1 //===-- HexagonISelDAGToDAG.cpp - A dag to dag inst selector for Hexagon --===//
2 //
3 // The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 //
10 // This file defines an instruction selector for the Hexagon target.
11 //
12 //===----------------------------------------------------------------------===//
13
14 #include "Hexagon.h"
15 #include "HexagonISelLowering.h"
16 #include "HexagonMachineFunctionInfo.h"
17 #include "HexagonTargetMachine.h"
18 #include "llvm/ADT/DenseMap.h"
19 #include "llvm/CodeGen/FunctionLoweringInfo.h"
20 #include "llvm/CodeGen/MachineInstrBuilder.h"
21 #include "llvm/CodeGen/SelectionDAGISel.h"
22 #include "llvm/IR/Intrinsics.h"
23 #include "llvm/Support/CommandLine.h"
24 #include "llvm/Support/Compiler.h"
25 #include "llvm/Support/Debug.h"
26 using namespace llvm;
27
28 #define DEBUG_TYPE "hexagon-isel"
29
30 static
31 cl::opt<unsigned>
32 MaxNumOfUsesForConstExtenders("ga-max-num-uses-for-constant-extenders",
33 cl::Hidden, cl::init(2),
34 cl::desc("Maximum number of uses of a global address such that we still us a"
35 "constant extended instruction"));
36
37 //===----------------------------------------------------------------------===//
38 // Instruction Selector Implementation
39 //===----------------------------------------------------------------------===//
40
41 namespace llvm {
42 void initializeHexagonDAGToDAGISelPass(PassRegistry&);
43 }
44
45 //===--------------------------------------------------------------------===//
46 /// HexagonDAGToDAGISel - Hexagon specific code to select Hexagon machine
47 /// instructions for SelectionDAG operations.
48 ///
49 namespace {
50 class HexagonDAGToDAGISel : public SelectionDAGISel {
51 const HexagonTargetMachine& HTM;
52 const HexagonSubtarget *HST;
53 public:
HexagonDAGToDAGISel(HexagonTargetMachine & tm,CodeGenOpt::Level OptLevel)54 explicit HexagonDAGToDAGISel(HexagonTargetMachine &tm,
55 CodeGenOpt::Level OptLevel)
56 : SelectionDAGISel(tm, OptLevel), HTM(tm) {
57 initializeHexagonDAGToDAGISelPass(*PassRegistry::getPassRegistry());
58 }
59
runOnMachineFunction(MachineFunction & MF)60 bool runOnMachineFunction(MachineFunction &MF) override {
61 // Reset the subtarget each time through.
62 HST = &MF.getSubtarget<HexagonSubtarget>();
63 SelectionDAGISel::runOnMachineFunction(MF);
64 return true;
65 }
66
67 virtual void PreprocessISelDAG() override;
68 virtual void EmitFunctionEntryCode() override;
69
70 SDNode *Select(SDNode *N) override;
71
72 // Complex Pattern Selectors.
73 inline bool SelectAddrGA(SDValue &N, SDValue &R);
74 inline bool SelectAddrGP(SDValue &N, SDValue &R);
75 bool SelectGlobalAddress(SDValue &N, SDValue &R, bool UseGP);
76 bool SelectAddrFI(SDValue &N, SDValue &R);
77
getPassName() const78 const char *getPassName() const override {
79 return "Hexagon DAG->DAG Pattern Instruction Selection";
80 }
81
82 SDNode *SelectFrameIndex(SDNode *N);
83 /// SelectInlineAsmMemoryOperand - Implement addressing mode selection for
84 /// inline asm expressions.
85 bool SelectInlineAsmMemoryOperand(const SDValue &Op,
86 unsigned ConstraintID,
87 std::vector<SDValue> &OutOps) override;
88 SDNode *SelectLoad(SDNode *N);
89 SDNode *SelectBaseOffsetLoad(LoadSDNode *LD, SDLoc dl);
90 SDNode *SelectIndexedLoad(LoadSDNode *LD, SDLoc dl);
91 SDNode *SelectIndexedLoadZeroExtend64(LoadSDNode *LD, unsigned Opcode,
92 SDLoc dl);
93 SDNode *SelectIndexedLoadSignExtend64(LoadSDNode *LD, unsigned Opcode,
94 SDLoc dl);
95 SDNode *SelectBaseOffsetStore(StoreSDNode *ST, SDLoc dl);
96 SDNode *SelectIndexedStore(StoreSDNode *ST, SDLoc dl);
97 SDNode *SelectStore(SDNode *N);
98 SDNode *SelectSHL(SDNode *N);
99 SDNode *SelectMul(SDNode *N);
100 SDNode *SelectZeroExtend(SDNode *N);
101 SDNode *SelectIntrinsicWChain(SDNode *N);
102 SDNode *SelectIntrinsicWOChain(SDNode *N);
103 SDNode *SelectConstant(SDNode *N);
104 SDNode *SelectConstantFP(SDNode *N);
105 SDNode *SelectAdd(SDNode *N);
106 SDNode *SelectBitOp(SDNode *N);
107 bool isConstExtProfitable(SDNode *N) const;
108
109 // XformMskToBitPosU5Imm - Returns the bit position which
110 // the single bit 32 bit mask represents.
111 // Used in Clr and Set bit immediate memops.
XformMskToBitPosU5Imm(uint32_t Imm,SDLoc DL)112 SDValue XformMskToBitPosU5Imm(uint32_t Imm, SDLoc DL) {
113 int32_t bitPos;
114 bitPos = Log2_32(Imm);
115 assert(bitPos >= 0 && bitPos < 32 &&
116 "Constant out of range for 32 BitPos Memops");
117 return CurDAG->getTargetConstant(bitPos, DL, MVT::i32);
118 }
119
120 // XformMskToBitPosU4Imm - Returns the bit position which the single-bit
121 // 16 bit mask represents. Used in Clr and Set bit immediate memops.
XformMskToBitPosU4Imm(uint16_t Imm,SDLoc DL)122 SDValue XformMskToBitPosU4Imm(uint16_t Imm, SDLoc DL) {
123 return XformMskToBitPosU5Imm(Imm, DL);
124 }
125
126 // XformMskToBitPosU3Imm - Returns the bit position which the single-bit
127 // 8 bit mask represents. Used in Clr and Set bit immediate memops.
XformMskToBitPosU3Imm(uint8_t Imm,SDLoc DL)128 SDValue XformMskToBitPosU3Imm(uint8_t Imm, SDLoc DL) {
129 return XformMskToBitPosU5Imm(Imm, DL);
130 }
131
132 // Return true if there is exactly one bit set in V, i.e., if V is one of the
133 // following integers: 2^0, 2^1, ..., 2^31.
ImmIsSingleBit(uint32_t v) const134 bool ImmIsSingleBit(uint32_t v) const {
135 return isPowerOf2_32(v);
136 }
137
138 // XformM5ToU5Imm - Return a target constant with the specified value, of
139 // type i32 where the negative literal is transformed into a positive literal
140 // for use in -= memops.
XformM5ToU5Imm(signed Imm,SDLoc DL)141 inline SDValue XformM5ToU5Imm(signed Imm, SDLoc DL) {
142 assert( (Imm >= -31 && Imm <= -1) && "Constant out of range for Memops");
143 return CurDAG->getTargetConstant( - Imm, DL, MVT::i32);
144 }
145
146 // XformU7ToU7M1Imm - Return a target constant decremented by 1, in range
147 // [1..128], used in cmpb.gtu instructions.
XformU7ToU7M1Imm(signed Imm,SDLoc DL)148 inline SDValue XformU7ToU7M1Imm(signed Imm, SDLoc DL) {
149 assert((Imm >= 1 && Imm <= 128) && "Constant out of range for cmpb op");
150 return CurDAG->getTargetConstant(Imm - 1, DL, MVT::i8);
151 }
152
153 // XformS8ToS8M1Imm - Return a target constant decremented by 1.
XformSToSM1Imm(signed Imm,SDLoc DL)154 inline SDValue XformSToSM1Imm(signed Imm, SDLoc DL) {
155 return CurDAG->getTargetConstant(Imm - 1, DL, MVT::i32);
156 }
157
158 // XformU8ToU8M1Imm - Return a target constant decremented by 1.
XformUToUM1Imm(unsigned Imm,SDLoc DL)159 inline SDValue XformUToUM1Imm(unsigned Imm, SDLoc DL) {
160 assert((Imm >= 1) && "Cannot decrement unsigned int less than 1");
161 return CurDAG->getTargetConstant(Imm - 1, DL, MVT::i32);
162 }
163
164 // XformSToSM2Imm - Return a target constant decremented by 2.
XformSToSM2Imm(unsigned Imm,SDLoc DL)165 inline SDValue XformSToSM2Imm(unsigned Imm, SDLoc DL) {
166 return CurDAG->getTargetConstant(Imm - 2, DL, MVT::i32);
167 }
168
169 // XformSToSM3Imm - Return a target constant decremented by 3.
XformSToSM3Imm(unsigned Imm,SDLoc DL)170 inline SDValue XformSToSM3Imm(unsigned Imm, SDLoc DL) {
171 return CurDAG->getTargetConstant(Imm - 3, DL, MVT::i32);
172 }
173
174 // Include the pieces autogenerated from the target description.
175 #include "HexagonGenDAGISel.inc"
176
177 private:
178 bool isValueExtension(const SDValue &Val, unsigned FromBits, SDValue &Src);
179 }; // end HexagonDAGToDAGISel
180 } // end anonymous namespace
181
182
183 /// createHexagonISelDag - This pass converts a legalized DAG into a
184 /// Hexagon-specific DAG, ready for instruction scheduling.
185 ///
186 namespace llvm {
createHexagonISelDag(HexagonTargetMachine & TM,CodeGenOpt::Level OptLevel)187 FunctionPass *createHexagonISelDag(HexagonTargetMachine &TM,
188 CodeGenOpt::Level OptLevel) {
189 return new HexagonDAGToDAGISel(TM, OptLevel);
190 }
191 }
192
initializePassOnce(PassRegistry & Registry)193 static void initializePassOnce(PassRegistry &Registry) {
194 const char *Name = "Hexagon DAG->DAG Pattern Instruction Selection";
195 PassInfo *PI = new PassInfo(Name, "hexagon-isel",
196 &SelectionDAGISel::ID, nullptr, false, false);
197 Registry.registerPass(*PI, true);
198 }
199
initializeHexagonDAGToDAGISelPass(PassRegistry & Registry)200 void llvm::initializeHexagonDAGToDAGISelPass(PassRegistry &Registry) {
201 CALL_ONCE_INITIALIZATION(initializePassOnce)
202 }
203
204
205 // Intrinsics that return a a predicate.
doesIntrinsicReturnPredicate(unsigned ID)206 static unsigned doesIntrinsicReturnPredicate(unsigned ID)
207 {
208 switch (ID) {
209 default:
210 return 0;
211 case Intrinsic::hexagon_C2_cmpeq:
212 case Intrinsic::hexagon_C2_cmpgt:
213 case Intrinsic::hexagon_C2_cmpgtu:
214 case Intrinsic::hexagon_C2_cmpgtup:
215 case Intrinsic::hexagon_C2_cmpgtp:
216 case Intrinsic::hexagon_C2_cmpeqp:
217 case Intrinsic::hexagon_C2_bitsset:
218 case Intrinsic::hexagon_C2_bitsclr:
219 case Intrinsic::hexagon_C2_cmpeqi:
220 case Intrinsic::hexagon_C2_cmpgti:
221 case Intrinsic::hexagon_C2_cmpgtui:
222 case Intrinsic::hexagon_C2_cmpgei:
223 case Intrinsic::hexagon_C2_cmpgeui:
224 case Intrinsic::hexagon_C2_cmplt:
225 case Intrinsic::hexagon_C2_cmpltu:
226 case Intrinsic::hexagon_C2_bitsclri:
227 case Intrinsic::hexagon_C2_and:
228 case Intrinsic::hexagon_C2_or:
229 case Intrinsic::hexagon_C2_xor:
230 case Intrinsic::hexagon_C2_andn:
231 case Intrinsic::hexagon_C2_not:
232 case Intrinsic::hexagon_C2_orn:
233 case Intrinsic::hexagon_C2_pxfer_map:
234 case Intrinsic::hexagon_C2_any8:
235 case Intrinsic::hexagon_C2_all8:
236 case Intrinsic::hexagon_A2_vcmpbeq:
237 case Intrinsic::hexagon_A2_vcmpbgtu:
238 case Intrinsic::hexagon_A2_vcmpheq:
239 case Intrinsic::hexagon_A2_vcmphgt:
240 case Intrinsic::hexagon_A2_vcmphgtu:
241 case Intrinsic::hexagon_A2_vcmpweq:
242 case Intrinsic::hexagon_A2_vcmpwgt:
243 case Intrinsic::hexagon_A2_vcmpwgtu:
244 case Intrinsic::hexagon_C2_tfrrp:
245 case Intrinsic::hexagon_S2_tstbit_i:
246 case Intrinsic::hexagon_S2_tstbit_r:
247 return 1;
248 }
249 }
250
SelectIndexedLoadSignExtend64(LoadSDNode * LD,unsigned Opcode,SDLoc dl)251 SDNode *HexagonDAGToDAGISel::SelectIndexedLoadSignExtend64(LoadSDNode *LD,
252 unsigned Opcode,
253 SDLoc dl) {
254 SDValue Chain = LD->getChain();
255 EVT LoadedVT = LD->getMemoryVT();
256 SDValue Base = LD->getBasePtr();
257 SDValue Offset = LD->getOffset();
258 SDNode *OffsetNode = Offset.getNode();
259 int32_t Val = cast<ConstantSDNode>(OffsetNode)->getSExtValue();
260
261 const HexagonInstrInfo &TII = *HST->getInstrInfo();
262 if (TII.isValidAutoIncImm(LoadedVT, Val)) {
263 SDValue TargetConst = CurDAG->getTargetConstant(Val, dl, MVT::i32);
264 SDNode *Result_1 = CurDAG->getMachineNode(Opcode, dl, MVT::i32, MVT::i32,
265 MVT::Other, Base, TargetConst,
266 Chain);
267 SDNode *Result_2 = CurDAG->getMachineNode(Hexagon::A2_sxtw, dl, MVT::i64,
268 SDValue(Result_1, 0));
269 MachineSDNode::mmo_iterator MemOp = MF->allocateMemRefsArray(1);
270 MemOp[0] = LD->getMemOperand();
271 cast<MachineSDNode>(Result_1)->setMemRefs(MemOp, MemOp + 1);
272 const SDValue Froms[] = { SDValue(LD, 0),
273 SDValue(LD, 1),
274 SDValue(LD, 2) };
275 const SDValue Tos[] = { SDValue(Result_2, 0),
276 SDValue(Result_1, 1),
277 SDValue(Result_1, 2) };
278 ReplaceUses(Froms, Tos, 3);
279 return Result_2;
280 }
281
282 SDValue TargetConst0 = CurDAG->getTargetConstant(0, dl, MVT::i32);
283 SDValue TargetConstVal = CurDAG->getTargetConstant(Val, dl, MVT::i32);
284 SDNode *Result_1 = CurDAG->getMachineNode(Opcode, dl, MVT::i32, MVT::Other,
285 Base, TargetConst0, Chain);
286 SDNode *Result_2 = CurDAG->getMachineNode(Hexagon::A2_sxtw, dl, MVT::i64,
287 SDValue(Result_1, 0));
288 SDNode* Result_3 = CurDAG->getMachineNode(Hexagon::A2_addi, dl, MVT::i32,
289 Base, TargetConstVal,
290 SDValue(Result_1, 1));
291 MachineSDNode::mmo_iterator MemOp = MF->allocateMemRefsArray(1);
292 MemOp[0] = LD->getMemOperand();
293 cast<MachineSDNode>(Result_1)->setMemRefs(MemOp, MemOp + 1);
294 const SDValue Froms[] = { SDValue(LD, 0),
295 SDValue(LD, 1),
296 SDValue(LD, 2) };
297 const SDValue Tos[] = { SDValue(Result_2, 0),
298 SDValue(Result_3, 0),
299 SDValue(Result_1, 1) };
300 ReplaceUses(Froms, Tos, 3);
301 return Result_2;
302 }
303
304
SelectIndexedLoadZeroExtend64(LoadSDNode * LD,unsigned Opcode,SDLoc dl)305 SDNode *HexagonDAGToDAGISel::SelectIndexedLoadZeroExtend64(LoadSDNode *LD,
306 unsigned Opcode,
307 SDLoc dl) {
308 SDValue Chain = LD->getChain();
309 EVT LoadedVT = LD->getMemoryVT();
310 SDValue Base = LD->getBasePtr();
311 SDValue Offset = LD->getOffset();
312 SDNode *OffsetNode = Offset.getNode();
313 int32_t Val = cast<ConstantSDNode>(OffsetNode)->getSExtValue();
314
315 const HexagonInstrInfo &TII = *HST->getInstrInfo();
316 if (TII.isValidAutoIncImm(LoadedVT, Val)) {
317 SDValue TargetConstVal = CurDAG->getTargetConstant(Val, dl, MVT::i32);
318 SDValue TargetConst0 = CurDAG->getTargetConstant(0, dl, MVT::i32);
319 SDNode *Result_1 = CurDAG->getMachineNode(Opcode, dl, MVT::i32,
320 MVT::i32, MVT::Other, Base,
321 TargetConstVal, Chain);
322 SDNode *Result_2 = CurDAG->getMachineNode(Hexagon::A4_combineir, dl,
323 MVT::i64, MVT::Other,
324 TargetConst0,
325 SDValue(Result_1,0));
326 MachineSDNode::mmo_iterator MemOp = MF->allocateMemRefsArray(1);
327 MemOp[0] = LD->getMemOperand();
328 cast<MachineSDNode>(Result_1)->setMemRefs(MemOp, MemOp + 1);
329 const SDValue Froms[] = { SDValue(LD, 0),
330 SDValue(LD, 1),
331 SDValue(LD, 2) };
332 const SDValue Tos[] = { SDValue(Result_2, 0),
333 SDValue(Result_1, 1),
334 SDValue(Result_1, 2) };
335 ReplaceUses(Froms, Tos, 3);
336 return Result_2;
337 }
338
339 // Generate an indirect load.
340 SDValue TargetConst0 = CurDAG->getTargetConstant(0, dl, MVT::i32);
341 SDValue TargetConstVal = CurDAG->getTargetConstant(Val, dl, MVT::i32);
342 SDNode *Result_1 = CurDAG->getMachineNode(Opcode, dl, MVT::i32,
343 MVT::Other, Base, TargetConst0,
344 Chain);
345 SDNode *Result_2 = CurDAG->getMachineNode(Hexagon::A4_combineir, dl,
346 MVT::i64, MVT::Other,
347 TargetConst0,
348 SDValue(Result_1,0));
349 // Add offset to base.
350 SDNode* Result_3 = CurDAG->getMachineNode(Hexagon::A2_addi, dl, MVT::i32,
351 Base, TargetConstVal,
352 SDValue(Result_1, 1));
353 MachineSDNode::mmo_iterator MemOp = MF->allocateMemRefsArray(1);
354 MemOp[0] = LD->getMemOperand();
355 cast<MachineSDNode>(Result_1)->setMemRefs(MemOp, MemOp + 1);
356 const SDValue Froms[] = { SDValue(LD, 0),
357 SDValue(LD, 1),
358 SDValue(LD, 2) };
359 const SDValue Tos[] = { SDValue(Result_2, 0), // Load value.
360 SDValue(Result_3, 0), // New address.
361 SDValue(Result_1, 1) };
362 ReplaceUses(Froms, Tos, 3);
363 return Result_2;
364 }
365
366
SelectIndexedLoad(LoadSDNode * LD,SDLoc dl)367 SDNode *HexagonDAGToDAGISel::SelectIndexedLoad(LoadSDNode *LD, SDLoc dl) {
368 SDValue Chain = LD->getChain();
369 SDValue Base = LD->getBasePtr();
370 SDValue Offset = LD->getOffset();
371 SDNode *OffsetNode = Offset.getNode();
372 // Get the constant value.
373 int32_t Val = cast<ConstantSDNode>(OffsetNode)->getSExtValue();
374 EVT LoadedVT = LD->getMemoryVT();
375 unsigned Opcode = 0;
376
377 // Check for zero extended loads. Treat any-extend loads as zero extended
378 // loads.
379 ISD::LoadExtType ExtType = LD->getExtensionType();
380 bool IsZeroExt = (ExtType == ISD::ZEXTLOAD || ExtType == ISD::EXTLOAD);
381
382 // Figure out the opcode.
383 const HexagonInstrInfo &TII = *HST->getInstrInfo();
384 if (LoadedVT == MVT::i64) {
385 if (TII.isValidAutoIncImm(LoadedVT, Val))
386 Opcode = Hexagon::L2_loadrd_pi;
387 else
388 Opcode = Hexagon::L2_loadrd_io;
389 } else if (LoadedVT == MVT::i32) {
390 if (TII.isValidAutoIncImm(LoadedVT, Val))
391 Opcode = Hexagon::L2_loadri_pi;
392 else
393 Opcode = Hexagon::L2_loadri_io;
394 } else if (LoadedVT == MVT::i16) {
395 if (TII.isValidAutoIncImm(LoadedVT, Val))
396 Opcode = IsZeroExt ? Hexagon::L2_loadruh_pi : Hexagon::L2_loadrh_pi;
397 else
398 Opcode = IsZeroExt ? Hexagon::L2_loadruh_io : Hexagon::L2_loadrh_io;
399 } else if (LoadedVT == MVT::i8) {
400 if (TII.isValidAutoIncImm(LoadedVT, Val))
401 Opcode = IsZeroExt ? Hexagon::L2_loadrub_pi : Hexagon::L2_loadrb_pi;
402 else
403 Opcode = IsZeroExt ? Hexagon::L2_loadrub_io : Hexagon::L2_loadrb_io;
404 } else
405 llvm_unreachable("unknown memory type");
406
407 // For zero extended i64 loads, we need to add combine instructions.
408 if (LD->getValueType(0) == MVT::i64 && IsZeroExt)
409 return SelectIndexedLoadZeroExtend64(LD, Opcode, dl);
410 // Handle sign extended i64 loads.
411 if (LD->getValueType(0) == MVT::i64 && ExtType == ISD::SEXTLOAD)
412 return SelectIndexedLoadSignExtend64(LD, Opcode, dl);
413
414 if (TII.isValidAutoIncImm(LoadedVT, Val)) {
415 SDValue TargetConstVal = CurDAG->getTargetConstant(Val, dl, MVT::i32);
416 SDNode* Result = CurDAG->getMachineNode(Opcode, dl,
417 LD->getValueType(0),
418 MVT::i32, MVT::Other, Base,
419 TargetConstVal, Chain);
420 MachineSDNode::mmo_iterator MemOp = MF->allocateMemRefsArray(1);
421 MemOp[0] = LD->getMemOperand();
422 cast<MachineSDNode>(Result)->setMemRefs(MemOp, MemOp + 1);
423 const SDValue Froms[] = { SDValue(LD, 0),
424 SDValue(LD, 1),
425 SDValue(LD, 2)
426 };
427 const SDValue Tos[] = { SDValue(Result, 0),
428 SDValue(Result, 1),
429 SDValue(Result, 2)
430 };
431 ReplaceUses(Froms, Tos, 3);
432 return Result;
433 } else {
434 SDValue TargetConst0 = CurDAG->getTargetConstant(0, dl, MVT::i32);
435 SDValue TargetConstVal = CurDAG->getTargetConstant(Val, dl, MVT::i32);
436 SDNode* Result_1 = CurDAG->getMachineNode(Opcode, dl,
437 LD->getValueType(0),
438 MVT::Other, Base, TargetConst0,
439 Chain);
440 SDNode* Result_2 = CurDAG->getMachineNode(Hexagon::A2_addi, dl, MVT::i32,
441 Base, TargetConstVal,
442 SDValue(Result_1, 1));
443 MachineSDNode::mmo_iterator MemOp = MF->allocateMemRefsArray(1);
444 MemOp[0] = LD->getMemOperand();
445 cast<MachineSDNode>(Result_1)->setMemRefs(MemOp, MemOp + 1);
446 const SDValue Froms[] = { SDValue(LD, 0),
447 SDValue(LD, 1),
448 SDValue(LD, 2)
449 };
450 const SDValue Tos[] = { SDValue(Result_1, 0),
451 SDValue(Result_2, 0),
452 SDValue(Result_1, 1)
453 };
454 ReplaceUses(Froms, Tos, 3);
455 return Result_1;
456 }
457 }
458
459
SelectLoad(SDNode * N)460 SDNode *HexagonDAGToDAGISel::SelectLoad(SDNode *N) {
461 SDNode *result;
462 SDLoc dl(N);
463 LoadSDNode *LD = cast<LoadSDNode>(N);
464 ISD::MemIndexedMode AM = LD->getAddressingMode();
465
466 // Handle indexed loads.
467 if (AM != ISD::UNINDEXED) {
468 result = SelectIndexedLoad(LD, dl);
469 } else {
470 result = SelectCode(LD);
471 }
472
473 return result;
474 }
475
476
SelectIndexedStore(StoreSDNode * ST,SDLoc dl)477 SDNode *HexagonDAGToDAGISel::SelectIndexedStore(StoreSDNode *ST, SDLoc dl) {
478 SDValue Chain = ST->getChain();
479 SDValue Base = ST->getBasePtr();
480 SDValue Offset = ST->getOffset();
481 SDValue Value = ST->getValue();
482 SDNode *OffsetNode = Offset.getNode();
483 // Get the constant value.
484 int32_t Val = cast<ConstantSDNode>(OffsetNode)->getSExtValue();
485 EVT StoredVT = ST->getMemoryVT();
486 EVT ValueVT = Value.getValueType();
487
488 // Offset value must be within representable range
489 // and must have correct alignment properties.
490 const HexagonInstrInfo &TII = *HST->getInstrInfo();
491 if (TII.isValidAutoIncImm(StoredVT, Val)) {
492 unsigned Opcode = 0;
493
494 // Figure out the post inc version of opcode.
495 if (StoredVT == MVT::i64) Opcode = Hexagon::S2_storerd_pi;
496 else if (StoredVT == MVT::i32) Opcode = Hexagon::S2_storeri_pi;
497 else if (StoredVT == MVT::i16) Opcode = Hexagon::S2_storerh_pi;
498 else if (StoredVT == MVT::i8) Opcode = Hexagon::S2_storerb_pi;
499 else llvm_unreachable("unknown memory type");
500
501 if (ST->isTruncatingStore() && ValueVT.getSizeInBits() == 64) {
502 assert(StoredVT.getSizeInBits() < 64 && "Not a truncating store");
503 Value = CurDAG->getTargetExtractSubreg(Hexagon::subreg_loreg,
504 dl, MVT::i32, Value);
505 }
506 SDValue Ops[] = {Base, CurDAG->getTargetConstant(Val, dl, MVT::i32), Value,
507 Chain};
508 // Build post increment store.
509 SDNode* Result = CurDAG->getMachineNode(Opcode, dl, MVT::i32,
510 MVT::Other, Ops);
511 MachineSDNode::mmo_iterator MemOp = MF->allocateMemRefsArray(1);
512 MemOp[0] = ST->getMemOperand();
513 cast<MachineSDNode>(Result)->setMemRefs(MemOp, MemOp + 1);
514
515 ReplaceUses(ST, Result);
516 ReplaceUses(SDValue(ST,1), SDValue(Result,1));
517 return Result;
518 }
519
520 // Note: Order of operands matches the def of instruction:
521 // def S2_storerd_io
522 // : STInst<(outs), (ins IntRegs:$base, imm:$offset, DoubleRegs:$src1), ...
523 // and it differs for POST_ST* for instance.
524 SDValue Ops[] = { Base, CurDAG->getTargetConstant(0, dl, MVT::i32), Value,
525 Chain};
526 unsigned Opcode = 0;
527
528 // Figure out the opcode.
529 if (StoredVT == MVT::i64) Opcode = Hexagon::S2_storerd_io;
530 else if (StoredVT == MVT::i32) Opcode = Hexagon::S2_storeri_io;
531 else if (StoredVT == MVT::i16) Opcode = Hexagon::S2_storerh_io;
532 else if (StoredVT == MVT::i8) Opcode = Hexagon::S2_storerb_io;
533 else llvm_unreachable("unknown memory type");
534
535 // Build regular store.
536 SDValue TargetConstVal = CurDAG->getTargetConstant(Val, dl, MVT::i32);
537 SDNode* Result_1 = CurDAG->getMachineNode(Opcode, dl, MVT::Other, Ops);
538 // Build splitted incriment instruction.
539 SDNode* Result_2 = CurDAG->getMachineNode(Hexagon::A2_addi, dl, MVT::i32,
540 Base,
541 TargetConstVal,
542 SDValue(Result_1, 0));
543 MachineSDNode::mmo_iterator MemOp = MF->allocateMemRefsArray(1);
544 MemOp[0] = ST->getMemOperand();
545 cast<MachineSDNode>(Result_1)->setMemRefs(MemOp, MemOp + 1);
546
547 ReplaceUses(SDValue(ST,0), SDValue(Result_2,0));
548 ReplaceUses(SDValue(ST,1), SDValue(Result_1,0));
549 return Result_2;
550 }
551
SelectStore(SDNode * N)552 SDNode *HexagonDAGToDAGISel::SelectStore(SDNode *N) {
553 SDLoc dl(N);
554 StoreSDNode *ST = cast<StoreSDNode>(N);
555 ISD::MemIndexedMode AM = ST->getAddressingMode();
556
557 // Handle indexed stores.
558 if (AM != ISD::UNINDEXED) {
559 return SelectIndexedStore(ST, dl);
560 }
561
562 return SelectCode(ST);
563 }
564
SelectMul(SDNode * N)565 SDNode *HexagonDAGToDAGISel::SelectMul(SDNode *N) {
566 SDLoc dl(N);
567
568 //
569 // %conv.i = sext i32 %tmp1 to i64
570 // %conv2.i = sext i32 %add to i64
571 // %mul.i = mul nsw i64 %conv2.i, %conv.i
572 //
573 // --- match with the following ---
574 //
575 // %mul.i = mpy (%tmp1, %add)
576 //
577
578 if (N->getValueType(0) == MVT::i64) {
579 // Shifting a i64 signed multiply.
580 SDValue MulOp0 = N->getOperand(0);
581 SDValue MulOp1 = N->getOperand(1);
582
583 SDValue OP0;
584 SDValue OP1;
585
586 // Handle sign_extend and sextload.
587 if (MulOp0.getOpcode() == ISD::SIGN_EXTEND) {
588 SDValue Sext0 = MulOp0.getOperand(0);
589 if (Sext0.getNode()->getValueType(0) != MVT::i32) {
590 return SelectCode(N);
591 }
592
593 OP0 = Sext0;
594 } else if (MulOp0.getOpcode() == ISD::LOAD) {
595 LoadSDNode *LD = cast<LoadSDNode>(MulOp0.getNode());
596 if (LD->getMemoryVT() != MVT::i32 ||
597 LD->getExtensionType() != ISD::SEXTLOAD ||
598 LD->getAddressingMode() != ISD::UNINDEXED) {
599 return SelectCode(N);
600 }
601
602 SDValue Chain = LD->getChain();
603 SDValue TargetConst0 = CurDAG->getTargetConstant(0, dl, MVT::i32);
604 OP0 = SDValue(CurDAG->getMachineNode(Hexagon::L2_loadri_io, dl, MVT::i32,
605 MVT::Other,
606 LD->getBasePtr(), TargetConst0,
607 Chain), 0);
608 } else {
609 return SelectCode(N);
610 }
611
612 // Same goes for the second operand.
613 if (MulOp1.getOpcode() == ISD::SIGN_EXTEND) {
614 SDValue Sext1 = MulOp1.getOperand(0);
615 if (Sext1.getNode()->getValueType(0) != MVT::i32) {
616 return SelectCode(N);
617 }
618
619 OP1 = Sext1;
620 } else if (MulOp1.getOpcode() == ISD::LOAD) {
621 LoadSDNode *LD = cast<LoadSDNode>(MulOp1.getNode());
622 if (LD->getMemoryVT() != MVT::i32 ||
623 LD->getExtensionType() != ISD::SEXTLOAD ||
624 LD->getAddressingMode() != ISD::UNINDEXED) {
625 return SelectCode(N);
626 }
627
628 SDValue Chain = LD->getChain();
629 SDValue TargetConst0 = CurDAG->getTargetConstant(0, dl, MVT::i32);
630 OP1 = SDValue(CurDAG->getMachineNode(Hexagon::L2_loadri_io, dl, MVT::i32,
631 MVT::Other,
632 LD->getBasePtr(), TargetConst0,
633 Chain), 0);
634 } else {
635 return SelectCode(N);
636 }
637
638 // Generate a mpy instruction.
639 SDNode *Result = CurDAG->getMachineNode(Hexagon::M2_dpmpyss_s0, dl, MVT::i64,
640 OP0, OP1);
641 ReplaceUses(N, Result);
642 return Result;
643 }
644
645 return SelectCode(N);
646 }
647
SelectSHL(SDNode * N)648 SDNode *HexagonDAGToDAGISel::SelectSHL(SDNode *N) {
649 SDLoc dl(N);
650 if (N->getValueType(0) == MVT::i32) {
651 SDValue Shl_0 = N->getOperand(0);
652 SDValue Shl_1 = N->getOperand(1);
653 // RHS is const.
654 if (Shl_1.getOpcode() == ISD::Constant) {
655 if (Shl_0.getOpcode() == ISD::MUL) {
656 SDValue Mul_0 = Shl_0.getOperand(0); // Val
657 SDValue Mul_1 = Shl_0.getOperand(1); // Const
658 // RHS of mul is const.
659 if (Mul_1.getOpcode() == ISD::Constant) {
660 int32_t ShlConst =
661 cast<ConstantSDNode>(Shl_1.getNode())->getSExtValue();
662 int32_t MulConst =
663 cast<ConstantSDNode>(Mul_1.getNode())->getSExtValue();
664 int32_t ValConst = MulConst << ShlConst;
665 SDValue Val = CurDAG->getTargetConstant(ValConst, dl,
666 MVT::i32);
667 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(Val.getNode()))
668 if (isInt<9>(CN->getSExtValue())) {
669 SDNode* Result =
670 CurDAG->getMachineNode(Hexagon::M2_mpysmi, dl,
671 MVT::i32, Mul_0, Val);
672 ReplaceUses(N, Result);
673 return Result;
674 }
675
676 }
677 } else if (Shl_0.getOpcode() == ISD::SUB) {
678 SDValue Sub_0 = Shl_0.getOperand(0); // Const 0
679 SDValue Sub_1 = Shl_0.getOperand(1); // Val
680 if (Sub_0.getOpcode() == ISD::Constant) {
681 int32_t SubConst =
682 cast<ConstantSDNode>(Sub_0.getNode())->getSExtValue();
683 if (SubConst == 0) {
684 if (Sub_1.getOpcode() == ISD::SHL) {
685 SDValue Shl2_0 = Sub_1.getOperand(0); // Val
686 SDValue Shl2_1 = Sub_1.getOperand(1); // Const
687 if (Shl2_1.getOpcode() == ISD::Constant) {
688 int32_t ShlConst =
689 cast<ConstantSDNode>(Shl_1.getNode())->getSExtValue();
690 int32_t Shl2Const =
691 cast<ConstantSDNode>(Shl2_1.getNode())->getSExtValue();
692 int32_t ValConst = 1 << (ShlConst+Shl2Const);
693 SDValue Val = CurDAG->getTargetConstant(-ValConst, dl,
694 MVT::i32);
695 if (ConstantSDNode *CN =
696 dyn_cast<ConstantSDNode>(Val.getNode()))
697 if (isInt<9>(CN->getSExtValue())) {
698 SDNode* Result =
699 CurDAG->getMachineNode(Hexagon::M2_mpysmi, dl, MVT::i32,
700 Shl2_0, Val);
701 ReplaceUses(N, Result);
702 return Result;
703 }
704 }
705 }
706 }
707 }
708 }
709 }
710 }
711 return SelectCode(N);
712 }
713
714
715 //
716 // If there is an zero_extend followed an intrinsic in DAG (this means - the
717 // result of the intrinsic is predicate); convert the zero_extend to
718 // transfer instruction.
719 //
720 // Zero extend -> transfer is lowered here. Otherwise, zero_extend will be
721 // converted into a MUX as predicate registers defined as 1 bit in the
722 // compiler. Architecture defines them as 8-bit registers.
723 // We want to preserve all the lower 8-bits and, not just 1 LSB bit.
724 //
SelectZeroExtend(SDNode * N)725 SDNode *HexagonDAGToDAGISel::SelectZeroExtend(SDNode *N) {
726 SDLoc dl(N);
727
728 SDValue Op0 = N->getOperand(0);
729 EVT OpVT = Op0.getValueType();
730 unsigned OpBW = OpVT.getSizeInBits();
731
732 // Special handling for zero-extending a vector of booleans.
733 if (OpVT.isVector() && OpVT.getVectorElementType() == MVT::i1 && OpBW <= 64) {
734 SDNode *Mask = CurDAG->getMachineNode(Hexagon::C2_mask, dl, MVT::i64, Op0);
735 unsigned NE = OpVT.getVectorNumElements();
736 EVT ExVT = N->getValueType(0);
737 unsigned ES = ExVT.getVectorElementType().getSizeInBits();
738 uint64_t MV = 0, Bit = 1;
739 for (unsigned i = 0; i < NE; ++i) {
740 MV |= Bit;
741 Bit <<= ES;
742 }
743 SDValue Ones = CurDAG->getTargetConstant(MV, dl, MVT::i64);
744 SDNode *OnesReg = CurDAG->getMachineNode(Hexagon::CONST64_Int_Real, dl,
745 MVT::i64, Ones);
746 if (ExVT.getSizeInBits() == 32) {
747 SDNode *And = CurDAG->getMachineNode(Hexagon::A2_andp, dl, MVT::i64,
748 SDValue(Mask,0), SDValue(OnesReg,0));
749 SDValue SubR = CurDAG->getTargetConstant(Hexagon::subreg_loreg, dl,
750 MVT::i32);
751 return CurDAG->getMachineNode(Hexagon::EXTRACT_SUBREG, dl, ExVT,
752 SDValue(And,0), SubR);
753 }
754 return CurDAG->getMachineNode(Hexagon::A2_andp, dl, ExVT,
755 SDValue(Mask,0), SDValue(OnesReg,0));
756 }
757
758 SDNode *IsIntrinsic = N->getOperand(0).getNode();
759 if ((IsIntrinsic->getOpcode() == ISD::INTRINSIC_WO_CHAIN)) {
760 unsigned ID =
761 cast<ConstantSDNode>(IsIntrinsic->getOperand(0))->getZExtValue();
762 if (doesIntrinsicReturnPredicate(ID)) {
763 // Now we need to differentiate target data types.
764 if (N->getValueType(0) == MVT::i64) {
765 // Convert the zero_extend to Rs = Pd followed by A2_combinew(0,Rs).
766 SDValue TargetConst0 = CurDAG->getTargetConstant(0, dl, MVT::i32);
767 SDNode *Result_1 = CurDAG->getMachineNode(Hexagon::C2_tfrpr, dl,
768 MVT::i32,
769 SDValue(IsIntrinsic, 0));
770 SDNode *Result_2 = CurDAG->getMachineNode(Hexagon::A2_tfrsi, dl,
771 MVT::i32,
772 TargetConst0);
773 SDNode *Result_3 = CurDAG->getMachineNode(Hexagon::A2_combinew, dl,
774 MVT::i64, MVT::Other,
775 SDValue(Result_2, 0),
776 SDValue(Result_1, 0));
777 ReplaceUses(N, Result_3);
778 return Result_3;
779 }
780 if (N->getValueType(0) == MVT::i32) {
781 // Convert the zero_extend to Rs = Pd
782 SDNode* RsPd = CurDAG->getMachineNode(Hexagon::C2_tfrpr, dl,
783 MVT::i32,
784 SDValue(IsIntrinsic, 0));
785 ReplaceUses(N, RsPd);
786 return RsPd;
787 }
788 llvm_unreachable("Unexpected value type");
789 }
790 }
791 return SelectCode(N);
792 }
793
794 //
795 // Checking for intrinsics circular load/store, and bitreverse load/store
796 // instrisics in order to select the correct lowered operation.
797 //
SelectIntrinsicWChain(SDNode * N)798 SDNode *HexagonDAGToDAGISel::SelectIntrinsicWChain(SDNode *N) {
799 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
800 if (IntNo == Intrinsic::hexagon_circ_ldd ||
801 IntNo == Intrinsic::hexagon_circ_ldw ||
802 IntNo == Intrinsic::hexagon_circ_lduh ||
803 IntNo == Intrinsic::hexagon_circ_ldh ||
804 IntNo == Intrinsic::hexagon_circ_ldub ||
805 IntNo == Intrinsic::hexagon_circ_ldb) {
806 SDLoc dl(N);
807 SDValue Chain = N->getOperand(0);
808 SDValue Base = N->getOperand(2);
809 SDValue Load = N->getOperand(3);
810 SDValue ModifierExpr = N->getOperand(4);
811 SDValue Offset = N->getOperand(5);
812
813 // We need to add the rerurn type for the load. This intrinsic has
814 // two return types, one for the load and one for the post-increment.
815 // Only the *_ld instructions push the extra return type, and bump the
816 // result node operand number correspondingly.
817 std::vector<EVT> ResTys;
818 unsigned opc;
819 unsigned memsize, align;
820 MVT MvtSize = MVT::i32;
821
822 if (IntNo == Intrinsic::hexagon_circ_ldd) {
823 ResTys.push_back(MVT::i32);
824 ResTys.push_back(MVT::i64);
825 opc = Hexagon::L2_loadrd_pci_pseudo;
826 memsize = 8;
827 align = 8;
828 } else if (IntNo == Intrinsic::hexagon_circ_ldw) {
829 ResTys.push_back(MVT::i32);
830 ResTys.push_back(MVT::i32);
831 opc = Hexagon::L2_loadri_pci_pseudo;
832 memsize = 4;
833 align = 4;
834 } else if (IntNo == Intrinsic::hexagon_circ_ldh) {
835 ResTys.push_back(MVT::i32);
836 ResTys.push_back(MVT::i32);
837 opc = Hexagon::L2_loadrh_pci_pseudo;
838 memsize = 2;
839 align = 2;
840 MvtSize = MVT::i16;
841 } else if (IntNo == Intrinsic::hexagon_circ_lduh) {
842 ResTys.push_back(MVT::i32);
843 ResTys.push_back(MVT::i32);
844 opc = Hexagon::L2_loadruh_pci_pseudo;
845 memsize = 2;
846 align = 2;
847 MvtSize = MVT::i16;
848 } else if (IntNo == Intrinsic::hexagon_circ_ldb) {
849 ResTys.push_back(MVT::i32);
850 ResTys.push_back(MVT::i32);
851 opc = Hexagon::L2_loadrb_pci_pseudo;
852 memsize = 1;
853 align = 1;
854 MvtSize = MVT::i8;
855 } else if (IntNo == Intrinsic::hexagon_circ_ldub) {
856 ResTys.push_back(MVT::i32);
857 ResTys.push_back(MVT::i32);
858 opc = Hexagon::L2_loadrub_pci_pseudo;
859 memsize = 1;
860 align = 1;
861 MvtSize = MVT::i8;
862 } else
863 llvm_unreachable("no opc");
864
865 ResTys.push_back(MVT::Other);
866
867 // Copy over the arguments, which are the same mostly.
868 SmallVector<SDValue, 5> Ops;
869 Ops.push_back(Base);
870 Ops.push_back(Load);
871 Ops.push_back(ModifierExpr);
872 int32_t Val = cast<ConstantSDNode>(Offset.getNode())->getSExtValue();
873 Ops.push_back(CurDAG->getTargetConstant(Val, dl, MVT::i32));
874 Ops.push_back(Chain);
875 SDNode* Result = CurDAG->getMachineNode(opc, dl, ResTys, Ops);
876
877 SDValue ST;
878 MachineMemOperand *Mem =
879 MF->getMachineMemOperand(MachinePointerInfo(),
880 MachineMemOperand::MOStore, memsize, align);
881 if (MvtSize != MVT::i32)
882 ST = CurDAG->getTruncStore(Chain, dl, SDValue(Result, 1), Load,
883 MvtSize, Mem);
884 else
885 ST = CurDAG->getStore(Chain, dl, SDValue(Result, 1), Load, Mem);
886
887 SDNode* Store = SelectStore(ST.getNode());
888
889 const SDValue Froms[] = { SDValue(N, 0),
890 SDValue(N, 1) };
891 const SDValue Tos[] = { SDValue(Result, 0),
892 SDValue(Store, 0) };
893 ReplaceUses(Froms, Tos, 2);
894 return Result;
895 }
896
897 if (IntNo == Intrinsic::hexagon_brev_ldd ||
898 IntNo == Intrinsic::hexagon_brev_ldw ||
899 IntNo == Intrinsic::hexagon_brev_ldh ||
900 IntNo == Intrinsic::hexagon_brev_lduh ||
901 IntNo == Intrinsic::hexagon_brev_ldb ||
902 IntNo == Intrinsic::hexagon_brev_ldub) {
903 SDLoc dl(N);
904 SDValue Chain = N->getOperand(0);
905 SDValue Base = N->getOperand(2);
906 SDValue Load = N->getOperand(3);
907 SDValue ModifierExpr = N->getOperand(4);
908
909 // We need to add the rerurn type for the load. This intrinsic has
910 // two return types, one for the load and one for the post-increment.
911 std::vector<EVT> ResTys;
912 unsigned opc;
913 unsigned memsize, align;
914 MVT MvtSize = MVT::i32;
915
916 if (IntNo == Intrinsic::hexagon_brev_ldd) {
917 ResTys.push_back(MVT::i32);
918 ResTys.push_back(MVT::i64);
919 opc = Hexagon::L2_loadrd_pbr_pseudo;
920 memsize = 8;
921 align = 8;
922 } else if (IntNo == Intrinsic::hexagon_brev_ldw) {
923 ResTys.push_back(MVT::i32);
924 ResTys.push_back(MVT::i32);
925 opc = Hexagon::L2_loadri_pbr_pseudo;
926 memsize = 4;
927 align = 4;
928 } else if (IntNo == Intrinsic::hexagon_brev_ldh) {
929 ResTys.push_back(MVT::i32);
930 ResTys.push_back(MVT::i32);
931 opc = Hexagon::L2_loadrh_pbr_pseudo;
932 memsize = 2;
933 align = 2;
934 MvtSize = MVT::i16;
935 } else if (IntNo == Intrinsic::hexagon_brev_lduh) {
936 ResTys.push_back(MVT::i32);
937 ResTys.push_back(MVT::i32);
938 opc = Hexagon::L2_loadruh_pbr_pseudo;
939 memsize = 2;
940 align = 2;
941 MvtSize = MVT::i16;
942 } else if (IntNo == Intrinsic::hexagon_brev_ldb) {
943 ResTys.push_back(MVT::i32);
944 ResTys.push_back(MVT::i32);
945 opc = Hexagon::L2_loadrb_pbr_pseudo;
946 memsize = 1;
947 align = 1;
948 MvtSize = MVT::i8;
949 } else if (IntNo == Intrinsic::hexagon_brev_ldub) {
950 ResTys.push_back(MVT::i32);
951 ResTys.push_back(MVT::i32);
952 opc = Hexagon::L2_loadrub_pbr_pseudo;
953 memsize = 1;
954 align = 1;
955 MvtSize = MVT::i8;
956 } else
957 llvm_unreachable("no opc");
958
959 ResTys.push_back(MVT::Other);
960
961 // Copy over the arguments, which are the same mostly.
962 SmallVector<SDValue, 4> Ops;
963 Ops.push_back(Base);
964 Ops.push_back(Load);
965 Ops.push_back(ModifierExpr);
966 Ops.push_back(Chain);
967 SDNode* Result = CurDAG->getMachineNode(opc, dl, ResTys, Ops);
968 SDValue ST;
969 MachineMemOperand *Mem =
970 MF->getMachineMemOperand(MachinePointerInfo(),
971 MachineMemOperand::MOStore, memsize, align);
972 if (MvtSize != MVT::i32)
973 ST = CurDAG->getTruncStore(Chain, dl, SDValue(Result, 1), Load,
974 MvtSize, Mem);
975 else
976 ST = CurDAG->getStore(Chain, dl, SDValue(Result, 1), Load, Mem);
977
978 SDNode* Store = SelectStore(ST.getNode());
979
980 const SDValue Froms[] = { SDValue(N, 0),
981 SDValue(N, 1) };
982 const SDValue Tos[] = { SDValue(Result, 0),
983 SDValue(Store, 0) };
984 ReplaceUses(Froms, Tos, 2);
985 return Result;
986 }
987
988 return SelectCode(N);
989 }
990
991 //
992 // Checking for intrinsics which have predicate registers as operand(s)
993 // and lowering to the actual intrinsic.
994 //
SelectIntrinsicWOChain(SDNode * N)995 SDNode *HexagonDAGToDAGISel::SelectIntrinsicWOChain(SDNode *N) {
996 unsigned IID = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
997 unsigned Bits;
998 switch (IID) {
999 case Intrinsic::hexagon_S2_vsplatrb:
1000 Bits = 8;
1001 break;
1002 case Intrinsic::hexagon_S2_vsplatrh:
1003 Bits = 16;
1004 break;
1005 default:
1006 return SelectCode(N);
1007 }
1008
1009 SDValue const &V = N->getOperand(1);
1010 SDValue U;
1011 if (isValueExtension(V, Bits, U)) {
1012 SDValue R = CurDAG->getNode(N->getOpcode(), SDLoc(N), N->getValueType(0),
1013 N->getOperand(0), U);
1014 return SelectCode(R.getNode());
1015 }
1016 return SelectCode(N);
1017 }
1018
1019 //
1020 // Map floating point constant values.
1021 //
SelectConstantFP(SDNode * N)1022 SDNode *HexagonDAGToDAGISel::SelectConstantFP(SDNode *N) {
1023 SDLoc dl(N);
1024 ConstantFPSDNode *CN = dyn_cast<ConstantFPSDNode>(N);
1025 APFloat APF = CN->getValueAPF();
1026 if (N->getValueType(0) == MVT::f32) {
1027 return CurDAG->getMachineNode(Hexagon::TFRI_f, dl, MVT::f32,
1028 CurDAG->getTargetConstantFP(APF.convertToFloat(), dl, MVT::f32));
1029 }
1030 else if (N->getValueType(0) == MVT::f64) {
1031 return CurDAG->getMachineNode(Hexagon::CONST64_Float_Real, dl, MVT::f64,
1032 CurDAG->getTargetConstantFP(APF.convertToDouble(), dl, MVT::f64));
1033 }
1034
1035 return SelectCode(N);
1036 }
1037
1038 //
1039 // Map predicate true (encoded as -1 in LLVM) to a XOR.
1040 //
SelectConstant(SDNode * N)1041 SDNode *HexagonDAGToDAGISel::SelectConstant(SDNode *N) {
1042 SDLoc dl(N);
1043 if (N->getValueType(0) == MVT::i1) {
1044 SDNode* Result = 0;
1045 int32_t Val = cast<ConstantSDNode>(N)->getSExtValue();
1046 if (Val == -1) {
1047 Result = CurDAG->getMachineNode(Hexagon::TFR_PdTrue, dl, MVT::i1);
1048 } else if (Val == 0) {
1049 Result = CurDAG->getMachineNode(Hexagon::TFR_PdFalse, dl, MVT::i1);
1050 }
1051 if (Result) {
1052 ReplaceUses(N, Result);
1053 return Result;
1054 }
1055 }
1056
1057 return SelectCode(N);
1058 }
1059
1060
1061 //
1062 // Map add followed by a asr -> asr +=.
1063 //
SelectAdd(SDNode * N)1064 SDNode *HexagonDAGToDAGISel::SelectAdd(SDNode *N) {
1065 SDLoc dl(N);
1066 if (N->getValueType(0) != MVT::i32) {
1067 return SelectCode(N);
1068 }
1069 // Identify nodes of the form: add(asr(...)).
1070 SDNode* Src1 = N->getOperand(0).getNode();
1071 if (Src1->getOpcode() != ISD::SRA || !Src1->hasOneUse()
1072 || Src1->getValueType(0) != MVT::i32) {
1073 return SelectCode(N);
1074 }
1075
1076 // Build Rd = Rd' + asr(Rs, Rt). The machine constraints will ensure that
1077 // Rd and Rd' are assigned to the same register
1078 SDNode* Result = CurDAG->getMachineNode(Hexagon::S2_asr_r_r_acc, dl, MVT::i32,
1079 N->getOperand(1),
1080 Src1->getOperand(0),
1081 Src1->getOperand(1));
1082 ReplaceUses(N, Result);
1083
1084 return Result;
1085 }
1086
1087 //
1088 // Map the following, where possible.
1089 // AND/FABS -> clrbit
1090 // OR -> setbit
1091 // XOR/FNEG ->toggle_bit.
1092 //
SelectBitOp(SDNode * N)1093 SDNode *HexagonDAGToDAGISel::SelectBitOp(SDNode *N) {
1094 SDLoc dl(N);
1095 EVT ValueVT = N->getValueType(0);
1096
1097 // We handle only 32 and 64-bit bit ops.
1098 if (!(ValueVT == MVT::i32 || ValueVT == MVT::i64 ||
1099 ValueVT == MVT::f32 || ValueVT == MVT::f64))
1100 return SelectCode(N);
1101
1102 // We handly only fabs and fneg for V5.
1103 unsigned Opc = N->getOpcode();
1104 if ((Opc == ISD::FABS || Opc == ISD::FNEG) && !HST->hasV5TOps())
1105 return SelectCode(N);
1106
1107 int64_t Val = 0;
1108 if (Opc != ISD::FABS && Opc != ISD::FNEG) {
1109 if (N->getOperand(1).getOpcode() == ISD::Constant)
1110 Val = cast<ConstantSDNode>((N)->getOperand(1))->getSExtValue();
1111 else
1112 return SelectCode(N);
1113 }
1114
1115 if (Opc == ISD::AND) {
1116 if (((ValueVT == MVT::i32) &&
1117 (!((Val & 0x80000000) || (Val & 0x7fffffff)))) ||
1118 ((ValueVT == MVT::i64) &&
1119 (!((Val & 0x8000000000000000) || (Val & 0x7fffffff)))))
1120 // If it's simple AND, do the normal op.
1121 return SelectCode(N);
1122 else
1123 Val = ~Val;
1124 }
1125
1126 // If OR or AND is being fed by shl, srl and, sra don't do this change,
1127 // because Hexagon provide |= &= on shl, srl, and sra.
1128 // Traverse the DAG to see if there is shl, srl and sra.
1129 if (Opc == ISD::OR || Opc == ISD::AND) {
1130 switch (N->getOperand(0)->getOpcode()) {
1131 default: break;
1132 case ISD::SRA:
1133 case ISD::SRL:
1134 case ISD::SHL:
1135 return SelectCode(N);
1136 }
1137 }
1138
1139 // Make sure it's power of 2.
1140 unsigned bitpos = 0;
1141 if (Opc != ISD::FABS && Opc != ISD::FNEG) {
1142 if (((ValueVT == MVT::i32) && !isPowerOf2_32(Val)) ||
1143 ((ValueVT == MVT::i64) && !isPowerOf2_64(Val)))
1144 return SelectCode(N);
1145
1146 // Get the bit position.
1147 bitpos = countTrailingZeros(uint64_t(Val));
1148 } else {
1149 // For fabs and fneg, it's always the 31st bit.
1150 bitpos = 31;
1151 }
1152
1153 unsigned BitOpc = 0;
1154 // Set the right opcode for bitwise operations.
1155 switch(Opc) {
1156 default: llvm_unreachable("Only bit-wise/abs/neg operations are allowed.");
1157 case ISD::AND:
1158 case ISD::FABS:
1159 BitOpc = Hexagon::S2_clrbit_i;
1160 break;
1161 case ISD::OR:
1162 BitOpc = Hexagon::S2_setbit_i;
1163 break;
1164 case ISD::XOR:
1165 case ISD::FNEG:
1166 BitOpc = Hexagon::S2_togglebit_i;
1167 break;
1168 }
1169
1170 SDNode *Result;
1171 // Get the right SDVal for the opcode.
1172 SDValue SDVal = CurDAG->getTargetConstant(bitpos, dl, MVT::i32);
1173
1174 if (ValueVT == MVT::i32 || ValueVT == MVT::f32) {
1175 Result = CurDAG->getMachineNode(BitOpc, dl, ValueVT,
1176 N->getOperand(0), SDVal);
1177 } else {
1178 // 64-bit gymnastic to use REG_SEQUENCE. But it's worth it.
1179 EVT SubValueVT;
1180 if (ValueVT == MVT::i64)
1181 SubValueVT = MVT::i32;
1182 else
1183 SubValueVT = MVT::f32;
1184
1185 SDNode *Reg = N->getOperand(0).getNode();
1186 SDValue RegClass = CurDAG->getTargetConstant(Hexagon::DoubleRegsRegClassID,
1187 dl, MVT::i64);
1188
1189 SDValue SubregHiIdx = CurDAG->getTargetConstant(Hexagon::subreg_hireg, dl,
1190 MVT::i32);
1191 SDValue SubregLoIdx = CurDAG->getTargetConstant(Hexagon::subreg_loreg, dl,
1192 MVT::i32);
1193
1194 SDValue SubregHI = CurDAG->getTargetExtractSubreg(Hexagon::subreg_hireg, dl,
1195 MVT::i32, SDValue(Reg, 0));
1196
1197 SDValue SubregLO = CurDAG->getTargetExtractSubreg(Hexagon::subreg_loreg, dl,
1198 MVT::i32, SDValue(Reg, 0));
1199
1200 // Clear/set/toggle hi or lo registers depending on the bit position.
1201 if (SubValueVT != MVT::f32 && bitpos < 32) {
1202 SDNode *Result0 = CurDAG->getMachineNode(BitOpc, dl, SubValueVT,
1203 SubregLO, SDVal);
1204 const SDValue Ops[] = { RegClass, SubregHI, SubregHiIdx,
1205 SDValue(Result0, 0), SubregLoIdx };
1206 Result = CurDAG->getMachineNode(TargetOpcode::REG_SEQUENCE,
1207 dl, ValueVT, Ops);
1208 } else {
1209 if (Opc != ISD::FABS && Opc != ISD::FNEG)
1210 SDVal = CurDAG->getTargetConstant(bitpos - 32, dl, MVT::i32);
1211 SDNode *Result0 = CurDAG->getMachineNode(BitOpc, dl, SubValueVT,
1212 SubregHI, SDVal);
1213 const SDValue Ops[] = { RegClass, SDValue(Result0, 0), SubregHiIdx,
1214 SubregLO, SubregLoIdx };
1215 Result = CurDAG->getMachineNode(TargetOpcode::REG_SEQUENCE,
1216 dl, ValueVT, Ops);
1217 }
1218 }
1219
1220 ReplaceUses(N, Result);
1221 return Result;
1222 }
1223
1224
SelectFrameIndex(SDNode * N)1225 SDNode *HexagonDAGToDAGISel::SelectFrameIndex(SDNode *N) {
1226 MachineFrameInfo *MFI = MF->getFrameInfo();
1227 const HexagonFrameLowering *HFI = HST->getFrameLowering();
1228 int FX = cast<FrameIndexSDNode>(N)->getIndex();
1229 unsigned StkA = HFI->getStackAlignment();
1230 unsigned MaxA = MFI->getMaxAlignment();
1231 SDValue FI = CurDAG->getTargetFrameIndex(FX, MVT::i32);
1232 SDLoc DL(N);
1233 SDValue Zero = CurDAG->getTargetConstant(0, DL, MVT::i32);
1234 SDNode *R = 0;
1235
1236 // Use TFR_FI when:
1237 // - the object is fixed, or
1238 // - there are no objects with higher-than-default alignment, or
1239 // - there are no dynamically allocated objects.
1240 // Otherwise, use TFR_FIA.
1241 if (FX < 0 || MaxA <= StkA || !MFI->hasVarSizedObjects()) {
1242 R = CurDAG->getMachineNode(Hexagon::TFR_FI, DL, MVT::i32, FI, Zero);
1243 } else {
1244 auto &HMFI = *MF->getInfo<HexagonMachineFunctionInfo>();
1245 unsigned AR = HMFI.getStackAlignBaseVReg();
1246 SDValue CH = CurDAG->getEntryNode();
1247 SDValue Ops[] = { CurDAG->getCopyFromReg(CH, DL, AR, MVT::i32), FI, Zero };
1248 R = CurDAG->getMachineNode(Hexagon::TFR_FIA, DL, MVT::i32, Ops);
1249 }
1250
1251 if (N->getHasDebugValue())
1252 CurDAG->TransferDbgValues(SDValue(N, 0), SDValue(R, 0));
1253 return R;
1254 }
1255
1256
Select(SDNode * N)1257 SDNode *HexagonDAGToDAGISel::Select(SDNode *N) {
1258 if (N->isMachineOpcode()) {
1259 N->setNodeId(-1);
1260 return nullptr; // Already selected.
1261 }
1262
1263 switch (N->getOpcode()) {
1264 case ISD::Constant:
1265 return SelectConstant(N);
1266
1267 case ISD::ConstantFP:
1268 return SelectConstantFP(N);
1269
1270 case ISD::FrameIndex:
1271 return SelectFrameIndex(N);
1272
1273 case ISD::ADD:
1274 return SelectAdd(N);
1275
1276 case ISD::SHL:
1277 return SelectSHL(N);
1278
1279 case ISD::LOAD:
1280 return SelectLoad(N);
1281
1282 case ISD::STORE:
1283 return SelectStore(N);
1284
1285 case ISD::MUL:
1286 return SelectMul(N);
1287
1288 case ISD::AND:
1289 case ISD::OR:
1290 case ISD::XOR:
1291 case ISD::FABS:
1292 case ISD::FNEG:
1293 return SelectBitOp(N);
1294
1295 case ISD::ZERO_EXTEND:
1296 return SelectZeroExtend(N);
1297
1298 case ISD::INTRINSIC_W_CHAIN:
1299 return SelectIntrinsicWChain(N);
1300
1301 case ISD::INTRINSIC_WO_CHAIN:
1302 return SelectIntrinsicWOChain(N);
1303 }
1304
1305 return SelectCode(N);
1306 }
1307
1308 bool HexagonDAGToDAGISel::
SelectInlineAsmMemoryOperand(const SDValue & Op,unsigned ConstraintID,std::vector<SDValue> & OutOps)1309 SelectInlineAsmMemoryOperand(const SDValue &Op, unsigned ConstraintID,
1310 std::vector<SDValue> &OutOps) {
1311 SDValue Inp = Op, Res;
1312
1313 switch (ConstraintID) {
1314 default:
1315 return true;
1316 case InlineAsm::Constraint_i:
1317 case InlineAsm::Constraint_o: // Offsetable.
1318 case InlineAsm::Constraint_v: // Not offsetable.
1319 case InlineAsm::Constraint_m: // Memory.
1320 if (SelectAddrFI(Inp, Res))
1321 OutOps.push_back(Res);
1322 else
1323 OutOps.push_back(Inp);
1324 break;
1325 }
1326
1327 OutOps.push_back(CurDAG->getTargetConstant(0, SDLoc(Op), MVT::i32));
1328 return false;
1329 }
1330
isConstExtProfitable(SDNode * N) const1331 bool HexagonDAGToDAGISel::isConstExtProfitable(SDNode *N) const {
1332 unsigned UseCount = 0;
1333 unsigned CallCount = 0;
1334 for (SDNode::use_iterator I = N->use_begin(), E = N->use_end(); I != E; ++I) {
1335 // Ignore call instructions.
1336 if (I->getOpcode() == ISD::CopyToReg)
1337 ++CallCount;
1338 UseCount++;
1339 }
1340
1341 return (UseCount <= 1) || (CallCount > 1);
1342
1343 }
1344
PreprocessISelDAG()1345 void HexagonDAGToDAGISel::PreprocessISelDAG() {
1346 SelectionDAG &DAG = *CurDAG;
1347 std::vector<SDNode*> Nodes;
1348 for (auto I = DAG.allnodes_begin(), E = DAG.allnodes_end(); I != E; ++I)
1349 Nodes.push_back(I);
1350
1351 // Simplify: (or (select c x 0) z) -> (select c (or x z) z)
1352 // (or (select c 0 y) z) -> (select c z (or y z))
1353 // This may not be the right thing for all targets, so do it here.
1354 for (auto I: Nodes) {
1355 if (I->getOpcode() != ISD::OR)
1356 continue;
1357
1358 auto IsZero = [] (const SDValue &V) -> bool {
1359 if (ConstantSDNode *SC = dyn_cast<ConstantSDNode>(V.getNode()))
1360 return SC->isNullValue();
1361 return false;
1362 };
1363 auto IsSelect0 = [IsZero] (const SDValue &Op) -> bool {
1364 if (Op.getOpcode() != ISD::SELECT)
1365 return false;
1366 return IsZero(Op.getOperand(1)) || IsZero(Op.getOperand(2));
1367 };
1368
1369 SDValue N0 = I->getOperand(0), N1 = I->getOperand(1);
1370 EVT VT = I->getValueType(0);
1371 bool SelN0 = IsSelect0(N0);
1372 SDValue SOp = SelN0 ? N0 : N1;
1373 SDValue VOp = SelN0 ? N1 : N0;
1374
1375 if (SOp.getOpcode() == ISD::SELECT && SOp.getNode()->hasOneUse()) {
1376 SDValue SC = SOp.getOperand(0);
1377 SDValue SX = SOp.getOperand(1);
1378 SDValue SY = SOp.getOperand(2);
1379 SDLoc DLS = SOp;
1380 if (IsZero(SY)) {
1381 SDValue NewOr = DAG.getNode(ISD::OR, DLS, VT, SX, VOp);
1382 SDValue NewSel = DAG.getNode(ISD::SELECT, DLS, VT, SC, NewOr, VOp);
1383 DAG.ReplaceAllUsesWith(I, NewSel.getNode());
1384 } else if (IsZero(SX)) {
1385 SDValue NewOr = DAG.getNode(ISD::OR, DLS, VT, SY, VOp);
1386 SDValue NewSel = DAG.getNode(ISD::SELECT, DLS, VT, SC, VOp, NewOr);
1387 DAG.ReplaceAllUsesWith(I, NewSel.getNode());
1388 }
1389 }
1390 }
1391 }
1392
EmitFunctionEntryCode()1393 void HexagonDAGToDAGISel::EmitFunctionEntryCode() {
1394 auto &HST = static_cast<const HexagonSubtarget&>(MF->getSubtarget());
1395 auto &HFI = *HST.getFrameLowering();
1396 if (!HFI.needsAligna(*MF))
1397 return;
1398
1399 MachineFrameInfo *MFI = MF->getFrameInfo();
1400 MachineBasicBlock *EntryBB = MF->begin();
1401 unsigned AR = FuncInfo->CreateReg(MVT::i32);
1402 unsigned MaxA = MFI->getMaxAlignment();
1403 auto &HII = *HST.getInstrInfo();
1404 BuildMI(EntryBB, DebugLoc(), HII.get(Hexagon::ALIGNA), AR)
1405 .addImm(MaxA);
1406 MF->getInfo<HexagonMachineFunctionInfo>()->setStackAlignBaseVReg(AR);
1407 }
1408
1409 // Match a frame index that can be used in an addressing mode.
SelectAddrFI(SDValue & N,SDValue & R)1410 bool HexagonDAGToDAGISel::SelectAddrFI(SDValue& N, SDValue &R) {
1411 if (N.getOpcode() != ISD::FrameIndex)
1412 return false;
1413 auto &HFI = *HST->getFrameLowering();
1414 MachineFrameInfo *MFI = MF->getFrameInfo();
1415 int FX = cast<FrameIndexSDNode>(N)->getIndex();
1416 if (!MFI->isFixedObjectIndex(FX) && HFI.needsAligna(*MF))
1417 return false;
1418 R = CurDAG->getTargetFrameIndex(FX, MVT::i32);
1419 return true;
1420 }
1421
SelectAddrGA(SDValue & N,SDValue & R)1422 inline bool HexagonDAGToDAGISel::SelectAddrGA(SDValue &N, SDValue &R) {
1423 return SelectGlobalAddress(N, R, false);
1424 }
1425
SelectAddrGP(SDValue & N,SDValue & R)1426 inline bool HexagonDAGToDAGISel::SelectAddrGP(SDValue &N, SDValue &R) {
1427 return SelectGlobalAddress(N, R, true);
1428 }
1429
SelectGlobalAddress(SDValue & N,SDValue & R,bool UseGP)1430 bool HexagonDAGToDAGISel::SelectGlobalAddress(SDValue &N, SDValue &R,
1431 bool UseGP) {
1432 switch (N.getOpcode()) {
1433 case ISD::ADD: {
1434 SDValue N0 = N.getOperand(0);
1435 SDValue N1 = N.getOperand(1);
1436 unsigned GAOpc = N0.getOpcode();
1437 if (UseGP && GAOpc != HexagonISD::CONST32_GP)
1438 return false;
1439 if (!UseGP && GAOpc != HexagonISD::CONST32)
1440 return false;
1441 if (ConstantSDNode *Const = dyn_cast<ConstantSDNode>(N1)) {
1442 SDValue Addr = N0.getOperand(0);
1443 if (GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(Addr)) {
1444 if (GA->getOpcode() == ISD::TargetGlobalAddress) {
1445 uint64_t NewOff = GA->getOffset() + (uint64_t)Const->getSExtValue();
1446 R = CurDAG->getTargetGlobalAddress(GA->getGlobal(), SDLoc(Const),
1447 N.getValueType(), NewOff);
1448 return true;
1449 }
1450 }
1451 }
1452 break;
1453 }
1454 case HexagonISD::CONST32:
1455 // The operand(0) of CONST32 is TargetGlobalAddress, which is what we
1456 // want in the instruction.
1457 if (!UseGP)
1458 R = N.getOperand(0);
1459 return !UseGP;
1460 case HexagonISD::CONST32_GP:
1461 if (UseGP)
1462 R = N.getOperand(0);
1463 return UseGP;
1464 default:
1465 return false;
1466 }
1467
1468 return false;
1469 }
1470
isValueExtension(const SDValue & Val,unsigned FromBits,SDValue & Src)1471 bool HexagonDAGToDAGISel::isValueExtension(const SDValue &Val,
1472 unsigned FromBits, SDValue &Src) {
1473 unsigned Opc = Val.getOpcode();
1474 switch (Opc) {
1475 case ISD::SIGN_EXTEND:
1476 case ISD::ZERO_EXTEND:
1477 case ISD::ANY_EXTEND: {
1478 SDValue const &Op0 = Val.getOperand(0);
1479 EVT T = Op0.getValueType();
1480 if (T.isInteger() && T.getSizeInBits() == FromBits) {
1481 Src = Op0;
1482 return true;
1483 }
1484 break;
1485 }
1486 case ISD::SIGN_EXTEND_INREG:
1487 case ISD::AssertSext:
1488 case ISD::AssertZext:
1489 if (Val.getOperand(0).getValueType().isInteger()) {
1490 VTSDNode *T = cast<VTSDNode>(Val.getOperand(1));
1491 if (T->getVT().getSizeInBits() == FromBits) {
1492 Src = Val.getOperand(0);
1493 return true;
1494 }
1495 }
1496 break;
1497 case ISD::AND: {
1498 // Check if this is an AND with "FromBits" of lower bits set to 1.
1499 uint64_t FromMask = (1 << FromBits) - 1;
1500 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Val.getOperand(0))) {
1501 if (C->getZExtValue() == FromMask) {
1502 Src = Val.getOperand(1);
1503 return true;
1504 }
1505 }
1506 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Val.getOperand(1))) {
1507 if (C->getZExtValue() == FromMask) {
1508 Src = Val.getOperand(0);
1509 return true;
1510 }
1511 }
1512 break;
1513 }
1514 case ISD::OR:
1515 case ISD::XOR: {
1516 // OR/XOR with the lower "FromBits" bits set to 0.
1517 uint64_t FromMask = (1 << FromBits) - 1;
1518 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Val.getOperand(0))) {
1519 if ((C->getZExtValue() & FromMask) == 0) {
1520 Src = Val.getOperand(1);
1521 return true;
1522 }
1523 }
1524 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Val.getOperand(1))) {
1525 if ((C->getZExtValue() & FromMask) == 0) {
1526 Src = Val.getOperand(0);
1527 return true;
1528 }
1529 }
1530 }
1531 default:
1532 break;
1533 }
1534 return false;
1535 }
1536