1 | //===-- XCoreISelLowering.cpp - XCore DAG Lowering Implementation ---------===// |
2 | // |
3 | // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. |
4 | // See https://llvm.org/LICENSE.txt for license information. |
5 | // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception |
6 | // |
7 | //===----------------------------------------------------------------------===// |
8 | // |
9 | // This file implements the XCoreTargetLowering class. |
10 | // |
11 | //===----------------------------------------------------------------------===// |
12 | |
13 | #include "XCoreISelLowering.h" |
14 | #include "XCore.h" |
15 | #include "XCoreMachineFunctionInfo.h" |
16 | #include "XCoreSubtarget.h" |
17 | #include "XCoreTargetMachine.h" |
18 | #include "XCoreTargetObjectFile.h" |
19 | #include "llvm/CodeGen/CallingConvLower.h" |
20 | #include "llvm/CodeGen/MachineFrameInfo.h" |
21 | #include "llvm/CodeGen/MachineFunction.h" |
22 | #include "llvm/CodeGen/MachineInstrBuilder.h" |
23 | #include "llvm/CodeGen/MachineJumpTableInfo.h" |
24 | #include "llvm/CodeGen/MachineRegisterInfo.h" |
25 | #include "llvm/CodeGen/ValueTypes.h" |
26 | #include "llvm/IR/CallingConv.h" |
27 | #include "llvm/IR/Constants.h" |
28 | #include "llvm/IR/DerivedTypes.h" |
29 | #include "llvm/IR/Function.h" |
30 | #include "llvm/IR/GlobalVariable.h" |
31 | #include "llvm/IR/Intrinsics.h" |
32 | #include "llvm/IR/IntrinsicsXCore.h" |
33 | #include "llvm/Support/Debug.h" |
34 | #include "llvm/Support/ErrorHandling.h" |
35 | #include "llvm/Support/KnownBits.h" |
36 | #include "llvm/Support/raw_ostream.h" |
37 | #include <algorithm> |
38 | |
39 | using namespace llvm; |
40 | |
41 | #define DEBUG_TYPE "xcore-lower" |
42 | |
43 | XCoreTargetLowering::XCoreTargetLowering(const TargetMachine &TM, |
44 | const XCoreSubtarget &Subtarget) |
45 | : TargetLowering(TM), TM(TM), Subtarget(Subtarget) { |
46 | |
47 | // Set up the register classes. |
48 | addRegisterClass(VT: MVT::i32, RC: &XCore::GRRegsRegClass); |
49 | |
50 | // Compute derived properties from the register classes |
51 | computeRegisterProperties(TRI: Subtarget.getRegisterInfo()); |
52 | |
53 | setStackPointerRegisterToSaveRestore(XCore::SP); |
54 | |
55 | setSchedulingPreference(Sched::Source); |
56 | |
57 | // Use i32 for setcc operations results (slt, sgt, ...). |
58 | setBooleanContents(ZeroOrOneBooleanContent); |
59 | setBooleanVectorContents(ZeroOrOneBooleanContent); // FIXME: Is this correct? |
60 | |
61 | // XCore does not have the NodeTypes below. |
62 | setOperationAction(Op: ISD::BR_CC, VT: MVT::i32, Action: Expand); |
63 | setOperationAction(Op: ISD::SELECT_CC, VT: MVT::i32, Action: Expand); |
64 | |
65 | // 64bit |
66 | setOperationAction(Op: ISD::ADD, VT: MVT::i64, Action: Custom); |
67 | setOperationAction(Op: ISD::SUB, VT: MVT::i64, Action: Custom); |
68 | setOperationAction(Op: ISD::SMUL_LOHI, VT: MVT::i32, Action: Custom); |
69 | setOperationAction(Op: ISD::UMUL_LOHI, VT: MVT::i32, Action: Custom); |
70 | setOperationAction(Op: ISD::MULHS, VT: MVT::i32, Action: Expand); |
71 | setOperationAction(Op: ISD::MULHU, VT: MVT::i32, Action: Expand); |
72 | setOperationAction(Op: ISD::SHL_PARTS, VT: MVT::i32, Action: Expand); |
73 | setOperationAction(Op: ISD::SRA_PARTS, VT: MVT::i32, Action: Expand); |
74 | setOperationAction(Op: ISD::SRL_PARTS, VT: MVT::i32, Action: Expand); |
75 | |
76 | // Bit Manipulation |
77 | setOperationAction(Op: ISD::CTPOP, VT: MVT::i32, Action: Expand); |
78 | setOperationAction(Op: ISD::ROTL , VT: MVT::i32, Action: Expand); |
79 | setOperationAction(Op: ISD::ROTR , VT: MVT::i32, Action: Expand); |
80 | setOperationAction(Op: ISD::BITREVERSE , VT: MVT::i32, Action: Legal); |
81 | |
82 | setOperationAction(Op: ISD::TRAP, VT: MVT::Other, Action: Legal); |
83 | |
84 | // Jump tables. |
85 | setOperationAction(Op: ISD::BR_JT, VT: MVT::Other, Action: Custom); |
86 | |
87 | setOperationAction(Op: ISD::GlobalAddress, VT: MVT::i32, Action: Custom); |
88 | setOperationAction(Op: ISD::BlockAddress, VT: MVT::i32 , Action: Custom); |
89 | |
90 | // Conversion of i64 -> double produces constantpool nodes |
91 | setOperationAction(Op: ISD::ConstantPool, VT: MVT::i32, Action: Custom); |
92 | |
93 | // Loads |
94 | for (MVT VT : MVT::integer_valuetypes()) { |
95 | setLoadExtAction(ExtType: ISD::EXTLOAD, ValVT: VT, MemVT: MVT::i1, Action: Promote); |
96 | setLoadExtAction(ExtType: ISD::ZEXTLOAD, ValVT: VT, MemVT: MVT::i1, Action: Promote); |
97 | setLoadExtAction(ExtType: ISD::SEXTLOAD, ValVT: VT, MemVT: MVT::i1, Action: Promote); |
98 | |
99 | setLoadExtAction(ExtType: ISD::SEXTLOAD, ValVT: VT, MemVT: MVT::i8, Action: Expand); |
100 | setLoadExtAction(ExtType: ISD::ZEXTLOAD, ValVT: VT, MemVT: MVT::i16, Action: Expand); |
101 | } |
102 | |
103 | // Custom expand misaligned loads / stores. |
104 | setOperationAction(Op: ISD::LOAD, VT: MVT::i32, Action: Custom); |
105 | setOperationAction(Op: ISD::STORE, VT: MVT::i32, Action: Custom); |
106 | |
107 | // Varargs |
108 | setOperationAction(Op: ISD::VAEND, VT: MVT::Other, Action: Expand); |
109 | setOperationAction(Op: ISD::VACOPY, VT: MVT::Other, Action: Expand); |
110 | setOperationAction(Op: ISD::VAARG, VT: MVT::Other, Action: Custom); |
111 | setOperationAction(Op: ISD::VASTART, VT: MVT::Other, Action: Custom); |
112 | |
113 | // Dynamic stack |
114 | setOperationAction(Op: ISD::STACKSAVE, VT: MVT::Other, Action: Expand); |
115 | setOperationAction(Op: ISD::STACKRESTORE, VT: MVT::Other, Action: Expand); |
116 | setOperationAction(Op: ISD::DYNAMIC_STACKALLOC, VT: MVT::i32, Action: Expand); |
117 | |
118 | // Exception handling |
119 | setOperationAction(Op: ISD::EH_RETURN, VT: MVT::Other, Action: Custom); |
120 | setOperationAction(Op: ISD::FRAME_TO_ARGS_OFFSET, VT: MVT::i32, Action: Custom); |
121 | |
122 | setOperationAction(Op: ISD::ATOMIC_FENCE, VT: MVT::Other, Action: Custom); |
123 | |
124 | // TRAMPOLINE is custom lowered. |
125 | setOperationAction(Op: ISD::INIT_TRAMPOLINE, VT: MVT::Other, Action: Custom); |
126 | setOperationAction(Op: ISD::ADJUST_TRAMPOLINE, VT: MVT::Other, Action: Custom); |
127 | |
128 | // We want to custom lower some of our intrinsics. |
129 | setOperationAction(Op: ISD::INTRINSIC_WO_CHAIN, VT: MVT::Other, Action: Custom); |
130 | |
131 | MaxStoresPerMemset = MaxStoresPerMemsetOptSize = 4; |
132 | MaxStoresPerMemmove = MaxStoresPerMemmoveOptSize |
133 | = MaxStoresPerMemcpy = MaxStoresPerMemcpyOptSize = 2; |
134 | |
135 | // We have target-specific dag combine patterns for the following nodes: |
136 | setTargetDAGCombine( |
137 | {ISD::STORE, ISD::ADD, ISD::INTRINSIC_VOID, ISD::INTRINSIC_W_CHAIN}); |
138 | |
139 | setMinFunctionAlignment(Align(2)); |
140 | setPrefFunctionAlignment(Align(4)); |
141 | |
142 | // This target doesn't implement native atomics. |
143 | setMaxAtomicSizeInBitsSupported(0); |
144 | } |
145 | |
146 | bool XCoreTargetLowering::isZExtFree(SDValue Val, EVT VT2) const { |
147 | if (Val.getOpcode() != ISD::LOAD) |
148 | return false; |
149 | |
150 | EVT VT1 = Val.getValueType(); |
151 | if (!VT1.isSimple() || !VT1.isInteger() || |
152 | !VT2.isSimple() || !VT2.isInteger()) |
153 | return false; |
154 | |
155 | switch (VT1.getSimpleVT().SimpleTy) { |
156 | default: break; |
157 | case MVT::i8: |
158 | return true; |
159 | } |
160 | |
161 | return false; |
162 | } |
163 | |
164 | SDValue XCoreTargetLowering:: |
165 | LowerOperation(SDValue Op, SelectionDAG &DAG) const { |
166 | switch (Op.getOpcode()) |
167 | { |
168 | case ISD::EH_RETURN: return LowerEH_RETURN(Op, DAG); |
169 | case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG); |
170 | case ISD::BlockAddress: return LowerBlockAddress(Op, DAG); |
171 | case ISD::ConstantPool: return LowerConstantPool(Op, DAG); |
172 | case ISD::BR_JT: return LowerBR_JT(Op, DAG); |
173 | case ISD::LOAD: return LowerLOAD(Op, DAG); |
174 | case ISD::STORE: return LowerSTORE(Op, DAG); |
175 | case ISD::VAARG: return LowerVAARG(Op, DAG); |
176 | case ISD::VASTART: return LowerVASTART(Op, DAG); |
177 | case ISD::SMUL_LOHI: return LowerSMUL_LOHI(Op, DAG); |
178 | case ISD::UMUL_LOHI: return LowerUMUL_LOHI(Op, DAG); |
179 | // FIXME: Remove these when LegalizeDAGTypes lands. |
180 | case ISD::ADD: |
181 | case ISD::SUB: return ExpandADDSUB(Op: Op.getNode(), DAG); |
182 | case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG); |
183 | case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG); |
184 | case ISD::FRAME_TO_ARGS_OFFSET: return LowerFRAME_TO_ARGS_OFFSET(Op, DAG); |
185 | case ISD::INIT_TRAMPOLINE: return LowerINIT_TRAMPOLINE(Op, DAG); |
186 | case ISD::ADJUST_TRAMPOLINE: return LowerADJUST_TRAMPOLINE(Op, DAG); |
187 | case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG); |
188 | case ISD::ATOMIC_FENCE: |
189 | return LowerATOMIC_FENCE(Op, DAG); |
190 | default: |
191 | llvm_unreachable("unimplemented operand" ); |
192 | } |
193 | } |
194 | |
195 | /// ReplaceNodeResults - Replace the results of node with an illegal result |
196 | /// type with new values built out of custom code. |
197 | void XCoreTargetLowering::ReplaceNodeResults(SDNode *N, |
198 | SmallVectorImpl<SDValue>&Results, |
199 | SelectionDAG &DAG) const { |
200 | switch (N->getOpcode()) { |
201 | default: |
202 | llvm_unreachable("Don't know how to custom expand this!" ); |
203 | case ISD::ADD: |
204 | case ISD::SUB: |
205 | Results.push_back(Elt: ExpandADDSUB(Op: N, DAG)); |
206 | return; |
207 | } |
208 | } |
209 | |
210 | //===----------------------------------------------------------------------===// |
211 | // Misc Lower Operation implementation |
212 | //===----------------------------------------------------------------------===// |
213 | |
214 | SDValue XCoreTargetLowering::getGlobalAddressWrapper(SDValue GA, |
215 | const GlobalValue *GV, |
216 | SelectionDAG &DAG) const { |
217 | // FIXME there is no actual debug info here |
218 | SDLoc dl(GA); |
219 | |
220 | if (GV->getValueType()->isFunctionTy()) |
221 | return DAG.getNode(Opcode: XCoreISD::PCRelativeWrapper, DL: dl, VT: MVT::i32, Operand: GA); |
222 | |
223 | const auto *GVar = dyn_cast<GlobalVariable>(Val: GV); |
224 | if ((GV->hasSection() && GV->getSection().starts_with(Prefix: ".cp." )) || |
225 | (GVar && GVar->isConstant() && GV->hasLocalLinkage())) |
226 | return DAG.getNode(Opcode: XCoreISD::CPRelativeWrapper, DL: dl, VT: MVT::i32, Operand: GA); |
227 | |
228 | return DAG.getNode(Opcode: XCoreISD::DPRelativeWrapper, DL: dl, VT: MVT::i32, Operand: GA); |
229 | } |
230 | |
231 | static bool IsSmallObject(const GlobalValue *GV, const XCoreTargetLowering &XTL) { |
232 | if (XTL.getTargetMachine().getCodeModel() == CodeModel::Small) |
233 | return true; |
234 | |
235 | Type *ObjType = GV->getValueType(); |
236 | if (!ObjType->isSized()) |
237 | return false; |
238 | |
239 | auto &DL = GV->getDataLayout(); |
240 | unsigned ObjSize = DL.getTypeAllocSize(Ty: ObjType); |
241 | return ObjSize < CodeModelLargeSize && ObjSize != 0; |
242 | } |
243 | |
244 | SDValue XCoreTargetLowering:: |
245 | LowerGlobalAddress(SDValue Op, SelectionDAG &DAG) const |
246 | { |
247 | const GlobalAddressSDNode *GN = cast<GlobalAddressSDNode>(Val&: Op); |
248 | const GlobalValue *GV = GN->getGlobal(); |
249 | SDLoc DL(GN); |
250 | int64_t Offset = GN->getOffset(); |
251 | if (IsSmallObject(GV, XTL: *this)) { |
252 | // We can only fold positive offsets that are a multiple of the word size. |
253 | int64_t FoldedOffset = std::max(a: Offset & ~3, b: (int64_t)0); |
254 | SDValue GA = DAG.getTargetGlobalAddress(GV, DL, VT: MVT::i32, offset: FoldedOffset); |
255 | GA = getGlobalAddressWrapper(GA, GV, DAG); |
256 | // Handle the rest of the offset. |
257 | if (Offset != FoldedOffset) { |
258 | SDValue Remaining = |
259 | DAG.getSignedConstant(Val: Offset - FoldedOffset, DL, VT: MVT::i32); |
260 | GA = DAG.getNode(Opcode: ISD::ADD, DL, VT: MVT::i32, N1: GA, N2: Remaining); |
261 | } |
262 | return GA; |
263 | } else { |
264 | // Ideally we would not fold in offset with an index <= 11. |
265 | Type *Ty = Type::getInt32Ty(C&: *DAG.getContext()); |
266 | Constant *Idx = ConstantInt::get(Ty, V: Offset); |
267 | Constant *GAI = ConstantExpr::getGetElementPtr( |
268 | Ty: Type::getInt8Ty(C&: *DAG.getContext()), C: const_cast<GlobalValue *>(GV), Idx); |
269 | SDValue CP = DAG.getConstantPool(C: GAI, VT: MVT::i32); |
270 | return DAG.getLoad(VT: getPointerTy(DL: DAG.getDataLayout()), dl: DL, |
271 | Chain: DAG.getEntryNode(), Ptr: CP, PtrInfo: MachinePointerInfo()); |
272 | } |
273 | } |
274 | |
275 | SDValue XCoreTargetLowering:: |
276 | LowerBlockAddress(SDValue Op, SelectionDAG &DAG) const |
277 | { |
278 | SDLoc DL(Op); |
279 | auto PtrVT = getPointerTy(DL: DAG.getDataLayout()); |
280 | const BlockAddress *BA = cast<BlockAddressSDNode>(Val&: Op)->getBlockAddress(); |
281 | SDValue Result = DAG.getTargetBlockAddress(BA, VT: PtrVT); |
282 | |
283 | return DAG.getNode(Opcode: XCoreISD::PCRelativeWrapper, DL, VT: PtrVT, Operand: Result); |
284 | } |
285 | |
286 | SDValue XCoreTargetLowering:: |
287 | LowerConstantPool(SDValue Op, SelectionDAG &DAG) const |
288 | { |
289 | ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Val&: Op); |
290 | // FIXME there isn't really debug info here |
291 | SDLoc dl(CP); |
292 | EVT PtrVT = Op.getValueType(); |
293 | SDValue Res; |
294 | if (CP->isMachineConstantPoolEntry()) { |
295 | Res = DAG.getTargetConstantPool(C: CP->getMachineCPVal(), VT: PtrVT, |
296 | Align: CP->getAlign(), Offset: CP->getOffset()); |
297 | } else { |
298 | Res = DAG.getTargetConstantPool(C: CP->getConstVal(), VT: PtrVT, Align: CP->getAlign(), |
299 | Offset: CP->getOffset()); |
300 | } |
301 | return DAG.getNode(Opcode: XCoreISD::CPRelativeWrapper, DL: dl, VT: MVT::i32, Operand: Res); |
302 | } |
303 | |
304 | unsigned XCoreTargetLowering::getJumpTableEncoding() const { |
305 | return MachineJumpTableInfo::EK_Inline; |
306 | } |
307 | |
308 | SDValue XCoreTargetLowering:: |
309 | LowerBR_JT(SDValue Op, SelectionDAG &DAG) const |
310 | { |
311 | SDValue Chain = Op.getOperand(i: 0); |
312 | SDValue Table = Op.getOperand(i: 1); |
313 | SDValue Index = Op.getOperand(i: 2); |
314 | SDLoc dl(Op); |
315 | JumpTableSDNode *JT = cast<JumpTableSDNode>(Val&: Table); |
316 | unsigned JTI = JT->getIndex(); |
317 | MachineFunction &MF = DAG.getMachineFunction(); |
318 | const MachineJumpTableInfo *MJTI = MF.getJumpTableInfo(); |
319 | SDValue TargetJT = DAG.getTargetJumpTable(JTI: JT->getIndex(), VT: MVT::i32); |
320 | |
321 | unsigned NumEntries = MJTI->getJumpTables()[JTI].MBBs.size(); |
322 | if (NumEntries <= 32) { |
323 | return DAG.getNode(Opcode: XCoreISD::BR_JT, DL: dl, VT: MVT::Other, N1: Chain, N2: TargetJT, N3: Index); |
324 | } |
325 | assert((NumEntries >> 31) == 0); |
326 | SDValue ScaledIndex = DAG.getNode(Opcode: ISD::SHL, DL: dl, VT: MVT::i32, N1: Index, |
327 | N2: DAG.getConstant(Val: 1, DL: dl, VT: MVT::i32)); |
328 | return DAG.getNode(Opcode: XCoreISD::BR_JT32, DL: dl, VT: MVT::Other, N1: Chain, N2: TargetJT, |
329 | N3: ScaledIndex); |
330 | } |
331 | |
332 | SDValue XCoreTargetLowering::lowerLoadWordFromAlignedBasePlusOffset( |
333 | const SDLoc &DL, SDValue Chain, SDValue Base, int64_t Offset, |
334 | SelectionDAG &DAG) const { |
335 | auto PtrVT = getPointerTy(DL: DAG.getDataLayout()); |
336 | if ((Offset & 0x3) == 0) { |
337 | return DAG.getLoad(VT: PtrVT, dl: DL, Chain, Ptr: Base, PtrInfo: MachinePointerInfo()); |
338 | } |
339 | // Lower to pair of consecutive word aligned loads plus some bit shifting. |
340 | int32_t HighOffset = alignTo(Value: Offset, Align: 4); |
341 | int32_t LowOffset = HighOffset - 4; |
342 | SDValue LowAddr, HighAddr; |
343 | if (GlobalAddressSDNode *GASD = |
344 | dyn_cast<GlobalAddressSDNode>(Val: Base.getNode())) { |
345 | LowAddr = DAG.getGlobalAddress(GV: GASD->getGlobal(), DL, VT: Base.getValueType(), |
346 | offset: LowOffset); |
347 | HighAddr = DAG.getGlobalAddress(GV: GASD->getGlobal(), DL, VT: Base.getValueType(), |
348 | offset: HighOffset); |
349 | } else { |
350 | LowAddr = DAG.getNode(Opcode: ISD::ADD, DL, VT: MVT::i32, N1: Base, |
351 | N2: DAG.getConstant(Val: LowOffset, DL, VT: MVT::i32)); |
352 | HighAddr = DAG.getNode(Opcode: ISD::ADD, DL, VT: MVT::i32, N1: Base, |
353 | N2: DAG.getConstant(Val: HighOffset, DL, VT: MVT::i32)); |
354 | } |
355 | SDValue LowShift = DAG.getConstant(Val: (Offset - LowOffset) * 8, DL, VT: MVT::i32); |
356 | SDValue HighShift = DAG.getConstant(Val: (HighOffset - Offset) * 8, DL, VT: MVT::i32); |
357 | |
358 | SDValue Low = DAG.getLoad(VT: PtrVT, dl: DL, Chain, Ptr: LowAddr, PtrInfo: MachinePointerInfo()); |
359 | SDValue High = DAG.getLoad(VT: PtrVT, dl: DL, Chain, Ptr: HighAddr, PtrInfo: MachinePointerInfo()); |
360 | SDValue LowShifted = DAG.getNode(Opcode: ISD::SRL, DL, VT: MVT::i32, N1: Low, N2: LowShift); |
361 | SDValue HighShifted = DAG.getNode(Opcode: ISD::SHL, DL, VT: MVT::i32, N1: High, N2: HighShift); |
362 | SDValue Result = DAG.getNode(Opcode: ISD::OR, DL, VT: MVT::i32, N1: LowShifted, N2: HighShifted); |
363 | Chain = DAG.getNode(Opcode: ISD::TokenFactor, DL, VT: MVT::Other, N1: Low.getValue(R: 1), |
364 | N2: High.getValue(R: 1)); |
365 | SDValue Ops[] = { Result, Chain }; |
366 | return DAG.getMergeValues(Ops, dl: DL); |
367 | } |
368 | |
369 | static bool isWordAligned(SDValue Value, SelectionDAG &DAG) |
370 | { |
371 | KnownBits Known = DAG.computeKnownBits(Op: Value); |
372 | return Known.countMinTrailingZeros() >= 2; |
373 | } |
374 | |
375 | SDValue XCoreTargetLowering::LowerLOAD(SDValue Op, SelectionDAG &DAG) const { |
376 | const TargetLowering &TLI = DAG.getTargetLoweringInfo(); |
377 | LLVMContext &Context = *DAG.getContext(); |
378 | LoadSDNode *LD = cast<LoadSDNode>(Val&: Op); |
379 | assert(LD->getExtensionType() == ISD::NON_EXTLOAD && |
380 | "Unexpected extension type" ); |
381 | assert(LD->getMemoryVT() == MVT::i32 && "Unexpected load EVT" ); |
382 | |
383 | if (allowsMemoryAccessForAlignment(Context, DL: DAG.getDataLayout(), |
384 | VT: LD->getMemoryVT(), MMO: *LD->getMemOperand())) |
385 | return SDValue(); |
386 | |
387 | SDValue Chain = LD->getChain(); |
388 | SDValue BasePtr = LD->getBasePtr(); |
389 | SDLoc DL(Op); |
390 | |
391 | if (!LD->isVolatile()) { |
392 | const GlobalValue *GV; |
393 | int64_t Offset = 0; |
394 | if (DAG.isBaseWithConstantOffset(Op: BasePtr) && |
395 | isWordAligned(Value: BasePtr->getOperand(Num: 0), DAG)) { |
396 | SDValue NewBasePtr = BasePtr->getOperand(Num: 0); |
397 | Offset = cast<ConstantSDNode>(Val: BasePtr->getOperand(Num: 1))->getSExtValue(); |
398 | return lowerLoadWordFromAlignedBasePlusOffset(DL, Chain, Base: NewBasePtr, |
399 | Offset, DAG); |
400 | } |
401 | if (TLI.isGAPlusOffset(N: BasePtr.getNode(), GA&: GV, Offset) && |
402 | GV->getPointerAlignment(DL: DAG.getDataLayout()) >= 4) { |
403 | SDValue NewBasePtr = DAG.getGlobalAddress(GV, DL, |
404 | VT: BasePtr->getValueType(ResNo: 0)); |
405 | return lowerLoadWordFromAlignedBasePlusOffset(DL, Chain, Base: NewBasePtr, |
406 | Offset, DAG); |
407 | } |
408 | } |
409 | |
410 | if (LD->getAlign() == Align(2)) { |
411 | SDValue Low = DAG.getExtLoad(ExtType: ISD::ZEXTLOAD, dl: DL, VT: MVT::i32, Chain, Ptr: BasePtr, |
412 | PtrInfo: LD->getPointerInfo(), MemVT: MVT::i16, Alignment: Align(2), |
413 | MMOFlags: LD->getMemOperand()->getFlags()); |
414 | SDValue HighAddr = DAG.getNode(Opcode: ISD::ADD, DL, VT: MVT::i32, N1: BasePtr, |
415 | N2: DAG.getConstant(Val: 2, DL, VT: MVT::i32)); |
416 | SDValue High = |
417 | DAG.getExtLoad(ExtType: ISD::EXTLOAD, dl: DL, VT: MVT::i32, Chain, Ptr: HighAddr, |
418 | PtrInfo: LD->getPointerInfo().getWithOffset(O: 2), MemVT: MVT::i16, |
419 | Alignment: Align(2), MMOFlags: LD->getMemOperand()->getFlags()); |
420 | SDValue HighShifted = DAG.getNode(Opcode: ISD::SHL, DL, VT: MVT::i32, N1: High, |
421 | N2: DAG.getConstant(Val: 16, DL, VT: MVT::i32)); |
422 | SDValue Result = DAG.getNode(Opcode: ISD::OR, DL, VT: MVT::i32, N1: Low, N2: HighShifted); |
423 | Chain = DAG.getNode(Opcode: ISD::TokenFactor, DL, VT: MVT::Other, N1: Low.getValue(R: 1), |
424 | N2: High.getValue(R: 1)); |
425 | SDValue Ops[] = { Result, Chain }; |
426 | return DAG.getMergeValues(Ops, dl: DL); |
427 | } |
428 | |
429 | // Lower to a call to __misaligned_load(BasePtr). |
430 | Type *IntPtrTy = DAG.getDataLayout().getIntPtrType(C&: Context); |
431 | TargetLowering::ArgListTy Args; |
432 | TargetLowering::ArgListEntry Entry; |
433 | |
434 | Entry.Ty = IntPtrTy; |
435 | Entry.Node = BasePtr; |
436 | Args.push_back(x: Entry); |
437 | |
438 | TargetLowering::CallLoweringInfo CLI(DAG); |
439 | CLI.setDebugLoc(DL).setChain(Chain).setLibCallee( |
440 | CC: CallingConv::C, ResultType: IntPtrTy, |
441 | Target: DAG.getExternalSymbol(Sym: "__misaligned_load" , |
442 | VT: getPointerTy(DL: DAG.getDataLayout())), |
443 | ArgsList: std::move(Args)); |
444 | |
445 | std::pair<SDValue, SDValue> CallResult = LowerCallTo(CLI); |
446 | SDValue Ops[] = { CallResult.first, CallResult.second }; |
447 | return DAG.getMergeValues(Ops, dl: DL); |
448 | } |
449 | |
450 | SDValue XCoreTargetLowering::LowerSTORE(SDValue Op, SelectionDAG &DAG) const { |
451 | LLVMContext &Context = *DAG.getContext(); |
452 | StoreSDNode *ST = cast<StoreSDNode>(Val&: Op); |
453 | assert(!ST->isTruncatingStore() && "Unexpected store type" ); |
454 | assert(ST->getMemoryVT() == MVT::i32 && "Unexpected store EVT" ); |
455 | |
456 | if (allowsMemoryAccessForAlignment(Context, DL: DAG.getDataLayout(), |
457 | VT: ST->getMemoryVT(), MMO: *ST->getMemOperand())) |
458 | return SDValue(); |
459 | |
460 | SDValue Chain = ST->getChain(); |
461 | SDValue BasePtr = ST->getBasePtr(); |
462 | SDValue Value = ST->getValue(); |
463 | SDLoc dl(Op); |
464 | |
465 | if (ST->getAlign() == Align(2)) { |
466 | SDValue Low = Value; |
467 | SDValue High = DAG.getNode(Opcode: ISD::SRL, DL: dl, VT: MVT::i32, N1: Value, |
468 | N2: DAG.getConstant(Val: 16, DL: dl, VT: MVT::i32)); |
469 | SDValue StoreLow = |
470 | DAG.getTruncStore(Chain, dl, Val: Low, Ptr: BasePtr, PtrInfo: ST->getPointerInfo(), |
471 | SVT: MVT::i16, Alignment: Align(2), MMOFlags: ST->getMemOperand()->getFlags()); |
472 | SDValue HighAddr = DAG.getNode(Opcode: ISD::ADD, DL: dl, VT: MVT::i32, N1: BasePtr, |
473 | N2: DAG.getConstant(Val: 2, DL: dl, VT: MVT::i32)); |
474 | SDValue StoreHigh = DAG.getTruncStore( |
475 | Chain, dl, Val: High, Ptr: HighAddr, PtrInfo: ST->getPointerInfo().getWithOffset(O: 2), |
476 | SVT: MVT::i16, Alignment: Align(2), MMOFlags: ST->getMemOperand()->getFlags()); |
477 | return DAG.getNode(Opcode: ISD::TokenFactor, DL: dl, VT: MVT::Other, N1: StoreLow, N2: StoreHigh); |
478 | } |
479 | |
480 | // Lower to a call to __misaligned_store(BasePtr, Value). |
481 | Type *IntPtrTy = DAG.getDataLayout().getIntPtrType(C&: Context); |
482 | TargetLowering::ArgListTy Args; |
483 | TargetLowering::ArgListEntry Entry; |
484 | |
485 | Entry.Ty = IntPtrTy; |
486 | Entry.Node = BasePtr; |
487 | Args.push_back(x: Entry); |
488 | |
489 | Entry.Node = Value; |
490 | Args.push_back(x: Entry); |
491 | |
492 | TargetLowering::CallLoweringInfo CLI(DAG); |
493 | CLI.setDebugLoc(dl).setChain(Chain).setCallee( |
494 | CC: CallingConv::C, ResultType: Type::getVoidTy(C&: Context), |
495 | Target: DAG.getExternalSymbol(Sym: "__misaligned_store" , |
496 | VT: getPointerTy(DL: DAG.getDataLayout())), |
497 | ArgsList: std::move(Args)); |
498 | |
499 | std::pair<SDValue, SDValue> CallResult = LowerCallTo(CLI); |
500 | return CallResult.second; |
501 | } |
502 | |
503 | SDValue XCoreTargetLowering:: |
504 | LowerSMUL_LOHI(SDValue Op, SelectionDAG &DAG) const |
505 | { |
506 | assert(Op.getValueType() == MVT::i32 && Op.getOpcode() == ISD::SMUL_LOHI && |
507 | "Unexpected operand to lower!" ); |
508 | SDLoc dl(Op); |
509 | SDValue LHS = Op.getOperand(i: 0); |
510 | SDValue RHS = Op.getOperand(i: 1); |
511 | SDValue Zero = DAG.getConstant(Val: 0, DL: dl, VT: MVT::i32); |
512 | SDValue Hi = DAG.getNode(Opcode: XCoreISD::MACCS, DL: dl, |
513 | VTList: DAG.getVTList(VT1: MVT::i32, VT2: MVT::i32), N1: Zero, N2: Zero, |
514 | N3: LHS, N4: RHS); |
515 | SDValue Lo(Hi.getNode(), 1); |
516 | SDValue Ops[] = { Lo, Hi }; |
517 | return DAG.getMergeValues(Ops, dl); |
518 | } |
519 | |
520 | SDValue XCoreTargetLowering:: |
521 | LowerUMUL_LOHI(SDValue Op, SelectionDAG &DAG) const |
522 | { |
523 | assert(Op.getValueType() == MVT::i32 && Op.getOpcode() == ISD::UMUL_LOHI && |
524 | "Unexpected operand to lower!" ); |
525 | SDLoc dl(Op); |
526 | SDValue LHS = Op.getOperand(i: 0); |
527 | SDValue RHS = Op.getOperand(i: 1); |
528 | SDValue Zero = DAG.getConstant(Val: 0, DL: dl, VT: MVT::i32); |
529 | SDValue Hi = DAG.getNode(Opcode: XCoreISD::LMUL, DL: dl, |
530 | VTList: DAG.getVTList(VT1: MVT::i32, VT2: MVT::i32), N1: LHS, N2: RHS, |
531 | N3: Zero, N4: Zero); |
532 | SDValue Lo(Hi.getNode(), 1); |
533 | SDValue Ops[] = { Lo, Hi }; |
534 | return DAG.getMergeValues(Ops, dl); |
535 | } |
536 | |
537 | /// isADDADDMUL - Return whether Op is in a form that is equivalent to |
538 | /// add(add(mul(x,y),a),b). If requireIntermediatesHaveOneUse is true then |
539 | /// each intermediate result in the calculation must also have a single use. |
540 | /// If the Op is in the correct form the constituent parts are written to Mul0, |
541 | /// Mul1, Addend0 and Addend1. |
542 | static bool |
543 | isADDADDMUL(SDValue Op, SDValue &Mul0, SDValue &Mul1, SDValue &Addend0, |
544 | SDValue &Addend1, bool requireIntermediatesHaveOneUse) |
545 | { |
546 | if (Op.getOpcode() != ISD::ADD) |
547 | return false; |
548 | SDValue N0 = Op.getOperand(i: 0); |
549 | SDValue N1 = Op.getOperand(i: 1); |
550 | SDValue AddOp; |
551 | SDValue OtherOp; |
552 | if (N0.getOpcode() == ISD::ADD) { |
553 | AddOp = N0; |
554 | OtherOp = N1; |
555 | } else if (N1.getOpcode() == ISD::ADD) { |
556 | AddOp = N1; |
557 | OtherOp = N0; |
558 | } else { |
559 | return false; |
560 | } |
561 | if (requireIntermediatesHaveOneUse && !AddOp.hasOneUse()) |
562 | return false; |
563 | if (OtherOp.getOpcode() == ISD::MUL) { |
564 | // add(add(a,b),mul(x,y)) |
565 | if (requireIntermediatesHaveOneUse && !OtherOp.hasOneUse()) |
566 | return false; |
567 | Mul0 = OtherOp.getOperand(i: 0); |
568 | Mul1 = OtherOp.getOperand(i: 1); |
569 | Addend0 = AddOp.getOperand(i: 0); |
570 | Addend1 = AddOp.getOperand(i: 1); |
571 | return true; |
572 | } |
573 | if (AddOp.getOperand(i: 0).getOpcode() == ISD::MUL) { |
574 | // add(add(mul(x,y),a),b) |
575 | if (requireIntermediatesHaveOneUse && !AddOp.getOperand(i: 0).hasOneUse()) |
576 | return false; |
577 | Mul0 = AddOp.getOperand(i: 0).getOperand(i: 0); |
578 | Mul1 = AddOp.getOperand(i: 0).getOperand(i: 1); |
579 | Addend0 = AddOp.getOperand(i: 1); |
580 | Addend1 = OtherOp; |
581 | return true; |
582 | } |
583 | if (AddOp.getOperand(i: 1).getOpcode() == ISD::MUL) { |
584 | // add(add(a,mul(x,y)),b) |
585 | if (requireIntermediatesHaveOneUse && !AddOp.getOperand(i: 1).hasOneUse()) |
586 | return false; |
587 | Mul0 = AddOp.getOperand(i: 1).getOperand(i: 0); |
588 | Mul1 = AddOp.getOperand(i: 1).getOperand(i: 1); |
589 | Addend0 = AddOp.getOperand(i: 0); |
590 | Addend1 = OtherOp; |
591 | return true; |
592 | } |
593 | return false; |
594 | } |
595 | |
596 | SDValue XCoreTargetLowering:: |
597 | TryExpandADDWithMul(SDNode *N, SelectionDAG &DAG) const |
598 | { |
599 | SDValue Mul; |
600 | SDValue Other; |
601 | if (N->getOperand(Num: 0).getOpcode() == ISD::MUL) { |
602 | Mul = N->getOperand(Num: 0); |
603 | Other = N->getOperand(Num: 1); |
604 | } else if (N->getOperand(Num: 1).getOpcode() == ISD::MUL) { |
605 | Mul = N->getOperand(Num: 1); |
606 | Other = N->getOperand(Num: 0); |
607 | } else { |
608 | return SDValue(); |
609 | } |
610 | SDLoc dl(N); |
611 | SDValue LL, RL, AddendL, AddendH; |
612 | LL = DAG.getNode(Opcode: ISD::EXTRACT_ELEMENT, DL: dl, VT: MVT::i32, |
613 | N1: Mul.getOperand(i: 0), N2: DAG.getConstant(Val: 0, DL: dl, VT: MVT::i32)); |
614 | RL = DAG.getNode(Opcode: ISD::EXTRACT_ELEMENT, DL: dl, VT: MVT::i32, |
615 | N1: Mul.getOperand(i: 1), N2: DAG.getConstant(Val: 0, DL: dl, VT: MVT::i32)); |
616 | AddendL = DAG.getNode(Opcode: ISD::EXTRACT_ELEMENT, DL: dl, VT: MVT::i32, |
617 | N1: Other, N2: DAG.getConstant(Val: 0, DL: dl, VT: MVT::i32)); |
618 | AddendH = DAG.getNode(Opcode: ISD::EXTRACT_ELEMENT, DL: dl, VT: MVT::i32, |
619 | N1: Other, N2: DAG.getConstant(Val: 1, DL: dl, VT: MVT::i32)); |
620 | APInt HighMask = APInt::getHighBitsSet(numBits: 64, hiBitsSet: 32); |
621 | unsigned LHSSB = DAG.ComputeNumSignBits(Op: Mul.getOperand(i: 0)); |
622 | unsigned RHSSB = DAG.ComputeNumSignBits(Op: Mul.getOperand(i: 1)); |
623 | if (DAG.MaskedValueIsZero(Op: Mul.getOperand(i: 0), Mask: HighMask) && |
624 | DAG.MaskedValueIsZero(Op: Mul.getOperand(i: 1), Mask: HighMask)) { |
625 | // The inputs are both zero-extended. |
626 | SDValue Hi = DAG.getNode(Opcode: XCoreISD::MACCU, DL: dl, |
627 | VTList: DAG.getVTList(VT1: MVT::i32, VT2: MVT::i32), N1: AddendH, |
628 | N2: AddendL, N3: LL, N4: RL); |
629 | SDValue Lo(Hi.getNode(), 1); |
630 | return DAG.getNode(Opcode: ISD::BUILD_PAIR, DL: dl, VT: MVT::i64, N1: Lo, N2: Hi); |
631 | } |
632 | if (LHSSB > 32 && RHSSB > 32) { |
633 | // The inputs are both sign-extended. |
634 | SDValue Hi = DAG.getNode(Opcode: XCoreISD::MACCS, DL: dl, |
635 | VTList: DAG.getVTList(VT1: MVT::i32, VT2: MVT::i32), N1: AddendH, |
636 | N2: AddendL, N3: LL, N4: RL); |
637 | SDValue Lo(Hi.getNode(), 1); |
638 | return DAG.getNode(Opcode: ISD::BUILD_PAIR, DL: dl, VT: MVT::i64, N1: Lo, N2: Hi); |
639 | } |
640 | SDValue LH, RH; |
641 | LH = DAG.getNode(Opcode: ISD::EXTRACT_ELEMENT, DL: dl, VT: MVT::i32, |
642 | N1: Mul.getOperand(i: 0), N2: DAG.getConstant(Val: 1, DL: dl, VT: MVT::i32)); |
643 | RH = DAG.getNode(Opcode: ISD::EXTRACT_ELEMENT, DL: dl, VT: MVT::i32, |
644 | N1: Mul.getOperand(i: 1), N2: DAG.getConstant(Val: 1, DL: dl, VT: MVT::i32)); |
645 | SDValue Hi = DAG.getNode(Opcode: XCoreISD::MACCU, DL: dl, |
646 | VTList: DAG.getVTList(VT1: MVT::i32, VT2: MVT::i32), N1: AddendH, |
647 | N2: AddendL, N3: LL, N4: RL); |
648 | SDValue Lo(Hi.getNode(), 1); |
649 | RH = DAG.getNode(Opcode: ISD::MUL, DL: dl, VT: MVT::i32, N1: LL, N2: RH); |
650 | LH = DAG.getNode(Opcode: ISD::MUL, DL: dl, VT: MVT::i32, N1: LH, N2: RL); |
651 | Hi = DAG.getNode(Opcode: ISD::ADD, DL: dl, VT: MVT::i32, N1: Hi, N2: RH); |
652 | Hi = DAG.getNode(Opcode: ISD::ADD, DL: dl, VT: MVT::i32, N1: Hi, N2: LH); |
653 | return DAG.getNode(Opcode: ISD::BUILD_PAIR, DL: dl, VT: MVT::i64, N1: Lo, N2: Hi); |
654 | } |
655 | |
656 | SDValue XCoreTargetLowering:: |
657 | ExpandADDSUB(SDNode *N, SelectionDAG &DAG) const |
658 | { |
659 | assert(N->getValueType(0) == MVT::i64 && |
660 | (N->getOpcode() == ISD::ADD || N->getOpcode() == ISD::SUB) && |
661 | "Unknown operand to lower!" ); |
662 | |
663 | if (N->getOpcode() == ISD::ADD) |
664 | if (SDValue Result = TryExpandADDWithMul(N, DAG)) |
665 | return Result; |
666 | |
667 | SDLoc dl(N); |
668 | |
669 | // Extract components |
670 | SDValue LHSL = DAG.getNode(Opcode: ISD::EXTRACT_ELEMENT, DL: dl, VT: MVT::i32, |
671 | N1: N->getOperand(Num: 0), |
672 | N2: DAG.getConstant(Val: 0, DL: dl, VT: MVT::i32)); |
673 | SDValue LHSH = DAG.getNode(Opcode: ISD::EXTRACT_ELEMENT, DL: dl, VT: MVT::i32, |
674 | N1: N->getOperand(Num: 0), |
675 | N2: DAG.getConstant(Val: 1, DL: dl, VT: MVT::i32)); |
676 | SDValue RHSL = DAG.getNode(Opcode: ISD::EXTRACT_ELEMENT, DL: dl, VT: MVT::i32, |
677 | N1: N->getOperand(Num: 1), |
678 | N2: DAG.getConstant(Val: 0, DL: dl, VT: MVT::i32)); |
679 | SDValue RHSH = DAG.getNode(Opcode: ISD::EXTRACT_ELEMENT, DL: dl, VT: MVT::i32, |
680 | N1: N->getOperand(Num: 1), |
681 | N2: DAG.getConstant(Val: 1, DL: dl, VT: MVT::i32)); |
682 | |
683 | // Expand |
684 | unsigned Opcode = (N->getOpcode() == ISD::ADD) ? XCoreISD::LADD : |
685 | XCoreISD::LSUB; |
686 | SDValue Zero = DAG.getConstant(Val: 0, DL: dl, VT: MVT::i32); |
687 | SDValue Lo = DAG.getNode(Opcode, DL: dl, VTList: DAG.getVTList(VT1: MVT::i32, VT2: MVT::i32), |
688 | N1: LHSL, N2: RHSL, N3: Zero); |
689 | SDValue Carry(Lo.getNode(), 1); |
690 | |
691 | SDValue Hi = DAG.getNode(Opcode, DL: dl, VTList: DAG.getVTList(VT1: MVT::i32, VT2: MVT::i32), |
692 | N1: LHSH, N2: RHSH, N3: Carry); |
693 | SDValue Ignored(Hi.getNode(), 1); |
694 | // Merge the pieces |
695 | return DAG.getNode(Opcode: ISD::BUILD_PAIR, DL: dl, VT: MVT::i64, N1: Lo, N2: Hi); |
696 | } |
697 | |
698 | SDValue XCoreTargetLowering:: |
699 | LowerVAARG(SDValue Op, SelectionDAG &DAG) const |
700 | { |
701 | // Whist llvm does not support aggregate varargs we can ignore |
702 | // the possibility of the ValueType being an implicit byVal vararg. |
703 | SDNode *Node = Op.getNode(); |
704 | EVT VT = Node->getValueType(ResNo: 0); // not an aggregate |
705 | SDValue InChain = Node->getOperand(Num: 0); |
706 | SDValue VAListPtr = Node->getOperand(Num: 1); |
707 | EVT PtrVT = VAListPtr.getValueType(); |
708 | const Value *SV = cast<SrcValueSDNode>(Val: Node->getOperand(Num: 2))->getValue(); |
709 | SDLoc dl(Node); |
710 | SDValue VAList = |
711 | DAG.getLoad(VT: PtrVT, dl, Chain: InChain, Ptr: VAListPtr, PtrInfo: MachinePointerInfo(SV)); |
712 | // Increment the pointer, VAList, to the next vararg |
713 | SDValue nextPtr = DAG.getNode(Opcode: ISD::ADD, DL: dl, VT: PtrVT, N1: VAList, |
714 | N2: DAG.getIntPtrConstant(Val: VT.getSizeInBits() / 8, |
715 | DL: dl)); |
716 | // Store the incremented VAList to the legalized pointer |
717 | InChain = DAG.getStore(Chain: VAList.getValue(R: 1), dl, Val: nextPtr, Ptr: VAListPtr, |
718 | PtrInfo: MachinePointerInfo(SV)); |
719 | // Load the actual argument out of the pointer VAList |
720 | return DAG.getLoad(VT, dl, Chain: InChain, Ptr: VAList, PtrInfo: MachinePointerInfo()); |
721 | } |
722 | |
723 | SDValue XCoreTargetLowering:: |
724 | LowerVASTART(SDValue Op, SelectionDAG &DAG) const |
725 | { |
726 | SDLoc dl(Op); |
727 | // vastart stores the address of the VarArgsFrameIndex slot into the |
728 | // memory location argument |
729 | MachineFunction &MF = DAG.getMachineFunction(); |
730 | XCoreFunctionInfo *XFI = MF.getInfo<XCoreFunctionInfo>(); |
731 | SDValue Addr = DAG.getFrameIndex(FI: XFI->getVarArgsFrameIndex(), VT: MVT::i32); |
732 | return DAG.getStore(Chain: Op.getOperand(i: 0), dl, Val: Addr, Ptr: Op.getOperand(i: 1), |
733 | PtrInfo: MachinePointerInfo()); |
734 | } |
735 | |
736 | SDValue XCoreTargetLowering::LowerFRAMEADDR(SDValue Op, |
737 | SelectionDAG &DAG) const { |
738 | // This nodes represent llvm.frameaddress on the DAG. |
739 | // It takes one operand, the index of the frame address to return. |
740 | // An index of zero corresponds to the current function's frame address. |
741 | // An index of one to the parent's frame address, and so on. |
742 | // Depths > 0 not supported yet! |
743 | if (Op.getConstantOperandVal(i: 0) > 0) |
744 | return SDValue(); |
745 | |
746 | MachineFunction &MF = DAG.getMachineFunction(); |
747 | const TargetRegisterInfo *RegInfo = Subtarget.getRegisterInfo(); |
748 | return DAG.getCopyFromReg(Chain: DAG.getEntryNode(), dl: SDLoc(Op), |
749 | Reg: RegInfo->getFrameRegister(MF), VT: MVT::i32); |
750 | } |
751 | |
752 | SDValue XCoreTargetLowering:: |
753 | LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) const { |
754 | // This nodes represent llvm.returnaddress on the DAG. |
755 | // It takes one operand, the index of the return address to return. |
756 | // An index of zero corresponds to the current function's return address. |
757 | // An index of one to the parent's return address, and so on. |
758 | // Depths > 0 not supported yet! |
759 | if (Op.getConstantOperandVal(i: 0) > 0) |
760 | return SDValue(); |
761 | |
762 | MachineFunction &MF = DAG.getMachineFunction(); |
763 | XCoreFunctionInfo *XFI = MF.getInfo<XCoreFunctionInfo>(); |
764 | int FI = XFI->createLRSpillSlot(MF); |
765 | SDValue FIN = DAG.getFrameIndex(FI, VT: MVT::i32); |
766 | return DAG.getLoad(VT: getPointerTy(DL: DAG.getDataLayout()), dl: SDLoc(Op), |
767 | Chain: DAG.getEntryNode(), Ptr: FIN, |
768 | PtrInfo: MachinePointerInfo::getFixedStack(MF, FI)); |
769 | } |
770 | |
771 | SDValue XCoreTargetLowering:: |
772 | LowerFRAME_TO_ARGS_OFFSET(SDValue Op, SelectionDAG &DAG) const { |
773 | // This node represents offset from frame pointer to first on-stack argument. |
774 | // This is needed for correct stack adjustment during unwind. |
775 | // However, we don't know the offset until after the frame has be finalised. |
776 | // This is done during the XCoreFTAOElim pass. |
777 | return DAG.getNode(Opcode: XCoreISD::FRAME_TO_ARGS_OFFSET, DL: SDLoc(Op), VT: MVT::i32); |
778 | } |
779 | |
780 | SDValue XCoreTargetLowering:: |
781 | LowerEH_RETURN(SDValue Op, SelectionDAG &DAG) const { |
782 | // OUTCHAIN = EH_RETURN(INCHAIN, OFFSET, HANDLER) |
783 | // This node represents 'eh_return' gcc dwarf builtin, which is used to |
784 | // return from exception. The general meaning is: adjust stack by OFFSET and |
785 | // pass execution to HANDLER. |
786 | MachineFunction &MF = DAG.getMachineFunction(); |
787 | SDValue Chain = Op.getOperand(i: 0); |
788 | SDValue Offset = Op.getOperand(i: 1); |
789 | SDValue Handler = Op.getOperand(i: 2); |
790 | SDLoc dl(Op); |
791 | |
792 | // Absolute SP = (FP + FrameToArgs) + Offset |
793 | const TargetRegisterInfo *RegInfo = Subtarget.getRegisterInfo(); |
794 | SDValue Stack = DAG.getCopyFromReg(Chain: DAG.getEntryNode(), dl, |
795 | Reg: RegInfo->getFrameRegister(MF), VT: MVT::i32); |
796 | SDValue FrameToArgs = DAG.getNode(Opcode: XCoreISD::FRAME_TO_ARGS_OFFSET, DL: dl, |
797 | VT: MVT::i32); |
798 | Stack = DAG.getNode(Opcode: ISD::ADD, DL: dl, VT: MVT::i32, N1: Stack, N2: FrameToArgs); |
799 | Stack = DAG.getNode(Opcode: ISD::ADD, DL: dl, VT: MVT::i32, N1: Stack, N2: Offset); |
800 | |
801 | // R0=ExceptionPointerRegister R1=ExceptionSelectorRegister |
802 | // which leaves 2 caller saved registers, R2 & R3 for us to use. |
803 | unsigned StackReg = XCore::R2; |
804 | unsigned HandlerReg = XCore::R3; |
805 | |
806 | SDValue OutChains[] = { |
807 | DAG.getCopyToReg(Chain, dl, Reg: StackReg, N: Stack), |
808 | DAG.getCopyToReg(Chain, dl, Reg: HandlerReg, N: Handler) |
809 | }; |
810 | |
811 | Chain = DAG.getNode(Opcode: ISD::TokenFactor, DL: dl, VT: MVT::Other, Ops: OutChains); |
812 | |
813 | return DAG.getNode(Opcode: XCoreISD::EH_RETURN, DL: dl, VT: MVT::Other, N1: Chain, |
814 | N2: DAG.getRegister(Reg: StackReg, VT: MVT::i32), |
815 | N3: DAG.getRegister(Reg: HandlerReg, VT: MVT::i32)); |
816 | |
817 | } |
818 | |
819 | SDValue XCoreTargetLowering:: |
820 | LowerADJUST_TRAMPOLINE(SDValue Op, SelectionDAG &DAG) const { |
821 | return Op.getOperand(i: 0); |
822 | } |
823 | |
824 | SDValue XCoreTargetLowering:: |
825 | LowerINIT_TRAMPOLINE(SDValue Op, SelectionDAG &DAG) const { |
826 | SDValue Chain = Op.getOperand(i: 0); |
827 | SDValue Trmp = Op.getOperand(i: 1); // trampoline |
828 | SDValue FPtr = Op.getOperand(i: 2); // nested function |
829 | SDValue Nest = Op.getOperand(i: 3); // 'nest' parameter value |
830 | |
831 | const Value *TrmpAddr = cast<SrcValueSDNode>(Val: Op.getOperand(i: 4))->getValue(); |
832 | |
833 | // .align 4 |
834 | // LDAPF_u10 r11, nest |
835 | // LDW_2rus r11, r11[0] |
836 | // STWSP_ru6 r11, sp[0] |
837 | // LDAPF_u10 r11, fptr |
838 | // LDW_2rus r11, r11[0] |
839 | // BAU_1r r11 |
840 | // nest: |
841 | // .word nest |
842 | // fptr: |
843 | // .word fptr |
844 | SDValue OutChains[5]; |
845 | |
846 | SDValue Addr = Trmp; |
847 | |
848 | SDLoc dl(Op); |
849 | OutChains[0] = |
850 | DAG.getStore(Chain, dl, Val: DAG.getConstant(Val: 0x0a3cd805, DL: dl, VT: MVT::i32), Ptr: Addr, |
851 | PtrInfo: MachinePointerInfo(TrmpAddr)); |
852 | |
853 | Addr = DAG.getNode(Opcode: ISD::ADD, DL: dl, VT: MVT::i32, N1: Trmp, |
854 | N2: DAG.getConstant(Val: 4, DL: dl, VT: MVT::i32)); |
855 | OutChains[1] = |
856 | DAG.getStore(Chain, dl, Val: DAG.getConstant(Val: 0xd80456c0, DL: dl, VT: MVT::i32), Ptr: Addr, |
857 | PtrInfo: MachinePointerInfo(TrmpAddr, 4)); |
858 | |
859 | Addr = DAG.getNode(Opcode: ISD::ADD, DL: dl, VT: MVT::i32, N1: Trmp, |
860 | N2: DAG.getConstant(Val: 8, DL: dl, VT: MVT::i32)); |
861 | OutChains[2] = |
862 | DAG.getStore(Chain, dl, Val: DAG.getConstant(Val: 0x27fb0a3c, DL: dl, VT: MVT::i32), Ptr: Addr, |
863 | PtrInfo: MachinePointerInfo(TrmpAddr, 8)); |
864 | |
865 | Addr = DAG.getNode(Opcode: ISD::ADD, DL: dl, VT: MVT::i32, N1: Trmp, |
866 | N2: DAG.getConstant(Val: 12, DL: dl, VT: MVT::i32)); |
867 | OutChains[3] = |
868 | DAG.getStore(Chain, dl, Val: Nest, Ptr: Addr, PtrInfo: MachinePointerInfo(TrmpAddr, 12)); |
869 | |
870 | Addr = DAG.getNode(Opcode: ISD::ADD, DL: dl, VT: MVT::i32, N1: Trmp, |
871 | N2: DAG.getConstant(Val: 16, DL: dl, VT: MVT::i32)); |
872 | OutChains[4] = |
873 | DAG.getStore(Chain, dl, Val: FPtr, Ptr: Addr, PtrInfo: MachinePointerInfo(TrmpAddr, 16)); |
874 | |
875 | return DAG.getNode(Opcode: ISD::TokenFactor, DL: dl, VT: MVT::Other, Ops: OutChains); |
876 | } |
877 | |
878 | SDValue XCoreTargetLowering:: |
879 | LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG) const { |
880 | SDLoc DL(Op); |
881 | unsigned IntNo = Op.getConstantOperandVal(i: 0); |
882 | switch (IntNo) { |
883 | case Intrinsic::xcore_crc8: |
884 | EVT VT = Op.getValueType(); |
885 | SDValue Data = |
886 | DAG.getNode(Opcode: XCoreISD::CRC8, DL, VTList: DAG.getVTList(VT1: VT, VT2: VT), |
887 | N1: Op.getOperand(i: 1), N2: Op.getOperand(i: 2) , N3: Op.getOperand(i: 3)); |
888 | SDValue Crc(Data.getNode(), 1); |
889 | SDValue Results[] = { Crc, Data }; |
890 | return DAG.getMergeValues(Ops: Results, dl: DL); |
891 | } |
892 | return SDValue(); |
893 | } |
894 | |
895 | SDValue XCoreTargetLowering:: |
896 | LowerATOMIC_FENCE(SDValue Op, SelectionDAG &DAG) const { |
897 | SDLoc DL(Op); |
898 | return DAG.getNode(Opcode: ISD::MEMBARRIER, DL, VT: MVT::Other, Operand: Op.getOperand(i: 0)); |
899 | } |
900 | |
901 | //===----------------------------------------------------------------------===// |
902 | // Calling Convention Implementation |
903 | //===----------------------------------------------------------------------===// |
904 | |
905 | #include "XCoreGenCallingConv.inc" |
906 | |
907 | //===----------------------------------------------------------------------===// |
908 | // Call Calling Convention Implementation |
909 | //===----------------------------------------------------------------------===// |
910 | |
911 | /// XCore call implementation |
912 | SDValue |
913 | XCoreTargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI, |
914 | SmallVectorImpl<SDValue> &InVals) const { |
915 | SelectionDAG &DAG = CLI.DAG; |
916 | SDLoc &dl = CLI.DL; |
917 | SmallVectorImpl<ISD::OutputArg> &Outs = CLI.Outs; |
918 | SmallVectorImpl<SDValue> &OutVals = CLI.OutVals; |
919 | SmallVectorImpl<ISD::InputArg> &Ins = CLI.Ins; |
920 | SDValue Chain = CLI.Chain; |
921 | SDValue Callee = CLI.Callee; |
922 | bool &isTailCall = CLI.IsTailCall; |
923 | CallingConv::ID CallConv = CLI.CallConv; |
924 | bool isVarArg = CLI.IsVarArg; |
925 | |
926 | // XCore target does not yet support tail call optimization. |
927 | isTailCall = false; |
928 | |
929 | // For now, only CallingConv::C implemented |
930 | switch (CallConv) |
931 | { |
932 | default: |
933 | report_fatal_error(reason: "Unsupported calling convention" ); |
934 | case CallingConv::Fast: |
935 | case CallingConv::C: |
936 | return LowerCCCCallTo(Chain, Callee, CallConv, isVarArg, isTailCall, |
937 | Outs, OutVals, Ins, dl, DAG, InVals); |
938 | } |
939 | } |
940 | |
941 | /// LowerCallResult - Lower the result values of a call into the |
942 | /// appropriate copies out of appropriate physical registers / memory locations. |
943 | static SDValue LowerCallResult(SDValue Chain, SDValue InGlue, |
944 | const SmallVectorImpl<CCValAssign> &RVLocs, |
945 | const SDLoc &dl, SelectionDAG &DAG, |
946 | SmallVectorImpl<SDValue> &InVals) { |
947 | SmallVector<std::pair<int, unsigned>, 4> ResultMemLocs; |
948 | // Copy results out of physical registers. |
949 | for (const CCValAssign &VA : RVLocs) { |
950 | if (VA.isRegLoc()) { |
951 | Chain = DAG.getCopyFromReg(Chain, dl, Reg: VA.getLocReg(), VT: VA.getValVT(), |
952 | Glue: InGlue).getValue(R: 1); |
953 | InGlue = Chain.getValue(R: 2); |
954 | InVals.push_back(Elt: Chain.getValue(R: 0)); |
955 | } else { |
956 | assert(VA.isMemLoc()); |
957 | ResultMemLocs.push_back(Elt: std::make_pair(x: VA.getLocMemOffset(), |
958 | y: InVals.size())); |
959 | // Reserve space for this result. |
960 | InVals.push_back(Elt: SDValue()); |
961 | } |
962 | } |
963 | |
964 | // Copy results out of memory. |
965 | SmallVector<SDValue, 4> MemOpChains; |
966 | for (unsigned i = 0, e = ResultMemLocs.size(); i != e; ++i) { |
967 | int offset = ResultMemLocs[i].first; |
968 | unsigned index = ResultMemLocs[i].second; |
969 | SDVTList VTs = DAG.getVTList(VT1: MVT::i32, VT2: MVT::Other); |
970 | SDValue Ops[] = { Chain, DAG.getConstant(Val: offset / 4, DL: dl, VT: MVT::i32) }; |
971 | SDValue load = DAG.getNode(Opcode: XCoreISD::LDWSP, DL: dl, VTList: VTs, Ops); |
972 | InVals[index] = load; |
973 | MemOpChains.push_back(Elt: load.getValue(R: 1)); |
974 | } |
975 | |
976 | // Transform all loads nodes into one single node because |
977 | // all load nodes are independent of each other. |
978 | if (!MemOpChains.empty()) |
979 | Chain = DAG.getNode(Opcode: ISD::TokenFactor, DL: dl, VT: MVT::Other, Ops: MemOpChains); |
980 | |
981 | return Chain; |
982 | } |
983 | |
984 | /// LowerCCCCallTo - functions arguments are copied from virtual |
985 | /// regs to (physical regs)/(stack frame), CALLSEQ_START and |
986 | /// CALLSEQ_END are emitted. |
987 | /// TODO: isTailCall, sret. |
988 | SDValue XCoreTargetLowering::LowerCCCCallTo( |
989 | SDValue Chain, SDValue Callee, CallingConv::ID CallConv, bool isVarArg, |
990 | bool isTailCall, const SmallVectorImpl<ISD::OutputArg> &Outs, |
991 | const SmallVectorImpl<SDValue> &OutVals, |
992 | const SmallVectorImpl<ISD::InputArg> &Ins, const SDLoc &dl, |
993 | SelectionDAG &DAG, SmallVectorImpl<SDValue> &InVals) const { |
994 | |
995 | // Analyze operands of the call, assigning locations to each operand. |
996 | SmallVector<CCValAssign, 16> ArgLocs; |
997 | CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), ArgLocs, |
998 | *DAG.getContext()); |
999 | |
1000 | // The ABI dictates there should be one stack slot available to the callee |
1001 | // on function entry (for saving lr). |
1002 | CCInfo.AllocateStack(Size: 4, Alignment: Align(4)); |
1003 | |
1004 | CCInfo.AnalyzeCallOperands(Outs, Fn: CC_XCore); |
1005 | |
1006 | SmallVector<CCValAssign, 16> RVLocs; |
1007 | // Analyze return values to determine the number of bytes of stack required. |
1008 | CCState RetCCInfo(CallConv, isVarArg, DAG.getMachineFunction(), RVLocs, |
1009 | *DAG.getContext()); |
1010 | RetCCInfo.AllocateStack(Size: CCInfo.getStackSize(), Alignment: Align(4)); |
1011 | RetCCInfo.AnalyzeCallResult(Ins, Fn: RetCC_XCore); |
1012 | |
1013 | // Get a count of how many bytes are to be pushed on the stack. |
1014 | unsigned NumBytes = RetCCInfo.getStackSize(); |
1015 | |
1016 | Chain = DAG.getCALLSEQ_START(Chain, InSize: NumBytes, OutSize: 0, DL: dl); |
1017 | |
1018 | SmallVector<std::pair<unsigned, SDValue>, 4> RegsToPass; |
1019 | SmallVector<SDValue, 12> MemOpChains; |
1020 | |
1021 | // Walk the register/memloc assignments, inserting copies/loads. |
1022 | for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) { |
1023 | CCValAssign &VA = ArgLocs[i]; |
1024 | SDValue Arg = OutVals[i]; |
1025 | |
1026 | // Promote the value if needed. |
1027 | switch (VA.getLocInfo()) { |
1028 | default: llvm_unreachable("Unknown loc info!" ); |
1029 | case CCValAssign::Full: break; |
1030 | case CCValAssign::SExt: |
1031 | Arg = DAG.getNode(Opcode: ISD::SIGN_EXTEND, DL: dl, VT: VA.getLocVT(), Operand: Arg); |
1032 | break; |
1033 | case CCValAssign::ZExt: |
1034 | Arg = DAG.getNode(Opcode: ISD::ZERO_EXTEND, DL: dl, VT: VA.getLocVT(), Operand: Arg); |
1035 | break; |
1036 | case CCValAssign::AExt: |
1037 | Arg = DAG.getNode(Opcode: ISD::ANY_EXTEND, DL: dl, VT: VA.getLocVT(), Operand: Arg); |
1038 | break; |
1039 | } |
1040 | |
1041 | // Arguments that can be passed on register must be kept at |
1042 | // RegsToPass vector |
1043 | if (VA.isRegLoc()) { |
1044 | RegsToPass.push_back(Elt: std::make_pair(x: VA.getLocReg(), y&: Arg)); |
1045 | } else { |
1046 | assert(VA.isMemLoc()); |
1047 | |
1048 | int Offset = VA.getLocMemOffset(); |
1049 | |
1050 | MemOpChains.push_back(Elt: DAG.getNode(Opcode: XCoreISD::STWSP, DL: dl, VT: MVT::Other, |
1051 | N1: Chain, N2: Arg, |
1052 | N3: DAG.getConstant(Val: Offset/4, DL: dl, |
1053 | VT: MVT::i32))); |
1054 | } |
1055 | } |
1056 | |
1057 | // Transform all store nodes into one single node because |
1058 | // all store nodes are independent of each other. |
1059 | if (!MemOpChains.empty()) |
1060 | Chain = DAG.getNode(Opcode: ISD::TokenFactor, DL: dl, VT: MVT::Other, Ops: MemOpChains); |
1061 | |
1062 | // Build a sequence of copy-to-reg nodes chained together with token |
1063 | // chain and flag operands which copy the outgoing args into registers. |
1064 | // The InGlue in necessary since all emitted instructions must be |
1065 | // stuck together. |
1066 | SDValue InGlue; |
1067 | for (const auto &[Reg, N] : RegsToPass) { |
1068 | Chain = DAG.getCopyToReg(Chain, dl, Reg, N, Glue: InGlue); |
1069 | InGlue = Chain.getValue(R: 1); |
1070 | } |
1071 | |
1072 | // If the callee is a GlobalAddress node (quite common, every direct call is) |
1073 | // turn it into a TargetGlobalAddress node so that legalize doesn't hack it. |
1074 | // Likewise ExternalSymbol -> TargetExternalSymbol. |
1075 | if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Val&: Callee)) |
1076 | Callee = DAG.getTargetGlobalAddress(GV: G->getGlobal(), DL: dl, VT: MVT::i32); |
1077 | else if (ExternalSymbolSDNode *E = dyn_cast<ExternalSymbolSDNode>(Val&: Callee)) |
1078 | Callee = DAG.getTargetExternalSymbol(Sym: E->getSymbol(), VT: MVT::i32); |
1079 | |
1080 | // XCoreBranchLink = #chain, #target_address, #opt_in_flags... |
1081 | // = Chain, Callee, Reg#1, Reg#2, ... |
1082 | // |
1083 | // Returns a chain & a flag for retval copy to use. |
1084 | SDVTList NodeTys = DAG.getVTList(VT1: MVT::Other, VT2: MVT::Glue); |
1085 | SmallVector<SDValue, 8> Ops; |
1086 | Ops.push_back(Elt: Chain); |
1087 | Ops.push_back(Elt: Callee); |
1088 | |
1089 | // Add argument registers to the end of the list so that they are |
1090 | // known live into the call. |
1091 | for (const auto &[Reg, N] : RegsToPass) |
1092 | Ops.push_back(Elt: DAG.getRegister(Reg, VT: N.getValueType())); |
1093 | |
1094 | if (InGlue.getNode()) |
1095 | Ops.push_back(Elt: InGlue); |
1096 | |
1097 | Chain = DAG.getNode(Opcode: XCoreISD::BL, DL: dl, VTList: NodeTys, Ops); |
1098 | InGlue = Chain.getValue(R: 1); |
1099 | |
1100 | // Create the CALLSEQ_END node. |
1101 | Chain = DAG.getCALLSEQ_END(Chain, Size1: NumBytes, Size2: 0, Glue: InGlue, DL: dl); |
1102 | InGlue = Chain.getValue(R: 1); |
1103 | |
1104 | // Handle result values, copying them out of physregs into vregs that we |
1105 | // return. |
1106 | return LowerCallResult(Chain, InGlue, RVLocs, dl, DAG, InVals); |
1107 | } |
1108 | |
1109 | //===----------------------------------------------------------------------===// |
1110 | // Formal Arguments Calling Convention Implementation |
1111 | //===----------------------------------------------------------------------===// |
1112 | |
1113 | namespace { |
1114 | struct ArgDataPair { SDValue SDV; ISD::ArgFlagsTy Flags; }; |
1115 | } |
1116 | |
1117 | /// XCore formal arguments implementation |
1118 | SDValue XCoreTargetLowering::LowerFormalArguments( |
1119 | SDValue Chain, CallingConv::ID CallConv, bool isVarArg, |
1120 | const SmallVectorImpl<ISD::InputArg> &Ins, const SDLoc &dl, |
1121 | SelectionDAG &DAG, SmallVectorImpl<SDValue> &InVals) const { |
1122 | switch (CallConv) |
1123 | { |
1124 | default: |
1125 | report_fatal_error(reason: "Unsupported calling convention" ); |
1126 | case CallingConv::C: |
1127 | case CallingConv::Fast: |
1128 | return LowerCCCArguments(Chain, CallConv, isVarArg, |
1129 | Ins, dl, DAG, InVals); |
1130 | } |
1131 | } |
1132 | |
1133 | /// LowerCCCArguments - transform physical registers into |
1134 | /// virtual registers and generate load operations for |
1135 | /// arguments places on the stack. |
1136 | /// TODO: sret |
1137 | SDValue XCoreTargetLowering::LowerCCCArguments( |
1138 | SDValue Chain, CallingConv::ID CallConv, bool isVarArg, |
1139 | const SmallVectorImpl<ISD::InputArg> &Ins, const SDLoc &dl, |
1140 | SelectionDAG &DAG, SmallVectorImpl<SDValue> &InVals) const { |
1141 | MachineFunction &MF = DAG.getMachineFunction(); |
1142 | MachineFrameInfo &MFI = MF.getFrameInfo(); |
1143 | MachineRegisterInfo &RegInfo = MF.getRegInfo(); |
1144 | XCoreFunctionInfo *XFI = MF.getInfo<XCoreFunctionInfo>(); |
1145 | |
1146 | // Assign locations to all of the incoming arguments. |
1147 | SmallVector<CCValAssign, 16> ArgLocs; |
1148 | CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), ArgLocs, |
1149 | *DAG.getContext()); |
1150 | |
1151 | CCInfo.AnalyzeFormalArguments(Ins, Fn: CC_XCore); |
1152 | |
1153 | unsigned StackSlotSize = XCoreFrameLowering::stackSlotSize(); |
1154 | |
1155 | unsigned LRSaveSize = StackSlotSize; |
1156 | |
1157 | if (!isVarArg) |
1158 | XFI->setReturnStackOffset(CCInfo.getStackSize() + LRSaveSize); |
1159 | |
1160 | // All getCopyFromReg ops must precede any getMemcpys to prevent the |
1161 | // scheduler clobbering a register before it has been copied. |
1162 | // The stages are: |
1163 | // 1. CopyFromReg (and load) arg & vararg registers. |
1164 | // 2. Chain CopyFromReg nodes into a TokenFactor. |
1165 | // 3. Memcpy 'byVal' args & push final InVals. |
1166 | // 4. Chain mem ops nodes into a TokenFactor. |
1167 | SmallVector<SDValue, 4> CFRegNode; |
1168 | SmallVector<ArgDataPair, 4> ArgData; |
1169 | SmallVector<SDValue, 4> MemOps; |
1170 | |
1171 | // 1a. CopyFromReg (and load) arg registers. |
1172 | for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) { |
1173 | |
1174 | CCValAssign &VA = ArgLocs[i]; |
1175 | SDValue ArgIn; |
1176 | |
1177 | if (VA.isRegLoc()) { |
1178 | // Arguments passed in registers |
1179 | EVT RegVT = VA.getLocVT(); |
1180 | switch (RegVT.getSimpleVT().SimpleTy) { |
1181 | default: |
1182 | { |
1183 | #ifndef NDEBUG |
1184 | errs() << "LowerFormalArguments Unhandled argument type: " |
1185 | << RegVT << "\n" ; |
1186 | #endif |
1187 | llvm_unreachable(nullptr); |
1188 | } |
1189 | case MVT::i32: |
1190 | Register VReg = RegInfo.createVirtualRegister(RegClass: &XCore::GRRegsRegClass); |
1191 | RegInfo.addLiveIn(Reg: VA.getLocReg(), vreg: VReg); |
1192 | ArgIn = DAG.getCopyFromReg(Chain, dl, Reg: VReg, VT: RegVT); |
1193 | CFRegNode.push_back(Elt: ArgIn.getValue(R: ArgIn->getNumValues() - 1)); |
1194 | } |
1195 | } else { |
1196 | // Only arguments passed on the stack should make it here. |
1197 | assert(VA.isMemLoc()); |
1198 | // Load the argument to a virtual register |
1199 | unsigned ObjSize = VA.getLocVT().getSizeInBits()/8; |
1200 | if (ObjSize > StackSlotSize) { |
1201 | errs() << "LowerFormalArguments Unhandled argument type: " |
1202 | << VA.getLocVT() << "\n" ; |
1203 | } |
1204 | // Create the frame index object for this incoming parameter... |
1205 | int FI = MFI.CreateFixedObject(Size: ObjSize, |
1206 | SPOffset: LRSaveSize + VA.getLocMemOffset(), |
1207 | IsImmutable: true); |
1208 | |
1209 | // Create the SelectionDAG nodes corresponding to a load |
1210 | //from this parameter |
1211 | SDValue FIN = DAG.getFrameIndex(FI, VT: MVT::i32); |
1212 | ArgIn = DAG.getLoad(VT: VA.getLocVT(), dl, Chain, Ptr: FIN, |
1213 | PtrInfo: MachinePointerInfo::getFixedStack(MF, FI)); |
1214 | } |
1215 | const ArgDataPair ADP = { .SDV: ArgIn, .Flags: Ins[i].Flags }; |
1216 | ArgData.push_back(Elt: ADP); |
1217 | } |
1218 | |
1219 | // 1b. CopyFromReg vararg registers. |
1220 | if (isVarArg) { |
1221 | // Argument registers |
1222 | static const MCPhysReg ArgRegs[] = { |
1223 | XCore::R0, XCore::R1, XCore::R2, XCore::R3 |
1224 | }; |
1225 | XCoreFunctionInfo *XFI = MF.getInfo<XCoreFunctionInfo>(); |
1226 | unsigned FirstVAReg = CCInfo.getFirstUnallocated(Regs: ArgRegs); |
1227 | if (FirstVAReg < std::size(ArgRegs)) { |
1228 | int offset = 0; |
1229 | // Save remaining registers, storing higher register numbers at a higher |
1230 | // address |
1231 | for (int i = std::size(ArgRegs) - 1; i >= (int)FirstVAReg; --i) { |
1232 | // Create a stack slot |
1233 | int FI = MFI.CreateFixedObject(Size: 4, SPOffset: offset, IsImmutable: true); |
1234 | if (i == (int)FirstVAReg) { |
1235 | XFI->setVarArgsFrameIndex(FI); |
1236 | } |
1237 | offset -= StackSlotSize; |
1238 | SDValue FIN = DAG.getFrameIndex(FI, VT: MVT::i32); |
1239 | // Move argument from phys reg -> virt reg |
1240 | Register VReg = RegInfo.createVirtualRegister(RegClass: &XCore::GRRegsRegClass); |
1241 | RegInfo.addLiveIn(Reg: ArgRegs[i], vreg: VReg); |
1242 | SDValue Val = DAG.getCopyFromReg(Chain, dl, Reg: VReg, VT: MVT::i32); |
1243 | CFRegNode.push_back(Elt: Val.getValue(R: Val->getNumValues() - 1)); |
1244 | // Move argument from virt reg -> stack |
1245 | SDValue Store = |
1246 | DAG.getStore(Chain: Val.getValue(R: 1), dl, Val, Ptr: FIN, PtrInfo: MachinePointerInfo()); |
1247 | MemOps.push_back(Elt: Store); |
1248 | } |
1249 | } else { |
1250 | // This will point to the next argument passed via stack. |
1251 | XFI->setVarArgsFrameIndex( |
1252 | MFI.CreateFixedObject(Size: 4, SPOffset: LRSaveSize + CCInfo.getStackSize(), IsImmutable: true)); |
1253 | } |
1254 | } |
1255 | |
1256 | // 2. chain CopyFromReg nodes into a TokenFactor. |
1257 | if (!CFRegNode.empty()) |
1258 | Chain = DAG.getNode(Opcode: ISD::TokenFactor, DL: dl, VT: MVT::Other, Ops: CFRegNode); |
1259 | |
1260 | // 3. Memcpy 'byVal' args & push final InVals. |
1261 | // Aggregates passed "byVal" need to be copied by the callee. |
1262 | // The callee will use a pointer to this copy, rather than the original |
1263 | // pointer. |
1264 | for (const ArgDataPair &ArgDI : ArgData) { |
1265 | if (ArgDI.Flags.isByVal() && ArgDI.Flags.getByValSize()) { |
1266 | unsigned Size = ArgDI.Flags.getByValSize(); |
1267 | Align Alignment = |
1268 | std::max(a: Align(StackSlotSize), b: ArgDI.Flags.getNonZeroByValAlign()); |
1269 | // Create a new object on the stack and copy the pointee into it. |
1270 | int FI = MFI.CreateStackObject(Size, Alignment, isSpillSlot: false); |
1271 | SDValue FIN = DAG.getFrameIndex(FI, VT: MVT::i32); |
1272 | InVals.push_back(Elt: FIN); |
1273 | MemOps.push_back(Elt: DAG.getMemcpy( |
1274 | Chain, dl, Dst: FIN, Src: ArgDI.SDV, Size: DAG.getConstant(Val: Size, DL: dl, VT: MVT::i32), |
1275 | Alignment, isVol: false, AlwaysInline: false, /*CI=*/nullptr, OverrideTailCall: std::nullopt, |
1276 | DstPtrInfo: MachinePointerInfo(), SrcPtrInfo: MachinePointerInfo())); |
1277 | } else { |
1278 | InVals.push_back(Elt: ArgDI.SDV); |
1279 | } |
1280 | } |
1281 | |
1282 | // 4, chain mem ops nodes into a TokenFactor. |
1283 | if (!MemOps.empty()) { |
1284 | MemOps.push_back(Elt: Chain); |
1285 | Chain = DAG.getNode(Opcode: ISD::TokenFactor, DL: dl, VT: MVT::Other, Ops: MemOps); |
1286 | } |
1287 | |
1288 | return Chain; |
1289 | } |
1290 | |
1291 | //===----------------------------------------------------------------------===// |
1292 | // Return Value Calling Convention Implementation |
1293 | //===----------------------------------------------------------------------===// |
1294 | |
1295 | bool XCoreTargetLowering:: |
1296 | CanLowerReturn(CallingConv::ID CallConv, MachineFunction &MF, |
1297 | bool isVarArg, |
1298 | const SmallVectorImpl<ISD::OutputArg> &Outs, |
1299 | LLVMContext &Context, const Type *RetTy) const { |
1300 | SmallVector<CCValAssign, 16> RVLocs; |
1301 | CCState CCInfo(CallConv, isVarArg, MF, RVLocs, Context); |
1302 | if (!CCInfo.CheckReturn(Outs, Fn: RetCC_XCore)) |
1303 | return false; |
1304 | if (CCInfo.getStackSize() != 0 && isVarArg) |
1305 | return false; |
1306 | return true; |
1307 | } |
1308 | |
1309 | SDValue |
1310 | XCoreTargetLowering::LowerReturn(SDValue Chain, CallingConv::ID CallConv, |
1311 | bool isVarArg, |
1312 | const SmallVectorImpl<ISD::OutputArg> &Outs, |
1313 | const SmallVectorImpl<SDValue> &OutVals, |
1314 | const SDLoc &dl, SelectionDAG &DAG) const { |
1315 | |
1316 | XCoreFunctionInfo *XFI = |
1317 | DAG.getMachineFunction().getInfo<XCoreFunctionInfo>(); |
1318 | MachineFrameInfo &MFI = DAG.getMachineFunction().getFrameInfo(); |
1319 | |
1320 | // CCValAssign - represent the assignment of |
1321 | // the return value to a location |
1322 | SmallVector<CCValAssign, 16> RVLocs; |
1323 | |
1324 | // CCState - Info about the registers and stack slot. |
1325 | CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), RVLocs, |
1326 | *DAG.getContext()); |
1327 | |
1328 | // Analyze return values. |
1329 | if (!isVarArg) |
1330 | CCInfo.AllocateStack(Size: XFI->getReturnStackOffset(), Alignment: Align(4)); |
1331 | |
1332 | CCInfo.AnalyzeReturn(Outs, Fn: RetCC_XCore); |
1333 | |
1334 | SDValue Glue; |
1335 | SmallVector<SDValue, 4> RetOps(1, Chain); |
1336 | |
1337 | // Return on XCore is always a "retsp 0" |
1338 | RetOps.push_back(Elt: DAG.getConstant(Val: 0, DL: dl, VT: MVT::i32)); |
1339 | |
1340 | SmallVector<SDValue, 4> MemOpChains; |
1341 | // Handle return values that must be copied to memory. |
1342 | for (unsigned i = 0, e = RVLocs.size(); i != e; ++i) { |
1343 | CCValAssign &VA = RVLocs[i]; |
1344 | if (VA.isRegLoc()) |
1345 | continue; |
1346 | assert(VA.isMemLoc()); |
1347 | if (isVarArg) { |
1348 | report_fatal_error(reason: "Can't return value from vararg function in memory" ); |
1349 | } |
1350 | |
1351 | int Offset = VA.getLocMemOffset(); |
1352 | unsigned ObjSize = VA.getLocVT().getSizeInBits() / 8; |
1353 | // Create the frame index object for the memory location. |
1354 | int FI = MFI.CreateFixedObject(Size: ObjSize, SPOffset: Offset, IsImmutable: false); |
1355 | |
1356 | // Create a SelectionDAG node corresponding to a store |
1357 | // to this memory location. |
1358 | SDValue FIN = DAG.getFrameIndex(FI, VT: MVT::i32); |
1359 | MemOpChains.push_back(Elt: DAG.getStore( |
1360 | Chain, dl, Val: OutVals[i], Ptr: FIN, |
1361 | PtrInfo: MachinePointerInfo::getFixedStack(MF&: DAG.getMachineFunction(), FI))); |
1362 | } |
1363 | |
1364 | // Transform all store nodes into one single node because |
1365 | // all stores are independent of each other. |
1366 | if (!MemOpChains.empty()) |
1367 | Chain = DAG.getNode(Opcode: ISD::TokenFactor, DL: dl, VT: MVT::Other, Ops: MemOpChains); |
1368 | |
1369 | // Now handle return values copied to registers. |
1370 | for (unsigned i = 0, e = RVLocs.size(); i != e; ++i) { |
1371 | CCValAssign &VA = RVLocs[i]; |
1372 | if (!VA.isRegLoc()) |
1373 | continue; |
1374 | // Copy the result values into the output registers. |
1375 | Chain = DAG.getCopyToReg(Chain, dl, Reg: VA.getLocReg(), N: OutVals[i], Glue); |
1376 | |
1377 | // guarantee that all emitted copies are |
1378 | // stuck together, avoiding something bad |
1379 | Glue = Chain.getValue(R: 1); |
1380 | RetOps.push_back(Elt: DAG.getRegister(Reg: VA.getLocReg(), VT: VA.getLocVT())); |
1381 | } |
1382 | |
1383 | RetOps[0] = Chain; // Update chain. |
1384 | |
1385 | // Add the glue if we have it. |
1386 | if (Glue.getNode()) |
1387 | RetOps.push_back(Elt: Glue); |
1388 | |
1389 | return DAG.getNode(Opcode: XCoreISD::RETSP, DL: dl, VT: MVT::Other, Ops: RetOps); |
1390 | } |
1391 | |
1392 | //===----------------------------------------------------------------------===// |
1393 | // Other Lowering Code |
1394 | //===----------------------------------------------------------------------===// |
1395 | |
1396 | MachineBasicBlock * |
1397 | XCoreTargetLowering::EmitInstrWithCustomInserter(MachineInstr &MI, |
1398 | MachineBasicBlock *BB) const { |
1399 | const TargetInstrInfo &TII = *Subtarget.getInstrInfo(); |
1400 | DebugLoc dl = MI.getDebugLoc(); |
1401 | assert((MI.getOpcode() == XCore::SELECT_CC) && |
1402 | "Unexpected instr type to insert" ); |
1403 | |
1404 | // To "insert" a SELECT_CC instruction, we actually have to insert the diamond |
1405 | // control-flow pattern. The incoming instruction knows the destination vreg |
1406 | // to set, the condition code register to branch on, the true/false values to |
1407 | // select between, and a branch opcode to use. |
1408 | const BasicBlock *LLVM_BB = BB->getBasicBlock(); |
1409 | MachineFunction::iterator It = ++BB->getIterator(); |
1410 | |
1411 | // thisMBB: |
1412 | // ... |
1413 | // TrueVal = ... |
1414 | // cmpTY ccX, r1, r2 |
1415 | // bCC copy1MBB |
1416 | // fallthrough --> copy0MBB |
1417 | MachineBasicBlock *thisMBB = BB; |
1418 | MachineFunction *F = BB->getParent(); |
1419 | MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(BB: LLVM_BB); |
1420 | MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(BB: LLVM_BB); |
1421 | F->insert(MBBI: It, MBB: copy0MBB); |
1422 | F->insert(MBBI: It, MBB: sinkMBB); |
1423 | |
1424 | // Transfer the remainder of BB and its successor edges to sinkMBB. |
1425 | sinkMBB->splice(Where: sinkMBB->begin(), Other: BB, |
1426 | From: std::next(x: MachineBasicBlock::iterator(MI)), To: BB->end()); |
1427 | sinkMBB->transferSuccessorsAndUpdatePHIs(FromMBB: BB); |
1428 | |
1429 | // Next, add the true and fallthrough blocks as its successors. |
1430 | BB->addSuccessor(Succ: copy0MBB); |
1431 | BB->addSuccessor(Succ: sinkMBB); |
1432 | |
1433 | BuildMI(BB, MIMD: dl, MCID: TII.get(Opcode: XCore::BRFT_lru6)) |
1434 | .addReg(RegNo: MI.getOperand(i: 1).getReg()) |
1435 | .addMBB(MBB: sinkMBB); |
1436 | |
1437 | // copy0MBB: |
1438 | // %FalseValue = ... |
1439 | // # fallthrough to sinkMBB |
1440 | BB = copy0MBB; |
1441 | |
1442 | // Update machine-CFG edges |
1443 | BB->addSuccessor(Succ: sinkMBB); |
1444 | |
1445 | // sinkMBB: |
1446 | // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ] |
1447 | // ... |
1448 | BB = sinkMBB; |
1449 | BuildMI(BB&: *BB, I: BB->begin(), MIMD: dl, MCID: TII.get(Opcode: XCore::PHI), DestReg: MI.getOperand(i: 0).getReg()) |
1450 | .addReg(RegNo: MI.getOperand(i: 3).getReg()) |
1451 | .addMBB(MBB: copy0MBB) |
1452 | .addReg(RegNo: MI.getOperand(i: 2).getReg()) |
1453 | .addMBB(MBB: thisMBB); |
1454 | |
1455 | MI.eraseFromParent(); // The pseudo instruction is gone now. |
1456 | return BB; |
1457 | } |
1458 | |
1459 | //===----------------------------------------------------------------------===// |
1460 | // Target Optimization Hooks |
1461 | //===----------------------------------------------------------------------===// |
1462 | |
1463 | SDValue XCoreTargetLowering::PerformDAGCombine(SDNode *N, |
1464 | DAGCombinerInfo &DCI) const { |
1465 | SelectionDAG &DAG = DCI.DAG; |
1466 | SDLoc dl(N); |
1467 | switch (N->getOpcode()) { |
1468 | default: break; |
1469 | case ISD::INTRINSIC_VOID: |
1470 | switch (N->getConstantOperandVal(Num: 1)) { |
1471 | case Intrinsic::xcore_outt: |
1472 | case Intrinsic::xcore_outct: |
1473 | case Intrinsic::xcore_chkct: { |
1474 | SDValue OutVal = N->getOperand(Num: 3); |
1475 | // These instructions ignore the high bits. |
1476 | if (OutVal.hasOneUse()) { |
1477 | unsigned BitWidth = OutVal.getValueSizeInBits(); |
1478 | APInt DemandedMask = APInt::getLowBitsSet(numBits: BitWidth, loBitsSet: 8); |
1479 | KnownBits Known; |
1480 | TargetLowering::TargetLoweringOpt TLO(DAG, !DCI.isBeforeLegalize(), |
1481 | !DCI.isBeforeLegalizeOps()); |
1482 | const TargetLowering &TLI = DAG.getTargetLoweringInfo(); |
1483 | if (TLI.ShrinkDemandedConstant(Op: OutVal, DemandedBits: DemandedMask, TLO) || |
1484 | TLI.SimplifyDemandedBits(Op: OutVal, DemandedBits: DemandedMask, Known, TLO)) |
1485 | DCI.CommitTargetLoweringOpt(TLO); |
1486 | } |
1487 | break; |
1488 | } |
1489 | case Intrinsic::xcore_setpt: { |
1490 | SDValue Time = N->getOperand(Num: 3); |
1491 | // This instruction ignores the high bits. |
1492 | if (Time.hasOneUse()) { |
1493 | unsigned BitWidth = Time.getValueSizeInBits(); |
1494 | APInt DemandedMask = APInt::getLowBitsSet(numBits: BitWidth, loBitsSet: 16); |
1495 | KnownBits Known; |
1496 | TargetLowering::TargetLoweringOpt TLO(DAG, !DCI.isBeforeLegalize(), |
1497 | !DCI.isBeforeLegalizeOps()); |
1498 | const TargetLowering &TLI = DAG.getTargetLoweringInfo(); |
1499 | if (TLI.ShrinkDemandedConstant(Op: Time, DemandedBits: DemandedMask, TLO) || |
1500 | TLI.SimplifyDemandedBits(Op: Time, DemandedBits: DemandedMask, Known, TLO)) |
1501 | DCI.CommitTargetLoweringOpt(TLO); |
1502 | } |
1503 | break; |
1504 | } |
1505 | } |
1506 | break; |
1507 | case XCoreISD::LADD: { |
1508 | SDValue N0 = N->getOperand(Num: 0); |
1509 | SDValue N1 = N->getOperand(Num: 1); |
1510 | SDValue N2 = N->getOperand(Num: 2); |
1511 | ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(Val&: N0); |
1512 | ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(Val&: N1); |
1513 | EVT VT = N0.getValueType(); |
1514 | |
1515 | // canonicalize constant to RHS |
1516 | if (N0C && !N1C) |
1517 | return DAG.getNode(Opcode: XCoreISD::LADD, DL: dl, VTList: DAG.getVTList(VT1: VT, VT2: VT), N1, N2: N0, N3: N2); |
1518 | |
1519 | // fold (ladd 0, 0, x) -> 0, x & 1 |
1520 | if (N0C && N0C->isZero() && N1C && N1C->isZero()) { |
1521 | SDValue Carry = DAG.getConstant(Val: 0, DL: dl, VT); |
1522 | SDValue Result = DAG.getNode(Opcode: ISD::AND, DL: dl, VT, N1: N2, |
1523 | N2: DAG.getConstant(Val: 1, DL: dl, VT)); |
1524 | SDValue Ops[] = { Result, Carry }; |
1525 | return DAG.getMergeValues(Ops, dl); |
1526 | } |
1527 | |
1528 | // fold (ladd x, 0, y) -> 0, add x, y iff carry is unused and y has only the |
1529 | // low bit set |
1530 | if (N1C && N1C->isZero() && N->hasNUsesOfValue(NUses: 0, Value: 1)) { |
1531 | APInt Mask = APInt::getHighBitsSet(numBits: VT.getSizeInBits(), |
1532 | hiBitsSet: VT.getSizeInBits() - 1); |
1533 | KnownBits Known = DAG.computeKnownBits(Op: N2); |
1534 | if ((Known.Zero & Mask) == Mask) { |
1535 | SDValue Carry = DAG.getConstant(Val: 0, DL: dl, VT); |
1536 | SDValue Result = DAG.getNode(Opcode: ISD::ADD, DL: dl, VT, N1: N0, N2); |
1537 | SDValue Ops[] = { Result, Carry }; |
1538 | return DAG.getMergeValues(Ops, dl); |
1539 | } |
1540 | } |
1541 | } |
1542 | break; |
1543 | case XCoreISD::LSUB: { |
1544 | SDValue N0 = N->getOperand(Num: 0); |
1545 | SDValue N1 = N->getOperand(Num: 1); |
1546 | SDValue N2 = N->getOperand(Num: 2); |
1547 | ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(Val&: N0); |
1548 | ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(Val&: N1); |
1549 | EVT VT = N0.getValueType(); |
1550 | |
1551 | // fold (lsub 0, 0, x) -> x, -x iff x has only the low bit set |
1552 | if (N0C && N0C->isZero() && N1C && N1C->isZero()) { |
1553 | APInt Mask = APInt::getHighBitsSet(numBits: VT.getSizeInBits(), |
1554 | hiBitsSet: VT.getSizeInBits() - 1); |
1555 | KnownBits Known = DAG.computeKnownBits(Op: N2); |
1556 | if ((Known.Zero & Mask) == Mask) { |
1557 | SDValue Borrow = N2; |
1558 | SDValue Result = DAG.getNode(Opcode: ISD::SUB, DL: dl, VT, |
1559 | N1: DAG.getConstant(Val: 0, DL: dl, VT), N2); |
1560 | SDValue Ops[] = { Result, Borrow }; |
1561 | return DAG.getMergeValues(Ops, dl); |
1562 | } |
1563 | } |
1564 | |
1565 | // fold (lsub x, 0, y) -> 0, sub x, y iff borrow is unused and y has only the |
1566 | // low bit set |
1567 | if (N1C && N1C->isZero() && N->hasNUsesOfValue(NUses: 0, Value: 1)) { |
1568 | APInt Mask = APInt::getHighBitsSet(numBits: VT.getSizeInBits(), |
1569 | hiBitsSet: VT.getSizeInBits() - 1); |
1570 | KnownBits Known = DAG.computeKnownBits(Op: N2); |
1571 | if ((Known.Zero & Mask) == Mask) { |
1572 | SDValue Borrow = DAG.getConstant(Val: 0, DL: dl, VT); |
1573 | SDValue Result = DAG.getNode(Opcode: ISD::SUB, DL: dl, VT, N1: N0, N2); |
1574 | SDValue Ops[] = { Result, Borrow }; |
1575 | return DAG.getMergeValues(Ops, dl); |
1576 | } |
1577 | } |
1578 | } |
1579 | break; |
1580 | case XCoreISD::LMUL: { |
1581 | SDValue N0 = N->getOperand(Num: 0); |
1582 | SDValue N1 = N->getOperand(Num: 1); |
1583 | SDValue N2 = N->getOperand(Num: 2); |
1584 | SDValue N3 = N->getOperand(Num: 3); |
1585 | ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(Val&: N0); |
1586 | ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(Val&: N1); |
1587 | EVT VT = N0.getValueType(); |
1588 | // Canonicalize multiplicative constant to RHS. If both multiplicative |
1589 | // operands are constant canonicalize smallest to RHS. |
1590 | if ((N0C && !N1C) || |
1591 | (N0C && N1C && N0C->getZExtValue() < N1C->getZExtValue())) |
1592 | return DAG.getNode(Opcode: XCoreISD::LMUL, DL: dl, VTList: DAG.getVTList(VT1: VT, VT2: VT), |
1593 | N1, N2: N0, N3: N2, N4: N3); |
1594 | |
1595 | // lmul(x, 0, a, b) |
1596 | if (N1C && N1C->isZero()) { |
1597 | // If the high result is unused fold to add(a, b) |
1598 | if (N->hasNUsesOfValue(NUses: 0, Value: 0)) { |
1599 | SDValue Lo = DAG.getNode(Opcode: ISD::ADD, DL: dl, VT, N1: N2, N2: N3); |
1600 | SDValue Ops[] = { Lo, Lo }; |
1601 | return DAG.getMergeValues(Ops, dl); |
1602 | } |
1603 | // Otherwise fold to ladd(a, b, 0) |
1604 | SDValue Result = |
1605 | DAG.getNode(Opcode: XCoreISD::LADD, DL: dl, VTList: DAG.getVTList(VT1: VT, VT2: VT), N1: N2, N2: N3, N3: N1); |
1606 | SDValue Carry(Result.getNode(), 1); |
1607 | SDValue Ops[] = { Carry, Result }; |
1608 | return DAG.getMergeValues(Ops, dl); |
1609 | } |
1610 | } |
1611 | break; |
1612 | case ISD::ADD: { |
1613 | // Fold 32 bit expressions such as add(add(mul(x,y),a),b) -> |
1614 | // lmul(x, y, a, b). The high result of lmul will be ignored. |
1615 | // This is only profitable if the intermediate results are unused |
1616 | // elsewhere. |
1617 | SDValue Mul0, Mul1, Addend0, Addend1; |
1618 | if (N->getValueType(ResNo: 0) == MVT::i32 && |
1619 | isADDADDMUL(Op: SDValue(N, 0), Mul0, Mul1, Addend0, Addend1, requireIntermediatesHaveOneUse: true)) { |
1620 | SDValue Ignored = DAG.getNode(Opcode: XCoreISD::LMUL, DL: dl, |
1621 | VTList: DAG.getVTList(VT1: MVT::i32, VT2: MVT::i32), N1: Mul0, |
1622 | N2: Mul1, N3: Addend0, N4: Addend1); |
1623 | SDValue Result(Ignored.getNode(), 1); |
1624 | return Result; |
1625 | } |
1626 | APInt HighMask = APInt::getHighBitsSet(numBits: 64, hiBitsSet: 32); |
1627 | // Fold 64 bit expression such as add(add(mul(x,y),a),b) -> |
1628 | // lmul(x, y, a, b) if all operands are zero-extended. We do this |
1629 | // before type legalization as it is messy to match the operands after |
1630 | // that. |
1631 | if (N->getValueType(ResNo: 0) == MVT::i64 && |
1632 | isADDADDMUL(Op: SDValue(N, 0), Mul0, Mul1, Addend0, Addend1, requireIntermediatesHaveOneUse: false) && |
1633 | DAG.MaskedValueIsZero(Op: Mul0, Mask: HighMask) && |
1634 | DAG.MaskedValueIsZero(Op: Mul1, Mask: HighMask) && |
1635 | DAG.MaskedValueIsZero(Op: Addend0, Mask: HighMask) && |
1636 | DAG.MaskedValueIsZero(Op: Addend1, Mask: HighMask)) { |
1637 | SDValue Mul0L = DAG.getNode(Opcode: ISD::EXTRACT_ELEMENT, DL: dl, VT: MVT::i32, |
1638 | N1: Mul0, N2: DAG.getConstant(Val: 0, DL: dl, VT: MVT::i32)); |
1639 | SDValue Mul1L = DAG.getNode(Opcode: ISD::EXTRACT_ELEMENT, DL: dl, VT: MVT::i32, |
1640 | N1: Mul1, N2: DAG.getConstant(Val: 0, DL: dl, VT: MVT::i32)); |
1641 | SDValue Addend0L = DAG.getNode(Opcode: ISD::EXTRACT_ELEMENT, DL: dl, VT: MVT::i32, |
1642 | N1: Addend0, N2: DAG.getConstant(Val: 0, DL: dl, VT: MVT::i32)); |
1643 | SDValue Addend1L = DAG.getNode(Opcode: ISD::EXTRACT_ELEMENT, DL: dl, VT: MVT::i32, |
1644 | N1: Addend1, N2: DAG.getConstant(Val: 0, DL: dl, VT: MVT::i32)); |
1645 | SDValue Hi = DAG.getNode(Opcode: XCoreISD::LMUL, DL: dl, |
1646 | VTList: DAG.getVTList(VT1: MVT::i32, VT2: MVT::i32), N1: Mul0L, N2: Mul1L, |
1647 | N3: Addend0L, N4: Addend1L); |
1648 | SDValue Lo(Hi.getNode(), 1); |
1649 | return DAG.getNode(Opcode: ISD::BUILD_PAIR, DL: dl, VT: MVT::i64, N1: Lo, N2: Hi); |
1650 | } |
1651 | } |
1652 | break; |
1653 | case ISD::STORE: { |
1654 | // Replace unaligned store of unaligned load with memmove. |
1655 | StoreSDNode *ST = cast<StoreSDNode>(Val: N); |
1656 | if (!DCI.isBeforeLegalize() || |
1657 | allowsMemoryAccessForAlignment(Context&: *DAG.getContext(), DL: DAG.getDataLayout(), |
1658 | VT: ST->getMemoryVT(), |
1659 | MMO: *ST->getMemOperand()) || |
1660 | ST->isVolatile() || ST->isIndexed()) { |
1661 | break; |
1662 | } |
1663 | SDValue Chain = ST->getChain(); |
1664 | |
1665 | unsigned StoreBits = ST->getMemoryVT().getStoreSizeInBits(); |
1666 | assert((StoreBits % 8) == 0 && |
1667 | "Store size in bits must be a multiple of 8" ); |
1668 | Align Alignment = ST->getAlign(); |
1669 | |
1670 | if (LoadSDNode *LD = dyn_cast<LoadSDNode>(Val: ST->getValue())) { |
1671 | if (LD->hasNUsesOfValue(NUses: 1, Value: 0) && ST->getMemoryVT() == LD->getMemoryVT() && |
1672 | LD->getAlign() == Alignment && |
1673 | !LD->isVolatile() && !LD->isIndexed() && |
1674 | Chain.reachesChainWithoutSideEffects(Dest: SDValue(LD, 1))) { |
1675 | bool isTail = isInTailCallPosition(DAG, Node: ST, Chain); |
1676 | return DAG.getMemmove(Chain, dl, Dst: ST->getBasePtr(), Src: LD->getBasePtr(), |
1677 | Size: DAG.getConstant(Val: StoreBits / 8, DL: dl, VT: MVT::i32), |
1678 | Alignment, isVol: false, CI: nullptr, OverrideTailCall: isTail, |
1679 | DstPtrInfo: ST->getPointerInfo(), SrcPtrInfo: LD->getPointerInfo()); |
1680 | } |
1681 | } |
1682 | break; |
1683 | } |
1684 | } |
1685 | return SDValue(); |
1686 | } |
1687 | |
1688 | void XCoreTargetLowering::computeKnownBitsForTargetNode(const SDValue Op, |
1689 | KnownBits &Known, |
1690 | const APInt &DemandedElts, |
1691 | const SelectionDAG &DAG, |
1692 | unsigned Depth) const { |
1693 | Known.resetAll(); |
1694 | switch (Op.getOpcode()) { |
1695 | default: break; |
1696 | case XCoreISD::LADD: |
1697 | case XCoreISD::LSUB: |
1698 | if (Op.getResNo() == 1) { |
1699 | // Top bits of carry / borrow are clear. |
1700 | Known.Zero = APInt::getHighBitsSet(numBits: Known.getBitWidth(), |
1701 | hiBitsSet: Known.getBitWidth() - 1); |
1702 | } |
1703 | break; |
1704 | case ISD::INTRINSIC_W_CHAIN: |
1705 | { |
1706 | unsigned IntNo = Op.getConstantOperandVal(i: 1); |
1707 | switch (IntNo) { |
1708 | case Intrinsic::xcore_getts: |
1709 | // High bits are known to be zero. |
1710 | Known.Zero = |
1711 | APInt::getHighBitsSet(numBits: Known.getBitWidth(), hiBitsSet: Known.getBitWidth() - 16); |
1712 | break; |
1713 | case Intrinsic::xcore_int: |
1714 | case Intrinsic::xcore_inct: |
1715 | // High bits are known to be zero. |
1716 | Known.Zero = |
1717 | APInt::getHighBitsSet(numBits: Known.getBitWidth(), hiBitsSet: Known.getBitWidth() - 8); |
1718 | break; |
1719 | case Intrinsic::xcore_testct: |
1720 | // Result is either 0 or 1. |
1721 | Known.Zero = |
1722 | APInt::getHighBitsSet(numBits: Known.getBitWidth(), hiBitsSet: Known.getBitWidth() - 1); |
1723 | break; |
1724 | case Intrinsic::xcore_testwct: |
1725 | // Result is in the range 0 - 4. |
1726 | Known.Zero = |
1727 | APInt::getHighBitsSet(numBits: Known.getBitWidth(), hiBitsSet: Known.getBitWidth() - 3); |
1728 | break; |
1729 | } |
1730 | } |
1731 | break; |
1732 | } |
1733 | } |
1734 | |
1735 | //===----------------------------------------------------------------------===// |
1736 | // Addressing mode description hooks |
1737 | //===----------------------------------------------------------------------===// |
1738 | |
1739 | static inline bool isImmUs(int64_t val) |
1740 | { |
1741 | return (val >= 0 && val <= 11); |
1742 | } |
1743 | |
1744 | static inline bool isImmUs2(int64_t val) |
1745 | { |
1746 | return (val%2 == 0 && isImmUs(val: val/2)); |
1747 | } |
1748 | |
1749 | static inline bool isImmUs4(int64_t val) |
1750 | { |
1751 | return (val%4 == 0 && isImmUs(val: val/4)); |
1752 | } |
1753 | |
1754 | /// isLegalAddressingMode - Return true if the addressing mode represented |
1755 | /// by AM is legal for this target, for a load/store of the specified type. |
1756 | bool XCoreTargetLowering::isLegalAddressingMode(const DataLayout &DL, |
1757 | const AddrMode &AM, Type *Ty, |
1758 | unsigned AS, |
1759 | Instruction *I) const { |
1760 | if (Ty->getTypeID() == Type::VoidTyID) |
1761 | return AM.Scale == 0 && isImmUs(val: AM.BaseOffs) && isImmUs4(val: AM.BaseOffs); |
1762 | |
1763 | unsigned Size = DL.getTypeAllocSize(Ty); |
1764 | if (AM.BaseGV) { |
1765 | return Size >= 4 && !AM.HasBaseReg && AM.Scale == 0 && |
1766 | AM.BaseOffs%4 == 0; |
1767 | } |
1768 | |
1769 | switch (Size) { |
1770 | case 1: |
1771 | // reg + imm |
1772 | if (AM.Scale == 0) { |
1773 | return isImmUs(val: AM.BaseOffs); |
1774 | } |
1775 | // reg + reg |
1776 | return AM.Scale == 1 && AM.BaseOffs == 0; |
1777 | case 2: |
1778 | case 3: |
1779 | // reg + imm |
1780 | if (AM.Scale == 0) { |
1781 | return isImmUs2(val: AM.BaseOffs); |
1782 | } |
1783 | // reg + reg<<1 |
1784 | return AM.Scale == 2 && AM.BaseOffs == 0; |
1785 | default: |
1786 | // reg + imm |
1787 | if (AM.Scale == 0) { |
1788 | return isImmUs4(val: AM.BaseOffs); |
1789 | } |
1790 | // reg + reg<<2 |
1791 | return AM.Scale == 4 && AM.BaseOffs == 0; |
1792 | } |
1793 | } |
1794 | |
1795 | //===----------------------------------------------------------------------===// |
1796 | // XCore Inline Assembly Support |
1797 | //===----------------------------------------------------------------------===// |
1798 | |
1799 | std::pair<unsigned, const TargetRegisterClass *> |
1800 | XCoreTargetLowering::getRegForInlineAsmConstraint(const TargetRegisterInfo *TRI, |
1801 | StringRef Constraint, |
1802 | MVT VT) const { |
1803 | if (Constraint.size() == 1) { |
1804 | switch (Constraint[0]) { |
1805 | default : break; |
1806 | case 'r': |
1807 | return std::make_pair(x: 0U, y: &XCore::GRRegsRegClass); |
1808 | } |
1809 | } |
1810 | // Use the default implementation in TargetLowering to convert the register |
1811 | // constraint into a member of a register class. |
1812 | return TargetLowering::getRegForInlineAsmConstraint(TRI, Constraint, VT); |
1813 | } |
1814 | |