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 "llvm/CodeGen/CallingConvLower.h" |
18 | #include "llvm/CodeGen/MachineFrameInfo.h" |
19 | #include "llvm/CodeGen/MachineFunction.h" |
20 | #include "llvm/CodeGen/MachineInstrBuilder.h" |
21 | #include "llvm/CodeGen/MachineRegisterInfo.h" |
22 | #include "llvm/CodeGen/TargetLoweringObjectFileImpl.h" |
23 | #include "llvm/CodeGen/ValueTypes.h" |
24 | #include "llvm/IR/DIBuilder.h" |
25 | #include "llvm/IR/DiagnosticInfo.h" |
26 | #include "llvm/IR/DiagnosticPrinter.h" |
27 | #include "llvm/IR/Module.h" |
28 | #include "llvm/Support/Debug.h" |
29 | #include "llvm/Support/ErrorHandling.h" |
30 | #include "llvm/Support/MathExtras.h" |
31 | #include "llvm/Support/raw_ostream.h" |
32 | |
33 | using namespace llvm; |
34 | |
35 | #define DEBUG_TYPE "bpf-lower" |
36 | |
37 | static cl::opt<bool> BPFExpandMemcpyInOrder("bpf-expand-memcpy-in-order" , |
38 | cl::Hidden, cl::init(Val: false), |
39 | cl::desc("Expand memcpy into load/store pairs in order" )); |
40 | |
41 | static void fail(const SDLoc &DL, SelectionDAG &DAG, const Twine &Msg, |
42 | SDValue Val = {}) { |
43 | std::string Str; |
44 | if (Val) { |
45 | raw_string_ostream OS(Str); |
46 | Val->print(OS); |
47 | OS << ' '; |
48 | } |
49 | MachineFunction &MF = DAG.getMachineFunction(); |
50 | DAG.getContext()->diagnose(DI: DiagnosticInfoUnsupported( |
51 | MF.getFunction(), Twine(Str).concat(Suffix: Msg), DL.getDebugLoc())); |
52 | } |
53 | |
54 | BPFTargetLowering::BPFTargetLowering(const TargetMachine &TM, |
55 | const BPFSubtarget &STI) |
56 | : TargetLowering(TM) { |
57 | |
58 | // Set up the register classes. |
59 | addRegisterClass(VT: MVT::i64, RC: &BPF::GPRRegClass); |
60 | if (STI.getHasAlu32()) |
61 | addRegisterClass(VT: MVT::i32, RC: &BPF::GPR32RegClass); |
62 | |
63 | // Compute derived properties from the register classes |
64 | computeRegisterProperties(TRI: STI.getRegisterInfo()); |
65 | |
66 | setStackPointerRegisterToSaveRestore(BPF::R11); |
67 | |
68 | setOperationAction(Op: ISD::BR_CC, VT: MVT::i64, Action: Custom); |
69 | setOperationAction(Op: ISD::BR_JT, VT: MVT::Other, Action: Expand); |
70 | setOperationAction(Op: ISD::BRIND, VT: MVT::Other, Action: Expand); |
71 | setOperationAction(Op: ISD::BRCOND, VT: MVT::Other, Action: Expand); |
72 | |
73 | setOperationAction(Op: ISD::TRAP, VT: MVT::Other, Action: Custom); |
74 | |
75 | setOperationAction(Ops: {ISD::GlobalAddress, ISD::ConstantPool}, VT: MVT::i64, Action: Custom); |
76 | |
77 | setOperationAction(Op: ISD::DYNAMIC_STACKALLOC, VT: MVT::i64, Action: Custom); |
78 | setOperationAction(Op: ISD::STACKSAVE, VT: MVT::Other, Action: Expand); |
79 | setOperationAction(Op: ISD::STACKRESTORE, VT: MVT::Other, Action: 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(Op: ISD::ATOMIC_LOAD_ADD, VT, Action: Custom); |
90 | } |
91 | |
92 | setOperationAction(Op: ISD::ATOMIC_LOAD_AND, VT, Action: Custom); |
93 | setOperationAction(Op: ISD::ATOMIC_LOAD_OR, VT, Action: Custom); |
94 | setOperationAction(Op: ISD::ATOMIC_LOAD_XOR, VT, Action: Custom); |
95 | setOperationAction(Op: ISD::ATOMIC_SWAP, VT, Action: Custom); |
96 | setOperationAction(Op: ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS, VT, Action: Custom); |
97 | } |
98 | |
99 | for (auto VT : {MVT::i32, MVT::i64}) { |
100 | setOperationAction(Op: ISD::ATOMIC_LOAD, VT, Action: Custom); |
101 | setOperationAction(Op: ISD::ATOMIC_STORE, VT, Action: Custom); |
102 | } |
103 | |
104 | for (auto VT : { MVT::i32, MVT::i64 }) { |
105 | if (VT == MVT::i32 && !STI.getHasAlu32()) |
106 | continue; |
107 | |
108 | setOperationAction(Op: ISD::SDIVREM, VT, Action: Expand); |
109 | setOperationAction(Op: ISD::UDIVREM, VT, Action: Expand); |
110 | if (!STI.hasSdivSmod()) { |
111 | setOperationAction(Op: ISD::SDIV, VT, Action: Custom); |
112 | setOperationAction(Op: ISD::SREM, VT, Action: Custom); |
113 | } |
114 | setOperationAction(Op: ISD::MULHU, VT, Action: Expand); |
115 | setOperationAction(Op: ISD::MULHS, VT, Action: Expand); |
116 | setOperationAction(Op: ISD::UMUL_LOHI, VT, Action: Expand); |
117 | setOperationAction(Op: ISD::SMUL_LOHI, VT, Action: Expand); |
118 | setOperationAction(Op: ISD::ROTR, VT, Action: Expand); |
119 | setOperationAction(Op: ISD::ROTL, VT, Action: Expand); |
120 | setOperationAction(Op: ISD::SHL_PARTS, VT, Action: Expand); |
121 | setOperationAction(Op: ISD::SRL_PARTS, VT, Action: Expand); |
122 | setOperationAction(Op: ISD::SRA_PARTS, VT, Action: Expand); |
123 | setOperationAction(Op: ISD::CTPOP, VT, Action: Expand); |
124 | setOperationAction(Op: ISD::CTTZ, VT, Action: Expand); |
125 | setOperationAction(Op: ISD::CTLZ, VT, Action: Expand); |
126 | setOperationAction(Op: ISD::CTTZ_ZERO_UNDEF, VT, Action: Expand); |
127 | setOperationAction(Op: ISD::CTLZ_ZERO_UNDEF, VT, Action: Expand); |
128 | |
129 | setOperationAction(Op: ISD::SETCC, VT, Action: Expand); |
130 | setOperationAction(Op: ISD::SELECT, VT, Action: Expand); |
131 | setOperationAction(Op: ISD::SELECT_CC, VT, Action: Custom); |
132 | } |
133 | |
134 | if (STI.getHasAlu32()) { |
135 | setOperationAction(Op: ISD::BSWAP, VT: MVT::i32, Action: Promote); |
136 | setOperationAction(Op: ISD::BR_CC, VT: MVT::i32, |
137 | Action: STI.getHasJmp32() ? Custom : Promote); |
138 | } |
139 | |
140 | setOperationAction(Op: ISD::SIGN_EXTEND_INREG, VT: MVT::i1, Action: Expand); |
141 | if (!STI.hasMovsx()) { |
142 | setOperationAction(Op: ISD::SIGN_EXTEND_INREG, VT: MVT::i8, Action: Expand); |
143 | setOperationAction(Op: ISD::SIGN_EXTEND_INREG, VT: MVT::i16, Action: Expand); |
144 | setOperationAction(Op: ISD::SIGN_EXTEND_INREG, VT: MVT::i32, Action: Expand); |
145 | } |
146 | |
147 | // Extended load operations for i1 types must be promoted |
148 | for (MVT VT : MVT::integer_valuetypes()) { |
149 | setLoadExtAction(ExtType: ISD::EXTLOAD, ValVT: VT, MemVT: MVT::i1, Action: Promote); |
150 | setLoadExtAction(ExtType: ISD::ZEXTLOAD, ValVT: VT, MemVT: MVT::i1, Action: Promote); |
151 | setLoadExtAction(ExtType: ISD::SEXTLOAD, ValVT: VT, MemVT: MVT::i1, Action: Promote); |
152 | |
153 | if (!STI.hasLdsx()) { |
154 | setLoadExtAction(ExtType: ISD::SEXTLOAD, ValVT: VT, MemVT: MVT::i8, Action: Expand); |
155 | setLoadExtAction(ExtType: ISD::SEXTLOAD, ValVT: VT, MemVT: MVT::i16, Action: Expand); |
156 | setLoadExtAction(ExtType: ISD::SEXTLOAD, ValVT: VT, MemVT: MVT::i32, Action: Expand); |
157 | } |
158 | } |
159 | |
160 | setBooleanContents(ZeroOrOneBooleanContent); |
161 | setMaxAtomicSizeInBitsSupported(64); |
162 | |
163 | // Function alignments |
164 | setMinFunctionAlignment(Align(8)); |
165 | setPrefFunctionAlignment(Align(8)); |
166 | |
167 | if (BPFExpandMemcpyInOrder) { |
168 | // LLVM generic code will try to expand memcpy into load/store pairs at this |
169 | // stage which is before quite a few IR optimization passes, therefore the |
170 | // loads and stores could potentially be moved apart from each other which |
171 | // will cause trouble to memcpy pattern matcher inside kernel eBPF JIT |
172 | // compilers. |
173 | // |
174 | // When -bpf-expand-memcpy-in-order specified, we want to defer the expand |
175 | // of memcpy to later stage in IR optimization pipeline so those load/store |
176 | // pairs won't be touched and could be kept in order. Hence, we set |
177 | // MaxStoresPerMem* to zero to disable the generic getMemcpyLoadsAndStores |
178 | // code path, and ask LLVM to use target expander EmitTargetCodeForMemcpy. |
179 | MaxStoresPerMemset = MaxStoresPerMemsetOptSize = 0; |
180 | MaxStoresPerMemcpy = MaxStoresPerMemcpyOptSize = 0; |
181 | MaxStoresPerMemmove = MaxStoresPerMemmoveOptSize = 0; |
182 | MaxLoadsPerMemcmp = 0; |
183 | } else { |
184 | // inline memcpy() for kernel to see explicit copy |
185 | unsigned CommonMaxStores = |
186 | STI.getSelectionDAGInfo()->getCommonMaxStoresPerMemFunc(); |
187 | |
188 | MaxStoresPerMemset = MaxStoresPerMemsetOptSize = CommonMaxStores; |
189 | MaxStoresPerMemcpy = MaxStoresPerMemcpyOptSize = CommonMaxStores; |
190 | MaxStoresPerMemmove = MaxStoresPerMemmoveOptSize = CommonMaxStores; |
191 | MaxLoadsPerMemcmp = MaxLoadsPerMemcmpOptSize = CommonMaxStores; |
192 | } |
193 | |
194 | // CPU/Feature control |
195 | HasAlu32 = STI.getHasAlu32(); |
196 | HasJmp32 = STI.getHasJmp32(); |
197 | HasJmpExt = STI.getHasJmpExt(); |
198 | HasMovsx = STI.hasMovsx(); |
199 | } |
200 | |
201 | bool BPFTargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const { |
202 | return false; |
203 | } |
204 | |
205 | bool BPFTargetLowering::isTruncateFree(Type *Ty1, Type *Ty2) const { |
206 | if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy()) |
207 | return false; |
208 | unsigned NumBits1 = Ty1->getPrimitiveSizeInBits(); |
209 | unsigned NumBits2 = Ty2->getPrimitiveSizeInBits(); |
210 | return NumBits1 > NumBits2; |
211 | } |
212 | |
213 | bool BPFTargetLowering::isTruncateFree(EVT VT1, EVT VT2) const { |
214 | if (!VT1.isInteger() || !VT2.isInteger()) |
215 | return false; |
216 | unsigned NumBits1 = VT1.getSizeInBits(); |
217 | unsigned NumBits2 = VT2.getSizeInBits(); |
218 | return NumBits1 > NumBits2; |
219 | } |
220 | |
221 | bool BPFTargetLowering::isZExtFree(Type *Ty1, Type *Ty2) const { |
222 | if (!getHasAlu32() || !Ty1->isIntegerTy() || !Ty2->isIntegerTy()) |
223 | return false; |
224 | unsigned NumBits1 = Ty1->getPrimitiveSizeInBits(); |
225 | unsigned NumBits2 = Ty2->getPrimitiveSizeInBits(); |
226 | return NumBits1 == 32 && NumBits2 == 64; |
227 | } |
228 | |
229 | bool BPFTargetLowering::isZExtFree(EVT VT1, EVT VT2) const { |
230 | if (!getHasAlu32() || !VT1.isInteger() || !VT2.isInteger()) |
231 | return false; |
232 | unsigned NumBits1 = VT1.getSizeInBits(); |
233 | unsigned NumBits2 = VT2.getSizeInBits(); |
234 | return NumBits1 == 32 && NumBits2 == 64; |
235 | } |
236 | |
237 | bool BPFTargetLowering::isZExtFree(SDValue Val, EVT VT2) const { |
238 | EVT VT1 = Val.getValueType(); |
239 | if (Val.getOpcode() == ISD::LOAD && VT1.isSimple() && VT2.isSimple()) { |
240 | MVT MT1 = VT1.getSimpleVT().SimpleTy; |
241 | MVT MT2 = VT2.getSimpleVT().SimpleTy; |
242 | if ((MT1 == MVT::i8 || MT1 == MVT::i16 || MT1 == MVT::i32) && |
243 | (MT2 == MVT::i32 || MT2 == MVT::i64)) |
244 | return true; |
245 | } |
246 | return TargetLoweringBase::isZExtFree(Val, VT2); |
247 | } |
248 | |
249 | BPFTargetLowering::ConstraintType |
250 | BPFTargetLowering::getConstraintType(StringRef Constraint) const { |
251 | if (Constraint.size() == 1) { |
252 | switch (Constraint[0]) { |
253 | default: |
254 | break; |
255 | case 'w': |
256 | return C_RegisterClass; |
257 | } |
258 | } |
259 | |
260 | return TargetLowering::getConstraintType(Constraint); |
261 | } |
262 | |
263 | std::pair<unsigned, const TargetRegisterClass *> |
264 | BPFTargetLowering::getRegForInlineAsmConstraint(const TargetRegisterInfo *TRI, |
265 | StringRef Constraint, |
266 | MVT VT) const { |
267 | if (Constraint.size() == 1) { |
268 | // GCC Constraint Letters |
269 | switch (Constraint[0]) { |
270 | case 'r': // GENERAL_REGS |
271 | return std::make_pair(x: 0U, y: &BPF::GPRRegClass); |
272 | case 'w': |
273 | if (HasAlu32) |
274 | return std::make_pair(x: 0U, y: &BPF::GPR32RegClass); |
275 | break; |
276 | default: |
277 | break; |
278 | } |
279 | } |
280 | |
281 | return TargetLowering::getRegForInlineAsmConstraint(TRI, Constraint, VT); |
282 | } |
283 | |
284 | void BPFTargetLowering::ReplaceNodeResults( |
285 | SDNode *N, SmallVectorImpl<SDValue> &Results, SelectionDAG &DAG) const { |
286 | const char *Msg; |
287 | uint32_t Opcode = N->getOpcode(); |
288 | switch (Opcode) { |
289 | default: |
290 | report_fatal_error(reason: "unhandled custom legalization: " + Twine(Opcode)); |
291 | case ISD::ATOMIC_LOAD_ADD: |
292 | case ISD::ATOMIC_LOAD_AND: |
293 | case ISD::ATOMIC_LOAD_OR: |
294 | case ISD::ATOMIC_LOAD_XOR: |
295 | case ISD::ATOMIC_SWAP: |
296 | case ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS: |
297 | if (HasAlu32 || Opcode == ISD::ATOMIC_LOAD_ADD) |
298 | Msg = "unsupported atomic operation, please use 32/64 bit version" ; |
299 | else |
300 | Msg = "unsupported atomic operation, please use 64 bit version" ; |
301 | break; |
302 | case ISD::ATOMIC_LOAD: |
303 | case ISD::ATOMIC_STORE: |
304 | return; |
305 | } |
306 | |
307 | SDLoc DL(N); |
308 | // We'll still produce a fatal error downstream, but this diagnostic is more |
309 | // user-friendly. |
310 | fail(DL, DAG, Msg); |
311 | } |
312 | |
313 | SDValue BPFTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const { |
314 | switch (Op.getOpcode()) { |
315 | default: |
316 | report_fatal_error(reason: "unimplemented opcode: " + Twine(Op.getOpcode())); |
317 | case ISD::BR_CC: |
318 | return LowerBR_CC(Op, DAG); |
319 | case ISD::GlobalAddress: |
320 | return LowerGlobalAddress(Op, DAG); |
321 | case ISD::ConstantPool: |
322 | return LowerConstantPool(Op, DAG); |
323 | case ISD::SELECT_CC: |
324 | return LowerSELECT_CC(Op, DAG); |
325 | case ISD::SDIV: |
326 | case ISD::SREM: |
327 | return LowerSDIVSREM(Op, DAG); |
328 | case ISD::DYNAMIC_STACKALLOC: |
329 | return LowerDYNAMIC_STACKALLOC(Op, DAG); |
330 | case ISD::ATOMIC_LOAD: |
331 | case ISD::ATOMIC_STORE: |
332 | return LowerATOMIC_LOAD_STORE(Op, DAG); |
333 | case ISD::TRAP: |
334 | return LowerTRAP(Op, DAG); |
335 | } |
336 | } |
337 | |
338 | // Calling Convention Implementation |
339 | #include "BPFGenCallingConv.inc" |
340 | |
341 | SDValue BPFTargetLowering::LowerFormalArguments( |
342 | SDValue Chain, CallingConv::ID CallConv, bool IsVarArg, |
343 | const SmallVectorImpl<ISD::InputArg> &Ins, const SDLoc &DL, |
344 | SelectionDAG &DAG, SmallVectorImpl<SDValue> &InVals) const { |
345 | switch (CallConv) { |
346 | default: |
347 | report_fatal_error(reason: "unimplemented calling convention: " + Twine(CallConv)); |
348 | case CallingConv::C: |
349 | case CallingConv::Fast: |
350 | break; |
351 | } |
352 | |
353 | MachineFunction &MF = DAG.getMachineFunction(); |
354 | MachineRegisterInfo &RegInfo = MF.getRegInfo(); |
355 | |
356 | // Assign locations to all of the incoming arguments. |
357 | SmallVector<CCValAssign, 16> ArgLocs; |
358 | CCState CCInfo(CallConv, IsVarArg, MF, ArgLocs, *DAG.getContext()); |
359 | CCInfo.AnalyzeFormalArguments(Ins, Fn: getHasAlu32() ? CC_BPF32 : CC_BPF64); |
360 | |
361 | bool HasMemArgs = false; |
362 | for (size_t I = 0; I < ArgLocs.size(); ++I) { |
363 | auto &VA = ArgLocs[I]; |
364 | |
365 | if (VA.isRegLoc()) { |
366 | // Arguments passed in registers |
367 | EVT RegVT = VA.getLocVT(); |
368 | MVT::SimpleValueType SimpleTy = RegVT.getSimpleVT().SimpleTy; |
369 | switch (SimpleTy) { |
370 | default: { |
371 | std::string Str; |
372 | { |
373 | raw_string_ostream OS(Str); |
374 | RegVT.print(OS); |
375 | } |
376 | report_fatal_error(reason: "unhandled argument type: " + Twine(Str)); |
377 | } |
378 | case MVT::i32: |
379 | case MVT::i64: |
380 | Register VReg = RegInfo.createVirtualRegister( |
381 | RegClass: SimpleTy == MVT::i64 ? &BPF::GPRRegClass : &BPF::GPR32RegClass); |
382 | RegInfo.addLiveIn(Reg: VA.getLocReg(), vreg: VReg); |
383 | SDValue ArgValue = DAG.getCopyFromReg(Chain, dl: DL, Reg: VReg, VT: RegVT); |
384 | |
385 | // If this is an value that has been promoted to wider types, insert an |
386 | // assert[sz]ext to capture this, then truncate to the right size. |
387 | if (VA.getLocInfo() == CCValAssign::SExt) |
388 | ArgValue = DAG.getNode(Opcode: ISD::AssertSext, DL, VT: RegVT, N1: ArgValue, |
389 | N2: DAG.getValueType(VA.getValVT())); |
390 | else if (VA.getLocInfo() == CCValAssign::ZExt) |
391 | ArgValue = DAG.getNode(Opcode: ISD::AssertZext, DL, VT: RegVT, N1: ArgValue, |
392 | N2: DAG.getValueType(VA.getValVT())); |
393 | |
394 | if (VA.getLocInfo() != CCValAssign::Full) |
395 | ArgValue = DAG.getNode(Opcode: ISD::TRUNCATE, DL, VT: VA.getValVT(), Operand: ArgValue); |
396 | |
397 | InVals.push_back(Elt: ArgValue); |
398 | |
399 | break; |
400 | } |
401 | } else { |
402 | if (VA.isMemLoc()) |
403 | HasMemArgs = true; |
404 | else |
405 | report_fatal_error(reason: "unhandled argument location" ); |
406 | InVals.push_back(Elt: DAG.getConstant(Val: 0, DL, VT: VA.getLocVT())); |
407 | } |
408 | } |
409 | if (HasMemArgs) |
410 | fail(DL, DAG, Msg: "stack arguments are not supported" ); |
411 | if (IsVarArg) |
412 | fail(DL, DAG, Msg: "variadic functions are not supported" ); |
413 | if (MF.getFunction().hasStructRetAttr()) |
414 | fail(DL, DAG, Msg: "aggregate returns are not supported" ); |
415 | |
416 | return Chain; |
417 | } |
418 | |
419 | const size_t BPFTargetLowering::MaxArgs = 5; |
420 | |
421 | static void resetRegMaskBit(const TargetRegisterInfo *TRI, uint32_t *RegMask, |
422 | MCRegister Reg) { |
423 | for (MCPhysReg SubReg : TRI->subregs_inclusive(Reg)) |
424 | RegMask[SubReg / 32] &= ~(1u << (SubReg % 32)); |
425 | } |
426 | |
427 | static uint32_t *regMaskFromTemplate(const TargetRegisterInfo *TRI, |
428 | MachineFunction &MF, |
429 | const uint32_t *BaseRegMask) { |
430 | uint32_t *RegMask = MF.allocateRegMask(); |
431 | unsigned RegMaskSize = MachineOperand::getRegMaskSize(NumRegs: TRI->getNumRegs()); |
432 | memcpy(dest: RegMask, src: BaseRegMask, n: sizeof(RegMask[0]) * RegMaskSize); |
433 | return RegMask; |
434 | } |
435 | |
436 | SDValue BPFTargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI, |
437 | SmallVectorImpl<SDValue> &InVals) const { |
438 | SelectionDAG &DAG = CLI.DAG; |
439 | auto &Outs = CLI.Outs; |
440 | auto &OutVals = CLI.OutVals; |
441 | auto &Ins = CLI.Ins; |
442 | SDValue Chain = CLI.Chain; |
443 | SDValue Callee = CLI.Callee; |
444 | bool &IsTailCall = CLI.IsTailCall; |
445 | CallingConv::ID CallConv = CLI.CallConv; |
446 | bool IsVarArg = CLI.IsVarArg; |
447 | MachineFunction &MF = DAG.getMachineFunction(); |
448 | |
449 | // BPF target does not support tail call optimization. |
450 | IsTailCall = false; |
451 | |
452 | switch (CallConv) { |
453 | default: |
454 | report_fatal_error(reason: "unsupported calling convention: " + Twine(CallConv)); |
455 | case CallingConv::Fast: |
456 | case CallingConv::C: |
457 | break; |
458 | } |
459 | |
460 | // Analyze operands of the call, assigning locations to each operand. |
461 | SmallVector<CCValAssign, 16> ArgLocs; |
462 | CCState CCInfo(CallConv, IsVarArg, MF, ArgLocs, *DAG.getContext()); |
463 | |
464 | CCInfo.AnalyzeCallOperands(Outs, Fn: getHasAlu32() ? CC_BPF32 : CC_BPF64); |
465 | |
466 | unsigned NumBytes = CCInfo.getStackSize(); |
467 | |
468 | if (Outs.size() > MaxArgs) |
469 | fail(DL: CLI.DL, DAG, Msg: "too many arguments" , Val: Callee); |
470 | |
471 | for (auto &Arg : Outs) { |
472 | ISD::ArgFlagsTy Flags = Arg.Flags; |
473 | if (!Flags.isByVal()) |
474 | continue; |
475 | fail(DL: CLI.DL, DAG, Msg: "pass by value not supported" , Val: Callee); |
476 | break; |
477 | } |
478 | |
479 | auto PtrVT = getPointerTy(DL: MF.getDataLayout()); |
480 | Chain = DAG.getCALLSEQ_START(Chain, InSize: NumBytes, OutSize: 0, DL: CLI.DL); |
481 | |
482 | SmallVector<std::pair<unsigned, SDValue>, MaxArgs> RegsToPass; |
483 | |
484 | // Walk arg assignments |
485 | for (size_t i = 0; i < std::min(a: ArgLocs.size(), b: MaxArgs); ++i) { |
486 | CCValAssign &VA = ArgLocs[i]; |
487 | SDValue &Arg = OutVals[i]; |
488 | |
489 | // Promote the value if needed. |
490 | switch (VA.getLocInfo()) { |
491 | default: |
492 | report_fatal_error(reason: "unhandled location info: " + Twine(VA.getLocInfo())); |
493 | case CCValAssign::Full: |
494 | break; |
495 | case CCValAssign::SExt: |
496 | Arg = DAG.getNode(Opcode: ISD::SIGN_EXTEND, DL: CLI.DL, VT: VA.getLocVT(), Operand: Arg); |
497 | break; |
498 | case CCValAssign::ZExt: |
499 | Arg = DAG.getNode(Opcode: ISD::ZERO_EXTEND, DL: CLI.DL, VT: VA.getLocVT(), Operand: Arg); |
500 | break; |
501 | case CCValAssign::AExt: |
502 | Arg = DAG.getNode(Opcode: ISD::ANY_EXTEND, DL: CLI.DL, VT: VA.getLocVT(), Operand: Arg); |
503 | break; |
504 | } |
505 | |
506 | // Push arguments into RegsToPass vector |
507 | if (VA.isRegLoc()) |
508 | RegsToPass.push_back(Elt: std::make_pair(x: VA.getLocReg(), y&: Arg)); |
509 | else |
510 | report_fatal_error(reason: "stack arguments are not supported" ); |
511 | } |
512 | |
513 | SDValue InGlue; |
514 | |
515 | // Build a sequence of copy-to-reg nodes chained together with token chain and |
516 | // flag operands which copy the outgoing args into registers. The InGlue in |
517 | // necessary since all emitted instructions must be stuck together. |
518 | for (auto &Reg : RegsToPass) { |
519 | Chain = DAG.getCopyToReg(Chain, dl: CLI.DL, Reg: Reg.first, N: Reg.second, Glue: InGlue); |
520 | InGlue = Chain.getValue(R: 1); |
521 | } |
522 | |
523 | // If the callee is a GlobalAddress node (quite common, every direct call is) |
524 | // turn it into a TargetGlobalAddress node so that legalize doesn't hack it. |
525 | // Likewise ExternalSymbol -> TargetExternalSymbol. |
526 | if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Val&: Callee)) { |
527 | Callee = DAG.getTargetGlobalAddress(GV: G->getGlobal(), DL: CLI.DL, VT: PtrVT, |
528 | offset: G->getOffset(), TargetFlags: 0); |
529 | } else if (ExternalSymbolSDNode *E = dyn_cast<ExternalSymbolSDNode>(Val&: Callee)) { |
530 | if (StringRef(E->getSymbol()) != BPF_TRAP) { |
531 | Callee = DAG.getTargetExternalSymbol(Sym: E->getSymbol(), VT: PtrVT, TargetFlags: 0); |
532 | fail(DL: CLI.DL, DAG, |
533 | Msg: Twine("A call to built-in function '" + StringRef(E->getSymbol()) + |
534 | "' is not supported." )); |
535 | } |
536 | } |
537 | |
538 | // Returns a chain & a flag for retval copy to use. |
539 | SDVTList NodeTys = DAG.getVTList(VT1: MVT::Other, VT2: MVT::Glue); |
540 | SmallVector<SDValue, 8> Ops; |
541 | Ops.push_back(Elt: Chain); |
542 | Ops.push_back(Elt: Callee); |
543 | |
544 | // Add argument registers to the end of the list so that they are |
545 | // known live into the call. |
546 | for (auto &Reg : RegsToPass) |
547 | Ops.push_back(Elt: DAG.getRegister(Reg: Reg.first, VT: Reg.second.getValueType())); |
548 | |
549 | bool HasFastCall = |
550 | (CLI.CB && isa<CallInst>(Val: CLI.CB) && CLI.CB->hasFnAttr(Kind: "bpf_fastcall" )); |
551 | const TargetRegisterInfo *TRI = MF.getSubtarget().getRegisterInfo(); |
552 | if (HasFastCall) { |
553 | uint32_t *RegMask = regMaskFromTemplate( |
554 | TRI, MF, BaseRegMask: TRI->getCallPreservedMask(MF, CallingConv::PreserveAll)); |
555 | for (auto const &RegPair : RegsToPass) |
556 | resetRegMaskBit(TRI, RegMask, Reg: RegPair.first); |
557 | if (!CLI.CB->getType()->isVoidTy()) |
558 | resetRegMaskBit(TRI, RegMask, Reg: BPF::R0); |
559 | Ops.push_back(Elt: DAG.getRegisterMask(RegMask)); |
560 | } else { |
561 | Ops.push_back( |
562 | Elt: DAG.getRegisterMask(RegMask: TRI->getCallPreservedMask(MF, CLI.CallConv))); |
563 | } |
564 | |
565 | if (InGlue.getNode()) |
566 | Ops.push_back(Elt: InGlue); |
567 | |
568 | Chain = DAG.getNode(Opcode: BPFISD::CALL, DL: CLI.DL, VTList: NodeTys, Ops); |
569 | InGlue = Chain.getValue(R: 1); |
570 | |
571 | DAG.addNoMergeSiteInfo(Node: Chain.getNode(), NoMerge: CLI.NoMerge); |
572 | |
573 | // Create the CALLSEQ_END node. |
574 | Chain = DAG.getCALLSEQ_END(Chain, Size1: NumBytes, Size2: 0, Glue: InGlue, DL: CLI.DL); |
575 | InGlue = Chain.getValue(R: 1); |
576 | |
577 | // Handle result values, copying them out of physregs into vregs that we |
578 | // return. |
579 | return LowerCallResult(Chain, InGlue, CallConv, IsVarArg, Ins, DL: CLI.DL, DAG, |
580 | InVals); |
581 | } |
582 | |
583 | SDValue |
584 | BPFTargetLowering::LowerReturn(SDValue Chain, CallingConv::ID CallConv, |
585 | bool IsVarArg, |
586 | const SmallVectorImpl<ISD::OutputArg> &Outs, |
587 | const SmallVectorImpl<SDValue> &OutVals, |
588 | const SDLoc &DL, SelectionDAG &DAG) const { |
589 | unsigned Opc = BPFISD::RET_GLUE; |
590 | |
591 | // CCValAssign - represent the assignment of the return value to a location |
592 | SmallVector<CCValAssign, 16> RVLocs; |
593 | MachineFunction &MF = DAG.getMachineFunction(); |
594 | |
595 | // CCState - Info about the registers and stack slot. |
596 | CCState CCInfo(CallConv, IsVarArg, MF, RVLocs, *DAG.getContext()); |
597 | |
598 | if (MF.getFunction().getReturnType()->isAggregateType()) { |
599 | fail(DL, DAG, Msg: "aggregate returns are not supported" ); |
600 | return DAG.getNode(Opcode: Opc, DL, VT: MVT::Other, Operand: Chain); |
601 | } |
602 | |
603 | // Analize return values. |
604 | CCInfo.AnalyzeReturn(Outs, Fn: getHasAlu32() ? RetCC_BPF32 : RetCC_BPF64); |
605 | |
606 | SDValue Glue; |
607 | SmallVector<SDValue, 4> RetOps(1, Chain); |
608 | |
609 | // Copy the result values into the output registers. |
610 | for (size_t i = 0; i != RVLocs.size(); ++i) { |
611 | CCValAssign &VA = RVLocs[i]; |
612 | if (!VA.isRegLoc()) |
613 | report_fatal_error(reason: "stack return values are not supported" ); |
614 | |
615 | Chain = DAG.getCopyToReg(Chain, dl: DL, Reg: VA.getLocReg(), N: OutVals[i], Glue); |
616 | |
617 | // Guarantee that all emitted copies are stuck together, |
618 | // avoiding something bad. |
619 | Glue = Chain.getValue(R: 1); |
620 | RetOps.push_back(Elt: DAG.getRegister(Reg: VA.getLocReg(), VT: VA.getLocVT())); |
621 | } |
622 | |
623 | RetOps[0] = Chain; // Update chain. |
624 | |
625 | // Add the glue if we have it. |
626 | if (Glue.getNode()) |
627 | RetOps.push_back(Elt: Glue); |
628 | |
629 | return DAG.getNode(Opcode: Opc, DL, VT: MVT::Other, Ops: RetOps); |
630 | } |
631 | |
632 | SDValue BPFTargetLowering::LowerCallResult( |
633 | SDValue Chain, SDValue InGlue, CallingConv::ID CallConv, bool IsVarArg, |
634 | const SmallVectorImpl<ISD::InputArg> &Ins, const SDLoc &DL, |
635 | SelectionDAG &DAG, SmallVectorImpl<SDValue> &InVals) const { |
636 | |
637 | MachineFunction &MF = DAG.getMachineFunction(); |
638 | // Assign locations to each value returned by this call. |
639 | SmallVector<CCValAssign, 16> RVLocs; |
640 | CCState CCInfo(CallConv, IsVarArg, MF, RVLocs, *DAG.getContext()); |
641 | |
642 | if (Ins.size() > 1) { |
643 | fail(DL, DAG, Msg: "only small returns supported" ); |
644 | for (auto &In : Ins) |
645 | InVals.push_back(Elt: DAG.getConstant(Val: 0, DL, VT: In.VT)); |
646 | return DAG.getCopyFromReg(Chain, dl: DL, Reg: 1, VT: Ins[0].VT, Glue: InGlue).getValue(R: 1); |
647 | } |
648 | |
649 | CCInfo.AnalyzeCallResult(Ins, Fn: getHasAlu32() ? RetCC_BPF32 : RetCC_BPF64); |
650 | |
651 | // Copy all of the result registers out of their specified physreg. |
652 | for (auto &Val : RVLocs) { |
653 | Chain = DAG.getCopyFromReg(Chain, dl: DL, Reg: Val.getLocReg(), |
654 | VT: Val.getValVT(), Glue: InGlue).getValue(R: 1); |
655 | InGlue = Chain.getValue(R: 2); |
656 | InVals.push_back(Elt: Chain.getValue(R: 0)); |
657 | } |
658 | |
659 | return Chain; |
660 | } |
661 | |
662 | static void NegateCC(SDValue &LHS, SDValue &RHS, ISD::CondCode &CC) { |
663 | switch (CC) { |
664 | default: |
665 | break; |
666 | case ISD::SETULT: |
667 | case ISD::SETULE: |
668 | case ISD::SETLT: |
669 | case ISD::SETLE: |
670 | CC = ISD::getSetCCSwappedOperands(Operation: CC); |
671 | std::swap(a&: LHS, b&: RHS); |
672 | break; |
673 | } |
674 | } |
675 | |
676 | SDValue BPFTargetLowering::LowerSDIVSREM(SDValue Op, SelectionDAG &DAG) const { |
677 | SDLoc DL(Op); |
678 | fail(DL, DAG, |
679 | Msg: "unsupported signed division, please convert to unsigned div/mod." ); |
680 | return DAG.getUNDEF(VT: Op->getValueType(ResNo: 0)); |
681 | } |
682 | |
683 | SDValue BPFTargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op, |
684 | SelectionDAG &DAG) const { |
685 | SDLoc DL(Op); |
686 | fail(DL, DAG, Msg: "unsupported dynamic stack allocation" ); |
687 | auto Ops = {DAG.getConstant(Val: 0, DL: SDLoc(), VT: Op.getValueType()), Op.getOperand(i: 0)}; |
688 | return DAG.getMergeValues(Ops, dl: SDLoc()); |
689 | } |
690 | |
691 | SDValue BPFTargetLowering::LowerBR_CC(SDValue Op, SelectionDAG &DAG) const { |
692 | SDValue Chain = Op.getOperand(i: 0); |
693 | ISD::CondCode CC = cast<CondCodeSDNode>(Val: Op.getOperand(i: 1))->get(); |
694 | SDValue LHS = Op.getOperand(i: 2); |
695 | SDValue RHS = Op.getOperand(i: 3); |
696 | SDValue Dest = Op.getOperand(i: 4); |
697 | SDLoc DL(Op); |
698 | |
699 | if (!getHasJmpExt()) |
700 | NegateCC(LHS, RHS, CC); |
701 | |
702 | return DAG.getNode(Opcode: BPFISD::BR_CC, DL, VT: Op.getValueType(), N1: Chain, N2: LHS, N3: RHS, |
703 | N4: DAG.getConstant(Val: CC, DL, VT: LHS.getValueType()), N5: Dest); |
704 | } |
705 | |
706 | SDValue BPFTargetLowering::LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const { |
707 | SDValue LHS = Op.getOperand(i: 0); |
708 | SDValue RHS = Op.getOperand(i: 1); |
709 | SDValue TrueV = Op.getOperand(i: 2); |
710 | SDValue FalseV = Op.getOperand(i: 3); |
711 | ISD::CondCode CC = cast<CondCodeSDNode>(Val: Op.getOperand(i: 4))->get(); |
712 | SDLoc DL(Op); |
713 | |
714 | if (!getHasJmpExt()) |
715 | NegateCC(LHS, RHS, CC); |
716 | |
717 | SDValue TargetCC = DAG.getConstant(Val: CC, DL, VT: LHS.getValueType()); |
718 | SDValue Ops[] = {LHS, RHS, TargetCC, TrueV, FalseV}; |
719 | |
720 | return DAG.getNode(Opcode: BPFISD::SELECT_CC, DL, VT: Op.getValueType(), Ops); |
721 | } |
722 | |
723 | SDValue BPFTargetLowering::LowerATOMIC_LOAD_STORE(SDValue Op, |
724 | SelectionDAG &DAG) const { |
725 | SDNode *N = Op.getNode(); |
726 | SDLoc DL(N); |
727 | |
728 | if (cast<AtomicSDNode>(Val: N)->getMergedOrdering() == |
729 | AtomicOrdering::SequentiallyConsistent) |
730 | fail(DL, DAG, |
731 | Msg: "sequentially consistent (seq_cst) " |
732 | "atomic load/store is not supported" ); |
733 | |
734 | return Op; |
735 | } |
736 | |
737 | static Function *createBPFUnreachable(Module *M) { |
738 | if (auto *Fn = M->getFunction(Name: BPF_TRAP)) |
739 | return Fn; |
740 | |
741 | FunctionType *FT = FunctionType::get(Result: Type::getVoidTy(C&: M->getContext()), isVarArg: false); |
742 | Function *NewF = |
743 | Function::Create(Ty: FT, Linkage: GlobalValue::ExternalWeakLinkage, N: BPF_TRAP, M); |
744 | NewF->setDSOLocal(true); |
745 | NewF->setCallingConv(CallingConv::C); |
746 | NewF->setSection(".ksyms" ); |
747 | |
748 | if (M->debug_compile_units().empty()) |
749 | return NewF; |
750 | |
751 | DIBuilder DBuilder(*M); |
752 | DITypeRefArray ParamTypes = |
753 | DBuilder.getOrCreateTypeArray(Elements: {nullptr /*void return*/}); |
754 | DISubroutineType *FuncType = DBuilder.createSubroutineType(ParameterTypes: ParamTypes); |
755 | DICompileUnit *CU = *M->debug_compile_units_begin(); |
756 | DISubprogram *SP = |
757 | DBuilder.createFunction(Scope: CU, Name: BPF_TRAP, LinkageName: BPF_TRAP, File: nullptr, LineNo: 0, Ty: FuncType, ScopeLine: 0, |
758 | Flags: DINode::FlagZero, SPFlags: DISubprogram::SPFlagZero); |
759 | NewF->setSubprogram(SP); |
760 | return NewF; |
761 | } |
762 | |
763 | SDValue BPFTargetLowering::LowerTRAP(SDValue Op, SelectionDAG &DAG) const { |
764 | MachineFunction &MF = DAG.getMachineFunction(); |
765 | TargetLowering::CallLoweringInfo CLI(DAG); |
766 | SmallVector<SDValue> InVals; |
767 | SDNode *N = Op.getNode(); |
768 | SDLoc DL(N); |
769 | |
770 | Function *Fn = createBPFUnreachable(M: MF.getFunction().getParent()); |
771 | auto PtrVT = getPointerTy(DL: MF.getDataLayout()); |
772 | CLI.Callee = DAG.getTargetGlobalAddress(GV: Fn, DL, VT: PtrVT); |
773 | CLI.Chain = N->getOperand(Num: 0); |
774 | CLI.IsTailCall = false; |
775 | CLI.CallConv = CallingConv::C; |
776 | CLI.IsVarArg = false; |
777 | CLI.DL = DL; |
778 | CLI.NoMerge = false; |
779 | CLI.DoesNotReturn = true; |
780 | return LowerCall(CLI, InVals); |
781 | } |
782 | |
783 | const char *BPFTargetLowering::getTargetNodeName(unsigned Opcode) const { |
784 | switch ((BPFISD::NodeType)Opcode) { |
785 | case BPFISD::FIRST_NUMBER: |
786 | break; |
787 | case BPFISD::RET_GLUE: |
788 | return "BPFISD::RET_GLUE" ; |
789 | case BPFISD::CALL: |
790 | return "BPFISD::CALL" ; |
791 | case BPFISD::SELECT_CC: |
792 | return "BPFISD::SELECT_CC" ; |
793 | case BPFISD::BR_CC: |
794 | return "BPFISD::BR_CC" ; |
795 | case BPFISD::Wrapper: |
796 | return "BPFISD::Wrapper" ; |
797 | case BPFISD::MEMCPY: |
798 | return "BPFISD::MEMCPY" ; |
799 | } |
800 | return nullptr; |
801 | } |
802 | |
803 | static SDValue getTargetNode(GlobalAddressSDNode *N, const SDLoc &DL, EVT Ty, |
804 | SelectionDAG &DAG, unsigned Flags) { |
805 | return DAG.getTargetGlobalAddress(GV: N->getGlobal(), DL, VT: Ty, offset: 0, TargetFlags: Flags); |
806 | } |
807 | |
808 | static SDValue getTargetNode(ConstantPoolSDNode *N, const SDLoc &DL, EVT Ty, |
809 | SelectionDAG &DAG, unsigned Flags) { |
810 | return DAG.getTargetConstantPool(C: N->getConstVal(), VT: Ty, Align: N->getAlign(), |
811 | Offset: N->getOffset(), TargetFlags: Flags); |
812 | } |
813 | |
814 | template <class NodeTy> |
815 | SDValue BPFTargetLowering::getAddr(NodeTy *N, SelectionDAG &DAG, |
816 | unsigned Flags) const { |
817 | SDLoc DL(N); |
818 | |
819 | SDValue GA = getTargetNode(N, DL, MVT::i64, DAG, Flags); |
820 | |
821 | return DAG.getNode(Opcode: BPFISD::Wrapper, DL, VT: MVT::i64, Operand: GA); |
822 | } |
823 | |
824 | SDValue BPFTargetLowering::LowerGlobalAddress(SDValue Op, |
825 | SelectionDAG &DAG) const { |
826 | GlobalAddressSDNode *N = cast<GlobalAddressSDNode>(Val&: Op); |
827 | if (N->getOffset() != 0) |
828 | report_fatal_error(reason: "invalid offset for global address: " + |
829 | Twine(N->getOffset())); |
830 | return getAddr(N, DAG); |
831 | } |
832 | |
833 | SDValue BPFTargetLowering::LowerConstantPool(SDValue Op, |
834 | SelectionDAG &DAG) const { |
835 | ConstantPoolSDNode *N = cast<ConstantPoolSDNode>(Val&: Op); |
836 | |
837 | return getAddr(N, DAG); |
838 | } |
839 | |
840 | unsigned |
841 | BPFTargetLowering::EmitSubregExt(MachineInstr &MI, MachineBasicBlock *BB, |
842 | unsigned Reg, bool isSigned) const { |
843 | const TargetInstrInfo &TII = *BB->getParent()->getSubtarget().getInstrInfo(); |
844 | const TargetRegisterClass *RC = getRegClassFor(VT: MVT::i64); |
845 | int RShiftOp = isSigned ? BPF::SRA_ri : BPF::SRL_ri; |
846 | MachineFunction *F = BB->getParent(); |
847 | DebugLoc DL = MI.getDebugLoc(); |
848 | |
849 | MachineRegisterInfo &RegInfo = F->getRegInfo(); |
850 | |
851 | if (!isSigned) { |
852 | Register PromotedReg0 = RegInfo.createVirtualRegister(RegClass: RC); |
853 | BuildMI(BB, MIMD: DL, MCID: TII.get(Opcode: BPF::MOV_32_64), DestReg: PromotedReg0).addReg(RegNo: Reg); |
854 | return PromotedReg0; |
855 | } |
856 | Register PromotedReg0 = RegInfo.createVirtualRegister(RegClass: RC); |
857 | Register PromotedReg1 = RegInfo.createVirtualRegister(RegClass: RC); |
858 | Register PromotedReg2 = RegInfo.createVirtualRegister(RegClass: RC); |
859 | if (HasMovsx) { |
860 | BuildMI(BB, MIMD: DL, MCID: TII.get(Opcode: BPF::MOVSX_rr_32), DestReg: PromotedReg0).addReg(RegNo: Reg); |
861 | } else { |
862 | BuildMI(BB, MIMD: DL, MCID: TII.get(Opcode: BPF::MOV_32_64), DestReg: PromotedReg0).addReg(RegNo: Reg); |
863 | BuildMI(BB, MIMD: DL, MCID: TII.get(Opcode: BPF::SLL_ri), DestReg: PromotedReg1) |
864 | .addReg(RegNo: PromotedReg0).addImm(Val: 32); |
865 | BuildMI(BB, MIMD: DL, MCID: TII.get(Opcode: RShiftOp), DestReg: PromotedReg2) |
866 | .addReg(RegNo: PromotedReg1).addImm(Val: 32); |
867 | } |
868 | |
869 | return PromotedReg2; |
870 | } |
871 | |
872 | MachineBasicBlock * |
873 | BPFTargetLowering::EmitInstrWithCustomInserterMemcpy(MachineInstr &MI, |
874 | MachineBasicBlock *BB) |
875 | const { |
876 | MachineFunction *MF = MI.getParent()->getParent(); |
877 | MachineRegisterInfo &MRI = MF->getRegInfo(); |
878 | MachineInstrBuilder MIB(*MF, MI); |
879 | unsigned ScratchReg; |
880 | |
881 | // This function does custom insertion during lowering BPFISD::MEMCPY which |
882 | // only has two register operands from memcpy semantics, the copy source |
883 | // address and the copy destination address. |
884 | // |
885 | // Because we will expand BPFISD::MEMCPY into load/store pairs, we will need |
886 | // a third scratch register to serve as the destination register of load and |
887 | // source register of store. |
888 | // |
889 | // The scratch register here is with the Define | Dead | EarlyClobber flags. |
890 | // The EarlyClobber flag has the semantic property that the operand it is |
891 | // attached to is clobbered before the rest of the inputs are read. Hence it |
892 | // must be unique among the operands to the instruction. The Define flag is |
893 | // needed to coerce the machine verifier that an Undef value isn't a problem |
894 | // as we anyway is loading memory into it. The Dead flag is needed as the |
895 | // value in scratch isn't supposed to be used by any other instruction. |
896 | ScratchReg = MRI.createVirtualRegister(RegClass: &BPF::GPRRegClass); |
897 | MIB.addReg(RegNo: ScratchReg, |
898 | flags: RegState::Define | RegState::Dead | RegState::EarlyClobber); |
899 | |
900 | return BB; |
901 | } |
902 | |
903 | MachineBasicBlock * |
904 | BPFTargetLowering::EmitInstrWithCustomInserter(MachineInstr &MI, |
905 | MachineBasicBlock *BB) const { |
906 | const TargetInstrInfo &TII = *BB->getParent()->getSubtarget().getInstrInfo(); |
907 | DebugLoc DL = MI.getDebugLoc(); |
908 | unsigned Opc = MI.getOpcode(); |
909 | bool isSelectRROp = (Opc == BPF::Select || |
910 | Opc == BPF::Select_64_32 || |
911 | Opc == BPF::Select_32 || |
912 | Opc == BPF::Select_32_64); |
913 | |
914 | bool isMemcpyOp = Opc == BPF::MEMCPY; |
915 | |
916 | #ifndef NDEBUG |
917 | bool isSelectRIOp = (Opc == BPF::Select_Ri || |
918 | Opc == BPF::Select_Ri_64_32 || |
919 | Opc == BPF::Select_Ri_32 || |
920 | Opc == BPF::Select_Ri_32_64); |
921 | |
922 | if (!(isSelectRROp || isSelectRIOp || isMemcpyOp)) |
923 | report_fatal_error("unhandled instruction type: " + Twine(Opc)); |
924 | #endif |
925 | |
926 | if (isMemcpyOp) |
927 | return EmitInstrWithCustomInserterMemcpy(MI, BB); |
928 | |
929 | bool is32BitCmp = (Opc == BPF::Select_32 || |
930 | Opc == BPF::Select_32_64 || |
931 | Opc == BPF::Select_Ri_32 || |
932 | Opc == BPF::Select_Ri_32_64); |
933 | |
934 | // To "insert" a SELECT instruction, we actually have to insert the diamond |
935 | // control-flow pattern. The incoming instruction knows the destination vreg |
936 | // to set, the condition code register to branch on, the true/false values to |
937 | // select between, and a branch opcode to use. |
938 | const BasicBlock *LLVM_BB = BB->getBasicBlock(); |
939 | MachineFunction::iterator I = ++BB->getIterator(); |
940 | |
941 | // ThisMBB: |
942 | // ... |
943 | // TrueVal = ... |
944 | // jmp_XX r1, r2 goto Copy1MBB |
945 | // fallthrough --> Copy0MBB |
946 | MachineBasicBlock *ThisMBB = BB; |
947 | MachineFunction *F = BB->getParent(); |
948 | MachineBasicBlock *Copy0MBB = F->CreateMachineBasicBlock(BB: LLVM_BB); |
949 | MachineBasicBlock *Copy1MBB = F->CreateMachineBasicBlock(BB: LLVM_BB); |
950 | |
951 | F->insert(MBBI: I, MBB: Copy0MBB); |
952 | F->insert(MBBI: I, MBB: Copy1MBB); |
953 | // Update machine-CFG edges by transferring all successors of the current |
954 | // block to the new block which will contain the Phi node for the select. |
955 | Copy1MBB->splice(Where: Copy1MBB->begin(), Other: BB, |
956 | From: std::next(x: MachineBasicBlock::iterator(MI)), To: BB->end()); |
957 | Copy1MBB->transferSuccessorsAndUpdatePHIs(FromMBB: BB); |
958 | // Next, add the true and fallthrough blocks as its successors. |
959 | BB->addSuccessor(Succ: Copy0MBB); |
960 | BB->addSuccessor(Succ: Copy1MBB); |
961 | |
962 | // Insert Branch if Flag |
963 | int CC = MI.getOperand(i: 3).getImm(); |
964 | int NewCC; |
965 | switch (CC) { |
966 | #define SET_NEWCC(X, Y) \ |
967 | case ISD::X: \ |
968 | if (is32BitCmp && HasJmp32) \ |
969 | NewCC = isSelectRROp ? BPF::Y##_rr_32 : BPF::Y##_ri_32; \ |
970 | else \ |
971 | NewCC = isSelectRROp ? BPF::Y##_rr : BPF::Y##_ri; \ |
972 | break |
973 | SET_NEWCC(SETGT, JSGT); |
974 | SET_NEWCC(SETUGT, JUGT); |
975 | SET_NEWCC(SETGE, JSGE); |
976 | SET_NEWCC(SETUGE, JUGE); |
977 | SET_NEWCC(SETEQ, JEQ); |
978 | SET_NEWCC(SETNE, JNE); |
979 | SET_NEWCC(SETLT, JSLT); |
980 | SET_NEWCC(SETULT, JULT); |
981 | SET_NEWCC(SETLE, JSLE); |
982 | SET_NEWCC(SETULE, JULE); |
983 | default: |
984 | report_fatal_error(reason: "unimplemented select CondCode " + Twine(CC)); |
985 | } |
986 | |
987 | Register LHS = MI.getOperand(i: 1).getReg(); |
988 | bool isSignedCmp = (CC == ISD::SETGT || |
989 | CC == ISD::SETGE || |
990 | CC == ISD::SETLT || |
991 | CC == ISD::SETLE); |
992 | |
993 | // eBPF at the moment only has 64-bit comparison. Any 32-bit comparison need |
994 | // to be promoted, however if the 32-bit comparison operands are destination |
995 | // registers then they are implicitly zero-extended already, there is no |
996 | // need of explicit zero-extend sequence for them. |
997 | // |
998 | // We simply do extension for all situations in this method, but we will |
999 | // try to remove those unnecessary in BPFMIPeephole pass. |
1000 | if (is32BitCmp && !HasJmp32) |
1001 | LHS = EmitSubregExt(MI, BB, Reg: LHS, isSigned: isSignedCmp); |
1002 | |
1003 | if (isSelectRROp) { |
1004 | Register RHS = MI.getOperand(i: 2).getReg(); |
1005 | |
1006 | if (is32BitCmp && !HasJmp32) |
1007 | RHS = EmitSubregExt(MI, BB, Reg: RHS, isSigned: isSignedCmp); |
1008 | |
1009 | BuildMI(BB, MIMD: DL, MCID: TII.get(Opcode: NewCC)).addReg(RegNo: LHS).addReg(RegNo: RHS).addMBB(MBB: Copy1MBB); |
1010 | } else { |
1011 | int64_t imm32 = MI.getOperand(i: 2).getImm(); |
1012 | // Check before we build J*_ri instruction. |
1013 | if (!isInt<32>(x: imm32)) |
1014 | report_fatal_error(reason: "immediate overflows 32 bits: " + Twine(imm32)); |
1015 | BuildMI(BB, MIMD: DL, MCID: TII.get(Opcode: NewCC)) |
1016 | .addReg(RegNo: LHS).addImm(Val: imm32).addMBB(MBB: Copy1MBB); |
1017 | } |
1018 | |
1019 | // Copy0MBB: |
1020 | // %FalseValue = ... |
1021 | // # fallthrough to Copy1MBB |
1022 | BB = Copy0MBB; |
1023 | |
1024 | // Update machine-CFG edges |
1025 | BB->addSuccessor(Succ: Copy1MBB); |
1026 | |
1027 | // Copy1MBB: |
1028 | // %Result = phi [ %FalseValue, Copy0MBB ], [ %TrueValue, ThisMBB ] |
1029 | // ... |
1030 | BB = Copy1MBB; |
1031 | BuildMI(BB&: *BB, I: BB->begin(), MIMD: DL, MCID: TII.get(Opcode: BPF::PHI), DestReg: MI.getOperand(i: 0).getReg()) |
1032 | .addReg(RegNo: MI.getOperand(i: 5).getReg()) |
1033 | .addMBB(MBB: Copy0MBB) |
1034 | .addReg(RegNo: MI.getOperand(i: 4).getReg()) |
1035 | .addMBB(MBB: ThisMBB); |
1036 | |
1037 | MI.eraseFromParent(); // The pseudo instruction is gone now. |
1038 | return BB; |
1039 | } |
1040 | |
1041 | EVT BPFTargetLowering::getSetCCResultType(const DataLayout &, LLVMContext &, |
1042 | EVT VT) const { |
1043 | return getHasAlu32() ? MVT::i32 : MVT::i64; |
1044 | } |
1045 | |
1046 | MVT BPFTargetLowering::getScalarShiftAmountTy(const DataLayout &DL, |
1047 | EVT VT) const { |
1048 | return (getHasAlu32() && VT == MVT::i32) ? MVT::i32 : MVT::i64; |
1049 | } |
1050 | |
1051 | bool BPFTargetLowering::isLegalAddressingMode(const DataLayout &DL, |
1052 | const AddrMode &AM, Type *Ty, |
1053 | unsigned AS, |
1054 | Instruction *I) const { |
1055 | // No global is ever allowed as a base. |
1056 | if (AM.BaseGV) |
1057 | return false; |
1058 | |
1059 | switch (AM.Scale) { |
1060 | case 0: // "r+i" or just "i", depending on HasBaseReg. |
1061 | break; |
1062 | case 1: |
1063 | if (!AM.HasBaseReg) // allow "r+i". |
1064 | break; |
1065 | return false; // disallow "r+r" or "r+r+i". |
1066 | default: |
1067 | return false; |
1068 | } |
1069 | |
1070 | return true; |
1071 | } |
1072 | |