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 CC_Hexagon(unsigned ValNo, MVT ValVT,
103 MVT LocVT, CCValAssign::LocInfo LocInfo,
104 ISD::ArgFlagsTy ArgFlags, CCState &State);
107 CC_Hexagon32(unsigned ValNo, MVT ValVT,
108 MVT LocVT, CCValAssign::LocInfo LocInfo,
109 ISD::ArgFlagsTy ArgFlags, CCState &State);
112 CC_Hexagon64(unsigned ValNo, MVT ValVT,
113 MVT LocVT, CCValAssign::LocInfo LocInfo,
114 ISD::ArgFlagsTy ArgFlags, CCState &State);
117 RetCC_Hexagon(unsigned ValNo, MVT ValVT,
118 MVT LocVT, CCValAssign::LocInfo LocInfo,
119 ISD::ArgFlagsTy ArgFlags, CCState &State);
122 RetCC_Hexagon32(unsigned ValNo, MVT ValVT,
123 MVT LocVT, CCValAssign::LocInfo LocInfo,
124 ISD::ArgFlagsTy ArgFlags, CCState &State);
127 RetCC_Hexagon64(unsigned ValNo, MVT ValVT,
128 MVT LocVT, CCValAssign::LocInfo LocInfo,
129 ISD::ArgFlagsTy ArgFlags, CCState &State);
132 CC_Hexagon_VarArg (unsigned ValNo, MVT ValVT,
133 MVT LocVT, CCValAssign::LocInfo LocInfo,
134 ISD::ArgFlagsTy ArgFlags, CCState &State) {
135 HexagonCCState &HState = static_cast<HexagonCCState &>(State);
137 if (ValNo < HState.getNumNamedVarArgParams()) {
138 // Deal with named arguments.
139 return CC_Hexagon(ValNo, ValVT, LocVT, LocInfo, ArgFlags, State);
142 // Deal with un-named arguments.
144 if (ArgFlags.isByVal()) {
145 // If pass-by-value, the size allocated on stack is decided
146 // by ArgFlags.getByValSize(), not by the size of LocVT.
147 ofst = State.AllocateStack(ArgFlags.getByValSize(),
148 ArgFlags.getByValAlign());
149 State.addLoc(CCValAssign::getMem(ValNo, ValVT, ofst, LocVT, LocInfo));
152 if (LocVT == MVT::i1 || LocVT == MVT::i8 || LocVT == MVT::i16) {
155 if (ArgFlags.isSExt())
156 LocInfo = CCValAssign::SExt;
157 else if (ArgFlags.isZExt())
158 LocInfo = CCValAssign::ZExt;
160 LocInfo = CCValAssign::AExt;
162 if (LocVT == MVT::i32 || LocVT == MVT::f32) {
163 ofst = State.AllocateStack(4, 4);
164 State.addLoc(CCValAssign::getMem(ValNo, ValVT, ofst, LocVT, LocInfo));
167 if (LocVT == MVT::i64 || LocVT == MVT::f64) {
168 ofst = State.AllocateStack(8, 8);
169 State.addLoc(CCValAssign::getMem(ValNo, ValVT, ofst, LocVT, LocInfo));
172 llvm_unreachable(nullptr);
177 CC_Hexagon (unsigned ValNo, MVT ValVT,
178 MVT LocVT, CCValAssign::LocInfo LocInfo,
179 ISD::ArgFlagsTy ArgFlags, CCState &State) {
181 if (ArgFlags.isByVal()) {
183 unsigned Offset = State.AllocateStack(ArgFlags.getByValSize(),
184 ArgFlags.getByValAlign());
185 State.addLoc(CCValAssign::getMem(ValNo, ValVT, Offset, LocVT, LocInfo));
189 if (LocVT == MVT::i1 || LocVT == MVT::i8 || LocVT == MVT::i16) {
192 if (ArgFlags.isSExt())
193 LocInfo = CCValAssign::SExt;
194 else if (ArgFlags.isZExt())
195 LocInfo = CCValAssign::ZExt;
197 LocInfo = CCValAssign::AExt;
198 } else if (LocVT == MVT::v4i8 || LocVT == MVT::v2i16) {
200 LocInfo = CCValAssign::BCvt;
201 } else if (LocVT == MVT::v8i8 || LocVT == MVT::v4i16 || LocVT == MVT::v2i32) {
203 LocInfo = CCValAssign::BCvt;
206 if (LocVT == MVT::i32 || LocVT == MVT::f32) {
207 if (!CC_Hexagon32(ValNo, ValVT, LocVT, LocInfo, ArgFlags, State))
211 if (LocVT == MVT::i64 || LocVT == MVT::f64) {
212 if (!CC_Hexagon64(ValNo, ValVT, LocVT, LocInfo, ArgFlags, State))
216 return true; // CC didn't match.
220 static bool CC_Hexagon32(unsigned ValNo, MVT ValVT,
221 MVT LocVT, CCValAssign::LocInfo LocInfo,
222 ISD::ArgFlagsTy ArgFlags, CCState &State) {
224 static const MCPhysReg RegList[] = {
225 Hexagon::R0, Hexagon::R1, Hexagon::R2, Hexagon::R3, Hexagon::R4,
228 if (unsigned Reg = State.AllocateReg(RegList)) {
229 State.addLoc(CCValAssign::getReg(ValNo, ValVT, Reg, LocVT, LocInfo));
233 unsigned Offset = State.AllocateStack(4, 4);
234 State.addLoc(CCValAssign::getMem(ValNo, ValVT, Offset, LocVT, LocInfo));
238 static bool CC_Hexagon64(unsigned ValNo, MVT ValVT,
239 MVT LocVT, CCValAssign::LocInfo LocInfo,
240 ISD::ArgFlagsTy ArgFlags, CCState &State) {
242 if (unsigned Reg = State.AllocateReg(Hexagon::D0)) {
243 State.addLoc(CCValAssign::getReg(ValNo, ValVT, Reg, LocVT, LocInfo));
247 static const MCPhysReg RegList1[] = {
248 Hexagon::D1, Hexagon::D2
250 static const MCPhysReg RegList2[] = {
251 Hexagon::R1, Hexagon::R3
253 if (unsigned Reg = State.AllocateReg(RegList1, RegList2)) {
254 State.addLoc(CCValAssign::getReg(ValNo, ValVT, Reg, LocVT, LocInfo));
258 unsigned Offset = State.AllocateStack(8, 8, Hexagon::D2);
259 State.addLoc(CCValAssign::getMem(ValNo, ValVT, Offset, LocVT, LocInfo));
263 static bool RetCC_Hexagon(unsigned ValNo, MVT ValVT,
264 MVT LocVT, CCValAssign::LocInfo LocInfo,
265 ISD::ArgFlagsTy ArgFlags, CCState &State) {
268 if (LocVT == MVT::i1 ||
273 if (ArgFlags.isSExt())
274 LocInfo = CCValAssign::SExt;
275 else if (ArgFlags.isZExt())
276 LocInfo = CCValAssign::ZExt;
278 LocInfo = CCValAssign::AExt;
279 } else if (LocVT == MVT::v4i8 || LocVT == MVT::v2i16) {
281 LocInfo = CCValAssign::BCvt;
282 } else if (LocVT == MVT::v8i8 || LocVT == MVT::v4i16 || LocVT == MVT::v2i32) {
284 LocInfo = CCValAssign::BCvt;
287 if (LocVT == MVT::i32 || LocVT == MVT::f32) {
288 if (!RetCC_Hexagon32(ValNo, ValVT, LocVT, LocInfo, ArgFlags, State))
292 if (LocVT == MVT::i64 || LocVT == MVT::f64) {
293 if (!RetCC_Hexagon64(ValNo, ValVT, LocVT, LocInfo, ArgFlags, State))
297 return true; // CC didn't match.
300 static bool RetCC_Hexagon32(unsigned ValNo, MVT ValVT,
301 MVT LocVT, CCValAssign::LocInfo LocInfo,
302 ISD::ArgFlagsTy ArgFlags, CCState &State) {
304 if (LocVT == MVT::i32 || LocVT == MVT::f32) {
305 if (unsigned Reg = State.AllocateReg(Hexagon::R0)) {
306 State.addLoc(CCValAssign::getReg(ValNo, ValVT, Reg, LocVT, LocInfo));
311 unsigned Offset = State.AllocateStack(4, 4);
312 State.addLoc(CCValAssign::getMem(ValNo, ValVT, Offset, LocVT, LocInfo));
316 static bool RetCC_Hexagon64(unsigned ValNo, MVT ValVT,
317 MVT LocVT, CCValAssign::LocInfo LocInfo,
318 ISD::ArgFlagsTy ArgFlags, CCState &State) {
319 if (LocVT == MVT::i64 || LocVT == MVT::f64) {
320 if (unsigned Reg = State.AllocateReg(Hexagon::D0)) {
321 State.addLoc(CCValAssign::getReg(ValNo, ValVT, Reg, LocVT, LocInfo));
326 unsigned Offset = State.AllocateStack(8, 8);
327 State.addLoc(CCValAssign::getMem(ValNo, ValVT, Offset, LocVT, LocInfo));
332 HexagonTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG)
337 /// CreateCopyOfByValArgument - Make a copy of an aggregate at address specified
338 /// by "Src" to address "Dst" of size "Size". Alignment information is
339 /// specified by the specific parameter attribute. The copy will be passed as
340 /// a byval function parameter. Sometimes what we are copying is the end of a
341 /// larger object, the part that does not fit in registers.
343 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
344 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
347 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), dl, MVT::i32);
348 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
349 /*isVolatile=*/false, /*AlwaysInline=*/false,
350 /*isTailCall=*/false,
351 MachinePointerInfo(), MachinePointerInfo());
355 // LowerReturn - Lower ISD::RET. If a struct is larger than 8 bytes and is
356 // passed by value, the function prototype is modified to return void and
357 // the value is stored in memory pointed by a pointer passed by caller.
359 HexagonTargetLowering::LowerReturn(SDValue Chain,
360 CallingConv::ID CallConv, bool isVarArg,
361 const SmallVectorImpl<ISD::OutputArg> &Outs,
362 const SmallVectorImpl<SDValue> &OutVals,
363 SDLoc dl, SelectionDAG &DAG) const {
365 // CCValAssign - represent the assignment of the return value to locations.
366 SmallVector<CCValAssign, 16> RVLocs;
368 // CCState - Info about the registers and stack slot.
369 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), RVLocs,
372 // Analyze return values of ISD::RET
373 CCInfo.AnalyzeReturn(Outs, RetCC_Hexagon);
376 SmallVector<SDValue, 4> RetOps(1, Chain);
378 // Copy the result values into the output registers.
379 for (unsigned i = 0; i != RVLocs.size(); ++i) {
380 CCValAssign &VA = RVLocs[i];
382 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), OutVals[i], Flag);
384 // Guarantee that all emitted copies are stuck together with flags.
385 Flag = Chain.getValue(1);
386 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
389 RetOps[0] = Chain; // Update chain.
391 // Add the flag if we have it.
393 RetOps.push_back(Flag);
395 return DAG.getNode(HexagonISD::RET_FLAG, dl, MVT::Other, RetOps);
398 bool HexagonTargetLowering::mayBeEmittedAsTailCall(CallInst *CI) const {
399 // If either no tail call or told not to tail call at all, don't.
401 CI->getParent()->getParent()->getFnAttribute("disable-tail-calls");
402 if (!CI->isTailCall() || Attr.getValueAsString() == "true")
408 /// LowerCallResult - Lower the result values of an ISD::CALL into the
409 /// appropriate copies out of appropriate physical registers. This assumes that
410 /// Chain/InFlag are the input chain/flag to use, and that TheCall is the call
411 /// being lowered. Returns a SDNode with the same number of values as the
414 HexagonTargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
415 CallingConv::ID CallConv, bool isVarArg,
417 SmallVectorImpl<ISD::InputArg> &Ins,
418 SDLoc dl, SelectionDAG &DAG,
419 SmallVectorImpl<SDValue> &InVals,
420 const SmallVectorImpl<SDValue> &OutVals,
421 SDValue Callee) const {
423 // Assign locations to each value returned by this call.
424 SmallVector<CCValAssign, 16> RVLocs;
426 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), RVLocs,
429 CCInfo.AnalyzeCallResult(Ins, RetCC_Hexagon);
431 // Copy all of the result registers out of their specified physreg.
432 for (unsigned i = 0; i != RVLocs.size(); ++i) {
433 Chain = DAG.getCopyFromReg(Chain, dl,
434 RVLocs[i].getLocReg(),
435 RVLocs[i].getValVT(), InFlag).getValue(1);
436 InFlag = Chain.getValue(2);
437 InVals.push_back(Chain.getValue(0));
443 /// LowerCall - Functions arguments are copied from virtual regs to
444 /// (physical regs)/(stack frame), CALLSEQ_START and CALLSEQ_END are emitted.
446 HexagonTargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI,
447 SmallVectorImpl<SDValue> &InVals) const {
448 SelectionDAG &DAG = CLI.DAG;
450 SmallVectorImpl<ISD::OutputArg> &Outs = CLI.Outs;
451 SmallVectorImpl<SDValue> &OutVals = CLI.OutVals;
452 SmallVectorImpl<ISD::InputArg> &Ins = CLI.Ins;
453 SDValue Chain = CLI.Chain;
454 SDValue Callee = CLI.Callee;
455 bool &isTailCall = CLI.IsTailCall;
456 CallingConv::ID CallConv = CLI.CallConv;
457 bool isVarArg = CLI.IsVarArg;
458 bool doesNotReturn = CLI.DoesNotReturn;
460 bool IsStructRet = (Outs.empty()) ? false : Outs[0].Flags.isSRet();
461 MachineFunction &MF = DAG.getMachineFunction();
462 auto PtrVT = getPointerTy(MF.getDataLayout());
464 // Check for varargs.
465 int NumNamedVarArgParams = -1;
466 if (GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(Callee))
468 const Function* CalleeFn = nullptr;
469 Callee = DAG.getTargetGlobalAddress(GA->getGlobal(), dl, MVT::i32);
470 if ((CalleeFn = dyn_cast<Function>(GA->getGlobal())))
472 // If a function has zero args and is a vararg function, that's
473 // disallowed so it must be an undeclared function. Do not assume
474 // varargs if the callee is undefined.
475 if (CalleeFn->isVarArg() &&
476 CalleeFn->getFunctionType()->getNumParams() != 0) {
477 NumNamedVarArgParams = CalleeFn->getFunctionType()->getNumParams();
482 // Analyze operands of the call, assigning locations to each operand.
483 SmallVector<CCValAssign, 16> ArgLocs;
484 HexagonCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), ArgLocs,
485 *DAG.getContext(), NumNamedVarArgParams);
488 CCInfo.AnalyzeCallOperands(Outs, CC_Hexagon_VarArg);
490 CCInfo.AnalyzeCallOperands(Outs, CC_Hexagon);
492 auto Attr = MF.getFunction()->getFnAttribute("disable-tail-calls");
493 if (Attr.getValueAsString() == "true")
497 bool StructAttrFlag = MF.getFunction()->hasStructRetAttr();
498 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv,
499 isVarArg, IsStructRet,
501 Outs, OutVals, Ins, DAG);
502 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
503 CCValAssign &VA = ArgLocs[i];
509 DEBUG(dbgs() << (isTailCall ? "Eligible for Tail Call\n"
510 : "Argument must be passed on stack. "
511 "Not eligible for Tail Call\n"));
513 // Get a count of how many bytes are to be pushed on the stack.
514 unsigned NumBytes = CCInfo.getNextStackOffset();
515 SmallVector<std::pair<unsigned, SDValue>, 16> RegsToPass;
516 SmallVector<SDValue, 8> MemOpChains;
518 auto &HRI = *Subtarget.getRegisterInfo();
520 DAG.getCopyFromReg(Chain, dl, HRI.getStackRegister(), PtrVT);
522 // Walk the register/memloc assignments, inserting copies/loads.
523 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
524 CCValAssign &VA = ArgLocs[i];
525 SDValue Arg = OutVals[i];
526 ISD::ArgFlagsTy Flags = Outs[i].Flags;
528 // Promote the value if needed.
529 switch (VA.getLocInfo()) {
531 // Loc info must be one of Full, SExt, ZExt, or AExt.
532 llvm_unreachable("Unknown loc info!");
533 case CCValAssign::BCvt:
534 case CCValAssign::Full:
536 case CCValAssign::SExt:
537 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), Arg);
539 case CCValAssign::ZExt:
540 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), Arg);
542 case CCValAssign::AExt:
543 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), Arg);
548 unsigned LocMemOffset = VA.getLocMemOffset();
549 SDValue MemAddr = DAG.getConstant(LocMemOffset, dl,
550 StackPtr.getValueType());
551 MemAddr = DAG.getNode(ISD::ADD, dl, MVT::i32, StackPtr, MemAddr);
552 if (Flags.isByVal()) {
553 // The argument is a struct passed by value. According to LLVM, "Arg"
555 MemOpChains.push_back(CreateCopyOfByValArgument(Arg, MemAddr, Chain,
558 MachinePointerInfo LocPI = MachinePointerInfo::getStack(LocMemOffset);
559 SDValue S = DAG.getStore(Chain, dl, Arg, MemAddr, LocPI, false,
561 MemOpChains.push_back(S);
566 // Arguments that can be passed on register must be kept at RegsToPass
569 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
572 // Transform all store nodes into one single node because all store
573 // nodes are independent of each other.
574 if (!MemOpChains.empty())
575 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOpChains);
578 SDValue C = DAG.getConstant(NumBytes, dl, PtrVT, true);
579 Chain = DAG.getCALLSEQ_START(Chain, C, dl);
582 // Build a sequence of copy-to-reg nodes chained together with token
583 // chain and flag operands which copy the outgoing args into registers.
584 // The InFlag in necessary since all emitted instructions must be
588 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
589 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
590 RegsToPass[i].second, InFlag);
591 InFlag = Chain.getValue(1);
594 // For tail calls lower the arguments to the 'real' stack slot.
596 // Force all the incoming stack arguments to be loaded from the stack
597 // before any new outgoing arguments are stored to the stack, because the
598 // outgoing stack slots may alias the incoming argument stack slots, and
599 // the alias isn't otherwise explicit. This is slightly more conservative
600 // than necessary, because it means that each store effectively depends
601 // on every argument instead of just those arguments it would clobber.
603 // Do not flag preceding copytoreg stuff together with the following stuff.
605 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
606 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
607 RegsToPass[i].second, InFlag);
608 InFlag = Chain.getValue(1);
613 // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every
614 // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol
615 // node so that legalize doesn't hack it.
616 if (flag_aligned_memcpy) {
617 const char *MemcpyName =
618 "__hexagon_memcpy_likely_aligned_min32bytes_mult8bytes";
619 Callee = DAG.getTargetExternalSymbol(MemcpyName, PtrVT);
620 flag_aligned_memcpy = false;
621 } else if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
622 Callee = DAG.getTargetGlobalAddress(G->getGlobal(), dl, PtrVT);
623 } else if (ExternalSymbolSDNode *S =
624 dyn_cast<ExternalSymbolSDNode>(Callee)) {
625 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), PtrVT);
628 // Returns a chain & a flag for retval copy to use.
629 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
630 SmallVector<SDValue, 8> Ops;
631 Ops.push_back(Chain);
632 Ops.push_back(Callee);
634 // Add argument registers to the end of the list so that they are
635 // known live into the call.
636 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
637 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
638 RegsToPass[i].second.getValueType()));
641 if (InFlag.getNode())
642 Ops.push_back(InFlag);
645 MF.getFrameInfo()->setHasTailCall();
646 return DAG.getNode(HexagonISD::TC_RETURN, dl, NodeTys, Ops);
649 int OpCode = doesNotReturn ? HexagonISD::CALLv3nr : HexagonISD::CALLv3;
650 Chain = DAG.getNode(OpCode, dl, NodeTys, Ops);
651 InFlag = Chain.getValue(1);
653 // Create the CALLSEQ_END node.
654 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, dl, true),
655 DAG.getIntPtrConstant(0, dl, true), InFlag, dl);
656 InFlag = Chain.getValue(1);
658 // Handle result values, copying them out of physregs into vregs that we
660 return LowerCallResult(Chain, InFlag, CallConv, isVarArg, Ins, dl, DAG,
661 InVals, OutVals, Callee);
664 static bool getIndexedAddressParts(SDNode *Ptr, EVT VT,
665 bool isSEXTLoad, SDValue &Base,
666 SDValue &Offset, bool &isInc,
668 if (Ptr->getOpcode() != ISD::ADD)
671 if (VT == MVT::i64 || VT == MVT::i32 || VT == MVT::i16 || VT == MVT::i8) {
672 isInc = (Ptr->getOpcode() == ISD::ADD);
673 Base = Ptr->getOperand(0);
674 Offset = Ptr->getOperand(1);
675 // Ensure that Offset is a constant.
676 return (isa<ConstantSDNode>(Offset));
682 // TODO: Put this function along with the other isS* functions in
683 // HexagonISelDAGToDAG.cpp into a common file. Or better still, use the
684 // functions defined in HexagonOperands.td.
685 static bool Is_PostInc_S4_Offset(SDNode * S, int ShiftAmount) {
686 ConstantSDNode *N = cast<ConstantSDNode>(S);
688 // immS4 predicate - True if the immediate fits in a 4-bit sign extended.
690 int64_t v = (int64_t)N->getSExtValue();
692 if (ShiftAmount > 0) {
694 v = v >> ShiftAmount;
696 return (v <= 7) && (v >= -8) && (m == 0);
699 /// getPostIndexedAddressParts - returns true by value, base pointer and
700 /// offset pointer and addressing mode by reference if this node can be
701 /// combined with a load / store to form a post-indexed load / store.
702 bool HexagonTargetLowering::getPostIndexedAddressParts(SDNode *N, SDNode *Op,
705 ISD::MemIndexedMode &AM,
706 SelectionDAG &DAG) const
710 bool isSEXTLoad = false;
712 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
713 VT = LD->getMemoryVT();
714 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD;
715 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
716 VT = ST->getMemoryVT();
717 if (ST->getValue().getValueType() == MVT::i64 && ST->isTruncatingStore()) {
725 bool isLegal = getIndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
727 // ShiftAmount = number of left-shifted bits in the Hexagon instruction.
728 int ShiftAmount = VT.getSizeInBits() / 16;
729 if (isLegal && Is_PostInc_S4_Offset(Offset.getNode(), ShiftAmount)) {
730 AM = isInc ? ISD::POST_INC : ISD::POST_DEC;
737 SDValue HexagonTargetLowering::LowerINLINEASM(SDValue Op,
738 SelectionDAG &DAG) const {
739 SDNode *Node = Op.getNode();
740 MachineFunction &MF = DAG.getMachineFunction();
741 auto &FuncInfo = *MF.getInfo<HexagonMachineFunctionInfo>();
742 switch (Node->getOpcode()) {
743 case ISD::INLINEASM: {
744 unsigned NumOps = Node->getNumOperands();
745 if (Node->getOperand(NumOps-1).getValueType() == MVT::Glue)
746 --NumOps; // Ignore the flag operand.
748 for (unsigned i = InlineAsm::Op_FirstOperand; i != NumOps;) {
749 if (FuncInfo.hasClobberLR())
752 cast<ConstantSDNode>(Node->getOperand(i))->getZExtValue();
753 unsigned NumVals = InlineAsm::getNumOperandRegisters(Flags);
754 ++i; // Skip the ID value.
756 switch (InlineAsm::getKind(Flags)) {
757 default: llvm_unreachable("Bad flags!");
758 case InlineAsm::Kind_RegDef:
759 case InlineAsm::Kind_RegUse:
760 case InlineAsm::Kind_Imm:
761 case InlineAsm::Kind_Clobber:
762 case InlineAsm::Kind_Mem: {
763 for (; NumVals; --NumVals, ++i) {}
766 case InlineAsm::Kind_RegDefEarlyClobber: {
767 for (; NumVals; --NumVals, ++i) {
769 cast<RegisterSDNode>(Node->getOperand(i))->getReg();
772 const HexagonRegisterInfo *QRI = Subtarget.getRegisterInfo();
773 if (Reg == QRI->getRARegister()) {
774 FuncInfo.setHasClobberLR(true);
789 // Taken from the XCore backend.
791 SDValue HexagonTargetLowering::
792 LowerBR_JT(SDValue Op, SelectionDAG &DAG) const
794 SDValue Chain = Op.getOperand(0);
795 SDValue Table = Op.getOperand(1);
796 SDValue Index = Op.getOperand(2);
798 JumpTableSDNode *JT = cast<JumpTableSDNode>(Table);
799 unsigned JTI = JT->getIndex();
800 MachineFunction &MF = DAG.getMachineFunction();
801 const MachineJumpTableInfo *MJTI = MF.getJumpTableInfo();
802 SDValue TargetJT = DAG.getTargetJumpTable(JT->getIndex(), MVT::i32);
804 // Mark all jump table targets as address taken.
805 const std::vector<MachineJumpTableEntry> &JTE = MJTI->getJumpTables();
806 const std::vector<MachineBasicBlock*> &JTBBs = JTE[JTI].MBBs;
807 for (unsigned i = 0, e = JTBBs.size(); i != e; ++i) {
808 MachineBasicBlock *MBB = JTBBs[i];
809 MBB->setHasAddressTaken();
810 // This line is needed to set the hasAddressTaken flag on the BasicBlock
812 BlockAddress::get(const_cast<BasicBlock *>(MBB->getBasicBlock()));
815 SDValue JumpTableBase = DAG.getNode(
816 HexagonISD::JT, dl, getPointerTy(DAG.getDataLayout()), TargetJT);
817 SDValue ShiftIndex = DAG.getNode(ISD::SHL, dl, MVT::i32, Index,
818 DAG.getConstant(2, dl, MVT::i32));
819 SDValue JTAddress = DAG.getNode(ISD::ADD, dl, MVT::i32, JumpTableBase,
821 SDValue LoadTarget = DAG.getLoad(MVT::i32, dl, Chain, JTAddress,
822 MachinePointerInfo(), false, false, false,
824 return DAG.getNode(HexagonISD::BR_JT, dl, MVT::Other, Chain, LoadTarget);
829 HexagonTargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
830 SelectionDAG &DAG) const {
831 SDValue Chain = Op.getOperand(0);
832 SDValue Size = Op.getOperand(1);
833 SDValue Align = Op.getOperand(2);
836 ConstantSDNode *AlignConst = dyn_cast<ConstantSDNode>(Align);
837 assert(AlignConst && "Non-constant Align in LowerDYNAMIC_STACKALLOC");
839 unsigned A = AlignConst->getSExtValue();
840 auto &HFI = *Subtarget.getFrameLowering();
841 // "Zero" means natural stack alignment.
843 A = HFI.getStackAlignment();
846 dbgs () << LLVM_FUNCTION_NAME << " Align: " << A << " Size: ";
847 Size.getNode()->dump(&DAG);
851 SDValue AC = DAG.getConstant(A, dl, MVT::i32);
852 SDVTList VTs = DAG.getVTList(MVT::i32, MVT::Other);
853 return DAG.getNode(HexagonISD::ALLOCA, dl, VTs, Chain, Size, AC);
857 HexagonTargetLowering::LowerFormalArguments(SDValue Chain,
858 CallingConv::ID CallConv,
861 SmallVectorImpl<ISD::InputArg> &Ins,
862 SDLoc dl, SelectionDAG &DAG,
863 SmallVectorImpl<SDValue> &InVals)
866 MachineFunction &MF = DAG.getMachineFunction();
867 MachineFrameInfo *MFI = MF.getFrameInfo();
868 MachineRegisterInfo &RegInfo = MF.getRegInfo();
869 auto &FuncInfo = *MF.getInfo<HexagonMachineFunctionInfo>();
871 // Assign locations to all of the incoming arguments.
872 SmallVector<CCValAssign, 16> ArgLocs;
873 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), ArgLocs,
876 CCInfo.AnalyzeFormalArguments(Ins, CC_Hexagon);
878 // For LLVM, in the case when returning a struct by value (>8byte),
879 // the first argument is a pointer that points to the location on caller's
880 // stack where the return value will be stored. For Hexagon, the location on
881 // caller's stack is passed only when the struct size is smaller than (and
882 // equal to) 8 bytes. If not, no address will be passed into callee and
883 // callee return the result direclty through R0/R1.
885 SmallVector<SDValue, 4> MemOps;
887 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
888 CCValAssign &VA = ArgLocs[i];
889 ISD::ArgFlagsTy Flags = Ins[i].Flags;
891 unsigned StackLocation;
894 if ( (VA.isRegLoc() && !Flags.isByVal())
895 || (VA.isRegLoc() && Flags.isByVal() && Flags.getByValSize() > 8)) {
896 // Arguments passed in registers
897 // 1. int, long long, ptr args that get allocated in register.
898 // 2. Large struct that gets an register to put its address in.
899 EVT RegVT = VA.getLocVT();
900 if (RegVT == MVT::i8 || RegVT == MVT::i16 ||
901 RegVT == MVT::i32 || RegVT == MVT::f32) {
903 RegInfo.createVirtualRegister(&Hexagon::IntRegsRegClass);
904 RegInfo.addLiveIn(VA.getLocReg(), VReg);
905 InVals.push_back(DAG.getCopyFromReg(Chain, dl, VReg, RegVT));
906 } else if (RegVT == MVT::i64 || RegVT == MVT::f64) {
908 RegInfo.createVirtualRegister(&Hexagon::DoubleRegsRegClass);
909 RegInfo.addLiveIn(VA.getLocReg(), VReg);
910 InVals.push_back(DAG.getCopyFromReg(Chain, dl, VReg, RegVT));
914 } else if (VA.isRegLoc() && Flags.isByVal() && Flags.getByValSize() <= 8) {
915 assert (0 && "ByValSize must be bigger than 8 bytes");
918 assert(VA.isMemLoc());
920 if (Flags.isByVal()) {
921 // If it's a byval parameter, then we need to compute the
922 // "real" size, not the size of the pointer.
923 ObjSize = Flags.getByValSize();
925 ObjSize = VA.getLocVT().getStoreSizeInBits() >> 3;
928 StackLocation = HEXAGON_LRFP_SIZE + VA.getLocMemOffset();
929 // Create the frame index object for this incoming parameter...
930 FI = MFI->CreateFixedObject(ObjSize, StackLocation, true);
932 // Create the SelectionDAG nodes cordl, responding to a load
933 // from this parameter.
934 SDValue FIN = DAG.getFrameIndex(FI, MVT::i32);
936 if (Flags.isByVal()) {
937 // If it's a pass-by-value aggregate, then do not dereference the stack
938 // location. Instead, we should generate a reference to the stack
940 InVals.push_back(FIN);
942 InVals.push_back(DAG.getLoad(VA.getLocVT(), dl, Chain, FIN,
943 MachinePointerInfo(), false, false,
950 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOps);
953 // This will point to the next argument passed via stack.
954 int FrameIndex = MFI->CreateFixedObject(Hexagon_PointerSize,
956 CCInfo.getNextStackOffset(),
958 FuncInfo.setVarArgsFrameIndex(FrameIndex);
965 HexagonTargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG) const {
966 // VASTART stores the address of the VarArgsFrameIndex slot into the
967 // memory location argument.
968 MachineFunction &MF = DAG.getMachineFunction();
969 HexagonMachineFunctionInfo *QFI = MF.getInfo<HexagonMachineFunctionInfo>();
970 SDValue Addr = DAG.getFrameIndex(QFI->getVarArgsFrameIndex(), MVT::i32);
971 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
972 return DAG.getStore(Op.getOperand(0), SDLoc(Op), Addr,
973 Op.getOperand(1), MachinePointerInfo(SV), false,
977 // Creates a SPLAT instruction for a constant value VAL.
978 static SDValue createSplat(SelectionDAG &DAG, SDLoc dl, EVT VT, SDValue Val) {
979 if (VT.getSimpleVT() == MVT::v4i8)
980 return DAG.getNode(HexagonISD::VSPLATB, dl, VT, Val);
982 if (VT.getSimpleVT() == MVT::v4i16)
983 return DAG.getNode(HexagonISD::VSPLATH, dl, VT, Val);
988 static bool isSExtFree(SDValue N) {
989 // A sign-extend of a truncate of a sign-extend is free.
990 if (N.getOpcode() == ISD::TRUNCATE &&
991 N.getOperand(0).getOpcode() == ISD::AssertSext)
993 // We have sign-extended loads.
994 if (N.getOpcode() == ISD::LOAD)
999 SDValue HexagonTargetLowering::LowerCTPOP(SDValue Op, SelectionDAG &DAG) const {
1001 SDValue InpVal = Op.getOperand(0);
1002 if (isa<ConstantSDNode>(InpVal)) {
1003 uint64_t V = cast<ConstantSDNode>(InpVal)->getZExtValue();
1004 return DAG.getTargetConstant(countPopulation(V), dl, MVT::i64);
1006 SDValue PopOut = DAG.getNode(HexagonISD::POPCOUNT, dl, MVT::i32, InpVal);
1007 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i64, PopOut);
1010 SDValue HexagonTargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) const {
1013 SDValue LHS = Op.getOperand(0);
1014 SDValue RHS = Op.getOperand(1);
1015 SDValue Cmp = Op.getOperand(2);
1016 ISD::CondCode CC = cast<CondCodeSDNode>(Cmp)->get();
1018 EVT VT = Op.getValueType();
1019 EVT LHSVT = LHS.getValueType();
1020 EVT RHSVT = RHS.getValueType();
1022 if (LHSVT == MVT::v2i16) {
1023 assert(ISD::isSignedIntSetCC(CC) || ISD::isUnsignedIntSetCC(CC));
1024 unsigned ExtOpc = ISD::isSignedIntSetCC(CC) ? ISD::SIGN_EXTEND
1026 SDValue LX = DAG.getNode(ExtOpc, dl, MVT::v2i32, LHS);
1027 SDValue RX = DAG.getNode(ExtOpc, dl, MVT::v2i32, RHS);
1028 SDValue SC = DAG.getNode(ISD::SETCC, dl, MVT::v2i1, LX, RX, Cmp);
1032 // Treat all other vector types as legal.
1036 // Equals and not equals should use sign-extend, not zero-extend, since
1037 // we can represent small negative values in the compare instructions.
1038 // The LLVM default is to use zero-extend arbitrarily in these cases.
1039 if ((CC == ISD::SETEQ || CC == ISD::SETNE) &&
1040 (RHSVT == MVT::i8 || RHSVT == MVT::i16) &&
1041 (LHSVT == MVT::i8 || LHSVT == MVT::i16)) {
1042 ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS);
1043 if (C && C->getAPIntValue().isNegative()) {
1044 LHS = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::i32, LHS);
1045 RHS = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::i32, RHS);
1046 return DAG.getNode(ISD::SETCC, dl, Op.getValueType(),
1047 LHS, RHS, Op.getOperand(2));
1049 if (isSExtFree(LHS) || isSExtFree(RHS)) {
1050 LHS = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::i32, LHS);
1051 RHS = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::i32, RHS);
1052 return DAG.getNode(ISD::SETCC, dl, Op.getValueType(),
1053 LHS, RHS, Op.getOperand(2));
1059 SDValue HexagonTargetLowering::LowerVSELECT(SDValue Op, SelectionDAG &DAG)
1061 SDValue PredOp = Op.getOperand(0);
1062 SDValue Op1 = Op.getOperand(1), Op2 = Op.getOperand(2);
1063 EVT OpVT = Op1.getValueType();
1066 if (OpVT == MVT::v2i16) {
1067 SDValue X1 = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::v2i32, Op1);
1068 SDValue X2 = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::v2i32, Op2);
1069 SDValue SL = DAG.getNode(ISD::VSELECT, DL, MVT::v2i32, PredOp, X1, X2);
1070 SDValue TR = DAG.getNode(ISD::TRUNCATE, DL, MVT::v2i16, SL);
1077 // Handle only specific vector loads.
1078 SDValue HexagonTargetLowering::LowerLOAD(SDValue Op, SelectionDAG &DAG) const {
1079 EVT VT = Op.getValueType();
1081 LoadSDNode *LoadNode = cast<LoadSDNode>(Op);
1082 SDValue Chain = LoadNode->getChain();
1083 SDValue Ptr = Op.getOperand(1);
1084 SDValue LoweredLoad;
1086 SDValue Base = LoadNode->getBasePtr();
1087 ISD::LoadExtType Ext = LoadNode->getExtensionType();
1088 unsigned Alignment = LoadNode->getAlignment();
1091 if(Ext == ISD::NON_EXTLOAD)
1092 Ext = ISD::ZEXTLOAD;
1094 if (VT == MVT::v4i16) {
1095 if (Alignment == 2) {
1098 Loads[0] = DAG.getExtLoad(Ext, DL, MVT::i32, Chain, Base,
1099 LoadNode->getPointerInfo(), MVT::i16,
1100 LoadNode->isVolatile(),
1101 LoadNode->isNonTemporal(),
1102 LoadNode->isInvariant(),
1105 SDValue Increment = DAG.getConstant(2, DL, MVT::i32);
1106 Ptr = DAG.getNode(ISD::ADD, DL, Base.getValueType(), Base, Increment);
1107 Loads[1] = DAG.getExtLoad(Ext, DL, MVT::i32, Chain, Ptr,
1108 LoadNode->getPointerInfo(), MVT::i16,
1109 LoadNode->isVolatile(),
1110 LoadNode->isNonTemporal(),
1111 LoadNode->isInvariant(),
1113 // SHL 16, then OR base and base+2.
1114 SDValue ShiftAmount = DAG.getConstant(16, DL, MVT::i32);
1115 SDValue Tmp1 = DAG.getNode(ISD::SHL, DL, MVT::i32, Loads[1], ShiftAmount);
1116 SDValue Tmp2 = DAG.getNode(ISD::OR, DL, MVT::i32, Tmp1, Loads[0]);
1118 Increment = DAG.getConstant(4, DL, MVT::i32);
1119 Ptr = DAG.getNode(ISD::ADD, DL, Base.getValueType(), Base, Increment);
1120 Loads[2] = DAG.getExtLoad(Ext, DL, MVT::i32, Chain, Ptr,
1121 LoadNode->getPointerInfo(), MVT::i16,
1122 LoadNode->isVolatile(),
1123 LoadNode->isNonTemporal(),
1124 LoadNode->isInvariant(),
1127 Increment = DAG.getConstant(6, DL, MVT::i32);
1128 Ptr = DAG.getNode(ISD::ADD, DL, Base.getValueType(), Base, Increment);
1129 Loads[3] = DAG.getExtLoad(Ext, DL, MVT::i32, Chain, Ptr,
1130 LoadNode->getPointerInfo(), MVT::i16,
1131 LoadNode->isVolatile(),
1132 LoadNode->isNonTemporal(),
1133 LoadNode->isInvariant(),
1135 // SHL 16, then OR base+4 and base+6.
1136 Tmp1 = DAG.getNode(ISD::SHL, DL, MVT::i32, Loads[3], ShiftAmount);
1137 SDValue Tmp4 = DAG.getNode(ISD::OR, DL, MVT::i32, Tmp1, Loads[2]);
1138 // Combine to i64. This could be optimised out later if we can
1139 // affect reg allocation of this code.
1140 Result = DAG.getNode(HexagonISD::COMBINE, DL, MVT::i64, Tmp4, Tmp2);
1141 LoadChain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other,
1142 Loads[0].getValue(1), Loads[1].getValue(1),
1143 Loads[2].getValue(1), Loads[3].getValue(1));
1145 // Perform default type expansion.
1146 Result = DAG.getLoad(MVT::i64, DL, Chain, Ptr, LoadNode->getPointerInfo(),
1147 LoadNode->isVolatile(), LoadNode->isNonTemporal(),
1148 LoadNode->isInvariant(), LoadNode->getAlignment());
1149 LoadChain = Result.getValue(1);
1152 llvm_unreachable("Custom lowering unsupported load");
1154 Result = DAG.getNode(ISD::BITCAST, DL, VT, Result);
1155 // Since we pretend to lower a load, we need the original chain
1156 // info attached to the result.
1157 SDValue Ops[] = { Result, LoadChain };
1159 return DAG.getMergeValues(Ops, DL);
1164 HexagonTargetLowering::LowerConstantPool(SDValue Op, SelectionDAG &DAG) const {
1165 EVT ValTy = Op.getValueType();
1167 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
1169 if (CP->isMachineConstantPoolEntry())
1170 Res = DAG.getTargetConstantPool(CP->getMachineCPVal(), ValTy,
1171 CP->getAlignment());
1173 Res = DAG.getTargetConstantPool(CP->getConstVal(), ValTy,
1174 CP->getAlignment());
1175 return DAG.getNode(HexagonISD::CP, dl, ValTy, Res);
1179 HexagonTargetLowering::LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) const {
1180 const HexagonRegisterInfo &HRI = *Subtarget.getRegisterInfo();
1181 MachineFunction &MF = DAG.getMachineFunction();
1182 MachineFrameInfo &MFI = *MF.getFrameInfo();
1183 MFI.setReturnAddressIsTaken(true);
1185 if (verifyReturnAddressArgumentIsConstant(Op, DAG))
1188 EVT VT = Op.getValueType();
1190 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
1192 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
1193 SDValue Offset = DAG.getConstant(4, dl, MVT::i32);
1194 return DAG.getLoad(VT, dl, DAG.getEntryNode(),
1195 DAG.getNode(ISD::ADD, dl, VT, FrameAddr, Offset),
1196 MachinePointerInfo(), false, false, false, 0);
1199 // Return LR, which contains the return address. Mark it an implicit live-in.
1200 unsigned Reg = MF.addLiveIn(HRI.getRARegister(), getRegClassFor(MVT::i32));
1201 return DAG.getCopyFromReg(DAG.getEntryNode(), dl, Reg, VT);
1205 HexagonTargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
1206 const HexagonRegisterInfo &HRI = *Subtarget.getRegisterInfo();
1207 MachineFrameInfo &MFI = *DAG.getMachineFunction().getFrameInfo();
1208 MFI.setFrameAddressIsTaken(true);
1210 EVT VT = Op.getValueType();
1212 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
1213 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl,
1214 HRI.getFrameRegister(), VT);
1216 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr,
1217 MachinePointerInfo(),
1218 false, false, false, 0);
1222 SDValue HexagonTargetLowering::LowerATOMIC_FENCE(SDValue Op,
1223 SelectionDAG& DAG) const {
1225 return DAG.getNode(HexagonISD::BARRIER, dl, MVT::Other, Op.getOperand(0));
1229 SDValue HexagonTargetLowering::LowerGLOBALADDRESS(SDValue Op,
1230 SelectionDAG &DAG) const {
1232 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
1233 int64_t Offset = cast<GlobalAddressSDNode>(Op)->getOffset();
1235 auto PtrVT = getPointerTy(DAG.getDataLayout());
1236 Result = DAG.getTargetGlobalAddress(GV, dl, PtrVT, Offset);
1238 const HexagonTargetObjectFile *TLOF =
1239 static_cast<const HexagonTargetObjectFile *>(
1240 getTargetMachine().getObjFileLowering());
1241 if (TLOF->IsGlobalInSmallSection(GV, getTargetMachine())) {
1242 return DAG.getNode(HexagonISD::CONST32_GP, dl, PtrVT, Result);
1245 return DAG.getNode(HexagonISD::CONST32, dl, PtrVT, Result);
1248 // Specifies that for loads and stores VT can be promoted to PromotedLdStVT.
1249 void HexagonTargetLowering::promoteLdStType(EVT VT, EVT PromotedLdStVT) {
1250 if (VT != PromotedLdStVT) {
1251 setOperationAction(ISD::LOAD, VT.getSimpleVT(), Promote);
1252 AddPromotedToType(ISD::LOAD, VT.getSimpleVT(),
1253 PromotedLdStVT.getSimpleVT());
1255 setOperationAction(ISD::STORE, VT.getSimpleVT(), Promote);
1256 AddPromotedToType(ISD::STORE, VT.getSimpleVT(),
1257 PromotedLdStVT.getSimpleVT());
1262 HexagonTargetLowering::LowerBlockAddress(SDValue Op, SelectionDAG &DAG) const {
1263 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
1264 SDValue BA_SD = DAG.getTargetBlockAddress(BA, MVT::i32);
1266 return DAG.getNode(HexagonISD::CONST32_GP, dl,
1267 getPointerTy(DAG.getDataLayout()), BA_SD);
1270 //===----------------------------------------------------------------------===//
1271 // TargetLowering Implementation
1272 //===----------------------------------------------------------------------===//
1274 HexagonTargetLowering::HexagonTargetLowering(const TargetMachine &TM,
1275 const HexagonSubtarget &STI)
1276 : TargetLowering(TM), HTM(static_cast<const HexagonTargetMachine&>(TM)),
1278 bool IsV4 = !Subtarget.hasV5TOps();
1279 auto &HRI = *Subtarget.getRegisterInfo();
1281 setPrefLoopAlignment(4);
1282 setPrefFunctionAlignment(4);
1283 setMinFunctionAlignment(2);
1284 setInsertFencesForAtomic(false);
1285 setExceptionPointerRegister(Hexagon::R0);
1286 setExceptionSelectorRegister(Hexagon::R1);
1287 setStackPointerRegisterToSaveRestore(HRI.getStackRegister());
1289 if (EnableHexSDNodeSched)
1290 setSchedulingPreference(Sched::VLIW);
1292 setSchedulingPreference(Sched::Source);
1294 // Limits for inline expansion of memcpy/memmove
1295 MaxStoresPerMemcpy = MaxStoresPerMemcpyCL;
1296 MaxStoresPerMemcpyOptSize = MaxStoresPerMemcpyOptSizeCL;
1297 MaxStoresPerMemmove = MaxStoresPerMemmoveCL;
1298 MaxStoresPerMemmoveOptSize = MaxStoresPerMemmoveOptSizeCL;
1299 MaxStoresPerMemset = MaxStoresPerMemsetCL;
1300 MaxStoresPerMemsetOptSize = MaxStoresPerMemsetOptSizeCL;
1303 // Set up register classes.
1306 addRegisterClass(MVT::i1, &Hexagon::PredRegsRegClass);
1307 addRegisterClass(MVT::v2i1, &Hexagon::PredRegsRegClass); // bbbbaaaa
1308 addRegisterClass(MVT::v4i1, &Hexagon::PredRegsRegClass); // ddccbbaa
1309 addRegisterClass(MVT::v8i1, &Hexagon::PredRegsRegClass); // hgfedcba
1310 addRegisterClass(MVT::i32, &Hexagon::IntRegsRegClass);
1311 addRegisterClass(MVT::v4i8, &Hexagon::IntRegsRegClass);
1312 addRegisterClass(MVT::v2i16, &Hexagon::IntRegsRegClass);
1313 addRegisterClass(MVT::i64, &Hexagon::DoubleRegsRegClass);
1314 addRegisterClass(MVT::v8i8, &Hexagon::DoubleRegsRegClass);
1315 addRegisterClass(MVT::v4i16, &Hexagon::DoubleRegsRegClass);
1316 addRegisterClass(MVT::v2i32, &Hexagon::DoubleRegsRegClass);
1318 if (Subtarget.hasV5TOps()) {
1319 addRegisterClass(MVT::f32, &Hexagon::IntRegsRegClass);
1320 addRegisterClass(MVT::f64, &Hexagon::DoubleRegsRegClass);
1324 // Handling of scalar operations.
1326 // All operations default to "legal", except:
1327 // - indexed loads and stores (pre-/post-incremented),
1328 // - ANY_EXTEND_VECTOR_INREG, ATOMIC_CMP_SWAP_WITH_SUCCESS, CONCAT_VECTORS,
1329 // ConstantFP, DEBUGTRAP, FCEIL, FCOPYSIGN, FEXP, FEXP2, FFLOOR, FGETSIGN,
1330 // FLOG, FLOG2, FLOG10, FMAXNUM, FMINNUM, FNEARBYINT, FRINT, FROUND, TRAP,
1331 // FTRUNC, PREFETCH, SIGN_EXTEND_VECTOR_INREG, ZERO_EXTEND_VECTOR_INREG,
1332 // which default to "expand" for at least one type.
1335 setOperationAction(ISD::ConstantFP, MVT::f32, Legal); // Default: expand
1336 setOperationAction(ISD::ConstantFP, MVT::f64, Legal); // Default: expand
1338 setOperationAction(ISD::ConstantPool, MVT::i32, Custom);
1339 setOperationAction(ISD::BUILD_PAIR, MVT::i64, Expand);
1340 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand);
1341 setOperationAction(ISD::INLINEASM, MVT::Other, Custom);
1342 setOperationAction(ISD::EH_RETURN, MVT::Other, Custom);
1343 setOperationAction(ISD::ATOMIC_FENCE, MVT::Other, Custom);
1345 // Custom legalize GlobalAddress nodes into CONST32.
1346 setOperationAction(ISD::GlobalAddress, MVT::i32, Custom);
1347 setOperationAction(ISD::GlobalAddress, MVT::i8, Custom);
1348 setOperationAction(ISD::BlockAddress, MVT::i32, Custom);
1350 // Hexagon needs to optimize cases with negative constants.
1351 setOperationAction(ISD::SETCC, MVT::i8, Custom);
1352 setOperationAction(ISD::SETCC, MVT::i16, Custom);
1354 // VASTART needs to be custom lowered to use the VarArgsFrameIndex.
1355 setOperationAction(ISD::VASTART, MVT::Other, Custom);
1356 setOperationAction(ISD::VAEND, MVT::Other, Expand);
1357 setOperationAction(ISD::VAARG, MVT::Other, Expand);
1359 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
1360 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
1361 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Custom);
1364 setOperationAction(ISD::BR_JT, MVT::Other, Custom);
1366 setOperationAction(ISD::BR_JT, MVT::Other, Expand);
1367 // Increase jump tables cutover to 5, was 4.
1368 setMinimumJumpTableEntries(MinimumJumpTables);
1370 // Hexagon has instructions for add/sub with carry. The problem with
1371 // modeling these instructions is that they produce 2 results: Rdd and Px.
1372 // To model the update of Px, we will have to use Defs[p0..p3] which will
1373 // cause any predicate live range to spill. So, we pretend we dont't have
1374 // these instructions.
1375 setOperationAction(ISD::ADDE, MVT::i8, Expand);
1376 setOperationAction(ISD::ADDE, MVT::i16, Expand);
1377 setOperationAction(ISD::ADDE, MVT::i32, Expand);
1378 setOperationAction(ISD::ADDE, MVT::i64, Expand);
1379 setOperationAction(ISD::SUBE, MVT::i8, Expand);
1380 setOperationAction(ISD::SUBE, MVT::i16, Expand);
1381 setOperationAction(ISD::SUBE, MVT::i32, Expand);
1382 setOperationAction(ISD::SUBE, MVT::i64, Expand);
1383 setOperationAction(ISD::ADDC, MVT::i8, Expand);
1384 setOperationAction(ISD::ADDC, MVT::i16, Expand);
1385 setOperationAction(ISD::ADDC, MVT::i32, Expand);
1386 setOperationAction(ISD::ADDC, MVT::i64, Expand);
1387 setOperationAction(ISD::SUBC, MVT::i8, Expand);
1388 setOperationAction(ISD::SUBC, MVT::i16, Expand);
1389 setOperationAction(ISD::SUBC, MVT::i32, Expand);
1390 setOperationAction(ISD::SUBC, MVT::i64, Expand);
1392 // Only add and sub that detect overflow are the saturating ones.
1393 for (MVT VT : MVT::integer_valuetypes()) {
1394 setOperationAction(ISD::UADDO, VT, Expand);
1395 setOperationAction(ISD::SADDO, VT, Expand);
1396 setOperationAction(ISD::USUBO, VT, Expand);
1397 setOperationAction(ISD::SSUBO, VT, Expand);
1400 setOperationAction(ISD::CTLZ, MVT::i8, Promote);
1401 setOperationAction(ISD::CTLZ, MVT::i16, Promote);
1402 setOperationAction(ISD::CTTZ, MVT::i8, Promote);
1403 setOperationAction(ISD::CTTZ, MVT::i16, Promote);
1404 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i8, Promote);
1405 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i16, Promote);
1406 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i8, Promote);
1407 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i16, Promote);
1409 // In V5, popcount can count # of 1s in i64 but returns i32.
1410 // On V4 it will be expanded (set later).
1411 setOperationAction(ISD::CTPOP, MVT::i8, Promote);
1412 setOperationAction(ISD::CTPOP, MVT::i16, Promote);
1413 setOperationAction(ISD::CTPOP, MVT::i32, Promote);
1414 setOperationAction(ISD::CTPOP, MVT::i64, Custom);
1416 // We custom lower i64 to i64 mul, so that it is not considered as a legal
1417 // operation. There is a pattern that will match i64 mul and transform it
1418 // to a series of instructions.
1419 setOperationAction(ISD::MUL, MVT::i64, Expand);
1420 setOperationAction(ISD::MULHS, MVT::i64, Expand);
1422 for (unsigned IntExpOp :
1423 {ISD::SDIV, ISD::UDIV, ISD::SREM, ISD::UREM, ISD::SDIVREM, ISD::UDIVREM,
1424 ISD::ROTL, ISD::ROTR, ISD::BSWAP, ISD::SHL_PARTS, ISD::SRA_PARTS,
1425 ISD::SRL_PARTS, ISD::SMUL_LOHI, ISD::UMUL_LOHI}) {
1426 setOperationAction(IntExpOp, MVT::i32, Expand);
1427 setOperationAction(IntExpOp, MVT::i64, Expand);
1430 for (unsigned FPExpOp :
1431 {ISD::FDIV, ISD::FREM, ISD::FSQRT, ISD::FSIN, ISD::FCOS, ISD::FSINCOS,
1432 ISD::FPOW, ISD::FCOPYSIGN}) {
1433 setOperationAction(FPExpOp, MVT::f32, Expand);
1434 setOperationAction(FPExpOp, MVT::f64, Expand);
1437 // No extending loads from i32.
1438 for (MVT VT : MVT::integer_valuetypes()) {
1439 setLoadExtAction(ISD::ZEXTLOAD, VT, MVT::i32, Expand);
1440 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::i32, Expand);
1441 setLoadExtAction(ISD::EXTLOAD, VT, MVT::i32, Expand);
1443 // Turn FP truncstore into trunc + store.
1444 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
1445 // Turn FP extload into load/fextend.
1446 for (MVT VT : MVT::fp_valuetypes())
1447 setLoadExtAction(ISD::EXTLOAD, VT, MVT::f32, Expand);
1449 // Expand BR_CC and SELECT_CC for all integer and fp types.
1450 for (MVT VT : MVT::integer_valuetypes()) {
1451 setOperationAction(ISD::BR_CC, VT, Expand);
1452 setOperationAction(ISD::SELECT_CC, VT, Expand);
1454 for (MVT VT : MVT::fp_valuetypes()) {
1455 setOperationAction(ISD::BR_CC, VT, Expand);
1456 setOperationAction(ISD::SELECT_CC, VT, Expand);
1458 setOperationAction(ISD::BR_CC, MVT::Other, Expand);
1461 // Handling of vector operations.
1464 // Custom lower v4i16 load only. Let v4i16 store to be
1465 // promoted for now.
1466 promoteLdStType(MVT::v4i8, MVT::i32);
1467 promoteLdStType(MVT::v2i16, MVT::i32);
1468 promoteLdStType(MVT::v8i8, MVT::i64);
1469 promoteLdStType(MVT::v2i32, MVT::i64);
1471 setOperationAction(ISD::LOAD, MVT::v4i16, Custom);
1472 setOperationAction(ISD::STORE, MVT::v4i16, Promote);
1473 AddPromotedToType(ISD::LOAD, MVT::v4i16, MVT::i64);
1474 AddPromotedToType(ISD::STORE, MVT::v4i16, MVT::i64);
1476 // Set the action for vector operations to "expand", then override it with
1477 // either "custom" or "legal" for specific cases.
1478 static unsigned VectExpOps[] = {
1479 // Integer arithmetic:
1480 ISD::ADD, ISD::SUB, ISD::MUL, ISD::SDIV, ISD::UDIV,
1481 ISD::SREM, ISD::UREM, ISD::SDIVREM, ISD::UDIVREM, ISD::ADDC,
1482 ISD::SUBC, ISD::SADDO, ISD::UADDO, ISD::SSUBO, ISD::USUBO,
1483 ISD::SMUL_LOHI, ISD::UMUL_LOHI,
1485 ISD::AND, ISD::OR, ISD::XOR, ISD::ROTL, ISD::ROTR,
1486 ISD::CTPOP, ISD::CTLZ, ISD::CTTZ, ISD::CTLZ_ZERO_UNDEF,
1487 ISD::CTTZ_ZERO_UNDEF,
1488 // Floating point arithmetic/math functions:
1489 ISD::FADD, ISD::FSUB, ISD::FMUL, ISD::FMA, ISD::FDIV,
1490 ISD::FREM, ISD::FNEG, ISD::FABS, ISD::FSQRT, ISD::FSIN,
1491 ISD::FCOS, ISD::FPOWI, ISD::FPOW, ISD::FLOG, ISD::FLOG2,
1492 ISD::FLOG10, ISD::FEXP, ISD::FEXP2, ISD::FCEIL, ISD::FTRUNC,
1493 ISD::FRINT, ISD::FNEARBYINT, ISD::FROUND, ISD::FFLOOR,
1494 ISD::FMINNUM, ISD::FMAXNUM, ISD::FSINCOS,
1496 ISD::SELECT, ISD::ConstantPool,
1498 ISD::BUILD_VECTOR, ISD::SCALAR_TO_VECTOR,
1499 ISD::EXTRACT_VECTOR_ELT, ISD::INSERT_VECTOR_ELT,
1500 ISD::EXTRACT_SUBVECTOR, ISD::INSERT_SUBVECTOR,
1501 ISD::CONCAT_VECTORS, ISD::VECTOR_SHUFFLE
1504 for (MVT VT : MVT::vector_valuetypes()) {
1505 for (unsigned VectExpOp : VectExpOps)
1506 setOperationAction(VectExpOp, VT, Expand);
1508 // Expand all extended loads and truncating stores:
1509 for (MVT TargetVT : MVT::vector_valuetypes()) {
1510 setLoadExtAction(ISD::EXTLOAD, TargetVT, VT, Expand);
1511 setTruncStoreAction(VT, TargetVT, Expand);
1514 setOperationAction(ISD::SRA, VT, Custom);
1515 setOperationAction(ISD::SHL, VT, Custom);
1516 setOperationAction(ISD::SRL, VT, Custom);
1519 // Types natively supported:
1520 for (MVT NativeVT : {MVT::v2i1, MVT::v4i1, MVT::v8i1, MVT::v32i1, MVT::v64i1,
1521 MVT::v4i8, MVT::v8i8, MVT::v2i16, MVT::v4i16, MVT::v1i32,
1522 MVT::v2i32, MVT::v1i64}) {
1523 setOperationAction(ISD::BUILD_VECTOR, NativeVT, Custom);
1524 setOperationAction(ISD::EXTRACT_VECTOR_ELT, NativeVT, Custom);
1525 setOperationAction(ISD::INSERT_VECTOR_ELT, NativeVT, Custom);
1526 setOperationAction(ISD::EXTRACT_SUBVECTOR, NativeVT, Custom);
1527 setOperationAction(ISD::INSERT_SUBVECTOR, NativeVT, Custom);
1528 setOperationAction(ISD::CONCAT_VECTORS, NativeVT, Custom);
1530 setOperationAction(ISD::ADD, NativeVT, Legal);
1531 setOperationAction(ISD::SUB, NativeVT, Legal);
1532 setOperationAction(ISD::MUL, NativeVT, Legal);
1533 setOperationAction(ISD::AND, NativeVT, Legal);
1534 setOperationAction(ISD::OR, NativeVT, Legal);
1535 setOperationAction(ISD::XOR, NativeVT, Legal);
1538 setOperationAction(ISD::SETCC, MVT::v2i16, Custom);
1539 setOperationAction(ISD::VSELECT, MVT::v2i16, Custom);
1540 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4i16, Custom);
1541 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v8i8, Custom);
1543 // Subtarget-specific operation actions.
1545 if (Subtarget.hasV5TOps()) {
1546 setOperationAction(ISD::FMA, MVT::f64, Expand);
1547 setOperationAction(ISD::FADD, MVT::f64, Expand);
1548 setOperationAction(ISD::FSUB, MVT::f64, Expand);
1549 setOperationAction(ISD::FMUL, MVT::f64, Expand);
1551 setOperationAction(ISD::FP_TO_UINT, MVT::i1, Promote);
1552 setOperationAction(ISD::FP_TO_UINT, MVT::i8, Promote);
1553 setOperationAction(ISD::FP_TO_UINT, MVT::i16, Promote);
1554 setOperationAction(ISD::FP_TO_SINT, MVT::i1, Promote);
1555 setOperationAction(ISD::FP_TO_SINT, MVT::i8, Promote);
1556 setOperationAction(ISD::FP_TO_SINT, MVT::i16, Promote);
1557 setOperationAction(ISD::UINT_TO_FP, MVT::i1, Promote);
1558 setOperationAction(ISD::UINT_TO_FP, MVT::i8, Promote);
1559 setOperationAction(ISD::UINT_TO_FP, MVT::i16, Promote);
1560 setOperationAction(ISD::SINT_TO_FP, MVT::i1, Promote);
1561 setOperationAction(ISD::SINT_TO_FP, MVT::i8, Promote);
1562 setOperationAction(ISD::SINT_TO_FP, MVT::i16, Promote);
1565 setOperationAction(ISD::SINT_TO_FP, MVT::i32, Expand);
1566 setOperationAction(ISD::SINT_TO_FP, MVT::i64, Expand);
1567 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Expand);
1568 setOperationAction(ISD::UINT_TO_FP, MVT::i64, Expand);
1569 setOperationAction(ISD::FP_TO_SINT, MVT::f64, Expand);
1570 setOperationAction(ISD::FP_TO_SINT, MVT::f32, Expand);
1571 setOperationAction(ISD::FP_EXTEND, MVT::f32, Expand);
1572 setOperationAction(ISD::FP_ROUND, MVT::f64, Expand);
1573 setCondCodeAction(ISD::SETUNE, MVT::f64, Expand);
1575 setOperationAction(ISD::CTPOP, MVT::i8, Expand);
1576 setOperationAction(ISD::CTPOP, MVT::i16, Expand);
1577 setOperationAction(ISD::CTPOP, MVT::i32, Expand);
1578 setOperationAction(ISD::CTPOP, MVT::i64, Expand);
1580 // Expand these operations for both f32 and f64:
1581 for (unsigned FPExpOpV4 :
1582 {ISD::FADD, ISD::FSUB, ISD::FMUL, ISD::FABS, ISD::FNEG, ISD::FMA}) {
1583 setOperationAction(FPExpOpV4, MVT::f32, Expand);
1584 setOperationAction(FPExpOpV4, MVT::f64, Expand);
1587 for (ISD::CondCode FPExpCCV4 :
1588 {ISD::SETOEQ, ISD::SETOGT, ISD::SETOLT, ISD::SETOGE, ISD::SETOLE,
1589 ISD::SETUO, ISD::SETO}) {
1590 setCondCodeAction(FPExpCCV4, MVT::f32, Expand);
1591 setCondCodeAction(FPExpCCV4, MVT::f64, Expand);
1595 // Handling of indexed loads/stores: default is "expand".
1597 for (MVT LSXTy : {MVT::i8, MVT::i16, MVT::i32, MVT::i64}) {
1598 setIndexedLoadAction(ISD::POST_INC, LSXTy, Legal);
1599 setIndexedStoreAction(ISD::POST_INC, LSXTy, Legal);
1602 computeRegisterProperties(&HRI);
1605 // Library calls for unsupported operations
1607 bool FastMath = EnableFastMath;
1609 setLibcallName(RTLIB::SDIV_I32, "__hexagon_divsi3");
1610 setLibcallName(RTLIB::SDIV_I64, "__hexagon_divdi3");
1611 setLibcallName(RTLIB::UDIV_I32, "__hexagon_udivsi3");
1612 setLibcallName(RTLIB::UDIV_I64, "__hexagon_udivdi3");
1613 setLibcallName(RTLIB::SREM_I32, "__hexagon_modsi3");
1614 setLibcallName(RTLIB::SREM_I64, "__hexagon_moddi3");
1615 setLibcallName(RTLIB::UREM_I32, "__hexagon_umodsi3");
1616 setLibcallName(RTLIB::UREM_I64, "__hexagon_umoddi3");
1618 setLibcallName(RTLIB::SINTTOFP_I128_F64, "__hexagon_floattidf");
1619 setLibcallName(RTLIB::SINTTOFP_I128_F32, "__hexagon_floattisf");
1620 setLibcallName(RTLIB::FPTOUINT_F32_I128, "__hexagon_fixunssfti");
1621 setLibcallName(RTLIB::FPTOUINT_F64_I128, "__hexagon_fixunsdfti");
1622 setLibcallName(RTLIB::FPTOSINT_F32_I128, "__hexagon_fixsfti");
1623 setLibcallName(RTLIB::FPTOSINT_F64_I128, "__hexagon_fixdfti");
1626 // Handle single-precision floating point operations on V4.
1628 setLibcallName(RTLIB::ADD_F32, "__hexagon_fast_addsf3");
1629 setLibcallName(RTLIB::SUB_F32, "__hexagon_fast_subsf3");
1630 setLibcallName(RTLIB::MUL_F32, "__hexagon_fast_mulsf3");
1631 setLibcallName(RTLIB::OGT_F32, "__hexagon_fast_gtsf2");
1632 setLibcallName(RTLIB::OLT_F32, "__hexagon_fast_ltsf2");
1633 // Double-precision compares.
1634 setLibcallName(RTLIB::OGT_F64, "__hexagon_fast_gtdf2");
1635 setLibcallName(RTLIB::OLT_F64, "__hexagon_fast_ltdf2");
1637 setLibcallName(RTLIB::ADD_F32, "__hexagon_addsf3");
1638 setLibcallName(RTLIB::SUB_F32, "__hexagon_subsf3");
1639 setLibcallName(RTLIB::MUL_F32, "__hexagon_mulsf3");
1640 setLibcallName(RTLIB::OGT_F32, "__hexagon_gtsf2");
1641 setLibcallName(RTLIB::OLT_F32, "__hexagon_ltsf2");
1642 // Double-precision compares.
1643 setLibcallName(RTLIB::OGT_F64, "__hexagon_gtdf2");
1644 setLibcallName(RTLIB::OLT_F64, "__hexagon_ltdf2");
1648 // This is the only fast library function for sqrtd.
1650 setLibcallName(RTLIB::SQRT_F64, "__hexagon_fast2_sqrtdf2");
1652 // Prefix is: nothing for "slow-math",
1653 // "fast2_" for V4 fast-math and V5+ fast-math double-precision
1654 // (actually, keep fast-math and fast-math2 separate for now)
1656 setLibcallName(RTLIB::ADD_F64, "__hexagon_fast_adddf3");
1657 setLibcallName(RTLIB::SUB_F64, "__hexagon_fast_subdf3");
1658 setLibcallName(RTLIB::MUL_F64, "__hexagon_fast_muldf3");
1659 setLibcallName(RTLIB::DIV_F64, "__hexagon_fast_divdf3");
1660 // Calling __hexagon_fast2_divsf3 with fast-math on V5 (ok).
1661 setLibcallName(RTLIB::DIV_F32, "__hexagon_fast_divsf3");
1663 setLibcallName(RTLIB::ADD_F64, "__hexagon_adddf3");
1664 setLibcallName(RTLIB::SUB_F64, "__hexagon_subdf3");
1665 setLibcallName(RTLIB::MUL_F64, "__hexagon_muldf3");
1666 setLibcallName(RTLIB::DIV_F64, "__hexagon_divdf3");
1667 setLibcallName(RTLIB::DIV_F32, "__hexagon_divsf3");
1670 if (Subtarget.hasV5TOps()) {
1672 setLibcallName(RTLIB::SQRT_F32, "__hexagon_fast2_sqrtf");
1674 setLibcallName(RTLIB::SQRT_F32, "__hexagon_sqrtf");
1677 setLibcallName(RTLIB::SINTTOFP_I32_F32, "__hexagon_floatsisf");
1678 setLibcallName(RTLIB::SINTTOFP_I32_F64, "__hexagon_floatsidf");
1679 setLibcallName(RTLIB::SINTTOFP_I64_F32, "__hexagon_floatdisf");
1680 setLibcallName(RTLIB::SINTTOFP_I64_F64, "__hexagon_floatdidf");
1681 setLibcallName(RTLIB::UINTTOFP_I32_F32, "__hexagon_floatunsisf");
1682 setLibcallName(RTLIB::UINTTOFP_I32_F64, "__hexagon_floatunsidf");
1683 setLibcallName(RTLIB::UINTTOFP_I64_F32, "__hexagon_floatundisf");
1684 setLibcallName(RTLIB::UINTTOFP_I64_F64, "__hexagon_floatundidf");
1685 setLibcallName(RTLIB::FPTOUINT_F32_I32, "__hexagon_fixunssfsi");
1686 setLibcallName(RTLIB::FPTOUINT_F32_I64, "__hexagon_fixunssfdi");
1687 setLibcallName(RTLIB::FPTOUINT_F64_I32, "__hexagon_fixunsdfsi");
1688 setLibcallName(RTLIB::FPTOUINT_F64_I64, "__hexagon_fixunsdfdi");
1689 setLibcallName(RTLIB::FPTOSINT_F32_I32, "__hexagon_fixsfsi");
1690 setLibcallName(RTLIB::FPTOSINT_F32_I64, "__hexagon_fixsfdi");
1691 setLibcallName(RTLIB::FPTOSINT_F64_I32, "__hexagon_fixdfsi");
1692 setLibcallName(RTLIB::FPTOSINT_F64_I64, "__hexagon_fixdfdi");
1693 setLibcallName(RTLIB::FPEXT_F32_F64, "__hexagon_extendsfdf2");
1694 setLibcallName(RTLIB::FPROUND_F64_F32, "__hexagon_truncdfsf2");
1695 setLibcallName(RTLIB::OEQ_F32, "__hexagon_eqsf2");
1696 setLibcallName(RTLIB::OEQ_F64, "__hexagon_eqdf2");
1697 setLibcallName(RTLIB::OGE_F32, "__hexagon_gesf2");
1698 setLibcallName(RTLIB::OGE_F64, "__hexagon_gedf2");
1699 setLibcallName(RTLIB::OLE_F32, "__hexagon_lesf2");
1700 setLibcallName(RTLIB::OLE_F64, "__hexagon_ledf2");
1701 setLibcallName(RTLIB::UNE_F32, "__hexagon_nesf2");
1702 setLibcallName(RTLIB::UNE_F64, "__hexagon_nedf2");
1703 setLibcallName(RTLIB::UO_F32, "__hexagon_unordsf2");
1704 setLibcallName(RTLIB::UO_F64, "__hexagon_unorddf2");
1705 setLibcallName(RTLIB::O_F32, "__hexagon_unordsf2");
1706 setLibcallName(RTLIB::O_F64, "__hexagon_unorddf2");
1709 // These cause problems when the shift amount is non-constant.
1710 setLibcallName(RTLIB::SHL_I128, nullptr);
1711 setLibcallName(RTLIB::SRL_I128, nullptr);
1712 setLibcallName(RTLIB::SRA_I128, nullptr);
1716 const char* HexagonTargetLowering::getTargetNodeName(unsigned Opcode) const {
1717 switch ((HexagonISD::NodeType)Opcode) {
1718 case HexagonISD::ALLOCA: return "HexagonISD::ALLOCA";
1719 case HexagonISD::ARGEXTEND: return "HexagonISD::ARGEXTEND";
1720 case HexagonISD::AT_GOT: return "HexagonISD::AT_GOT";
1721 case HexagonISD::AT_PCREL: return "HexagonISD::AT_PCREL";
1722 case HexagonISD::BARRIER: return "HexagonISD::BARRIER";
1723 case HexagonISD::BR_JT: return "HexagonISD::BR_JT";
1724 case HexagonISD::CALLR: return "HexagonISD::CALLR";
1725 case HexagonISD::CALLv3nr: return "HexagonISD::CALLv3nr";
1726 case HexagonISD::CALLv3: return "HexagonISD::CALLv3";
1727 case HexagonISD::COMBINE: return "HexagonISD::COMBINE";
1728 case HexagonISD::CONST32_GP: return "HexagonISD::CONST32_GP";
1729 case HexagonISD::CONST32: return "HexagonISD::CONST32";
1730 case HexagonISD::CP: return "HexagonISD::CP";
1731 case HexagonISD::DCFETCH: return "HexagonISD::DCFETCH";
1732 case HexagonISD::EH_RETURN: return "HexagonISD::EH_RETURN";
1733 case HexagonISD::EXTRACTU: return "HexagonISD::EXTRACTU";
1734 case HexagonISD::EXTRACTURP: return "HexagonISD::EXTRACTURP";
1735 case HexagonISD::FCONST32: return "HexagonISD::FCONST32";
1736 case HexagonISD::INSERT: return "HexagonISD::INSERT";
1737 case HexagonISD::INSERTRP: return "HexagonISD::INSERTRP";
1738 case HexagonISD::JT: return "HexagonISD::JT";
1739 case HexagonISD::PACKHL: return "HexagonISD::PACKHL";
1740 case HexagonISD::PIC_ADD: return "HexagonISD::PIC_ADD";
1741 case HexagonISD::POPCOUNT: return "HexagonISD::POPCOUNT";
1742 case HexagonISD::RET_FLAG: return "HexagonISD::RET_FLAG";
1743 case HexagonISD::SHUFFEB: return "HexagonISD::SHUFFEB";
1744 case HexagonISD::SHUFFEH: return "HexagonISD::SHUFFEH";
1745 case HexagonISD::SHUFFOB: return "HexagonISD::SHUFFOB";
1746 case HexagonISD::SHUFFOH: return "HexagonISD::SHUFFOH";
1747 case HexagonISD::TC_RETURN: return "HexagonISD::TC_RETURN";
1748 case HexagonISD::VCMPBEQ: return "HexagonISD::VCMPBEQ";
1749 case HexagonISD::VCMPBGT: return "HexagonISD::VCMPBGT";
1750 case HexagonISD::VCMPBGTU: return "HexagonISD::VCMPBGTU";
1751 case HexagonISD::VCMPHEQ: return "HexagonISD::VCMPHEQ";
1752 case HexagonISD::VCMPHGT: return "HexagonISD::VCMPHGT";
1753 case HexagonISD::VCMPHGTU: return "HexagonISD::VCMPHGTU";
1754 case HexagonISD::VCMPWEQ: return "HexagonISD::VCMPWEQ";
1755 case HexagonISD::VCMPWGT: return "HexagonISD::VCMPWGT";
1756 case HexagonISD::VCMPWGTU: return "HexagonISD::VCMPWGTU";
1757 case HexagonISD::VSHLH: return "HexagonISD::VSHLH";
1758 case HexagonISD::VSHLW: return "HexagonISD::VSHLW";
1759 case HexagonISD::VSPLATB: return "HexagonISD::VSPLTB";
1760 case HexagonISD::VSPLATH: return "HexagonISD::VSPLATH";
1761 case HexagonISD::VSRAH: return "HexagonISD::VSRAH";
1762 case HexagonISD::VSRAW: return "HexagonISD::VSRAW";
1763 case HexagonISD::VSRLH: return "HexagonISD::VSRLH";
1764 case HexagonISD::VSRLW: return "HexagonISD::VSRLW";
1765 case HexagonISD::VSXTBH: return "HexagonISD::VSXTBH";
1766 case HexagonISD::VSXTBW: return "HexagonISD::VSXTBW";
1767 case HexagonISD::OP_END: break;
1772 bool HexagonTargetLowering::isTruncateFree(Type *Ty1, Type *Ty2) const {
1773 EVT MTy1 = EVT::getEVT(Ty1);
1774 EVT MTy2 = EVT::getEVT(Ty2);
1775 if (!MTy1.isSimple() || !MTy2.isSimple())
1777 return (MTy1.getSimpleVT() == MVT::i64) && (MTy2.getSimpleVT() == MVT::i32);
1780 bool HexagonTargetLowering::isTruncateFree(EVT VT1, EVT VT2) const {
1781 if (!VT1.isSimple() || !VT2.isSimple())
1783 return (VT1.getSimpleVT() == MVT::i64) && (VT2.getSimpleVT() == MVT::i32);
1786 // shouldExpandBuildVectorWithShuffles
1787 // Should we expand the build vector with shuffles?
1789 HexagonTargetLowering::shouldExpandBuildVectorWithShuffles(EVT VT,
1790 unsigned DefinedValues) const {
1792 // Hexagon vector shuffle operates on element sizes of bytes or halfwords
1793 EVT EltVT = VT.getVectorElementType();
1794 int EltBits = EltVT.getSizeInBits();
1795 if ((EltBits != 8) && (EltBits != 16))
1798 return TargetLowering::shouldExpandBuildVectorWithShuffles(VT, DefinedValues);
1801 // LowerVECTOR_SHUFFLE - Lower a vector shuffle (V1, V2, V3). V1 and
1802 // V2 are the two vectors to select data from, V3 is the permutation.
1803 static SDValue LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) {
1804 const ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(Op);
1805 SDValue V1 = Op.getOperand(0);
1806 SDValue V2 = Op.getOperand(1);
1808 EVT VT = Op.getValueType();
1810 if (V2.getOpcode() == ISD::UNDEF)
1813 if (SVN->isSplat()) {
1814 int Lane = SVN->getSplatIndex();
1815 if (Lane == -1) Lane = 0;
1817 // Test if V1 is a SCALAR_TO_VECTOR.
1818 if (Lane == 0 && V1.getOpcode() == ISD::SCALAR_TO_VECTOR)
1819 return createSplat(DAG, dl, VT, V1.getOperand(0));
1821 // Test if V1 is a BUILD_VECTOR which is equivalent to a SCALAR_TO_VECTOR
1822 // (and probably will turn into a SCALAR_TO_VECTOR once legalization
1824 if (Lane == 0 && V1.getOpcode() == ISD::BUILD_VECTOR &&
1825 !isa<ConstantSDNode>(V1.getOperand(0))) {
1826 bool IsScalarToVector = true;
1827 for (unsigned i = 1, e = V1.getNumOperands(); i != e; ++i)
1828 if (V1.getOperand(i).getOpcode() != ISD::UNDEF) {
1829 IsScalarToVector = false;
1832 if (IsScalarToVector)
1833 return createSplat(DAG, dl, VT, V1.getOperand(0));
1835 return createSplat(DAG, dl, VT, DAG.getConstant(Lane, dl, MVT::i32));
1838 // FIXME: We need to support more general vector shuffles. See
1839 // below the comment from the ARM backend that deals in the general
1840 // case with the vector shuffles. For now, let expand handle these.
1843 // If the shuffle is not directly supported and it has 4 elements, use
1844 // the PerfectShuffle-generated table to synthesize it from other shuffles.
1847 // If BUILD_VECTOR has same base element repeated several times,
1849 static bool isCommonSplatElement(BuildVectorSDNode *BVN) {
1850 unsigned NElts = BVN->getNumOperands();
1851 SDValue V0 = BVN->getOperand(0);
1853 for (unsigned i = 1, e = NElts; i != e; ++i) {
1854 if (BVN->getOperand(i) != V0)
1860 // LowerVECTOR_SHIFT - Lower a vector shift. Try to convert
1861 // <VT> = SHL/SRA/SRL <VT> by <VT> to Hexagon specific
1862 // <VT> = SHL/SRA/SRL <VT> by <IT/i32>.
1863 static SDValue LowerVECTOR_SHIFT(SDValue Op, SelectionDAG &DAG) {
1864 BuildVectorSDNode *BVN = 0;
1865 SDValue V1 = Op.getOperand(0);
1866 SDValue V2 = Op.getOperand(1);
1869 EVT VT = Op.getValueType();
1871 if ((BVN = dyn_cast<BuildVectorSDNode>(V1.getNode())) &&
1872 isCommonSplatElement(BVN))
1874 else if ((BVN = dyn_cast<BuildVectorSDNode>(V2.getNode())) &&
1875 isCommonSplatElement(BVN))
1880 SDValue CommonSplat = BVN->getOperand(0);
1883 if (VT.getSimpleVT() == MVT::v4i16) {
1884 switch (Op.getOpcode()) {
1886 Result = DAG.getNode(HexagonISD::VSRAH, dl, VT, V3, CommonSplat);
1889 Result = DAG.getNode(HexagonISD::VSHLH, dl, VT, V3, CommonSplat);
1892 Result = DAG.getNode(HexagonISD::VSRLH, dl, VT, V3, CommonSplat);
1897 } else if (VT.getSimpleVT() == MVT::v2i32) {
1898 switch (Op.getOpcode()) {
1900 Result = DAG.getNode(HexagonISD::VSRAW, dl, VT, V3, CommonSplat);
1903 Result = DAG.getNode(HexagonISD::VSHLW, dl, VT, V3, CommonSplat);
1906 Result = DAG.getNode(HexagonISD::VSRLW, dl, VT, V3, CommonSplat);
1915 return DAG.getNode(ISD::BITCAST, dl, VT, Result);
1919 HexagonTargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) const {
1920 BuildVectorSDNode *BVN = cast<BuildVectorSDNode>(Op.getNode());
1922 EVT VT = Op.getValueType();
1924 unsigned Size = VT.getSizeInBits();
1926 // A vector larger than 64 bits cannot be represented in Hexagon.
1927 // Expand will split the vector.
1931 APInt APSplatBits, APSplatUndef;
1932 unsigned SplatBitSize;
1934 unsigned NElts = BVN->getNumOperands();
1936 // Try to generate a SPLAT instruction.
1937 if ((VT.getSimpleVT() == MVT::v4i8 || VT.getSimpleVT() == MVT::v4i16) &&
1938 (BVN->isConstantSplat(APSplatBits, APSplatUndef, SplatBitSize,
1939 HasAnyUndefs, 0, true) && SplatBitSize <= 16)) {
1940 unsigned SplatBits = APSplatBits.getZExtValue();
1941 int32_t SextVal = ((int32_t) (SplatBits << (32 - SplatBitSize)) >>
1942 (32 - SplatBitSize));
1943 return createSplat(DAG, dl, VT, DAG.getConstant(SextVal, dl, MVT::i32));
1946 // Try to generate COMBINE to build v2i32 vectors.
1947 if (VT.getSimpleVT() == MVT::v2i32) {
1948 SDValue V0 = BVN->getOperand(0);
1949 SDValue V1 = BVN->getOperand(1);
1951 if (V0.getOpcode() == ISD::UNDEF)
1952 V0 = DAG.getConstant(0, dl, MVT::i32);
1953 if (V1.getOpcode() == ISD::UNDEF)
1954 V1 = DAG.getConstant(0, dl, MVT::i32);
1956 ConstantSDNode *C0 = dyn_cast<ConstantSDNode>(V0);
1957 ConstantSDNode *C1 = dyn_cast<ConstantSDNode>(V1);
1958 // If the element isn't a constant, it is in a register:
1959 // generate a COMBINE Register Register instruction.
1961 return DAG.getNode(HexagonISD::COMBINE, dl, VT, V1, V0);
1963 // If one of the operands is an 8 bit integer constant, generate
1964 // a COMBINE Immediate Immediate instruction.
1965 if (isInt<8>(C0->getSExtValue()) ||
1966 isInt<8>(C1->getSExtValue()))
1967 return DAG.getNode(HexagonISD::COMBINE, dl, VT, V1, V0);
1970 // Try to generate a S2_packhl to build v2i16 vectors.
1971 if (VT.getSimpleVT() == MVT::v2i16) {
1972 for (unsigned i = 0, e = NElts; i != e; ++i) {
1973 if (BVN->getOperand(i).getOpcode() == ISD::UNDEF)
1975 ConstantSDNode *Cst = dyn_cast<ConstantSDNode>(BVN->getOperand(i));
1976 // If the element isn't a constant, it is in a register:
1977 // generate a S2_packhl instruction.
1979 SDValue pack = DAG.getNode(HexagonISD::PACKHL, dl, MVT::v4i16,
1980 BVN->getOperand(1), BVN->getOperand(0));
1982 return DAG.getTargetExtractSubreg(Hexagon::subreg_loreg, dl, MVT::v2i16,
1988 // In the general case, generate a CONST32 or a CONST64 for constant vectors,
1989 // and insert_vector_elt for all the other cases.
1991 unsigned EltSize = Size / NElts;
1993 uint64_t Mask = ~uint64_t(0ULL) >> (64 - EltSize);
1994 bool HasNonConstantElements = false;
1996 for (unsigned i = 0, e = NElts; i != e; ++i) {
1997 // LLVM's BUILD_VECTOR operands are in Little Endian mode, whereas Hexagon's
1998 // combine, const64, etc. are Big Endian.
1999 unsigned OpIdx = NElts - i - 1;
2000 SDValue Operand = BVN->getOperand(OpIdx);
2001 if (Operand.getOpcode() == ISD::UNDEF)
2005 if (ConstantSDNode *Cst = dyn_cast<ConstantSDNode>(Operand))
2006 Val = Cst->getSExtValue();
2008 HasNonConstantElements = true;
2011 Res = (Res << EltSize) | Val;
2015 ConstVal = DAG.getConstant(Res, dl, MVT::i64);
2017 ConstVal = DAG.getConstant(Res, dl, MVT::i32);
2019 // When there are non constant operands, add them with INSERT_VECTOR_ELT to
2020 // ConstVal, the constant part of the vector.
2021 if (HasNonConstantElements) {
2022 EVT EltVT = VT.getVectorElementType();
2023 SDValue Width = DAG.getConstant(EltVT.getSizeInBits(), dl, MVT::i64);
2024 SDValue Shifted = DAG.getNode(ISD::SHL, dl, MVT::i64, Width,
2025 DAG.getConstant(32, dl, MVT::i64));
2027 for (unsigned i = 0, e = NElts; i != e; ++i) {
2028 // LLVM's BUILD_VECTOR operands are in Little Endian mode, whereas Hexagon
2030 unsigned OpIdx = NElts - i - 1;
2031 SDValue Operand = BVN->getOperand(OpIdx);
2032 if (isa<ConstantSDNode>(Operand))
2033 // This operand is already in ConstVal.
2036 if (VT.getSizeInBits() == 64 &&
2037 Operand.getValueType().getSizeInBits() == 32) {
2038 SDValue C = DAG.getConstant(0, dl, MVT::i32);
2039 Operand = DAG.getNode(HexagonISD::COMBINE, dl, VT, C, Operand);
2042 SDValue Idx = DAG.getConstant(OpIdx, dl, MVT::i64);
2043 SDValue Offset = DAG.getNode(ISD::MUL, dl, MVT::i64, Idx, Width);
2044 SDValue Combined = DAG.getNode(ISD::OR, dl, MVT::i64, Shifted, Offset);
2045 const SDValue Ops[] = {ConstVal, Operand, Combined};
2047 if (VT.getSizeInBits() == 32)
2048 ConstVal = DAG.getNode(HexagonISD::INSERTRP, dl, MVT::i32, Ops);
2050 ConstVal = DAG.getNode(HexagonISD::INSERTRP, dl, MVT::i64, Ops);
2054 return DAG.getNode(ISD::BITCAST, dl, VT, ConstVal);
2058 HexagonTargetLowering::LowerCONCAT_VECTORS(SDValue Op,
2059 SelectionDAG &DAG) const {
2061 EVT VT = Op.getValueType();
2062 unsigned NElts = Op.getNumOperands();
2063 SDValue Vec = Op.getOperand(0);
2064 EVT VecVT = Vec.getValueType();
2065 SDValue Width = DAG.getConstant(VecVT.getSizeInBits(), dl, MVT::i64);
2066 SDValue Shifted = DAG.getNode(ISD::SHL, dl, MVT::i64, Width,
2067 DAG.getConstant(32, dl, MVT::i64));
2068 SDValue ConstVal = DAG.getConstant(0, dl, MVT::i64);
2070 ConstantSDNode *W = dyn_cast<ConstantSDNode>(Width);
2071 ConstantSDNode *S = dyn_cast<ConstantSDNode>(Shifted);
2073 if ((VecVT.getSimpleVT() == MVT::v2i16) && (NElts == 2) && W && S) {
2074 if ((W->getZExtValue() == 32) && ((S->getZExtValue() >> 32) == 32)) {
2075 // We are trying to concat two v2i16 to a single v4i16.
2076 SDValue Vec0 = Op.getOperand(1);
2077 SDValue Combined = DAG.getNode(HexagonISD::COMBINE, dl, VT, Vec0, Vec);
2078 return DAG.getNode(ISD::BITCAST, dl, VT, Combined);
2082 if ((VecVT.getSimpleVT() == MVT::v4i8) && (NElts == 2) && W && S) {
2083 if ((W->getZExtValue() == 32) && ((S->getZExtValue() >> 32) == 32)) {
2084 // We are trying to concat two v4i8 to a single v8i8.
2085 SDValue Vec0 = Op.getOperand(1);
2086 SDValue Combined = DAG.getNode(HexagonISD::COMBINE, dl, VT, Vec0, Vec);
2087 return DAG.getNode(ISD::BITCAST, dl, VT, Combined);
2091 for (unsigned i = 0, e = NElts; i != e; ++i) {
2092 unsigned OpIdx = NElts - i - 1;
2093 SDValue Operand = Op.getOperand(OpIdx);
2095 if (VT.getSizeInBits() == 64 &&
2096 Operand.getValueType().getSizeInBits() == 32) {
2097 SDValue C = DAG.getConstant(0, dl, MVT::i32);
2098 Operand = DAG.getNode(HexagonISD::COMBINE, dl, VT, C, Operand);
2101 SDValue Idx = DAG.getConstant(OpIdx, dl, MVT::i64);
2102 SDValue Offset = DAG.getNode(ISD::MUL, dl, MVT::i64, Idx, Width);
2103 SDValue Combined = DAG.getNode(ISD::OR, dl, MVT::i64, Shifted, Offset);
2104 const SDValue Ops[] = {ConstVal, Operand, Combined};
2106 if (VT.getSizeInBits() == 32)
2107 ConstVal = DAG.getNode(HexagonISD::INSERTRP, dl, MVT::i32, Ops);
2109 ConstVal = DAG.getNode(HexagonISD::INSERTRP, dl, MVT::i64, Ops);
2112 return DAG.getNode(ISD::BITCAST, dl, VT, ConstVal);
2116 HexagonTargetLowering::LowerEXTRACT_VECTOR(SDValue Op,
2117 SelectionDAG &DAG) const {
2118 EVT VT = Op.getValueType();
2119 int VTN = VT.isVector() ? VT.getVectorNumElements() : 1;
2121 SDValue Idx = Op.getOperand(1);
2122 SDValue Vec = Op.getOperand(0);
2123 EVT VecVT = Vec.getValueType();
2124 EVT EltVT = VecVT.getVectorElementType();
2125 int EltSize = EltVT.getSizeInBits();
2126 SDValue Width = DAG.getConstant(Op.getOpcode() == ISD::EXTRACT_VECTOR_ELT ?
2127 EltSize : VTN * EltSize, dl, MVT::i64);
2129 // Constant element number.
2130 if (ConstantSDNode *CI = dyn_cast<ConstantSDNode>(Idx)) {
2131 uint64_t X = CI->getZExtValue();
2132 SDValue Offset = DAG.getConstant(X * EltSize, dl, MVT::i32);
2133 const SDValue Ops[] = {Vec, Width, Offset};
2135 ConstantSDNode *CW = dyn_cast<ConstantSDNode>(Width);
2136 assert(CW && "Non constant width in LowerEXTRACT_VECTOR");
2139 MVT SVT = VecVT.getSimpleVT();
2140 uint64_t W = CW->getZExtValue();
2143 // Translate this node into EXTRACT_SUBREG.
2144 unsigned Subreg = (X == 0) ? Hexagon::subreg_loreg : 0;
2147 Subreg = Hexagon::subreg_loreg;
2148 else if (SVT == MVT::v2i32 && X == 1)
2149 Subreg = Hexagon::subreg_hireg;
2150 else if (SVT == MVT::v4i16 && X == 2)
2151 Subreg = Hexagon::subreg_hireg;
2152 else if (SVT == MVT::v8i8 && X == 4)
2153 Subreg = Hexagon::subreg_hireg;
2155 llvm_unreachable("Bad offset");
2156 N = DAG.getTargetExtractSubreg(Subreg, dl, MVT::i32, Vec);
2158 } else if (VecVT.getSizeInBits() == 32) {
2159 N = DAG.getNode(HexagonISD::EXTRACTU, dl, MVT::i32, Ops);
2161 N = DAG.getNode(HexagonISD::EXTRACTU, dl, MVT::i64, Ops);
2162 if (VT.getSizeInBits() == 32)
2163 N = DAG.getTargetExtractSubreg(Hexagon::subreg_loreg, dl, MVT::i32, N);
2166 return DAG.getNode(ISD::BITCAST, dl, VT, N);
2169 // Variable element number.
2170 SDValue Offset = DAG.getNode(ISD::MUL, dl, MVT::i32, Idx,
2171 DAG.getConstant(EltSize, dl, MVT::i32));
2172 SDValue Shifted = DAG.getNode(ISD::SHL, dl, MVT::i64, Width,
2173 DAG.getConstant(32, dl, MVT::i64));
2174 SDValue Combined = DAG.getNode(ISD::OR, dl, MVT::i64, Shifted, Offset);
2176 const SDValue Ops[] = {Vec, Combined};
2179 if (VecVT.getSizeInBits() == 32) {
2180 N = DAG.getNode(HexagonISD::EXTRACTURP, dl, MVT::i32, Ops);
2182 N = DAG.getNode(HexagonISD::EXTRACTURP, dl, MVT::i64, Ops);
2183 if (VT.getSizeInBits() == 32)
2184 N = DAG.getTargetExtractSubreg(Hexagon::subreg_loreg, dl, MVT::i32, N);
2186 return DAG.getNode(ISD::BITCAST, dl, VT, N);
2190 HexagonTargetLowering::LowerINSERT_VECTOR(SDValue Op,
2191 SelectionDAG &DAG) const {
2192 EVT VT = Op.getValueType();
2193 int VTN = VT.isVector() ? VT.getVectorNumElements() : 1;
2195 SDValue Vec = Op.getOperand(0);
2196 SDValue Val = Op.getOperand(1);
2197 SDValue Idx = Op.getOperand(2);
2198 EVT VecVT = Vec.getValueType();
2199 EVT EltVT = VecVT.getVectorElementType();
2200 int EltSize = EltVT.getSizeInBits();
2201 SDValue Width = DAG.getConstant(Op.getOpcode() == ISD::INSERT_VECTOR_ELT ?
2202 EltSize : VTN * EltSize, dl, MVT::i64);
2204 if (ConstantSDNode *C = cast<ConstantSDNode>(Idx)) {
2205 SDValue Offset = DAG.getConstant(C->getSExtValue() * EltSize, dl, MVT::i32);
2206 const SDValue Ops[] = {Vec, Val, Width, Offset};
2209 if (VT.getSizeInBits() == 32)
2210 N = DAG.getNode(HexagonISD::INSERT, dl, MVT::i32, Ops);
2212 N = DAG.getNode(HexagonISD::INSERT, dl, MVT::i64, Ops);
2214 return DAG.getNode(ISD::BITCAST, dl, VT, N);
2217 // Variable element number.
2218 SDValue Offset = DAG.getNode(ISD::MUL, dl, MVT::i32, Idx,
2219 DAG.getConstant(EltSize, dl, MVT::i32));
2220 SDValue Shifted = DAG.getNode(ISD::SHL, dl, MVT::i64, Width,
2221 DAG.getConstant(32, dl, MVT::i64));
2222 SDValue Combined = DAG.getNode(ISD::OR, dl, MVT::i64, Shifted, Offset);
2224 if (VT.getSizeInBits() == 64 &&
2225 Val.getValueType().getSizeInBits() == 32) {
2226 SDValue C = DAG.getConstant(0, dl, MVT::i32);
2227 Val = DAG.getNode(HexagonISD::COMBINE, dl, VT, C, Val);
2230 const SDValue Ops[] = {Vec, Val, Combined};
2233 if (VT.getSizeInBits() == 32)
2234 N = DAG.getNode(HexagonISD::INSERTRP, dl, MVT::i32, Ops);
2236 N = DAG.getNode(HexagonISD::INSERTRP, dl, MVT::i64, Ops);
2238 return DAG.getNode(ISD::BITCAST, dl, VT, N);
2242 HexagonTargetLowering::allowTruncateForTailCall(Type *Ty1, Type *Ty2) const {
2243 // Assuming the caller does not have either a signext or zeroext modifier, and
2244 // only one value is accepted, any reasonable truncation is allowed.
2245 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
2248 // FIXME: in principle up to 64-bit could be made safe, but it would be very
2249 // fragile at the moment: any support for multiple value returns would be
2250 // liable to disallow tail calls involving i64 -> iN truncation in many cases.
2251 return Ty1->getPrimitiveSizeInBits() <= 32;
2255 HexagonTargetLowering::LowerEH_RETURN(SDValue Op, SelectionDAG &DAG) const {
2256 SDValue Chain = Op.getOperand(0);
2257 SDValue Offset = Op.getOperand(1);
2258 SDValue Handler = Op.getOperand(2);
2260 auto PtrVT = getPointerTy(DAG.getDataLayout());
2262 // Mark function as containing a call to EH_RETURN.
2263 HexagonMachineFunctionInfo *FuncInfo =
2264 DAG.getMachineFunction().getInfo<HexagonMachineFunctionInfo>();
2265 FuncInfo->setHasEHReturn();
2267 unsigned OffsetReg = Hexagon::R28;
2270 DAG.getNode(ISD::ADD, dl, PtrVT, DAG.getRegister(Hexagon::R30, PtrVT),
2271 DAG.getIntPtrConstant(4, dl));
2272 Chain = DAG.getStore(Chain, dl, Handler, StoreAddr, MachinePointerInfo(),
2274 Chain = DAG.getCopyToReg(Chain, dl, OffsetReg, Offset);
2276 // Not needed we already use it as explict input to EH_RETURN.
2277 // MF.getRegInfo().addLiveOut(OffsetReg);
2279 return DAG.getNode(HexagonISD::EH_RETURN, dl, MVT::Other, Chain);
2283 HexagonTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
2284 unsigned Opc = Op.getOpcode();
2288 Op.getNode()->dumpr(&DAG);
2289 if (Opc > HexagonISD::OP_BEGIN && Opc < HexagonISD::OP_END)
2290 errs() << "Check for a non-legal type in this operation\n";
2292 llvm_unreachable("Should not custom lower this!");
2293 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
2294 case ISD::INSERT_SUBVECTOR: return LowerINSERT_VECTOR(Op, DAG);
2295 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR(Op, DAG);
2296 case ISD::EXTRACT_SUBVECTOR: return LowerEXTRACT_VECTOR(Op, DAG);
2297 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR(Op, DAG);
2298 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
2299 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
2302 case ISD::SRL: return LowerVECTOR_SHIFT(Op, DAG);
2303 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
2304 case ISD::EH_RETURN: return LowerEH_RETURN(Op, DAG);
2305 // Frame & Return address. Currently unimplemented.
2306 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
2307 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
2308 case ISD::ATOMIC_FENCE: return LowerATOMIC_FENCE(Op, DAG);
2309 case ISD::GlobalAddress: return LowerGLOBALADDRESS(Op, DAG);
2310 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
2311 case ISD::VASTART: return LowerVASTART(Op, DAG);
2312 case ISD::BR_JT: return LowerBR_JT(Op, DAG);
2313 // Custom lower some vector loads.
2314 case ISD::LOAD: return LowerLOAD(Op, DAG);
2315 case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
2316 case ISD::SETCC: return LowerSETCC(Op, DAG);
2317 case ISD::VSELECT: return LowerVSELECT(Op, DAG);
2318 case ISD::CTPOP: return LowerCTPOP(Op, DAG);
2319 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
2320 case ISD::INLINEASM: return LowerINLINEASM(Op, DAG);
2325 HexagonTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
2326 MachineBasicBlock *BB)
2328 switch (MI->getOpcode()) {
2329 case Hexagon::ALLOCA: {
2330 MachineFunction *MF = BB->getParent();
2331 auto *FuncInfo = MF->getInfo<HexagonMachineFunctionInfo>();
2332 FuncInfo->addAllocaAdjustInst(MI);
2335 default: llvm_unreachable("Unexpected instr type to insert");
2339 //===----------------------------------------------------------------------===//
2340 // Inline Assembly Support
2341 //===----------------------------------------------------------------------===//
2343 std::pair<unsigned, const TargetRegisterClass *>
2344 HexagonTargetLowering::getRegForInlineAsmConstraint(
2345 const TargetRegisterInfo *TRI, StringRef Constraint, MVT VT) const {
2346 if (Constraint.size() == 1) {
2347 switch (Constraint[0]) {
2349 switch (VT.SimpleTy) {
2351 llvm_unreachable("getRegForInlineAsmConstraint Unhandled data type");
2356 return std::make_pair(0U, &Hexagon::IntRegsRegClass);
2359 return std::make_pair(0U, &Hexagon::DoubleRegsRegClass);
2362 llvm_unreachable("Unknown asm register class");
2366 return TargetLowering::getRegForInlineAsmConstraint(TRI, Constraint, VT);
2369 /// isFPImmLegal - Returns true if the target can instruction select the
2370 /// specified FP immediate natively. If false, the legalizer will
2371 /// materialize the FP immediate as a load from a constant pool.
2372 bool HexagonTargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
2373 return Subtarget.hasV5TOps();
2376 /// isLegalAddressingMode - Return true if the addressing mode represented by
2377 /// AM is legal for this target, for a load/store of the specified type.
2378 bool HexagonTargetLowering::isLegalAddressingMode(const DataLayout &DL,
2379 const AddrMode &AM, Type *Ty,
2380 unsigned AS) const {
2381 // Allows a signed-extended 11-bit immediate field.
2382 if (AM.BaseOffs <= -(1LL << 13) || AM.BaseOffs >= (1LL << 13)-1)
2385 // No global is ever allowed as a base.
2389 int Scale = AM.Scale;
2390 if (Scale < 0) Scale = -Scale;
2392 case 0: // No scale reg, "r+i", "r", or just "i".
2394 default: // No scaled addressing mode.
2400 /// isLegalICmpImmediate - Return true if the specified immediate is legal
2401 /// icmp immediate, that is the target has icmp instructions which can compare
2402 /// a register against the immediate without having to materialize the
2403 /// immediate into a register.
2404 bool HexagonTargetLowering::isLegalICmpImmediate(int64_t Imm) const {
2405 return Imm >= -512 && Imm <= 511;
2408 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
2409 /// for tail call optimization. Targets which want to do tail call
2410 /// optimization should implement this function.
2411 bool HexagonTargetLowering::IsEligibleForTailCallOptimization(
2413 CallingConv::ID CalleeCC,
2415 bool isCalleeStructRet,
2416 bool isCallerStructRet,
2417 const SmallVectorImpl<ISD::OutputArg> &Outs,
2418 const SmallVectorImpl<SDValue> &OutVals,
2419 const SmallVectorImpl<ISD::InputArg> &Ins,
2420 SelectionDAG& DAG) const {
2421 const Function *CallerF = DAG.getMachineFunction().getFunction();
2422 CallingConv::ID CallerCC = CallerF->getCallingConv();
2423 bool CCMatch = CallerCC == CalleeCC;
2425 // ***************************************************************************
2426 // Look for obvious safe cases to perform tail call optimization that do not
2427 // require ABI changes.
2428 // ***************************************************************************
2430 // If this is a tail call via a function pointer, then don't do it!
2431 if (!(dyn_cast<GlobalAddressSDNode>(Callee))
2432 && !(dyn_cast<ExternalSymbolSDNode>(Callee))) {
2436 // Do not optimize if the calling conventions do not match.
2440 // Do not tail call optimize vararg calls.
2444 // Also avoid tail call optimization if either caller or callee uses struct
2445 // return semantics.
2446 if (isCalleeStructRet || isCallerStructRet)
2449 // In addition to the cases above, we also disable Tail Call Optimization if
2450 // the calling convention code that at least one outgoing argument needs to
2451 // go on the stack. We cannot check that here because at this point that
2452 // information is not available.
2456 // Return true when the given node fits in a positive half word.
2457 bool llvm::isPositiveHalfWord(SDNode *N) {
2458 ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N);
2459 if (CN && CN->getSExtValue() > 0 && isInt<16>(CN->getSExtValue()))
2462 switch (N->getOpcode()) {
2465 case ISD::SIGN_EXTEND_INREG:
2470 Value *HexagonTargetLowering::emitLoadLinked(IRBuilder<> &Builder, Value *Addr,
2471 AtomicOrdering Ord) const {
2472 BasicBlock *BB = Builder.GetInsertBlock();
2473 Module *M = BB->getParent()->getParent();
2474 Type *Ty = cast<PointerType>(Addr->getType())->getElementType();
2475 unsigned SZ = Ty->getPrimitiveSizeInBits();
2476 assert((SZ == 32 || SZ == 64) && "Only 32/64-bit atomic loads supported");
2477 Intrinsic::ID IntID = (SZ == 32) ? Intrinsic::hexagon_L2_loadw_locked
2478 : Intrinsic::hexagon_L4_loadd_locked;
2479 Value *Fn = Intrinsic::getDeclaration(M, IntID);
2480 return Builder.CreateCall(Fn, Addr, "larx");
2483 /// Perform a store-conditional operation to Addr. Return the status of the
2484 /// store. This should be 0 if the store succeeded, non-zero otherwise.
2485 Value *HexagonTargetLowering::emitStoreConditional(IRBuilder<> &Builder,
2486 Value *Val, Value *Addr, AtomicOrdering Ord) const {
2487 BasicBlock *BB = Builder.GetInsertBlock();
2488 Module *M = BB->getParent()->getParent();
2489 Type *Ty = Val->getType();
2490 unsigned SZ = Ty->getPrimitiveSizeInBits();
2491 assert((SZ == 32 || SZ == 64) && "Only 32/64-bit atomic stores supported");
2492 Intrinsic::ID IntID = (SZ == 32) ? Intrinsic::hexagon_S2_storew_locked
2493 : Intrinsic::hexagon_S4_stored_locked;
2494 Value *Fn = Intrinsic::getDeclaration(M, IntID);
2495 Value *Call = Builder.CreateCall(Fn, {Addr, Val}, "stcx");
2496 Value *Cmp = Builder.CreateICmpEQ(Call, Builder.getInt32(0), "");
2497 Value *Ext = Builder.CreateZExt(Cmp, Type::getInt32Ty(M->getContext()));
2501 bool HexagonTargetLowering::shouldExpandAtomicLoadInIR(LoadInst *LI) const {
2502 // Do not expand loads and stores that don't exceed 64 bits.
2503 return LI->getType()->getPrimitiveSizeInBits() > 64;
2506 bool HexagonTargetLowering::shouldExpandAtomicStoreInIR(StoreInst *SI) const {
2507 // Do not expand loads and stores that don't exceed 64 bits.
2508 return SI->getValueOperand()->getType()->getPrimitiveSizeInBits() > 64;