1 | //===-- AVRISelLowering.cpp - AVR 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 AVR uses to lower LLVM code into a |
10 | // selection DAG. |
11 | // |
12 | //===----------------------------------------------------------------------===// |
13 | |
14 | #include "AVRISelLowering.h" |
15 | |
16 | #include "llvm/ADT/ArrayRef.h" |
17 | #include "llvm/ADT/StringSwitch.h" |
18 | #include "llvm/CodeGen/CallingConvLower.h" |
19 | #include "llvm/CodeGen/MachineFrameInfo.h" |
20 | #include "llvm/CodeGen/MachineInstrBuilder.h" |
21 | #include "llvm/CodeGen/MachineRegisterInfo.h" |
22 | #include "llvm/CodeGen/SelectionDAG.h" |
23 | #include "llvm/CodeGen/TargetLoweringObjectFileImpl.h" |
24 | #include "llvm/IR/Function.h" |
25 | #include "llvm/Support/ErrorHandling.h" |
26 | |
27 | #include "AVR.h" |
28 | #include "AVRMachineFunctionInfo.h" |
29 | #include "AVRSubtarget.h" |
30 | #include "AVRTargetMachine.h" |
31 | #include "MCTargetDesc/AVRMCTargetDesc.h" |
32 | |
33 | namespace llvm { |
34 | |
35 | AVRTargetLowering::AVRTargetLowering(const AVRTargetMachine &TM, |
36 | const AVRSubtarget &STI) |
37 | : TargetLowering(TM), Subtarget(STI) { |
38 | // Set up the register classes. |
39 | addRegisterClass(VT: MVT::i8, RC: &AVR::GPR8RegClass); |
40 | addRegisterClass(VT: MVT::i16, RC: &AVR::DREGSRegClass); |
41 | |
42 | // Compute derived properties from the register classes. |
43 | computeRegisterProperties(TRI: Subtarget.getRegisterInfo()); |
44 | |
45 | setBooleanContents(ZeroOrOneBooleanContent); |
46 | setBooleanVectorContents(ZeroOrOneBooleanContent); |
47 | setSchedulingPreference(Sched::RegPressure); |
48 | setStackPointerRegisterToSaveRestore(AVR::SP); |
49 | setSupportsUnalignedAtomics(true); |
50 | |
51 | setOperationAction(Op: ISD::GlobalAddress, VT: MVT::i16, Action: Custom); |
52 | setOperationAction(Op: ISD::BlockAddress, VT: MVT::i16, Action: Custom); |
53 | |
54 | setOperationAction(Op: ISD::STACKSAVE, VT: MVT::Other, Action: Expand); |
55 | setOperationAction(Op: ISD::STACKRESTORE, VT: MVT::Other, Action: Expand); |
56 | setOperationAction(Op: ISD::DYNAMIC_STACKALLOC, VT: MVT::i8, Action: Expand); |
57 | setOperationAction(Op: ISD::DYNAMIC_STACKALLOC, VT: MVT::i16, Action: Expand); |
58 | |
59 | setOperationAction(Op: ISD::INLINEASM, VT: MVT::Other, Action: Custom); |
60 | |
61 | for (MVT VT : MVT::integer_valuetypes()) { |
62 | for (auto N : {ISD::EXTLOAD, ISD::SEXTLOAD, ISD::ZEXTLOAD}) { |
63 | setLoadExtAction(ExtType: N, ValVT: VT, MemVT: MVT::i1, Action: Promote); |
64 | setLoadExtAction(ExtType: N, ValVT: VT, MemVT: MVT::i8, Action: Expand); |
65 | } |
66 | } |
67 | |
68 | setTruncStoreAction(ValVT: MVT::i16, MemVT: MVT::i8, Action: Expand); |
69 | |
70 | for (MVT VT : MVT::integer_valuetypes()) { |
71 | setOperationAction(Op: ISD::ADDC, VT, Action: Legal); |
72 | setOperationAction(Op: ISD::SUBC, VT, Action: Legal); |
73 | setOperationAction(Op: ISD::ADDE, VT, Action: Legal); |
74 | setOperationAction(Op: ISD::SUBE, VT, Action: Legal); |
75 | } |
76 | |
77 | // sub (x, imm) gets canonicalized to add (x, -imm), so for illegal types |
78 | // revert into a sub since we don't have an add with immediate instruction. |
79 | setOperationAction(Op: ISD::ADD, VT: MVT::i32, Action: Custom); |
80 | setOperationAction(Op: ISD::ADD, VT: MVT::i64, Action: Custom); |
81 | |
82 | // our shift instructions are only able to shift 1 bit at a time, so handle |
83 | // this in a custom way. |
84 | setOperationAction(Op: ISD::SRA, VT: MVT::i8, Action: Custom); |
85 | setOperationAction(Op: ISD::SHL, VT: MVT::i8, Action: Custom); |
86 | setOperationAction(Op: ISD::SRL, VT: MVT::i8, Action: Custom); |
87 | setOperationAction(Op: ISD::SRA, VT: MVT::i16, Action: Custom); |
88 | setOperationAction(Op: ISD::SHL, VT: MVT::i16, Action: Custom); |
89 | setOperationAction(Op: ISD::SRL, VT: MVT::i16, Action: Custom); |
90 | setOperationAction(Op: ISD::SRA, VT: MVT::i32, Action: Custom); |
91 | setOperationAction(Op: ISD::SHL, VT: MVT::i32, Action: Custom); |
92 | setOperationAction(Op: ISD::SRL, VT: MVT::i32, Action: Custom); |
93 | setOperationAction(Op: ISD::SHL_PARTS, VT: MVT::i16, Action: Expand); |
94 | setOperationAction(Op: ISD::SRA_PARTS, VT: MVT::i16, Action: Expand); |
95 | setOperationAction(Op: ISD::SRL_PARTS, VT: MVT::i16, Action: Expand); |
96 | |
97 | setOperationAction(Op: ISD::ROTL, VT: MVT::i8, Action: Custom); |
98 | setOperationAction(Op: ISD::ROTL, VT: MVT::i16, Action: Expand); |
99 | setOperationAction(Op: ISD::ROTR, VT: MVT::i8, Action: Custom); |
100 | setOperationAction(Op: ISD::ROTR, VT: MVT::i16, Action: Expand); |
101 | |
102 | setOperationAction(Op: ISD::BR_CC, VT: MVT::i8, Action: Custom); |
103 | setOperationAction(Op: ISD::BR_CC, VT: MVT::i16, Action: Custom); |
104 | setOperationAction(Op: ISD::BR_CC, VT: MVT::i32, Action: Custom); |
105 | setOperationAction(Op: ISD::BR_CC, VT: MVT::i64, Action: Custom); |
106 | setOperationAction(Op: ISD::BRCOND, VT: MVT::Other, Action: Expand); |
107 | |
108 | setOperationAction(Op: ISD::SELECT_CC, VT: MVT::i8, Action: Custom); |
109 | setOperationAction(Op: ISD::SELECT_CC, VT: MVT::i16, Action: Custom); |
110 | setOperationAction(Op: ISD::SELECT_CC, VT: MVT::i32, Action: Expand); |
111 | setOperationAction(Op: ISD::SELECT_CC, VT: MVT::i64, Action: Expand); |
112 | setOperationAction(Op: ISD::SETCC, VT: MVT::i8, Action: Custom); |
113 | setOperationAction(Op: ISD::SETCC, VT: MVT::i16, Action: Custom); |
114 | setOperationAction(Op: ISD::SETCC, VT: MVT::i32, Action: Custom); |
115 | setOperationAction(Op: ISD::SETCC, VT: MVT::i64, Action: Custom); |
116 | setOperationAction(Op: ISD::SELECT, VT: MVT::i8, Action: Expand); |
117 | setOperationAction(Op: ISD::SELECT, VT: MVT::i16, Action: Expand); |
118 | |
119 | setOperationAction(Op: ISD::BSWAP, VT: MVT::i16, Action: Expand); |
120 | |
121 | // Add support for postincrement and predecrement load/stores. |
122 | setIndexedLoadAction(IdxModes: ISD::POST_INC, VT: MVT::i8, Action: Legal); |
123 | setIndexedLoadAction(IdxModes: ISD::POST_INC, VT: MVT::i16, Action: Legal); |
124 | setIndexedLoadAction(IdxModes: ISD::PRE_DEC, VT: MVT::i8, Action: Legal); |
125 | setIndexedLoadAction(IdxModes: ISD::PRE_DEC, VT: MVT::i16, Action: Legal); |
126 | setIndexedStoreAction(IdxModes: ISD::POST_INC, VT: MVT::i8, Action: Legal); |
127 | setIndexedStoreAction(IdxModes: ISD::POST_INC, VT: MVT::i16, Action: Legal); |
128 | setIndexedStoreAction(IdxModes: ISD::PRE_DEC, VT: MVT::i8, Action: Legal); |
129 | setIndexedStoreAction(IdxModes: ISD::PRE_DEC, VT: MVT::i16, Action: Legal); |
130 | |
131 | setOperationAction(Op: ISD::BR_JT, VT: MVT::Other, Action: Expand); |
132 | |
133 | setOperationAction(Op: ISD::VASTART, VT: MVT::Other, Action: Custom); |
134 | setOperationAction(Op: ISD::VAEND, VT: MVT::Other, Action: Expand); |
135 | setOperationAction(Op: ISD::VAARG, VT: MVT::Other, Action: Expand); |
136 | setOperationAction(Op: ISD::VACOPY, VT: MVT::Other, Action: Expand); |
137 | |
138 | // Atomic operations which must be lowered to rtlib calls |
139 | for (MVT VT : MVT::integer_valuetypes()) { |
140 | setOperationAction(Op: ISD::ATOMIC_SWAP, VT, Action: Expand); |
141 | setOperationAction(Op: ISD::ATOMIC_CMP_SWAP, VT, Action: Expand); |
142 | setOperationAction(Op: ISD::ATOMIC_LOAD_NAND, VT, Action: Expand); |
143 | setOperationAction(Op: ISD::ATOMIC_LOAD_MAX, VT, Action: Expand); |
144 | setOperationAction(Op: ISD::ATOMIC_LOAD_MIN, VT, Action: Expand); |
145 | setOperationAction(Op: ISD::ATOMIC_LOAD_UMAX, VT, Action: Expand); |
146 | setOperationAction(Op: ISD::ATOMIC_LOAD_UMIN, VT, Action: Expand); |
147 | } |
148 | |
149 | // Division/remainder |
150 | setOperationAction(Op: ISD::UDIV, VT: MVT::i8, Action: Expand); |
151 | setOperationAction(Op: ISD::UDIV, VT: MVT::i16, Action: Expand); |
152 | setOperationAction(Op: ISD::UREM, VT: MVT::i8, Action: Expand); |
153 | setOperationAction(Op: ISD::UREM, VT: MVT::i16, Action: Expand); |
154 | setOperationAction(Op: ISD::SDIV, VT: MVT::i8, Action: Expand); |
155 | setOperationAction(Op: ISD::SDIV, VT: MVT::i16, Action: Expand); |
156 | setOperationAction(Op: ISD::SREM, VT: MVT::i8, Action: Expand); |
157 | setOperationAction(Op: ISD::SREM, VT: MVT::i16, Action: Expand); |
158 | |
159 | // Make division and modulus custom |
160 | setOperationAction(Op: ISD::UDIVREM, VT: MVT::i8, Action: Custom); |
161 | setOperationAction(Op: ISD::UDIVREM, VT: MVT::i16, Action: Custom); |
162 | setOperationAction(Op: ISD::UDIVREM, VT: MVT::i32, Action: Custom); |
163 | setOperationAction(Op: ISD::SDIVREM, VT: MVT::i8, Action: Custom); |
164 | setOperationAction(Op: ISD::SDIVREM, VT: MVT::i16, Action: Custom); |
165 | setOperationAction(Op: ISD::SDIVREM, VT: MVT::i32, Action: Custom); |
166 | |
167 | // Do not use MUL. The AVR instructions are closer to SMUL_LOHI &co. |
168 | setOperationAction(Op: ISD::MUL, VT: MVT::i8, Action: Expand); |
169 | setOperationAction(Op: ISD::MUL, VT: MVT::i16, Action: Expand); |
170 | |
171 | // Expand 16 bit multiplications. |
172 | setOperationAction(Op: ISD::SMUL_LOHI, VT: MVT::i16, Action: Expand); |
173 | setOperationAction(Op: ISD::UMUL_LOHI, VT: MVT::i16, Action: Expand); |
174 | |
175 | // Expand multiplications to libcalls when there is |
176 | // no hardware MUL. |
177 | if (!Subtarget.supportsMultiplication()) { |
178 | setOperationAction(Op: ISD::SMUL_LOHI, VT: MVT::i8, Action: Expand); |
179 | setOperationAction(Op: ISD::UMUL_LOHI, VT: MVT::i8, Action: Expand); |
180 | } |
181 | |
182 | for (MVT VT : MVT::integer_valuetypes()) { |
183 | setOperationAction(Op: ISD::MULHS, VT, Action: Expand); |
184 | setOperationAction(Op: ISD::MULHU, VT, Action: Expand); |
185 | } |
186 | |
187 | for (MVT VT : MVT::integer_valuetypes()) { |
188 | setOperationAction(Op: ISD::CTPOP, VT, Action: Expand); |
189 | setOperationAction(Op: ISD::CTLZ, VT, Action: Expand); |
190 | setOperationAction(Op: ISD::CTTZ, VT, Action: Expand); |
191 | } |
192 | |
193 | for (MVT VT : MVT::integer_valuetypes()) { |
194 | setOperationAction(Op: ISD::SIGN_EXTEND_INREG, VT, Action: Expand); |
195 | // TODO: The generated code is pretty poor. Investigate using the |
196 | // same "shift and subtract with carry" trick that we do for |
197 | // extending 8-bit to 16-bit. This may require infrastructure |
198 | // improvements in how we treat 16-bit "registers" to be feasible. |
199 | } |
200 | |
201 | setMinFunctionAlignment(Align(2)); |
202 | setMinimumJumpTableEntries(UINT_MAX); |
203 | } |
204 | |
205 | EVT AVRTargetLowering::getSetCCResultType(const DataLayout &DL, LLVMContext &, |
206 | EVT VT) const { |
207 | assert(!VT.isVector() && "No AVR SetCC type for vectors!" ); |
208 | return MVT::i8; |
209 | } |
210 | |
211 | SDValue AVRTargetLowering::LowerShifts(SDValue Op, SelectionDAG &DAG) const { |
212 | unsigned Opc8; |
213 | const SDNode *N = Op.getNode(); |
214 | EVT VT = Op.getValueType(); |
215 | SDLoc dl(N); |
216 | assert(llvm::has_single_bit<uint32_t>(VT.getSizeInBits()) && |
217 | "Expected power-of-2 shift amount" ); |
218 | |
219 | if (VT.getSizeInBits() == 32) { |
220 | if (!isa<ConstantSDNode>(Val: N->getOperand(Num: 1))) { |
221 | // 32-bit shifts are converted to a loop in IR. |
222 | // This should be unreachable. |
223 | report_fatal_error(reason: "Expected a constant shift amount!" ); |
224 | } |
225 | SDVTList ResTys = DAG.getVTList(VT1: MVT::i16, VT2: MVT::i16); |
226 | SDValue SrcLo = |
227 | DAG.getNode(Opcode: ISD::EXTRACT_ELEMENT, DL: dl, VT: MVT::i16, N1: Op.getOperand(i: 0), |
228 | N2: DAG.getConstant(Val: 0, DL: dl, VT: MVT::i16)); |
229 | SDValue SrcHi = |
230 | DAG.getNode(Opcode: ISD::EXTRACT_ELEMENT, DL: dl, VT: MVT::i16, N1: Op.getOperand(i: 0), |
231 | N2: DAG.getConstant(Val: 1, DL: dl, VT: MVT::i16)); |
232 | uint64_t ShiftAmount = N->getConstantOperandVal(Num: 1); |
233 | if (ShiftAmount == 16) { |
234 | // Special case these two operations because they appear to be used by the |
235 | // generic codegen parts to lower 32-bit numbers. |
236 | // TODO: perhaps we can lower shift amounts bigger than 16 to a 16-bit |
237 | // shift of a part of the 32-bit value? |
238 | switch (Op.getOpcode()) { |
239 | case ISD::SHL: { |
240 | SDValue Zero = DAG.getConstant(Val: 0, DL: dl, VT: MVT::i16); |
241 | return DAG.getNode(Opcode: ISD::BUILD_PAIR, DL: dl, VT: MVT::i32, N1: Zero, N2: SrcLo); |
242 | } |
243 | case ISD::SRL: { |
244 | SDValue Zero = DAG.getConstant(Val: 0, DL: dl, VT: MVT::i16); |
245 | return DAG.getNode(Opcode: ISD::BUILD_PAIR, DL: dl, VT: MVT::i32, N1: SrcHi, N2: Zero); |
246 | } |
247 | } |
248 | } |
249 | SDValue Cnt = DAG.getTargetConstant(Val: ShiftAmount, DL: dl, VT: MVT::i8); |
250 | unsigned Opc; |
251 | switch (Op.getOpcode()) { |
252 | default: |
253 | llvm_unreachable("Invalid 32-bit shift opcode!" ); |
254 | case ISD::SHL: |
255 | Opc = AVRISD::LSLW; |
256 | break; |
257 | case ISD::SRL: |
258 | Opc = AVRISD::LSRW; |
259 | break; |
260 | case ISD::SRA: |
261 | Opc = AVRISD::ASRW; |
262 | break; |
263 | } |
264 | SDValue Result = DAG.getNode(Opcode: Opc, DL: dl, VTList: ResTys, N1: SrcLo, N2: SrcHi, N3: Cnt); |
265 | return DAG.getNode(Opcode: ISD::BUILD_PAIR, DL: dl, VT: MVT::i32, N1: Result.getValue(R: 0), |
266 | N2: Result.getValue(R: 1)); |
267 | } |
268 | |
269 | // Expand non-constant shifts to loops. |
270 | if (!isa<ConstantSDNode>(Val: N->getOperand(Num: 1))) { |
271 | switch (Op.getOpcode()) { |
272 | default: |
273 | llvm_unreachable("Invalid shift opcode!" ); |
274 | case ISD::SHL: |
275 | return DAG.getNode(Opcode: AVRISD::LSLLOOP, DL: dl, VT, N1: N->getOperand(Num: 0), |
276 | N2: N->getOperand(Num: 1)); |
277 | case ISD::SRL: |
278 | return DAG.getNode(Opcode: AVRISD::LSRLOOP, DL: dl, VT, N1: N->getOperand(Num: 0), |
279 | N2: N->getOperand(Num: 1)); |
280 | case ISD::ROTL: { |
281 | SDValue Amt = N->getOperand(Num: 1); |
282 | EVT AmtVT = Amt.getValueType(); |
283 | Amt = DAG.getNode(Opcode: ISD::AND, DL: dl, VT: AmtVT, N1: Amt, |
284 | N2: DAG.getConstant(Val: VT.getSizeInBits() - 1, DL: dl, VT: AmtVT)); |
285 | return DAG.getNode(Opcode: AVRISD::ROLLOOP, DL: dl, VT, N1: N->getOperand(Num: 0), N2: Amt); |
286 | } |
287 | case ISD::ROTR: { |
288 | SDValue Amt = N->getOperand(Num: 1); |
289 | EVT AmtVT = Amt.getValueType(); |
290 | Amt = DAG.getNode(Opcode: ISD::AND, DL: dl, VT: AmtVT, N1: Amt, |
291 | N2: DAG.getConstant(Val: VT.getSizeInBits() - 1, DL: dl, VT: AmtVT)); |
292 | return DAG.getNode(Opcode: AVRISD::RORLOOP, DL: dl, VT, N1: N->getOperand(Num: 0), N2: Amt); |
293 | } |
294 | case ISD::SRA: |
295 | return DAG.getNode(Opcode: AVRISD::ASRLOOP, DL: dl, VT, N1: N->getOperand(Num: 0), |
296 | N2: N->getOperand(Num: 1)); |
297 | } |
298 | } |
299 | |
300 | uint64_t ShiftAmount = N->getConstantOperandVal(Num: 1); |
301 | SDValue Victim = N->getOperand(Num: 0); |
302 | |
303 | switch (Op.getOpcode()) { |
304 | case ISD::SRA: |
305 | Opc8 = AVRISD::ASR; |
306 | break; |
307 | case ISD::ROTL: |
308 | Opc8 = AVRISD::ROL; |
309 | ShiftAmount = ShiftAmount % VT.getSizeInBits(); |
310 | break; |
311 | case ISD::ROTR: |
312 | Opc8 = AVRISD::ROR; |
313 | ShiftAmount = ShiftAmount % VT.getSizeInBits(); |
314 | break; |
315 | case ISD::SRL: |
316 | Opc8 = AVRISD::LSR; |
317 | break; |
318 | case ISD::SHL: |
319 | Opc8 = AVRISD::LSL; |
320 | break; |
321 | default: |
322 | llvm_unreachable("Invalid shift opcode" ); |
323 | } |
324 | |
325 | // Optimize int8/int16 shifts. |
326 | if (VT.getSizeInBits() == 8) { |
327 | if (Op.getOpcode() == ISD::SHL && 4 <= ShiftAmount && ShiftAmount < 7) { |
328 | // Optimize LSL when 4 <= ShiftAmount <= 6. |
329 | Victim = DAG.getNode(Opcode: AVRISD::SWAP, DL: dl, VT, Operand: Victim); |
330 | Victim = |
331 | DAG.getNode(Opcode: ISD::AND, DL: dl, VT, N1: Victim, N2: DAG.getConstant(Val: 0xf0, DL: dl, VT)); |
332 | ShiftAmount -= 4; |
333 | } else if (Op.getOpcode() == ISD::SRL && 4 <= ShiftAmount && |
334 | ShiftAmount < 7) { |
335 | // Optimize LSR when 4 <= ShiftAmount <= 6. |
336 | Victim = DAG.getNode(Opcode: AVRISD::SWAP, DL: dl, VT, Operand: Victim); |
337 | Victim = |
338 | DAG.getNode(Opcode: ISD::AND, DL: dl, VT, N1: Victim, N2: DAG.getConstant(Val: 0x0f, DL: dl, VT)); |
339 | ShiftAmount -= 4; |
340 | } else if (Op.getOpcode() == ISD::SHL && ShiftAmount == 7) { |
341 | // Optimize LSL when ShiftAmount == 7. |
342 | Victim = DAG.getNode(Opcode: AVRISD::LSLBN, DL: dl, VT, N1: Victim, |
343 | N2: DAG.getConstant(Val: 7, DL: dl, VT)); |
344 | ShiftAmount = 0; |
345 | } else if (Op.getOpcode() == ISD::SRL && ShiftAmount == 7) { |
346 | // Optimize LSR when ShiftAmount == 7. |
347 | Victim = DAG.getNode(Opcode: AVRISD::LSRBN, DL: dl, VT, N1: Victim, |
348 | N2: DAG.getConstant(Val: 7, DL: dl, VT)); |
349 | ShiftAmount = 0; |
350 | } else if (Op.getOpcode() == ISD::SRA && ShiftAmount == 6) { |
351 | // Optimize ASR when ShiftAmount == 6. |
352 | Victim = DAG.getNode(Opcode: AVRISD::ASRBN, DL: dl, VT, N1: Victim, |
353 | N2: DAG.getConstant(Val: 6, DL: dl, VT)); |
354 | ShiftAmount = 0; |
355 | } else if (Op.getOpcode() == ISD::SRA && ShiftAmount == 7) { |
356 | // Optimize ASR when ShiftAmount == 7. |
357 | Victim = DAG.getNode(Opcode: AVRISD::ASRBN, DL: dl, VT, N1: Victim, |
358 | N2: DAG.getConstant(Val: 7, DL: dl, VT)); |
359 | ShiftAmount = 0; |
360 | } else if (Op.getOpcode() == ISD::ROTL && ShiftAmount == 3) { |
361 | // Optimize left rotation 3 bits to swap then right rotation 1 bit. |
362 | Victim = DAG.getNode(Opcode: AVRISD::SWAP, DL: dl, VT, Operand: Victim); |
363 | Victim = DAG.getNode(Opcode: AVRISD::ROR, DL: dl, VT, Operand: Victim); |
364 | ShiftAmount = 0; |
365 | } else if (Op.getOpcode() == ISD::ROTR && ShiftAmount == 3) { |
366 | // Optimize right rotation 3 bits to swap then left rotation 1 bit. |
367 | Victim = DAG.getNode(Opcode: AVRISD::SWAP, DL: dl, VT, Operand: Victim); |
368 | Victim = DAG.getNode(Opcode: AVRISD::ROL, DL: dl, VT, Operand: Victim); |
369 | ShiftAmount = 0; |
370 | } else if (Op.getOpcode() == ISD::ROTL && ShiftAmount == 7) { |
371 | // Optimize left rotation 7 bits to right rotation 1 bit. |
372 | Victim = DAG.getNode(Opcode: AVRISD::ROR, DL: dl, VT, Operand: Victim); |
373 | ShiftAmount = 0; |
374 | } else if (Op.getOpcode() == ISD::ROTR && ShiftAmount == 7) { |
375 | // Optimize right rotation 7 bits to left rotation 1 bit. |
376 | Victim = DAG.getNode(Opcode: AVRISD::ROL, DL: dl, VT, Operand: Victim); |
377 | ShiftAmount = 0; |
378 | } else if ((Op.getOpcode() == ISD::ROTR || Op.getOpcode() == ISD::ROTL) && |
379 | ShiftAmount >= 4) { |
380 | // Optimize left/right rotation with the SWAP instruction. |
381 | Victim = DAG.getNode(Opcode: AVRISD::SWAP, DL: dl, VT, Operand: Victim); |
382 | ShiftAmount -= 4; |
383 | } |
384 | } else if (VT.getSizeInBits() == 16) { |
385 | if (Op.getOpcode() == ISD::SRA) |
386 | // Special optimization for int16 arithmetic right shift. |
387 | switch (ShiftAmount) { |
388 | case 15: |
389 | Victim = DAG.getNode(Opcode: AVRISD::ASRWN, DL: dl, VT, N1: Victim, |
390 | N2: DAG.getConstant(Val: 15, DL: dl, VT)); |
391 | ShiftAmount = 0; |
392 | break; |
393 | case 14: |
394 | Victim = DAG.getNode(Opcode: AVRISD::ASRWN, DL: dl, VT, N1: Victim, |
395 | N2: DAG.getConstant(Val: 14, DL: dl, VT)); |
396 | ShiftAmount = 0; |
397 | break; |
398 | case 7: |
399 | Victim = DAG.getNode(Opcode: AVRISD::ASRWN, DL: dl, VT, N1: Victim, |
400 | N2: DAG.getConstant(Val: 7, DL: dl, VT)); |
401 | ShiftAmount = 0; |
402 | break; |
403 | default: |
404 | break; |
405 | } |
406 | if (4 <= ShiftAmount && ShiftAmount < 8) |
407 | switch (Op.getOpcode()) { |
408 | case ISD::SHL: |
409 | Victim = DAG.getNode(Opcode: AVRISD::LSLWN, DL: dl, VT, N1: Victim, |
410 | N2: DAG.getConstant(Val: 4, DL: dl, VT)); |
411 | ShiftAmount -= 4; |
412 | break; |
413 | case ISD::SRL: |
414 | Victim = DAG.getNode(Opcode: AVRISD::LSRWN, DL: dl, VT, N1: Victim, |
415 | N2: DAG.getConstant(Val: 4, DL: dl, VT)); |
416 | ShiftAmount -= 4; |
417 | break; |
418 | default: |
419 | break; |
420 | } |
421 | else if (8 <= ShiftAmount && ShiftAmount < 12) |
422 | switch (Op.getOpcode()) { |
423 | case ISD::SHL: |
424 | Victim = DAG.getNode(Opcode: AVRISD::LSLWN, DL: dl, VT, N1: Victim, |
425 | N2: DAG.getConstant(Val: 8, DL: dl, VT)); |
426 | ShiftAmount -= 8; |
427 | // Only operate on the higher byte for remaining shift bits. |
428 | Opc8 = AVRISD::LSLHI; |
429 | break; |
430 | case ISD::SRL: |
431 | Victim = DAG.getNode(Opcode: AVRISD::LSRWN, DL: dl, VT, N1: Victim, |
432 | N2: DAG.getConstant(Val: 8, DL: dl, VT)); |
433 | ShiftAmount -= 8; |
434 | // Only operate on the lower byte for remaining shift bits. |
435 | Opc8 = AVRISD::LSRLO; |
436 | break; |
437 | case ISD::SRA: |
438 | Victim = DAG.getNode(Opcode: AVRISD::ASRWN, DL: dl, VT, N1: Victim, |
439 | N2: DAG.getConstant(Val: 8, DL: dl, VT)); |
440 | ShiftAmount -= 8; |
441 | // Only operate on the lower byte for remaining shift bits. |
442 | Opc8 = AVRISD::ASRLO; |
443 | break; |
444 | default: |
445 | break; |
446 | } |
447 | else if (12 <= ShiftAmount) |
448 | switch (Op.getOpcode()) { |
449 | case ISD::SHL: |
450 | Victim = DAG.getNode(Opcode: AVRISD::LSLWN, DL: dl, VT, N1: Victim, |
451 | N2: DAG.getConstant(Val: 12, DL: dl, VT)); |
452 | ShiftAmount -= 12; |
453 | // Only operate on the higher byte for remaining shift bits. |
454 | Opc8 = AVRISD::LSLHI; |
455 | break; |
456 | case ISD::SRL: |
457 | Victim = DAG.getNode(Opcode: AVRISD::LSRWN, DL: dl, VT, N1: Victim, |
458 | N2: DAG.getConstant(Val: 12, DL: dl, VT)); |
459 | ShiftAmount -= 12; |
460 | // Only operate on the lower byte for remaining shift bits. |
461 | Opc8 = AVRISD::LSRLO; |
462 | break; |
463 | case ISD::SRA: |
464 | Victim = DAG.getNode(Opcode: AVRISD::ASRWN, DL: dl, VT, N1: Victim, |
465 | N2: DAG.getConstant(Val: 8, DL: dl, VT)); |
466 | ShiftAmount -= 8; |
467 | // Only operate on the lower byte for remaining shift bits. |
468 | Opc8 = AVRISD::ASRLO; |
469 | break; |
470 | default: |
471 | break; |
472 | } |
473 | } |
474 | |
475 | while (ShiftAmount--) { |
476 | Victim = DAG.getNode(Opcode: Opc8, DL: dl, VT, Operand: Victim); |
477 | } |
478 | |
479 | return Victim; |
480 | } |
481 | |
482 | SDValue AVRTargetLowering::LowerDivRem(SDValue Op, SelectionDAG &DAG) const { |
483 | unsigned Opcode = Op->getOpcode(); |
484 | assert((Opcode == ISD::SDIVREM || Opcode == ISD::UDIVREM) && |
485 | "Invalid opcode for Div/Rem lowering" ); |
486 | bool IsSigned = (Opcode == ISD::SDIVREM); |
487 | EVT VT = Op->getValueType(ResNo: 0); |
488 | Type *Ty = VT.getTypeForEVT(Context&: *DAG.getContext()); |
489 | |
490 | RTLIB::Libcall LC; |
491 | switch (VT.getSimpleVT().SimpleTy) { |
492 | default: |
493 | llvm_unreachable("Unexpected request for libcall!" ); |
494 | case MVT::i8: |
495 | LC = IsSigned ? RTLIB::SDIVREM_I8 : RTLIB::UDIVREM_I8; |
496 | break; |
497 | case MVT::i16: |
498 | LC = IsSigned ? RTLIB::SDIVREM_I16 : RTLIB::UDIVREM_I16; |
499 | break; |
500 | case MVT::i32: |
501 | LC = IsSigned ? RTLIB::SDIVREM_I32 : RTLIB::UDIVREM_I32; |
502 | break; |
503 | } |
504 | |
505 | SDValue InChain = DAG.getEntryNode(); |
506 | |
507 | TargetLowering::ArgListTy Args; |
508 | TargetLowering::ArgListEntry Entry; |
509 | for (SDValue const &Value : Op->op_values()) { |
510 | Entry.Node = Value; |
511 | Entry.Ty = Value.getValueType().getTypeForEVT(Context&: *DAG.getContext()); |
512 | Entry.IsSExt = IsSigned; |
513 | Entry.IsZExt = !IsSigned; |
514 | Args.push_back(x: Entry); |
515 | } |
516 | |
517 | SDValue Callee = DAG.getExternalSymbol(Sym: getLibcallName(Call: LC), |
518 | VT: getPointerTy(DL: DAG.getDataLayout())); |
519 | |
520 | Type *RetTy = (Type *)StructType::get(elt1: Ty, elts: Ty); |
521 | |
522 | SDLoc dl(Op); |
523 | TargetLowering::CallLoweringInfo CLI(DAG); |
524 | CLI.setDebugLoc(dl) |
525 | .setChain(InChain) |
526 | .setLibCallee(CC: getLibcallCallingConv(Call: LC), ResultType: RetTy, Target: Callee, ArgsList: std::move(Args)) |
527 | .setInRegister() |
528 | .setSExtResult(IsSigned) |
529 | .setZExtResult(!IsSigned); |
530 | |
531 | std::pair<SDValue, SDValue> CallInfo = LowerCallTo(CLI); |
532 | return CallInfo.first; |
533 | } |
534 | |
535 | SDValue AVRTargetLowering::LowerGlobalAddress(SDValue Op, |
536 | SelectionDAG &DAG) const { |
537 | auto DL = DAG.getDataLayout(); |
538 | |
539 | const GlobalValue *GV = cast<GlobalAddressSDNode>(Val&: Op)->getGlobal(); |
540 | int64_t Offset = cast<GlobalAddressSDNode>(Val&: Op)->getOffset(); |
541 | |
542 | // Create the TargetGlobalAddress node, folding in the constant offset. |
543 | SDValue Result = |
544 | DAG.getTargetGlobalAddress(GV, DL: SDLoc(Op), VT: getPointerTy(DL), offset: Offset); |
545 | return DAG.getNode(Opcode: AVRISD::WRAPPER, DL: SDLoc(Op), VT: getPointerTy(DL), Operand: Result); |
546 | } |
547 | |
548 | SDValue AVRTargetLowering::LowerBlockAddress(SDValue Op, |
549 | SelectionDAG &DAG) const { |
550 | auto DL = DAG.getDataLayout(); |
551 | const BlockAddress *BA = cast<BlockAddressSDNode>(Val&: Op)->getBlockAddress(); |
552 | |
553 | SDValue Result = DAG.getTargetBlockAddress(BA, VT: getPointerTy(DL)); |
554 | |
555 | return DAG.getNode(Opcode: AVRISD::WRAPPER, DL: SDLoc(Op), VT: getPointerTy(DL), Operand: Result); |
556 | } |
557 | |
558 | /// IntCCToAVRCC - Convert a DAG integer condition code to an AVR CC. |
559 | static AVRCC::CondCodes intCCToAVRCC(ISD::CondCode CC) { |
560 | switch (CC) { |
561 | default: |
562 | llvm_unreachable("Unknown condition code!" ); |
563 | case ISD::SETEQ: |
564 | return AVRCC::COND_EQ; |
565 | case ISD::SETNE: |
566 | return AVRCC::COND_NE; |
567 | case ISD::SETGE: |
568 | return AVRCC::COND_GE; |
569 | case ISD::SETLT: |
570 | return AVRCC::COND_LT; |
571 | case ISD::SETUGE: |
572 | return AVRCC::COND_SH; |
573 | case ISD::SETULT: |
574 | return AVRCC::COND_LO; |
575 | } |
576 | } |
577 | |
578 | /// Returns appropriate CP/CPI/CPC nodes code for the given 8/16-bit operands. |
579 | SDValue AVRTargetLowering::getAVRCmp(SDValue LHS, SDValue RHS, |
580 | SelectionDAG &DAG, SDLoc DL) const { |
581 | assert((LHS.getSimpleValueType() == RHS.getSimpleValueType()) && |
582 | "LHS and RHS have different types" ); |
583 | assert(((LHS.getSimpleValueType() == MVT::i16) || |
584 | (LHS.getSimpleValueType() == MVT::i8)) && |
585 | "invalid comparison type" ); |
586 | |
587 | SDValue Cmp; |
588 | |
589 | if (LHS.getSimpleValueType() == MVT::i16 && isa<ConstantSDNode>(Val: RHS)) { |
590 | uint64_t Imm = RHS->getAsZExtVal(); |
591 | // Generate a CPI/CPC pair if RHS is a 16-bit constant. Use the zero |
592 | // register for the constant RHS if its lower or higher byte is zero. |
593 | SDValue LHSlo = DAG.getNode(Opcode: ISD::EXTRACT_ELEMENT, DL, VT: MVT::i8, N1: LHS, |
594 | N2: DAG.getIntPtrConstant(Val: 0, DL)); |
595 | SDValue LHShi = DAG.getNode(Opcode: ISD::EXTRACT_ELEMENT, DL, VT: MVT::i8, N1: LHS, |
596 | N2: DAG.getIntPtrConstant(Val: 1, DL)); |
597 | SDValue RHSlo = (Imm & 0xff) == 0 |
598 | ? DAG.getRegister(Reg: Subtarget.getZeroRegister(), VT: MVT::i8) |
599 | : DAG.getNode(Opcode: ISD::EXTRACT_ELEMENT, DL, VT: MVT::i8, N1: RHS, |
600 | N2: DAG.getIntPtrConstant(Val: 0, DL)); |
601 | SDValue RHShi = (Imm & 0xff00) == 0 |
602 | ? DAG.getRegister(Reg: Subtarget.getZeroRegister(), VT: MVT::i8) |
603 | : DAG.getNode(Opcode: ISD::EXTRACT_ELEMENT, DL, VT: MVT::i8, N1: RHS, |
604 | N2: DAG.getIntPtrConstant(Val: 1, DL)); |
605 | Cmp = DAG.getNode(Opcode: AVRISD::CMP, DL, VT: MVT::Glue, N1: LHSlo, N2: RHSlo); |
606 | Cmp = DAG.getNode(Opcode: AVRISD::CMPC, DL, VT: MVT::Glue, N1: LHShi, N2: RHShi, N3: Cmp); |
607 | } else if (RHS.getSimpleValueType() == MVT::i16 && isa<ConstantSDNode>(Val: LHS)) { |
608 | // Generate a CPI/CPC pair if LHS is a 16-bit constant. Use the zero |
609 | // register for the constant LHS if its lower or higher byte is zero. |
610 | uint64_t Imm = LHS->getAsZExtVal(); |
611 | SDValue LHSlo = (Imm & 0xff) == 0 |
612 | ? DAG.getRegister(Reg: Subtarget.getZeroRegister(), VT: MVT::i8) |
613 | : DAG.getNode(Opcode: ISD::EXTRACT_ELEMENT, DL, VT: MVT::i8, N1: LHS, |
614 | N2: DAG.getIntPtrConstant(Val: 0, DL)); |
615 | SDValue LHShi = (Imm & 0xff00) == 0 |
616 | ? DAG.getRegister(Reg: Subtarget.getZeroRegister(), VT: MVT::i8) |
617 | : DAG.getNode(Opcode: ISD::EXTRACT_ELEMENT, DL, VT: MVT::i8, N1: LHS, |
618 | N2: DAG.getIntPtrConstant(Val: 1, DL)); |
619 | SDValue RHSlo = DAG.getNode(Opcode: ISD::EXTRACT_ELEMENT, DL, VT: MVT::i8, N1: RHS, |
620 | N2: DAG.getIntPtrConstant(Val: 0, DL)); |
621 | SDValue RHShi = DAG.getNode(Opcode: ISD::EXTRACT_ELEMENT, DL, VT: MVT::i8, N1: RHS, |
622 | N2: DAG.getIntPtrConstant(Val: 1, DL)); |
623 | Cmp = DAG.getNode(Opcode: AVRISD::CMP, DL, VT: MVT::Glue, N1: LHSlo, N2: RHSlo); |
624 | Cmp = DAG.getNode(Opcode: AVRISD::CMPC, DL, VT: MVT::Glue, N1: LHShi, N2: RHShi, N3: Cmp); |
625 | } else { |
626 | // Generate ordinary 16-bit comparison. |
627 | Cmp = DAG.getNode(Opcode: AVRISD::CMP, DL, VT: MVT::Glue, N1: LHS, N2: RHS); |
628 | } |
629 | |
630 | return Cmp; |
631 | } |
632 | |
633 | /// Returns appropriate AVR CMP/CMPC nodes and corresponding condition code for |
634 | /// the given operands. |
635 | SDValue AVRTargetLowering::getAVRCmp(SDValue LHS, SDValue RHS, ISD::CondCode CC, |
636 | SDValue &AVRcc, SelectionDAG &DAG, |
637 | SDLoc DL) const { |
638 | SDValue Cmp; |
639 | EVT VT = LHS.getValueType(); |
640 | bool UseTest = false; |
641 | |
642 | switch (CC) { |
643 | default: |
644 | break; |
645 | case ISD::SETLE: { |
646 | // Swap operands and reverse the branching condition. |
647 | std::swap(a&: LHS, b&: RHS); |
648 | CC = ISD::SETGE; |
649 | break; |
650 | } |
651 | case ISD::SETGT: { |
652 | if (const ConstantSDNode *C = dyn_cast<ConstantSDNode>(Val&: RHS)) { |
653 | switch (C->getSExtValue()) { |
654 | case -1: { |
655 | // When doing lhs > -1 use a tst instruction on the top part of lhs |
656 | // and use brpl instead of using a chain of cp/cpc. |
657 | UseTest = true; |
658 | AVRcc = DAG.getConstant(Val: AVRCC::COND_PL, DL, VT: MVT::i8); |
659 | break; |
660 | } |
661 | case 0: { |
662 | // Turn lhs > 0 into 0 < lhs since 0 can be materialized with |
663 | // __zero_reg__ in lhs. |
664 | RHS = LHS; |
665 | LHS = DAG.getConstant(Val: 0, DL, VT); |
666 | CC = ISD::SETLT; |
667 | break; |
668 | } |
669 | default: { |
670 | // Turn lhs < rhs with lhs constant into rhs >= lhs+1, this allows |
671 | // us to fold the constant into the cmp instruction. |
672 | RHS = DAG.getConstant(Val: C->getSExtValue() + 1, DL, VT); |
673 | CC = ISD::SETGE; |
674 | break; |
675 | } |
676 | } |
677 | break; |
678 | } |
679 | // Swap operands and reverse the branching condition. |
680 | std::swap(a&: LHS, b&: RHS); |
681 | CC = ISD::SETLT; |
682 | break; |
683 | } |
684 | case ISD::SETLT: { |
685 | if (const ConstantSDNode *C = dyn_cast<ConstantSDNode>(Val&: RHS)) { |
686 | switch (C->getSExtValue()) { |
687 | case 1: { |
688 | // Turn lhs < 1 into 0 >= lhs since 0 can be materialized with |
689 | // __zero_reg__ in lhs. |
690 | RHS = LHS; |
691 | LHS = DAG.getConstant(Val: 0, DL, VT); |
692 | CC = ISD::SETGE; |
693 | break; |
694 | } |
695 | case 0: { |
696 | // When doing lhs < 0 use a tst instruction on the top part of lhs |
697 | // and use brmi instead of using a chain of cp/cpc. |
698 | UseTest = true; |
699 | AVRcc = DAG.getConstant(Val: AVRCC::COND_MI, DL, VT: MVT::i8); |
700 | break; |
701 | } |
702 | } |
703 | } |
704 | break; |
705 | } |
706 | case ISD::SETULE: { |
707 | // Swap operands and reverse the branching condition. |
708 | std::swap(a&: LHS, b&: RHS); |
709 | CC = ISD::SETUGE; |
710 | break; |
711 | } |
712 | case ISD::SETUGT: { |
713 | // Turn lhs < rhs with lhs constant into rhs >= lhs+1, this allows us to |
714 | // fold the constant into the cmp instruction. |
715 | if (const ConstantSDNode *C = dyn_cast<ConstantSDNode>(Val&: RHS)) { |
716 | RHS = DAG.getConstant(Val: C->getSExtValue() + 1, DL, VT); |
717 | CC = ISD::SETUGE; |
718 | break; |
719 | } |
720 | // Swap operands and reverse the branching condition. |
721 | std::swap(a&: LHS, b&: RHS); |
722 | CC = ISD::SETULT; |
723 | break; |
724 | } |
725 | } |
726 | |
727 | // Expand 32 and 64 bit comparisons with custom CMP and CMPC nodes instead of |
728 | // using the default and/or/xor expansion code which is much longer. |
729 | if (VT == MVT::i32) { |
730 | SDValue LHSlo = DAG.getNode(Opcode: ISD::EXTRACT_ELEMENT, DL, VT: MVT::i16, N1: LHS, |
731 | N2: DAG.getIntPtrConstant(Val: 0, DL)); |
732 | SDValue LHShi = DAG.getNode(Opcode: ISD::EXTRACT_ELEMENT, DL, VT: MVT::i16, N1: LHS, |
733 | N2: DAG.getIntPtrConstant(Val: 1, DL)); |
734 | SDValue RHSlo = DAG.getNode(Opcode: ISD::EXTRACT_ELEMENT, DL, VT: MVT::i16, N1: RHS, |
735 | N2: DAG.getIntPtrConstant(Val: 0, DL)); |
736 | SDValue RHShi = DAG.getNode(Opcode: ISD::EXTRACT_ELEMENT, DL, VT: MVT::i16, N1: RHS, |
737 | N2: DAG.getIntPtrConstant(Val: 1, DL)); |
738 | |
739 | if (UseTest) { |
740 | // When using tst we only care about the highest part. |
741 | SDValue Top = DAG.getNode(Opcode: ISD::EXTRACT_ELEMENT, DL, VT: MVT::i8, N1: LHShi, |
742 | N2: DAG.getIntPtrConstant(Val: 1, DL)); |
743 | Cmp = DAG.getNode(Opcode: AVRISD::TST, DL, VT: MVT::Glue, Operand: Top); |
744 | } else { |
745 | Cmp = getAVRCmp(LHS: LHSlo, RHS: RHSlo, DAG, DL); |
746 | Cmp = DAG.getNode(Opcode: AVRISD::CMPC, DL, VT: MVT::Glue, N1: LHShi, N2: RHShi, N3: Cmp); |
747 | } |
748 | } else if (VT == MVT::i64) { |
749 | SDValue LHS_0 = DAG.getNode(Opcode: ISD::EXTRACT_ELEMENT, DL, VT: MVT::i32, N1: LHS, |
750 | N2: DAG.getIntPtrConstant(Val: 0, DL)); |
751 | SDValue LHS_1 = DAG.getNode(Opcode: ISD::EXTRACT_ELEMENT, DL, VT: MVT::i32, N1: LHS, |
752 | N2: DAG.getIntPtrConstant(Val: 1, DL)); |
753 | |
754 | SDValue LHS0 = DAG.getNode(Opcode: ISD::EXTRACT_ELEMENT, DL, VT: MVT::i16, N1: LHS_0, |
755 | N2: DAG.getIntPtrConstant(Val: 0, DL)); |
756 | SDValue LHS1 = DAG.getNode(Opcode: ISD::EXTRACT_ELEMENT, DL, VT: MVT::i16, N1: LHS_0, |
757 | N2: DAG.getIntPtrConstant(Val: 1, DL)); |
758 | SDValue LHS2 = DAG.getNode(Opcode: ISD::EXTRACT_ELEMENT, DL, VT: MVT::i16, N1: LHS_1, |
759 | N2: DAG.getIntPtrConstant(Val: 0, DL)); |
760 | SDValue LHS3 = DAG.getNode(Opcode: ISD::EXTRACT_ELEMENT, DL, VT: MVT::i16, N1: LHS_1, |
761 | N2: DAG.getIntPtrConstant(Val: 1, DL)); |
762 | |
763 | SDValue RHS_0 = DAG.getNode(Opcode: ISD::EXTRACT_ELEMENT, DL, VT: MVT::i32, N1: RHS, |
764 | N2: DAG.getIntPtrConstant(Val: 0, DL)); |
765 | SDValue RHS_1 = DAG.getNode(Opcode: ISD::EXTRACT_ELEMENT, DL, VT: MVT::i32, N1: RHS, |
766 | N2: DAG.getIntPtrConstant(Val: 1, DL)); |
767 | |
768 | SDValue RHS0 = DAG.getNode(Opcode: ISD::EXTRACT_ELEMENT, DL, VT: MVT::i16, N1: RHS_0, |
769 | N2: DAG.getIntPtrConstant(Val: 0, DL)); |
770 | SDValue RHS1 = DAG.getNode(Opcode: ISD::EXTRACT_ELEMENT, DL, VT: MVT::i16, N1: RHS_0, |
771 | N2: DAG.getIntPtrConstant(Val: 1, DL)); |
772 | SDValue RHS2 = DAG.getNode(Opcode: ISD::EXTRACT_ELEMENT, DL, VT: MVT::i16, N1: RHS_1, |
773 | N2: DAG.getIntPtrConstant(Val: 0, DL)); |
774 | SDValue RHS3 = DAG.getNode(Opcode: ISD::EXTRACT_ELEMENT, DL, VT: MVT::i16, N1: RHS_1, |
775 | N2: DAG.getIntPtrConstant(Val: 1, DL)); |
776 | |
777 | if (UseTest) { |
778 | // When using tst we only care about the highest part. |
779 | SDValue Top = DAG.getNode(Opcode: ISD::EXTRACT_ELEMENT, DL, VT: MVT::i8, N1: LHS3, |
780 | N2: DAG.getIntPtrConstant(Val: 1, DL)); |
781 | Cmp = DAG.getNode(Opcode: AVRISD::TST, DL, VT: MVT::Glue, Operand: Top); |
782 | } else { |
783 | Cmp = getAVRCmp(LHS: LHS0, RHS: RHS0, DAG, DL); |
784 | Cmp = DAG.getNode(Opcode: AVRISD::CMPC, DL, VT: MVT::Glue, N1: LHS1, N2: RHS1, N3: Cmp); |
785 | Cmp = DAG.getNode(Opcode: AVRISD::CMPC, DL, VT: MVT::Glue, N1: LHS2, N2: RHS2, N3: Cmp); |
786 | Cmp = DAG.getNode(Opcode: AVRISD::CMPC, DL, VT: MVT::Glue, N1: LHS3, N2: RHS3, N3: Cmp); |
787 | } |
788 | } else if (VT == MVT::i8 || VT == MVT::i16) { |
789 | if (UseTest) { |
790 | // When using tst we only care about the highest part. |
791 | Cmp = DAG.getNode(Opcode: AVRISD::TST, DL, VT: MVT::Glue, |
792 | Operand: (VT == MVT::i8) |
793 | ? LHS |
794 | : DAG.getNode(Opcode: ISD::EXTRACT_ELEMENT, DL, VT: MVT::i8, |
795 | N1: LHS, N2: DAG.getIntPtrConstant(Val: 1, DL))); |
796 | } else { |
797 | Cmp = getAVRCmp(LHS, RHS, DAG, DL); |
798 | } |
799 | } else { |
800 | llvm_unreachable("Invalid comparison size" ); |
801 | } |
802 | |
803 | // When using a test instruction AVRcc is already set. |
804 | if (!UseTest) { |
805 | AVRcc = DAG.getConstant(Val: intCCToAVRCC(CC), DL, VT: MVT::i8); |
806 | } |
807 | |
808 | return Cmp; |
809 | } |
810 | |
811 | SDValue AVRTargetLowering::LowerBR_CC(SDValue Op, SelectionDAG &DAG) const { |
812 | SDValue Chain = Op.getOperand(i: 0); |
813 | ISD::CondCode CC = cast<CondCodeSDNode>(Val: Op.getOperand(i: 1))->get(); |
814 | SDValue LHS = Op.getOperand(i: 2); |
815 | SDValue RHS = Op.getOperand(i: 3); |
816 | SDValue Dest = Op.getOperand(i: 4); |
817 | SDLoc dl(Op); |
818 | |
819 | SDValue TargetCC; |
820 | SDValue Cmp = getAVRCmp(LHS, RHS, CC, AVRcc&: TargetCC, DAG, DL: dl); |
821 | |
822 | return DAG.getNode(Opcode: AVRISD::BRCOND, DL: dl, VT: MVT::Other, N1: Chain, N2: Dest, N3: TargetCC, |
823 | N4: Cmp); |
824 | } |
825 | |
826 | SDValue AVRTargetLowering::LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const { |
827 | SDValue LHS = Op.getOperand(i: 0); |
828 | SDValue RHS = Op.getOperand(i: 1); |
829 | SDValue TrueV = Op.getOperand(i: 2); |
830 | SDValue FalseV = Op.getOperand(i: 3); |
831 | ISD::CondCode CC = cast<CondCodeSDNode>(Val: Op.getOperand(i: 4))->get(); |
832 | SDLoc dl(Op); |
833 | |
834 | SDValue TargetCC; |
835 | SDValue Cmp = getAVRCmp(LHS, RHS, CC, AVRcc&: TargetCC, DAG, DL: dl); |
836 | |
837 | SDValue Ops[] = {TrueV, FalseV, TargetCC, Cmp}; |
838 | |
839 | return DAG.getNode(Opcode: AVRISD::SELECT_CC, DL: dl, VT: Op.getValueType(), Ops); |
840 | } |
841 | |
842 | SDValue AVRTargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) const { |
843 | SDValue LHS = Op.getOperand(i: 0); |
844 | SDValue RHS = Op.getOperand(i: 1); |
845 | ISD::CondCode CC = cast<CondCodeSDNode>(Val: Op.getOperand(i: 2))->get(); |
846 | SDLoc DL(Op); |
847 | |
848 | SDValue TargetCC; |
849 | SDValue Cmp = getAVRCmp(LHS, RHS, CC, AVRcc&: TargetCC, DAG, DL); |
850 | |
851 | SDValue TrueV = DAG.getConstant(Val: 1, DL, VT: Op.getValueType()); |
852 | SDValue FalseV = DAG.getConstant(Val: 0, DL, VT: Op.getValueType()); |
853 | SDValue Ops[] = {TrueV, FalseV, TargetCC, Cmp}; |
854 | |
855 | return DAG.getNode(Opcode: AVRISD::SELECT_CC, DL, VT: Op.getValueType(), Ops); |
856 | } |
857 | |
858 | SDValue AVRTargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG) const { |
859 | const MachineFunction &MF = DAG.getMachineFunction(); |
860 | const AVRMachineFunctionInfo *AFI = MF.getInfo<AVRMachineFunctionInfo>(); |
861 | const Value *SV = cast<SrcValueSDNode>(Val: Op.getOperand(i: 2))->getValue(); |
862 | auto DL = DAG.getDataLayout(); |
863 | SDLoc dl(Op); |
864 | |
865 | // Vastart just stores the address of the VarArgsFrameIndex slot into the |
866 | // memory location argument. |
867 | SDValue FI = DAG.getFrameIndex(FI: AFI->getVarArgsFrameIndex(), VT: getPointerTy(DL)); |
868 | |
869 | return DAG.getStore(Chain: Op.getOperand(i: 0), dl, Val: FI, Ptr: Op.getOperand(i: 1), |
870 | PtrInfo: MachinePointerInfo(SV)); |
871 | } |
872 | |
873 | // Modify the existing ISD::INLINEASM node to add the implicit zero register. |
874 | SDValue AVRTargetLowering::LowerINLINEASM(SDValue Op, SelectionDAG &DAG) const { |
875 | SDValue ZeroReg = DAG.getRegister(Reg: Subtarget.getZeroRegister(), VT: MVT::i8); |
876 | if (Op.getOperand(i: Op.getNumOperands() - 1) == ZeroReg || |
877 | Op.getOperand(i: Op.getNumOperands() - 2) == ZeroReg) { |
878 | // Zero register has already been added. Don't add it again. |
879 | // If this isn't handled, we get called over and over again. |
880 | return Op; |
881 | } |
882 | |
883 | // Get a list of operands to the new INLINEASM node. This is mostly a copy, |
884 | // with some edits. |
885 | // Add the following operands at the end (but before the glue node, if it's |
886 | // there): |
887 | // - The flags of the implicit zero register operand. |
888 | // - The implicit zero register operand itself. |
889 | SDLoc dl(Op); |
890 | SmallVector<SDValue, 8> Ops; |
891 | SDNode *N = Op.getNode(); |
892 | SDValue Glue; |
893 | for (unsigned I = 0; I < N->getNumOperands(); I++) { |
894 | SDValue Operand = N->getOperand(Num: I); |
895 | if (Operand.getValueType() == MVT::Glue) { |
896 | // The glue operand always needs to be at the end, so we need to treat it |
897 | // specially. |
898 | Glue = Operand; |
899 | } else { |
900 | Ops.push_back(Elt: Operand); |
901 | } |
902 | } |
903 | InlineAsm::Flag Flags(InlineAsm::Kind::RegUse, 1); |
904 | Ops.push_back(Elt: DAG.getTargetConstant(Val: Flags, DL: dl, VT: MVT::i32)); |
905 | Ops.push_back(Elt: ZeroReg); |
906 | if (Glue) { |
907 | Ops.push_back(Elt: Glue); |
908 | } |
909 | |
910 | // Replace the current INLINEASM node with a new one that has the zero |
911 | // register as implicit parameter. |
912 | SDValue New = DAG.getNode(Opcode: N->getOpcode(), DL: dl, VTList: N->getVTList(), Ops); |
913 | DAG.ReplaceAllUsesOfValueWith(From: Op, To: New); |
914 | DAG.ReplaceAllUsesOfValueWith(From: Op.getValue(R: 1), To: New.getValue(R: 1)); |
915 | |
916 | return New; |
917 | } |
918 | |
919 | SDValue AVRTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const { |
920 | switch (Op.getOpcode()) { |
921 | default: |
922 | llvm_unreachable("Don't know how to custom lower this!" ); |
923 | case ISD::SHL: |
924 | case ISD::SRA: |
925 | case ISD::SRL: |
926 | case ISD::ROTL: |
927 | case ISD::ROTR: |
928 | return LowerShifts(Op, DAG); |
929 | case ISD::GlobalAddress: |
930 | return LowerGlobalAddress(Op, DAG); |
931 | case ISD::BlockAddress: |
932 | return LowerBlockAddress(Op, DAG); |
933 | case ISD::BR_CC: |
934 | return LowerBR_CC(Op, DAG); |
935 | case ISD::SELECT_CC: |
936 | return LowerSELECT_CC(Op, DAG); |
937 | case ISD::SETCC: |
938 | return LowerSETCC(Op, DAG); |
939 | case ISD::VASTART: |
940 | return LowerVASTART(Op, DAG); |
941 | case ISD::SDIVREM: |
942 | case ISD::UDIVREM: |
943 | return LowerDivRem(Op, DAG); |
944 | case ISD::INLINEASM: |
945 | return LowerINLINEASM(Op, DAG); |
946 | } |
947 | |
948 | return SDValue(); |
949 | } |
950 | |
951 | /// Replace a node with an illegal result type |
952 | /// with a new node built out of custom code. |
953 | void AVRTargetLowering::ReplaceNodeResults(SDNode *N, |
954 | SmallVectorImpl<SDValue> &Results, |
955 | SelectionDAG &DAG) const { |
956 | SDLoc DL(N); |
957 | |
958 | switch (N->getOpcode()) { |
959 | case ISD::ADD: { |
960 | // Convert add (x, imm) into sub (x, -imm). |
961 | if (const ConstantSDNode *C = dyn_cast<ConstantSDNode>(Val: N->getOperand(Num: 1))) { |
962 | SDValue Sub = DAG.getNode( |
963 | Opcode: ISD::SUB, DL, VT: N->getValueType(ResNo: 0), N1: N->getOperand(Num: 0), |
964 | N2: DAG.getConstant(Val: -C->getAPIntValue(), DL, VT: C->getValueType(ResNo: 0))); |
965 | Results.push_back(Elt: Sub); |
966 | } |
967 | break; |
968 | } |
969 | default: { |
970 | SDValue Res = LowerOperation(Op: SDValue(N, 0), DAG); |
971 | |
972 | for (unsigned I = 0, E = Res->getNumValues(); I != E; ++I) |
973 | Results.push_back(Elt: Res.getValue(R: I)); |
974 | |
975 | break; |
976 | } |
977 | } |
978 | } |
979 | |
980 | /// Return true if the addressing mode represented |
981 | /// by AM is legal for this target, for a load/store of the specified type. |
982 | bool AVRTargetLowering::isLegalAddressingMode(const DataLayout &DL, |
983 | const AddrMode &AM, Type *Ty, |
984 | unsigned AS, |
985 | Instruction *I) const { |
986 | int64_t Offs = AM.BaseOffs; |
987 | |
988 | // Allow absolute addresses. |
989 | if (AM.BaseGV && !AM.HasBaseReg && AM.Scale == 0 && Offs == 0) { |
990 | return true; |
991 | } |
992 | |
993 | // Flash memory instructions only allow zero offsets. |
994 | if (isa<PointerType>(Val: Ty) && AS == AVR::ProgramMemory) { |
995 | return false; |
996 | } |
997 | |
998 | // Allow reg+<6bit> offset. |
999 | if (Offs < 0) |
1000 | Offs = -Offs; |
1001 | if (AM.BaseGV == nullptr && AM.HasBaseReg && AM.Scale == 0 && |
1002 | isUInt<6>(x: Offs)) { |
1003 | return true; |
1004 | } |
1005 | |
1006 | return false; |
1007 | } |
1008 | |
1009 | /// Returns true by value, base pointer and |
1010 | /// offset pointer and addressing mode by reference if the node's address |
1011 | /// can be legally represented as pre-indexed load / store address. |
1012 | bool AVRTargetLowering::getPreIndexedAddressParts(SDNode *N, SDValue &Base, |
1013 | SDValue &Offset, |
1014 | ISD::MemIndexedMode &AM, |
1015 | SelectionDAG &DAG) const { |
1016 | EVT VT; |
1017 | const SDNode *Op; |
1018 | SDLoc DL(N); |
1019 | |
1020 | if (const LoadSDNode *LD = dyn_cast<LoadSDNode>(Val: N)) { |
1021 | VT = LD->getMemoryVT(); |
1022 | Op = LD->getBasePtr().getNode(); |
1023 | if (LD->getExtensionType() != ISD::NON_EXTLOAD) |
1024 | return false; |
1025 | if (AVR::isProgramMemoryAccess(N: LD)) { |
1026 | return false; |
1027 | } |
1028 | } else if (const StoreSDNode *ST = dyn_cast<StoreSDNode>(Val: N)) { |
1029 | VT = ST->getMemoryVT(); |
1030 | Op = ST->getBasePtr().getNode(); |
1031 | if (AVR::isProgramMemoryAccess(N: ST)) { |
1032 | return false; |
1033 | } |
1034 | } else { |
1035 | return false; |
1036 | } |
1037 | |
1038 | if (VT != MVT::i8 && VT != MVT::i16) { |
1039 | return false; |
1040 | } |
1041 | |
1042 | if (Op->getOpcode() != ISD::ADD && Op->getOpcode() != ISD::SUB) { |
1043 | return false; |
1044 | } |
1045 | |
1046 | if (const ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Val: Op->getOperand(Num: 1))) { |
1047 | int RHSC = RHS->getSExtValue(); |
1048 | if (Op->getOpcode() == ISD::SUB) |
1049 | RHSC = -RHSC; |
1050 | |
1051 | if ((VT == MVT::i16 && RHSC != -2) || (VT == MVT::i8 && RHSC != -1)) { |
1052 | return false; |
1053 | } |
1054 | |
1055 | Base = Op->getOperand(Num: 0); |
1056 | Offset = DAG.getSignedConstant(Val: RHSC, DL, VT: MVT::i8); |
1057 | AM = ISD::PRE_DEC; |
1058 | |
1059 | return true; |
1060 | } |
1061 | |
1062 | return false; |
1063 | } |
1064 | |
1065 | /// Returns true by value, base pointer and |
1066 | /// offset pointer and addressing mode by reference if this node can be |
1067 | /// combined with a load / store to form a post-indexed load / store. |
1068 | bool AVRTargetLowering::getPostIndexedAddressParts(SDNode *N, SDNode *Op, |
1069 | SDValue &Base, |
1070 | SDValue &Offset, |
1071 | ISD::MemIndexedMode &AM, |
1072 | SelectionDAG &DAG) const { |
1073 | EVT VT; |
1074 | SDValue Ptr; |
1075 | SDLoc DL(N); |
1076 | |
1077 | if (const LoadSDNode *LD = dyn_cast<LoadSDNode>(Val: N)) { |
1078 | VT = LD->getMemoryVT(); |
1079 | Ptr = LD->getBasePtr(); |
1080 | if (LD->getExtensionType() != ISD::NON_EXTLOAD) |
1081 | return false; |
1082 | } else if (const StoreSDNode *ST = dyn_cast<StoreSDNode>(Val: N)) { |
1083 | VT = ST->getMemoryVT(); |
1084 | Ptr = ST->getBasePtr(); |
1085 | // We can not store to program memory. |
1086 | if (AVR::isProgramMemoryAccess(N: ST)) |
1087 | return false; |
1088 | // Since the high byte need to be stored first, we can not emit |
1089 | // i16 post increment store like: |
1090 | // st X+, r24 |
1091 | // st X+, r25 |
1092 | if (VT == MVT::i16 && !Subtarget.hasLowByteFirst()) |
1093 | return false; |
1094 | } else { |
1095 | return false; |
1096 | } |
1097 | |
1098 | if (VT != MVT::i8 && VT != MVT::i16) { |
1099 | return false; |
1100 | } |
1101 | |
1102 | if (Op->getOpcode() != ISD::ADD && Op->getOpcode() != ISD::SUB) { |
1103 | return false; |
1104 | } |
1105 | |
1106 | if (const ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Val: Op->getOperand(Num: 1))) { |
1107 | int RHSC = RHS->getSExtValue(); |
1108 | if (Op->getOpcode() == ISD::SUB) |
1109 | RHSC = -RHSC; |
1110 | if ((VT == MVT::i16 && RHSC != 2) || (VT == MVT::i8 && RHSC != 1)) { |
1111 | return false; |
1112 | } |
1113 | |
1114 | // FIXME: We temporarily disable post increment load from program memory, |
1115 | // due to bug https://github.com/llvm/llvm-project/issues/59914. |
1116 | if (const LoadSDNode *LD = dyn_cast<LoadSDNode>(Val: N)) |
1117 | if (AVR::isProgramMemoryAccess(N: LD)) |
1118 | return false; |
1119 | |
1120 | Base = Op->getOperand(Num: 0); |
1121 | |
1122 | // Post-indexing updates the base, so it's not a valid transform |
1123 | // if that's not the same as the load's pointer. |
1124 | if (Ptr != Base) |
1125 | return false; |
1126 | |
1127 | Offset = DAG.getConstant(Val: RHSC, DL, VT: MVT::i8); |
1128 | AM = ISD::POST_INC; |
1129 | |
1130 | return true; |
1131 | } |
1132 | |
1133 | return false; |
1134 | } |
1135 | |
1136 | bool AVRTargetLowering::isOffsetFoldingLegal( |
1137 | const GlobalAddressSDNode *GA) const { |
1138 | return true; |
1139 | } |
1140 | |
1141 | //===----------------------------------------------------------------------===// |
1142 | // Formal Arguments Calling Convention Implementation |
1143 | //===----------------------------------------------------------------------===// |
1144 | |
1145 | #include "AVRGenCallingConv.inc" |
1146 | |
1147 | /// Registers for calling conventions, ordered in reverse as required by ABI. |
1148 | /// Both arrays must be of the same length. |
1149 | static const MCPhysReg RegList8AVR[] = { |
1150 | AVR::R25, AVR::R24, AVR::R23, AVR::R22, AVR::R21, AVR::R20, |
1151 | AVR::R19, AVR::R18, AVR::R17, AVR::R16, AVR::R15, AVR::R14, |
1152 | AVR::R13, AVR::R12, AVR::R11, AVR::R10, AVR::R9, AVR::R8}; |
1153 | static const MCPhysReg RegList8Tiny[] = {AVR::R25, AVR::R24, AVR::R23, |
1154 | AVR::R22, AVR::R21, AVR::R20}; |
1155 | static const MCPhysReg RegList16AVR[] = { |
1156 | AVR::R26R25, AVR::R25R24, AVR::R24R23, AVR::R23R22, AVR::R22R21, |
1157 | AVR::R21R20, AVR::R20R19, AVR::R19R18, AVR::R18R17, AVR::R17R16, |
1158 | AVR::R16R15, AVR::R15R14, AVR::R14R13, AVR::R13R12, AVR::R12R11, |
1159 | AVR::R11R10, AVR::R10R9, AVR::R9R8}; |
1160 | static const MCPhysReg RegList16Tiny[] = {AVR::R26R25, AVR::R25R24, |
1161 | AVR::R24R23, AVR::R23R22, |
1162 | AVR::R22R21, AVR::R21R20}; |
1163 | |
1164 | static_assert(std::size(RegList8AVR) == std::size(RegList16AVR), |
1165 | "8-bit and 16-bit register arrays must be of equal length" ); |
1166 | static_assert(std::size(RegList8Tiny) == std::size(RegList16Tiny), |
1167 | "8-bit and 16-bit register arrays must be of equal length" ); |
1168 | |
1169 | /// Analyze incoming and outgoing function arguments. We need custom C++ code |
1170 | /// to handle special constraints in the ABI. |
1171 | /// In addition, all pieces of a certain argument have to be passed either |
1172 | /// using registers or the stack but never mixing both. |
1173 | template <typename ArgT> |
1174 | static void analyzeArguments(TargetLowering::CallLoweringInfo *CLI, |
1175 | const Function *F, const DataLayout *TD, |
1176 | const SmallVectorImpl<ArgT> &Args, |
1177 | SmallVectorImpl<CCValAssign> &ArgLocs, |
1178 | CCState &CCInfo, bool Tiny) { |
1179 | // Choose the proper register list for argument passing according to the ABI. |
1180 | ArrayRef<MCPhysReg> RegList8; |
1181 | ArrayRef<MCPhysReg> RegList16; |
1182 | if (Tiny) { |
1183 | RegList8 = ArrayRef(RegList8Tiny); |
1184 | RegList16 = ArrayRef(RegList16Tiny); |
1185 | } else { |
1186 | RegList8 = ArrayRef(RegList8AVR); |
1187 | RegList16 = ArrayRef(RegList16AVR); |
1188 | } |
1189 | |
1190 | unsigned NumArgs = Args.size(); |
1191 | // This is the index of the last used register, in RegList*. |
1192 | // -1 means R26 (R26 is never actually used in CC). |
1193 | int RegLastIdx = -1; |
1194 | // Once a value is passed to the stack it will always be used |
1195 | bool UseStack = false; |
1196 | for (unsigned i = 0; i != NumArgs;) { |
1197 | MVT VT = Args[i].VT; |
1198 | // We have to count the number of bytes for each function argument, that is |
1199 | // those Args with the same OrigArgIndex. This is important in case the |
1200 | // function takes an aggregate type. |
1201 | // Current argument will be between [i..j). |
1202 | unsigned ArgIndex = Args[i].OrigArgIndex; |
1203 | unsigned TotalBytes = VT.getStoreSize(); |
1204 | unsigned j = i + 1; |
1205 | for (; j != NumArgs; ++j) { |
1206 | if (Args[j].OrigArgIndex != ArgIndex) |
1207 | break; |
1208 | TotalBytes += Args[j].VT.getStoreSize(); |
1209 | } |
1210 | // Round up to even number of bytes. |
1211 | TotalBytes = alignTo(Value: TotalBytes, Align: 2); |
1212 | // Skip zero sized arguments |
1213 | if (TotalBytes == 0) |
1214 | continue; |
1215 | // The index of the first register to be used |
1216 | unsigned RegIdx = RegLastIdx + TotalBytes; |
1217 | RegLastIdx = RegIdx; |
1218 | // If there are not enough registers, use the stack |
1219 | if (RegIdx >= RegList8.size()) { |
1220 | UseStack = true; |
1221 | } |
1222 | for (; i != j; ++i) { |
1223 | MVT VT = Args[i].VT; |
1224 | |
1225 | if (UseStack) { |
1226 | auto evt = EVT(VT).getTypeForEVT(Context&: CCInfo.getContext()); |
1227 | unsigned Offset = CCInfo.AllocateStack(Size: TD->getTypeAllocSize(Ty: evt), |
1228 | Alignment: TD->getABITypeAlign(Ty: evt)); |
1229 | CCInfo.addLoc( |
1230 | V: CCValAssign::getMem(ValNo: i, ValVT: VT, Offset, LocVT: VT, HTP: CCValAssign::Full)); |
1231 | } else { |
1232 | unsigned Reg; |
1233 | if (VT == MVT::i8) { |
1234 | Reg = CCInfo.AllocateReg(Reg: RegList8[RegIdx]); |
1235 | } else if (VT == MVT::i16) { |
1236 | Reg = CCInfo.AllocateReg(Reg: RegList16[RegIdx]); |
1237 | } else { |
1238 | llvm_unreachable( |
1239 | "calling convention can only manage i8 and i16 types" ); |
1240 | } |
1241 | assert(Reg && "register not available in calling convention" ); |
1242 | CCInfo.addLoc(V: CCValAssign::getReg(ValNo: i, ValVT: VT, Reg, LocVT: VT, HTP: CCValAssign::Full)); |
1243 | // Registers inside a particular argument are sorted in increasing order |
1244 | // (remember the array is reversed). |
1245 | RegIdx -= VT.getStoreSize(); |
1246 | } |
1247 | } |
1248 | } |
1249 | } |
1250 | |
1251 | /// Count the total number of bytes needed to pass or return these arguments. |
1252 | template <typename ArgT> |
1253 | static unsigned |
1254 | getTotalArgumentsSizeInBytes(const SmallVectorImpl<ArgT> &Args) { |
1255 | unsigned TotalBytes = 0; |
1256 | |
1257 | for (const ArgT &Arg : Args) { |
1258 | TotalBytes += Arg.VT.getStoreSize(); |
1259 | } |
1260 | return TotalBytes; |
1261 | } |
1262 | |
1263 | /// Analyze incoming and outgoing value of returning from a function. |
1264 | /// The algorithm is similar to analyzeArguments, but there can only be |
1265 | /// one value, possibly an aggregate, and it is limited to 8 bytes. |
1266 | template <typename ArgT> |
1267 | static void analyzeReturnValues(const SmallVectorImpl<ArgT> &Args, |
1268 | CCState &CCInfo, bool Tiny) { |
1269 | unsigned NumArgs = Args.size(); |
1270 | unsigned TotalBytes = getTotalArgumentsSizeInBytes(Args); |
1271 | // CanLowerReturn() guarantees this assertion. |
1272 | if (Tiny) |
1273 | assert(TotalBytes <= 4 && |
1274 | "return values greater than 4 bytes cannot be lowered on AVRTiny" ); |
1275 | else |
1276 | assert(TotalBytes <= 8 && |
1277 | "return values greater than 8 bytes cannot be lowered on AVR" ); |
1278 | |
1279 | // Choose the proper register list for argument passing according to the ABI. |
1280 | ArrayRef<MCPhysReg> RegList8; |
1281 | ArrayRef<MCPhysReg> RegList16; |
1282 | if (Tiny) { |
1283 | RegList8 = ArrayRef(RegList8Tiny); |
1284 | RegList16 = ArrayRef(RegList16Tiny); |
1285 | } else { |
1286 | RegList8 = ArrayRef(RegList8AVR); |
1287 | RegList16 = ArrayRef(RegList16AVR); |
1288 | } |
1289 | |
1290 | // GCC-ABI says that the size is rounded up to the next even number, |
1291 | // but actually once it is more than 4 it will always round up to 8. |
1292 | if (TotalBytes > 4) { |
1293 | TotalBytes = 8; |
1294 | } else { |
1295 | TotalBytes = alignTo(Value: TotalBytes, Align: 2); |
1296 | } |
1297 | |
1298 | // The index of the first register to use. |
1299 | int RegIdx = TotalBytes - 1; |
1300 | for (unsigned i = 0; i != NumArgs; ++i) { |
1301 | MVT VT = Args[i].VT; |
1302 | unsigned Reg; |
1303 | if (VT == MVT::i8) { |
1304 | Reg = CCInfo.AllocateReg(Reg: RegList8[RegIdx]); |
1305 | } else if (VT == MVT::i16) { |
1306 | Reg = CCInfo.AllocateReg(Reg: RegList16[RegIdx]); |
1307 | } else { |
1308 | llvm_unreachable("calling convention can only manage i8 and i16 types" ); |
1309 | } |
1310 | assert(Reg && "register not available in calling convention" ); |
1311 | CCInfo.addLoc(V: CCValAssign::getReg(ValNo: i, ValVT: VT, Reg, LocVT: VT, HTP: CCValAssign::Full)); |
1312 | // Registers sort in increasing order |
1313 | RegIdx -= VT.getStoreSize(); |
1314 | } |
1315 | } |
1316 | |
1317 | SDValue AVRTargetLowering::LowerFormalArguments( |
1318 | SDValue Chain, CallingConv::ID CallConv, bool isVarArg, |
1319 | const SmallVectorImpl<ISD::InputArg> &Ins, const SDLoc &dl, |
1320 | SelectionDAG &DAG, SmallVectorImpl<SDValue> &InVals) const { |
1321 | MachineFunction &MF = DAG.getMachineFunction(); |
1322 | MachineFrameInfo &MFI = MF.getFrameInfo(); |
1323 | auto DL = DAG.getDataLayout(); |
1324 | |
1325 | // Assign locations to all of the incoming arguments. |
1326 | SmallVector<CCValAssign, 16> ArgLocs; |
1327 | CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), ArgLocs, |
1328 | *DAG.getContext()); |
1329 | |
1330 | // Variadic functions do not need all the analysis below. |
1331 | if (isVarArg) { |
1332 | CCInfo.AnalyzeFormalArguments(Ins, Fn: ArgCC_AVR_Vararg); |
1333 | } else { |
1334 | analyzeArguments(CLI: nullptr, F: &MF.getFunction(), TD: &DL, Args: Ins, ArgLocs, CCInfo, |
1335 | Tiny: Subtarget.hasTinyEncoding()); |
1336 | } |
1337 | |
1338 | SDValue ArgValue; |
1339 | for (CCValAssign &VA : ArgLocs) { |
1340 | |
1341 | // Arguments stored on registers. |
1342 | if (VA.isRegLoc()) { |
1343 | EVT RegVT = VA.getLocVT(); |
1344 | const TargetRegisterClass *RC; |
1345 | if (RegVT == MVT::i8) { |
1346 | RC = &AVR::GPR8RegClass; |
1347 | } else if (RegVT == MVT::i16) { |
1348 | RC = &AVR::DREGSRegClass; |
1349 | } else { |
1350 | llvm_unreachable("Unknown argument type!" ); |
1351 | } |
1352 | |
1353 | Register Reg = MF.addLiveIn(PReg: VA.getLocReg(), RC); |
1354 | ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, VT: RegVT); |
1355 | |
1356 | // :NOTE: Clang should not promote any i8 into i16 but for safety the |
1357 | // following code will handle zexts or sexts generated by other |
1358 | // front ends. Otherwise: |
1359 | // If this is an 8 bit value, it is really passed promoted |
1360 | // to 16 bits. Insert an assert[sz]ext to capture this, then |
1361 | // truncate to the right size. |
1362 | switch (VA.getLocInfo()) { |
1363 | default: |
1364 | llvm_unreachable("Unknown loc info!" ); |
1365 | case CCValAssign::Full: |
1366 | break; |
1367 | case CCValAssign::BCvt: |
1368 | ArgValue = DAG.getNode(Opcode: ISD::BITCAST, DL: dl, VT: VA.getValVT(), Operand: ArgValue); |
1369 | break; |
1370 | case CCValAssign::SExt: |
1371 | ArgValue = DAG.getNode(Opcode: ISD::AssertSext, DL: dl, VT: RegVT, N1: ArgValue, |
1372 | N2: DAG.getValueType(VA.getValVT())); |
1373 | ArgValue = DAG.getNode(Opcode: ISD::TRUNCATE, DL: dl, VT: VA.getValVT(), Operand: ArgValue); |
1374 | break; |
1375 | case CCValAssign::ZExt: |
1376 | ArgValue = DAG.getNode(Opcode: ISD::AssertZext, DL: dl, VT: RegVT, N1: ArgValue, |
1377 | N2: DAG.getValueType(VA.getValVT())); |
1378 | ArgValue = DAG.getNode(Opcode: ISD::TRUNCATE, DL: dl, VT: VA.getValVT(), Operand: ArgValue); |
1379 | break; |
1380 | } |
1381 | |
1382 | InVals.push_back(Elt: ArgValue); |
1383 | } else { |
1384 | // Only arguments passed on the stack should make it here. |
1385 | assert(VA.isMemLoc()); |
1386 | |
1387 | EVT LocVT = VA.getLocVT(); |
1388 | |
1389 | // Create the frame index object for this incoming parameter. |
1390 | int FI = MFI.CreateFixedObject(Size: LocVT.getSizeInBits() / 8, |
1391 | SPOffset: VA.getLocMemOffset(), IsImmutable: true); |
1392 | |
1393 | // Create the SelectionDAG nodes corresponding to a load |
1394 | // from this parameter. |
1395 | SDValue FIN = DAG.getFrameIndex(FI, VT: getPointerTy(DL)); |
1396 | InVals.push_back(Elt: DAG.getLoad(VT: LocVT, dl, Chain, Ptr: FIN, |
1397 | PtrInfo: MachinePointerInfo::getFixedStack(MF, FI))); |
1398 | } |
1399 | } |
1400 | |
1401 | // If the function takes variable number of arguments, make a frame index for |
1402 | // the start of the first vararg value... for expansion of llvm.va_start. |
1403 | if (isVarArg) { |
1404 | unsigned StackSize = CCInfo.getStackSize(); |
1405 | AVRMachineFunctionInfo *AFI = MF.getInfo<AVRMachineFunctionInfo>(); |
1406 | |
1407 | AFI->setVarArgsFrameIndex(MFI.CreateFixedObject(Size: 2, SPOffset: StackSize, IsImmutable: true)); |
1408 | } |
1409 | |
1410 | return Chain; |
1411 | } |
1412 | |
1413 | //===----------------------------------------------------------------------===// |
1414 | // Call Calling Convention Implementation |
1415 | //===----------------------------------------------------------------------===// |
1416 | |
1417 | SDValue AVRTargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI, |
1418 | SmallVectorImpl<SDValue> &InVals) const { |
1419 | SelectionDAG &DAG = CLI.DAG; |
1420 | SDLoc &DL = CLI.DL; |
1421 | SmallVectorImpl<ISD::OutputArg> &Outs = CLI.Outs; |
1422 | SmallVectorImpl<SDValue> &OutVals = CLI.OutVals; |
1423 | SmallVectorImpl<ISD::InputArg> &Ins = CLI.Ins; |
1424 | SDValue Chain = CLI.Chain; |
1425 | SDValue Callee = CLI.Callee; |
1426 | bool &isTailCall = CLI.IsTailCall; |
1427 | CallingConv::ID CallConv = CLI.CallConv; |
1428 | bool isVarArg = CLI.IsVarArg; |
1429 | |
1430 | MachineFunction &MF = DAG.getMachineFunction(); |
1431 | |
1432 | // AVR does not yet support tail call optimization. |
1433 | isTailCall = false; |
1434 | |
1435 | // Analyze operands of the call, assigning locations to each operand. |
1436 | SmallVector<CCValAssign, 16> ArgLocs; |
1437 | CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), ArgLocs, |
1438 | *DAG.getContext()); |
1439 | |
1440 | // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every |
1441 | // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol |
1442 | // node so that legalize doesn't hack it. |
1443 | const Function *F = nullptr; |
1444 | if (const GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Val&: Callee)) { |
1445 | const GlobalValue *GV = G->getGlobal(); |
1446 | if (isa<Function>(Val: GV)) |
1447 | F = cast<Function>(Val: GV); |
1448 | Callee = |
1449 | DAG.getTargetGlobalAddress(GV, DL, VT: getPointerTy(DL: DAG.getDataLayout())); |
1450 | } else if (const ExternalSymbolSDNode *ES = |
1451 | dyn_cast<ExternalSymbolSDNode>(Val&: Callee)) { |
1452 | Callee = DAG.getTargetExternalSymbol(Sym: ES->getSymbol(), |
1453 | VT: getPointerTy(DL: DAG.getDataLayout())); |
1454 | } |
1455 | |
1456 | // Variadic functions do not need all the analysis below. |
1457 | if (isVarArg) { |
1458 | CCInfo.AnalyzeCallOperands(Outs, Fn: ArgCC_AVR_Vararg); |
1459 | } else { |
1460 | analyzeArguments(CLI: &CLI, F, TD: &DAG.getDataLayout(), Args: Outs, ArgLocs, CCInfo, |
1461 | Tiny: Subtarget.hasTinyEncoding()); |
1462 | } |
1463 | |
1464 | // Get a count of how many bytes are to be pushed on the stack. |
1465 | unsigned NumBytes = CCInfo.getStackSize(); |
1466 | |
1467 | Chain = DAG.getCALLSEQ_START(Chain, InSize: NumBytes, OutSize: 0, DL); |
1468 | |
1469 | SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass; |
1470 | |
1471 | // First, walk the register assignments, inserting copies. |
1472 | unsigned AI, AE; |
1473 | bool HasStackArgs = false; |
1474 | for (AI = 0, AE = ArgLocs.size(); AI != AE; ++AI) { |
1475 | CCValAssign &VA = ArgLocs[AI]; |
1476 | EVT RegVT = VA.getLocVT(); |
1477 | SDValue Arg = OutVals[AI]; |
1478 | |
1479 | // Promote the value if needed. With Clang this should not happen. |
1480 | switch (VA.getLocInfo()) { |
1481 | default: |
1482 | llvm_unreachable("Unknown loc info!" ); |
1483 | case CCValAssign::Full: |
1484 | break; |
1485 | case CCValAssign::SExt: |
1486 | Arg = DAG.getNode(Opcode: ISD::SIGN_EXTEND, DL, VT: RegVT, Operand: Arg); |
1487 | break; |
1488 | case CCValAssign::ZExt: |
1489 | Arg = DAG.getNode(Opcode: ISD::ZERO_EXTEND, DL, VT: RegVT, Operand: Arg); |
1490 | break; |
1491 | case CCValAssign::AExt: |
1492 | Arg = DAG.getNode(Opcode: ISD::ANY_EXTEND, DL, VT: RegVT, Operand: Arg); |
1493 | break; |
1494 | case CCValAssign::BCvt: |
1495 | Arg = DAG.getNode(Opcode: ISD::BITCAST, DL, VT: RegVT, Operand: Arg); |
1496 | break; |
1497 | } |
1498 | |
1499 | // Stop when we encounter a stack argument, we need to process them |
1500 | // in reverse order in the loop below. |
1501 | if (VA.isMemLoc()) { |
1502 | HasStackArgs = true; |
1503 | break; |
1504 | } |
1505 | |
1506 | // Arguments that can be passed on registers must be kept in the RegsToPass |
1507 | // vector. |
1508 | RegsToPass.push_back(Elt: std::make_pair(x: VA.getLocReg(), y&: Arg)); |
1509 | } |
1510 | |
1511 | // Second, stack arguments have to walked. |
1512 | // Previously this code created chained stores but those chained stores appear |
1513 | // to be unchained in the legalization phase. Therefore, do not attempt to |
1514 | // chain them here. In fact, chaining them here somehow causes the first and |
1515 | // second store to be reversed which is the exact opposite of the intended |
1516 | // effect. |
1517 | if (HasStackArgs) { |
1518 | SmallVector<SDValue, 8> MemOpChains; |
1519 | for (; AI != AE; AI++) { |
1520 | CCValAssign &VA = ArgLocs[AI]; |
1521 | SDValue Arg = OutVals[AI]; |
1522 | |
1523 | assert(VA.isMemLoc()); |
1524 | |
1525 | // SP points to one stack slot further so add one to adjust it. |
1526 | SDValue PtrOff = DAG.getNode( |
1527 | Opcode: ISD::ADD, DL, VT: getPointerTy(DL: DAG.getDataLayout()), |
1528 | N1: DAG.getRegister(Reg: AVR::SP, VT: getPointerTy(DL: DAG.getDataLayout())), |
1529 | N2: DAG.getIntPtrConstant(Val: VA.getLocMemOffset() + 1, DL)); |
1530 | |
1531 | MemOpChains.push_back( |
1532 | Elt: DAG.getStore(Chain, dl: DL, Val: Arg, Ptr: PtrOff, |
1533 | PtrInfo: MachinePointerInfo::getStack(MF, Offset: VA.getLocMemOffset()))); |
1534 | } |
1535 | |
1536 | if (!MemOpChains.empty()) |
1537 | Chain = DAG.getNode(Opcode: ISD::TokenFactor, DL, VT: MVT::Other, Ops: MemOpChains); |
1538 | } |
1539 | |
1540 | // Build a sequence of copy-to-reg nodes chained together with token chain and |
1541 | // flag operands which copy the outgoing args into registers. The InGlue in |
1542 | // necessary since all emited instructions must be stuck together. |
1543 | SDValue InGlue; |
1544 | for (auto Reg : RegsToPass) { |
1545 | Chain = DAG.getCopyToReg(Chain, dl: DL, Reg: Reg.first, N: Reg.second, Glue: InGlue); |
1546 | InGlue = Chain.getValue(R: 1); |
1547 | } |
1548 | |
1549 | // Returns a chain & a flag for retval copy to use. |
1550 | SDVTList NodeTys = DAG.getVTList(VT1: MVT::Other, VT2: MVT::Glue); |
1551 | SmallVector<SDValue, 8> Ops; |
1552 | Ops.push_back(Elt: Chain); |
1553 | Ops.push_back(Elt: Callee); |
1554 | |
1555 | // Add argument registers to the end of the list so that they are known live |
1556 | // into the call. |
1557 | for (auto Reg : RegsToPass) { |
1558 | Ops.push_back(Elt: DAG.getRegister(Reg: Reg.first, VT: Reg.second.getValueType())); |
1559 | } |
1560 | |
1561 | // The zero register (usually R1) must be passed as an implicit register so |
1562 | // that this register is correctly zeroed in interrupts. |
1563 | Ops.push_back(Elt: DAG.getRegister(Reg: Subtarget.getZeroRegister(), VT: MVT::i8)); |
1564 | |
1565 | // Add a register mask operand representing the call-preserved registers. |
1566 | const TargetRegisterInfo *TRI = Subtarget.getRegisterInfo(); |
1567 | const uint32_t *Mask = |
1568 | TRI->getCallPreservedMask(MF: DAG.getMachineFunction(), CallConv); |
1569 | assert(Mask && "Missing call preserved mask for calling convention" ); |
1570 | Ops.push_back(Elt: DAG.getRegisterMask(RegMask: Mask)); |
1571 | |
1572 | if (InGlue.getNode()) { |
1573 | Ops.push_back(Elt: InGlue); |
1574 | } |
1575 | |
1576 | Chain = DAG.getNode(Opcode: AVRISD::CALL, DL, VTList: NodeTys, Ops); |
1577 | InGlue = Chain.getValue(R: 1); |
1578 | |
1579 | // Create the CALLSEQ_END node. |
1580 | Chain = DAG.getCALLSEQ_END(Chain, Size1: NumBytes, Size2: 0, Glue: InGlue, DL); |
1581 | |
1582 | if (!Ins.empty()) { |
1583 | InGlue = Chain.getValue(R: 1); |
1584 | } |
1585 | |
1586 | // Handle result values, copying them out of physregs into vregs that we |
1587 | // return. |
1588 | return LowerCallResult(Chain, InGlue, CallConv, isVarArg, Ins, dl: DL, DAG, |
1589 | InVals); |
1590 | } |
1591 | |
1592 | /// Lower the result values of a call into the |
1593 | /// appropriate copies out of appropriate physical registers. |
1594 | /// |
1595 | SDValue AVRTargetLowering::LowerCallResult( |
1596 | SDValue Chain, SDValue InGlue, CallingConv::ID CallConv, bool isVarArg, |
1597 | const SmallVectorImpl<ISD::InputArg> &Ins, const SDLoc &dl, |
1598 | SelectionDAG &DAG, SmallVectorImpl<SDValue> &InVals) const { |
1599 | |
1600 | // Assign locations to each value returned by this call. |
1601 | SmallVector<CCValAssign, 16> RVLocs; |
1602 | CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), RVLocs, |
1603 | *DAG.getContext()); |
1604 | |
1605 | // Handle runtime calling convs. |
1606 | if (CallConv == CallingConv::AVR_BUILTIN) { |
1607 | CCInfo.AnalyzeCallResult(Ins, Fn: RetCC_AVR_BUILTIN); |
1608 | } else { |
1609 | analyzeReturnValues(Args: Ins, CCInfo, Tiny: Subtarget.hasTinyEncoding()); |
1610 | } |
1611 | |
1612 | // Copy all of the result registers out of their specified physreg. |
1613 | for (CCValAssign const &RVLoc : RVLocs) { |
1614 | Chain = DAG.getCopyFromReg(Chain, dl, Reg: RVLoc.getLocReg(), VT: RVLoc.getValVT(), |
1615 | Glue: InGlue) |
1616 | .getValue(R: 1); |
1617 | InGlue = Chain.getValue(R: 2); |
1618 | InVals.push_back(Elt: Chain.getValue(R: 0)); |
1619 | } |
1620 | |
1621 | return Chain; |
1622 | } |
1623 | |
1624 | //===----------------------------------------------------------------------===// |
1625 | // Return Value Calling Convention Implementation |
1626 | //===----------------------------------------------------------------------===// |
1627 | |
1628 | bool AVRTargetLowering::CanLowerReturn( |
1629 | CallingConv::ID CallConv, MachineFunction &MF, bool isVarArg, |
1630 | const SmallVectorImpl<ISD::OutputArg> &Outs, LLVMContext &Context, |
1631 | const Type *RetTy) const { |
1632 | if (CallConv == CallingConv::AVR_BUILTIN) { |
1633 | SmallVector<CCValAssign, 16> RVLocs; |
1634 | CCState CCInfo(CallConv, isVarArg, MF, RVLocs, Context); |
1635 | return CCInfo.CheckReturn(Outs, Fn: RetCC_AVR_BUILTIN); |
1636 | } |
1637 | |
1638 | unsigned TotalBytes = getTotalArgumentsSizeInBytes(Args: Outs); |
1639 | return TotalBytes <= (unsigned)(Subtarget.hasTinyEncoding() ? 4 : 8); |
1640 | } |
1641 | |
1642 | SDValue |
1643 | AVRTargetLowering::LowerReturn(SDValue Chain, CallingConv::ID CallConv, |
1644 | bool isVarArg, |
1645 | const SmallVectorImpl<ISD::OutputArg> &Outs, |
1646 | const SmallVectorImpl<SDValue> &OutVals, |
1647 | const SDLoc &dl, SelectionDAG &DAG) const { |
1648 | // CCValAssign - represent the assignment of the return value to locations. |
1649 | SmallVector<CCValAssign, 16> RVLocs; |
1650 | |
1651 | // CCState - Info about the registers and stack slot. |
1652 | CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), RVLocs, |
1653 | *DAG.getContext()); |
1654 | |
1655 | MachineFunction &MF = DAG.getMachineFunction(); |
1656 | |
1657 | // Analyze return values. |
1658 | if (CallConv == CallingConv::AVR_BUILTIN) { |
1659 | CCInfo.AnalyzeReturn(Outs, Fn: RetCC_AVR_BUILTIN); |
1660 | } else { |
1661 | analyzeReturnValues(Args: Outs, CCInfo, Tiny: Subtarget.hasTinyEncoding()); |
1662 | } |
1663 | |
1664 | SDValue Glue; |
1665 | SmallVector<SDValue, 4> RetOps(1, Chain); |
1666 | // Copy the result values into the output registers. |
1667 | for (unsigned i = 0, e = RVLocs.size(); i != e; ++i) { |
1668 | CCValAssign &VA = RVLocs[i]; |
1669 | assert(VA.isRegLoc() && "Can only return in registers!" ); |
1670 | |
1671 | Chain = DAG.getCopyToReg(Chain, dl, Reg: VA.getLocReg(), N: OutVals[i], Glue); |
1672 | |
1673 | // Guarantee that all emitted copies are stuck together with flags. |
1674 | Glue = Chain.getValue(R: 1); |
1675 | RetOps.push_back(Elt: DAG.getRegister(Reg: VA.getLocReg(), VT: VA.getLocVT())); |
1676 | } |
1677 | |
1678 | // Don't emit the ret/reti instruction when the naked attribute is present in |
1679 | // the function being compiled. |
1680 | if (MF.getFunction().getAttributes().hasFnAttr(Kind: Attribute::Naked)) { |
1681 | return Chain; |
1682 | } |
1683 | |
1684 | const AVRMachineFunctionInfo *AFI = MF.getInfo<AVRMachineFunctionInfo>(); |
1685 | |
1686 | if (!AFI->isInterruptOrSignalHandler()) { |
1687 | // The return instruction has an implicit zero register operand: it must |
1688 | // contain zero on return. |
1689 | // This is not needed in interrupts however, where the zero register is |
1690 | // handled specially (only pushed/popped when needed). |
1691 | RetOps.push_back(Elt: DAG.getRegister(Reg: Subtarget.getZeroRegister(), VT: MVT::i8)); |
1692 | } |
1693 | |
1694 | unsigned RetOpc = |
1695 | AFI->isInterruptOrSignalHandler() ? AVRISD::RETI_GLUE : AVRISD::RET_GLUE; |
1696 | |
1697 | RetOps[0] = Chain; // Update chain. |
1698 | |
1699 | if (Glue.getNode()) { |
1700 | RetOps.push_back(Elt: Glue); |
1701 | } |
1702 | |
1703 | return DAG.getNode(Opcode: RetOpc, DL: dl, VT: MVT::Other, Ops: RetOps); |
1704 | } |
1705 | |
1706 | //===----------------------------------------------------------------------===// |
1707 | // Custom Inserters |
1708 | //===----------------------------------------------------------------------===// |
1709 | |
1710 | MachineBasicBlock *AVRTargetLowering::insertShift(MachineInstr &MI, |
1711 | MachineBasicBlock *BB, |
1712 | bool Tiny) const { |
1713 | unsigned Opc; |
1714 | const TargetRegisterClass *RC; |
1715 | bool HasRepeatedOperand = false; |
1716 | MachineFunction *F = BB->getParent(); |
1717 | MachineRegisterInfo &RI = F->getRegInfo(); |
1718 | const TargetInstrInfo &TII = *Subtarget.getInstrInfo(); |
1719 | DebugLoc dl = MI.getDebugLoc(); |
1720 | |
1721 | switch (MI.getOpcode()) { |
1722 | default: |
1723 | llvm_unreachable("Invalid shift opcode!" ); |
1724 | case AVR::Lsl8: |
1725 | Opc = AVR::ADDRdRr; // LSL is an alias of ADD Rd, Rd |
1726 | RC = &AVR::GPR8RegClass; |
1727 | HasRepeatedOperand = true; |
1728 | break; |
1729 | case AVR::Lsl16: |
1730 | Opc = AVR::LSLWRd; |
1731 | RC = &AVR::DREGSRegClass; |
1732 | break; |
1733 | case AVR::Asr8: |
1734 | Opc = AVR::ASRRd; |
1735 | RC = &AVR::GPR8RegClass; |
1736 | break; |
1737 | case AVR::Asr16: |
1738 | Opc = AVR::ASRWRd; |
1739 | RC = &AVR::DREGSRegClass; |
1740 | break; |
1741 | case AVR::Lsr8: |
1742 | Opc = AVR::LSRRd; |
1743 | RC = &AVR::GPR8RegClass; |
1744 | break; |
1745 | case AVR::Lsr16: |
1746 | Opc = AVR::LSRWRd; |
1747 | RC = &AVR::DREGSRegClass; |
1748 | break; |
1749 | case AVR::Rol8: |
1750 | Opc = Tiny ? AVR::ROLBRdR17 : AVR::ROLBRdR1; |
1751 | RC = &AVR::GPR8RegClass; |
1752 | break; |
1753 | case AVR::Rol16: |
1754 | Opc = AVR::ROLWRd; |
1755 | RC = &AVR::DREGSRegClass; |
1756 | break; |
1757 | case AVR::Ror8: |
1758 | Opc = AVR::RORBRd; |
1759 | RC = &AVR::GPR8RegClass; |
1760 | break; |
1761 | case AVR::Ror16: |
1762 | Opc = AVR::RORWRd; |
1763 | RC = &AVR::DREGSRegClass; |
1764 | break; |
1765 | } |
1766 | |
1767 | const BasicBlock *LLVM_BB = BB->getBasicBlock(); |
1768 | |
1769 | MachineFunction::iterator I; |
1770 | for (I = BB->getIterator(); I != F->end() && &(*I) != BB; ++I) |
1771 | ; |
1772 | if (I != F->end()) |
1773 | ++I; |
1774 | |
1775 | // Create loop block. |
1776 | MachineBasicBlock *LoopBB = F->CreateMachineBasicBlock(BB: LLVM_BB); |
1777 | MachineBasicBlock *CheckBB = F->CreateMachineBasicBlock(BB: LLVM_BB); |
1778 | MachineBasicBlock *RemBB = F->CreateMachineBasicBlock(BB: LLVM_BB); |
1779 | |
1780 | F->insert(MBBI: I, MBB: LoopBB); |
1781 | F->insert(MBBI: I, MBB: CheckBB); |
1782 | F->insert(MBBI: I, MBB: RemBB); |
1783 | |
1784 | // Update machine-CFG edges by transferring all successors of the current |
1785 | // block to the block containing instructions after shift. |
1786 | RemBB->splice(Where: RemBB->begin(), Other: BB, From: std::next(x: MachineBasicBlock::iterator(MI)), |
1787 | To: BB->end()); |
1788 | RemBB->transferSuccessorsAndUpdatePHIs(FromMBB: BB); |
1789 | |
1790 | // Add edges BB => LoopBB => CheckBB => RemBB, CheckBB => LoopBB. |
1791 | BB->addSuccessor(Succ: CheckBB); |
1792 | LoopBB->addSuccessor(Succ: CheckBB); |
1793 | CheckBB->addSuccessor(Succ: LoopBB); |
1794 | CheckBB->addSuccessor(Succ: RemBB); |
1795 | |
1796 | Register ShiftAmtReg = RI.createVirtualRegister(RegClass: &AVR::GPR8RegClass); |
1797 | Register ShiftAmtReg2 = RI.createVirtualRegister(RegClass: &AVR::GPR8RegClass); |
1798 | Register ShiftReg = RI.createVirtualRegister(RegClass: RC); |
1799 | Register ShiftReg2 = RI.createVirtualRegister(RegClass: RC); |
1800 | Register ShiftAmtSrcReg = MI.getOperand(i: 2).getReg(); |
1801 | Register SrcReg = MI.getOperand(i: 1).getReg(); |
1802 | Register DstReg = MI.getOperand(i: 0).getReg(); |
1803 | |
1804 | // BB: |
1805 | // rjmp CheckBB |
1806 | BuildMI(BB, MIMD: dl, MCID: TII.get(Opcode: AVR::RJMPk)).addMBB(MBB: CheckBB); |
1807 | |
1808 | // LoopBB: |
1809 | // ShiftReg2 = shift ShiftReg |
1810 | auto ShiftMI = BuildMI(BB: LoopBB, MIMD: dl, MCID: TII.get(Opcode: Opc), DestReg: ShiftReg2).addReg(RegNo: ShiftReg); |
1811 | if (HasRepeatedOperand) |
1812 | ShiftMI.addReg(RegNo: ShiftReg); |
1813 | |
1814 | // CheckBB: |
1815 | // ShiftReg = phi [%SrcReg, BB], [%ShiftReg2, LoopBB] |
1816 | // ShiftAmt = phi [%N, BB], [%ShiftAmt2, LoopBB] |
1817 | // DestReg = phi [%SrcReg, BB], [%ShiftReg, LoopBB] |
1818 | // ShiftAmt2 = ShiftAmt - 1; |
1819 | // if (ShiftAmt2 >= 0) goto LoopBB; |
1820 | BuildMI(BB: CheckBB, MIMD: dl, MCID: TII.get(Opcode: AVR::PHI), DestReg: ShiftReg) |
1821 | .addReg(RegNo: SrcReg) |
1822 | .addMBB(MBB: BB) |
1823 | .addReg(RegNo: ShiftReg2) |
1824 | .addMBB(MBB: LoopBB); |
1825 | BuildMI(BB: CheckBB, MIMD: dl, MCID: TII.get(Opcode: AVR::PHI), DestReg: ShiftAmtReg) |
1826 | .addReg(RegNo: ShiftAmtSrcReg) |
1827 | .addMBB(MBB: BB) |
1828 | .addReg(RegNo: ShiftAmtReg2) |
1829 | .addMBB(MBB: LoopBB); |
1830 | BuildMI(BB: CheckBB, MIMD: dl, MCID: TII.get(Opcode: AVR::PHI), DestReg: DstReg) |
1831 | .addReg(RegNo: SrcReg) |
1832 | .addMBB(MBB: BB) |
1833 | .addReg(RegNo: ShiftReg2) |
1834 | .addMBB(MBB: LoopBB); |
1835 | |
1836 | BuildMI(BB: CheckBB, MIMD: dl, MCID: TII.get(Opcode: AVR::DECRd), DestReg: ShiftAmtReg2).addReg(RegNo: ShiftAmtReg); |
1837 | BuildMI(BB: CheckBB, MIMD: dl, MCID: TII.get(Opcode: AVR::BRPLk)).addMBB(MBB: LoopBB); |
1838 | |
1839 | MI.eraseFromParent(); // The pseudo instruction is gone now. |
1840 | return RemBB; |
1841 | } |
1842 | |
1843 | // Do a multibyte AVR shift. Insert shift instructions and put the output |
1844 | // registers in the Regs array. |
1845 | // Because AVR does not have a normal shift instruction (only a single bit shift |
1846 | // instruction), we have to emulate this behavior with other instructions. |
1847 | // It first tries large steps (moving registers around) and then smaller steps |
1848 | // like single bit shifts. |
1849 | // Large shifts actually reduce the number of shifted registers, so the below |
1850 | // algorithms have to work independently of the number of registers that are |
1851 | // shifted. |
1852 | // For more information and background, see this blogpost: |
1853 | // https://aykevl.nl/2021/02/avr-bitshift |
1854 | static void insertMultibyteShift(MachineInstr &MI, MachineBasicBlock *BB, |
1855 | MutableArrayRef<std::pair<Register, int>> Regs, |
1856 | ISD::NodeType Opc, int64_t ShiftAmt) { |
1857 | const TargetInstrInfo &TII = *BB->getParent()->getSubtarget().getInstrInfo(); |
1858 | const AVRSubtarget &STI = BB->getParent()->getSubtarget<AVRSubtarget>(); |
1859 | MachineRegisterInfo &MRI = BB->getParent()->getRegInfo(); |
1860 | const DebugLoc &dl = MI.getDebugLoc(); |
1861 | |
1862 | const bool ShiftLeft = Opc == ISD::SHL; |
1863 | const bool ArithmeticShift = Opc == ISD::SRA; |
1864 | |
1865 | // Zero a register, for use in later operations. |
1866 | Register ZeroReg = MRI.createVirtualRegister(RegClass: &AVR::GPR8RegClass); |
1867 | BuildMI(BB&: *BB, I&: MI, MIMD: dl, MCID: TII.get(Opcode: AVR::COPY), DestReg: ZeroReg) |
1868 | .addReg(RegNo: STI.getZeroRegister()); |
1869 | |
1870 | // Do a shift modulo 6 or 7. This is a bit more complicated than most shifts |
1871 | // and is hard to compose with the rest, so these are special cased. |
1872 | // The basic idea is to shift one or two bits in the opposite direction and |
1873 | // then move registers around to get the correct end result. |
1874 | if (ShiftLeft && (ShiftAmt % 8) >= 6) { |
1875 | // Left shift modulo 6 or 7. |
1876 | |
1877 | // Create a slice of the registers we're going to modify, to ease working |
1878 | // with them. |
1879 | size_t ShiftRegsOffset = ShiftAmt / 8; |
1880 | size_t ShiftRegsSize = Regs.size() - ShiftRegsOffset; |
1881 | MutableArrayRef<std::pair<Register, int>> ShiftRegs = |
1882 | Regs.slice(N: ShiftRegsOffset, M: ShiftRegsSize); |
1883 | |
1884 | // Shift one to the right, keeping the least significant bit as the carry |
1885 | // bit. |
1886 | insertMultibyteShift(MI, BB, Regs: ShiftRegs, Opc: ISD::SRL, ShiftAmt: 1); |
1887 | |
1888 | // Rotate the least significant bit from the carry bit into a new register |
1889 | // (that starts out zero). |
1890 | Register LowByte = MRI.createVirtualRegister(RegClass: &AVR::GPR8RegClass); |
1891 | BuildMI(BB&: *BB, I&: MI, MIMD: dl, MCID: TII.get(Opcode: AVR::RORRd), DestReg: LowByte).addReg(RegNo: ZeroReg); |
1892 | |
1893 | // Shift one more to the right if this is a modulo-6 shift. |
1894 | if (ShiftAmt % 8 == 6) { |
1895 | insertMultibyteShift(MI, BB, Regs: ShiftRegs, Opc: ISD::SRL, ShiftAmt: 1); |
1896 | Register NewLowByte = MRI.createVirtualRegister(RegClass: &AVR::GPR8RegClass); |
1897 | BuildMI(BB&: *BB, I&: MI, MIMD: dl, MCID: TII.get(Opcode: AVR::RORRd), DestReg: NewLowByte).addReg(RegNo: LowByte); |
1898 | LowByte = NewLowByte; |
1899 | } |
1900 | |
1901 | // Move all registers to the left, zeroing the bottom registers as needed. |
1902 | for (size_t I = 0; I < Regs.size(); I++) { |
1903 | int ShiftRegsIdx = I + 1; |
1904 | if (ShiftRegsIdx < (int)ShiftRegs.size()) { |
1905 | Regs[I] = ShiftRegs[ShiftRegsIdx]; |
1906 | } else if (ShiftRegsIdx == (int)ShiftRegs.size()) { |
1907 | Regs[I] = std::pair(LowByte, 0); |
1908 | } else { |
1909 | Regs[I] = std::pair(ZeroReg, 0); |
1910 | } |
1911 | } |
1912 | |
1913 | return; |
1914 | } |
1915 | |
1916 | // Right shift modulo 6 or 7. |
1917 | if (!ShiftLeft && (ShiftAmt % 8) >= 6) { |
1918 | // Create a view on the registers we're going to modify, to ease working |
1919 | // with them. |
1920 | size_t ShiftRegsSize = Regs.size() - (ShiftAmt / 8); |
1921 | MutableArrayRef<std::pair<Register, int>> ShiftRegs = |
1922 | Regs.slice(N: 0, M: ShiftRegsSize); |
1923 | |
1924 | // Shift one to the left. |
1925 | insertMultibyteShift(MI, BB, Regs: ShiftRegs, Opc: ISD::SHL, ShiftAmt: 1); |
1926 | |
1927 | // Sign or zero extend the most significant register into a new register. |
1928 | // The HighByte is the byte that still has one (or two) bits from the |
1929 | // original value. The ExtByte is purely a zero/sign extend byte (all bits |
1930 | // are either 0 or 1). |
1931 | Register HighByte = MRI.createVirtualRegister(RegClass: &AVR::GPR8RegClass); |
1932 | Register ExtByte = 0; |
1933 | if (ArithmeticShift) { |
1934 | // Sign-extend bit that was shifted out last. |
1935 | BuildMI(BB&: *BB, I&: MI, MIMD: dl, MCID: TII.get(Opcode: AVR::SBCRdRr), DestReg: HighByte) |
1936 | .addReg(RegNo: HighByte, flags: RegState::Undef) |
1937 | .addReg(RegNo: HighByte, flags: RegState::Undef); |
1938 | ExtByte = HighByte; |
1939 | // The highest bit of the original value is the same as the zero-extend |
1940 | // byte, so HighByte and ExtByte are the same. |
1941 | } else { |
1942 | // Use the zero register for zero extending. |
1943 | ExtByte = ZeroReg; |
1944 | // Rotate most significant bit into a new register (that starts out zero). |
1945 | BuildMI(BB&: *BB, I&: MI, MIMD: dl, MCID: TII.get(Opcode: AVR::ADCRdRr), DestReg: HighByte) |
1946 | .addReg(RegNo: ExtByte) |
1947 | .addReg(RegNo: ExtByte); |
1948 | } |
1949 | |
1950 | // Shift one more to the left for modulo 6 shifts. |
1951 | if (ShiftAmt % 8 == 6) { |
1952 | insertMultibyteShift(MI, BB, Regs: ShiftRegs, Opc: ISD::SHL, ShiftAmt: 1); |
1953 | // Shift the topmost bit into the HighByte. |
1954 | Register NewExt = MRI.createVirtualRegister(RegClass: &AVR::GPR8RegClass); |
1955 | BuildMI(BB&: *BB, I&: MI, MIMD: dl, MCID: TII.get(Opcode: AVR::ADCRdRr), DestReg: NewExt) |
1956 | .addReg(RegNo: HighByte) |
1957 | .addReg(RegNo: HighByte); |
1958 | HighByte = NewExt; |
1959 | } |
1960 | |
1961 | // Move all to the right, while sign or zero extending. |
1962 | for (int I = Regs.size() - 1; I >= 0; I--) { |
1963 | int ShiftRegsIdx = I - (Regs.size() - ShiftRegs.size()) - 1; |
1964 | if (ShiftRegsIdx >= 0) { |
1965 | Regs[I] = ShiftRegs[ShiftRegsIdx]; |
1966 | } else if (ShiftRegsIdx == -1) { |
1967 | Regs[I] = std::pair(HighByte, 0); |
1968 | } else { |
1969 | Regs[I] = std::pair(ExtByte, 0); |
1970 | } |
1971 | } |
1972 | |
1973 | return; |
1974 | } |
1975 | |
1976 | // For shift amounts of at least one register, simply rename the registers and |
1977 | // zero the bottom registers. |
1978 | while (ShiftLeft && ShiftAmt >= 8) { |
1979 | // Move all registers one to the left. |
1980 | for (size_t I = 0; I < Regs.size() - 1; I++) { |
1981 | Regs[I] = Regs[I + 1]; |
1982 | } |
1983 | |
1984 | // Zero the least significant register. |
1985 | Regs[Regs.size() - 1] = std::pair(ZeroReg, 0); |
1986 | |
1987 | // Continue shifts with the leftover registers. |
1988 | Regs = Regs.drop_back(N: 1); |
1989 | |
1990 | ShiftAmt -= 8; |
1991 | } |
1992 | |
1993 | // And again, the same for right shifts. |
1994 | Register ShrExtendReg = 0; |
1995 | if (!ShiftLeft && ShiftAmt >= 8) { |
1996 | if (ArithmeticShift) { |
1997 | // Sign extend the most significant register into ShrExtendReg. |
1998 | ShrExtendReg = MRI.createVirtualRegister(RegClass: &AVR::GPR8RegClass); |
1999 | Register Tmp = MRI.createVirtualRegister(RegClass: &AVR::GPR8RegClass); |
2000 | BuildMI(BB&: *BB, I&: MI, MIMD: dl, MCID: TII.get(Opcode: AVR::ADDRdRr), DestReg: Tmp) |
2001 | .addReg(RegNo: Regs[0].first, flags: 0, SubReg: Regs[0].second) |
2002 | .addReg(RegNo: Regs[0].first, flags: 0, SubReg: Regs[0].second); |
2003 | BuildMI(BB&: *BB, I&: MI, MIMD: dl, MCID: TII.get(Opcode: AVR::SBCRdRr), DestReg: ShrExtendReg) |
2004 | .addReg(RegNo: Tmp) |
2005 | .addReg(RegNo: Tmp); |
2006 | } else { |
2007 | ShrExtendReg = ZeroReg; |
2008 | } |
2009 | for (; ShiftAmt >= 8; ShiftAmt -= 8) { |
2010 | // Move all registers one to the right. |
2011 | for (size_t I = Regs.size() - 1; I != 0; I--) { |
2012 | Regs[I] = Regs[I - 1]; |
2013 | } |
2014 | |
2015 | // Zero or sign extend the most significant register. |
2016 | Regs[0] = std::pair(ShrExtendReg, 0); |
2017 | |
2018 | // Continue shifts with the leftover registers. |
2019 | Regs = Regs.drop_front(N: 1); |
2020 | } |
2021 | } |
2022 | |
2023 | // The bigger shifts are already handled above. |
2024 | assert((ShiftAmt < 8) && "Unexpect shift amount" ); |
2025 | |
2026 | // Shift by four bits, using a complicated swap/eor/andi/eor sequence. |
2027 | // It only works for logical shifts because the bits shifted in are all |
2028 | // zeroes. |
2029 | // To shift a single byte right, it produces code like this: |
2030 | // swap r0 |
2031 | // andi r0, 0x0f |
2032 | // For a two-byte (16-bit) shift, it adds the following instructions to shift |
2033 | // the upper byte into the lower byte: |
2034 | // swap r1 |
2035 | // eor r0, r1 |
2036 | // andi r1, 0x0f |
2037 | // eor r0, r1 |
2038 | // For bigger shifts, it repeats the above sequence. For example, for a 3-byte |
2039 | // (24-bit) shift it adds: |
2040 | // swap r2 |
2041 | // eor r1, r2 |
2042 | // andi r2, 0x0f |
2043 | // eor r1, r2 |
2044 | if (!ArithmeticShift && ShiftAmt >= 4) { |
2045 | Register Prev = 0; |
2046 | for (size_t I = 0; I < Regs.size(); I++) { |
2047 | size_t Idx = ShiftLeft ? I : Regs.size() - I - 1; |
2048 | Register SwapReg = MRI.createVirtualRegister(RegClass: &AVR::LD8RegClass); |
2049 | BuildMI(BB&: *BB, I&: MI, MIMD: dl, MCID: TII.get(Opcode: AVR::SWAPRd), DestReg: SwapReg) |
2050 | .addReg(RegNo: Regs[Idx].first, flags: 0, SubReg: Regs[Idx].second); |
2051 | if (I != 0) { |
2052 | Register R = MRI.createVirtualRegister(RegClass: &AVR::GPR8RegClass); |
2053 | BuildMI(BB&: *BB, I&: MI, MIMD: dl, MCID: TII.get(Opcode: AVR::EORRdRr), DestReg: R) |
2054 | .addReg(RegNo: Prev) |
2055 | .addReg(RegNo: SwapReg); |
2056 | Prev = R; |
2057 | } |
2058 | Register AndReg = MRI.createVirtualRegister(RegClass: &AVR::LD8RegClass); |
2059 | BuildMI(BB&: *BB, I&: MI, MIMD: dl, MCID: TII.get(Opcode: AVR::ANDIRdK), DestReg: AndReg) |
2060 | .addReg(RegNo: SwapReg) |
2061 | .addImm(Val: ShiftLeft ? 0xf0 : 0x0f); |
2062 | if (I != 0) { |
2063 | Register R = MRI.createVirtualRegister(RegClass: &AVR::GPR8RegClass); |
2064 | BuildMI(BB&: *BB, I&: MI, MIMD: dl, MCID: TII.get(Opcode: AVR::EORRdRr), DestReg: R) |
2065 | .addReg(RegNo: Prev) |
2066 | .addReg(RegNo: AndReg); |
2067 | size_t PrevIdx = ShiftLeft ? Idx - 1 : Idx + 1; |
2068 | Regs[PrevIdx] = std::pair(R, 0); |
2069 | } |
2070 | Prev = AndReg; |
2071 | Regs[Idx] = std::pair(AndReg, 0); |
2072 | } |
2073 | ShiftAmt -= 4; |
2074 | } |
2075 | |
2076 | // Shift by one. This is the fallback that always works, and the shift |
2077 | // operation that is used for 1, 2, and 3 bit shifts. |
2078 | while (ShiftLeft && ShiftAmt) { |
2079 | // Shift one to the left. |
2080 | for (ssize_t I = Regs.size() - 1; I >= 0; I--) { |
2081 | Register Out = MRI.createVirtualRegister(RegClass: &AVR::GPR8RegClass); |
2082 | Register In = Regs[I].first; |
2083 | Register InSubreg = Regs[I].second; |
2084 | if (I == (ssize_t)Regs.size() - 1) { // first iteration |
2085 | BuildMI(BB&: *BB, I&: MI, MIMD: dl, MCID: TII.get(Opcode: AVR::ADDRdRr), DestReg: Out) |
2086 | .addReg(RegNo: In, flags: 0, SubReg: InSubreg) |
2087 | .addReg(RegNo: In, flags: 0, SubReg: InSubreg); |
2088 | } else { |
2089 | BuildMI(BB&: *BB, I&: MI, MIMD: dl, MCID: TII.get(Opcode: AVR::ADCRdRr), DestReg: Out) |
2090 | .addReg(RegNo: In, flags: 0, SubReg: InSubreg) |
2091 | .addReg(RegNo: In, flags: 0, SubReg: InSubreg); |
2092 | } |
2093 | Regs[I] = std::pair(Out, 0); |
2094 | } |
2095 | ShiftAmt--; |
2096 | } |
2097 | while (!ShiftLeft && ShiftAmt) { |
2098 | // Shift one to the right. |
2099 | for (size_t I = 0; I < Regs.size(); I++) { |
2100 | Register Out = MRI.createVirtualRegister(RegClass: &AVR::GPR8RegClass); |
2101 | Register In = Regs[I].first; |
2102 | Register InSubreg = Regs[I].second; |
2103 | if (I == 0) { |
2104 | unsigned Opc = ArithmeticShift ? AVR::ASRRd : AVR::LSRRd; |
2105 | BuildMI(BB&: *BB, I&: MI, MIMD: dl, MCID: TII.get(Opcode: Opc), DestReg: Out).addReg(RegNo: In, flags: 0, SubReg: InSubreg); |
2106 | } else { |
2107 | BuildMI(BB&: *BB, I&: MI, MIMD: dl, MCID: TII.get(Opcode: AVR::RORRd), DestReg: Out).addReg(RegNo: In, flags: 0, SubReg: InSubreg); |
2108 | } |
2109 | Regs[I] = std::pair(Out, 0); |
2110 | } |
2111 | ShiftAmt--; |
2112 | } |
2113 | |
2114 | if (ShiftAmt != 0) { |
2115 | llvm_unreachable("don't know how to shift!" ); // sanity check |
2116 | } |
2117 | } |
2118 | |
2119 | // Do a wide (32-bit) shift. |
2120 | MachineBasicBlock * |
2121 | AVRTargetLowering::insertWideShift(MachineInstr &MI, |
2122 | MachineBasicBlock *BB) const { |
2123 | const TargetInstrInfo &TII = *Subtarget.getInstrInfo(); |
2124 | const DebugLoc &dl = MI.getDebugLoc(); |
2125 | |
2126 | // How much to shift to the right (meaning: a negative number indicates a left |
2127 | // shift). |
2128 | int64_t ShiftAmt = MI.getOperand(i: 4).getImm(); |
2129 | ISD::NodeType Opc; |
2130 | switch (MI.getOpcode()) { |
2131 | case AVR::Lsl32: |
2132 | Opc = ISD::SHL; |
2133 | break; |
2134 | case AVR::Lsr32: |
2135 | Opc = ISD::SRL; |
2136 | break; |
2137 | case AVR::Asr32: |
2138 | Opc = ISD::SRA; |
2139 | break; |
2140 | } |
2141 | |
2142 | // Read the input registers, with the most significant register at index 0. |
2143 | std::array<std::pair<Register, int>, 4> Registers = { |
2144 | std::pair(MI.getOperand(i: 3).getReg(), AVR::sub_hi), |
2145 | std::pair(MI.getOperand(i: 3).getReg(), AVR::sub_lo), |
2146 | std::pair(MI.getOperand(i: 2).getReg(), AVR::sub_hi), |
2147 | std::pair(MI.getOperand(i: 2).getReg(), AVR::sub_lo), |
2148 | }; |
2149 | |
2150 | // Do the shift. The registers are modified in-place. |
2151 | insertMultibyteShift(MI, BB, Regs: Registers, Opc, ShiftAmt); |
2152 | |
2153 | // Combine the 8-bit registers into 16-bit register pairs. |
2154 | // This done either from LSB to MSB or from MSB to LSB, depending on the |
2155 | // shift. It's an optimization so that the register allocator will use the |
2156 | // fewest movs possible (which order we use isn't a correctness issue, just an |
2157 | // optimization issue). |
2158 | // - lsl prefers starting from the most significant byte (2nd case). |
2159 | // - lshr prefers starting from the least significant byte (1st case). |
2160 | // - for ashr it depends on the number of shifted bytes. |
2161 | // Some shift operations still don't get the most optimal mov sequences even |
2162 | // with this distinction. TODO: figure out why and try to fix it (but we're |
2163 | // already equal to or faster than avr-gcc in all cases except ashr 8). |
2164 | if (Opc != ISD::SHL && |
2165 | (Opc != ISD::SRA || (ShiftAmt < 16 || ShiftAmt >= 22))) { |
2166 | // Use the resulting registers starting with the least significant byte. |
2167 | BuildMI(BB&: *BB, I&: MI, MIMD: dl, MCID: TII.get(Opcode: AVR::REG_SEQUENCE), DestReg: MI.getOperand(i: 0).getReg()) |
2168 | .addReg(RegNo: Registers[3].first, flags: 0, SubReg: Registers[3].second) |
2169 | .addImm(Val: AVR::sub_lo) |
2170 | .addReg(RegNo: Registers[2].first, flags: 0, SubReg: Registers[2].second) |
2171 | .addImm(Val: AVR::sub_hi); |
2172 | BuildMI(BB&: *BB, I&: MI, MIMD: dl, MCID: TII.get(Opcode: AVR::REG_SEQUENCE), DestReg: MI.getOperand(i: 1).getReg()) |
2173 | .addReg(RegNo: Registers[1].first, flags: 0, SubReg: Registers[1].second) |
2174 | .addImm(Val: AVR::sub_lo) |
2175 | .addReg(RegNo: Registers[0].first, flags: 0, SubReg: Registers[0].second) |
2176 | .addImm(Val: AVR::sub_hi); |
2177 | } else { |
2178 | // Use the resulting registers starting with the most significant byte. |
2179 | BuildMI(BB&: *BB, I&: MI, MIMD: dl, MCID: TII.get(Opcode: AVR::REG_SEQUENCE), DestReg: MI.getOperand(i: 1).getReg()) |
2180 | .addReg(RegNo: Registers[0].first, flags: 0, SubReg: Registers[0].second) |
2181 | .addImm(Val: AVR::sub_hi) |
2182 | .addReg(RegNo: Registers[1].first, flags: 0, SubReg: Registers[1].second) |
2183 | .addImm(Val: AVR::sub_lo); |
2184 | BuildMI(BB&: *BB, I&: MI, MIMD: dl, MCID: TII.get(Opcode: AVR::REG_SEQUENCE), DestReg: MI.getOperand(i: 0).getReg()) |
2185 | .addReg(RegNo: Registers[2].first, flags: 0, SubReg: Registers[2].second) |
2186 | .addImm(Val: AVR::sub_hi) |
2187 | .addReg(RegNo: Registers[3].first, flags: 0, SubReg: Registers[3].second) |
2188 | .addImm(Val: AVR::sub_lo); |
2189 | } |
2190 | |
2191 | // Remove the pseudo instruction. |
2192 | MI.eraseFromParent(); |
2193 | return BB; |
2194 | } |
2195 | |
2196 | static bool isCopyMulResult(MachineBasicBlock::iterator const &I) { |
2197 | if (I->getOpcode() == AVR::COPY) { |
2198 | Register SrcReg = I->getOperand(i: 1).getReg(); |
2199 | return (SrcReg == AVR::R0 || SrcReg == AVR::R1); |
2200 | } |
2201 | |
2202 | return false; |
2203 | } |
2204 | |
2205 | // The mul instructions wreak havock on our zero_reg R1. We need to clear it |
2206 | // after the result has been evacuated. This is probably not the best way to do |
2207 | // it, but it works for now. |
2208 | MachineBasicBlock *AVRTargetLowering::insertMul(MachineInstr &MI, |
2209 | MachineBasicBlock *BB) const { |
2210 | const TargetInstrInfo &TII = *Subtarget.getInstrInfo(); |
2211 | MachineBasicBlock::iterator I(MI); |
2212 | ++I; // in any case insert *after* the mul instruction |
2213 | if (isCopyMulResult(I)) |
2214 | ++I; |
2215 | if (isCopyMulResult(I)) |
2216 | ++I; |
2217 | BuildMI(BB&: *BB, I, MIMD: MI.getDebugLoc(), MCID: TII.get(Opcode: AVR::EORRdRr), DestReg: AVR::R1) |
2218 | .addReg(RegNo: AVR::R1) |
2219 | .addReg(RegNo: AVR::R1); |
2220 | return BB; |
2221 | } |
2222 | |
2223 | // Insert a read from the zero register. |
2224 | MachineBasicBlock * |
2225 | AVRTargetLowering::insertCopyZero(MachineInstr &MI, |
2226 | MachineBasicBlock *BB) const { |
2227 | const TargetInstrInfo &TII = *Subtarget.getInstrInfo(); |
2228 | MachineBasicBlock::iterator I(MI); |
2229 | BuildMI(BB&: *BB, I, MIMD: MI.getDebugLoc(), MCID: TII.get(Opcode: AVR::COPY)) |
2230 | .add(MO: MI.getOperand(i: 0)) |
2231 | .addReg(RegNo: Subtarget.getZeroRegister()); |
2232 | MI.eraseFromParent(); |
2233 | return BB; |
2234 | } |
2235 | |
2236 | // Lower atomicrmw operation to disable interrupts, do operation, and restore |
2237 | // interrupts. This works because all AVR microcontrollers are single core. |
2238 | MachineBasicBlock *AVRTargetLowering::insertAtomicArithmeticOp( |
2239 | MachineInstr &MI, MachineBasicBlock *BB, unsigned Opcode, int Width) const { |
2240 | MachineRegisterInfo &MRI = BB->getParent()->getRegInfo(); |
2241 | const TargetInstrInfo &TII = *Subtarget.getInstrInfo(); |
2242 | MachineBasicBlock::iterator I(MI); |
2243 | DebugLoc dl = MI.getDebugLoc(); |
2244 | |
2245 | // Example instruction sequence, for an atomic 8-bit add: |
2246 | // ldi r25, 5 |
2247 | // in r0, SREG |
2248 | // cli |
2249 | // ld r24, X |
2250 | // add r25, r24 |
2251 | // st X, r25 |
2252 | // out SREG, r0 |
2253 | |
2254 | const TargetRegisterClass *RC = |
2255 | (Width == 8) ? &AVR::GPR8RegClass : &AVR::DREGSRegClass; |
2256 | unsigned LoadOpcode = (Width == 8) ? AVR::LDRdPtr : AVR::LDWRdPtr; |
2257 | unsigned StoreOpcode = (Width == 8) ? AVR::STPtrRr : AVR::STWPtrRr; |
2258 | |
2259 | // Disable interrupts. |
2260 | BuildMI(BB&: *BB, I, MIMD: dl, MCID: TII.get(Opcode: AVR::INRdA), DestReg: Subtarget.getTmpRegister()) |
2261 | .addImm(Val: Subtarget.getIORegSREG()); |
2262 | BuildMI(BB&: *BB, I, MIMD: dl, MCID: TII.get(Opcode: AVR::BCLRs)).addImm(Val: 7); |
2263 | |
2264 | // Load the original value. |
2265 | BuildMI(BB&: *BB, I, MIMD: dl, MCID: TII.get(Opcode: LoadOpcode), DestReg: MI.getOperand(i: 0).getReg()) |
2266 | .add(MO: MI.getOperand(i: 1)); |
2267 | |
2268 | // Do the arithmetic operation. |
2269 | Register Result = MRI.createVirtualRegister(RegClass: RC); |
2270 | BuildMI(BB&: *BB, I, MIMD: dl, MCID: TII.get(Opcode), DestReg: Result) |
2271 | .addReg(RegNo: MI.getOperand(i: 0).getReg()) |
2272 | .add(MO: MI.getOperand(i: 2)); |
2273 | |
2274 | // Store the result. |
2275 | BuildMI(BB&: *BB, I, MIMD: dl, MCID: TII.get(Opcode: StoreOpcode)) |
2276 | .add(MO: MI.getOperand(i: 1)) |
2277 | .addReg(RegNo: Result); |
2278 | |
2279 | // Restore interrupts. |
2280 | BuildMI(BB&: *BB, I, MIMD: dl, MCID: TII.get(Opcode: AVR::OUTARr)) |
2281 | .addImm(Val: Subtarget.getIORegSREG()) |
2282 | .addReg(RegNo: Subtarget.getTmpRegister()); |
2283 | |
2284 | // Remove the pseudo instruction. |
2285 | MI.eraseFromParent(); |
2286 | return BB; |
2287 | } |
2288 | |
2289 | MachineBasicBlock * |
2290 | AVRTargetLowering::EmitInstrWithCustomInserter(MachineInstr &MI, |
2291 | MachineBasicBlock *MBB) const { |
2292 | int Opc = MI.getOpcode(); |
2293 | const AVRSubtarget &STI = MBB->getParent()->getSubtarget<AVRSubtarget>(); |
2294 | |
2295 | // Pseudo shift instructions with a non constant shift amount are expanded |
2296 | // into a loop. |
2297 | switch (Opc) { |
2298 | case AVR::Lsl8: |
2299 | case AVR::Lsl16: |
2300 | case AVR::Lsr8: |
2301 | case AVR::Lsr16: |
2302 | case AVR::Rol8: |
2303 | case AVR::Rol16: |
2304 | case AVR::Ror8: |
2305 | case AVR::Ror16: |
2306 | case AVR::Asr8: |
2307 | case AVR::Asr16: |
2308 | return insertShift(MI, BB: MBB, Tiny: STI.hasTinyEncoding()); |
2309 | case AVR::Lsl32: |
2310 | case AVR::Lsr32: |
2311 | case AVR::Asr32: |
2312 | return insertWideShift(MI, BB: MBB); |
2313 | case AVR::MULRdRr: |
2314 | case AVR::MULSRdRr: |
2315 | return insertMul(MI, BB: MBB); |
2316 | case AVR::CopyZero: |
2317 | return insertCopyZero(MI, BB: MBB); |
2318 | case AVR::AtomicLoadAdd8: |
2319 | return insertAtomicArithmeticOp(MI, BB: MBB, Opcode: AVR::ADDRdRr, Width: 8); |
2320 | case AVR::AtomicLoadAdd16: |
2321 | return insertAtomicArithmeticOp(MI, BB: MBB, Opcode: AVR::ADDWRdRr, Width: 16); |
2322 | case AVR::AtomicLoadSub8: |
2323 | return insertAtomicArithmeticOp(MI, BB: MBB, Opcode: AVR::SUBRdRr, Width: 8); |
2324 | case AVR::AtomicLoadSub16: |
2325 | return insertAtomicArithmeticOp(MI, BB: MBB, Opcode: AVR::SUBWRdRr, Width: 16); |
2326 | case AVR::AtomicLoadAnd8: |
2327 | return insertAtomicArithmeticOp(MI, BB: MBB, Opcode: AVR::ANDRdRr, Width: 8); |
2328 | case AVR::AtomicLoadAnd16: |
2329 | return insertAtomicArithmeticOp(MI, BB: MBB, Opcode: AVR::ANDWRdRr, Width: 16); |
2330 | case AVR::AtomicLoadOr8: |
2331 | return insertAtomicArithmeticOp(MI, BB: MBB, Opcode: AVR::ORRdRr, Width: 8); |
2332 | case AVR::AtomicLoadOr16: |
2333 | return insertAtomicArithmeticOp(MI, BB: MBB, Opcode: AVR::ORWRdRr, Width: 16); |
2334 | case AVR::AtomicLoadXor8: |
2335 | return insertAtomicArithmeticOp(MI, BB: MBB, Opcode: AVR::EORRdRr, Width: 8); |
2336 | case AVR::AtomicLoadXor16: |
2337 | return insertAtomicArithmeticOp(MI, BB: MBB, Opcode: AVR::EORWRdRr, Width: 16); |
2338 | } |
2339 | |
2340 | assert((Opc == AVR::Select16 || Opc == AVR::Select8) && |
2341 | "Unexpected instr type to insert" ); |
2342 | |
2343 | const AVRInstrInfo &TII = (const AVRInstrInfo &)*MI.getParent() |
2344 | ->getParent() |
2345 | ->getSubtarget() |
2346 | .getInstrInfo(); |
2347 | DebugLoc dl = MI.getDebugLoc(); |
2348 | |
2349 | // To "insert" a SELECT instruction, we insert the diamond |
2350 | // control-flow pattern. The incoming instruction knows the |
2351 | // destination vreg to set, the condition code register to branch |
2352 | // on, the true/false values to select between, and a branch opcode |
2353 | // to use. |
2354 | |
2355 | MachineFunction *MF = MBB->getParent(); |
2356 | const BasicBlock *LLVM_BB = MBB->getBasicBlock(); |
2357 | MachineBasicBlock *FallThrough = MBB->getFallThrough(); |
2358 | |
2359 | // If the current basic block falls through to another basic block, |
2360 | // we must insert an unconditional branch to the fallthrough destination |
2361 | // if we are to insert basic blocks at the prior fallthrough point. |
2362 | if (FallThrough != nullptr) { |
2363 | BuildMI(BB: MBB, MIMD: dl, MCID: TII.get(Opcode: AVR::RJMPk)).addMBB(MBB: FallThrough); |
2364 | } |
2365 | |
2366 | MachineBasicBlock *trueMBB = MF->CreateMachineBasicBlock(BB: LLVM_BB); |
2367 | MachineBasicBlock *falseMBB = MF->CreateMachineBasicBlock(BB: LLVM_BB); |
2368 | |
2369 | MachineFunction::iterator I; |
2370 | for (I = MF->begin(); I != MF->end() && &(*I) != MBB; ++I) |
2371 | ; |
2372 | if (I != MF->end()) |
2373 | ++I; |
2374 | MF->insert(MBBI: I, MBB: trueMBB); |
2375 | MF->insert(MBBI: I, MBB: falseMBB); |
2376 | |
2377 | // Set the call frame size on entry to the new basic blocks. |
2378 | unsigned CallFrameSize = TII.getCallFrameSizeAt(MI); |
2379 | trueMBB->setCallFrameSize(CallFrameSize); |
2380 | falseMBB->setCallFrameSize(CallFrameSize); |
2381 | |
2382 | // Transfer remaining instructions and all successors of the current |
2383 | // block to the block which will contain the Phi node for the |
2384 | // select. |
2385 | trueMBB->splice(Where: trueMBB->begin(), Other: MBB, |
2386 | From: std::next(x: MachineBasicBlock::iterator(MI)), To: MBB->end()); |
2387 | trueMBB->transferSuccessorsAndUpdatePHIs(FromMBB: MBB); |
2388 | |
2389 | AVRCC::CondCodes CC = (AVRCC::CondCodes)MI.getOperand(i: 3).getImm(); |
2390 | BuildMI(BB: MBB, MIMD: dl, MCID: TII.getBrCond(CC)).addMBB(MBB: trueMBB); |
2391 | BuildMI(BB: MBB, MIMD: dl, MCID: TII.get(Opcode: AVR::RJMPk)).addMBB(MBB: falseMBB); |
2392 | MBB->addSuccessor(Succ: falseMBB); |
2393 | MBB->addSuccessor(Succ: trueMBB); |
2394 | |
2395 | // Unconditionally flow back to the true block |
2396 | BuildMI(BB: falseMBB, MIMD: dl, MCID: TII.get(Opcode: AVR::RJMPk)).addMBB(MBB: trueMBB); |
2397 | falseMBB->addSuccessor(Succ: trueMBB); |
2398 | |
2399 | // Set up the Phi node to determine where we came from |
2400 | BuildMI(BB&: *trueMBB, I: trueMBB->begin(), MIMD: dl, MCID: TII.get(Opcode: AVR::PHI), |
2401 | DestReg: MI.getOperand(i: 0).getReg()) |
2402 | .addReg(RegNo: MI.getOperand(i: 1).getReg()) |
2403 | .addMBB(MBB) |
2404 | .addReg(RegNo: MI.getOperand(i: 2).getReg()) |
2405 | .addMBB(MBB: falseMBB); |
2406 | |
2407 | MI.eraseFromParent(); // The pseudo instruction is gone now. |
2408 | return trueMBB; |
2409 | } |
2410 | |
2411 | //===----------------------------------------------------------------------===// |
2412 | // Inline Asm Support |
2413 | //===----------------------------------------------------------------------===// |
2414 | |
2415 | AVRTargetLowering::ConstraintType |
2416 | AVRTargetLowering::getConstraintType(StringRef Constraint) const { |
2417 | if (Constraint.size() == 1) { |
2418 | // See http://www.nongnu.org/avr-libc/user-manual/inline_asm.html |
2419 | switch (Constraint[0]) { |
2420 | default: |
2421 | break; |
2422 | case 'a': // Simple upper registers |
2423 | case 'b': // Base pointer registers pairs |
2424 | case 'd': // Upper register |
2425 | case 'l': // Lower registers |
2426 | case 'e': // Pointer register pairs |
2427 | case 'q': // Stack pointer register |
2428 | case 'r': // Any register |
2429 | case 'w': // Special upper register pairs |
2430 | return C_RegisterClass; |
2431 | case 't': // Temporary register |
2432 | case 'x': |
2433 | case 'X': // Pointer register pair X |
2434 | case 'y': |
2435 | case 'Y': // Pointer register pair Y |
2436 | case 'z': |
2437 | case 'Z': // Pointer register pair Z |
2438 | return C_Register; |
2439 | case 'Q': // A memory address based on Y or Z pointer with displacement. |
2440 | return C_Memory; |
2441 | case 'G': // Floating point constant |
2442 | case 'I': // 6-bit positive integer constant |
2443 | case 'J': // 6-bit negative integer constant |
2444 | case 'K': // Integer constant (Range: 2) |
2445 | case 'L': // Integer constant (Range: 0) |
2446 | case 'M': // 8-bit integer constant |
2447 | case 'N': // Integer constant (Range: -1) |
2448 | case 'O': // Integer constant (Range: 8, 16, 24) |
2449 | case 'P': // Integer constant (Range: 1) |
2450 | case 'R': // Integer constant (Range: -6 to 5)x |
2451 | return C_Immediate; |
2452 | } |
2453 | } |
2454 | |
2455 | return TargetLowering::getConstraintType(Constraint); |
2456 | } |
2457 | |
2458 | InlineAsm::ConstraintCode |
2459 | AVRTargetLowering::getInlineAsmMemConstraint(StringRef ConstraintCode) const { |
2460 | // Not sure if this is actually the right thing to do, but we got to do |
2461 | // *something* [agnat] |
2462 | switch (ConstraintCode[0]) { |
2463 | case 'Q': |
2464 | return InlineAsm::ConstraintCode::Q; |
2465 | } |
2466 | return TargetLowering::getInlineAsmMemConstraint(ConstraintCode); |
2467 | } |
2468 | |
2469 | AVRTargetLowering::ConstraintWeight |
2470 | AVRTargetLowering::getSingleConstraintMatchWeight( |
2471 | AsmOperandInfo &info, const char *constraint) const { |
2472 | ConstraintWeight weight = CW_Invalid; |
2473 | Value *CallOperandVal = info.CallOperandVal; |
2474 | |
2475 | // If we don't have a value, we can't do a match, |
2476 | // but allow it at the lowest weight. |
2477 | // (this behaviour has been copied from the ARM backend) |
2478 | if (!CallOperandVal) { |
2479 | return CW_Default; |
2480 | } |
2481 | |
2482 | // Look at the constraint type. |
2483 | switch (*constraint) { |
2484 | default: |
2485 | weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint); |
2486 | break; |
2487 | case 'd': |
2488 | case 'r': |
2489 | case 'l': |
2490 | weight = CW_Register; |
2491 | break; |
2492 | case 'a': |
2493 | case 'b': |
2494 | case 'e': |
2495 | case 'q': |
2496 | case 't': |
2497 | case 'w': |
2498 | case 'x': |
2499 | case 'X': |
2500 | case 'y': |
2501 | case 'Y': |
2502 | case 'z': |
2503 | case 'Z': |
2504 | weight = CW_SpecificReg; |
2505 | break; |
2506 | case 'G': |
2507 | if (const ConstantFP *C = dyn_cast<ConstantFP>(Val: CallOperandVal)) { |
2508 | if (C->isZero()) { |
2509 | weight = CW_Constant; |
2510 | } |
2511 | } |
2512 | break; |
2513 | case 'I': |
2514 | if (const ConstantInt *C = dyn_cast<ConstantInt>(Val: CallOperandVal)) { |
2515 | if (isUInt<6>(x: C->getZExtValue())) { |
2516 | weight = CW_Constant; |
2517 | } |
2518 | } |
2519 | break; |
2520 | case 'J': |
2521 | if (const ConstantInt *C = dyn_cast<ConstantInt>(Val: CallOperandVal)) { |
2522 | if ((C->getSExtValue() >= -63) && (C->getSExtValue() <= 0)) { |
2523 | weight = CW_Constant; |
2524 | } |
2525 | } |
2526 | break; |
2527 | case 'K': |
2528 | if (const ConstantInt *C = dyn_cast<ConstantInt>(Val: CallOperandVal)) { |
2529 | if (C->getZExtValue() == 2) { |
2530 | weight = CW_Constant; |
2531 | } |
2532 | } |
2533 | break; |
2534 | case 'L': |
2535 | if (const ConstantInt *C = dyn_cast<ConstantInt>(Val: CallOperandVal)) { |
2536 | if (C->getZExtValue() == 0) { |
2537 | weight = CW_Constant; |
2538 | } |
2539 | } |
2540 | break; |
2541 | case 'M': |
2542 | if (const ConstantInt *C = dyn_cast<ConstantInt>(Val: CallOperandVal)) { |
2543 | if (isUInt<8>(x: C->getZExtValue())) { |
2544 | weight = CW_Constant; |
2545 | } |
2546 | } |
2547 | break; |
2548 | case 'N': |
2549 | if (const ConstantInt *C = dyn_cast<ConstantInt>(Val: CallOperandVal)) { |
2550 | if (C->getSExtValue() == -1) { |
2551 | weight = CW_Constant; |
2552 | } |
2553 | } |
2554 | break; |
2555 | case 'O': |
2556 | if (const ConstantInt *C = dyn_cast<ConstantInt>(Val: CallOperandVal)) { |
2557 | if ((C->getZExtValue() == 8) || (C->getZExtValue() == 16) || |
2558 | (C->getZExtValue() == 24)) { |
2559 | weight = CW_Constant; |
2560 | } |
2561 | } |
2562 | break; |
2563 | case 'P': |
2564 | if (const ConstantInt *C = dyn_cast<ConstantInt>(Val: CallOperandVal)) { |
2565 | if (C->getZExtValue() == 1) { |
2566 | weight = CW_Constant; |
2567 | } |
2568 | } |
2569 | break; |
2570 | case 'R': |
2571 | if (const ConstantInt *C = dyn_cast<ConstantInt>(Val: CallOperandVal)) { |
2572 | if ((C->getSExtValue() >= -6) && (C->getSExtValue() <= 5)) { |
2573 | weight = CW_Constant; |
2574 | } |
2575 | } |
2576 | break; |
2577 | case 'Q': |
2578 | weight = CW_Memory; |
2579 | break; |
2580 | } |
2581 | |
2582 | return weight; |
2583 | } |
2584 | |
2585 | std::pair<unsigned, const TargetRegisterClass *> |
2586 | AVRTargetLowering::getRegForInlineAsmConstraint(const TargetRegisterInfo *TRI, |
2587 | StringRef Constraint, |
2588 | MVT VT) const { |
2589 | if (Constraint.size() == 1) { |
2590 | switch (Constraint[0]) { |
2591 | case 'a': // Simple upper registers r16..r23. |
2592 | if (VT == MVT::i8) |
2593 | return std::make_pair(x: 0U, y: &AVR::LD8loRegClass); |
2594 | else if (VT == MVT::i16) |
2595 | return std::make_pair(x: 0U, y: &AVR::DREGSLD8loRegClass); |
2596 | break; |
2597 | case 'b': // Base pointer registers: y, z. |
2598 | if (VT == MVT::i8 || VT == MVT::i16) |
2599 | return std::make_pair(x: 0U, y: &AVR::PTRDISPREGSRegClass); |
2600 | break; |
2601 | case 'd': // Upper registers r16..r31. |
2602 | if (VT == MVT::i8) |
2603 | return std::make_pair(x: 0U, y: &AVR::LD8RegClass); |
2604 | else if (VT == MVT::i16) |
2605 | return std::make_pair(x: 0U, y: &AVR::DLDREGSRegClass); |
2606 | break; |
2607 | case 'l': // Lower registers r0..r15. |
2608 | if (VT == MVT::i8) |
2609 | return std::make_pair(x: 0U, y: &AVR::GPR8loRegClass); |
2610 | else if (VT == MVT::i16) |
2611 | return std::make_pair(x: 0U, y: &AVR::DREGSloRegClass); |
2612 | break; |
2613 | case 'e': // Pointer register pairs: x, y, z. |
2614 | if (VT == MVT::i8 || VT == MVT::i16) |
2615 | return std::make_pair(x: 0U, y: &AVR::PTRREGSRegClass); |
2616 | break; |
2617 | case 'q': // Stack pointer register: SPH:SPL. |
2618 | return std::make_pair(x: 0U, y: &AVR::GPRSPRegClass); |
2619 | case 'r': // Any register: r0..r31. |
2620 | if (VT == MVT::i8) |
2621 | return std::make_pair(x: 0U, y: &AVR::GPR8RegClass); |
2622 | else if (VT == MVT::i16) |
2623 | return std::make_pair(x: 0U, y: &AVR::DREGSRegClass); |
2624 | break; |
2625 | case 't': // Temporary register: r0. |
2626 | if (VT == MVT::i8) |
2627 | return std::make_pair(x: unsigned(Subtarget.getTmpRegister()), |
2628 | y: &AVR::GPR8RegClass); |
2629 | break; |
2630 | case 'w': // Special upper register pairs: r24, r26, r28, r30. |
2631 | if (VT == MVT::i8 || VT == MVT::i16) |
2632 | return std::make_pair(x: 0U, y: &AVR::IWREGSRegClass); |
2633 | break; |
2634 | case 'x': // Pointer register pair X: r27:r26. |
2635 | case 'X': |
2636 | if (VT == MVT::i8 || VT == MVT::i16) |
2637 | return std::make_pair(x: unsigned(AVR::R27R26), y: &AVR::PTRREGSRegClass); |
2638 | break; |
2639 | case 'y': // Pointer register pair Y: r29:r28. |
2640 | case 'Y': |
2641 | if (VT == MVT::i8 || VT == MVT::i16) |
2642 | return std::make_pair(x: unsigned(AVR::R29R28), y: &AVR::PTRREGSRegClass); |
2643 | break; |
2644 | case 'z': // Pointer register pair Z: r31:r30. |
2645 | case 'Z': |
2646 | if (VT == MVT::i8 || VT == MVT::i16) |
2647 | return std::make_pair(x: unsigned(AVR::R31R30), y: &AVR::PTRREGSRegClass); |
2648 | break; |
2649 | default: |
2650 | break; |
2651 | } |
2652 | } |
2653 | |
2654 | return TargetLowering::getRegForInlineAsmConstraint( |
2655 | TRI: Subtarget.getRegisterInfo(), Constraint, VT); |
2656 | } |
2657 | |
2658 | void AVRTargetLowering::LowerAsmOperandForConstraint(SDValue Op, |
2659 | StringRef Constraint, |
2660 | std::vector<SDValue> &Ops, |
2661 | SelectionDAG &DAG) const { |
2662 | SDValue Result; |
2663 | SDLoc DL(Op); |
2664 | EVT Ty = Op.getValueType(); |
2665 | |
2666 | // Currently only support length 1 constraints. |
2667 | if (Constraint.size() != 1) { |
2668 | return; |
2669 | } |
2670 | |
2671 | char ConstraintLetter = Constraint[0]; |
2672 | switch (ConstraintLetter) { |
2673 | default: |
2674 | break; |
2675 | // Deal with integers first: |
2676 | case 'I': |
2677 | case 'J': |
2678 | case 'K': |
2679 | case 'L': |
2680 | case 'M': |
2681 | case 'N': |
2682 | case 'O': |
2683 | case 'P': |
2684 | case 'R': { |
2685 | const ConstantSDNode *C = dyn_cast<ConstantSDNode>(Val&: Op); |
2686 | if (!C) { |
2687 | return; |
2688 | } |
2689 | |
2690 | int64_t CVal64 = C->getSExtValue(); |
2691 | uint64_t CUVal64 = C->getZExtValue(); |
2692 | switch (ConstraintLetter) { |
2693 | case 'I': // 0..63 |
2694 | if (!isUInt<6>(x: CUVal64)) |
2695 | return; |
2696 | Result = DAG.getTargetConstant(Val: CUVal64, DL, VT: Ty); |
2697 | break; |
2698 | case 'J': // -63..0 |
2699 | if (CVal64 < -63 || CVal64 > 0) |
2700 | return; |
2701 | Result = DAG.getTargetConstant(Val: CVal64, DL, VT: Ty); |
2702 | break; |
2703 | case 'K': // 2 |
2704 | if (CUVal64 != 2) |
2705 | return; |
2706 | Result = DAG.getTargetConstant(Val: CUVal64, DL, VT: Ty); |
2707 | break; |
2708 | case 'L': // 0 |
2709 | if (CUVal64 != 0) |
2710 | return; |
2711 | Result = DAG.getTargetConstant(Val: CUVal64, DL, VT: Ty); |
2712 | break; |
2713 | case 'M': // 0..255 |
2714 | if (!isUInt<8>(x: CUVal64)) |
2715 | return; |
2716 | // i8 type may be printed as a negative number, |
2717 | // e.g. 254 would be printed as -2, |
2718 | // so we force it to i16 at least. |
2719 | if (Ty.getSimpleVT() == MVT::i8) { |
2720 | Ty = MVT::i16; |
2721 | } |
2722 | Result = DAG.getTargetConstant(Val: CUVal64, DL, VT: Ty); |
2723 | break; |
2724 | case 'N': // -1 |
2725 | if (CVal64 != -1) |
2726 | return; |
2727 | Result = DAG.getTargetConstant(Val: CVal64, DL, VT: Ty); |
2728 | break; |
2729 | case 'O': // 8, 16, 24 |
2730 | if (CUVal64 != 8 && CUVal64 != 16 && CUVal64 != 24) |
2731 | return; |
2732 | Result = DAG.getTargetConstant(Val: CUVal64, DL, VT: Ty); |
2733 | break; |
2734 | case 'P': // 1 |
2735 | if (CUVal64 != 1) |
2736 | return; |
2737 | Result = DAG.getTargetConstant(Val: CUVal64, DL, VT: Ty); |
2738 | break; |
2739 | case 'R': // -6..5 |
2740 | if (CVal64 < -6 || CVal64 > 5) |
2741 | return; |
2742 | Result = DAG.getTargetConstant(Val: CVal64, DL, VT: Ty); |
2743 | break; |
2744 | } |
2745 | |
2746 | break; |
2747 | } |
2748 | case 'G': |
2749 | const ConstantFPSDNode *FC = dyn_cast<ConstantFPSDNode>(Val&: Op); |
2750 | if (!FC || !FC->isZero()) |
2751 | return; |
2752 | // Soften float to i8 0 |
2753 | Result = DAG.getTargetConstant(Val: 0, DL, VT: MVT::i8); |
2754 | break; |
2755 | } |
2756 | |
2757 | if (Result.getNode()) { |
2758 | Ops.push_back(x: Result); |
2759 | return; |
2760 | } |
2761 | |
2762 | return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG); |
2763 | } |
2764 | |
2765 | Register AVRTargetLowering::getRegisterByName(const char *RegName, LLT VT, |
2766 | const MachineFunction &MF) const { |
2767 | Register Reg; |
2768 | |
2769 | if (VT == LLT::scalar(SizeInBits: 8)) { |
2770 | Reg = StringSwitch<unsigned>(RegName) |
2771 | .Case(S: "r0" , Value: AVR::R0) |
2772 | .Case(S: "r1" , Value: AVR::R1) |
2773 | .Default(Value: 0); |
2774 | } else { |
2775 | Reg = StringSwitch<unsigned>(RegName) |
2776 | .Case(S: "r0" , Value: AVR::R1R0) |
2777 | .Case(S: "sp" , Value: AVR::SP) |
2778 | .Default(Value: 0); |
2779 | } |
2780 | |
2781 | if (Reg) |
2782 | return Reg; |
2783 | |
2784 | report_fatal_error( |
2785 | reason: Twine("Invalid register name \"" + StringRef(RegName) + "\"." )); |
2786 | } |
2787 | |
2788 | } // end of namespace llvm |
2789 | |