[ORC] Add std::tuple support to SimplePackedSerialization.
[llvm-project.git] / llvm / lib / Target / AVR / AVRISelLowering.cpp
blobce50ed0bcfd74413ba4118b48fe8b5e22ffe48bf
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.
12 //===----------------------------------------------------------------------===//
14 #include "AVRISelLowering.h"
16 #include "llvm/ADT/StringSwitch.h"
17 #include "llvm/ADT/STLExtras.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"
27 #include "AVR.h"
28 #include "AVRMachineFunctionInfo.h"
29 #include "AVRSubtarget.h"
30 #include "AVRTargetMachine.h"
31 #include "MCTargetDesc/AVRMCTargetDesc.h"
33 namespace llvm {
35 AVRTargetLowering::AVRTargetLowering(const AVRTargetMachine &TM,
36 const AVRSubtarget &STI)
37 : TargetLowering(TM), Subtarget(STI) {
38 // Set up the register classes.
39 addRegisterClass(MVT::i8, &AVR::GPR8RegClass);
40 addRegisterClass(MVT::i16, &AVR::DREGSRegClass);
42 // Compute derived properties from the register classes.
43 computeRegisterProperties(Subtarget.getRegisterInfo());
45 setBooleanContents(ZeroOrOneBooleanContent);
46 setBooleanVectorContents(ZeroOrOneBooleanContent);
47 setSchedulingPreference(Sched::RegPressure);
48 setStackPointerRegisterToSaveRestore(AVR::SP);
49 setSupportsUnalignedAtomics(true);
51 setOperationAction(ISD::GlobalAddress, MVT::i16, Custom);
52 setOperationAction(ISD::BlockAddress, MVT::i16, Custom);
54 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
55 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
56 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i8, Expand);
57 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i16, Expand);
59 for (MVT VT : MVT::integer_valuetypes()) {
60 for (auto N : {ISD::EXTLOAD, ISD::SEXTLOAD, ISD::ZEXTLOAD}) {
61 setLoadExtAction(N, VT, MVT::i1, Promote);
62 setLoadExtAction(N, VT, MVT::i8, Expand);
66 setTruncStoreAction(MVT::i16, MVT::i8, Expand);
68 for (MVT VT : MVT::integer_valuetypes()) {
69 setOperationAction(ISD::ADDC, VT, Legal);
70 setOperationAction(ISD::SUBC, VT, Legal);
71 setOperationAction(ISD::ADDE, VT, Legal);
72 setOperationAction(ISD::SUBE, VT, Legal);
75 // sub (x, imm) gets canonicalized to add (x, -imm), so for illegal types
76 // revert into a sub since we don't have an add with immediate instruction.
77 setOperationAction(ISD::ADD, MVT::i32, Custom);
78 setOperationAction(ISD::ADD, MVT::i64, Custom);
80 // our shift instructions are only able to shift 1 bit at a time, so handle
81 // this in a custom way.
82 setOperationAction(ISD::SRA, MVT::i8, Custom);
83 setOperationAction(ISD::SHL, MVT::i8, Custom);
84 setOperationAction(ISD::SRL, MVT::i8, Custom);
85 setOperationAction(ISD::SRA, MVT::i16, Custom);
86 setOperationAction(ISD::SHL, MVT::i16, Custom);
87 setOperationAction(ISD::SRL, MVT::i16, Custom);
88 setOperationAction(ISD::SHL_PARTS, MVT::i16, Expand);
89 setOperationAction(ISD::SRA_PARTS, MVT::i16, Expand);
90 setOperationAction(ISD::SRL_PARTS, MVT::i16, Expand);
92 setOperationAction(ISD::ROTL, MVT::i8, Custom);
93 setOperationAction(ISD::ROTL, MVT::i16, Expand);
94 setOperationAction(ISD::ROTR, MVT::i8, Custom);
95 setOperationAction(ISD::ROTR, MVT::i16, Expand);
97 setOperationAction(ISD::BR_CC, MVT::i8, Custom);
98 setOperationAction(ISD::BR_CC, MVT::i16, Custom);
99 setOperationAction(ISD::BR_CC, MVT::i32, Custom);
100 setOperationAction(ISD::BR_CC, MVT::i64, Custom);
101 setOperationAction(ISD::BRCOND, MVT::Other, Expand);
103 setOperationAction(ISD::SELECT_CC, MVT::i8, Custom);
104 setOperationAction(ISD::SELECT_CC, MVT::i16, Custom);
105 setOperationAction(ISD::SELECT_CC, MVT::i32, Expand);
106 setOperationAction(ISD::SELECT_CC, MVT::i64, Expand);
107 setOperationAction(ISD::SETCC, MVT::i8, Custom);
108 setOperationAction(ISD::SETCC, MVT::i16, Custom);
109 setOperationAction(ISD::SETCC, MVT::i32, Custom);
110 setOperationAction(ISD::SETCC, MVT::i64, Custom);
111 setOperationAction(ISD::SELECT, MVT::i8, Expand);
112 setOperationAction(ISD::SELECT, MVT::i16, Expand);
114 setOperationAction(ISD::BSWAP, MVT::i16, Expand);
116 // Add support for postincrement and predecrement load/stores.
117 setIndexedLoadAction(ISD::POST_INC, MVT::i8, Legal);
118 setIndexedLoadAction(ISD::POST_INC, MVT::i16, Legal);
119 setIndexedLoadAction(ISD::PRE_DEC, MVT::i8, Legal);
120 setIndexedLoadAction(ISD::PRE_DEC, MVT::i16, Legal);
121 setIndexedStoreAction(ISD::POST_INC, MVT::i8, Legal);
122 setIndexedStoreAction(ISD::POST_INC, MVT::i16, Legal);
123 setIndexedStoreAction(ISD::PRE_DEC, MVT::i8, Legal);
124 setIndexedStoreAction(ISD::PRE_DEC, MVT::i16, Legal);
126 setOperationAction(ISD::BR_JT, MVT::Other, Expand);
128 setOperationAction(ISD::VASTART, MVT::Other, Custom);
129 setOperationAction(ISD::VAEND, MVT::Other, Expand);
130 setOperationAction(ISD::VAARG, MVT::Other, Expand);
131 setOperationAction(ISD::VACOPY, MVT::Other, Expand);
133 // Atomic operations which must be lowered to rtlib calls
134 for (MVT VT : MVT::integer_valuetypes()) {
135 setOperationAction(ISD::ATOMIC_SWAP, VT, Expand);
136 setOperationAction(ISD::ATOMIC_CMP_SWAP, VT, Expand);
137 setOperationAction(ISD::ATOMIC_LOAD_NAND, VT, Expand);
138 setOperationAction(ISD::ATOMIC_LOAD_MAX, VT, Expand);
139 setOperationAction(ISD::ATOMIC_LOAD_MIN, VT, Expand);
140 setOperationAction(ISD::ATOMIC_LOAD_UMAX, VT, Expand);
141 setOperationAction(ISD::ATOMIC_LOAD_UMIN, VT, Expand);
144 // Division/remainder
145 setOperationAction(ISD::UDIV, MVT::i8, Expand);
146 setOperationAction(ISD::UDIV, MVT::i16, Expand);
147 setOperationAction(ISD::UREM, MVT::i8, Expand);
148 setOperationAction(ISD::UREM, MVT::i16, Expand);
149 setOperationAction(ISD::SDIV, MVT::i8, Expand);
150 setOperationAction(ISD::SDIV, MVT::i16, Expand);
151 setOperationAction(ISD::SREM, MVT::i8, Expand);
152 setOperationAction(ISD::SREM, MVT::i16, Expand);
154 // Make division and modulus custom
155 setOperationAction(ISD::UDIVREM, MVT::i8, Custom);
156 setOperationAction(ISD::UDIVREM, MVT::i16, Custom);
157 setOperationAction(ISD::UDIVREM, MVT::i32, Custom);
158 setOperationAction(ISD::SDIVREM, MVT::i8, Custom);
159 setOperationAction(ISD::SDIVREM, MVT::i16, Custom);
160 setOperationAction(ISD::SDIVREM, MVT::i32, Custom);
162 // Do not use MUL. The AVR instructions are closer to SMUL_LOHI &co.
163 setOperationAction(ISD::MUL, MVT::i8, Expand);
164 setOperationAction(ISD::MUL, MVT::i16, Expand);
166 // Expand 16 bit multiplications.
167 setOperationAction(ISD::SMUL_LOHI, MVT::i16, Expand);
168 setOperationAction(ISD::UMUL_LOHI, MVT::i16, Expand);
170 // Expand multiplications to libcalls when there is
171 // no hardware MUL.
172 if (!Subtarget.supportsMultiplication()) {
173 setOperationAction(ISD::SMUL_LOHI, MVT::i8, Expand);
174 setOperationAction(ISD::UMUL_LOHI, MVT::i8, Expand);
177 for (MVT VT : MVT::integer_valuetypes()) {
178 setOperationAction(ISD::MULHS, VT, Expand);
179 setOperationAction(ISD::MULHU, VT, Expand);
182 for (MVT VT : MVT::integer_valuetypes()) {
183 setOperationAction(ISD::CTPOP, VT, Expand);
184 setOperationAction(ISD::CTLZ, VT, Expand);
185 setOperationAction(ISD::CTTZ, VT, Expand);
188 for (MVT VT : MVT::integer_valuetypes()) {
189 setOperationAction(ISD::SIGN_EXTEND_INREG, VT, Expand);
190 // TODO: The generated code is pretty poor. Investigate using the
191 // same "shift and subtract with carry" trick that we do for
192 // extending 8-bit to 16-bit. This may require infrastructure
193 // improvements in how we treat 16-bit "registers" to be feasible.
196 // Division rtlib functions (not supported), use divmod functions instead
197 setLibcallName(RTLIB::SDIV_I8, nullptr);
198 setLibcallName(RTLIB::SDIV_I16, nullptr);
199 setLibcallName(RTLIB::SDIV_I32, nullptr);
200 setLibcallName(RTLIB::UDIV_I8, nullptr);
201 setLibcallName(RTLIB::UDIV_I16, nullptr);
202 setLibcallName(RTLIB::UDIV_I32, nullptr);
204 // Modulus rtlib functions (not supported), use divmod functions instead
205 setLibcallName(RTLIB::SREM_I8, nullptr);
206 setLibcallName(RTLIB::SREM_I16, nullptr);
207 setLibcallName(RTLIB::SREM_I32, nullptr);
208 setLibcallName(RTLIB::UREM_I8, nullptr);
209 setLibcallName(RTLIB::UREM_I16, nullptr);
210 setLibcallName(RTLIB::UREM_I32, nullptr);
212 // Division and modulus rtlib functions
213 setLibcallName(RTLIB::SDIVREM_I8, "__divmodqi4");
214 setLibcallName(RTLIB::SDIVREM_I16, "__divmodhi4");
215 setLibcallName(RTLIB::SDIVREM_I32, "__divmodsi4");
216 setLibcallName(RTLIB::UDIVREM_I8, "__udivmodqi4");
217 setLibcallName(RTLIB::UDIVREM_I16, "__udivmodhi4");
218 setLibcallName(RTLIB::UDIVREM_I32, "__udivmodsi4");
220 // Several of the runtime library functions use a special calling conv
221 setLibcallCallingConv(RTLIB::SDIVREM_I8, CallingConv::AVR_BUILTIN);
222 setLibcallCallingConv(RTLIB::SDIVREM_I16, CallingConv::AVR_BUILTIN);
223 setLibcallCallingConv(RTLIB::UDIVREM_I8, CallingConv::AVR_BUILTIN);
224 setLibcallCallingConv(RTLIB::UDIVREM_I16, CallingConv::AVR_BUILTIN);
226 // Trigonometric rtlib functions
227 setLibcallName(RTLIB::SIN_F32, "sin");
228 setLibcallName(RTLIB::COS_F32, "cos");
230 setMinFunctionAlignment(Align(2));
231 setMinimumJumpTableEntries(UINT_MAX);
234 const char *AVRTargetLowering::getTargetNodeName(unsigned Opcode) const {
235 #define NODE(name) \
236 case AVRISD::name: \
237 return #name
239 switch (Opcode) {
240 default:
241 return nullptr;
242 NODE(RET_FLAG);
243 NODE(RETI_FLAG);
244 NODE(CALL);
245 NODE(WRAPPER);
246 NODE(LSL);
247 NODE(LSR);
248 NODE(ROL);
249 NODE(ROR);
250 NODE(ASR);
251 NODE(LSLLOOP);
252 NODE(LSRLOOP);
253 NODE(ROLLOOP);
254 NODE(RORLOOP);
255 NODE(ASRLOOP);
256 NODE(BRCOND);
257 NODE(CMP);
258 NODE(CMPC);
259 NODE(TST);
260 NODE(SELECT_CC);
261 #undef NODE
265 EVT AVRTargetLowering::getSetCCResultType(const DataLayout &DL, LLVMContext &,
266 EVT VT) const {
267 assert(!VT.isVector() && "No AVR SetCC type for vectors!");
268 return MVT::i8;
271 SDValue AVRTargetLowering::LowerShifts(SDValue Op, SelectionDAG &DAG) const {
272 //:TODO: this function has to be completely rewritten to produce optimal
273 // code, for now it's producing very long but correct code.
274 unsigned Opc8;
275 const SDNode *N = Op.getNode();
276 EVT VT = Op.getValueType();
277 SDLoc dl(N);
278 assert(isPowerOf2_32(VT.getSizeInBits()) &&
279 "Expected power-of-2 shift amount");
281 // Expand non-constant shifts to loops.
282 if (!isa<ConstantSDNode>(N->getOperand(1))) {
283 switch (Op.getOpcode()) {
284 default:
285 llvm_unreachable("Invalid shift opcode!");
286 case ISD::SHL:
287 return DAG.getNode(AVRISD::LSLLOOP, dl, VT, N->getOperand(0),
288 N->getOperand(1));
289 case ISD::SRL:
290 return DAG.getNode(AVRISD::LSRLOOP, dl, VT, N->getOperand(0),
291 N->getOperand(1));
292 case ISD::ROTL: {
293 SDValue Amt = N->getOperand(1);
294 EVT AmtVT = Amt.getValueType();
295 Amt = DAG.getNode(ISD::AND, dl, AmtVT, Amt,
296 DAG.getConstant(VT.getSizeInBits() - 1, dl, AmtVT));
297 return DAG.getNode(AVRISD::ROLLOOP, dl, VT, N->getOperand(0), Amt);
299 case ISD::ROTR: {
300 SDValue Amt = N->getOperand(1);
301 EVT AmtVT = Amt.getValueType();
302 Amt = DAG.getNode(ISD::AND, dl, AmtVT, Amt,
303 DAG.getConstant(VT.getSizeInBits() - 1, dl, AmtVT));
304 return DAG.getNode(AVRISD::RORLOOP, dl, VT, N->getOperand(0), Amt);
306 case ISD::SRA:
307 return DAG.getNode(AVRISD::ASRLOOP, dl, VT, N->getOperand(0),
308 N->getOperand(1));
312 uint64_t ShiftAmount = cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
313 SDValue Victim = N->getOperand(0);
315 switch (Op.getOpcode()) {
316 case ISD::SRA:
317 Opc8 = AVRISD::ASR;
318 break;
319 case ISD::ROTL:
320 Opc8 = AVRISD::ROL;
321 ShiftAmount = ShiftAmount % VT.getSizeInBits();
322 break;
323 case ISD::ROTR:
324 Opc8 = AVRISD::ROR;
325 ShiftAmount = ShiftAmount % VT.getSizeInBits();
326 break;
327 case ISD::SRL:
328 Opc8 = AVRISD::LSR;
329 break;
330 case ISD::SHL:
331 Opc8 = AVRISD::LSL;
332 break;
333 default:
334 llvm_unreachable("Invalid shift opcode");
337 // Optimize int8/int16 shifts.
338 if (VT.getSizeInBits() == 8) {
339 if (Op.getOpcode() == ISD::SHL && 4 <= ShiftAmount && ShiftAmount < 7) {
340 // Optimize LSL when 4 <= ShiftAmount <= 6.
341 Victim = DAG.getNode(AVRISD::SWAP, dl, VT, Victim);
342 Victim =
343 DAG.getNode(ISD::AND, dl, VT, Victim, DAG.getConstant(0xf0, dl, VT));
344 ShiftAmount -= 4;
345 } else if (Op.getOpcode() == ISD::SRL && 4 <= ShiftAmount &&
346 ShiftAmount < 7) {
347 // Optimize LSR when 4 <= ShiftAmount <= 6.
348 Victim = DAG.getNode(AVRISD::SWAP, dl, VT, Victim);
349 Victim =
350 DAG.getNode(ISD::AND, dl, VT, Victim, DAG.getConstant(0x0f, dl, VT));
351 ShiftAmount -= 4;
352 } else if (Op.getOpcode() == ISD::SHL && ShiftAmount == 7) {
353 // Optimize LSL when ShiftAmount == 7.
354 Victim = DAG.getNode(AVRISD::LSLBN, dl, VT, Victim,
355 DAG.getConstant(7, dl, VT));
356 ShiftAmount = 0;
357 } else if (Op.getOpcode() == ISD::SRL && ShiftAmount == 7) {
358 // Optimize LSR when ShiftAmount == 7.
359 Victim = DAG.getNode(AVRISD::LSRBN, dl, VT, Victim,
360 DAG.getConstant(7, dl, VT));
361 ShiftAmount = 0;
362 } else if (Op.getOpcode() == ISD::SRA && ShiftAmount == 7) {
363 // Optimize ASR when ShiftAmount == 7.
364 Victim = DAG.getNode(AVRISD::ASRBN, dl, VT, Victim,
365 DAG.getConstant(7, dl, VT));
366 ShiftAmount = 0;
368 } else if (VT.getSizeInBits() == 16) {
369 if (4 <= ShiftAmount && ShiftAmount < 8)
370 switch (Op.getOpcode()) {
371 case ISD::SHL:
372 Victim = DAG.getNode(AVRISD::LSLWN, dl, VT, Victim,
373 DAG.getConstant(4, dl, VT));
374 ShiftAmount -= 4;
375 break;
376 case ISD::SRL:
377 Victim = DAG.getNode(AVRISD::LSRWN, dl, VT, Victim,
378 DAG.getConstant(4, dl, VT));
379 ShiftAmount -= 4;
380 break;
381 default:
382 break;
384 else if (8 <= ShiftAmount && ShiftAmount < 12)
385 switch (Op.getOpcode()) {
386 case ISD::SHL:
387 Victim = DAG.getNode(AVRISD::LSLWN, dl, VT, Victim,
388 DAG.getConstant(8, dl, VT));
389 ShiftAmount -= 8;
390 break;
391 case ISD::SRL:
392 Victim = DAG.getNode(AVRISD::LSRWN, dl, VT, Victim,
393 DAG.getConstant(8, dl, VT));
394 ShiftAmount -= 8;
395 break;
396 case ISD::SRA:
397 Victim = DAG.getNode(AVRISD::ASRWN, dl, VT, Victim,
398 DAG.getConstant(8, dl, VT));
399 ShiftAmount -= 8;
400 break;
401 default:
402 break;
404 else if (12 <= ShiftAmount)
405 switch (Op.getOpcode()) {
406 case ISD::SHL:
407 Victim = DAG.getNode(AVRISD::LSLWN, dl, VT, Victim,
408 DAG.getConstant(12, dl, VT));
409 ShiftAmount -= 12;
410 break;
411 case ISD::SRL:
412 Victim = DAG.getNode(AVRISD::LSRWN, dl, VT, Victim,
413 DAG.getConstant(12, dl, VT));
414 ShiftAmount -= 12;
415 break;
416 default:
417 break;
421 while (ShiftAmount--) {
422 Victim = DAG.getNode(Opc8, dl, VT, Victim);
425 return Victim;
428 SDValue AVRTargetLowering::LowerDivRem(SDValue Op, SelectionDAG &DAG) const {
429 unsigned Opcode = Op->getOpcode();
430 assert((Opcode == ISD::SDIVREM || Opcode == ISD::UDIVREM) &&
431 "Invalid opcode for Div/Rem lowering");
432 bool IsSigned = (Opcode == ISD::SDIVREM);
433 EVT VT = Op->getValueType(0);
434 Type *Ty = VT.getTypeForEVT(*DAG.getContext());
436 RTLIB::Libcall LC;
437 switch (VT.getSimpleVT().SimpleTy) {
438 default:
439 llvm_unreachable("Unexpected request for libcall!");
440 case MVT::i8:
441 LC = IsSigned ? RTLIB::SDIVREM_I8 : RTLIB::UDIVREM_I8;
442 break;
443 case MVT::i16:
444 LC = IsSigned ? RTLIB::SDIVREM_I16 : RTLIB::UDIVREM_I16;
445 break;
446 case MVT::i32:
447 LC = IsSigned ? RTLIB::SDIVREM_I32 : RTLIB::UDIVREM_I32;
448 break;
451 SDValue InChain = DAG.getEntryNode();
453 TargetLowering::ArgListTy Args;
454 TargetLowering::ArgListEntry Entry;
455 for (SDValue const &Value : Op->op_values()) {
456 Entry.Node = Value;
457 Entry.Ty = Value.getValueType().getTypeForEVT(*DAG.getContext());
458 Entry.IsSExt = IsSigned;
459 Entry.IsZExt = !IsSigned;
460 Args.push_back(Entry);
463 SDValue Callee = DAG.getExternalSymbol(getLibcallName(LC),
464 getPointerTy(DAG.getDataLayout()));
466 Type *RetTy = (Type *)StructType::get(Ty, Ty);
468 SDLoc dl(Op);
469 TargetLowering::CallLoweringInfo CLI(DAG);
470 CLI.setDebugLoc(dl)
471 .setChain(InChain)
472 .setLibCallee(getLibcallCallingConv(LC), RetTy, Callee, std::move(Args))
473 .setInRegister()
474 .setSExtResult(IsSigned)
475 .setZExtResult(!IsSigned);
477 std::pair<SDValue, SDValue> CallInfo = LowerCallTo(CLI);
478 return CallInfo.first;
481 SDValue AVRTargetLowering::LowerGlobalAddress(SDValue Op,
482 SelectionDAG &DAG) const {
483 auto DL = DAG.getDataLayout();
485 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
486 int64_t Offset = cast<GlobalAddressSDNode>(Op)->getOffset();
488 // Create the TargetGlobalAddress node, folding in the constant offset.
489 SDValue Result =
490 DAG.getTargetGlobalAddress(GV, SDLoc(Op), getPointerTy(DL), Offset);
491 return DAG.getNode(AVRISD::WRAPPER, SDLoc(Op), getPointerTy(DL), Result);
494 SDValue AVRTargetLowering::LowerBlockAddress(SDValue Op,
495 SelectionDAG &DAG) const {
496 auto DL = DAG.getDataLayout();
497 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
499 SDValue Result = DAG.getTargetBlockAddress(BA, getPointerTy(DL));
501 return DAG.getNode(AVRISD::WRAPPER, SDLoc(Op), getPointerTy(DL), Result);
504 /// IntCCToAVRCC - Convert a DAG integer condition code to an AVR CC.
505 static AVRCC::CondCodes intCCToAVRCC(ISD::CondCode CC) {
506 switch (CC) {
507 default:
508 llvm_unreachable("Unknown condition code!");
509 case ISD::SETEQ:
510 return AVRCC::COND_EQ;
511 case ISD::SETNE:
512 return AVRCC::COND_NE;
513 case ISD::SETGE:
514 return AVRCC::COND_GE;
515 case ISD::SETLT:
516 return AVRCC::COND_LT;
517 case ISD::SETUGE:
518 return AVRCC::COND_SH;
519 case ISD::SETULT:
520 return AVRCC::COND_LO;
524 /// Returns appropriate CP/CPI/CPC nodes code for the given 8/16-bit operands.
525 SDValue AVRTargetLowering::getAVRCmp(SDValue LHS, SDValue RHS,
526 SelectionDAG &DAG, SDLoc DL) const {
527 assert((LHS.getSimpleValueType() == RHS.getSimpleValueType()) &&
528 "LHS and RHS have different types");
529 assert(((LHS.getSimpleValueType() == MVT::i16) ||
530 (LHS.getSimpleValueType() == MVT::i8)) && "invalid comparison type");
532 SDValue Cmp;
534 if (LHS.getSimpleValueType() == MVT::i16 && isa<ConstantSDNode>(RHS)) {
535 // Generate a CPI/CPC pair if RHS is a 16-bit constant.
536 SDValue LHSlo = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i8, LHS,
537 DAG.getIntPtrConstant(0, DL));
538 SDValue LHShi = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i8, LHS,
539 DAG.getIntPtrConstant(1, DL));
540 SDValue RHSlo = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i8, RHS,
541 DAG.getIntPtrConstant(0, DL));
542 SDValue RHShi = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i8, RHS,
543 DAG.getIntPtrConstant(1, DL));
544 Cmp = DAG.getNode(AVRISD::CMP, DL, MVT::Glue, LHSlo, RHSlo);
545 Cmp = DAG.getNode(AVRISD::CMPC, DL, MVT::Glue, LHShi, RHShi, Cmp);
546 } else {
547 // Generate ordinary 16-bit comparison.
548 Cmp = DAG.getNode(AVRISD::CMP, DL, MVT::Glue, LHS, RHS);
551 return Cmp;
554 /// Returns appropriate AVR CMP/CMPC nodes and corresponding condition code for
555 /// the given operands.
556 SDValue AVRTargetLowering::getAVRCmp(SDValue LHS, SDValue RHS, ISD::CondCode CC,
557 SDValue &AVRcc, SelectionDAG &DAG,
558 SDLoc DL) const {
559 SDValue Cmp;
560 EVT VT = LHS.getValueType();
561 bool UseTest = false;
563 switch (CC) {
564 default:
565 break;
566 case ISD::SETLE: {
567 // Swap operands and reverse the branching condition.
568 std::swap(LHS, RHS);
569 CC = ISD::SETGE;
570 break;
572 case ISD::SETGT: {
573 if (const ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS)) {
574 switch (C->getSExtValue()) {
575 case -1: {
576 // When doing lhs > -1 use a tst instruction on the top part of lhs
577 // and use brpl instead of using a chain of cp/cpc.
578 UseTest = true;
579 AVRcc = DAG.getConstant(AVRCC::COND_PL, DL, MVT::i8);
580 break;
582 case 0: {
583 // Turn lhs > 0 into 0 < lhs since 0 can be materialized with
584 // __zero_reg__ in lhs.
585 RHS = LHS;
586 LHS = DAG.getConstant(0, DL, VT);
587 CC = ISD::SETLT;
588 break;
590 default: {
591 // Turn lhs < rhs with lhs constant into rhs >= lhs+1, this allows
592 // us to fold the constant into the cmp instruction.
593 RHS = DAG.getConstant(C->getSExtValue() + 1, DL, VT);
594 CC = ISD::SETGE;
595 break;
598 break;
600 // Swap operands and reverse the branching condition.
601 std::swap(LHS, RHS);
602 CC = ISD::SETLT;
603 break;
605 case ISD::SETLT: {
606 if (const ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS)) {
607 switch (C->getSExtValue()) {
608 case 1: {
609 // Turn lhs < 1 into 0 >= lhs since 0 can be materialized with
610 // __zero_reg__ in lhs.
611 RHS = LHS;
612 LHS = DAG.getConstant(0, DL, VT);
613 CC = ISD::SETGE;
614 break;
616 case 0: {
617 // When doing lhs < 0 use a tst instruction on the top part of lhs
618 // and use brmi instead of using a chain of cp/cpc.
619 UseTest = true;
620 AVRcc = DAG.getConstant(AVRCC::COND_MI, DL, MVT::i8);
621 break;
625 break;
627 case ISD::SETULE: {
628 // Swap operands and reverse the branching condition.
629 std::swap(LHS, RHS);
630 CC = ISD::SETUGE;
631 break;
633 case ISD::SETUGT: {
634 // Turn lhs < rhs with lhs constant into rhs >= lhs+1, this allows us to
635 // fold the constant into the cmp instruction.
636 if (const ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS)) {
637 RHS = DAG.getConstant(C->getSExtValue() + 1, DL, VT);
638 CC = ISD::SETUGE;
639 break;
641 // Swap operands and reverse the branching condition.
642 std::swap(LHS, RHS);
643 CC = ISD::SETULT;
644 break;
648 // Expand 32 and 64 bit comparisons with custom CMP and CMPC nodes instead of
649 // using the default and/or/xor expansion code which is much longer.
650 if (VT == MVT::i32) {
651 SDValue LHSlo = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i16, LHS,
652 DAG.getIntPtrConstant(0, DL));
653 SDValue LHShi = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i16, LHS,
654 DAG.getIntPtrConstant(1, DL));
655 SDValue RHSlo = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i16, RHS,
656 DAG.getIntPtrConstant(0, DL));
657 SDValue RHShi = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i16, RHS,
658 DAG.getIntPtrConstant(1, DL));
660 if (UseTest) {
661 // When using tst we only care about the highest part.
662 SDValue Top = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i8, LHShi,
663 DAG.getIntPtrConstant(1, DL));
664 Cmp = DAG.getNode(AVRISD::TST, DL, MVT::Glue, Top);
665 } else {
666 Cmp = getAVRCmp(LHSlo, RHSlo, DAG, DL);
667 Cmp = DAG.getNode(AVRISD::CMPC, DL, MVT::Glue, LHShi, RHShi, Cmp);
669 } else if (VT == MVT::i64) {
670 SDValue LHS_0 = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i32, LHS,
671 DAG.getIntPtrConstant(0, DL));
672 SDValue LHS_1 = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i32, LHS,
673 DAG.getIntPtrConstant(1, DL));
675 SDValue LHS0 = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i16, LHS_0,
676 DAG.getIntPtrConstant(0, DL));
677 SDValue LHS1 = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i16, LHS_0,
678 DAG.getIntPtrConstant(1, DL));
679 SDValue LHS2 = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i16, LHS_1,
680 DAG.getIntPtrConstant(0, DL));
681 SDValue LHS3 = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i16, LHS_1,
682 DAG.getIntPtrConstant(1, DL));
684 SDValue RHS_0 = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i32, RHS,
685 DAG.getIntPtrConstant(0, DL));
686 SDValue RHS_1 = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i32, RHS,
687 DAG.getIntPtrConstant(1, DL));
689 SDValue RHS0 = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i16, RHS_0,
690 DAG.getIntPtrConstant(0, DL));
691 SDValue RHS1 = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i16, RHS_0,
692 DAG.getIntPtrConstant(1, DL));
693 SDValue RHS2 = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i16, RHS_1,
694 DAG.getIntPtrConstant(0, DL));
695 SDValue RHS3 = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i16, RHS_1,
696 DAG.getIntPtrConstant(1, DL));
698 if (UseTest) {
699 // When using tst we only care about the highest part.
700 SDValue Top = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i8, LHS3,
701 DAG.getIntPtrConstant(1, DL));
702 Cmp = DAG.getNode(AVRISD::TST, DL, MVT::Glue, Top);
703 } else {
704 Cmp = getAVRCmp(LHS0, RHS0, DAG, DL);
705 Cmp = DAG.getNode(AVRISD::CMPC, DL, MVT::Glue, LHS1, RHS1, Cmp);
706 Cmp = DAG.getNode(AVRISD::CMPC, DL, MVT::Glue, LHS2, RHS2, Cmp);
707 Cmp = DAG.getNode(AVRISD::CMPC, DL, MVT::Glue, LHS3, RHS3, Cmp);
709 } else if (VT == MVT::i8 || VT == MVT::i16) {
710 if (UseTest) {
711 // When using tst we only care about the highest part.
712 Cmp = DAG.getNode(AVRISD::TST, DL, MVT::Glue,
713 (VT == MVT::i8)
714 ? LHS
715 : DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i8,
716 LHS, DAG.getIntPtrConstant(1, DL)));
717 } else {
718 Cmp = getAVRCmp(LHS, RHS, DAG, DL);
720 } else {
721 llvm_unreachable("Invalid comparison size");
724 // When using a test instruction AVRcc is already set.
725 if (!UseTest) {
726 AVRcc = DAG.getConstant(intCCToAVRCC(CC), DL, MVT::i8);
729 return Cmp;
732 SDValue AVRTargetLowering::LowerBR_CC(SDValue Op, SelectionDAG &DAG) const {
733 SDValue Chain = Op.getOperand(0);
734 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
735 SDValue LHS = Op.getOperand(2);
736 SDValue RHS = Op.getOperand(3);
737 SDValue Dest = Op.getOperand(4);
738 SDLoc dl(Op);
740 SDValue TargetCC;
741 SDValue Cmp = getAVRCmp(LHS, RHS, CC, TargetCC, DAG, dl);
743 return DAG.getNode(AVRISD::BRCOND, dl, MVT::Other, Chain, Dest, TargetCC,
744 Cmp);
747 SDValue AVRTargetLowering::LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const {
748 SDValue LHS = Op.getOperand(0);
749 SDValue RHS = Op.getOperand(1);
750 SDValue TrueV = Op.getOperand(2);
751 SDValue FalseV = Op.getOperand(3);
752 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
753 SDLoc dl(Op);
755 SDValue TargetCC;
756 SDValue Cmp = getAVRCmp(LHS, RHS, CC, TargetCC, DAG, dl);
758 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::Glue);
759 SDValue Ops[] = {TrueV, FalseV, TargetCC, Cmp};
761 return DAG.getNode(AVRISD::SELECT_CC, dl, VTs, Ops);
764 SDValue AVRTargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) const {
765 SDValue LHS = Op.getOperand(0);
766 SDValue RHS = Op.getOperand(1);
767 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
768 SDLoc DL(Op);
770 SDValue TargetCC;
771 SDValue Cmp = getAVRCmp(LHS, RHS, CC, TargetCC, DAG, DL);
773 SDValue TrueV = DAG.getConstant(1, DL, Op.getValueType());
774 SDValue FalseV = DAG.getConstant(0, DL, Op.getValueType());
775 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::Glue);
776 SDValue Ops[] = {TrueV, FalseV, TargetCC, Cmp};
778 return DAG.getNode(AVRISD::SELECT_CC, DL, VTs, Ops);
781 SDValue AVRTargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG) const {
782 const MachineFunction &MF = DAG.getMachineFunction();
783 const AVRMachineFunctionInfo *AFI = MF.getInfo<AVRMachineFunctionInfo>();
784 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
785 auto DL = DAG.getDataLayout();
786 SDLoc dl(Op);
788 // Vastart just stores the address of the VarArgsFrameIndex slot into the
789 // memory location argument.
790 SDValue FI = DAG.getFrameIndex(AFI->getVarArgsFrameIndex(), getPointerTy(DL));
792 return DAG.getStore(Op.getOperand(0), dl, FI, Op.getOperand(1),
793 MachinePointerInfo(SV));
796 SDValue AVRTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
797 switch (Op.getOpcode()) {
798 default:
799 llvm_unreachable("Don't know how to custom lower this!");
800 case ISD::SHL:
801 case ISD::SRA:
802 case ISD::SRL:
803 case ISD::ROTL:
804 case ISD::ROTR:
805 return LowerShifts(Op, DAG);
806 case ISD::GlobalAddress:
807 return LowerGlobalAddress(Op, DAG);
808 case ISD::BlockAddress:
809 return LowerBlockAddress(Op, DAG);
810 case ISD::BR_CC:
811 return LowerBR_CC(Op, DAG);
812 case ISD::SELECT_CC:
813 return LowerSELECT_CC(Op, DAG);
814 case ISD::SETCC:
815 return LowerSETCC(Op, DAG);
816 case ISD::VASTART:
817 return LowerVASTART(Op, DAG);
818 case ISD::SDIVREM:
819 case ISD::UDIVREM:
820 return LowerDivRem(Op, DAG);
823 return SDValue();
826 /// Replace a node with an illegal result type
827 /// with a new node built out of custom code.
828 void AVRTargetLowering::ReplaceNodeResults(SDNode *N,
829 SmallVectorImpl<SDValue> &Results,
830 SelectionDAG &DAG) const {
831 SDLoc DL(N);
833 switch (N->getOpcode()) {
834 case ISD::ADD: {
835 // Convert add (x, imm) into sub (x, -imm).
836 if (const ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1))) {
837 SDValue Sub = DAG.getNode(
838 ISD::SUB, DL, N->getValueType(0), N->getOperand(0),
839 DAG.getConstant(-C->getAPIntValue(), DL, C->getValueType(0)));
840 Results.push_back(Sub);
842 break;
844 default: {
845 SDValue Res = LowerOperation(SDValue(N, 0), DAG);
847 for (unsigned I = 0, E = Res->getNumValues(); I != E; ++I)
848 Results.push_back(Res.getValue(I));
850 break;
855 /// Return true if the addressing mode represented
856 /// by AM is legal for this target, for a load/store of the specified type.
857 bool AVRTargetLowering::isLegalAddressingMode(const DataLayout &DL,
858 const AddrMode &AM, Type *Ty,
859 unsigned AS, Instruction *I) const {
860 int64_t Offs = AM.BaseOffs;
862 // Allow absolute addresses.
863 if (AM.BaseGV && !AM.HasBaseReg && AM.Scale == 0 && Offs == 0) {
864 return true;
867 // Flash memory instructions only allow zero offsets.
868 if (isa<PointerType>(Ty) && AS == AVR::ProgramMemory) {
869 return false;
872 // Allow reg+<6bit> offset.
873 if (Offs < 0)
874 Offs = -Offs;
875 if (AM.BaseGV == 0 && AM.HasBaseReg && AM.Scale == 0 && isUInt<6>(Offs)) {
876 return true;
879 return false;
882 /// Returns true by value, base pointer and
883 /// offset pointer and addressing mode by reference if the node's address
884 /// can be legally represented as pre-indexed load / store address.
885 bool AVRTargetLowering::getPreIndexedAddressParts(SDNode *N, SDValue &Base,
886 SDValue &Offset,
887 ISD::MemIndexedMode &AM,
888 SelectionDAG &DAG) const {
889 EVT VT;
890 const SDNode *Op;
891 SDLoc DL(N);
893 if (const LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
894 VT = LD->getMemoryVT();
895 Op = LD->getBasePtr().getNode();
896 if (LD->getExtensionType() != ISD::NON_EXTLOAD)
897 return false;
898 if (AVR::isProgramMemoryAccess(LD)) {
899 return false;
901 } else if (const StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
902 VT = ST->getMemoryVT();
903 Op = ST->getBasePtr().getNode();
904 if (AVR::isProgramMemoryAccess(ST)) {
905 return false;
907 } else {
908 return false;
911 if (VT != MVT::i8 && VT != MVT::i16) {
912 return false;
915 if (Op->getOpcode() != ISD::ADD && Op->getOpcode() != ISD::SUB) {
916 return false;
919 if (const ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Op->getOperand(1))) {
920 int RHSC = RHS->getSExtValue();
921 if (Op->getOpcode() == ISD::SUB)
922 RHSC = -RHSC;
924 if ((VT == MVT::i16 && RHSC != -2) || (VT == MVT::i8 && RHSC != -1)) {
925 return false;
928 Base = Op->getOperand(0);
929 Offset = DAG.getConstant(RHSC, DL, MVT::i8);
930 AM = ISD::PRE_DEC;
932 return true;
935 return false;
938 /// Returns true by value, base pointer and
939 /// offset pointer and addressing mode by reference if this node can be
940 /// combined with a load / store to form a post-indexed load / store.
941 bool AVRTargetLowering::getPostIndexedAddressParts(SDNode *N, SDNode *Op,
942 SDValue &Base,
943 SDValue &Offset,
944 ISD::MemIndexedMode &AM,
945 SelectionDAG &DAG) const {
946 EVT VT;
947 SDLoc DL(N);
949 if (const LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
950 VT = LD->getMemoryVT();
951 if (LD->getExtensionType() != ISD::NON_EXTLOAD)
952 return false;
953 } else if (const StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
954 VT = ST->getMemoryVT();
955 if (AVR::isProgramMemoryAccess(ST)) {
956 return false;
958 } else {
959 return false;
962 if (VT != MVT::i8 && VT != MVT::i16) {
963 return false;
966 if (Op->getOpcode() != ISD::ADD && Op->getOpcode() != ISD::SUB) {
967 return false;
970 if (const ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Op->getOperand(1))) {
971 int RHSC = RHS->getSExtValue();
972 if (Op->getOpcode() == ISD::SUB)
973 RHSC = -RHSC;
974 if ((VT == MVT::i16 && RHSC != 2) || (VT == MVT::i8 && RHSC != 1)) {
975 return false;
978 Base = Op->getOperand(0);
979 Offset = DAG.getConstant(RHSC, DL, MVT::i8);
980 AM = ISD::POST_INC;
982 return true;
985 return false;
988 bool AVRTargetLowering::isOffsetFoldingLegal(
989 const GlobalAddressSDNode *GA) const {
990 return true;
993 //===----------------------------------------------------------------------===//
994 // Formal Arguments Calling Convention Implementation
995 //===----------------------------------------------------------------------===//
997 #include "AVRGenCallingConv.inc"
999 /// Registers for calling conventions, ordered in reverse as required by ABI.
1000 /// Both arrays must be of the same length.
1001 static const MCPhysReg RegList8[] = {
1002 AVR::R25, AVR::R24, AVR::R23, AVR::R22, AVR::R21, AVR::R20,
1003 AVR::R19, AVR::R18, AVR::R17, AVR::R16, AVR::R15, AVR::R14,
1004 AVR::R13, AVR::R12, AVR::R11, AVR::R10, AVR::R9, AVR::R8};
1005 static const MCPhysReg RegList16[] = {
1006 AVR::R26R25, AVR::R25R24, AVR::R24R23, AVR::R23R22,
1007 AVR::R22R21, AVR::R21R20, AVR::R20R19, AVR::R19R18,
1008 AVR::R18R17, AVR::R17R16, AVR::R16R15, AVR::R15R14,
1009 AVR::R14R13, AVR::R13R12, AVR::R12R11, AVR::R11R10,
1010 AVR::R10R9, AVR::R9R8};
1012 static_assert(array_lengthof(RegList8) == array_lengthof(RegList16),
1013 "8-bit and 16-bit register arrays must be of equal length");
1015 /// Analyze incoming and outgoing function arguments. We need custom C++ code
1016 /// to handle special constraints in the ABI.
1017 /// In addition, all pieces of a certain argument have to be passed either
1018 /// using registers or the stack but never mixing both.
1019 template <typename ArgT>
1020 static void
1021 analyzeArguments(TargetLowering::CallLoweringInfo *CLI, const Function *F,
1022 const DataLayout *TD, const SmallVectorImpl<ArgT> &Args,
1023 SmallVectorImpl<CCValAssign> &ArgLocs, CCState &CCInfo) {
1024 unsigned NumArgs = Args.size();
1025 // This is the index of the last used register, in RegList*.
1026 // -1 means R26 (R26 is never actually used in CC).
1027 int RegLastIdx = -1;
1028 // Once a value is passed to the stack it will always be used
1029 bool UseStack = false;
1030 for (unsigned i = 0; i != NumArgs;) {
1031 MVT VT = Args[i].VT;
1032 // We have to count the number of bytes for each function argument, that is
1033 // those Args with the same OrigArgIndex. This is important in case the
1034 // function takes an aggregate type.
1035 // Current argument will be between [i..j).
1036 unsigned ArgIndex = Args[i].OrigArgIndex;
1037 unsigned TotalBytes = VT.getStoreSize();
1038 unsigned j = i + 1;
1039 for (; j != NumArgs; ++j) {
1040 if (Args[j].OrigArgIndex != ArgIndex)
1041 break;
1042 TotalBytes += Args[j].VT.getStoreSize();
1044 // Round up to even number of bytes.
1045 TotalBytes = alignTo(TotalBytes, 2);
1046 // Skip zero sized arguments
1047 if (TotalBytes == 0)
1048 continue;
1049 // The index of the first register to be used
1050 unsigned RegIdx = RegLastIdx + TotalBytes;
1051 RegLastIdx = RegIdx;
1052 // If there are not enough registers, use the stack
1053 if (RegIdx >= array_lengthof(RegList8)) {
1054 UseStack = true;
1056 for (; i != j; ++i) {
1057 MVT VT = Args[i].VT;
1059 if (UseStack) {
1060 auto evt = EVT(VT).getTypeForEVT(CCInfo.getContext());
1061 unsigned Offset = CCInfo.AllocateStack(TD->getTypeAllocSize(evt),
1062 TD->getABITypeAlign(evt));
1063 CCInfo.addLoc(
1064 CCValAssign::getMem(i, VT, Offset, VT, CCValAssign::Full));
1065 } else {
1066 unsigned Reg;
1067 if (VT == MVT::i8) {
1068 Reg = CCInfo.AllocateReg(RegList8[RegIdx]);
1069 } else if (VT == MVT::i16) {
1070 Reg = CCInfo.AllocateReg(RegList16[RegIdx]);
1071 } else {
1072 llvm_unreachable(
1073 "calling convention can only manage i8 and i16 types");
1075 assert(Reg && "register not available in calling convention");
1076 CCInfo.addLoc(CCValAssign::getReg(i, VT, Reg, VT, CCValAssign::Full));
1077 // Registers inside a particular argument are sorted in increasing order
1078 // (remember the array is reversed).
1079 RegIdx -= VT.getStoreSize();
1085 /// Count the total number of bytes needed to pass or return these arguments.
1086 template <typename ArgT>
1087 static unsigned getTotalArgumentsSizeInBytes(const SmallVectorImpl<ArgT> &Args) {
1088 unsigned TotalBytes = 0;
1090 for (const ArgT& Arg : Args) {
1091 TotalBytes += Arg.VT.getStoreSize();
1093 return TotalBytes;
1096 /// Analyze incoming and outgoing value of returning from a function.
1097 /// The algorithm is similar to analyzeArguments, but there can only be
1098 /// one value, possibly an aggregate, and it is limited to 8 bytes.
1099 template <typename ArgT>
1100 static void analyzeReturnValues(const SmallVectorImpl<ArgT> &Args,
1101 CCState &CCInfo) {
1102 unsigned NumArgs = Args.size();
1103 unsigned TotalBytes = getTotalArgumentsSizeInBytes(Args);
1104 // CanLowerReturn() guarantees this assertion.
1105 assert(TotalBytes <= 8 && "return values greater than 8 bytes cannot be lowered");
1107 // GCC-ABI says that the size is rounded up to the next even number,
1108 // but actually once it is more than 4 it will always round up to 8.
1109 if (TotalBytes > 4) {
1110 TotalBytes = 8;
1111 } else {
1112 TotalBytes = alignTo(TotalBytes, 2);
1115 // The index of the first register to use.
1116 int RegIdx = TotalBytes - 1;
1117 for (unsigned i = 0; i != NumArgs; ++i) {
1118 MVT VT = Args[i].VT;
1119 unsigned Reg;
1120 if (VT == MVT::i8) {
1121 Reg = CCInfo.AllocateReg(RegList8[RegIdx]);
1122 } else if (VT == MVT::i16) {
1123 Reg = CCInfo.AllocateReg(RegList16[RegIdx]);
1124 } else {
1125 llvm_unreachable("calling convention can only manage i8 and i16 types");
1127 assert(Reg && "register not available in calling convention");
1128 CCInfo.addLoc(CCValAssign::getReg(i, VT, Reg, VT, CCValAssign::Full));
1129 // Registers sort in increasing order
1130 RegIdx -= VT.getStoreSize();
1134 SDValue AVRTargetLowering::LowerFormalArguments(
1135 SDValue Chain, CallingConv::ID CallConv, bool isVarArg,
1136 const SmallVectorImpl<ISD::InputArg> &Ins, const SDLoc &dl,
1137 SelectionDAG &DAG, SmallVectorImpl<SDValue> &InVals) const {
1138 MachineFunction &MF = DAG.getMachineFunction();
1139 MachineFrameInfo &MFI = MF.getFrameInfo();
1140 auto DL = DAG.getDataLayout();
1142 // Assign locations to all of the incoming arguments.
1143 SmallVector<CCValAssign, 16> ArgLocs;
1144 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), ArgLocs,
1145 *DAG.getContext());
1147 // Variadic functions do not need all the analysis below.
1148 if (isVarArg) {
1149 CCInfo.AnalyzeFormalArguments(Ins, ArgCC_AVR_Vararg);
1150 } else {
1151 analyzeArguments(nullptr, &MF.getFunction(), &DL, Ins, ArgLocs, CCInfo);
1154 SDValue ArgValue;
1155 for (CCValAssign &VA : ArgLocs) {
1157 // Arguments stored on registers.
1158 if (VA.isRegLoc()) {
1159 EVT RegVT = VA.getLocVT();
1160 const TargetRegisterClass *RC;
1161 if (RegVT == MVT::i8) {
1162 RC = &AVR::GPR8RegClass;
1163 } else if (RegVT == MVT::i16) {
1164 RC = &AVR::DREGSRegClass;
1165 } else {
1166 llvm_unreachable("Unknown argument type!");
1169 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
1170 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
1172 // :NOTE: Clang should not promote any i8 into i16 but for safety the
1173 // following code will handle zexts or sexts generated by other
1174 // front ends. Otherwise:
1175 // If this is an 8 bit value, it is really passed promoted
1176 // to 16 bits. Insert an assert[sz]ext to capture this, then
1177 // truncate to the right size.
1178 switch (VA.getLocInfo()) {
1179 default:
1180 llvm_unreachable("Unknown loc info!");
1181 case CCValAssign::Full:
1182 break;
1183 case CCValAssign::BCvt:
1184 ArgValue = DAG.getNode(ISD::BITCAST, dl, VA.getValVT(), ArgValue);
1185 break;
1186 case CCValAssign::SExt:
1187 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
1188 DAG.getValueType(VA.getValVT()));
1189 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
1190 break;
1191 case CCValAssign::ZExt:
1192 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
1193 DAG.getValueType(VA.getValVT()));
1194 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
1195 break;
1198 InVals.push_back(ArgValue);
1199 } else {
1200 // Sanity check.
1201 assert(VA.isMemLoc());
1203 EVT LocVT = VA.getLocVT();
1205 // Create the frame index object for this incoming parameter.
1206 int FI = MFI.CreateFixedObject(LocVT.getSizeInBits() / 8,
1207 VA.getLocMemOffset(), true);
1209 // Create the SelectionDAG nodes corresponding to a load
1210 // from this parameter.
1211 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy(DL));
1212 InVals.push_back(DAG.getLoad(LocVT, dl, Chain, FIN,
1213 MachinePointerInfo::getFixedStack(MF, FI)));
1217 // If the function takes variable number of arguments, make a frame index for
1218 // the start of the first vararg value... for expansion of llvm.va_start.
1219 if (isVarArg) {
1220 unsigned StackSize = CCInfo.getNextStackOffset();
1221 AVRMachineFunctionInfo *AFI = MF.getInfo<AVRMachineFunctionInfo>();
1223 AFI->setVarArgsFrameIndex(MFI.CreateFixedObject(2, StackSize, true));
1226 return Chain;
1229 //===----------------------------------------------------------------------===//
1230 // Call Calling Convention Implementation
1231 //===----------------------------------------------------------------------===//
1233 SDValue AVRTargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI,
1234 SmallVectorImpl<SDValue> &InVals) const {
1235 SelectionDAG &DAG = CLI.DAG;
1236 SDLoc &DL = CLI.DL;
1237 SmallVectorImpl<ISD::OutputArg> &Outs = CLI.Outs;
1238 SmallVectorImpl<SDValue> &OutVals = CLI.OutVals;
1239 SmallVectorImpl<ISD::InputArg> &Ins = CLI.Ins;
1240 SDValue Chain = CLI.Chain;
1241 SDValue Callee = CLI.Callee;
1242 bool &isTailCall = CLI.IsTailCall;
1243 CallingConv::ID CallConv = CLI.CallConv;
1244 bool isVarArg = CLI.IsVarArg;
1246 MachineFunction &MF = DAG.getMachineFunction();
1248 // AVR does not yet support tail call optimization.
1249 isTailCall = false;
1251 // Analyze operands of the call, assigning locations to each operand.
1252 SmallVector<CCValAssign, 16> ArgLocs;
1253 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), ArgLocs,
1254 *DAG.getContext());
1256 // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every
1257 // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol
1258 // node so that legalize doesn't hack it.
1259 const Function *F = nullptr;
1260 if (const GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1261 const GlobalValue *GV = G->getGlobal();
1263 F = cast<Function>(GV);
1264 Callee =
1265 DAG.getTargetGlobalAddress(GV, DL, getPointerTy(DAG.getDataLayout()));
1266 } else if (const ExternalSymbolSDNode *ES =
1267 dyn_cast<ExternalSymbolSDNode>(Callee)) {
1268 Callee = DAG.getTargetExternalSymbol(ES->getSymbol(),
1269 getPointerTy(DAG.getDataLayout()));
1272 // Variadic functions do not need all the analysis below.
1273 if (isVarArg) {
1274 CCInfo.AnalyzeCallOperands(Outs, ArgCC_AVR_Vararg);
1275 } else {
1276 analyzeArguments(&CLI, F, &DAG.getDataLayout(), Outs, ArgLocs, CCInfo);
1279 // Get a count of how many bytes are to be pushed on the stack.
1280 unsigned NumBytes = CCInfo.getNextStackOffset();
1282 Chain = DAG.getCALLSEQ_START(Chain, NumBytes, 0, DL);
1284 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
1286 // First, walk the register assignments, inserting copies.
1287 unsigned AI, AE;
1288 bool HasStackArgs = false;
1289 for (AI = 0, AE = ArgLocs.size(); AI != AE; ++AI) {
1290 CCValAssign &VA = ArgLocs[AI];
1291 EVT RegVT = VA.getLocVT();
1292 SDValue Arg = OutVals[AI];
1294 // Promote the value if needed. With Clang this should not happen.
1295 switch (VA.getLocInfo()) {
1296 default:
1297 llvm_unreachable("Unknown loc info!");
1298 case CCValAssign::Full:
1299 break;
1300 case CCValAssign::SExt:
1301 Arg = DAG.getNode(ISD::SIGN_EXTEND, DL, RegVT, Arg);
1302 break;
1303 case CCValAssign::ZExt:
1304 Arg = DAG.getNode(ISD::ZERO_EXTEND, DL, RegVT, Arg);
1305 break;
1306 case CCValAssign::AExt:
1307 Arg = DAG.getNode(ISD::ANY_EXTEND, DL, RegVT, Arg);
1308 break;
1309 case CCValAssign::BCvt:
1310 Arg = DAG.getNode(ISD::BITCAST, DL, RegVT, Arg);
1311 break;
1314 // Stop when we encounter a stack argument, we need to process them
1315 // in reverse order in the loop below.
1316 if (VA.isMemLoc()) {
1317 HasStackArgs = true;
1318 break;
1321 // Arguments that can be passed on registers must be kept in the RegsToPass
1322 // vector.
1323 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1326 // Second, stack arguments have to walked.
1327 // Previously this code created chained stores but those chained stores appear
1328 // to be unchained in the legalization phase. Therefore, do not attempt to
1329 // chain them here. In fact, chaining them here somehow causes the first and
1330 // second store to be reversed which is the exact opposite of the intended
1331 // effect.
1332 if (HasStackArgs) {
1333 SmallVector<SDValue, 8> MemOpChains;
1334 for (; AI != AE; AI++) {
1335 CCValAssign &VA = ArgLocs[AI];
1336 SDValue Arg = OutVals[AI];
1338 assert(VA.isMemLoc());
1340 // SP points to one stack slot further so add one to adjust it.
1341 SDValue PtrOff = DAG.getNode(
1342 ISD::ADD, DL, getPointerTy(DAG.getDataLayout()),
1343 DAG.getRegister(AVR::SP, getPointerTy(DAG.getDataLayout())),
1344 DAG.getIntPtrConstant(VA.getLocMemOffset() + 1, DL));
1346 MemOpChains.push_back(
1347 DAG.getStore(Chain, DL, Arg, PtrOff,
1348 MachinePointerInfo::getStack(MF, VA.getLocMemOffset())));
1351 if (!MemOpChains.empty())
1352 Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other, MemOpChains);
1355 // Build a sequence of copy-to-reg nodes chained together with token chain and
1356 // flag operands which copy the outgoing args into registers. The InFlag in
1357 // necessary since all emited instructions must be stuck together.
1358 SDValue InFlag;
1359 for (auto Reg : RegsToPass) {
1360 Chain = DAG.getCopyToReg(Chain, DL, Reg.first, Reg.second, InFlag);
1361 InFlag = Chain.getValue(1);
1364 // Returns a chain & a flag for retval copy to use.
1365 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
1366 SmallVector<SDValue, 8> Ops;
1367 Ops.push_back(Chain);
1368 Ops.push_back(Callee);
1370 // Add argument registers to the end of the list so that they are known live
1371 // into the call.
1372 for (auto Reg : RegsToPass) {
1373 Ops.push_back(DAG.getRegister(Reg.first, Reg.second.getValueType()));
1376 // Add a register mask operand representing the call-preserved registers.
1377 const TargetRegisterInfo *TRI = Subtarget.getRegisterInfo();
1378 const uint32_t *Mask =
1379 TRI->getCallPreservedMask(DAG.getMachineFunction(), CallConv);
1380 assert(Mask && "Missing call preserved mask for calling convention");
1381 Ops.push_back(DAG.getRegisterMask(Mask));
1383 if (InFlag.getNode()) {
1384 Ops.push_back(InFlag);
1387 Chain = DAG.getNode(AVRISD::CALL, DL, NodeTys, Ops);
1388 InFlag = Chain.getValue(1);
1390 // Create the CALLSEQ_END node.
1391 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, DL, true),
1392 DAG.getIntPtrConstant(0, DL, true), InFlag, DL);
1394 if (!Ins.empty()) {
1395 InFlag = Chain.getValue(1);
1398 // Handle result values, copying them out of physregs into vregs that we
1399 // return.
1400 return LowerCallResult(Chain, InFlag, CallConv, isVarArg, Ins, DL, DAG,
1401 InVals);
1404 /// Lower the result values of a call into the
1405 /// appropriate copies out of appropriate physical registers.
1407 SDValue AVRTargetLowering::LowerCallResult(
1408 SDValue Chain, SDValue InFlag, CallingConv::ID CallConv, bool isVarArg,
1409 const SmallVectorImpl<ISD::InputArg> &Ins, const SDLoc &dl, SelectionDAG &DAG,
1410 SmallVectorImpl<SDValue> &InVals) const {
1412 // Assign locations to each value returned by this call.
1413 SmallVector<CCValAssign, 16> RVLocs;
1414 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), RVLocs,
1415 *DAG.getContext());
1417 // Handle runtime calling convs.
1418 if (CallConv == CallingConv::AVR_BUILTIN) {
1419 CCInfo.AnalyzeCallResult(Ins, RetCC_AVR_BUILTIN);
1420 } else {
1421 analyzeReturnValues(Ins, CCInfo);
1424 // Copy all of the result registers out of their specified physreg.
1425 for (CCValAssign const &RVLoc : RVLocs) {
1426 Chain = DAG.getCopyFromReg(Chain, dl, RVLoc.getLocReg(), RVLoc.getValVT(),
1427 InFlag)
1428 .getValue(1);
1429 InFlag = Chain.getValue(2);
1430 InVals.push_back(Chain.getValue(0));
1433 return Chain;
1436 //===----------------------------------------------------------------------===//
1437 // Return Value Calling Convention Implementation
1438 //===----------------------------------------------------------------------===//
1440 bool AVRTargetLowering::CanLowerReturn(
1441 CallingConv::ID CallConv, MachineFunction &MF, bool isVarArg,
1442 const SmallVectorImpl<ISD::OutputArg> &Outs, LLVMContext &Context) const {
1443 if (CallConv == CallingConv::AVR_BUILTIN) {
1444 SmallVector<CCValAssign, 16> RVLocs;
1445 CCState CCInfo(CallConv, isVarArg, MF, RVLocs, Context);
1446 return CCInfo.CheckReturn(Outs, RetCC_AVR_BUILTIN);
1449 unsigned TotalBytes = getTotalArgumentsSizeInBytes(Outs);
1450 return TotalBytes <= 8;
1453 SDValue
1454 AVRTargetLowering::LowerReturn(SDValue Chain, CallingConv::ID CallConv,
1455 bool isVarArg,
1456 const SmallVectorImpl<ISD::OutputArg> &Outs,
1457 const SmallVectorImpl<SDValue> &OutVals,
1458 const SDLoc &dl, SelectionDAG &DAG) const {
1459 // CCValAssign - represent the assignment of the return value to locations.
1460 SmallVector<CCValAssign, 16> RVLocs;
1462 // CCState - Info about the registers and stack slot.
1463 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), RVLocs,
1464 *DAG.getContext());
1466 MachineFunction &MF = DAG.getMachineFunction();
1468 // Analyze return values.
1469 if (CallConv == CallingConv::AVR_BUILTIN) {
1470 CCInfo.AnalyzeReturn(Outs, RetCC_AVR_BUILTIN);
1471 } else {
1472 analyzeReturnValues(Outs, CCInfo);
1475 SDValue Flag;
1476 SmallVector<SDValue, 4> RetOps(1, Chain);
1477 // Copy the result values into the output registers.
1478 for (unsigned i = 0, e = RVLocs.size(); i != e; ++i) {
1479 CCValAssign &VA = RVLocs[i];
1480 assert(VA.isRegLoc() && "Can only return in registers!");
1482 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), OutVals[i], Flag);
1484 // Guarantee that all emitted copies are stuck together with flags.
1485 Flag = Chain.getValue(1);
1486 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
1489 // Don't emit the ret/reti instruction when the naked attribute is present in
1490 // the function being compiled.
1491 if (MF.getFunction().getAttributes().hasFnAttr(Attribute::Naked)) {
1492 return Chain;
1495 const AVRMachineFunctionInfo *AFI = MF.getInfo<AVRMachineFunctionInfo>();
1497 unsigned RetOpc =
1498 AFI->isInterruptOrSignalHandler()
1499 ? AVRISD::RETI_FLAG
1500 : AVRISD::RET_FLAG;
1502 RetOps[0] = Chain; // Update chain.
1504 if (Flag.getNode()) {
1505 RetOps.push_back(Flag);
1508 return DAG.getNode(RetOpc, dl, MVT::Other, RetOps);
1511 //===----------------------------------------------------------------------===//
1512 // Custom Inserters
1513 //===----------------------------------------------------------------------===//
1515 MachineBasicBlock *AVRTargetLowering::insertShift(MachineInstr &MI,
1516 MachineBasicBlock *BB) const {
1517 unsigned Opc;
1518 const TargetRegisterClass *RC;
1519 bool HasRepeatedOperand = false;
1520 MachineFunction *F = BB->getParent();
1521 MachineRegisterInfo &RI = F->getRegInfo();
1522 const TargetInstrInfo &TII = *Subtarget.getInstrInfo();
1523 DebugLoc dl = MI.getDebugLoc();
1525 switch (MI.getOpcode()) {
1526 default:
1527 llvm_unreachable("Invalid shift opcode!");
1528 case AVR::Lsl8:
1529 Opc = AVR::ADDRdRr; // LSL is an alias of ADD Rd, Rd
1530 RC = &AVR::GPR8RegClass;
1531 HasRepeatedOperand = true;
1532 break;
1533 case AVR::Lsl16:
1534 Opc = AVR::LSLWRd;
1535 RC = &AVR::DREGSRegClass;
1536 break;
1537 case AVR::Asr8:
1538 Opc = AVR::ASRRd;
1539 RC = &AVR::GPR8RegClass;
1540 break;
1541 case AVR::Asr16:
1542 Opc = AVR::ASRWRd;
1543 RC = &AVR::DREGSRegClass;
1544 break;
1545 case AVR::Lsr8:
1546 Opc = AVR::LSRRd;
1547 RC = &AVR::GPR8RegClass;
1548 break;
1549 case AVR::Lsr16:
1550 Opc = AVR::LSRWRd;
1551 RC = &AVR::DREGSRegClass;
1552 break;
1553 case AVR::Rol8:
1554 Opc = AVR::ROLBRd;
1555 RC = &AVR::GPR8RegClass;
1556 break;
1557 case AVR::Rol16:
1558 Opc = AVR::ROLWRd;
1559 RC = &AVR::DREGSRegClass;
1560 break;
1561 case AVR::Ror8:
1562 Opc = AVR::RORBRd;
1563 RC = &AVR::GPR8RegClass;
1564 break;
1565 case AVR::Ror16:
1566 Opc = AVR::RORWRd;
1567 RC = &AVR::DREGSRegClass;
1568 break;
1571 const BasicBlock *LLVM_BB = BB->getBasicBlock();
1573 MachineFunction::iterator I;
1574 for (I = BB->getIterator(); I != F->end() && &(*I) != BB; ++I);
1575 if (I != F->end()) ++I;
1577 // Create loop block.
1578 MachineBasicBlock *LoopBB = F->CreateMachineBasicBlock(LLVM_BB);
1579 MachineBasicBlock *CheckBB = F->CreateMachineBasicBlock(LLVM_BB);
1580 MachineBasicBlock *RemBB = F->CreateMachineBasicBlock(LLVM_BB);
1582 F->insert(I, LoopBB);
1583 F->insert(I, CheckBB);
1584 F->insert(I, RemBB);
1586 // Update machine-CFG edges by transferring all successors of the current
1587 // block to the block containing instructions after shift.
1588 RemBB->splice(RemBB->begin(), BB, std::next(MachineBasicBlock::iterator(MI)),
1589 BB->end());
1590 RemBB->transferSuccessorsAndUpdatePHIs(BB);
1592 // Add edges BB => LoopBB => CheckBB => RemBB, CheckBB => LoopBB.
1593 BB->addSuccessor(CheckBB);
1594 LoopBB->addSuccessor(CheckBB);
1595 CheckBB->addSuccessor(LoopBB);
1596 CheckBB->addSuccessor(RemBB);
1598 Register ShiftAmtReg = RI.createVirtualRegister(&AVR::GPR8RegClass);
1599 Register ShiftAmtReg2 = RI.createVirtualRegister(&AVR::GPR8RegClass);
1600 Register ShiftReg = RI.createVirtualRegister(RC);
1601 Register ShiftReg2 = RI.createVirtualRegister(RC);
1602 Register ShiftAmtSrcReg = MI.getOperand(2).getReg();
1603 Register SrcReg = MI.getOperand(1).getReg();
1604 Register DstReg = MI.getOperand(0).getReg();
1606 // BB:
1607 // rjmp CheckBB
1608 BuildMI(BB, dl, TII.get(AVR::RJMPk)).addMBB(CheckBB);
1610 // LoopBB:
1611 // ShiftReg2 = shift ShiftReg
1612 auto ShiftMI = BuildMI(LoopBB, dl, TII.get(Opc), ShiftReg2).addReg(ShiftReg);
1613 if (HasRepeatedOperand)
1614 ShiftMI.addReg(ShiftReg);
1616 // CheckBB:
1617 // ShiftReg = phi [%SrcReg, BB], [%ShiftReg2, LoopBB]
1618 // ShiftAmt = phi [%N, BB], [%ShiftAmt2, LoopBB]
1619 // DestReg = phi [%SrcReg, BB], [%ShiftReg, LoopBB]
1620 // ShiftAmt2 = ShiftAmt - 1;
1621 // if (ShiftAmt2 >= 0) goto LoopBB;
1622 BuildMI(CheckBB, dl, TII.get(AVR::PHI), ShiftReg)
1623 .addReg(SrcReg)
1624 .addMBB(BB)
1625 .addReg(ShiftReg2)
1626 .addMBB(LoopBB);
1627 BuildMI(CheckBB, dl, TII.get(AVR::PHI), ShiftAmtReg)
1628 .addReg(ShiftAmtSrcReg)
1629 .addMBB(BB)
1630 .addReg(ShiftAmtReg2)
1631 .addMBB(LoopBB);
1632 BuildMI(CheckBB, dl, TII.get(AVR::PHI), DstReg)
1633 .addReg(SrcReg)
1634 .addMBB(BB)
1635 .addReg(ShiftReg2)
1636 .addMBB(LoopBB);
1638 BuildMI(CheckBB, dl, TII.get(AVR::DECRd), ShiftAmtReg2)
1639 .addReg(ShiftAmtReg);
1640 BuildMI(CheckBB, dl, TII.get(AVR::BRPLk)).addMBB(LoopBB);
1642 MI.eraseFromParent(); // The pseudo instruction is gone now.
1643 return RemBB;
1646 static bool isCopyMulResult(MachineBasicBlock::iterator const &I) {
1647 if (I->getOpcode() == AVR::COPY) {
1648 Register SrcReg = I->getOperand(1).getReg();
1649 return (SrcReg == AVR::R0 || SrcReg == AVR::R1);
1652 return false;
1655 // The mul instructions wreak havock on our zero_reg R1. We need to clear it
1656 // after the result has been evacuated. This is probably not the best way to do
1657 // it, but it works for now.
1658 MachineBasicBlock *AVRTargetLowering::insertMul(MachineInstr &MI,
1659 MachineBasicBlock *BB) const {
1660 const TargetInstrInfo &TII = *Subtarget.getInstrInfo();
1661 MachineBasicBlock::iterator I(MI);
1662 ++I; // in any case insert *after* the mul instruction
1663 if (isCopyMulResult(I))
1664 ++I;
1665 if (isCopyMulResult(I))
1666 ++I;
1667 BuildMI(*BB, I, MI.getDebugLoc(), TII.get(AVR::EORRdRr), AVR::R1)
1668 .addReg(AVR::R1)
1669 .addReg(AVR::R1);
1670 return BB;
1673 MachineBasicBlock *
1674 AVRTargetLowering::EmitInstrWithCustomInserter(MachineInstr &MI,
1675 MachineBasicBlock *MBB) const {
1676 int Opc = MI.getOpcode();
1678 // Pseudo shift instructions with a non constant shift amount are expanded
1679 // into a loop.
1680 switch (Opc) {
1681 case AVR::Lsl8:
1682 case AVR::Lsl16:
1683 case AVR::Lsr8:
1684 case AVR::Lsr16:
1685 case AVR::Rol8:
1686 case AVR::Rol16:
1687 case AVR::Ror8:
1688 case AVR::Ror16:
1689 case AVR::Asr8:
1690 case AVR::Asr16:
1691 return insertShift(MI, MBB);
1692 case AVR::MULRdRr:
1693 case AVR::MULSRdRr:
1694 return insertMul(MI, MBB);
1697 assert((Opc == AVR::Select16 || Opc == AVR::Select8) &&
1698 "Unexpected instr type to insert");
1700 const AVRInstrInfo &TII = (const AVRInstrInfo &)*MI.getParent()
1701 ->getParent()
1702 ->getSubtarget()
1703 .getInstrInfo();
1704 DebugLoc dl = MI.getDebugLoc();
1706 // To "insert" a SELECT instruction, we insert the diamond
1707 // control-flow pattern. The incoming instruction knows the
1708 // destination vreg to set, the condition code register to branch
1709 // on, the true/false values to select between, and a branch opcode
1710 // to use.
1712 MachineFunction *MF = MBB->getParent();
1713 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
1714 MachineBasicBlock *FallThrough = MBB->getFallThrough();
1716 // If the current basic block falls through to another basic block,
1717 // we must insert an unconditional branch to the fallthrough destination
1718 // if we are to insert basic blocks at the prior fallthrough point.
1719 if (FallThrough != nullptr) {
1720 BuildMI(MBB, dl, TII.get(AVR::RJMPk)).addMBB(FallThrough);
1723 MachineBasicBlock *trueMBB = MF->CreateMachineBasicBlock(LLVM_BB);
1724 MachineBasicBlock *falseMBB = MF->CreateMachineBasicBlock(LLVM_BB);
1726 MachineFunction::iterator I;
1727 for (I = MF->begin(); I != MF->end() && &(*I) != MBB; ++I);
1728 if (I != MF->end()) ++I;
1729 MF->insert(I, trueMBB);
1730 MF->insert(I, falseMBB);
1732 // Transfer remaining instructions and all successors of the current
1733 // block to the block which will contain the Phi node for the
1734 // select.
1735 trueMBB->splice(trueMBB->begin(), MBB,
1736 std::next(MachineBasicBlock::iterator(MI)), MBB->end());
1737 trueMBB->transferSuccessorsAndUpdatePHIs(MBB);
1739 AVRCC::CondCodes CC = (AVRCC::CondCodes)MI.getOperand(3).getImm();
1740 BuildMI(MBB, dl, TII.getBrCond(CC)).addMBB(trueMBB);
1741 BuildMI(MBB, dl, TII.get(AVR::RJMPk)).addMBB(falseMBB);
1742 MBB->addSuccessor(falseMBB);
1743 MBB->addSuccessor(trueMBB);
1745 // Unconditionally flow back to the true block
1746 BuildMI(falseMBB, dl, TII.get(AVR::RJMPk)).addMBB(trueMBB);
1747 falseMBB->addSuccessor(trueMBB);
1749 // Set up the Phi node to determine where we came from
1750 BuildMI(*trueMBB, trueMBB->begin(), dl, TII.get(AVR::PHI), MI.getOperand(0).getReg())
1751 .addReg(MI.getOperand(1).getReg())
1752 .addMBB(MBB)
1753 .addReg(MI.getOperand(2).getReg())
1754 .addMBB(falseMBB) ;
1756 MI.eraseFromParent(); // The pseudo instruction is gone now.
1757 return trueMBB;
1760 //===----------------------------------------------------------------------===//
1761 // Inline Asm Support
1762 //===----------------------------------------------------------------------===//
1764 AVRTargetLowering::ConstraintType
1765 AVRTargetLowering::getConstraintType(StringRef Constraint) const {
1766 if (Constraint.size() == 1) {
1767 // See http://www.nongnu.org/avr-libc/user-manual/inline_asm.html
1768 switch (Constraint[0]) {
1769 default:
1770 break;
1771 case 'a': // Simple upper registers
1772 case 'b': // Base pointer registers pairs
1773 case 'd': // Upper register
1774 case 'l': // Lower registers
1775 case 'e': // Pointer register pairs
1776 case 'q': // Stack pointer register
1777 case 'r': // Any register
1778 case 'w': // Special upper register pairs
1779 return C_RegisterClass;
1780 case 't': // Temporary register
1781 case 'x': case 'X': // Pointer register pair X
1782 case 'y': case 'Y': // Pointer register pair Y
1783 case 'z': case 'Z': // Pointer register pair Z
1784 return C_Register;
1785 case 'Q': // A memory address based on Y or Z pointer with displacement.
1786 return C_Memory;
1787 case 'G': // Floating point constant
1788 case 'I': // 6-bit positive integer constant
1789 case 'J': // 6-bit negative integer constant
1790 case 'K': // Integer constant (Range: 2)
1791 case 'L': // Integer constant (Range: 0)
1792 case 'M': // 8-bit integer constant
1793 case 'N': // Integer constant (Range: -1)
1794 case 'O': // Integer constant (Range: 8, 16, 24)
1795 case 'P': // Integer constant (Range: 1)
1796 case 'R': // Integer constant (Range: -6 to 5)x
1797 return C_Immediate;
1801 return TargetLowering::getConstraintType(Constraint);
1804 unsigned
1805 AVRTargetLowering::getInlineAsmMemConstraint(StringRef ConstraintCode) const {
1806 // Not sure if this is actually the right thing to do, but we got to do
1807 // *something* [agnat]
1808 switch (ConstraintCode[0]) {
1809 case 'Q':
1810 return InlineAsm::Constraint_Q;
1812 return TargetLowering::getInlineAsmMemConstraint(ConstraintCode);
1815 AVRTargetLowering::ConstraintWeight
1816 AVRTargetLowering::getSingleConstraintMatchWeight(
1817 AsmOperandInfo &info, const char *constraint) const {
1818 ConstraintWeight weight = CW_Invalid;
1819 Value *CallOperandVal = info.CallOperandVal;
1821 // If we don't have a value, we can't do a match,
1822 // but allow it at the lowest weight.
1823 // (this behaviour has been copied from the ARM backend)
1824 if (!CallOperandVal) {
1825 return CW_Default;
1828 // Look at the constraint type.
1829 switch (*constraint) {
1830 default:
1831 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
1832 break;
1833 case 'd':
1834 case 'r':
1835 case 'l':
1836 weight = CW_Register;
1837 break;
1838 case 'a':
1839 case 'b':
1840 case 'e':
1841 case 'q':
1842 case 't':
1843 case 'w':
1844 case 'x': case 'X':
1845 case 'y': case 'Y':
1846 case 'z': case 'Z':
1847 weight = CW_SpecificReg;
1848 break;
1849 case 'G':
1850 if (const ConstantFP *C = dyn_cast<ConstantFP>(CallOperandVal)) {
1851 if (C->isZero()) {
1852 weight = CW_Constant;
1855 break;
1856 case 'I':
1857 if (const ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
1858 if (isUInt<6>(C->getZExtValue())) {
1859 weight = CW_Constant;
1862 break;
1863 case 'J':
1864 if (const ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
1865 if ((C->getSExtValue() >= -63) && (C->getSExtValue() <= 0)) {
1866 weight = CW_Constant;
1869 break;
1870 case 'K':
1871 if (const ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
1872 if (C->getZExtValue() == 2) {
1873 weight = CW_Constant;
1876 break;
1877 case 'L':
1878 if (const ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
1879 if (C->getZExtValue() == 0) {
1880 weight = CW_Constant;
1883 break;
1884 case 'M':
1885 if (const ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
1886 if (isUInt<8>(C->getZExtValue())) {
1887 weight = CW_Constant;
1890 break;
1891 case 'N':
1892 if (const ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
1893 if (C->getSExtValue() == -1) {
1894 weight = CW_Constant;
1897 break;
1898 case 'O':
1899 if (const ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
1900 if ((C->getZExtValue() == 8) || (C->getZExtValue() == 16) ||
1901 (C->getZExtValue() == 24)) {
1902 weight = CW_Constant;
1905 break;
1906 case 'P':
1907 if (const ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
1908 if (C->getZExtValue() == 1) {
1909 weight = CW_Constant;
1912 break;
1913 case 'R':
1914 if (const ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
1915 if ((C->getSExtValue() >= -6) && (C->getSExtValue() <= 5)) {
1916 weight = CW_Constant;
1919 break;
1920 case 'Q':
1921 weight = CW_Memory;
1922 break;
1925 return weight;
1928 std::pair<unsigned, const TargetRegisterClass *>
1929 AVRTargetLowering::getRegForInlineAsmConstraint(const TargetRegisterInfo *TRI,
1930 StringRef Constraint,
1931 MVT VT) const {
1932 if (Constraint.size() == 1) {
1933 switch (Constraint[0]) {
1934 case 'a': // Simple upper registers r16..r23.
1935 if (VT == MVT::i8)
1936 return std::make_pair(0U, &AVR::LD8loRegClass);
1937 else if (VT == MVT::i16)
1938 return std::make_pair(0U, &AVR::DREGSLD8loRegClass);
1939 break;
1940 case 'b': // Base pointer registers: y, z.
1941 if (VT == MVT::i8 || VT == MVT::i16)
1942 return std::make_pair(0U, &AVR::PTRDISPREGSRegClass);
1943 break;
1944 case 'd': // Upper registers r16..r31.
1945 if (VT == MVT::i8)
1946 return std::make_pair(0U, &AVR::LD8RegClass);
1947 else if (VT == MVT::i16)
1948 return std::make_pair(0U, &AVR::DLDREGSRegClass);
1949 break;
1950 case 'l': // Lower registers r0..r15.
1951 if (VT == MVT::i8)
1952 return std::make_pair(0U, &AVR::GPR8loRegClass);
1953 else if (VT == MVT::i16)
1954 return std::make_pair(0U, &AVR::DREGSloRegClass);
1955 break;
1956 case 'e': // Pointer register pairs: x, y, z.
1957 if (VT == MVT::i8 || VT == MVT::i16)
1958 return std::make_pair(0U, &AVR::PTRREGSRegClass);
1959 break;
1960 case 'q': // Stack pointer register: SPH:SPL.
1961 return std::make_pair(0U, &AVR::GPRSPRegClass);
1962 case 'r': // Any register: r0..r31.
1963 if (VT == MVT::i8)
1964 return std::make_pair(0U, &AVR::GPR8RegClass);
1965 else if (VT == MVT::i16)
1966 return std::make_pair(0U, &AVR::DREGSRegClass);
1967 break;
1968 case 't': // Temporary register: r0.
1969 if (VT == MVT::i8)
1970 return std::make_pair(unsigned(AVR::R0), &AVR::GPR8RegClass);
1971 break;
1972 case 'w': // Special upper register pairs: r24, r26, r28, r30.
1973 if (VT == MVT::i8 || VT == MVT::i16)
1974 return std::make_pair(0U, &AVR::IWREGSRegClass);
1975 break;
1976 case 'x': // Pointer register pair X: r27:r26.
1977 case 'X':
1978 if (VT == MVT::i8 || VT == MVT::i16)
1979 return std::make_pair(unsigned(AVR::R27R26), &AVR::PTRREGSRegClass);
1980 break;
1981 case 'y': // Pointer register pair Y: r29:r28.
1982 case 'Y':
1983 if (VT == MVT::i8 || VT == MVT::i16)
1984 return std::make_pair(unsigned(AVR::R29R28), &AVR::PTRREGSRegClass);
1985 break;
1986 case 'z': // Pointer register pair Z: r31:r30.
1987 case 'Z':
1988 if (VT == MVT::i8 || VT == MVT::i16)
1989 return std::make_pair(unsigned(AVR::R31R30), &AVR::PTRREGSRegClass);
1990 break;
1991 default:
1992 break;
1996 return TargetLowering::getRegForInlineAsmConstraint(
1997 Subtarget.getRegisterInfo(), Constraint, VT);
2000 void AVRTargetLowering::LowerAsmOperandForConstraint(SDValue Op,
2001 std::string &Constraint,
2002 std::vector<SDValue> &Ops,
2003 SelectionDAG &DAG) const {
2004 SDValue Result(0, 0);
2005 SDLoc DL(Op);
2006 EVT Ty = Op.getValueType();
2008 // Currently only support length 1 constraints.
2009 if (Constraint.length() != 1) {
2010 return;
2013 char ConstraintLetter = Constraint[0];
2014 switch (ConstraintLetter) {
2015 default:
2016 break;
2017 // Deal with integers first:
2018 case 'I':
2019 case 'J':
2020 case 'K':
2021 case 'L':
2022 case 'M':
2023 case 'N':
2024 case 'O':
2025 case 'P':
2026 case 'R': {
2027 const ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
2028 if (!C) {
2029 return;
2032 int64_t CVal64 = C->getSExtValue();
2033 uint64_t CUVal64 = C->getZExtValue();
2034 switch (ConstraintLetter) {
2035 case 'I': // 0..63
2036 if (!isUInt<6>(CUVal64))
2037 return;
2038 Result = DAG.getTargetConstant(CUVal64, DL, Ty);
2039 break;
2040 case 'J': // -63..0
2041 if (CVal64 < -63 || CVal64 > 0)
2042 return;
2043 Result = DAG.getTargetConstant(CVal64, DL, Ty);
2044 break;
2045 case 'K': // 2
2046 if (CUVal64 != 2)
2047 return;
2048 Result = DAG.getTargetConstant(CUVal64, DL, Ty);
2049 break;
2050 case 'L': // 0
2051 if (CUVal64 != 0)
2052 return;
2053 Result = DAG.getTargetConstant(CUVal64, DL, Ty);
2054 break;
2055 case 'M': // 0..255
2056 if (!isUInt<8>(CUVal64))
2057 return;
2058 // i8 type may be printed as a negative number,
2059 // e.g. 254 would be printed as -2,
2060 // so we force it to i16 at least.
2061 if (Ty.getSimpleVT() == MVT::i8) {
2062 Ty = MVT::i16;
2064 Result = DAG.getTargetConstant(CUVal64, DL, Ty);
2065 break;
2066 case 'N': // -1
2067 if (CVal64 != -1)
2068 return;
2069 Result = DAG.getTargetConstant(CVal64, DL, Ty);
2070 break;
2071 case 'O': // 8, 16, 24
2072 if (CUVal64 != 8 && CUVal64 != 16 && CUVal64 != 24)
2073 return;
2074 Result = DAG.getTargetConstant(CUVal64, DL, Ty);
2075 break;
2076 case 'P': // 1
2077 if (CUVal64 != 1)
2078 return;
2079 Result = DAG.getTargetConstant(CUVal64, DL, Ty);
2080 break;
2081 case 'R': // -6..5
2082 if (CVal64 < -6 || CVal64 > 5)
2083 return;
2084 Result = DAG.getTargetConstant(CVal64, DL, Ty);
2085 break;
2088 break;
2090 case 'G':
2091 const ConstantFPSDNode *FC = dyn_cast<ConstantFPSDNode>(Op);
2092 if (!FC || !FC->isZero())
2093 return;
2094 // Soften float to i8 0
2095 Result = DAG.getTargetConstant(0, DL, MVT::i8);
2096 break;
2099 if (Result.getNode()) {
2100 Ops.push_back(Result);
2101 return;
2104 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
2107 Register AVRTargetLowering::getRegisterByName(const char *RegName, LLT VT,
2108 const MachineFunction &MF) const {
2109 Register Reg;
2111 if (VT == LLT::scalar(8)) {
2112 Reg = StringSwitch<unsigned>(RegName)
2113 .Case("r0", AVR::R0)
2114 .Case("r1", AVR::R1)
2115 .Default(0);
2116 } else {
2117 Reg = StringSwitch<unsigned>(RegName)
2118 .Case("r0", AVR::R1R0)
2119 .Case("sp", AVR::SP)
2120 .Default(0);
2123 if (Reg)
2124 return Reg;
2126 report_fatal_error(
2127 Twine("Invalid register name \"" + StringRef(RegName) + "\"."));
2130 } // end of namespace llvm