1 | //===-- X86FastISel.cpp - X86 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 X86-specific support for the FastISel class. Much |
10 | // of the target-specific code is generated by tablegen in the file |
11 | // X86GenFastISel.inc, which is #included here. |
12 | // |
13 | //===----------------------------------------------------------------------===// |
14 | |
15 | #include "X86.h" |
16 | #include "X86CallingConv.h" |
17 | #include "X86InstrBuilder.h" |
18 | #include "X86InstrInfo.h" |
19 | #include "X86MachineFunctionInfo.h" |
20 | #include "X86RegisterInfo.h" |
21 | #include "X86Subtarget.h" |
22 | #include "X86TargetMachine.h" |
23 | #include "llvm/Analysis/BranchProbabilityInfo.h" |
24 | #include "llvm/CodeGen/FastISel.h" |
25 | #include "llvm/CodeGen/FunctionLoweringInfo.h" |
26 | #include "llvm/CodeGen/MachineConstantPool.h" |
27 | #include "llvm/CodeGen/MachineFrameInfo.h" |
28 | #include "llvm/CodeGen/MachineRegisterInfo.h" |
29 | #include "llvm/IR/CallingConv.h" |
30 | #include "llvm/IR/DebugInfo.h" |
31 | #include "llvm/IR/DerivedTypes.h" |
32 | #include "llvm/IR/GetElementPtrTypeIterator.h" |
33 | #include "llvm/IR/GlobalAlias.h" |
34 | #include "llvm/IR/GlobalVariable.h" |
35 | #include "llvm/IR/Instructions.h" |
36 | #include "llvm/IR/IntrinsicInst.h" |
37 | #include "llvm/IR/IntrinsicsX86.h" |
38 | #include "llvm/IR/Operator.h" |
39 | #include "llvm/MC/MCAsmInfo.h" |
40 | #include "llvm/MC/MCSymbol.h" |
41 | #include "llvm/Support/ErrorHandling.h" |
42 | #include "llvm/Target/TargetOptions.h" |
43 | using namespace llvm; |
44 | |
45 | namespace { |
46 | |
47 | class X86FastISel final : public FastISel { |
48 | /// Subtarget - Keep a pointer to the X86Subtarget around so that we can |
49 | /// make the right decision when generating code for different targets. |
50 | const X86Subtarget *Subtarget; |
51 | |
52 | public: |
53 | explicit X86FastISel(FunctionLoweringInfo &funcInfo, |
54 | const TargetLibraryInfo *libInfo) |
55 | : FastISel(funcInfo, libInfo) { |
56 | Subtarget = &funcInfo.MF->getSubtarget<X86Subtarget>(); |
57 | } |
58 | |
59 | bool fastSelectInstruction(const Instruction *I) override; |
60 | |
61 | /// The specified machine instr operand is a vreg, and that |
62 | /// vreg is being provided by the specified load instruction. If possible, |
63 | /// try to fold the load as an operand to the instruction, returning true if |
64 | /// possible. |
65 | bool tryToFoldLoadIntoMI(MachineInstr *MI, unsigned OpNo, |
66 | const LoadInst *LI) override; |
67 | |
68 | bool fastLowerArguments() override; |
69 | bool fastLowerCall(CallLoweringInfo &CLI) override; |
70 | bool fastLowerIntrinsicCall(const IntrinsicInst *II) override; |
71 | |
72 | #include "X86GenFastISel.inc" |
73 | |
74 | private: |
75 | bool X86FastEmitCompare(const Value *LHS, const Value *RHS, EVT VT, |
76 | const DebugLoc &DL); |
77 | |
78 | bool X86FastEmitLoad(MVT VT, X86AddressMode &AM, MachineMemOperand *MMO, |
79 | unsigned &ResultReg, unsigned Alignment = 1); |
80 | |
81 | bool X86FastEmitStore(EVT VT, const Value *Val, X86AddressMode &AM, |
82 | MachineMemOperand *MMO = nullptr, bool Aligned = false); |
83 | bool X86FastEmitStore(EVT VT, unsigned ValReg, X86AddressMode &AM, |
84 | MachineMemOperand *MMO = nullptr, bool Aligned = false); |
85 | |
86 | bool X86FastEmitExtend(ISD::NodeType Opc, EVT DstVT, unsigned Src, EVT SrcVT, |
87 | unsigned &ResultReg); |
88 | |
89 | bool X86SelectAddress(const Value *V, X86AddressMode &AM); |
90 | bool X86SelectCallAddress(const Value *V, X86AddressMode &AM); |
91 | |
92 | bool X86SelectLoad(const Instruction *I); |
93 | |
94 | bool X86SelectStore(const Instruction *I); |
95 | |
96 | bool X86SelectRet(const Instruction *I); |
97 | |
98 | bool X86SelectCmp(const Instruction *I); |
99 | |
100 | bool X86SelectZExt(const Instruction *I); |
101 | |
102 | bool X86SelectSExt(const Instruction *I); |
103 | |
104 | bool X86SelectBranch(const Instruction *I); |
105 | |
106 | bool X86SelectShift(const Instruction *I); |
107 | |
108 | bool X86SelectDivRem(const Instruction *I); |
109 | |
110 | bool X86FastEmitCMoveSelect(MVT RetVT, const Instruction *I); |
111 | |
112 | bool X86FastEmitSSESelect(MVT RetVT, const Instruction *I); |
113 | |
114 | bool X86FastEmitPseudoSelect(MVT RetVT, const Instruction *I); |
115 | |
116 | bool X86SelectSelect(const Instruction *I); |
117 | |
118 | bool X86SelectTrunc(const Instruction *I); |
119 | |
120 | bool X86SelectFPExtOrFPTrunc(const Instruction *I, unsigned Opc, |
121 | const TargetRegisterClass *RC); |
122 | |
123 | bool X86SelectFPExt(const Instruction *I); |
124 | bool X86SelectFPTrunc(const Instruction *I); |
125 | bool X86SelectSIToFP(const Instruction *I); |
126 | bool X86SelectUIToFP(const Instruction *I); |
127 | bool X86SelectIntToFP(const Instruction *I, bool IsSigned); |
128 | |
129 | const X86InstrInfo *getInstrInfo() const { |
130 | return Subtarget->getInstrInfo(); |
131 | } |
132 | const X86TargetMachine *getTargetMachine() const { |
133 | return static_cast<const X86TargetMachine *>(&TM); |
134 | } |
135 | |
136 | bool handleConstantAddresses(const Value *V, X86AddressMode &AM); |
137 | |
138 | unsigned X86MaterializeInt(const ConstantInt *CI, MVT VT); |
139 | unsigned X86MaterializeFP(const ConstantFP *CFP, MVT VT); |
140 | unsigned X86MaterializeGV(const GlobalValue *GV, MVT VT); |
141 | unsigned fastMaterializeConstant(const Constant *C) override; |
142 | |
143 | unsigned fastMaterializeAlloca(const AllocaInst *C) override; |
144 | |
145 | unsigned fastMaterializeFloatZero(const ConstantFP *CF) override; |
146 | |
147 | /// isScalarFPTypeInSSEReg - Return true if the specified scalar FP type is |
148 | /// computed in an SSE register, not on the X87 floating point stack. |
149 | bool isScalarFPTypeInSSEReg(EVT VT) const { |
150 | return (VT == MVT::f64 && Subtarget->hasSSE2()) || |
151 | (VT == MVT::f32 && Subtarget->hasSSE1()) || VT == MVT::f16; |
152 | } |
153 | |
154 | bool isTypeLegal(Type *Ty, MVT &VT, bool AllowI1 = false); |
155 | |
156 | bool IsMemcpySmall(uint64_t Len); |
157 | |
158 | bool TryEmitSmallMemcpy(X86AddressMode DestAM, |
159 | X86AddressMode SrcAM, uint64_t Len); |
160 | |
161 | bool foldX86XALUIntrinsic(X86::CondCode &CC, const Instruction *I, |
162 | const Value *Cond); |
163 | |
164 | const MachineInstrBuilder &addFullAddress(const MachineInstrBuilder &MIB, |
165 | X86AddressMode &AM); |
166 | |
167 | unsigned fastEmitInst_rrrr(unsigned MachineInstOpcode, |
168 | const TargetRegisterClass *RC, unsigned Op0, |
169 | unsigned Op1, unsigned Op2, unsigned Op3); |
170 | }; |
171 | |
172 | } // end anonymous namespace. |
173 | |
174 | static std::pair<unsigned, bool> |
175 | getX86SSEConditionCode(CmpInst::Predicate Predicate) { |
176 | unsigned CC; |
177 | bool NeedSwap = false; |
178 | |
179 | // SSE Condition code mapping: |
180 | // 0 - EQ |
181 | // 1 - LT |
182 | // 2 - LE |
183 | // 3 - UNORD |
184 | // 4 - NEQ |
185 | // 5 - NLT |
186 | // 6 - NLE |
187 | // 7 - ORD |
188 | switch (Predicate) { |
189 | default: llvm_unreachable("Unexpected predicate" ); |
190 | case CmpInst::FCMP_OEQ: CC = 0; break; |
191 | case CmpInst::FCMP_OGT: NeedSwap = true; [[fallthrough]]; |
192 | case CmpInst::FCMP_OLT: CC = 1; break; |
193 | case CmpInst::FCMP_OGE: NeedSwap = true; [[fallthrough]]; |
194 | case CmpInst::FCMP_OLE: CC = 2; break; |
195 | case CmpInst::FCMP_UNO: CC = 3; break; |
196 | case CmpInst::FCMP_UNE: CC = 4; break; |
197 | case CmpInst::FCMP_ULE: NeedSwap = true; [[fallthrough]]; |
198 | case CmpInst::FCMP_UGE: CC = 5; break; |
199 | case CmpInst::FCMP_ULT: NeedSwap = true; [[fallthrough]]; |
200 | case CmpInst::FCMP_UGT: CC = 6; break; |
201 | case CmpInst::FCMP_ORD: CC = 7; break; |
202 | case CmpInst::FCMP_UEQ: CC = 8; break; |
203 | case CmpInst::FCMP_ONE: CC = 12; break; |
204 | } |
205 | |
206 | return std::make_pair(x&: CC, y&: NeedSwap); |
207 | } |
208 | |
209 | /// Adds a complex addressing mode to the given machine instr builder. |
210 | /// Note, this will constrain the index register. If its not possible to |
211 | /// constrain the given index register, then a new one will be created. The |
212 | /// IndexReg field of the addressing mode will be updated to match in this case. |
213 | const MachineInstrBuilder & |
214 | X86FastISel::addFullAddress(const MachineInstrBuilder &MIB, |
215 | X86AddressMode &AM) { |
216 | // First constrain the index register. It needs to be a GR64_NOSP. |
217 | AM.IndexReg = constrainOperandRegClass(II: MIB->getDesc(), Op: AM.IndexReg, |
218 | OpNum: MIB->getNumOperands() + |
219 | X86::AddrIndexReg); |
220 | return ::addFullAddress(MIB, AM); |
221 | } |
222 | |
223 | /// Check if it is possible to fold the condition from the XALU intrinsic |
224 | /// into the user. The condition code will only be updated on success. |
225 | bool X86FastISel::foldX86XALUIntrinsic(X86::CondCode &CC, const Instruction *I, |
226 | const Value *Cond) { |
227 | if (!isa<ExtractValueInst>(Val: Cond)) |
228 | return false; |
229 | |
230 | const auto *EV = cast<ExtractValueInst>(Val: Cond); |
231 | if (!isa<IntrinsicInst>(Val: EV->getAggregateOperand())) |
232 | return false; |
233 | |
234 | const auto *II = cast<IntrinsicInst>(Val: EV->getAggregateOperand()); |
235 | MVT RetVT; |
236 | const Function *Callee = II->getCalledFunction(); |
237 | Type *RetTy = |
238 | cast<StructType>(Val: Callee->getReturnType())->getTypeAtIndex(N: 0U); |
239 | if (!isTypeLegal(Ty: RetTy, VT&: RetVT)) |
240 | return false; |
241 | |
242 | if (RetVT != MVT::i32 && RetVT != MVT::i64) |
243 | return false; |
244 | |
245 | X86::CondCode TmpCC; |
246 | switch (II->getIntrinsicID()) { |
247 | default: return false; |
248 | case Intrinsic::sadd_with_overflow: |
249 | case Intrinsic::ssub_with_overflow: |
250 | case Intrinsic::smul_with_overflow: |
251 | case Intrinsic::umul_with_overflow: TmpCC = X86::COND_O; break; |
252 | case Intrinsic::uadd_with_overflow: |
253 | case Intrinsic::usub_with_overflow: TmpCC = X86::COND_B; break; |
254 | } |
255 | |
256 | // Check if both instructions are in the same basic block. |
257 | if (II->getParent() != I->getParent()) |
258 | return false; |
259 | |
260 | // Make sure nothing is in the way |
261 | BasicBlock::const_iterator Start(I); |
262 | BasicBlock::const_iterator End(II); |
263 | for (auto Itr = std::prev(x: Start); Itr != End; --Itr) { |
264 | // We only expect extractvalue instructions between the intrinsic and the |
265 | // instruction to be selected. |
266 | if (!isa<ExtractValueInst>(Val: Itr)) |
267 | return false; |
268 | |
269 | // Check that the extractvalue operand comes from the intrinsic. |
270 | const auto *EVI = cast<ExtractValueInst>(Val&: Itr); |
271 | if (EVI->getAggregateOperand() != II) |
272 | return false; |
273 | } |
274 | |
275 | // Make sure no potentially eflags clobbering phi moves can be inserted in |
276 | // between. |
277 | auto HasPhis = [](const BasicBlock *Succ) { return !Succ->phis().empty(); }; |
278 | if (I->isTerminator() && llvm::any_of(Range: successors(I), P: HasPhis)) |
279 | return false; |
280 | |
281 | // Make sure there are no potentially eflags clobbering constant |
282 | // materializations in between. |
283 | if (llvm::any_of(Range: I->operands(), P: [](Value *V) { return isa<Constant>(Val: V); })) |
284 | return false; |
285 | |
286 | CC = TmpCC; |
287 | return true; |
288 | } |
289 | |
290 | bool X86FastISel::isTypeLegal(Type *Ty, MVT &VT, bool AllowI1) { |
291 | EVT evt = TLI.getValueType(DL, Ty, /*AllowUnknown=*/true); |
292 | if (evt == MVT::Other || !evt.isSimple()) |
293 | // Unhandled type. Halt "fast" selection and bail. |
294 | return false; |
295 | |
296 | VT = evt.getSimpleVT(); |
297 | // For now, require SSE/SSE2 for performing floating-point operations, |
298 | // since x87 requires additional work. |
299 | if (VT == MVT::f64 && !Subtarget->hasSSE2()) |
300 | return false; |
301 | if (VT == MVT::f32 && !Subtarget->hasSSE1()) |
302 | return false; |
303 | // Similarly, no f80 support yet. |
304 | if (VT == MVT::f80) |
305 | return false; |
306 | // We only handle legal types. For example, on x86-32 the instruction |
307 | // selector contains all of the 64-bit instructions from x86-64, |
308 | // under the assumption that i64 won't be used if the target doesn't |
309 | // support it. |
310 | return (AllowI1 && VT == MVT::i1) || TLI.isTypeLegal(VT); |
311 | } |
312 | |
313 | /// X86FastEmitLoad - Emit a machine instruction to load a value of type VT. |
314 | /// The address is either pre-computed, i.e. Ptr, or a GlobalAddress, i.e. GV. |
315 | /// Return true and the result register by reference if it is possible. |
316 | bool X86FastISel::X86FastEmitLoad(MVT VT, X86AddressMode &AM, |
317 | MachineMemOperand *MMO, unsigned &ResultReg, |
318 | unsigned Alignment) { |
319 | bool HasSSE1 = Subtarget->hasSSE1(); |
320 | bool HasSSE2 = Subtarget->hasSSE2(); |
321 | bool HasSSE41 = Subtarget->hasSSE41(); |
322 | bool HasAVX = Subtarget->hasAVX(); |
323 | bool HasAVX2 = Subtarget->hasAVX2(); |
324 | bool HasAVX512 = Subtarget->hasAVX512(); |
325 | bool HasVLX = Subtarget->hasVLX(); |
326 | bool IsNonTemporal = MMO && MMO->isNonTemporal(); |
327 | |
328 | // Treat i1 loads the same as i8 loads. Masking will be done when storing. |
329 | if (VT == MVT::i1) |
330 | VT = MVT::i8; |
331 | |
332 | // Get opcode and regclass of the output for the given load instruction. |
333 | unsigned Opc = 0; |
334 | switch (VT.SimpleTy) { |
335 | default: return false; |
336 | case MVT::i8: |
337 | Opc = X86::MOV8rm; |
338 | break; |
339 | case MVT::i16: |
340 | Opc = X86::MOV16rm; |
341 | break; |
342 | case MVT::i32: |
343 | Opc = X86::MOV32rm; |
344 | break; |
345 | case MVT::i64: |
346 | // Must be in x86-64 mode. |
347 | Opc = X86::MOV64rm; |
348 | break; |
349 | case MVT::f32: |
350 | Opc = HasAVX512 ? X86::VMOVSSZrm_alt |
351 | : HasAVX ? X86::VMOVSSrm_alt |
352 | : HasSSE1 ? X86::MOVSSrm_alt |
353 | : X86::LD_Fp32m; |
354 | break; |
355 | case MVT::f64: |
356 | Opc = HasAVX512 ? X86::VMOVSDZrm_alt |
357 | : HasAVX ? X86::VMOVSDrm_alt |
358 | : HasSSE2 ? X86::MOVSDrm_alt |
359 | : X86::LD_Fp64m; |
360 | break; |
361 | case MVT::f80: |
362 | // No f80 support yet. |
363 | return false; |
364 | case MVT::v4f32: |
365 | if (IsNonTemporal && Alignment >= 16 && HasSSE41) |
366 | Opc = HasVLX ? X86::VMOVNTDQAZ128rm : |
367 | HasAVX ? X86::VMOVNTDQArm : X86::MOVNTDQArm; |
368 | else if (Alignment >= 16) |
369 | Opc = HasVLX ? X86::VMOVAPSZ128rm : |
370 | HasAVX ? X86::VMOVAPSrm : X86::MOVAPSrm; |
371 | else |
372 | Opc = HasVLX ? X86::VMOVUPSZ128rm : |
373 | HasAVX ? X86::VMOVUPSrm : X86::MOVUPSrm; |
374 | break; |
375 | case MVT::v2f64: |
376 | if (IsNonTemporal && Alignment >= 16 && HasSSE41) |
377 | Opc = HasVLX ? X86::VMOVNTDQAZ128rm : |
378 | HasAVX ? X86::VMOVNTDQArm : X86::MOVNTDQArm; |
379 | else if (Alignment >= 16) |
380 | Opc = HasVLX ? X86::VMOVAPDZ128rm : |
381 | HasAVX ? X86::VMOVAPDrm : X86::MOVAPDrm; |
382 | else |
383 | Opc = HasVLX ? X86::VMOVUPDZ128rm : |
384 | HasAVX ? X86::VMOVUPDrm : X86::MOVUPDrm; |
385 | break; |
386 | case MVT::v4i32: |
387 | case MVT::v2i64: |
388 | case MVT::v8i16: |
389 | case MVT::v16i8: |
390 | if (IsNonTemporal && Alignment >= 16 && HasSSE41) |
391 | Opc = HasVLX ? X86::VMOVNTDQAZ128rm : |
392 | HasAVX ? X86::VMOVNTDQArm : X86::MOVNTDQArm; |
393 | else if (Alignment >= 16) |
394 | Opc = HasVLX ? X86::VMOVDQA64Z128rm : |
395 | HasAVX ? X86::VMOVDQArm : X86::MOVDQArm; |
396 | else |
397 | Opc = HasVLX ? X86::VMOVDQU64Z128rm : |
398 | HasAVX ? X86::VMOVDQUrm : X86::MOVDQUrm; |
399 | break; |
400 | case MVT::v8f32: |
401 | assert(HasAVX); |
402 | if (IsNonTemporal && Alignment >= 32 && HasAVX2) |
403 | Opc = HasVLX ? X86::VMOVNTDQAZ256rm : X86::VMOVNTDQAYrm; |
404 | else if (IsNonTemporal && Alignment >= 16) |
405 | return false; // Force split for X86::VMOVNTDQArm |
406 | else if (Alignment >= 32) |
407 | Opc = HasVLX ? X86::VMOVAPSZ256rm : X86::VMOVAPSYrm; |
408 | else |
409 | Opc = HasVLX ? X86::VMOVUPSZ256rm : X86::VMOVUPSYrm; |
410 | break; |
411 | case MVT::v4f64: |
412 | assert(HasAVX); |
413 | if (IsNonTemporal && Alignment >= 32 && HasAVX2) |
414 | Opc = HasVLX ? X86::VMOVNTDQAZ256rm : X86::VMOVNTDQAYrm; |
415 | else if (IsNonTemporal && Alignment >= 16) |
416 | return false; // Force split for X86::VMOVNTDQArm |
417 | else if (Alignment >= 32) |
418 | Opc = HasVLX ? X86::VMOVAPDZ256rm : X86::VMOVAPDYrm; |
419 | else |
420 | Opc = HasVLX ? X86::VMOVUPDZ256rm : X86::VMOVUPDYrm; |
421 | break; |
422 | case MVT::v8i32: |
423 | case MVT::v4i64: |
424 | case MVT::v16i16: |
425 | case MVT::v32i8: |
426 | assert(HasAVX); |
427 | if (IsNonTemporal && Alignment >= 32 && HasAVX2) |
428 | Opc = HasVLX ? X86::VMOVNTDQAZ256rm : X86::VMOVNTDQAYrm; |
429 | else if (IsNonTemporal && Alignment >= 16) |
430 | return false; // Force split for X86::VMOVNTDQArm |
431 | else if (Alignment >= 32) |
432 | Opc = HasVLX ? X86::VMOVDQA64Z256rm : X86::VMOVDQAYrm; |
433 | else |
434 | Opc = HasVLX ? X86::VMOVDQU64Z256rm : X86::VMOVDQUYrm; |
435 | break; |
436 | case MVT::v16f32: |
437 | assert(HasAVX512); |
438 | if (IsNonTemporal && Alignment >= 64) |
439 | Opc = X86::VMOVNTDQAZrm; |
440 | else |
441 | Opc = (Alignment >= 64) ? X86::VMOVAPSZrm : X86::VMOVUPSZrm; |
442 | break; |
443 | case MVT::v8f64: |
444 | assert(HasAVX512); |
445 | if (IsNonTemporal && Alignment >= 64) |
446 | Opc = X86::VMOVNTDQAZrm; |
447 | else |
448 | Opc = (Alignment >= 64) ? X86::VMOVAPDZrm : X86::VMOVUPDZrm; |
449 | break; |
450 | case MVT::v8i64: |
451 | case MVT::v16i32: |
452 | case MVT::v32i16: |
453 | case MVT::v64i8: |
454 | assert(HasAVX512); |
455 | // Note: There are a lot more choices based on type with AVX-512, but |
456 | // there's really no advantage when the load isn't masked. |
457 | if (IsNonTemporal && Alignment >= 64) |
458 | Opc = X86::VMOVNTDQAZrm; |
459 | else |
460 | Opc = (Alignment >= 64) ? X86::VMOVDQA64Zrm : X86::VMOVDQU64Zrm; |
461 | break; |
462 | } |
463 | |
464 | const TargetRegisterClass *RC = TLI.getRegClassFor(VT); |
465 | |
466 | ResultReg = createResultReg(RC); |
467 | MachineInstrBuilder MIB = |
468 | BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, MCID: TII.get(Opcode: Opc), DestReg: ResultReg); |
469 | addFullAddress(MIB, AM); |
470 | if (MMO) |
471 | MIB->addMemOperand(MF&: *FuncInfo.MF, MO: MMO); |
472 | return true; |
473 | } |
474 | |
475 | /// X86FastEmitStore - Emit a machine instruction to store a value Val of |
476 | /// type VT. The address is either pre-computed, consisted of a base ptr, Ptr |
477 | /// and a displacement offset, or a GlobalAddress, |
478 | /// i.e. V. Return true if it is possible. |
479 | bool X86FastISel::X86FastEmitStore(EVT VT, unsigned ValReg, X86AddressMode &AM, |
480 | MachineMemOperand *MMO, bool Aligned) { |
481 | bool HasSSE1 = Subtarget->hasSSE1(); |
482 | bool HasSSE2 = Subtarget->hasSSE2(); |
483 | bool HasSSE4A = Subtarget->hasSSE4A(); |
484 | bool HasAVX = Subtarget->hasAVX(); |
485 | bool HasAVX512 = Subtarget->hasAVX512(); |
486 | bool HasVLX = Subtarget->hasVLX(); |
487 | bool IsNonTemporal = MMO && MMO->isNonTemporal(); |
488 | |
489 | // Get opcode and regclass of the output for the given store instruction. |
490 | unsigned Opc = 0; |
491 | switch (VT.getSimpleVT().SimpleTy) { |
492 | case MVT::f80: // No f80 support yet. |
493 | default: return false; |
494 | case MVT::i1: { |
495 | // Mask out all but lowest bit. |
496 | Register AndResult = createResultReg(RC: &X86::GR8RegClass); |
497 | BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, |
498 | MCID: TII.get(Opcode: X86::AND8ri), DestReg: AndResult) |
499 | .addReg(RegNo: ValReg).addImm(Val: 1); |
500 | ValReg = AndResult; |
501 | [[fallthrough]]; // handle i1 as i8. |
502 | } |
503 | case MVT::i8: Opc = X86::MOV8mr; break; |
504 | case MVT::i16: Opc = X86::MOV16mr; break; |
505 | case MVT::i32: |
506 | Opc = (IsNonTemporal && HasSSE2) ? X86::MOVNTImr : X86::MOV32mr; |
507 | break; |
508 | case MVT::i64: |
509 | // Must be in x86-64 mode. |
510 | Opc = (IsNonTemporal && HasSSE2) ? X86::MOVNTI_64mr : X86::MOV64mr; |
511 | break; |
512 | case MVT::f32: |
513 | if (HasSSE1) { |
514 | if (IsNonTemporal && HasSSE4A) |
515 | Opc = X86::MOVNTSS; |
516 | else |
517 | Opc = HasAVX512 ? X86::VMOVSSZmr : |
518 | HasAVX ? X86::VMOVSSmr : X86::MOVSSmr; |
519 | } else |
520 | Opc = X86::ST_Fp32m; |
521 | break; |
522 | case MVT::f64: |
523 | if (HasSSE2) { |
524 | if (IsNonTemporal && HasSSE4A) |
525 | Opc = X86::MOVNTSD; |
526 | else |
527 | Opc = HasAVX512 ? X86::VMOVSDZmr : |
528 | HasAVX ? X86::VMOVSDmr : X86::MOVSDmr; |
529 | } else |
530 | Opc = X86::ST_Fp64m; |
531 | break; |
532 | case MVT::x86mmx: |
533 | Opc = (IsNonTemporal && HasSSE1) ? X86::MMX_MOVNTQmr : X86::MMX_MOVQ64mr; |
534 | break; |
535 | case MVT::v4f32: |
536 | if (Aligned) { |
537 | if (IsNonTemporal) |
538 | Opc = HasVLX ? X86::VMOVNTPSZ128mr : |
539 | HasAVX ? X86::VMOVNTPSmr : X86::MOVNTPSmr; |
540 | else |
541 | Opc = HasVLX ? X86::VMOVAPSZ128mr : |
542 | HasAVX ? X86::VMOVAPSmr : X86::MOVAPSmr; |
543 | } else |
544 | Opc = HasVLX ? X86::VMOVUPSZ128mr : |
545 | HasAVX ? X86::VMOVUPSmr : X86::MOVUPSmr; |
546 | break; |
547 | case MVT::v2f64: |
548 | if (Aligned) { |
549 | if (IsNonTemporal) |
550 | Opc = HasVLX ? X86::VMOVNTPDZ128mr : |
551 | HasAVX ? X86::VMOVNTPDmr : X86::MOVNTPDmr; |
552 | else |
553 | Opc = HasVLX ? X86::VMOVAPDZ128mr : |
554 | HasAVX ? X86::VMOVAPDmr : X86::MOVAPDmr; |
555 | } else |
556 | Opc = HasVLX ? X86::VMOVUPDZ128mr : |
557 | HasAVX ? X86::VMOVUPDmr : X86::MOVUPDmr; |
558 | break; |
559 | case MVT::v4i32: |
560 | case MVT::v2i64: |
561 | case MVT::v8i16: |
562 | case MVT::v16i8: |
563 | if (Aligned) { |
564 | if (IsNonTemporal) |
565 | Opc = HasVLX ? X86::VMOVNTDQZ128mr : |
566 | HasAVX ? X86::VMOVNTDQmr : X86::MOVNTDQmr; |
567 | else |
568 | Opc = HasVLX ? X86::VMOVDQA64Z128mr : |
569 | HasAVX ? X86::VMOVDQAmr : X86::MOVDQAmr; |
570 | } else |
571 | Opc = HasVLX ? X86::VMOVDQU64Z128mr : |
572 | HasAVX ? X86::VMOVDQUmr : X86::MOVDQUmr; |
573 | break; |
574 | case MVT::v8f32: |
575 | assert(HasAVX); |
576 | if (Aligned) { |
577 | if (IsNonTemporal) |
578 | Opc = HasVLX ? X86::VMOVNTPSZ256mr : X86::VMOVNTPSYmr; |
579 | else |
580 | Opc = HasVLX ? X86::VMOVAPSZ256mr : X86::VMOVAPSYmr; |
581 | } else |
582 | Opc = HasVLX ? X86::VMOVUPSZ256mr : X86::VMOVUPSYmr; |
583 | break; |
584 | case MVT::v4f64: |
585 | assert(HasAVX); |
586 | if (Aligned) { |
587 | if (IsNonTemporal) |
588 | Opc = HasVLX ? X86::VMOVNTPDZ256mr : X86::VMOVNTPDYmr; |
589 | else |
590 | Opc = HasVLX ? X86::VMOVAPDZ256mr : X86::VMOVAPDYmr; |
591 | } else |
592 | Opc = HasVLX ? X86::VMOVUPDZ256mr : X86::VMOVUPDYmr; |
593 | break; |
594 | case MVT::v8i32: |
595 | case MVT::v4i64: |
596 | case MVT::v16i16: |
597 | case MVT::v32i8: |
598 | assert(HasAVX); |
599 | if (Aligned) { |
600 | if (IsNonTemporal) |
601 | Opc = HasVLX ? X86::VMOVNTDQZ256mr : X86::VMOVNTDQYmr; |
602 | else |
603 | Opc = HasVLX ? X86::VMOVDQA64Z256mr : X86::VMOVDQAYmr; |
604 | } else |
605 | Opc = HasVLX ? X86::VMOVDQU64Z256mr : X86::VMOVDQUYmr; |
606 | break; |
607 | case MVT::v16f32: |
608 | assert(HasAVX512); |
609 | if (Aligned) |
610 | Opc = IsNonTemporal ? X86::VMOVNTPSZmr : X86::VMOVAPSZmr; |
611 | else |
612 | Opc = X86::VMOVUPSZmr; |
613 | break; |
614 | case MVT::v8f64: |
615 | assert(HasAVX512); |
616 | if (Aligned) { |
617 | Opc = IsNonTemporal ? X86::VMOVNTPDZmr : X86::VMOVAPDZmr; |
618 | } else |
619 | Opc = X86::VMOVUPDZmr; |
620 | break; |
621 | case MVT::v8i64: |
622 | case MVT::v16i32: |
623 | case MVT::v32i16: |
624 | case MVT::v64i8: |
625 | assert(HasAVX512); |
626 | // Note: There are a lot more choices based on type with AVX-512, but |
627 | // there's really no advantage when the store isn't masked. |
628 | if (Aligned) |
629 | Opc = IsNonTemporal ? X86::VMOVNTDQZmr : X86::VMOVDQA64Zmr; |
630 | else |
631 | Opc = X86::VMOVDQU64Zmr; |
632 | break; |
633 | } |
634 | |
635 | const MCInstrDesc &Desc = TII.get(Opcode: Opc); |
636 | // Some of the instructions in the previous switch use FR128 instead |
637 | // of FR32 for ValReg. Make sure the register we feed the instruction |
638 | // matches its register class constraints. |
639 | // Note: This is fine to do a copy from FR32 to FR128, this is the |
640 | // same registers behind the scene and actually why it did not trigger |
641 | // any bugs before. |
642 | ValReg = constrainOperandRegClass(II: Desc, Op: ValReg, OpNum: Desc.getNumOperands() - 1); |
643 | MachineInstrBuilder MIB = |
644 | BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, MCID: Desc); |
645 | addFullAddress(MIB, AM).addReg(RegNo: ValReg); |
646 | if (MMO) |
647 | MIB->addMemOperand(MF&: *FuncInfo.MF, MO: MMO); |
648 | |
649 | return true; |
650 | } |
651 | |
652 | bool X86FastISel::X86FastEmitStore(EVT VT, const Value *Val, |
653 | X86AddressMode &AM, |
654 | MachineMemOperand *MMO, bool Aligned) { |
655 | // Handle 'null' like i32/i64 0. |
656 | if (isa<ConstantPointerNull>(Val)) |
657 | Val = Constant::getNullValue(Ty: DL.getIntPtrType(C&: Val->getContext())); |
658 | |
659 | // If this is a store of a simple constant, fold the constant into the store. |
660 | if (const ConstantInt *CI = dyn_cast<ConstantInt>(Val)) { |
661 | unsigned Opc = 0; |
662 | bool Signed = true; |
663 | switch (VT.getSimpleVT().SimpleTy) { |
664 | default: break; |
665 | case MVT::i1: |
666 | Signed = false; |
667 | [[fallthrough]]; // Handle as i8. |
668 | case MVT::i8: Opc = X86::MOV8mi; break; |
669 | case MVT::i16: Opc = X86::MOV16mi; break; |
670 | case MVT::i32: Opc = X86::MOV32mi; break; |
671 | case MVT::i64: |
672 | // Must be a 32-bit sign extended value. |
673 | if (isInt<32>(x: CI->getSExtValue())) |
674 | Opc = X86::MOV64mi32; |
675 | break; |
676 | } |
677 | |
678 | if (Opc) { |
679 | MachineInstrBuilder MIB = |
680 | BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, MCID: TII.get(Opcode: Opc)); |
681 | addFullAddress(MIB, AM).addImm(Val: Signed ? (uint64_t) CI->getSExtValue() |
682 | : CI->getZExtValue()); |
683 | if (MMO) |
684 | MIB->addMemOperand(MF&: *FuncInfo.MF, MO: MMO); |
685 | return true; |
686 | } |
687 | } |
688 | |
689 | Register ValReg = getRegForValue(V: Val); |
690 | if (ValReg == 0) |
691 | return false; |
692 | |
693 | return X86FastEmitStore(VT, ValReg, AM, MMO, Aligned); |
694 | } |
695 | |
696 | /// X86FastEmitExtend - Emit a machine instruction to extend a value Src of |
697 | /// type SrcVT to type DstVT using the specified extension opcode Opc (e.g. |
698 | /// ISD::SIGN_EXTEND). |
699 | bool X86FastISel::X86FastEmitExtend(ISD::NodeType Opc, EVT DstVT, |
700 | unsigned Src, EVT SrcVT, |
701 | unsigned &ResultReg) { |
702 | unsigned RR = fastEmit_r(VT: SrcVT.getSimpleVT(), RetVT: DstVT.getSimpleVT(), Opcode: Opc, Op0: Src); |
703 | if (RR == 0) |
704 | return false; |
705 | |
706 | ResultReg = RR; |
707 | return true; |
708 | } |
709 | |
710 | bool X86FastISel::handleConstantAddresses(const Value *V, X86AddressMode &AM) { |
711 | // Handle constant address. |
712 | if (const GlobalValue *GV = dyn_cast<GlobalValue>(Val: V)) { |
713 | // Can't handle alternate code models yet. |
714 | if (TM.getCodeModel() != CodeModel::Small && |
715 | TM.getCodeModel() != CodeModel::Medium) |
716 | return false; |
717 | |
718 | // Can't handle large objects yet. |
719 | if (TM.isLargeGlobalValue(GV)) |
720 | return false; |
721 | |
722 | // Can't handle TLS yet. |
723 | if (GV->isThreadLocal()) |
724 | return false; |
725 | |
726 | // Can't handle !absolute_symbol references yet. |
727 | if (GV->isAbsoluteSymbolRef()) |
728 | return false; |
729 | |
730 | // RIP-relative addresses can't have additional register operands, so if |
731 | // we've already folded stuff into the addressing mode, just force the |
732 | // global value into its own register, which we can use as the basereg. |
733 | if (!Subtarget->isPICStyleRIPRel() || |
734 | (AM.Base.Reg == 0 && AM.IndexReg == 0)) { |
735 | // Okay, we've committed to selecting this global. Set up the address. |
736 | AM.GV = GV; |
737 | |
738 | // Allow the subtarget to classify the global. |
739 | unsigned char GVFlags = Subtarget->classifyGlobalReference(GV); |
740 | |
741 | // If this reference is relative to the pic base, set it now. |
742 | if (isGlobalRelativeToPICBase(TargetFlag: GVFlags)) { |
743 | // FIXME: How do we know Base.Reg is free?? |
744 | AM.Base.Reg = getInstrInfo()->getGlobalBaseReg(MF: FuncInfo.MF); |
745 | } |
746 | |
747 | // Unless the ABI requires an extra load, return a direct reference to |
748 | // the global. |
749 | if (!isGlobalStubReference(TargetFlag: GVFlags)) { |
750 | if (Subtarget->isPICStyleRIPRel()) { |
751 | // Use rip-relative addressing if we can. Above we verified that the |
752 | // base and index registers are unused. |
753 | assert(AM.Base.Reg == 0 && AM.IndexReg == 0); |
754 | AM.Base.Reg = X86::RIP; |
755 | } |
756 | AM.GVOpFlags = GVFlags; |
757 | return true; |
758 | } |
759 | |
760 | // Ok, we need to do a load from a stub. If we've already loaded from |
761 | // this stub, reuse the loaded pointer, otherwise emit the load now. |
762 | DenseMap<const Value *, Register>::iterator I = LocalValueMap.find(Val: V); |
763 | Register LoadReg; |
764 | if (I != LocalValueMap.end() && I->second) { |
765 | LoadReg = I->second; |
766 | } else { |
767 | // Issue load from stub. |
768 | unsigned Opc = 0; |
769 | const TargetRegisterClass *RC = nullptr; |
770 | X86AddressMode StubAM; |
771 | StubAM.Base.Reg = AM.Base.Reg; |
772 | StubAM.GV = GV; |
773 | StubAM.GVOpFlags = GVFlags; |
774 | |
775 | // Prepare for inserting code in the local-value area. |
776 | SavePoint SaveInsertPt = enterLocalValueArea(); |
777 | |
778 | if (TLI.getPointerTy(DL) == MVT::i64) { |
779 | Opc = X86::MOV64rm; |
780 | RC = &X86::GR64RegClass; |
781 | } else { |
782 | Opc = X86::MOV32rm; |
783 | RC = &X86::GR32RegClass; |
784 | } |
785 | |
786 | if (Subtarget->isPICStyleRIPRel() || GVFlags == X86II::MO_GOTPCREL || |
787 | GVFlags == X86II::MO_GOTPCREL_NORELAX) |
788 | StubAM.Base.Reg = X86::RIP; |
789 | |
790 | LoadReg = createResultReg(RC); |
791 | MachineInstrBuilder LoadMI = |
792 | BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, MCID: TII.get(Opcode: Opc), DestReg: LoadReg); |
793 | addFullAddress(MIB: LoadMI, AM&: StubAM); |
794 | |
795 | // Ok, back to normal mode. |
796 | leaveLocalValueArea(Old: SaveInsertPt); |
797 | |
798 | // Prevent loading GV stub multiple times in same MBB. |
799 | LocalValueMap[V] = LoadReg; |
800 | } |
801 | |
802 | // Now construct the final address. Note that the Disp, Scale, |
803 | // and Index values may already be set here. |
804 | AM.Base.Reg = LoadReg; |
805 | AM.GV = nullptr; |
806 | return true; |
807 | } |
808 | } |
809 | |
810 | // If all else fails, try to materialize the value in a register. |
811 | if (!AM.GV || !Subtarget->isPICStyleRIPRel()) { |
812 | if (AM.Base.Reg == 0) { |
813 | AM.Base.Reg = getRegForValue(V); |
814 | return AM.Base.Reg != 0; |
815 | } |
816 | if (AM.IndexReg == 0) { |
817 | assert(AM.Scale == 1 && "Scale with no index!" ); |
818 | AM.IndexReg = getRegForValue(V); |
819 | return AM.IndexReg != 0; |
820 | } |
821 | } |
822 | |
823 | return false; |
824 | } |
825 | |
826 | /// X86SelectAddress - Attempt to fill in an address from the given value. |
827 | /// |
828 | bool X86FastISel::X86SelectAddress(const Value *V, X86AddressMode &AM) { |
829 | SmallVector<const Value *, 32> GEPs; |
830 | redo_gep: |
831 | const User *U = nullptr; |
832 | unsigned Opcode = Instruction::UserOp1; |
833 | if (const Instruction *I = dyn_cast<Instruction>(Val: V)) { |
834 | // Don't walk into other basic blocks; it's possible we haven't |
835 | // visited them yet, so the instructions may not yet be assigned |
836 | // virtual registers. |
837 | if (FuncInfo.StaticAllocaMap.count(Val: static_cast<const AllocaInst *>(V)) || |
838 | FuncInfo.MBBMap[I->getParent()] == FuncInfo.MBB) { |
839 | Opcode = I->getOpcode(); |
840 | U = I; |
841 | } |
842 | } else if (const ConstantExpr *C = dyn_cast<ConstantExpr>(Val: V)) { |
843 | Opcode = C->getOpcode(); |
844 | U = C; |
845 | } |
846 | |
847 | if (PointerType *Ty = dyn_cast<PointerType>(Val: V->getType())) |
848 | if (Ty->getAddressSpace() > 255) |
849 | // Fast instruction selection doesn't support the special |
850 | // address spaces. |
851 | return false; |
852 | |
853 | switch (Opcode) { |
854 | default: break; |
855 | case Instruction::BitCast: |
856 | // Look past bitcasts. |
857 | return X86SelectAddress(V: U->getOperand(i: 0), AM); |
858 | |
859 | case Instruction::IntToPtr: |
860 | // Look past no-op inttoptrs. |
861 | if (TLI.getValueType(DL, Ty: U->getOperand(i: 0)->getType()) == |
862 | TLI.getPointerTy(DL)) |
863 | return X86SelectAddress(V: U->getOperand(i: 0), AM); |
864 | break; |
865 | |
866 | case Instruction::PtrToInt: |
867 | // Look past no-op ptrtoints. |
868 | if (TLI.getValueType(DL, Ty: U->getType()) == TLI.getPointerTy(DL)) |
869 | return X86SelectAddress(V: U->getOperand(i: 0), AM); |
870 | break; |
871 | |
872 | case Instruction::Alloca: { |
873 | // Do static allocas. |
874 | const AllocaInst *A = cast<AllocaInst>(Val: V); |
875 | DenseMap<const AllocaInst *, int>::iterator SI = |
876 | FuncInfo.StaticAllocaMap.find(Val: A); |
877 | if (SI != FuncInfo.StaticAllocaMap.end()) { |
878 | AM.BaseType = X86AddressMode::FrameIndexBase; |
879 | AM.Base.FrameIndex = SI->second; |
880 | return true; |
881 | } |
882 | break; |
883 | } |
884 | |
885 | case Instruction::Add: { |
886 | // Adds of constants are common and easy enough. |
887 | if (const ConstantInt *CI = dyn_cast<ConstantInt>(Val: U->getOperand(i: 1))) { |
888 | uint64_t Disp = (int32_t)AM.Disp + (uint64_t)CI->getSExtValue(); |
889 | // They have to fit in the 32-bit signed displacement field though. |
890 | if (isInt<32>(x: Disp)) { |
891 | AM.Disp = (uint32_t)Disp; |
892 | return X86SelectAddress(V: U->getOperand(i: 0), AM); |
893 | } |
894 | } |
895 | break; |
896 | } |
897 | |
898 | case Instruction::GetElementPtr: { |
899 | X86AddressMode SavedAM = AM; |
900 | |
901 | // Pattern-match simple GEPs. |
902 | uint64_t Disp = (int32_t)AM.Disp; |
903 | unsigned IndexReg = AM.IndexReg; |
904 | unsigned Scale = AM.Scale; |
905 | gep_type_iterator GTI = gep_type_begin(GEP: U); |
906 | // Iterate through the indices, folding what we can. Constants can be |
907 | // folded, and one dynamic index can be handled, if the scale is supported. |
908 | for (User::const_op_iterator i = U->op_begin() + 1, e = U->op_end(); |
909 | i != e; ++i, ++GTI) { |
910 | const Value *Op = *i; |
911 | if (StructType *STy = GTI.getStructTypeOrNull()) { |
912 | const StructLayout *SL = DL.getStructLayout(Ty: STy); |
913 | Disp += SL->getElementOffset(Idx: cast<ConstantInt>(Val: Op)->getZExtValue()); |
914 | continue; |
915 | } |
916 | |
917 | // A array/variable index is always of the form i*S where S is the |
918 | // constant scale size. See if we can push the scale into immediates. |
919 | uint64_t S = GTI.getSequentialElementStride(DL); |
920 | for (;;) { |
921 | if (const ConstantInt *CI = dyn_cast<ConstantInt>(Val: Op)) { |
922 | // Constant-offset addressing. |
923 | Disp += CI->getSExtValue() * S; |
924 | break; |
925 | } |
926 | if (canFoldAddIntoGEP(GEP: U, Add: Op)) { |
927 | // A compatible add with a constant operand. Fold the constant. |
928 | ConstantInt *CI = |
929 | cast<ConstantInt>(Val: cast<AddOperator>(Val: Op)->getOperand(i_nocapture: 1)); |
930 | Disp += CI->getSExtValue() * S; |
931 | // Iterate on the other operand. |
932 | Op = cast<AddOperator>(Val: Op)->getOperand(i_nocapture: 0); |
933 | continue; |
934 | } |
935 | if (IndexReg == 0 && |
936 | (!AM.GV || !Subtarget->isPICStyleRIPRel()) && |
937 | (S == 1 || S == 2 || S == 4 || S == 8)) { |
938 | // Scaled-index addressing. |
939 | Scale = S; |
940 | IndexReg = getRegForGEPIndex(Idx: Op); |
941 | if (IndexReg == 0) |
942 | return false; |
943 | break; |
944 | } |
945 | // Unsupported. |
946 | goto unsupported_gep; |
947 | } |
948 | } |
949 | |
950 | // Check for displacement overflow. |
951 | if (!isInt<32>(x: Disp)) |
952 | break; |
953 | |
954 | AM.IndexReg = IndexReg; |
955 | AM.Scale = Scale; |
956 | AM.Disp = (uint32_t)Disp; |
957 | GEPs.push_back(Elt: V); |
958 | |
959 | if (const GetElementPtrInst *GEP = |
960 | dyn_cast<GetElementPtrInst>(Val: U->getOperand(i: 0))) { |
961 | // Ok, the GEP indices were covered by constant-offset and scaled-index |
962 | // addressing. Update the address state and move on to examining the base. |
963 | V = GEP; |
964 | goto redo_gep; |
965 | } else if (X86SelectAddress(V: U->getOperand(i: 0), AM)) { |
966 | return true; |
967 | } |
968 | |
969 | // If we couldn't merge the gep value into this addr mode, revert back to |
970 | // our address and just match the value instead of completely failing. |
971 | AM = SavedAM; |
972 | |
973 | for (const Value *I : reverse(C&: GEPs)) |
974 | if (handleConstantAddresses(V: I, AM)) |
975 | return true; |
976 | |
977 | return false; |
978 | unsupported_gep: |
979 | // Ok, the GEP indices weren't all covered. |
980 | break; |
981 | } |
982 | } |
983 | |
984 | return handleConstantAddresses(V, AM); |
985 | } |
986 | |
987 | /// X86SelectCallAddress - Attempt to fill in an address from the given value. |
988 | /// |
989 | bool X86FastISel::X86SelectCallAddress(const Value *V, X86AddressMode &AM) { |
990 | const User *U = nullptr; |
991 | unsigned Opcode = Instruction::UserOp1; |
992 | const Instruction *I = dyn_cast<Instruction>(Val: V); |
993 | // Record if the value is defined in the same basic block. |
994 | // |
995 | // This information is crucial to know whether or not folding an |
996 | // operand is valid. |
997 | // Indeed, FastISel generates or reuses a virtual register for all |
998 | // operands of all instructions it selects. Obviously, the definition and |
999 | // its uses must use the same virtual register otherwise the produced |
1000 | // code is incorrect. |
1001 | // Before instruction selection, FunctionLoweringInfo::set sets the virtual |
1002 | // registers for values that are alive across basic blocks. This ensures |
1003 | // that the values are consistently set between across basic block, even |
1004 | // if different instruction selection mechanisms are used (e.g., a mix of |
1005 | // SDISel and FastISel). |
1006 | // For values local to a basic block, the instruction selection process |
1007 | // generates these virtual registers with whatever method is appropriate |
1008 | // for its needs. In particular, FastISel and SDISel do not share the way |
1009 | // local virtual registers are set. |
1010 | // Therefore, this is impossible (or at least unsafe) to share values |
1011 | // between basic blocks unless they use the same instruction selection |
1012 | // method, which is not guarantee for X86. |
1013 | // Moreover, things like hasOneUse could not be used accurately, if we |
1014 | // allow to reference values across basic blocks whereas they are not |
1015 | // alive across basic blocks initially. |
1016 | bool InMBB = true; |
1017 | if (I) { |
1018 | Opcode = I->getOpcode(); |
1019 | U = I; |
1020 | InMBB = I->getParent() == FuncInfo.MBB->getBasicBlock(); |
1021 | } else if (const ConstantExpr *C = dyn_cast<ConstantExpr>(Val: V)) { |
1022 | Opcode = C->getOpcode(); |
1023 | U = C; |
1024 | } |
1025 | |
1026 | switch (Opcode) { |
1027 | default: break; |
1028 | case Instruction::BitCast: |
1029 | // Look past bitcasts if its operand is in the same BB. |
1030 | if (InMBB) |
1031 | return X86SelectCallAddress(V: U->getOperand(i: 0), AM); |
1032 | break; |
1033 | |
1034 | case Instruction::IntToPtr: |
1035 | // Look past no-op inttoptrs if its operand is in the same BB. |
1036 | if (InMBB && |
1037 | TLI.getValueType(DL, Ty: U->getOperand(i: 0)->getType()) == |
1038 | TLI.getPointerTy(DL)) |
1039 | return X86SelectCallAddress(V: U->getOperand(i: 0), AM); |
1040 | break; |
1041 | |
1042 | case Instruction::PtrToInt: |
1043 | // Look past no-op ptrtoints if its operand is in the same BB. |
1044 | if (InMBB && TLI.getValueType(DL, Ty: U->getType()) == TLI.getPointerTy(DL)) |
1045 | return X86SelectCallAddress(V: U->getOperand(i: 0), AM); |
1046 | break; |
1047 | } |
1048 | |
1049 | // Handle constant address. |
1050 | if (const GlobalValue *GV = dyn_cast<GlobalValue>(Val: V)) { |
1051 | // Can't handle alternate code models yet. |
1052 | if (TM.getCodeModel() != CodeModel::Small && |
1053 | TM.getCodeModel() != CodeModel::Medium) |
1054 | return false; |
1055 | |
1056 | // RIP-relative addresses can't have additional register operands. |
1057 | if (Subtarget->isPICStyleRIPRel() && |
1058 | (AM.Base.Reg != 0 || AM.IndexReg != 0)) |
1059 | return false; |
1060 | |
1061 | // Can't handle TLS. |
1062 | if (const GlobalVariable *GVar = dyn_cast<GlobalVariable>(Val: GV)) |
1063 | if (GVar->isThreadLocal()) |
1064 | return false; |
1065 | |
1066 | // Okay, we've committed to selecting this global. Set up the basic address. |
1067 | AM.GV = GV; |
1068 | |
1069 | // Return a direct reference to the global. Fastisel can handle calls to |
1070 | // functions that require loads, such as dllimport and nonlazybind |
1071 | // functions. |
1072 | if (Subtarget->isPICStyleRIPRel()) { |
1073 | // Use rip-relative addressing if we can. Above we verified that the |
1074 | // base and index registers are unused. |
1075 | assert(AM.Base.Reg == 0 && AM.IndexReg == 0); |
1076 | AM.Base.Reg = X86::RIP; |
1077 | } else { |
1078 | AM.GVOpFlags = Subtarget->classifyLocalReference(GV: nullptr); |
1079 | } |
1080 | |
1081 | return true; |
1082 | } |
1083 | |
1084 | // If all else fails, try to materialize the value in a register. |
1085 | if (!AM.GV || !Subtarget->isPICStyleRIPRel()) { |
1086 | auto GetCallRegForValue = [this](const Value *V) { |
1087 | Register Reg = getRegForValue(V); |
1088 | |
1089 | // In 64-bit mode, we need a 64-bit register even if pointers are 32 bits. |
1090 | if (Reg && Subtarget->isTarget64BitILP32()) { |
1091 | Register CopyReg = createResultReg(RC: &X86::GR32RegClass); |
1092 | BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, MCID: TII.get(Opcode: X86::MOV32rr), |
1093 | DestReg: CopyReg) |
1094 | .addReg(RegNo: Reg); |
1095 | |
1096 | Register ExtReg = createResultReg(RC: &X86::GR64RegClass); |
1097 | BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, |
1098 | MCID: TII.get(Opcode: TargetOpcode::SUBREG_TO_REG), DestReg: ExtReg) |
1099 | .addImm(Val: 0) |
1100 | .addReg(RegNo: CopyReg) |
1101 | .addImm(Val: X86::sub_32bit); |
1102 | Reg = ExtReg; |
1103 | } |
1104 | |
1105 | return Reg; |
1106 | }; |
1107 | |
1108 | if (AM.Base.Reg == 0) { |
1109 | AM.Base.Reg = GetCallRegForValue(V); |
1110 | return AM.Base.Reg != 0; |
1111 | } |
1112 | if (AM.IndexReg == 0) { |
1113 | assert(AM.Scale == 1 && "Scale with no index!" ); |
1114 | AM.IndexReg = GetCallRegForValue(V); |
1115 | return AM.IndexReg != 0; |
1116 | } |
1117 | } |
1118 | |
1119 | return false; |
1120 | } |
1121 | |
1122 | |
1123 | /// X86SelectStore - Select and emit code to implement store instructions. |
1124 | bool X86FastISel::X86SelectStore(const Instruction *I) { |
1125 | // Atomic stores need special handling. |
1126 | const StoreInst *S = cast<StoreInst>(Val: I); |
1127 | |
1128 | if (S->isAtomic()) |
1129 | return false; |
1130 | |
1131 | const Value *PtrV = I->getOperand(i: 1); |
1132 | if (TLI.supportSwiftError()) { |
1133 | // Swifterror values can come from either a function parameter with |
1134 | // swifterror attribute or an alloca with swifterror attribute. |
1135 | if (const Argument *Arg = dyn_cast<Argument>(Val: PtrV)) { |
1136 | if (Arg->hasSwiftErrorAttr()) |
1137 | return false; |
1138 | } |
1139 | |
1140 | if (const AllocaInst *Alloca = dyn_cast<AllocaInst>(Val: PtrV)) { |
1141 | if (Alloca->isSwiftError()) |
1142 | return false; |
1143 | } |
1144 | } |
1145 | |
1146 | const Value *Val = S->getValueOperand(); |
1147 | const Value *Ptr = S->getPointerOperand(); |
1148 | |
1149 | MVT VT; |
1150 | if (!isTypeLegal(Ty: Val->getType(), VT, /*AllowI1=*/true)) |
1151 | return false; |
1152 | |
1153 | Align Alignment = S->getAlign(); |
1154 | Align ABIAlignment = DL.getABITypeAlign(Ty: Val->getType()); |
1155 | bool Aligned = Alignment >= ABIAlignment; |
1156 | |
1157 | X86AddressMode AM; |
1158 | if (!X86SelectAddress(V: Ptr, AM)) |
1159 | return false; |
1160 | |
1161 | return X86FastEmitStore(VT, Val, AM, MMO: createMachineMemOperandFor(I), Aligned); |
1162 | } |
1163 | |
1164 | /// X86SelectRet - Select and emit code to implement ret instructions. |
1165 | bool X86FastISel::X86SelectRet(const Instruction *I) { |
1166 | const ReturnInst *Ret = cast<ReturnInst>(Val: I); |
1167 | const Function &F = *I->getParent()->getParent(); |
1168 | const X86MachineFunctionInfo *X86MFInfo = |
1169 | FuncInfo.MF->getInfo<X86MachineFunctionInfo>(); |
1170 | |
1171 | if (!FuncInfo.CanLowerReturn) |
1172 | return false; |
1173 | |
1174 | if (TLI.supportSwiftError() && |
1175 | F.getAttributes().hasAttrSomewhere(Kind: Attribute::SwiftError)) |
1176 | return false; |
1177 | |
1178 | if (TLI.supportSplitCSR(MF: FuncInfo.MF)) |
1179 | return false; |
1180 | |
1181 | CallingConv::ID CC = F.getCallingConv(); |
1182 | if (CC != CallingConv::C && |
1183 | CC != CallingConv::Fast && |
1184 | CC != CallingConv::Tail && |
1185 | CC != CallingConv::SwiftTail && |
1186 | CC != CallingConv::X86_FastCall && |
1187 | CC != CallingConv::X86_StdCall && |
1188 | CC != CallingConv::X86_ThisCall && |
1189 | CC != CallingConv::X86_64_SysV && |
1190 | CC != CallingConv::Win64) |
1191 | return false; |
1192 | |
1193 | // Don't handle popping bytes if they don't fit the ret's immediate. |
1194 | if (!isUInt<16>(x: X86MFInfo->getBytesToPopOnReturn())) |
1195 | return false; |
1196 | |
1197 | // fastcc with -tailcallopt is intended to provide a guaranteed |
1198 | // tail call optimization. Fastisel doesn't know how to do that. |
1199 | if ((CC == CallingConv::Fast && TM.Options.GuaranteedTailCallOpt) || |
1200 | CC == CallingConv::Tail || CC == CallingConv::SwiftTail) |
1201 | return false; |
1202 | |
1203 | // Let SDISel handle vararg functions. |
1204 | if (F.isVarArg()) |
1205 | return false; |
1206 | |
1207 | // Build a list of return value registers. |
1208 | SmallVector<unsigned, 4> RetRegs; |
1209 | |
1210 | if (Ret->getNumOperands() > 0) { |
1211 | SmallVector<ISD::OutputArg, 4> Outs; |
1212 | GetReturnInfo(CC, ReturnType: F.getReturnType(), attr: F.getAttributes(), Outs, TLI, DL); |
1213 | |
1214 | // Analyze operands of the call, assigning locations to each operand. |
1215 | SmallVector<CCValAssign, 16> ValLocs; |
1216 | CCState CCInfo(CC, F.isVarArg(), *FuncInfo.MF, ValLocs, I->getContext()); |
1217 | CCInfo.AnalyzeReturn(Outs, Fn: RetCC_X86); |
1218 | |
1219 | const Value *RV = Ret->getOperand(i_nocapture: 0); |
1220 | Register Reg = getRegForValue(V: RV); |
1221 | if (Reg == 0) |
1222 | return false; |
1223 | |
1224 | // Only handle a single return value for now. |
1225 | if (ValLocs.size() != 1) |
1226 | return false; |
1227 | |
1228 | CCValAssign &VA = ValLocs[0]; |
1229 | |
1230 | // Don't bother handling odd stuff for now. |
1231 | if (VA.getLocInfo() != CCValAssign::Full) |
1232 | return false; |
1233 | // Only handle register returns for now. |
1234 | if (!VA.isRegLoc()) |
1235 | return false; |
1236 | |
1237 | // The calling-convention tables for x87 returns don't tell |
1238 | // the whole story. |
1239 | if (VA.getLocReg() == X86::FP0 || VA.getLocReg() == X86::FP1) |
1240 | return false; |
1241 | |
1242 | unsigned SrcReg = Reg + VA.getValNo(); |
1243 | EVT SrcVT = TLI.getValueType(DL, Ty: RV->getType()); |
1244 | EVT DstVT = VA.getValVT(); |
1245 | // Special handling for extended integers. |
1246 | if (SrcVT != DstVT) { |
1247 | if (SrcVT != MVT::i1 && SrcVT != MVT::i8 && SrcVT != MVT::i16) |
1248 | return false; |
1249 | |
1250 | if (!Outs[0].Flags.isZExt() && !Outs[0].Flags.isSExt()) |
1251 | return false; |
1252 | |
1253 | if (SrcVT == MVT::i1) { |
1254 | if (Outs[0].Flags.isSExt()) |
1255 | return false; |
1256 | SrcReg = fastEmitZExtFromI1(VT: MVT::i8, Op0: SrcReg); |
1257 | SrcVT = MVT::i8; |
1258 | } |
1259 | if (SrcVT != DstVT) { |
1260 | unsigned Op = |
1261 | Outs[0].Flags.isZExt() ? ISD::ZERO_EXTEND : ISD::SIGN_EXTEND; |
1262 | SrcReg = |
1263 | fastEmit_r(VT: SrcVT.getSimpleVT(), RetVT: DstVT.getSimpleVT(), Opcode: Op, Op0: SrcReg); |
1264 | } |
1265 | } |
1266 | |
1267 | // Make the copy. |
1268 | Register DstReg = VA.getLocReg(); |
1269 | const TargetRegisterClass *SrcRC = MRI.getRegClass(Reg: SrcReg); |
1270 | // Avoid a cross-class copy. This is very unlikely. |
1271 | if (!SrcRC->contains(Reg: DstReg)) |
1272 | return false; |
1273 | BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, |
1274 | MCID: TII.get(Opcode: TargetOpcode::COPY), DestReg: DstReg).addReg(RegNo: SrcReg); |
1275 | |
1276 | // Add register to return instruction. |
1277 | RetRegs.push_back(Elt: VA.getLocReg()); |
1278 | } |
1279 | |
1280 | // Swift calling convention does not require we copy the sret argument |
1281 | // into %rax/%eax for the return, and SRetReturnReg is not set for Swift. |
1282 | |
1283 | // All x86 ABIs require that for returning structs by value we copy |
1284 | // the sret argument into %rax/%eax (depending on ABI) for the return. |
1285 | // We saved the argument into a virtual register in the entry block, |
1286 | // so now we copy the value out and into %rax/%eax. |
1287 | if (F.hasStructRetAttr() && CC != CallingConv::Swift && |
1288 | CC != CallingConv::SwiftTail) { |
1289 | Register Reg = X86MFInfo->getSRetReturnReg(); |
1290 | assert(Reg && |
1291 | "SRetReturnReg should have been set in LowerFormalArguments()!" ); |
1292 | unsigned RetReg = Subtarget->isTarget64BitLP64() ? X86::RAX : X86::EAX; |
1293 | BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, |
1294 | MCID: TII.get(Opcode: TargetOpcode::COPY), DestReg: RetReg).addReg(RegNo: Reg); |
1295 | RetRegs.push_back(Elt: RetReg); |
1296 | } |
1297 | |
1298 | // Now emit the RET. |
1299 | MachineInstrBuilder MIB; |
1300 | if (X86MFInfo->getBytesToPopOnReturn()) { |
1301 | MIB = BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, |
1302 | MCID: TII.get(Opcode: Subtarget->is64Bit() ? X86::RETI64 : X86::RETI32)) |
1303 | .addImm(Val: X86MFInfo->getBytesToPopOnReturn()); |
1304 | } else { |
1305 | MIB = BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, |
1306 | MCID: TII.get(Opcode: Subtarget->is64Bit() ? X86::RET64 : X86::RET32)); |
1307 | } |
1308 | for (unsigned Reg : RetRegs) |
1309 | MIB.addReg(RegNo: Reg, flags: RegState::Implicit); |
1310 | return true; |
1311 | } |
1312 | |
1313 | /// X86SelectLoad - Select and emit code to implement load instructions. |
1314 | /// |
1315 | bool X86FastISel::X86SelectLoad(const Instruction *I) { |
1316 | const LoadInst *LI = cast<LoadInst>(Val: I); |
1317 | |
1318 | // Atomic loads need special handling. |
1319 | if (LI->isAtomic()) |
1320 | return false; |
1321 | |
1322 | const Value *SV = I->getOperand(i: 0); |
1323 | if (TLI.supportSwiftError()) { |
1324 | // Swifterror values can come from either a function parameter with |
1325 | // swifterror attribute or an alloca with swifterror attribute. |
1326 | if (const Argument *Arg = dyn_cast<Argument>(Val: SV)) { |
1327 | if (Arg->hasSwiftErrorAttr()) |
1328 | return false; |
1329 | } |
1330 | |
1331 | if (const AllocaInst *Alloca = dyn_cast<AllocaInst>(Val: SV)) { |
1332 | if (Alloca->isSwiftError()) |
1333 | return false; |
1334 | } |
1335 | } |
1336 | |
1337 | MVT VT; |
1338 | if (!isTypeLegal(Ty: LI->getType(), VT, /*AllowI1=*/true)) |
1339 | return false; |
1340 | |
1341 | const Value *Ptr = LI->getPointerOperand(); |
1342 | |
1343 | X86AddressMode AM; |
1344 | if (!X86SelectAddress(V: Ptr, AM)) |
1345 | return false; |
1346 | |
1347 | unsigned ResultReg = 0; |
1348 | if (!X86FastEmitLoad(VT, AM, MMO: createMachineMemOperandFor(I: LI), ResultReg, |
1349 | Alignment: LI->getAlign().value())) |
1350 | return false; |
1351 | |
1352 | updateValueMap(I, Reg: ResultReg); |
1353 | return true; |
1354 | } |
1355 | |
1356 | static unsigned X86ChooseCmpOpcode(EVT VT, const X86Subtarget *Subtarget) { |
1357 | bool HasAVX512 = Subtarget->hasAVX512(); |
1358 | bool HasAVX = Subtarget->hasAVX(); |
1359 | bool HasSSE1 = Subtarget->hasSSE1(); |
1360 | bool HasSSE2 = Subtarget->hasSSE2(); |
1361 | |
1362 | switch (VT.getSimpleVT().SimpleTy) { |
1363 | default: return 0; |
1364 | case MVT::i8: return X86::CMP8rr; |
1365 | case MVT::i16: return X86::CMP16rr; |
1366 | case MVT::i32: return X86::CMP32rr; |
1367 | case MVT::i64: return X86::CMP64rr; |
1368 | case MVT::f32: |
1369 | return HasAVX512 ? X86::VUCOMISSZrr |
1370 | : HasAVX ? X86::VUCOMISSrr |
1371 | : HasSSE1 ? X86::UCOMISSrr |
1372 | : 0; |
1373 | case MVT::f64: |
1374 | return HasAVX512 ? X86::VUCOMISDZrr |
1375 | : HasAVX ? X86::VUCOMISDrr |
1376 | : HasSSE2 ? X86::UCOMISDrr |
1377 | : 0; |
1378 | } |
1379 | } |
1380 | |
1381 | /// If we have a comparison with RHS as the RHS of the comparison, return an |
1382 | /// opcode that works for the compare (e.g. CMP32ri) otherwise return 0. |
1383 | static unsigned X86ChooseCmpImmediateOpcode(EVT VT, const ConstantInt *RHSC) { |
1384 | switch (VT.getSimpleVT().SimpleTy) { |
1385 | // Otherwise, we can't fold the immediate into this comparison. |
1386 | default: |
1387 | return 0; |
1388 | case MVT::i8: |
1389 | return X86::CMP8ri; |
1390 | case MVT::i16: |
1391 | return X86::CMP16ri; |
1392 | case MVT::i32: |
1393 | return X86::CMP32ri; |
1394 | case MVT::i64: |
1395 | // 64-bit comparisons are only valid if the immediate fits in a 32-bit sext |
1396 | // field. |
1397 | return isInt<32>(x: RHSC->getSExtValue()) ? X86::CMP64ri32 : 0; |
1398 | } |
1399 | } |
1400 | |
1401 | bool X86FastISel::X86FastEmitCompare(const Value *Op0, const Value *Op1, EVT VT, |
1402 | const DebugLoc &CurMIMD) { |
1403 | Register Op0Reg = getRegForValue(V: Op0); |
1404 | if (Op0Reg == 0) return false; |
1405 | |
1406 | // Handle 'null' like i32/i64 0. |
1407 | if (isa<ConstantPointerNull>(Val: Op1)) |
1408 | Op1 = Constant::getNullValue(Ty: DL.getIntPtrType(C&: Op0->getContext())); |
1409 | |
1410 | // We have two options: compare with register or immediate. If the RHS of |
1411 | // the compare is an immediate that we can fold into this compare, use |
1412 | // CMPri, otherwise use CMPrr. |
1413 | if (const ConstantInt *Op1C = dyn_cast<ConstantInt>(Val: Op1)) { |
1414 | if (unsigned CompareImmOpc = X86ChooseCmpImmediateOpcode(VT, RHSC: Op1C)) { |
1415 | BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD: CurMIMD, MCID: TII.get(Opcode: CompareImmOpc)) |
1416 | .addReg(RegNo: Op0Reg) |
1417 | .addImm(Val: Op1C->getSExtValue()); |
1418 | return true; |
1419 | } |
1420 | } |
1421 | |
1422 | unsigned CompareOpc = X86ChooseCmpOpcode(VT, Subtarget); |
1423 | if (CompareOpc == 0) return false; |
1424 | |
1425 | Register Op1Reg = getRegForValue(V: Op1); |
1426 | if (Op1Reg == 0) return false; |
1427 | BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD: CurMIMD, MCID: TII.get(Opcode: CompareOpc)) |
1428 | .addReg(RegNo: Op0Reg) |
1429 | .addReg(RegNo: Op1Reg); |
1430 | |
1431 | return true; |
1432 | } |
1433 | |
1434 | bool X86FastISel::X86SelectCmp(const Instruction *I) { |
1435 | const CmpInst *CI = cast<CmpInst>(Val: I); |
1436 | |
1437 | MVT VT; |
1438 | if (!isTypeLegal(Ty: I->getOperand(i: 0)->getType(), VT)) |
1439 | return false; |
1440 | |
1441 | // Below code only works for scalars. |
1442 | if (VT.isVector()) |
1443 | return false; |
1444 | |
1445 | // Try to optimize or fold the cmp. |
1446 | CmpInst::Predicate Predicate = optimizeCmpPredicate(CI); |
1447 | unsigned ResultReg = 0; |
1448 | switch (Predicate) { |
1449 | default: break; |
1450 | case CmpInst::FCMP_FALSE: { |
1451 | ResultReg = createResultReg(RC: &X86::GR32RegClass); |
1452 | BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, MCID: TII.get(Opcode: X86::MOV32r0), |
1453 | DestReg: ResultReg); |
1454 | ResultReg = fastEmitInst_extractsubreg(RetVT: MVT::i8, Op0: ResultReg, Idx: X86::sub_8bit); |
1455 | if (!ResultReg) |
1456 | return false; |
1457 | break; |
1458 | } |
1459 | case CmpInst::FCMP_TRUE: { |
1460 | ResultReg = createResultReg(RC: &X86::GR8RegClass); |
1461 | BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, MCID: TII.get(Opcode: X86::MOV8ri), |
1462 | DestReg: ResultReg).addImm(Val: 1); |
1463 | break; |
1464 | } |
1465 | } |
1466 | |
1467 | if (ResultReg) { |
1468 | updateValueMap(I, Reg: ResultReg); |
1469 | return true; |
1470 | } |
1471 | |
1472 | const Value *LHS = CI->getOperand(i_nocapture: 0); |
1473 | const Value *RHS = CI->getOperand(i_nocapture: 1); |
1474 | |
1475 | // The optimizer might have replaced fcmp oeq %x, %x with fcmp ord %x, 0.0. |
1476 | // We don't have to materialize a zero constant for this case and can just use |
1477 | // %x again on the RHS. |
1478 | if (Predicate == CmpInst::FCMP_ORD || Predicate == CmpInst::FCMP_UNO) { |
1479 | const auto *RHSC = dyn_cast<ConstantFP>(Val: RHS); |
1480 | if (RHSC && RHSC->isNullValue()) |
1481 | RHS = LHS; |
1482 | } |
1483 | |
1484 | // FCMP_OEQ and FCMP_UNE cannot be checked with a single instruction. |
1485 | static const uint16_t SETFOpcTable[2][3] = { |
1486 | { X86::COND_E, X86::COND_NP, X86::AND8rr }, |
1487 | { X86::COND_NE, X86::COND_P, X86::OR8rr } |
1488 | }; |
1489 | const uint16_t *SETFOpc = nullptr; |
1490 | switch (Predicate) { |
1491 | default: break; |
1492 | case CmpInst::FCMP_OEQ: SETFOpc = &SETFOpcTable[0][0]; break; |
1493 | case CmpInst::FCMP_UNE: SETFOpc = &SETFOpcTable[1][0]; break; |
1494 | } |
1495 | |
1496 | ResultReg = createResultReg(RC: &X86::GR8RegClass); |
1497 | if (SETFOpc) { |
1498 | if (!X86FastEmitCompare(Op0: LHS, Op1: RHS, VT, CurMIMD: I->getDebugLoc())) |
1499 | return false; |
1500 | |
1501 | Register FlagReg1 = createResultReg(RC: &X86::GR8RegClass); |
1502 | Register FlagReg2 = createResultReg(RC: &X86::GR8RegClass); |
1503 | BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, MCID: TII.get(Opcode: X86::SETCCr), |
1504 | DestReg: FlagReg1).addImm(Val: SETFOpc[0]); |
1505 | BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, MCID: TII.get(Opcode: X86::SETCCr), |
1506 | DestReg: FlagReg2).addImm(Val: SETFOpc[1]); |
1507 | BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, MCID: TII.get(Opcode: SETFOpc[2]), |
1508 | DestReg: ResultReg).addReg(RegNo: FlagReg1).addReg(RegNo: FlagReg2); |
1509 | updateValueMap(I, Reg: ResultReg); |
1510 | return true; |
1511 | } |
1512 | |
1513 | X86::CondCode CC; |
1514 | bool SwapArgs; |
1515 | std::tie(args&: CC, args&: SwapArgs) = X86::getX86ConditionCode(Predicate); |
1516 | assert(CC <= X86::LAST_VALID_COND && "Unexpected condition code." ); |
1517 | |
1518 | if (SwapArgs) |
1519 | std::swap(a&: LHS, b&: RHS); |
1520 | |
1521 | // Emit a compare of LHS/RHS. |
1522 | if (!X86FastEmitCompare(Op0: LHS, Op1: RHS, VT, CurMIMD: I->getDebugLoc())) |
1523 | return false; |
1524 | |
1525 | BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, MCID: TII.get(Opcode: X86::SETCCr), |
1526 | DestReg: ResultReg).addImm(Val: CC); |
1527 | updateValueMap(I, Reg: ResultReg); |
1528 | return true; |
1529 | } |
1530 | |
1531 | bool X86FastISel::X86SelectZExt(const Instruction *I) { |
1532 | EVT DstVT = TLI.getValueType(DL, Ty: I->getType()); |
1533 | if (!TLI.isTypeLegal(VT: DstVT)) |
1534 | return false; |
1535 | |
1536 | Register ResultReg = getRegForValue(V: I->getOperand(i: 0)); |
1537 | if (ResultReg == 0) |
1538 | return false; |
1539 | |
1540 | // Handle zero-extension from i1 to i8, which is common. |
1541 | MVT SrcVT = TLI.getSimpleValueType(DL, Ty: I->getOperand(i: 0)->getType()); |
1542 | if (SrcVT == MVT::i1) { |
1543 | // Set the high bits to zero. |
1544 | ResultReg = fastEmitZExtFromI1(VT: MVT::i8, Op0: ResultReg); |
1545 | SrcVT = MVT::i8; |
1546 | |
1547 | if (ResultReg == 0) |
1548 | return false; |
1549 | } |
1550 | |
1551 | if (DstVT == MVT::i64) { |
1552 | // Handle extension to 64-bits via sub-register shenanigans. |
1553 | unsigned MovInst; |
1554 | |
1555 | switch (SrcVT.SimpleTy) { |
1556 | case MVT::i8: MovInst = X86::MOVZX32rr8; break; |
1557 | case MVT::i16: MovInst = X86::MOVZX32rr16; break; |
1558 | case MVT::i32: MovInst = X86::MOV32rr; break; |
1559 | default: llvm_unreachable("Unexpected zext to i64 source type" ); |
1560 | } |
1561 | |
1562 | Register Result32 = createResultReg(RC: &X86::GR32RegClass); |
1563 | BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, MCID: TII.get(Opcode: MovInst), DestReg: Result32) |
1564 | .addReg(RegNo: ResultReg); |
1565 | |
1566 | ResultReg = createResultReg(RC: &X86::GR64RegClass); |
1567 | BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, MCID: TII.get(Opcode: TargetOpcode::SUBREG_TO_REG), |
1568 | DestReg: ResultReg) |
1569 | .addImm(Val: 0).addReg(RegNo: Result32).addImm(Val: X86::sub_32bit); |
1570 | } else if (DstVT == MVT::i16) { |
1571 | // i8->i16 doesn't exist in the autogenerated isel table. Need to zero |
1572 | // extend to 32-bits and then extract down to 16-bits. |
1573 | Register Result32 = createResultReg(RC: &X86::GR32RegClass); |
1574 | BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, MCID: TII.get(Opcode: X86::MOVZX32rr8), |
1575 | DestReg: Result32).addReg(RegNo: ResultReg); |
1576 | |
1577 | ResultReg = fastEmitInst_extractsubreg(RetVT: MVT::i16, Op0: Result32, Idx: X86::sub_16bit); |
1578 | } else if (DstVT != MVT::i8) { |
1579 | ResultReg = fastEmit_r(VT: MVT::i8, RetVT: DstVT.getSimpleVT(), Opcode: ISD::ZERO_EXTEND, |
1580 | Op0: ResultReg); |
1581 | if (ResultReg == 0) |
1582 | return false; |
1583 | } |
1584 | |
1585 | updateValueMap(I, Reg: ResultReg); |
1586 | return true; |
1587 | } |
1588 | |
1589 | bool X86FastISel::X86SelectSExt(const Instruction *I) { |
1590 | EVT DstVT = TLI.getValueType(DL, Ty: I->getType()); |
1591 | if (!TLI.isTypeLegal(VT: DstVT)) |
1592 | return false; |
1593 | |
1594 | Register ResultReg = getRegForValue(V: I->getOperand(i: 0)); |
1595 | if (ResultReg == 0) |
1596 | return false; |
1597 | |
1598 | // Handle sign-extension from i1 to i8. |
1599 | MVT SrcVT = TLI.getSimpleValueType(DL, Ty: I->getOperand(i: 0)->getType()); |
1600 | if (SrcVT == MVT::i1) { |
1601 | // Set the high bits to zero. |
1602 | Register ZExtReg = fastEmitZExtFromI1(VT: MVT::i8, Op0: ResultReg); |
1603 | if (ZExtReg == 0) |
1604 | return false; |
1605 | |
1606 | // Negate the result to make an 8-bit sign extended value. |
1607 | ResultReg = createResultReg(RC: &X86::GR8RegClass); |
1608 | BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, MCID: TII.get(Opcode: X86::NEG8r), |
1609 | DestReg: ResultReg).addReg(RegNo: ZExtReg); |
1610 | |
1611 | SrcVT = MVT::i8; |
1612 | } |
1613 | |
1614 | if (DstVT == MVT::i16) { |
1615 | // i8->i16 doesn't exist in the autogenerated isel table. Need to sign |
1616 | // extend to 32-bits and then extract down to 16-bits. |
1617 | Register Result32 = createResultReg(RC: &X86::GR32RegClass); |
1618 | BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, MCID: TII.get(Opcode: X86::MOVSX32rr8), |
1619 | DestReg: Result32).addReg(RegNo: ResultReg); |
1620 | |
1621 | ResultReg = fastEmitInst_extractsubreg(RetVT: MVT::i16, Op0: Result32, Idx: X86::sub_16bit); |
1622 | } else if (DstVT != MVT::i8) { |
1623 | ResultReg = fastEmit_r(VT: MVT::i8, RetVT: DstVT.getSimpleVT(), Opcode: ISD::SIGN_EXTEND, |
1624 | Op0: ResultReg); |
1625 | if (ResultReg == 0) |
1626 | return false; |
1627 | } |
1628 | |
1629 | updateValueMap(I, Reg: ResultReg); |
1630 | return true; |
1631 | } |
1632 | |
1633 | bool X86FastISel::X86SelectBranch(const Instruction *I) { |
1634 | // Unconditional branches are selected by tablegen-generated code. |
1635 | // Handle a conditional branch. |
1636 | const BranchInst *BI = cast<BranchInst>(Val: I); |
1637 | MachineBasicBlock *TrueMBB = FuncInfo.MBBMap[BI->getSuccessor(i: 0)]; |
1638 | MachineBasicBlock *FalseMBB = FuncInfo.MBBMap[BI->getSuccessor(i: 1)]; |
1639 | |
1640 | // Fold the common case of a conditional branch with a comparison |
1641 | // in the same block (values defined on other blocks may not have |
1642 | // initialized registers). |
1643 | X86::CondCode CC; |
1644 | if (const CmpInst *CI = dyn_cast<CmpInst>(Val: BI->getCondition())) { |
1645 | if (CI->hasOneUse() && CI->getParent() == I->getParent()) { |
1646 | EVT VT = TLI.getValueType(DL, Ty: CI->getOperand(i_nocapture: 0)->getType()); |
1647 | |
1648 | // Try to optimize or fold the cmp. |
1649 | CmpInst::Predicate Predicate = optimizeCmpPredicate(CI); |
1650 | switch (Predicate) { |
1651 | default: break; |
1652 | case CmpInst::FCMP_FALSE: fastEmitBranch(MSucc: FalseMBB, DbgLoc: MIMD.getDL()); return true; |
1653 | case CmpInst::FCMP_TRUE: fastEmitBranch(MSucc: TrueMBB, DbgLoc: MIMD.getDL()); return true; |
1654 | } |
1655 | |
1656 | const Value *CmpLHS = CI->getOperand(i_nocapture: 0); |
1657 | const Value *CmpRHS = CI->getOperand(i_nocapture: 1); |
1658 | |
1659 | // The optimizer might have replaced fcmp oeq %x, %x with fcmp ord %x, |
1660 | // 0.0. |
1661 | // We don't have to materialize a zero constant for this case and can just |
1662 | // use %x again on the RHS. |
1663 | if (Predicate == CmpInst::FCMP_ORD || Predicate == CmpInst::FCMP_UNO) { |
1664 | const auto *CmpRHSC = dyn_cast<ConstantFP>(Val: CmpRHS); |
1665 | if (CmpRHSC && CmpRHSC->isNullValue()) |
1666 | CmpRHS = CmpLHS; |
1667 | } |
1668 | |
1669 | // Try to take advantage of fallthrough opportunities. |
1670 | if (FuncInfo.MBB->isLayoutSuccessor(MBB: TrueMBB)) { |
1671 | std::swap(a&: TrueMBB, b&: FalseMBB); |
1672 | Predicate = CmpInst::getInversePredicate(pred: Predicate); |
1673 | } |
1674 | |
1675 | // FCMP_OEQ and FCMP_UNE cannot be expressed with a single flag/condition |
1676 | // code check. Instead two branch instructions are required to check all |
1677 | // the flags. First we change the predicate to a supported condition code, |
1678 | // which will be the first branch. Later one we will emit the second |
1679 | // branch. |
1680 | bool = false; |
1681 | switch (Predicate) { |
1682 | default: break; |
1683 | case CmpInst::FCMP_OEQ: |
1684 | std::swap(a&: TrueMBB, b&: FalseMBB); |
1685 | [[fallthrough]]; |
1686 | case CmpInst::FCMP_UNE: |
1687 | NeedExtraBranch = true; |
1688 | Predicate = CmpInst::FCMP_ONE; |
1689 | break; |
1690 | } |
1691 | |
1692 | bool SwapArgs; |
1693 | std::tie(args&: CC, args&: SwapArgs) = X86::getX86ConditionCode(Predicate); |
1694 | assert(CC <= X86::LAST_VALID_COND && "Unexpected condition code." ); |
1695 | |
1696 | if (SwapArgs) |
1697 | std::swap(a&: CmpLHS, b&: CmpRHS); |
1698 | |
1699 | // Emit a compare of the LHS and RHS, setting the flags. |
1700 | if (!X86FastEmitCompare(Op0: CmpLHS, Op1: CmpRHS, VT, CurMIMD: CI->getDebugLoc())) |
1701 | return false; |
1702 | |
1703 | BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, MCID: TII.get(Opcode: X86::JCC_1)) |
1704 | .addMBB(MBB: TrueMBB).addImm(Val: CC); |
1705 | |
1706 | // X86 requires a second branch to handle UNE (and OEQ, which is mapped |
1707 | // to UNE above). |
1708 | if (NeedExtraBranch) { |
1709 | BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, MCID: TII.get(Opcode: X86::JCC_1)) |
1710 | .addMBB(MBB: TrueMBB).addImm(Val: X86::COND_P); |
1711 | } |
1712 | |
1713 | finishCondBranch(BranchBB: BI->getParent(), TrueMBB, FalseMBB); |
1714 | return true; |
1715 | } |
1716 | } else if (TruncInst *TI = dyn_cast<TruncInst>(Val: BI->getCondition())) { |
1717 | // Handle things like "%cond = trunc i32 %X to i1 / br i1 %cond", which |
1718 | // typically happen for _Bool and C++ bools. |
1719 | MVT SourceVT; |
1720 | if (TI->hasOneUse() && TI->getParent() == I->getParent() && |
1721 | isTypeLegal(Ty: TI->getOperand(i_nocapture: 0)->getType(), VT&: SourceVT)) { |
1722 | unsigned TestOpc = 0; |
1723 | switch (SourceVT.SimpleTy) { |
1724 | default: break; |
1725 | case MVT::i8: TestOpc = X86::TEST8ri; break; |
1726 | case MVT::i16: TestOpc = X86::TEST16ri; break; |
1727 | case MVT::i32: TestOpc = X86::TEST32ri; break; |
1728 | case MVT::i64: TestOpc = X86::TEST64ri32; break; |
1729 | } |
1730 | if (TestOpc) { |
1731 | Register OpReg = getRegForValue(V: TI->getOperand(i_nocapture: 0)); |
1732 | if (OpReg == 0) return false; |
1733 | |
1734 | BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, MCID: TII.get(Opcode: TestOpc)) |
1735 | .addReg(RegNo: OpReg).addImm(Val: 1); |
1736 | |
1737 | unsigned JmpCond = X86::COND_NE; |
1738 | if (FuncInfo.MBB->isLayoutSuccessor(MBB: TrueMBB)) { |
1739 | std::swap(a&: TrueMBB, b&: FalseMBB); |
1740 | JmpCond = X86::COND_E; |
1741 | } |
1742 | |
1743 | BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, MCID: TII.get(Opcode: X86::JCC_1)) |
1744 | .addMBB(MBB: TrueMBB).addImm(Val: JmpCond); |
1745 | |
1746 | finishCondBranch(BranchBB: BI->getParent(), TrueMBB, FalseMBB); |
1747 | return true; |
1748 | } |
1749 | } |
1750 | } else if (foldX86XALUIntrinsic(CC, I: BI, Cond: BI->getCondition())) { |
1751 | // Fake request the condition, otherwise the intrinsic might be completely |
1752 | // optimized away. |
1753 | Register TmpReg = getRegForValue(V: BI->getCondition()); |
1754 | if (TmpReg == 0) |
1755 | return false; |
1756 | |
1757 | BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, MCID: TII.get(Opcode: X86::JCC_1)) |
1758 | .addMBB(MBB: TrueMBB).addImm(Val: CC); |
1759 | finishCondBranch(BranchBB: BI->getParent(), TrueMBB, FalseMBB); |
1760 | return true; |
1761 | } |
1762 | |
1763 | // Otherwise do a clumsy setcc and re-test it. |
1764 | // Note that i1 essentially gets ANY_EXTEND'ed to i8 where it isn't used |
1765 | // in an explicit cast, so make sure to handle that correctly. |
1766 | Register OpReg = getRegForValue(V: BI->getCondition()); |
1767 | if (OpReg == 0) return false; |
1768 | |
1769 | // In case OpReg is a K register, COPY to a GPR |
1770 | if (MRI.getRegClass(Reg: OpReg) == &X86::VK1RegClass) { |
1771 | unsigned KOpReg = OpReg; |
1772 | OpReg = createResultReg(RC: &X86::GR32RegClass); |
1773 | BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, |
1774 | MCID: TII.get(Opcode: TargetOpcode::COPY), DestReg: OpReg) |
1775 | .addReg(RegNo: KOpReg); |
1776 | OpReg = fastEmitInst_extractsubreg(RetVT: MVT::i8, Op0: OpReg, Idx: X86::sub_8bit); |
1777 | } |
1778 | BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, MCID: TII.get(Opcode: X86::TEST8ri)) |
1779 | .addReg(RegNo: OpReg) |
1780 | .addImm(Val: 1); |
1781 | BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, MCID: TII.get(Opcode: X86::JCC_1)) |
1782 | .addMBB(MBB: TrueMBB).addImm(Val: X86::COND_NE); |
1783 | finishCondBranch(BranchBB: BI->getParent(), TrueMBB, FalseMBB); |
1784 | return true; |
1785 | } |
1786 | |
1787 | bool X86FastISel::X86SelectShift(const Instruction *I) { |
1788 | unsigned CReg = 0, OpReg = 0; |
1789 | const TargetRegisterClass *RC = nullptr; |
1790 | if (I->getType()->isIntegerTy(Bitwidth: 8)) { |
1791 | CReg = X86::CL; |
1792 | RC = &X86::GR8RegClass; |
1793 | switch (I->getOpcode()) { |
1794 | case Instruction::LShr: OpReg = X86::SHR8rCL; break; |
1795 | case Instruction::AShr: OpReg = X86::SAR8rCL; break; |
1796 | case Instruction::Shl: OpReg = X86::SHL8rCL; break; |
1797 | default: return false; |
1798 | } |
1799 | } else if (I->getType()->isIntegerTy(Bitwidth: 16)) { |
1800 | CReg = X86::CX; |
1801 | RC = &X86::GR16RegClass; |
1802 | switch (I->getOpcode()) { |
1803 | default: llvm_unreachable("Unexpected shift opcode" ); |
1804 | case Instruction::LShr: OpReg = X86::SHR16rCL; break; |
1805 | case Instruction::AShr: OpReg = X86::SAR16rCL; break; |
1806 | case Instruction::Shl: OpReg = X86::SHL16rCL; break; |
1807 | } |
1808 | } else if (I->getType()->isIntegerTy(Bitwidth: 32)) { |
1809 | CReg = X86::ECX; |
1810 | RC = &X86::GR32RegClass; |
1811 | switch (I->getOpcode()) { |
1812 | default: llvm_unreachable("Unexpected shift opcode" ); |
1813 | case Instruction::LShr: OpReg = X86::SHR32rCL; break; |
1814 | case Instruction::AShr: OpReg = X86::SAR32rCL; break; |
1815 | case Instruction::Shl: OpReg = X86::SHL32rCL; break; |
1816 | } |
1817 | } else if (I->getType()->isIntegerTy(Bitwidth: 64)) { |
1818 | CReg = X86::RCX; |
1819 | RC = &X86::GR64RegClass; |
1820 | switch (I->getOpcode()) { |
1821 | default: llvm_unreachable("Unexpected shift opcode" ); |
1822 | case Instruction::LShr: OpReg = X86::SHR64rCL; break; |
1823 | case Instruction::AShr: OpReg = X86::SAR64rCL; break; |
1824 | case Instruction::Shl: OpReg = X86::SHL64rCL; break; |
1825 | } |
1826 | } else { |
1827 | return false; |
1828 | } |
1829 | |
1830 | MVT VT; |
1831 | if (!isTypeLegal(Ty: I->getType(), VT)) |
1832 | return false; |
1833 | |
1834 | Register Op0Reg = getRegForValue(V: I->getOperand(i: 0)); |
1835 | if (Op0Reg == 0) return false; |
1836 | |
1837 | Register Op1Reg = getRegForValue(V: I->getOperand(i: 1)); |
1838 | if (Op1Reg == 0) return false; |
1839 | BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, MCID: TII.get(Opcode: TargetOpcode::COPY), |
1840 | DestReg: CReg).addReg(RegNo: Op1Reg); |
1841 | |
1842 | // The shift instruction uses X86::CL. If we defined a super-register |
1843 | // of X86::CL, emit a subreg KILL to precisely describe what we're doing here. |
1844 | if (CReg != X86::CL) |
1845 | BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, |
1846 | MCID: TII.get(Opcode: TargetOpcode::KILL), DestReg: X86::CL) |
1847 | .addReg(RegNo: CReg, flags: RegState::Kill); |
1848 | |
1849 | Register ResultReg = createResultReg(RC); |
1850 | BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, MCID: TII.get(Opcode: OpReg), DestReg: ResultReg) |
1851 | .addReg(RegNo: Op0Reg); |
1852 | updateValueMap(I, Reg: ResultReg); |
1853 | return true; |
1854 | } |
1855 | |
1856 | bool X86FastISel::X86SelectDivRem(const Instruction *I) { |
1857 | const static unsigned NumTypes = 4; // i8, i16, i32, i64 |
1858 | const static unsigned NumOps = 4; // SDiv, SRem, UDiv, URem |
1859 | const static bool S = true; // IsSigned |
1860 | const static bool U = false; // !IsSigned |
1861 | const static unsigned Copy = TargetOpcode::COPY; |
1862 | // For the X86 DIV/IDIV instruction, in most cases the dividend |
1863 | // (numerator) must be in a specific register pair highreg:lowreg, |
1864 | // producing the quotient in lowreg and the remainder in highreg. |
1865 | // For most data types, to set up the instruction, the dividend is |
1866 | // copied into lowreg, and lowreg is sign-extended or zero-extended |
1867 | // into highreg. The exception is i8, where the dividend is defined |
1868 | // as a single register rather than a register pair, and we |
1869 | // therefore directly sign-extend or zero-extend the dividend into |
1870 | // lowreg, instead of copying, and ignore the highreg. |
1871 | const static struct DivRemEntry { |
1872 | // The following portion depends only on the data type. |
1873 | const TargetRegisterClass *RC; |
1874 | unsigned LowInReg; // low part of the register pair |
1875 | unsigned HighInReg; // high part of the register pair |
1876 | // The following portion depends on both the data type and the operation. |
1877 | struct DivRemResult { |
1878 | unsigned OpDivRem; // The specific DIV/IDIV opcode to use. |
1879 | unsigned OpSignExtend; // Opcode for sign-extending lowreg into |
1880 | // highreg, or copying a zero into highreg. |
1881 | unsigned OpCopy; // Opcode for copying dividend into lowreg, or |
1882 | // zero/sign-extending into lowreg for i8. |
1883 | unsigned DivRemResultReg; // Register containing the desired result. |
1884 | bool IsOpSigned; // Whether to use signed or unsigned form. |
1885 | } ResultTable[NumOps]; |
1886 | } OpTable[NumTypes] = { |
1887 | { .RC: &X86::GR8RegClass, .LowInReg: X86::AX, .HighInReg: 0, .ResultTable: { |
1888 | { .OpDivRem: X86::IDIV8r, .OpSignExtend: 0, .OpCopy: X86::MOVSX16rr8, .DivRemResultReg: X86::AL, .IsOpSigned: S }, // SDiv |
1889 | { .OpDivRem: X86::IDIV8r, .OpSignExtend: 0, .OpCopy: X86::MOVSX16rr8, .DivRemResultReg: X86::AH, .IsOpSigned: S }, // SRem |
1890 | { .OpDivRem: X86::DIV8r, .OpSignExtend: 0, .OpCopy: X86::MOVZX16rr8, .DivRemResultReg: X86::AL, .IsOpSigned: U }, // UDiv |
1891 | { .OpDivRem: X86::DIV8r, .OpSignExtend: 0, .OpCopy: X86::MOVZX16rr8, .DivRemResultReg: X86::AH, .IsOpSigned: U }, // URem |
1892 | } |
1893 | }, // i8 |
1894 | { .RC: &X86::GR16RegClass, .LowInReg: X86::AX, .HighInReg: X86::DX, .ResultTable: { |
1895 | { .OpDivRem: X86::IDIV16r, .OpSignExtend: X86::CWD, .OpCopy: Copy, .DivRemResultReg: X86::AX, .IsOpSigned: S }, // SDiv |
1896 | { .OpDivRem: X86::IDIV16r, .OpSignExtend: X86::CWD, .OpCopy: Copy, .DivRemResultReg: X86::DX, .IsOpSigned: S }, // SRem |
1897 | { .OpDivRem: X86::DIV16r, .OpSignExtend: X86::MOV32r0, .OpCopy: Copy, .DivRemResultReg: X86::AX, .IsOpSigned: U }, // UDiv |
1898 | { .OpDivRem: X86::DIV16r, .OpSignExtend: X86::MOV32r0, .OpCopy: Copy, .DivRemResultReg: X86::DX, .IsOpSigned: U }, // URem |
1899 | } |
1900 | }, // i16 |
1901 | { .RC: &X86::GR32RegClass, .LowInReg: X86::EAX, .HighInReg: X86::EDX, .ResultTable: { |
1902 | { .OpDivRem: X86::IDIV32r, .OpSignExtend: X86::CDQ, .OpCopy: Copy, .DivRemResultReg: X86::EAX, .IsOpSigned: S }, // SDiv |
1903 | { .OpDivRem: X86::IDIV32r, .OpSignExtend: X86::CDQ, .OpCopy: Copy, .DivRemResultReg: X86::EDX, .IsOpSigned: S }, // SRem |
1904 | { .OpDivRem: X86::DIV32r, .OpSignExtend: X86::MOV32r0, .OpCopy: Copy, .DivRemResultReg: X86::EAX, .IsOpSigned: U }, // UDiv |
1905 | { .OpDivRem: X86::DIV32r, .OpSignExtend: X86::MOV32r0, .OpCopy: Copy, .DivRemResultReg: X86::EDX, .IsOpSigned: U }, // URem |
1906 | } |
1907 | }, // i32 |
1908 | { .RC: &X86::GR64RegClass, .LowInReg: X86::RAX, .HighInReg: X86::RDX, .ResultTable: { |
1909 | { .OpDivRem: X86::IDIV64r, .OpSignExtend: X86::CQO, .OpCopy: Copy, .DivRemResultReg: X86::RAX, .IsOpSigned: S }, // SDiv |
1910 | { .OpDivRem: X86::IDIV64r, .OpSignExtend: X86::CQO, .OpCopy: Copy, .DivRemResultReg: X86::RDX, .IsOpSigned: S }, // SRem |
1911 | { .OpDivRem: X86::DIV64r, .OpSignExtend: X86::MOV32r0, .OpCopy: Copy, .DivRemResultReg: X86::RAX, .IsOpSigned: U }, // UDiv |
1912 | { .OpDivRem: X86::DIV64r, .OpSignExtend: X86::MOV32r0, .OpCopy: Copy, .DivRemResultReg: X86::RDX, .IsOpSigned: U }, // URem |
1913 | } |
1914 | }, // i64 |
1915 | }; |
1916 | |
1917 | MVT VT; |
1918 | if (!isTypeLegal(Ty: I->getType(), VT)) |
1919 | return false; |
1920 | |
1921 | unsigned TypeIndex, OpIndex; |
1922 | switch (VT.SimpleTy) { |
1923 | default: return false; |
1924 | case MVT::i8: TypeIndex = 0; break; |
1925 | case MVT::i16: TypeIndex = 1; break; |
1926 | case MVT::i32: TypeIndex = 2; break; |
1927 | case MVT::i64: TypeIndex = 3; |
1928 | if (!Subtarget->is64Bit()) |
1929 | return false; |
1930 | break; |
1931 | } |
1932 | |
1933 | switch (I->getOpcode()) { |
1934 | default: llvm_unreachable("Unexpected div/rem opcode" ); |
1935 | case Instruction::SDiv: OpIndex = 0; break; |
1936 | case Instruction::SRem: OpIndex = 1; break; |
1937 | case Instruction::UDiv: OpIndex = 2; break; |
1938 | case Instruction::URem: OpIndex = 3; break; |
1939 | } |
1940 | |
1941 | const DivRemEntry &TypeEntry = OpTable[TypeIndex]; |
1942 | const DivRemEntry::DivRemResult &OpEntry = TypeEntry.ResultTable[OpIndex]; |
1943 | Register Op0Reg = getRegForValue(V: I->getOperand(i: 0)); |
1944 | if (Op0Reg == 0) |
1945 | return false; |
1946 | Register Op1Reg = getRegForValue(V: I->getOperand(i: 1)); |
1947 | if (Op1Reg == 0) |
1948 | return false; |
1949 | |
1950 | // Move op0 into low-order input register. |
1951 | BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, |
1952 | MCID: TII.get(Opcode: OpEntry.OpCopy), DestReg: TypeEntry.LowInReg).addReg(RegNo: Op0Reg); |
1953 | // Zero-extend or sign-extend into high-order input register. |
1954 | if (OpEntry.OpSignExtend) { |
1955 | if (OpEntry.IsOpSigned) |
1956 | BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, |
1957 | MCID: TII.get(Opcode: OpEntry.OpSignExtend)); |
1958 | else { |
1959 | Register Zero32 = createResultReg(RC: &X86::GR32RegClass); |
1960 | BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, |
1961 | MCID: TII.get(Opcode: X86::MOV32r0), DestReg: Zero32); |
1962 | |
1963 | // Copy the zero into the appropriate sub/super/identical physical |
1964 | // register. Unfortunately the operations needed are not uniform enough |
1965 | // to fit neatly into the table above. |
1966 | if (VT == MVT::i16) { |
1967 | BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, |
1968 | MCID: TII.get(Opcode: Copy), DestReg: TypeEntry.HighInReg) |
1969 | .addReg(RegNo: Zero32, flags: 0, SubReg: X86::sub_16bit); |
1970 | } else if (VT == MVT::i32) { |
1971 | BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, |
1972 | MCID: TII.get(Opcode: Copy), DestReg: TypeEntry.HighInReg) |
1973 | .addReg(RegNo: Zero32); |
1974 | } else if (VT == MVT::i64) { |
1975 | BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, |
1976 | MCID: TII.get(Opcode: TargetOpcode::SUBREG_TO_REG), DestReg: TypeEntry.HighInReg) |
1977 | .addImm(Val: 0).addReg(RegNo: Zero32).addImm(Val: X86::sub_32bit); |
1978 | } |
1979 | } |
1980 | } |
1981 | // Generate the DIV/IDIV instruction. |
1982 | BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, |
1983 | MCID: TII.get(Opcode: OpEntry.OpDivRem)).addReg(RegNo: Op1Reg); |
1984 | // For i8 remainder, we can't reference ah directly, as we'll end |
1985 | // up with bogus copies like %r9b = COPY %ah. Reference ax |
1986 | // instead to prevent ah references in a rex instruction. |
1987 | // |
1988 | // The current assumption of the fast register allocator is that isel |
1989 | // won't generate explicit references to the GR8_NOREX registers. If |
1990 | // the allocator and/or the backend get enhanced to be more robust in |
1991 | // that regard, this can be, and should be, removed. |
1992 | unsigned ResultReg = 0; |
1993 | if ((I->getOpcode() == Instruction::SRem || |
1994 | I->getOpcode() == Instruction::URem) && |
1995 | OpEntry.DivRemResultReg == X86::AH && Subtarget->is64Bit()) { |
1996 | Register SourceSuperReg = createResultReg(RC: &X86::GR16RegClass); |
1997 | Register ResultSuperReg = createResultReg(RC: &X86::GR16RegClass); |
1998 | BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, |
1999 | MCID: TII.get(Opcode: Copy), DestReg: SourceSuperReg).addReg(RegNo: X86::AX); |
2000 | |
2001 | // Shift AX right by 8 bits instead of using AH. |
2002 | BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, MCID: TII.get(Opcode: X86::SHR16ri), |
2003 | DestReg: ResultSuperReg).addReg(RegNo: SourceSuperReg).addImm(Val: 8); |
2004 | |
2005 | // Now reference the 8-bit subreg of the result. |
2006 | ResultReg = fastEmitInst_extractsubreg(RetVT: MVT::i8, Op0: ResultSuperReg, |
2007 | Idx: X86::sub_8bit); |
2008 | } |
2009 | // Copy the result out of the physreg if we haven't already. |
2010 | if (!ResultReg) { |
2011 | ResultReg = createResultReg(RC: TypeEntry.RC); |
2012 | BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, MCID: TII.get(Opcode: Copy), DestReg: ResultReg) |
2013 | .addReg(RegNo: OpEntry.DivRemResultReg); |
2014 | } |
2015 | updateValueMap(I, Reg: ResultReg); |
2016 | |
2017 | return true; |
2018 | } |
2019 | |
2020 | /// Emit a conditional move instruction (if the are supported) to lower |
2021 | /// the select. |
2022 | bool X86FastISel::X86FastEmitCMoveSelect(MVT RetVT, const Instruction *I) { |
2023 | // Check if the subtarget supports these instructions. |
2024 | if (!Subtarget->canUseCMOV()) |
2025 | return false; |
2026 | |
2027 | // FIXME: Add support for i8. |
2028 | if (RetVT < MVT::i16 || RetVT > MVT::i64) |
2029 | return false; |
2030 | |
2031 | const Value *Cond = I->getOperand(i: 0); |
2032 | const TargetRegisterClass *RC = TLI.getRegClassFor(VT: RetVT); |
2033 | bool NeedTest = true; |
2034 | X86::CondCode CC = X86::COND_NE; |
2035 | |
2036 | // Optimize conditions coming from a compare if both instructions are in the |
2037 | // same basic block (values defined in other basic blocks may not have |
2038 | // initialized registers). |
2039 | const auto *CI = dyn_cast<CmpInst>(Val: Cond); |
2040 | if (CI && (CI->getParent() == I->getParent())) { |
2041 | CmpInst::Predicate Predicate = optimizeCmpPredicate(CI); |
2042 | |
2043 | // FCMP_OEQ and FCMP_UNE cannot be checked with a single instruction. |
2044 | static const uint16_t SETFOpcTable[2][3] = { |
2045 | { X86::COND_NP, X86::COND_E, X86::TEST8rr }, |
2046 | { X86::COND_P, X86::COND_NE, X86::OR8rr } |
2047 | }; |
2048 | const uint16_t *SETFOpc = nullptr; |
2049 | switch (Predicate) { |
2050 | default: break; |
2051 | case CmpInst::FCMP_OEQ: |
2052 | SETFOpc = &SETFOpcTable[0][0]; |
2053 | Predicate = CmpInst::ICMP_NE; |
2054 | break; |
2055 | case CmpInst::FCMP_UNE: |
2056 | SETFOpc = &SETFOpcTable[1][0]; |
2057 | Predicate = CmpInst::ICMP_NE; |
2058 | break; |
2059 | } |
2060 | |
2061 | bool NeedSwap; |
2062 | std::tie(args&: CC, args&: NeedSwap) = X86::getX86ConditionCode(Predicate); |
2063 | assert(CC <= X86::LAST_VALID_COND && "Unexpected condition code." ); |
2064 | |
2065 | const Value *CmpLHS = CI->getOperand(i_nocapture: 0); |
2066 | const Value *CmpRHS = CI->getOperand(i_nocapture: 1); |
2067 | if (NeedSwap) |
2068 | std::swap(a&: CmpLHS, b&: CmpRHS); |
2069 | |
2070 | EVT CmpVT = TLI.getValueType(DL, Ty: CmpLHS->getType()); |
2071 | // Emit a compare of the LHS and RHS, setting the flags. |
2072 | if (!X86FastEmitCompare(Op0: CmpLHS, Op1: CmpRHS, VT: CmpVT, CurMIMD: CI->getDebugLoc())) |
2073 | return false; |
2074 | |
2075 | if (SETFOpc) { |
2076 | Register FlagReg1 = createResultReg(RC: &X86::GR8RegClass); |
2077 | Register FlagReg2 = createResultReg(RC: &X86::GR8RegClass); |
2078 | BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, MCID: TII.get(Opcode: X86::SETCCr), |
2079 | DestReg: FlagReg1).addImm(Val: SETFOpc[0]); |
2080 | BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, MCID: TII.get(Opcode: X86::SETCCr), |
2081 | DestReg: FlagReg2).addImm(Val: SETFOpc[1]); |
2082 | auto const &II = TII.get(Opcode: SETFOpc[2]); |
2083 | if (II.getNumDefs()) { |
2084 | Register TmpReg = createResultReg(RC: &X86::GR8RegClass); |
2085 | BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, MCID: II, DestReg: TmpReg) |
2086 | .addReg(RegNo: FlagReg2).addReg(RegNo: FlagReg1); |
2087 | } else { |
2088 | BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, MCID: II) |
2089 | .addReg(RegNo: FlagReg2).addReg(RegNo: FlagReg1); |
2090 | } |
2091 | } |
2092 | NeedTest = false; |
2093 | } else if (foldX86XALUIntrinsic(CC, I, Cond)) { |
2094 | // Fake request the condition, otherwise the intrinsic might be completely |
2095 | // optimized away. |
2096 | Register TmpReg = getRegForValue(V: Cond); |
2097 | if (TmpReg == 0) |
2098 | return false; |
2099 | |
2100 | NeedTest = false; |
2101 | } |
2102 | |
2103 | if (NeedTest) { |
2104 | // Selects operate on i1, however, CondReg is 8 bits width and may contain |
2105 | // garbage. Indeed, only the less significant bit is supposed to be |
2106 | // accurate. If we read more than the lsb, we may see non-zero values |
2107 | // whereas lsb is zero. Therefore, we have to truncate Op0Reg to i1 for |
2108 | // the select. This is achieved by performing TEST against 1. |
2109 | Register CondReg = getRegForValue(V: Cond); |
2110 | if (CondReg == 0) |
2111 | return false; |
2112 | |
2113 | // In case OpReg is a K register, COPY to a GPR |
2114 | if (MRI.getRegClass(Reg: CondReg) == &X86::VK1RegClass) { |
2115 | unsigned KCondReg = CondReg; |
2116 | CondReg = createResultReg(RC: &X86::GR32RegClass); |
2117 | BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, |
2118 | MCID: TII.get(Opcode: TargetOpcode::COPY), DestReg: CondReg) |
2119 | .addReg(RegNo: KCondReg); |
2120 | CondReg = fastEmitInst_extractsubreg(RetVT: MVT::i8, Op0: CondReg, Idx: X86::sub_8bit); |
2121 | } |
2122 | BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, MCID: TII.get(Opcode: X86::TEST8ri)) |
2123 | .addReg(RegNo: CondReg) |
2124 | .addImm(Val: 1); |
2125 | } |
2126 | |
2127 | const Value *LHS = I->getOperand(i: 1); |
2128 | const Value *RHS = I->getOperand(i: 2); |
2129 | |
2130 | Register RHSReg = getRegForValue(V: RHS); |
2131 | Register LHSReg = getRegForValue(V: LHS); |
2132 | if (!LHSReg || !RHSReg) |
2133 | return false; |
2134 | |
2135 | const TargetRegisterInfo &TRI = *Subtarget->getRegisterInfo(); |
2136 | unsigned Opc = X86::getCMovOpcode(RegBytes: TRI.getRegSizeInBits(RC: *RC) / 8, HasMemoryOperand: false, |
2137 | HasNDD: Subtarget->hasNDD()); |
2138 | Register ResultReg = fastEmitInst_rri(MachineInstOpcode: Opc, RC, Op0: RHSReg, Op1: LHSReg, Imm: CC); |
2139 | updateValueMap(I, Reg: ResultReg); |
2140 | return true; |
2141 | } |
2142 | |
2143 | /// Emit SSE or AVX instructions to lower the select. |
2144 | /// |
2145 | /// Try to use SSE1/SSE2 instructions to simulate a select without branches. |
2146 | /// This lowers fp selects into a CMP/AND/ANDN/OR sequence when the necessary |
2147 | /// SSE instructions are available. If AVX is available, try to use a VBLENDV. |
2148 | bool X86FastISel::X86FastEmitSSESelect(MVT RetVT, const Instruction *I) { |
2149 | // Optimize conditions coming from a compare if both instructions are in the |
2150 | // same basic block (values defined in other basic blocks may not have |
2151 | // initialized registers). |
2152 | const auto *CI = dyn_cast<FCmpInst>(Val: I->getOperand(i: 0)); |
2153 | if (!CI || (CI->getParent() != I->getParent())) |
2154 | return false; |
2155 | |
2156 | if (I->getType() != CI->getOperand(i_nocapture: 0)->getType() || |
2157 | !((Subtarget->hasSSE1() && RetVT == MVT::f32) || |
2158 | (Subtarget->hasSSE2() && RetVT == MVT::f64))) |
2159 | return false; |
2160 | |
2161 | const Value *CmpLHS = CI->getOperand(i_nocapture: 0); |
2162 | const Value *CmpRHS = CI->getOperand(i_nocapture: 1); |
2163 | CmpInst::Predicate Predicate = optimizeCmpPredicate(CI); |
2164 | |
2165 | // The optimizer might have replaced fcmp oeq %x, %x with fcmp ord %x, 0.0. |
2166 | // We don't have to materialize a zero constant for this case and can just use |
2167 | // %x again on the RHS. |
2168 | if (Predicate == CmpInst::FCMP_ORD || Predicate == CmpInst::FCMP_UNO) { |
2169 | const auto *CmpRHSC = dyn_cast<ConstantFP>(Val: CmpRHS); |
2170 | if (CmpRHSC && CmpRHSC->isNullValue()) |
2171 | CmpRHS = CmpLHS; |
2172 | } |
2173 | |
2174 | unsigned CC; |
2175 | bool NeedSwap; |
2176 | std::tie(args&: CC, args&: NeedSwap) = getX86SSEConditionCode(Predicate); |
2177 | if (CC > 7 && !Subtarget->hasAVX()) |
2178 | return false; |
2179 | |
2180 | if (NeedSwap) |
2181 | std::swap(a&: CmpLHS, b&: CmpRHS); |
2182 | |
2183 | const Value *LHS = I->getOperand(i: 1); |
2184 | const Value *RHS = I->getOperand(i: 2); |
2185 | |
2186 | Register LHSReg = getRegForValue(V: LHS); |
2187 | Register RHSReg = getRegForValue(V: RHS); |
2188 | Register CmpLHSReg = getRegForValue(V: CmpLHS); |
2189 | Register CmpRHSReg = getRegForValue(V: CmpRHS); |
2190 | if (!LHSReg || !RHSReg || !CmpLHSReg || !CmpRHSReg) |
2191 | return false; |
2192 | |
2193 | const TargetRegisterClass *RC = TLI.getRegClassFor(VT: RetVT); |
2194 | unsigned ResultReg; |
2195 | |
2196 | if (Subtarget->hasAVX512()) { |
2197 | // If we have AVX512 we can use a mask compare and masked movss/sd. |
2198 | const TargetRegisterClass *VR128X = &X86::VR128XRegClass; |
2199 | const TargetRegisterClass *VK1 = &X86::VK1RegClass; |
2200 | |
2201 | unsigned CmpOpcode = |
2202 | (RetVT == MVT::f32) ? X86::VCMPSSZrri : X86::VCMPSDZrri; |
2203 | Register CmpReg = fastEmitInst_rri(MachineInstOpcode: CmpOpcode, RC: VK1, Op0: CmpLHSReg, Op1: CmpRHSReg, |
2204 | Imm: CC); |
2205 | |
2206 | // Need an IMPLICIT_DEF for the input that is used to generate the upper |
2207 | // bits of the result register since its not based on any of the inputs. |
2208 | Register ImplicitDefReg = createResultReg(RC: VR128X); |
2209 | BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, |
2210 | MCID: TII.get(Opcode: TargetOpcode::IMPLICIT_DEF), DestReg: ImplicitDefReg); |
2211 | |
2212 | // Place RHSReg is the passthru of the masked movss/sd operation and put |
2213 | // LHS in the input. The mask input comes from the compare. |
2214 | unsigned MovOpcode = |
2215 | (RetVT == MVT::f32) ? X86::VMOVSSZrrk : X86::VMOVSDZrrk; |
2216 | unsigned MovReg = fastEmitInst_rrrr(MachineInstOpcode: MovOpcode, RC: VR128X, Op0: RHSReg, Op1: CmpReg, |
2217 | Op2: ImplicitDefReg, Op3: LHSReg); |
2218 | |
2219 | ResultReg = createResultReg(RC); |
2220 | BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, |
2221 | MCID: TII.get(Opcode: TargetOpcode::COPY), DestReg: ResultReg).addReg(RegNo: MovReg); |
2222 | |
2223 | } else if (Subtarget->hasAVX()) { |
2224 | const TargetRegisterClass *VR128 = &X86::VR128RegClass; |
2225 | |
2226 | // If we have AVX, create 1 blendv instead of 3 logic instructions. |
2227 | // Blendv was introduced with SSE 4.1, but the 2 register form implicitly |
2228 | // uses XMM0 as the selection register. That may need just as many |
2229 | // instructions as the AND/ANDN/OR sequence due to register moves, so |
2230 | // don't bother. |
2231 | unsigned CmpOpcode = |
2232 | (RetVT == MVT::f32) ? X86::VCMPSSrri : X86::VCMPSDrri; |
2233 | unsigned BlendOpcode = |
2234 | (RetVT == MVT::f32) ? X86::VBLENDVPSrrr : X86::VBLENDVPDrrr; |
2235 | |
2236 | Register CmpReg = fastEmitInst_rri(MachineInstOpcode: CmpOpcode, RC, Op0: CmpLHSReg, Op1: CmpRHSReg, |
2237 | Imm: CC); |
2238 | Register VBlendReg = fastEmitInst_rrr(MachineInstOpcode: BlendOpcode, RC: VR128, Op0: RHSReg, Op1: LHSReg, |
2239 | Op2: CmpReg); |
2240 | ResultReg = createResultReg(RC); |
2241 | BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, |
2242 | MCID: TII.get(Opcode: TargetOpcode::COPY), DestReg: ResultReg).addReg(RegNo: VBlendReg); |
2243 | } else { |
2244 | // Choose the SSE instruction sequence based on data type (float or double). |
2245 | static const uint16_t OpcTable[2][4] = { |
2246 | { X86::CMPSSrri, X86::ANDPSrr, X86::ANDNPSrr, X86::ORPSrr }, |
2247 | { X86::CMPSDrri, X86::ANDPDrr, X86::ANDNPDrr, X86::ORPDrr } |
2248 | }; |
2249 | |
2250 | const uint16_t *Opc = nullptr; |
2251 | switch (RetVT.SimpleTy) { |
2252 | default: return false; |
2253 | case MVT::f32: Opc = &OpcTable[0][0]; break; |
2254 | case MVT::f64: Opc = &OpcTable[1][0]; break; |
2255 | } |
2256 | |
2257 | const TargetRegisterClass *VR128 = &X86::VR128RegClass; |
2258 | Register CmpReg = fastEmitInst_rri(MachineInstOpcode: Opc[0], RC, Op0: CmpLHSReg, Op1: CmpRHSReg, Imm: CC); |
2259 | Register AndReg = fastEmitInst_rr(MachineInstOpcode: Opc[1], RC: VR128, Op0: CmpReg, Op1: LHSReg); |
2260 | Register AndNReg = fastEmitInst_rr(MachineInstOpcode: Opc[2], RC: VR128, Op0: CmpReg, Op1: RHSReg); |
2261 | Register OrReg = fastEmitInst_rr(MachineInstOpcode: Opc[3], RC: VR128, Op0: AndNReg, Op1: AndReg); |
2262 | ResultReg = createResultReg(RC); |
2263 | BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, |
2264 | MCID: TII.get(Opcode: TargetOpcode::COPY), DestReg: ResultReg).addReg(RegNo: OrReg); |
2265 | } |
2266 | updateValueMap(I, Reg: ResultReg); |
2267 | return true; |
2268 | } |
2269 | |
2270 | bool X86FastISel::X86FastEmitPseudoSelect(MVT RetVT, const Instruction *I) { |
2271 | // These are pseudo CMOV instructions and will be later expanded into control- |
2272 | // flow. |
2273 | unsigned Opc; |
2274 | switch (RetVT.SimpleTy) { |
2275 | default: return false; |
2276 | case MVT::i8: Opc = X86::CMOV_GR8; break; |
2277 | case MVT::i16: Opc = X86::CMOV_GR16; break; |
2278 | case MVT::i32: Opc = X86::CMOV_GR32; break; |
2279 | case MVT::f16: |
2280 | Opc = Subtarget->hasAVX512() ? X86::CMOV_FR16X : X86::CMOV_FR16; break; |
2281 | case MVT::f32: |
2282 | Opc = Subtarget->hasAVX512() ? X86::CMOV_FR32X : X86::CMOV_FR32; break; |
2283 | case MVT::f64: |
2284 | Opc = Subtarget->hasAVX512() ? X86::CMOV_FR64X : X86::CMOV_FR64; break; |
2285 | } |
2286 | |
2287 | const Value *Cond = I->getOperand(i: 0); |
2288 | X86::CondCode CC = X86::COND_NE; |
2289 | |
2290 | // Optimize conditions coming from a compare if both instructions are in the |
2291 | // same basic block (values defined in other basic blocks may not have |
2292 | // initialized registers). |
2293 | const auto *CI = dyn_cast<CmpInst>(Val: Cond); |
2294 | if (CI && (CI->getParent() == I->getParent())) { |
2295 | bool NeedSwap; |
2296 | std::tie(args&: CC, args&: NeedSwap) = X86::getX86ConditionCode(Predicate: CI->getPredicate()); |
2297 | if (CC > X86::LAST_VALID_COND) |
2298 | return false; |
2299 | |
2300 | const Value *CmpLHS = CI->getOperand(i_nocapture: 0); |
2301 | const Value *CmpRHS = CI->getOperand(i_nocapture: 1); |
2302 | |
2303 | if (NeedSwap) |
2304 | std::swap(a&: CmpLHS, b&: CmpRHS); |
2305 | |
2306 | EVT CmpVT = TLI.getValueType(DL, Ty: CmpLHS->getType()); |
2307 | if (!X86FastEmitCompare(Op0: CmpLHS, Op1: CmpRHS, VT: CmpVT, CurMIMD: CI->getDebugLoc())) |
2308 | return false; |
2309 | } else { |
2310 | Register CondReg = getRegForValue(V: Cond); |
2311 | if (CondReg == 0) |
2312 | return false; |
2313 | |
2314 | // In case OpReg is a K register, COPY to a GPR |
2315 | if (MRI.getRegClass(Reg: CondReg) == &X86::VK1RegClass) { |
2316 | unsigned KCondReg = CondReg; |
2317 | CondReg = createResultReg(RC: &X86::GR32RegClass); |
2318 | BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, |
2319 | MCID: TII.get(Opcode: TargetOpcode::COPY), DestReg: CondReg) |
2320 | .addReg(RegNo: KCondReg); |
2321 | CondReg = fastEmitInst_extractsubreg(RetVT: MVT::i8, Op0: CondReg, Idx: X86::sub_8bit); |
2322 | } |
2323 | BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, MCID: TII.get(Opcode: X86::TEST8ri)) |
2324 | .addReg(RegNo: CondReg) |
2325 | .addImm(Val: 1); |
2326 | } |
2327 | |
2328 | const Value *LHS = I->getOperand(i: 1); |
2329 | const Value *RHS = I->getOperand(i: 2); |
2330 | |
2331 | Register LHSReg = getRegForValue(V: LHS); |
2332 | Register RHSReg = getRegForValue(V: RHS); |
2333 | if (!LHSReg || !RHSReg) |
2334 | return false; |
2335 | |
2336 | const TargetRegisterClass *RC = TLI.getRegClassFor(VT: RetVT); |
2337 | |
2338 | Register ResultReg = |
2339 | fastEmitInst_rri(MachineInstOpcode: Opc, RC, Op0: RHSReg, Op1: LHSReg, Imm: CC); |
2340 | updateValueMap(I, Reg: ResultReg); |
2341 | return true; |
2342 | } |
2343 | |
2344 | bool X86FastISel::X86SelectSelect(const Instruction *I) { |
2345 | MVT RetVT; |
2346 | if (!isTypeLegal(Ty: I->getType(), VT&: RetVT)) |
2347 | return false; |
2348 | |
2349 | // Check if we can fold the select. |
2350 | if (const auto *CI = dyn_cast<CmpInst>(Val: I->getOperand(i: 0))) { |
2351 | CmpInst::Predicate Predicate = optimizeCmpPredicate(CI); |
2352 | const Value *Opnd = nullptr; |
2353 | switch (Predicate) { |
2354 | default: break; |
2355 | case CmpInst::FCMP_FALSE: Opnd = I->getOperand(i: 2); break; |
2356 | case CmpInst::FCMP_TRUE: Opnd = I->getOperand(i: 1); break; |
2357 | } |
2358 | // No need for a select anymore - this is an unconditional move. |
2359 | if (Opnd) { |
2360 | Register OpReg = getRegForValue(V: Opnd); |
2361 | if (OpReg == 0) |
2362 | return false; |
2363 | const TargetRegisterClass *RC = TLI.getRegClassFor(VT: RetVT); |
2364 | Register ResultReg = createResultReg(RC); |
2365 | BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, |
2366 | MCID: TII.get(Opcode: TargetOpcode::COPY), DestReg: ResultReg) |
2367 | .addReg(RegNo: OpReg); |
2368 | updateValueMap(I, Reg: ResultReg); |
2369 | return true; |
2370 | } |
2371 | } |
2372 | |
2373 | // First try to use real conditional move instructions. |
2374 | if (X86FastEmitCMoveSelect(RetVT, I)) |
2375 | return true; |
2376 | |
2377 | // Try to use a sequence of SSE instructions to simulate a conditional move. |
2378 | if (X86FastEmitSSESelect(RetVT, I)) |
2379 | return true; |
2380 | |
2381 | // Fall-back to pseudo conditional move instructions, which will be later |
2382 | // converted to control-flow. |
2383 | if (X86FastEmitPseudoSelect(RetVT, I)) |
2384 | return true; |
2385 | |
2386 | return false; |
2387 | } |
2388 | |
2389 | // Common code for X86SelectSIToFP and X86SelectUIToFP. |
2390 | bool X86FastISel::X86SelectIntToFP(const Instruction *I, bool IsSigned) { |
2391 | // The target-independent selection algorithm in FastISel already knows how |
2392 | // to select a SINT_TO_FP if the target is SSE but not AVX. |
2393 | // Early exit if the subtarget doesn't have AVX. |
2394 | // Unsigned conversion requires avx512. |
2395 | bool HasAVX512 = Subtarget->hasAVX512(); |
2396 | if (!Subtarget->hasAVX() || (!IsSigned && !HasAVX512)) |
2397 | return false; |
2398 | |
2399 | // TODO: We could sign extend narrower types. |
2400 | EVT SrcVT = TLI.getValueType(DL, Ty: I->getOperand(i: 0)->getType()); |
2401 | if (SrcVT != MVT::i32 && SrcVT != MVT::i64) |
2402 | return false; |
2403 | |
2404 | // Select integer to float/double conversion. |
2405 | Register OpReg = getRegForValue(V: I->getOperand(i: 0)); |
2406 | if (OpReg == 0) |
2407 | return false; |
2408 | |
2409 | unsigned Opcode; |
2410 | |
2411 | static const uint16_t SCvtOpc[2][2][2] = { |
2412 | { { X86::VCVTSI2SSrr, X86::VCVTSI642SSrr }, |
2413 | { X86::VCVTSI2SDrr, X86::VCVTSI642SDrr } }, |
2414 | { { X86::VCVTSI2SSZrr, X86::VCVTSI642SSZrr }, |
2415 | { X86::VCVTSI2SDZrr, X86::VCVTSI642SDZrr } }, |
2416 | }; |
2417 | static const uint16_t UCvtOpc[2][2] = { |
2418 | { X86::VCVTUSI2SSZrr, X86::VCVTUSI642SSZrr }, |
2419 | { X86::VCVTUSI2SDZrr, X86::VCVTUSI642SDZrr }, |
2420 | }; |
2421 | bool Is64Bit = SrcVT == MVT::i64; |
2422 | |
2423 | if (I->getType()->isDoubleTy()) { |
2424 | // s/uitofp int -> double |
2425 | Opcode = IsSigned ? SCvtOpc[HasAVX512][1][Is64Bit] : UCvtOpc[1][Is64Bit]; |
2426 | } else if (I->getType()->isFloatTy()) { |
2427 | // s/uitofp int -> float |
2428 | Opcode = IsSigned ? SCvtOpc[HasAVX512][0][Is64Bit] : UCvtOpc[0][Is64Bit]; |
2429 | } else |
2430 | return false; |
2431 | |
2432 | MVT DstVT = TLI.getValueType(DL, Ty: I->getType()).getSimpleVT(); |
2433 | const TargetRegisterClass *RC = TLI.getRegClassFor(VT: DstVT); |
2434 | Register ImplicitDefReg = createResultReg(RC); |
2435 | BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, |
2436 | MCID: TII.get(Opcode: TargetOpcode::IMPLICIT_DEF), DestReg: ImplicitDefReg); |
2437 | Register ResultReg = fastEmitInst_rr(MachineInstOpcode: Opcode, RC, Op0: ImplicitDefReg, Op1: OpReg); |
2438 | updateValueMap(I, Reg: ResultReg); |
2439 | return true; |
2440 | } |
2441 | |
2442 | bool X86FastISel::X86SelectSIToFP(const Instruction *I) { |
2443 | return X86SelectIntToFP(I, /*IsSigned*/true); |
2444 | } |
2445 | |
2446 | bool X86FastISel::X86SelectUIToFP(const Instruction *I) { |
2447 | return X86SelectIntToFP(I, /*IsSigned*/false); |
2448 | } |
2449 | |
2450 | // Helper method used by X86SelectFPExt and X86SelectFPTrunc. |
2451 | bool X86FastISel::X86SelectFPExtOrFPTrunc(const Instruction *I, |
2452 | unsigned TargetOpc, |
2453 | const TargetRegisterClass *RC) { |
2454 | assert((I->getOpcode() == Instruction::FPExt || |
2455 | I->getOpcode() == Instruction::FPTrunc) && |
2456 | "Instruction must be an FPExt or FPTrunc!" ); |
2457 | bool HasAVX = Subtarget->hasAVX(); |
2458 | |
2459 | Register OpReg = getRegForValue(V: I->getOperand(i: 0)); |
2460 | if (OpReg == 0) |
2461 | return false; |
2462 | |
2463 | unsigned ImplicitDefReg; |
2464 | if (HasAVX) { |
2465 | ImplicitDefReg = createResultReg(RC); |
2466 | BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, |
2467 | MCID: TII.get(Opcode: TargetOpcode::IMPLICIT_DEF), DestReg: ImplicitDefReg); |
2468 | |
2469 | } |
2470 | |
2471 | Register ResultReg = createResultReg(RC); |
2472 | MachineInstrBuilder MIB; |
2473 | MIB = BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, MCID: TII.get(Opcode: TargetOpc), |
2474 | DestReg: ResultReg); |
2475 | |
2476 | if (HasAVX) |
2477 | MIB.addReg(RegNo: ImplicitDefReg); |
2478 | |
2479 | MIB.addReg(RegNo: OpReg); |
2480 | updateValueMap(I, Reg: ResultReg); |
2481 | return true; |
2482 | } |
2483 | |
2484 | bool X86FastISel::X86SelectFPExt(const Instruction *I) { |
2485 | if (Subtarget->hasSSE2() && I->getType()->isDoubleTy() && |
2486 | I->getOperand(i: 0)->getType()->isFloatTy()) { |
2487 | bool HasAVX512 = Subtarget->hasAVX512(); |
2488 | // fpext from float to double. |
2489 | unsigned Opc = |
2490 | HasAVX512 ? X86::VCVTSS2SDZrr |
2491 | : Subtarget->hasAVX() ? X86::VCVTSS2SDrr : X86::CVTSS2SDrr; |
2492 | return X86SelectFPExtOrFPTrunc(I, TargetOpc: Opc, RC: TLI.getRegClassFor(VT: MVT::f64)); |
2493 | } |
2494 | |
2495 | return false; |
2496 | } |
2497 | |
2498 | bool X86FastISel::X86SelectFPTrunc(const Instruction *I) { |
2499 | if (Subtarget->hasSSE2() && I->getType()->isFloatTy() && |
2500 | I->getOperand(i: 0)->getType()->isDoubleTy()) { |
2501 | bool HasAVX512 = Subtarget->hasAVX512(); |
2502 | // fptrunc from double to float. |
2503 | unsigned Opc = |
2504 | HasAVX512 ? X86::VCVTSD2SSZrr |
2505 | : Subtarget->hasAVX() ? X86::VCVTSD2SSrr : X86::CVTSD2SSrr; |
2506 | return X86SelectFPExtOrFPTrunc(I, TargetOpc: Opc, RC: TLI.getRegClassFor(VT: MVT::f32)); |
2507 | } |
2508 | |
2509 | return false; |
2510 | } |
2511 | |
2512 | bool X86FastISel::X86SelectTrunc(const Instruction *I) { |
2513 | EVT SrcVT = TLI.getValueType(DL, Ty: I->getOperand(i: 0)->getType()); |
2514 | EVT DstVT = TLI.getValueType(DL, Ty: I->getType()); |
2515 | |
2516 | // This code only handles truncation to byte. |
2517 | if (DstVT != MVT::i8 && DstVT != MVT::i1) |
2518 | return false; |
2519 | if (!TLI.isTypeLegal(VT: SrcVT)) |
2520 | return false; |
2521 | |
2522 | Register InputReg = getRegForValue(V: I->getOperand(i: 0)); |
2523 | if (!InputReg) |
2524 | // Unhandled operand. Halt "fast" selection and bail. |
2525 | return false; |
2526 | |
2527 | if (SrcVT == MVT::i8) { |
2528 | // Truncate from i8 to i1; no code needed. |
2529 | updateValueMap(I, Reg: InputReg); |
2530 | return true; |
2531 | } |
2532 | |
2533 | // Issue an extract_subreg. |
2534 | Register ResultReg = fastEmitInst_extractsubreg(RetVT: MVT::i8, Op0: InputReg, |
2535 | Idx: X86::sub_8bit); |
2536 | if (!ResultReg) |
2537 | return false; |
2538 | |
2539 | updateValueMap(I, Reg: ResultReg); |
2540 | return true; |
2541 | } |
2542 | |
2543 | bool X86FastISel::IsMemcpySmall(uint64_t Len) { |
2544 | return Len <= (Subtarget->is64Bit() ? 32 : 16); |
2545 | } |
2546 | |
2547 | bool X86FastISel::TryEmitSmallMemcpy(X86AddressMode DestAM, |
2548 | X86AddressMode SrcAM, uint64_t Len) { |
2549 | |
2550 | // Make sure we don't bloat code by inlining very large memcpy's. |
2551 | if (!IsMemcpySmall(Len)) |
2552 | return false; |
2553 | |
2554 | bool i64Legal = Subtarget->is64Bit(); |
2555 | |
2556 | // We don't care about alignment here since we just emit integer accesses. |
2557 | while (Len) { |
2558 | MVT VT; |
2559 | if (Len >= 8 && i64Legal) |
2560 | VT = MVT::i64; |
2561 | else if (Len >= 4) |
2562 | VT = MVT::i32; |
2563 | else if (Len >= 2) |
2564 | VT = MVT::i16; |
2565 | else |
2566 | VT = MVT::i8; |
2567 | |
2568 | unsigned Reg; |
2569 | bool RV = X86FastEmitLoad(VT, AM&: SrcAM, MMO: nullptr, ResultReg&: Reg); |
2570 | RV &= X86FastEmitStore(VT, ValReg: Reg, AM&: DestAM); |
2571 | assert(RV && "Failed to emit load or store??" ); |
2572 | (void)RV; |
2573 | |
2574 | unsigned Size = VT.getSizeInBits()/8; |
2575 | Len -= Size; |
2576 | DestAM.Disp += Size; |
2577 | SrcAM.Disp += Size; |
2578 | } |
2579 | |
2580 | return true; |
2581 | } |
2582 | |
2583 | bool X86FastISel::fastLowerIntrinsicCall(const IntrinsicInst *II) { |
2584 | // FIXME: Handle more intrinsics. |
2585 | switch (II->getIntrinsicID()) { |
2586 | default: return false; |
2587 | case Intrinsic::convert_from_fp16: |
2588 | case Intrinsic::convert_to_fp16: { |
2589 | if (Subtarget->useSoftFloat() || !Subtarget->hasF16C()) |
2590 | return false; |
2591 | |
2592 | const Value *Op = II->getArgOperand(i: 0); |
2593 | Register InputReg = getRegForValue(V: Op); |
2594 | if (InputReg == 0) |
2595 | return false; |
2596 | |
2597 | // F16C only allows converting from float to half and from half to float. |
2598 | bool IsFloatToHalf = II->getIntrinsicID() == Intrinsic::convert_to_fp16; |
2599 | if (IsFloatToHalf) { |
2600 | if (!Op->getType()->isFloatTy()) |
2601 | return false; |
2602 | } else { |
2603 | if (!II->getType()->isFloatTy()) |
2604 | return false; |
2605 | } |
2606 | |
2607 | unsigned ResultReg = 0; |
2608 | const TargetRegisterClass *RC = TLI.getRegClassFor(VT: MVT::v8i16); |
2609 | if (IsFloatToHalf) { |
2610 | // 'InputReg' is implicitly promoted from register class FR32 to |
2611 | // register class VR128 by method 'constrainOperandRegClass' which is |
2612 | // directly called by 'fastEmitInst_ri'. |
2613 | // Instruction VCVTPS2PHrr takes an extra immediate operand which is |
2614 | // used to provide rounding control: use MXCSR.RC, encoded as 0b100. |
2615 | // It's consistent with the other FP instructions, which are usually |
2616 | // controlled by MXCSR. |
2617 | unsigned Opc = Subtarget->hasVLX() ? X86::VCVTPS2PHZ128rr |
2618 | : X86::VCVTPS2PHrr; |
2619 | InputReg = fastEmitInst_ri(MachineInstOpcode: Opc, RC, Op0: InputReg, Imm: 4); |
2620 | |
2621 | // Move the lower 32-bits of ResultReg to another register of class GR32. |
2622 | Opc = Subtarget->hasAVX512() ? X86::VMOVPDI2DIZrr |
2623 | : X86::VMOVPDI2DIrr; |
2624 | ResultReg = createResultReg(RC: &X86::GR32RegClass); |
2625 | BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, MCID: TII.get(Opcode: Opc), DestReg: ResultReg) |
2626 | .addReg(RegNo: InputReg, flags: RegState::Kill); |
2627 | |
2628 | // The result value is in the lower 16-bits of ResultReg. |
2629 | unsigned RegIdx = X86::sub_16bit; |
2630 | ResultReg = fastEmitInst_extractsubreg(RetVT: MVT::i16, Op0: ResultReg, Idx: RegIdx); |
2631 | } else { |
2632 | assert(Op->getType()->isIntegerTy(16) && "Expected a 16-bit integer!" ); |
2633 | // Explicitly zero-extend the input to 32-bit. |
2634 | InputReg = fastEmit_r(VT: MVT::i16, RetVT: MVT::i32, Opcode: ISD::ZERO_EXTEND, Op0: InputReg); |
2635 | |
2636 | // The following SCALAR_TO_VECTOR will be expanded into a VMOVDI2PDIrr. |
2637 | InputReg = fastEmit_r(VT: MVT::i32, RetVT: MVT::v4i32, Opcode: ISD::SCALAR_TO_VECTOR, |
2638 | Op0: InputReg); |
2639 | |
2640 | unsigned Opc = Subtarget->hasVLX() ? X86::VCVTPH2PSZ128rr |
2641 | : X86::VCVTPH2PSrr; |
2642 | InputReg = fastEmitInst_r(MachineInstOpcode: Opc, RC, Op0: InputReg); |
2643 | |
2644 | // The result value is in the lower 32-bits of ResultReg. |
2645 | // Emit an explicit copy from register class VR128 to register class FR32. |
2646 | ResultReg = createResultReg(RC: TLI.getRegClassFor(VT: MVT::f32)); |
2647 | BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, |
2648 | MCID: TII.get(Opcode: TargetOpcode::COPY), DestReg: ResultReg) |
2649 | .addReg(RegNo: InputReg, flags: RegState::Kill); |
2650 | } |
2651 | |
2652 | updateValueMap(I: II, Reg: ResultReg); |
2653 | return true; |
2654 | } |
2655 | case Intrinsic::frameaddress: { |
2656 | MachineFunction *MF = FuncInfo.MF; |
2657 | if (MF->getTarget().getMCAsmInfo()->usesWindowsCFI()) |
2658 | return false; |
2659 | |
2660 | Type *RetTy = II->getCalledFunction()->getReturnType(); |
2661 | |
2662 | MVT VT; |
2663 | if (!isTypeLegal(Ty: RetTy, VT)) |
2664 | return false; |
2665 | |
2666 | unsigned Opc; |
2667 | const TargetRegisterClass *RC = nullptr; |
2668 | |
2669 | switch (VT.SimpleTy) { |
2670 | default: llvm_unreachable("Invalid result type for frameaddress." ); |
2671 | case MVT::i32: Opc = X86::MOV32rm; RC = &X86::GR32RegClass; break; |
2672 | case MVT::i64: Opc = X86::MOV64rm; RC = &X86::GR64RegClass; break; |
2673 | } |
2674 | |
2675 | // This needs to be set before we call getPtrSizedFrameRegister, otherwise |
2676 | // we get the wrong frame register. |
2677 | MachineFrameInfo &MFI = MF->getFrameInfo(); |
2678 | MFI.setFrameAddressIsTaken(true); |
2679 | |
2680 | const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo(); |
2681 | unsigned FrameReg = RegInfo->getPtrSizedFrameRegister(MF: *MF); |
2682 | assert(((FrameReg == X86::RBP && VT == MVT::i64) || |
2683 | (FrameReg == X86::EBP && VT == MVT::i32)) && |
2684 | "Invalid Frame Register!" ); |
2685 | |
2686 | // Always make a copy of the frame register to a vreg first, so that we |
2687 | // never directly reference the frame register (the TwoAddressInstruction- |
2688 | // Pass doesn't like that). |
2689 | Register SrcReg = createResultReg(RC); |
2690 | BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, |
2691 | MCID: TII.get(Opcode: TargetOpcode::COPY), DestReg: SrcReg).addReg(RegNo: FrameReg); |
2692 | |
2693 | // Now recursively load from the frame address. |
2694 | // movq (%rbp), %rax |
2695 | // movq (%rax), %rax |
2696 | // movq (%rax), %rax |
2697 | // ... |
2698 | unsigned Depth = cast<ConstantInt>(Val: II->getOperand(i_nocapture: 0))->getZExtValue(); |
2699 | while (Depth--) { |
2700 | Register DestReg = createResultReg(RC); |
2701 | addDirectMem(MIB: BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, |
2702 | MCID: TII.get(Opcode: Opc), DestReg), Reg: SrcReg); |
2703 | SrcReg = DestReg; |
2704 | } |
2705 | |
2706 | updateValueMap(I: II, Reg: SrcReg); |
2707 | return true; |
2708 | } |
2709 | case Intrinsic::memcpy: { |
2710 | const MemCpyInst *MCI = cast<MemCpyInst>(Val: II); |
2711 | // Don't handle volatile or variable length memcpys. |
2712 | if (MCI->isVolatile()) |
2713 | return false; |
2714 | |
2715 | if (isa<ConstantInt>(Val: MCI->getLength())) { |
2716 | // Small memcpy's are common enough that we want to do them |
2717 | // without a call if possible. |
2718 | uint64_t Len = cast<ConstantInt>(Val: MCI->getLength())->getZExtValue(); |
2719 | if (IsMemcpySmall(Len)) { |
2720 | X86AddressMode DestAM, SrcAM; |
2721 | if (!X86SelectAddress(V: MCI->getRawDest(), AM&: DestAM) || |
2722 | !X86SelectAddress(V: MCI->getRawSource(), AM&: SrcAM)) |
2723 | return false; |
2724 | TryEmitSmallMemcpy(DestAM, SrcAM, Len); |
2725 | return true; |
2726 | } |
2727 | } |
2728 | |
2729 | unsigned SizeWidth = Subtarget->is64Bit() ? 64 : 32; |
2730 | if (!MCI->getLength()->getType()->isIntegerTy(Bitwidth: SizeWidth)) |
2731 | return false; |
2732 | |
2733 | if (MCI->getSourceAddressSpace() > 255 || MCI->getDestAddressSpace() > 255) |
2734 | return false; |
2735 | |
2736 | return lowerCallTo(CI: II, SymName: "memcpy" , NumArgs: II->arg_size() - 1); |
2737 | } |
2738 | case Intrinsic::memset: { |
2739 | const MemSetInst *MSI = cast<MemSetInst>(Val: II); |
2740 | |
2741 | if (MSI->isVolatile()) |
2742 | return false; |
2743 | |
2744 | unsigned SizeWidth = Subtarget->is64Bit() ? 64 : 32; |
2745 | if (!MSI->getLength()->getType()->isIntegerTy(Bitwidth: SizeWidth)) |
2746 | return false; |
2747 | |
2748 | if (MSI->getDestAddressSpace() > 255) |
2749 | return false; |
2750 | |
2751 | return lowerCallTo(CI: II, SymName: "memset" , NumArgs: II->arg_size() - 1); |
2752 | } |
2753 | case Intrinsic::stackprotector: { |
2754 | // Emit code to store the stack guard onto the stack. |
2755 | EVT PtrTy = TLI.getPointerTy(DL); |
2756 | |
2757 | const Value *Op1 = II->getArgOperand(i: 0); // The guard's value. |
2758 | const AllocaInst *Slot = cast<AllocaInst>(Val: II->getArgOperand(i: 1)); |
2759 | |
2760 | MFI.setStackProtectorIndex(FuncInfo.StaticAllocaMap[Slot]); |
2761 | |
2762 | // Grab the frame index. |
2763 | X86AddressMode AM; |
2764 | if (!X86SelectAddress(V: Slot, AM)) return false; |
2765 | if (!X86FastEmitStore(VT: PtrTy, Val: Op1, AM)) return false; |
2766 | return true; |
2767 | } |
2768 | case Intrinsic::dbg_declare: { |
2769 | const DbgDeclareInst *DI = cast<DbgDeclareInst>(Val: II); |
2770 | X86AddressMode AM; |
2771 | assert(DI->getAddress() && "Null address should be checked earlier!" ); |
2772 | if (!X86SelectAddress(V: DI->getAddress(), AM)) |
2773 | return false; |
2774 | const MCInstrDesc &II = TII.get(Opcode: TargetOpcode::DBG_VALUE); |
2775 | assert(DI->getVariable()->isValidLocationForIntrinsic(MIMD.getDL()) && |
2776 | "Expected inlined-at fields to agree" ); |
2777 | addFullAddress(MIB: BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, MCID: II), AM) |
2778 | .addImm(Val: 0) |
2779 | .addMetadata(MD: DI->getVariable()) |
2780 | .addMetadata(MD: DI->getExpression()); |
2781 | return true; |
2782 | } |
2783 | case Intrinsic::trap: { |
2784 | BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, MCID: TII.get(Opcode: X86::TRAP)); |
2785 | return true; |
2786 | } |
2787 | case Intrinsic::sqrt: { |
2788 | if (!Subtarget->hasSSE1()) |
2789 | return false; |
2790 | |
2791 | Type *RetTy = II->getCalledFunction()->getReturnType(); |
2792 | |
2793 | MVT VT; |
2794 | if (!isTypeLegal(Ty: RetTy, VT)) |
2795 | return false; |
2796 | |
2797 | // Unfortunately we can't use fastEmit_r, because the AVX version of FSQRT |
2798 | // is not generated by FastISel yet. |
2799 | // FIXME: Update this code once tablegen can handle it. |
2800 | static const uint16_t SqrtOpc[3][2] = { |
2801 | { X86::SQRTSSr, X86::SQRTSDr }, |
2802 | { X86::VSQRTSSr, X86::VSQRTSDr }, |
2803 | { X86::VSQRTSSZr, X86::VSQRTSDZr }, |
2804 | }; |
2805 | unsigned AVXLevel = Subtarget->hasAVX512() ? 2 : |
2806 | Subtarget->hasAVX() ? 1 : |
2807 | 0; |
2808 | unsigned Opc; |
2809 | switch (VT.SimpleTy) { |
2810 | default: return false; |
2811 | case MVT::f32: Opc = SqrtOpc[AVXLevel][0]; break; |
2812 | case MVT::f64: Opc = SqrtOpc[AVXLevel][1]; break; |
2813 | } |
2814 | |
2815 | const Value *SrcVal = II->getArgOperand(i: 0); |
2816 | Register SrcReg = getRegForValue(V: SrcVal); |
2817 | |
2818 | if (SrcReg == 0) |
2819 | return false; |
2820 | |
2821 | const TargetRegisterClass *RC = TLI.getRegClassFor(VT); |
2822 | unsigned ImplicitDefReg = 0; |
2823 | if (AVXLevel > 0) { |
2824 | ImplicitDefReg = createResultReg(RC); |
2825 | BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, |
2826 | MCID: TII.get(Opcode: TargetOpcode::IMPLICIT_DEF), DestReg: ImplicitDefReg); |
2827 | } |
2828 | |
2829 | Register ResultReg = createResultReg(RC); |
2830 | MachineInstrBuilder MIB; |
2831 | MIB = BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, MCID: TII.get(Opcode: Opc), |
2832 | DestReg: ResultReg); |
2833 | |
2834 | if (ImplicitDefReg) |
2835 | MIB.addReg(RegNo: ImplicitDefReg); |
2836 | |
2837 | MIB.addReg(RegNo: SrcReg); |
2838 | |
2839 | updateValueMap(I: II, Reg: ResultReg); |
2840 | return true; |
2841 | } |
2842 | case Intrinsic::sadd_with_overflow: |
2843 | case Intrinsic::uadd_with_overflow: |
2844 | case Intrinsic::ssub_with_overflow: |
2845 | case Intrinsic::usub_with_overflow: |
2846 | case Intrinsic::smul_with_overflow: |
2847 | case Intrinsic::umul_with_overflow: { |
2848 | // This implements the basic lowering of the xalu with overflow intrinsics |
2849 | // into add/sub/mul followed by either seto or setb. |
2850 | const Function *Callee = II->getCalledFunction(); |
2851 | auto *Ty = cast<StructType>(Val: Callee->getReturnType()); |
2852 | Type *RetTy = Ty->getTypeAtIndex(N: 0U); |
2853 | assert(Ty->getTypeAtIndex(1)->isIntegerTy() && |
2854 | Ty->getTypeAtIndex(1)->getScalarSizeInBits() == 1 && |
2855 | "Overflow value expected to be an i1" ); |
2856 | |
2857 | MVT VT; |
2858 | if (!isTypeLegal(Ty: RetTy, VT)) |
2859 | return false; |
2860 | |
2861 | if (VT < MVT::i8 || VT > MVT::i64) |
2862 | return false; |
2863 | |
2864 | const Value *LHS = II->getArgOperand(i: 0); |
2865 | const Value *RHS = II->getArgOperand(i: 1); |
2866 | |
2867 | // Canonicalize immediate to the RHS. |
2868 | if (isa<ConstantInt>(Val: LHS) && !isa<ConstantInt>(Val: RHS) && II->isCommutative()) |
2869 | std::swap(a&: LHS, b&: RHS); |
2870 | |
2871 | unsigned BaseOpc, CondCode; |
2872 | switch (II->getIntrinsicID()) { |
2873 | default: llvm_unreachable("Unexpected intrinsic!" ); |
2874 | case Intrinsic::sadd_with_overflow: |
2875 | BaseOpc = ISD::ADD; CondCode = X86::COND_O; break; |
2876 | case Intrinsic::uadd_with_overflow: |
2877 | BaseOpc = ISD::ADD; CondCode = X86::COND_B; break; |
2878 | case Intrinsic::ssub_with_overflow: |
2879 | BaseOpc = ISD::SUB; CondCode = X86::COND_O; break; |
2880 | case Intrinsic::usub_with_overflow: |
2881 | BaseOpc = ISD::SUB; CondCode = X86::COND_B; break; |
2882 | case Intrinsic::smul_with_overflow: |
2883 | BaseOpc = X86ISD::SMUL; CondCode = X86::COND_O; break; |
2884 | case Intrinsic::umul_with_overflow: |
2885 | BaseOpc = X86ISD::UMUL; CondCode = X86::COND_O; break; |
2886 | } |
2887 | |
2888 | Register LHSReg = getRegForValue(V: LHS); |
2889 | if (LHSReg == 0) |
2890 | return false; |
2891 | |
2892 | unsigned ResultReg = 0; |
2893 | // Check if we have an immediate version. |
2894 | if (const auto *CI = dyn_cast<ConstantInt>(Val: RHS)) { |
2895 | static const uint16_t Opc[2][4] = { |
2896 | { X86::INC8r, X86::INC16r, X86::INC32r, X86::INC64r }, |
2897 | { X86::DEC8r, X86::DEC16r, X86::DEC32r, X86::DEC64r } |
2898 | }; |
2899 | |
2900 | if (CI->isOne() && (BaseOpc == ISD::ADD || BaseOpc == ISD::SUB) && |
2901 | CondCode == X86::COND_O) { |
2902 | // We can use INC/DEC. |
2903 | ResultReg = createResultReg(RC: TLI.getRegClassFor(VT)); |
2904 | bool IsDec = BaseOpc == ISD::SUB; |
2905 | BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, |
2906 | MCID: TII.get(Opcode: Opc[IsDec][VT.SimpleTy-MVT::i8]), DestReg: ResultReg) |
2907 | .addReg(RegNo: LHSReg); |
2908 | } else |
2909 | ResultReg = fastEmit_ri(VT, RetVT: VT, Opcode: BaseOpc, Op0: LHSReg, imm1: CI->getZExtValue()); |
2910 | } |
2911 | |
2912 | unsigned RHSReg; |
2913 | if (!ResultReg) { |
2914 | RHSReg = getRegForValue(V: RHS); |
2915 | if (RHSReg == 0) |
2916 | return false; |
2917 | ResultReg = fastEmit_rr(VT, RetVT: VT, Opcode: BaseOpc, Op0: LHSReg, Op1: RHSReg); |
2918 | } |
2919 | |
2920 | // FastISel doesn't have a pattern for all X86::MUL*r and X86::IMUL*r. Emit |
2921 | // it manually. |
2922 | if (BaseOpc == X86ISD::UMUL && !ResultReg) { |
2923 | static const uint16_t MULOpc[] = |
2924 | { X86::MUL8r, X86::MUL16r, X86::MUL32r, X86::MUL64r }; |
2925 | static const MCPhysReg Reg[] = { X86::AL, X86::AX, X86::EAX, X86::RAX }; |
2926 | // First copy the first operand into RAX, which is an implicit input to |
2927 | // the X86::MUL*r instruction. |
2928 | BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, |
2929 | MCID: TII.get(Opcode: TargetOpcode::COPY), DestReg: Reg[VT.SimpleTy-MVT::i8]) |
2930 | .addReg(RegNo: LHSReg); |
2931 | ResultReg = fastEmitInst_r(MachineInstOpcode: MULOpc[VT.SimpleTy-MVT::i8], |
2932 | RC: TLI.getRegClassFor(VT), Op0: RHSReg); |
2933 | } else if (BaseOpc == X86ISD::SMUL && !ResultReg) { |
2934 | static const uint16_t MULOpc[] = |
2935 | { X86::IMUL8r, X86::IMUL16rr, X86::IMUL32rr, X86::IMUL64rr }; |
2936 | if (VT == MVT::i8) { |
2937 | // Copy the first operand into AL, which is an implicit input to the |
2938 | // X86::IMUL8r instruction. |
2939 | BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, |
2940 | MCID: TII.get(Opcode: TargetOpcode::COPY), DestReg: X86::AL) |
2941 | .addReg(RegNo: LHSReg); |
2942 | ResultReg = fastEmitInst_r(MachineInstOpcode: MULOpc[0], RC: TLI.getRegClassFor(VT), Op0: RHSReg); |
2943 | } else |
2944 | ResultReg = fastEmitInst_rr(MachineInstOpcode: MULOpc[VT.SimpleTy-MVT::i8], |
2945 | RC: TLI.getRegClassFor(VT), Op0: LHSReg, Op1: RHSReg); |
2946 | } |
2947 | |
2948 | if (!ResultReg) |
2949 | return false; |
2950 | |
2951 | // Assign to a GPR since the overflow return value is lowered to a SETcc. |
2952 | Register ResultReg2 = createResultReg(RC: &X86::GR8RegClass); |
2953 | assert((ResultReg+1) == ResultReg2 && "Nonconsecutive result registers." ); |
2954 | BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, MCID: TII.get(Opcode: X86::SETCCr), |
2955 | DestReg: ResultReg2).addImm(Val: CondCode); |
2956 | |
2957 | updateValueMap(I: II, Reg: ResultReg, NumRegs: 2); |
2958 | return true; |
2959 | } |
2960 | case Intrinsic::x86_sse_cvttss2si: |
2961 | case Intrinsic::x86_sse_cvttss2si64: |
2962 | case Intrinsic::x86_sse2_cvttsd2si: |
2963 | case Intrinsic::x86_sse2_cvttsd2si64: { |
2964 | bool IsInputDouble; |
2965 | switch (II->getIntrinsicID()) { |
2966 | default: llvm_unreachable("Unexpected intrinsic." ); |
2967 | case Intrinsic::x86_sse_cvttss2si: |
2968 | case Intrinsic::x86_sse_cvttss2si64: |
2969 | if (!Subtarget->hasSSE1()) |
2970 | return false; |
2971 | IsInputDouble = false; |
2972 | break; |
2973 | case Intrinsic::x86_sse2_cvttsd2si: |
2974 | case Intrinsic::x86_sse2_cvttsd2si64: |
2975 | if (!Subtarget->hasSSE2()) |
2976 | return false; |
2977 | IsInputDouble = true; |
2978 | break; |
2979 | } |
2980 | |
2981 | Type *RetTy = II->getCalledFunction()->getReturnType(); |
2982 | MVT VT; |
2983 | if (!isTypeLegal(Ty: RetTy, VT)) |
2984 | return false; |
2985 | |
2986 | static const uint16_t CvtOpc[3][2][2] = { |
2987 | { { X86::CVTTSS2SIrr, X86::CVTTSS2SI64rr }, |
2988 | { X86::CVTTSD2SIrr, X86::CVTTSD2SI64rr } }, |
2989 | { { X86::VCVTTSS2SIrr, X86::VCVTTSS2SI64rr }, |
2990 | { X86::VCVTTSD2SIrr, X86::VCVTTSD2SI64rr } }, |
2991 | { { X86::VCVTTSS2SIZrr, X86::VCVTTSS2SI64Zrr }, |
2992 | { X86::VCVTTSD2SIZrr, X86::VCVTTSD2SI64Zrr } }, |
2993 | }; |
2994 | unsigned AVXLevel = Subtarget->hasAVX512() ? 2 : |
2995 | Subtarget->hasAVX() ? 1 : |
2996 | 0; |
2997 | unsigned Opc; |
2998 | switch (VT.SimpleTy) { |
2999 | default: llvm_unreachable("Unexpected result type." ); |
3000 | case MVT::i32: Opc = CvtOpc[AVXLevel][IsInputDouble][0]; break; |
3001 | case MVT::i64: Opc = CvtOpc[AVXLevel][IsInputDouble][1]; break; |
3002 | } |
3003 | |
3004 | // Check if we can fold insertelement instructions into the convert. |
3005 | const Value *Op = II->getArgOperand(i: 0); |
3006 | while (auto *IE = dyn_cast<InsertElementInst>(Val: Op)) { |
3007 | const Value *Index = IE->getOperand(i_nocapture: 2); |
3008 | if (!isa<ConstantInt>(Val: Index)) |
3009 | break; |
3010 | unsigned Idx = cast<ConstantInt>(Val: Index)->getZExtValue(); |
3011 | |
3012 | if (Idx == 0) { |
3013 | Op = IE->getOperand(i_nocapture: 1); |
3014 | break; |
3015 | } |
3016 | Op = IE->getOperand(i_nocapture: 0); |
3017 | } |
3018 | |
3019 | Register Reg = getRegForValue(V: Op); |
3020 | if (Reg == 0) |
3021 | return false; |
3022 | |
3023 | Register ResultReg = createResultReg(RC: TLI.getRegClassFor(VT)); |
3024 | BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, MCID: TII.get(Opcode: Opc), DestReg: ResultReg) |
3025 | .addReg(RegNo: Reg); |
3026 | |
3027 | updateValueMap(I: II, Reg: ResultReg); |
3028 | return true; |
3029 | } |
3030 | case Intrinsic::x86_sse42_crc32_32_8: |
3031 | case Intrinsic::x86_sse42_crc32_32_16: |
3032 | case Intrinsic::x86_sse42_crc32_32_32: |
3033 | case Intrinsic::x86_sse42_crc32_64_64: { |
3034 | if (!Subtarget->hasCRC32()) |
3035 | return false; |
3036 | |
3037 | Type *RetTy = II->getCalledFunction()->getReturnType(); |
3038 | |
3039 | MVT VT; |
3040 | if (!isTypeLegal(Ty: RetTy, VT)) |
3041 | return false; |
3042 | |
3043 | unsigned Opc; |
3044 | const TargetRegisterClass *RC = nullptr; |
3045 | |
3046 | switch (II->getIntrinsicID()) { |
3047 | default: |
3048 | llvm_unreachable("Unexpected intrinsic." ); |
3049 | #define GET_EGPR_IF_ENABLED(OPC) Subtarget->hasEGPR() ? OPC##_EVEX : OPC |
3050 | case Intrinsic::x86_sse42_crc32_32_8: |
3051 | Opc = GET_EGPR_IF_ENABLED(X86::CRC32r32r8); |
3052 | RC = &X86::GR32RegClass; |
3053 | break; |
3054 | case Intrinsic::x86_sse42_crc32_32_16: |
3055 | Opc = GET_EGPR_IF_ENABLED(X86::CRC32r32r16); |
3056 | RC = &X86::GR32RegClass; |
3057 | break; |
3058 | case Intrinsic::x86_sse42_crc32_32_32: |
3059 | Opc = GET_EGPR_IF_ENABLED(X86::CRC32r32r32); |
3060 | RC = &X86::GR32RegClass; |
3061 | break; |
3062 | case Intrinsic::x86_sse42_crc32_64_64: |
3063 | Opc = GET_EGPR_IF_ENABLED(X86::CRC32r64r64); |
3064 | RC = &X86::GR64RegClass; |
3065 | break; |
3066 | #undef GET_EGPR_IF_ENABLED |
3067 | } |
3068 | |
3069 | const Value *LHS = II->getArgOperand(i: 0); |
3070 | const Value *RHS = II->getArgOperand(i: 1); |
3071 | |
3072 | Register LHSReg = getRegForValue(V: LHS); |
3073 | Register RHSReg = getRegForValue(V: RHS); |
3074 | if (!LHSReg || !RHSReg) |
3075 | return false; |
3076 | |
3077 | Register ResultReg = fastEmitInst_rr(MachineInstOpcode: Opc, RC, Op0: LHSReg, Op1: RHSReg); |
3078 | if (!ResultReg) |
3079 | return false; |
3080 | |
3081 | updateValueMap(I: II, Reg: ResultReg); |
3082 | return true; |
3083 | } |
3084 | } |
3085 | } |
3086 | |
3087 | bool X86FastISel::fastLowerArguments() { |
3088 | if (!FuncInfo.CanLowerReturn) |
3089 | return false; |
3090 | |
3091 | const Function *F = FuncInfo.Fn; |
3092 | if (F->isVarArg()) |
3093 | return false; |
3094 | |
3095 | CallingConv::ID CC = F->getCallingConv(); |
3096 | if (CC != CallingConv::C) |
3097 | return false; |
3098 | |
3099 | if (Subtarget->isCallingConvWin64(CC)) |
3100 | return false; |
3101 | |
3102 | if (!Subtarget->is64Bit()) |
3103 | return false; |
3104 | |
3105 | if (Subtarget->useSoftFloat()) |
3106 | return false; |
3107 | |
3108 | // Only handle simple cases. i.e. Up to 6 i32/i64 scalar arguments. |
3109 | unsigned GPRCnt = 0; |
3110 | unsigned FPRCnt = 0; |
3111 | for (auto const &Arg : F->args()) { |
3112 | if (Arg.hasAttribute(Kind: Attribute::ByVal) || |
3113 | Arg.hasAttribute(Kind: Attribute::InReg) || |
3114 | Arg.hasAttribute(Kind: Attribute::StructRet) || |
3115 | Arg.hasAttribute(Kind: Attribute::SwiftSelf) || |
3116 | Arg.hasAttribute(Kind: Attribute::SwiftAsync) || |
3117 | Arg.hasAttribute(Kind: Attribute::SwiftError) || |
3118 | Arg.hasAttribute(Kind: Attribute::Nest)) |
3119 | return false; |
3120 | |
3121 | Type *ArgTy = Arg.getType(); |
3122 | if (ArgTy->isStructTy() || ArgTy->isArrayTy() || ArgTy->isVectorTy()) |
3123 | return false; |
3124 | |
3125 | EVT ArgVT = TLI.getValueType(DL, Ty: ArgTy); |
3126 | if (!ArgVT.isSimple()) return false; |
3127 | switch (ArgVT.getSimpleVT().SimpleTy) { |
3128 | default: return false; |
3129 | case MVT::i32: |
3130 | case MVT::i64: |
3131 | ++GPRCnt; |
3132 | break; |
3133 | case MVT::f32: |
3134 | case MVT::f64: |
3135 | if (!Subtarget->hasSSE1()) |
3136 | return false; |
3137 | ++FPRCnt; |
3138 | break; |
3139 | } |
3140 | |
3141 | if (GPRCnt > 6) |
3142 | return false; |
3143 | |
3144 | if (FPRCnt > 8) |
3145 | return false; |
3146 | } |
3147 | |
3148 | static const MCPhysReg GPR32ArgRegs[] = { |
3149 | X86::EDI, X86::ESI, X86::EDX, X86::ECX, X86::R8D, X86::R9D |
3150 | }; |
3151 | static const MCPhysReg GPR64ArgRegs[] = { |
3152 | X86::RDI, X86::RSI, X86::RDX, X86::RCX, X86::R8 , X86::R9 |
3153 | }; |
3154 | static const MCPhysReg XMMArgRegs[] = { |
3155 | X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3, |
3156 | X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7 |
3157 | }; |
3158 | |
3159 | unsigned GPRIdx = 0; |
3160 | unsigned FPRIdx = 0; |
3161 | for (auto const &Arg : F->args()) { |
3162 | MVT VT = TLI.getSimpleValueType(DL, Ty: Arg.getType()); |
3163 | const TargetRegisterClass *RC = TLI.getRegClassFor(VT); |
3164 | unsigned SrcReg; |
3165 | switch (VT.SimpleTy) { |
3166 | default: llvm_unreachable("Unexpected value type." ); |
3167 | case MVT::i32: SrcReg = GPR32ArgRegs[GPRIdx++]; break; |
3168 | case MVT::i64: SrcReg = GPR64ArgRegs[GPRIdx++]; break; |
3169 | case MVT::f32: [[fallthrough]]; |
3170 | case MVT::f64: SrcReg = XMMArgRegs[FPRIdx++]; break; |
3171 | } |
3172 | Register DstReg = FuncInfo.MF->addLiveIn(PReg: SrcReg, RC); |
3173 | // FIXME: Unfortunately it's necessary to emit a copy from the livein copy. |
3174 | // Without this, EmitLiveInCopies may eliminate the livein if its only |
3175 | // use is a bitcast (which isn't turned into an instruction). |
3176 | Register ResultReg = createResultReg(RC); |
3177 | BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, |
3178 | MCID: TII.get(Opcode: TargetOpcode::COPY), DestReg: ResultReg) |
3179 | .addReg(RegNo: DstReg, flags: getKillRegState(B: true)); |
3180 | updateValueMap(I: &Arg, Reg: ResultReg); |
3181 | } |
3182 | return true; |
3183 | } |
3184 | |
3185 | static unsigned computeBytesPoppedByCalleeForSRet(const X86Subtarget *Subtarget, |
3186 | CallingConv::ID CC, |
3187 | const CallBase *CB) { |
3188 | if (Subtarget->is64Bit()) |
3189 | return 0; |
3190 | if (Subtarget->getTargetTriple().isOSMSVCRT()) |
3191 | return 0; |
3192 | if (CC == CallingConv::Fast || CC == CallingConv::GHC || |
3193 | CC == CallingConv::HiPE || CC == CallingConv::Tail || |
3194 | CC == CallingConv::SwiftTail) |
3195 | return 0; |
3196 | |
3197 | if (CB) |
3198 | if (CB->arg_empty() || !CB->paramHasAttr(ArgNo: 0, Kind: Attribute::StructRet) || |
3199 | CB->paramHasAttr(ArgNo: 0, Kind: Attribute::InReg) || Subtarget->isTargetMCU()) |
3200 | return 0; |
3201 | |
3202 | return 4; |
3203 | } |
3204 | |
3205 | bool X86FastISel::fastLowerCall(CallLoweringInfo &CLI) { |
3206 | auto &OutVals = CLI.OutVals; |
3207 | auto &OutFlags = CLI.OutFlags; |
3208 | auto &OutRegs = CLI.OutRegs; |
3209 | auto &Ins = CLI.Ins; |
3210 | auto &InRegs = CLI.InRegs; |
3211 | CallingConv::ID CC = CLI.CallConv; |
3212 | bool &IsTailCall = CLI.IsTailCall; |
3213 | bool IsVarArg = CLI.IsVarArg; |
3214 | const Value *Callee = CLI.Callee; |
3215 | MCSymbol *Symbol = CLI.Symbol; |
3216 | const auto *CB = CLI.CB; |
3217 | |
3218 | bool Is64Bit = Subtarget->is64Bit(); |
3219 | bool IsWin64 = Subtarget->isCallingConvWin64(CC); |
3220 | |
3221 | // Call / invoke instructions with NoCfCheck attribute require special |
3222 | // handling. |
3223 | if (CB && CB->doesNoCfCheck()) |
3224 | return false; |
3225 | |
3226 | // Functions with no_caller_saved_registers that need special handling. |
3227 | if ((CB && isa<CallInst>(Val: CB) && CB->hasFnAttr(Kind: "no_caller_saved_registers" ))) |
3228 | return false; |
3229 | |
3230 | // Functions with no_callee_saved_registers that need special handling. |
3231 | if ((CB && CB->hasFnAttr(Kind: "no_callee_saved_registers" ))) |
3232 | return false; |
3233 | |
3234 | // Indirect calls with CFI checks need special handling. |
3235 | if (CB && CB->isIndirectCall() && CB->getOperandBundle(ID: LLVMContext::OB_kcfi)) |
3236 | return false; |
3237 | |
3238 | // Functions using thunks for indirect calls need to use SDISel. |
3239 | if (Subtarget->useIndirectThunkCalls()) |
3240 | return false; |
3241 | |
3242 | // Handle only C and fastcc calling conventions for now. |
3243 | switch (CC) { |
3244 | default: return false; |
3245 | case CallingConv::C: |
3246 | case CallingConv::Fast: |
3247 | case CallingConv::Tail: |
3248 | case CallingConv::Swift: |
3249 | case CallingConv::SwiftTail: |
3250 | case CallingConv::X86_FastCall: |
3251 | case CallingConv::X86_StdCall: |
3252 | case CallingConv::X86_ThisCall: |
3253 | case CallingConv::Win64: |
3254 | case CallingConv::X86_64_SysV: |
3255 | case CallingConv::CFGuard_Check: |
3256 | break; |
3257 | } |
3258 | |
3259 | // Allow SelectionDAG isel to handle tail calls. |
3260 | if (IsTailCall) |
3261 | return false; |
3262 | |
3263 | // fastcc with -tailcallopt is intended to provide a guaranteed |
3264 | // tail call optimization. Fastisel doesn't know how to do that. |
3265 | if ((CC == CallingConv::Fast && TM.Options.GuaranteedTailCallOpt) || |
3266 | CC == CallingConv::Tail || CC == CallingConv::SwiftTail) |
3267 | return false; |
3268 | |
3269 | // Don't know how to handle Win64 varargs yet. Nothing special needed for |
3270 | // x86-32. Special handling for x86-64 is implemented. |
3271 | if (IsVarArg && IsWin64) |
3272 | return false; |
3273 | |
3274 | // Don't know about inalloca yet. |
3275 | if (CLI.CB && CLI.CB->hasInAllocaArgument()) |
3276 | return false; |
3277 | |
3278 | for (auto Flag : CLI.OutFlags) |
3279 | if (Flag.isSwiftError() || Flag.isPreallocated()) |
3280 | return false; |
3281 | |
3282 | SmallVector<MVT, 16> OutVTs; |
3283 | SmallVector<unsigned, 16> ArgRegs; |
3284 | |
3285 | // If this is a constant i1/i8/i16 argument, promote to i32 to avoid an extra |
3286 | // instruction. This is safe because it is common to all FastISel supported |
3287 | // calling conventions on x86. |
3288 | for (int i = 0, e = OutVals.size(); i != e; ++i) { |
3289 | Value *&Val = OutVals[i]; |
3290 | ISD::ArgFlagsTy Flags = OutFlags[i]; |
3291 | if (auto *CI = dyn_cast<ConstantInt>(Val)) { |
3292 | if (CI->getBitWidth() < 32) { |
3293 | if (Flags.isSExt()) |
3294 | Val = ConstantInt::get(Context&: CI->getContext(), V: CI->getValue().sext(width: 32)); |
3295 | else |
3296 | Val = ConstantInt::get(Context&: CI->getContext(), V: CI->getValue().zext(width: 32)); |
3297 | } |
3298 | } |
3299 | |
3300 | // Passing bools around ends up doing a trunc to i1 and passing it. |
3301 | // Codegen this as an argument + "and 1". |
3302 | MVT VT; |
3303 | auto *TI = dyn_cast<TruncInst>(Val); |
3304 | unsigned ResultReg; |
3305 | if (TI && TI->getType()->isIntegerTy(Bitwidth: 1) && CLI.CB && |
3306 | (TI->getParent() == CLI.CB->getParent()) && TI->hasOneUse()) { |
3307 | Value *PrevVal = TI->getOperand(i_nocapture: 0); |
3308 | ResultReg = getRegForValue(V: PrevVal); |
3309 | |
3310 | if (!ResultReg) |
3311 | return false; |
3312 | |
3313 | if (!isTypeLegal(Ty: PrevVal->getType(), VT)) |
3314 | return false; |
3315 | |
3316 | ResultReg = fastEmit_ri(VT, RetVT: VT, Opcode: ISD::AND, Op0: ResultReg, imm1: 1); |
3317 | } else { |
3318 | if (!isTypeLegal(Ty: Val->getType(), VT) || |
3319 | (VT.isVector() && VT.getVectorElementType() == MVT::i1)) |
3320 | return false; |
3321 | ResultReg = getRegForValue(V: Val); |
3322 | } |
3323 | |
3324 | if (!ResultReg) |
3325 | return false; |
3326 | |
3327 | ArgRegs.push_back(Elt: ResultReg); |
3328 | OutVTs.push_back(Elt: VT); |
3329 | } |
3330 | |
3331 | // Analyze operands of the call, assigning locations to each operand. |
3332 | SmallVector<CCValAssign, 16> ArgLocs; |
3333 | CCState CCInfo(CC, IsVarArg, *FuncInfo.MF, ArgLocs, CLI.RetTy->getContext()); |
3334 | |
3335 | // Allocate shadow area for Win64 |
3336 | if (IsWin64) |
3337 | CCInfo.AllocateStack(Size: 32, Alignment: Align(8)); |
3338 | |
3339 | CCInfo.AnalyzeCallOperands(ArgVTs&: OutVTs, Flags&: OutFlags, Fn: CC_X86); |
3340 | |
3341 | // Get a count of how many bytes are to be pushed on the stack. |
3342 | unsigned NumBytes = CCInfo.getAlignedCallFrameSize(); |
3343 | |
3344 | // Issue CALLSEQ_START |
3345 | unsigned AdjStackDown = TII.getCallFrameSetupOpcode(); |
3346 | BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, MCID: TII.get(Opcode: AdjStackDown)) |
3347 | .addImm(Val: NumBytes).addImm(Val: 0).addImm(Val: 0); |
3348 | |
3349 | // Walk the register/memloc assignments, inserting copies/loads. |
3350 | const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo(); |
3351 | for (const CCValAssign &VA : ArgLocs) { |
3352 | const Value *ArgVal = OutVals[VA.getValNo()]; |
3353 | MVT ArgVT = OutVTs[VA.getValNo()]; |
3354 | |
3355 | if (ArgVT == MVT::x86mmx) |
3356 | return false; |
3357 | |
3358 | unsigned ArgReg = ArgRegs[VA.getValNo()]; |
3359 | |
3360 | // Promote the value if needed. |
3361 | switch (VA.getLocInfo()) { |
3362 | case CCValAssign::Full: break; |
3363 | case CCValAssign::SExt: { |
3364 | assert(VA.getLocVT().isInteger() && !VA.getLocVT().isVector() && |
3365 | "Unexpected extend" ); |
3366 | |
3367 | if (ArgVT == MVT::i1) |
3368 | return false; |
3369 | |
3370 | bool Emitted = X86FastEmitExtend(Opc: ISD::SIGN_EXTEND, DstVT: VA.getLocVT(), Src: ArgReg, |
3371 | SrcVT: ArgVT, ResultReg&: ArgReg); |
3372 | assert(Emitted && "Failed to emit a sext!" ); (void)Emitted; |
3373 | ArgVT = VA.getLocVT(); |
3374 | break; |
3375 | } |
3376 | case CCValAssign::ZExt: { |
3377 | assert(VA.getLocVT().isInteger() && !VA.getLocVT().isVector() && |
3378 | "Unexpected extend" ); |
3379 | |
3380 | // Handle zero-extension from i1 to i8, which is common. |
3381 | if (ArgVT == MVT::i1) { |
3382 | // Set the high bits to zero. |
3383 | ArgReg = fastEmitZExtFromI1(VT: MVT::i8, Op0: ArgReg); |
3384 | ArgVT = MVT::i8; |
3385 | |
3386 | if (ArgReg == 0) |
3387 | return false; |
3388 | } |
3389 | |
3390 | bool Emitted = X86FastEmitExtend(Opc: ISD::ZERO_EXTEND, DstVT: VA.getLocVT(), Src: ArgReg, |
3391 | SrcVT: ArgVT, ResultReg&: ArgReg); |
3392 | assert(Emitted && "Failed to emit a zext!" ); (void)Emitted; |
3393 | ArgVT = VA.getLocVT(); |
3394 | break; |
3395 | } |
3396 | case CCValAssign::AExt: { |
3397 | assert(VA.getLocVT().isInteger() && !VA.getLocVT().isVector() && |
3398 | "Unexpected extend" ); |
3399 | bool Emitted = X86FastEmitExtend(Opc: ISD::ANY_EXTEND, DstVT: VA.getLocVT(), Src: ArgReg, |
3400 | SrcVT: ArgVT, ResultReg&: ArgReg); |
3401 | if (!Emitted) |
3402 | Emitted = X86FastEmitExtend(Opc: ISD::ZERO_EXTEND, DstVT: VA.getLocVT(), Src: ArgReg, |
3403 | SrcVT: ArgVT, ResultReg&: ArgReg); |
3404 | if (!Emitted) |
3405 | Emitted = X86FastEmitExtend(Opc: ISD::SIGN_EXTEND, DstVT: VA.getLocVT(), Src: ArgReg, |
3406 | SrcVT: ArgVT, ResultReg&: ArgReg); |
3407 | |
3408 | assert(Emitted && "Failed to emit a aext!" ); (void)Emitted; |
3409 | ArgVT = VA.getLocVT(); |
3410 | break; |
3411 | } |
3412 | case CCValAssign::BCvt: { |
3413 | ArgReg = fastEmit_r(VT: ArgVT, RetVT: VA.getLocVT(), Opcode: ISD::BITCAST, Op0: ArgReg); |
3414 | assert(ArgReg && "Failed to emit a bitcast!" ); |
3415 | ArgVT = VA.getLocVT(); |
3416 | break; |
3417 | } |
3418 | case CCValAssign::VExt: |
3419 | // VExt has not been implemented, so this should be impossible to reach |
3420 | // for now. However, fallback to Selection DAG isel once implemented. |
3421 | return false; |
3422 | case CCValAssign::AExtUpper: |
3423 | case CCValAssign::SExtUpper: |
3424 | case CCValAssign::ZExtUpper: |
3425 | case CCValAssign::FPExt: |
3426 | case CCValAssign::Trunc: |
3427 | llvm_unreachable("Unexpected loc info!" ); |
3428 | case CCValAssign::Indirect: |
3429 | // FIXME: Indirect doesn't need extending, but fast-isel doesn't fully |
3430 | // support this. |
3431 | return false; |
3432 | } |
3433 | |
3434 | if (VA.isRegLoc()) { |
3435 | BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, |
3436 | MCID: TII.get(Opcode: TargetOpcode::COPY), DestReg: VA.getLocReg()).addReg(RegNo: ArgReg); |
3437 | OutRegs.push_back(Elt: VA.getLocReg()); |
3438 | } else { |
3439 | assert(VA.isMemLoc() && "Unknown value location!" ); |
3440 | |
3441 | // Don't emit stores for undef values. |
3442 | if (isa<UndefValue>(Val: ArgVal)) |
3443 | continue; |
3444 | |
3445 | unsigned LocMemOffset = VA.getLocMemOffset(); |
3446 | X86AddressMode AM; |
3447 | AM.Base.Reg = RegInfo->getStackRegister(); |
3448 | AM.Disp = LocMemOffset; |
3449 | ISD::ArgFlagsTy Flags = OutFlags[VA.getValNo()]; |
3450 | Align Alignment = DL.getABITypeAlign(Ty: ArgVal->getType()); |
3451 | MachineMemOperand *MMO = FuncInfo.MF->getMachineMemOperand( |
3452 | PtrInfo: MachinePointerInfo::getStack(MF&: *FuncInfo.MF, Offset: LocMemOffset), |
3453 | F: MachineMemOperand::MOStore, Size: ArgVT.getStoreSize(), BaseAlignment: Alignment); |
3454 | if (Flags.isByVal()) { |
3455 | X86AddressMode SrcAM; |
3456 | SrcAM.Base.Reg = ArgReg; |
3457 | if (!TryEmitSmallMemcpy(DestAM: AM, SrcAM, Len: Flags.getByValSize())) |
3458 | return false; |
3459 | } else if (isa<ConstantInt>(Val: ArgVal) || isa<ConstantPointerNull>(Val: ArgVal)) { |
3460 | // If this is a really simple value, emit this with the Value* version |
3461 | // of X86FastEmitStore. If it isn't simple, we don't want to do this, |
3462 | // as it can cause us to reevaluate the argument. |
3463 | if (!X86FastEmitStore(VT: ArgVT, Val: ArgVal, AM, MMO)) |
3464 | return false; |
3465 | } else { |
3466 | if (!X86FastEmitStore(VT: ArgVT, ValReg: ArgReg, AM, MMO)) |
3467 | return false; |
3468 | } |
3469 | } |
3470 | } |
3471 | |
3472 | // ELF / PIC requires GOT in the EBX register before function calls via PLT |
3473 | // GOT pointer. |
3474 | if (Subtarget->isPICStyleGOT()) { |
3475 | unsigned Base = getInstrInfo()->getGlobalBaseReg(MF: FuncInfo.MF); |
3476 | BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, |
3477 | MCID: TII.get(Opcode: TargetOpcode::COPY), DestReg: X86::EBX).addReg(RegNo: Base); |
3478 | } |
3479 | |
3480 | if (Is64Bit && IsVarArg && !IsWin64) { |
3481 | // From AMD64 ABI document: |
3482 | // For calls that may call functions that use varargs or stdargs |
3483 | // (prototype-less calls or calls to functions containing ellipsis (...) in |
3484 | // the declaration) %al is used as hidden argument to specify the number |
3485 | // of SSE registers used. The contents of %al do not need to match exactly |
3486 | // the number of registers, but must be an ubound on the number of SSE |
3487 | // registers used and is in the range 0 - 8 inclusive. |
3488 | |
3489 | // Count the number of XMM registers allocated. |
3490 | static const MCPhysReg XMMArgRegs[] = { |
3491 | X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3, |
3492 | X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7 |
3493 | }; |
3494 | unsigned NumXMMRegs = CCInfo.getFirstUnallocated(Regs: XMMArgRegs); |
3495 | assert((Subtarget->hasSSE1() || !NumXMMRegs) |
3496 | && "SSE registers cannot be used when SSE is disabled" ); |
3497 | BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, MCID: TII.get(Opcode: X86::MOV8ri), |
3498 | DestReg: X86::AL).addImm(Val: NumXMMRegs); |
3499 | } |
3500 | |
3501 | // Materialize callee address in a register. FIXME: GV address can be |
3502 | // handled with a CALLpcrel32 instead. |
3503 | X86AddressMode CalleeAM; |
3504 | if (!X86SelectCallAddress(V: Callee, AM&: CalleeAM)) |
3505 | return false; |
3506 | |
3507 | unsigned CalleeOp = 0; |
3508 | const GlobalValue *GV = nullptr; |
3509 | if (CalleeAM.GV != nullptr) { |
3510 | GV = CalleeAM.GV; |
3511 | } else if (CalleeAM.Base.Reg != 0) { |
3512 | CalleeOp = CalleeAM.Base.Reg; |
3513 | } else |
3514 | return false; |
3515 | |
3516 | // Issue the call. |
3517 | MachineInstrBuilder MIB; |
3518 | if (CalleeOp) { |
3519 | // Register-indirect call. |
3520 | unsigned CallOpc = Is64Bit ? X86::CALL64r : X86::CALL32r; |
3521 | MIB = BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, MCID: TII.get(Opcode: CallOpc)) |
3522 | .addReg(RegNo: CalleeOp); |
3523 | } else { |
3524 | // Direct call. |
3525 | assert(GV && "Not a direct call" ); |
3526 | // See if we need any target-specific flags on the GV operand. |
3527 | unsigned char OpFlags = Subtarget->classifyGlobalFunctionReference(GV); |
3528 | if (OpFlags == X86II::MO_PLT && !Is64Bit && |
3529 | TM.getRelocationModel() == Reloc::Static && isa<Function>(Val: GV) && |
3530 | cast<Function>(Val: GV)->isIntrinsic()) |
3531 | OpFlags = X86II::MO_NO_FLAG; |
3532 | |
3533 | // This will be a direct call, or an indirect call through memory for |
3534 | // NonLazyBind calls or dllimport calls. |
3535 | bool NeedLoad = OpFlags == X86II::MO_DLLIMPORT || |
3536 | OpFlags == X86II::MO_GOTPCREL || |
3537 | OpFlags == X86II::MO_GOTPCREL_NORELAX || |
3538 | OpFlags == X86II::MO_COFFSTUB; |
3539 | unsigned CallOpc = NeedLoad |
3540 | ? (Is64Bit ? X86::CALL64m : X86::CALL32m) |
3541 | : (Is64Bit ? X86::CALL64pcrel32 : X86::CALLpcrel32); |
3542 | |
3543 | MIB = BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, MCID: TII.get(Opcode: CallOpc)); |
3544 | if (NeedLoad) |
3545 | MIB.addReg(RegNo: Is64Bit ? X86::RIP : 0).addImm(Val: 1).addReg(RegNo: 0); |
3546 | if (Symbol) |
3547 | MIB.addSym(Sym: Symbol, TargetFlags: OpFlags); |
3548 | else |
3549 | MIB.addGlobalAddress(GV, Offset: 0, TargetFlags: OpFlags); |
3550 | if (NeedLoad) |
3551 | MIB.addReg(RegNo: 0); |
3552 | } |
3553 | |
3554 | // Add a register mask operand representing the call-preserved registers. |
3555 | // Proper defs for return values will be added by setPhysRegsDeadExcept(). |
3556 | MIB.addRegMask(Mask: TRI.getCallPreservedMask(MF: *FuncInfo.MF, CC)); |
3557 | |
3558 | // Add an implicit use GOT pointer in EBX. |
3559 | if (Subtarget->isPICStyleGOT()) |
3560 | MIB.addReg(RegNo: X86::EBX, flags: RegState::Implicit); |
3561 | |
3562 | if (Is64Bit && IsVarArg && !IsWin64) |
3563 | MIB.addReg(RegNo: X86::AL, flags: RegState::Implicit); |
3564 | |
3565 | // Add implicit physical register uses to the call. |
3566 | for (auto Reg : OutRegs) |
3567 | MIB.addReg(RegNo: Reg, flags: RegState::Implicit); |
3568 | |
3569 | // Issue CALLSEQ_END |
3570 | unsigned NumBytesForCalleeToPop = |
3571 | X86::isCalleePop(CallingConv: CC, is64Bit: Subtarget->is64Bit(), IsVarArg, |
3572 | GuaranteeTCO: TM.Options.GuaranteedTailCallOpt) |
3573 | ? NumBytes // Callee pops everything. |
3574 | : computeBytesPoppedByCalleeForSRet(Subtarget, CC, CB: CLI.CB); |
3575 | unsigned AdjStackUp = TII.getCallFrameDestroyOpcode(); |
3576 | BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, MCID: TII.get(Opcode: AdjStackUp)) |
3577 | .addImm(Val: NumBytes).addImm(Val: NumBytesForCalleeToPop); |
3578 | |
3579 | // Now handle call return values. |
3580 | SmallVector<CCValAssign, 16> RVLocs; |
3581 | CCState CCRetInfo(CC, IsVarArg, *FuncInfo.MF, RVLocs, |
3582 | CLI.RetTy->getContext()); |
3583 | CCRetInfo.AnalyzeCallResult(Ins, Fn: RetCC_X86); |
3584 | |
3585 | // Copy all of the result registers out of their specified physreg. |
3586 | Register ResultReg = FuncInfo.CreateRegs(Ty: CLI.RetTy); |
3587 | for (unsigned i = 0; i != RVLocs.size(); ++i) { |
3588 | CCValAssign &VA = RVLocs[i]; |
3589 | EVT CopyVT = VA.getValVT(); |
3590 | unsigned CopyReg = ResultReg + i; |
3591 | Register SrcReg = VA.getLocReg(); |
3592 | |
3593 | // If this is x86-64, and we disabled SSE, we can't return FP values |
3594 | if ((CopyVT == MVT::f32 || CopyVT == MVT::f64) && |
3595 | ((Is64Bit || Ins[i].Flags.isInReg()) && !Subtarget->hasSSE1())) { |
3596 | report_fatal_error(reason: "SSE register return with SSE disabled" ); |
3597 | } |
3598 | |
3599 | // If we prefer to use the value in xmm registers, copy it out as f80 and |
3600 | // use a truncate to move it from fp stack reg to xmm reg. |
3601 | if ((SrcReg == X86::FP0 || SrcReg == X86::FP1) && |
3602 | isScalarFPTypeInSSEReg(VT: VA.getValVT())) { |
3603 | CopyVT = MVT::f80; |
3604 | CopyReg = createResultReg(RC: &X86::RFP80RegClass); |
3605 | } |
3606 | |
3607 | // Copy out the result. |
3608 | BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, |
3609 | MCID: TII.get(Opcode: TargetOpcode::COPY), DestReg: CopyReg).addReg(RegNo: SrcReg); |
3610 | InRegs.push_back(Elt: VA.getLocReg()); |
3611 | |
3612 | // Round the f80 to the right size, which also moves it to the appropriate |
3613 | // xmm register. This is accomplished by storing the f80 value in memory |
3614 | // and then loading it back. |
3615 | if (CopyVT != VA.getValVT()) { |
3616 | EVT ResVT = VA.getValVT(); |
3617 | unsigned Opc = ResVT == MVT::f32 ? X86::ST_Fp80m32 : X86::ST_Fp80m64; |
3618 | unsigned MemSize = ResVT.getSizeInBits()/8; |
3619 | int FI = MFI.CreateStackObject(Size: MemSize, Alignment: Align(MemSize), isSpillSlot: false); |
3620 | addFrameReference(MIB: BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, |
3621 | MCID: TII.get(Opcode: Opc)), FI) |
3622 | .addReg(RegNo: CopyReg); |
3623 | Opc = ResVT == MVT::f32 ? X86::MOVSSrm_alt : X86::MOVSDrm_alt; |
3624 | addFrameReference(MIB: BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, |
3625 | MCID: TII.get(Opcode: Opc), DestReg: ResultReg + i), FI); |
3626 | } |
3627 | } |
3628 | |
3629 | CLI.ResultReg = ResultReg; |
3630 | CLI.NumResultRegs = RVLocs.size(); |
3631 | CLI.Call = MIB; |
3632 | |
3633 | return true; |
3634 | } |
3635 | |
3636 | bool |
3637 | X86FastISel::fastSelectInstruction(const Instruction *I) { |
3638 | switch (I->getOpcode()) { |
3639 | default: break; |
3640 | case Instruction::Load: |
3641 | return X86SelectLoad(I); |
3642 | case Instruction::Store: |
3643 | return X86SelectStore(I); |
3644 | case Instruction::Ret: |
3645 | return X86SelectRet(I); |
3646 | case Instruction::ICmp: |
3647 | case Instruction::FCmp: |
3648 | return X86SelectCmp(I); |
3649 | case Instruction::ZExt: |
3650 | return X86SelectZExt(I); |
3651 | case Instruction::SExt: |
3652 | return X86SelectSExt(I); |
3653 | case Instruction::Br: |
3654 | return X86SelectBranch(I); |
3655 | case Instruction::LShr: |
3656 | case Instruction::AShr: |
3657 | case Instruction::Shl: |
3658 | return X86SelectShift(I); |
3659 | case Instruction::SDiv: |
3660 | case Instruction::UDiv: |
3661 | case Instruction::SRem: |
3662 | case Instruction::URem: |
3663 | return X86SelectDivRem(I); |
3664 | case Instruction::Select: |
3665 | return X86SelectSelect(I); |
3666 | case Instruction::Trunc: |
3667 | return X86SelectTrunc(I); |
3668 | case Instruction::FPExt: |
3669 | return X86SelectFPExt(I); |
3670 | case Instruction::FPTrunc: |
3671 | return X86SelectFPTrunc(I); |
3672 | case Instruction::SIToFP: |
3673 | return X86SelectSIToFP(I); |
3674 | case Instruction::UIToFP: |
3675 | return X86SelectUIToFP(I); |
3676 | case Instruction::IntToPtr: // Deliberate fall-through. |
3677 | case Instruction::PtrToInt: { |
3678 | EVT SrcVT = TLI.getValueType(DL, Ty: I->getOperand(i: 0)->getType()); |
3679 | EVT DstVT = TLI.getValueType(DL, Ty: I->getType()); |
3680 | if (DstVT.bitsGT(VT: SrcVT)) |
3681 | return X86SelectZExt(I); |
3682 | if (DstVT.bitsLT(VT: SrcVT)) |
3683 | return X86SelectTrunc(I); |
3684 | Register Reg = getRegForValue(V: I->getOperand(i: 0)); |
3685 | if (Reg == 0) return false; |
3686 | updateValueMap(I, Reg); |
3687 | return true; |
3688 | } |
3689 | case Instruction::BitCast: { |
3690 | // Select SSE2/AVX bitcasts between 128/256/512 bit vector types. |
3691 | if (!Subtarget->hasSSE2()) |
3692 | return false; |
3693 | |
3694 | MVT SrcVT, DstVT; |
3695 | if (!isTypeLegal(Ty: I->getOperand(i: 0)->getType(), VT&: SrcVT) || |
3696 | !isTypeLegal(Ty: I->getType(), VT&: DstVT)) |
3697 | return false; |
3698 | |
3699 | // Only allow vectors that use xmm/ymm/zmm. |
3700 | if (!SrcVT.isVector() || !DstVT.isVector() || |
3701 | SrcVT.getVectorElementType() == MVT::i1 || |
3702 | DstVT.getVectorElementType() == MVT::i1) |
3703 | return false; |
3704 | |
3705 | Register Reg = getRegForValue(V: I->getOperand(i: 0)); |
3706 | if (!Reg) |
3707 | return false; |
3708 | |
3709 | // Emit a reg-reg copy so we don't propagate cached known bits information |
3710 | // with the wrong VT if we fall out of fast isel after selecting this. |
3711 | const TargetRegisterClass *DstClass = TLI.getRegClassFor(VT: DstVT); |
3712 | Register ResultReg = createResultReg(RC: DstClass); |
3713 | BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, |
3714 | MCID: TII.get(Opcode: TargetOpcode::COPY), DestReg: ResultReg).addReg(RegNo: Reg); |
3715 | |
3716 | updateValueMap(I, Reg: ResultReg); |
3717 | return true; |
3718 | } |
3719 | } |
3720 | |
3721 | return false; |
3722 | } |
3723 | |
3724 | unsigned X86FastISel::X86MaterializeInt(const ConstantInt *CI, MVT VT) { |
3725 | if (VT > MVT::i64) |
3726 | return 0; |
3727 | |
3728 | uint64_t Imm = CI->getZExtValue(); |
3729 | if (Imm == 0) { |
3730 | Register SrcReg = fastEmitInst_(MachineInstOpcode: X86::MOV32r0, RC: &X86::GR32RegClass); |
3731 | switch (VT.SimpleTy) { |
3732 | default: llvm_unreachable("Unexpected value type" ); |
3733 | case MVT::i1: |
3734 | case MVT::i8: |
3735 | return fastEmitInst_extractsubreg(RetVT: MVT::i8, Op0: SrcReg, Idx: X86::sub_8bit); |
3736 | case MVT::i16: |
3737 | return fastEmitInst_extractsubreg(RetVT: MVT::i16, Op0: SrcReg, Idx: X86::sub_16bit); |
3738 | case MVT::i32: |
3739 | return SrcReg; |
3740 | case MVT::i64: { |
3741 | Register ResultReg = createResultReg(RC: &X86::GR64RegClass); |
3742 | BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, |
3743 | MCID: TII.get(Opcode: TargetOpcode::SUBREG_TO_REG), DestReg: ResultReg) |
3744 | .addImm(Val: 0).addReg(RegNo: SrcReg).addImm(Val: X86::sub_32bit); |
3745 | return ResultReg; |
3746 | } |
3747 | } |
3748 | } |
3749 | |
3750 | unsigned Opc = 0; |
3751 | switch (VT.SimpleTy) { |
3752 | default: llvm_unreachable("Unexpected value type" ); |
3753 | case MVT::i1: |
3754 | VT = MVT::i8; |
3755 | [[fallthrough]]; |
3756 | case MVT::i8: Opc = X86::MOV8ri; break; |
3757 | case MVT::i16: Opc = X86::MOV16ri; break; |
3758 | case MVT::i32: Opc = X86::MOV32ri; break; |
3759 | case MVT::i64: { |
3760 | if (isUInt<32>(x: Imm)) |
3761 | Opc = X86::MOV32ri64; |
3762 | else if (isInt<32>(x: Imm)) |
3763 | Opc = X86::MOV64ri32; |
3764 | else |
3765 | Opc = X86::MOV64ri; |
3766 | break; |
3767 | } |
3768 | } |
3769 | return fastEmitInst_i(MachineInstOpcode: Opc, RC: TLI.getRegClassFor(VT), Imm); |
3770 | } |
3771 | |
3772 | unsigned X86FastISel::X86MaterializeFP(const ConstantFP *CFP, MVT VT) { |
3773 | if (CFP->isNullValue()) |
3774 | return fastMaterializeFloatZero(CF: CFP); |
3775 | |
3776 | // Can't handle alternate code models yet. |
3777 | CodeModel::Model CM = TM.getCodeModel(); |
3778 | if (CM != CodeModel::Small && CM != CodeModel::Medium && |
3779 | CM != CodeModel::Large) |
3780 | return 0; |
3781 | |
3782 | // Get opcode and regclass of the output for the given load instruction. |
3783 | unsigned Opc = 0; |
3784 | bool HasSSE1 = Subtarget->hasSSE1(); |
3785 | bool HasSSE2 = Subtarget->hasSSE2(); |
3786 | bool HasAVX = Subtarget->hasAVX(); |
3787 | bool HasAVX512 = Subtarget->hasAVX512(); |
3788 | switch (VT.SimpleTy) { |
3789 | default: return 0; |
3790 | case MVT::f32: |
3791 | Opc = HasAVX512 ? X86::VMOVSSZrm_alt |
3792 | : HasAVX ? X86::VMOVSSrm_alt |
3793 | : HasSSE1 ? X86::MOVSSrm_alt |
3794 | : X86::LD_Fp32m; |
3795 | break; |
3796 | case MVT::f64: |
3797 | Opc = HasAVX512 ? X86::VMOVSDZrm_alt |
3798 | : HasAVX ? X86::VMOVSDrm_alt |
3799 | : HasSSE2 ? X86::MOVSDrm_alt |
3800 | : X86::LD_Fp64m; |
3801 | break; |
3802 | case MVT::f80: |
3803 | // No f80 support yet. |
3804 | return 0; |
3805 | } |
3806 | |
3807 | // MachineConstantPool wants an explicit alignment. |
3808 | Align Alignment = DL.getPrefTypeAlign(Ty: CFP->getType()); |
3809 | |
3810 | // x86-32 PIC requires a PIC base register for constant pools. |
3811 | unsigned PICBase = 0; |
3812 | unsigned char OpFlag = Subtarget->classifyLocalReference(GV: nullptr); |
3813 | if (OpFlag == X86II::MO_PIC_BASE_OFFSET) |
3814 | PICBase = getInstrInfo()->getGlobalBaseReg(MF: FuncInfo.MF); |
3815 | else if (OpFlag == X86II::MO_GOTOFF) |
3816 | PICBase = getInstrInfo()->getGlobalBaseReg(MF: FuncInfo.MF); |
3817 | else if (Subtarget->is64Bit() && TM.getCodeModel() != CodeModel::Large) |
3818 | PICBase = X86::RIP; |
3819 | |
3820 | // Create the load from the constant pool. |
3821 | unsigned CPI = MCP.getConstantPoolIndex(C: CFP, Alignment); |
3822 | Register ResultReg = createResultReg(RC: TLI.getRegClassFor(VT: VT.SimpleTy)); |
3823 | |
3824 | // Large code model only applies to 64-bit mode. |
3825 | if (Subtarget->is64Bit() && CM == CodeModel::Large) { |
3826 | Register AddrReg = createResultReg(RC: &X86::GR64RegClass); |
3827 | BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, MCID: TII.get(Opcode: X86::MOV64ri), |
3828 | DestReg: AddrReg) |
3829 | .addConstantPoolIndex(Idx: CPI, Offset: 0, TargetFlags: OpFlag); |
3830 | MachineInstrBuilder MIB = BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, |
3831 | MCID: TII.get(Opcode: Opc), DestReg: ResultReg); |
3832 | addRegReg(MIB, Reg1: AddrReg, isKill1: false, Reg2: PICBase, isKill2: false); |
3833 | MachineMemOperand *MMO = FuncInfo.MF->getMachineMemOperand( |
3834 | PtrInfo: MachinePointerInfo::getConstantPool(MF&: *FuncInfo.MF), |
3835 | F: MachineMemOperand::MOLoad, Size: DL.getPointerSize(), BaseAlignment: Alignment); |
3836 | MIB->addMemOperand(MF&: *FuncInfo.MF, MO: MMO); |
3837 | return ResultReg; |
3838 | } |
3839 | |
3840 | addConstantPoolReference(MIB: BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, |
3841 | MCID: TII.get(Opcode: Opc), DestReg: ResultReg), |
3842 | CPI, GlobalBaseReg: PICBase, OpFlags: OpFlag); |
3843 | return ResultReg; |
3844 | } |
3845 | |
3846 | unsigned X86FastISel::X86MaterializeGV(const GlobalValue *GV, MVT VT) { |
3847 | // Can't handle large GlobalValues yet. |
3848 | if (TM.getCodeModel() != CodeModel::Small && |
3849 | TM.getCodeModel() != CodeModel::Medium) |
3850 | return 0; |
3851 | if (TM.isLargeGlobalValue(GV)) |
3852 | return 0; |
3853 | |
3854 | // Materialize addresses with LEA/MOV instructions. |
3855 | X86AddressMode AM; |
3856 | if (X86SelectAddress(V: GV, AM)) { |
3857 | // If the expression is just a basereg, then we're done, otherwise we need |
3858 | // to emit an LEA. |
3859 | if (AM.BaseType == X86AddressMode::RegBase && |
3860 | AM.IndexReg == 0 && AM.Disp == 0 && AM.GV == nullptr) |
3861 | return AM.Base.Reg; |
3862 | |
3863 | Register ResultReg = createResultReg(RC: TLI.getRegClassFor(VT)); |
3864 | if (TM.getRelocationModel() == Reloc::Static && |
3865 | TLI.getPointerTy(DL) == MVT::i64) { |
3866 | // The displacement code could be more than 32 bits away so we need to use |
3867 | // an instruction with a 64 bit immediate |
3868 | BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, MCID: TII.get(Opcode: X86::MOV64ri), |
3869 | DestReg: ResultReg) |
3870 | .addGlobalAddress(GV); |
3871 | } else { |
3872 | unsigned Opc = |
3873 | TLI.getPointerTy(DL) == MVT::i32 |
3874 | ? (Subtarget->isTarget64BitILP32() ? X86::LEA64_32r : X86::LEA32r) |
3875 | : X86::LEA64r; |
3876 | addFullAddress(MIB: BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, |
3877 | MCID: TII.get(Opcode: Opc), DestReg: ResultReg), AM); |
3878 | } |
3879 | return ResultReg; |
3880 | } |
3881 | return 0; |
3882 | } |
3883 | |
3884 | unsigned X86FastISel::fastMaterializeConstant(const Constant *C) { |
3885 | EVT CEVT = TLI.getValueType(DL, Ty: C->getType(), AllowUnknown: true); |
3886 | |
3887 | // Only handle simple types. |
3888 | if (!CEVT.isSimple()) |
3889 | return 0; |
3890 | MVT VT = CEVT.getSimpleVT(); |
3891 | |
3892 | if (const auto *CI = dyn_cast<ConstantInt>(Val: C)) |
3893 | return X86MaterializeInt(CI, VT); |
3894 | if (const auto *CFP = dyn_cast<ConstantFP>(Val: C)) |
3895 | return X86MaterializeFP(CFP, VT); |
3896 | if (const auto *GV = dyn_cast<GlobalValue>(Val: C)) |
3897 | return X86MaterializeGV(GV, VT); |
3898 | if (isa<UndefValue>(Val: C)) { |
3899 | unsigned Opc = 0; |
3900 | switch (VT.SimpleTy) { |
3901 | default: |
3902 | break; |
3903 | case MVT::f32: |
3904 | if (!Subtarget->hasSSE1()) |
3905 | Opc = X86::LD_Fp032; |
3906 | break; |
3907 | case MVT::f64: |
3908 | if (!Subtarget->hasSSE2()) |
3909 | Opc = X86::LD_Fp064; |
3910 | break; |
3911 | case MVT::f80: |
3912 | Opc = X86::LD_Fp080; |
3913 | break; |
3914 | } |
3915 | |
3916 | if (Opc) { |
3917 | Register ResultReg = createResultReg(RC: TLI.getRegClassFor(VT)); |
3918 | BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, MCID: TII.get(Opcode: Opc), |
3919 | DestReg: ResultReg); |
3920 | return ResultReg; |
3921 | } |
3922 | } |
3923 | |
3924 | return 0; |
3925 | } |
3926 | |
3927 | unsigned X86FastISel::fastMaterializeAlloca(const AllocaInst *C) { |
3928 | // Fail on dynamic allocas. At this point, getRegForValue has already |
3929 | // checked its CSE maps, so if we're here trying to handle a dynamic |
3930 | // alloca, we're not going to succeed. X86SelectAddress has a |
3931 | // check for dynamic allocas, because it's called directly from |
3932 | // various places, but targetMaterializeAlloca also needs a check |
3933 | // in order to avoid recursion between getRegForValue, |
3934 | // X86SelectAddrss, and targetMaterializeAlloca. |
3935 | if (!FuncInfo.StaticAllocaMap.count(Val: C)) |
3936 | return 0; |
3937 | assert(C->isStaticAlloca() && "dynamic alloca in the static alloca map?" ); |
3938 | |
3939 | X86AddressMode AM; |
3940 | if (!X86SelectAddress(V: C, AM)) |
3941 | return 0; |
3942 | unsigned Opc = |
3943 | TLI.getPointerTy(DL) == MVT::i32 |
3944 | ? (Subtarget->isTarget64BitILP32() ? X86::LEA64_32r : X86::LEA32r) |
3945 | : X86::LEA64r; |
3946 | const TargetRegisterClass *RC = TLI.getRegClassFor(VT: TLI.getPointerTy(DL)); |
3947 | Register ResultReg = createResultReg(RC); |
3948 | addFullAddress(MIB: BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, |
3949 | MCID: TII.get(Opcode: Opc), DestReg: ResultReg), AM); |
3950 | return ResultReg; |
3951 | } |
3952 | |
3953 | unsigned X86FastISel::fastMaterializeFloatZero(const ConstantFP *CF) { |
3954 | MVT VT; |
3955 | if (!isTypeLegal(Ty: CF->getType(), VT)) |
3956 | return 0; |
3957 | |
3958 | // Get opcode and regclass for the given zero. |
3959 | bool HasSSE1 = Subtarget->hasSSE1(); |
3960 | bool HasSSE2 = Subtarget->hasSSE2(); |
3961 | bool HasAVX512 = Subtarget->hasAVX512(); |
3962 | unsigned Opc = 0; |
3963 | switch (VT.SimpleTy) { |
3964 | default: return 0; |
3965 | case MVT::f16: |
3966 | Opc = HasAVX512 ? X86::AVX512_FsFLD0SH : X86::FsFLD0SH; |
3967 | break; |
3968 | case MVT::f32: |
3969 | Opc = HasAVX512 ? X86::AVX512_FsFLD0SS |
3970 | : HasSSE1 ? X86::FsFLD0SS |
3971 | : X86::LD_Fp032; |
3972 | break; |
3973 | case MVT::f64: |
3974 | Opc = HasAVX512 ? X86::AVX512_FsFLD0SD |
3975 | : HasSSE2 ? X86::FsFLD0SD |
3976 | : X86::LD_Fp064; |
3977 | break; |
3978 | case MVT::f80: |
3979 | // No f80 support yet. |
3980 | return 0; |
3981 | } |
3982 | |
3983 | Register ResultReg = createResultReg(RC: TLI.getRegClassFor(VT)); |
3984 | BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, MCID: TII.get(Opcode: Opc), DestReg: ResultReg); |
3985 | return ResultReg; |
3986 | } |
3987 | |
3988 | |
3989 | bool X86FastISel::tryToFoldLoadIntoMI(MachineInstr *MI, unsigned OpNo, |
3990 | const LoadInst *LI) { |
3991 | const Value *Ptr = LI->getPointerOperand(); |
3992 | X86AddressMode AM; |
3993 | if (!X86SelectAddress(V: Ptr, AM)) |
3994 | return false; |
3995 | |
3996 | const X86InstrInfo &XII = (const X86InstrInfo &)TII; |
3997 | |
3998 | unsigned Size = DL.getTypeAllocSize(Ty: LI->getType()); |
3999 | |
4000 | SmallVector<MachineOperand, 8> AddrOps; |
4001 | AM.getFullAddress(MO&: AddrOps); |
4002 | |
4003 | MachineInstr *Result = XII.foldMemoryOperandImpl( |
4004 | MF&: *FuncInfo.MF, MI&: *MI, OpNum: OpNo, MOs: AddrOps, InsertPt: FuncInfo.InsertPt, Size, Alignment: LI->getAlign(), |
4005 | /*AllowCommute=*/true); |
4006 | if (!Result) |
4007 | return false; |
4008 | |
4009 | // The index register could be in the wrong register class. Unfortunately, |
4010 | // foldMemoryOperandImpl could have commuted the instruction so its not enough |
4011 | // to just look at OpNo + the offset to the index reg. We actually need to |
4012 | // scan the instruction to find the index reg and see if its the correct reg |
4013 | // class. |
4014 | unsigned OperandNo = 0; |
4015 | for (MachineInstr::mop_iterator I = Result->operands_begin(), |
4016 | E = Result->operands_end(); I != E; ++I, ++OperandNo) { |
4017 | MachineOperand &MO = *I; |
4018 | if (!MO.isReg() || MO.isDef() || MO.getReg() != AM.IndexReg) |
4019 | continue; |
4020 | // Found the index reg, now try to rewrite it. |
4021 | Register IndexReg = constrainOperandRegClass(II: Result->getDesc(), |
4022 | Op: MO.getReg(), OpNum: OperandNo); |
4023 | if (IndexReg == MO.getReg()) |
4024 | continue; |
4025 | MO.setReg(IndexReg); |
4026 | } |
4027 | |
4028 | Result->addMemOperand(MF&: *FuncInfo.MF, MO: createMachineMemOperandFor(I: LI)); |
4029 | Result->cloneInstrSymbols(MF&: *FuncInfo.MF, MI: *MI); |
4030 | MachineBasicBlock::iterator I(MI); |
4031 | removeDeadCode(I, E: std::next(x: I)); |
4032 | return true; |
4033 | } |
4034 | |
4035 | unsigned X86FastISel::fastEmitInst_rrrr(unsigned MachineInstOpcode, |
4036 | const TargetRegisterClass *RC, |
4037 | unsigned Op0, unsigned Op1, |
4038 | unsigned Op2, unsigned Op3) { |
4039 | const MCInstrDesc &II = TII.get(Opcode: MachineInstOpcode); |
4040 | |
4041 | Register ResultReg = createResultReg(RC); |
4042 | Op0 = constrainOperandRegClass(II, Op: Op0, OpNum: II.getNumDefs()); |
4043 | Op1 = constrainOperandRegClass(II, Op: Op1, OpNum: II.getNumDefs() + 1); |
4044 | Op2 = constrainOperandRegClass(II, Op: Op2, OpNum: II.getNumDefs() + 2); |
4045 | Op3 = constrainOperandRegClass(II, Op: Op3, OpNum: II.getNumDefs() + 3); |
4046 | |
4047 | if (II.getNumDefs() >= 1) |
4048 | BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, MCID: II, DestReg: ResultReg) |
4049 | .addReg(RegNo: Op0) |
4050 | .addReg(RegNo: Op1) |
4051 | .addReg(RegNo: Op2) |
4052 | .addReg(RegNo: Op3); |
4053 | else { |
4054 | BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, MCID: II) |
4055 | .addReg(RegNo: Op0) |
4056 | .addReg(RegNo: Op1) |
4057 | .addReg(RegNo: Op2) |
4058 | .addReg(RegNo: Op3); |
4059 | BuildMI(BB&: *FuncInfo.MBB, I: FuncInfo.InsertPt, MIMD, MCID: TII.get(Opcode: TargetOpcode::COPY), |
4060 | DestReg: ResultReg) |
4061 | .addReg(RegNo: II.implicit_defs()[0]); |
4062 | } |
4063 | return ResultReg; |
4064 | } |
4065 | |
4066 | |
4067 | namespace llvm { |
4068 | FastISel *X86::createFastISel(FunctionLoweringInfo &funcInfo, |
4069 | const TargetLibraryInfo *libInfo) { |
4070 | return new X86FastISel(funcInfo, libInfo); |
4071 | } |
4072 | } |
4073 | |