1 | //===- LegalizeDAG.cpp - Implement SelectionDAG::Legalize -----------------===// |
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 SelectionDAG::Legalize method. |
10 | // |
11 | //===----------------------------------------------------------------------===// |
12 | |
13 | #include "llvm/ADT/APFloat.h" |
14 | #include "llvm/ADT/APInt.h" |
15 | #include "llvm/ADT/ArrayRef.h" |
16 | #include "llvm/ADT/FloatingPointMode.h" |
17 | #include "llvm/ADT/SetVector.h" |
18 | #include "llvm/ADT/SmallPtrSet.h" |
19 | #include "llvm/ADT/SmallSet.h" |
20 | #include "llvm/ADT/SmallVector.h" |
21 | #include "llvm/Analysis/ConstantFolding.h" |
22 | #include "llvm/Analysis/TargetLibraryInfo.h" |
23 | #include "llvm/CodeGen/ISDOpcodes.h" |
24 | #include "llvm/CodeGen/MachineFrameInfo.h" |
25 | #include "llvm/CodeGen/MachineFunction.h" |
26 | #include "llvm/CodeGen/MachineJumpTableInfo.h" |
27 | #include "llvm/CodeGen/MachineMemOperand.h" |
28 | #include "llvm/CodeGen/RuntimeLibcallUtil.h" |
29 | #include "llvm/CodeGen/SelectionDAG.h" |
30 | #include "llvm/CodeGen/SelectionDAGNodes.h" |
31 | #include "llvm/CodeGen/TargetFrameLowering.h" |
32 | #include "llvm/CodeGen/TargetLowering.h" |
33 | #include "llvm/CodeGen/TargetSubtargetInfo.h" |
34 | #include "llvm/CodeGen/ValueTypes.h" |
35 | #include "llvm/CodeGenTypes/MachineValueType.h" |
36 | #include "llvm/IR/CallingConv.h" |
37 | #include "llvm/IR/Constants.h" |
38 | #include "llvm/IR/DataLayout.h" |
39 | #include "llvm/IR/DerivedTypes.h" |
40 | #include "llvm/IR/Function.h" |
41 | #include "llvm/IR/Metadata.h" |
42 | #include "llvm/IR/Type.h" |
43 | #include "llvm/Support/Casting.h" |
44 | #include "llvm/Support/Compiler.h" |
45 | #include "llvm/Support/Debug.h" |
46 | #include "llvm/Support/ErrorHandling.h" |
47 | #include "llvm/Support/MathExtras.h" |
48 | #include "llvm/Support/raw_ostream.h" |
49 | #include "llvm/Target/TargetMachine.h" |
50 | #include "llvm/Target/TargetOptions.h" |
51 | #include <cassert> |
52 | #include <cstdint> |
53 | #include <tuple> |
54 | #include <utility> |
55 | |
56 | using namespace llvm; |
57 | |
58 | #define DEBUG_TYPE "legalizedag" |
59 | |
60 | namespace { |
61 | |
62 | /// Keeps track of state when getting the sign of a floating-point value as an |
63 | /// integer. |
64 | struct FloatSignAsInt { |
65 | EVT FloatVT; |
66 | SDValue Chain; |
67 | SDValue FloatPtr; |
68 | SDValue IntPtr; |
69 | MachinePointerInfo IntPointerInfo; |
70 | MachinePointerInfo FloatPointerInfo; |
71 | SDValue IntValue; |
72 | APInt SignMask; |
73 | uint8_t SignBit; |
74 | }; |
75 | |
76 | //===----------------------------------------------------------------------===// |
77 | /// This takes an arbitrary SelectionDAG as input and |
78 | /// hacks on it until the target machine can handle it. This involves |
79 | /// eliminating value sizes the machine cannot handle (promoting small sizes to |
80 | /// large sizes or splitting up large values into small values) as well as |
81 | /// eliminating operations the machine cannot handle. |
82 | /// |
83 | /// This code also does a small amount of optimization and recognition of idioms |
84 | /// as part of its processing. For example, if a target does not support a |
85 | /// 'setcc' instruction efficiently, but does support 'brcc' instruction, this |
86 | /// will attempt merge setcc and brc instructions into brcc's. |
87 | class SelectionDAGLegalize { |
88 | const TargetMachine &TM; |
89 | const TargetLowering &TLI; |
90 | SelectionDAG &DAG; |
91 | |
92 | /// The set of nodes which have already been legalized. We hold a |
93 | /// reference to it in order to update as necessary on node deletion. |
94 | SmallPtrSetImpl<SDNode *> &LegalizedNodes; |
95 | |
96 | /// A set of all the nodes updated during legalization. |
97 | SmallSetVector<SDNode *, 16> *UpdatedNodes; |
98 | |
99 | EVT getSetCCResultType(EVT VT) const { |
100 | return TLI.getSetCCResultType(DL: DAG.getDataLayout(), Context&: *DAG.getContext(), VT); |
101 | } |
102 | |
103 | // Libcall insertion helpers. |
104 | |
105 | public: |
106 | SelectionDAGLegalize(SelectionDAG &DAG, |
107 | SmallPtrSetImpl<SDNode *> &LegalizedNodes, |
108 | SmallSetVector<SDNode *, 16> *UpdatedNodes = nullptr) |
109 | : TM(DAG.getTarget()), TLI(DAG.getTargetLoweringInfo()), DAG(DAG), |
110 | LegalizedNodes(LegalizedNodes), UpdatedNodes(UpdatedNodes) {} |
111 | |
112 | /// Legalizes the given operation. |
113 | void LegalizeOp(SDNode *Node); |
114 | |
115 | private: |
116 | SDValue OptimizeFloatStore(StoreSDNode *ST); |
117 | |
118 | void LegalizeLoadOps(SDNode *Node); |
119 | void LegalizeStoreOps(SDNode *Node); |
120 | |
121 | SDValue ExpandINSERT_VECTOR_ELT(SDValue Op); |
122 | |
123 | /// Return a vector shuffle operation which |
124 | /// performs the same shuffe in terms of order or result bytes, but on a type |
125 | /// whose vector element type is narrower than the original shuffle type. |
126 | /// e.g. <v4i32> <0, 1, 0, 1> -> v8i16 <0, 1, 2, 3, 0, 1, 2, 3> |
127 | SDValue ShuffleWithNarrowerEltType(EVT NVT, EVT VT, const SDLoc &dl, |
128 | SDValue N1, SDValue N2, |
129 | ArrayRef<int> Mask) const; |
130 | |
131 | std::pair<SDValue, SDValue> ExpandLibCall(RTLIB::Libcall LC, SDNode *Node, |
132 | TargetLowering::ArgListTy &&Args, |
133 | bool IsSigned, EVT RetVT); |
134 | std::pair<SDValue, SDValue> ExpandLibCall(RTLIB::Libcall LC, SDNode *Node, bool isSigned); |
135 | |
136 | void ExpandFPLibCall(SDNode *Node, RTLIB::Libcall LC, |
137 | SmallVectorImpl<SDValue> &Results); |
138 | void ExpandFPLibCall(SDNode *Node, RTLIB::Libcall Call_F32, |
139 | RTLIB::Libcall Call_F64, RTLIB::Libcall Call_F80, |
140 | RTLIB::Libcall Call_F128, |
141 | RTLIB::Libcall Call_PPCF128, |
142 | SmallVectorImpl<SDValue> &Results); |
143 | SDValue ExpandIntLibCall(SDNode *Node, bool isSigned, |
144 | RTLIB::Libcall Call_I8, |
145 | RTLIB::Libcall Call_I16, |
146 | RTLIB::Libcall Call_I32, |
147 | RTLIB::Libcall Call_I64, |
148 | RTLIB::Libcall Call_I128); |
149 | void ExpandArgFPLibCall(SDNode *Node, |
150 | RTLIB::Libcall Call_F32, RTLIB::Libcall Call_F64, |
151 | RTLIB::Libcall Call_F80, RTLIB::Libcall Call_F128, |
152 | RTLIB::Libcall Call_PPCF128, |
153 | SmallVectorImpl<SDValue> &Results); |
154 | SDValue ExpandBitCountingLibCall(SDNode *Node, RTLIB::Libcall CallI32, |
155 | RTLIB::Libcall CallI64, |
156 | RTLIB::Libcall CallI128); |
157 | void ExpandDivRemLibCall(SDNode *Node, SmallVectorImpl<SDValue> &Results); |
158 | void ExpandSinCosLibCall(SDNode *Node, SmallVectorImpl<SDValue> &Results); |
159 | |
160 | SDValue EmitStackConvert(SDValue SrcOp, EVT SlotVT, EVT DestVT, |
161 | const SDLoc &dl); |
162 | SDValue EmitStackConvert(SDValue SrcOp, EVT SlotVT, EVT DestVT, |
163 | const SDLoc &dl, SDValue ChainIn); |
164 | SDValue ExpandBUILD_VECTOR(SDNode *Node); |
165 | SDValue ExpandSPLAT_VECTOR(SDNode *Node); |
166 | SDValue ExpandSCALAR_TO_VECTOR(SDNode *Node); |
167 | void ExpandDYNAMIC_STACKALLOC(SDNode *Node, |
168 | SmallVectorImpl<SDValue> &Results); |
169 | void getSignAsIntValue(FloatSignAsInt &State, const SDLoc &DL, |
170 | SDValue Value) const; |
171 | SDValue modifySignAsInt(const FloatSignAsInt &State, const SDLoc &DL, |
172 | SDValue NewIntValue) const; |
173 | SDValue ExpandFCOPYSIGN(SDNode *Node) const; |
174 | SDValue ExpandFABS(SDNode *Node) const; |
175 | SDValue ExpandFNEG(SDNode *Node) const; |
176 | SDValue expandLdexp(SDNode *Node) const; |
177 | SDValue expandFrexp(SDNode *Node) const; |
178 | |
179 | SDValue ExpandLegalINT_TO_FP(SDNode *Node, SDValue &Chain); |
180 | void PromoteLegalINT_TO_FP(SDNode *N, const SDLoc &dl, |
181 | SmallVectorImpl<SDValue> &Results); |
182 | void PromoteLegalFP_TO_INT(SDNode *N, const SDLoc &dl, |
183 | SmallVectorImpl<SDValue> &Results); |
184 | SDValue PromoteLegalFP_TO_INT_SAT(SDNode *Node, const SDLoc &dl); |
185 | |
186 | /// Implements vector reduce operation promotion. |
187 | /// |
188 | /// All vector operands are promoted to a vector type with larger element |
189 | /// type, and the start value is promoted to a larger scalar type. Then the |
190 | /// result is truncated back to the original scalar type. |
191 | SDValue PromoteReduction(SDNode *Node); |
192 | |
193 | SDValue ExpandPARITY(SDValue Op, const SDLoc &dl); |
194 | |
195 | SDValue ExpandExtractFromVectorThroughStack(SDValue Op); |
196 | SDValue ExpandInsertToVectorThroughStack(SDValue Op); |
197 | SDValue ExpandVectorBuildThroughStack(SDNode* Node); |
198 | SDValue ExpandConcatVectors(SDNode *Node); |
199 | |
200 | SDValue ExpandConstantFP(ConstantFPSDNode *CFP, bool UseCP); |
201 | SDValue ExpandConstant(ConstantSDNode *CP); |
202 | |
203 | // if ExpandNode returns false, LegalizeOp falls back to ConvertNodeToLibcall |
204 | bool ExpandNode(SDNode *Node); |
205 | void ConvertNodeToLibcall(SDNode *Node); |
206 | void PromoteNode(SDNode *Node); |
207 | |
208 | public: |
209 | // Node replacement helpers |
210 | |
211 | void ReplacedNode(SDNode *N) { |
212 | LegalizedNodes.erase(Ptr: N); |
213 | if (UpdatedNodes) |
214 | UpdatedNodes->insert(X: N); |
215 | } |
216 | |
217 | void ReplaceNode(SDNode *Old, SDNode *New) { |
218 | LLVM_DEBUG(dbgs() << " ... replacing: " ; Old->dump(&DAG); |
219 | dbgs() << " with: " ; New->dump(&DAG)); |
220 | |
221 | assert(Old->getNumValues() == New->getNumValues() && |
222 | "Replacing one node with another that produces a different number " |
223 | "of values!" ); |
224 | DAG.ReplaceAllUsesWith(From: Old, To: New); |
225 | if (UpdatedNodes) |
226 | UpdatedNodes->insert(X: New); |
227 | ReplacedNode(N: Old); |
228 | } |
229 | |
230 | void ReplaceNode(SDValue Old, SDValue New) { |
231 | LLVM_DEBUG(dbgs() << " ... replacing: " ; Old->dump(&DAG); |
232 | dbgs() << " with: " ; New->dump(&DAG)); |
233 | |
234 | DAG.ReplaceAllUsesWith(From: Old, To: New); |
235 | if (UpdatedNodes) |
236 | UpdatedNodes->insert(X: New.getNode()); |
237 | ReplacedNode(N: Old.getNode()); |
238 | } |
239 | |
240 | void ReplaceNode(SDNode *Old, const SDValue *New) { |
241 | LLVM_DEBUG(dbgs() << " ... replacing: " ; Old->dump(&DAG)); |
242 | |
243 | DAG.ReplaceAllUsesWith(From: Old, To: New); |
244 | for (unsigned i = 0, e = Old->getNumValues(); i != e; ++i) { |
245 | LLVM_DEBUG(dbgs() << (i == 0 ? " with: " : " and: " ); |
246 | New[i]->dump(&DAG)); |
247 | if (UpdatedNodes) |
248 | UpdatedNodes->insert(X: New[i].getNode()); |
249 | } |
250 | ReplacedNode(N: Old); |
251 | } |
252 | |
253 | void ReplaceNodeWithValue(SDValue Old, SDValue New) { |
254 | LLVM_DEBUG(dbgs() << " ... replacing: " ; Old->dump(&DAG); |
255 | dbgs() << " with: " ; New->dump(&DAG)); |
256 | |
257 | DAG.ReplaceAllUsesOfValueWith(From: Old, To: New); |
258 | if (UpdatedNodes) |
259 | UpdatedNodes->insert(X: New.getNode()); |
260 | ReplacedNode(N: Old.getNode()); |
261 | } |
262 | }; |
263 | |
264 | } // end anonymous namespace |
265 | |
266 | // Helper function that generates an MMO that considers the alignment of the |
267 | // stack, and the size of the stack object |
268 | static MachineMemOperand *getStackAlignedMMO(SDValue StackPtr, |
269 | MachineFunction &MF, |
270 | bool isObjectScalable) { |
271 | auto &MFI = MF.getFrameInfo(); |
272 | int FI = cast<FrameIndexSDNode>(Val&: StackPtr)->getIndex(); |
273 | MachinePointerInfo PtrInfo = MachinePointerInfo::getFixedStack(MF, FI); |
274 | LocationSize ObjectSize = isObjectScalable |
275 | ? LocationSize::beforeOrAfterPointer() |
276 | : LocationSize::precise(Value: MFI.getObjectSize(ObjectIdx: FI)); |
277 | return MF.getMachineMemOperand(PtrInfo, F: MachineMemOperand::MOStore, |
278 | Size: ObjectSize, BaseAlignment: MFI.getObjectAlign(ObjectIdx: FI)); |
279 | } |
280 | |
281 | /// Return a vector shuffle operation which |
282 | /// performs the same shuffle in terms of order or result bytes, but on a type |
283 | /// whose vector element type is narrower than the original shuffle type. |
284 | /// e.g. <v4i32> <0, 1, 0, 1> -> v8i16 <0, 1, 2, 3, 0, 1, 2, 3> |
285 | SDValue SelectionDAGLegalize::ShuffleWithNarrowerEltType( |
286 | EVT NVT, EVT VT, const SDLoc &dl, SDValue N1, SDValue N2, |
287 | ArrayRef<int> Mask) const { |
288 | unsigned NumMaskElts = VT.getVectorNumElements(); |
289 | unsigned NumDestElts = NVT.getVectorNumElements(); |
290 | unsigned NumEltsGrowth = NumDestElts / NumMaskElts; |
291 | |
292 | assert(NumEltsGrowth && "Cannot promote to vector type with fewer elts!" ); |
293 | |
294 | if (NumEltsGrowth == 1) |
295 | return DAG.getVectorShuffle(VT: NVT, dl, N1, N2, Mask); |
296 | |
297 | SmallVector<int, 8> NewMask; |
298 | for (unsigned i = 0; i != NumMaskElts; ++i) { |
299 | int Idx = Mask[i]; |
300 | for (unsigned j = 0; j != NumEltsGrowth; ++j) { |
301 | if (Idx < 0) |
302 | NewMask.push_back(Elt: -1); |
303 | else |
304 | NewMask.push_back(Elt: Idx * NumEltsGrowth + j); |
305 | } |
306 | } |
307 | assert(NewMask.size() == NumDestElts && "Non-integer NumEltsGrowth?" ); |
308 | assert(TLI.isShuffleMaskLegal(NewMask, NVT) && "Shuffle not legal?" ); |
309 | return DAG.getVectorShuffle(VT: NVT, dl, N1, N2, Mask: NewMask); |
310 | } |
311 | |
312 | /// Expands the ConstantFP node to an integer constant or |
313 | /// a load from the constant pool. |
314 | SDValue |
315 | SelectionDAGLegalize::ExpandConstantFP(ConstantFPSDNode *CFP, bool UseCP) { |
316 | bool Extend = false; |
317 | SDLoc dl(CFP); |
318 | |
319 | // If a FP immediate is precise when represented as a float and if the |
320 | // target can do an extending load from float to double, we put it into |
321 | // the constant pool as a float, even if it's is statically typed as a |
322 | // double. This shrinks FP constants and canonicalizes them for targets where |
323 | // an FP extending load is the same cost as a normal load (such as on the x87 |
324 | // fp stack or PPC FP unit). |
325 | EVT VT = CFP->getValueType(ResNo: 0); |
326 | ConstantFP *LLVMC = const_cast<ConstantFP*>(CFP->getConstantFPValue()); |
327 | if (!UseCP) { |
328 | assert((VT == MVT::f64 || VT == MVT::f32) && "Invalid type expansion" ); |
329 | return DAG.getConstant(Val: LLVMC->getValueAPF().bitcastToAPInt(), DL: dl, |
330 | VT: (VT == MVT::f64) ? MVT::i64 : MVT::i32); |
331 | } |
332 | |
333 | APFloat APF = CFP->getValueAPF(); |
334 | EVT OrigVT = VT; |
335 | EVT SVT = VT; |
336 | |
337 | // We don't want to shrink SNaNs. Converting the SNaN back to its real type |
338 | // can cause it to be changed into a QNaN on some platforms (e.g. on SystemZ). |
339 | if (!APF.isSignaling()) { |
340 | while (SVT != MVT::f32 && SVT != MVT::f16 && SVT != MVT::bf16) { |
341 | SVT = (MVT::SimpleValueType)(SVT.getSimpleVT().SimpleTy - 1); |
342 | if (ConstantFPSDNode::isValueValidForType(VT: SVT, Val: APF) && |
343 | // Only do this if the target has a native EXTLOAD instruction from |
344 | // smaller type. |
345 | TLI.isLoadExtLegal(ExtType: ISD::EXTLOAD, ValVT: OrigVT, MemVT: SVT) && |
346 | TLI.ShouldShrinkFPConstant(OrigVT)) { |
347 | Type *SType = SVT.getTypeForEVT(Context&: *DAG.getContext()); |
348 | LLVMC = cast<ConstantFP>(Val: ConstantFoldCastOperand( |
349 | Opcode: Instruction::FPTrunc, C: LLVMC, DestTy: SType, DL: DAG.getDataLayout())); |
350 | VT = SVT; |
351 | Extend = true; |
352 | } |
353 | } |
354 | } |
355 | |
356 | SDValue CPIdx = |
357 | DAG.getConstantPool(C: LLVMC, VT: TLI.getPointerTy(DL: DAG.getDataLayout())); |
358 | Align Alignment = cast<ConstantPoolSDNode>(Val&: CPIdx)->getAlign(); |
359 | if (Extend) { |
360 | SDValue Result = DAG.getExtLoad( |
361 | ExtType: ISD::EXTLOAD, dl, VT: OrigVT, Chain: DAG.getEntryNode(), Ptr: CPIdx, |
362 | PtrInfo: MachinePointerInfo::getConstantPool(MF&: DAG.getMachineFunction()), MemVT: VT, |
363 | Alignment); |
364 | return Result; |
365 | } |
366 | SDValue Result = DAG.getLoad( |
367 | VT: OrigVT, dl, Chain: DAG.getEntryNode(), Ptr: CPIdx, |
368 | PtrInfo: MachinePointerInfo::getConstantPool(MF&: DAG.getMachineFunction()), Alignment); |
369 | return Result; |
370 | } |
371 | |
372 | /// Expands the Constant node to a load from the constant pool. |
373 | SDValue SelectionDAGLegalize::ExpandConstant(ConstantSDNode *CP) { |
374 | SDLoc dl(CP); |
375 | EVT VT = CP->getValueType(ResNo: 0); |
376 | SDValue CPIdx = DAG.getConstantPool(C: CP->getConstantIntValue(), |
377 | VT: TLI.getPointerTy(DL: DAG.getDataLayout())); |
378 | Align Alignment = cast<ConstantPoolSDNode>(Val&: CPIdx)->getAlign(); |
379 | SDValue Result = DAG.getLoad( |
380 | VT, dl, Chain: DAG.getEntryNode(), Ptr: CPIdx, |
381 | PtrInfo: MachinePointerInfo::getConstantPool(MF&: DAG.getMachineFunction()), Alignment); |
382 | return Result; |
383 | } |
384 | |
385 | SDValue SelectionDAGLegalize::ExpandINSERT_VECTOR_ELT(SDValue Op) { |
386 | SDValue Vec = Op.getOperand(i: 0); |
387 | SDValue Val = Op.getOperand(i: 1); |
388 | SDValue Idx = Op.getOperand(i: 2); |
389 | SDLoc dl(Op); |
390 | |
391 | if (ConstantSDNode *InsertPos = dyn_cast<ConstantSDNode>(Val&: Idx)) { |
392 | // SCALAR_TO_VECTOR requires that the type of the value being inserted |
393 | // match the element type of the vector being created, except for |
394 | // integers in which case the inserted value can be over width. |
395 | EVT EltVT = Vec.getValueType().getVectorElementType(); |
396 | if (Val.getValueType() == EltVT || |
397 | (EltVT.isInteger() && Val.getValueType().bitsGE(VT: EltVT))) { |
398 | SDValue ScVec = DAG.getNode(Opcode: ISD::SCALAR_TO_VECTOR, DL: dl, |
399 | VT: Vec.getValueType(), Operand: Val); |
400 | |
401 | unsigned NumElts = Vec.getValueType().getVectorNumElements(); |
402 | // We generate a shuffle of InVec and ScVec, so the shuffle mask |
403 | // should be 0,1,2,3,4,5... with the appropriate element replaced with |
404 | // elt 0 of the RHS. |
405 | SmallVector<int, 8> ShufOps; |
406 | for (unsigned i = 0; i != NumElts; ++i) |
407 | ShufOps.push_back(Elt: i != InsertPos->getZExtValue() ? i : NumElts); |
408 | |
409 | return DAG.getVectorShuffle(VT: Vec.getValueType(), dl, N1: Vec, N2: ScVec, Mask: ShufOps); |
410 | } |
411 | } |
412 | return ExpandInsertToVectorThroughStack(Op); |
413 | } |
414 | |
415 | SDValue SelectionDAGLegalize::OptimizeFloatStore(StoreSDNode* ST) { |
416 | if (!ISD::isNormalStore(N: ST)) |
417 | return SDValue(); |
418 | |
419 | LLVM_DEBUG(dbgs() << "Optimizing float store operations\n" ); |
420 | // Turn 'store float 1.0, Ptr' -> 'store int 0x12345678, Ptr' |
421 | // FIXME: move this to the DAG Combiner! Note that we can't regress due |
422 | // to phase ordering between legalized code and the dag combiner. This |
423 | // probably means that we need to integrate dag combiner and legalizer |
424 | // together. |
425 | // We generally can't do this one for long doubles. |
426 | SDValue Chain = ST->getChain(); |
427 | SDValue Ptr = ST->getBasePtr(); |
428 | SDValue Value = ST->getValue(); |
429 | MachineMemOperand::Flags MMOFlags = ST->getMemOperand()->getFlags(); |
430 | AAMDNodes AAInfo = ST->getAAInfo(); |
431 | SDLoc dl(ST); |
432 | |
433 | // Don't optimise TargetConstantFP |
434 | if (Value.getOpcode() == ISD::TargetConstantFP) |
435 | return SDValue(); |
436 | |
437 | if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Val&: Value)) { |
438 | if (CFP->getValueType(ResNo: 0) == MVT::f32 && |
439 | TLI.isTypeLegal(VT: MVT::i32)) { |
440 | SDValue Con = DAG.getConstant(Val: CFP->getValueAPF(). |
441 | bitcastToAPInt().zextOrTrunc(width: 32), |
442 | DL: SDLoc(CFP), VT: MVT::i32); |
443 | return DAG.getStore(Chain, dl, Val: Con, Ptr, PtrInfo: ST->getPointerInfo(), |
444 | Alignment: ST->getBaseAlign(), MMOFlags, AAInfo); |
445 | } |
446 | |
447 | if (CFP->getValueType(ResNo: 0) == MVT::f64 && |
448 | !TLI.isFPImmLegal(CFP->getValueAPF(), MVT::f64)) { |
449 | // If this target supports 64-bit registers, do a single 64-bit store. |
450 | if (TLI.isTypeLegal(VT: MVT::i64)) { |
451 | SDValue Con = DAG.getConstant(Val: CFP->getValueAPF().bitcastToAPInt(). |
452 | zextOrTrunc(width: 64), DL: SDLoc(CFP), VT: MVT::i64); |
453 | return DAG.getStore(Chain, dl, Val: Con, Ptr, PtrInfo: ST->getPointerInfo(), |
454 | Alignment: ST->getBaseAlign(), MMOFlags, AAInfo); |
455 | } |
456 | |
457 | if (TLI.isTypeLegal(VT: MVT::i32) && !ST->isVolatile()) { |
458 | // Otherwise, if the target supports 32-bit registers, use 2 32-bit |
459 | // stores. If the target supports neither 32- nor 64-bits, this |
460 | // xform is certainly not worth it. |
461 | const APInt &IntVal = CFP->getValueAPF().bitcastToAPInt(); |
462 | SDValue Lo = DAG.getConstant(Val: IntVal.trunc(width: 32), DL: dl, VT: MVT::i32); |
463 | SDValue Hi = DAG.getConstant(Val: IntVal.lshr(shiftAmt: 32).trunc(width: 32), DL: dl, VT: MVT::i32); |
464 | if (DAG.getDataLayout().isBigEndian()) |
465 | std::swap(a&: Lo, b&: Hi); |
466 | |
467 | Lo = DAG.getStore(Chain, dl, Val: Lo, Ptr, PtrInfo: ST->getPointerInfo(), |
468 | Alignment: ST->getBaseAlign(), MMOFlags, AAInfo); |
469 | Ptr = DAG.getMemBasePlusOffset(Base: Ptr, Offset: TypeSize::getFixed(ExactSize: 4), DL: dl); |
470 | Hi = DAG.getStore(Chain, dl, Val: Hi, Ptr, |
471 | PtrInfo: ST->getPointerInfo().getWithOffset(O: 4), |
472 | Alignment: ST->getBaseAlign(), MMOFlags, AAInfo); |
473 | |
474 | return DAG.getNode(Opcode: ISD::TokenFactor, DL: dl, VT: MVT::Other, N1: Lo, N2: Hi); |
475 | } |
476 | } |
477 | } |
478 | return SDValue(); |
479 | } |
480 | |
481 | void SelectionDAGLegalize::LegalizeStoreOps(SDNode *Node) { |
482 | StoreSDNode *ST = cast<StoreSDNode>(Val: Node); |
483 | SDValue Chain = ST->getChain(); |
484 | SDValue Ptr = ST->getBasePtr(); |
485 | SDLoc dl(Node); |
486 | |
487 | MachineMemOperand::Flags MMOFlags = ST->getMemOperand()->getFlags(); |
488 | AAMDNodes AAInfo = ST->getAAInfo(); |
489 | |
490 | if (!ST->isTruncatingStore()) { |
491 | LLVM_DEBUG(dbgs() << "Legalizing store operation\n" ); |
492 | if (SDNode *OptStore = OptimizeFloatStore(ST).getNode()) { |
493 | ReplaceNode(Old: ST, New: OptStore); |
494 | return; |
495 | } |
496 | |
497 | SDValue Value = ST->getValue(); |
498 | MVT VT = Value.getSimpleValueType(); |
499 | switch (TLI.getOperationAction(Op: ISD::STORE, VT)) { |
500 | default: llvm_unreachable("This action is not supported yet!" ); |
501 | case TargetLowering::Legal: { |
502 | // If this is an unaligned store and the target doesn't support it, |
503 | // expand it. |
504 | EVT MemVT = ST->getMemoryVT(); |
505 | const DataLayout &DL = DAG.getDataLayout(); |
506 | if (!TLI.allowsMemoryAccessForAlignment(Context&: *DAG.getContext(), DL, VT: MemVT, |
507 | MMO: *ST->getMemOperand())) { |
508 | LLVM_DEBUG(dbgs() << "Expanding unsupported unaligned store\n" ); |
509 | SDValue Result = TLI.expandUnalignedStore(ST, DAG); |
510 | ReplaceNode(Old: SDValue(ST, 0), New: Result); |
511 | } else |
512 | LLVM_DEBUG(dbgs() << "Legal store\n" ); |
513 | break; |
514 | } |
515 | case TargetLowering::Custom: { |
516 | LLVM_DEBUG(dbgs() << "Trying custom lowering\n" ); |
517 | SDValue Res = TLI.LowerOperation(Op: SDValue(Node, 0), DAG); |
518 | if (Res && Res != SDValue(Node, 0)) |
519 | ReplaceNode(Old: SDValue(Node, 0), New: Res); |
520 | return; |
521 | } |
522 | case TargetLowering::Promote: { |
523 | MVT NVT = TLI.getTypeToPromoteTo(Op: ISD::STORE, VT); |
524 | assert(NVT.getSizeInBits() == VT.getSizeInBits() && |
525 | "Can only promote stores to same size type" ); |
526 | Value = DAG.getNode(Opcode: ISD::BITCAST, DL: dl, VT: NVT, Operand: Value); |
527 | SDValue Result = DAG.getStore(Chain, dl, Val: Value, Ptr, PtrInfo: ST->getPointerInfo(), |
528 | Alignment: ST->getBaseAlign(), MMOFlags, AAInfo); |
529 | ReplaceNode(Old: SDValue(Node, 0), New: Result); |
530 | break; |
531 | } |
532 | } |
533 | return; |
534 | } |
535 | |
536 | LLVM_DEBUG(dbgs() << "Legalizing truncating store operations\n" ); |
537 | SDValue Value = ST->getValue(); |
538 | EVT StVT = ST->getMemoryVT(); |
539 | TypeSize StWidth = StVT.getSizeInBits(); |
540 | TypeSize StSize = StVT.getStoreSizeInBits(); |
541 | auto &DL = DAG.getDataLayout(); |
542 | |
543 | if (StWidth != StSize) { |
544 | // Promote to a byte-sized store with upper bits zero if not |
545 | // storing an integral number of bytes. For example, promote |
546 | // TRUNCSTORE:i1 X -> TRUNCSTORE:i8 (and X, 1) |
547 | EVT NVT = EVT::getIntegerVT(Context&: *DAG.getContext(), BitWidth: StSize.getFixedValue()); |
548 | Value = DAG.getZeroExtendInReg(Op: Value, DL: dl, VT: StVT); |
549 | SDValue Result = |
550 | DAG.getTruncStore(Chain, dl, Val: Value, Ptr, PtrInfo: ST->getPointerInfo(), SVT: NVT, |
551 | Alignment: ST->getBaseAlign(), MMOFlags, AAInfo); |
552 | ReplaceNode(Old: SDValue(Node, 0), New: Result); |
553 | } else if (!StVT.isVector() && !isPowerOf2_64(Value: StWidth.getFixedValue())) { |
554 | // If not storing a power-of-2 number of bits, expand as two stores. |
555 | assert(!StVT.isVector() && "Unsupported truncstore!" ); |
556 | unsigned StWidthBits = StWidth.getFixedValue(); |
557 | unsigned LogStWidth = Log2_32(Value: StWidthBits); |
558 | assert(LogStWidth < 32); |
559 | unsigned RoundWidth = 1 << LogStWidth; |
560 | assert(RoundWidth < StWidthBits); |
561 | unsigned = StWidthBits - RoundWidth; |
562 | assert(ExtraWidth < RoundWidth); |
563 | assert(!(RoundWidth % 8) && !(ExtraWidth % 8) && |
564 | "Store size not an integral number of bytes!" ); |
565 | EVT RoundVT = EVT::getIntegerVT(Context&: *DAG.getContext(), BitWidth: RoundWidth); |
566 | EVT = EVT::getIntegerVT(Context&: *DAG.getContext(), BitWidth: ExtraWidth); |
567 | SDValue Lo, Hi; |
568 | unsigned IncrementSize; |
569 | |
570 | if (DL.isLittleEndian()) { |
571 | // TRUNCSTORE:i24 X -> TRUNCSTORE:i16 X, TRUNCSTORE@+2:i8 (srl X, 16) |
572 | // Store the bottom RoundWidth bits. |
573 | Lo = DAG.getTruncStore(Chain, dl, Val: Value, Ptr, PtrInfo: ST->getPointerInfo(), |
574 | SVT: RoundVT, Alignment: ST->getBaseAlign(), MMOFlags, AAInfo); |
575 | |
576 | // Store the remaining ExtraWidth bits. |
577 | IncrementSize = RoundWidth / 8; |
578 | Ptr = |
579 | DAG.getMemBasePlusOffset(Base: Ptr, Offset: TypeSize::getFixed(ExactSize: IncrementSize), DL: dl); |
580 | Hi = DAG.getNode( |
581 | Opcode: ISD::SRL, DL: dl, VT: Value.getValueType(), N1: Value, |
582 | N2: DAG.getConstant(Val: RoundWidth, DL: dl, |
583 | VT: TLI.getShiftAmountTy(LHSTy: Value.getValueType(), DL))); |
584 | Hi = DAG.getTruncStore(Chain, dl, Val: Hi, Ptr, |
585 | PtrInfo: ST->getPointerInfo().getWithOffset(O: IncrementSize), |
586 | SVT: ExtraVT, Alignment: ST->getBaseAlign(), MMOFlags, AAInfo); |
587 | } else { |
588 | // Big endian - avoid unaligned stores. |
589 | // TRUNCSTORE:i24 X -> TRUNCSTORE:i16 (srl X, 8), TRUNCSTORE@+2:i8 X |
590 | // Store the top RoundWidth bits. |
591 | Hi = DAG.getNode( |
592 | Opcode: ISD::SRL, DL: dl, VT: Value.getValueType(), N1: Value, |
593 | N2: DAG.getConstant(Val: ExtraWidth, DL: dl, |
594 | VT: TLI.getShiftAmountTy(LHSTy: Value.getValueType(), DL))); |
595 | Hi = DAG.getTruncStore(Chain, dl, Val: Hi, Ptr, PtrInfo: ST->getPointerInfo(), SVT: RoundVT, |
596 | Alignment: ST->getBaseAlign(), MMOFlags, AAInfo); |
597 | |
598 | // Store the remaining ExtraWidth bits. |
599 | IncrementSize = RoundWidth / 8; |
600 | Ptr = DAG.getNode(Opcode: ISD::ADD, DL: dl, VT: Ptr.getValueType(), N1: Ptr, |
601 | N2: DAG.getConstant(Val: IncrementSize, DL: dl, |
602 | VT: Ptr.getValueType())); |
603 | Lo = DAG.getTruncStore(Chain, dl, Val: Value, Ptr, |
604 | PtrInfo: ST->getPointerInfo().getWithOffset(O: IncrementSize), |
605 | SVT: ExtraVT, Alignment: ST->getBaseAlign(), MMOFlags, AAInfo); |
606 | } |
607 | |
608 | // The order of the stores doesn't matter. |
609 | SDValue Result = DAG.getNode(Opcode: ISD::TokenFactor, DL: dl, VT: MVT::Other, N1: Lo, N2: Hi); |
610 | ReplaceNode(Old: SDValue(Node, 0), New: Result); |
611 | } else { |
612 | switch (TLI.getTruncStoreAction(ValVT: ST->getValue().getValueType(), MemVT: StVT)) { |
613 | default: llvm_unreachable("This action is not supported yet!" ); |
614 | case TargetLowering::Legal: { |
615 | EVT MemVT = ST->getMemoryVT(); |
616 | // If this is an unaligned store and the target doesn't support it, |
617 | // expand it. |
618 | if (!TLI.allowsMemoryAccessForAlignment(Context&: *DAG.getContext(), DL, VT: MemVT, |
619 | MMO: *ST->getMemOperand())) { |
620 | SDValue Result = TLI.expandUnalignedStore(ST, DAG); |
621 | ReplaceNode(Old: SDValue(ST, 0), New: Result); |
622 | } |
623 | break; |
624 | } |
625 | case TargetLowering::Custom: { |
626 | SDValue Res = TLI.LowerOperation(Op: SDValue(Node, 0), DAG); |
627 | if (Res && Res != SDValue(Node, 0)) |
628 | ReplaceNode(Old: SDValue(Node, 0), New: Res); |
629 | return; |
630 | } |
631 | case TargetLowering::Expand: |
632 | assert(!StVT.isVector() && |
633 | "Vector Stores are handled in LegalizeVectorOps" ); |
634 | |
635 | SDValue Result; |
636 | |
637 | // TRUNCSTORE:i16 i32 -> STORE i16 |
638 | if (TLI.isTypeLegal(VT: StVT)) { |
639 | Value = DAG.getNode(Opcode: ISD::TRUNCATE, DL: dl, VT: StVT, Operand: Value); |
640 | Result = DAG.getStore(Chain, dl, Val: Value, Ptr, PtrInfo: ST->getPointerInfo(), |
641 | Alignment: ST->getBaseAlign(), MMOFlags, AAInfo); |
642 | } else { |
643 | // The in-memory type isn't legal. Truncate to the type it would promote |
644 | // to, and then do a truncstore. |
645 | Value = DAG.getNode(Opcode: ISD::TRUNCATE, DL: dl, |
646 | VT: TLI.getTypeToTransformTo(Context&: *DAG.getContext(), VT: StVT), |
647 | Operand: Value); |
648 | Result = DAG.getTruncStore(Chain, dl, Val: Value, Ptr, PtrInfo: ST->getPointerInfo(), |
649 | SVT: StVT, Alignment: ST->getBaseAlign(), MMOFlags, AAInfo); |
650 | } |
651 | |
652 | ReplaceNode(Old: SDValue(Node, 0), New: Result); |
653 | break; |
654 | } |
655 | } |
656 | } |
657 | |
658 | void SelectionDAGLegalize::LegalizeLoadOps(SDNode *Node) { |
659 | LoadSDNode *LD = cast<LoadSDNode>(Val: Node); |
660 | SDValue Chain = LD->getChain(); // The chain. |
661 | SDValue Ptr = LD->getBasePtr(); // The base pointer. |
662 | SDValue Value; // The value returned by the load op. |
663 | SDLoc dl(Node); |
664 | |
665 | ISD::LoadExtType ExtType = LD->getExtensionType(); |
666 | if (ExtType == ISD::NON_EXTLOAD) { |
667 | LLVM_DEBUG(dbgs() << "Legalizing non-extending load operation\n" ); |
668 | MVT VT = Node->getSimpleValueType(ResNo: 0); |
669 | SDValue RVal = SDValue(Node, 0); |
670 | SDValue RChain = SDValue(Node, 1); |
671 | |
672 | switch (TLI.getOperationAction(Op: Node->getOpcode(), VT)) { |
673 | default: llvm_unreachable("This action is not supported yet!" ); |
674 | case TargetLowering::Legal: { |
675 | EVT MemVT = LD->getMemoryVT(); |
676 | const DataLayout &DL = DAG.getDataLayout(); |
677 | // If this is an unaligned load and the target doesn't support it, |
678 | // expand it. |
679 | if (!TLI.allowsMemoryAccessForAlignment(Context&: *DAG.getContext(), DL, VT: MemVT, |
680 | MMO: *LD->getMemOperand())) { |
681 | std::tie(args&: RVal, args&: RChain) = TLI.expandUnalignedLoad(LD, DAG); |
682 | } |
683 | break; |
684 | } |
685 | case TargetLowering::Custom: |
686 | if (SDValue Res = TLI.LowerOperation(Op: RVal, DAG)) { |
687 | RVal = Res; |
688 | RChain = Res.getValue(R: 1); |
689 | } |
690 | break; |
691 | |
692 | case TargetLowering::Promote: { |
693 | MVT NVT = TLI.getTypeToPromoteTo(Op: Node->getOpcode(), VT); |
694 | assert(NVT.getSizeInBits() == VT.getSizeInBits() && |
695 | "Can only promote loads to same size type" ); |
696 | |
697 | // If the range metadata type does not match the legalized memory |
698 | // operation type, remove the range metadata. |
699 | if (const MDNode *MD = LD->getRanges()) { |
700 | ConstantInt *Lower = mdconst::extract<ConstantInt>(MD: MD->getOperand(I: 0)); |
701 | if (Lower->getBitWidth() != NVT.getScalarSizeInBits() || |
702 | !NVT.isInteger()) |
703 | LD->getMemOperand()->clearRanges(); |
704 | } |
705 | SDValue Res = DAG.getLoad(VT: NVT, dl, Chain, Ptr, MMO: LD->getMemOperand()); |
706 | RVal = DAG.getNode(Opcode: ISD::BITCAST, DL: dl, VT, Operand: Res); |
707 | RChain = Res.getValue(R: 1); |
708 | break; |
709 | } |
710 | } |
711 | if (RChain.getNode() != Node) { |
712 | assert(RVal.getNode() != Node && "Load must be completely replaced" ); |
713 | DAG.ReplaceAllUsesOfValueWith(From: SDValue(Node, 0), To: RVal); |
714 | DAG.ReplaceAllUsesOfValueWith(From: SDValue(Node, 1), To: RChain); |
715 | if (UpdatedNodes) { |
716 | UpdatedNodes->insert(X: RVal.getNode()); |
717 | UpdatedNodes->insert(X: RChain.getNode()); |
718 | } |
719 | ReplacedNode(N: Node); |
720 | } |
721 | return; |
722 | } |
723 | |
724 | LLVM_DEBUG(dbgs() << "Legalizing extending load operation\n" ); |
725 | EVT SrcVT = LD->getMemoryVT(); |
726 | TypeSize SrcWidth = SrcVT.getSizeInBits(); |
727 | MachineMemOperand::Flags MMOFlags = LD->getMemOperand()->getFlags(); |
728 | AAMDNodes AAInfo = LD->getAAInfo(); |
729 | |
730 | if (SrcWidth != SrcVT.getStoreSizeInBits() && |
731 | // Some targets pretend to have an i1 loading operation, and actually |
732 | // load an i8. This trick is correct for ZEXTLOAD because the top 7 |
733 | // bits are guaranteed to be zero; it helps the optimizers understand |
734 | // that these bits are zero. It is also useful for EXTLOAD, since it |
735 | // tells the optimizers that those bits are undefined. It would be |
736 | // nice to have an effective generic way of getting these benefits... |
737 | // Until such a way is found, don't insist on promoting i1 here. |
738 | (SrcVT != MVT::i1 || |
739 | TLI.getLoadExtAction(ExtType, ValVT: Node->getValueType(ResNo: 0), MemVT: MVT::i1) == |
740 | TargetLowering::Promote)) { |
741 | // Promote to a byte-sized load if not loading an integral number of |
742 | // bytes. For example, promote EXTLOAD:i20 -> EXTLOAD:i24. |
743 | unsigned NewWidth = SrcVT.getStoreSizeInBits(); |
744 | EVT NVT = EVT::getIntegerVT(Context&: *DAG.getContext(), BitWidth: NewWidth); |
745 | SDValue Ch; |
746 | |
747 | // The extra bits are guaranteed to be zero, since we stored them that |
748 | // way. A zext load from NVT thus automatically gives zext from SrcVT. |
749 | |
750 | ISD::LoadExtType NewExtType = |
751 | ExtType == ISD::ZEXTLOAD ? ISD::ZEXTLOAD : ISD::EXTLOAD; |
752 | |
753 | SDValue Result = DAG.getExtLoad(ExtType: NewExtType, dl, VT: Node->getValueType(ResNo: 0), |
754 | Chain, Ptr, PtrInfo: LD->getPointerInfo(), MemVT: NVT, |
755 | Alignment: LD->getBaseAlign(), MMOFlags, AAInfo); |
756 | |
757 | Ch = Result.getValue(R: 1); // The chain. |
758 | |
759 | if (ExtType == ISD::SEXTLOAD) |
760 | // Having the top bits zero doesn't help when sign extending. |
761 | Result = DAG.getNode(Opcode: ISD::SIGN_EXTEND_INREG, DL: dl, |
762 | VT: Result.getValueType(), |
763 | N1: Result, N2: DAG.getValueType(SrcVT)); |
764 | else if (ExtType == ISD::ZEXTLOAD || NVT == Result.getValueType()) |
765 | // All the top bits are guaranteed to be zero - inform the optimizers. |
766 | Result = DAG.getNode(Opcode: ISD::AssertZext, DL: dl, |
767 | VT: Result.getValueType(), N1: Result, |
768 | N2: DAG.getValueType(SrcVT)); |
769 | |
770 | Value = Result; |
771 | Chain = Ch; |
772 | } else if (!isPowerOf2_64(Value: SrcWidth.getKnownMinValue())) { |
773 | // If not loading a power-of-2 number of bits, expand as two loads. |
774 | assert(!SrcVT.isVector() && "Unsupported extload!" ); |
775 | unsigned SrcWidthBits = SrcWidth.getFixedValue(); |
776 | unsigned LogSrcWidth = Log2_32(Value: SrcWidthBits); |
777 | assert(LogSrcWidth < 32); |
778 | unsigned RoundWidth = 1 << LogSrcWidth; |
779 | assert(RoundWidth < SrcWidthBits); |
780 | unsigned = SrcWidthBits - RoundWidth; |
781 | assert(ExtraWidth < RoundWidth); |
782 | assert(!(RoundWidth % 8) && !(ExtraWidth % 8) && |
783 | "Load size not an integral number of bytes!" ); |
784 | EVT RoundVT = EVT::getIntegerVT(Context&: *DAG.getContext(), BitWidth: RoundWidth); |
785 | EVT = EVT::getIntegerVT(Context&: *DAG.getContext(), BitWidth: ExtraWidth); |
786 | SDValue Lo, Hi, Ch; |
787 | unsigned IncrementSize; |
788 | auto &DL = DAG.getDataLayout(); |
789 | |
790 | if (DL.isLittleEndian()) { |
791 | // EXTLOAD:i24 -> ZEXTLOAD:i16 | (shl EXTLOAD@+2:i8, 16) |
792 | // Load the bottom RoundWidth bits. |
793 | Lo = DAG.getExtLoad(ExtType: ISD::ZEXTLOAD, dl, VT: Node->getValueType(ResNo: 0), Chain, Ptr, |
794 | PtrInfo: LD->getPointerInfo(), MemVT: RoundVT, Alignment: LD->getBaseAlign(), |
795 | MMOFlags, AAInfo); |
796 | |
797 | // Load the remaining ExtraWidth bits. |
798 | IncrementSize = RoundWidth / 8; |
799 | Ptr = |
800 | DAG.getMemBasePlusOffset(Base: Ptr, Offset: TypeSize::getFixed(ExactSize: IncrementSize), DL: dl); |
801 | Hi = DAG.getExtLoad(ExtType, dl, VT: Node->getValueType(ResNo: 0), Chain, Ptr, |
802 | PtrInfo: LD->getPointerInfo().getWithOffset(O: IncrementSize), |
803 | MemVT: ExtraVT, Alignment: LD->getBaseAlign(), MMOFlags, AAInfo); |
804 | |
805 | // Build a factor node to remember that this load is independent of |
806 | // the other one. |
807 | Ch = DAG.getNode(Opcode: ISD::TokenFactor, DL: dl, VT: MVT::Other, N1: Lo.getValue(R: 1), |
808 | N2: Hi.getValue(R: 1)); |
809 | |
810 | // Move the top bits to the right place. |
811 | Hi = DAG.getNode( |
812 | Opcode: ISD::SHL, DL: dl, VT: Hi.getValueType(), N1: Hi, |
813 | N2: DAG.getConstant(Val: RoundWidth, DL: dl, |
814 | VT: TLI.getShiftAmountTy(LHSTy: Hi.getValueType(), DL))); |
815 | |
816 | // Join the hi and lo parts. |
817 | Value = DAG.getNode(Opcode: ISD::OR, DL: dl, VT: Node->getValueType(ResNo: 0), N1: Lo, N2: Hi); |
818 | } else { |
819 | // Big endian - avoid unaligned loads. |
820 | // EXTLOAD:i24 -> (shl EXTLOAD:i16, 8) | ZEXTLOAD@+2:i8 |
821 | // Load the top RoundWidth bits. |
822 | Hi = DAG.getExtLoad(ExtType, dl, VT: Node->getValueType(ResNo: 0), Chain, Ptr, |
823 | PtrInfo: LD->getPointerInfo(), MemVT: RoundVT, Alignment: LD->getBaseAlign(), |
824 | MMOFlags, AAInfo); |
825 | |
826 | // Load the remaining ExtraWidth bits. |
827 | IncrementSize = RoundWidth / 8; |
828 | Ptr = |
829 | DAG.getMemBasePlusOffset(Base: Ptr, Offset: TypeSize::getFixed(ExactSize: IncrementSize), DL: dl); |
830 | Lo = DAG.getExtLoad(ExtType: ISD::ZEXTLOAD, dl, VT: Node->getValueType(ResNo: 0), Chain, Ptr, |
831 | PtrInfo: LD->getPointerInfo().getWithOffset(O: IncrementSize), |
832 | MemVT: ExtraVT, Alignment: LD->getBaseAlign(), MMOFlags, AAInfo); |
833 | |
834 | // Build a factor node to remember that this load is independent of |
835 | // the other one. |
836 | Ch = DAG.getNode(Opcode: ISD::TokenFactor, DL: dl, VT: MVT::Other, N1: Lo.getValue(R: 1), |
837 | N2: Hi.getValue(R: 1)); |
838 | |
839 | // Move the top bits to the right place. |
840 | Hi = DAG.getNode( |
841 | Opcode: ISD::SHL, DL: dl, VT: Hi.getValueType(), N1: Hi, |
842 | N2: DAG.getConstant(Val: ExtraWidth, DL: dl, |
843 | VT: TLI.getShiftAmountTy(LHSTy: Hi.getValueType(), DL))); |
844 | |
845 | // Join the hi and lo parts. |
846 | Value = DAG.getNode(Opcode: ISD::OR, DL: dl, VT: Node->getValueType(ResNo: 0), N1: Lo, N2: Hi); |
847 | } |
848 | |
849 | Chain = Ch; |
850 | } else { |
851 | bool isCustom = false; |
852 | switch (TLI.getLoadExtAction(ExtType, ValVT: Node->getValueType(ResNo: 0), |
853 | MemVT: SrcVT.getSimpleVT())) { |
854 | default: llvm_unreachable("This action is not supported yet!" ); |
855 | case TargetLowering::Custom: |
856 | isCustom = true; |
857 | [[fallthrough]]; |
858 | case TargetLowering::Legal: |
859 | Value = SDValue(Node, 0); |
860 | Chain = SDValue(Node, 1); |
861 | |
862 | if (isCustom) { |
863 | if (SDValue Res = TLI.LowerOperation(Op: SDValue(Node, 0), DAG)) { |
864 | Value = Res; |
865 | Chain = Res.getValue(R: 1); |
866 | } |
867 | } else { |
868 | // If this is an unaligned load and the target doesn't support it, |
869 | // expand it. |
870 | EVT MemVT = LD->getMemoryVT(); |
871 | const DataLayout &DL = DAG.getDataLayout(); |
872 | if (!TLI.allowsMemoryAccess(Context&: *DAG.getContext(), DL, VT: MemVT, |
873 | MMO: *LD->getMemOperand())) { |
874 | std::tie(args&: Value, args&: Chain) = TLI.expandUnalignedLoad(LD, DAG); |
875 | } |
876 | } |
877 | break; |
878 | |
879 | case TargetLowering::Expand: { |
880 | EVT DestVT = Node->getValueType(ResNo: 0); |
881 | if (!TLI.isLoadExtLegal(ExtType: ISD::EXTLOAD, ValVT: DestVT, MemVT: SrcVT)) { |
882 | // If the source type is not legal, see if there is a legal extload to |
883 | // an intermediate type that we can then extend further. |
884 | EVT LoadVT = TLI.getRegisterType(VT: SrcVT.getSimpleVT()); |
885 | if ((LoadVT.isFloatingPoint() == SrcVT.isFloatingPoint()) && |
886 | (TLI.isTypeLegal(VT: SrcVT) || // Same as SrcVT == LoadVT? |
887 | TLI.isLoadExtLegal(ExtType, ValVT: LoadVT, MemVT: SrcVT))) { |
888 | // If we are loading a legal type, this is a non-extload followed by a |
889 | // full extend. |
890 | ISD::LoadExtType MidExtType = |
891 | (LoadVT == SrcVT) ? ISD::NON_EXTLOAD : ExtType; |
892 | |
893 | SDValue Load = DAG.getExtLoad(ExtType: MidExtType, dl, VT: LoadVT, Chain, Ptr, |
894 | MemVT: SrcVT, MMO: LD->getMemOperand()); |
895 | unsigned ExtendOp = |
896 | ISD::getExtForLoadExtType(IsFP: SrcVT.isFloatingPoint(), ExtType); |
897 | Value = DAG.getNode(Opcode: ExtendOp, DL: dl, VT: Node->getValueType(ResNo: 0), Operand: Load); |
898 | Chain = Load.getValue(R: 1); |
899 | break; |
900 | } |
901 | |
902 | // Handle the special case of fp16 extloads. EXTLOAD doesn't have the |
903 | // normal undefined upper bits behavior to allow using an in-reg extend |
904 | // with the illegal FP type, so load as an integer and do the |
905 | // from-integer conversion. |
906 | EVT SVT = SrcVT.getScalarType(); |
907 | if (SVT == MVT::f16 || SVT == MVT::bf16) { |
908 | EVT ISrcVT = SrcVT.changeTypeToInteger(); |
909 | EVT IDestVT = DestVT.changeTypeToInteger(); |
910 | EVT ILoadVT = TLI.getRegisterType(VT: IDestVT.getSimpleVT()); |
911 | |
912 | SDValue Result = DAG.getExtLoad(ExtType: ISD::ZEXTLOAD, dl, VT: ILoadVT, Chain, |
913 | Ptr, MemVT: ISrcVT, MMO: LD->getMemOperand()); |
914 | Value = |
915 | DAG.getNode(Opcode: SVT == MVT::f16 ? ISD::FP16_TO_FP : ISD::BF16_TO_FP, |
916 | DL: dl, VT: DestVT, Operand: Result); |
917 | Chain = Result.getValue(R: 1); |
918 | break; |
919 | } |
920 | } |
921 | |
922 | assert(!SrcVT.isVector() && |
923 | "Vector Loads are handled in LegalizeVectorOps" ); |
924 | |
925 | // FIXME: This does not work for vectors on most targets. Sign- |
926 | // and zero-extend operations are currently folded into extending |
927 | // loads, whether they are legal or not, and then we end up here |
928 | // without any support for legalizing them. |
929 | assert(ExtType != ISD::EXTLOAD && |
930 | "EXTLOAD should always be supported!" ); |
931 | // Turn the unsupported load into an EXTLOAD followed by an |
932 | // explicit zero/sign extend inreg. |
933 | SDValue Result = DAG.getExtLoad(ExtType: ISD::EXTLOAD, dl, |
934 | VT: Node->getValueType(ResNo: 0), |
935 | Chain, Ptr, MemVT: SrcVT, |
936 | MMO: LD->getMemOperand()); |
937 | SDValue ValRes; |
938 | if (ExtType == ISD::SEXTLOAD) |
939 | ValRes = DAG.getNode(Opcode: ISD::SIGN_EXTEND_INREG, DL: dl, |
940 | VT: Result.getValueType(), |
941 | N1: Result, N2: DAG.getValueType(SrcVT)); |
942 | else |
943 | ValRes = DAG.getZeroExtendInReg(Op: Result, DL: dl, VT: SrcVT); |
944 | Value = ValRes; |
945 | Chain = Result.getValue(R: 1); |
946 | break; |
947 | } |
948 | } |
949 | } |
950 | |
951 | // Since loads produce two values, make sure to remember that we legalized |
952 | // both of them. |
953 | if (Chain.getNode() != Node) { |
954 | assert(Value.getNode() != Node && "Load must be completely replaced" ); |
955 | DAG.ReplaceAllUsesOfValueWith(From: SDValue(Node, 0), To: Value); |
956 | DAG.ReplaceAllUsesOfValueWith(From: SDValue(Node, 1), To: Chain); |
957 | if (UpdatedNodes) { |
958 | UpdatedNodes->insert(X: Value.getNode()); |
959 | UpdatedNodes->insert(X: Chain.getNode()); |
960 | } |
961 | ReplacedNode(N: Node); |
962 | } |
963 | } |
964 | |
965 | /// Return a legal replacement for the given operation, with all legal operands. |
966 | void SelectionDAGLegalize::LegalizeOp(SDNode *Node) { |
967 | LLVM_DEBUG(dbgs() << "\nLegalizing: " ; Node->dump(&DAG)); |
968 | |
969 | // Allow illegal target nodes and illegal registers. |
970 | if (Node->getOpcode() == ISD::TargetConstant || |
971 | Node->getOpcode() == ISD::Register) |
972 | return; |
973 | |
974 | #ifndef NDEBUG |
975 | for (unsigned i = 0, e = Node->getNumValues(); i != e; ++i) |
976 | assert(TLI.getTypeAction(*DAG.getContext(), Node->getValueType(i)) == |
977 | TargetLowering::TypeLegal && |
978 | "Unexpected illegal type!" ); |
979 | |
980 | for (const SDValue &Op : Node->op_values()) |
981 | assert((TLI.getTypeAction(*DAG.getContext(), Op.getValueType()) == |
982 | TargetLowering::TypeLegal || |
983 | Op.getOpcode() == ISD::TargetConstant || |
984 | Op.getOpcode() == ISD::Register) && |
985 | "Unexpected illegal type!" ); |
986 | #endif |
987 | |
988 | // Figure out the correct action; the way to query this varies by opcode |
989 | TargetLowering::LegalizeAction Action = TargetLowering::Legal; |
990 | bool SimpleFinishLegalizing = true; |
991 | switch (Node->getOpcode()) { |
992 | // TODO: Currently, POISON is being lowered to UNDEF here. However, there is |
993 | // an open concern that this transformation may not be ideal, as targets |
994 | // should ideally handle POISON directly. Changing this behavior would require |
995 | // adding support for POISON in TableGen, which is a large change. |
996 | // Additionally, many existing test cases rely on the current behavior (e.g., |
997 | // llvm/test/CodeGen/PowerPC/vec_shuffle.ll). A broader discussion and |
998 | // incremental changes might be needed to properly |
999 | // support POISON without breaking existing targets and tests. |
1000 | case ISD::POISON: { |
1001 | SDValue UndefNode = DAG.getUNDEF(VT: Node->getValueType(ResNo: 0)); |
1002 | ReplaceNode(Old: Node, New: UndefNode.getNode()); |
1003 | break; |
1004 | } |
1005 | case ISD::INTRINSIC_W_CHAIN: |
1006 | case ISD::INTRINSIC_WO_CHAIN: |
1007 | case ISD::INTRINSIC_VOID: |
1008 | case ISD::STACKSAVE: |
1009 | Action = TLI.getOperationAction(Op: Node->getOpcode(), VT: MVT::Other); |
1010 | break; |
1011 | case ISD::GET_DYNAMIC_AREA_OFFSET: |
1012 | Action = TLI.getOperationAction(Op: Node->getOpcode(), |
1013 | VT: Node->getValueType(ResNo: 0)); |
1014 | break; |
1015 | case ISD::VAARG: |
1016 | Action = TLI.getOperationAction(Op: Node->getOpcode(), |
1017 | VT: Node->getValueType(ResNo: 0)); |
1018 | if (Action != TargetLowering::Promote) |
1019 | Action = TLI.getOperationAction(Op: Node->getOpcode(), VT: MVT::Other); |
1020 | break; |
1021 | case ISD::SET_FPENV: |
1022 | case ISD::SET_FPMODE: |
1023 | Action = TLI.getOperationAction(Op: Node->getOpcode(), |
1024 | VT: Node->getOperand(Num: 1).getValueType()); |
1025 | break; |
1026 | case ISD::FP_TO_FP16: |
1027 | case ISD::FP_TO_BF16: |
1028 | case ISD::SINT_TO_FP: |
1029 | case ISD::UINT_TO_FP: |
1030 | case ISD::EXTRACT_VECTOR_ELT: |
1031 | case ISD::LROUND: |
1032 | case ISD::LLROUND: |
1033 | case ISD::LRINT: |
1034 | case ISD::LLRINT: |
1035 | Action = TLI.getOperationAction(Op: Node->getOpcode(), |
1036 | VT: Node->getOperand(Num: 0).getValueType()); |
1037 | break; |
1038 | case ISD::STRICT_FP_TO_FP16: |
1039 | case ISD::STRICT_FP_TO_BF16: |
1040 | case ISD::STRICT_SINT_TO_FP: |
1041 | case ISD::STRICT_UINT_TO_FP: |
1042 | case ISD::STRICT_LRINT: |
1043 | case ISD::STRICT_LLRINT: |
1044 | case ISD::STRICT_LROUND: |
1045 | case ISD::STRICT_LLROUND: |
1046 | // These pseudo-ops are the same as the other STRICT_ ops except |
1047 | // they are registered with setOperationAction() using the input type |
1048 | // instead of the output type. |
1049 | Action = TLI.getOperationAction(Op: Node->getOpcode(), |
1050 | VT: Node->getOperand(Num: 1).getValueType()); |
1051 | break; |
1052 | case ISD::SIGN_EXTEND_INREG: { |
1053 | EVT InnerType = cast<VTSDNode>(Val: Node->getOperand(Num: 1))->getVT(); |
1054 | Action = TLI.getOperationAction(Op: Node->getOpcode(), VT: InnerType); |
1055 | break; |
1056 | } |
1057 | case ISD::ATOMIC_STORE: |
1058 | Action = TLI.getOperationAction(Op: Node->getOpcode(), |
1059 | VT: Node->getOperand(Num: 1).getValueType()); |
1060 | break; |
1061 | case ISD::SELECT_CC: |
1062 | case ISD::STRICT_FSETCC: |
1063 | case ISD::STRICT_FSETCCS: |
1064 | case ISD::SETCC: |
1065 | case ISD::SETCCCARRY: |
1066 | case ISD::VP_SETCC: |
1067 | case ISD::BR_CC: { |
1068 | unsigned Opc = Node->getOpcode(); |
1069 | unsigned CCOperand = Opc == ISD::SELECT_CC ? 4 |
1070 | : Opc == ISD::STRICT_FSETCC ? 3 |
1071 | : Opc == ISD::STRICT_FSETCCS ? 3 |
1072 | : Opc == ISD::SETCCCARRY ? 3 |
1073 | : (Opc == ISD::SETCC || Opc == ISD::VP_SETCC) ? 2 |
1074 | : 1; |
1075 | unsigned CompareOperand = Opc == ISD::BR_CC ? 2 |
1076 | : Opc == ISD::STRICT_FSETCC ? 1 |
1077 | : Opc == ISD::STRICT_FSETCCS ? 1 |
1078 | : 0; |
1079 | MVT OpVT = Node->getOperand(Num: CompareOperand).getSimpleValueType(); |
1080 | ISD::CondCode CCCode = |
1081 | cast<CondCodeSDNode>(Val: Node->getOperand(Num: CCOperand))->get(); |
1082 | Action = TLI.getCondCodeAction(CC: CCCode, VT: OpVT); |
1083 | if (Action == TargetLowering::Legal) { |
1084 | if (Node->getOpcode() == ISD::SELECT_CC) |
1085 | Action = TLI.getOperationAction(Op: Node->getOpcode(), |
1086 | VT: Node->getValueType(ResNo: 0)); |
1087 | else |
1088 | Action = TLI.getOperationAction(Op: Node->getOpcode(), VT: OpVT); |
1089 | } |
1090 | break; |
1091 | } |
1092 | case ISD::LOAD: |
1093 | case ISD::STORE: |
1094 | // FIXME: Model these properly. LOAD and STORE are complicated, and |
1095 | // STORE expects the unlegalized operand in some cases. |
1096 | SimpleFinishLegalizing = false; |
1097 | break; |
1098 | case ISD::CALLSEQ_START: |
1099 | case ISD::CALLSEQ_END: |
1100 | // FIXME: This shouldn't be necessary. These nodes have special properties |
1101 | // dealing with the recursive nature of legalization. Removing this |
1102 | // special case should be done as part of making LegalizeDAG non-recursive. |
1103 | SimpleFinishLegalizing = false; |
1104 | break; |
1105 | case ISD::EXTRACT_ELEMENT: |
1106 | case ISD::GET_ROUNDING: |
1107 | case ISD::MERGE_VALUES: |
1108 | case ISD::EH_RETURN: |
1109 | case ISD::FRAME_TO_ARGS_OFFSET: |
1110 | case ISD::EH_DWARF_CFA: |
1111 | case ISD::EH_SJLJ_SETJMP: |
1112 | case ISD::EH_SJLJ_LONGJMP: |
1113 | case ISD::EH_SJLJ_SETUP_DISPATCH: |
1114 | // These operations lie about being legal: when they claim to be legal, |
1115 | // they should actually be expanded. |
1116 | Action = TLI.getOperationAction(Op: Node->getOpcode(), VT: Node->getValueType(ResNo: 0)); |
1117 | if (Action == TargetLowering::Legal) |
1118 | Action = TargetLowering::Expand; |
1119 | break; |
1120 | case ISD::INIT_TRAMPOLINE: |
1121 | case ISD::ADJUST_TRAMPOLINE: |
1122 | case ISD::FRAMEADDR: |
1123 | case ISD::RETURNADDR: |
1124 | case ISD::ADDROFRETURNADDR: |
1125 | case ISD::SPONENTRY: |
1126 | // These operations lie about being legal: when they claim to be legal, |
1127 | // they should actually be custom-lowered. |
1128 | Action = TLI.getOperationAction(Op: Node->getOpcode(), VT: Node->getValueType(ResNo: 0)); |
1129 | if (Action == TargetLowering::Legal) |
1130 | Action = TargetLowering::Custom; |
1131 | break; |
1132 | case ISD::CLEAR_CACHE: |
1133 | // This operation is typically going to be LibCall unless the target wants |
1134 | // something differrent. |
1135 | Action = TLI.getOperationAction(Op: Node->getOpcode(), VT: Node->getValueType(ResNo: 0)); |
1136 | break; |
1137 | case ISD::READCYCLECOUNTER: |
1138 | case ISD::READSTEADYCOUNTER: |
1139 | // READCYCLECOUNTER and READSTEADYCOUNTER return a i64, even if type |
1140 | // legalization might have expanded that to several smaller types. |
1141 | Action = TLI.getOperationAction(Op: Node->getOpcode(), VT: MVT::i64); |
1142 | break; |
1143 | case ISD::READ_REGISTER: |
1144 | case ISD::WRITE_REGISTER: |
1145 | // Named register is legal in the DAG, but blocked by register name |
1146 | // selection if not implemented by target (to chose the correct register) |
1147 | // They'll be converted to Copy(To/From)Reg. |
1148 | Action = TargetLowering::Legal; |
1149 | break; |
1150 | case ISD::UBSANTRAP: |
1151 | Action = TLI.getOperationAction(Op: Node->getOpcode(), VT: Node->getValueType(ResNo: 0)); |
1152 | if (Action == TargetLowering::Expand) { |
1153 | // replace ISD::UBSANTRAP with ISD::TRAP |
1154 | SDValue NewVal; |
1155 | NewVal = DAG.getNode(Opcode: ISD::TRAP, DL: SDLoc(Node), VTList: Node->getVTList(), |
1156 | N: Node->getOperand(Num: 0)); |
1157 | ReplaceNode(Old: Node, New: NewVal.getNode()); |
1158 | LegalizeOp(Node: NewVal.getNode()); |
1159 | return; |
1160 | } |
1161 | break; |
1162 | case ISD::DEBUGTRAP: |
1163 | Action = TLI.getOperationAction(Op: Node->getOpcode(), VT: Node->getValueType(ResNo: 0)); |
1164 | if (Action == TargetLowering::Expand) { |
1165 | // replace ISD::DEBUGTRAP with ISD::TRAP |
1166 | SDValue NewVal; |
1167 | NewVal = DAG.getNode(Opcode: ISD::TRAP, DL: SDLoc(Node), VTList: Node->getVTList(), |
1168 | N: Node->getOperand(Num: 0)); |
1169 | ReplaceNode(Old: Node, New: NewVal.getNode()); |
1170 | LegalizeOp(Node: NewVal.getNode()); |
1171 | return; |
1172 | } |
1173 | break; |
1174 | case ISD::SADDSAT: |
1175 | case ISD::UADDSAT: |
1176 | case ISD::SSUBSAT: |
1177 | case ISD::USUBSAT: |
1178 | case ISD::SSHLSAT: |
1179 | case ISD::USHLSAT: |
1180 | case ISD::SCMP: |
1181 | case ISD::UCMP: |
1182 | case ISD::FP_TO_SINT_SAT: |
1183 | case ISD::FP_TO_UINT_SAT: |
1184 | Action = TLI.getOperationAction(Op: Node->getOpcode(), VT: Node->getValueType(ResNo: 0)); |
1185 | break; |
1186 | case ISD::SMULFIX: |
1187 | case ISD::SMULFIXSAT: |
1188 | case ISD::UMULFIX: |
1189 | case ISD::UMULFIXSAT: |
1190 | case ISD::SDIVFIX: |
1191 | case ISD::SDIVFIXSAT: |
1192 | case ISD::UDIVFIX: |
1193 | case ISD::UDIVFIXSAT: { |
1194 | unsigned Scale = Node->getConstantOperandVal(Num: 2); |
1195 | Action = TLI.getFixedPointOperationAction(Op: Node->getOpcode(), |
1196 | VT: Node->getValueType(ResNo: 0), Scale); |
1197 | break; |
1198 | } |
1199 | case ISD::MSCATTER: |
1200 | Action = TLI.getOperationAction(Op: Node->getOpcode(), |
1201 | VT: cast<MaskedScatterSDNode>(Val: Node)->getValue().getValueType()); |
1202 | break; |
1203 | case ISD::MSTORE: |
1204 | Action = TLI.getOperationAction(Op: Node->getOpcode(), |
1205 | VT: cast<MaskedStoreSDNode>(Val: Node)->getValue().getValueType()); |
1206 | break; |
1207 | case ISD::VP_SCATTER: |
1208 | Action = TLI.getOperationAction( |
1209 | Op: Node->getOpcode(), |
1210 | VT: cast<VPScatterSDNode>(Val: Node)->getValue().getValueType()); |
1211 | break; |
1212 | case ISD::VP_STORE: |
1213 | Action = TLI.getOperationAction( |
1214 | Op: Node->getOpcode(), |
1215 | VT: cast<VPStoreSDNode>(Val: Node)->getValue().getValueType()); |
1216 | break; |
1217 | case ISD::EXPERIMENTAL_VP_STRIDED_STORE: |
1218 | Action = TLI.getOperationAction( |
1219 | Op: Node->getOpcode(), |
1220 | VT: cast<VPStridedStoreSDNode>(Val: Node)->getValue().getValueType()); |
1221 | break; |
1222 | case ISD::VECREDUCE_FADD: |
1223 | case ISD::VECREDUCE_FMUL: |
1224 | case ISD::VECREDUCE_ADD: |
1225 | case ISD::VECREDUCE_MUL: |
1226 | case ISD::VECREDUCE_AND: |
1227 | case ISD::VECREDUCE_OR: |
1228 | case ISD::VECREDUCE_XOR: |
1229 | case ISD::VECREDUCE_SMAX: |
1230 | case ISD::VECREDUCE_SMIN: |
1231 | case ISD::VECREDUCE_UMAX: |
1232 | case ISD::VECREDUCE_UMIN: |
1233 | case ISD::VECREDUCE_FMAX: |
1234 | case ISD::VECREDUCE_FMIN: |
1235 | case ISD::VECREDUCE_FMAXIMUM: |
1236 | case ISD::VECREDUCE_FMINIMUM: |
1237 | case ISD::IS_FPCLASS: |
1238 | Action = TLI.getOperationAction( |
1239 | Op: Node->getOpcode(), VT: Node->getOperand(Num: 0).getValueType()); |
1240 | break; |
1241 | case ISD::VECREDUCE_SEQ_FADD: |
1242 | case ISD::VECREDUCE_SEQ_FMUL: |
1243 | case ISD::VP_REDUCE_FADD: |
1244 | case ISD::VP_REDUCE_FMUL: |
1245 | case ISD::VP_REDUCE_ADD: |
1246 | case ISD::VP_REDUCE_MUL: |
1247 | case ISD::VP_REDUCE_AND: |
1248 | case ISD::VP_REDUCE_OR: |
1249 | case ISD::VP_REDUCE_XOR: |
1250 | case ISD::VP_REDUCE_SMAX: |
1251 | case ISD::VP_REDUCE_SMIN: |
1252 | case ISD::VP_REDUCE_UMAX: |
1253 | case ISD::VP_REDUCE_UMIN: |
1254 | case ISD::VP_REDUCE_FMAX: |
1255 | case ISD::VP_REDUCE_FMIN: |
1256 | case ISD::VP_REDUCE_FMAXIMUM: |
1257 | case ISD::VP_REDUCE_FMINIMUM: |
1258 | case ISD::VP_REDUCE_SEQ_FADD: |
1259 | case ISD::VP_REDUCE_SEQ_FMUL: |
1260 | Action = TLI.getOperationAction( |
1261 | Op: Node->getOpcode(), VT: Node->getOperand(Num: 1).getValueType()); |
1262 | break; |
1263 | case ISD::VP_CTTZ_ELTS: |
1264 | case ISD::VP_CTTZ_ELTS_ZERO_UNDEF: |
1265 | Action = TLI.getOperationAction(Op: Node->getOpcode(), |
1266 | VT: Node->getOperand(Num: 0).getValueType()); |
1267 | break; |
1268 | case ISD::EXPERIMENTAL_VECTOR_HISTOGRAM: |
1269 | Action = TLI.getOperationAction( |
1270 | Op: Node->getOpcode(), |
1271 | VT: cast<MaskedHistogramSDNode>(Val: Node)->getIndex().getValueType()); |
1272 | break; |
1273 | default: |
1274 | if (Node->getOpcode() >= ISD::BUILTIN_OP_END) { |
1275 | Action = TLI.getCustomOperationAction(Op&: *Node); |
1276 | } else { |
1277 | Action = TLI.getOperationAction(Op: Node->getOpcode(), VT: Node->getValueType(ResNo: 0)); |
1278 | } |
1279 | break; |
1280 | } |
1281 | |
1282 | if (SimpleFinishLegalizing) { |
1283 | SDNode *NewNode = Node; |
1284 | switch (Node->getOpcode()) { |
1285 | default: break; |
1286 | case ISD::SHL: |
1287 | case ISD::SRL: |
1288 | case ISD::SRA: |
1289 | case ISD::ROTL: |
1290 | case ISD::ROTR: { |
1291 | // Legalizing shifts/rotates requires adjusting the shift amount |
1292 | // to the appropriate width. |
1293 | SDValue Op0 = Node->getOperand(Num: 0); |
1294 | SDValue Op1 = Node->getOperand(Num: 1); |
1295 | if (!Op1.getValueType().isVector()) { |
1296 | SDValue SAO = DAG.getShiftAmountOperand(LHSTy: Op0.getValueType(), Op: Op1); |
1297 | // The getShiftAmountOperand() may create a new operand node or |
1298 | // return the existing one. If new operand is created we need |
1299 | // to update the parent node. |
1300 | // Do not try to legalize SAO here! It will be automatically legalized |
1301 | // in the next round. |
1302 | if (SAO != Op1) |
1303 | NewNode = DAG.UpdateNodeOperands(N: Node, Op1: Op0, Op2: SAO); |
1304 | } |
1305 | } |
1306 | break; |
1307 | case ISD::FSHL: |
1308 | case ISD::FSHR: |
1309 | case ISD::SRL_PARTS: |
1310 | case ISD::SRA_PARTS: |
1311 | case ISD::SHL_PARTS: { |
1312 | // Legalizing shifts/rotates requires adjusting the shift amount |
1313 | // to the appropriate width. |
1314 | SDValue Op0 = Node->getOperand(Num: 0); |
1315 | SDValue Op1 = Node->getOperand(Num: 1); |
1316 | SDValue Op2 = Node->getOperand(Num: 2); |
1317 | if (!Op2.getValueType().isVector()) { |
1318 | SDValue SAO = DAG.getShiftAmountOperand(LHSTy: Op0.getValueType(), Op: Op2); |
1319 | // The getShiftAmountOperand() may create a new operand node or |
1320 | // return the existing one. If new operand is created we need |
1321 | // to update the parent node. |
1322 | if (SAO != Op2) |
1323 | NewNode = DAG.UpdateNodeOperands(N: Node, Op1: Op0, Op2: Op1, Op3: SAO); |
1324 | } |
1325 | break; |
1326 | } |
1327 | } |
1328 | |
1329 | if (NewNode != Node) { |
1330 | ReplaceNode(Old: Node, New: NewNode); |
1331 | Node = NewNode; |
1332 | } |
1333 | switch (Action) { |
1334 | case TargetLowering::Legal: |
1335 | LLVM_DEBUG(dbgs() << "Legal node: nothing to do\n" ); |
1336 | return; |
1337 | case TargetLowering::Custom: |
1338 | LLVM_DEBUG(dbgs() << "Trying custom legalization\n" ); |
1339 | // FIXME: The handling for custom lowering with multiple results is |
1340 | // a complete mess. |
1341 | if (SDValue Res = TLI.LowerOperation(Op: SDValue(Node, 0), DAG)) { |
1342 | if (!(Res.getNode() != Node || Res.getResNo() != 0)) |
1343 | return; |
1344 | |
1345 | if (Node->getNumValues() == 1) { |
1346 | // Verify the new types match the original. Glue is waived because |
1347 | // ISD::ADDC can be legalized by replacing Glue with an integer type. |
1348 | assert((Res.getValueType() == Node->getValueType(0) || |
1349 | Node->getValueType(0) == MVT::Glue) && |
1350 | "Type mismatch for custom legalized operation" ); |
1351 | LLVM_DEBUG(dbgs() << "Successfully custom legalized node\n" ); |
1352 | // We can just directly replace this node with the lowered value. |
1353 | ReplaceNode(Old: SDValue(Node, 0), New: Res); |
1354 | return; |
1355 | } |
1356 | |
1357 | SmallVector<SDValue, 8> ResultVals; |
1358 | for (unsigned i = 0, e = Node->getNumValues(); i != e; ++i) { |
1359 | // Verify the new types match the original. Glue is waived because |
1360 | // ISD::ADDC can be legalized by replacing Glue with an integer type. |
1361 | assert((Res->getValueType(i) == Node->getValueType(i) || |
1362 | Node->getValueType(i) == MVT::Glue) && |
1363 | "Type mismatch for custom legalized operation" ); |
1364 | ResultVals.push_back(Elt: Res.getValue(R: i)); |
1365 | } |
1366 | LLVM_DEBUG(dbgs() << "Successfully custom legalized node\n" ); |
1367 | ReplaceNode(Old: Node, New: ResultVals.data()); |
1368 | return; |
1369 | } |
1370 | LLVM_DEBUG(dbgs() << "Could not custom legalize node\n" ); |
1371 | [[fallthrough]]; |
1372 | case TargetLowering::Expand: |
1373 | if (ExpandNode(Node)) |
1374 | return; |
1375 | [[fallthrough]]; |
1376 | case TargetLowering::LibCall: |
1377 | ConvertNodeToLibcall(Node); |
1378 | return; |
1379 | case TargetLowering::Promote: |
1380 | PromoteNode(Node); |
1381 | return; |
1382 | } |
1383 | } |
1384 | |
1385 | switch (Node->getOpcode()) { |
1386 | default: |
1387 | #ifndef NDEBUG |
1388 | dbgs() << "NODE: " ; |
1389 | Node->dump( &DAG); |
1390 | dbgs() << "\n" ; |
1391 | #endif |
1392 | llvm_unreachable("Do not know how to legalize this operator!" ); |
1393 | |
1394 | case ISD::CALLSEQ_START: |
1395 | case ISD::CALLSEQ_END: |
1396 | break; |
1397 | case ISD::LOAD: |
1398 | return LegalizeLoadOps(Node); |
1399 | case ISD::STORE: |
1400 | return LegalizeStoreOps(Node); |
1401 | } |
1402 | } |
1403 | |
1404 | SDValue SelectionDAGLegalize::ExpandExtractFromVectorThroughStack(SDValue Op) { |
1405 | SDValue Vec = Op.getOperand(i: 0); |
1406 | SDValue Idx = Op.getOperand(i: 1); |
1407 | SDLoc dl(Op); |
1408 | |
1409 | // Before we generate a new store to a temporary stack slot, see if there is |
1410 | // already one that we can use. There often is because when we scalarize |
1411 | // vector operations (using SelectionDAG::UnrollVectorOp for example) a whole |
1412 | // series of EXTRACT_VECTOR_ELT nodes are generated, one for each element in |
1413 | // the vector. If all are expanded here, we don't want one store per vector |
1414 | // element. |
1415 | |
1416 | // Caches for hasPredecessorHelper |
1417 | SmallPtrSet<const SDNode *, 32> Visited; |
1418 | SmallVector<const SDNode *, 16> Worklist; |
1419 | Visited.insert(Ptr: Op.getNode()); |
1420 | Worklist.push_back(Elt: Idx.getNode()); |
1421 | SDValue StackPtr, Ch; |
1422 | for (SDNode *User : Vec.getNode()->users()) { |
1423 | if (StoreSDNode *ST = dyn_cast<StoreSDNode>(Val: User)) { |
1424 | if (ST->isIndexed() || ST->isTruncatingStore() || |
1425 | ST->getValue() != Vec) |
1426 | continue; |
1427 | |
1428 | // Make sure that nothing else could have stored into the destination of |
1429 | // this store. |
1430 | if (!ST->getChain().reachesChainWithoutSideEffects(Dest: DAG.getEntryNode())) |
1431 | continue; |
1432 | |
1433 | // If the index is dependent on the store we will introduce a cycle when |
1434 | // creating the load (the load uses the index, and by replacing the chain |
1435 | // we will make the index dependent on the load). Also, the store might be |
1436 | // dependent on the extractelement and introduce a cycle when creating |
1437 | // the load. |
1438 | if (SDNode::hasPredecessorHelper(N: ST, Visited, Worklist) || |
1439 | ST->hasPredecessor(N: Op.getNode())) |
1440 | continue; |
1441 | |
1442 | StackPtr = ST->getBasePtr(); |
1443 | Ch = SDValue(ST, 0); |
1444 | break; |
1445 | } |
1446 | } |
1447 | |
1448 | EVT VecVT = Vec.getValueType(); |
1449 | |
1450 | if (!Ch.getNode()) { |
1451 | // Store the value to a temporary stack slot, then LOAD the returned part. |
1452 | StackPtr = DAG.CreateStackTemporary(VT: VecVT); |
1453 | MachineMemOperand *StoreMMO = getStackAlignedMMO( |
1454 | StackPtr, MF&: DAG.getMachineFunction(), isObjectScalable: VecVT.isScalableVector()); |
1455 | Ch = DAG.getStore(Chain: DAG.getEntryNode(), dl, Val: Vec, Ptr: StackPtr, MMO: StoreMMO); |
1456 | } |
1457 | |
1458 | SDValue NewLoad; |
1459 | Align ElementAlignment = |
1460 | std::min(a: cast<StoreSDNode>(Val&: Ch)->getAlign(), |
1461 | b: DAG.getDataLayout().getPrefTypeAlign( |
1462 | Ty: Op.getValueType().getTypeForEVT(Context&: *DAG.getContext()))); |
1463 | |
1464 | if (Op.getValueType().isVector()) { |
1465 | StackPtr = TLI.getVectorSubVecPointer(DAG, VecPtr: StackPtr, VecVT, |
1466 | SubVecVT: Op.getValueType(), Index: Idx); |
1467 | NewLoad = DAG.getLoad(VT: Op.getValueType(), dl, Chain: Ch, Ptr: StackPtr, |
1468 | PtrInfo: MachinePointerInfo(), Alignment: ElementAlignment); |
1469 | } else { |
1470 | StackPtr = TLI.getVectorElementPointer(DAG, VecPtr: StackPtr, VecVT, Index: Idx); |
1471 | NewLoad = DAG.getExtLoad(ExtType: ISD::EXTLOAD, dl, VT: Op.getValueType(), Chain: Ch, Ptr: StackPtr, |
1472 | PtrInfo: MachinePointerInfo(), MemVT: VecVT.getVectorElementType(), |
1473 | Alignment: ElementAlignment); |
1474 | } |
1475 | |
1476 | // Replace the chain going out of the store, by the one out of the load. |
1477 | DAG.ReplaceAllUsesOfValueWith(From: Ch, To: SDValue(NewLoad.getNode(), 1)); |
1478 | |
1479 | // We introduced a cycle though, so update the loads operands, making sure |
1480 | // to use the original store's chain as an incoming chain. |
1481 | SmallVector<SDValue, 6> NewLoadOperands(NewLoad->ops()); |
1482 | NewLoadOperands[0] = Ch; |
1483 | NewLoad = |
1484 | SDValue(DAG.UpdateNodeOperands(N: NewLoad.getNode(), Ops: NewLoadOperands), 0); |
1485 | return NewLoad; |
1486 | } |
1487 | |
1488 | SDValue SelectionDAGLegalize::ExpandInsertToVectorThroughStack(SDValue Op) { |
1489 | assert(Op.getValueType().isVector() && "Non-vector insert subvector!" ); |
1490 | |
1491 | SDValue Vec = Op.getOperand(i: 0); |
1492 | SDValue Part = Op.getOperand(i: 1); |
1493 | SDValue Idx = Op.getOperand(i: 2); |
1494 | SDLoc dl(Op); |
1495 | |
1496 | // Store the value to a temporary stack slot, then LOAD the returned part. |
1497 | EVT VecVT = Vec.getValueType(); |
1498 | EVT PartVT = Part.getValueType(); |
1499 | SDValue StackPtr = DAG.CreateStackTemporary(VT: VecVT); |
1500 | int FI = cast<FrameIndexSDNode>(Val: StackPtr.getNode())->getIndex(); |
1501 | MachinePointerInfo PtrInfo = |
1502 | MachinePointerInfo::getFixedStack(MF&: DAG.getMachineFunction(), FI); |
1503 | |
1504 | // First store the whole vector. |
1505 | Align BaseVecAlignment = |
1506 | DAG.getMachineFunction().getFrameInfo().getObjectAlign(ObjectIdx: FI); |
1507 | SDValue Ch = DAG.getStore(Chain: DAG.getEntryNode(), dl, Val: Vec, Ptr: StackPtr, PtrInfo, |
1508 | Alignment: BaseVecAlignment); |
1509 | |
1510 | // Freeze the index so we don't poison the clamping code we're about to emit. |
1511 | Idx = DAG.getFreeze(V: Idx); |
1512 | |
1513 | Type *PartTy = PartVT.getTypeForEVT(Context&: *DAG.getContext()); |
1514 | Align PartAlignment = DAG.getDataLayout().getPrefTypeAlign(Ty: PartTy); |
1515 | |
1516 | // Then store the inserted part. |
1517 | if (PartVT.isVector()) { |
1518 | SDValue SubStackPtr = |
1519 | TLI.getVectorSubVecPointer(DAG, VecPtr: StackPtr, VecVT, SubVecVT: PartVT, Index: Idx); |
1520 | |
1521 | // Store the subvector. |
1522 | Ch = DAG.getStore( |
1523 | Chain: Ch, dl, Val: Part, Ptr: SubStackPtr, |
1524 | PtrInfo: MachinePointerInfo::getUnknownStack(MF&: DAG.getMachineFunction()), |
1525 | Alignment: PartAlignment); |
1526 | } else { |
1527 | SDValue SubStackPtr = |
1528 | TLI.getVectorElementPointer(DAG, VecPtr: StackPtr, VecVT, Index: Idx); |
1529 | |
1530 | // Store the scalar value. |
1531 | Ch = DAG.getTruncStore( |
1532 | Chain: Ch, dl, Val: Part, Ptr: SubStackPtr, |
1533 | PtrInfo: MachinePointerInfo::getUnknownStack(MF&: DAG.getMachineFunction()), |
1534 | SVT: VecVT.getVectorElementType(), Alignment: PartAlignment); |
1535 | } |
1536 | |
1537 | assert(cast<StoreSDNode>(Ch)->getAlign() == PartAlignment && |
1538 | "ElementAlignment does not match!" ); |
1539 | |
1540 | // Finally, load the updated vector. |
1541 | return DAG.getLoad(VT: Op.getValueType(), dl, Chain: Ch, Ptr: StackPtr, PtrInfo, |
1542 | Alignment: BaseVecAlignment); |
1543 | } |
1544 | |
1545 | SDValue SelectionDAGLegalize::ExpandConcatVectors(SDNode *Node) { |
1546 | assert(Node->getOpcode() == ISD::CONCAT_VECTORS && "Unexpected opcode!" ); |
1547 | SDLoc DL(Node); |
1548 | SmallVector<SDValue, 16> Ops; |
1549 | unsigned NumOperands = Node->getNumOperands(); |
1550 | MVT VectorIdxType = TLI.getVectorIdxTy(DL: DAG.getDataLayout()); |
1551 | EVT VectorValueType = Node->getOperand(Num: 0).getValueType(); |
1552 | unsigned NumSubElem = VectorValueType.getVectorNumElements(); |
1553 | EVT ElementValueType = TLI.getTypeToTransformTo( |
1554 | Context&: *DAG.getContext(), VT: VectorValueType.getVectorElementType()); |
1555 | for (unsigned I = 0; I < NumOperands; ++I) { |
1556 | SDValue SubOp = Node->getOperand(Num: I); |
1557 | for (unsigned Idx = 0; Idx < NumSubElem; ++Idx) { |
1558 | Ops.push_back(Elt: DAG.getNode(Opcode: ISD::EXTRACT_VECTOR_ELT, DL, VT: ElementValueType, |
1559 | N1: SubOp, |
1560 | N2: DAG.getConstant(Val: Idx, DL, VT: VectorIdxType))); |
1561 | } |
1562 | } |
1563 | return DAG.getBuildVector(VT: Node->getValueType(ResNo: 0), DL, Ops); |
1564 | } |
1565 | |
1566 | SDValue SelectionDAGLegalize::ExpandVectorBuildThroughStack(SDNode* Node) { |
1567 | assert((Node->getOpcode() == ISD::BUILD_VECTOR || |
1568 | Node->getOpcode() == ISD::CONCAT_VECTORS) && |
1569 | "Unexpected opcode!" ); |
1570 | |
1571 | // We can't handle this case efficiently. Allocate a sufficiently |
1572 | // aligned object on the stack, store each operand into it, then load |
1573 | // the result as a vector. |
1574 | // Create the stack frame object. |
1575 | EVT VT = Node->getValueType(ResNo: 0); |
1576 | EVT MemVT = isa<BuildVectorSDNode>(Val: Node) ? VT.getVectorElementType() |
1577 | : Node->getOperand(Num: 0).getValueType(); |
1578 | SDLoc dl(Node); |
1579 | SDValue FIPtr = DAG.CreateStackTemporary(VT); |
1580 | int FI = cast<FrameIndexSDNode>(Val: FIPtr.getNode())->getIndex(); |
1581 | MachinePointerInfo PtrInfo = |
1582 | MachinePointerInfo::getFixedStack(MF&: DAG.getMachineFunction(), FI); |
1583 | |
1584 | // Emit a store of each element to the stack slot. |
1585 | SmallVector<SDValue, 8> Stores; |
1586 | unsigned TypeByteSize = MemVT.getSizeInBits() / 8; |
1587 | assert(TypeByteSize > 0 && "Vector element type too small for stack store!" ); |
1588 | |
1589 | // If the destination vector element type of a BUILD_VECTOR is narrower than |
1590 | // the source element type, only store the bits necessary. |
1591 | bool Truncate = isa<BuildVectorSDNode>(Val: Node) && |
1592 | MemVT.bitsLT(VT: Node->getOperand(Num: 0).getValueType()); |
1593 | |
1594 | // Store (in the right endianness) the elements to memory. |
1595 | for (unsigned i = 0, e = Node->getNumOperands(); i != e; ++i) { |
1596 | // Ignore undef elements. |
1597 | if (Node->getOperand(Num: i).isUndef()) continue; |
1598 | |
1599 | unsigned Offset = TypeByteSize*i; |
1600 | |
1601 | SDValue Idx = |
1602 | DAG.getMemBasePlusOffset(Base: FIPtr, Offset: TypeSize::getFixed(ExactSize: Offset), DL: dl); |
1603 | |
1604 | if (Truncate) |
1605 | Stores.push_back(Elt: DAG.getTruncStore(Chain: DAG.getEntryNode(), dl, |
1606 | Val: Node->getOperand(Num: i), Ptr: Idx, |
1607 | PtrInfo: PtrInfo.getWithOffset(O: Offset), SVT: MemVT)); |
1608 | else |
1609 | Stores.push_back(Elt: DAG.getStore(Chain: DAG.getEntryNode(), dl, Val: Node->getOperand(Num: i), |
1610 | Ptr: Idx, PtrInfo: PtrInfo.getWithOffset(O: Offset))); |
1611 | } |
1612 | |
1613 | SDValue StoreChain; |
1614 | if (!Stores.empty()) // Not all undef elements? |
1615 | StoreChain = DAG.getNode(Opcode: ISD::TokenFactor, DL: dl, VT: MVT::Other, Ops: Stores); |
1616 | else |
1617 | StoreChain = DAG.getEntryNode(); |
1618 | |
1619 | // Result is a load from the stack slot. |
1620 | return DAG.getLoad(VT, dl, Chain: StoreChain, Ptr: FIPtr, PtrInfo); |
1621 | } |
1622 | |
1623 | /// Bitcast a floating-point value to an integer value. Only bitcast the part |
1624 | /// containing the sign bit if the target has no integer value capable of |
1625 | /// holding all bits of the floating-point value. |
1626 | void SelectionDAGLegalize::getSignAsIntValue(FloatSignAsInt &State, |
1627 | const SDLoc &DL, |
1628 | SDValue Value) const { |
1629 | EVT FloatVT = Value.getValueType(); |
1630 | unsigned NumBits = FloatVT.getScalarSizeInBits(); |
1631 | State.FloatVT = FloatVT; |
1632 | EVT IVT = EVT::getIntegerVT(Context&: *DAG.getContext(), BitWidth: NumBits); |
1633 | // Convert to an integer of the same size. |
1634 | if (TLI.isTypeLegal(VT: IVT)) { |
1635 | State.IntValue = DAG.getNode(Opcode: ISD::BITCAST, DL, VT: IVT, Operand: Value); |
1636 | State.SignMask = APInt::getSignMask(BitWidth: NumBits); |
1637 | State.SignBit = NumBits - 1; |
1638 | return; |
1639 | } |
1640 | |
1641 | auto &DataLayout = DAG.getDataLayout(); |
1642 | // Store the float to memory, then load the sign part out as an integer. |
1643 | MVT LoadTy = TLI.getRegisterType(VT: MVT::i8); |
1644 | // First create a temporary that is aligned for both the load and store. |
1645 | SDValue StackPtr = DAG.CreateStackTemporary(VT1: FloatVT, VT2: LoadTy); |
1646 | int FI = cast<FrameIndexSDNode>(Val: StackPtr.getNode())->getIndex(); |
1647 | // Then store the float to it. |
1648 | State.FloatPtr = StackPtr; |
1649 | MachineFunction &MF = DAG.getMachineFunction(); |
1650 | State.FloatPointerInfo = MachinePointerInfo::getFixedStack(MF, FI); |
1651 | State.Chain = DAG.getStore(Chain: DAG.getEntryNode(), dl: DL, Val: Value, Ptr: State.FloatPtr, |
1652 | PtrInfo: State.FloatPointerInfo); |
1653 | |
1654 | SDValue IntPtr; |
1655 | if (DataLayout.isBigEndian()) { |
1656 | assert(FloatVT.isByteSized() && "Unsupported floating point type!" ); |
1657 | // Load out a legal integer with the same sign bit as the float. |
1658 | IntPtr = StackPtr; |
1659 | State.IntPointerInfo = State.FloatPointerInfo; |
1660 | } else { |
1661 | // Advance the pointer so that the loaded byte will contain the sign bit. |
1662 | unsigned ByteOffset = (NumBits / 8) - 1; |
1663 | IntPtr = |
1664 | DAG.getMemBasePlusOffset(Base: StackPtr, Offset: TypeSize::getFixed(ExactSize: ByteOffset), DL); |
1665 | State.IntPointerInfo = MachinePointerInfo::getFixedStack(MF, FI, |
1666 | Offset: ByteOffset); |
1667 | } |
1668 | |
1669 | State.IntPtr = IntPtr; |
1670 | State.IntValue = DAG.getExtLoad(ExtType: ISD::EXTLOAD, dl: DL, VT: LoadTy, Chain: State.Chain, Ptr: IntPtr, |
1671 | PtrInfo: State.IntPointerInfo, MemVT: MVT::i8); |
1672 | State.SignMask = APInt::getOneBitSet(numBits: LoadTy.getScalarSizeInBits(), BitNo: 7); |
1673 | State.SignBit = 7; |
1674 | } |
1675 | |
1676 | /// Replace the integer value produced by getSignAsIntValue() with a new value |
1677 | /// and cast the result back to a floating-point type. |
1678 | SDValue SelectionDAGLegalize::modifySignAsInt(const FloatSignAsInt &State, |
1679 | const SDLoc &DL, |
1680 | SDValue NewIntValue) const { |
1681 | if (!State.Chain) |
1682 | return DAG.getNode(Opcode: ISD::BITCAST, DL, VT: State.FloatVT, Operand: NewIntValue); |
1683 | |
1684 | // Override the part containing the sign bit in the value stored on the stack. |
1685 | SDValue Chain = DAG.getTruncStore(Chain: State.Chain, dl: DL, Val: NewIntValue, Ptr: State.IntPtr, |
1686 | PtrInfo: State.IntPointerInfo, SVT: MVT::i8); |
1687 | return DAG.getLoad(VT: State.FloatVT, dl: DL, Chain, Ptr: State.FloatPtr, |
1688 | PtrInfo: State.FloatPointerInfo); |
1689 | } |
1690 | |
1691 | SDValue SelectionDAGLegalize::ExpandFCOPYSIGN(SDNode *Node) const { |
1692 | SDLoc DL(Node); |
1693 | SDValue Mag = Node->getOperand(Num: 0); |
1694 | SDValue Sign = Node->getOperand(Num: 1); |
1695 | |
1696 | // Get sign bit into an integer value. |
1697 | FloatSignAsInt SignAsInt; |
1698 | getSignAsIntValue(State&: SignAsInt, DL, Value: Sign); |
1699 | |
1700 | EVT IntVT = SignAsInt.IntValue.getValueType(); |
1701 | SDValue SignMask = DAG.getConstant(Val: SignAsInt.SignMask, DL, VT: IntVT); |
1702 | SDValue SignBit = DAG.getNode(Opcode: ISD::AND, DL, VT: IntVT, N1: SignAsInt.IntValue, |
1703 | N2: SignMask); |
1704 | |
1705 | // If FABS is legal transform |
1706 | // FCOPYSIGN(x, y) => SignBit(y) ? -FABS(x) : FABS(x) |
1707 | EVT FloatVT = Mag.getValueType(); |
1708 | if (TLI.isOperationLegalOrCustom(Op: ISD::FABS, VT: FloatVT) && |
1709 | TLI.isOperationLegalOrCustom(Op: ISD::FNEG, VT: FloatVT)) { |
1710 | SDValue AbsValue = DAG.getNode(Opcode: ISD::FABS, DL, VT: FloatVT, Operand: Mag); |
1711 | SDValue NegValue = DAG.getNode(Opcode: ISD::FNEG, DL, VT: FloatVT, Operand: AbsValue); |
1712 | SDValue Cond = DAG.getSetCC(DL, VT: getSetCCResultType(VT: IntVT), LHS: SignBit, |
1713 | RHS: DAG.getConstant(Val: 0, DL, VT: IntVT), Cond: ISD::SETNE); |
1714 | return DAG.getSelect(DL, VT: FloatVT, Cond, LHS: NegValue, RHS: AbsValue); |
1715 | } |
1716 | |
1717 | // Transform Mag value to integer, and clear the sign bit. |
1718 | FloatSignAsInt MagAsInt; |
1719 | getSignAsIntValue(State&: MagAsInt, DL, Value: Mag); |
1720 | EVT MagVT = MagAsInt.IntValue.getValueType(); |
1721 | SDValue ClearSignMask = DAG.getConstant(Val: ~MagAsInt.SignMask, DL, VT: MagVT); |
1722 | SDValue ClearedSign = DAG.getNode(Opcode: ISD::AND, DL, VT: MagVT, N1: MagAsInt.IntValue, |
1723 | N2: ClearSignMask); |
1724 | |
1725 | // Get the signbit at the right position for MagAsInt. |
1726 | int ShiftAmount = SignAsInt.SignBit - MagAsInt.SignBit; |
1727 | EVT ShiftVT = IntVT; |
1728 | if (SignBit.getScalarValueSizeInBits() < |
1729 | ClearedSign.getScalarValueSizeInBits()) { |
1730 | SignBit = DAG.getNode(Opcode: ISD::ZERO_EXTEND, DL, VT: MagVT, Operand: SignBit); |
1731 | ShiftVT = MagVT; |
1732 | } |
1733 | if (ShiftAmount > 0) { |
1734 | SDValue ShiftCnst = DAG.getConstant(Val: ShiftAmount, DL, VT: ShiftVT); |
1735 | SignBit = DAG.getNode(Opcode: ISD::SRL, DL, VT: ShiftVT, N1: SignBit, N2: ShiftCnst); |
1736 | } else if (ShiftAmount < 0) { |
1737 | SDValue ShiftCnst = DAG.getConstant(Val: -ShiftAmount, DL, VT: ShiftVT); |
1738 | SignBit = DAG.getNode(Opcode: ISD::SHL, DL, VT: ShiftVT, N1: SignBit, N2: ShiftCnst); |
1739 | } |
1740 | if (SignBit.getScalarValueSizeInBits() > |
1741 | ClearedSign.getScalarValueSizeInBits()) { |
1742 | SignBit = DAG.getNode(Opcode: ISD::TRUNCATE, DL, VT: MagVT, Operand: SignBit); |
1743 | } |
1744 | |
1745 | // Store the part with the modified sign and convert back to float. |
1746 | SDValue CopiedSign = DAG.getNode(Opcode: ISD::OR, DL, VT: MagVT, N1: ClearedSign, N2: SignBit, |
1747 | Flags: SDNodeFlags::Disjoint); |
1748 | |
1749 | return modifySignAsInt(State: MagAsInt, DL, NewIntValue: CopiedSign); |
1750 | } |
1751 | |
1752 | SDValue SelectionDAGLegalize::ExpandFNEG(SDNode *Node) const { |
1753 | // Get the sign bit as an integer. |
1754 | SDLoc DL(Node); |
1755 | FloatSignAsInt SignAsInt; |
1756 | getSignAsIntValue(State&: SignAsInt, DL, Value: Node->getOperand(Num: 0)); |
1757 | EVT IntVT = SignAsInt.IntValue.getValueType(); |
1758 | |
1759 | // Flip the sign. |
1760 | SDValue SignMask = DAG.getConstant(Val: SignAsInt.SignMask, DL, VT: IntVT); |
1761 | SDValue SignFlip = |
1762 | DAG.getNode(Opcode: ISD::XOR, DL, VT: IntVT, N1: SignAsInt.IntValue, N2: SignMask); |
1763 | |
1764 | // Convert back to float. |
1765 | return modifySignAsInt(State: SignAsInt, DL, NewIntValue: SignFlip); |
1766 | } |
1767 | |
1768 | SDValue SelectionDAGLegalize::ExpandFABS(SDNode *Node) const { |
1769 | SDLoc DL(Node); |
1770 | SDValue Value = Node->getOperand(Num: 0); |
1771 | |
1772 | // Transform FABS(x) => FCOPYSIGN(x, 0.0) if FCOPYSIGN is legal. |
1773 | EVT FloatVT = Value.getValueType(); |
1774 | if (TLI.isOperationLegalOrCustom(Op: ISD::FCOPYSIGN, VT: FloatVT)) { |
1775 | SDValue Zero = DAG.getConstantFP(Val: 0.0, DL, VT: FloatVT); |
1776 | return DAG.getNode(Opcode: ISD::FCOPYSIGN, DL, VT: FloatVT, N1: Value, N2: Zero); |
1777 | } |
1778 | |
1779 | // Transform value to integer, clear the sign bit and transform back. |
1780 | FloatSignAsInt ValueAsInt; |
1781 | getSignAsIntValue(State&: ValueAsInt, DL, Value); |
1782 | EVT IntVT = ValueAsInt.IntValue.getValueType(); |
1783 | SDValue ClearSignMask = DAG.getConstant(Val: ~ValueAsInt.SignMask, DL, VT: IntVT); |
1784 | SDValue ClearedSign = DAG.getNode(Opcode: ISD::AND, DL, VT: IntVT, N1: ValueAsInt.IntValue, |
1785 | N2: ClearSignMask); |
1786 | return modifySignAsInt(State: ValueAsInt, DL, NewIntValue: ClearedSign); |
1787 | } |
1788 | |
1789 | void SelectionDAGLegalize::ExpandDYNAMIC_STACKALLOC(SDNode* Node, |
1790 | SmallVectorImpl<SDValue> &Results) { |
1791 | Register SPReg = TLI.getStackPointerRegisterToSaveRestore(); |
1792 | assert(SPReg && "Target cannot require DYNAMIC_STACKALLOC expansion and" |
1793 | " not tell us which reg is the stack pointer!" ); |
1794 | SDLoc dl(Node); |
1795 | EVT VT = Node->getValueType(ResNo: 0); |
1796 | SDValue Tmp1 = SDValue(Node, 0); |
1797 | SDValue Tmp2 = SDValue(Node, 1); |
1798 | SDValue Tmp3 = Node->getOperand(Num: 2); |
1799 | SDValue Chain = Tmp1.getOperand(i: 0); |
1800 | |
1801 | // Chain the dynamic stack allocation so that it doesn't modify the stack |
1802 | // pointer when other instructions are using the stack. |
1803 | Chain = DAG.getCALLSEQ_START(Chain, InSize: 0, OutSize: 0, DL: dl); |
1804 | |
1805 | SDValue Size = Tmp2.getOperand(i: 1); |
1806 | SDValue SP = DAG.getCopyFromReg(Chain, dl, Reg: SPReg, VT); |
1807 | Chain = SP.getValue(R: 1); |
1808 | Align Alignment = cast<ConstantSDNode>(Val&: Tmp3)->getAlignValue(); |
1809 | const TargetFrameLowering *TFL = DAG.getSubtarget().getFrameLowering(); |
1810 | unsigned Opc = |
1811 | TFL->getStackGrowthDirection() == TargetFrameLowering::StackGrowsUp ? |
1812 | ISD::ADD : ISD::SUB; |
1813 | |
1814 | Align StackAlign = TFL->getStackAlign(); |
1815 | Tmp1 = DAG.getNode(Opcode: Opc, DL: dl, VT, N1: SP, N2: Size); // Value |
1816 | if (Alignment > StackAlign) |
1817 | Tmp1 = DAG.getNode(Opcode: ISD::AND, DL: dl, VT, N1: Tmp1, |
1818 | N2: DAG.getSignedConstant(Val: -Alignment.value(), DL: dl, VT)); |
1819 | Chain = DAG.getCopyToReg(Chain, dl, Reg: SPReg, N: Tmp1); // Output chain |
1820 | |
1821 | Tmp2 = DAG.getCALLSEQ_END(Chain, Size1: 0, Size2: 0, Glue: SDValue(), DL: dl); |
1822 | |
1823 | Results.push_back(Elt: Tmp1); |
1824 | Results.push_back(Elt: Tmp2); |
1825 | } |
1826 | |
1827 | /// Emit a store/load combination to the stack. This stores |
1828 | /// SrcOp to a stack slot of type SlotVT, truncating it if needed. It then does |
1829 | /// a load from the stack slot to DestVT, extending it if needed. |
1830 | /// The resultant code need not be legal. |
1831 | SDValue SelectionDAGLegalize::EmitStackConvert(SDValue SrcOp, EVT SlotVT, |
1832 | EVT DestVT, const SDLoc &dl) { |
1833 | return EmitStackConvert(SrcOp, SlotVT, DestVT, dl, ChainIn: DAG.getEntryNode()); |
1834 | } |
1835 | |
1836 | SDValue SelectionDAGLegalize::EmitStackConvert(SDValue SrcOp, EVT SlotVT, |
1837 | EVT DestVT, const SDLoc &dl, |
1838 | SDValue Chain) { |
1839 | EVT SrcVT = SrcOp.getValueType(); |
1840 | Type *DestType = DestVT.getTypeForEVT(Context&: *DAG.getContext()); |
1841 | Align DestAlign = DAG.getDataLayout().getPrefTypeAlign(Ty: DestType); |
1842 | |
1843 | // Don't convert with stack if the load/store is expensive. |
1844 | if ((SrcVT.bitsGT(VT: SlotVT) && |
1845 | !TLI.isTruncStoreLegalOrCustom(ValVT: SrcOp.getValueType(), MemVT: SlotVT)) || |
1846 | (SlotVT.bitsLT(VT: DestVT) && |
1847 | !TLI.isLoadExtLegalOrCustom(ExtType: ISD::EXTLOAD, ValVT: DestVT, MemVT: SlotVT))) |
1848 | return SDValue(); |
1849 | |
1850 | // Create the stack frame object. |
1851 | Align SrcAlign = DAG.getDataLayout().getPrefTypeAlign( |
1852 | Ty: SrcOp.getValueType().getTypeForEVT(Context&: *DAG.getContext())); |
1853 | SDValue FIPtr = DAG.CreateStackTemporary(Bytes: SlotVT.getStoreSize(), Alignment: SrcAlign); |
1854 | |
1855 | FrameIndexSDNode *StackPtrFI = cast<FrameIndexSDNode>(Val&: FIPtr); |
1856 | int SPFI = StackPtrFI->getIndex(); |
1857 | MachinePointerInfo PtrInfo = |
1858 | MachinePointerInfo::getFixedStack(MF&: DAG.getMachineFunction(), FI: SPFI); |
1859 | |
1860 | // Emit a store to the stack slot. Use a truncstore if the input value is |
1861 | // later than DestVT. |
1862 | SDValue Store; |
1863 | |
1864 | if (SrcVT.bitsGT(VT: SlotVT)) |
1865 | Store = DAG.getTruncStore(Chain, dl, Val: SrcOp, Ptr: FIPtr, PtrInfo, |
1866 | SVT: SlotVT, Alignment: SrcAlign); |
1867 | else { |
1868 | assert(SrcVT.bitsEq(SlotVT) && "Invalid store" ); |
1869 | Store = DAG.getStore(Chain, dl, Val: SrcOp, Ptr: FIPtr, PtrInfo, Alignment: SrcAlign); |
1870 | } |
1871 | |
1872 | // Result is a load from the stack slot. |
1873 | if (SlotVT.bitsEq(VT: DestVT)) |
1874 | return DAG.getLoad(VT: DestVT, dl, Chain: Store, Ptr: FIPtr, PtrInfo, Alignment: DestAlign); |
1875 | |
1876 | assert(SlotVT.bitsLT(DestVT) && "Unknown extension!" ); |
1877 | return DAG.getExtLoad(ExtType: ISD::EXTLOAD, dl, VT: DestVT, Chain: Store, Ptr: FIPtr, PtrInfo, MemVT: SlotVT, |
1878 | Alignment: DestAlign); |
1879 | } |
1880 | |
1881 | SDValue SelectionDAGLegalize::ExpandSCALAR_TO_VECTOR(SDNode *Node) { |
1882 | SDLoc dl(Node); |
1883 | // Create a vector sized/aligned stack slot, store the value to element #0, |
1884 | // then load the whole vector back out. |
1885 | SDValue StackPtr = DAG.CreateStackTemporary(VT: Node->getValueType(ResNo: 0)); |
1886 | |
1887 | FrameIndexSDNode *StackPtrFI = cast<FrameIndexSDNode>(Val&: StackPtr); |
1888 | int SPFI = StackPtrFI->getIndex(); |
1889 | |
1890 | SDValue Ch = DAG.getTruncStore( |
1891 | Chain: DAG.getEntryNode(), dl, Val: Node->getOperand(Num: 0), Ptr: StackPtr, |
1892 | PtrInfo: MachinePointerInfo::getFixedStack(MF&: DAG.getMachineFunction(), FI: SPFI), |
1893 | SVT: Node->getValueType(ResNo: 0).getVectorElementType()); |
1894 | return DAG.getLoad( |
1895 | VT: Node->getValueType(ResNo: 0), dl, Chain: Ch, Ptr: StackPtr, |
1896 | PtrInfo: MachinePointerInfo::getFixedStack(MF&: DAG.getMachineFunction(), FI: SPFI)); |
1897 | } |
1898 | |
1899 | static bool |
1900 | ExpandBVWithShuffles(SDNode *Node, SelectionDAG &DAG, |
1901 | const TargetLowering &TLI, SDValue &Res) { |
1902 | unsigned NumElems = Node->getNumOperands(); |
1903 | SDLoc dl(Node); |
1904 | EVT VT = Node->getValueType(ResNo: 0); |
1905 | |
1906 | // Try to group the scalars into pairs, shuffle the pairs together, then |
1907 | // shuffle the pairs of pairs together, etc. until the vector has |
1908 | // been built. This will work only if all of the necessary shuffle masks |
1909 | // are legal. |
1910 | |
1911 | // We do this in two phases; first to check the legality of the shuffles, |
1912 | // and next, assuming that all shuffles are legal, to create the new nodes. |
1913 | for (int Phase = 0; Phase < 2; ++Phase) { |
1914 | SmallVector<std::pair<SDValue, SmallVector<int, 16>>, 16> IntermedVals, |
1915 | NewIntermedVals; |
1916 | for (unsigned i = 0; i < NumElems; ++i) { |
1917 | SDValue V = Node->getOperand(Num: i); |
1918 | if (V.isUndef()) |
1919 | continue; |
1920 | |
1921 | SDValue Vec; |
1922 | if (Phase) |
1923 | Vec = DAG.getNode(Opcode: ISD::SCALAR_TO_VECTOR, DL: dl, VT, Operand: V); |
1924 | IntermedVals.push_back(Elt: std::make_pair(x&: Vec, y: SmallVector<int, 16>(1, i))); |
1925 | } |
1926 | |
1927 | while (IntermedVals.size() > 2) { |
1928 | NewIntermedVals.clear(); |
1929 | for (unsigned i = 0, e = (IntermedVals.size() & ~1u); i < e; i += 2) { |
1930 | // This vector and the next vector are shuffled together (simply to |
1931 | // append the one to the other). |
1932 | SmallVector<int, 16> ShuffleVec(NumElems, -1); |
1933 | |
1934 | SmallVector<int, 16> FinalIndices; |
1935 | FinalIndices.reserve(N: IntermedVals[i].second.size() + |
1936 | IntermedVals[i+1].second.size()); |
1937 | |
1938 | int k = 0; |
1939 | for (unsigned j = 0, f = IntermedVals[i].second.size(); j != f; |
1940 | ++j, ++k) { |
1941 | ShuffleVec[k] = j; |
1942 | FinalIndices.push_back(Elt: IntermedVals[i].second[j]); |
1943 | } |
1944 | for (unsigned j = 0, f = IntermedVals[i+1].second.size(); j != f; |
1945 | ++j, ++k) { |
1946 | ShuffleVec[k] = NumElems + j; |
1947 | FinalIndices.push_back(Elt: IntermedVals[i+1].second[j]); |
1948 | } |
1949 | |
1950 | SDValue Shuffle; |
1951 | if (Phase) |
1952 | Shuffle = DAG.getVectorShuffle(VT, dl, N1: IntermedVals[i].first, |
1953 | N2: IntermedVals[i+1].first, |
1954 | Mask: ShuffleVec); |
1955 | else if (!TLI.isShuffleMaskLegal(ShuffleVec, VT)) |
1956 | return false; |
1957 | NewIntermedVals.push_back( |
1958 | Elt: std::make_pair(x&: Shuffle, y: std::move(FinalIndices))); |
1959 | } |
1960 | |
1961 | // If we had an odd number of defined values, then append the last |
1962 | // element to the array of new vectors. |
1963 | if ((IntermedVals.size() & 1) != 0) |
1964 | NewIntermedVals.push_back(Elt: IntermedVals.back()); |
1965 | |
1966 | IntermedVals.swap(RHS&: NewIntermedVals); |
1967 | } |
1968 | |
1969 | assert(IntermedVals.size() <= 2 && IntermedVals.size() > 0 && |
1970 | "Invalid number of intermediate vectors" ); |
1971 | SDValue Vec1 = IntermedVals[0].first; |
1972 | SDValue Vec2; |
1973 | if (IntermedVals.size() > 1) |
1974 | Vec2 = IntermedVals[1].first; |
1975 | else if (Phase) |
1976 | Vec2 = DAG.getUNDEF(VT); |
1977 | |
1978 | SmallVector<int, 16> ShuffleVec(NumElems, -1); |
1979 | for (unsigned i = 0, e = IntermedVals[0].second.size(); i != e; ++i) |
1980 | ShuffleVec[IntermedVals[0].second[i]] = i; |
1981 | for (unsigned i = 0, e = IntermedVals[1].second.size(); i != e; ++i) |
1982 | ShuffleVec[IntermedVals[1].second[i]] = NumElems + i; |
1983 | |
1984 | if (Phase) |
1985 | Res = DAG.getVectorShuffle(VT, dl, N1: Vec1, N2: Vec2, Mask: ShuffleVec); |
1986 | else if (!TLI.isShuffleMaskLegal(ShuffleVec, VT)) |
1987 | return false; |
1988 | } |
1989 | |
1990 | return true; |
1991 | } |
1992 | |
1993 | /// Expand a BUILD_VECTOR node on targets that don't |
1994 | /// support the operation, but do support the resultant vector type. |
1995 | SDValue SelectionDAGLegalize::ExpandBUILD_VECTOR(SDNode *Node) { |
1996 | unsigned NumElems = Node->getNumOperands(); |
1997 | SDValue Value1, Value2; |
1998 | SDLoc dl(Node); |
1999 | EVT VT = Node->getValueType(ResNo: 0); |
2000 | EVT OpVT = Node->getOperand(Num: 0).getValueType(); |
2001 | EVT EltVT = VT.getVectorElementType(); |
2002 | |
2003 | // If the only non-undef value is the low element, turn this into a |
2004 | // SCALAR_TO_VECTOR node. If this is { X, X, X, X }, determine X. |
2005 | bool isOnlyLowElement = true; |
2006 | bool MoreThanTwoValues = false; |
2007 | bool isConstant = true; |
2008 | for (unsigned i = 0; i < NumElems; ++i) { |
2009 | SDValue V = Node->getOperand(Num: i); |
2010 | if (V.isUndef()) |
2011 | continue; |
2012 | if (i > 0) |
2013 | isOnlyLowElement = false; |
2014 | if (!isa<ConstantFPSDNode>(Val: V) && !isa<ConstantSDNode>(Val: V)) |
2015 | isConstant = false; |
2016 | |
2017 | if (!Value1.getNode()) { |
2018 | Value1 = V; |
2019 | } else if (!Value2.getNode()) { |
2020 | if (V != Value1) |
2021 | Value2 = V; |
2022 | } else if (V != Value1 && V != Value2) { |
2023 | MoreThanTwoValues = true; |
2024 | } |
2025 | } |
2026 | |
2027 | if (!Value1.getNode()) |
2028 | return DAG.getUNDEF(VT); |
2029 | |
2030 | if (isOnlyLowElement) |
2031 | return DAG.getNode(Opcode: ISD::SCALAR_TO_VECTOR, DL: dl, VT, Operand: Node->getOperand(Num: 0)); |
2032 | |
2033 | // If all elements are constants, create a load from the constant pool. |
2034 | if (isConstant) { |
2035 | SmallVector<Constant*, 16> CV; |
2036 | for (unsigned i = 0, e = NumElems; i != e; ++i) { |
2037 | if (ConstantFPSDNode *V = |
2038 | dyn_cast<ConstantFPSDNode>(Val: Node->getOperand(Num: i))) { |
2039 | CV.push_back(Elt: const_cast<ConstantFP *>(V->getConstantFPValue())); |
2040 | } else if (ConstantSDNode *V = |
2041 | dyn_cast<ConstantSDNode>(Val: Node->getOperand(Num: i))) { |
2042 | if (OpVT==EltVT) |
2043 | CV.push_back(Elt: const_cast<ConstantInt *>(V->getConstantIntValue())); |
2044 | else { |
2045 | // If OpVT and EltVT don't match, EltVT is not legal and the |
2046 | // element values have been promoted/truncated earlier. Undo this; |
2047 | // we don't want a v16i8 to become a v16i32 for example. |
2048 | const ConstantInt *CI = V->getConstantIntValue(); |
2049 | CV.push_back(Elt: ConstantInt::get(Ty: EltVT.getTypeForEVT(Context&: *DAG.getContext()), |
2050 | V: CI->getZExtValue())); |
2051 | } |
2052 | } else { |
2053 | assert(Node->getOperand(i).isUndef()); |
2054 | Type *OpNTy = EltVT.getTypeForEVT(Context&: *DAG.getContext()); |
2055 | CV.push_back(Elt: UndefValue::get(T: OpNTy)); |
2056 | } |
2057 | } |
2058 | Constant *CP = ConstantVector::get(V: CV); |
2059 | SDValue CPIdx = |
2060 | DAG.getConstantPool(C: CP, VT: TLI.getPointerTy(DL: DAG.getDataLayout())); |
2061 | Align Alignment = cast<ConstantPoolSDNode>(Val&: CPIdx)->getAlign(); |
2062 | return DAG.getLoad( |
2063 | VT, dl, Chain: DAG.getEntryNode(), Ptr: CPIdx, |
2064 | PtrInfo: MachinePointerInfo::getConstantPool(MF&: DAG.getMachineFunction()), |
2065 | Alignment); |
2066 | } |
2067 | |
2068 | SmallSet<SDValue, 16> DefinedValues; |
2069 | for (unsigned i = 0; i < NumElems; ++i) { |
2070 | if (Node->getOperand(Num: i).isUndef()) |
2071 | continue; |
2072 | DefinedValues.insert(V: Node->getOperand(Num: i)); |
2073 | } |
2074 | |
2075 | if (TLI.shouldExpandBuildVectorWithShuffles(VT, DefinedValues: DefinedValues.size())) { |
2076 | if (!MoreThanTwoValues) { |
2077 | SmallVector<int, 8> ShuffleVec(NumElems, -1); |
2078 | for (unsigned i = 0; i < NumElems; ++i) { |
2079 | SDValue V = Node->getOperand(Num: i); |
2080 | if (V.isUndef()) |
2081 | continue; |
2082 | ShuffleVec[i] = V == Value1 ? 0 : NumElems; |
2083 | } |
2084 | if (TLI.isShuffleMaskLegal(ShuffleVec, Node->getValueType(ResNo: 0))) { |
2085 | // Get the splatted value into the low element of a vector register. |
2086 | SDValue Vec1 = DAG.getNode(Opcode: ISD::SCALAR_TO_VECTOR, DL: dl, VT, Operand: Value1); |
2087 | SDValue Vec2; |
2088 | if (Value2.getNode()) |
2089 | Vec2 = DAG.getNode(Opcode: ISD::SCALAR_TO_VECTOR, DL: dl, VT, Operand: Value2); |
2090 | else |
2091 | Vec2 = DAG.getUNDEF(VT); |
2092 | |
2093 | // Return shuffle(LowValVec, undef, <0,0,0,0>) |
2094 | return DAG.getVectorShuffle(VT, dl, N1: Vec1, N2: Vec2, Mask: ShuffleVec); |
2095 | } |
2096 | } else { |
2097 | SDValue Res; |
2098 | if (ExpandBVWithShuffles(Node, DAG, TLI, Res)) |
2099 | return Res; |
2100 | } |
2101 | } |
2102 | |
2103 | // Otherwise, we can't handle this case efficiently. |
2104 | return ExpandVectorBuildThroughStack(Node); |
2105 | } |
2106 | |
2107 | SDValue SelectionDAGLegalize::ExpandSPLAT_VECTOR(SDNode *Node) { |
2108 | SDLoc DL(Node); |
2109 | EVT VT = Node->getValueType(ResNo: 0); |
2110 | SDValue SplatVal = Node->getOperand(Num: 0); |
2111 | |
2112 | return DAG.getSplatBuildVector(VT, DL, Op: SplatVal); |
2113 | } |
2114 | |
2115 | // Expand a node into a call to a libcall, returning the value as the first |
2116 | // result and the chain as the second. If the result value does not fit into a |
2117 | // register, return the lo part and set the hi part to the by-reg argument in |
2118 | // the first. If it does fit into a single register, return the result and |
2119 | // leave the Hi part unset. |
2120 | std::pair<SDValue, SDValue> |
2121 | SelectionDAGLegalize::ExpandLibCall(RTLIB::Libcall LC, SDNode *Node, |
2122 | TargetLowering::ArgListTy &&Args, |
2123 | bool IsSigned, EVT RetVT) { |
2124 | EVT CodePtrTy = TLI.getPointerTy(DL: DAG.getDataLayout()); |
2125 | SDValue Callee; |
2126 | if (const char *LibcallName = TLI.getLibcallName(Call: LC)) |
2127 | Callee = DAG.getExternalSymbol(Sym: LibcallName, VT: CodePtrTy); |
2128 | else { |
2129 | Callee = DAG.getUNDEF(VT: CodePtrTy); |
2130 | DAG.getContext()->emitError(ErrorStr: Twine("no libcall available for " ) + |
2131 | Node->getOperationName(G: &DAG)); |
2132 | } |
2133 | |
2134 | Type *RetTy = RetVT.getTypeForEVT(Context&: *DAG.getContext()); |
2135 | |
2136 | // By default, the input chain to this libcall is the entry node of the |
2137 | // function. If the libcall is going to be emitted as a tail call then |
2138 | // TLI.isUsedByReturnOnly will change it to the right chain if the return |
2139 | // node which is being folded has a non-entry input chain. |
2140 | SDValue InChain = DAG.getEntryNode(); |
2141 | |
2142 | // isTailCall may be true since the callee does not reference caller stack |
2143 | // frame. Check if it's in the right position and that the return types match. |
2144 | SDValue TCChain = InChain; |
2145 | const Function &F = DAG.getMachineFunction().getFunction(); |
2146 | bool isTailCall = |
2147 | TLI.isInTailCallPosition(DAG, Node, Chain&: TCChain) && |
2148 | (RetTy == F.getReturnType() || F.getReturnType()->isVoidTy()); |
2149 | if (isTailCall) |
2150 | InChain = TCChain; |
2151 | |
2152 | TargetLowering::CallLoweringInfo CLI(DAG); |
2153 | bool signExtend = TLI.shouldSignExtendTypeInLibCall(Ty: RetTy, IsSigned); |
2154 | CLI.setDebugLoc(SDLoc(Node)) |
2155 | .setChain(InChain) |
2156 | .setLibCallee(CC: TLI.getLibcallCallingConv(Call: LC), ResultType: RetTy, Target: Callee, |
2157 | ArgsList: std::move(Args)) |
2158 | .setTailCall(isTailCall) |
2159 | .setSExtResult(signExtend) |
2160 | .setZExtResult(!signExtend) |
2161 | .setIsPostTypeLegalization(true); |
2162 | |
2163 | std::pair<SDValue, SDValue> CallInfo = TLI.LowerCallTo(CLI); |
2164 | |
2165 | if (!CallInfo.second.getNode()) { |
2166 | LLVM_DEBUG(dbgs() << "Created tailcall: " ; DAG.getRoot().dump(&DAG)); |
2167 | // It's a tailcall, return the chain (which is the DAG root). |
2168 | return {DAG.getRoot(), DAG.getRoot()}; |
2169 | } |
2170 | |
2171 | LLVM_DEBUG(dbgs() << "Created libcall: " ; CallInfo.first.dump(&DAG)); |
2172 | return CallInfo; |
2173 | } |
2174 | |
2175 | std::pair<SDValue, SDValue> SelectionDAGLegalize::ExpandLibCall(RTLIB::Libcall LC, SDNode *Node, |
2176 | bool isSigned) { |
2177 | TargetLowering::ArgListTy Args; |
2178 | TargetLowering::ArgListEntry Entry; |
2179 | for (const SDValue &Op : Node->op_values()) { |
2180 | EVT ArgVT = Op.getValueType(); |
2181 | Type *ArgTy = ArgVT.getTypeForEVT(Context&: *DAG.getContext()); |
2182 | Entry.Node = Op; |
2183 | Entry.Ty = ArgTy; |
2184 | Entry.IsSExt = TLI.shouldSignExtendTypeInLibCall(Ty: ArgTy, IsSigned: isSigned); |
2185 | Entry.IsZExt = !Entry.IsSExt; |
2186 | Args.push_back(x: Entry); |
2187 | } |
2188 | |
2189 | return ExpandLibCall(LC, Node, Args: std::move(Args), IsSigned: isSigned, |
2190 | RetVT: Node->getValueType(ResNo: 0)); |
2191 | } |
2192 | |
2193 | void SelectionDAGLegalize::ExpandFPLibCall(SDNode* Node, |
2194 | RTLIB::Libcall LC, |
2195 | SmallVectorImpl<SDValue> &Results) { |
2196 | if (LC == RTLIB::UNKNOWN_LIBCALL) |
2197 | llvm_unreachable("Can't create an unknown libcall!" ); |
2198 | |
2199 | if (Node->isStrictFPOpcode()) { |
2200 | EVT RetVT = Node->getValueType(ResNo: 0); |
2201 | SmallVector<SDValue, 4> Ops(drop_begin(RangeOrContainer: Node->ops())); |
2202 | TargetLowering::MakeLibCallOptions CallOptions; |
2203 | CallOptions.IsPostTypeLegalization = true; |
2204 | // FIXME: This doesn't support tail calls. |
2205 | std::pair<SDValue, SDValue> Tmp = TLI.makeLibCall(DAG, LC, RetVT, |
2206 | Ops, CallOptions, |
2207 | dl: SDLoc(Node), |
2208 | Chain: Node->getOperand(Num: 0)); |
2209 | Results.push_back(Elt: Tmp.first); |
2210 | Results.push_back(Elt: Tmp.second); |
2211 | } else { |
2212 | bool IsSignedArgument = Node->getOpcode() == ISD::FLDEXP; |
2213 | SDValue Tmp = ExpandLibCall(LC, Node, isSigned: IsSignedArgument).first; |
2214 | Results.push_back(Elt: Tmp); |
2215 | } |
2216 | } |
2217 | |
2218 | /// Expand the node to a libcall based on the result type. |
2219 | void SelectionDAGLegalize::ExpandFPLibCall(SDNode* Node, |
2220 | RTLIB::Libcall Call_F32, |
2221 | RTLIB::Libcall Call_F64, |
2222 | RTLIB::Libcall Call_F80, |
2223 | RTLIB::Libcall Call_F128, |
2224 | RTLIB::Libcall Call_PPCF128, |
2225 | SmallVectorImpl<SDValue> &Results) { |
2226 | RTLIB::Libcall LC = RTLIB::getFPLibCall(VT: Node->getSimpleValueType(ResNo: 0), |
2227 | Call_F32, Call_F64, Call_F80, |
2228 | Call_F128, Call_PPCF128); |
2229 | ExpandFPLibCall(Node, LC, Results); |
2230 | } |
2231 | |
2232 | SDValue SelectionDAGLegalize::ExpandIntLibCall(SDNode* Node, bool isSigned, |
2233 | RTLIB::Libcall Call_I8, |
2234 | RTLIB::Libcall Call_I16, |
2235 | RTLIB::Libcall Call_I32, |
2236 | RTLIB::Libcall Call_I64, |
2237 | RTLIB::Libcall Call_I128) { |
2238 | RTLIB::Libcall LC; |
2239 | switch (Node->getSimpleValueType(ResNo: 0).SimpleTy) { |
2240 | default: llvm_unreachable("Unexpected request for libcall!" ); |
2241 | case MVT::i8: LC = Call_I8; break; |
2242 | case MVT::i16: LC = Call_I16; break; |
2243 | case MVT::i32: LC = Call_I32; break; |
2244 | case MVT::i64: LC = Call_I64; break; |
2245 | case MVT::i128: LC = Call_I128; break; |
2246 | } |
2247 | return ExpandLibCall(LC, Node, isSigned).first; |
2248 | } |
2249 | |
2250 | /// Expand the node to a libcall based on first argument type (for instance |
2251 | /// lround and its variant). |
2252 | void SelectionDAGLegalize::ExpandArgFPLibCall(SDNode* Node, |
2253 | RTLIB::Libcall Call_F32, |
2254 | RTLIB::Libcall Call_F64, |
2255 | RTLIB::Libcall Call_F80, |
2256 | RTLIB::Libcall Call_F128, |
2257 | RTLIB::Libcall Call_PPCF128, |
2258 | SmallVectorImpl<SDValue> &Results) { |
2259 | EVT InVT = Node->getOperand(Num: Node->isStrictFPOpcode() ? 1 : 0).getValueType(); |
2260 | RTLIB::Libcall LC = RTLIB::getFPLibCall(VT: InVT.getSimpleVT(), |
2261 | Call_F32, Call_F64, Call_F80, |
2262 | Call_F128, Call_PPCF128); |
2263 | ExpandFPLibCall(Node, LC, Results); |
2264 | } |
2265 | |
2266 | SDValue SelectionDAGLegalize::ExpandBitCountingLibCall( |
2267 | SDNode *Node, RTLIB::Libcall CallI32, RTLIB::Libcall CallI64, |
2268 | RTLIB::Libcall CallI128) { |
2269 | RTLIB::Libcall LC; |
2270 | switch (Node->getSimpleValueType(ResNo: 0).SimpleTy) { |
2271 | default: |
2272 | llvm_unreachable("Unexpected request for libcall!" ); |
2273 | case MVT::i32: |
2274 | LC = CallI32; |
2275 | break; |
2276 | case MVT::i64: |
2277 | LC = CallI64; |
2278 | break; |
2279 | case MVT::i128: |
2280 | LC = CallI128; |
2281 | break; |
2282 | } |
2283 | |
2284 | // Bit-counting libcalls have one unsigned argument and return `int`. |
2285 | // Note that `int` may be illegal on this target; ExpandLibCall will |
2286 | // take care of promoting it to a legal type. |
2287 | SDValue Op = Node->getOperand(Num: 0); |
2288 | EVT IntVT = |
2289 | EVT::getIntegerVT(Context&: *DAG.getContext(), BitWidth: DAG.getLibInfo().getIntSize()); |
2290 | |
2291 | TargetLowering::ArgListEntry Arg; |
2292 | EVT ArgVT = Op.getValueType(); |
2293 | Type *ArgTy = ArgVT.getTypeForEVT(Context&: *DAG.getContext()); |
2294 | Arg.Node = Op; |
2295 | Arg.Ty = ArgTy; |
2296 | Arg.IsSExt = TLI.shouldSignExtendTypeInLibCall(Ty: ArgTy, /*IsSigned=*/false); |
2297 | Arg.IsZExt = !Arg.IsSExt; |
2298 | |
2299 | SDValue Res = ExpandLibCall(LC, Node, Args: TargetLowering::ArgListTy{Arg}, |
2300 | /*IsSigned=*/true, RetVT: IntVT) |
2301 | .first; |
2302 | |
2303 | // If ExpandLibCall created a tail call, the result was already |
2304 | // of the correct type. Otherwise, we need to sign extend it. |
2305 | if (Res.getValueType() != MVT::Other) |
2306 | Res = DAG.getSExtOrTrunc(Op: Res, DL: SDLoc(Node), VT: Node->getValueType(ResNo: 0)); |
2307 | return Res; |
2308 | } |
2309 | |
2310 | /// Issue libcalls to __{u}divmod to compute div / rem pairs. |
2311 | void |
2312 | SelectionDAGLegalize::ExpandDivRemLibCall(SDNode *Node, |
2313 | SmallVectorImpl<SDValue> &Results) { |
2314 | unsigned Opcode = Node->getOpcode(); |
2315 | bool isSigned = Opcode == ISD::SDIVREM; |
2316 | |
2317 | RTLIB::Libcall LC; |
2318 | switch (Node->getSimpleValueType(ResNo: 0).SimpleTy) { |
2319 | default: llvm_unreachable("Unexpected request for libcall!" ); |
2320 | case MVT::i8: LC= isSigned ? RTLIB::SDIVREM_I8 : RTLIB::UDIVREM_I8; break; |
2321 | case MVT::i16: LC= isSigned ? RTLIB::SDIVREM_I16 : RTLIB::UDIVREM_I16; break; |
2322 | case MVT::i32: LC= isSigned ? RTLIB::SDIVREM_I32 : RTLIB::UDIVREM_I32; break; |
2323 | case MVT::i64: LC= isSigned ? RTLIB::SDIVREM_I64 : RTLIB::UDIVREM_I64; break; |
2324 | case MVT::i128: LC= isSigned ? RTLIB::SDIVREM_I128:RTLIB::UDIVREM_I128; break; |
2325 | } |
2326 | |
2327 | // The input chain to this libcall is the entry node of the function. |
2328 | // Legalizing the call will automatically add the previous call to the |
2329 | // dependence. |
2330 | SDValue InChain = DAG.getEntryNode(); |
2331 | |
2332 | EVT RetVT = Node->getValueType(ResNo: 0); |
2333 | Type *RetTy = RetVT.getTypeForEVT(Context&: *DAG.getContext()); |
2334 | |
2335 | TargetLowering::ArgListTy Args; |
2336 | TargetLowering::ArgListEntry Entry; |
2337 | for (const SDValue &Op : Node->op_values()) { |
2338 | EVT ArgVT = Op.getValueType(); |
2339 | Type *ArgTy = ArgVT.getTypeForEVT(Context&: *DAG.getContext()); |
2340 | Entry.Node = Op; |
2341 | Entry.Ty = ArgTy; |
2342 | Entry.IsSExt = isSigned; |
2343 | Entry.IsZExt = !isSigned; |
2344 | Args.push_back(x: Entry); |
2345 | } |
2346 | |
2347 | // Also pass the return address of the remainder. |
2348 | SDValue FIPtr = DAG.CreateStackTemporary(VT: RetVT); |
2349 | Entry.Node = FIPtr; |
2350 | Entry.Ty = PointerType::getUnqual(C&: RetTy->getContext()); |
2351 | Entry.IsSExt = isSigned; |
2352 | Entry.IsZExt = !isSigned; |
2353 | Args.push_back(x: Entry); |
2354 | |
2355 | SDValue Callee = DAG.getExternalSymbol(Sym: TLI.getLibcallName(Call: LC), |
2356 | VT: TLI.getPointerTy(DL: DAG.getDataLayout())); |
2357 | |
2358 | SDLoc dl(Node); |
2359 | TargetLowering::CallLoweringInfo CLI(DAG); |
2360 | CLI.setDebugLoc(dl) |
2361 | .setChain(InChain) |
2362 | .setLibCallee(CC: TLI.getLibcallCallingConv(Call: LC), ResultType: RetTy, Target: Callee, |
2363 | ArgsList: std::move(Args)) |
2364 | .setSExtResult(isSigned) |
2365 | .setZExtResult(!isSigned); |
2366 | |
2367 | std::pair<SDValue, SDValue> CallInfo = TLI.LowerCallTo(CLI); |
2368 | |
2369 | // Remainder is loaded back from the stack frame. |
2370 | SDValue Rem = |
2371 | DAG.getLoad(VT: RetVT, dl, Chain: CallInfo.second, Ptr: FIPtr, PtrInfo: MachinePointerInfo()); |
2372 | Results.push_back(Elt: CallInfo.first); |
2373 | Results.push_back(Elt: Rem); |
2374 | } |
2375 | |
2376 | /// Return true if sincos libcall is available. |
2377 | static bool isSinCosLibcallAvailable(SDNode *Node, const TargetLowering &TLI) { |
2378 | RTLIB::Libcall LC = RTLIB::getSINCOS(RetVT: Node->getSimpleValueType(ResNo: 0).SimpleTy); |
2379 | return TLI.getLibcallName(Call: LC) != nullptr; |
2380 | } |
2381 | |
2382 | /// Only issue sincos libcall if both sin and cos are needed. |
2383 | static bool useSinCos(SDNode *Node) { |
2384 | unsigned OtherOpcode = Node->getOpcode() == ISD::FSIN |
2385 | ? ISD::FCOS : ISD::FSIN; |
2386 | |
2387 | SDValue Op0 = Node->getOperand(Num: 0); |
2388 | for (const SDNode *User : Op0.getNode()->users()) { |
2389 | if (User == Node) |
2390 | continue; |
2391 | // The other user might have been turned into sincos already. |
2392 | if (User->getOpcode() == OtherOpcode || User->getOpcode() == ISD::FSINCOS) |
2393 | return true; |
2394 | } |
2395 | return false; |
2396 | } |
2397 | |
2398 | SDValue SelectionDAGLegalize::expandLdexp(SDNode *Node) const { |
2399 | SDLoc dl(Node); |
2400 | EVT VT = Node->getValueType(ResNo: 0); |
2401 | SDValue X = Node->getOperand(Num: 0); |
2402 | SDValue N = Node->getOperand(Num: 1); |
2403 | EVT ExpVT = N.getValueType(); |
2404 | EVT AsIntVT = VT.changeTypeToInteger(); |
2405 | if (AsIntVT == EVT()) // TODO: How to handle f80? |
2406 | return SDValue(); |
2407 | |
2408 | if (Node->getOpcode() == ISD::STRICT_FLDEXP) // TODO |
2409 | return SDValue(); |
2410 | |
2411 | SDNodeFlags NSW; |
2412 | NSW.setNoSignedWrap(true); |
2413 | SDNodeFlags NUW_NSW; |
2414 | NUW_NSW.setNoUnsignedWrap(true); |
2415 | NUW_NSW.setNoSignedWrap(true); |
2416 | |
2417 | EVT SetCCVT = |
2418 | TLI.getSetCCResultType(DL: DAG.getDataLayout(), Context&: *DAG.getContext(), VT: ExpVT); |
2419 | const fltSemantics &FltSem = VT.getFltSemantics(); |
2420 | |
2421 | const APFloat::ExponentType MaxExpVal = APFloat::semanticsMaxExponent(FltSem); |
2422 | const APFloat::ExponentType MinExpVal = APFloat::semanticsMinExponent(FltSem); |
2423 | const int Precision = APFloat::semanticsPrecision(FltSem); |
2424 | |
2425 | const SDValue MaxExp = DAG.getSignedConstant(Val: MaxExpVal, DL: dl, VT: ExpVT); |
2426 | const SDValue MinExp = DAG.getSignedConstant(Val: MinExpVal, DL: dl, VT: ExpVT); |
2427 | |
2428 | const SDValue DoubleMaxExp = DAG.getSignedConstant(Val: 2 * MaxExpVal, DL: dl, VT: ExpVT); |
2429 | |
2430 | const APFloat One(FltSem, "1.0" ); |
2431 | APFloat ScaleUpK = scalbn(X: One, Exp: MaxExpVal, RM: APFloat::rmNearestTiesToEven); |
2432 | |
2433 | // Offset by precision to avoid denormal range. |
2434 | APFloat ScaleDownK = |
2435 | scalbn(X: One, Exp: MinExpVal + Precision, RM: APFloat::rmNearestTiesToEven); |
2436 | |
2437 | // TODO: Should really introduce control flow and use a block for the > |
2438 | // MaxExp, < MinExp cases |
2439 | |
2440 | // First, handle exponents Exp > MaxExp and scale down. |
2441 | SDValue NGtMaxExp = DAG.getSetCC(DL: dl, VT: SetCCVT, LHS: N, RHS: MaxExp, Cond: ISD::SETGT); |
2442 | |
2443 | SDValue DecN0 = DAG.getNode(Opcode: ISD::SUB, DL: dl, VT: ExpVT, N1: N, N2: MaxExp, Flags: NSW); |
2444 | SDValue ClampMaxVal = DAG.getConstant(Val: 3 * MaxExpVal, DL: dl, VT: ExpVT); |
2445 | SDValue ClampN_Big = DAG.getNode(Opcode: ISD::SMIN, DL: dl, VT: ExpVT, N1: N, N2: ClampMaxVal); |
2446 | SDValue DecN1 = |
2447 | DAG.getNode(Opcode: ISD::SUB, DL: dl, VT: ExpVT, N1: ClampN_Big, N2: DoubleMaxExp, Flags: NSW); |
2448 | |
2449 | SDValue ScaleUpTwice = |
2450 | DAG.getSetCC(DL: dl, VT: SetCCVT, LHS: N, RHS: DoubleMaxExp, Cond: ISD::SETUGT); |
2451 | |
2452 | const SDValue ScaleUpVal = DAG.getConstantFP(Val: ScaleUpK, DL: dl, VT); |
2453 | SDValue ScaleUp0 = DAG.getNode(Opcode: ISD::FMUL, DL: dl, VT, N1: X, N2: ScaleUpVal); |
2454 | SDValue ScaleUp1 = DAG.getNode(Opcode: ISD::FMUL, DL: dl, VT, N1: ScaleUp0, N2: ScaleUpVal); |
2455 | |
2456 | SDValue SelectN_Big = |
2457 | DAG.getNode(Opcode: ISD::SELECT, DL: dl, VT: ExpVT, N1: ScaleUpTwice, N2: DecN1, N3: DecN0); |
2458 | SDValue SelectX_Big = |
2459 | DAG.getNode(Opcode: ISD::SELECT, DL: dl, VT, N1: ScaleUpTwice, N2: ScaleUp1, N3: ScaleUp0); |
2460 | |
2461 | // Now handle exponents Exp < MinExp |
2462 | SDValue NLtMinExp = DAG.getSetCC(DL: dl, VT: SetCCVT, LHS: N, RHS: MinExp, Cond: ISD::SETLT); |
2463 | |
2464 | SDValue Increment0 = DAG.getConstant(Val: -(MinExpVal + Precision), DL: dl, VT: ExpVT); |
2465 | SDValue Increment1 = DAG.getConstant(Val: -2 * (MinExpVal + Precision), DL: dl, VT: ExpVT); |
2466 | |
2467 | SDValue IncN0 = DAG.getNode(Opcode: ISD::ADD, DL: dl, VT: ExpVT, N1: N, N2: Increment0, Flags: NUW_NSW); |
2468 | |
2469 | SDValue ClampMinVal = |
2470 | DAG.getSignedConstant(Val: 3 * MinExpVal + 2 * Precision, DL: dl, VT: ExpVT); |
2471 | SDValue ClampN_Small = DAG.getNode(Opcode: ISD::SMAX, DL: dl, VT: ExpVT, N1: N, N2: ClampMinVal); |
2472 | SDValue IncN1 = |
2473 | DAG.getNode(Opcode: ISD::ADD, DL: dl, VT: ExpVT, N1: ClampN_Small, N2: Increment1, Flags: NSW); |
2474 | |
2475 | const SDValue ScaleDownVal = DAG.getConstantFP(Val: ScaleDownK, DL: dl, VT); |
2476 | SDValue ScaleDown0 = DAG.getNode(Opcode: ISD::FMUL, DL: dl, VT, N1: X, N2: ScaleDownVal); |
2477 | SDValue ScaleDown1 = DAG.getNode(Opcode: ISD::FMUL, DL: dl, VT, N1: ScaleDown0, N2: ScaleDownVal); |
2478 | |
2479 | SDValue ScaleDownTwice = DAG.getSetCC( |
2480 | DL: dl, VT: SetCCVT, LHS: N, |
2481 | RHS: DAG.getSignedConstant(Val: 2 * MinExpVal + Precision, DL: dl, VT: ExpVT), Cond: ISD::SETULT); |
2482 | |
2483 | SDValue SelectN_Small = |
2484 | DAG.getNode(Opcode: ISD::SELECT, DL: dl, VT: ExpVT, N1: ScaleDownTwice, N2: IncN1, N3: IncN0); |
2485 | SDValue SelectX_Small = |
2486 | DAG.getNode(Opcode: ISD::SELECT, DL: dl, VT, N1: ScaleDownTwice, N2: ScaleDown1, N3: ScaleDown0); |
2487 | |
2488 | // Now combine the two out of range exponent handling cases with the base |
2489 | // case. |
2490 | SDValue NewX = DAG.getNode( |
2491 | Opcode: ISD::SELECT, DL: dl, VT, N1: NGtMaxExp, N2: SelectX_Big, |
2492 | N3: DAG.getNode(Opcode: ISD::SELECT, DL: dl, VT, N1: NLtMinExp, N2: SelectX_Small, N3: X)); |
2493 | |
2494 | SDValue NewN = DAG.getNode( |
2495 | Opcode: ISD::SELECT, DL: dl, VT: ExpVT, N1: NGtMaxExp, N2: SelectN_Big, |
2496 | N3: DAG.getNode(Opcode: ISD::SELECT, DL: dl, VT: ExpVT, N1: NLtMinExp, N2: SelectN_Small, N3: N)); |
2497 | |
2498 | SDValue BiasedN = DAG.getNode(Opcode: ISD::ADD, DL: dl, VT: ExpVT, N1: NewN, N2: MaxExp, Flags: NSW); |
2499 | |
2500 | SDValue ExponentShiftAmt = |
2501 | DAG.getShiftAmountConstant(Val: Precision - 1, VT: ExpVT, DL: dl); |
2502 | SDValue CastExpToValTy = DAG.getZExtOrTrunc(Op: BiasedN, DL: dl, VT: AsIntVT); |
2503 | |
2504 | SDValue AsInt = DAG.getNode(Opcode: ISD::SHL, DL: dl, VT: AsIntVT, N1: CastExpToValTy, |
2505 | N2: ExponentShiftAmt, Flags: NUW_NSW); |
2506 | SDValue AsFP = DAG.getNode(Opcode: ISD::BITCAST, DL: dl, VT, Operand: AsInt); |
2507 | return DAG.getNode(Opcode: ISD::FMUL, DL: dl, VT, N1: NewX, N2: AsFP); |
2508 | } |
2509 | |
2510 | SDValue SelectionDAGLegalize::expandFrexp(SDNode *Node) const { |
2511 | SDLoc dl(Node); |
2512 | SDValue Val = Node->getOperand(Num: 0); |
2513 | EVT VT = Val.getValueType(); |
2514 | EVT ExpVT = Node->getValueType(ResNo: 1); |
2515 | EVT AsIntVT = VT.changeTypeToInteger(); |
2516 | if (AsIntVT == EVT()) // TODO: How to handle f80? |
2517 | return SDValue(); |
2518 | |
2519 | const fltSemantics &FltSem = VT.getFltSemantics(); |
2520 | const APFloat::ExponentType MinExpVal = APFloat::semanticsMinExponent(FltSem); |
2521 | const unsigned Precision = APFloat::semanticsPrecision(FltSem); |
2522 | const unsigned BitSize = VT.getScalarSizeInBits(); |
2523 | |
2524 | // TODO: Could introduce control flow and skip over the denormal handling. |
2525 | |
2526 | // scale_up = fmul value, scalbn(1.0, precision + 1) |
2527 | // extracted_exp = (bitcast value to uint) >> precision - 1 |
2528 | // biased_exp = extracted_exp + min_exp |
2529 | // extracted_fract = (bitcast value to uint) & (fract_mask | sign_mask) |
2530 | // |
2531 | // is_denormal = val < smallest_normalized |
2532 | // computed_fract = is_denormal ? scale_up : extracted_fract |
2533 | // computed_exp = is_denormal ? biased_exp + (-precision - 1) : biased_exp |
2534 | // |
2535 | // result_0 = (!isfinite(val) || iszero(val)) ? val : computed_fract |
2536 | // result_1 = (!isfinite(val) || iszero(val)) ? 0 : computed_exp |
2537 | |
2538 | SDValue NegSmallestNormalizedInt = DAG.getConstant( |
2539 | Val: APFloat::getSmallestNormalized(Sem: FltSem, Negative: true).bitcastToAPInt(), DL: dl, |
2540 | VT: AsIntVT); |
2541 | |
2542 | SDValue SmallestNormalizedInt = DAG.getConstant( |
2543 | Val: APFloat::getSmallestNormalized(Sem: FltSem, Negative: false).bitcastToAPInt(), DL: dl, |
2544 | VT: AsIntVT); |
2545 | |
2546 | // Masks out the exponent bits. |
2547 | SDValue ExpMask = |
2548 | DAG.getConstant(Val: APFloat::getInf(Sem: FltSem).bitcastToAPInt(), DL: dl, VT: AsIntVT); |
2549 | |
2550 | // Mask out the exponent part of the value. |
2551 | // |
2552 | // e.g, for f32 FractSignMaskVal = 0x807fffff |
2553 | APInt FractSignMaskVal = APInt::getBitsSet(numBits: BitSize, loBit: 0, hiBit: Precision - 1); |
2554 | FractSignMaskVal.setBit(BitSize - 1); // Set the sign bit |
2555 | |
2556 | APInt SignMaskVal = APInt::getSignedMaxValue(numBits: BitSize); |
2557 | SDValue SignMask = DAG.getConstant(Val: SignMaskVal, DL: dl, VT: AsIntVT); |
2558 | |
2559 | SDValue FractSignMask = DAG.getConstant(Val: FractSignMaskVal, DL: dl, VT: AsIntVT); |
2560 | |
2561 | const APFloat One(FltSem, "1.0" ); |
2562 | // Scale a possible denormal input. |
2563 | // e.g., for f64, 0x1p+54 |
2564 | APFloat ScaleUpKVal = |
2565 | scalbn(X: One, Exp: Precision + 1, RM: APFloat::rmNearestTiesToEven); |
2566 | |
2567 | SDValue ScaleUpK = DAG.getConstantFP(Val: ScaleUpKVal, DL: dl, VT); |
2568 | SDValue ScaleUp = DAG.getNode(Opcode: ISD::FMUL, DL: dl, VT, N1: Val, N2: ScaleUpK); |
2569 | |
2570 | EVT SetCCVT = |
2571 | TLI.getSetCCResultType(DL: DAG.getDataLayout(), Context&: *DAG.getContext(), VT); |
2572 | |
2573 | SDValue AsInt = DAG.getNode(Opcode: ISD::BITCAST, DL: dl, VT: AsIntVT, Operand: Val); |
2574 | |
2575 | SDValue Abs = DAG.getNode(Opcode: ISD::AND, DL: dl, VT: AsIntVT, N1: AsInt, N2: SignMask); |
2576 | |
2577 | SDValue AddNegSmallestNormal = |
2578 | DAG.getNode(Opcode: ISD::ADD, DL: dl, VT: AsIntVT, N1: Abs, N2: NegSmallestNormalizedInt); |
2579 | SDValue DenormOrZero = DAG.getSetCC(DL: dl, VT: SetCCVT, LHS: AddNegSmallestNormal, |
2580 | RHS: NegSmallestNormalizedInt, Cond: ISD::SETULE); |
2581 | |
2582 | SDValue IsDenormal = |
2583 | DAG.getSetCC(DL: dl, VT: SetCCVT, LHS: Abs, RHS: SmallestNormalizedInt, Cond: ISD::SETULT); |
2584 | |
2585 | SDValue MinExp = DAG.getSignedConstant(Val: MinExpVal, DL: dl, VT: ExpVT); |
2586 | SDValue Zero = DAG.getConstant(Val: 0, DL: dl, VT: ExpVT); |
2587 | |
2588 | SDValue ScaledAsInt = DAG.getNode(Opcode: ISD::BITCAST, DL: dl, VT: AsIntVT, Operand: ScaleUp); |
2589 | SDValue ScaledSelect = |
2590 | DAG.getNode(Opcode: ISD::SELECT, DL: dl, VT: AsIntVT, N1: IsDenormal, N2: ScaledAsInt, N3: AsInt); |
2591 | |
2592 | SDValue ExpMaskScaled = |
2593 | DAG.getNode(Opcode: ISD::AND, DL: dl, VT: AsIntVT, N1: ScaledAsInt, N2: ExpMask); |
2594 | |
2595 | SDValue ScaledValue = |
2596 | DAG.getNode(Opcode: ISD::SELECT, DL: dl, VT: AsIntVT, N1: IsDenormal, N2: ExpMaskScaled, N3: Abs); |
2597 | |
2598 | // Extract the exponent bits. |
2599 | SDValue ExponentShiftAmt = |
2600 | DAG.getShiftAmountConstant(Val: Precision - 1, VT: AsIntVT, DL: dl); |
2601 | SDValue ShiftedExp = |
2602 | DAG.getNode(Opcode: ISD::SRL, DL: dl, VT: AsIntVT, N1: ScaledValue, N2: ExponentShiftAmt); |
2603 | SDValue Exp = DAG.getSExtOrTrunc(Op: ShiftedExp, DL: dl, VT: ExpVT); |
2604 | |
2605 | SDValue NormalBiasedExp = DAG.getNode(Opcode: ISD::ADD, DL: dl, VT: ExpVT, N1: Exp, N2: MinExp); |
2606 | SDValue DenormalOffset = DAG.getConstant(Val: -Precision - 1, DL: dl, VT: ExpVT); |
2607 | SDValue DenormalExpBias = |
2608 | DAG.getNode(Opcode: ISD::SELECT, DL: dl, VT: ExpVT, N1: IsDenormal, N2: DenormalOffset, N3: Zero); |
2609 | |
2610 | SDValue MaskedFractAsInt = |
2611 | DAG.getNode(Opcode: ISD::AND, DL: dl, VT: AsIntVT, N1: ScaledSelect, N2: FractSignMask); |
2612 | const APFloat Half(FltSem, "0.5" ); |
2613 | SDValue FPHalf = DAG.getConstant(Val: Half.bitcastToAPInt(), DL: dl, VT: AsIntVT); |
2614 | SDValue Or = DAG.getNode(Opcode: ISD::OR, DL: dl, VT: AsIntVT, N1: MaskedFractAsInt, N2: FPHalf); |
2615 | SDValue MaskedFract = DAG.getNode(Opcode: ISD::BITCAST, DL: dl, VT, Operand: Or); |
2616 | |
2617 | SDValue ComputedExp = |
2618 | DAG.getNode(Opcode: ISD::ADD, DL: dl, VT: ExpVT, N1: NormalBiasedExp, N2: DenormalExpBias); |
2619 | |
2620 | SDValue Result0 = |
2621 | DAG.getNode(Opcode: ISD::SELECT, DL: dl, VT, N1: DenormOrZero, N2: Val, N3: MaskedFract); |
2622 | |
2623 | SDValue Result1 = |
2624 | DAG.getNode(Opcode: ISD::SELECT, DL: dl, VT: ExpVT, N1: DenormOrZero, N2: Zero, N3: ComputedExp); |
2625 | |
2626 | return DAG.getMergeValues(Ops: {Result0, Result1}, dl); |
2627 | } |
2628 | |
2629 | /// This function is responsible for legalizing a |
2630 | /// INT_TO_FP operation of the specified operand when the target requests that |
2631 | /// we expand it. At this point, we know that the result and operand types are |
2632 | /// legal for the target. |
2633 | SDValue SelectionDAGLegalize::ExpandLegalINT_TO_FP(SDNode *Node, |
2634 | SDValue &Chain) { |
2635 | bool isSigned = (Node->getOpcode() == ISD::STRICT_SINT_TO_FP || |
2636 | Node->getOpcode() == ISD::SINT_TO_FP); |
2637 | EVT DestVT = Node->getValueType(ResNo: 0); |
2638 | SDLoc dl(Node); |
2639 | unsigned OpNo = Node->isStrictFPOpcode() ? 1 : 0; |
2640 | SDValue Op0 = Node->getOperand(Num: OpNo); |
2641 | EVT SrcVT = Op0.getValueType(); |
2642 | |
2643 | // TODO: Should any fast-math-flags be set for the created nodes? |
2644 | LLVM_DEBUG(dbgs() << "Legalizing INT_TO_FP\n" ); |
2645 | if (SrcVT == MVT::i32 && TLI.isTypeLegal(VT: MVT::f64) && |
2646 | (DestVT.bitsLE(VT: MVT::f64) || |
2647 | TLI.isOperationLegal(Op: Node->isStrictFPOpcode() ? ISD::STRICT_FP_EXTEND |
2648 | : ISD::FP_EXTEND, |
2649 | VT: DestVT))) { |
2650 | LLVM_DEBUG(dbgs() << "32-bit [signed|unsigned] integer to float/double " |
2651 | "expansion\n" ); |
2652 | |
2653 | // Get the stack frame index of a 8 byte buffer. |
2654 | SDValue StackSlot = DAG.CreateStackTemporary(VT: MVT::f64); |
2655 | |
2656 | SDValue Lo = Op0; |
2657 | // if signed map to unsigned space |
2658 | if (isSigned) { |
2659 | // Invert sign bit (signed to unsigned mapping). |
2660 | Lo = DAG.getNode(Opcode: ISD::XOR, DL: dl, VT: MVT::i32, N1: Lo, |
2661 | N2: DAG.getConstant(Val: 0x80000000u, DL: dl, VT: MVT::i32)); |
2662 | } |
2663 | // Initial hi portion of constructed double. |
2664 | SDValue Hi = DAG.getConstant(Val: 0x43300000u, DL: dl, VT: MVT::i32); |
2665 | |
2666 | // If this a big endian target, swap the lo and high data. |
2667 | if (DAG.getDataLayout().isBigEndian()) |
2668 | std::swap(a&: Lo, b&: Hi); |
2669 | |
2670 | SDValue MemChain = DAG.getEntryNode(); |
2671 | |
2672 | // Store the lo of the constructed double. |
2673 | SDValue Store1 = DAG.getStore(Chain: MemChain, dl, Val: Lo, Ptr: StackSlot, |
2674 | PtrInfo: MachinePointerInfo()); |
2675 | // Store the hi of the constructed double. |
2676 | SDValue HiPtr = |
2677 | DAG.getMemBasePlusOffset(Base: StackSlot, Offset: TypeSize::getFixed(ExactSize: 4), DL: dl); |
2678 | SDValue Store2 = |
2679 | DAG.getStore(Chain: MemChain, dl, Val: Hi, Ptr: HiPtr, PtrInfo: MachinePointerInfo()); |
2680 | MemChain = DAG.getNode(Opcode: ISD::TokenFactor, DL: dl, VT: MVT::Other, N1: Store1, N2: Store2); |
2681 | |
2682 | // load the constructed double |
2683 | SDValue Load = |
2684 | DAG.getLoad(VT: MVT::f64, dl, Chain: MemChain, Ptr: StackSlot, PtrInfo: MachinePointerInfo()); |
2685 | // FP constant to bias correct the final result |
2686 | SDValue Bias = DAG.getConstantFP( |
2687 | Val: isSigned ? llvm::bit_cast<double>(from: 0x4330000080000000ULL) |
2688 | : llvm::bit_cast<double>(from: 0x4330000000000000ULL), |
2689 | DL: dl, VT: MVT::f64); |
2690 | // Subtract the bias and get the final result. |
2691 | SDValue Sub; |
2692 | SDValue Result; |
2693 | if (Node->isStrictFPOpcode()) { |
2694 | Sub = DAG.getNode(Opcode: ISD::STRICT_FSUB, DL: dl, ResultTys: {MVT::f64, MVT::Other}, |
2695 | Ops: {Node->getOperand(Num: 0), Load, Bias}); |
2696 | Chain = Sub.getValue(R: 1); |
2697 | if (DestVT != Sub.getValueType()) { |
2698 | std::pair<SDValue, SDValue> ResultPair; |
2699 | ResultPair = |
2700 | DAG.getStrictFPExtendOrRound(Op: Sub, Chain, DL: dl, VT: DestVT); |
2701 | Result = ResultPair.first; |
2702 | Chain = ResultPair.second; |
2703 | } |
2704 | else |
2705 | Result = Sub; |
2706 | } else { |
2707 | Sub = DAG.getNode(Opcode: ISD::FSUB, DL: dl, VT: MVT::f64, N1: Load, N2: Bias); |
2708 | Result = DAG.getFPExtendOrRound(Op: Sub, DL: dl, VT: DestVT); |
2709 | } |
2710 | return Result; |
2711 | } |
2712 | |
2713 | if (isSigned) |
2714 | return SDValue(); |
2715 | |
2716 | // TODO: Generalize this for use with other types. |
2717 | if (((SrcVT == MVT::i32 || SrcVT == MVT::i64) && DestVT == MVT::f32) || |
2718 | (SrcVT == MVT::i64 && DestVT == MVT::f64)) { |
2719 | LLVM_DEBUG(dbgs() << "Converting unsigned i32/i64 to f32/f64\n" ); |
2720 | // For unsigned conversions, convert them to signed conversions using the |
2721 | // algorithm from the x86_64 __floatundisf in compiler_rt. That method |
2722 | // should be valid for i32->f32 as well. |
2723 | |
2724 | // More generally this transform should be valid if there are 3 more bits |
2725 | // in the integer type than the significand. Rounding uses the first bit |
2726 | // after the width of the significand and the OR of all bits after that. So |
2727 | // we need to be able to OR the shifted out bit into one of the bits that |
2728 | // participate in the OR. |
2729 | |
2730 | // TODO: This really should be implemented using a branch rather than a |
2731 | // select. We happen to get lucky and machinesink does the right |
2732 | // thing most of the time. This would be a good candidate for a |
2733 | // pseudo-op, or, even better, for whole-function isel. |
2734 | EVT SetCCVT = getSetCCResultType(VT: SrcVT); |
2735 | |
2736 | SDValue SignBitTest = DAG.getSetCC( |
2737 | DL: dl, VT: SetCCVT, LHS: Op0, RHS: DAG.getConstant(Val: 0, DL: dl, VT: SrcVT), Cond: ISD::SETLT); |
2738 | |
2739 | EVT ShiftVT = TLI.getShiftAmountTy(LHSTy: SrcVT, DL: DAG.getDataLayout()); |
2740 | SDValue ShiftConst = DAG.getConstant(Val: 1, DL: dl, VT: ShiftVT); |
2741 | SDValue Shr = DAG.getNode(Opcode: ISD::SRL, DL: dl, VT: SrcVT, N1: Op0, N2: ShiftConst); |
2742 | SDValue AndConst = DAG.getConstant(Val: 1, DL: dl, VT: SrcVT); |
2743 | SDValue And = DAG.getNode(Opcode: ISD::AND, DL: dl, VT: SrcVT, N1: Op0, N2: AndConst); |
2744 | SDValue Or = DAG.getNode(Opcode: ISD::OR, DL: dl, VT: SrcVT, N1: And, N2: Shr); |
2745 | |
2746 | SDValue Slow, Fast; |
2747 | if (Node->isStrictFPOpcode()) { |
2748 | // In strict mode, we must avoid spurious exceptions, and therefore |
2749 | // must make sure to only emit a single STRICT_SINT_TO_FP. |
2750 | SDValue InCvt = DAG.getSelect(DL: dl, VT: SrcVT, Cond: SignBitTest, LHS: Or, RHS: Op0); |
2751 | Fast = DAG.getNode(Opcode: ISD::STRICT_SINT_TO_FP, DL: dl, ResultTys: { DestVT, MVT::Other }, |
2752 | Ops: { Node->getOperand(Num: 0), InCvt }); |
2753 | Slow = DAG.getNode(Opcode: ISD::STRICT_FADD, DL: dl, ResultTys: { DestVT, MVT::Other }, |
2754 | Ops: { Fast.getValue(R: 1), Fast, Fast }); |
2755 | Chain = Slow.getValue(R: 1); |
2756 | // The STRICT_SINT_TO_FP inherits the exception mode from the |
2757 | // incoming STRICT_UINT_TO_FP node; the STRICT_FADD node can |
2758 | // never raise any exception. |
2759 | SDNodeFlags Flags; |
2760 | Flags.setNoFPExcept(Node->getFlags().hasNoFPExcept()); |
2761 | Fast->setFlags(Flags); |
2762 | Flags.setNoFPExcept(true); |
2763 | Slow->setFlags(Flags); |
2764 | } else { |
2765 | SDValue SignCvt = DAG.getNode(Opcode: ISD::SINT_TO_FP, DL: dl, VT: DestVT, Operand: Or); |
2766 | Slow = DAG.getNode(Opcode: ISD::FADD, DL: dl, VT: DestVT, N1: SignCvt, N2: SignCvt); |
2767 | Fast = DAG.getNode(Opcode: ISD::SINT_TO_FP, DL: dl, VT: DestVT, Operand: Op0); |
2768 | } |
2769 | |
2770 | return DAG.getSelect(DL: dl, VT: DestVT, Cond: SignBitTest, LHS: Slow, RHS: Fast); |
2771 | } |
2772 | |
2773 | // Don't expand it if there isn't cheap fadd. |
2774 | if (!TLI.isOperationLegalOrCustom( |
2775 | Op: Node->isStrictFPOpcode() ? ISD::STRICT_FADD : ISD::FADD, VT: DestVT)) |
2776 | return SDValue(); |
2777 | |
2778 | // The following optimization is valid only if every value in SrcVT (when |
2779 | // treated as signed) is representable in DestVT. Check that the mantissa |
2780 | // size of DestVT is >= than the number of bits in SrcVT -1. |
2781 | assert(APFloat::semanticsPrecision(DestVT.getFltSemantics()) >= |
2782 | SrcVT.getSizeInBits() - 1 && |
2783 | "Cannot perform lossless SINT_TO_FP!" ); |
2784 | |
2785 | SDValue Tmp1; |
2786 | if (Node->isStrictFPOpcode()) { |
2787 | Tmp1 = DAG.getNode(Opcode: ISD::STRICT_SINT_TO_FP, DL: dl, ResultTys: { DestVT, MVT::Other }, |
2788 | Ops: { Node->getOperand(Num: 0), Op0 }); |
2789 | } else |
2790 | Tmp1 = DAG.getNode(Opcode: ISD::SINT_TO_FP, DL: dl, VT: DestVT, Operand: Op0); |
2791 | |
2792 | SDValue SignSet = DAG.getSetCC(DL: dl, VT: getSetCCResultType(VT: SrcVT), LHS: Op0, |
2793 | RHS: DAG.getConstant(Val: 0, DL: dl, VT: SrcVT), Cond: ISD::SETLT); |
2794 | SDValue Zero = DAG.getIntPtrConstant(Val: 0, DL: dl), |
2795 | Four = DAG.getIntPtrConstant(Val: 4, DL: dl); |
2796 | SDValue CstOffset = DAG.getSelect(DL: dl, VT: Zero.getValueType(), |
2797 | Cond: SignSet, LHS: Four, RHS: Zero); |
2798 | |
2799 | // If the sign bit of the integer is set, the large number will be treated |
2800 | // as a negative number. To counteract this, the dynamic code adds an |
2801 | // offset depending on the data type. |
2802 | uint64_t FF; |
2803 | switch (SrcVT.getSimpleVT().SimpleTy) { |
2804 | default: |
2805 | return SDValue(); |
2806 | case MVT::i8 : FF = 0x43800000ULL; break; // 2^8 (as a float) |
2807 | case MVT::i16: FF = 0x47800000ULL; break; // 2^16 (as a float) |
2808 | case MVT::i32: FF = 0x4F800000ULL; break; // 2^32 (as a float) |
2809 | case MVT::i64: FF = 0x5F800000ULL; break; // 2^64 (as a float) |
2810 | } |
2811 | if (DAG.getDataLayout().isLittleEndian()) |
2812 | FF <<= 32; |
2813 | Constant *FudgeFactor = ConstantInt::get( |
2814 | Ty: Type::getInt64Ty(C&: *DAG.getContext()), V: FF); |
2815 | |
2816 | SDValue CPIdx = |
2817 | DAG.getConstantPool(C: FudgeFactor, VT: TLI.getPointerTy(DL: DAG.getDataLayout())); |
2818 | Align Alignment = cast<ConstantPoolSDNode>(Val&: CPIdx)->getAlign(); |
2819 | CPIdx = DAG.getNode(Opcode: ISD::ADD, DL: dl, VT: CPIdx.getValueType(), N1: CPIdx, N2: CstOffset); |
2820 | Alignment = commonAlignment(A: Alignment, Offset: 4); |
2821 | SDValue FudgeInReg; |
2822 | if (DestVT == MVT::f32) |
2823 | FudgeInReg = DAG.getLoad( |
2824 | VT: MVT::f32, dl, Chain: DAG.getEntryNode(), Ptr: CPIdx, |
2825 | PtrInfo: MachinePointerInfo::getConstantPool(MF&: DAG.getMachineFunction()), |
2826 | Alignment); |
2827 | else { |
2828 | SDValue Load = DAG.getExtLoad( |
2829 | ExtType: ISD::EXTLOAD, dl, VT: DestVT, Chain: DAG.getEntryNode(), Ptr: CPIdx, |
2830 | PtrInfo: MachinePointerInfo::getConstantPool(MF&: DAG.getMachineFunction()), MemVT: MVT::f32, |
2831 | Alignment); |
2832 | HandleSDNode Handle(Load); |
2833 | LegalizeOp(Node: Load.getNode()); |
2834 | FudgeInReg = Handle.getValue(); |
2835 | } |
2836 | |
2837 | if (Node->isStrictFPOpcode()) { |
2838 | SDValue Result = DAG.getNode(Opcode: ISD::STRICT_FADD, DL: dl, ResultTys: { DestVT, MVT::Other }, |
2839 | Ops: { Tmp1.getValue(R: 1), Tmp1, FudgeInReg }); |
2840 | Chain = Result.getValue(R: 1); |
2841 | return Result; |
2842 | } |
2843 | |
2844 | return DAG.getNode(Opcode: ISD::FADD, DL: dl, VT: DestVT, N1: Tmp1, N2: FudgeInReg); |
2845 | } |
2846 | |
2847 | /// This function is responsible for legalizing a |
2848 | /// *INT_TO_FP operation of the specified operand when the target requests that |
2849 | /// we promote it. At this point, we know that the result and operand types are |
2850 | /// legal for the target, and that there is a legal UINT_TO_FP or SINT_TO_FP |
2851 | /// operation that takes a larger input. |
2852 | void SelectionDAGLegalize::PromoteLegalINT_TO_FP( |
2853 | SDNode *N, const SDLoc &dl, SmallVectorImpl<SDValue> &Results) { |
2854 | bool IsStrict = N->isStrictFPOpcode(); |
2855 | bool IsSigned = N->getOpcode() == ISD::SINT_TO_FP || |
2856 | N->getOpcode() == ISD::STRICT_SINT_TO_FP; |
2857 | EVT DestVT = N->getValueType(ResNo: 0); |
2858 | SDValue LegalOp = N->getOperand(Num: IsStrict ? 1 : 0); |
2859 | unsigned UIntOp = IsStrict ? ISD::STRICT_UINT_TO_FP : ISD::UINT_TO_FP; |
2860 | unsigned SIntOp = IsStrict ? ISD::STRICT_SINT_TO_FP : ISD::SINT_TO_FP; |
2861 | |
2862 | // First step, figure out the appropriate *INT_TO_FP operation to use. |
2863 | EVT NewInTy = LegalOp.getValueType(); |
2864 | |
2865 | unsigned OpToUse = 0; |
2866 | |
2867 | // Scan for the appropriate larger type to use. |
2868 | while (true) { |
2869 | NewInTy = (MVT::SimpleValueType)(NewInTy.getSimpleVT().SimpleTy+1); |
2870 | assert(NewInTy.isInteger() && "Ran out of possibilities!" ); |
2871 | |
2872 | // If the target supports SINT_TO_FP of this type, use it. |
2873 | if (TLI.isOperationLegalOrCustom(Op: SIntOp, VT: NewInTy)) { |
2874 | OpToUse = SIntOp; |
2875 | break; |
2876 | } |
2877 | if (IsSigned) |
2878 | continue; |
2879 | |
2880 | // If the target supports UINT_TO_FP of this type, use it. |
2881 | if (TLI.isOperationLegalOrCustom(Op: UIntOp, VT: NewInTy)) { |
2882 | OpToUse = UIntOp; |
2883 | break; |
2884 | } |
2885 | |
2886 | // Otherwise, try a larger type. |
2887 | } |
2888 | |
2889 | // Okay, we found the operation and type to use. Zero extend our input to the |
2890 | // desired type then run the operation on it. |
2891 | if (IsStrict) { |
2892 | SDValue Res = |
2893 | DAG.getNode(Opcode: OpToUse, DL: dl, ResultTys: {DestVT, MVT::Other}, |
2894 | Ops: {N->getOperand(Num: 0), |
2895 | DAG.getNode(Opcode: IsSigned ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND, |
2896 | DL: dl, VT: NewInTy, Operand: LegalOp)}); |
2897 | Results.push_back(Elt: Res); |
2898 | Results.push_back(Elt: Res.getValue(R: 1)); |
2899 | return; |
2900 | } |
2901 | |
2902 | Results.push_back( |
2903 | Elt: DAG.getNode(Opcode: OpToUse, DL: dl, VT: DestVT, |
2904 | Operand: DAG.getNode(Opcode: IsSigned ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND, |
2905 | DL: dl, VT: NewInTy, Operand: LegalOp))); |
2906 | } |
2907 | |
2908 | /// This function is responsible for legalizing a |
2909 | /// FP_TO_*INT operation of the specified operand when the target requests that |
2910 | /// we promote it. At this point, we know that the result and operand types are |
2911 | /// legal for the target, and that there is a legal FP_TO_UINT or FP_TO_SINT |
2912 | /// operation that returns a larger result. |
2913 | void SelectionDAGLegalize::PromoteLegalFP_TO_INT(SDNode *N, const SDLoc &dl, |
2914 | SmallVectorImpl<SDValue> &Results) { |
2915 | bool IsStrict = N->isStrictFPOpcode(); |
2916 | bool IsSigned = N->getOpcode() == ISD::FP_TO_SINT || |
2917 | N->getOpcode() == ISD::STRICT_FP_TO_SINT; |
2918 | EVT DestVT = N->getValueType(ResNo: 0); |
2919 | SDValue LegalOp = N->getOperand(Num: IsStrict ? 1 : 0); |
2920 | // First step, figure out the appropriate FP_TO*INT operation to use. |
2921 | EVT NewOutTy = DestVT; |
2922 | |
2923 | unsigned OpToUse = 0; |
2924 | |
2925 | // Scan for the appropriate larger type to use. |
2926 | while (true) { |
2927 | NewOutTy = (MVT::SimpleValueType)(NewOutTy.getSimpleVT().SimpleTy+1); |
2928 | assert(NewOutTy.isInteger() && "Ran out of possibilities!" ); |
2929 | |
2930 | // A larger signed type can hold all unsigned values of the requested type, |
2931 | // so using FP_TO_SINT is valid |
2932 | OpToUse = IsStrict ? ISD::STRICT_FP_TO_SINT : ISD::FP_TO_SINT; |
2933 | if (TLI.isOperationLegalOrCustom(Op: OpToUse, VT: NewOutTy)) |
2934 | break; |
2935 | |
2936 | // However, if the value may be < 0.0, we *must* use some FP_TO_SINT. |
2937 | OpToUse = IsStrict ? ISD::STRICT_FP_TO_UINT : ISD::FP_TO_UINT; |
2938 | if (!IsSigned && TLI.isOperationLegalOrCustom(Op: OpToUse, VT: NewOutTy)) |
2939 | break; |
2940 | |
2941 | // Otherwise, try a larger type. |
2942 | } |
2943 | |
2944 | // Okay, we found the operation and type to use. |
2945 | SDValue Operation; |
2946 | if (IsStrict) { |
2947 | SDVTList VTs = DAG.getVTList(VT1: NewOutTy, VT2: MVT::Other); |
2948 | Operation = DAG.getNode(Opcode: OpToUse, DL: dl, VTList: VTs, N1: N->getOperand(Num: 0), N2: LegalOp); |
2949 | } else |
2950 | Operation = DAG.getNode(Opcode: OpToUse, DL: dl, VT: NewOutTy, Operand: LegalOp); |
2951 | |
2952 | // Truncate the result of the extended FP_TO_*INT operation to the desired |
2953 | // size. |
2954 | SDValue Trunc = DAG.getNode(Opcode: ISD::TRUNCATE, DL: dl, VT: DestVT, Operand: Operation); |
2955 | Results.push_back(Elt: Trunc); |
2956 | if (IsStrict) |
2957 | Results.push_back(Elt: Operation.getValue(R: 1)); |
2958 | } |
2959 | |
2960 | /// Promote FP_TO_*INT_SAT operation to a larger result type. At this point |
2961 | /// the result and operand types are legal and there must be a legal |
2962 | /// FP_TO_*INT_SAT operation for a larger result type. |
2963 | SDValue SelectionDAGLegalize::PromoteLegalFP_TO_INT_SAT(SDNode *Node, |
2964 | const SDLoc &dl) { |
2965 | unsigned Opcode = Node->getOpcode(); |
2966 | |
2967 | // Scan for the appropriate larger type to use. |
2968 | EVT NewOutTy = Node->getValueType(ResNo: 0); |
2969 | while (true) { |
2970 | NewOutTy = (MVT::SimpleValueType)(NewOutTy.getSimpleVT().SimpleTy + 1); |
2971 | assert(NewOutTy.isInteger() && "Ran out of possibilities!" ); |
2972 | |
2973 | if (TLI.isOperationLegalOrCustom(Op: Opcode, VT: NewOutTy)) |
2974 | break; |
2975 | } |
2976 | |
2977 | // Saturation width is determined by second operand, so we don't have to |
2978 | // perform any fixup and can directly truncate the result. |
2979 | SDValue Result = DAG.getNode(Opcode, DL: dl, VT: NewOutTy, N1: Node->getOperand(Num: 0), |
2980 | N2: Node->getOperand(Num: 1)); |
2981 | return DAG.getNode(Opcode: ISD::TRUNCATE, DL: dl, VT: Node->getValueType(ResNo: 0), Operand: Result); |
2982 | } |
2983 | |
2984 | /// Open code the operations for PARITY of the specified operation. |
2985 | SDValue SelectionDAGLegalize::ExpandPARITY(SDValue Op, const SDLoc &dl) { |
2986 | EVT VT = Op.getValueType(); |
2987 | EVT ShVT = TLI.getShiftAmountTy(LHSTy: VT, DL: DAG.getDataLayout()); |
2988 | unsigned Sz = VT.getScalarSizeInBits(); |
2989 | |
2990 | // If CTPOP is legal, use it. Otherwise use shifts and xor. |
2991 | SDValue Result; |
2992 | if (TLI.isOperationLegalOrPromote(Op: ISD::CTPOP, VT)) { |
2993 | Result = DAG.getNode(Opcode: ISD::CTPOP, DL: dl, VT, Operand: Op); |
2994 | } else { |
2995 | Result = Op; |
2996 | for (unsigned i = Log2_32_Ceil(Value: Sz); i != 0;) { |
2997 | SDValue Shift = DAG.getNode(Opcode: ISD::SRL, DL: dl, VT, N1: Result, |
2998 | N2: DAG.getConstant(Val: 1ULL << (--i), DL: dl, VT: ShVT)); |
2999 | Result = DAG.getNode(Opcode: ISD::XOR, DL: dl, VT, N1: Result, N2: Shift); |
3000 | } |
3001 | } |
3002 | |
3003 | return DAG.getNode(Opcode: ISD::AND, DL: dl, VT, N1: Result, N2: DAG.getConstant(Val: 1, DL: dl, VT)); |
3004 | } |
3005 | |
3006 | SDValue SelectionDAGLegalize::PromoteReduction(SDNode *Node) { |
3007 | bool IsVPOpcode = ISD::isVPOpcode(Opcode: Node->getOpcode()); |
3008 | MVT VecVT = IsVPOpcode ? Node->getOperand(Num: 1).getSimpleValueType() |
3009 | : Node->getOperand(Num: 0).getSimpleValueType(); |
3010 | MVT NewVecVT = TLI.getTypeToPromoteTo(Op: Node->getOpcode(), VT: VecVT); |
3011 | MVT ScalarVT = Node->getSimpleValueType(ResNo: 0); |
3012 | MVT NewScalarVT = NewVecVT.getVectorElementType(); |
3013 | |
3014 | SDLoc DL(Node); |
3015 | SmallVector<SDValue, 4> Operands(Node->getNumOperands()); |
3016 | |
3017 | // FIXME: Support integer. |
3018 | assert(Node->getOperand(0).getValueType().isFloatingPoint() && |
3019 | "Only FP promotion is supported" ); |
3020 | |
3021 | for (unsigned j = 0; j != Node->getNumOperands(); ++j) |
3022 | if (Node->getOperand(Num: j).getValueType().isVector() && |
3023 | !(IsVPOpcode && |
3024 | ISD::getVPMaskIdx(Opcode: Node->getOpcode()) == j)) { // Skip mask operand. |
3025 | // promote the vector operand. |
3026 | // FIXME: Support integer. |
3027 | assert(Node->getOperand(j).getValueType().isFloatingPoint() && |
3028 | "Only FP promotion is supported" ); |
3029 | Operands[j] = |
3030 | DAG.getNode(Opcode: ISD::FP_EXTEND, DL, VT: NewVecVT, Operand: Node->getOperand(Num: j)); |
3031 | } else if (Node->getOperand(Num: j).getValueType().isFloatingPoint()) { |
3032 | // promote the initial value. |
3033 | Operands[j] = |
3034 | DAG.getNode(Opcode: ISD::FP_EXTEND, DL, VT: NewScalarVT, Operand: Node->getOperand(Num: j)); |
3035 | } else { |
3036 | Operands[j] = Node->getOperand(Num: j); // Skip VL operand. |
3037 | } |
3038 | |
3039 | SDValue Res = DAG.getNode(Opcode: Node->getOpcode(), DL, VT: NewScalarVT, Ops: Operands, |
3040 | Flags: Node->getFlags()); |
3041 | |
3042 | assert(ScalarVT.isFloatingPoint() && "Only FP promotion is supported" ); |
3043 | return DAG.getNode(Opcode: ISD::FP_ROUND, DL, VT: ScalarVT, N1: Res, |
3044 | N2: DAG.getIntPtrConstant(Val: 0, DL, /*isTarget=*/true)); |
3045 | } |
3046 | |
3047 | bool SelectionDAGLegalize::ExpandNode(SDNode *Node) { |
3048 | LLVM_DEBUG(dbgs() << "Trying to expand node\n" ); |
3049 | SmallVector<SDValue, 8> Results; |
3050 | SDLoc dl(Node); |
3051 | SDValue Tmp1, Tmp2, Tmp3, Tmp4; |
3052 | bool NeedInvert; |
3053 | switch (Node->getOpcode()) { |
3054 | case ISD::ABS: |
3055 | if ((Tmp1 = TLI.expandABS(N: Node, DAG))) |
3056 | Results.push_back(Elt: Tmp1); |
3057 | break; |
3058 | case ISD::ABDS: |
3059 | case ISD::ABDU: |
3060 | if ((Tmp1 = TLI.expandABD(N: Node, DAG))) |
3061 | Results.push_back(Elt: Tmp1); |
3062 | break; |
3063 | case ISD::AVGCEILS: |
3064 | case ISD::AVGCEILU: |
3065 | case ISD::AVGFLOORS: |
3066 | case ISD::AVGFLOORU: |
3067 | if ((Tmp1 = TLI.expandAVG(N: Node, DAG))) |
3068 | Results.push_back(Elt: Tmp1); |
3069 | break; |
3070 | case ISD::CTPOP: |
3071 | if ((Tmp1 = TLI.expandCTPOP(N: Node, DAG))) |
3072 | Results.push_back(Elt: Tmp1); |
3073 | break; |
3074 | case ISD::CTLZ: |
3075 | case ISD::CTLZ_ZERO_UNDEF: |
3076 | if ((Tmp1 = TLI.expandCTLZ(N: Node, DAG))) |
3077 | Results.push_back(Elt: Tmp1); |
3078 | break; |
3079 | case ISD::CTTZ: |
3080 | case ISD::CTTZ_ZERO_UNDEF: |
3081 | if ((Tmp1 = TLI.expandCTTZ(N: Node, DAG))) |
3082 | Results.push_back(Elt: Tmp1); |
3083 | break; |
3084 | case ISD::BITREVERSE: |
3085 | if ((Tmp1 = TLI.expandBITREVERSE(N: Node, DAG))) |
3086 | Results.push_back(Elt: Tmp1); |
3087 | break; |
3088 | case ISD::BSWAP: |
3089 | if ((Tmp1 = TLI.expandBSWAP(N: Node, DAG))) |
3090 | Results.push_back(Elt: Tmp1); |
3091 | break; |
3092 | case ISD::PARITY: |
3093 | Results.push_back(Elt: ExpandPARITY(Op: Node->getOperand(Num: 0), dl)); |
3094 | break; |
3095 | case ISD::FRAMEADDR: |
3096 | case ISD::RETURNADDR: |
3097 | case ISD::FRAME_TO_ARGS_OFFSET: |
3098 | Results.push_back(Elt: DAG.getConstant(Val: 0, DL: dl, VT: Node->getValueType(ResNo: 0))); |
3099 | break; |
3100 | case ISD::EH_DWARF_CFA: { |
3101 | SDValue CfaArg = DAG.getSExtOrTrunc(Op: Node->getOperand(Num: 0), DL: dl, |
3102 | VT: TLI.getPointerTy(DL: DAG.getDataLayout())); |
3103 | SDValue Offset = DAG.getNode(Opcode: ISD::ADD, DL: dl, |
3104 | VT: CfaArg.getValueType(), |
3105 | N1: DAG.getNode(Opcode: ISD::FRAME_TO_ARGS_OFFSET, DL: dl, |
3106 | VT: CfaArg.getValueType()), |
3107 | N2: CfaArg); |
3108 | SDValue FA = DAG.getNode( |
3109 | Opcode: ISD::FRAMEADDR, DL: dl, VT: TLI.getPointerTy(DL: DAG.getDataLayout()), |
3110 | Operand: DAG.getConstant(Val: 0, DL: dl, VT: TLI.getPointerTy(DL: DAG.getDataLayout()))); |
3111 | Results.push_back(Elt: DAG.getNode(Opcode: ISD::ADD, DL: dl, VT: FA.getValueType(), |
3112 | N1: FA, N2: Offset)); |
3113 | break; |
3114 | } |
3115 | case ISD::GET_ROUNDING: |
3116 | Results.push_back(Elt: DAG.getConstant(Val: 1, DL: dl, VT: Node->getValueType(ResNo: 0))); |
3117 | Results.push_back(Elt: Node->getOperand(Num: 0)); |
3118 | break; |
3119 | case ISD::EH_RETURN: |
3120 | case ISD::EH_LABEL: |
3121 | case ISD::PREFETCH: |
3122 | case ISD::VAEND: |
3123 | case ISD::EH_SJLJ_LONGJMP: |
3124 | // If the target didn't expand these, there's nothing to do, so just |
3125 | // preserve the chain and be done. |
3126 | Results.push_back(Elt: Node->getOperand(Num: 0)); |
3127 | break; |
3128 | case ISD::READCYCLECOUNTER: |
3129 | case ISD::READSTEADYCOUNTER: |
3130 | // If the target didn't expand this, just return 'zero' and preserve the |
3131 | // chain. |
3132 | Results.append(NumInputs: Node->getNumValues() - 1, |
3133 | Elt: DAG.getConstant(Val: 0, DL: dl, VT: Node->getValueType(ResNo: 0))); |
3134 | Results.push_back(Elt: Node->getOperand(Num: 0)); |
3135 | break; |
3136 | case ISD::EH_SJLJ_SETJMP: |
3137 | // If the target didn't expand this, just return 'zero' and preserve the |
3138 | // chain. |
3139 | Results.push_back(Elt: DAG.getConstant(Val: 0, DL: dl, VT: MVT::i32)); |
3140 | Results.push_back(Elt: Node->getOperand(Num: 0)); |
3141 | break; |
3142 | case ISD::ATOMIC_LOAD: { |
3143 | // There is no libcall for atomic load; fake it with ATOMIC_CMP_SWAP. |
3144 | SDValue Zero = DAG.getConstant(Val: 0, DL: dl, VT: Node->getValueType(ResNo: 0)); |
3145 | SDVTList VTs = DAG.getVTList(VT1: Node->getValueType(ResNo: 0), VT2: MVT::Other); |
3146 | SDValue Swap = DAG.getAtomicCmpSwap( |
3147 | Opcode: ISD::ATOMIC_CMP_SWAP, dl, MemVT: cast<AtomicSDNode>(Val: Node)->getMemoryVT(), VTs, |
3148 | Chain: Node->getOperand(Num: 0), Ptr: Node->getOperand(Num: 1), Cmp: Zero, Swp: Zero, |
3149 | MMO: cast<AtomicSDNode>(Val: Node)->getMemOperand()); |
3150 | Results.push_back(Elt: Swap.getValue(R: 0)); |
3151 | Results.push_back(Elt: Swap.getValue(R: 1)); |
3152 | break; |
3153 | } |
3154 | case ISD::ATOMIC_STORE: { |
3155 | // There is no libcall for atomic store; fake it with ATOMIC_SWAP. |
3156 | SDValue Swap = DAG.getAtomic( |
3157 | Opcode: ISD::ATOMIC_SWAP, dl, MemVT: cast<AtomicSDNode>(Val: Node)->getMemoryVT(), |
3158 | Chain: Node->getOperand(Num: 0), Ptr: Node->getOperand(Num: 2), Val: Node->getOperand(Num: 1), |
3159 | MMO: cast<AtomicSDNode>(Val: Node)->getMemOperand()); |
3160 | Results.push_back(Elt: Swap.getValue(R: 1)); |
3161 | break; |
3162 | } |
3163 | case ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS: { |
3164 | // Expanding an ATOMIC_CMP_SWAP_WITH_SUCCESS produces an ATOMIC_CMP_SWAP and |
3165 | // splits out the success value as a comparison. Expanding the resulting |
3166 | // ATOMIC_CMP_SWAP will produce a libcall. |
3167 | SDVTList VTs = DAG.getVTList(VT1: Node->getValueType(ResNo: 0), VT2: MVT::Other); |
3168 | SDValue Res = DAG.getAtomicCmpSwap( |
3169 | Opcode: ISD::ATOMIC_CMP_SWAP, dl, MemVT: cast<AtomicSDNode>(Val: Node)->getMemoryVT(), VTs, |
3170 | Chain: Node->getOperand(Num: 0), Ptr: Node->getOperand(Num: 1), Cmp: Node->getOperand(Num: 2), |
3171 | Swp: Node->getOperand(Num: 3), MMO: cast<MemSDNode>(Val: Node)->getMemOperand()); |
3172 | |
3173 | SDValue ExtRes = Res; |
3174 | SDValue LHS = Res; |
3175 | SDValue RHS = Node->getOperand(Num: 1); |
3176 | |
3177 | EVT AtomicType = cast<AtomicSDNode>(Val: Node)->getMemoryVT(); |
3178 | EVT OuterType = Node->getValueType(ResNo: 0); |
3179 | switch (TLI.getExtendForAtomicOps()) { |
3180 | case ISD::SIGN_EXTEND: |
3181 | LHS = DAG.getNode(Opcode: ISD::AssertSext, DL: dl, VT: OuterType, N1: Res, |
3182 | N2: DAG.getValueType(AtomicType)); |
3183 | RHS = DAG.getNode(Opcode: ISD::SIGN_EXTEND_INREG, DL: dl, VT: OuterType, |
3184 | N1: Node->getOperand(Num: 2), N2: DAG.getValueType(AtomicType)); |
3185 | ExtRes = LHS; |
3186 | break; |
3187 | case ISD::ZERO_EXTEND: |
3188 | LHS = DAG.getNode(Opcode: ISD::AssertZext, DL: dl, VT: OuterType, N1: Res, |
3189 | N2: DAG.getValueType(AtomicType)); |
3190 | RHS = DAG.getZeroExtendInReg(Op: Node->getOperand(Num: 2), DL: dl, VT: AtomicType); |
3191 | ExtRes = LHS; |
3192 | break; |
3193 | case ISD::ANY_EXTEND: |
3194 | LHS = DAG.getZeroExtendInReg(Op: Res, DL: dl, VT: AtomicType); |
3195 | RHS = DAG.getZeroExtendInReg(Op: Node->getOperand(Num: 2), DL: dl, VT: AtomicType); |
3196 | break; |
3197 | default: |
3198 | llvm_unreachable("Invalid atomic op extension" ); |
3199 | } |
3200 | |
3201 | SDValue Success = |
3202 | DAG.getSetCC(DL: dl, VT: Node->getValueType(ResNo: 1), LHS, RHS, Cond: ISD::SETEQ); |
3203 | |
3204 | Results.push_back(Elt: ExtRes.getValue(R: 0)); |
3205 | Results.push_back(Elt: Success); |
3206 | Results.push_back(Elt: Res.getValue(R: 1)); |
3207 | break; |
3208 | } |
3209 | case ISD::ATOMIC_LOAD_SUB: { |
3210 | SDLoc DL(Node); |
3211 | EVT VT = Node->getValueType(ResNo: 0); |
3212 | SDValue RHS = Node->getOperand(Num: 2); |
3213 | AtomicSDNode *AN = cast<AtomicSDNode>(Val: Node); |
3214 | if (RHS->getOpcode() == ISD::SIGN_EXTEND_INREG && |
3215 | cast<VTSDNode>(Val: RHS->getOperand(Num: 1))->getVT() == AN->getMemoryVT()) |
3216 | RHS = RHS->getOperand(Num: 0); |
3217 | SDValue NewRHS = |
3218 | DAG.getNode(Opcode: ISD::SUB, DL, VT, N1: DAG.getConstant(Val: 0, DL, VT), N2: RHS); |
3219 | SDValue Res = DAG.getAtomic(Opcode: ISD::ATOMIC_LOAD_ADD, dl: DL, MemVT: AN->getMemoryVT(), |
3220 | Chain: Node->getOperand(Num: 0), Ptr: Node->getOperand(Num: 1), |
3221 | Val: NewRHS, MMO: AN->getMemOperand()); |
3222 | Results.push_back(Elt: Res); |
3223 | Results.push_back(Elt: Res.getValue(R: 1)); |
3224 | break; |
3225 | } |
3226 | case ISD::DYNAMIC_STACKALLOC: |
3227 | ExpandDYNAMIC_STACKALLOC(Node, Results); |
3228 | break; |
3229 | case ISD::MERGE_VALUES: |
3230 | for (unsigned i = 0; i < Node->getNumValues(); i++) |
3231 | Results.push_back(Elt: Node->getOperand(Num: i)); |
3232 | break; |
3233 | case ISD::POISON: |
3234 | case ISD::UNDEF: { |
3235 | EVT VT = Node->getValueType(ResNo: 0); |
3236 | if (VT.isInteger()) |
3237 | Results.push_back(Elt: DAG.getConstant(Val: 0, DL: dl, VT)); |
3238 | else { |
3239 | assert(VT.isFloatingPoint() && "Unknown value type!" ); |
3240 | Results.push_back(Elt: DAG.getConstantFP(Val: 0, DL: dl, VT)); |
3241 | } |
3242 | break; |
3243 | } |
3244 | case ISD::STRICT_FP_ROUND: |
3245 | // When strict mode is enforced we can't do expansion because it |
3246 | // does not honor the "strict" properties. Only libcall is allowed. |
3247 | if (TLI.isStrictFPEnabled()) |
3248 | break; |
3249 | // We might as well mutate to FP_ROUND when FP_ROUND operation is legal |
3250 | // since this operation is more efficient than stack operation. |
3251 | if (TLI.getStrictFPOperationAction(Op: Node->getOpcode(), |
3252 | VT: Node->getValueType(ResNo: 0)) |
3253 | == TargetLowering::Legal) |
3254 | break; |
3255 | // We fall back to use stack operation when the FP_ROUND operation |
3256 | // isn't available. |
3257 | if ((Tmp1 = EmitStackConvert(SrcOp: Node->getOperand(Num: 1), SlotVT: Node->getValueType(ResNo: 0), |
3258 | DestVT: Node->getValueType(ResNo: 0), dl, |
3259 | Chain: Node->getOperand(Num: 0)))) { |
3260 | ReplaceNode(Old: Node, New: Tmp1.getNode()); |
3261 | LLVM_DEBUG(dbgs() << "Successfully expanded STRICT_FP_ROUND node\n" ); |
3262 | return true; |
3263 | } |
3264 | break; |
3265 | case ISD::FP_ROUND: { |
3266 | if ((Tmp1 = TLI.expandFP_ROUND(Node, DAG))) { |
3267 | Results.push_back(Elt: Tmp1); |
3268 | break; |
3269 | } |
3270 | |
3271 | [[fallthrough]]; |
3272 | } |
3273 | case ISD::BITCAST: |
3274 | if ((Tmp1 = EmitStackConvert(SrcOp: Node->getOperand(Num: 0), SlotVT: Node->getValueType(ResNo: 0), |
3275 | DestVT: Node->getValueType(ResNo: 0), dl))) |
3276 | Results.push_back(Elt: Tmp1); |
3277 | break; |
3278 | case ISD::STRICT_FP_EXTEND: |
3279 | // When strict mode is enforced we can't do expansion because it |
3280 | // does not honor the "strict" properties. Only libcall is allowed. |
3281 | if (TLI.isStrictFPEnabled()) |
3282 | break; |
3283 | // We might as well mutate to FP_EXTEND when FP_EXTEND operation is legal |
3284 | // since this operation is more efficient than stack operation. |
3285 | if (TLI.getStrictFPOperationAction(Op: Node->getOpcode(), |
3286 | VT: Node->getValueType(ResNo: 0)) |
3287 | == TargetLowering::Legal) |
3288 | break; |
3289 | // We fall back to use stack operation when the FP_EXTEND operation |
3290 | // isn't available. |
3291 | if ((Tmp1 = EmitStackConvert( |
3292 | SrcOp: Node->getOperand(Num: 1), SlotVT: Node->getOperand(Num: 1).getValueType(), |
3293 | DestVT: Node->getValueType(ResNo: 0), dl, Chain: Node->getOperand(Num: 0)))) { |
3294 | ReplaceNode(Old: Node, New: Tmp1.getNode()); |
3295 | LLVM_DEBUG(dbgs() << "Successfully expanded STRICT_FP_EXTEND node\n" ); |
3296 | return true; |
3297 | } |
3298 | break; |
3299 | case ISD::FP_EXTEND: { |
3300 | SDValue Op = Node->getOperand(Num: 0); |
3301 | EVT SrcVT = Op.getValueType(); |
3302 | EVT DstVT = Node->getValueType(ResNo: 0); |
3303 | if (SrcVT.getScalarType() == MVT::bf16) { |
3304 | Results.push_back(Elt: DAG.getNode(Opcode: ISD::BF16_TO_FP, DL: SDLoc(Node), VT: DstVT, Operand: Op)); |
3305 | break; |
3306 | } |
3307 | |
3308 | if ((Tmp1 = EmitStackConvert(SrcOp: Op, SlotVT: SrcVT, DestVT: DstVT, dl))) |
3309 | Results.push_back(Elt: Tmp1); |
3310 | break; |
3311 | } |
3312 | case ISD::BF16_TO_FP: { |
3313 | // Always expand bf16 to f32 casts, they lower to ext + shift. |
3314 | // |
3315 | // Note that the operand of this code can be bf16 or an integer type in case |
3316 | // bf16 is not supported on the target and was softened. |
3317 | SDValue Op = Node->getOperand(Num: 0); |
3318 | if (Op.getValueType() == MVT::bf16) { |
3319 | Op = DAG.getNode(Opcode: ISD::ANY_EXTEND, DL: dl, VT: MVT::i32, |
3320 | Operand: DAG.getNode(Opcode: ISD::BITCAST, DL: dl, VT: MVT::i16, Operand: Op)); |
3321 | } else { |
3322 | Op = DAG.getAnyExtOrTrunc(Op, DL: dl, VT: MVT::i32); |
3323 | } |
3324 | Op = DAG.getNode( |
3325 | Opcode: ISD::SHL, DL: dl, VT: MVT::i32, N1: Op, |
3326 | N2: DAG.getConstant(Val: 16, DL: dl, |
3327 | VT: TLI.getShiftAmountTy(LHSTy: MVT::i32, DL: DAG.getDataLayout()))); |
3328 | Op = DAG.getNode(Opcode: ISD::BITCAST, DL: dl, VT: MVT::f32, Operand: Op); |
3329 | // Add fp_extend in case the output is bigger than f32. |
3330 | if (Node->getValueType(ResNo: 0) != MVT::f32) |
3331 | Op = DAG.getNode(Opcode: ISD::FP_EXTEND, DL: dl, VT: Node->getValueType(ResNo: 0), Operand: Op); |
3332 | Results.push_back(Elt: Op); |
3333 | break; |
3334 | } |
3335 | case ISD::FP_TO_BF16: { |
3336 | SDValue Op = Node->getOperand(Num: 0); |
3337 | if (Op.getValueType() != MVT::f32) |
3338 | Op = DAG.getNode(Opcode: ISD::FP_ROUND, DL: dl, VT: MVT::f32, N1: Op, |
3339 | N2: DAG.getIntPtrConstant(Val: 0, DL: dl, /*isTarget=*/true)); |
3340 | // Certain SNaNs will turn into infinities if we do a simple shift right. |
3341 | if (!DAG.isKnownNeverSNaN(Op)) { |
3342 | Op = DAG.getNode(Opcode: ISD::FCANONICALIZE, DL: dl, VT: MVT::f32, Operand: Op, Flags: Node->getFlags()); |
3343 | } |
3344 | Op = DAG.getNode( |
3345 | Opcode: ISD::SRL, DL: dl, VT: MVT::i32, N1: DAG.getNode(Opcode: ISD::BITCAST, DL: dl, VT: MVT::i32, Operand: Op), |
3346 | N2: DAG.getConstant(Val: 16, DL: dl, |
3347 | VT: TLI.getShiftAmountTy(LHSTy: MVT::i32, DL: DAG.getDataLayout()))); |
3348 | // The result of this node can be bf16 or an integer type in case bf16 is |
3349 | // not supported on the target and was softened to i16 for storage. |
3350 | if (Node->getValueType(ResNo: 0) == MVT::bf16) { |
3351 | Op = DAG.getNode(Opcode: ISD::BITCAST, DL: dl, VT: MVT::bf16, |
3352 | Operand: DAG.getNode(Opcode: ISD::TRUNCATE, DL: dl, VT: MVT::i16, Operand: Op)); |
3353 | } else { |
3354 | Op = DAG.getAnyExtOrTrunc(Op, DL: dl, VT: Node->getValueType(ResNo: 0)); |
3355 | } |
3356 | Results.push_back(Elt: Op); |
3357 | break; |
3358 | } |
3359 | case ISD::SIGN_EXTEND_INREG: { |
3360 | EVT = cast<VTSDNode>(Val: Node->getOperand(Num: 1))->getVT(); |
3361 | EVT VT = Node->getValueType(ResNo: 0); |
3362 | |
3363 | // An in-register sign-extend of a boolean is a negation: |
3364 | // 'true' (1) sign-extended is -1. |
3365 | // 'false' (0) sign-extended is 0. |
3366 | // However, we must mask the high bits of the source operand because the |
3367 | // SIGN_EXTEND_INREG does not guarantee that the high bits are already zero. |
3368 | |
3369 | // TODO: Do this for vectors too? |
3370 | if (ExtraVT.isScalarInteger() && ExtraVT.getSizeInBits() == 1) { |
3371 | SDValue One = DAG.getConstant(Val: 1, DL: dl, VT); |
3372 | SDValue And = DAG.getNode(Opcode: ISD::AND, DL: dl, VT, N1: Node->getOperand(Num: 0), N2: One); |
3373 | SDValue Zero = DAG.getConstant(Val: 0, DL: dl, VT); |
3374 | SDValue Neg = DAG.getNode(Opcode: ISD::SUB, DL: dl, VT, N1: Zero, N2: And); |
3375 | Results.push_back(Elt: Neg); |
3376 | break; |
3377 | } |
3378 | |
3379 | // NOTE: we could fall back on load/store here too for targets without |
3380 | // SRA. However, it is doubtful that any exist. |
3381 | EVT ShiftAmountTy = TLI.getShiftAmountTy(LHSTy: VT, DL: DAG.getDataLayout()); |
3382 | unsigned BitsDiff = VT.getScalarSizeInBits() - |
3383 | ExtraVT.getScalarSizeInBits(); |
3384 | SDValue ShiftCst = DAG.getConstant(Val: BitsDiff, DL: dl, VT: ShiftAmountTy); |
3385 | Tmp1 = DAG.getNode(Opcode: ISD::SHL, DL: dl, VT: Node->getValueType(ResNo: 0), |
3386 | N1: Node->getOperand(Num: 0), N2: ShiftCst); |
3387 | Tmp1 = DAG.getNode(Opcode: ISD::SRA, DL: dl, VT: Node->getValueType(ResNo: 0), N1: Tmp1, N2: ShiftCst); |
3388 | Results.push_back(Elt: Tmp1); |
3389 | break; |
3390 | } |
3391 | case ISD::UINT_TO_FP: |
3392 | case ISD::STRICT_UINT_TO_FP: |
3393 | if (TLI.expandUINT_TO_FP(N: Node, Result&: Tmp1, Chain&: Tmp2, DAG)) { |
3394 | Results.push_back(Elt: Tmp1); |
3395 | if (Node->isStrictFPOpcode()) |
3396 | Results.push_back(Elt: Tmp2); |
3397 | break; |
3398 | } |
3399 | [[fallthrough]]; |
3400 | case ISD::SINT_TO_FP: |
3401 | case ISD::STRICT_SINT_TO_FP: |
3402 | if ((Tmp1 = ExpandLegalINT_TO_FP(Node, Chain&: Tmp2))) { |
3403 | Results.push_back(Elt: Tmp1); |
3404 | if (Node->isStrictFPOpcode()) |
3405 | Results.push_back(Elt: Tmp2); |
3406 | } |
3407 | break; |
3408 | case ISD::FP_TO_SINT: |
3409 | if (TLI.expandFP_TO_SINT(N: Node, Result&: Tmp1, DAG)) |
3410 | Results.push_back(Elt: Tmp1); |
3411 | break; |
3412 | case ISD::STRICT_FP_TO_SINT: |
3413 | if (TLI.expandFP_TO_SINT(N: Node, Result&: Tmp1, DAG)) { |
3414 | ReplaceNode(Old: Node, New: Tmp1.getNode()); |
3415 | LLVM_DEBUG(dbgs() << "Successfully expanded STRICT_FP_TO_SINT node\n" ); |
3416 | return true; |
3417 | } |
3418 | break; |
3419 | case ISD::FP_TO_UINT: |
3420 | if (TLI.expandFP_TO_UINT(N: Node, Result&: Tmp1, Chain&: Tmp2, DAG)) |
3421 | Results.push_back(Elt: Tmp1); |
3422 | break; |
3423 | case ISD::STRICT_FP_TO_UINT: |
3424 | if (TLI.expandFP_TO_UINT(N: Node, Result&: Tmp1, Chain&: Tmp2, DAG)) { |
3425 | // Relink the chain. |
3426 | DAG.ReplaceAllUsesOfValueWith(From: SDValue(Node,1), To: Tmp2); |
3427 | // Replace the new UINT result. |
3428 | ReplaceNodeWithValue(Old: SDValue(Node, 0), New: Tmp1); |
3429 | LLVM_DEBUG(dbgs() << "Successfully expanded STRICT_FP_TO_UINT node\n" ); |
3430 | return true; |
3431 | } |
3432 | break; |
3433 | case ISD::FP_TO_SINT_SAT: |
3434 | case ISD::FP_TO_UINT_SAT: |
3435 | Results.push_back(Elt: TLI.expandFP_TO_INT_SAT(N: Node, DAG)); |
3436 | break; |
3437 | case ISD::LROUND: |
3438 | case ISD::LLROUND: { |
3439 | SDValue Arg = Node->getOperand(Num: 0); |
3440 | EVT ArgVT = Arg.getValueType(); |
3441 | EVT ResVT = Node->getValueType(ResNo: 0); |
3442 | SDLoc dl(Node); |
3443 | SDValue RoundNode = DAG.getNode(Opcode: ISD::FROUND, DL: dl, VT: ArgVT, Operand: Arg); |
3444 | Results.push_back(Elt: DAG.getNode(Opcode: ISD::FP_TO_SINT, DL: dl, VT: ResVT, Operand: RoundNode)); |
3445 | break; |
3446 | } |
3447 | case ISD::VAARG: |
3448 | Results.push_back(Elt: DAG.expandVAArg(Node)); |
3449 | Results.push_back(Elt: Results[0].getValue(R: 1)); |
3450 | break; |
3451 | case ISD::VACOPY: |
3452 | Results.push_back(Elt: DAG.expandVACopy(Node)); |
3453 | break; |
3454 | case ISD::EXTRACT_VECTOR_ELT: |
3455 | if (Node->getOperand(Num: 0).getValueType().getVectorElementCount().isScalar()) |
3456 | // This must be an access of the only element. Return it. |
3457 | Tmp1 = DAG.getNode(Opcode: ISD::BITCAST, DL: dl, VT: Node->getValueType(ResNo: 0), |
3458 | Operand: Node->getOperand(Num: 0)); |
3459 | else |
3460 | Tmp1 = ExpandExtractFromVectorThroughStack(Op: SDValue(Node, 0)); |
3461 | Results.push_back(Elt: Tmp1); |
3462 | break; |
3463 | case ISD::EXTRACT_SUBVECTOR: |
3464 | Results.push_back(Elt: ExpandExtractFromVectorThroughStack(Op: SDValue(Node, 0))); |
3465 | break; |
3466 | case ISD::INSERT_SUBVECTOR: |
3467 | Results.push_back(Elt: ExpandInsertToVectorThroughStack(Op: SDValue(Node, 0))); |
3468 | break; |
3469 | case ISD::CONCAT_VECTORS: |
3470 | if (EVT VectorValueType = Node->getOperand(Num: 0).getValueType(); |
3471 | VectorValueType.isScalableVector() || |
3472 | TLI.isOperationExpand(Op: ISD::EXTRACT_VECTOR_ELT, VT: VectorValueType)) |
3473 | Results.push_back(Elt: ExpandVectorBuildThroughStack(Node)); |
3474 | else |
3475 | Results.push_back(Elt: ExpandConcatVectors(Node)); |
3476 | break; |
3477 | case ISD::SCALAR_TO_VECTOR: |
3478 | Results.push_back(Elt: ExpandSCALAR_TO_VECTOR(Node)); |
3479 | break; |
3480 | case ISD::INSERT_VECTOR_ELT: |
3481 | Results.push_back(Elt: ExpandINSERT_VECTOR_ELT(Op: SDValue(Node, 0))); |
3482 | break; |
3483 | case ISD::VECTOR_SHUFFLE: { |
3484 | SmallVector<int, 32> NewMask; |
3485 | ArrayRef<int> Mask = cast<ShuffleVectorSDNode>(Val: Node)->getMask(); |
3486 | |
3487 | EVT VT = Node->getValueType(ResNo: 0); |
3488 | EVT EltVT = VT.getVectorElementType(); |
3489 | SDValue Op0 = Node->getOperand(Num: 0); |
3490 | SDValue Op1 = Node->getOperand(Num: 1); |
3491 | if (!TLI.isTypeLegal(VT: EltVT)) { |
3492 | EVT NewEltVT = TLI.getTypeToTransformTo(Context&: *DAG.getContext(), VT: EltVT); |
3493 | |
3494 | // BUILD_VECTOR operands are allowed to be wider than the element type. |
3495 | // But if NewEltVT is smaller that EltVT the BUILD_VECTOR does not accept |
3496 | // it. |
3497 | if (NewEltVT.bitsLT(VT: EltVT)) { |
3498 | // Convert shuffle node. |
3499 | // If original node was v4i64 and the new EltVT is i32, |
3500 | // cast operands to v8i32 and re-build the mask. |
3501 | |
3502 | // Calculate new VT, the size of the new VT should be equal to original. |
3503 | EVT NewVT = |
3504 | EVT::getVectorVT(Context&: *DAG.getContext(), VT: NewEltVT, |
3505 | NumElements: VT.getSizeInBits() / NewEltVT.getSizeInBits()); |
3506 | assert(NewVT.bitsEq(VT)); |
3507 | |
3508 | // cast operands to new VT |
3509 | Op0 = DAG.getNode(Opcode: ISD::BITCAST, DL: dl, VT: NewVT, Operand: Op0); |
3510 | Op1 = DAG.getNode(Opcode: ISD::BITCAST, DL: dl, VT: NewVT, Operand: Op1); |
3511 | |
3512 | // Convert the shuffle mask |
3513 | unsigned int factor = |
3514 | NewVT.getVectorNumElements()/VT.getVectorNumElements(); |
3515 | |
3516 | // EltVT gets smaller |
3517 | assert(factor > 0); |
3518 | |
3519 | for (unsigned i = 0; i < VT.getVectorNumElements(); ++i) { |
3520 | if (Mask[i] < 0) { |
3521 | for (unsigned fi = 0; fi < factor; ++fi) |
3522 | NewMask.push_back(Elt: Mask[i]); |
3523 | } |
3524 | else { |
3525 | for (unsigned fi = 0; fi < factor; ++fi) |
3526 | NewMask.push_back(Elt: Mask[i]*factor+fi); |
3527 | } |
3528 | } |
3529 | Mask = NewMask; |
3530 | VT = NewVT; |
3531 | } |
3532 | EltVT = NewEltVT; |
3533 | } |
3534 | unsigned NumElems = VT.getVectorNumElements(); |
3535 | SmallVector<SDValue, 16> Ops; |
3536 | for (unsigned i = 0; i != NumElems; ++i) { |
3537 | if (Mask[i] < 0) { |
3538 | Ops.push_back(Elt: DAG.getUNDEF(VT: EltVT)); |
3539 | continue; |
3540 | } |
3541 | unsigned Idx = Mask[i]; |
3542 | if (Idx < NumElems) |
3543 | Ops.push_back(Elt: DAG.getNode(Opcode: ISD::EXTRACT_VECTOR_ELT, DL: dl, VT: EltVT, N1: Op0, |
3544 | N2: DAG.getVectorIdxConstant(Val: Idx, DL: dl))); |
3545 | else |
3546 | Ops.push_back( |
3547 | Elt: DAG.getNode(Opcode: ISD::EXTRACT_VECTOR_ELT, DL: dl, VT: EltVT, N1: Op1, |
3548 | N2: DAG.getVectorIdxConstant(Val: Idx - NumElems, DL: dl))); |
3549 | } |
3550 | |
3551 | Tmp1 = DAG.getBuildVector(VT, DL: dl, Ops); |
3552 | // We may have changed the BUILD_VECTOR type. Cast it back to the Node type. |
3553 | Tmp1 = DAG.getNode(Opcode: ISD::BITCAST, DL: dl, VT: Node->getValueType(ResNo: 0), Operand: Tmp1); |
3554 | Results.push_back(Elt: Tmp1); |
3555 | break; |
3556 | } |
3557 | case ISD::VECTOR_SPLICE: { |
3558 | Results.push_back(Elt: TLI.expandVectorSplice(Node, DAG)); |
3559 | break; |
3560 | } |
3561 | case ISD::VECTOR_DEINTERLEAVE: { |
3562 | unsigned Factor = Node->getNumOperands(); |
3563 | if (Factor <= 2 || !isPowerOf2_32(Value: Factor)) |
3564 | break; |
3565 | SmallVector<SDValue, 8> Ops(Node->ops()); |
3566 | EVT VecVT = Node->getValueType(ResNo: 0); |
3567 | SmallVector<EVT> HalfVTs(Factor / 2, VecVT); |
3568 | // Deinterleave at Factor/2 so each result contains two factors interleaved: |
3569 | // a0b0 c0d0 a1b1 c1d1 -> [a0c0 b0d0] [a1c1 b1d1] |
3570 | SDValue L = DAG.getNode(Opcode: ISD::VECTOR_DEINTERLEAVE, DL: dl, ResultTys: HalfVTs, |
3571 | Ops: ArrayRef(Ops).take_front(N: Factor / 2)); |
3572 | SDValue R = DAG.getNode(Opcode: ISD::VECTOR_DEINTERLEAVE, DL: dl, ResultTys: HalfVTs, |
3573 | Ops: ArrayRef(Ops).take_back(N: Factor / 2)); |
3574 | Results.resize(N: Factor); |
3575 | // Deinterleave the 2 factors out: |
3576 | // [a0c0 a1c1] [b0d0 b1d1] -> a0a1 b0b1 c0c1 d0d1 |
3577 | for (unsigned I = 0; I < Factor / 2; I++) { |
3578 | SDValue Deinterleave = |
3579 | DAG.getNode(Opcode: ISD::VECTOR_DEINTERLEAVE, DL: dl, ResultTys: {VecVT, VecVT}, |
3580 | Ops: {L.getValue(R: I), R.getValue(R: I)}); |
3581 | Results[I] = Deinterleave.getValue(R: 0); |
3582 | Results[I + Factor / 2] = Deinterleave.getValue(R: 1); |
3583 | } |
3584 | break; |
3585 | } |
3586 | case ISD::VECTOR_INTERLEAVE: { |
3587 | unsigned Factor = Node->getNumOperands(); |
3588 | if (Factor <= 2 || !isPowerOf2_32(Value: Factor)) |
3589 | break; |
3590 | EVT VecVT = Node->getValueType(ResNo: 0); |
3591 | SmallVector<EVT> HalfVTs(Factor / 2, VecVT); |
3592 | SmallVector<SDValue, 8> LOps, ROps; |
3593 | // Interleave so we have 2 factors per result: |
3594 | // a0a1 b0b1 c0c1 d0d1 -> [a0c0 b0d0] [a1c1 b1d1] |
3595 | for (unsigned I = 0; I < Factor / 2; I++) { |
3596 | SDValue Interleave = |
3597 | DAG.getNode(Opcode: ISD::VECTOR_INTERLEAVE, DL: dl, ResultTys: {VecVT, VecVT}, |
3598 | Ops: {Node->getOperand(Num: I), Node->getOperand(Num: I + Factor / 2)}); |
3599 | LOps.push_back(Elt: Interleave.getValue(R: 0)); |
3600 | ROps.push_back(Elt: Interleave.getValue(R: 1)); |
3601 | } |
3602 | // Interleave at Factor/2: |
3603 | // [a0c0 b0d0] [a1c1 b1d1] -> a0b0 c0d0 a1b1 c1d1 |
3604 | SDValue L = DAG.getNode(Opcode: ISD::VECTOR_INTERLEAVE, DL: dl, ResultTys: HalfVTs, Ops: LOps); |
3605 | SDValue R = DAG.getNode(Opcode: ISD::VECTOR_INTERLEAVE, DL: dl, ResultTys: HalfVTs, Ops: ROps); |
3606 | for (unsigned I = 0; I < Factor / 2; I++) |
3607 | Results.push_back(Elt: L.getValue(R: I)); |
3608 | for (unsigned I = 0; I < Factor / 2; I++) |
3609 | Results.push_back(Elt: R.getValue(R: I)); |
3610 | break; |
3611 | } |
3612 | case ISD::EXTRACT_ELEMENT: { |
3613 | EVT OpTy = Node->getOperand(Num: 0).getValueType(); |
3614 | if (Node->getConstantOperandVal(Num: 1)) { |
3615 | // 1 -> Hi |
3616 | Tmp1 = DAG.getNode(Opcode: ISD::SRL, DL: dl, VT: OpTy, N1: Node->getOperand(Num: 0), |
3617 | N2: DAG.getConstant(Val: OpTy.getSizeInBits() / 2, DL: dl, |
3618 | VT: TLI.getShiftAmountTy( |
3619 | LHSTy: Node->getOperand(Num: 0).getValueType(), |
3620 | DL: DAG.getDataLayout()))); |
3621 | Tmp1 = DAG.getNode(Opcode: ISD::TRUNCATE, DL: dl, VT: Node->getValueType(ResNo: 0), Operand: Tmp1); |
3622 | } else { |
3623 | // 0 -> Lo |
3624 | Tmp1 = DAG.getNode(Opcode: ISD::TRUNCATE, DL: dl, VT: Node->getValueType(ResNo: 0), |
3625 | Operand: Node->getOperand(Num: 0)); |
3626 | } |
3627 | Results.push_back(Elt: Tmp1); |
3628 | break; |
3629 | } |
3630 | case ISD::STACKSAVE: |
3631 | // Expand to CopyFromReg if the target set |
3632 | // StackPointerRegisterToSaveRestore. |
3633 | if (Register SP = TLI.getStackPointerRegisterToSaveRestore()) { |
3634 | Results.push_back(Elt: DAG.getCopyFromReg(Chain: Node->getOperand(Num: 0), dl, Reg: SP, |
3635 | VT: Node->getValueType(ResNo: 0))); |
3636 | Results.push_back(Elt: Results[0].getValue(R: 1)); |
3637 | } else { |
3638 | Results.push_back(Elt: DAG.getUNDEF(VT: Node->getValueType(ResNo: 0))); |
3639 | Results.push_back(Elt: Node->getOperand(Num: 0)); |
3640 | } |
3641 | break; |
3642 | case ISD::STACKRESTORE: |
3643 | // Expand to CopyToReg if the target set |
3644 | // StackPointerRegisterToSaveRestore. |
3645 | if (Register SP = TLI.getStackPointerRegisterToSaveRestore()) { |
3646 | Results.push_back(Elt: DAG.getCopyToReg(Chain: Node->getOperand(Num: 0), dl, Reg: SP, |
3647 | N: Node->getOperand(Num: 1))); |
3648 | } else { |
3649 | Results.push_back(Elt: Node->getOperand(Num: 0)); |
3650 | } |
3651 | break; |
3652 | case ISD::GET_DYNAMIC_AREA_OFFSET: |
3653 | Results.push_back(Elt: DAG.getConstant(Val: 0, DL: dl, VT: Node->getValueType(ResNo: 0))); |
3654 | Results.push_back(Elt: Results[0].getValue(R: 0)); |
3655 | break; |
3656 | case ISD::FCOPYSIGN: |
3657 | Results.push_back(Elt: ExpandFCOPYSIGN(Node)); |
3658 | break; |
3659 | case ISD::FNEG: |
3660 | Results.push_back(Elt: ExpandFNEG(Node)); |
3661 | break; |
3662 | case ISD::FABS: |
3663 | Results.push_back(Elt: ExpandFABS(Node)); |
3664 | break; |
3665 | case ISD::IS_FPCLASS: { |
3666 | auto Test = static_cast<FPClassTest>(Node->getConstantOperandVal(Num: 1)); |
3667 | if (SDValue Expanded = |
3668 | TLI.expandIS_FPCLASS(ResultVT: Node->getValueType(ResNo: 0), Op: Node->getOperand(Num: 0), |
3669 | Test, Flags: Node->getFlags(), DL: SDLoc(Node), DAG)) |
3670 | Results.push_back(Elt: Expanded); |
3671 | break; |
3672 | } |
3673 | case ISD::SMIN: |
3674 | case ISD::SMAX: |
3675 | case ISD::UMIN: |
3676 | case ISD::UMAX: { |
3677 | // Expand Y = MAX(A, B) -> Y = (A > B) ? A : B |
3678 | ISD::CondCode Pred; |
3679 | switch (Node->getOpcode()) { |
3680 | default: llvm_unreachable("How did we get here?" ); |
3681 | case ISD::SMAX: Pred = ISD::SETGT; break; |
3682 | case ISD::SMIN: Pred = ISD::SETLT; break; |
3683 | case ISD::UMAX: Pred = ISD::SETUGT; break; |
3684 | case ISD::UMIN: Pred = ISD::SETULT; break; |
3685 | } |
3686 | Tmp1 = Node->getOperand(Num: 0); |
3687 | Tmp2 = Node->getOperand(Num: 1); |
3688 | Tmp1 = DAG.getSelectCC(DL: dl, LHS: Tmp1, RHS: Tmp2, True: Tmp1, False: Tmp2, Cond: Pred); |
3689 | Results.push_back(Elt: Tmp1); |
3690 | break; |
3691 | } |
3692 | case ISD::FMINNUM: |
3693 | case ISD::FMAXNUM: { |
3694 | if (SDValue Expanded = TLI.expandFMINNUM_FMAXNUM(N: Node, DAG)) |
3695 | Results.push_back(Elt: Expanded); |
3696 | break; |
3697 | } |
3698 | case ISD::FMINIMUM: |
3699 | case ISD::FMAXIMUM: { |
3700 | if (SDValue Expanded = TLI.expandFMINIMUM_FMAXIMUM(N: Node, DAG)) |
3701 | Results.push_back(Elt: Expanded); |
3702 | break; |
3703 | } |
3704 | case ISD::FMINIMUMNUM: |
3705 | case ISD::FMAXIMUMNUM: { |
3706 | Results.push_back(Elt: TLI.expandFMINIMUMNUM_FMAXIMUMNUM(N: Node, DAG)); |
3707 | break; |
3708 | } |
3709 | case ISD::FSIN: |
3710 | case ISD::FCOS: { |
3711 | EVT VT = Node->getValueType(ResNo: 0); |
3712 | // Turn fsin / fcos into ISD::FSINCOS node if there are a pair of fsin / |
3713 | // fcos which share the same operand and both are used. |
3714 | if ((TLI.isOperationLegalOrCustom(Op: ISD::FSINCOS, VT) || |
3715 | isSinCosLibcallAvailable(Node, TLI)) |
3716 | && useSinCos(Node)) { |
3717 | SDVTList VTs = DAG.getVTList(VT1: VT, VT2: VT); |
3718 | Tmp1 = DAG.getNode(Opcode: ISD::FSINCOS, DL: dl, VTList: VTs, N: Node->getOperand(Num: 0)); |
3719 | if (Node->getOpcode() == ISD::FCOS) |
3720 | Tmp1 = Tmp1.getValue(R: 1); |
3721 | Results.push_back(Elt: Tmp1); |
3722 | } |
3723 | break; |
3724 | } |
3725 | case ISD::FLDEXP: |
3726 | case ISD::STRICT_FLDEXP: { |
3727 | EVT VT = Node->getValueType(ResNo: 0); |
3728 | RTLIB::Libcall LC = RTLIB::getLDEXP(RetVT: VT); |
3729 | // Use the LibCall instead, it is very likely faster |
3730 | // FIXME: Use separate LibCall action. |
3731 | if (TLI.getLibcallName(Call: LC)) |
3732 | break; |
3733 | |
3734 | if (SDValue Expanded = expandLdexp(Node)) { |
3735 | Results.push_back(Elt: Expanded); |
3736 | if (Node->getOpcode() == ISD::STRICT_FLDEXP) |
3737 | Results.push_back(Elt: Expanded.getValue(R: 1)); |
3738 | } |
3739 | |
3740 | break; |
3741 | } |
3742 | case ISD::FFREXP: { |
3743 | RTLIB::Libcall LC = RTLIB::getFREXP(RetVT: Node->getValueType(ResNo: 0)); |
3744 | // Use the LibCall instead, it is very likely faster |
3745 | // FIXME: Use separate LibCall action. |
3746 | if (TLI.getLibcallName(Call: LC)) |
3747 | break; |
3748 | |
3749 | if (SDValue Expanded = expandFrexp(Node)) { |
3750 | Results.push_back(Elt: Expanded); |
3751 | Results.push_back(Elt: Expanded.getValue(R: 1)); |
3752 | } |
3753 | break; |
3754 | } |
3755 | case ISD::FSINCOS: { |
3756 | if (isSinCosLibcallAvailable(Node, TLI)) |
3757 | break; |
3758 | EVT VT = Node->getValueType(ResNo: 0); |
3759 | SDValue Op = Node->getOperand(Num: 0); |
3760 | SDNodeFlags Flags = Node->getFlags(); |
3761 | Tmp1 = DAG.getNode(Opcode: ISD::FSIN, DL: dl, VT, Operand: Op, Flags); |
3762 | Tmp2 = DAG.getNode(Opcode: ISD::FCOS, DL: dl, VT, Operand: Op, Flags); |
3763 | Results.append(IL: {Tmp1, Tmp2}); |
3764 | break; |
3765 | } |
3766 | case ISD::FMAD: |
3767 | llvm_unreachable("Illegal fmad should never be formed" ); |
3768 | |
3769 | case ISD::FP16_TO_FP: |
3770 | if (Node->getValueType(ResNo: 0) != MVT::f32) { |
3771 | // We can extend to types bigger than f32 in two steps without changing |
3772 | // the result. Since "f16 -> f32" is much more commonly available, give |
3773 | // CodeGen the option of emitting that before resorting to a libcall. |
3774 | SDValue Res = |
3775 | DAG.getNode(Opcode: ISD::FP16_TO_FP, DL: dl, VT: MVT::f32, Operand: Node->getOperand(Num: 0)); |
3776 | Results.push_back( |
3777 | Elt: DAG.getNode(Opcode: ISD::FP_EXTEND, DL: dl, VT: Node->getValueType(ResNo: 0), Operand: Res)); |
3778 | } |
3779 | break; |
3780 | case ISD::STRICT_BF16_TO_FP: |
3781 | case ISD::STRICT_FP16_TO_FP: |
3782 | if (Node->getValueType(ResNo: 0) != MVT::f32) { |
3783 | // We can extend to types bigger than f32 in two steps without changing |
3784 | // the result. Since "f16 -> f32" is much more commonly available, give |
3785 | // CodeGen the option of emitting that before resorting to a libcall. |
3786 | SDValue Res = DAG.getNode(Opcode: Node->getOpcode(), DL: dl, ResultTys: {MVT::f32, MVT::Other}, |
3787 | Ops: {Node->getOperand(Num: 0), Node->getOperand(Num: 1)}); |
3788 | Res = DAG.getNode(Opcode: ISD::STRICT_FP_EXTEND, DL: dl, |
3789 | ResultTys: {Node->getValueType(ResNo: 0), MVT::Other}, |
3790 | Ops: {Res.getValue(R: 1), Res}); |
3791 | Results.push_back(Elt: Res); |
3792 | Results.push_back(Elt: Res.getValue(R: 1)); |
3793 | } |
3794 | break; |
3795 | case ISD::FP_TO_FP16: |
3796 | LLVM_DEBUG(dbgs() << "Legalizing FP_TO_FP16\n" ); |
3797 | if (!TLI.useSoftFloat() && TM.Options.UnsafeFPMath) { |
3798 | SDValue Op = Node->getOperand(Num: 0); |
3799 | MVT SVT = Op.getSimpleValueType(); |
3800 | if ((SVT == MVT::f64 || SVT == MVT::f80) && |
3801 | TLI.isOperationLegalOrCustom(Op: ISD::FP_TO_FP16, VT: MVT::f32)) { |
3802 | // Under fastmath, we can expand this node into a fround followed by |
3803 | // a float-half conversion. |
3804 | SDValue FloatVal = |
3805 | DAG.getNode(Opcode: ISD::FP_ROUND, DL: dl, VT: MVT::f32, N1: Op, |
3806 | N2: DAG.getIntPtrConstant(Val: 0, DL: dl, /*isTarget=*/true)); |
3807 | Results.push_back( |
3808 | Elt: DAG.getNode(Opcode: ISD::FP_TO_FP16, DL: dl, VT: Node->getValueType(ResNo: 0), Operand: FloatVal)); |
3809 | } |
3810 | } |
3811 | break; |
3812 | case ISD::ConstantFP: { |
3813 | ConstantFPSDNode *CFP = cast<ConstantFPSDNode>(Val: Node); |
3814 | // Check to see if this FP immediate is already legal. |
3815 | // If this is a legal constant, turn it into a TargetConstantFP node. |
3816 | if (!TLI.isFPImmLegal(CFP->getValueAPF(), Node->getValueType(ResNo: 0), |
3817 | ForCodeSize: DAG.shouldOptForSize())) |
3818 | Results.push_back(Elt: ExpandConstantFP(CFP, UseCP: true)); |
3819 | break; |
3820 | } |
3821 | case ISD::Constant: { |
3822 | ConstantSDNode *CP = cast<ConstantSDNode>(Val: Node); |
3823 | Results.push_back(Elt: ExpandConstant(CP)); |
3824 | break; |
3825 | } |
3826 | case ISD::FSUB: { |
3827 | EVT VT = Node->getValueType(ResNo: 0); |
3828 | if (TLI.isOperationLegalOrCustom(Op: ISD::FADD, VT) && |
3829 | TLI.isOperationLegalOrCustom(Op: ISD::FNEG, VT)) { |
3830 | const SDNodeFlags Flags = Node->getFlags(); |
3831 | Tmp1 = DAG.getNode(Opcode: ISD::FNEG, DL: dl, VT, Operand: Node->getOperand(Num: 1)); |
3832 | Tmp1 = DAG.getNode(Opcode: ISD::FADD, DL: dl, VT, N1: Node->getOperand(Num: 0), N2: Tmp1, Flags); |
3833 | Results.push_back(Elt: Tmp1); |
3834 | } |
3835 | break; |
3836 | } |
3837 | case ISD::SUB: { |
3838 | EVT VT = Node->getValueType(ResNo: 0); |
3839 | assert(TLI.isOperationLegalOrCustom(ISD::ADD, VT) && |
3840 | TLI.isOperationLegalOrCustom(ISD::XOR, VT) && |
3841 | "Don't know how to expand this subtraction!" ); |
3842 | Tmp1 = DAG.getNOT(DL: dl, Val: Node->getOperand(Num: 1), VT); |
3843 | Tmp1 = DAG.getNode(Opcode: ISD::ADD, DL: dl, VT, N1: Tmp1, N2: DAG.getConstant(Val: 1, DL: dl, VT)); |
3844 | Results.push_back(Elt: DAG.getNode(Opcode: ISD::ADD, DL: dl, VT, N1: Node->getOperand(Num: 0), N2: Tmp1)); |
3845 | break; |
3846 | } |
3847 | case ISD::UREM: |
3848 | case ISD::SREM: |
3849 | if (TLI.expandREM(Node, Result&: Tmp1, DAG)) |
3850 | Results.push_back(Elt: Tmp1); |
3851 | break; |
3852 | case ISD::UDIV: |
3853 | case ISD::SDIV: { |
3854 | bool isSigned = Node->getOpcode() == ISD::SDIV; |
3855 | unsigned DivRemOpc = isSigned ? ISD::SDIVREM : ISD::UDIVREM; |
3856 | EVT VT = Node->getValueType(ResNo: 0); |
3857 | if (TLI.isOperationLegalOrCustom(Op: DivRemOpc, VT)) { |
3858 | SDVTList VTs = DAG.getVTList(VT1: VT, VT2: VT); |
3859 | Tmp1 = DAG.getNode(Opcode: DivRemOpc, DL: dl, VTList: VTs, N1: Node->getOperand(Num: 0), |
3860 | N2: Node->getOperand(Num: 1)); |
3861 | Results.push_back(Elt: Tmp1); |
3862 | } |
3863 | break; |
3864 | } |
3865 | case ISD::MULHU: |
3866 | case ISD::MULHS: { |
3867 | unsigned ExpandOpcode = |
3868 | Node->getOpcode() == ISD::MULHU ? ISD::UMUL_LOHI : ISD::SMUL_LOHI; |
3869 | EVT VT = Node->getValueType(ResNo: 0); |
3870 | SDVTList VTs = DAG.getVTList(VT1: VT, VT2: VT); |
3871 | |
3872 | Tmp1 = DAG.getNode(Opcode: ExpandOpcode, DL: dl, VTList: VTs, N1: Node->getOperand(Num: 0), |
3873 | N2: Node->getOperand(Num: 1)); |
3874 | Results.push_back(Elt: Tmp1.getValue(R: 1)); |
3875 | break; |
3876 | } |
3877 | case ISD::UMUL_LOHI: |
3878 | case ISD::SMUL_LOHI: { |
3879 | SDValue LHS = Node->getOperand(Num: 0); |
3880 | SDValue RHS = Node->getOperand(Num: 1); |
3881 | MVT VT = LHS.getSimpleValueType(); |
3882 | unsigned MULHOpcode = |
3883 | Node->getOpcode() == ISD::UMUL_LOHI ? ISD::MULHU : ISD::MULHS; |
3884 | |
3885 | if (TLI.isOperationLegalOrCustom(Op: MULHOpcode, VT)) { |
3886 | Results.push_back(Elt: DAG.getNode(Opcode: ISD::MUL, DL: dl, VT, N1: LHS, N2: RHS)); |
3887 | Results.push_back(Elt: DAG.getNode(Opcode: MULHOpcode, DL: dl, VT, N1: LHS, N2: RHS)); |
3888 | break; |
3889 | } |
3890 | |
3891 | SmallVector<SDValue, 4> Halves; |
3892 | EVT HalfType = EVT(VT).getHalfSizedIntegerVT(Context&: *DAG.getContext()); |
3893 | assert(TLI.isTypeLegal(HalfType)); |
3894 | if (TLI.expandMUL_LOHI(Opcode: Node->getOpcode(), VT, dl, LHS, RHS, Result&: Halves, |
3895 | HiLoVT: HalfType, DAG, |
3896 | Kind: TargetLowering::MulExpansionKind::Always)) { |
3897 | for (unsigned i = 0; i < 2; ++i) { |
3898 | SDValue Lo = DAG.getNode(Opcode: ISD::ZERO_EXTEND, DL: dl, VT, Operand: Halves[2 * i]); |
3899 | SDValue Hi = DAG.getNode(Opcode: ISD::ANY_EXTEND, DL: dl, VT, Operand: Halves[2 * i + 1]); |
3900 | SDValue Shift = DAG.getConstant( |
3901 | Val: HalfType.getScalarSizeInBits(), DL: dl, |
3902 | VT: TLI.getShiftAmountTy(LHSTy: HalfType, DL: DAG.getDataLayout())); |
3903 | Hi = DAG.getNode(Opcode: ISD::SHL, DL: dl, VT, N1: Hi, N2: Shift); |
3904 | Results.push_back(Elt: DAG.getNode(Opcode: ISD::OR, DL: dl, VT, N1: Lo, N2: Hi)); |
3905 | } |
3906 | break; |
3907 | } |
3908 | break; |
3909 | } |
3910 | case ISD::MUL: { |
3911 | EVT VT = Node->getValueType(ResNo: 0); |
3912 | SDVTList VTs = DAG.getVTList(VT1: VT, VT2: VT); |
3913 | // See if multiply or divide can be lowered using two-result operations. |
3914 | // We just need the low half of the multiply; try both the signed |
3915 | // and unsigned forms. If the target supports both SMUL_LOHI and |
3916 | // UMUL_LOHI, form a preference by checking which forms of plain |
3917 | // MULH it supports. |
3918 | bool HasSMUL_LOHI = TLI.isOperationLegalOrCustom(Op: ISD::SMUL_LOHI, VT); |
3919 | bool HasUMUL_LOHI = TLI.isOperationLegalOrCustom(Op: ISD::UMUL_LOHI, VT); |
3920 | bool HasMULHS = TLI.isOperationLegalOrCustom(Op: ISD::MULHS, VT); |
3921 | bool HasMULHU = TLI.isOperationLegalOrCustom(Op: ISD::MULHU, VT); |
3922 | unsigned OpToUse = 0; |
3923 | if (HasSMUL_LOHI && !HasMULHS) { |
3924 | OpToUse = ISD::SMUL_LOHI; |
3925 | } else if (HasUMUL_LOHI && !HasMULHU) { |
3926 | OpToUse = ISD::UMUL_LOHI; |
3927 | } else if (HasSMUL_LOHI) { |
3928 | OpToUse = ISD::SMUL_LOHI; |
3929 | } else if (HasUMUL_LOHI) { |
3930 | OpToUse = ISD::UMUL_LOHI; |
3931 | } |
3932 | if (OpToUse) { |
3933 | Results.push_back(Elt: DAG.getNode(Opcode: OpToUse, DL: dl, VTList: VTs, N1: Node->getOperand(Num: 0), |
3934 | N2: Node->getOperand(Num: 1))); |
3935 | break; |
3936 | } |
3937 | |
3938 | SDValue Lo, Hi; |
3939 | EVT HalfType = VT.getHalfSizedIntegerVT(Context&: *DAG.getContext()); |
3940 | if (TLI.isOperationLegalOrCustom(Op: ISD::ZERO_EXTEND, VT) && |
3941 | TLI.isOperationLegalOrCustom(Op: ISD::ANY_EXTEND, VT) && |
3942 | TLI.isOperationLegalOrCustom(Op: ISD::SHL, VT) && |
3943 | TLI.isOperationLegalOrCustom(Op: ISD::OR, VT) && |
3944 | TLI.expandMUL(N: Node, Lo, Hi, HiLoVT: HalfType, DAG, |
3945 | Kind: TargetLowering::MulExpansionKind::OnlyLegalOrCustom)) { |
3946 | Lo = DAG.getNode(Opcode: ISD::ZERO_EXTEND, DL: dl, VT, Operand: Lo); |
3947 | Hi = DAG.getNode(Opcode: ISD::ANY_EXTEND, DL: dl, VT, Operand: Hi); |
3948 | SDValue Shift = |
3949 | DAG.getConstant(Val: HalfType.getSizeInBits(), DL: dl, |
3950 | VT: TLI.getShiftAmountTy(LHSTy: HalfType, DL: DAG.getDataLayout())); |
3951 | Hi = DAG.getNode(Opcode: ISD::SHL, DL: dl, VT, N1: Hi, N2: Shift); |
3952 | Results.push_back(Elt: DAG.getNode(Opcode: ISD::OR, DL: dl, VT, N1: Lo, N2: Hi)); |
3953 | } |
3954 | break; |
3955 | } |
3956 | case ISD::FSHL: |
3957 | case ISD::FSHR: |
3958 | if (SDValue Expanded = TLI.expandFunnelShift(N: Node, DAG)) |
3959 | Results.push_back(Elt: Expanded); |
3960 | break; |
3961 | case ISD::ROTL: |
3962 | case ISD::ROTR: |
3963 | if (SDValue Expanded = TLI.expandROT(N: Node, AllowVectorOps: true /*AllowVectorOps*/, DAG)) |
3964 | Results.push_back(Elt: Expanded); |
3965 | break; |
3966 | case ISD::SADDSAT: |
3967 | case ISD::UADDSAT: |
3968 | case ISD::SSUBSAT: |
3969 | case ISD::USUBSAT: |
3970 | Results.push_back(Elt: TLI.expandAddSubSat(Node, DAG)); |
3971 | break; |
3972 | case ISD::SCMP: |
3973 | case ISD::UCMP: |
3974 | Results.push_back(Elt: TLI.expandCMP(Node, DAG)); |
3975 | break; |
3976 | case ISD::SSHLSAT: |
3977 | case ISD::USHLSAT: |
3978 | Results.push_back(Elt: TLI.expandShlSat(Node, DAG)); |
3979 | break; |
3980 | case ISD::SMULFIX: |
3981 | case ISD::SMULFIXSAT: |
3982 | case ISD::UMULFIX: |
3983 | case ISD::UMULFIXSAT: |
3984 | Results.push_back(Elt: TLI.expandFixedPointMul(Node, DAG)); |
3985 | break; |
3986 | case ISD::SDIVFIX: |
3987 | case ISD::SDIVFIXSAT: |
3988 | case ISD::UDIVFIX: |
3989 | case ISD::UDIVFIXSAT: |
3990 | if (SDValue V = TLI.expandFixedPointDiv(Opcode: Node->getOpcode(), dl: SDLoc(Node), |
3991 | LHS: Node->getOperand(Num: 0), |
3992 | RHS: Node->getOperand(Num: 1), |
3993 | Scale: Node->getConstantOperandVal(Num: 2), |
3994 | DAG)) { |
3995 | Results.push_back(Elt: V); |
3996 | break; |
3997 | } |
3998 | // FIXME: We might want to retry here with a wider type if we fail, if that |
3999 | // type is legal. |
4000 | // FIXME: Technically, so long as we only have sdivfixes where BW+Scale is |
4001 | // <= 128 (which is the case for all of the default Embedded-C types), |
4002 | // we will only get here with types and scales that we could always expand |
4003 | // if we were allowed to generate libcalls to division functions of illegal |
4004 | // type. But we cannot do that. |
4005 | llvm_unreachable("Cannot expand DIVFIX!" ); |
4006 | case ISD::UADDO_CARRY: |
4007 | case ISD::USUBO_CARRY: { |
4008 | SDValue LHS = Node->getOperand(Num: 0); |
4009 | SDValue RHS = Node->getOperand(Num: 1); |
4010 | SDValue Carry = Node->getOperand(Num: 2); |
4011 | |
4012 | bool IsAdd = Node->getOpcode() == ISD::UADDO_CARRY; |
4013 | |
4014 | // Initial add of the 2 operands. |
4015 | unsigned Op = IsAdd ? ISD::ADD : ISD::SUB; |
4016 | EVT VT = LHS.getValueType(); |
4017 | SDValue Sum = DAG.getNode(Opcode: Op, DL: dl, VT, N1: LHS, N2: RHS); |
4018 | |
4019 | // Initial check for overflow. |
4020 | EVT CarryType = Node->getValueType(ResNo: 1); |
4021 | EVT SetCCType = getSetCCResultType(VT: Node->getValueType(ResNo: 0)); |
4022 | ISD::CondCode CC = IsAdd ? ISD::SETULT : ISD::SETUGT; |
4023 | SDValue Overflow = DAG.getSetCC(DL: dl, VT: SetCCType, LHS: Sum, RHS: LHS, Cond: CC); |
4024 | |
4025 | // Add of the sum and the carry. |
4026 | SDValue One = DAG.getConstant(Val: 1, DL: dl, VT); |
4027 | SDValue CarryExt = |
4028 | DAG.getNode(Opcode: ISD::AND, DL: dl, VT, N1: DAG.getZExtOrTrunc(Op: Carry, DL: dl, VT), N2: One); |
4029 | SDValue Sum2 = DAG.getNode(Opcode: Op, DL: dl, VT, N1: Sum, N2: CarryExt); |
4030 | |
4031 | // Second check for overflow. If we are adding, we can only overflow if the |
4032 | // initial sum is all 1s ang the carry is set, resulting in a new sum of 0. |
4033 | // If we are subtracting, we can only overflow if the initial sum is 0 and |
4034 | // the carry is set, resulting in a new sum of all 1s. |
4035 | SDValue Zero = DAG.getConstant(Val: 0, DL: dl, VT); |
4036 | SDValue Overflow2 = |
4037 | IsAdd ? DAG.getSetCC(DL: dl, VT: SetCCType, LHS: Sum2, RHS: Zero, Cond: ISD::SETEQ) |
4038 | : DAG.getSetCC(DL: dl, VT: SetCCType, LHS: Sum, RHS: Zero, Cond: ISD::SETEQ); |
4039 | Overflow2 = DAG.getNode(Opcode: ISD::AND, DL: dl, VT: SetCCType, N1: Overflow2, |
4040 | N2: DAG.getZExtOrTrunc(Op: Carry, DL: dl, VT: SetCCType)); |
4041 | |
4042 | SDValue ResultCarry = |
4043 | DAG.getNode(Opcode: ISD::OR, DL: dl, VT: SetCCType, N1: Overflow, N2: Overflow2); |
4044 | |
4045 | Results.push_back(Elt: Sum2); |
4046 | Results.push_back(Elt: DAG.getBoolExtOrTrunc(Op: ResultCarry, SL: dl, VT: CarryType, OpVT: VT)); |
4047 | break; |
4048 | } |
4049 | case ISD::SADDO: |
4050 | case ISD::SSUBO: { |
4051 | SDValue Result, Overflow; |
4052 | TLI.expandSADDSUBO(Node, Result, Overflow, DAG); |
4053 | Results.push_back(Elt: Result); |
4054 | Results.push_back(Elt: Overflow); |
4055 | break; |
4056 | } |
4057 | case ISD::UADDO: |
4058 | case ISD::USUBO: { |
4059 | SDValue Result, Overflow; |
4060 | TLI.expandUADDSUBO(Node, Result, Overflow, DAG); |
4061 | Results.push_back(Elt: Result); |
4062 | Results.push_back(Elt: Overflow); |
4063 | break; |
4064 | } |
4065 | case ISD::UMULO: |
4066 | case ISD::SMULO: { |
4067 | SDValue Result, Overflow; |
4068 | if (TLI.expandMULO(Node, Result, Overflow, DAG)) { |
4069 | Results.push_back(Elt: Result); |
4070 | Results.push_back(Elt: Overflow); |
4071 | } |
4072 | break; |
4073 | } |
4074 | case ISD::BUILD_PAIR: { |
4075 | EVT PairTy = Node->getValueType(ResNo: 0); |
4076 | Tmp1 = DAG.getNode(Opcode: ISD::ZERO_EXTEND, DL: dl, VT: PairTy, Operand: Node->getOperand(Num: 0)); |
4077 | Tmp2 = DAG.getNode(Opcode: ISD::ANY_EXTEND, DL: dl, VT: PairTy, Operand: Node->getOperand(Num: 1)); |
4078 | Tmp2 = DAG.getNode( |
4079 | Opcode: ISD::SHL, DL: dl, VT: PairTy, N1: Tmp2, |
4080 | N2: DAG.getConstant(Val: PairTy.getSizeInBits() / 2, DL: dl, |
4081 | VT: TLI.getShiftAmountTy(LHSTy: PairTy, DL: DAG.getDataLayout()))); |
4082 | Results.push_back(Elt: DAG.getNode(Opcode: ISD::OR, DL: dl, VT: PairTy, N1: Tmp1, N2: Tmp2)); |
4083 | break; |
4084 | } |
4085 | case ISD::SELECT: |
4086 | Tmp1 = Node->getOperand(Num: 0); |
4087 | Tmp2 = Node->getOperand(Num: 1); |
4088 | Tmp3 = Node->getOperand(Num: 2); |
4089 | if (Tmp1.getOpcode() == ISD::SETCC) { |
4090 | Tmp1 = DAG.getSelectCC(DL: dl, LHS: Tmp1.getOperand(i: 0), RHS: Tmp1.getOperand(i: 1), |
4091 | True: Tmp2, False: Tmp3, |
4092 | Cond: cast<CondCodeSDNode>(Val: Tmp1.getOperand(i: 2))->get()); |
4093 | } else { |
4094 | Tmp1 = DAG.getSelectCC(DL: dl, LHS: Tmp1, |
4095 | RHS: DAG.getConstant(Val: 0, DL: dl, VT: Tmp1.getValueType()), |
4096 | True: Tmp2, False: Tmp3, Cond: ISD::SETNE); |
4097 | } |
4098 | Tmp1->setFlags(Node->getFlags()); |
4099 | Results.push_back(Elt: Tmp1); |
4100 | break; |
4101 | case ISD::BR_JT: { |
4102 | SDValue Chain = Node->getOperand(Num: 0); |
4103 | SDValue Table = Node->getOperand(Num: 1); |
4104 | SDValue Index = Node->getOperand(Num: 2); |
4105 | int JTI = cast<JumpTableSDNode>(Val: Table.getNode())->getIndex(); |
4106 | |
4107 | const DataLayout &TD = DAG.getDataLayout(); |
4108 | EVT PTy = TLI.getPointerTy(DL: TD); |
4109 | |
4110 | unsigned EntrySize = |
4111 | DAG.getMachineFunction().getJumpTableInfo()->getEntrySize(TD); |
4112 | |
4113 | // For power-of-two jumptable entry sizes convert multiplication to a shift. |
4114 | // This transformation needs to be done here since otherwise the MIPS |
4115 | // backend will end up emitting a three instruction multiply sequence |
4116 | // instead of a single shift and MSP430 will call a runtime function. |
4117 | if (llvm::isPowerOf2_32(Value: EntrySize)) |
4118 | Index = DAG.getNode( |
4119 | Opcode: ISD::SHL, DL: dl, VT: Index.getValueType(), N1: Index, |
4120 | N2: DAG.getConstant(Val: llvm::Log2_32(Value: EntrySize), DL: dl, VT: Index.getValueType())); |
4121 | else |
4122 | Index = DAG.getNode(Opcode: ISD::MUL, DL: dl, VT: Index.getValueType(), N1: Index, |
4123 | N2: DAG.getConstant(Val: EntrySize, DL: dl, VT: Index.getValueType())); |
4124 | SDValue Addr = DAG.getMemBasePlusOffset(Base: Table, Offset: Index, DL: dl); |
4125 | |
4126 | EVT MemVT = EVT::getIntegerVT(Context&: *DAG.getContext(), BitWidth: EntrySize * 8); |
4127 | SDValue LD = DAG.getExtLoad( |
4128 | ExtType: ISD::SEXTLOAD, dl, VT: PTy, Chain, Ptr: Addr, |
4129 | PtrInfo: MachinePointerInfo::getJumpTable(MF&: DAG.getMachineFunction()), MemVT); |
4130 | Addr = LD; |
4131 | if (TLI.isJumpTableRelative()) { |
4132 | // For PIC, the sequence is: |
4133 | // BRIND(RelocBase + load(Jumptable + index)) |
4134 | // RelocBase can be JumpTable, GOT or some sort of global base. |
4135 | Addr = DAG.getMemBasePlusOffset(Base: TLI.getPICJumpTableRelocBase(Table, DAG), |
4136 | Offset: Addr, DL: dl); |
4137 | } |
4138 | |
4139 | Tmp1 = TLI.expandIndirectJTBranch(dl, Value: LD.getValue(R: 1), Addr, JTI, DAG); |
4140 | Results.push_back(Elt: Tmp1); |
4141 | break; |
4142 | } |
4143 | case ISD::BRCOND: |
4144 | // Expand brcond's setcc into its constituent parts and create a BR_CC |
4145 | // Node. |
4146 | Tmp1 = Node->getOperand(Num: 0); |
4147 | Tmp2 = Node->getOperand(Num: 1); |
4148 | if (Tmp2.getOpcode() == ISD::SETCC && |
4149 | TLI.isOperationLegalOrCustom(Op: ISD::BR_CC, |
4150 | VT: Tmp2.getOperand(i: 0).getValueType())) { |
4151 | Tmp1 = DAG.getNode(Opcode: ISD::BR_CC, DL: dl, VT: MVT::Other, N1: Tmp1, N2: Tmp2.getOperand(i: 2), |
4152 | N3: Tmp2.getOperand(i: 0), N4: Tmp2.getOperand(i: 1), |
4153 | N5: Node->getOperand(Num: 2)); |
4154 | } else { |
4155 | // We test only the i1 bit. Skip the AND if UNDEF or another AND. |
4156 | if (Tmp2.isUndef() || |
4157 | (Tmp2.getOpcode() == ISD::AND && isOneConstant(V: Tmp2.getOperand(i: 1)))) |
4158 | Tmp3 = Tmp2; |
4159 | else |
4160 | Tmp3 = DAG.getNode(Opcode: ISD::AND, DL: dl, VT: Tmp2.getValueType(), N1: Tmp2, |
4161 | N2: DAG.getConstant(Val: 1, DL: dl, VT: Tmp2.getValueType())); |
4162 | Tmp1 = DAG.getNode(Opcode: ISD::BR_CC, DL: dl, VT: MVT::Other, N1: Tmp1, |
4163 | N2: DAG.getCondCode(Cond: ISD::SETNE), N3: Tmp3, |
4164 | N4: DAG.getConstant(Val: 0, DL: dl, VT: Tmp3.getValueType()), |
4165 | N5: Node->getOperand(Num: 2)); |
4166 | } |
4167 | Results.push_back(Elt: Tmp1); |
4168 | break; |
4169 | case ISD::SETCC: |
4170 | case ISD::VP_SETCC: |
4171 | case ISD::STRICT_FSETCC: |
4172 | case ISD::STRICT_FSETCCS: { |
4173 | bool IsVP = Node->getOpcode() == ISD::VP_SETCC; |
4174 | bool IsStrict = Node->getOpcode() == ISD::STRICT_FSETCC || |
4175 | Node->getOpcode() == ISD::STRICT_FSETCCS; |
4176 | bool IsSignaling = Node->getOpcode() == ISD::STRICT_FSETCCS; |
4177 | SDValue Chain = IsStrict ? Node->getOperand(Num: 0) : SDValue(); |
4178 | unsigned Offset = IsStrict ? 1 : 0; |
4179 | Tmp1 = Node->getOperand(Num: 0 + Offset); |
4180 | Tmp2 = Node->getOperand(Num: 1 + Offset); |
4181 | Tmp3 = Node->getOperand(Num: 2 + Offset); |
4182 | SDValue Mask, EVL; |
4183 | if (IsVP) { |
4184 | Mask = Node->getOperand(Num: 3 + Offset); |
4185 | EVL = Node->getOperand(Num: 4 + Offset); |
4186 | } |
4187 | bool Legalized = TLI.LegalizeSetCCCondCode( |
4188 | DAG, VT: Node->getValueType(ResNo: 0), LHS&: Tmp1, RHS&: Tmp2, CC&: Tmp3, Mask, EVL, NeedInvert, dl, |
4189 | Chain, IsSignaling); |
4190 | |
4191 | if (Legalized) { |
4192 | // If we expanded the SETCC by swapping LHS and RHS, or by inverting the |
4193 | // condition code, create a new SETCC node. |
4194 | if (Tmp3.getNode()) { |
4195 | if (IsStrict) { |
4196 | Tmp1 = DAG.getNode(Opcode: Node->getOpcode(), DL: dl, VTList: Node->getVTList(), |
4197 | Ops: {Chain, Tmp1, Tmp2, Tmp3}, Flags: Node->getFlags()); |
4198 | Chain = Tmp1.getValue(R: 1); |
4199 | } else if (IsVP) { |
4200 | Tmp1 = DAG.getNode(Opcode: Node->getOpcode(), DL: dl, VT: Node->getValueType(ResNo: 0), |
4201 | Ops: {Tmp1, Tmp2, Tmp3, Mask, EVL}, Flags: Node->getFlags()); |
4202 | } else { |
4203 | Tmp1 = DAG.getNode(Opcode: Node->getOpcode(), DL: dl, VT: Node->getValueType(ResNo: 0), N1: Tmp1, |
4204 | N2: Tmp2, N3: Tmp3, Flags: Node->getFlags()); |
4205 | } |
4206 | } |
4207 | |
4208 | // If we expanded the SETCC by inverting the condition code, then wrap |
4209 | // the existing SETCC in a NOT to restore the intended condition. |
4210 | if (NeedInvert) { |
4211 | if (!IsVP) |
4212 | Tmp1 = DAG.getLogicalNOT(DL: dl, Val: Tmp1, VT: Tmp1->getValueType(ResNo: 0)); |
4213 | else |
4214 | Tmp1 = |
4215 | DAG.getVPLogicalNOT(DL: dl, Val: Tmp1, Mask, EVL, VT: Tmp1->getValueType(ResNo: 0)); |
4216 | } |
4217 | |
4218 | Results.push_back(Elt: Tmp1); |
4219 | if (IsStrict) |
4220 | Results.push_back(Elt: Chain); |
4221 | |
4222 | break; |
4223 | } |
4224 | |
4225 | // FIXME: It seems Legalized is false iff CCCode is Legal. I don't |
4226 | // understand if this code is useful for strict nodes. |
4227 | assert(!IsStrict && "Don't know how to expand for strict nodes." ); |
4228 | |
4229 | // Otherwise, SETCC for the given comparison type must be completely |
4230 | // illegal; expand it into a SELECT_CC. |
4231 | // FIXME: This drops the mask/evl for VP_SETCC. |
4232 | EVT VT = Node->getValueType(ResNo: 0); |
4233 | EVT Tmp1VT = Tmp1.getValueType(); |
4234 | Tmp1 = DAG.getNode(Opcode: ISD::SELECT_CC, DL: dl, VT, N1: Tmp1, N2: Tmp2, |
4235 | N3: DAG.getBoolConstant(V: true, DL: dl, VT, OpVT: Tmp1VT), |
4236 | N4: DAG.getBoolConstant(V: false, DL: dl, VT, OpVT: Tmp1VT), N5: Tmp3); |
4237 | Tmp1->setFlags(Node->getFlags()); |
4238 | Results.push_back(Elt: Tmp1); |
4239 | break; |
4240 | } |
4241 | case ISD::SELECT_CC: { |
4242 | // TODO: need to add STRICT_SELECT_CC and STRICT_SELECT_CCS |
4243 | Tmp1 = Node->getOperand(Num: 0); // LHS |
4244 | Tmp2 = Node->getOperand(Num: 1); // RHS |
4245 | Tmp3 = Node->getOperand(Num: 2); // True |
4246 | Tmp4 = Node->getOperand(Num: 3); // False |
4247 | EVT VT = Node->getValueType(ResNo: 0); |
4248 | SDValue Chain; |
4249 | SDValue CC = Node->getOperand(Num: 4); |
4250 | ISD::CondCode CCOp = cast<CondCodeSDNode>(Val&: CC)->get(); |
4251 | |
4252 | if (TLI.isCondCodeLegalOrCustom(CC: CCOp, VT: Tmp1.getSimpleValueType())) { |
4253 | // If the condition code is legal, then we need to expand this |
4254 | // node using SETCC and SELECT. |
4255 | EVT CmpVT = Tmp1.getValueType(); |
4256 | assert(!TLI.isOperationExpand(ISD::SELECT, VT) && |
4257 | "Cannot expand ISD::SELECT_CC when ISD::SELECT also needs to be " |
4258 | "expanded." ); |
4259 | EVT CCVT = getSetCCResultType(VT: CmpVT); |
4260 | SDValue Cond = DAG.getNode(Opcode: ISD::SETCC, DL: dl, VT: CCVT, N1: Tmp1, N2: Tmp2, N3: CC, Flags: Node->getFlags()); |
4261 | Results.push_back( |
4262 | Elt: DAG.getSelect(DL: dl, VT, Cond, LHS: Tmp3, RHS: Tmp4, Flags: Node->getFlags())); |
4263 | break; |
4264 | } |
4265 | |
4266 | // SELECT_CC is legal, so the condition code must not be. |
4267 | bool Legalized = false; |
4268 | // Try to legalize by inverting the condition. This is for targets that |
4269 | // might support an ordered version of a condition, but not the unordered |
4270 | // version (or vice versa). |
4271 | ISD::CondCode InvCC = ISD::getSetCCInverse(Operation: CCOp, Type: Tmp1.getValueType()); |
4272 | if (TLI.isCondCodeLegalOrCustom(CC: InvCC, VT: Tmp1.getSimpleValueType())) { |
4273 | // Use the new condition code and swap true and false |
4274 | Legalized = true; |
4275 | Tmp1 = DAG.getSelectCC(DL: dl, LHS: Tmp1, RHS: Tmp2, True: Tmp4, False: Tmp3, Cond: InvCC); |
4276 | Tmp1->setFlags(Node->getFlags()); |
4277 | } else { |
4278 | // If The inverse is not legal, then try to swap the arguments using |
4279 | // the inverse condition code. |
4280 | ISD::CondCode SwapInvCC = ISD::getSetCCSwappedOperands(Operation: InvCC); |
4281 | if (TLI.isCondCodeLegalOrCustom(CC: SwapInvCC, VT: Tmp1.getSimpleValueType())) { |
4282 | // The swapped inverse condition is legal, so swap true and false, |
4283 | // lhs and rhs. |
4284 | Legalized = true; |
4285 | Tmp1 = DAG.getSelectCC(DL: dl, LHS: Tmp2, RHS: Tmp1, True: Tmp4, False: Tmp3, Cond: SwapInvCC); |
4286 | Tmp1->setFlags(Node->getFlags()); |
4287 | } |
4288 | } |
4289 | |
4290 | if (!Legalized) { |
4291 | Legalized = TLI.LegalizeSetCCCondCode( |
4292 | DAG, VT: getSetCCResultType(VT: Tmp1.getValueType()), LHS&: Tmp1, RHS&: Tmp2, CC, |
4293 | /*Mask*/ SDValue(), /*EVL*/ SDValue(), NeedInvert, dl, Chain); |
4294 | |
4295 | assert(Legalized && "Can't legalize SELECT_CC with legal condition!" ); |
4296 | |
4297 | // If we expanded the SETCC by inverting the condition code, then swap |
4298 | // the True/False operands to match. |
4299 | if (NeedInvert) |
4300 | std::swap(a&: Tmp3, b&: Tmp4); |
4301 | |
4302 | // If we expanded the SETCC by swapping LHS and RHS, or by inverting the |
4303 | // condition code, create a new SELECT_CC node. |
4304 | if (CC.getNode()) { |
4305 | Tmp1 = DAG.getNode(Opcode: ISD::SELECT_CC, DL: dl, VT: Node->getValueType(ResNo: 0), |
4306 | N1: Tmp1, N2: Tmp2, N3: Tmp3, N4: Tmp4, N5: CC); |
4307 | } else { |
4308 | Tmp2 = DAG.getConstant(Val: 0, DL: dl, VT: Tmp1.getValueType()); |
4309 | CC = DAG.getCondCode(Cond: ISD::SETNE); |
4310 | Tmp1 = DAG.getNode(Opcode: ISD::SELECT_CC, DL: dl, VT: Node->getValueType(ResNo: 0), N1: Tmp1, |
4311 | N2: Tmp2, N3: Tmp3, N4: Tmp4, N5: CC); |
4312 | } |
4313 | Tmp1->setFlags(Node->getFlags()); |
4314 | } |
4315 | Results.push_back(Elt: Tmp1); |
4316 | break; |
4317 | } |
4318 | case ISD::BR_CC: { |
4319 | // TODO: need to add STRICT_BR_CC and STRICT_BR_CCS |
4320 | SDValue Chain; |
4321 | Tmp1 = Node->getOperand(Num: 0); // Chain |
4322 | Tmp2 = Node->getOperand(Num: 2); // LHS |
4323 | Tmp3 = Node->getOperand(Num: 3); // RHS |
4324 | Tmp4 = Node->getOperand(Num: 1); // CC |
4325 | |
4326 | bool Legalized = TLI.LegalizeSetCCCondCode( |
4327 | DAG, VT: getSetCCResultType(VT: Tmp2.getValueType()), LHS&: Tmp2, RHS&: Tmp3, CC&: Tmp4, |
4328 | /*Mask*/ SDValue(), /*EVL*/ SDValue(), NeedInvert, dl, Chain); |
4329 | (void)Legalized; |
4330 | assert(Legalized && "Can't legalize BR_CC with legal condition!" ); |
4331 | |
4332 | // If we expanded the SETCC by swapping LHS and RHS, create a new BR_CC |
4333 | // node. |
4334 | if (Tmp4.getNode()) { |
4335 | assert(!NeedInvert && "Don't know how to invert BR_CC!" ); |
4336 | |
4337 | Tmp1 = DAG.getNode(Opcode: ISD::BR_CC, DL: dl, VT: Node->getValueType(ResNo: 0), N1: Tmp1, |
4338 | N2: Tmp4, N3: Tmp2, N4: Tmp3, N5: Node->getOperand(Num: 4)); |
4339 | } else { |
4340 | Tmp3 = DAG.getConstant(Val: 0, DL: dl, VT: Tmp2.getValueType()); |
4341 | Tmp4 = DAG.getCondCode(Cond: NeedInvert ? ISD::SETEQ : ISD::SETNE); |
4342 | Tmp1 = DAG.getNode(Opcode: ISD::BR_CC, DL: dl, VT: Node->getValueType(ResNo: 0), N1: Tmp1, N2: Tmp4, |
4343 | N3: Tmp2, N4: Tmp3, N5: Node->getOperand(Num: 4)); |
4344 | } |
4345 | Results.push_back(Elt: Tmp1); |
4346 | break; |
4347 | } |
4348 | case ISD::BUILD_VECTOR: |
4349 | Results.push_back(Elt: ExpandBUILD_VECTOR(Node)); |
4350 | break; |
4351 | case ISD::SPLAT_VECTOR: |
4352 | Results.push_back(Elt: ExpandSPLAT_VECTOR(Node)); |
4353 | break; |
4354 | case ISD::SRA: |
4355 | case ISD::SRL: |
4356 | case ISD::SHL: { |
4357 | // Scalarize vector SRA/SRL/SHL. |
4358 | EVT VT = Node->getValueType(ResNo: 0); |
4359 | assert(VT.isVector() && "Unable to legalize non-vector shift" ); |
4360 | assert(TLI.isTypeLegal(VT.getScalarType())&& "Element type must be legal" ); |
4361 | unsigned NumElem = VT.getVectorNumElements(); |
4362 | |
4363 | SmallVector<SDValue, 8> Scalars; |
4364 | for (unsigned Idx = 0; Idx < NumElem; Idx++) { |
4365 | SDValue Ex = |
4366 | DAG.getNode(Opcode: ISD::EXTRACT_VECTOR_ELT, DL: dl, VT: VT.getScalarType(), |
4367 | N1: Node->getOperand(Num: 0), N2: DAG.getVectorIdxConstant(Val: Idx, DL: dl)); |
4368 | SDValue Sh = |
4369 | DAG.getNode(Opcode: ISD::EXTRACT_VECTOR_ELT, DL: dl, VT: VT.getScalarType(), |
4370 | N1: Node->getOperand(Num: 1), N2: DAG.getVectorIdxConstant(Val: Idx, DL: dl)); |
4371 | Scalars.push_back(Elt: DAG.getNode(Opcode: Node->getOpcode(), DL: dl, |
4372 | VT: VT.getScalarType(), N1: Ex, N2: Sh)); |
4373 | } |
4374 | |
4375 | SDValue Result = DAG.getBuildVector(VT: Node->getValueType(ResNo: 0), DL: dl, Ops: Scalars); |
4376 | Results.push_back(Elt: Result); |
4377 | break; |
4378 | } |
4379 | case ISD::VECREDUCE_FADD: |
4380 | case ISD::VECREDUCE_FMUL: |
4381 | case ISD::VECREDUCE_ADD: |
4382 | case ISD::VECREDUCE_MUL: |
4383 | case ISD::VECREDUCE_AND: |
4384 | case ISD::VECREDUCE_OR: |
4385 | case ISD::VECREDUCE_XOR: |
4386 | case ISD::VECREDUCE_SMAX: |
4387 | case ISD::VECREDUCE_SMIN: |
4388 | case ISD::VECREDUCE_UMAX: |
4389 | case ISD::VECREDUCE_UMIN: |
4390 | case ISD::VECREDUCE_FMAX: |
4391 | case ISD::VECREDUCE_FMIN: |
4392 | case ISD::VECREDUCE_FMAXIMUM: |
4393 | case ISD::VECREDUCE_FMINIMUM: |
4394 | Results.push_back(Elt: TLI.expandVecReduce(Node, DAG)); |
4395 | break; |
4396 | case ISD::VP_CTTZ_ELTS: |
4397 | case ISD::VP_CTTZ_ELTS_ZERO_UNDEF: |
4398 | Results.push_back(Elt: TLI.expandVPCTTZElements(N: Node, DAG)); |
4399 | break; |
4400 | case ISD::CLEAR_CACHE: |
4401 | // The default expansion of llvm.clear_cache is simply a no-op for those |
4402 | // targets where it is not needed. |
4403 | Results.push_back(Elt: Node->getOperand(Num: 0)); |
4404 | break; |
4405 | case ISD::LRINT: |
4406 | case ISD::LLRINT: { |
4407 | SDValue Arg = Node->getOperand(Num: 0); |
4408 | EVT ArgVT = Arg.getValueType(); |
4409 | EVT ResVT = Node->getValueType(ResNo: 0); |
4410 | SDLoc dl(Node); |
4411 | SDValue RoundNode = DAG.getNode(Opcode: ISD::FRINT, DL: dl, VT: ArgVT, Operand: Arg); |
4412 | Results.push_back(Elt: DAG.getNode(Opcode: ISD::FP_TO_SINT, DL: dl, VT: ResVT, Operand: RoundNode)); |
4413 | break; |
4414 | } |
4415 | case ISD::ADDRSPACECAST: |
4416 | Results.push_back(Elt: DAG.UnrollVectorOp(N: Node)); |
4417 | break; |
4418 | case ISD::GLOBAL_OFFSET_TABLE: |
4419 | case ISD::GlobalAddress: |
4420 | case ISD::GlobalTLSAddress: |
4421 | case ISD::ExternalSymbol: |
4422 | case ISD::ConstantPool: |
4423 | case ISD::JumpTable: |
4424 | case ISD::INTRINSIC_W_CHAIN: |
4425 | case ISD::INTRINSIC_WO_CHAIN: |
4426 | case ISD::INTRINSIC_VOID: |
4427 | // FIXME: Custom lowering for these operations shouldn't return null! |
4428 | // Return true so that we don't call ConvertNodeToLibcall which also won't |
4429 | // do anything. |
4430 | return true; |
4431 | } |
4432 | |
4433 | if (!TLI.isStrictFPEnabled() && Results.empty() && Node->isStrictFPOpcode()) { |
4434 | // FIXME: We were asked to expand a strict floating-point operation, |
4435 | // but there is currently no expansion implemented that would preserve |
4436 | // the "strict" properties. For now, we just fall back to the non-strict |
4437 | // version if that is legal on the target. The actual mutation of the |
4438 | // operation will happen in SelectionDAGISel::DoInstructionSelection. |
4439 | switch (Node->getOpcode()) { |
4440 | default: |
4441 | if (TLI.getStrictFPOperationAction(Op: Node->getOpcode(), |
4442 | VT: Node->getValueType(ResNo: 0)) |
4443 | == TargetLowering::Legal) |
4444 | return true; |
4445 | break; |
4446 | case ISD::STRICT_FSUB: { |
4447 | if (TLI.getStrictFPOperationAction( |
4448 | Op: ISD::STRICT_FSUB, VT: Node->getValueType(ResNo: 0)) == TargetLowering::Legal) |
4449 | return true; |
4450 | if (TLI.getStrictFPOperationAction( |
4451 | Op: ISD::STRICT_FADD, VT: Node->getValueType(ResNo: 0)) != TargetLowering::Legal) |
4452 | break; |
4453 | |
4454 | EVT VT = Node->getValueType(ResNo: 0); |
4455 | const SDNodeFlags Flags = Node->getFlags(); |
4456 | SDValue Neg = DAG.getNode(Opcode: ISD::FNEG, DL: dl, VT, Operand: Node->getOperand(Num: 2), Flags); |
4457 | SDValue Fadd = DAG.getNode(Opcode: ISD::STRICT_FADD, DL: dl, VTList: Node->getVTList(), |
4458 | Ops: {Node->getOperand(Num: 0), Node->getOperand(Num: 1), Neg}, |
4459 | Flags); |
4460 | |
4461 | Results.push_back(Elt: Fadd); |
4462 | Results.push_back(Elt: Fadd.getValue(R: 1)); |
4463 | break; |
4464 | } |
4465 | case ISD::STRICT_SINT_TO_FP: |
4466 | case ISD::STRICT_UINT_TO_FP: |
4467 | case ISD::STRICT_LRINT: |
4468 | case ISD::STRICT_LLRINT: |
4469 | case ISD::STRICT_LROUND: |
4470 | case ISD::STRICT_LLROUND: |
4471 | // These are registered by the operand type instead of the value |
4472 | // type. Reflect that here. |
4473 | if (TLI.getStrictFPOperationAction(Op: Node->getOpcode(), |
4474 | VT: Node->getOperand(Num: 1).getValueType()) |
4475 | == TargetLowering::Legal) |
4476 | return true; |
4477 | break; |
4478 | } |
4479 | } |
4480 | |
4481 | // Replace the original node with the legalized result. |
4482 | if (Results.empty()) { |
4483 | LLVM_DEBUG(dbgs() << "Cannot expand node\n" ); |
4484 | return false; |
4485 | } |
4486 | |
4487 | LLVM_DEBUG(dbgs() << "Successfully expanded node\n" ); |
4488 | ReplaceNode(Old: Node, New: Results.data()); |
4489 | return true; |
4490 | } |
4491 | |
4492 | void SelectionDAGLegalize::ConvertNodeToLibcall(SDNode *Node) { |
4493 | LLVM_DEBUG(dbgs() << "Trying to convert node to libcall\n" ); |
4494 | SmallVector<SDValue, 8> Results; |
4495 | SDLoc dl(Node); |
4496 | TargetLowering::MakeLibCallOptions CallOptions; |
4497 | CallOptions.IsPostTypeLegalization = true; |
4498 | // FIXME: Check flags on the node to see if we can use a finite call. |
4499 | unsigned Opc = Node->getOpcode(); |
4500 | switch (Opc) { |
4501 | case ISD::ATOMIC_FENCE: { |
4502 | // If the target didn't lower this, lower it to '__sync_synchronize()' call |
4503 | // FIXME: handle "fence singlethread" more efficiently. |
4504 | TargetLowering::ArgListTy Args; |
4505 | |
4506 | TargetLowering::CallLoweringInfo CLI(DAG); |
4507 | CLI.setDebugLoc(dl) |
4508 | .setChain(Node->getOperand(Num: 0)) |
4509 | .setLibCallee( |
4510 | CC: CallingConv::C, ResultType: Type::getVoidTy(C&: *DAG.getContext()), |
4511 | Target: DAG.getExternalSymbol(Sym: "__sync_synchronize" , |
4512 | VT: TLI.getPointerTy(DL: DAG.getDataLayout())), |
4513 | ArgsList: std::move(Args)); |
4514 | |
4515 | std::pair<SDValue, SDValue> CallResult = TLI.LowerCallTo(CLI); |
4516 | |
4517 | Results.push_back(Elt: CallResult.second); |
4518 | break; |
4519 | } |
4520 | // By default, atomic intrinsics are marked Legal and lowered. Targets |
4521 | // which don't support them directly, however, may want libcalls, in which |
4522 | // case they mark them Expand, and we get here. |
4523 | case ISD::ATOMIC_SWAP: |
4524 | case ISD::ATOMIC_LOAD_ADD: |
4525 | case ISD::ATOMIC_LOAD_SUB: |
4526 | case ISD::ATOMIC_LOAD_AND: |
4527 | case ISD::ATOMIC_LOAD_CLR: |
4528 | case ISD::ATOMIC_LOAD_OR: |
4529 | case ISD::ATOMIC_LOAD_XOR: |
4530 | case ISD::ATOMIC_LOAD_NAND: |
4531 | case ISD::ATOMIC_LOAD_MIN: |
4532 | case ISD::ATOMIC_LOAD_MAX: |
4533 | case ISD::ATOMIC_LOAD_UMIN: |
4534 | case ISD::ATOMIC_LOAD_UMAX: |
4535 | case ISD::ATOMIC_CMP_SWAP: { |
4536 | MVT VT = cast<AtomicSDNode>(Val: Node)->getMemoryVT().getSimpleVT(); |
4537 | AtomicOrdering Order = cast<AtomicSDNode>(Val: Node)->getMergedOrdering(); |
4538 | RTLIB::Libcall LC = RTLIB::getOUTLINE_ATOMIC(Opc, Order, VT); |
4539 | EVT RetVT = Node->getValueType(ResNo: 0); |
4540 | SmallVector<SDValue, 4> Ops; |
4541 | if (TLI.getLibcallName(Call: LC)) { |
4542 | // If outline atomic available, prepare its arguments and expand. |
4543 | Ops.append(in_start: Node->op_begin() + 2, in_end: Node->op_end()); |
4544 | Ops.push_back(Elt: Node->getOperand(Num: 1)); |
4545 | |
4546 | } else { |
4547 | LC = RTLIB::getSYNC(Opc, VT); |
4548 | assert(LC != RTLIB::UNKNOWN_LIBCALL && |
4549 | "Unexpected atomic op or value type!" ); |
4550 | // Arguments for expansion to sync libcall |
4551 | Ops.append(in_start: Node->op_begin() + 1, in_end: Node->op_end()); |
4552 | } |
4553 | std::pair<SDValue, SDValue> Tmp = TLI.makeLibCall(DAG, LC, RetVT, |
4554 | Ops, CallOptions, |
4555 | dl: SDLoc(Node), |
4556 | Chain: Node->getOperand(Num: 0)); |
4557 | Results.push_back(Elt: Tmp.first); |
4558 | Results.push_back(Elt: Tmp.second); |
4559 | break; |
4560 | } |
4561 | case ISD::TRAP: { |
4562 | // If this operation is not supported, lower it to 'abort()' call |
4563 | TargetLowering::ArgListTy Args; |
4564 | TargetLowering::CallLoweringInfo CLI(DAG); |
4565 | CLI.setDebugLoc(dl) |
4566 | .setChain(Node->getOperand(Num: 0)) |
4567 | .setLibCallee(CC: CallingConv::C, ResultType: Type::getVoidTy(C&: *DAG.getContext()), |
4568 | Target: DAG.getExternalSymbol( |
4569 | Sym: "abort" , VT: TLI.getPointerTy(DL: DAG.getDataLayout())), |
4570 | ArgsList: std::move(Args)); |
4571 | std::pair<SDValue, SDValue> CallResult = TLI.LowerCallTo(CLI); |
4572 | |
4573 | Results.push_back(Elt: CallResult.second); |
4574 | break; |
4575 | } |
4576 | case ISD::CLEAR_CACHE: { |
4577 | SDValue InputChain = Node->getOperand(Num: 0); |
4578 | SDValue StartVal = Node->getOperand(Num: 1); |
4579 | SDValue EndVal = Node->getOperand(Num: 2); |
4580 | std::pair<SDValue, SDValue> Tmp = TLI.makeLibCall( |
4581 | DAG, LC: RTLIB::CLEAR_CACHE, RetVT: MVT::isVoid, Ops: {StartVal, EndVal}, CallOptions, |
4582 | dl: SDLoc(Node), Chain: InputChain); |
4583 | Results.push_back(Elt: Tmp.second); |
4584 | break; |
4585 | } |
4586 | case ISD::FMINNUM: |
4587 | case ISD::STRICT_FMINNUM: |
4588 | ExpandFPLibCall(Node, Call_F32: RTLIB::FMIN_F32, Call_F64: RTLIB::FMIN_F64, |
4589 | Call_F80: RTLIB::FMIN_F80, Call_F128: RTLIB::FMIN_F128, |
4590 | Call_PPCF128: RTLIB::FMIN_PPCF128, Results); |
4591 | break; |
4592 | // FIXME: We do not have libcalls for FMAXIMUM and FMINIMUM. So, we cannot use |
4593 | // libcall legalization for these nodes, but there is no default expasion for |
4594 | // these nodes either (see PR63267 for example). |
4595 | case ISD::FMAXNUM: |
4596 | case ISD::STRICT_FMAXNUM: |
4597 | ExpandFPLibCall(Node, Call_F32: RTLIB::FMAX_F32, Call_F64: RTLIB::FMAX_F64, |
4598 | Call_F80: RTLIB::FMAX_F80, Call_F128: RTLIB::FMAX_F128, |
4599 | Call_PPCF128: RTLIB::FMAX_PPCF128, Results); |
4600 | break; |
4601 | case ISD::FMINIMUMNUM: |
4602 | ExpandFPLibCall(Node, Call_F32: RTLIB::FMINIMUM_NUM_F32, Call_F64: RTLIB::FMINIMUM_NUM_F64, |
4603 | Call_F80: RTLIB::FMINIMUM_NUM_F80, Call_F128: RTLIB::FMINIMUM_NUM_F128, |
4604 | Call_PPCF128: RTLIB::FMINIMUM_NUM_PPCF128, Results); |
4605 | break; |
4606 | case ISD::FMAXIMUMNUM: |
4607 | ExpandFPLibCall(Node, Call_F32: RTLIB::FMAXIMUM_NUM_F32, Call_F64: RTLIB::FMAXIMUM_NUM_F64, |
4608 | Call_F80: RTLIB::FMAXIMUM_NUM_F80, Call_F128: RTLIB::FMAXIMUM_NUM_F128, |
4609 | Call_PPCF128: RTLIB::FMAXIMUM_NUM_PPCF128, Results); |
4610 | break; |
4611 | case ISD::FSQRT: |
4612 | case ISD::STRICT_FSQRT: |
4613 | ExpandFPLibCall(Node, Call_F32: RTLIB::SQRT_F32, Call_F64: RTLIB::SQRT_F64, |
4614 | Call_F80: RTLIB::SQRT_F80, Call_F128: RTLIB::SQRT_F128, |
4615 | Call_PPCF128: RTLIB::SQRT_PPCF128, Results); |
4616 | break; |
4617 | case ISD::FCBRT: |
4618 | ExpandFPLibCall(Node, Call_F32: RTLIB::CBRT_F32, Call_F64: RTLIB::CBRT_F64, |
4619 | Call_F80: RTLIB::CBRT_F80, Call_F128: RTLIB::CBRT_F128, |
4620 | Call_PPCF128: RTLIB::CBRT_PPCF128, Results); |
4621 | break; |
4622 | case ISD::FSIN: |
4623 | case ISD::STRICT_FSIN: |
4624 | ExpandFPLibCall(Node, Call_F32: RTLIB::SIN_F32, Call_F64: RTLIB::SIN_F64, |
4625 | Call_F80: RTLIB::SIN_F80, Call_F128: RTLIB::SIN_F128, |
4626 | Call_PPCF128: RTLIB::SIN_PPCF128, Results); |
4627 | break; |
4628 | case ISD::FCOS: |
4629 | case ISD::STRICT_FCOS: |
4630 | ExpandFPLibCall(Node, Call_F32: RTLIB::COS_F32, Call_F64: RTLIB::COS_F64, |
4631 | Call_F80: RTLIB::COS_F80, Call_F128: RTLIB::COS_F128, |
4632 | Call_PPCF128: RTLIB::COS_PPCF128, Results); |
4633 | break; |
4634 | case ISD::FTAN: |
4635 | case ISD::STRICT_FTAN: |
4636 | ExpandFPLibCall(Node, Call_F32: RTLIB::TAN_F32, Call_F64: RTLIB::TAN_F64, Call_F80: RTLIB::TAN_F80, |
4637 | Call_F128: RTLIB::TAN_F128, Call_PPCF128: RTLIB::TAN_PPCF128, Results); |
4638 | break; |
4639 | case ISD::FASIN: |
4640 | case ISD::STRICT_FASIN: |
4641 | ExpandFPLibCall(Node, Call_F32: RTLIB::ASIN_F32, Call_F64: RTLIB::ASIN_F64, Call_F80: RTLIB::ASIN_F80, |
4642 | Call_F128: RTLIB::ASIN_F128, Call_PPCF128: RTLIB::ASIN_PPCF128, Results); |
4643 | break; |
4644 | case ISD::FACOS: |
4645 | case ISD::STRICT_FACOS: |
4646 | ExpandFPLibCall(Node, Call_F32: RTLIB::ACOS_F32, Call_F64: RTLIB::ACOS_F64, Call_F80: RTLIB::ACOS_F80, |
4647 | Call_F128: RTLIB::ACOS_F128, Call_PPCF128: RTLIB::ACOS_PPCF128, Results); |
4648 | break; |
4649 | case ISD::FATAN: |
4650 | case ISD::STRICT_FATAN: |
4651 | ExpandFPLibCall(Node, Call_F32: RTLIB::ATAN_F32, Call_F64: RTLIB::ATAN_F64, Call_F80: RTLIB::ATAN_F80, |
4652 | Call_F128: RTLIB::ATAN_F128, Call_PPCF128: RTLIB::ATAN_PPCF128, Results); |
4653 | break; |
4654 | case ISD::FATAN2: |
4655 | case ISD::STRICT_FATAN2: |
4656 | ExpandFPLibCall(Node, Call_F32: RTLIB::ATAN2_F32, Call_F64: RTLIB::ATAN2_F64, Call_F80: RTLIB::ATAN2_F80, |
4657 | Call_F128: RTLIB::ATAN2_F128, Call_PPCF128: RTLIB::ATAN2_PPCF128, Results); |
4658 | break; |
4659 | case ISD::FSINH: |
4660 | case ISD::STRICT_FSINH: |
4661 | ExpandFPLibCall(Node, Call_F32: RTLIB::SINH_F32, Call_F64: RTLIB::SINH_F64, Call_F80: RTLIB::SINH_F80, |
4662 | Call_F128: RTLIB::SINH_F128, Call_PPCF128: RTLIB::SINH_PPCF128, Results); |
4663 | break; |
4664 | case ISD::FCOSH: |
4665 | case ISD::STRICT_FCOSH: |
4666 | ExpandFPLibCall(Node, Call_F32: RTLIB::COSH_F32, Call_F64: RTLIB::COSH_F64, Call_F80: RTLIB::COSH_F80, |
4667 | Call_F128: RTLIB::COSH_F128, Call_PPCF128: RTLIB::COSH_PPCF128, Results); |
4668 | break; |
4669 | case ISD::FTANH: |
4670 | case ISD::STRICT_FTANH: |
4671 | ExpandFPLibCall(Node, Call_F32: RTLIB::TANH_F32, Call_F64: RTLIB::TANH_F64, Call_F80: RTLIB::TANH_F80, |
4672 | Call_F128: RTLIB::TANH_F128, Call_PPCF128: RTLIB::TANH_PPCF128, Results); |
4673 | break; |
4674 | case ISD::FSINCOS: |
4675 | case ISD::FSINCOSPI: { |
4676 | EVT VT = Node->getValueType(ResNo: 0); |
4677 | RTLIB::Libcall LC = Node->getOpcode() == ISD::FSINCOS |
4678 | ? RTLIB::getSINCOS(RetVT: VT) |
4679 | : RTLIB::getSINCOSPI(RetVT: VT); |
4680 | bool Expanded = DAG.expandMultipleResultFPLibCall(LC, Node, Results); |
4681 | if (!Expanded) |
4682 | llvm_unreachable("Expected scalar FSINCOS[PI] to expand to libcall!" ); |
4683 | break; |
4684 | } |
4685 | case ISD::FLOG: |
4686 | case ISD::STRICT_FLOG: |
4687 | ExpandFPLibCall(Node, Call_F32: RTLIB::LOG_F32, Call_F64: RTLIB::LOG_F64, Call_F80: RTLIB::LOG_F80, |
4688 | Call_F128: RTLIB::LOG_F128, Call_PPCF128: RTLIB::LOG_PPCF128, Results); |
4689 | break; |
4690 | case ISD::FLOG2: |
4691 | case ISD::STRICT_FLOG2: |
4692 | ExpandFPLibCall(Node, Call_F32: RTLIB::LOG2_F32, Call_F64: RTLIB::LOG2_F64, Call_F80: RTLIB::LOG2_F80, |
4693 | Call_F128: RTLIB::LOG2_F128, Call_PPCF128: RTLIB::LOG2_PPCF128, Results); |
4694 | break; |
4695 | case ISD::FLOG10: |
4696 | case ISD::STRICT_FLOG10: |
4697 | ExpandFPLibCall(Node, Call_F32: RTLIB::LOG10_F32, Call_F64: RTLIB::LOG10_F64, Call_F80: RTLIB::LOG10_F80, |
4698 | Call_F128: RTLIB::LOG10_F128, Call_PPCF128: RTLIB::LOG10_PPCF128, Results); |
4699 | break; |
4700 | case ISD::FEXP: |
4701 | case ISD::STRICT_FEXP: |
4702 | ExpandFPLibCall(Node, Call_F32: RTLIB::EXP_F32, Call_F64: RTLIB::EXP_F64, Call_F80: RTLIB::EXP_F80, |
4703 | Call_F128: RTLIB::EXP_F128, Call_PPCF128: RTLIB::EXP_PPCF128, Results); |
4704 | break; |
4705 | case ISD::FEXP2: |
4706 | case ISD::STRICT_FEXP2: |
4707 | ExpandFPLibCall(Node, Call_F32: RTLIB::EXP2_F32, Call_F64: RTLIB::EXP2_F64, Call_F80: RTLIB::EXP2_F80, |
4708 | Call_F128: RTLIB::EXP2_F128, Call_PPCF128: RTLIB::EXP2_PPCF128, Results); |
4709 | break; |
4710 | case ISD::FEXP10: |
4711 | ExpandFPLibCall(Node, Call_F32: RTLIB::EXP10_F32, Call_F64: RTLIB::EXP10_F64, Call_F80: RTLIB::EXP10_F80, |
4712 | Call_F128: RTLIB::EXP10_F128, Call_PPCF128: RTLIB::EXP10_PPCF128, Results); |
4713 | break; |
4714 | case ISD::FTRUNC: |
4715 | case ISD::STRICT_FTRUNC: |
4716 | ExpandFPLibCall(Node, Call_F32: RTLIB::TRUNC_F32, Call_F64: RTLIB::TRUNC_F64, |
4717 | Call_F80: RTLIB::TRUNC_F80, Call_F128: RTLIB::TRUNC_F128, |
4718 | Call_PPCF128: RTLIB::TRUNC_PPCF128, Results); |
4719 | break; |
4720 | case ISD::FFLOOR: |
4721 | case ISD::STRICT_FFLOOR: |
4722 | ExpandFPLibCall(Node, Call_F32: RTLIB::FLOOR_F32, Call_F64: RTLIB::FLOOR_F64, |
4723 | Call_F80: RTLIB::FLOOR_F80, Call_F128: RTLIB::FLOOR_F128, |
4724 | Call_PPCF128: RTLIB::FLOOR_PPCF128, Results); |
4725 | break; |
4726 | case ISD::FCEIL: |
4727 | case ISD::STRICT_FCEIL: |
4728 | ExpandFPLibCall(Node, Call_F32: RTLIB::CEIL_F32, Call_F64: RTLIB::CEIL_F64, |
4729 | Call_F80: RTLIB::CEIL_F80, Call_F128: RTLIB::CEIL_F128, |
4730 | Call_PPCF128: RTLIB::CEIL_PPCF128, Results); |
4731 | break; |
4732 | case ISD::FRINT: |
4733 | case ISD::STRICT_FRINT: |
4734 | ExpandFPLibCall(Node, Call_F32: RTLIB::RINT_F32, Call_F64: RTLIB::RINT_F64, |
4735 | Call_F80: RTLIB::RINT_F80, Call_F128: RTLIB::RINT_F128, |
4736 | Call_PPCF128: RTLIB::RINT_PPCF128, Results); |
4737 | break; |
4738 | case ISD::FNEARBYINT: |
4739 | case ISD::STRICT_FNEARBYINT: |
4740 | ExpandFPLibCall(Node, Call_F32: RTLIB::NEARBYINT_F32, |
4741 | Call_F64: RTLIB::NEARBYINT_F64, |
4742 | Call_F80: RTLIB::NEARBYINT_F80, |
4743 | Call_F128: RTLIB::NEARBYINT_F128, |
4744 | Call_PPCF128: RTLIB::NEARBYINT_PPCF128, Results); |
4745 | break; |
4746 | case ISD::FROUND: |
4747 | case ISD::STRICT_FROUND: |
4748 | ExpandFPLibCall(Node, Call_F32: RTLIB::ROUND_F32, |
4749 | Call_F64: RTLIB::ROUND_F64, |
4750 | Call_F80: RTLIB::ROUND_F80, |
4751 | Call_F128: RTLIB::ROUND_F128, |
4752 | Call_PPCF128: RTLIB::ROUND_PPCF128, Results); |
4753 | break; |
4754 | case ISD::FROUNDEVEN: |
4755 | case ISD::STRICT_FROUNDEVEN: |
4756 | ExpandFPLibCall(Node, Call_F32: RTLIB::ROUNDEVEN_F32, |
4757 | Call_F64: RTLIB::ROUNDEVEN_F64, |
4758 | Call_F80: RTLIB::ROUNDEVEN_F80, |
4759 | Call_F128: RTLIB::ROUNDEVEN_F128, |
4760 | Call_PPCF128: RTLIB::ROUNDEVEN_PPCF128, Results); |
4761 | break; |
4762 | case ISD::FLDEXP: |
4763 | case ISD::STRICT_FLDEXP: |
4764 | ExpandFPLibCall(Node, Call_F32: RTLIB::LDEXP_F32, Call_F64: RTLIB::LDEXP_F64, Call_F80: RTLIB::LDEXP_F80, |
4765 | Call_F128: RTLIB::LDEXP_F128, Call_PPCF128: RTLIB::LDEXP_PPCF128, Results); |
4766 | break; |
4767 | case ISD::FMODF: |
4768 | case ISD::FFREXP: { |
4769 | EVT VT = Node->getValueType(ResNo: 0); |
4770 | RTLIB::Libcall LC = Node->getOpcode() == ISD::FMODF ? RTLIB::getMODF(RetVT: VT) |
4771 | : RTLIB::getFREXP(RetVT: VT); |
4772 | bool Expanded = DAG.expandMultipleResultFPLibCall(LC, Node, Results, |
4773 | /*CallRetResNo=*/0); |
4774 | if (!Expanded) |
4775 | llvm_unreachable("Expected scalar FFREXP/FMODF to expand to libcall!" ); |
4776 | break; |
4777 | } |
4778 | case ISD::FPOWI: |
4779 | case ISD::STRICT_FPOWI: { |
4780 | RTLIB::Libcall LC = RTLIB::getPOWI(RetVT: Node->getSimpleValueType(ResNo: 0)); |
4781 | assert(LC != RTLIB::UNKNOWN_LIBCALL && "Unexpected fpowi." ); |
4782 | if (!TLI.getLibcallName(Call: LC)) { |
4783 | // Some targets don't have a powi libcall; use pow instead. |
4784 | if (Node->isStrictFPOpcode()) { |
4785 | SDValue Exponent = |
4786 | DAG.getNode(Opcode: ISD::STRICT_SINT_TO_FP, DL: SDLoc(Node), |
4787 | ResultTys: {Node->getValueType(ResNo: 0), Node->getValueType(ResNo: 1)}, |
4788 | Ops: {Node->getOperand(Num: 0), Node->getOperand(Num: 2)}); |
4789 | SDValue FPOW = |
4790 | DAG.getNode(Opcode: ISD::STRICT_FPOW, DL: SDLoc(Node), |
4791 | ResultTys: {Node->getValueType(ResNo: 0), Node->getValueType(ResNo: 1)}, |
4792 | Ops: {Exponent.getValue(R: 1), Node->getOperand(Num: 1), Exponent}); |
4793 | Results.push_back(Elt: FPOW); |
4794 | Results.push_back(Elt: FPOW.getValue(R: 1)); |
4795 | } else { |
4796 | SDValue Exponent = |
4797 | DAG.getNode(Opcode: ISD::SINT_TO_FP, DL: SDLoc(Node), VT: Node->getValueType(ResNo: 0), |
4798 | Operand: Node->getOperand(Num: 1)); |
4799 | Results.push_back(Elt: DAG.getNode(Opcode: ISD::FPOW, DL: SDLoc(Node), |
4800 | VT: Node->getValueType(ResNo: 0), |
4801 | N1: Node->getOperand(Num: 0), N2: Exponent)); |
4802 | } |
4803 | break; |
4804 | } |
4805 | unsigned Offset = Node->isStrictFPOpcode() ? 1 : 0; |
4806 | bool ExponentHasSizeOfInt = |
4807 | DAG.getLibInfo().getIntSize() == |
4808 | Node->getOperand(Num: 1 + Offset).getValueType().getSizeInBits(); |
4809 | if (!ExponentHasSizeOfInt) { |
4810 | // If the exponent does not match with sizeof(int) a libcall to |
4811 | // RTLIB::POWI would use the wrong type for the argument. |
4812 | DAG.getContext()->emitError(ErrorStr: "POWI exponent does not match sizeof(int)" ); |
4813 | Results.push_back(Elt: DAG.getUNDEF(VT: Node->getValueType(ResNo: 0))); |
4814 | break; |
4815 | } |
4816 | ExpandFPLibCall(Node, LC, Results); |
4817 | break; |
4818 | } |
4819 | case ISD::FPOW: |
4820 | case ISD::STRICT_FPOW: |
4821 | ExpandFPLibCall(Node, Call_F32: RTLIB::POW_F32, Call_F64: RTLIB::POW_F64, Call_F80: RTLIB::POW_F80, |
4822 | Call_F128: RTLIB::POW_F128, Call_PPCF128: RTLIB::POW_PPCF128, Results); |
4823 | break; |
4824 | case ISD::LROUND: |
4825 | case ISD::STRICT_LROUND: |
4826 | ExpandArgFPLibCall(Node, Call_F32: RTLIB::LROUND_F32, |
4827 | Call_F64: RTLIB::LROUND_F64, Call_F80: RTLIB::LROUND_F80, |
4828 | Call_F128: RTLIB::LROUND_F128, |
4829 | Call_PPCF128: RTLIB::LROUND_PPCF128, Results); |
4830 | break; |
4831 | case ISD::LLROUND: |
4832 | case ISD::STRICT_LLROUND: |
4833 | ExpandArgFPLibCall(Node, Call_F32: RTLIB::LLROUND_F32, |
4834 | Call_F64: RTLIB::LLROUND_F64, Call_F80: RTLIB::LLROUND_F80, |
4835 | Call_F128: RTLIB::LLROUND_F128, |
4836 | Call_PPCF128: RTLIB::LLROUND_PPCF128, Results); |
4837 | break; |
4838 | case ISD::LRINT: |
4839 | case ISD::STRICT_LRINT: |
4840 | ExpandArgFPLibCall(Node, Call_F32: RTLIB::LRINT_F32, |
4841 | Call_F64: RTLIB::LRINT_F64, Call_F80: RTLIB::LRINT_F80, |
4842 | Call_F128: RTLIB::LRINT_F128, |
4843 | Call_PPCF128: RTLIB::LRINT_PPCF128, Results); |
4844 | break; |
4845 | case ISD::LLRINT: |
4846 | case ISD::STRICT_LLRINT: |
4847 | ExpandArgFPLibCall(Node, Call_F32: RTLIB::LLRINT_F32, |
4848 | Call_F64: RTLIB::LLRINT_F64, Call_F80: RTLIB::LLRINT_F80, |
4849 | Call_F128: RTLIB::LLRINT_F128, |
4850 | Call_PPCF128: RTLIB::LLRINT_PPCF128, Results); |
4851 | break; |
4852 | case ISD::FDIV: |
4853 | case ISD::STRICT_FDIV: |
4854 | ExpandFPLibCall(Node, Call_F32: RTLIB::DIV_F32, Call_F64: RTLIB::DIV_F64, |
4855 | Call_F80: RTLIB::DIV_F80, Call_F128: RTLIB::DIV_F128, |
4856 | Call_PPCF128: RTLIB::DIV_PPCF128, Results); |
4857 | break; |
4858 | case ISD::FREM: |
4859 | case ISD::STRICT_FREM: |
4860 | ExpandFPLibCall(Node, Call_F32: RTLIB::REM_F32, Call_F64: RTLIB::REM_F64, |
4861 | Call_F80: RTLIB::REM_F80, Call_F128: RTLIB::REM_F128, |
4862 | Call_PPCF128: RTLIB::REM_PPCF128, Results); |
4863 | break; |
4864 | case ISD::FMA: |
4865 | case ISD::STRICT_FMA: |
4866 | ExpandFPLibCall(Node, Call_F32: RTLIB::FMA_F32, Call_F64: RTLIB::FMA_F64, |
4867 | Call_F80: RTLIB::FMA_F80, Call_F128: RTLIB::FMA_F128, |
4868 | Call_PPCF128: RTLIB::FMA_PPCF128, Results); |
4869 | break; |
4870 | case ISD::FADD: |
4871 | case ISD::STRICT_FADD: |
4872 | ExpandFPLibCall(Node, Call_F32: RTLIB::ADD_F32, Call_F64: RTLIB::ADD_F64, |
4873 | Call_F80: RTLIB::ADD_F80, Call_F128: RTLIB::ADD_F128, |
4874 | Call_PPCF128: RTLIB::ADD_PPCF128, Results); |
4875 | break; |
4876 | case ISD::FMUL: |
4877 | case ISD::STRICT_FMUL: |
4878 | ExpandFPLibCall(Node, Call_F32: RTLIB::MUL_F32, Call_F64: RTLIB::MUL_F64, |
4879 | Call_F80: RTLIB::MUL_F80, Call_F128: RTLIB::MUL_F128, |
4880 | Call_PPCF128: RTLIB::MUL_PPCF128, Results); |
4881 | break; |
4882 | case ISD::FP16_TO_FP: |
4883 | if (Node->getValueType(ResNo: 0) == MVT::f32) { |
4884 | Results.push_back(Elt: ExpandLibCall(LC: RTLIB::FPEXT_F16_F32, Node, isSigned: false).first); |
4885 | } |
4886 | break; |
4887 | case ISD::STRICT_BF16_TO_FP: |
4888 | if (Node->getValueType(ResNo: 0) == MVT::f32) { |
4889 | std::pair<SDValue, SDValue> Tmp = TLI.makeLibCall( |
4890 | DAG, LC: RTLIB::FPEXT_BF16_F32, RetVT: MVT::f32, Ops: Node->getOperand(Num: 1), |
4891 | CallOptions, dl: SDLoc(Node), Chain: Node->getOperand(Num: 0)); |
4892 | Results.push_back(Elt: Tmp.first); |
4893 | Results.push_back(Elt: Tmp.second); |
4894 | } |
4895 | break; |
4896 | case ISD::STRICT_FP16_TO_FP: { |
4897 | if (Node->getValueType(ResNo: 0) == MVT::f32) { |
4898 | std::pair<SDValue, SDValue> Tmp = TLI.makeLibCall( |
4899 | DAG, LC: RTLIB::FPEXT_F16_F32, RetVT: MVT::f32, Ops: Node->getOperand(Num: 1), CallOptions, |
4900 | dl: SDLoc(Node), Chain: Node->getOperand(Num: 0)); |
4901 | Results.push_back(Elt: Tmp.first); |
4902 | Results.push_back(Elt: Tmp.second); |
4903 | } |
4904 | break; |
4905 | } |
4906 | case ISD::FP_TO_FP16: { |
4907 | RTLIB::Libcall LC = |
4908 | RTLIB::getFPROUND(OpVT: Node->getOperand(Num: 0).getValueType(), RetVT: MVT::f16); |
4909 | assert(LC != RTLIB::UNKNOWN_LIBCALL && "Unable to expand fp_to_fp16" ); |
4910 | Results.push_back(Elt: ExpandLibCall(LC, Node, isSigned: false).first); |
4911 | break; |
4912 | } |
4913 | case ISD::FP_TO_BF16: { |
4914 | RTLIB::Libcall LC = |
4915 | RTLIB::getFPROUND(OpVT: Node->getOperand(Num: 0).getValueType(), RetVT: MVT::bf16); |
4916 | assert(LC != RTLIB::UNKNOWN_LIBCALL && "Unable to expand fp_to_bf16" ); |
4917 | Results.push_back(Elt: ExpandLibCall(LC, Node, isSigned: false).first); |
4918 | break; |
4919 | } |
4920 | case ISD::STRICT_SINT_TO_FP: |
4921 | case ISD::STRICT_UINT_TO_FP: |
4922 | case ISD::SINT_TO_FP: |
4923 | case ISD::UINT_TO_FP: { |
4924 | // TODO - Common the code with DAGTypeLegalizer::SoftenFloatRes_XINT_TO_FP |
4925 | bool IsStrict = Node->isStrictFPOpcode(); |
4926 | bool Signed = Node->getOpcode() == ISD::SINT_TO_FP || |
4927 | Node->getOpcode() == ISD::STRICT_SINT_TO_FP; |
4928 | EVT SVT = Node->getOperand(Num: IsStrict ? 1 : 0).getValueType(); |
4929 | EVT RVT = Node->getValueType(ResNo: 0); |
4930 | EVT NVT = EVT(); |
4931 | SDLoc dl(Node); |
4932 | |
4933 | // Even if the input is legal, no libcall may exactly match, eg. we don't |
4934 | // have i1 -> fp conversions. So, it needs to be promoted to a larger type, |
4935 | // eg: i13 -> fp. Then, look for an appropriate libcall. |
4936 | RTLIB::Libcall LC = RTLIB::UNKNOWN_LIBCALL; |
4937 | for (unsigned t = MVT::FIRST_INTEGER_VALUETYPE; |
4938 | t <= MVT::LAST_INTEGER_VALUETYPE && LC == RTLIB::UNKNOWN_LIBCALL; |
4939 | ++t) { |
4940 | NVT = (MVT::SimpleValueType)t; |
4941 | // The source needs to big enough to hold the operand. |
4942 | if (NVT.bitsGE(VT: SVT)) |
4943 | LC = Signed ? RTLIB::getSINTTOFP(OpVT: NVT, RetVT: RVT) |
4944 | : RTLIB::getUINTTOFP(OpVT: NVT, RetVT: RVT); |
4945 | } |
4946 | assert(LC != RTLIB::UNKNOWN_LIBCALL && "Unable to legalize as libcall" ); |
4947 | |
4948 | SDValue Chain = IsStrict ? Node->getOperand(Num: 0) : SDValue(); |
4949 | // Sign/zero extend the argument if the libcall takes a larger type. |
4950 | SDValue Op = DAG.getNode(Opcode: Signed ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND, DL: dl, |
4951 | VT: NVT, Operand: Node->getOperand(Num: IsStrict ? 1 : 0)); |
4952 | CallOptions.setIsSigned(Signed); |
4953 | std::pair<SDValue, SDValue> Tmp = |
4954 | TLI.makeLibCall(DAG, LC, RetVT: RVT, Ops: Op, CallOptions, dl, Chain); |
4955 | Results.push_back(Elt: Tmp.first); |
4956 | if (IsStrict) |
4957 | Results.push_back(Elt: Tmp.second); |
4958 | break; |
4959 | } |
4960 | case ISD::FP_TO_SINT: |
4961 | case ISD::FP_TO_UINT: |
4962 | case ISD::STRICT_FP_TO_SINT: |
4963 | case ISD::STRICT_FP_TO_UINT: { |
4964 | // TODO - Common the code with DAGTypeLegalizer::SoftenFloatOp_FP_TO_XINT. |
4965 | bool IsStrict = Node->isStrictFPOpcode(); |
4966 | bool Signed = Node->getOpcode() == ISD::FP_TO_SINT || |
4967 | Node->getOpcode() == ISD::STRICT_FP_TO_SINT; |
4968 | |
4969 | SDValue Op = Node->getOperand(Num: IsStrict ? 1 : 0); |
4970 | EVT SVT = Op.getValueType(); |
4971 | EVT RVT = Node->getValueType(ResNo: 0); |
4972 | EVT NVT = EVT(); |
4973 | SDLoc dl(Node); |
4974 | |
4975 | // Even if the result is legal, no libcall may exactly match, eg. we don't |
4976 | // have fp -> i1 conversions. So, it needs to be promoted to a larger type, |
4977 | // eg: fp -> i32. Then, look for an appropriate libcall. |
4978 | RTLIB::Libcall LC = RTLIB::UNKNOWN_LIBCALL; |
4979 | for (unsigned IntVT = MVT::FIRST_INTEGER_VALUETYPE; |
4980 | IntVT <= MVT::LAST_INTEGER_VALUETYPE && LC == RTLIB::UNKNOWN_LIBCALL; |
4981 | ++IntVT) { |
4982 | NVT = (MVT::SimpleValueType)IntVT; |
4983 | // The type needs to big enough to hold the result. |
4984 | if (NVT.bitsGE(VT: RVT)) |
4985 | LC = Signed ? RTLIB::getFPTOSINT(OpVT: SVT, RetVT: NVT) |
4986 | : RTLIB::getFPTOUINT(OpVT: SVT, RetVT: NVT); |
4987 | } |
4988 | assert(LC != RTLIB::UNKNOWN_LIBCALL && "Unable to legalize as libcall" ); |
4989 | |
4990 | SDValue Chain = IsStrict ? Node->getOperand(Num: 0) : SDValue(); |
4991 | std::pair<SDValue, SDValue> Tmp = |
4992 | TLI.makeLibCall(DAG, LC, RetVT: NVT, Ops: Op, CallOptions, dl, Chain); |
4993 | |
4994 | // Truncate the result if the libcall returns a larger type. |
4995 | Results.push_back(Elt: DAG.getNode(Opcode: ISD::TRUNCATE, DL: dl, VT: RVT, Operand: Tmp.first)); |
4996 | if (IsStrict) |
4997 | Results.push_back(Elt: Tmp.second); |
4998 | break; |
4999 | } |
5000 | |
5001 | case ISD::FP_ROUND: |
5002 | case ISD::STRICT_FP_ROUND: { |
5003 | // X = FP_ROUND(Y, TRUNC) |
5004 | // TRUNC is a flag, which is always an integer that is zero or one. |
5005 | // If TRUNC is 0, this is a normal rounding, if it is 1, this FP_ROUND |
5006 | // is known to not change the value of Y. |
5007 | // We can only expand it into libcall if the TRUNC is 0. |
5008 | bool IsStrict = Node->isStrictFPOpcode(); |
5009 | SDValue Op = Node->getOperand(Num: IsStrict ? 1 : 0); |
5010 | SDValue Chain = IsStrict ? Node->getOperand(Num: 0) : SDValue(); |
5011 | EVT VT = Node->getValueType(ResNo: 0); |
5012 | assert(cast<ConstantSDNode>(Node->getOperand(IsStrict ? 2 : 1))->isZero() && |
5013 | "Unable to expand as libcall if it is not normal rounding" ); |
5014 | |
5015 | RTLIB::Libcall LC = RTLIB::getFPROUND(OpVT: Op.getValueType(), RetVT: VT); |
5016 | assert(LC != RTLIB::UNKNOWN_LIBCALL && "Unable to legalize as libcall" ); |
5017 | |
5018 | std::pair<SDValue, SDValue> Tmp = |
5019 | TLI.makeLibCall(DAG, LC, RetVT: VT, Ops: Op, CallOptions, dl: SDLoc(Node), Chain); |
5020 | Results.push_back(Elt: Tmp.first); |
5021 | if (IsStrict) |
5022 | Results.push_back(Elt: Tmp.second); |
5023 | break; |
5024 | } |
5025 | case ISD::FP_EXTEND: { |
5026 | Results.push_back( |
5027 | Elt: ExpandLibCall(LC: RTLIB::getFPEXT(OpVT: Node->getOperand(Num: 0).getValueType(), |
5028 | RetVT: Node->getValueType(ResNo: 0)), |
5029 | Node, isSigned: false).first); |
5030 | break; |
5031 | } |
5032 | case ISD::STRICT_FP_EXTEND: |
5033 | case ISD::STRICT_FP_TO_FP16: |
5034 | case ISD::STRICT_FP_TO_BF16: { |
5035 | RTLIB::Libcall LC = RTLIB::UNKNOWN_LIBCALL; |
5036 | if (Node->getOpcode() == ISD::STRICT_FP_TO_FP16) |
5037 | LC = RTLIB::getFPROUND(OpVT: Node->getOperand(Num: 1).getValueType(), RetVT: MVT::f16); |
5038 | else if (Node->getOpcode() == ISD::STRICT_FP_TO_BF16) |
5039 | LC = RTLIB::getFPROUND(OpVT: Node->getOperand(Num: 1).getValueType(), RetVT: MVT::bf16); |
5040 | else |
5041 | LC = RTLIB::getFPEXT(OpVT: Node->getOperand(Num: 1).getValueType(), |
5042 | RetVT: Node->getValueType(ResNo: 0)); |
5043 | |
5044 | assert(LC != RTLIB::UNKNOWN_LIBCALL && "Unable to legalize as libcall" ); |
5045 | |
5046 | std::pair<SDValue, SDValue> Tmp = |
5047 | TLI.makeLibCall(DAG, LC, RetVT: Node->getValueType(ResNo: 0), Ops: Node->getOperand(Num: 1), |
5048 | CallOptions, dl: SDLoc(Node), Chain: Node->getOperand(Num: 0)); |
5049 | Results.push_back(Elt: Tmp.first); |
5050 | Results.push_back(Elt: Tmp.second); |
5051 | break; |
5052 | } |
5053 | case ISD::FSUB: |
5054 | case ISD::STRICT_FSUB: |
5055 | ExpandFPLibCall(Node, Call_F32: RTLIB::SUB_F32, Call_F64: RTLIB::SUB_F64, |
5056 | Call_F80: RTLIB::SUB_F80, Call_F128: RTLIB::SUB_F128, |
5057 | Call_PPCF128: RTLIB::SUB_PPCF128, Results); |
5058 | break; |
5059 | case ISD::SREM: |
5060 | Results.push_back(Elt: ExpandIntLibCall(Node, isSigned: true, |
5061 | Call_I8: RTLIB::SREM_I8, |
5062 | Call_I16: RTLIB::SREM_I16, Call_I32: RTLIB::SREM_I32, |
5063 | Call_I64: RTLIB::SREM_I64, Call_I128: RTLIB::SREM_I128)); |
5064 | break; |
5065 | case ISD::UREM: |
5066 | Results.push_back(Elt: ExpandIntLibCall(Node, isSigned: false, |
5067 | Call_I8: RTLIB::UREM_I8, |
5068 | Call_I16: RTLIB::UREM_I16, Call_I32: RTLIB::UREM_I32, |
5069 | Call_I64: RTLIB::UREM_I64, Call_I128: RTLIB::UREM_I128)); |
5070 | break; |
5071 | case ISD::SDIV: |
5072 | Results.push_back(Elt: ExpandIntLibCall(Node, isSigned: true, |
5073 | Call_I8: RTLIB::SDIV_I8, |
5074 | Call_I16: RTLIB::SDIV_I16, Call_I32: RTLIB::SDIV_I32, |
5075 | Call_I64: RTLIB::SDIV_I64, Call_I128: RTLIB::SDIV_I128)); |
5076 | break; |
5077 | case ISD::UDIV: |
5078 | Results.push_back(Elt: ExpandIntLibCall(Node, isSigned: false, |
5079 | Call_I8: RTLIB::UDIV_I8, |
5080 | Call_I16: RTLIB::UDIV_I16, Call_I32: RTLIB::UDIV_I32, |
5081 | Call_I64: RTLIB::UDIV_I64, Call_I128: RTLIB::UDIV_I128)); |
5082 | break; |
5083 | case ISD::SDIVREM: |
5084 | case ISD::UDIVREM: |
5085 | // Expand into divrem libcall |
5086 | ExpandDivRemLibCall(Node, Results); |
5087 | break; |
5088 | case ISD::MUL: |
5089 | Results.push_back(Elt: ExpandIntLibCall(Node, isSigned: false, |
5090 | Call_I8: RTLIB::MUL_I8, |
5091 | Call_I16: RTLIB::MUL_I16, Call_I32: RTLIB::MUL_I32, |
5092 | Call_I64: RTLIB::MUL_I64, Call_I128: RTLIB::MUL_I128)); |
5093 | break; |
5094 | case ISD::CTLZ_ZERO_UNDEF: |
5095 | Results.push_back(Elt: ExpandBitCountingLibCall( |
5096 | Node, CallI32: RTLIB::CTLZ_I32, CallI64: RTLIB::CTLZ_I64, CallI128: RTLIB::CTLZ_I128)); |
5097 | break; |
5098 | case ISD::CTPOP: |
5099 | Results.push_back(Elt: ExpandBitCountingLibCall( |
5100 | Node, CallI32: RTLIB::CTPOP_I32, CallI64: RTLIB::CTPOP_I64, CallI128: RTLIB::CTPOP_I128)); |
5101 | break; |
5102 | case ISD::RESET_FPENV: { |
5103 | // It is legalized to call 'fesetenv(FE_DFL_ENV)'. On most targets |
5104 | // FE_DFL_ENV is defined as '((const fenv_t *) -1)' in glibc. |
5105 | EVT PtrTy = TLI.getPointerTy(DL: DAG.getDataLayout()); |
5106 | SDValue Ptr = DAG.getAllOnesConstant(DL: dl, VT: PtrTy); |
5107 | SDValue Chain = Node->getOperand(Num: 0); |
5108 | Results.push_back( |
5109 | Elt: DAG.makeStateFunctionCall(LibFunc: RTLIB::FESETENV, Ptr, InChain: Chain, DLoc: dl)); |
5110 | break; |
5111 | } |
5112 | case ISD::GET_FPENV_MEM: { |
5113 | SDValue Chain = Node->getOperand(Num: 0); |
5114 | SDValue EnvPtr = Node->getOperand(Num: 1); |
5115 | Results.push_back( |
5116 | Elt: DAG.makeStateFunctionCall(LibFunc: RTLIB::FEGETENV, Ptr: EnvPtr, InChain: Chain, DLoc: dl)); |
5117 | break; |
5118 | } |
5119 | case ISD::SET_FPENV_MEM: { |
5120 | SDValue Chain = Node->getOperand(Num: 0); |
5121 | SDValue EnvPtr = Node->getOperand(Num: 1); |
5122 | Results.push_back( |
5123 | Elt: DAG.makeStateFunctionCall(LibFunc: RTLIB::FESETENV, Ptr: EnvPtr, InChain: Chain, DLoc: dl)); |
5124 | break; |
5125 | } |
5126 | case ISD::GET_FPMODE: { |
5127 | // Call fegetmode, which saves control modes into a stack slot. Then load |
5128 | // the value to return from the stack. |
5129 | EVT ModeVT = Node->getValueType(ResNo: 0); |
5130 | SDValue StackPtr = DAG.CreateStackTemporary(VT: ModeVT); |
5131 | int SPFI = cast<FrameIndexSDNode>(Val: StackPtr.getNode())->getIndex(); |
5132 | SDValue Chain = DAG.makeStateFunctionCall(LibFunc: RTLIB::FEGETMODE, Ptr: StackPtr, |
5133 | InChain: Node->getOperand(Num: 0), DLoc: dl); |
5134 | SDValue LdInst = DAG.getLoad( |
5135 | VT: ModeVT, dl, Chain, Ptr: StackPtr, |
5136 | PtrInfo: MachinePointerInfo::getFixedStack(MF&: DAG.getMachineFunction(), FI: SPFI)); |
5137 | Results.push_back(Elt: LdInst); |
5138 | Results.push_back(Elt: LdInst.getValue(R: 1)); |
5139 | break; |
5140 | } |
5141 | case ISD::SET_FPMODE: { |
5142 | // Move control modes to stack slot and then call fesetmode with the pointer |
5143 | // to the slot as argument. |
5144 | SDValue Mode = Node->getOperand(Num: 1); |
5145 | EVT ModeVT = Mode.getValueType(); |
5146 | SDValue StackPtr = DAG.CreateStackTemporary(VT: ModeVT); |
5147 | int SPFI = cast<FrameIndexSDNode>(Val: StackPtr.getNode())->getIndex(); |
5148 | SDValue StInst = DAG.getStore( |
5149 | Chain: Node->getOperand(Num: 0), dl, Val: Mode, Ptr: StackPtr, |
5150 | PtrInfo: MachinePointerInfo::getFixedStack(MF&: DAG.getMachineFunction(), FI: SPFI)); |
5151 | Results.push_back( |
5152 | Elt: DAG.makeStateFunctionCall(LibFunc: RTLIB::FESETMODE, Ptr: StackPtr, InChain: StInst, DLoc: dl)); |
5153 | break; |
5154 | } |
5155 | case ISD::RESET_FPMODE: { |
5156 | // It is legalized to a call 'fesetmode(FE_DFL_MODE)'. On most targets |
5157 | // FE_DFL_MODE is defined as '((const femode_t *) -1)' in glibc. If not, the |
5158 | // target must provide custom lowering. |
5159 | const DataLayout &DL = DAG.getDataLayout(); |
5160 | EVT PtrTy = TLI.getPointerTy(DL); |
5161 | SDValue Mode = DAG.getAllOnesConstant(DL: dl, VT: PtrTy); |
5162 | Results.push_back(Elt: DAG.makeStateFunctionCall(LibFunc: RTLIB::FESETMODE, Ptr: Mode, |
5163 | InChain: Node->getOperand(Num: 0), DLoc: dl)); |
5164 | break; |
5165 | } |
5166 | } |
5167 | |
5168 | // Replace the original node with the legalized result. |
5169 | if (!Results.empty()) { |
5170 | LLVM_DEBUG(dbgs() << "Successfully converted node to libcall\n" ); |
5171 | ReplaceNode(Old: Node, New: Results.data()); |
5172 | } else |
5173 | LLVM_DEBUG(dbgs() << "Could not convert node to libcall\n" ); |
5174 | } |
5175 | |
5176 | // Determine the vector type to use in place of an original scalar element when |
5177 | // promoting equally sized vectors. |
5178 | static MVT getPromotedVectorElementType(const TargetLowering &TLI, |
5179 | MVT EltVT, MVT NewEltVT) { |
5180 | unsigned OldEltsPerNewElt = EltVT.getSizeInBits() / NewEltVT.getSizeInBits(); |
5181 | MVT MidVT = OldEltsPerNewElt == 1 |
5182 | ? NewEltVT |
5183 | : MVT::getVectorVT(VT: NewEltVT, NumElements: OldEltsPerNewElt); |
5184 | assert(TLI.isTypeLegal(MidVT) && "unexpected" ); |
5185 | return MidVT; |
5186 | } |
5187 | |
5188 | void SelectionDAGLegalize::PromoteNode(SDNode *Node) { |
5189 | LLVM_DEBUG(dbgs() << "Trying to promote node\n" ); |
5190 | SmallVector<SDValue, 8> Results; |
5191 | MVT OVT = Node->getSimpleValueType(ResNo: 0); |
5192 | if (Node->getOpcode() == ISD::UINT_TO_FP || |
5193 | Node->getOpcode() == ISD::SINT_TO_FP || |
5194 | Node->getOpcode() == ISD::SETCC || |
5195 | Node->getOpcode() == ISD::EXTRACT_VECTOR_ELT || |
5196 | Node->getOpcode() == ISD::INSERT_VECTOR_ELT || |
5197 | Node->getOpcode() == ISD::VECREDUCE_FMAX || |
5198 | Node->getOpcode() == ISD::VECREDUCE_FMIN || |
5199 | Node->getOpcode() == ISD::VECREDUCE_FMAXIMUM || |
5200 | Node->getOpcode() == ISD::VECREDUCE_FMINIMUM) { |
5201 | OVT = Node->getOperand(Num: 0).getSimpleValueType(); |
5202 | } |
5203 | if (Node->getOpcode() == ISD::ATOMIC_STORE || |
5204 | Node->getOpcode() == ISD::STRICT_UINT_TO_FP || |
5205 | Node->getOpcode() == ISD::STRICT_SINT_TO_FP || |
5206 | Node->getOpcode() == ISD::STRICT_FSETCC || |
5207 | Node->getOpcode() == ISD::STRICT_FSETCCS || |
5208 | Node->getOpcode() == ISD::VP_REDUCE_FADD || |
5209 | Node->getOpcode() == ISD::VP_REDUCE_FMUL || |
5210 | Node->getOpcode() == ISD::VP_REDUCE_FMAX || |
5211 | Node->getOpcode() == ISD::VP_REDUCE_FMIN || |
5212 | Node->getOpcode() == ISD::VP_REDUCE_FMAXIMUM || |
5213 | Node->getOpcode() == ISD::VP_REDUCE_FMINIMUM || |
5214 | Node->getOpcode() == ISD::VP_REDUCE_SEQ_FADD) |
5215 | OVT = Node->getOperand(Num: 1).getSimpleValueType(); |
5216 | if (Node->getOpcode() == ISD::BR_CC || |
5217 | Node->getOpcode() == ISD::SELECT_CC) |
5218 | OVT = Node->getOperand(Num: 2).getSimpleValueType(); |
5219 | // Preserve fast math flags |
5220 | SDNodeFlags FastMathFlags = Node->getFlags() & SDNodeFlags::FastMathFlags; |
5221 | SelectionDAG::FlagInserter FlagsInserter(DAG, FastMathFlags); |
5222 | MVT NVT = TLI.getTypeToPromoteTo(Op: Node->getOpcode(), VT: OVT); |
5223 | SDLoc dl(Node); |
5224 | SDValue Tmp1, Tmp2, Tmp3, Tmp4; |
5225 | switch (Node->getOpcode()) { |
5226 | case ISD::CTTZ: |
5227 | case ISD::CTTZ_ZERO_UNDEF: |
5228 | case ISD::CTLZ: |
5229 | case ISD::CTPOP: { |
5230 | // Zero extend the argument unless its cttz, then use any_extend. |
5231 | if (Node->getOpcode() == ISD::CTTZ || |
5232 | Node->getOpcode() == ISD::CTTZ_ZERO_UNDEF) |
5233 | Tmp1 = DAG.getNode(Opcode: ISD::ANY_EXTEND, DL: dl, VT: NVT, Operand: Node->getOperand(Num: 0)); |
5234 | else |
5235 | Tmp1 = DAG.getNode(Opcode: ISD::ZERO_EXTEND, DL: dl, VT: NVT, Operand: Node->getOperand(Num: 0)); |
5236 | |
5237 | unsigned NewOpc = Node->getOpcode(); |
5238 | if (NewOpc == ISD::CTTZ) { |
5239 | // The count is the same in the promoted type except if the original |
5240 | // value was zero. This can be handled by setting the bit just off |
5241 | // the top of the original type. |
5242 | auto TopBit = APInt::getOneBitSet(numBits: NVT.getSizeInBits(), |
5243 | BitNo: OVT.getSizeInBits()); |
5244 | Tmp1 = DAG.getNode(Opcode: ISD::OR, DL: dl, VT: NVT, N1: Tmp1, |
5245 | N2: DAG.getConstant(Val: TopBit, DL: dl, VT: NVT)); |
5246 | NewOpc = ISD::CTTZ_ZERO_UNDEF; |
5247 | } |
5248 | // Perform the larger operation. For CTPOP and CTTZ_ZERO_UNDEF, this is |
5249 | // already the correct result. |
5250 | Tmp1 = DAG.getNode(Opcode: NewOpc, DL: dl, VT: NVT, Operand: Tmp1); |
5251 | if (NewOpc == ISD::CTLZ) { |
5252 | // Tmp1 = Tmp1 - (sizeinbits(NVT) - sizeinbits(Old VT)) |
5253 | Tmp1 = DAG.getNode(Opcode: ISD::SUB, DL: dl, VT: NVT, N1: Tmp1, |
5254 | N2: DAG.getConstant(Val: NVT.getSizeInBits() - |
5255 | OVT.getSizeInBits(), DL: dl, VT: NVT)); |
5256 | } |
5257 | Results.push_back( |
5258 | Elt: DAG.getNode(Opcode: ISD::TRUNCATE, DL: dl, VT: OVT, Operand: Tmp1, Flags: SDNodeFlags::NoWrap)); |
5259 | break; |
5260 | } |
5261 | case ISD::CTLZ_ZERO_UNDEF: { |
5262 | // We know that the argument is unlikely to be zero, hence we can take a |
5263 | // different approach as compared to ISD::CTLZ |
5264 | |
5265 | // Any Extend the argument |
5266 | auto AnyExtendedNode = |
5267 | DAG.getNode(Opcode: ISD::ANY_EXTEND, DL: dl, VT: NVT, Operand: Node->getOperand(Num: 0)); |
5268 | |
5269 | // Tmp1 = Tmp1 << (sizeinbits(NVT) - sizeinbits(Old VT)) |
5270 | auto ShiftConstant = DAG.getShiftAmountConstant( |
5271 | Val: NVT.getSizeInBits() - OVT.getSizeInBits(), VT: NVT, DL: dl); |
5272 | auto LeftShiftResult = |
5273 | DAG.getNode(Opcode: ISD::SHL, DL: dl, VT: NVT, N1: AnyExtendedNode, N2: ShiftConstant); |
5274 | |
5275 | // Perform the larger operation |
5276 | auto CTLZResult = DAG.getNode(Opcode: Node->getOpcode(), DL: dl, VT: NVT, Operand: LeftShiftResult); |
5277 | Results.push_back(Elt: DAG.getNode(Opcode: ISD::TRUNCATE, DL: dl, VT: OVT, Operand: CTLZResult)); |
5278 | break; |
5279 | } |
5280 | case ISD::BITREVERSE: |
5281 | case ISD::BSWAP: { |
5282 | unsigned DiffBits = NVT.getSizeInBits() - OVT.getSizeInBits(); |
5283 | Tmp1 = DAG.getNode(Opcode: ISD::ZERO_EXTEND, DL: dl, VT: NVT, Operand: Node->getOperand(Num: 0)); |
5284 | Tmp1 = DAG.getNode(Opcode: Node->getOpcode(), DL: dl, VT: NVT, Operand: Tmp1); |
5285 | Tmp1 = DAG.getNode( |
5286 | Opcode: ISD::SRL, DL: dl, VT: NVT, N1: Tmp1, |
5287 | N2: DAG.getConstant(Val: DiffBits, DL: dl, |
5288 | VT: TLI.getShiftAmountTy(LHSTy: NVT, DL: DAG.getDataLayout()))); |
5289 | |
5290 | Results.push_back(Elt: DAG.getNode(Opcode: ISD::TRUNCATE, DL: dl, VT: OVT, Operand: Tmp1)); |
5291 | break; |
5292 | } |
5293 | case ISD::FP_TO_UINT: |
5294 | case ISD::STRICT_FP_TO_UINT: |
5295 | case ISD::FP_TO_SINT: |
5296 | case ISD::STRICT_FP_TO_SINT: |
5297 | PromoteLegalFP_TO_INT(N: Node, dl, Results); |
5298 | break; |
5299 | case ISD::FP_TO_UINT_SAT: |
5300 | case ISD::FP_TO_SINT_SAT: |
5301 | Results.push_back(Elt: PromoteLegalFP_TO_INT_SAT(Node, dl)); |
5302 | break; |
5303 | case ISD::UINT_TO_FP: |
5304 | case ISD::STRICT_UINT_TO_FP: |
5305 | case ISD::SINT_TO_FP: |
5306 | case ISD::STRICT_SINT_TO_FP: |
5307 | PromoteLegalINT_TO_FP(N: Node, dl, Results); |
5308 | break; |
5309 | case ISD::VAARG: { |
5310 | SDValue Chain = Node->getOperand(Num: 0); // Get the chain. |
5311 | SDValue Ptr = Node->getOperand(Num: 1); // Get the pointer. |
5312 | |
5313 | unsigned TruncOp; |
5314 | if (OVT.isVector()) { |
5315 | TruncOp = ISD::BITCAST; |
5316 | } else { |
5317 | assert(OVT.isInteger() |
5318 | && "VAARG promotion is supported only for vectors or integer types" ); |
5319 | TruncOp = ISD::TRUNCATE; |
5320 | } |
5321 | |
5322 | // Perform the larger operation, then convert back |
5323 | Tmp1 = DAG.getVAArg(VT: NVT, dl, Chain, Ptr, SV: Node->getOperand(Num: 2), |
5324 | Align: Node->getConstantOperandVal(Num: 3)); |
5325 | Chain = Tmp1.getValue(R: 1); |
5326 | |
5327 | Tmp2 = DAG.getNode(Opcode: TruncOp, DL: dl, VT: OVT, Operand: Tmp1); |
5328 | |
5329 | // Modified the chain result - switch anything that used the old chain to |
5330 | // use the new one. |
5331 | DAG.ReplaceAllUsesOfValueWith(From: SDValue(Node, 0), To: Tmp2); |
5332 | DAG.ReplaceAllUsesOfValueWith(From: SDValue(Node, 1), To: Chain); |
5333 | if (UpdatedNodes) { |
5334 | UpdatedNodes->insert(X: Tmp2.getNode()); |
5335 | UpdatedNodes->insert(X: Chain.getNode()); |
5336 | } |
5337 | ReplacedNode(N: Node); |
5338 | break; |
5339 | } |
5340 | case ISD::MUL: |
5341 | case ISD::SDIV: |
5342 | case ISD::SREM: |
5343 | case ISD::UDIV: |
5344 | case ISD::UREM: |
5345 | case ISD::SMIN: |
5346 | case ISD::SMAX: |
5347 | case ISD::UMIN: |
5348 | case ISD::UMAX: |
5349 | case ISD::AND: |
5350 | case ISD::OR: |
5351 | case ISD::XOR: { |
5352 | unsigned ExtOp, TruncOp; |
5353 | if (OVT.isVector()) { |
5354 | ExtOp = ISD::BITCAST; |
5355 | TruncOp = ISD::BITCAST; |
5356 | } else { |
5357 | assert(OVT.isInteger() && "Cannot promote logic operation" ); |
5358 | |
5359 | switch (Node->getOpcode()) { |
5360 | default: |
5361 | ExtOp = ISD::ANY_EXTEND; |
5362 | break; |
5363 | case ISD::SDIV: |
5364 | case ISD::SREM: |
5365 | case ISD::SMIN: |
5366 | case ISD::SMAX: |
5367 | ExtOp = ISD::SIGN_EXTEND; |
5368 | break; |
5369 | case ISD::UDIV: |
5370 | case ISD::UREM: |
5371 | ExtOp = ISD::ZERO_EXTEND; |
5372 | break; |
5373 | case ISD::UMIN: |
5374 | case ISD::UMAX: |
5375 | if (TLI.isSExtCheaperThanZExt(FromTy: OVT, ToTy: NVT)) |
5376 | ExtOp = ISD::SIGN_EXTEND; |
5377 | else |
5378 | ExtOp = ISD::ZERO_EXTEND; |
5379 | break; |
5380 | } |
5381 | TruncOp = ISD::TRUNCATE; |
5382 | } |
5383 | // Promote each of the values to the new type. |
5384 | Tmp1 = DAG.getNode(Opcode: ExtOp, DL: dl, VT: NVT, Operand: Node->getOperand(Num: 0)); |
5385 | Tmp2 = DAG.getNode(Opcode: ExtOp, DL: dl, VT: NVT, Operand: Node->getOperand(Num: 1)); |
5386 | // Perform the larger operation, then convert back |
5387 | Tmp1 = DAG.getNode(Opcode: Node->getOpcode(), DL: dl, VT: NVT, N1: Tmp1, N2: Tmp2); |
5388 | Results.push_back(Elt: DAG.getNode(Opcode: TruncOp, DL: dl, VT: OVT, Operand: Tmp1)); |
5389 | break; |
5390 | } |
5391 | case ISD::UMUL_LOHI: |
5392 | case ISD::SMUL_LOHI: { |
5393 | // Promote to a multiply in a wider integer type. |
5394 | unsigned ExtOp = Node->getOpcode() == ISD::UMUL_LOHI ? ISD::ZERO_EXTEND |
5395 | : ISD::SIGN_EXTEND; |
5396 | Tmp1 = DAG.getNode(Opcode: ExtOp, DL: dl, VT: NVT, Operand: Node->getOperand(Num: 0)); |
5397 | Tmp2 = DAG.getNode(Opcode: ExtOp, DL: dl, VT: NVT, Operand: Node->getOperand(Num: 1)); |
5398 | Tmp1 = DAG.getNode(Opcode: ISD::MUL, DL: dl, VT: NVT, N1: Tmp1, N2: Tmp2); |
5399 | |
5400 | auto &DL = DAG.getDataLayout(); |
5401 | unsigned OriginalSize = OVT.getScalarSizeInBits(); |
5402 | Tmp2 = DAG.getNode( |
5403 | Opcode: ISD::SRL, DL: dl, VT: NVT, N1: Tmp1, |
5404 | N2: DAG.getConstant(Val: OriginalSize, DL: dl, VT: TLI.getScalarShiftAmountTy(DL, NVT))); |
5405 | Results.push_back(Elt: DAG.getNode(Opcode: ISD::TRUNCATE, DL: dl, VT: OVT, Operand: Tmp1)); |
5406 | Results.push_back(Elt: DAG.getNode(Opcode: ISD::TRUNCATE, DL: dl, VT: OVT, Operand: Tmp2)); |
5407 | break; |
5408 | } |
5409 | case ISD::SELECT: { |
5410 | unsigned ExtOp, TruncOp; |
5411 | if (Node->getValueType(ResNo: 0).isVector() || |
5412 | Node->getValueType(ResNo: 0).getSizeInBits() == NVT.getSizeInBits()) { |
5413 | ExtOp = ISD::BITCAST; |
5414 | TruncOp = ISD::BITCAST; |
5415 | } else if (Node->getValueType(ResNo: 0).isInteger()) { |
5416 | ExtOp = ISD::ANY_EXTEND; |
5417 | TruncOp = ISD::TRUNCATE; |
5418 | } else { |
5419 | ExtOp = ISD::FP_EXTEND; |
5420 | TruncOp = ISD::FP_ROUND; |
5421 | } |
5422 | Tmp1 = Node->getOperand(Num: 0); |
5423 | // Promote each of the values to the new type. |
5424 | Tmp2 = DAG.getNode(Opcode: ExtOp, DL: dl, VT: NVT, Operand: Node->getOperand(Num: 1)); |
5425 | Tmp3 = DAG.getNode(Opcode: ExtOp, DL: dl, VT: NVT, Operand: Node->getOperand(Num: 2)); |
5426 | // Perform the larger operation, then round down. |
5427 | Tmp1 = DAG.getSelect(DL: dl, VT: NVT, Cond: Tmp1, LHS: Tmp2, RHS: Tmp3); |
5428 | if (TruncOp != ISD::FP_ROUND) |
5429 | Tmp1 = DAG.getNode(Opcode: TruncOp, DL: dl, VT: Node->getValueType(ResNo: 0), Operand: Tmp1); |
5430 | else |
5431 | Tmp1 = DAG.getNode(Opcode: TruncOp, DL: dl, VT: Node->getValueType(ResNo: 0), N1: Tmp1, |
5432 | N2: DAG.getIntPtrConstant(Val: 0, DL: dl, /*isTarget=*/true)); |
5433 | Results.push_back(Elt: Tmp1); |
5434 | break; |
5435 | } |
5436 | case ISD::VECTOR_SHUFFLE: { |
5437 | ArrayRef<int> Mask = cast<ShuffleVectorSDNode>(Val: Node)->getMask(); |
5438 | |
5439 | // Cast the two input vectors. |
5440 | Tmp1 = DAG.getNode(Opcode: ISD::BITCAST, DL: dl, VT: NVT, Operand: Node->getOperand(Num: 0)); |
5441 | Tmp2 = DAG.getNode(Opcode: ISD::BITCAST, DL: dl, VT: NVT, Operand: Node->getOperand(Num: 1)); |
5442 | |
5443 | // Convert the shuffle mask to the right # elements. |
5444 | Tmp1 = ShuffleWithNarrowerEltType(NVT, VT: OVT, dl, N1: Tmp1, N2: Tmp2, Mask); |
5445 | Tmp1 = DAG.getNode(Opcode: ISD::BITCAST, DL: dl, VT: OVT, Operand: Tmp1); |
5446 | Results.push_back(Elt: Tmp1); |
5447 | break; |
5448 | } |
5449 | case ISD::VECTOR_SPLICE: { |
5450 | Tmp1 = DAG.getNode(Opcode: ISD::ANY_EXTEND, DL: dl, VT: NVT, Operand: Node->getOperand(Num: 0)); |
5451 | Tmp2 = DAG.getNode(Opcode: ISD::ANY_EXTEND, DL: dl, VT: NVT, Operand: Node->getOperand(Num: 1)); |
5452 | Tmp3 = DAG.getNode(Opcode: ISD::VECTOR_SPLICE, DL: dl, VT: NVT, N1: Tmp1, N2: Tmp2, |
5453 | N3: Node->getOperand(Num: 2)); |
5454 | Results.push_back(Elt: DAG.getNode(Opcode: ISD::TRUNCATE, DL: dl, VT: OVT, Operand: Tmp3)); |
5455 | break; |
5456 | } |
5457 | case ISD::SELECT_CC: { |
5458 | SDValue Cond = Node->getOperand(Num: 4); |
5459 | ISD::CondCode CCCode = cast<CondCodeSDNode>(Val&: Cond)->get(); |
5460 | // Type of the comparison operands. |
5461 | MVT CVT = Node->getSimpleValueType(ResNo: 0); |
5462 | assert(CVT == OVT && "not handled" ); |
5463 | |
5464 | unsigned ExtOp = ISD::FP_EXTEND; |
5465 | if (NVT.isInteger()) { |
5466 | ExtOp = isSignedIntSetCC(Code: CCCode) ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND; |
5467 | } |
5468 | |
5469 | // Promote the comparison operands, if needed. |
5470 | if (TLI.isCondCodeLegal(CC: CCCode, VT: CVT)) { |
5471 | Tmp1 = Node->getOperand(Num: 0); |
5472 | Tmp2 = Node->getOperand(Num: 1); |
5473 | } else { |
5474 | Tmp1 = DAG.getNode(Opcode: ExtOp, DL: dl, VT: NVT, Operand: Node->getOperand(Num: 0)); |
5475 | Tmp2 = DAG.getNode(Opcode: ExtOp, DL: dl, VT: NVT, Operand: Node->getOperand(Num: 1)); |
5476 | } |
5477 | // Cast the true/false operands. |
5478 | Tmp3 = DAG.getNode(Opcode: ExtOp, DL: dl, VT: NVT, Operand: Node->getOperand(Num: 2)); |
5479 | Tmp4 = DAG.getNode(Opcode: ExtOp, DL: dl, VT: NVT, Operand: Node->getOperand(Num: 3)); |
5480 | |
5481 | Tmp1 = DAG.getNode(Opcode: ISD::SELECT_CC, DL: dl, VT: NVT, Ops: {Tmp1, Tmp2, Tmp3, Tmp4, Cond}, |
5482 | Flags: Node->getFlags()); |
5483 | |
5484 | // Cast the result back to the original type. |
5485 | if (ExtOp != ISD::FP_EXTEND) |
5486 | Tmp1 = DAG.getNode(Opcode: ISD::TRUNCATE, DL: dl, VT: OVT, Operand: Tmp1); |
5487 | else |
5488 | Tmp1 = DAG.getNode(Opcode: ISD::FP_ROUND, DL: dl, VT: OVT, N1: Tmp1, |
5489 | N2: DAG.getIntPtrConstant(Val: 0, DL: dl, /*isTarget=*/true)); |
5490 | |
5491 | Results.push_back(Elt: Tmp1); |
5492 | break; |
5493 | } |
5494 | case ISD::SETCC: |
5495 | case ISD::STRICT_FSETCC: |
5496 | case ISD::STRICT_FSETCCS: { |
5497 | unsigned ExtOp = ISD::FP_EXTEND; |
5498 | if (NVT.isInteger()) { |
5499 | ISD::CondCode CCCode = cast<CondCodeSDNode>(Val: Node->getOperand(Num: 2))->get(); |
5500 | if (isSignedIntSetCC(Code: CCCode) || |
5501 | TLI.isSExtCheaperThanZExt(FromTy: Node->getOperand(Num: 0).getValueType(), ToTy: NVT)) |
5502 | ExtOp = ISD::SIGN_EXTEND; |
5503 | else |
5504 | ExtOp = ISD::ZERO_EXTEND; |
5505 | } |
5506 | if (Node->isStrictFPOpcode()) { |
5507 | SDValue InChain = Node->getOperand(Num: 0); |
5508 | std::tie(args&: Tmp1, args: std::ignore) = |
5509 | DAG.getStrictFPExtendOrRound(Op: Node->getOperand(Num: 1), Chain: InChain, DL: dl, VT: NVT); |
5510 | std::tie(args&: Tmp2, args: std::ignore) = |
5511 | DAG.getStrictFPExtendOrRound(Op: Node->getOperand(Num: 2), Chain: InChain, DL: dl, VT: NVT); |
5512 | SmallVector<SDValue, 2> TmpChains = {Tmp1.getValue(R: 1), Tmp2.getValue(R: 1)}; |
5513 | SDValue OutChain = DAG.getTokenFactor(DL: dl, Vals&: TmpChains); |
5514 | SDVTList VTs = DAG.getVTList(VT1: Node->getValueType(ResNo: 0), VT2: MVT::Other); |
5515 | Results.push_back(Elt: DAG.getNode(Opcode: Node->getOpcode(), DL: dl, VTList: VTs, |
5516 | Ops: {OutChain, Tmp1, Tmp2, Node->getOperand(Num: 3)}, |
5517 | Flags: Node->getFlags())); |
5518 | Results.push_back(Elt: Results.back().getValue(R: 1)); |
5519 | break; |
5520 | } |
5521 | Tmp1 = DAG.getNode(Opcode: ExtOp, DL: dl, VT: NVT, Operand: Node->getOperand(Num: 0)); |
5522 | Tmp2 = DAG.getNode(Opcode: ExtOp, DL: dl, VT: NVT, Operand: Node->getOperand(Num: 1)); |
5523 | Results.push_back(Elt: DAG.getNode(Opcode: ISD::SETCC, DL: dl, VT: Node->getValueType(ResNo: 0), N1: Tmp1, |
5524 | N2: Tmp2, N3: Node->getOperand(Num: 2), Flags: Node->getFlags())); |
5525 | break; |
5526 | } |
5527 | case ISD::BR_CC: { |
5528 | unsigned ExtOp = ISD::FP_EXTEND; |
5529 | if (NVT.isInteger()) { |
5530 | ISD::CondCode CCCode = |
5531 | cast<CondCodeSDNode>(Val: Node->getOperand(Num: 1))->get(); |
5532 | ExtOp = isSignedIntSetCC(Code: CCCode) ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND; |
5533 | } |
5534 | Tmp1 = DAG.getNode(Opcode: ExtOp, DL: dl, VT: NVT, Operand: Node->getOperand(Num: 2)); |
5535 | Tmp2 = DAG.getNode(Opcode: ExtOp, DL: dl, VT: NVT, Operand: Node->getOperand(Num: 3)); |
5536 | Results.push_back(Elt: DAG.getNode(Opcode: ISD::BR_CC, DL: dl, VT: Node->getValueType(ResNo: 0), |
5537 | N1: Node->getOperand(Num: 0), N2: Node->getOperand(Num: 1), |
5538 | N3: Tmp1, N4: Tmp2, N5: Node->getOperand(Num: 4))); |
5539 | break; |
5540 | } |
5541 | case ISD::FADD: |
5542 | case ISD::FSUB: |
5543 | case ISD::FMUL: |
5544 | case ISD::FDIV: |
5545 | case ISD::FREM: |
5546 | case ISD::FMINNUM: |
5547 | case ISD::FMAXNUM: |
5548 | case ISD::FMINIMUM: |
5549 | case ISD::FMAXIMUM: |
5550 | case ISD::FMINIMUMNUM: |
5551 | case ISD::FMAXIMUMNUM: |
5552 | case ISD::FPOW: |
5553 | case ISD::FATAN2: |
5554 | Tmp1 = DAG.getNode(Opcode: ISD::FP_EXTEND, DL: dl, VT: NVT, Operand: Node->getOperand(Num: 0)); |
5555 | Tmp2 = DAG.getNode(Opcode: ISD::FP_EXTEND, DL: dl, VT: NVT, Operand: Node->getOperand(Num: 1)); |
5556 | Tmp3 = DAG.getNode(Opcode: Node->getOpcode(), DL: dl, VT: NVT, N1: Tmp1, N2: Tmp2); |
5557 | Results.push_back( |
5558 | Elt: DAG.getNode(Opcode: ISD::FP_ROUND, DL: dl, VT: OVT, N1: Tmp3, |
5559 | N2: DAG.getIntPtrConstant(Val: 0, DL: dl, /*isTarget=*/true))); |
5560 | break; |
5561 | |
5562 | case ISD::STRICT_FMINIMUM: |
5563 | case ISD::STRICT_FMAXIMUM: { |
5564 | SDValue InChain = Node->getOperand(Num: 0); |
5565 | SDVTList VTs = DAG.getVTList(VT1: NVT, VT2: MVT::Other); |
5566 | Tmp1 = DAG.getNode(Opcode: ISD::STRICT_FP_EXTEND, DL: dl, VTList: VTs, N1: InChain, |
5567 | N2: Node->getOperand(Num: 1)); |
5568 | Tmp2 = DAG.getNode(Opcode: ISD::STRICT_FP_EXTEND, DL: dl, VTList: VTs, N1: InChain, |
5569 | N2: Node->getOperand(Num: 2)); |
5570 | SmallVector<SDValue, 4> Ops = {InChain, Tmp1, Tmp2}; |
5571 | Tmp3 = DAG.getNode(Opcode: Node->getOpcode(), DL: dl, VTList: VTs, Ops, Flags: Node->getFlags()); |
5572 | Tmp4 = DAG.getNode(Opcode: ISD::STRICT_FP_ROUND, DL: dl, VTList: DAG.getVTList(VT1: OVT, VT2: MVT::Other), |
5573 | N1: InChain, N2: Tmp3, |
5574 | N3: DAG.getIntPtrConstant(Val: 0, DL: dl, /*isTarget=*/true)); |
5575 | Results.push_back(Elt: Tmp4); |
5576 | Results.push_back(Elt: Tmp4.getValue(R: 1)); |
5577 | break; |
5578 | } |
5579 | |
5580 | case ISD::STRICT_FADD: |
5581 | case ISD::STRICT_FSUB: |
5582 | case ISD::STRICT_FMUL: |
5583 | case ISD::STRICT_FDIV: |
5584 | case ISD::STRICT_FMINNUM: |
5585 | case ISD::STRICT_FMAXNUM: |
5586 | case ISD::STRICT_FREM: |
5587 | case ISD::STRICT_FPOW: |
5588 | case ISD::STRICT_FATAN2: |
5589 | Tmp1 = DAG.getNode(Opcode: ISD::STRICT_FP_EXTEND, DL: dl, ResultTys: {NVT, MVT::Other}, |
5590 | Ops: {Node->getOperand(Num: 0), Node->getOperand(Num: 1)}); |
5591 | Tmp2 = DAG.getNode(Opcode: ISD::STRICT_FP_EXTEND, DL: dl, ResultTys: {NVT, MVT::Other}, |
5592 | Ops: {Node->getOperand(Num: 0), Node->getOperand(Num: 2)}); |
5593 | Tmp3 = DAG.getNode(Opcode: ISD::TokenFactor, DL: dl, VT: MVT::Other, N1: Tmp1.getValue(R: 1), |
5594 | N2: Tmp2.getValue(R: 1)); |
5595 | Tmp1 = DAG.getNode(Opcode: Node->getOpcode(), DL: dl, ResultTys: {NVT, MVT::Other}, |
5596 | Ops: {Tmp3, Tmp1, Tmp2}); |
5597 | Tmp1 = DAG.getNode(Opcode: ISD::STRICT_FP_ROUND, DL: dl, ResultTys: {OVT, MVT::Other}, |
5598 | Ops: {Tmp1.getValue(R: 1), Tmp1, |
5599 | DAG.getIntPtrConstant(Val: 0, DL: dl, /*isTarget=*/true)}); |
5600 | Results.push_back(Elt: Tmp1); |
5601 | Results.push_back(Elt: Tmp1.getValue(R: 1)); |
5602 | break; |
5603 | case ISD::FMA: |
5604 | Tmp1 = DAG.getNode(Opcode: ISD::FP_EXTEND, DL: dl, VT: NVT, Operand: Node->getOperand(Num: 0)); |
5605 | Tmp2 = DAG.getNode(Opcode: ISD::FP_EXTEND, DL: dl, VT: NVT, Operand: Node->getOperand(Num: 1)); |
5606 | Tmp3 = DAG.getNode(Opcode: ISD::FP_EXTEND, DL: dl, VT: NVT, Operand: Node->getOperand(Num: 2)); |
5607 | Results.push_back( |
5608 | Elt: DAG.getNode(Opcode: ISD::FP_ROUND, DL: dl, VT: OVT, |
5609 | N1: DAG.getNode(Opcode: Node->getOpcode(), DL: dl, VT: NVT, N1: Tmp1, N2: Tmp2, N3: Tmp3), |
5610 | N2: DAG.getIntPtrConstant(Val: 0, DL: dl, /*isTarget=*/true))); |
5611 | break; |
5612 | case ISD::STRICT_FMA: |
5613 | Tmp1 = DAG.getNode(Opcode: ISD::STRICT_FP_EXTEND, DL: dl, ResultTys: {NVT, MVT::Other}, |
5614 | Ops: {Node->getOperand(Num: 0), Node->getOperand(Num: 1)}); |
5615 | Tmp2 = DAG.getNode(Opcode: ISD::STRICT_FP_EXTEND, DL: dl, ResultTys: {NVT, MVT::Other}, |
5616 | Ops: {Node->getOperand(Num: 0), Node->getOperand(Num: 2)}); |
5617 | Tmp3 = DAG.getNode(Opcode: ISD::STRICT_FP_EXTEND, DL: dl, ResultTys: {NVT, MVT::Other}, |
5618 | Ops: {Node->getOperand(Num: 0), Node->getOperand(Num: 3)}); |
5619 | Tmp4 = DAG.getNode(Opcode: ISD::TokenFactor, DL: dl, VT: MVT::Other, N1: Tmp1.getValue(R: 1), |
5620 | N2: Tmp2.getValue(R: 1), N3: Tmp3.getValue(R: 1)); |
5621 | Tmp4 = DAG.getNode(Opcode: Node->getOpcode(), DL: dl, ResultTys: {NVT, MVT::Other}, |
5622 | Ops: {Tmp4, Tmp1, Tmp2, Tmp3}); |
5623 | Tmp4 = DAG.getNode(Opcode: ISD::STRICT_FP_ROUND, DL: dl, ResultTys: {OVT, MVT::Other}, |
5624 | Ops: {Tmp4.getValue(R: 1), Tmp4, |
5625 | DAG.getIntPtrConstant(Val: 0, DL: dl, /*isTarget=*/true)}); |
5626 | Results.push_back(Elt: Tmp4); |
5627 | Results.push_back(Elt: Tmp4.getValue(R: 1)); |
5628 | break; |
5629 | case ISD::FCOPYSIGN: |
5630 | case ISD::FLDEXP: |
5631 | case ISD::FPOWI: { |
5632 | Tmp1 = DAG.getNode(Opcode: ISD::FP_EXTEND, DL: dl, VT: NVT, Operand: Node->getOperand(Num: 0)); |
5633 | Tmp2 = Node->getOperand(Num: 1); |
5634 | Tmp3 = DAG.getNode(Opcode: Node->getOpcode(), DL: dl, VT: NVT, N1: Tmp1, N2: Tmp2); |
5635 | |
5636 | // fcopysign doesn't change anything but the sign bit, so |
5637 | // (fp_round (fcopysign (fpext a), b)) |
5638 | // is as precise as |
5639 | // (fp_round (fpext a)) |
5640 | // which is a no-op. Mark it as a TRUNCating FP_ROUND. |
5641 | const bool isTrunc = (Node->getOpcode() == ISD::FCOPYSIGN); |
5642 | Results.push_back( |
5643 | Elt: DAG.getNode(Opcode: ISD::FP_ROUND, DL: dl, VT: OVT, N1: Tmp3, |
5644 | N2: DAG.getIntPtrConstant(Val: isTrunc, DL: dl, /*isTarget=*/true))); |
5645 | break; |
5646 | } |
5647 | case ISD::STRICT_FLDEXP: { |
5648 | Tmp1 = DAG.getNode(Opcode: ISD::STRICT_FP_EXTEND, DL: dl, ResultTys: {NVT, MVT::Other}, |
5649 | Ops: {Node->getOperand(Num: 0), Node->getOperand(Num: 1)}); |
5650 | Tmp2 = Node->getOperand(Num: 2); |
5651 | Tmp3 = DAG.getNode(Opcode: ISD::STRICT_FLDEXP, DL: dl, ResultTys: {NVT, MVT::Other}, |
5652 | Ops: {Tmp1.getValue(R: 1), Tmp1, Tmp2}); |
5653 | Tmp4 = DAG.getNode(Opcode: ISD::STRICT_FP_ROUND, DL: dl, ResultTys: {OVT, MVT::Other}, |
5654 | Ops: {Tmp3.getValue(R: 1), Tmp3, |
5655 | DAG.getIntPtrConstant(Val: 0, DL: dl, /*isTarget=*/true)}); |
5656 | Results.push_back(Elt: Tmp4); |
5657 | Results.push_back(Elt: Tmp4.getValue(R: 1)); |
5658 | break; |
5659 | } |
5660 | case ISD::STRICT_FPOWI: |
5661 | Tmp1 = DAG.getNode(Opcode: ISD::STRICT_FP_EXTEND, DL: dl, ResultTys: {NVT, MVT::Other}, |
5662 | Ops: {Node->getOperand(Num: 0), Node->getOperand(Num: 1)}); |
5663 | Tmp2 = DAG.getNode(Opcode: Node->getOpcode(), DL: dl, ResultTys: {NVT, MVT::Other}, |
5664 | Ops: {Tmp1.getValue(R: 1), Tmp1, Node->getOperand(Num: 2)}); |
5665 | Tmp3 = DAG.getNode(Opcode: ISD::STRICT_FP_ROUND, DL: dl, ResultTys: {OVT, MVT::Other}, |
5666 | Ops: {Tmp2.getValue(R: 1), Tmp2, |
5667 | DAG.getIntPtrConstant(Val: 0, DL: dl, /*isTarget=*/true)}); |
5668 | Results.push_back(Elt: Tmp3); |
5669 | Results.push_back(Elt: Tmp3.getValue(R: 1)); |
5670 | break; |
5671 | case ISD::FFREXP: { |
5672 | Tmp1 = DAG.getNode(Opcode: ISD::FP_EXTEND, DL: dl, VT: NVT, Operand: Node->getOperand(Num: 0)); |
5673 | Tmp2 = DAG.getNode(Opcode: ISD::FFREXP, DL: dl, ResultTys: {NVT, Node->getValueType(ResNo: 1)}, Ops: Tmp1); |
5674 | |
5675 | Results.push_back( |
5676 | Elt: DAG.getNode(Opcode: ISD::FP_ROUND, DL: dl, VT: OVT, N1: Tmp2, |
5677 | N2: DAG.getIntPtrConstant(Val: 0, DL: dl, /*isTarget=*/true))); |
5678 | |
5679 | Results.push_back(Elt: Tmp2.getValue(R: 1)); |
5680 | break; |
5681 | } |
5682 | case ISD::FMODF: |
5683 | case ISD::FSINCOS: |
5684 | case ISD::FSINCOSPI: { |
5685 | Tmp1 = DAG.getNode(Opcode: ISD::FP_EXTEND, DL: dl, VT: NVT, Operand: Node->getOperand(Num: 0)); |
5686 | Tmp2 = DAG.getNode(Opcode: Node->getOpcode(), DL: dl, VTList: DAG.getVTList(VT1: NVT, VT2: NVT), N: Tmp1); |
5687 | Tmp3 = DAG.getIntPtrConstant(Val: 0, DL: dl, /*isTarget=*/true); |
5688 | for (unsigned ResNum = 0; ResNum < Node->getNumValues(); ResNum++) |
5689 | Results.push_back( |
5690 | Elt: DAG.getNode(Opcode: ISD::FP_ROUND, DL: dl, VT: OVT, N1: Tmp2.getValue(R: ResNum), N2: Tmp3)); |
5691 | break; |
5692 | } |
5693 | case ISD::FFLOOR: |
5694 | case ISD::FCEIL: |
5695 | case ISD::FRINT: |
5696 | case ISD::FNEARBYINT: |
5697 | case ISD::FROUND: |
5698 | case ISD::FROUNDEVEN: |
5699 | case ISD::FTRUNC: |
5700 | case ISD::FNEG: |
5701 | case ISD::FSQRT: |
5702 | case ISD::FSIN: |
5703 | case ISD::FCOS: |
5704 | case ISD::FTAN: |
5705 | case ISD::FASIN: |
5706 | case ISD::FACOS: |
5707 | case ISD::FATAN: |
5708 | case ISD::FSINH: |
5709 | case ISD::FCOSH: |
5710 | case ISD::FTANH: |
5711 | case ISD::FLOG: |
5712 | case ISD::FLOG2: |
5713 | case ISD::FLOG10: |
5714 | case ISD::FABS: |
5715 | case ISD::FEXP: |
5716 | case ISD::FEXP2: |
5717 | case ISD::FEXP10: |
5718 | case ISD::FCANONICALIZE: |
5719 | Tmp1 = DAG.getNode(Opcode: ISD::FP_EXTEND, DL: dl, VT: NVT, Operand: Node->getOperand(Num: 0)); |
5720 | Tmp2 = DAG.getNode(Opcode: Node->getOpcode(), DL: dl, VT: NVT, Operand: Tmp1); |
5721 | Results.push_back( |
5722 | Elt: DAG.getNode(Opcode: ISD::FP_ROUND, DL: dl, VT: OVT, N1: Tmp2, |
5723 | N2: DAG.getIntPtrConstant(Val: 0, DL: dl, /*isTarget=*/true))); |
5724 | break; |
5725 | case ISD::STRICT_FFLOOR: |
5726 | case ISD::STRICT_FCEIL: |
5727 | case ISD::STRICT_FRINT: |
5728 | case ISD::STRICT_FNEARBYINT: |
5729 | case ISD::STRICT_FROUND: |
5730 | case ISD::STRICT_FROUNDEVEN: |
5731 | case ISD::STRICT_FTRUNC: |
5732 | case ISD::STRICT_FSQRT: |
5733 | case ISD::STRICT_FSIN: |
5734 | case ISD::STRICT_FCOS: |
5735 | case ISD::STRICT_FTAN: |
5736 | case ISD::STRICT_FASIN: |
5737 | case ISD::STRICT_FACOS: |
5738 | case ISD::STRICT_FATAN: |
5739 | case ISD::STRICT_FSINH: |
5740 | case ISD::STRICT_FCOSH: |
5741 | case ISD::STRICT_FTANH: |
5742 | case ISD::STRICT_FLOG: |
5743 | case ISD::STRICT_FLOG2: |
5744 | case ISD::STRICT_FLOG10: |
5745 | case ISD::STRICT_FEXP: |
5746 | case ISD::STRICT_FEXP2: |
5747 | Tmp1 = DAG.getNode(Opcode: ISD::STRICT_FP_EXTEND, DL: dl, ResultTys: {NVT, MVT::Other}, |
5748 | Ops: {Node->getOperand(Num: 0), Node->getOperand(Num: 1)}); |
5749 | Tmp2 = DAG.getNode(Opcode: Node->getOpcode(), DL: dl, ResultTys: {NVT, MVT::Other}, |
5750 | Ops: {Tmp1.getValue(R: 1), Tmp1}); |
5751 | Tmp3 = DAG.getNode(Opcode: ISD::STRICT_FP_ROUND, DL: dl, ResultTys: {OVT, MVT::Other}, |
5752 | Ops: {Tmp2.getValue(R: 1), Tmp2, |
5753 | DAG.getIntPtrConstant(Val: 0, DL: dl, /*isTarget=*/true)}); |
5754 | Results.push_back(Elt: Tmp3); |
5755 | Results.push_back(Elt: Tmp3.getValue(R: 1)); |
5756 | break; |
5757 | case ISD::BUILD_VECTOR: { |
5758 | MVT EltVT = OVT.getVectorElementType(); |
5759 | MVT NewEltVT = NVT.getVectorElementType(); |
5760 | |
5761 | // Handle bitcasts to a different vector type with the same total bit size |
5762 | // |
5763 | // e.g. v2i64 = build_vector i64:x, i64:y => v4i32 |
5764 | // => |
5765 | // v4i32 = concat_vectors (v2i32 (bitcast i64:x)), (v2i32 (bitcast i64:y)) |
5766 | |
5767 | assert(NVT.isVector() && OVT.getSizeInBits() == NVT.getSizeInBits() && |
5768 | "Invalid promote type for build_vector" ); |
5769 | assert(NewEltVT.bitsLE(EltVT) && "not handled" ); |
5770 | |
5771 | MVT MidVT = getPromotedVectorElementType(TLI, EltVT, NewEltVT); |
5772 | |
5773 | SmallVector<SDValue, 8> NewOps; |
5774 | for (const SDValue &Op : Node->op_values()) |
5775 | NewOps.push_back(Elt: DAG.getNode(Opcode: ISD::BITCAST, DL: SDLoc(Op), VT: MidVT, Operand: Op)); |
5776 | |
5777 | SDLoc SL(Node); |
5778 | SDValue Concat = |
5779 | DAG.getNode(Opcode: MidVT == NewEltVT ? ISD::BUILD_VECTOR : ISD::CONCAT_VECTORS, |
5780 | DL: SL, VT: NVT, Ops: NewOps); |
5781 | SDValue CvtVec = DAG.getNode(Opcode: ISD::BITCAST, DL: SL, VT: OVT, Operand: Concat); |
5782 | Results.push_back(Elt: CvtVec); |
5783 | break; |
5784 | } |
5785 | case ISD::EXTRACT_VECTOR_ELT: { |
5786 | MVT EltVT = OVT.getVectorElementType(); |
5787 | MVT NewEltVT = NVT.getVectorElementType(); |
5788 | |
5789 | // Handle bitcasts to a different vector type with the same total bit size. |
5790 | // |
5791 | // e.g. v2i64 = extract_vector_elt x:v2i64, y:i32 |
5792 | // => |
5793 | // v4i32:castx = bitcast x:v2i64 |
5794 | // |
5795 | // i64 = bitcast |
5796 | // (v2i32 build_vector (i32 (extract_vector_elt castx, (2 * y))), |
5797 | // (i32 (extract_vector_elt castx, (2 * y + 1))) |
5798 | // |
5799 | |
5800 | assert(NVT.isVector() && OVT.getSizeInBits() == NVT.getSizeInBits() && |
5801 | "Invalid promote type for extract_vector_elt" ); |
5802 | assert(NewEltVT.bitsLT(EltVT) && "not handled" ); |
5803 | |
5804 | MVT MidVT = getPromotedVectorElementType(TLI, EltVT, NewEltVT); |
5805 | unsigned NewEltsPerOldElt = MidVT.getVectorNumElements(); |
5806 | |
5807 | SDValue Idx = Node->getOperand(Num: 1); |
5808 | EVT IdxVT = Idx.getValueType(); |
5809 | SDLoc SL(Node); |
5810 | SDValue Factor = DAG.getConstant(Val: NewEltsPerOldElt, DL: SL, VT: IdxVT); |
5811 | SDValue NewBaseIdx = DAG.getNode(Opcode: ISD::MUL, DL: SL, VT: IdxVT, N1: Idx, N2: Factor); |
5812 | |
5813 | SDValue CastVec = DAG.getNode(Opcode: ISD::BITCAST, DL: SL, VT: NVT, Operand: Node->getOperand(Num: 0)); |
5814 | |
5815 | SmallVector<SDValue, 8> NewOps; |
5816 | for (unsigned I = 0; I < NewEltsPerOldElt; ++I) { |
5817 | SDValue IdxOffset = DAG.getConstant(Val: I, DL: SL, VT: IdxVT); |
5818 | SDValue TmpIdx = DAG.getNode(Opcode: ISD::ADD, DL: SL, VT: IdxVT, N1: NewBaseIdx, N2: IdxOffset); |
5819 | |
5820 | SDValue Elt = DAG.getNode(Opcode: ISD::EXTRACT_VECTOR_ELT, DL: SL, VT: NewEltVT, |
5821 | N1: CastVec, N2: TmpIdx); |
5822 | NewOps.push_back(Elt); |
5823 | } |
5824 | |
5825 | SDValue NewVec = DAG.getBuildVector(VT: MidVT, DL: SL, Ops: NewOps); |
5826 | Results.push_back(Elt: DAG.getNode(Opcode: ISD::BITCAST, DL: SL, VT: EltVT, Operand: NewVec)); |
5827 | break; |
5828 | } |
5829 | case ISD::INSERT_VECTOR_ELT: { |
5830 | MVT EltVT = OVT.getVectorElementType(); |
5831 | MVT NewEltVT = NVT.getVectorElementType(); |
5832 | |
5833 | // Handle bitcasts to a different vector type with the same total bit size |
5834 | // |
5835 | // e.g. v2i64 = insert_vector_elt x:v2i64, y:i64, z:i32 |
5836 | // => |
5837 | // v4i32:castx = bitcast x:v2i64 |
5838 | // v2i32:casty = bitcast y:i64 |
5839 | // |
5840 | // v2i64 = bitcast |
5841 | // (v4i32 insert_vector_elt |
5842 | // (v4i32 insert_vector_elt v4i32:castx, |
5843 | // (extract_vector_elt casty, 0), 2 * z), |
5844 | // (extract_vector_elt casty, 1), (2 * z + 1)) |
5845 | |
5846 | assert(NVT.isVector() && OVT.getSizeInBits() == NVT.getSizeInBits() && |
5847 | "Invalid promote type for insert_vector_elt" ); |
5848 | assert(NewEltVT.bitsLT(EltVT) && "not handled" ); |
5849 | |
5850 | MVT MidVT = getPromotedVectorElementType(TLI, EltVT, NewEltVT); |
5851 | unsigned NewEltsPerOldElt = MidVT.getVectorNumElements(); |
5852 | |
5853 | SDValue Val = Node->getOperand(Num: 1); |
5854 | SDValue Idx = Node->getOperand(Num: 2); |
5855 | EVT IdxVT = Idx.getValueType(); |
5856 | SDLoc SL(Node); |
5857 | |
5858 | SDValue Factor = DAG.getConstant(Val: NewEltsPerOldElt, DL: SDLoc(), VT: IdxVT); |
5859 | SDValue NewBaseIdx = DAG.getNode(Opcode: ISD::MUL, DL: SL, VT: IdxVT, N1: Idx, N2: Factor); |
5860 | |
5861 | SDValue CastVec = DAG.getNode(Opcode: ISD::BITCAST, DL: SL, VT: NVT, Operand: Node->getOperand(Num: 0)); |
5862 | SDValue CastVal = DAG.getNode(Opcode: ISD::BITCAST, DL: SL, VT: MidVT, Operand: Val); |
5863 | |
5864 | SDValue NewVec = CastVec; |
5865 | for (unsigned I = 0; I < NewEltsPerOldElt; ++I) { |
5866 | SDValue IdxOffset = DAG.getConstant(Val: I, DL: SL, VT: IdxVT); |
5867 | SDValue InEltIdx = DAG.getNode(Opcode: ISD::ADD, DL: SL, VT: IdxVT, N1: NewBaseIdx, N2: IdxOffset); |
5868 | |
5869 | SDValue Elt = DAG.getNode(Opcode: ISD::EXTRACT_VECTOR_ELT, DL: SL, VT: NewEltVT, |
5870 | N1: CastVal, N2: IdxOffset); |
5871 | |
5872 | NewVec = DAG.getNode(Opcode: ISD::INSERT_VECTOR_ELT, DL: SL, VT: NVT, |
5873 | N1: NewVec, N2: Elt, N3: InEltIdx); |
5874 | } |
5875 | |
5876 | Results.push_back(Elt: DAG.getNode(Opcode: ISD::BITCAST, DL: SL, VT: OVT, Operand: NewVec)); |
5877 | break; |
5878 | } |
5879 | case ISD::SCALAR_TO_VECTOR: { |
5880 | MVT EltVT = OVT.getVectorElementType(); |
5881 | MVT NewEltVT = NVT.getVectorElementType(); |
5882 | |
5883 | // Handle bitcasts to different vector type with the same total bit size. |
5884 | // |
5885 | // e.g. v2i64 = scalar_to_vector x:i64 |
5886 | // => |
5887 | // concat_vectors (v2i32 bitcast x:i64), (v2i32 undef) |
5888 | // |
5889 | |
5890 | MVT MidVT = getPromotedVectorElementType(TLI, EltVT, NewEltVT); |
5891 | SDValue Val = Node->getOperand(Num: 0); |
5892 | SDLoc SL(Node); |
5893 | |
5894 | SDValue CastVal = DAG.getNode(Opcode: ISD::BITCAST, DL: SL, VT: MidVT, Operand: Val); |
5895 | SDValue Undef = DAG.getUNDEF(VT: MidVT); |
5896 | |
5897 | SmallVector<SDValue, 8> NewElts; |
5898 | NewElts.push_back(Elt: CastVal); |
5899 | for (unsigned I = 1, NElts = OVT.getVectorNumElements(); I != NElts; ++I) |
5900 | NewElts.push_back(Elt: Undef); |
5901 | |
5902 | SDValue Concat = DAG.getNode(Opcode: ISD::CONCAT_VECTORS, DL: SL, VT: NVT, Ops: NewElts); |
5903 | SDValue CvtVec = DAG.getNode(Opcode: ISD::BITCAST, DL: SL, VT: OVT, Operand: Concat); |
5904 | Results.push_back(Elt: CvtVec); |
5905 | break; |
5906 | } |
5907 | case ISD::ATOMIC_SWAP: |
5908 | case ISD::ATOMIC_STORE: { |
5909 | AtomicSDNode *AM = cast<AtomicSDNode>(Val: Node); |
5910 | SDLoc SL(Node); |
5911 | SDValue CastVal = DAG.getNode(Opcode: ISD::BITCAST, DL: SL, VT: NVT, Operand: AM->getVal()); |
5912 | assert(NVT.getSizeInBits() == OVT.getSizeInBits() && |
5913 | "unexpected promotion type" ); |
5914 | assert(AM->getMemoryVT().getSizeInBits() == NVT.getSizeInBits() && |
5915 | "unexpected atomic_swap with illegal type" ); |
5916 | |
5917 | SDValue Op0 = AM->getBasePtr(); |
5918 | SDValue Op1 = CastVal; |
5919 | |
5920 | // ATOMIC_STORE uses a swapped operand order from every other AtomicSDNode, |
5921 | // but really it should merge with ISD::STORE. |
5922 | if (AM->getOpcode() == ISD::ATOMIC_STORE) |
5923 | std::swap(a&: Op0, b&: Op1); |
5924 | |
5925 | SDValue NewAtomic = DAG.getAtomic(Opcode: AM->getOpcode(), dl: SL, MemVT: NVT, Chain: AM->getChain(), |
5926 | Ptr: Op0, Val: Op1, MMO: AM->getMemOperand()); |
5927 | |
5928 | if (AM->getOpcode() != ISD::ATOMIC_STORE) { |
5929 | Results.push_back(Elt: DAG.getNode(Opcode: ISD::BITCAST, DL: SL, VT: OVT, Operand: NewAtomic)); |
5930 | Results.push_back(Elt: NewAtomic.getValue(R: 1)); |
5931 | } else |
5932 | Results.push_back(Elt: NewAtomic); |
5933 | break; |
5934 | } |
5935 | case ISD::ATOMIC_LOAD: { |
5936 | AtomicSDNode *AM = cast<AtomicSDNode>(Val: Node); |
5937 | SDLoc SL(Node); |
5938 | assert(NVT.getSizeInBits() == OVT.getSizeInBits() && |
5939 | "unexpected promotion type" ); |
5940 | assert(AM->getMemoryVT().getSizeInBits() == NVT.getSizeInBits() && |
5941 | "unexpected atomic_load with illegal type" ); |
5942 | |
5943 | SDValue NewAtomic = |
5944 | DAG.getAtomic(Opcode: ISD::ATOMIC_LOAD, dl: SL, MemVT: NVT, VTList: DAG.getVTList(VT1: NVT, VT2: MVT::Other), |
5945 | Ops: {AM->getChain(), AM->getBasePtr()}, MMO: AM->getMemOperand()); |
5946 | Results.push_back(Elt: DAG.getNode(Opcode: ISD::BITCAST, DL: SL, VT: OVT, Operand: NewAtomic)); |
5947 | Results.push_back(Elt: NewAtomic.getValue(R: 1)); |
5948 | break; |
5949 | } |
5950 | case ISD::SPLAT_VECTOR: { |
5951 | SDValue Scalar = Node->getOperand(Num: 0); |
5952 | MVT ScalarType = Scalar.getSimpleValueType(); |
5953 | MVT NewScalarType = NVT.getVectorElementType(); |
5954 | if (ScalarType.isInteger()) { |
5955 | Tmp1 = DAG.getNode(Opcode: ISD::ANY_EXTEND, DL: dl, VT: NewScalarType, Operand: Scalar); |
5956 | Tmp2 = DAG.getNode(Opcode: Node->getOpcode(), DL: dl, VT: NVT, Operand: Tmp1); |
5957 | Results.push_back(Elt: DAG.getNode(Opcode: ISD::TRUNCATE, DL: dl, VT: OVT, Operand: Tmp2)); |
5958 | break; |
5959 | } |
5960 | Tmp1 = DAG.getNode(Opcode: ISD::FP_EXTEND, DL: dl, VT: NewScalarType, Operand: Scalar); |
5961 | Tmp2 = DAG.getNode(Opcode: Node->getOpcode(), DL: dl, VT: NVT, Operand: Tmp1); |
5962 | Results.push_back( |
5963 | Elt: DAG.getNode(Opcode: ISD::FP_ROUND, DL: dl, VT: OVT, N1: Tmp2, |
5964 | N2: DAG.getIntPtrConstant(Val: 0, DL: dl, /*isTarget=*/true))); |
5965 | break; |
5966 | } |
5967 | case ISD::VECREDUCE_FMAX: |
5968 | case ISD::VECREDUCE_FMIN: |
5969 | case ISD::VECREDUCE_FMAXIMUM: |
5970 | case ISD::VECREDUCE_FMINIMUM: |
5971 | case ISD::VP_REDUCE_FMAX: |
5972 | case ISD::VP_REDUCE_FMIN: |
5973 | case ISD::VP_REDUCE_FMAXIMUM: |
5974 | case ISD::VP_REDUCE_FMINIMUM: |
5975 | Results.push_back(Elt: PromoteReduction(Node)); |
5976 | break; |
5977 | } |
5978 | |
5979 | // Replace the original node with the legalized result. |
5980 | if (!Results.empty()) { |
5981 | LLVM_DEBUG(dbgs() << "Successfully promoted node\n" ); |
5982 | ReplaceNode(Old: Node, New: Results.data()); |
5983 | } else |
5984 | LLVM_DEBUG(dbgs() << "Could not promote node\n" ); |
5985 | } |
5986 | |
5987 | /// This is the entry point for the file. |
5988 | void SelectionDAG::Legalize() { |
5989 | AssignTopologicalOrder(); |
5990 | |
5991 | SmallPtrSet<SDNode *, 16> LegalizedNodes; |
5992 | // Use a delete listener to remove nodes which were deleted during |
5993 | // legalization from LegalizeNodes. This is needed to handle the situation |
5994 | // where a new node is allocated by the object pool to the same address of a |
5995 | // previously deleted node. |
5996 | DAGNodeDeletedListener DeleteListener( |
5997 | *this, |
5998 | [&LegalizedNodes](SDNode *N, SDNode *E) { LegalizedNodes.erase(Ptr: N); }); |
5999 | |
6000 | SelectionDAGLegalize Legalizer(*this, LegalizedNodes); |
6001 | |
6002 | // Visit all the nodes. We start in topological order, so that we see |
6003 | // nodes with their original operands intact. Legalization can produce |
6004 | // new nodes which may themselves need to be legalized. Iterate until all |
6005 | // nodes have been legalized. |
6006 | while (true) { |
6007 | bool AnyLegalized = false; |
6008 | for (auto NI = allnodes_end(); NI != allnodes_begin();) { |
6009 | --NI; |
6010 | |
6011 | SDNode *N = &*NI; |
6012 | if (N->use_empty() && N != getRoot().getNode()) { |
6013 | ++NI; |
6014 | DeleteNode(N); |
6015 | continue; |
6016 | } |
6017 | |
6018 | if (LegalizedNodes.insert(Ptr: N).second) { |
6019 | AnyLegalized = true; |
6020 | Legalizer.LegalizeOp(Node: N); |
6021 | |
6022 | if (N->use_empty() && N != getRoot().getNode()) { |
6023 | ++NI; |
6024 | DeleteNode(N); |
6025 | } |
6026 | } |
6027 | } |
6028 | if (!AnyLegalized) |
6029 | break; |
6030 | |
6031 | } |
6032 | |
6033 | // Remove dead nodes now. |
6034 | RemoveDeadNodes(); |
6035 | } |
6036 | |
6037 | bool SelectionDAG::LegalizeOp(SDNode *N, |
6038 | SmallSetVector<SDNode *, 16> &UpdatedNodes) { |
6039 | SmallPtrSet<SDNode *, 16> LegalizedNodes; |
6040 | SelectionDAGLegalize Legalizer(*this, LegalizedNodes, &UpdatedNodes); |
6041 | |
6042 | // Directly insert the node in question, and legalize it. This will recurse |
6043 | // as needed through operands. |
6044 | LegalizedNodes.insert(Ptr: N); |
6045 | Legalizer.LegalizeOp(Node: N); |
6046 | |
6047 | return LegalizedNodes.count(Ptr: N); |
6048 | } |
6049 | |