1 //===-- HexagonISelLowering.cpp - Hexagon DAG Lowering Implementation -----===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file implements the interfaces that Hexagon uses to lower LLVM code
11 // into a selection DAG.
13 //===----------------------------------------------------------------------===//
15 #include "HexagonISelLowering.h"
16 #include "HexagonMachineFunctionInfo.h"
17 #include "HexagonSubtarget.h"
18 #include "HexagonTargetMachine.h"
19 #include "HexagonTargetObjectFile.h"
20 #include "llvm/CodeGen/CallingConvLower.h"
21 #include "llvm/CodeGen/MachineFrameInfo.h"
22 #include "llvm/CodeGen/MachineFunction.h"
23 #include "llvm/CodeGen/MachineInstrBuilder.h"
24 #include "llvm/CodeGen/MachineJumpTableInfo.h"
25 #include "llvm/CodeGen/MachineRegisterInfo.h"
26 #include "llvm/CodeGen/SelectionDAGISel.h"
27 #include "llvm/CodeGen/ValueTypes.h"
28 #include "llvm/IR/CallingConv.h"
29 #include "llvm/IR/DerivedTypes.h"
30 #include "llvm/IR/Function.h"
31 #include "llvm/IR/GlobalAlias.h"
32 #include "llvm/IR/GlobalVariable.h"
33 #include "llvm/IR/InlineAsm.h"
34 #include "llvm/IR/Intrinsics.h"
35 #include "llvm/Support/CommandLine.h"
36 #include "llvm/Support/Debug.h"
37 #include "llvm/Support/ErrorHandling.h"
38 #include "llvm/Support/raw_ostream.h"
42 #define DEBUG_TYPE "hexagon-lowering"
45 EmitJumpTables("hexagon-emit-jump-tables", cl::init(true), cl::Hidden,
46 cl::desc("Control jump table emission on Hexagon target"));
48 static cl::opt<bool> EnableHexSDNodeSched("enable-hexagon-sdnode-sched",
49 cl::Hidden, cl::ZeroOrMore, cl::init(false),
50 cl::desc("Enable Hexagon SDNode scheduling"));
52 static cl::opt<bool> EnableFastMath("ffast-math",
53 cl::Hidden, cl::ZeroOrMore, cl::init(false),
54 cl::desc("Enable Fast Math processing"));
56 static cl::opt<int> MinimumJumpTables("minimum-jump-tables",
57 cl::Hidden, cl::ZeroOrMore, cl::init(5),
58 cl::desc("Set minimum jump tables"));
60 static cl::opt<int> MaxStoresPerMemcpyCL("max-store-memcpy",
61 cl::Hidden, cl::ZeroOrMore, cl::init(6),
62 cl::desc("Max #stores to inline memcpy"));
64 static cl::opt<int> MaxStoresPerMemcpyOptSizeCL("max-store-memcpy-Os",
65 cl::Hidden, cl::ZeroOrMore, cl::init(4),
66 cl::desc("Max #stores to inline memcpy"));
68 static cl::opt<int> MaxStoresPerMemmoveCL("max-store-memmove",
69 cl::Hidden, cl::ZeroOrMore, cl::init(6),
70 cl::desc("Max #stores to inline memmove"));
72 static cl::opt<int> MaxStoresPerMemmoveOptSizeCL("max-store-memmove-Os",
73 cl::Hidden, cl::ZeroOrMore, cl::init(4),
74 cl::desc("Max #stores to inline memmove"));
76 static cl::opt<int> MaxStoresPerMemsetCL("max-store-memset",
77 cl::Hidden, cl::ZeroOrMore, cl::init(8),
78 cl::desc("Max #stores to inline memset"));
80 static cl::opt<int> MaxStoresPerMemsetOptSizeCL("max-store-memset-Os",
81 cl::Hidden, cl::ZeroOrMore, cl::init(4),
82 cl::desc("Max #stores to inline memset"));
86 class HexagonCCState : public CCState {
87 unsigned NumNamedVarArgParams;
90 HexagonCCState(CallingConv::ID CC, bool isVarArg, MachineFunction &MF,
91 SmallVectorImpl<CCValAssign> &locs, LLVMContext &C,
92 int NumNamedVarArgParams)
93 : CCState(CC, isVarArg, MF, locs, C),
94 NumNamedVarArgParams(NumNamedVarArgParams) {}
96 unsigned getNumNamedVarArgParams() const { return NumNamedVarArgParams; }
100 // Implement calling convention for Hexagon.
102 static bool IsHvxVectorType(MVT ty);
105 CC_Hexagon(unsigned ValNo, MVT ValVT,
106 MVT LocVT, CCValAssign::LocInfo LocInfo,
107 ISD::ArgFlagsTy ArgFlags, CCState &State);
110 CC_Hexagon32(unsigned ValNo, MVT ValVT,
111 MVT LocVT, CCValAssign::LocInfo LocInfo,
112 ISD::ArgFlagsTy ArgFlags, CCState &State);
115 CC_Hexagon64(unsigned ValNo, MVT ValVT,
116 MVT LocVT, CCValAssign::LocInfo LocInfo,
117 ISD::ArgFlagsTy ArgFlags, CCState &State);
120 CC_HexagonVector(unsigned ValNo, MVT ValVT,
121 MVT LocVT, CCValAssign::LocInfo LocInfo,
122 ISD::ArgFlagsTy ArgFlags, CCState &State);
125 RetCC_Hexagon(unsigned ValNo, MVT ValVT,
126 MVT LocVT, CCValAssign::LocInfo LocInfo,
127 ISD::ArgFlagsTy ArgFlags, CCState &State);
130 RetCC_Hexagon32(unsigned ValNo, MVT ValVT,
131 MVT LocVT, CCValAssign::LocInfo LocInfo,
132 ISD::ArgFlagsTy ArgFlags, CCState &State);
135 RetCC_Hexagon64(unsigned ValNo, MVT ValVT,
136 MVT LocVT, CCValAssign::LocInfo LocInfo,
137 ISD::ArgFlagsTy ArgFlags, CCState &State);
140 RetCC_HexagonVector(unsigned ValNo, MVT ValVT,
141 MVT LocVT, CCValAssign::LocInfo LocInfo,
142 ISD::ArgFlagsTy ArgFlags, CCState &State);
145 CC_Hexagon_VarArg (unsigned ValNo, MVT ValVT,
146 MVT LocVT, CCValAssign::LocInfo LocInfo,
147 ISD::ArgFlagsTy ArgFlags, CCState &State) {
148 HexagonCCState &HState = static_cast<HexagonCCState &>(State);
150 if (ValNo < HState.getNumNamedVarArgParams()) {
151 // Deal with named arguments.
152 return CC_Hexagon(ValNo, ValVT, LocVT, LocInfo, ArgFlags, State);
155 // Deal with un-named arguments.
157 if (ArgFlags.isByVal()) {
158 // If pass-by-value, the size allocated on stack is decided
159 // by ArgFlags.getByValSize(), not by the size of LocVT.
160 ofst = State.AllocateStack(ArgFlags.getByValSize(),
161 ArgFlags.getByValAlign());
162 State.addLoc(CCValAssign::getMem(ValNo, ValVT, ofst, LocVT, LocInfo));
165 if (LocVT == MVT::i1 || LocVT == MVT::i8 || LocVT == MVT::i16) {
168 if (ArgFlags.isSExt())
169 LocInfo = CCValAssign::SExt;
170 else if (ArgFlags.isZExt())
171 LocInfo = CCValAssign::ZExt;
173 LocInfo = CCValAssign::AExt;
175 if (LocVT == MVT::i32 || LocVT == MVT::f32) {
176 ofst = State.AllocateStack(4, 4);
177 State.addLoc(CCValAssign::getMem(ValNo, ValVT, ofst, LocVT, LocInfo));
180 if (LocVT == MVT::i64 || LocVT == MVT::f64) {
181 ofst = State.AllocateStack(8, 8);
182 State.addLoc(CCValAssign::getMem(ValNo, ValVT, ofst, LocVT, LocInfo));
185 if (LocVT == MVT::v2i64 || LocVT == MVT::v4i32 || LocVT == MVT::v8i16 ||
186 LocVT == MVT::v16i8) {
187 ofst = State.AllocateStack(16, 16);
188 State.addLoc(CCValAssign::getMem(ValNo, ValVT, ofst, LocVT, LocInfo));
191 if (LocVT == MVT::v4i64 || LocVT == MVT::v8i32 || LocVT == MVT::v16i16 ||
192 LocVT == MVT::v32i8) {
193 ofst = State.AllocateStack(32, 32);
194 State.addLoc(CCValAssign::getMem(ValNo, ValVT, ofst, LocVT, LocInfo));
197 if (LocVT == MVT::v8i64 || LocVT == MVT::v16i32 || LocVT == MVT::v32i16 ||
198 LocVT == MVT::v64i8 || LocVT == MVT::v512i1) {
199 ofst = State.AllocateStack(64, 64);
200 State.addLoc(CCValAssign::getMem(ValNo, ValVT, ofst, LocVT, LocInfo));
203 if (LocVT == MVT::v16i64 || LocVT == MVT::v32i32 || LocVT == MVT::v64i16 ||
204 LocVT == MVT::v128i8 || LocVT == MVT::v1024i1) {
205 ofst = State.AllocateStack(128, 128);
206 State.addLoc(CCValAssign::getMem(ValNo, ValVT, ofst, LocVT, LocInfo));
209 if (LocVT == MVT::v32i64 || LocVT == MVT::v64i32 || LocVT == MVT::v128i16 ||
210 LocVT == MVT::v256i8) {
211 ofst = State.AllocateStack(256, 256);
212 State.addLoc(CCValAssign::getMem(ValNo, ValVT, ofst, LocVT, LocInfo));
216 llvm_unreachable(nullptr);
220 static bool CC_Hexagon (unsigned ValNo, MVT ValVT, MVT LocVT,
221 CCValAssign::LocInfo LocInfo, ISD::ArgFlagsTy ArgFlags, CCState &State) {
222 if (ArgFlags.isByVal()) {
224 unsigned Offset = State.AllocateStack(ArgFlags.getByValSize(),
225 ArgFlags.getByValAlign());
226 State.addLoc(CCValAssign::getMem(ValNo, ValVT, Offset, LocVT, LocInfo));
230 if (LocVT == MVT::i1 || LocVT == MVT::i8 || LocVT == MVT::i16) {
233 if (ArgFlags.isSExt())
234 LocInfo = CCValAssign::SExt;
235 else if (ArgFlags.isZExt())
236 LocInfo = CCValAssign::ZExt;
238 LocInfo = CCValAssign::AExt;
239 } else if (LocVT == MVT::v4i8 || LocVT == MVT::v2i16) {
241 LocInfo = CCValAssign::BCvt;
242 } else if (LocVT == MVT::v8i8 || LocVT == MVT::v4i16 || LocVT == MVT::v2i32) {
244 LocInfo = CCValAssign::BCvt;
247 if (LocVT == MVT::i32 || LocVT == MVT::f32) {
248 if (!CC_Hexagon32(ValNo, ValVT, LocVT, LocInfo, ArgFlags, State))
252 if (LocVT == MVT::i64 || LocVT == MVT::f64) {
253 if (!CC_Hexagon64(ValNo, ValVT, LocVT, LocInfo, ArgFlags, State))
257 if (LocVT == MVT::v8i32 || LocVT == MVT::v16i16 || LocVT == MVT::v32i8) {
258 unsigned Offset = State.AllocateStack(ArgFlags.getByValSize(), 32);
259 State.addLoc(CCValAssign::getMem(ValNo, ValVT, Offset, LocVT, LocInfo));
263 if (IsHvxVectorType(LocVT)) {
264 if (!CC_HexagonVector(ValNo, ValVT, LocVT, LocInfo, ArgFlags, State))
268 return true; // CC didn't match.
272 static bool CC_Hexagon32(unsigned ValNo, MVT ValVT,
273 MVT LocVT, CCValAssign::LocInfo LocInfo,
274 ISD::ArgFlagsTy ArgFlags, CCState &State) {
276 static const MCPhysReg RegList[] = {
277 Hexagon::R0, Hexagon::R1, Hexagon::R2, Hexagon::R3, Hexagon::R4,
280 if (unsigned Reg = State.AllocateReg(RegList)) {
281 State.addLoc(CCValAssign::getReg(ValNo, ValVT, Reg, LocVT, LocInfo));
285 unsigned Offset = State.AllocateStack(4, 4);
286 State.addLoc(CCValAssign::getMem(ValNo, ValVT, Offset, LocVT, LocInfo));
290 static bool CC_Hexagon64(unsigned ValNo, MVT ValVT,
291 MVT LocVT, CCValAssign::LocInfo LocInfo,
292 ISD::ArgFlagsTy ArgFlags, CCState &State) {
294 if (unsigned Reg = State.AllocateReg(Hexagon::D0)) {
295 State.addLoc(CCValAssign::getReg(ValNo, ValVT, Reg, LocVT, LocInfo));
299 static const MCPhysReg RegList1[] = {
300 Hexagon::D1, Hexagon::D2
302 static const MCPhysReg RegList2[] = {
303 Hexagon::R1, Hexagon::R3
305 if (unsigned Reg = State.AllocateReg(RegList1, RegList2)) {
306 State.addLoc(CCValAssign::getReg(ValNo, ValVT, Reg, LocVT, LocInfo));
310 unsigned Offset = State.AllocateStack(8, 8, Hexagon::D2);
311 State.addLoc(CCValAssign::getMem(ValNo, ValVT, Offset, LocVT, LocInfo));
315 static bool CC_HexagonVector(unsigned ValNo, MVT ValVT,
316 MVT LocVT, CCValAssign::LocInfo LocInfo,
317 ISD::ArgFlagsTy ArgFlags, CCState &State) {
319 static const uint16_t VecLstS[] = { Hexagon::V0, Hexagon::V1,
320 Hexagon::V2, Hexagon::V3,
321 Hexagon::V4, Hexagon::V5,
322 Hexagon::V6, Hexagon::V7,
323 Hexagon::V8, Hexagon::V9,
324 Hexagon::V10, Hexagon::V11,
325 Hexagon::V12, Hexagon::V13,
326 Hexagon::V14, Hexagon::V15};
327 static const uint16_t VecLstD[] = { Hexagon::W0, Hexagon::W1,
328 Hexagon::W2, Hexagon::W3,
329 Hexagon::W4, Hexagon::W5,
330 Hexagon::W6, Hexagon::W7};
331 auto &MF = State.getMachineFunction();
332 auto &HST = MF.getSubtarget<HexagonSubtarget>();
333 bool UseHVX = HST.useHVXOps();
334 bool UseHVXDbl = HST.useHVXDblOps();
336 if ((UseHVX && !UseHVXDbl) &&
337 (LocVT == MVT::v8i64 || LocVT == MVT::v16i32 || LocVT == MVT::v32i16 ||
338 LocVT == MVT::v64i8 || LocVT == MVT::v512i1)) {
339 if (unsigned Reg = State.AllocateReg(VecLstS)) {
340 State.addLoc(CCValAssign::getReg(ValNo, ValVT, Reg, LocVT, LocInfo));
343 unsigned Offset = State.AllocateStack(64, 64);
344 State.addLoc(CCValAssign::getMem(ValNo, ValVT, Offset, LocVT, LocInfo));
347 if ((UseHVX && !UseHVXDbl) &&
348 (LocVT == MVT::v16i64 || LocVT == MVT::v32i32 || LocVT == MVT::v64i16 ||
349 LocVT == MVT::v128i8)) {
350 if (unsigned Reg = State.AllocateReg(VecLstD)) {
351 State.addLoc(CCValAssign::getReg(ValNo, ValVT, Reg, LocVT, LocInfo));
354 unsigned Offset = State.AllocateStack(128, 128);
355 State.addLoc(CCValAssign::getMem(ValNo, ValVT, Offset, LocVT, LocInfo));
359 if ((UseHVX && UseHVXDbl) &&
360 (LocVT == MVT::v32i64 || LocVT == MVT::v64i32 || LocVT == MVT::v128i16 ||
361 LocVT == MVT::v256i8)) {
362 if (unsigned Reg = State.AllocateReg(VecLstD)) {
363 State.addLoc(CCValAssign::getReg(ValNo, ValVT, Reg, LocVT, LocInfo));
366 unsigned Offset = State.AllocateStack(256, 256);
367 State.addLoc(CCValAssign::getMem(ValNo, ValVT, Offset, LocVT, LocInfo));
370 if ((UseHVX && UseHVXDbl) &&
371 (LocVT == MVT::v16i64 || LocVT == MVT::v32i32 || LocVT == MVT::v64i16 ||
372 LocVT == MVT::v128i8 || LocVT == MVT::v1024i1)) {
373 if (unsigned Reg = State.AllocateReg(VecLstS)) {
374 State.addLoc(CCValAssign::getReg(ValNo, ValVT, Reg, LocVT, LocInfo));
377 unsigned Offset = State.AllocateStack(128, 128);
378 State.addLoc(CCValAssign::getMem(ValNo, ValVT, Offset, LocVT, LocInfo));
384 static bool RetCC_Hexagon(unsigned ValNo, MVT ValVT,
385 MVT LocVT, CCValAssign::LocInfo LocInfo,
386 ISD::ArgFlagsTy ArgFlags, CCState &State) {
387 auto &MF = State.getMachineFunction();
388 auto &HST = MF.getSubtarget<HexagonSubtarget>();
389 bool UseHVX = HST.useHVXOps();
390 bool UseHVXDbl = HST.useHVXDblOps();
392 if (LocVT == MVT::i1 ||
397 if (ArgFlags.isSExt())
398 LocInfo = CCValAssign::SExt;
399 else if (ArgFlags.isZExt())
400 LocInfo = CCValAssign::ZExt;
402 LocInfo = CCValAssign::AExt;
403 } else if (LocVT == MVT::v4i8 || LocVT == MVT::v2i16) {
405 LocInfo = CCValAssign::BCvt;
406 } else if (LocVT == MVT::v8i8 || LocVT == MVT::v4i16 || LocVT == MVT::v2i32) {
408 LocInfo = CCValAssign::BCvt;
409 } else if (LocVT == MVT::v64i8 || LocVT == MVT::v32i16 ||
410 LocVT == MVT::v16i32 || LocVT == MVT::v8i64 ||
411 LocVT == MVT::v512i1) {
414 LocInfo = CCValAssign::Full;
415 } else if (LocVT == MVT::v128i8 || LocVT == MVT::v64i16 ||
416 LocVT == MVT::v32i32 || LocVT == MVT::v16i64 ||
417 (LocVT == MVT::v1024i1 && UseHVX && UseHVXDbl)) {
420 LocInfo = CCValAssign::Full;
421 } else if (LocVT == MVT::v256i8 || LocVT == MVT::v128i16 ||
422 LocVT == MVT::v64i32 || LocVT == MVT::v32i64) {
425 LocInfo = CCValAssign::Full;
427 if (LocVT == MVT::i32 || LocVT == MVT::f32) {
428 if (!RetCC_Hexagon32(ValNo, ValVT, LocVT, LocInfo, ArgFlags, State))
432 if (LocVT == MVT::i64 || LocVT == MVT::f64) {
433 if (!RetCC_Hexagon64(ValNo, ValVT, LocVT, LocInfo, ArgFlags, State))
436 if (LocVT == MVT::v16i32 || LocVT == MVT::v32i32 || LocVT == MVT::v64i32) {
437 if (!RetCC_HexagonVector(ValNo, ValVT, LocVT, LocInfo, ArgFlags, State))
440 return true; // CC didn't match.
443 static bool RetCC_Hexagon32(unsigned ValNo, MVT ValVT,
444 MVT LocVT, CCValAssign::LocInfo LocInfo,
445 ISD::ArgFlagsTy ArgFlags, CCState &State) {
447 if (LocVT == MVT::i32 || LocVT == MVT::f32) {
448 if (unsigned Reg = State.AllocateReg(Hexagon::R0)) {
449 State.addLoc(CCValAssign::getReg(ValNo, ValVT, Reg, LocVT, LocInfo));
454 unsigned Offset = State.AllocateStack(4, 4);
455 State.addLoc(CCValAssign::getMem(ValNo, ValVT, Offset, LocVT, LocInfo));
459 static bool RetCC_Hexagon64(unsigned ValNo, MVT ValVT,
460 MVT LocVT, CCValAssign::LocInfo LocInfo,
461 ISD::ArgFlagsTy ArgFlags, CCState &State) {
462 if (LocVT == MVT::i64 || LocVT == MVT::f64) {
463 if (unsigned Reg = State.AllocateReg(Hexagon::D0)) {
464 State.addLoc(CCValAssign::getReg(ValNo, ValVT, Reg, LocVT, LocInfo));
469 unsigned Offset = State.AllocateStack(8, 8);
470 State.addLoc(CCValAssign::getMem(ValNo, ValVT, Offset, LocVT, LocInfo));
474 static bool RetCC_HexagonVector(unsigned ValNo, MVT ValVT,
475 MVT LocVT, CCValAssign::LocInfo LocInfo,
476 ISD::ArgFlagsTy ArgFlags, CCState &State) {
477 auto &MF = State.getMachineFunction();
478 auto &HST = MF.getSubtarget<HexagonSubtarget>();
479 bool UseHVX = HST.useHVXOps();
480 bool UseHVXDbl = HST.useHVXDblOps();
482 unsigned OffSiz = 64;
483 if (LocVT == MVT::v16i32) {
484 if (unsigned Reg = State.AllocateReg(Hexagon::V0)) {
485 State.addLoc(CCValAssign::getReg(ValNo, ValVT, Reg, LocVT, LocInfo));
488 } else if (LocVT == MVT::v32i32) {
489 unsigned Req = (UseHVX && UseHVXDbl) ? Hexagon::V0 : Hexagon::W0;
490 if (unsigned Reg = State.AllocateReg(Req)) {
491 State.addLoc(CCValAssign::getReg(ValNo, ValVT, Reg, LocVT, LocInfo));
495 } else if (LocVT == MVT::v64i32) {
496 if (unsigned Reg = State.AllocateReg(Hexagon::W0)) {
497 State.addLoc(CCValAssign::getReg(ValNo, ValVT, Reg, LocVT, LocInfo));
503 unsigned Offset = State.AllocateStack(OffSiz, OffSiz);
504 State.addLoc(CCValAssign::getMem(ValNo, ValVT, Offset, LocVT, LocInfo));
509 HexagonTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG)
514 /// CreateCopyOfByValArgument - Make a copy of an aggregate at address specified
515 /// by "Src" to address "Dst" of size "Size". Alignment information is
516 /// specified by the specific parameter attribute. The copy will be passed as
517 /// a byval function parameter. Sometimes what we are copying is the end of a
518 /// larger object, the part that does not fit in registers.
520 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
521 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
524 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), dl, MVT::i32);
525 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
526 /*isVolatile=*/false, /*AlwaysInline=*/false,
527 /*isTailCall=*/false,
528 MachinePointerInfo(), MachinePointerInfo());
531 static bool IsHvxVectorType(MVT ty) {
532 return (ty == MVT::v8i64 || ty == MVT::v16i32 || ty == MVT::v32i16 ||
534 ty == MVT::v16i64 || ty == MVT::v32i32 || ty == MVT::v64i16 ||
536 ty == MVT::v32i64 || ty == MVT::v64i32 || ty == MVT::v128i16 ||
538 ty == MVT::v512i1 || ty == MVT::v1024i1);
541 // LowerReturn - Lower ISD::RET. If a struct is larger than 8 bytes and is
542 // passed by value, the function prototype is modified to return void and
543 // the value is stored in memory pointed by a pointer passed by caller.
545 HexagonTargetLowering::LowerReturn(SDValue Chain,
546 CallingConv::ID CallConv, bool isVarArg,
547 const SmallVectorImpl<ISD::OutputArg> &Outs,
548 const SmallVectorImpl<SDValue> &OutVals,
549 SDLoc dl, SelectionDAG &DAG) const {
551 // CCValAssign - represent the assignment of the return value to locations.
552 SmallVector<CCValAssign, 16> RVLocs;
554 // CCState - Info about the registers and stack slot.
555 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), RVLocs,
558 // Analyze return values of ISD::RET
559 CCInfo.AnalyzeReturn(Outs, RetCC_Hexagon);
562 SmallVector<SDValue, 4> RetOps(1, Chain);
564 // Copy the result values into the output registers.
565 for (unsigned i = 0; i != RVLocs.size(); ++i) {
566 CCValAssign &VA = RVLocs[i];
568 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), OutVals[i], Flag);
570 // Guarantee that all emitted copies are stuck together with flags.
571 Flag = Chain.getValue(1);
572 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
575 RetOps[0] = Chain; // Update chain.
577 // Add the flag if we have it.
579 RetOps.push_back(Flag);
581 return DAG.getNode(HexagonISD::RET_FLAG, dl, MVT::Other, RetOps);
584 bool HexagonTargetLowering::mayBeEmittedAsTailCall(CallInst *CI) const {
585 // If either no tail call or told not to tail call at all, don't.
587 CI->getParent()->getParent()->getFnAttribute("disable-tail-calls");
588 if (!CI->isTailCall() || Attr.getValueAsString() == "true")
594 /// LowerCallResult - Lower the result values of an ISD::CALL into the
595 /// appropriate copies out of appropriate physical registers. This assumes that
596 /// Chain/InFlag are the input chain/flag to use, and that TheCall is the call
597 /// being lowered. Returns a SDNode with the same number of values as the
600 HexagonTargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
601 CallingConv::ID CallConv, bool isVarArg,
603 SmallVectorImpl<ISD::InputArg> &Ins,
604 SDLoc dl, SelectionDAG &DAG,
605 SmallVectorImpl<SDValue> &InVals,
606 const SmallVectorImpl<SDValue> &OutVals,
607 SDValue Callee) const {
609 // Assign locations to each value returned by this call.
610 SmallVector<CCValAssign, 16> RVLocs;
612 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), RVLocs,
615 CCInfo.AnalyzeCallResult(Ins, RetCC_Hexagon);
617 // Copy all of the result registers out of their specified physreg.
618 for (unsigned i = 0; i != RVLocs.size(); ++i) {
619 Chain = DAG.getCopyFromReg(Chain, dl,
620 RVLocs[i].getLocReg(),
621 RVLocs[i].getValVT(), InFlag).getValue(1);
622 InFlag = Chain.getValue(2);
623 InVals.push_back(Chain.getValue(0));
629 /// LowerCall - Functions arguments are copied from virtual regs to
630 /// (physical regs)/(stack frame), CALLSEQ_START and CALLSEQ_END are emitted.
632 HexagonTargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI,
633 SmallVectorImpl<SDValue> &InVals) const {
634 SelectionDAG &DAG = CLI.DAG;
636 SmallVectorImpl<ISD::OutputArg> &Outs = CLI.Outs;
637 SmallVectorImpl<SDValue> &OutVals = CLI.OutVals;
638 SmallVectorImpl<ISD::InputArg> &Ins = CLI.Ins;
639 SDValue Chain = CLI.Chain;
640 SDValue Callee = CLI.Callee;
641 bool &isTailCall = CLI.IsTailCall;
642 CallingConv::ID CallConv = CLI.CallConv;
643 bool isVarArg = CLI.IsVarArg;
644 bool doesNotReturn = CLI.DoesNotReturn;
646 bool IsStructRet = (Outs.empty()) ? false : Outs[0].Flags.isSRet();
647 MachineFunction &MF = DAG.getMachineFunction();
648 auto PtrVT = getPointerTy(MF.getDataLayout());
650 // Check for varargs.
651 int NumNamedVarArgParams = -1;
652 if (GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(Callee))
654 const Function* CalleeFn = nullptr;
655 Callee = DAG.getTargetGlobalAddress(GA->getGlobal(), dl, MVT::i32);
656 if ((CalleeFn = dyn_cast<Function>(GA->getGlobal())))
658 // If a function has zero args and is a vararg function, that's
659 // disallowed so it must be an undeclared function. Do not assume
660 // varargs if the callee is undefined.
661 if (CalleeFn->isVarArg() &&
662 CalleeFn->getFunctionType()->getNumParams() != 0) {
663 NumNamedVarArgParams = CalleeFn->getFunctionType()->getNumParams();
668 // Analyze operands of the call, assigning locations to each operand.
669 SmallVector<CCValAssign, 16> ArgLocs;
670 HexagonCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), ArgLocs,
671 *DAG.getContext(), NumNamedVarArgParams);
674 CCInfo.AnalyzeCallOperands(Outs, CC_Hexagon_VarArg);
676 CCInfo.AnalyzeCallOperands(Outs, CC_Hexagon);
678 auto Attr = MF.getFunction()->getFnAttribute("disable-tail-calls");
679 if (Attr.getValueAsString() == "true")
683 bool StructAttrFlag = MF.getFunction()->hasStructRetAttr();
684 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv,
685 isVarArg, IsStructRet,
687 Outs, OutVals, Ins, DAG);
688 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
689 CCValAssign &VA = ArgLocs[i];
695 DEBUG(dbgs() << (isTailCall ? "Eligible for Tail Call\n"
696 : "Argument must be passed on stack. "
697 "Not eligible for Tail Call\n"));
699 // Get a count of how many bytes are to be pushed on the stack.
700 unsigned NumBytes = CCInfo.getNextStackOffset();
701 SmallVector<std::pair<unsigned, SDValue>, 16> RegsToPass;
702 SmallVector<SDValue, 8> MemOpChains;
704 auto &HRI = *Subtarget.getRegisterInfo();
706 DAG.getCopyFromReg(Chain, dl, HRI.getStackRegister(), PtrVT);
708 bool NeedsArgAlign = false;
709 unsigned LargestAlignSeen = 0;
710 // Walk the register/memloc assignments, inserting copies/loads.
711 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
712 CCValAssign &VA = ArgLocs[i];
713 SDValue Arg = OutVals[i];
714 ISD::ArgFlagsTy Flags = Outs[i].Flags;
715 // Record if we need > 8 byte alignment on an argument.
716 bool ArgAlign = IsHvxVectorType(VA.getValVT());
717 NeedsArgAlign |= ArgAlign;
719 // Promote the value if needed.
720 switch (VA.getLocInfo()) {
722 // Loc info must be one of Full, SExt, ZExt, or AExt.
723 llvm_unreachable("Unknown loc info!");
724 case CCValAssign::BCvt:
725 case CCValAssign::Full:
727 case CCValAssign::SExt:
728 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), Arg);
730 case CCValAssign::ZExt:
731 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), Arg);
733 case CCValAssign::AExt:
734 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), Arg);
739 unsigned LocMemOffset = VA.getLocMemOffset();
740 SDValue MemAddr = DAG.getConstant(LocMemOffset, dl,
741 StackPtr.getValueType());
742 MemAddr = DAG.getNode(ISD::ADD, dl, MVT::i32, StackPtr, MemAddr);
744 LargestAlignSeen = std::max(LargestAlignSeen,
745 VA.getLocVT().getStoreSizeInBits() >> 3);
746 if (Flags.isByVal()) {
747 // The argument is a struct passed by value. According to LLVM, "Arg"
749 MemOpChains.push_back(CreateCopyOfByValArgument(Arg, MemAddr, Chain,
752 MachinePointerInfo LocPI = MachinePointerInfo::getStack(
753 DAG.getMachineFunction(), LocMemOffset);
754 SDValue S = DAG.getStore(Chain, dl, Arg, MemAddr, LocPI, false,
756 MemOpChains.push_back(S);
761 // Arguments that can be passed on register must be kept at RegsToPass
764 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
767 if (NeedsArgAlign && Subtarget.hasV60TOps()) {
768 DEBUG(dbgs() << "Function needs byte stack align due to call args\n");
769 MachineFrameInfo* MFI = DAG.getMachineFunction().getFrameInfo();
770 // V6 vectors passed by value have 64 or 128 byte alignment depending
771 // on whether we are 64 byte vector mode or 128 byte.
772 bool UseHVXDbl = Subtarget.useHVXDblOps();
773 assert(Subtarget.useHVXOps());
774 const unsigned ObjAlign = UseHVXDbl ? 128 : 64;
775 LargestAlignSeen = std::max(LargestAlignSeen, ObjAlign);
776 MFI->ensureMaxAlignment(LargestAlignSeen);
778 // Transform all store nodes into one single node because all store
779 // nodes are independent of each other.
780 if (!MemOpChains.empty())
781 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOpChains);
784 SDValue C = DAG.getConstant(NumBytes, dl, PtrVT, true);
785 Chain = DAG.getCALLSEQ_START(Chain, C, dl);
788 // Build a sequence of copy-to-reg nodes chained together with token
789 // chain and flag operands which copy the outgoing args into registers.
790 // The InFlag in necessary since all emitted instructions must be
794 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
795 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
796 RegsToPass[i].second, InFlag);
797 InFlag = Chain.getValue(1);
800 // For tail calls lower the arguments to the 'real' stack slot.
802 // Force all the incoming stack arguments to be loaded from the stack
803 // before any new outgoing arguments are stored to the stack, because the
804 // outgoing stack slots may alias the incoming argument stack slots, and
805 // the alias isn't otherwise explicit. This is slightly more conservative
806 // than necessary, because it means that each store effectively depends
807 // on every argument instead of just those arguments it would clobber.
809 // Do not flag preceding copytoreg stuff together with the following stuff.
811 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
812 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
813 RegsToPass[i].second, InFlag);
814 InFlag = Chain.getValue(1);
819 // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every
820 // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol
821 // node so that legalize doesn't hack it.
822 if (flag_aligned_memcpy) {
823 const char *MemcpyName =
824 "__hexagon_memcpy_likely_aligned_min32bytes_mult8bytes";
825 Callee = DAG.getTargetExternalSymbol(MemcpyName, PtrVT);
826 flag_aligned_memcpy = false;
827 } else if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
828 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), dl, PtrVT);
829 } else if (ExternalSymbolSDNode *S =
830 dyn_cast<ExternalSymbolSDNode>(Callee)) {
831 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), PtrVT);
834 // Returns a chain & a flag for retval copy to use.
835 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
836 SmallVector<SDValue, 8> Ops;
837 Ops.push_back(Chain);
838 Ops.push_back(Callee);
840 // Add argument registers to the end of the list so that they are
841 // known live into the call.
842 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
843 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
844 RegsToPass[i].second.getValueType()));
847 if (InFlag.getNode())
848 Ops.push_back(InFlag);
851 MF.getFrameInfo()->setHasTailCall();
852 return DAG.getNode(HexagonISD::TC_RETURN, dl, NodeTys, Ops);
855 int OpCode = doesNotReturn ? HexagonISD::CALLv3nr : HexagonISD::CALLv3;
856 Chain = DAG.getNode(OpCode, dl, NodeTys, Ops);
857 InFlag = Chain.getValue(1);
859 // Create the CALLSEQ_END node.
860 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, dl, true),
861 DAG.getIntPtrConstant(0, dl, true), InFlag, dl);
862 InFlag = Chain.getValue(1);
864 // Handle result values, copying them out of physregs into vregs that we
866 return LowerCallResult(Chain, InFlag, CallConv, isVarArg, Ins, dl, DAG,
867 InVals, OutVals, Callee);
870 static bool getIndexedAddressParts(SDNode *Ptr, EVT VT,
871 bool isSEXTLoad, SDValue &Base,
872 SDValue &Offset, bool &isInc,
874 if (Ptr->getOpcode() != ISD::ADD)
877 auto &HST = static_cast<const HexagonSubtarget&>(DAG.getSubtarget());
878 bool UseHVX = HST.useHVXOps();
879 bool UseHVXDbl = HST.useHVXDblOps();
881 bool ValidHVXDblType =
882 (UseHVX && UseHVXDbl) && (VT == MVT::v32i32 || VT == MVT::v16i64 ||
883 VT == MVT::v64i16 || VT == MVT::v128i8);
885 UseHVX && !UseHVXDbl && (VT == MVT::v16i32 || VT == MVT::v8i64 ||
886 VT == MVT::v32i16 || VT == MVT::v64i8);
888 if (ValidHVXDblType || ValidHVXType ||
889 VT == MVT::i64 || VT == MVT::i32 || VT == MVT::i16 || VT == MVT::i8) {
890 isInc = (Ptr->getOpcode() == ISD::ADD);
891 Base = Ptr->getOperand(0);
892 Offset = Ptr->getOperand(1);
893 // Ensure that Offset is a constant.
894 return (isa<ConstantSDNode>(Offset));
900 /// getPostIndexedAddressParts - returns true by value, base pointer and
901 /// offset pointer and addressing mode by reference if this node can be
902 /// combined with a load / store to form a post-indexed load / store.
903 bool HexagonTargetLowering::getPostIndexedAddressParts(SDNode *N, SDNode *Op,
906 ISD::MemIndexedMode &AM,
907 SelectionDAG &DAG) const
911 bool isSEXTLoad = false;
913 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
914 VT = LD->getMemoryVT();
915 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
916 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
917 VT = ST->getMemoryVT();
918 if (ST->getValue().getValueType() == MVT::i64 && ST->isTruncatingStore()) {
926 bool isLegal = getIndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
929 auto &HII = *Subtarget.getInstrInfo();
930 int32_t OffsetVal = cast<ConstantSDNode>(Offset.getNode())->getSExtValue();
931 if (HII.isValidAutoIncImm(VT, OffsetVal)) {
932 AM = isInc ? ISD::POST_INC : ISD::POST_DEC;
940 SDValue HexagonTargetLowering::LowerINLINEASM(SDValue Op,
941 SelectionDAG &DAG) const {
942 SDNode *Node = Op.getNode();
943 MachineFunction &MF = DAG.getMachineFunction();
944 auto &FuncInfo = *MF.getInfo<HexagonMachineFunctionInfo>();
945 switch (Node->getOpcode()) {
946 case ISD::INLINEASM: {
947 unsigned NumOps = Node->getNumOperands();
948 if (Node->getOperand(NumOps-1).getValueType() == MVT::Glue)
949 --NumOps; // Ignore the flag operand.
951 for (unsigned i = InlineAsm::Op_FirstOperand; i != NumOps;) {
952 if (FuncInfo.hasClobberLR())
955 cast<ConstantSDNode>(Node->getOperand(i))->getZExtValue();
956 unsigned NumVals = InlineAsm::getNumOperandRegisters(Flags);
957 ++i; // Skip the ID value.
959 switch (InlineAsm::getKind(Flags)) {
960 default: llvm_unreachable("Bad flags!");
961 case InlineAsm::Kind_RegDef:
962 case InlineAsm::Kind_RegUse:
963 case InlineAsm::Kind_Imm:
964 case InlineAsm::Kind_Clobber:
965 case InlineAsm::Kind_Mem: {
966 for (; NumVals; --NumVals, ++i) {}
969 case InlineAsm::Kind_RegDefEarlyClobber: {
970 for (; NumVals; --NumVals, ++i) {
972 cast<RegisterSDNode>(Node->getOperand(i))->getReg();
975 const HexagonRegisterInfo *QRI = Subtarget.getRegisterInfo();
976 if (Reg == QRI->getRARegister()) {
977 FuncInfo.setHasClobberLR(true);
992 // Taken from the XCore backend.
994 SDValue HexagonTargetLowering::
995 LowerBR_JT(SDValue Op, SelectionDAG &DAG) const
997 SDValue Chain = Op.getOperand(0);
998 SDValue Table = Op.getOperand(1);
999 SDValue Index = Op.getOperand(2);
1001 JumpTableSDNode *JT = cast<JumpTableSDNode>(Table);
1002 unsigned JTI = JT->getIndex();
1003 MachineFunction &MF = DAG.getMachineFunction();
1004 const MachineJumpTableInfo *MJTI = MF.getJumpTableInfo();
1005 SDValue TargetJT = DAG.getTargetJumpTable(JT->getIndex(), MVT::i32);
1007 // Mark all jump table targets as address taken.
1008 const std::vector<MachineJumpTableEntry> &JTE = MJTI->getJumpTables();
1009 const std::vector<MachineBasicBlock*> &JTBBs = JTE[JTI].MBBs;
1010 for (unsigned i = 0, e = JTBBs.size(); i != e; ++i) {
1011 MachineBasicBlock *MBB = JTBBs[i];
1012 MBB->setHasAddressTaken();
1013 // This line is needed to set the hasAddressTaken flag on the BasicBlock
1015 BlockAddress::get(const_cast<BasicBlock *>(MBB->getBasicBlock()));
1018 SDValue JumpTableBase = DAG.getNode(
1019 HexagonISD::JT, dl, getPointerTy(DAG.getDataLayout()), TargetJT);
1020 SDValue ShiftIndex = DAG.getNode(ISD::SHL, dl, MVT::i32, Index,
1021 DAG.getConstant(2, dl, MVT::i32));
1022 SDValue JTAddress = DAG.getNode(ISD::ADD, dl, MVT::i32, JumpTableBase,
1024 SDValue LoadTarget = DAG.getLoad(MVT::i32, dl, Chain, JTAddress,
1025 MachinePointerInfo(), false, false, false,
1027 return DAG.getNode(HexagonISD::BR_JT, dl, MVT::Other, Chain, LoadTarget);
1032 HexagonTargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
1033 SelectionDAG &DAG) const {
1034 SDValue Chain = Op.getOperand(0);
1035 SDValue Size = Op.getOperand(1);
1036 SDValue Align = Op.getOperand(2);
1039 ConstantSDNode *AlignConst = dyn_cast<ConstantSDNode>(Align);
1040 assert(AlignConst && "Non-constant Align in LowerDYNAMIC_STACKALLOC");
1042 unsigned A = AlignConst->getSExtValue();
1043 auto &HFI = *Subtarget.getFrameLowering();
1044 // "Zero" means natural stack alignment.
1046 A = HFI.getStackAlignment();
1049 dbgs () << LLVM_FUNCTION_NAME << " Align: " << A << " Size: ";
1050 Size.getNode()->dump(&DAG);
1054 SDValue AC = DAG.getConstant(A, dl, MVT::i32);
1055 SDVTList VTs = DAG.getVTList(MVT::i32, MVT::Other);
1056 SDValue AA = DAG.getNode(HexagonISD::ALLOCA, dl, VTs, Chain, Size, AC);
1057 if (Op.getNode()->getHasDebugValue())
1058 DAG.TransferDbgValues(Op, AA);
1063 HexagonTargetLowering::LowerFormalArguments(SDValue Chain,
1064 CallingConv::ID CallConv,
1067 SmallVectorImpl<ISD::InputArg> &Ins,
1068 SDLoc dl, SelectionDAG &DAG,
1069 SmallVectorImpl<SDValue> &InVals)
1072 MachineFunction &MF = DAG.getMachineFunction();
1073 MachineFrameInfo *MFI = MF.getFrameInfo();
1074 MachineRegisterInfo &RegInfo = MF.getRegInfo();
1075 auto &FuncInfo = *MF.getInfo<HexagonMachineFunctionInfo>();
1077 // Assign locations to all of the incoming arguments.
1078 SmallVector<CCValAssign, 16> ArgLocs;
1079 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), ArgLocs,
1082 CCInfo.AnalyzeFormalArguments(Ins, CC_Hexagon);
1084 // For LLVM, in the case when returning a struct by value (>8byte),
1085 // the first argument is a pointer that points to the location on caller's
1086 // stack where the return value will be stored. For Hexagon, the location on
1087 // caller's stack is passed only when the struct size is smaller than (and
1088 // equal to) 8 bytes. If not, no address will be passed into callee and
1089 // callee return the result direclty through R0/R1.
1091 SmallVector<SDValue, 8> MemOps;
1092 bool UseHVX = Subtarget.useHVXOps(), UseHVXDbl = Subtarget.useHVXDblOps();
1094 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1095 CCValAssign &VA = ArgLocs[i];
1096 ISD::ArgFlagsTy Flags = Ins[i].Flags;
1098 unsigned StackLocation;
1101 if ( (VA.isRegLoc() && !Flags.isByVal())
1102 || (VA.isRegLoc() && Flags.isByVal() && Flags.getByValSize() > 8)) {
1103 // Arguments passed in registers
1104 // 1. int, long long, ptr args that get allocated in register.
1105 // 2. Large struct that gets an register to put its address in.
1106 EVT RegVT = VA.getLocVT();
1107 if (RegVT == MVT::i8 || RegVT == MVT::i16 ||
1108 RegVT == MVT::i32 || RegVT == MVT::f32) {
1110 RegInfo.createVirtualRegister(&Hexagon::IntRegsRegClass);
1111 RegInfo.addLiveIn(VA.getLocReg(), VReg);
1112 InVals.push_back(DAG.getCopyFromReg(Chain, dl, VReg, RegVT));
1113 } else if (RegVT == MVT::i64 || RegVT == MVT::f64) {
1115 RegInfo.createVirtualRegister(&Hexagon::DoubleRegsRegClass);
1116 RegInfo.addLiveIn(VA.getLocReg(), VReg);
1117 InVals.push_back(DAG.getCopyFromReg(Chain, dl, VReg, RegVT));
1120 } else if ((RegVT == MVT::v8i64 || RegVT == MVT::v16i32 ||
1121 RegVT == MVT::v32i16 || RegVT == MVT::v64i8)) {
1123 RegInfo.createVirtualRegister(&Hexagon::VectorRegsRegClass);
1124 RegInfo.addLiveIn(VA.getLocReg(), VReg);
1125 InVals.push_back(DAG.getCopyFromReg(Chain, dl, VReg, RegVT));
1126 } else if (UseHVX && UseHVXDbl &&
1127 ((RegVT == MVT::v16i64 || RegVT == MVT::v32i32 ||
1128 RegVT == MVT::v64i16 || RegVT == MVT::v128i8))) {
1130 RegInfo.createVirtualRegister(&Hexagon::VectorRegs128BRegClass);
1131 RegInfo.addLiveIn(VA.getLocReg(), VReg);
1132 InVals.push_back(DAG.getCopyFromReg(Chain, dl, VReg, RegVT));
1135 } else if ((RegVT == MVT::v16i64 || RegVT == MVT::v32i32 ||
1136 RegVT == MVT::v64i16 || RegVT == MVT::v128i8)) {
1138 RegInfo.createVirtualRegister(&Hexagon::VecDblRegsRegClass);
1139 RegInfo.addLiveIn(VA.getLocReg(), VReg);
1140 InVals.push_back(DAG.getCopyFromReg(Chain, dl, VReg, RegVT));
1141 } else if (UseHVX && UseHVXDbl &&
1142 ((RegVT == MVT::v32i64 || RegVT == MVT::v64i32 ||
1143 RegVT == MVT::v128i16 || RegVT == MVT::v256i8))) {
1145 RegInfo.createVirtualRegister(&Hexagon::VecDblRegs128BRegClass);
1146 RegInfo.addLiveIn(VA.getLocReg(), VReg);
1147 InVals.push_back(DAG.getCopyFromReg(Chain, dl, VReg, RegVT));
1148 } else if (RegVT == MVT::v512i1 || RegVT == MVT::v1024i1) {
1149 assert(0 && "need to support VecPred regs");
1151 RegInfo.createVirtualRegister(&Hexagon::VecPredRegsRegClass);
1152 RegInfo.addLiveIn(VA.getLocReg(), VReg);
1153 InVals.push_back(DAG.getCopyFromReg(Chain, dl, VReg, RegVT));
1157 } else if (VA.isRegLoc() && Flags.isByVal() && Flags.getByValSize() <= 8) {
1158 assert (0 && "ByValSize must be bigger than 8 bytes");
1161 assert(VA.isMemLoc());
1163 if (Flags.isByVal()) {
1164 // If it's a byval parameter, then we need to compute the
1165 // "real" size, not the size of the pointer.
1166 ObjSize = Flags.getByValSize();
1168 ObjSize = VA.getLocVT().getStoreSizeInBits() >> 3;
1171 StackLocation = HEXAGON_LRFP_SIZE + VA.getLocMemOffset();
1172 // Create the frame index object for this incoming parameter...
1173 FI = MFI->CreateFixedObject(ObjSize, StackLocation, true);
1175 // Create the SelectionDAG nodes cordl, responding to a load
1176 // from this parameter.
1177 SDValue FIN = DAG.getFrameIndex(FI, MVT::i32);
1179 if (Flags.isByVal()) {
1180 // If it's a pass-by-value aggregate, then do not dereference the stack
1181 // location. Instead, we should generate a reference to the stack
1183 InVals.push_back(FIN);
1185 InVals.push_back(DAG.getLoad(VA.getLocVT(), dl, Chain, FIN,
1186 MachinePointerInfo(), false, false,
1192 if (!MemOps.empty())
1193 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOps);
1196 // This will point to the next argument passed via stack.
1197 int FrameIndex = MFI->CreateFixedObject(Hexagon_PointerSize,
1199 CCInfo.getNextStackOffset(),
1201 FuncInfo.setVarArgsFrameIndex(FrameIndex);
1208 HexagonTargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG) const {
1209 // VASTART stores the address of the VarArgsFrameIndex slot into the
1210 // memory location argument.
1211 MachineFunction &MF = DAG.getMachineFunction();
1212 HexagonMachineFunctionInfo *QFI = MF.getInfo<HexagonMachineFunctionInfo>();
1213 SDValue Addr = DAG.getFrameIndex(QFI->getVarArgsFrameIndex(), MVT::i32);
1214 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
1215 return DAG.getStore(Op.getOperand(0), SDLoc(Op), Addr,
1216 Op.getOperand(1), MachinePointerInfo(SV), false,
1220 // Creates a SPLAT instruction for a constant value VAL.
1221 static SDValue createSplat(SelectionDAG &DAG, SDLoc dl, EVT VT, SDValue Val) {
1222 if (VT.getSimpleVT() == MVT::v4i8)
1223 return DAG.getNode(HexagonISD::VSPLATB, dl, VT, Val);
1225 if (VT.getSimpleVT() == MVT::v4i16)
1226 return DAG.getNode(HexagonISD::VSPLATH, dl, VT, Val);
1231 static bool isSExtFree(SDValue N) {
1232 // A sign-extend of a truncate of a sign-extend is free.
1233 if (N.getOpcode() == ISD::TRUNCATE &&
1234 N.getOperand(0).getOpcode() == ISD::AssertSext)
1236 // We have sign-extended loads.
1237 if (N.getOpcode() == ISD::LOAD)
1242 SDValue HexagonTargetLowering::LowerCTPOP(SDValue Op, SelectionDAG &DAG) const {
1244 SDValue InpVal = Op.getOperand(0);
1245 if (isa<ConstantSDNode>(InpVal)) {
1246 uint64_t V = cast<ConstantSDNode>(InpVal)->getZExtValue();
1247 return DAG.getTargetConstant(countPopulation(V), dl, MVT::i64);
1249 SDValue PopOut = DAG.getNode(HexagonISD::POPCOUNT, dl, MVT::i32, InpVal);
1250 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i64, PopOut);
1253 SDValue HexagonTargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) const {
1256 SDValue LHS = Op.getOperand(0);
1257 SDValue RHS = Op.getOperand(1);
1258 SDValue Cmp = Op.getOperand(2);
1259 ISD::CondCode CC = cast<CondCodeSDNode>(Cmp)->get();
1261 EVT VT = Op.getValueType();
1262 EVT LHSVT = LHS.getValueType();
1263 EVT RHSVT = RHS.getValueType();
1265 if (LHSVT == MVT::v2i16) {
1266 assert(ISD::isSignedIntSetCC(CC) || ISD::isUnsignedIntSetCC(CC));
1267 unsigned ExtOpc = ISD::isSignedIntSetCC(CC) ? ISD::SIGN_EXTEND
1269 SDValue LX = DAG.getNode(ExtOpc, dl, MVT::v2i32, LHS);
1270 SDValue RX = DAG.getNode(ExtOpc, dl, MVT::v2i32, RHS);
1271 SDValue SC = DAG.getNode(ISD::SETCC, dl, MVT::v2i1, LX, RX, Cmp);
1275 // Treat all other vector types as legal.
1279 // Equals and not equals should use sign-extend, not zero-extend, since
1280 // we can represent small negative values in the compare instructions.
1281 // The LLVM default is to use zero-extend arbitrarily in these cases.
1282 if ((CC == ISD::SETEQ || CC == ISD::SETNE) &&
1283 (RHSVT == MVT::i8 || RHSVT == MVT::i16) &&
1284 (LHSVT == MVT::i8 || LHSVT == MVT::i16)) {
1285 ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS);
1286 if (C && C->getAPIntValue().isNegative()) {
1287 LHS = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::i32, LHS);
1288 RHS = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::i32, RHS);
1289 return DAG.getNode(ISD::SETCC, dl, Op.getValueType(),
1290 LHS, RHS, Op.getOperand(2));
1292 if (isSExtFree(LHS) || isSExtFree(RHS)) {
1293 LHS = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::i32, LHS);
1294 RHS = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::i32, RHS);
1295 return DAG.getNode(ISD::SETCC, dl, Op.getValueType(),
1296 LHS, RHS, Op.getOperand(2));
1302 SDValue HexagonTargetLowering::LowerVSELECT(SDValue Op, SelectionDAG &DAG)
1304 SDValue PredOp = Op.getOperand(0);
1305 SDValue Op1 = Op.getOperand(1), Op2 = Op.getOperand(2);
1306 EVT OpVT = Op1.getValueType();
1309 if (OpVT == MVT::v2i16) {
1310 SDValue X1 = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::v2i32, Op1);
1311 SDValue X2 = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::v2i32, Op2);
1312 SDValue SL = DAG.getNode(ISD::VSELECT, DL, MVT::v2i32, PredOp, X1, X2);
1313 SDValue TR = DAG.getNode(ISD::TRUNCATE, DL, MVT::v2i16, SL);
1320 // Handle only specific vector loads.
1321 SDValue HexagonTargetLowering::LowerLOAD(SDValue Op, SelectionDAG &DAG) const {
1322 EVT VT = Op.getValueType();
1324 LoadSDNode *LoadNode = cast<LoadSDNode>(Op);
1325 SDValue Chain = LoadNode->getChain();
1326 SDValue Ptr = Op.getOperand(1);
1327 SDValue LoweredLoad;
1329 SDValue Base = LoadNode->getBasePtr();
1330 ISD::LoadExtType Ext = LoadNode->getExtensionType();
1331 unsigned Alignment = LoadNode->getAlignment();
1334 if(Ext == ISD::NON_EXTLOAD)
1335 Ext = ISD::ZEXTLOAD;
1337 if (VT == MVT::v4i16) {
1338 if (Alignment == 2) {
1341 Loads[0] = DAG.getExtLoad(Ext, DL, MVT::i32, Chain, Base,
1342 LoadNode->getPointerInfo(), MVT::i16,
1343 LoadNode->isVolatile(),
1344 LoadNode->isNonTemporal(),
1345 LoadNode->isInvariant(),
1348 SDValue Increment = DAG.getConstant(2, DL, MVT::i32);
1349 Ptr = DAG.getNode(ISD::ADD, DL, Base.getValueType(), Base, Increment);
1350 Loads[1] = DAG.getExtLoad(Ext, DL, MVT::i32, Chain, Ptr,
1351 LoadNode->getPointerInfo(), MVT::i16,
1352 LoadNode->isVolatile(),
1353 LoadNode->isNonTemporal(),
1354 LoadNode->isInvariant(),
1356 // SHL 16, then OR base and base+2.
1357 SDValue ShiftAmount = DAG.getConstant(16, DL, MVT::i32);
1358 SDValue Tmp1 = DAG.getNode(ISD::SHL, DL, MVT::i32, Loads[1], ShiftAmount);
1359 SDValue Tmp2 = DAG.getNode(ISD::OR, DL, MVT::i32, Tmp1, Loads[0]);
1361 Increment = DAG.getConstant(4, DL, MVT::i32);
1362 Ptr = DAG.getNode(ISD::ADD, DL, Base.getValueType(), Base, Increment);
1363 Loads[2] = DAG.getExtLoad(Ext, DL, MVT::i32, Chain, Ptr,
1364 LoadNode->getPointerInfo(), MVT::i16,
1365 LoadNode->isVolatile(),
1366 LoadNode->isNonTemporal(),
1367 LoadNode->isInvariant(),
1370 Increment = DAG.getConstant(6, DL, MVT::i32);
1371 Ptr = DAG.getNode(ISD::ADD, DL, Base.getValueType(), Base, Increment);
1372 Loads[3] = DAG.getExtLoad(Ext, DL, MVT::i32, Chain, Ptr,
1373 LoadNode->getPointerInfo(), MVT::i16,
1374 LoadNode->isVolatile(),
1375 LoadNode->isNonTemporal(),
1376 LoadNode->isInvariant(),
1378 // SHL 16, then OR base+4 and base+6.
1379 Tmp1 = DAG.getNode(ISD::SHL, DL, MVT::i32, Loads[3], ShiftAmount);
1380 SDValue Tmp4 = DAG.getNode(ISD::OR, DL, MVT::i32, Tmp1, Loads[2]);
1381 // Combine to i64. This could be optimised out later if we can
1382 // affect reg allocation of this code.
1383 Result = DAG.getNode(HexagonISD::COMBINE, DL, MVT::i64, Tmp4, Tmp2);
1384 LoadChain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other,
1385 Loads[0].getValue(1), Loads[1].getValue(1),
1386 Loads[2].getValue(1), Loads[3].getValue(1));
1388 // Perform default type expansion.
1389 Result = DAG.getLoad(MVT::i64, DL, Chain, Ptr, LoadNode->getPointerInfo(),
1390 LoadNode->isVolatile(), LoadNode->isNonTemporal(),
1391 LoadNode->isInvariant(), LoadNode->getAlignment());
1392 LoadChain = Result.getValue(1);
1395 llvm_unreachable("Custom lowering unsupported load");
1397 Result = DAG.getNode(ISD::BITCAST, DL, VT, Result);
1398 // Since we pretend to lower a load, we need the original chain
1399 // info attached to the result.
1400 SDValue Ops[] = { Result, LoadChain };
1402 return DAG.getMergeValues(Ops, DL);
1407 HexagonTargetLowering::LowerConstantPool(SDValue Op, SelectionDAG &DAG) const {
1408 EVT ValTy = Op.getValueType();
1410 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
1412 if (CP->isMachineConstantPoolEntry())
1413 Res = DAG.getTargetConstantPool(CP->getMachineCPVal(), ValTy,
1414 CP->getAlignment());
1416 Res = DAG.getTargetConstantPool(CP->getConstVal(), ValTy,
1417 CP->getAlignment());
1418 return DAG.getNode(HexagonISD::CP, dl, ValTy, Res);
1422 HexagonTargetLowering::LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) const {
1423 const HexagonRegisterInfo &HRI = *Subtarget.getRegisterInfo();
1424 MachineFunction &MF = DAG.getMachineFunction();
1425 MachineFrameInfo &MFI = *MF.getFrameInfo();
1426 MFI.setReturnAddressIsTaken(true);
1428 if (verifyReturnAddressArgumentIsConstant(Op, DAG))
1431 EVT VT = Op.getValueType();
1433 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
1435 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
1436 SDValue Offset = DAG.getConstant(4, dl, MVT::i32);
1437 return DAG.getLoad(VT, dl, DAG.getEntryNode(),
1438 DAG.getNode(ISD::ADD, dl, VT, FrameAddr, Offset),
1439 MachinePointerInfo(), false, false, false, 0);
1442 // Return LR, which contains the return address. Mark it an implicit live-in.
1443 unsigned Reg = MF.addLiveIn(HRI.getRARegister(), getRegClassFor(MVT::i32));
1444 return DAG.getCopyFromReg(DAG.getEntryNode(), dl, Reg, VT);
1448 HexagonTargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
1449 const HexagonRegisterInfo &HRI = *Subtarget.getRegisterInfo();
1450 MachineFrameInfo &MFI = *DAG.getMachineFunction().getFrameInfo();
1451 MFI.setFrameAddressIsTaken(true);
1453 EVT VT = Op.getValueType();
1455 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
1456 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl,
1457 HRI.getFrameRegister(), VT);
1459 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr,
1460 MachinePointerInfo(),
1461 false, false, false, 0);
1465 SDValue HexagonTargetLowering::LowerATOMIC_FENCE(SDValue Op,
1466 SelectionDAG& DAG) const {
1468 return DAG.getNode(HexagonISD::BARRIER, dl, MVT::Other, Op.getOperand(0));
1472 SDValue HexagonTargetLowering::LowerGLOBALADDRESS(SDValue Op,
1473 SelectionDAG &DAG) const {
1475 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
1476 int64_t Offset = cast<GlobalAddressSDNode>(Op)->getOffset();
1478 auto PtrVT = getPointerTy(DAG.getDataLayout());
1479 Result = DAG.getTargetGlobalAddress(GV, dl, PtrVT, Offset);
1481 const HexagonTargetObjectFile *TLOF =
1482 static_cast<const HexagonTargetObjectFile *>(
1483 getTargetMachine().getObjFileLowering());
1484 if (TLOF->IsGlobalInSmallSection(GV, getTargetMachine())) {
1485 return DAG.getNode(HexagonISD::CONST32_GP, dl, PtrVT, Result);
1488 return DAG.getNode(HexagonISD::CONST32, dl, PtrVT, Result);
1491 // Specifies that for loads and stores VT can be promoted to PromotedLdStVT.
1492 void HexagonTargetLowering::promoteLdStType(EVT VT, EVT PromotedLdStVT) {
1493 if (VT != PromotedLdStVT) {
1494 setOperationAction(ISD::LOAD, VT.getSimpleVT(), Promote);
1495 AddPromotedToType(ISD::LOAD, VT.getSimpleVT(),
1496 PromotedLdStVT.getSimpleVT());
1498 setOperationAction(ISD::STORE, VT.getSimpleVT(), Promote);
1499 AddPromotedToType(ISD::STORE, VT.getSimpleVT(),
1500 PromotedLdStVT.getSimpleVT());
1505 HexagonTargetLowering::LowerBlockAddress(SDValue Op, SelectionDAG &DAG) const {
1506 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
1507 SDValue BA_SD = DAG.getTargetBlockAddress(BA, MVT::i32);
1509 return DAG.getNode(HexagonISD::CONST32_GP, dl,
1510 getPointerTy(DAG.getDataLayout()), BA_SD);
1513 //===----------------------------------------------------------------------===//
1514 // TargetLowering Implementation
1515 //===----------------------------------------------------------------------===//
1517 HexagonTargetLowering::HexagonTargetLowering(const TargetMachine &TM,
1518 const HexagonSubtarget &ST)
1519 : TargetLowering(TM), HTM(static_cast<const HexagonTargetMachine&>(TM)),
1521 bool IsV4 = !Subtarget.hasV5TOps();
1522 auto &HRI = *Subtarget.getRegisterInfo();
1523 bool UseHVX = Subtarget.useHVXOps(), UseHVXDbl = Subtarget.useHVXDblOps();
1525 setPrefLoopAlignment(4);
1526 setPrefFunctionAlignment(4);
1527 setMinFunctionAlignment(2);
1528 setInsertFencesForAtomic(false);
1529 setStackPointerRegisterToSaveRestore(HRI.getStackRegister());
1531 if (EnableHexSDNodeSched)
1532 setSchedulingPreference(Sched::VLIW);
1534 setSchedulingPreference(Sched::Source);
1536 // Limits for inline expansion of memcpy/memmove
1537 MaxStoresPerMemcpy = MaxStoresPerMemcpyCL;
1538 MaxStoresPerMemcpyOptSize = MaxStoresPerMemcpyOptSizeCL;
1539 MaxStoresPerMemmove = MaxStoresPerMemmoveCL;
1540 MaxStoresPerMemmoveOptSize = MaxStoresPerMemmoveOptSizeCL;
1541 MaxStoresPerMemset = MaxStoresPerMemsetCL;
1542 MaxStoresPerMemsetOptSize = MaxStoresPerMemsetOptSizeCL;
1545 // Set up register classes.
1548 addRegisterClass(MVT::i1, &Hexagon::PredRegsRegClass);
1549 addRegisterClass(MVT::v2i1, &Hexagon::PredRegsRegClass); // bbbbaaaa
1550 addRegisterClass(MVT::v4i1, &Hexagon::PredRegsRegClass); // ddccbbaa
1551 addRegisterClass(MVT::v8i1, &Hexagon::PredRegsRegClass); // hgfedcba
1552 addRegisterClass(MVT::i32, &Hexagon::IntRegsRegClass);
1553 addRegisterClass(MVT::v4i8, &Hexagon::IntRegsRegClass);
1554 addRegisterClass(MVT::v2i16, &Hexagon::IntRegsRegClass);
1555 addRegisterClass(MVT::i64, &Hexagon::DoubleRegsRegClass);
1556 addRegisterClass(MVT::v8i8, &Hexagon::DoubleRegsRegClass);
1557 addRegisterClass(MVT::v4i16, &Hexagon::DoubleRegsRegClass);
1558 addRegisterClass(MVT::v2i32, &Hexagon::DoubleRegsRegClass);
1560 if (Subtarget.hasV5TOps()) {
1561 addRegisterClass(MVT::f32, &Hexagon::IntRegsRegClass);
1562 addRegisterClass(MVT::f64, &Hexagon::DoubleRegsRegClass);
1565 if (Subtarget.hasV60TOps()) {
1566 if (Subtarget.useHVXSglOps()) {
1567 addRegisterClass(MVT::v64i8, &Hexagon::VectorRegsRegClass);
1568 addRegisterClass(MVT::v32i16, &Hexagon::VectorRegsRegClass);
1569 addRegisterClass(MVT::v16i32, &Hexagon::VectorRegsRegClass);
1570 addRegisterClass(MVT::v8i64, &Hexagon::VectorRegsRegClass);
1571 addRegisterClass(MVT::v128i8, &Hexagon::VecDblRegsRegClass);
1572 addRegisterClass(MVT::v64i16, &Hexagon::VecDblRegsRegClass);
1573 addRegisterClass(MVT::v32i32, &Hexagon::VecDblRegsRegClass);
1574 addRegisterClass(MVT::v16i64, &Hexagon::VecDblRegsRegClass);
1575 addRegisterClass(MVT::v512i1, &Hexagon::VecPredRegsRegClass);
1576 } else if (Subtarget.useHVXDblOps()) {
1577 addRegisterClass(MVT::v128i8, &Hexagon::VectorRegs128BRegClass);
1578 addRegisterClass(MVT::v64i16, &Hexagon::VectorRegs128BRegClass);
1579 addRegisterClass(MVT::v32i32, &Hexagon::VectorRegs128BRegClass);
1580 addRegisterClass(MVT::v16i64, &Hexagon::VectorRegs128BRegClass);
1581 addRegisterClass(MVT::v256i8, &Hexagon::VecDblRegs128BRegClass);
1582 addRegisterClass(MVT::v128i16, &Hexagon::VecDblRegs128BRegClass);
1583 addRegisterClass(MVT::v64i32, &Hexagon::VecDblRegs128BRegClass);
1584 addRegisterClass(MVT::v32i64, &Hexagon::VecDblRegs128BRegClass);
1585 addRegisterClass(MVT::v1024i1, &Hexagon::VecPredRegs128BRegClass);
1591 // Handling of scalar operations.
1593 // All operations default to "legal", except:
1594 // - indexed loads and stores (pre-/post-incremented),
1595 // - ANY_EXTEND_VECTOR_INREG, ATOMIC_CMP_SWAP_WITH_SUCCESS, CONCAT_VECTORS,
1596 // ConstantFP, DEBUGTRAP, FCEIL, FCOPYSIGN, FEXP, FEXP2, FFLOOR, FGETSIGN,
1597 // FLOG, FLOG2, FLOG10, FMAXNUM, FMINNUM, FNEARBYINT, FRINT, FROUND, TRAP,
1598 // FTRUNC, PREFETCH, SIGN_EXTEND_VECTOR_INREG, ZERO_EXTEND_VECTOR_INREG,
1599 // which default to "expand" for at least one type.
1602 setOperationAction(ISD::ConstantFP, MVT::f32, Legal); // Default: expand
1603 setOperationAction(ISD::ConstantFP, MVT::f64, Legal); // Default: expand
1605 setOperationAction(ISD::ConstantPool, MVT::i32, Custom);
1606 setOperationAction(ISD::BUILD_PAIR, MVT::i64, Expand);
1607 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand);
1608 setOperationAction(ISD::INLINEASM, MVT::Other, Custom);
1609 setOperationAction(ISD::EH_RETURN, MVT::Other, Custom);
1610 setOperationAction(ISD::ATOMIC_FENCE, MVT::Other, Custom);
1612 // Custom legalize GlobalAddress nodes into CONST32.
1613 setOperationAction(ISD::GlobalAddress, MVT::i32, Custom);
1614 setOperationAction(ISD::GlobalAddress, MVT::i8, Custom);
1615 setOperationAction(ISD::BlockAddress, MVT::i32, Custom);
1617 // Hexagon needs to optimize cases with negative constants.
1618 setOperationAction(ISD::SETCC, MVT::i8, Custom);
1619 setOperationAction(ISD::SETCC, MVT::i16, Custom);
1621 // VASTART needs to be custom lowered to use the VarArgsFrameIndex.
1622 setOperationAction(ISD::VASTART, MVT::Other, Custom);
1623 setOperationAction(ISD::VAEND, MVT::Other, Expand);
1624 setOperationAction(ISD::VAARG, MVT::Other, Expand);
1626 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
1627 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
1628 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Custom);
1631 setOperationAction(ISD::BR_JT, MVT::Other, Custom);
1633 setOperationAction(ISD::BR_JT, MVT::Other, Expand);
1634 // Increase jump tables cutover to 5, was 4.
1635 setMinimumJumpTableEntries(MinimumJumpTables);
1637 // Hexagon has instructions for add/sub with carry. The problem with
1638 // modeling these instructions is that they produce 2 results: Rdd and Px.
1639 // To model the update of Px, we will have to use Defs[p0..p3] which will
1640 // cause any predicate live range to spill. So, we pretend we dont't have
1641 // these instructions.
1642 setOperationAction(ISD::ADDE, MVT::i8, Expand);
1643 setOperationAction(ISD::ADDE, MVT::i16, Expand);
1644 setOperationAction(ISD::ADDE, MVT::i32, Expand);
1645 setOperationAction(ISD::ADDE, MVT::i64, Expand);
1646 setOperationAction(ISD::SUBE, MVT::i8, Expand);
1647 setOperationAction(ISD::SUBE, MVT::i16, Expand);
1648 setOperationAction(ISD::SUBE, MVT::i32, Expand);
1649 setOperationAction(ISD::SUBE, MVT::i64, Expand);
1650 setOperationAction(ISD::ADDC, MVT::i8, Expand);
1651 setOperationAction(ISD::ADDC, MVT::i16, Expand);
1652 setOperationAction(ISD::ADDC, MVT::i32, Expand);
1653 setOperationAction(ISD::ADDC, MVT::i64, Expand);
1654 setOperationAction(ISD::SUBC, MVT::i8, Expand);
1655 setOperationAction(ISD::SUBC, MVT::i16, Expand);
1656 setOperationAction(ISD::SUBC, MVT::i32, Expand);
1657 setOperationAction(ISD::SUBC, MVT::i64, Expand);
1659 // Only add and sub that detect overflow are the saturating ones.
1660 for (MVT VT : MVT::integer_valuetypes()) {
1661 setOperationAction(ISD::UADDO, VT, Expand);
1662 setOperationAction(ISD::SADDO, VT, Expand);
1663 setOperationAction(ISD::USUBO, VT, Expand);
1664 setOperationAction(ISD::SSUBO, VT, Expand);
1667 setOperationAction(ISD::CTLZ, MVT::i8, Promote);
1668 setOperationAction(ISD::CTLZ, MVT::i16, Promote);
1669 setOperationAction(ISD::CTTZ, MVT::i8, Promote);
1670 setOperationAction(ISD::CTTZ, MVT::i16, Promote);
1671 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i8, Promote);
1672 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i16, Promote);
1673 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i8, Promote);
1674 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i16, Promote);
1676 // In V5, popcount can count # of 1s in i64 but returns i32.
1677 // On V4 it will be expanded (set later).
1678 setOperationAction(ISD::CTPOP, MVT::i8, Promote);
1679 setOperationAction(ISD::CTPOP, MVT::i16, Promote);
1680 setOperationAction(ISD::CTPOP, MVT::i32, Promote);
1681 setOperationAction(ISD::CTPOP, MVT::i64, Custom);
1683 // We custom lower i64 to i64 mul, so that it is not considered as a legal
1684 // operation. There is a pattern that will match i64 mul and transform it
1685 // to a series of instructions.
1686 setOperationAction(ISD::MUL, MVT::i64, Expand);
1687 setOperationAction(ISD::MULHS, MVT::i64, Expand);
1689 for (unsigned IntExpOp :
1690 { ISD::SDIV, ISD::UDIV, ISD::SREM, ISD::UREM,
1691 ISD::SDIVREM, ISD::UDIVREM, ISD::ROTL, ISD::ROTR,
1692 ISD::BSWAP, ISD::SHL_PARTS, ISD::SRA_PARTS, ISD::SRL_PARTS,
1693 ISD::SMUL_LOHI, ISD::UMUL_LOHI }) {
1694 setOperationAction(IntExpOp, MVT::i32, Expand);
1695 setOperationAction(IntExpOp, MVT::i64, Expand);
1698 for (unsigned FPExpOp :
1699 {ISD::FDIV, ISD::FREM, ISD::FSQRT, ISD::FSIN, ISD::FCOS, ISD::FSINCOS,
1700 ISD::FPOW, ISD::FCOPYSIGN}) {
1701 setOperationAction(FPExpOp, MVT::f32, Expand);
1702 setOperationAction(FPExpOp, MVT::f64, Expand);
1705 // No extending loads from i32.
1706 for (MVT VT : MVT::integer_valuetypes()) {
1707 setLoadExtAction(ISD::ZEXTLOAD, VT, MVT::i32, Expand);
1708 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::i32, Expand);
1709 setLoadExtAction(ISD::EXTLOAD, VT, MVT::i32, Expand);
1711 // Turn FP truncstore into trunc + store.
1712 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
1713 // Turn FP extload into load/fextend.
1714 for (MVT VT : MVT::fp_valuetypes())
1715 setLoadExtAction(ISD::EXTLOAD, VT, MVT::f32, Expand);
1717 // Expand BR_CC and SELECT_CC for all integer and fp types.
1718 for (MVT VT : MVT::integer_valuetypes()) {
1719 setOperationAction(ISD::BR_CC, VT, Expand);
1720 setOperationAction(ISD::SELECT_CC, VT, Expand);
1722 for (MVT VT : MVT::fp_valuetypes()) {
1723 setOperationAction(ISD::BR_CC, VT, Expand);
1724 setOperationAction(ISD::SELECT_CC, VT, Expand);
1726 setOperationAction(ISD::BR_CC, MVT::Other, Expand);
1729 // Handling of vector operations.
1732 // Custom lower v4i16 load only. Let v4i16 store to be
1733 // promoted for now.
1734 promoteLdStType(MVT::v4i8, MVT::i32);
1735 promoteLdStType(MVT::v2i16, MVT::i32);
1736 promoteLdStType(MVT::v8i8, MVT::i64);
1737 promoteLdStType(MVT::v2i32, MVT::i64);
1739 setOperationAction(ISD::LOAD, MVT::v4i16, Custom);
1740 setOperationAction(ISD::STORE, MVT::v4i16, Promote);
1741 AddPromotedToType(ISD::LOAD, MVT::v4i16, MVT::i64);
1742 AddPromotedToType(ISD::STORE, MVT::v4i16, MVT::i64);
1744 // Set the action for vector operations to "expand", then override it with
1745 // either "custom" or "legal" for specific cases.
1746 static const unsigned VectExpOps[] = {
1747 // Integer arithmetic:
1748 ISD::ADD, ISD::SUB, ISD::MUL, ISD::SDIV, ISD::UDIV,
1749 ISD::SREM, ISD::UREM, ISD::SDIVREM, ISD::UDIVREM, ISD::ADDC,
1750 ISD::SUBC, ISD::SADDO, ISD::UADDO, ISD::SSUBO, ISD::USUBO,
1751 ISD::SMUL_LOHI, ISD::UMUL_LOHI,
1753 ISD::AND, ISD::OR, ISD::XOR, ISD::ROTL, ISD::ROTR,
1754 ISD::CTPOP, ISD::CTLZ, ISD::CTTZ, ISD::CTLZ_ZERO_UNDEF,
1755 ISD::CTTZ_ZERO_UNDEF,
1756 // Floating point arithmetic/math functions:
1757 ISD::FADD, ISD::FSUB, ISD::FMUL, ISD::FMA, ISD::FDIV,
1758 ISD::FREM, ISD::FNEG, ISD::FABS, ISD::FSQRT, ISD::FSIN,
1759 ISD::FCOS, ISD::FPOWI, ISD::FPOW, ISD::FLOG, ISD::FLOG2,
1760 ISD::FLOG10, ISD::FEXP, ISD::FEXP2, ISD::FCEIL, ISD::FTRUNC,
1761 ISD::FRINT, ISD::FNEARBYINT, ISD::FROUND, ISD::FFLOOR,
1762 ISD::FMINNUM, ISD::FMAXNUM, ISD::FSINCOS,
1764 ISD::SELECT, ISD::ConstantPool,
1766 ISD::BUILD_VECTOR, ISD::SCALAR_TO_VECTOR,
1767 ISD::EXTRACT_VECTOR_ELT, ISD::INSERT_VECTOR_ELT,
1768 ISD::EXTRACT_SUBVECTOR, ISD::INSERT_SUBVECTOR,
1769 ISD::CONCAT_VECTORS, ISD::VECTOR_SHUFFLE
1772 for (MVT VT : MVT::vector_valuetypes()) {
1773 for (unsigned VectExpOp : VectExpOps)
1774 setOperationAction(VectExpOp, VT, Expand);
1776 // Expand all extended loads and truncating stores:
1777 for (MVT TargetVT : MVT::vector_valuetypes()) {
1778 setLoadExtAction(ISD::EXTLOAD, TargetVT, VT, Expand);
1779 setTruncStoreAction(VT, TargetVT, Expand);
1782 setOperationAction(ISD::SRA, VT, Custom);
1783 setOperationAction(ISD::SHL, VT, Custom);
1784 setOperationAction(ISD::SRL, VT, Custom);
1787 // Types natively supported:
1788 for (MVT NativeVT : {MVT::v2i1, MVT::v4i1, MVT::v8i1, MVT::v32i1, MVT::v64i1,
1789 MVT::v4i8, MVT::v8i8, MVT::v2i16, MVT::v4i16, MVT::v1i32,
1790 MVT::v2i32, MVT::v1i64}) {
1791 setOperationAction(ISD::BUILD_VECTOR, NativeVT, Custom);
1792 setOperationAction(ISD::EXTRACT_VECTOR_ELT, NativeVT, Custom);
1793 setOperationAction(ISD::INSERT_VECTOR_ELT, NativeVT, Custom);
1794 setOperationAction(ISD::EXTRACT_SUBVECTOR, NativeVT, Custom);
1795 setOperationAction(ISD::INSERT_SUBVECTOR, NativeVT, Custom);
1796 setOperationAction(ISD::CONCAT_VECTORS, NativeVT, Custom);
1798 setOperationAction(ISD::ADD, NativeVT, Legal);
1799 setOperationAction(ISD::SUB, NativeVT, Legal);
1800 setOperationAction(ISD::MUL, NativeVT, Legal);
1801 setOperationAction(ISD::AND, NativeVT, Legal);
1802 setOperationAction(ISD::OR, NativeVT, Legal);
1803 setOperationAction(ISD::XOR, NativeVT, Legal);
1806 setOperationAction(ISD::SETCC, MVT::v2i16, Custom);
1807 setOperationAction(ISD::VSELECT, MVT::v2i16, Custom);
1808 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4i16, Custom);
1809 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v8i8, Custom);
1812 setOperationAction(ISD::CONCAT_VECTORS, MVT::v128i8, Custom);
1813 setOperationAction(ISD::CONCAT_VECTORS, MVT::v64i16, Custom);
1814 setOperationAction(ISD::CONCAT_VECTORS, MVT::v32i32, Custom);
1815 setOperationAction(ISD::CONCAT_VECTORS, MVT::v16i64, Custom);
1818 setOperationAction(ISD::CONCAT_VECTORS, MVT::v256i8, Custom);
1819 setOperationAction(ISD::CONCAT_VECTORS, MVT::v128i16, Custom);
1820 setOperationAction(ISD::CONCAT_VECTORS, MVT::v64i32, Custom);
1821 setOperationAction(ISD::CONCAT_VECTORS, MVT::v32i64, Custom);
1824 // Subtarget-specific operation actions.
1826 if (Subtarget.hasV5TOps()) {
1827 setOperationAction(ISD::FMA, MVT::f64, Expand);
1828 setOperationAction(ISD::FADD, MVT::f64, Expand);
1829 setOperationAction(ISD::FSUB, MVT::f64, Expand);
1830 setOperationAction(ISD::FMUL, MVT::f64, Expand);
1832 setOperationAction(ISD::FP_TO_UINT, MVT::i1, Promote);
1833 setOperationAction(ISD::FP_TO_UINT, MVT::i8, Promote);
1834 setOperationAction(ISD::FP_TO_UINT, MVT::i16, Promote);
1835 setOperationAction(ISD::FP_TO_SINT, MVT::i1, Promote);
1836 setOperationAction(ISD::FP_TO_SINT, MVT::i8, Promote);
1837 setOperationAction(ISD::FP_TO_SINT, MVT::i16, Promote);
1838 setOperationAction(ISD::UINT_TO_FP, MVT::i1, Promote);
1839 setOperationAction(ISD::UINT_TO_FP, MVT::i8, Promote);
1840 setOperationAction(ISD::UINT_TO_FP, MVT::i16, Promote);
1841 setOperationAction(ISD::SINT_TO_FP, MVT::i1, Promote);
1842 setOperationAction(ISD::SINT_TO_FP, MVT::i8, Promote);
1843 setOperationAction(ISD::SINT_TO_FP, MVT::i16, Promote);
1846 setOperationAction(ISD::SINT_TO_FP, MVT::i32, Expand);
1847 setOperationAction(ISD::SINT_TO_FP, MVT::i64, Expand);
1848 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Expand);
1849 setOperationAction(ISD::UINT_TO_FP, MVT::i64, Expand);
1850 setOperationAction(ISD::FP_TO_SINT, MVT::f64, Expand);
1851 setOperationAction(ISD::FP_TO_SINT, MVT::f32, Expand);
1852 setOperationAction(ISD::FP_EXTEND, MVT::f32, Expand);
1853 setOperationAction(ISD::FP_ROUND, MVT::f64, Expand);
1854 setCondCodeAction(ISD::SETUNE, MVT::f64, Expand);
1856 setOperationAction(ISD::CTPOP, MVT::i8, Expand);
1857 setOperationAction(ISD::CTPOP, MVT::i16, Expand);
1858 setOperationAction(ISD::CTPOP, MVT::i32, Expand);
1859 setOperationAction(ISD::CTPOP, MVT::i64, Expand);
1861 // Expand these operations for both f32 and f64:
1862 for (unsigned FPExpOpV4 :
1863 {ISD::FADD, ISD::FSUB, ISD::FMUL, ISD::FABS, ISD::FNEG, ISD::FMA}) {
1864 setOperationAction(FPExpOpV4, MVT::f32, Expand);
1865 setOperationAction(FPExpOpV4, MVT::f64, Expand);
1868 for (ISD::CondCode FPExpCCV4 :
1869 {ISD::SETOEQ, ISD::SETOGT, ISD::SETOLT, ISD::SETOGE, ISD::SETOLE,
1870 ISD::SETUO, ISD::SETO}) {
1871 setCondCodeAction(FPExpCCV4, MVT::f32, Expand);
1872 setCondCodeAction(FPExpCCV4, MVT::f64, Expand);
1876 // Handling of indexed loads/stores: default is "expand".
1878 for (MVT LSXTy : {MVT::i8, MVT::i16, MVT::i32, MVT::i64}) {
1879 setIndexedLoadAction(ISD::POST_INC, LSXTy, Legal);
1880 setIndexedStoreAction(ISD::POST_INC, LSXTy, Legal);
1884 for (MVT VT : {MVT::v128i8, MVT::v64i16, MVT::v32i32, MVT::v16i64}) {
1885 setIndexedLoadAction(ISD::POST_INC, VT, Legal);
1886 setIndexedStoreAction(ISD::POST_INC, VT, Legal);
1890 computeRegisterProperties(&HRI);
1893 // Library calls for unsupported operations
1895 bool FastMath = EnableFastMath;
1897 setLibcallName(RTLIB::SDIV_I32, "__hexagon_divsi3");
1898 setLibcallName(RTLIB::SDIV_I64, "__hexagon_divdi3");
1899 setLibcallName(RTLIB::UDIV_I32, "__hexagon_udivsi3");
1900 setLibcallName(RTLIB::UDIV_I64, "__hexagon_udivdi3");
1901 setLibcallName(RTLIB::SREM_I32, "__hexagon_modsi3");
1902 setLibcallName(RTLIB::SREM_I64, "__hexagon_moddi3");
1903 setLibcallName(RTLIB::UREM_I32, "__hexagon_umodsi3");
1904 setLibcallName(RTLIB::UREM_I64, "__hexagon_umoddi3");
1906 setLibcallName(RTLIB::SINTTOFP_I128_F64, "__hexagon_floattidf");
1907 setLibcallName(RTLIB::SINTTOFP_I128_F32, "__hexagon_floattisf");
1908 setLibcallName(RTLIB::FPTOUINT_F32_I128, "__hexagon_fixunssfti");
1909 setLibcallName(RTLIB::FPTOUINT_F64_I128, "__hexagon_fixunsdfti");
1910 setLibcallName(RTLIB::FPTOSINT_F32_I128, "__hexagon_fixsfti");
1911 setLibcallName(RTLIB::FPTOSINT_F64_I128, "__hexagon_fixdfti");
1914 // Handle single-precision floating point operations on V4.
1916 setLibcallName(RTLIB::ADD_F32, "__hexagon_fast_addsf3");
1917 setLibcallName(RTLIB::SUB_F32, "__hexagon_fast_subsf3");
1918 setLibcallName(RTLIB::MUL_F32, "__hexagon_fast_mulsf3");
1919 setLibcallName(RTLIB::OGT_F32, "__hexagon_fast_gtsf2");
1920 setLibcallName(RTLIB::OLT_F32, "__hexagon_fast_ltsf2");
1921 // Double-precision compares.
1922 setLibcallName(RTLIB::OGT_F64, "__hexagon_fast_gtdf2");
1923 setLibcallName(RTLIB::OLT_F64, "__hexagon_fast_ltdf2");
1925 setLibcallName(RTLIB::ADD_F32, "__hexagon_addsf3");
1926 setLibcallName(RTLIB::SUB_F32, "__hexagon_subsf3");
1927 setLibcallName(RTLIB::MUL_F32, "__hexagon_mulsf3");
1928 setLibcallName(RTLIB::OGT_F32, "__hexagon_gtsf2");
1929 setLibcallName(RTLIB::OLT_F32, "__hexagon_ltsf2");
1930 // Double-precision compares.
1931 setLibcallName(RTLIB::OGT_F64, "__hexagon_gtdf2");
1932 setLibcallName(RTLIB::OLT_F64, "__hexagon_ltdf2");
1936 // This is the only fast library function for sqrtd.
1938 setLibcallName(RTLIB::SQRT_F64, "__hexagon_fast2_sqrtdf2");
1940 // Prefix is: nothing for "slow-math",
1941 // "fast2_" for V4 fast-math and V5+ fast-math double-precision
1942 // (actually, keep fast-math and fast-math2 separate for now)
1944 setLibcallName(RTLIB::ADD_F64, "__hexagon_fast_adddf3");
1945 setLibcallName(RTLIB::SUB_F64, "__hexagon_fast_subdf3");
1946 setLibcallName(RTLIB::MUL_F64, "__hexagon_fast_muldf3");
1947 setLibcallName(RTLIB::DIV_F64, "__hexagon_fast_divdf3");
1948 // Calling __hexagon_fast2_divsf3 with fast-math on V5 (ok).
1949 setLibcallName(RTLIB::DIV_F32, "__hexagon_fast_divsf3");
1951 setLibcallName(RTLIB::ADD_F64, "__hexagon_adddf3");
1952 setLibcallName(RTLIB::SUB_F64, "__hexagon_subdf3");
1953 setLibcallName(RTLIB::MUL_F64, "__hexagon_muldf3");
1954 setLibcallName(RTLIB::DIV_F64, "__hexagon_divdf3");
1955 setLibcallName(RTLIB::DIV_F32, "__hexagon_divsf3");
1958 if (Subtarget.hasV5TOps()) {
1960 setLibcallName(RTLIB::SQRT_F32, "__hexagon_fast2_sqrtf");
1962 setLibcallName(RTLIB::SQRT_F32, "__hexagon_sqrtf");
1965 setLibcallName(RTLIB::SINTTOFP_I32_F32, "__hexagon_floatsisf");
1966 setLibcallName(RTLIB::SINTTOFP_I32_F64, "__hexagon_floatsidf");
1967 setLibcallName(RTLIB::SINTTOFP_I64_F32, "__hexagon_floatdisf");
1968 setLibcallName(RTLIB::SINTTOFP_I64_F64, "__hexagon_floatdidf");
1969 setLibcallName(RTLIB::UINTTOFP_I32_F32, "__hexagon_floatunsisf");
1970 setLibcallName(RTLIB::UINTTOFP_I32_F64, "__hexagon_floatunsidf");
1971 setLibcallName(RTLIB::UINTTOFP_I64_F32, "__hexagon_floatundisf");
1972 setLibcallName(RTLIB::UINTTOFP_I64_F64, "__hexagon_floatundidf");
1973 setLibcallName(RTLIB::FPTOUINT_F32_I32, "__hexagon_fixunssfsi");
1974 setLibcallName(RTLIB::FPTOUINT_F32_I64, "__hexagon_fixunssfdi");
1975 setLibcallName(RTLIB::FPTOUINT_F64_I32, "__hexagon_fixunsdfsi");
1976 setLibcallName(RTLIB::FPTOUINT_F64_I64, "__hexagon_fixunsdfdi");
1977 setLibcallName(RTLIB::FPTOSINT_F32_I32, "__hexagon_fixsfsi");
1978 setLibcallName(RTLIB::FPTOSINT_F32_I64, "__hexagon_fixsfdi");
1979 setLibcallName(RTLIB::FPTOSINT_F64_I32, "__hexagon_fixdfsi");
1980 setLibcallName(RTLIB::FPTOSINT_F64_I64, "__hexagon_fixdfdi");
1981 setLibcallName(RTLIB::FPEXT_F32_F64, "__hexagon_extendsfdf2");
1982 setLibcallName(RTLIB::FPROUND_F64_F32, "__hexagon_truncdfsf2");
1983 setLibcallName(RTLIB::OEQ_F32, "__hexagon_eqsf2");
1984 setLibcallName(RTLIB::OEQ_F64, "__hexagon_eqdf2");
1985 setLibcallName(RTLIB::OGE_F32, "__hexagon_gesf2");
1986 setLibcallName(RTLIB::OGE_F64, "__hexagon_gedf2");
1987 setLibcallName(RTLIB::OLE_F32, "__hexagon_lesf2");
1988 setLibcallName(RTLIB::OLE_F64, "__hexagon_ledf2");
1989 setLibcallName(RTLIB::UNE_F32, "__hexagon_nesf2");
1990 setLibcallName(RTLIB::UNE_F64, "__hexagon_nedf2");
1991 setLibcallName(RTLIB::UO_F32, "__hexagon_unordsf2");
1992 setLibcallName(RTLIB::UO_F64, "__hexagon_unorddf2");
1993 setLibcallName(RTLIB::O_F32, "__hexagon_unordsf2");
1994 setLibcallName(RTLIB::O_F64, "__hexagon_unorddf2");
1997 // These cause problems when the shift amount is non-constant.
1998 setLibcallName(RTLIB::SHL_I128, nullptr);
1999 setLibcallName(RTLIB::SRL_I128, nullptr);
2000 setLibcallName(RTLIB::SRA_I128, nullptr);
2004 const char* HexagonTargetLowering::getTargetNodeName(unsigned Opcode) const {
2005 switch ((HexagonISD::NodeType)Opcode) {
2006 case HexagonISD::ALLOCA: return "HexagonISD::ALLOCA";
2007 case HexagonISD::ARGEXTEND: return "HexagonISD::ARGEXTEND";
2008 case HexagonISD::AT_GOT: return "HexagonISD::AT_GOT";
2009 case HexagonISD::AT_PCREL: return "HexagonISD::AT_PCREL";
2010 case HexagonISD::BARRIER: return "HexagonISD::BARRIER";
2011 case HexagonISD::BR_JT: return "HexagonISD::BR_JT";
2012 case HexagonISD::CALLR: return "HexagonISD::CALLR";
2013 case HexagonISD::CALLv3nr: return "HexagonISD::CALLv3nr";
2014 case HexagonISD::CALLv3: return "HexagonISD::CALLv3";
2015 case HexagonISD::COMBINE: return "HexagonISD::COMBINE";
2016 case HexagonISD::CONST32_GP: return "HexagonISD::CONST32_GP";
2017 case HexagonISD::CONST32: return "HexagonISD::CONST32";
2018 case HexagonISD::CP: return "HexagonISD::CP";
2019 case HexagonISD::DCFETCH: return "HexagonISD::DCFETCH";
2020 case HexagonISD::EH_RETURN: return "HexagonISD::EH_RETURN";
2021 case HexagonISD::EXTRACTU: return "HexagonISD::EXTRACTU";
2022 case HexagonISD::EXTRACTURP: return "HexagonISD::EXTRACTURP";
2023 case HexagonISD::FCONST32: return "HexagonISD::FCONST32";
2024 case HexagonISD::INSERT: return "HexagonISD::INSERT";
2025 case HexagonISD::INSERTRP: return "HexagonISD::INSERTRP";
2026 case HexagonISD::JT: return "HexagonISD::JT";
2027 case HexagonISD::PACKHL: return "HexagonISD::PACKHL";
2028 case HexagonISD::PIC_ADD: return "HexagonISD::PIC_ADD";
2029 case HexagonISD::POPCOUNT: return "HexagonISD::POPCOUNT";
2030 case HexagonISD::RET_FLAG: return "HexagonISD::RET_FLAG";
2031 case HexagonISD::SHUFFEB: return "HexagonISD::SHUFFEB";
2032 case HexagonISD::SHUFFEH: return "HexagonISD::SHUFFEH";
2033 case HexagonISD::SHUFFOB: return "HexagonISD::SHUFFOB";
2034 case HexagonISD::SHUFFOH: return "HexagonISD::SHUFFOH";
2035 case HexagonISD::TC_RETURN: return "HexagonISD::TC_RETURN";
2036 case HexagonISD::VCMPBEQ: return "HexagonISD::VCMPBEQ";
2037 case HexagonISD::VCMPBGT: return "HexagonISD::VCMPBGT";
2038 case HexagonISD::VCMPBGTU: return "HexagonISD::VCMPBGTU";
2039 case HexagonISD::VCMPHEQ: return "HexagonISD::VCMPHEQ";
2040 case HexagonISD::VCMPHGT: return "HexagonISD::VCMPHGT";
2041 case HexagonISD::VCMPHGTU: return "HexagonISD::VCMPHGTU";
2042 case HexagonISD::VCMPWEQ: return "HexagonISD::VCMPWEQ";
2043 case HexagonISD::VCMPWGT: return "HexagonISD::VCMPWGT";
2044 case HexagonISD::VCMPWGTU: return "HexagonISD::VCMPWGTU";
2045 case HexagonISD::VSHLH: return "HexagonISD::VSHLH";
2046 case HexagonISD::VSHLW: return "HexagonISD::VSHLW";
2047 case HexagonISD::VSPLATB: return "HexagonISD::VSPLTB";
2048 case HexagonISD::VSPLATH: return "HexagonISD::VSPLATH";
2049 case HexagonISD::VSRAH: return "HexagonISD::VSRAH";
2050 case HexagonISD::VSRAW: return "HexagonISD::VSRAW";
2051 case HexagonISD::VSRLH: return "HexagonISD::VSRLH";
2052 case HexagonISD::VSRLW: return "HexagonISD::VSRLW";
2053 case HexagonISD::VSXTBH: return "HexagonISD::VSXTBH";
2054 case HexagonISD::VSXTBW: return "HexagonISD::VSXTBW";
2055 case HexagonISD::OP_END: break;
2060 bool HexagonTargetLowering::isTruncateFree(Type *Ty1, Type *Ty2) const {
2061 EVT MTy1 = EVT::getEVT(Ty1);
2062 EVT MTy2 = EVT::getEVT(Ty2);
2063 if (!MTy1.isSimple() || !MTy2.isSimple())
2065 return (MTy1.getSimpleVT() == MVT::i64) && (MTy2.getSimpleVT() == MVT::i32);
2068 bool HexagonTargetLowering::isTruncateFree(EVT VT1, EVT VT2) const {
2069 if (!VT1.isSimple() || !VT2.isSimple())
2071 return (VT1.getSimpleVT() == MVT::i64) && (VT2.getSimpleVT() == MVT::i32);
2074 // shouldExpandBuildVectorWithShuffles
2075 // Should we expand the build vector with shuffles?
2077 HexagonTargetLowering::shouldExpandBuildVectorWithShuffles(EVT VT,
2078 unsigned DefinedValues) const {
2080 // Hexagon vector shuffle operates on element sizes of bytes or halfwords
2081 EVT EltVT = VT.getVectorElementType();
2082 int EltBits = EltVT.getSizeInBits();
2083 if ((EltBits != 8) && (EltBits != 16))
2086 return TargetLowering::shouldExpandBuildVectorWithShuffles(VT, DefinedValues);
2089 // LowerVECTOR_SHUFFLE - Lower a vector shuffle (V1, V2, V3). V1 and
2090 // V2 are the two vectors to select data from, V3 is the permutation.
2091 static SDValue LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) {
2092 const ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(Op);
2093 SDValue V1 = Op.getOperand(0);
2094 SDValue V2 = Op.getOperand(1);
2096 EVT VT = Op.getValueType();
2098 if (V2.getOpcode() == ISD::UNDEF)
2101 if (SVN->isSplat()) {
2102 int Lane = SVN->getSplatIndex();
2103 if (Lane == -1) Lane = 0;
2105 // Test if V1 is a SCALAR_TO_VECTOR.
2106 if (Lane == 0 && V1.getOpcode() == ISD::SCALAR_TO_VECTOR)
2107 return createSplat(DAG, dl, VT, V1.getOperand(0));
2109 // Test if V1 is a BUILD_VECTOR which is equivalent to a SCALAR_TO_VECTOR
2110 // (and probably will turn into a SCALAR_TO_VECTOR once legalization
2112 if (Lane == 0 && V1.getOpcode() == ISD::BUILD_VECTOR &&
2113 !isa<ConstantSDNode>(V1.getOperand(0))) {
2114 bool IsScalarToVector = true;
2115 for (unsigned i = 1, e = V1.getNumOperands(); i != e; ++i)
2116 if (V1.getOperand(i).getOpcode() != ISD::UNDEF) {
2117 IsScalarToVector = false;
2120 if (IsScalarToVector)
2121 return createSplat(DAG, dl, VT, V1.getOperand(0));
2123 return createSplat(DAG, dl, VT, DAG.getConstant(Lane, dl, MVT::i32));
2126 // FIXME: We need to support more general vector shuffles. See
2127 // below the comment from the ARM backend that deals in the general
2128 // case with the vector shuffles. For now, let expand handle these.
2131 // If the shuffle is not directly supported and it has 4 elements, use
2132 // the PerfectShuffle-generated table to synthesize it from other shuffles.
2135 // If BUILD_VECTOR has same base element repeated several times,
2137 static bool isCommonSplatElement(BuildVectorSDNode *BVN) {
2138 unsigned NElts = BVN->getNumOperands();
2139 SDValue V0 = BVN->getOperand(0);
2141 for (unsigned i = 1, e = NElts; i != e; ++i) {
2142 if (BVN->getOperand(i) != V0)
2148 // LowerVECTOR_SHIFT - Lower a vector shift. Try to convert
2149 // <VT> = SHL/SRA/SRL <VT> by <VT> to Hexagon specific
2150 // <VT> = SHL/SRA/SRL <VT> by <IT/i32>.
2151 static SDValue LowerVECTOR_SHIFT(SDValue Op, SelectionDAG &DAG) {
2152 BuildVectorSDNode *BVN = 0;
2153 SDValue V1 = Op.getOperand(0);
2154 SDValue V2 = Op.getOperand(1);
2157 EVT VT = Op.getValueType();
2159 if ((BVN = dyn_cast<BuildVectorSDNode>(V1.getNode())) &&
2160 isCommonSplatElement(BVN))
2162 else if ((BVN = dyn_cast<BuildVectorSDNode>(V2.getNode())) &&
2163 isCommonSplatElement(BVN))
2168 SDValue CommonSplat = BVN->getOperand(0);
2171 if (VT.getSimpleVT() == MVT::v4i16) {
2172 switch (Op.getOpcode()) {
2174 Result = DAG.getNode(HexagonISD::VSRAH, dl, VT, V3, CommonSplat);
2177 Result = DAG.getNode(HexagonISD::VSHLH, dl, VT, V3, CommonSplat);
2180 Result = DAG.getNode(HexagonISD::VSRLH, dl, VT, V3, CommonSplat);
2185 } else if (VT.getSimpleVT() == MVT::v2i32) {
2186 switch (Op.getOpcode()) {
2188 Result = DAG.getNode(HexagonISD::VSRAW, dl, VT, V3, CommonSplat);
2191 Result = DAG.getNode(HexagonISD::VSHLW, dl, VT, V3, CommonSplat);
2194 Result = DAG.getNode(HexagonISD::VSRLW, dl, VT, V3, CommonSplat);
2203 return DAG.getNode(ISD::BITCAST, dl, VT, Result);
2207 HexagonTargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) const {
2208 BuildVectorSDNode *BVN = cast<BuildVectorSDNode>(Op.getNode());
2210 EVT VT = Op.getValueType();
2212 unsigned Size = VT.getSizeInBits();
2214 // A vector larger than 64 bits cannot be represented in Hexagon.
2215 // Expand will split the vector.
2219 APInt APSplatBits, APSplatUndef;
2220 unsigned SplatBitSize;
2222 unsigned NElts = BVN->getNumOperands();
2224 // Try to generate a SPLAT instruction.
2225 if ((VT.getSimpleVT() == MVT::v4i8 || VT.getSimpleVT() == MVT::v4i16) &&
2226 (BVN->isConstantSplat(APSplatBits, APSplatUndef, SplatBitSize,
2227 HasAnyUndefs, 0, true) && SplatBitSize <= 16)) {
2228 unsigned SplatBits = APSplatBits.getZExtValue();
2229 int32_t SextVal = ((int32_t) (SplatBits << (32 - SplatBitSize)) >>
2230 (32 - SplatBitSize));
2231 return createSplat(DAG, dl, VT, DAG.getConstant(SextVal, dl, MVT::i32));
2234 // Try to generate COMBINE to build v2i32 vectors.
2235 if (VT.getSimpleVT() == MVT::v2i32) {
2236 SDValue V0 = BVN->getOperand(0);
2237 SDValue V1 = BVN->getOperand(1);
2239 if (V0.getOpcode() == ISD::UNDEF)
2240 V0 = DAG.getConstant(0, dl, MVT::i32);
2241 if (V1.getOpcode() == ISD::UNDEF)
2242 V1 = DAG.getConstant(0, dl, MVT::i32);
2244 ConstantSDNode *C0 = dyn_cast<ConstantSDNode>(V0);
2245 ConstantSDNode *C1 = dyn_cast<ConstantSDNode>(V1);
2246 // If the element isn't a constant, it is in a register:
2247 // generate a COMBINE Register Register instruction.
2249 return DAG.getNode(HexagonISD::COMBINE, dl, VT, V1, V0);
2251 // If one of the operands is an 8 bit integer constant, generate
2252 // a COMBINE Immediate Immediate instruction.
2253 if (isInt<8>(C0->getSExtValue()) ||
2254 isInt<8>(C1->getSExtValue()))
2255 return DAG.getNode(HexagonISD::COMBINE, dl, VT, V1, V0);
2258 // Try to generate a S2_packhl to build v2i16 vectors.
2259 if (VT.getSimpleVT() == MVT::v2i16) {
2260 for (unsigned i = 0, e = NElts; i != e; ++i) {
2261 if (BVN->getOperand(i).getOpcode() == ISD::UNDEF)
2263 ConstantSDNode *Cst = dyn_cast<ConstantSDNode>(BVN->getOperand(i));
2264 // If the element isn't a constant, it is in a register:
2265 // generate a S2_packhl instruction.
2267 SDValue pack = DAG.getNode(HexagonISD::PACKHL, dl, MVT::v4i16,
2268 BVN->getOperand(1), BVN->getOperand(0));
2270 return DAG.getTargetExtractSubreg(Hexagon::subreg_loreg, dl, MVT::v2i16,
2276 // In the general case, generate a CONST32 or a CONST64 for constant vectors,
2277 // and insert_vector_elt for all the other cases.
2279 unsigned EltSize = Size / NElts;
2281 uint64_t Mask = ~uint64_t(0ULL) >> (64 - EltSize);
2282 bool HasNonConstantElements = false;
2284 for (unsigned i = 0, e = NElts; i != e; ++i) {
2285 // LLVM's BUILD_VECTOR operands are in Little Endian mode, whereas Hexagon's
2286 // combine, const64, etc. are Big Endian.
2287 unsigned OpIdx = NElts - i - 1;
2288 SDValue Operand = BVN->getOperand(OpIdx);
2289 if (Operand.getOpcode() == ISD::UNDEF)
2293 if (ConstantSDNode *Cst = dyn_cast<ConstantSDNode>(Operand))
2294 Val = Cst->getSExtValue();
2296 HasNonConstantElements = true;
2299 Res = (Res << EltSize) | Val;
2303 ConstVal = DAG.getConstant(Res, dl, MVT::i64);
2305 ConstVal = DAG.getConstant(Res, dl, MVT::i32);
2307 // When there are non constant operands, add them with INSERT_VECTOR_ELT to
2308 // ConstVal, the constant part of the vector.
2309 if (HasNonConstantElements) {
2310 EVT EltVT = VT.getVectorElementType();
2311 SDValue Width = DAG.getConstant(EltVT.getSizeInBits(), dl, MVT::i64);
2312 SDValue Shifted = DAG.getNode(ISD::SHL, dl, MVT::i64, Width,
2313 DAG.getConstant(32, dl, MVT::i64));
2315 for (unsigned i = 0, e = NElts; i != e; ++i) {
2316 // LLVM's BUILD_VECTOR operands are in Little Endian mode, whereas Hexagon
2318 unsigned OpIdx = NElts - i - 1;
2319 SDValue Operand = BVN->getOperand(OpIdx);
2320 if (isa<ConstantSDNode>(Operand))
2321 // This operand is already in ConstVal.
2324 if (VT.getSizeInBits() == 64 &&
2325 Operand.getValueType().getSizeInBits() == 32) {
2326 SDValue C = DAG.getConstant(0, dl, MVT::i32);
2327 Operand = DAG.getNode(HexagonISD::COMBINE, dl, VT, C, Operand);
2330 SDValue Idx = DAG.getConstant(OpIdx, dl, MVT::i64);
2331 SDValue Offset = DAG.getNode(ISD::MUL, dl, MVT::i64, Idx, Width);
2332 SDValue Combined = DAG.getNode(ISD::OR, dl, MVT::i64, Shifted, Offset);
2333 const SDValue Ops[] = {ConstVal, Operand, Combined};
2335 if (VT.getSizeInBits() == 32)
2336 ConstVal = DAG.getNode(HexagonISD::INSERTRP, dl, MVT::i32, Ops);
2338 ConstVal = DAG.getNode(HexagonISD::INSERTRP, dl, MVT::i64, Ops);
2342 return DAG.getNode(ISD::BITCAST, dl, VT, ConstVal);
2346 HexagonTargetLowering::LowerCONCAT_VECTORS(SDValue Op,
2347 SelectionDAG &DAG) const {
2349 EVT VT = Op.getValueType();
2350 unsigned NElts = Op.getNumOperands();
2351 SDValue Vec = Op.getOperand(0);
2352 EVT VecVT = Vec.getValueType();
2353 SDValue Width = DAG.getConstant(VecVT.getSizeInBits(), dl, MVT::i64);
2354 SDValue Shifted = DAG.getNode(ISD::SHL, dl, MVT::i64, Width,
2355 DAG.getConstant(32, dl, MVT::i64));
2356 SDValue ConstVal = DAG.getConstant(0, dl, MVT::i64);
2358 ConstantSDNode *W = dyn_cast<ConstantSDNode>(Width);
2359 ConstantSDNode *S = dyn_cast<ConstantSDNode>(Shifted);
2361 if ((VecVT.getSimpleVT() == MVT::v2i16) && (NElts == 2) && W && S) {
2362 if ((W->getZExtValue() == 32) && ((S->getZExtValue() >> 32) == 32)) {
2363 // We are trying to concat two v2i16 to a single v4i16.
2364 SDValue Vec0 = Op.getOperand(1);
2365 SDValue Combined = DAG.getNode(HexagonISD::COMBINE, dl, VT, Vec0, Vec);
2366 return DAG.getNode(ISD::BITCAST, dl, VT, Combined);
2370 if ((VecVT.getSimpleVT() == MVT::v4i8) && (NElts == 2) && W && S) {
2371 if ((W->getZExtValue() == 32) && ((S->getZExtValue() >> 32) == 32)) {
2372 // We are trying to concat two v4i8 to a single v8i8.
2373 SDValue Vec0 = Op.getOperand(1);
2374 SDValue Combined = DAG.getNode(HexagonISD::COMBINE, dl, VT, Vec0, Vec);
2375 return DAG.getNode(ISD::BITCAST, dl, VT, Combined);
2379 for (unsigned i = 0, e = NElts; i != e; ++i) {
2380 unsigned OpIdx = NElts - i - 1;
2381 SDValue Operand = Op.getOperand(OpIdx);
2383 if (VT.getSizeInBits() == 64 &&
2384 Operand.getValueType().getSizeInBits() == 32) {
2385 SDValue C = DAG.getConstant(0, dl, MVT::i32);
2386 Operand = DAG.getNode(HexagonISD::COMBINE, dl, VT, C, Operand);
2389 SDValue Idx = DAG.getConstant(OpIdx, dl, MVT::i64);
2390 SDValue Offset = DAG.getNode(ISD::MUL, dl, MVT::i64, Idx, Width);
2391 SDValue Combined = DAG.getNode(ISD::OR, dl, MVT::i64, Shifted, Offset);
2392 const SDValue Ops[] = {ConstVal, Operand, Combined};
2394 if (VT.getSizeInBits() == 32)
2395 ConstVal = DAG.getNode(HexagonISD::INSERTRP, dl, MVT::i32, Ops);
2397 ConstVal = DAG.getNode(HexagonISD::INSERTRP, dl, MVT::i64, Ops);
2400 return DAG.getNode(ISD::BITCAST, dl, VT, ConstVal);
2404 HexagonTargetLowering::LowerEXTRACT_VECTOR(SDValue Op,
2405 SelectionDAG &DAG) const {
2406 EVT VT = Op.getValueType();
2407 int VTN = VT.isVector() ? VT.getVectorNumElements() : 1;
2409 SDValue Idx = Op.getOperand(1);
2410 SDValue Vec = Op.getOperand(0);
2411 EVT VecVT = Vec.getValueType();
2412 EVT EltVT = VecVT.getVectorElementType();
2413 int EltSize = EltVT.getSizeInBits();
2414 SDValue Width = DAG.getConstant(Op.getOpcode() == ISD::EXTRACT_VECTOR_ELT ?
2415 EltSize : VTN * EltSize, dl, MVT::i64);
2417 // Constant element number.
2418 if (ConstantSDNode *CI = dyn_cast<ConstantSDNode>(Idx)) {
2419 uint64_t X = CI->getZExtValue();
2420 SDValue Offset = DAG.getConstant(X * EltSize, dl, MVT::i32);
2421 const SDValue Ops[] = {Vec, Width, Offset};
2423 ConstantSDNode *CW = dyn_cast<ConstantSDNode>(Width);
2424 assert(CW && "Non constant width in LowerEXTRACT_VECTOR");
2427 MVT SVT = VecVT.getSimpleVT();
2428 uint64_t W = CW->getZExtValue();
2431 // Translate this node into EXTRACT_SUBREG.
2432 unsigned Subreg = (X == 0) ? Hexagon::subreg_loreg : 0;
2435 Subreg = Hexagon::subreg_loreg;
2436 else if (SVT == MVT::v2i32 && X == 1)
2437 Subreg = Hexagon::subreg_hireg;
2438 else if (SVT == MVT::v4i16 && X == 2)
2439 Subreg = Hexagon::subreg_hireg;
2440 else if (SVT == MVT::v8i8 && X == 4)
2441 Subreg = Hexagon::subreg_hireg;
2443 llvm_unreachable("Bad offset");
2444 N = DAG.getTargetExtractSubreg(Subreg, dl, MVT::i32, Vec);
2446 } else if (VecVT.getSizeInBits() == 32) {
2447 N = DAG.getNode(HexagonISD::EXTRACTU, dl, MVT::i32, Ops);
2449 N = DAG.getNode(HexagonISD::EXTRACTU, dl, MVT::i64, Ops);
2450 if (VT.getSizeInBits() == 32)
2451 N = DAG.getTargetExtractSubreg(Hexagon::subreg_loreg, dl, MVT::i32, N);
2454 return DAG.getNode(ISD::BITCAST, dl, VT, N);
2457 // Variable element number.
2458 SDValue Offset = DAG.getNode(ISD::MUL, dl, MVT::i32, Idx,
2459 DAG.getConstant(EltSize, dl, MVT::i32));
2460 SDValue Shifted = DAG.getNode(ISD::SHL, dl, MVT::i64, Width,
2461 DAG.getConstant(32, dl, MVT::i64));
2462 SDValue Combined = DAG.getNode(ISD::OR, dl, MVT::i64, Shifted, Offset);
2464 const SDValue Ops[] = {Vec, Combined};
2467 if (VecVT.getSizeInBits() == 32) {
2468 N = DAG.getNode(HexagonISD::EXTRACTURP, dl, MVT::i32, Ops);
2470 N = DAG.getNode(HexagonISD::EXTRACTURP, dl, MVT::i64, Ops);
2471 if (VT.getSizeInBits() == 32)
2472 N = DAG.getTargetExtractSubreg(Hexagon::subreg_loreg, dl, MVT::i32, N);
2474 return DAG.getNode(ISD::BITCAST, dl, VT, N);
2478 HexagonTargetLowering::LowerINSERT_VECTOR(SDValue Op,
2479 SelectionDAG &DAG) const {
2480 EVT VT = Op.getValueType();
2481 int VTN = VT.isVector() ? VT.getVectorNumElements() : 1;
2483 SDValue Vec = Op.getOperand(0);
2484 SDValue Val = Op.getOperand(1);
2485 SDValue Idx = Op.getOperand(2);
2486 EVT VecVT = Vec.getValueType();
2487 EVT EltVT = VecVT.getVectorElementType();
2488 int EltSize = EltVT.getSizeInBits();
2489 SDValue Width = DAG.getConstant(Op.getOpcode() == ISD::INSERT_VECTOR_ELT ?
2490 EltSize : VTN * EltSize, dl, MVT::i64);
2492 if (ConstantSDNode *C = cast<ConstantSDNode>(Idx)) {
2493 SDValue Offset = DAG.getConstant(C->getSExtValue() * EltSize, dl, MVT::i32);
2494 const SDValue Ops[] = {Vec, Val, Width, Offset};
2497 if (VT.getSizeInBits() == 32)
2498 N = DAG.getNode(HexagonISD::INSERT, dl, MVT::i32, Ops);
2500 N = DAG.getNode(HexagonISD::INSERT, dl, MVT::i64, Ops);
2502 return DAG.getNode(ISD::BITCAST, dl, VT, N);
2505 // Variable element number.
2506 SDValue Offset = DAG.getNode(ISD::MUL, dl, MVT::i32, Idx,
2507 DAG.getConstant(EltSize, dl, MVT::i32));
2508 SDValue Shifted = DAG.getNode(ISD::SHL, dl, MVT::i64, Width,
2509 DAG.getConstant(32, dl, MVT::i64));
2510 SDValue Combined = DAG.getNode(ISD::OR, dl, MVT::i64, Shifted, Offset);
2512 if (VT.getSizeInBits() == 64 &&
2513 Val.getValueType().getSizeInBits() == 32) {
2514 SDValue C = DAG.getConstant(0, dl, MVT::i32);
2515 Val = DAG.getNode(HexagonISD::COMBINE, dl, VT, C, Val);
2518 const SDValue Ops[] = {Vec, Val, Combined};
2521 if (VT.getSizeInBits() == 32)
2522 N = DAG.getNode(HexagonISD::INSERTRP, dl, MVT::i32, Ops);
2524 N = DAG.getNode(HexagonISD::INSERTRP, dl, MVT::i64, Ops);
2526 return DAG.getNode(ISD::BITCAST, dl, VT, N);
2530 HexagonTargetLowering::allowTruncateForTailCall(Type *Ty1, Type *Ty2) const {
2531 // Assuming the caller does not have either a signext or zeroext modifier, and
2532 // only one value is accepted, any reasonable truncation is allowed.
2533 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
2536 // FIXME: in principle up to 64-bit could be made safe, but it would be very
2537 // fragile at the moment: any support for multiple value returns would be
2538 // liable to disallow tail calls involving i64 -> iN truncation in many cases.
2539 return Ty1->getPrimitiveSizeInBits() <= 32;
2543 HexagonTargetLowering::LowerEH_RETURN(SDValue Op, SelectionDAG &DAG) const {
2544 SDValue Chain = Op.getOperand(0);
2545 SDValue Offset = Op.getOperand(1);
2546 SDValue Handler = Op.getOperand(2);
2548 auto PtrVT = getPointerTy(DAG.getDataLayout());
2550 // Mark function as containing a call to EH_RETURN.
2551 HexagonMachineFunctionInfo *FuncInfo =
2552 DAG.getMachineFunction().getInfo<HexagonMachineFunctionInfo>();
2553 FuncInfo->setHasEHReturn();
2555 unsigned OffsetReg = Hexagon::R28;
2558 DAG.getNode(ISD::ADD, dl, PtrVT, DAG.getRegister(Hexagon::R30, PtrVT),
2559 DAG.getIntPtrConstant(4, dl));
2560 Chain = DAG.getStore(Chain, dl, Handler, StoreAddr, MachinePointerInfo(),
2562 Chain = DAG.getCopyToReg(Chain, dl, OffsetReg, Offset);
2564 // Not needed we already use it as explict input to EH_RETURN.
2565 // MF.getRegInfo().addLiveOut(OffsetReg);
2567 return DAG.getNode(HexagonISD::EH_RETURN, dl, MVT::Other, Chain);
2571 HexagonTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
2572 unsigned Opc = Op.getOpcode();
2576 Op.getNode()->dumpr(&DAG);
2577 if (Opc > HexagonISD::OP_BEGIN && Opc < HexagonISD::OP_END)
2578 errs() << "Check for a non-legal type in this operation\n";
2580 llvm_unreachable("Should not custom lower this!");
2581 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
2582 case ISD::INSERT_SUBVECTOR: return LowerINSERT_VECTOR(Op, DAG);
2583 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR(Op, DAG);
2584 case ISD::EXTRACT_SUBVECTOR: return LowerEXTRACT_VECTOR(Op, DAG);
2585 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR(Op, DAG);
2586 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
2587 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
2590 case ISD::SRL: return LowerVECTOR_SHIFT(Op, DAG);
2591 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
2592 case ISD::EH_RETURN: return LowerEH_RETURN(Op, DAG);
2593 // Frame & Return address. Currently unimplemented.
2594 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
2595 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
2596 case ISD::ATOMIC_FENCE: return LowerATOMIC_FENCE(Op, DAG);
2597 case ISD::GlobalAddress: return LowerGLOBALADDRESS(Op, DAG);
2598 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
2599 case ISD::VASTART: return LowerVASTART(Op, DAG);
2600 case ISD::BR_JT: return LowerBR_JT(Op, DAG);
2601 // Custom lower some vector loads.
2602 case ISD::LOAD: return LowerLOAD(Op, DAG);
2603 case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
2604 case ISD::SETCC: return LowerSETCC(Op, DAG);
2605 case ISD::VSELECT: return LowerVSELECT(Op, DAG);
2606 case ISD::CTPOP: return LowerCTPOP(Op, DAG);
2607 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
2608 case ISD::INLINEASM: return LowerINLINEASM(Op, DAG);
2613 HexagonTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
2614 MachineBasicBlock *BB)
2616 switch (MI->getOpcode()) {
2617 case Hexagon::ALLOCA: {
2618 MachineFunction *MF = BB->getParent();
2619 auto *FuncInfo = MF->getInfo<HexagonMachineFunctionInfo>();
2620 FuncInfo->addAllocaAdjustInst(MI);
2623 default: llvm_unreachable("Unexpected instr type to insert");
2627 //===----------------------------------------------------------------------===//
2628 // Inline Assembly Support
2629 //===----------------------------------------------------------------------===//
2631 std::pair<unsigned, const TargetRegisterClass *>
2632 HexagonTargetLowering::getRegForInlineAsmConstraint(
2633 const TargetRegisterInfo *TRI, StringRef Constraint, MVT VT) const {
2634 bool UseHVX = Subtarget.useHVXOps(), UseHVXDbl = Subtarget.useHVXDblOps();
2636 if (Constraint.size() == 1) {
2637 switch (Constraint[0]) {
2639 switch (VT.SimpleTy) {
2641 llvm_unreachable("getRegForInlineAsmConstraint Unhandled data type");
2646 return std::make_pair(0U, &Hexagon::IntRegsRegClass);
2649 return std::make_pair(0U, &Hexagon::DoubleRegsRegClass);
2652 switch (VT.SimpleTy) {
2654 llvm_unreachable("getRegForInlineAsmConstraint Unhandled data type");
2661 return std::make_pair(0U, &Hexagon::VecPredRegsRegClass);
2664 switch (VT.SimpleTy) {
2666 llvm_unreachable("getRegForInlineAsmConstraint Unhandled data type");
2671 return std::make_pair(0U, &Hexagon::VectorRegsRegClass);
2676 if (Subtarget.hasV60TOps() && UseHVX && UseHVXDbl)
2677 return std::make_pair(0U, &Hexagon::VectorRegs128BRegClass);
2679 return std::make_pair(0U, &Hexagon::VecDblRegsRegClass);
2684 return std::make_pair(0U, &Hexagon::VecDblRegs128BRegClass);
2688 llvm_unreachable("Unknown asm register class");
2692 return TargetLowering::getRegForInlineAsmConstraint(TRI, Constraint, VT);
2695 /// isFPImmLegal - Returns true if the target can instruction select the
2696 /// specified FP immediate natively. If false, the legalizer will
2697 /// materialize the FP immediate as a load from a constant pool.
2698 bool HexagonTargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
2699 return Subtarget.hasV5TOps();
2702 /// isLegalAddressingMode - Return true if the addressing mode represented by
2703 /// AM is legal for this target, for a load/store of the specified type.
2704 bool HexagonTargetLowering::isLegalAddressingMode(const DataLayout &DL,
2705 const AddrMode &AM, Type *Ty,
2706 unsigned AS) const {
2707 // Allows a signed-extended 11-bit immediate field.
2708 if (AM.BaseOffs <= -(1LL << 13) || AM.BaseOffs >= (1LL << 13)-1)
2711 // No global is ever allowed as a base.
2715 int Scale = AM.Scale;
2716 if (Scale < 0) Scale = -Scale;
2718 case 0: // No scale reg, "r+i", "r", or just "i".
2720 default: // No scaled addressing mode.
2726 /// isLegalICmpImmediate - Return true if the specified immediate is legal
2727 /// icmp immediate, that is the target has icmp instructions which can compare
2728 /// a register against the immediate without having to materialize the
2729 /// immediate into a register.
2730 bool HexagonTargetLowering::isLegalICmpImmediate(int64_t Imm) const {
2731 return Imm >= -512 && Imm <= 511;
2734 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
2735 /// for tail call optimization. Targets which want to do tail call
2736 /// optimization should implement this function.
2737 bool HexagonTargetLowering::IsEligibleForTailCallOptimization(
2739 CallingConv::ID CalleeCC,
2741 bool isCalleeStructRet,
2742 bool isCallerStructRet,
2743 const SmallVectorImpl<ISD::OutputArg> &Outs,
2744 const SmallVectorImpl<SDValue> &OutVals,
2745 const SmallVectorImpl<ISD::InputArg> &Ins,
2746 SelectionDAG& DAG) const {
2747 const Function *CallerF = DAG.getMachineFunction().getFunction();
2748 CallingConv::ID CallerCC = CallerF->getCallingConv();
2749 bool CCMatch = CallerCC == CalleeCC;
2751 // ***************************************************************************
2752 // Look for obvious safe cases to perform tail call optimization that do not
2753 // require ABI changes.
2754 // ***************************************************************************
2756 // If this is a tail call via a function pointer, then don't do it!
2757 if (!(isa<GlobalAddressSDNode>(Callee)) &&
2758 !(isa<ExternalSymbolSDNode>(Callee))) {
2762 // Do not optimize if the calling conventions do not match.
2766 // Do not tail call optimize vararg calls.
2770 // Also avoid tail call optimization if either caller or callee uses struct
2771 // return semantics.
2772 if (isCalleeStructRet || isCallerStructRet)
2775 // In addition to the cases above, we also disable Tail Call Optimization if
2776 // the calling convention code that at least one outgoing argument needs to
2777 // go on the stack. We cannot check that here because at this point that
2778 // information is not available.
2782 // Return true when the given node fits in a positive half word.
2783 bool llvm::isPositiveHalfWord(SDNode *N) {
2784 ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N);
2785 if (CN && CN->getSExtValue() > 0 && isInt<16>(CN->getSExtValue()))
2788 switch (N->getOpcode()) {
2791 case ISD::SIGN_EXTEND_INREG:
2796 std::pair<const TargetRegisterClass*, uint8_t>
2797 HexagonTargetLowering::findRepresentativeClass(const TargetRegisterInfo *TRI,
2799 const TargetRegisterClass *RRC = nullptr;
2802 switch (VT.SimpleTy) {
2804 return TargetLowering::findRepresentativeClass(TRI, VT);
2809 RRC = &Hexagon::VectorRegsRegClass;
2815 if (Subtarget.hasV60TOps() && Subtarget.useHVXOps() &&
2816 Subtarget.useHVXDblOps())
2817 RRC = &Hexagon::VectorRegs128BRegClass;
2819 RRC = &Hexagon::VecDblRegsRegClass;
2825 RRC = &Hexagon::VecDblRegs128BRegClass;
2828 return std::make_pair(RRC, Cost);
2831 Value *HexagonTargetLowering::emitLoadLinked(IRBuilder<> &Builder, Value *Addr,
2832 AtomicOrdering Ord) const {
2833 BasicBlock *BB = Builder.GetInsertBlock();
2834 Module *M = BB->getParent()->getParent();
2835 Type *Ty = cast<PointerType>(Addr->getType())->getElementType();
2836 unsigned SZ = Ty->getPrimitiveSizeInBits();
2837 assert((SZ == 32 || SZ == 64) && "Only 32/64-bit atomic loads supported");
2838 Intrinsic::ID IntID = (SZ == 32) ? Intrinsic::hexagon_L2_loadw_locked
2839 : Intrinsic::hexagon_L4_loadd_locked;
2840 Value *Fn = Intrinsic::getDeclaration(M, IntID);
2841 return Builder.CreateCall(Fn, Addr, "larx");
2844 /// Perform a store-conditional operation to Addr. Return the status of the
2845 /// store. This should be 0 if the store succeeded, non-zero otherwise.
2846 Value *HexagonTargetLowering::emitStoreConditional(IRBuilder<> &Builder,
2847 Value *Val, Value *Addr, AtomicOrdering Ord) const {
2848 BasicBlock *BB = Builder.GetInsertBlock();
2849 Module *M = BB->getParent()->getParent();
2850 Type *Ty = Val->getType();
2851 unsigned SZ = Ty->getPrimitiveSizeInBits();
2852 assert((SZ == 32 || SZ == 64) && "Only 32/64-bit atomic stores supported");
2853 Intrinsic::ID IntID = (SZ == 32) ? Intrinsic::hexagon_S2_storew_locked
2854 : Intrinsic::hexagon_S4_stored_locked;
2855 Value *Fn = Intrinsic::getDeclaration(M, IntID);
2856 Value *Call = Builder.CreateCall(Fn, {Addr, Val}, "stcx");
2857 Value *Cmp = Builder.CreateICmpEQ(Call, Builder.getInt32(0), "");
2858 Value *Ext = Builder.CreateZExt(Cmp, Type::getInt32Ty(M->getContext()));
2862 TargetLowering::AtomicExpansionKind
2863 HexagonTargetLowering::shouldExpandAtomicLoadInIR(LoadInst *LI) const {
2864 // Do not expand loads and stores that don't exceed 64 bits.
2865 return LI->getType()->getPrimitiveSizeInBits() > 64
2866 ? AtomicExpansionKind::LLSC
2867 : AtomicExpansionKind::None;
2870 bool HexagonTargetLowering::shouldExpandAtomicStoreInIR(StoreInst *SI) const {
2871 // Do not expand loads and stores that don't exceed 64 bits.
2872 return SI->getValueOperand()->getType()->getPrimitiveSizeInBits() > 64;