1 | //===-- RISCVISelDAGToDAG.cpp - A dag to dag inst selector for RISC-V -----===// |
2 | // |
3 | // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. |
4 | // See https://llvm.org/LICENSE.txt for license information. |
5 | // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception |
6 | // |
7 | //===----------------------------------------------------------------------===// |
8 | // |
9 | // This file defines an instruction selector for the RISC-V target. |
10 | // |
11 | //===----------------------------------------------------------------------===// |
12 | |
13 | #include "RISCVISelDAGToDAG.h" |
14 | #include "MCTargetDesc/RISCVBaseInfo.h" |
15 | #include "MCTargetDesc/RISCVMCTargetDesc.h" |
16 | #include "MCTargetDesc/RISCVMatInt.h" |
17 | #include "RISCVISelLowering.h" |
18 | #include "RISCVInstrInfo.h" |
19 | #include "RISCVSelectionDAGInfo.h" |
20 | #include "llvm/CodeGen/MachineFrameInfo.h" |
21 | #include "llvm/IR/IntrinsicsRISCV.h" |
22 | #include "llvm/Support/Alignment.h" |
23 | #include "llvm/Support/Debug.h" |
24 | #include "llvm/Support/MathExtras.h" |
25 | #include "llvm/Support/raw_ostream.h" |
26 | |
27 | using namespace llvm; |
28 | |
29 | #define DEBUG_TYPE "riscv-isel" |
30 | #define PASS_NAME "RISC-V DAG->DAG Pattern Instruction Selection" |
31 | |
32 | static cl::opt<bool> UsePseudoMovImm( |
33 | "riscv-use-rematerializable-movimm" , cl::Hidden, |
34 | cl::desc("Use a rematerializable pseudoinstruction for 2 instruction " |
35 | "constant materialization" ), |
36 | cl::init(Val: false)); |
37 | |
38 | #define GET_DAGISEL_BODY RISCVDAGToDAGISel |
39 | #include "RISCVGenDAGISel.inc" |
40 | |
41 | void RISCVDAGToDAGISel::PreprocessISelDAG() { |
42 | SelectionDAG::allnodes_iterator Position = CurDAG->allnodes_end(); |
43 | |
44 | bool MadeChange = false; |
45 | while (Position != CurDAG->allnodes_begin()) { |
46 | SDNode *N = &*--Position; |
47 | if (N->use_empty()) |
48 | continue; |
49 | |
50 | SDValue Result; |
51 | switch (N->getOpcode()) { |
52 | case ISD::SPLAT_VECTOR: { |
53 | // Convert integer SPLAT_VECTOR to VMV_V_X_VL and floating-point |
54 | // SPLAT_VECTOR to VFMV_V_F_VL to reduce isel burden. |
55 | MVT VT = N->getSimpleValueType(ResNo: 0); |
56 | unsigned Opc = |
57 | VT.isInteger() ? RISCVISD::VMV_V_X_VL : RISCVISD::VFMV_V_F_VL; |
58 | SDLoc DL(N); |
59 | SDValue VL = CurDAG->getRegister(Reg: RISCV::X0, VT: Subtarget->getXLenVT()); |
60 | SDValue Src = N->getOperand(Num: 0); |
61 | if (VT.isInteger()) |
62 | Src = CurDAG->getNode(Opcode: ISD::ANY_EXTEND, DL, VT: Subtarget->getXLenVT(), |
63 | Operand: N->getOperand(Num: 0)); |
64 | Result = CurDAG->getNode(Opcode: Opc, DL, VT, N1: CurDAG->getUNDEF(VT), N2: Src, N3: VL); |
65 | break; |
66 | } |
67 | case RISCVISD::SPLAT_VECTOR_SPLIT_I64_VL: { |
68 | // Lower SPLAT_VECTOR_SPLIT_I64 to two scalar stores and a stride 0 vector |
69 | // load. Done after lowering and combining so that we have a chance to |
70 | // optimize this to VMV_V_X_VL when the upper bits aren't needed. |
71 | assert(N->getNumOperands() == 4 && "Unexpected number of operands" ); |
72 | MVT VT = N->getSimpleValueType(ResNo: 0); |
73 | SDValue Passthru = N->getOperand(Num: 0); |
74 | SDValue Lo = N->getOperand(Num: 1); |
75 | SDValue Hi = N->getOperand(Num: 2); |
76 | SDValue VL = N->getOperand(Num: 3); |
77 | assert(VT.getVectorElementType() == MVT::i64 && VT.isScalableVector() && |
78 | Lo.getValueType() == MVT::i32 && Hi.getValueType() == MVT::i32 && |
79 | "Unexpected VTs!" ); |
80 | MachineFunction &MF = CurDAG->getMachineFunction(); |
81 | SDLoc DL(N); |
82 | |
83 | // Create temporary stack for each expanding node. |
84 | SDValue StackSlot = |
85 | CurDAG->CreateStackTemporary(Bytes: TypeSize::getFixed(ExactSize: 8), Alignment: Align(8)); |
86 | int FI = cast<FrameIndexSDNode>(Val: StackSlot.getNode())->getIndex(); |
87 | MachinePointerInfo MPI = MachinePointerInfo::getFixedStack(MF, FI); |
88 | |
89 | SDValue Chain = CurDAG->getEntryNode(); |
90 | Lo = CurDAG->getStore(Chain, dl: DL, Val: Lo, Ptr: StackSlot, PtrInfo: MPI, Alignment: Align(8)); |
91 | |
92 | SDValue OffsetSlot = |
93 | CurDAG->getMemBasePlusOffset(Base: StackSlot, Offset: TypeSize::getFixed(ExactSize: 4), DL); |
94 | Hi = CurDAG->getStore(Chain, dl: DL, Val: Hi, Ptr: OffsetSlot, PtrInfo: MPI.getWithOffset(O: 4), |
95 | Alignment: Align(8)); |
96 | |
97 | Chain = CurDAG->getNode(Opcode: ISD::TokenFactor, DL, VT: MVT::Other, N1: Lo, N2: Hi); |
98 | |
99 | SDVTList VTs = CurDAG->getVTList(VTs: {VT, MVT::Other}); |
100 | SDValue IntID = |
101 | CurDAG->getTargetConstant(Val: Intrinsic::riscv_vlse, DL, VT: MVT::i64); |
102 | SDValue Ops[] = {Chain, |
103 | IntID, |
104 | Passthru, |
105 | StackSlot, |
106 | CurDAG->getRegister(Reg: RISCV::X0, VT: MVT::i64), |
107 | VL}; |
108 | |
109 | Result = CurDAG->getMemIntrinsicNode(Opcode: ISD::INTRINSIC_W_CHAIN, dl: DL, VTList: VTs, Ops, |
110 | MemVT: MVT::i64, PtrInfo: MPI, Alignment: Align(8), |
111 | Flags: MachineMemOperand::MOLoad); |
112 | break; |
113 | } |
114 | case ISD::FP_EXTEND: { |
115 | // We only have vector patterns for riscv_fpextend_vl in isel. |
116 | SDLoc DL(N); |
117 | MVT VT = N->getSimpleValueType(ResNo: 0); |
118 | if (!VT.isVector()) |
119 | break; |
120 | SDValue VLMAX = CurDAG->getRegister(Reg: RISCV::X0, VT: Subtarget->getXLenVT()); |
121 | SDValue TrueMask = CurDAG->getNode( |
122 | Opcode: RISCVISD::VMSET_VL, DL, VT: VT.changeVectorElementType(EltVT: MVT::i1), Operand: VLMAX); |
123 | Result = CurDAG->getNode(Opcode: RISCVISD::FP_EXTEND_VL, DL, VT, N1: N->getOperand(Num: 0), |
124 | N2: TrueMask, N3: VLMAX); |
125 | break; |
126 | } |
127 | } |
128 | |
129 | if (Result) { |
130 | LLVM_DEBUG(dbgs() << "RISC-V DAG preprocessing replacing:\nOld: " ); |
131 | LLVM_DEBUG(N->dump(CurDAG)); |
132 | LLVM_DEBUG(dbgs() << "\nNew: " ); |
133 | LLVM_DEBUG(Result->dump(CurDAG)); |
134 | LLVM_DEBUG(dbgs() << "\n" ); |
135 | |
136 | CurDAG->ReplaceAllUsesOfValueWith(From: SDValue(N, 0), To: Result); |
137 | MadeChange = true; |
138 | } |
139 | } |
140 | |
141 | if (MadeChange) |
142 | CurDAG->RemoveDeadNodes(); |
143 | } |
144 | |
145 | void RISCVDAGToDAGISel::PostprocessISelDAG() { |
146 | HandleSDNode Dummy(CurDAG->getRoot()); |
147 | SelectionDAG::allnodes_iterator Position = CurDAG->allnodes_end(); |
148 | |
149 | bool MadeChange = false; |
150 | while (Position != CurDAG->allnodes_begin()) { |
151 | SDNode *N = &*--Position; |
152 | // Skip dead nodes and any non-machine opcodes. |
153 | if (N->use_empty() || !N->isMachineOpcode()) |
154 | continue; |
155 | |
156 | MadeChange |= doPeepholeSExtW(Node: N); |
157 | |
158 | // FIXME: This is here only because the VMerge transform doesn't |
159 | // know how to handle masked true inputs. Once that has been moved |
160 | // to post-ISEL, this can be deleted as well. |
161 | MadeChange |= doPeepholeMaskedRVV(Node: cast<MachineSDNode>(Val: N)); |
162 | } |
163 | |
164 | CurDAG->setRoot(Dummy.getValue()); |
165 | |
166 | MadeChange |= doPeepholeMergeVVMFold(); |
167 | |
168 | // After we're done with everything else, convert IMPLICIT_DEF |
169 | // passthru operands to NoRegister. This is required to workaround |
170 | // an optimization deficiency in MachineCSE. This really should |
171 | // be merged back into each of the patterns (i.e. there's no good |
172 | // reason not to go directly to NoReg), but is being done this way |
173 | // to allow easy backporting. |
174 | MadeChange |= doPeepholeNoRegPassThru(); |
175 | |
176 | if (MadeChange) |
177 | CurDAG->RemoveDeadNodes(); |
178 | } |
179 | |
180 | static SDValue selectImmSeq(SelectionDAG *CurDAG, const SDLoc &DL, const MVT VT, |
181 | RISCVMatInt::InstSeq &Seq) { |
182 | SDValue SrcReg = CurDAG->getRegister(Reg: RISCV::X0, VT); |
183 | for (const RISCVMatInt::Inst &Inst : Seq) { |
184 | SDValue SDImm = CurDAG->getSignedTargetConstant(Val: Inst.getImm(), DL, VT); |
185 | SDNode *Result = nullptr; |
186 | switch (Inst.getOpndKind()) { |
187 | case RISCVMatInt::Imm: |
188 | Result = CurDAG->getMachineNode(Opcode: Inst.getOpcode(), dl: DL, VT, Op1: SDImm); |
189 | break; |
190 | case RISCVMatInt::RegX0: |
191 | Result = CurDAG->getMachineNode(Opcode: Inst.getOpcode(), dl: DL, VT, Op1: SrcReg, |
192 | Op2: CurDAG->getRegister(Reg: RISCV::X0, VT)); |
193 | break; |
194 | case RISCVMatInt::RegReg: |
195 | Result = CurDAG->getMachineNode(Opcode: Inst.getOpcode(), dl: DL, VT, Op1: SrcReg, Op2: SrcReg); |
196 | break; |
197 | case RISCVMatInt::RegImm: |
198 | Result = CurDAG->getMachineNode(Opcode: Inst.getOpcode(), dl: DL, VT, Op1: SrcReg, Op2: SDImm); |
199 | break; |
200 | } |
201 | |
202 | // Only the first instruction has X0 as its source. |
203 | SrcReg = SDValue(Result, 0); |
204 | } |
205 | |
206 | return SrcReg; |
207 | } |
208 | |
209 | static SDValue selectImm(SelectionDAG *CurDAG, const SDLoc &DL, const MVT VT, |
210 | int64_t Imm, const RISCVSubtarget &Subtarget) { |
211 | RISCVMatInt::InstSeq Seq = RISCVMatInt::generateInstSeq(Val: Imm, STI: Subtarget); |
212 | |
213 | // Use a rematerializable pseudo instruction for short sequences if enabled. |
214 | if (Seq.size() == 2 && UsePseudoMovImm) |
215 | return SDValue( |
216 | CurDAG->getMachineNode(Opcode: RISCV::PseudoMovImm, dl: DL, VT, |
217 | Op1: CurDAG->getSignedTargetConstant(Val: Imm, DL, VT)), |
218 | 0); |
219 | |
220 | // See if we can create this constant as (ADD (SLLI X, C), X) where X is at |
221 | // worst an LUI+ADDIW. This will require an extra register, but avoids a |
222 | // constant pool. |
223 | // If we have Zba we can use (ADD_UW X, (SLLI X, 32)) to handle cases where |
224 | // low and high 32 bits are the same and bit 31 and 63 are set. |
225 | if (Seq.size() > 3) { |
226 | unsigned ShiftAmt, AddOpc; |
227 | RISCVMatInt::InstSeq SeqLo = |
228 | RISCVMatInt::generateTwoRegInstSeq(Val: Imm, STI: Subtarget, ShiftAmt, AddOpc); |
229 | if (!SeqLo.empty() && (SeqLo.size() + 2) < Seq.size()) { |
230 | SDValue Lo = selectImmSeq(CurDAG, DL, VT, Seq&: SeqLo); |
231 | |
232 | SDValue SLLI = SDValue( |
233 | CurDAG->getMachineNode(Opcode: RISCV::SLLI, dl: DL, VT, Op1: Lo, |
234 | Op2: CurDAG->getTargetConstant(Val: ShiftAmt, DL, VT)), |
235 | 0); |
236 | return SDValue(CurDAG->getMachineNode(Opcode: AddOpc, dl: DL, VT, Op1: Lo, Op2: SLLI), 0); |
237 | } |
238 | } |
239 | |
240 | // Otherwise, use the original sequence. |
241 | return selectImmSeq(CurDAG, DL, VT, Seq); |
242 | } |
243 | |
244 | void RISCVDAGToDAGISel::addVectorLoadStoreOperands( |
245 | SDNode *Node, unsigned Log2SEW, const SDLoc &DL, unsigned CurOp, |
246 | bool IsMasked, bool IsStridedOrIndexed, SmallVectorImpl<SDValue> &Operands, |
247 | bool IsLoad, MVT *IndexVT) { |
248 | SDValue Chain = Node->getOperand(Num: 0); |
249 | |
250 | Operands.push_back(Elt: Node->getOperand(Num: CurOp++)); // Base pointer. |
251 | |
252 | if (IsStridedOrIndexed) { |
253 | Operands.push_back(Elt: Node->getOperand(Num: CurOp++)); // Index. |
254 | if (IndexVT) |
255 | *IndexVT = Operands.back()->getSimpleValueType(ResNo: 0); |
256 | } |
257 | |
258 | if (IsMasked) { |
259 | SDValue Mask = Node->getOperand(Num: CurOp++); |
260 | Operands.push_back(Elt: Mask); |
261 | } |
262 | SDValue VL; |
263 | selectVLOp(N: Node->getOperand(Num: CurOp++), VL); |
264 | Operands.push_back(Elt: VL); |
265 | |
266 | MVT XLenVT = Subtarget->getXLenVT(); |
267 | SDValue SEWOp = CurDAG->getTargetConstant(Val: Log2SEW, DL, VT: XLenVT); |
268 | Operands.push_back(Elt: SEWOp); |
269 | |
270 | // At the IR layer, all the masked load intrinsics have policy operands, |
271 | // none of the others do. All have passthru operands. For our pseudos, |
272 | // all loads have policy operands. |
273 | if (IsLoad) { |
274 | uint64_t Policy = RISCVVType::MASK_AGNOSTIC; |
275 | if (IsMasked) |
276 | Policy = Node->getConstantOperandVal(Num: CurOp++); |
277 | SDValue PolicyOp = CurDAG->getTargetConstant(Val: Policy, DL, VT: XLenVT); |
278 | Operands.push_back(Elt: PolicyOp); |
279 | } |
280 | |
281 | Operands.push_back(Elt: Chain); // Chain. |
282 | } |
283 | |
284 | void RISCVDAGToDAGISel::selectVLSEG(SDNode *Node, unsigned NF, bool IsMasked, |
285 | bool IsStrided) { |
286 | SDLoc DL(Node); |
287 | MVT VT = Node->getSimpleValueType(ResNo: 0); |
288 | unsigned Log2SEW = Node->getConstantOperandVal(Num: Node->getNumOperands() - 1); |
289 | RISCVVType::VLMUL LMUL = RISCVTargetLowering::getLMUL(VT); |
290 | |
291 | unsigned CurOp = 2; |
292 | SmallVector<SDValue, 8> Operands; |
293 | |
294 | Operands.push_back(Elt: Node->getOperand(Num: CurOp++)); |
295 | |
296 | addVectorLoadStoreOperands(Node, Log2SEW, DL, CurOp, IsMasked, IsStridedOrIndexed: IsStrided, |
297 | Operands, /*IsLoad=*/true); |
298 | |
299 | const RISCV::VLSEGPseudo *P = |
300 | RISCV::getVLSEGPseudo(NF, Masked: IsMasked, Strided: IsStrided, /*FF*/ false, Log2SEW, |
301 | LMUL: static_cast<unsigned>(LMUL)); |
302 | MachineSDNode *Load = |
303 | CurDAG->getMachineNode(Opcode: P->Pseudo, dl: DL, VT1: MVT::Untyped, VT2: MVT::Other, Ops: Operands); |
304 | |
305 | if (auto *MemOp = dyn_cast<MemSDNode>(Val: Node)) |
306 | CurDAG->setNodeMemRefs(N: Load, NewMemRefs: {MemOp->getMemOperand()}); |
307 | |
308 | ReplaceUses(F: SDValue(Node, 0), T: SDValue(Load, 0)); |
309 | ReplaceUses(F: SDValue(Node, 1), T: SDValue(Load, 1)); |
310 | CurDAG->RemoveDeadNode(N: Node); |
311 | } |
312 | |
313 | void RISCVDAGToDAGISel::selectVLSEGFF(SDNode *Node, unsigned NF, |
314 | bool IsMasked) { |
315 | SDLoc DL(Node); |
316 | MVT VT = Node->getSimpleValueType(ResNo: 0); |
317 | MVT XLenVT = Subtarget->getXLenVT(); |
318 | unsigned Log2SEW = Node->getConstantOperandVal(Num: Node->getNumOperands() - 1); |
319 | RISCVVType::VLMUL LMUL = RISCVTargetLowering::getLMUL(VT); |
320 | |
321 | unsigned CurOp = 2; |
322 | SmallVector<SDValue, 7> Operands; |
323 | |
324 | Operands.push_back(Elt: Node->getOperand(Num: CurOp++)); |
325 | |
326 | addVectorLoadStoreOperands(Node, Log2SEW, DL, CurOp, IsMasked, |
327 | /*IsStridedOrIndexed*/ false, Operands, |
328 | /*IsLoad=*/true); |
329 | |
330 | const RISCV::VLSEGPseudo *P = |
331 | RISCV::getVLSEGPseudo(NF, Masked: IsMasked, /*Strided*/ false, /*FF*/ true, |
332 | Log2SEW, LMUL: static_cast<unsigned>(LMUL)); |
333 | MachineSDNode *Load = CurDAG->getMachineNode(Opcode: P->Pseudo, dl: DL, VT1: MVT::Untyped, |
334 | VT2: XLenVT, VT3: MVT::Other, Ops: Operands); |
335 | |
336 | if (auto *MemOp = dyn_cast<MemSDNode>(Val: Node)) |
337 | CurDAG->setNodeMemRefs(N: Load, NewMemRefs: {MemOp->getMemOperand()}); |
338 | |
339 | ReplaceUses(F: SDValue(Node, 0), T: SDValue(Load, 0)); // Result |
340 | ReplaceUses(F: SDValue(Node, 1), T: SDValue(Load, 1)); // VL |
341 | ReplaceUses(F: SDValue(Node, 2), T: SDValue(Load, 2)); // Chain |
342 | CurDAG->RemoveDeadNode(N: Node); |
343 | } |
344 | |
345 | void RISCVDAGToDAGISel::selectVLXSEG(SDNode *Node, unsigned NF, bool IsMasked, |
346 | bool IsOrdered) { |
347 | SDLoc DL(Node); |
348 | MVT VT = Node->getSimpleValueType(ResNo: 0); |
349 | unsigned Log2SEW = Node->getConstantOperandVal(Num: Node->getNumOperands() - 1); |
350 | RISCVVType::VLMUL LMUL = RISCVTargetLowering::getLMUL(VT); |
351 | |
352 | unsigned CurOp = 2; |
353 | SmallVector<SDValue, 8> Operands; |
354 | |
355 | Operands.push_back(Elt: Node->getOperand(Num: CurOp++)); |
356 | |
357 | MVT IndexVT; |
358 | addVectorLoadStoreOperands(Node, Log2SEW, DL, CurOp, IsMasked, |
359 | /*IsStridedOrIndexed*/ true, Operands, |
360 | /*IsLoad=*/true, IndexVT: &IndexVT); |
361 | |
362 | #ifndef NDEBUG |
363 | // Number of element = RVVBitsPerBlock * LMUL / SEW |
364 | unsigned ContainedTyNumElts = RISCV::RVVBitsPerBlock >> Log2SEW; |
365 | auto DecodedLMUL = RISCVVType::decodeVLMUL(LMUL); |
366 | if (DecodedLMUL.second) |
367 | ContainedTyNumElts /= DecodedLMUL.first; |
368 | else |
369 | ContainedTyNumElts *= DecodedLMUL.first; |
370 | assert(ContainedTyNumElts == IndexVT.getVectorMinNumElements() && |
371 | "Element count mismatch" ); |
372 | #endif |
373 | |
374 | RISCVVType::VLMUL IndexLMUL = RISCVTargetLowering::getLMUL(VT: IndexVT); |
375 | unsigned IndexLog2EEW = Log2_32(Value: IndexVT.getScalarSizeInBits()); |
376 | if (IndexLog2EEW == 6 && !Subtarget->is64Bit()) { |
377 | report_fatal_error(reason: "The V extension does not support EEW=64 for index " |
378 | "values when XLEN=32" ); |
379 | } |
380 | const RISCV::VLXSEGPseudo *P = RISCV::getVLXSEGPseudo( |
381 | NF, Masked: IsMasked, Ordered: IsOrdered, Log2SEW: IndexLog2EEW, LMUL: static_cast<unsigned>(LMUL), |
382 | IndexLMUL: static_cast<unsigned>(IndexLMUL)); |
383 | MachineSDNode *Load = |
384 | CurDAG->getMachineNode(Opcode: P->Pseudo, dl: DL, VT1: MVT::Untyped, VT2: MVT::Other, Ops: Operands); |
385 | |
386 | if (auto *MemOp = dyn_cast<MemSDNode>(Val: Node)) |
387 | CurDAG->setNodeMemRefs(N: Load, NewMemRefs: {MemOp->getMemOperand()}); |
388 | |
389 | ReplaceUses(F: SDValue(Node, 0), T: SDValue(Load, 0)); |
390 | ReplaceUses(F: SDValue(Node, 1), T: SDValue(Load, 1)); |
391 | CurDAG->RemoveDeadNode(N: Node); |
392 | } |
393 | |
394 | void RISCVDAGToDAGISel::selectVSSEG(SDNode *Node, unsigned NF, bool IsMasked, |
395 | bool IsStrided) { |
396 | SDLoc DL(Node); |
397 | MVT VT = Node->getOperand(Num: 2)->getSimpleValueType(ResNo: 0); |
398 | unsigned Log2SEW = Node->getConstantOperandVal(Num: Node->getNumOperands() - 1); |
399 | RISCVVType::VLMUL LMUL = RISCVTargetLowering::getLMUL(VT); |
400 | |
401 | unsigned CurOp = 2; |
402 | SmallVector<SDValue, 8> Operands; |
403 | |
404 | Operands.push_back(Elt: Node->getOperand(Num: CurOp++)); |
405 | |
406 | addVectorLoadStoreOperands(Node, Log2SEW, DL, CurOp, IsMasked, IsStridedOrIndexed: IsStrided, |
407 | Operands); |
408 | |
409 | const RISCV::VSSEGPseudo *P = RISCV::getVSSEGPseudo( |
410 | NF, Masked: IsMasked, Strided: IsStrided, Log2SEW, LMUL: static_cast<unsigned>(LMUL)); |
411 | MachineSDNode *Store = |
412 | CurDAG->getMachineNode(Opcode: P->Pseudo, dl: DL, VT: Node->getValueType(ResNo: 0), Ops: Operands); |
413 | |
414 | if (auto *MemOp = dyn_cast<MemSDNode>(Val: Node)) |
415 | CurDAG->setNodeMemRefs(N: Store, NewMemRefs: {MemOp->getMemOperand()}); |
416 | |
417 | ReplaceNode(F: Node, T: Store); |
418 | } |
419 | |
420 | void RISCVDAGToDAGISel::selectVSXSEG(SDNode *Node, unsigned NF, bool IsMasked, |
421 | bool IsOrdered) { |
422 | SDLoc DL(Node); |
423 | MVT VT = Node->getOperand(Num: 2)->getSimpleValueType(ResNo: 0); |
424 | unsigned Log2SEW = Node->getConstantOperandVal(Num: Node->getNumOperands() - 1); |
425 | RISCVVType::VLMUL LMUL = RISCVTargetLowering::getLMUL(VT); |
426 | |
427 | unsigned CurOp = 2; |
428 | SmallVector<SDValue, 8> Operands; |
429 | |
430 | Operands.push_back(Elt: Node->getOperand(Num: CurOp++)); |
431 | |
432 | MVT IndexVT; |
433 | addVectorLoadStoreOperands(Node, Log2SEW, DL, CurOp, IsMasked, |
434 | /*IsStridedOrIndexed*/ true, Operands, |
435 | /*IsLoad=*/false, IndexVT: &IndexVT); |
436 | |
437 | #ifndef NDEBUG |
438 | // Number of element = RVVBitsPerBlock * LMUL / SEW |
439 | unsigned ContainedTyNumElts = RISCV::RVVBitsPerBlock >> Log2SEW; |
440 | auto DecodedLMUL = RISCVVType::decodeVLMUL(LMUL); |
441 | if (DecodedLMUL.second) |
442 | ContainedTyNumElts /= DecodedLMUL.first; |
443 | else |
444 | ContainedTyNumElts *= DecodedLMUL.first; |
445 | assert(ContainedTyNumElts == IndexVT.getVectorMinNumElements() && |
446 | "Element count mismatch" ); |
447 | #endif |
448 | |
449 | RISCVVType::VLMUL IndexLMUL = RISCVTargetLowering::getLMUL(VT: IndexVT); |
450 | unsigned IndexLog2EEW = Log2_32(Value: IndexVT.getScalarSizeInBits()); |
451 | if (IndexLog2EEW == 6 && !Subtarget->is64Bit()) { |
452 | report_fatal_error(reason: "The V extension does not support EEW=64 for index " |
453 | "values when XLEN=32" ); |
454 | } |
455 | const RISCV::VSXSEGPseudo *P = RISCV::getVSXSEGPseudo( |
456 | NF, Masked: IsMasked, Ordered: IsOrdered, Log2SEW: IndexLog2EEW, LMUL: static_cast<unsigned>(LMUL), |
457 | IndexLMUL: static_cast<unsigned>(IndexLMUL)); |
458 | MachineSDNode *Store = |
459 | CurDAG->getMachineNode(Opcode: P->Pseudo, dl: DL, VT: Node->getValueType(ResNo: 0), Ops: Operands); |
460 | |
461 | if (auto *MemOp = dyn_cast<MemSDNode>(Val: Node)) |
462 | CurDAG->setNodeMemRefs(N: Store, NewMemRefs: {MemOp->getMemOperand()}); |
463 | |
464 | ReplaceNode(F: Node, T: Store); |
465 | } |
466 | |
467 | void RISCVDAGToDAGISel::selectVSETVLI(SDNode *Node) { |
468 | if (!Subtarget->hasVInstructions()) |
469 | return; |
470 | |
471 | assert(Node->getOpcode() == ISD::INTRINSIC_WO_CHAIN && "Unexpected opcode" ); |
472 | |
473 | SDLoc DL(Node); |
474 | MVT XLenVT = Subtarget->getXLenVT(); |
475 | |
476 | unsigned IntNo = Node->getConstantOperandVal(Num: 0); |
477 | |
478 | assert((IntNo == Intrinsic::riscv_vsetvli || |
479 | IntNo == Intrinsic::riscv_vsetvlimax) && |
480 | "Unexpected vsetvli intrinsic" ); |
481 | |
482 | bool VLMax = IntNo == Intrinsic::riscv_vsetvlimax; |
483 | unsigned Offset = (VLMax ? 1 : 2); |
484 | |
485 | assert(Node->getNumOperands() == Offset + 2 && |
486 | "Unexpected number of operands" ); |
487 | |
488 | unsigned SEW = |
489 | RISCVVType::decodeVSEW(VSEW: Node->getConstantOperandVal(Num: Offset) & 0x7); |
490 | RISCVVType::VLMUL VLMul = static_cast<RISCVVType::VLMUL>( |
491 | Node->getConstantOperandVal(Num: Offset + 1) & 0x7); |
492 | |
493 | unsigned VTypeI = RISCVVType::encodeVTYPE(VLMUL: VLMul, SEW, /*TailAgnostic*/ true, |
494 | /*MaskAgnostic*/ true); |
495 | SDValue VTypeIOp = CurDAG->getTargetConstant(Val: VTypeI, DL, VT: XLenVT); |
496 | |
497 | SDValue VLOperand; |
498 | unsigned Opcode = RISCV::PseudoVSETVLI; |
499 | if (auto *C = dyn_cast<ConstantSDNode>(Val: Node->getOperand(Num: 1))) { |
500 | if (auto VLEN = Subtarget->getRealVLen()) |
501 | if (*VLEN / RISCVVType::getSEWLMULRatio(SEW, VLMul) == C->getZExtValue()) |
502 | VLMax = true; |
503 | } |
504 | if (VLMax || isAllOnesConstant(V: Node->getOperand(Num: 1))) { |
505 | VLOperand = CurDAG->getRegister(Reg: RISCV::X0, VT: XLenVT); |
506 | Opcode = RISCV::PseudoVSETVLIX0; |
507 | } else { |
508 | VLOperand = Node->getOperand(Num: 1); |
509 | |
510 | if (auto *C = dyn_cast<ConstantSDNode>(Val&: VLOperand)) { |
511 | uint64_t AVL = C->getZExtValue(); |
512 | if (isUInt<5>(x: AVL)) { |
513 | SDValue VLImm = CurDAG->getTargetConstant(Val: AVL, DL, VT: XLenVT); |
514 | ReplaceNode(F: Node, T: CurDAG->getMachineNode(Opcode: RISCV::PseudoVSETIVLI, dl: DL, |
515 | VT: XLenVT, Op1: VLImm, Op2: VTypeIOp)); |
516 | return; |
517 | } |
518 | } |
519 | } |
520 | |
521 | ReplaceNode(F: Node, |
522 | T: CurDAG->getMachineNode(Opcode, dl: DL, VT: XLenVT, Op1: VLOperand, Op2: VTypeIOp)); |
523 | } |
524 | |
525 | bool RISCVDAGToDAGISel::tryShrinkShlLogicImm(SDNode *Node) { |
526 | MVT VT = Node->getSimpleValueType(ResNo: 0); |
527 | unsigned Opcode = Node->getOpcode(); |
528 | assert((Opcode == ISD::AND || Opcode == ISD::OR || Opcode == ISD::XOR) && |
529 | "Unexpected opcode" ); |
530 | SDLoc DL(Node); |
531 | |
532 | // For operations of the form (x << C1) op C2, check if we can use |
533 | // ANDI/ORI/XORI by transforming it into (x op (C2>>C1)) << C1. |
534 | SDValue N0 = Node->getOperand(Num: 0); |
535 | SDValue N1 = Node->getOperand(Num: 1); |
536 | |
537 | ConstantSDNode *Cst = dyn_cast<ConstantSDNode>(Val&: N1); |
538 | if (!Cst) |
539 | return false; |
540 | |
541 | int64_t Val = Cst->getSExtValue(); |
542 | |
543 | // Check if immediate can already use ANDI/ORI/XORI. |
544 | if (isInt<12>(x: Val)) |
545 | return false; |
546 | |
547 | SDValue Shift = N0; |
548 | |
549 | // If Val is simm32 and we have a sext_inreg from i32, then the binop |
550 | // produces at least 33 sign bits. We can peek through the sext_inreg and use |
551 | // a SLLIW at the end. |
552 | bool SignExt = false; |
553 | if (isInt<32>(x: Val) && N0.getOpcode() == ISD::SIGN_EXTEND_INREG && |
554 | N0.hasOneUse() && cast<VTSDNode>(Val: N0.getOperand(i: 1))->getVT() == MVT::i32) { |
555 | SignExt = true; |
556 | Shift = N0.getOperand(i: 0); |
557 | } |
558 | |
559 | if (Shift.getOpcode() != ISD::SHL || !Shift.hasOneUse()) |
560 | return false; |
561 | |
562 | ConstantSDNode *ShlCst = dyn_cast<ConstantSDNode>(Val: Shift.getOperand(i: 1)); |
563 | if (!ShlCst) |
564 | return false; |
565 | |
566 | uint64_t ShAmt = ShlCst->getZExtValue(); |
567 | |
568 | // Make sure that we don't change the operation by removing bits. |
569 | // This only matters for OR and XOR, AND is unaffected. |
570 | uint64_t RemovedBitsMask = maskTrailingOnes<uint64_t>(N: ShAmt); |
571 | if (Opcode != ISD::AND && (Val & RemovedBitsMask) != 0) |
572 | return false; |
573 | |
574 | int64_t ShiftedVal = Val >> ShAmt; |
575 | if (!isInt<12>(x: ShiftedVal)) |
576 | return false; |
577 | |
578 | // If we peeked through a sext_inreg, make sure the shift is valid for SLLIW. |
579 | if (SignExt && ShAmt >= 32) |
580 | return false; |
581 | |
582 | // Ok, we can reorder to get a smaller immediate. |
583 | unsigned BinOpc; |
584 | switch (Opcode) { |
585 | default: llvm_unreachable("Unexpected opcode" ); |
586 | case ISD::AND: BinOpc = RISCV::ANDI; break; |
587 | case ISD::OR: BinOpc = RISCV::ORI; break; |
588 | case ISD::XOR: BinOpc = RISCV::XORI; break; |
589 | } |
590 | |
591 | unsigned ShOpc = SignExt ? RISCV::SLLIW : RISCV::SLLI; |
592 | |
593 | SDNode *BinOp = CurDAG->getMachineNode( |
594 | Opcode: BinOpc, dl: DL, VT, Op1: Shift.getOperand(i: 0), |
595 | Op2: CurDAG->getSignedTargetConstant(Val: ShiftedVal, DL, VT)); |
596 | SDNode *SLLI = |
597 | CurDAG->getMachineNode(Opcode: ShOpc, dl: DL, VT, Op1: SDValue(BinOp, 0), |
598 | Op2: CurDAG->getTargetConstant(Val: ShAmt, DL, VT)); |
599 | ReplaceNode(F: Node, T: SLLI); |
600 | return true; |
601 | } |
602 | |
603 | bool RISCVDAGToDAGISel::(SDNode *Node) { |
604 | unsigned Opc; |
605 | |
606 | if (Subtarget->hasVendorXTHeadBb()) |
607 | Opc = RISCV::TH_EXT; |
608 | else if (Subtarget->hasVendorXAndesPerf()) |
609 | Opc = RISCV::NDS_BFOS; |
610 | else if (Subtarget->hasVendorXqcibm()) |
611 | Opc = RISCV::QC_EXT; |
612 | else |
613 | // Only supported with XTHeadBb/XAndesPerf/Xqcibm at the moment. |
614 | return false; |
615 | |
616 | auto *N1C = dyn_cast<ConstantSDNode>(Val: Node->getOperand(Num: 1)); |
617 | if (!N1C) |
618 | return false; |
619 | |
620 | SDValue N0 = Node->getOperand(Num: 0); |
621 | if (!N0.hasOneUse()) |
622 | return false; |
623 | |
624 | auto = [&](SDValue N0, unsigned Msb, unsigned Lsb, |
625 | const SDLoc &DL, MVT VT) { |
626 | if (Opc == RISCV::QC_EXT) { |
627 | // QC.EXT X, width, shamt |
628 | // shamt is the same as Lsb |
629 | // width is the number of bits to extract from the Lsb |
630 | Msb = Msb - Lsb + 1; |
631 | } |
632 | return CurDAG->getMachineNode(Opcode: Opc, dl: DL, VT, Op1: N0.getOperand(i: 0), |
633 | Op2: CurDAG->getTargetConstant(Val: Msb, DL, VT), |
634 | Op3: CurDAG->getTargetConstant(Val: Lsb, DL, VT)); |
635 | }; |
636 | |
637 | SDLoc DL(Node); |
638 | MVT VT = Node->getSimpleValueType(ResNo: 0); |
639 | const unsigned RightShAmt = N1C->getZExtValue(); |
640 | |
641 | // Transform (sra (shl X, C1) C2) with C1 < C2 |
642 | // -> (SignedBitfieldExtract X, msb, lsb) |
643 | if (N0.getOpcode() == ISD::SHL) { |
644 | auto *N01C = dyn_cast<ConstantSDNode>(Val: N0->getOperand(Num: 1)); |
645 | if (!N01C) |
646 | return false; |
647 | |
648 | const unsigned LeftShAmt = N01C->getZExtValue(); |
649 | // Make sure that this is a bitfield extraction (i.e., the shift-right |
650 | // amount can not be less than the left-shift). |
651 | if (LeftShAmt > RightShAmt) |
652 | return false; |
653 | |
654 | const unsigned MsbPlusOne = VT.getSizeInBits() - LeftShAmt; |
655 | const unsigned Msb = MsbPlusOne - 1; |
656 | const unsigned Lsb = RightShAmt - LeftShAmt; |
657 | |
658 | SDNode *Sbe = BitfieldExtract(N0, Msb, Lsb, DL, VT); |
659 | ReplaceNode(F: Node, T: Sbe); |
660 | return true; |
661 | } |
662 | |
663 | // Transform (sra (sext_inreg X, _), C) -> |
664 | // (SignedBitfieldExtract X, msb, lsb) |
665 | if (N0.getOpcode() == ISD::SIGN_EXTEND_INREG) { |
666 | unsigned ExtSize = |
667 | cast<VTSDNode>(Val: N0.getOperand(i: 1))->getVT().getSizeInBits(); |
668 | |
669 | // ExtSize of 32 should use sraiw via tablegen pattern. |
670 | if (ExtSize == 32) |
671 | return false; |
672 | |
673 | const unsigned Msb = ExtSize - 1; |
674 | // If the shift-right amount is greater than Msb, it means that extracts |
675 | // the X[Msb] bit and sign-extend it. |
676 | const unsigned Lsb = RightShAmt > Msb ? Msb : RightShAmt; |
677 | |
678 | SDNode *Sbe = BitfieldExtract(N0, Msb, Lsb, DL, VT); |
679 | ReplaceNode(F: Node, T: Sbe); |
680 | return true; |
681 | } |
682 | |
683 | return false; |
684 | } |
685 | |
686 | bool RISCVDAGToDAGISel::trySignedBitfieldInsertInSign(SDNode *Node) { |
687 | // Only supported with XAndesPerf at the moment. |
688 | if (!Subtarget->hasVendorXAndesPerf()) |
689 | return false; |
690 | |
691 | auto *N1C = dyn_cast<ConstantSDNode>(Val: Node->getOperand(Num: 1)); |
692 | if (!N1C) |
693 | return false; |
694 | |
695 | SDValue N0 = Node->getOperand(Num: 0); |
696 | if (!N0.hasOneUse()) |
697 | return false; |
698 | |
699 | auto BitfieldInsert = [&](SDValue N0, unsigned Msb, unsigned Lsb, |
700 | const SDLoc &DL, MVT VT) { |
701 | unsigned Opc = RISCV::NDS_BFOS; |
702 | // If the Lsb is equal to the Msb, then the Lsb should be 0. |
703 | if (Lsb == Msb) |
704 | Lsb = 0; |
705 | return CurDAG->getMachineNode(Opcode: Opc, dl: DL, VT, Op1: N0.getOperand(i: 0), |
706 | Op2: CurDAG->getTargetConstant(Val: Lsb, DL, VT), |
707 | Op3: CurDAG->getTargetConstant(Val: Msb, DL, VT)); |
708 | }; |
709 | |
710 | SDLoc DL(Node); |
711 | MVT VT = Node->getSimpleValueType(ResNo: 0); |
712 | const unsigned RightShAmt = N1C->getZExtValue(); |
713 | |
714 | // Transform (sra (shl X, C1) C2) with C1 > C2 |
715 | // -> (NDS.BFOS X, lsb, msb) |
716 | if (N0.getOpcode() == ISD::SHL) { |
717 | auto *N01C = dyn_cast<ConstantSDNode>(Val: N0->getOperand(Num: 1)); |
718 | if (!N01C) |
719 | return false; |
720 | |
721 | const unsigned LeftShAmt = N01C->getZExtValue(); |
722 | // Make sure that this is a bitfield insertion (i.e., the shift-right |
723 | // amount should be less than the left-shift). |
724 | if (LeftShAmt <= RightShAmt) |
725 | return false; |
726 | |
727 | const unsigned MsbPlusOne = VT.getSizeInBits() - RightShAmt; |
728 | const unsigned Msb = MsbPlusOne - 1; |
729 | const unsigned Lsb = LeftShAmt - RightShAmt; |
730 | |
731 | SDNode *Sbi = BitfieldInsert(N0, Msb, Lsb, DL, VT); |
732 | ReplaceNode(F: Node, T: Sbi); |
733 | return true; |
734 | } |
735 | |
736 | return false; |
737 | } |
738 | |
739 | bool RISCVDAGToDAGISel::(SDNode *Node, |
740 | const SDLoc &DL, MVT VT, |
741 | SDValue X, unsigned Msb, |
742 | unsigned Lsb) { |
743 | unsigned Opc; |
744 | |
745 | if (Subtarget->hasVendorXTHeadBb()) { |
746 | Opc = RISCV::TH_EXTU; |
747 | } else if (Subtarget->hasVendorXAndesPerf()) { |
748 | Opc = RISCV::NDS_BFOZ; |
749 | } else if (Subtarget->hasVendorXqcibm()) { |
750 | Opc = RISCV::QC_EXTU; |
751 | // QC.EXTU X, width, shamt |
752 | // shamt is the same as Lsb |
753 | // width is the number of bits to extract from the Lsb |
754 | Msb = Msb - Lsb + 1; |
755 | } else { |
756 | // Only supported with XTHeadBb/XAndesPerf/Xqcibm at the moment. |
757 | return false; |
758 | } |
759 | |
760 | SDNode *Ube = CurDAG->getMachineNode(Opcode: Opc, dl: DL, VT, Op1: X, |
761 | Op2: CurDAG->getTargetConstant(Val: Msb, DL, VT), |
762 | Op3: CurDAG->getTargetConstant(Val: Lsb, DL, VT)); |
763 | ReplaceNode(F: Node, T: Ube); |
764 | return true; |
765 | } |
766 | |
767 | bool RISCVDAGToDAGISel::tryUnsignedBitfieldInsertInZero(SDNode *Node, |
768 | const SDLoc &DL, MVT VT, |
769 | SDValue X, unsigned Msb, |
770 | unsigned Lsb) { |
771 | // Only supported with XAndesPerf at the moment. |
772 | if (!Subtarget->hasVendorXAndesPerf()) |
773 | return false; |
774 | |
775 | unsigned Opc = RISCV::NDS_BFOZ; |
776 | |
777 | // If the Lsb is equal to the Msb, then the Lsb should be 0. |
778 | if (Lsb == Msb) |
779 | Lsb = 0; |
780 | SDNode *Ubi = CurDAG->getMachineNode(Opcode: Opc, dl: DL, VT, Op1: X, |
781 | Op2: CurDAG->getTargetConstant(Val: Lsb, DL, VT), |
782 | Op3: CurDAG->getTargetConstant(Val: Msb, DL, VT)); |
783 | ReplaceNode(F: Node, T: Ubi); |
784 | return true; |
785 | } |
786 | |
787 | bool RISCVDAGToDAGISel::tryIndexedLoad(SDNode *Node) { |
788 | // Target does not support indexed loads. |
789 | if (!Subtarget->hasVendorXTHeadMemIdx()) |
790 | return false; |
791 | |
792 | LoadSDNode *Ld = cast<LoadSDNode>(Val: Node); |
793 | ISD::MemIndexedMode AM = Ld->getAddressingMode(); |
794 | if (AM == ISD::UNINDEXED) |
795 | return false; |
796 | |
797 | const ConstantSDNode *C = dyn_cast<ConstantSDNode>(Val: Ld->getOffset()); |
798 | if (!C) |
799 | return false; |
800 | |
801 | EVT LoadVT = Ld->getMemoryVT(); |
802 | assert((AM == ISD::PRE_INC || AM == ISD::POST_INC) && |
803 | "Unexpected addressing mode" ); |
804 | bool IsPre = AM == ISD::PRE_INC; |
805 | bool IsPost = AM == ISD::POST_INC; |
806 | int64_t Offset = C->getSExtValue(); |
807 | |
808 | // The constants that can be encoded in the THeadMemIdx instructions |
809 | // are of the form (sign_extend(imm5) << imm2). |
810 | unsigned Shift; |
811 | for (Shift = 0; Shift < 4; Shift++) |
812 | if (isInt<5>(x: Offset >> Shift) && ((Offset % (1LL << Shift)) == 0)) |
813 | break; |
814 | |
815 | // Constant cannot be encoded. |
816 | if (Shift == 4) |
817 | return false; |
818 | |
819 | bool IsZExt = (Ld->getExtensionType() == ISD::ZEXTLOAD); |
820 | unsigned Opcode; |
821 | if (LoadVT == MVT::i8 && IsPre) |
822 | Opcode = IsZExt ? RISCV::TH_LBUIB : RISCV::TH_LBIB; |
823 | else if (LoadVT == MVT::i8 && IsPost) |
824 | Opcode = IsZExt ? RISCV::TH_LBUIA : RISCV::TH_LBIA; |
825 | else if (LoadVT == MVT::i16 && IsPre) |
826 | Opcode = IsZExt ? RISCV::TH_LHUIB : RISCV::TH_LHIB; |
827 | else if (LoadVT == MVT::i16 && IsPost) |
828 | Opcode = IsZExt ? RISCV::TH_LHUIA : RISCV::TH_LHIA; |
829 | else if (LoadVT == MVT::i32 && IsPre) |
830 | Opcode = IsZExt ? RISCV::TH_LWUIB : RISCV::TH_LWIB; |
831 | else if (LoadVT == MVT::i32 && IsPost) |
832 | Opcode = IsZExt ? RISCV::TH_LWUIA : RISCV::TH_LWIA; |
833 | else if (LoadVT == MVT::i64 && IsPre) |
834 | Opcode = RISCV::TH_LDIB; |
835 | else if (LoadVT == MVT::i64 && IsPost) |
836 | Opcode = RISCV::TH_LDIA; |
837 | else |
838 | return false; |
839 | |
840 | EVT Ty = Ld->getOffset().getValueType(); |
841 | SDValue Ops[] = { |
842 | Ld->getBasePtr(), |
843 | CurDAG->getSignedTargetConstant(Val: Offset >> Shift, DL: SDLoc(Node), VT: Ty), |
844 | CurDAG->getTargetConstant(Val: Shift, DL: SDLoc(Node), VT: Ty), Ld->getChain()}; |
845 | SDNode *New = CurDAG->getMachineNode(Opcode, dl: SDLoc(Node), VT1: Ld->getValueType(ResNo: 0), |
846 | VT2: Ld->getValueType(ResNo: 1), VT3: MVT::Other, Ops); |
847 | |
848 | MachineMemOperand *MemOp = cast<MemSDNode>(Val: Node)->getMemOperand(); |
849 | CurDAG->setNodeMemRefs(N: cast<MachineSDNode>(Val: New), NewMemRefs: {MemOp}); |
850 | |
851 | ReplaceNode(F: Node, T: New); |
852 | |
853 | return true; |
854 | } |
855 | |
856 | void RISCVDAGToDAGISel::selectSF_VC_X_SE(SDNode *Node) { |
857 | if (!Subtarget->hasVInstructions()) |
858 | return; |
859 | |
860 | assert(Node->getOpcode() == ISD::INTRINSIC_VOID && "Unexpected opcode" ); |
861 | |
862 | SDLoc DL(Node); |
863 | unsigned IntNo = Node->getConstantOperandVal(Num: 1); |
864 | |
865 | assert((IntNo == Intrinsic::riscv_sf_vc_x_se || |
866 | IntNo == Intrinsic::riscv_sf_vc_i_se) && |
867 | "Unexpected vsetvli intrinsic" ); |
868 | |
869 | // imm, imm, imm, simm5/scalar, sew, log2lmul, vl |
870 | unsigned Log2SEW = Log2_32(Value: Node->getConstantOperandVal(Num: 6)); |
871 | SDValue SEWOp = |
872 | CurDAG->getTargetConstant(Val: Log2SEW, DL, VT: Subtarget->getXLenVT()); |
873 | SmallVector<SDValue, 8> Operands = {Node->getOperand(Num: 2), Node->getOperand(Num: 3), |
874 | Node->getOperand(Num: 4), Node->getOperand(Num: 5), |
875 | Node->getOperand(Num: 8), SEWOp, |
876 | Node->getOperand(Num: 0)}; |
877 | |
878 | unsigned Opcode; |
879 | auto *LMulSDNode = cast<ConstantSDNode>(Val: Node->getOperand(Num: 7)); |
880 | switch (LMulSDNode->getSExtValue()) { |
881 | case 5: |
882 | Opcode = IntNo == Intrinsic::riscv_sf_vc_x_se ? RISCV::PseudoSF_VC_X_SE_MF8 |
883 | : RISCV::PseudoSF_VC_I_SE_MF8; |
884 | break; |
885 | case 6: |
886 | Opcode = IntNo == Intrinsic::riscv_sf_vc_x_se ? RISCV::PseudoSF_VC_X_SE_MF4 |
887 | : RISCV::PseudoSF_VC_I_SE_MF4; |
888 | break; |
889 | case 7: |
890 | Opcode = IntNo == Intrinsic::riscv_sf_vc_x_se ? RISCV::PseudoSF_VC_X_SE_MF2 |
891 | : RISCV::PseudoSF_VC_I_SE_MF2; |
892 | break; |
893 | case 0: |
894 | Opcode = IntNo == Intrinsic::riscv_sf_vc_x_se ? RISCV::PseudoSF_VC_X_SE_M1 |
895 | : RISCV::PseudoSF_VC_I_SE_M1; |
896 | break; |
897 | case 1: |
898 | Opcode = IntNo == Intrinsic::riscv_sf_vc_x_se ? RISCV::PseudoSF_VC_X_SE_M2 |
899 | : RISCV::PseudoSF_VC_I_SE_M2; |
900 | break; |
901 | case 2: |
902 | Opcode = IntNo == Intrinsic::riscv_sf_vc_x_se ? RISCV::PseudoSF_VC_X_SE_M4 |
903 | : RISCV::PseudoSF_VC_I_SE_M4; |
904 | break; |
905 | case 3: |
906 | Opcode = IntNo == Intrinsic::riscv_sf_vc_x_se ? RISCV::PseudoSF_VC_X_SE_M8 |
907 | : RISCV::PseudoSF_VC_I_SE_M8; |
908 | break; |
909 | } |
910 | |
911 | ReplaceNode(F: Node, T: CurDAG->getMachineNode( |
912 | Opcode, dl: DL, VT: Node->getSimpleValueType(ResNo: 0), Ops: Operands)); |
913 | } |
914 | |
915 | static unsigned getSegInstNF(unsigned Intrinsic) { |
916 | #define INST_NF_CASE(NAME, NF) \ |
917 | case Intrinsic::riscv_##NAME##NF: \ |
918 | return NF; |
919 | #define INST_NF_CASE_MASK(NAME, NF) \ |
920 | case Intrinsic::riscv_##NAME##NF##_mask: \ |
921 | return NF; |
922 | #define INST_NF_CASE_FF(NAME, NF) \ |
923 | case Intrinsic::riscv_##NAME##NF##ff: \ |
924 | return NF; |
925 | #define INST_NF_CASE_FF_MASK(NAME, NF) \ |
926 | case Intrinsic::riscv_##NAME##NF##ff_mask: \ |
927 | return NF; |
928 | #define INST_ALL_NF_CASE_BASE(MACRO_NAME, NAME) \ |
929 | MACRO_NAME(NAME, 2) \ |
930 | MACRO_NAME(NAME, 3) \ |
931 | MACRO_NAME(NAME, 4) \ |
932 | MACRO_NAME(NAME, 5) \ |
933 | MACRO_NAME(NAME, 6) \ |
934 | MACRO_NAME(NAME, 7) \ |
935 | MACRO_NAME(NAME, 8) |
936 | #define INST_ALL_NF_CASE(NAME) \ |
937 | INST_ALL_NF_CASE_BASE(INST_NF_CASE, NAME) \ |
938 | INST_ALL_NF_CASE_BASE(INST_NF_CASE_MASK, NAME) |
939 | #define INST_ALL_NF_CASE_WITH_FF(NAME) \ |
940 | INST_ALL_NF_CASE(NAME) \ |
941 | INST_ALL_NF_CASE_BASE(INST_NF_CASE_FF, NAME) \ |
942 | INST_ALL_NF_CASE_BASE(INST_NF_CASE_FF_MASK, NAME) |
943 | switch (Intrinsic) { |
944 | default: |
945 | llvm_unreachable("Unexpected segment load/store intrinsic" ); |
946 | INST_ALL_NF_CASE_WITH_FF(vlseg) |
947 | INST_ALL_NF_CASE(vlsseg) |
948 | INST_ALL_NF_CASE(vloxseg) |
949 | INST_ALL_NF_CASE(vluxseg) |
950 | INST_ALL_NF_CASE(vsseg) |
951 | INST_ALL_NF_CASE(vssseg) |
952 | INST_ALL_NF_CASE(vsoxseg) |
953 | INST_ALL_NF_CASE(vsuxseg) |
954 | } |
955 | } |
956 | |
957 | void RISCVDAGToDAGISel::Select(SDNode *Node) { |
958 | // If we have a custom node, we have already selected. |
959 | if (Node->isMachineOpcode()) { |
960 | LLVM_DEBUG(dbgs() << "== " ; Node->dump(CurDAG); dbgs() << "\n" ); |
961 | Node->setNodeId(-1); |
962 | return; |
963 | } |
964 | |
965 | // Instruction Selection not handled by the auto-generated tablegen selection |
966 | // should be handled here. |
967 | unsigned Opcode = Node->getOpcode(); |
968 | MVT XLenVT = Subtarget->getXLenVT(); |
969 | SDLoc DL(Node); |
970 | MVT VT = Node->getSimpleValueType(ResNo: 0); |
971 | |
972 | bool HasBitTest = Subtarget->hasStdExtZbs() || Subtarget->hasVendorXTHeadBs(); |
973 | |
974 | switch (Opcode) { |
975 | case ISD::Constant: { |
976 | assert((VT == Subtarget->getXLenVT() || VT == MVT::i32) && "Unexpected VT" ); |
977 | auto *ConstNode = cast<ConstantSDNode>(Val: Node); |
978 | if (ConstNode->isZero()) { |
979 | SDValue New = |
980 | CurDAG->getCopyFromReg(Chain: CurDAG->getEntryNode(), dl: DL, Reg: RISCV::X0, VT); |
981 | ReplaceNode(F: Node, T: New.getNode()); |
982 | return; |
983 | } |
984 | int64_t Imm = ConstNode->getSExtValue(); |
985 | // If only the lower 8 bits are used, try to convert this to a simm6 by |
986 | // sign-extending bit 7. This is neutral without the C extension, and |
987 | // allows C.LI to be used if C is present. |
988 | if (isUInt<8>(x: Imm) && isInt<6>(x: SignExtend64<8>(x: Imm)) && hasAllBUsers(Node)) |
989 | Imm = SignExtend64<8>(x: Imm); |
990 | // If the upper XLen-16 bits are not used, try to convert this to a simm12 |
991 | // by sign extending bit 15. |
992 | if (isUInt<16>(x: Imm) && isInt<12>(x: SignExtend64<16>(x: Imm)) && |
993 | hasAllHUsers(Node)) |
994 | Imm = SignExtend64<16>(x: Imm); |
995 | // If the upper 32-bits are not used try to convert this into a simm32 by |
996 | // sign extending bit 32. |
997 | if (!isInt<32>(x: Imm) && isUInt<32>(x: Imm) && hasAllWUsers(Node)) |
998 | Imm = SignExtend64<32>(x: Imm); |
999 | |
1000 | ReplaceNode(F: Node, T: selectImm(CurDAG, DL, VT, Imm, Subtarget: *Subtarget).getNode()); |
1001 | return; |
1002 | } |
1003 | case ISD::ConstantFP: { |
1004 | const APFloat &APF = cast<ConstantFPSDNode>(Val: Node)->getValueAPF(); |
1005 | |
1006 | bool Is64Bit = Subtarget->is64Bit(); |
1007 | bool HasZdinx = Subtarget->hasStdExtZdinx(); |
1008 | |
1009 | bool NegZeroF64 = APF.isNegZero() && VT == MVT::f64; |
1010 | SDValue Imm; |
1011 | // For +0.0 or f64 -0.0 we need to start from X0. For all others, we will |
1012 | // create an integer immediate. |
1013 | if (APF.isPosZero() || NegZeroF64) { |
1014 | if (VT == MVT::f64 && HasZdinx && !Is64Bit) |
1015 | Imm = CurDAG->getRegister(Reg: RISCV::X0_Pair, VT: MVT::f64); |
1016 | else |
1017 | Imm = CurDAG->getRegister(Reg: RISCV::X0, VT: XLenVT); |
1018 | } else { |
1019 | Imm = selectImm(CurDAG, DL, VT: XLenVT, Imm: APF.bitcastToAPInt().getSExtValue(), |
1020 | Subtarget: *Subtarget); |
1021 | } |
1022 | |
1023 | unsigned Opc; |
1024 | switch (VT.SimpleTy) { |
1025 | default: |
1026 | llvm_unreachable("Unexpected size" ); |
1027 | case MVT::bf16: |
1028 | assert(Subtarget->hasStdExtZfbfmin()); |
1029 | Opc = RISCV::FMV_H_X; |
1030 | break; |
1031 | case MVT::f16: |
1032 | Opc = Subtarget->hasStdExtZhinxmin() ? RISCV::COPY : RISCV::FMV_H_X; |
1033 | break; |
1034 | case MVT::f32: |
1035 | Opc = Subtarget->hasStdExtZfinx() ? RISCV::COPY : RISCV::FMV_W_X; |
1036 | break; |
1037 | case MVT::f64: |
1038 | // For RV32, we can't move from a GPR, we need to convert instead. This |
1039 | // should only happen for +0.0 and -0.0. |
1040 | assert((Subtarget->is64Bit() || APF.isZero()) && "Unexpected constant" ); |
1041 | if (HasZdinx) |
1042 | Opc = RISCV::COPY; |
1043 | else |
1044 | Opc = Is64Bit ? RISCV::FMV_D_X : RISCV::FCVT_D_W; |
1045 | break; |
1046 | } |
1047 | |
1048 | SDNode *Res; |
1049 | if (VT.SimpleTy == MVT::f16 && Opc == RISCV::COPY) { |
1050 | Res = |
1051 | CurDAG->getTargetExtractSubreg(SRIdx: RISCV::sub_16, DL, VT, Operand: Imm).getNode(); |
1052 | } else if (VT.SimpleTy == MVT::f32 && Opc == RISCV::COPY) { |
1053 | Res = |
1054 | CurDAG->getTargetExtractSubreg(SRIdx: RISCV::sub_32, DL, VT, Operand: Imm).getNode(); |
1055 | } else if (Opc == RISCV::FCVT_D_W_IN32X || Opc == RISCV::FCVT_D_W) |
1056 | Res = CurDAG->getMachineNode( |
1057 | Opcode: Opc, dl: DL, VT, Op1: Imm, |
1058 | Op2: CurDAG->getTargetConstant(Val: RISCVFPRndMode::RNE, DL, VT: XLenVT)); |
1059 | else |
1060 | Res = CurDAG->getMachineNode(Opcode: Opc, dl: DL, VT, Op1: Imm); |
1061 | |
1062 | // For f64 -0.0, we need to insert a fneg.d idiom. |
1063 | if (NegZeroF64) { |
1064 | Opc = RISCV::FSGNJN_D; |
1065 | if (HasZdinx) |
1066 | Opc = Is64Bit ? RISCV::FSGNJN_D_INX : RISCV::FSGNJN_D_IN32X; |
1067 | Res = |
1068 | CurDAG->getMachineNode(Opcode: Opc, dl: DL, VT, Op1: SDValue(Res, 0), Op2: SDValue(Res, 0)); |
1069 | } |
1070 | |
1071 | ReplaceNode(F: Node, T: Res); |
1072 | return; |
1073 | } |
1074 | case RISCVISD::BuildGPRPair: |
1075 | case RISCVISD::BuildPairF64: { |
1076 | if (Opcode == RISCVISD::BuildPairF64 && !Subtarget->hasStdExtZdinx()) |
1077 | break; |
1078 | |
1079 | assert((!Subtarget->is64Bit() || Opcode == RISCVISD::BuildGPRPair) && |
1080 | "BuildPairF64 only handled here on rv32i_zdinx" ); |
1081 | |
1082 | SDValue Ops[] = { |
1083 | CurDAG->getTargetConstant(Val: RISCV::GPRPairRegClassID, DL, VT: MVT::i32), |
1084 | Node->getOperand(Num: 0), |
1085 | CurDAG->getTargetConstant(Val: RISCV::sub_gpr_even, DL, VT: MVT::i32), |
1086 | Node->getOperand(Num: 1), |
1087 | CurDAG->getTargetConstant(Val: RISCV::sub_gpr_odd, DL, VT: MVT::i32)}; |
1088 | |
1089 | SDNode *N = CurDAG->getMachineNode(Opcode: TargetOpcode::REG_SEQUENCE, dl: DL, VT, Ops); |
1090 | ReplaceNode(F: Node, T: N); |
1091 | return; |
1092 | } |
1093 | case RISCVISD::SplitGPRPair: |
1094 | case RISCVISD::SplitF64: { |
1095 | if (Subtarget->hasStdExtZdinx() || Opcode != RISCVISD::SplitF64) { |
1096 | assert((!Subtarget->is64Bit() || Opcode == RISCVISD::SplitGPRPair) && |
1097 | "SplitF64 only handled here on rv32i_zdinx" ); |
1098 | |
1099 | if (!SDValue(Node, 0).use_empty()) { |
1100 | SDValue Lo = CurDAG->getTargetExtractSubreg(SRIdx: RISCV::sub_gpr_even, DL, |
1101 | VT: Node->getValueType(ResNo: 0), |
1102 | Operand: Node->getOperand(Num: 0)); |
1103 | ReplaceUses(F: SDValue(Node, 0), T: Lo); |
1104 | } |
1105 | |
1106 | if (!SDValue(Node, 1).use_empty()) { |
1107 | SDValue Hi = CurDAG->getTargetExtractSubreg( |
1108 | SRIdx: RISCV::sub_gpr_odd, DL, VT: Node->getValueType(ResNo: 1), Operand: Node->getOperand(Num: 0)); |
1109 | ReplaceUses(F: SDValue(Node, 1), T: Hi); |
1110 | } |
1111 | |
1112 | CurDAG->RemoveDeadNode(N: Node); |
1113 | return; |
1114 | } |
1115 | |
1116 | assert(Opcode != RISCVISD::SplitGPRPair && |
1117 | "SplitGPRPair should already be handled" ); |
1118 | |
1119 | if (!Subtarget->hasStdExtZfa()) |
1120 | break; |
1121 | assert(Subtarget->hasStdExtD() && !Subtarget->is64Bit() && |
1122 | "Unexpected subtarget" ); |
1123 | |
1124 | // With Zfa, lower to fmv.x.w and fmvh.x.d. |
1125 | if (!SDValue(Node, 0).use_empty()) { |
1126 | SDNode *Lo = CurDAG->getMachineNode(Opcode: RISCV::FMV_X_W_FPR64, dl: DL, VT, |
1127 | Op1: Node->getOperand(Num: 0)); |
1128 | ReplaceUses(F: SDValue(Node, 0), T: SDValue(Lo, 0)); |
1129 | } |
1130 | if (!SDValue(Node, 1).use_empty()) { |
1131 | SDNode *Hi = CurDAG->getMachineNode(Opcode: RISCV::FMVH_X_D, dl: DL, VT, |
1132 | Op1: Node->getOperand(Num: 0)); |
1133 | ReplaceUses(F: SDValue(Node, 1), T: SDValue(Hi, 0)); |
1134 | } |
1135 | |
1136 | CurDAG->RemoveDeadNode(N: Node); |
1137 | return; |
1138 | } |
1139 | case ISD::SHL: { |
1140 | auto *N1C = dyn_cast<ConstantSDNode>(Val: Node->getOperand(Num: 1)); |
1141 | if (!N1C) |
1142 | break; |
1143 | SDValue N0 = Node->getOperand(Num: 0); |
1144 | if (N0.getOpcode() != ISD::AND || !N0.hasOneUse() || |
1145 | !isa<ConstantSDNode>(Val: N0.getOperand(i: 1))) |
1146 | break; |
1147 | unsigned ShAmt = N1C->getZExtValue(); |
1148 | uint64_t Mask = N0.getConstantOperandVal(i: 1); |
1149 | |
1150 | if (isShiftedMask_64(Value: Mask)) { |
1151 | unsigned XLen = Subtarget->getXLen(); |
1152 | unsigned LeadingZeros = XLen - llvm::bit_width(Value: Mask); |
1153 | unsigned TrailingZeros = llvm::countr_zero(Val: Mask); |
1154 | if (ShAmt <= 32 && TrailingZeros > 0 && LeadingZeros == 32) { |
1155 | // Optimize (shl (and X, C2), C) -> (slli (srliw X, C3), C3+C) |
1156 | // where C2 has 32 leading zeros and C3 trailing zeros. |
1157 | SDNode *SRLIW = CurDAG->getMachineNode( |
1158 | Opcode: RISCV::SRLIW, dl: DL, VT, Op1: N0->getOperand(Num: 0), |
1159 | Op2: CurDAG->getTargetConstant(Val: TrailingZeros, DL, VT)); |
1160 | SDNode *SLLI = CurDAG->getMachineNode( |
1161 | Opcode: RISCV::SLLI, dl: DL, VT, Op1: SDValue(SRLIW, 0), |
1162 | Op2: CurDAG->getTargetConstant(Val: TrailingZeros + ShAmt, DL, VT)); |
1163 | ReplaceNode(F: Node, T: SLLI); |
1164 | return; |
1165 | } |
1166 | if (TrailingZeros == 0 && LeadingZeros > ShAmt && |
1167 | XLen - LeadingZeros > 11 && LeadingZeros != 32) { |
1168 | // Optimize (shl (and X, C2), C) -> (srli (slli X, C4), C4-C) |
1169 | // where C2 has C4 leading zeros and no trailing zeros. |
1170 | // This is profitable if the "and" was to be lowered to |
1171 | // (srli (slli X, C4), C4) and not (andi X, C2). |
1172 | // For "LeadingZeros == 32": |
1173 | // - with Zba it's just (slli.uw X, C) |
1174 | // - without Zba a tablegen pattern applies the very same |
1175 | // transform as we would have done here |
1176 | SDNode *SLLI = CurDAG->getMachineNode( |
1177 | Opcode: RISCV::SLLI, dl: DL, VT, Op1: N0->getOperand(Num: 0), |
1178 | Op2: CurDAG->getTargetConstant(Val: LeadingZeros, DL, VT)); |
1179 | SDNode *SRLI = CurDAG->getMachineNode( |
1180 | Opcode: RISCV::SRLI, dl: DL, VT, Op1: SDValue(SLLI, 0), |
1181 | Op2: CurDAG->getTargetConstant(Val: LeadingZeros - ShAmt, DL, VT)); |
1182 | ReplaceNode(F: Node, T: SRLI); |
1183 | return; |
1184 | } |
1185 | } |
1186 | break; |
1187 | } |
1188 | case ISD::SRL: { |
1189 | auto *N1C = dyn_cast<ConstantSDNode>(Val: Node->getOperand(Num: 1)); |
1190 | if (!N1C) |
1191 | break; |
1192 | SDValue N0 = Node->getOperand(Num: 0); |
1193 | if (N0.getOpcode() != ISD::AND || !isa<ConstantSDNode>(Val: N0.getOperand(i: 1))) |
1194 | break; |
1195 | unsigned ShAmt = N1C->getZExtValue(); |
1196 | uint64_t Mask = N0.getConstantOperandVal(i: 1); |
1197 | |
1198 | // Optimize (srl (and X, C2), C) -> (slli (srliw X, C3), C3-C) where C2 has |
1199 | // 32 leading zeros and C3 trailing zeros. |
1200 | if (isShiftedMask_64(Value: Mask) && N0.hasOneUse()) { |
1201 | unsigned XLen = Subtarget->getXLen(); |
1202 | unsigned LeadingZeros = XLen - llvm::bit_width(Value: Mask); |
1203 | unsigned TrailingZeros = llvm::countr_zero(Val: Mask); |
1204 | if (LeadingZeros == 32 && TrailingZeros > ShAmt) { |
1205 | SDNode *SRLIW = CurDAG->getMachineNode( |
1206 | Opcode: RISCV::SRLIW, dl: DL, VT, Op1: N0->getOperand(Num: 0), |
1207 | Op2: CurDAG->getTargetConstant(Val: TrailingZeros, DL, VT)); |
1208 | SDNode *SLLI = CurDAG->getMachineNode( |
1209 | Opcode: RISCV::SLLI, dl: DL, VT, Op1: SDValue(SRLIW, 0), |
1210 | Op2: CurDAG->getTargetConstant(Val: TrailingZeros - ShAmt, DL, VT)); |
1211 | ReplaceNode(F: Node, T: SLLI); |
1212 | return; |
1213 | } |
1214 | } |
1215 | |
1216 | // Optimize (srl (and X, C2), C) -> |
1217 | // (srli (slli X, (XLen-C3), (XLen-C3) + C) |
1218 | // Where C2 is a mask with C3 trailing ones. |
1219 | // Taking into account that the C2 may have had lower bits unset by |
1220 | // SimplifyDemandedBits. This avoids materializing the C2 immediate. |
1221 | // This pattern occurs when type legalizing right shifts for types with |
1222 | // less than XLen bits. |
1223 | Mask |= maskTrailingOnes<uint64_t>(N: ShAmt); |
1224 | if (!isMask_64(Value: Mask)) |
1225 | break; |
1226 | unsigned TrailingOnes = llvm::countr_one(Value: Mask); |
1227 | if (ShAmt >= TrailingOnes) |
1228 | break; |
1229 | // If the mask has 32 trailing ones, use SRLI on RV32 or SRLIW on RV64. |
1230 | if (TrailingOnes == 32) { |
1231 | SDNode *SRLI = CurDAG->getMachineNode( |
1232 | Opcode: Subtarget->is64Bit() ? RISCV::SRLIW : RISCV::SRLI, dl: DL, VT, |
1233 | Op1: N0->getOperand(Num: 0), Op2: CurDAG->getTargetConstant(Val: ShAmt, DL, VT)); |
1234 | ReplaceNode(F: Node, T: SRLI); |
1235 | return; |
1236 | } |
1237 | |
1238 | // Only do the remaining transforms if the AND has one use. |
1239 | if (!N0.hasOneUse()) |
1240 | break; |
1241 | |
1242 | // If C2 is (1 << ShAmt) use bexti or th.tst if possible. |
1243 | if (HasBitTest && ShAmt + 1 == TrailingOnes) { |
1244 | SDNode *BEXTI = CurDAG->getMachineNode( |
1245 | Opcode: Subtarget->hasStdExtZbs() ? RISCV::BEXTI : RISCV::TH_TST, dl: DL, VT, |
1246 | Op1: N0->getOperand(Num: 0), Op2: CurDAG->getTargetConstant(Val: ShAmt, DL, VT)); |
1247 | ReplaceNode(F: Node, T: BEXTI); |
1248 | return; |
1249 | } |
1250 | |
1251 | const unsigned Msb = TrailingOnes - 1; |
1252 | const unsigned Lsb = ShAmt; |
1253 | if (tryUnsignedBitfieldExtract(Node, DL, VT, X: N0->getOperand(Num: 0), Msb, Lsb)) |
1254 | return; |
1255 | |
1256 | unsigned LShAmt = Subtarget->getXLen() - TrailingOnes; |
1257 | SDNode *SLLI = |
1258 | CurDAG->getMachineNode(Opcode: RISCV::SLLI, dl: DL, VT, Op1: N0->getOperand(Num: 0), |
1259 | Op2: CurDAG->getTargetConstant(Val: LShAmt, DL, VT)); |
1260 | SDNode *SRLI = CurDAG->getMachineNode( |
1261 | Opcode: RISCV::SRLI, dl: DL, VT, Op1: SDValue(SLLI, 0), |
1262 | Op2: CurDAG->getTargetConstant(Val: LShAmt + ShAmt, DL, VT)); |
1263 | ReplaceNode(F: Node, T: SRLI); |
1264 | return; |
1265 | } |
1266 | case ISD::SRA: { |
1267 | if (trySignedBitfieldExtract(Node)) |
1268 | return; |
1269 | |
1270 | if (trySignedBitfieldInsertInSign(Node)) |
1271 | return; |
1272 | |
1273 | // Optimize (sra (sext_inreg X, i16), C) -> |
1274 | // (srai (slli X, (XLen-16), (XLen-16) + C) |
1275 | // And (sra (sext_inreg X, i8), C) -> |
1276 | // (srai (slli X, (XLen-8), (XLen-8) + C) |
1277 | // This can occur when Zbb is enabled, which makes sext_inreg i16/i8 legal. |
1278 | // This transform matches the code we get without Zbb. The shifts are more |
1279 | // compressible, and this can help expose CSE opportunities in the sdiv by |
1280 | // constant optimization. |
1281 | auto *N1C = dyn_cast<ConstantSDNode>(Val: Node->getOperand(Num: 1)); |
1282 | if (!N1C) |
1283 | break; |
1284 | SDValue N0 = Node->getOperand(Num: 0); |
1285 | if (N0.getOpcode() != ISD::SIGN_EXTEND_INREG || !N0.hasOneUse()) |
1286 | break; |
1287 | unsigned ShAmt = N1C->getZExtValue(); |
1288 | unsigned ExtSize = |
1289 | cast<VTSDNode>(Val: N0.getOperand(i: 1))->getVT().getSizeInBits(); |
1290 | // ExtSize of 32 should use sraiw via tablegen pattern. |
1291 | if (ExtSize >= 32 || ShAmt >= ExtSize) |
1292 | break; |
1293 | unsigned LShAmt = Subtarget->getXLen() - ExtSize; |
1294 | SDNode *SLLI = |
1295 | CurDAG->getMachineNode(Opcode: RISCV::SLLI, dl: DL, VT, Op1: N0->getOperand(Num: 0), |
1296 | Op2: CurDAG->getTargetConstant(Val: LShAmt, DL, VT)); |
1297 | SDNode *SRAI = CurDAG->getMachineNode( |
1298 | Opcode: RISCV::SRAI, dl: DL, VT, Op1: SDValue(SLLI, 0), |
1299 | Op2: CurDAG->getTargetConstant(Val: LShAmt + ShAmt, DL, VT)); |
1300 | ReplaceNode(F: Node, T: SRAI); |
1301 | return; |
1302 | } |
1303 | case ISD::OR: |
1304 | case ISD::XOR: |
1305 | if (tryShrinkShlLogicImm(Node)) |
1306 | return; |
1307 | |
1308 | break; |
1309 | case ISD::AND: { |
1310 | auto *N1C = dyn_cast<ConstantSDNode>(Val: Node->getOperand(Num: 1)); |
1311 | if (!N1C) |
1312 | break; |
1313 | |
1314 | SDValue N0 = Node->getOperand(Num: 0); |
1315 | |
1316 | bool LeftShift = N0.getOpcode() == ISD::SHL; |
1317 | if (LeftShift || N0.getOpcode() == ISD::SRL) { |
1318 | auto *C = dyn_cast<ConstantSDNode>(Val: N0.getOperand(i: 1)); |
1319 | if (!C) |
1320 | break; |
1321 | unsigned C2 = C->getZExtValue(); |
1322 | unsigned XLen = Subtarget->getXLen(); |
1323 | assert((C2 > 0 && C2 < XLen) && "Unexpected shift amount!" ); |
1324 | |
1325 | // Keep track of whether this is a c.andi. If we can't use c.andi, the |
1326 | // shift pair might offer more compression opportunities. |
1327 | // TODO: We could check for C extension here, but we don't have many lit |
1328 | // tests with the C extension enabled so not checking gets better |
1329 | // coverage. |
1330 | // TODO: What if ANDI faster than shift? |
1331 | bool IsCANDI = isInt<6>(x: N1C->getSExtValue()); |
1332 | |
1333 | uint64_t C1 = N1C->getZExtValue(); |
1334 | |
1335 | // Clear irrelevant bits in the mask. |
1336 | if (LeftShift) |
1337 | C1 &= maskTrailingZeros<uint64_t>(N: C2); |
1338 | else |
1339 | C1 &= maskTrailingOnes<uint64_t>(N: XLen - C2); |
1340 | |
1341 | // Some transforms should only be done if the shift has a single use or |
1342 | // the AND would become (srli (slli X, 32), 32) |
1343 | bool OneUseOrZExtW = N0.hasOneUse() || C1 == UINT64_C(0xFFFFFFFF); |
1344 | |
1345 | SDValue X = N0.getOperand(i: 0); |
1346 | |
1347 | // Turn (and (srl x, c2) c1) -> (srli (slli x, c3-c2), c3) if c1 is a mask |
1348 | // with c3 leading zeros. |
1349 | if (!LeftShift && isMask_64(Value: C1)) { |
1350 | unsigned Leading = XLen - llvm::bit_width(Value: C1); |
1351 | if (C2 < Leading) { |
1352 | // If the number of leading zeros is C2+32 this can be SRLIW. |
1353 | if (C2 + 32 == Leading) { |
1354 | SDNode *SRLIW = CurDAG->getMachineNode( |
1355 | Opcode: RISCV::SRLIW, dl: DL, VT, Op1: X, Op2: CurDAG->getTargetConstant(Val: C2, DL, VT)); |
1356 | ReplaceNode(F: Node, T: SRLIW); |
1357 | return; |
1358 | } |
1359 | |
1360 | // (and (srl (sexti32 Y), c2), c1) -> (srliw (sraiw Y, 31), c3 - 32) |
1361 | // if c1 is a mask with c3 leading zeros and c2 >= 32 and c3-c2==1. |
1362 | // |
1363 | // This pattern occurs when (i32 (srl (sra 31), c3 - 32)) is type |
1364 | // legalized and goes through DAG combine. |
1365 | if (C2 >= 32 && (Leading - C2) == 1 && N0.hasOneUse() && |
1366 | X.getOpcode() == ISD::SIGN_EXTEND_INREG && |
1367 | cast<VTSDNode>(Val: X.getOperand(i: 1))->getVT() == MVT::i32) { |
1368 | SDNode *SRAIW = |
1369 | CurDAG->getMachineNode(Opcode: RISCV::SRAIW, dl: DL, VT, Op1: X.getOperand(i: 0), |
1370 | Op2: CurDAG->getTargetConstant(Val: 31, DL, VT)); |
1371 | SDNode *SRLIW = CurDAG->getMachineNode( |
1372 | Opcode: RISCV::SRLIW, dl: DL, VT, Op1: SDValue(SRAIW, 0), |
1373 | Op2: CurDAG->getTargetConstant(Val: Leading - 32, DL, VT)); |
1374 | ReplaceNode(F: Node, T: SRLIW); |
1375 | return; |
1376 | } |
1377 | |
1378 | // Try to use an unsigned bitfield extract (e.g., th.extu) if |
1379 | // available. |
1380 | // Transform (and (srl x, C2), C1) |
1381 | // -> (<bfextract> x, msb, lsb) |
1382 | // |
1383 | // Make sure to keep this below the SRLIW cases, as we always want to |
1384 | // prefer the more common instruction. |
1385 | const unsigned Msb = llvm::bit_width(Value: C1) + C2 - 1; |
1386 | const unsigned Lsb = C2; |
1387 | if (tryUnsignedBitfieldExtract(Node, DL, VT, X, Msb, Lsb)) |
1388 | return; |
1389 | |
1390 | // (srli (slli x, c3-c2), c3). |
1391 | // Skip if we could use (zext.w (sraiw X, C2)). |
1392 | bool Skip = Subtarget->hasStdExtZba() && Leading == 32 && |
1393 | X.getOpcode() == ISD::SIGN_EXTEND_INREG && |
1394 | cast<VTSDNode>(Val: X.getOperand(i: 1))->getVT() == MVT::i32; |
1395 | // Also Skip if we can use bexti or th.tst. |
1396 | Skip |= HasBitTest && Leading == XLen - 1; |
1397 | if (OneUseOrZExtW && !Skip) { |
1398 | SDNode *SLLI = CurDAG->getMachineNode( |
1399 | Opcode: RISCV::SLLI, dl: DL, VT, Op1: X, |
1400 | Op2: CurDAG->getTargetConstant(Val: Leading - C2, DL, VT)); |
1401 | SDNode *SRLI = CurDAG->getMachineNode( |
1402 | Opcode: RISCV::SRLI, dl: DL, VT, Op1: SDValue(SLLI, 0), |
1403 | Op2: CurDAG->getTargetConstant(Val: Leading, DL, VT)); |
1404 | ReplaceNode(F: Node, T: SRLI); |
1405 | return; |
1406 | } |
1407 | } |
1408 | } |
1409 | |
1410 | // Turn (and (shl x, c2), c1) -> (srli (slli c2+c3), c3) if c1 is a mask |
1411 | // shifted by c2 bits with c3 leading zeros. |
1412 | if (LeftShift && isShiftedMask_64(Value: C1)) { |
1413 | unsigned Leading = XLen - llvm::bit_width(Value: C1); |
1414 | |
1415 | if (C2 + Leading < XLen && |
1416 | C1 == (maskTrailingOnes<uint64_t>(N: XLen - (C2 + Leading)) << C2)) { |
1417 | // Use slli.uw when possible. |
1418 | if ((XLen - (C2 + Leading)) == 32 && Subtarget->hasStdExtZba()) { |
1419 | SDNode *SLLI_UW = |
1420 | CurDAG->getMachineNode(Opcode: RISCV::SLLI_UW, dl: DL, VT, Op1: X, |
1421 | Op2: CurDAG->getTargetConstant(Val: C2, DL, VT)); |
1422 | ReplaceNode(F: Node, T: SLLI_UW); |
1423 | return; |
1424 | } |
1425 | |
1426 | // Try to use an unsigned bitfield insert (e.g., nds.bfoz) if |
1427 | // available. |
1428 | // Transform (and (shl x, c2), c1) |
1429 | // -> (<bfinsert> x, msb, lsb) |
1430 | // e.g. |
1431 | // (and (shl x, 12), 0x00fff000) |
1432 | // If XLen = 32 and C2 = 12, then |
1433 | // Msb = 32 - 8 - 1 = 23 and Lsb = 12 |
1434 | const unsigned Msb = XLen - Leading - 1; |
1435 | const unsigned Lsb = C2; |
1436 | if (tryUnsignedBitfieldInsertInZero(Node, DL, VT, X, Msb, Lsb)) |
1437 | return; |
1438 | |
1439 | // (srli (slli c2+c3), c3) |
1440 | if (OneUseOrZExtW && !IsCANDI) { |
1441 | SDNode *SLLI = CurDAG->getMachineNode( |
1442 | Opcode: RISCV::SLLI, dl: DL, VT, Op1: X, |
1443 | Op2: CurDAG->getTargetConstant(Val: C2 + Leading, DL, VT)); |
1444 | SDNode *SRLI = CurDAG->getMachineNode( |
1445 | Opcode: RISCV::SRLI, dl: DL, VT, Op1: SDValue(SLLI, 0), |
1446 | Op2: CurDAG->getTargetConstant(Val: Leading, DL, VT)); |
1447 | ReplaceNode(F: Node, T: SRLI); |
1448 | return; |
1449 | } |
1450 | } |
1451 | } |
1452 | |
1453 | // Turn (and (shr x, c2), c1) -> (slli (srli x, c2+c3), c3) if c1 is a |
1454 | // shifted mask with c2 leading zeros and c3 trailing zeros. |
1455 | if (!LeftShift && isShiftedMask_64(Value: C1)) { |
1456 | unsigned Leading = XLen - llvm::bit_width(Value: C1); |
1457 | unsigned Trailing = llvm::countr_zero(Val: C1); |
1458 | if (Leading == C2 && C2 + Trailing < XLen && OneUseOrZExtW && |
1459 | !IsCANDI) { |
1460 | unsigned SrliOpc = RISCV::SRLI; |
1461 | // If the input is zexti32 we should use SRLIW. |
1462 | if (X.getOpcode() == ISD::AND && |
1463 | isa<ConstantSDNode>(Val: X.getOperand(i: 1)) && |
1464 | X.getConstantOperandVal(i: 1) == UINT64_C(0xFFFFFFFF)) { |
1465 | SrliOpc = RISCV::SRLIW; |
1466 | X = X.getOperand(i: 0); |
1467 | } |
1468 | SDNode *SRLI = CurDAG->getMachineNode( |
1469 | Opcode: SrliOpc, dl: DL, VT, Op1: X, |
1470 | Op2: CurDAG->getTargetConstant(Val: C2 + Trailing, DL, VT)); |
1471 | SDNode *SLLI = CurDAG->getMachineNode( |
1472 | Opcode: RISCV::SLLI, dl: DL, VT, Op1: SDValue(SRLI, 0), |
1473 | Op2: CurDAG->getTargetConstant(Val: Trailing, DL, VT)); |
1474 | ReplaceNode(F: Node, T: SLLI); |
1475 | return; |
1476 | } |
1477 | // If the leading zero count is C2+32, we can use SRLIW instead of SRLI. |
1478 | if (Leading > 32 && (Leading - 32) == C2 && C2 + Trailing < 32 && |
1479 | OneUseOrZExtW && !IsCANDI) { |
1480 | SDNode *SRLIW = CurDAG->getMachineNode( |
1481 | Opcode: RISCV::SRLIW, dl: DL, VT, Op1: X, |
1482 | Op2: CurDAG->getTargetConstant(Val: C2 + Trailing, DL, VT)); |
1483 | SDNode *SLLI = CurDAG->getMachineNode( |
1484 | Opcode: RISCV::SLLI, dl: DL, VT, Op1: SDValue(SRLIW, 0), |
1485 | Op2: CurDAG->getTargetConstant(Val: Trailing, DL, VT)); |
1486 | ReplaceNode(F: Node, T: SLLI); |
1487 | return; |
1488 | } |
1489 | // If we have 32 bits in the mask, we can use SLLI_UW instead of SLLI. |
1490 | if (Trailing > 0 && Leading + Trailing == 32 && C2 + Trailing < XLen && |
1491 | OneUseOrZExtW && Subtarget->hasStdExtZba()) { |
1492 | SDNode *SRLI = CurDAG->getMachineNode( |
1493 | Opcode: RISCV::SRLI, dl: DL, VT, Op1: X, |
1494 | Op2: CurDAG->getTargetConstant(Val: C2 + Trailing, DL, VT)); |
1495 | SDNode *SLLI_UW = CurDAG->getMachineNode( |
1496 | Opcode: RISCV::SLLI_UW, dl: DL, VT, Op1: SDValue(SRLI, 0), |
1497 | Op2: CurDAG->getTargetConstant(Val: Trailing, DL, VT)); |
1498 | ReplaceNode(F: Node, T: SLLI_UW); |
1499 | return; |
1500 | } |
1501 | } |
1502 | |
1503 | // Turn (and (shl x, c2), c1) -> (slli (srli x, c3-c2), c3) if c1 is a |
1504 | // shifted mask with no leading zeros and c3 trailing zeros. |
1505 | if (LeftShift && isShiftedMask_64(Value: C1)) { |
1506 | unsigned Leading = XLen - llvm::bit_width(Value: C1); |
1507 | unsigned Trailing = llvm::countr_zero(Val: C1); |
1508 | if (Leading == 0 && C2 < Trailing && OneUseOrZExtW && !IsCANDI) { |
1509 | SDNode *SRLI = CurDAG->getMachineNode( |
1510 | Opcode: RISCV::SRLI, dl: DL, VT, Op1: X, |
1511 | Op2: CurDAG->getTargetConstant(Val: Trailing - C2, DL, VT)); |
1512 | SDNode *SLLI = CurDAG->getMachineNode( |
1513 | Opcode: RISCV::SLLI, dl: DL, VT, Op1: SDValue(SRLI, 0), |
1514 | Op2: CurDAG->getTargetConstant(Val: Trailing, DL, VT)); |
1515 | ReplaceNode(F: Node, T: SLLI); |
1516 | return; |
1517 | } |
1518 | // If we have (32-C2) leading zeros, we can use SRLIW instead of SRLI. |
1519 | if (C2 < Trailing && Leading + C2 == 32 && OneUseOrZExtW && !IsCANDI) { |
1520 | SDNode *SRLIW = CurDAG->getMachineNode( |
1521 | Opcode: RISCV::SRLIW, dl: DL, VT, Op1: X, |
1522 | Op2: CurDAG->getTargetConstant(Val: Trailing - C2, DL, VT)); |
1523 | SDNode *SLLI = CurDAG->getMachineNode( |
1524 | Opcode: RISCV::SLLI, dl: DL, VT, Op1: SDValue(SRLIW, 0), |
1525 | Op2: CurDAG->getTargetConstant(Val: Trailing, DL, VT)); |
1526 | ReplaceNode(F: Node, T: SLLI); |
1527 | return; |
1528 | } |
1529 | |
1530 | // If we have 32 bits in the mask, we can use SLLI_UW instead of SLLI. |
1531 | if (C2 < Trailing && Leading + Trailing == 32 && OneUseOrZExtW && |
1532 | Subtarget->hasStdExtZba()) { |
1533 | SDNode *SRLI = CurDAG->getMachineNode( |
1534 | Opcode: RISCV::SRLI, dl: DL, VT, Op1: X, |
1535 | Op2: CurDAG->getTargetConstant(Val: Trailing - C2, DL, VT)); |
1536 | SDNode *SLLI_UW = CurDAG->getMachineNode( |
1537 | Opcode: RISCV::SLLI_UW, dl: DL, VT, Op1: SDValue(SRLI, 0), |
1538 | Op2: CurDAG->getTargetConstant(Val: Trailing, DL, VT)); |
1539 | ReplaceNode(F: Node, T: SLLI_UW); |
1540 | return; |
1541 | } |
1542 | } |
1543 | } |
1544 | |
1545 | const uint64_t C1 = N1C->getZExtValue(); |
1546 | |
1547 | if (N0.getOpcode() == ISD::SRA && isa<ConstantSDNode>(Val: N0.getOperand(i: 1)) && |
1548 | N0.hasOneUse()) { |
1549 | unsigned C2 = N0.getConstantOperandVal(i: 1); |
1550 | unsigned XLen = Subtarget->getXLen(); |
1551 | assert((C2 > 0 && C2 < XLen) && "Unexpected shift amount!" ); |
1552 | |
1553 | SDValue X = N0.getOperand(i: 0); |
1554 | |
1555 | // Prefer SRAIW + ANDI when possible. |
1556 | bool Skip = C2 > 32 && isInt<12>(x: N1C->getSExtValue()) && |
1557 | X.getOpcode() == ISD::SHL && |
1558 | isa<ConstantSDNode>(Val: X.getOperand(i: 1)) && |
1559 | X.getConstantOperandVal(i: 1) == 32; |
1560 | // Turn (and (sra x, c2), c1) -> (srli (srai x, c2-c3), c3) if c1 is a |
1561 | // mask with c3 leading zeros and c2 is larger than c3. |
1562 | if (isMask_64(Value: C1) && !Skip) { |
1563 | unsigned Leading = XLen - llvm::bit_width(Value: C1); |
1564 | if (C2 > Leading) { |
1565 | SDNode *SRAI = CurDAG->getMachineNode( |
1566 | Opcode: RISCV::SRAI, dl: DL, VT, Op1: X, |
1567 | Op2: CurDAG->getTargetConstant(Val: C2 - Leading, DL, VT)); |
1568 | SDNode *SRLI = CurDAG->getMachineNode( |
1569 | Opcode: RISCV::SRLI, dl: DL, VT, Op1: SDValue(SRAI, 0), |
1570 | Op2: CurDAG->getTargetConstant(Val: Leading, DL, VT)); |
1571 | ReplaceNode(F: Node, T: SRLI); |
1572 | return; |
1573 | } |
1574 | } |
1575 | |
1576 | // Look for (and (sra y, c2), c1) where c1 is a shifted mask with c3 |
1577 | // leading zeros and c4 trailing zeros. If c2 is greater than c3, we can |
1578 | // use (slli (srli (srai y, c2 - c3), c3 + c4), c4). |
1579 | if (isShiftedMask_64(Value: C1) && !Skip) { |
1580 | unsigned Leading = XLen - llvm::bit_width(Value: C1); |
1581 | unsigned Trailing = llvm::countr_zero(Val: C1); |
1582 | if (C2 > Leading && Leading > 0 && Trailing > 0) { |
1583 | SDNode *SRAI = CurDAG->getMachineNode( |
1584 | Opcode: RISCV::SRAI, dl: DL, VT, Op1: N0.getOperand(i: 0), |
1585 | Op2: CurDAG->getTargetConstant(Val: C2 - Leading, DL, VT)); |
1586 | SDNode *SRLI = CurDAG->getMachineNode( |
1587 | Opcode: RISCV::SRLI, dl: DL, VT, Op1: SDValue(SRAI, 0), |
1588 | Op2: CurDAG->getTargetConstant(Val: Leading + Trailing, DL, VT)); |
1589 | SDNode *SLLI = CurDAG->getMachineNode( |
1590 | Opcode: RISCV::SLLI, dl: DL, VT, Op1: SDValue(SRLI, 0), |
1591 | Op2: CurDAG->getTargetConstant(Val: Trailing, DL, VT)); |
1592 | ReplaceNode(F: Node, T: SLLI); |
1593 | return; |
1594 | } |
1595 | } |
1596 | } |
1597 | |
1598 | // If C1 masks off the upper bits only (but can't be formed as an |
1599 | // ANDI), use an unsigned bitfield extract (e.g., th.extu), if |
1600 | // available. |
1601 | // Transform (and x, C1) |
1602 | // -> (<bfextract> x, msb, lsb) |
1603 | if (isMask_64(Value: C1) && !isInt<12>(x: N1C->getSExtValue())) { |
1604 | const unsigned Msb = llvm::bit_width(Value: C1) - 1; |
1605 | if (tryUnsignedBitfieldExtract(Node, DL, VT, X: N0, Msb, Lsb: 0)) |
1606 | return; |
1607 | } |
1608 | |
1609 | if (tryShrinkShlLogicImm(Node)) |
1610 | return; |
1611 | |
1612 | break; |
1613 | } |
1614 | case ISD::MUL: { |
1615 | // Special case for calculating (mul (and X, C2), C1) where the full product |
1616 | // fits in XLen bits. We can shift X left by the number of leading zeros in |
1617 | // C2 and shift C1 left by XLen-lzcnt(C2). This will ensure the final |
1618 | // product has XLen trailing zeros, putting it in the output of MULHU. This |
1619 | // can avoid materializing a constant in a register for C2. |
1620 | |
1621 | // RHS should be a constant. |
1622 | auto *N1C = dyn_cast<ConstantSDNode>(Val: Node->getOperand(Num: 1)); |
1623 | if (!N1C || !N1C->hasOneUse()) |
1624 | break; |
1625 | |
1626 | // LHS should be an AND with constant. |
1627 | SDValue N0 = Node->getOperand(Num: 0); |
1628 | if (N0.getOpcode() != ISD::AND || !isa<ConstantSDNode>(Val: N0.getOperand(i: 1))) |
1629 | break; |
1630 | |
1631 | uint64_t C2 = N0.getConstantOperandVal(i: 1); |
1632 | |
1633 | // Constant should be a mask. |
1634 | if (!isMask_64(Value: C2)) |
1635 | break; |
1636 | |
1637 | // If this can be an ANDI or ZEXT.H, don't do this if the ANDI/ZEXT has |
1638 | // multiple users or the constant is a simm12. This prevents inserting a |
1639 | // shift and still have uses of the AND/ZEXT. Shifting a simm12 will likely |
1640 | // make it more costly to materialize. Otherwise, using a SLLI might allow |
1641 | // it to be compressed. |
1642 | bool IsANDIOrZExt = |
1643 | isInt<12>(x: C2) || |
1644 | (C2 == UINT64_C(0xFFFF) && Subtarget->hasStdExtZbb()); |
1645 | // With XTHeadBb, we can use TH.EXTU. |
1646 | IsANDIOrZExt |= C2 == UINT64_C(0xFFFF) && Subtarget->hasVendorXTHeadBb(); |
1647 | if (IsANDIOrZExt && (isInt<12>(x: N1C->getSExtValue()) || !N0.hasOneUse())) |
1648 | break; |
1649 | // If this can be a ZEXT.w, don't do this if the ZEXT has multiple users or |
1650 | // the constant is a simm32. |
1651 | bool IsZExtW = C2 == UINT64_C(0xFFFFFFFF) && Subtarget->hasStdExtZba(); |
1652 | // With XTHeadBb, we can use TH.EXTU. |
1653 | IsZExtW |= C2 == UINT64_C(0xFFFFFFFF) && Subtarget->hasVendorXTHeadBb(); |
1654 | if (IsZExtW && (isInt<32>(x: N1C->getSExtValue()) || !N0.hasOneUse())) |
1655 | break; |
1656 | |
1657 | // We need to shift left the AND input and C1 by a total of XLen bits. |
1658 | |
1659 | // How far left do we need to shift the AND input? |
1660 | unsigned XLen = Subtarget->getXLen(); |
1661 | unsigned LeadingZeros = XLen - llvm::bit_width(Value: C2); |
1662 | |
1663 | // The constant gets shifted by the remaining amount unless that would |
1664 | // shift bits out. |
1665 | uint64_t C1 = N1C->getZExtValue(); |
1666 | unsigned ConstantShift = XLen - LeadingZeros; |
1667 | if (ConstantShift > (XLen - llvm::bit_width(Value: C1))) |
1668 | break; |
1669 | |
1670 | uint64_t ShiftedC1 = C1 << ConstantShift; |
1671 | // If this RV32, we need to sign extend the constant. |
1672 | if (XLen == 32) |
1673 | ShiftedC1 = SignExtend64<32>(x: ShiftedC1); |
1674 | |
1675 | // Create (mulhu (slli X, lzcnt(C2)), C1 << (XLen - lzcnt(C2))). |
1676 | SDNode *Imm = selectImm(CurDAG, DL, VT, Imm: ShiftedC1, Subtarget: *Subtarget).getNode(); |
1677 | SDNode *SLLI = |
1678 | CurDAG->getMachineNode(Opcode: RISCV::SLLI, dl: DL, VT, Op1: N0.getOperand(i: 0), |
1679 | Op2: CurDAG->getTargetConstant(Val: LeadingZeros, DL, VT)); |
1680 | SDNode *MULHU = CurDAG->getMachineNode(Opcode: RISCV::MULHU, dl: DL, VT, |
1681 | Op1: SDValue(SLLI, 0), Op2: SDValue(Imm, 0)); |
1682 | ReplaceNode(F: Node, T: MULHU); |
1683 | return; |
1684 | } |
1685 | case ISD::LOAD: { |
1686 | if (tryIndexedLoad(Node)) |
1687 | return; |
1688 | |
1689 | if (Subtarget->hasVendorXCVmem() && !Subtarget->is64Bit()) { |
1690 | // We match post-incrementing load here |
1691 | LoadSDNode *Load = cast<LoadSDNode>(Val: Node); |
1692 | if (Load->getAddressingMode() != ISD::POST_INC) |
1693 | break; |
1694 | |
1695 | SDValue Chain = Node->getOperand(Num: 0); |
1696 | SDValue Base = Node->getOperand(Num: 1); |
1697 | SDValue Offset = Node->getOperand(Num: 2); |
1698 | |
1699 | bool Simm12 = false; |
1700 | bool SignExtend = Load->getExtensionType() == ISD::SEXTLOAD; |
1701 | |
1702 | if (auto ConstantOffset = dyn_cast<ConstantSDNode>(Val&: Offset)) { |
1703 | int ConstantVal = ConstantOffset->getSExtValue(); |
1704 | Simm12 = isInt<12>(x: ConstantVal); |
1705 | if (Simm12) |
1706 | Offset = CurDAG->getTargetConstant(Val: ConstantVal, DL: SDLoc(Offset), |
1707 | VT: Offset.getValueType()); |
1708 | } |
1709 | |
1710 | unsigned Opcode = 0; |
1711 | switch (Load->getMemoryVT().getSimpleVT().SimpleTy) { |
1712 | case MVT::i8: |
1713 | if (Simm12 && SignExtend) |
1714 | Opcode = RISCV::CV_LB_ri_inc; |
1715 | else if (Simm12 && !SignExtend) |
1716 | Opcode = RISCV::CV_LBU_ri_inc; |
1717 | else if (!Simm12 && SignExtend) |
1718 | Opcode = RISCV::CV_LB_rr_inc; |
1719 | else |
1720 | Opcode = RISCV::CV_LBU_rr_inc; |
1721 | break; |
1722 | case MVT::i16: |
1723 | if (Simm12 && SignExtend) |
1724 | Opcode = RISCV::CV_LH_ri_inc; |
1725 | else if (Simm12 && !SignExtend) |
1726 | Opcode = RISCV::CV_LHU_ri_inc; |
1727 | else if (!Simm12 && SignExtend) |
1728 | Opcode = RISCV::CV_LH_rr_inc; |
1729 | else |
1730 | Opcode = RISCV::CV_LHU_rr_inc; |
1731 | break; |
1732 | case MVT::i32: |
1733 | if (Simm12) |
1734 | Opcode = RISCV::CV_LW_ri_inc; |
1735 | else |
1736 | Opcode = RISCV::CV_LW_rr_inc; |
1737 | break; |
1738 | default: |
1739 | break; |
1740 | } |
1741 | if (!Opcode) |
1742 | break; |
1743 | |
1744 | ReplaceNode(F: Node, T: CurDAG->getMachineNode(Opcode, dl: DL, VT1: XLenVT, VT2: XLenVT, |
1745 | VT3: Chain.getSimpleValueType(), Op1: Base, |
1746 | Op2: Offset, Op3: Chain)); |
1747 | return; |
1748 | } |
1749 | break; |
1750 | } |
1751 | case RISCVISD::LD_RV32: { |
1752 | assert(Subtarget->hasStdExtZilsd() && "LD_RV32 is only used with Zilsd" ); |
1753 | |
1754 | SDValue Base, Offset; |
1755 | SDValue Chain = Node->getOperand(Num: 0); |
1756 | SDValue Addr = Node->getOperand(Num: 1); |
1757 | SelectAddrRegImm(Addr, Base, Offset); |
1758 | |
1759 | SDValue Ops[] = {Base, Offset, Chain}; |
1760 | MachineSDNode *New = CurDAG->getMachineNode( |
1761 | Opcode: RISCV::LD_RV32, dl: DL, ResultTys: {MVT::Untyped, MVT::Other}, Ops); |
1762 | SDValue Lo = CurDAG->getTargetExtractSubreg(SRIdx: RISCV::sub_gpr_even, DL, |
1763 | VT: MVT::i32, Operand: SDValue(New, 0)); |
1764 | SDValue Hi = CurDAG->getTargetExtractSubreg(SRIdx: RISCV::sub_gpr_odd, DL, |
1765 | VT: MVT::i32, Operand: SDValue(New, 0)); |
1766 | CurDAG->setNodeMemRefs(N: New, NewMemRefs: {cast<MemSDNode>(Val: Node)->getMemOperand()}); |
1767 | ReplaceUses(F: SDValue(Node, 0), T: Lo); |
1768 | ReplaceUses(F: SDValue(Node, 1), T: Hi); |
1769 | ReplaceUses(F: SDValue(Node, 2), T: SDValue(New, 1)); |
1770 | CurDAG->RemoveDeadNode(N: Node); |
1771 | return; |
1772 | } |
1773 | case RISCVISD::SD_RV32: { |
1774 | SDValue Base, Offset; |
1775 | SDValue Chain = Node->getOperand(Num: 0); |
1776 | SDValue Addr = Node->getOperand(Num: 3); |
1777 | SelectAddrRegImm(Addr, Base, Offset); |
1778 | |
1779 | SDValue Lo = Node->getOperand(Num: 1); |
1780 | SDValue Hi = Node->getOperand(Num: 2); |
1781 | |
1782 | SDValue RegPair; |
1783 | // Peephole to use X0_Pair for storing zero. |
1784 | if (isNullConstant(V: Lo) && isNullConstant(V: Hi)) { |
1785 | RegPair = CurDAG->getRegister(Reg: RISCV::X0_Pair, VT: MVT::Untyped); |
1786 | } else { |
1787 | SDValue Ops[] = { |
1788 | CurDAG->getTargetConstant(Val: RISCV::GPRPairRegClassID, DL, VT: MVT::i32), Lo, |
1789 | CurDAG->getTargetConstant(Val: RISCV::sub_gpr_even, DL, VT: MVT::i32), Hi, |
1790 | CurDAG->getTargetConstant(Val: RISCV::sub_gpr_odd, DL, VT: MVT::i32)}; |
1791 | |
1792 | RegPair = SDValue(CurDAG->getMachineNode(Opcode: TargetOpcode::REG_SEQUENCE, dl: DL, |
1793 | VT: MVT::Untyped, Ops), |
1794 | 0); |
1795 | } |
1796 | |
1797 | MachineSDNode *New = CurDAG->getMachineNode(Opcode: RISCV::SD_RV32, dl: DL, VT: MVT::Other, |
1798 | Ops: {RegPair, Base, Offset, Chain}); |
1799 | CurDAG->setNodeMemRefs(N: New, NewMemRefs: {cast<MemSDNode>(Val: Node)->getMemOperand()}); |
1800 | ReplaceUses(F: SDValue(Node, 0), T: SDValue(New, 0)); |
1801 | CurDAG->RemoveDeadNode(N: Node); |
1802 | return; |
1803 | } |
1804 | case ISD::INTRINSIC_WO_CHAIN: { |
1805 | unsigned IntNo = Node->getConstantOperandVal(Num: 0); |
1806 | switch (IntNo) { |
1807 | // By default we do not custom select any intrinsic. |
1808 | default: |
1809 | break; |
1810 | case Intrinsic::riscv_vmsgeu: |
1811 | case Intrinsic::riscv_vmsge: { |
1812 | SDValue Src1 = Node->getOperand(Num: 1); |
1813 | SDValue Src2 = Node->getOperand(Num: 2); |
1814 | bool IsUnsigned = IntNo == Intrinsic::riscv_vmsgeu; |
1815 | bool IsCmpConstant = false; |
1816 | bool IsCmpMinimum = false; |
1817 | // Only custom select scalar second operand. |
1818 | if (Src2.getValueType() != XLenVT) |
1819 | break; |
1820 | // Small constants are handled with patterns. |
1821 | int64_t CVal = 0; |
1822 | MVT Src1VT = Src1.getSimpleValueType(); |
1823 | if (auto *C = dyn_cast<ConstantSDNode>(Val&: Src2)) { |
1824 | IsCmpConstant = true; |
1825 | CVal = C->getSExtValue(); |
1826 | if (CVal >= -15 && CVal <= 16) { |
1827 | if (!IsUnsigned || CVal != 0) |
1828 | break; |
1829 | IsCmpMinimum = true; |
1830 | } else if (!IsUnsigned && CVal == APInt::getSignedMinValue( |
1831 | numBits: Src1VT.getScalarSizeInBits()) |
1832 | .getSExtValue()) { |
1833 | IsCmpMinimum = true; |
1834 | } |
1835 | } |
1836 | unsigned VMSLTOpcode, VMNANDOpcode, VMSetOpcode, VMSGTOpcode; |
1837 | switch (RISCVTargetLowering::getLMUL(VT: Src1VT)) { |
1838 | default: |
1839 | llvm_unreachable("Unexpected LMUL!" ); |
1840 | #define CASE_VMSLT_OPCODES(lmulenum, suffix) \ |
1841 | case RISCVVType::lmulenum: \ |
1842 | VMSLTOpcode = IsUnsigned ? RISCV::PseudoVMSLTU_VX_##suffix \ |
1843 | : RISCV::PseudoVMSLT_VX_##suffix; \ |
1844 | VMSGTOpcode = IsUnsigned ? RISCV::PseudoVMSGTU_VX_##suffix \ |
1845 | : RISCV::PseudoVMSGT_VX_##suffix; \ |
1846 | break; |
1847 | CASE_VMSLT_OPCODES(LMUL_F8, MF8) |
1848 | CASE_VMSLT_OPCODES(LMUL_F4, MF4) |
1849 | CASE_VMSLT_OPCODES(LMUL_F2, MF2) |
1850 | CASE_VMSLT_OPCODES(LMUL_1, M1) |
1851 | CASE_VMSLT_OPCODES(LMUL_2, M2) |
1852 | CASE_VMSLT_OPCODES(LMUL_4, M4) |
1853 | CASE_VMSLT_OPCODES(LMUL_8, M8) |
1854 | #undef CASE_VMSLT_OPCODES |
1855 | } |
1856 | // Mask operations use the LMUL from the mask type. |
1857 | switch (RISCVTargetLowering::getLMUL(VT)) { |
1858 | default: |
1859 | llvm_unreachable("Unexpected LMUL!" ); |
1860 | #define CASE_VMNAND_VMSET_OPCODES(lmulenum, suffix) \ |
1861 | case RISCVVType::lmulenum: \ |
1862 | VMNANDOpcode = RISCV::PseudoVMNAND_MM_##suffix; \ |
1863 | VMSetOpcode = RISCV::PseudoVMSET_M_##suffix; \ |
1864 | break; |
1865 | CASE_VMNAND_VMSET_OPCODES(LMUL_F8, B64) |
1866 | CASE_VMNAND_VMSET_OPCODES(LMUL_F4, B32) |
1867 | CASE_VMNAND_VMSET_OPCODES(LMUL_F2, B16) |
1868 | CASE_VMNAND_VMSET_OPCODES(LMUL_1, B8) |
1869 | CASE_VMNAND_VMSET_OPCODES(LMUL_2, B4) |
1870 | CASE_VMNAND_VMSET_OPCODES(LMUL_4, B2) |
1871 | CASE_VMNAND_VMSET_OPCODES(LMUL_8, B1) |
1872 | #undef CASE_VMNAND_VMSET_OPCODES |
1873 | } |
1874 | SDValue SEW = CurDAG->getTargetConstant( |
1875 | Val: Log2_32(Value: Src1VT.getScalarSizeInBits()), DL, VT: XLenVT); |
1876 | SDValue MaskSEW = CurDAG->getTargetConstant(Val: 0, DL, VT: XLenVT); |
1877 | SDValue VL; |
1878 | selectVLOp(N: Node->getOperand(Num: 3), VL); |
1879 | |
1880 | // If vmsge(u) with minimum value, expand it to vmset. |
1881 | if (IsCmpMinimum) { |
1882 | ReplaceNode(F: Node, |
1883 | T: CurDAG->getMachineNode(Opcode: VMSetOpcode, dl: DL, VT, Op1: VL, Op2: MaskSEW)); |
1884 | return; |
1885 | } |
1886 | |
1887 | if (IsCmpConstant) { |
1888 | SDValue Imm = |
1889 | selectImm(CurDAG, DL: SDLoc(Src2), VT: XLenVT, Imm: CVal - 1, Subtarget: *Subtarget); |
1890 | |
1891 | ReplaceNode(F: Node, T: CurDAG->getMachineNode(Opcode: VMSGTOpcode, dl: DL, VT, |
1892 | Ops: {Src1, Imm, VL, SEW})); |
1893 | return; |
1894 | } |
1895 | |
1896 | // Expand to |
1897 | // vmslt{u}.vx vd, va, x; vmnand.mm vd, vd, vd |
1898 | SDValue Cmp = SDValue( |
1899 | CurDAG->getMachineNode(Opcode: VMSLTOpcode, dl: DL, VT, Ops: {Src1, Src2, VL, SEW}), |
1900 | 0); |
1901 | ReplaceNode(F: Node, T: CurDAG->getMachineNode(Opcode: VMNANDOpcode, dl: DL, VT, |
1902 | Ops: {Cmp, Cmp, VL, MaskSEW})); |
1903 | return; |
1904 | } |
1905 | case Intrinsic::riscv_vmsgeu_mask: |
1906 | case Intrinsic::riscv_vmsge_mask: { |
1907 | SDValue Src1 = Node->getOperand(Num: 2); |
1908 | SDValue Src2 = Node->getOperand(Num: 3); |
1909 | bool IsUnsigned = IntNo == Intrinsic::riscv_vmsgeu_mask; |
1910 | bool IsCmpConstant = false; |
1911 | bool IsCmpMinimum = false; |
1912 | // Only custom select scalar second operand. |
1913 | if (Src2.getValueType() != XLenVT) |
1914 | break; |
1915 | // Small constants are handled with patterns. |
1916 | MVT Src1VT = Src1.getSimpleValueType(); |
1917 | int64_t CVal = 0; |
1918 | if (auto *C = dyn_cast<ConstantSDNode>(Val&: Src2)) { |
1919 | IsCmpConstant = true; |
1920 | CVal = C->getSExtValue(); |
1921 | if (CVal >= -15 && CVal <= 16) { |
1922 | if (!IsUnsigned || CVal != 0) |
1923 | break; |
1924 | IsCmpMinimum = true; |
1925 | } else if (!IsUnsigned && CVal == APInt::getSignedMinValue( |
1926 | numBits: Src1VT.getScalarSizeInBits()) |
1927 | .getSExtValue()) { |
1928 | IsCmpMinimum = true; |
1929 | } |
1930 | } |
1931 | unsigned VMSLTOpcode, VMSLTMaskOpcode, VMXOROpcode, VMANDNOpcode, |
1932 | VMOROpcode, VMSGTMaskOpcode; |
1933 | switch (RISCVTargetLowering::getLMUL(VT: Src1VT)) { |
1934 | default: |
1935 | llvm_unreachable("Unexpected LMUL!" ); |
1936 | #define CASE_VMSLT_OPCODES(lmulenum, suffix) \ |
1937 | case RISCVVType::lmulenum: \ |
1938 | VMSLTOpcode = IsUnsigned ? RISCV::PseudoVMSLTU_VX_##suffix \ |
1939 | : RISCV::PseudoVMSLT_VX_##suffix; \ |
1940 | VMSLTMaskOpcode = IsUnsigned ? RISCV::PseudoVMSLTU_VX_##suffix##_MASK \ |
1941 | : RISCV::PseudoVMSLT_VX_##suffix##_MASK; \ |
1942 | VMSGTMaskOpcode = IsUnsigned ? RISCV::PseudoVMSGTU_VX_##suffix##_MASK \ |
1943 | : RISCV::PseudoVMSGT_VX_##suffix##_MASK; \ |
1944 | break; |
1945 | CASE_VMSLT_OPCODES(LMUL_F8, MF8) |
1946 | CASE_VMSLT_OPCODES(LMUL_F4, MF4) |
1947 | CASE_VMSLT_OPCODES(LMUL_F2, MF2) |
1948 | CASE_VMSLT_OPCODES(LMUL_1, M1) |
1949 | CASE_VMSLT_OPCODES(LMUL_2, M2) |
1950 | CASE_VMSLT_OPCODES(LMUL_4, M4) |
1951 | CASE_VMSLT_OPCODES(LMUL_8, M8) |
1952 | #undef CASE_VMSLT_OPCODES |
1953 | } |
1954 | // Mask operations use the LMUL from the mask type. |
1955 | switch (RISCVTargetLowering::getLMUL(VT)) { |
1956 | default: |
1957 | llvm_unreachable("Unexpected LMUL!" ); |
1958 | #define CASE_VMXOR_VMANDN_VMOR_OPCODES(lmulenum, suffix) \ |
1959 | case RISCVVType::lmulenum: \ |
1960 | VMXOROpcode = RISCV::PseudoVMXOR_MM_##suffix; \ |
1961 | VMANDNOpcode = RISCV::PseudoVMANDN_MM_##suffix; \ |
1962 | VMOROpcode = RISCV::PseudoVMOR_MM_##suffix; \ |
1963 | break; |
1964 | CASE_VMXOR_VMANDN_VMOR_OPCODES(LMUL_F8, B64) |
1965 | CASE_VMXOR_VMANDN_VMOR_OPCODES(LMUL_F4, B32) |
1966 | CASE_VMXOR_VMANDN_VMOR_OPCODES(LMUL_F2, B16) |
1967 | CASE_VMXOR_VMANDN_VMOR_OPCODES(LMUL_1, B8) |
1968 | CASE_VMXOR_VMANDN_VMOR_OPCODES(LMUL_2, B4) |
1969 | CASE_VMXOR_VMANDN_VMOR_OPCODES(LMUL_4, B2) |
1970 | CASE_VMXOR_VMANDN_VMOR_OPCODES(LMUL_8, B1) |
1971 | #undef CASE_VMXOR_VMANDN_VMOR_OPCODES |
1972 | } |
1973 | SDValue SEW = CurDAG->getTargetConstant( |
1974 | Val: Log2_32(Value: Src1VT.getScalarSizeInBits()), DL, VT: XLenVT); |
1975 | SDValue MaskSEW = CurDAG->getTargetConstant(Val: 0, DL, VT: XLenVT); |
1976 | SDValue VL; |
1977 | selectVLOp(N: Node->getOperand(Num: 5), VL); |
1978 | SDValue MaskedOff = Node->getOperand(Num: 1); |
1979 | SDValue Mask = Node->getOperand(Num: 4); |
1980 | |
1981 | // If vmsge(u) with minimum value, expand it to vmor mask, maskedoff. |
1982 | if (IsCmpMinimum) { |
1983 | // We don't need vmor if the MaskedOff and the Mask are the same |
1984 | // value. |
1985 | if (Mask == MaskedOff) { |
1986 | ReplaceUses(F: Node, T: Mask.getNode()); |
1987 | return; |
1988 | } |
1989 | ReplaceNode(F: Node, |
1990 | T: CurDAG->getMachineNode(Opcode: VMOROpcode, dl: DL, VT, |
1991 | Ops: {Mask, MaskedOff, VL, MaskSEW})); |
1992 | return; |
1993 | } |
1994 | |
1995 | // If the MaskedOff value and the Mask are the same value use |
1996 | // vmslt{u}.vx vt, va, x; vmandn.mm vd, vd, vt |
1997 | // This avoids needing to copy v0 to vd before starting the next sequence. |
1998 | if (Mask == MaskedOff) { |
1999 | SDValue Cmp = SDValue( |
2000 | CurDAG->getMachineNode(Opcode: VMSLTOpcode, dl: DL, VT, Ops: {Src1, Src2, VL, SEW}), |
2001 | 0); |
2002 | ReplaceNode(F: Node, T: CurDAG->getMachineNode(Opcode: VMANDNOpcode, dl: DL, VT, |
2003 | Ops: {Mask, Cmp, VL, MaskSEW})); |
2004 | return; |
2005 | } |
2006 | |
2007 | SDValue PolicyOp = |
2008 | CurDAG->getTargetConstant(Val: RISCVVType::TAIL_AGNOSTIC, DL, VT: XLenVT); |
2009 | |
2010 | if (IsCmpConstant) { |
2011 | SDValue Imm = |
2012 | selectImm(CurDAG, DL: SDLoc(Src2), VT: XLenVT, Imm: CVal - 1, Subtarget: *Subtarget); |
2013 | |
2014 | ReplaceNode(F: Node, T: CurDAG->getMachineNode( |
2015 | Opcode: VMSGTMaskOpcode, dl: DL, VT, |
2016 | Ops: {MaskedOff, Src1, Imm, Mask, VL, SEW, PolicyOp})); |
2017 | return; |
2018 | } |
2019 | |
2020 | // Otherwise use |
2021 | // vmslt{u}.vx vd, va, x, v0.t; vmxor.mm vd, vd, v0 |
2022 | // The result is mask undisturbed. |
2023 | // We use the same instructions to emulate mask agnostic behavior, because |
2024 | // the agnostic result can be either undisturbed or all 1. |
2025 | SDValue Cmp = SDValue(CurDAG->getMachineNode(Opcode: VMSLTMaskOpcode, dl: DL, VT, |
2026 | Ops: {MaskedOff, Src1, Src2, Mask, |
2027 | VL, SEW, PolicyOp}), |
2028 | 0); |
2029 | // vmxor.mm vd, vd, v0 is used to update active value. |
2030 | ReplaceNode(F: Node, T: CurDAG->getMachineNode(Opcode: VMXOROpcode, dl: DL, VT, |
2031 | Ops: {Cmp, Mask, VL, MaskSEW})); |
2032 | return; |
2033 | } |
2034 | case Intrinsic::riscv_vsetvli: |
2035 | case Intrinsic::riscv_vsetvlimax: |
2036 | return selectVSETVLI(Node); |
2037 | } |
2038 | break; |
2039 | } |
2040 | case ISD::INTRINSIC_W_CHAIN: { |
2041 | unsigned IntNo = Node->getConstantOperandVal(Num: 1); |
2042 | switch (IntNo) { |
2043 | // By default we do not custom select any intrinsic. |
2044 | default: |
2045 | break; |
2046 | case Intrinsic::riscv_vlseg2: |
2047 | case Intrinsic::riscv_vlseg3: |
2048 | case Intrinsic::riscv_vlseg4: |
2049 | case Intrinsic::riscv_vlseg5: |
2050 | case Intrinsic::riscv_vlseg6: |
2051 | case Intrinsic::riscv_vlseg7: |
2052 | case Intrinsic::riscv_vlseg8: { |
2053 | selectVLSEG(Node, NF: getSegInstNF(Intrinsic: IntNo), /*IsMasked*/ false, |
2054 | /*IsStrided*/ false); |
2055 | return; |
2056 | } |
2057 | case Intrinsic::riscv_vlseg2_mask: |
2058 | case Intrinsic::riscv_vlseg3_mask: |
2059 | case Intrinsic::riscv_vlseg4_mask: |
2060 | case Intrinsic::riscv_vlseg5_mask: |
2061 | case Intrinsic::riscv_vlseg6_mask: |
2062 | case Intrinsic::riscv_vlseg7_mask: |
2063 | case Intrinsic::riscv_vlseg8_mask: { |
2064 | selectVLSEG(Node, NF: getSegInstNF(Intrinsic: IntNo), /*IsMasked*/ true, |
2065 | /*IsStrided*/ false); |
2066 | return; |
2067 | } |
2068 | case Intrinsic::riscv_vlsseg2: |
2069 | case Intrinsic::riscv_vlsseg3: |
2070 | case Intrinsic::riscv_vlsseg4: |
2071 | case Intrinsic::riscv_vlsseg5: |
2072 | case Intrinsic::riscv_vlsseg6: |
2073 | case Intrinsic::riscv_vlsseg7: |
2074 | case Intrinsic::riscv_vlsseg8: { |
2075 | selectVLSEG(Node, NF: getSegInstNF(Intrinsic: IntNo), /*IsMasked*/ false, |
2076 | /*IsStrided*/ true); |
2077 | return; |
2078 | } |
2079 | case Intrinsic::riscv_vlsseg2_mask: |
2080 | case Intrinsic::riscv_vlsseg3_mask: |
2081 | case Intrinsic::riscv_vlsseg4_mask: |
2082 | case Intrinsic::riscv_vlsseg5_mask: |
2083 | case Intrinsic::riscv_vlsseg6_mask: |
2084 | case Intrinsic::riscv_vlsseg7_mask: |
2085 | case Intrinsic::riscv_vlsseg8_mask: { |
2086 | selectVLSEG(Node, NF: getSegInstNF(Intrinsic: IntNo), /*IsMasked*/ true, |
2087 | /*IsStrided*/ true); |
2088 | return; |
2089 | } |
2090 | case Intrinsic::riscv_vloxseg2: |
2091 | case Intrinsic::riscv_vloxseg3: |
2092 | case Intrinsic::riscv_vloxseg4: |
2093 | case Intrinsic::riscv_vloxseg5: |
2094 | case Intrinsic::riscv_vloxseg6: |
2095 | case Intrinsic::riscv_vloxseg7: |
2096 | case Intrinsic::riscv_vloxseg8: |
2097 | selectVLXSEG(Node, NF: getSegInstNF(Intrinsic: IntNo), /*IsMasked*/ false, |
2098 | /*IsOrdered*/ true); |
2099 | return; |
2100 | case Intrinsic::riscv_vluxseg2: |
2101 | case Intrinsic::riscv_vluxseg3: |
2102 | case Intrinsic::riscv_vluxseg4: |
2103 | case Intrinsic::riscv_vluxseg5: |
2104 | case Intrinsic::riscv_vluxseg6: |
2105 | case Intrinsic::riscv_vluxseg7: |
2106 | case Intrinsic::riscv_vluxseg8: |
2107 | selectVLXSEG(Node, NF: getSegInstNF(Intrinsic: IntNo), /*IsMasked*/ false, |
2108 | /*IsOrdered*/ false); |
2109 | return; |
2110 | case Intrinsic::riscv_vloxseg2_mask: |
2111 | case Intrinsic::riscv_vloxseg3_mask: |
2112 | case Intrinsic::riscv_vloxseg4_mask: |
2113 | case Intrinsic::riscv_vloxseg5_mask: |
2114 | case Intrinsic::riscv_vloxseg6_mask: |
2115 | case Intrinsic::riscv_vloxseg7_mask: |
2116 | case Intrinsic::riscv_vloxseg8_mask: |
2117 | selectVLXSEG(Node, NF: getSegInstNF(Intrinsic: IntNo), /*IsMasked*/ true, |
2118 | /*IsOrdered*/ true); |
2119 | return; |
2120 | case Intrinsic::riscv_vluxseg2_mask: |
2121 | case Intrinsic::riscv_vluxseg3_mask: |
2122 | case Intrinsic::riscv_vluxseg4_mask: |
2123 | case Intrinsic::riscv_vluxseg5_mask: |
2124 | case Intrinsic::riscv_vluxseg6_mask: |
2125 | case Intrinsic::riscv_vluxseg7_mask: |
2126 | case Intrinsic::riscv_vluxseg8_mask: |
2127 | selectVLXSEG(Node, NF: getSegInstNF(Intrinsic: IntNo), /*IsMasked*/ true, |
2128 | /*IsOrdered*/ false); |
2129 | return; |
2130 | case Intrinsic::riscv_vlseg8ff: |
2131 | case Intrinsic::riscv_vlseg7ff: |
2132 | case Intrinsic::riscv_vlseg6ff: |
2133 | case Intrinsic::riscv_vlseg5ff: |
2134 | case Intrinsic::riscv_vlseg4ff: |
2135 | case Intrinsic::riscv_vlseg3ff: |
2136 | case Intrinsic::riscv_vlseg2ff: { |
2137 | selectVLSEGFF(Node, NF: getSegInstNF(Intrinsic: IntNo), /*IsMasked*/ false); |
2138 | return; |
2139 | } |
2140 | case Intrinsic::riscv_vlseg8ff_mask: |
2141 | case Intrinsic::riscv_vlseg7ff_mask: |
2142 | case Intrinsic::riscv_vlseg6ff_mask: |
2143 | case Intrinsic::riscv_vlseg5ff_mask: |
2144 | case Intrinsic::riscv_vlseg4ff_mask: |
2145 | case Intrinsic::riscv_vlseg3ff_mask: |
2146 | case Intrinsic::riscv_vlseg2ff_mask: { |
2147 | selectVLSEGFF(Node, NF: getSegInstNF(Intrinsic: IntNo), /*IsMasked*/ true); |
2148 | return; |
2149 | } |
2150 | case Intrinsic::riscv_vloxei: |
2151 | case Intrinsic::riscv_vloxei_mask: |
2152 | case Intrinsic::riscv_vluxei: |
2153 | case Intrinsic::riscv_vluxei_mask: { |
2154 | bool IsMasked = IntNo == Intrinsic::riscv_vloxei_mask || |
2155 | IntNo == Intrinsic::riscv_vluxei_mask; |
2156 | bool IsOrdered = IntNo == Intrinsic::riscv_vloxei || |
2157 | IntNo == Intrinsic::riscv_vloxei_mask; |
2158 | |
2159 | MVT VT = Node->getSimpleValueType(ResNo: 0); |
2160 | unsigned Log2SEW = Log2_32(Value: VT.getScalarSizeInBits()); |
2161 | |
2162 | unsigned CurOp = 2; |
2163 | SmallVector<SDValue, 8> Operands; |
2164 | Operands.push_back(Elt: Node->getOperand(Num: CurOp++)); |
2165 | |
2166 | MVT IndexVT; |
2167 | addVectorLoadStoreOperands(Node, Log2SEW, DL, CurOp, IsMasked, |
2168 | /*IsStridedOrIndexed*/ true, Operands, |
2169 | /*IsLoad=*/true, IndexVT: &IndexVT); |
2170 | |
2171 | assert(VT.getVectorElementCount() == IndexVT.getVectorElementCount() && |
2172 | "Element count mismatch" ); |
2173 | |
2174 | RISCVVType::VLMUL LMUL = RISCVTargetLowering::getLMUL(VT); |
2175 | RISCVVType::VLMUL IndexLMUL = RISCVTargetLowering::getLMUL(VT: IndexVT); |
2176 | unsigned IndexLog2EEW = Log2_32(Value: IndexVT.getScalarSizeInBits()); |
2177 | if (IndexLog2EEW == 6 && !Subtarget->is64Bit()) { |
2178 | report_fatal_error(reason: "The V extension does not support EEW=64 for index " |
2179 | "values when XLEN=32" ); |
2180 | } |
2181 | const RISCV::VLX_VSXPseudo *P = RISCV::getVLXPseudo( |
2182 | Masked: IsMasked, Ordered: IsOrdered, Log2SEW: IndexLog2EEW, LMUL: static_cast<unsigned>(LMUL), |
2183 | IndexLMUL: static_cast<unsigned>(IndexLMUL)); |
2184 | MachineSDNode *Load = |
2185 | CurDAG->getMachineNode(Opcode: P->Pseudo, dl: DL, VTs: Node->getVTList(), Ops: Operands); |
2186 | |
2187 | if (auto *MemOp = dyn_cast<MemSDNode>(Val: Node)) |
2188 | CurDAG->setNodeMemRefs(N: Load, NewMemRefs: {MemOp->getMemOperand()}); |
2189 | |
2190 | ReplaceNode(F: Node, T: Load); |
2191 | return; |
2192 | } |
2193 | case Intrinsic::riscv_vlm: |
2194 | case Intrinsic::riscv_vle: |
2195 | case Intrinsic::riscv_vle_mask: |
2196 | case Intrinsic::riscv_vlse: |
2197 | case Intrinsic::riscv_vlse_mask: { |
2198 | bool IsMasked = IntNo == Intrinsic::riscv_vle_mask || |
2199 | IntNo == Intrinsic::riscv_vlse_mask; |
2200 | bool IsStrided = |
2201 | IntNo == Intrinsic::riscv_vlse || IntNo == Intrinsic::riscv_vlse_mask; |
2202 | |
2203 | MVT VT = Node->getSimpleValueType(ResNo: 0); |
2204 | unsigned Log2SEW = Log2_32(Value: VT.getScalarSizeInBits()); |
2205 | |
2206 | // The riscv_vlm intrinsic are always tail agnostic and no passthru |
2207 | // operand at the IR level. In pseudos, they have both policy and |
2208 | // passthru operand. The passthru operand is needed to track the |
2209 | // "tail undefined" state, and the policy is there just for |
2210 | // for consistency - it will always be "don't care" for the |
2211 | // unmasked form. |
2212 | bool HasPassthruOperand = IntNo != Intrinsic::riscv_vlm; |
2213 | unsigned CurOp = 2; |
2214 | SmallVector<SDValue, 8> Operands; |
2215 | if (HasPassthruOperand) |
2216 | Operands.push_back(Elt: Node->getOperand(Num: CurOp++)); |
2217 | else { |
2218 | // We eagerly lower to implicit_def (instead of undef), as we |
2219 | // otherwise fail to select nodes such as: nxv1i1 = undef |
2220 | SDNode *Passthru = |
2221 | CurDAG->getMachineNode(Opcode: TargetOpcode::IMPLICIT_DEF, dl: DL, VT); |
2222 | Operands.push_back(Elt: SDValue(Passthru, 0)); |
2223 | } |
2224 | addVectorLoadStoreOperands(Node, Log2SEW, DL, CurOp, IsMasked, IsStridedOrIndexed: IsStrided, |
2225 | Operands, /*IsLoad=*/true); |
2226 | |
2227 | RISCVVType::VLMUL LMUL = RISCVTargetLowering::getLMUL(VT); |
2228 | const RISCV::VLEPseudo *P = |
2229 | RISCV::getVLEPseudo(Masked: IsMasked, Strided: IsStrided, /*FF*/ false, Log2SEW, |
2230 | LMUL: static_cast<unsigned>(LMUL)); |
2231 | MachineSDNode *Load = |
2232 | CurDAG->getMachineNode(Opcode: P->Pseudo, dl: DL, VTs: Node->getVTList(), Ops: Operands); |
2233 | |
2234 | if (auto *MemOp = dyn_cast<MemSDNode>(Val: Node)) |
2235 | CurDAG->setNodeMemRefs(N: Load, NewMemRefs: {MemOp->getMemOperand()}); |
2236 | |
2237 | ReplaceNode(F: Node, T: Load); |
2238 | return; |
2239 | } |
2240 | case Intrinsic::riscv_vleff: |
2241 | case Intrinsic::riscv_vleff_mask: { |
2242 | bool IsMasked = IntNo == Intrinsic::riscv_vleff_mask; |
2243 | |
2244 | MVT VT = Node->getSimpleValueType(ResNo: 0); |
2245 | unsigned Log2SEW = Log2_32(Value: VT.getScalarSizeInBits()); |
2246 | |
2247 | unsigned CurOp = 2; |
2248 | SmallVector<SDValue, 7> Operands; |
2249 | Operands.push_back(Elt: Node->getOperand(Num: CurOp++)); |
2250 | addVectorLoadStoreOperands(Node, Log2SEW, DL, CurOp, IsMasked, |
2251 | /*IsStridedOrIndexed*/ false, Operands, |
2252 | /*IsLoad=*/true); |
2253 | |
2254 | RISCVVType::VLMUL LMUL = RISCVTargetLowering::getLMUL(VT); |
2255 | const RISCV::VLEPseudo *P = |
2256 | RISCV::getVLEPseudo(Masked: IsMasked, /*Strided*/ false, /*FF*/ true, |
2257 | Log2SEW, LMUL: static_cast<unsigned>(LMUL)); |
2258 | MachineSDNode *Load = CurDAG->getMachineNode( |
2259 | Opcode: P->Pseudo, dl: DL, VTs: Node->getVTList(), Ops: Operands); |
2260 | if (auto *MemOp = dyn_cast<MemSDNode>(Val: Node)) |
2261 | CurDAG->setNodeMemRefs(N: Load, NewMemRefs: {MemOp->getMemOperand()}); |
2262 | |
2263 | ReplaceNode(F: Node, T: Load); |
2264 | return; |
2265 | } |
2266 | } |
2267 | break; |
2268 | } |
2269 | case ISD::INTRINSIC_VOID: { |
2270 | unsigned IntNo = Node->getConstantOperandVal(Num: 1); |
2271 | switch (IntNo) { |
2272 | case Intrinsic::riscv_vsseg2: |
2273 | case Intrinsic::riscv_vsseg3: |
2274 | case Intrinsic::riscv_vsseg4: |
2275 | case Intrinsic::riscv_vsseg5: |
2276 | case Intrinsic::riscv_vsseg6: |
2277 | case Intrinsic::riscv_vsseg7: |
2278 | case Intrinsic::riscv_vsseg8: { |
2279 | selectVSSEG(Node, NF: getSegInstNF(Intrinsic: IntNo), /*IsMasked*/ false, |
2280 | /*IsStrided*/ false); |
2281 | return; |
2282 | } |
2283 | case Intrinsic::riscv_vsseg2_mask: |
2284 | case Intrinsic::riscv_vsseg3_mask: |
2285 | case Intrinsic::riscv_vsseg4_mask: |
2286 | case Intrinsic::riscv_vsseg5_mask: |
2287 | case Intrinsic::riscv_vsseg6_mask: |
2288 | case Intrinsic::riscv_vsseg7_mask: |
2289 | case Intrinsic::riscv_vsseg8_mask: { |
2290 | selectVSSEG(Node, NF: getSegInstNF(Intrinsic: IntNo), /*IsMasked*/ true, |
2291 | /*IsStrided*/ false); |
2292 | return; |
2293 | } |
2294 | case Intrinsic::riscv_vssseg2: |
2295 | case Intrinsic::riscv_vssseg3: |
2296 | case Intrinsic::riscv_vssseg4: |
2297 | case Intrinsic::riscv_vssseg5: |
2298 | case Intrinsic::riscv_vssseg6: |
2299 | case Intrinsic::riscv_vssseg7: |
2300 | case Intrinsic::riscv_vssseg8: { |
2301 | selectVSSEG(Node, NF: getSegInstNF(Intrinsic: IntNo), /*IsMasked*/ false, |
2302 | /*IsStrided*/ true); |
2303 | return; |
2304 | } |
2305 | case Intrinsic::riscv_vssseg2_mask: |
2306 | case Intrinsic::riscv_vssseg3_mask: |
2307 | case Intrinsic::riscv_vssseg4_mask: |
2308 | case Intrinsic::riscv_vssseg5_mask: |
2309 | case Intrinsic::riscv_vssseg6_mask: |
2310 | case Intrinsic::riscv_vssseg7_mask: |
2311 | case Intrinsic::riscv_vssseg8_mask: { |
2312 | selectVSSEG(Node, NF: getSegInstNF(Intrinsic: IntNo), /*IsMasked*/ true, |
2313 | /*IsStrided*/ true); |
2314 | return; |
2315 | } |
2316 | case Intrinsic::riscv_vsoxseg2: |
2317 | case Intrinsic::riscv_vsoxseg3: |
2318 | case Intrinsic::riscv_vsoxseg4: |
2319 | case Intrinsic::riscv_vsoxseg5: |
2320 | case Intrinsic::riscv_vsoxseg6: |
2321 | case Intrinsic::riscv_vsoxseg7: |
2322 | case Intrinsic::riscv_vsoxseg8: |
2323 | selectVSXSEG(Node, NF: getSegInstNF(Intrinsic: IntNo), /*IsMasked*/ false, |
2324 | /*IsOrdered*/ true); |
2325 | return; |
2326 | case Intrinsic::riscv_vsuxseg2: |
2327 | case Intrinsic::riscv_vsuxseg3: |
2328 | case Intrinsic::riscv_vsuxseg4: |
2329 | case Intrinsic::riscv_vsuxseg5: |
2330 | case Intrinsic::riscv_vsuxseg6: |
2331 | case Intrinsic::riscv_vsuxseg7: |
2332 | case Intrinsic::riscv_vsuxseg8: |
2333 | selectVSXSEG(Node, NF: getSegInstNF(Intrinsic: IntNo), /*IsMasked*/ false, |
2334 | /*IsOrdered*/ false); |
2335 | return; |
2336 | case Intrinsic::riscv_vsoxseg2_mask: |
2337 | case Intrinsic::riscv_vsoxseg3_mask: |
2338 | case Intrinsic::riscv_vsoxseg4_mask: |
2339 | case Intrinsic::riscv_vsoxseg5_mask: |
2340 | case Intrinsic::riscv_vsoxseg6_mask: |
2341 | case Intrinsic::riscv_vsoxseg7_mask: |
2342 | case Intrinsic::riscv_vsoxseg8_mask: |
2343 | selectVSXSEG(Node, NF: getSegInstNF(Intrinsic: IntNo), /*IsMasked*/ true, |
2344 | /*IsOrdered*/ true); |
2345 | return; |
2346 | case Intrinsic::riscv_vsuxseg2_mask: |
2347 | case Intrinsic::riscv_vsuxseg3_mask: |
2348 | case Intrinsic::riscv_vsuxseg4_mask: |
2349 | case Intrinsic::riscv_vsuxseg5_mask: |
2350 | case Intrinsic::riscv_vsuxseg6_mask: |
2351 | case Intrinsic::riscv_vsuxseg7_mask: |
2352 | case Intrinsic::riscv_vsuxseg8_mask: |
2353 | selectVSXSEG(Node, NF: getSegInstNF(Intrinsic: IntNo), /*IsMasked*/ true, |
2354 | /*IsOrdered*/ false); |
2355 | return; |
2356 | case Intrinsic::riscv_vsoxei: |
2357 | case Intrinsic::riscv_vsoxei_mask: |
2358 | case Intrinsic::riscv_vsuxei: |
2359 | case Intrinsic::riscv_vsuxei_mask: { |
2360 | bool IsMasked = IntNo == Intrinsic::riscv_vsoxei_mask || |
2361 | IntNo == Intrinsic::riscv_vsuxei_mask; |
2362 | bool IsOrdered = IntNo == Intrinsic::riscv_vsoxei || |
2363 | IntNo == Intrinsic::riscv_vsoxei_mask; |
2364 | |
2365 | MVT VT = Node->getOperand(Num: 2)->getSimpleValueType(ResNo: 0); |
2366 | unsigned Log2SEW = Log2_32(Value: VT.getScalarSizeInBits()); |
2367 | |
2368 | unsigned CurOp = 2; |
2369 | SmallVector<SDValue, 8> Operands; |
2370 | Operands.push_back(Elt: Node->getOperand(Num: CurOp++)); // Store value. |
2371 | |
2372 | MVT IndexVT; |
2373 | addVectorLoadStoreOperands(Node, Log2SEW, DL, CurOp, IsMasked, |
2374 | /*IsStridedOrIndexed*/ true, Operands, |
2375 | /*IsLoad=*/false, IndexVT: &IndexVT); |
2376 | |
2377 | assert(VT.getVectorElementCount() == IndexVT.getVectorElementCount() && |
2378 | "Element count mismatch" ); |
2379 | |
2380 | RISCVVType::VLMUL LMUL = RISCVTargetLowering::getLMUL(VT); |
2381 | RISCVVType::VLMUL IndexLMUL = RISCVTargetLowering::getLMUL(VT: IndexVT); |
2382 | unsigned IndexLog2EEW = Log2_32(Value: IndexVT.getScalarSizeInBits()); |
2383 | if (IndexLog2EEW == 6 && !Subtarget->is64Bit()) { |
2384 | report_fatal_error(reason: "The V extension does not support EEW=64 for index " |
2385 | "values when XLEN=32" ); |
2386 | } |
2387 | const RISCV::VLX_VSXPseudo *P = RISCV::getVSXPseudo( |
2388 | Masked: IsMasked, Ordered: IsOrdered, Log2SEW: IndexLog2EEW, |
2389 | LMUL: static_cast<unsigned>(LMUL), IndexLMUL: static_cast<unsigned>(IndexLMUL)); |
2390 | MachineSDNode *Store = |
2391 | CurDAG->getMachineNode(Opcode: P->Pseudo, dl: DL, VTs: Node->getVTList(), Ops: Operands); |
2392 | |
2393 | if (auto *MemOp = dyn_cast<MemSDNode>(Val: Node)) |
2394 | CurDAG->setNodeMemRefs(N: Store, NewMemRefs: {MemOp->getMemOperand()}); |
2395 | |
2396 | ReplaceNode(F: Node, T: Store); |
2397 | return; |
2398 | } |
2399 | case Intrinsic::riscv_vsm: |
2400 | case Intrinsic::riscv_vse: |
2401 | case Intrinsic::riscv_vse_mask: |
2402 | case Intrinsic::riscv_vsse: |
2403 | case Intrinsic::riscv_vsse_mask: { |
2404 | bool IsMasked = IntNo == Intrinsic::riscv_vse_mask || |
2405 | IntNo == Intrinsic::riscv_vsse_mask; |
2406 | bool IsStrided = |
2407 | IntNo == Intrinsic::riscv_vsse || IntNo == Intrinsic::riscv_vsse_mask; |
2408 | |
2409 | MVT VT = Node->getOperand(Num: 2)->getSimpleValueType(ResNo: 0); |
2410 | unsigned Log2SEW = Log2_32(Value: VT.getScalarSizeInBits()); |
2411 | |
2412 | unsigned CurOp = 2; |
2413 | SmallVector<SDValue, 8> Operands; |
2414 | Operands.push_back(Elt: Node->getOperand(Num: CurOp++)); // Store value. |
2415 | |
2416 | addVectorLoadStoreOperands(Node, Log2SEW, DL, CurOp, IsMasked, IsStridedOrIndexed: IsStrided, |
2417 | Operands); |
2418 | |
2419 | RISCVVType::VLMUL LMUL = RISCVTargetLowering::getLMUL(VT); |
2420 | const RISCV::VSEPseudo *P = RISCV::getVSEPseudo( |
2421 | Masked: IsMasked, Strided: IsStrided, Log2SEW, LMUL: static_cast<unsigned>(LMUL)); |
2422 | MachineSDNode *Store = |
2423 | CurDAG->getMachineNode(Opcode: P->Pseudo, dl: DL, VTs: Node->getVTList(), Ops: Operands); |
2424 | if (auto *MemOp = dyn_cast<MemSDNode>(Val: Node)) |
2425 | CurDAG->setNodeMemRefs(N: Store, NewMemRefs: {MemOp->getMemOperand()}); |
2426 | |
2427 | ReplaceNode(F: Node, T: Store); |
2428 | return; |
2429 | } |
2430 | case Intrinsic::riscv_sf_vc_x_se: |
2431 | case Intrinsic::riscv_sf_vc_i_se: |
2432 | selectSF_VC_X_SE(Node); |
2433 | return; |
2434 | } |
2435 | break; |
2436 | } |
2437 | case ISD::BITCAST: { |
2438 | MVT SrcVT = Node->getOperand(Num: 0).getSimpleValueType(); |
2439 | // Just drop bitcasts between vectors if both are fixed or both are |
2440 | // scalable. |
2441 | if ((VT.isScalableVector() && SrcVT.isScalableVector()) || |
2442 | (VT.isFixedLengthVector() && SrcVT.isFixedLengthVector())) { |
2443 | ReplaceUses(F: SDValue(Node, 0), T: Node->getOperand(Num: 0)); |
2444 | CurDAG->RemoveDeadNode(N: Node); |
2445 | return; |
2446 | } |
2447 | break; |
2448 | } |
2449 | case ISD::INSERT_SUBVECTOR: |
2450 | case RISCVISD::TUPLE_INSERT: { |
2451 | SDValue V = Node->getOperand(Num: 0); |
2452 | SDValue SubV = Node->getOperand(Num: 1); |
2453 | SDLoc DL(SubV); |
2454 | auto Idx = Node->getConstantOperandVal(Num: 2); |
2455 | MVT SubVecVT = SubV.getSimpleValueType(); |
2456 | |
2457 | const RISCVTargetLowering &TLI = *Subtarget->getTargetLowering(); |
2458 | MVT SubVecContainerVT = SubVecVT; |
2459 | // Establish the correct scalable-vector types for any fixed-length type. |
2460 | if (SubVecVT.isFixedLengthVector()) { |
2461 | SubVecContainerVT = TLI.getContainerForFixedLengthVector(VT: SubVecVT); |
2462 | TypeSize VecRegSize = TypeSize::getScalable(MinimumSize: RISCV::RVVBitsPerBlock); |
2463 | [[maybe_unused]] bool ExactlyVecRegSized = |
2464 | Subtarget->expandVScale(X: SubVecVT.getSizeInBits()) |
2465 | .isKnownMultipleOf(RHS: Subtarget->expandVScale(X: VecRegSize)); |
2466 | assert(isPowerOf2_64(Subtarget->expandVScale(SubVecVT.getSizeInBits()) |
2467 | .getKnownMinValue())); |
2468 | assert(Idx == 0 && (ExactlyVecRegSized || V.isUndef())); |
2469 | } |
2470 | MVT ContainerVT = VT; |
2471 | if (VT.isFixedLengthVector()) |
2472 | ContainerVT = TLI.getContainerForFixedLengthVector(VT); |
2473 | |
2474 | const auto *TRI = Subtarget->getRegisterInfo(); |
2475 | unsigned SubRegIdx; |
2476 | std::tie(args&: SubRegIdx, args&: Idx) = |
2477 | RISCVTargetLowering::decomposeSubvectorInsertExtractToSubRegs( |
2478 | VecVT: ContainerVT, SubVecVT: SubVecContainerVT, InsertExtractIdx: Idx, TRI); |
2479 | |
2480 | // If the Idx hasn't been completely eliminated then this is a subvector |
2481 | // insert which doesn't naturally align to a vector register. These must |
2482 | // be handled using instructions to manipulate the vector registers. |
2483 | if (Idx != 0) |
2484 | break; |
2485 | |
2486 | RISCVVType::VLMUL SubVecLMUL = |
2487 | RISCVTargetLowering::getLMUL(VT: SubVecContainerVT); |
2488 | [[maybe_unused]] bool IsSubVecPartReg = |
2489 | SubVecLMUL == RISCVVType::VLMUL::LMUL_F2 || |
2490 | SubVecLMUL == RISCVVType::VLMUL::LMUL_F4 || |
2491 | SubVecLMUL == RISCVVType::VLMUL::LMUL_F8; |
2492 | assert((V.getValueType().isRISCVVectorTuple() || !IsSubVecPartReg || |
2493 | V.isUndef()) && |
2494 | "Expecting lowering to have created legal INSERT_SUBVECTORs when " |
2495 | "the subvector is smaller than a full-sized register" ); |
2496 | |
2497 | // If we haven't set a SubRegIdx, then we must be going between |
2498 | // equally-sized LMUL groups (e.g. VR -> VR). This can be done as a copy. |
2499 | if (SubRegIdx == RISCV::NoSubRegister) { |
2500 | unsigned InRegClassID = |
2501 | RISCVTargetLowering::getRegClassIDForVecVT(VT: ContainerVT); |
2502 | assert(RISCVTargetLowering::getRegClassIDForVecVT(SubVecContainerVT) == |
2503 | InRegClassID && |
2504 | "Unexpected subvector extraction" ); |
2505 | SDValue RC = CurDAG->getTargetConstant(Val: InRegClassID, DL, VT: XLenVT); |
2506 | SDNode *NewNode = CurDAG->getMachineNode(Opcode: TargetOpcode::COPY_TO_REGCLASS, |
2507 | dl: DL, VT, Op1: SubV, Op2: RC); |
2508 | ReplaceNode(F: Node, T: NewNode); |
2509 | return; |
2510 | } |
2511 | |
2512 | SDValue Insert = CurDAG->getTargetInsertSubreg(SRIdx: SubRegIdx, DL, VT, Operand: V, Subreg: SubV); |
2513 | ReplaceNode(F: Node, T: Insert.getNode()); |
2514 | return; |
2515 | } |
2516 | case ISD::EXTRACT_SUBVECTOR: |
2517 | case RISCVISD::TUPLE_EXTRACT: { |
2518 | SDValue V = Node->getOperand(Num: 0); |
2519 | auto Idx = Node->getConstantOperandVal(Num: 1); |
2520 | MVT InVT = V.getSimpleValueType(); |
2521 | SDLoc DL(V); |
2522 | |
2523 | const RISCVTargetLowering &TLI = *Subtarget->getTargetLowering(); |
2524 | MVT SubVecContainerVT = VT; |
2525 | // Establish the correct scalable-vector types for any fixed-length type. |
2526 | if (VT.isFixedLengthVector()) { |
2527 | assert(Idx == 0); |
2528 | SubVecContainerVT = TLI.getContainerForFixedLengthVector(VT); |
2529 | } |
2530 | if (InVT.isFixedLengthVector()) |
2531 | InVT = TLI.getContainerForFixedLengthVector(VT: InVT); |
2532 | |
2533 | const auto *TRI = Subtarget->getRegisterInfo(); |
2534 | unsigned SubRegIdx; |
2535 | std::tie(args&: SubRegIdx, args&: Idx) = |
2536 | RISCVTargetLowering::decomposeSubvectorInsertExtractToSubRegs( |
2537 | VecVT: InVT, SubVecVT: SubVecContainerVT, InsertExtractIdx: Idx, TRI); |
2538 | |
2539 | // If the Idx hasn't been completely eliminated then this is a subvector |
2540 | // extract which doesn't naturally align to a vector register. These must |
2541 | // be handled using instructions to manipulate the vector registers. |
2542 | if (Idx != 0) |
2543 | break; |
2544 | |
2545 | // If we haven't set a SubRegIdx, then we must be going between |
2546 | // equally-sized LMUL types (e.g. VR -> VR). This can be done as a copy. |
2547 | if (SubRegIdx == RISCV::NoSubRegister) { |
2548 | unsigned InRegClassID = RISCVTargetLowering::getRegClassIDForVecVT(VT: InVT); |
2549 | assert(RISCVTargetLowering::getRegClassIDForVecVT(SubVecContainerVT) == |
2550 | InRegClassID && |
2551 | "Unexpected subvector extraction" ); |
2552 | SDValue RC = CurDAG->getTargetConstant(Val: InRegClassID, DL, VT: XLenVT); |
2553 | SDNode *NewNode = |
2554 | CurDAG->getMachineNode(Opcode: TargetOpcode::COPY_TO_REGCLASS, dl: DL, VT, Op1: V, Op2: RC); |
2555 | ReplaceNode(F: Node, T: NewNode); |
2556 | return; |
2557 | } |
2558 | |
2559 | SDValue = CurDAG->getTargetExtractSubreg(SRIdx: SubRegIdx, DL, VT, Operand: V); |
2560 | ReplaceNode(F: Node, T: Extract.getNode()); |
2561 | return; |
2562 | } |
2563 | case RISCVISD::VMV_S_X_VL: |
2564 | case RISCVISD::VFMV_S_F_VL: |
2565 | case RISCVISD::VMV_V_X_VL: |
2566 | case RISCVISD::VFMV_V_F_VL: { |
2567 | // Try to match splat of a scalar load to a strided load with stride of x0. |
2568 | bool IsScalarMove = Node->getOpcode() == RISCVISD::VMV_S_X_VL || |
2569 | Node->getOpcode() == RISCVISD::VFMV_S_F_VL; |
2570 | if (!Node->getOperand(Num: 0).isUndef()) |
2571 | break; |
2572 | SDValue Src = Node->getOperand(Num: 1); |
2573 | auto *Ld = dyn_cast<LoadSDNode>(Val&: Src); |
2574 | // Can't fold load update node because the second |
2575 | // output is used so that load update node can't be removed. |
2576 | if (!Ld || Ld->isIndexed()) |
2577 | break; |
2578 | EVT MemVT = Ld->getMemoryVT(); |
2579 | // The memory VT should be the same size as the element type. |
2580 | if (MemVT.getStoreSize() != VT.getVectorElementType().getStoreSize()) |
2581 | break; |
2582 | if (!IsProfitableToFold(N: Src, U: Node, Root: Node) || |
2583 | !IsLegalToFold(N: Src, U: Node, Root: Node, OptLevel: TM.getOptLevel())) |
2584 | break; |
2585 | |
2586 | SDValue VL; |
2587 | if (IsScalarMove) { |
2588 | // We could deal with more VL if we update the VSETVLI insert pass to |
2589 | // avoid introducing more VSETVLI. |
2590 | if (!isOneConstant(V: Node->getOperand(Num: 2))) |
2591 | break; |
2592 | selectVLOp(N: Node->getOperand(Num: 2), VL); |
2593 | } else |
2594 | selectVLOp(N: Node->getOperand(Num: 2), VL); |
2595 | |
2596 | unsigned Log2SEW = Log2_32(Value: VT.getScalarSizeInBits()); |
2597 | SDValue SEW = CurDAG->getTargetConstant(Val: Log2SEW, DL, VT: XLenVT); |
2598 | |
2599 | // If VL=1, then we don't need to do a strided load and can just do a |
2600 | // regular load. |
2601 | bool IsStrided = !isOneConstant(V: VL); |
2602 | |
2603 | // Only do a strided load if we have optimized zero-stride vector load. |
2604 | if (IsStrided && !Subtarget->hasOptimizedZeroStrideLoad()) |
2605 | break; |
2606 | |
2607 | SmallVector<SDValue> Operands = { |
2608 | SDValue(CurDAG->getMachineNode(Opcode: TargetOpcode::IMPLICIT_DEF, dl: DL, VT), 0), |
2609 | Ld->getBasePtr()}; |
2610 | if (IsStrided) |
2611 | Operands.push_back(Elt: CurDAG->getRegister(Reg: RISCV::X0, VT: XLenVT)); |
2612 | uint64_t Policy = RISCVVType::MASK_AGNOSTIC | RISCVVType::TAIL_AGNOSTIC; |
2613 | SDValue PolicyOp = CurDAG->getTargetConstant(Val: Policy, DL, VT: XLenVT); |
2614 | Operands.append(IL: {VL, SEW, PolicyOp, Ld->getChain()}); |
2615 | |
2616 | RISCVVType::VLMUL LMUL = RISCVTargetLowering::getLMUL(VT); |
2617 | const RISCV::VLEPseudo *P = RISCV::getVLEPseudo( |
2618 | /*IsMasked*/ Masked: false, Strided: IsStrided, /*FF*/ false, |
2619 | Log2SEW, LMUL: static_cast<unsigned>(LMUL)); |
2620 | MachineSDNode *Load = |
2621 | CurDAG->getMachineNode(Opcode: P->Pseudo, dl: DL, ResultTys: {VT, MVT::Other}, Ops: Operands); |
2622 | // Update the chain. |
2623 | ReplaceUses(F: Src.getValue(R: 1), T: SDValue(Load, 1)); |
2624 | // Record the mem-refs |
2625 | CurDAG->setNodeMemRefs(N: Load, NewMemRefs: {Ld->getMemOperand()}); |
2626 | // Replace the splat with the vlse. |
2627 | ReplaceNode(F: Node, T: Load); |
2628 | return; |
2629 | } |
2630 | case ISD::PREFETCH: |
2631 | unsigned Locality = Node->getConstantOperandVal(Num: 3); |
2632 | if (Locality > 2) |
2633 | break; |
2634 | |
2635 | if (auto *LoadStoreMem = dyn_cast<MemSDNode>(Val: Node)) { |
2636 | MachineMemOperand *MMO = LoadStoreMem->getMemOperand(); |
2637 | MMO->setFlags(MachineMemOperand::MONonTemporal); |
2638 | |
2639 | int NontemporalLevel = 0; |
2640 | switch (Locality) { |
2641 | case 0: |
2642 | NontemporalLevel = 3; // NTL.ALL |
2643 | break; |
2644 | case 1: |
2645 | NontemporalLevel = 1; // NTL.PALL |
2646 | break; |
2647 | case 2: |
2648 | NontemporalLevel = 0; // NTL.P1 |
2649 | break; |
2650 | default: |
2651 | llvm_unreachable("unexpected locality value." ); |
2652 | } |
2653 | |
2654 | if (NontemporalLevel & 0b1) |
2655 | MMO->setFlags(MONontemporalBit0); |
2656 | if (NontemporalLevel & 0b10) |
2657 | MMO->setFlags(MONontemporalBit1); |
2658 | } |
2659 | break; |
2660 | } |
2661 | |
2662 | // Select the default instruction. |
2663 | SelectCode(N: Node); |
2664 | } |
2665 | |
2666 | bool RISCVDAGToDAGISel::SelectInlineAsmMemoryOperand( |
2667 | const SDValue &Op, InlineAsm::ConstraintCode ConstraintID, |
2668 | std::vector<SDValue> &OutOps) { |
2669 | // Always produce a register and immediate operand, as expected by |
2670 | // RISCVAsmPrinter::PrintAsmMemoryOperand. |
2671 | switch (ConstraintID) { |
2672 | case InlineAsm::ConstraintCode::o: |
2673 | case InlineAsm::ConstraintCode::m: { |
2674 | SDValue Op0, Op1; |
2675 | [[maybe_unused]] bool Found = SelectAddrRegImm(Addr: Op, Base&: Op0, Offset&: Op1); |
2676 | assert(Found && "SelectAddrRegImm should always succeed" ); |
2677 | OutOps.push_back(x: Op0); |
2678 | OutOps.push_back(x: Op1); |
2679 | return false; |
2680 | } |
2681 | case InlineAsm::ConstraintCode::A: |
2682 | OutOps.push_back(x: Op); |
2683 | OutOps.push_back( |
2684 | x: CurDAG->getTargetConstant(Val: 0, DL: SDLoc(Op), VT: Subtarget->getXLenVT())); |
2685 | return false; |
2686 | default: |
2687 | report_fatal_error(reason: "Unexpected asm memory constraint " + |
2688 | InlineAsm::getMemConstraintName(C: ConstraintID)); |
2689 | } |
2690 | |
2691 | return true; |
2692 | } |
2693 | |
2694 | bool RISCVDAGToDAGISel::SelectAddrFrameIndex(SDValue Addr, SDValue &Base, |
2695 | SDValue &Offset) { |
2696 | if (auto *FIN = dyn_cast<FrameIndexSDNode>(Val&: Addr)) { |
2697 | Base = CurDAG->getTargetFrameIndex(FI: FIN->getIndex(), VT: Subtarget->getXLenVT()); |
2698 | Offset = CurDAG->getTargetConstant(Val: 0, DL: SDLoc(Addr), VT: Subtarget->getXLenVT()); |
2699 | return true; |
2700 | } |
2701 | |
2702 | return false; |
2703 | } |
2704 | |
2705 | // Fold constant addresses. |
2706 | static bool selectConstantAddr(SelectionDAG *CurDAG, const SDLoc &DL, |
2707 | const MVT VT, const RISCVSubtarget *Subtarget, |
2708 | SDValue Addr, SDValue &Base, SDValue &Offset, |
2709 | bool IsPrefetch = false) { |
2710 | if (!isa<ConstantSDNode>(Val: Addr)) |
2711 | return false; |
2712 | |
2713 | int64_t CVal = cast<ConstantSDNode>(Val&: Addr)->getSExtValue(); |
2714 | |
2715 | // If the constant is a simm12, we can fold the whole constant and use X0 as |
2716 | // the base. If the constant can be materialized with LUI+simm12, use LUI as |
2717 | // the base. We can't use generateInstSeq because it favors LUI+ADDIW. |
2718 | int64_t Lo12 = SignExtend64<12>(x: CVal); |
2719 | int64_t Hi = (uint64_t)CVal - (uint64_t)Lo12; |
2720 | if (!Subtarget->is64Bit() || isInt<32>(x: Hi)) { |
2721 | if (IsPrefetch && (Lo12 & 0b11111) != 0) |
2722 | return false; |
2723 | if (Hi) { |
2724 | int64_t Hi20 = (Hi >> 12) & 0xfffff; |
2725 | Base = SDValue( |
2726 | CurDAG->getMachineNode(Opcode: RISCV::LUI, dl: DL, VT, |
2727 | Op1: CurDAG->getTargetConstant(Val: Hi20, DL, VT)), |
2728 | 0); |
2729 | } else { |
2730 | Base = CurDAG->getRegister(Reg: RISCV::X0, VT); |
2731 | } |
2732 | Offset = CurDAG->getSignedTargetConstant(Val: Lo12, DL, VT); |
2733 | return true; |
2734 | } |
2735 | |
2736 | // Ask how constant materialization would handle this constant. |
2737 | RISCVMatInt::InstSeq Seq = RISCVMatInt::generateInstSeq(Val: CVal, STI: *Subtarget); |
2738 | |
2739 | // If the last instruction would be an ADDI, we can fold its immediate and |
2740 | // emit the rest of the sequence as the base. |
2741 | if (Seq.back().getOpcode() != RISCV::ADDI) |
2742 | return false; |
2743 | Lo12 = Seq.back().getImm(); |
2744 | if (IsPrefetch && (Lo12 & 0b11111) != 0) |
2745 | return false; |
2746 | |
2747 | // Drop the last instruction. |
2748 | Seq.pop_back(); |
2749 | assert(!Seq.empty() && "Expected more instructions in sequence" ); |
2750 | |
2751 | Base = selectImmSeq(CurDAG, DL, VT, Seq); |
2752 | Offset = CurDAG->getSignedTargetConstant(Val: Lo12, DL, VT); |
2753 | return true; |
2754 | } |
2755 | |
2756 | // Is this ADD instruction only used as the base pointer of scalar loads and |
2757 | // stores? |
2758 | static bool isWorthFoldingAdd(SDValue Add) { |
2759 | for (auto *User : Add->users()) { |
2760 | if (User->getOpcode() != ISD::LOAD && User->getOpcode() != ISD::STORE && |
2761 | User->getOpcode() != ISD::ATOMIC_LOAD && |
2762 | User->getOpcode() != ISD::ATOMIC_STORE) |
2763 | return false; |
2764 | EVT VT = cast<MemSDNode>(Val: User)->getMemoryVT(); |
2765 | if (!VT.isScalarInteger() && VT != MVT::f16 && VT != MVT::f32 && |
2766 | VT != MVT::f64) |
2767 | return false; |
2768 | // Don't allow stores of the value. It must be used as the address. |
2769 | if (User->getOpcode() == ISD::STORE && |
2770 | cast<StoreSDNode>(Val: User)->getValue() == Add) |
2771 | return false; |
2772 | if (User->getOpcode() == ISD::ATOMIC_STORE && |
2773 | cast<AtomicSDNode>(Val: User)->getVal() == Add) |
2774 | return false; |
2775 | if (isStrongerThanMonotonic(AO: cast<MemSDNode>(Val: User)->getSuccessOrdering())) |
2776 | return false; |
2777 | } |
2778 | |
2779 | return true; |
2780 | } |
2781 | |
2782 | bool RISCVDAGToDAGISel::SelectAddrRegRegScale(SDValue Addr, |
2783 | unsigned MaxShiftAmount, |
2784 | SDValue &Base, SDValue &Index, |
2785 | SDValue &Scale) { |
2786 | EVT VT = Addr.getSimpleValueType(); |
2787 | auto UnwrapShl = [this, VT, MaxShiftAmount](SDValue N, SDValue &Index, |
2788 | SDValue &Shift) { |
2789 | uint64_t ShiftAmt = 0; |
2790 | Index = N; |
2791 | |
2792 | if (N.getOpcode() == ISD::SHL && isa<ConstantSDNode>(Val: N.getOperand(i: 1))) { |
2793 | // Only match shifts by a value in range [0, MaxShiftAmount]. |
2794 | if (N.getConstantOperandVal(i: 1) <= MaxShiftAmount) { |
2795 | Index = N.getOperand(i: 0); |
2796 | ShiftAmt = N.getConstantOperandVal(i: 1); |
2797 | } |
2798 | } |
2799 | |
2800 | Shift = CurDAG->getTargetConstant(Val: ShiftAmt, DL: SDLoc(N), VT); |
2801 | return ShiftAmt != 0; |
2802 | }; |
2803 | |
2804 | if (Addr.getOpcode() == ISD::ADD) { |
2805 | if (auto *C1 = dyn_cast<ConstantSDNode>(Val: Addr.getOperand(i: 1))) { |
2806 | SDValue AddrB = Addr.getOperand(i: 0); |
2807 | if (AddrB.getOpcode() == ISD::ADD && |
2808 | UnwrapShl(AddrB.getOperand(i: 0), Index, Scale) && |
2809 | !isa<ConstantSDNode>(Val: AddrB.getOperand(i: 1)) && |
2810 | isInt<12>(x: C1->getSExtValue())) { |
2811 | // (add (add (shl A C2) B) C1) -> (add (add B C1) (shl A C2)) |
2812 | SDValue C1Val = |
2813 | CurDAG->getTargetConstant(Val: C1->getZExtValue(), DL: SDLoc(Addr), VT); |
2814 | Base = SDValue(CurDAG->getMachineNode(Opcode: RISCV::ADDI, dl: SDLoc(Addr), VT, |
2815 | Op1: AddrB.getOperand(i: 1), Op2: C1Val), |
2816 | 0); |
2817 | return true; |
2818 | } |
2819 | } else if (UnwrapShl(Addr.getOperand(i: 0), Index, Scale)) { |
2820 | Base = Addr.getOperand(i: 1); |
2821 | return true; |
2822 | } else { |
2823 | UnwrapShl(Addr.getOperand(i: 1), Index, Scale); |
2824 | Base = Addr.getOperand(i: 0); |
2825 | return true; |
2826 | } |
2827 | } |
2828 | |
2829 | return false; |
2830 | } |
2831 | |
2832 | bool RISCVDAGToDAGISel::SelectAddrRegImm(SDValue Addr, SDValue &Base, |
2833 | SDValue &Offset) { |
2834 | if (SelectAddrFrameIndex(Addr, Base, Offset)) |
2835 | return true; |
2836 | |
2837 | SDLoc DL(Addr); |
2838 | MVT VT = Addr.getSimpleValueType(); |
2839 | |
2840 | if (Addr.getOpcode() == RISCVISD::ADD_LO) { |
2841 | Base = Addr.getOperand(i: 0); |
2842 | Offset = Addr.getOperand(i: 1); |
2843 | return true; |
2844 | } |
2845 | |
2846 | if (CurDAG->isBaseWithConstantOffset(Op: Addr)) { |
2847 | int64_t CVal = cast<ConstantSDNode>(Val: Addr.getOperand(i: 1))->getSExtValue(); |
2848 | if (isInt<12>(x: CVal) && isInt<12>(x: CVal)) { |
2849 | Base = Addr.getOperand(i: 0); |
2850 | if (Base.getOpcode() == RISCVISD::ADD_LO) { |
2851 | SDValue LoOperand = Base.getOperand(i: 1); |
2852 | if (auto *GA = dyn_cast<GlobalAddressSDNode>(Val&: LoOperand)) { |
2853 | // If the Lo in (ADD_LO hi, lo) is a global variable's address |
2854 | // (its low part, really), then we can rely on the alignment of that |
2855 | // variable to provide a margin of safety before low part can overflow |
2856 | // the 12 bits of the load/store offset. Check if CVal falls within |
2857 | // that margin; if so (low part + CVal) can't overflow. |
2858 | const DataLayout &DL = CurDAG->getDataLayout(); |
2859 | Align Alignment = commonAlignment( |
2860 | A: GA->getGlobal()->getPointerAlignment(DL), Offset: GA->getOffset()); |
2861 | if ((CVal == 0 || Alignment > CVal)) { |
2862 | int64_t CombinedOffset = CVal + GA->getOffset(); |
2863 | Base = Base.getOperand(i: 0); |
2864 | Offset = CurDAG->getTargetGlobalAddress( |
2865 | GV: GA->getGlobal(), DL: SDLoc(LoOperand), VT: LoOperand.getValueType(), |
2866 | offset: CombinedOffset, TargetFlags: GA->getTargetFlags()); |
2867 | return true; |
2868 | } |
2869 | } |
2870 | } |
2871 | |
2872 | if (auto *FIN = dyn_cast<FrameIndexSDNode>(Val&: Base)) |
2873 | Base = CurDAG->getTargetFrameIndex(FI: FIN->getIndex(), VT); |
2874 | Offset = CurDAG->getSignedTargetConstant(Val: CVal, DL, VT); |
2875 | return true; |
2876 | } |
2877 | } |
2878 | |
2879 | // Handle ADD with large immediates. |
2880 | if (Addr.getOpcode() == ISD::ADD && isa<ConstantSDNode>(Val: Addr.getOperand(i: 1))) { |
2881 | int64_t CVal = cast<ConstantSDNode>(Val: Addr.getOperand(i: 1))->getSExtValue(); |
2882 | assert(!(isInt<12>(CVal) && isInt<12>(CVal)) && |
2883 | "simm12 not already handled?" ); |
2884 | |
2885 | // Handle immediates in the range [-4096,-2049] or [2048, 4094]. We can use |
2886 | // an ADDI for part of the offset and fold the rest into the load/store. |
2887 | // This mirrors the AddiPair PatFrag in RISCVInstrInfo.td. |
2888 | if (CVal >= -4096 && CVal <= 4094) { |
2889 | int64_t Adj = CVal < 0 ? -2048 : 2047; |
2890 | Base = SDValue( |
2891 | CurDAG->getMachineNode(Opcode: RISCV::ADDI, dl: DL, VT, Op1: Addr.getOperand(i: 0), |
2892 | Op2: CurDAG->getSignedTargetConstant(Val: Adj, DL, VT)), |
2893 | 0); |
2894 | Offset = CurDAG->getSignedTargetConstant(Val: CVal - Adj, DL, VT); |
2895 | return true; |
2896 | } |
2897 | |
2898 | // For larger immediates, we might be able to save one instruction from |
2899 | // constant materialization by folding the Lo12 bits of the immediate into |
2900 | // the address. We should only do this if the ADD is only used by loads and |
2901 | // stores that can fold the lo12 bits. Otherwise, the ADD will get iseled |
2902 | // separately with the full materialized immediate creating extra |
2903 | // instructions. |
2904 | if (isWorthFoldingAdd(Add: Addr) && |
2905 | selectConstantAddr(CurDAG, DL, VT, Subtarget, Addr: Addr.getOperand(i: 1), Base, |
2906 | Offset, /*IsPrefetch=*/false)) { |
2907 | // Insert an ADD instruction with the materialized Hi52 bits. |
2908 | Base = SDValue( |
2909 | CurDAG->getMachineNode(Opcode: RISCV::ADD, dl: DL, VT, Op1: Addr.getOperand(i: 0), Op2: Base), |
2910 | 0); |
2911 | return true; |
2912 | } |
2913 | } |
2914 | |
2915 | if (selectConstantAddr(CurDAG, DL, VT, Subtarget, Addr, Base, Offset, |
2916 | /*IsPrefetch=*/false)) |
2917 | return true; |
2918 | |
2919 | Base = Addr; |
2920 | Offset = CurDAG->getTargetConstant(Val: 0, DL, VT); |
2921 | return true; |
2922 | } |
2923 | |
2924 | /// Similar to SelectAddrRegImm, except that the offset restricted for |
2925 | /// unsinged nine bits. |
2926 | bool RISCVDAGToDAGISel::SelectAddrRegImm9(SDValue Addr, SDValue &Base, |
2927 | SDValue &Offset) { |
2928 | if (SelectAddrFrameIndex(Addr, Base, Offset)) |
2929 | return true; |
2930 | |
2931 | SDLoc DL(Addr); |
2932 | MVT VT = Addr.getSimpleValueType(); |
2933 | |
2934 | if (CurDAG->isBaseWithConstantOffset(Op: Addr)) { |
2935 | int64_t CVal = cast<ConstantSDNode>(Val: Addr.getOperand(i: 1))->getSExtValue(); |
2936 | if (isUInt<9>(x: CVal)) { |
2937 | Base = Addr.getOperand(i: 0); |
2938 | |
2939 | if (auto *FIN = dyn_cast<FrameIndexSDNode>(Val&: Base)) |
2940 | Base = CurDAG->getTargetFrameIndex(FI: FIN->getIndex(), VT); |
2941 | Offset = CurDAG->getSignedTargetConstant(Val: CVal, DL, VT); |
2942 | return true; |
2943 | } |
2944 | } |
2945 | |
2946 | Base = Addr; |
2947 | Offset = CurDAG->getTargetConstant(Val: 0, DL, VT); |
2948 | return true; |
2949 | } |
2950 | |
2951 | /// Similar to SelectAddrRegImm, except that the least significant 5 bits of |
2952 | /// Offset should be all zeros. |
2953 | bool RISCVDAGToDAGISel::SelectAddrRegImmLsb00000(SDValue Addr, SDValue &Base, |
2954 | SDValue &Offset) { |
2955 | if (SelectAddrFrameIndex(Addr, Base, Offset)) |
2956 | return true; |
2957 | |
2958 | SDLoc DL(Addr); |
2959 | MVT VT = Addr.getSimpleValueType(); |
2960 | |
2961 | if (CurDAG->isBaseWithConstantOffset(Op: Addr)) { |
2962 | int64_t CVal = cast<ConstantSDNode>(Val: Addr.getOperand(i: 1))->getSExtValue(); |
2963 | if (isInt<12>(x: CVal)) { |
2964 | Base = Addr.getOperand(i: 0); |
2965 | |
2966 | // Early-out if not a valid offset. |
2967 | if ((CVal & 0b11111) != 0) { |
2968 | Base = Addr; |
2969 | Offset = CurDAG->getTargetConstant(Val: 0, DL, VT); |
2970 | return true; |
2971 | } |
2972 | |
2973 | if (auto *FIN = dyn_cast<FrameIndexSDNode>(Val&: Base)) |
2974 | Base = CurDAG->getTargetFrameIndex(FI: FIN->getIndex(), VT); |
2975 | Offset = CurDAG->getSignedTargetConstant(Val: CVal, DL, VT); |
2976 | return true; |
2977 | } |
2978 | } |
2979 | |
2980 | // Handle ADD with large immediates. |
2981 | if (Addr.getOpcode() == ISD::ADD && isa<ConstantSDNode>(Val: Addr.getOperand(i: 1))) { |
2982 | int64_t CVal = cast<ConstantSDNode>(Val: Addr.getOperand(i: 1))->getSExtValue(); |
2983 | assert(!isInt<12>(CVal) && "simm12 not already handled?" ); |
2984 | |
2985 | // Handle immediates in the range [-4096,-2049] or [2017, 4065]. We can save |
2986 | // one instruction by folding adjustment (-2048 or 2016) into the address. |
2987 | if ((-2049 >= CVal && CVal >= -4096) || (4065 >= CVal && CVal >= 2017)) { |
2988 | int64_t Adj = CVal < 0 ? -2048 : 2016; |
2989 | int64_t AdjustedOffset = CVal - Adj; |
2990 | Base = |
2991 | SDValue(CurDAG->getMachineNode( |
2992 | Opcode: RISCV::ADDI, dl: DL, VT, Op1: Addr.getOperand(i: 0), |
2993 | Op2: CurDAG->getSignedTargetConstant(Val: AdjustedOffset, DL, VT)), |
2994 | 0); |
2995 | Offset = CurDAG->getSignedTargetConstant(Val: Adj, DL, VT); |
2996 | return true; |
2997 | } |
2998 | |
2999 | if (selectConstantAddr(CurDAG, DL, VT, Subtarget, Addr: Addr.getOperand(i: 1), Base, |
3000 | Offset, /*IsPrefetch=*/true)) { |
3001 | // Insert an ADD instruction with the materialized Hi52 bits. |
3002 | Base = SDValue( |
3003 | CurDAG->getMachineNode(Opcode: RISCV::ADD, dl: DL, VT, Op1: Addr.getOperand(i: 0), Op2: Base), |
3004 | 0); |
3005 | return true; |
3006 | } |
3007 | } |
3008 | |
3009 | if (selectConstantAddr(CurDAG, DL, VT, Subtarget, Addr, Base, Offset, |
3010 | /*IsPrefetch=*/true)) |
3011 | return true; |
3012 | |
3013 | Base = Addr; |
3014 | Offset = CurDAG->getTargetConstant(Val: 0, DL, VT); |
3015 | return true; |
3016 | } |
3017 | |
3018 | bool RISCVDAGToDAGISel::SelectAddrRegReg(SDValue Addr, SDValue &Base, |
3019 | SDValue &Offset) { |
3020 | if (Addr.getOpcode() != ISD::ADD) |
3021 | return false; |
3022 | |
3023 | if (isa<ConstantSDNode>(Val: Addr.getOperand(i: 1))) |
3024 | return false; |
3025 | |
3026 | Base = Addr.getOperand(i: 0); |
3027 | Offset = Addr.getOperand(i: 1); |
3028 | return true; |
3029 | } |
3030 | |
3031 | bool RISCVDAGToDAGISel::selectShiftMask(SDValue N, unsigned ShiftWidth, |
3032 | SDValue &ShAmt) { |
3033 | ShAmt = N; |
3034 | |
3035 | // Peek through zext. |
3036 | if (ShAmt->getOpcode() == ISD::ZERO_EXTEND) |
3037 | ShAmt = ShAmt.getOperand(i: 0); |
3038 | |
3039 | // Shift instructions on RISC-V only read the lower 5 or 6 bits of the shift |
3040 | // amount. If there is an AND on the shift amount, we can bypass it if it |
3041 | // doesn't affect any of those bits. |
3042 | if (ShAmt.getOpcode() == ISD::AND && |
3043 | isa<ConstantSDNode>(Val: ShAmt.getOperand(i: 1))) { |
3044 | const APInt &AndMask = ShAmt.getConstantOperandAPInt(i: 1); |
3045 | |
3046 | // Since the max shift amount is a power of 2 we can subtract 1 to make a |
3047 | // mask that covers the bits needed to represent all shift amounts. |
3048 | assert(isPowerOf2_32(ShiftWidth) && "Unexpected max shift amount!" ); |
3049 | APInt ShMask(AndMask.getBitWidth(), ShiftWidth - 1); |
3050 | |
3051 | if (ShMask.isSubsetOf(RHS: AndMask)) { |
3052 | ShAmt = ShAmt.getOperand(i: 0); |
3053 | } else { |
3054 | // SimplifyDemandedBits may have optimized the mask so try restoring any |
3055 | // bits that are known zero. |
3056 | KnownBits Known = CurDAG->computeKnownBits(Op: ShAmt.getOperand(i: 0)); |
3057 | if (!ShMask.isSubsetOf(RHS: AndMask | Known.Zero)) |
3058 | return true; |
3059 | ShAmt = ShAmt.getOperand(i: 0); |
3060 | } |
3061 | } |
3062 | |
3063 | if (ShAmt.getOpcode() == ISD::ADD && |
3064 | isa<ConstantSDNode>(Val: ShAmt.getOperand(i: 1))) { |
3065 | uint64_t Imm = ShAmt.getConstantOperandVal(i: 1); |
3066 | // If we are shifting by X+N where N == 0 mod Size, then just shift by X |
3067 | // to avoid the ADD. |
3068 | if (Imm != 0 && Imm % ShiftWidth == 0) { |
3069 | ShAmt = ShAmt.getOperand(i: 0); |
3070 | return true; |
3071 | } |
3072 | } else if (ShAmt.getOpcode() == ISD::SUB && |
3073 | isa<ConstantSDNode>(Val: ShAmt.getOperand(i: 0))) { |
3074 | uint64_t Imm = ShAmt.getConstantOperandVal(i: 0); |
3075 | // If we are shifting by N-X where N == 0 mod Size, then just shift by -X to |
3076 | // generate a NEG instead of a SUB of a constant. |
3077 | if (Imm != 0 && Imm % ShiftWidth == 0) { |
3078 | SDLoc DL(ShAmt); |
3079 | EVT VT = ShAmt.getValueType(); |
3080 | SDValue Zero = CurDAG->getRegister(Reg: RISCV::X0, VT); |
3081 | unsigned NegOpc = VT == MVT::i64 ? RISCV::SUBW : RISCV::SUB; |
3082 | MachineSDNode *Neg = CurDAG->getMachineNode(Opcode: NegOpc, dl: DL, VT, Op1: Zero, |
3083 | Op2: ShAmt.getOperand(i: 1)); |
3084 | ShAmt = SDValue(Neg, 0); |
3085 | return true; |
3086 | } |
3087 | // If we are shifting by N-X where N == -1 mod Size, then just shift by ~X |
3088 | // to generate a NOT instead of a SUB of a constant. |
3089 | if (Imm % ShiftWidth == ShiftWidth - 1) { |
3090 | SDLoc DL(ShAmt); |
3091 | EVT VT = ShAmt.getValueType(); |
3092 | MachineSDNode *Not = CurDAG->getMachineNode( |
3093 | Opcode: RISCV::XORI, dl: DL, VT, Op1: ShAmt.getOperand(i: 1), |
3094 | Op2: CurDAG->getAllOnesConstant(DL, VT, /*isTarget=*/IsTarget: true)); |
3095 | ShAmt = SDValue(Not, 0); |
3096 | return true; |
3097 | } |
3098 | } |
3099 | |
3100 | return true; |
3101 | } |
3102 | |
3103 | /// RISC-V doesn't have general instructions for integer setne/seteq, but we can |
3104 | /// check for equality with 0. This function emits instructions that convert the |
3105 | /// seteq/setne into something that can be compared with 0. |
3106 | /// \p ExpectedCCVal indicates the condition code to attempt to match (e.g. |
3107 | /// ISD::SETNE). |
3108 | bool RISCVDAGToDAGISel::selectSETCC(SDValue N, ISD::CondCode ExpectedCCVal, |
3109 | SDValue &Val) { |
3110 | assert(ISD::isIntEqualitySetCC(ExpectedCCVal) && |
3111 | "Unexpected condition code!" ); |
3112 | |
3113 | // We're looking for a setcc. |
3114 | if (N->getOpcode() != ISD::SETCC) |
3115 | return false; |
3116 | |
3117 | // Must be an equality comparison. |
3118 | ISD::CondCode CCVal = cast<CondCodeSDNode>(Val: N->getOperand(Num: 2))->get(); |
3119 | if (CCVal != ExpectedCCVal) |
3120 | return false; |
3121 | |
3122 | SDValue LHS = N->getOperand(Num: 0); |
3123 | SDValue RHS = N->getOperand(Num: 1); |
3124 | |
3125 | if (!LHS.getValueType().isScalarInteger()) |
3126 | return false; |
3127 | |
3128 | // If the RHS side is 0, we don't need any extra instructions, return the LHS. |
3129 | if (isNullConstant(V: RHS)) { |
3130 | Val = LHS; |
3131 | return true; |
3132 | } |
3133 | |
3134 | SDLoc DL(N); |
3135 | |
3136 | if (auto *C = dyn_cast<ConstantSDNode>(Val&: RHS)) { |
3137 | int64_t CVal = C->getSExtValue(); |
3138 | // If the RHS is -2048, we can use xori to produce 0 if the LHS is -2048 and |
3139 | // non-zero otherwise. |
3140 | if (CVal == -2048) { |
3141 | Val = SDValue( |
3142 | CurDAG->getMachineNode( |
3143 | Opcode: RISCV::XORI, dl: DL, VT: N->getValueType(ResNo: 0), Op1: LHS, |
3144 | Op2: CurDAG->getSignedTargetConstant(Val: CVal, DL, VT: N->getValueType(ResNo: 0))), |
3145 | 0); |
3146 | return true; |
3147 | } |
3148 | // If the RHS is [-2047,2048], we can use addi with -RHS to produce 0 if the |
3149 | // LHS is equal to the RHS and non-zero otherwise. |
3150 | if (isInt<12>(x: CVal) || CVal == 2048) { |
3151 | Val = SDValue( |
3152 | CurDAG->getMachineNode( |
3153 | Opcode: RISCV::ADDI, dl: DL, VT: N->getValueType(ResNo: 0), Op1: LHS, |
3154 | Op2: CurDAG->getSignedTargetConstant(Val: -CVal, DL, VT: N->getValueType(ResNo: 0))), |
3155 | 0); |
3156 | return true; |
3157 | } |
3158 | if (isPowerOf2_64(Value: CVal) && Subtarget->hasStdExtZbs()) { |
3159 | Val = SDValue( |
3160 | CurDAG->getMachineNode( |
3161 | Opcode: RISCV::BINVI, dl: DL, VT: N->getValueType(ResNo: 0), Op1: LHS, |
3162 | Op2: CurDAG->getTargetConstant(Val: Log2_64(Value: CVal), DL, VT: N->getValueType(ResNo: 0))), |
3163 | 0); |
3164 | return true; |
3165 | } |
3166 | // Same as the addi case above but for larger immediates (signed 26-bit) use |
3167 | // the QC_E_ADDI instruction from the Xqcilia extension, if available. Avoid |
3168 | // anything which can be done with a single lui as it might be compressible. |
3169 | if (Subtarget->hasVendorXqcilia() && isInt<26>(x: CVal) && |
3170 | (CVal & 0xFFF) != 0) { |
3171 | Val = SDValue( |
3172 | CurDAG->getMachineNode( |
3173 | Opcode: RISCV::QC_E_ADDI, dl: DL, VT: N->getValueType(ResNo: 0), Op1: LHS, |
3174 | Op2: CurDAG->getSignedTargetConstant(Val: -CVal, DL, VT: N->getValueType(ResNo: 0))), |
3175 | 0); |
3176 | return true; |
3177 | } |
3178 | } |
3179 | |
3180 | // If nothing else we can XOR the LHS and RHS to produce zero if they are |
3181 | // equal and a non-zero value if they aren't. |
3182 | Val = SDValue( |
3183 | CurDAG->getMachineNode(Opcode: RISCV::XOR, dl: DL, VT: N->getValueType(ResNo: 0), Op1: LHS, Op2: RHS), 0); |
3184 | return true; |
3185 | } |
3186 | |
3187 | bool RISCVDAGToDAGISel::selectSExtBits(SDValue N, unsigned Bits, SDValue &Val) { |
3188 | if (N.getOpcode() == ISD::SIGN_EXTEND_INREG && |
3189 | cast<VTSDNode>(Val: N.getOperand(i: 1))->getVT().getSizeInBits() == Bits) { |
3190 | Val = N.getOperand(i: 0); |
3191 | return true; |
3192 | } |
3193 | |
3194 | auto UnwrapShlSra = [](SDValue N, unsigned ShiftAmt) { |
3195 | if (N.getOpcode() != ISD::SRA || !isa<ConstantSDNode>(Val: N.getOperand(i: 1))) |
3196 | return N; |
3197 | |
3198 | SDValue N0 = N.getOperand(i: 0); |
3199 | if (N0.getOpcode() == ISD::SHL && isa<ConstantSDNode>(Val: N0.getOperand(i: 1)) && |
3200 | N.getConstantOperandVal(i: 1) == ShiftAmt && |
3201 | N0.getConstantOperandVal(i: 1) == ShiftAmt) |
3202 | return N0.getOperand(i: 0); |
3203 | |
3204 | return N; |
3205 | }; |
3206 | |
3207 | MVT VT = N.getSimpleValueType(); |
3208 | if (CurDAG->ComputeNumSignBits(Op: N) > (VT.getSizeInBits() - Bits)) { |
3209 | Val = UnwrapShlSra(N, VT.getSizeInBits() - Bits); |
3210 | return true; |
3211 | } |
3212 | |
3213 | return false; |
3214 | } |
3215 | |
3216 | bool RISCVDAGToDAGISel::selectZExtBits(SDValue N, unsigned Bits, SDValue &Val) { |
3217 | if (N.getOpcode() == ISD::AND) { |
3218 | auto *C = dyn_cast<ConstantSDNode>(Val: N.getOperand(i: 1)); |
3219 | if (C && C->getZExtValue() == maskTrailingOnes<uint64_t>(N: Bits)) { |
3220 | Val = N.getOperand(i: 0); |
3221 | return true; |
3222 | } |
3223 | } |
3224 | MVT VT = N.getSimpleValueType(); |
3225 | APInt Mask = APInt::getBitsSetFrom(numBits: VT.getSizeInBits(), loBit: Bits); |
3226 | if (CurDAG->MaskedValueIsZero(Op: N, Mask)) { |
3227 | Val = N; |
3228 | return true; |
3229 | } |
3230 | |
3231 | return false; |
3232 | } |
3233 | |
3234 | /// Look for various patterns that can be done with a SHL that can be folded |
3235 | /// into a SHXADD. \p ShAmt contains 1, 2, or 3 and is set based on which |
3236 | /// SHXADD we are trying to match. |
3237 | bool RISCVDAGToDAGISel::selectSHXADDOp(SDValue N, unsigned ShAmt, |
3238 | SDValue &Val) { |
3239 | if (N.getOpcode() == ISD::AND && isa<ConstantSDNode>(Val: N.getOperand(i: 1))) { |
3240 | SDValue N0 = N.getOperand(i: 0); |
3241 | |
3242 | if (bool LeftShift = N0.getOpcode() == ISD::SHL; |
3243 | (LeftShift || N0.getOpcode() == ISD::SRL) && |
3244 | isa<ConstantSDNode>(Val: N0.getOperand(i: 1))) { |
3245 | uint64_t Mask = N.getConstantOperandVal(i: 1); |
3246 | unsigned C2 = N0.getConstantOperandVal(i: 1); |
3247 | |
3248 | unsigned XLen = Subtarget->getXLen(); |
3249 | if (LeftShift) |
3250 | Mask &= maskTrailingZeros<uint64_t>(N: C2); |
3251 | else |
3252 | Mask &= maskTrailingOnes<uint64_t>(N: XLen - C2); |
3253 | |
3254 | if (isShiftedMask_64(Value: Mask)) { |
3255 | unsigned Leading = XLen - llvm::bit_width(Value: Mask); |
3256 | unsigned Trailing = llvm::countr_zero(Val: Mask); |
3257 | if (Trailing != ShAmt) |
3258 | return false; |
3259 | |
3260 | unsigned Opcode; |
3261 | // Look for (and (shl y, c2), c1) where c1 is a shifted mask with no |
3262 | // leading zeros and c3 trailing zeros. We can use an SRLI by c3-c2 |
3263 | // followed by a SHXADD with c3 for the X amount. |
3264 | if (LeftShift && Leading == 0 && C2 < Trailing) |
3265 | Opcode = RISCV::SRLI; |
3266 | // Look for (and (shl y, c2), c1) where c1 is a shifted mask with 32-c2 |
3267 | // leading zeros and c3 trailing zeros. We can use an SRLIW by c3-c2 |
3268 | // followed by a SHXADD with c3 for the X amount. |
3269 | else if (LeftShift && Leading == 32 - C2 && C2 < Trailing) |
3270 | Opcode = RISCV::SRLIW; |
3271 | // Look for (and (shr y, c2), c1) where c1 is a shifted mask with c2 |
3272 | // leading zeros and c3 trailing zeros. We can use an SRLI by c2+c3 |
3273 | // followed by a SHXADD using c3 for the X amount. |
3274 | else if (!LeftShift && Leading == C2) |
3275 | Opcode = RISCV::SRLI; |
3276 | // Look for (and (shr y, c2), c1) where c1 is a shifted mask with 32+c2 |
3277 | // leading zeros and c3 trailing zeros. We can use an SRLIW by c2+c3 |
3278 | // followed by a SHXADD using c3 for the X amount. |
3279 | else if (!LeftShift && Leading == 32 + C2) |
3280 | Opcode = RISCV::SRLIW; |
3281 | else |
3282 | return false; |
3283 | |
3284 | SDLoc DL(N); |
3285 | EVT VT = N.getValueType(); |
3286 | ShAmt = LeftShift ? Trailing - C2 : Trailing + C2; |
3287 | Val = SDValue( |
3288 | CurDAG->getMachineNode(Opcode, dl: DL, VT, Op1: N0.getOperand(i: 0), |
3289 | Op2: CurDAG->getTargetConstant(Val: ShAmt, DL, VT)), |
3290 | 0); |
3291 | return true; |
3292 | } |
3293 | } else if (N0.getOpcode() == ISD::SRA && N0.hasOneUse() && |
3294 | isa<ConstantSDNode>(Val: N0.getOperand(i: 1))) { |
3295 | uint64_t Mask = N.getConstantOperandVal(i: 1); |
3296 | unsigned C2 = N0.getConstantOperandVal(i: 1); |
3297 | |
3298 | // Look for (and (sra y, c2), c1) where c1 is a shifted mask with c3 |
3299 | // leading zeros and c4 trailing zeros. If c2 is greater than c3, we can |
3300 | // use (srli (srai y, c2 - c3), c3 + c4) followed by a SHXADD with c4 as |
3301 | // the X amount. |
3302 | if (isShiftedMask_64(Value: Mask)) { |
3303 | unsigned XLen = Subtarget->getXLen(); |
3304 | unsigned Leading = XLen - llvm::bit_width(Value: Mask); |
3305 | unsigned Trailing = llvm::countr_zero(Val: Mask); |
3306 | if (C2 > Leading && Leading > 0 && Trailing == ShAmt) { |
3307 | SDLoc DL(N); |
3308 | EVT VT = N.getValueType(); |
3309 | Val = SDValue(CurDAG->getMachineNode( |
3310 | Opcode: RISCV::SRAI, dl: DL, VT, Op1: N0.getOperand(i: 0), |
3311 | Op2: CurDAG->getTargetConstant(Val: C2 - Leading, DL, VT)), |
3312 | 0); |
3313 | Val = SDValue(CurDAG->getMachineNode( |
3314 | Opcode: RISCV::SRLI, dl: DL, VT, Op1: Val, |
3315 | Op2: CurDAG->getTargetConstant(Val: Leading + ShAmt, DL, VT)), |
3316 | 0); |
3317 | return true; |
3318 | } |
3319 | } |
3320 | } |
3321 | } else if (bool LeftShift = N.getOpcode() == ISD::SHL; |
3322 | (LeftShift || N.getOpcode() == ISD::SRL) && |
3323 | isa<ConstantSDNode>(Val: N.getOperand(i: 1))) { |
3324 | SDValue N0 = N.getOperand(i: 0); |
3325 | if (N0.getOpcode() == ISD::AND && N0.hasOneUse() && |
3326 | isa<ConstantSDNode>(Val: N0.getOperand(i: 1))) { |
3327 | uint64_t Mask = N0.getConstantOperandVal(i: 1); |
3328 | if (isShiftedMask_64(Value: Mask)) { |
3329 | unsigned C1 = N.getConstantOperandVal(i: 1); |
3330 | unsigned XLen = Subtarget->getXLen(); |
3331 | unsigned Leading = XLen - llvm::bit_width(Value: Mask); |
3332 | unsigned Trailing = llvm::countr_zero(Val: Mask); |
3333 | // Look for (shl (and X, Mask), C1) where Mask has 32 leading zeros and |
3334 | // C3 trailing zeros. If C1+C3==ShAmt we can use SRLIW+SHXADD. |
3335 | if (LeftShift && Leading == 32 && Trailing > 0 && |
3336 | (Trailing + C1) == ShAmt) { |
3337 | SDLoc DL(N); |
3338 | EVT VT = N.getValueType(); |
3339 | Val = SDValue(CurDAG->getMachineNode( |
3340 | Opcode: RISCV::SRLIW, dl: DL, VT, Op1: N0.getOperand(i: 0), |
3341 | Op2: CurDAG->getTargetConstant(Val: Trailing, DL, VT)), |
3342 | 0); |
3343 | return true; |
3344 | } |
3345 | // Look for (srl (and X, Mask), C1) where Mask has 32 leading zeros and |
3346 | // C3 trailing zeros. If C3-C1==ShAmt we can use SRLIW+SHXADD. |
3347 | if (!LeftShift && Leading == 32 && Trailing > C1 && |
3348 | (Trailing - C1) == ShAmt) { |
3349 | SDLoc DL(N); |
3350 | EVT VT = N.getValueType(); |
3351 | Val = SDValue(CurDAG->getMachineNode( |
3352 | Opcode: RISCV::SRLIW, dl: DL, VT, Op1: N0.getOperand(i: 0), |
3353 | Op2: CurDAG->getTargetConstant(Val: Trailing, DL, VT)), |
3354 | 0); |
3355 | return true; |
3356 | } |
3357 | } |
3358 | } |
3359 | } |
3360 | |
3361 | return false; |
3362 | } |
3363 | |
3364 | /// Look for various patterns that can be done with a SHL that can be folded |
3365 | /// into a SHXADD_UW. \p ShAmt contains 1, 2, or 3 and is set based on which |
3366 | /// SHXADD_UW we are trying to match. |
3367 | bool RISCVDAGToDAGISel::selectSHXADD_UWOp(SDValue N, unsigned ShAmt, |
3368 | SDValue &Val) { |
3369 | if (N.getOpcode() == ISD::AND && isa<ConstantSDNode>(Val: N.getOperand(i: 1)) && |
3370 | N.hasOneUse()) { |
3371 | SDValue N0 = N.getOperand(i: 0); |
3372 | if (N0.getOpcode() == ISD::SHL && isa<ConstantSDNode>(Val: N0.getOperand(i: 1)) && |
3373 | N0.hasOneUse()) { |
3374 | uint64_t Mask = N.getConstantOperandVal(i: 1); |
3375 | unsigned C2 = N0.getConstantOperandVal(i: 1); |
3376 | |
3377 | Mask &= maskTrailingZeros<uint64_t>(N: C2); |
3378 | |
3379 | // Look for (and (shl y, c2), c1) where c1 is a shifted mask with |
3380 | // 32-ShAmt leading zeros and c2 trailing zeros. We can use SLLI by |
3381 | // c2-ShAmt followed by SHXADD_UW with ShAmt for the X amount. |
3382 | if (isShiftedMask_64(Value: Mask)) { |
3383 | unsigned Leading = llvm::countl_zero(Val: Mask); |
3384 | unsigned Trailing = llvm::countr_zero(Val: Mask); |
3385 | if (Leading == 32 - ShAmt && Trailing == C2 && Trailing > ShAmt) { |
3386 | SDLoc DL(N); |
3387 | EVT VT = N.getValueType(); |
3388 | Val = SDValue(CurDAG->getMachineNode( |
3389 | Opcode: RISCV::SLLI, dl: DL, VT, Op1: N0.getOperand(i: 0), |
3390 | Op2: CurDAG->getTargetConstant(Val: C2 - ShAmt, DL, VT)), |
3391 | 0); |
3392 | return true; |
3393 | } |
3394 | } |
3395 | } |
3396 | } |
3397 | |
3398 | return false; |
3399 | } |
3400 | |
3401 | bool RISCVDAGToDAGISel::orIsAdd(const SDNode *N) const { |
3402 | if (N->getFlags().hasDisjoint()) |
3403 | return true; |
3404 | KnownBits Known0 = CurDAG->computeKnownBits(Op: N->getOperand(Num: 0), Depth: 0); |
3405 | KnownBits Known1 = CurDAG->computeKnownBits(Op: N->getOperand(Num: 1), Depth: 0); |
3406 | return KnownBits::haveNoCommonBitsSet(LHS: Known0, RHS: Known1); |
3407 | } |
3408 | |
3409 | bool RISCVDAGToDAGISel::selectImm64IfCheaper(int64_t Imm, int64_t OrigImm, |
3410 | SDValue N, SDValue &Val) { |
3411 | int OrigCost = RISCVMatInt::getIntMatCost(Val: APInt(64, OrigImm), Size: 64, STI: *Subtarget, |
3412 | /*CompressionCost=*/true); |
3413 | int Cost = RISCVMatInt::getIntMatCost(Val: APInt(64, Imm), Size: 64, STI: *Subtarget, |
3414 | /*CompressionCost=*/true); |
3415 | if (OrigCost <= Cost) |
3416 | return false; |
3417 | |
3418 | Val = selectImm(CurDAG, DL: SDLoc(N), VT: N->getSimpleValueType(ResNo: 0), Imm, Subtarget: *Subtarget); |
3419 | return true; |
3420 | } |
3421 | |
3422 | bool RISCVDAGToDAGISel::selectZExtImm32(SDValue N, SDValue &Val) { |
3423 | if (!isa<ConstantSDNode>(Val: N)) |
3424 | return false; |
3425 | int64_t Imm = cast<ConstantSDNode>(Val&: N)->getSExtValue(); |
3426 | if ((Imm >> 31) != 1) |
3427 | return false; |
3428 | |
3429 | for (const SDNode *U : N->users()) { |
3430 | switch (U->getOpcode()) { |
3431 | case ISD::ADD: |
3432 | break; |
3433 | case ISD::OR: |
3434 | if (orIsAdd(N: U)) |
3435 | break; |
3436 | return false; |
3437 | default: |
3438 | return false; |
3439 | } |
3440 | } |
3441 | |
3442 | return selectImm64IfCheaper(Imm: 0xffffffff00000000 | Imm, OrigImm: Imm, N, Val); |
3443 | } |
3444 | |
3445 | bool RISCVDAGToDAGISel::selectNegImm(SDValue N, SDValue &Val) { |
3446 | if (!isa<ConstantSDNode>(Val: N)) |
3447 | return false; |
3448 | int64_t Imm = cast<ConstantSDNode>(Val&: N)->getSExtValue(); |
3449 | if (isInt<32>(x: Imm)) |
3450 | return false; |
3451 | |
3452 | for (const SDNode *U : N->users()) { |
3453 | switch (U->getOpcode()) { |
3454 | case ISD::ADD: |
3455 | break; |
3456 | case RISCVISD::VMV_V_X_VL: |
3457 | if (!all_of(Range: U->users(), P: [](const SDNode *V) { |
3458 | return V->getOpcode() == ISD::ADD || |
3459 | V->getOpcode() == RISCVISD::ADD_VL; |
3460 | })) |
3461 | return false; |
3462 | break; |
3463 | default: |
3464 | return false; |
3465 | } |
3466 | } |
3467 | |
3468 | return selectImm64IfCheaper(Imm: -Imm, OrigImm: Imm, N, Val); |
3469 | } |
3470 | |
3471 | bool RISCVDAGToDAGISel::selectInvLogicImm(SDValue N, SDValue &Val) { |
3472 | if (!isa<ConstantSDNode>(Val: N)) |
3473 | return false; |
3474 | int64_t Imm = cast<ConstantSDNode>(Val&: N)->getSExtValue(); |
3475 | |
3476 | // For 32-bit signed constants, we can only substitute LUI+ADDI with LUI. |
3477 | if (isInt<32>(x: Imm) && ((Imm & 0xfff) != 0xfff || Imm == -1)) |
3478 | return false; |
3479 | |
3480 | // Abandon this transform if the constant is needed elsewhere. |
3481 | for (const SDNode *U : N->users()) { |
3482 | switch (U->getOpcode()) { |
3483 | case ISD::AND: |
3484 | case ISD::OR: |
3485 | case ISD::XOR: |
3486 | if (!(Subtarget->hasStdExtZbb() || Subtarget->hasStdExtZbkb())) |
3487 | return false; |
3488 | break; |
3489 | case RISCVISD::VMV_V_X_VL: |
3490 | if (!Subtarget->hasStdExtZvkb()) |
3491 | return false; |
3492 | if (!all_of(Range: U->users(), P: [](const SDNode *V) { |
3493 | return V->getOpcode() == ISD::AND || |
3494 | V->getOpcode() == RISCVISD::AND_VL; |
3495 | })) |
3496 | return false; |
3497 | break; |
3498 | default: |
3499 | return false; |
3500 | } |
3501 | } |
3502 | |
3503 | if (isInt<32>(x: Imm)) { |
3504 | Val = |
3505 | selectImm(CurDAG, DL: SDLoc(N), VT: N->getSimpleValueType(ResNo: 0), Imm: ~Imm, Subtarget: *Subtarget); |
3506 | return true; |
3507 | } |
3508 | |
3509 | // For 64-bit constants, the instruction sequences get complex, |
3510 | // so we select inverted only if it's cheaper. |
3511 | return selectImm64IfCheaper(Imm: ~Imm, OrigImm: Imm, N, Val); |
3512 | } |
3513 | |
3514 | static bool vectorPseudoHasAllNBitUsers(SDNode *User, unsigned UserOpNo, |
3515 | unsigned Bits, |
3516 | const TargetInstrInfo *TII) { |
3517 | unsigned MCOpcode = RISCV::getRVVMCOpcode(RVVPseudoOpcode: User->getMachineOpcode()); |
3518 | |
3519 | if (!MCOpcode) |
3520 | return false; |
3521 | |
3522 | const MCInstrDesc &MCID = TII->get(Opcode: User->getMachineOpcode()); |
3523 | const uint64_t TSFlags = MCID.TSFlags; |
3524 | if (!RISCVII::hasSEWOp(TSFlags)) |
3525 | return false; |
3526 | assert(RISCVII::hasVLOp(TSFlags)); |
3527 | |
3528 | unsigned ChainOpIdx = User->getNumOperands() - 1; |
3529 | bool HasChainOp = User->getOperand(Num: ChainOpIdx).getValueType() == MVT::Other; |
3530 | bool HasVecPolicyOp = RISCVII::hasVecPolicyOp(TSFlags); |
3531 | unsigned VLIdx = User->getNumOperands() - HasVecPolicyOp - HasChainOp - 2; |
3532 | const unsigned Log2SEW = User->getConstantOperandVal(Num: VLIdx + 1); |
3533 | |
3534 | if (UserOpNo == VLIdx) |
3535 | return false; |
3536 | |
3537 | auto NumDemandedBits = |
3538 | RISCV::getVectorLowDemandedScalarBits(Opcode: MCOpcode, Log2SEW); |
3539 | return NumDemandedBits && Bits >= *NumDemandedBits; |
3540 | } |
3541 | |
3542 | // Return true if all users of this SDNode* only consume the lower \p Bits. |
3543 | // This can be used to form W instructions for add/sub/mul/shl even when the |
3544 | // root isn't a sext_inreg. This can allow the ADDW/SUBW/MULW/SLLIW to CSE if |
3545 | // SimplifyDemandedBits has made it so some users see a sext_inreg and some |
3546 | // don't. The sext_inreg+add/sub/mul/shl will get selected, but still leave |
3547 | // the add/sub/mul/shl to become non-W instructions. By checking the users we |
3548 | // may be able to use a W instruction and CSE with the other instruction if |
3549 | // this has happened. We could try to detect that the CSE opportunity exists |
3550 | // before doing this, but that would be more complicated. |
3551 | bool RISCVDAGToDAGISel::hasAllNBitUsers(SDNode *Node, unsigned Bits, |
3552 | const unsigned Depth) const { |
3553 | assert((Node->getOpcode() == ISD::ADD || Node->getOpcode() == ISD::SUB || |
3554 | Node->getOpcode() == ISD::MUL || Node->getOpcode() == ISD::SHL || |
3555 | Node->getOpcode() == ISD::SRL || Node->getOpcode() == ISD::AND || |
3556 | Node->getOpcode() == ISD::OR || Node->getOpcode() == ISD::XOR || |
3557 | Node->getOpcode() == ISD::SIGN_EXTEND_INREG || |
3558 | isa<ConstantSDNode>(Node) || Depth != 0) && |
3559 | "Unexpected opcode" ); |
3560 | |
3561 | if (Depth >= SelectionDAG::MaxRecursionDepth) |
3562 | return false; |
3563 | |
3564 | // The PatFrags that call this may run before RISCVGenDAGISel.inc has checked |
3565 | // the VT. Ensure the type is scalar to avoid wasting time on vectors. |
3566 | if (Depth == 0 && !Node->getValueType(ResNo: 0).isScalarInteger()) |
3567 | return false; |
3568 | |
3569 | for (SDUse &Use : Node->uses()) { |
3570 | SDNode *User = Use.getUser(); |
3571 | // Users of this node should have already been instruction selected |
3572 | if (!User->isMachineOpcode()) |
3573 | return false; |
3574 | |
3575 | // TODO: Add more opcodes? |
3576 | switch (User->getMachineOpcode()) { |
3577 | default: |
3578 | if (vectorPseudoHasAllNBitUsers(User, UserOpNo: Use.getOperandNo(), Bits, TII)) |
3579 | break; |
3580 | return false; |
3581 | case RISCV::ADDW: |
3582 | case RISCV::ADDIW: |
3583 | case RISCV::SUBW: |
3584 | case RISCV::MULW: |
3585 | case RISCV::SLLW: |
3586 | case RISCV::SLLIW: |
3587 | case RISCV::SRAW: |
3588 | case RISCV::SRAIW: |
3589 | case RISCV::SRLW: |
3590 | case RISCV::SRLIW: |
3591 | case RISCV::DIVW: |
3592 | case RISCV::DIVUW: |
3593 | case RISCV::REMW: |
3594 | case RISCV::REMUW: |
3595 | case RISCV::ROLW: |
3596 | case RISCV::RORW: |
3597 | case RISCV::RORIW: |
3598 | case RISCV::CLZW: |
3599 | case RISCV::CTZW: |
3600 | case RISCV::CPOPW: |
3601 | case RISCV::SLLI_UW: |
3602 | case RISCV::FMV_W_X: |
3603 | case RISCV::FCVT_H_W: |
3604 | case RISCV::FCVT_H_W_INX: |
3605 | case RISCV::FCVT_H_WU: |
3606 | case RISCV::FCVT_H_WU_INX: |
3607 | case RISCV::FCVT_S_W: |
3608 | case RISCV::FCVT_S_W_INX: |
3609 | case RISCV::FCVT_S_WU: |
3610 | case RISCV::FCVT_S_WU_INX: |
3611 | case RISCV::FCVT_D_W: |
3612 | case RISCV::FCVT_D_W_INX: |
3613 | case RISCV::FCVT_D_WU: |
3614 | case RISCV::FCVT_D_WU_INX: |
3615 | case RISCV::TH_REVW: |
3616 | case RISCV::TH_SRRIW: |
3617 | if (Bits >= 32) |
3618 | break; |
3619 | return false; |
3620 | case RISCV::SLL: |
3621 | case RISCV::SRA: |
3622 | case RISCV::SRL: |
3623 | case RISCV::ROL: |
3624 | case RISCV::ROR: |
3625 | case RISCV::BSET: |
3626 | case RISCV::BCLR: |
3627 | case RISCV::BINV: |
3628 | // Shift amount operands only use log2(Xlen) bits. |
3629 | if (Use.getOperandNo() == 1 && Bits >= Log2_32(Value: Subtarget->getXLen())) |
3630 | break; |
3631 | return false; |
3632 | case RISCV::SLLI: |
3633 | // SLLI only uses the lower (XLen - ShAmt) bits. |
3634 | if (Bits >= Subtarget->getXLen() - User->getConstantOperandVal(Num: 1)) |
3635 | break; |
3636 | return false; |
3637 | case RISCV::ANDI: |
3638 | if (Bits >= (unsigned)llvm::bit_width(Value: User->getConstantOperandVal(Num: 1))) |
3639 | break; |
3640 | goto RecCheck; |
3641 | case RISCV::ORI: { |
3642 | uint64_t Imm = cast<ConstantSDNode>(Val: User->getOperand(Num: 1))->getSExtValue(); |
3643 | if (Bits >= (unsigned)llvm::bit_width<uint64_t>(Value: ~Imm)) |
3644 | break; |
3645 | [[fallthrough]]; |
3646 | } |
3647 | case RISCV::AND: |
3648 | case RISCV::OR: |
3649 | case RISCV::XOR: |
3650 | case RISCV::XORI: |
3651 | case RISCV::ANDN: |
3652 | case RISCV::ORN: |
3653 | case RISCV::XNOR: |
3654 | case RISCV::SH1ADD: |
3655 | case RISCV::SH2ADD: |
3656 | case RISCV::SH3ADD: |
3657 | RecCheck: |
3658 | if (hasAllNBitUsers(Node: User, Bits, Depth: Depth + 1)) |
3659 | break; |
3660 | return false; |
3661 | case RISCV::SRLI: { |
3662 | unsigned ShAmt = User->getConstantOperandVal(Num: 1); |
3663 | // If we are shifting right by less than Bits, and users don't demand any |
3664 | // bits that were shifted into [Bits-1:0], then we can consider this as an |
3665 | // N-Bit user. |
3666 | if (Bits > ShAmt && hasAllNBitUsers(Node: User, Bits: Bits - ShAmt, Depth: Depth + 1)) |
3667 | break; |
3668 | return false; |
3669 | } |
3670 | case RISCV::SEXT_B: |
3671 | case RISCV::PACKH: |
3672 | if (Bits >= 8) |
3673 | break; |
3674 | return false; |
3675 | case RISCV::SEXT_H: |
3676 | case RISCV::FMV_H_X: |
3677 | case RISCV::ZEXT_H_RV32: |
3678 | case RISCV::ZEXT_H_RV64: |
3679 | case RISCV::PACKW: |
3680 | if (Bits >= 16) |
3681 | break; |
3682 | return false; |
3683 | case RISCV::PACK: |
3684 | if (Bits >= (Subtarget->getXLen() / 2)) |
3685 | break; |
3686 | return false; |
3687 | case RISCV::ADD_UW: |
3688 | case RISCV::SH1ADD_UW: |
3689 | case RISCV::SH2ADD_UW: |
3690 | case RISCV::SH3ADD_UW: |
3691 | // The first operand to add.uw/shXadd.uw is implicitly zero extended from |
3692 | // 32 bits. |
3693 | if (Use.getOperandNo() == 0 && Bits >= 32) |
3694 | break; |
3695 | return false; |
3696 | case RISCV::SB: |
3697 | if (Use.getOperandNo() == 0 && Bits >= 8) |
3698 | break; |
3699 | return false; |
3700 | case RISCV::SH: |
3701 | if (Use.getOperandNo() == 0 && Bits >= 16) |
3702 | break; |
3703 | return false; |
3704 | case RISCV::SW: |
3705 | if (Use.getOperandNo() == 0 && Bits >= 32) |
3706 | break; |
3707 | return false; |
3708 | } |
3709 | } |
3710 | |
3711 | return true; |
3712 | } |
3713 | |
3714 | // Select a constant that can be represented as (sign_extend(imm5) << imm2). |
3715 | bool RISCVDAGToDAGISel::selectSimm5Shl2(SDValue N, SDValue &Simm5, |
3716 | SDValue &Shl2) { |
3717 | if (auto *C = dyn_cast<ConstantSDNode>(Val&: N)) { |
3718 | int64_t Offset = C->getSExtValue(); |
3719 | unsigned Shift; |
3720 | for (Shift = 0; Shift < 4; Shift++) |
3721 | if (isInt<5>(x: Offset >> Shift) && ((Offset % (1LL << Shift)) == 0)) |
3722 | break; |
3723 | |
3724 | // Constant cannot be encoded. |
3725 | if (Shift == 4) |
3726 | return false; |
3727 | |
3728 | EVT Ty = N->getValueType(ResNo: 0); |
3729 | Simm5 = CurDAG->getSignedTargetConstant(Val: Offset >> Shift, DL: SDLoc(N), VT: Ty); |
3730 | Shl2 = CurDAG->getTargetConstant(Val: Shift, DL: SDLoc(N), VT: Ty); |
3731 | return true; |
3732 | } |
3733 | |
3734 | return false; |
3735 | } |
3736 | |
3737 | // Select VL as a 5 bit immediate or a value that will become a register. This |
3738 | // allows us to choose between VSETIVLI or VSETVLI later. |
3739 | bool RISCVDAGToDAGISel::selectVLOp(SDValue N, SDValue &VL) { |
3740 | auto *C = dyn_cast<ConstantSDNode>(Val&: N); |
3741 | if (C && isUInt<5>(x: C->getZExtValue())) { |
3742 | VL = CurDAG->getTargetConstant(Val: C->getZExtValue(), DL: SDLoc(N), |
3743 | VT: N->getValueType(ResNo: 0)); |
3744 | } else if (C && C->isAllOnes()) { |
3745 | // Treat all ones as VLMax. |
3746 | VL = CurDAG->getSignedTargetConstant(Val: RISCV::VLMaxSentinel, DL: SDLoc(N), |
3747 | VT: N->getValueType(ResNo: 0)); |
3748 | } else if (isa<RegisterSDNode>(Val: N) && |
3749 | cast<RegisterSDNode>(Val&: N)->getReg() == RISCV::X0) { |
3750 | // All our VL operands use an operand that allows GPRNoX0 or an immediate |
3751 | // as the register class. Convert X0 to a special immediate to pass the |
3752 | // MachineVerifier. This is recognized specially by the vsetvli insertion |
3753 | // pass. |
3754 | VL = CurDAG->getSignedTargetConstant(Val: RISCV::VLMaxSentinel, DL: SDLoc(N), |
3755 | VT: N->getValueType(ResNo: 0)); |
3756 | } else { |
3757 | VL = N; |
3758 | } |
3759 | |
3760 | return true; |
3761 | } |
3762 | |
3763 | static SDValue findVSplat(SDValue N) { |
3764 | if (N.getOpcode() == ISD::INSERT_SUBVECTOR) { |
3765 | if (!N.getOperand(i: 0).isUndef()) |
3766 | return SDValue(); |
3767 | N = N.getOperand(i: 1); |
3768 | } |
3769 | SDValue Splat = N; |
3770 | if ((Splat.getOpcode() != RISCVISD::VMV_V_X_VL && |
3771 | Splat.getOpcode() != RISCVISD::VMV_S_X_VL) || |
3772 | !Splat.getOperand(i: 0).isUndef()) |
3773 | return SDValue(); |
3774 | assert(Splat.getNumOperands() == 3 && "Unexpected number of operands" ); |
3775 | return Splat; |
3776 | } |
3777 | |
3778 | bool RISCVDAGToDAGISel::selectVSplat(SDValue N, SDValue &SplatVal) { |
3779 | SDValue Splat = findVSplat(N); |
3780 | if (!Splat) |
3781 | return false; |
3782 | |
3783 | SplatVal = Splat.getOperand(i: 1); |
3784 | return true; |
3785 | } |
3786 | |
3787 | static bool selectVSplatImmHelper(SDValue N, SDValue &SplatVal, |
3788 | SelectionDAG &DAG, |
3789 | const RISCVSubtarget &Subtarget, |
3790 | std::function<bool(int64_t)> ValidateImm, |
3791 | bool Decrement = false) { |
3792 | SDValue Splat = findVSplat(N); |
3793 | if (!Splat || !isa<ConstantSDNode>(Val: Splat.getOperand(i: 1))) |
3794 | return false; |
3795 | |
3796 | const unsigned SplatEltSize = Splat.getScalarValueSizeInBits(); |
3797 | assert(Subtarget.getXLenVT() == Splat.getOperand(1).getSimpleValueType() && |
3798 | "Unexpected splat operand type" ); |
3799 | |
3800 | // The semantics of RISCVISD::VMV_V_X_VL is that when the operand |
3801 | // type is wider than the resulting vector element type: an implicit |
3802 | // truncation first takes place. Therefore, perform a manual |
3803 | // truncation/sign-extension in order to ignore any truncated bits and catch |
3804 | // any zero-extended immediate. |
3805 | // For example, we wish to match (i8 -1) -> (XLenVT 255) as a simm5 by first |
3806 | // sign-extending to (XLenVT -1). |
3807 | APInt SplatConst = Splat.getConstantOperandAPInt(i: 1).sextOrTrunc(width: SplatEltSize); |
3808 | |
3809 | int64_t SplatImm = SplatConst.getSExtValue(); |
3810 | |
3811 | if (!ValidateImm(SplatImm)) |
3812 | return false; |
3813 | |
3814 | if (Decrement) |
3815 | SplatImm -= 1; |
3816 | |
3817 | SplatVal = |
3818 | DAG.getSignedTargetConstant(Val: SplatImm, DL: SDLoc(N), VT: Subtarget.getXLenVT()); |
3819 | return true; |
3820 | } |
3821 | |
3822 | bool RISCVDAGToDAGISel::selectVSplatSimm5(SDValue N, SDValue &SplatVal) { |
3823 | return selectVSplatImmHelper(N, SplatVal, DAG&: *CurDAG, Subtarget: *Subtarget, |
3824 | ValidateImm: [](int64_t Imm) { return isInt<5>(x: Imm); }); |
3825 | } |
3826 | |
3827 | bool RISCVDAGToDAGISel::selectVSplatSimm5Plus1(SDValue N, SDValue &SplatVal) { |
3828 | return selectVSplatImmHelper( |
3829 | N, SplatVal, DAG&: *CurDAG, Subtarget: *Subtarget, |
3830 | ValidateImm: [](int64_t Imm) { return (isInt<5>(x: Imm) && Imm != -16) || Imm == 16; }, |
3831 | /*Decrement=*/true); |
3832 | } |
3833 | |
3834 | bool RISCVDAGToDAGISel::selectVSplatSimm5Plus1NoDec(SDValue N, SDValue &SplatVal) { |
3835 | return selectVSplatImmHelper( |
3836 | N, SplatVal, DAG&: *CurDAG, Subtarget: *Subtarget, |
3837 | ValidateImm: [](int64_t Imm) { return (isInt<5>(x: Imm) && Imm != -16) || Imm == 16; }, |
3838 | /*Decrement=*/false); |
3839 | } |
3840 | |
3841 | bool RISCVDAGToDAGISel::selectVSplatSimm5Plus1NonZero(SDValue N, |
3842 | SDValue &SplatVal) { |
3843 | return selectVSplatImmHelper( |
3844 | N, SplatVal, DAG&: *CurDAG, Subtarget: *Subtarget, |
3845 | ValidateImm: [](int64_t Imm) { |
3846 | return Imm != 0 && ((isInt<5>(x: Imm) && Imm != -16) || Imm == 16); |
3847 | }, |
3848 | /*Decrement=*/true); |
3849 | } |
3850 | |
3851 | bool RISCVDAGToDAGISel::selectVSplatUimm(SDValue N, unsigned Bits, |
3852 | SDValue &SplatVal) { |
3853 | return selectVSplatImmHelper( |
3854 | N, SplatVal, DAG&: *CurDAG, Subtarget: *Subtarget, |
3855 | ValidateImm: [Bits](int64_t Imm) { return isUIntN(N: Bits, x: Imm); }); |
3856 | } |
3857 | |
3858 | bool RISCVDAGToDAGISel::selectVSplatImm64Neg(SDValue N, SDValue &SplatVal) { |
3859 | SDValue Splat = findVSplat(N); |
3860 | return Splat && selectNegImm(N: Splat.getOperand(i: 1), Val&: SplatVal); |
3861 | } |
3862 | |
3863 | bool RISCVDAGToDAGISel::selectLow8BitsVSplat(SDValue N, SDValue &SplatVal) { |
3864 | auto IsExtOrTrunc = [](SDValue N) { |
3865 | switch (N->getOpcode()) { |
3866 | case ISD::SIGN_EXTEND: |
3867 | case ISD::ZERO_EXTEND: |
3868 | // There's no passthru on these _VL nodes so any VL/mask is ok, since any |
3869 | // inactive elements will be undef. |
3870 | case RISCVISD::TRUNCATE_VECTOR_VL: |
3871 | case RISCVISD::VSEXT_VL: |
3872 | case RISCVISD::VZEXT_VL: |
3873 | return true; |
3874 | default: |
3875 | return false; |
3876 | } |
3877 | }; |
3878 | |
3879 | // We can have multiple nested nodes, so unravel them all if needed. |
3880 | while (IsExtOrTrunc(N)) { |
3881 | if (!N.hasOneUse() || N.getScalarValueSizeInBits() < 8) |
3882 | return false; |
3883 | N = N->getOperand(Num: 0); |
3884 | } |
3885 | |
3886 | return selectVSplat(N, SplatVal); |
3887 | } |
3888 | |
3889 | bool RISCVDAGToDAGISel::selectScalarFPAsInt(SDValue N, SDValue &Imm) { |
3890 | // Allow bitcasts from XLenVT -> FP. |
3891 | if (N.getOpcode() == ISD::BITCAST && |
3892 | N.getOperand(i: 0).getValueType() == Subtarget->getXLenVT()) { |
3893 | Imm = N.getOperand(i: 0); |
3894 | return true; |
3895 | } |
3896 | // Allow moves from XLenVT to FP. |
3897 | if (N.getOpcode() == RISCVISD::FMV_H_X || |
3898 | N.getOpcode() == RISCVISD::FMV_W_X_RV64) { |
3899 | Imm = N.getOperand(i: 0); |
3900 | return true; |
3901 | } |
3902 | |
3903 | // Otherwise, look for FP constants that can materialized with scalar int. |
3904 | ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Val: N.getNode()); |
3905 | if (!CFP) |
3906 | return false; |
3907 | const APFloat &APF = CFP->getValueAPF(); |
3908 | // td can handle +0.0 already. |
3909 | if (APF.isPosZero()) |
3910 | return false; |
3911 | |
3912 | MVT VT = CFP->getSimpleValueType(ResNo: 0); |
3913 | |
3914 | MVT XLenVT = Subtarget->getXLenVT(); |
3915 | if (VT == MVT::f64 && !Subtarget->is64Bit()) { |
3916 | assert(APF.isNegZero() && "Unexpected constant." ); |
3917 | return false; |
3918 | } |
3919 | SDLoc DL(N); |
3920 | Imm = selectImm(CurDAG, DL, VT: XLenVT, Imm: APF.bitcastToAPInt().getSExtValue(), |
3921 | Subtarget: *Subtarget); |
3922 | return true; |
3923 | } |
3924 | |
3925 | bool RISCVDAGToDAGISel::selectRVVSimm5(SDValue N, unsigned Width, |
3926 | SDValue &Imm) { |
3927 | if (auto *C = dyn_cast<ConstantSDNode>(Val&: N)) { |
3928 | int64_t ImmVal = SignExtend64(X: C->getSExtValue(), B: Width); |
3929 | |
3930 | if (!isInt<5>(x: ImmVal)) |
3931 | return false; |
3932 | |
3933 | Imm = CurDAG->getSignedTargetConstant(Val: ImmVal, DL: SDLoc(N), |
3934 | VT: Subtarget->getXLenVT()); |
3935 | return true; |
3936 | } |
3937 | |
3938 | return false; |
3939 | } |
3940 | |
3941 | // Try to remove sext.w if the input is a W instruction or can be made into |
3942 | // a W instruction cheaply. |
3943 | bool RISCVDAGToDAGISel::doPeepholeSExtW(SDNode *N) { |
3944 | // Look for the sext.w pattern, addiw rd, rs1, 0. |
3945 | if (N->getMachineOpcode() != RISCV::ADDIW || |
3946 | !isNullConstant(V: N->getOperand(Num: 1))) |
3947 | return false; |
3948 | |
3949 | SDValue N0 = N->getOperand(Num: 0); |
3950 | if (!N0.isMachineOpcode()) |
3951 | return false; |
3952 | |
3953 | switch (N0.getMachineOpcode()) { |
3954 | default: |
3955 | break; |
3956 | case RISCV::ADD: |
3957 | case RISCV::ADDI: |
3958 | case RISCV::SUB: |
3959 | case RISCV::MUL: |
3960 | case RISCV::SLLI: { |
3961 | // Convert sext.w+add/sub/mul to their W instructions. This will create |
3962 | // a new independent instruction. This improves latency. |
3963 | unsigned Opc; |
3964 | switch (N0.getMachineOpcode()) { |
3965 | default: |
3966 | llvm_unreachable("Unexpected opcode!" ); |
3967 | case RISCV::ADD: Opc = RISCV::ADDW; break; |
3968 | case RISCV::ADDI: Opc = RISCV::ADDIW; break; |
3969 | case RISCV::SUB: Opc = RISCV::SUBW; break; |
3970 | case RISCV::MUL: Opc = RISCV::MULW; break; |
3971 | case RISCV::SLLI: Opc = RISCV::SLLIW; break; |
3972 | } |
3973 | |
3974 | SDValue N00 = N0.getOperand(i: 0); |
3975 | SDValue N01 = N0.getOperand(i: 1); |
3976 | |
3977 | // Shift amount needs to be uimm5. |
3978 | if (N0.getMachineOpcode() == RISCV::SLLI && |
3979 | !isUInt<5>(x: cast<ConstantSDNode>(Val&: N01)->getSExtValue())) |
3980 | break; |
3981 | |
3982 | SDNode *Result = |
3983 | CurDAG->getMachineNode(Opcode: Opc, dl: SDLoc(N), VT: N->getValueType(ResNo: 0), |
3984 | Op1: N00, Op2: N01); |
3985 | ReplaceUses(F: N, T: Result); |
3986 | return true; |
3987 | } |
3988 | case RISCV::ADDW: |
3989 | case RISCV::ADDIW: |
3990 | case RISCV::SUBW: |
3991 | case RISCV::MULW: |
3992 | case RISCV::SLLIW: |
3993 | case RISCV::PACKW: |
3994 | case RISCV::TH_MULAW: |
3995 | case RISCV::TH_MULAH: |
3996 | case RISCV::TH_MULSW: |
3997 | case RISCV::TH_MULSH: |
3998 | if (N0.getValueType() == MVT::i32) |
3999 | break; |
4000 | |
4001 | // Result is already sign extended just remove the sext.w. |
4002 | // NOTE: We only handle the nodes that are selected with hasAllWUsers. |
4003 | ReplaceUses(F: N, T: N0.getNode()); |
4004 | return true; |
4005 | } |
4006 | |
4007 | return false; |
4008 | } |
4009 | |
4010 | static bool usesAllOnesMask(SDValue MaskOp) { |
4011 | const auto IsVMSet = [](unsigned Opc) { |
4012 | return Opc == RISCV::PseudoVMSET_M_B1 || Opc == RISCV::PseudoVMSET_M_B16 || |
4013 | Opc == RISCV::PseudoVMSET_M_B2 || Opc == RISCV::PseudoVMSET_M_B32 || |
4014 | Opc == RISCV::PseudoVMSET_M_B4 || Opc == RISCV::PseudoVMSET_M_B64 || |
4015 | Opc == RISCV::PseudoVMSET_M_B8; |
4016 | }; |
4017 | |
4018 | // TODO: Check that the VMSET is the expected bitwidth? The pseudo has |
4019 | // undefined behaviour if it's the wrong bitwidth, so we could choose to |
4020 | // assume that it's all-ones? Same applies to its VL. |
4021 | return MaskOp->isMachineOpcode() && IsVMSet(MaskOp.getMachineOpcode()); |
4022 | } |
4023 | |
4024 | static bool isImplicitDef(SDValue V) { |
4025 | if (!V.isMachineOpcode()) |
4026 | return false; |
4027 | if (V.getMachineOpcode() == TargetOpcode::REG_SEQUENCE) { |
4028 | for (unsigned I = 1; I < V.getNumOperands(); I += 2) |
4029 | if (!isImplicitDef(V: V.getOperand(i: I))) |
4030 | return false; |
4031 | return true; |
4032 | } |
4033 | return V.getMachineOpcode() == TargetOpcode::IMPLICIT_DEF; |
4034 | } |
4035 | |
4036 | // Optimize masked RVV pseudo instructions with a known all-ones mask to their |
4037 | // corresponding "unmasked" pseudo versions. |
4038 | bool RISCVDAGToDAGISel::doPeepholeMaskedRVV(MachineSDNode *N) { |
4039 | const RISCV::RISCVMaskedPseudoInfo *I = |
4040 | RISCV::getMaskedPseudoInfo(MaskedPseudo: N->getMachineOpcode()); |
4041 | if (!I) |
4042 | return false; |
4043 | |
4044 | unsigned MaskOpIdx = I->MaskOpIdx; |
4045 | if (!usesAllOnesMask(MaskOp: N->getOperand(Num: MaskOpIdx))) |
4046 | return false; |
4047 | |
4048 | // There are two classes of pseudos in the table - compares and |
4049 | // everything else. See the comment on RISCVMaskedPseudo for details. |
4050 | const unsigned Opc = I->UnmaskedPseudo; |
4051 | const MCInstrDesc &MCID = TII->get(Opcode: Opc); |
4052 | const bool HasPassthru = RISCVII::isFirstDefTiedToFirstUse(Desc: MCID); |
4053 | |
4054 | const MCInstrDesc &MaskedMCID = TII->get(Opcode: N->getMachineOpcode()); |
4055 | const bool MaskedHasPassthru = RISCVII::isFirstDefTiedToFirstUse(Desc: MaskedMCID); |
4056 | |
4057 | assert((RISCVII::hasVecPolicyOp(MaskedMCID.TSFlags) || |
4058 | !RISCVII::hasVecPolicyOp(MCID.TSFlags)) && |
4059 | "Unmasked pseudo has policy but masked pseudo doesn't?" ); |
4060 | assert(RISCVII::hasVecPolicyOp(MCID.TSFlags) == HasPassthru && |
4061 | "Unexpected pseudo structure" ); |
4062 | assert(!(HasPassthru && !MaskedHasPassthru) && |
4063 | "Unmasked pseudo has passthru but masked pseudo doesn't?" ); |
4064 | |
4065 | SmallVector<SDValue, 8> Ops; |
4066 | // Skip the passthru operand at index 0 if the unmasked don't have one. |
4067 | bool ShouldSkip = !HasPassthru && MaskedHasPassthru; |
4068 | bool DropPolicy = !RISCVII::hasVecPolicyOp(TSFlags: MCID.TSFlags) && |
4069 | RISCVII::hasVecPolicyOp(TSFlags: MaskedMCID.TSFlags); |
4070 | bool HasChainOp = |
4071 | N->getOperand(Num: N->getNumOperands() - 1).getValueType() == MVT::Other; |
4072 | unsigned LastOpNum = N->getNumOperands() - 1 - HasChainOp; |
4073 | for (unsigned I = ShouldSkip, E = N->getNumOperands(); I != E; I++) { |
4074 | // Skip the mask |
4075 | SDValue Op = N->getOperand(Num: I); |
4076 | if (I == MaskOpIdx) |
4077 | continue; |
4078 | if (DropPolicy && I == LastOpNum) |
4079 | continue; |
4080 | Ops.push_back(Elt: Op); |
4081 | } |
4082 | |
4083 | MachineSDNode *Result = |
4084 | CurDAG->getMachineNode(Opcode: Opc, dl: SDLoc(N), VTs: N->getVTList(), Ops); |
4085 | |
4086 | if (!N->memoperands_empty()) |
4087 | CurDAG->setNodeMemRefs(N: Result, NewMemRefs: N->memoperands()); |
4088 | |
4089 | Result->setFlags(N->getFlags()); |
4090 | ReplaceUses(F: N, T: Result); |
4091 | |
4092 | return true; |
4093 | } |
4094 | |
4095 | static bool IsVMerge(SDNode *N) { |
4096 | return RISCV::getRVVMCOpcode(RVVPseudoOpcode: N->getMachineOpcode()) == RISCV::VMERGE_VVM; |
4097 | } |
4098 | |
4099 | // Try to fold away VMERGE_VVM instructions into their true operands: |
4100 | // |
4101 | // %true = PseudoVADD_VV ... |
4102 | // %x = PseudoVMERGE_VVM %false, %false, %true, %mask |
4103 | // -> |
4104 | // %x = PseudoVADD_VV_MASK %false, ..., %mask |
4105 | // |
4106 | // We can only fold if vmerge's passthru operand, vmerge's false operand and |
4107 | // %true's passthru operand (if it has one) are the same. This is because we |
4108 | // have to consolidate them into one passthru operand in the result. |
4109 | // |
4110 | // If %true is masked, then we can use its mask instead of vmerge's if vmerge's |
4111 | // mask is all ones. |
4112 | // |
4113 | // The resulting VL is the minimum of the two VLs. |
4114 | // |
4115 | // The resulting policy is the effective policy the vmerge would have had, |
4116 | // i.e. whether or not it's passthru operand was implicit-def. |
4117 | bool RISCVDAGToDAGISel::performCombineVMergeAndVOps(SDNode *N) { |
4118 | SDValue Passthru, False, True, VL, Mask; |
4119 | assert(IsVMerge(N)); |
4120 | Passthru = N->getOperand(Num: 0); |
4121 | False = N->getOperand(Num: 1); |
4122 | True = N->getOperand(Num: 2); |
4123 | Mask = N->getOperand(Num: 3); |
4124 | VL = N->getOperand(Num: 4); |
4125 | |
4126 | // If the EEW of True is different from vmerge's SEW, then we can't fold. |
4127 | if (True.getSimpleValueType() != N->getSimpleValueType(ResNo: 0)) |
4128 | return false; |
4129 | |
4130 | // We require that either passthru and false are the same, or that passthru |
4131 | // is undefined. |
4132 | if (Passthru != False && !isImplicitDef(V: Passthru)) |
4133 | return false; |
4134 | |
4135 | assert(True.getResNo() == 0 && |
4136 | "Expect True is the first output of an instruction." ); |
4137 | |
4138 | // Need N is the exactly one using True. |
4139 | if (!True.hasOneUse()) |
4140 | return false; |
4141 | |
4142 | if (!True.isMachineOpcode()) |
4143 | return false; |
4144 | |
4145 | unsigned TrueOpc = True.getMachineOpcode(); |
4146 | const MCInstrDesc &TrueMCID = TII->get(Opcode: TrueOpc); |
4147 | uint64_t TrueTSFlags = TrueMCID.TSFlags; |
4148 | bool HasTiedDest = RISCVII::isFirstDefTiedToFirstUse(Desc: TrueMCID); |
4149 | |
4150 | const RISCV::RISCVMaskedPseudoInfo *Info = |
4151 | RISCV::lookupMaskedIntrinsicByUnmasked(UnmaskedPseudo: TrueOpc); |
4152 | if (!Info) |
4153 | return false; |
4154 | |
4155 | // If True has a passthru operand then it needs to be the same as vmerge's |
4156 | // False, since False will be used for the result's passthru operand. |
4157 | if (HasTiedDest && !isImplicitDef(V: True->getOperand(Num: 0))) { |
4158 | SDValue PassthruOpTrue = True->getOperand(Num: 0); |
4159 | if (False != PassthruOpTrue) |
4160 | return false; |
4161 | } |
4162 | |
4163 | // Skip if True has side effect. |
4164 | if (TII->get(Opcode: TrueOpc).hasUnmodeledSideEffects()) |
4165 | return false; |
4166 | |
4167 | unsigned TrueChainOpIdx = True.getNumOperands() - 1; |
4168 | bool HasChainOp = |
4169 | True.getOperand(i: TrueChainOpIdx).getValueType() == MVT::Other; |
4170 | |
4171 | if (HasChainOp) { |
4172 | // Avoid creating cycles in the DAG. We must ensure that none of the other |
4173 | // operands depend on True through it's Chain. |
4174 | SmallVector<const SDNode *, 4> LoopWorklist; |
4175 | SmallPtrSet<const SDNode *, 16> Visited; |
4176 | LoopWorklist.push_back(Elt: False.getNode()); |
4177 | LoopWorklist.push_back(Elt: Mask.getNode()); |
4178 | LoopWorklist.push_back(Elt: VL.getNode()); |
4179 | if (SDNode::hasPredecessorHelper(N: True.getNode(), Visited, Worklist&: LoopWorklist)) |
4180 | return false; |
4181 | } |
4182 | |
4183 | // The vector policy operand may be present for masked intrinsics |
4184 | bool HasVecPolicyOp = RISCVII::hasVecPolicyOp(TSFlags: TrueTSFlags); |
4185 | unsigned TrueVLIndex = |
4186 | True.getNumOperands() - HasVecPolicyOp - HasChainOp - 2; |
4187 | SDValue TrueVL = True.getOperand(i: TrueVLIndex); |
4188 | SDValue SEW = True.getOperand(i: TrueVLIndex + 1); |
4189 | |
4190 | auto GetMinVL = [](SDValue LHS, SDValue RHS) { |
4191 | if (LHS == RHS) |
4192 | return LHS; |
4193 | if (isAllOnesConstant(V: LHS)) |
4194 | return RHS; |
4195 | if (isAllOnesConstant(V: RHS)) |
4196 | return LHS; |
4197 | auto *CLHS = dyn_cast<ConstantSDNode>(Val&: LHS); |
4198 | auto *CRHS = dyn_cast<ConstantSDNode>(Val&: RHS); |
4199 | if (!CLHS || !CRHS) |
4200 | return SDValue(); |
4201 | return CLHS->getZExtValue() <= CRHS->getZExtValue() ? LHS : RHS; |
4202 | }; |
4203 | |
4204 | // Because N and True must have the same passthru operand (or True's operand |
4205 | // is implicit_def), the "effective" body is the minimum of their VLs. |
4206 | SDValue OrigVL = VL; |
4207 | VL = GetMinVL(TrueVL, VL); |
4208 | if (!VL) |
4209 | return false; |
4210 | |
4211 | // Some operations produce different elementwise results depending on the |
4212 | // active elements, like viota.m or vredsum. This transformation is illegal |
4213 | // for these if we change the active elements (i.e. mask or VL). |
4214 | const MCInstrDesc &TrueBaseMCID = TII->get(Opcode: RISCV::getRVVMCOpcode(RVVPseudoOpcode: TrueOpc)); |
4215 | if (RISCVII::elementsDependOnVL(TSFlags: TrueBaseMCID.TSFlags) && (TrueVL != VL)) |
4216 | return false; |
4217 | if (RISCVII::elementsDependOnMask(TSFlags: TrueBaseMCID.TSFlags) && |
4218 | (Mask && !usesAllOnesMask(MaskOp: Mask))) |
4219 | return false; |
4220 | |
4221 | // Make sure it doesn't raise any observable fp exceptions, since changing the |
4222 | // active elements will affect how fflags is set. |
4223 | if (mayRaiseFPException(Node: True.getNode()) && !True->getFlags().hasNoFPExcept()) |
4224 | return false; |
4225 | |
4226 | SDLoc DL(N); |
4227 | |
4228 | unsigned MaskedOpc = Info->MaskedPseudo; |
4229 | #ifndef NDEBUG |
4230 | const MCInstrDesc &MaskedMCID = TII->get(MaskedOpc); |
4231 | assert(RISCVII::hasVecPolicyOp(MaskedMCID.TSFlags) && |
4232 | "Expected instructions with mask have policy operand." ); |
4233 | assert(MaskedMCID.getOperandConstraint(MaskedMCID.getNumDefs(), |
4234 | MCOI::TIED_TO) == 0 && |
4235 | "Expected instructions with mask have a tied dest." ); |
4236 | #endif |
4237 | |
4238 | // Use a tumu policy, relaxing it to tail agnostic provided that the passthru |
4239 | // operand is undefined. |
4240 | // |
4241 | // However, if the VL became smaller than what the vmerge had originally, then |
4242 | // elements past VL that were previously in the vmerge's body will have moved |
4243 | // to the tail. In that case we always need to use tail undisturbed to |
4244 | // preserve them. |
4245 | bool MergeVLShrunk = VL != OrigVL; |
4246 | uint64_t Policy = (isImplicitDef(V: Passthru) && !MergeVLShrunk) |
4247 | ? RISCVVType::TAIL_AGNOSTIC |
4248 | : /*TUMU*/ 0; |
4249 | SDValue PolicyOp = |
4250 | CurDAG->getTargetConstant(Val: Policy, DL, VT: Subtarget->getXLenVT()); |
4251 | |
4252 | |
4253 | SmallVector<SDValue, 8> Ops; |
4254 | Ops.push_back(Elt: False); |
4255 | |
4256 | const bool HasRoundingMode = RISCVII::hasRoundModeOp(TSFlags: TrueTSFlags); |
4257 | const unsigned NormalOpsEnd = TrueVLIndex - HasRoundingMode; |
4258 | Ops.append(in_start: True->op_begin() + HasTiedDest, in_end: True->op_begin() + NormalOpsEnd); |
4259 | |
4260 | Ops.push_back(Elt: Mask); |
4261 | |
4262 | // For unmasked "VOp" with rounding mode operand, that is interfaces like |
4263 | // (..., rm, vl) or (..., rm, vl, policy). |
4264 | // Its masked version is (..., vm, rm, vl, policy). |
4265 | // Check the rounding mode pseudo nodes under RISCVInstrInfoVPseudos.td |
4266 | if (HasRoundingMode) |
4267 | Ops.push_back(Elt: True->getOperand(Num: TrueVLIndex - 1)); |
4268 | |
4269 | Ops.append(IL: {VL, SEW, PolicyOp}); |
4270 | |
4271 | // Result node should have chain operand of True. |
4272 | if (HasChainOp) |
4273 | Ops.push_back(Elt: True.getOperand(i: TrueChainOpIdx)); |
4274 | |
4275 | MachineSDNode *Result = |
4276 | CurDAG->getMachineNode(Opcode: MaskedOpc, dl: DL, VTs: True->getVTList(), Ops); |
4277 | Result->setFlags(True->getFlags()); |
4278 | |
4279 | if (!cast<MachineSDNode>(Val&: True)->memoperands_empty()) |
4280 | CurDAG->setNodeMemRefs(N: Result, NewMemRefs: cast<MachineSDNode>(Val&: True)->memoperands()); |
4281 | |
4282 | // Replace vmerge.vvm node by Result. |
4283 | ReplaceUses(F: SDValue(N, 0), T: SDValue(Result, 0)); |
4284 | |
4285 | // Replace another value of True. E.g. chain and VL. |
4286 | for (unsigned Idx = 1; Idx < True->getNumValues(); ++Idx) |
4287 | ReplaceUses(F: True.getValue(R: Idx), T: SDValue(Result, Idx)); |
4288 | |
4289 | return true; |
4290 | } |
4291 | |
4292 | bool RISCVDAGToDAGISel::doPeepholeMergeVVMFold() { |
4293 | bool MadeChange = false; |
4294 | SelectionDAG::allnodes_iterator Position = CurDAG->allnodes_end(); |
4295 | |
4296 | while (Position != CurDAG->allnodes_begin()) { |
4297 | SDNode *N = &*--Position; |
4298 | if (N->use_empty() || !N->isMachineOpcode()) |
4299 | continue; |
4300 | |
4301 | if (IsVMerge(N)) |
4302 | MadeChange |= performCombineVMergeAndVOps(N); |
4303 | } |
4304 | return MadeChange; |
4305 | } |
4306 | |
4307 | /// If our passthru is an implicit_def, use noreg instead. This side |
4308 | /// steps issues with MachineCSE not being able to CSE expressions with |
4309 | /// IMPLICIT_DEF operands while preserving the semantic intent. See |
4310 | /// pr64282 for context. Note that this transform is the last one |
4311 | /// performed at ISEL DAG to DAG. |
4312 | bool RISCVDAGToDAGISel::doPeepholeNoRegPassThru() { |
4313 | bool MadeChange = false; |
4314 | SelectionDAG::allnodes_iterator Position = CurDAG->allnodes_end(); |
4315 | |
4316 | while (Position != CurDAG->allnodes_begin()) { |
4317 | SDNode *N = &*--Position; |
4318 | if (N->use_empty() || !N->isMachineOpcode()) |
4319 | continue; |
4320 | |
4321 | const unsigned Opc = N->getMachineOpcode(); |
4322 | if (!RISCVVPseudosTable::getPseudoInfo(Pseudo: Opc) || |
4323 | !RISCVII::isFirstDefTiedToFirstUse(Desc: TII->get(Opcode: Opc)) || |
4324 | !isImplicitDef(V: N->getOperand(Num: 0))) |
4325 | continue; |
4326 | |
4327 | SmallVector<SDValue> Ops; |
4328 | Ops.push_back(Elt: CurDAG->getRegister(Reg: RISCV::NoRegister, VT: N->getValueType(ResNo: 0))); |
4329 | for (unsigned I = 1, E = N->getNumOperands(); I != E; I++) { |
4330 | SDValue Op = N->getOperand(Num: I); |
4331 | Ops.push_back(Elt: Op); |
4332 | } |
4333 | |
4334 | MachineSDNode *Result = |
4335 | CurDAG->getMachineNode(Opcode: Opc, dl: SDLoc(N), VTs: N->getVTList(), Ops); |
4336 | Result->setFlags(N->getFlags()); |
4337 | CurDAG->setNodeMemRefs(N: Result, NewMemRefs: cast<MachineSDNode>(Val: N)->memoperands()); |
4338 | ReplaceUses(F: N, T: Result); |
4339 | MadeChange = true; |
4340 | } |
4341 | return MadeChange; |
4342 | } |
4343 | |
4344 | |
4345 | // This pass converts a legalized DAG into a RISCV-specific DAG, ready |
4346 | // for instruction scheduling. |
4347 | FunctionPass *llvm::createRISCVISelDag(RISCVTargetMachine &TM, |
4348 | CodeGenOptLevel OptLevel) { |
4349 | return new RISCVDAGToDAGISelLegacy(TM, OptLevel); |
4350 | } |
4351 | |
4352 | char RISCVDAGToDAGISelLegacy::ID = 0; |
4353 | |
4354 | RISCVDAGToDAGISelLegacy::RISCVDAGToDAGISelLegacy(RISCVTargetMachine &TM, |
4355 | CodeGenOptLevel OptLevel) |
4356 | : SelectionDAGISelLegacy( |
4357 | ID, std::make_unique<RISCVDAGToDAGISel>(args&: TM, args&: OptLevel)) {} |
4358 | |
4359 | INITIALIZE_PASS(RISCVDAGToDAGISelLegacy, DEBUG_TYPE, PASS_NAME, false, false) |
4360 | |