[InstCombine] Signed saturation patterns
[llvm-core.git] / lib / Target / Lanai / LanaiISelLowering.cpp
blob43933d062a7eff5c6754c5ba090b74014982d539
1 //===-- LanaiISelLowering.cpp - Lanai 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 implements the LanaiTargetLowering class.
11 //===----------------------------------------------------------------------===//
13 #include "LanaiISelLowering.h"
14 #include "Lanai.h"
15 #include "LanaiCondCode.h"
16 #include "LanaiMachineFunctionInfo.h"
17 #include "LanaiSubtarget.h"
18 #include "LanaiTargetObjectFile.h"
19 #include "MCTargetDesc/LanaiBaseInfo.h"
20 #include "llvm/ADT/APInt.h"
21 #include "llvm/ADT/ArrayRef.h"
22 #include "llvm/ADT/SmallVector.h"
23 #include "llvm/ADT/StringRef.h"
24 #include "llvm/ADT/StringSwitch.h"
25 #include "llvm/CodeGen/CallingConvLower.h"
26 #include "llvm/CodeGen/MachineFrameInfo.h"
27 #include "llvm/CodeGen/MachineFunction.h"
28 #include "llvm/CodeGen/MachineMemOperand.h"
29 #include "llvm/CodeGen/MachineRegisterInfo.h"
30 #include "llvm/CodeGen/RuntimeLibcalls.h"
31 #include "llvm/CodeGen/SelectionDAG.h"
32 #include "llvm/CodeGen/SelectionDAGNodes.h"
33 #include "llvm/CodeGen/TargetCallingConv.h"
34 #include "llvm/CodeGen/ValueTypes.h"
35 #include "llvm/IR/CallingConv.h"
36 #include "llvm/IR/DerivedTypes.h"
37 #include "llvm/IR/Function.h"
38 #include "llvm/IR/GlobalValue.h"
39 #include "llvm/Support/Casting.h"
40 #include "llvm/Support/CodeGen.h"
41 #include "llvm/Support/CommandLine.h"
42 #include "llvm/Support/Debug.h"
43 #include "llvm/Support/ErrorHandling.h"
44 #include "llvm/Support/KnownBits.h"
45 #include "llvm/Support/MachineValueType.h"
46 #include "llvm/Support/MathExtras.h"
47 #include "llvm/Support/raw_ostream.h"
48 #include "llvm/Target/TargetMachine.h"
49 #include <cassert>
50 #include <cmath>
51 #include <cstdint>
52 #include <cstdlib>
53 #include <utility>
55 #define DEBUG_TYPE "lanai-lower"
57 using namespace llvm;
59 // Limit on number of instructions the lowered multiplication may have before a
60 // call to the library function should be generated instead. The threshold is
61 // currently set to 14 as this was the smallest threshold that resulted in all
62 // constant multiplications being lowered. A threshold of 5 covered all cases
63 // except for one multiplication which required 14. mulsi3 requires 16
64 // instructions (including the prologue and epilogue but excluding instructions
65 // at call site). Until we can inline mulsi3, generating at most 14 instructions
66 // will be faster than invoking mulsi3.
67 static cl::opt<int> LanaiLowerConstantMulThreshold(
68 "lanai-constant-mul-threshold", cl::Hidden,
69 cl::desc("Maximum number of instruction to generate when lowering constant "
70 "multiplication instead of calling library function [default=14]"),
71 cl::init(14));
73 LanaiTargetLowering::LanaiTargetLowering(const TargetMachine &TM,
74 const LanaiSubtarget &STI)
75 : TargetLowering(TM) {
76 // Set up the register classes.
77 addRegisterClass(MVT::i32, &Lanai::GPRRegClass);
79 // Compute derived properties from the register classes
80 TRI = STI.getRegisterInfo();
81 computeRegisterProperties(TRI);
83 setStackPointerRegisterToSaveRestore(Lanai::SP);
85 setOperationAction(ISD::BR_CC, MVT::i32, Custom);
86 setOperationAction(ISD::BR_JT, MVT::Other, Expand);
87 setOperationAction(ISD::BRCOND, MVT::Other, Expand);
88 setOperationAction(ISD::SETCC, MVT::i32, Custom);
89 setOperationAction(ISD::SELECT, MVT::i32, Expand);
90 setOperationAction(ISD::SELECT_CC, MVT::i32, Custom);
92 setOperationAction(ISD::GlobalAddress, MVT::i32, Custom);
93 setOperationAction(ISD::BlockAddress, MVT::i32, Custom);
94 setOperationAction(ISD::JumpTable, MVT::i32, Custom);
95 setOperationAction(ISD::ConstantPool, MVT::i32, Custom);
97 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Custom);
98 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
99 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
101 setOperationAction(ISD::VASTART, MVT::Other, Custom);
102 setOperationAction(ISD::VAARG, MVT::Other, Expand);
103 setOperationAction(ISD::VACOPY, MVT::Other, Expand);
104 setOperationAction(ISD::VAEND, MVT::Other, Expand);
106 setOperationAction(ISD::SDIV, MVT::i32, Expand);
107 setOperationAction(ISD::UDIV, MVT::i32, Expand);
108 setOperationAction(ISD::SDIVREM, MVT::i32, Expand);
109 setOperationAction(ISD::UDIVREM, MVT::i32, Expand);
110 setOperationAction(ISD::SREM, MVT::i32, Expand);
111 setOperationAction(ISD::UREM, MVT::i32, Expand);
113 setOperationAction(ISD::MUL, MVT::i32, Custom);
114 setOperationAction(ISD::MULHU, MVT::i32, Expand);
115 setOperationAction(ISD::MULHS, MVT::i32, Expand);
116 setOperationAction(ISD::UMUL_LOHI, MVT::i32, Expand);
117 setOperationAction(ISD::SMUL_LOHI, MVT::i32, Expand);
119 setOperationAction(ISD::ROTR, MVT::i32, Expand);
120 setOperationAction(ISD::ROTL, MVT::i32, Expand);
121 setOperationAction(ISD::SHL_PARTS, MVT::i32, Custom);
122 setOperationAction(ISD::SRL_PARTS, MVT::i32, Custom);
123 setOperationAction(ISD::SRA_PARTS, MVT::i32, Expand);
125 setOperationAction(ISD::BSWAP, MVT::i32, Expand);
126 setOperationAction(ISD::CTPOP, MVT::i32, Legal);
127 setOperationAction(ISD::CTLZ, MVT::i32, Legal);
128 setOperationAction(ISD::CTTZ, MVT::i32, Legal);
130 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand);
131 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8, Expand);
132 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16, Expand);
134 // Extended load operations for i1 types must be promoted
135 for (MVT VT : MVT::integer_valuetypes()) {
136 setLoadExtAction(ISD::EXTLOAD, VT, MVT::i1, Promote);
137 setLoadExtAction(ISD::ZEXTLOAD, VT, MVT::i1, Promote);
138 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::i1, Promote);
141 setTargetDAGCombine(ISD::ADD);
142 setTargetDAGCombine(ISD::SUB);
143 setTargetDAGCombine(ISD::AND);
144 setTargetDAGCombine(ISD::OR);
145 setTargetDAGCombine(ISD::XOR);
147 // Function alignments
148 setMinFunctionAlignment(Align(4));
149 setPrefFunctionAlignment(Align(4));
151 setJumpIsExpensive(true);
153 // TODO: Setting the minimum jump table entries needed before a
154 // switch is transformed to a jump table to 100 to avoid creating jump tables
155 // as this was causing bad performance compared to a large group of if
156 // statements. Re-evaluate this on new benchmarks.
157 setMinimumJumpTableEntries(100);
159 // Use fast calling convention for library functions.
160 for (int I = 0; I < RTLIB::UNKNOWN_LIBCALL; ++I) {
161 setLibcallCallingConv(static_cast<RTLIB::Libcall>(I), CallingConv::Fast);
164 MaxStoresPerMemset = 16; // For @llvm.memset -> sequence of stores
165 MaxStoresPerMemsetOptSize = 8;
166 MaxStoresPerMemcpy = 16; // For @llvm.memcpy -> sequence of stores
167 MaxStoresPerMemcpyOptSize = 8;
168 MaxStoresPerMemmove = 16; // For @llvm.memmove -> sequence of stores
169 MaxStoresPerMemmoveOptSize = 8;
171 // Booleans always contain 0 or 1.
172 setBooleanContents(ZeroOrOneBooleanContent);
175 SDValue LanaiTargetLowering::LowerOperation(SDValue Op,
176 SelectionDAG &DAG) const {
177 switch (Op.getOpcode()) {
178 case ISD::MUL:
179 return LowerMUL(Op, DAG);
180 case ISD::BR_CC:
181 return LowerBR_CC(Op, DAG);
182 case ISD::ConstantPool:
183 return LowerConstantPool(Op, DAG);
184 case ISD::GlobalAddress:
185 return LowerGlobalAddress(Op, DAG);
186 case ISD::BlockAddress:
187 return LowerBlockAddress(Op, DAG);
188 case ISD::JumpTable:
189 return LowerJumpTable(Op, DAG);
190 case ISD::SELECT_CC:
191 return LowerSELECT_CC(Op, DAG);
192 case ISD::SETCC:
193 return LowerSETCC(Op, DAG);
194 case ISD::SHL_PARTS:
195 return LowerSHL_PARTS(Op, DAG);
196 case ISD::SRL_PARTS:
197 return LowerSRL_PARTS(Op, DAG);
198 case ISD::VASTART:
199 return LowerVASTART(Op, DAG);
200 case ISD::DYNAMIC_STACKALLOC:
201 return LowerDYNAMIC_STACKALLOC(Op, DAG);
202 case ISD::RETURNADDR:
203 return LowerRETURNADDR(Op, DAG);
204 case ISD::FRAMEADDR:
205 return LowerFRAMEADDR(Op, DAG);
206 default:
207 llvm_unreachable("unimplemented operand");
211 //===----------------------------------------------------------------------===//
212 // Lanai Inline Assembly Support
213 //===----------------------------------------------------------------------===//
215 Register LanaiTargetLowering::getRegisterByName(
216 const char *RegName, EVT /*VT*/,
217 const MachineFunction & /*MF*/) const {
218 // Only unallocatable registers should be matched here.
219 Register Reg = StringSwitch<unsigned>(RegName)
220 .Case("pc", Lanai::PC)
221 .Case("sp", Lanai::SP)
222 .Case("fp", Lanai::FP)
223 .Case("rr1", Lanai::RR1)
224 .Case("r10", Lanai::R10)
225 .Case("rr2", Lanai::RR2)
226 .Case("r11", Lanai::R11)
227 .Case("rca", Lanai::RCA)
228 .Default(0);
230 if (Reg)
231 return Reg;
232 report_fatal_error("Invalid register name global variable");
235 std::pair<unsigned, const TargetRegisterClass *>
236 LanaiTargetLowering::getRegForInlineAsmConstraint(const TargetRegisterInfo *TRI,
237 StringRef Constraint,
238 MVT VT) const {
239 if (Constraint.size() == 1)
240 // GCC Constraint Letters
241 switch (Constraint[0]) {
242 case 'r': // GENERAL_REGS
243 return std::make_pair(0U, &Lanai::GPRRegClass);
244 default:
245 break;
248 return TargetLowering::getRegForInlineAsmConstraint(TRI, Constraint, VT);
251 // Examine constraint type and operand type and determine a weight value.
252 // This object must already have been set up with the operand type
253 // and the current alternative constraint selected.
254 TargetLowering::ConstraintWeight
255 LanaiTargetLowering::getSingleConstraintMatchWeight(
256 AsmOperandInfo &Info, const char *Constraint) const {
257 ConstraintWeight Weight = CW_Invalid;
258 Value *CallOperandVal = Info.CallOperandVal;
259 // If we don't have a value, we can't do a match,
260 // but allow it at the lowest weight.
261 if (CallOperandVal == nullptr)
262 return CW_Default;
263 // Look at the constraint type.
264 switch (*Constraint) {
265 case 'I': // signed 16 bit immediate
266 case 'J': // integer zero
267 case 'K': // unsigned 16 bit immediate
268 case 'L': // immediate in the range 0 to 31
269 case 'M': // signed 32 bit immediate where lower 16 bits are 0
270 case 'N': // signed 26 bit immediate
271 case 'O': // integer zero
272 if (isa<ConstantInt>(CallOperandVal))
273 Weight = CW_Constant;
274 break;
275 default:
276 Weight = TargetLowering::getSingleConstraintMatchWeight(Info, Constraint);
277 break;
279 return Weight;
282 // LowerAsmOperandForConstraint - Lower the specified operand into the Ops
283 // vector. If it is invalid, don't add anything to Ops.
284 void LanaiTargetLowering::LowerAsmOperandForConstraint(
285 SDValue Op, std::string &Constraint, std::vector<SDValue> &Ops,
286 SelectionDAG &DAG) const {
287 SDValue Result(nullptr, 0);
289 // Only support length 1 constraints for now.
290 if (Constraint.length() > 1)
291 return;
293 char ConstraintLetter = Constraint[0];
294 switch (ConstraintLetter) {
295 case 'I': // Signed 16 bit constant
296 // If this fails, the parent routine will give an error
297 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
298 if (isInt<16>(C->getSExtValue())) {
299 Result = DAG.getTargetConstant(C->getSExtValue(), SDLoc(C),
300 Op.getValueType());
301 break;
304 return;
305 case 'J': // integer zero
306 case 'O':
307 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
308 if (C->getZExtValue() == 0) {
309 Result = DAG.getTargetConstant(0, SDLoc(C), Op.getValueType());
310 break;
313 return;
314 case 'K': // unsigned 16 bit immediate
315 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
316 if (isUInt<16>(C->getZExtValue())) {
317 Result = DAG.getTargetConstant(C->getSExtValue(), SDLoc(C),
318 Op.getValueType());
319 break;
322 return;
323 case 'L': // immediate in the range 0 to 31
324 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
325 if (C->getZExtValue() <= 31) {
326 Result = DAG.getTargetConstant(C->getZExtValue(), SDLoc(C),
327 Op.getValueType());
328 break;
331 return;
332 case 'M': // signed 32 bit immediate where lower 16 bits are 0
333 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
334 int64_t Val = C->getSExtValue();
335 if ((isInt<32>(Val)) && ((Val & 0xffff) == 0)) {
336 Result = DAG.getTargetConstant(Val, SDLoc(C), Op.getValueType());
337 break;
340 return;
341 case 'N': // signed 26 bit immediate
342 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
343 int64_t Val = C->getSExtValue();
344 if ((Val >= -33554432) && (Val <= 33554431)) {
345 Result = DAG.getTargetConstant(Val, SDLoc(C), Op.getValueType());
346 break;
349 return;
350 default:
351 break; // This will fall through to the generic implementation
354 if (Result.getNode()) {
355 Ops.push_back(Result);
356 return;
359 TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
362 //===----------------------------------------------------------------------===//
363 // Calling Convention Implementation
364 //===----------------------------------------------------------------------===//
366 #include "LanaiGenCallingConv.inc"
368 static unsigned NumFixedArgs;
369 static bool CC_Lanai32_VarArg(unsigned ValNo, MVT ValVT, MVT LocVT,
370 CCValAssign::LocInfo LocInfo,
371 ISD::ArgFlagsTy ArgFlags, CCState &State) {
372 // Handle fixed arguments with default CC.
373 // Note: Both the default and fast CC handle VarArg the same and hence the
374 // calling convention of the function is not considered here.
375 if (ValNo < NumFixedArgs) {
376 return CC_Lanai32(ValNo, ValVT, LocVT, LocInfo, ArgFlags, State);
379 // Promote i8/i16 args to i32
380 if (LocVT == MVT::i8 || LocVT == MVT::i16) {
381 LocVT = MVT::i32;
382 if (ArgFlags.isSExt())
383 LocInfo = CCValAssign::SExt;
384 else if (ArgFlags.isZExt())
385 LocInfo = CCValAssign::ZExt;
386 else
387 LocInfo = CCValAssign::AExt;
390 // VarArgs get passed on stack
391 unsigned Offset = State.AllocateStack(4, 4);
392 State.addLoc(CCValAssign::getMem(ValNo, ValVT, Offset, LocVT, LocInfo));
393 return false;
396 SDValue LanaiTargetLowering::LowerFormalArguments(
397 SDValue Chain, CallingConv::ID CallConv, bool IsVarArg,
398 const SmallVectorImpl<ISD::InputArg> &Ins, const SDLoc &DL,
399 SelectionDAG &DAG, SmallVectorImpl<SDValue> &InVals) const {
400 switch (CallConv) {
401 case CallingConv::C:
402 case CallingConv::Fast:
403 return LowerCCCArguments(Chain, CallConv, IsVarArg, Ins, DL, DAG, InVals);
404 default:
405 report_fatal_error("Unsupported calling convention");
409 SDValue LanaiTargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI,
410 SmallVectorImpl<SDValue> &InVals) const {
411 SelectionDAG &DAG = CLI.DAG;
412 SDLoc &DL = CLI.DL;
413 SmallVectorImpl<ISD::OutputArg> &Outs = CLI.Outs;
414 SmallVectorImpl<SDValue> &OutVals = CLI.OutVals;
415 SmallVectorImpl<ISD::InputArg> &Ins = CLI.Ins;
416 SDValue Chain = CLI.Chain;
417 SDValue Callee = CLI.Callee;
418 bool &IsTailCall = CLI.IsTailCall;
419 CallingConv::ID CallConv = CLI.CallConv;
420 bool IsVarArg = CLI.IsVarArg;
422 // Lanai target does not yet support tail call optimization.
423 IsTailCall = false;
425 switch (CallConv) {
426 case CallingConv::Fast:
427 case CallingConv::C:
428 return LowerCCCCallTo(Chain, Callee, CallConv, IsVarArg, IsTailCall, Outs,
429 OutVals, Ins, DL, DAG, InVals);
430 default:
431 report_fatal_error("Unsupported calling convention");
435 // LowerCCCArguments - transform physical registers into virtual registers and
436 // generate load operations for arguments places on the stack.
437 SDValue LanaiTargetLowering::LowerCCCArguments(
438 SDValue Chain, CallingConv::ID CallConv, bool IsVarArg,
439 const SmallVectorImpl<ISD::InputArg> &Ins, const SDLoc &DL,
440 SelectionDAG &DAG, SmallVectorImpl<SDValue> &InVals) const {
441 MachineFunction &MF = DAG.getMachineFunction();
442 MachineFrameInfo &MFI = MF.getFrameInfo();
443 MachineRegisterInfo &RegInfo = MF.getRegInfo();
444 LanaiMachineFunctionInfo *LanaiMFI = MF.getInfo<LanaiMachineFunctionInfo>();
446 // Assign locations to all of the incoming arguments.
447 SmallVector<CCValAssign, 16> ArgLocs;
448 CCState CCInfo(CallConv, IsVarArg, DAG.getMachineFunction(), ArgLocs,
449 *DAG.getContext());
450 if (CallConv == CallingConv::Fast) {
451 CCInfo.AnalyzeFormalArguments(Ins, CC_Lanai32_Fast);
452 } else {
453 CCInfo.AnalyzeFormalArguments(Ins, CC_Lanai32);
456 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
457 CCValAssign &VA = ArgLocs[i];
458 if (VA.isRegLoc()) {
459 // Arguments passed in registers
460 EVT RegVT = VA.getLocVT();
461 switch (RegVT.getSimpleVT().SimpleTy) {
462 case MVT::i32: {
463 Register VReg = RegInfo.createVirtualRegister(&Lanai::GPRRegClass);
464 RegInfo.addLiveIn(VA.getLocReg(), VReg);
465 SDValue ArgValue = DAG.getCopyFromReg(Chain, DL, VReg, RegVT);
467 // If this is an 8/16-bit value, it is really passed promoted to 32
468 // bits. Insert an assert[sz]ext to capture this, then truncate to the
469 // right size.
470 if (VA.getLocInfo() == CCValAssign::SExt)
471 ArgValue = DAG.getNode(ISD::AssertSext, DL, RegVT, ArgValue,
472 DAG.getValueType(VA.getValVT()));
473 else if (VA.getLocInfo() == CCValAssign::ZExt)
474 ArgValue = DAG.getNode(ISD::AssertZext, DL, RegVT, ArgValue,
475 DAG.getValueType(VA.getValVT()));
477 if (VA.getLocInfo() != CCValAssign::Full)
478 ArgValue = DAG.getNode(ISD::TRUNCATE, DL, VA.getValVT(), ArgValue);
480 InVals.push_back(ArgValue);
481 break;
483 default:
484 LLVM_DEBUG(dbgs() << "LowerFormalArguments Unhandled argument type: "
485 << RegVT.getEVTString() << "\n");
486 llvm_unreachable("unhandled argument type");
488 } else {
489 // Sanity check
490 assert(VA.isMemLoc());
491 // Load the argument to a virtual register
492 unsigned ObjSize = VA.getLocVT().getSizeInBits() / 8;
493 // Check that the argument fits in stack slot
494 if (ObjSize > 4) {
495 errs() << "LowerFormalArguments Unhandled argument type: "
496 << EVT(VA.getLocVT()).getEVTString() << "\n";
498 // Create the frame index object for this incoming parameter...
499 int FI = MFI.CreateFixedObject(ObjSize, VA.getLocMemOffset(), true);
501 // Create the SelectionDAG nodes corresponding to a load
502 // from this parameter
503 SDValue FIN = DAG.getFrameIndex(FI, MVT::i32);
504 InVals.push_back(DAG.getLoad(
505 VA.getLocVT(), DL, Chain, FIN,
506 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), FI)));
510 // The Lanai ABI for returning structs by value requires that we copy
511 // the sret argument into rv for the return. Save the argument into
512 // a virtual register so that we can access it from the return points.
513 if (MF.getFunction().hasStructRetAttr()) {
514 unsigned Reg = LanaiMFI->getSRetReturnReg();
515 if (!Reg) {
516 Reg = MF.getRegInfo().createVirtualRegister(getRegClassFor(MVT::i32));
517 LanaiMFI->setSRetReturnReg(Reg);
519 SDValue Copy = DAG.getCopyToReg(DAG.getEntryNode(), DL, Reg, InVals[0]);
520 Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other, Copy, Chain);
523 if (IsVarArg) {
524 // Record the frame index of the first variable argument
525 // which is a value necessary to VASTART.
526 int FI = MFI.CreateFixedObject(4, CCInfo.getNextStackOffset(), true);
527 LanaiMFI->setVarArgsFrameIndex(FI);
530 return Chain;
533 SDValue
534 LanaiTargetLowering::LowerReturn(SDValue Chain, CallingConv::ID CallConv,
535 bool IsVarArg,
536 const SmallVectorImpl<ISD::OutputArg> &Outs,
537 const SmallVectorImpl<SDValue> &OutVals,
538 const SDLoc &DL, SelectionDAG &DAG) const {
539 // CCValAssign - represent the assignment of the return value to a location
540 SmallVector<CCValAssign, 16> RVLocs;
542 // CCState - Info about the registers and stack slot.
543 CCState CCInfo(CallConv, IsVarArg, DAG.getMachineFunction(), RVLocs,
544 *DAG.getContext());
546 // Analize return values.
547 CCInfo.AnalyzeReturn(Outs, RetCC_Lanai32);
549 SDValue Flag;
550 SmallVector<SDValue, 4> RetOps(1, Chain);
552 // Copy the result values into the output registers.
553 for (unsigned i = 0; i != RVLocs.size(); ++i) {
554 CCValAssign &VA = RVLocs[i];
555 assert(VA.isRegLoc() && "Can only return in registers!");
557 Chain = DAG.getCopyToReg(Chain, DL, VA.getLocReg(), OutVals[i], Flag);
559 // Guarantee that all emitted copies are stuck together with flags.
560 Flag = Chain.getValue(1);
561 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
564 // The Lanai ABI for returning structs by value requires that we copy
565 // the sret argument into rv for the return. We saved the argument into
566 // a virtual register in the entry block, so now we copy the value out
567 // and into rv.
568 if (DAG.getMachineFunction().getFunction().hasStructRetAttr()) {
569 MachineFunction &MF = DAG.getMachineFunction();
570 LanaiMachineFunctionInfo *LanaiMFI = MF.getInfo<LanaiMachineFunctionInfo>();
571 unsigned Reg = LanaiMFI->getSRetReturnReg();
572 assert(Reg &&
573 "SRetReturnReg should have been set in LowerFormalArguments().");
574 SDValue Val =
575 DAG.getCopyFromReg(Chain, DL, Reg, getPointerTy(DAG.getDataLayout()));
577 Chain = DAG.getCopyToReg(Chain, DL, Lanai::RV, Val, Flag);
578 Flag = Chain.getValue(1);
579 RetOps.push_back(
580 DAG.getRegister(Lanai::RV, getPointerTy(DAG.getDataLayout())));
583 RetOps[0] = Chain; // Update chain
585 unsigned Opc = LanaiISD::RET_FLAG;
586 if (Flag.getNode())
587 RetOps.push_back(Flag);
589 // Return Void
590 return DAG.getNode(Opc, DL, MVT::Other,
591 ArrayRef<SDValue>(&RetOps[0], RetOps.size()));
594 // LowerCCCCallTo - functions arguments are copied from virtual regs to
595 // (physical regs)/(stack frame), CALLSEQ_START and CALLSEQ_END are emitted.
596 SDValue LanaiTargetLowering::LowerCCCCallTo(
597 SDValue Chain, SDValue Callee, CallingConv::ID CallConv, bool IsVarArg,
598 bool /*IsTailCall*/, const SmallVectorImpl<ISD::OutputArg> &Outs,
599 const SmallVectorImpl<SDValue> &OutVals,
600 const SmallVectorImpl<ISD::InputArg> &Ins, const SDLoc &DL,
601 SelectionDAG &DAG, SmallVectorImpl<SDValue> &InVals) const {
602 // Analyze operands of the call, assigning locations to each operand.
603 SmallVector<CCValAssign, 16> ArgLocs;
604 CCState CCInfo(CallConv, IsVarArg, DAG.getMachineFunction(), ArgLocs,
605 *DAG.getContext());
606 GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee);
607 MachineFrameInfo &MFI = DAG.getMachineFunction().getFrameInfo();
609 NumFixedArgs = 0;
610 if (IsVarArg && G) {
611 const Function *CalleeFn = dyn_cast<Function>(G->getGlobal());
612 if (CalleeFn)
613 NumFixedArgs = CalleeFn->getFunctionType()->getNumParams();
615 if (NumFixedArgs)
616 CCInfo.AnalyzeCallOperands(Outs, CC_Lanai32_VarArg);
617 else {
618 if (CallConv == CallingConv::Fast)
619 CCInfo.AnalyzeCallOperands(Outs, CC_Lanai32_Fast);
620 else
621 CCInfo.AnalyzeCallOperands(Outs, CC_Lanai32);
624 // Get a count of how many bytes are to be pushed on the stack.
625 unsigned NumBytes = CCInfo.getNextStackOffset();
627 // Create local copies for byval args.
628 SmallVector<SDValue, 8> ByValArgs;
629 for (unsigned I = 0, E = Outs.size(); I != E; ++I) {
630 ISD::ArgFlagsTy Flags = Outs[I].Flags;
631 if (!Flags.isByVal())
632 continue;
634 SDValue Arg = OutVals[I];
635 unsigned Size = Flags.getByValSize();
636 unsigned Align = Flags.getByValAlign();
638 int FI = MFI.CreateStackObject(Size, Align, false);
639 SDValue FIPtr = DAG.getFrameIndex(FI, getPointerTy(DAG.getDataLayout()));
640 SDValue SizeNode = DAG.getConstant(Size, DL, MVT::i32);
642 Chain = DAG.getMemcpy(Chain, DL, FIPtr, Arg, SizeNode, Align,
643 /*IsVolatile=*/false,
644 /*AlwaysInline=*/false,
645 /*isTailCall=*/false, MachinePointerInfo(),
646 MachinePointerInfo());
647 ByValArgs.push_back(FIPtr);
650 Chain = DAG.getCALLSEQ_START(Chain, NumBytes, 0, DL);
652 SmallVector<std::pair<unsigned, SDValue>, 4> RegsToPass;
653 SmallVector<SDValue, 12> MemOpChains;
654 SDValue StackPtr;
656 // Walk the register/memloc assignments, inserting copies/loads.
657 for (unsigned I = 0, J = 0, E = ArgLocs.size(); I != E; ++I) {
658 CCValAssign &VA = ArgLocs[I];
659 SDValue Arg = OutVals[I];
660 ISD::ArgFlagsTy Flags = Outs[I].Flags;
662 // Promote the value if needed.
663 switch (VA.getLocInfo()) {
664 case CCValAssign::Full:
665 break;
666 case CCValAssign::SExt:
667 Arg = DAG.getNode(ISD::SIGN_EXTEND, DL, VA.getLocVT(), Arg);
668 break;
669 case CCValAssign::ZExt:
670 Arg = DAG.getNode(ISD::ZERO_EXTEND, DL, VA.getLocVT(), Arg);
671 break;
672 case CCValAssign::AExt:
673 Arg = DAG.getNode(ISD::ANY_EXTEND, DL, VA.getLocVT(), Arg);
674 break;
675 default:
676 llvm_unreachable("Unknown loc info!");
679 // Use local copy if it is a byval arg.
680 if (Flags.isByVal())
681 Arg = ByValArgs[J++];
683 // Arguments that can be passed on register must be kept at RegsToPass
684 // vector
685 if (VA.isRegLoc()) {
686 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
687 } else {
688 assert(VA.isMemLoc());
690 if (StackPtr.getNode() == nullptr)
691 StackPtr = DAG.getCopyFromReg(Chain, DL, Lanai::SP,
692 getPointerTy(DAG.getDataLayout()));
694 SDValue PtrOff =
695 DAG.getNode(ISD::ADD, DL, getPointerTy(DAG.getDataLayout()), StackPtr,
696 DAG.getIntPtrConstant(VA.getLocMemOffset(), DL));
698 MemOpChains.push_back(
699 DAG.getStore(Chain, DL, Arg, PtrOff, MachinePointerInfo()));
703 // Transform all store nodes into one single node because all store nodes are
704 // independent of each other.
705 if (!MemOpChains.empty())
706 Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other,
707 ArrayRef<SDValue>(&MemOpChains[0], MemOpChains.size()));
709 SDValue InFlag;
711 // Build a sequence of copy-to-reg nodes chained together with token chain and
712 // flag operands which copy the outgoing args into registers. The InFlag in
713 // necessary since all emitted instructions must be stuck together.
714 for (unsigned I = 0, E = RegsToPass.size(); I != E; ++I) {
715 Chain = DAG.getCopyToReg(Chain, DL, RegsToPass[I].first,
716 RegsToPass[I].second, InFlag);
717 InFlag = Chain.getValue(1);
720 // If the callee is a GlobalAddress node (quite common, every direct call is)
721 // turn it into a TargetGlobalAddress node so that legalize doesn't hack it.
722 // Likewise ExternalSymbol -> TargetExternalSymbol.
723 uint8_t OpFlag = LanaiII::MO_NO_FLAG;
724 if (G) {
725 Callee = DAG.getTargetGlobalAddress(
726 G->getGlobal(), DL, getPointerTy(DAG.getDataLayout()), 0, OpFlag);
727 } else if (ExternalSymbolSDNode *E = dyn_cast<ExternalSymbolSDNode>(Callee)) {
728 Callee = DAG.getTargetExternalSymbol(
729 E->getSymbol(), getPointerTy(DAG.getDataLayout()), OpFlag);
732 // Returns a chain & a flag for retval copy to use.
733 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
734 SmallVector<SDValue, 8> Ops;
735 Ops.push_back(Chain);
736 Ops.push_back(Callee);
738 // Add a register mask operand representing the call-preserved registers.
739 // TODO: Should return-twice functions be handled?
740 const uint32_t *Mask =
741 TRI->getCallPreservedMask(DAG.getMachineFunction(), CallConv);
742 assert(Mask && "Missing call preserved mask for calling convention");
743 Ops.push_back(DAG.getRegisterMask(Mask));
745 // Add argument registers to the end of the list so that they are
746 // known live into the call.
747 for (unsigned I = 0, E = RegsToPass.size(); I != E; ++I)
748 Ops.push_back(DAG.getRegister(RegsToPass[I].first,
749 RegsToPass[I].second.getValueType()));
751 if (InFlag.getNode())
752 Ops.push_back(InFlag);
754 Chain = DAG.getNode(LanaiISD::CALL, DL, NodeTys,
755 ArrayRef<SDValue>(&Ops[0], Ops.size()));
756 InFlag = Chain.getValue(1);
758 // Create the CALLSEQ_END node.
759 Chain = DAG.getCALLSEQ_END(
760 Chain,
761 DAG.getConstant(NumBytes, DL, getPointerTy(DAG.getDataLayout()), true),
762 DAG.getConstant(0, DL, getPointerTy(DAG.getDataLayout()), true), InFlag,
763 DL);
764 InFlag = Chain.getValue(1);
766 // Handle result values, copying them out of physregs into vregs that we
767 // return.
768 return LowerCallResult(Chain, InFlag, CallConv, IsVarArg, Ins, DL, DAG,
769 InVals);
772 // LowerCallResult - Lower the result values of a call into the
773 // appropriate copies out of appropriate physical registers.
774 SDValue LanaiTargetLowering::LowerCallResult(
775 SDValue Chain, SDValue InFlag, CallingConv::ID CallConv, bool IsVarArg,
776 const SmallVectorImpl<ISD::InputArg> &Ins, const SDLoc &DL,
777 SelectionDAG &DAG, SmallVectorImpl<SDValue> &InVals) const {
778 // Assign locations to each value returned by this call.
779 SmallVector<CCValAssign, 16> RVLocs;
780 CCState CCInfo(CallConv, IsVarArg, DAG.getMachineFunction(), RVLocs,
781 *DAG.getContext());
783 CCInfo.AnalyzeCallResult(Ins, RetCC_Lanai32);
785 // Copy all of the result registers out of their specified physreg.
786 for (unsigned I = 0; I != RVLocs.size(); ++I) {
787 Chain = DAG.getCopyFromReg(Chain, DL, RVLocs[I].getLocReg(),
788 RVLocs[I].getValVT(), InFlag)
789 .getValue(1);
790 InFlag = Chain.getValue(2);
791 InVals.push_back(Chain.getValue(0));
794 return Chain;
797 //===----------------------------------------------------------------------===//
798 // Custom Lowerings
799 //===----------------------------------------------------------------------===//
801 static LPCC::CondCode IntCondCCodeToICC(SDValue CC, const SDLoc &DL,
802 SDValue &RHS, SelectionDAG &DAG) {
803 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
805 // For integer, only the SETEQ, SETNE, SETLT, SETLE, SETGT, SETGE, SETULT,
806 // SETULE, SETUGT, and SETUGE opcodes are used (see CodeGen/ISDOpcodes.h)
807 // and Lanai only supports integer comparisons, so only provide definitions
808 // for them.
809 switch (SetCCOpcode) {
810 case ISD::SETEQ:
811 return LPCC::ICC_EQ;
812 case ISD::SETGT:
813 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS))
814 if (RHSC->getZExtValue() == 0xFFFFFFFF) {
815 // X > -1 -> X >= 0 -> is_plus(X)
816 RHS = DAG.getConstant(0, DL, RHS.getValueType());
817 return LPCC::ICC_PL;
819 return LPCC::ICC_GT;
820 case ISD::SETUGT:
821 return LPCC::ICC_UGT;
822 case ISD::SETLT:
823 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS))
824 if (RHSC->getZExtValue() == 0)
825 // X < 0 -> is_minus(X)
826 return LPCC::ICC_MI;
827 return LPCC::ICC_LT;
828 case ISD::SETULT:
829 return LPCC::ICC_ULT;
830 case ISD::SETLE:
831 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS))
832 if (RHSC->getZExtValue() == 0xFFFFFFFF) {
833 // X <= -1 -> X < 0 -> is_minus(X)
834 RHS = DAG.getConstant(0, DL, RHS.getValueType());
835 return LPCC::ICC_MI;
837 return LPCC::ICC_LE;
838 case ISD::SETULE:
839 return LPCC::ICC_ULE;
840 case ISD::SETGE:
841 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS))
842 if (RHSC->getZExtValue() == 0)
843 // X >= 0 -> is_plus(X)
844 return LPCC::ICC_PL;
845 return LPCC::ICC_GE;
846 case ISD::SETUGE:
847 return LPCC::ICC_UGE;
848 case ISD::SETNE:
849 return LPCC::ICC_NE;
850 case ISD::SETONE:
851 case ISD::SETUNE:
852 case ISD::SETOGE:
853 case ISD::SETOLE:
854 case ISD::SETOLT:
855 case ISD::SETOGT:
856 case ISD::SETOEQ:
857 case ISD::SETUEQ:
858 case ISD::SETO:
859 case ISD::SETUO:
860 llvm_unreachable("Unsupported comparison.");
861 default:
862 llvm_unreachable("Unknown integer condition code!");
866 SDValue LanaiTargetLowering::LowerBR_CC(SDValue Op, SelectionDAG &DAG) const {
867 SDValue Chain = Op.getOperand(0);
868 SDValue Cond = Op.getOperand(1);
869 SDValue LHS = Op.getOperand(2);
870 SDValue RHS = Op.getOperand(3);
871 SDValue Dest = Op.getOperand(4);
872 SDLoc DL(Op);
874 LPCC::CondCode CC = IntCondCCodeToICC(Cond, DL, RHS, DAG);
875 SDValue TargetCC = DAG.getConstant(CC, DL, MVT::i32);
876 SDValue Flag =
877 DAG.getNode(LanaiISD::SET_FLAG, DL, MVT::Glue, LHS, RHS, TargetCC);
879 return DAG.getNode(LanaiISD::BR_CC, DL, Op.getValueType(), Chain, Dest,
880 TargetCC, Flag);
883 SDValue LanaiTargetLowering::LowerMUL(SDValue Op, SelectionDAG &DAG) const {
884 EVT VT = Op->getValueType(0);
885 if (VT != MVT::i32)
886 return SDValue();
888 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op->getOperand(1));
889 if (!C)
890 return SDValue();
892 int64_t MulAmt = C->getSExtValue();
893 int32_t HighestOne = -1;
894 uint32_t NonzeroEntries = 0;
895 int SignedDigit[32] = {0};
897 // Convert to non-adjacent form (NAF) signed-digit representation.
898 // NAF is a signed-digit form where no adjacent digits are non-zero. It is the
899 // minimal Hamming weight representation of a number (on average 1/3 of the
900 // digits will be non-zero vs 1/2 for regular binary representation). And as
901 // the non-zero digits will be the only digits contributing to the instruction
902 // count, this is desirable. The next loop converts it to NAF (following the
903 // approach in 'Guide to Elliptic Curve Cryptography' [ISBN: 038795273X]) by
904 // choosing the non-zero coefficients such that the resulting quotient is
905 // divisible by 2 which will cause the next coefficient to be zero.
906 int64_t E = std::abs(MulAmt);
907 int S = (MulAmt < 0 ? -1 : 1);
908 int I = 0;
909 while (E > 0) {
910 int ZI = 0;
911 if (E % 2 == 1) {
912 ZI = 2 - (E % 4);
913 if (ZI != 0)
914 ++NonzeroEntries;
916 SignedDigit[I] = S * ZI;
917 if (SignedDigit[I] == 1)
918 HighestOne = I;
919 E = (E - ZI) / 2;
920 ++I;
923 // Compute number of instructions required. Due to differences in lowering
924 // between the different processors this count is not exact.
925 // Start by assuming a shift and a add/sub for every non-zero entry (hence
926 // every non-zero entry requires 1 shift and 1 add/sub except for the first
927 // entry).
928 int32_t InstrRequired = 2 * NonzeroEntries - 1;
929 // Correct possible over-adding due to shift by 0 (which is not emitted).
930 if (std::abs(MulAmt) % 2 == 1)
931 --InstrRequired;
932 // Return if the form generated would exceed the instruction threshold.
933 if (InstrRequired > LanaiLowerConstantMulThreshold)
934 return SDValue();
936 SDValue Res;
937 SDLoc DL(Op);
938 SDValue V = Op->getOperand(0);
940 // Initialize the running sum. Set the running sum to the maximal shifted
941 // positive value (i.e., largest i such that zi == 1 and MulAmt has V<<i as a
942 // term NAF).
943 if (HighestOne == -1)
944 Res = DAG.getConstant(0, DL, MVT::i32);
945 else {
946 Res = DAG.getNode(ISD::SHL, DL, VT, V,
947 DAG.getConstant(HighestOne, DL, MVT::i32));
948 SignedDigit[HighestOne] = 0;
951 // Assemble multiplication from shift, add, sub using NAF form and running
952 // sum.
953 for (unsigned int I = 0; I < sizeof(SignedDigit) / sizeof(SignedDigit[0]);
954 ++I) {
955 if (SignedDigit[I] == 0)
956 continue;
958 // Shifted multiplicand (v<<i).
959 SDValue Op =
960 DAG.getNode(ISD::SHL, DL, VT, V, DAG.getConstant(I, DL, MVT::i32));
961 if (SignedDigit[I] == 1)
962 Res = DAG.getNode(ISD::ADD, DL, VT, Res, Op);
963 else if (SignedDigit[I] == -1)
964 Res = DAG.getNode(ISD::SUB, DL, VT, Res, Op);
966 return Res;
969 SDValue LanaiTargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) const {
970 SDValue LHS = Op.getOperand(0);
971 SDValue RHS = Op.getOperand(1);
972 SDValue Cond = Op.getOperand(2);
973 SDLoc DL(Op);
975 LPCC::CondCode CC = IntCondCCodeToICC(Cond, DL, RHS, DAG);
976 SDValue TargetCC = DAG.getConstant(CC, DL, MVT::i32);
977 SDValue Flag =
978 DAG.getNode(LanaiISD::SET_FLAG, DL, MVT::Glue, LHS, RHS, TargetCC);
980 return DAG.getNode(LanaiISD::SETCC, DL, Op.getValueType(), TargetCC, Flag);
983 SDValue LanaiTargetLowering::LowerSELECT_CC(SDValue Op,
984 SelectionDAG &DAG) const {
985 SDValue LHS = Op.getOperand(0);
986 SDValue RHS = Op.getOperand(1);
987 SDValue TrueV = Op.getOperand(2);
988 SDValue FalseV = Op.getOperand(3);
989 SDValue Cond = Op.getOperand(4);
990 SDLoc DL(Op);
992 LPCC::CondCode CC = IntCondCCodeToICC(Cond, DL, RHS, DAG);
993 SDValue TargetCC = DAG.getConstant(CC, DL, MVT::i32);
994 SDValue Flag =
995 DAG.getNode(LanaiISD::SET_FLAG, DL, MVT::Glue, LHS, RHS, TargetCC);
997 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::Glue);
998 return DAG.getNode(LanaiISD::SELECT_CC, DL, VTs, TrueV, FalseV, TargetCC,
999 Flag);
1002 SDValue LanaiTargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG) const {
1003 MachineFunction &MF = DAG.getMachineFunction();
1004 LanaiMachineFunctionInfo *FuncInfo = MF.getInfo<LanaiMachineFunctionInfo>();
1006 SDLoc DL(Op);
1007 SDValue FI = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(),
1008 getPointerTy(DAG.getDataLayout()));
1010 // vastart just stores the address of the VarArgsFrameIndex slot into the
1011 // memory location argument.
1012 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
1013 return DAG.getStore(Op.getOperand(0), DL, FI, Op.getOperand(1),
1014 MachinePointerInfo(SV));
1017 SDValue LanaiTargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
1018 SelectionDAG &DAG) const {
1019 SDValue Chain = Op.getOperand(0);
1020 SDValue Size = Op.getOperand(1);
1021 SDLoc DL(Op);
1023 unsigned SPReg = getStackPointerRegisterToSaveRestore();
1025 // Get a reference to the stack pointer.
1026 SDValue StackPointer = DAG.getCopyFromReg(Chain, DL, SPReg, MVT::i32);
1028 // Subtract the dynamic size from the actual stack size to
1029 // obtain the new stack size.
1030 SDValue Sub = DAG.getNode(ISD::SUB, DL, MVT::i32, StackPointer, Size);
1032 // For Lanai, the outgoing memory arguments area should be on top of the
1033 // alloca area on the stack i.e., the outgoing memory arguments should be
1034 // at a lower address than the alloca area. Move the alloca area down the
1035 // stack by adding back the space reserved for outgoing arguments to SP
1036 // here.
1038 // We do not know what the size of the outgoing args is at this point.
1039 // So, we add a pseudo instruction ADJDYNALLOC that will adjust the
1040 // stack pointer. We replace this instruction with on that has the correct,
1041 // known offset in emitPrologue().
1042 SDValue ArgAdjust = DAG.getNode(LanaiISD::ADJDYNALLOC, DL, MVT::i32, Sub);
1044 // The Sub result contains the new stack start address, so it
1045 // must be placed in the stack pointer register.
1046 SDValue CopyChain = DAG.getCopyToReg(Chain, DL, SPReg, Sub);
1048 SDValue Ops[2] = {ArgAdjust, CopyChain};
1049 return DAG.getMergeValues(Ops, DL);
1052 SDValue LanaiTargetLowering::LowerRETURNADDR(SDValue Op,
1053 SelectionDAG &DAG) const {
1054 MachineFunction &MF = DAG.getMachineFunction();
1055 MachineFrameInfo &MFI = MF.getFrameInfo();
1056 MFI.setReturnAddressIsTaken(true);
1058 EVT VT = Op.getValueType();
1059 SDLoc DL(Op);
1060 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
1061 if (Depth) {
1062 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
1063 const unsigned Offset = -4;
1064 SDValue Ptr = DAG.getNode(ISD::ADD, DL, VT, FrameAddr,
1065 DAG.getIntPtrConstant(Offset, DL));
1066 return DAG.getLoad(VT, DL, DAG.getEntryNode(), Ptr, MachinePointerInfo());
1069 // Return the link register, which contains the return address.
1070 // Mark it an implicit live-in.
1071 unsigned Reg = MF.addLiveIn(TRI->getRARegister(), getRegClassFor(MVT::i32));
1072 return DAG.getCopyFromReg(DAG.getEntryNode(), DL, Reg, VT);
1075 SDValue LanaiTargetLowering::LowerFRAMEADDR(SDValue Op,
1076 SelectionDAG &DAG) const {
1077 MachineFrameInfo &MFI = DAG.getMachineFunction().getFrameInfo();
1078 MFI.setFrameAddressIsTaken(true);
1080 EVT VT = Op.getValueType();
1081 SDLoc DL(Op);
1082 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), DL, Lanai::FP, VT);
1083 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
1084 while (Depth--) {
1085 const unsigned Offset = -8;
1086 SDValue Ptr = DAG.getNode(ISD::ADD, DL, VT, FrameAddr,
1087 DAG.getIntPtrConstant(Offset, DL));
1088 FrameAddr =
1089 DAG.getLoad(VT, DL, DAG.getEntryNode(), Ptr, MachinePointerInfo());
1091 return FrameAddr;
1094 const char *LanaiTargetLowering::getTargetNodeName(unsigned Opcode) const {
1095 switch (Opcode) {
1096 case LanaiISD::ADJDYNALLOC:
1097 return "LanaiISD::ADJDYNALLOC";
1098 case LanaiISD::RET_FLAG:
1099 return "LanaiISD::RET_FLAG";
1100 case LanaiISD::CALL:
1101 return "LanaiISD::CALL";
1102 case LanaiISD::SELECT_CC:
1103 return "LanaiISD::SELECT_CC";
1104 case LanaiISD::SETCC:
1105 return "LanaiISD::SETCC";
1106 case LanaiISD::SUBBF:
1107 return "LanaiISD::SUBBF";
1108 case LanaiISD::SET_FLAG:
1109 return "LanaiISD::SET_FLAG";
1110 case LanaiISD::BR_CC:
1111 return "LanaiISD::BR_CC";
1112 case LanaiISD::Wrapper:
1113 return "LanaiISD::Wrapper";
1114 case LanaiISD::HI:
1115 return "LanaiISD::HI";
1116 case LanaiISD::LO:
1117 return "LanaiISD::LO";
1118 case LanaiISD::SMALL:
1119 return "LanaiISD::SMALL";
1120 default:
1121 return nullptr;
1125 SDValue LanaiTargetLowering::LowerConstantPool(SDValue Op,
1126 SelectionDAG &DAG) const {
1127 SDLoc DL(Op);
1128 ConstantPoolSDNode *N = cast<ConstantPoolSDNode>(Op);
1129 const Constant *C = N->getConstVal();
1130 const LanaiTargetObjectFile *TLOF =
1131 static_cast<const LanaiTargetObjectFile *>(
1132 getTargetMachine().getObjFileLowering());
1134 // If the code model is small or constant will be placed in the small section,
1135 // then assume address will fit in 21-bits.
1136 if (getTargetMachine().getCodeModel() == CodeModel::Small ||
1137 TLOF->isConstantInSmallSection(DAG.getDataLayout(), C)) {
1138 SDValue Small = DAG.getTargetConstantPool(
1139 C, MVT::i32, N->getAlignment(), N->getOffset(), LanaiII::MO_NO_FLAG);
1140 return DAG.getNode(ISD::OR, DL, MVT::i32,
1141 DAG.getRegister(Lanai::R0, MVT::i32),
1142 DAG.getNode(LanaiISD::SMALL, DL, MVT::i32, Small));
1143 } else {
1144 uint8_t OpFlagHi = LanaiII::MO_ABS_HI;
1145 uint8_t OpFlagLo = LanaiII::MO_ABS_LO;
1147 SDValue Hi = DAG.getTargetConstantPool(C, MVT::i32, N->getAlignment(),
1148 N->getOffset(), OpFlagHi);
1149 SDValue Lo = DAG.getTargetConstantPool(C, MVT::i32, N->getAlignment(),
1150 N->getOffset(), OpFlagLo);
1151 Hi = DAG.getNode(LanaiISD::HI, DL, MVT::i32, Hi);
1152 Lo = DAG.getNode(LanaiISD::LO, DL, MVT::i32, Lo);
1153 SDValue Result = DAG.getNode(ISD::OR, DL, MVT::i32, Hi, Lo);
1154 return Result;
1158 SDValue LanaiTargetLowering::LowerGlobalAddress(SDValue Op,
1159 SelectionDAG &DAG) const {
1160 SDLoc DL(Op);
1161 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
1162 int64_t Offset = cast<GlobalAddressSDNode>(Op)->getOffset();
1164 const LanaiTargetObjectFile *TLOF =
1165 static_cast<const LanaiTargetObjectFile *>(
1166 getTargetMachine().getObjFileLowering());
1168 // If the code model is small or global variable will be placed in the small
1169 // section, then assume address will fit in 21-bits.
1170 const GlobalObject *GO = GV->getBaseObject();
1171 if (TLOF->isGlobalInSmallSection(GO, getTargetMachine())) {
1172 SDValue Small = DAG.getTargetGlobalAddress(
1173 GV, DL, getPointerTy(DAG.getDataLayout()), Offset, LanaiII::MO_NO_FLAG);
1174 return DAG.getNode(ISD::OR, DL, MVT::i32,
1175 DAG.getRegister(Lanai::R0, MVT::i32),
1176 DAG.getNode(LanaiISD::SMALL, DL, MVT::i32, Small));
1177 } else {
1178 uint8_t OpFlagHi = LanaiII::MO_ABS_HI;
1179 uint8_t OpFlagLo = LanaiII::MO_ABS_LO;
1181 // Create the TargetGlobalAddress node, folding in the constant offset.
1182 SDValue Hi = DAG.getTargetGlobalAddress(
1183 GV, DL, getPointerTy(DAG.getDataLayout()), Offset, OpFlagHi);
1184 SDValue Lo = DAG.getTargetGlobalAddress(
1185 GV, DL, getPointerTy(DAG.getDataLayout()), Offset, OpFlagLo);
1186 Hi = DAG.getNode(LanaiISD::HI, DL, MVT::i32, Hi);
1187 Lo = DAG.getNode(LanaiISD::LO, DL, MVT::i32, Lo);
1188 return DAG.getNode(ISD::OR, DL, MVT::i32, Hi, Lo);
1192 SDValue LanaiTargetLowering::LowerBlockAddress(SDValue Op,
1193 SelectionDAG &DAG) const {
1194 SDLoc DL(Op);
1195 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
1197 uint8_t OpFlagHi = LanaiII::MO_ABS_HI;
1198 uint8_t OpFlagLo = LanaiII::MO_ABS_LO;
1200 SDValue Hi = DAG.getBlockAddress(BA, MVT::i32, true, OpFlagHi);
1201 SDValue Lo = DAG.getBlockAddress(BA, MVT::i32, true, OpFlagLo);
1202 Hi = DAG.getNode(LanaiISD::HI, DL, MVT::i32, Hi);
1203 Lo = DAG.getNode(LanaiISD::LO, DL, MVT::i32, Lo);
1204 SDValue Result = DAG.getNode(ISD::OR, DL, MVT::i32, Hi, Lo);
1205 return Result;
1208 SDValue LanaiTargetLowering::LowerJumpTable(SDValue Op,
1209 SelectionDAG &DAG) const {
1210 SDLoc DL(Op);
1211 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
1213 // If the code model is small assume address will fit in 21-bits.
1214 if (getTargetMachine().getCodeModel() == CodeModel::Small) {
1215 SDValue Small = DAG.getTargetJumpTable(
1216 JT->getIndex(), getPointerTy(DAG.getDataLayout()), LanaiII::MO_NO_FLAG);
1217 return DAG.getNode(ISD::OR, DL, MVT::i32,
1218 DAG.getRegister(Lanai::R0, MVT::i32),
1219 DAG.getNode(LanaiISD::SMALL, DL, MVT::i32, Small));
1220 } else {
1221 uint8_t OpFlagHi = LanaiII::MO_ABS_HI;
1222 uint8_t OpFlagLo = LanaiII::MO_ABS_LO;
1224 SDValue Hi = DAG.getTargetJumpTable(
1225 JT->getIndex(), getPointerTy(DAG.getDataLayout()), OpFlagHi);
1226 SDValue Lo = DAG.getTargetJumpTable(
1227 JT->getIndex(), getPointerTy(DAG.getDataLayout()), OpFlagLo);
1228 Hi = DAG.getNode(LanaiISD::HI, DL, MVT::i32, Hi);
1229 Lo = DAG.getNode(LanaiISD::LO, DL, MVT::i32, Lo);
1230 SDValue Result = DAG.getNode(ISD::OR, DL, MVT::i32, Hi, Lo);
1231 return Result;
1235 SDValue LanaiTargetLowering::LowerSHL_PARTS(SDValue Op,
1236 SelectionDAG &DAG) const {
1237 EVT VT = Op.getValueType();
1238 unsigned VTBits = VT.getSizeInBits();
1239 SDLoc dl(Op);
1240 assert(Op.getNumOperands() == 3 && "Unexpected SHL!");
1241 SDValue ShOpLo = Op.getOperand(0);
1242 SDValue ShOpHi = Op.getOperand(1);
1243 SDValue ShAmt = Op.getOperand(2);
1245 // Performs the following for (ShOpLo + (ShOpHi << 32)) << ShAmt:
1246 // LoBitsForHi = (ShAmt == 0) ? 0 : (ShOpLo >> (32-ShAmt))
1247 // HiBitsForHi = ShOpHi << ShAmt
1248 // Hi = (ShAmt >= 32) ? (ShOpLo << (ShAmt-32)) : (LoBitsForHi | HiBitsForHi)
1249 // Lo = (ShAmt >= 32) ? 0 : (ShOpLo << ShAmt)
1250 // return (Hi << 32) | Lo;
1252 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
1253 DAG.getConstant(VTBits, dl, MVT::i32), ShAmt);
1254 SDValue LoBitsForHi = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, RevShAmt);
1256 // If ShAmt == 0, we just calculated "(SRL ShOpLo, 32)" which is "undef". We
1257 // wanted 0, so CSEL it directly.
1258 SDValue Zero = DAG.getConstant(0, dl, MVT::i32);
1259 SDValue SetCC = DAG.getSetCC(dl, MVT::i32, ShAmt, Zero, ISD::SETEQ);
1260 LoBitsForHi = DAG.getSelect(dl, MVT::i32, SetCC, Zero, LoBitsForHi);
1262 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
1263 DAG.getConstant(VTBits, dl, MVT::i32));
1264 SDValue HiBitsForHi = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, ShAmt);
1265 SDValue HiForNormalShift =
1266 DAG.getNode(ISD::OR, dl, VT, LoBitsForHi, HiBitsForHi);
1268 SDValue HiForBigShift = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ExtraShAmt);
1270 SetCC = DAG.getSetCC(dl, MVT::i32, ExtraShAmt, Zero, ISD::SETGE);
1271 SDValue Hi =
1272 DAG.getSelect(dl, MVT::i32, SetCC, HiForBigShift, HiForNormalShift);
1274 // Lanai shifts of larger than register sizes are wrapped rather than
1275 // clamped, so we can't just emit "lo << b" if b is too big.
1276 SDValue LoForNormalShift = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
1277 SDValue Lo = DAG.getSelect(
1278 dl, MVT::i32, SetCC, DAG.getConstant(0, dl, MVT::i32), LoForNormalShift);
1280 SDValue Ops[2] = {Lo, Hi};
1281 return DAG.getMergeValues(Ops, dl);
1284 SDValue LanaiTargetLowering::LowerSRL_PARTS(SDValue Op,
1285 SelectionDAG &DAG) const {
1286 MVT VT = Op.getSimpleValueType();
1287 unsigned VTBits = VT.getSizeInBits();
1288 SDLoc dl(Op);
1289 SDValue ShOpLo = Op.getOperand(0);
1290 SDValue ShOpHi = Op.getOperand(1);
1291 SDValue ShAmt = Op.getOperand(2);
1293 // Performs the following for a >> b:
1294 // unsigned r_high = a_high >> b;
1295 // r_high = (32 - b <= 0) ? 0 : r_high;
1297 // unsigned r_low = a_low >> b;
1298 // r_low = (32 - b <= 0) ? r_high : r_low;
1299 // r_low = (b == 0) ? r_low : r_low | (a_high << (32 - b));
1300 // return (unsigned long long)r_high << 32 | r_low;
1301 // Note: This takes advantage of Lanai's shift behavior to avoid needing to
1302 // mask the shift amount.
1304 SDValue Zero = DAG.getConstant(0, dl, MVT::i32);
1305 SDValue NegatedPlus32 = DAG.getNode(
1306 ISD::SUB, dl, MVT::i32, DAG.getConstant(VTBits, dl, MVT::i32), ShAmt);
1307 SDValue SetCC = DAG.getSetCC(dl, MVT::i32, NegatedPlus32, Zero, ISD::SETLE);
1309 SDValue Hi = DAG.getNode(ISD::SRL, dl, MVT::i32, ShOpHi, ShAmt);
1310 Hi = DAG.getSelect(dl, MVT::i32, SetCC, Zero, Hi);
1312 SDValue Lo = DAG.getNode(ISD::SRL, dl, MVT::i32, ShOpLo, ShAmt);
1313 Lo = DAG.getSelect(dl, MVT::i32, SetCC, Hi, Lo);
1314 SDValue CarryBits =
1315 DAG.getNode(ISD::SHL, dl, MVT::i32, ShOpHi, NegatedPlus32);
1316 SDValue ShiftIsZero = DAG.getSetCC(dl, MVT::i32, ShAmt, Zero, ISD::SETEQ);
1317 Lo = DAG.getSelect(dl, MVT::i32, ShiftIsZero, Lo,
1318 DAG.getNode(ISD::OR, dl, MVT::i32, Lo, CarryBits));
1320 SDValue Ops[2] = {Lo, Hi};
1321 return DAG.getMergeValues(Ops, dl);
1324 // Helper function that checks if N is a null or all ones constant.
1325 static inline bool isZeroOrAllOnes(SDValue N, bool AllOnes) {
1326 return AllOnes ? isAllOnesConstant(N) : isNullConstant(N);
1329 // Return true if N is conditionally 0 or all ones.
1330 // Detects these expressions where cc is an i1 value:
1332 // (select cc 0, y) [AllOnes=0]
1333 // (select cc y, 0) [AllOnes=0]
1334 // (zext cc) [AllOnes=0]
1335 // (sext cc) [AllOnes=0/1]
1336 // (select cc -1, y) [AllOnes=1]
1337 // (select cc y, -1) [AllOnes=1]
1339 // * AllOnes determines whether to check for an all zero (AllOnes false) or an
1340 // all ones operand (AllOnes true).
1341 // * Invert is set when N is the all zero/ones constant when CC is false.
1342 // * OtherOp is set to the alternative value of N.
1344 // For example, for (select cc X, Y) and AllOnes = 0 if:
1345 // * X = 0, Invert = False and OtherOp = Y
1346 // * Y = 0, Invert = True and OtherOp = X
1347 static bool isConditionalZeroOrAllOnes(SDNode *N, bool AllOnes, SDValue &CC,
1348 bool &Invert, SDValue &OtherOp,
1349 SelectionDAG &DAG) {
1350 switch (N->getOpcode()) {
1351 default:
1352 return false;
1353 case ISD::SELECT: {
1354 CC = N->getOperand(0);
1355 SDValue N1 = N->getOperand(1);
1356 SDValue N2 = N->getOperand(2);
1357 if (isZeroOrAllOnes(N1, AllOnes)) {
1358 Invert = false;
1359 OtherOp = N2;
1360 return true;
1362 if (isZeroOrAllOnes(N2, AllOnes)) {
1363 Invert = true;
1364 OtherOp = N1;
1365 return true;
1367 return false;
1369 case ISD::ZERO_EXTEND: {
1370 // (zext cc) can never be the all ones value.
1371 if (AllOnes)
1372 return false;
1373 CC = N->getOperand(0);
1374 if (CC.getValueType() != MVT::i1)
1375 return false;
1376 SDLoc dl(N);
1377 EVT VT = N->getValueType(0);
1378 OtherOp = DAG.getConstant(1, dl, VT);
1379 Invert = true;
1380 return true;
1382 case ISD::SIGN_EXTEND: {
1383 CC = N->getOperand(0);
1384 if (CC.getValueType() != MVT::i1)
1385 return false;
1386 SDLoc dl(N);
1387 EVT VT = N->getValueType(0);
1388 Invert = !AllOnes;
1389 if (AllOnes)
1390 // When looking for an AllOnes constant, N is an sext, and the 'other'
1391 // value is 0.
1392 OtherOp = DAG.getConstant(0, dl, VT);
1393 else
1394 OtherOp =
1395 DAG.getConstant(APInt::getAllOnesValue(VT.getSizeInBits()), dl, VT);
1396 return true;
1401 // Combine a constant select operand into its use:
1403 // (add (select cc, 0, c), x) -> (select cc, x, (add, x, c))
1404 // (sub x, (select cc, 0, c)) -> (select cc, x, (sub, x, c))
1405 // (and (select cc, -1, c), x) -> (select cc, x, (and, x, c)) [AllOnes=1]
1406 // (or (select cc, 0, c), x) -> (select cc, x, (or, x, c))
1407 // (xor (select cc, 0, c), x) -> (select cc, x, (xor, x, c))
1409 // The transform is rejected if the select doesn't have a constant operand that
1410 // is null, or all ones when AllOnes is set.
1412 // Also recognize sext/zext from i1:
1414 // (add (zext cc), x) -> (select cc (add x, 1), x)
1415 // (add (sext cc), x) -> (select cc (add x, -1), x)
1417 // These transformations eventually create predicated instructions.
1418 static SDValue combineSelectAndUse(SDNode *N, SDValue Slct, SDValue OtherOp,
1419 TargetLowering::DAGCombinerInfo &DCI,
1420 bool AllOnes) {
1421 SelectionDAG &DAG = DCI.DAG;
1422 EVT VT = N->getValueType(0);
1423 SDValue NonConstantVal;
1424 SDValue CCOp;
1425 bool SwapSelectOps;
1426 if (!isConditionalZeroOrAllOnes(Slct.getNode(), AllOnes, CCOp, SwapSelectOps,
1427 NonConstantVal, DAG))
1428 return SDValue();
1430 // Slct is now know to be the desired identity constant when CC is true.
1431 SDValue TrueVal = OtherOp;
1432 SDValue FalseVal =
1433 DAG.getNode(N->getOpcode(), SDLoc(N), VT, OtherOp, NonConstantVal);
1434 // Unless SwapSelectOps says CC should be false.
1435 if (SwapSelectOps)
1436 std::swap(TrueVal, FalseVal);
1438 return DAG.getNode(ISD::SELECT, SDLoc(N), VT, CCOp, TrueVal, FalseVal);
1441 // Attempt combineSelectAndUse on each operand of a commutative operator N.
1442 static SDValue
1443 combineSelectAndUseCommutative(SDNode *N, TargetLowering::DAGCombinerInfo &DCI,
1444 bool AllOnes) {
1445 SDValue N0 = N->getOperand(0);
1446 SDValue N1 = N->getOperand(1);
1447 if (N0.getNode()->hasOneUse())
1448 if (SDValue Result = combineSelectAndUse(N, N0, N1, DCI, AllOnes))
1449 return Result;
1450 if (N1.getNode()->hasOneUse())
1451 if (SDValue Result = combineSelectAndUse(N, N1, N0, DCI, AllOnes))
1452 return Result;
1453 return SDValue();
1456 // PerformSUBCombine - Target-specific dag combine xforms for ISD::SUB.
1457 static SDValue PerformSUBCombine(SDNode *N,
1458 TargetLowering::DAGCombinerInfo &DCI) {
1459 SDValue N0 = N->getOperand(0);
1460 SDValue N1 = N->getOperand(1);
1462 // fold (sub x, (select cc, 0, c)) -> (select cc, x, (sub, x, c))
1463 if (N1.getNode()->hasOneUse())
1464 if (SDValue Result = combineSelectAndUse(N, N1, N0, DCI, /*AllOnes=*/false))
1465 return Result;
1467 return SDValue();
1470 SDValue LanaiTargetLowering::PerformDAGCombine(SDNode *N,
1471 DAGCombinerInfo &DCI) const {
1472 switch (N->getOpcode()) {
1473 default:
1474 break;
1475 case ISD::ADD:
1476 case ISD::OR:
1477 case ISD::XOR:
1478 return combineSelectAndUseCommutative(N, DCI, /*AllOnes=*/false);
1479 case ISD::AND:
1480 return combineSelectAndUseCommutative(N, DCI, /*AllOnes=*/true);
1481 case ISD::SUB:
1482 return PerformSUBCombine(N, DCI);
1485 return SDValue();
1488 void LanaiTargetLowering::computeKnownBitsForTargetNode(
1489 const SDValue Op, KnownBits &Known, const APInt &DemandedElts,
1490 const SelectionDAG &DAG, unsigned Depth) const {
1491 unsigned BitWidth = Known.getBitWidth();
1492 switch (Op.getOpcode()) {
1493 default:
1494 break;
1495 case LanaiISD::SETCC:
1496 Known = KnownBits(BitWidth);
1497 Known.Zero.setBits(1, BitWidth);
1498 break;
1499 case LanaiISD::SELECT_CC:
1500 KnownBits Known2;
1501 Known = DAG.computeKnownBits(Op->getOperand(0), Depth + 1);
1502 Known2 = DAG.computeKnownBits(Op->getOperand(1), Depth + 1);
1503 Known.Zero &= Known2.Zero;
1504 Known.One &= Known2.One;
1505 break;