xref: /freebsd/contrib/llvm-project/llvm/lib/Target/X86/X86FlagsCopyLowering.cpp (revision d37eb51047221dc3322b34db1038ff3aa533883f)
1 //====- X86FlagsCopyLowering.cpp - Lowers COPY nodes of EFLAGS ------------===//
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 /// \file
9 ///
10 /// Lowers COPY nodes of EFLAGS by directly extracting and preserving individual
11 /// flag bits.
12 ///
13 /// We have to do this by carefully analyzing and rewriting the usage of the
14 /// copied EFLAGS register because there is no general way to rematerialize the
15 /// entire EFLAGS register safely and efficiently. Using `popf` both forces
16 /// dynamic stack adjustment and can create correctness issues due to IF, TF,
17 /// and other non-status flags being overwritten. Using sequences involving
18 /// SAHF don't work on all x86 processors and are often quite slow compared to
19 /// directly testing a single status preserved in its own GPR.
20 ///
21 //===----------------------------------------------------------------------===//
22 
23 #include "X86.h"
24 #include "X86InstrBuilder.h"
25 #include "X86InstrInfo.h"
26 #include "X86Subtarget.h"
27 #include "llvm/ADT/PostOrderIterator.h"
28 #include "llvm/ADT/STLExtras.h"
29 #include "llvm/ADT/ScopeExit.h"
30 #include "llvm/ADT/SmallPtrSet.h"
31 #include "llvm/ADT/SmallVector.h"
32 #include "llvm/ADT/Statistic.h"
33 #include "llvm/CodeGen/MachineBasicBlock.h"
34 #include "llvm/CodeGen/MachineConstantPool.h"
35 #include "llvm/CodeGen/MachineDominators.h"
36 #include "llvm/CodeGen/MachineFunction.h"
37 #include "llvm/CodeGen/MachineFunctionPass.h"
38 #include "llvm/CodeGen/MachineInstr.h"
39 #include "llvm/CodeGen/MachineInstrBuilder.h"
40 #include "llvm/CodeGen/MachineModuleInfo.h"
41 #include "llvm/CodeGen/MachineOperand.h"
42 #include "llvm/CodeGen/MachineRegisterInfo.h"
43 #include "llvm/CodeGen/MachineSSAUpdater.h"
44 #include "llvm/CodeGen/TargetInstrInfo.h"
45 #include "llvm/CodeGen/TargetRegisterInfo.h"
46 #include "llvm/CodeGen/TargetSchedule.h"
47 #include "llvm/CodeGen/TargetSubtargetInfo.h"
48 #include "llvm/IR/DebugLoc.h"
49 #include "llvm/MC/MCSchedule.h"
50 #include "llvm/Pass.h"
51 #include "llvm/Support/CommandLine.h"
52 #include "llvm/Support/Debug.h"
53 #include "llvm/Support/raw_ostream.h"
54 #include <algorithm>
55 #include <cassert>
56 #include <iterator>
57 #include <utility>
58 
59 using namespace llvm;
60 
61 #define PASS_KEY "x86-flags-copy-lowering"
62 #define DEBUG_TYPE PASS_KEY
63 
64 STATISTIC(NumCopiesEliminated, "Number of copies of EFLAGS eliminated");
65 STATISTIC(NumSetCCsInserted, "Number of setCC instructions inserted");
66 STATISTIC(NumTestsInserted, "Number of test instructions inserted");
67 STATISTIC(NumAddsInserted, "Number of adds instructions inserted");
68 
69 namespace {
70 
71 // Convenient array type for storing registers associated with each condition.
72 using CondRegArray = std::array<unsigned, X86::LAST_VALID_COND + 1>;
73 
74 class X86FlagsCopyLoweringPass : public MachineFunctionPass {
75 public:
76   X86FlagsCopyLoweringPass() : MachineFunctionPass(ID) { }
77 
78   StringRef getPassName() const override { return "X86 EFLAGS copy lowering"; }
79   bool runOnMachineFunction(MachineFunction &MF) override;
80   void getAnalysisUsage(AnalysisUsage &AU) const override;
81 
82   /// Pass identification, replacement for typeid.
83   static char ID;
84 
85 private:
86   MachineRegisterInfo *MRI = nullptr;
87   const X86Subtarget *Subtarget = nullptr;
88   const X86InstrInfo *TII = nullptr;
89   const TargetRegisterInfo *TRI = nullptr;
90   const TargetRegisterClass *PromoteRC = nullptr;
91   MachineDominatorTree *MDT = nullptr;
92 
93   CondRegArray collectCondsInRegs(MachineBasicBlock &MBB,
94                                   MachineBasicBlock::iterator CopyDefI);
95 
96   Register promoteCondToReg(MachineBasicBlock &MBB,
97                             MachineBasicBlock::iterator TestPos,
98                             const DebugLoc &TestLoc, X86::CondCode Cond);
99   std::pair<unsigned, bool> getCondOrInverseInReg(
100       MachineBasicBlock &TestMBB, MachineBasicBlock::iterator TestPos,
101       const DebugLoc &TestLoc, X86::CondCode Cond, CondRegArray &CondRegs);
102   void insertTest(MachineBasicBlock &MBB, MachineBasicBlock::iterator Pos,
103                   const DebugLoc &Loc, unsigned Reg);
104 
105   void rewriteArithmetic(MachineBasicBlock &TestMBB,
106                          MachineBasicBlock::iterator TestPos,
107                          const DebugLoc &TestLoc, MachineInstr &MI,
108                          MachineOperand &FlagUse, CondRegArray &CondRegs);
109   void rewriteCMov(MachineBasicBlock &TestMBB,
110                    MachineBasicBlock::iterator TestPos, const DebugLoc &TestLoc,
111                    MachineInstr &CMovI, MachineOperand &FlagUse,
112                    CondRegArray &CondRegs);
113   void rewriteFCMov(MachineBasicBlock &TestMBB,
114                     MachineBasicBlock::iterator TestPos,
115                     const DebugLoc &TestLoc, MachineInstr &CMovI,
116                     MachineOperand &FlagUse, CondRegArray &CondRegs);
117   void rewriteCondJmp(MachineBasicBlock &TestMBB,
118                       MachineBasicBlock::iterator TestPos,
119                       const DebugLoc &TestLoc, MachineInstr &JmpI,
120                       CondRegArray &CondRegs);
121   void rewriteCopy(MachineInstr &MI, MachineOperand &FlagUse,
122                    MachineInstr &CopyDefI);
123   void rewriteSetCC(MachineBasicBlock &TestMBB,
124                     MachineBasicBlock::iterator TestPos,
125                     const DebugLoc &TestLoc, MachineInstr &SetCCI,
126                     MachineOperand &FlagUse, CondRegArray &CondRegs);
127 };
128 
129 } // end anonymous namespace
130 
131 INITIALIZE_PASS_BEGIN(X86FlagsCopyLoweringPass, DEBUG_TYPE,
132                       "X86 EFLAGS copy lowering", false, false)
133 INITIALIZE_PASS_END(X86FlagsCopyLoweringPass, DEBUG_TYPE,
134                     "X86 EFLAGS copy lowering", false, false)
135 
136 FunctionPass *llvm::createX86FlagsCopyLoweringPass() {
137   return new X86FlagsCopyLoweringPass();
138 }
139 
140 char X86FlagsCopyLoweringPass::ID = 0;
141 
142 void X86FlagsCopyLoweringPass::getAnalysisUsage(AnalysisUsage &AU) const {
143   AU.addRequired<MachineDominatorTree>();
144   MachineFunctionPass::getAnalysisUsage(AU);
145 }
146 
147 namespace {
148 /// An enumeration of the arithmetic instruction mnemonics which have
149 /// interesting flag semantics.
150 ///
151 /// We can map instruction opcodes into these mnemonics to make it easy to
152 /// dispatch with specific functionality.
153 enum class FlagArithMnemonic {
154   ADC,
155   RCL,
156   RCR,
157   SBB,
158   SETB,
159 };
160 } // namespace
161 
162 static FlagArithMnemonic getMnemonicFromOpcode(unsigned Opcode) {
163   switch (Opcode) {
164   default:
165     report_fatal_error("No support for lowering a copy into EFLAGS when used "
166                        "by this instruction!");
167 
168 #define LLVM_EXPAND_INSTR_SIZES(MNEMONIC, SUFFIX)                              \
169   case X86::MNEMONIC##8##SUFFIX:                                               \
170   case X86::MNEMONIC##16##SUFFIX:                                              \
171   case X86::MNEMONIC##32##SUFFIX:                                              \
172   case X86::MNEMONIC##64##SUFFIX:
173 
174 #define LLVM_EXPAND_ADC_SBB_INSTR(MNEMONIC)                                    \
175   LLVM_EXPAND_INSTR_SIZES(MNEMONIC, rr)                                        \
176   LLVM_EXPAND_INSTR_SIZES(MNEMONIC, rr_REV)                                    \
177   LLVM_EXPAND_INSTR_SIZES(MNEMONIC, rm)                                        \
178   LLVM_EXPAND_INSTR_SIZES(MNEMONIC, mr)                                        \
179   case X86::MNEMONIC##8ri:                                                     \
180   case X86::MNEMONIC##16ri8:                                                   \
181   case X86::MNEMONIC##32ri8:                                                   \
182   case X86::MNEMONIC##64ri8:                                                   \
183   case X86::MNEMONIC##16ri:                                                    \
184   case X86::MNEMONIC##32ri:                                                    \
185   case X86::MNEMONIC##64ri32:                                                  \
186   case X86::MNEMONIC##8mi:                                                     \
187   case X86::MNEMONIC##16mi8:                                                   \
188   case X86::MNEMONIC##32mi8:                                                   \
189   case X86::MNEMONIC##64mi8:                                                   \
190   case X86::MNEMONIC##16mi:                                                    \
191   case X86::MNEMONIC##32mi:                                                    \
192   case X86::MNEMONIC##64mi32:                                                  \
193   case X86::MNEMONIC##8i8:                                                     \
194   case X86::MNEMONIC##16i16:                                                   \
195   case X86::MNEMONIC##32i32:                                                   \
196   case X86::MNEMONIC##64i32:
197 
198     LLVM_EXPAND_ADC_SBB_INSTR(ADC)
199     return FlagArithMnemonic::ADC;
200 
201     LLVM_EXPAND_ADC_SBB_INSTR(SBB)
202     return FlagArithMnemonic::SBB;
203 
204 #undef LLVM_EXPAND_ADC_SBB_INSTR
205 
206     LLVM_EXPAND_INSTR_SIZES(RCL, rCL)
207     LLVM_EXPAND_INSTR_SIZES(RCL, r1)
208     LLVM_EXPAND_INSTR_SIZES(RCL, ri)
209     return FlagArithMnemonic::RCL;
210 
211     LLVM_EXPAND_INSTR_SIZES(RCR, rCL)
212     LLVM_EXPAND_INSTR_SIZES(RCR, r1)
213     LLVM_EXPAND_INSTR_SIZES(RCR, ri)
214     return FlagArithMnemonic::RCR;
215 
216 #undef LLVM_EXPAND_INSTR_SIZES
217 
218   case X86::SETB_C32r:
219   case X86::SETB_C64r:
220     return FlagArithMnemonic::SETB;
221   }
222 }
223 
224 static MachineBasicBlock &splitBlock(MachineBasicBlock &MBB,
225                                      MachineInstr &SplitI,
226                                      const X86InstrInfo &TII) {
227   MachineFunction &MF = *MBB.getParent();
228 
229   assert(SplitI.getParent() == &MBB &&
230          "Split instruction must be in the split block!");
231   assert(SplitI.isBranch() &&
232          "Only designed to split a tail of branch instructions!");
233   assert(X86::getCondFromBranch(SplitI) != X86::COND_INVALID &&
234          "Must split on an actual jCC instruction!");
235 
236   // Dig out the previous instruction to the split point.
237   MachineInstr &PrevI = *std::prev(SplitI.getIterator());
238   assert(PrevI.isBranch() && "Must split after a branch!");
239   assert(X86::getCondFromBranch(PrevI) != X86::COND_INVALID &&
240          "Must split after an actual jCC instruction!");
241   assert(!std::prev(PrevI.getIterator())->isTerminator() &&
242          "Must only have this one terminator prior to the split!");
243 
244   // Grab the one successor edge that will stay in `MBB`.
245   MachineBasicBlock &UnsplitSucc = *PrevI.getOperand(0).getMBB();
246 
247   // Analyze the original block to see if we are actually splitting an edge
248   // into two edges. This can happen when we have multiple conditional jumps to
249   // the same successor.
250   bool IsEdgeSplit =
251       std::any_of(SplitI.getIterator(), MBB.instr_end(),
252                   [&](MachineInstr &MI) {
253                     assert(MI.isTerminator() &&
254                            "Should only have spliced terminators!");
255                     return llvm::any_of(
256                         MI.operands(), [&](MachineOperand &MOp) {
257                           return MOp.isMBB() && MOp.getMBB() == &UnsplitSucc;
258                         });
259                   }) ||
260       MBB.getFallThrough() == &UnsplitSucc;
261 
262   MachineBasicBlock &NewMBB = *MF.CreateMachineBasicBlock();
263 
264   // Insert the new block immediately after the current one. Any existing
265   // fallthrough will be sunk into this new block anyways.
266   MF.insert(std::next(MachineFunction::iterator(&MBB)), &NewMBB);
267 
268   // Splice the tail of instructions into the new block.
269   NewMBB.splice(NewMBB.end(), &MBB, SplitI.getIterator(), MBB.end());
270 
271   // Copy the necessary succesors (and their probability info) into the new
272   // block.
273   for (auto SI = MBB.succ_begin(), SE = MBB.succ_end(); SI != SE; ++SI)
274     if (IsEdgeSplit || *SI != &UnsplitSucc)
275       NewMBB.copySuccessor(&MBB, SI);
276   // Normalize the probabilities if we didn't end up splitting the edge.
277   if (!IsEdgeSplit)
278     NewMBB.normalizeSuccProbs();
279 
280   // Now replace all of the moved successors in the original block with the new
281   // block. This will merge their probabilities.
282   for (MachineBasicBlock *Succ : NewMBB.successors())
283     if (Succ != &UnsplitSucc)
284       MBB.replaceSuccessor(Succ, &NewMBB);
285 
286   // We should always end up replacing at least one successor.
287   assert(MBB.isSuccessor(&NewMBB) &&
288          "Failed to make the new block a successor!");
289 
290   // Now update all the PHIs.
291   for (MachineBasicBlock *Succ : NewMBB.successors()) {
292     for (MachineInstr &MI : *Succ) {
293       if (!MI.isPHI())
294         break;
295 
296       for (int OpIdx = 1, NumOps = MI.getNumOperands(); OpIdx < NumOps;
297            OpIdx += 2) {
298         MachineOperand &OpV = MI.getOperand(OpIdx);
299         MachineOperand &OpMBB = MI.getOperand(OpIdx + 1);
300         assert(OpMBB.isMBB() && "Block operand to a PHI is not a block!");
301         if (OpMBB.getMBB() != &MBB)
302           continue;
303 
304         // Replace the operand for unsplit successors
305         if (!IsEdgeSplit || Succ != &UnsplitSucc) {
306           OpMBB.setMBB(&NewMBB);
307 
308           // We have to continue scanning as there may be multiple entries in
309           // the PHI.
310           continue;
311         }
312 
313         // When we have split the edge append a new successor.
314         MI.addOperand(MF, OpV);
315         MI.addOperand(MF, MachineOperand::CreateMBB(&NewMBB));
316         break;
317       }
318     }
319   }
320 
321   return NewMBB;
322 }
323 
324 static X86::CondCode getCondFromFCMOV(unsigned Opcode) {
325   switch (Opcode) {
326   default: return X86::COND_INVALID;
327   case X86::CMOVBE_Fp32:  case X86::CMOVBE_Fp64:  case X86::CMOVBE_Fp80:
328     return X86::COND_BE;
329   case X86::CMOVB_Fp32:   case X86::CMOVB_Fp64:   case X86::CMOVB_Fp80:
330     return X86::COND_B;
331   case X86::CMOVE_Fp32:   case X86::CMOVE_Fp64:   case X86::CMOVE_Fp80:
332     return X86::COND_E;
333   case X86::CMOVNBE_Fp32: case X86::CMOVNBE_Fp64: case X86::CMOVNBE_Fp80:
334     return X86::COND_A;
335   case X86::CMOVNB_Fp32:  case X86::CMOVNB_Fp64:  case X86::CMOVNB_Fp80:
336     return X86::COND_AE;
337   case X86::CMOVNE_Fp32:  case X86::CMOVNE_Fp64:  case X86::CMOVNE_Fp80:
338     return X86::COND_NE;
339   case X86::CMOVNP_Fp32:  case X86::CMOVNP_Fp64:  case X86::CMOVNP_Fp80:
340     return X86::COND_NP;
341   case X86::CMOVP_Fp32:   case X86::CMOVP_Fp64:   case X86::CMOVP_Fp80:
342     return X86::COND_P;
343   }
344 }
345 
346 bool X86FlagsCopyLoweringPass::runOnMachineFunction(MachineFunction &MF) {
347   LLVM_DEBUG(dbgs() << "********** " << getPassName() << " : " << MF.getName()
348                     << " **********\n");
349 
350   Subtarget = &MF.getSubtarget<X86Subtarget>();
351   MRI = &MF.getRegInfo();
352   TII = Subtarget->getInstrInfo();
353   TRI = Subtarget->getRegisterInfo();
354   MDT = &getAnalysis<MachineDominatorTree>();
355   PromoteRC = &X86::GR8RegClass;
356 
357   if (MF.begin() == MF.end())
358     // Nothing to do for a degenerate empty function...
359     return false;
360 
361   // Collect the copies in RPO so that when there are chains where a copy is in
362   // turn copied again we visit the first one first. This ensures we can find
363   // viable locations for testing the original EFLAGS that dominate all the
364   // uses across complex CFGs.
365   SmallVector<MachineInstr *, 4> Copies;
366   ReversePostOrderTraversal<MachineFunction *> RPOT(&MF);
367   for (MachineBasicBlock *MBB : RPOT)
368     for (MachineInstr &MI : *MBB)
369       if (MI.getOpcode() == TargetOpcode::COPY &&
370           MI.getOperand(0).getReg() == X86::EFLAGS)
371         Copies.push_back(&MI);
372 
373   for (MachineInstr *CopyI : Copies) {
374     MachineBasicBlock &MBB = *CopyI->getParent();
375 
376     MachineOperand &VOp = CopyI->getOperand(1);
377     assert(VOp.isReg() &&
378            "The input to the copy for EFLAGS should always be a register!");
379     MachineInstr &CopyDefI = *MRI->getVRegDef(VOp.getReg());
380     if (CopyDefI.getOpcode() != TargetOpcode::COPY) {
381       // FIXME: The big likely candidate here are PHI nodes. We could in theory
382       // handle PHI nodes, but it gets really, really hard. Insanely hard. Hard
383       // enough that it is probably better to change every other part of LLVM
384       // to avoid creating them. The issue is that once we have PHIs we won't
385       // know which original EFLAGS value we need to capture with our setCCs
386       // below. The end result will be computing a complete set of setCCs that
387       // we *might* want, computing them in every place where we copy *out* of
388       // EFLAGS and then doing SSA formation on all of them to insert necessary
389       // PHI nodes and consume those here. Then hoping that somehow we DCE the
390       // unnecessary ones. This DCE seems very unlikely to be successful and so
391       // we will almost certainly end up with a glut of dead setCC
392       // instructions. Until we have a motivating test case and fail to avoid
393       // it by changing other parts of LLVM's lowering, we refuse to handle
394       // this complex case here.
395       LLVM_DEBUG(
396           dbgs() << "ERROR: Encountered unexpected def of an eflags copy: ";
397           CopyDefI.dump());
398       report_fatal_error(
399           "Cannot lower EFLAGS copy unless it is defined in turn by a copy!");
400     }
401 
402     auto Cleanup = make_scope_exit([&] {
403       // All uses of the EFLAGS copy are now rewritten, kill the copy into
404       // eflags and if dead the copy from.
405       CopyI->eraseFromParent();
406       if (MRI->use_empty(CopyDefI.getOperand(0).getReg()))
407         CopyDefI.eraseFromParent();
408       ++NumCopiesEliminated;
409     });
410 
411     MachineOperand &DOp = CopyI->getOperand(0);
412     assert(DOp.isDef() && "Expected register def!");
413     assert(DOp.getReg() == X86::EFLAGS && "Unexpected copy def register!");
414     if (DOp.isDead())
415       continue;
416 
417     MachineBasicBlock *TestMBB = CopyDefI.getParent();
418     auto TestPos = CopyDefI.getIterator();
419     DebugLoc TestLoc = CopyDefI.getDebugLoc();
420 
421     LLVM_DEBUG(dbgs() << "Rewriting copy: "; CopyI->dump());
422 
423     // Walk up across live-in EFLAGS to find where they were actually def'ed.
424     //
425     // This copy's def may just be part of a region of blocks covered by
426     // a single def of EFLAGS and we want to find the top of that region where
427     // possible.
428     //
429     // This is essentially a search for a *candidate* reaching definition
430     // location. We don't need to ever find the actual reaching definition here,
431     // but we want to walk up the dominator tree to find the highest point which
432     // would be viable for such a definition.
433     auto HasEFLAGSClobber = [&](MachineBasicBlock::iterator Begin,
434                                 MachineBasicBlock::iterator End) {
435       // Scan backwards as we expect these to be relatively short and often find
436       // a clobber near the end.
437       return llvm::any_of(
438           llvm::reverse(llvm::make_range(Begin, End)), [&](MachineInstr &MI) {
439             // Flag any instruction (other than the copy we are
440             // currently rewriting) that defs EFLAGS.
441             return &MI != CopyI && MI.findRegisterDefOperand(X86::EFLAGS);
442           });
443     };
444     auto HasEFLAGSClobberPath = [&](MachineBasicBlock *BeginMBB,
445                                     MachineBasicBlock *EndMBB) {
446       assert(MDT->dominates(BeginMBB, EndMBB) &&
447              "Only support paths down the dominator tree!");
448       SmallPtrSet<MachineBasicBlock *, 4> Visited;
449       SmallVector<MachineBasicBlock *, 4> Worklist;
450       // We terminate at the beginning. No need to scan it.
451       Visited.insert(BeginMBB);
452       Worklist.push_back(EndMBB);
453       do {
454         auto *MBB = Worklist.pop_back_val();
455         for (auto *PredMBB : MBB->predecessors()) {
456           if (!Visited.insert(PredMBB).second)
457             continue;
458           if (HasEFLAGSClobber(PredMBB->begin(), PredMBB->end()))
459             return true;
460           // Enqueue this block to walk its predecessors.
461           Worklist.push_back(PredMBB);
462         }
463       } while (!Worklist.empty());
464       // No clobber found along a path from the begin to end.
465       return false;
466     };
467     while (TestMBB->isLiveIn(X86::EFLAGS) && !TestMBB->pred_empty() &&
468            !HasEFLAGSClobber(TestMBB->begin(), TestPos)) {
469       // Find the nearest common dominator of the predecessors, as
470       // that will be the best candidate to hoist into.
471       MachineBasicBlock *HoistMBB =
472           std::accumulate(std::next(TestMBB->pred_begin()), TestMBB->pred_end(),
473                           *TestMBB->pred_begin(),
474                           [&](MachineBasicBlock *LHS, MachineBasicBlock *RHS) {
475                             return MDT->findNearestCommonDominator(LHS, RHS);
476                           });
477 
478       // Now we need to scan all predecessors that may be reached along paths to
479       // the hoist block. A clobber anywhere in any of these blocks the hoist.
480       // Note that this even handles loops because we require *no* clobbers.
481       if (HasEFLAGSClobberPath(HoistMBB, TestMBB))
482         break;
483 
484       // We also need the terminators to not sneakily clobber flags.
485       if (HasEFLAGSClobber(HoistMBB->getFirstTerminator()->getIterator(),
486                            HoistMBB->instr_end()))
487         break;
488 
489       // We found a viable location, hoist our test position to it.
490       TestMBB = HoistMBB;
491       TestPos = TestMBB->getFirstTerminator()->getIterator();
492       // Clear the debug location as it would just be confusing after hoisting.
493       TestLoc = DebugLoc();
494     }
495     LLVM_DEBUG({
496       auto DefIt = llvm::find_if(
497           llvm::reverse(llvm::make_range(TestMBB->instr_begin(), TestPos)),
498           [&](MachineInstr &MI) {
499             return MI.findRegisterDefOperand(X86::EFLAGS);
500           });
501       if (DefIt.base() != TestMBB->instr_begin()) {
502         dbgs() << "  Using EFLAGS defined by: ";
503         DefIt->dump();
504       } else {
505         dbgs() << "  Using live-in flags for BB:\n";
506         TestMBB->dump();
507       }
508     });
509 
510     // While rewriting uses, we buffer jumps and rewrite them in a second pass
511     // because doing so will perturb the CFG that we are walking to find the
512     // uses in the first place.
513     SmallVector<MachineInstr *, 4> JmpIs;
514 
515     // Gather the condition flags that have already been preserved in
516     // registers. We do this from scratch each time as we expect there to be
517     // very few of them and we expect to not revisit the same copy definition
518     // many times. If either of those change sufficiently we could build a map
519     // of these up front instead.
520     CondRegArray CondRegs = collectCondsInRegs(*TestMBB, TestPos);
521 
522     // Collect the basic blocks we need to scan. Typically this will just be
523     // a single basic block but we may have to scan multiple blocks if the
524     // EFLAGS copy lives into successors.
525     SmallVector<MachineBasicBlock *, 2> Blocks;
526     SmallPtrSet<MachineBasicBlock *, 2> VisitedBlocks;
527     Blocks.push_back(&MBB);
528 
529     do {
530       MachineBasicBlock &UseMBB = *Blocks.pop_back_val();
531 
532       // Track when if/when we find a kill of the flags in this block.
533       bool FlagsKilled = false;
534 
535       // In most cases, we walk from the beginning to the end of the block. But
536       // when the block is the same block as the copy is from, we will visit it
537       // twice. The first time we start from the copy and go to the end. The
538       // second time we start from the beginning and go to the copy. This lets
539       // us handle copies inside of cycles.
540       // FIXME: This loop is *super* confusing. This is at least in part
541       // a symptom of all of this routine needing to be refactored into
542       // documentable components. Once done, there may be a better way to write
543       // this loop.
544       for (auto MII = (&UseMBB == &MBB && !VisitedBlocks.count(&UseMBB))
545                           ? std::next(CopyI->getIterator())
546                           : UseMBB.instr_begin(),
547                 MIE = UseMBB.instr_end();
548            MII != MIE;) {
549         MachineInstr &MI = *MII++;
550         // If we are in the original copy block and encounter either the copy
551         // def or the copy itself, break so that we don't re-process any part of
552         // the block or process the instructions in the range that was copied
553         // over.
554         if (&MI == CopyI || &MI == &CopyDefI) {
555           assert(&UseMBB == &MBB && VisitedBlocks.count(&MBB) &&
556                  "Should only encounter these on the second pass over the "
557                  "original block.");
558           break;
559         }
560 
561         MachineOperand *FlagUse = MI.findRegisterUseOperand(X86::EFLAGS);
562         if (!FlagUse) {
563           if (MI.findRegisterDefOperand(X86::EFLAGS)) {
564             // If EFLAGS are defined, it's as-if they were killed. We can stop
565             // scanning here.
566             //
567             // NB!!! Many instructions only modify some flags. LLVM currently
568             // models this as clobbering all flags, but if that ever changes
569             // this will need to be carefully updated to handle that more
570             // complex logic.
571             FlagsKilled = true;
572             break;
573           }
574           continue;
575         }
576 
577         LLVM_DEBUG(dbgs() << "  Rewriting use: "; MI.dump());
578 
579         // Check the kill flag before we rewrite as that may change it.
580         if (FlagUse->isKill())
581           FlagsKilled = true;
582 
583         // Once we encounter a branch, the rest of the instructions must also be
584         // branches. We can't rewrite in place here, so we handle them below.
585         //
586         // Note that we don't have to handle tail calls here, even conditional
587         // tail calls, as those are not introduced into the X86 MI until post-RA
588         // branch folding or black placement. As a consequence, we get to deal
589         // with the simpler formulation of conditional branches followed by tail
590         // calls.
591         if (X86::getCondFromBranch(MI) != X86::COND_INVALID) {
592           auto JmpIt = MI.getIterator();
593           do {
594             JmpIs.push_back(&*JmpIt);
595             ++JmpIt;
596           } while (JmpIt != UseMBB.instr_end() &&
597                    X86::getCondFromBranch(*JmpIt) !=
598                        X86::COND_INVALID);
599           break;
600         }
601 
602         // Otherwise we can just rewrite in-place.
603         if (X86::getCondFromCMov(MI) != X86::COND_INVALID) {
604           rewriteCMov(*TestMBB, TestPos, TestLoc, MI, *FlagUse, CondRegs);
605         } else if (getCondFromFCMOV(MI.getOpcode()) != X86::COND_INVALID) {
606           rewriteFCMov(*TestMBB, TestPos, TestLoc, MI, *FlagUse, CondRegs);
607         } else if (X86::getCondFromSETCC(MI) != X86::COND_INVALID) {
608           rewriteSetCC(*TestMBB, TestPos, TestLoc, MI, *FlagUse, CondRegs);
609         } else if (MI.getOpcode() == TargetOpcode::COPY) {
610           rewriteCopy(MI, *FlagUse, CopyDefI);
611         } else {
612           // We assume all other instructions that use flags also def them.
613           assert(MI.findRegisterDefOperand(X86::EFLAGS) &&
614                  "Expected a def of EFLAGS for this instruction!");
615 
616           // NB!!! Several arithmetic instructions only *partially* update
617           // flags. Theoretically, we could generate MI code sequences that
618           // would rely on this fact and observe different flags independently.
619           // But currently LLVM models all of these instructions as clobbering
620           // all the flags in an undef way. We rely on that to simplify the
621           // logic.
622           FlagsKilled = true;
623 
624           // Generically handle remaining uses as arithmetic instructions.
625           rewriteArithmetic(*TestMBB, TestPos, TestLoc, MI, *FlagUse,
626                             CondRegs);
627         }
628 
629         // If this was the last use of the flags, we're done.
630         if (FlagsKilled)
631           break;
632       }
633 
634       // If the flags were killed, we're done with this block.
635       if (FlagsKilled)
636         continue;
637 
638       // Otherwise we need to scan successors for ones where the flags live-in
639       // and queue those up for processing.
640       for (MachineBasicBlock *SuccMBB : UseMBB.successors())
641         if (SuccMBB->isLiveIn(X86::EFLAGS) &&
642             VisitedBlocks.insert(SuccMBB).second) {
643           // We currently don't do any PHI insertion and so we require that the
644           // test basic block dominates all of the use basic blocks. Further, we
645           // can't have a cycle from the test block back to itself as that would
646           // create a cycle requiring a PHI to break it.
647           //
648           // We could in theory do PHI insertion here if it becomes useful by
649           // just taking undef values in along every edge that we don't trace
650           // this EFLAGS copy along. This isn't as bad as fully general PHI
651           // insertion, but still seems like a great deal of complexity.
652           //
653           // Because it is theoretically possible that some earlier MI pass or
654           // other lowering transformation could induce this to happen, we do
655           // a hard check even in non-debug builds here.
656           if (SuccMBB == TestMBB || !MDT->dominates(TestMBB, SuccMBB)) {
657             LLVM_DEBUG({
658               dbgs()
659                   << "ERROR: Encountered use that is not dominated by our test "
660                      "basic block! Rewriting this would require inserting PHI "
661                      "nodes to track the flag state across the CFG.\n\nTest "
662                      "block:\n";
663               TestMBB->dump();
664               dbgs() << "Use block:\n";
665               SuccMBB->dump();
666             });
667             report_fatal_error(
668                 "Cannot lower EFLAGS copy when original copy def "
669                 "does not dominate all uses.");
670           }
671 
672           Blocks.push_back(SuccMBB);
673 
674           // After this, EFLAGS will be recreated before each use.
675           SuccMBB->removeLiveIn(X86::EFLAGS);
676         }
677     } while (!Blocks.empty());
678 
679     // Now rewrite the jumps that use the flags. These we handle specially
680     // because if there are multiple jumps in a single basic block we'll have
681     // to do surgery on the CFG.
682     MachineBasicBlock *LastJmpMBB = nullptr;
683     for (MachineInstr *JmpI : JmpIs) {
684       // Past the first jump within a basic block we need to split the blocks
685       // apart.
686       if (JmpI->getParent() == LastJmpMBB)
687         splitBlock(*JmpI->getParent(), *JmpI, *TII);
688       else
689         LastJmpMBB = JmpI->getParent();
690 
691       rewriteCondJmp(*TestMBB, TestPos, TestLoc, *JmpI, CondRegs);
692     }
693 
694     // FIXME: Mark the last use of EFLAGS before the copy's def as a kill if
695     // the copy's def operand is itself a kill.
696   }
697 
698 #ifndef NDEBUG
699   for (MachineBasicBlock &MBB : MF)
700     for (MachineInstr &MI : MBB)
701       if (MI.getOpcode() == TargetOpcode::COPY &&
702           (MI.getOperand(0).getReg() == X86::EFLAGS ||
703            MI.getOperand(1).getReg() == X86::EFLAGS)) {
704         LLVM_DEBUG(dbgs() << "ERROR: Found a COPY involving EFLAGS: ";
705                    MI.dump());
706         llvm_unreachable("Unlowered EFLAGS copy!");
707       }
708 #endif
709 
710   return true;
711 }
712 
713 /// Collect any conditions that have already been set in registers so that we
714 /// can re-use them rather than adding duplicates.
715 CondRegArray X86FlagsCopyLoweringPass::collectCondsInRegs(
716     MachineBasicBlock &MBB, MachineBasicBlock::iterator TestPos) {
717   CondRegArray CondRegs = {};
718 
719   // Scan backwards across the range of instructions with live EFLAGS.
720   for (MachineInstr &MI :
721        llvm::reverse(llvm::make_range(MBB.begin(), TestPos))) {
722     X86::CondCode Cond = X86::getCondFromSETCC(MI);
723     if (Cond != X86::COND_INVALID && !MI.mayStore() &&
724         MI.getOperand(0).isReg() && MI.getOperand(0).getReg().isVirtual()) {
725       assert(MI.getOperand(0).isDef() &&
726              "A non-storing SETcc should always define a register!");
727       CondRegs[Cond] = MI.getOperand(0).getReg();
728     }
729 
730     // Stop scanning when we see the first definition of the EFLAGS as prior to
731     // this we would potentially capture the wrong flag state.
732     if (MI.findRegisterDefOperand(X86::EFLAGS))
733       break;
734   }
735   return CondRegs;
736 }
737 
738 Register X86FlagsCopyLoweringPass::promoteCondToReg(
739     MachineBasicBlock &TestMBB, MachineBasicBlock::iterator TestPos,
740     const DebugLoc &TestLoc, X86::CondCode Cond) {
741   Register Reg = MRI->createVirtualRegister(PromoteRC);
742   auto SetI = BuildMI(TestMBB, TestPos, TestLoc,
743                       TII->get(X86::SETCCr), Reg).addImm(Cond);
744   (void)SetI;
745   LLVM_DEBUG(dbgs() << "    save cond: "; SetI->dump());
746   ++NumSetCCsInserted;
747   return Reg;
748 }
749 
750 std::pair<unsigned, bool> X86FlagsCopyLoweringPass::getCondOrInverseInReg(
751     MachineBasicBlock &TestMBB, MachineBasicBlock::iterator TestPos,
752     const DebugLoc &TestLoc, X86::CondCode Cond, CondRegArray &CondRegs) {
753   unsigned &CondReg = CondRegs[Cond];
754   unsigned &InvCondReg = CondRegs[X86::GetOppositeBranchCondition(Cond)];
755   if (!CondReg && !InvCondReg)
756     CondReg = promoteCondToReg(TestMBB, TestPos, TestLoc, Cond);
757 
758   if (CondReg)
759     return {CondReg, false};
760   else
761     return {InvCondReg, true};
762 }
763 
764 void X86FlagsCopyLoweringPass::insertTest(MachineBasicBlock &MBB,
765                                           MachineBasicBlock::iterator Pos,
766                                           const DebugLoc &Loc, unsigned Reg) {
767   auto TestI =
768       BuildMI(MBB, Pos, Loc, TII->get(X86::TEST8rr)).addReg(Reg).addReg(Reg);
769   (void)TestI;
770   LLVM_DEBUG(dbgs() << "    test cond: "; TestI->dump());
771   ++NumTestsInserted;
772 }
773 
774 void X86FlagsCopyLoweringPass::rewriteArithmetic(
775     MachineBasicBlock &TestMBB, MachineBasicBlock::iterator TestPos,
776     const DebugLoc &TestLoc, MachineInstr &MI, MachineOperand &FlagUse,
777     CondRegArray &CondRegs) {
778   // Arithmetic is either reading CF or OF. Figure out which condition we need
779   // to preserve in a register.
780   X86::CondCode Cond = X86::COND_INVALID;
781 
782   // The addend to use to reset CF or OF when added to the flag value.
783   int Addend = 0;
784 
785   switch (getMnemonicFromOpcode(MI.getOpcode())) {
786   case FlagArithMnemonic::ADC:
787   case FlagArithMnemonic::RCL:
788   case FlagArithMnemonic::RCR:
789   case FlagArithMnemonic::SBB:
790   case FlagArithMnemonic::SETB:
791     Cond = X86::COND_B; // CF == 1
792     // Set up an addend that when one is added will need a carry due to not
793     // having a higher bit available.
794     Addend = 255;
795     break;
796   }
797 
798   // Now get a register that contains the value of the flag input to the
799   // arithmetic. We require exactly this flag to simplify the arithmetic
800   // required to materialize it back into the flag.
801   unsigned &CondReg = CondRegs[Cond];
802   if (!CondReg)
803     CondReg = promoteCondToReg(TestMBB, TestPos, TestLoc, Cond);
804 
805   MachineBasicBlock &MBB = *MI.getParent();
806 
807   // Insert an instruction that will set the flag back to the desired value.
808   Register TmpReg = MRI->createVirtualRegister(PromoteRC);
809   auto AddI =
810       BuildMI(MBB, MI.getIterator(), MI.getDebugLoc(), TII->get(X86::ADD8ri))
811           .addDef(TmpReg, RegState::Dead)
812           .addReg(CondReg)
813           .addImm(Addend);
814   (void)AddI;
815   LLVM_DEBUG(dbgs() << "    add cond: "; AddI->dump());
816   ++NumAddsInserted;
817   FlagUse.setIsKill(true);
818 }
819 
820 void X86FlagsCopyLoweringPass::rewriteCMov(MachineBasicBlock &TestMBB,
821                                            MachineBasicBlock::iterator TestPos,
822                                            const DebugLoc &TestLoc,
823                                            MachineInstr &CMovI,
824                                            MachineOperand &FlagUse,
825                                            CondRegArray &CondRegs) {
826   // First get the register containing this specific condition.
827   X86::CondCode Cond = X86::getCondFromCMov(CMovI);
828   unsigned CondReg;
829   bool Inverted;
830   std::tie(CondReg, Inverted) =
831       getCondOrInverseInReg(TestMBB, TestPos, TestLoc, Cond, CondRegs);
832 
833   MachineBasicBlock &MBB = *CMovI.getParent();
834 
835   // Insert a direct test of the saved register.
836   insertTest(MBB, CMovI.getIterator(), CMovI.getDebugLoc(), CondReg);
837 
838   // Rewrite the CMov to use the !ZF flag from the test, and then kill its use
839   // of the flags afterward.
840   CMovI.getOperand(CMovI.getDesc().getNumOperands() - 1)
841       .setImm(Inverted ? X86::COND_E : X86::COND_NE);
842   FlagUse.setIsKill(true);
843   LLVM_DEBUG(dbgs() << "    fixed cmov: "; CMovI.dump());
844 }
845 
846 void X86FlagsCopyLoweringPass::rewriteFCMov(MachineBasicBlock &TestMBB,
847                                             MachineBasicBlock::iterator TestPos,
848                                             const DebugLoc &TestLoc,
849                                             MachineInstr &CMovI,
850                                             MachineOperand &FlagUse,
851                                             CondRegArray &CondRegs) {
852   // First get the register containing this specific condition.
853   X86::CondCode Cond = getCondFromFCMOV(CMovI.getOpcode());
854   unsigned CondReg;
855   bool Inverted;
856   std::tie(CondReg, Inverted) =
857       getCondOrInverseInReg(TestMBB, TestPos, TestLoc, Cond, CondRegs);
858 
859   MachineBasicBlock &MBB = *CMovI.getParent();
860 
861   // Insert a direct test of the saved register.
862   insertTest(MBB, CMovI.getIterator(), CMovI.getDebugLoc(), CondReg);
863 
864   auto getFCMOVOpcode = [](unsigned Opcode, bool Inverted) {
865     switch (Opcode) {
866     default: llvm_unreachable("Unexpected opcode!");
867     case X86::CMOVBE_Fp32: case X86::CMOVNBE_Fp32:
868     case X86::CMOVB_Fp32:  case X86::CMOVNB_Fp32:
869     case X86::CMOVE_Fp32:  case X86::CMOVNE_Fp32:
870     case X86::CMOVP_Fp32:  case X86::CMOVNP_Fp32:
871       return Inverted ? X86::CMOVE_Fp32 : X86::CMOVNE_Fp32;
872     case X86::CMOVBE_Fp64: case X86::CMOVNBE_Fp64:
873     case X86::CMOVB_Fp64:  case X86::CMOVNB_Fp64:
874     case X86::CMOVE_Fp64:  case X86::CMOVNE_Fp64:
875     case X86::CMOVP_Fp64:  case X86::CMOVNP_Fp64:
876       return Inverted ? X86::CMOVE_Fp64 : X86::CMOVNE_Fp64;
877     case X86::CMOVBE_Fp80: case X86::CMOVNBE_Fp80:
878     case X86::CMOVB_Fp80:  case X86::CMOVNB_Fp80:
879     case X86::CMOVE_Fp80:  case X86::CMOVNE_Fp80:
880     case X86::CMOVP_Fp80:  case X86::CMOVNP_Fp80:
881       return Inverted ? X86::CMOVE_Fp80 : X86::CMOVNE_Fp80;
882     }
883   };
884 
885   // Rewrite the CMov to use the !ZF flag from the test.
886   CMovI.setDesc(TII->get(getFCMOVOpcode(CMovI.getOpcode(), Inverted)));
887   FlagUse.setIsKill(true);
888   LLVM_DEBUG(dbgs() << "    fixed fcmov: "; CMovI.dump());
889 }
890 
891 void X86FlagsCopyLoweringPass::rewriteCondJmp(
892     MachineBasicBlock &TestMBB, MachineBasicBlock::iterator TestPos,
893     const DebugLoc &TestLoc, MachineInstr &JmpI, CondRegArray &CondRegs) {
894   // First get the register containing this specific condition.
895   X86::CondCode Cond = X86::getCondFromBranch(JmpI);
896   unsigned CondReg;
897   bool Inverted;
898   std::tie(CondReg, Inverted) =
899       getCondOrInverseInReg(TestMBB, TestPos, TestLoc, Cond, CondRegs);
900 
901   MachineBasicBlock &JmpMBB = *JmpI.getParent();
902 
903   // Insert a direct test of the saved register.
904   insertTest(JmpMBB, JmpI.getIterator(), JmpI.getDebugLoc(), CondReg);
905 
906   // Rewrite the jump to use the !ZF flag from the test, and kill its use of
907   // flags afterward.
908   JmpI.getOperand(1).setImm(Inverted ? X86::COND_E : X86::COND_NE);
909   JmpI.findRegisterUseOperand(X86::EFLAGS)->setIsKill(true);
910   LLVM_DEBUG(dbgs() << "    fixed jCC: "; JmpI.dump());
911 }
912 
913 void X86FlagsCopyLoweringPass::rewriteCopy(MachineInstr &MI,
914                                            MachineOperand &FlagUse,
915                                            MachineInstr &CopyDefI) {
916   // Just replace this copy with the original copy def.
917   MRI->replaceRegWith(MI.getOperand(0).getReg(),
918                       CopyDefI.getOperand(0).getReg());
919   MI.eraseFromParent();
920 }
921 
922 void X86FlagsCopyLoweringPass::rewriteSetCC(MachineBasicBlock &TestMBB,
923                                             MachineBasicBlock::iterator TestPos,
924                                             const DebugLoc &TestLoc,
925                                             MachineInstr &SetCCI,
926                                             MachineOperand &FlagUse,
927                                             CondRegArray &CondRegs) {
928   X86::CondCode Cond = X86::getCondFromSETCC(SetCCI);
929   // Note that we can't usefully rewrite this to the inverse without complex
930   // analysis of the users of the setCC. Largely we rely on duplicates which
931   // could have been avoided already being avoided here.
932   unsigned &CondReg = CondRegs[Cond];
933   if (!CondReg)
934     CondReg = promoteCondToReg(TestMBB, TestPos, TestLoc, Cond);
935 
936   // Rewriting a register def is trivial: we just replace the register and
937   // remove the setcc.
938   if (!SetCCI.mayStore()) {
939     assert(SetCCI.getOperand(0).isReg() &&
940            "Cannot have a non-register defined operand to SETcc!");
941     Register OldReg = SetCCI.getOperand(0).getReg();
942     // Drop Kill flags on the old register before replacing. CondReg may have
943     // a longer live range.
944     MRI->clearKillFlags(OldReg);
945     MRI->replaceRegWith(OldReg, CondReg);
946     SetCCI.eraseFromParent();
947     return;
948   }
949 
950   // Otherwise, we need to emit a store.
951   auto MIB = BuildMI(*SetCCI.getParent(), SetCCI.getIterator(),
952                      SetCCI.getDebugLoc(), TII->get(X86::MOV8mr));
953   // Copy the address operands.
954   for (int i = 0; i < X86::AddrNumOperands; ++i)
955     MIB.add(SetCCI.getOperand(i));
956 
957   MIB.addReg(CondReg);
958 
959   MIB.setMemRefs(SetCCI.memoperands());
960 
961   SetCCI.eraseFromParent();
962 }
963