1 | //===-- PPCFastISel.cpp - PowerPC FastISel implementation -----------------===// |
2 | // |
3 | // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. |
4 | // See https://llvm.org/LICENSE.txt for license information. |
5 | // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception |
6 | // |
7 | //===----------------------------------------------------------------------===// |
8 | // |
9 | // This file defines the PowerPC-specific support for the FastISel class. Some |
10 | // of the target-specific code is generated by tablegen in the file |
11 | // PPCGenFastISel.inc, which is #included here. |
12 | // |
13 | //===----------------------------------------------------------------------===// |
14 | |
15 | #include "MCTargetDesc/PPCPredicates.h" |
16 | #include "PPC.h" |
17 | #include "PPCCallingConv.h" |
18 | #include "PPCISelLowering.h" |
19 | #include "PPCMachineFunctionInfo.h" |
20 | #include "PPCSubtarget.h" |
21 | #include "llvm/CodeGen/CallingConvLower.h" |
22 | #include "llvm/CodeGen/FastISel.h" |
23 | #include "llvm/CodeGen/FunctionLoweringInfo.h" |
24 | #include "llvm/CodeGen/MachineConstantPool.h" |
25 | #include "llvm/CodeGen/MachineFrameInfo.h" |
26 | #include "llvm/CodeGen/MachineInstrBuilder.h" |
27 | #include "llvm/CodeGen/MachineRegisterInfo.h" |
28 | #include "llvm/CodeGen/TargetLowering.h" |
29 | #include "llvm/IR/CallingConv.h" |
30 | #include "llvm/IR/GetElementPtrTypeIterator.h" |
31 | #include "llvm/IR/GlobalVariable.h" |
32 | #include "llvm/IR/Operator.h" |
33 | #include "llvm/Target/TargetMachine.h" |
34 | |
35 | //===----------------------------------------------------------------------===// |
36 | // |
37 | // TBD: |
38 | // fastLowerArguments: Handle simple cases. |
39 | // PPCMaterializeGV: Handle TLS. |
40 | // SelectCall: Handle function pointers. |
41 | // SelectCall: Handle multi-register return values. |
42 | // SelectCall: Optimize away nops for local calls. |
43 | // processCallArgs: Handle bit-converted arguments. |
44 | // finishCall: Handle multi-register return values. |
45 | // PPCComputeAddress: Handle parameter references as FrameIndex's. |
46 | // PPCEmitCmp: Handle immediate as operand 1. |
47 | // SelectCall: Handle small byval arguments. |
48 | // SelectIntrinsicCall: Implement. |
49 | // SelectSelect: Implement. |
50 | // Consider factoring isTypeLegal into the base class. |
51 | // Implement switches and jump tables. |
52 | // |
53 | //===----------------------------------------------------------------------===// |
54 | using namespace llvm; |
55 | |
56 | #define DEBUG_TYPE "ppcfastisel" |
57 | |
58 | namespace { |
59 | |
60 | struct Address { |
61 | enum { |
62 | RegBase, |
63 | FrameIndexBase |
64 | } BaseType; |
65 | |
66 | union { |
67 | unsigned Reg; |
68 | int FI; |
69 | } Base; |
70 | |
71 | int64_t Offset; |
72 | |
73 | // Innocuous defaults for our address. |
74 | Address() |
75 | : BaseType(RegBase), Offset(0) { |
76 | Base.Reg = 0; |
77 | } |
78 | }; |
79 | |
80 | class PPCFastISel final : public FastISel { |
81 | |
82 | const TargetMachine &TM; |
83 | const PPCSubtarget *Subtarget; |
84 | PPCFunctionInfo *PPCFuncInfo; |
85 | const TargetInstrInfo &TII; |
86 | const TargetLowering &TLI; |
87 | LLVMContext *Context; |
88 | |
89 | public: |
90 | explicit PPCFastISel(FunctionLoweringInfo &FuncInfo, |
91 | const TargetLibraryInfo *LibInfo) |
92 | : FastISel(FuncInfo, LibInfo), TM(FuncInfo.MF->getTarget()), |
93 | Subtarget(&FuncInfo.MF->getSubtarget<PPCSubtarget>()), |
94 | PPCFuncInfo(FuncInfo.MF->getInfo<PPCFunctionInfo>()), |
95 | TII(*Subtarget->getInstrInfo()), TLI(*Subtarget->getTargetLowering()), |
96 | Context(&FuncInfo.Fn->getContext()) {} |
97 | |
98 | // Backend specific FastISel code. |
99 | private: |
100 | bool fastSelectInstruction(const Instruction *I) override; |
101 | Register fastMaterializeConstant(const Constant *C) override; |
102 | Register fastMaterializeAlloca(const AllocaInst *AI) override; |
103 | bool tryToFoldLoadIntoMI(MachineInstr *MI, unsigned OpNo, |
104 | const LoadInst *LI) override; |
105 | bool fastLowerArguments() override; |
106 | Register fastEmit_i(MVT Ty, MVT RetTy, unsigned Opc, uint64_t Imm) override; |
107 | Register fastEmitInst_ri(unsigned MachineInstOpcode, |
108 | const TargetRegisterClass *RC, Register Op0, |
109 | uint64_t Imm); |
110 | Register fastEmitInst_r(unsigned MachineInstOpcode, |
111 | const TargetRegisterClass *RC, Register Op0); |
112 | Register fastEmitInst_rr(unsigned MachineInstOpcode, |
113 | const TargetRegisterClass *RC, Register Op0, |
114 | Register Op1); |
115 | |
116 | bool fastLowerCall(CallLoweringInfo &CLI) override; |
117 | |
118 | // Instruction selection routines. |
119 | private: |
120 | bool SelectLoad(const Instruction *I); |
121 | bool SelectStore(const Instruction *I); |
122 | bool SelectBranch(const Instruction *I); |
123 | bool SelectIndirectBr(const Instruction *I); |
124 | bool SelectFPExt(const Instruction *I); |
125 | bool SelectFPTrunc(const Instruction *I); |
126 | bool SelectIToFP(const Instruction *I, bool IsSigned); |
127 | bool SelectFPToI(const Instruction *I, bool IsSigned); |
128 | bool SelectBinaryIntOp(const Instruction *I, unsigned ISDOpcode); |
129 | bool SelectRet(const Instruction *I); |
130 | bool SelectTrunc(const Instruction *I); |
131 | bool SelectIntExt(const Instruction *I); |
132 | |
133 | // Utility routines. |
134 | private: |
135 | bool isTypeLegal(Type *Ty, MVT &VT); |
136 | bool isLoadTypeLegal(Type *Ty, MVT &VT); |
137 | bool isValueAvailable(const Value *V) const; |
138 | bool isVSFRCRegClass(const TargetRegisterClass *RC) const { |
139 | return RC->getID() == PPC::VSFRCRegClassID; |
140 | } |
141 | bool isVSSRCRegClass(const TargetRegisterClass *RC) const { |
142 | return RC->getID() == PPC::VSSRCRegClassID; |
143 | } |
144 | Register copyRegToRegClass(const TargetRegisterClass *ToRC, Register SrcReg, |
145 | unsigned Flag = 0, unsigned SubReg = 0) { |
146 | Register TmpReg = createResultReg(RC: ToRC); |
147 | BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, |
148 | MCID: TII.get(Opcode: TargetOpcode::COPY), DestReg: TmpReg).addReg(RegNo: SrcReg, flags: Flag, SubReg); |
149 | return TmpReg; |
150 | } |
151 | bool PPCEmitCmp(const Value *Src1Value, const Value *Src2Value, bool isZExt, |
152 | Register DestReg, const PPC::Predicate Pred); |
153 | bool PPCEmitLoad(MVT VT, Register &ResultReg, Address &Addr, |
154 | const TargetRegisterClass *RC, bool IsZExt = true, |
155 | unsigned FP64LoadOpc = PPC::LFD); |
156 | bool PPCEmitStore(MVT VT, Register SrcReg, Address &Addr); |
157 | bool PPCComputeAddress(const Value *Obj, Address &Addr); |
158 | void PPCSimplifyAddress(Address &Addr, bool &UseOffset, Register &IndexReg); |
159 | bool PPCEmitIntExt(MVT SrcVT, Register SrcReg, MVT DestVT, Register DestReg, |
160 | bool IsZExt); |
161 | Register PPCMaterializeFP(const ConstantFP *CFP, MVT VT); |
162 | Register PPCMaterializeGV(const GlobalValue *GV, MVT VT); |
163 | Register PPCMaterializeInt(const ConstantInt *CI, MVT VT, |
164 | bool UseSExt = true); |
165 | Register PPCMaterialize32BitInt(int64_t Imm, const TargetRegisterClass *RC); |
166 | Register PPCMaterialize64BitInt(int64_t Imm, const TargetRegisterClass *RC); |
167 | Register PPCMoveToIntReg(const Instruction *I, MVT VT, Register SrcReg, |
168 | bool IsSigned); |
169 | Register PPCMoveToFPReg(MVT VT, Register SrcReg, bool IsSigned); |
170 | |
171 | // Call handling routines. |
172 | private: |
173 | bool processCallArgs(SmallVectorImpl<Value *> &Args, |
174 | SmallVectorImpl<Register> &ArgRegs, |
175 | SmallVectorImpl<MVT> &ArgVTs, |
176 | SmallVectorImpl<ISD::ArgFlagsTy> &ArgFlags, |
177 | SmallVectorImpl<unsigned> &RegArgs, CallingConv::ID CC, |
178 | unsigned &NumBytes, bool IsVarArg); |
179 | bool finishCall(MVT RetVT, CallLoweringInfo &CLI, unsigned &NumBytes); |
180 | |
181 | private: |
182 | #include "PPCGenFastISel.inc" |
183 | |
184 | }; |
185 | |
186 | } // end anonymous namespace |
187 | |
188 | static std::optional<PPC::Predicate> getComparePred(CmpInst::Predicate Pred) { |
189 | switch (Pred) { |
190 | // These are not representable with any single compare. |
191 | case CmpInst::FCMP_FALSE: |
192 | case CmpInst::FCMP_TRUE: |
193 | // Major concern about the following 6 cases is NaN result. The comparison |
194 | // result consists of 4 bits, indicating lt, eq, gt and un (unordered), |
195 | // only one of which will be set. The result is generated by fcmpu |
196 | // instruction. However, bc instruction only inspects one of the first 3 |
197 | // bits, so when un is set, bc instruction may jump to an undesired |
198 | // place. |
199 | // |
200 | // More specifically, if we expect an unordered comparison and un is set, we |
201 | // expect to always go to true branch; in such case UEQ, UGT and ULT still |
202 | // give false, which are undesired; but UNE, UGE, ULE happen to give true, |
203 | // since they are tested by inspecting !eq, !lt, !gt, respectively. |
204 | // |
205 | // Similarly, for ordered comparison, when un is set, we always expect the |
206 | // result to be false. In such case OGT, OLT and OEQ is good, since they are |
207 | // actually testing GT, LT, and EQ respectively, which are false. OGE, OLE |
208 | // and ONE are tested through !lt, !gt and !eq, and these are true. |
209 | case CmpInst::FCMP_UEQ: |
210 | case CmpInst::FCMP_UGT: |
211 | case CmpInst::FCMP_ULT: |
212 | case CmpInst::FCMP_OGE: |
213 | case CmpInst::FCMP_OLE: |
214 | case CmpInst::FCMP_ONE: |
215 | default: |
216 | return std::nullopt; |
217 | |
218 | case CmpInst::FCMP_OEQ: |
219 | case CmpInst::ICMP_EQ: |
220 | return PPC::PRED_EQ; |
221 | |
222 | case CmpInst::FCMP_OGT: |
223 | case CmpInst::ICMP_UGT: |
224 | case CmpInst::ICMP_SGT: |
225 | return PPC::PRED_GT; |
226 | |
227 | case CmpInst::FCMP_UGE: |
228 | case CmpInst::ICMP_UGE: |
229 | case CmpInst::ICMP_SGE: |
230 | return PPC::PRED_GE; |
231 | |
232 | case CmpInst::FCMP_OLT: |
233 | case CmpInst::ICMP_ULT: |
234 | case CmpInst::ICMP_SLT: |
235 | return PPC::PRED_LT; |
236 | |
237 | case CmpInst::FCMP_ULE: |
238 | case CmpInst::ICMP_ULE: |
239 | case CmpInst::ICMP_SLE: |
240 | return PPC::PRED_LE; |
241 | |
242 | case CmpInst::FCMP_UNE: |
243 | case CmpInst::ICMP_NE: |
244 | return PPC::PRED_NE; |
245 | |
246 | case CmpInst::FCMP_ORD: |
247 | return PPC::PRED_NU; |
248 | |
249 | case CmpInst::FCMP_UNO: |
250 | return PPC::PRED_UN; |
251 | } |
252 | } |
253 | |
254 | // Determine whether the type Ty is simple enough to be handled by |
255 | // fast-isel, and return its equivalent machine type in VT. |
256 | // FIXME: Copied directly from ARM -- factor into base class? |
257 | bool PPCFastISel::isTypeLegal(Type *Ty, MVT &VT) { |
258 | EVT Evt = TLI.getValueType(DL, Ty, AllowUnknown: true); |
259 | |
260 | // Only handle simple types. |
261 | if (Evt == MVT::Other || !Evt.isSimple()) return false; |
262 | VT = Evt.getSimpleVT(); |
263 | |
264 | // Handle all legal types, i.e. a register that will directly hold this |
265 | // value. |
266 | return TLI.isTypeLegal(VT); |
267 | } |
268 | |
269 | // Determine whether the type Ty is simple enough to be handled by |
270 | // fast-isel as a load target, and return its equivalent machine type in VT. |
271 | bool PPCFastISel::isLoadTypeLegal(Type *Ty, MVT &VT) { |
272 | if (isTypeLegal(Ty, VT)) return true; |
273 | |
274 | // If this is a type than can be sign or zero-extended to a basic operation |
275 | // go ahead and accept it now. |
276 | if (VT == MVT::i8 || VT == MVT::i16 || VT == MVT::i32) { |
277 | return true; |
278 | } |
279 | |
280 | return false; |
281 | } |
282 | |
283 | bool PPCFastISel::isValueAvailable(const Value *V) const { |
284 | if (!isa<Instruction>(Val: V)) |
285 | return true; |
286 | |
287 | const auto *I = cast<Instruction>(Val: V); |
288 | return FuncInfo.getMBB(BB: I->getParent()) == FuncInfo.MBB; |
289 | } |
290 | |
291 | // Given a value Obj, create an Address object Addr that represents its |
292 | // address. Return false if we can't handle it. |
293 | bool PPCFastISel::PPCComputeAddress(const Value *Obj, Address &Addr) { |
294 | const User *U = nullptr; |
295 | unsigned Opcode = Instruction::UserOp1; |
296 | if (const Instruction *I = dyn_cast<Instruction>(Val: Obj)) { |
297 | // Don't walk into other basic blocks unless the object is an alloca from |
298 | // another block, otherwise it may not have a virtual register assigned. |
299 | if (FuncInfo.StaticAllocaMap.count(Val: static_cast<const AllocaInst *>(Obj)) || |
300 | FuncInfo.getMBB(BB: I->getParent()) == FuncInfo.MBB) { |
301 | Opcode = I->getOpcode(); |
302 | U = I; |
303 | } |
304 | } else if (const ConstantExpr *C = dyn_cast<ConstantExpr>(Val: Obj)) { |
305 | Opcode = C->getOpcode(); |
306 | U = C; |
307 | } |
308 | |
309 | switch (Opcode) { |
310 | default: |
311 | break; |
312 | case Instruction::BitCast: |
313 | // Look through bitcasts. |
314 | return PPCComputeAddress(Obj: U->getOperand(i: 0), Addr); |
315 | case Instruction::IntToPtr: |
316 | // Look past no-op inttoptrs. |
317 | if (TLI.getValueType(DL, Ty: U->getOperand(i: 0)->getType()) == |
318 | TLI.getPointerTy(DL)) |
319 | return PPCComputeAddress(Obj: U->getOperand(i: 0), Addr); |
320 | break; |
321 | case Instruction::PtrToInt: |
322 | // Look past no-op ptrtoints. |
323 | if (TLI.getValueType(DL, Ty: U->getType()) == TLI.getPointerTy(DL)) |
324 | return PPCComputeAddress(Obj: U->getOperand(i: 0), Addr); |
325 | break; |
326 | case Instruction::GetElementPtr: { |
327 | Address SavedAddr = Addr; |
328 | int64_t TmpOffset = Addr.Offset; |
329 | |
330 | // Iterate through the GEP folding the constants into offsets where |
331 | // we can. |
332 | gep_type_iterator GTI = gep_type_begin(GEP: U); |
333 | for (User::const_op_iterator II = U->op_begin() + 1, IE = U->op_end(); |
334 | II != IE; ++II, ++GTI) { |
335 | const Value *Op = *II; |
336 | if (StructType *STy = GTI.getStructTypeOrNull()) { |
337 | const StructLayout *SL = DL.getStructLayout(Ty: STy); |
338 | unsigned Idx = cast<ConstantInt>(Val: Op)->getZExtValue(); |
339 | TmpOffset += SL->getElementOffset(Idx); |
340 | } else { |
341 | uint64_t S = GTI.getSequentialElementStride(DL); |
342 | for (;;) { |
343 | if (const ConstantInt *CI = dyn_cast<ConstantInt>(Val: Op)) { |
344 | // Constant-offset addressing. |
345 | TmpOffset += CI->getSExtValue() * S; |
346 | break; |
347 | } |
348 | if (canFoldAddIntoGEP(GEP: U, Add: Op)) { |
349 | // A compatible add with a constant operand. Fold the constant. |
350 | ConstantInt *CI = |
351 | cast<ConstantInt>(Val: cast<AddOperator>(Val: Op)->getOperand(i_nocapture: 1)); |
352 | TmpOffset += CI->getSExtValue() * S; |
353 | // Iterate on the other operand. |
354 | Op = cast<AddOperator>(Val: Op)->getOperand(i_nocapture: 0); |
355 | continue; |
356 | } |
357 | // Unsupported |
358 | goto unsupported_gep; |
359 | } |
360 | } |
361 | } |
362 | |
363 | // Try to grab the base operand now. |
364 | Addr.Offset = TmpOffset; |
365 | if (PPCComputeAddress(Obj: U->getOperand(i: 0), Addr)) return true; |
366 | |
367 | // We failed, restore everything and try the other options. |
368 | Addr = SavedAddr; |
369 | |
370 | unsupported_gep: |
371 | break; |
372 | } |
373 | case Instruction::Alloca: { |
374 | const AllocaInst *AI = cast<AllocaInst>(Val: Obj); |
375 | DenseMap<const AllocaInst*, int>::iterator SI = |
376 | FuncInfo.StaticAllocaMap.find(Val: AI); |
377 | if (SI != FuncInfo.StaticAllocaMap.end()) { |
378 | Addr.BaseType = Address::FrameIndexBase; |
379 | Addr.Base.FI = SI->second; |
380 | return true; |
381 | } |
382 | break; |
383 | } |
384 | } |
385 | |
386 | // FIXME: References to parameters fall through to the behavior |
387 | // below. They should be able to reference a frame index since |
388 | // they are stored to the stack, so we can get "ld rx, offset(r1)" |
389 | // instead of "addi ry, r1, offset / ld rx, 0(ry)". Obj will |
390 | // just contain the parameter. Try to handle this with a FI. |
391 | |
392 | // Try to get this in a register if nothing else has worked. |
393 | if (Addr.Base.Reg == 0) |
394 | Addr.Base.Reg = getRegForValue(V: Obj); |
395 | |
396 | // Prevent assignment of base register to X0, which is inappropriate |
397 | // for loads and stores alike. |
398 | if (Addr.Base.Reg != 0) |
399 | MRI.setRegClass(Reg: Addr.Base.Reg, RC: &PPC::G8RC_and_G8RC_NOX0RegClass); |
400 | |
401 | return Addr.Base.Reg != 0; |
402 | } |
403 | |
404 | // Fix up some addresses that can't be used directly. For example, if |
405 | // an offset won't fit in an instruction field, we may need to move it |
406 | // into an index register. |
407 | void PPCFastISel::PPCSimplifyAddress(Address &Addr, bool &UseOffset, |
408 | Register &IndexReg) { |
409 | |
410 | // Check whether the offset fits in the instruction field. |
411 | if (!isInt<16>(x: Addr.Offset)) |
412 | UseOffset = false; |
413 | |
414 | // If this is a stack pointer and the offset needs to be simplified then |
415 | // put the alloca address into a register, set the base type back to |
416 | // register and continue. This should almost never happen. |
417 | if (!UseOffset && Addr.BaseType == Address::FrameIndexBase) { |
418 | Register ResultReg = createResultReg(RC: &PPC::G8RC_and_G8RC_NOX0RegClass); |
419 | BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, MCID: TII.get(Opcode: PPC::ADDI8), |
420 | DestReg: ResultReg).addFrameIndex(Idx: Addr.Base.FI).addImm(Val: 0); |
421 | Addr.Base.Reg = ResultReg; |
422 | Addr.BaseType = Address::RegBase; |
423 | } |
424 | |
425 | if (!UseOffset) { |
426 | IntegerType *OffsetTy = Type::getInt64Ty(C&: *Context); |
427 | const ConstantInt *Offset = ConstantInt::getSigned(Ty: OffsetTy, V: Addr.Offset); |
428 | IndexReg = PPCMaterializeInt(CI: Offset, VT: MVT::i64); |
429 | assert(IndexReg && "Unexpected error in PPCMaterializeInt!" ); |
430 | } |
431 | } |
432 | |
433 | // Emit a load instruction if possible, returning true if we succeeded, |
434 | // otherwise false. See commentary below for how the register class of |
435 | // the load is determined. |
436 | bool PPCFastISel::PPCEmitLoad(MVT VT, Register &ResultReg, Address &Addr, |
437 | const TargetRegisterClass *RC, |
438 | bool IsZExt, unsigned FP64LoadOpc) { |
439 | unsigned Opc; |
440 | bool UseOffset = true; |
441 | bool HasSPE = Subtarget->hasSPE(); |
442 | |
443 | // If ResultReg is given, it determines the register class of the load. |
444 | // Otherwise, RC is the register class to use. If the result of the |
445 | // load isn't anticipated in this block, both may be zero, in which |
446 | // case we must make a conservative guess. In particular, don't assign |
447 | // R0 or X0 to the result register, as the result may be used in a load, |
448 | // store, add-immediate, or isel that won't permit this. (Though |
449 | // perhaps the spill and reload of live-exit values would handle this?) |
450 | const TargetRegisterClass *UseRC = |
451 | (ResultReg ? MRI.getRegClass(Reg: ResultReg) : |
452 | (RC ? RC : |
453 | (VT == MVT::f64 ? (HasSPE ? &PPC::SPERCRegClass : &PPC::F8RCRegClass) : |
454 | (VT == MVT::f32 ? (HasSPE ? &PPC::GPRCRegClass : &PPC::F4RCRegClass) : |
455 | (VT == MVT::i64 ? &PPC::G8RC_and_G8RC_NOX0RegClass : |
456 | &PPC::GPRC_and_GPRC_NOR0RegClass))))); |
457 | |
458 | bool Is32BitInt = UseRC->hasSuperClassEq(RC: &PPC::GPRCRegClass); |
459 | |
460 | switch (VT.SimpleTy) { |
461 | default: // e.g., vector types not handled |
462 | return false; |
463 | case MVT::i8: |
464 | Opc = Is32BitInt ? PPC::LBZ : PPC::LBZ8; |
465 | break; |
466 | case MVT::i16: |
467 | Opc = (IsZExt ? (Is32BitInt ? PPC::LHZ : PPC::LHZ8) |
468 | : (Is32BitInt ? PPC::LHA : PPC::LHA8)); |
469 | break; |
470 | case MVT::i32: |
471 | Opc = (IsZExt ? (Is32BitInt ? PPC::LWZ : PPC::LWZ8) |
472 | : (Is32BitInt ? PPC::LWA_32 : PPC::LWA)); |
473 | if ((Opc == PPC::LWA || Opc == PPC::LWA_32) && ((Addr.Offset & 3) != 0)) |
474 | UseOffset = false; |
475 | break; |
476 | case MVT::i64: |
477 | Opc = PPC::LD; |
478 | assert(UseRC->hasSuperClassEq(&PPC::G8RCRegClass) && |
479 | "64-bit load with 32-bit target??" ); |
480 | UseOffset = ((Addr.Offset & 3) == 0); |
481 | break; |
482 | case MVT::f32: |
483 | Opc = Subtarget->hasSPE() ? PPC::SPELWZ : PPC::LFS; |
484 | break; |
485 | case MVT::f64: |
486 | Opc = FP64LoadOpc; |
487 | break; |
488 | } |
489 | |
490 | // If necessary, materialize the offset into a register and use |
491 | // the indexed form. Also handle stack pointers with special needs. |
492 | Register IndexReg; |
493 | PPCSimplifyAddress(Addr, UseOffset, IndexReg); |
494 | |
495 | // If this is a potential VSX load with an offset of 0, a VSX indexed load can |
496 | // be used. |
497 | bool IsVSSRC = isVSSRCRegClass(RC: UseRC); |
498 | bool IsVSFRC = isVSFRCRegClass(RC: UseRC); |
499 | bool Is32VSXLoad = IsVSSRC && Opc == PPC::LFS; |
500 | bool Is64VSXLoad = IsVSFRC && Opc == PPC::LFD; |
501 | if ((Is32VSXLoad || Is64VSXLoad) && |
502 | (Addr.BaseType != Address::FrameIndexBase) && UseOffset && |
503 | (Addr.Offset == 0)) { |
504 | UseOffset = false; |
505 | } |
506 | |
507 | if (!ResultReg) |
508 | ResultReg = createResultReg(RC: UseRC); |
509 | |
510 | // Note: If we still have a frame index here, we know the offset is |
511 | // in range, as otherwise PPCSimplifyAddress would have converted it |
512 | // into a RegBase. |
513 | if (Addr.BaseType == Address::FrameIndexBase) { |
514 | // VSX only provides an indexed load. |
515 | if (Is32VSXLoad || Is64VSXLoad) return false; |
516 | |
517 | MachineMemOperand *MMO = FuncInfo.MF->getMachineMemOperand( |
518 | PtrInfo: MachinePointerInfo::getFixedStack(MF&: *FuncInfo.MF, FI: Addr.Base.FI, |
519 | Offset: Addr.Offset), |
520 | F: MachineMemOperand::MOLoad, Size: MFI.getObjectSize(ObjectIdx: Addr.Base.FI), |
521 | BaseAlignment: MFI.getObjectAlign(ObjectIdx: Addr.Base.FI)); |
522 | |
523 | BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, MCID: TII.get(Opcode: Opc), DestReg: ResultReg) |
524 | .addImm(Val: Addr.Offset).addFrameIndex(Idx: Addr.Base.FI).addMemOperand(MMO); |
525 | |
526 | // Base reg with offset in range. |
527 | } else if (UseOffset) { |
528 | // VSX only provides an indexed load. |
529 | if (Is32VSXLoad || Is64VSXLoad) return false; |
530 | |
531 | BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, MCID: TII.get(Opcode: Opc), DestReg: ResultReg) |
532 | .addImm(Val: Addr.Offset).addReg(RegNo: Addr.Base.Reg); |
533 | |
534 | // Indexed form. |
535 | } else { |
536 | // Get the RR opcode corresponding to the RI one. FIXME: It would be |
537 | // preferable to use the ImmToIdxMap from PPCRegisterInfo.cpp, but it |
538 | // is hard to get at. |
539 | switch (Opc) { |
540 | default: llvm_unreachable("Unexpected opcode!" ); |
541 | case PPC::LBZ: Opc = PPC::LBZX; break; |
542 | case PPC::LBZ8: Opc = PPC::LBZX8; break; |
543 | case PPC::LHZ: Opc = PPC::LHZX; break; |
544 | case PPC::LHZ8: Opc = PPC::LHZX8; break; |
545 | case PPC::LHA: Opc = PPC::LHAX; break; |
546 | case PPC::LHA8: Opc = PPC::LHAX8; break; |
547 | case PPC::LWZ: Opc = PPC::LWZX; break; |
548 | case PPC::LWZ8: Opc = PPC::LWZX8; break; |
549 | case PPC::LWA: Opc = PPC::LWAX; break; |
550 | case PPC::LWA_32: Opc = PPC::LWAX_32; break; |
551 | case PPC::LD: Opc = PPC::LDX; break; |
552 | case PPC::LFS: Opc = IsVSSRC ? PPC::LXSSPX : PPC::LFSX; break; |
553 | case PPC::LFD: Opc = IsVSFRC ? PPC::LXSDX : PPC::LFDX; break; |
554 | case PPC::EVLDD: Opc = PPC::EVLDDX; break; |
555 | case PPC::SPELWZ: Opc = PPC::SPELWZX; break; |
556 | } |
557 | |
558 | auto MIB = BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, MCID: TII.get(Opcode: Opc), |
559 | DestReg: ResultReg); |
560 | |
561 | // If we have an index register defined we use it in the store inst, |
562 | // otherwise we use X0 as base as it makes the vector instructions to |
563 | // use zero in the computation of the effective address regardless the |
564 | // content of the register. |
565 | if (IndexReg) |
566 | MIB.addReg(RegNo: Addr.Base.Reg).addReg(RegNo: IndexReg); |
567 | else |
568 | MIB.addReg(RegNo: PPC::ZERO8).addReg(RegNo: Addr.Base.Reg); |
569 | } |
570 | |
571 | return true; |
572 | } |
573 | |
574 | // Attempt to fast-select a load instruction. |
575 | bool PPCFastISel::SelectLoad(const Instruction *I) { |
576 | // FIXME: No atomic loads are supported. |
577 | if (cast<LoadInst>(Val: I)->isAtomic()) |
578 | return false; |
579 | |
580 | // Verify we have a legal type before going any further. |
581 | MVT VT; |
582 | if (!isLoadTypeLegal(Ty: I->getType(), VT)) |
583 | return false; |
584 | |
585 | // See if we can handle this address. |
586 | Address Addr; |
587 | if (!PPCComputeAddress(Obj: I->getOperand(i: 0), Addr)) |
588 | return false; |
589 | |
590 | // Look at the currently assigned register for this instruction |
591 | // to determine the required register class. This is necessary |
592 | // to constrain RA from using R0/X0 when this is not legal. |
593 | Register AssignedReg = FuncInfo.ValueMap[I]; |
594 | const TargetRegisterClass *RC = |
595 | AssignedReg ? MRI.getRegClass(Reg: AssignedReg) : nullptr; |
596 | |
597 | Register ResultReg = 0; |
598 | if (!PPCEmitLoad(VT, ResultReg, Addr, RC, IsZExt: true, |
599 | FP64LoadOpc: Subtarget->hasSPE() ? PPC::EVLDD : PPC::LFD)) |
600 | return false; |
601 | updateValueMap(I, Reg: ResultReg); |
602 | return true; |
603 | } |
604 | |
605 | // Emit a store instruction to store SrcReg at Addr. |
606 | bool PPCFastISel::PPCEmitStore(MVT VT, Register SrcReg, Address &Addr) { |
607 | assert(SrcReg && "Nothing to store!" ); |
608 | unsigned Opc; |
609 | bool UseOffset = true; |
610 | |
611 | const TargetRegisterClass *RC = MRI.getRegClass(Reg: SrcReg); |
612 | bool Is32BitInt = RC->hasSuperClassEq(RC: &PPC::GPRCRegClass); |
613 | |
614 | switch (VT.SimpleTy) { |
615 | default: // e.g., vector types not handled |
616 | return false; |
617 | case MVT::i8: |
618 | Opc = Is32BitInt ? PPC::STB : PPC::STB8; |
619 | break; |
620 | case MVT::i16: |
621 | Opc = Is32BitInt ? PPC::STH : PPC::STH8; |
622 | break; |
623 | case MVT::i32: |
624 | assert(Is32BitInt && "Not GPRC for i32??" ); |
625 | Opc = PPC::STW; |
626 | break; |
627 | case MVT::i64: |
628 | Opc = PPC::STD; |
629 | UseOffset = ((Addr.Offset & 3) == 0); |
630 | break; |
631 | case MVT::f32: |
632 | Opc = Subtarget->hasSPE() ? PPC::SPESTW : PPC::STFS; |
633 | break; |
634 | case MVT::f64: |
635 | Opc = Subtarget->hasSPE() ? PPC::EVSTDD : PPC::STFD; |
636 | break; |
637 | } |
638 | |
639 | // If necessary, materialize the offset into a register and use |
640 | // the indexed form. Also handle stack pointers with special needs. |
641 | Register IndexReg; |
642 | PPCSimplifyAddress(Addr, UseOffset, IndexReg); |
643 | |
644 | // If this is a potential VSX store with an offset of 0, a VSX indexed store |
645 | // can be used. |
646 | bool IsVSSRC = isVSSRCRegClass(RC); |
647 | bool IsVSFRC = isVSFRCRegClass(RC); |
648 | bool Is32VSXStore = IsVSSRC && Opc == PPC::STFS; |
649 | bool Is64VSXStore = IsVSFRC && Opc == PPC::STFD; |
650 | if ((Is32VSXStore || Is64VSXStore) && |
651 | (Addr.BaseType != Address::FrameIndexBase) && UseOffset && |
652 | (Addr.Offset == 0)) { |
653 | UseOffset = false; |
654 | } |
655 | |
656 | // Note: If we still have a frame index here, we know the offset is |
657 | // in range, as otherwise PPCSimplifyAddress would have converted it |
658 | // into a RegBase. |
659 | if (Addr.BaseType == Address::FrameIndexBase) { |
660 | // VSX only provides an indexed store. |
661 | if (Is32VSXStore || Is64VSXStore) return false; |
662 | |
663 | MachineMemOperand *MMO = FuncInfo.MF->getMachineMemOperand( |
664 | PtrInfo: MachinePointerInfo::getFixedStack(MF&: *FuncInfo.MF, FI: Addr.Base.FI, |
665 | Offset: Addr.Offset), |
666 | F: MachineMemOperand::MOStore, Size: MFI.getObjectSize(ObjectIdx: Addr.Base.FI), |
667 | BaseAlignment: MFI.getObjectAlign(ObjectIdx: Addr.Base.FI)); |
668 | |
669 | BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, MCID: TII.get(Opcode: Opc)) |
670 | .addReg(RegNo: SrcReg) |
671 | .addImm(Val: Addr.Offset) |
672 | .addFrameIndex(Idx: Addr.Base.FI) |
673 | .addMemOperand(MMO); |
674 | |
675 | // Base reg with offset in range. |
676 | } else if (UseOffset) { |
677 | // VSX only provides an indexed store. |
678 | if (Is32VSXStore || Is64VSXStore) |
679 | return false; |
680 | |
681 | BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, MCID: TII.get(Opcode: Opc)) |
682 | .addReg(RegNo: SrcReg).addImm(Val: Addr.Offset).addReg(RegNo: Addr.Base.Reg); |
683 | |
684 | // Indexed form. |
685 | } else { |
686 | // Get the RR opcode corresponding to the RI one. FIXME: It would be |
687 | // preferable to use the ImmToIdxMap from PPCRegisterInfo.cpp, but it |
688 | // is hard to get at. |
689 | switch (Opc) { |
690 | default: llvm_unreachable("Unexpected opcode!" ); |
691 | case PPC::STB: Opc = PPC::STBX; break; |
692 | case PPC::STH : Opc = PPC::STHX; break; |
693 | case PPC::STW : Opc = PPC::STWX; break; |
694 | case PPC::STB8: Opc = PPC::STBX8; break; |
695 | case PPC::STH8: Opc = PPC::STHX8; break; |
696 | case PPC::STW8: Opc = PPC::STWX8; break; |
697 | case PPC::STD: Opc = PPC::STDX; break; |
698 | case PPC::STFS: Opc = IsVSSRC ? PPC::STXSSPX : PPC::STFSX; break; |
699 | case PPC::STFD: Opc = IsVSFRC ? PPC::STXSDX : PPC::STFDX; break; |
700 | case PPC::EVSTDD: Opc = PPC::EVSTDDX; break; |
701 | case PPC::SPESTW: Opc = PPC::SPESTWX; break; |
702 | } |
703 | |
704 | auto MIB = BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, MCID: TII.get(Opcode: Opc)) |
705 | .addReg(RegNo: SrcReg); |
706 | |
707 | // If we have an index register defined we use it in the store inst, |
708 | // otherwise we use X0 as base as it makes the vector instructions to |
709 | // use zero in the computation of the effective address regardless the |
710 | // content of the register. |
711 | if (IndexReg) |
712 | MIB.addReg(RegNo: Addr.Base.Reg).addReg(RegNo: IndexReg); |
713 | else |
714 | MIB.addReg(RegNo: PPC::ZERO8).addReg(RegNo: Addr.Base.Reg); |
715 | } |
716 | |
717 | return true; |
718 | } |
719 | |
720 | // Attempt to fast-select a store instruction. |
721 | bool PPCFastISel::SelectStore(const Instruction *I) { |
722 | Value *Op0 = I->getOperand(i: 0); |
723 | Register SrcReg; |
724 | |
725 | // FIXME: No atomics loads are supported. |
726 | if (cast<StoreInst>(Val: I)->isAtomic()) |
727 | return false; |
728 | |
729 | // Verify we have a legal type before going any further. |
730 | MVT VT; |
731 | if (!isLoadTypeLegal(Ty: Op0->getType(), VT)) |
732 | return false; |
733 | |
734 | // Get the value to be stored into a register. |
735 | SrcReg = getRegForValue(V: Op0); |
736 | if (!SrcReg) |
737 | return false; |
738 | |
739 | // See if we can handle this address. |
740 | Address Addr; |
741 | if (!PPCComputeAddress(Obj: I->getOperand(i: 1), Addr)) |
742 | return false; |
743 | |
744 | if (!PPCEmitStore(VT, SrcReg, Addr)) |
745 | return false; |
746 | |
747 | return true; |
748 | } |
749 | |
750 | // Attempt to fast-select a branch instruction. |
751 | bool PPCFastISel::SelectBranch(const Instruction *I) { |
752 | const BranchInst *BI = cast<BranchInst>(Val: I); |
753 | MachineBasicBlock *BrBB = FuncInfo.MBB; |
754 | MachineBasicBlock *TBB = FuncInfo.getMBB(BB: BI->getSuccessor(i: 0)); |
755 | MachineBasicBlock *FBB = FuncInfo.getMBB(BB: BI->getSuccessor(i: 1)); |
756 | |
757 | // For now, just try the simplest case where it's fed by a compare. |
758 | if (const CmpInst *CI = dyn_cast<CmpInst>(Val: BI->getCondition())) { |
759 | if (isValueAvailable(V: CI)) { |
760 | std::optional<PPC::Predicate> OptPPCPred = |
761 | getComparePred(Pred: CI->getPredicate()); |
762 | if (!OptPPCPred) |
763 | return false; |
764 | |
765 | PPC::Predicate PPCPred = *OptPPCPred; |
766 | |
767 | // Take advantage of fall-through opportunities. |
768 | if (FuncInfo.MBB->isLayoutSuccessor(MBB: TBB)) { |
769 | std::swap(a&: TBB, b&: FBB); |
770 | PPCPred = PPC::InvertPredicate(Opcode: PPCPred); |
771 | } |
772 | |
773 | Register CondReg = createResultReg(RC: &PPC::CRRCRegClass); |
774 | |
775 | if (!PPCEmitCmp(Src1Value: CI->getOperand(i_nocapture: 0), Src2Value: CI->getOperand(i_nocapture: 1), isZExt: CI->isUnsigned(), |
776 | DestReg: CondReg, Pred: PPCPred)) |
777 | return false; |
778 | |
779 | BuildMI(BB&: *BrBB, I: FuncInfo.InsertPt, MIMD, MCID: TII.get(Opcode: PPC::BCC)) |
780 | .addImm(Val: Subtarget->hasSPE() ? PPC::PRED_SPE : PPCPred) |
781 | .addReg(RegNo: CondReg) |
782 | .addMBB(MBB: TBB); |
783 | finishCondBranch(BranchBB: BI->getParent(), TrueMBB: TBB, FalseMBB: FBB); |
784 | return true; |
785 | } |
786 | } else if (const ConstantInt *CI = |
787 | dyn_cast<ConstantInt>(Val: BI->getCondition())) { |
788 | uint64_t Imm = CI->getZExtValue(); |
789 | MachineBasicBlock *Target = (Imm == 0) ? FBB : TBB; |
790 | fastEmitBranch(MSucc: Target, DbgLoc: MIMD.getDL()); |
791 | return true; |
792 | } |
793 | |
794 | // FIXME: ARM looks for a case where the block containing the compare |
795 | // has been split from the block containing the branch. If this happens, |
796 | // there is a vreg available containing the result of the compare. I'm |
797 | // not sure we can do much, as we've lost the predicate information with |
798 | // the compare instruction -- we have a 4-bit CR but don't know which bit |
799 | // to test here. |
800 | return false; |
801 | } |
802 | |
803 | // Attempt to emit a compare of the two source values. Signed and unsigned |
804 | // comparisons are supported. Return false if we can't handle it. |
805 | bool PPCFastISel::PPCEmitCmp(const Value *SrcValue1, const Value *SrcValue2, |
806 | bool IsZExt, Register DestReg, |
807 | const PPC::Predicate Pred) { |
808 | Type *Ty = SrcValue1->getType(); |
809 | EVT SrcEVT = TLI.getValueType(DL, Ty, AllowUnknown: true); |
810 | if (!SrcEVT.isSimple()) |
811 | return false; |
812 | MVT SrcVT = SrcEVT.getSimpleVT(); |
813 | |
814 | if (SrcVT == MVT::i1 && Subtarget->useCRBits()) |
815 | return false; |
816 | |
817 | // See if operand 2 is an immediate encodeable in the compare. |
818 | // FIXME: Operands are not in canonical order at -O0, so an immediate |
819 | // operand in position 1 is a lost opportunity for now. We are |
820 | // similar to ARM in this regard. |
821 | int64_t Imm = 0; |
822 | bool UseImm = false; |
823 | const bool HasSPE = Subtarget->hasSPE(); |
824 | |
825 | // Only 16-bit integer constants can be represented in compares for |
826 | // PowerPC. Others will be materialized into a register. |
827 | if (const ConstantInt *ConstInt = dyn_cast<ConstantInt>(Val: SrcValue2)) { |
828 | if (SrcVT == MVT::i64 || SrcVT == MVT::i32 || SrcVT == MVT::i16 || |
829 | SrcVT == MVT::i8 || SrcVT == MVT::i1) { |
830 | const APInt &CIVal = ConstInt->getValue(); |
831 | Imm = (IsZExt) ? (int64_t)CIVal.getZExtValue() : |
832 | (int64_t)CIVal.getSExtValue(); |
833 | if ((IsZExt && isUInt<16>(x: Imm)) || (!IsZExt && isInt<16>(x: Imm))) |
834 | UseImm = true; |
835 | } |
836 | } |
837 | |
838 | Register SrcReg1 = getRegForValue(V: SrcValue1); |
839 | if (!SrcReg1) |
840 | return false; |
841 | |
842 | Register SrcReg2; |
843 | if (!UseImm) { |
844 | SrcReg2 = getRegForValue(V: SrcValue2); |
845 | if (!SrcReg2) |
846 | return false; |
847 | } |
848 | |
849 | unsigned CmpOpc; |
850 | bool NeedsExt = false; |
851 | |
852 | auto RC1 = MRI.getRegClass(Reg: SrcReg1); |
853 | auto RC2 = SrcReg2 != 0 ? MRI.getRegClass(Reg: SrcReg2) : nullptr; |
854 | |
855 | switch (SrcVT.SimpleTy) { |
856 | default: return false; |
857 | case MVT::f32: |
858 | if (HasSPE) { |
859 | switch (Pred) { |
860 | default: return false; |
861 | case PPC::PRED_EQ: |
862 | CmpOpc = PPC::EFSCMPEQ; |
863 | break; |
864 | case PPC::PRED_LT: |
865 | CmpOpc = PPC::EFSCMPLT; |
866 | break; |
867 | case PPC::PRED_GT: |
868 | CmpOpc = PPC::EFSCMPGT; |
869 | break; |
870 | } |
871 | } else { |
872 | CmpOpc = PPC::FCMPUS; |
873 | if (isVSSRCRegClass(RC: RC1)) |
874 | SrcReg1 = copyRegToRegClass(ToRC: &PPC::F4RCRegClass, SrcReg: SrcReg1); |
875 | if (RC2 && isVSSRCRegClass(RC: RC2)) |
876 | SrcReg2 = copyRegToRegClass(ToRC: &PPC::F4RCRegClass, SrcReg: SrcReg2); |
877 | } |
878 | break; |
879 | case MVT::f64: |
880 | if (HasSPE) { |
881 | switch (Pred) { |
882 | default: return false; |
883 | case PPC::PRED_EQ: |
884 | CmpOpc = PPC::EFDCMPEQ; |
885 | break; |
886 | case PPC::PRED_LT: |
887 | CmpOpc = PPC::EFDCMPLT; |
888 | break; |
889 | case PPC::PRED_GT: |
890 | CmpOpc = PPC::EFDCMPGT; |
891 | break; |
892 | } |
893 | } else if (isVSFRCRegClass(RC: RC1) || (RC2 && isVSFRCRegClass(RC: RC2))) { |
894 | CmpOpc = PPC::XSCMPUDP; |
895 | } else { |
896 | CmpOpc = PPC::FCMPUD; |
897 | } |
898 | break; |
899 | case MVT::i1: |
900 | case MVT::i8: |
901 | case MVT::i16: |
902 | NeedsExt = true; |
903 | [[fallthrough]]; |
904 | case MVT::i32: |
905 | if (!UseImm) |
906 | CmpOpc = IsZExt ? PPC::CMPLW : PPC::CMPW; |
907 | else |
908 | CmpOpc = IsZExt ? PPC::CMPLWI : PPC::CMPWI; |
909 | break; |
910 | case MVT::i64: |
911 | if (!UseImm) |
912 | CmpOpc = IsZExt ? PPC::CMPLD : PPC::CMPD; |
913 | else |
914 | CmpOpc = IsZExt ? PPC::CMPLDI : PPC::CMPDI; |
915 | break; |
916 | } |
917 | |
918 | if (NeedsExt) { |
919 | Register ExtReg = createResultReg(RC: &PPC::GPRCRegClass); |
920 | if (!PPCEmitIntExt(SrcVT, SrcReg: SrcReg1, DestVT: MVT::i32, DestReg: ExtReg, IsZExt)) |
921 | return false; |
922 | SrcReg1 = ExtReg; |
923 | |
924 | if (!UseImm) { |
925 | Register ExtReg = createResultReg(RC: &PPC::GPRCRegClass); |
926 | if (!PPCEmitIntExt(SrcVT, SrcReg: SrcReg2, DestVT: MVT::i32, DestReg: ExtReg, IsZExt)) |
927 | return false; |
928 | SrcReg2 = ExtReg; |
929 | } |
930 | } |
931 | |
932 | if (!UseImm) |
933 | BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, MCID: TII.get(Opcode: CmpOpc), DestReg) |
934 | .addReg(RegNo: SrcReg1).addReg(RegNo: SrcReg2); |
935 | else |
936 | BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, MCID: TII.get(Opcode: CmpOpc), DestReg) |
937 | .addReg(RegNo: SrcReg1).addImm(Val: Imm); |
938 | |
939 | return true; |
940 | } |
941 | |
942 | // Attempt to fast-select a floating-point extend instruction. |
943 | bool PPCFastISel::SelectFPExt(const Instruction *I) { |
944 | Value *Src = I->getOperand(i: 0); |
945 | EVT SrcVT = TLI.getValueType(DL, Ty: Src->getType(), AllowUnknown: true); |
946 | EVT DestVT = TLI.getValueType(DL, Ty: I->getType(), AllowUnknown: true); |
947 | |
948 | if (SrcVT != MVT::f32 || DestVT != MVT::f64) |
949 | return false; |
950 | |
951 | Register SrcReg = getRegForValue(V: Src); |
952 | if (!SrcReg) |
953 | return false; |
954 | |
955 | // No code is generated for a FP extend. |
956 | updateValueMap(I, Reg: SrcReg); |
957 | return true; |
958 | } |
959 | |
960 | // Attempt to fast-select a floating-point truncate instruction. |
961 | bool PPCFastISel::SelectFPTrunc(const Instruction *I) { |
962 | Value *Src = I->getOperand(i: 0); |
963 | EVT SrcVT = TLI.getValueType(DL, Ty: Src->getType(), AllowUnknown: true); |
964 | EVT DestVT = TLI.getValueType(DL, Ty: I->getType(), AllowUnknown: true); |
965 | |
966 | if (SrcVT != MVT::f64 || DestVT != MVT::f32) |
967 | return false; |
968 | |
969 | Register SrcReg = getRegForValue(V: Src); |
970 | if (!SrcReg) |
971 | return false; |
972 | |
973 | // Round the result to single precision. |
974 | Register DestReg; |
975 | auto RC = MRI.getRegClass(Reg: SrcReg); |
976 | if (Subtarget->hasSPE()) { |
977 | DestReg = createResultReg(RC: &PPC::GPRCRegClass); |
978 | BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, MCID: TII.get(Opcode: PPC::EFSCFD), |
979 | DestReg) |
980 | .addReg(RegNo: SrcReg); |
981 | } else if (Subtarget->hasP8Vector() && isVSFRCRegClass(RC)) { |
982 | DestReg = createResultReg(RC: &PPC::VSSRCRegClass); |
983 | BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, MCID: TII.get(Opcode: PPC::XSRSP), |
984 | DestReg) |
985 | .addReg(RegNo: SrcReg); |
986 | } else { |
987 | SrcReg = copyRegToRegClass(ToRC: &PPC::F8RCRegClass, SrcReg); |
988 | DestReg = createResultReg(RC: &PPC::F4RCRegClass); |
989 | BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, |
990 | MCID: TII.get(Opcode: PPC::FRSP), DestReg) |
991 | .addReg(RegNo: SrcReg); |
992 | } |
993 | |
994 | updateValueMap(I, Reg: DestReg); |
995 | return true; |
996 | } |
997 | |
998 | // Move an i32 or i64 value in a GPR to an f64 value in an FPR. |
999 | // FIXME: When direct register moves are implemented (see PowerISA 2.07), |
1000 | // those should be used instead of moving via a stack slot when the |
1001 | // subtarget permits. |
1002 | // FIXME: The code here is sloppy for the 4-byte case. Can use a 4-byte |
1003 | // stack slot and 4-byte store/load sequence. Or just sext the 4-byte |
1004 | // case to 8 bytes which produces tighter code but wastes stack space. |
1005 | Register PPCFastISel::PPCMoveToFPReg(MVT SrcVT, Register SrcReg, |
1006 | bool IsSigned) { |
1007 | |
1008 | // If necessary, extend 32-bit int to 64-bit. |
1009 | if (SrcVT == MVT::i32) { |
1010 | Register TmpReg = createResultReg(RC: &PPC::G8RCRegClass); |
1011 | if (!PPCEmitIntExt(SrcVT: MVT::i32, SrcReg, DestVT: MVT::i64, DestReg: TmpReg, IsZExt: !IsSigned)) |
1012 | return Register(); |
1013 | SrcReg = TmpReg; |
1014 | } |
1015 | |
1016 | // Get a stack slot 8 bytes wide, aligned on an 8-byte boundary. |
1017 | Address Addr; |
1018 | Addr.BaseType = Address::FrameIndexBase; |
1019 | Addr.Base.FI = MFI.CreateStackObject(Size: 8, Alignment: Align(8), isSpillSlot: false); |
1020 | |
1021 | // Store the value from the GPR. |
1022 | if (!PPCEmitStore(VT: MVT::i64, SrcReg, Addr)) |
1023 | return Register(); |
1024 | |
1025 | // Load the integer value into an FPR. The kind of load used depends |
1026 | // on a number of conditions. |
1027 | unsigned LoadOpc = PPC::LFD; |
1028 | |
1029 | if (SrcVT == MVT::i32) { |
1030 | if (!IsSigned) { |
1031 | LoadOpc = PPC::LFIWZX; |
1032 | Addr.Offset = (Subtarget->isLittleEndian()) ? 0 : 4; |
1033 | } else if (Subtarget->hasLFIWAX()) { |
1034 | LoadOpc = PPC::LFIWAX; |
1035 | Addr.Offset = (Subtarget->isLittleEndian()) ? 0 : 4; |
1036 | } |
1037 | } |
1038 | |
1039 | const TargetRegisterClass *RC = &PPC::F8RCRegClass; |
1040 | Register ResultReg; |
1041 | if (!PPCEmitLoad(VT: MVT::f64, ResultReg, Addr, RC, IsZExt: !IsSigned, FP64LoadOpc: LoadOpc)) |
1042 | return Register(); |
1043 | |
1044 | return ResultReg; |
1045 | } |
1046 | |
1047 | // Attempt to fast-select an integer-to-floating-point conversion. |
1048 | // FIXME: Once fast-isel has better support for VSX, conversions using |
1049 | // direct moves should be implemented. |
1050 | bool PPCFastISel::SelectIToFP(const Instruction *I, bool IsSigned) { |
1051 | MVT DstVT; |
1052 | Type *DstTy = I->getType(); |
1053 | if (!isTypeLegal(Ty: DstTy, VT&: DstVT)) |
1054 | return false; |
1055 | |
1056 | if (DstVT != MVT::f32 && DstVT != MVT::f64) |
1057 | return false; |
1058 | |
1059 | Value *Src = I->getOperand(i: 0); |
1060 | EVT SrcEVT = TLI.getValueType(DL, Ty: Src->getType(), AllowUnknown: true); |
1061 | if (!SrcEVT.isSimple()) |
1062 | return false; |
1063 | |
1064 | MVT SrcVT = SrcEVT.getSimpleVT(); |
1065 | |
1066 | if (SrcVT != MVT::i8 && SrcVT != MVT::i16 && |
1067 | SrcVT != MVT::i32 && SrcVT != MVT::i64) |
1068 | return false; |
1069 | |
1070 | Register SrcReg = getRegForValue(V: Src); |
1071 | if (!SrcReg) |
1072 | return false; |
1073 | |
1074 | // Shortcut for SPE. Doesn't need to store/load, since it's all in the GPRs |
1075 | if (Subtarget->hasSPE()) { |
1076 | unsigned Opc; |
1077 | if (DstVT == MVT::f32) |
1078 | Opc = IsSigned ? PPC::EFSCFSI : PPC::EFSCFUI; |
1079 | else |
1080 | Opc = IsSigned ? PPC::EFDCFSI : PPC::EFDCFUI; |
1081 | |
1082 | Register DestReg = createResultReg(RC: &PPC::SPERCRegClass); |
1083 | // Generate the convert. |
1084 | BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, MCID: TII.get(Opcode: Opc), DestReg) |
1085 | .addReg(RegNo: SrcReg); |
1086 | updateValueMap(I, Reg: DestReg); |
1087 | return true; |
1088 | } |
1089 | |
1090 | // We can only lower an unsigned convert if we have the newer |
1091 | // floating-point conversion operations. |
1092 | if (!IsSigned && !Subtarget->hasFPCVT()) |
1093 | return false; |
1094 | |
1095 | // FIXME: For now we require the newer floating-point conversion operations |
1096 | // (which are present only on P7 and A2 server models) when converting |
1097 | // to single-precision float. Otherwise we have to generate a lot of |
1098 | // fiddly code to avoid double rounding. If necessary, the fiddly code |
1099 | // can be found in PPCTargetLowering::LowerINT_TO_FP(). |
1100 | if (DstVT == MVT::f32 && !Subtarget->hasFPCVT()) |
1101 | return false; |
1102 | |
1103 | // Extend the input if necessary. |
1104 | if (SrcVT == MVT::i8 || SrcVT == MVT::i16) { |
1105 | Register TmpReg = createResultReg(RC: &PPC::G8RCRegClass); |
1106 | if (!PPCEmitIntExt(SrcVT, SrcReg, DestVT: MVT::i64, DestReg: TmpReg, IsZExt: !IsSigned)) |
1107 | return false; |
1108 | SrcVT = MVT::i64; |
1109 | SrcReg = TmpReg; |
1110 | } |
1111 | |
1112 | // Move the integer value to an FPR. |
1113 | Register FPReg = PPCMoveToFPReg(SrcVT, SrcReg, IsSigned); |
1114 | if (!FPReg) |
1115 | return false; |
1116 | |
1117 | // Determine the opcode for the conversion. |
1118 | const TargetRegisterClass *RC = &PPC::F8RCRegClass; |
1119 | Register DestReg = createResultReg(RC); |
1120 | unsigned Opc; |
1121 | |
1122 | if (DstVT == MVT::f32) |
1123 | Opc = IsSigned ? PPC::FCFIDS : PPC::FCFIDUS; |
1124 | else |
1125 | Opc = IsSigned ? PPC::FCFID : PPC::FCFIDU; |
1126 | |
1127 | // Generate the convert. |
1128 | BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, MCID: TII.get(Opcode: Opc), DestReg) |
1129 | .addReg(RegNo: FPReg); |
1130 | |
1131 | updateValueMap(I, Reg: DestReg); |
1132 | return true; |
1133 | } |
1134 | |
1135 | // Move the floating-point value in SrcReg into an integer destination |
1136 | // register, and return the register (or zero if we can't handle it). |
1137 | // FIXME: When direct register moves are implemented (see PowerISA 2.07), |
1138 | // those should be used instead of moving via a stack slot when the |
1139 | // subtarget permits. |
1140 | Register PPCFastISel::PPCMoveToIntReg(const Instruction *I, MVT VT, |
1141 | Register SrcReg, bool IsSigned) { |
1142 | // Get a stack slot 8 bytes wide, aligned on an 8-byte boundary. |
1143 | // Note that if have STFIWX available, we could use a 4-byte stack |
1144 | // slot for i32, but this being fast-isel we'll just go with the |
1145 | // easiest code gen possible. |
1146 | Address Addr; |
1147 | Addr.BaseType = Address::FrameIndexBase; |
1148 | Addr.Base.FI = MFI.CreateStackObject(Size: 8, Alignment: Align(8), isSpillSlot: false); |
1149 | |
1150 | // Store the value from the FPR. |
1151 | if (!PPCEmitStore(VT: MVT::f64, SrcReg, Addr)) |
1152 | return Register(); |
1153 | |
1154 | // Reload it into a GPR. If we want an i32 on big endian, modify the |
1155 | // address to have a 4-byte offset so we load from the right place. |
1156 | if (VT == MVT::i32) |
1157 | Addr.Offset = (Subtarget->isLittleEndian()) ? 0 : 4; |
1158 | |
1159 | // Look at the currently assigned register for this instruction |
1160 | // to determine the required register class. |
1161 | Register AssignedReg = FuncInfo.ValueMap[I]; |
1162 | const TargetRegisterClass *RC = |
1163 | AssignedReg ? MRI.getRegClass(Reg: AssignedReg) : nullptr; |
1164 | |
1165 | Register ResultReg; |
1166 | if (!PPCEmitLoad(VT, ResultReg, Addr, RC, IsZExt: !IsSigned)) |
1167 | return Register(); |
1168 | |
1169 | return ResultReg; |
1170 | } |
1171 | |
1172 | // Attempt to fast-select a floating-point-to-integer conversion. |
1173 | // FIXME: Once fast-isel has better support for VSX, conversions using |
1174 | // direct moves should be implemented. |
1175 | bool PPCFastISel::SelectFPToI(const Instruction *I, bool IsSigned) { |
1176 | MVT DstVT, SrcVT; |
1177 | Type *DstTy = I->getType(); |
1178 | if (!isTypeLegal(Ty: DstTy, VT&: DstVT)) |
1179 | return false; |
1180 | |
1181 | if (DstVT != MVT::i32 && DstVT != MVT::i64) |
1182 | return false; |
1183 | |
1184 | // If we don't have FCTIDUZ, or SPE, and we need it, punt to SelectionDAG. |
1185 | if (DstVT == MVT::i64 && !IsSigned && !Subtarget->hasFPCVT() && |
1186 | !Subtarget->hasSPE()) |
1187 | return false; |
1188 | |
1189 | Value *Src = I->getOperand(i: 0); |
1190 | Type *SrcTy = Src->getType(); |
1191 | if (!isTypeLegal(Ty: SrcTy, VT&: SrcVT)) |
1192 | return false; |
1193 | |
1194 | if (SrcVT != MVT::f32 && SrcVT != MVT::f64) |
1195 | return false; |
1196 | |
1197 | Register SrcReg = getRegForValue(V: Src); |
1198 | if (!SrcReg) |
1199 | return false; |
1200 | |
1201 | // Convert f32 to f64 or convert VSSRC to VSFRC if necessary. This is just a |
1202 | // meaningless copy to get the register class right. |
1203 | const TargetRegisterClass *InRC = MRI.getRegClass(Reg: SrcReg); |
1204 | if (InRC == &PPC::F4RCRegClass) |
1205 | SrcReg = copyRegToRegClass(ToRC: &PPC::F8RCRegClass, SrcReg); |
1206 | else if (InRC == &PPC::VSSRCRegClass) |
1207 | SrcReg = copyRegToRegClass(ToRC: &PPC::VSFRCRegClass, SrcReg); |
1208 | |
1209 | // Determine the opcode for the conversion, which takes place |
1210 | // entirely within FPRs or VSRs. |
1211 | Register DestReg; |
1212 | unsigned Opc; |
1213 | auto RC = MRI.getRegClass(Reg: SrcReg); |
1214 | |
1215 | if (Subtarget->hasSPE()) { |
1216 | DestReg = createResultReg(RC: &PPC::GPRCRegClass); |
1217 | if (IsSigned) |
1218 | Opc = InRC == &PPC::GPRCRegClass ? PPC::EFSCTSIZ : PPC::EFDCTSIZ; |
1219 | else |
1220 | Opc = InRC == &PPC::GPRCRegClass ? PPC::EFSCTUIZ : PPC::EFDCTUIZ; |
1221 | } else if (isVSFRCRegClass(RC)) { |
1222 | DestReg = createResultReg(RC: &PPC::VSFRCRegClass); |
1223 | if (DstVT == MVT::i32) |
1224 | Opc = IsSigned ? PPC::XSCVDPSXWS : PPC::XSCVDPUXWS; |
1225 | else |
1226 | Opc = IsSigned ? PPC::XSCVDPSXDS : PPC::XSCVDPUXDS; |
1227 | } else { |
1228 | DestReg = createResultReg(RC: &PPC::F8RCRegClass); |
1229 | if (DstVT == MVT::i32) |
1230 | if (IsSigned) |
1231 | Opc = PPC::FCTIWZ; |
1232 | else |
1233 | Opc = Subtarget->hasFPCVT() ? PPC::FCTIWUZ : PPC::FCTIDZ; |
1234 | else |
1235 | Opc = IsSigned ? PPC::FCTIDZ : PPC::FCTIDUZ; |
1236 | } |
1237 | |
1238 | // Generate the convert. |
1239 | BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, MCID: TII.get(Opcode: Opc), DestReg) |
1240 | .addReg(RegNo: SrcReg); |
1241 | |
1242 | // Now move the integer value from a float register to an integer register. |
1243 | Register IntReg = Subtarget->hasSPE() |
1244 | ? DestReg |
1245 | : PPCMoveToIntReg(I, VT: DstVT, SrcReg: DestReg, IsSigned); |
1246 | |
1247 | if (!IntReg) |
1248 | return false; |
1249 | |
1250 | updateValueMap(I, Reg: IntReg); |
1251 | return true; |
1252 | } |
1253 | |
1254 | // Attempt to fast-select a binary integer operation that isn't already |
1255 | // handled automatically. |
1256 | bool PPCFastISel::SelectBinaryIntOp(const Instruction *I, unsigned ISDOpcode) { |
1257 | EVT DestVT = TLI.getValueType(DL, Ty: I->getType(), AllowUnknown: true); |
1258 | |
1259 | // We can get here in the case when we have a binary operation on a non-legal |
1260 | // type and the target independent selector doesn't know how to handle it. |
1261 | if (DestVT != MVT::i16 && DestVT != MVT::i8) |
1262 | return false; |
1263 | |
1264 | // Look at the currently assigned register for this instruction |
1265 | // to determine the required register class. If there is no register, |
1266 | // make a conservative choice (don't assign R0). |
1267 | Register AssignedReg = FuncInfo.ValueMap[I]; |
1268 | const TargetRegisterClass *RC = |
1269 | (AssignedReg ? MRI.getRegClass(Reg: AssignedReg) : |
1270 | &PPC::GPRC_and_GPRC_NOR0RegClass); |
1271 | bool IsGPRC = RC->hasSuperClassEq(RC: &PPC::GPRCRegClass); |
1272 | |
1273 | unsigned Opc; |
1274 | switch (ISDOpcode) { |
1275 | default: return false; |
1276 | case ISD::ADD: |
1277 | Opc = IsGPRC ? PPC::ADD4 : PPC::ADD8; |
1278 | break; |
1279 | case ISD::OR: |
1280 | Opc = IsGPRC ? PPC::OR : PPC::OR8; |
1281 | break; |
1282 | case ISD::SUB: |
1283 | Opc = IsGPRC ? PPC::SUBF : PPC::SUBF8; |
1284 | break; |
1285 | } |
1286 | |
1287 | Register ResultReg = createResultReg(RC: RC ? RC : &PPC::G8RCRegClass); |
1288 | Register SrcReg1 = getRegForValue(V: I->getOperand(i: 0)); |
1289 | if (!SrcReg1) |
1290 | return false; |
1291 | |
1292 | // Handle case of small immediate operand. |
1293 | if (const ConstantInt *ConstInt = dyn_cast<ConstantInt>(Val: I->getOperand(i: 1))) { |
1294 | const APInt &CIVal = ConstInt->getValue(); |
1295 | int Imm = (int)CIVal.getSExtValue(); |
1296 | bool UseImm = true; |
1297 | if (isInt<16>(x: Imm)) { |
1298 | switch (Opc) { |
1299 | default: |
1300 | llvm_unreachable("Missing case!" ); |
1301 | case PPC::ADD4: |
1302 | Opc = PPC::ADDI; |
1303 | MRI.setRegClass(Reg: SrcReg1, RC: &PPC::GPRC_and_GPRC_NOR0RegClass); |
1304 | break; |
1305 | case PPC::ADD8: |
1306 | Opc = PPC::ADDI8; |
1307 | MRI.setRegClass(Reg: SrcReg1, RC: &PPC::G8RC_and_G8RC_NOX0RegClass); |
1308 | break; |
1309 | case PPC::OR: |
1310 | Opc = PPC::ORI; |
1311 | break; |
1312 | case PPC::OR8: |
1313 | Opc = PPC::ORI8; |
1314 | break; |
1315 | case PPC::SUBF: |
1316 | if (Imm == -32768) |
1317 | UseImm = false; |
1318 | else { |
1319 | Opc = PPC::ADDI; |
1320 | MRI.setRegClass(Reg: SrcReg1, RC: &PPC::GPRC_and_GPRC_NOR0RegClass); |
1321 | Imm = -Imm; |
1322 | } |
1323 | break; |
1324 | case PPC::SUBF8: |
1325 | if (Imm == -32768) |
1326 | UseImm = false; |
1327 | else { |
1328 | Opc = PPC::ADDI8; |
1329 | MRI.setRegClass(Reg: SrcReg1, RC: &PPC::G8RC_and_G8RC_NOX0RegClass); |
1330 | Imm = -Imm; |
1331 | } |
1332 | break; |
1333 | } |
1334 | |
1335 | if (UseImm) { |
1336 | BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, MCID: TII.get(Opcode: Opc), |
1337 | DestReg: ResultReg) |
1338 | .addReg(RegNo: SrcReg1) |
1339 | .addImm(Val: Imm); |
1340 | updateValueMap(I, Reg: ResultReg); |
1341 | return true; |
1342 | } |
1343 | } |
1344 | } |
1345 | |
1346 | // Reg-reg case. |
1347 | Register SrcReg2 = getRegForValue(V: I->getOperand(i: 1)); |
1348 | if (!SrcReg2) |
1349 | return false; |
1350 | |
1351 | // Reverse operands for subtract-from. |
1352 | if (ISDOpcode == ISD::SUB) |
1353 | std::swap(a&: SrcReg1, b&: SrcReg2); |
1354 | |
1355 | BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, MCID: TII.get(Opcode: Opc), DestReg: ResultReg) |
1356 | .addReg(RegNo: SrcReg1).addReg(RegNo: SrcReg2); |
1357 | updateValueMap(I, Reg: ResultReg); |
1358 | return true; |
1359 | } |
1360 | |
1361 | // Handle arguments to a call that we're attempting to fast-select. |
1362 | // Return false if the arguments are too complex for us at the moment. |
1363 | bool PPCFastISel::processCallArgs(SmallVectorImpl<Value *> &Args, |
1364 | SmallVectorImpl<Register> &ArgRegs, |
1365 | SmallVectorImpl<MVT> &ArgVTs, |
1366 | SmallVectorImpl<ISD::ArgFlagsTy> &ArgFlags, |
1367 | SmallVectorImpl<unsigned> &RegArgs, |
1368 | CallingConv::ID CC, unsigned &NumBytes, |
1369 | bool IsVarArg) { |
1370 | SmallVector<CCValAssign, 16> ArgLocs; |
1371 | CCState CCInfo(CC, IsVarArg, *FuncInfo.MF, ArgLocs, *Context); |
1372 | |
1373 | // Reserve space for the linkage area on the stack. |
1374 | unsigned LinkageSize = Subtarget->getFrameLowering()->getLinkageSize(); |
1375 | CCInfo.AllocateStack(Size: LinkageSize, Alignment: Align(8)); |
1376 | |
1377 | CCInfo.AnalyzeCallOperands(ArgVTs, Flags&: ArgFlags, Fn: CC_PPC64_ELF_FIS); |
1378 | |
1379 | // Bail out if we can't handle any of the arguments. |
1380 | for (const CCValAssign &VA : ArgLocs) { |
1381 | MVT ArgVT = ArgVTs[VA.getValNo()]; |
1382 | |
1383 | // Skip vector arguments for now, as well as long double and |
1384 | // uint128_t, and anything that isn't passed in a register. |
1385 | if (ArgVT.isVector() || ArgVT.getSizeInBits() > 64 || ArgVT == MVT::i1 || |
1386 | !VA.isRegLoc() || VA.needsCustom()) |
1387 | return false; |
1388 | |
1389 | // Skip bit-converted arguments for now. |
1390 | if (VA.getLocInfo() == CCValAssign::BCvt) |
1391 | return false; |
1392 | } |
1393 | |
1394 | // Get a count of how many bytes are to be pushed onto the stack. |
1395 | NumBytes = CCInfo.getStackSize(); |
1396 | |
1397 | // The prolog code of the callee may store up to 8 GPR argument registers to |
1398 | // the stack, allowing va_start to index over them in memory if its varargs. |
1399 | // Because we cannot tell if this is needed on the caller side, we have to |
1400 | // conservatively assume that it is needed. As such, make sure we have at |
1401 | // least enough stack space for the caller to store the 8 GPRs. |
1402 | // FIXME: On ELFv2, it may be unnecessary to allocate the parameter area. |
1403 | NumBytes = std::max(a: NumBytes, b: LinkageSize + 64); |
1404 | |
1405 | // Issue CALLSEQ_START. |
1406 | BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, |
1407 | MCID: TII.get(Opcode: TII.getCallFrameSetupOpcode())) |
1408 | .addImm(Val: NumBytes).addImm(Val: 0); |
1409 | |
1410 | // Prepare to assign register arguments. Every argument uses up a |
1411 | // GPR protocol register even if it's passed in a floating-point |
1412 | // register (unless we're using the fast calling convention). |
1413 | unsigned NextGPR = PPC::X3; |
1414 | unsigned NextFPR = PPC::F1; |
1415 | |
1416 | // Process arguments. |
1417 | for (const CCValAssign &VA : ArgLocs) { |
1418 | Register Arg = ArgRegs[VA.getValNo()]; |
1419 | MVT ArgVT = ArgVTs[VA.getValNo()]; |
1420 | |
1421 | // Handle argument promotion and bitcasts. |
1422 | switch (VA.getLocInfo()) { |
1423 | default: |
1424 | llvm_unreachable("Unknown loc info!" ); |
1425 | case CCValAssign::Full: |
1426 | break; |
1427 | case CCValAssign::SExt: { |
1428 | MVT DestVT = VA.getLocVT(); |
1429 | const TargetRegisterClass *RC = |
1430 | (DestVT == MVT::i64) ? &PPC::G8RCRegClass : &PPC::GPRCRegClass; |
1431 | Register TmpReg = createResultReg(RC); |
1432 | if (!PPCEmitIntExt(SrcVT: ArgVT, SrcReg: Arg, DestVT, DestReg: TmpReg, /*IsZExt*/false)) |
1433 | llvm_unreachable("Failed to emit a sext!" ); |
1434 | ArgVT = DestVT; |
1435 | Arg = TmpReg; |
1436 | break; |
1437 | } |
1438 | case CCValAssign::AExt: |
1439 | case CCValAssign::ZExt: { |
1440 | MVT DestVT = VA.getLocVT(); |
1441 | const TargetRegisterClass *RC = |
1442 | (DestVT == MVT::i64) ? &PPC::G8RCRegClass : &PPC::GPRCRegClass; |
1443 | Register TmpReg = createResultReg(RC); |
1444 | if (!PPCEmitIntExt(SrcVT: ArgVT, SrcReg: Arg, DestVT, DestReg: TmpReg, /*IsZExt*/true)) |
1445 | llvm_unreachable("Failed to emit a zext!" ); |
1446 | ArgVT = DestVT; |
1447 | Arg = TmpReg; |
1448 | break; |
1449 | } |
1450 | case CCValAssign::BCvt: { |
1451 | // FIXME: Not yet handled. |
1452 | llvm_unreachable("Should have bailed before getting here!" ); |
1453 | break; |
1454 | } |
1455 | } |
1456 | |
1457 | // Copy this argument to the appropriate register. |
1458 | unsigned ArgReg; |
1459 | if (ArgVT == MVT::f32 || ArgVT == MVT::f64) { |
1460 | ArgReg = NextFPR++; |
1461 | if (CC != CallingConv::Fast) |
1462 | ++NextGPR; |
1463 | } else |
1464 | ArgReg = NextGPR++; |
1465 | |
1466 | BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, |
1467 | MCID: TII.get(Opcode: TargetOpcode::COPY), DestReg: ArgReg).addReg(RegNo: Arg); |
1468 | RegArgs.push_back(Elt: ArgReg); |
1469 | } |
1470 | |
1471 | return true; |
1472 | } |
1473 | |
1474 | // For a call that we've determined we can fast-select, finish the |
1475 | // call sequence and generate a copy to obtain the return value (if any). |
1476 | bool PPCFastISel::finishCall(MVT RetVT, CallLoweringInfo &CLI, unsigned &NumBytes) { |
1477 | CallingConv::ID CC = CLI.CallConv; |
1478 | |
1479 | // Issue CallSEQ_END. |
1480 | BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, |
1481 | MCID: TII.get(Opcode: TII.getCallFrameDestroyOpcode())) |
1482 | .addImm(Val: NumBytes).addImm(Val: 0); |
1483 | |
1484 | // Next, generate a copy to obtain the return value. |
1485 | // FIXME: No multi-register return values yet, though I don't foresee |
1486 | // any real difficulties there. |
1487 | if (RetVT != MVT::isVoid) { |
1488 | SmallVector<CCValAssign, 16> RVLocs; |
1489 | CCState CCInfo(CC, false, *FuncInfo.MF, RVLocs, *Context); |
1490 | CCInfo.AnalyzeCallResult(VT: RetVT, Fn: RetCC_PPC64_ELF_FIS); |
1491 | CCValAssign &VA = RVLocs[0]; |
1492 | assert(RVLocs.size() == 1 && "No support for multi-reg return values!" ); |
1493 | assert(VA.isRegLoc() && "Can only return in registers!" ); |
1494 | |
1495 | MVT DestVT = VA.getValVT(); |
1496 | MVT CopyVT = DestVT; |
1497 | |
1498 | // Ints smaller than a register still arrive in a full 64-bit |
1499 | // register, so make sure we recognize this. |
1500 | if (RetVT == MVT::i8 || RetVT == MVT::i16 || RetVT == MVT::i32) |
1501 | CopyVT = MVT::i64; |
1502 | |
1503 | Register SourcePhysReg = VA.getLocReg(); |
1504 | Register ResultReg; |
1505 | |
1506 | if (RetVT == CopyVT) { |
1507 | const TargetRegisterClass *CpyRC = TLI.getRegClassFor(VT: CopyVT); |
1508 | ResultReg = copyRegToRegClass(ToRC: CpyRC, SrcReg: SourcePhysReg); |
1509 | |
1510 | // If necessary, round the floating result to single precision. |
1511 | } else if (CopyVT == MVT::f64) { |
1512 | ResultReg = createResultReg(RC: TLI.getRegClassFor(VT: RetVT)); |
1513 | BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, MCID: TII.get(Opcode: PPC::FRSP), |
1514 | DestReg: ResultReg).addReg(RegNo: SourcePhysReg); |
1515 | |
1516 | // If only the low half of a general register is needed, generate |
1517 | // a GPRC copy instead of a G8RC copy. (EXTRACT_SUBREG can't be |
1518 | // used along the fast-isel path (not lowered), and downstream logic |
1519 | // also doesn't like a direct subreg copy on a physical reg.) |
1520 | } else if (RetVT == MVT::i8 || RetVT == MVT::i16 || RetVT == MVT::i32) { |
1521 | // Convert physical register from G8RC to GPRC. |
1522 | SourcePhysReg = (SourcePhysReg - PPC::X0) + PPC::R0; |
1523 | ResultReg = copyRegToRegClass(ToRC: &PPC::GPRCRegClass, SrcReg: SourcePhysReg); |
1524 | } |
1525 | |
1526 | assert(ResultReg && "ResultReg unset!" ); |
1527 | CLI.InRegs.push_back(Elt: SourcePhysReg); |
1528 | CLI.ResultReg = ResultReg; |
1529 | CLI.NumResultRegs = 1; |
1530 | } |
1531 | |
1532 | return true; |
1533 | } |
1534 | |
1535 | bool PPCFastISel::fastLowerCall(CallLoweringInfo &CLI) { |
1536 | CallingConv::ID CC = CLI.CallConv; |
1537 | bool IsTailCall = CLI.IsTailCall; |
1538 | bool IsVarArg = CLI.IsVarArg; |
1539 | const Value *Callee = CLI.Callee; |
1540 | const MCSymbol *Symbol = CLI.Symbol; |
1541 | |
1542 | if (!Callee && !Symbol) |
1543 | return false; |
1544 | |
1545 | // Allow SelectionDAG isel to handle tail calls and long calls. |
1546 | if (IsTailCall || Subtarget->useLongCalls()) |
1547 | return false; |
1548 | |
1549 | // Let SDISel handle vararg functions. |
1550 | if (IsVarArg) |
1551 | return false; |
1552 | |
1553 | // If this is a PC-Rel function, let SDISel handle the call. |
1554 | if (Subtarget->isUsingPCRelativeCalls()) |
1555 | return false; |
1556 | |
1557 | // Handle simple calls for now, with legal return types and |
1558 | // those that can be extended. |
1559 | Type *RetTy = CLI.RetTy; |
1560 | MVT RetVT; |
1561 | if (RetTy->isVoidTy()) |
1562 | RetVT = MVT::isVoid; |
1563 | else if (!isTypeLegal(Ty: RetTy, VT&: RetVT) && RetVT != MVT::i16 && |
1564 | RetVT != MVT::i8) |
1565 | return false; |
1566 | else if (RetVT == MVT::i1 && Subtarget->useCRBits()) |
1567 | // We can't handle boolean returns when CR bits are in use. |
1568 | return false; |
1569 | |
1570 | // FIXME: No multi-register return values yet. |
1571 | if (RetVT != MVT::isVoid && RetVT != MVT::i8 && RetVT != MVT::i16 && |
1572 | RetVT != MVT::i32 && RetVT != MVT::i64 && RetVT != MVT::f32 && |
1573 | RetVT != MVT::f64) { |
1574 | SmallVector<CCValAssign, 16> RVLocs; |
1575 | CCState CCInfo(CC, IsVarArg, *FuncInfo.MF, RVLocs, *Context); |
1576 | CCInfo.AnalyzeCallResult(VT: RetVT, Fn: RetCC_PPC64_ELF_FIS); |
1577 | if (RVLocs.size() > 1) |
1578 | return false; |
1579 | } |
1580 | |
1581 | // Bail early if more than 8 arguments, as we only currently |
1582 | // handle arguments passed in registers. |
1583 | unsigned NumArgs = CLI.OutVals.size(); |
1584 | if (NumArgs > 8) |
1585 | return false; |
1586 | |
1587 | // Set up the argument vectors. |
1588 | SmallVector<Value*, 8> Args; |
1589 | SmallVector<Register, 8> ArgRegs; |
1590 | SmallVector<MVT, 8> ArgVTs; |
1591 | SmallVector<ISD::ArgFlagsTy, 8> ArgFlags; |
1592 | |
1593 | Args.reserve(N: NumArgs); |
1594 | ArgRegs.reserve(N: NumArgs); |
1595 | ArgVTs.reserve(N: NumArgs); |
1596 | ArgFlags.reserve(N: NumArgs); |
1597 | |
1598 | for (unsigned i = 0, ie = NumArgs; i != ie; ++i) { |
1599 | // Only handle easy calls for now. It would be reasonably easy |
1600 | // to handle <= 8-byte structures passed ByVal in registers, but we |
1601 | // have to ensure they are right-justified in the register. |
1602 | ISD::ArgFlagsTy Flags = CLI.OutFlags[i]; |
1603 | if (Flags.isInReg() || Flags.isSRet() || Flags.isNest() || Flags.isByVal()) |
1604 | return false; |
1605 | |
1606 | Value *ArgValue = CLI.OutVals[i]; |
1607 | Type *ArgTy = ArgValue->getType(); |
1608 | MVT ArgVT; |
1609 | if (!isTypeLegal(Ty: ArgTy, VT&: ArgVT) && ArgVT != MVT::i16 && ArgVT != MVT::i8) |
1610 | return false; |
1611 | |
1612 | // FIXME: FastISel cannot handle non-simple types yet, including 128-bit FP |
1613 | // types, which is passed through vector register. Skip these types and |
1614 | // fallback to default SelectionDAG based selection. |
1615 | if (ArgVT.isVector() || ArgVT == MVT::f128) |
1616 | return false; |
1617 | |
1618 | Register Arg = getRegForValue(V: ArgValue); |
1619 | if (!Arg) |
1620 | return false; |
1621 | |
1622 | Args.push_back(Elt: ArgValue); |
1623 | ArgRegs.push_back(Elt: Arg); |
1624 | ArgVTs.push_back(Elt: ArgVT); |
1625 | ArgFlags.push_back(Elt: Flags); |
1626 | } |
1627 | |
1628 | // Process the arguments. |
1629 | SmallVector<unsigned, 8> RegArgs; |
1630 | unsigned NumBytes; |
1631 | |
1632 | if (!processCallArgs(Args, ArgRegs, ArgVTs, ArgFlags, |
1633 | RegArgs, CC, NumBytes, IsVarArg)) |
1634 | return false; |
1635 | |
1636 | MachineInstrBuilder MIB; |
1637 | // FIXME: No handling for function pointers yet. This requires |
1638 | // implementing the function descriptor (OPD) setup. |
1639 | const GlobalValue *GV = dyn_cast<GlobalValue>(Val: Callee); |
1640 | if (!GV) { |
1641 | // patchpoints are a special case; they always dispatch to a pointer value. |
1642 | // However, we don't actually want to generate the indirect call sequence |
1643 | // here (that will be generated, as necessary, during asm printing), and |
1644 | // the call we generate here will be erased by FastISel::selectPatchpoint, |
1645 | // so don't try very hard... |
1646 | if (CLI.IsPatchPoint) |
1647 | MIB = BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, MCID: TII.get(Opcode: PPC::NOP)); |
1648 | else |
1649 | return false; |
1650 | } else { |
1651 | // Build direct call with NOP for TOC restore. |
1652 | // FIXME: We can and should optimize away the NOP for local calls. |
1653 | MIB = BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, |
1654 | MCID: TII.get(Opcode: PPC::BL8_NOP)); |
1655 | // Add callee. |
1656 | MIB.addGlobalAddress(GV); |
1657 | } |
1658 | |
1659 | // Add implicit physical register uses to the call. |
1660 | for (unsigned Reg : RegArgs) |
1661 | MIB.addReg(RegNo: Reg, flags: RegState::Implicit); |
1662 | |
1663 | // Direct calls, in both the ELF V1 and V2 ABIs, need the TOC register live |
1664 | // into the call. |
1665 | PPCFuncInfo->setUsesTOCBasePtr(); |
1666 | MIB.addReg(RegNo: PPC::X2, flags: RegState::Implicit); |
1667 | |
1668 | // Add a register mask with the call-preserved registers. Proper |
1669 | // defs for return values will be added by setPhysRegsDeadExcept(). |
1670 | MIB.addRegMask(Mask: TRI.getCallPreservedMask(MF: *FuncInfo.MF, CC)); |
1671 | |
1672 | CLI.Call = MIB; |
1673 | |
1674 | // Finish off the call including any return values. |
1675 | return finishCall(RetVT, CLI, NumBytes); |
1676 | } |
1677 | |
1678 | // Attempt to fast-select a return instruction. |
1679 | bool PPCFastISel::SelectRet(const Instruction *I) { |
1680 | |
1681 | if (!FuncInfo.CanLowerReturn) |
1682 | return false; |
1683 | |
1684 | const ReturnInst *Ret = cast<ReturnInst>(Val: I); |
1685 | const Function &F = *I->getParent()->getParent(); |
1686 | |
1687 | // Build a list of return value registers. |
1688 | SmallVector<Register, 4> RetRegs; |
1689 | CallingConv::ID CC = F.getCallingConv(); |
1690 | |
1691 | if (Ret->getNumOperands() > 0) { |
1692 | SmallVector<ISD::OutputArg, 4> Outs; |
1693 | GetReturnInfo(CC, ReturnType: F.getReturnType(), attr: F.getAttributes(), Outs, TLI, DL); |
1694 | |
1695 | // Analyze operands of the call, assigning locations to each operand. |
1696 | SmallVector<CCValAssign, 16> ValLocs; |
1697 | CCState CCInfo(CC, F.isVarArg(), *FuncInfo.MF, ValLocs, *Context); |
1698 | CCInfo.AnalyzeReturn(Outs, Fn: RetCC_PPC64_ELF_FIS); |
1699 | const Value *RV = Ret->getOperand(i_nocapture: 0); |
1700 | |
1701 | // FIXME: Only one output register for now. |
1702 | if (ValLocs.size() > 1) |
1703 | return false; |
1704 | |
1705 | // Special case for returning a constant integer of any size - materialize |
1706 | // the constant as an i64 and copy it to the return register. |
1707 | if (const ConstantInt *CI = dyn_cast<ConstantInt>(Val: RV)) { |
1708 | CCValAssign &VA = ValLocs[0]; |
1709 | |
1710 | Register RetReg = VA.getLocReg(); |
1711 | // We still need to worry about properly extending the sign. For example, |
1712 | // we could have only a single bit or a constant that needs zero |
1713 | // extension rather than sign extension. Make sure we pass the return |
1714 | // value extension property to integer materialization. |
1715 | Register SrcReg = |
1716 | PPCMaterializeInt(CI, VT: MVT::i64, UseSExt: VA.getLocInfo() != CCValAssign::ZExt); |
1717 | |
1718 | BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, |
1719 | MCID: TII.get(Opcode: TargetOpcode::COPY), DestReg: RetReg).addReg(RegNo: SrcReg); |
1720 | |
1721 | RetRegs.push_back(Elt: RetReg); |
1722 | |
1723 | } else { |
1724 | Register Reg = getRegForValue(V: RV); |
1725 | |
1726 | if (!Reg) |
1727 | return false; |
1728 | |
1729 | // Copy the result values into the output registers. |
1730 | for (unsigned i = 0; i < ValLocs.size(); ++i) { |
1731 | |
1732 | CCValAssign &VA = ValLocs[i]; |
1733 | assert(VA.isRegLoc() && "Can only return in registers!" ); |
1734 | RetRegs.push_back(Elt: VA.getLocReg()); |
1735 | Register SrcReg = Reg + VA.getValNo(); |
1736 | |
1737 | EVT RVEVT = TLI.getValueType(DL, Ty: RV->getType()); |
1738 | if (!RVEVT.isSimple()) |
1739 | return false; |
1740 | MVT RVVT = RVEVT.getSimpleVT(); |
1741 | MVT DestVT = VA.getLocVT(); |
1742 | |
1743 | if (RVVT != DestVT && RVVT != MVT::i8 && |
1744 | RVVT != MVT::i16 && RVVT != MVT::i32) |
1745 | return false; |
1746 | |
1747 | if (RVVT != DestVT) { |
1748 | switch (VA.getLocInfo()) { |
1749 | default: |
1750 | llvm_unreachable("Unknown loc info!" ); |
1751 | case CCValAssign::Full: |
1752 | llvm_unreachable("Full value assign but types don't match?" ); |
1753 | case CCValAssign::AExt: |
1754 | case CCValAssign::ZExt: { |
1755 | const TargetRegisterClass *RC = |
1756 | (DestVT == MVT::i64) ? &PPC::G8RCRegClass : &PPC::GPRCRegClass; |
1757 | Register TmpReg = createResultReg(RC); |
1758 | if (!PPCEmitIntExt(SrcVT: RVVT, SrcReg, DestVT, DestReg: TmpReg, IsZExt: true)) |
1759 | return false; |
1760 | SrcReg = TmpReg; |
1761 | break; |
1762 | } |
1763 | case CCValAssign::SExt: { |
1764 | const TargetRegisterClass *RC = |
1765 | (DestVT == MVT::i64) ? &PPC::G8RCRegClass : &PPC::GPRCRegClass; |
1766 | Register TmpReg = createResultReg(RC); |
1767 | if (!PPCEmitIntExt(SrcVT: RVVT, SrcReg, DestVT, DestReg: TmpReg, IsZExt: false)) |
1768 | return false; |
1769 | SrcReg = TmpReg; |
1770 | break; |
1771 | } |
1772 | } |
1773 | } |
1774 | |
1775 | BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, |
1776 | MCID: TII.get(Opcode: TargetOpcode::COPY), DestReg: RetRegs[i]) |
1777 | .addReg(RegNo: SrcReg); |
1778 | } |
1779 | } |
1780 | } |
1781 | |
1782 | MachineInstrBuilder MIB = BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, |
1783 | MCID: TII.get(Opcode: PPC::BLR8)); |
1784 | |
1785 | for (Register Reg : RetRegs) |
1786 | MIB.addReg(RegNo: Reg, flags: RegState::Implicit); |
1787 | |
1788 | return true; |
1789 | } |
1790 | |
1791 | // Attempt to emit an integer extend of SrcReg into DestReg. Both |
1792 | // signed and zero extensions are supported. Return false if we |
1793 | // can't handle it. |
1794 | bool PPCFastISel::PPCEmitIntExt(MVT SrcVT, Register SrcReg, MVT DestVT, |
1795 | Register DestReg, bool IsZExt) { |
1796 | if (DestVT != MVT::i32 && DestVT != MVT::i64) |
1797 | return false; |
1798 | if (SrcVT != MVT::i8 && SrcVT != MVT::i16 && SrcVT != MVT::i32) |
1799 | return false; |
1800 | |
1801 | // Signed extensions use EXTSB, EXTSH, EXTSW. |
1802 | if (!IsZExt) { |
1803 | unsigned Opc; |
1804 | if (SrcVT == MVT::i8) |
1805 | Opc = (DestVT == MVT::i32) ? PPC::EXTSB : PPC::EXTSB8_32_64; |
1806 | else if (SrcVT == MVT::i16) |
1807 | Opc = (DestVT == MVT::i32) ? PPC::EXTSH : PPC::EXTSH8_32_64; |
1808 | else { |
1809 | assert(DestVT == MVT::i64 && "Signed extend from i32 to i32??" ); |
1810 | Opc = PPC::EXTSW_32_64; |
1811 | } |
1812 | BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, MCID: TII.get(Opcode: Opc), DestReg) |
1813 | .addReg(RegNo: SrcReg); |
1814 | |
1815 | // Unsigned 32-bit extensions use RLWINM. |
1816 | } else if (DestVT == MVT::i32) { |
1817 | unsigned MB; |
1818 | if (SrcVT == MVT::i8) |
1819 | MB = 24; |
1820 | else { |
1821 | assert(SrcVT == MVT::i16 && "Unsigned extend from i32 to i32??" ); |
1822 | MB = 16; |
1823 | } |
1824 | BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, MCID: TII.get(Opcode: PPC::RLWINM), |
1825 | DestReg) |
1826 | .addReg(RegNo: SrcReg).addImm(/*SH=*/Val: 0).addImm(Val: MB).addImm(/*ME=*/Val: 31); |
1827 | |
1828 | // Unsigned 64-bit extensions use RLDICL (with a 32-bit source). |
1829 | } else { |
1830 | unsigned MB; |
1831 | if (SrcVT == MVT::i8) |
1832 | MB = 56; |
1833 | else if (SrcVT == MVT::i16) |
1834 | MB = 48; |
1835 | else |
1836 | MB = 32; |
1837 | BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, |
1838 | MCID: TII.get(Opcode: PPC::RLDICL_32_64), DestReg) |
1839 | .addReg(RegNo: SrcReg).addImm(/*SH=*/Val: 0).addImm(Val: MB); |
1840 | } |
1841 | |
1842 | return true; |
1843 | } |
1844 | |
1845 | // Attempt to fast-select an indirect branch instruction. |
1846 | bool PPCFastISel::SelectIndirectBr(const Instruction *I) { |
1847 | Register AddrReg = getRegForValue(V: I->getOperand(i: 0)); |
1848 | if (!AddrReg) |
1849 | return false; |
1850 | |
1851 | BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, MCID: TII.get(Opcode: PPC::MTCTR8)) |
1852 | .addReg(RegNo: AddrReg); |
1853 | BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, MCID: TII.get(Opcode: PPC::BCTR8)); |
1854 | |
1855 | const IndirectBrInst *IB = cast<IndirectBrInst>(Val: I); |
1856 | for (const BasicBlock *SuccBB : IB->successors()) |
1857 | FuncInfo.MBB->addSuccessor(Succ: FuncInfo.getMBB(BB: SuccBB)); |
1858 | |
1859 | return true; |
1860 | } |
1861 | |
1862 | // Attempt to fast-select an integer truncate instruction. |
1863 | bool PPCFastISel::SelectTrunc(const Instruction *I) { |
1864 | Value *Src = I->getOperand(i: 0); |
1865 | EVT SrcVT = TLI.getValueType(DL, Ty: Src->getType(), AllowUnknown: true); |
1866 | EVT DestVT = TLI.getValueType(DL, Ty: I->getType(), AllowUnknown: true); |
1867 | |
1868 | if (SrcVT != MVT::i64 && SrcVT != MVT::i32 && SrcVT != MVT::i16) |
1869 | return false; |
1870 | |
1871 | if (DestVT != MVT::i32 && DestVT != MVT::i16 && DestVT != MVT::i8) |
1872 | return false; |
1873 | |
1874 | Register SrcReg = getRegForValue(V: Src); |
1875 | if (!SrcReg) |
1876 | return false; |
1877 | |
1878 | // The only interesting case is when we need to switch register classes. |
1879 | if (SrcVT == MVT::i64) |
1880 | SrcReg = copyRegToRegClass(ToRC: &PPC::GPRCRegClass, SrcReg, Flag: 0, SubReg: PPC::sub_32); |
1881 | |
1882 | updateValueMap(I, Reg: SrcReg); |
1883 | return true; |
1884 | } |
1885 | |
1886 | // Attempt to fast-select an integer extend instruction. |
1887 | bool PPCFastISel::SelectIntExt(const Instruction *I) { |
1888 | Type *DestTy = I->getType(); |
1889 | Value *Src = I->getOperand(i: 0); |
1890 | Type *SrcTy = Src->getType(); |
1891 | |
1892 | bool IsZExt = isa<ZExtInst>(Val: I); |
1893 | Register SrcReg = getRegForValue(V: Src); |
1894 | if (!SrcReg) return false; |
1895 | |
1896 | EVT SrcEVT, DestEVT; |
1897 | SrcEVT = TLI.getValueType(DL, Ty: SrcTy, AllowUnknown: true); |
1898 | DestEVT = TLI.getValueType(DL, Ty: DestTy, AllowUnknown: true); |
1899 | if (!SrcEVT.isSimple()) |
1900 | return false; |
1901 | if (!DestEVT.isSimple()) |
1902 | return false; |
1903 | |
1904 | MVT SrcVT = SrcEVT.getSimpleVT(); |
1905 | MVT DestVT = DestEVT.getSimpleVT(); |
1906 | |
1907 | // If we know the register class needed for the result of this |
1908 | // instruction, use it. Otherwise pick the register class of the |
1909 | // correct size that does not contain X0/R0, since we don't know |
1910 | // whether downstream uses permit that assignment. |
1911 | Register AssignedReg = FuncInfo.ValueMap[I]; |
1912 | const TargetRegisterClass *RC = |
1913 | (AssignedReg ? MRI.getRegClass(Reg: AssignedReg) : |
1914 | (DestVT == MVT::i64 ? &PPC::G8RC_and_G8RC_NOX0RegClass : |
1915 | &PPC::GPRC_and_GPRC_NOR0RegClass)); |
1916 | Register ResultReg = createResultReg(RC); |
1917 | |
1918 | if (!PPCEmitIntExt(SrcVT, SrcReg, DestVT, DestReg: ResultReg, IsZExt)) |
1919 | return false; |
1920 | |
1921 | updateValueMap(I, Reg: ResultReg); |
1922 | return true; |
1923 | } |
1924 | |
1925 | // Attempt to fast-select an instruction that wasn't handled by |
1926 | // the table-generated machinery. |
1927 | bool PPCFastISel::fastSelectInstruction(const Instruction *I) { |
1928 | |
1929 | switch (I->getOpcode()) { |
1930 | case Instruction::Load: |
1931 | return SelectLoad(I); |
1932 | case Instruction::Store: |
1933 | return SelectStore(I); |
1934 | case Instruction::Br: |
1935 | return SelectBranch(I); |
1936 | case Instruction::IndirectBr: |
1937 | return SelectIndirectBr(I); |
1938 | case Instruction::FPExt: |
1939 | return SelectFPExt(I); |
1940 | case Instruction::FPTrunc: |
1941 | return SelectFPTrunc(I); |
1942 | case Instruction::SIToFP: |
1943 | return SelectIToFP(I, /*IsSigned*/ true); |
1944 | case Instruction::UIToFP: |
1945 | return SelectIToFP(I, /*IsSigned*/ false); |
1946 | case Instruction::FPToSI: |
1947 | return SelectFPToI(I, /*IsSigned*/ true); |
1948 | case Instruction::FPToUI: |
1949 | return SelectFPToI(I, /*IsSigned*/ false); |
1950 | case Instruction::Add: |
1951 | return SelectBinaryIntOp(I, ISDOpcode: ISD::ADD); |
1952 | case Instruction::Or: |
1953 | return SelectBinaryIntOp(I, ISDOpcode: ISD::OR); |
1954 | case Instruction::Sub: |
1955 | return SelectBinaryIntOp(I, ISDOpcode: ISD::SUB); |
1956 | case Instruction::Ret: |
1957 | return SelectRet(I); |
1958 | case Instruction::Trunc: |
1959 | return SelectTrunc(I); |
1960 | case Instruction::ZExt: |
1961 | case Instruction::SExt: |
1962 | return SelectIntExt(I); |
1963 | // Here add other flavors of Instruction::XXX that automated |
1964 | // cases don't catch. For example, switches are terminators |
1965 | // that aren't yet handled. |
1966 | default: |
1967 | break; |
1968 | } |
1969 | return false; |
1970 | } |
1971 | |
1972 | // Materialize a floating-point constant into a register, and return |
1973 | // the register number (or zero if we failed to handle it). |
1974 | Register PPCFastISel::PPCMaterializeFP(const ConstantFP *CFP, MVT VT) { |
1975 | // If this is a PC-Rel function, let SDISel handle constant pool. |
1976 | if (Subtarget->isUsingPCRelativeCalls()) |
1977 | return Register(); |
1978 | |
1979 | // No plans to handle long double here. |
1980 | if (VT != MVT::f32 && VT != MVT::f64) |
1981 | return Register(); |
1982 | |
1983 | // All FP constants are loaded from the constant pool. |
1984 | Align Alignment = DL.getPrefTypeAlign(Ty: CFP->getType()); |
1985 | unsigned Idx = MCP.getConstantPoolIndex(C: cast<Constant>(Val: CFP), Alignment); |
1986 | const bool HasSPE = Subtarget->hasSPE(); |
1987 | const TargetRegisterClass *RC; |
1988 | if (HasSPE) |
1989 | RC = ((VT == MVT::f32) ? &PPC::GPRCRegClass : &PPC::SPERCRegClass); |
1990 | else |
1991 | RC = ((VT == MVT::f32) ? &PPC::F4RCRegClass : &PPC::F8RCRegClass); |
1992 | |
1993 | Register DestReg = createResultReg(RC); |
1994 | CodeModel::Model CModel = TM.getCodeModel(); |
1995 | |
1996 | MachineMemOperand *MMO = FuncInfo.MF->getMachineMemOperand( |
1997 | PtrInfo: MachinePointerInfo::getConstantPool(MF&: *FuncInfo.MF), |
1998 | F: MachineMemOperand::MOLoad, Size: (VT == MVT::f32) ? 4 : 8, BaseAlignment: Alignment); |
1999 | |
2000 | unsigned Opc; |
2001 | |
2002 | if (HasSPE) |
2003 | Opc = ((VT == MVT::f32) ? PPC::SPELWZ : PPC::EVLDD); |
2004 | else |
2005 | Opc = ((VT == MVT::f32) ? PPC::LFS : PPC::LFD); |
2006 | |
2007 | Register TmpReg = createResultReg(RC: &PPC::G8RC_and_G8RC_NOX0RegClass); |
2008 | |
2009 | PPCFuncInfo->setUsesTOCBasePtr(); |
2010 | // For small code model, generate a LF[SD](0, LDtocCPT(Idx, X2)). |
2011 | if (CModel == CodeModel::Small) { |
2012 | BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, MCID: TII.get(Opcode: PPC::LDtocCPT), |
2013 | DestReg: TmpReg) |
2014 | .addConstantPoolIndex(Idx).addReg(RegNo: PPC::X2); |
2015 | BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, MCID: TII.get(Opcode: Opc), DestReg) |
2016 | .addImm(Val: 0).addReg(RegNo: TmpReg).addMemOperand(MMO); |
2017 | } else { |
2018 | // Otherwise we generate LF[SD](Idx[lo], ADDIStocHA8(X2, Idx)). |
2019 | BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, MCID: TII.get(Opcode: PPC::ADDIStocHA8), |
2020 | DestReg: TmpReg).addReg(RegNo: PPC::X2).addConstantPoolIndex(Idx); |
2021 | // But for large code model, we must generate a LDtocL followed |
2022 | // by the LF[SD]. |
2023 | if (CModel == CodeModel::Large) { |
2024 | Register TmpReg2 = createResultReg(RC: &PPC::G8RC_and_G8RC_NOX0RegClass); |
2025 | BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, MCID: TII.get(Opcode: PPC::LDtocL), |
2026 | DestReg: TmpReg2).addConstantPoolIndex(Idx).addReg(RegNo: TmpReg); |
2027 | BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, MCID: TII.get(Opcode: Opc), DestReg) |
2028 | .addImm(Val: 0) |
2029 | .addReg(RegNo: TmpReg2); |
2030 | } else |
2031 | BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, MCID: TII.get(Opcode: Opc), DestReg) |
2032 | .addConstantPoolIndex(Idx, Offset: 0, TargetFlags: PPCII::MO_TOC_LO) |
2033 | .addReg(RegNo: TmpReg) |
2034 | .addMemOperand(MMO); |
2035 | } |
2036 | |
2037 | return DestReg; |
2038 | } |
2039 | |
2040 | // Materialize the address of a global value into a register, and return |
2041 | // the register number (or zero if we failed to handle it). |
2042 | Register PPCFastISel::PPCMaterializeGV(const GlobalValue *GV, MVT VT) { |
2043 | // If this is a PC-Rel function, let SDISel handle GV materialization. |
2044 | if (Subtarget->isUsingPCRelativeCalls()) |
2045 | return Register(); |
2046 | |
2047 | assert(VT == MVT::i64 && "Non-address!" ); |
2048 | const TargetRegisterClass *RC = &PPC::G8RC_and_G8RC_NOX0RegClass; |
2049 | Register DestReg = createResultReg(RC); |
2050 | |
2051 | // Global values may be plain old object addresses, TLS object |
2052 | // addresses, constant pool entries, or jump tables. How we generate |
2053 | // code for these may depend on small, medium, or large code model. |
2054 | CodeModel::Model CModel = TM.getCodeModel(); |
2055 | |
2056 | // FIXME: Jump tables are not yet required because fast-isel doesn't |
2057 | // handle switches; if that changes, we need them as well. For now, |
2058 | // what follows assumes everything's a generic (or TLS) global address. |
2059 | |
2060 | // FIXME: We don't yet handle the complexity of TLS. |
2061 | if (GV->isThreadLocal()) |
2062 | return Register(); |
2063 | |
2064 | PPCFuncInfo->setUsesTOCBasePtr(); |
2065 | bool IsAIXTocData = TM.getTargetTriple().isOSAIX() && |
2066 | isa<GlobalVariable>(Val: GV) && |
2067 | cast<GlobalVariable>(Val: GV)->hasAttribute(Kind: "toc-data" ); |
2068 | |
2069 | // For small code model, generate a simple TOC load. |
2070 | if (CModel == CodeModel::Small) { |
2071 | auto MIB = BuildMI( |
2072 | BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, |
2073 | MCID: IsAIXTocData ? TII.get(Opcode: PPC::ADDItoc8) : TII.get(Opcode: PPC::LDtoc), DestReg); |
2074 | if (IsAIXTocData) |
2075 | MIB.addReg(RegNo: PPC::X2).addGlobalAddress(GV); |
2076 | else |
2077 | MIB.addGlobalAddress(GV).addReg(RegNo: PPC::X2); |
2078 | } else { |
2079 | // If the address is an externally defined symbol, a symbol with common |
2080 | // or externally available linkage, a non-local function address, or a |
2081 | // jump table address (not yet needed), or if we are generating code |
2082 | // for large code model, we generate: |
2083 | // LDtocL(GV, ADDIStocHA8(%x2, GV)) |
2084 | // Otherwise we generate: |
2085 | // ADDItocL8(ADDIStocHA8(%x2, GV), GV) |
2086 | // Either way, start with the ADDIStocHA8: |
2087 | Register HighPartReg = createResultReg(RC); |
2088 | BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, MCID: TII.get(Opcode: PPC::ADDIStocHA8), |
2089 | DestReg: HighPartReg).addReg(RegNo: PPC::X2).addGlobalAddress(GV); |
2090 | |
2091 | if (Subtarget->isGVIndirectSymbol(GV)) { |
2092 | assert(!IsAIXTocData && "TOC data should always be direct." ); |
2093 | BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, MCID: TII.get(Opcode: PPC::LDtocL), |
2094 | DestReg).addGlobalAddress(GV).addReg(RegNo: HighPartReg); |
2095 | } else { |
2096 | // Otherwise generate the ADDItocL8. |
2097 | BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, MCID: TII.get(Opcode: PPC::ADDItocL8), |
2098 | DestReg) |
2099 | .addReg(RegNo: HighPartReg) |
2100 | .addGlobalAddress(GV); |
2101 | } |
2102 | } |
2103 | |
2104 | return DestReg; |
2105 | } |
2106 | |
2107 | // Materialize a 32-bit integer constant into a register, and return |
2108 | // the register number (or zero if we failed to handle it). |
2109 | Register PPCFastISel::PPCMaterialize32BitInt(int64_t Imm, |
2110 | const TargetRegisterClass *RC) { |
2111 | unsigned Lo = Imm & 0xFFFF; |
2112 | unsigned Hi = (Imm >> 16) & 0xFFFF; |
2113 | |
2114 | Register ResultReg = createResultReg(RC); |
2115 | bool IsGPRC = RC->hasSuperClassEq(RC: &PPC::GPRCRegClass); |
2116 | |
2117 | if (isInt<16>(x: Imm)) |
2118 | BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, |
2119 | MCID: TII.get(Opcode: IsGPRC ? PPC::LI : PPC::LI8), DestReg: ResultReg) |
2120 | .addImm(Val: Imm); |
2121 | else if (Lo) { |
2122 | // Both Lo and Hi have nonzero bits. |
2123 | Register TmpReg = createResultReg(RC); |
2124 | BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, |
2125 | MCID: TII.get(Opcode: IsGPRC ? PPC::LIS : PPC::LIS8), DestReg: TmpReg) |
2126 | .addImm(Val: Hi); |
2127 | BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, |
2128 | MCID: TII.get(Opcode: IsGPRC ? PPC::ORI : PPC::ORI8), DestReg: ResultReg) |
2129 | .addReg(RegNo: TmpReg).addImm(Val: Lo); |
2130 | } else |
2131 | // Just Hi bits. |
2132 | BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, |
2133 | MCID: TII.get(Opcode: IsGPRC ? PPC::LIS : PPC::LIS8), DestReg: ResultReg) |
2134 | .addImm(Val: Hi); |
2135 | |
2136 | return ResultReg; |
2137 | } |
2138 | |
2139 | // Materialize a 64-bit integer constant into a register, and return |
2140 | // the register number (or zero if we failed to handle it). |
2141 | Register PPCFastISel::PPCMaterialize64BitInt(int64_t Imm, |
2142 | const TargetRegisterClass *RC) { |
2143 | unsigned Remainder = 0; |
2144 | unsigned Shift = 0; |
2145 | |
2146 | // If the value doesn't fit in 32 bits, see if we can shift it |
2147 | // so that it fits in 32 bits. |
2148 | if (!isInt<32>(x: Imm)) { |
2149 | Shift = llvm::countr_zero<uint64_t>(Val: Imm); |
2150 | int64_t ImmSh = static_cast<uint64_t>(Imm) >> Shift; |
2151 | |
2152 | if (isInt<32>(x: ImmSh)) |
2153 | Imm = ImmSh; |
2154 | else { |
2155 | Remainder = Imm; |
2156 | Shift = 32; |
2157 | Imm >>= 32; |
2158 | } |
2159 | } |
2160 | |
2161 | // Handle the high-order 32 bits (if shifted) or the whole 32 bits |
2162 | // (if not shifted). |
2163 | Register TmpReg1 = PPCMaterialize32BitInt(Imm, RC); |
2164 | if (!Shift) |
2165 | return TmpReg1; |
2166 | |
2167 | // If upper 32 bits were not zero, we've built them and need to shift |
2168 | // them into place. |
2169 | Register TmpReg2; |
2170 | if (Imm) { |
2171 | TmpReg2 = createResultReg(RC); |
2172 | BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, MCID: TII.get(Opcode: PPC::RLDICR), |
2173 | DestReg: TmpReg2).addReg(RegNo: TmpReg1).addImm(Val: Shift).addImm(Val: 63 - Shift); |
2174 | } else |
2175 | TmpReg2 = TmpReg1; |
2176 | |
2177 | Register TmpReg3; |
2178 | unsigned Hi, Lo; |
2179 | if ((Hi = (Remainder >> 16) & 0xFFFF)) { |
2180 | TmpReg3 = createResultReg(RC); |
2181 | BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, MCID: TII.get(Opcode: PPC::ORIS8), |
2182 | DestReg: TmpReg3).addReg(RegNo: TmpReg2).addImm(Val: Hi); |
2183 | } else |
2184 | TmpReg3 = TmpReg2; |
2185 | |
2186 | if ((Lo = Remainder & 0xFFFF)) { |
2187 | Register ResultReg = createResultReg(RC); |
2188 | BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, MCID: TII.get(Opcode: PPC::ORI8), |
2189 | DestReg: ResultReg).addReg(RegNo: TmpReg3).addImm(Val: Lo); |
2190 | return ResultReg; |
2191 | } |
2192 | |
2193 | return TmpReg3; |
2194 | } |
2195 | |
2196 | // Materialize an integer constant into a register, and return |
2197 | // the register number (or zero if we failed to handle it). |
2198 | Register PPCFastISel::PPCMaterializeInt(const ConstantInt *CI, MVT VT, |
2199 | bool UseSExt) { |
2200 | // If we're using CR bit registers for i1 values, handle that as a special |
2201 | // case first. |
2202 | if (VT == MVT::i1 && Subtarget->useCRBits()) { |
2203 | Register ImmReg = createResultReg(RC: &PPC::CRBITRCRegClass); |
2204 | BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, |
2205 | MCID: TII.get(Opcode: CI->isZero() ? PPC::CRUNSET : PPC::CRSET), DestReg: ImmReg); |
2206 | return ImmReg; |
2207 | } |
2208 | |
2209 | if (VT != MVT::i64 && VT != MVT::i32 && VT != MVT::i16 && VT != MVT::i8 && |
2210 | VT != MVT::i1) |
2211 | return Register(); |
2212 | |
2213 | const TargetRegisterClass *RC = |
2214 | ((VT == MVT::i64) ? &PPC::G8RCRegClass : &PPC::GPRCRegClass); |
2215 | int64_t Imm = UseSExt ? CI->getSExtValue() : CI->getZExtValue(); |
2216 | |
2217 | // If the constant is in range, use a load-immediate. |
2218 | // Since LI will sign extend the constant we need to make sure that for |
2219 | // our zeroext constants that the sign extended constant fits into 16-bits - |
2220 | // a range of 0..0x7fff. |
2221 | if (isInt<16>(x: Imm)) { |
2222 | unsigned Opc = (VT == MVT::i64) ? PPC::LI8 : PPC::LI; |
2223 | Register ImmReg = createResultReg(RC); |
2224 | BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, MCID: TII.get(Opcode: Opc), DestReg: ImmReg) |
2225 | .addImm(Val: Imm); |
2226 | return ImmReg; |
2227 | } |
2228 | |
2229 | // Construct the constant piecewise. |
2230 | if (VT == MVT::i64) |
2231 | return PPCMaterialize64BitInt(Imm, RC); |
2232 | else if (VT == MVT::i32) |
2233 | return PPCMaterialize32BitInt(Imm, RC); |
2234 | |
2235 | return Register(); |
2236 | } |
2237 | |
2238 | // Materialize a constant into a register, and return the register |
2239 | // number (or zero if we failed to handle it). |
2240 | Register PPCFastISel::fastMaterializeConstant(const Constant *C) { |
2241 | EVT CEVT = TLI.getValueType(DL, Ty: C->getType(), AllowUnknown: true); |
2242 | |
2243 | // Only handle simple types. |
2244 | if (!CEVT.isSimple()) |
2245 | return Register(); |
2246 | MVT VT = CEVT.getSimpleVT(); |
2247 | |
2248 | if (const ConstantFP *CFP = dyn_cast<ConstantFP>(Val: C)) |
2249 | return PPCMaterializeFP(CFP, VT); |
2250 | else if (const GlobalValue *GV = dyn_cast<GlobalValue>(Val: C)) |
2251 | return PPCMaterializeGV(GV, VT); |
2252 | else if (const ConstantInt *CI = dyn_cast<ConstantInt>(Val: C)) |
2253 | // Note that the code in FunctionLoweringInfo::ComputePHILiveOutRegInfo |
2254 | // assumes that constant PHI operands will be zero extended, and failure to |
2255 | // match that assumption will cause problems if we sign extend here but |
2256 | // some user of a PHI is in a block for which we fall back to full SDAG |
2257 | // instruction selection. |
2258 | return PPCMaterializeInt(CI, VT, UseSExt: false); |
2259 | |
2260 | return Register(); |
2261 | } |
2262 | |
2263 | // Materialize the address created by an alloca into a register, and |
2264 | // return the register number (or zero if we failed to handle it). |
2265 | Register PPCFastISel::fastMaterializeAlloca(const AllocaInst *AI) { |
2266 | DenseMap<const AllocaInst *, int>::iterator SI = |
2267 | FuncInfo.StaticAllocaMap.find(Val: AI); |
2268 | |
2269 | // Don't handle dynamic allocas. |
2270 | if (SI == FuncInfo.StaticAllocaMap.end()) |
2271 | return Register(); |
2272 | |
2273 | MVT VT; |
2274 | if (!isLoadTypeLegal(Ty: AI->getType(), VT)) |
2275 | return Register(); |
2276 | |
2277 | if (SI != FuncInfo.StaticAllocaMap.end()) { |
2278 | Register ResultReg = createResultReg(RC: &PPC::G8RC_and_G8RC_NOX0RegClass); |
2279 | BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, MCID: TII.get(Opcode: PPC::ADDI8), |
2280 | DestReg: ResultReg).addFrameIndex(Idx: SI->second).addImm(Val: 0); |
2281 | return ResultReg; |
2282 | } |
2283 | |
2284 | return Register(); |
2285 | } |
2286 | |
2287 | // Fold loads into extends when possible. |
2288 | // FIXME: We can have multiple redundant extend/trunc instructions |
2289 | // following a load. The folding only picks up one. Extend this |
2290 | // to check subsequent instructions for the same pattern and remove |
2291 | // them. Thus ResultReg should be the def reg for the last redundant |
2292 | // instruction in a chain, and all intervening instructions can be |
2293 | // removed from parent. Change test/CodeGen/PowerPC/fast-isel-fold.ll |
2294 | // to add ELF64-NOT: rldicl to the appropriate tests when this works. |
2295 | bool PPCFastISel::tryToFoldLoadIntoMI(MachineInstr *MI, unsigned OpNo, |
2296 | const LoadInst *LI) { |
2297 | // Verify we have a legal type before going any further. |
2298 | MVT VT; |
2299 | if (!isLoadTypeLegal(Ty: LI->getType(), VT)) |
2300 | return false; |
2301 | |
2302 | // Combine load followed by zero- or sign-extend. |
2303 | bool IsZExt = false; |
2304 | switch(MI->getOpcode()) { |
2305 | default: |
2306 | return false; |
2307 | |
2308 | case PPC::RLDICL: |
2309 | case PPC::RLDICL_32_64: { |
2310 | IsZExt = true; |
2311 | unsigned MB = MI->getOperand(i: 3).getImm(); |
2312 | if ((VT == MVT::i8 && MB <= 56) || |
2313 | (VT == MVT::i16 && MB <= 48) || |
2314 | (VT == MVT::i32 && MB <= 32)) |
2315 | break; |
2316 | return false; |
2317 | } |
2318 | |
2319 | case PPC::RLWINM: |
2320 | case PPC::RLWINM8: { |
2321 | IsZExt = true; |
2322 | unsigned MB = MI->getOperand(i: 3).getImm(); |
2323 | if ((VT == MVT::i8 && MB <= 24) || |
2324 | (VT == MVT::i16 && MB <= 16)) |
2325 | break; |
2326 | return false; |
2327 | } |
2328 | |
2329 | case PPC::EXTSB: |
2330 | case PPC::EXTSB8: |
2331 | case PPC::EXTSB8_32_64: |
2332 | /* There is no sign-extending load-byte instruction. */ |
2333 | return false; |
2334 | |
2335 | case PPC::EXTSH: |
2336 | case PPC::EXTSH8: |
2337 | case PPC::EXTSH8_32_64: { |
2338 | if (VT != MVT::i16 && VT != MVT::i8) |
2339 | return false; |
2340 | break; |
2341 | } |
2342 | |
2343 | case PPC::EXTSW: |
2344 | case PPC::EXTSW_32: |
2345 | case PPC::EXTSW_32_64: { |
2346 | if (VT != MVT::i32 && VT != MVT::i16 && VT != MVT::i8) |
2347 | return false; |
2348 | break; |
2349 | } |
2350 | } |
2351 | |
2352 | // See if we can handle this address. |
2353 | Address Addr; |
2354 | if (!PPCComputeAddress(Obj: LI->getOperand(i_nocapture: 0), Addr)) |
2355 | return false; |
2356 | |
2357 | Register ResultReg = MI->getOperand(i: 0).getReg(); |
2358 | |
2359 | if (!PPCEmitLoad(VT, ResultReg, Addr, RC: nullptr, IsZExt, |
2360 | FP64LoadOpc: Subtarget->hasSPE() ? PPC::EVLDD : PPC::LFD)) |
2361 | return false; |
2362 | |
2363 | MachineBasicBlock::iterator I(MI); |
2364 | removeDeadCode(I, E: std::next(x: I)); |
2365 | return true; |
2366 | } |
2367 | |
2368 | // Attempt to lower call arguments in a faster way than done by |
2369 | // the selection DAG code. |
2370 | bool PPCFastISel::fastLowerArguments() { |
2371 | // Defer to normal argument lowering for now. It's reasonably |
2372 | // efficient. Consider doing something like ARM to handle the |
2373 | // case where all args fit in registers, no varargs, no float |
2374 | // or vector args. |
2375 | return false; |
2376 | } |
2377 | |
2378 | // Handle materializing integer constants into a register. This is not |
2379 | // automatically generated for PowerPC, so must be explicitly created here. |
2380 | Register PPCFastISel::fastEmit_i(MVT Ty, MVT VT, unsigned Opc, uint64_t Imm) { |
2381 | |
2382 | if (Opc != ISD::Constant) |
2383 | return Register(); |
2384 | |
2385 | // If we're using CR bit registers for i1 values, handle that as a special |
2386 | // case first. |
2387 | if (VT == MVT::i1 && Subtarget->useCRBits()) { |
2388 | Register ImmReg = createResultReg(RC: &PPC::CRBITRCRegClass); |
2389 | BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, |
2390 | MCID: TII.get(Opcode: Imm == 0 ? PPC::CRUNSET : PPC::CRSET), DestReg: ImmReg); |
2391 | return ImmReg; |
2392 | } |
2393 | |
2394 | if (VT != MVT::i64 && VT != MVT::i32 && VT != MVT::i16 && VT != MVT::i8 && |
2395 | VT != MVT::i1) |
2396 | return Register(); |
2397 | |
2398 | const TargetRegisterClass *RC = ((VT == MVT::i64) ? &PPC::G8RCRegClass : |
2399 | &PPC::GPRCRegClass); |
2400 | if (VT == MVT::i64) |
2401 | return PPCMaterialize64BitInt(Imm, RC); |
2402 | else |
2403 | return PPCMaterialize32BitInt(Imm, RC); |
2404 | } |
2405 | |
2406 | // Override for ADDI and ADDI8 to set the correct register class |
2407 | // on RHS operand 0. The automatic infrastructure naively assumes |
2408 | // GPRC for i32 and G8RC for i64; the concept of "no R0" is lost |
2409 | // for these cases. At the moment, none of the other automatically |
2410 | // generated RI instructions require special treatment. However, once |
2411 | // SelectSelect is implemented, "isel" requires similar handling. |
2412 | // |
2413 | // Also be conservative about the output register class. Avoid |
2414 | // assigning R0 or X0 to the output register for GPRC and G8RC |
2415 | // register classes, as any such result could be used in ADDI, etc., |
2416 | // where those regs have another meaning. |
2417 | Register PPCFastISel::fastEmitInst_ri(unsigned MachineInstOpcode, |
2418 | const TargetRegisterClass *RC, |
2419 | Register Op0, uint64_t Imm) { |
2420 | if (MachineInstOpcode == PPC::ADDI) |
2421 | MRI.setRegClass(Reg: Op0, RC: &PPC::GPRC_and_GPRC_NOR0RegClass); |
2422 | else if (MachineInstOpcode == PPC::ADDI8) |
2423 | MRI.setRegClass(Reg: Op0, RC: &PPC::G8RC_and_G8RC_NOX0RegClass); |
2424 | |
2425 | const TargetRegisterClass *UseRC = |
2426 | (RC == &PPC::GPRCRegClass ? &PPC::GPRC_and_GPRC_NOR0RegClass : |
2427 | (RC == &PPC::G8RCRegClass ? &PPC::G8RC_and_G8RC_NOX0RegClass : RC)); |
2428 | |
2429 | return FastISel::fastEmitInst_ri(MachineInstOpcode, RC: UseRC, Op0, Imm); |
2430 | } |
2431 | |
2432 | // Override for instructions with one register operand to avoid use of |
2433 | // R0/X0. The automatic infrastructure isn't aware of the context so |
2434 | // we must be conservative. |
2435 | Register PPCFastISel::fastEmitInst_r(unsigned MachineInstOpcode, |
2436 | const TargetRegisterClass *RC, |
2437 | Register Op0) { |
2438 | const TargetRegisterClass *UseRC = |
2439 | (RC == &PPC::GPRCRegClass ? &PPC::GPRC_and_GPRC_NOR0RegClass : |
2440 | (RC == &PPC::G8RCRegClass ? &PPC::G8RC_and_G8RC_NOX0RegClass : RC)); |
2441 | |
2442 | return FastISel::fastEmitInst_r(MachineInstOpcode, RC: UseRC, Op0); |
2443 | } |
2444 | |
2445 | // Override for instructions with two register operands to avoid use |
2446 | // of R0/X0. The automatic infrastructure isn't aware of the context |
2447 | // so we must be conservative. |
2448 | Register PPCFastISel::fastEmitInst_rr(unsigned MachineInstOpcode, |
2449 | const TargetRegisterClass *RC, |
2450 | Register Op0, Register Op1) { |
2451 | const TargetRegisterClass *UseRC = |
2452 | (RC == &PPC::GPRCRegClass ? &PPC::GPRC_and_GPRC_NOR0RegClass : |
2453 | (RC == &PPC::G8RCRegClass ? &PPC::G8RC_and_G8RC_NOX0RegClass : RC)); |
2454 | |
2455 | return FastISel::fastEmitInst_rr(MachineInstOpcode, RC: UseRC, Op0, Op1); |
2456 | } |
2457 | |
2458 | namespace llvm { |
2459 | // Create the fast instruction selector for PowerPC64 ELF. |
2460 | FastISel *PPC::createFastISel(FunctionLoweringInfo &FuncInfo, |
2461 | const TargetLibraryInfo *LibInfo) { |
2462 | // Only available on 64-bit for now. |
2463 | const PPCSubtarget &Subtarget = FuncInfo.MF->getSubtarget<PPCSubtarget>(); |
2464 | if (Subtarget.isPPC64()) |
2465 | return new PPCFastISel(FuncInfo, LibInfo); |
2466 | return nullptr; |
2467 | } |
2468 | } |
2469 | |