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/SmallPtrSet.h"
30 #include "llvm/ADT/SmallSet.h"
31 #include "llvm/ADT/SmallVector.h"
32 #include "llvm/ADT/Statistic.h"
33 #include "llvm/CodeGen/EdgeBundles.h"
34 #include "llvm/CodeGen/LivePhysRegs.h"
35 #include "llvm/CodeGen/MachineFunctionPass.h"
36 #include "llvm/CodeGen/MachineInstrBuilder.h"
37 #include "llvm/CodeGen/MachineRegisterInfo.h"
38 #include "llvm/CodeGen/Passes.h"
39 #include "llvm/CodeGen/TargetInstrInfo.h"
40 #include "llvm/CodeGen/TargetSubtargetInfo.h"
41 #include "llvm/Config/llvm-config.h"
42 #include "llvm/IR/InlineAsm.h"
43 #include "llvm/InitializePasses.h"
44 #include "llvm/Support/Debug.h"
45 #include "llvm/Support/ErrorHandling.h"
46 #include "llvm/Support/raw_ostream.h"
47 #include "llvm/Target/TargetMachine.h"
48 #include <algorithm>
49 #include <bitset>
50 using namespace llvm;
51 
52 #define DEBUG_TYPE "x86-codegen"
53 
54 STATISTIC(NumFXCH, "Number of fxch instructions inserted");
55 STATISTIC(NumFP  , "Number of floating point instructions");
56 
57 namespace {
58   const unsigned ScratchFPReg = 7;
59 
60   struct FPS : public MachineFunctionPass {
61     static char ID;
FPS__anon8063afeb0111::FPS62     FPS() : MachineFunctionPass(ID) {
63       // This is really only to keep valgrind quiet.
64       // The logic in isLive() is too much for it.
65       memset(Stack, 0, sizeof(Stack));
66       memset(RegMap, 0, sizeof(RegMap));
67     }
68 
getAnalysisUsage__anon8063afeb0111::FPS69     void getAnalysisUsage(AnalysisUsage &AU) const override {
70       AU.setPreservesCFG();
71       AU.addRequired<EdgeBundles>();
72       AU.addPreservedID(MachineLoopInfoID);
73       AU.addPreservedID(MachineDominatorsID);
74       MachineFunctionPass::getAnalysisUsage(AU);
75     }
76 
77     bool runOnMachineFunction(MachineFunction &MF) override;
78 
getRequiredProperties__anon8063afeb0111::FPS79     MachineFunctionProperties getRequiredProperties() const override {
80       return MachineFunctionProperties().set(
81           MachineFunctionProperties::Property::NoVRegs);
82     }
83 
getPassName__anon8063afeb0111::FPS84     StringRef getPassName() const override { return "X86 FP Stackifier"; }
85 
86   private:
87     const TargetInstrInfo *TII = nullptr; // Machine instruction info.
88 
89     // Two CFG edges are related if they leave the same block, or enter the same
90     // block. The transitive closure of an edge under this relation is a
91     // LiveBundle. It represents a set of CFG edges where the live FP stack
92     // registers must be allocated identically in the x87 stack.
93     //
94     // A LiveBundle is usually all the edges leaving a block, or all the edges
95     // entering a block, but it can contain more edges if critical edges are
96     // present.
97     //
98     // The set of live FP registers in a LiveBundle is calculated by bundleCFG,
99     // but the exact mapping of FP registers to stack slots is fixed later.
100     struct LiveBundle {
101       // Bit mask of live FP registers. Bit 0 = FP0, bit 1 = FP1, &c.
102       unsigned Mask;
103 
104       // Number of pre-assigned live registers in FixStack. This is 0 when the
105       // stack order has not yet been fixed.
106       unsigned FixCount;
107 
108       // Assigned stack order for live-in registers.
109       // FixStack[i] == getStackEntry(i) for all i < FixCount.
110       unsigned char FixStack[8];
111 
LiveBundle__anon8063afeb0111::FPS::LiveBundle112       LiveBundle() : Mask(0), FixCount(0) {}
113 
114       // Have the live registers been assigned a stack order yet?
isFixed__anon8063afeb0111::FPS::LiveBundle115       bool isFixed() const { return !Mask || FixCount; }
116     };
117 
118     // Numbered LiveBundle structs. LiveBundles[0] is used for all CFG edges
119     // with no live FP registers.
120     SmallVector<LiveBundle, 8> LiveBundles;
121 
122     // The edge bundle analysis provides indices into the LiveBundles vector.
123     EdgeBundles *Bundles = nullptr;
124 
125     // Return a bitmask of FP registers in block's live-in list.
calcLiveInMask__anon8063afeb0111::FPS126     static unsigned calcLiveInMask(MachineBasicBlock *MBB, bool RemoveFPs) {
127       unsigned Mask = 0;
128       for (MachineBasicBlock::livein_iterator I = MBB->livein_begin();
129            I != MBB->livein_end(); ) {
130         MCPhysReg Reg = I->PhysReg;
131         static_assert(X86::FP6 - X86::FP0 == 6, "sequential regnums");
132         if (Reg >= X86::FP0 && Reg <= X86::FP6) {
133           Mask |= 1 << (Reg - X86::FP0);
134           if (RemoveFPs) {
135             I = MBB->removeLiveIn(I);
136             continue;
137           }
138         }
139         ++I;
140       }
141       return Mask;
142     }
143 
144     // Partition all the CFG edges into LiveBundles.
145     void bundleCFGRecomputeKillFlags(MachineFunction &MF);
146 
147     MachineBasicBlock *MBB = nullptr;     // Current basic block
148 
149     // The hardware keeps track of how many FP registers are live, so we have
150     // to model that exactly. Usually, each live register corresponds to an
151     // FP<n> register, but when dealing with calls, returns, and inline
152     // assembly, it is sometimes necessary to have live scratch registers.
153     unsigned Stack[8];          // FP<n> Registers in each stack slot...
154     unsigned StackTop = 0;      // The current top of the FP stack.
155 
156     enum {
157       NumFPRegs = 8             // Including scratch pseudo-registers.
158     };
159 
160     // For each live FP<n> register, point to its Stack[] entry.
161     // The first entries correspond to FP0-FP6, the rest are scratch registers
162     // used when we need slightly different live registers than what the
163     // register allocator thinks.
164     unsigned RegMap[NumFPRegs];
165 
166     // Set up our stack model to match the incoming registers to MBB.
167     void setupBlockStack();
168 
169     // Shuffle live registers to match the expectations of successor blocks.
170     void finishBlockStack();
171 
172 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
dumpStack__anon8063afeb0111::FPS173     void dumpStack() const {
174       dbgs() << "Stack contents:";
175       for (unsigned i = 0; i != StackTop; ++i) {
176         dbgs() << " FP" << Stack[i];
177         assert(RegMap[Stack[i]] == i && "Stack[] doesn't match RegMap[]!");
178       }
179     }
180 #endif
181 
182     /// getSlot - Return the stack slot number a particular register number is
183     /// in.
getSlot__anon8063afeb0111::FPS184     unsigned getSlot(unsigned RegNo) const {
185       assert(RegNo < NumFPRegs && "Regno out of range!");
186       return RegMap[RegNo];
187     }
188 
189     /// isLive - Is RegNo currently live in the stack?
isLive__anon8063afeb0111::FPS190     bool isLive(unsigned RegNo) const {
191       unsigned Slot = getSlot(RegNo);
192       return Slot < StackTop && Stack[Slot] == RegNo;
193     }
194 
195     /// getStackEntry - Return the X86::FP<n> register in register ST(i).
getStackEntry__anon8063afeb0111::FPS196     unsigned getStackEntry(unsigned STi) const {
197       if (STi >= StackTop)
198         report_fatal_error("Access past stack top!");
199       return Stack[StackTop-1-STi];
200     }
201 
202     /// getSTReg - Return the X86::ST(i) register which contains the specified
203     /// FP<RegNo> register.
getSTReg__anon8063afeb0111::FPS204     unsigned getSTReg(unsigned RegNo) const {
205       return StackTop - 1 - getSlot(RegNo) + X86::ST0;
206     }
207 
208     // pushReg - Push the specified FP<n> register onto the stack.
pushReg__anon8063afeb0111::FPS209     void pushReg(unsigned Reg) {
210       assert(Reg < NumFPRegs && "Register number out of range!");
211       if (StackTop >= 8)
212         report_fatal_error("Stack overflow!");
213       Stack[StackTop] = Reg;
214       RegMap[Reg] = StackTop++;
215     }
216 
217     // popReg - Pop a register from the stack.
popReg__anon8063afeb0111::FPS218     void popReg() {
219       if (StackTop == 0)
220         report_fatal_error("Cannot pop empty stack!");
221       RegMap[Stack[--StackTop]] = ~0;     // Update state
222     }
223 
isAtTop__anon8063afeb0111::FPS224     bool isAtTop(unsigned RegNo) const { return getSlot(RegNo) == StackTop-1; }
moveToTop__anon8063afeb0111::FPS225     void moveToTop(unsigned RegNo, MachineBasicBlock::iterator I) {
226       DebugLoc dl = I == MBB->end() ? DebugLoc() : I->getDebugLoc();
227       if (isAtTop(RegNo)) return;
228 
229       unsigned STReg = getSTReg(RegNo);
230       unsigned RegOnTop = getStackEntry(0);
231 
232       // Swap the slots the regs are in.
233       std::swap(RegMap[RegNo], RegMap[RegOnTop]);
234 
235       // Swap stack slot contents.
236       if (RegMap[RegOnTop] >= StackTop)
237         report_fatal_error("Access past stack top!");
238       std::swap(Stack[RegMap[RegOnTop]], Stack[StackTop-1]);
239 
240       // Emit an fxch to update the runtime processors version of the state.
241       BuildMI(*MBB, I, dl, TII->get(X86::XCH_F)).addReg(STReg);
242       ++NumFXCH;
243     }
244 
duplicateToTop__anon8063afeb0111::FPS245     void duplicateToTop(unsigned RegNo, unsigned AsReg,
246                         MachineBasicBlock::iterator I) {
247       DebugLoc dl = I == MBB->end() ? DebugLoc() : I->getDebugLoc();
248       unsigned STReg = getSTReg(RegNo);
249       pushReg(AsReg);   // New register on top of stack
250 
251       BuildMI(*MBB, I, dl, TII->get(X86::LD_Frr)).addReg(STReg);
252     }
253 
254     /// popStackAfter - Pop the current value off of the top of the FP stack
255     /// after the specified instruction.
256     void popStackAfter(MachineBasicBlock::iterator &I);
257 
258     /// freeStackSlotAfter - Free the specified register from the register
259     /// stack, so that it is no longer in a register.  If the register is
260     /// currently at the top of the stack, we just pop the current instruction,
261     /// otherwise we store the current top-of-stack into the specified slot,
262     /// then pop the top of stack.
263     void freeStackSlotAfter(MachineBasicBlock::iterator &I, unsigned Reg);
264 
265     /// freeStackSlotBefore - Just the pop, no folding. Return the inserted
266     /// instruction.
267     MachineBasicBlock::iterator
268     freeStackSlotBefore(MachineBasicBlock::iterator I, unsigned FPRegNo);
269 
270     /// Adjust the live registers to be the set in Mask.
271     void adjustLiveRegs(unsigned Mask, MachineBasicBlock::iterator I);
272 
273     /// Shuffle the top FixCount stack entries such that FP reg FixStack[0] is
274     /// st(0), FP reg FixStack[1] is st(1) etc.
275     void shuffleStackTop(const unsigned char *FixStack, unsigned FixCount,
276                          MachineBasicBlock::iterator I);
277 
278     bool processBasicBlock(MachineFunction &MF, MachineBasicBlock &MBB);
279 
280     void handleCall(MachineBasicBlock::iterator &I);
281     void handleReturn(MachineBasicBlock::iterator &I);
282     void handleZeroArgFP(MachineBasicBlock::iterator &I);
283     void handleOneArgFP(MachineBasicBlock::iterator &I);
284     void handleOneArgFPRW(MachineBasicBlock::iterator &I);
285     void handleTwoArgFP(MachineBasicBlock::iterator &I);
286     void handleCompareFP(MachineBasicBlock::iterator &I);
287     void handleCondMovFP(MachineBasicBlock::iterator &I);
288     void handleSpecialFP(MachineBasicBlock::iterator &I);
289 
290     // Check if a COPY instruction is using FP registers.
isFPCopy__anon8063afeb0111::FPS291     static bool isFPCopy(MachineInstr &MI) {
292       Register DstReg = MI.getOperand(0).getReg();
293       Register SrcReg = MI.getOperand(1).getReg();
294 
295       return X86::RFP80RegClass.contains(DstReg) ||
296         X86::RFP80RegClass.contains(SrcReg);
297     }
298 
299     void setKillFlags(MachineBasicBlock &MBB) const;
300   };
301 }
302 
303 char FPS::ID = 0;
304 
305 INITIALIZE_PASS_BEGIN(FPS, DEBUG_TYPE, "X86 FP Stackifier",
306                       false, false)
INITIALIZE_PASS_DEPENDENCY(EdgeBundles)307 INITIALIZE_PASS_DEPENDENCY(EdgeBundles)
308 INITIALIZE_PASS_END(FPS, DEBUG_TYPE, "X86 FP Stackifier",
309                     false, false)
310 
311 FunctionPass *llvm::createX86FloatingPointStackifierPass() { return new FPS(); }
312 
313 /// getFPReg - Return the X86::FPx register number for the specified operand.
314 /// For example, this returns 3 for X86::FP3.
getFPReg(const MachineOperand & MO)315 static unsigned getFPReg(const MachineOperand &MO) {
316   assert(MO.isReg() && "Expected an FP register!");
317   Register Reg = MO.getReg();
318   assert(Reg >= X86::FP0 && Reg <= X86::FP6 && "Expected FP register!");
319   return Reg - X86::FP0;
320 }
321 
322 /// runOnMachineFunction - Loop over all of the basic blocks, transforming FP
323 /// register references into FP stack references.
324 ///
runOnMachineFunction(MachineFunction & MF)325 bool FPS::runOnMachineFunction(MachineFunction &MF) {
326   // We only need to run this pass if there are any FP registers used in this
327   // function.  If it is all integer, there is nothing for us to do!
328   bool FPIsUsed = false;
329 
330   static_assert(X86::FP6 == X86::FP0+6, "Register enums aren't sorted right!");
331   const MachineRegisterInfo &MRI = MF.getRegInfo();
332   for (unsigned i = 0; i <= 6; ++i)
333     if (!MRI.reg_nodbg_empty(X86::FP0 + i)) {
334       FPIsUsed = true;
335       break;
336     }
337 
338   // Early exit.
339   if (!FPIsUsed) return false;
340 
341   Bundles = &getAnalysis<EdgeBundles>();
342   TII = MF.getSubtarget().getInstrInfo();
343 
344   // Prepare cross-MBB liveness.
345   bundleCFGRecomputeKillFlags(MF);
346 
347   StackTop = 0;
348 
349   // Process the function in depth first order so that we process at least one
350   // of the predecessors for every reachable block in the function.
351   df_iterator_default_set<MachineBasicBlock*> Processed;
352   MachineBasicBlock *Entry = &MF.front();
353 
354   LiveBundle &Bundle =
355     LiveBundles[Bundles->getBundle(Entry->getNumber(), false)];
356 
357   // In regcall convention, some FP registers may not be passed through
358   // the stack, so they will need to be assigned to the stack first
359   if ((Entry->getParent()->getFunction().getCallingConv() ==
360     CallingConv::X86_RegCall) && (Bundle.Mask && !Bundle.FixCount)) {
361     // In the register calling convention, up to one FP argument could be
362     // saved in the first FP register.
363     // If bundle.mask is non-zero and Bundle.FixCount is zero, it means
364     // that the FP registers contain arguments.
365     // The actual value is passed in FP0.
366     // Here we fix the stack and mark FP0 as pre-assigned register.
367     assert((Bundle.Mask & 0xFE) == 0 &&
368       "Only FP0 could be passed as an argument");
369     Bundle.FixCount = 1;
370     Bundle.FixStack[0] = 0;
371   }
372 
373   bool Changed = false;
374   for (MachineBasicBlock *BB : depth_first_ext(Entry, Processed))
375     Changed |= processBasicBlock(MF, *BB);
376 
377   // Process any unreachable blocks in arbitrary order now.
378   if (MF.size() != Processed.size())
379     for (MachineBasicBlock &BB : MF)
380       if (Processed.insert(&BB).second)
381         Changed |= processBasicBlock(MF, BB);
382 
383   LiveBundles.clear();
384 
385   return Changed;
386 }
387 
388 /// bundleCFG - Scan all the basic blocks to determine consistent live-in and
389 /// live-out sets for the FP registers. Consistent means that the set of
390 /// registers live-out from a block is identical to the live-in set of all
391 /// successors. This is not enforced by the normal live-in lists since
392 /// registers may be implicitly defined, or not used by all successors.
bundleCFGRecomputeKillFlags(MachineFunction & MF)393 void FPS::bundleCFGRecomputeKillFlags(MachineFunction &MF) {
394   assert(LiveBundles.empty() && "Stale data in LiveBundles");
395   LiveBundles.resize(Bundles->getNumBundles());
396 
397   // Gather the actual live-in masks for all MBBs.
398   for (MachineBasicBlock &MBB : MF) {
399     setKillFlags(MBB);
400 
401     const unsigned Mask = calcLiveInMask(&MBB, false);
402     if (!Mask)
403       continue;
404     // Update MBB ingoing bundle mask.
405     LiveBundles[Bundles->getBundle(MBB.getNumber(), false)].Mask |= Mask;
406   }
407 }
408 
409 /// processBasicBlock - Loop over all of the instructions in the basic block,
410 /// transforming FP instructions into their stack form.
411 ///
processBasicBlock(MachineFunction & MF,MachineBasicBlock & BB)412 bool FPS::processBasicBlock(MachineFunction &MF, MachineBasicBlock &BB) {
413   bool Changed = false;
414   MBB = &BB;
415 
416   setupBlockStack();
417 
418   for (MachineBasicBlock::iterator I = BB.begin(); I != BB.end(); ++I) {
419     MachineInstr &MI = *I;
420     uint64_t Flags = MI.getDesc().TSFlags;
421 
422     unsigned FPInstClass = Flags & X86II::FPTypeMask;
423     if (MI.isInlineAsm())
424       FPInstClass = X86II::SpecialFP;
425 
426     if (MI.isCopy() && isFPCopy(MI))
427       FPInstClass = X86II::SpecialFP;
428 
429     if (MI.isImplicitDef() &&
430         X86::RFP80RegClass.contains(MI.getOperand(0).getReg()))
431       FPInstClass = X86II::SpecialFP;
432 
433     if (MI.isCall())
434       FPInstClass = X86II::SpecialFP;
435 
436     if (FPInstClass == X86II::NotFP)
437       continue;  // Efficiently ignore non-fp insts!
438 
439     MachineInstr *PrevMI = nullptr;
440     if (I != BB.begin())
441       PrevMI = &*std::prev(I);
442 
443     ++NumFP;  // Keep track of # of pseudo instrs
444     LLVM_DEBUG(dbgs() << "\nFPInst:\t" << MI);
445 
446     // Get dead variables list now because the MI pointer may be deleted as part
447     // of processing!
448     SmallVector<unsigned, 8> DeadRegs;
449     for (unsigned i = 0, e = MI.getNumOperands(); i != e; ++i) {
450       const MachineOperand &MO = MI.getOperand(i);
451       if (MO.isReg() && MO.isDead())
452         DeadRegs.push_back(MO.getReg());
453     }
454 
455     switch (FPInstClass) {
456     case X86II::ZeroArgFP:  handleZeroArgFP(I); break;
457     case X86II::OneArgFP:   handleOneArgFP(I);  break;  // fstp ST(0)
458     case X86II::OneArgFPRW: handleOneArgFPRW(I); break; // ST(0) = fsqrt(ST(0))
459     case X86II::TwoArgFP:   handleTwoArgFP(I);  break;
460     case X86II::CompareFP:  handleCompareFP(I); break;
461     case X86II::CondMovFP:  handleCondMovFP(I); break;
462     case X86II::SpecialFP:  handleSpecialFP(I); break;
463     default: llvm_unreachable("Unknown FP Type!");
464     }
465 
466     // Check to see if any of the values defined by this instruction are dead
467     // after definition.  If so, pop them.
468     for (unsigned i = 0, e = DeadRegs.size(); i != e; ++i) {
469       unsigned Reg = DeadRegs[i];
470       // Check if Reg is live on the stack. An inline-asm register operand that
471       // is in the clobber list and marked dead might not be live on the stack.
472       static_assert(X86::FP7 - X86::FP0 == 7, "sequential FP regnumbers");
473       if (Reg >= X86::FP0 && Reg <= X86::FP6 && isLive(Reg-X86::FP0)) {
474         LLVM_DEBUG(dbgs() << "Register FP#" << Reg - X86::FP0 << " is dead!\n");
475         freeStackSlotAfter(I, Reg-X86::FP0);
476       }
477     }
478 
479     // Print out all of the instructions expanded to if -debug
480     LLVM_DEBUG({
481       MachineBasicBlock::iterator PrevI = PrevMI;
482       if (I == PrevI) {
483         dbgs() << "Just deleted pseudo instruction\n";
484       } else {
485         MachineBasicBlock::iterator Start = I;
486         // Rewind to first instruction newly inserted.
487         while (Start != BB.begin() && std::prev(Start) != PrevI)
488           --Start;
489         dbgs() << "Inserted instructions:\n\t";
490         Start->print(dbgs());
491         while (++Start != std::next(I)) {
492         }
493       }
494       dumpStack();
495     });
496     (void)PrevMI;
497 
498     Changed = true;
499   }
500 
501   finishBlockStack();
502 
503   return Changed;
504 }
505 
506 /// setupBlockStack - Use the live bundles to set up our model of the stack
507 /// to match predecessors' live out stack.
setupBlockStack()508 void FPS::setupBlockStack() {
509   LLVM_DEBUG(dbgs() << "\nSetting up live-ins for " << printMBBReference(*MBB)
510                     << " derived from " << MBB->getName() << ".\n");
511   StackTop = 0;
512   // Get the live-in bundle for MBB.
513   const LiveBundle &Bundle =
514     LiveBundles[Bundles->getBundle(MBB->getNumber(), false)];
515 
516   if (!Bundle.Mask) {
517     LLVM_DEBUG(dbgs() << "Block has no FP live-ins.\n");
518     return;
519   }
520 
521   // Depth-first iteration should ensure that we always have an assigned stack.
522   assert(Bundle.isFixed() && "Reached block before any predecessors");
523 
524   // Push the fixed live-in registers.
525   for (unsigned i = Bundle.FixCount; i > 0; --i) {
526     LLVM_DEBUG(dbgs() << "Live-in st(" << (i - 1) << "): %fp"
527                       << unsigned(Bundle.FixStack[i - 1]) << '\n');
528     pushReg(Bundle.FixStack[i-1]);
529   }
530 
531   // Kill off unwanted live-ins. This can happen with a critical edge.
532   // FIXME: We could keep these live registers around as zombies. They may need
533   // to be revived at the end of a short block. It might save a few instrs.
534   unsigned Mask = calcLiveInMask(MBB, /*RemoveFPs=*/true);
535   adjustLiveRegs(Mask, MBB->begin());
536   LLVM_DEBUG(MBB->dump());
537 }
538 
539 /// finishBlockStack - Revive live-outs that are implicitly defined out of
540 /// MBB. Shuffle live registers to match the expected fixed stack of any
541 /// predecessors, and ensure that all predecessors are expecting the same
542 /// stack.
finishBlockStack()543 void FPS::finishBlockStack() {
544   // The RET handling below takes care of return blocks for us.
545   if (MBB->succ_empty())
546     return;
547 
548   LLVM_DEBUG(dbgs() << "Setting up live-outs for " << printMBBReference(*MBB)
549                     << " derived from " << MBB->getName() << ".\n");
550 
551   // Get MBB's live-out bundle.
552   unsigned BundleIdx = Bundles->getBundle(MBB->getNumber(), true);
553   LiveBundle &Bundle = LiveBundles[BundleIdx];
554 
555   // We may need to kill and define some registers to match successors.
556   // FIXME: This can probably be combined with the shuffle below.
557   MachineBasicBlock::iterator Term = MBB->getFirstTerminator();
558   adjustLiveRegs(Bundle.Mask, Term);
559 
560   if (!Bundle.Mask) {
561     LLVM_DEBUG(dbgs() << "No live-outs.\n");
562     return;
563   }
564 
565   // Has the stack order been fixed yet?
566   LLVM_DEBUG(dbgs() << "LB#" << BundleIdx << ": ");
567   if (Bundle.isFixed()) {
568     LLVM_DEBUG(dbgs() << "Shuffling stack to match.\n");
569     shuffleStackTop(Bundle.FixStack, Bundle.FixCount, Term);
570   } else {
571     // Not fixed yet, we get to choose.
572     LLVM_DEBUG(dbgs() << "Fixing stack order now.\n");
573     Bundle.FixCount = StackTop;
574     for (unsigned i = 0; i < StackTop; ++i)
575       Bundle.FixStack[i] = getStackEntry(i);
576   }
577 }
578 
579 
580 //===----------------------------------------------------------------------===//
581 // Efficient Lookup Table Support
582 //===----------------------------------------------------------------------===//
583 
584 namespace {
585   struct TableEntry {
586     uint16_t from;
587     uint16_t to;
operator <__anon8063afeb0311::TableEntry588     bool operator<(const TableEntry &TE) const { return from < TE.from; }
operator <(const TableEntry & TE,unsigned V)589     friend bool operator<(const TableEntry &TE, unsigned V) {
590       return TE.from < V;
591     }
operator <(unsigned V,const TableEntry & TE)592     friend bool LLVM_ATTRIBUTE_UNUSED operator<(unsigned V,
593                                                 const TableEntry &TE) {
594       return V < TE.from;
595     }
596   };
597 }
598 
Lookup(ArrayRef<TableEntry> Table,unsigned Opcode)599 static int Lookup(ArrayRef<TableEntry> Table, unsigned Opcode) {
600   const TableEntry *I = llvm::lower_bound(Table, Opcode);
601   if (I != Table.end() && I->from == Opcode)
602     return I->to;
603   return -1;
604 }
605 
606 #ifdef NDEBUG
607 #define ASSERT_SORTED(TABLE)
608 #else
609 #define ASSERT_SORTED(TABLE)                                                   \
610   {                                                                            \
611     static std::atomic<bool> TABLE##Checked(false);                            \
612     if (!TABLE##Checked.load(std::memory_order_relaxed)) {                     \
613       assert(std::is_sorted(std::begin(TABLE), std::end(TABLE)) &&             \
614              "All lookup tables must be sorted for efficient access!");        \
615       TABLE##Checked.store(true, std::memory_order_relaxed);                   \
616     }                                                                          \
617   }
618 #endif
619 
620 //===----------------------------------------------------------------------===//
621 // Register File -> Register Stack Mapping Methods
622 //===----------------------------------------------------------------------===//
623 
624 // OpcodeTable - Sorted map of register instructions to their stack version.
625 // The first element is an register file pseudo instruction, the second is the
626 // concrete X86 instruction which uses the register stack.
627 //
628 static const TableEntry OpcodeTable[] = {
629   { X86::ABS_Fp32     , X86::ABS_F     },
630   { X86::ABS_Fp64     , X86::ABS_F     },
631   { X86::ABS_Fp80     , X86::ABS_F     },
632   { X86::ADD_Fp32m    , X86::ADD_F32m  },
633   { X86::ADD_Fp64m    , X86::ADD_F64m  },
634   { X86::ADD_Fp64m32  , X86::ADD_F32m  },
635   { X86::ADD_Fp80m32  , X86::ADD_F32m  },
636   { X86::ADD_Fp80m64  , X86::ADD_F64m  },
637   { X86::ADD_FpI16m32 , X86::ADD_FI16m },
638   { X86::ADD_FpI16m64 , X86::ADD_FI16m },
639   { X86::ADD_FpI16m80 , X86::ADD_FI16m },
640   { X86::ADD_FpI32m32 , X86::ADD_FI32m },
641   { X86::ADD_FpI32m64 , X86::ADD_FI32m },
642   { X86::ADD_FpI32m80 , X86::ADD_FI32m },
643   { X86::CHS_Fp32     , X86::CHS_F     },
644   { X86::CHS_Fp64     , X86::CHS_F     },
645   { X86::CHS_Fp80     , X86::CHS_F     },
646   { X86::CMOVBE_Fp32  , X86::CMOVBE_F  },
647   { X86::CMOVBE_Fp64  , X86::CMOVBE_F  },
648   { X86::CMOVBE_Fp80  , X86::CMOVBE_F  },
649   { X86::CMOVB_Fp32   , X86::CMOVB_F   },
650   { X86::CMOVB_Fp64   , X86::CMOVB_F  },
651   { X86::CMOVB_Fp80   , X86::CMOVB_F  },
652   { X86::CMOVE_Fp32   , X86::CMOVE_F  },
653   { X86::CMOVE_Fp64   , X86::CMOVE_F   },
654   { X86::CMOVE_Fp80   , X86::CMOVE_F   },
655   { X86::CMOVNBE_Fp32 , X86::CMOVNBE_F },
656   { X86::CMOVNBE_Fp64 , X86::CMOVNBE_F },
657   { X86::CMOVNBE_Fp80 , X86::CMOVNBE_F },
658   { X86::CMOVNB_Fp32  , X86::CMOVNB_F  },
659   { X86::CMOVNB_Fp64  , X86::CMOVNB_F  },
660   { X86::CMOVNB_Fp80  , X86::CMOVNB_F  },
661   { X86::CMOVNE_Fp32  , X86::CMOVNE_F  },
662   { X86::CMOVNE_Fp64  , X86::CMOVNE_F  },
663   { X86::CMOVNE_Fp80  , X86::CMOVNE_F  },
664   { X86::CMOVNP_Fp32  , X86::CMOVNP_F  },
665   { X86::CMOVNP_Fp64  , X86::CMOVNP_F  },
666   { X86::CMOVNP_Fp80  , X86::CMOVNP_F  },
667   { X86::CMOVP_Fp32   , X86::CMOVP_F   },
668   { X86::CMOVP_Fp64   , X86::CMOVP_F   },
669   { X86::CMOVP_Fp80   , X86::CMOVP_F   },
670   { X86::COM_FpIr32   , X86::COM_FIr   },
671   { X86::COM_FpIr64   , X86::COM_FIr   },
672   { X86::COM_FpIr80   , X86::COM_FIr   },
673   { X86::COM_Fpr32    , X86::COM_FST0r },
674   { X86::COM_Fpr64    , X86::COM_FST0r },
675   { X86::COM_Fpr80    , X86::COM_FST0r },
676   { X86::DIVR_Fp32m   , X86::DIVR_F32m },
677   { X86::DIVR_Fp64m   , X86::DIVR_F64m },
678   { X86::DIVR_Fp64m32 , X86::DIVR_F32m },
679   { X86::DIVR_Fp80m32 , X86::DIVR_F32m },
680   { X86::DIVR_Fp80m64 , X86::DIVR_F64m },
681   { X86::DIVR_FpI16m32, X86::DIVR_FI16m},
682   { X86::DIVR_FpI16m64, X86::DIVR_FI16m},
683   { X86::DIVR_FpI16m80, X86::DIVR_FI16m},
684   { X86::DIVR_FpI32m32, X86::DIVR_FI32m},
685   { X86::DIVR_FpI32m64, X86::DIVR_FI32m},
686   { X86::DIVR_FpI32m80, X86::DIVR_FI32m},
687   { X86::DIV_Fp32m    , X86::DIV_F32m  },
688   { X86::DIV_Fp64m    , X86::DIV_F64m  },
689   { X86::DIV_Fp64m32  , X86::DIV_F32m  },
690   { X86::DIV_Fp80m32  , X86::DIV_F32m  },
691   { X86::DIV_Fp80m64  , X86::DIV_F64m  },
692   { X86::DIV_FpI16m32 , X86::DIV_FI16m },
693   { X86::DIV_FpI16m64 , X86::DIV_FI16m },
694   { X86::DIV_FpI16m80 , X86::DIV_FI16m },
695   { X86::DIV_FpI32m32 , X86::DIV_FI32m },
696   { X86::DIV_FpI32m64 , X86::DIV_FI32m },
697   { X86::DIV_FpI32m80 , X86::DIV_FI32m },
698   { X86::ILD_Fp16m32  , X86::ILD_F16m  },
699   { X86::ILD_Fp16m64  , X86::ILD_F16m  },
700   { X86::ILD_Fp16m80  , X86::ILD_F16m  },
701   { X86::ILD_Fp32m32  , X86::ILD_F32m  },
702   { X86::ILD_Fp32m64  , X86::ILD_F32m  },
703   { X86::ILD_Fp32m80  , X86::ILD_F32m  },
704   { X86::ILD_Fp64m32  , X86::ILD_F64m  },
705   { X86::ILD_Fp64m64  , X86::ILD_F64m  },
706   { X86::ILD_Fp64m80  , X86::ILD_F64m  },
707   { X86::ISTT_Fp16m32 , X86::ISTT_FP16m},
708   { X86::ISTT_Fp16m64 , X86::ISTT_FP16m},
709   { X86::ISTT_Fp16m80 , X86::ISTT_FP16m},
710   { X86::ISTT_Fp32m32 , X86::ISTT_FP32m},
711   { X86::ISTT_Fp32m64 , X86::ISTT_FP32m},
712   { X86::ISTT_Fp32m80 , X86::ISTT_FP32m},
713   { X86::ISTT_Fp64m32 , X86::ISTT_FP64m},
714   { X86::ISTT_Fp64m64 , X86::ISTT_FP64m},
715   { X86::ISTT_Fp64m80 , X86::ISTT_FP64m},
716   { X86::IST_Fp16m32  , X86::IST_F16m  },
717   { X86::IST_Fp16m64  , X86::IST_F16m  },
718   { X86::IST_Fp16m80  , X86::IST_F16m  },
719   { X86::IST_Fp32m32  , X86::IST_F32m  },
720   { X86::IST_Fp32m64  , X86::IST_F32m  },
721   { X86::IST_Fp32m80  , X86::IST_F32m  },
722   { X86::IST_Fp64m32  , X86::IST_FP64m },
723   { X86::IST_Fp64m64  , X86::IST_FP64m },
724   { X86::IST_Fp64m80  , X86::IST_FP64m },
725   { X86::LD_Fp032     , X86::LD_F0     },
726   { X86::LD_Fp064     , X86::LD_F0     },
727   { X86::LD_Fp080     , X86::LD_F0     },
728   { X86::LD_Fp132     , X86::LD_F1     },
729   { X86::LD_Fp164     , X86::LD_F1     },
730   { X86::LD_Fp180     , X86::LD_F1     },
731   { X86::LD_Fp32m     , X86::LD_F32m   },
732   { X86::LD_Fp32m64   , X86::LD_F32m   },
733   { X86::LD_Fp32m80   , X86::LD_F32m   },
734   { X86::LD_Fp64m     , X86::LD_F64m   },
735   { X86::LD_Fp64m80   , X86::LD_F64m   },
736   { X86::LD_Fp80m     , X86::LD_F80m   },
737   { X86::MUL_Fp32m    , X86::MUL_F32m  },
738   { X86::MUL_Fp64m    , X86::MUL_F64m  },
739   { X86::MUL_Fp64m32  , X86::MUL_F32m  },
740   { X86::MUL_Fp80m32  , X86::MUL_F32m  },
741   { X86::MUL_Fp80m64  , X86::MUL_F64m  },
742   { X86::MUL_FpI16m32 , X86::MUL_FI16m },
743   { X86::MUL_FpI16m64 , X86::MUL_FI16m },
744   { X86::MUL_FpI16m80 , X86::MUL_FI16m },
745   { X86::MUL_FpI32m32 , X86::MUL_FI32m },
746   { X86::MUL_FpI32m64 , X86::MUL_FI32m },
747   { X86::MUL_FpI32m80 , X86::MUL_FI32m },
748   { X86::SQRT_Fp32    , X86::SQRT_F    },
749   { X86::SQRT_Fp64    , X86::SQRT_F    },
750   { X86::SQRT_Fp80    , X86::SQRT_F    },
751   { X86::ST_Fp32m     , X86::ST_F32m   },
752   { X86::ST_Fp64m     , X86::ST_F64m   },
753   { X86::ST_Fp64m32   , X86::ST_F32m   },
754   { X86::ST_Fp80m32   , X86::ST_F32m   },
755   { X86::ST_Fp80m64   , X86::ST_F64m   },
756   { X86::ST_FpP80m    , X86::ST_FP80m  },
757   { X86::SUBR_Fp32m   , X86::SUBR_F32m },
758   { X86::SUBR_Fp64m   , X86::SUBR_F64m },
759   { X86::SUBR_Fp64m32 , X86::SUBR_F32m },
760   { X86::SUBR_Fp80m32 , X86::SUBR_F32m },
761   { X86::SUBR_Fp80m64 , X86::SUBR_F64m },
762   { X86::SUBR_FpI16m32, X86::SUBR_FI16m},
763   { X86::SUBR_FpI16m64, X86::SUBR_FI16m},
764   { X86::SUBR_FpI16m80, X86::SUBR_FI16m},
765   { X86::SUBR_FpI32m32, X86::SUBR_FI32m},
766   { X86::SUBR_FpI32m64, X86::SUBR_FI32m},
767   { X86::SUBR_FpI32m80, X86::SUBR_FI32m},
768   { X86::SUB_Fp32m    , X86::SUB_F32m  },
769   { X86::SUB_Fp64m    , X86::SUB_F64m  },
770   { X86::SUB_Fp64m32  , X86::SUB_F32m  },
771   { X86::SUB_Fp80m32  , X86::SUB_F32m  },
772   { X86::SUB_Fp80m64  , X86::SUB_F64m  },
773   { X86::SUB_FpI16m32 , X86::SUB_FI16m },
774   { X86::SUB_FpI16m64 , X86::SUB_FI16m },
775   { X86::SUB_FpI16m80 , X86::SUB_FI16m },
776   { X86::SUB_FpI32m32 , X86::SUB_FI32m },
777   { X86::SUB_FpI32m64 , X86::SUB_FI32m },
778   { X86::SUB_FpI32m80 , X86::SUB_FI32m },
779   { X86::TST_Fp32     , X86::TST_F     },
780   { X86::TST_Fp64     , X86::TST_F     },
781   { X86::TST_Fp80     , X86::TST_F     },
782   { X86::UCOM_FpIr32  , X86::UCOM_FIr  },
783   { X86::UCOM_FpIr64  , X86::UCOM_FIr  },
784   { X86::UCOM_FpIr80  , X86::UCOM_FIr  },
785   { X86::UCOM_Fpr32   , X86::UCOM_Fr   },
786   { X86::UCOM_Fpr64   , X86::UCOM_Fr   },
787   { X86::UCOM_Fpr80   , X86::UCOM_Fr   },
788 };
789 
getConcreteOpcode(unsigned Opcode)790 static unsigned getConcreteOpcode(unsigned Opcode) {
791   ASSERT_SORTED(OpcodeTable);
792   int Opc = Lookup(OpcodeTable, Opcode);
793   assert(Opc != -1 && "FP Stack instruction not in OpcodeTable!");
794   return Opc;
795 }
796 
797 //===----------------------------------------------------------------------===//
798 // Helper Methods
799 //===----------------------------------------------------------------------===//
800 
801 // PopTable - Sorted map of instructions to their popping version.  The first
802 // element is an instruction, the second is the version which pops.
803 //
804 static const TableEntry PopTable[] = {
805   { X86::ADD_FrST0 , X86::ADD_FPrST0  },
806 
807   { X86::COMP_FST0r, X86::FCOMPP      },
808   { X86::COM_FIr   , X86::COM_FIPr    },
809   { X86::COM_FST0r , X86::COMP_FST0r  },
810 
811   { X86::DIVR_FrST0, X86::DIVR_FPrST0 },
812   { X86::DIV_FrST0 , X86::DIV_FPrST0  },
813 
814   { X86::IST_F16m  , X86::IST_FP16m   },
815   { X86::IST_F32m  , X86::IST_FP32m   },
816 
817   { X86::MUL_FrST0 , X86::MUL_FPrST0  },
818 
819   { X86::ST_F32m   , X86::ST_FP32m    },
820   { X86::ST_F64m   , X86::ST_FP64m    },
821   { X86::ST_Frr    , X86::ST_FPrr     },
822 
823   { X86::SUBR_FrST0, X86::SUBR_FPrST0 },
824   { X86::SUB_FrST0 , X86::SUB_FPrST0  },
825 
826   { X86::UCOM_FIr  , X86::UCOM_FIPr   },
827 
828   { X86::UCOM_FPr  , X86::UCOM_FPPr   },
829   { X86::UCOM_Fr   , X86::UCOM_FPr    },
830 };
831 
832 /// popStackAfter - Pop the current value off of the top of the FP stack after
833 /// the specified instruction.  This attempts to be sneaky and combine the pop
834 /// into the instruction itself if possible.  The iterator is left pointing to
835 /// the last instruction, be it a new pop instruction inserted, or the old
836 /// instruction if it was modified in place.
837 ///
popStackAfter(MachineBasicBlock::iterator & I)838 void FPS::popStackAfter(MachineBasicBlock::iterator &I) {
839   MachineInstr &MI = *I;
840   const DebugLoc &dl = MI.getDebugLoc();
841   ASSERT_SORTED(PopTable);
842 
843   popReg();
844 
845   // Check to see if there is a popping version of this instruction...
846   int Opcode = Lookup(PopTable, I->getOpcode());
847   if (Opcode != -1) {
848     I->setDesc(TII->get(Opcode));
849     if (Opcode == X86::FCOMPP || Opcode == X86::UCOM_FPPr)
850       I->RemoveOperand(0);
851   } else {    // Insert an explicit pop
852     I = BuildMI(*MBB, ++I, dl, TII->get(X86::ST_FPrr)).addReg(X86::ST0);
853   }
854 }
855 
856 /// freeStackSlotAfter - Free the specified register from the register stack, so
857 /// that it is no longer in a register.  If the register is currently at the top
858 /// of the stack, we just pop the current instruction, otherwise we store the
859 /// current top-of-stack into the specified slot, then pop the top of stack.
freeStackSlotAfter(MachineBasicBlock::iterator & I,unsigned FPRegNo)860 void FPS::freeStackSlotAfter(MachineBasicBlock::iterator &I, unsigned FPRegNo) {
861   if (getStackEntry(0) == FPRegNo) {  // already at the top of stack? easy.
862     popStackAfter(I);
863     return;
864   }
865 
866   // Otherwise, store the top of stack into the dead slot, killing the operand
867   // without having to add in an explicit xchg then pop.
868   //
869   I = freeStackSlotBefore(++I, FPRegNo);
870 }
871 
872 /// freeStackSlotBefore - Free the specified register without trying any
873 /// folding.
874 MachineBasicBlock::iterator
freeStackSlotBefore(MachineBasicBlock::iterator I,unsigned FPRegNo)875 FPS::freeStackSlotBefore(MachineBasicBlock::iterator I, unsigned FPRegNo) {
876   unsigned STReg    = getSTReg(FPRegNo);
877   unsigned OldSlot  = getSlot(FPRegNo);
878   unsigned TopReg   = Stack[StackTop-1];
879   Stack[OldSlot]    = TopReg;
880   RegMap[TopReg]    = OldSlot;
881   RegMap[FPRegNo]   = ~0;
882   Stack[--StackTop] = ~0;
883   return BuildMI(*MBB, I, DebugLoc(), TII->get(X86::ST_FPrr))
884       .addReg(STReg)
885       .getInstr();
886 }
887 
888 /// adjustLiveRegs - Kill and revive registers such that exactly the FP
889 /// registers with a bit in Mask are live.
adjustLiveRegs(unsigned Mask,MachineBasicBlock::iterator I)890 void FPS::adjustLiveRegs(unsigned Mask, MachineBasicBlock::iterator I) {
891   unsigned Defs = Mask;
892   unsigned Kills = 0;
893   for (unsigned i = 0; i < StackTop; ++i) {
894     unsigned RegNo = Stack[i];
895     if (!(Defs & (1 << RegNo)))
896       // This register is live, but we don't want it.
897       Kills |= (1 << RegNo);
898     else
899       // We don't need to imp-def this live register.
900       Defs &= ~(1 << RegNo);
901   }
902   assert((Kills & Defs) == 0 && "Register needs killing and def'ing?");
903 
904   // Produce implicit-defs for free by using killed registers.
905   while (Kills && Defs) {
906     unsigned KReg = countTrailingZeros(Kills);
907     unsigned DReg = countTrailingZeros(Defs);
908     LLVM_DEBUG(dbgs() << "Renaming %fp" << KReg << " as imp %fp" << DReg
909                       << "\n");
910     std::swap(Stack[getSlot(KReg)], Stack[getSlot(DReg)]);
911     std::swap(RegMap[KReg], RegMap[DReg]);
912     Kills &= ~(1 << KReg);
913     Defs &= ~(1 << DReg);
914   }
915 
916   // Kill registers by popping.
917   if (Kills && I != MBB->begin()) {
918     MachineBasicBlock::iterator I2 = std::prev(I);
919     while (StackTop) {
920       unsigned KReg = getStackEntry(0);
921       if (!(Kills & (1 << KReg)))
922         break;
923       LLVM_DEBUG(dbgs() << "Popping %fp" << KReg << "\n");
924       popStackAfter(I2);
925       Kills &= ~(1 << KReg);
926     }
927   }
928 
929   // Manually kill the rest.
930   while (Kills) {
931     unsigned KReg = countTrailingZeros(Kills);
932     LLVM_DEBUG(dbgs() << "Killing %fp" << KReg << "\n");
933     freeStackSlotBefore(I, KReg);
934     Kills &= ~(1 << KReg);
935   }
936 
937   // Load zeros for all the imp-defs.
938   while(Defs) {
939     unsigned DReg = countTrailingZeros(Defs);
940     LLVM_DEBUG(dbgs() << "Defining %fp" << DReg << " as 0\n");
941     BuildMI(*MBB, I, DebugLoc(), TII->get(X86::LD_F0));
942     pushReg(DReg);
943     Defs &= ~(1 << DReg);
944   }
945 
946   // Now we should have the correct registers live.
947   LLVM_DEBUG(dumpStack());
948   assert(StackTop == countPopulation(Mask) && "Live count mismatch");
949 }
950 
951 /// shuffleStackTop - emit fxch instructions before I to shuffle the top
952 /// FixCount entries into the order given by FixStack.
953 /// FIXME: Is there a better algorithm than insertion sort?
shuffleStackTop(const unsigned char * FixStack,unsigned FixCount,MachineBasicBlock::iterator I)954 void FPS::shuffleStackTop(const unsigned char *FixStack,
955                           unsigned FixCount,
956                           MachineBasicBlock::iterator I) {
957   // Move items into place, starting from the desired stack bottom.
958   while (FixCount--) {
959     // Old register at position FixCount.
960     unsigned OldReg = getStackEntry(FixCount);
961     // Desired register at position FixCount.
962     unsigned Reg = FixStack[FixCount];
963     if (Reg == OldReg)
964       continue;
965     // (Reg st0) (OldReg st0) = (Reg OldReg st0)
966     moveToTop(Reg, I);
967     if (FixCount > 0)
968       moveToTop(OldReg, I);
969   }
970   LLVM_DEBUG(dumpStack());
971 }
972 
973 
974 //===----------------------------------------------------------------------===//
975 // Instruction transformation implementation
976 //===----------------------------------------------------------------------===//
977 
handleCall(MachineBasicBlock::iterator & I)978 void FPS::handleCall(MachineBasicBlock::iterator &I) {
979   MachineInstr &MI = *I;
980   unsigned STReturns = 0;
981 
982   for (unsigned i = 0, e = MI.getNumOperands(); i != e; ++i) {
983     MachineOperand &Op = MI.getOperand(i);
984     if (!Op.isReg() || Op.getReg() < X86::FP0 || Op.getReg() > X86::FP6)
985       continue;
986 
987     assert(Op.isImplicit() && "Expected implicit def/use");
988 
989     if (Op.isDef())
990       STReturns |= 1 << getFPReg(Op);
991 
992     // Remove the operand so that later passes don't see it.
993     MI.RemoveOperand(i);
994     --i;
995     --e;
996   }
997 
998   unsigned N = countTrailingOnes(STReturns);
999 
1000   // FP registers used for function return must be consecutive starting at
1001   // FP0
1002   assert(STReturns == 0 || (isMask_32(STReturns) && N <= 2));
1003 
1004   // Reset the FP Stack - It is required because of possible leftovers from
1005   // passed arguments. The caller should assume that the FP stack is
1006   // returned empty (unless the callee returns values on FP stack).
1007   while (StackTop > 0)
1008     popReg();
1009 
1010   for (unsigned I = 0; I < N; ++I)
1011     pushReg(N - I - 1);
1012 }
1013 
1014 /// If RET has an FP register use operand, pass the first one in ST(0) and
1015 /// the second one in ST(1).
handleReturn(MachineBasicBlock::iterator & I)1016 void FPS::handleReturn(MachineBasicBlock::iterator &I) {
1017   MachineInstr &MI = *I;
1018 
1019   // Find the register operands.
1020   unsigned FirstFPRegOp = ~0U, SecondFPRegOp = ~0U;
1021   unsigned LiveMask = 0;
1022 
1023   for (unsigned i = 0, e = MI.getNumOperands(); i != e; ++i) {
1024     MachineOperand &Op = MI.getOperand(i);
1025     if (!Op.isReg() || Op.getReg() < X86::FP0 || Op.getReg() > X86::FP6)
1026       continue;
1027     // FP Register uses must be kills unless there are two uses of the same
1028     // register, in which case only one will be a kill.
1029     assert(Op.isUse() &&
1030            (Op.isKill() ||                    // Marked kill.
1031             getFPReg(Op) == FirstFPRegOp ||   // Second instance.
1032             MI.killsRegister(Op.getReg())) && // Later use is marked kill.
1033            "Ret only defs operands, and values aren't live beyond it");
1034 
1035     if (FirstFPRegOp == ~0U)
1036       FirstFPRegOp = getFPReg(Op);
1037     else {
1038       assert(SecondFPRegOp == ~0U && "More than two fp operands!");
1039       SecondFPRegOp = getFPReg(Op);
1040     }
1041     LiveMask |= (1 << getFPReg(Op));
1042 
1043     // Remove the operand so that later passes don't see it.
1044     MI.RemoveOperand(i);
1045     --i;
1046     --e;
1047   }
1048 
1049   // We may have been carrying spurious live-ins, so make sure only the
1050   // returned registers are left live.
1051   adjustLiveRegs(LiveMask, MI);
1052   if (!LiveMask) return;  // Quick check to see if any are possible.
1053 
1054   // There are only four possibilities here:
1055   // 1) we are returning a single FP value.  In this case, it has to be in
1056   //    ST(0) already, so just declare success by removing the value from the
1057   //    FP Stack.
1058   if (SecondFPRegOp == ~0U) {
1059     // Assert that the top of stack contains the right FP register.
1060     assert(StackTop == 1 && FirstFPRegOp == getStackEntry(0) &&
1061            "Top of stack not the right register for RET!");
1062 
1063     // Ok, everything is good, mark the value as not being on the stack
1064     // anymore so that our assertion about the stack being empty at end of
1065     // block doesn't fire.
1066     StackTop = 0;
1067     return;
1068   }
1069 
1070   // Otherwise, we are returning two values:
1071   // 2) If returning the same value for both, we only have one thing in the FP
1072   //    stack.  Consider:  RET FP1, FP1
1073   if (StackTop == 1) {
1074     assert(FirstFPRegOp == SecondFPRegOp && FirstFPRegOp == getStackEntry(0)&&
1075            "Stack misconfiguration for RET!");
1076 
1077     // Duplicate the TOS so that we return it twice.  Just pick some other FPx
1078     // register to hold it.
1079     unsigned NewReg = ScratchFPReg;
1080     duplicateToTop(FirstFPRegOp, NewReg, MI);
1081     FirstFPRegOp = NewReg;
1082   }
1083 
1084   /// Okay we know we have two different FPx operands now:
1085   assert(StackTop == 2 && "Must have two values live!");
1086 
1087   /// 3) If SecondFPRegOp is currently in ST(0) and FirstFPRegOp is currently
1088   ///    in ST(1).  In this case, emit an fxch.
1089   if (getStackEntry(0) == SecondFPRegOp) {
1090     assert(getStackEntry(1) == FirstFPRegOp && "Unknown regs live");
1091     moveToTop(FirstFPRegOp, MI);
1092   }
1093 
1094   /// 4) Finally, FirstFPRegOp must be in ST(0) and SecondFPRegOp must be in
1095   /// ST(1).  Just remove both from our understanding of the stack and return.
1096   assert(getStackEntry(0) == FirstFPRegOp && "Unknown regs live");
1097   assert(getStackEntry(1) == SecondFPRegOp && "Unknown regs live");
1098   StackTop = 0;
1099 }
1100 
1101 /// handleZeroArgFP - ST(0) = fld0    ST(0) = flds <mem>
1102 ///
handleZeroArgFP(MachineBasicBlock::iterator & I)1103 void FPS::handleZeroArgFP(MachineBasicBlock::iterator &I) {
1104   MachineInstr &MI = *I;
1105   unsigned DestReg = getFPReg(MI.getOperand(0));
1106 
1107   // Change from the pseudo instruction to the concrete instruction.
1108   MI.RemoveOperand(0); // Remove the explicit ST(0) operand
1109   MI.setDesc(TII->get(getConcreteOpcode(MI.getOpcode())));
1110   MI.addOperand(
1111       MachineOperand::CreateReg(X86::ST0, /*isDef*/ true, /*isImp*/ true));
1112 
1113   // Result gets pushed on the stack.
1114   pushReg(DestReg);
1115 }
1116 
1117 /// handleOneArgFP - fst <mem>, ST(0)
1118 ///
handleOneArgFP(MachineBasicBlock::iterator & I)1119 void FPS::handleOneArgFP(MachineBasicBlock::iterator &I) {
1120   MachineInstr &MI = *I;
1121   unsigned NumOps = MI.getDesc().getNumOperands();
1122   assert((NumOps == X86::AddrNumOperands + 1 || NumOps == 1) &&
1123          "Can only handle fst* & ftst instructions!");
1124 
1125   // Is this the last use of the source register?
1126   unsigned Reg = getFPReg(MI.getOperand(NumOps - 1));
1127   bool KillsSrc = MI.killsRegister(X86::FP0 + Reg);
1128 
1129   // FISTP64m is strange because there isn't a non-popping versions.
1130   // If we have one _and_ we don't want to pop the operand, duplicate the value
1131   // on the stack instead of moving it.  This ensure that popping the value is
1132   // always ok.
1133   // Ditto FISTTP16m, FISTTP32m, FISTTP64m, ST_FpP80m.
1134   //
1135   if (!KillsSrc && (MI.getOpcode() == X86::IST_Fp64m32 ||
1136                     MI.getOpcode() == X86::ISTT_Fp16m32 ||
1137                     MI.getOpcode() == X86::ISTT_Fp32m32 ||
1138                     MI.getOpcode() == X86::ISTT_Fp64m32 ||
1139                     MI.getOpcode() == X86::IST_Fp64m64 ||
1140                     MI.getOpcode() == X86::ISTT_Fp16m64 ||
1141                     MI.getOpcode() == X86::ISTT_Fp32m64 ||
1142                     MI.getOpcode() == X86::ISTT_Fp64m64 ||
1143                     MI.getOpcode() == X86::IST_Fp64m80 ||
1144                     MI.getOpcode() == X86::ISTT_Fp16m80 ||
1145                     MI.getOpcode() == X86::ISTT_Fp32m80 ||
1146                     MI.getOpcode() == X86::ISTT_Fp64m80 ||
1147                     MI.getOpcode() == X86::ST_FpP80m)) {
1148     duplicateToTop(Reg, ScratchFPReg, I);
1149   } else {
1150     moveToTop(Reg, I);            // Move to the top of the stack...
1151   }
1152 
1153   // Convert from the pseudo instruction to the concrete instruction.
1154   MI.RemoveOperand(NumOps - 1); // Remove explicit ST(0) operand
1155   MI.setDesc(TII->get(getConcreteOpcode(MI.getOpcode())));
1156   MI.addOperand(
1157       MachineOperand::CreateReg(X86::ST0, /*isDef*/ false, /*isImp*/ true));
1158 
1159   if (MI.getOpcode() == X86::IST_FP64m || MI.getOpcode() == X86::ISTT_FP16m ||
1160       MI.getOpcode() == X86::ISTT_FP32m || MI.getOpcode() == X86::ISTT_FP64m ||
1161       MI.getOpcode() == X86::ST_FP80m) {
1162     if (StackTop == 0)
1163       report_fatal_error("Stack empty??");
1164     --StackTop;
1165   } else if (KillsSrc) { // Last use of operand?
1166     popStackAfter(I);
1167   }
1168 }
1169 
1170 
1171 /// handleOneArgFPRW: Handle instructions that read from the top of stack and
1172 /// replace the value with a newly computed value.  These instructions may have
1173 /// non-fp operands after their FP operands.
1174 ///
1175 ///  Examples:
1176 ///     R1 = fchs R2
1177 ///     R1 = fadd R2, [mem]
1178 ///
handleOneArgFPRW(MachineBasicBlock::iterator & I)1179 void FPS::handleOneArgFPRW(MachineBasicBlock::iterator &I) {
1180   MachineInstr &MI = *I;
1181 #ifndef NDEBUG
1182   unsigned NumOps = MI.getDesc().getNumOperands();
1183   assert(NumOps >= 2 && "FPRW instructions must have 2 ops!!");
1184 #endif
1185 
1186   // Is this the last use of the source register?
1187   unsigned Reg = getFPReg(MI.getOperand(1));
1188   bool KillsSrc = MI.killsRegister(X86::FP0 + Reg);
1189 
1190   if (KillsSrc) {
1191     // If this is the last use of the source register, just make sure it's on
1192     // the top of the stack.
1193     moveToTop(Reg, I);
1194     if (StackTop == 0)
1195       report_fatal_error("Stack cannot be empty!");
1196     --StackTop;
1197     pushReg(getFPReg(MI.getOperand(0)));
1198   } else {
1199     // If this is not the last use of the source register, _copy_ it to the top
1200     // of the stack.
1201     duplicateToTop(Reg, getFPReg(MI.getOperand(0)), I);
1202   }
1203 
1204   // Change from the pseudo instruction to the concrete instruction.
1205   MI.RemoveOperand(1); // Drop the source operand.
1206   MI.RemoveOperand(0); // Drop the destination operand.
1207   MI.setDesc(TII->get(getConcreteOpcode(MI.getOpcode())));
1208 }
1209 
1210 
1211 //===----------------------------------------------------------------------===//
1212 // Define tables of various ways to map pseudo instructions
1213 //
1214 
1215 // ForwardST0Table - Map: A = B op C  into: ST(0) = ST(0) op ST(i)
1216 static const TableEntry ForwardST0Table[] = {
1217   { X86::ADD_Fp32  , X86::ADD_FST0r },
1218   { X86::ADD_Fp64  , X86::ADD_FST0r },
1219   { X86::ADD_Fp80  , X86::ADD_FST0r },
1220   { X86::DIV_Fp32  , X86::DIV_FST0r },
1221   { X86::DIV_Fp64  , X86::DIV_FST0r },
1222   { X86::DIV_Fp80  , X86::DIV_FST0r },
1223   { X86::MUL_Fp32  , X86::MUL_FST0r },
1224   { X86::MUL_Fp64  , X86::MUL_FST0r },
1225   { X86::MUL_Fp80  , X86::MUL_FST0r },
1226   { X86::SUB_Fp32  , X86::SUB_FST0r },
1227   { X86::SUB_Fp64  , X86::SUB_FST0r },
1228   { X86::SUB_Fp80  , X86::SUB_FST0r },
1229 };
1230 
1231 // ReverseST0Table - Map: A = B op C  into: ST(0) = ST(i) op ST(0)
1232 static const TableEntry ReverseST0Table[] = {
1233   { X86::ADD_Fp32  , X86::ADD_FST0r  },   // commutative
1234   { X86::ADD_Fp64  , X86::ADD_FST0r  },   // commutative
1235   { X86::ADD_Fp80  , X86::ADD_FST0r  },   // commutative
1236   { X86::DIV_Fp32  , X86::DIVR_FST0r },
1237   { X86::DIV_Fp64  , X86::DIVR_FST0r },
1238   { X86::DIV_Fp80  , X86::DIVR_FST0r },
1239   { X86::MUL_Fp32  , X86::MUL_FST0r  },   // commutative
1240   { X86::MUL_Fp64  , X86::MUL_FST0r  },   // commutative
1241   { X86::MUL_Fp80  , X86::MUL_FST0r  },   // commutative
1242   { X86::SUB_Fp32  , X86::SUBR_FST0r },
1243   { X86::SUB_Fp64  , X86::SUBR_FST0r },
1244   { X86::SUB_Fp80  , X86::SUBR_FST0r },
1245 };
1246 
1247 // ForwardSTiTable - Map: A = B op C  into: ST(i) = ST(0) op ST(i)
1248 static const TableEntry ForwardSTiTable[] = {
1249   { X86::ADD_Fp32  , X86::ADD_FrST0  },   // commutative
1250   { X86::ADD_Fp64  , X86::ADD_FrST0  },   // commutative
1251   { X86::ADD_Fp80  , X86::ADD_FrST0  },   // commutative
1252   { X86::DIV_Fp32  , X86::DIVR_FrST0 },
1253   { X86::DIV_Fp64  , X86::DIVR_FrST0 },
1254   { X86::DIV_Fp80  , X86::DIVR_FrST0 },
1255   { X86::MUL_Fp32  , X86::MUL_FrST0  },   // commutative
1256   { X86::MUL_Fp64  , X86::MUL_FrST0  },   // commutative
1257   { X86::MUL_Fp80  , X86::MUL_FrST0  },   // commutative
1258   { X86::SUB_Fp32  , X86::SUBR_FrST0 },
1259   { X86::SUB_Fp64  , X86::SUBR_FrST0 },
1260   { X86::SUB_Fp80  , X86::SUBR_FrST0 },
1261 };
1262 
1263 // ReverseSTiTable - Map: A = B op C  into: ST(i) = ST(i) op ST(0)
1264 static const TableEntry ReverseSTiTable[] = {
1265   { X86::ADD_Fp32  , X86::ADD_FrST0 },
1266   { X86::ADD_Fp64  , X86::ADD_FrST0 },
1267   { X86::ADD_Fp80  , X86::ADD_FrST0 },
1268   { X86::DIV_Fp32  , X86::DIV_FrST0 },
1269   { X86::DIV_Fp64  , X86::DIV_FrST0 },
1270   { X86::DIV_Fp80  , X86::DIV_FrST0 },
1271   { X86::MUL_Fp32  , X86::MUL_FrST0 },
1272   { X86::MUL_Fp64  , X86::MUL_FrST0 },
1273   { X86::MUL_Fp80  , X86::MUL_FrST0 },
1274   { X86::SUB_Fp32  , X86::SUB_FrST0 },
1275   { X86::SUB_Fp64  , X86::SUB_FrST0 },
1276   { X86::SUB_Fp80  , X86::SUB_FrST0 },
1277 };
1278 
1279 
1280 /// handleTwoArgFP - Handle instructions like FADD and friends which are virtual
1281 /// instructions which need to be simplified and possibly transformed.
1282 ///
1283 /// Result: ST(0) = fsub  ST(0), ST(i)
1284 ///         ST(i) = fsub  ST(0), ST(i)
1285 ///         ST(0) = fsubr ST(0), ST(i)
1286 ///         ST(i) = fsubr ST(0), ST(i)
1287 ///
handleTwoArgFP(MachineBasicBlock::iterator & I)1288 void FPS::handleTwoArgFP(MachineBasicBlock::iterator &I) {
1289   ASSERT_SORTED(ForwardST0Table); ASSERT_SORTED(ReverseST0Table);
1290   ASSERT_SORTED(ForwardSTiTable); ASSERT_SORTED(ReverseSTiTable);
1291   MachineInstr &MI = *I;
1292 
1293   unsigned NumOperands = MI.getDesc().getNumOperands();
1294   assert(NumOperands == 3 && "Illegal TwoArgFP instruction!");
1295   unsigned Dest = getFPReg(MI.getOperand(0));
1296   unsigned Op0 = getFPReg(MI.getOperand(NumOperands - 2));
1297   unsigned Op1 = getFPReg(MI.getOperand(NumOperands - 1));
1298   bool KillsOp0 = MI.killsRegister(X86::FP0 + Op0);
1299   bool KillsOp1 = MI.killsRegister(X86::FP0 + Op1);
1300   DebugLoc dl = MI.getDebugLoc();
1301 
1302   unsigned TOS = getStackEntry(0);
1303 
1304   // One of our operands must be on the top of the stack.  If neither is yet, we
1305   // need to move one.
1306   if (Op0 != TOS && Op1 != TOS) {   // No operand at TOS?
1307     // We can choose to move either operand to the top of the stack.  If one of
1308     // the operands is killed by this instruction, we want that one so that we
1309     // can update right on top of the old version.
1310     if (KillsOp0) {
1311       moveToTop(Op0, I);         // Move dead operand to TOS.
1312       TOS = Op0;
1313     } else if (KillsOp1) {
1314       moveToTop(Op1, I);
1315       TOS = Op1;
1316     } else {
1317       // All of the operands are live after this instruction executes, so we
1318       // cannot update on top of any operand.  Because of this, we must
1319       // duplicate one of the stack elements to the top.  It doesn't matter
1320       // which one we pick.
1321       //
1322       duplicateToTop(Op0, Dest, I);
1323       Op0 = TOS = Dest;
1324       KillsOp0 = true;
1325     }
1326   } else if (!KillsOp0 && !KillsOp1) {
1327     // If we DO have one of our operands at the top of the stack, but we don't
1328     // have a dead operand, we must duplicate one of the operands to a new slot
1329     // on the stack.
1330     duplicateToTop(Op0, Dest, I);
1331     Op0 = TOS = Dest;
1332     KillsOp0 = true;
1333   }
1334 
1335   // Now we know that one of our operands is on the top of the stack, and at
1336   // least one of our operands is killed by this instruction.
1337   assert((TOS == Op0 || TOS == Op1) && (KillsOp0 || KillsOp1) &&
1338          "Stack conditions not set up right!");
1339 
1340   // We decide which form to use based on what is on the top of the stack, and
1341   // which operand is killed by this instruction.
1342   ArrayRef<TableEntry> InstTable;
1343   bool isForward = TOS == Op0;
1344   bool updateST0 = (TOS == Op0 && !KillsOp1) || (TOS == Op1 && !KillsOp0);
1345   if (updateST0) {
1346     if (isForward)
1347       InstTable = ForwardST0Table;
1348     else
1349       InstTable = ReverseST0Table;
1350   } else {
1351     if (isForward)
1352       InstTable = ForwardSTiTable;
1353     else
1354       InstTable = ReverseSTiTable;
1355   }
1356 
1357   int Opcode = Lookup(InstTable, MI.getOpcode());
1358   assert(Opcode != -1 && "Unknown TwoArgFP pseudo instruction!");
1359 
1360   // NotTOS - The register which is not on the top of stack...
1361   unsigned NotTOS = (TOS == Op0) ? Op1 : Op0;
1362 
1363   // Replace the old instruction with a new instruction
1364   MBB->remove(&*I++);
1365   I = BuildMI(*MBB, I, dl, TII->get(Opcode)).addReg(getSTReg(NotTOS));
1366 
1367   if (!MI.mayRaiseFPException())
1368     I->setFlag(MachineInstr::MIFlag::NoFPExcept);
1369 
1370   // If both operands are killed, pop one off of the stack in addition to
1371   // overwriting the other one.
1372   if (KillsOp0 && KillsOp1 && Op0 != Op1) {
1373     assert(!updateST0 && "Should have updated other operand!");
1374     popStackAfter(I);   // Pop the top of stack
1375   }
1376 
1377   // Update stack information so that we know the destination register is now on
1378   // the stack.
1379   unsigned UpdatedSlot = getSlot(updateST0 ? TOS : NotTOS);
1380   assert(UpdatedSlot < StackTop && Dest < 7);
1381   Stack[UpdatedSlot]   = Dest;
1382   RegMap[Dest]         = UpdatedSlot;
1383   MBB->getParent()->DeleteMachineInstr(&MI); // Remove the old instruction
1384 }
1385 
1386 /// handleCompareFP - Handle FUCOM and FUCOMI instructions, which have two FP
1387 /// register arguments and no explicit destinations.
1388 ///
handleCompareFP(MachineBasicBlock::iterator & I)1389 void FPS::handleCompareFP(MachineBasicBlock::iterator &I) {
1390   MachineInstr &MI = *I;
1391 
1392   unsigned NumOperands = MI.getDesc().getNumOperands();
1393   assert(NumOperands == 2 && "Illegal FUCOM* instruction!");
1394   unsigned Op0 = getFPReg(MI.getOperand(NumOperands - 2));
1395   unsigned Op1 = getFPReg(MI.getOperand(NumOperands - 1));
1396   bool KillsOp0 = MI.killsRegister(X86::FP0 + Op0);
1397   bool KillsOp1 = MI.killsRegister(X86::FP0 + Op1);
1398 
1399   // Make sure the first operand is on the top of stack, the other one can be
1400   // anywhere.
1401   moveToTop(Op0, I);
1402 
1403   // Change from the pseudo instruction to the concrete instruction.
1404   MI.getOperand(0).setReg(getSTReg(Op1));
1405   MI.RemoveOperand(1);
1406   MI.setDesc(TII->get(getConcreteOpcode(MI.getOpcode())));
1407 
1408   // If any of the operands are killed by this instruction, free them.
1409   if (KillsOp0) freeStackSlotAfter(I, Op0);
1410   if (KillsOp1 && Op0 != Op1) freeStackSlotAfter(I, Op1);
1411 }
1412 
1413 /// handleCondMovFP - Handle two address conditional move instructions.  These
1414 /// instructions move a st(i) register to st(0) iff a condition is true.  These
1415 /// instructions require that the first operand is at the top of the stack, but
1416 /// otherwise don't modify the stack at all.
handleCondMovFP(MachineBasicBlock::iterator & I)1417 void FPS::handleCondMovFP(MachineBasicBlock::iterator &I) {
1418   MachineInstr &MI = *I;
1419 
1420   unsigned Op0 = getFPReg(MI.getOperand(0));
1421   unsigned Op1 = getFPReg(MI.getOperand(2));
1422   bool KillsOp1 = MI.killsRegister(X86::FP0 + Op1);
1423 
1424   // The first operand *must* be on the top of the stack.
1425   moveToTop(Op0, I);
1426 
1427   // Change the second operand to the stack register that the operand is in.
1428   // Change from the pseudo instruction to the concrete instruction.
1429   MI.RemoveOperand(0);
1430   MI.RemoveOperand(1);
1431   MI.getOperand(0).setReg(getSTReg(Op1));
1432   MI.setDesc(TII->get(getConcreteOpcode(MI.getOpcode())));
1433 
1434   // If we kill the second operand, make sure to pop it from the stack.
1435   if (Op0 != Op1 && KillsOp1) {
1436     // Get this value off of the register stack.
1437     freeStackSlotAfter(I, Op1);
1438   }
1439 }
1440 
1441 
1442 /// handleSpecialFP - Handle special instructions which behave unlike other
1443 /// floating point instructions.  This is primarily intended for use by pseudo
1444 /// instructions.
1445 ///
handleSpecialFP(MachineBasicBlock::iterator & Inst)1446 void FPS::handleSpecialFP(MachineBasicBlock::iterator &Inst) {
1447   MachineInstr &MI = *Inst;
1448 
1449   if (MI.isCall()) {
1450     handleCall(Inst);
1451     return;
1452   }
1453 
1454   if (MI.isReturn()) {
1455     handleReturn(Inst);
1456     return;
1457   }
1458 
1459   switch (MI.getOpcode()) {
1460   default: llvm_unreachable("Unknown SpecialFP instruction!");
1461   case TargetOpcode::COPY: {
1462     // We handle three kinds of copies: FP <- FP, FP <- ST, and ST <- FP.
1463     const MachineOperand &MO1 = MI.getOperand(1);
1464     const MachineOperand &MO0 = MI.getOperand(0);
1465     bool KillsSrc = MI.killsRegister(MO1.getReg());
1466 
1467     // FP <- FP copy.
1468     unsigned DstFP = getFPReg(MO0);
1469     unsigned SrcFP = getFPReg(MO1);
1470     assert(isLive(SrcFP) && "Cannot copy dead register");
1471     if (KillsSrc) {
1472       // If the input operand is killed, we can just change the owner of the
1473       // incoming stack slot into the result.
1474       unsigned Slot = getSlot(SrcFP);
1475       Stack[Slot] = DstFP;
1476       RegMap[DstFP] = Slot;
1477     } else {
1478       // For COPY we just duplicate the specified value to a new stack slot.
1479       // This could be made better, but would require substantial changes.
1480       duplicateToTop(SrcFP, DstFP, Inst);
1481     }
1482     break;
1483   }
1484 
1485   case TargetOpcode::IMPLICIT_DEF: {
1486     // All FP registers must be explicitly defined, so load a 0 instead.
1487     unsigned Reg = MI.getOperand(0).getReg() - X86::FP0;
1488     LLVM_DEBUG(dbgs() << "Emitting LD_F0 for implicit FP" << Reg << '\n');
1489     BuildMI(*MBB, Inst, MI.getDebugLoc(), TII->get(X86::LD_F0));
1490     pushReg(Reg);
1491     break;
1492   }
1493 
1494   case TargetOpcode::INLINEASM:
1495   case TargetOpcode::INLINEASM_BR: {
1496     // The inline asm MachineInstr currently only *uses* FP registers for the
1497     // 'f' constraint.  These should be turned into the current ST(x) register
1498     // in the machine instr.
1499     //
1500     // There are special rules for x87 inline assembly. The compiler must know
1501     // exactly how many registers are popped and pushed implicitly by the asm.
1502     // Otherwise it is not possible to restore the stack state after the inline
1503     // asm.
1504     //
1505     // There are 3 kinds of input operands:
1506     //
1507     // 1. Popped inputs. These must appear at the stack top in ST0-STn. A
1508     //    popped input operand must be in a fixed stack slot, and it is either
1509     //    tied to an output operand, or in the clobber list. The MI has ST use
1510     //    and def operands for these inputs.
1511     //
1512     // 2. Fixed inputs. These inputs appear in fixed stack slots, but are
1513     //    preserved by the inline asm. The fixed stack slots must be STn-STm
1514     //    following the popped inputs. A fixed input operand cannot be tied to
1515     //    an output or appear in the clobber list. The MI has ST use operands
1516     //    and no defs for these inputs.
1517     //
1518     // 3. Preserved inputs. These inputs use the "f" constraint which is
1519     //    represented as an FP register. The inline asm won't change these
1520     //    stack slots.
1521     //
1522     // Outputs must be in ST registers, FP outputs are not allowed. Clobbered
1523     // registers do not count as output operands. The inline asm changes the
1524     // stack as if it popped all the popped inputs and then pushed all the
1525     // output operands.
1526 
1527     // Scan the assembly for ST registers used, defined and clobbered. We can
1528     // only tell clobbers from defs by looking at the asm descriptor.
1529     unsigned STUses = 0, STDefs = 0, STClobbers = 0, STDeadDefs = 0;
1530     unsigned NumOps = 0;
1531     SmallSet<unsigned, 1> FRegIdx;
1532     unsigned RCID;
1533 
1534     for (unsigned i = InlineAsm::MIOp_FirstOperand, e = MI.getNumOperands();
1535          i != e && MI.getOperand(i).isImm(); i += 1 + NumOps) {
1536       unsigned Flags = MI.getOperand(i).getImm();
1537 
1538       NumOps = InlineAsm::getNumOperandRegisters(Flags);
1539       if (NumOps != 1)
1540         continue;
1541       const MachineOperand &MO = MI.getOperand(i + 1);
1542       if (!MO.isReg())
1543         continue;
1544       unsigned STReg = MO.getReg() - X86::FP0;
1545       if (STReg >= 8)
1546         continue;
1547 
1548       // If the flag has a register class constraint, this must be an operand
1549       // with constraint "f". Record its index and continue.
1550       if (InlineAsm::hasRegClassConstraint(Flags, RCID)) {
1551         FRegIdx.insert(i + 1);
1552         continue;
1553       }
1554 
1555       switch (InlineAsm::getKind(Flags)) {
1556       case InlineAsm::Kind_RegUse:
1557         STUses |= (1u << STReg);
1558         break;
1559       case InlineAsm::Kind_RegDef:
1560       case InlineAsm::Kind_RegDefEarlyClobber:
1561         STDefs |= (1u << STReg);
1562         if (MO.isDead())
1563           STDeadDefs |= (1u << STReg);
1564         break;
1565       case InlineAsm::Kind_Clobber:
1566         STClobbers |= (1u << STReg);
1567         break;
1568       default:
1569         break;
1570       }
1571     }
1572 
1573     if (STUses && !isMask_32(STUses))
1574       MI.emitError("fixed input regs must be last on the x87 stack");
1575     unsigned NumSTUses = countTrailingOnes(STUses);
1576 
1577     // Defs must be contiguous from the stack top. ST0-STn.
1578     if (STDefs && !isMask_32(STDefs)) {
1579       MI.emitError("output regs must be last on the x87 stack");
1580       STDefs = NextPowerOf2(STDefs) - 1;
1581     }
1582     unsigned NumSTDefs = countTrailingOnes(STDefs);
1583 
1584     // So must the clobbered stack slots. ST0-STm, m >= n.
1585     if (STClobbers && !isMask_32(STDefs | STClobbers))
1586       MI.emitError("clobbers must be last on the x87 stack");
1587 
1588     // Popped inputs are the ones that are also clobbered or defined.
1589     unsigned STPopped = STUses & (STDefs | STClobbers);
1590     if (STPopped && !isMask_32(STPopped))
1591       MI.emitError("implicitly popped regs must be last on the x87 stack");
1592     unsigned NumSTPopped = countTrailingOnes(STPopped);
1593 
1594     LLVM_DEBUG(dbgs() << "Asm uses " << NumSTUses << " fixed regs, pops "
1595                       << NumSTPopped << ", and defines " << NumSTDefs
1596                       << " regs.\n");
1597 
1598 #ifndef NDEBUG
1599     // If any input operand uses constraint "f", all output register
1600     // constraints must be early-clobber defs.
1601     for (unsigned I = 0, E = MI.getNumOperands(); I < E; ++I)
1602       if (FRegIdx.count(I)) {
1603         assert((1 << getFPReg(MI.getOperand(I)) & STDefs) == 0 &&
1604                "Operands with constraint \"f\" cannot overlap with defs");
1605       }
1606 #endif
1607 
1608     // Collect all FP registers (register operands with constraints "t", "u",
1609     // and "f") to kill afer the instruction.
1610     unsigned FPKills = ((1u << NumFPRegs) - 1) & ~0xff;
1611     for (unsigned i = 0, e = MI.getNumOperands(); i != e; ++i) {
1612       MachineOperand &Op = MI.getOperand(i);
1613       if (!Op.isReg() || Op.getReg() < X86::FP0 || Op.getReg() > X86::FP6)
1614         continue;
1615       unsigned FPReg = getFPReg(Op);
1616 
1617       // If we kill this operand, make sure to pop it from the stack after the
1618       // asm.  We just remember it for now, and pop them all off at the end in
1619       // a batch.
1620       if (Op.isUse() && Op.isKill())
1621         FPKills |= 1U << FPReg;
1622     }
1623 
1624     // Do not include registers that are implicitly popped by defs/clobbers.
1625     FPKills &= ~(STDefs | STClobbers);
1626 
1627     // Now we can rearrange the live registers to match what was requested.
1628     unsigned char STUsesArray[8];
1629 
1630     for (unsigned I = 0; I < NumSTUses; ++I)
1631       STUsesArray[I] = I;
1632 
1633     shuffleStackTop(STUsesArray, NumSTUses, Inst);
1634     LLVM_DEBUG({
1635       dbgs() << "Before asm: ";
1636       dumpStack();
1637     });
1638 
1639     // With the stack layout fixed, rewrite the FP registers.
1640     for (unsigned i = 0, e = MI.getNumOperands(); i != e; ++i) {
1641       MachineOperand &Op = MI.getOperand(i);
1642       if (!Op.isReg() || Op.getReg() < X86::FP0 || Op.getReg() > X86::FP6)
1643         continue;
1644 
1645       unsigned FPReg = getFPReg(Op);
1646 
1647       if (FRegIdx.count(i))
1648         // Operand with constraint "f".
1649         Op.setReg(getSTReg(FPReg));
1650       else
1651         // Operand with a single register class constraint ("t" or "u").
1652         Op.setReg(X86::ST0 + FPReg);
1653     }
1654 
1655     // Simulate the inline asm popping its inputs and pushing its outputs.
1656     StackTop -= NumSTPopped;
1657 
1658     for (unsigned i = 0; i < NumSTDefs; ++i)
1659       pushReg(NumSTDefs - i - 1);
1660 
1661     // If this asm kills any FP registers (is the last use of them) we must
1662     // explicitly emit pop instructions for them.  Do this now after the asm has
1663     // executed so that the ST(x) numbers are not off (which would happen if we
1664     // did this inline with operand rewriting).
1665     //
1666     // Note: this might be a non-optimal pop sequence.  We might be able to do
1667     // better by trying to pop in stack order or something.
1668     while (FPKills) {
1669       unsigned FPReg = countTrailingZeros(FPKills);
1670       if (isLive(FPReg))
1671         freeStackSlotAfter(Inst, FPReg);
1672       FPKills &= ~(1U << FPReg);
1673     }
1674 
1675     // Don't delete the inline asm!
1676     return;
1677   }
1678   }
1679 
1680   Inst = MBB->erase(Inst);  // Remove the pseudo instruction
1681 
1682   // We want to leave I pointing to the previous instruction, but what if we
1683   // just erased the first instruction?
1684   if (Inst == MBB->begin()) {
1685     LLVM_DEBUG(dbgs() << "Inserting dummy KILL\n");
1686     Inst = BuildMI(*MBB, Inst, DebugLoc(), TII->get(TargetOpcode::KILL));
1687   } else
1688     --Inst;
1689 }
1690 
setKillFlags(MachineBasicBlock & MBB) const1691 void FPS::setKillFlags(MachineBasicBlock &MBB) const {
1692   const TargetRegisterInfo &TRI =
1693       *MBB.getParent()->getSubtarget().getRegisterInfo();
1694   LivePhysRegs LPR(TRI);
1695 
1696   LPR.addLiveOuts(MBB);
1697 
1698   for (MachineBasicBlock::reverse_iterator I = MBB.rbegin(), E = MBB.rend();
1699        I != E; ++I) {
1700     if (I->isDebugInstr())
1701       continue;
1702 
1703     std::bitset<8> Defs;
1704     SmallVector<MachineOperand *, 2> Uses;
1705     MachineInstr &MI = *I;
1706 
1707     for (auto &MO : I->operands()) {
1708       if (!MO.isReg())
1709         continue;
1710 
1711       unsigned Reg = MO.getReg() - X86::FP0;
1712 
1713       if (Reg >= 8)
1714         continue;
1715 
1716       if (MO.isDef()) {
1717         Defs.set(Reg);
1718         if (!LPR.contains(MO.getReg()))
1719           MO.setIsDead();
1720       } else
1721         Uses.push_back(&MO);
1722     }
1723 
1724     for (auto *MO : Uses)
1725       if (Defs.test(getFPReg(*MO)) || !LPR.contains(MO->getReg()))
1726         MO->setIsKill();
1727 
1728     LPR.stepBackward(MI);
1729   }
1730 }
1731