xref: /freebsd/contrib/llvm-project/llvm/lib/Target/BPF/BPFISelLowering.cpp (revision d9a42747950146bf03cda7f6e25d219253f8a57a)
1 //===-- BPFISelLowering.cpp - BPF DAG Lowering Implementation  ------------===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 //
9 // This file defines the interfaces that BPF uses to lower LLVM code into a
10 // selection DAG.
11 //
12 //===----------------------------------------------------------------------===//
13 
14 #include "BPFISelLowering.h"
15 #include "BPF.h"
16 #include "BPFSubtarget.h"
17 #include "BPFTargetMachine.h"
18 #include "llvm/CodeGen/CallingConvLower.h"
19 #include "llvm/CodeGen/MachineFrameInfo.h"
20 #include "llvm/CodeGen/MachineFunction.h"
21 #include "llvm/CodeGen/MachineInstrBuilder.h"
22 #include "llvm/CodeGen/MachineRegisterInfo.h"
23 #include "llvm/CodeGen/TargetLoweringObjectFileImpl.h"
24 #include "llvm/CodeGen/ValueTypes.h"
25 #include "llvm/IR/DiagnosticInfo.h"
26 #include "llvm/IR/DiagnosticPrinter.h"
27 #include "llvm/Support/Debug.h"
28 #include "llvm/Support/ErrorHandling.h"
29 #include "llvm/Support/raw_ostream.h"
30 using namespace llvm;
31 
32 #define DEBUG_TYPE "bpf-lower"
33 
34 static cl::opt<bool> BPFExpandMemcpyInOrder("bpf-expand-memcpy-in-order",
35   cl::Hidden, cl::init(false),
36   cl::desc("Expand memcpy into load/store pairs in order"));
37 
38 static void fail(const SDLoc &DL, SelectionDAG &DAG, const Twine &Msg) {
39   MachineFunction &MF = DAG.getMachineFunction();
40   DAG.getContext()->diagnose(
41       DiagnosticInfoUnsupported(MF.getFunction(), Msg, DL.getDebugLoc()));
42 }
43 
44 static void fail(const SDLoc &DL, SelectionDAG &DAG, const char *Msg,
45                  SDValue Val) {
46   MachineFunction &MF = DAG.getMachineFunction();
47   std::string Str;
48   raw_string_ostream OS(Str);
49   OS << Msg;
50   Val->print(OS);
51   OS.flush();
52   DAG.getContext()->diagnose(
53       DiagnosticInfoUnsupported(MF.getFunction(), Str, DL.getDebugLoc()));
54 }
55 
56 BPFTargetLowering::BPFTargetLowering(const TargetMachine &TM,
57                                      const BPFSubtarget &STI)
58     : TargetLowering(TM) {
59 
60   // Set up the register classes.
61   addRegisterClass(MVT::i64, &BPF::GPRRegClass);
62   if (STI.getHasAlu32())
63     addRegisterClass(MVT::i32, &BPF::GPR32RegClass);
64 
65   // Compute derived properties from the register classes
66   computeRegisterProperties(STI.getRegisterInfo());
67 
68   setStackPointerRegisterToSaveRestore(BPF::R11);
69 
70   setOperationAction(ISD::BR_CC, MVT::i64, Custom);
71   setOperationAction(ISD::BR_JT, MVT::Other, Expand);
72   setOperationAction(ISD::BRIND, MVT::Other, Expand);
73   setOperationAction(ISD::BRCOND, MVT::Other, Expand);
74 
75   setOperationAction(ISD::GlobalAddress, MVT::i64, Custom);
76 
77   setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i64, Custom);
78   setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
79   setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
80 
81   // Set unsupported atomic operations as Custom so
82   // we can emit better error messages than fatal error
83   // from selectiondag.
84   for (auto VT : {MVT::i8, MVT::i16, MVT::i32}) {
85     if (VT == MVT::i32) {
86       if (STI.getHasAlu32())
87         continue;
88     } else {
89       setOperationAction(ISD::ATOMIC_LOAD_ADD, VT, Custom);
90     }
91 
92     setOperationAction(ISD::ATOMIC_LOAD_AND, VT, Custom);
93     setOperationAction(ISD::ATOMIC_LOAD_OR, VT, Custom);
94     setOperationAction(ISD::ATOMIC_LOAD_XOR, VT, Custom);
95     setOperationAction(ISD::ATOMIC_SWAP, VT, Custom);
96     setOperationAction(ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS, VT, Custom);
97   }
98 
99   for (auto VT : { MVT::i32, MVT::i64 }) {
100     if (VT == MVT::i32 && !STI.getHasAlu32())
101       continue;
102 
103     setOperationAction(ISD::SDIVREM, VT, Expand);
104     setOperationAction(ISD::UDIVREM, VT, Expand);
105     setOperationAction(ISD::SREM, VT, Expand);
106     setOperationAction(ISD::MULHU, VT, Expand);
107     setOperationAction(ISD::MULHS, VT, Expand);
108     setOperationAction(ISD::UMUL_LOHI, VT, Expand);
109     setOperationAction(ISD::SMUL_LOHI, VT, Expand);
110     setOperationAction(ISD::ROTR, VT, Expand);
111     setOperationAction(ISD::ROTL, VT, Expand);
112     setOperationAction(ISD::SHL_PARTS, VT, Expand);
113     setOperationAction(ISD::SRL_PARTS, VT, Expand);
114     setOperationAction(ISD::SRA_PARTS, VT, Expand);
115     setOperationAction(ISD::CTPOP, VT, Expand);
116 
117     setOperationAction(ISD::SETCC, VT, Expand);
118     setOperationAction(ISD::SELECT, VT, Expand);
119     setOperationAction(ISD::SELECT_CC, VT, Custom);
120   }
121 
122   if (STI.getHasAlu32()) {
123     setOperationAction(ISD::BSWAP, MVT::i32, Promote);
124     setOperationAction(ISD::BR_CC, MVT::i32,
125                        STI.getHasJmp32() ? Custom : Promote);
126   }
127 
128   setOperationAction(ISD::CTTZ, MVT::i64, Custom);
129   setOperationAction(ISD::CTLZ, MVT::i64, Custom);
130   setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i64, Custom);
131   setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i64, Custom);
132 
133   setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand);
134   setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8, Expand);
135   setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16, Expand);
136   setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i32, Expand);
137 
138   // Extended load operations for i1 types must be promoted
139   for (MVT VT : MVT::integer_valuetypes()) {
140     setLoadExtAction(ISD::EXTLOAD, VT, MVT::i1, Promote);
141     setLoadExtAction(ISD::ZEXTLOAD, VT, MVT::i1, Promote);
142     setLoadExtAction(ISD::SEXTLOAD, VT, MVT::i1, Promote);
143 
144     setLoadExtAction(ISD::SEXTLOAD, VT, MVT::i8, Expand);
145     setLoadExtAction(ISD::SEXTLOAD, VT, MVT::i16, Expand);
146     setLoadExtAction(ISD::SEXTLOAD, VT, MVT::i32, Expand);
147   }
148 
149   setBooleanContents(ZeroOrOneBooleanContent);
150 
151   // Function alignments
152   setMinFunctionAlignment(Align(8));
153   setPrefFunctionAlignment(Align(8));
154 
155   if (BPFExpandMemcpyInOrder) {
156     // LLVM generic code will try to expand memcpy into load/store pairs at this
157     // stage which is before quite a few IR optimization passes, therefore the
158     // loads and stores could potentially be moved apart from each other which
159     // will cause trouble to memcpy pattern matcher inside kernel eBPF JIT
160     // compilers.
161     //
162     // When -bpf-expand-memcpy-in-order specified, we want to defer the expand
163     // of memcpy to later stage in IR optimization pipeline so those load/store
164     // pairs won't be touched and could be kept in order. Hence, we set
165     // MaxStoresPerMem* to zero to disable the generic getMemcpyLoadsAndStores
166     // code path, and ask LLVM to use target expander EmitTargetCodeForMemcpy.
167     MaxStoresPerMemset = MaxStoresPerMemsetOptSize = 0;
168     MaxStoresPerMemcpy = MaxStoresPerMemcpyOptSize = 0;
169     MaxStoresPerMemmove = MaxStoresPerMemmoveOptSize = 0;
170     MaxLoadsPerMemcmp = 0;
171   } else {
172     // inline memcpy() for kernel to see explicit copy
173     unsigned CommonMaxStores =
174       STI.getSelectionDAGInfo()->getCommonMaxStoresPerMemFunc();
175 
176     MaxStoresPerMemset = MaxStoresPerMemsetOptSize = CommonMaxStores;
177     MaxStoresPerMemcpy = MaxStoresPerMemcpyOptSize = CommonMaxStores;
178     MaxStoresPerMemmove = MaxStoresPerMemmoveOptSize = CommonMaxStores;
179     MaxLoadsPerMemcmp = MaxLoadsPerMemcmpOptSize = CommonMaxStores;
180   }
181 
182   // CPU/Feature control
183   HasAlu32 = STI.getHasAlu32();
184   HasJmp32 = STI.getHasJmp32();
185   HasJmpExt = STI.getHasJmpExt();
186 }
187 
188 bool BPFTargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
189   return false;
190 }
191 
192 bool BPFTargetLowering::isTruncateFree(Type *Ty1, Type *Ty2) const {
193   if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
194     return false;
195   unsigned NumBits1 = Ty1->getPrimitiveSizeInBits();
196   unsigned NumBits2 = Ty2->getPrimitiveSizeInBits();
197   return NumBits1 > NumBits2;
198 }
199 
200 bool BPFTargetLowering::isTruncateFree(EVT VT1, EVT VT2) const {
201   if (!VT1.isInteger() || !VT2.isInteger())
202     return false;
203   unsigned NumBits1 = VT1.getSizeInBits();
204   unsigned NumBits2 = VT2.getSizeInBits();
205   return NumBits1 > NumBits2;
206 }
207 
208 bool BPFTargetLowering::isZExtFree(Type *Ty1, Type *Ty2) const {
209   if (!getHasAlu32() || !Ty1->isIntegerTy() || !Ty2->isIntegerTy())
210     return false;
211   unsigned NumBits1 = Ty1->getPrimitiveSizeInBits();
212   unsigned NumBits2 = Ty2->getPrimitiveSizeInBits();
213   return NumBits1 == 32 && NumBits2 == 64;
214 }
215 
216 bool BPFTargetLowering::isZExtFree(EVT VT1, EVT VT2) const {
217   if (!getHasAlu32() || !VT1.isInteger() || !VT2.isInteger())
218     return false;
219   unsigned NumBits1 = VT1.getSizeInBits();
220   unsigned NumBits2 = VT2.getSizeInBits();
221   return NumBits1 == 32 && NumBits2 == 64;
222 }
223 
224 BPFTargetLowering::ConstraintType
225 BPFTargetLowering::getConstraintType(StringRef Constraint) const {
226   if (Constraint.size() == 1) {
227     switch (Constraint[0]) {
228     default:
229       break;
230     case 'w':
231       return C_RegisterClass;
232     }
233   }
234 
235   return TargetLowering::getConstraintType(Constraint);
236 }
237 
238 std::pair<unsigned, const TargetRegisterClass *>
239 BPFTargetLowering::getRegForInlineAsmConstraint(const TargetRegisterInfo *TRI,
240                                                 StringRef Constraint,
241                                                 MVT VT) const {
242   if (Constraint.size() == 1)
243     // GCC Constraint Letters
244     switch (Constraint[0]) {
245     case 'r': // GENERAL_REGS
246       return std::make_pair(0U, &BPF::GPRRegClass);
247     case 'w':
248       if (HasAlu32)
249         return std::make_pair(0U, &BPF::GPR32RegClass);
250       break;
251     default:
252       break;
253     }
254 
255   return TargetLowering::getRegForInlineAsmConstraint(TRI, Constraint, VT);
256 }
257 
258 void BPFTargetLowering::ReplaceNodeResults(
259   SDNode *N, SmallVectorImpl<SDValue> &Results, SelectionDAG &DAG) const {
260   const char *err_msg;
261   uint32_t Opcode = N->getOpcode();
262   switch (Opcode) {
263   default:
264     report_fatal_error("Unhandled custom legalization");
265   case ISD::ATOMIC_LOAD_ADD:
266   case ISD::ATOMIC_LOAD_AND:
267   case ISD::ATOMIC_LOAD_OR:
268   case ISD::ATOMIC_LOAD_XOR:
269   case ISD::ATOMIC_SWAP:
270   case ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS:
271     if (HasAlu32 || Opcode == ISD::ATOMIC_LOAD_ADD)
272       err_msg = "Unsupported atomic operations, please use 32/64 bit version";
273     else
274       err_msg = "Unsupported atomic operations, please use 64 bit version";
275     break;
276   }
277 
278   SDLoc DL(N);
279   fail(DL, DAG, err_msg);
280 }
281 
282 SDValue BPFTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
283   switch (Op.getOpcode()) {
284   case ISD::BR_CC:
285     return LowerBR_CC(Op, DAG);
286   case ISD::GlobalAddress:
287     return LowerGlobalAddress(Op, DAG);
288   case ISD::SELECT_CC:
289     return LowerSELECT_CC(Op, DAG);
290   case ISD::DYNAMIC_STACKALLOC:
291     report_fatal_error("Unsupported dynamic stack allocation");
292   default:
293     llvm_unreachable("unimplemented operand");
294   }
295 }
296 
297 // Calling Convention Implementation
298 #include "BPFGenCallingConv.inc"
299 
300 SDValue BPFTargetLowering::LowerFormalArguments(
301     SDValue Chain, CallingConv::ID CallConv, bool IsVarArg,
302     const SmallVectorImpl<ISD::InputArg> &Ins, const SDLoc &DL,
303     SelectionDAG &DAG, SmallVectorImpl<SDValue> &InVals) const {
304   switch (CallConv) {
305   default:
306     report_fatal_error("Unsupported calling convention");
307   case CallingConv::C:
308   case CallingConv::Fast:
309     break;
310   }
311 
312   MachineFunction &MF = DAG.getMachineFunction();
313   MachineRegisterInfo &RegInfo = MF.getRegInfo();
314 
315   // Assign locations to all of the incoming arguments.
316   SmallVector<CCValAssign, 16> ArgLocs;
317   CCState CCInfo(CallConv, IsVarArg, MF, ArgLocs, *DAG.getContext());
318   CCInfo.AnalyzeFormalArguments(Ins, getHasAlu32() ? CC_BPF32 : CC_BPF64);
319 
320   for (auto &VA : ArgLocs) {
321     if (VA.isRegLoc()) {
322       // Arguments passed in registers
323       EVT RegVT = VA.getLocVT();
324       MVT::SimpleValueType SimpleTy = RegVT.getSimpleVT().SimpleTy;
325       switch (SimpleTy) {
326       default: {
327         errs() << "LowerFormalArguments Unhandled argument type: "
328                << RegVT.getEVTString() << '\n';
329         llvm_unreachable(nullptr);
330       }
331       case MVT::i32:
332       case MVT::i64:
333         Register VReg = RegInfo.createVirtualRegister(
334             SimpleTy == MVT::i64 ? &BPF::GPRRegClass : &BPF::GPR32RegClass);
335         RegInfo.addLiveIn(VA.getLocReg(), VReg);
336         SDValue ArgValue = DAG.getCopyFromReg(Chain, DL, VReg, RegVT);
337 
338         // If this is an value that has been promoted to wider types, insert an
339         // assert[sz]ext to capture this, then truncate to the right size.
340         if (VA.getLocInfo() == CCValAssign::SExt)
341           ArgValue = DAG.getNode(ISD::AssertSext, DL, RegVT, ArgValue,
342                                  DAG.getValueType(VA.getValVT()));
343         else if (VA.getLocInfo() == CCValAssign::ZExt)
344           ArgValue = DAG.getNode(ISD::AssertZext, DL, RegVT, ArgValue,
345                                  DAG.getValueType(VA.getValVT()));
346 
347         if (VA.getLocInfo() != CCValAssign::Full)
348           ArgValue = DAG.getNode(ISD::TRUNCATE, DL, VA.getValVT(), ArgValue);
349 
350         InVals.push_back(ArgValue);
351 
352 	break;
353       }
354     } else {
355       fail(DL, DAG, "defined with too many args");
356       InVals.push_back(DAG.getConstant(0, DL, VA.getLocVT()));
357     }
358   }
359 
360   if (IsVarArg || MF.getFunction().hasStructRetAttr()) {
361     fail(DL, DAG, "functions with VarArgs or StructRet are not supported");
362   }
363 
364   return Chain;
365 }
366 
367 const unsigned BPFTargetLowering::MaxArgs = 5;
368 
369 SDValue BPFTargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI,
370                                      SmallVectorImpl<SDValue> &InVals) const {
371   SelectionDAG &DAG = CLI.DAG;
372   auto &Outs = CLI.Outs;
373   auto &OutVals = CLI.OutVals;
374   auto &Ins = CLI.Ins;
375   SDValue Chain = CLI.Chain;
376   SDValue Callee = CLI.Callee;
377   bool &IsTailCall = CLI.IsTailCall;
378   CallingConv::ID CallConv = CLI.CallConv;
379   bool IsVarArg = CLI.IsVarArg;
380   MachineFunction &MF = DAG.getMachineFunction();
381 
382   // BPF target does not support tail call optimization.
383   IsTailCall = false;
384 
385   switch (CallConv) {
386   default:
387     report_fatal_error("Unsupported calling convention");
388   case CallingConv::Fast:
389   case CallingConv::C:
390     break;
391   }
392 
393   // Analyze operands of the call, assigning locations to each operand.
394   SmallVector<CCValAssign, 16> ArgLocs;
395   CCState CCInfo(CallConv, IsVarArg, MF, ArgLocs, *DAG.getContext());
396 
397   CCInfo.AnalyzeCallOperands(Outs, getHasAlu32() ? CC_BPF32 : CC_BPF64);
398 
399   unsigned NumBytes = CCInfo.getNextStackOffset();
400 
401   if (Outs.size() > MaxArgs)
402     fail(CLI.DL, DAG, "too many args to ", Callee);
403 
404   for (auto &Arg : Outs) {
405     ISD::ArgFlagsTy Flags = Arg.Flags;
406     if (!Flags.isByVal())
407       continue;
408 
409     fail(CLI.DL, DAG, "pass by value not supported ", Callee);
410   }
411 
412   auto PtrVT = getPointerTy(MF.getDataLayout());
413   Chain = DAG.getCALLSEQ_START(Chain, NumBytes, 0, CLI.DL);
414 
415   SmallVector<std::pair<unsigned, SDValue>, MaxArgs> RegsToPass;
416 
417   // Walk arg assignments
418   for (unsigned i = 0,
419                 e = std::min(static_cast<unsigned>(ArgLocs.size()), MaxArgs);
420        i != e; ++i) {
421     CCValAssign &VA = ArgLocs[i];
422     SDValue Arg = OutVals[i];
423 
424     // Promote the value if needed.
425     switch (VA.getLocInfo()) {
426     default:
427       llvm_unreachable("Unknown loc info");
428     case CCValAssign::Full:
429       break;
430     case CCValAssign::SExt:
431       Arg = DAG.getNode(ISD::SIGN_EXTEND, CLI.DL, VA.getLocVT(), Arg);
432       break;
433     case CCValAssign::ZExt:
434       Arg = DAG.getNode(ISD::ZERO_EXTEND, CLI.DL, VA.getLocVT(), Arg);
435       break;
436     case CCValAssign::AExt:
437       Arg = DAG.getNode(ISD::ANY_EXTEND, CLI.DL, VA.getLocVT(), Arg);
438       break;
439     }
440 
441     // Push arguments into RegsToPass vector
442     if (VA.isRegLoc())
443       RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
444     else
445       llvm_unreachable("call arg pass bug");
446   }
447 
448   SDValue InFlag;
449 
450   // Build a sequence of copy-to-reg nodes chained together with token chain and
451   // flag operands which copy the outgoing args into registers.  The InFlag in
452   // necessary since all emitted instructions must be stuck together.
453   for (auto &Reg : RegsToPass) {
454     Chain = DAG.getCopyToReg(Chain, CLI.DL, Reg.first, Reg.second, InFlag);
455     InFlag = Chain.getValue(1);
456   }
457 
458   // If the callee is a GlobalAddress node (quite common, every direct call is)
459   // turn it into a TargetGlobalAddress node so that legalize doesn't hack it.
460   // Likewise ExternalSymbol -> TargetExternalSymbol.
461   if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
462     Callee = DAG.getTargetGlobalAddress(G->getGlobal(), CLI.DL, PtrVT,
463                                         G->getOffset(), 0);
464   } else if (ExternalSymbolSDNode *E = dyn_cast<ExternalSymbolSDNode>(Callee)) {
465     Callee = DAG.getTargetExternalSymbol(E->getSymbol(), PtrVT, 0);
466     fail(CLI.DL, DAG, Twine("A call to built-in function '"
467                             + StringRef(E->getSymbol())
468                             + "' is not supported."));
469   }
470 
471   // Returns a chain & a flag for retval copy to use.
472   SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
473   SmallVector<SDValue, 8> Ops;
474   Ops.push_back(Chain);
475   Ops.push_back(Callee);
476 
477   // Add argument registers to the end of the list so that they are
478   // known live into the call.
479   for (auto &Reg : RegsToPass)
480     Ops.push_back(DAG.getRegister(Reg.first, Reg.second.getValueType()));
481 
482   if (InFlag.getNode())
483     Ops.push_back(InFlag);
484 
485   Chain = DAG.getNode(BPFISD::CALL, CLI.DL, NodeTys, Ops);
486   InFlag = Chain.getValue(1);
487 
488   // Create the CALLSEQ_END node.
489   Chain = DAG.getCALLSEQ_END(
490       Chain, DAG.getConstant(NumBytes, CLI.DL, PtrVT, true),
491       DAG.getConstant(0, CLI.DL, PtrVT, true), InFlag, CLI.DL);
492   InFlag = Chain.getValue(1);
493 
494   // Handle result values, copying them out of physregs into vregs that we
495   // return.
496   return LowerCallResult(Chain, InFlag, CallConv, IsVarArg, Ins, CLI.DL, DAG,
497                          InVals);
498 }
499 
500 SDValue
501 BPFTargetLowering::LowerReturn(SDValue Chain, CallingConv::ID CallConv,
502                                bool IsVarArg,
503                                const SmallVectorImpl<ISD::OutputArg> &Outs,
504                                const SmallVectorImpl<SDValue> &OutVals,
505                                const SDLoc &DL, SelectionDAG &DAG) const {
506   unsigned Opc = BPFISD::RET_FLAG;
507 
508   // CCValAssign - represent the assignment of the return value to a location
509   SmallVector<CCValAssign, 16> RVLocs;
510   MachineFunction &MF = DAG.getMachineFunction();
511 
512   // CCState - Info about the registers and stack slot.
513   CCState CCInfo(CallConv, IsVarArg, MF, RVLocs, *DAG.getContext());
514 
515   if (MF.getFunction().getReturnType()->isAggregateType()) {
516     fail(DL, DAG, "only integer returns supported");
517     return DAG.getNode(Opc, DL, MVT::Other, Chain);
518   }
519 
520   // Analize return values.
521   CCInfo.AnalyzeReturn(Outs, getHasAlu32() ? RetCC_BPF32 : RetCC_BPF64);
522 
523   SDValue Flag;
524   SmallVector<SDValue, 4> RetOps(1, Chain);
525 
526   // Copy the result values into the output registers.
527   for (unsigned i = 0; i != RVLocs.size(); ++i) {
528     CCValAssign &VA = RVLocs[i];
529     assert(VA.isRegLoc() && "Can only return in registers!");
530 
531     Chain = DAG.getCopyToReg(Chain, DL, VA.getLocReg(), OutVals[i], Flag);
532 
533     // Guarantee that all emitted copies are stuck together,
534     // avoiding something bad.
535     Flag = Chain.getValue(1);
536     RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
537   }
538 
539   RetOps[0] = Chain; // Update chain.
540 
541   // Add the flag if we have it.
542   if (Flag.getNode())
543     RetOps.push_back(Flag);
544 
545   return DAG.getNode(Opc, DL, MVT::Other, RetOps);
546 }
547 
548 SDValue BPFTargetLowering::LowerCallResult(
549     SDValue Chain, SDValue InFlag, CallingConv::ID CallConv, bool IsVarArg,
550     const SmallVectorImpl<ISD::InputArg> &Ins, const SDLoc &DL,
551     SelectionDAG &DAG, SmallVectorImpl<SDValue> &InVals) const {
552 
553   MachineFunction &MF = DAG.getMachineFunction();
554   // Assign locations to each value returned by this call.
555   SmallVector<CCValAssign, 16> RVLocs;
556   CCState CCInfo(CallConv, IsVarArg, MF, RVLocs, *DAG.getContext());
557 
558   if (Ins.size() >= 2) {
559     fail(DL, DAG, "only small returns supported");
560     for (unsigned i = 0, e = Ins.size(); i != e; ++i)
561       InVals.push_back(DAG.getConstant(0, DL, Ins[i].VT));
562     return DAG.getCopyFromReg(Chain, DL, 1, Ins[0].VT, InFlag).getValue(1);
563   }
564 
565   CCInfo.AnalyzeCallResult(Ins, getHasAlu32() ? RetCC_BPF32 : RetCC_BPF64);
566 
567   // Copy all of the result registers out of their specified physreg.
568   for (auto &Val : RVLocs) {
569     Chain = DAG.getCopyFromReg(Chain, DL, Val.getLocReg(),
570                                Val.getValVT(), InFlag).getValue(1);
571     InFlag = Chain.getValue(2);
572     InVals.push_back(Chain.getValue(0));
573   }
574 
575   return Chain;
576 }
577 
578 static void NegateCC(SDValue &LHS, SDValue &RHS, ISD::CondCode &CC) {
579   switch (CC) {
580   default:
581     break;
582   case ISD::SETULT:
583   case ISD::SETULE:
584   case ISD::SETLT:
585   case ISD::SETLE:
586     CC = ISD::getSetCCSwappedOperands(CC);
587     std::swap(LHS, RHS);
588     break;
589   }
590 }
591 
592 SDValue BPFTargetLowering::LowerBR_CC(SDValue Op, SelectionDAG &DAG) const {
593   SDValue Chain = Op.getOperand(0);
594   ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
595   SDValue LHS = Op.getOperand(2);
596   SDValue RHS = Op.getOperand(3);
597   SDValue Dest = Op.getOperand(4);
598   SDLoc DL(Op);
599 
600   if (!getHasJmpExt())
601     NegateCC(LHS, RHS, CC);
602 
603   return DAG.getNode(BPFISD::BR_CC, DL, Op.getValueType(), Chain, LHS, RHS,
604                      DAG.getConstant(CC, DL, LHS.getValueType()), Dest);
605 }
606 
607 SDValue BPFTargetLowering::LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const {
608   SDValue LHS = Op.getOperand(0);
609   SDValue RHS = Op.getOperand(1);
610   SDValue TrueV = Op.getOperand(2);
611   SDValue FalseV = Op.getOperand(3);
612   ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
613   SDLoc DL(Op);
614 
615   if (!getHasJmpExt())
616     NegateCC(LHS, RHS, CC);
617 
618   SDValue TargetCC = DAG.getConstant(CC, DL, LHS.getValueType());
619   SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::Glue);
620   SDValue Ops[] = {LHS, RHS, TargetCC, TrueV, FalseV};
621 
622   return DAG.getNode(BPFISD::SELECT_CC, DL, VTs, Ops);
623 }
624 
625 const char *BPFTargetLowering::getTargetNodeName(unsigned Opcode) const {
626   switch ((BPFISD::NodeType)Opcode) {
627   case BPFISD::FIRST_NUMBER:
628     break;
629   case BPFISD::RET_FLAG:
630     return "BPFISD::RET_FLAG";
631   case BPFISD::CALL:
632     return "BPFISD::CALL";
633   case BPFISD::SELECT_CC:
634     return "BPFISD::SELECT_CC";
635   case BPFISD::BR_CC:
636     return "BPFISD::BR_CC";
637   case BPFISD::Wrapper:
638     return "BPFISD::Wrapper";
639   case BPFISD::MEMCPY:
640     return "BPFISD::MEMCPY";
641   }
642   return nullptr;
643 }
644 
645 SDValue BPFTargetLowering::LowerGlobalAddress(SDValue Op,
646                                               SelectionDAG &DAG) const {
647   auto N = cast<GlobalAddressSDNode>(Op);
648   assert(N->getOffset() == 0 && "Invalid offset for global address");
649 
650   SDLoc DL(Op);
651   const GlobalValue *GV = N->getGlobal();
652   SDValue GA = DAG.getTargetGlobalAddress(GV, DL, MVT::i64);
653 
654   return DAG.getNode(BPFISD::Wrapper, DL, MVT::i64, GA);
655 }
656 
657 unsigned
658 BPFTargetLowering::EmitSubregExt(MachineInstr &MI, MachineBasicBlock *BB,
659                                  unsigned Reg, bool isSigned) const {
660   const TargetInstrInfo &TII = *BB->getParent()->getSubtarget().getInstrInfo();
661   const TargetRegisterClass *RC = getRegClassFor(MVT::i64);
662   int RShiftOp = isSigned ? BPF::SRA_ri : BPF::SRL_ri;
663   MachineFunction *F = BB->getParent();
664   DebugLoc DL = MI.getDebugLoc();
665 
666   MachineRegisterInfo &RegInfo = F->getRegInfo();
667 
668   if (!isSigned) {
669     Register PromotedReg0 = RegInfo.createVirtualRegister(RC);
670     BuildMI(BB, DL, TII.get(BPF::MOV_32_64), PromotedReg0).addReg(Reg);
671     return PromotedReg0;
672   }
673   Register PromotedReg0 = RegInfo.createVirtualRegister(RC);
674   Register PromotedReg1 = RegInfo.createVirtualRegister(RC);
675   Register PromotedReg2 = RegInfo.createVirtualRegister(RC);
676   BuildMI(BB, DL, TII.get(BPF::MOV_32_64), PromotedReg0).addReg(Reg);
677   BuildMI(BB, DL, TII.get(BPF::SLL_ri), PromotedReg1)
678     .addReg(PromotedReg0).addImm(32);
679   BuildMI(BB, DL, TII.get(RShiftOp), PromotedReg2)
680     .addReg(PromotedReg1).addImm(32);
681 
682   return PromotedReg2;
683 }
684 
685 MachineBasicBlock *
686 BPFTargetLowering::EmitInstrWithCustomInserterMemcpy(MachineInstr &MI,
687                                                      MachineBasicBlock *BB)
688                                                      const {
689   MachineFunction *MF = MI.getParent()->getParent();
690   MachineRegisterInfo &MRI = MF->getRegInfo();
691   MachineInstrBuilder MIB(*MF, MI);
692   unsigned ScratchReg;
693 
694   // This function does custom insertion during lowering BPFISD::MEMCPY which
695   // only has two register operands from memcpy semantics, the copy source
696   // address and the copy destination address.
697   //
698   // Because we will expand BPFISD::MEMCPY into load/store pairs, we will need
699   // a third scratch register to serve as the destination register of load and
700   // source register of store.
701   //
702   // The scratch register here is with the Define | Dead | EarlyClobber flags.
703   // The EarlyClobber flag has the semantic property that the operand it is
704   // attached to is clobbered before the rest of the inputs are read. Hence it
705   // must be unique among the operands to the instruction. The Define flag is
706   // needed to coerce the machine verifier that an Undef value isn't a problem
707   // as we anyway is loading memory into it. The Dead flag is needed as the
708   // value in scratch isn't supposed to be used by any other instruction.
709   ScratchReg = MRI.createVirtualRegister(&BPF::GPRRegClass);
710   MIB.addReg(ScratchReg,
711              RegState::Define | RegState::Dead | RegState::EarlyClobber);
712 
713   return BB;
714 }
715 
716 MachineBasicBlock *
717 BPFTargetLowering::EmitInstrWithCustomInserter(MachineInstr &MI,
718                                                MachineBasicBlock *BB) const {
719   const TargetInstrInfo &TII = *BB->getParent()->getSubtarget().getInstrInfo();
720   DebugLoc DL = MI.getDebugLoc();
721   unsigned Opc = MI.getOpcode();
722   bool isSelectRROp = (Opc == BPF::Select ||
723                        Opc == BPF::Select_64_32 ||
724                        Opc == BPF::Select_32 ||
725                        Opc == BPF::Select_32_64);
726 
727   bool isMemcpyOp = Opc == BPF::MEMCPY;
728 
729 #ifndef NDEBUG
730   bool isSelectRIOp = (Opc == BPF::Select_Ri ||
731                        Opc == BPF::Select_Ri_64_32 ||
732                        Opc == BPF::Select_Ri_32 ||
733                        Opc == BPF::Select_Ri_32_64);
734 
735 
736   assert((isSelectRROp || isSelectRIOp || isMemcpyOp) &&
737          "Unexpected instr type to insert");
738 #endif
739 
740   if (isMemcpyOp)
741     return EmitInstrWithCustomInserterMemcpy(MI, BB);
742 
743   bool is32BitCmp = (Opc == BPF::Select_32 ||
744                      Opc == BPF::Select_32_64 ||
745                      Opc == BPF::Select_Ri_32 ||
746                      Opc == BPF::Select_Ri_32_64);
747 
748   // To "insert" a SELECT instruction, we actually have to insert the diamond
749   // control-flow pattern.  The incoming instruction knows the destination vreg
750   // to set, the condition code register to branch on, the true/false values to
751   // select between, and a branch opcode to use.
752   const BasicBlock *LLVM_BB = BB->getBasicBlock();
753   MachineFunction::iterator I = ++BB->getIterator();
754 
755   // ThisMBB:
756   // ...
757   //  TrueVal = ...
758   //  jmp_XX r1, r2 goto Copy1MBB
759   //  fallthrough --> Copy0MBB
760   MachineBasicBlock *ThisMBB = BB;
761   MachineFunction *F = BB->getParent();
762   MachineBasicBlock *Copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
763   MachineBasicBlock *Copy1MBB = F->CreateMachineBasicBlock(LLVM_BB);
764 
765   F->insert(I, Copy0MBB);
766   F->insert(I, Copy1MBB);
767   // Update machine-CFG edges by transferring all successors of the current
768   // block to the new block which will contain the Phi node for the select.
769   Copy1MBB->splice(Copy1MBB->begin(), BB,
770                    std::next(MachineBasicBlock::iterator(MI)), BB->end());
771   Copy1MBB->transferSuccessorsAndUpdatePHIs(BB);
772   // Next, add the true and fallthrough blocks as its successors.
773   BB->addSuccessor(Copy0MBB);
774   BB->addSuccessor(Copy1MBB);
775 
776   // Insert Branch if Flag
777   int CC = MI.getOperand(3).getImm();
778   int NewCC;
779   switch (CC) {
780 #define SET_NEWCC(X, Y) \
781   case ISD::X: \
782     if (is32BitCmp && HasJmp32) \
783       NewCC = isSelectRROp ? BPF::Y##_rr_32 : BPF::Y##_ri_32; \
784     else \
785       NewCC = isSelectRROp ? BPF::Y##_rr : BPF::Y##_ri; \
786     break
787   SET_NEWCC(SETGT, JSGT);
788   SET_NEWCC(SETUGT, JUGT);
789   SET_NEWCC(SETGE, JSGE);
790   SET_NEWCC(SETUGE, JUGE);
791   SET_NEWCC(SETEQ, JEQ);
792   SET_NEWCC(SETNE, JNE);
793   SET_NEWCC(SETLT, JSLT);
794   SET_NEWCC(SETULT, JULT);
795   SET_NEWCC(SETLE, JSLE);
796   SET_NEWCC(SETULE, JULE);
797   default:
798     report_fatal_error("unimplemented select CondCode " + Twine(CC));
799   }
800 
801   Register LHS = MI.getOperand(1).getReg();
802   bool isSignedCmp = (CC == ISD::SETGT ||
803                       CC == ISD::SETGE ||
804                       CC == ISD::SETLT ||
805                       CC == ISD::SETLE);
806 
807   // eBPF at the moment only has 64-bit comparison. Any 32-bit comparison need
808   // to be promoted, however if the 32-bit comparison operands are destination
809   // registers then they are implicitly zero-extended already, there is no
810   // need of explicit zero-extend sequence for them.
811   //
812   // We simply do extension for all situations in this method, but we will
813   // try to remove those unnecessary in BPFMIPeephole pass.
814   if (is32BitCmp && !HasJmp32)
815     LHS = EmitSubregExt(MI, BB, LHS, isSignedCmp);
816 
817   if (isSelectRROp) {
818     Register RHS = MI.getOperand(2).getReg();
819 
820     if (is32BitCmp && !HasJmp32)
821       RHS = EmitSubregExt(MI, BB, RHS, isSignedCmp);
822 
823     BuildMI(BB, DL, TII.get(NewCC)).addReg(LHS).addReg(RHS).addMBB(Copy1MBB);
824   } else {
825     int64_t imm32 = MI.getOperand(2).getImm();
826     // Check before we build J*_ri instruction.
827     assert (isInt<32>(imm32));
828     BuildMI(BB, DL, TII.get(NewCC))
829         .addReg(LHS).addImm(imm32).addMBB(Copy1MBB);
830   }
831 
832   // Copy0MBB:
833   //  %FalseValue = ...
834   //  # fallthrough to Copy1MBB
835   BB = Copy0MBB;
836 
837   // Update machine-CFG edges
838   BB->addSuccessor(Copy1MBB);
839 
840   // Copy1MBB:
841   //  %Result = phi [ %FalseValue, Copy0MBB ], [ %TrueValue, ThisMBB ]
842   // ...
843   BB = Copy1MBB;
844   BuildMI(*BB, BB->begin(), DL, TII.get(BPF::PHI), MI.getOperand(0).getReg())
845       .addReg(MI.getOperand(5).getReg())
846       .addMBB(Copy0MBB)
847       .addReg(MI.getOperand(4).getReg())
848       .addMBB(ThisMBB);
849 
850   MI.eraseFromParent(); // The pseudo instruction is gone now.
851   return BB;
852 }
853 
854 EVT BPFTargetLowering::getSetCCResultType(const DataLayout &, LLVMContext &,
855                                           EVT VT) const {
856   return getHasAlu32() ? MVT::i32 : MVT::i64;
857 }
858 
859 MVT BPFTargetLowering::getScalarShiftAmountTy(const DataLayout &DL,
860                                               EVT VT) const {
861   return (getHasAlu32() && VT == MVT::i32) ? MVT::i32 : MVT::i64;
862 }
863 
864 bool BPFTargetLowering::isLegalAddressingMode(const DataLayout &DL,
865                                               const AddrMode &AM, Type *Ty,
866                                               unsigned AS,
867                                               Instruction *I) const {
868   // No global is ever allowed as a base.
869   if (AM.BaseGV)
870     return false;
871 
872   switch (AM.Scale) {
873   case 0: // "r+i" or just "i", depending on HasBaseReg.
874     break;
875   case 1:
876     if (!AM.HasBaseReg) // allow "r+i".
877       break;
878     return false; // disallow "r+r" or "r+r+i".
879   default:
880     return false;
881   }
882 
883   return true;
884 }
885