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/MathExtras.h"
30#include "llvm/Support/raw_ostream.h"
31
32using namespace llvm;
33
34#define DEBUG_TYPE "bpf-lower"
35
36static cl::opt<bool> BPFExpandMemcpyInOrder("bpf-expand-memcpy-in-order",
37 cl::Hidden, cl::init(Val: false),
38 cl::desc("Expand memcpy into load/store pairs in order"));
39
40static void fail(const SDLoc &DL, SelectionDAG &DAG, const Twine &Msg,
41 SDValue Val = {}) {
42 std::string Str;
43 if (Val) {
44 raw_string_ostream OS(Str);
45 Val->print(OS);
46 OS << ' ';
47 }
48 MachineFunction &MF = DAG.getMachineFunction();
49 DAG.getContext()->diagnose(DI: DiagnosticInfoUnsupported(
50 MF.getFunction(), Twine(Str).concat(Suffix: Msg), DL.getDebugLoc()));
51}
52
53BPFTargetLowering::BPFTargetLowering(const TargetMachine &TM,
54 const BPFSubtarget &STI)
55 : TargetLowering(TM) {
56
57 // Set up the register classes.
58 addRegisterClass(MVT::i64, &BPF::GPRRegClass);
59 if (STI.getHasAlu32())
60 addRegisterClass(MVT::i32, &BPF::GPR32RegClass);
61
62 // Compute derived properties from the register classes
63 computeRegisterProperties(STI.getRegisterInfo());
64
65 setStackPointerRegisterToSaveRestore(BPF::R11);
66
67 setOperationAction(ISD::BR_CC, MVT::i64, Custom);
68 setOperationAction(ISD::BR_JT, MVT::Other, Expand);
69 setOperationAction(ISD::BRIND, MVT::Other, Expand);
70 setOperationAction(ISD::BRCOND, MVT::Other, Expand);
71
72 setOperationAction({ISD::GlobalAddress, ISD::ConstantPool}, MVT::i64, Custom);
73
74 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i64, Custom);
75 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
76 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
77
78 // Set unsupported atomic operations as Custom so
79 // we can emit better error messages than fatal error
80 // from selectiondag.
81 for (auto VT : {MVT::i8, MVT::i16, MVT::i32}) {
82 if (VT == MVT::i32) {
83 if (STI.getHasAlu32())
84 continue;
85 } else {
86 setOperationAction(ISD::ATOMIC_LOAD_ADD, VT, Custom);
87 }
88
89 setOperationAction(ISD::ATOMIC_LOAD_AND, VT, Custom);
90 setOperationAction(ISD::ATOMIC_LOAD_OR, VT, Custom);
91 setOperationAction(ISD::ATOMIC_LOAD_XOR, VT, Custom);
92 setOperationAction(ISD::ATOMIC_SWAP, VT, Custom);
93 setOperationAction(ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS, VT, Custom);
94 }
95
96 for (auto VT : { MVT::i32, MVT::i64 }) {
97 if (VT == MVT::i32 && !STI.getHasAlu32())
98 continue;
99
100 setOperationAction(ISD::SDIVREM, VT, Expand);
101 setOperationAction(ISD::UDIVREM, VT, Expand);
102 if (!STI.hasSdivSmod()) {
103 setOperationAction(ISD::SDIV, VT, Custom);
104 setOperationAction(ISD::SREM, VT, Custom);
105 }
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 setOperationAction(ISD::CTTZ, VT, Expand);
117 setOperationAction(ISD::CTLZ, VT, Expand);
118 setOperationAction(ISD::CTTZ_ZERO_UNDEF, VT, Expand);
119 setOperationAction(ISD::CTLZ_ZERO_UNDEF, VT, Expand);
120
121 setOperationAction(ISD::SETCC, VT, Expand);
122 setOperationAction(ISD::SELECT, VT, Expand);
123 setOperationAction(ISD::SELECT_CC, VT, Custom);
124 }
125
126 if (STI.getHasAlu32()) {
127 setOperationAction(ISD::BSWAP, MVT::i32, Promote);
128 setOperationAction(ISD::BR_CC, MVT::i32,
129 STI.getHasJmp32() ? Custom : Promote);
130 }
131
132 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand);
133 if (!STI.hasMovsx()) {
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
139 // Extended load operations for i1 types must be promoted
140 for (MVT VT : MVT::integer_valuetypes()) {
141 setLoadExtAction(ISD::EXTLOAD, VT, MVT::i1, Promote);
142 setLoadExtAction(ISD::ZEXTLOAD, VT, MVT::i1, Promote);
143 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::i1, Promote);
144
145 if (!STI.hasLdsx()) {
146 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::i8, Expand);
147 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::i16, Expand);
148 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::i32, Expand);
149 }
150 }
151
152 setBooleanContents(ZeroOrOneBooleanContent);
153 setMaxAtomicSizeInBitsSupported(64);
154
155 // Function alignments
156 setMinFunctionAlignment(Align(8));
157 setPrefFunctionAlignment(Align(8));
158
159 if (BPFExpandMemcpyInOrder) {
160 // LLVM generic code will try to expand memcpy into load/store pairs at this
161 // stage which is before quite a few IR optimization passes, therefore the
162 // loads and stores could potentially be moved apart from each other which
163 // will cause trouble to memcpy pattern matcher inside kernel eBPF JIT
164 // compilers.
165 //
166 // When -bpf-expand-memcpy-in-order specified, we want to defer the expand
167 // of memcpy to later stage in IR optimization pipeline so those load/store
168 // pairs won't be touched and could be kept in order. Hence, we set
169 // MaxStoresPerMem* to zero to disable the generic getMemcpyLoadsAndStores
170 // code path, and ask LLVM to use target expander EmitTargetCodeForMemcpy.
171 MaxStoresPerMemset = MaxStoresPerMemsetOptSize = 0;
172 MaxStoresPerMemcpy = MaxStoresPerMemcpyOptSize = 0;
173 MaxStoresPerMemmove = MaxStoresPerMemmoveOptSize = 0;
174 MaxLoadsPerMemcmp = 0;
175 } else {
176 // inline memcpy() for kernel to see explicit copy
177 unsigned CommonMaxStores =
178 STI.getSelectionDAGInfo()->getCommonMaxStoresPerMemFunc();
179
180 MaxStoresPerMemset = MaxStoresPerMemsetOptSize = CommonMaxStores;
181 MaxStoresPerMemcpy = MaxStoresPerMemcpyOptSize = CommonMaxStores;
182 MaxStoresPerMemmove = MaxStoresPerMemmoveOptSize = CommonMaxStores;
183 MaxLoadsPerMemcmp = MaxLoadsPerMemcmpOptSize = CommonMaxStores;
184 }
185
186 // CPU/Feature control
187 HasAlu32 = STI.getHasAlu32();
188 HasJmp32 = STI.getHasJmp32();
189 HasJmpExt = STI.getHasJmpExt();
190 HasMovsx = STI.hasMovsx();
191}
192
193bool BPFTargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
194 return false;
195}
196
197bool BPFTargetLowering::isTruncateFree(Type *Ty1, Type *Ty2) const {
198 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
199 return false;
200 unsigned NumBits1 = Ty1->getPrimitiveSizeInBits();
201 unsigned NumBits2 = Ty2->getPrimitiveSizeInBits();
202 return NumBits1 > NumBits2;
203}
204
205bool BPFTargetLowering::isTruncateFree(EVT VT1, EVT VT2) const {
206 if (!VT1.isInteger() || !VT2.isInteger())
207 return false;
208 unsigned NumBits1 = VT1.getSizeInBits();
209 unsigned NumBits2 = VT2.getSizeInBits();
210 return NumBits1 > NumBits2;
211}
212
213bool BPFTargetLowering::isZExtFree(Type *Ty1, Type *Ty2) const {
214 if (!getHasAlu32() || !Ty1->isIntegerTy() || !Ty2->isIntegerTy())
215 return false;
216 unsigned NumBits1 = Ty1->getPrimitiveSizeInBits();
217 unsigned NumBits2 = Ty2->getPrimitiveSizeInBits();
218 return NumBits1 == 32 && NumBits2 == 64;
219}
220
221bool BPFTargetLowering::isZExtFree(EVT VT1, EVT VT2) const {
222 if (!getHasAlu32() || !VT1.isInteger() || !VT2.isInteger())
223 return false;
224 unsigned NumBits1 = VT1.getSizeInBits();
225 unsigned NumBits2 = VT2.getSizeInBits();
226 return NumBits1 == 32 && NumBits2 == 64;
227}
228
229bool BPFTargetLowering::isZExtFree(SDValue Val, EVT VT2) const {
230 EVT VT1 = Val.getValueType();
231 if (Val.getOpcode() == ISD::LOAD && VT1.isSimple() && VT2.isSimple()) {
232 MVT MT1 = VT1.getSimpleVT().SimpleTy;
233 MVT MT2 = VT2.getSimpleVT().SimpleTy;
234 if ((MT1 == MVT::i8 || MT1 == MVT::i16 || MT1 == MVT::i32) &&
235 (MT2 == MVT::i32 || MT2 == MVT::i64))
236 return true;
237 }
238 return TargetLoweringBase::isZExtFree(Val, VT2);
239}
240
241BPFTargetLowering::ConstraintType
242BPFTargetLowering::getConstraintType(StringRef Constraint) const {
243 if (Constraint.size() == 1) {
244 switch (Constraint[0]) {
245 default:
246 break;
247 case 'w':
248 return C_RegisterClass;
249 }
250 }
251
252 return TargetLowering::getConstraintType(Constraint);
253}
254
255std::pair<unsigned, const TargetRegisterClass *>
256BPFTargetLowering::getRegForInlineAsmConstraint(const TargetRegisterInfo *TRI,
257 StringRef Constraint,
258 MVT VT) const {
259 if (Constraint.size() == 1) {
260 // GCC Constraint Letters
261 switch (Constraint[0]) {
262 case 'r': // GENERAL_REGS
263 return std::make_pair(0U, &BPF::GPRRegClass);
264 case 'w':
265 if (HasAlu32)
266 return std::make_pair(0U, &BPF::GPR32RegClass);
267 break;
268 default:
269 break;
270 }
271 }
272
273 return TargetLowering::getRegForInlineAsmConstraint(TRI, Constraint, VT);
274}
275
276void BPFTargetLowering::ReplaceNodeResults(
277 SDNode *N, SmallVectorImpl<SDValue> &Results, SelectionDAG &DAG) const {
278 const char *Msg;
279 uint32_t Opcode = N->getOpcode();
280 switch (Opcode) {
281 default:
282 report_fatal_error(reason: "unhandled custom legalization: " + Twine(Opcode));
283 case ISD::ATOMIC_LOAD_ADD:
284 case ISD::ATOMIC_LOAD_AND:
285 case ISD::ATOMIC_LOAD_OR:
286 case ISD::ATOMIC_LOAD_XOR:
287 case ISD::ATOMIC_SWAP:
288 case ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS:
289 if (HasAlu32 || Opcode == ISD::ATOMIC_LOAD_ADD)
290 Msg = "unsupported atomic operation, please use 32/64 bit version";
291 else
292 Msg = "unsupported atomic operation, please use 64 bit version";
293 break;
294 }
295
296 SDLoc DL(N);
297 // We'll still produce a fatal error downstream, but this diagnostic is more
298 // user-friendly.
299 fail(DL, DAG, Msg);
300}
301
302SDValue BPFTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
303 switch (Op.getOpcode()) {
304 default:
305 report_fatal_error(reason: "unimplemented opcode: " + Twine(Op.getOpcode()));
306 case ISD::BR_CC:
307 return LowerBR_CC(Op, DAG);
308 case ISD::GlobalAddress:
309 return LowerGlobalAddress(Op, DAG);
310 case ISD::ConstantPool:
311 return LowerConstantPool(Op, DAG);
312 case ISD::SELECT_CC:
313 return LowerSELECT_CC(Op, DAG);
314 case ISD::SDIV:
315 case ISD::SREM:
316 return LowerSDIVSREM(Op, DAG);
317 case ISD::DYNAMIC_STACKALLOC:
318 return LowerDYNAMIC_STACKALLOC(Op, DAG);
319 }
320}
321
322// Calling Convention Implementation
323#include "BPFGenCallingConv.inc"
324
325SDValue BPFTargetLowering::LowerFormalArguments(
326 SDValue Chain, CallingConv::ID CallConv, bool IsVarArg,
327 const SmallVectorImpl<ISD::InputArg> &Ins, const SDLoc &DL,
328 SelectionDAG &DAG, SmallVectorImpl<SDValue> &InVals) const {
329 switch (CallConv) {
330 default:
331 report_fatal_error(reason: "unimplemented calling convention: " + Twine(CallConv));
332 case CallingConv::C:
333 case CallingConv::Fast:
334 break;
335 }
336
337 MachineFunction &MF = DAG.getMachineFunction();
338 MachineRegisterInfo &RegInfo = MF.getRegInfo();
339
340 // Assign locations to all of the incoming arguments.
341 SmallVector<CCValAssign, 16> ArgLocs;
342 CCState CCInfo(CallConv, IsVarArg, MF, ArgLocs, *DAG.getContext());
343 CCInfo.AnalyzeFormalArguments(Ins, getHasAlu32() ? CC_BPF32 : CC_BPF64);
344
345 bool HasMemArgs = false;
346 for (size_t I = 0; I < ArgLocs.size(); ++I) {
347 auto &VA = ArgLocs[I];
348
349 if (VA.isRegLoc()) {
350 // Arguments passed in registers
351 EVT RegVT = VA.getLocVT();
352 MVT::SimpleValueType SimpleTy = RegVT.getSimpleVT().SimpleTy;
353 switch (SimpleTy) {
354 default: {
355 std::string Str;
356 {
357 raw_string_ostream OS(Str);
358 RegVT.print(OS);
359 }
360 report_fatal_error(reason: "unhandled argument type: " + Twine(Str));
361 }
362 case MVT::i32:
363 case MVT::i64:
364 Register VReg = RegInfo.createVirtualRegister(
365 SimpleTy == MVT::i64 ? &BPF::GPRRegClass : &BPF::GPR32RegClass);
366 RegInfo.addLiveIn(Reg: VA.getLocReg(), vreg: VReg);
367 SDValue ArgValue = DAG.getCopyFromReg(Chain, dl: DL, Reg: VReg, VT: RegVT);
368
369 // If this is an value that has been promoted to wider types, insert an
370 // assert[sz]ext to capture this, then truncate to the right size.
371 if (VA.getLocInfo() == CCValAssign::SExt)
372 ArgValue = DAG.getNode(ISD::AssertSext, DL, RegVT, ArgValue,
373 DAG.getValueType(VA.getValVT()));
374 else if (VA.getLocInfo() == CCValAssign::ZExt)
375 ArgValue = DAG.getNode(ISD::AssertZext, DL, RegVT, ArgValue,
376 DAG.getValueType(VA.getValVT()));
377
378 if (VA.getLocInfo() != CCValAssign::Full)
379 ArgValue = DAG.getNode(ISD::TRUNCATE, DL, VA.getValVT(), ArgValue);
380
381 InVals.push_back(Elt: ArgValue);
382
383 break;
384 }
385 } else {
386 if (VA.isMemLoc())
387 HasMemArgs = true;
388 else
389 report_fatal_error(reason: "unhandled argument location");
390 InVals.push_back(Elt: DAG.getConstant(0, DL, VA.getLocVT()));
391 }
392 }
393 if (HasMemArgs)
394 fail(DL, DAG, Msg: "stack arguments are not supported");
395 if (IsVarArg)
396 fail(DL, DAG, Msg: "variadic functions are not supported");
397 if (MF.getFunction().hasStructRetAttr())
398 fail(DL, DAG, Msg: "aggregate returns are not supported");
399
400 return Chain;
401}
402
403const size_t BPFTargetLowering::MaxArgs = 5;
404
405SDValue BPFTargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI,
406 SmallVectorImpl<SDValue> &InVals) const {
407 SelectionDAG &DAG = CLI.DAG;
408 auto &Outs = CLI.Outs;
409 auto &OutVals = CLI.OutVals;
410 auto &Ins = CLI.Ins;
411 SDValue Chain = CLI.Chain;
412 SDValue Callee = CLI.Callee;
413 bool &IsTailCall = CLI.IsTailCall;
414 CallingConv::ID CallConv = CLI.CallConv;
415 bool IsVarArg = CLI.IsVarArg;
416 MachineFunction &MF = DAG.getMachineFunction();
417
418 // BPF target does not support tail call optimization.
419 IsTailCall = false;
420
421 switch (CallConv) {
422 default:
423 report_fatal_error(reason: "unsupported calling convention: " + Twine(CallConv));
424 case CallingConv::Fast:
425 case CallingConv::C:
426 break;
427 }
428
429 // Analyze operands of the call, assigning locations to each operand.
430 SmallVector<CCValAssign, 16> ArgLocs;
431 CCState CCInfo(CallConv, IsVarArg, MF, ArgLocs, *DAG.getContext());
432
433 CCInfo.AnalyzeCallOperands(Outs, getHasAlu32() ? CC_BPF32 : CC_BPF64);
434
435 unsigned NumBytes = CCInfo.getStackSize();
436
437 if (Outs.size() > MaxArgs)
438 fail(DL: CLI.DL, DAG, Msg: "too many arguments", Val: Callee);
439
440 for (auto &Arg : Outs) {
441 ISD::ArgFlagsTy Flags = Arg.Flags;
442 if (!Flags.isByVal())
443 continue;
444 fail(DL: CLI.DL, DAG, Msg: "pass by value not supported", Val: Callee);
445 break;
446 }
447
448 auto PtrVT = getPointerTy(DL: MF.getDataLayout());
449 Chain = DAG.getCALLSEQ_START(Chain, InSize: NumBytes, OutSize: 0, DL: CLI.DL);
450
451 SmallVector<std::pair<unsigned, SDValue>, MaxArgs> RegsToPass;
452
453 // Walk arg assignments
454 for (size_t i = 0; i < std::min(a: ArgLocs.size(), b: MaxArgs); ++i) {
455 CCValAssign &VA = ArgLocs[i];
456 SDValue &Arg = OutVals[i];
457
458 // Promote the value if needed.
459 switch (VA.getLocInfo()) {
460 default:
461 report_fatal_error(reason: "unhandled location info: " + Twine(VA.getLocInfo()));
462 case CCValAssign::Full:
463 break;
464 case CCValAssign::SExt:
465 Arg = DAG.getNode(Opcode: ISD::SIGN_EXTEND, DL: CLI.DL, VT: VA.getLocVT(), Operand: Arg);
466 break;
467 case CCValAssign::ZExt:
468 Arg = DAG.getNode(Opcode: ISD::ZERO_EXTEND, DL: CLI.DL, VT: VA.getLocVT(), Operand: Arg);
469 break;
470 case CCValAssign::AExt:
471 Arg = DAG.getNode(Opcode: ISD::ANY_EXTEND, DL: CLI.DL, VT: VA.getLocVT(), Operand: Arg);
472 break;
473 }
474
475 // Push arguments into RegsToPass vector
476 if (VA.isRegLoc())
477 RegsToPass.push_back(Elt: std::make_pair(x: VA.getLocReg(), y&: Arg));
478 else
479 report_fatal_error(reason: "stack arguments are not supported");
480 }
481
482 SDValue InGlue;
483
484 // Build a sequence of copy-to-reg nodes chained together with token chain and
485 // flag operands which copy the outgoing args into registers. The InGlue in
486 // necessary since all emitted instructions must be stuck together.
487 for (auto &Reg : RegsToPass) {
488 Chain = DAG.getCopyToReg(Chain, dl: CLI.DL, Reg: Reg.first, N: Reg.second, Glue: InGlue);
489 InGlue = Chain.getValue(R: 1);
490 }
491
492 // If the callee is a GlobalAddress node (quite common, every direct call is)
493 // turn it into a TargetGlobalAddress node so that legalize doesn't hack it.
494 // Likewise ExternalSymbol -> TargetExternalSymbol.
495 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Val&: Callee)) {
496 Callee = DAG.getTargetGlobalAddress(GV: G->getGlobal(), DL: CLI.DL, VT: PtrVT,
497 offset: G->getOffset(), TargetFlags: 0);
498 } else if (ExternalSymbolSDNode *E = dyn_cast<ExternalSymbolSDNode>(Val&: Callee)) {
499 Callee = DAG.getTargetExternalSymbol(Sym: E->getSymbol(), VT: PtrVT, TargetFlags: 0);
500 fail(DL: CLI.DL, DAG,
501 Msg: Twine("A call to built-in function '" + StringRef(E->getSymbol()) +
502 "' is not supported."));
503 }
504
505 // Returns a chain & a flag for retval copy to use.
506 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
507 SmallVector<SDValue, 8> Ops;
508 Ops.push_back(Elt: Chain);
509 Ops.push_back(Elt: Callee);
510
511 // Add argument registers to the end of the list so that they are
512 // known live into the call.
513 for (auto &Reg : RegsToPass)
514 Ops.push_back(Elt: DAG.getRegister(Reg: Reg.first, VT: Reg.second.getValueType()));
515
516 if (InGlue.getNode())
517 Ops.push_back(Elt: InGlue);
518
519 Chain = DAG.getNode(Opcode: BPFISD::CALL, DL: CLI.DL, VTList: NodeTys, Ops);
520 InGlue = Chain.getValue(R: 1);
521
522 DAG.addNoMergeSiteInfo(Node: Chain.getNode(), NoMerge: CLI.NoMerge);
523
524 // Create the CALLSEQ_END node.
525 Chain = DAG.getCALLSEQ_END(Chain, Size1: NumBytes, Size2: 0, Glue: InGlue, DL: CLI.DL);
526 InGlue = Chain.getValue(R: 1);
527
528 // Handle result values, copying them out of physregs into vregs that we
529 // return.
530 return LowerCallResult(Chain, InGlue, CallConv, IsVarArg, Ins, DL: CLI.DL, DAG,
531 InVals);
532}
533
534SDValue
535BPFTargetLowering::LowerReturn(SDValue Chain, CallingConv::ID CallConv,
536 bool IsVarArg,
537 const SmallVectorImpl<ISD::OutputArg> &Outs,
538 const SmallVectorImpl<SDValue> &OutVals,
539 const SDLoc &DL, SelectionDAG &DAG) const {
540 unsigned Opc = BPFISD::RET_GLUE;
541
542 // CCValAssign - represent the assignment of the return value to a location
543 SmallVector<CCValAssign, 16> RVLocs;
544 MachineFunction &MF = DAG.getMachineFunction();
545
546 // CCState - Info about the registers and stack slot.
547 CCState CCInfo(CallConv, IsVarArg, MF, RVLocs, *DAG.getContext());
548
549 if (MF.getFunction().getReturnType()->isAggregateType()) {
550 fail(DL, DAG, Msg: "aggregate returns are not supported");
551 return DAG.getNode(Opc, DL, MVT::Other, Chain);
552 }
553
554 // Analize return values.
555 CCInfo.AnalyzeReturn(Outs, getHasAlu32() ? RetCC_BPF32 : RetCC_BPF64);
556
557 SDValue Glue;
558 SmallVector<SDValue, 4> RetOps(1, Chain);
559
560 // Copy the result values into the output registers.
561 for (size_t i = 0; i != RVLocs.size(); ++i) {
562 CCValAssign &VA = RVLocs[i];
563 if (!VA.isRegLoc())
564 report_fatal_error(reason: "stack return values are not supported");
565
566 Chain = DAG.getCopyToReg(Chain, dl: DL, Reg: VA.getLocReg(), N: OutVals[i], Glue);
567
568 // Guarantee that all emitted copies are stuck together,
569 // avoiding something bad.
570 Glue = Chain.getValue(R: 1);
571 RetOps.push_back(Elt: DAG.getRegister(Reg: VA.getLocReg(), VT: VA.getLocVT()));
572 }
573
574 RetOps[0] = Chain; // Update chain.
575
576 // Add the glue if we have it.
577 if (Glue.getNode())
578 RetOps.push_back(Elt: Glue);
579
580 return DAG.getNode(Opc, DL, MVT::Other, RetOps);
581}
582
583SDValue BPFTargetLowering::LowerCallResult(
584 SDValue Chain, SDValue InGlue, CallingConv::ID CallConv, bool IsVarArg,
585 const SmallVectorImpl<ISD::InputArg> &Ins, const SDLoc &DL,
586 SelectionDAG &DAG, SmallVectorImpl<SDValue> &InVals) const {
587
588 MachineFunction &MF = DAG.getMachineFunction();
589 // Assign locations to each value returned by this call.
590 SmallVector<CCValAssign, 16> RVLocs;
591 CCState CCInfo(CallConv, IsVarArg, MF, RVLocs, *DAG.getContext());
592
593 if (Ins.size() > 1) {
594 fail(DL, DAG, Msg: "only small returns supported");
595 for (auto &In : Ins)
596 InVals.push_back(Elt: DAG.getConstant(Val: 0, DL, VT: In.VT));
597 return DAG.getCopyFromReg(Chain, dl: DL, Reg: 1, VT: Ins[0].VT, Glue: InGlue).getValue(R: 1);
598 }
599
600 CCInfo.AnalyzeCallResult(Ins, getHasAlu32() ? RetCC_BPF32 : RetCC_BPF64);
601
602 // Copy all of the result registers out of their specified physreg.
603 for (auto &Val : RVLocs) {
604 Chain = DAG.getCopyFromReg(Chain, dl: DL, Reg: Val.getLocReg(),
605 VT: Val.getValVT(), Glue: InGlue).getValue(R: 1);
606 InGlue = Chain.getValue(R: 2);
607 InVals.push_back(Elt: Chain.getValue(R: 0));
608 }
609
610 return Chain;
611}
612
613static void NegateCC(SDValue &LHS, SDValue &RHS, ISD::CondCode &CC) {
614 switch (CC) {
615 default:
616 break;
617 case ISD::SETULT:
618 case ISD::SETULE:
619 case ISD::SETLT:
620 case ISD::SETLE:
621 CC = ISD::getSetCCSwappedOperands(Operation: CC);
622 std::swap(a&: LHS, b&: RHS);
623 break;
624 }
625}
626
627SDValue BPFTargetLowering::LowerSDIVSREM(SDValue Op, SelectionDAG &DAG) const {
628 SDLoc DL(Op);
629 fail(DL, DAG,
630 Msg: "unsupported signed division, please convert to unsigned div/mod.");
631 return DAG.getUNDEF(VT: Op->getValueType(ResNo: 0));
632}
633
634SDValue BPFTargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
635 SelectionDAG &DAG) const {
636 SDLoc DL(Op);
637 fail(DL, DAG, Msg: "unsupported dynamic stack allocation");
638 auto Ops = {DAG.getConstant(Val: 0, DL: SDLoc(), VT: Op.getValueType()), Op.getOperand(i: 0)};
639 return DAG.getMergeValues(Ops, dl: SDLoc());
640}
641
642SDValue BPFTargetLowering::LowerBR_CC(SDValue Op, SelectionDAG &DAG) const {
643 SDValue Chain = Op.getOperand(i: 0);
644 ISD::CondCode CC = cast<CondCodeSDNode>(Val: Op.getOperand(i: 1))->get();
645 SDValue LHS = Op.getOperand(i: 2);
646 SDValue RHS = Op.getOperand(i: 3);
647 SDValue Dest = Op.getOperand(i: 4);
648 SDLoc DL(Op);
649
650 if (!getHasJmpExt())
651 NegateCC(LHS, RHS, CC);
652
653 return DAG.getNode(Opcode: BPFISD::BR_CC, DL, VT: Op.getValueType(), N1: Chain, N2: LHS, N3: RHS,
654 N4: DAG.getConstant(Val: CC, DL, VT: LHS.getValueType()), N5: Dest);
655}
656
657SDValue BPFTargetLowering::LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const {
658 SDValue LHS = Op.getOperand(i: 0);
659 SDValue RHS = Op.getOperand(i: 1);
660 SDValue TrueV = Op.getOperand(i: 2);
661 SDValue FalseV = Op.getOperand(i: 3);
662 ISD::CondCode CC = cast<CondCodeSDNode>(Val: Op.getOperand(i: 4))->get();
663 SDLoc DL(Op);
664
665 if (!getHasJmpExt())
666 NegateCC(LHS, RHS, CC);
667
668 SDValue TargetCC = DAG.getConstant(Val: CC, DL, VT: LHS.getValueType());
669 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::Glue);
670 SDValue Ops[] = {LHS, RHS, TargetCC, TrueV, FalseV};
671
672 return DAG.getNode(Opcode: BPFISD::SELECT_CC, DL, VTList: VTs, Ops);
673}
674
675const char *BPFTargetLowering::getTargetNodeName(unsigned Opcode) const {
676 switch ((BPFISD::NodeType)Opcode) {
677 case BPFISD::FIRST_NUMBER:
678 break;
679 case BPFISD::RET_GLUE:
680 return "BPFISD::RET_GLUE";
681 case BPFISD::CALL:
682 return "BPFISD::CALL";
683 case BPFISD::SELECT_CC:
684 return "BPFISD::SELECT_CC";
685 case BPFISD::BR_CC:
686 return "BPFISD::BR_CC";
687 case BPFISD::Wrapper:
688 return "BPFISD::Wrapper";
689 case BPFISD::MEMCPY:
690 return "BPFISD::MEMCPY";
691 }
692 return nullptr;
693}
694
695static SDValue getTargetNode(GlobalAddressSDNode *N, const SDLoc &DL, EVT Ty,
696 SelectionDAG &DAG, unsigned Flags) {
697 return DAG.getTargetGlobalAddress(GV: N->getGlobal(), DL, VT: Ty, offset: 0, TargetFlags: Flags);
698}
699
700static SDValue getTargetNode(ConstantPoolSDNode *N, const SDLoc &DL, EVT Ty,
701 SelectionDAG &DAG, unsigned Flags) {
702 return DAG.getTargetConstantPool(C: N->getConstVal(), VT: Ty, Align: N->getAlign(),
703 Offset: N->getOffset(), TargetFlags: Flags);
704}
705
706template <class NodeTy>
707SDValue BPFTargetLowering::getAddr(NodeTy *N, SelectionDAG &DAG,
708 unsigned Flags) const {
709 SDLoc DL(N);
710
711 SDValue GA = getTargetNode(N, DL, MVT::i64, DAG, Flags);
712
713 return DAG.getNode(BPFISD::Wrapper, DL, MVT::i64, GA);
714}
715
716SDValue BPFTargetLowering::LowerGlobalAddress(SDValue Op,
717 SelectionDAG &DAG) const {
718 GlobalAddressSDNode *N = cast<GlobalAddressSDNode>(Val&: Op);
719 if (N->getOffset() != 0)
720 report_fatal_error(reason: "invalid offset for global address: " +
721 Twine(N->getOffset()));
722 return getAddr(N, DAG);
723}
724
725SDValue BPFTargetLowering::LowerConstantPool(SDValue Op,
726 SelectionDAG &DAG) const {
727 ConstantPoolSDNode *N = cast<ConstantPoolSDNode>(Val&: Op);
728
729 return getAddr(N, DAG);
730}
731
732unsigned
733BPFTargetLowering::EmitSubregExt(MachineInstr &MI, MachineBasicBlock *BB,
734 unsigned Reg, bool isSigned) const {
735 const TargetInstrInfo &TII = *BB->getParent()->getSubtarget().getInstrInfo();
736 const TargetRegisterClass *RC = getRegClassFor(MVT::i64);
737 int RShiftOp = isSigned ? BPF::SRA_ri : BPF::SRL_ri;
738 MachineFunction *F = BB->getParent();
739 DebugLoc DL = MI.getDebugLoc();
740
741 MachineRegisterInfo &RegInfo = F->getRegInfo();
742
743 if (!isSigned) {
744 Register PromotedReg0 = RegInfo.createVirtualRegister(RegClass: RC);
745 BuildMI(BB, DL, TII.get(BPF::MOV_32_64), PromotedReg0).addReg(Reg);
746 return PromotedReg0;
747 }
748 Register PromotedReg0 = RegInfo.createVirtualRegister(RegClass: RC);
749 Register PromotedReg1 = RegInfo.createVirtualRegister(RegClass: RC);
750 Register PromotedReg2 = RegInfo.createVirtualRegister(RegClass: RC);
751 if (HasMovsx) {
752 BuildMI(BB, DL, TII.get(BPF::MOVSX_rr_32), PromotedReg0).addReg(Reg);
753 } else {
754 BuildMI(BB, DL, TII.get(BPF::MOV_32_64), PromotedReg0).addReg(Reg);
755 BuildMI(BB, DL, TII.get(BPF::SLL_ri), PromotedReg1)
756 .addReg(PromotedReg0).addImm(32);
757 BuildMI(BB, MIMD: DL, MCID: TII.get(Opcode: RShiftOp), DestReg: PromotedReg2)
758 .addReg(RegNo: PromotedReg1).addImm(Val: 32);
759 }
760
761 return PromotedReg2;
762}
763
764MachineBasicBlock *
765BPFTargetLowering::EmitInstrWithCustomInserterMemcpy(MachineInstr &MI,
766 MachineBasicBlock *BB)
767 const {
768 MachineFunction *MF = MI.getParent()->getParent();
769 MachineRegisterInfo &MRI = MF->getRegInfo();
770 MachineInstrBuilder MIB(*MF, MI);
771 unsigned ScratchReg;
772
773 // This function does custom insertion during lowering BPFISD::MEMCPY which
774 // only has two register operands from memcpy semantics, the copy source
775 // address and the copy destination address.
776 //
777 // Because we will expand BPFISD::MEMCPY into load/store pairs, we will need
778 // a third scratch register to serve as the destination register of load and
779 // source register of store.
780 //
781 // The scratch register here is with the Define | Dead | EarlyClobber flags.
782 // The EarlyClobber flag has the semantic property that the operand it is
783 // attached to is clobbered before the rest of the inputs are read. Hence it
784 // must be unique among the operands to the instruction. The Define flag is
785 // needed to coerce the machine verifier that an Undef value isn't a problem
786 // as we anyway is loading memory into it. The Dead flag is needed as the
787 // value in scratch isn't supposed to be used by any other instruction.
788 ScratchReg = MRI.createVirtualRegister(&BPF::GPRRegClass);
789 MIB.addReg(RegNo: ScratchReg,
790 flags: RegState::Define | RegState::Dead | RegState::EarlyClobber);
791
792 return BB;
793}
794
795MachineBasicBlock *
796BPFTargetLowering::EmitInstrWithCustomInserter(MachineInstr &MI,
797 MachineBasicBlock *BB) const {
798 const TargetInstrInfo &TII = *BB->getParent()->getSubtarget().getInstrInfo();
799 DebugLoc DL = MI.getDebugLoc();
800 unsigned Opc = MI.getOpcode();
801 bool isSelectRROp = (Opc == BPF::Select ||
802 Opc == BPF::Select_64_32 ||
803 Opc == BPF::Select_32 ||
804 Opc == BPF::Select_32_64);
805
806 bool isMemcpyOp = Opc == BPF::MEMCPY;
807
808#ifndef NDEBUG
809 bool isSelectRIOp = (Opc == BPF::Select_Ri ||
810 Opc == BPF::Select_Ri_64_32 ||
811 Opc == BPF::Select_Ri_32 ||
812 Opc == BPF::Select_Ri_32_64);
813
814 if (!(isSelectRROp || isSelectRIOp || isMemcpyOp))
815 report_fatal_error(reason: "unhandled instruction type: " + Twine(Opc));
816#endif
817
818 if (isMemcpyOp)
819 return EmitInstrWithCustomInserterMemcpy(MI, BB);
820
821 bool is32BitCmp = (Opc == BPF::Select_32 ||
822 Opc == BPF::Select_32_64 ||
823 Opc == BPF::Select_Ri_32 ||
824 Opc == BPF::Select_Ri_32_64);
825
826 // To "insert" a SELECT instruction, we actually have to insert the diamond
827 // control-flow pattern. The incoming instruction knows the destination vreg
828 // to set, the condition code register to branch on, the true/false values to
829 // select between, and a branch opcode to use.
830 const BasicBlock *LLVM_BB = BB->getBasicBlock();
831 MachineFunction::iterator I = ++BB->getIterator();
832
833 // ThisMBB:
834 // ...
835 // TrueVal = ...
836 // jmp_XX r1, r2 goto Copy1MBB
837 // fallthrough --> Copy0MBB
838 MachineBasicBlock *ThisMBB = BB;
839 MachineFunction *F = BB->getParent();
840 MachineBasicBlock *Copy0MBB = F->CreateMachineBasicBlock(BB: LLVM_BB);
841 MachineBasicBlock *Copy1MBB = F->CreateMachineBasicBlock(BB: LLVM_BB);
842
843 F->insert(MBBI: I, MBB: Copy0MBB);
844 F->insert(MBBI: I, MBB: Copy1MBB);
845 // Update machine-CFG edges by transferring all successors of the current
846 // block to the new block which will contain the Phi node for the select.
847 Copy1MBB->splice(Where: Copy1MBB->begin(), Other: BB,
848 From: std::next(x: MachineBasicBlock::iterator(MI)), To: BB->end());
849 Copy1MBB->transferSuccessorsAndUpdatePHIs(FromMBB: BB);
850 // Next, add the true and fallthrough blocks as its successors.
851 BB->addSuccessor(Succ: Copy0MBB);
852 BB->addSuccessor(Succ: Copy1MBB);
853
854 // Insert Branch if Flag
855 int CC = MI.getOperand(i: 3).getImm();
856 int NewCC;
857 switch (CC) {
858#define SET_NEWCC(X, Y) \
859 case ISD::X: \
860 if (is32BitCmp && HasJmp32) \
861 NewCC = isSelectRROp ? BPF::Y##_rr_32 : BPF::Y##_ri_32; \
862 else \
863 NewCC = isSelectRROp ? BPF::Y##_rr : BPF::Y##_ri; \
864 break
865 SET_NEWCC(SETGT, JSGT);
866 SET_NEWCC(SETUGT, JUGT);
867 SET_NEWCC(SETGE, JSGE);
868 SET_NEWCC(SETUGE, JUGE);
869 SET_NEWCC(SETEQ, JEQ);
870 SET_NEWCC(SETNE, JNE);
871 SET_NEWCC(SETLT, JSLT);
872 SET_NEWCC(SETULT, JULT);
873 SET_NEWCC(SETLE, JSLE);
874 SET_NEWCC(SETULE, JULE);
875 default:
876 report_fatal_error(reason: "unimplemented select CondCode " + Twine(CC));
877 }
878
879 Register LHS = MI.getOperand(i: 1).getReg();
880 bool isSignedCmp = (CC == ISD::SETGT ||
881 CC == ISD::SETGE ||
882 CC == ISD::SETLT ||
883 CC == ISD::SETLE);
884
885 // eBPF at the moment only has 64-bit comparison. Any 32-bit comparison need
886 // to be promoted, however if the 32-bit comparison operands are destination
887 // registers then they are implicitly zero-extended already, there is no
888 // need of explicit zero-extend sequence for them.
889 //
890 // We simply do extension for all situations in this method, but we will
891 // try to remove those unnecessary in BPFMIPeephole pass.
892 if (is32BitCmp && !HasJmp32)
893 LHS = EmitSubregExt(MI, BB, Reg: LHS, isSigned: isSignedCmp);
894
895 if (isSelectRROp) {
896 Register RHS = MI.getOperand(i: 2).getReg();
897
898 if (is32BitCmp && !HasJmp32)
899 RHS = EmitSubregExt(MI, BB, Reg: RHS, isSigned: isSignedCmp);
900
901 BuildMI(BB, MIMD: DL, MCID: TII.get(Opcode: NewCC)).addReg(RegNo: LHS).addReg(RegNo: RHS).addMBB(MBB: Copy1MBB);
902 } else {
903 int64_t imm32 = MI.getOperand(i: 2).getImm();
904 // Check before we build J*_ri instruction.
905 if (!isInt<32>(x: imm32))
906 report_fatal_error(reason: "immediate overflows 32 bits: " + Twine(imm32));
907 BuildMI(BB, MIMD: DL, MCID: TII.get(Opcode: NewCC))
908 .addReg(RegNo: LHS).addImm(Val: imm32).addMBB(MBB: Copy1MBB);
909 }
910
911 // Copy0MBB:
912 // %FalseValue = ...
913 // # fallthrough to Copy1MBB
914 BB = Copy0MBB;
915
916 // Update machine-CFG edges
917 BB->addSuccessor(Succ: Copy1MBB);
918
919 // Copy1MBB:
920 // %Result = phi [ %FalseValue, Copy0MBB ], [ %TrueValue, ThisMBB ]
921 // ...
922 BB = Copy1MBB;
923 BuildMI(*BB, BB->begin(), DL, TII.get(BPF::PHI), MI.getOperand(0).getReg())
924 .addReg(MI.getOperand(5).getReg())
925 .addMBB(Copy0MBB)
926 .addReg(MI.getOperand(4).getReg())
927 .addMBB(ThisMBB);
928
929 MI.eraseFromParent(); // The pseudo instruction is gone now.
930 return BB;
931}
932
933EVT BPFTargetLowering::getSetCCResultType(const DataLayout &, LLVMContext &,
934 EVT VT) const {
935 return getHasAlu32() ? MVT::i32 : MVT::i64;
936}
937
938MVT BPFTargetLowering::getScalarShiftAmountTy(const DataLayout &DL,
939 EVT VT) const {
940 return (getHasAlu32() && VT == MVT::i32) ? MVT::i32 : MVT::i64;
941}
942
943bool BPFTargetLowering::isLegalAddressingMode(const DataLayout &DL,
944 const AddrMode &AM, Type *Ty,
945 unsigned AS,
946 Instruction *I) const {
947 // No global is ever allowed as a base.
948 if (AM.BaseGV)
949 return false;
950
951 switch (AM.Scale) {
952 case 0: // "r+i" or just "i", depending on HasBaseReg.
953 break;
954 case 1:
955 if (!AM.HasBaseReg) // allow "r+i".
956 break;
957 return false; // disallow "r+r" or "r+r+i".
958 default:
959 return false;
960 }
961
962 return true;
963}
964

source code of llvm/lib/Target/BPF/BPFISelLowering.cpp