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 "HexagonTargetMachine.h"
17 #include "llvm/ADT/DenseMap.h"
18 #include "llvm/CodeGen/SelectionDAGISel.h"
19 #include "llvm/IR/Intrinsics.h"
20 #include "llvm/Support/CommandLine.h"
21 #include "llvm/Support/Compiler.h"
22 #include "llvm/Support/Debug.h"
23 using namespace llvm;
24
25 #define DEBUG_TYPE "hexagon-isel"
26
27 static
28 cl::opt<unsigned>
29 MaxNumOfUsesForConstExtenders("ga-max-num-uses-for-constant-extenders",
30 cl::Hidden, cl::init(2),
31 cl::desc("Maximum number of uses of a global address such that we still us a"
32 "constant extended instruction"));
33
34 //===----------------------------------------------------------------------===//
35 // Instruction Selector Implementation
36 //===----------------------------------------------------------------------===//
37
38 namespace llvm {
39 void initializeHexagonDAGToDAGISelPass(PassRegistry&);
40 }
41
42 //===--------------------------------------------------------------------===//
43 /// HexagonDAGToDAGISel - Hexagon specific code to select Hexagon machine
44 /// instructions for SelectionDAG operations.
45 ///
46 namespace {
47 class HexagonDAGToDAGISel : public SelectionDAGISel {
48 const HexagonTargetMachine& HTM;
49 const HexagonSubtarget *HST;
50 public:
HexagonDAGToDAGISel(HexagonTargetMachine & tm,CodeGenOpt::Level OptLevel)51 explicit HexagonDAGToDAGISel(HexagonTargetMachine &tm,
52 CodeGenOpt::Level OptLevel)
53 : SelectionDAGISel(tm, OptLevel), HTM(tm) {
54 initializeHexagonDAGToDAGISelPass(*PassRegistry::getPassRegistry());
55 }
56
runOnMachineFunction(MachineFunction & MF)57 bool runOnMachineFunction(MachineFunction &MF) override {
58 // Reset the subtarget each time through.
59 HST = &MF.getSubtarget<HexagonSubtarget>();
60 SelectionDAGISel::runOnMachineFunction(MF);
61 return true;
62 }
63
64 virtual void PreprocessISelDAG() override;
65
66 SDNode *Select(SDNode *N) override;
67
68 // Complex Pattern Selectors.
69 inline bool SelectAddrGA(SDValue &N, SDValue &R);
70 inline bool SelectAddrGP(SDValue &N, SDValue &R);
71 bool SelectGlobalAddress(SDValue &N, SDValue &R, bool UseGP);
72 bool SelectAddrFI(SDValue &N, SDValue &R);
73
getPassName() const74 const char *getPassName() const override {
75 return "Hexagon DAG->DAG Pattern Instruction Selection";
76 }
77
78 SDNode *SelectFrameIndex(SDNode *N);
79 /// SelectInlineAsmMemoryOperand - Implement addressing mode selection for
80 /// inline asm expressions.
81 bool SelectInlineAsmMemoryOperand(const SDValue &Op,
82 unsigned ConstraintID,
83 std::vector<SDValue> &OutOps) override;
84 SDNode *SelectLoad(SDNode *N);
85 SDNode *SelectBaseOffsetLoad(LoadSDNode *LD, SDLoc dl);
86 SDNode *SelectIndexedLoad(LoadSDNode *LD, SDLoc dl);
87 SDNode *SelectIndexedLoadZeroExtend64(LoadSDNode *LD, unsigned Opcode,
88 SDLoc dl);
89 SDNode *SelectIndexedLoadSignExtend64(LoadSDNode *LD, unsigned Opcode,
90 SDLoc dl);
91 SDNode *SelectBaseOffsetStore(StoreSDNode *ST, SDLoc dl);
92 SDNode *SelectIndexedStore(StoreSDNode *ST, SDLoc dl);
93 SDNode *SelectStore(SDNode *N);
94 SDNode *SelectSHL(SDNode *N);
95 SDNode *SelectMul(SDNode *N);
96 SDNode *SelectZeroExtend(SDNode *N);
97 SDNode *SelectIntrinsicWChain(SDNode *N);
98 SDNode *SelectIntrinsicWOChain(SDNode *N);
99 SDNode *SelectConstant(SDNode *N);
100 SDNode *SelectConstantFP(SDNode *N);
101 SDNode *SelectAdd(SDNode *N);
102 SDNode *SelectBitOp(SDNode *N);
103
104 // XformMskToBitPosU5Imm - Returns the bit position which
105 // the single bit 32 bit mask represents.
106 // Used in Clr and Set bit immediate memops.
XformMskToBitPosU5Imm(uint32_t Imm)107 SDValue XformMskToBitPosU5Imm(uint32_t Imm) {
108 int32_t bitPos;
109 bitPos = Log2_32(Imm);
110 assert(bitPos >= 0 && bitPos < 32 &&
111 "Constant out of range for 32 BitPos Memops");
112 return CurDAG->getTargetConstant(bitPos, MVT::i32);
113 }
114
115 // XformMskToBitPosU4Imm - Returns the bit position which the single-bit
116 // 16 bit mask represents. Used in Clr and Set bit immediate memops.
XformMskToBitPosU4Imm(uint16_t Imm)117 SDValue XformMskToBitPosU4Imm(uint16_t Imm) {
118 return XformMskToBitPosU5Imm(Imm);
119 }
120
121 // XformMskToBitPosU3Imm - Returns the bit position which the single-bit
122 // 8 bit mask represents. Used in Clr and Set bit immediate memops.
XformMskToBitPosU3Imm(uint8_t Imm)123 SDValue XformMskToBitPosU3Imm(uint8_t Imm) {
124 return XformMskToBitPosU5Imm(Imm);
125 }
126
127 // Return true if there is exactly one bit set in V, i.e., if V is one of the
128 // following integers: 2^0, 2^1, ..., 2^31.
ImmIsSingleBit(uint32_t v) const129 bool ImmIsSingleBit(uint32_t v) const {
130 return isPowerOf2_32(v);
131 }
132
133 // XformM5ToU5Imm - Return a target constant with the specified value, of
134 // type i32 where the negative literal is transformed into a positive literal
135 // for use in -= memops.
XformM5ToU5Imm(signed Imm)136 inline SDValue XformM5ToU5Imm(signed Imm) {
137 assert( (Imm >= -31 && Imm <= -1) && "Constant out of range for Memops");
138 return CurDAG->getTargetConstant( - Imm, MVT::i32);
139 }
140
141 // XformU7ToU7M1Imm - Return a target constant decremented by 1, in range
142 // [1..128], used in cmpb.gtu instructions.
XformU7ToU7M1Imm(signed Imm)143 inline SDValue XformU7ToU7M1Imm(signed Imm) {
144 assert((Imm >= 1 && Imm <= 128) && "Constant out of range for cmpb op");
145 return CurDAG->getTargetConstant(Imm - 1, MVT::i8);
146 }
147
148 // XformS8ToS8M1Imm - Return a target constant decremented by 1.
XformSToSM1Imm(signed Imm)149 inline SDValue XformSToSM1Imm(signed Imm) {
150 return CurDAG->getTargetConstant(Imm - 1, MVT::i32);
151 }
152
153 // XformU8ToU8M1Imm - Return a target constant decremented by 1.
XformUToUM1Imm(unsigned Imm)154 inline SDValue XformUToUM1Imm(unsigned Imm) {
155 assert((Imm >= 1) && "Cannot decrement unsigned int less than 1");
156 return CurDAG->getTargetConstant(Imm - 1, MVT::i32);
157 }
158
159 // XformSToSM2Imm - Return a target constant decremented by 2.
XformSToSM2Imm(unsigned Imm)160 inline SDValue XformSToSM2Imm(unsigned Imm) {
161 return CurDAG->getTargetConstant(Imm - 2, MVT::i32);
162 }
163
164 // XformSToSM3Imm - Return a target constant decremented by 3.
XformSToSM3Imm(unsigned Imm)165 inline SDValue XformSToSM3Imm(unsigned Imm) {
166 return CurDAG->getTargetConstant(Imm - 3, MVT::i32);
167 }
168
169 // Include the pieces autogenerated from the target description.
170 #include "HexagonGenDAGISel.inc"
171
172 private:
173 bool isValueExtension(const SDValue &Val, unsigned FromBits, SDValue &Src);
174 }; // end HexagonDAGToDAGISel
175 } // end anonymous namespace
176
177
178 /// createHexagonISelDag - This pass converts a legalized DAG into a
179 /// Hexagon-specific DAG, ready for instruction scheduling.
180 ///
181 namespace llvm {
createHexagonISelDag(HexagonTargetMachine & TM,CodeGenOpt::Level OptLevel)182 FunctionPass *createHexagonISelDag(HexagonTargetMachine &TM,
183 CodeGenOpt::Level OptLevel) {
184 return new HexagonDAGToDAGISel(TM, OptLevel);
185 }
186 }
187
initializePassOnce(PassRegistry & Registry)188 static void initializePassOnce(PassRegistry &Registry) {
189 const char *Name = "Hexagon DAG->DAG Pattern Instruction Selection";
190 PassInfo *PI = new PassInfo(Name, "hexagon-isel",
191 &SelectionDAGISel::ID, nullptr, false, false);
192 Registry.registerPass(*PI, true);
193 }
194
initializeHexagonDAGToDAGISelPass(PassRegistry & Registry)195 void llvm::initializeHexagonDAGToDAGISelPass(PassRegistry &Registry) {
196 CALL_ONCE_INITIALIZATION(initializePassOnce)
197 }
198
199
200 // Intrinsics that return a a predicate.
doesIntrinsicReturnPredicate(unsigned ID)201 static unsigned doesIntrinsicReturnPredicate(unsigned ID)
202 {
203 switch (ID) {
204 default:
205 return 0;
206 case Intrinsic::hexagon_C2_cmpeq:
207 case Intrinsic::hexagon_C2_cmpgt:
208 case Intrinsic::hexagon_C2_cmpgtu:
209 case Intrinsic::hexagon_C2_cmpgtup:
210 case Intrinsic::hexagon_C2_cmpgtp:
211 case Intrinsic::hexagon_C2_cmpeqp:
212 case Intrinsic::hexagon_C2_bitsset:
213 case Intrinsic::hexagon_C2_bitsclr:
214 case Intrinsic::hexagon_C2_cmpeqi:
215 case Intrinsic::hexagon_C2_cmpgti:
216 case Intrinsic::hexagon_C2_cmpgtui:
217 case Intrinsic::hexagon_C2_cmpgei:
218 case Intrinsic::hexagon_C2_cmpgeui:
219 case Intrinsic::hexagon_C2_cmplt:
220 case Intrinsic::hexagon_C2_cmpltu:
221 case Intrinsic::hexagon_C2_bitsclri:
222 case Intrinsic::hexagon_C2_and:
223 case Intrinsic::hexagon_C2_or:
224 case Intrinsic::hexagon_C2_xor:
225 case Intrinsic::hexagon_C2_andn:
226 case Intrinsic::hexagon_C2_not:
227 case Intrinsic::hexagon_C2_orn:
228 case Intrinsic::hexagon_C2_pxfer_map:
229 case Intrinsic::hexagon_C2_any8:
230 case Intrinsic::hexagon_C2_all8:
231 case Intrinsic::hexagon_A2_vcmpbeq:
232 case Intrinsic::hexagon_A2_vcmpbgtu:
233 case Intrinsic::hexagon_A2_vcmpheq:
234 case Intrinsic::hexagon_A2_vcmphgt:
235 case Intrinsic::hexagon_A2_vcmphgtu:
236 case Intrinsic::hexagon_A2_vcmpweq:
237 case Intrinsic::hexagon_A2_vcmpwgt:
238 case Intrinsic::hexagon_A2_vcmpwgtu:
239 case Intrinsic::hexagon_C2_tfrrp:
240 case Intrinsic::hexagon_S2_tstbit_i:
241 case Intrinsic::hexagon_S2_tstbit_r:
242 return 1;
243 }
244 }
245
SelectIndexedLoadSignExtend64(LoadSDNode * LD,unsigned Opcode,SDLoc dl)246 SDNode *HexagonDAGToDAGISel::SelectIndexedLoadSignExtend64(LoadSDNode *LD,
247 unsigned Opcode,
248 SDLoc dl) {
249 SDValue Chain = LD->getChain();
250 EVT LoadedVT = LD->getMemoryVT();
251 SDValue Base = LD->getBasePtr();
252 SDValue Offset = LD->getOffset();
253 SDNode *OffsetNode = Offset.getNode();
254 int32_t Val = cast<ConstantSDNode>(OffsetNode)->getSExtValue();
255
256 const HexagonInstrInfo &TII = *HST->getInstrInfo();
257 if (TII.isValidAutoIncImm(LoadedVT, Val)) {
258 SDValue TargetConst = CurDAG->getTargetConstant(Val, MVT::i32);
259 SDNode *Result_1 = CurDAG->getMachineNode(Opcode, dl, MVT::i32, MVT::i32,
260 MVT::Other, Base, TargetConst,
261 Chain);
262 SDNode *Result_2 = CurDAG->getMachineNode(Hexagon::A2_sxtw, dl, MVT::i64,
263 SDValue(Result_1, 0));
264 MachineSDNode::mmo_iterator MemOp = MF->allocateMemRefsArray(1);
265 MemOp[0] = LD->getMemOperand();
266 cast<MachineSDNode>(Result_1)->setMemRefs(MemOp, MemOp + 1);
267 const SDValue Froms[] = { SDValue(LD, 0),
268 SDValue(LD, 1),
269 SDValue(LD, 2) };
270 const SDValue Tos[] = { SDValue(Result_2, 0),
271 SDValue(Result_1, 1),
272 SDValue(Result_1, 2) };
273 ReplaceUses(Froms, Tos, 3);
274 return Result_2;
275 }
276
277 SDValue TargetConst0 = CurDAG->getTargetConstant(0, MVT::i32);
278 SDValue TargetConstVal = CurDAG->getTargetConstant(Val, MVT::i32);
279 SDNode *Result_1 = CurDAG->getMachineNode(Opcode, dl, MVT::i32, MVT::Other,
280 Base, TargetConst0, Chain);
281 SDNode *Result_2 = CurDAG->getMachineNode(Hexagon::A2_sxtw, dl, MVT::i64,
282 SDValue(Result_1, 0));
283 SDNode* Result_3 = CurDAG->getMachineNode(Hexagon::A2_addi, dl, MVT::i32,
284 Base, TargetConstVal,
285 SDValue(Result_1, 1));
286 MachineSDNode::mmo_iterator MemOp = MF->allocateMemRefsArray(1);
287 MemOp[0] = LD->getMemOperand();
288 cast<MachineSDNode>(Result_1)->setMemRefs(MemOp, MemOp + 1);
289 const SDValue Froms[] = { SDValue(LD, 0),
290 SDValue(LD, 1),
291 SDValue(LD, 2) };
292 const SDValue Tos[] = { SDValue(Result_2, 0),
293 SDValue(Result_3, 0),
294 SDValue(Result_1, 1) };
295 ReplaceUses(Froms, Tos, 3);
296 return Result_2;
297 }
298
299
SelectIndexedLoadZeroExtend64(LoadSDNode * LD,unsigned Opcode,SDLoc dl)300 SDNode *HexagonDAGToDAGISel::SelectIndexedLoadZeroExtend64(LoadSDNode *LD,
301 unsigned Opcode,
302 SDLoc dl) {
303 SDValue Chain = LD->getChain();
304 EVT LoadedVT = LD->getMemoryVT();
305 SDValue Base = LD->getBasePtr();
306 SDValue Offset = LD->getOffset();
307 SDNode *OffsetNode = Offset.getNode();
308 int32_t Val = cast<ConstantSDNode>(OffsetNode)->getSExtValue();
309
310 const HexagonInstrInfo &TII = *HST->getInstrInfo();
311 if (TII.isValidAutoIncImm(LoadedVT, Val)) {
312 SDValue TargetConstVal = CurDAG->getTargetConstant(Val, MVT::i32);
313 SDValue TargetConst0 = CurDAG->getTargetConstant(0, MVT::i32);
314 SDNode *Result_1 = CurDAG->getMachineNode(Opcode, dl, MVT::i32,
315 MVT::i32, MVT::Other, Base,
316 TargetConstVal, Chain);
317 SDNode *Result_2 = CurDAG->getMachineNode(Hexagon::A4_combineir, dl,
318 MVT::i64, MVT::Other,
319 TargetConst0,
320 SDValue(Result_1,0));
321 MachineSDNode::mmo_iterator MemOp = MF->allocateMemRefsArray(1);
322 MemOp[0] = LD->getMemOperand();
323 cast<MachineSDNode>(Result_1)->setMemRefs(MemOp, MemOp + 1);
324 const SDValue Froms[] = { SDValue(LD, 0),
325 SDValue(LD, 1),
326 SDValue(LD, 2) };
327 const SDValue Tos[] = { SDValue(Result_2, 0),
328 SDValue(Result_1, 1),
329 SDValue(Result_1, 2) };
330 ReplaceUses(Froms, Tos, 3);
331 return Result_2;
332 }
333
334 // Generate an indirect load.
335 SDValue TargetConst0 = CurDAG->getTargetConstant(0, MVT::i32);
336 SDValue TargetConstVal = CurDAG->getTargetConstant(Val, MVT::i32);
337 SDNode *Result_1 = CurDAG->getMachineNode(Opcode, dl, MVT::i32,
338 MVT::Other, Base, TargetConst0,
339 Chain);
340 SDNode *Result_2 = CurDAG->getMachineNode(Hexagon::A4_combineir, dl,
341 MVT::i64, MVT::Other,
342 TargetConst0,
343 SDValue(Result_1,0));
344 // Add offset to base.
345 SDNode* Result_3 = CurDAG->getMachineNode(Hexagon::A2_addi, dl, MVT::i32,
346 Base, TargetConstVal,
347 SDValue(Result_1, 1));
348 MachineSDNode::mmo_iterator MemOp = MF->allocateMemRefsArray(1);
349 MemOp[0] = LD->getMemOperand();
350 cast<MachineSDNode>(Result_1)->setMemRefs(MemOp, MemOp + 1);
351 const SDValue Froms[] = { SDValue(LD, 0),
352 SDValue(LD, 1),
353 SDValue(LD, 2) };
354 const SDValue Tos[] = { SDValue(Result_2, 0), // Load value.
355 SDValue(Result_3, 0), // New address.
356 SDValue(Result_1, 1) };
357 ReplaceUses(Froms, Tos, 3);
358 return Result_2;
359 }
360
361
SelectIndexedLoad(LoadSDNode * LD,SDLoc dl)362 SDNode *HexagonDAGToDAGISel::SelectIndexedLoad(LoadSDNode *LD, SDLoc dl) {
363 SDValue Chain = LD->getChain();
364 SDValue Base = LD->getBasePtr();
365 SDValue Offset = LD->getOffset();
366 SDNode *OffsetNode = Offset.getNode();
367 // Get the constant value.
368 int32_t Val = cast<ConstantSDNode>(OffsetNode)->getSExtValue();
369 EVT LoadedVT = LD->getMemoryVT();
370 unsigned Opcode = 0;
371
372 // Check for zero extended loads. Treat any-extend loads as zero extended
373 // loads.
374 ISD::LoadExtType ExtType = LD->getExtensionType();
375 bool IsZeroExt = (ExtType == ISD::ZEXTLOAD || ExtType == ISD::EXTLOAD);
376
377 // Figure out the opcode.
378 const HexagonInstrInfo &TII = *HST->getInstrInfo();
379 if (LoadedVT == MVT::i64) {
380 if (TII.isValidAutoIncImm(LoadedVT, Val))
381 Opcode = Hexagon::L2_loadrd_pi;
382 else
383 Opcode = Hexagon::L2_loadrd_io;
384 } else if (LoadedVT == MVT::i32) {
385 if (TII.isValidAutoIncImm(LoadedVT, Val))
386 Opcode = Hexagon::L2_loadri_pi;
387 else
388 Opcode = Hexagon::L2_loadri_io;
389 } else if (LoadedVT == MVT::i16) {
390 if (TII.isValidAutoIncImm(LoadedVT, Val))
391 Opcode = IsZeroExt ? Hexagon::L2_loadruh_pi : Hexagon::L2_loadrh_pi;
392 else
393 Opcode = IsZeroExt ? Hexagon::L2_loadruh_io : Hexagon::L2_loadrh_io;
394 } else if (LoadedVT == MVT::i8) {
395 if (TII.isValidAutoIncImm(LoadedVT, Val))
396 Opcode = IsZeroExt ? Hexagon::L2_loadrub_pi : Hexagon::L2_loadrb_pi;
397 else
398 Opcode = IsZeroExt ? Hexagon::L2_loadrub_io : Hexagon::L2_loadrb_io;
399 } else
400 llvm_unreachable("unknown memory type");
401
402 // For zero extended i64 loads, we need to add combine instructions.
403 if (LD->getValueType(0) == MVT::i64 && IsZeroExt)
404 return SelectIndexedLoadZeroExtend64(LD, Opcode, dl);
405 // Handle sign extended i64 loads.
406 if (LD->getValueType(0) == MVT::i64 && ExtType == ISD::SEXTLOAD)
407 return SelectIndexedLoadSignExtend64(LD, Opcode, dl);
408
409 if (TII.isValidAutoIncImm(LoadedVT, Val)) {
410 SDValue TargetConstVal = CurDAG->getTargetConstant(Val, MVT::i32);
411 SDNode* Result = CurDAG->getMachineNode(Opcode, dl,
412 LD->getValueType(0),
413 MVT::i32, MVT::Other, Base,
414 TargetConstVal, Chain);
415 MachineSDNode::mmo_iterator MemOp = MF->allocateMemRefsArray(1);
416 MemOp[0] = LD->getMemOperand();
417 cast<MachineSDNode>(Result)->setMemRefs(MemOp, MemOp + 1);
418 const SDValue Froms[] = { SDValue(LD, 0),
419 SDValue(LD, 1),
420 SDValue(LD, 2)
421 };
422 const SDValue Tos[] = { SDValue(Result, 0),
423 SDValue(Result, 1),
424 SDValue(Result, 2)
425 };
426 ReplaceUses(Froms, Tos, 3);
427 return Result;
428 } else {
429 SDValue TargetConst0 = CurDAG->getTargetConstant(0, MVT::i32);
430 SDValue TargetConstVal = CurDAG->getTargetConstant(Val, MVT::i32);
431 SDNode* Result_1 = CurDAG->getMachineNode(Opcode, dl,
432 LD->getValueType(0),
433 MVT::Other, Base, TargetConst0,
434 Chain);
435 SDNode* Result_2 = CurDAG->getMachineNode(Hexagon::A2_addi, dl, MVT::i32,
436 Base, TargetConstVal,
437 SDValue(Result_1, 1));
438 MachineSDNode::mmo_iterator MemOp = MF->allocateMemRefsArray(1);
439 MemOp[0] = LD->getMemOperand();
440 cast<MachineSDNode>(Result_1)->setMemRefs(MemOp, MemOp + 1);
441 const SDValue Froms[] = { SDValue(LD, 0),
442 SDValue(LD, 1),
443 SDValue(LD, 2)
444 };
445 const SDValue Tos[] = { SDValue(Result_1, 0),
446 SDValue(Result_2, 0),
447 SDValue(Result_1, 1)
448 };
449 ReplaceUses(Froms, Tos, 3);
450 return Result_1;
451 }
452 }
453
454
SelectLoad(SDNode * N)455 SDNode *HexagonDAGToDAGISel::SelectLoad(SDNode *N) {
456 SDNode *result;
457 SDLoc dl(N);
458 LoadSDNode *LD = cast<LoadSDNode>(N);
459 ISD::MemIndexedMode AM = LD->getAddressingMode();
460
461 // Handle indexed loads.
462 if (AM != ISD::UNINDEXED) {
463 result = SelectIndexedLoad(LD, dl);
464 } else {
465 result = SelectCode(LD);
466 }
467
468 return result;
469 }
470
471
SelectIndexedStore(StoreSDNode * ST,SDLoc dl)472 SDNode *HexagonDAGToDAGISel::SelectIndexedStore(StoreSDNode *ST, SDLoc dl) {
473 SDValue Chain = ST->getChain();
474 SDValue Base = ST->getBasePtr();
475 SDValue Offset = ST->getOffset();
476 SDValue Value = ST->getValue();
477 SDNode *OffsetNode = Offset.getNode();
478 // Get the constant value.
479 int32_t Val = cast<ConstantSDNode>(OffsetNode)->getSExtValue();
480 EVT StoredVT = ST->getMemoryVT();
481 EVT ValueVT = Value.getValueType();
482
483 // Offset value must be within representable range
484 // and must have correct alignment properties.
485 const HexagonInstrInfo &TII = *HST->getInstrInfo();
486 if (TII.isValidAutoIncImm(StoredVT, Val)) {
487 unsigned Opcode = 0;
488
489 // Figure out the post inc version of opcode.
490 if (StoredVT == MVT::i64) Opcode = Hexagon::S2_storerd_pi;
491 else if (StoredVT == MVT::i32) Opcode = Hexagon::S2_storeri_pi;
492 else if (StoredVT == MVT::i16) Opcode = Hexagon::S2_storerh_pi;
493 else if (StoredVT == MVT::i8) Opcode = Hexagon::S2_storerb_pi;
494 else llvm_unreachable("unknown memory type");
495
496 if (ST->isTruncatingStore() && ValueVT.getSizeInBits() == 64) {
497 assert(StoredVT.getSizeInBits() < 64 && "Not a truncating store");
498 Value = CurDAG->getTargetExtractSubreg(Hexagon::subreg_loreg,
499 dl, MVT::i32, Value);
500 }
501 SDValue Ops[] = {Base, CurDAG->getTargetConstant(Val, MVT::i32), Value,
502 Chain};
503 // Build post increment store.
504 SDNode* Result = CurDAG->getMachineNode(Opcode, dl, MVT::i32,
505 MVT::Other, Ops);
506 MachineSDNode::mmo_iterator MemOp = MF->allocateMemRefsArray(1);
507 MemOp[0] = ST->getMemOperand();
508 cast<MachineSDNode>(Result)->setMemRefs(MemOp, MemOp + 1);
509
510 ReplaceUses(ST, Result);
511 ReplaceUses(SDValue(ST,1), SDValue(Result,1));
512 return Result;
513 }
514
515 // Note: Order of operands matches the def of instruction:
516 // def S2_storerd_io
517 // : STInst<(outs), (ins IntRegs:$base, imm:$offset, DoubleRegs:$src1), ...
518 // and it differs for POST_ST* for instance.
519 SDValue Ops[] = { Base, CurDAG->getTargetConstant(0, MVT::i32), Value,
520 Chain};
521 unsigned Opcode = 0;
522
523 // Figure out the opcode.
524 if (StoredVT == MVT::i64) Opcode = Hexagon::S2_storerd_io;
525 else if (StoredVT == MVT::i32) Opcode = Hexagon::S2_storeri_io;
526 else if (StoredVT == MVT::i16) Opcode = Hexagon::S2_storerh_io;
527 else if (StoredVT == MVT::i8) Opcode = Hexagon::S2_storerb_io;
528 else llvm_unreachable("unknown memory type");
529
530 // Build regular store.
531 SDValue TargetConstVal = CurDAG->getTargetConstant(Val, MVT::i32);
532 SDNode* Result_1 = CurDAG->getMachineNode(Opcode, dl, MVT::Other, Ops);
533 // Build splitted incriment instruction.
534 SDNode* Result_2 = CurDAG->getMachineNode(Hexagon::A2_addi, dl, MVT::i32,
535 Base,
536 TargetConstVal,
537 SDValue(Result_1, 0));
538 MachineSDNode::mmo_iterator MemOp = MF->allocateMemRefsArray(1);
539 MemOp[0] = ST->getMemOperand();
540 cast<MachineSDNode>(Result_1)->setMemRefs(MemOp, MemOp + 1);
541
542 ReplaceUses(SDValue(ST,0), SDValue(Result_2,0));
543 ReplaceUses(SDValue(ST,1), SDValue(Result_1,0));
544 return Result_2;
545 }
546
SelectStore(SDNode * N)547 SDNode *HexagonDAGToDAGISel::SelectStore(SDNode *N) {
548 SDLoc dl(N);
549 StoreSDNode *ST = cast<StoreSDNode>(N);
550 ISD::MemIndexedMode AM = ST->getAddressingMode();
551
552 // Handle indexed stores.
553 if (AM != ISD::UNINDEXED) {
554 return SelectIndexedStore(ST, dl);
555 }
556
557 return SelectCode(ST);
558 }
559
SelectMul(SDNode * N)560 SDNode *HexagonDAGToDAGISel::SelectMul(SDNode *N) {
561 SDLoc dl(N);
562
563 //
564 // %conv.i = sext i32 %tmp1 to i64
565 // %conv2.i = sext i32 %add to i64
566 // %mul.i = mul nsw i64 %conv2.i, %conv.i
567 //
568 // --- match with the following ---
569 //
570 // %mul.i = mpy (%tmp1, %add)
571 //
572
573 if (N->getValueType(0) == MVT::i64) {
574 // Shifting a i64 signed multiply.
575 SDValue MulOp0 = N->getOperand(0);
576 SDValue MulOp1 = N->getOperand(1);
577
578 SDValue OP0;
579 SDValue OP1;
580
581 // Handle sign_extend and sextload.
582 if (MulOp0.getOpcode() == ISD::SIGN_EXTEND) {
583 SDValue Sext0 = MulOp0.getOperand(0);
584 if (Sext0.getNode()->getValueType(0) != MVT::i32) {
585 return SelectCode(N);
586 }
587
588 OP0 = Sext0;
589 } else if (MulOp0.getOpcode() == ISD::LOAD) {
590 LoadSDNode *LD = cast<LoadSDNode>(MulOp0.getNode());
591 if (LD->getMemoryVT() != MVT::i32 ||
592 LD->getExtensionType() != ISD::SEXTLOAD ||
593 LD->getAddressingMode() != ISD::UNINDEXED) {
594 return SelectCode(N);
595 }
596
597 SDValue Chain = LD->getChain();
598 SDValue TargetConst0 = CurDAG->getTargetConstant(0, MVT::i32);
599 OP0 = SDValue(CurDAG->getMachineNode(Hexagon::L2_loadri_io, dl, MVT::i32,
600 MVT::Other,
601 LD->getBasePtr(), TargetConst0,
602 Chain), 0);
603 } else {
604 return SelectCode(N);
605 }
606
607 // Same goes for the second operand.
608 if (MulOp1.getOpcode() == ISD::SIGN_EXTEND) {
609 SDValue Sext1 = MulOp1.getOperand(0);
610 if (Sext1.getNode()->getValueType(0) != MVT::i32) {
611 return SelectCode(N);
612 }
613
614 OP1 = Sext1;
615 } else if (MulOp1.getOpcode() == ISD::LOAD) {
616 LoadSDNode *LD = cast<LoadSDNode>(MulOp1.getNode());
617 if (LD->getMemoryVT() != MVT::i32 ||
618 LD->getExtensionType() != ISD::SEXTLOAD ||
619 LD->getAddressingMode() != ISD::UNINDEXED) {
620 return SelectCode(N);
621 }
622
623 SDValue Chain = LD->getChain();
624 SDValue TargetConst0 = CurDAG->getTargetConstant(0, MVT::i32);
625 OP1 = SDValue(CurDAG->getMachineNode(Hexagon::L2_loadri_io, dl, MVT::i32,
626 MVT::Other,
627 LD->getBasePtr(), TargetConst0,
628 Chain), 0);
629 } else {
630 return SelectCode(N);
631 }
632
633 // Generate a mpy instruction.
634 SDNode *Result = CurDAG->getMachineNode(Hexagon::M2_dpmpyss_s0, dl, MVT::i64,
635 OP0, OP1);
636 ReplaceUses(N, Result);
637 return Result;
638 }
639
640 return SelectCode(N);
641 }
642
SelectSHL(SDNode * N)643 SDNode *HexagonDAGToDAGISel::SelectSHL(SDNode *N) {
644 SDLoc dl(N);
645 if (N->getValueType(0) == MVT::i32) {
646 SDValue Shl_0 = N->getOperand(0);
647 SDValue Shl_1 = N->getOperand(1);
648 // RHS is const.
649 if (Shl_1.getOpcode() == ISD::Constant) {
650 if (Shl_0.getOpcode() == ISD::MUL) {
651 SDValue Mul_0 = Shl_0.getOperand(0); // Val
652 SDValue Mul_1 = Shl_0.getOperand(1); // Const
653 // RHS of mul is const.
654 if (Mul_1.getOpcode() == ISD::Constant) {
655 int32_t ShlConst =
656 cast<ConstantSDNode>(Shl_1.getNode())->getSExtValue();
657 int32_t MulConst =
658 cast<ConstantSDNode>(Mul_1.getNode())->getSExtValue();
659 int32_t ValConst = MulConst << ShlConst;
660 SDValue Val = CurDAG->getTargetConstant(ValConst,
661 MVT::i32);
662 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(Val.getNode()))
663 if (isInt<9>(CN->getSExtValue())) {
664 SDNode* Result =
665 CurDAG->getMachineNode(Hexagon::M2_mpysmi, dl,
666 MVT::i32, Mul_0, Val);
667 ReplaceUses(N, Result);
668 return Result;
669 }
670
671 }
672 } else if (Shl_0.getOpcode() == ISD::SUB) {
673 SDValue Sub_0 = Shl_0.getOperand(0); // Const 0
674 SDValue Sub_1 = Shl_0.getOperand(1); // Val
675 if (Sub_0.getOpcode() == ISD::Constant) {
676 int32_t SubConst =
677 cast<ConstantSDNode>(Sub_0.getNode())->getSExtValue();
678 if (SubConst == 0) {
679 if (Sub_1.getOpcode() == ISD::SHL) {
680 SDValue Shl2_0 = Sub_1.getOperand(0); // Val
681 SDValue Shl2_1 = Sub_1.getOperand(1); // Const
682 if (Shl2_1.getOpcode() == ISD::Constant) {
683 int32_t ShlConst =
684 cast<ConstantSDNode>(Shl_1.getNode())->getSExtValue();
685 int32_t Shl2Const =
686 cast<ConstantSDNode>(Shl2_1.getNode())->getSExtValue();
687 int32_t ValConst = 1 << (ShlConst+Shl2Const);
688 SDValue Val = CurDAG->getTargetConstant(-ValConst, MVT::i32);
689 if (ConstantSDNode *CN =
690 dyn_cast<ConstantSDNode>(Val.getNode()))
691 if (isInt<9>(CN->getSExtValue())) {
692 SDNode* Result =
693 CurDAG->getMachineNode(Hexagon::M2_mpysmi, dl, MVT::i32,
694 Shl2_0, Val);
695 ReplaceUses(N, Result);
696 return Result;
697 }
698 }
699 }
700 }
701 }
702 }
703 }
704 }
705 return SelectCode(N);
706 }
707
708
709 //
710 // If there is an zero_extend followed an intrinsic in DAG (this means - the
711 // result of the intrinsic is predicate); convert the zero_extend to
712 // transfer instruction.
713 //
714 // Zero extend -> transfer is lowered here. Otherwise, zero_extend will be
715 // converted into a MUX as predicate registers defined as 1 bit in the
716 // compiler. Architecture defines them as 8-bit registers.
717 // We want to preserve all the lower 8-bits and, not just 1 LSB bit.
718 //
SelectZeroExtend(SDNode * N)719 SDNode *HexagonDAGToDAGISel::SelectZeroExtend(SDNode *N) {
720 SDLoc dl(N);
721
722 SDValue Op0 = N->getOperand(0);
723 EVT OpVT = Op0.getValueType();
724 unsigned OpBW = OpVT.getSizeInBits();
725
726 // Special handling for zero-extending a vector of booleans.
727 if (OpVT.isVector() && OpVT.getVectorElementType() == MVT::i1 && OpBW <= 64) {
728 SDNode *Mask = CurDAG->getMachineNode(Hexagon::C2_mask, dl, MVT::i64, Op0);
729 unsigned NE = OpVT.getVectorNumElements();
730 EVT ExVT = N->getValueType(0);
731 unsigned ES = ExVT.getVectorElementType().getSizeInBits();
732 uint64_t MV = 0, Bit = 1;
733 for (unsigned i = 0; i < NE; ++i) {
734 MV |= Bit;
735 Bit <<= ES;
736 }
737 SDValue Ones = CurDAG->getTargetConstant(MV, MVT::i64);
738 SDNode *OnesReg = CurDAG->getMachineNode(Hexagon::CONST64_Int_Real, dl,
739 MVT::i64, Ones);
740 if (ExVT.getSizeInBits() == 32) {
741 SDNode *And = CurDAG->getMachineNode(Hexagon::A2_andp, dl, MVT::i64,
742 SDValue(Mask,0), SDValue(OnesReg,0));
743 SDValue SubR = CurDAG->getTargetConstant(Hexagon::subreg_loreg, MVT::i32);
744 return CurDAG->getMachineNode(Hexagon::EXTRACT_SUBREG, dl, ExVT,
745 SDValue(And,0), SubR);
746 }
747 return CurDAG->getMachineNode(Hexagon::A2_andp, dl, ExVT,
748 SDValue(Mask,0), SDValue(OnesReg,0));
749 }
750
751 SDNode *IsIntrinsic = N->getOperand(0).getNode();
752 if ((IsIntrinsic->getOpcode() == ISD::INTRINSIC_WO_CHAIN)) {
753 unsigned ID =
754 cast<ConstantSDNode>(IsIntrinsic->getOperand(0))->getZExtValue();
755 if (doesIntrinsicReturnPredicate(ID)) {
756 // Now we need to differentiate target data types.
757 if (N->getValueType(0) == MVT::i64) {
758 // Convert the zero_extend to Rs = Pd followed by A2_combinew(0,Rs).
759 SDValue TargetConst0 = CurDAG->getTargetConstant(0, MVT::i32);
760 SDNode *Result_1 = CurDAG->getMachineNode(Hexagon::C2_tfrpr, dl,
761 MVT::i32,
762 SDValue(IsIntrinsic, 0));
763 SDNode *Result_2 = CurDAG->getMachineNode(Hexagon::A2_tfrsi, dl,
764 MVT::i32,
765 TargetConst0);
766 SDNode *Result_3 = CurDAG->getMachineNode(Hexagon::A2_combinew, dl,
767 MVT::i64, MVT::Other,
768 SDValue(Result_2, 0),
769 SDValue(Result_1, 0));
770 ReplaceUses(N, Result_3);
771 return Result_3;
772 }
773 if (N->getValueType(0) == MVT::i32) {
774 // Convert the zero_extend to Rs = Pd
775 SDNode* RsPd = CurDAG->getMachineNode(Hexagon::C2_tfrpr, dl,
776 MVT::i32,
777 SDValue(IsIntrinsic, 0));
778 ReplaceUses(N, RsPd);
779 return RsPd;
780 }
781 llvm_unreachable("Unexpected value type");
782 }
783 }
784 return SelectCode(N);
785 }
786
787 //
788 // Checking for intrinsics circular load/store, and bitreverse load/store
789 // instrisics in order to select the correct lowered operation.
790 //
SelectIntrinsicWChain(SDNode * N)791 SDNode *HexagonDAGToDAGISel::SelectIntrinsicWChain(SDNode *N) {
792 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
793 if (IntNo == Intrinsic::hexagon_circ_ldd ||
794 IntNo == Intrinsic::hexagon_circ_ldw ||
795 IntNo == Intrinsic::hexagon_circ_lduh ||
796 IntNo == Intrinsic::hexagon_circ_ldh ||
797 IntNo == Intrinsic::hexagon_circ_ldub ||
798 IntNo == Intrinsic::hexagon_circ_ldb) {
799 SDLoc dl(N);
800 SDValue Chain = N->getOperand(0);
801 SDValue Base = N->getOperand(2);
802 SDValue Load = N->getOperand(3);
803 SDValue ModifierExpr = N->getOperand(4);
804 SDValue Offset = N->getOperand(5);
805
806 // We need to add the rerurn type for the load. This intrinsic has
807 // two return types, one for the load and one for the post-increment.
808 // Only the *_ld instructions push the extra return type, and bump the
809 // result node operand number correspondingly.
810 std::vector<EVT> ResTys;
811 unsigned opc;
812 unsigned memsize, align;
813 MVT MvtSize = MVT::i32;
814
815 if (IntNo == Intrinsic::hexagon_circ_ldd) {
816 ResTys.push_back(MVT::i32);
817 ResTys.push_back(MVT::i64);
818 opc = Hexagon::L2_loadrd_pci_pseudo;
819 memsize = 8;
820 align = 8;
821 } else if (IntNo == Intrinsic::hexagon_circ_ldw) {
822 ResTys.push_back(MVT::i32);
823 ResTys.push_back(MVT::i32);
824 opc = Hexagon::L2_loadri_pci_pseudo;
825 memsize = 4;
826 align = 4;
827 } else if (IntNo == Intrinsic::hexagon_circ_ldh) {
828 ResTys.push_back(MVT::i32);
829 ResTys.push_back(MVT::i32);
830 opc = Hexagon::L2_loadrh_pci_pseudo;
831 memsize = 2;
832 align = 2;
833 MvtSize = MVT::i16;
834 } else if (IntNo == Intrinsic::hexagon_circ_lduh) {
835 ResTys.push_back(MVT::i32);
836 ResTys.push_back(MVT::i32);
837 opc = Hexagon::L2_loadruh_pci_pseudo;
838 memsize = 2;
839 align = 2;
840 MvtSize = MVT::i16;
841 } else if (IntNo == Intrinsic::hexagon_circ_ldb) {
842 ResTys.push_back(MVT::i32);
843 ResTys.push_back(MVT::i32);
844 opc = Hexagon::L2_loadrb_pci_pseudo;
845 memsize = 1;
846 align = 1;
847 MvtSize = MVT::i8;
848 } else if (IntNo == Intrinsic::hexagon_circ_ldub) {
849 ResTys.push_back(MVT::i32);
850 ResTys.push_back(MVT::i32);
851 opc = Hexagon::L2_loadrub_pci_pseudo;
852 memsize = 1;
853 align = 1;
854 MvtSize = MVT::i8;
855 } else
856 llvm_unreachable("no opc");
857
858 ResTys.push_back(MVT::Other);
859
860 // Copy over the arguments, which are the same mostly.
861 SmallVector<SDValue, 5> Ops;
862 Ops.push_back(Base);
863 Ops.push_back(Load);
864 Ops.push_back(ModifierExpr);
865 int32_t Val = cast<ConstantSDNode>(Offset.getNode())->getSExtValue();
866 Ops.push_back(CurDAG->getTargetConstant(Val, MVT::i32));
867 Ops.push_back(Chain);
868 SDNode* Result = CurDAG->getMachineNode(opc, dl, ResTys, Ops);
869
870 SDValue ST;
871 MachineMemOperand *Mem =
872 MF->getMachineMemOperand(MachinePointerInfo(),
873 MachineMemOperand::MOStore, memsize, align);
874 if (MvtSize != MVT::i32)
875 ST = CurDAG->getTruncStore(Chain, dl, SDValue(Result, 1), Load,
876 MvtSize, Mem);
877 else
878 ST = CurDAG->getStore(Chain, dl, SDValue(Result, 1), Load, Mem);
879
880 SDNode* Store = SelectStore(ST.getNode());
881
882 const SDValue Froms[] = { SDValue(N, 0),
883 SDValue(N, 1) };
884 const SDValue Tos[] = { SDValue(Result, 0),
885 SDValue(Store, 0) };
886 ReplaceUses(Froms, Tos, 2);
887 return Result;
888 }
889
890 if (IntNo == Intrinsic::hexagon_brev_ldd ||
891 IntNo == Intrinsic::hexagon_brev_ldw ||
892 IntNo == Intrinsic::hexagon_brev_ldh ||
893 IntNo == Intrinsic::hexagon_brev_lduh ||
894 IntNo == Intrinsic::hexagon_brev_ldb ||
895 IntNo == Intrinsic::hexagon_brev_ldub) {
896 SDLoc dl(N);
897 SDValue Chain = N->getOperand(0);
898 SDValue Base = N->getOperand(2);
899 SDValue Load = N->getOperand(3);
900 SDValue ModifierExpr = N->getOperand(4);
901
902 // We need to add the rerurn type for the load. This intrinsic has
903 // two return types, one for the load and one for the post-increment.
904 std::vector<EVT> ResTys;
905 unsigned opc;
906 unsigned memsize, align;
907 MVT MvtSize = MVT::i32;
908
909 if (IntNo == Intrinsic::hexagon_brev_ldd) {
910 ResTys.push_back(MVT::i32);
911 ResTys.push_back(MVT::i64);
912 opc = Hexagon::L2_loadrd_pbr_pseudo;
913 memsize = 8;
914 align = 8;
915 } else if (IntNo == Intrinsic::hexagon_brev_ldw) {
916 ResTys.push_back(MVT::i32);
917 ResTys.push_back(MVT::i32);
918 opc = Hexagon::L2_loadri_pbr_pseudo;
919 memsize = 4;
920 align = 4;
921 } else if (IntNo == Intrinsic::hexagon_brev_ldh) {
922 ResTys.push_back(MVT::i32);
923 ResTys.push_back(MVT::i32);
924 opc = Hexagon::L2_loadrh_pbr_pseudo;
925 memsize = 2;
926 align = 2;
927 MvtSize = MVT::i16;
928 } else if (IntNo == Intrinsic::hexagon_brev_lduh) {
929 ResTys.push_back(MVT::i32);
930 ResTys.push_back(MVT::i32);
931 opc = Hexagon::L2_loadruh_pbr_pseudo;
932 memsize = 2;
933 align = 2;
934 MvtSize = MVT::i16;
935 } else if (IntNo == Intrinsic::hexagon_brev_ldb) {
936 ResTys.push_back(MVT::i32);
937 ResTys.push_back(MVT::i32);
938 opc = Hexagon::L2_loadrb_pbr_pseudo;
939 memsize = 1;
940 align = 1;
941 MvtSize = MVT::i8;
942 } else if (IntNo == Intrinsic::hexagon_brev_ldub) {
943 ResTys.push_back(MVT::i32);
944 ResTys.push_back(MVT::i32);
945 opc = Hexagon::L2_loadrub_pbr_pseudo;
946 memsize = 1;
947 align = 1;
948 MvtSize = MVT::i8;
949 } else
950 llvm_unreachable("no opc");
951
952 ResTys.push_back(MVT::Other);
953
954 // Copy over the arguments, which are the same mostly.
955 SmallVector<SDValue, 4> Ops;
956 Ops.push_back(Base);
957 Ops.push_back(Load);
958 Ops.push_back(ModifierExpr);
959 Ops.push_back(Chain);
960 SDNode* Result = CurDAG->getMachineNode(opc, dl, ResTys, Ops);
961 SDValue ST;
962 MachineMemOperand *Mem =
963 MF->getMachineMemOperand(MachinePointerInfo(),
964 MachineMemOperand::MOStore, memsize, align);
965 if (MvtSize != MVT::i32)
966 ST = CurDAG->getTruncStore(Chain, dl, SDValue(Result, 1), Load,
967 MvtSize, Mem);
968 else
969 ST = CurDAG->getStore(Chain, dl, SDValue(Result, 1), Load, Mem);
970
971 SDNode* Store = SelectStore(ST.getNode());
972
973 const SDValue Froms[] = { SDValue(N, 0),
974 SDValue(N, 1) };
975 const SDValue Tos[] = { SDValue(Result, 0),
976 SDValue(Store, 0) };
977 ReplaceUses(Froms, Tos, 2);
978 return Result;
979 }
980
981 return SelectCode(N);
982 }
983
984 //
985 // Checking for intrinsics which have predicate registers as operand(s)
986 // and lowering to the actual intrinsic.
987 //
SelectIntrinsicWOChain(SDNode * N)988 SDNode *HexagonDAGToDAGISel::SelectIntrinsicWOChain(SDNode *N) {
989 unsigned IID = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
990 unsigned Bits;
991 switch (IID) {
992 case Intrinsic::hexagon_S2_vsplatrb:
993 Bits = 8;
994 break;
995 case Intrinsic::hexagon_S2_vsplatrh:
996 Bits = 16;
997 break;
998 default:
999 return SelectCode(N);
1000 }
1001
1002 SDValue const &V = N->getOperand(1);
1003 SDValue U;
1004 if (isValueExtension(V, Bits, U)) {
1005 SDValue R = CurDAG->getNode(N->getOpcode(), SDLoc(N), N->getValueType(0),
1006 N->getOperand(0), U);
1007 return SelectCode(R.getNode());
1008 }
1009 return SelectCode(N);
1010 }
1011
1012 //
1013 // Map floating point constant values.
1014 //
SelectConstantFP(SDNode * N)1015 SDNode *HexagonDAGToDAGISel::SelectConstantFP(SDNode *N) {
1016 SDLoc dl(N);
1017 ConstantFPSDNode *CN = dyn_cast<ConstantFPSDNode>(N);
1018 APFloat APF = CN->getValueAPF();
1019 if (N->getValueType(0) == MVT::f32) {
1020 return CurDAG->getMachineNode(Hexagon::TFRI_f, dl, MVT::f32,
1021 CurDAG->getTargetConstantFP(APF.convertToFloat(), MVT::f32));
1022 }
1023 else if (N->getValueType(0) == MVT::f64) {
1024 return CurDAG->getMachineNode(Hexagon::CONST64_Float_Real, dl, MVT::f64,
1025 CurDAG->getTargetConstantFP(APF.convertToDouble(), MVT::f64));
1026 }
1027
1028 return SelectCode(N);
1029 }
1030
1031 //
1032 // Map predicate true (encoded as -1 in LLVM) to a XOR.
1033 //
SelectConstant(SDNode * N)1034 SDNode *HexagonDAGToDAGISel::SelectConstant(SDNode *N) {
1035 SDLoc dl(N);
1036 if (N->getValueType(0) == MVT::i1) {
1037 SDNode* Result = 0;
1038 int32_t Val = cast<ConstantSDNode>(N)->getSExtValue();
1039 if (Val == -1) {
1040 Result = CurDAG->getMachineNode(Hexagon::TFR_PdTrue, dl, MVT::i1);
1041 } else if (Val == 0) {
1042 Result = CurDAG->getMachineNode(Hexagon::TFR_PdFalse, dl, MVT::i1);
1043 }
1044 if (Result) {
1045 ReplaceUses(N, Result);
1046 return Result;
1047 }
1048 }
1049
1050 return SelectCode(N);
1051 }
1052
1053
1054 //
1055 // Map add followed by a asr -> asr +=.
1056 //
SelectAdd(SDNode * N)1057 SDNode *HexagonDAGToDAGISel::SelectAdd(SDNode *N) {
1058 SDLoc dl(N);
1059 if (N->getValueType(0) != MVT::i32) {
1060 return SelectCode(N);
1061 }
1062 // Identify nodes of the form: add(asr(...)).
1063 SDNode* Src1 = N->getOperand(0).getNode();
1064 if (Src1->getOpcode() != ISD::SRA || !Src1->hasOneUse()
1065 || Src1->getValueType(0) != MVT::i32) {
1066 return SelectCode(N);
1067 }
1068
1069 // Build Rd = Rd' + asr(Rs, Rt). The machine constraints will ensure that
1070 // Rd and Rd' are assigned to the same register
1071 SDNode* Result = CurDAG->getMachineNode(Hexagon::S2_asr_r_r_acc, dl, MVT::i32,
1072 N->getOperand(1),
1073 Src1->getOperand(0),
1074 Src1->getOperand(1));
1075 ReplaceUses(N, Result);
1076
1077 return Result;
1078 }
1079
1080 //
1081 // Map the following, where possible.
1082 // AND/FABS -> clrbit
1083 // OR -> setbit
1084 // XOR/FNEG ->toggle_bit.
1085 //
SelectBitOp(SDNode * N)1086 SDNode *HexagonDAGToDAGISel::SelectBitOp(SDNode *N) {
1087 SDLoc dl(N);
1088 EVT ValueVT = N->getValueType(0);
1089
1090 // We handle only 32 and 64-bit bit ops.
1091 if (!(ValueVT == MVT::i32 || ValueVT == MVT::i64 ||
1092 ValueVT == MVT::f32 || ValueVT == MVT::f64))
1093 return SelectCode(N);
1094
1095 // We handly only fabs and fneg for V5.
1096 unsigned Opc = N->getOpcode();
1097 if ((Opc == ISD::FABS || Opc == ISD::FNEG) && !HST->hasV5TOps())
1098 return SelectCode(N);
1099
1100 int64_t Val = 0;
1101 if (Opc != ISD::FABS && Opc != ISD::FNEG) {
1102 if (N->getOperand(1).getOpcode() == ISD::Constant)
1103 Val = cast<ConstantSDNode>((N)->getOperand(1))->getSExtValue();
1104 else
1105 return SelectCode(N);
1106 }
1107
1108 if (Opc == ISD::AND) {
1109 if (((ValueVT == MVT::i32) &&
1110 (!((Val & 0x80000000) || (Val & 0x7fffffff)))) ||
1111 ((ValueVT == MVT::i64) &&
1112 (!((Val & 0x8000000000000000) || (Val & 0x7fffffff)))))
1113 // If it's simple AND, do the normal op.
1114 return SelectCode(N);
1115 else
1116 Val = ~Val;
1117 }
1118
1119 // If OR or AND is being fed by shl, srl and, sra don't do this change,
1120 // because Hexagon provide |= &= on shl, srl, and sra.
1121 // Traverse the DAG to see if there is shl, srl and sra.
1122 if (Opc == ISD::OR || Opc == ISD::AND) {
1123 switch (N->getOperand(0)->getOpcode()) {
1124 default: break;
1125 case ISD::SRA:
1126 case ISD::SRL:
1127 case ISD::SHL:
1128 return SelectCode(N);
1129 }
1130 }
1131
1132 // Make sure it's power of 2.
1133 unsigned bitpos = 0;
1134 if (Opc != ISD::FABS && Opc != ISD::FNEG) {
1135 if (((ValueVT == MVT::i32) && !isPowerOf2_32(Val)) ||
1136 ((ValueVT == MVT::i64) && !isPowerOf2_64(Val)))
1137 return SelectCode(N);
1138
1139 // Get the bit position.
1140 bitpos = countTrailingZeros(uint64_t(Val));
1141 } else {
1142 // For fabs and fneg, it's always the 31st bit.
1143 bitpos = 31;
1144 }
1145
1146 unsigned BitOpc = 0;
1147 // Set the right opcode for bitwise operations.
1148 switch(Opc) {
1149 default: llvm_unreachable("Only bit-wise/abs/neg operations are allowed.");
1150 case ISD::AND:
1151 case ISD::FABS:
1152 BitOpc = Hexagon::S2_clrbit_i;
1153 break;
1154 case ISD::OR:
1155 BitOpc = Hexagon::S2_setbit_i;
1156 break;
1157 case ISD::XOR:
1158 case ISD::FNEG:
1159 BitOpc = Hexagon::S2_togglebit_i;
1160 break;
1161 }
1162
1163 SDNode *Result;
1164 // Get the right SDVal for the opcode.
1165 SDValue SDVal = CurDAG->getTargetConstant(bitpos, MVT::i32);
1166
1167 if (ValueVT == MVT::i32 || ValueVT == MVT::f32) {
1168 Result = CurDAG->getMachineNode(BitOpc, dl, ValueVT,
1169 N->getOperand(0), SDVal);
1170 } else {
1171 // 64-bit gymnastic to use REG_SEQUENCE. But it's worth it.
1172 EVT SubValueVT;
1173 if (ValueVT == MVT::i64)
1174 SubValueVT = MVT::i32;
1175 else
1176 SubValueVT = MVT::f32;
1177
1178 SDNode *Reg = N->getOperand(0).getNode();
1179 SDValue RegClass = CurDAG->getTargetConstant(Hexagon::DoubleRegsRegClassID,
1180 MVT::i64);
1181
1182 SDValue SubregHiIdx = CurDAG->getTargetConstant(Hexagon::subreg_hireg,
1183 MVT::i32);
1184 SDValue SubregLoIdx = CurDAG->getTargetConstant(Hexagon::subreg_loreg,
1185 MVT::i32);
1186
1187 SDValue SubregHI = CurDAG->getTargetExtractSubreg(Hexagon::subreg_hireg, dl,
1188 MVT::i32, SDValue(Reg, 0));
1189
1190 SDValue SubregLO = CurDAG->getTargetExtractSubreg(Hexagon::subreg_loreg, dl,
1191 MVT::i32, SDValue(Reg, 0));
1192
1193 // Clear/set/toggle hi or lo registers depending on the bit position.
1194 if (SubValueVT != MVT::f32 && bitpos < 32) {
1195 SDNode *Result0 = CurDAG->getMachineNode(BitOpc, dl, SubValueVT,
1196 SubregLO, SDVal);
1197 const SDValue Ops[] = { RegClass, SubregHI, SubregHiIdx,
1198 SDValue(Result0, 0), SubregLoIdx };
1199 Result = CurDAG->getMachineNode(TargetOpcode::REG_SEQUENCE,
1200 dl, ValueVT, Ops);
1201 } else {
1202 if (Opc != ISD::FABS && Opc != ISD::FNEG)
1203 SDVal = CurDAG->getTargetConstant(bitpos-32, MVT::i32);
1204 SDNode *Result0 = CurDAG->getMachineNode(BitOpc, dl, SubValueVT,
1205 SubregHI, SDVal);
1206 const SDValue Ops[] = { RegClass, SDValue(Result0, 0), SubregHiIdx,
1207 SubregLO, SubregLoIdx };
1208 Result = CurDAG->getMachineNode(TargetOpcode::REG_SEQUENCE,
1209 dl, ValueVT, Ops);
1210 }
1211 }
1212
1213 ReplaceUses(N, Result);
1214 return Result;
1215 }
1216
1217
SelectFrameIndex(SDNode * N)1218 SDNode *HexagonDAGToDAGISel::SelectFrameIndex(SDNode *N) {
1219 int FX = cast<FrameIndexSDNode>(N)->getIndex();
1220 SDValue FI = CurDAG->getTargetFrameIndex(FX, MVT::i32);
1221 SDValue Zero = CurDAG->getTargetConstant(0, MVT::i32);
1222 SDLoc DL(N);
1223
1224 SDNode *R = CurDAG->getMachineNode(Hexagon::TFR_FI, DL, MVT::i32, FI, Zero);
1225
1226 if (N->getHasDebugValue())
1227 CurDAG->TransferDbgValues(SDValue(N, 0), SDValue(R, 0));
1228 return R;
1229 }
1230
1231
Select(SDNode * N)1232 SDNode *HexagonDAGToDAGISel::Select(SDNode *N) {
1233 if (N->isMachineOpcode()) {
1234 N->setNodeId(-1);
1235 return nullptr; // Already selected.
1236 }
1237
1238 switch (N->getOpcode()) {
1239 case ISD::Constant:
1240 return SelectConstant(N);
1241
1242 case ISD::ConstantFP:
1243 return SelectConstantFP(N);
1244
1245 case ISD::FrameIndex:
1246 return SelectFrameIndex(N);
1247
1248 case ISD::ADD:
1249 return SelectAdd(N);
1250
1251 case ISD::SHL:
1252 return SelectSHL(N);
1253
1254 case ISD::LOAD:
1255 return SelectLoad(N);
1256
1257 case ISD::STORE:
1258 return SelectStore(N);
1259
1260 case ISD::MUL:
1261 return SelectMul(N);
1262
1263 case ISD::AND:
1264 case ISD::OR:
1265 case ISD::XOR:
1266 case ISD::FABS:
1267 case ISD::FNEG:
1268 return SelectBitOp(N);
1269
1270 case ISD::ZERO_EXTEND:
1271 return SelectZeroExtend(N);
1272
1273 case ISD::INTRINSIC_W_CHAIN:
1274 return SelectIntrinsicWChain(N);
1275
1276 case ISD::INTRINSIC_WO_CHAIN:
1277 return SelectIntrinsicWOChain(N);
1278 }
1279
1280 return SelectCode(N);
1281 }
1282
1283
1284 bool HexagonDAGToDAGISel::
SelectInlineAsmMemoryOperand(const SDValue & Op,unsigned ConstraintID,std::vector<SDValue> & OutOps)1285 SelectInlineAsmMemoryOperand(const SDValue &Op, unsigned ConstraintID,
1286 std::vector<SDValue> &OutOps) {
1287 SDValue Inp = Op, Res;
1288
1289 switch (ConstraintID) {
1290 default:
1291 return true;
1292 case InlineAsm::Constraint_i:
1293 case InlineAsm::Constraint_o: // Offsetable.
1294 case InlineAsm::Constraint_v: // Not offsetable.
1295 case InlineAsm::Constraint_m: // Memory.
1296 if (SelectAddrFI(Inp, Res))
1297 OutOps.push_back(Res);
1298 else
1299 OutOps.push_back(Inp);
1300 break;
1301 }
1302
1303 OutOps.push_back(CurDAG->getTargetConstant(0, MVT::i32));
1304 return false;
1305 }
1306
PreprocessISelDAG()1307 void HexagonDAGToDAGISel::PreprocessISelDAG() {
1308 SelectionDAG &DAG = *CurDAG;
1309 std::vector<SDNode*> Nodes;
1310 for (auto I = DAG.allnodes_begin(), E = DAG.allnodes_end(); I != E; ++I)
1311 Nodes.push_back(I);
1312
1313 // Simplify: (or (select c x 0) z) -> (select c (or x z) z)
1314 // (or (select c 0 y) z) -> (select c z (or y z))
1315 // This may not be the right thing for all targets, so do it here.
1316 for (auto I: Nodes) {
1317 if (I->getOpcode() != ISD::OR)
1318 continue;
1319
1320 auto IsZero = [] (const SDValue &V) -> bool {
1321 if (ConstantSDNode *SC = dyn_cast<ConstantSDNode>(V.getNode()))
1322 return SC->isNullValue();
1323 return false;
1324 };
1325 auto IsSelect0 = [IsZero] (const SDValue &Op) -> bool {
1326 if (Op.getOpcode() != ISD::SELECT)
1327 return false;
1328 return IsZero(Op.getOperand(1)) || IsZero(Op.getOperand(2));
1329 };
1330
1331 SDValue N0 = I->getOperand(0), N1 = I->getOperand(1);
1332 EVT VT = I->getValueType(0);
1333 bool SelN0 = IsSelect0(N0);
1334 SDValue SOp = SelN0 ? N0 : N1;
1335 SDValue VOp = SelN0 ? N1 : N0;
1336
1337 if (SOp.getOpcode() == ISD::SELECT && SOp.getNode()->hasOneUse()) {
1338 SDValue SC = SOp.getOperand(0);
1339 SDValue SX = SOp.getOperand(1);
1340 SDValue SY = SOp.getOperand(2);
1341 SDLoc DLS = SOp;
1342 if (IsZero(SY)) {
1343 SDValue NewOr = DAG.getNode(ISD::OR, DLS, VT, SX, VOp);
1344 SDValue NewSel = DAG.getNode(ISD::SELECT, DLS, VT, SC, NewOr, VOp);
1345 DAG.ReplaceAllUsesWith(I, NewSel.getNode());
1346 } else if (IsZero(SX)) {
1347 SDValue NewOr = DAG.getNode(ISD::OR, DLS, VT, SY, VOp);
1348 SDValue NewSel = DAG.getNode(ISD::SELECT, DLS, VT, SC, VOp, NewOr);
1349 DAG.ReplaceAllUsesWith(I, NewSel.getNode());
1350 }
1351 }
1352 }
1353 }
1354
1355
SelectAddrFI(SDValue & N,SDValue & R)1356 bool HexagonDAGToDAGISel::SelectAddrFI(SDValue& N, SDValue &R) {
1357 if (N.getOpcode() != ISD::FrameIndex)
1358 return false;
1359 FrameIndexSDNode *FX = cast<FrameIndexSDNode>(N);
1360 R = CurDAG->getTargetFrameIndex(FX->getIndex(), MVT::i32);
1361 return true;
1362 }
1363
SelectAddrGA(SDValue & N,SDValue & R)1364 inline bool HexagonDAGToDAGISel::SelectAddrGA(SDValue &N, SDValue &R) {
1365 return SelectGlobalAddress(N, R, false);
1366 }
1367
SelectAddrGP(SDValue & N,SDValue & R)1368 inline bool HexagonDAGToDAGISel::SelectAddrGP(SDValue &N, SDValue &R) {
1369 return SelectGlobalAddress(N, R, true);
1370 }
1371
SelectGlobalAddress(SDValue & N,SDValue & R,bool UseGP)1372 bool HexagonDAGToDAGISel::SelectGlobalAddress(SDValue &N, SDValue &R,
1373 bool UseGP) {
1374 switch (N.getOpcode()) {
1375 case ISD::ADD: {
1376 SDValue N0 = N.getOperand(0);
1377 SDValue N1 = N.getOperand(1);
1378 unsigned GAOpc = N0.getOpcode();
1379 if (UseGP && GAOpc != HexagonISD::CONST32_GP)
1380 return false;
1381 if (!UseGP && GAOpc != HexagonISD::CONST32)
1382 return false;
1383 if (ConstantSDNode *Const = dyn_cast<ConstantSDNode>(N1)) {
1384 SDValue Addr = N0.getOperand(0);
1385 if (GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(Addr)) {
1386 if (GA->getOpcode() == ISD::TargetGlobalAddress) {
1387 uint64_t NewOff = GA->getOffset() + (uint64_t)Const->getSExtValue();
1388 R = CurDAG->getTargetGlobalAddress(GA->getGlobal(), SDLoc(Const),
1389 N.getValueType(), NewOff);
1390 return true;
1391 }
1392 }
1393 }
1394 break;
1395 }
1396 case HexagonISD::CONST32:
1397 // The operand(0) of CONST32 is TargetGlobalAddress, which is what we
1398 // want in the instruction.
1399 if (!UseGP)
1400 R = N.getOperand(0);
1401 return !UseGP;
1402 case HexagonISD::CONST32_GP:
1403 if (UseGP)
1404 R = N.getOperand(0);
1405 return UseGP;
1406 default:
1407 return false;
1408 }
1409
1410 return false;
1411 }
1412
isValueExtension(const SDValue & Val,unsigned FromBits,SDValue & Src)1413 bool HexagonDAGToDAGISel::isValueExtension(const SDValue &Val,
1414 unsigned FromBits, SDValue &Src) {
1415 unsigned Opc = Val.getOpcode();
1416 switch (Opc) {
1417 case ISD::SIGN_EXTEND:
1418 case ISD::ZERO_EXTEND:
1419 case ISD::ANY_EXTEND: {
1420 SDValue const &Op0 = Val.getOperand(0);
1421 EVT T = Op0.getValueType();
1422 if (T.isInteger() && T.getSizeInBits() == FromBits) {
1423 Src = Op0;
1424 return true;
1425 }
1426 break;
1427 }
1428 case ISD::SIGN_EXTEND_INREG:
1429 case ISD::AssertSext:
1430 case ISD::AssertZext:
1431 if (Val.getOperand(0).getValueType().isInteger()) {
1432 VTSDNode *T = cast<VTSDNode>(Val.getOperand(1));
1433 if (T->getVT().getSizeInBits() == FromBits) {
1434 Src = Val.getOperand(0);
1435 return true;
1436 }
1437 }
1438 break;
1439 case ISD::AND: {
1440 // Check if this is an AND with "FromBits" of lower bits set to 1.
1441 uint64_t FromMask = (1 << FromBits) - 1;
1442 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Val.getOperand(0))) {
1443 if (C->getZExtValue() == FromMask) {
1444 Src = Val.getOperand(1);
1445 return true;
1446 }
1447 }
1448 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Val.getOperand(1))) {
1449 if (C->getZExtValue() == FromMask) {
1450 Src = Val.getOperand(0);
1451 return true;
1452 }
1453 }
1454 break;
1455 }
1456 case ISD::OR:
1457 case ISD::XOR: {
1458 // OR/XOR with the lower "FromBits" bits set to 0.
1459 uint64_t FromMask = (1 << FromBits) - 1;
1460 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Val.getOperand(0))) {
1461 if ((C->getZExtValue() & FromMask) == 0) {
1462 Src = Val.getOperand(1);
1463 return true;
1464 }
1465 }
1466 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Val.getOperand(1))) {
1467 if ((C->getZExtValue() & FromMask) == 0) {
1468 Src = Val.getOperand(0);
1469 return true;
1470 }
1471 }
1472 }
1473 default:
1474 break;
1475 }
1476 return false;
1477 }
1478