | } |
927 | Offset = CurDAG->getRegister(Reg: 0, VT: MVT::i32); |
928 | Opc = CurDAG->getTargetConstant(Val: ARM_AM::getAM3Opc(Opc: ARM_AM::add, Offset: 0), DL: SDLoc(N), |
929 | VT: MVT::i32); |
930 | return true; |
931 | } |
932 | |
933 | // If the RHS is +/- imm8, fold into addr mode. |
934 | int RHSC; |
935 | if (isScaledConstantInRange(Node: N.getOperand(i: 1), /*Scale=*/1, |
936 | RangeMin: -256 + 1, RangeMax: 256, ScaledConstant&: RHSC)) { // 8 bits. |
937 | Base = N.getOperand(i: 0); |
938 | if (Base.getOpcode() == ISD::FrameIndex) { |
939 | int FI = cast<FrameIndexSDNode>(Val&: Base)->getIndex(); |
940 | Base = CurDAG->getTargetFrameIndex( |
941 | FI, VT: TLI->getPointerTy(DL: CurDAG->getDataLayout())); |
942 | } |
943 | Offset = CurDAG->getRegister(Reg: 0, VT: MVT::i32); |
944 | |
945 | ARM_AM::AddrOpc AddSub = ARM_AM::add; |
946 | if (RHSC < 0) { |
947 | AddSub = ARM_AM::sub; |
948 | RHSC = -RHSC; |
949 | } |
950 | Opc = CurDAG->getTargetConstant(Val: ARM_AM::getAM3Opc(Opc: AddSub, Offset: RHSC), DL: SDLoc(N), |
951 | VT: MVT::i32); |
952 | return true; |
953 | } |
954 | |
955 | Base = N.getOperand(i: 0); |
956 | Offset = N.getOperand(i: 1); |
957 | Opc = CurDAG->getTargetConstant(Val: ARM_AM::getAM3Opc(Opc: ARM_AM::add, Offset: 0), DL: SDLoc(N), |
958 | VT: MVT::i32); |
959 | return true; |
960 | } |
961 | |
962 | bool ARMDAGToDAGISel::SelectAddrMode3Offset(SDNode *Op, SDValue N, |
963 | SDValue &Offset, SDValue &Opc) { |
964 | unsigned Opcode = Op->getOpcode(); |
965 | ISD::MemIndexedMode AM = (Opcode == ISD::LOAD) |
966 | ? cast<LoadSDNode>(Val: Op)->getAddressingMode() |
967 | : cast<StoreSDNode>(Val: Op)->getAddressingMode(); |
968 | ARM_AM::AddrOpc AddSub = (AM == ISD::PRE_INC || AM == ISD::POST_INC) |
969 | ? ARM_AM::add : ARM_AM::sub; |
970 | int Val; |
971 | if (isScaledConstantInRange(Node: N, /*Scale=*/1, RangeMin: 0, RangeMax: 256, ScaledConstant&: Val)) { // 12 bits. |
972 | Offset = CurDAG->getRegister(Reg: 0, VT: MVT::i32); |
973 | Opc = CurDAG->getTargetConstant(Val: ARM_AM::getAM3Opc(Opc: AddSub, Offset: Val), DL: SDLoc(Op), |
974 | VT: MVT::i32); |
975 | return true; |
976 | } |
977 | |
978 | Offset = N; |
979 | Opc = CurDAG->getTargetConstant(Val: ARM_AM::getAM3Opc(Opc: AddSub, Offset: 0), DL: SDLoc(Op), |
980 | VT: MVT::i32); |
981 | return true; |
982 | } |
983 | |
984 | bool ARMDAGToDAGISel::IsAddressingMode5(SDValue N, SDValue &Base, SDValue &Offset, |
985 | bool FP16) { |
986 | if (!CurDAG->isBaseWithConstantOffset(Op: N)) { |
987 | Base = N; |
988 | if (N.getOpcode() == ISD::FrameIndex) { |
989 | int FI = cast<FrameIndexSDNode>(Val&: N)->getIndex(); |
990 | Base = CurDAG->getTargetFrameIndex( |
991 | FI, VT: TLI->getPointerTy(DL: CurDAG->getDataLayout())); |
992 | } else if (N.getOpcode() == ARMISD::Wrapper && |
993 | N.getOperand(i: 0).getOpcode() != ISD::TargetGlobalAddress && |
994 | N.getOperand(i: 0).getOpcode() != ISD::TargetExternalSymbol && |
995 | N.getOperand(i: 0).getOpcode() != ISD::TargetGlobalTLSAddress) { |
996 | Base = N.getOperand(i: 0); |
997 | } |
998 | Offset = CurDAG->getTargetConstant(Val: ARM_AM::getAM5Opc(Opc: ARM_AM::add, Offset: 0), |
999 | DL: SDLoc(N), VT: MVT::i32); |
1000 | return true; |
1001 | } |
1002 | |
1003 | // If the RHS is +/- imm8, fold into addr mode. |
1004 | int RHSC; |
1005 | const int Scale = FP16 ? 2 : 4; |
1006 | |
1007 | if (isScaledConstantInRange(Node: N.getOperand(i: 1), Scale, RangeMin: -255, RangeMax: 256, ScaledConstant&: RHSC)) { |
1008 | Base = N.getOperand(i: 0); |
1009 | if (Base.getOpcode() == ISD::FrameIndex) { |
1010 | int FI = cast<FrameIndexSDNode>(Val&: Base)->getIndex(); |
1011 | Base = CurDAG->getTargetFrameIndex( |
1012 | FI, VT: TLI->getPointerTy(DL: CurDAG->getDataLayout())); |
1013 | } |
1014 | |
1015 | ARM_AM::AddrOpc AddSub = ARM_AM::add; |
1016 | if (RHSC < 0) { |
1017 | AddSub = ARM_AM::sub; |
1018 | RHSC = -RHSC; |
1019 | } |
1020 | |
1021 | if (FP16) |
1022 | Offset = CurDAG->getTargetConstant(Val: ARM_AM::getAM5FP16Opc(Opc: AddSub, Offset: RHSC), |
1023 | DL: SDLoc(N), VT: MVT::i32); |
1024 | else |
1025 | Offset = CurDAG->getTargetConstant(Val: ARM_AM::getAM5Opc(Opc: AddSub, Offset: RHSC), |
1026 | DL: SDLoc(N), VT: MVT::i32); |
1027 | |
1028 | return true; |
1029 | } |
1030 | |
1031 | Base = N; |
1032 | |
1033 | if (FP16) |
1034 | Offset = CurDAG->getTargetConstant(Val: ARM_AM::getAM5FP16Opc(Opc: ARM_AM::add, Offset: 0), |
1035 | DL: SDLoc(N), VT: MVT::i32); |
1036 | else |
1037 | Offset = CurDAG->getTargetConstant(Val: ARM_AM::getAM5Opc(Opc: ARM_AM::add, Offset: 0), |
1038 | DL: SDLoc(N), VT: MVT::i32); |
1039 | |
1040 | return true; |
1041 | } |
1042 | |
1043 | bool ARMDAGToDAGISel::SelectAddrMode5(SDValue N, |
1044 | SDValue &Base, SDValue &Offset) { |
1045 | return IsAddressingMode5(N, Base, Offset, /*FP16=*/ false); |
1046 | } |
1047 | |
1048 | bool ARMDAGToDAGISel::SelectAddrMode5FP16(SDValue N, |
1049 | SDValue &Base, SDValue &Offset) { |
1050 | return IsAddressingMode5(N, Base, Offset, /*FP16=*/ true); |
1051 | } |
1052 | |
1053 | bool ARMDAGToDAGISel::SelectAddrMode6(SDNode *Parent, SDValue N, SDValue &Addr, |
1054 | SDValue &Align) { |
1055 | Addr = N; |
1056 | |
1057 | unsigned Alignment = 0; |
1058 | |
1059 | MemSDNode *MemN = cast<MemSDNode>(Val: Parent); |
1060 | |
1061 | if (isa<LSBaseSDNode>(Val: MemN) || |
1062 | ((MemN->getOpcode() == ARMISD::VST1_UPD || |
1063 | MemN->getOpcode() == ARMISD::VLD1_UPD) && |
1064 | MemN->getConstantOperandVal(Num: MemN->getNumOperands() - 1) == 1)) { |
1065 | // This case occurs only for VLD1-lane/dup and VST1-lane instructions. |
1066 | // The maximum alignment is equal to the memory size being referenced. |
1067 | llvm::Align MMOAlign = MemN->getAlign(); |
1068 | unsigned MemSize = MemN->getMemoryVT().getSizeInBits() / 8; |
1069 | if (MMOAlign.value() >= MemSize && MemSize > 1) |
1070 | Alignment = MemSize; |
1071 | } else { |
1072 | // All other uses of addrmode6 are for intrinsics. For now just record |
1073 | // the raw alignment value; it will be refined later based on the legal |
1074 | // alignment operands for the intrinsic. |
1075 | Alignment = MemN->getAlign().value(); |
1076 | } |
1077 | |
1078 | Align = CurDAG->getTargetConstant(Val: Alignment, DL: SDLoc(N), VT: MVT::i32); |
1079 | return true; |
1080 | } |
1081 | |
1082 | bool ARMDAGToDAGISel::SelectAddrMode6Offset(SDNode *Op, SDValue N, |
1083 | SDValue &Offset) { |
1084 | LSBaseSDNode *LdSt = cast<LSBaseSDNode>(Val: Op); |
1085 | ISD::MemIndexedMode AM = LdSt->getAddressingMode(); |
1086 | if (AM != ISD::POST_INC) |
1087 | return false; |
1088 | Offset = N; |
1089 | if (ConstantSDNode *NC = dyn_cast<ConstantSDNode>(Val&: N)) { |
1090 | if (NC->getZExtValue() * 8 == LdSt->getMemoryVT().getSizeInBits()) |
1091 | Offset = CurDAG->getRegister(Reg: 0, VT: MVT::i32); |
1092 | } |
1093 | return true; |
1094 | } |
1095 | |
1096 | bool ARMDAGToDAGISel::SelectAddrModePC(SDValue N, |
1097 | SDValue &Offset, SDValue &Label) { |
1098 | if (N.getOpcode() == ARMISD::PIC_ADD && N.hasOneUse()) { |
1099 | Offset = N.getOperand(i: 0); |
1100 | SDValue N1 = N.getOperand(i: 1); |
1101 | Label = CurDAG->getTargetConstant(Val: N1->getAsZExtVal(), DL: SDLoc(N), VT: MVT::i32); |
1102 | return true; |
1103 | } |
1104 | |
1105 | return false; |
1106 | } |
1107 | |
1108 | |
1109 | //===----------------------------------------------------------------------===// |
1110 | // Thumb Addressing Modes |
1111 | //===----------------------------------------------------------------------===// |
1112 | |
1113 | static bool shouldUseZeroOffsetLdSt(SDValue N) { |
1114 | // Negative numbers are difficult to materialise in thumb1. If we are |
1115 | // selecting the add of a negative, instead try to select ri with a zero |
1116 | // offset, so create the add node directly which will become a sub. |
1117 | if (N.getOpcode() != ISD::ADD) |
1118 | return false; |
1119 | |
1120 | // Look for an imm which is not legal for ld/st, but is legal for sub. |
1121 | if (auto C = dyn_cast<ConstantSDNode>(Val: N.getOperand(i: 1))) |
1122 | return C->getSExtValue() < 0 && C->getSExtValue() >= -255; |
1123 | |
1124 | return false; |
1125 | } |
1126 | |
1127 | bool ARMDAGToDAGISel::SelectThumbAddrModeRRSext(SDValue N, SDValue &Base, |
1128 | SDValue &Offset) { |
1129 | if (N.getOpcode() != ISD::ADD && !CurDAG->isBaseWithConstantOffset(Op: N)) { |
1130 | if (!isNullConstant(V: N)) |
1131 | return false; |
1132 | |
1133 | Base = Offset = N; |
1134 | return true; |
1135 | } |
1136 | |
1137 | Base = N.getOperand(i: 0); |
1138 | Offset = N.getOperand(i: 1); |
1139 | return true; |
1140 | } |
1141 | |
1142 | bool ARMDAGToDAGISel::SelectThumbAddrModeRR(SDValue N, SDValue &Base, |
1143 | SDValue &Offset) { |
1144 | if (shouldUseZeroOffsetLdSt(N)) |
1145 | return false; // Select ri instead |
1146 | return SelectThumbAddrModeRRSext(N, Base, Offset); |
1147 | } |
1148 | |
1149 | bool |
1150 | ARMDAGToDAGISel::SelectThumbAddrModeImm5S(SDValue N, unsigned Scale, |
1151 | SDValue &Base, SDValue &OffImm) { |
1152 | if (shouldUseZeroOffsetLdSt(N)) { |
1153 | Base = N; |
1154 | OffImm = CurDAG->getTargetConstant(Val: 0, DL: SDLoc(N), VT: MVT::i32); |
1155 | return true; |
1156 | } |
1157 | |
1158 | if (!CurDAG->isBaseWithConstantOffset(Op: N)) { |
1159 | if (N.getOpcode() == ISD::ADD) { |
1160 | return false; // We want to select register offset instead |
1161 | } else if (N.getOpcode() == ARMISD::Wrapper && |
1162 | N.getOperand(i: 0).getOpcode() != ISD::TargetGlobalAddress && |
1163 | N.getOperand(i: 0).getOpcode() != ISD::TargetExternalSymbol && |
1164 | N.getOperand(i: 0).getOpcode() != ISD::TargetConstantPool && |
1165 | N.getOperand(i: 0).getOpcode() != ISD::TargetGlobalTLSAddress) { |
1166 | Base = N.getOperand(i: 0); |
1167 | } else { |
1168 | Base = N; |
1169 | } |
1170 | |
1171 | OffImm = CurDAG->getTargetConstant(Val: 0, DL: SDLoc(N), VT: MVT::i32); |
1172 | return true; |
1173 | } |
1174 | |
1175 | // If the RHS is + imm5 * scale, fold into addr mode. |
1176 | int RHSC; |
1177 | if (isScaledConstantInRange(Node: N.getOperand(i: 1), Scale, RangeMin: 0, RangeMax: 32, ScaledConstant&: RHSC)) { |
1178 | Base = N.getOperand(i: 0); |
1179 | OffImm = CurDAG->getSignedTargetConstant(Val: RHSC, DL: SDLoc(N), VT: MVT::i32); |
1180 | return true; |
1181 | } |
1182 | |
1183 | // Offset is too large, so use register offset instead. |
1184 | return false; |
1185 | } |
1186 | |
1187 | bool |
1188 | ARMDAGToDAGISel::SelectThumbAddrModeImm5S4(SDValue N, SDValue &Base, |
1189 | SDValue &OffImm) { |
1190 | return SelectThumbAddrModeImm5S(N, Scale: 4, Base, OffImm); |
1191 | } |
1192 | |
1193 | bool |
1194 | ARMDAGToDAGISel::SelectThumbAddrModeImm5S2(SDValue N, SDValue &Base, |
1195 | SDValue &OffImm) { |
1196 | return SelectThumbAddrModeImm5S(N, Scale: 2, Base, OffImm); |
1197 | } |
1198 | |
1199 | bool |
1200 | ARMDAGToDAGISel::SelectThumbAddrModeImm5S1(SDValue N, SDValue &Base, |
1201 | SDValue &OffImm) { |
1202 | return SelectThumbAddrModeImm5S(N, Scale: 1, Base, OffImm); |
1203 | } |
1204 | |
1205 | bool ARMDAGToDAGISel::SelectThumbAddrModeSP(SDValue N, |
1206 | SDValue &Base, SDValue &OffImm) { |
1207 | if (N.getOpcode() == ISD::FrameIndex) { |
1208 | int FI = cast<FrameIndexSDNode>(Val&: N)->getIndex(); |
1209 | // Only multiples of 4 are allowed for the offset, so the frame object |
1210 | // alignment must be at least 4. |
1211 | MachineFrameInfo &MFI = MF->getFrameInfo(); |
1212 | if (MFI.getObjectAlign(ObjectIdx: FI) < Align(4)) |
1213 | MFI.setObjectAlignment(ObjectIdx: FI, Alignment: Align(4)); |
1214 | Base = CurDAG->getTargetFrameIndex( |
1215 | FI, VT: TLI->getPointerTy(DL: CurDAG->getDataLayout())); |
1216 | OffImm = CurDAG->getTargetConstant(Val: 0, DL: SDLoc(N), VT: MVT::i32); |
1217 | return true; |
1218 | } |
1219 | |
1220 | if (!CurDAG->isBaseWithConstantOffset(Op: N)) |
1221 | return false; |
1222 | |
1223 | if (N.getOperand(i: 0).getOpcode() == ISD::FrameIndex) { |
1224 | // If the RHS is + imm8 * scale, fold into addr mode. |
1225 | int RHSC; |
1226 | if (isScaledConstantInRange(Node: N.getOperand(i: 1), /*Scale=*/4, RangeMin: 0, RangeMax: 256, ScaledConstant&: RHSC)) { |
1227 | Base = N.getOperand(i: 0); |
1228 | int FI = cast<FrameIndexSDNode>(Val&: Base)->getIndex(); |
1229 | // Make sure the offset is inside the object, or we might fail to |
1230 | // allocate an emergency spill slot. (An out-of-range access is UB, but |
1231 | // it could show up anyway.) |
1232 | MachineFrameInfo &MFI = MF->getFrameInfo(); |
1233 | if (RHSC * 4 < MFI.getObjectSize(ObjectIdx: FI)) { |
1234 | // For LHS+RHS to result in an offset that's a multiple of 4 the object |
1235 | // indexed by the LHS must be 4-byte aligned. |
1236 | if (!MFI.isFixedObjectIndex(ObjectIdx: FI) && MFI.getObjectAlign(ObjectIdx: FI) < Align(4)) |
1237 | MFI.setObjectAlignment(ObjectIdx: FI, Alignment: Align(4)); |
1238 | if (MFI.getObjectAlign(ObjectIdx: FI) >= Align(4)) { |
1239 | Base = CurDAG->getTargetFrameIndex( |
1240 | FI, VT: TLI->getPointerTy(DL: CurDAG->getDataLayout())); |
1241 | OffImm = CurDAG->getSignedTargetConstant(Val: RHSC, DL: SDLoc(N), VT: MVT::i32); |
1242 | return true; |
1243 | } |
1244 | } |
1245 | } |
1246 | } |
1247 | |
1248 | return false; |
1249 | } |
1250 | |
1251 | template <unsigned Shift> |
1252 | bool ARMDAGToDAGISel::SelectTAddrModeImm7(SDValue N, SDValue &Base, |
1253 | SDValue &OffImm) { |
1254 | if (N.getOpcode() == ISD::SUB || CurDAG->isBaseWithConstantOffset(Op: N)) { |
1255 | int RHSC; |
1256 | if (isScaledConstantInRange(Node: N.getOperand(i: 1), Scale: 1 << Shift, RangeMin: -0x7f, RangeMax: 0x80, |
1257 | ScaledConstant&: RHSC)) { |
1258 | Base = N.getOperand(i: 0); |
1259 | if (N.getOpcode() == ISD::SUB) |
1260 | RHSC = -RHSC; |
1261 | OffImm = CurDAG->getSignedTargetConstant(Val: RHSC * (1 << Shift), DL: SDLoc(N), |
1262 | VT: MVT::i32); |
1263 | return true; |
1264 | } |
1265 | } |
1266 | |
1267 | // Base only. |
1268 | Base = N; |
1269 | OffImm = CurDAG->getTargetConstant(Val: 0, DL: SDLoc(N), VT: MVT::i32); |
1270 | return true; |
1271 | } |
1272 | |
1273 | |
1274 | //===----------------------------------------------------------------------===// |
1275 | // Thumb 2 Addressing Modes |
1276 | //===----------------------------------------------------------------------===// |
1277 | |
1278 | |
1279 | bool ARMDAGToDAGISel::SelectT2AddrModeImm12(SDValue N, |
1280 | SDValue &Base, SDValue &OffImm) { |
1281 | // Match simple R + imm12 operands. |
1282 | |
1283 | // Base only. |
1284 | if (N.getOpcode() != ISD::ADD && N.getOpcode() != ISD::SUB && |
1285 | !CurDAG->isBaseWithConstantOffset(Op: N)) { |
1286 | if (N.getOpcode() == ISD::FrameIndex) { |
1287 | // Match frame index. |
1288 | int FI = cast<FrameIndexSDNode>(Val&: N)->getIndex(); |
1289 | Base = CurDAG->getTargetFrameIndex( |
1290 | FI, VT: TLI->getPointerTy(DL: CurDAG->getDataLayout())); |
1291 | OffImm = CurDAG->getTargetConstant(Val: 0, DL: SDLoc(N), VT: MVT::i32); |
1292 | return true; |
1293 | } |
1294 | |
1295 | if (N.getOpcode() == ARMISD::Wrapper && |
1296 | N.getOperand(i: 0).getOpcode() != ISD::TargetGlobalAddress && |
1297 | N.getOperand(i: 0).getOpcode() != ISD::TargetExternalSymbol && |
1298 | N.getOperand(i: 0).getOpcode() != ISD::TargetGlobalTLSAddress) { |
1299 | Base = N.getOperand(i: 0); |
1300 | if (Base.getOpcode() == ISD::TargetConstantPool) |
1301 | return false; // We want to select t2LDRpci instead. |
1302 | } else |
1303 | Base = N; |
1304 | OffImm = CurDAG->getTargetConstant(Val: 0, DL: SDLoc(N), VT: MVT::i32); |
1305 | return true; |
1306 | } |
1307 | |
1308 | if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Val: N.getOperand(i: 1))) { |
1309 | if (SelectT2AddrModeImm8(N, Base, OffImm)) |
1310 | // Let t2LDRi8 handle (R - imm8). |
1311 | return false; |
1312 | |
1313 | int RHSC = (int)RHS->getZExtValue(); |
1314 | if (N.getOpcode() == ISD::SUB) |
1315 | RHSC = -RHSC; |
1316 | |
1317 | if (RHSC >= 0 && RHSC < 0x1000) { // 12 bits (unsigned) |
1318 | Base = N.getOperand(i: 0); |
1319 | if (Base.getOpcode() == ISD::FrameIndex) { |
1320 | int FI = cast<FrameIndexSDNode>(Val&: Base)->getIndex(); |
1321 | Base = CurDAG->getTargetFrameIndex( |
1322 | FI, VT: TLI->getPointerTy(DL: CurDAG->getDataLayout())); |
1323 | } |
1324 | OffImm = CurDAG->getSignedTargetConstant(Val: RHSC, DL: SDLoc(N), VT: MVT::i32); |
1325 | return true; |
1326 | } |
1327 | } |
1328 | |
1329 | // Base only. |
1330 | Base = N; |
1331 | OffImm = CurDAG->getTargetConstant(Val: 0, DL: SDLoc(N), VT: MVT::i32); |
1332 | return true; |
1333 | } |
1334 | |
1335 | template <unsigned Shift> |
1336 | bool ARMDAGToDAGISel::SelectT2AddrModeImm8(SDValue N, SDValue &Base, |
1337 | SDValue &OffImm) { |
1338 | if (N.getOpcode() == ISD::SUB || CurDAG->isBaseWithConstantOffset(Op: N)) { |
1339 | int RHSC; |
1340 | if (isScaledConstantInRange(Node: N.getOperand(i: 1), Scale: 1 << Shift, RangeMin: -255, RangeMax: 256, ScaledConstant&: RHSC)) { |
1341 | Base = N.getOperand(i: 0); |
1342 | if (Base.getOpcode() == ISD::FrameIndex) { |
1343 | int FI = cast<FrameIndexSDNode>(Val&: Base)->getIndex(); |
1344 | Base = CurDAG->getTargetFrameIndex( |
1345 | FI, VT: TLI->getPointerTy(DL: CurDAG->getDataLayout())); |
1346 | } |
1347 | |
1348 | if (N.getOpcode() == ISD::SUB) |
1349 | RHSC = -RHSC; |
1350 | OffImm = CurDAG->getSignedTargetConstant(Val: RHSC * (1 << Shift), DL: SDLoc(N), |
1351 | VT: MVT::i32); |
1352 | return true; |
1353 | } |
1354 | } |
1355 | |
1356 | // Base only. |
1357 | Base = N; |
1358 | OffImm = CurDAG->getTargetConstant(Val: 0, DL: SDLoc(N), VT: MVT::i32); |
1359 | return true; |
1360 | } |
1361 | |
1362 | bool ARMDAGToDAGISel::SelectT2AddrModeImm8(SDValue N, |
1363 | SDValue &Base, SDValue &OffImm) { |
1364 | // Match simple R - imm8 operands. |
1365 | if (N.getOpcode() != ISD::ADD && N.getOpcode() != ISD::SUB && |
1366 | !CurDAG->isBaseWithConstantOffset(Op: N)) |
1367 | return false; |
1368 | |
1369 | if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Val: N.getOperand(i: 1))) { |
1370 | int RHSC = (int)RHS->getSExtValue(); |
1371 | if (N.getOpcode() == ISD::SUB) |
1372 | RHSC = -RHSC; |
1373 | |
1374 | if ((RHSC >= -255) && (RHSC < 0)) { // 8 bits (always negative) |
1375 | Base = N.getOperand(i: 0); |
1376 | if (Base.getOpcode() == ISD::FrameIndex) { |
1377 | int FI = cast<FrameIndexSDNode>(Val&: Base)->getIndex(); |
1378 | Base = CurDAG->getTargetFrameIndex( |
1379 | FI, VT: TLI->getPointerTy(DL: CurDAG->getDataLayout())); |
1380 | } |
1381 | OffImm = CurDAG->getSignedTargetConstant(Val: RHSC, DL: SDLoc(N), VT: MVT::i32); |
1382 | return true; |
1383 | } |
1384 | } |
1385 | |
1386 | return false; |
1387 | } |
1388 | |
1389 | bool ARMDAGToDAGISel::SelectT2AddrModeImm8Offset(SDNode *Op, SDValue N, |
1390 | SDValue &OffImm){ |
1391 | unsigned Opcode = Op->getOpcode(); |
1392 | ISD::MemIndexedMode AM = (Opcode == ISD::LOAD) |
1393 | ? cast<LoadSDNode>(Val: Op)->getAddressingMode() |
1394 | : cast<StoreSDNode>(Val: Op)->getAddressingMode(); |
1395 | int RHSC; |
1396 | if (isScaledConstantInRange(Node: N, /*Scale=*/1, RangeMin: 0, RangeMax: 0x100, ScaledConstant&: RHSC)) { // 8 bits. |
1397 | OffImm = ((AM == ISD::PRE_INC) || (AM == ISD::POST_INC)) |
1398 | ? CurDAG->getSignedTargetConstant(Val: RHSC, DL: SDLoc(N), VT: MVT::i32) |
1399 | : CurDAG->getSignedTargetConstant(Val: -RHSC, DL: SDLoc(N), VT: MVT::i32); |
1400 | return true; |
1401 | } |
1402 | |
1403 | return false; |
1404 | } |
1405 | |
1406 | template <unsigned Shift> |
1407 | bool ARMDAGToDAGISel::SelectT2AddrModeImm7(SDValue N, SDValue &Base, |
1408 | SDValue &OffImm) { |
1409 | if (N.getOpcode() == ISD::SUB || CurDAG->isBaseWithConstantOffset(Op: N)) { |
1410 | int RHSC; |
1411 | if (isScaledConstantInRange(Node: N.getOperand(i: 1), Scale: 1 << Shift, RangeMin: -0x7f, RangeMax: 0x80, |
1412 | ScaledConstant&: RHSC)) { |
1413 | Base = N.getOperand(i: 0); |
1414 | if (Base.getOpcode() == ISD::FrameIndex) { |
1415 | int FI = cast<FrameIndexSDNode>(Val&: Base)->getIndex(); |
1416 | Base = CurDAG->getTargetFrameIndex( |
1417 | FI, VT: TLI->getPointerTy(DL: CurDAG->getDataLayout())); |
1418 | } |
1419 | |
1420 | if (N.getOpcode() == ISD::SUB) |
1421 | RHSC = -RHSC; |
1422 | OffImm = CurDAG->getSignedTargetConstant(Val: RHSC * (1 << Shift), DL: SDLoc(N), |
1423 | VT: MVT::i32); |
1424 | return true; |
1425 | } |
1426 | } |
1427 | |
1428 | // Base only. |
1429 | Base = N; |
1430 | OffImm = CurDAG->getTargetConstant(Val: 0, DL: SDLoc(N), VT: MVT::i32); |
1431 | return true; |
1432 | } |
1433 | |
1434 | template <unsigned Shift> |
1435 | bool ARMDAGToDAGISel::SelectT2AddrModeImm7Offset(SDNode *Op, SDValue N, |
1436 | SDValue &OffImm) { |
1437 | return SelectT2AddrModeImm7Offset(Op, N, OffImm, Shift); |
1438 | } |
1439 | |
1440 | bool ARMDAGToDAGISel::SelectT2AddrModeImm7Offset(SDNode *Op, SDValue N, |
1441 | SDValue &OffImm, |
1442 | unsigned Shift) { |
1443 | unsigned Opcode = Op->getOpcode(); |
1444 | ISD::MemIndexedMode AM; |
1445 | switch (Opcode) { |
1446 | case ISD::LOAD: |
1447 | AM = cast<LoadSDNode>(Val: Op)->getAddressingMode(); |
1448 | break; |
1449 | case ISD::STORE: |
1450 | AM = cast<StoreSDNode>(Val: Op)->getAddressingMode(); |
1451 | break; |
1452 | case ISD::MLOAD: |
1453 | AM = cast<MaskedLoadSDNode>(Val: Op)->getAddressingMode(); |
1454 | break; |
1455 | case ISD::MSTORE: |
1456 | AM = cast<MaskedStoreSDNode>(Val: Op)->getAddressingMode(); |
1457 | break; |
1458 | default: |
1459 | llvm_unreachable("Unexpected Opcode for Imm7Offset" ); |
1460 | } |
1461 | |
1462 | int RHSC; |
1463 | // 7 bit constant, shifted by Shift. |
1464 | if (isScaledConstantInRange(Node: N, Scale: 1 << Shift, RangeMin: 0, RangeMax: 0x80, ScaledConstant&: RHSC)) { |
1465 | OffImm = ((AM == ISD::PRE_INC) || (AM == ISD::POST_INC)) |
1466 | ? CurDAG->getSignedTargetConstant(Val: RHSC * (1 << Shift), |
1467 | DL: SDLoc(N), VT: MVT::i32) |
1468 | : CurDAG->getSignedTargetConstant(Val: -RHSC * (1 << Shift), |
1469 | DL: SDLoc(N), VT: MVT::i32); |
1470 | return true; |
1471 | } |
1472 | return false; |
1473 | } |
1474 | |
1475 | template <int Min, int Max> |
1476 | bool ARMDAGToDAGISel::SelectImmediateInRange(SDValue N, SDValue &OffImm) { |
1477 | int Val; |
1478 | if (isScaledConstantInRange(Node: N, Scale: 1, RangeMin: Min, RangeMax: Max, ScaledConstant&: Val)) { |
1479 | OffImm = CurDAG->getSignedTargetConstant(Val, DL: SDLoc(N), VT: MVT::i32); |
1480 | return true; |
1481 | } |
1482 | return false; |
1483 | } |
1484 | |
1485 | bool ARMDAGToDAGISel::SelectT2AddrModeSoReg(SDValue N, |
1486 | SDValue &Base, |
1487 | SDValue &OffReg, SDValue &ShImm) { |
1488 | // (R - imm8) should be handled by t2LDRi8. The rest are handled by t2LDRi12. |
1489 | if (N.getOpcode() != ISD::ADD && !CurDAG->isBaseWithConstantOffset(Op: N)) |
1490 | return false; |
1491 | |
1492 | // Leave (R + imm12) for t2LDRi12, (R - imm8) for t2LDRi8. |
1493 | if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Val: N.getOperand(i: 1))) { |
1494 | int RHSC = (int)RHS->getZExtValue(); |
1495 | if (RHSC >= 0 && RHSC < 0x1000) // 12 bits (unsigned) |
1496 | return false; |
1497 | else if (RHSC < 0 && RHSC >= -255) // 8 bits |
1498 | return false; |
1499 | } |
1500 | |
1501 | // Look for (R + R) or (R + (R << [1,2,3])). |
1502 | unsigned ShAmt = 0; |
1503 | Base = N.getOperand(i: 0); |
1504 | OffReg = N.getOperand(i: 1); |
1505 | |
1506 | // Swap if it is ((R << c) + R). |
1507 | ARM_AM::ShiftOpc ShOpcVal = ARM_AM::getShiftOpcForNode(Opcode: OffReg.getOpcode()); |
1508 | if (ShOpcVal != ARM_AM::lsl) { |
1509 | ShOpcVal = ARM_AM::getShiftOpcForNode(Opcode: Base.getOpcode()); |
1510 | if (ShOpcVal == ARM_AM::lsl) |
1511 | std::swap(a&: Base, b&: OffReg); |
1512 | } |
1513 | |
1514 | if (ShOpcVal == ARM_AM::lsl) { |
1515 | // Check to see if the RHS of the shift is a constant, if not, we can't fold |
1516 | // it. |
1517 | if (ConstantSDNode *Sh = dyn_cast<ConstantSDNode>(Val: OffReg.getOperand(i: 1))) { |
1518 | ShAmt = Sh->getZExtValue(); |
1519 | if (ShAmt < 4 && isShifterOpProfitable(Shift: OffReg, ShOpcVal, ShAmt)) |
1520 | OffReg = OffReg.getOperand(i: 0); |
1521 | else { |
1522 | ShAmt = 0; |
1523 | } |
1524 | } |
1525 | } |
1526 | |
1527 | // If OffReg is a multiply-by-constant and it's profitable to extract a shift |
1528 | // and use it in a shifted operand do so. |
1529 | if (OffReg.getOpcode() == ISD::MUL && N.hasOneUse()) { |
1530 | unsigned PowerOfTwo = 0; |
1531 | SDValue NewMulConst; |
1532 | if (canExtractShiftFromMul(N: OffReg, MaxShift: 3, PowerOfTwo, NewMulConst)) { |
1533 | HandleSDNode Handle(OffReg); |
1534 | replaceDAGValue(N: OffReg.getOperand(i: 1), M: NewMulConst); |
1535 | OffReg = Handle.getValue(); |
1536 | ShAmt = PowerOfTwo; |
1537 | } |
1538 | } |
1539 | |
1540 | ShImm = CurDAG->getTargetConstant(Val: ShAmt, DL: SDLoc(N), VT: MVT::i32); |
1541 | |
1542 | return true; |
1543 | } |
1544 | |
1545 | bool ARMDAGToDAGISel::SelectT2AddrModeExclusive(SDValue N, SDValue &Base, |
1546 | SDValue &OffImm) { |
1547 | // This *must* succeed since it's used for the irreplaceable ldrex and strex |
1548 | // instructions. |
1549 | Base = N; |
1550 | OffImm = CurDAG->getTargetConstant(Val: 0, DL: SDLoc(N), VT: MVT::i32); |
1551 | |
1552 | if (N.getOpcode() != ISD::ADD || !CurDAG->isBaseWithConstantOffset(Op: N)) |
1553 | return true; |
1554 | |
1555 | ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Val: N.getOperand(i: 1)); |
1556 | if (!RHS) |
1557 | return true; |
1558 | |
1559 | uint32_t RHSC = (int)RHS->getZExtValue(); |
1560 | if (RHSC > 1020 || RHSC % 4 != 0) |
1561 | return true; |
1562 | |
1563 | Base = N.getOperand(i: 0); |
1564 | if (Base.getOpcode() == ISD::FrameIndex) { |
1565 | int FI = cast<FrameIndexSDNode>(Val&: Base)->getIndex(); |
1566 | Base = CurDAG->getTargetFrameIndex( |
1567 | FI, VT: TLI->getPointerTy(DL: CurDAG->getDataLayout())); |
1568 | } |
1569 | |
1570 | OffImm = CurDAG->getTargetConstant(Val: RHSC/4, DL: SDLoc(N), VT: MVT::i32); |
1571 | return true; |
1572 | } |
1573 | |
1574 | //===--------------------------------------------------------------------===// |
1575 | |
1576 | /// getAL - Returns a ARMCC::AL immediate node. |
1577 | static inline SDValue getAL(SelectionDAG *CurDAG, const SDLoc &dl) { |
1578 | return CurDAG->getTargetConstant(Val: (uint64_t)ARMCC::AL, DL: dl, VT: MVT::i32); |
1579 | } |
1580 | |
1581 | void ARMDAGToDAGISel::transferMemOperands(SDNode *N, SDNode *Result) { |
1582 | MachineMemOperand *MemOp = cast<MemSDNode>(Val: N)->getMemOperand(); |
1583 | CurDAG->setNodeMemRefs(N: cast<MachineSDNode>(Val: Result), NewMemRefs: {MemOp}); |
1584 | } |
1585 | |
1586 | bool ARMDAGToDAGISel::tryARMIndexedLoad(SDNode *N) { |
1587 | LoadSDNode *LD = cast<LoadSDNode>(Val: N); |
1588 | ISD::MemIndexedMode AM = LD->getAddressingMode(); |
1589 | if (AM == ISD::UNINDEXED) |
1590 | return false; |
1591 | |
1592 | EVT LoadedVT = LD->getMemoryVT(); |
1593 | SDValue Offset, AMOpc; |
1594 | bool isPre = (AM == ISD::PRE_INC) || (AM == ISD::PRE_DEC); |
1595 | unsigned Opcode = 0; |
1596 | bool Match = false; |
1597 | if (LoadedVT == MVT::i32 && isPre && |
1598 | SelectAddrMode2OffsetImmPre(Op: N, N: LD->getOffset(), Offset, Opc&: AMOpc)) { |
1599 | Opcode = ARM::LDR_PRE_IMM; |
1600 | Match = true; |
1601 | } else if (LoadedVT == MVT::i32 && !isPre && |
1602 | SelectAddrMode2OffsetImm(Op: N, N: LD->getOffset(), Offset, Opc&: AMOpc)) { |
1603 | Opcode = ARM::LDR_POST_IMM; |
1604 | Match = true; |
1605 | } else if (LoadedVT == MVT::i32 && |
1606 | SelectAddrMode2OffsetReg(Op: N, N: LD->getOffset(), Offset, Opc&: AMOpc)) { |
1607 | Opcode = isPre ? ARM::LDR_PRE_REG : ARM::LDR_POST_REG; |
1608 | Match = true; |
1609 | |
1610 | } else if (LoadedVT == MVT::i16 && |
1611 | SelectAddrMode3Offset(Op: N, N: LD->getOffset(), Offset, Opc&: AMOpc)) { |
1612 | Match = true; |
1613 | Opcode = (LD->getExtensionType() == ISD::SEXTLOAD) |
1614 | ? (isPre ? ARM::LDRSH_PRE : ARM::LDRSH_POST) |
1615 | : (isPre ? ARM::LDRH_PRE : ARM::LDRH_POST); |
1616 | } else if (LoadedVT == MVT::i8 || LoadedVT == MVT::i1) { |
1617 | if (LD->getExtensionType() == ISD::SEXTLOAD) { |
1618 | if (SelectAddrMode3Offset(Op: N, N: LD->getOffset(), Offset, Opc&: AMOpc)) { |
1619 | Match = true; |
1620 | Opcode = isPre ? ARM::LDRSB_PRE : ARM::LDRSB_POST; |
1621 | } |
1622 | } else { |
1623 | if (isPre && |
1624 | SelectAddrMode2OffsetImmPre(Op: N, N: LD->getOffset(), Offset, Opc&: AMOpc)) { |
1625 | Match = true; |
1626 | Opcode = ARM::LDRB_PRE_IMM; |
1627 | } else if (!isPre && |
1628 | SelectAddrMode2OffsetImm(Op: N, N: LD->getOffset(), Offset, Opc&: AMOpc)) { |
1629 | Match = true; |
1630 | Opcode = ARM::LDRB_POST_IMM; |
1631 | } else if (SelectAddrMode2OffsetReg(Op: N, N: LD->getOffset(), Offset, Opc&: AMOpc)) { |
1632 | Match = true; |
1633 | Opcode = isPre ? ARM::LDRB_PRE_REG : ARM::LDRB_POST_REG; |
1634 | } |
1635 | } |
1636 | } |
1637 | |
1638 | if (Match) { |
1639 | if (Opcode == ARM::LDR_PRE_IMM || Opcode == ARM::LDRB_PRE_IMM) { |
1640 | SDValue Chain = LD->getChain(); |
1641 | SDValue Base = LD->getBasePtr(); |
1642 | SDValue Ops[]= { Base, AMOpc, getAL(CurDAG, dl: SDLoc(N)), |
1643 | CurDAG->getRegister(Reg: 0, VT: MVT::i32), Chain }; |
1644 | SDNode *New = CurDAG->getMachineNode(Opcode, dl: SDLoc(N), VT1: MVT::i32, VT2: MVT::i32, |
1645 | VT3: MVT::Other, Ops); |
1646 | transferMemOperands(N, Result: New); |
1647 | ReplaceNode(F: N, T: New); |
1648 | return true; |
1649 | } else { |
1650 | SDValue Chain = LD->getChain(); |
1651 | SDValue Base = LD->getBasePtr(); |
1652 | SDValue Ops[]= { Base, Offset, AMOpc, getAL(CurDAG, dl: SDLoc(N)), |
1653 | CurDAG->getRegister(Reg: 0, VT: MVT::i32), Chain }; |
1654 | SDNode *New = CurDAG->getMachineNode(Opcode, dl: SDLoc(N), VT1: MVT::i32, VT2: MVT::i32, |
1655 | VT3: MVT::Other, Ops); |
1656 | transferMemOperands(N, Result: New); |
1657 | ReplaceNode(F: N, T: New); |
1658 | return true; |
1659 | } |
1660 | } |
1661 | |
1662 | return false; |
1663 | } |
1664 | |
1665 | bool ARMDAGToDAGISel::tryT1IndexedLoad(SDNode *N) { |
1666 | LoadSDNode *LD = cast<LoadSDNode>(Val: N); |
1667 | EVT LoadedVT = LD->getMemoryVT(); |
1668 | ISD::MemIndexedMode AM = LD->getAddressingMode(); |
1669 | if (AM != ISD::POST_INC || LD->getExtensionType() != ISD::NON_EXTLOAD || |
1670 | LoadedVT.getSimpleVT().SimpleTy != MVT::i32) |
1671 | return false; |
1672 | |
1673 | auto *COffs = dyn_cast<ConstantSDNode>(Val: LD->getOffset()); |
1674 | if (!COffs || COffs->getZExtValue() != 4) |
1675 | return false; |
1676 | |
1677 | // A T1 post-indexed load is just a single register LDM: LDM r0!, {r1}. |
1678 | // The encoding of LDM is not how the rest of ISel expects a post-inc load to |
1679 | // look however, so we use a pseudo here and switch it for a tLDMIA_UPD after |
1680 | // ISel. |
1681 | SDValue Chain = LD->getChain(); |
1682 | SDValue Base = LD->getBasePtr(); |
1683 | SDValue Ops[]= { Base, getAL(CurDAG, dl: SDLoc(N)), |
1684 | CurDAG->getRegister(Reg: 0, VT: MVT::i32), Chain }; |
1685 | SDNode *New = CurDAG->getMachineNode(Opcode: ARM::tLDR_postidx, dl: SDLoc(N), VT1: MVT::i32, |
1686 | VT2: MVT::i32, VT3: MVT::Other, Ops); |
1687 | transferMemOperands(N, Result: New); |
1688 | ReplaceNode(F: N, T: New); |
1689 | return true; |
1690 | } |
1691 | |
1692 | bool ARMDAGToDAGISel::tryT2IndexedLoad(SDNode *N) { |
1693 | LoadSDNode *LD = cast<LoadSDNode>(Val: N); |
1694 | ISD::MemIndexedMode AM = LD->getAddressingMode(); |
1695 | if (AM == ISD::UNINDEXED) |
1696 | return false; |
1697 | |
1698 | EVT LoadedVT = LD->getMemoryVT(); |
1699 | bool isSExtLd = LD->getExtensionType() == ISD::SEXTLOAD; |
1700 | SDValue Offset; |
1701 | bool isPre = (AM == ISD::PRE_INC) || (AM == ISD::PRE_DEC); |
1702 | unsigned Opcode = 0; |
1703 | bool Match = false; |
1704 | if (SelectT2AddrModeImm8Offset(Op: N, N: LD->getOffset(), OffImm&: Offset)) { |
1705 | switch (LoadedVT.getSimpleVT().SimpleTy) { |
1706 | case MVT::i32: |
1707 | Opcode = isPre ? ARM::t2LDR_PRE : ARM::t2LDR_POST; |
1708 | break; |
1709 | case MVT::i16: |
1710 | if (isSExtLd) |
1711 | Opcode = isPre ? ARM::t2LDRSH_PRE : ARM::t2LDRSH_POST; |
1712 | else |
1713 | Opcode = isPre ? ARM::t2LDRH_PRE : ARM::t2LDRH_POST; |
1714 | break; |
1715 | case MVT::i8: |
1716 | case MVT::i1: |
1717 | if (isSExtLd) |
1718 | Opcode = isPre ? ARM::t2LDRSB_PRE : ARM::t2LDRSB_POST; |
1719 | else |
1720 | Opcode = isPre ? ARM::t2LDRB_PRE : ARM::t2LDRB_POST; |
1721 | break; |
1722 | default: |
1723 | return false; |
1724 | } |
1725 | Match = true; |
1726 | } |
1727 | |
1728 | if (Match) { |
1729 | SDValue Chain = LD->getChain(); |
1730 | SDValue Base = LD->getBasePtr(); |
1731 | SDValue Ops[]= { Base, Offset, getAL(CurDAG, dl: SDLoc(N)), |
1732 | CurDAG->getRegister(Reg: 0, VT: MVT::i32), Chain }; |
1733 | SDNode *New = CurDAG->getMachineNode(Opcode, dl: SDLoc(N), VT1: MVT::i32, VT2: MVT::i32, |
1734 | VT3: MVT::Other, Ops); |
1735 | transferMemOperands(N, Result: New); |
1736 | ReplaceNode(F: N, T: New); |
1737 | return true; |
1738 | } |
1739 | |
1740 | return false; |
1741 | } |
1742 | |
1743 | bool ARMDAGToDAGISel::tryMVEIndexedLoad(SDNode *N) { |
1744 | EVT LoadedVT; |
1745 | unsigned Opcode = 0; |
1746 | bool isSExtLd, isPre; |
1747 | Align Alignment; |
1748 | ARMVCC::VPTCodes Pred; |
1749 | SDValue PredReg; |
1750 | SDValue Chain, Base, Offset; |
1751 | |
1752 | if (LoadSDNode *LD = dyn_cast<LoadSDNode>(Val: N)) { |
1753 | ISD::MemIndexedMode AM = LD->getAddressingMode(); |
1754 | if (AM == ISD::UNINDEXED) |
1755 | return false; |
1756 | LoadedVT = LD->getMemoryVT(); |
1757 | if (!LoadedVT.isVector()) |
1758 | return false; |
1759 | |
1760 | Chain = LD->getChain(); |
1761 | Base = LD->getBasePtr(); |
1762 | Offset = LD->getOffset(); |
1763 | Alignment = LD->getAlign(); |
1764 | isSExtLd = LD->getExtensionType() == ISD::SEXTLOAD; |
1765 | isPre = (AM == ISD::PRE_INC) || (AM == ISD::PRE_DEC); |
1766 | Pred = ARMVCC::None; |
1767 | PredReg = CurDAG->getRegister(Reg: 0, VT: MVT::i32); |
1768 | } else if (MaskedLoadSDNode *LD = dyn_cast<MaskedLoadSDNode>(Val: N)) { |
1769 | ISD::MemIndexedMode AM = LD->getAddressingMode(); |
1770 | if (AM == ISD::UNINDEXED) |
1771 | return false; |
1772 | LoadedVT = LD->getMemoryVT(); |
1773 | if (!LoadedVT.isVector()) |
1774 | return false; |
1775 | |
1776 | Chain = LD->getChain(); |
1777 | Base = LD->getBasePtr(); |
1778 | Offset = LD->getOffset(); |
1779 | Alignment = LD->getAlign(); |
1780 | isSExtLd = LD->getExtensionType() == ISD::SEXTLOAD; |
1781 | isPre = (AM == ISD::PRE_INC) || (AM == ISD::PRE_DEC); |
1782 | Pred = ARMVCC::Then; |
1783 | PredReg = LD->getMask(); |
1784 | } else |
1785 | llvm_unreachable("Expected a Load or a Masked Load!" ); |
1786 | |
1787 | // We allow LE non-masked loads to change the type (for example use a vldrb.8 |
1788 | // as opposed to a vldrw.32). This can allow extra addressing modes or |
1789 | // alignments for what is otherwise an equivalent instruction. |
1790 | bool CanChangeType = Subtarget->isLittle() && !isa<MaskedLoadSDNode>(Val: N); |
1791 | |
1792 | SDValue NewOffset; |
1793 | if (Alignment >= Align(2) && LoadedVT == MVT::v4i16 && |
1794 | SelectT2AddrModeImm7Offset(Op: N, N: Offset, OffImm&: NewOffset, Shift: 1)) { |
1795 | if (isSExtLd) |
1796 | Opcode = isPre ? ARM::MVE_VLDRHS32_pre : ARM::MVE_VLDRHS32_post; |
1797 | else |
1798 | Opcode = isPre ? ARM::MVE_VLDRHU32_pre : ARM::MVE_VLDRHU32_post; |
1799 | } else if (LoadedVT == MVT::v8i8 && |
1800 | SelectT2AddrModeImm7Offset(Op: N, N: Offset, OffImm&: NewOffset, Shift: 0)) { |
1801 | if (isSExtLd) |
1802 | Opcode = isPre ? ARM::MVE_VLDRBS16_pre : ARM::MVE_VLDRBS16_post; |
1803 | else |
1804 | Opcode = isPre ? ARM::MVE_VLDRBU16_pre : ARM::MVE_VLDRBU16_post; |
1805 | } else if (LoadedVT == MVT::v4i8 && |
1806 | SelectT2AddrModeImm7Offset(Op: N, N: Offset, OffImm&: NewOffset, Shift: 0)) { |
1807 | if (isSExtLd) |
1808 | Opcode = isPre ? ARM::MVE_VLDRBS32_pre : ARM::MVE_VLDRBS32_post; |
1809 | else |
1810 | Opcode = isPre ? ARM::MVE_VLDRBU32_pre : ARM::MVE_VLDRBU32_post; |
1811 | } else if (Alignment >= Align(4) && |
1812 | (CanChangeType || LoadedVT == MVT::v4i32 || |
1813 | LoadedVT == MVT::v4f32) && |
1814 | SelectT2AddrModeImm7Offset(Op: N, N: Offset, OffImm&: NewOffset, Shift: 2)) |
1815 | Opcode = isPre ? ARM::MVE_VLDRWU32_pre : ARM::MVE_VLDRWU32_post; |
1816 | else if (Alignment >= Align(2) && |
1817 | (CanChangeType || LoadedVT == MVT::v8i16 || |
1818 | LoadedVT == MVT::v8f16) && |
1819 | SelectT2AddrModeImm7Offset(Op: N, N: Offset, OffImm&: NewOffset, Shift: 1)) |
1820 | Opcode = isPre ? ARM::MVE_VLDRHU16_pre : ARM::MVE_VLDRHU16_post; |
1821 | else if ((CanChangeType || LoadedVT == MVT::v16i8) && |
1822 | SelectT2AddrModeImm7Offset(Op: N, N: Offset, OffImm&: NewOffset, Shift: 0)) |
1823 | Opcode = isPre ? ARM::MVE_VLDRBU8_pre : ARM::MVE_VLDRBU8_post; |
1824 | else |
1825 | return false; |
1826 | |
1827 | SDValue Ops[] = {Base, |
1828 | NewOffset, |
1829 | CurDAG->getTargetConstant(Val: Pred, DL: SDLoc(N), VT: MVT::i32), |
1830 | PredReg, |
1831 | CurDAG->getRegister(Reg: 0, VT: MVT::i32), // tp_reg |
1832 | Chain}; |
1833 | SDNode *New = CurDAG->getMachineNode(Opcode, dl: SDLoc(N), VT1: MVT::i32, |
1834 | VT2: N->getValueType(ResNo: 0), VT3: MVT::Other, Ops); |
1835 | transferMemOperands(N, Result: New); |
1836 | ReplaceUses(F: SDValue(N, 0), T: SDValue(New, 1)); |
1837 | ReplaceUses(F: SDValue(N, 1), T: SDValue(New, 0)); |
1838 | ReplaceUses(F: SDValue(N, 2), T: SDValue(New, 2)); |
1839 | CurDAG->RemoveDeadNode(N); |
1840 | return true; |
1841 | } |
1842 | |
1843 | /// Form a GPRPair pseudo register from a pair of GPR regs. |
1844 | SDNode *ARMDAGToDAGISel::createGPRPairNode(EVT VT, SDValue V0, SDValue V1) { |
1845 | SDLoc dl(V0.getNode()); |
1846 | SDValue RegClass = |
1847 | CurDAG->getTargetConstant(Val: ARM::GPRPairRegClassID, DL: dl, VT: MVT::i32); |
1848 | SDValue SubReg0 = CurDAG->getTargetConstant(Val: ARM::gsub_0, DL: dl, VT: MVT::i32); |
1849 | SDValue SubReg1 = CurDAG->getTargetConstant(Val: ARM::gsub_1, DL: dl, VT: MVT::i32); |
1850 | const SDValue Ops[] = { RegClass, V0, SubReg0, V1, SubReg1 }; |
1851 | return CurDAG->getMachineNode(Opcode: TargetOpcode::REG_SEQUENCE, dl, VT, Ops); |
1852 | } |
1853 | |
1854 | /// Form a D register from a pair of S registers. |
1855 | SDNode *ARMDAGToDAGISel::createSRegPairNode(EVT VT, SDValue V0, SDValue V1) { |
1856 | SDLoc dl(V0.getNode()); |
1857 | SDValue RegClass = |
1858 | CurDAG->getTargetConstant(Val: ARM::DPR_VFP2RegClassID, DL: dl, VT: MVT::i32); |
1859 | SDValue SubReg0 = CurDAG->getTargetConstant(Val: ARM::ssub_0, DL: dl, VT: MVT::i32); |
1860 | SDValue SubReg1 = CurDAG->getTargetConstant(Val: ARM::ssub_1, DL: dl, VT: MVT::i32); |
1861 | const SDValue Ops[] = { RegClass, V0, SubReg0, V1, SubReg1 }; |
1862 | return CurDAG->getMachineNode(Opcode: TargetOpcode::REG_SEQUENCE, dl, VT, Ops); |
1863 | } |
1864 | |
1865 | /// Form a quad register from a pair of D registers. |
1866 | SDNode *ARMDAGToDAGISel::createDRegPairNode(EVT VT, SDValue V0, SDValue V1) { |
1867 | SDLoc dl(V0.getNode()); |
1868 | SDValue RegClass = CurDAG->getTargetConstant(Val: ARM::QPRRegClassID, DL: dl, |
1869 | VT: MVT::i32); |
1870 | SDValue SubReg0 = CurDAG->getTargetConstant(Val: ARM::dsub_0, DL: dl, VT: MVT::i32); |
1871 | SDValue SubReg1 = CurDAG->getTargetConstant(Val: ARM::dsub_1, DL: dl, VT: MVT::i32); |
1872 | const SDValue Ops[] = { RegClass, V0, SubReg0, V1, SubReg1 }; |
1873 | return CurDAG->getMachineNode(Opcode: TargetOpcode::REG_SEQUENCE, dl, VT, Ops); |
1874 | } |
1875 | |
1876 | /// Form 4 consecutive D registers from a pair of Q registers. |
1877 | SDNode *ARMDAGToDAGISel::createQRegPairNode(EVT VT, SDValue V0, SDValue V1) { |
1878 | SDLoc dl(V0.getNode()); |
1879 | SDValue RegClass = CurDAG->getTargetConstant(Val: ARM::QQPRRegClassID, DL: dl, |
1880 | VT: MVT::i32); |
1881 | SDValue SubReg0 = CurDAG->getTargetConstant(Val: ARM::qsub_0, DL: dl, VT: MVT::i32); |
1882 | SDValue SubReg1 = CurDAG->getTargetConstant(Val: ARM::qsub_1, DL: dl, VT: MVT::i32); |
1883 | const SDValue Ops[] = { RegClass, V0, SubReg0, V1, SubReg1 }; |
1884 | return CurDAG->getMachineNode(Opcode: TargetOpcode::REG_SEQUENCE, dl, VT, Ops); |
1885 | } |
1886 | |
1887 | /// Form 4 consecutive S registers. |
1888 | SDNode *ARMDAGToDAGISel::createQuadSRegsNode(EVT VT, SDValue V0, SDValue V1, |
1889 | SDValue V2, SDValue V3) { |
1890 | SDLoc dl(V0.getNode()); |
1891 | SDValue RegClass = |
1892 | CurDAG->getTargetConstant(Val: ARM::QPR_VFP2RegClassID, DL: dl, VT: MVT::i32); |
1893 | SDValue SubReg0 = CurDAG->getTargetConstant(Val: ARM::ssub_0, DL: dl, VT: MVT::i32); |
1894 | SDValue SubReg1 = CurDAG->getTargetConstant(Val: ARM::ssub_1, DL: dl, VT: MVT::i32); |
1895 | SDValue SubReg2 = CurDAG->getTargetConstant(Val: ARM::ssub_2, DL: dl, VT: MVT::i32); |
1896 | SDValue SubReg3 = CurDAG->getTargetConstant(Val: ARM::ssub_3, DL: dl, VT: MVT::i32); |
1897 | const SDValue Ops[] = { RegClass, V0, SubReg0, V1, SubReg1, |
1898 | V2, SubReg2, V3, SubReg3 }; |
1899 | return CurDAG->getMachineNode(Opcode: TargetOpcode::REG_SEQUENCE, dl, VT, Ops); |
1900 | } |
1901 | |
1902 | /// Form 4 consecutive D registers. |
1903 | SDNode *ARMDAGToDAGISel::createQuadDRegsNode(EVT VT, SDValue V0, SDValue V1, |
1904 | SDValue V2, SDValue V3) { |
1905 | SDLoc dl(V0.getNode()); |
1906 | SDValue RegClass = CurDAG->getTargetConstant(Val: ARM::QQPRRegClassID, DL: dl, |
1907 | VT: MVT::i32); |
1908 | SDValue SubReg0 = CurDAG->getTargetConstant(Val: ARM::dsub_0, DL: dl, VT: MVT::i32); |
1909 | SDValue SubReg1 = CurDAG->getTargetConstant(Val: ARM::dsub_1, DL: dl, VT: MVT::i32); |
1910 | SDValue SubReg2 = CurDAG->getTargetConstant(Val: ARM::dsub_2, DL: dl, VT: MVT::i32); |
1911 | SDValue SubReg3 = CurDAG->getTargetConstant(Val: ARM::dsub_3, DL: dl, VT: MVT::i32); |
1912 | const SDValue Ops[] = { RegClass, V0, SubReg0, V1, SubReg1, |
1913 | V2, SubReg2, V3, SubReg3 }; |
1914 | return CurDAG->getMachineNode(Opcode: TargetOpcode::REG_SEQUENCE, dl, VT, Ops); |
1915 | } |
1916 | |
1917 | /// Form 4 consecutive Q registers. |
1918 | SDNode *ARMDAGToDAGISel::createQuadQRegsNode(EVT VT, SDValue V0, SDValue V1, |
1919 | SDValue V2, SDValue V3) { |
1920 | SDLoc dl(V0.getNode()); |
1921 | SDValue RegClass = CurDAG->getTargetConstant(Val: ARM::QQQQPRRegClassID, DL: dl, |
1922 | VT: MVT::i32); |
1923 | SDValue SubReg0 = CurDAG->getTargetConstant(Val: ARM::qsub_0, DL: dl, VT: MVT::i32); |
1924 | SDValue SubReg1 = CurDAG->getTargetConstant(Val: ARM::qsub_1, DL: dl, VT: MVT::i32); |
1925 | SDValue SubReg2 = CurDAG->getTargetConstant(Val: ARM::qsub_2, DL: dl, VT: MVT::i32); |
1926 | SDValue SubReg3 = CurDAG->getTargetConstant(Val: ARM::qsub_3, DL: dl, VT: MVT::i32); |
1927 | const SDValue Ops[] = { RegClass, V0, SubReg0, V1, SubReg1, |
1928 | V2, SubReg2, V3, SubReg3 }; |
1929 | return CurDAG->getMachineNode(Opcode: TargetOpcode::REG_SEQUENCE, dl, VT, Ops); |
1930 | } |
1931 | |
1932 | /// GetVLDSTAlign - Get the alignment (in bytes) for the alignment operand |
1933 | /// of a NEON VLD or VST instruction. The supported values depend on the |
1934 | /// number of registers being loaded. |
1935 | SDValue ARMDAGToDAGISel::GetVLDSTAlign(SDValue Align, const SDLoc &dl, |
1936 | unsigned NumVecs, bool is64BitVector) { |
1937 | unsigned NumRegs = NumVecs; |
1938 | if (!is64BitVector && NumVecs < 3) |
1939 | NumRegs *= 2; |
1940 | |
1941 | unsigned Alignment = Align->getAsZExtVal(); |
1942 | if (Alignment >= 32 && NumRegs == 4) |
1943 | Alignment = 32; |
1944 | else if (Alignment >= 16 && (NumRegs == 2 || NumRegs == 4)) |
1945 | Alignment = 16; |
1946 | else if (Alignment >= 8) |
1947 | Alignment = 8; |
1948 | else |
1949 | Alignment = 0; |
1950 | |
1951 | return CurDAG->getTargetConstant(Val: Alignment, DL: dl, VT: MVT::i32); |
1952 | } |
1953 | |
1954 | static bool isVLDfixed(unsigned Opc) |
1955 | { |
1956 | switch (Opc) { |
1957 | default: return false; |
1958 | case ARM::VLD1d8wb_fixed : return true; |
1959 | case ARM::VLD1d16wb_fixed : return true; |
1960 | case ARM::VLD1d64Qwb_fixed : return true; |
1961 | case ARM::VLD1d32wb_fixed : return true; |
1962 | case ARM::VLD1d64wb_fixed : return true; |
1963 | case ARM::VLD1d8TPseudoWB_fixed : return true; |
1964 | case ARM::VLD1d16TPseudoWB_fixed : return true; |
1965 | case ARM::VLD1d32TPseudoWB_fixed : return true; |
1966 | case ARM::VLD1d64TPseudoWB_fixed : return true; |
1967 | case ARM::VLD1d8QPseudoWB_fixed : return true; |
1968 | case ARM::VLD1d16QPseudoWB_fixed : return true; |
1969 | case ARM::VLD1d32QPseudoWB_fixed : return true; |
1970 | case ARM::VLD1d64QPseudoWB_fixed : return true; |
1971 | case ARM::VLD1q8wb_fixed : return true; |
1972 | case ARM::VLD1q16wb_fixed : return true; |
1973 | case ARM::VLD1q32wb_fixed : return true; |
1974 | case ARM::VLD1q64wb_fixed : return true; |
1975 | case ARM::VLD1DUPd8wb_fixed : return true; |
1976 | case ARM::VLD1DUPd16wb_fixed : return true; |
1977 | case ARM::VLD1DUPd32wb_fixed : return true; |
1978 | case ARM::VLD1DUPq8wb_fixed : return true; |
1979 | case ARM::VLD1DUPq16wb_fixed : return true; |
1980 | case ARM::VLD1DUPq32wb_fixed : return true; |
1981 | case ARM::VLD2d8wb_fixed : return true; |
1982 | case ARM::VLD2d16wb_fixed : return true; |
1983 | case ARM::VLD2d32wb_fixed : return true; |
1984 | case ARM::VLD2q8PseudoWB_fixed : return true; |
1985 | case ARM::VLD2q16PseudoWB_fixed : return true; |
1986 | case ARM::VLD2q32PseudoWB_fixed : return true; |
1987 | case ARM::VLD2DUPd8wb_fixed : return true; |
1988 | case ARM::VLD2DUPd16wb_fixed : return true; |
1989 | case ARM::VLD2DUPd32wb_fixed : return true; |
1990 | case ARM::VLD2DUPq8OddPseudoWB_fixed: return true; |
1991 | case ARM::VLD2DUPq16OddPseudoWB_fixed: return true; |
1992 | case ARM::VLD2DUPq32OddPseudoWB_fixed: return true; |
1993 | } |
1994 | } |
1995 | |
1996 | static bool isVSTfixed(unsigned Opc) |
1997 | { |
1998 | switch (Opc) { |
1999 | default: return false; |
2000 | case ARM::VST1d8wb_fixed : return true; |
2001 | case ARM::VST1d16wb_fixed : return true; |
2002 | case ARM::VST1d32wb_fixed : return true; |
2003 | case ARM::VST1d64wb_fixed : return true; |
2004 | case ARM::VST1q8wb_fixed : return true; |
2005 | case ARM::VST1q16wb_fixed : return true; |
2006 | case ARM::VST1q32wb_fixed : return true; |
2007 | case ARM::VST1q64wb_fixed : return true; |
2008 | case ARM::VST1d8TPseudoWB_fixed : return true; |
2009 | case ARM::VST1d16TPseudoWB_fixed : return true; |
2010 | case ARM::VST1d32TPseudoWB_fixed : return true; |
2011 | case ARM::VST1d64TPseudoWB_fixed : return true; |
2012 | case ARM::VST1d8QPseudoWB_fixed : return true; |
2013 | case ARM::VST1d16QPseudoWB_fixed : return true; |
2014 | case ARM::VST1d32QPseudoWB_fixed : return true; |
2015 | case ARM::VST1d64QPseudoWB_fixed : return true; |
2016 | case ARM::VST2d8wb_fixed : return true; |
2017 | case ARM::VST2d16wb_fixed : return true; |
2018 | case ARM::VST2d32wb_fixed : return true; |
2019 | case ARM::VST2q8PseudoWB_fixed : return true; |
2020 | case ARM::VST2q16PseudoWB_fixed : return true; |
2021 | case ARM::VST2q32PseudoWB_fixed : return true; |
2022 | } |
2023 | } |
2024 | |
2025 | // Get the register stride update opcode of a VLD/VST instruction that |
2026 | // is otherwise equivalent to the given fixed stride updating instruction. |
2027 | static unsigned getVLDSTRegisterUpdateOpcode(unsigned Opc) { |
2028 | assert((isVLDfixed(Opc) || isVSTfixed(Opc)) |
2029 | && "Incorrect fixed stride updating instruction." ); |
2030 | switch (Opc) { |
2031 | default: break; |
2032 | case ARM::VLD1d8wb_fixed: return ARM::VLD1d8wb_register; |
2033 | case ARM::VLD1d16wb_fixed: return ARM::VLD1d16wb_register; |
2034 | case ARM::VLD1d32wb_fixed: return ARM::VLD1d32wb_register; |
2035 | case ARM::VLD1d64wb_fixed: return ARM::VLD1d64wb_register; |
2036 | case ARM::VLD1q8wb_fixed: return ARM::VLD1q8wb_register; |
2037 | case ARM::VLD1q16wb_fixed: return ARM::VLD1q16wb_register; |
2038 | case ARM::VLD1q32wb_fixed: return ARM::VLD1q32wb_register; |
2039 | case ARM::VLD1q64wb_fixed: return ARM::VLD1q64wb_register; |
2040 | case ARM::VLD1d64Twb_fixed: return ARM::VLD1d64Twb_register; |
2041 | case ARM::VLD1d64Qwb_fixed: return ARM::VLD1d64Qwb_register; |
2042 | case ARM::VLD1d8TPseudoWB_fixed: return ARM::VLD1d8TPseudoWB_register; |
2043 | case ARM::VLD1d16TPseudoWB_fixed: return ARM::VLD1d16TPseudoWB_register; |
2044 | case ARM::VLD1d32TPseudoWB_fixed: return ARM::VLD1d32TPseudoWB_register; |
2045 | case ARM::VLD1d64TPseudoWB_fixed: return ARM::VLD1d64TPseudoWB_register; |
2046 | case ARM::VLD1d8QPseudoWB_fixed: return ARM::VLD1d8QPseudoWB_register; |
2047 | case ARM::VLD1d16QPseudoWB_fixed: return ARM::VLD1d16QPseudoWB_register; |
2048 | case ARM::VLD1d32QPseudoWB_fixed: return ARM::VLD1d32QPseudoWB_register; |
2049 | case ARM::VLD1d64QPseudoWB_fixed: return ARM::VLD1d64QPseudoWB_register; |
2050 | case ARM::VLD1DUPd8wb_fixed : return ARM::VLD1DUPd8wb_register; |
2051 | case ARM::VLD1DUPd16wb_fixed : return ARM::VLD1DUPd16wb_register; |
2052 | case ARM::VLD1DUPd32wb_fixed : return ARM::VLD1DUPd32wb_register; |
2053 | case ARM::VLD1DUPq8wb_fixed : return ARM::VLD1DUPq8wb_register; |
2054 | case ARM::VLD1DUPq16wb_fixed : return ARM::VLD1DUPq16wb_register; |
2055 | case ARM::VLD1DUPq32wb_fixed : return ARM::VLD1DUPq32wb_register; |
2056 | case ARM::VLD2DUPq8OddPseudoWB_fixed: return ARM::VLD2DUPq8OddPseudoWB_register; |
2057 | case ARM::VLD2DUPq16OddPseudoWB_fixed: return ARM::VLD2DUPq16OddPseudoWB_register; |
2058 | case ARM::VLD2DUPq32OddPseudoWB_fixed: return ARM::VLD2DUPq32OddPseudoWB_register; |
2059 | |
2060 | case ARM::VST1d8wb_fixed: return ARM::VST1d8wb_register; |
2061 | case ARM::VST1d16wb_fixed: return ARM::VST1d16wb_register; |
2062 | case ARM::VST1d32wb_fixed: return ARM::VST1d32wb_register; |
2063 | case ARM::VST1d64wb_fixed: return ARM::VST1d64wb_register; |
2064 | case ARM::VST1q8wb_fixed: return ARM::VST1q8wb_register; |
2065 | case ARM::VST1q16wb_fixed: return ARM::VST1q16wb_register; |
2066 | case ARM::VST1q32wb_fixed: return ARM::VST1q32wb_register; |
2067 | case ARM::VST1q64wb_fixed: return ARM::VST1q64wb_register; |
2068 | case ARM::VST1d8TPseudoWB_fixed: return ARM::VST1d8TPseudoWB_register; |
2069 | case ARM::VST1d16TPseudoWB_fixed: return ARM::VST1d16TPseudoWB_register; |
2070 | case ARM::VST1d32TPseudoWB_fixed: return ARM::VST1d32TPseudoWB_register; |
2071 | case ARM::VST1d64TPseudoWB_fixed: return ARM::VST1d64TPseudoWB_register; |
2072 | case ARM::VST1d8QPseudoWB_fixed: return ARM::VST1d8QPseudoWB_register; |
2073 | case ARM::VST1d16QPseudoWB_fixed: return ARM::VST1d16QPseudoWB_register; |
2074 | case ARM::VST1d32QPseudoWB_fixed: return ARM::VST1d32QPseudoWB_register; |
2075 | case ARM::VST1d64QPseudoWB_fixed: return ARM::VST1d64QPseudoWB_register; |
2076 | |
2077 | case ARM::VLD2d8wb_fixed: return ARM::VLD2d8wb_register; |
2078 | case ARM::VLD2d16wb_fixed: return ARM::VLD2d16wb_register; |
2079 | case ARM::VLD2d32wb_fixed: return ARM::VLD2d32wb_register; |
2080 | case ARM::VLD2q8PseudoWB_fixed: return ARM::VLD2q8PseudoWB_register; |
2081 | case ARM::VLD2q16PseudoWB_fixed: return ARM::VLD2q16PseudoWB_register; |
2082 | case ARM::VLD2q32PseudoWB_fixed: return ARM::VLD2q32PseudoWB_register; |
2083 | |
2084 | case ARM::VST2d8wb_fixed: return ARM::VST2d8wb_register; |
2085 | case ARM::VST2d16wb_fixed: return ARM::VST2d16wb_register; |
2086 | case ARM::VST2d32wb_fixed: return ARM::VST2d32wb_register; |
2087 | case ARM::VST2q8PseudoWB_fixed: return ARM::VST2q8PseudoWB_register; |
2088 | case ARM::VST2q16PseudoWB_fixed: return ARM::VST2q16PseudoWB_register; |
2089 | case ARM::VST2q32PseudoWB_fixed: return ARM::VST2q32PseudoWB_register; |
2090 | |
2091 | case ARM::VLD2DUPd8wb_fixed: return ARM::VLD2DUPd8wb_register; |
2092 | case ARM::VLD2DUPd16wb_fixed: return ARM::VLD2DUPd16wb_register; |
2093 | case ARM::VLD2DUPd32wb_fixed: return ARM::VLD2DUPd32wb_register; |
2094 | } |
2095 | return Opc; // If not one we handle, return it unchanged. |
2096 | } |
2097 | |
2098 | /// Returns true if the given increment is a Constant known to be equal to the |
2099 | /// access size performed by a NEON load/store. This means the "[rN]!" form can |
2100 | /// be used. |
2101 | static bool isPerfectIncrement(SDValue Inc, EVT VecTy, unsigned NumVecs) { |
2102 | auto C = dyn_cast<ConstantSDNode>(Val&: Inc); |
2103 | return C && C->getZExtValue() == VecTy.getSizeInBits() / 8 * NumVecs; |
2104 | } |
2105 | |
2106 | void ARMDAGToDAGISel::SelectVLD(SDNode *N, bool isUpdating, unsigned NumVecs, |
2107 | const uint16_t *DOpcodes, |
2108 | const uint16_t *QOpcodes0, |
2109 | const uint16_t *QOpcodes1) { |
2110 | assert(Subtarget->hasNEON()); |
2111 | assert(NumVecs >= 1 && NumVecs <= 4 && "VLD NumVecs out-of-range" ); |
2112 | SDLoc dl(N); |
2113 | |
2114 | SDValue MemAddr, Align; |
2115 | bool IsIntrinsic = !isUpdating; // By coincidence, all supported updating |
2116 | // nodes are not intrinsics. |
2117 | unsigned AddrOpIdx = IsIntrinsic ? 2 : 1; |
2118 | if (!SelectAddrMode6(Parent: N, N: N->getOperand(Num: AddrOpIdx), Addr&: MemAddr, Align)) |
2119 | return; |
2120 | |
2121 | SDValue Chain = N->getOperand(Num: 0); |
2122 | EVT VT = N->getValueType(ResNo: 0); |
2123 | bool is64BitVector = VT.is64BitVector(); |
2124 | Align = GetVLDSTAlign(Align, dl, NumVecs, is64BitVector); |
2125 | |
2126 | unsigned OpcodeIndex; |
2127 | switch (VT.getSimpleVT().SimpleTy) { |
2128 | default: llvm_unreachable("unhandled vld type" ); |
2129 | // Double-register operations: |
2130 | case MVT::v8i8: OpcodeIndex = 0; break; |
2131 | case MVT::v4f16: |
2132 | case MVT::v4bf16: |
2133 | case MVT::v4i16: OpcodeIndex = 1; break; |
2134 | case MVT::v2f32: |
2135 | case MVT::v2i32: OpcodeIndex = 2; break; |
2136 | case MVT::v1i64: OpcodeIndex = 3; break; |
2137 | // Quad-register operations: |
2138 | case MVT::v16i8: OpcodeIndex = 0; break; |
2139 | case MVT::v8f16: |
2140 | case MVT::v8bf16: |
2141 | case MVT::v8i16: OpcodeIndex = 1; break; |
2142 | case MVT::v4f32: |
2143 | case MVT::v4i32: OpcodeIndex = 2; break; |
2144 | case MVT::v2f64: |
2145 | case MVT::v2i64: OpcodeIndex = 3; break; |
2146 | } |
2147 | |
2148 | EVT ResTy; |
2149 | if (NumVecs == 1) |
2150 | ResTy = VT; |
2151 | else { |
2152 | unsigned ResTyElts = (NumVecs == 3) ? 4 : NumVecs; |
2153 | if (!is64BitVector) |
2154 | ResTyElts *= 2; |
2155 | ResTy = EVT::getVectorVT(Context&: *CurDAG->getContext(), VT: MVT::i64, NumElements: ResTyElts); |
2156 | } |
2157 | std::vector<EVT> ResTys; |
2158 | ResTys.push_back(x: ResTy); |
2159 | if (isUpdating) |
2160 | ResTys.push_back(x: MVT::i32); |
2161 | ResTys.push_back(x: MVT::Other); |
2162 | |
2163 | SDValue Pred = getAL(CurDAG, dl); |
2164 | SDValue Reg0 = CurDAG->getRegister(Reg: 0, VT: MVT::i32); |
2165 | SDNode *VLd; |
2166 | SmallVector<SDValue, 7> Ops; |
2167 | |
2168 | // Double registers and VLD1/VLD2 quad registers are directly supported. |
2169 | if (is64BitVector || NumVecs <= 2) { |
2170 | unsigned Opc = (is64BitVector ? DOpcodes[OpcodeIndex] : |
2171 | QOpcodes0[OpcodeIndex]); |
2172 | Ops.push_back(Elt: MemAddr); |
2173 | Ops.push_back(Elt: Align); |
2174 | if (isUpdating) { |
2175 | SDValue Inc = N->getOperand(Num: AddrOpIdx + 1); |
2176 | bool IsImmUpdate = isPerfectIncrement(Inc, VecTy: VT, NumVecs); |
2177 | if (!IsImmUpdate) { |
2178 | // We use a VLD1 for v1i64 even if the pseudo says vld2/3/4, so |
2179 | // check for the opcode rather than the number of vector elements. |
2180 | if (isVLDfixed(Opc)) |
2181 | Opc = getVLDSTRegisterUpdateOpcode(Opc); |
2182 | Ops.push_back(Elt: Inc); |
2183 | // VLD1/VLD2 fixed increment does not need Reg0 so only include it in |
2184 | // the operands if not such an opcode. |
2185 | } else if (!isVLDfixed(Opc)) |
2186 | Ops.push_back(Elt: Reg0); |
2187 | } |
2188 | Ops.push_back(Elt: Pred); |
2189 | Ops.push_back(Elt: Reg0); |
2190 | Ops.push_back(Elt: Chain); |
2191 | VLd = CurDAG->getMachineNode(Opcode: Opc, dl, ResultTys: ResTys, Ops); |
2192 | |
2193 | } else { |
2194 | // Otherwise, quad registers are loaded with two separate instructions, |
2195 | // where one loads the even registers and the other loads the odd registers. |
2196 | EVT AddrTy = MemAddr.getValueType(); |
2197 | |
2198 | // Load the even subregs. This is always an updating load, so that it |
2199 | // provides the address to the second load for the odd subregs. |
2200 | SDValue ImplDef = |
2201 | SDValue(CurDAG->getMachineNode(Opcode: TargetOpcode::IMPLICIT_DEF, dl, VT: ResTy), 0); |
2202 | const SDValue OpsA[] = { MemAddr, Align, Reg0, ImplDef, Pred, Reg0, Chain }; |
2203 | SDNode *VLdA = CurDAG->getMachineNode(Opcode: QOpcodes0[OpcodeIndex], dl, |
2204 | VT1: ResTy, VT2: AddrTy, VT3: MVT::Other, Ops: OpsA); |
2205 | Chain = SDValue(VLdA, 2); |
2206 | |
2207 | // Load the odd subregs. |
2208 | Ops.push_back(Elt: SDValue(VLdA, 1)); |
2209 | Ops.push_back(Elt: Align); |
2210 | if (isUpdating) { |
2211 | SDValue Inc = N->getOperand(Num: AddrOpIdx + 1); |
2212 | assert(isa<ConstantSDNode>(Inc.getNode()) && |
2213 | "only constant post-increment update allowed for VLD3/4" ); |
2214 | (void)Inc; |
2215 | Ops.push_back(Elt: Reg0); |
2216 | } |
2217 | Ops.push_back(Elt: SDValue(VLdA, 0)); |
2218 | Ops.push_back(Elt: Pred); |
2219 | Ops.push_back(Elt: Reg0); |
2220 | Ops.push_back(Elt: Chain); |
2221 | VLd = CurDAG->getMachineNode(Opcode: QOpcodes1[OpcodeIndex], dl, ResultTys: ResTys, Ops); |
2222 | } |
2223 | |
2224 | // Transfer memoperands. |
2225 | MachineMemOperand *MemOp = cast<MemIntrinsicSDNode>(Val: N)->getMemOperand(); |
2226 | CurDAG->setNodeMemRefs(N: cast<MachineSDNode>(Val: VLd), NewMemRefs: {MemOp}); |
2227 | |
2228 | if (NumVecs == 1) { |
2229 | ReplaceNode(F: N, T: VLd); |
2230 | return; |
2231 | } |
2232 | |
2233 | // Extract out the subregisters. |
2234 | SDValue SuperReg = SDValue(VLd, 0); |
2235 | static_assert(ARM::dsub_7 == ARM::dsub_0 + 7 && |
2236 | ARM::qsub_3 == ARM::qsub_0 + 3, |
2237 | "Unexpected subreg numbering" ); |
2238 | unsigned Sub0 = (is64BitVector ? ARM::dsub_0 : ARM::qsub_0); |
2239 | for (unsigned Vec = 0; Vec < NumVecs; ++Vec) |
2240 | ReplaceUses(F: SDValue(N, Vec), |
2241 | T: CurDAG->getTargetExtractSubreg(SRIdx: Sub0 + Vec, DL: dl, VT, Operand: SuperReg)); |
2242 | ReplaceUses(F: SDValue(N, NumVecs), T: SDValue(VLd, 1)); |
2243 | if (isUpdating) |
2244 | ReplaceUses(F: SDValue(N, NumVecs + 1), T: SDValue(VLd, 2)); |
2245 | CurDAG->RemoveDeadNode(N); |
2246 | } |
2247 | |
2248 | void ARMDAGToDAGISel::SelectVST(SDNode *N, bool isUpdating, unsigned NumVecs, |
2249 | const uint16_t *DOpcodes, |
2250 | const uint16_t *QOpcodes0, |
2251 | const uint16_t *QOpcodes1) { |
2252 | assert(Subtarget->hasNEON()); |
2253 | assert(NumVecs >= 1 && NumVecs <= 4 && "VST NumVecs out-of-range" ); |
2254 | SDLoc dl(N); |
2255 | |
2256 | SDValue MemAddr, Align; |
2257 | bool IsIntrinsic = !isUpdating; // By coincidence, all supported updating |
2258 | // nodes are not intrinsics. |
2259 | unsigned AddrOpIdx = IsIntrinsic ? 2 : 1; |
2260 | unsigned Vec0Idx = 3; // AddrOpIdx + (isUpdating ? 2 : 1) |
2261 | if (!SelectAddrMode6(Parent: N, N: N->getOperand(Num: AddrOpIdx), Addr&: MemAddr, Align)) |
2262 | return; |
2263 | |
2264 | MachineMemOperand *MemOp = cast<MemIntrinsicSDNode>(Val: N)->getMemOperand(); |
2265 | |
2266 | SDValue Chain = N->getOperand(Num: 0); |
2267 | EVT VT = N->getOperand(Num: Vec0Idx).getValueType(); |
2268 | bool is64BitVector = VT.is64BitVector(); |
2269 | Align = GetVLDSTAlign(Align, dl, NumVecs, is64BitVector); |
2270 | |
2271 | unsigned OpcodeIndex; |
2272 | switch (VT.getSimpleVT().SimpleTy) { |
2273 | default: llvm_unreachable("unhandled vst type" ); |
2274 | // Double-register operations: |
2275 | case MVT::v8i8: OpcodeIndex = 0; break; |
2276 | case MVT::v4f16: |
2277 | case MVT::v4bf16: |
2278 | case MVT::v4i16: OpcodeIndex = 1; break; |
2279 | case MVT::v2f32: |
2280 | case MVT::v2i32: OpcodeIndex = 2; break; |
2281 | case MVT::v1i64: OpcodeIndex = 3; break; |
2282 | // Quad-register operations: |
2283 | case MVT::v16i8: OpcodeIndex = 0; break; |
2284 | case MVT::v8f16: |
2285 | case MVT::v8bf16: |
2286 | case MVT::v8i16: OpcodeIndex = 1; break; |
2287 | case MVT::v4f32: |
2288 | case MVT::v4i32: OpcodeIndex = 2; break; |
2289 | case MVT::v2f64: |
2290 | case MVT::v2i64: OpcodeIndex = 3; break; |
2291 | } |
2292 | |
2293 | std::vector<EVT> ResTys; |
2294 | if (isUpdating) |
2295 | ResTys.push_back(x: MVT::i32); |
2296 | ResTys.push_back(x: MVT::Other); |
2297 | |
2298 | SDValue Pred = getAL(CurDAG, dl); |
2299 | SDValue Reg0 = CurDAG->getRegister(Reg: 0, VT: MVT::i32); |
2300 | SmallVector<SDValue, 7> Ops; |
2301 | |
2302 | // Double registers and VST1/VST2 quad registers are directly supported. |
2303 | if (is64BitVector || NumVecs <= 2) { |
2304 | SDValue SrcReg; |
2305 | if (NumVecs == 1) { |
2306 | SrcReg = N->getOperand(Num: Vec0Idx); |
2307 | } else if (is64BitVector) { |
2308 | // Form a REG_SEQUENCE to force register allocation. |
2309 | SDValue V0 = N->getOperand(Num: Vec0Idx + 0); |
2310 | SDValue V1 = N->getOperand(Num: Vec0Idx + 1); |
2311 | if (NumVecs == 2) |
2312 | SrcReg = SDValue(createDRegPairNode(VT: MVT::v2i64, V0, V1), 0); |
2313 | else { |
2314 | SDValue V2 = N->getOperand(Num: Vec0Idx + 2); |
2315 | // If it's a vst3, form a quad D-register and leave the last part as |
2316 | // an undef. |
2317 | SDValue V3 = (NumVecs == 3) |
2318 | ? SDValue(CurDAG->getMachineNode(Opcode: TargetOpcode::IMPLICIT_DEF,dl,VT), 0) |
2319 | : N->getOperand(Num: Vec0Idx + 3); |
2320 | SrcReg = SDValue(createQuadDRegsNode(VT: MVT::v4i64, V0, V1, V2, V3), 0); |
2321 | } |
2322 | } else { |
2323 | // Form a QQ register. |
2324 | SDValue Q0 = N->getOperand(Num: Vec0Idx); |
2325 | SDValue Q1 = N->getOperand(Num: Vec0Idx + 1); |
2326 | SrcReg = SDValue(createQRegPairNode(VT: MVT::v4i64, V0: Q0, V1: Q1), 0); |
2327 | } |
2328 | |
2329 | unsigned Opc = (is64BitVector ? DOpcodes[OpcodeIndex] : |
2330 | QOpcodes0[OpcodeIndex]); |
2331 | Ops.push_back(Elt: MemAddr); |
2332 | Ops.push_back(Elt: Align); |
2333 | if (isUpdating) { |
2334 | SDValue Inc = N->getOperand(Num: AddrOpIdx + 1); |
2335 | bool IsImmUpdate = isPerfectIncrement(Inc, VecTy: VT, NumVecs); |
2336 | if (!IsImmUpdate) { |
2337 | // We use a VST1 for v1i64 even if the pseudo says VST2/3/4, so |
2338 | // check for the opcode rather than the number of vector elements. |
2339 | if (isVSTfixed(Opc)) |
2340 | Opc = getVLDSTRegisterUpdateOpcode(Opc); |
2341 | Ops.push_back(Elt: Inc); |
2342 | } |
2343 | // VST1/VST2 fixed increment does not need Reg0 so only include it in |
2344 | // the operands if not such an opcode. |
2345 | else if (!isVSTfixed(Opc)) |
2346 | Ops.push_back(Elt: Reg0); |
2347 | } |
2348 | Ops.push_back(Elt: SrcReg); |
2349 | Ops.push_back(Elt: Pred); |
2350 | Ops.push_back(Elt: Reg0); |
2351 | Ops.push_back(Elt: Chain); |
2352 | SDNode *VSt = CurDAG->getMachineNode(Opcode: Opc, dl, ResultTys: ResTys, Ops); |
2353 | |
2354 | // Transfer memoperands. |
2355 | CurDAG->setNodeMemRefs(N: cast<MachineSDNode>(Val: VSt), NewMemRefs: {MemOp}); |
2356 | |
2357 | ReplaceNode(F: N, T: VSt); |
2358 | return; |
2359 | } |
2360 | |
2361 | // Otherwise, quad registers are stored with two separate instructions, |
2362 | // where one stores the even registers and the other stores the odd registers. |
2363 | |
2364 | // Form the QQQQ REG_SEQUENCE. |
2365 | SDValue V0 = N->getOperand(Num: Vec0Idx + 0); |
2366 | SDValue V1 = N->getOperand(Num: Vec0Idx + 1); |
2367 | SDValue V2 = N->getOperand(Num: Vec0Idx + 2); |
2368 | SDValue V3 = (NumVecs == 3) |
2369 | ? SDValue(CurDAG->getMachineNode(Opcode: TargetOpcode::IMPLICIT_DEF, dl, VT), 0) |
2370 | : N->getOperand(Num: Vec0Idx + 3); |
2371 | SDValue RegSeq = SDValue(createQuadQRegsNode(VT: MVT::v8i64, V0, V1, V2, V3), 0); |
2372 | |
2373 | // Store the even D registers. This is always an updating store, so that it |
2374 | // provides the address to the second store for the odd subregs. |
2375 | const SDValue OpsA[] = { MemAddr, Align, Reg0, RegSeq, Pred, Reg0, Chain }; |
2376 | SDNode *VStA = CurDAG->getMachineNode(Opcode: QOpcodes0[OpcodeIndex], dl, |
2377 | VT1: MemAddr.getValueType(), |
2378 | VT2: MVT::Other, Ops: OpsA); |
2379 | CurDAG->setNodeMemRefs(N: cast<MachineSDNode>(Val: VStA), NewMemRefs: {MemOp}); |
2380 | Chain = SDValue(VStA, 1); |
2381 | |
2382 | // Store the odd D registers. |
2383 | Ops.push_back(Elt: SDValue(VStA, 0)); |
2384 | Ops.push_back(Elt: Align); |
2385 | if (isUpdating) { |
2386 | SDValue Inc = N->getOperand(Num: AddrOpIdx + 1); |
2387 | assert(isa<ConstantSDNode>(Inc.getNode()) && |
2388 | "only constant post-increment update allowed for VST3/4" ); |
2389 | (void)Inc; |
2390 | Ops.push_back(Elt: Reg0); |
2391 | } |
2392 | Ops.push_back(Elt: RegSeq); |
2393 | Ops.push_back(Elt: Pred); |
2394 | Ops.push_back(Elt: Reg0); |
2395 | Ops.push_back(Elt: Chain); |
2396 | SDNode *VStB = CurDAG->getMachineNode(Opcode: QOpcodes1[OpcodeIndex], dl, ResultTys: ResTys, |
2397 | Ops); |
2398 | CurDAG->setNodeMemRefs(N: cast<MachineSDNode>(Val: VStB), NewMemRefs: {MemOp}); |
2399 | ReplaceNode(F: N, T: VStB); |
2400 | } |
2401 | |
2402 | void ARMDAGToDAGISel::SelectVLDSTLane(SDNode *N, bool IsLoad, bool isUpdating, |
2403 | unsigned NumVecs, |
2404 | const uint16_t *DOpcodes, |
2405 | const uint16_t *QOpcodes) { |
2406 | assert(Subtarget->hasNEON()); |
2407 | assert(NumVecs >=2 && NumVecs <= 4 && "VLDSTLane NumVecs out-of-range" ); |
2408 | SDLoc dl(N); |
2409 | |
2410 | SDValue MemAddr, Align; |
2411 | bool IsIntrinsic = !isUpdating; // By coincidence, all supported updating |
2412 | // nodes are not intrinsics. |
2413 | unsigned AddrOpIdx = IsIntrinsic ? 2 : 1; |
2414 | unsigned Vec0Idx = 3; // AddrOpIdx + (isUpdating ? 2 : 1) |
2415 | if (!SelectAddrMode6(Parent: N, N: N->getOperand(Num: AddrOpIdx), Addr&: MemAddr, Align)) |
2416 | return; |
2417 | |
2418 | MachineMemOperand *MemOp = cast<MemIntrinsicSDNode>(Val: N)->getMemOperand(); |
2419 | |
2420 | SDValue Chain = N->getOperand(Num: 0); |
2421 | unsigned Lane = N->getConstantOperandVal(Num: Vec0Idx + NumVecs); |
2422 | EVT VT = N->getOperand(Num: Vec0Idx).getValueType(); |
2423 | bool is64BitVector = VT.is64BitVector(); |
2424 | |
2425 | unsigned Alignment = 0; |
2426 | if (NumVecs != 3) { |
2427 | Alignment = Align->getAsZExtVal(); |
2428 | unsigned NumBytes = NumVecs * VT.getScalarSizeInBits() / 8; |
2429 | if (Alignment > NumBytes) |
2430 | Alignment = NumBytes; |
2431 | if (Alignment < 8 && Alignment < NumBytes) |
2432 | Alignment = 0; |
2433 | // Alignment must be a power of two; make sure of that. |
2434 | Alignment = (Alignment & -Alignment); |
2435 | if (Alignment == 1) |
2436 | Alignment = 0; |
2437 | } |
2438 | Align = CurDAG->getTargetConstant(Val: Alignment, DL: dl, VT: MVT::i32); |
2439 | |
2440 | unsigned OpcodeIndex; |
2441 | switch (VT.getSimpleVT().SimpleTy) { |
2442 | default: llvm_unreachable("unhandled vld/vst lane type" ); |
2443 | // Double-register operations: |
2444 | case MVT::v8i8: OpcodeIndex = 0; break; |
2445 | case MVT::v4f16: |
2446 | case MVT::v4bf16: |
2447 | case MVT::v4i16: OpcodeIndex = 1; break; |
2448 | case MVT::v2f32: |
2449 | case MVT::v2i32: OpcodeIndex = 2; break; |
2450 | // Quad-register operations: |
2451 | case MVT::v8f16: |
2452 | case MVT::v8bf16: |
2453 | case MVT::v8i16: OpcodeIndex = 0; break; |
2454 | case MVT::v4f32: |
2455 | case MVT::v4i32: OpcodeIndex = 1; break; |
2456 | } |
2457 | |
2458 | std::vector<EVT> ResTys; |
2459 | if (IsLoad) { |
2460 | unsigned ResTyElts = (NumVecs == 3) ? 4 : NumVecs; |
2461 | if (!is64BitVector) |
2462 | ResTyElts *= 2; |
2463 | ResTys.push_back(x: EVT::getVectorVT(Context&: *CurDAG->getContext(), |
2464 | VT: MVT::i64, NumElements: ResTyElts)); |
2465 | } |
2466 | if (isUpdating) |
2467 | ResTys.push_back(x: MVT::i32); |
2468 | ResTys.push_back(x: MVT::Other); |
2469 | |
2470 | SDValue Pred = getAL(CurDAG, dl); |
2471 | SDValue Reg0 = CurDAG->getRegister(Reg: 0, VT: MVT::i32); |
2472 | |
2473 | SmallVector<SDValue, 8> Ops; |
2474 | Ops.push_back(Elt: MemAddr); |
2475 | Ops.push_back(Elt: Align); |
2476 | if (isUpdating) { |
2477 | SDValue Inc = N->getOperand(Num: AddrOpIdx + 1); |
2478 | bool IsImmUpdate = |
2479 | isPerfectIncrement(Inc, VecTy: VT.getVectorElementType(), NumVecs); |
2480 | Ops.push_back(Elt: IsImmUpdate ? Reg0 : Inc); |
2481 | } |
2482 | |
2483 | SDValue SuperReg; |
2484 | SDValue V0 = N->getOperand(Num: Vec0Idx + 0); |
2485 | SDValue V1 = N->getOperand(Num: Vec0Idx + 1); |
2486 | if (NumVecs == 2) { |
2487 | if (is64BitVector) |
2488 | SuperReg = SDValue(createDRegPairNode(VT: MVT::v2i64, V0, V1), 0); |
2489 | else |
2490 | SuperReg = SDValue(createQRegPairNode(VT: MVT::v4i64, V0, V1), 0); |
2491 | } else { |
2492 | SDValue V2 = N->getOperand(Num: Vec0Idx + 2); |
2493 | SDValue V3 = (NumVecs == 3) |
2494 | ? SDValue(CurDAG->getMachineNode(Opcode: TargetOpcode::IMPLICIT_DEF, dl, VT), 0) |
2495 | : N->getOperand(Num: Vec0Idx + 3); |
2496 | if (is64BitVector) |
2497 | SuperReg = SDValue(createQuadDRegsNode(VT: MVT::v4i64, V0, V1, V2, V3), 0); |
2498 | else |
2499 | SuperReg = SDValue(createQuadQRegsNode(VT: MVT::v8i64, V0, V1, V2, V3), 0); |
2500 | } |
2501 | Ops.push_back(Elt: SuperReg); |
2502 | Ops.push_back(Elt: getI32Imm(Imm: Lane, dl)); |
2503 | Ops.push_back(Elt: Pred); |
2504 | Ops.push_back(Elt: Reg0); |
2505 | Ops.push_back(Elt: Chain); |
2506 | |
2507 | unsigned Opc = (is64BitVector ? DOpcodes[OpcodeIndex] : |
2508 | QOpcodes[OpcodeIndex]); |
2509 | SDNode *VLdLn = CurDAG->getMachineNode(Opcode: Opc, dl, ResultTys: ResTys, Ops); |
2510 | CurDAG->setNodeMemRefs(N: cast<MachineSDNode>(Val: VLdLn), NewMemRefs: {MemOp}); |
2511 | if (!IsLoad) { |
2512 | ReplaceNode(F: N, T: VLdLn); |
2513 | return; |
2514 | } |
2515 | |
2516 | // Extract the subregisters. |
2517 | SuperReg = SDValue(VLdLn, 0); |
2518 | static_assert(ARM::dsub_7 == ARM::dsub_0 + 7 && |
2519 | ARM::qsub_3 == ARM::qsub_0 + 3, |
2520 | "Unexpected subreg numbering" ); |
2521 | unsigned Sub0 = is64BitVector ? ARM::dsub_0 : ARM::qsub_0; |
2522 | for (unsigned Vec = 0; Vec < NumVecs; ++Vec) |
2523 | ReplaceUses(F: SDValue(N, Vec), |
2524 | T: CurDAG->getTargetExtractSubreg(SRIdx: Sub0 + Vec, DL: dl, VT, Operand: SuperReg)); |
2525 | ReplaceUses(F: SDValue(N, NumVecs), T: SDValue(VLdLn, 1)); |
2526 | if (isUpdating) |
2527 | ReplaceUses(F: SDValue(N, NumVecs + 1), T: SDValue(VLdLn, 2)); |
2528 | CurDAG->RemoveDeadNode(N); |
2529 | } |
2530 | |
2531 | template <typename SDValueVector> |
2532 | void ARMDAGToDAGISel::AddMVEPredicateToOps(SDValueVector &Ops, SDLoc Loc, |
2533 | SDValue PredicateMask) { |
2534 | Ops.push_back(CurDAG->getTargetConstant(Val: ARMVCC::Then, DL: Loc, VT: MVT::i32)); |
2535 | Ops.push_back(PredicateMask); |
2536 | Ops.push_back(CurDAG->getRegister(Reg: 0, VT: MVT::i32)); // tp_reg |
2537 | } |
2538 | |
2539 | template <typename SDValueVector> |
2540 | void ARMDAGToDAGISel::AddMVEPredicateToOps(SDValueVector &Ops, SDLoc Loc, |
2541 | SDValue PredicateMask, |
2542 | SDValue Inactive) { |
2543 | Ops.push_back(CurDAG->getTargetConstant(Val: ARMVCC::Then, DL: Loc, VT: MVT::i32)); |
2544 | Ops.push_back(PredicateMask); |
2545 | Ops.push_back(CurDAG->getRegister(Reg: 0, VT: MVT::i32)); // tp_reg |
2546 | Ops.push_back(Inactive); |
2547 | } |
2548 | |
2549 | template <typename SDValueVector> |
2550 | void ARMDAGToDAGISel::AddEmptyMVEPredicateToOps(SDValueVector &Ops, SDLoc Loc) { |
2551 | Ops.push_back(CurDAG->getTargetConstant(Val: ARMVCC::None, DL: Loc, VT: MVT::i32)); |
2552 | Ops.push_back(CurDAG->getRegister(Reg: 0, VT: MVT::i32)); |
2553 | Ops.push_back(CurDAG->getRegister(Reg: 0, VT: MVT::i32)); // tp_reg |
2554 | } |
2555 | |
2556 | template <typename SDValueVector> |
2557 | void ARMDAGToDAGISel::AddEmptyMVEPredicateToOps(SDValueVector &Ops, SDLoc Loc, |
2558 | EVT InactiveTy) { |
2559 | Ops.push_back(CurDAG->getTargetConstant(Val: ARMVCC::None, DL: Loc, VT: MVT::i32)); |
2560 | Ops.push_back(CurDAG->getRegister(Reg: 0, VT: MVT::i32)); |
2561 | Ops.push_back(CurDAG->getRegister(Reg: 0, VT: MVT::i32)); // tp_reg |
2562 | Ops.push_back(SDValue( |
2563 | CurDAG->getMachineNode(Opcode: TargetOpcode::IMPLICIT_DEF, dl: Loc, VT: InactiveTy), 0)); |
2564 | } |
2565 | |
2566 | void ARMDAGToDAGISel::SelectMVE_WB(SDNode *N, const uint16_t *Opcodes, |
2567 | bool Predicated) { |
2568 | SDLoc Loc(N); |
2569 | SmallVector<SDValue, 8> Ops; |
2570 | |
2571 | uint16_t Opcode; |
2572 | switch (N->getValueType(ResNo: 1).getVectorElementType().getSizeInBits()) { |
2573 | case 32: |
2574 | Opcode = Opcodes[0]; |
2575 | break; |
2576 | case 64: |
2577 | Opcode = Opcodes[1]; |
2578 | break; |
2579 | default: |
2580 | llvm_unreachable("bad vector element size in SelectMVE_WB" ); |
2581 | } |
2582 | |
2583 | Ops.push_back(Elt: N->getOperand(Num: 2)); // vector of base addresses |
2584 | |
2585 | int32_t ImmValue = N->getConstantOperandVal(Num: 3); |
2586 | Ops.push_back(Elt: getI32Imm(Imm: ImmValue, dl: Loc)); // immediate offset |
2587 | |
2588 | if (Predicated) |
2589 | AddMVEPredicateToOps(Ops, Loc, PredicateMask: N->getOperand(Num: 4)); |
2590 | else |
2591 | AddEmptyMVEPredicateToOps(Ops, Loc); |
2592 | |
2593 | Ops.push_back(Elt: N->getOperand(Num: 0)); // chain |
2594 | |
2595 | SmallVector<EVT, 8> VTs; |
2596 | VTs.push_back(Elt: N->getValueType(ResNo: 1)); |
2597 | VTs.push_back(Elt: N->getValueType(ResNo: 0)); |
2598 | VTs.push_back(Elt: N->getValueType(ResNo: 2)); |
2599 | |
2600 | SDNode *New = CurDAG->getMachineNode(Opcode, dl: SDLoc(N), ResultTys: VTs, Ops); |
2601 | ReplaceUses(F: SDValue(N, 0), T: SDValue(New, 1)); |
2602 | ReplaceUses(F: SDValue(N, 1), T: SDValue(New, 0)); |
2603 | ReplaceUses(F: SDValue(N, 2), T: SDValue(New, 2)); |
2604 | transferMemOperands(N, Result: New); |
2605 | CurDAG->RemoveDeadNode(N); |
2606 | } |
2607 | |
2608 | void ARMDAGToDAGISel::SelectMVE_LongShift(SDNode *N, uint16_t Opcode, |
2609 | bool Immediate, |
2610 | bool HasSaturationOperand) { |
2611 | SDLoc Loc(N); |
2612 | SmallVector<SDValue, 8> Ops; |
2613 | |
2614 | // Two 32-bit halves of the value to be shifted |
2615 | Ops.push_back(Elt: N->getOperand(Num: 1)); |
2616 | Ops.push_back(Elt: N->getOperand(Num: 2)); |
2617 | |
2618 | // The shift count |
2619 | if (Immediate) { |
2620 | int32_t ImmValue = N->getConstantOperandVal(Num: 3); |
2621 | Ops.push_back(Elt: getI32Imm(Imm: ImmValue, dl: Loc)); // immediate shift count |
2622 | } else { |
2623 | Ops.push_back(Elt: N->getOperand(Num: 3)); |
2624 | } |
2625 | |
2626 | // The immediate saturation operand, if any |
2627 | if (HasSaturationOperand) { |
2628 | int32_t SatOp = N->getConstantOperandVal(Num: 4); |
2629 | int SatBit = (SatOp == 64 ? 0 : 1); |
2630 | Ops.push_back(Elt: getI32Imm(Imm: SatBit, dl: Loc)); |
2631 | } |
2632 | |
2633 | // MVE scalar shifts are IT-predicable, so include the standard |
2634 | // predicate arguments. |
2635 | Ops.push_back(Elt: getAL(CurDAG, dl: Loc)); |
2636 | Ops.push_back(Elt: CurDAG->getRegister(Reg: 0, VT: MVT::i32)); |
2637 | |
2638 | CurDAG->SelectNodeTo(N, MachineOpc: Opcode, VTs: N->getVTList(), Ops: ArrayRef(Ops)); |
2639 | } |
2640 | |
2641 | void ARMDAGToDAGISel::SelectMVE_VADCSBC(SDNode *N, uint16_t OpcodeWithCarry, |
2642 | uint16_t OpcodeWithNoCarry, |
2643 | bool Add, bool Predicated) { |
2644 | SDLoc Loc(N); |
2645 | SmallVector<SDValue, 8> Ops; |
2646 | uint16_t Opcode; |
2647 | |
2648 | unsigned FirstInputOp = Predicated ? 2 : 1; |
2649 | |
2650 | // Two input vectors and the input carry flag |
2651 | Ops.push_back(Elt: N->getOperand(Num: FirstInputOp)); |
2652 | Ops.push_back(Elt: N->getOperand(Num: FirstInputOp + 1)); |
2653 | SDValue CarryIn = N->getOperand(Num: FirstInputOp + 2); |
2654 | ConstantSDNode *CarryInConstant = dyn_cast<ConstantSDNode>(Val&: CarryIn); |
2655 | uint32_t CarryMask = 1 << 29; |
2656 | uint32_t CarryExpected = Add ? 0 : CarryMask; |
2657 | if (CarryInConstant && |
2658 | (CarryInConstant->getZExtValue() & CarryMask) == CarryExpected) { |
2659 | Opcode = OpcodeWithNoCarry; |
2660 | } else { |
2661 | Ops.push_back(Elt: CarryIn); |
2662 | Opcode = OpcodeWithCarry; |
2663 | } |
2664 | |
2665 | if (Predicated) |
2666 | AddMVEPredicateToOps(Ops, Loc, |
2667 | PredicateMask: N->getOperand(Num: FirstInputOp + 3), // predicate |
2668 | Inactive: N->getOperand(Num: FirstInputOp - 1)); // inactive |
2669 | else |
2670 | AddEmptyMVEPredicateToOps(Ops, Loc, InactiveTy: N->getValueType(ResNo: 0)); |
2671 | |
2672 | CurDAG->SelectNodeTo(N, MachineOpc: Opcode, VTs: N->getVTList(), Ops: ArrayRef(Ops)); |
2673 | } |
2674 | |
2675 | void ARMDAGToDAGISel::SelectMVE_VSHLC(SDNode *N, bool Predicated) { |
2676 | SDLoc Loc(N); |
2677 | SmallVector<SDValue, 8> Ops; |
2678 | |
2679 | // One vector input, followed by a 32-bit word of bits to shift in |
2680 | // and then an immediate shift count |
2681 | Ops.push_back(Elt: N->getOperand(Num: 1)); |
2682 | Ops.push_back(Elt: N->getOperand(Num: 2)); |
2683 | int32_t ImmValue = N->getConstantOperandVal(Num: 3); |
2684 | Ops.push_back(Elt: getI32Imm(Imm: ImmValue, dl: Loc)); // immediate shift count |
2685 | |
2686 | if (Predicated) |
2687 | AddMVEPredicateToOps(Ops, Loc, PredicateMask: N->getOperand(Num: 4)); |
2688 | else |
2689 | AddEmptyMVEPredicateToOps(Ops, Loc); |
2690 | |
2691 | CurDAG->SelectNodeTo(N, MachineOpc: ARM::MVE_VSHLC, VTs: N->getVTList(), Ops: ArrayRef(Ops)); |
2692 | } |
2693 | |
2694 | static bool SDValueToConstBool(SDValue SDVal) { |
2695 | assert(isa<ConstantSDNode>(SDVal) && "expected a compile-time constant" ); |
2696 | ConstantSDNode *SDValConstant = dyn_cast<ConstantSDNode>(Val&: SDVal); |
2697 | uint64_t Value = SDValConstant->getZExtValue(); |
2698 | assert((Value == 0 || Value == 1) && "expected value 0 or 1" ); |
2699 | return Value; |
2700 | } |
2701 | |
2702 | void ARMDAGToDAGISel::SelectBaseMVE_VMLLDAV(SDNode *N, bool Predicated, |
2703 | const uint16_t *OpcodesS, |
2704 | const uint16_t *OpcodesU, |
2705 | size_t Stride, size_t TySize) { |
2706 | assert(TySize < Stride && "Invalid TySize" ); |
2707 | bool IsUnsigned = SDValueToConstBool(SDVal: N->getOperand(Num: 1)); |
2708 | bool IsSub = SDValueToConstBool(SDVal: N->getOperand(Num: 2)); |
2709 | bool IsExchange = SDValueToConstBool(SDVal: N->getOperand(Num: 3)); |
2710 | if (IsUnsigned) { |
2711 | assert(!IsSub && |
2712 | "Unsigned versions of vmlsldav[a]/vrmlsldavh[a] do not exist" ); |
2713 | assert(!IsExchange && |
2714 | "Unsigned versions of vmlaldav[a]x/vrmlaldavh[a]x do not exist" ); |
2715 | } |
2716 | |
2717 | auto OpIsZero = [N](size_t OpNo) { |
2718 | return isNullConstant(V: N->getOperand(Num: OpNo)); |
2719 | }; |
2720 | |
2721 | // If the input accumulator value is not zero, select an instruction with |
2722 | // accumulator, otherwise select an instruction without accumulator |
2723 | bool IsAccum = !(OpIsZero(4) && OpIsZero(5)); |
2724 | |
2725 | const uint16_t *Opcodes = IsUnsigned ? OpcodesU : OpcodesS; |
2726 | if (IsSub) |
2727 | Opcodes += 4 * Stride; |
2728 | if (IsExchange) |
2729 | Opcodes += 2 * Stride; |
2730 | if (IsAccum) |
2731 | Opcodes += Stride; |
2732 | uint16_t Opcode = Opcodes[TySize]; |
2733 | |
2734 | SDLoc Loc(N); |
2735 | SmallVector<SDValue, 8> Ops; |
2736 | // Push the accumulator operands, if they are used |
2737 | if (IsAccum) { |
2738 | Ops.push_back(Elt: N->getOperand(Num: 4)); |
2739 | Ops.push_back(Elt: N->getOperand(Num: 5)); |
2740 | } |
2741 | // Push the two vector operands |
2742 | Ops.push_back(Elt: N->getOperand(Num: 6)); |
2743 | Ops.push_back(Elt: N->getOperand(Num: 7)); |
2744 | |
2745 | if (Predicated) |
2746 | AddMVEPredicateToOps(Ops, Loc, PredicateMask: N->getOperand(Num: 8)); |
2747 | else |
2748 | AddEmptyMVEPredicateToOps(Ops, Loc); |
2749 | |
2750 | CurDAG->SelectNodeTo(N, MachineOpc: Opcode, VTs: N->getVTList(), Ops: ArrayRef(Ops)); |
2751 | } |
2752 | |
2753 | void ARMDAGToDAGISel::SelectMVE_VMLLDAV(SDNode *N, bool Predicated, |
2754 | const uint16_t *OpcodesS, |
2755 | const uint16_t *OpcodesU) { |
2756 | EVT VecTy = N->getOperand(Num: 6).getValueType(); |
2757 | size_t SizeIndex; |
2758 | switch (VecTy.getVectorElementType().getSizeInBits()) { |
2759 | case 16: |
2760 | SizeIndex = 0; |
2761 | break; |
2762 | case 32: |
2763 | SizeIndex = 1; |
2764 | break; |
2765 | default: |
2766 | llvm_unreachable("bad vector element size" ); |
2767 | } |
2768 | |
2769 | SelectBaseMVE_VMLLDAV(N, Predicated, OpcodesS, OpcodesU, Stride: 2, TySize: SizeIndex); |
2770 | } |
2771 | |
2772 | void ARMDAGToDAGISel::SelectMVE_VRMLLDAVH(SDNode *N, bool Predicated, |
2773 | const uint16_t *OpcodesS, |
2774 | const uint16_t *OpcodesU) { |
2775 | assert( |
2776 | N->getOperand(6).getValueType().getVectorElementType().getSizeInBits() == |
2777 | 32 && |
2778 | "bad vector element size" ); |
2779 | SelectBaseMVE_VMLLDAV(N, Predicated, OpcodesS, OpcodesU, Stride: 1, TySize: 0); |
2780 | } |
2781 | |
2782 | void ARMDAGToDAGISel::SelectMVE_VLD(SDNode *N, unsigned NumVecs, |
2783 | const uint16_t *const *Opcodes, |
2784 | bool HasWriteback) { |
2785 | EVT VT = N->getValueType(ResNo: 0); |
2786 | SDLoc Loc(N); |
2787 | |
2788 | const uint16_t *OurOpcodes; |
2789 | switch (VT.getVectorElementType().getSizeInBits()) { |
2790 | case 8: |
2791 | OurOpcodes = Opcodes[0]; |
2792 | break; |
2793 | case 16: |
2794 | OurOpcodes = Opcodes[1]; |
2795 | break; |
2796 | case 32: |
2797 | OurOpcodes = Opcodes[2]; |
2798 | break; |
2799 | default: |
2800 | llvm_unreachable("bad vector element size in SelectMVE_VLD" ); |
2801 | } |
2802 | |
2803 | EVT DataTy = EVT::getVectorVT(Context&: *CurDAG->getContext(), VT: MVT::i64, NumElements: NumVecs * 2); |
2804 | SmallVector<EVT, 4> ResultTys = {DataTy, MVT::Other}; |
2805 | unsigned PtrOperand = HasWriteback ? 1 : 2; |
2806 | |
2807 | auto Data = SDValue( |
2808 | CurDAG->getMachineNode(Opcode: TargetOpcode::IMPLICIT_DEF, dl: Loc, VT: DataTy), 0); |
2809 | SDValue Chain = N->getOperand(Num: 0); |
2810 | // Add a MVE_VLDn instruction for each Vec, except the last |
2811 | for (unsigned Stage = 0; Stage < NumVecs - 1; ++Stage) { |
2812 | SDValue Ops[] = {Data, N->getOperand(Num: PtrOperand), Chain}; |
2813 | auto LoadInst = |
2814 | CurDAG->getMachineNode(Opcode: OurOpcodes[Stage], dl: Loc, ResultTys, Ops); |
2815 | Data = SDValue(LoadInst, 0); |
2816 | Chain = SDValue(LoadInst, 1); |
2817 | transferMemOperands(N, Result: LoadInst); |
2818 | } |
2819 | // The last may need a writeback on it |
2820 | if (HasWriteback) |
2821 | ResultTys = {DataTy, MVT::i32, MVT::Other}; |
2822 | SDValue Ops[] = {Data, N->getOperand(Num: PtrOperand), Chain}; |
2823 | auto LoadInst = |
2824 | CurDAG->getMachineNode(Opcode: OurOpcodes[NumVecs - 1], dl: Loc, ResultTys, Ops); |
2825 | transferMemOperands(N, Result: LoadInst); |
2826 | |
2827 | unsigned i; |
2828 | for (i = 0; i < NumVecs; i++) |
2829 | ReplaceUses(F: SDValue(N, i), |
2830 | T: CurDAG->getTargetExtractSubreg(SRIdx: ARM::qsub_0 + i, DL: Loc, VT, |
2831 | Operand: SDValue(LoadInst, 0))); |
2832 | if (HasWriteback) |
2833 | ReplaceUses(F: SDValue(N, i++), T: SDValue(LoadInst, 1)); |
2834 | ReplaceUses(F: SDValue(N, i), T: SDValue(LoadInst, HasWriteback ? 2 : 1)); |
2835 | CurDAG->RemoveDeadNode(N); |
2836 | } |
2837 | |
2838 | void ARMDAGToDAGISel::SelectMVE_VxDUP(SDNode *N, const uint16_t *Opcodes, |
2839 | bool Wrapping, bool Predicated) { |
2840 | EVT VT = N->getValueType(ResNo: 0); |
2841 | SDLoc Loc(N); |
2842 | |
2843 | uint16_t Opcode; |
2844 | switch (VT.getScalarSizeInBits()) { |
2845 | case 8: |
2846 | Opcode = Opcodes[0]; |
2847 | break; |
2848 | case 16: |
2849 | Opcode = Opcodes[1]; |
2850 | break; |
2851 | case 32: |
2852 | Opcode = Opcodes[2]; |
2853 | break; |
2854 | default: |
2855 | llvm_unreachable("bad vector element size in SelectMVE_VxDUP" ); |
2856 | } |
2857 | |
2858 | SmallVector<SDValue, 8> Ops; |
2859 | unsigned OpIdx = 1; |
2860 | |
2861 | SDValue Inactive; |
2862 | if (Predicated) |
2863 | Inactive = N->getOperand(Num: OpIdx++); |
2864 | |
2865 | Ops.push_back(Elt: N->getOperand(Num: OpIdx++)); // base |
2866 | if (Wrapping) |
2867 | Ops.push_back(Elt: N->getOperand(Num: OpIdx++)); // limit |
2868 | |
2869 | SDValue ImmOp = N->getOperand(Num: OpIdx++); // step |
2870 | int ImmValue = ImmOp->getAsZExtVal(); |
2871 | Ops.push_back(Elt: getI32Imm(Imm: ImmValue, dl: Loc)); |
2872 | |
2873 | if (Predicated) |
2874 | AddMVEPredicateToOps(Ops, Loc, PredicateMask: N->getOperand(Num: OpIdx), Inactive); |
2875 | else |
2876 | AddEmptyMVEPredicateToOps(Ops, Loc, InactiveTy: N->getValueType(ResNo: 0)); |
2877 | |
2878 | CurDAG->SelectNodeTo(N, MachineOpc: Opcode, VTs: N->getVTList(), Ops: ArrayRef(Ops)); |
2879 | } |
2880 | |
2881 | void ARMDAGToDAGISel::SelectCDE_CXxD(SDNode *N, uint16_t Opcode, |
2882 | size_t , bool HasAccum) { |
2883 | bool IsBigEndian = CurDAG->getDataLayout().isBigEndian(); |
2884 | SDLoc Loc(N); |
2885 | SmallVector<SDValue, 8> Ops; |
2886 | |
2887 | unsigned OpIdx = 1; |
2888 | |
2889 | // Convert and append the immediate operand designating the coprocessor. |
2890 | SDValue ImmCorpoc = N->getOperand(Num: OpIdx++); |
2891 | uint32_t ImmCoprocVal = ImmCorpoc->getAsZExtVal(); |
2892 | Ops.push_back(Elt: getI32Imm(Imm: ImmCoprocVal, dl: Loc)); |
2893 | |
2894 | // For accumulating variants copy the low and high order parts of the |
2895 | // accumulator into a register pair and add it to the operand vector. |
2896 | if (HasAccum) { |
2897 | SDValue AccLo = N->getOperand(Num: OpIdx++); |
2898 | SDValue AccHi = N->getOperand(Num: OpIdx++); |
2899 | if (IsBigEndian) |
2900 | std::swap(a&: AccLo, b&: AccHi); |
2901 | Ops.push_back(Elt: SDValue(createGPRPairNode(VT: MVT::Untyped, V0: AccLo, V1: AccHi), 0)); |
2902 | } |
2903 | |
2904 | // Copy extra operands as-is. |
2905 | for (size_t I = 0; I < NumExtraOps; I++) |
2906 | Ops.push_back(Elt: N->getOperand(Num: OpIdx++)); |
2907 | |
2908 | // Convert and append the immediate operand |
2909 | SDValue Imm = N->getOperand(Num: OpIdx); |
2910 | uint32_t ImmVal = Imm->getAsZExtVal(); |
2911 | Ops.push_back(Elt: getI32Imm(Imm: ImmVal, dl: Loc)); |
2912 | |
2913 | // Accumulating variants are IT-predicable, add predicate operands. |
2914 | if (HasAccum) { |
2915 | SDValue Pred = getAL(CurDAG, dl: Loc); |
2916 | SDValue PredReg = CurDAG->getRegister(Reg: 0, VT: MVT::i32); |
2917 | Ops.push_back(Elt: Pred); |
2918 | Ops.push_back(Elt: PredReg); |
2919 | } |
2920 | |
2921 | // Create the CDE intruction |
2922 | SDNode *InstrNode = CurDAG->getMachineNode(Opcode, dl: Loc, VT: MVT::Untyped, Ops); |
2923 | SDValue ResultPair = SDValue(InstrNode, 0); |
2924 | |
2925 | // The original intrinsic had two outputs, and the output of the dual-register |
2926 | // CDE instruction is a register pair. We need to extract the two subregisters |
2927 | // and replace all uses of the original outputs with the extracted |
2928 | // subregisters. |
2929 | uint16_t SubRegs[2] = {ARM::gsub_0, ARM::gsub_1}; |
2930 | if (IsBigEndian) |
2931 | std::swap(a&: SubRegs[0], b&: SubRegs[1]); |
2932 | |
2933 | for (size_t ResIdx = 0; ResIdx < 2; ResIdx++) { |
2934 | if (SDValue(N, ResIdx).use_empty()) |
2935 | continue; |
2936 | SDValue SubReg = CurDAG->getTargetExtractSubreg(SRIdx: SubRegs[ResIdx], DL: Loc, |
2937 | VT: MVT::i32, Operand: ResultPair); |
2938 | ReplaceUses(F: SDValue(N, ResIdx), T: SubReg); |
2939 | } |
2940 | |
2941 | CurDAG->RemoveDeadNode(N); |
2942 | } |
2943 | |
2944 | void ARMDAGToDAGISel::SelectVLDDup(SDNode *N, bool IsIntrinsic, |
2945 | bool isUpdating, unsigned NumVecs, |
2946 | const uint16_t *DOpcodes, |
2947 | const uint16_t *QOpcodes0, |
2948 | const uint16_t *QOpcodes1) { |
2949 | assert(Subtarget->hasNEON()); |
2950 | assert(NumVecs >= 1 && NumVecs <= 4 && "VLDDup NumVecs out-of-range" ); |
2951 | SDLoc dl(N); |
2952 | |
2953 | SDValue MemAddr, Align; |
2954 | unsigned AddrOpIdx = IsIntrinsic ? 2 : 1; |
2955 | if (!SelectAddrMode6(Parent: N, N: N->getOperand(Num: AddrOpIdx), Addr&: MemAddr, Align)) |
2956 | return; |
2957 | |
2958 | SDValue Chain = N->getOperand(Num: 0); |
2959 | EVT VT = N->getValueType(ResNo: 0); |
2960 | bool is64BitVector = VT.is64BitVector(); |
2961 | |
2962 | unsigned Alignment = 0; |
2963 | if (NumVecs != 3) { |
2964 | Alignment = Align->getAsZExtVal(); |
2965 | unsigned NumBytes = NumVecs * VT.getScalarSizeInBits() / 8; |
2966 | if (Alignment > NumBytes) |
2967 | Alignment = NumBytes; |
2968 | if (Alignment < 8 && Alignment < NumBytes) |
2969 | Alignment = 0; |
2970 | // Alignment must be a power of two; make sure of that. |
2971 | Alignment = (Alignment & -Alignment); |
2972 | if (Alignment == 1) |
2973 | Alignment = 0; |
2974 | } |
2975 | Align = CurDAG->getTargetConstant(Val: Alignment, DL: dl, VT: MVT::i32); |
2976 | |
2977 | unsigned OpcodeIndex; |
2978 | switch (VT.getSimpleVT().SimpleTy) { |
2979 | default: llvm_unreachable("unhandled vld-dup type" ); |
2980 | case MVT::v8i8: |
2981 | case MVT::v16i8: OpcodeIndex = 0; break; |
2982 | case MVT::v4i16: |
2983 | case MVT::v8i16: |
2984 | case MVT::v4f16: |
2985 | case MVT::v8f16: |
2986 | case MVT::v4bf16: |
2987 | case MVT::v8bf16: |
2988 | OpcodeIndex = 1; break; |
2989 | case MVT::v2f32: |
2990 | case MVT::v2i32: |
2991 | case MVT::v4f32: |
2992 | case MVT::v4i32: OpcodeIndex = 2; break; |
2993 | case MVT::v1f64: |
2994 | case MVT::v1i64: OpcodeIndex = 3; break; |
2995 | } |
2996 | |
2997 | unsigned ResTyElts = (NumVecs == 3) ? 4 : NumVecs; |
2998 | if (!is64BitVector) |
2999 | ResTyElts *= 2; |
3000 | EVT ResTy = EVT::getVectorVT(Context&: *CurDAG->getContext(), VT: MVT::i64, NumElements: ResTyElts); |
3001 | |
3002 | std::vector<EVT> ResTys; |
3003 | ResTys.push_back(x: ResTy); |
3004 | if (isUpdating) |
3005 | ResTys.push_back(x: MVT::i32); |
3006 | ResTys.push_back(x: MVT::Other); |
3007 | |
3008 | SDValue Pred = getAL(CurDAG, dl); |
3009 | SDValue Reg0 = CurDAG->getRegister(Reg: 0, VT: MVT::i32); |
3010 | |
3011 | SmallVector<SDValue, 6> Ops; |
3012 | Ops.push_back(Elt: MemAddr); |
3013 | Ops.push_back(Elt: Align); |
3014 | unsigned Opc = is64BitVector ? DOpcodes[OpcodeIndex] |
3015 | : (NumVecs == 1) ? QOpcodes0[OpcodeIndex] |
3016 | : QOpcodes1[OpcodeIndex]; |
3017 | if (isUpdating) { |
3018 | SDValue Inc = N->getOperand(Num: 2); |
3019 | bool IsImmUpdate = |
3020 | isPerfectIncrement(Inc, VecTy: VT.getVectorElementType(), NumVecs); |
3021 | if (IsImmUpdate) { |
3022 | if (!isVLDfixed(Opc)) |
3023 | Ops.push_back(Elt: Reg0); |
3024 | } else { |
3025 | if (isVLDfixed(Opc)) |
3026 | Opc = getVLDSTRegisterUpdateOpcode(Opc); |
3027 | Ops.push_back(Elt: Inc); |
3028 | } |
3029 | } |
3030 | if (is64BitVector || NumVecs == 1) { |
3031 | // Double registers and VLD1 quad registers are directly supported. |
3032 | } else { |
3033 | SDValue ImplDef = SDValue( |
3034 | CurDAG->getMachineNode(Opcode: TargetOpcode::IMPLICIT_DEF, dl, VT: ResTy), 0); |
3035 | const SDValue OpsA[] = {MemAddr, Align, ImplDef, Pred, Reg0, Chain}; |
3036 | SDNode *VLdA = CurDAG->getMachineNode(Opcode: QOpcodes0[OpcodeIndex], dl, VT1: ResTy, |
3037 | VT2: MVT::Other, Ops: OpsA); |
3038 | Ops.push_back(Elt: SDValue(VLdA, 0)); |
3039 | Chain = SDValue(VLdA, 1); |
3040 | } |
3041 | |
3042 | Ops.push_back(Elt: Pred); |
3043 | Ops.push_back(Elt: Reg0); |
3044 | Ops.push_back(Elt: Chain); |
3045 | |
3046 | SDNode *VLdDup = CurDAG->getMachineNode(Opcode: Opc, dl, ResultTys: ResTys, Ops); |
3047 | |
3048 | // Transfer memoperands. |
3049 | MachineMemOperand *MemOp = cast<MemIntrinsicSDNode>(Val: N)->getMemOperand(); |
3050 | CurDAG->setNodeMemRefs(N: cast<MachineSDNode>(Val: VLdDup), NewMemRefs: {MemOp}); |
3051 | |
3052 | // Extract the subregisters. |
3053 | if (NumVecs == 1) { |
3054 | ReplaceUses(F: SDValue(N, 0), T: SDValue(VLdDup, 0)); |
3055 | } else { |
3056 | SDValue SuperReg = SDValue(VLdDup, 0); |
3057 | static_assert(ARM::dsub_7 == ARM::dsub_0 + 7, "Unexpected subreg numbering" ); |
3058 | unsigned SubIdx = is64BitVector ? ARM::dsub_0 : ARM::qsub_0; |
3059 | for (unsigned Vec = 0; Vec != NumVecs; ++Vec) { |
3060 | ReplaceUses(F: SDValue(N, Vec), |
3061 | T: CurDAG->getTargetExtractSubreg(SRIdx: SubIdx+Vec, DL: dl, VT, Operand: SuperReg)); |
3062 | } |
3063 | } |
3064 | ReplaceUses(F: SDValue(N, NumVecs), T: SDValue(VLdDup, 1)); |
3065 | if (isUpdating) |
3066 | ReplaceUses(F: SDValue(N, NumVecs + 1), T: SDValue(VLdDup, 2)); |
3067 | CurDAG->RemoveDeadNode(N); |
3068 | } |
3069 | |
3070 | bool ARMDAGToDAGISel::tryInsertVectorElt(SDNode *N) { |
3071 | if (!Subtarget->hasMVEIntegerOps()) |
3072 | return false; |
3073 | |
3074 | SDLoc dl(N); |
3075 | |
3076 | // We are trying to use VMOV/VMOVX/VINS to more efficiently lower insert and |
3077 | // extracts of v8f16 and v8i16 vectors. Check that we have two adjacent |
3078 | // inserts of the correct type: |
3079 | SDValue Ins1 = SDValue(N, 0); |
3080 | SDValue Ins2 = N->getOperand(Num: 0); |
3081 | EVT VT = Ins1.getValueType(); |
3082 | if (Ins2.getOpcode() != ISD::INSERT_VECTOR_ELT || !Ins2.hasOneUse() || |
3083 | !isa<ConstantSDNode>(Val: Ins1.getOperand(i: 2)) || |
3084 | !isa<ConstantSDNode>(Val: Ins2.getOperand(i: 2)) || |
3085 | (VT != MVT::v8f16 && VT != MVT::v8i16) || (Ins2.getValueType() != VT)) |
3086 | return false; |
3087 | |
3088 | unsigned Lane1 = Ins1.getConstantOperandVal(i: 2); |
3089 | unsigned Lane2 = Ins2.getConstantOperandVal(i: 2); |
3090 | if (Lane2 % 2 != 0 || Lane1 != Lane2 + 1) |
3091 | return false; |
3092 | |
3093 | // If the inserted values will be able to use T/B already, leave it to the |
3094 | // existing tablegen patterns. For example VCVTT/VCVTB. |
3095 | SDValue Val1 = Ins1.getOperand(i: 1); |
3096 | SDValue Val2 = Ins2.getOperand(i: 1); |
3097 | if (Val1.getOpcode() == ISD::FP_ROUND || Val2.getOpcode() == ISD::FP_ROUND) |
3098 | return false; |
3099 | |
3100 | // Check if the inserted values are both extracts. |
3101 | if ((Val1.getOpcode() == ISD::EXTRACT_VECTOR_ELT || |
3102 | Val1.getOpcode() == ARMISD::VGETLANEu) && |
3103 | (Val2.getOpcode() == ISD::EXTRACT_VECTOR_ELT || |
3104 | Val2.getOpcode() == ARMISD::VGETLANEu) && |
3105 | isa<ConstantSDNode>(Val: Val1.getOperand(i: 1)) && |
3106 | isa<ConstantSDNode>(Val: Val2.getOperand(i: 1)) && |
3107 | (Val1.getOperand(i: 0).getValueType() == MVT::v8f16 || |
3108 | Val1.getOperand(i: 0).getValueType() == MVT::v8i16) && |
3109 | (Val2.getOperand(i: 0).getValueType() == MVT::v8f16 || |
3110 | Val2.getOperand(i: 0).getValueType() == MVT::v8i16)) { |
3111 | unsigned = Val1.getConstantOperandVal(i: 1); |
3112 | unsigned = Val2.getConstantOperandVal(i: 1); |
3113 | |
3114 | // If the two extracted lanes are from the same place and adjacent, this |
3115 | // simplifies into a f32 lane move. |
3116 | if (Val1.getOperand(i: 0) == Val2.getOperand(i: 0) && ExtractLane2 % 2 == 0 && |
3117 | ExtractLane1 == ExtractLane2 + 1) { |
3118 | SDValue NewExt = CurDAG->getTargetExtractSubreg( |
3119 | SRIdx: ARM::ssub_0 + ExtractLane2 / 2, DL: dl, VT: MVT::f32, Operand: Val1.getOperand(i: 0)); |
3120 | SDValue NewIns = CurDAG->getTargetInsertSubreg( |
3121 | SRIdx: ARM::ssub_0 + Lane2 / 2, DL: dl, VT, Operand: Ins2.getOperand(i: 0), |
3122 | Subreg: NewExt); |
3123 | ReplaceUses(F: Ins1, T: NewIns); |
3124 | return true; |
3125 | } |
3126 | |
3127 | // Else v8i16 pattern of an extract and an insert, with a optional vmovx for |
3128 | // extracting odd lanes. |
3129 | if (VT == MVT::v8i16 && Subtarget->hasFullFP16()) { |
3130 | SDValue Inp1 = CurDAG->getTargetExtractSubreg( |
3131 | SRIdx: ARM::ssub_0 + ExtractLane1 / 2, DL: dl, VT: MVT::f32, Operand: Val1.getOperand(i: 0)); |
3132 | SDValue Inp2 = CurDAG->getTargetExtractSubreg( |
3133 | SRIdx: ARM::ssub_0 + ExtractLane2 / 2, DL: dl, VT: MVT::f32, Operand: Val2.getOperand(i: 0)); |
3134 | if (ExtractLane1 % 2 != 0) |
3135 | Inp1 = SDValue(CurDAG->getMachineNode(Opcode: ARM::VMOVH, dl, VT: MVT::f32, Op1: Inp1), 0); |
3136 | if (ExtractLane2 % 2 != 0) |
3137 | Inp2 = SDValue(CurDAG->getMachineNode(Opcode: ARM::VMOVH, dl, VT: MVT::f32, Op1: Inp2), 0); |
3138 | SDNode *VINS = CurDAG->getMachineNode(Opcode: ARM::VINSH, dl, VT: MVT::f32, Op1: Inp2, Op2: Inp1); |
3139 | SDValue NewIns = |
3140 | CurDAG->getTargetInsertSubreg(SRIdx: ARM::ssub_0 + Lane2 / 2, DL: dl, VT: MVT::v4f32, |
3141 | Operand: Ins2.getOperand(i: 0), Subreg: SDValue(VINS, 0)); |
3142 | ReplaceUses(F: Ins1, T: NewIns); |
3143 | return true; |
3144 | } |
3145 | } |
3146 | |
3147 | // The inserted values are not extracted - if they are f16 then insert them |
3148 | // directly using a VINS. |
3149 | if (VT == MVT::v8f16 && Subtarget->hasFullFP16()) { |
3150 | SDNode *VINS = CurDAG->getMachineNode(Opcode: ARM::VINSH, dl, VT: MVT::f32, Op1: Val2, Op2: Val1); |
3151 | SDValue NewIns = |
3152 | CurDAG->getTargetInsertSubreg(SRIdx: ARM::ssub_0 + Lane2 / 2, DL: dl, VT: MVT::v4f32, |
3153 | Operand: Ins2.getOperand(i: 0), Subreg: SDValue(VINS, 0)); |
3154 | ReplaceUses(F: Ins1, T: NewIns); |
3155 | return true; |
3156 | } |
3157 | |
3158 | return false; |
3159 | } |
3160 | |
3161 | bool ARMDAGToDAGISel::transformFixedFloatingPointConversion(SDNode *N, |
3162 | SDNode *FMul, |
3163 | bool IsUnsigned, |
3164 | bool FixedToFloat) { |
3165 | auto Type = N->getValueType(ResNo: 0); |
3166 | unsigned ScalarBits = Type.getScalarSizeInBits(); |
3167 | if (ScalarBits > 32) |
3168 | return false; |
3169 | |
3170 | SDNodeFlags FMulFlags = FMul->getFlags(); |
3171 | // The fixed-point vcvt and vcvt+vmul are not always equivalent if inf is |
3172 | // allowed in 16 bit unsigned floats |
3173 | if (ScalarBits == 16 && !FMulFlags.hasNoInfs() && IsUnsigned) |
3174 | return false; |
3175 | |
3176 | SDValue ImmNode = FMul->getOperand(Num: 1); |
3177 | SDValue VecVal = FMul->getOperand(Num: 0); |
3178 | if (VecVal->getOpcode() == ISD::UINT_TO_FP || |
3179 | VecVal->getOpcode() == ISD::SINT_TO_FP) |
3180 | VecVal = VecVal->getOperand(Num: 0); |
3181 | |
3182 | if (VecVal.getValueType().getScalarSizeInBits() != ScalarBits) |
3183 | return false; |
3184 | |
3185 | if (ImmNode.getOpcode() == ISD::BITCAST) { |
3186 | if (ImmNode.getValueType().getScalarSizeInBits() != ScalarBits) |
3187 | return false; |
3188 | ImmNode = ImmNode.getOperand(i: 0); |
3189 | } |
3190 | |
3191 | if (ImmNode.getValueType().getScalarSizeInBits() != ScalarBits) |
3192 | return false; |
3193 | |
3194 | APFloat ImmAPF(0.0f); |
3195 | switch (ImmNode.getOpcode()) { |
3196 | case ARMISD::VMOVIMM: |
3197 | case ARMISD::VDUP: { |
3198 | if (!isa<ConstantSDNode>(Val: ImmNode.getOperand(i: 0))) |
3199 | return false; |
3200 | unsigned Imm = ImmNode.getConstantOperandVal(i: 0); |
3201 | if (ImmNode.getOpcode() == ARMISD::VMOVIMM) |
3202 | Imm = ARM_AM::decodeVMOVModImm(ModImm: Imm, EltBits&: ScalarBits); |
3203 | ImmAPF = |
3204 | APFloat(ScalarBits == 32 ? APFloat::IEEEsingle() : APFloat::IEEEhalf(), |
3205 | APInt(ScalarBits, Imm)); |
3206 | break; |
3207 | } |
3208 | case ARMISD::VMOVFPIMM: { |
3209 | ImmAPF = APFloat(ARM_AM::getFPImmFloat(Imm: ImmNode.getConstantOperandVal(i: 0))); |
3210 | break; |
3211 | } |
3212 | default: |
3213 | return false; |
3214 | } |
3215 | |
3216 | // Where n is the number of fractional bits, multiplying by 2^n will convert |
3217 | // from float to fixed and multiplying by 2^-n will convert from fixed to |
3218 | // float. Taking log2 of the factor (after taking the inverse in the case of |
3219 | // float to fixed) will give n. |
3220 | APFloat ToConvert = ImmAPF; |
3221 | if (FixedToFloat) { |
3222 | if (!ImmAPF.getExactInverse(inv: &ToConvert)) |
3223 | return false; |
3224 | } |
3225 | APSInt Converted(64, false); |
3226 | bool IsExact; |
3227 | ToConvert.convertToInteger(Result&: Converted, RM: llvm::RoundingMode::NearestTiesToEven, |
3228 | IsExact: &IsExact); |
3229 | if (!IsExact || !Converted.isPowerOf2()) |
3230 | return false; |
3231 | |
3232 | unsigned FracBits = Converted.logBase2(); |
3233 | if (FracBits > ScalarBits) |
3234 | return false; |
3235 | |
3236 | SmallVector<SDValue, 3> Ops{ |
3237 | VecVal, CurDAG->getConstant(Val: FracBits, DL: SDLoc(N), VT: MVT::i32)}; |
3238 | AddEmptyMVEPredicateToOps(Ops, Loc: SDLoc(N), InactiveTy: Type); |
3239 | |
3240 | unsigned int Opcode; |
3241 | switch (ScalarBits) { |
3242 | case 16: |
3243 | if (FixedToFloat) |
3244 | Opcode = IsUnsigned ? ARM::MVE_VCVTf16u16_fix : ARM::MVE_VCVTf16s16_fix; |
3245 | else |
3246 | Opcode = IsUnsigned ? ARM::MVE_VCVTu16f16_fix : ARM::MVE_VCVTs16f16_fix; |
3247 | break; |
3248 | case 32: |
3249 | if (FixedToFloat) |
3250 | Opcode = IsUnsigned ? ARM::MVE_VCVTf32u32_fix : ARM::MVE_VCVTf32s32_fix; |
3251 | else |
3252 | Opcode = IsUnsigned ? ARM::MVE_VCVTu32f32_fix : ARM::MVE_VCVTs32f32_fix; |
3253 | break; |
3254 | default: |
3255 | llvm_unreachable("unexpected number of scalar bits" ); |
3256 | break; |
3257 | } |
3258 | |
3259 | ReplaceNode(F: N, T: CurDAG->getMachineNode(Opcode, dl: SDLoc(N), VT: Type, Ops)); |
3260 | return true; |
3261 | } |
3262 | |
3263 | bool ARMDAGToDAGISel::tryFP_TO_INT(SDNode *N, SDLoc dl) { |
3264 | // Transform a floating-point to fixed-point conversion to a VCVT |
3265 | if (!Subtarget->hasMVEFloatOps()) |
3266 | return false; |
3267 | EVT Type = N->getValueType(ResNo: 0); |
3268 | if (!Type.isVector()) |
3269 | return false; |
3270 | unsigned int ScalarBits = Type.getScalarSizeInBits(); |
3271 | |
3272 | bool IsUnsigned = N->getOpcode() == ISD::FP_TO_UINT || |
3273 | N->getOpcode() == ISD::FP_TO_UINT_SAT; |
3274 | SDNode *Node = N->getOperand(Num: 0).getNode(); |
3275 | |
3276 | // floating-point to fixed-point with one fractional bit gets turned into an |
3277 | // FP_TO_[U|S]INT(FADD (x, x)) rather than an FP_TO_[U|S]INT(FMUL (x, y)) |
3278 | if (Node->getOpcode() == ISD::FADD) { |
3279 | if (Node->getOperand(Num: 0) != Node->getOperand(Num: 1)) |
3280 | return false; |
3281 | SDNodeFlags Flags = Node->getFlags(); |
3282 | // The fixed-point vcvt and vcvt+vmul are not always equivalent if inf is |
3283 | // allowed in 16 bit unsigned floats |
3284 | if (ScalarBits == 16 && !Flags.hasNoInfs() && IsUnsigned) |
3285 | return false; |
3286 | |
3287 | unsigned Opcode; |
3288 | switch (ScalarBits) { |
3289 | case 16: |
3290 | Opcode = IsUnsigned ? ARM::MVE_VCVTu16f16_fix : ARM::MVE_VCVTs16f16_fix; |
3291 | break; |
3292 | case 32: |
3293 | Opcode = IsUnsigned ? ARM::MVE_VCVTu32f32_fix : ARM::MVE_VCVTs32f32_fix; |
3294 | break; |
3295 | } |
3296 | SmallVector<SDValue, 3> Ops{Node->getOperand(Num: 0), |
3297 | CurDAG->getConstant(Val: 1, DL: dl, VT: MVT::i32)}; |
3298 | AddEmptyMVEPredicateToOps(Ops, Loc: dl, InactiveTy: Type); |
3299 | |
3300 | ReplaceNode(F: N, T: CurDAG->getMachineNode(Opcode, dl, VT: Type, Ops)); |
3301 | return true; |
3302 | } |
3303 | |
3304 | if (Node->getOpcode() != ISD::FMUL) |
3305 | return false; |
3306 | |
3307 | return transformFixedFloatingPointConversion(N, FMul: Node, IsUnsigned, FixedToFloat: false); |
3308 | } |
3309 | |
3310 | bool ARMDAGToDAGISel::tryFMULFixed(SDNode *N, SDLoc dl) { |
3311 | // Transform a fixed-point to floating-point conversion to a VCVT |
3312 | if (!Subtarget->hasMVEFloatOps()) |
3313 | return false; |
3314 | auto Type = N->getValueType(ResNo: 0); |
3315 | if (!Type.isVector()) |
3316 | return false; |
3317 | |
3318 | auto LHS = N->getOperand(Num: 0); |
3319 | if (LHS.getOpcode() != ISD::SINT_TO_FP && LHS.getOpcode() != ISD::UINT_TO_FP) |
3320 | return false; |
3321 | |
3322 | return transformFixedFloatingPointConversion( |
3323 | N, FMul: N, IsUnsigned: LHS.getOpcode() == ISD::UINT_TO_FP, FixedToFloat: true); |
3324 | } |
3325 | |
3326 | bool ARMDAGToDAGISel::(SDNode *N, bool isSigned) { |
3327 | if (!Subtarget->hasV6T2Ops()) |
3328 | return false; |
3329 | |
3330 | unsigned Opc = isSigned |
3331 | ? (Subtarget->isThumb() ? ARM::t2SBFX : ARM::SBFX) |
3332 | : (Subtarget->isThumb() ? ARM::t2UBFX : ARM::UBFX); |
3333 | SDLoc dl(N); |
3334 | |
3335 | // For unsigned extracts, check for a shift right and mask |
3336 | unsigned And_imm = 0; |
3337 | if (N->getOpcode() == ISD::AND) { |
3338 | if (isOpcWithIntImmediate(N, Opc: ISD::AND, Imm&: And_imm)) { |
3339 | |
3340 | // The immediate is a mask of the low bits iff imm & (imm+1) == 0 |
3341 | if (And_imm & (And_imm + 1)) |
3342 | return false; |
3343 | |
3344 | unsigned Srl_imm = 0; |
3345 | if (isOpcWithIntImmediate(N: N->getOperand(Num: 0).getNode(), Opc: ISD::SRL, |
3346 | Imm&: Srl_imm)) { |
3347 | assert(Srl_imm > 0 && Srl_imm < 32 && "bad amount in shift node!" ); |
3348 | |
3349 | // Mask off the unnecessary bits of the AND immediate; normally |
3350 | // DAGCombine will do this, but that might not happen if |
3351 | // targetShrinkDemandedConstant chooses a different immediate. |
3352 | And_imm &= -1U >> Srl_imm; |
3353 | |
3354 | // Note: The width operand is encoded as width-1. |
3355 | unsigned Width = llvm::countr_one(Value: And_imm) - 1; |
3356 | unsigned LSB = Srl_imm; |
3357 | |
3358 | SDValue Reg0 = CurDAG->getRegister(Reg: 0, VT: MVT::i32); |
3359 | |
3360 | if ((LSB + Width + 1) == N->getValueType(ResNo: 0).getSizeInBits()) { |
3361 | // It's cheaper to use a right shift to extract the top bits. |
3362 | if (Subtarget->isThumb()) { |
3363 | Opc = isSigned ? ARM::t2ASRri : ARM::t2LSRri; |
3364 | SDValue Ops[] = { N->getOperand(Num: 0).getOperand(i: 0), |
3365 | CurDAG->getTargetConstant(Val: LSB, DL: dl, VT: MVT::i32), |
3366 | getAL(CurDAG, dl), Reg0, Reg0 }; |
3367 | CurDAG->SelectNodeTo(N, MachineOpc: Opc, VT: MVT::i32, Ops); |
3368 | return true; |
3369 | } |
3370 | |
3371 | // ARM models shift instructions as MOVsi with shifter operand. |
3372 | ARM_AM::ShiftOpc ShOpcVal = ARM_AM::getShiftOpcForNode(Opcode: ISD::SRL); |
3373 | SDValue ShOpc = |
3374 | CurDAG->getTargetConstant(Val: ARM_AM::getSORegOpc(ShOp: ShOpcVal, Imm: LSB), DL: dl, |
3375 | VT: MVT::i32); |
3376 | SDValue Ops[] = { N->getOperand(Num: 0).getOperand(i: 0), ShOpc, |
3377 | getAL(CurDAG, dl), Reg0, Reg0 }; |
3378 | CurDAG->SelectNodeTo(N, MachineOpc: ARM::MOVsi, VT: MVT::i32, Ops); |
3379 | return true; |
3380 | } |
3381 | |
3382 | assert(LSB + Width + 1 <= 32 && "Shouldn't create an invalid ubfx" ); |
3383 | SDValue Ops[] = { N->getOperand(Num: 0).getOperand(i: 0), |
3384 | CurDAG->getTargetConstant(Val: LSB, DL: dl, VT: MVT::i32), |
3385 | CurDAG->getTargetConstant(Val: Width, DL: dl, VT: MVT::i32), |
3386 | getAL(CurDAG, dl), Reg0 }; |
3387 | CurDAG->SelectNodeTo(N, MachineOpc: Opc, VT: MVT::i32, Ops); |
3388 | return true; |
3389 | } |
3390 | } |
3391 | return false; |
3392 | } |
3393 | |
3394 | // Otherwise, we're looking for a shift of a shift |
3395 | unsigned Shl_imm = 0; |
3396 | if (isOpcWithIntImmediate(N: N->getOperand(Num: 0).getNode(), Opc: ISD::SHL, Imm&: Shl_imm)) { |
3397 | assert(Shl_imm > 0 && Shl_imm < 32 && "bad amount in shift node!" ); |
3398 | unsigned Srl_imm = 0; |
3399 | if (isInt32Immediate(N: N->getOperand(Num: 1), Imm&: Srl_imm)) { |
3400 | assert(Srl_imm > 0 && Srl_imm < 32 && "bad amount in shift node!" ); |
3401 | // Note: The width operand is encoded as width-1. |
3402 | unsigned Width = 32 - Srl_imm - 1; |
3403 | int LSB = Srl_imm - Shl_imm; |
3404 | if (LSB < 0) |
3405 | return false; |
3406 | SDValue Reg0 = CurDAG->getRegister(Reg: 0, VT: MVT::i32); |
3407 | assert(LSB + Width + 1 <= 32 && "Shouldn't create an invalid ubfx" ); |
3408 | SDValue Ops[] = { N->getOperand(Num: 0).getOperand(i: 0), |
3409 | CurDAG->getTargetConstant(Val: LSB, DL: dl, VT: MVT::i32), |
3410 | CurDAG->getTargetConstant(Val: Width, DL: dl, VT: MVT::i32), |
3411 | getAL(CurDAG, dl), Reg0 }; |
3412 | CurDAG->SelectNodeTo(N, MachineOpc: Opc, VT: MVT::i32, Ops); |
3413 | return true; |
3414 | } |
3415 | } |
3416 | |
3417 | // Or we are looking for a shift of an and, with a mask operand |
3418 | if (isOpcWithIntImmediate(N: N->getOperand(Num: 0).getNode(), Opc: ISD::AND, Imm&: And_imm) && |
3419 | isShiftedMask_32(Value: And_imm)) { |
3420 | unsigned Srl_imm = 0; |
3421 | unsigned LSB = llvm::countr_zero(Val: And_imm); |
3422 | // Shift must be the same as the ands lsb |
3423 | if (isInt32Immediate(N: N->getOperand(Num: 1), Imm&: Srl_imm) && Srl_imm == LSB) { |
3424 | assert(Srl_imm > 0 && Srl_imm < 32 && "bad amount in shift node!" ); |
3425 | unsigned MSB = llvm::Log2_32(Value: And_imm); |
3426 | // Note: The width operand is encoded as width-1. |
3427 | unsigned Width = MSB - LSB; |
3428 | SDValue Reg0 = CurDAG->getRegister(Reg: 0, VT: MVT::i32); |
3429 | assert(Srl_imm + Width + 1 <= 32 && "Shouldn't create an invalid ubfx" ); |
3430 | SDValue Ops[] = { N->getOperand(Num: 0).getOperand(i: 0), |
3431 | CurDAG->getTargetConstant(Val: Srl_imm, DL: dl, VT: MVT::i32), |
3432 | CurDAG->getTargetConstant(Val: Width, DL: dl, VT: MVT::i32), |
3433 | getAL(CurDAG, dl), Reg0 }; |
3434 | CurDAG->SelectNodeTo(N, MachineOpc: Opc, VT: MVT::i32, Ops); |
3435 | return true; |
3436 | } |
3437 | } |
3438 | |
3439 | if (N->getOpcode() == ISD::SIGN_EXTEND_INREG) { |
3440 | unsigned Width = cast<VTSDNode>(Val: N->getOperand(Num: 1))->getVT().getSizeInBits(); |
3441 | unsigned LSB = 0; |
3442 | if (!isOpcWithIntImmediate(N: N->getOperand(Num: 0).getNode(), Opc: ISD::SRL, Imm&: LSB) && |
3443 | !isOpcWithIntImmediate(N: N->getOperand(Num: 0).getNode(), Opc: ISD::SRA, Imm&: LSB)) |
3444 | return false; |
3445 | |
3446 | if (LSB + Width > 32) |
3447 | return false; |
3448 | |
3449 | SDValue Reg0 = CurDAG->getRegister(Reg: 0, VT: MVT::i32); |
3450 | assert(LSB + Width <= 32 && "Shouldn't create an invalid ubfx" ); |
3451 | SDValue Ops[] = { N->getOperand(Num: 0).getOperand(i: 0), |
3452 | CurDAG->getTargetConstant(Val: LSB, DL: dl, VT: MVT::i32), |
3453 | CurDAG->getTargetConstant(Val: Width - 1, DL: dl, VT: MVT::i32), |
3454 | getAL(CurDAG, dl), Reg0 }; |
3455 | CurDAG->SelectNodeTo(N, MachineOpc: Opc, VT: MVT::i32, Ops); |
3456 | return true; |
3457 | } |
3458 | |
3459 | return false; |
3460 | } |
3461 | |
3462 | /// Target-specific DAG combining for ISD::SUB. |
3463 | /// Target-independent combining lowers SELECT_CC nodes of the form |
3464 | /// select_cc setg[ge] X, 0, X, -X |
3465 | /// select_cc setgt X, -1, X, -X |
3466 | /// select_cc setl[te] X, 0, -X, X |
3467 | /// select_cc setlt X, 1, -X, X |
3468 | /// which represent Integer ABS into: |
3469 | /// Y = sra (X, size(X)-1); sub (xor (X, Y), Y) |
3470 | /// ARM instruction selection detects the latter and matches it to |
3471 | /// ARM::ABS or ARM::t2ABS machine node. |
3472 | bool ARMDAGToDAGISel::tryABSOp(SDNode *N){ |
3473 | SDValue SUBSrc0 = N->getOperand(Num: 0); |
3474 | SDValue SUBSrc1 = N->getOperand(Num: 1); |
3475 | EVT VT = N->getValueType(ResNo: 0); |
3476 | |
3477 | if (Subtarget->isThumb1Only()) |
3478 | return false; |
3479 | |
3480 | if (SUBSrc0.getOpcode() != ISD::XOR || SUBSrc1.getOpcode() != ISD::SRA) |
3481 | return false; |
3482 | |
3483 | SDValue XORSrc0 = SUBSrc0.getOperand(i: 0); |
3484 | SDValue XORSrc1 = SUBSrc0.getOperand(i: 1); |
3485 | SDValue SRASrc0 = SUBSrc1.getOperand(i: 0); |
3486 | SDValue SRASrc1 = SUBSrc1.getOperand(i: 1); |
3487 | ConstantSDNode *SRAConstant = dyn_cast<ConstantSDNode>(Val&: SRASrc1); |
3488 | EVT XType = SRASrc0.getValueType(); |
3489 | unsigned Size = XType.getSizeInBits() - 1; |
3490 | |
3491 | if (XORSrc1 == SUBSrc1 && XORSrc0 == SRASrc0 && XType.isInteger() && |
3492 | SRAConstant != nullptr && Size == SRAConstant->getZExtValue()) { |
3493 | unsigned Opcode = Subtarget->isThumb2() ? ARM::t2ABS : ARM::ABS; |
3494 | CurDAG->SelectNodeTo(N, MachineOpc: Opcode, VT, Op1: XORSrc0); |
3495 | return true; |
3496 | } |
3497 | |
3498 | return false; |
3499 | } |
3500 | |
3501 | /// We've got special pseudo-instructions for these |
3502 | void ARMDAGToDAGISel::SelectCMP_SWAP(SDNode *N) { |
3503 | unsigned Opcode; |
3504 | EVT MemTy = cast<MemSDNode>(Val: N)->getMemoryVT(); |
3505 | if (MemTy == MVT::i8) |
3506 | Opcode = Subtarget->isThumb() ? ARM::tCMP_SWAP_8 : ARM::CMP_SWAP_8; |
3507 | else if (MemTy == MVT::i16) |
3508 | Opcode = Subtarget->isThumb() ? ARM::tCMP_SWAP_16 : ARM::CMP_SWAP_16; |
3509 | else if (MemTy == MVT::i32) |
3510 | Opcode = Subtarget->isThumb() ? ARM::tCMP_SWAP_32 : ARM::CMP_SWAP_32; |
3511 | else |
3512 | llvm_unreachable("Unknown AtomicCmpSwap type" ); |
3513 | |
3514 | SDValue Ops[] = {N->getOperand(Num: 1), N->getOperand(Num: 2), N->getOperand(Num: 3), |
3515 | N->getOperand(Num: 0)}; |
3516 | SDNode *CmpSwap = CurDAG->getMachineNode( |
3517 | Opcode, dl: SDLoc(N), |
3518 | VTs: CurDAG->getVTList(VT1: MVT::i32, VT2: MVT::i32, VT3: MVT::Other), Ops); |
3519 | |
3520 | MachineMemOperand *MemOp = cast<MemSDNode>(Val: N)->getMemOperand(); |
3521 | CurDAG->setNodeMemRefs(N: cast<MachineSDNode>(Val: CmpSwap), NewMemRefs: {MemOp}); |
3522 | |
3523 | ReplaceUses(F: SDValue(N, 0), T: SDValue(CmpSwap, 0)); |
3524 | ReplaceUses(F: SDValue(N, 1), T: SDValue(CmpSwap, 2)); |
3525 | CurDAG->RemoveDeadNode(N); |
3526 | } |
3527 | |
3528 | static std::optional<std::pair<unsigned, unsigned>> |
3529 | getContiguousRangeOfSetBits(const APInt &A) { |
3530 | unsigned FirstOne = A.getBitWidth() - A.countl_zero() - 1; |
3531 | unsigned LastOne = A.countr_zero(); |
3532 | if (A.popcount() != (FirstOne - LastOne + 1)) |
3533 | return std::nullopt; |
3534 | return std::make_pair(x&: FirstOne, y&: LastOne); |
3535 | } |
3536 | |
3537 | void ARMDAGToDAGISel::SelectCMPZ(SDNode *N, bool &SwitchEQNEToPLMI) { |
3538 | assert(N->getOpcode() == ARMISD::CMPZ); |
3539 | SwitchEQNEToPLMI = false; |
3540 | |
3541 | if (!Subtarget->isThumb()) |
3542 | // FIXME: Work out whether it is profitable to do this in A32 mode - LSL and |
3543 | // LSR don't exist as standalone instructions - they need the barrel shifter. |
3544 | return; |
3545 | |
3546 | // select (cmpz (and X, C), #0) -> (LSLS X) or (LSRS X) or (LSRS (LSLS X)) |
3547 | SDValue And = N->getOperand(Num: 0); |
3548 | if (!And->hasOneUse()) |
3549 | return; |
3550 | |
3551 | SDValue Zero = N->getOperand(Num: 1); |
3552 | if (!isNullConstant(V: Zero) || And->getOpcode() != ISD::AND) |
3553 | return; |
3554 | SDValue X = And.getOperand(i: 0); |
3555 | auto C = dyn_cast<ConstantSDNode>(Val: And.getOperand(i: 1)); |
3556 | |
3557 | if (!C) |
3558 | return; |
3559 | auto Range = getContiguousRangeOfSetBits(A: C->getAPIntValue()); |
3560 | if (!Range) |
3561 | return; |
3562 | |
3563 | // There are several ways to lower this: |
3564 | SDNode *NewN; |
3565 | SDLoc dl(N); |
3566 | |
3567 | auto EmitShift = [&](unsigned Opc, SDValue Src, unsigned Imm) -> SDNode* { |
3568 | if (Subtarget->isThumb2()) { |
3569 | Opc = (Opc == ARM::tLSLri) ? ARM::t2LSLri : ARM::t2LSRri; |
3570 | SDValue Ops[] = { Src, CurDAG->getTargetConstant(Val: Imm, DL: dl, VT: MVT::i32), |
3571 | getAL(CurDAG, dl), CurDAG->getRegister(Reg: 0, VT: MVT::i32), |
3572 | CurDAG->getRegister(Reg: 0, VT: MVT::i32) }; |
3573 | return CurDAG->getMachineNode(Opcode: Opc, dl, VT: MVT::i32, Ops); |
3574 | } else { |
3575 | SDValue Ops[] = {CurDAG->getRegister(Reg: ARM::CPSR, VT: MVT::i32), Src, |
3576 | CurDAG->getTargetConstant(Val: Imm, DL: dl, VT: MVT::i32), |
3577 | getAL(CurDAG, dl), CurDAG->getRegister(Reg: 0, VT: MVT::i32)}; |
3578 | return CurDAG->getMachineNode(Opcode: Opc, dl, VT: MVT::i32, Ops); |
3579 | } |
3580 | }; |
3581 | |
3582 | if (Range->second == 0) { |
3583 | // 1. Mask includes the LSB -> Simply shift the top N bits off |
3584 | NewN = EmitShift(ARM::tLSLri, X, 31 - Range->first); |
3585 | ReplaceNode(F: And.getNode(), T: NewN); |
3586 | } else if (Range->first == 31) { |
3587 | // 2. Mask includes the MSB -> Simply shift the bottom N bits off |
3588 | NewN = EmitShift(ARM::tLSRri, X, Range->second); |
3589 | ReplaceNode(F: And.getNode(), T: NewN); |
3590 | } else if (Range->first == Range->second) { |
3591 | // 3. Only one bit is set. We can shift this into the sign bit and use a |
3592 | // PL/MI comparison. This is not safe if CMPZ has multiple uses because |
3593 | // only one of them (the one currently being selected) will be switched |
3594 | // to use the new condition code. |
3595 | if (!N->hasOneUse()) |
3596 | return; |
3597 | NewN = EmitShift(ARM::tLSLri, X, 31 - Range->first); |
3598 | ReplaceNode(F: And.getNode(), T: NewN); |
3599 | |
3600 | SwitchEQNEToPLMI = true; |
3601 | } else if (!Subtarget->hasV6T2Ops()) { |
3602 | // 4. Do a double shift to clear bottom and top bits, but only in |
3603 | // thumb-1 mode as in thumb-2 we can use UBFX. |
3604 | NewN = EmitShift(ARM::tLSLri, X, 31 - Range->first); |
3605 | NewN = EmitShift(ARM::tLSRri, SDValue(NewN, 0), |
3606 | Range->second + (31 - Range->first)); |
3607 | ReplaceNode(F: And.getNode(), T: NewN); |
3608 | } |
3609 | } |
3610 | |
3611 | static unsigned getVectorShuffleOpcode(EVT VT, unsigned Opc64[3], |
3612 | unsigned Opc128[3]) { |
3613 | assert((VT.is64BitVector() || VT.is128BitVector()) && |
3614 | "Unexpected vector shuffle length" ); |
3615 | switch (VT.getScalarSizeInBits()) { |
3616 | default: |
3617 | llvm_unreachable("Unexpected vector shuffle element size" ); |
3618 | case 8: |
3619 | return VT.is64BitVector() ? Opc64[0] : Opc128[0]; |
3620 | case 16: |
3621 | return VT.is64BitVector() ? Opc64[1] : Opc128[1]; |
3622 | case 32: |
3623 | return VT.is64BitVector() ? Opc64[2] : Opc128[2]; |
3624 | } |
3625 | } |
3626 | |
3627 | void ARMDAGToDAGISel::Select(SDNode *N) { |
3628 | SDLoc dl(N); |
3629 | |
3630 | if (N->isMachineOpcode()) { |
3631 | N->setNodeId(-1); |
3632 | return; // Already selected. |
3633 | } |
3634 | |
3635 | switch (N->getOpcode()) { |
3636 | default: break; |
3637 | case ISD::STORE: { |
3638 | // For Thumb1, match an sp-relative store in C++. This is a little |
3639 | // unfortunate, but I don't think I can make the chain check work |
3640 | // otherwise. (The chain of the store has to be the same as the chain |
3641 | // of the CopyFromReg, or else we can't replace the CopyFromReg with |
3642 | // a direct reference to "SP".) |
3643 | // |
3644 | // This is only necessary on Thumb1 because Thumb1 sp-relative stores use |
3645 | // a different addressing mode from other four-byte stores. |
3646 | // |
3647 | // This pattern usually comes up with call arguments. |
3648 | StoreSDNode *ST = cast<StoreSDNode>(Val: N); |
3649 | SDValue Ptr = ST->getBasePtr(); |
3650 | if (Subtarget->isThumb1Only() && ST->isUnindexed()) { |
3651 | int RHSC = 0; |
3652 | if (Ptr.getOpcode() == ISD::ADD && |
3653 | isScaledConstantInRange(Node: Ptr.getOperand(i: 1), /*Scale=*/4, RangeMin: 0, RangeMax: 256, ScaledConstant&: RHSC)) |
3654 | Ptr = Ptr.getOperand(i: 0); |
3655 | |
3656 | if (Ptr.getOpcode() == ISD::CopyFromReg && |
3657 | cast<RegisterSDNode>(Val: Ptr.getOperand(i: 1))->getReg() == ARM::SP && |
3658 | Ptr.getOperand(i: 0) == ST->getChain()) { |
3659 | SDValue Ops[] = {ST->getValue(), |
3660 | CurDAG->getRegister(Reg: ARM::SP, VT: MVT::i32), |
3661 | CurDAG->getTargetConstant(Val: RHSC, DL: dl, VT: MVT::i32), |
3662 | getAL(CurDAG, dl), |
3663 | CurDAG->getRegister(Reg: 0, VT: MVT::i32), |
3664 | ST->getChain()}; |
3665 | MachineSDNode *ResNode = |
3666 | CurDAG->getMachineNode(Opcode: ARM::tSTRspi, dl, VT: MVT::Other, Ops); |
3667 | MachineMemOperand *MemOp = ST->getMemOperand(); |
3668 | CurDAG->setNodeMemRefs(N: cast<MachineSDNode>(Val: ResNode), NewMemRefs: {MemOp}); |
3669 | ReplaceNode(F: N, T: ResNode); |
3670 | return; |
3671 | } |
3672 | } |
3673 | break; |
3674 | } |
3675 | case ISD::WRITE_REGISTER: |
3676 | if (tryWriteRegister(N)) |
3677 | return; |
3678 | break; |
3679 | case ISD::READ_REGISTER: |
3680 | if (tryReadRegister(N)) |
3681 | return; |
3682 | break; |
3683 | case ISD::INLINEASM: |
3684 | case ISD::INLINEASM_BR: |
3685 | if (tryInlineAsm(N)) |
3686 | return; |
3687 | break; |
3688 | case ISD::SUB: |
3689 | // Select special operations if SUB node forms integer ABS pattern |
3690 | if (tryABSOp(N)) |
3691 | return; |
3692 | // Other cases are autogenerated. |
3693 | break; |
3694 | case ISD::Constant: { |
3695 | unsigned Val = N->getAsZExtVal(); |
3696 | // If we can't materialize the constant we need to use a literal pool |
3697 | if (ConstantMaterializationCost(Val, Subtarget) > 2 && |
3698 | !Subtarget->genExecuteOnly()) { |
3699 | SDValue CPIdx = CurDAG->getTargetConstantPool( |
3700 | C: ConstantInt::get(Ty: Type::getInt32Ty(C&: *CurDAG->getContext()), V: Val), |
3701 | VT: TLI->getPointerTy(DL: CurDAG->getDataLayout())); |
3702 | |
3703 | SDNode *ResNode; |
3704 | if (Subtarget->isThumb()) { |
3705 | SDValue Ops[] = { |
3706 | CPIdx, |
3707 | getAL(CurDAG, dl), |
3708 | CurDAG->getRegister(Reg: 0, VT: MVT::i32), |
3709 | CurDAG->getEntryNode() |
3710 | }; |
3711 | ResNode = CurDAG->getMachineNode(Opcode: ARM::tLDRpci, dl, VT1: MVT::i32, VT2: MVT::Other, |
3712 | Ops); |
3713 | } else { |
3714 | SDValue Ops[] = { |
3715 | CPIdx, |
3716 | CurDAG->getTargetConstant(Val: 0, DL: dl, VT: MVT::i32), |
3717 | getAL(CurDAG, dl), |
3718 | CurDAG->getRegister(Reg: 0, VT: MVT::i32), |
3719 | CurDAG->getEntryNode() |
3720 | }; |
3721 | ResNode = CurDAG->getMachineNode(Opcode: ARM::LDRcp, dl, VT1: MVT::i32, VT2: MVT::Other, |
3722 | Ops); |
3723 | } |
3724 | // Annotate the Node with memory operand information so that MachineInstr |
3725 | // queries work properly. This e.g. gives the register allocation the |
3726 | // required information for rematerialization. |
3727 | MachineFunction& MF = CurDAG->getMachineFunction(); |
3728 | MachineMemOperand *MemOp = |
3729 | MF.getMachineMemOperand(PtrInfo: MachinePointerInfo::getConstantPool(MF), |
3730 | F: MachineMemOperand::MOLoad, Size: 4, BaseAlignment: Align(4)); |
3731 | |
3732 | CurDAG->setNodeMemRefs(N: cast<MachineSDNode>(Val: ResNode), NewMemRefs: {MemOp}); |
3733 | |
3734 | ReplaceNode(F: N, T: ResNode); |
3735 | return; |
3736 | } |
3737 | |
3738 | // Other cases are autogenerated. |
3739 | break; |
3740 | } |
3741 | case ISD::FrameIndex: { |
3742 | // Selects to ADDri FI, 0 which in turn will become ADDri SP, imm. |
3743 | int FI = cast<FrameIndexSDNode>(Val: N)->getIndex(); |
3744 | SDValue TFI = CurDAG->getTargetFrameIndex( |
3745 | FI, VT: TLI->getPointerTy(DL: CurDAG->getDataLayout())); |
3746 | if (Subtarget->isThumb1Only()) { |
3747 | // Set the alignment of the frame object to 4, to avoid having to generate |
3748 | // more than one ADD |
3749 | MachineFrameInfo &MFI = MF->getFrameInfo(); |
3750 | if (MFI.getObjectAlign(ObjectIdx: FI) < Align(4)) |
3751 | MFI.setObjectAlignment(ObjectIdx: FI, Alignment: Align(4)); |
3752 | CurDAG->SelectNodeTo(N, MachineOpc: ARM::tADDframe, VT: MVT::i32, Op1: TFI, |
3753 | Op2: CurDAG->getTargetConstant(Val: 0, DL: dl, VT: MVT::i32)); |
3754 | return; |
3755 | } else { |
3756 | unsigned Opc = ((Subtarget->isThumb() && Subtarget->hasThumb2()) ? |
3757 | ARM::t2ADDri : ARM::ADDri); |
3758 | SDValue Ops[] = { TFI, CurDAG->getTargetConstant(Val: 0, DL: dl, VT: MVT::i32), |
3759 | getAL(CurDAG, dl), CurDAG->getRegister(Reg: 0, VT: MVT::i32), |
3760 | CurDAG->getRegister(Reg: 0, VT: MVT::i32) }; |
3761 | CurDAG->SelectNodeTo(N, MachineOpc: Opc, VT: MVT::i32, Ops); |
3762 | return; |
3763 | } |
3764 | } |
3765 | case ISD::INSERT_VECTOR_ELT: { |
3766 | if (tryInsertVectorElt(N)) |
3767 | return; |
3768 | break; |
3769 | } |
3770 | case ISD::SRL: |
3771 | if (tryV6T2BitfieldExtractOp(N, isSigned: false)) |
3772 | return; |
3773 | break; |
3774 | case ISD::SIGN_EXTEND_INREG: |
3775 | case ISD::SRA: |
3776 | if (tryV6T2BitfieldExtractOp(N, isSigned: true)) |
3777 | return; |
3778 | break; |
3779 | case ISD::FP_TO_UINT: |
3780 | case ISD::FP_TO_SINT: |
3781 | case ISD::FP_TO_UINT_SAT: |
3782 | case ISD::FP_TO_SINT_SAT: |
3783 | if (tryFP_TO_INT(N, dl)) |
3784 | return; |
3785 | break; |
3786 | case ISD::FMUL: |
3787 | if (tryFMULFixed(N, dl)) |
3788 | return; |
3789 | break; |
3790 | case ISD::MUL: |
3791 | if (Subtarget->isThumb1Only()) |
3792 | break; |
3793 | if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Val: N->getOperand(Num: 1))) { |
3794 | unsigned RHSV = C->getZExtValue(); |
3795 | if (!RHSV) break; |
3796 | if (isPowerOf2_32(Value: RHSV-1)) { // 2^n+1? |
3797 | unsigned ShImm = Log2_32(Value: RHSV-1); |
3798 | if (ShImm >= 32) |
3799 | break; |
3800 | SDValue V = N->getOperand(Num: 0); |
3801 | ShImm = ARM_AM::getSORegOpc(ShOp: ARM_AM::lsl, Imm: ShImm); |
3802 | SDValue ShImmOp = CurDAG->getTargetConstant(Val: ShImm, DL: dl, VT: MVT::i32); |
3803 | SDValue Reg0 = CurDAG->getRegister(Reg: 0, VT: MVT::i32); |
3804 | if (Subtarget->isThumb()) { |
3805 | SDValue Ops[] = { V, V, ShImmOp, getAL(CurDAG, dl), Reg0, Reg0 }; |
3806 | CurDAG->SelectNodeTo(N, MachineOpc: ARM::t2ADDrs, VT: MVT::i32, Ops); |
3807 | return; |
3808 | } else { |
3809 | SDValue Ops[] = { V, V, Reg0, ShImmOp, getAL(CurDAG, dl), Reg0, |
3810 | Reg0 }; |
3811 | CurDAG->SelectNodeTo(N, MachineOpc: ARM::ADDrsi, VT: MVT::i32, Ops); |
3812 | return; |
3813 | } |
3814 | } |
3815 | if (isPowerOf2_32(Value: RHSV+1)) { // 2^n-1? |
3816 | unsigned ShImm = Log2_32(Value: RHSV+1); |
3817 | if (ShImm >= 32) |
3818 | break; |
3819 | SDValue V = N->getOperand(Num: 0); |
3820 | ShImm = ARM_AM::getSORegOpc(ShOp: ARM_AM::lsl, Imm: ShImm); |
3821 | SDValue ShImmOp = CurDAG->getTargetConstant(Val: ShImm, DL: dl, VT: MVT::i32); |
3822 | SDValue Reg0 = CurDAG->getRegister(Reg: 0, VT: MVT::i32); |
3823 | if (Subtarget->isThumb()) { |
3824 | SDValue Ops[] = { V, V, ShImmOp, getAL(CurDAG, dl), Reg0, Reg0 }; |
3825 | CurDAG->SelectNodeTo(N, MachineOpc: ARM::t2RSBrs, VT: MVT::i32, Ops); |
3826 | return; |
3827 | } else { |
3828 | SDValue Ops[] = { V, V, Reg0, ShImmOp, getAL(CurDAG, dl), Reg0, |
3829 | Reg0 }; |
3830 | CurDAG->SelectNodeTo(N, MachineOpc: ARM::RSBrsi, VT: MVT::i32, Ops); |
3831 | return; |
3832 | } |
3833 | } |
3834 | } |
3835 | break; |
3836 | case ISD::AND: { |
3837 | // Check for unsigned bitfield extract |
3838 | if (tryV6T2BitfieldExtractOp(N, isSigned: false)) |
3839 | return; |
3840 | |
3841 | // If an immediate is used in an AND node, it is possible that the immediate |
3842 | // can be more optimally materialized when negated. If this is the case we |
3843 | // can negate the immediate and use a BIC instead. |
3844 | auto *N1C = dyn_cast<ConstantSDNode>(Val: N->getOperand(Num: 1)); |
3845 | if (N1C && N1C->hasOneUse() && Subtarget->isThumb()) { |
3846 | uint32_t Imm = (uint32_t) N1C->getZExtValue(); |
3847 | |
3848 | // In Thumb2 mode, an AND can take a 12-bit immediate. If this |
3849 | // immediate can be negated and fit in the immediate operand of |
3850 | // a t2BIC, don't do any manual transform here as this can be |
3851 | // handled by the generic ISel machinery. |
3852 | bool PreferImmediateEncoding = |
3853 | Subtarget->hasThumb2() && (is_t2_so_imm(Imm) || is_t2_so_imm_not(Imm)); |
3854 | if (!PreferImmediateEncoding && |
3855 | ConstantMaterializationCost(Val: Imm, Subtarget) > |
3856 | ConstantMaterializationCost(Val: ~Imm, Subtarget)) { |
3857 | // The current immediate costs more to materialize than a negated |
3858 | // immediate, so negate the immediate and use a BIC. |
3859 | SDValue NewImm = CurDAG->getConstant(Val: ~Imm, DL: dl, VT: MVT::i32); |
3860 | // If the new constant didn't exist before, reposition it in the topological |
3861 | // ordering so it is just before N. Otherwise, don't touch its location. |
3862 | if (NewImm->getNodeId() == -1) |
3863 | CurDAG->RepositionNode(Position: N->getIterator(), N: NewImm.getNode()); |
3864 | |
3865 | if (!Subtarget->hasThumb2()) { |
3866 | SDValue Ops[] = {CurDAG->getRegister(Reg: ARM::CPSR, VT: MVT::i32), |
3867 | N->getOperand(Num: 0), NewImm, getAL(CurDAG, dl), |
3868 | CurDAG->getRegister(Reg: 0, VT: MVT::i32)}; |
3869 | ReplaceNode(F: N, T: CurDAG->getMachineNode(Opcode: ARM::tBIC, dl, VT: MVT::i32, Ops)); |
3870 | return; |
3871 | } else { |
3872 | SDValue Ops[] = {N->getOperand(Num: 0), NewImm, getAL(CurDAG, dl), |
3873 | CurDAG->getRegister(Reg: 0, VT: MVT::i32), |
3874 | CurDAG->getRegister(Reg: 0, VT: MVT::i32)}; |
3875 | ReplaceNode(F: N, |
3876 | T: CurDAG->getMachineNode(Opcode: ARM::t2BICrr, dl, VT: MVT::i32, Ops)); |
3877 | return; |
3878 | } |
3879 | } |
3880 | } |
3881 | |
3882 | // (and (or x, c2), c1) and top 16-bits of c1 and c2 match, lower 16-bits |
3883 | // of c1 are 0xffff, and lower 16-bit of c2 are 0. That is, the top 16-bits |
3884 | // are entirely contributed by c2 and lower 16-bits are entirely contributed |
3885 | // by x. That's equal to (or (and x, 0xffff), (and c1, 0xffff0000)). |
3886 | // Select it to: "movt x, ((c1 & 0xffff) >> 16) |
3887 | EVT VT = N->getValueType(ResNo: 0); |
3888 | if (VT != MVT::i32) |
3889 | break; |
3890 | unsigned Opc = (Subtarget->isThumb() && Subtarget->hasThumb2()) |
3891 | ? ARM::t2MOVTi16 |
3892 | : (Subtarget->hasV6T2Ops() ? ARM::MOVTi16 : 0); |
3893 | if (!Opc) |
3894 | break; |
3895 | SDValue N0 = N->getOperand(Num: 0), N1 = N->getOperand(Num: 1); |
3896 | N1C = dyn_cast<ConstantSDNode>(Val&: N1); |
3897 | if (!N1C) |
3898 | break; |
3899 | if (N0.getOpcode() == ISD::OR && N0.getNode()->hasOneUse()) { |
3900 | SDValue N2 = N0.getOperand(i: 1); |
3901 | ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(Val&: N2); |
3902 | if (!N2C) |
3903 | break; |
3904 | unsigned N1CVal = N1C->getZExtValue(); |
3905 | unsigned N2CVal = N2C->getZExtValue(); |
3906 | if ((N1CVal & 0xffff0000U) == (N2CVal & 0xffff0000U) && |
3907 | (N1CVal & 0xffffU) == 0xffffU && |
3908 | (N2CVal & 0xffffU) == 0x0U) { |
3909 | SDValue Imm16 = CurDAG->getTargetConstant(Val: (N2CVal & 0xFFFF0000U) >> 16, |
3910 | DL: dl, VT: MVT::i32); |
3911 | SDValue Ops[] = { N0.getOperand(i: 0), Imm16, |
3912 | getAL(CurDAG, dl), CurDAG->getRegister(Reg: 0, VT: MVT::i32) }; |
3913 | ReplaceNode(F: N, T: CurDAG->getMachineNode(Opcode: Opc, dl, VT, Ops)); |
3914 | return; |
3915 | } |
3916 | } |
3917 | |
3918 | break; |
3919 | } |
3920 | case ARMISD::UMAAL: { |
3921 | unsigned Opc = Subtarget->isThumb() ? ARM::t2UMAAL : ARM::UMAAL; |
3922 | SDValue Ops[] = { N->getOperand(Num: 0), N->getOperand(Num: 1), |
3923 | N->getOperand(Num: 2), N->getOperand(Num: 3), |
3924 | getAL(CurDAG, dl), |
3925 | CurDAG->getRegister(Reg: 0, VT: MVT::i32) }; |
3926 | ReplaceNode(F: N, T: CurDAG->getMachineNode(Opcode: Opc, dl, VT1: MVT::i32, VT2: MVT::i32, Ops)); |
3927 | return; |
3928 | } |
3929 | case ARMISD::UMLAL:{ |
3930 | if (Subtarget->isThumb()) { |
3931 | SDValue Ops[] = { N->getOperand(Num: 0), N->getOperand(Num: 1), N->getOperand(Num: 2), |
3932 | N->getOperand(Num: 3), getAL(CurDAG, dl), |
3933 | CurDAG->getRegister(Reg: 0, VT: MVT::i32)}; |
3934 | ReplaceNode( |
3935 | F: N, T: CurDAG->getMachineNode(Opcode: ARM::t2UMLAL, dl, VT1: MVT::i32, VT2: MVT::i32, Ops)); |
3936 | return; |
3937 | }else{ |
3938 | SDValue Ops[] = { N->getOperand(Num: 0), N->getOperand(Num: 1), N->getOperand(Num: 2), |
3939 | N->getOperand(Num: 3), getAL(CurDAG, dl), |
3940 | CurDAG->getRegister(Reg: 0, VT: MVT::i32), |
3941 | CurDAG->getRegister(Reg: 0, VT: MVT::i32) }; |
3942 | ReplaceNode(F: N, T: CurDAG->getMachineNode( |
3943 | Opcode: Subtarget->hasV6Ops() ? ARM::UMLAL : ARM::UMLALv5, dl, |
3944 | VT1: MVT::i32, VT2: MVT::i32, Ops)); |
3945 | return; |
3946 | } |
3947 | } |
3948 | case ARMISD::SMLAL:{ |
3949 | if (Subtarget->isThumb()) { |
3950 | SDValue Ops[] = { N->getOperand(Num: 0), N->getOperand(Num: 1), N->getOperand(Num: 2), |
3951 | N->getOperand(Num: 3), getAL(CurDAG, dl), |
3952 | CurDAG->getRegister(Reg: 0, VT: MVT::i32)}; |
3953 | ReplaceNode( |
3954 | F: N, T: CurDAG->getMachineNode(Opcode: ARM::t2SMLAL, dl, VT1: MVT::i32, VT2: MVT::i32, Ops)); |
3955 | return; |
3956 | }else{ |
3957 | SDValue Ops[] = { N->getOperand(Num: 0), N->getOperand(Num: 1), N->getOperand(Num: 2), |
3958 | N->getOperand(Num: 3), getAL(CurDAG, dl), |
3959 | CurDAG->getRegister(Reg: 0, VT: MVT::i32), |
3960 | CurDAG->getRegister(Reg: 0, VT: MVT::i32) }; |
3961 | ReplaceNode(F: N, T: CurDAG->getMachineNode( |
3962 | Opcode: Subtarget->hasV6Ops() ? ARM::SMLAL : ARM::SMLALv5, dl, |
3963 | VT1: MVT::i32, VT2: MVT::i32, Ops)); |
3964 | return; |
3965 | } |
3966 | } |
3967 | case ARMISD::SUBE: { |
3968 | if (!Subtarget->hasV6Ops() || !Subtarget->hasDSP()) |
3969 | break; |
3970 | // Look for a pattern to match SMMLS |
3971 | // (sube a, (smul_loHi a, b), (subc 0, (smul_LOhi(a, b)))) |
3972 | if (N->getOperand(Num: 1).getOpcode() != ISD::SMUL_LOHI || |
3973 | N->getOperand(Num: 2).getOpcode() != ARMISD::SUBC || |
3974 | !SDValue(N, 1).use_empty()) |
3975 | break; |
3976 | |
3977 | if (Subtarget->isThumb()) |
3978 | assert(Subtarget->hasThumb2() && |
3979 | "This pattern should not be generated for Thumb" ); |
3980 | |
3981 | SDValue SmulLoHi = N->getOperand(Num: 1); |
3982 | SDValue Subc = N->getOperand(Num: 2); |
3983 | SDValue Zero = Subc.getOperand(i: 0); |
3984 | |
3985 | if (!isNullConstant(V: Zero) || Subc.getOperand(i: 1) != SmulLoHi.getValue(R: 0) || |
3986 | N->getOperand(Num: 1) != SmulLoHi.getValue(R: 1) || |
3987 | N->getOperand(Num: 2) != Subc.getValue(R: 1)) |
3988 | break; |
3989 | |
3990 | unsigned Opc = Subtarget->isThumb2() ? ARM::t2SMMLS : ARM::SMMLS; |
3991 | SDValue Ops[] = { SmulLoHi.getOperand(i: 0), SmulLoHi.getOperand(i: 1), |
3992 | N->getOperand(Num: 0), getAL(CurDAG, dl), |
3993 | CurDAG->getRegister(Reg: 0, VT: MVT::i32) }; |
3994 | ReplaceNode(F: N, T: CurDAG->getMachineNode(Opcode: Opc, dl, VT: MVT::i32, Ops)); |
3995 | return; |
3996 | } |
3997 | case ISD::LOAD: { |
3998 | if (Subtarget->hasMVEIntegerOps() && tryMVEIndexedLoad(N)) |
3999 | return; |
4000 | if (Subtarget->isThumb() && Subtarget->hasThumb2()) { |
4001 | if (tryT2IndexedLoad(N)) |
4002 | return; |
4003 | } else if (Subtarget->isThumb()) { |
4004 | if (tryT1IndexedLoad(N)) |
4005 | return; |
4006 | } else if (tryARMIndexedLoad(N)) |
4007 | return; |
4008 | // Other cases are autogenerated. |
4009 | break; |
4010 | } |
4011 | case ISD::MLOAD: |
4012 | if (Subtarget->hasMVEIntegerOps() && tryMVEIndexedLoad(N)) |
4013 | return; |
4014 | // Other cases are autogenerated. |
4015 | break; |
4016 | case ARMISD::WLSSETUP: { |
4017 | SDNode *New = CurDAG->getMachineNode(Opcode: ARM::t2WhileLoopSetup, dl, VT: MVT::i32, |
4018 | Op1: N->getOperand(Num: 0)); |
4019 | ReplaceUses(F: N, T: New); |
4020 | CurDAG->RemoveDeadNode(N); |
4021 | return; |
4022 | } |
4023 | case ARMISD::WLS: { |
4024 | SDNode *New = CurDAG->getMachineNode(Opcode: ARM::t2WhileLoopStart, dl, VT: MVT::Other, |
4025 | Op1: N->getOperand(Num: 1), Op2: N->getOperand(Num: 2), |
4026 | Op3: N->getOperand(Num: 0)); |
4027 | ReplaceUses(F: N, T: New); |
4028 | CurDAG->RemoveDeadNode(N); |
4029 | return; |
4030 | } |
4031 | case ARMISD::LE: { |
4032 | SDValue Ops[] = { N->getOperand(Num: 1), |
4033 | N->getOperand(Num: 2), |
4034 | N->getOperand(Num: 0) }; |
4035 | unsigned Opc = ARM::t2LoopEnd; |
4036 | SDNode *New = CurDAG->getMachineNode(Opcode: Opc, dl, VT: MVT::Other, Ops); |
4037 | ReplaceUses(F: N, T: New); |
4038 | CurDAG->RemoveDeadNode(N); |
4039 | return; |
4040 | } |
4041 | case ARMISD::LDRD: { |
4042 | if (Subtarget->isThumb2()) |
4043 | break; // TableGen handles isel in this case. |
4044 | SDValue Base, RegOffset, ImmOffset; |
4045 | const SDValue &Chain = N->getOperand(Num: 0); |
4046 | const SDValue &Addr = N->getOperand(Num: 1); |
4047 | SelectAddrMode3(N: Addr, Base, Offset&: RegOffset, Opc&: ImmOffset); |
4048 | if (RegOffset != CurDAG->getRegister(Reg: 0, VT: MVT::i32)) { |
4049 | // The register-offset variant of LDRD mandates that the register |
4050 | // allocated to RegOffset is not reused in any of the remaining operands. |
4051 | // This restriction is currently not enforced. Therefore emitting this |
4052 | // variant is explicitly avoided. |
4053 | Base = Addr; |
4054 | RegOffset = CurDAG->getRegister(Reg: 0, VT: MVT::i32); |
4055 | } |
4056 | SDValue Ops[] = {Base, RegOffset, ImmOffset, Chain}; |
4057 | SDNode *New = CurDAG->getMachineNode(Opcode: ARM::LOADDUAL, dl, |
4058 | ResultTys: {MVT::Untyped, MVT::Other}, Ops); |
4059 | SDValue Lo = CurDAG->getTargetExtractSubreg(SRIdx: ARM::gsub_0, DL: dl, VT: MVT::i32, |
4060 | Operand: SDValue(New, 0)); |
4061 | SDValue Hi = CurDAG->getTargetExtractSubreg(SRIdx: ARM::gsub_1, DL: dl, VT: MVT::i32, |
4062 | Operand: SDValue(New, 0)); |
4063 | transferMemOperands(N, Result: New); |
4064 | ReplaceUses(F: SDValue(N, 0), T: Lo); |
4065 | ReplaceUses(F: SDValue(N, 1), T: Hi); |
4066 | ReplaceUses(F: SDValue(N, 2), T: SDValue(New, 1)); |
4067 | CurDAG->RemoveDeadNode(N); |
4068 | return; |
4069 | } |
4070 | case ARMISD::STRD: { |
4071 | if (Subtarget->isThumb2()) |
4072 | break; // TableGen handles isel in this case. |
4073 | SDValue Base, RegOffset, ImmOffset; |
4074 | const SDValue &Chain = N->getOperand(Num: 0); |
4075 | const SDValue &Addr = N->getOperand(Num: 3); |
4076 | SelectAddrMode3(N: Addr, Base, Offset&: RegOffset, Opc&: ImmOffset); |
4077 | if (RegOffset != CurDAG->getRegister(Reg: 0, VT: MVT::i32)) { |
4078 | // The register-offset variant of STRD mandates that the register |
4079 | // allocated to RegOffset is not reused in any of the remaining operands. |
4080 | // This restriction is currently not enforced. Therefore emitting this |
4081 | // variant is explicitly avoided. |
4082 | Base = Addr; |
4083 | RegOffset = CurDAG->getRegister(Reg: 0, VT: MVT::i32); |
4084 | } |
4085 | SDNode *RegPair = |
4086 | createGPRPairNode(VT: MVT::Untyped, V0: N->getOperand(Num: 1), V1: N->getOperand(Num: 2)); |
4087 | SDValue Ops[] = {SDValue(RegPair, 0), Base, RegOffset, ImmOffset, Chain}; |
4088 | SDNode *New = CurDAG->getMachineNode(Opcode: ARM::STOREDUAL, dl, VT: MVT::Other, Ops); |
4089 | transferMemOperands(N, Result: New); |
4090 | ReplaceUses(F: SDValue(N, 0), T: SDValue(New, 0)); |
4091 | CurDAG->RemoveDeadNode(N); |
4092 | return; |
4093 | } |
4094 | case ARMISD::LOOP_DEC: { |
4095 | SDValue Ops[] = { N->getOperand(Num: 1), |
4096 | N->getOperand(Num: 2), |
4097 | N->getOperand(Num: 0) }; |
4098 | SDNode *Dec = |
4099 | CurDAG->getMachineNode(Opcode: ARM::t2LoopDec, dl, |
4100 | VTs: CurDAG->getVTList(VT1: MVT::i32, VT2: MVT::Other), Ops); |
4101 | ReplaceUses(F: N, T: Dec); |
4102 | CurDAG->RemoveDeadNode(N); |
4103 | return; |
4104 | } |
4105 | case ARMISD::BRCOND: { |
4106 | // Pattern: (ARMbrcond:void (bb:Other):$dst, (imm:i32):$cc) |
4107 | // Emits: (Bcc:void (bb:Other):$dst, (imm:i32):$cc) |
4108 | // Pattern complexity = 6 cost = 1 size = 0 |
4109 | |
4110 | // Pattern: (ARMbrcond:void (bb:Other):$dst, (imm:i32):$cc) |
4111 | // Emits: (tBcc:void (bb:Other):$dst, (imm:i32):$cc) |
4112 | // Pattern complexity = 6 cost = 1 size = 0 |
4113 | |
4114 | // Pattern: (ARMbrcond:void (bb:Other):$dst, (imm:i32):$cc) |
4115 | // Emits: (t2Bcc:void (bb:Other):$dst, (imm:i32):$cc) |
4116 | // Pattern complexity = 6 cost = 1 size = 0 |
4117 | |
4118 | unsigned Opc = Subtarget->isThumb() ? |
4119 | ((Subtarget->hasThumb2()) ? ARM::t2Bcc : ARM::tBcc) : ARM::Bcc; |
4120 | SDValue Chain = N->getOperand(Num: 0); |
4121 | SDValue N1 = N->getOperand(Num: 1); |
4122 | SDValue N2 = N->getOperand(Num: 2); |
4123 | SDValue Flags = N->getOperand(Num: 3); |
4124 | assert(N1.getOpcode() == ISD::BasicBlock); |
4125 | assert(N2.getOpcode() == ISD::Constant); |
4126 | |
4127 | unsigned CC = (unsigned)N2->getAsZExtVal(); |
4128 | |
4129 | if (Flags.getOpcode() == ARMISD::CMPZ) { |
4130 | if (Flags.getOperand(i: 0).getOpcode() == ISD::INTRINSIC_W_CHAIN) { |
4131 | SDValue Int = Flags.getOperand(i: 0); |
4132 | uint64_t ID = Int->getConstantOperandVal(Num: 1); |
4133 | |
4134 | // Handle low-overhead loops. |
4135 | if (ID == Intrinsic::loop_decrement_reg) { |
4136 | SDValue Elements = Int.getOperand(i: 2); |
4137 | SDValue Size = CurDAG->getTargetConstant(Val: Int.getConstantOperandVal(i: 3), |
4138 | DL: dl, VT: MVT::i32); |
4139 | |
4140 | SDValue Args[] = { Elements, Size, Int.getOperand(i: 0) }; |
4141 | SDNode *LoopDec = |
4142 | CurDAG->getMachineNode(Opcode: ARM::t2LoopDec, dl, |
4143 | VTs: CurDAG->getVTList(VT1: MVT::i32, VT2: MVT::Other), |
4144 | Ops: Args); |
4145 | ReplaceUses(F: Int.getNode(), T: LoopDec); |
4146 | |
4147 | SDValue EndArgs[] = { SDValue(LoopDec, 0), N1, Chain }; |
4148 | SDNode *LoopEnd = |
4149 | CurDAG->getMachineNode(Opcode: ARM::t2LoopEnd, dl, VT: MVT::Other, Ops: EndArgs); |
4150 | |
4151 | ReplaceUses(F: N, T: LoopEnd); |
4152 | CurDAG->RemoveDeadNode(N); |
4153 | CurDAG->RemoveDeadNode(N: Flags.getNode()); |
4154 | CurDAG->RemoveDeadNode(N: Int.getNode()); |
4155 | return; |
4156 | } |
4157 | } |
4158 | |
4159 | bool SwitchEQNEToPLMI; |
4160 | SelectCMPZ(N: Flags.getNode(), SwitchEQNEToPLMI); |
4161 | Flags = N->getOperand(Num: 3); |
4162 | |
4163 | if (SwitchEQNEToPLMI) { |
4164 | switch ((ARMCC::CondCodes)CC) { |
4165 | default: llvm_unreachable("CMPZ must be either NE or EQ!" ); |
4166 | case ARMCC::NE: |
4167 | CC = (unsigned)ARMCC::MI; |
4168 | break; |
4169 | case ARMCC::EQ: |
4170 | CC = (unsigned)ARMCC::PL; |
4171 | break; |
4172 | } |
4173 | } |
4174 | } |
4175 | |
4176 | SDValue Tmp2 = CurDAG->getTargetConstant(Val: CC, DL: dl, VT: MVT::i32); |
4177 | Chain = CurDAG->getCopyToReg(Chain, dl, Reg: ARM::CPSR, N: Flags, Glue: SDValue()); |
4178 | SDValue Ops[] = {N1, Tmp2, CurDAG->getRegister(Reg: ARM::CPSR, VT: MVT::i32), Chain, |
4179 | Chain.getValue(R: 1)}; |
4180 | CurDAG->SelectNodeTo(N, MachineOpc: Opc, VT: MVT::Other, Ops); |
4181 | return; |
4182 | } |
4183 | |
4184 | case ARMISD::CMPZ: { |
4185 | // select (CMPZ X, #-C) -> (CMPZ (ADDS X, #C), #0) |
4186 | // This allows us to avoid materializing the expensive negative constant. |
4187 | // The CMPZ #0 is useless and will be peepholed away but we need to keep |
4188 | // it for its flags output. |
4189 | SDValue X = N->getOperand(Num: 0); |
4190 | auto *C = dyn_cast<ConstantSDNode>(Val: N->getOperand(Num: 1).getNode()); |
4191 | if (C && C->getSExtValue() < 0 && Subtarget->isThumb()) { |
4192 | int64_t Addend = -C->getSExtValue(); |
4193 | |
4194 | SDNode *Add = nullptr; |
4195 | // ADDS can be better than CMN if the immediate fits in a |
4196 | // 16-bit ADDS, which means either [0,256) for tADDi8 or [0,8) for tADDi3. |
4197 | // Outside that range we can just use a CMN which is 32-bit but has a |
4198 | // 12-bit immediate range. |
4199 | if (Addend < 1<<8) { |
4200 | if (Subtarget->isThumb2()) { |
4201 | SDValue Ops[] = { X, CurDAG->getTargetConstant(Val: Addend, DL: dl, VT: MVT::i32), |
4202 | getAL(CurDAG, dl), CurDAG->getRegister(Reg: 0, VT: MVT::i32), |
4203 | CurDAG->getRegister(Reg: 0, VT: MVT::i32) }; |
4204 | Add = CurDAG->getMachineNode(Opcode: ARM::t2ADDri, dl, VT: MVT::i32, Ops); |
4205 | } else { |
4206 | unsigned Opc = (Addend < 1<<3) ? ARM::tADDi3 : ARM::tADDi8; |
4207 | SDValue Ops[] = {CurDAG->getRegister(Reg: ARM::CPSR, VT: MVT::i32), X, |
4208 | CurDAG->getTargetConstant(Val: Addend, DL: dl, VT: MVT::i32), |
4209 | getAL(CurDAG, dl), CurDAG->getRegister(Reg: 0, VT: MVT::i32)}; |
4210 | Add = CurDAG->getMachineNode(Opcode: Opc, dl, VT: MVT::i32, Ops); |
4211 | } |
4212 | } |
4213 | if (Add) { |
4214 | SDValue Ops2[] = {SDValue(Add, 0), CurDAG->getConstant(Val: 0, DL: dl, VT: MVT::i32)}; |
4215 | CurDAG->MorphNodeTo(N, Opc: ARMISD::CMPZ, VTs: N->getVTList(), Ops: Ops2); |
4216 | } |
4217 | } |
4218 | // Other cases are autogenerated. |
4219 | break; |
4220 | } |
4221 | |
4222 | case ARMISD::CMOV: { |
4223 | SDValue Flags = N->getOperand(Num: 3); |
4224 | |
4225 | if (Flags.getOpcode() == ARMISD::CMPZ) { |
4226 | bool SwitchEQNEToPLMI; |
4227 | SelectCMPZ(N: Flags.getNode(), SwitchEQNEToPLMI); |
4228 | |
4229 | if (SwitchEQNEToPLMI) { |
4230 | SDValue ARMcc = N->getOperand(Num: 2); |
4231 | ARMCC::CondCodes CC = (ARMCC::CondCodes)ARMcc->getAsZExtVal(); |
4232 | |
4233 | switch (CC) { |
4234 | default: llvm_unreachable("CMPZ must be either NE or EQ!" ); |
4235 | case ARMCC::NE: |
4236 | CC = ARMCC::MI; |
4237 | break; |
4238 | case ARMCC::EQ: |
4239 | CC = ARMCC::PL; |
4240 | break; |
4241 | } |
4242 | SDValue NewARMcc = CurDAG->getConstant(Val: (unsigned)CC, DL: dl, VT: MVT::i32); |
4243 | SDValue Ops[] = {N->getOperand(Num: 0), N->getOperand(Num: 1), NewARMcc, |
4244 | N->getOperand(Num: 3)}; |
4245 | CurDAG->MorphNodeTo(N, Opc: ARMISD::CMOV, VTs: N->getVTList(), Ops); |
4246 | } |
4247 | } |
4248 | // Other cases are autogenerated. |
4249 | break; |
4250 | } |
4251 | case ARMISD::VZIP: { |
4252 | EVT VT = N->getValueType(ResNo: 0); |
4253 | // vzip.32 Dd, Dm is a pseudo-instruction expanded to vtrn.32 Dd, Dm. |
4254 | unsigned Opc64[] = {ARM::VZIPd8, ARM::VZIPd16, ARM::VTRNd32}; |
4255 | unsigned Opc128[] = {ARM::VZIPq8, ARM::VZIPq16, ARM::VZIPq32}; |
4256 | unsigned Opc = getVectorShuffleOpcode(VT, Opc64, Opc128); |
4257 | SDValue Pred = getAL(CurDAG, dl); |
4258 | SDValue PredReg = CurDAG->getRegister(Reg: 0, VT: MVT::i32); |
4259 | SDValue Ops[] = {N->getOperand(Num: 0), N->getOperand(Num: 1), Pred, PredReg}; |
4260 | ReplaceNode(F: N, T: CurDAG->getMachineNode(Opcode: Opc, dl, VT1: VT, VT2: VT, Ops)); |
4261 | return; |
4262 | } |
4263 | case ARMISD::VUZP: { |
4264 | EVT VT = N->getValueType(ResNo: 0); |
4265 | // vuzp.32 Dd, Dm is a pseudo-instruction expanded to vtrn.32 Dd, Dm. |
4266 | unsigned Opc64[] = {ARM::VUZPd8, ARM::VUZPd16, ARM::VTRNd32}; |
4267 | unsigned Opc128[] = {ARM::VUZPq8, ARM::VUZPq16, ARM::VUZPq32}; |
4268 | unsigned Opc = getVectorShuffleOpcode(VT, Opc64, Opc128); |
4269 | SDValue Pred = getAL(CurDAG, dl); |
4270 | SDValue PredReg = CurDAG->getRegister(Reg: 0, VT: MVT::i32); |
4271 | SDValue Ops[] = {N->getOperand(Num: 0), N->getOperand(Num: 1), Pred, PredReg}; |
4272 | ReplaceNode(F: N, T: CurDAG->getMachineNode(Opcode: Opc, dl, VT1: VT, VT2: VT, Ops)); |
4273 | return; |
4274 | } |
4275 | case ARMISD::VTRN: { |
4276 | EVT VT = N->getValueType(ResNo: 0); |
4277 | unsigned Opc64[] = {ARM::VTRNd8, ARM::VTRNd16, ARM::VTRNd32}; |
4278 | unsigned Opc128[] = {ARM::VTRNq8, ARM::VTRNq16, ARM::VTRNq32}; |
4279 | unsigned Opc = getVectorShuffleOpcode(VT, Opc64, Opc128); |
4280 | SDValue Pred = getAL(CurDAG, dl); |
4281 | SDValue PredReg = CurDAG->getRegister(Reg: 0, VT: MVT::i32); |
4282 | SDValue Ops[] = {N->getOperand(Num: 0), N->getOperand(Num: 1), Pred, PredReg}; |
4283 | ReplaceNode(F: N, T: CurDAG->getMachineNode(Opcode: Opc, dl, VT1: VT, VT2: VT, Ops)); |
4284 | return; |
4285 | } |
4286 | case ARMISD::BUILD_VECTOR: { |
4287 | EVT VecVT = N->getValueType(ResNo: 0); |
4288 | EVT EltVT = VecVT.getVectorElementType(); |
4289 | unsigned NumElts = VecVT.getVectorNumElements(); |
4290 | if (EltVT == MVT::f64) { |
4291 | assert(NumElts == 2 && "unexpected type for BUILD_VECTOR" ); |
4292 | ReplaceNode( |
4293 | F: N, T: createDRegPairNode(VT: VecVT, V0: N->getOperand(Num: 0), V1: N->getOperand(Num: 1))); |
4294 | return; |
4295 | } |
4296 | assert(EltVT == MVT::f32 && "unexpected type for BUILD_VECTOR" ); |
4297 | if (NumElts == 2) { |
4298 | ReplaceNode( |
4299 | F: N, T: createSRegPairNode(VT: VecVT, V0: N->getOperand(Num: 0), V1: N->getOperand(Num: 1))); |
4300 | return; |
4301 | } |
4302 | assert(NumElts == 4 && "unexpected type for BUILD_VECTOR" ); |
4303 | ReplaceNode(F: N, |
4304 | T: createQuadSRegsNode(VT: VecVT, V0: N->getOperand(Num: 0), V1: N->getOperand(Num: 1), |
4305 | V2: N->getOperand(Num: 2), V3: N->getOperand(Num: 3))); |
4306 | return; |
4307 | } |
4308 | |
4309 | case ARMISD::VLD1DUP: { |
4310 | static const uint16_t DOpcodes[] = { ARM::VLD1DUPd8, ARM::VLD1DUPd16, |
4311 | ARM::VLD1DUPd32 }; |
4312 | static const uint16_t QOpcodes[] = { ARM::VLD1DUPq8, ARM::VLD1DUPq16, |
4313 | ARM::VLD1DUPq32 }; |
4314 | SelectVLDDup(N, /* IsIntrinsic= */ false, isUpdating: false, NumVecs: 1, DOpcodes, QOpcodes0: QOpcodes); |
4315 | return; |
4316 | } |
4317 | |
4318 | case ARMISD::VLD2DUP: { |
4319 | static const uint16_t Opcodes[] = { ARM::VLD2DUPd8, ARM::VLD2DUPd16, |
4320 | ARM::VLD2DUPd32 }; |
4321 | SelectVLDDup(N, /* IsIntrinsic= */ false, isUpdating: false, NumVecs: 2, DOpcodes: Opcodes); |
4322 | return; |
4323 | } |
4324 | |
4325 | case ARMISD::VLD3DUP: { |
4326 | static const uint16_t Opcodes[] = { ARM::VLD3DUPd8Pseudo, |
4327 | ARM::VLD3DUPd16Pseudo, |
4328 | ARM::VLD3DUPd32Pseudo }; |
4329 | SelectVLDDup(N, /* IsIntrinsic= */ false, isUpdating: false, NumVecs: 3, DOpcodes: Opcodes); |
4330 | return; |
4331 | } |
4332 | |
4333 | case ARMISD::VLD4DUP: { |
4334 | static const uint16_t Opcodes[] = { ARM::VLD4DUPd8Pseudo, |
4335 | ARM::VLD4DUPd16Pseudo, |
4336 | ARM::VLD4DUPd32Pseudo }; |
4337 | SelectVLDDup(N, /* IsIntrinsic= */ false, isUpdating: false, NumVecs: 4, DOpcodes: Opcodes); |
4338 | return; |
4339 | } |
4340 | |
4341 | case ARMISD::VLD1DUP_UPD: { |
4342 | static const uint16_t DOpcodes[] = { ARM::VLD1DUPd8wb_fixed, |
4343 | ARM::VLD1DUPd16wb_fixed, |
4344 | ARM::VLD1DUPd32wb_fixed }; |
4345 | static const uint16_t QOpcodes[] = { ARM::VLD1DUPq8wb_fixed, |
4346 | ARM::VLD1DUPq16wb_fixed, |
4347 | ARM::VLD1DUPq32wb_fixed }; |
4348 | SelectVLDDup(N, /* IsIntrinsic= */ false, isUpdating: true, NumVecs: 1, DOpcodes, QOpcodes0: QOpcodes); |
4349 | return; |
4350 | } |
4351 | |
4352 | case ARMISD::VLD2DUP_UPD: { |
4353 | static const uint16_t DOpcodes[] = { ARM::VLD2DUPd8wb_fixed, |
4354 | ARM::VLD2DUPd16wb_fixed, |
4355 | ARM::VLD2DUPd32wb_fixed, |
4356 | ARM::VLD1q64wb_fixed }; |
4357 | static const uint16_t QOpcodes0[] = { ARM::VLD2DUPq8EvenPseudo, |
4358 | ARM::VLD2DUPq16EvenPseudo, |
4359 | ARM::VLD2DUPq32EvenPseudo }; |
4360 | static const uint16_t QOpcodes1[] = { ARM::VLD2DUPq8OddPseudoWB_fixed, |
4361 | ARM::VLD2DUPq16OddPseudoWB_fixed, |
4362 | ARM::VLD2DUPq32OddPseudoWB_fixed }; |
4363 | SelectVLDDup(N, /* IsIntrinsic= */ false, isUpdating: true, NumVecs: 2, DOpcodes, QOpcodes0, QOpcodes1); |
4364 | return; |
4365 | } |
4366 | |
4367 | case ARMISD::VLD3DUP_UPD: { |
4368 | static const uint16_t DOpcodes[] = { ARM::VLD3DUPd8Pseudo_UPD, |
4369 | ARM::VLD3DUPd16Pseudo_UPD, |
4370 | ARM::VLD3DUPd32Pseudo_UPD, |
4371 | ARM::VLD1d64TPseudoWB_fixed }; |
4372 | static const uint16_t QOpcodes0[] = { ARM::VLD3DUPq8EvenPseudo, |
4373 | ARM::VLD3DUPq16EvenPseudo, |
4374 | ARM::VLD3DUPq32EvenPseudo }; |
4375 | static const uint16_t QOpcodes1[] = { ARM::VLD3DUPq8OddPseudo_UPD, |
4376 | ARM::VLD3DUPq16OddPseudo_UPD, |
4377 | ARM::VLD3DUPq32OddPseudo_UPD }; |
4378 | SelectVLDDup(N, /* IsIntrinsic= */ false, isUpdating: true, NumVecs: 3, DOpcodes, QOpcodes0, QOpcodes1); |
4379 | return; |
4380 | } |
4381 | |
4382 | case ARMISD::VLD4DUP_UPD: { |
4383 | static const uint16_t DOpcodes[] = { ARM::VLD4DUPd8Pseudo_UPD, |
4384 | ARM::VLD4DUPd16Pseudo_UPD, |
4385 | ARM::VLD4DUPd32Pseudo_UPD, |
4386 | ARM::VLD1d64QPseudoWB_fixed }; |
4387 | static const uint16_t QOpcodes0[] = { ARM::VLD4DUPq8EvenPseudo, |
4388 | ARM::VLD4DUPq16EvenPseudo, |
4389 | ARM::VLD4DUPq32EvenPseudo }; |
4390 | static const uint16_t QOpcodes1[] = { ARM::VLD4DUPq8OddPseudo_UPD, |
4391 | ARM::VLD4DUPq16OddPseudo_UPD, |
4392 | ARM::VLD4DUPq32OddPseudo_UPD }; |
4393 | SelectVLDDup(N, /* IsIntrinsic= */ false, isUpdating: true, NumVecs: 4, DOpcodes, QOpcodes0, QOpcodes1); |
4394 | return; |
4395 | } |
4396 | |
4397 | case ARMISD::VLD1_UPD: { |
4398 | static const uint16_t DOpcodes[] = { ARM::VLD1d8wb_fixed, |
4399 | ARM::VLD1d16wb_fixed, |
4400 | ARM::VLD1d32wb_fixed, |
4401 | ARM::VLD1d64wb_fixed }; |
4402 | static const uint16_t QOpcodes[] = { ARM::VLD1q8wb_fixed, |
4403 | ARM::VLD1q16wb_fixed, |
4404 | ARM::VLD1q32wb_fixed, |
4405 | ARM::VLD1q64wb_fixed }; |
4406 | SelectVLD(N, isUpdating: true, NumVecs: 1, DOpcodes, QOpcodes0: QOpcodes, QOpcodes1: nullptr); |
4407 | return; |
4408 | } |
4409 | |
4410 | case ARMISD::VLD2_UPD: { |
4411 | if (Subtarget->hasNEON()) { |
4412 | static const uint16_t DOpcodes[] = { |
4413 | ARM::VLD2d8wb_fixed, ARM::VLD2d16wb_fixed, ARM::VLD2d32wb_fixed, |
4414 | ARM::VLD1q64wb_fixed}; |
4415 | static const uint16_t QOpcodes[] = {ARM::VLD2q8PseudoWB_fixed, |
4416 | ARM::VLD2q16PseudoWB_fixed, |
4417 | ARM::VLD2q32PseudoWB_fixed}; |
4418 | SelectVLD(N, isUpdating: true, NumVecs: 2, DOpcodes, QOpcodes0: QOpcodes, QOpcodes1: nullptr); |
4419 | } else { |
4420 | static const uint16_t Opcodes8[] = {ARM::MVE_VLD20_8, |
4421 | ARM::MVE_VLD21_8_wb}; |
4422 | static const uint16_t Opcodes16[] = {ARM::MVE_VLD20_16, |
4423 | ARM::MVE_VLD21_16_wb}; |
4424 | static const uint16_t Opcodes32[] = {ARM::MVE_VLD20_32, |
4425 | ARM::MVE_VLD21_32_wb}; |
4426 | static const uint16_t *const Opcodes[] = {Opcodes8, Opcodes16, Opcodes32}; |
4427 | SelectMVE_VLD(N, NumVecs: 2, Opcodes, HasWriteback: true); |
4428 | } |
4429 | return; |
4430 | } |
4431 | |
4432 | case ARMISD::VLD3_UPD: { |
4433 | static const uint16_t DOpcodes[] = { ARM::VLD3d8Pseudo_UPD, |
4434 | ARM::VLD3d16Pseudo_UPD, |
4435 | ARM::VLD3d32Pseudo_UPD, |
4436 | ARM::VLD1d64TPseudoWB_fixed}; |
4437 | static const uint16_t QOpcodes0[] = { ARM::VLD3q8Pseudo_UPD, |
4438 | ARM::VLD3q16Pseudo_UPD, |
4439 | ARM::VLD3q32Pseudo_UPD }; |
4440 | static const uint16_t QOpcodes1[] = { ARM::VLD3q8oddPseudo_UPD, |
4441 | ARM::VLD3q16oddPseudo_UPD, |
4442 | ARM::VLD3q32oddPseudo_UPD }; |
4443 | SelectVLD(N, isUpdating: true, NumVecs: 3, DOpcodes, QOpcodes0, QOpcodes1); |
4444 | return; |
4445 | } |
4446 | |
4447 | case ARMISD::VLD4_UPD: { |
4448 | if (Subtarget->hasNEON()) { |
4449 | static const uint16_t DOpcodes[] = { |
4450 | ARM::VLD4d8Pseudo_UPD, ARM::VLD4d16Pseudo_UPD, ARM::VLD4d32Pseudo_UPD, |
4451 | ARM::VLD1d64QPseudoWB_fixed}; |
4452 | static const uint16_t QOpcodes0[] = {ARM::VLD4q8Pseudo_UPD, |
4453 | ARM::VLD4q16Pseudo_UPD, |
4454 | ARM::VLD4q32Pseudo_UPD}; |
4455 | static const uint16_t QOpcodes1[] = {ARM::VLD4q8oddPseudo_UPD, |
4456 | ARM::VLD4q16oddPseudo_UPD, |
4457 | ARM::VLD4q32oddPseudo_UPD}; |
4458 | SelectVLD(N, isUpdating: true, NumVecs: 4, DOpcodes, QOpcodes0, QOpcodes1); |
4459 | } else { |
4460 | static const uint16_t Opcodes8[] = {ARM::MVE_VLD40_8, ARM::MVE_VLD41_8, |
4461 | ARM::MVE_VLD42_8, |
4462 | ARM::MVE_VLD43_8_wb}; |
4463 | static const uint16_t Opcodes16[] = {ARM::MVE_VLD40_16, ARM::MVE_VLD41_16, |
4464 | ARM::MVE_VLD42_16, |
4465 | ARM::MVE_VLD43_16_wb}; |
4466 | static const uint16_t Opcodes32[] = {ARM::MVE_VLD40_32, ARM::MVE_VLD41_32, |
4467 | ARM::MVE_VLD42_32, |
4468 | ARM::MVE_VLD43_32_wb}; |
4469 | static const uint16_t *const Opcodes[] = {Opcodes8, Opcodes16, Opcodes32}; |
4470 | SelectMVE_VLD(N, NumVecs: 4, Opcodes, HasWriteback: true); |
4471 | } |
4472 | return; |
4473 | } |
4474 | |
4475 | case ARMISD::VLD1x2_UPD: { |
4476 | if (Subtarget->hasNEON()) { |
4477 | static const uint16_t DOpcodes[] = { |
4478 | ARM::VLD1q8wb_fixed, ARM::VLD1q16wb_fixed, ARM::VLD1q32wb_fixed, |
4479 | ARM::VLD1q64wb_fixed}; |
4480 | static const uint16_t QOpcodes[] = { |
4481 | ARM::VLD1d8QPseudoWB_fixed, ARM::VLD1d16QPseudoWB_fixed, |
4482 | ARM::VLD1d32QPseudoWB_fixed, ARM::VLD1d64QPseudoWB_fixed}; |
4483 | SelectVLD(N, isUpdating: true, NumVecs: 2, DOpcodes, QOpcodes0: QOpcodes, QOpcodes1: nullptr); |
4484 | return; |
4485 | } |
4486 | break; |
4487 | } |
4488 | |
4489 | case ARMISD::VLD1x3_UPD: { |
4490 | if (Subtarget->hasNEON()) { |
4491 | static const uint16_t DOpcodes[] = { |
4492 | ARM::VLD1d8TPseudoWB_fixed, ARM::VLD1d16TPseudoWB_fixed, |
4493 | ARM::VLD1d32TPseudoWB_fixed, ARM::VLD1d64TPseudoWB_fixed}; |
4494 | static const uint16_t QOpcodes0[] = { |
4495 | ARM::VLD1q8LowTPseudo_UPD, ARM::VLD1q16LowTPseudo_UPD, |
4496 | ARM::VLD1q32LowTPseudo_UPD, ARM::VLD1q64LowTPseudo_UPD}; |
4497 | static const uint16_t QOpcodes1[] = { |
4498 | ARM::VLD1q8HighTPseudo_UPD, ARM::VLD1q16HighTPseudo_UPD, |
4499 | ARM::VLD1q32HighTPseudo_UPD, ARM::VLD1q64HighTPseudo_UPD}; |
4500 | SelectVLD(N, isUpdating: true, NumVecs: 3, DOpcodes, QOpcodes0, QOpcodes1); |
4501 | return; |
4502 | } |
4503 | break; |
4504 | } |
4505 | |
4506 | case ARMISD::VLD1x4_UPD: { |
4507 | if (Subtarget->hasNEON()) { |
4508 | static const uint16_t DOpcodes[] = { |
4509 | ARM::VLD1d8QPseudoWB_fixed, ARM::VLD1d16QPseudoWB_fixed, |
4510 | ARM::VLD1d32QPseudoWB_fixed, ARM::VLD1d64QPseudoWB_fixed}; |
4511 | static const uint16_t QOpcodes0[] = { |
4512 | ARM::VLD1q8LowQPseudo_UPD, ARM::VLD1q16LowQPseudo_UPD, |
4513 | ARM::VLD1q32LowQPseudo_UPD, ARM::VLD1q64LowQPseudo_UPD}; |
4514 | static const uint16_t QOpcodes1[] = { |
4515 | ARM::VLD1q8HighQPseudo_UPD, ARM::VLD1q16HighQPseudo_UPD, |
4516 | ARM::VLD1q32HighQPseudo_UPD, ARM::VLD1q64HighQPseudo_UPD}; |
4517 | SelectVLD(N, isUpdating: true, NumVecs: 4, DOpcodes, QOpcodes0, QOpcodes1); |
4518 | return; |
4519 | } |
4520 | break; |
4521 | } |
4522 | |
4523 | case ARMISD::VLD2LN_UPD: { |
4524 | static const uint16_t DOpcodes[] = { ARM::VLD2LNd8Pseudo_UPD, |
4525 | ARM::VLD2LNd16Pseudo_UPD, |
4526 | ARM::VLD2LNd32Pseudo_UPD }; |
4527 | static const uint16_t QOpcodes[] = { ARM::VLD2LNq16Pseudo_UPD, |
4528 | ARM::VLD2LNq32Pseudo_UPD }; |
4529 | SelectVLDSTLane(N, IsLoad: true, isUpdating: true, NumVecs: 2, DOpcodes, QOpcodes); |
4530 | return; |
4531 | } |
4532 | |
4533 | case ARMISD::VLD3LN_UPD: { |
4534 | static const uint16_t DOpcodes[] = { ARM::VLD3LNd8Pseudo_UPD, |
4535 | ARM::VLD3LNd16Pseudo_UPD, |
4536 | ARM::VLD3LNd32Pseudo_UPD }; |
4537 | static const uint16_t QOpcodes[] = { ARM::VLD3LNq16Pseudo_UPD, |
4538 | ARM::VLD3LNq32Pseudo_UPD }; |
4539 | SelectVLDSTLane(N, IsLoad: true, isUpdating: true, NumVecs: 3, DOpcodes, QOpcodes); |
4540 | return; |
4541 | } |
4542 | |
4543 | case ARMISD::VLD4LN_UPD: { |
4544 | static const uint16_t DOpcodes[] = { ARM::VLD4LNd8Pseudo_UPD, |
4545 | ARM::VLD4LNd16Pseudo_UPD, |
4546 | ARM::VLD4LNd32Pseudo_UPD }; |
4547 | static const uint16_t QOpcodes[] = { ARM::VLD4LNq16Pseudo_UPD, |
4548 | ARM::VLD4LNq32Pseudo_UPD }; |
4549 | SelectVLDSTLane(N, IsLoad: true, isUpdating: true, NumVecs: 4, DOpcodes, QOpcodes); |
4550 | return; |
4551 | } |
4552 | |
4553 | case ARMISD::VST1_UPD: { |
4554 | static const uint16_t DOpcodes[] = { ARM::VST1d8wb_fixed, |
4555 | ARM::VST1d16wb_fixed, |
4556 | ARM::VST1d32wb_fixed, |
4557 | ARM::VST1d64wb_fixed }; |
4558 | static const uint16_t QOpcodes[] = { ARM::VST1q8wb_fixed, |
4559 | ARM::VST1q16wb_fixed, |
4560 | ARM::VST1q32wb_fixed, |
4561 | ARM::VST1q64wb_fixed }; |
4562 | SelectVST(N, isUpdating: true, NumVecs: 1, DOpcodes, QOpcodes0: QOpcodes, QOpcodes1: nullptr); |
4563 | return; |
4564 | } |
4565 | |
4566 | case ARMISD::VST2_UPD: { |
4567 | if (Subtarget->hasNEON()) { |
4568 | static const uint16_t DOpcodes[] = { |
4569 | ARM::VST2d8wb_fixed, ARM::VST2d16wb_fixed, ARM::VST2d32wb_fixed, |
4570 | ARM::VST1q64wb_fixed}; |
4571 | static const uint16_t QOpcodes[] = {ARM::VST2q8PseudoWB_fixed, |
4572 | ARM::VST2q16PseudoWB_fixed, |
4573 | ARM::VST2q32PseudoWB_fixed}; |
4574 | SelectVST(N, isUpdating: true, NumVecs: 2, DOpcodes, QOpcodes0: QOpcodes, QOpcodes1: nullptr); |
4575 | return; |
4576 | } |
4577 | break; |
4578 | } |
4579 | |
4580 | case ARMISD::VST3_UPD: { |
4581 | static const uint16_t DOpcodes[] = { ARM::VST3d8Pseudo_UPD, |
4582 | ARM::VST3d16Pseudo_UPD, |
4583 | ARM::VST3d32Pseudo_UPD, |
4584 | ARM::VST1d64TPseudoWB_fixed}; |
4585 | static const uint16_t QOpcodes0[] = { ARM::VST3q8Pseudo_UPD, |
4586 | ARM::VST3q16Pseudo_UPD, |
4587 | ARM::VST3q32Pseudo_UPD }; |
4588 | static const uint16_t QOpcodes1[] = { ARM::VST3q8oddPseudo_UPD, |
4589 | ARM::VST3q16oddPseudo_UPD, |
4590 | ARM::VST3q32oddPseudo_UPD }; |
4591 | SelectVST(N, isUpdating: true, NumVecs: 3, DOpcodes, QOpcodes0, QOpcodes1); |
4592 | return; |
4593 | } |
4594 | |
4595 | case ARMISD::VST4_UPD: { |
4596 | if (Subtarget->hasNEON()) { |
4597 | static const uint16_t DOpcodes[] = { |
4598 | ARM::VST4d8Pseudo_UPD, ARM::VST4d16Pseudo_UPD, ARM::VST4d32Pseudo_UPD, |
4599 | ARM::VST1d64QPseudoWB_fixed}; |
4600 | static const uint16_t QOpcodes0[] = {ARM::VST4q8Pseudo_UPD, |
4601 | ARM::VST4q16Pseudo_UPD, |
4602 | ARM::VST4q32Pseudo_UPD}; |
4603 | static const uint16_t QOpcodes1[] = {ARM::VST4q8oddPseudo_UPD, |
4604 | ARM::VST4q16oddPseudo_UPD, |
4605 | ARM::VST4q32oddPseudo_UPD}; |
4606 | SelectVST(N, isUpdating: true, NumVecs: 4, DOpcodes, QOpcodes0, QOpcodes1); |
4607 | return; |
4608 | } |
4609 | break; |
4610 | } |
4611 | |
4612 | case ARMISD::VST1x2_UPD: { |
4613 | if (Subtarget->hasNEON()) { |
4614 | static const uint16_t DOpcodes[] = { ARM::VST1q8wb_fixed, |
4615 | ARM::VST1q16wb_fixed, |
4616 | ARM::VST1q32wb_fixed, |
4617 | ARM::VST1q64wb_fixed}; |
4618 | static const uint16_t QOpcodes[] = { ARM::VST1d8QPseudoWB_fixed, |
4619 | ARM::VST1d16QPseudoWB_fixed, |
4620 | ARM::VST1d32QPseudoWB_fixed, |
4621 | ARM::VST1d64QPseudoWB_fixed }; |
4622 | SelectVST(N, isUpdating: true, NumVecs: 2, DOpcodes, QOpcodes0: QOpcodes, QOpcodes1: nullptr); |
4623 | return; |
4624 | } |
4625 | break; |
4626 | } |
4627 | |
4628 | case ARMISD::VST1x3_UPD: { |
4629 | if (Subtarget->hasNEON()) { |
4630 | static const uint16_t DOpcodes[] = { ARM::VST1d8TPseudoWB_fixed, |
4631 | ARM::VST1d16TPseudoWB_fixed, |
4632 | ARM::VST1d32TPseudoWB_fixed, |
4633 | ARM::VST1d64TPseudoWB_fixed }; |
4634 | static const uint16_t QOpcodes0[] = { ARM::VST1q8LowTPseudo_UPD, |
4635 | ARM::VST1q16LowTPseudo_UPD, |
4636 | ARM::VST1q32LowTPseudo_UPD, |
4637 | ARM::VST1q64LowTPseudo_UPD }; |
4638 | static const uint16_t QOpcodes1[] = { ARM::VST1q8HighTPseudo_UPD, |
4639 | ARM::VST1q16HighTPseudo_UPD, |
4640 | ARM::VST1q32HighTPseudo_UPD, |
4641 | ARM::VST1q64HighTPseudo_UPD }; |
4642 | SelectVST(N, isUpdating: true, NumVecs: 3, DOpcodes, QOpcodes0, QOpcodes1); |
4643 | return; |
4644 | } |
4645 | break; |
4646 | } |
4647 | |
4648 | case ARMISD::VST1x4_UPD: { |
4649 | if (Subtarget->hasNEON()) { |
4650 | static const uint16_t DOpcodes[] = { ARM::VST1d8QPseudoWB_fixed, |
4651 | ARM::VST1d16QPseudoWB_fixed, |
4652 | ARM::VST1d32QPseudoWB_fixed, |
4653 | ARM::VST1d64QPseudoWB_fixed }; |
4654 | static const uint16_t QOpcodes0[] = { ARM::VST1q8LowQPseudo_UPD, |
4655 | ARM::VST1q16LowQPseudo_UPD, |
4656 | ARM::VST1q32LowQPseudo_UPD, |
4657 | ARM::VST1q64LowQPseudo_UPD }; |
4658 | static const uint16_t QOpcodes1[] = { ARM::VST1q8HighQPseudo_UPD, |
4659 | ARM::VST1q16HighQPseudo_UPD, |
4660 | ARM::VST1q32HighQPseudo_UPD, |
4661 | ARM::VST1q64HighQPseudo_UPD }; |
4662 | SelectVST(N, isUpdating: true, NumVecs: 4, DOpcodes, QOpcodes0, QOpcodes1); |
4663 | return; |
4664 | } |
4665 | break; |
4666 | } |
4667 | case ARMISD::VST2LN_UPD: { |
4668 | static const uint16_t DOpcodes[] = { ARM::VST2LNd8Pseudo_UPD, |
4669 | ARM::VST2LNd16Pseudo_UPD, |
4670 | ARM::VST2LNd32Pseudo_UPD }; |
4671 | static const uint16_t QOpcodes[] = { ARM::VST2LNq16Pseudo_UPD, |
4672 | ARM::VST2LNq32Pseudo_UPD }; |
4673 | SelectVLDSTLane(N, IsLoad: false, isUpdating: true, NumVecs: 2, DOpcodes, QOpcodes); |
4674 | return; |
4675 | } |
4676 | |
4677 | case ARMISD::VST3LN_UPD: { |
4678 | static const uint16_t DOpcodes[] = { ARM::VST3LNd8Pseudo_UPD, |
4679 | ARM::VST3LNd16Pseudo_UPD, |
4680 | ARM::VST3LNd32Pseudo_UPD }; |
4681 | static const uint16_t QOpcodes[] = { ARM::VST3LNq16Pseudo_UPD, |
4682 | ARM::VST3LNq32Pseudo_UPD }; |
4683 | SelectVLDSTLane(N, IsLoad: false, isUpdating: true, NumVecs: 3, DOpcodes, QOpcodes); |
4684 | return; |
4685 | } |
4686 | |
4687 | case ARMISD::VST4LN_UPD: { |
4688 | static const uint16_t DOpcodes[] = { ARM::VST4LNd8Pseudo_UPD, |
4689 | ARM::VST4LNd16Pseudo_UPD, |
4690 | ARM::VST4LNd32Pseudo_UPD }; |
4691 | static const uint16_t QOpcodes[] = { ARM::VST4LNq16Pseudo_UPD, |
4692 | ARM::VST4LNq32Pseudo_UPD }; |
4693 | SelectVLDSTLane(N, IsLoad: false, isUpdating: true, NumVecs: 4, DOpcodes, QOpcodes); |
4694 | return; |
4695 | } |
4696 | |
4697 | case ISD::INTRINSIC_VOID: |
4698 | case ISD::INTRINSIC_W_CHAIN: { |
4699 | unsigned IntNo = N->getConstantOperandVal(Num: 1); |
4700 | switch (IntNo) { |
4701 | default: |
4702 | break; |
4703 | |
4704 | case Intrinsic::arm_mrrc: |
4705 | case Intrinsic::arm_mrrc2: { |
4706 | SDLoc dl(N); |
4707 | SDValue Chain = N->getOperand(Num: 0); |
4708 | unsigned Opc; |
4709 | |
4710 | if (Subtarget->isThumb()) |
4711 | Opc = (IntNo == Intrinsic::arm_mrrc ? ARM::t2MRRC : ARM::t2MRRC2); |
4712 | else |
4713 | Opc = (IntNo == Intrinsic::arm_mrrc ? ARM::MRRC : ARM::MRRC2); |
4714 | |
4715 | SmallVector<SDValue, 5> Ops; |
4716 | Ops.push_back(Elt: getI32Imm(Imm: N->getConstantOperandVal(Num: 2), dl)); /* coproc */ |
4717 | Ops.push_back(Elt: getI32Imm(Imm: N->getConstantOperandVal(Num: 3), dl)); /* opc */ |
4718 | Ops.push_back(Elt: getI32Imm(Imm: N->getConstantOperandVal(Num: 4), dl)); /* CRm */ |
4719 | |
4720 | // The mrrc2 instruction in ARM doesn't allow predicates, the top 4 bits of the encoded |
4721 | // instruction will always be '1111' but it is possible in assembly language to specify |
4722 | // AL as a predicate to mrrc2 but it doesn't make any difference to the encoded instruction. |
4723 | if (Opc != ARM::MRRC2) { |
4724 | Ops.push_back(Elt: getAL(CurDAG, dl)); |
4725 | Ops.push_back(Elt: CurDAG->getRegister(Reg: 0, VT: MVT::i32)); |
4726 | } |
4727 | |
4728 | Ops.push_back(Elt: Chain); |
4729 | |
4730 | // Writes to two registers. |
4731 | const EVT RetType[] = {MVT::i32, MVT::i32, MVT::Other}; |
4732 | |
4733 | ReplaceNode(F: N, T: CurDAG->getMachineNode(Opcode: Opc, dl, ResultTys: RetType, Ops)); |
4734 | return; |
4735 | } |
4736 | case Intrinsic::arm_ldaexd: |
4737 | case Intrinsic::arm_ldrexd: { |
4738 | SDLoc dl(N); |
4739 | SDValue Chain = N->getOperand(Num: 0); |
4740 | SDValue MemAddr = N->getOperand(Num: 2); |
4741 | bool isThumb = Subtarget->isThumb() && Subtarget->hasV8MBaselineOps(); |
4742 | |
4743 | bool IsAcquire = IntNo == Intrinsic::arm_ldaexd; |
4744 | unsigned NewOpc = isThumb ? (IsAcquire ? ARM::t2LDAEXD : ARM::t2LDREXD) |
4745 | : (IsAcquire ? ARM::LDAEXD : ARM::LDREXD); |
4746 | |
4747 | // arm_ldrexd returns a i64 value in {i32, i32} |
4748 | std::vector<EVT> ResTys; |
4749 | if (isThumb) { |
4750 | ResTys.push_back(x: MVT::i32); |
4751 | ResTys.push_back(x: MVT::i32); |
4752 | } else |
4753 | ResTys.push_back(x: MVT::Untyped); |
4754 | ResTys.push_back(x: MVT::Other); |
4755 | |
4756 | // Place arguments in the right order. |
4757 | SDValue Ops[] = {MemAddr, getAL(CurDAG, dl), |
4758 | CurDAG->getRegister(Reg: 0, VT: MVT::i32), Chain}; |
4759 | SDNode *Ld = CurDAG->getMachineNode(Opcode: NewOpc, dl, ResultTys: ResTys, Ops); |
4760 | // Transfer memoperands. |
4761 | MachineMemOperand *MemOp = cast<MemIntrinsicSDNode>(Val: N)->getMemOperand(); |
4762 | CurDAG->setNodeMemRefs(N: cast<MachineSDNode>(Val: Ld), NewMemRefs: {MemOp}); |
4763 | |
4764 | // Remap uses. |
4765 | SDValue OutChain = isThumb ? SDValue(Ld, 2) : SDValue(Ld, 1); |
4766 | if (!SDValue(N, 0).use_empty()) { |
4767 | SDValue Result; |
4768 | if (isThumb) |
4769 | Result = SDValue(Ld, 0); |
4770 | else { |
4771 | SDValue SubRegIdx = |
4772 | CurDAG->getTargetConstant(Val: ARM::gsub_0, DL: dl, VT: MVT::i32); |
4773 | SDNode *ResNode = CurDAG->getMachineNode(Opcode: TargetOpcode::EXTRACT_SUBREG, |
4774 | dl, VT: MVT::i32, Op1: SDValue(Ld, 0), Op2: SubRegIdx); |
4775 | Result = SDValue(ResNode,0); |
4776 | } |
4777 | ReplaceUses(F: SDValue(N, 0), T: Result); |
4778 | } |
4779 | if (!SDValue(N, 1).use_empty()) { |
4780 | SDValue Result; |
4781 | if (isThumb) |
4782 | Result = SDValue(Ld, 1); |
4783 | else { |
4784 | SDValue SubRegIdx = |
4785 | CurDAG->getTargetConstant(Val: ARM::gsub_1, DL: dl, VT: MVT::i32); |
4786 | SDNode *ResNode = CurDAG->getMachineNode(Opcode: TargetOpcode::EXTRACT_SUBREG, |
4787 | dl, VT: MVT::i32, Op1: SDValue(Ld, 0), Op2: SubRegIdx); |
4788 | Result = SDValue(ResNode,0); |
4789 | } |
4790 | ReplaceUses(F: SDValue(N, 1), T: Result); |
4791 | } |
4792 | ReplaceUses(F: SDValue(N, 2), T: OutChain); |
4793 | CurDAG->RemoveDeadNode(N); |
4794 | return; |
4795 | } |
4796 | case Intrinsic::arm_stlexd: |
4797 | case Intrinsic::arm_strexd: { |
4798 | SDLoc dl(N); |
4799 | SDValue Chain = N->getOperand(Num: 0); |
4800 | SDValue Val0 = N->getOperand(Num: 2); |
4801 | SDValue Val1 = N->getOperand(Num: 3); |
4802 | SDValue MemAddr = N->getOperand(Num: 4); |
4803 | |
4804 | // Store exclusive double return a i32 value which is the return status |
4805 | // of the issued store. |
4806 | const EVT ResTys[] = {MVT::i32, MVT::Other}; |
4807 | |
4808 | bool isThumb = Subtarget->isThumb() && Subtarget->hasThumb2(); |
4809 | // Place arguments in the right order. |
4810 | SmallVector<SDValue, 7> Ops; |
4811 | if (isThumb) { |
4812 | Ops.push_back(Elt: Val0); |
4813 | Ops.push_back(Elt: Val1); |
4814 | } else |
4815 | // arm_strexd uses GPRPair. |
4816 | Ops.push_back(Elt: SDValue(createGPRPairNode(VT: MVT::Untyped, V0: Val0, V1: Val1), 0)); |
4817 | Ops.push_back(Elt: MemAddr); |
4818 | Ops.push_back(Elt: getAL(CurDAG, dl)); |
4819 | Ops.push_back(Elt: CurDAG->getRegister(Reg: 0, VT: MVT::i32)); |
4820 | Ops.push_back(Elt: Chain); |
4821 | |
4822 | bool IsRelease = IntNo == Intrinsic::arm_stlexd; |
4823 | unsigned NewOpc = isThumb ? (IsRelease ? ARM::t2STLEXD : ARM::t2STREXD) |
4824 | : (IsRelease ? ARM::STLEXD : ARM::STREXD); |
4825 | |
4826 | SDNode *St = CurDAG->getMachineNode(Opcode: NewOpc, dl, ResultTys: ResTys, Ops); |
4827 | // Transfer memoperands. |
4828 | MachineMemOperand *MemOp = cast<MemIntrinsicSDNode>(Val: N)->getMemOperand(); |
4829 | CurDAG->setNodeMemRefs(N: cast<MachineSDNode>(Val: St), NewMemRefs: {MemOp}); |
4830 | |
4831 | ReplaceNode(F: N, T: St); |
4832 | return; |
4833 | } |
4834 | |
4835 | case Intrinsic::arm_neon_vld1: { |
4836 | static const uint16_t DOpcodes[] = { ARM::VLD1d8, ARM::VLD1d16, |
4837 | ARM::VLD1d32, ARM::VLD1d64 }; |
4838 | static const uint16_t QOpcodes[] = { ARM::VLD1q8, ARM::VLD1q16, |
4839 | ARM::VLD1q32, ARM::VLD1q64}; |
4840 | SelectVLD(N, isUpdating: false, NumVecs: 1, DOpcodes, QOpcodes0: QOpcodes, QOpcodes1: nullptr); |
4841 | return; |
4842 | } |
4843 | |
4844 | case Intrinsic::arm_neon_vld1x2: { |
4845 | static const uint16_t DOpcodes[] = { ARM::VLD1q8, ARM::VLD1q16, |
4846 | ARM::VLD1q32, ARM::VLD1q64 }; |
4847 | static const uint16_t QOpcodes[] = { ARM::VLD1d8QPseudo, |
4848 | ARM::VLD1d16QPseudo, |
4849 | ARM::VLD1d32QPseudo, |
4850 | ARM::VLD1d64QPseudo }; |
4851 | SelectVLD(N, isUpdating: false, NumVecs: 2, DOpcodes, QOpcodes0: QOpcodes, QOpcodes1: nullptr); |
4852 | return; |
4853 | } |
4854 | |
4855 | case Intrinsic::arm_neon_vld1x3: { |
4856 | static const uint16_t DOpcodes[] = { ARM::VLD1d8TPseudo, |
4857 | ARM::VLD1d16TPseudo, |
4858 | ARM::VLD1d32TPseudo, |
4859 | ARM::VLD1d64TPseudo }; |
4860 | static const uint16_t QOpcodes0[] = { ARM::VLD1q8LowTPseudo_UPD, |
4861 | ARM::VLD1q16LowTPseudo_UPD, |
4862 | ARM::VLD1q32LowTPseudo_UPD, |
4863 | ARM::VLD1q64LowTPseudo_UPD }; |
4864 | static const uint16_t QOpcodes1[] = { ARM::VLD1q8HighTPseudo, |
4865 | ARM::VLD1q16HighTPseudo, |
4866 | ARM::VLD1q32HighTPseudo, |
4867 | ARM::VLD1q64HighTPseudo }; |
4868 | SelectVLD(N, isUpdating: false, NumVecs: 3, DOpcodes, QOpcodes0, QOpcodes1); |
4869 | return; |
4870 | } |
4871 | |
4872 | case Intrinsic::arm_neon_vld1x4: { |
4873 | static const uint16_t DOpcodes[] = { ARM::VLD1d8QPseudo, |
4874 | ARM::VLD1d16QPseudo, |
4875 | ARM::VLD1d32QPseudo, |
4876 | ARM::VLD1d64QPseudo }; |
4877 | static const uint16_t QOpcodes0[] = { ARM::VLD1q8LowQPseudo_UPD, |
4878 | ARM::VLD1q16LowQPseudo_UPD, |
4879 | ARM::VLD1q32LowQPseudo_UPD, |
4880 | ARM::VLD1q64LowQPseudo_UPD }; |
4881 | static const uint16_t QOpcodes1[] = { ARM::VLD1q8HighQPseudo, |
4882 | ARM::VLD1q16HighQPseudo, |
4883 | ARM::VLD1q32HighQPseudo, |
4884 | ARM::VLD1q64HighQPseudo }; |
4885 | SelectVLD(N, isUpdating: false, NumVecs: 4, DOpcodes, QOpcodes0, QOpcodes1); |
4886 | return; |
4887 | } |
4888 | |
4889 | case Intrinsic::arm_neon_vld2: { |
4890 | static const uint16_t DOpcodes[] = { ARM::VLD2d8, ARM::VLD2d16, |
4891 | ARM::VLD2d32, ARM::VLD1q64 }; |
4892 | static const uint16_t QOpcodes[] = { ARM::VLD2q8Pseudo, ARM::VLD2q16Pseudo, |
4893 | ARM::VLD2q32Pseudo }; |
4894 | SelectVLD(N, isUpdating: false, NumVecs: 2, DOpcodes, QOpcodes0: QOpcodes, QOpcodes1: nullptr); |
4895 | return; |
4896 | } |
4897 | |
4898 | case Intrinsic::arm_neon_vld3: { |
4899 | static const uint16_t DOpcodes[] = { ARM::VLD3d8Pseudo, |
4900 | ARM::VLD3d16Pseudo, |
4901 | ARM::VLD3d32Pseudo, |
4902 | ARM::VLD1d64TPseudo }; |
4903 | static const uint16_t QOpcodes0[] = { ARM::VLD3q8Pseudo_UPD, |
4904 | ARM::VLD3q16Pseudo_UPD, |
4905 | ARM::VLD3q32Pseudo_UPD }; |
4906 | static const uint16_t QOpcodes1[] = { ARM::VLD3q8oddPseudo, |
4907 | ARM::VLD3q16oddPseudo, |
4908 | ARM::VLD3q32oddPseudo }; |
4909 | SelectVLD(N, isUpdating: false, NumVecs: 3, DOpcodes, QOpcodes0, QOpcodes1); |
4910 | return; |
4911 | } |
4912 | |
4913 | case Intrinsic::arm_neon_vld4: { |
4914 | static const uint16_t DOpcodes[] = { ARM::VLD4d8Pseudo, |
4915 | ARM::VLD4d16Pseudo, |
4916 | ARM::VLD4d32Pseudo, |
4917 | ARM::VLD1d64QPseudo }; |
4918 | static const uint16_t QOpcodes0[] = { ARM::VLD4q8Pseudo_UPD, |
4919 | ARM::VLD4q16Pseudo_UPD, |
4920 | ARM::VLD4q32Pseudo_UPD }; |
4921 | static const uint16_t QOpcodes1[] = { ARM::VLD4q8oddPseudo, |
4922 | ARM::VLD4q16oddPseudo, |
4923 | ARM::VLD4q32oddPseudo }; |
4924 | SelectVLD(N, isUpdating: false, NumVecs: 4, DOpcodes, QOpcodes0, QOpcodes1); |
4925 | return; |
4926 | } |
4927 | |
4928 | case Intrinsic::arm_neon_vld2dup: { |
4929 | static const uint16_t DOpcodes[] = { ARM::VLD2DUPd8, ARM::VLD2DUPd16, |
4930 | ARM::VLD2DUPd32, ARM::VLD1q64 }; |
4931 | static const uint16_t QOpcodes0[] = { ARM::VLD2DUPq8EvenPseudo, |
4932 | ARM::VLD2DUPq16EvenPseudo, |
4933 | ARM::VLD2DUPq32EvenPseudo }; |
4934 | static const uint16_t QOpcodes1[] = { ARM::VLD2DUPq8OddPseudo, |
4935 | ARM::VLD2DUPq16OddPseudo, |
4936 | ARM::VLD2DUPq32OddPseudo }; |
4937 | SelectVLDDup(N, /* IsIntrinsic= */ true, isUpdating: false, NumVecs: 2, |
4938 | DOpcodes, QOpcodes0, QOpcodes1); |
4939 | return; |
4940 | } |
4941 | |
4942 | case Intrinsic::arm_neon_vld3dup: { |
4943 | static const uint16_t DOpcodes[] = { ARM::VLD3DUPd8Pseudo, |
4944 | ARM::VLD3DUPd16Pseudo, |
4945 | ARM::VLD3DUPd32Pseudo, |
4946 | ARM::VLD1d64TPseudo }; |
4947 | static const uint16_t QOpcodes0[] = { ARM::VLD3DUPq8EvenPseudo, |
4948 | ARM::VLD3DUPq16EvenPseudo, |
4949 | ARM::VLD3DUPq32EvenPseudo }; |
4950 | static const uint16_t QOpcodes1[] = { ARM::VLD3DUPq8OddPseudo, |
4951 | ARM::VLD3DUPq16OddPseudo, |
4952 | ARM::VLD3DUPq32OddPseudo }; |
4953 | SelectVLDDup(N, /* IsIntrinsic= */ true, isUpdating: false, NumVecs: 3, |
4954 | DOpcodes, QOpcodes0, QOpcodes1); |
4955 | return; |
4956 | } |
4957 | |
4958 | case Intrinsic::arm_neon_vld4dup: { |
4959 | static const uint16_t DOpcodes[] = { ARM::VLD4DUPd8Pseudo, |
4960 | ARM::VLD4DUPd16Pseudo, |
4961 | ARM::VLD4DUPd32Pseudo, |
4962 | ARM::VLD1d64QPseudo }; |
4963 | static const uint16_t QOpcodes0[] = { ARM::VLD4DUPq8EvenPseudo, |
4964 | ARM::VLD4DUPq16EvenPseudo, |
4965 | ARM::VLD4DUPq32EvenPseudo }; |
4966 | static const uint16_t QOpcodes1[] = { ARM::VLD4DUPq8OddPseudo, |
4967 | ARM::VLD4DUPq16OddPseudo, |
4968 | ARM::VLD4DUPq32OddPseudo }; |
4969 | SelectVLDDup(N, /* IsIntrinsic= */ true, isUpdating: false, NumVecs: 4, |
4970 | DOpcodes, QOpcodes0, QOpcodes1); |
4971 | return; |
4972 | } |
4973 | |
4974 | case Intrinsic::arm_neon_vld2lane: { |
4975 | static const uint16_t DOpcodes[] = { ARM::VLD2LNd8Pseudo, |
4976 | ARM::VLD2LNd16Pseudo, |
4977 | ARM::VLD2LNd32Pseudo }; |
4978 | static const uint16_t QOpcodes[] = { ARM::VLD2LNq16Pseudo, |
4979 | ARM::VLD2LNq32Pseudo }; |
4980 | SelectVLDSTLane(N, IsLoad: true, isUpdating: false, NumVecs: 2, DOpcodes, QOpcodes); |
4981 | return; |
4982 | } |
4983 | |
4984 | case Intrinsic::arm_neon_vld3lane: { |
4985 | static const uint16_t DOpcodes[] = { ARM::VLD3LNd8Pseudo, |
4986 | ARM::VLD3LNd16Pseudo, |
4987 | ARM::VLD3LNd32Pseudo }; |
4988 | static const uint16_t QOpcodes[] = { ARM::VLD3LNq16Pseudo, |
4989 | ARM::VLD3LNq32Pseudo }; |
4990 | SelectVLDSTLane(N, IsLoad: true, isUpdating: false, NumVecs: 3, DOpcodes, QOpcodes); |
4991 | return; |
4992 | } |
4993 | |
4994 | case Intrinsic::arm_neon_vld4lane: { |
4995 | static const uint16_t DOpcodes[] = { ARM::VLD4LNd8Pseudo, |
4996 | ARM::VLD4LNd16Pseudo, |
4997 | ARM::VLD4LNd32Pseudo }; |
4998 | static const uint16_t QOpcodes[] = { ARM::VLD4LNq16Pseudo, |
4999 | ARM::VLD4LNq32Pseudo }; |
5000 | SelectVLDSTLane(N, IsLoad: true, isUpdating: false, NumVecs: 4, DOpcodes, QOpcodes); |
5001 | return; |
5002 | } |
5003 | |
5004 | case Intrinsic::arm_neon_vst1: { |
5005 | static const uint16_t DOpcodes[] = { ARM::VST1d8, ARM::VST1d16, |
5006 | ARM::VST1d32, ARM::VST1d64 }; |
5007 | static const uint16_t QOpcodes[] = { ARM::VST1q8, ARM::VST1q16, |
5008 | ARM::VST1q32, ARM::VST1q64 }; |
5009 | SelectVST(N, isUpdating: false, NumVecs: 1, DOpcodes, QOpcodes0: QOpcodes, QOpcodes1: nullptr); |
5010 | return; |
5011 | } |
5012 | |
5013 | case Intrinsic::arm_neon_vst1x2: { |
5014 | static const uint16_t DOpcodes[] = { ARM::VST1q8, ARM::VST1q16, |
5015 | ARM::VST1q32, ARM::VST1q64 }; |
5016 | static const uint16_t QOpcodes[] = { ARM::VST1d8QPseudo, |
5017 | ARM::VST1d16QPseudo, |
5018 | ARM::VST1d32QPseudo, |
5019 | ARM::VST1d64QPseudo }; |
5020 | SelectVST(N, isUpdating: false, NumVecs: 2, DOpcodes, QOpcodes0: QOpcodes, QOpcodes1: nullptr); |
5021 | return; |
5022 | } |
5023 | |
5024 | case Intrinsic::arm_neon_vst1x3: { |
5025 | static const uint16_t DOpcodes[] = { ARM::VST1d8TPseudo, |
5026 | ARM::VST1d16TPseudo, |
5027 | ARM::VST1d32TPseudo, |
5028 | ARM::VST1d64TPseudo }; |
5029 | static const uint16_t QOpcodes0[] = { ARM::VST1q8LowTPseudo_UPD, |
5030 | ARM::VST1q16LowTPseudo_UPD, |
5031 | ARM::VST1q32LowTPseudo_UPD, |
5032 | ARM::VST1q64LowTPseudo_UPD }; |
5033 | static const uint16_t QOpcodes1[] = { ARM::VST1q8HighTPseudo, |
5034 | ARM::VST1q16HighTPseudo, |
5035 | ARM::VST1q32HighTPseudo, |
5036 | ARM::VST1q64HighTPseudo }; |
5037 | SelectVST(N, isUpdating: false, NumVecs: 3, DOpcodes, QOpcodes0, QOpcodes1); |
5038 | return; |
5039 | } |
5040 | |
5041 | case Intrinsic::arm_neon_vst1x4: { |
5042 | static const uint16_t DOpcodes[] = { ARM::VST1d8QPseudo, |
5043 | ARM::VST1d16QPseudo, |
5044 | ARM::VST1d32QPseudo, |
5045 | ARM::VST1d64QPseudo }; |
5046 | static const uint16_t QOpcodes0[] = { ARM::VST1q8LowQPseudo_UPD, |
5047 | ARM::VST1q16LowQPseudo_UPD, |
5048 | ARM::VST1q32LowQPseudo_UPD, |
5049 | ARM::VST1q64LowQPseudo_UPD }; |
5050 | static const uint16_t QOpcodes1[] = { ARM::VST1q8HighQPseudo, |
5051 | ARM::VST1q16HighQPseudo, |
5052 | ARM::VST1q32HighQPseudo, |
5053 | ARM::VST1q64HighQPseudo }; |
5054 | SelectVST(N, isUpdating: false, NumVecs: 4, DOpcodes, QOpcodes0, QOpcodes1); |
5055 | return; |
5056 | } |
5057 | |
5058 | case Intrinsic::arm_neon_vst2: { |
5059 | static const uint16_t DOpcodes[] = { ARM::VST2d8, ARM::VST2d16, |
5060 | ARM::VST2d32, ARM::VST1q64 }; |
5061 | static const uint16_t QOpcodes[] = { ARM::VST2q8Pseudo, ARM::VST2q16Pseudo, |
5062 | ARM::VST2q32Pseudo }; |
5063 | SelectVST(N, isUpdating: false, NumVecs: 2, DOpcodes, QOpcodes0: QOpcodes, QOpcodes1: nullptr); |
5064 | return; |
5065 | } |
5066 | |
5067 | case Intrinsic::arm_neon_vst3: { |
5068 | static const uint16_t DOpcodes[] = { ARM::VST3d8Pseudo, |
5069 | ARM::VST3d16Pseudo, |
5070 | ARM::VST3d32Pseudo, |
5071 | ARM::VST1d64TPseudo }; |
5072 | static const uint16_t QOpcodes0[] = { ARM::VST3q8Pseudo_UPD, |
5073 | ARM::VST3q16Pseudo_UPD, |
5074 | ARM::VST3q32Pseudo_UPD }; |
5075 | static const uint16_t QOpcodes1[] = { ARM::VST3q8oddPseudo, |
5076 | ARM::VST3q16oddPseudo, |
5077 | ARM::VST3q32oddPseudo }; |
5078 | SelectVST(N, isUpdating: false, NumVecs: 3, DOpcodes, QOpcodes0, QOpcodes1); |
5079 | return; |
5080 | } |
5081 | |
5082 | case Intrinsic::arm_neon_vst4: { |
5083 | static const uint16_t DOpcodes[] = { ARM::VST4d8Pseudo, |
5084 | ARM::VST4d16Pseudo, |
5085 | ARM::VST4d32Pseudo, |
5086 | ARM::VST1d64QPseudo }; |
5087 | static const uint16_t QOpcodes0[] = { ARM::VST4q8Pseudo_UPD, |
5088 | ARM::VST4q16Pseudo_UPD, |
5089 | ARM::VST4q32Pseudo_UPD }; |
5090 | static const uint16_t QOpcodes1[] = { ARM::VST4q8oddPseudo, |
5091 | ARM::VST4q16oddPseudo, |
5092 | ARM::VST4q32oddPseudo }; |
5093 | SelectVST(N, isUpdating: false, NumVecs: 4, DOpcodes, QOpcodes0, QOpcodes1); |
5094 | return; |
5095 | } |
5096 | |
5097 | case Intrinsic::arm_neon_vst2lane: { |
5098 | static const uint16_t DOpcodes[] = { ARM::VST2LNd8Pseudo, |
5099 | ARM::VST2LNd16Pseudo, |
5100 | ARM::VST2LNd32Pseudo }; |
5101 | static const uint16_t QOpcodes[] = { ARM::VST2LNq16Pseudo, |
5102 | ARM::VST2LNq32Pseudo }; |
5103 | SelectVLDSTLane(N, IsLoad: false, isUpdating: false, NumVecs: 2, DOpcodes, QOpcodes); |
5104 | return; |
5105 | } |
5106 | |
5107 | case Intrinsic::arm_neon_vst3lane: { |
5108 | static const uint16_t DOpcodes[] = { ARM::VST3LNd8Pseudo, |
5109 | ARM::VST3LNd16Pseudo, |
5110 | ARM::VST3LNd32Pseudo }; |
5111 | static const uint16_t QOpcodes[] = { ARM::VST3LNq16Pseudo, |
5112 | ARM::VST3LNq32Pseudo }; |
5113 | SelectVLDSTLane(N, IsLoad: false, isUpdating: false, NumVecs: 3, DOpcodes, QOpcodes); |
5114 | return; |
5115 | } |
5116 | |
5117 | case Intrinsic::arm_neon_vst4lane: { |
5118 | static const uint16_t DOpcodes[] = { ARM::VST4LNd8Pseudo, |
5119 | ARM::VST4LNd16Pseudo, |
5120 | ARM::VST4LNd32Pseudo }; |
5121 | static const uint16_t QOpcodes[] = { ARM::VST4LNq16Pseudo, |
5122 | ARM::VST4LNq32Pseudo }; |
5123 | SelectVLDSTLane(N, IsLoad: false, isUpdating: false, NumVecs: 4, DOpcodes, QOpcodes); |
5124 | return; |
5125 | } |
5126 | |
5127 | case Intrinsic::arm_mve_vldr_gather_base_wb: |
5128 | case Intrinsic::arm_mve_vldr_gather_base_wb_predicated: { |
5129 | static const uint16_t Opcodes[] = {ARM::MVE_VLDRWU32_qi_pre, |
5130 | ARM::MVE_VLDRDU64_qi_pre}; |
5131 | SelectMVE_WB(N, Opcodes, |
5132 | Predicated: IntNo == Intrinsic::arm_mve_vldr_gather_base_wb_predicated); |
5133 | return; |
5134 | } |
5135 | |
5136 | case Intrinsic::arm_mve_vld2q: { |
5137 | static const uint16_t Opcodes8[] = {ARM::MVE_VLD20_8, ARM::MVE_VLD21_8}; |
5138 | static const uint16_t Opcodes16[] = {ARM::MVE_VLD20_16, |
5139 | ARM::MVE_VLD21_16}; |
5140 | static const uint16_t Opcodes32[] = {ARM::MVE_VLD20_32, |
5141 | ARM::MVE_VLD21_32}; |
5142 | static const uint16_t *const Opcodes[] = {Opcodes8, Opcodes16, Opcodes32}; |
5143 | SelectMVE_VLD(N, NumVecs: 2, Opcodes, HasWriteback: false); |
5144 | return; |
5145 | } |
5146 | |
5147 | case Intrinsic::arm_mve_vld4q: { |
5148 | static const uint16_t Opcodes8[] = {ARM::MVE_VLD40_8, ARM::MVE_VLD41_8, |
5149 | ARM::MVE_VLD42_8, ARM::MVE_VLD43_8}; |
5150 | static const uint16_t Opcodes16[] = {ARM::MVE_VLD40_16, ARM::MVE_VLD41_16, |
5151 | ARM::MVE_VLD42_16, |
5152 | ARM::MVE_VLD43_16}; |
5153 | static const uint16_t Opcodes32[] = {ARM::MVE_VLD40_32, ARM::MVE_VLD41_32, |
5154 | ARM::MVE_VLD42_32, |
5155 | ARM::MVE_VLD43_32}; |
5156 | static const uint16_t *const Opcodes[] = {Opcodes8, Opcodes16, Opcodes32}; |
5157 | SelectMVE_VLD(N, NumVecs: 4, Opcodes, HasWriteback: false); |
5158 | return; |
5159 | } |
5160 | } |
5161 | break; |
5162 | } |
5163 | |
5164 | case ISD::INTRINSIC_WO_CHAIN: { |
5165 | unsigned IntNo = N->getConstantOperandVal(Num: 0); |
5166 | switch (IntNo) { |
5167 | default: |
5168 | break; |
5169 | |
5170 | // Scalar f32 -> bf16 |
5171 | case Intrinsic::arm_neon_vcvtbfp2bf: { |
5172 | SDLoc dl(N); |
5173 | const SDValue &Src = N->getOperand(Num: 1); |
5174 | llvm::EVT DestTy = N->getValueType(ResNo: 0); |
5175 | SDValue Pred = getAL(CurDAG, dl); |
5176 | SDValue Reg0 = CurDAG->getRegister(Reg: 0, VT: MVT::i32); |
5177 | SDValue Ops[] = { Src, Src, Pred, Reg0 }; |
5178 | CurDAG->SelectNodeTo(N, MachineOpc: ARM::BF16_VCVTB, VT: DestTy, Ops); |
5179 | return; |
5180 | } |
5181 | |
5182 | // Vector v4f32 -> v4bf16 |
5183 | case Intrinsic::arm_neon_vcvtfp2bf: { |
5184 | SDLoc dl(N); |
5185 | const SDValue &Src = N->getOperand(Num: 1); |
5186 | SDValue Pred = getAL(CurDAG, dl); |
5187 | SDValue Reg0 = CurDAG->getRegister(Reg: 0, VT: MVT::i32); |
5188 | SDValue Ops[] = { Src, Pred, Reg0 }; |
5189 | CurDAG->SelectNodeTo(N, MachineOpc: ARM::BF16_VCVT, VT: MVT::v4bf16, Ops); |
5190 | return; |
5191 | } |
5192 | |
5193 | case Intrinsic::arm_mve_urshrl: |
5194 | SelectMVE_LongShift(N, Opcode: ARM::MVE_URSHRL, Immediate: true, HasSaturationOperand: false); |
5195 | return; |
5196 | case Intrinsic::arm_mve_uqshll: |
5197 | SelectMVE_LongShift(N, Opcode: ARM::MVE_UQSHLL, Immediate: true, HasSaturationOperand: false); |
5198 | return; |
5199 | case Intrinsic::arm_mve_srshrl: |
5200 | SelectMVE_LongShift(N, Opcode: ARM::MVE_SRSHRL, Immediate: true, HasSaturationOperand: false); |
5201 | return; |
5202 | case Intrinsic::arm_mve_sqshll: |
5203 | SelectMVE_LongShift(N, Opcode: ARM::MVE_SQSHLL, Immediate: true, HasSaturationOperand: false); |
5204 | return; |
5205 | case Intrinsic::arm_mve_uqrshll: |
5206 | SelectMVE_LongShift(N, Opcode: ARM::MVE_UQRSHLL, Immediate: false, HasSaturationOperand: true); |
5207 | return; |
5208 | case Intrinsic::arm_mve_sqrshrl: |
5209 | SelectMVE_LongShift(N, Opcode: ARM::MVE_SQRSHRL, Immediate: false, HasSaturationOperand: true); |
5210 | return; |
5211 | |
5212 | case Intrinsic::arm_mve_vadc: |
5213 | case Intrinsic::arm_mve_vadc_predicated: |
5214 | SelectMVE_VADCSBC(N, OpcodeWithCarry: ARM::MVE_VADC, OpcodeWithNoCarry: ARM::MVE_VADCI, Add: true, |
5215 | Predicated: IntNo == Intrinsic::arm_mve_vadc_predicated); |
5216 | return; |
5217 | case Intrinsic::arm_mve_vsbc: |
5218 | case Intrinsic::arm_mve_vsbc_predicated: |
5219 | SelectMVE_VADCSBC(N, OpcodeWithCarry: ARM::MVE_VSBC, OpcodeWithNoCarry: ARM::MVE_VSBCI, Add: false, |
5220 | Predicated: IntNo == Intrinsic::arm_mve_vsbc_predicated); |
5221 | return; |
5222 | case Intrinsic::arm_mve_vshlc: |
5223 | case Intrinsic::arm_mve_vshlc_predicated: |
5224 | SelectMVE_VSHLC(N, Predicated: IntNo == Intrinsic::arm_mve_vshlc_predicated); |
5225 | return; |
5226 | |
5227 | case Intrinsic::arm_mve_vmlldava: |
5228 | case Intrinsic::arm_mve_vmlldava_predicated: { |
5229 | static const uint16_t OpcodesU[] = { |
5230 | ARM::MVE_VMLALDAVu16, ARM::MVE_VMLALDAVu32, |
5231 | ARM::MVE_VMLALDAVau16, ARM::MVE_VMLALDAVau32, |
5232 | }; |
5233 | static const uint16_t OpcodesS[] = { |
5234 | ARM::MVE_VMLALDAVs16, ARM::MVE_VMLALDAVs32, |
5235 | ARM::MVE_VMLALDAVas16, ARM::MVE_VMLALDAVas32, |
5236 | ARM::MVE_VMLALDAVxs16, ARM::MVE_VMLALDAVxs32, |
5237 | ARM::MVE_VMLALDAVaxs16, ARM::MVE_VMLALDAVaxs32, |
5238 | ARM::MVE_VMLSLDAVs16, ARM::MVE_VMLSLDAVs32, |
5239 | ARM::MVE_VMLSLDAVas16, ARM::MVE_VMLSLDAVas32, |
5240 | ARM::MVE_VMLSLDAVxs16, ARM::MVE_VMLSLDAVxs32, |
5241 | ARM::MVE_VMLSLDAVaxs16, ARM::MVE_VMLSLDAVaxs32, |
5242 | }; |
5243 | SelectMVE_VMLLDAV(N, Predicated: IntNo == Intrinsic::arm_mve_vmlldava_predicated, |
5244 | OpcodesS, OpcodesU); |
5245 | return; |
5246 | } |
5247 | |
5248 | case Intrinsic::arm_mve_vrmlldavha: |
5249 | case Intrinsic::arm_mve_vrmlldavha_predicated: { |
5250 | static const uint16_t OpcodesU[] = { |
5251 | ARM::MVE_VRMLALDAVHu32, ARM::MVE_VRMLALDAVHau32, |
5252 | }; |
5253 | static const uint16_t OpcodesS[] = { |
5254 | ARM::MVE_VRMLALDAVHs32, ARM::MVE_VRMLALDAVHas32, |
5255 | ARM::MVE_VRMLALDAVHxs32, ARM::MVE_VRMLALDAVHaxs32, |
5256 | ARM::MVE_VRMLSLDAVHs32, ARM::MVE_VRMLSLDAVHas32, |
5257 | ARM::MVE_VRMLSLDAVHxs32, ARM::MVE_VRMLSLDAVHaxs32, |
5258 | }; |
5259 | SelectMVE_VRMLLDAVH(N, Predicated: IntNo == Intrinsic::arm_mve_vrmlldavha_predicated, |
5260 | OpcodesS, OpcodesU); |
5261 | return; |
5262 | } |
5263 | |
5264 | case Intrinsic::arm_mve_vidup: |
5265 | case Intrinsic::arm_mve_vidup_predicated: { |
5266 | static const uint16_t Opcodes[] = { |
5267 | ARM::MVE_VIDUPu8, ARM::MVE_VIDUPu16, ARM::MVE_VIDUPu32, |
5268 | }; |
5269 | SelectMVE_VxDUP(N, Opcodes, Wrapping: false, |
5270 | Predicated: IntNo == Intrinsic::arm_mve_vidup_predicated); |
5271 | return; |
5272 | } |
5273 | |
5274 | case Intrinsic::arm_mve_vddup: |
5275 | case Intrinsic::arm_mve_vddup_predicated: { |
5276 | static const uint16_t Opcodes[] = { |
5277 | ARM::MVE_VDDUPu8, ARM::MVE_VDDUPu16, ARM::MVE_VDDUPu32, |
5278 | }; |
5279 | SelectMVE_VxDUP(N, Opcodes, Wrapping: false, |
5280 | Predicated: IntNo == Intrinsic::arm_mve_vddup_predicated); |
5281 | return; |
5282 | } |
5283 | |
5284 | case Intrinsic::arm_mve_viwdup: |
5285 | case Intrinsic::arm_mve_viwdup_predicated: { |
5286 | static const uint16_t Opcodes[] = { |
5287 | ARM::MVE_VIWDUPu8, ARM::MVE_VIWDUPu16, ARM::MVE_VIWDUPu32, |
5288 | }; |
5289 | SelectMVE_VxDUP(N, Opcodes, Wrapping: true, |
5290 | Predicated: IntNo == Intrinsic::arm_mve_viwdup_predicated); |
5291 | return; |
5292 | } |
5293 | |
5294 | case Intrinsic::arm_mve_vdwdup: |
5295 | case Intrinsic::arm_mve_vdwdup_predicated: { |
5296 | static const uint16_t Opcodes[] = { |
5297 | ARM::MVE_VDWDUPu8, ARM::MVE_VDWDUPu16, ARM::MVE_VDWDUPu32, |
5298 | }; |
5299 | SelectMVE_VxDUP(N, Opcodes, Wrapping: true, |
5300 | Predicated: IntNo == Intrinsic::arm_mve_vdwdup_predicated); |
5301 | return; |
5302 | } |
5303 | |
5304 | case Intrinsic::arm_cde_cx1d: |
5305 | case Intrinsic::arm_cde_cx1da: |
5306 | case Intrinsic::arm_cde_cx2d: |
5307 | case Intrinsic::arm_cde_cx2da: |
5308 | case Intrinsic::arm_cde_cx3d: |
5309 | case Intrinsic::arm_cde_cx3da: { |
5310 | bool HasAccum = IntNo == Intrinsic::arm_cde_cx1da || |
5311 | IntNo == Intrinsic::arm_cde_cx2da || |
5312 | IntNo == Intrinsic::arm_cde_cx3da; |
5313 | size_t ; |
5314 | uint16_t Opcode; |
5315 | switch (IntNo) { |
5316 | case Intrinsic::arm_cde_cx1d: |
5317 | case Intrinsic::arm_cde_cx1da: |
5318 | NumExtraOps = 0; |
5319 | Opcode = HasAccum ? ARM::CDE_CX1DA : ARM::CDE_CX1D; |
5320 | break; |
5321 | case Intrinsic::arm_cde_cx2d: |
5322 | case Intrinsic::arm_cde_cx2da: |
5323 | NumExtraOps = 1; |
5324 | Opcode = HasAccum ? ARM::CDE_CX2DA : ARM::CDE_CX2D; |
5325 | break; |
5326 | case Intrinsic::arm_cde_cx3d: |
5327 | case Intrinsic::arm_cde_cx3da: |
5328 | NumExtraOps = 2; |
5329 | Opcode = HasAccum ? ARM::CDE_CX3DA : ARM::CDE_CX3D; |
5330 | break; |
5331 | default: |
5332 | llvm_unreachable("Unexpected opcode" ); |
5333 | } |
5334 | SelectCDE_CXxD(N, Opcode, NumExtraOps, HasAccum); |
5335 | return; |
5336 | } |
5337 | } |
5338 | break; |
5339 | } |
5340 | |
5341 | case ISD::ATOMIC_CMP_SWAP: |
5342 | SelectCMP_SWAP(N); |
5343 | return; |
5344 | } |
5345 | |
5346 | SelectCode(N); |
5347 | } |
5348 | |
5349 | // Inspect a register string of the form |
5350 | // cp<coprocessor>:<opc1>:c<CRn>:c<CRm>:<opc2> (32bit) or |
5351 | // cp<coprocessor>:<opc1>:c<CRm> (64bit) inspect the fields of the string |
5352 | // and obtain the integer operands from them, adding these operands to the |
5353 | // provided vector. |
5354 | static void getIntOperandsFromRegisterString(StringRef RegString, |
5355 | SelectionDAG *CurDAG, |
5356 | const SDLoc &DL, |
5357 | std::vector<SDValue> &Ops) { |
5358 | SmallVector<StringRef, 5> Fields; |
5359 | RegString.split(A&: Fields, Separator: ':'); |
5360 | |
5361 | if (Fields.size() > 1) { |
5362 | bool AllIntFields = true; |
5363 | |
5364 | for (StringRef Field : Fields) { |
5365 | // Need to trim out leading 'cp' characters and get the integer field. |
5366 | unsigned IntField; |
5367 | AllIntFields &= !Field.trim(Chars: "CPcp" ).getAsInteger(Radix: 10, Result&: IntField); |
5368 | Ops.push_back(x: CurDAG->getTargetConstant(Val: IntField, DL, VT: MVT::i32)); |
5369 | } |
5370 | |
5371 | assert(AllIntFields && |
5372 | "Unexpected non-integer value in special register string." ); |
5373 | (void)AllIntFields; |
5374 | } |
5375 | } |
5376 | |
5377 | // Maps a Banked Register string to its mask value. The mask value returned is |
5378 | // for use in the MRSbanked / MSRbanked instruction nodes as the Banked Register |
5379 | // mask operand, which expresses which register is to be used, e.g. r8, and in |
5380 | // which mode it is to be used, e.g. usr. Returns -1 to signify that the string |
5381 | // was invalid. |
5382 | static inline int getBankedRegisterMask(StringRef RegString) { |
5383 | auto TheReg = ARMBankedReg::lookupBankedRegByName(Name: RegString.lower()); |
5384 | if (!TheReg) |
5385 | return -1; |
5386 | return TheReg->Encoding; |
5387 | } |
5388 | |
5389 | // The flags here are common to those allowed for apsr in the A class cores and |
5390 | // those allowed for the special registers in the M class cores. Returns a |
5391 | // value representing which flags were present, -1 if invalid. |
5392 | static inline int getMClassFlagsMask(StringRef Flags) { |
5393 | return StringSwitch<int>(Flags) |
5394 | .Case(S: "" , Value: 0x2) // no flags means nzcvq for psr registers, and 0x2 is |
5395 | // correct when flags are not permitted |
5396 | .Case(S: "g" , Value: 0x1) |
5397 | .Case(S: "nzcvq" , Value: 0x2) |
5398 | .Case(S: "nzcvqg" , Value: 0x3) |
5399 | .Default(Value: -1); |
5400 | } |
5401 | |
5402 | // Maps MClass special registers string to its value for use in the |
5403 | // t2MRS_M/t2MSR_M instruction nodes as the SYSm value operand. |
5404 | // Returns -1 to signify that the string was invalid. |
5405 | static int getMClassRegisterMask(StringRef Reg, const ARMSubtarget *Subtarget) { |
5406 | auto TheReg = ARMSysReg::lookupMClassSysRegByName(Name: Reg); |
5407 | const FeatureBitset &FeatureBits = Subtarget->getFeatureBits(); |
5408 | if (!TheReg || !TheReg->hasRequiredFeatures(ActiveFeatures: FeatureBits)) |
5409 | return -1; |
5410 | return (int)(TheReg->Encoding & 0xFFF); // SYSm value |
5411 | } |
5412 | |
5413 | static int getARClassRegisterMask(StringRef Reg, StringRef Flags) { |
5414 | // The mask operand contains the special register (R Bit) in bit 4, whether |
5415 | // the register is spsr (R bit is 1) or one of cpsr/apsr (R bit is 0), and |
5416 | // bits 3-0 contains the fields to be accessed in the special register, set by |
5417 | // the flags provided with the register. |
5418 | int Mask = 0; |
5419 | if (Reg == "apsr" ) { |
5420 | // The flags permitted for apsr are the same flags that are allowed in |
5421 | // M class registers. We get the flag value and then shift the flags into |
5422 | // the correct place to combine with the mask. |
5423 | Mask = getMClassFlagsMask(Flags); |
5424 | if (Mask == -1) |
5425 | return -1; |
5426 | return Mask << 2; |
5427 | } |
5428 | |
5429 | if (Reg != "cpsr" && Reg != "spsr" ) { |
5430 | return -1; |
5431 | } |
5432 | |
5433 | // This is the same as if the flags were "fc" |
5434 | if (Flags.empty() || Flags == "all" ) |
5435 | return Mask | 0x9; |
5436 | |
5437 | // Inspect the supplied flags string and set the bits in the mask for |
5438 | // the relevant and valid flags allowed for cpsr and spsr. |
5439 | for (char Flag : Flags) { |
5440 | int FlagVal; |
5441 | switch (Flag) { |
5442 | case 'c': |
5443 | FlagVal = 0x1; |
5444 | break; |
5445 | case 'x': |
5446 | FlagVal = 0x2; |
5447 | break; |
5448 | case 's': |
5449 | FlagVal = 0x4; |
5450 | break; |
5451 | case 'f': |
5452 | FlagVal = 0x8; |
5453 | break; |
5454 | default: |
5455 | FlagVal = 0; |
5456 | } |
5457 | |
5458 | // This avoids allowing strings where the same flag bit appears twice. |
5459 | if (!FlagVal || (Mask & FlagVal)) |
5460 | return -1; |
5461 | Mask |= FlagVal; |
5462 | } |
5463 | |
5464 | // If the register is spsr then we need to set the R bit. |
5465 | if (Reg == "spsr" ) |
5466 | Mask |= 0x10; |
5467 | |
5468 | return Mask; |
5469 | } |
5470 | |
5471 | // Lower the read_register intrinsic to ARM specific DAG nodes |
5472 | // using the supplied metadata string to select the instruction node to use |
5473 | // and the registers/masks to construct as operands for the node. |
5474 | bool ARMDAGToDAGISel::tryReadRegister(SDNode *N){ |
5475 | const auto *MD = cast<MDNodeSDNode>(Val: N->getOperand(Num: 1)); |
5476 | const auto *RegString = cast<MDString>(Val: MD->getMD()->getOperand(I: 0)); |
5477 | bool IsThumb2 = Subtarget->isThumb2(); |
5478 | SDLoc DL(N); |
5479 | |
5480 | std::vector<SDValue> Ops; |
5481 | getIntOperandsFromRegisterString(RegString: RegString->getString(), CurDAG, DL, Ops); |
5482 | |
5483 | if (!Ops.empty()) { |
5484 | // If the special register string was constructed of fields (as defined |
5485 | // in the ACLE) then need to lower to MRC node (32 bit) or |
5486 | // MRRC node(64 bit), we can make the distinction based on the number of |
5487 | // operands we have. |
5488 | unsigned Opcode; |
5489 | SmallVector<EVT, 3> ResTypes; |
5490 | if (Ops.size() == 5){ |
5491 | Opcode = IsThumb2 ? ARM::t2MRC : ARM::MRC; |
5492 | ResTypes.append(IL: { MVT::i32, MVT::Other }); |
5493 | } else { |
5494 | assert(Ops.size() == 3 && |
5495 | "Invalid number of fields in special register string." ); |
5496 | Opcode = IsThumb2 ? ARM::t2MRRC : ARM::MRRC; |
5497 | ResTypes.append(IL: { MVT::i32, MVT::i32, MVT::Other }); |
5498 | } |
5499 | |
5500 | Ops.push_back(x: getAL(CurDAG, dl: DL)); |
5501 | Ops.push_back(x: CurDAG->getRegister(Reg: 0, VT: MVT::i32)); |
5502 | Ops.push_back(x: N->getOperand(Num: 0)); |
5503 | ReplaceNode(F: N, T: CurDAG->getMachineNode(Opcode, dl: DL, ResultTys: ResTypes, Ops)); |
5504 | return true; |
5505 | } |
5506 | |
5507 | std::string SpecialReg = RegString->getString().lower(); |
5508 | |
5509 | int BankedReg = getBankedRegisterMask(RegString: SpecialReg); |
5510 | if (BankedReg != -1) { |
5511 | Ops = { CurDAG->getTargetConstant(Val: BankedReg, DL, VT: MVT::i32), |
5512 | getAL(CurDAG, dl: DL), CurDAG->getRegister(Reg: 0, VT: MVT::i32), |
5513 | N->getOperand(Num: 0) }; |
5514 | ReplaceNode( |
5515 | F: N, T: CurDAG->getMachineNode(Opcode: IsThumb2 ? ARM::t2MRSbanked : ARM::MRSbanked, |
5516 | dl: DL, VT1: MVT::i32, VT2: MVT::Other, Ops)); |
5517 | return true; |
5518 | } |
5519 | |
5520 | // The VFP registers are read by creating SelectionDAG nodes with opcodes |
5521 | // corresponding to the register that is being read from. So we switch on the |
5522 | // string to find which opcode we need to use. |
5523 | unsigned Opcode = StringSwitch<unsigned>(SpecialReg) |
5524 | .Case(S: "fpscr" , Value: ARM::VMRS) |
5525 | .Case(S: "fpexc" , Value: ARM::VMRS_FPEXC) |
5526 | .Case(S: "fpsid" , Value: ARM::VMRS_FPSID) |
5527 | .Case(S: "mvfr0" , Value: ARM::VMRS_MVFR0) |
5528 | .Case(S: "mvfr1" , Value: ARM::VMRS_MVFR1) |
5529 | .Case(S: "mvfr2" , Value: ARM::VMRS_MVFR2) |
5530 | .Case(S: "fpinst" , Value: ARM::VMRS_FPINST) |
5531 | .Case(S: "fpinst2" , Value: ARM::VMRS_FPINST2) |
5532 | .Default(Value: 0); |
5533 | |
5534 | // If an opcode was found then we can lower the read to a VFP instruction. |
5535 | if (Opcode) { |
5536 | if (!Subtarget->hasVFP2Base()) |
5537 | return false; |
5538 | if (Opcode == ARM::VMRS_MVFR2 && !Subtarget->hasFPARMv8Base()) |
5539 | return false; |
5540 | |
5541 | Ops = { getAL(CurDAG, dl: DL), CurDAG->getRegister(Reg: 0, VT: MVT::i32), |
5542 | N->getOperand(Num: 0) }; |
5543 | ReplaceNode(F: N, |
5544 | T: CurDAG->getMachineNode(Opcode, dl: DL, VT1: MVT::i32, VT2: MVT::Other, Ops)); |
5545 | return true; |
5546 | } |
5547 | |
5548 | // If the target is M Class then need to validate that the register string |
5549 | // is an acceptable value, so check that a mask can be constructed from the |
5550 | // string. |
5551 | if (Subtarget->isMClass()) { |
5552 | int SYSmValue = getMClassRegisterMask(Reg: SpecialReg, Subtarget); |
5553 | if (SYSmValue == -1) |
5554 | return false; |
5555 | |
5556 | SDValue Ops[] = { CurDAG->getTargetConstant(Val: SYSmValue, DL, VT: MVT::i32), |
5557 | getAL(CurDAG, dl: DL), CurDAG->getRegister(Reg: 0, VT: MVT::i32), |
5558 | N->getOperand(Num: 0) }; |
5559 | ReplaceNode( |
5560 | F: N, T: CurDAG->getMachineNode(Opcode: ARM::t2MRS_M, dl: DL, VT1: MVT::i32, VT2: MVT::Other, Ops)); |
5561 | return true; |
5562 | } |
5563 | |
5564 | // Here we know the target is not M Class so we need to check if it is one |
5565 | // of the remaining possible values which are apsr, cpsr or spsr. |
5566 | if (SpecialReg == "apsr" || SpecialReg == "cpsr" ) { |
5567 | Ops = { getAL(CurDAG, dl: DL), CurDAG->getRegister(Reg: 0, VT: MVT::i32), |
5568 | N->getOperand(Num: 0) }; |
5569 | ReplaceNode(F: N, T: CurDAG->getMachineNode(Opcode: IsThumb2 ? ARM::t2MRS_AR : ARM::MRS, |
5570 | dl: DL, VT1: MVT::i32, VT2: MVT::Other, Ops)); |
5571 | return true; |
5572 | } |
5573 | |
5574 | if (SpecialReg == "spsr" ) { |
5575 | Ops = { getAL(CurDAG, dl: DL), CurDAG->getRegister(Reg: 0, VT: MVT::i32), |
5576 | N->getOperand(Num: 0) }; |
5577 | ReplaceNode( |
5578 | F: N, T: CurDAG->getMachineNode(Opcode: IsThumb2 ? ARM::t2MRSsys_AR : ARM::MRSsys, dl: DL, |
5579 | VT1: MVT::i32, VT2: MVT::Other, Ops)); |
5580 | return true; |
5581 | } |
5582 | |
5583 | return false; |
5584 | } |
5585 | |
5586 | // Lower the write_register intrinsic to ARM specific DAG nodes |
5587 | // using the supplied metadata string to select the instruction node to use |
5588 | // and the registers/masks to use in the nodes |
5589 | bool ARMDAGToDAGISel::tryWriteRegister(SDNode *N){ |
5590 | const auto *MD = cast<MDNodeSDNode>(Val: N->getOperand(Num: 1)); |
5591 | const auto *RegString = cast<MDString>(Val: MD->getMD()->getOperand(I: 0)); |
5592 | bool IsThumb2 = Subtarget->isThumb2(); |
5593 | SDLoc DL(N); |
5594 | |
5595 | std::vector<SDValue> Ops; |
5596 | getIntOperandsFromRegisterString(RegString: RegString->getString(), CurDAG, DL, Ops); |
5597 | |
5598 | if (!Ops.empty()) { |
5599 | // If the special register string was constructed of fields (as defined |
5600 | // in the ACLE) then need to lower to MCR node (32 bit) or |
5601 | // MCRR node(64 bit), we can make the distinction based on the number of |
5602 | // operands we have. |
5603 | unsigned Opcode; |
5604 | if (Ops.size() == 5) { |
5605 | Opcode = IsThumb2 ? ARM::t2MCR : ARM::MCR; |
5606 | Ops.insert(position: Ops.begin()+2, x: N->getOperand(Num: 2)); |
5607 | } else { |
5608 | assert(Ops.size() == 3 && |
5609 | "Invalid number of fields in special register string." ); |
5610 | Opcode = IsThumb2 ? ARM::t2MCRR : ARM::MCRR; |
5611 | SDValue WriteValue[] = { N->getOperand(Num: 2), N->getOperand(Num: 3) }; |
5612 | Ops.insert(position: Ops.begin()+2, first: WriteValue, last: WriteValue+2); |
5613 | } |
5614 | |
5615 | Ops.push_back(x: getAL(CurDAG, dl: DL)); |
5616 | Ops.push_back(x: CurDAG->getRegister(Reg: 0, VT: MVT::i32)); |
5617 | Ops.push_back(x: N->getOperand(Num: 0)); |
5618 | |
5619 | ReplaceNode(F: N, T: CurDAG->getMachineNode(Opcode, dl: DL, VT: MVT::Other, Ops)); |
5620 | return true; |
5621 | } |
5622 | |
5623 | std::string SpecialReg = RegString->getString().lower(); |
5624 | int BankedReg = getBankedRegisterMask(RegString: SpecialReg); |
5625 | if (BankedReg != -1) { |
5626 | Ops = { CurDAG->getTargetConstant(Val: BankedReg, DL, VT: MVT::i32), N->getOperand(Num: 2), |
5627 | getAL(CurDAG, dl: DL), CurDAG->getRegister(Reg: 0, VT: MVT::i32), |
5628 | N->getOperand(Num: 0) }; |
5629 | ReplaceNode( |
5630 | F: N, T: CurDAG->getMachineNode(Opcode: IsThumb2 ? ARM::t2MSRbanked : ARM::MSRbanked, |
5631 | dl: DL, VT: MVT::Other, Ops)); |
5632 | return true; |
5633 | } |
5634 | |
5635 | // The VFP registers are written to by creating SelectionDAG nodes with |
5636 | // opcodes corresponding to the register that is being written. So we switch |
5637 | // on the string to find which opcode we need to use. |
5638 | unsigned Opcode = StringSwitch<unsigned>(SpecialReg) |
5639 | .Case(S: "fpscr" , Value: ARM::VMSR) |
5640 | .Case(S: "fpexc" , Value: ARM::VMSR_FPEXC) |
5641 | .Case(S: "fpsid" , Value: ARM::VMSR_FPSID) |
5642 | .Case(S: "fpinst" , Value: ARM::VMSR_FPINST) |
5643 | .Case(S: "fpinst2" , Value: ARM::VMSR_FPINST2) |
5644 | .Default(Value: 0); |
5645 | |
5646 | if (Opcode) { |
5647 | if (!Subtarget->hasVFP2Base()) |
5648 | return false; |
5649 | Ops = { N->getOperand(Num: 2), getAL(CurDAG, dl: DL), |
5650 | CurDAG->getRegister(Reg: 0, VT: MVT::i32), N->getOperand(Num: 0) }; |
5651 | ReplaceNode(F: N, T: CurDAG->getMachineNode(Opcode, dl: DL, VT: MVT::Other, Ops)); |
5652 | return true; |
5653 | } |
5654 | |
5655 | std::pair<StringRef, StringRef> Fields; |
5656 | Fields = StringRef(SpecialReg).rsplit(Separator: '_'); |
5657 | std::string Reg = Fields.first.str(); |
5658 | StringRef Flags = Fields.second; |
5659 | |
5660 | // If the target was M Class then need to validate the special register value |
5661 | // and retrieve the mask for use in the instruction node. |
5662 | if (Subtarget->isMClass()) { |
5663 | int SYSmValue = getMClassRegisterMask(Reg: SpecialReg, Subtarget); |
5664 | if (SYSmValue == -1) |
5665 | return false; |
5666 | |
5667 | SDValue Ops[] = { CurDAG->getTargetConstant(Val: SYSmValue, DL, VT: MVT::i32), |
5668 | N->getOperand(Num: 2), getAL(CurDAG, dl: DL), |
5669 | CurDAG->getRegister(Reg: 0, VT: MVT::i32), N->getOperand(Num: 0) }; |
5670 | ReplaceNode(F: N, T: CurDAG->getMachineNode(Opcode: ARM::t2MSR_M, dl: DL, VT: MVT::Other, Ops)); |
5671 | return true; |
5672 | } |
5673 | |
5674 | // We then check to see if a valid mask can be constructed for one of the |
5675 | // register string values permitted for the A and R class cores. These values |
5676 | // are apsr, spsr and cpsr; these are also valid on older cores. |
5677 | int Mask = getARClassRegisterMask(Reg, Flags); |
5678 | if (Mask != -1) { |
5679 | Ops = { CurDAG->getTargetConstant(Val: Mask, DL, VT: MVT::i32), N->getOperand(Num: 2), |
5680 | getAL(CurDAG, dl: DL), CurDAG->getRegister(Reg: 0, VT: MVT::i32), |
5681 | N->getOperand(Num: 0) }; |
5682 | ReplaceNode(F: N, T: CurDAG->getMachineNode(Opcode: IsThumb2 ? ARM::t2MSR_AR : ARM::MSR, |
5683 | dl: DL, VT: MVT::Other, Ops)); |
5684 | return true; |
5685 | } |
5686 | |
5687 | return false; |
5688 | } |
5689 | |
5690 | bool ARMDAGToDAGISel::tryInlineAsm(SDNode *N){ |
5691 | std::vector<SDValue> AsmNodeOperands; |
5692 | InlineAsm::Flag Flag; |
5693 | bool Changed = false; |
5694 | unsigned NumOps = N->getNumOperands(); |
5695 | |
5696 | // Normally, i64 data is bounded to two arbitrary GRPs for "%r" constraint. |
5697 | // However, some instrstions (e.g. ldrexd/strexd in ARM mode) require |
5698 | // (even/even+1) GPRs and use %n and %Hn to refer to the individual regs |
5699 | // respectively. Since there is no constraint to explicitly specify a |
5700 | // reg pair, we use GPRPair reg class for "%r" for 64-bit data. For Thumb, |
5701 | // the 64-bit data may be referred by H, Q, R modifiers, so we still pack |
5702 | // them into a GPRPair. |
5703 | |
5704 | SDLoc dl(N); |
5705 | SDValue Glue = N->getGluedNode() ? N->getOperand(Num: NumOps - 1) : SDValue(); |
5706 | |
5707 | SmallVector<bool, 8> OpChanged; |
5708 | // Glue node will be appended late. |
5709 | for(unsigned i = 0, e = N->getGluedNode() ? NumOps - 1 : NumOps; i < e; ++i) { |
5710 | SDValue op = N->getOperand(Num: i); |
5711 | AsmNodeOperands.push_back(x: op); |
5712 | |
5713 | if (i < InlineAsm::Op_FirstOperand) |
5714 | continue; |
5715 | |
5716 | if (const auto *C = dyn_cast<ConstantSDNode>(Val: N->getOperand(Num: i))) |
5717 | Flag = InlineAsm::Flag(C->getZExtValue()); |
5718 | else |
5719 | continue; |
5720 | |
5721 | // Immediate operands to inline asm in the SelectionDAG are modeled with |
5722 | // two operands. The first is a constant of value InlineAsm::Kind::Imm, and |
5723 | // the second is a constant with the value of the immediate. If we get here |
5724 | // and we have a Kind::Imm, skip the next operand, and continue. |
5725 | if (Flag.isImmKind()) { |
5726 | SDValue op = N->getOperand(Num: ++i); |
5727 | AsmNodeOperands.push_back(x: op); |
5728 | continue; |
5729 | } |
5730 | |
5731 | const unsigned NumRegs = Flag.getNumOperandRegisters(); |
5732 | if (NumRegs) |
5733 | OpChanged.push_back(Elt: false); |
5734 | |
5735 | unsigned DefIdx = 0; |
5736 | bool IsTiedToChangedOp = false; |
5737 | // If it's a use that is tied with a previous def, it has no |
5738 | // reg class constraint. |
5739 | if (Changed && Flag.isUseOperandTiedToDef(Idx&: DefIdx)) |
5740 | IsTiedToChangedOp = OpChanged[DefIdx]; |
5741 | |
5742 | // Memory operands to inline asm in the SelectionDAG are modeled with two |
5743 | // operands: a constant of value InlineAsm::Kind::Mem followed by the input |
5744 | // operand. If we get here and we have a Kind::Mem, skip the next operand |
5745 | // (so it doesn't get misinterpreted), and continue. We do this here because |
5746 | // it's important to update the OpChanged array correctly before moving on. |
5747 | if (Flag.isMemKind()) { |
5748 | SDValue op = N->getOperand(Num: ++i); |
5749 | AsmNodeOperands.push_back(x: op); |
5750 | continue; |
5751 | } |
5752 | |
5753 | if (!Flag.isRegUseKind() && !Flag.isRegDefKind() && |
5754 | !Flag.isRegDefEarlyClobberKind()) |
5755 | continue; |
5756 | |
5757 | unsigned RC; |
5758 | const bool HasRC = Flag.hasRegClassConstraint(RC); |
5759 | if ((!IsTiedToChangedOp && (!HasRC || RC != ARM::GPRRegClassID)) |
5760 | || NumRegs != 2) |
5761 | continue; |
5762 | |
5763 | assert((i+2 < NumOps) && "Invalid number of operands in inline asm" ); |
5764 | SDValue V0 = N->getOperand(Num: i+1); |
5765 | SDValue V1 = N->getOperand(Num: i+2); |
5766 | Register Reg0 = cast<RegisterSDNode>(Val&: V0)->getReg(); |
5767 | Register Reg1 = cast<RegisterSDNode>(Val&: V1)->getReg(); |
5768 | SDValue PairedReg; |
5769 | MachineRegisterInfo &MRI = MF->getRegInfo(); |
5770 | |
5771 | if (Flag.isRegDefKind() || Flag.isRegDefEarlyClobberKind()) { |
5772 | // Replace the two GPRs with 1 GPRPair and copy values from GPRPair to |
5773 | // the original GPRs. |
5774 | |
5775 | Register GPVR = MRI.createVirtualRegister(RegClass: &ARM::GPRPairRegClass); |
5776 | PairedReg = CurDAG->getRegister(Reg: GPVR, VT: MVT::Untyped); |
5777 | SDValue Chain = SDValue(N,0); |
5778 | |
5779 | SDNode *GU = N->getGluedUser(); |
5780 | SDValue RegCopy = CurDAG->getCopyFromReg(Chain, dl, Reg: GPVR, VT: MVT::Untyped, |
5781 | Glue: Chain.getValue(R: 1)); |
5782 | |
5783 | // Extract values from a GPRPair reg and copy to the original GPR reg. |
5784 | SDValue Sub0 = CurDAG->getTargetExtractSubreg(SRIdx: ARM::gsub_0, DL: dl, VT: MVT::i32, |
5785 | Operand: RegCopy); |
5786 | SDValue Sub1 = CurDAG->getTargetExtractSubreg(SRIdx: ARM::gsub_1, DL: dl, VT: MVT::i32, |
5787 | Operand: RegCopy); |
5788 | SDValue T0 = CurDAG->getCopyToReg(Chain: Sub0, dl, Reg: Reg0, N: Sub0, |
5789 | Glue: RegCopy.getValue(R: 1)); |
5790 | SDValue T1 = CurDAG->getCopyToReg(Chain: Sub1, dl, Reg: Reg1, N: Sub1, Glue: T0.getValue(R: 1)); |
5791 | |
5792 | // Update the original glue user. |
5793 | std::vector<SDValue> Ops(GU->op_begin(), GU->op_end()-1); |
5794 | Ops.push_back(x: T1.getValue(R: 1)); |
5795 | CurDAG->UpdateNodeOperands(N: GU, Ops); |
5796 | } else { |
5797 | // For Kind == InlineAsm::Kind::RegUse, we first copy two GPRs into a |
5798 | // GPRPair and then pass the GPRPair to the inline asm. |
5799 | SDValue Chain = AsmNodeOperands[InlineAsm::Op_InputChain]; |
5800 | |
5801 | // As REG_SEQ doesn't take RegisterSDNode, we copy them first. |
5802 | SDValue T0 = CurDAG->getCopyFromReg(Chain, dl, Reg: Reg0, VT: MVT::i32, |
5803 | Glue: Chain.getValue(R: 1)); |
5804 | SDValue T1 = CurDAG->getCopyFromReg(Chain, dl, Reg: Reg1, VT: MVT::i32, |
5805 | Glue: T0.getValue(R: 1)); |
5806 | SDValue Pair = SDValue(createGPRPairNode(VT: MVT::Untyped, V0: T0, V1: T1), 0); |
5807 | |
5808 | // Copy REG_SEQ into a GPRPair-typed VR and replace the original two |
5809 | // i32 VRs of inline asm with it. |
5810 | Register GPVR = MRI.createVirtualRegister(RegClass: &ARM::GPRPairRegClass); |
5811 | PairedReg = CurDAG->getRegister(Reg: GPVR, VT: MVT::Untyped); |
5812 | Chain = CurDAG->getCopyToReg(Chain: T1, dl, Reg: GPVR, N: Pair, Glue: T1.getValue(R: 1)); |
5813 | |
5814 | AsmNodeOperands[InlineAsm::Op_InputChain] = Chain; |
5815 | Glue = Chain.getValue(R: 1); |
5816 | } |
5817 | |
5818 | Changed = true; |
5819 | |
5820 | if(PairedReg.getNode()) { |
5821 | OpChanged[OpChanged.size() -1 ] = true; |
5822 | Flag = InlineAsm::Flag(Flag.getKind(), 1 /* RegNum*/); |
5823 | if (IsTiedToChangedOp) |
5824 | Flag.setMatchingOp(DefIdx); |
5825 | else |
5826 | Flag.setRegClass(ARM::GPRPairRegClassID); |
5827 | // Replace the current flag. |
5828 | AsmNodeOperands[AsmNodeOperands.size() -1] = CurDAG->getTargetConstant( |
5829 | Val: Flag, DL: dl, VT: MVT::i32); |
5830 | // Add the new register node and skip the original two GPRs. |
5831 | AsmNodeOperands.push_back(x: PairedReg); |
5832 | // Skip the next two GPRs. |
5833 | i += 2; |
5834 | } |
5835 | } |
5836 | |
5837 | if (Glue.getNode()) |
5838 | AsmNodeOperands.push_back(x: Glue); |
5839 | if (!Changed) |
5840 | return false; |
5841 | |
5842 | SDValue New = CurDAG->getNode(Opcode: N->getOpcode(), DL: SDLoc(N), |
5843 | VTList: CurDAG->getVTList(VT1: MVT::Other, VT2: MVT::Glue), Ops: AsmNodeOperands); |
5844 | New->setNodeId(-1); |
5845 | ReplaceNode(F: N, T: New.getNode()); |
5846 | return true; |
5847 | } |
5848 | |
5849 | bool ARMDAGToDAGISel::SelectInlineAsmMemoryOperand( |
5850 | const SDValue &Op, InlineAsm::ConstraintCode ConstraintID, |
5851 | std::vector<SDValue> &OutOps) { |
5852 | switch(ConstraintID) { |
5853 | default: |
5854 | llvm_unreachable("Unexpected asm memory constraint" ); |
5855 | case InlineAsm::ConstraintCode::m: |
5856 | case InlineAsm::ConstraintCode::o: |
5857 | case InlineAsm::ConstraintCode::Q: |
5858 | case InlineAsm::ConstraintCode::Um: |
5859 | case InlineAsm::ConstraintCode::Un: |
5860 | case InlineAsm::ConstraintCode::Uq: |
5861 | case InlineAsm::ConstraintCode::Us: |
5862 | case InlineAsm::ConstraintCode::Ut: |
5863 | case InlineAsm::ConstraintCode::Uv: |
5864 | case InlineAsm::ConstraintCode::Uy: |
5865 | // Require the address to be in a register. That is safe for all ARM |
5866 | // variants and it is hard to do anything much smarter without knowing |
5867 | // how the operand is used. |
5868 | OutOps.push_back(x: Op); |
5869 | return false; |
5870 | } |
5871 | return true; |
5872 | } |
5873 | |
5874 | /// createARMISelDag - This pass converts a legalized DAG into a |
5875 | /// ARM-specific DAG, ready for instruction scheduling. |
5876 | /// |
5877 | FunctionPass *llvm::createARMISelDag(ARMBaseTargetMachine &TM, |
5878 | CodeGenOptLevel OptLevel) { |
5879 | return new ARMDAGToDAGISelLegacy(TM, OptLevel); |
5880 | } |
5881 | |