1 //===- FastISel.cpp - Implementation of the FastISel class ----------------===//
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 contains the implementation of the FastISel class.
10 //
11 // "Fast" instruction selection is designed to emit very poor code quickly.
12 // Also, it is not designed to be able to do much lowering, so most illegal
13 // types (e.g. i64 on 32-bit targets) and operations are not supported.  It is
14 // also not intended to be able to do much optimization, except in a few cases
15 // where doing optimizations reduces overall compile time.  For example, folding
16 // constants into immediate fields is often done, because it's cheap and it
17 // reduces the number of instructions later phases have to examine.
18 //
19 // "Fast" instruction selection is able to fail gracefully and transfer
20 // control to the SelectionDAG selector for operations that it doesn't
21 // support.  In many cases, this allows us to avoid duplicating a lot of
22 // the complicated lowering logic that SelectionDAG currently has.
23 //
24 // The intended use for "fast" instruction selection is "-O0" mode
25 // compilation, where the quality of the generated code is irrelevant when
26 // weighed against the speed at which the code can be generated.  Also,
27 // at -O0, the LLVM optimizers are not running, and this makes the
28 // compile time of codegen a much higher portion of the overall compile
29 // time.  Despite its limitations, "fast" instruction selection is able to
30 // handle enough code on its own to provide noticeable overall speedups
31 // in -O0 compiles.
32 //
33 // Basic operations are supported in a target-independent way, by reading
34 // the same instruction descriptions that the SelectionDAG selector reads,
35 // and identifying simple arithmetic operations that can be directly selected
36 // from simple operators.  More complicated operations currently require
37 // target-specific code.
38 //
39 //===----------------------------------------------------------------------===//
40 
41 #include "llvm/CodeGen/FastISel.h"
42 #include "llvm/ADT/APFloat.h"
43 #include "llvm/ADT/APSInt.h"
44 #include "llvm/ADT/DenseMap.h"
45 #include "llvm/ADT/Optional.h"
46 #include "llvm/ADT/SmallPtrSet.h"
47 #include "llvm/ADT/SmallString.h"
48 #include "llvm/ADT/SmallVector.h"
49 #include "llvm/ADT/Statistic.h"
50 #include "llvm/Analysis/BranchProbabilityInfo.h"
51 #include "llvm/Analysis/TargetLibraryInfo.h"
52 #include "llvm/CodeGen/Analysis.h"
53 #include "llvm/CodeGen/FunctionLoweringInfo.h"
54 #include "llvm/CodeGen/ISDOpcodes.h"
55 #include "llvm/CodeGen/MachineBasicBlock.h"
56 #include "llvm/CodeGen/MachineFrameInfo.h"
57 #include "llvm/CodeGen/MachineInstr.h"
58 #include "llvm/CodeGen/MachineInstrBuilder.h"
59 #include "llvm/CodeGen/MachineMemOperand.h"
60 #include "llvm/CodeGen/MachineModuleInfo.h"
61 #include "llvm/CodeGen/MachineOperand.h"
62 #include "llvm/CodeGen/MachineRegisterInfo.h"
63 #include "llvm/CodeGen/StackMaps.h"
64 #include "llvm/CodeGen/TargetInstrInfo.h"
65 #include "llvm/CodeGen/TargetLowering.h"
66 #include "llvm/CodeGen/TargetSubtargetInfo.h"
67 #include "llvm/CodeGen/ValueTypes.h"
68 #include "llvm/IR/Argument.h"
69 #include "llvm/IR/Attributes.h"
70 #include "llvm/IR/BasicBlock.h"
71 #include "llvm/IR/CallingConv.h"
72 #include "llvm/IR/Constant.h"
73 #include "llvm/IR/Constants.h"
74 #include "llvm/IR/DataLayout.h"
75 #include "llvm/IR/DebugInfo.h"
76 #include "llvm/IR/DebugLoc.h"
77 #include "llvm/IR/DerivedTypes.h"
78 #include "llvm/IR/Function.h"
79 #include "llvm/IR/GetElementPtrTypeIterator.h"
80 #include "llvm/IR/GlobalValue.h"
81 #include "llvm/IR/InlineAsm.h"
82 #include "llvm/IR/InstrTypes.h"
83 #include "llvm/IR/Instruction.h"
84 #include "llvm/IR/Instructions.h"
85 #include "llvm/IR/IntrinsicInst.h"
86 #include "llvm/IR/LLVMContext.h"
87 #include "llvm/IR/Mangler.h"
88 #include "llvm/IR/Metadata.h"
89 #include "llvm/IR/Operator.h"
90 #include "llvm/IR/PatternMatch.h"
91 #include "llvm/IR/Type.h"
92 #include "llvm/IR/User.h"
93 #include "llvm/IR/Value.h"
94 #include "llvm/MC/MCContext.h"
95 #include "llvm/MC/MCInstrDesc.h"
96 #include "llvm/MC/MCRegisterInfo.h"
97 #include "llvm/Support/Casting.h"
98 #include "llvm/Support/Debug.h"
99 #include "llvm/Support/ErrorHandling.h"
100 #include "llvm/Support/MachineValueType.h"
101 #include "llvm/Support/MathExtras.h"
102 #include "llvm/Support/raw_ostream.h"
103 #include "llvm/Target/TargetMachine.h"
104 #include "llvm/Target/TargetOptions.h"
105 #include <algorithm>
106 #include <cassert>
107 #include <cstdint>
108 #include <iterator>
109 #include <utility>
110 
111 using namespace llvm;
112 using namespace PatternMatch;
113 
114 #define DEBUG_TYPE "isel"
115 
116 // FIXME: Remove this after the feature has proven reliable.
117 static cl::opt<bool> SinkLocalValues("fast-isel-sink-local-values",
118                                      cl::init(true), cl::Hidden,
119                                      cl::desc("Sink local values in FastISel"));
120 
121 STATISTIC(NumFastIselSuccessIndependent, "Number of insts selected by "
122                                          "target-independent selector");
123 STATISTIC(NumFastIselSuccessTarget, "Number of insts selected by "
124                                     "target-specific selector");
125 STATISTIC(NumFastIselDead, "Number of dead insts removed on failure");
126 
127 /// Set the current block to which generated machine instructions will be
128 /// appended.
startNewBlock()129 void FastISel::startNewBlock() {
130   assert(LocalValueMap.empty() &&
131          "local values should be cleared after finishing a BB");
132 
133   // Instructions are appended to FuncInfo.MBB. If the basic block already
134   // contains labels or copies, use the last instruction as the last local
135   // value.
136   EmitStartPt = nullptr;
137   if (!FuncInfo.MBB->empty())
138     EmitStartPt = &FuncInfo.MBB->back();
139   LastLocalValue = EmitStartPt;
140 }
141 
142 /// Flush the local CSE map and sink anything we can.
finishBasicBlock()143 void FastISel::finishBasicBlock() { flushLocalValueMap(); }
144 
lowerArguments()145 bool FastISel::lowerArguments() {
146   if (!FuncInfo.CanLowerReturn)
147     // Fallback to SDISel argument lowering code to deal with sret pointer
148     // parameter.
149     return false;
150 
151   if (!fastLowerArguments())
152     return false;
153 
154   // Enter arguments into ValueMap for uses in non-entry BBs.
155   for (Function::const_arg_iterator I = FuncInfo.Fn->arg_begin(),
156                                     E = FuncInfo.Fn->arg_end();
157        I != E; ++I) {
158     DenseMap<const Value *, Register>::iterator VI = LocalValueMap.find(&*I);
159     assert(VI != LocalValueMap.end() && "Missed an argument?");
160     FuncInfo.ValueMap[&*I] = VI->second;
161   }
162   return true;
163 }
164 
165 /// Return the defined register if this instruction defines exactly one
166 /// virtual register and uses no other virtual registers. Otherwise return 0.
findSinkableLocalRegDef(MachineInstr & MI)167 static Register findSinkableLocalRegDef(MachineInstr &MI) {
168   Register RegDef;
169   for (const MachineOperand &MO : MI.operands()) {
170     if (!MO.isReg())
171       continue;
172     if (MO.isDef()) {
173       if (RegDef)
174         return 0;
175       RegDef = MO.getReg();
176     } else if (MO.getReg().isVirtual()) {
177       // This is another use of a vreg. Don't try to sink it.
178       return Register();
179     }
180   }
181   return RegDef;
182 }
183 
flushLocalValueMap()184 void FastISel::flushLocalValueMap() {
185   // Try to sink local values down to their first use so that we can give them a
186   // better debug location. This has the side effect of shrinking local value
187   // live ranges, which helps out fast regalloc.
188   if (SinkLocalValues && LastLocalValue != EmitStartPt) {
189     // Sink local value materialization instructions between EmitStartPt and
190     // LastLocalValue. Visit them bottom-up, starting from LastLocalValue, to
191     // avoid inserting into the range that we're iterating over.
192     MachineBasicBlock::reverse_iterator RE =
193         EmitStartPt ? MachineBasicBlock::reverse_iterator(EmitStartPt)
194                     : FuncInfo.MBB->rend();
195     MachineBasicBlock::reverse_iterator RI(LastLocalValue);
196 
197     InstOrderMap OrderMap;
198     for (; RI != RE;) {
199       MachineInstr &LocalMI = *RI;
200       ++RI;
201       bool Store = true;
202       if (!LocalMI.isSafeToMove(nullptr, Store))
203         continue;
204       Register DefReg = findSinkableLocalRegDef(LocalMI);
205       if (DefReg == 0)
206         continue;
207 
208       sinkLocalValueMaterialization(LocalMI, DefReg, OrderMap);
209     }
210   }
211 
212   LocalValueMap.clear();
213   LastLocalValue = EmitStartPt;
214   recomputeInsertPt();
215   SavedInsertPt = FuncInfo.InsertPt;
216   LastFlushPoint = FuncInfo.InsertPt;
217 }
218 
isRegUsedByPhiNodes(Register DefReg,FunctionLoweringInfo & FuncInfo)219 static bool isRegUsedByPhiNodes(Register DefReg,
220                                 FunctionLoweringInfo &FuncInfo) {
221   for (auto &P : FuncInfo.PHINodesToUpdate)
222     if (P.second == DefReg)
223       return true;
224   return false;
225 }
226 
isTerminatingEHLabel(MachineBasicBlock * MBB,MachineInstr & MI)227 static bool isTerminatingEHLabel(MachineBasicBlock *MBB, MachineInstr &MI) {
228   // Ignore non-EH labels.
229   if (!MI.isEHLabel())
230     return false;
231 
232   // Any EH label outside a landing pad must be for an invoke. Consider it a
233   // terminator.
234   if (!MBB->isEHPad())
235     return true;
236 
237   // If this is a landingpad, the first non-phi instruction will be an EH_LABEL.
238   // Don't consider that label to be a terminator.
239   return MI.getIterator() != MBB->getFirstNonPHI();
240 }
241 
242 /// Build a map of instruction orders. Return the first terminator and its
243 /// order. Consider EH_LABEL instructions to be terminators as well, since local
244 /// values for phis after invokes must be materialized before the call.
initialize(MachineBasicBlock * MBB,MachineBasicBlock::iterator LastFlushPoint)245 void FastISel::InstOrderMap::initialize(
246     MachineBasicBlock *MBB, MachineBasicBlock::iterator LastFlushPoint) {
247   unsigned Order = 0;
248   for (MachineInstr &I : *MBB) {
249     if (!FirstTerminator &&
250         (I.isTerminator() || isTerminatingEHLabel(MBB, I))) {
251       FirstTerminator = &I;
252       FirstTerminatorOrder = Order;
253     }
254     Orders[&I] = Order++;
255 
256     // We don't need to order instructions past the last flush point.
257     if (I.getIterator() == LastFlushPoint)
258       break;
259   }
260 }
261 
sinkLocalValueMaterialization(MachineInstr & LocalMI,Register DefReg,InstOrderMap & OrderMap)262 void FastISel::sinkLocalValueMaterialization(MachineInstr &LocalMI,
263                                              Register DefReg,
264                                              InstOrderMap &OrderMap) {
265   // If this register is used by a register fixup, MRI will not contain all
266   // the uses until after register fixups, so don't attempt to sink or DCE
267   // this instruction. Register fixups typically come from no-op cast
268   // instructions, which replace the cast instruction vreg with the local
269   // value vreg.
270   if (FuncInfo.RegsWithFixups.count(DefReg))
271     return;
272 
273   // We can DCE this instruction if there are no uses and it wasn't a
274   // materialized for a successor PHI node.
275   bool UsedByPHI = isRegUsedByPhiNodes(DefReg, FuncInfo);
276   if (!UsedByPHI && MRI.use_nodbg_empty(DefReg)) {
277     if (EmitStartPt == &LocalMI)
278       EmitStartPt = EmitStartPt->getPrevNode();
279     LLVM_DEBUG(dbgs() << "removing dead local value materialization "
280                       << LocalMI);
281     OrderMap.Orders.erase(&LocalMI);
282     LocalMI.eraseFromParent();
283     return;
284   }
285 
286   // Number the instructions if we haven't yet so we can efficiently find the
287   // earliest use.
288   if (OrderMap.Orders.empty())
289     OrderMap.initialize(FuncInfo.MBB, LastFlushPoint);
290 
291   // Find the first user in the BB.
292   MachineInstr *FirstUser = nullptr;
293   unsigned FirstOrder = std::numeric_limits<unsigned>::max();
294   for (MachineInstr &UseInst : MRI.use_nodbg_instructions(DefReg)) {
295     auto I = OrderMap.Orders.find(&UseInst);
296     assert(I != OrderMap.Orders.end() &&
297            "local value used by instruction outside local region");
298     unsigned UseOrder = I->second;
299     if (UseOrder < FirstOrder) {
300       FirstOrder = UseOrder;
301       FirstUser = &UseInst;
302     }
303   }
304 
305   // The insertion point will be the first terminator or the first user,
306   // whichever came first. If there was no terminator, this must be a
307   // fallthrough block and the insertion point is the end of the block.
308   MachineBasicBlock::instr_iterator SinkPos;
309   if (UsedByPHI && OrderMap.FirstTerminatorOrder < FirstOrder) {
310     FirstOrder = OrderMap.FirstTerminatorOrder;
311     SinkPos = OrderMap.FirstTerminator->getIterator();
312   } else if (FirstUser) {
313     SinkPos = FirstUser->getIterator();
314   } else {
315     assert(UsedByPHI && "must be users if not used by a phi");
316     SinkPos = FuncInfo.MBB->instr_end();
317   }
318 
319   // Collect all DBG_VALUEs before the new insertion position so that we can
320   // sink them.
321   SmallVector<MachineInstr *, 1> DbgValues;
322   for (MachineInstr &DbgVal : MRI.use_instructions(DefReg)) {
323     if (!DbgVal.isDebugValue())
324       continue;
325     unsigned UseOrder = OrderMap.Orders[&DbgVal];
326     if (UseOrder < FirstOrder)
327       DbgValues.push_back(&DbgVal);
328   }
329 
330   // Sink LocalMI before SinkPos and assign it the same DebugLoc.
331   LLVM_DEBUG(dbgs() << "sinking local value to first use " << LocalMI);
332   FuncInfo.MBB->remove(&LocalMI);
333   FuncInfo.MBB->insert(SinkPos, &LocalMI);
334   if (SinkPos != FuncInfo.MBB->end())
335     LocalMI.setDebugLoc(SinkPos->getDebugLoc());
336 
337   // Sink any debug values that we've collected.
338   for (MachineInstr *DI : DbgValues) {
339     FuncInfo.MBB->remove(DI);
340     FuncInfo.MBB->insert(SinkPos, DI);
341   }
342 }
343 
hasTrivialKill(const Value * V)344 bool FastISel::hasTrivialKill(const Value *V) {
345   // Don't consider constants or arguments to have trivial kills.
346   const Instruction *I = dyn_cast<Instruction>(V);
347   if (!I)
348     return false;
349 
350   // No-op casts are trivially coalesced by fast-isel.
351   if (const auto *Cast = dyn_cast<CastInst>(I))
352     if (Cast->isNoopCast(DL) && !hasTrivialKill(Cast->getOperand(0)))
353       return false;
354 
355   // Even the value might have only one use in the LLVM IR, it is possible that
356   // FastISel might fold the use into another instruction and now there is more
357   // than one use at the Machine Instruction level.
358   Register Reg = lookUpRegForValue(V);
359   if (Reg && !MRI.use_empty(Reg))
360     return false;
361 
362   // GEPs with all zero indices are trivially coalesced by fast-isel.
363   if (const auto *GEP = dyn_cast<GetElementPtrInst>(I))
364     if (GEP->hasAllZeroIndices() && !hasTrivialKill(GEP->getOperand(0)))
365       return false;
366 
367   // Only instructions with a single use in the same basic block are considered
368   // to have trivial kills.
369   return I->hasOneUse() &&
370          !(I->getOpcode() == Instruction::BitCast ||
371            I->getOpcode() == Instruction::PtrToInt ||
372            I->getOpcode() == Instruction::IntToPtr) &&
373          cast<Instruction>(*I->user_begin())->getParent() == I->getParent();
374 }
375 
getRegForValue(const Value * V)376 Register FastISel::getRegForValue(const Value *V) {
377   EVT RealVT = TLI.getValueType(DL, V->getType(), /*AllowUnknown=*/true);
378   // Don't handle non-simple values in FastISel.
379   if (!RealVT.isSimple())
380     return Register();
381 
382   // Ignore illegal types. We must do this before looking up the value
383   // in ValueMap because Arguments are given virtual registers regardless
384   // of whether FastISel can handle them.
385   MVT VT = RealVT.getSimpleVT();
386   if (!TLI.isTypeLegal(VT)) {
387     // Handle integer promotions, though, because they're common and easy.
388     if (VT == MVT::i1 || VT == MVT::i8 || VT == MVT::i16)
389       VT = TLI.getTypeToTransformTo(V->getContext(), VT).getSimpleVT();
390     else
391       return Register();
392   }
393 
394   // Look up the value to see if we already have a register for it.
395   Register Reg = lookUpRegForValue(V);
396   if (Reg)
397     return Reg;
398 
399   // In bottom-up mode, just create the virtual register which will be used
400   // to hold the value. It will be materialized later.
401   if (isa<Instruction>(V) &&
402       (!isa<AllocaInst>(V) ||
403        !FuncInfo.StaticAllocaMap.count(cast<AllocaInst>(V))))
404     return FuncInfo.InitializeRegForValue(V);
405 
406   SavePoint SaveInsertPt = enterLocalValueArea();
407 
408   // Materialize the value in a register. Emit any instructions in the
409   // local value area.
410   Reg = materializeRegForValue(V, VT);
411 
412   leaveLocalValueArea(SaveInsertPt);
413 
414   return Reg;
415 }
416 
materializeConstant(const Value * V,MVT VT)417 Register FastISel::materializeConstant(const Value *V, MVT VT) {
418   Register Reg;
419   if (const auto *CI = dyn_cast<ConstantInt>(V)) {
420     if (CI->getValue().getActiveBits() <= 64)
421       Reg = fastEmit_i(VT, VT, ISD::Constant, CI->getZExtValue());
422   } else if (isa<AllocaInst>(V))
423     Reg = fastMaterializeAlloca(cast<AllocaInst>(V));
424   else if (isa<ConstantPointerNull>(V))
425     // Translate this as an integer zero so that it can be
426     // local-CSE'd with actual integer zeros.
427     Reg =
428         getRegForValue(Constant::getNullValue(DL.getIntPtrType(V->getType())));
429   else if (const auto *CF = dyn_cast<ConstantFP>(V)) {
430     if (CF->isNullValue())
431       Reg = fastMaterializeFloatZero(CF);
432     else
433       // Try to emit the constant directly.
434       Reg = fastEmit_f(VT, VT, ISD::ConstantFP, CF);
435 
436     if (!Reg) {
437       // Try to emit the constant by using an integer constant with a cast.
438       const APFloat &Flt = CF->getValueAPF();
439       EVT IntVT = TLI.getPointerTy(DL);
440       uint32_t IntBitWidth = IntVT.getSizeInBits();
441       APSInt SIntVal(IntBitWidth, /*isUnsigned=*/false);
442       bool isExact;
443       (void)Flt.convertToInteger(SIntVal, APFloat::rmTowardZero, &isExact);
444       if (isExact) {
445         Register IntegerReg =
446             getRegForValue(ConstantInt::get(V->getContext(), SIntVal));
447         if (IntegerReg)
448           Reg = fastEmit_r(IntVT.getSimpleVT(), VT, ISD::SINT_TO_FP, IntegerReg,
449                            /*Kill=*/false);
450       }
451     }
452   } else if (const auto *Op = dyn_cast<Operator>(V)) {
453     if (!selectOperator(Op, Op->getOpcode()))
454       if (!isa<Instruction>(Op) ||
455           !fastSelectInstruction(cast<Instruction>(Op)))
456         return 0;
457     Reg = lookUpRegForValue(Op);
458   } else if (isa<UndefValue>(V)) {
459     Reg = createResultReg(TLI.getRegClassFor(VT));
460     BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
461             TII.get(TargetOpcode::IMPLICIT_DEF), Reg);
462   }
463   return Reg;
464 }
465 
466 /// Helper for getRegForValue. This function is called when the value isn't
467 /// already available in a register and must be materialized with new
468 /// instructions.
materializeRegForValue(const Value * V,MVT VT)469 Register FastISel::materializeRegForValue(const Value *V, MVT VT) {
470   Register Reg;
471   // Give the target-specific code a try first.
472   if (isa<Constant>(V))
473     Reg = fastMaterializeConstant(cast<Constant>(V));
474 
475   // If target-specific code couldn't or didn't want to handle the value, then
476   // give target-independent code a try.
477   if (!Reg)
478     Reg = materializeConstant(V, VT);
479 
480   // Don't cache constant materializations in the general ValueMap.
481   // To do so would require tracking what uses they dominate.
482   if (Reg) {
483     LocalValueMap[V] = Reg;
484     LastLocalValue = MRI.getVRegDef(Reg);
485   }
486   return Reg;
487 }
488 
lookUpRegForValue(const Value * V)489 Register FastISel::lookUpRegForValue(const Value *V) {
490   // Look up the value to see if we already have a register for it. We
491   // cache values defined by Instructions across blocks, and other values
492   // only locally. This is because Instructions already have the SSA
493   // def-dominates-use requirement enforced.
494   DenseMap<const Value *, Register>::iterator I = FuncInfo.ValueMap.find(V);
495   if (I != FuncInfo.ValueMap.end())
496     return I->second;
497   return LocalValueMap[V];
498 }
499 
updateValueMap(const Value * I,Register Reg,unsigned NumRegs)500 void FastISel::updateValueMap(const Value *I, Register Reg, unsigned NumRegs) {
501   if (!isa<Instruction>(I)) {
502     LocalValueMap[I] = Reg;
503     return;
504   }
505 
506   Register &AssignedReg = FuncInfo.ValueMap[I];
507   if (!AssignedReg)
508     // Use the new register.
509     AssignedReg = Reg;
510   else if (Reg != AssignedReg) {
511     // Arrange for uses of AssignedReg to be replaced by uses of Reg.
512     for (unsigned i = 0; i < NumRegs; i++) {
513       FuncInfo.RegFixups[AssignedReg + i] = Reg + i;
514       FuncInfo.RegsWithFixups.insert(Reg + i);
515     }
516 
517     AssignedReg = Reg;
518   }
519 }
520 
getRegForGEPIndex(const Value * Idx)521 std::pair<Register, bool> FastISel::getRegForGEPIndex(const Value *Idx) {
522   Register IdxN = getRegForValue(Idx);
523   if (!IdxN)
524     // Unhandled operand. Halt "fast" selection and bail.
525     return std::pair<Register, bool>(Register(), false);
526 
527   bool IdxNIsKill = hasTrivialKill(Idx);
528 
529   // If the index is smaller or larger than intptr_t, truncate or extend it.
530   MVT PtrVT = TLI.getPointerTy(DL);
531   EVT IdxVT = EVT::getEVT(Idx->getType(), /*HandleUnknown=*/false);
532   if (IdxVT.bitsLT(PtrVT)) {
533     IdxN = fastEmit_r(IdxVT.getSimpleVT(), PtrVT, ISD::SIGN_EXTEND, IdxN,
534                       IdxNIsKill);
535     IdxNIsKill = true;
536   } else if (IdxVT.bitsGT(PtrVT)) {
537     IdxN =
538         fastEmit_r(IdxVT.getSimpleVT(), PtrVT, ISD::TRUNCATE, IdxN, IdxNIsKill);
539     IdxNIsKill = true;
540   }
541   return std::pair<Register, bool>(IdxN, IdxNIsKill);
542 }
543 
recomputeInsertPt()544 void FastISel::recomputeInsertPt() {
545   if (getLastLocalValue()) {
546     FuncInfo.InsertPt = getLastLocalValue();
547     FuncInfo.MBB = FuncInfo.InsertPt->getParent();
548     ++FuncInfo.InsertPt;
549   } else
550     FuncInfo.InsertPt = FuncInfo.MBB->getFirstNonPHI();
551 
552   // Now skip past any EH_LABELs, which must remain at the beginning.
553   while (FuncInfo.InsertPt != FuncInfo.MBB->end() &&
554          FuncInfo.InsertPt->getOpcode() == TargetOpcode::EH_LABEL)
555     ++FuncInfo.InsertPt;
556 }
557 
removeDeadCode(MachineBasicBlock::iterator I,MachineBasicBlock::iterator E)558 void FastISel::removeDeadCode(MachineBasicBlock::iterator I,
559                               MachineBasicBlock::iterator E) {
560   assert(I.isValid() && E.isValid() && std::distance(I, E) > 0 &&
561          "Invalid iterator!");
562   while (I != E) {
563     if (LastFlushPoint == I)
564       LastFlushPoint = E;
565     if (SavedInsertPt == I)
566       SavedInsertPt = E;
567     if (EmitStartPt == I)
568       EmitStartPt = E.isValid() ? &*E : nullptr;
569     if (LastLocalValue == I)
570       LastLocalValue = E.isValid() ? &*E : nullptr;
571 
572     MachineInstr *Dead = &*I;
573     ++I;
574     Dead->eraseFromParent();
575     ++NumFastIselDead;
576   }
577   recomputeInsertPt();
578 }
579 
enterLocalValueArea()580 FastISel::SavePoint FastISel::enterLocalValueArea() {
581   MachineBasicBlock::iterator OldInsertPt = FuncInfo.InsertPt;
582   DebugLoc OldDL = DbgLoc;
583   recomputeInsertPt();
584   DbgLoc = DebugLoc();
585   SavePoint SP = {OldInsertPt, OldDL};
586   return SP;
587 }
588 
leaveLocalValueArea(SavePoint OldInsertPt)589 void FastISel::leaveLocalValueArea(SavePoint OldInsertPt) {
590   if (FuncInfo.InsertPt != FuncInfo.MBB->begin())
591     LastLocalValue = &*std::prev(FuncInfo.InsertPt);
592 
593   // Restore the previous insert position.
594   FuncInfo.InsertPt = OldInsertPt.InsertPt;
595   DbgLoc = OldInsertPt.DL;
596 }
597 
selectBinaryOp(const User * I,unsigned ISDOpcode)598 bool FastISel::selectBinaryOp(const User *I, unsigned ISDOpcode) {
599   EVT VT = EVT::getEVT(I->getType(), /*HandleUnknown=*/true);
600   if (VT == MVT::Other || !VT.isSimple())
601     // Unhandled type. Halt "fast" selection and bail.
602     return false;
603 
604   // We only handle legal types. For example, on x86-32 the instruction
605   // selector contains all of the 64-bit instructions from x86-64,
606   // under the assumption that i64 won't be used if the target doesn't
607   // support it.
608   if (!TLI.isTypeLegal(VT)) {
609     // MVT::i1 is special. Allow AND, OR, or XOR because they
610     // don't require additional zeroing, which makes them easy.
611     if (VT == MVT::i1 && (ISDOpcode == ISD::AND || ISDOpcode == ISD::OR ||
612                           ISDOpcode == ISD::XOR))
613       VT = TLI.getTypeToTransformTo(I->getContext(), VT);
614     else
615       return false;
616   }
617 
618   // Check if the first operand is a constant, and handle it as "ri".  At -O0,
619   // we don't have anything that canonicalizes operand order.
620   if (const auto *CI = dyn_cast<ConstantInt>(I->getOperand(0)))
621     if (isa<Instruction>(I) && cast<Instruction>(I)->isCommutative()) {
622       Register Op1 = getRegForValue(I->getOperand(1));
623       if (!Op1)
624         return false;
625       bool Op1IsKill = hasTrivialKill(I->getOperand(1));
626 
627       Register ResultReg =
628           fastEmit_ri_(VT.getSimpleVT(), ISDOpcode, Op1, Op1IsKill,
629                        CI->getZExtValue(), VT.getSimpleVT());
630       if (!ResultReg)
631         return false;
632 
633       // We successfully emitted code for the given LLVM Instruction.
634       updateValueMap(I, ResultReg);
635       return true;
636     }
637 
638   Register Op0 = getRegForValue(I->getOperand(0));
639   if (!Op0) // Unhandled operand. Halt "fast" selection and bail.
640     return false;
641   bool Op0IsKill = hasTrivialKill(I->getOperand(0));
642 
643   // Check if the second operand is a constant and handle it appropriately.
644   if (const auto *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
645     uint64_t Imm = CI->getSExtValue();
646 
647     // Transform "sdiv exact X, 8" -> "sra X, 3".
648     if (ISDOpcode == ISD::SDIV && isa<BinaryOperator>(I) &&
649         cast<BinaryOperator>(I)->isExact() && isPowerOf2_64(Imm)) {
650       Imm = Log2_64(Imm);
651       ISDOpcode = ISD::SRA;
652     }
653 
654     // Transform "urem x, pow2" -> "and x, pow2-1".
655     if (ISDOpcode == ISD::UREM && isa<BinaryOperator>(I) &&
656         isPowerOf2_64(Imm)) {
657       --Imm;
658       ISDOpcode = ISD::AND;
659     }
660 
661     Register ResultReg = fastEmit_ri_(VT.getSimpleVT(), ISDOpcode, Op0,
662                                       Op0IsKill, Imm, VT.getSimpleVT());
663     if (!ResultReg)
664       return false;
665 
666     // We successfully emitted code for the given LLVM Instruction.
667     updateValueMap(I, ResultReg);
668     return true;
669   }
670 
671   Register Op1 = getRegForValue(I->getOperand(1));
672   if (!Op1) // Unhandled operand. Halt "fast" selection and bail.
673     return false;
674   bool Op1IsKill = hasTrivialKill(I->getOperand(1));
675 
676   // Now we have both operands in registers. Emit the instruction.
677   Register ResultReg = fastEmit_rr(VT.getSimpleVT(), VT.getSimpleVT(),
678                                    ISDOpcode, Op0, Op0IsKill, Op1, Op1IsKill);
679   if (!ResultReg)
680     // Target-specific code wasn't able to find a machine opcode for
681     // the given ISD opcode and type. Halt "fast" selection and bail.
682     return false;
683 
684   // We successfully emitted code for the given LLVM Instruction.
685   updateValueMap(I, ResultReg);
686   return true;
687 }
688 
selectGetElementPtr(const User * I)689 bool FastISel::selectGetElementPtr(const User *I) {
690   Register N = getRegForValue(I->getOperand(0));
691   if (!N) // Unhandled operand. Halt "fast" selection and bail.
692     return false;
693 
694   // FIXME: The code below does not handle vector GEPs. Halt "fast" selection
695   // and bail.
696   if (isa<VectorType>(I->getType()))
697     return false;
698 
699   bool NIsKill = hasTrivialKill(I->getOperand(0));
700 
701   // Keep a running tab of the total offset to coalesce multiple N = N + Offset
702   // into a single N = N + TotalOffset.
703   uint64_t TotalOffs = 0;
704   // FIXME: What's a good SWAG number for MaxOffs?
705   uint64_t MaxOffs = 2048;
706   MVT VT = TLI.getPointerTy(DL);
707   for (gep_type_iterator GTI = gep_type_begin(I), E = gep_type_end(I);
708        GTI != E; ++GTI) {
709     const Value *Idx = GTI.getOperand();
710     if (StructType *StTy = GTI.getStructTypeOrNull()) {
711       uint64_t Field = cast<ConstantInt>(Idx)->getZExtValue();
712       if (Field) {
713         // N = N + Offset
714         TotalOffs += DL.getStructLayout(StTy)->getElementOffset(Field);
715         if (TotalOffs >= MaxOffs) {
716           N = fastEmit_ri_(VT, ISD::ADD, N, NIsKill, TotalOffs, VT);
717           if (!N) // Unhandled operand. Halt "fast" selection and bail.
718             return false;
719           NIsKill = true;
720           TotalOffs = 0;
721         }
722       }
723     } else {
724       Type *Ty = GTI.getIndexedType();
725 
726       // If this is a constant subscript, handle it quickly.
727       if (const auto *CI = dyn_cast<ConstantInt>(Idx)) {
728         if (CI->isZero())
729           continue;
730         // N = N + Offset
731         uint64_t IdxN = CI->getValue().sextOrTrunc(64).getSExtValue();
732         TotalOffs += DL.getTypeAllocSize(Ty) * IdxN;
733         if (TotalOffs >= MaxOffs) {
734           N = fastEmit_ri_(VT, ISD::ADD, N, NIsKill, TotalOffs, VT);
735           if (!N) // Unhandled operand. Halt "fast" selection and bail.
736             return false;
737           NIsKill = true;
738           TotalOffs = 0;
739         }
740         continue;
741       }
742       if (TotalOffs) {
743         N = fastEmit_ri_(VT, ISD::ADD, N, NIsKill, TotalOffs, VT);
744         if (!N) // Unhandled operand. Halt "fast" selection and bail.
745           return false;
746         NIsKill = true;
747         TotalOffs = 0;
748       }
749 
750       // N = N + Idx * ElementSize;
751       uint64_t ElementSize = DL.getTypeAllocSize(Ty);
752       std::pair<Register, bool> Pair = getRegForGEPIndex(Idx);
753       Register IdxN = Pair.first;
754       bool IdxNIsKill = Pair.second;
755       if (!IdxN) // Unhandled operand. Halt "fast" selection and bail.
756         return false;
757 
758       if (ElementSize != 1) {
759         IdxN = fastEmit_ri_(VT, ISD::MUL, IdxN, IdxNIsKill, ElementSize, VT);
760         if (!IdxN) // Unhandled operand. Halt "fast" selection and bail.
761           return false;
762         IdxNIsKill = true;
763       }
764       N = fastEmit_rr(VT, VT, ISD::ADD, N, NIsKill, IdxN, IdxNIsKill);
765       if (!N) // Unhandled operand. Halt "fast" selection and bail.
766         return false;
767     }
768   }
769   if (TotalOffs) {
770     N = fastEmit_ri_(VT, ISD::ADD, N, NIsKill, TotalOffs, VT);
771     if (!N) // Unhandled operand. Halt "fast" selection and bail.
772       return false;
773   }
774 
775   // We successfully emitted code for the given LLVM Instruction.
776   updateValueMap(I, N);
777   return true;
778 }
779 
addStackMapLiveVars(SmallVectorImpl<MachineOperand> & Ops,const CallInst * CI,unsigned StartIdx)780 bool FastISel::addStackMapLiveVars(SmallVectorImpl<MachineOperand> &Ops,
781                                    const CallInst *CI, unsigned StartIdx) {
782   for (unsigned i = StartIdx, e = CI->getNumArgOperands(); i != e; ++i) {
783     Value *Val = CI->getArgOperand(i);
784     // Check for constants and encode them with a StackMaps::ConstantOp prefix.
785     if (const auto *C = dyn_cast<ConstantInt>(Val)) {
786       Ops.push_back(MachineOperand::CreateImm(StackMaps::ConstantOp));
787       Ops.push_back(MachineOperand::CreateImm(C->getSExtValue()));
788     } else if (isa<ConstantPointerNull>(Val)) {
789       Ops.push_back(MachineOperand::CreateImm(StackMaps::ConstantOp));
790       Ops.push_back(MachineOperand::CreateImm(0));
791     } else if (auto *AI = dyn_cast<AllocaInst>(Val)) {
792       // Values coming from a stack location also require a special encoding,
793       // but that is added later on by the target specific frame index
794       // elimination implementation.
795       auto SI = FuncInfo.StaticAllocaMap.find(AI);
796       if (SI != FuncInfo.StaticAllocaMap.end())
797         Ops.push_back(MachineOperand::CreateFI(SI->second));
798       else
799         return false;
800     } else {
801       Register Reg = getRegForValue(Val);
802       if (!Reg)
803         return false;
804       Ops.push_back(MachineOperand::CreateReg(Reg, /*isDef=*/false));
805     }
806   }
807   return true;
808 }
809 
selectStackmap(const CallInst * I)810 bool FastISel::selectStackmap(const CallInst *I) {
811   // void @llvm.experimental.stackmap(i64 <id>, i32 <numShadowBytes>,
812   //                                  [live variables...])
813   assert(I->getCalledFunction()->getReturnType()->isVoidTy() &&
814          "Stackmap cannot return a value.");
815 
816   // The stackmap intrinsic only records the live variables (the arguments
817   // passed to it) and emits NOPS (if requested). Unlike the patchpoint
818   // intrinsic, this won't be lowered to a function call. This means we don't
819   // have to worry about calling conventions and target-specific lowering code.
820   // Instead we perform the call lowering right here.
821   //
822   // CALLSEQ_START(0, 0...)
823   // STACKMAP(id, nbytes, ...)
824   // CALLSEQ_END(0, 0)
825   //
826   SmallVector<MachineOperand, 32> Ops;
827 
828   // Add the <id> and <numBytes> constants.
829   assert(isa<ConstantInt>(I->getOperand(PatchPointOpers::IDPos)) &&
830          "Expected a constant integer.");
831   const auto *ID = cast<ConstantInt>(I->getOperand(PatchPointOpers::IDPos));
832   Ops.push_back(MachineOperand::CreateImm(ID->getZExtValue()));
833 
834   assert(isa<ConstantInt>(I->getOperand(PatchPointOpers::NBytesPos)) &&
835          "Expected a constant integer.");
836   const auto *NumBytes =
837       cast<ConstantInt>(I->getOperand(PatchPointOpers::NBytesPos));
838   Ops.push_back(MachineOperand::CreateImm(NumBytes->getZExtValue()));
839 
840   // Push live variables for the stack map (skipping the first two arguments
841   // <id> and <numBytes>).
842   if (!addStackMapLiveVars(Ops, I, 2))
843     return false;
844 
845   // We are not adding any register mask info here, because the stackmap doesn't
846   // clobber anything.
847 
848   // Add scratch registers as implicit def and early clobber.
849   CallingConv::ID CC = I->getCallingConv();
850   const MCPhysReg *ScratchRegs = TLI.getScratchRegisters(CC);
851   for (unsigned i = 0; ScratchRegs[i]; ++i)
852     Ops.push_back(MachineOperand::CreateReg(
853         ScratchRegs[i], /*isDef=*/true, /*isImp=*/true, /*isKill=*/false,
854         /*isDead=*/false, /*isUndef=*/false, /*isEarlyClobber=*/true));
855 
856   // Issue CALLSEQ_START
857   unsigned AdjStackDown = TII.getCallFrameSetupOpcode();
858   auto Builder =
859       BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AdjStackDown));
860   const MCInstrDesc &MCID = Builder.getInstr()->getDesc();
861   for (unsigned I = 0, E = MCID.getNumOperands(); I < E; ++I)
862     Builder.addImm(0);
863 
864   // Issue STACKMAP.
865   MachineInstrBuilder MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
866                                     TII.get(TargetOpcode::STACKMAP));
867   for (auto const &MO : Ops)
868     MIB.add(MO);
869 
870   // Issue CALLSEQ_END
871   unsigned AdjStackUp = TII.getCallFrameDestroyOpcode();
872   BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AdjStackUp))
873       .addImm(0)
874       .addImm(0);
875 
876   // Inform the Frame Information that we have a stackmap in this function.
877   FuncInfo.MF->getFrameInfo().setHasStackMap();
878 
879   return true;
880 }
881 
882 /// Lower an argument list according to the target calling convention.
883 ///
884 /// This is a helper for lowering intrinsics that follow a target calling
885 /// convention or require stack pointer adjustment. Only a subset of the
886 /// intrinsic's operands need to participate in the calling convention.
lowerCallOperands(const CallInst * CI,unsigned ArgIdx,unsigned NumArgs,const Value * Callee,bool ForceRetVoidTy,CallLoweringInfo & CLI)887 bool FastISel::lowerCallOperands(const CallInst *CI, unsigned ArgIdx,
888                                  unsigned NumArgs, const Value *Callee,
889                                  bool ForceRetVoidTy, CallLoweringInfo &CLI) {
890   ArgListTy Args;
891   Args.reserve(NumArgs);
892 
893   // Populate the argument list.
894   for (unsigned ArgI = ArgIdx, ArgE = ArgIdx + NumArgs; ArgI != ArgE; ++ArgI) {
895     Value *V = CI->getOperand(ArgI);
896 
897     assert(!V->getType()->isEmptyTy() && "Empty type passed to intrinsic.");
898 
899     ArgListEntry Entry;
900     Entry.Val = V;
901     Entry.Ty = V->getType();
902     Entry.setAttributes(CI, ArgI);
903     Args.push_back(Entry);
904   }
905 
906   Type *RetTy = ForceRetVoidTy ? Type::getVoidTy(CI->getType()->getContext())
907                                : CI->getType();
908   CLI.setCallee(CI->getCallingConv(), RetTy, Callee, std::move(Args), NumArgs);
909 
910   return lowerCallTo(CLI);
911 }
912 
setCallee(const DataLayout & DL,MCContext & Ctx,CallingConv::ID CC,Type * ResultTy,StringRef Target,ArgListTy && ArgsList,unsigned FixedArgs)913 FastISel::CallLoweringInfo &FastISel::CallLoweringInfo::setCallee(
914     const DataLayout &DL, MCContext &Ctx, CallingConv::ID CC, Type *ResultTy,
915     StringRef Target, ArgListTy &&ArgsList, unsigned FixedArgs) {
916   SmallString<32> MangledName;
917   Mangler::getNameWithPrefix(MangledName, Target, DL);
918   MCSymbol *Sym = Ctx.getOrCreateSymbol(MangledName);
919   return setCallee(CC, ResultTy, Sym, std::move(ArgsList), FixedArgs);
920 }
921 
selectPatchpoint(const CallInst * I)922 bool FastISel::selectPatchpoint(const CallInst *I) {
923   // void|i64 @llvm.experimental.patchpoint.void|i64(i64 <id>,
924   //                                                 i32 <numBytes>,
925   //                                                 i8* <target>,
926   //                                                 i32 <numArgs>,
927   //                                                 [Args...],
928   //                                                 [live variables...])
929   CallingConv::ID CC = I->getCallingConv();
930   bool IsAnyRegCC = CC == CallingConv::AnyReg;
931   bool HasDef = !I->getType()->isVoidTy();
932   Value *Callee = I->getOperand(PatchPointOpers::TargetPos)->stripPointerCasts();
933 
934   // Get the real number of arguments participating in the call <numArgs>
935   assert(isa<ConstantInt>(I->getOperand(PatchPointOpers::NArgPos)) &&
936          "Expected a constant integer.");
937   const auto *NumArgsVal =
938       cast<ConstantInt>(I->getOperand(PatchPointOpers::NArgPos));
939   unsigned NumArgs = NumArgsVal->getZExtValue();
940 
941   // Skip the four meta args: <id>, <numNopBytes>, <target>, <numArgs>
942   // This includes all meta-operands up to but not including CC.
943   unsigned NumMetaOpers = PatchPointOpers::CCPos;
944   assert(I->getNumArgOperands() >= NumMetaOpers + NumArgs &&
945          "Not enough arguments provided to the patchpoint intrinsic");
946 
947   // For AnyRegCC the arguments are lowered later on manually.
948   unsigned NumCallArgs = IsAnyRegCC ? 0 : NumArgs;
949   CallLoweringInfo CLI;
950   CLI.setIsPatchPoint();
951   if (!lowerCallOperands(I, NumMetaOpers, NumCallArgs, Callee, IsAnyRegCC, CLI))
952     return false;
953 
954   assert(CLI.Call && "No call instruction specified.");
955 
956   SmallVector<MachineOperand, 32> Ops;
957 
958   // Add an explicit result reg if we use the anyreg calling convention.
959   if (IsAnyRegCC && HasDef) {
960     assert(CLI.NumResultRegs == 0 && "Unexpected result register.");
961     CLI.ResultReg = createResultReg(TLI.getRegClassFor(MVT::i64));
962     CLI.NumResultRegs = 1;
963     Ops.push_back(MachineOperand::CreateReg(CLI.ResultReg, /*isDef=*/true));
964   }
965 
966   // Add the <id> and <numBytes> constants.
967   assert(isa<ConstantInt>(I->getOperand(PatchPointOpers::IDPos)) &&
968          "Expected a constant integer.");
969   const auto *ID = cast<ConstantInt>(I->getOperand(PatchPointOpers::IDPos));
970   Ops.push_back(MachineOperand::CreateImm(ID->getZExtValue()));
971 
972   assert(isa<ConstantInt>(I->getOperand(PatchPointOpers::NBytesPos)) &&
973          "Expected a constant integer.");
974   const auto *NumBytes =
975       cast<ConstantInt>(I->getOperand(PatchPointOpers::NBytesPos));
976   Ops.push_back(MachineOperand::CreateImm(NumBytes->getZExtValue()));
977 
978   // Add the call target.
979   if (const auto *C = dyn_cast<IntToPtrInst>(Callee)) {
980     uint64_t CalleeConstAddr =
981       cast<ConstantInt>(C->getOperand(0))->getZExtValue();
982     Ops.push_back(MachineOperand::CreateImm(CalleeConstAddr));
983   } else if (const auto *C = dyn_cast<ConstantExpr>(Callee)) {
984     if (C->getOpcode() == Instruction::IntToPtr) {
985       uint64_t CalleeConstAddr =
986         cast<ConstantInt>(C->getOperand(0))->getZExtValue();
987       Ops.push_back(MachineOperand::CreateImm(CalleeConstAddr));
988     } else
989       llvm_unreachable("Unsupported ConstantExpr.");
990   } else if (const auto *GV = dyn_cast<GlobalValue>(Callee)) {
991     Ops.push_back(MachineOperand::CreateGA(GV, 0));
992   } else if (isa<ConstantPointerNull>(Callee))
993     Ops.push_back(MachineOperand::CreateImm(0));
994   else
995     llvm_unreachable("Unsupported callee address.");
996 
997   // Adjust <numArgs> to account for any arguments that have been passed on
998   // the stack instead.
999   unsigned NumCallRegArgs = IsAnyRegCC ? NumArgs : CLI.OutRegs.size();
1000   Ops.push_back(MachineOperand::CreateImm(NumCallRegArgs));
1001 
1002   // Add the calling convention
1003   Ops.push_back(MachineOperand::CreateImm((unsigned)CC));
1004 
1005   // Add the arguments we omitted previously. The register allocator should
1006   // place these in any free register.
1007   if (IsAnyRegCC) {
1008     for (unsigned i = NumMetaOpers, e = NumMetaOpers + NumArgs; i != e; ++i) {
1009       Register Reg = getRegForValue(I->getArgOperand(i));
1010       if (!Reg)
1011         return false;
1012       Ops.push_back(MachineOperand::CreateReg(Reg, /*isDef=*/false));
1013     }
1014   }
1015 
1016   // Push the arguments from the call instruction.
1017   for (auto Reg : CLI.OutRegs)
1018     Ops.push_back(MachineOperand::CreateReg(Reg, /*isDef=*/false));
1019 
1020   // Push live variables for the stack map.
1021   if (!addStackMapLiveVars(Ops, I, NumMetaOpers + NumArgs))
1022     return false;
1023 
1024   // Push the register mask info.
1025   Ops.push_back(MachineOperand::CreateRegMask(
1026       TRI.getCallPreservedMask(*FuncInfo.MF, CC)));
1027 
1028   // Add scratch registers as implicit def and early clobber.
1029   const MCPhysReg *ScratchRegs = TLI.getScratchRegisters(CC);
1030   for (unsigned i = 0; ScratchRegs[i]; ++i)
1031     Ops.push_back(MachineOperand::CreateReg(
1032         ScratchRegs[i], /*isDef=*/true, /*isImp=*/true, /*isKill=*/false,
1033         /*isDead=*/false, /*isUndef=*/false, /*isEarlyClobber=*/true));
1034 
1035   // Add implicit defs (return values).
1036   for (auto Reg : CLI.InRegs)
1037     Ops.push_back(MachineOperand::CreateReg(Reg, /*isDef=*/true,
1038                                             /*isImp=*/true));
1039 
1040   // Insert the patchpoint instruction before the call generated by the target.
1041   MachineInstrBuilder MIB = BuildMI(*FuncInfo.MBB, CLI.Call, DbgLoc,
1042                                     TII.get(TargetOpcode::PATCHPOINT));
1043 
1044   for (auto &MO : Ops)
1045     MIB.add(MO);
1046 
1047   MIB->setPhysRegsDeadExcept(CLI.InRegs, TRI);
1048 
1049   // Delete the original call instruction.
1050   CLI.Call->eraseFromParent();
1051 
1052   // Inform the Frame Information that we have a patchpoint in this function.
1053   FuncInfo.MF->getFrameInfo().setHasPatchPoint();
1054 
1055   if (CLI.NumResultRegs)
1056     updateValueMap(I, CLI.ResultReg, CLI.NumResultRegs);
1057   return true;
1058 }
1059 
selectXRayCustomEvent(const CallInst * I)1060 bool FastISel::selectXRayCustomEvent(const CallInst *I) {
1061   const auto &Triple = TM.getTargetTriple();
1062   if (Triple.getArch() != Triple::x86_64 || !Triple.isOSLinux())
1063     return true; // don't do anything to this instruction.
1064   SmallVector<MachineOperand, 8> Ops;
1065   Ops.push_back(MachineOperand::CreateReg(getRegForValue(I->getArgOperand(0)),
1066                                           /*isDef=*/false));
1067   Ops.push_back(MachineOperand::CreateReg(getRegForValue(I->getArgOperand(1)),
1068                                           /*isDef=*/false));
1069   MachineInstrBuilder MIB =
1070       BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
1071               TII.get(TargetOpcode::PATCHABLE_EVENT_CALL));
1072   for (auto &MO : Ops)
1073     MIB.add(MO);
1074 
1075   // Insert the Patchable Event Call instruction, that gets lowered properly.
1076   return true;
1077 }
1078 
selectXRayTypedEvent(const CallInst * I)1079 bool FastISel::selectXRayTypedEvent(const CallInst *I) {
1080   const auto &Triple = TM.getTargetTriple();
1081   if (Triple.getArch() != Triple::x86_64 || !Triple.isOSLinux())
1082     return true; // don't do anything to this instruction.
1083   SmallVector<MachineOperand, 8> Ops;
1084   Ops.push_back(MachineOperand::CreateReg(getRegForValue(I->getArgOperand(0)),
1085                                           /*isDef=*/false));
1086   Ops.push_back(MachineOperand::CreateReg(getRegForValue(I->getArgOperand(1)),
1087                                           /*isDef=*/false));
1088   Ops.push_back(MachineOperand::CreateReg(getRegForValue(I->getArgOperand(2)),
1089                                           /*isDef=*/false));
1090   MachineInstrBuilder MIB =
1091       BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
1092               TII.get(TargetOpcode::PATCHABLE_TYPED_EVENT_CALL));
1093   for (auto &MO : Ops)
1094     MIB.add(MO);
1095 
1096   // Insert the Patchable Typed Event Call instruction, that gets lowered properly.
1097   return true;
1098 }
1099 
1100 /// Returns an AttributeList representing the attributes applied to the return
1101 /// value of the given call.
getReturnAttrs(FastISel::CallLoweringInfo & CLI)1102 static AttributeList getReturnAttrs(FastISel::CallLoweringInfo &CLI) {
1103   SmallVector<Attribute::AttrKind, 2> Attrs;
1104   if (CLI.RetSExt)
1105     Attrs.push_back(Attribute::SExt);
1106   if (CLI.RetZExt)
1107     Attrs.push_back(Attribute::ZExt);
1108   if (CLI.IsInReg)
1109     Attrs.push_back(Attribute::InReg);
1110 
1111   return AttributeList::get(CLI.RetTy->getContext(), AttributeList::ReturnIndex,
1112                             Attrs);
1113 }
1114 
lowerCallTo(const CallInst * CI,const char * SymName,unsigned NumArgs)1115 bool FastISel::lowerCallTo(const CallInst *CI, const char *SymName,
1116                            unsigned NumArgs) {
1117   MCContext &Ctx = MF->getContext();
1118   SmallString<32> MangledName;
1119   Mangler::getNameWithPrefix(MangledName, SymName, DL);
1120   MCSymbol *Sym = Ctx.getOrCreateSymbol(MangledName);
1121   return lowerCallTo(CI, Sym, NumArgs);
1122 }
1123 
lowerCallTo(const CallInst * CI,MCSymbol * Symbol,unsigned NumArgs)1124 bool FastISel::lowerCallTo(const CallInst *CI, MCSymbol *Symbol,
1125                            unsigned NumArgs) {
1126   FunctionType *FTy = CI->getFunctionType();
1127   Type *RetTy = CI->getType();
1128 
1129   ArgListTy Args;
1130   Args.reserve(NumArgs);
1131 
1132   // Populate the argument list.
1133   // Attributes for args start at offset 1, after the return attribute.
1134   for (unsigned ArgI = 0; ArgI != NumArgs; ++ArgI) {
1135     Value *V = CI->getOperand(ArgI);
1136 
1137     assert(!V->getType()->isEmptyTy() && "Empty type passed to intrinsic.");
1138 
1139     ArgListEntry Entry;
1140     Entry.Val = V;
1141     Entry.Ty = V->getType();
1142     Entry.setAttributes(CI, ArgI);
1143     Args.push_back(Entry);
1144   }
1145   TLI.markLibCallAttributes(MF, CI->getCallingConv(), Args);
1146 
1147   CallLoweringInfo CLI;
1148   CLI.setCallee(RetTy, FTy, Symbol, std::move(Args), *CI, NumArgs);
1149 
1150   return lowerCallTo(CLI);
1151 }
1152 
lowerCallTo(CallLoweringInfo & CLI)1153 bool FastISel::lowerCallTo(CallLoweringInfo &CLI) {
1154   // Handle the incoming return values from the call.
1155   CLI.clearIns();
1156   SmallVector<EVT, 4> RetTys;
1157   ComputeValueVTs(TLI, DL, CLI.RetTy, RetTys);
1158 
1159   SmallVector<ISD::OutputArg, 4> Outs;
1160   GetReturnInfo(CLI.CallConv, CLI.RetTy, getReturnAttrs(CLI), Outs, TLI, DL);
1161 
1162   bool CanLowerReturn = TLI.CanLowerReturn(
1163       CLI.CallConv, *FuncInfo.MF, CLI.IsVarArg, Outs, CLI.RetTy->getContext());
1164 
1165   // FIXME: sret demotion isn't supported yet - bail out.
1166   if (!CanLowerReturn)
1167     return false;
1168 
1169   for (unsigned I = 0, E = RetTys.size(); I != E; ++I) {
1170     EVT VT = RetTys[I];
1171     MVT RegisterVT = TLI.getRegisterType(CLI.RetTy->getContext(), VT);
1172     unsigned NumRegs = TLI.getNumRegisters(CLI.RetTy->getContext(), VT);
1173     for (unsigned i = 0; i != NumRegs; ++i) {
1174       ISD::InputArg MyFlags;
1175       MyFlags.VT = RegisterVT;
1176       MyFlags.ArgVT = VT;
1177       MyFlags.Used = CLI.IsReturnValueUsed;
1178       if (CLI.RetSExt)
1179         MyFlags.Flags.setSExt();
1180       if (CLI.RetZExt)
1181         MyFlags.Flags.setZExt();
1182       if (CLI.IsInReg)
1183         MyFlags.Flags.setInReg();
1184       CLI.Ins.push_back(MyFlags);
1185     }
1186   }
1187 
1188   // Handle all of the outgoing arguments.
1189   CLI.clearOuts();
1190   for (auto &Arg : CLI.getArgs()) {
1191     Type *FinalType = Arg.Ty;
1192     if (Arg.IsByVal)
1193       FinalType = cast<PointerType>(Arg.Ty)->getElementType();
1194     bool NeedsRegBlock = TLI.functionArgumentNeedsConsecutiveRegisters(
1195         FinalType, CLI.CallConv, CLI.IsVarArg);
1196 
1197     ISD::ArgFlagsTy Flags;
1198     if (Arg.IsZExt)
1199       Flags.setZExt();
1200     if (Arg.IsSExt)
1201       Flags.setSExt();
1202     if (Arg.IsInReg)
1203       Flags.setInReg();
1204     if (Arg.IsSRet)
1205       Flags.setSRet();
1206     if (Arg.IsSwiftSelf)
1207       Flags.setSwiftSelf();
1208     if (Arg.IsSwiftError)
1209       Flags.setSwiftError();
1210     if (Arg.IsCFGuardTarget)
1211       Flags.setCFGuardTarget();
1212     if (Arg.IsByVal)
1213       Flags.setByVal();
1214     if (Arg.IsInAlloca) {
1215       Flags.setInAlloca();
1216       // Set the byval flag for CCAssignFn callbacks that don't know about
1217       // inalloca. This way we can know how many bytes we should've allocated
1218       // and how many bytes a callee cleanup function will pop.  If we port
1219       // inalloca to more targets, we'll have to add custom inalloca handling in
1220       // the various CC lowering callbacks.
1221       Flags.setByVal();
1222     }
1223     if (Arg.IsPreallocated) {
1224       Flags.setPreallocated();
1225       // Set the byval flag for CCAssignFn callbacks that don't know about
1226       // preallocated. This way we can know how many bytes we should've
1227       // allocated and how many bytes a callee cleanup function will pop.  If we
1228       // port preallocated to more targets, we'll have to add custom
1229       // preallocated handling in the various CC lowering callbacks.
1230       Flags.setByVal();
1231     }
1232     if (Arg.IsByVal || Arg.IsInAlloca || Arg.IsPreallocated) {
1233       PointerType *Ty = cast<PointerType>(Arg.Ty);
1234       Type *ElementTy = Ty->getElementType();
1235       unsigned FrameSize =
1236           DL.getTypeAllocSize(Arg.ByValType ? Arg.ByValType : ElementTy);
1237 
1238       // For ByVal, alignment should come from FE. BE will guess if this info
1239       // is not there, but there are cases it cannot get right.
1240       MaybeAlign FrameAlign = Arg.Alignment;
1241       if (!FrameAlign)
1242         FrameAlign = Align(TLI.getByValTypeAlignment(ElementTy, DL));
1243       Flags.setByValSize(FrameSize);
1244       Flags.setByValAlign(*FrameAlign);
1245     }
1246     if (Arg.IsNest)
1247       Flags.setNest();
1248     if (NeedsRegBlock)
1249       Flags.setInConsecutiveRegs();
1250     Flags.setOrigAlign(DL.getABITypeAlign(Arg.Ty));
1251 
1252     CLI.OutVals.push_back(Arg.Val);
1253     CLI.OutFlags.push_back(Flags);
1254   }
1255 
1256   if (!fastLowerCall(CLI))
1257     return false;
1258 
1259   // Set all unused physreg defs as dead.
1260   assert(CLI.Call && "No call instruction specified.");
1261   CLI.Call->setPhysRegsDeadExcept(CLI.InRegs, TRI);
1262 
1263   if (CLI.NumResultRegs && CLI.CB)
1264     updateValueMap(CLI.CB, CLI.ResultReg, CLI.NumResultRegs);
1265 
1266   // Set labels for heapallocsite call.
1267   if (CLI.CB)
1268     if (MDNode *MD = CLI.CB->getMetadata("heapallocsite"))
1269       CLI.Call->setHeapAllocMarker(*MF, MD);
1270 
1271   return true;
1272 }
1273 
lowerCall(const CallInst * CI)1274 bool FastISel::lowerCall(const CallInst *CI) {
1275   FunctionType *FuncTy = CI->getFunctionType();
1276   Type *RetTy = CI->getType();
1277 
1278   ArgListTy Args;
1279   ArgListEntry Entry;
1280   Args.reserve(CI->arg_size());
1281 
1282   for (auto i = CI->arg_begin(), e = CI->arg_end(); i != e; ++i) {
1283     Value *V = *i;
1284 
1285     // Skip empty types
1286     if (V->getType()->isEmptyTy())
1287       continue;
1288 
1289     Entry.Val = V;
1290     Entry.Ty = V->getType();
1291 
1292     // Skip the first return-type Attribute to get to params.
1293     Entry.setAttributes(CI, i - CI->arg_begin());
1294     Args.push_back(Entry);
1295   }
1296 
1297   // Check if target-independent constraints permit a tail call here.
1298   // Target-dependent constraints are checked within fastLowerCall.
1299   bool IsTailCall = CI->isTailCall();
1300   if (IsTailCall && !isInTailCallPosition(*CI, TM))
1301     IsTailCall = false;
1302   if (IsTailCall && MF->getFunction()
1303                             .getFnAttribute("disable-tail-calls")
1304                             .getValueAsString() == "true")
1305     IsTailCall = false;
1306 
1307   CallLoweringInfo CLI;
1308   CLI.setCallee(RetTy, FuncTy, CI->getCalledOperand(), std::move(Args), *CI)
1309       .setTailCall(IsTailCall);
1310 
1311   return lowerCallTo(CLI);
1312 }
1313 
selectCall(const User * I)1314 bool FastISel::selectCall(const User *I) {
1315   const CallInst *Call = cast<CallInst>(I);
1316 
1317   // Handle simple inline asms.
1318   if (const InlineAsm *IA = dyn_cast<InlineAsm>(Call->getCalledOperand())) {
1319     // If the inline asm has side effects, then make sure that no local value
1320     // lives across by flushing the local value map.
1321     if (IA->hasSideEffects())
1322       flushLocalValueMap();
1323 
1324     // Don't attempt to handle constraints.
1325     if (!IA->getConstraintString().empty())
1326       return false;
1327 
1328     unsigned ExtraInfo = 0;
1329     if (IA->hasSideEffects())
1330       ExtraInfo |= InlineAsm::Extra_HasSideEffects;
1331     if (IA->isAlignStack())
1332       ExtraInfo |= InlineAsm::Extra_IsAlignStack;
1333     if (Call->isConvergent())
1334       ExtraInfo |= InlineAsm::Extra_IsConvergent;
1335     ExtraInfo |= IA->getDialect() * InlineAsm::Extra_AsmDialect;
1336 
1337     MachineInstrBuilder MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
1338                                       TII.get(TargetOpcode::INLINEASM));
1339     MIB.addExternalSymbol(IA->getAsmString().c_str());
1340     MIB.addImm(ExtraInfo);
1341 
1342     const MDNode *SrcLoc = Call->getMetadata("srcloc");
1343     if (SrcLoc)
1344       MIB.addMetadata(SrcLoc);
1345 
1346     return true;
1347   }
1348 
1349   // Handle intrinsic function calls.
1350   if (const auto *II = dyn_cast<IntrinsicInst>(Call))
1351     return selectIntrinsicCall(II);
1352 
1353   // Usually, it does not make sense to initialize a value,
1354   // make an unrelated function call and use the value, because
1355   // it tends to be spilled on the stack. So, we move the pointer
1356   // to the last local value to the beginning of the block, so that
1357   // all the values which have already been materialized,
1358   // appear after the call. It also makes sense to skip intrinsics
1359   // since they tend to be inlined.
1360   flushLocalValueMap();
1361 
1362   return lowerCall(Call);
1363 }
1364 
selectIntrinsicCall(const IntrinsicInst * II)1365 bool FastISel::selectIntrinsicCall(const IntrinsicInst *II) {
1366   switch (II->getIntrinsicID()) {
1367   default:
1368     break;
1369   // At -O0 we don't care about the lifetime intrinsics.
1370   case Intrinsic::lifetime_start:
1371   case Intrinsic::lifetime_end:
1372   // The donothing intrinsic does, well, nothing.
1373   case Intrinsic::donothing:
1374   // Neither does the sideeffect intrinsic.
1375   case Intrinsic::sideeffect:
1376   // Neither does the assume intrinsic; it's also OK not to codegen its operand.
1377   case Intrinsic::assume:
1378     return true;
1379   case Intrinsic::dbg_declare: {
1380     const DbgDeclareInst *DI = cast<DbgDeclareInst>(II);
1381     assert(DI->getVariable() && "Missing variable");
1382     if (!FuncInfo.MF->getMMI().hasDebugInfo()) {
1383       LLVM_DEBUG(dbgs() << "Dropping debug info for " << *DI
1384                         << " (!hasDebugInfo)\n");
1385       return true;
1386     }
1387 
1388     const Value *Address = DI->getAddress();
1389     if (!Address || isa<UndefValue>(Address)) {
1390       LLVM_DEBUG(dbgs() << "Dropping debug info for " << *DI
1391                         << " (bad/undef address)\n");
1392       return true;
1393     }
1394 
1395     // Byval arguments with frame indices were already handled after argument
1396     // lowering and before isel.
1397     const auto *Arg =
1398         dyn_cast<Argument>(Address->stripInBoundsConstantOffsets());
1399     if (Arg && FuncInfo.getArgumentFrameIndex(Arg) != INT_MAX)
1400       return true;
1401 
1402     Optional<MachineOperand> Op;
1403     if (Register Reg = lookUpRegForValue(Address))
1404       Op = MachineOperand::CreateReg(Reg, false);
1405 
1406     // If we have a VLA that has a "use" in a metadata node that's then used
1407     // here but it has no other uses, then we have a problem. E.g.,
1408     //
1409     //   int foo (const int *x) {
1410     //     char a[*x];
1411     //     return 0;
1412     //   }
1413     //
1414     // If we assign 'a' a vreg and fast isel later on has to use the selection
1415     // DAG isel, it will want to copy the value to the vreg. However, there are
1416     // no uses, which goes counter to what selection DAG isel expects.
1417     if (!Op && !Address->use_empty() && isa<Instruction>(Address) &&
1418         (!isa<AllocaInst>(Address) ||
1419          !FuncInfo.StaticAllocaMap.count(cast<AllocaInst>(Address))))
1420       Op = MachineOperand::CreateReg(FuncInfo.InitializeRegForValue(Address),
1421                                      false);
1422 
1423     if (Op) {
1424       assert(DI->getVariable()->isValidLocationForIntrinsic(DbgLoc) &&
1425              "Expected inlined-at fields to agree");
1426       // A dbg.declare describes the address of a source variable, so lower it
1427       // into an indirect DBG_VALUE.
1428       BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
1429               TII.get(TargetOpcode::DBG_VALUE), /*IsIndirect*/ true,
1430               *Op, DI->getVariable(), DI->getExpression());
1431     } else {
1432       // We can't yet handle anything else here because it would require
1433       // generating code, thus altering codegen because of debug info.
1434       LLVM_DEBUG(dbgs() << "Dropping debug info for " << *DI
1435                         << " (no materialized reg for address)\n");
1436     }
1437     return true;
1438   }
1439   case Intrinsic::dbg_value: {
1440     // This form of DBG_VALUE is target-independent.
1441     const DbgValueInst *DI = cast<DbgValueInst>(II);
1442     const MCInstrDesc &II = TII.get(TargetOpcode::DBG_VALUE);
1443     const Value *V = DI->getValue();
1444     assert(DI->getVariable()->isValidLocationForIntrinsic(DbgLoc) &&
1445            "Expected inlined-at fields to agree");
1446     if (!V || isa<UndefValue>(V)) {
1447       // Currently the optimizer can produce this; insert an undef to
1448       // help debugging.
1449       BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, II, false, 0U,
1450               DI->getVariable(), DI->getExpression());
1451     } else if (const auto *CI = dyn_cast<ConstantInt>(V)) {
1452       if (CI->getBitWidth() > 64)
1453         BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, II)
1454             .addCImm(CI)
1455             .addImm(0U)
1456             .addMetadata(DI->getVariable())
1457             .addMetadata(DI->getExpression());
1458       else
1459         BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, II)
1460             .addImm(CI->getZExtValue())
1461             .addImm(0U)
1462             .addMetadata(DI->getVariable())
1463             .addMetadata(DI->getExpression());
1464     } else if (const auto *CF = dyn_cast<ConstantFP>(V)) {
1465       BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, II)
1466           .addFPImm(CF)
1467           .addImm(0U)
1468           .addMetadata(DI->getVariable())
1469           .addMetadata(DI->getExpression());
1470     } else if (Register Reg = lookUpRegForValue(V)) {
1471       // FIXME: This does not handle register-indirect values at offset 0.
1472       bool IsIndirect = false;
1473       BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, II, IsIndirect, Reg,
1474               DI->getVariable(), DI->getExpression());
1475     } else {
1476       // We don't know how to handle other cases, so we drop.
1477       LLVM_DEBUG(dbgs() << "Dropping debug info for " << *DI << "\n");
1478     }
1479     return true;
1480   }
1481   case Intrinsic::dbg_label: {
1482     const DbgLabelInst *DI = cast<DbgLabelInst>(II);
1483     assert(DI->getLabel() && "Missing label");
1484     if (!FuncInfo.MF->getMMI().hasDebugInfo()) {
1485       LLVM_DEBUG(dbgs() << "Dropping debug info for " << *DI << "\n");
1486       return true;
1487     }
1488 
1489     BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
1490             TII.get(TargetOpcode::DBG_LABEL)).addMetadata(DI->getLabel());
1491     return true;
1492   }
1493   case Intrinsic::objectsize:
1494     llvm_unreachable("llvm.objectsize.* should have been lowered already");
1495 
1496   case Intrinsic::is_constant:
1497     llvm_unreachable("llvm.is.constant.* should have been lowered already");
1498 
1499   case Intrinsic::launder_invariant_group:
1500   case Intrinsic::strip_invariant_group:
1501   case Intrinsic::expect: {
1502     Register ResultReg = getRegForValue(II->getArgOperand(0));
1503     if (!ResultReg)
1504       return false;
1505     updateValueMap(II, ResultReg);
1506     return true;
1507   }
1508   case Intrinsic::experimental_stackmap:
1509     return selectStackmap(II);
1510   case Intrinsic::experimental_patchpoint_void:
1511   case Intrinsic::experimental_patchpoint_i64:
1512     return selectPatchpoint(II);
1513 
1514   case Intrinsic::xray_customevent:
1515     return selectXRayCustomEvent(II);
1516   case Intrinsic::xray_typedevent:
1517     return selectXRayTypedEvent(II);
1518   }
1519 
1520   return fastLowerIntrinsicCall(II);
1521 }
1522 
selectCast(const User * I,unsigned Opcode)1523 bool FastISel::selectCast(const User *I, unsigned Opcode) {
1524   EVT SrcVT = TLI.getValueType(DL, I->getOperand(0)->getType());
1525   EVT DstVT = TLI.getValueType(DL, I->getType());
1526 
1527   if (SrcVT == MVT::Other || !SrcVT.isSimple() || DstVT == MVT::Other ||
1528       !DstVT.isSimple())
1529     // Unhandled type. Halt "fast" selection and bail.
1530     return false;
1531 
1532   // Check if the destination type is legal.
1533   if (!TLI.isTypeLegal(DstVT))
1534     return false;
1535 
1536   // Check if the source operand is legal.
1537   if (!TLI.isTypeLegal(SrcVT))
1538     return false;
1539 
1540   Register InputReg = getRegForValue(I->getOperand(0));
1541   if (!InputReg)
1542     // Unhandled operand.  Halt "fast" selection and bail.
1543     return false;
1544 
1545   bool InputRegIsKill = hasTrivialKill(I->getOperand(0));
1546 
1547   Register ResultReg = fastEmit_r(SrcVT.getSimpleVT(), DstVT.getSimpleVT(),
1548                                   Opcode, InputReg, InputRegIsKill);
1549   if (!ResultReg)
1550     return false;
1551 
1552   updateValueMap(I, ResultReg);
1553   return true;
1554 }
1555 
selectBitCast(const User * I)1556 bool FastISel::selectBitCast(const User *I) {
1557   // If the bitcast doesn't change the type, just use the operand value.
1558   if (I->getType() == I->getOperand(0)->getType()) {
1559     Register Reg = getRegForValue(I->getOperand(0));
1560     if (!Reg)
1561       return false;
1562     updateValueMap(I, Reg);
1563     return true;
1564   }
1565 
1566   // Bitcasts of other values become reg-reg copies or BITCAST operators.
1567   EVT SrcEVT = TLI.getValueType(DL, I->getOperand(0)->getType());
1568   EVT DstEVT = TLI.getValueType(DL, I->getType());
1569   if (SrcEVT == MVT::Other || DstEVT == MVT::Other ||
1570       !TLI.isTypeLegal(SrcEVT) || !TLI.isTypeLegal(DstEVT))
1571     // Unhandled type. Halt "fast" selection and bail.
1572     return false;
1573 
1574   MVT SrcVT = SrcEVT.getSimpleVT();
1575   MVT DstVT = DstEVT.getSimpleVT();
1576   Register Op0 = getRegForValue(I->getOperand(0));
1577   if (!Op0) // Unhandled operand. Halt "fast" selection and bail.
1578     return false;
1579   bool Op0IsKill = hasTrivialKill(I->getOperand(0));
1580 
1581   // First, try to perform the bitcast by inserting a reg-reg copy.
1582   Register ResultReg;
1583   if (SrcVT == DstVT) {
1584     const TargetRegisterClass *SrcClass = TLI.getRegClassFor(SrcVT);
1585     const TargetRegisterClass *DstClass = TLI.getRegClassFor(DstVT);
1586     // Don't attempt a cross-class copy. It will likely fail.
1587     if (SrcClass == DstClass) {
1588       ResultReg = createResultReg(DstClass);
1589       BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
1590               TII.get(TargetOpcode::COPY), ResultReg).addReg(Op0);
1591     }
1592   }
1593 
1594   // If the reg-reg copy failed, select a BITCAST opcode.
1595   if (!ResultReg)
1596     ResultReg = fastEmit_r(SrcVT, DstVT, ISD::BITCAST, Op0, Op0IsKill);
1597 
1598   if (!ResultReg)
1599     return false;
1600 
1601   updateValueMap(I, ResultReg);
1602   return true;
1603 }
1604 
selectFreeze(const User * I)1605 bool FastISel::selectFreeze(const User *I) {
1606   Register Reg = getRegForValue(I->getOperand(0));
1607   if (!Reg)
1608     // Unhandled operand.
1609     return false;
1610 
1611   EVT ETy = TLI.getValueType(DL, I->getOperand(0)->getType());
1612   if (ETy == MVT::Other || !TLI.isTypeLegal(ETy))
1613     // Unhandled type, bail out.
1614     return false;
1615 
1616   MVT Ty = ETy.getSimpleVT();
1617   const TargetRegisterClass *TyRegClass = TLI.getRegClassFor(Ty);
1618   Register ResultReg = createResultReg(TyRegClass);
1619   BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
1620           TII.get(TargetOpcode::COPY), ResultReg).addReg(Reg);
1621 
1622   updateValueMap(I, ResultReg);
1623   return true;
1624 }
1625 
1626 // Remove local value instructions starting from the instruction after
1627 // SavedLastLocalValue to the current function insert point.
removeDeadLocalValueCode(MachineInstr * SavedLastLocalValue)1628 void FastISel::removeDeadLocalValueCode(MachineInstr *SavedLastLocalValue)
1629 {
1630   MachineInstr *CurLastLocalValue = getLastLocalValue();
1631   if (CurLastLocalValue != SavedLastLocalValue) {
1632     // Find the first local value instruction to be deleted.
1633     // This is the instruction after SavedLastLocalValue if it is non-NULL.
1634     // Otherwise it's the first instruction in the block.
1635     MachineBasicBlock::iterator FirstDeadInst(SavedLastLocalValue);
1636     if (SavedLastLocalValue)
1637       ++FirstDeadInst;
1638     else
1639       FirstDeadInst = FuncInfo.MBB->getFirstNonPHI();
1640     setLastLocalValue(SavedLastLocalValue);
1641     removeDeadCode(FirstDeadInst, FuncInfo.InsertPt);
1642   }
1643 }
1644 
selectInstruction(const Instruction * I)1645 bool FastISel::selectInstruction(const Instruction *I) {
1646   MachineInstr *SavedLastLocalValue = getLastLocalValue();
1647   // Just before the terminator instruction, insert instructions to
1648   // feed PHI nodes in successor blocks.
1649   if (I->isTerminator()) {
1650     if (!handlePHINodesInSuccessorBlocks(I->getParent())) {
1651       // PHI node handling may have generated local value instructions,
1652       // even though it failed to handle all PHI nodes.
1653       // We remove these instructions because SelectionDAGISel will generate
1654       // them again.
1655       removeDeadLocalValueCode(SavedLastLocalValue);
1656       return false;
1657     }
1658   }
1659 
1660   // FastISel does not handle any operand bundles except OB_funclet.
1661   if (auto *Call = dyn_cast<CallBase>(I))
1662     for (unsigned i = 0, e = Call->getNumOperandBundles(); i != e; ++i)
1663       if (Call->getOperandBundleAt(i).getTagID() != LLVMContext::OB_funclet)
1664         return false;
1665 
1666   DbgLoc = I->getDebugLoc();
1667 
1668   SavedInsertPt = FuncInfo.InsertPt;
1669 
1670   if (const auto *Call = dyn_cast<CallInst>(I)) {
1671     const Function *F = Call->getCalledFunction();
1672     LibFunc Func;
1673 
1674     // As a special case, don't handle calls to builtin library functions that
1675     // may be translated directly to target instructions.
1676     if (F && !F->hasLocalLinkage() && F->hasName() &&
1677         LibInfo->getLibFunc(F->getName(), Func) &&
1678         LibInfo->hasOptimizedCodeGen(Func))
1679       return false;
1680 
1681     // Don't handle Intrinsic::trap if a trap function is specified.
1682     if (F && F->getIntrinsicID() == Intrinsic::trap &&
1683         Call->hasFnAttr("trap-func-name"))
1684       return false;
1685   }
1686 
1687   // First, try doing target-independent selection.
1688   if (!SkipTargetIndependentISel) {
1689     if (selectOperator(I, I->getOpcode())) {
1690       ++NumFastIselSuccessIndependent;
1691       DbgLoc = DebugLoc();
1692       return true;
1693     }
1694     // Remove dead code.
1695     recomputeInsertPt();
1696     if (SavedInsertPt != FuncInfo.InsertPt)
1697       removeDeadCode(FuncInfo.InsertPt, SavedInsertPt);
1698     SavedInsertPt = FuncInfo.InsertPt;
1699   }
1700   // Next, try calling the target to attempt to handle the instruction.
1701   if (fastSelectInstruction(I)) {
1702     ++NumFastIselSuccessTarget;
1703     DbgLoc = DebugLoc();
1704     return true;
1705   }
1706   // Remove dead code.
1707   recomputeInsertPt();
1708   if (SavedInsertPt != FuncInfo.InsertPt)
1709     removeDeadCode(FuncInfo.InsertPt, SavedInsertPt);
1710 
1711   DbgLoc = DebugLoc();
1712   // Undo phi node updates, because they will be added again by SelectionDAG.
1713   if (I->isTerminator()) {
1714     // PHI node handling may have generated local value instructions.
1715     // We remove them because SelectionDAGISel will generate them again.
1716     removeDeadLocalValueCode(SavedLastLocalValue);
1717     FuncInfo.PHINodesToUpdate.resize(FuncInfo.OrigNumPHINodesToUpdate);
1718   }
1719   return false;
1720 }
1721 
1722 /// Emit an unconditional branch to the given block, unless it is the immediate
1723 /// (fall-through) successor, and update the CFG.
fastEmitBranch(MachineBasicBlock * MSucc,const DebugLoc & DbgLoc)1724 void FastISel::fastEmitBranch(MachineBasicBlock *MSucc,
1725                               const DebugLoc &DbgLoc) {
1726   if (FuncInfo.MBB->getBasicBlock()->sizeWithoutDebug() > 1 &&
1727       FuncInfo.MBB->isLayoutSuccessor(MSucc)) {
1728     // For more accurate line information if this is the only non-debug
1729     // instruction in the block then emit it, otherwise we have the
1730     // unconditional fall-through case, which needs no instructions.
1731   } else {
1732     // The unconditional branch case.
1733     TII.insertBranch(*FuncInfo.MBB, MSucc, nullptr,
1734                      SmallVector<MachineOperand, 0>(), DbgLoc);
1735   }
1736   if (FuncInfo.BPI) {
1737     auto BranchProbability = FuncInfo.BPI->getEdgeProbability(
1738         FuncInfo.MBB->getBasicBlock(), MSucc->getBasicBlock());
1739     FuncInfo.MBB->addSuccessor(MSucc, BranchProbability);
1740   } else
1741     FuncInfo.MBB->addSuccessorWithoutProb(MSucc);
1742 }
1743 
finishCondBranch(const BasicBlock * BranchBB,MachineBasicBlock * TrueMBB,MachineBasicBlock * FalseMBB)1744 void FastISel::finishCondBranch(const BasicBlock *BranchBB,
1745                                 MachineBasicBlock *TrueMBB,
1746                                 MachineBasicBlock *FalseMBB) {
1747   // Add TrueMBB as successor unless it is equal to the FalseMBB: This can
1748   // happen in degenerate IR and MachineIR forbids to have a block twice in the
1749   // successor/predecessor lists.
1750   if (TrueMBB != FalseMBB) {
1751     if (FuncInfo.BPI) {
1752       auto BranchProbability =
1753           FuncInfo.BPI->getEdgeProbability(BranchBB, TrueMBB->getBasicBlock());
1754       FuncInfo.MBB->addSuccessor(TrueMBB, BranchProbability);
1755     } else
1756       FuncInfo.MBB->addSuccessorWithoutProb(TrueMBB);
1757   }
1758 
1759   fastEmitBranch(FalseMBB, DbgLoc);
1760 }
1761 
1762 /// Emit an FNeg operation.
selectFNeg(const User * I,const Value * In)1763 bool FastISel::selectFNeg(const User *I, const Value *In) {
1764   Register OpReg = getRegForValue(In);
1765   if (!OpReg)
1766     return false;
1767   bool OpRegIsKill = hasTrivialKill(In);
1768 
1769   // If the target has ISD::FNEG, use it.
1770   EVT VT = TLI.getValueType(DL, I->getType());
1771   Register ResultReg = fastEmit_r(VT.getSimpleVT(), VT.getSimpleVT(), ISD::FNEG,
1772                                   OpReg, OpRegIsKill);
1773   if (ResultReg) {
1774     updateValueMap(I, ResultReg);
1775     return true;
1776   }
1777 
1778   // Bitcast the value to integer, twiddle the sign bit with xor,
1779   // and then bitcast it back to floating-point.
1780   if (VT.getSizeInBits() > 64)
1781     return false;
1782   EVT IntVT = EVT::getIntegerVT(I->getContext(), VT.getSizeInBits());
1783   if (!TLI.isTypeLegal(IntVT))
1784     return false;
1785 
1786   Register IntReg = fastEmit_r(VT.getSimpleVT(), IntVT.getSimpleVT(),
1787                                ISD::BITCAST, OpReg, OpRegIsKill);
1788   if (!IntReg)
1789     return false;
1790 
1791   Register IntResultReg = fastEmit_ri_(
1792       IntVT.getSimpleVT(), ISD::XOR, IntReg, /*IsKill=*/true,
1793       UINT64_C(1) << (VT.getSizeInBits() - 1), IntVT.getSimpleVT());
1794   if (!IntResultReg)
1795     return false;
1796 
1797   ResultReg = fastEmit_r(IntVT.getSimpleVT(), VT.getSimpleVT(), ISD::BITCAST,
1798                          IntResultReg, /*IsKill=*/true);
1799   if (!ResultReg)
1800     return false;
1801 
1802   updateValueMap(I, ResultReg);
1803   return true;
1804 }
1805 
selectExtractValue(const User * U)1806 bool FastISel::selectExtractValue(const User *U) {
1807   const ExtractValueInst *EVI = dyn_cast<ExtractValueInst>(U);
1808   if (!EVI)
1809     return false;
1810 
1811   // Make sure we only try to handle extracts with a legal result.  But also
1812   // allow i1 because it's easy.
1813   EVT RealVT = TLI.getValueType(DL, EVI->getType(), /*AllowUnknown=*/true);
1814   if (!RealVT.isSimple())
1815     return false;
1816   MVT VT = RealVT.getSimpleVT();
1817   if (!TLI.isTypeLegal(VT) && VT != MVT::i1)
1818     return false;
1819 
1820   const Value *Op0 = EVI->getOperand(0);
1821   Type *AggTy = Op0->getType();
1822 
1823   // Get the base result register.
1824   unsigned ResultReg;
1825   DenseMap<const Value *, Register>::iterator I = FuncInfo.ValueMap.find(Op0);
1826   if (I != FuncInfo.ValueMap.end())
1827     ResultReg = I->second;
1828   else if (isa<Instruction>(Op0))
1829     ResultReg = FuncInfo.InitializeRegForValue(Op0);
1830   else
1831     return false; // fast-isel can't handle aggregate constants at the moment
1832 
1833   // Get the actual result register, which is an offset from the base register.
1834   unsigned VTIndex = ComputeLinearIndex(AggTy, EVI->getIndices());
1835 
1836   SmallVector<EVT, 4> AggValueVTs;
1837   ComputeValueVTs(TLI, DL, AggTy, AggValueVTs);
1838 
1839   for (unsigned i = 0; i < VTIndex; i++)
1840     ResultReg += TLI.getNumRegisters(FuncInfo.Fn->getContext(), AggValueVTs[i]);
1841 
1842   updateValueMap(EVI, ResultReg);
1843   return true;
1844 }
1845 
selectOperator(const User * I,unsigned Opcode)1846 bool FastISel::selectOperator(const User *I, unsigned Opcode) {
1847   switch (Opcode) {
1848   case Instruction::Add:
1849     return selectBinaryOp(I, ISD::ADD);
1850   case Instruction::FAdd:
1851     return selectBinaryOp(I, ISD::FADD);
1852   case Instruction::Sub:
1853     return selectBinaryOp(I, ISD::SUB);
1854   case Instruction::FSub: {
1855     // FNeg is currently represented in LLVM IR as a special case of FSub.
1856     Value *X;
1857     if (match(I, m_FNeg(m_Value(X))))
1858        return selectFNeg(I, X);
1859     return selectBinaryOp(I, ISD::FSUB);
1860   }
1861   case Instruction::Mul:
1862     return selectBinaryOp(I, ISD::MUL);
1863   case Instruction::FMul:
1864     return selectBinaryOp(I, ISD::FMUL);
1865   case Instruction::SDiv:
1866     return selectBinaryOp(I, ISD::SDIV);
1867   case Instruction::UDiv:
1868     return selectBinaryOp(I, ISD::UDIV);
1869   case Instruction::FDiv:
1870     return selectBinaryOp(I, ISD::FDIV);
1871   case Instruction::SRem:
1872     return selectBinaryOp(I, ISD::SREM);
1873   case Instruction::URem:
1874     return selectBinaryOp(I, ISD::UREM);
1875   case Instruction::FRem:
1876     return selectBinaryOp(I, ISD::FREM);
1877   case Instruction::Shl:
1878     return selectBinaryOp(I, ISD::SHL);
1879   case Instruction::LShr:
1880     return selectBinaryOp(I, ISD::SRL);
1881   case Instruction::AShr:
1882     return selectBinaryOp(I, ISD::SRA);
1883   case Instruction::And:
1884     return selectBinaryOp(I, ISD::AND);
1885   case Instruction::Or:
1886     return selectBinaryOp(I, ISD::OR);
1887   case Instruction::Xor:
1888     return selectBinaryOp(I, ISD::XOR);
1889 
1890   case Instruction::FNeg:
1891     return selectFNeg(I, I->getOperand(0));
1892 
1893   case Instruction::GetElementPtr:
1894     return selectGetElementPtr(I);
1895 
1896   case Instruction::Br: {
1897     const BranchInst *BI = cast<BranchInst>(I);
1898 
1899     if (BI->isUnconditional()) {
1900       const BasicBlock *LLVMSucc = BI->getSuccessor(0);
1901       MachineBasicBlock *MSucc = FuncInfo.MBBMap[LLVMSucc];
1902       fastEmitBranch(MSucc, BI->getDebugLoc());
1903       return true;
1904     }
1905 
1906     // Conditional branches are not handed yet.
1907     // Halt "fast" selection and bail.
1908     return false;
1909   }
1910 
1911   case Instruction::Unreachable:
1912     if (TM.Options.TrapUnreachable)
1913       return fastEmit_(MVT::Other, MVT::Other, ISD::TRAP) != 0;
1914     else
1915       return true;
1916 
1917   case Instruction::Alloca:
1918     // FunctionLowering has the static-sized case covered.
1919     if (FuncInfo.StaticAllocaMap.count(cast<AllocaInst>(I)))
1920       return true;
1921 
1922     // Dynamic-sized alloca is not handled yet.
1923     return false;
1924 
1925   case Instruction::Call:
1926     // On AIX, call lowering uses the DAG-ISEL path currently so that the
1927     // callee of the direct function call instruction will be mapped to the
1928     // symbol for the function's entry point, which is distinct from the
1929     // function descriptor symbol. The latter is the symbol whose XCOFF symbol
1930     // name is the C-linkage name of the source level function.
1931     if (TM.getTargetTriple().isOSAIX())
1932       return false;
1933     return selectCall(I);
1934 
1935   case Instruction::BitCast:
1936     return selectBitCast(I);
1937 
1938   case Instruction::FPToSI:
1939     return selectCast(I, ISD::FP_TO_SINT);
1940   case Instruction::ZExt:
1941     return selectCast(I, ISD::ZERO_EXTEND);
1942   case Instruction::SExt:
1943     return selectCast(I, ISD::SIGN_EXTEND);
1944   case Instruction::Trunc:
1945     return selectCast(I, ISD::TRUNCATE);
1946   case Instruction::SIToFP:
1947     return selectCast(I, ISD::SINT_TO_FP);
1948 
1949   case Instruction::IntToPtr: // Deliberate fall-through.
1950   case Instruction::PtrToInt: {
1951     EVT SrcVT = TLI.getValueType(DL, I->getOperand(0)->getType());
1952     EVT DstVT = TLI.getValueType(DL, I->getType());
1953     if (DstVT.bitsGT(SrcVT))
1954       return selectCast(I, ISD::ZERO_EXTEND);
1955     if (DstVT.bitsLT(SrcVT))
1956       return selectCast(I, ISD::TRUNCATE);
1957     Register Reg = getRegForValue(I->getOperand(0));
1958     if (!Reg)
1959       return false;
1960     updateValueMap(I, Reg);
1961     return true;
1962   }
1963 
1964   case Instruction::ExtractValue:
1965     return selectExtractValue(I);
1966 
1967   case Instruction::Freeze:
1968     return selectFreeze(I);
1969 
1970   case Instruction::PHI:
1971     llvm_unreachable("FastISel shouldn't visit PHI nodes!");
1972 
1973   default:
1974     // Unhandled instruction. Halt "fast" selection and bail.
1975     return false;
1976   }
1977 }
1978 
FastISel(FunctionLoweringInfo & FuncInfo,const TargetLibraryInfo * LibInfo,bool SkipTargetIndependentISel)1979 FastISel::FastISel(FunctionLoweringInfo &FuncInfo,
1980                    const TargetLibraryInfo *LibInfo,
1981                    bool SkipTargetIndependentISel)
1982     : FuncInfo(FuncInfo), MF(FuncInfo.MF), MRI(FuncInfo.MF->getRegInfo()),
1983       MFI(FuncInfo.MF->getFrameInfo()), MCP(*FuncInfo.MF->getConstantPool()),
1984       TM(FuncInfo.MF->getTarget()), DL(MF->getDataLayout()),
1985       TII(*MF->getSubtarget().getInstrInfo()),
1986       TLI(*MF->getSubtarget().getTargetLowering()),
1987       TRI(*MF->getSubtarget().getRegisterInfo()), LibInfo(LibInfo),
1988       SkipTargetIndependentISel(SkipTargetIndependentISel),
1989       LastLocalValue(nullptr), EmitStartPt(nullptr) {}
1990 
1991 FastISel::~FastISel() = default;
1992 
fastLowerArguments()1993 bool FastISel::fastLowerArguments() { return false; }
1994 
fastLowerCall(CallLoweringInfo &)1995 bool FastISel::fastLowerCall(CallLoweringInfo & /*CLI*/) { return false; }
1996 
fastLowerIntrinsicCall(const IntrinsicInst *)1997 bool FastISel::fastLowerIntrinsicCall(const IntrinsicInst * /*II*/) {
1998   return false;
1999 }
2000 
fastEmit_(MVT,MVT,unsigned)2001 unsigned FastISel::fastEmit_(MVT, MVT, unsigned) { return 0; }
2002 
fastEmit_r(MVT,MVT,unsigned,unsigned,bool)2003 unsigned FastISel::fastEmit_r(MVT, MVT, unsigned, unsigned /*Op0*/,
2004                               bool /*Op0IsKill*/) {
2005   return 0;
2006 }
2007 
fastEmit_rr(MVT,MVT,unsigned,unsigned,bool,unsigned,bool)2008 unsigned FastISel::fastEmit_rr(MVT, MVT, unsigned, unsigned /*Op0*/,
2009                                bool /*Op0IsKill*/, unsigned /*Op1*/,
2010                                bool /*Op1IsKill*/) {
2011   return 0;
2012 }
2013 
fastEmit_i(MVT,MVT,unsigned,uint64_t)2014 unsigned FastISel::fastEmit_i(MVT, MVT, unsigned, uint64_t /*Imm*/) {
2015   return 0;
2016 }
2017 
fastEmit_f(MVT,MVT,unsigned,const ConstantFP *)2018 unsigned FastISel::fastEmit_f(MVT, MVT, unsigned,
2019                               const ConstantFP * /*FPImm*/) {
2020   return 0;
2021 }
2022 
fastEmit_ri(MVT,MVT,unsigned,unsigned,bool,uint64_t)2023 unsigned FastISel::fastEmit_ri(MVT, MVT, unsigned, unsigned /*Op0*/,
2024                                bool /*Op0IsKill*/, uint64_t /*Imm*/) {
2025   return 0;
2026 }
2027 
2028 /// This method is a wrapper of fastEmit_ri. It first tries to emit an
2029 /// instruction with an immediate operand using fastEmit_ri.
2030 /// If that fails, it materializes the immediate into a register and try
2031 /// fastEmit_rr instead.
fastEmit_ri_(MVT VT,unsigned Opcode,unsigned Op0,bool Op0IsKill,uint64_t Imm,MVT ImmType)2032 Register FastISel::fastEmit_ri_(MVT VT, unsigned Opcode, unsigned Op0,
2033                                 bool Op0IsKill, uint64_t Imm, MVT ImmType) {
2034   // If this is a multiply by a power of two, emit this as a shift left.
2035   if (Opcode == ISD::MUL && isPowerOf2_64(Imm)) {
2036     Opcode = ISD::SHL;
2037     Imm = Log2_64(Imm);
2038   } else if (Opcode == ISD::UDIV && isPowerOf2_64(Imm)) {
2039     // div x, 8 -> srl x, 3
2040     Opcode = ISD::SRL;
2041     Imm = Log2_64(Imm);
2042   }
2043 
2044   // Horrible hack (to be removed), check to make sure shift amounts are
2045   // in-range.
2046   if ((Opcode == ISD::SHL || Opcode == ISD::SRA || Opcode == ISD::SRL) &&
2047       Imm >= VT.getSizeInBits())
2048     return 0;
2049 
2050   // First check if immediate type is legal. If not, we can't use the ri form.
2051   Register ResultReg = fastEmit_ri(VT, VT, Opcode, Op0, Op0IsKill, Imm);
2052   if (ResultReg)
2053     return ResultReg;
2054   Register MaterialReg = fastEmit_i(ImmType, ImmType, ISD::Constant, Imm);
2055   bool IsImmKill = true;
2056   if (!MaterialReg) {
2057     // This is a bit ugly/slow, but failing here means falling out of
2058     // fast-isel, which would be very slow.
2059     IntegerType *ITy =
2060         IntegerType::get(FuncInfo.Fn->getContext(), VT.getSizeInBits());
2061     MaterialReg = getRegForValue(ConstantInt::get(ITy, Imm));
2062     if (!MaterialReg)
2063       return 0;
2064     // FIXME: If the materialized register here has no uses yet then this
2065     // will be the first use and we should be able to mark it as killed.
2066     // However, the local value area for materialising constant expressions
2067     // grows down, not up, which means that any constant expressions we generate
2068     // later which also use 'Imm' could be after this instruction and therefore
2069     // after this kill.
2070     IsImmKill = false;
2071   }
2072   return fastEmit_rr(VT, VT, Opcode, Op0, Op0IsKill, MaterialReg, IsImmKill);
2073 }
2074 
createResultReg(const TargetRegisterClass * RC)2075 Register FastISel::createResultReg(const TargetRegisterClass *RC) {
2076   return MRI.createVirtualRegister(RC);
2077 }
2078 
constrainOperandRegClass(const MCInstrDesc & II,Register Op,unsigned OpNum)2079 Register FastISel::constrainOperandRegClass(const MCInstrDesc &II, Register Op,
2080                                             unsigned OpNum) {
2081   if (Op.isVirtual()) {
2082     const TargetRegisterClass *RegClass =
2083         TII.getRegClass(II, OpNum, &TRI, *FuncInfo.MF);
2084     if (!MRI.constrainRegClass(Op, RegClass)) {
2085       // If it's not legal to COPY between the register classes, something
2086       // has gone very wrong before we got here.
2087       Register NewOp = createResultReg(RegClass);
2088       BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
2089               TII.get(TargetOpcode::COPY), NewOp).addReg(Op);
2090       return NewOp;
2091     }
2092   }
2093   return Op;
2094 }
2095 
fastEmitInst_(unsigned MachineInstOpcode,const TargetRegisterClass * RC)2096 Register FastISel::fastEmitInst_(unsigned MachineInstOpcode,
2097                                  const TargetRegisterClass *RC) {
2098   Register ResultReg = createResultReg(RC);
2099   const MCInstrDesc &II = TII.get(MachineInstOpcode);
2100 
2101   BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, II, ResultReg);
2102   return ResultReg;
2103 }
2104 
fastEmitInst_r(unsigned MachineInstOpcode,const TargetRegisterClass * RC,unsigned Op0,bool Op0IsKill)2105 Register FastISel::fastEmitInst_r(unsigned MachineInstOpcode,
2106                                   const TargetRegisterClass *RC, unsigned Op0,
2107                                   bool Op0IsKill) {
2108   const MCInstrDesc &II = TII.get(MachineInstOpcode);
2109 
2110   Register ResultReg = createResultReg(RC);
2111   Op0 = constrainOperandRegClass(II, Op0, II.getNumDefs());
2112 
2113   if (II.getNumDefs() >= 1)
2114     BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, II, ResultReg)
2115         .addReg(Op0, getKillRegState(Op0IsKill));
2116   else {
2117     BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, II)
2118         .addReg(Op0, getKillRegState(Op0IsKill));
2119     BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
2120             TII.get(TargetOpcode::COPY), ResultReg).addReg(II.ImplicitDefs[0]);
2121   }
2122 
2123   return ResultReg;
2124 }
2125 
fastEmitInst_rr(unsigned MachineInstOpcode,const TargetRegisterClass * RC,unsigned Op0,bool Op0IsKill,unsigned Op1,bool Op1IsKill)2126 Register FastISel::fastEmitInst_rr(unsigned MachineInstOpcode,
2127                                    const TargetRegisterClass *RC, unsigned Op0,
2128                                    bool Op0IsKill, unsigned Op1,
2129                                    bool Op1IsKill) {
2130   const MCInstrDesc &II = TII.get(MachineInstOpcode);
2131 
2132   Register ResultReg = createResultReg(RC);
2133   Op0 = constrainOperandRegClass(II, Op0, II.getNumDefs());
2134   Op1 = constrainOperandRegClass(II, Op1, II.getNumDefs() + 1);
2135 
2136   if (II.getNumDefs() >= 1)
2137     BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, II, ResultReg)
2138         .addReg(Op0, getKillRegState(Op0IsKill))
2139         .addReg(Op1, getKillRegState(Op1IsKill));
2140   else {
2141     BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, II)
2142         .addReg(Op0, getKillRegState(Op0IsKill))
2143         .addReg(Op1, getKillRegState(Op1IsKill));
2144     BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
2145             TII.get(TargetOpcode::COPY), ResultReg).addReg(II.ImplicitDefs[0]);
2146   }
2147   return ResultReg;
2148 }
2149 
fastEmitInst_rrr(unsigned MachineInstOpcode,const TargetRegisterClass * RC,unsigned Op0,bool Op0IsKill,unsigned Op1,bool Op1IsKill,unsigned Op2,bool Op2IsKill)2150 Register FastISel::fastEmitInst_rrr(unsigned MachineInstOpcode,
2151                                     const TargetRegisterClass *RC, unsigned Op0,
2152                                     bool Op0IsKill, unsigned Op1,
2153                                     bool Op1IsKill, unsigned Op2,
2154                                     bool Op2IsKill) {
2155   const MCInstrDesc &II = TII.get(MachineInstOpcode);
2156 
2157   Register ResultReg = createResultReg(RC);
2158   Op0 = constrainOperandRegClass(II, Op0, II.getNumDefs());
2159   Op1 = constrainOperandRegClass(II, Op1, II.getNumDefs() + 1);
2160   Op2 = constrainOperandRegClass(II, Op2, II.getNumDefs() + 2);
2161 
2162   if (II.getNumDefs() >= 1)
2163     BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, II, ResultReg)
2164         .addReg(Op0, getKillRegState(Op0IsKill))
2165         .addReg(Op1, getKillRegState(Op1IsKill))
2166         .addReg(Op2, getKillRegState(Op2IsKill));
2167   else {
2168     BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, II)
2169         .addReg(Op0, getKillRegState(Op0IsKill))
2170         .addReg(Op1, getKillRegState(Op1IsKill))
2171         .addReg(Op2, getKillRegState(Op2IsKill));
2172     BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
2173             TII.get(TargetOpcode::COPY), ResultReg).addReg(II.ImplicitDefs[0]);
2174   }
2175   return ResultReg;
2176 }
2177 
fastEmitInst_ri(unsigned MachineInstOpcode,const TargetRegisterClass * RC,unsigned Op0,bool Op0IsKill,uint64_t Imm)2178 Register FastISel::fastEmitInst_ri(unsigned MachineInstOpcode,
2179                                    const TargetRegisterClass *RC, unsigned Op0,
2180                                    bool Op0IsKill, uint64_t Imm) {
2181   const MCInstrDesc &II = TII.get(MachineInstOpcode);
2182 
2183   Register ResultReg = createResultReg(RC);
2184   Op0 = constrainOperandRegClass(II, Op0, II.getNumDefs());
2185 
2186   if (II.getNumDefs() >= 1)
2187     BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, II, ResultReg)
2188         .addReg(Op0, getKillRegState(Op0IsKill))
2189         .addImm(Imm);
2190   else {
2191     BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, II)
2192         .addReg(Op0, getKillRegState(Op0IsKill))
2193         .addImm(Imm);
2194     BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
2195             TII.get(TargetOpcode::COPY), ResultReg).addReg(II.ImplicitDefs[0]);
2196   }
2197   return ResultReg;
2198 }
2199 
fastEmitInst_rii(unsigned MachineInstOpcode,const TargetRegisterClass * RC,unsigned Op0,bool Op0IsKill,uint64_t Imm1,uint64_t Imm2)2200 Register FastISel::fastEmitInst_rii(unsigned MachineInstOpcode,
2201                                     const TargetRegisterClass *RC, unsigned Op0,
2202                                     bool Op0IsKill, uint64_t Imm1,
2203                                     uint64_t Imm2) {
2204   const MCInstrDesc &II = TII.get(MachineInstOpcode);
2205 
2206   Register ResultReg = createResultReg(RC);
2207   Op0 = constrainOperandRegClass(II, Op0, II.getNumDefs());
2208 
2209   if (II.getNumDefs() >= 1)
2210     BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, II, ResultReg)
2211         .addReg(Op0, getKillRegState(Op0IsKill))
2212         .addImm(Imm1)
2213         .addImm(Imm2);
2214   else {
2215     BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, II)
2216         .addReg(Op0, getKillRegState(Op0IsKill))
2217         .addImm(Imm1)
2218         .addImm(Imm2);
2219     BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
2220             TII.get(TargetOpcode::COPY), ResultReg).addReg(II.ImplicitDefs[0]);
2221   }
2222   return ResultReg;
2223 }
2224 
fastEmitInst_f(unsigned MachineInstOpcode,const TargetRegisterClass * RC,const ConstantFP * FPImm)2225 Register FastISel::fastEmitInst_f(unsigned MachineInstOpcode,
2226                                   const TargetRegisterClass *RC,
2227                                   const ConstantFP *FPImm) {
2228   const MCInstrDesc &II = TII.get(MachineInstOpcode);
2229 
2230   Register ResultReg = createResultReg(RC);
2231 
2232   if (II.getNumDefs() >= 1)
2233     BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, II, ResultReg)
2234         .addFPImm(FPImm);
2235   else {
2236     BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, II)
2237         .addFPImm(FPImm);
2238     BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
2239             TII.get(TargetOpcode::COPY), ResultReg).addReg(II.ImplicitDefs[0]);
2240   }
2241   return ResultReg;
2242 }
2243 
fastEmitInst_rri(unsigned MachineInstOpcode,const TargetRegisterClass * RC,unsigned Op0,bool Op0IsKill,unsigned Op1,bool Op1IsKill,uint64_t Imm)2244 Register FastISel::fastEmitInst_rri(unsigned MachineInstOpcode,
2245                                     const TargetRegisterClass *RC, unsigned Op0,
2246                                     bool Op0IsKill, unsigned Op1,
2247                                     bool Op1IsKill, uint64_t Imm) {
2248   const MCInstrDesc &II = TII.get(MachineInstOpcode);
2249 
2250   Register ResultReg = createResultReg(RC);
2251   Op0 = constrainOperandRegClass(II, Op0, II.getNumDefs());
2252   Op1 = constrainOperandRegClass(II, Op1, II.getNumDefs() + 1);
2253 
2254   if (II.getNumDefs() >= 1)
2255     BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, II, ResultReg)
2256         .addReg(Op0, getKillRegState(Op0IsKill))
2257         .addReg(Op1, getKillRegState(Op1IsKill))
2258         .addImm(Imm);
2259   else {
2260     BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, II)
2261         .addReg(Op0, getKillRegState(Op0IsKill))
2262         .addReg(Op1, getKillRegState(Op1IsKill))
2263         .addImm(Imm);
2264     BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
2265             TII.get(TargetOpcode::COPY), ResultReg).addReg(II.ImplicitDefs[0]);
2266   }
2267   return ResultReg;
2268 }
2269 
fastEmitInst_i(unsigned MachineInstOpcode,const TargetRegisterClass * RC,uint64_t Imm)2270 Register FastISel::fastEmitInst_i(unsigned MachineInstOpcode,
2271                                   const TargetRegisterClass *RC, uint64_t Imm) {
2272   Register ResultReg = createResultReg(RC);
2273   const MCInstrDesc &II = TII.get(MachineInstOpcode);
2274 
2275   if (II.getNumDefs() >= 1)
2276     BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, II, ResultReg)
2277         .addImm(Imm);
2278   else {
2279     BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, II).addImm(Imm);
2280     BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
2281             TII.get(TargetOpcode::COPY), ResultReg).addReg(II.ImplicitDefs[0]);
2282   }
2283   return ResultReg;
2284 }
2285 
fastEmitInst_extractsubreg(MVT RetVT,unsigned Op0,bool Op0IsKill,uint32_t Idx)2286 Register FastISel::fastEmitInst_extractsubreg(MVT RetVT, unsigned Op0,
2287                                               bool Op0IsKill, uint32_t Idx) {
2288   Register ResultReg = createResultReg(TLI.getRegClassFor(RetVT));
2289   assert(Register::isVirtualRegister(Op0) &&
2290          "Cannot yet extract from physregs");
2291   const TargetRegisterClass *RC = MRI.getRegClass(Op0);
2292   MRI.constrainRegClass(Op0, TRI.getSubClassWithSubReg(RC, Idx));
2293   BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(TargetOpcode::COPY),
2294           ResultReg).addReg(Op0, getKillRegState(Op0IsKill), Idx);
2295   return ResultReg;
2296 }
2297 
2298 /// Emit MachineInstrs to compute the value of Op with all but the least
2299 /// significant bit set to zero.
fastEmitZExtFromI1(MVT VT,unsigned Op0,bool Op0IsKill)2300 Register FastISel::fastEmitZExtFromI1(MVT VT, unsigned Op0, bool Op0IsKill) {
2301   return fastEmit_ri(VT, VT, ISD::AND, Op0, Op0IsKill, 1);
2302 }
2303 
2304 /// HandlePHINodesInSuccessorBlocks - Handle PHI nodes in successor blocks.
2305 /// Emit code to ensure constants are copied into registers when needed.
2306 /// Remember the virtual registers that need to be added to the Machine PHI
2307 /// nodes as input.  We cannot just directly add them, because expansion
2308 /// might result in multiple MBB's for one BB.  As such, the start of the
2309 /// BB might correspond to a different MBB than the end.
handlePHINodesInSuccessorBlocks(const BasicBlock * LLVMBB)2310 bool FastISel::handlePHINodesInSuccessorBlocks(const BasicBlock *LLVMBB) {
2311   const Instruction *TI = LLVMBB->getTerminator();
2312 
2313   SmallPtrSet<MachineBasicBlock *, 4> SuccsHandled;
2314   FuncInfo.OrigNumPHINodesToUpdate = FuncInfo.PHINodesToUpdate.size();
2315 
2316   // Check successor nodes' PHI nodes that expect a constant to be available
2317   // from this block.
2318   for (unsigned succ = 0, e = TI->getNumSuccessors(); succ != e; ++succ) {
2319     const BasicBlock *SuccBB = TI->getSuccessor(succ);
2320     if (!isa<PHINode>(SuccBB->begin()))
2321       continue;
2322     MachineBasicBlock *SuccMBB = FuncInfo.MBBMap[SuccBB];
2323 
2324     // If this terminator has multiple identical successors (common for
2325     // switches), only handle each succ once.
2326     if (!SuccsHandled.insert(SuccMBB).second)
2327       continue;
2328 
2329     MachineBasicBlock::iterator MBBI = SuccMBB->begin();
2330 
2331     // At this point we know that there is a 1-1 correspondence between LLVM PHI
2332     // nodes and Machine PHI nodes, but the incoming operands have not been
2333     // emitted yet.
2334     for (const PHINode &PN : SuccBB->phis()) {
2335       // Ignore dead phi's.
2336       if (PN.use_empty())
2337         continue;
2338 
2339       // Only handle legal types. Two interesting things to note here. First,
2340       // by bailing out early, we may leave behind some dead instructions,
2341       // since SelectionDAG's HandlePHINodesInSuccessorBlocks will insert its
2342       // own moves. Second, this check is necessary because FastISel doesn't
2343       // use CreateRegs to create registers, so it always creates
2344       // exactly one register for each non-void instruction.
2345       EVT VT = TLI.getValueType(DL, PN.getType(), /*AllowUnknown=*/true);
2346       if (VT == MVT::Other || !TLI.isTypeLegal(VT)) {
2347         // Handle integer promotions, though, because they're common and easy.
2348         if (!(VT == MVT::i1 || VT == MVT::i8 || VT == MVT::i16)) {
2349           FuncInfo.PHINodesToUpdate.resize(FuncInfo.OrigNumPHINodesToUpdate);
2350           return false;
2351         }
2352       }
2353 
2354       const Value *PHIOp = PN.getIncomingValueForBlock(LLVMBB);
2355 
2356       // Set the DebugLoc for the copy. Prefer the location of the operand
2357       // if there is one; use the location of the PHI otherwise.
2358       DbgLoc = PN.getDebugLoc();
2359       if (const auto *Inst = dyn_cast<Instruction>(PHIOp))
2360         DbgLoc = Inst->getDebugLoc();
2361 
2362       Register Reg = getRegForValue(PHIOp);
2363       if (!Reg) {
2364         FuncInfo.PHINodesToUpdate.resize(FuncInfo.OrigNumPHINodesToUpdate);
2365         return false;
2366       }
2367       FuncInfo.PHINodesToUpdate.push_back(std::make_pair(&*MBBI++, Reg));
2368       DbgLoc = DebugLoc();
2369     }
2370   }
2371 
2372   return true;
2373 }
2374 
tryToFoldLoad(const LoadInst * LI,const Instruction * FoldInst)2375 bool FastISel::tryToFoldLoad(const LoadInst *LI, const Instruction *FoldInst) {
2376   assert(LI->hasOneUse() &&
2377          "tryToFoldLoad expected a LoadInst with a single use");
2378   // We know that the load has a single use, but don't know what it is.  If it
2379   // isn't one of the folded instructions, then we can't succeed here.  Handle
2380   // this by scanning the single-use users of the load until we get to FoldInst.
2381   unsigned MaxUsers = 6; // Don't scan down huge single-use chains of instrs.
2382 
2383   const Instruction *TheUser = LI->user_back();
2384   while (TheUser != FoldInst && // Scan up until we find FoldInst.
2385          // Stay in the right block.
2386          TheUser->getParent() == FoldInst->getParent() &&
2387          --MaxUsers) { // Don't scan too far.
2388     // If there are multiple or no uses of this instruction, then bail out.
2389     if (!TheUser->hasOneUse())
2390       return false;
2391 
2392     TheUser = TheUser->user_back();
2393   }
2394 
2395   // If we didn't find the fold instruction, then we failed to collapse the
2396   // sequence.
2397   if (TheUser != FoldInst)
2398     return false;
2399 
2400   // Don't try to fold volatile loads.  Target has to deal with alignment
2401   // constraints.
2402   if (LI->isVolatile())
2403     return false;
2404 
2405   // Figure out which vreg this is going into.  If there is no assigned vreg yet
2406   // then there actually was no reference to it.  Perhaps the load is referenced
2407   // by a dead instruction.
2408   Register LoadReg = getRegForValue(LI);
2409   if (!LoadReg)
2410     return false;
2411 
2412   // We can't fold if this vreg has no uses or more than one use.  Multiple uses
2413   // may mean that the instruction got lowered to multiple MIs, or the use of
2414   // the loaded value ended up being multiple operands of the result.
2415   if (!MRI.hasOneUse(LoadReg))
2416     return false;
2417 
2418   MachineRegisterInfo::reg_iterator RI = MRI.reg_begin(LoadReg);
2419   MachineInstr *User = RI->getParent();
2420 
2421   // Set the insertion point properly.  Folding the load can cause generation of
2422   // other random instructions (like sign extends) for addressing modes; make
2423   // sure they get inserted in a logical place before the new instruction.
2424   FuncInfo.InsertPt = User;
2425   FuncInfo.MBB = User->getParent();
2426 
2427   // Ask the target to try folding the load.
2428   return tryToFoldLoadIntoMI(User, RI.getOperandNo(), LI);
2429 }
2430 
canFoldAddIntoGEP(const User * GEP,const Value * Add)2431 bool FastISel::canFoldAddIntoGEP(const User *GEP, const Value *Add) {
2432   // Must be an add.
2433   if (!isa<AddOperator>(Add))
2434     return false;
2435   // Type size needs to match.
2436   if (DL.getTypeSizeInBits(GEP->getType()) !=
2437       DL.getTypeSizeInBits(Add->getType()))
2438     return false;
2439   // Must be in the same basic block.
2440   if (isa<Instruction>(Add) &&
2441       FuncInfo.MBBMap[cast<Instruction>(Add)->getParent()] != FuncInfo.MBB)
2442     return false;
2443   // Must have a constant operand.
2444   return isa<ConstantInt>(cast<AddOperator>(Add)->getOperand(1));
2445 }
2446 
2447 MachineMemOperand *
createMachineMemOperandFor(const Instruction * I) const2448 FastISel::createMachineMemOperandFor(const Instruction *I) const {
2449   const Value *Ptr;
2450   Type *ValTy;
2451   MaybeAlign Alignment;
2452   MachineMemOperand::Flags Flags;
2453   bool IsVolatile;
2454 
2455   if (const auto *LI = dyn_cast<LoadInst>(I)) {
2456     Alignment = LI->getAlign();
2457     IsVolatile = LI->isVolatile();
2458     Flags = MachineMemOperand::MOLoad;
2459     Ptr = LI->getPointerOperand();
2460     ValTy = LI->getType();
2461   } else if (const auto *SI = dyn_cast<StoreInst>(I)) {
2462     Alignment = SI->getAlign();
2463     IsVolatile = SI->isVolatile();
2464     Flags = MachineMemOperand::MOStore;
2465     Ptr = SI->getPointerOperand();
2466     ValTy = SI->getValueOperand()->getType();
2467   } else
2468     return nullptr;
2469 
2470   bool IsNonTemporal = I->hasMetadata(LLVMContext::MD_nontemporal);
2471   bool IsInvariant = I->hasMetadata(LLVMContext::MD_invariant_load);
2472   bool IsDereferenceable = I->hasMetadata(LLVMContext::MD_dereferenceable);
2473   const MDNode *Ranges = I->getMetadata(LLVMContext::MD_range);
2474 
2475   AAMDNodes AAInfo;
2476   I->getAAMetadata(AAInfo);
2477 
2478   if (!Alignment) // Ensure that codegen never sees alignment 0.
2479     Alignment = DL.getABITypeAlign(ValTy);
2480 
2481   unsigned Size = DL.getTypeStoreSize(ValTy);
2482 
2483   if (IsVolatile)
2484     Flags |= MachineMemOperand::MOVolatile;
2485   if (IsNonTemporal)
2486     Flags |= MachineMemOperand::MONonTemporal;
2487   if (IsDereferenceable)
2488     Flags |= MachineMemOperand::MODereferenceable;
2489   if (IsInvariant)
2490     Flags |= MachineMemOperand::MOInvariant;
2491 
2492   return FuncInfo.MF->getMachineMemOperand(MachinePointerInfo(Ptr), Flags, Size,
2493                                            *Alignment, AAInfo, Ranges);
2494 }
2495 
optimizeCmpPredicate(const CmpInst * CI) const2496 CmpInst::Predicate FastISel::optimizeCmpPredicate(const CmpInst *CI) const {
2497   // If both operands are the same, then try to optimize or fold the cmp.
2498   CmpInst::Predicate Predicate = CI->getPredicate();
2499   if (CI->getOperand(0) != CI->getOperand(1))
2500     return Predicate;
2501 
2502   switch (Predicate) {
2503   default: llvm_unreachable("Invalid predicate!");
2504   case CmpInst::FCMP_FALSE: Predicate = CmpInst::FCMP_FALSE; break;
2505   case CmpInst::FCMP_OEQ:   Predicate = CmpInst::FCMP_ORD;   break;
2506   case CmpInst::FCMP_OGT:   Predicate = CmpInst::FCMP_FALSE; break;
2507   case CmpInst::FCMP_OGE:   Predicate = CmpInst::FCMP_ORD;   break;
2508   case CmpInst::FCMP_OLT:   Predicate = CmpInst::FCMP_FALSE; break;
2509   case CmpInst::FCMP_OLE:   Predicate = CmpInst::FCMP_ORD;   break;
2510   case CmpInst::FCMP_ONE:   Predicate = CmpInst::FCMP_FALSE; break;
2511   case CmpInst::FCMP_ORD:   Predicate = CmpInst::FCMP_ORD;   break;
2512   case CmpInst::FCMP_UNO:   Predicate = CmpInst::FCMP_UNO;   break;
2513   case CmpInst::FCMP_UEQ:   Predicate = CmpInst::FCMP_TRUE;  break;
2514   case CmpInst::FCMP_UGT:   Predicate = CmpInst::FCMP_UNO;   break;
2515   case CmpInst::FCMP_UGE:   Predicate = CmpInst::FCMP_TRUE;  break;
2516   case CmpInst::FCMP_ULT:   Predicate = CmpInst::FCMP_UNO;   break;
2517   case CmpInst::FCMP_ULE:   Predicate = CmpInst::FCMP_TRUE;  break;
2518   case CmpInst::FCMP_UNE:   Predicate = CmpInst::FCMP_UNO;   break;
2519   case CmpInst::FCMP_TRUE:  Predicate = CmpInst::FCMP_TRUE;  break;
2520 
2521   case CmpInst::ICMP_EQ:    Predicate = CmpInst::FCMP_TRUE;  break;
2522   case CmpInst::ICMP_NE:    Predicate = CmpInst::FCMP_FALSE; break;
2523   case CmpInst::ICMP_UGT:   Predicate = CmpInst::FCMP_FALSE; break;
2524   case CmpInst::ICMP_UGE:   Predicate = CmpInst::FCMP_TRUE;  break;
2525   case CmpInst::ICMP_ULT:   Predicate = CmpInst::FCMP_FALSE; break;
2526   case CmpInst::ICMP_ULE:   Predicate = CmpInst::FCMP_TRUE;  break;
2527   case CmpInst::ICMP_SGT:   Predicate = CmpInst::FCMP_FALSE; break;
2528   case CmpInst::ICMP_SGE:   Predicate = CmpInst::FCMP_TRUE;  break;
2529   case CmpInst::ICMP_SLT:   Predicate = CmpInst::FCMP_FALSE; break;
2530   case CmpInst::ICMP_SLE:   Predicate = CmpInst::FCMP_TRUE;  break;
2531   }
2532 
2533   return Predicate;
2534 }
2535