1 //===-- RISCVISelLowering.cpp - RISCV DAG Lowering Implementation --------===//
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
7 //===----------------------------------------------------------------------===//
9 // This file defines the interfaces that RISCV uses to lower LLVM code into a
12 //===----------------------------------------------------------------------===//
14 #include "RISCVISelLowering.h"
16 #include "RISCVMachineFunctionInfo.h"
17 #include "RISCVRegisterInfo.h"
18 #include "RISCVSubtarget.h"
19 #include "RISCVTargetMachine.h"
20 #include "Utils/RISCVMatInt.h"
21 #include "llvm/ADT/SmallSet.h"
22 #include "llvm/ADT/Statistic.h"
23 #include "llvm/CodeGen/CallingConvLower.h"
24 #include "llvm/CodeGen/MachineFrameInfo.h"
25 #include "llvm/CodeGen/MachineFunction.h"
26 #include "llvm/CodeGen/MachineInstrBuilder.h"
27 #include "llvm/CodeGen/MachineRegisterInfo.h"
28 #include "llvm/CodeGen/SelectionDAGISel.h"
29 #include "llvm/CodeGen/TargetLoweringObjectFileImpl.h"
30 #include "llvm/CodeGen/ValueTypes.h"
31 #include "llvm/IR/DiagnosticInfo.h"
32 #include "llvm/IR/DiagnosticPrinter.h"
33 #include "llvm/Support/Debug.h"
34 #include "llvm/Support/ErrorHandling.h"
35 #include "llvm/Support/raw_ostream.h"
39 #define DEBUG_TYPE "riscv-lower"
41 STATISTIC(NumTailCalls, "Number of tail calls");
43 RISCVTargetLowering::RISCVTargetLowering(const TargetMachine &TM,
44 const RISCVSubtarget &STI)
45 : TargetLowering(TM), Subtarget(STI) {
47 if (Subtarget.isRV32E())
48 report_fatal_error("Codegen not yet implemented for RV32E");
50 RISCVABI::ABI ABI = Subtarget.getTargetABI();
51 assert(ABI != RISCVABI::ABI_Unknown && "Improperly initialised target ABI");
55 report_fatal_error("Don't know how to lower this ABI");
56 case RISCVABI::ABI_ILP32:
57 case RISCVABI::ABI_ILP32F:
58 case RISCVABI::ABI_ILP32D:
59 case RISCVABI::ABI_LP64:
60 case RISCVABI::ABI_LP64F:
61 case RISCVABI::ABI_LP64D:
65 MVT XLenVT = Subtarget.getXLenVT();
67 // Set up the register classes.
68 addRegisterClass(XLenVT, &RISCV::GPRRegClass);
70 if (Subtarget.hasStdExtF())
71 addRegisterClass(MVT::f32, &RISCV::FPR32RegClass);
72 if (Subtarget.hasStdExtD())
73 addRegisterClass(MVT::f64, &RISCV::FPR64RegClass);
75 // Compute derived properties from the register classes.
76 computeRegisterProperties(STI.getRegisterInfo());
78 setStackPointerRegisterToSaveRestore(RISCV::X2);
80 for (auto N : {ISD::EXTLOAD, ISD::SEXTLOAD, ISD::ZEXTLOAD})
81 setLoadExtAction(N, XLenVT, MVT::i1, Promote);
83 // TODO: add all necessary setOperationAction calls.
84 setOperationAction(ISD::DYNAMIC_STACKALLOC, XLenVT, Expand);
86 setOperationAction(ISD::BR_JT, MVT::Other, Expand);
87 setOperationAction(ISD::BR_CC, XLenVT, Expand);
88 setOperationAction(ISD::SELECT, XLenVT, Custom);
89 setOperationAction(ISD::SELECT_CC, XLenVT, Expand);
91 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
92 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
94 setOperationAction(ISD::VASTART, MVT::Other, Custom);
95 setOperationAction(ISD::VAARG, MVT::Other, Expand);
96 setOperationAction(ISD::VACOPY, MVT::Other, Expand);
97 setOperationAction(ISD::VAEND, MVT::Other, Expand);
99 for (auto VT : {MVT::i1, MVT::i8, MVT::i16})
100 setOperationAction(ISD::SIGN_EXTEND_INREG, VT, Expand);
102 if (Subtarget.is64Bit()) {
103 setOperationAction(ISD::SHL, MVT::i32, Custom);
104 setOperationAction(ISD::SRA, MVT::i32, Custom);
105 setOperationAction(ISD::SRL, MVT::i32, Custom);
108 if (!Subtarget.hasStdExtM()) {
109 setOperationAction(ISD::MUL, XLenVT, Expand);
110 setOperationAction(ISD::MULHS, XLenVT, Expand);
111 setOperationAction(ISD::MULHU, XLenVT, Expand);
112 setOperationAction(ISD::SDIV, XLenVT, Expand);
113 setOperationAction(ISD::UDIV, XLenVT, Expand);
114 setOperationAction(ISD::SREM, XLenVT, Expand);
115 setOperationAction(ISD::UREM, XLenVT, Expand);
118 if (Subtarget.is64Bit() && Subtarget.hasStdExtM()) {
119 setOperationAction(ISD::SDIV, MVT::i32, Custom);
120 setOperationAction(ISD::UDIV, MVT::i32, Custom);
121 setOperationAction(ISD::UREM, MVT::i32, Custom);
124 setOperationAction(ISD::SDIVREM, XLenVT, Expand);
125 setOperationAction(ISD::UDIVREM, XLenVT, Expand);
126 setOperationAction(ISD::SMUL_LOHI, XLenVT, Expand);
127 setOperationAction(ISD::UMUL_LOHI, XLenVT, Expand);
129 setOperationAction(ISD::SHL_PARTS, XLenVT, Custom);
130 setOperationAction(ISD::SRL_PARTS, XLenVT, Custom);
131 setOperationAction(ISD::SRA_PARTS, XLenVT, Custom);
133 setOperationAction(ISD::ROTL, XLenVT, Expand);
134 setOperationAction(ISD::ROTR, XLenVT, Expand);
135 setOperationAction(ISD::BSWAP, XLenVT, Expand);
136 setOperationAction(ISD::CTTZ, XLenVT, Expand);
137 setOperationAction(ISD::CTLZ, XLenVT, Expand);
138 setOperationAction(ISD::CTPOP, XLenVT, Expand);
140 ISD::CondCode FPCCToExtend[] = {
141 ISD::SETOGT, ISD::SETOGE, ISD::SETONE, ISD::SETUEQ, ISD::SETUGT,
142 ISD::SETUGE, ISD::SETULT, ISD::SETULE, ISD::SETUNE, ISD::SETGT,
143 ISD::SETGE, ISD::SETNE};
145 ISD::NodeType FPOpToExtend[] = {
146 ISD::FSIN, ISD::FCOS, ISD::FSINCOS, ISD::FPOW, ISD::FREM};
148 if (Subtarget.hasStdExtF()) {
149 setOperationAction(ISD::FMINNUM, MVT::f32, Legal);
150 setOperationAction(ISD::FMAXNUM, MVT::f32, Legal);
151 for (auto CC : FPCCToExtend)
152 setCondCodeAction(CC, MVT::f32, Expand);
153 setOperationAction(ISD::SELECT_CC, MVT::f32, Expand);
154 setOperationAction(ISD::SELECT, MVT::f32, Custom);
155 setOperationAction(ISD::BR_CC, MVT::f32, Expand);
156 for (auto Op : FPOpToExtend)
157 setOperationAction(Op, MVT::f32, Expand);
160 if (Subtarget.hasStdExtF() && Subtarget.is64Bit())
161 setOperationAction(ISD::BITCAST, MVT::i32, Custom);
163 if (Subtarget.hasStdExtD()) {
164 setOperationAction(ISD::FMINNUM, MVT::f64, Legal);
165 setOperationAction(ISD::FMAXNUM, MVT::f64, Legal);
166 for (auto CC : FPCCToExtend)
167 setCondCodeAction(CC, MVT::f64, Expand);
168 setOperationAction(ISD::SELECT_CC, MVT::f64, Expand);
169 setOperationAction(ISD::SELECT, MVT::f64, Custom);
170 setOperationAction(ISD::BR_CC, MVT::f64, Expand);
171 setLoadExtAction(ISD::EXTLOAD, MVT::f64, MVT::f32, Expand);
172 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
173 for (auto Op : FPOpToExtend)
174 setOperationAction(Op, MVT::f64, Expand);
177 setOperationAction(ISD::GlobalAddress, XLenVT, Custom);
178 setOperationAction(ISD::BlockAddress, XLenVT, Custom);
179 setOperationAction(ISD::ConstantPool, XLenVT, Custom);
181 setOperationAction(ISD::GlobalTLSAddress, XLenVT, Custom);
183 // TODO: On M-mode only targets, the cycle[h] CSR may not be present.
184 // Unfortunately this can't be determined just from the ISA naming string.
185 setOperationAction(ISD::READCYCLECOUNTER, MVT::i64,
186 Subtarget.is64Bit() ? Legal : Custom);
188 if (Subtarget.hasStdExtA()) {
189 setMaxAtomicSizeInBitsSupported(Subtarget.getXLen());
190 setMinCmpXchgSizeInBits(32);
192 setMaxAtomicSizeInBitsSupported(0);
195 setBooleanContents(ZeroOrOneBooleanContent);
197 // Function alignments (log2).
198 unsigned FunctionAlignment = Subtarget.hasStdExtC() ? 1 : 2;
199 setMinFunctionAlignment(FunctionAlignment);
200 setPrefFunctionAlignment(FunctionAlignment);
202 // Effectively disable jump table generation.
203 setMinimumJumpTableEntries(INT_MAX);
206 EVT RISCVTargetLowering::getSetCCResultType(const DataLayout &DL, LLVMContext &,
209 return getPointerTy(DL);
210 return VT.changeVectorElementTypeToInteger();
213 bool RISCVTargetLowering::getTgtMemIntrinsic(IntrinsicInfo &Info,
216 unsigned Intrinsic) const {
220 case Intrinsic::riscv_masked_atomicrmw_xchg_i32:
221 case Intrinsic::riscv_masked_atomicrmw_add_i32:
222 case Intrinsic::riscv_masked_atomicrmw_sub_i32:
223 case Intrinsic::riscv_masked_atomicrmw_nand_i32:
224 case Intrinsic::riscv_masked_atomicrmw_max_i32:
225 case Intrinsic::riscv_masked_atomicrmw_min_i32:
226 case Intrinsic::riscv_masked_atomicrmw_umax_i32:
227 case Intrinsic::riscv_masked_atomicrmw_umin_i32:
228 case Intrinsic::riscv_masked_cmpxchg_i32:
229 PointerType *PtrTy = cast<PointerType>(I.getArgOperand(0)->getType());
230 Info.opc = ISD::INTRINSIC_W_CHAIN;
231 Info.memVT = MVT::getVT(PtrTy->getElementType());
232 Info.ptrVal = I.getArgOperand(0);
235 Info.flags = MachineMemOperand::MOLoad | MachineMemOperand::MOStore |
236 MachineMemOperand::MOVolatile;
241 bool RISCVTargetLowering::isLegalAddressingMode(const DataLayout &DL,
242 const AddrMode &AM, Type *Ty,
244 Instruction *I) const {
245 // No global is ever allowed as a base.
249 // Require a 12-bit signed offset.
250 if (!isInt<12>(AM.BaseOffs))
254 case 0: // "r+i" or just "i", depending on HasBaseReg.
257 if (!AM.HasBaseReg) // allow "r+i".
259 return false; // disallow "r+r" or "r+r+i".
267 bool RISCVTargetLowering::isLegalICmpImmediate(int64_t Imm) const {
268 return isInt<12>(Imm);
271 bool RISCVTargetLowering::isLegalAddImmediate(int64_t Imm) const {
272 return isInt<12>(Imm);
275 // On RV32, 64-bit integers are split into their high and low parts and held
276 // in two different registers, so the trunc is free since the low register can
278 bool RISCVTargetLowering::isTruncateFree(Type *SrcTy, Type *DstTy) const {
279 if (Subtarget.is64Bit() || !SrcTy->isIntegerTy() || !DstTy->isIntegerTy())
281 unsigned SrcBits = SrcTy->getPrimitiveSizeInBits();
282 unsigned DestBits = DstTy->getPrimitiveSizeInBits();
283 return (SrcBits == 64 && DestBits == 32);
286 bool RISCVTargetLowering::isTruncateFree(EVT SrcVT, EVT DstVT) const {
287 if (Subtarget.is64Bit() || SrcVT.isVector() || DstVT.isVector() ||
288 !SrcVT.isInteger() || !DstVT.isInteger())
290 unsigned SrcBits = SrcVT.getSizeInBits();
291 unsigned DestBits = DstVT.getSizeInBits();
292 return (SrcBits == 64 && DestBits == 32);
295 bool RISCVTargetLowering::isZExtFree(SDValue Val, EVT VT2) const {
296 // Zexts are free if they can be combined with a load.
297 if (auto *LD = dyn_cast<LoadSDNode>(Val)) {
298 EVT MemVT = LD->getMemoryVT();
299 if ((MemVT == MVT::i8 || MemVT == MVT::i16 ||
300 (Subtarget.is64Bit() && MemVT == MVT::i32)) &&
301 (LD->getExtensionType() == ISD::NON_EXTLOAD ||
302 LD->getExtensionType() == ISD::ZEXTLOAD))
306 return TargetLowering::isZExtFree(Val, VT2);
309 bool RISCVTargetLowering::isSExtCheaperThanZExt(EVT SrcVT, EVT DstVT) const {
310 return Subtarget.is64Bit() && SrcVT == MVT::i32 && DstVT == MVT::i64;
313 bool RISCVTargetLowering::hasBitPreservingFPLogic(EVT VT) const {
314 return (VT == MVT::f32 && Subtarget.hasStdExtF()) ||
315 (VT == MVT::f64 && Subtarget.hasStdExtD());
318 // Changes the condition code and swaps operands if necessary, so the SetCC
319 // operation matches one of the comparisons supported directly in the RISC-V
321 static void normaliseSetCC(SDValue &LHS, SDValue &RHS, ISD::CondCode &CC) {
329 CC = ISD::getSetCCSwappedOperands(CC);
335 // Return the RISC-V branch opcode that matches the given DAG integer
336 // condition code. The CondCode must be one of those supported by the RISC-V
337 // ISA (see normaliseSetCC).
338 static unsigned getBranchOpcodeForIntCondCode(ISD::CondCode CC) {
341 llvm_unreachable("Unsupported CondCode");
357 SDValue RISCVTargetLowering::LowerOperation(SDValue Op,
358 SelectionDAG &DAG) const {
359 switch (Op.getOpcode()) {
361 report_fatal_error("unimplemented operand");
362 case ISD::GlobalAddress:
363 return lowerGlobalAddress(Op, DAG);
364 case ISD::BlockAddress:
365 return lowerBlockAddress(Op, DAG);
366 case ISD::ConstantPool:
367 return lowerConstantPool(Op, DAG);
368 case ISD::GlobalTLSAddress:
369 return lowerGlobalTLSAddress(Op, DAG);
371 return lowerSELECT(Op, DAG);
373 return lowerVASTART(Op, DAG);
375 return lowerFRAMEADDR(Op, DAG);
376 case ISD::RETURNADDR:
377 return lowerRETURNADDR(Op, DAG);
379 return lowerShiftLeftParts(Op, DAG);
381 return lowerShiftRightParts(Op, DAG, true);
383 return lowerShiftRightParts(Op, DAG, false);
385 assert(Subtarget.is64Bit() && Subtarget.hasStdExtF() &&
386 "Unexpected custom legalisation");
388 SDValue Op0 = Op.getOperand(0);
389 if (Op.getValueType() != MVT::f32 || Op0.getValueType() != MVT::i32)
391 SDValue NewOp0 = DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i64, Op0);
392 SDValue FPConv = DAG.getNode(RISCVISD::FMV_W_X_RV64, DL, MVT::f32, NewOp0);
398 static SDValue getTargetNode(GlobalAddressSDNode *N, SDLoc DL, EVT Ty,
399 SelectionDAG &DAG, unsigned Flags) {
400 return DAG.getTargetGlobalAddress(N->getGlobal(), DL, Ty, 0, Flags);
403 static SDValue getTargetNode(BlockAddressSDNode *N, SDLoc DL, EVT Ty,
404 SelectionDAG &DAG, unsigned Flags) {
405 return DAG.getTargetBlockAddress(N->getBlockAddress(), Ty, N->getOffset(),
409 static SDValue getTargetNode(ConstantPoolSDNode *N, SDLoc DL, EVT Ty,
410 SelectionDAG &DAG, unsigned Flags) {
411 return DAG.getTargetConstantPool(N->getConstVal(), Ty, N->getAlignment(),
412 N->getOffset(), Flags);
415 template <class NodeTy>
416 SDValue RISCVTargetLowering::getAddr(NodeTy *N, SelectionDAG &DAG,
417 bool IsLocal) const {
419 EVT Ty = getPointerTy(DAG.getDataLayout());
421 if (isPositionIndependent()) {
422 SDValue Addr = getTargetNode(N, DL, Ty, DAG, 0);
424 // Use PC-relative addressing to access the symbol. This generates the
425 // pattern (PseudoLLA sym), which expands to (addi (auipc %pcrel_hi(sym))
426 // %pcrel_lo(auipc)).
427 return SDValue(DAG.getMachineNode(RISCV::PseudoLLA, DL, Ty, Addr), 0);
429 // Use PC-relative addressing to access the GOT for this symbol, then load
430 // the address from the GOT. This generates the pattern (PseudoLA sym),
431 // which expands to (ld (addi (auipc %got_pcrel_hi(sym)) %pcrel_lo(auipc))).
432 return SDValue(DAG.getMachineNode(RISCV::PseudoLA, DL, Ty, Addr), 0);
435 switch (getTargetMachine().getCodeModel()) {
437 report_fatal_error("Unsupported code model for lowering");
438 case CodeModel::Small: {
439 // Generate a sequence for accessing addresses within the first 2 GiB of
440 // address space. This generates the pattern (addi (lui %hi(sym)) %lo(sym)).
441 SDValue AddrHi = getTargetNode(N, DL, Ty, DAG, RISCVII::MO_HI);
442 SDValue AddrLo = getTargetNode(N, DL, Ty, DAG, RISCVII::MO_LO);
443 SDValue MNHi = SDValue(DAG.getMachineNode(RISCV::LUI, DL, Ty, AddrHi), 0);
444 return SDValue(DAG.getMachineNode(RISCV::ADDI, DL, Ty, MNHi, AddrLo), 0);
446 case CodeModel::Medium: {
447 // Generate a sequence for accessing addresses within any 2GiB range within
448 // the address space. This generates the pattern (PseudoLLA sym), which
449 // expands to (addi (auipc %pcrel_hi(sym)) %pcrel_lo(auipc)).
450 SDValue Addr = getTargetNode(N, DL, Ty, DAG, 0);
451 return SDValue(DAG.getMachineNode(RISCV::PseudoLLA, DL, Ty, Addr), 0);
456 SDValue RISCVTargetLowering::lowerGlobalAddress(SDValue Op,
457 SelectionDAG &DAG) const {
459 EVT Ty = Op.getValueType();
460 GlobalAddressSDNode *N = cast<GlobalAddressSDNode>(Op);
461 int64_t Offset = N->getOffset();
462 MVT XLenVT = Subtarget.getXLenVT();
464 const GlobalValue *GV = N->getGlobal();
465 bool IsLocal = getTargetMachine().shouldAssumeDSOLocal(*GV->getParent(), GV);
466 SDValue Addr = getAddr(N, DAG, IsLocal);
468 // In order to maximise the opportunity for common subexpression elimination,
469 // emit a separate ADD node for the global address offset instead of folding
470 // it in the global address node. Later peephole optimisations may choose to
471 // fold it back in when profitable.
473 return DAG.getNode(ISD::ADD, DL, Ty, Addr,
474 DAG.getConstant(Offset, DL, XLenVT));
478 SDValue RISCVTargetLowering::lowerBlockAddress(SDValue Op,
479 SelectionDAG &DAG) const {
480 BlockAddressSDNode *N = cast<BlockAddressSDNode>(Op);
482 return getAddr(N, DAG);
485 SDValue RISCVTargetLowering::lowerConstantPool(SDValue Op,
486 SelectionDAG &DAG) const {
487 ConstantPoolSDNode *N = cast<ConstantPoolSDNode>(Op);
489 return getAddr(N, DAG);
492 SDValue RISCVTargetLowering::getStaticTLSAddr(GlobalAddressSDNode *N,
496 EVT Ty = getPointerTy(DAG.getDataLayout());
497 const GlobalValue *GV = N->getGlobal();
498 MVT XLenVT = Subtarget.getXLenVT();
501 // Use PC-relative addressing to access the GOT for this TLS symbol, then
502 // load the address from the GOT and add the thread pointer. This generates
503 // the pattern (PseudoLA_TLS_IE sym), which expands to
504 // (ld (auipc %tls_ie_pcrel_hi(sym)) %pcrel_lo(auipc)).
505 SDValue Addr = DAG.getTargetGlobalAddress(GV, DL, Ty, 0, 0);
507 SDValue(DAG.getMachineNode(RISCV::PseudoLA_TLS_IE, DL, Ty, Addr), 0);
509 // Add the thread pointer.
510 SDValue TPReg = DAG.getRegister(RISCV::X4, XLenVT);
511 return DAG.getNode(ISD::ADD, DL, Ty, Load, TPReg);
514 // Generate a sequence for accessing the address relative to the thread
515 // pointer, with the appropriate adjustment for the thread pointer offset.
516 // This generates the pattern
517 // (add (add_tprel (lui %tprel_hi(sym)) tp %tprel_add(sym)) %tprel_lo(sym))
519 DAG.getTargetGlobalAddress(GV, DL, Ty, 0, RISCVII::MO_TPREL_HI);
521 DAG.getTargetGlobalAddress(GV, DL, Ty, 0, RISCVII::MO_TPREL_ADD);
523 DAG.getTargetGlobalAddress(GV, DL, Ty, 0, RISCVII::MO_TPREL_LO);
525 SDValue MNHi = SDValue(DAG.getMachineNode(RISCV::LUI, DL, Ty, AddrHi), 0);
526 SDValue TPReg = DAG.getRegister(RISCV::X4, XLenVT);
527 SDValue MNAdd = SDValue(
528 DAG.getMachineNode(RISCV::PseudoAddTPRel, DL, Ty, MNHi, TPReg, AddrAdd),
530 return SDValue(DAG.getMachineNode(RISCV::ADDI, DL, Ty, MNAdd, AddrLo), 0);
533 SDValue RISCVTargetLowering::getDynamicTLSAddr(GlobalAddressSDNode *N,
534 SelectionDAG &DAG) const {
536 EVT Ty = getPointerTy(DAG.getDataLayout());
537 IntegerType *CallTy = Type::getIntNTy(*DAG.getContext(), Ty.getSizeInBits());
538 const GlobalValue *GV = N->getGlobal();
540 // Use a PC-relative addressing mode to access the global dynamic GOT address.
541 // This generates the pattern (PseudoLA_TLS_GD sym), which expands to
542 // (addi (auipc %tls_gd_pcrel_hi(sym)) %pcrel_lo(auipc)).
543 SDValue Addr = DAG.getTargetGlobalAddress(GV, DL, Ty, 0, 0);
545 SDValue(DAG.getMachineNode(RISCV::PseudoLA_TLS_GD, DL, Ty, Addr), 0);
547 // Prepare argument list to generate call.
552 Args.push_back(Entry);
554 // Setup call to __tls_get_addr.
555 TargetLowering::CallLoweringInfo CLI(DAG);
557 .setChain(DAG.getEntryNode())
558 .setLibCallee(CallingConv::C, CallTy,
559 DAG.getExternalSymbol("__tls_get_addr", Ty),
562 return LowerCallTo(CLI).first;
565 SDValue RISCVTargetLowering::lowerGlobalTLSAddress(SDValue Op,
566 SelectionDAG &DAG) const {
568 EVT Ty = Op.getValueType();
569 GlobalAddressSDNode *N = cast<GlobalAddressSDNode>(Op);
570 int64_t Offset = N->getOffset();
571 MVT XLenVT = Subtarget.getXLenVT();
573 TLSModel::Model Model = getTargetMachine().getTLSModel(N->getGlobal());
577 case TLSModel::LocalExec:
578 Addr = getStaticTLSAddr(N, DAG, /*UseGOT=*/false);
580 case TLSModel::InitialExec:
581 Addr = getStaticTLSAddr(N, DAG, /*UseGOT=*/true);
583 case TLSModel::LocalDynamic:
584 case TLSModel::GeneralDynamic:
585 Addr = getDynamicTLSAddr(N, DAG);
589 // In order to maximise the opportunity for common subexpression elimination,
590 // emit a separate ADD node for the global address offset instead of folding
591 // it in the global address node. Later peephole optimisations may choose to
592 // fold it back in when profitable.
594 return DAG.getNode(ISD::ADD, DL, Ty, Addr,
595 DAG.getConstant(Offset, DL, XLenVT));
599 SDValue RISCVTargetLowering::lowerSELECT(SDValue Op, SelectionDAG &DAG) const {
600 SDValue CondV = Op.getOperand(0);
601 SDValue TrueV = Op.getOperand(1);
602 SDValue FalseV = Op.getOperand(2);
604 MVT XLenVT = Subtarget.getXLenVT();
606 // If the result type is XLenVT and CondV is the output of a SETCC node
607 // which also operated on XLenVT inputs, then merge the SETCC node into the
608 // lowered RISCVISD::SELECT_CC to take advantage of the integer
609 // compare+branch instructions. i.e.:
610 // (select (setcc lhs, rhs, cc), truev, falsev)
611 // -> (riscvisd::select_cc lhs, rhs, cc, truev, falsev)
612 if (Op.getSimpleValueType() == XLenVT && CondV.getOpcode() == ISD::SETCC &&
613 CondV.getOperand(0).getSimpleValueType() == XLenVT) {
614 SDValue LHS = CondV.getOperand(0);
615 SDValue RHS = CondV.getOperand(1);
616 auto CC = cast<CondCodeSDNode>(CondV.getOperand(2));
617 ISD::CondCode CCVal = CC->get();
619 normaliseSetCC(LHS, RHS, CCVal);
621 SDValue TargetCC = DAG.getConstant(CCVal, DL, XLenVT);
622 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::Glue);
623 SDValue Ops[] = {LHS, RHS, TargetCC, TrueV, FalseV};
624 return DAG.getNode(RISCVISD::SELECT_CC, DL, VTs, Ops);
628 // (select condv, truev, falsev)
629 // -> (riscvisd::select_cc condv, zero, setne, truev, falsev)
630 SDValue Zero = DAG.getConstant(0, DL, XLenVT);
631 SDValue SetNE = DAG.getConstant(ISD::SETNE, DL, XLenVT);
633 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::Glue);
634 SDValue Ops[] = {CondV, Zero, SetNE, TrueV, FalseV};
636 return DAG.getNode(RISCVISD::SELECT_CC, DL, VTs, Ops);
639 SDValue RISCVTargetLowering::lowerVASTART(SDValue Op, SelectionDAG &DAG) const {
640 MachineFunction &MF = DAG.getMachineFunction();
641 RISCVMachineFunctionInfo *FuncInfo = MF.getInfo<RISCVMachineFunctionInfo>();
644 SDValue FI = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(),
645 getPointerTy(MF.getDataLayout()));
647 // vastart just stores the address of the VarArgsFrameIndex slot into the
648 // memory location argument.
649 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
650 return DAG.getStore(Op.getOperand(0), DL, FI, Op.getOperand(1),
651 MachinePointerInfo(SV));
654 SDValue RISCVTargetLowering::lowerFRAMEADDR(SDValue Op,
655 SelectionDAG &DAG) const {
656 const RISCVRegisterInfo &RI = *Subtarget.getRegisterInfo();
657 MachineFunction &MF = DAG.getMachineFunction();
658 MachineFrameInfo &MFI = MF.getFrameInfo();
659 MFI.setFrameAddressIsTaken(true);
660 unsigned FrameReg = RI.getFrameRegister(MF);
661 int XLenInBytes = Subtarget.getXLen() / 8;
663 EVT VT = Op.getValueType();
665 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), DL, FrameReg, VT);
666 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
668 int Offset = -(XLenInBytes * 2);
669 SDValue Ptr = DAG.getNode(ISD::ADD, DL, VT, FrameAddr,
670 DAG.getIntPtrConstant(Offset, DL));
672 DAG.getLoad(VT, DL, DAG.getEntryNode(), Ptr, MachinePointerInfo());
677 SDValue RISCVTargetLowering::lowerRETURNADDR(SDValue Op,
678 SelectionDAG &DAG) const {
679 const RISCVRegisterInfo &RI = *Subtarget.getRegisterInfo();
680 MachineFunction &MF = DAG.getMachineFunction();
681 MachineFrameInfo &MFI = MF.getFrameInfo();
682 MFI.setReturnAddressIsTaken(true);
683 MVT XLenVT = Subtarget.getXLenVT();
684 int XLenInBytes = Subtarget.getXLen() / 8;
686 if (verifyReturnAddressArgumentIsConstant(Op, DAG))
689 EVT VT = Op.getValueType();
691 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
693 int Off = -XLenInBytes;
694 SDValue FrameAddr = lowerFRAMEADDR(Op, DAG);
695 SDValue Offset = DAG.getConstant(Off, DL, VT);
696 return DAG.getLoad(VT, DL, DAG.getEntryNode(),
697 DAG.getNode(ISD::ADD, DL, VT, FrameAddr, Offset),
698 MachinePointerInfo());
701 // Return the value of the return address register, marking it an implicit
703 unsigned Reg = MF.addLiveIn(RI.getRARegister(), getRegClassFor(XLenVT));
704 return DAG.getCopyFromReg(DAG.getEntryNode(), DL, Reg, XLenVT);
707 SDValue RISCVTargetLowering::lowerShiftLeftParts(SDValue Op,
708 SelectionDAG &DAG) const {
710 SDValue Lo = Op.getOperand(0);
711 SDValue Hi = Op.getOperand(1);
712 SDValue Shamt = Op.getOperand(2);
713 EVT VT = Lo.getValueType();
715 // if Shamt-XLEN < 0: // Shamt < XLEN
717 // Hi = (Hi << Shamt) | ((Lo >>u 1) >>u (XLEN-1 - Shamt))
720 // Hi = Lo << (Shamt-XLEN)
722 SDValue Zero = DAG.getConstant(0, DL, VT);
723 SDValue One = DAG.getConstant(1, DL, VT);
724 SDValue MinusXLen = DAG.getConstant(-(int)Subtarget.getXLen(), DL, VT);
725 SDValue XLenMinus1 = DAG.getConstant(Subtarget.getXLen() - 1, DL, VT);
726 SDValue ShamtMinusXLen = DAG.getNode(ISD::ADD, DL, VT, Shamt, MinusXLen);
727 SDValue XLenMinus1Shamt = DAG.getNode(ISD::SUB, DL, VT, XLenMinus1, Shamt);
729 SDValue LoTrue = DAG.getNode(ISD::SHL, DL, VT, Lo, Shamt);
730 SDValue ShiftRight1Lo = DAG.getNode(ISD::SRL, DL, VT, Lo, One);
731 SDValue ShiftRightLo =
732 DAG.getNode(ISD::SRL, DL, VT, ShiftRight1Lo, XLenMinus1Shamt);
733 SDValue ShiftLeftHi = DAG.getNode(ISD::SHL, DL, VT, Hi, Shamt);
734 SDValue HiTrue = DAG.getNode(ISD::OR, DL, VT, ShiftLeftHi, ShiftRightLo);
735 SDValue HiFalse = DAG.getNode(ISD::SHL, DL, VT, Lo, ShamtMinusXLen);
737 SDValue CC = DAG.getSetCC(DL, VT, ShamtMinusXLen, Zero, ISD::SETLT);
739 Lo = DAG.getNode(ISD::SELECT, DL, VT, CC, LoTrue, Zero);
740 Hi = DAG.getNode(ISD::SELECT, DL, VT, CC, HiTrue, HiFalse);
742 SDValue Parts[2] = {Lo, Hi};
743 return DAG.getMergeValues(Parts, DL);
746 SDValue RISCVTargetLowering::lowerShiftRightParts(SDValue Op, SelectionDAG &DAG,
749 SDValue Lo = Op.getOperand(0);
750 SDValue Hi = Op.getOperand(1);
751 SDValue Shamt = Op.getOperand(2);
752 EVT VT = Lo.getValueType();
755 // if Shamt-XLEN < 0: // Shamt < XLEN
756 // Lo = (Lo >>u Shamt) | ((Hi << 1) << (XLEN-1 - Shamt))
759 // Lo = Hi >>s (Shamt-XLEN);
760 // Hi = Hi >>s (XLEN-1)
763 // if Shamt-XLEN < 0: // Shamt < XLEN
764 // Lo = (Lo >>u Shamt) | ((Hi << 1) << (XLEN-1 - Shamt))
767 // Lo = Hi >>u (Shamt-XLEN);
770 unsigned ShiftRightOp = IsSRA ? ISD::SRA : ISD::SRL;
772 SDValue Zero = DAG.getConstant(0, DL, VT);
773 SDValue One = DAG.getConstant(1, DL, VT);
774 SDValue MinusXLen = DAG.getConstant(-(int)Subtarget.getXLen(), DL, VT);
775 SDValue XLenMinus1 = DAG.getConstant(Subtarget.getXLen() - 1, DL, VT);
776 SDValue ShamtMinusXLen = DAG.getNode(ISD::ADD, DL, VT, Shamt, MinusXLen);
777 SDValue XLenMinus1Shamt = DAG.getNode(ISD::SUB, DL, VT, XLenMinus1, Shamt);
779 SDValue ShiftRightLo = DAG.getNode(ISD::SRL, DL, VT, Lo, Shamt);
780 SDValue ShiftLeftHi1 = DAG.getNode(ISD::SHL, DL, VT, Hi, One);
781 SDValue ShiftLeftHi =
782 DAG.getNode(ISD::SHL, DL, VT, ShiftLeftHi1, XLenMinus1Shamt);
783 SDValue LoTrue = DAG.getNode(ISD::OR, DL, VT, ShiftRightLo, ShiftLeftHi);
784 SDValue HiTrue = DAG.getNode(ShiftRightOp, DL, VT, Hi, Shamt);
785 SDValue LoFalse = DAG.getNode(ShiftRightOp, DL, VT, Hi, ShamtMinusXLen);
787 IsSRA ? DAG.getNode(ISD::SRA, DL, VT, Hi, XLenMinus1) : Zero;
789 SDValue CC = DAG.getSetCC(DL, VT, ShamtMinusXLen, Zero, ISD::SETLT);
791 Lo = DAG.getNode(ISD::SELECT, DL, VT, CC, LoTrue, LoFalse);
792 Hi = DAG.getNode(ISD::SELECT, DL, VT, CC, HiTrue, HiFalse);
794 SDValue Parts[2] = {Lo, Hi};
795 return DAG.getMergeValues(Parts, DL);
798 // Returns the opcode of the target-specific SDNode that implements the 32-bit
799 // form of the given Opcode.
800 static RISCVISD::NodeType getRISCVWOpcode(unsigned Opcode) {
803 llvm_unreachable("Unexpected opcode");
805 return RISCVISD::SLLW;
807 return RISCVISD::SRAW;
809 return RISCVISD::SRLW;
811 return RISCVISD::DIVW;
813 return RISCVISD::DIVUW;
815 return RISCVISD::REMUW;
819 // Converts the given 32-bit operation to a target-specific SelectionDAG node.
820 // Because i32 isn't a legal type for RV64, these operations would otherwise
821 // be promoted to i64, making it difficult to select the SLLW/DIVUW/.../*W
822 // later one because the fact the operation was originally of type i32 is
824 static SDValue customLegalizeToWOp(SDNode *N, SelectionDAG &DAG) {
826 RISCVISD::NodeType WOpcode = getRISCVWOpcode(N->getOpcode());
827 SDValue NewOp0 = DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i64, N->getOperand(0));
828 SDValue NewOp1 = DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i64, N->getOperand(1));
829 SDValue NewRes = DAG.getNode(WOpcode, DL, MVT::i64, NewOp0, NewOp1);
830 // ReplaceNodeResults requires we maintain the same type for the return value.
831 return DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, NewRes);
834 void RISCVTargetLowering::ReplaceNodeResults(SDNode *N,
835 SmallVectorImpl<SDValue> &Results,
836 SelectionDAG &DAG) const {
838 switch (N->getOpcode()) {
840 llvm_unreachable("Don't know how to custom type legalize this operation!");
841 case ISD::READCYCLECOUNTER: {
842 assert(!Subtarget.is64Bit() &&
843 "READCYCLECOUNTER only has custom type legalization on riscv32");
845 SDVTList VTs = DAG.getVTList(MVT::i32, MVT::i32, MVT::Other);
847 DAG.getNode(RISCVISD::READ_CYCLE_WIDE, DL, VTs, N->getOperand(0));
849 Results.push_back(RCW);
850 Results.push_back(RCW.getValue(1));
851 Results.push_back(RCW.getValue(2));
857 assert(N->getValueType(0) == MVT::i32 && Subtarget.is64Bit() &&
858 "Unexpected custom legalisation");
859 if (N->getOperand(1).getOpcode() == ISD::Constant)
861 Results.push_back(customLegalizeToWOp(N, DAG));
866 assert(N->getValueType(0) == MVT::i32 && Subtarget.is64Bit() &&
867 Subtarget.hasStdExtM() && "Unexpected custom legalisation");
868 if (N->getOperand(0).getOpcode() == ISD::Constant ||
869 N->getOperand(1).getOpcode() == ISD::Constant)
871 Results.push_back(customLegalizeToWOp(N, DAG));
874 assert(N->getValueType(0) == MVT::i32 && Subtarget.is64Bit() &&
875 Subtarget.hasStdExtF() && "Unexpected custom legalisation");
877 SDValue Op0 = N->getOperand(0);
878 if (Op0.getValueType() != MVT::f32)
881 DAG.getNode(RISCVISD::FMV_X_ANYEXTW_RV64, DL, MVT::i64, Op0);
882 Results.push_back(DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, FPConv));
888 SDValue RISCVTargetLowering::PerformDAGCombine(SDNode *N,
889 DAGCombinerInfo &DCI) const {
890 SelectionDAG &DAG = DCI.DAG;
892 switch (N->getOpcode()) {
895 case RISCVISD::SplitF64: {
896 SDValue Op0 = N->getOperand(0);
897 // If the input to SplitF64 is just BuildPairF64 then the operation is
898 // redundant. Instead, use BuildPairF64's operands directly.
899 if (Op0->getOpcode() == RISCVISD::BuildPairF64)
900 return DCI.CombineTo(N, Op0.getOperand(0), Op0.getOperand(1));
904 // It's cheaper to materialise two 32-bit integers than to load a double
905 // from the constant pool and transfer it to integer registers through the
907 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(Op0)) {
908 APInt V = C->getValueAPF().bitcastToAPInt();
909 SDValue Lo = DAG.getConstant(V.trunc(32), DL, MVT::i32);
910 SDValue Hi = DAG.getConstant(V.lshr(32).trunc(32), DL, MVT::i32);
911 return DCI.CombineTo(N, Lo, Hi);
914 // This is a target-specific version of a DAGCombine performed in
915 // DAGCombiner::visitBITCAST. It performs the equivalent of:
916 // fold (bitconvert (fneg x)) -> (xor (bitconvert x), signbit)
917 // fold (bitconvert (fabs x)) -> (and (bitconvert x), (not signbit))
918 if (!(Op0.getOpcode() == ISD::FNEG || Op0.getOpcode() == ISD::FABS) ||
919 !Op0.getNode()->hasOneUse())
921 SDValue NewSplitF64 =
922 DAG.getNode(RISCVISD::SplitF64, DL, DAG.getVTList(MVT::i32, MVT::i32),
924 SDValue Lo = NewSplitF64.getValue(0);
925 SDValue Hi = NewSplitF64.getValue(1);
926 APInt SignBit = APInt::getSignMask(32);
927 if (Op0.getOpcode() == ISD::FNEG) {
928 SDValue NewHi = DAG.getNode(ISD::XOR, DL, MVT::i32, Hi,
929 DAG.getConstant(SignBit, DL, MVT::i32));
930 return DCI.CombineTo(N, Lo, NewHi);
932 assert(Op0.getOpcode() == ISD::FABS);
933 SDValue NewHi = DAG.getNode(ISD::AND, DL, MVT::i32, Hi,
934 DAG.getConstant(~SignBit, DL, MVT::i32));
935 return DCI.CombineTo(N, Lo, NewHi);
939 case RISCVISD::SRLW: {
940 // Only the lower 32 bits of LHS and lower 5 bits of RHS are read.
941 SDValue LHS = N->getOperand(0);
942 SDValue RHS = N->getOperand(1);
943 APInt LHSMask = APInt::getLowBitsSet(LHS.getValueSizeInBits(), 32);
944 APInt RHSMask = APInt::getLowBitsSet(RHS.getValueSizeInBits(), 5);
945 if ((SimplifyDemandedBits(N->getOperand(0), LHSMask, DCI)) ||
946 (SimplifyDemandedBits(N->getOperand(1), RHSMask, DCI)))
950 case RISCVISD::FMV_X_ANYEXTW_RV64: {
952 SDValue Op0 = N->getOperand(0);
953 // If the input to FMV_X_ANYEXTW_RV64 is just FMV_W_X_RV64 then the
954 // conversion is unnecessary and can be replaced with an ANY_EXTEND
955 // of the FMV_W_X_RV64 operand.
956 if (Op0->getOpcode() == RISCVISD::FMV_W_X_RV64) {
958 DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i64, Op0.getOperand(0));
959 return DCI.CombineTo(N, AExtOp);
962 // This is a target-specific version of a DAGCombine performed in
963 // DAGCombiner::visitBITCAST. It performs the equivalent of:
964 // fold (bitconvert (fneg x)) -> (xor (bitconvert x), signbit)
965 // fold (bitconvert (fabs x)) -> (and (bitconvert x), (not signbit))
966 if (!(Op0.getOpcode() == ISD::FNEG || Op0.getOpcode() == ISD::FABS) ||
967 !Op0.getNode()->hasOneUse())
969 SDValue NewFMV = DAG.getNode(RISCVISD::FMV_X_ANYEXTW_RV64, DL, MVT::i64,
971 APInt SignBit = APInt::getSignMask(32).sext(64);
972 if (Op0.getOpcode() == ISD::FNEG) {
973 return DCI.CombineTo(N,
974 DAG.getNode(ISD::XOR, DL, MVT::i64, NewFMV,
975 DAG.getConstant(SignBit, DL, MVT::i64)));
977 assert(Op0.getOpcode() == ISD::FABS);
978 return DCI.CombineTo(N,
979 DAG.getNode(ISD::AND, DL, MVT::i64, NewFMV,
980 DAG.getConstant(~SignBit, DL, MVT::i64)));
987 bool RISCVTargetLowering::isDesirableToCommuteWithShift(
988 const SDNode *N, CombineLevel Level) const {
989 // The following folds are only desirable if `(OP _, c1 << c2)` can be
990 // materialised in fewer instructions than `(OP _, c1)`:
992 // (shl (add x, c1), c2) -> (add (shl x, c2), c1 << c2)
993 // (shl (or x, c1), c2) -> (or (shl x, c2), c1 << c2)
994 SDValue N0 = N->getOperand(0);
995 EVT Ty = N0.getValueType();
996 if (Ty.isScalarInteger() &&
997 (N0.getOpcode() == ISD::ADD || N0.getOpcode() == ISD::OR)) {
998 auto *C1 = dyn_cast<ConstantSDNode>(N0->getOperand(1));
999 auto *C2 = dyn_cast<ConstantSDNode>(N->getOperand(1));
1001 APInt C1Int = C1->getAPIntValue();
1002 APInt ShiftedC1Int = C1Int << C2->getAPIntValue();
1004 // We can materialise `c1 << c2` into an add immediate, so it's "free",
1005 // and the combine should happen, to potentially allow further combines
1007 if (ShiftedC1Int.getMinSignedBits() <= 64 &&
1008 isLegalAddImmediate(ShiftedC1Int.getSExtValue()))
1011 // We can materialise `c1` in an add immediate, so it's "free", and the
1012 // combine should be prevented.
1013 if (C1Int.getMinSignedBits() <= 64 &&
1014 isLegalAddImmediate(C1Int.getSExtValue()))
1017 // Neither constant will fit into an immediate, so find materialisation
1019 int C1Cost = RISCVMatInt::getIntMatCost(C1Int, Ty.getSizeInBits(),
1020 Subtarget.is64Bit());
1021 int ShiftedC1Cost = RISCVMatInt::getIntMatCost(
1022 ShiftedC1Int, Ty.getSizeInBits(), Subtarget.is64Bit());
1024 // Materialising `c1` is cheaper than materialising `c1 << c2`, so the
1025 // combine should be prevented.
1026 if (C1Cost < ShiftedC1Cost)
1033 unsigned RISCVTargetLowering::ComputeNumSignBitsForTargetNode(
1034 SDValue Op, const APInt &DemandedElts, const SelectionDAG &DAG,
1035 unsigned Depth) const {
1036 switch (Op.getOpcode()) {
1039 case RISCVISD::SLLW:
1040 case RISCVISD::SRAW:
1041 case RISCVISD::SRLW:
1042 case RISCVISD::DIVW:
1043 case RISCVISD::DIVUW:
1044 case RISCVISD::REMUW:
1045 // TODO: As the result is sign-extended, this is conservatively correct. A
1046 // more precise answer could be calculated for SRAW depending on known
1047 // bits in the shift amount.
1054 MachineBasicBlock *emitReadCycleWidePseudo(MachineInstr &MI,
1055 MachineBasicBlock *BB) {
1056 assert(MI.getOpcode() == RISCV::ReadCycleWide && "Unexpected instruction");
1058 // To read the 64-bit cycle CSR on a 32-bit target, we read the two halves.
1059 // Should the count have wrapped while it was being read, we need to try
1063 // rdcycleh x3 # load high word of cycle
1064 // rdcycle x2 # load low word of cycle
1065 // rdcycleh x4 # load high word of cycle
1066 // bne x3, x4, read # check if high word reads match, otherwise try again
1069 MachineFunction &MF = *BB->getParent();
1070 const BasicBlock *LLVM_BB = BB->getBasicBlock();
1071 MachineFunction::iterator It = ++BB->getIterator();
1073 MachineBasicBlock *LoopMBB = MF.CreateMachineBasicBlock(LLVM_BB);
1074 MF.insert(It, LoopMBB);
1076 MachineBasicBlock *DoneMBB = MF.CreateMachineBasicBlock(LLVM_BB);
1077 MF.insert(It, DoneMBB);
1079 // Transfer the remainder of BB and its successor edges to DoneMBB.
1080 DoneMBB->splice(DoneMBB->begin(), BB,
1081 std::next(MachineBasicBlock::iterator(MI)), BB->end());
1082 DoneMBB->transferSuccessorsAndUpdatePHIs(BB);
1084 BB->addSuccessor(LoopMBB);
1086 MachineRegisterInfo &RegInfo = MF.getRegInfo();
1087 unsigned ReadAgainReg = RegInfo.createVirtualRegister(&RISCV::GPRRegClass);
1088 unsigned LoReg = MI.getOperand(0).getReg();
1089 unsigned HiReg = MI.getOperand(1).getReg();
1090 DebugLoc DL = MI.getDebugLoc();
1092 const TargetInstrInfo *TII = MF.getSubtarget().getInstrInfo();
1093 BuildMI(LoopMBB, DL, TII->get(RISCV::CSRRS), HiReg)
1094 .addImm(RISCVSysReg::lookupSysRegByName("CYCLEH")->Encoding)
1096 BuildMI(LoopMBB, DL, TII->get(RISCV::CSRRS), LoReg)
1097 .addImm(RISCVSysReg::lookupSysRegByName("CYCLE")->Encoding)
1099 BuildMI(LoopMBB, DL, TII->get(RISCV::CSRRS), ReadAgainReg)
1100 .addImm(RISCVSysReg::lookupSysRegByName("CYCLEH")->Encoding)
1103 BuildMI(LoopMBB, DL, TII->get(RISCV::BNE))
1105 .addReg(ReadAgainReg)
1108 LoopMBB->addSuccessor(LoopMBB);
1109 LoopMBB->addSuccessor(DoneMBB);
1111 MI.eraseFromParent();
1116 static MachineBasicBlock *emitSplitF64Pseudo(MachineInstr &MI,
1117 MachineBasicBlock *BB) {
1118 assert(MI.getOpcode() == RISCV::SplitF64Pseudo && "Unexpected instruction");
1120 MachineFunction &MF = *BB->getParent();
1121 DebugLoc DL = MI.getDebugLoc();
1122 const TargetInstrInfo &TII = *MF.getSubtarget().getInstrInfo();
1123 const TargetRegisterInfo *RI = MF.getSubtarget().getRegisterInfo();
1124 unsigned LoReg = MI.getOperand(0).getReg();
1125 unsigned HiReg = MI.getOperand(1).getReg();
1126 unsigned SrcReg = MI.getOperand(2).getReg();
1127 const TargetRegisterClass *SrcRC = &RISCV::FPR64RegClass;
1128 int FI = MF.getInfo<RISCVMachineFunctionInfo>()->getMoveF64FrameIndex();
1130 TII.storeRegToStackSlot(*BB, MI, SrcReg, MI.getOperand(2).isKill(), FI, SrcRC,
1132 MachineMemOperand *MMO =
1133 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(MF, FI),
1134 MachineMemOperand::MOLoad, 8, 8);
1135 BuildMI(*BB, MI, DL, TII.get(RISCV::LW), LoReg)
1138 .addMemOperand(MMO);
1139 BuildMI(*BB, MI, DL, TII.get(RISCV::LW), HiReg)
1142 .addMemOperand(MMO);
1143 MI.eraseFromParent(); // The pseudo instruction is gone now.
1147 static MachineBasicBlock *emitBuildPairF64Pseudo(MachineInstr &MI,
1148 MachineBasicBlock *BB) {
1149 assert(MI.getOpcode() == RISCV::BuildPairF64Pseudo &&
1150 "Unexpected instruction");
1152 MachineFunction &MF = *BB->getParent();
1153 DebugLoc DL = MI.getDebugLoc();
1154 const TargetInstrInfo &TII = *MF.getSubtarget().getInstrInfo();
1155 const TargetRegisterInfo *RI = MF.getSubtarget().getRegisterInfo();
1156 unsigned DstReg = MI.getOperand(0).getReg();
1157 unsigned LoReg = MI.getOperand(1).getReg();
1158 unsigned HiReg = MI.getOperand(2).getReg();
1159 const TargetRegisterClass *DstRC = &RISCV::FPR64RegClass;
1160 int FI = MF.getInfo<RISCVMachineFunctionInfo>()->getMoveF64FrameIndex();
1162 MachineMemOperand *MMO =
1163 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(MF, FI),
1164 MachineMemOperand::MOStore, 8, 8);
1165 BuildMI(*BB, MI, DL, TII.get(RISCV::SW))
1166 .addReg(LoReg, getKillRegState(MI.getOperand(1).isKill()))
1169 .addMemOperand(MMO);
1170 BuildMI(*BB, MI, DL, TII.get(RISCV::SW))
1171 .addReg(HiReg, getKillRegState(MI.getOperand(2).isKill()))
1174 .addMemOperand(MMO);
1175 TII.loadRegFromStackSlot(*BB, MI, DstReg, FI, DstRC, RI);
1176 MI.eraseFromParent(); // The pseudo instruction is gone now.
1180 static bool isSelectPseudo(MachineInstr &MI) {
1181 switch (MI.getOpcode()) {
1184 case RISCV::Select_GPR_Using_CC_GPR:
1185 case RISCV::Select_FPR32_Using_CC_GPR:
1186 case RISCV::Select_FPR64_Using_CC_GPR:
1191 static MachineBasicBlock *emitSelectPseudo(MachineInstr &MI,
1192 MachineBasicBlock *BB) {
1193 // To "insert" Select_* instructions, we actually have to insert the triangle
1194 // control-flow pattern. The incoming instructions know the destination vreg
1195 // to set, the condition code register to branch on, the true/false values to
1196 // select between, and the condcode to use to select the appropriate branch.
1198 // We produce the following control flow:
1205 // When we find a sequence of selects we attempt to optimize their emission
1206 // by sharing the control flow. Currently we only handle cases where we have
1207 // multiple selects with the exact same condition (same LHS, RHS and CC).
1208 // The selects may be interleaved with other instructions if the other
1209 // instructions meet some requirements we deem safe:
1210 // - They are debug instructions. Otherwise,
1211 // - They do not have side-effects, do not access memory and their inputs do
1212 // not depend on the results of the select pseudo-instructions.
1213 // The TrueV/FalseV operands of the selects cannot depend on the result of
1214 // previous selects in the sequence.
1215 // These conditions could be further relaxed. See the X86 target for a
1216 // related approach and more information.
1217 unsigned LHS = MI.getOperand(1).getReg();
1218 unsigned RHS = MI.getOperand(2).getReg();
1219 auto CC = static_cast<ISD::CondCode>(MI.getOperand(3).getImm());
1221 SmallVector<MachineInstr *, 4> SelectDebugValues;
1222 SmallSet<unsigned, 4> SelectDests;
1223 SelectDests.insert(MI.getOperand(0).getReg());
1225 MachineInstr *LastSelectPseudo = &MI;
1227 for (auto E = BB->end(), SequenceMBBI = MachineBasicBlock::iterator(MI);
1228 SequenceMBBI != E; ++SequenceMBBI) {
1229 if (SequenceMBBI->isDebugInstr())
1231 else if (isSelectPseudo(*SequenceMBBI)) {
1232 if (SequenceMBBI->getOperand(1).getReg() != LHS ||
1233 SequenceMBBI->getOperand(2).getReg() != RHS ||
1234 SequenceMBBI->getOperand(3).getImm() != CC ||
1235 SelectDests.count(SequenceMBBI->getOperand(4).getReg()) ||
1236 SelectDests.count(SequenceMBBI->getOperand(5).getReg()))
1238 LastSelectPseudo = &*SequenceMBBI;
1239 SequenceMBBI->collectDebugValues(SelectDebugValues);
1240 SelectDests.insert(SequenceMBBI->getOperand(0).getReg());
1242 if (SequenceMBBI->hasUnmodeledSideEffects() ||
1243 SequenceMBBI->mayLoadOrStore())
1245 if (llvm::any_of(SequenceMBBI->operands(), [&](MachineOperand &MO) {
1246 return MO.isReg() && MO.isUse() && SelectDests.count(MO.getReg());
1252 const TargetInstrInfo &TII = *BB->getParent()->getSubtarget().getInstrInfo();
1253 const BasicBlock *LLVM_BB = BB->getBasicBlock();
1254 DebugLoc DL = MI.getDebugLoc();
1255 MachineFunction::iterator I = ++BB->getIterator();
1257 MachineBasicBlock *HeadMBB = BB;
1258 MachineFunction *F = BB->getParent();
1259 MachineBasicBlock *TailMBB = F->CreateMachineBasicBlock(LLVM_BB);
1260 MachineBasicBlock *IfFalseMBB = F->CreateMachineBasicBlock(LLVM_BB);
1262 F->insert(I, IfFalseMBB);
1263 F->insert(I, TailMBB);
1265 // Transfer debug instructions associated with the selects to TailMBB.
1266 for (MachineInstr *DebugInstr : SelectDebugValues) {
1267 TailMBB->push_back(DebugInstr->removeFromParent());
1270 // Move all instructions after the sequence to TailMBB.
1271 TailMBB->splice(TailMBB->end(), HeadMBB,
1272 std::next(LastSelectPseudo->getIterator()), HeadMBB->end());
1273 // Update machine-CFG edges by transferring all successors of the current
1274 // block to the new block which will contain the Phi nodes for the selects.
1275 TailMBB->transferSuccessorsAndUpdatePHIs(HeadMBB);
1276 // Set the successors for HeadMBB.
1277 HeadMBB->addSuccessor(IfFalseMBB);
1278 HeadMBB->addSuccessor(TailMBB);
1280 // Insert appropriate branch.
1281 unsigned Opcode = getBranchOpcodeForIntCondCode(CC);
1283 BuildMI(HeadMBB, DL, TII.get(Opcode))
1288 // IfFalseMBB just falls through to TailMBB.
1289 IfFalseMBB->addSuccessor(TailMBB);
1291 // Create PHIs for all of the select pseudo-instructions.
1292 auto SelectMBBI = MI.getIterator();
1293 auto SelectEnd = std::next(LastSelectPseudo->getIterator());
1294 auto InsertionPoint = TailMBB->begin();
1295 while (SelectMBBI != SelectEnd) {
1296 auto Next = std::next(SelectMBBI);
1297 if (isSelectPseudo(*SelectMBBI)) {
1298 // %Result = phi [ %TrueValue, HeadMBB ], [ %FalseValue, IfFalseMBB ]
1299 BuildMI(*TailMBB, InsertionPoint, SelectMBBI->getDebugLoc(),
1300 TII.get(RISCV::PHI), SelectMBBI->getOperand(0).getReg())
1301 .addReg(SelectMBBI->getOperand(4).getReg())
1303 .addReg(SelectMBBI->getOperand(5).getReg())
1304 .addMBB(IfFalseMBB);
1305 SelectMBBI->eraseFromParent();
1310 F->getProperties().reset(MachineFunctionProperties::Property::NoPHIs);
1315 RISCVTargetLowering::EmitInstrWithCustomInserter(MachineInstr &MI,
1316 MachineBasicBlock *BB) const {
1317 switch (MI.getOpcode()) {
1319 llvm_unreachable("Unexpected instr type to insert");
1320 case RISCV::ReadCycleWide:
1321 assert(!Subtarget.is64Bit() &&
1322 "ReadCycleWrite is only to be used on riscv32");
1323 return emitReadCycleWidePseudo(MI, BB);
1324 case RISCV::Select_GPR_Using_CC_GPR:
1325 case RISCV::Select_FPR32_Using_CC_GPR:
1326 case RISCV::Select_FPR64_Using_CC_GPR:
1327 return emitSelectPseudo(MI, BB);
1328 case RISCV::BuildPairF64Pseudo:
1329 return emitBuildPairF64Pseudo(MI, BB);
1330 case RISCV::SplitF64Pseudo:
1331 return emitSplitF64Pseudo(MI, BB);
1335 // Calling Convention Implementation.
1336 // The expectations for frontend ABI lowering vary from target to target.
1337 // Ideally, an LLVM frontend would be able to avoid worrying about many ABI
1338 // details, but this is a longer term goal. For now, we simply try to keep the
1339 // role of the frontend as simple and well-defined as possible. The rules can
1340 // be summarised as:
1341 // * Never split up large scalar arguments. We handle them here.
1342 // * If a hardfloat calling convention is being used, and the struct may be
1343 // passed in a pair of registers (fp+fp, int+fp), and both registers are
1344 // available, then pass as two separate arguments. If either the GPRs or FPRs
1345 // are exhausted, then pass according to the rule below.
1346 // * If a struct could never be passed in registers or directly in a stack
1347 // slot (as it is larger than 2*XLEN and the floating point rules don't
1348 // apply), then pass it using a pointer with the byval attribute.
1349 // * If a struct is less than 2*XLEN, then coerce to either a two-element
1350 // word-sized array or a 2*XLEN scalar (depending on alignment).
1351 // * The frontend can determine whether a struct is returned by reference or
1352 // not based on its size and fields. If it will be returned by reference, the
1353 // frontend must modify the prototype so a pointer with the sret annotation is
1354 // passed as the first argument. This is not necessary for large scalar
1356 // * Struct return values and varargs should be coerced to structs containing
1357 // register-size fields in the same situations they would be for fixed
1360 static const MCPhysReg ArgGPRs[] = {
1361 RISCV::X10, RISCV::X11, RISCV::X12, RISCV::X13,
1362 RISCV::X14, RISCV::X15, RISCV::X16, RISCV::X17
1364 static const MCPhysReg ArgFPR32s[] = {
1365 RISCV::F10_32, RISCV::F11_32, RISCV::F12_32, RISCV::F13_32,
1366 RISCV::F14_32, RISCV::F15_32, RISCV::F16_32, RISCV::F17_32
1368 static const MCPhysReg ArgFPR64s[] = {
1369 RISCV::F10_64, RISCV::F11_64, RISCV::F12_64, RISCV::F13_64,
1370 RISCV::F14_64, RISCV::F15_64, RISCV::F16_64, RISCV::F17_64
1373 // Pass a 2*XLEN argument that has been split into two XLEN values through
1374 // registers or the stack as necessary.
1375 static bool CC_RISCVAssign2XLen(unsigned XLen, CCState &State, CCValAssign VA1,
1376 ISD::ArgFlagsTy ArgFlags1, unsigned ValNo2,
1377 MVT ValVT2, MVT LocVT2,
1378 ISD::ArgFlagsTy ArgFlags2) {
1379 unsigned XLenInBytes = XLen / 8;
1380 if (unsigned Reg = State.AllocateReg(ArgGPRs)) {
1381 // At least one half can be passed via register.
1382 State.addLoc(CCValAssign::getReg(VA1.getValNo(), VA1.getValVT(), Reg,
1383 VA1.getLocVT(), CCValAssign::Full));
1385 // Both halves must be passed on the stack, with proper alignment.
1386 unsigned StackAlign = std::max(XLenInBytes, ArgFlags1.getOrigAlign());
1388 CCValAssign::getMem(VA1.getValNo(), VA1.getValVT(),
1389 State.AllocateStack(XLenInBytes, StackAlign),
1390 VA1.getLocVT(), CCValAssign::Full));
1391 State.addLoc(CCValAssign::getMem(
1392 ValNo2, ValVT2, State.AllocateStack(XLenInBytes, XLenInBytes), LocVT2,
1393 CCValAssign::Full));
1397 if (unsigned Reg = State.AllocateReg(ArgGPRs)) {
1398 // The second half can also be passed via register.
1400 CCValAssign::getReg(ValNo2, ValVT2, Reg, LocVT2, CCValAssign::Full));
1402 // The second half is passed via the stack, without additional alignment.
1403 State.addLoc(CCValAssign::getMem(
1404 ValNo2, ValVT2, State.AllocateStack(XLenInBytes, XLenInBytes), LocVT2,
1405 CCValAssign::Full));
1411 // Implements the RISC-V calling convention. Returns true upon failure.
1412 static bool CC_RISCV(const DataLayout &DL, RISCVABI::ABI ABI, unsigned ValNo,
1413 MVT ValVT, MVT LocVT, CCValAssign::LocInfo LocInfo,
1414 ISD::ArgFlagsTy ArgFlags, CCState &State, bool IsFixed,
1415 bool IsRet, Type *OrigTy) {
1416 unsigned XLen = DL.getLargestLegalIntTypeSizeInBits();
1417 assert(XLen == 32 || XLen == 64);
1418 MVT XLenVT = XLen == 32 ? MVT::i32 : MVT::i64;
1420 // Any return value split in to more than two values can't be returned
1422 if (IsRet && ValNo > 1)
1425 // UseGPRForF32 if targeting one of the soft-float ABIs, if passing a
1426 // variadic argument, or if no F32 argument registers are available.
1427 bool UseGPRForF32 = true;
1428 // UseGPRForF64 if targeting soft-float ABIs or an FLEN=32 ABI, if passing a
1429 // variadic argument, or if no F64 argument registers are available.
1430 bool UseGPRForF64 = true;
1434 llvm_unreachable("Unexpected ABI");
1435 case RISCVABI::ABI_ILP32:
1436 case RISCVABI::ABI_LP64:
1438 case RISCVABI::ABI_ILP32F:
1439 case RISCVABI::ABI_LP64F:
1440 UseGPRForF32 = !IsFixed;
1442 case RISCVABI::ABI_ILP32D:
1443 case RISCVABI::ABI_LP64D:
1444 UseGPRForF32 = !IsFixed;
1445 UseGPRForF64 = !IsFixed;
1449 if (State.getFirstUnallocated(ArgFPR32s) == array_lengthof(ArgFPR32s))
1450 UseGPRForF32 = true;
1451 if (State.getFirstUnallocated(ArgFPR64s) == array_lengthof(ArgFPR64s))
1452 UseGPRForF64 = true;
1454 // From this point on, rely on UseGPRForF32, UseGPRForF64 and similar local
1455 // variables rather than directly checking against the target ABI.
1457 if (UseGPRForF32 && ValVT == MVT::f32) {
1459 LocInfo = CCValAssign::BCvt;
1460 } else if (UseGPRForF64 && XLen == 64 && ValVT == MVT::f64) {
1462 LocInfo = CCValAssign::BCvt;
1465 // If this is a variadic argument, the RISC-V calling convention requires
1466 // that it is assigned an 'even' or 'aligned' register if it has 8-byte
1467 // alignment (RV32) or 16-byte alignment (RV64). An aligned register should
1468 // be used regardless of whether the original argument was split during
1469 // legalisation or not. The argument will not be passed by registers if the
1470 // original type is larger than 2*XLEN, so the register alignment rule does
1472 unsigned TwoXLenInBytes = (2 * XLen) / 8;
1473 if (!IsFixed && ArgFlags.getOrigAlign() == TwoXLenInBytes &&
1474 DL.getTypeAllocSize(OrigTy) == TwoXLenInBytes) {
1475 unsigned RegIdx = State.getFirstUnallocated(ArgGPRs);
1476 // Skip 'odd' register if necessary.
1477 if (RegIdx != array_lengthof(ArgGPRs) && RegIdx % 2 == 1)
1478 State.AllocateReg(ArgGPRs);
1481 SmallVectorImpl<CCValAssign> &PendingLocs = State.getPendingLocs();
1482 SmallVectorImpl<ISD::ArgFlagsTy> &PendingArgFlags =
1483 State.getPendingArgFlags();
1485 assert(PendingLocs.size() == PendingArgFlags.size() &&
1486 "PendingLocs and PendingArgFlags out of sync");
1488 // Handle passing f64 on RV32D with a soft float ABI or when floating point
1489 // registers are exhausted.
1490 if (UseGPRForF64 && XLen == 32 && ValVT == MVT::f64) {
1491 assert(!ArgFlags.isSplit() && PendingLocs.empty() &&
1492 "Can't lower f64 if it is split");
1493 // Depending on available argument GPRS, f64 may be passed in a pair of
1494 // GPRs, split between a GPR and the stack, or passed completely on the
1495 // stack. LowerCall/LowerFormalArguments/LowerReturn must recognise these
1497 unsigned Reg = State.AllocateReg(ArgGPRs);
1500 unsigned StackOffset = State.AllocateStack(8, 8);
1502 CCValAssign::getMem(ValNo, ValVT, StackOffset, LocVT, LocInfo));
1505 if (!State.AllocateReg(ArgGPRs))
1506 State.AllocateStack(4, 4);
1507 State.addLoc(CCValAssign::getReg(ValNo, ValVT, Reg, LocVT, LocInfo));
1511 // Split arguments might be passed indirectly, so keep track of the pending
1513 if (ArgFlags.isSplit() || !PendingLocs.empty()) {
1515 LocInfo = CCValAssign::Indirect;
1516 PendingLocs.push_back(
1517 CCValAssign::getPending(ValNo, ValVT, LocVT, LocInfo));
1518 PendingArgFlags.push_back(ArgFlags);
1519 if (!ArgFlags.isSplitEnd()) {
1524 // If the split argument only had two elements, it should be passed directly
1525 // in registers or on the stack.
1526 if (ArgFlags.isSplitEnd() && PendingLocs.size() <= 2) {
1527 assert(PendingLocs.size() == 2 && "Unexpected PendingLocs.size()");
1528 // Apply the normal calling convention rules to the first half of the
1530 CCValAssign VA = PendingLocs[0];
1531 ISD::ArgFlagsTy AF = PendingArgFlags[0];
1532 PendingLocs.clear();
1533 PendingArgFlags.clear();
1534 return CC_RISCVAssign2XLen(XLen, State, VA, AF, ValNo, ValVT, LocVT,
1538 // Allocate to a register if possible, or else a stack slot.
1540 if (ValVT == MVT::f32 && !UseGPRForF32)
1541 Reg = State.AllocateReg(ArgFPR32s, ArgFPR64s);
1542 else if (ValVT == MVT::f64 && !UseGPRForF64)
1543 Reg = State.AllocateReg(ArgFPR64s, ArgFPR32s);
1545 Reg = State.AllocateReg(ArgGPRs);
1546 unsigned StackOffset = Reg ? 0 : State.AllocateStack(XLen / 8, XLen / 8);
1548 // If we reach this point and PendingLocs is non-empty, we must be at the
1549 // end of a split argument that must be passed indirectly.
1550 if (!PendingLocs.empty()) {
1551 assert(ArgFlags.isSplitEnd() && "Expected ArgFlags.isSplitEnd()");
1552 assert(PendingLocs.size() > 2 && "Unexpected PendingLocs.size()");
1554 for (auto &It : PendingLocs) {
1556 It.convertToReg(Reg);
1558 It.convertToMem(StackOffset);
1561 PendingLocs.clear();
1562 PendingArgFlags.clear();
1566 assert((!UseGPRForF32 || !UseGPRForF64 || LocVT == XLenVT) &&
1567 "Expected an XLenVT at this stage");
1570 State.addLoc(CCValAssign::getReg(ValNo, ValVT, Reg, LocVT, LocInfo));
1574 // When an f32 or f64 is passed on the stack, no bit-conversion is needed.
1575 if (ValVT == MVT::f32 || ValVT == MVT::f64) {
1577 LocInfo = CCValAssign::Full;
1579 State.addLoc(CCValAssign::getMem(ValNo, ValVT, StackOffset, LocVT, LocInfo));
1583 void RISCVTargetLowering::analyzeInputArgs(
1584 MachineFunction &MF, CCState &CCInfo,
1585 const SmallVectorImpl<ISD::InputArg> &Ins, bool IsRet) const {
1586 unsigned NumArgs = Ins.size();
1587 FunctionType *FType = MF.getFunction().getFunctionType();
1589 for (unsigned i = 0; i != NumArgs; ++i) {
1590 MVT ArgVT = Ins[i].VT;
1591 ISD::ArgFlagsTy ArgFlags = Ins[i].Flags;
1593 Type *ArgTy = nullptr;
1595 ArgTy = FType->getReturnType();
1596 else if (Ins[i].isOrigArg())
1597 ArgTy = FType->getParamType(Ins[i].getOrigArgIndex());
1599 RISCVABI::ABI ABI = MF.getSubtarget<RISCVSubtarget>().getTargetABI();
1600 if (CC_RISCV(MF.getDataLayout(), ABI, i, ArgVT, ArgVT, CCValAssign::Full,
1601 ArgFlags, CCInfo, /*IsRet=*/true, IsRet, ArgTy)) {
1602 LLVM_DEBUG(dbgs() << "InputArg #" << i << " has unhandled type "
1603 << EVT(ArgVT).getEVTString() << '\n');
1604 llvm_unreachable(nullptr);
1609 void RISCVTargetLowering::analyzeOutputArgs(
1610 MachineFunction &MF, CCState &CCInfo,
1611 const SmallVectorImpl<ISD::OutputArg> &Outs, bool IsRet,
1612 CallLoweringInfo *CLI) const {
1613 unsigned NumArgs = Outs.size();
1615 for (unsigned i = 0; i != NumArgs; i++) {
1616 MVT ArgVT = Outs[i].VT;
1617 ISD::ArgFlagsTy ArgFlags = Outs[i].Flags;
1618 Type *OrigTy = CLI ? CLI->getArgs()[Outs[i].OrigArgIndex].Ty : nullptr;
1620 RISCVABI::ABI ABI = MF.getSubtarget<RISCVSubtarget>().getTargetABI();
1621 if (CC_RISCV(MF.getDataLayout(), ABI, i, ArgVT, ArgVT, CCValAssign::Full,
1622 ArgFlags, CCInfo, Outs[i].IsFixed, IsRet, OrigTy)) {
1623 LLVM_DEBUG(dbgs() << "OutputArg #" << i << " has unhandled type "
1624 << EVT(ArgVT).getEVTString() << "\n");
1625 llvm_unreachable(nullptr);
1630 // Convert Val to a ValVT. Should not be called for CCValAssign::Indirect
1632 static SDValue convertLocVTToValVT(SelectionDAG &DAG, SDValue Val,
1633 const CCValAssign &VA, const SDLoc &DL) {
1634 switch (VA.getLocInfo()) {
1636 llvm_unreachable("Unexpected CCValAssign::LocInfo");
1637 case CCValAssign::Full:
1639 case CCValAssign::BCvt:
1640 if (VA.getLocVT() == MVT::i64 && VA.getValVT() == MVT::f32) {
1641 Val = DAG.getNode(RISCVISD::FMV_W_X_RV64, DL, MVT::f32, Val);
1644 Val = DAG.getNode(ISD::BITCAST, DL, VA.getValVT(), Val);
1650 // The caller is responsible for loading the full value if the argument is
1651 // passed with CCValAssign::Indirect.
1652 static SDValue unpackFromRegLoc(SelectionDAG &DAG, SDValue Chain,
1653 const CCValAssign &VA, const SDLoc &DL) {
1654 MachineFunction &MF = DAG.getMachineFunction();
1655 MachineRegisterInfo &RegInfo = MF.getRegInfo();
1656 EVT LocVT = VA.getLocVT();
1658 const TargetRegisterClass *RC;
1660 switch (LocVT.getSimpleVT().SimpleTy) {
1662 llvm_unreachable("Unexpected register type");
1665 RC = &RISCV::GPRRegClass;
1668 RC = &RISCV::FPR32RegClass;
1671 RC = &RISCV::FPR64RegClass;
1675 unsigned VReg = RegInfo.createVirtualRegister(RC);
1676 RegInfo.addLiveIn(VA.getLocReg(), VReg);
1677 Val = DAG.getCopyFromReg(Chain, DL, VReg, LocVT);
1679 if (VA.getLocInfo() == CCValAssign::Indirect)
1682 return convertLocVTToValVT(DAG, Val, VA, DL);
1685 static SDValue convertValVTToLocVT(SelectionDAG &DAG, SDValue Val,
1686 const CCValAssign &VA, const SDLoc &DL) {
1687 EVT LocVT = VA.getLocVT();
1689 switch (VA.getLocInfo()) {
1691 llvm_unreachable("Unexpected CCValAssign::LocInfo");
1692 case CCValAssign::Full:
1694 case CCValAssign::BCvt:
1695 if (VA.getLocVT() == MVT::i64 && VA.getValVT() == MVT::f32) {
1696 Val = DAG.getNode(RISCVISD::FMV_X_ANYEXTW_RV64, DL, MVT::i64, Val);
1699 Val = DAG.getNode(ISD::BITCAST, DL, LocVT, Val);
1705 // The caller is responsible for loading the full value if the argument is
1706 // passed with CCValAssign::Indirect.
1707 static SDValue unpackFromMemLoc(SelectionDAG &DAG, SDValue Chain,
1708 const CCValAssign &VA, const SDLoc &DL) {
1709 MachineFunction &MF = DAG.getMachineFunction();
1710 MachineFrameInfo &MFI = MF.getFrameInfo();
1711 EVT LocVT = VA.getLocVT();
1712 EVT ValVT = VA.getValVT();
1713 EVT PtrVT = MVT::getIntegerVT(DAG.getDataLayout().getPointerSizeInBits(0));
1714 int FI = MFI.CreateFixedObject(ValVT.getSizeInBits() / 8,
1715 VA.getLocMemOffset(), /*Immutable=*/true);
1716 SDValue FIN = DAG.getFrameIndex(FI, PtrVT);
1719 ISD::LoadExtType ExtType;
1720 switch (VA.getLocInfo()) {
1722 llvm_unreachable("Unexpected CCValAssign::LocInfo");
1723 case CCValAssign::Full:
1724 case CCValAssign::Indirect:
1725 case CCValAssign::BCvt:
1726 ExtType = ISD::NON_EXTLOAD;
1729 Val = DAG.getExtLoad(
1730 ExtType, DL, LocVT, Chain, FIN,
1731 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), FI), ValVT);
1735 static SDValue unpackF64OnRV32DSoftABI(SelectionDAG &DAG, SDValue Chain,
1736 const CCValAssign &VA, const SDLoc &DL) {
1737 assert(VA.getLocVT() == MVT::i32 && VA.getValVT() == MVT::f64 &&
1739 MachineFunction &MF = DAG.getMachineFunction();
1740 MachineFrameInfo &MFI = MF.getFrameInfo();
1741 MachineRegisterInfo &RegInfo = MF.getRegInfo();
1743 if (VA.isMemLoc()) {
1744 // f64 is passed on the stack.
1745 int FI = MFI.CreateFixedObject(8, VA.getLocMemOffset(), /*Immutable=*/true);
1746 SDValue FIN = DAG.getFrameIndex(FI, MVT::i32);
1747 return DAG.getLoad(MVT::f64, DL, Chain, FIN,
1748 MachinePointerInfo::getFixedStack(MF, FI));
1751 assert(VA.isRegLoc() && "Expected register VA assignment");
1753 unsigned LoVReg = RegInfo.createVirtualRegister(&RISCV::GPRRegClass);
1754 RegInfo.addLiveIn(VA.getLocReg(), LoVReg);
1755 SDValue Lo = DAG.getCopyFromReg(Chain, DL, LoVReg, MVT::i32);
1757 if (VA.getLocReg() == RISCV::X17) {
1758 // Second half of f64 is passed on the stack.
1759 int FI = MFI.CreateFixedObject(4, 0, /*Immutable=*/true);
1760 SDValue FIN = DAG.getFrameIndex(FI, MVT::i32);
1761 Hi = DAG.getLoad(MVT::i32, DL, Chain, FIN,
1762 MachinePointerInfo::getFixedStack(MF, FI));
1764 // Second half of f64 is passed in another GPR.
1765 unsigned HiVReg = RegInfo.createVirtualRegister(&RISCV::GPRRegClass);
1766 RegInfo.addLiveIn(VA.getLocReg() + 1, HiVReg);
1767 Hi = DAG.getCopyFromReg(Chain, DL, HiVReg, MVT::i32);
1769 return DAG.getNode(RISCVISD::BuildPairF64, DL, MVT::f64, Lo, Hi);
1772 // Transform physical registers into virtual registers.
1773 SDValue RISCVTargetLowering::LowerFormalArguments(
1774 SDValue Chain, CallingConv::ID CallConv, bool IsVarArg,
1775 const SmallVectorImpl<ISD::InputArg> &Ins, const SDLoc &DL,
1776 SelectionDAG &DAG, SmallVectorImpl<SDValue> &InVals) const {
1780 report_fatal_error("Unsupported calling convention");
1781 case CallingConv::C:
1782 case CallingConv::Fast:
1786 MachineFunction &MF = DAG.getMachineFunction();
1788 const Function &Func = MF.getFunction();
1789 if (Func.hasFnAttribute("interrupt")) {
1790 if (!Func.arg_empty())
1792 "Functions with the interrupt attribute cannot have arguments!");
1795 MF.getFunction().getFnAttribute("interrupt").getValueAsString();
1797 if (!(Kind == "user" || Kind == "supervisor" || Kind == "machine"))
1799 "Function interrupt attribute argument not supported!");
1802 EVT PtrVT = getPointerTy(DAG.getDataLayout());
1803 MVT XLenVT = Subtarget.getXLenVT();
1804 unsigned XLenInBytes = Subtarget.getXLen() / 8;
1805 // Used with vargs to acumulate store chains.
1806 std::vector<SDValue> OutChains;
1808 // Assign locations to all of the incoming arguments.
1809 SmallVector<CCValAssign, 16> ArgLocs;
1810 CCState CCInfo(CallConv, IsVarArg, MF, ArgLocs, *DAG.getContext());
1811 analyzeInputArgs(MF, CCInfo, Ins, /*IsRet=*/false);
1813 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1814 CCValAssign &VA = ArgLocs[i];
1816 // Passing f64 on RV32D with a soft float ABI must be handled as a special
1818 if (VA.getLocVT() == MVT::i32 && VA.getValVT() == MVT::f64)
1819 ArgValue = unpackF64OnRV32DSoftABI(DAG, Chain, VA, DL);
1820 else if (VA.isRegLoc())
1821 ArgValue = unpackFromRegLoc(DAG, Chain, VA, DL);
1823 ArgValue = unpackFromMemLoc(DAG, Chain, VA, DL);
1825 if (VA.getLocInfo() == CCValAssign::Indirect) {
1826 // If the original argument was split and passed by reference (e.g. i128
1827 // on RV32), we need to load all parts of it here (using the same
1829 InVals.push_back(DAG.getLoad(VA.getValVT(), DL, Chain, ArgValue,
1830 MachinePointerInfo()));
1831 unsigned ArgIndex = Ins[i].OrigArgIndex;
1832 assert(Ins[i].PartOffset == 0);
1833 while (i + 1 != e && Ins[i + 1].OrigArgIndex == ArgIndex) {
1834 CCValAssign &PartVA = ArgLocs[i + 1];
1835 unsigned PartOffset = Ins[i + 1].PartOffset;
1836 SDValue Address = DAG.getNode(ISD::ADD, DL, PtrVT, ArgValue,
1837 DAG.getIntPtrConstant(PartOffset, DL));
1838 InVals.push_back(DAG.getLoad(PartVA.getValVT(), DL, Chain, Address,
1839 MachinePointerInfo()));
1844 InVals.push_back(ArgValue);
1848 ArrayRef<MCPhysReg> ArgRegs = makeArrayRef(ArgGPRs);
1849 unsigned Idx = CCInfo.getFirstUnallocated(ArgRegs);
1850 const TargetRegisterClass *RC = &RISCV::GPRRegClass;
1851 MachineFrameInfo &MFI = MF.getFrameInfo();
1852 MachineRegisterInfo &RegInfo = MF.getRegInfo();
1853 RISCVMachineFunctionInfo *RVFI = MF.getInfo<RISCVMachineFunctionInfo>();
1855 // Offset of the first variable argument from stack pointer, and size of
1856 // the vararg save area. For now, the varargs save area is either zero or
1857 // large enough to hold a0-a7.
1858 int VaArgOffset, VarArgsSaveSize;
1860 // If all registers are allocated, then all varargs must be passed on the
1861 // stack and we don't need to save any argregs.
1862 if (ArgRegs.size() == Idx) {
1863 VaArgOffset = CCInfo.getNextStackOffset();
1864 VarArgsSaveSize = 0;
1866 VarArgsSaveSize = XLenInBytes * (ArgRegs.size() - Idx);
1867 VaArgOffset = -VarArgsSaveSize;
1870 // Record the frame index of the first variable argument
1871 // which is a value necessary to VASTART.
1872 int FI = MFI.CreateFixedObject(XLenInBytes, VaArgOffset, true);
1873 RVFI->setVarArgsFrameIndex(FI);
1875 // If saving an odd number of registers then create an extra stack slot to
1876 // ensure that the frame pointer is 2*XLEN-aligned, which in turn ensures
1877 // offsets to even-numbered registered remain 2*XLEN-aligned.
1879 FI = MFI.CreateFixedObject(XLenInBytes, VaArgOffset - (int)XLenInBytes,
1881 VarArgsSaveSize += XLenInBytes;
1884 // Copy the integer registers that may have been used for passing varargs
1885 // to the vararg save area.
1886 for (unsigned I = Idx; I < ArgRegs.size();
1887 ++I, VaArgOffset += XLenInBytes) {
1888 const unsigned Reg = RegInfo.createVirtualRegister(RC);
1889 RegInfo.addLiveIn(ArgRegs[I], Reg);
1890 SDValue ArgValue = DAG.getCopyFromReg(Chain, DL, Reg, XLenVT);
1891 FI = MFI.CreateFixedObject(XLenInBytes, VaArgOffset, true);
1892 SDValue PtrOff = DAG.getFrameIndex(FI, getPointerTy(DAG.getDataLayout()));
1893 SDValue Store = DAG.getStore(Chain, DL, ArgValue, PtrOff,
1894 MachinePointerInfo::getFixedStack(MF, FI));
1895 cast<StoreSDNode>(Store.getNode())
1897 ->setValue((Value *)nullptr);
1898 OutChains.push_back(Store);
1900 RVFI->setVarArgsSaveSize(VarArgsSaveSize);
1903 // All stores are grouped in one node to allow the matching between
1904 // the size of Ins and InVals. This only happens for vararg functions.
1905 if (!OutChains.empty()) {
1906 OutChains.push_back(Chain);
1907 Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other, OutChains);
1913 /// isEligibleForTailCallOptimization - Check whether the call is eligible
1914 /// for tail call optimization.
1915 /// Note: This is modelled after ARM's IsEligibleForTailCallOptimization.
1916 bool RISCVTargetLowering::isEligibleForTailCallOptimization(
1917 CCState &CCInfo, CallLoweringInfo &CLI, MachineFunction &MF,
1918 const SmallVector<CCValAssign, 16> &ArgLocs) const {
1920 auto &Callee = CLI.Callee;
1921 auto CalleeCC = CLI.CallConv;
1922 auto IsVarArg = CLI.IsVarArg;
1923 auto &Outs = CLI.Outs;
1924 auto &Caller = MF.getFunction();
1925 auto CallerCC = Caller.getCallingConv();
1927 // Do not tail call opt functions with "disable-tail-calls" attribute.
1928 if (Caller.getFnAttribute("disable-tail-calls").getValueAsString() == "true")
1931 // Exception-handling functions need a special set of instructions to
1932 // indicate a return to the hardware. Tail-calling another function would
1933 // probably break this.
1934 // TODO: The "interrupt" attribute isn't currently defined by RISC-V. This
1935 // should be expanded as new function attributes are introduced.
1936 if (Caller.hasFnAttribute("interrupt"))
1939 // Do not tail call opt functions with varargs.
1943 // Do not tail call opt if the stack is used to pass parameters.
1944 if (CCInfo.getNextStackOffset() != 0)
1947 // Do not tail call opt if any parameters need to be passed indirectly.
1948 // Since long doubles (fp128) and i128 are larger than 2*XLEN, they are
1949 // passed indirectly. So the address of the value will be passed in a
1950 // register, or if not available, then the address is put on the stack. In
1951 // order to pass indirectly, space on the stack often needs to be allocated
1952 // in order to store the value. In this case the CCInfo.getNextStackOffset()
1953 // != 0 check is not enough and we need to check if any CCValAssign ArgsLocs
1954 // are passed CCValAssign::Indirect.
1955 for (auto &VA : ArgLocs)
1956 if (VA.getLocInfo() == CCValAssign::Indirect)
1959 // Do not tail call opt if either caller or callee uses struct return
1961 auto IsCallerStructRet = Caller.hasStructRetAttr();
1962 auto IsCalleeStructRet = Outs.empty() ? false : Outs[0].Flags.isSRet();
1963 if (IsCallerStructRet || IsCalleeStructRet)
1966 // Externally-defined functions with weak linkage should not be
1967 // tail-called. The behaviour of branch instructions in this situation (as
1968 // used for tail calls) is implementation-defined, so we cannot rely on the
1969 // linker replacing the tail call with a return.
1970 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1971 const GlobalValue *GV = G->getGlobal();
1972 if (GV->hasExternalWeakLinkage())
1976 // The callee has to preserve all registers the caller needs to preserve.
1977 const RISCVRegisterInfo *TRI = Subtarget.getRegisterInfo();
1978 const uint32_t *CallerPreserved = TRI->getCallPreservedMask(MF, CallerCC);
1979 if (CalleeCC != CallerCC) {
1980 const uint32_t *CalleePreserved = TRI->getCallPreservedMask(MF, CalleeCC);
1981 if (!TRI->regmaskSubsetEqual(CallerPreserved, CalleePreserved))
1985 // Byval parameters hand the function a pointer directly into the stack area
1986 // we want to reuse during a tail call. Working around this *is* possible
1987 // but less efficient and uglier in LowerCall.
1988 for (auto &Arg : Outs)
1989 if (Arg.Flags.isByVal())
1995 // Lower a call to a callseq_start + CALL + callseq_end chain, and add input
1996 // and output parameter nodes.
1997 SDValue RISCVTargetLowering::LowerCall(CallLoweringInfo &CLI,
1998 SmallVectorImpl<SDValue> &InVals) const {
1999 SelectionDAG &DAG = CLI.DAG;
2001 SmallVectorImpl<ISD::OutputArg> &Outs = CLI.Outs;
2002 SmallVectorImpl<SDValue> &OutVals = CLI.OutVals;
2003 SmallVectorImpl<ISD::InputArg> &Ins = CLI.Ins;
2004 SDValue Chain = CLI.Chain;
2005 SDValue Callee = CLI.Callee;
2006 bool &IsTailCall = CLI.IsTailCall;
2007 CallingConv::ID CallConv = CLI.CallConv;
2008 bool IsVarArg = CLI.IsVarArg;
2009 EVT PtrVT = getPointerTy(DAG.getDataLayout());
2010 MVT XLenVT = Subtarget.getXLenVT();
2012 MachineFunction &MF = DAG.getMachineFunction();
2014 // Analyze the operands of the call, assigning locations to each operand.
2015 SmallVector<CCValAssign, 16> ArgLocs;
2016 CCState ArgCCInfo(CallConv, IsVarArg, MF, ArgLocs, *DAG.getContext());
2017 analyzeOutputArgs(MF, ArgCCInfo, Outs, /*IsRet=*/false, &CLI);
2019 // Check if it's really possible to do a tail call.
2021 IsTailCall = isEligibleForTailCallOptimization(ArgCCInfo, CLI, MF, ArgLocs);
2025 else if (CLI.CS && CLI.CS.isMustTailCall())
2026 report_fatal_error("failed to perform tail call elimination on a call "
2027 "site marked musttail");
2029 // Get a count of how many bytes are to be pushed on the stack.
2030 unsigned NumBytes = ArgCCInfo.getNextStackOffset();
2032 // Create local copies for byval args
2033 SmallVector<SDValue, 8> ByValArgs;
2034 for (unsigned i = 0, e = Outs.size(); i != e; ++i) {
2035 ISD::ArgFlagsTy Flags = Outs[i].Flags;
2036 if (!Flags.isByVal())
2039 SDValue Arg = OutVals[i];
2040 unsigned Size = Flags.getByValSize();
2041 unsigned Align = Flags.getByValAlign();
2043 int FI = MF.getFrameInfo().CreateStackObject(Size, Align, /*isSS=*/false);
2044 SDValue FIPtr = DAG.getFrameIndex(FI, getPointerTy(DAG.getDataLayout()));
2045 SDValue SizeNode = DAG.getConstant(Size, DL, XLenVT);
2047 Chain = DAG.getMemcpy(Chain, DL, FIPtr, Arg, SizeNode, Align,
2048 /*IsVolatile=*/false,
2049 /*AlwaysInline=*/false,
2050 IsTailCall, MachinePointerInfo(),
2051 MachinePointerInfo());
2052 ByValArgs.push_back(FIPtr);
2056 Chain = DAG.getCALLSEQ_START(Chain, NumBytes, 0, CLI.DL);
2058 // Copy argument values to their designated locations.
2059 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
2060 SmallVector<SDValue, 8> MemOpChains;
2062 for (unsigned i = 0, j = 0, e = ArgLocs.size(); i != e; ++i) {
2063 CCValAssign &VA = ArgLocs[i];
2064 SDValue ArgValue = OutVals[i];
2065 ISD::ArgFlagsTy Flags = Outs[i].Flags;
2067 // Handle passing f64 on RV32D with a soft float ABI as a special case.
2068 bool IsF64OnRV32DSoftABI =
2069 VA.getLocVT() == MVT::i32 && VA.getValVT() == MVT::f64;
2070 if (IsF64OnRV32DSoftABI && VA.isRegLoc()) {
2071 SDValue SplitF64 = DAG.getNode(
2072 RISCVISD::SplitF64, DL, DAG.getVTList(MVT::i32, MVT::i32), ArgValue);
2073 SDValue Lo = SplitF64.getValue(0);
2074 SDValue Hi = SplitF64.getValue(1);
2076 unsigned RegLo = VA.getLocReg();
2077 RegsToPass.push_back(std::make_pair(RegLo, Lo));
2079 if (RegLo == RISCV::X17) {
2080 // Second half of f64 is passed on the stack.
2081 // Work out the address of the stack slot.
2082 if (!StackPtr.getNode())
2083 StackPtr = DAG.getCopyFromReg(Chain, DL, RISCV::X2, PtrVT);
2085 MemOpChains.push_back(
2086 DAG.getStore(Chain, DL, Hi, StackPtr, MachinePointerInfo()));
2088 // Second half of f64 is passed in another GPR.
2089 unsigned RegHigh = RegLo + 1;
2090 RegsToPass.push_back(std::make_pair(RegHigh, Hi));
2095 // IsF64OnRV32DSoftABI && VA.isMemLoc() is handled below in the same way
2096 // as any other MemLoc.
2098 // Promote the value if needed.
2099 // For now, only handle fully promoted and indirect arguments.
2100 if (VA.getLocInfo() == CCValAssign::Indirect) {
2101 // Store the argument in a stack slot and pass its address.
2102 SDValue SpillSlot = DAG.CreateStackTemporary(Outs[i].ArgVT);
2103 int FI = cast<FrameIndexSDNode>(SpillSlot)->getIndex();
2104 MemOpChains.push_back(
2105 DAG.getStore(Chain, DL, ArgValue, SpillSlot,
2106 MachinePointerInfo::getFixedStack(MF, FI)));
2107 // If the original argument was split (e.g. i128), we need
2108 // to store all parts of it here (and pass just one address).
2109 unsigned ArgIndex = Outs[i].OrigArgIndex;
2110 assert(Outs[i].PartOffset == 0);
2111 while (i + 1 != e && Outs[i + 1].OrigArgIndex == ArgIndex) {
2112 SDValue PartValue = OutVals[i + 1];
2113 unsigned PartOffset = Outs[i + 1].PartOffset;
2114 SDValue Address = DAG.getNode(ISD::ADD, DL, PtrVT, SpillSlot,
2115 DAG.getIntPtrConstant(PartOffset, DL));
2116 MemOpChains.push_back(
2117 DAG.getStore(Chain, DL, PartValue, Address,
2118 MachinePointerInfo::getFixedStack(MF, FI)));
2121 ArgValue = SpillSlot;
2123 ArgValue = convertValVTToLocVT(DAG, ArgValue, VA, DL);
2126 // Use local copy if it is a byval arg.
2127 if (Flags.isByVal())
2128 ArgValue = ByValArgs[j++];
2130 if (VA.isRegLoc()) {
2131 // Queue up the argument copies and emit them at the end.
2132 RegsToPass.push_back(std::make_pair(VA.getLocReg(), ArgValue));
2134 assert(VA.isMemLoc() && "Argument not register or memory");
2135 assert(!IsTailCall && "Tail call not allowed if stack is used "
2136 "for passing parameters");
2138 // Work out the address of the stack slot.
2139 if (!StackPtr.getNode())
2140 StackPtr = DAG.getCopyFromReg(Chain, DL, RISCV::X2, PtrVT);
2142 DAG.getNode(ISD::ADD, DL, PtrVT, StackPtr,
2143 DAG.getIntPtrConstant(VA.getLocMemOffset(), DL));
2146 MemOpChains.push_back(
2147 DAG.getStore(Chain, DL, ArgValue, Address, MachinePointerInfo()));
2151 // Join the stores, which are independent of one another.
2152 if (!MemOpChains.empty())
2153 Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other, MemOpChains);
2157 // Build a sequence of copy-to-reg nodes, chained and glued together.
2158 for (auto &Reg : RegsToPass) {
2159 Chain = DAG.getCopyToReg(Chain, DL, Reg.first, Reg.second, Glue);
2160 Glue = Chain.getValue(1);
2163 // If the callee is a GlobalAddress/ExternalSymbol node, turn it into a
2164 // TargetGlobalAddress/TargetExternalSymbol node so that legalize won't
2165 // split it and then direct call can be matched by PseudoCALL.
2166 if (GlobalAddressSDNode *S = dyn_cast<GlobalAddressSDNode>(Callee)) {
2167 const GlobalValue *GV = S->getGlobal();
2169 unsigned OpFlags = RISCVII::MO_CALL;
2170 if (!getTargetMachine().shouldAssumeDSOLocal(*GV->getParent(), GV))
2171 OpFlags = RISCVII::MO_PLT;
2173 Callee = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, OpFlags);
2174 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
2175 unsigned OpFlags = RISCVII::MO_CALL;
2177 if (!getTargetMachine().shouldAssumeDSOLocal(*MF.getFunction().getParent(),
2179 OpFlags = RISCVII::MO_PLT;
2181 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), PtrVT, OpFlags);
2184 // The first call operand is the chain and the second is the target address.
2185 SmallVector<SDValue, 8> Ops;
2186 Ops.push_back(Chain);
2187 Ops.push_back(Callee);
2189 // Add argument registers to the end of the list so that they are
2190 // known live into the call.
2191 for (auto &Reg : RegsToPass)
2192 Ops.push_back(DAG.getRegister(Reg.first, Reg.second.getValueType()));
2195 // Add a register mask operand representing the call-preserved registers.
2196 const TargetRegisterInfo *TRI = Subtarget.getRegisterInfo();
2197 const uint32_t *Mask = TRI->getCallPreservedMask(MF, CallConv);
2198 assert(Mask && "Missing call preserved mask for calling convention");
2199 Ops.push_back(DAG.getRegisterMask(Mask));
2202 // Glue the call to the argument copies, if any.
2204 Ops.push_back(Glue);
2207 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
2210 MF.getFrameInfo().setHasTailCall();
2211 return DAG.getNode(RISCVISD::TAIL, DL, NodeTys, Ops);
2214 Chain = DAG.getNode(RISCVISD::CALL, DL, NodeTys, Ops);
2215 Glue = Chain.getValue(1);
2217 // Mark the end of the call, which is glued to the call itself.
2218 Chain = DAG.getCALLSEQ_END(Chain,
2219 DAG.getConstant(NumBytes, DL, PtrVT, true),
2220 DAG.getConstant(0, DL, PtrVT, true),
2222 Glue = Chain.getValue(1);
2224 // Assign locations to each value returned by this call.
2225 SmallVector<CCValAssign, 16> RVLocs;
2226 CCState RetCCInfo(CallConv, IsVarArg, MF, RVLocs, *DAG.getContext());
2227 analyzeInputArgs(MF, RetCCInfo, Ins, /*IsRet=*/true);
2229 // Copy all of the result registers out of their specified physreg.
2230 for (auto &VA : RVLocs) {
2231 // Copy the value out
2233 DAG.getCopyFromReg(Chain, DL, VA.getLocReg(), VA.getLocVT(), Glue);
2234 // Glue the RetValue to the end of the call sequence
2235 Chain = RetValue.getValue(1);
2236 Glue = RetValue.getValue(2);
2238 if (VA.getLocVT() == MVT::i32 && VA.getValVT() == MVT::f64) {
2239 assert(VA.getLocReg() == ArgGPRs[0] && "Unexpected reg assignment");
2241 DAG.getCopyFromReg(Chain, DL, ArgGPRs[1], MVT::i32, Glue);
2242 Chain = RetValue2.getValue(1);
2243 Glue = RetValue2.getValue(2);
2244 RetValue = DAG.getNode(RISCVISD::BuildPairF64, DL, MVT::f64, RetValue,
2248 RetValue = convertLocVTToValVT(DAG, RetValue, VA, DL);
2250 InVals.push_back(RetValue);
2256 bool RISCVTargetLowering::CanLowerReturn(
2257 CallingConv::ID CallConv, MachineFunction &MF, bool IsVarArg,
2258 const SmallVectorImpl<ISD::OutputArg> &Outs, LLVMContext &Context) const {
2259 SmallVector<CCValAssign, 16> RVLocs;
2260 CCState CCInfo(CallConv, IsVarArg, MF, RVLocs, Context);
2261 for (unsigned i = 0, e = Outs.size(); i != e; ++i) {
2262 MVT VT = Outs[i].VT;
2263 ISD::ArgFlagsTy ArgFlags = Outs[i].Flags;
2264 RISCVABI::ABI ABI = MF.getSubtarget<RISCVSubtarget>().getTargetABI();
2265 if (CC_RISCV(MF.getDataLayout(), ABI, i, VT, VT, CCValAssign::Full,
2266 ArgFlags, CCInfo, /*IsFixed=*/true, /*IsRet=*/true, nullptr))
2273 RISCVTargetLowering::LowerReturn(SDValue Chain, CallingConv::ID CallConv,
2275 const SmallVectorImpl<ISD::OutputArg> &Outs,
2276 const SmallVectorImpl<SDValue> &OutVals,
2277 const SDLoc &DL, SelectionDAG &DAG) const {
2278 // Stores the assignment of the return value to a location.
2279 SmallVector<CCValAssign, 16> RVLocs;
2281 // Info about the registers and stack slot.
2282 CCState CCInfo(CallConv, IsVarArg, DAG.getMachineFunction(), RVLocs,
2285 analyzeOutputArgs(DAG.getMachineFunction(), CCInfo, Outs, /*IsRet=*/true,
2289 SmallVector<SDValue, 4> RetOps(1, Chain);
2291 // Copy the result values into the output registers.
2292 for (unsigned i = 0, e = RVLocs.size(); i < e; ++i) {
2293 SDValue Val = OutVals[i];
2294 CCValAssign &VA = RVLocs[i];
2295 assert(VA.isRegLoc() && "Can only return in registers!");
2297 if (VA.getLocVT() == MVT::i32 && VA.getValVT() == MVT::f64) {
2298 // Handle returning f64 on RV32D with a soft float ABI.
2299 assert(VA.isRegLoc() && "Expected return via registers");
2300 SDValue SplitF64 = DAG.getNode(RISCVISD::SplitF64, DL,
2301 DAG.getVTList(MVT::i32, MVT::i32), Val);
2302 SDValue Lo = SplitF64.getValue(0);
2303 SDValue Hi = SplitF64.getValue(1);
2304 unsigned RegLo = VA.getLocReg();
2305 unsigned RegHi = RegLo + 1;
2306 Chain = DAG.getCopyToReg(Chain, DL, RegLo, Lo, Glue);
2307 Glue = Chain.getValue(1);
2308 RetOps.push_back(DAG.getRegister(RegLo, MVT::i32));
2309 Chain = DAG.getCopyToReg(Chain, DL, RegHi, Hi, Glue);
2310 Glue = Chain.getValue(1);
2311 RetOps.push_back(DAG.getRegister(RegHi, MVT::i32));
2313 // Handle a 'normal' return.
2314 Val = convertValVTToLocVT(DAG, Val, VA, DL);
2315 Chain = DAG.getCopyToReg(Chain, DL, VA.getLocReg(), Val, Glue);
2317 // Guarantee that all emitted copies are stuck together.
2318 Glue = Chain.getValue(1);
2319 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
2323 RetOps[0] = Chain; // Update chain.
2325 // Add the glue node if we have it.
2326 if (Glue.getNode()) {
2327 RetOps.push_back(Glue);
2330 // Interrupt service routines use different return instructions.
2331 const Function &Func = DAG.getMachineFunction().getFunction();
2332 if (Func.hasFnAttribute("interrupt")) {
2333 if (!Func.getReturnType()->isVoidTy())
2335 "Functions with the interrupt attribute must have void return type!");
2337 MachineFunction &MF = DAG.getMachineFunction();
2339 MF.getFunction().getFnAttribute("interrupt").getValueAsString();
2343 RetOpc = RISCVISD::URET_FLAG;
2344 else if (Kind == "supervisor")
2345 RetOpc = RISCVISD::SRET_FLAG;
2347 RetOpc = RISCVISD::MRET_FLAG;
2349 return DAG.getNode(RetOpc, DL, MVT::Other, RetOps);
2352 return DAG.getNode(RISCVISD::RET_FLAG, DL, MVT::Other, RetOps);
2355 const char *RISCVTargetLowering::getTargetNodeName(unsigned Opcode) const {
2356 switch ((RISCVISD::NodeType)Opcode) {
2357 case RISCVISD::FIRST_NUMBER:
2359 case RISCVISD::RET_FLAG:
2360 return "RISCVISD::RET_FLAG";
2361 case RISCVISD::URET_FLAG:
2362 return "RISCVISD::URET_FLAG";
2363 case RISCVISD::SRET_FLAG:
2364 return "RISCVISD::SRET_FLAG";
2365 case RISCVISD::MRET_FLAG:
2366 return "RISCVISD::MRET_FLAG";
2367 case RISCVISD::CALL:
2368 return "RISCVISD::CALL";
2369 case RISCVISD::SELECT_CC:
2370 return "RISCVISD::SELECT_CC";
2371 case RISCVISD::BuildPairF64:
2372 return "RISCVISD::BuildPairF64";
2373 case RISCVISD::SplitF64:
2374 return "RISCVISD::SplitF64";
2375 case RISCVISD::TAIL:
2376 return "RISCVISD::TAIL";
2377 case RISCVISD::SLLW:
2378 return "RISCVISD::SLLW";
2379 case RISCVISD::SRAW:
2380 return "RISCVISD::SRAW";
2381 case RISCVISD::SRLW:
2382 return "RISCVISD::SRLW";
2383 case RISCVISD::DIVW:
2384 return "RISCVISD::DIVW";
2385 case RISCVISD::DIVUW:
2386 return "RISCVISD::DIVUW";
2387 case RISCVISD::REMUW:
2388 return "RISCVISD::REMUW";
2389 case RISCVISD::FMV_W_X_RV64:
2390 return "RISCVISD::FMV_W_X_RV64";
2391 case RISCVISD::FMV_X_ANYEXTW_RV64:
2392 return "RISCVISD::FMV_X_ANYEXTW_RV64";
2393 case RISCVISD::READ_CYCLE_WIDE:
2394 return "RISCVISD::READ_CYCLE_WIDE";
2399 /// getConstraintType - Given a constraint letter, return the type of
2400 /// constraint it is for this target.
2401 RISCVTargetLowering::ConstraintType
2402 RISCVTargetLowering::getConstraintType(StringRef Constraint) const {
2403 if (Constraint.size() == 1) {
2404 switch (Constraint[0]) {
2408 return C_RegisterClass;
2417 return TargetLowering::getConstraintType(Constraint);
2420 std::pair<unsigned, const TargetRegisterClass *>
2421 RISCVTargetLowering::getRegForInlineAsmConstraint(const TargetRegisterInfo *TRI,
2422 StringRef Constraint,
2424 // First, see if this is a constraint that directly corresponds to a
2425 // RISCV register class.
2426 if (Constraint.size() == 1) {
2427 switch (Constraint[0]) {
2429 return std::make_pair(0U, &RISCV::GPRRegClass);
2431 if (Subtarget.hasStdExtF() && VT == MVT::f32)
2432 return std::make_pair(0U, &RISCV::FPR32RegClass);
2433 if (Subtarget.hasStdExtD() && VT == MVT::f64)
2434 return std::make_pair(0U, &RISCV::FPR64RegClass);
2441 return TargetLowering::getRegForInlineAsmConstraint(TRI, Constraint, VT);
2445 RISCVTargetLowering::getInlineAsmMemConstraint(StringRef ConstraintCode) const {
2446 // Currently only support length 1 constraints.
2447 if (ConstraintCode.size() == 1) {
2448 switch (ConstraintCode[0]) {
2450 return InlineAsm::Constraint_A;
2456 return TargetLowering::getInlineAsmMemConstraint(ConstraintCode);
2459 void RISCVTargetLowering::LowerAsmOperandForConstraint(
2460 SDValue Op, std::string &Constraint, std::vector<SDValue> &Ops,
2461 SelectionDAG &DAG) const {
2462 // Currently only support length 1 constraints.
2463 if (Constraint.length() == 1) {
2464 switch (Constraint[0]) {
2466 // Validate & create a 12-bit signed immediate operand.
2467 if (auto *C = dyn_cast<ConstantSDNode>(Op)) {
2468 uint64_t CVal = C->getSExtValue();
2469 if (isInt<12>(CVal))
2471 DAG.getTargetConstant(CVal, SDLoc(Op), Subtarget.getXLenVT()));
2475 // Validate & create an integer zero operand.
2476 if (auto *C = dyn_cast<ConstantSDNode>(Op))
2477 if (C->getZExtValue() == 0)
2479 DAG.getTargetConstant(0, SDLoc(Op), Subtarget.getXLenVT()));
2482 // Validate & create a 5-bit unsigned immediate operand.
2483 if (auto *C = dyn_cast<ConstantSDNode>(Op)) {
2484 uint64_t CVal = C->getZExtValue();
2485 if (isUInt<5>(CVal))
2487 DAG.getTargetConstant(CVal, SDLoc(Op), Subtarget.getXLenVT()));
2494 TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
2497 Instruction *RISCVTargetLowering::emitLeadingFence(IRBuilder<> &Builder,
2499 AtomicOrdering Ord) const {
2500 if (isa<LoadInst>(Inst) && Ord == AtomicOrdering::SequentiallyConsistent)
2501 return Builder.CreateFence(Ord);
2502 if (isa<StoreInst>(Inst) && isReleaseOrStronger(Ord))
2503 return Builder.CreateFence(AtomicOrdering::Release);
2507 Instruction *RISCVTargetLowering::emitTrailingFence(IRBuilder<> &Builder,
2509 AtomicOrdering Ord) const {
2510 if (isa<LoadInst>(Inst) && isAcquireOrStronger(Ord))
2511 return Builder.CreateFence(AtomicOrdering::Acquire);
2515 TargetLowering::AtomicExpansionKind
2516 RISCVTargetLowering::shouldExpandAtomicRMWInIR(AtomicRMWInst *AI) const {
2517 // atomicrmw {fadd,fsub} must be expanded to use compare-exchange, as floating
2518 // point operations can't be used in an lr/sc sequence without breaking the
2519 // forward-progress guarantee.
2520 if (AI->isFloatingPointOperation())
2521 return AtomicExpansionKind::CmpXChg;
2523 unsigned Size = AI->getType()->getPrimitiveSizeInBits();
2524 if (Size == 8 || Size == 16)
2525 return AtomicExpansionKind::MaskedIntrinsic;
2526 return AtomicExpansionKind::None;
2529 static Intrinsic::ID
2530 getIntrinsicForMaskedAtomicRMWBinOp(unsigned XLen, AtomicRMWInst::BinOp BinOp) {
2534 llvm_unreachable("Unexpected AtomicRMW BinOp");
2535 case AtomicRMWInst::Xchg:
2536 return Intrinsic::riscv_masked_atomicrmw_xchg_i32;
2537 case AtomicRMWInst::Add:
2538 return Intrinsic::riscv_masked_atomicrmw_add_i32;
2539 case AtomicRMWInst::Sub:
2540 return Intrinsic::riscv_masked_atomicrmw_sub_i32;
2541 case AtomicRMWInst::Nand:
2542 return Intrinsic::riscv_masked_atomicrmw_nand_i32;
2543 case AtomicRMWInst::Max:
2544 return Intrinsic::riscv_masked_atomicrmw_max_i32;
2545 case AtomicRMWInst::Min:
2546 return Intrinsic::riscv_masked_atomicrmw_min_i32;
2547 case AtomicRMWInst::UMax:
2548 return Intrinsic::riscv_masked_atomicrmw_umax_i32;
2549 case AtomicRMWInst::UMin:
2550 return Intrinsic::riscv_masked_atomicrmw_umin_i32;
2557 llvm_unreachable("Unexpected AtomicRMW BinOp");
2558 case AtomicRMWInst::Xchg:
2559 return Intrinsic::riscv_masked_atomicrmw_xchg_i64;
2560 case AtomicRMWInst::Add:
2561 return Intrinsic::riscv_masked_atomicrmw_add_i64;
2562 case AtomicRMWInst::Sub:
2563 return Intrinsic::riscv_masked_atomicrmw_sub_i64;
2564 case AtomicRMWInst::Nand:
2565 return Intrinsic::riscv_masked_atomicrmw_nand_i64;
2566 case AtomicRMWInst::Max:
2567 return Intrinsic::riscv_masked_atomicrmw_max_i64;
2568 case AtomicRMWInst::Min:
2569 return Intrinsic::riscv_masked_atomicrmw_min_i64;
2570 case AtomicRMWInst::UMax:
2571 return Intrinsic::riscv_masked_atomicrmw_umax_i64;
2572 case AtomicRMWInst::UMin:
2573 return Intrinsic::riscv_masked_atomicrmw_umin_i64;
2577 llvm_unreachable("Unexpected XLen\n");
2580 Value *RISCVTargetLowering::emitMaskedAtomicRMWIntrinsic(
2581 IRBuilder<> &Builder, AtomicRMWInst *AI, Value *AlignedAddr, Value *Incr,
2582 Value *Mask, Value *ShiftAmt, AtomicOrdering Ord) const {
2583 unsigned XLen = Subtarget.getXLen();
2585 Builder.getIntN(XLen, static_cast<uint64_t>(AI->getOrdering()));
2586 Type *Tys[] = {AlignedAddr->getType()};
2587 Function *LrwOpScwLoop = Intrinsic::getDeclaration(
2589 getIntrinsicForMaskedAtomicRMWBinOp(XLen, AI->getOperation()), Tys);
2592 Incr = Builder.CreateSExt(Incr, Builder.getInt64Ty());
2593 Mask = Builder.CreateSExt(Mask, Builder.getInt64Ty());
2594 ShiftAmt = Builder.CreateSExt(ShiftAmt, Builder.getInt64Ty());
2599 // Must pass the shift amount needed to sign extend the loaded value prior
2600 // to performing a signed comparison for min/max. ShiftAmt is the number of
2601 // bits to shift the value into position. Pass XLen-ShiftAmt-ValWidth, which
2602 // is the number of bits to left+right shift the value in order to
2604 if (AI->getOperation() == AtomicRMWInst::Min ||
2605 AI->getOperation() == AtomicRMWInst::Max) {
2606 const DataLayout &DL = AI->getModule()->getDataLayout();
2608 DL.getTypeStoreSizeInBits(AI->getValOperand()->getType());
2610 Builder.CreateSub(Builder.getIntN(XLen, XLen - ValWidth), ShiftAmt);
2611 Result = Builder.CreateCall(LrwOpScwLoop,
2612 {AlignedAddr, Incr, Mask, SextShamt, Ordering});
2615 Builder.CreateCall(LrwOpScwLoop, {AlignedAddr, Incr, Mask, Ordering});
2619 Result = Builder.CreateTrunc(Result, Builder.getInt32Ty());
2623 TargetLowering::AtomicExpansionKind
2624 RISCVTargetLowering::shouldExpandAtomicCmpXchgInIR(
2625 AtomicCmpXchgInst *CI) const {
2626 unsigned Size = CI->getCompareOperand()->getType()->getPrimitiveSizeInBits();
2627 if (Size == 8 || Size == 16)
2628 return AtomicExpansionKind::MaskedIntrinsic;
2629 return AtomicExpansionKind::None;
2632 Value *RISCVTargetLowering::emitMaskedAtomicCmpXchgIntrinsic(
2633 IRBuilder<> &Builder, AtomicCmpXchgInst *CI, Value *AlignedAddr,
2634 Value *CmpVal, Value *NewVal, Value *Mask, AtomicOrdering Ord) const {
2635 unsigned XLen = Subtarget.getXLen();
2636 Value *Ordering = Builder.getIntN(XLen, static_cast<uint64_t>(Ord));
2637 Intrinsic::ID CmpXchgIntrID = Intrinsic::riscv_masked_cmpxchg_i32;
2639 CmpVal = Builder.CreateSExt(CmpVal, Builder.getInt64Ty());
2640 NewVal = Builder.CreateSExt(NewVal, Builder.getInt64Ty());
2641 Mask = Builder.CreateSExt(Mask, Builder.getInt64Ty());
2642 CmpXchgIntrID = Intrinsic::riscv_masked_cmpxchg_i64;
2644 Type *Tys[] = {AlignedAddr->getType()};
2645 Function *MaskedCmpXchg =
2646 Intrinsic::getDeclaration(CI->getModule(), CmpXchgIntrID, Tys);
2647 Value *Result = Builder.CreateCall(
2648 MaskedCmpXchg, {AlignedAddr, CmpVal, NewVal, Mask, Ordering});
2650 Result = Builder.CreateTrunc(Result, Builder.getInt32Ty());
2654 unsigned RISCVTargetLowering::getExceptionPointerRegister(
2655 const Constant *PersonalityFn) const {
2659 unsigned RISCVTargetLowering::getExceptionSelectorRegister(
2660 const Constant *PersonalityFn) const {