xref: /freebsd/contrib/llvm-project/llvm/lib/Target/X86/X86FloatingPoint.cpp (revision 5036d9652a5701d00e9e40ea942c278e9f77d33d)
1 //===-- X86FloatingPoint.cpp - Floating point Reg -> Stack converter ------===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 //
9 // This file defines the pass which converts floating point instructions from
10 // pseudo registers into register stack instructions.  This pass uses live
11 // variable information to indicate where the FPn registers are used and their
12 // lifetimes.
13 //
14 // The x87 hardware tracks liveness of the stack registers, so it is necessary
15 // to implement exact liveness tracking between basic blocks. The CFG edges are
16 // partitioned into bundles where the same FP registers must be live in
17 // identical stack positions. Instructions are inserted at the end of each basic
18 // block to rearrange the live registers to match the outgoing bundle.
19 //
20 // This approach avoids splitting critical edges at the potential cost of more
21 // live register shuffling instructions when critical edges are present.
22 //
23 //===----------------------------------------------------------------------===//
24 
25 #include "X86.h"
26 #include "X86InstrInfo.h"
27 #include "llvm/ADT/DepthFirstIterator.h"
28 #include "llvm/ADT/STLExtras.h"
29 #include "llvm/ADT/SmallSet.h"
30 #include "llvm/ADT/SmallVector.h"
31 #include "llvm/ADT/Statistic.h"
32 #include "llvm/CodeGen/EdgeBundles.h"
33 #include "llvm/CodeGen/LiveRegUnits.h"
34 #include "llvm/CodeGen/MachineFunctionPass.h"
35 #include "llvm/CodeGen/MachineInstrBuilder.h"
36 #include "llvm/CodeGen/MachineRegisterInfo.h"
37 #include "llvm/CodeGen/Passes.h"
38 #include "llvm/CodeGen/TargetInstrInfo.h"
39 #include "llvm/CodeGen/TargetSubtargetInfo.h"
40 #include "llvm/Config/llvm-config.h"
41 #include "llvm/IR/InlineAsm.h"
42 #include "llvm/InitializePasses.h"
43 #include "llvm/Support/Debug.h"
44 #include "llvm/Support/ErrorHandling.h"
45 #include "llvm/Support/raw_ostream.h"
46 #include "llvm/Target/TargetMachine.h"
47 #include <algorithm>
48 #include <bitset>
49 using namespace llvm;
50 
51 #define DEBUG_TYPE "x86-codegen"
52 
53 STATISTIC(NumFXCH, "Number of fxch instructions inserted");
54 STATISTIC(NumFP  , "Number of floating point instructions");
55 
56 namespace {
57   const unsigned ScratchFPReg = 7;
58 
59   struct FPS : public MachineFunctionPass {
60     static char ID;
61     FPS() : MachineFunctionPass(ID) {
62       // This is really only to keep valgrind quiet.
63       // The logic in isLive() is too much for it.
64       memset(Stack, 0, sizeof(Stack));
65       memset(RegMap, 0, sizeof(RegMap));
66     }
67 
68     void getAnalysisUsage(AnalysisUsage &AU) const override {
69       AU.setPreservesCFG();
70       AU.addRequired<EdgeBundles>();
71       AU.addPreservedID(MachineLoopInfoID);
72       AU.addPreservedID(MachineDominatorsID);
73       MachineFunctionPass::getAnalysisUsage(AU);
74     }
75 
76     bool runOnMachineFunction(MachineFunction &MF) override;
77 
78     MachineFunctionProperties getRequiredProperties() const override {
79       return MachineFunctionProperties().set(
80           MachineFunctionProperties::Property::NoVRegs);
81     }
82 
83     StringRef getPassName() const override { return "X86 FP Stackifier"; }
84 
85   private:
86     const TargetInstrInfo *TII = nullptr; // Machine instruction info.
87 
88     // Two CFG edges are related if they leave the same block, or enter the same
89     // block. The transitive closure of an edge under this relation is a
90     // LiveBundle. It represents a set of CFG edges where the live FP stack
91     // registers must be allocated identically in the x87 stack.
92     //
93     // A LiveBundle is usually all the edges leaving a block, or all the edges
94     // entering a block, but it can contain more edges if critical edges are
95     // present.
96     //
97     // The set of live FP registers in a LiveBundle is calculated by bundleCFG,
98     // but the exact mapping of FP registers to stack slots is fixed later.
99     struct LiveBundle {
100       // Bit mask of live FP registers. Bit 0 = FP0, bit 1 = FP1, &c.
101       unsigned Mask = 0;
102 
103       // Number of pre-assigned live registers in FixStack. This is 0 when the
104       // stack order has not yet been fixed.
105       unsigned FixCount = 0;
106 
107       // Assigned stack order for live-in registers.
108       // FixStack[i] == getStackEntry(i) for all i < FixCount.
109       unsigned char FixStack[8];
110 
111       LiveBundle() = default;
112 
113       // Have the live registers been assigned a stack order yet?
114       bool isFixed() const { return !Mask || FixCount; }
115     };
116 
117     // Numbered LiveBundle structs. LiveBundles[0] is used for all CFG edges
118     // with no live FP registers.
119     SmallVector<LiveBundle, 8> LiveBundles;
120 
121     // The edge bundle analysis provides indices into the LiveBundles vector.
122     EdgeBundles *Bundles = nullptr;
123 
124     // Return a bitmask of FP registers in block's live-in list.
125     static unsigned calcLiveInMask(MachineBasicBlock *MBB, bool RemoveFPs) {
126       unsigned Mask = 0;
127       for (MachineBasicBlock::livein_iterator I = MBB->livein_begin();
128            I != MBB->livein_end(); ) {
129         MCPhysReg Reg = I->PhysReg;
130         static_assert(X86::FP6 - X86::FP0 == 6, "sequential regnums");
131         if (Reg >= X86::FP0 && Reg <= X86::FP6) {
132           Mask |= 1 << (Reg - X86::FP0);
133           if (RemoveFPs) {
134             I = MBB->removeLiveIn(I);
135             continue;
136           }
137         }
138         ++I;
139       }
140       return Mask;
141     }
142 
143     // Partition all the CFG edges into LiveBundles.
144     void bundleCFGRecomputeKillFlags(MachineFunction &MF);
145 
146     MachineBasicBlock *MBB = nullptr;     // Current basic block
147 
148     // The hardware keeps track of how many FP registers are live, so we have
149     // to model that exactly. Usually, each live register corresponds to an
150     // FP<n> register, but when dealing with calls, returns, and inline
151     // assembly, it is sometimes necessary to have live scratch registers.
152     unsigned Stack[8];          // FP<n> Registers in each stack slot...
153     unsigned StackTop = 0;      // The current top of the FP stack.
154 
155     enum {
156       NumFPRegs = 8             // Including scratch pseudo-registers.
157     };
158 
159     // For each live FP<n> register, point to its Stack[] entry.
160     // The first entries correspond to FP0-FP6, the rest are scratch registers
161     // used when we need slightly different live registers than what the
162     // register allocator thinks.
163     unsigned RegMap[NumFPRegs];
164 
165     // Set up our stack model to match the incoming registers to MBB.
166     void setupBlockStack();
167 
168     // Shuffle live registers to match the expectations of successor blocks.
169     void finishBlockStack();
170 
171 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
172     void dumpStack() const {
173       dbgs() << "Stack contents:";
174       for (unsigned i = 0; i != StackTop; ++i) {
175         dbgs() << " FP" << Stack[i];
176         assert(RegMap[Stack[i]] == i && "Stack[] doesn't match RegMap[]!");
177       }
178     }
179 #endif
180 
181     /// getSlot - Return the stack slot number a particular register number is
182     /// in.
183     unsigned getSlot(unsigned RegNo) const {
184       assert(RegNo < NumFPRegs && "Regno out of range!");
185       return RegMap[RegNo];
186     }
187 
188     /// isLive - Is RegNo currently live in the stack?
189     bool isLive(unsigned RegNo) const {
190       unsigned Slot = getSlot(RegNo);
191       return Slot < StackTop && Stack[Slot] == RegNo;
192     }
193 
194     /// getStackEntry - Return the X86::FP<n> register in register ST(i).
195     unsigned getStackEntry(unsigned STi) const {
196       if (STi >= StackTop)
197         report_fatal_error("Access past stack top!");
198       return Stack[StackTop-1-STi];
199     }
200 
201     /// getSTReg - Return the X86::ST(i) register which contains the specified
202     /// FP<RegNo> register.
203     unsigned getSTReg(unsigned RegNo) const {
204       return StackTop - 1 - getSlot(RegNo) + X86::ST0;
205     }
206 
207     // pushReg - Push the specified FP<n> register onto the stack.
208     void pushReg(unsigned Reg) {
209       assert(Reg < NumFPRegs && "Register number out of range!");
210       if (StackTop >= 8)
211         report_fatal_error("Stack overflow!");
212       Stack[StackTop] = Reg;
213       RegMap[Reg] = StackTop++;
214     }
215 
216     // popReg - Pop a register from the stack.
217     void popReg() {
218       if (StackTop == 0)
219         report_fatal_error("Cannot pop empty stack!");
220       RegMap[Stack[--StackTop]] = ~0;     // Update state
221     }
222 
223     bool isAtTop(unsigned RegNo) const { return getSlot(RegNo) == StackTop-1; }
224     void moveToTop(unsigned RegNo, MachineBasicBlock::iterator I) {
225       DebugLoc dl = I == MBB->end() ? DebugLoc() : I->getDebugLoc();
226       if (isAtTop(RegNo)) return;
227 
228       unsigned STReg = getSTReg(RegNo);
229       unsigned RegOnTop = getStackEntry(0);
230 
231       // Swap the slots the regs are in.
232       std::swap(RegMap[RegNo], RegMap[RegOnTop]);
233 
234       // Swap stack slot contents.
235       if (RegMap[RegOnTop] >= StackTop)
236         report_fatal_error("Access past stack top!");
237       std::swap(Stack[RegMap[RegOnTop]], Stack[StackTop-1]);
238 
239       // Emit an fxch to update the runtime processors version of the state.
240       BuildMI(*MBB, I, dl, TII->get(X86::XCH_F)).addReg(STReg);
241       ++NumFXCH;
242     }
243 
244     void duplicateToTop(unsigned RegNo, unsigned AsReg,
245                         MachineBasicBlock::iterator I) {
246       DebugLoc dl = I == MBB->end() ? DebugLoc() : I->getDebugLoc();
247       unsigned STReg = getSTReg(RegNo);
248       pushReg(AsReg);   // New register on top of stack
249 
250       BuildMI(*MBB, I, dl, TII->get(X86::LD_Frr)).addReg(STReg);
251     }
252 
253     /// popStackAfter - Pop the current value off of the top of the FP stack
254     /// after the specified instruction.
255     void popStackAfter(MachineBasicBlock::iterator &I);
256 
257     /// freeStackSlotAfter - Free the specified register from the register
258     /// stack, so that it is no longer in a register.  If the register is
259     /// currently at the top of the stack, we just pop the current instruction,
260     /// otherwise we store the current top-of-stack into the specified slot,
261     /// then pop the top of stack.
262     void freeStackSlotAfter(MachineBasicBlock::iterator &I, unsigned Reg);
263 
264     /// freeStackSlotBefore - Just the pop, no folding. Return the inserted
265     /// instruction.
266     MachineBasicBlock::iterator
267     freeStackSlotBefore(MachineBasicBlock::iterator I, unsigned FPRegNo);
268 
269     /// Adjust the live registers to be the set in Mask.
270     void adjustLiveRegs(unsigned Mask, MachineBasicBlock::iterator I);
271 
272     /// Shuffle the top FixCount stack entries such that FP reg FixStack[0] is
273     /// st(0), FP reg FixStack[1] is st(1) etc.
274     void shuffleStackTop(const unsigned char *FixStack, unsigned FixCount,
275                          MachineBasicBlock::iterator I);
276 
277     bool processBasicBlock(MachineFunction &MF, MachineBasicBlock &MBB);
278 
279     void handleCall(MachineBasicBlock::iterator &I);
280     void handleReturn(MachineBasicBlock::iterator &I);
281     void handleZeroArgFP(MachineBasicBlock::iterator &I);
282     void handleOneArgFP(MachineBasicBlock::iterator &I);
283     void handleOneArgFPRW(MachineBasicBlock::iterator &I);
284     void handleTwoArgFP(MachineBasicBlock::iterator &I);
285     void handleCompareFP(MachineBasicBlock::iterator &I);
286     void handleCondMovFP(MachineBasicBlock::iterator &I);
287     void handleSpecialFP(MachineBasicBlock::iterator &I);
288 
289     // Check if a COPY instruction is using FP registers.
290     static bool isFPCopy(MachineInstr &MI) {
291       Register DstReg = MI.getOperand(0).getReg();
292       Register SrcReg = MI.getOperand(1).getReg();
293 
294       return X86::RFP80RegClass.contains(DstReg) ||
295         X86::RFP80RegClass.contains(SrcReg);
296     }
297 
298     void setKillFlags(MachineBasicBlock &MBB) const;
299   };
300 }
301 
302 char FPS::ID = 0;
303 
304 INITIALIZE_PASS_BEGIN(FPS, DEBUG_TYPE, "X86 FP Stackifier",
305                       false, false)
306 INITIALIZE_PASS_DEPENDENCY(EdgeBundles)
307 INITIALIZE_PASS_END(FPS, DEBUG_TYPE, "X86 FP Stackifier",
308                     false, false)
309 
310 FunctionPass *llvm::createX86FloatingPointStackifierPass() { return new FPS(); }
311 
312 /// getFPReg - Return the X86::FPx register number for the specified operand.
313 /// For example, this returns 3 for X86::FP3.
314 static unsigned getFPReg(const MachineOperand &MO) {
315   assert(MO.isReg() && "Expected an FP register!");
316   Register Reg = MO.getReg();
317   assert(Reg >= X86::FP0 && Reg <= X86::FP6 && "Expected FP register!");
318   return Reg - X86::FP0;
319 }
320 
321 /// runOnMachineFunction - Loop over all of the basic blocks, transforming FP
322 /// register references into FP stack references.
323 ///
324 bool FPS::runOnMachineFunction(MachineFunction &MF) {
325   // We only need to run this pass if there are any FP registers used in this
326   // function.  If it is all integer, there is nothing for us to do!
327   bool FPIsUsed = false;
328 
329   static_assert(X86::FP6 == X86::FP0+6, "Register enums aren't sorted right!");
330   const MachineRegisterInfo &MRI = MF.getRegInfo();
331   for (unsigned i = 0; i <= 6; ++i)
332     if (!MRI.reg_nodbg_empty(X86::FP0 + i)) {
333       FPIsUsed = true;
334       break;
335     }
336 
337   // Early exit.
338   if (!FPIsUsed) return false;
339 
340   Bundles = &getAnalysis<EdgeBundles>();
341   TII = MF.getSubtarget().getInstrInfo();
342 
343   // Prepare cross-MBB liveness.
344   bundleCFGRecomputeKillFlags(MF);
345 
346   StackTop = 0;
347 
348   // Process the function in depth first order so that we process at least one
349   // of the predecessors for every reachable block in the function.
350   df_iterator_default_set<MachineBasicBlock*> Processed;
351   MachineBasicBlock *Entry = &MF.front();
352 
353   LiveBundle &Bundle =
354     LiveBundles[Bundles->getBundle(Entry->getNumber(), false)];
355 
356   // In regcall convention, some FP registers may not be passed through
357   // the stack, so they will need to be assigned to the stack first
358   if ((Entry->getParent()->getFunction().getCallingConv() ==
359     CallingConv::X86_RegCall) && (Bundle.Mask && !Bundle.FixCount)) {
360     // In the register calling convention, up to one FP argument could be
361     // saved in the first FP register.
362     // If bundle.mask is non-zero and Bundle.FixCount is zero, it means
363     // that the FP registers contain arguments.
364     // The actual value is passed in FP0.
365     // Here we fix the stack and mark FP0 as pre-assigned register.
366     assert((Bundle.Mask & 0xFE) == 0 &&
367       "Only FP0 could be passed as an argument");
368     Bundle.FixCount = 1;
369     Bundle.FixStack[0] = 0;
370   }
371 
372   bool Changed = false;
373   for (MachineBasicBlock *BB : depth_first_ext(Entry, Processed))
374     Changed |= processBasicBlock(MF, *BB);
375 
376   // Process any unreachable blocks in arbitrary order now.
377   if (MF.size() != Processed.size())
378     for (MachineBasicBlock &BB : MF)
379       if (Processed.insert(&BB).second)
380         Changed |= processBasicBlock(MF, BB);
381 
382   LiveBundles.clear();
383 
384   return Changed;
385 }
386 
387 /// bundleCFG - Scan all the basic blocks to determine consistent live-in and
388 /// live-out sets for the FP registers. Consistent means that the set of
389 /// registers live-out from a block is identical to the live-in set of all
390 /// successors. This is not enforced by the normal live-in lists since
391 /// registers may be implicitly defined, or not used by all successors.
392 void FPS::bundleCFGRecomputeKillFlags(MachineFunction &MF) {
393   assert(LiveBundles.empty() && "Stale data in LiveBundles");
394   LiveBundles.resize(Bundles->getNumBundles());
395 
396   // Gather the actual live-in masks for all MBBs.
397   for (MachineBasicBlock &MBB : MF) {
398     setKillFlags(MBB);
399 
400     const unsigned Mask = calcLiveInMask(&MBB, false);
401     if (!Mask)
402       continue;
403     // Update MBB ingoing bundle mask.
404     LiveBundles[Bundles->getBundle(MBB.getNumber(), false)].Mask |= Mask;
405   }
406 }
407 
408 /// processBasicBlock - Loop over all of the instructions in the basic block,
409 /// transforming FP instructions into their stack form.
410 ///
411 bool FPS::processBasicBlock(MachineFunction &MF, MachineBasicBlock &BB) {
412   bool Changed = false;
413   MBB = &BB;
414 
415   setupBlockStack();
416 
417   for (MachineBasicBlock::iterator I = BB.begin(); I != BB.end(); ++I) {
418     MachineInstr &MI = *I;
419     uint64_t Flags = MI.getDesc().TSFlags;
420 
421     unsigned FPInstClass = Flags & X86II::FPTypeMask;
422     if (MI.isInlineAsm())
423       FPInstClass = X86II::SpecialFP;
424 
425     if (MI.isCopy() && isFPCopy(MI))
426       FPInstClass = X86II::SpecialFP;
427 
428     if (MI.isImplicitDef() &&
429         X86::RFP80RegClass.contains(MI.getOperand(0).getReg()))
430       FPInstClass = X86II::SpecialFP;
431 
432     if (MI.isCall())
433       FPInstClass = X86II::SpecialFP;
434 
435     if (FPInstClass == X86II::NotFP)
436       continue;  // Efficiently ignore non-fp insts!
437 
438     MachineInstr *PrevMI = nullptr;
439     if (I != BB.begin())
440       PrevMI = &*std::prev(I);
441 
442     ++NumFP;  // Keep track of # of pseudo instrs
443     LLVM_DEBUG(dbgs() << "\nFPInst:\t" << MI);
444 
445     // Get dead variables list now because the MI pointer may be deleted as part
446     // of processing!
447     SmallVector<unsigned, 8> DeadRegs;
448     for (const MachineOperand &MO : MI.operands())
449       if (MO.isReg() && MO.isDead())
450         DeadRegs.push_back(MO.getReg());
451 
452     switch (FPInstClass) {
453     case X86II::ZeroArgFP:  handleZeroArgFP(I); break;
454     case X86II::OneArgFP:   handleOneArgFP(I);  break;  // fstp ST(0)
455     case X86II::OneArgFPRW: handleOneArgFPRW(I); break; // ST(0) = fsqrt(ST(0))
456     case X86II::TwoArgFP:   handleTwoArgFP(I);  break;
457     case X86II::CompareFP:  handleCompareFP(I); break;
458     case X86II::CondMovFP:  handleCondMovFP(I); break;
459     case X86II::SpecialFP:  handleSpecialFP(I); break;
460     default: llvm_unreachable("Unknown FP Type!");
461     }
462 
463     // Check to see if any of the values defined by this instruction are dead
464     // after definition.  If so, pop them.
465     for (unsigned Reg : DeadRegs) {
466       // Check if Reg is live on the stack. An inline-asm register operand that
467       // is in the clobber list and marked dead might not be live on the stack.
468       static_assert(X86::FP7 - X86::FP0 == 7, "sequential FP regnumbers");
469       if (Reg >= X86::FP0 && Reg <= X86::FP6 && isLive(Reg-X86::FP0)) {
470         LLVM_DEBUG(dbgs() << "Register FP#" << Reg - X86::FP0 << " is dead!\n");
471         freeStackSlotAfter(I, Reg-X86::FP0);
472       }
473     }
474 
475     // Print out all of the instructions expanded to if -debug
476     LLVM_DEBUG({
477       MachineBasicBlock::iterator PrevI = PrevMI;
478       if (I == PrevI) {
479         dbgs() << "Just deleted pseudo instruction\n";
480       } else {
481         MachineBasicBlock::iterator Start = I;
482         // Rewind to first instruction newly inserted.
483         while (Start != BB.begin() && std::prev(Start) != PrevI)
484           --Start;
485         dbgs() << "Inserted instructions:\n\t";
486         Start->print(dbgs());
487         while (++Start != std::next(I)) {
488         }
489       }
490       dumpStack();
491     });
492     (void)PrevMI;
493 
494     Changed = true;
495   }
496 
497   finishBlockStack();
498 
499   return Changed;
500 }
501 
502 /// setupBlockStack - Use the live bundles to set up our model of the stack
503 /// to match predecessors' live out stack.
504 void FPS::setupBlockStack() {
505   LLVM_DEBUG(dbgs() << "\nSetting up live-ins for " << printMBBReference(*MBB)
506                     << " derived from " << MBB->getName() << ".\n");
507   StackTop = 0;
508   // Get the live-in bundle for MBB.
509   const LiveBundle &Bundle =
510     LiveBundles[Bundles->getBundle(MBB->getNumber(), false)];
511 
512   if (!Bundle.Mask) {
513     LLVM_DEBUG(dbgs() << "Block has no FP live-ins.\n");
514     return;
515   }
516 
517   // Depth-first iteration should ensure that we always have an assigned stack.
518   assert(Bundle.isFixed() && "Reached block before any predecessors");
519 
520   // Push the fixed live-in registers.
521   for (unsigned i = Bundle.FixCount; i > 0; --i) {
522     LLVM_DEBUG(dbgs() << "Live-in st(" << (i - 1) << "): %fp"
523                       << unsigned(Bundle.FixStack[i - 1]) << '\n');
524     pushReg(Bundle.FixStack[i-1]);
525   }
526 
527   // Kill off unwanted live-ins. This can happen with a critical edge.
528   // FIXME: We could keep these live registers around as zombies. They may need
529   // to be revived at the end of a short block. It might save a few instrs.
530   unsigned Mask = calcLiveInMask(MBB, /*RemoveFPs=*/true);
531   adjustLiveRegs(Mask, MBB->begin());
532   LLVM_DEBUG(MBB->dump());
533 }
534 
535 /// finishBlockStack - Revive live-outs that are implicitly defined out of
536 /// MBB. Shuffle live registers to match the expected fixed stack of any
537 /// predecessors, and ensure that all predecessors are expecting the same
538 /// stack.
539 void FPS::finishBlockStack() {
540   // The RET handling below takes care of return blocks for us.
541   if (MBB->succ_empty())
542     return;
543 
544   LLVM_DEBUG(dbgs() << "Setting up live-outs for " << printMBBReference(*MBB)
545                     << " derived from " << MBB->getName() << ".\n");
546 
547   // Get MBB's live-out bundle.
548   unsigned BundleIdx = Bundles->getBundle(MBB->getNumber(), true);
549   LiveBundle &Bundle = LiveBundles[BundleIdx];
550 
551   // We may need to kill and define some registers to match successors.
552   // FIXME: This can probably be combined with the shuffle below.
553   MachineBasicBlock::iterator Term = MBB->getFirstTerminator();
554   adjustLiveRegs(Bundle.Mask, Term);
555 
556   if (!Bundle.Mask) {
557     LLVM_DEBUG(dbgs() << "No live-outs.\n");
558     return;
559   }
560 
561   // Has the stack order been fixed yet?
562   LLVM_DEBUG(dbgs() << "LB#" << BundleIdx << ": ");
563   if (Bundle.isFixed()) {
564     LLVM_DEBUG(dbgs() << "Shuffling stack to match.\n");
565     shuffleStackTop(Bundle.FixStack, Bundle.FixCount, Term);
566   } else {
567     // Not fixed yet, we get to choose.
568     LLVM_DEBUG(dbgs() << "Fixing stack order now.\n");
569     Bundle.FixCount = StackTop;
570     for (unsigned i = 0; i < StackTop; ++i)
571       Bundle.FixStack[i] = getStackEntry(i);
572   }
573 }
574 
575 
576 //===----------------------------------------------------------------------===//
577 // Efficient Lookup Table Support
578 //===----------------------------------------------------------------------===//
579 
580 namespace {
581   struct TableEntry {
582     uint16_t from;
583     uint16_t to;
584     bool operator<(const TableEntry &TE) const { return from < TE.from; }
585     friend bool operator<(const TableEntry &TE, unsigned V) {
586       return TE.from < V;
587     }
588     friend bool LLVM_ATTRIBUTE_UNUSED operator<(unsigned V,
589                                                 const TableEntry &TE) {
590       return V < TE.from;
591     }
592   };
593 }
594 
595 static int Lookup(ArrayRef<TableEntry> Table, unsigned Opcode) {
596   const TableEntry *I = llvm::lower_bound(Table, Opcode);
597   if (I != Table.end() && I->from == Opcode)
598     return I->to;
599   return -1;
600 }
601 
602 #ifdef NDEBUG
603 #define ASSERT_SORTED(TABLE)
604 #else
605 #define ASSERT_SORTED(TABLE)                                                   \
606   {                                                                            \
607     static std::atomic<bool> TABLE##Checked(false);                            \
608     if (!TABLE##Checked.load(std::memory_order_relaxed)) {                     \
609       assert(is_sorted(TABLE) &&                                               \
610              "All lookup tables must be sorted for efficient access!");        \
611       TABLE##Checked.store(true, std::memory_order_relaxed);                   \
612     }                                                                          \
613   }
614 #endif
615 
616 //===----------------------------------------------------------------------===//
617 // Register File -> Register Stack Mapping Methods
618 //===----------------------------------------------------------------------===//
619 
620 // OpcodeTable - Sorted map of register instructions to their stack version.
621 // The first element is an register file pseudo instruction, the second is the
622 // concrete X86 instruction which uses the register stack.
623 //
624 static const TableEntry OpcodeTable[] = {
625   { X86::ABS_Fp32     , X86::ABS_F     },
626   { X86::ABS_Fp64     , X86::ABS_F     },
627   { X86::ABS_Fp80     , X86::ABS_F     },
628   { X86::ADD_Fp32m    , X86::ADD_F32m  },
629   { X86::ADD_Fp64m    , X86::ADD_F64m  },
630   { X86::ADD_Fp64m32  , X86::ADD_F32m  },
631   { X86::ADD_Fp80m32  , X86::ADD_F32m  },
632   { X86::ADD_Fp80m64  , X86::ADD_F64m  },
633   { X86::ADD_FpI16m32 , X86::ADD_FI16m },
634   { X86::ADD_FpI16m64 , X86::ADD_FI16m },
635   { X86::ADD_FpI16m80 , X86::ADD_FI16m },
636   { X86::ADD_FpI32m32 , X86::ADD_FI32m },
637   { X86::ADD_FpI32m64 , X86::ADD_FI32m },
638   { X86::ADD_FpI32m80 , X86::ADD_FI32m },
639   { X86::CHS_Fp32     , X86::CHS_F     },
640   { X86::CHS_Fp64     , X86::CHS_F     },
641   { X86::CHS_Fp80     , X86::CHS_F     },
642   { X86::CMOVBE_Fp32  , X86::CMOVBE_F  },
643   { X86::CMOVBE_Fp64  , X86::CMOVBE_F  },
644   { X86::CMOVBE_Fp80  , X86::CMOVBE_F  },
645   { X86::CMOVB_Fp32   , X86::CMOVB_F   },
646   { X86::CMOVB_Fp64   , X86::CMOVB_F  },
647   { X86::CMOVB_Fp80   , X86::CMOVB_F  },
648   { X86::CMOVE_Fp32   , X86::CMOVE_F  },
649   { X86::CMOVE_Fp64   , X86::CMOVE_F   },
650   { X86::CMOVE_Fp80   , X86::CMOVE_F   },
651   { X86::CMOVNBE_Fp32 , X86::CMOVNBE_F },
652   { X86::CMOVNBE_Fp64 , X86::CMOVNBE_F },
653   { X86::CMOVNBE_Fp80 , X86::CMOVNBE_F },
654   { X86::CMOVNB_Fp32  , X86::CMOVNB_F  },
655   { X86::CMOVNB_Fp64  , X86::CMOVNB_F  },
656   { X86::CMOVNB_Fp80  , X86::CMOVNB_F  },
657   { X86::CMOVNE_Fp32  , X86::CMOVNE_F  },
658   { X86::CMOVNE_Fp64  , X86::CMOVNE_F  },
659   { X86::CMOVNE_Fp80  , X86::CMOVNE_F  },
660   { X86::CMOVNP_Fp32  , X86::CMOVNP_F  },
661   { X86::CMOVNP_Fp64  , X86::CMOVNP_F  },
662   { X86::CMOVNP_Fp80  , X86::CMOVNP_F  },
663   { X86::CMOVP_Fp32   , X86::CMOVP_F   },
664   { X86::CMOVP_Fp64   , X86::CMOVP_F   },
665   { X86::CMOVP_Fp80   , X86::CMOVP_F   },
666   { X86::COM_FpIr32   , X86::COM_FIr   },
667   { X86::COM_FpIr64   , X86::COM_FIr   },
668   { X86::COM_FpIr80   , X86::COM_FIr   },
669   { X86::COM_Fpr32    , X86::COM_FST0r },
670   { X86::COM_Fpr64    , X86::COM_FST0r },
671   { X86::COM_Fpr80    , X86::COM_FST0r },
672   { X86::DIVR_Fp32m   , X86::DIVR_F32m },
673   { X86::DIVR_Fp64m   , X86::DIVR_F64m },
674   { X86::DIVR_Fp64m32 , X86::DIVR_F32m },
675   { X86::DIVR_Fp80m32 , X86::DIVR_F32m },
676   { X86::DIVR_Fp80m64 , X86::DIVR_F64m },
677   { X86::DIVR_FpI16m32, X86::DIVR_FI16m},
678   { X86::DIVR_FpI16m64, X86::DIVR_FI16m},
679   { X86::DIVR_FpI16m80, X86::DIVR_FI16m},
680   { X86::DIVR_FpI32m32, X86::DIVR_FI32m},
681   { X86::DIVR_FpI32m64, X86::DIVR_FI32m},
682   { X86::DIVR_FpI32m80, X86::DIVR_FI32m},
683   { X86::DIV_Fp32m    , X86::DIV_F32m  },
684   { X86::DIV_Fp64m    , X86::DIV_F64m  },
685   { X86::DIV_Fp64m32  , X86::DIV_F32m  },
686   { X86::DIV_Fp80m32  , X86::DIV_F32m  },
687   { X86::DIV_Fp80m64  , X86::DIV_F64m  },
688   { X86::DIV_FpI16m32 , X86::DIV_FI16m },
689   { X86::DIV_FpI16m64 , X86::DIV_FI16m },
690   { X86::DIV_FpI16m80 , X86::DIV_FI16m },
691   { X86::DIV_FpI32m32 , X86::DIV_FI32m },
692   { X86::DIV_FpI32m64 , X86::DIV_FI32m },
693   { X86::DIV_FpI32m80 , X86::DIV_FI32m },
694   { X86::ILD_Fp16m32  , X86::ILD_F16m  },
695   { X86::ILD_Fp16m64  , X86::ILD_F16m  },
696   { X86::ILD_Fp16m80  , X86::ILD_F16m  },
697   { X86::ILD_Fp32m32  , X86::ILD_F32m  },
698   { X86::ILD_Fp32m64  , X86::ILD_F32m  },
699   { X86::ILD_Fp32m80  , X86::ILD_F32m  },
700   { X86::ILD_Fp64m32  , X86::ILD_F64m  },
701   { X86::ILD_Fp64m64  , X86::ILD_F64m  },
702   { X86::ILD_Fp64m80  , X86::ILD_F64m  },
703   { X86::ISTT_Fp16m32 , X86::ISTT_FP16m},
704   { X86::ISTT_Fp16m64 , X86::ISTT_FP16m},
705   { X86::ISTT_Fp16m80 , X86::ISTT_FP16m},
706   { X86::ISTT_Fp32m32 , X86::ISTT_FP32m},
707   { X86::ISTT_Fp32m64 , X86::ISTT_FP32m},
708   { X86::ISTT_Fp32m80 , X86::ISTT_FP32m},
709   { X86::ISTT_Fp64m32 , X86::ISTT_FP64m},
710   { X86::ISTT_Fp64m64 , X86::ISTT_FP64m},
711   { X86::ISTT_Fp64m80 , X86::ISTT_FP64m},
712   { X86::IST_Fp16m32  , X86::IST_F16m  },
713   { X86::IST_Fp16m64  , X86::IST_F16m  },
714   { X86::IST_Fp16m80  , X86::IST_F16m  },
715   { X86::IST_Fp32m32  , X86::IST_F32m  },
716   { X86::IST_Fp32m64  , X86::IST_F32m  },
717   { X86::IST_Fp32m80  , X86::IST_F32m  },
718   { X86::IST_Fp64m32  , X86::IST_FP64m },
719   { X86::IST_Fp64m64  , X86::IST_FP64m },
720   { X86::IST_Fp64m80  , X86::IST_FP64m },
721   { X86::LD_Fp032     , X86::LD_F0     },
722   { X86::LD_Fp064     , X86::LD_F0     },
723   { X86::LD_Fp080     , X86::LD_F0     },
724   { X86::LD_Fp132     , X86::LD_F1     },
725   { X86::LD_Fp164     , X86::LD_F1     },
726   { X86::LD_Fp180     , X86::LD_F1     },
727   { X86::LD_Fp32m     , X86::LD_F32m   },
728   { X86::LD_Fp32m64   , X86::LD_F32m   },
729   { X86::LD_Fp32m80   , X86::LD_F32m   },
730   { X86::LD_Fp64m     , X86::LD_F64m   },
731   { X86::LD_Fp64m80   , X86::LD_F64m   },
732   { X86::LD_Fp80m     , X86::LD_F80m   },
733   { X86::MUL_Fp32m    , X86::MUL_F32m  },
734   { X86::MUL_Fp64m    , X86::MUL_F64m  },
735   { X86::MUL_Fp64m32  , X86::MUL_F32m  },
736   { X86::MUL_Fp80m32  , X86::MUL_F32m  },
737   { X86::MUL_Fp80m64  , X86::MUL_F64m  },
738   { X86::MUL_FpI16m32 , X86::MUL_FI16m },
739   { X86::MUL_FpI16m64 , X86::MUL_FI16m },
740   { X86::MUL_FpI16m80 , X86::MUL_FI16m },
741   { X86::MUL_FpI32m32 , X86::MUL_FI32m },
742   { X86::MUL_FpI32m64 , X86::MUL_FI32m },
743   { X86::MUL_FpI32m80 , X86::MUL_FI32m },
744   { X86::SQRT_Fp32    , X86::SQRT_F    },
745   { X86::SQRT_Fp64    , X86::SQRT_F    },
746   { X86::SQRT_Fp80    , X86::SQRT_F    },
747   { X86::ST_Fp32m     , X86::ST_F32m   },
748   { X86::ST_Fp64m     , X86::ST_F64m   },
749   { X86::ST_Fp64m32   , X86::ST_F32m   },
750   { X86::ST_Fp80m32   , X86::ST_F32m   },
751   { X86::ST_Fp80m64   , X86::ST_F64m   },
752   { X86::ST_FpP80m    , X86::ST_FP80m  },
753   { X86::SUBR_Fp32m   , X86::SUBR_F32m },
754   { X86::SUBR_Fp64m   , X86::SUBR_F64m },
755   { X86::SUBR_Fp64m32 , X86::SUBR_F32m },
756   { X86::SUBR_Fp80m32 , X86::SUBR_F32m },
757   { X86::SUBR_Fp80m64 , X86::SUBR_F64m },
758   { X86::SUBR_FpI16m32, X86::SUBR_FI16m},
759   { X86::SUBR_FpI16m64, X86::SUBR_FI16m},
760   { X86::SUBR_FpI16m80, X86::SUBR_FI16m},
761   { X86::SUBR_FpI32m32, X86::SUBR_FI32m},
762   { X86::SUBR_FpI32m64, X86::SUBR_FI32m},
763   { X86::SUBR_FpI32m80, X86::SUBR_FI32m},
764   { X86::SUB_Fp32m    , X86::SUB_F32m  },
765   { X86::SUB_Fp64m    , X86::SUB_F64m  },
766   { X86::SUB_Fp64m32  , X86::SUB_F32m  },
767   { X86::SUB_Fp80m32  , X86::SUB_F32m  },
768   { X86::SUB_Fp80m64  , X86::SUB_F64m  },
769   { X86::SUB_FpI16m32 , X86::SUB_FI16m },
770   { X86::SUB_FpI16m64 , X86::SUB_FI16m },
771   { X86::SUB_FpI16m80 , X86::SUB_FI16m },
772   { X86::SUB_FpI32m32 , X86::SUB_FI32m },
773   { X86::SUB_FpI32m64 , X86::SUB_FI32m },
774   { X86::SUB_FpI32m80 , X86::SUB_FI32m },
775   { X86::TST_Fp32     , X86::TST_F     },
776   { X86::TST_Fp64     , X86::TST_F     },
777   { X86::TST_Fp80     , X86::TST_F     },
778   { X86::UCOM_FpIr32  , X86::UCOM_FIr  },
779   { X86::UCOM_FpIr64  , X86::UCOM_FIr  },
780   { X86::UCOM_FpIr80  , X86::UCOM_FIr  },
781   { X86::UCOM_Fpr32   , X86::UCOM_Fr   },
782   { X86::UCOM_Fpr64   , X86::UCOM_Fr   },
783   { X86::UCOM_Fpr80   , X86::UCOM_Fr   },
784   { X86::XAM_Fp32     , X86::XAM_F     },
785   { X86::XAM_Fp64     , X86::XAM_F     },
786   { X86::XAM_Fp80     , X86::XAM_F     },
787 };
788 
789 static unsigned getConcreteOpcode(unsigned Opcode) {
790   ASSERT_SORTED(OpcodeTable);
791   int Opc = Lookup(OpcodeTable, Opcode);
792   assert(Opc != -1 && "FP Stack instruction not in OpcodeTable!");
793   return Opc;
794 }
795 
796 //===----------------------------------------------------------------------===//
797 // Helper Methods
798 //===----------------------------------------------------------------------===//
799 
800 // PopTable - Sorted map of instructions to their popping version.  The first
801 // element is an instruction, the second is the version which pops.
802 //
803 static const TableEntry PopTable[] = {
804   { X86::ADD_FrST0 , X86::ADD_FPrST0  },
805 
806   { X86::COMP_FST0r, X86::FCOMPP      },
807   { X86::COM_FIr   , X86::COM_FIPr    },
808   { X86::COM_FST0r , X86::COMP_FST0r  },
809 
810   { X86::DIVR_FrST0, X86::DIVR_FPrST0 },
811   { X86::DIV_FrST0 , X86::DIV_FPrST0  },
812 
813   { X86::IST_F16m  , X86::IST_FP16m   },
814   { X86::IST_F32m  , X86::IST_FP32m   },
815 
816   { X86::MUL_FrST0 , X86::MUL_FPrST0  },
817 
818   { X86::ST_F32m   , X86::ST_FP32m    },
819   { X86::ST_F64m   , X86::ST_FP64m    },
820   { X86::ST_Frr    , X86::ST_FPrr     },
821 
822   { X86::SUBR_FrST0, X86::SUBR_FPrST0 },
823   { X86::SUB_FrST0 , X86::SUB_FPrST0  },
824 
825   { X86::UCOM_FIr  , X86::UCOM_FIPr   },
826 
827   { X86::UCOM_FPr  , X86::UCOM_FPPr   },
828   { X86::UCOM_Fr   , X86::UCOM_FPr    },
829 };
830 
831 static bool doesInstructionSetFPSW(MachineInstr &MI) {
832   if (const MachineOperand *MO =
833           MI.findRegisterDefOperand(X86::FPSW, /*TRI=*/nullptr))
834     if (!MO->isDead())
835       return true;
836   return false;
837 }
838 
839 static MachineBasicBlock::iterator
840 getNextFPInstruction(MachineBasicBlock::iterator I) {
841   MachineBasicBlock &MBB = *I->getParent();
842   while (++I != MBB.end()) {
843     MachineInstr &MI = *I;
844     if (X86::isX87Instruction(MI))
845       return I;
846   }
847   return MBB.end();
848 }
849 
850 /// popStackAfter - Pop the current value off of the top of the FP stack after
851 /// the specified instruction.  This attempts to be sneaky and combine the pop
852 /// into the instruction itself if possible.  The iterator is left pointing to
853 /// the last instruction, be it a new pop instruction inserted, or the old
854 /// instruction if it was modified in place.
855 ///
856 void FPS::popStackAfter(MachineBasicBlock::iterator &I) {
857   MachineInstr &MI = *I;
858   const DebugLoc &dl = MI.getDebugLoc();
859   ASSERT_SORTED(PopTable);
860 
861   popReg();
862 
863   // Check to see if there is a popping version of this instruction...
864   int Opcode = Lookup(PopTable, I->getOpcode());
865   if (Opcode != -1) {
866     I->setDesc(TII->get(Opcode));
867     if (Opcode == X86::FCOMPP || Opcode == X86::UCOM_FPPr)
868       I->removeOperand(0);
869     MI.dropDebugNumber();
870   } else {    // Insert an explicit pop
871     // If this instruction sets FPSW, which is read in following instruction,
872     // insert pop after that reader.
873     if (doesInstructionSetFPSW(MI)) {
874       MachineBasicBlock &MBB = *MI.getParent();
875       MachineBasicBlock::iterator Next = getNextFPInstruction(I);
876       if (Next != MBB.end() && Next->readsRegister(X86::FPSW, /*TRI=*/nullptr))
877         I = Next;
878     }
879     I = BuildMI(*MBB, ++I, dl, TII->get(X86::ST_FPrr)).addReg(X86::ST0);
880   }
881 }
882 
883 /// freeStackSlotAfter - Free the specified register from the register stack, so
884 /// that it is no longer in a register.  If the register is currently at the top
885 /// of the stack, we just pop the current instruction, otherwise we store the
886 /// current top-of-stack into the specified slot, then pop the top of stack.
887 void FPS::freeStackSlotAfter(MachineBasicBlock::iterator &I, unsigned FPRegNo) {
888   if (getStackEntry(0) == FPRegNo) {  // already at the top of stack? easy.
889     popStackAfter(I);
890     return;
891   }
892 
893   // Otherwise, store the top of stack into the dead slot, killing the operand
894   // without having to add in an explicit xchg then pop.
895   //
896   I = freeStackSlotBefore(++I, FPRegNo);
897 }
898 
899 /// freeStackSlotBefore - Free the specified register without trying any
900 /// folding.
901 MachineBasicBlock::iterator
902 FPS::freeStackSlotBefore(MachineBasicBlock::iterator I, unsigned FPRegNo) {
903   unsigned STReg    = getSTReg(FPRegNo);
904   unsigned OldSlot  = getSlot(FPRegNo);
905   unsigned TopReg   = Stack[StackTop-1];
906   Stack[OldSlot]    = TopReg;
907   RegMap[TopReg]    = OldSlot;
908   RegMap[FPRegNo]   = ~0;
909   Stack[--StackTop] = ~0;
910   return BuildMI(*MBB, I, DebugLoc(), TII->get(X86::ST_FPrr))
911       .addReg(STReg)
912       .getInstr();
913 }
914 
915 /// adjustLiveRegs - Kill and revive registers such that exactly the FP
916 /// registers with a bit in Mask are live.
917 void FPS::adjustLiveRegs(unsigned Mask, MachineBasicBlock::iterator I) {
918   unsigned Defs = Mask;
919   unsigned Kills = 0;
920   for (unsigned i = 0; i < StackTop; ++i) {
921     unsigned RegNo = Stack[i];
922     if (!(Defs & (1 << RegNo)))
923       // This register is live, but we don't want it.
924       Kills |= (1 << RegNo);
925     else
926       // We don't need to imp-def this live register.
927       Defs &= ~(1 << RegNo);
928   }
929   assert((Kills & Defs) == 0 && "Register needs killing and def'ing?");
930 
931   // Produce implicit-defs for free by using killed registers.
932   while (Kills && Defs) {
933     unsigned KReg = llvm::countr_zero(Kills);
934     unsigned DReg = llvm::countr_zero(Defs);
935     LLVM_DEBUG(dbgs() << "Renaming %fp" << KReg << " as imp %fp" << DReg
936                       << "\n");
937     std::swap(Stack[getSlot(KReg)], Stack[getSlot(DReg)]);
938     std::swap(RegMap[KReg], RegMap[DReg]);
939     Kills &= ~(1 << KReg);
940     Defs &= ~(1 << DReg);
941   }
942 
943   // Kill registers by popping.
944   if (Kills && I != MBB->begin()) {
945     MachineBasicBlock::iterator I2 = std::prev(I);
946     while (StackTop) {
947       unsigned KReg = getStackEntry(0);
948       if (!(Kills & (1 << KReg)))
949         break;
950       LLVM_DEBUG(dbgs() << "Popping %fp" << KReg << "\n");
951       popStackAfter(I2);
952       Kills &= ~(1 << KReg);
953     }
954   }
955 
956   // Manually kill the rest.
957   while (Kills) {
958     unsigned KReg = llvm::countr_zero(Kills);
959     LLVM_DEBUG(dbgs() << "Killing %fp" << KReg << "\n");
960     freeStackSlotBefore(I, KReg);
961     Kills &= ~(1 << KReg);
962   }
963 
964   // Load zeros for all the imp-defs.
965   while(Defs) {
966     unsigned DReg = llvm::countr_zero(Defs);
967     LLVM_DEBUG(dbgs() << "Defining %fp" << DReg << " as 0\n");
968     BuildMI(*MBB, I, DebugLoc(), TII->get(X86::LD_F0));
969     pushReg(DReg);
970     Defs &= ~(1 << DReg);
971   }
972 
973   // Now we should have the correct registers live.
974   LLVM_DEBUG(dumpStack());
975   assert(StackTop == (unsigned)llvm::popcount(Mask) && "Live count mismatch");
976 }
977 
978 /// shuffleStackTop - emit fxch instructions before I to shuffle the top
979 /// FixCount entries into the order given by FixStack.
980 /// FIXME: Is there a better algorithm than insertion sort?
981 void FPS::shuffleStackTop(const unsigned char *FixStack,
982                           unsigned FixCount,
983                           MachineBasicBlock::iterator I) {
984   // Move items into place, starting from the desired stack bottom.
985   while (FixCount--) {
986     // Old register at position FixCount.
987     unsigned OldReg = getStackEntry(FixCount);
988     // Desired register at position FixCount.
989     unsigned Reg = FixStack[FixCount];
990     if (Reg == OldReg)
991       continue;
992     // (Reg st0) (OldReg st0) = (Reg OldReg st0)
993     moveToTop(Reg, I);
994     if (FixCount > 0)
995       moveToTop(OldReg, I);
996   }
997   LLVM_DEBUG(dumpStack());
998 }
999 
1000 
1001 //===----------------------------------------------------------------------===//
1002 // Instruction transformation implementation
1003 //===----------------------------------------------------------------------===//
1004 
1005 void FPS::handleCall(MachineBasicBlock::iterator &I) {
1006   MachineInstr &MI = *I;
1007   unsigned STReturns = 0;
1008 
1009   bool ClobbersFPStack = false;
1010   for (unsigned i = 0, e = MI.getNumOperands(); i != e; ++i) {
1011     MachineOperand &Op = MI.getOperand(i);
1012     // Check if this call clobbers the FP stack.
1013     // is sufficient.
1014     if (Op.isRegMask()) {
1015       bool ClobbersFP0 = Op.clobbersPhysReg(X86::FP0);
1016 #ifndef NDEBUG
1017       static_assert(X86::FP7 - X86::FP0 == 7, "sequential FP regnumbers");
1018       for (unsigned i = 1; i != 8; ++i)
1019         assert(Op.clobbersPhysReg(X86::FP0 + i) == ClobbersFP0 &&
1020                "Inconsistent FP register clobber");
1021 #endif
1022 
1023       if (ClobbersFP0)
1024         ClobbersFPStack = true;
1025     }
1026 
1027     if (!Op.isReg() || Op.getReg() < X86::FP0 || Op.getReg() > X86::FP6)
1028       continue;
1029 
1030     assert(Op.isImplicit() && "Expected implicit def/use");
1031 
1032     if (Op.isDef())
1033       STReturns |= 1 << getFPReg(Op);
1034 
1035     // Remove the operand so that later passes don't see it.
1036     MI.removeOperand(i);
1037     --i;
1038     --e;
1039   }
1040 
1041   // Most calls should have a regmask that clobbers the FP registers. If it
1042   // isn't present then the register allocator didn't spill the FP registers
1043   // so they are still on the stack.
1044   assert((ClobbersFPStack || STReturns == 0) &&
1045          "ST returns without FP stack clobber");
1046   if (!ClobbersFPStack)
1047     return;
1048 
1049   unsigned N = llvm::countr_one(STReturns);
1050 
1051   // FP registers used for function return must be consecutive starting at
1052   // FP0
1053   assert(STReturns == 0 || (isMask_32(STReturns) && N <= 2));
1054 
1055   // Reset the FP Stack - It is required because of possible leftovers from
1056   // passed arguments. The caller should assume that the FP stack is
1057   // returned empty (unless the callee returns values on FP stack).
1058   while (StackTop > 0)
1059     popReg();
1060 
1061   for (unsigned I = 0; I < N; ++I)
1062     pushReg(N - I - 1);
1063 
1064   // If this call has been modified, drop all variable values defined by it.
1065   // We can't track them once they've been stackified.
1066   if (STReturns)
1067     I->dropDebugNumber();
1068 }
1069 
1070 /// If RET has an FP register use operand, pass the first one in ST(0) and
1071 /// the second one in ST(1).
1072 void FPS::handleReturn(MachineBasicBlock::iterator &I) {
1073   MachineInstr &MI = *I;
1074 
1075   // Find the register operands.
1076   unsigned FirstFPRegOp = ~0U, SecondFPRegOp = ~0U;
1077   unsigned LiveMask = 0;
1078 
1079   for (unsigned i = 0, e = MI.getNumOperands(); i != e; ++i) {
1080     MachineOperand &Op = MI.getOperand(i);
1081     if (!Op.isReg() || Op.getReg() < X86::FP0 || Op.getReg() > X86::FP6)
1082       continue;
1083     // FP Register uses must be kills unless there are two uses of the same
1084     // register, in which case only one will be a kill.
1085     assert(Op.isUse() &&
1086            (Op.isKill() ||                  // Marked kill.
1087             getFPReg(Op) == FirstFPRegOp || // Second instance.
1088             MI.killsRegister(Op.getReg(),
1089                              /*TRI=*/nullptr)) && // Later use is marked kill.
1090            "Ret only defs operands, and values aren't live beyond it");
1091 
1092     if (FirstFPRegOp == ~0U)
1093       FirstFPRegOp = getFPReg(Op);
1094     else {
1095       assert(SecondFPRegOp == ~0U && "More than two fp operands!");
1096       SecondFPRegOp = getFPReg(Op);
1097     }
1098     LiveMask |= (1 << getFPReg(Op));
1099 
1100     // Remove the operand so that later passes don't see it.
1101     MI.removeOperand(i);
1102     --i;
1103     --e;
1104   }
1105 
1106   // We may have been carrying spurious live-ins, so make sure only the
1107   // returned registers are left live.
1108   adjustLiveRegs(LiveMask, MI);
1109   if (!LiveMask) return;  // Quick check to see if any are possible.
1110 
1111   // There are only four possibilities here:
1112   // 1) we are returning a single FP value.  In this case, it has to be in
1113   //    ST(0) already, so just declare success by removing the value from the
1114   //    FP Stack.
1115   if (SecondFPRegOp == ~0U) {
1116     // Assert that the top of stack contains the right FP register.
1117     assert(StackTop == 1 && FirstFPRegOp == getStackEntry(0) &&
1118            "Top of stack not the right register for RET!");
1119 
1120     // Ok, everything is good, mark the value as not being on the stack
1121     // anymore so that our assertion about the stack being empty at end of
1122     // block doesn't fire.
1123     StackTop = 0;
1124     return;
1125   }
1126 
1127   // Otherwise, we are returning two values:
1128   // 2) If returning the same value for both, we only have one thing in the FP
1129   //    stack.  Consider:  RET FP1, FP1
1130   if (StackTop == 1) {
1131     assert(FirstFPRegOp == SecondFPRegOp && FirstFPRegOp == getStackEntry(0)&&
1132            "Stack misconfiguration for RET!");
1133 
1134     // Duplicate the TOS so that we return it twice.  Just pick some other FPx
1135     // register to hold it.
1136     unsigned NewReg = ScratchFPReg;
1137     duplicateToTop(FirstFPRegOp, NewReg, MI);
1138     FirstFPRegOp = NewReg;
1139   }
1140 
1141   /// Okay we know we have two different FPx operands now:
1142   assert(StackTop == 2 && "Must have two values live!");
1143 
1144   /// 3) If SecondFPRegOp is currently in ST(0) and FirstFPRegOp is currently
1145   ///    in ST(1).  In this case, emit an fxch.
1146   if (getStackEntry(0) == SecondFPRegOp) {
1147     assert(getStackEntry(1) == FirstFPRegOp && "Unknown regs live");
1148     moveToTop(FirstFPRegOp, MI);
1149   }
1150 
1151   /// 4) Finally, FirstFPRegOp must be in ST(0) and SecondFPRegOp must be in
1152   /// ST(1).  Just remove both from our understanding of the stack and return.
1153   assert(getStackEntry(0) == FirstFPRegOp && "Unknown regs live");
1154   assert(getStackEntry(1) == SecondFPRegOp && "Unknown regs live");
1155   StackTop = 0;
1156 }
1157 
1158 /// handleZeroArgFP - ST(0) = fld0    ST(0) = flds <mem>
1159 ///
1160 void FPS::handleZeroArgFP(MachineBasicBlock::iterator &I) {
1161   MachineInstr &MI = *I;
1162   unsigned DestReg = getFPReg(MI.getOperand(0));
1163 
1164   // Change from the pseudo instruction to the concrete instruction.
1165   MI.removeOperand(0); // Remove the explicit ST(0) operand
1166   MI.setDesc(TII->get(getConcreteOpcode(MI.getOpcode())));
1167   MI.addOperand(
1168       MachineOperand::CreateReg(X86::ST0, /*isDef*/ true, /*isImp*/ true));
1169 
1170   // Result gets pushed on the stack.
1171   pushReg(DestReg);
1172 
1173   MI.dropDebugNumber();
1174 }
1175 
1176 /// handleOneArgFP - fst <mem>, ST(0)
1177 ///
1178 void FPS::handleOneArgFP(MachineBasicBlock::iterator &I) {
1179   MachineInstr &MI = *I;
1180   unsigned NumOps = MI.getDesc().getNumOperands();
1181   assert((NumOps == X86::AddrNumOperands + 1 || NumOps == 1) &&
1182          "Can only handle fst* & ftst instructions!");
1183 
1184   // Is this the last use of the source register?
1185   unsigned Reg = getFPReg(MI.getOperand(NumOps - 1));
1186   bool KillsSrc = MI.killsRegister(X86::FP0 + Reg, /*TRI=*/nullptr);
1187 
1188   // FISTP64m is strange because there isn't a non-popping versions.
1189   // If we have one _and_ we don't want to pop the operand, duplicate the value
1190   // on the stack instead of moving it.  This ensure that popping the value is
1191   // always ok.
1192   // Ditto FISTTP16m, FISTTP32m, FISTTP64m, ST_FpP80m.
1193   //
1194   if (!KillsSrc && (MI.getOpcode() == X86::IST_Fp64m32 ||
1195                     MI.getOpcode() == X86::ISTT_Fp16m32 ||
1196                     MI.getOpcode() == X86::ISTT_Fp32m32 ||
1197                     MI.getOpcode() == X86::ISTT_Fp64m32 ||
1198                     MI.getOpcode() == X86::IST_Fp64m64 ||
1199                     MI.getOpcode() == X86::ISTT_Fp16m64 ||
1200                     MI.getOpcode() == X86::ISTT_Fp32m64 ||
1201                     MI.getOpcode() == X86::ISTT_Fp64m64 ||
1202                     MI.getOpcode() == X86::IST_Fp64m80 ||
1203                     MI.getOpcode() == X86::ISTT_Fp16m80 ||
1204                     MI.getOpcode() == X86::ISTT_Fp32m80 ||
1205                     MI.getOpcode() == X86::ISTT_Fp64m80 ||
1206                     MI.getOpcode() == X86::ST_FpP80m)) {
1207     duplicateToTop(Reg, ScratchFPReg, I);
1208   } else {
1209     moveToTop(Reg, I);            // Move to the top of the stack...
1210   }
1211 
1212   // Convert from the pseudo instruction to the concrete instruction.
1213   MI.removeOperand(NumOps - 1); // Remove explicit ST(0) operand
1214   MI.setDesc(TII->get(getConcreteOpcode(MI.getOpcode())));
1215   MI.addOperand(
1216       MachineOperand::CreateReg(X86::ST0, /*isDef*/ false, /*isImp*/ true));
1217 
1218   if (MI.getOpcode() == X86::IST_FP64m || MI.getOpcode() == X86::ISTT_FP16m ||
1219       MI.getOpcode() == X86::ISTT_FP32m || MI.getOpcode() == X86::ISTT_FP64m ||
1220       MI.getOpcode() == X86::ST_FP80m) {
1221     if (StackTop == 0)
1222       report_fatal_error("Stack empty??");
1223     --StackTop;
1224   } else if (KillsSrc) { // Last use of operand?
1225     popStackAfter(I);
1226   }
1227 
1228   MI.dropDebugNumber();
1229 }
1230 
1231 
1232 /// handleOneArgFPRW: Handle instructions that read from the top of stack and
1233 /// replace the value with a newly computed value.  These instructions may have
1234 /// non-fp operands after their FP operands.
1235 ///
1236 ///  Examples:
1237 ///     R1 = fchs R2
1238 ///     R1 = fadd R2, [mem]
1239 ///
1240 void FPS::handleOneArgFPRW(MachineBasicBlock::iterator &I) {
1241   MachineInstr &MI = *I;
1242 #ifndef NDEBUG
1243   unsigned NumOps = MI.getDesc().getNumOperands();
1244   assert(NumOps >= 2 && "FPRW instructions must have 2 ops!!");
1245 #endif
1246 
1247   // Is this the last use of the source register?
1248   unsigned Reg = getFPReg(MI.getOperand(1));
1249   bool KillsSrc = MI.killsRegister(X86::FP0 + Reg, /*TRI=*/nullptr);
1250 
1251   if (KillsSrc) {
1252     // If this is the last use of the source register, just make sure it's on
1253     // the top of the stack.
1254     moveToTop(Reg, I);
1255     if (StackTop == 0)
1256       report_fatal_error("Stack cannot be empty!");
1257     --StackTop;
1258     pushReg(getFPReg(MI.getOperand(0)));
1259   } else {
1260     // If this is not the last use of the source register, _copy_ it to the top
1261     // of the stack.
1262     duplicateToTop(Reg, getFPReg(MI.getOperand(0)), I);
1263   }
1264 
1265   // Change from the pseudo instruction to the concrete instruction.
1266   MI.removeOperand(1); // Drop the source operand.
1267   MI.removeOperand(0); // Drop the destination operand.
1268   MI.setDesc(TII->get(getConcreteOpcode(MI.getOpcode())));
1269   MI.dropDebugNumber();
1270 }
1271 
1272 
1273 //===----------------------------------------------------------------------===//
1274 // Define tables of various ways to map pseudo instructions
1275 //
1276 
1277 // ForwardST0Table - Map: A = B op C  into: ST(0) = ST(0) op ST(i)
1278 static const TableEntry ForwardST0Table[] = {
1279   { X86::ADD_Fp32  , X86::ADD_FST0r },
1280   { X86::ADD_Fp64  , X86::ADD_FST0r },
1281   { X86::ADD_Fp80  , X86::ADD_FST0r },
1282   { X86::DIV_Fp32  , X86::DIV_FST0r },
1283   { X86::DIV_Fp64  , X86::DIV_FST0r },
1284   { X86::DIV_Fp80  , X86::DIV_FST0r },
1285   { X86::MUL_Fp32  , X86::MUL_FST0r },
1286   { X86::MUL_Fp64  , X86::MUL_FST0r },
1287   { X86::MUL_Fp80  , X86::MUL_FST0r },
1288   { X86::SUB_Fp32  , X86::SUB_FST0r },
1289   { X86::SUB_Fp64  , X86::SUB_FST0r },
1290   { X86::SUB_Fp80  , X86::SUB_FST0r },
1291 };
1292 
1293 // ReverseST0Table - Map: A = B op C  into: ST(0) = ST(i) op ST(0)
1294 static const TableEntry ReverseST0Table[] = {
1295   { X86::ADD_Fp32  , X86::ADD_FST0r  },   // commutative
1296   { X86::ADD_Fp64  , X86::ADD_FST0r  },   // commutative
1297   { X86::ADD_Fp80  , X86::ADD_FST0r  },   // commutative
1298   { X86::DIV_Fp32  , X86::DIVR_FST0r },
1299   { X86::DIV_Fp64  , X86::DIVR_FST0r },
1300   { X86::DIV_Fp80  , X86::DIVR_FST0r },
1301   { X86::MUL_Fp32  , X86::MUL_FST0r  },   // commutative
1302   { X86::MUL_Fp64  , X86::MUL_FST0r  },   // commutative
1303   { X86::MUL_Fp80  , X86::MUL_FST0r  },   // commutative
1304   { X86::SUB_Fp32  , X86::SUBR_FST0r },
1305   { X86::SUB_Fp64  , X86::SUBR_FST0r },
1306   { X86::SUB_Fp80  , X86::SUBR_FST0r },
1307 };
1308 
1309 // ForwardSTiTable - Map: A = B op C  into: ST(i) = ST(0) op ST(i)
1310 static const TableEntry ForwardSTiTable[] = {
1311   { X86::ADD_Fp32  , X86::ADD_FrST0  },   // commutative
1312   { X86::ADD_Fp64  , X86::ADD_FrST0  },   // commutative
1313   { X86::ADD_Fp80  , X86::ADD_FrST0  },   // commutative
1314   { X86::DIV_Fp32  , X86::DIVR_FrST0 },
1315   { X86::DIV_Fp64  , X86::DIVR_FrST0 },
1316   { X86::DIV_Fp80  , X86::DIVR_FrST0 },
1317   { X86::MUL_Fp32  , X86::MUL_FrST0  },   // commutative
1318   { X86::MUL_Fp64  , X86::MUL_FrST0  },   // commutative
1319   { X86::MUL_Fp80  , X86::MUL_FrST0  },   // commutative
1320   { X86::SUB_Fp32  , X86::SUBR_FrST0 },
1321   { X86::SUB_Fp64  , X86::SUBR_FrST0 },
1322   { X86::SUB_Fp80  , X86::SUBR_FrST0 },
1323 };
1324 
1325 // ReverseSTiTable - Map: A = B op C  into: ST(i) = ST(i) op ST(0)
1326 static const TableEntry ReverseSTiTable[] = {
1327   { X86::ADD_Fp32  , X86::ADD_FrST0 },
1328   { X86::ADD_Fp64  , X86::ADD_FrST0 },
1329   { X86::ADD_Fp80  , X86::ADD_FrST0 },
1330   { X86::DIV_Fp32  , X86::DIV_FrST0 },
1331   { X86::DIV_Fp64  , X86::DIV_FrST0 },
1332   { X86::DIV_Fp80  , X86::DIV_FrST0 },
1333   { X86::MUL_Fp32  , X86::MUL_FrST0 },
1334   { X86::MUL_Fp64  , X86::MUL_FrST0 },
1335   { X86::MUL_Fp80  , X86::MUL_FrST0 },
1336   { X86::SUB_Fp32  , X86::SUB_FrST0 },
1337   { X86::SUB_Fp64  , X86::SUB_FrST0 },
1338   { X86::SUB_Fp80  , X86::SUB_FrST0 },
1339 };
1340 
1341 
1342 /// handleTwoArgFP - Handle instructions like FADD and friends which are virtual
1343 /// instructions which need to be simplified and possibly transformed.
1344 ///
1345 /// Result: ST(0) = fsub  ST(0), ST(i)
1346 ///         ST(i) = fsub  ST(0), ST(i)
1347 ///         ST(0) = fsubr ST(0), ST(i)
1348 ///         ST(i) = fsubr ST(0), ST(i)
1349 ///
1350 void FPS::handleTwoArgFP(MachineBasicBlock::iterator &I) {
1351   ASSERT_SORTED(ForwardST0Table); ASSERT_SORTED(ReverseST0Table);
1352   ASSERT_SORTED(ForwardSTiTable); ASSERT_SORTED(ReverseSTiTable);
1353   MachineInstr &MI = *I;
1354 
1355   unsigned NumOperands = MI.getDesc().getNumOperands();
1356   assert(NumOperands == 3 && "Illegal TwoArgFP instruction!");
1357   unsigned Dest = getFPReg(MI.getOperand(0));
1358   unsigned Op0 = getFPReg(MI.getOperand(NumOperands - 2));
1359   unsigned Op1 = getFPReg(MI.getOperand(NumOperands - 1));
1360   bool KillsOp0 = MI.killsRegister(X86::FP0 + Op0, /*TRI=*/nullptr);
1361   bool KillsOp1 = MI.killsRegister(X86::FP0 + Op1, /*TRI=*/nullptr);
1362   const DebugLoc &dl = MI.getDebugLoc();
1363 
1364   unsigned TOS = getStackEntry(0);
1365 
1366   // One of our operands must be on the top of the stack.  If neither is yet, we
1367   // need to move one.
1368   if (Op0 != TOS && Op1 != TOS) {   // No operand at TOS?
1369     // We can choose to move either operand to the top of the stack.  If one of
1370     // the operands is killed by this instruction, we want that one so that we
1371     // can update right on top of the old version.
1372     if (KillsOp0) {
1373       moveToTop(Op0, I);         // Move dead operand to TOS.
1374       TOS = Op0;
1375     } else if (KillsOp1) {
1376       moveToTop(Op1, I);
1377       TOS = Op1;
1378     } else {
1379       // All of the operands are live after this instruction executes, so we
1380       // cannot update on top of any operand.  Because of this, we must
1381       // duplicate one of the stack elements to the top.  It doesn't matter
1382       // which one we pick.
1383       //
1384       duplicateToTop(Op0, Dest, I);
1385       Op0 = TOS = Dest;
1386       KillsOp0 = true;
1387     }
1388   } else if (!KillsOp0 && !KillsOp1) {
1389     // If we DO have one of our operands at the top of the stack, but we don't
1390     // have a dead operand, we must duplicate one of the operands to a new slot
1391     // on the stack.
1392     duplicateToTop(Op0, Dest, I);
1393     Op0 = TOS = Dest;
1394     KillsOp0 = true;
1395   }
1396 
1397   // Now we know that one of our operands is on the top of the stack, and at
1398   // least one of our operands is killed by this instruction.
1399   assert((TOS == Op0 || TOS == Op1) && (KillsOp0 || KillsOp1) &&
1400          "Stack conditions not set up right!");
1401 
1402   // We decide which form to use based on what is on the top of the stack, and
1403   // which operand is killed by this instruction.
1404   ArrayRef<TableEntry> InstTable;
1405   bool isForward = TOS == Op0;
1406   bool updateST0 = (TOS == Op0 && !KillsOp1) || (TOS == Op1 && !KillsOp0);
1407   if (updateST0) {
1408     if (isForward)
1409       InstTable = ForwardST0Table;
1410     else
1411       InstTable = ReverseST0Table;
1412   } else {
1413     if (isForward)
1414       InstTable = ForwardSTiTable;
1415     else
1416       InstTable = ReverseSTiTable;
1417   }
1418 
1419   int Opcode = Lookup(InstTable, MI.getOpcode());
1420   assert(Opcode != -1 && "Unknown TwoArgFP pseudo instruction!");
1421 
1422   // NotTOS - The register which is not on the top of stack...
1423   unsigned NotTOS = (TOS == Op0) ? Op1 : Op0;
1424 
1425   // Replace the old instruction with a new instruction
1426   MBB->remove(&*I++);
1427   I = BuildMI(*MBB, I, dl, TII->get(Opcode)).addReg(getSTReg(NotTOS));
1428 
1429   if (!MI.mayRaiseFPException())
1430     I->setFlag(MachineInstr::MIFlag::NoFPExcept);
1431 
1432   // If both operands are killed, pop one off of the stack in addition to
1433   // overwriting the other one.
1434   if (KillsOp0 && KillsOp1 && Op0 != Op1) {
1435     assert(!updateST0 && "Should have updated other operand!");
1436     popStackAfter(I);   // Pop the top of stack
1437   }
1438 
1439   // Update stack information so that we know the destination register is now on
1440   // the stack.
1441   unsigned UpdatedSlot = getSlot(updateST0 ? TOS : NotTOS);
1442   assert(UpdatedSlot < StackTop && Dest < 7);
1443   Stack[UpdatedSlot]   = Dest;
1444   RegMap[Dest]         = UpdatedSlot;
1445   MBB->getParent()->deleteMachineInstr(&MI); // Remove the old instruction
1446 }
1447 
1448 /// handleCompareFP - Handle FUCOM and FUCOMI instructions, which have two FP
1449 /// register arguments and no explicit destinations.
1450 ///
1451 void FPS::handleCompareFP(MachineBasicBlock::iterator &I) {
1452   MachineInstr &MI = *I;
1453 
1454   unsigned NumOperands = MI.getDesc().getNumOperands();
1455   assert(NumOperands == 2 && "Illegal FUCOM* instruction!");
1456   unsigned Op0 = getFPReg(MI.getOperand(NumOperands - 2));
1457   unsigned Op1 = getFPReg(MI.getOperand(NumOperands - 1));
1458   bool KillsOp0 = MI.killsRegister(X86::FP0 + Op0, /*TRI=*/nullptr);
1459   bool KillsOp1 = MI.killsRegister(X86::FP0 + Op1, /*TRI=*/nullptr);
1460 
1461   // Make sure the first operand is on the top of stack, the other one can be
1462   // anywhere.
1463   moveToTop(Op0, I);
1464 
1465   // Change from the pseudo instruction to the concrete instruction.
1466   MI.getOperand(0).setReg(getSTReg(Op1));
1467   MI.removeOperand(1);
1468   MI.setDesc(TII->get(getConcreteOpcode(MI.getOpcode())));
1469   MI.dropDebugNumber();
1470 
1471   // If any of the operands are killed by this instruction, free them.
1472   if (KillsOp0) freeStackSlotAfter(I, Op0);
1473   if (KillsOp1 && Op0 != Op1) freeStackSlotAfter(I, Op1);
1474 }
1475 
1476 /// handleCondMovFP - Handle two address conditional move instructions.  These
1477 /// instructions move a st(i) register to st(0) iff a condition is true.  These
1478 /// instructions require that the first operand is at the top of the stack, but
1479 /// otherwise don't modify the stack at all.
1480 void FPS::handleCondMovFP(MachineBasicBlock::iterator &I) {
1481   MachineInstr &MI = *I;
1482 
1483   unsigned Op0 = getFPReg(MI.getOperand(0));
1484   unsigned Op1 = getFPReg(MI.getOperand(2));
1485   bool KillsOp1 = MI.killsRegister(X86::FP0 + Op1, /*TRI=*/nullptr);
1486 
1487   // The first operand *must* be on the top of the stack.
1488   moveToTop(Op0, I);
1489 
1490   // Change the second operand to the stack register that the operand is in.
1491   // Change from the pseudo instruction to the concrete instruction.
1492   MI.removeOperand(0);
1493   MI.removeOperand(1);
1494   MI.getOperand(0).setReg(getSTReg(Op1));
1495   MI.setDesc(TII->get(getConcreteOpcode(MI.getOpcode())));
1496   MI.dropDebugNumber();
1497 
1498   // If we kill the second operand, make sure to pop it from the stack.
1499   if (Op0 != Op1 && KillsOp1) {
1500     // Get this value off of the register stack.
1501     freeStackSlotAfter(I, Op1);
1502   }
1503 }
1504 
1505 
1506 /// handleSpecialFP - Handle special instructions which behave unlike other
1507 /// floating point instructions.  This is primarily intended for use by pseudo
1508 /// instructions.
1509 ///
1510 void FPS::handleSpecialFP(MachineBasicBlock::iterator &Inst) {
1511   MachineInstr &MI = *Inst;
1512 
1513   if (MI.isCall()) {
1514     handleCall(Inst);
1515     return;
1516   }
1517 
1518   if (MI.isReturn()) {
1519     handleReturn(Inst);
1520     return;
1521   }
1522 
1523   switch (MI.getOpcode()) {
1524   default: llvm_unreachable("Unknown SpecialFP instruction!");
1525   case TargetOpcode::COPY: {
1526     // We handle three kinds of copies: FP <- FP, FP <- ST, and ST <- FP.
1527     const MachineOperand &MO1 = MI.getOperand(1);
1528     const MachineOperand &MO0 = MI.getOperand(0);
1529     bool KillsSrc = MI.killsRegister(MO1.getReg(), /*TRI=*/nullptr);
1530 
1531     // FP <- FP copy.
1532     unsigned DstFP = getFPReg(MO0);
1533     unsigned SrcFP = getFPReg(MO1);
1534     assert(isLive(SrcFP) && "Cannot copy dead register");
1535     if (KillsSrc) {
1536       // If the input operand is killed, we can just change the owner of the
1537       // incoming stack slot into the result.
1538       unsigned Slot = getSlot(SrcFP);
1539       Stack[Slot] = DstFP;
1540       RegMap[DstFP] = Slot;
1541     } else {
1542       // For COPY we just duplicate the specified value to a new stack slot.
1543       // This could be made better, but would require substantial changes.
1544       duplicateToTop(SrcFP, DstFP, Inst);
1545     }
1546     break;
1547   }
1548 
1549   case TargetOpcode::IMPLICIT_DEF: {
1550     // All FP registers must be explicitly defined, so load a 0 instead.
1551     unsigned Reg = MI.getOperand(0).getReg() - X86::FP0;
1552     LLVM_DEBUG(dbgs() << "Emitting LD_F0 for implicit FP" << Reg << '\n');
1553     BuildMI(*MBB, Inst, MI.getDebugLoc(), TII->get(X86::LD_F0));
1554     pushReg(Reg);
1555     break;
1556   }
1557 
1558   case TargetOpcode::INLINEASM:
1559   case TargetOpcode::INLINEASM_BR: {
1560     // The inline asm MachineInstr currently only *uses* FP registers for the
1561     // 'f' constraint.  These should be turned into the current ST(x) register
1562     // in the machine instr.
1563     //
1564     // There are special rules for x87 inline assembly. The compiler must know
1565     // exactly how many registers are popped and pushed implicitly by the asm.
1566     // Otherwise it is not possible to restore the stack state after the inline
1567     // asm.
1568     //
1569     // There are 3 kinds of input operands:
1570     //
1571     // 1. Popped inputs. These must appear at the stack top in ST0-STn. A
1572     //    popped input operand must be in a fixed stack slot, and it is either
1573     //    tied to an output operand, or in the clobber list. The MI has ST use
1574     //    and def operands for these inputs.
1575     //
1576     // 2. Fixed inputs. These inputs appear in fixed stack slots, but are
1577     //    preserved by the inline asm. The fixed stack slots must be STn-STm
1578     //    following the popped inputs. A fixed input operand cannot be tied to
1579     //    an output or appear in the clobber list. The MI has ST use operands
1580     //    and no defs for these inputs.
1581     //
1582     // 3. Preserved inputs. These inputs use the "f" constraint which is
1583     //    represented as an FP register. The inline asm won't change these
1584     //    stack slots.
1585     //
1586     // Outputs must be in ST registers, FP outputs are not allowed. Clobbered
1587     // registers do not count as output operands. The inline asm changes the
1588     // stack as if it popped all the popped inputs and then pushed all the
1589     // output operands.
1590 
1591     // Scan the assembly for ST registers used, defined and clobbered. We can
1592     // only tell clobbers from defs by looking at the asm descriptor.
1593     unsigned STUses = 0, STDefs = 0, STClobbers = 0;
1594     unsigned NumOps = 0;
1595     SmallSet<unsigned, 1> FRegIdx;
1596     unsigned RCID;
1597 
1598     for (unsigned i = InlineAsm::MIOp_FirstOperand, e = MI.getNumOperands();
1599          i != e && MI.getOperand(i).isImm(); i += 1 + NumOps) {
1600       unsigned Flags = MI.getOperand(i).getImm();
1601       const InlineAsm::Flag F(Flags);
1602 
1603       NumOps = F.getNumOperandRegisters();
1604       if (NumOps != 1)
1605         continue;
1606       const MachineOperand &MO = MI.getOperand(i + 1);
1607       if (!MO.isReg())
1608         continue;
1609       unsigned STReg = MO.getReg() - X86::FP0;
1610       if (STReg >= 8)
1611         continue;
1612 
1613       // If the flag has a register class constraint, this must be an operand
1614       // with constraint "f". Record its index and continue.
1615       if (F.hasRegClassConstraint(RCID)) {
1616         FRegIdx.insert(i + 1);
1617         continue;
1618       }
1619 
1620       switch (F.getKind()) {
1621       case InlineAsm::Kind::RegUse:
1622         STUses |= (1u << STReg);
1623         break;
1624       case InlineAsm::Kind::RegDef:
1625       case InlineAsm::Kind::RegDefEarlyClobber:
1626         STDefs |= (1u << STReg);
1627         break;
1628       case InlineAsm::Kind::Clobber:
1629         STClobbers |= (1u << STReg);
1630         break;
1631       default:
1632         break;
1633       }
1634     }
1635 
1636     if (STUses && !isMask_32(STUses))
1637       MI.emitError("fixed input regs must be last on the x87 stack");
1638     unsigned NumSTUses = llvm::countr_one(STUses);
1639 
1640     // Defs must be contiguous from the stack top. ST0-STn.
1641     if (STDefs && !isMask_32(STDefs)) {
1642       MI.emitError("output regs must be last on the x87 stack");
1643       STDefs = NextPowerOf2(STDefs) - 1;
1644     }
1645     unsigned NumSTDefs = llvm::countr_one(STDefs);
1646 
1647     // So must the clobbered stack slots. ST0-STm, m >= n.
1648     if (STClobbers && !isMask_32(STDefs | STClobbers))
1649       MI.emitError("clobbers must be last on the x87 stack");
1650 
1651     // Popped inputs are the ones that are also clobbered or defined.
1652     unsigned STPopped = STUses & (STDefs | STClobbers);
1653     if (STPopped && !isMask_32(STPopped))
1654       MI.emitError("implicitly popped regs must be last on the x87 stack");
1655     unsigned NumSTPopped = llvm::countr_one(STPopped);
1656 
1657     LLVM_DEBUG(dbgs() << "Asm uses " << NumSTUses << " fixed regs, pops "
1658                       << NumSTPopped << ", and defines " << NumSTDefs
1659                       << " regs.\n");
1660 
1661 #ifndef NDEBUG
1662     // If any input operand uses constraint "f", all output register
1663     // constraints must be early-clobber defs.
1664     for (unsigned I = 0, E = MI.getNumOperands(); I < E; ++I)
1665       if (FRegIdx.count(I)) {
1666         assert((1 << getFPReg(MI.getOperand(I)) & STDefs) == 0 &&
1667                "Operands with constraint \"f\" cannot overlap with defs");
1668       }
1669 #endif
1670 
1671     // Collect all FP registers (register operands with constraints "t", "u",
1672     // and "f") to kill afer the instruction.
1673     unsigned FPKills = ((1u << NumFPRegs) - 1) & ~0xff;
1674     for (const MachineOperand &Op : MI.operands()) {
1675       if (!Op.isReg() || Op.getReg() < X86::FP0 || Op.getReg() > X86::FP6)
1676         continue;
1677       unsigned FPReg = getFPReg(Op);
1678 
1679       // If we kill this operand, make sure to pop it from the stack after the
1680       // asm.  We just remember it for now, and pop them all off at the end in
1681       // a batch.
1682       if (Op.isUse() && Op.isKill())
1683         FPKills |= 1U << FPReg;
1684     }
1685 
1686     // Do not include registers that are implicitly popped by defs/clobbers.
1687     FPKills &= ~(STDefs | STClobbers);
1688 
1689     // Now we can rearrange the live registers to match what was requested.
1690     unsigned char STUsesArray[8];
1691 
1692     for (unsigned I = 0; I < NumSTUses; ++I)
1693       STUsesArray[I] = I;
1694 
1695     shuffleStackTop(STUsesArray, NumSTUses, Inst);
1696     LLVM_DEBUG({
1697       dbgs() << "Before asm: ";
1698       dumpStack();
1699     });
1700 
1701     // With the stack layout fixed, rewrite the FP registers.
1702     for (unsigned i = 0, e = MI.getNumOperands(); i != e; ++i) {
1703       MachineOperand &Op = MI.getOperand(i);
1704       if (!Op.isReg() || Op.getReg() < X86::FP0 || Op.getReg() > X86::FP6)
1705         continue;
1706 
1707       unsigned FPReg = getFPReg(Op);
1708 
1709       if (FRegIdx.count(i))
1710         // Operand with constraint "f".
1711         Op.setReg(getSTReg(FPReg));
1712       else
1713         // Operand with a single register class constraint ("t" or "u").
1714         Op.setReg(X86::ST0 + FPReg);
1715     }
1716 
1717     // Simulate the inline asm popping its inputs and pushing its outputs.
1718     StackTop -= NumSTPopped;
1719 
1720     for (unsigned i = 0; i < NumSTDefs; ++i)
1721       pushReg(NumSTDefs - i - 1);
1722 
1723     // If this asm kills any FP registers (is the last use of them) we must
1724     // explicitly emit pop instructions for them.  Do this now after the asm has
1725     // executed so that the ST(x) numbers are not off (which would happen if we
1726     // did this inline with operand rewriting).
1727     //
1728     // Note: this might be a non-optimal pop sequence.  We might be able to do
1729     // better by trying to pop in stack order or something.
1730     while (FPKills) {
1731       unsigned FPReg = llvm::countr_zero(FPKills);
1732       if (isLive(FPReg))
1733         freeStackSlotAfter(Inst, FPReg);
1734       FPKills &= ~(1U << FPReg);
1735     }
1736 
1737     // Don't delete the inline asm!
1738     return;
1739   }
1740   }
1741 
1742   Inst = MBB->erase(Inst);  // Remove the pseudo instruction
1743 
1744   // We want to leave I pointing to the previous instruction, but what if we
1745   // just erased the first instruction?
1746   if (Inst == MBB->begin()) {
1747     LLVM_DEBUG(dbgs() << "Inserting dummy KILL\n");
1748     Inst = BuildMI(*MBB, Inst, DebugLoc(), TII->get(TargetOpcode::KILL));
1749   } else
1750     --Inst;
1751 }
1752 
1753 void FPS::setKillFlags(MachineBasicBlock &MBB) const {
1754   const TargetRegisterInfo &TRI =
1755       *MBB.getParent()->getSubtarget().getRegisterInfo();
1756   LiveRegUnits LPR(TRI);
1757 
1758   LPR.addLiveOuts(MBB);
1759 
1760   for (MachineInstr &MI : llvm::reverse(MBB)) {
1761     if (MI.isDebugInstr())
1762       continue;
1763 
1764     std::bitset<8> Defs;
1765     SmallVector<MachineOperand *, 2> Uses;
1766 
1767     for (auto &MO : MI.operands()) {
1768       if (!MO.isReg())
1769         continue;
1770 
1771       unsigned Reg = MO.getReg() - X86::FP0;
1772 
1773       if (Reg >= 8)
1774         continue;
1775 
1776       if (MO.isDef()) {
1777         Defs.set(Reg);
1778         if (LPR.available(MO.getReg()))
1779           MO.setIsDead();
1780       } else
1781         Uses.push_back(&MO);
1782     }
1783 
1784     for (auto *MO : Uses)
1785       if (Defs.test(getFPReg(*MO)) || LPR.available(MO->getReg()))
1786         MO->setIsKill();
1787 
1788     LPR.stepBackward(MI);
1789   }
1790 }
1791