1 //===-- X86FloatingPoint.cpp - Floating point Reg -> Stack converter ------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file defines the pass which converts floating point instructions from
11 // pseudo registers into register stack instructions. This pass uses live
12 // variable information to indicate where the FPn registers are used and their
15 // The x87 hardware tracks liveness of the stack registers, so it is necessary
16 // to implement exact liveness tracking between basic blocks. The CFG edges are
17 // partitioned into bundles where the same FP registers must be live in
18 // identical stack positions. Instructions are inserted at the end of each basic
19 // block to rearrange the live registers to match the outgoing bundle.
21 // This approach avoids splitting critical edges at the potential cost of more
22 // live register shuffling instructions when critical edges are present.
24 //===----------------------------------------------------------------------===//
26 #define DEBUG_TYPE "x86-codegen"
28 #include "X86InstrInfo.h"
29 #include "llvm/ADT/DepthFirstIterator.h"
30 #include "llvm/ADT/STLExtras.h"
31 #include "llvm/ADT/SmallPtrSet.h"
32 #include "llvm/ADT/SmallVector.h"
33 #include "llvm/ADT/Statistic.h"
34 #include "llvm/CodeGen/EdgeBundles.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/IR/InlineAsm.h"
40 #include "llvm/Support/Debug.h"
41 #include "llvm/Support/ErrorHandling.h"
42 #include "llvm/Support/raw_ostream.h"
43 #include "llvm/Target/TargetInstrInfo.h"
44 #include "llvm/Target/TargetMachine.h"
48 STATISTIC(NumFXCH, "Number of fxch instructions inserted");
49 STATISTIC(NumFP , "Number of floating point instructions");
52 struct FPS : public MachineFunctionPass {
54 FPS() : MachineFunctionPass(ID) {
55 initializeEdgeBundlesPass(*PassRegistry::getPassRegistry());
56 // This is really only to keep valgrind quiet.
57 // The logic in isLive() is too much for it.
58 memset(Stack, 0, sizeof(Stack));
59 memset(RegMap, 0, sizeof(RegMap));
62 void getAnalysisUsage(AnalysisUsage &AU) const override {
64 AU.addRequired<EdgeBundles>();
65 AU.addPreservedID(MachineLoopInfoID);
66 AU.addPreservedID(MachineDominatorsID);
67 MachineFunctionPass::getAnalysisUsage(AU);
70 bool runOnMachineFunction(MachineFunction &MF) override;
72 const char *getPassName() const override { return "X86 FP Stackifier"; }
75 const TargetInstrInfo *TII; // Machine instruction info.
77 // Two CFG edges are related if they leave the same block, or enter the same
78 // block. The transitive closure of an edge under this relation is a
79 // LiveBundle. It represents a set of CFG edges where the live FP stack
80 // registers must be allocated identically in the x87 stack.
82 // A LiveBundle is usually all the edges leaving a block, or all the edges
83 // entering a block, but it can contain more edges if critical edges are
86 // The set of live FP registers in a LiveBundle is calculated by bundleCFG,
87 // but the exact mapping of FP registers to stack slots is fixed later.
89 // Bit mask of live FP registers. Bit 0 = FP0, bit 1 = FP1, &c.
92 // Number of pre-assigned live registers in FixStack. This is 0 when the
93 // stack order has not yet been fixed.
96 // Assigned stack order for live-in registers.
97 // FixStack[i] == getStackEntry(i) for all i < FixCount.
98 unsigned char FixStack[8];
100 LiveBundle() : Mask(0), FixCount(0) {}
102 // Have the live registers been assigned a stack order yet?
103 bool isFixed() const { return !Mask || FixCount; }
106 // Numbered LiveBundle structs. LiveBundles[0] is used for all CFG edges
107 // with no live FP registers.
108 SmallVector<LiveBundle, 8> LiveBundles;
110 // The edge bundle analysis provides indices into the LiveBundles vector.
111 EdgeBundles *Bundles;
113 // Return a bitmask of FP registers in block's live-in list.
114 static unsigned calcLiveInMask(MachineBasicBlock *MBB) {
116 for (MachineBasicBlock::livein_iterator I = MBB->livein_begin(),
117 E = MBB->livein_end(); I != E; ++I) {
119 if (Reg < X86::FP0 || Reg > X86::FP6)
121 Mask |= 1 << (Reg - X86::FP0);
126 // Partition all the CFG edges into LiveBundles.
127 void bundleCFG(MachineFunction &MF);
129 MachineBasicBlock *MBB; // Current basic block
131 // The hardware keeps track of how many FP registers are live, so we have
132 // to model that exactly. Usually, each live register corresponds to an
133 // FP<n> register, but when dealing with calls, returns, and inline
134 // assembly, it is sometimes necessary to have live scratch registers.
135 unsigned Stack[8]; // FP<n> Registers in each stack slot...
136 unsigned StackTop; // The current top of the FP stack.
139 NumFPRegs = 16 // Including scratch pseudo-registers.
142 // For each live FP<n> register, point to its Stack[] entry.
143 // The first entries correspond to FP0-FP6, the rest are scratch registers
144 // used when we need slightly different live registers than what the
145 // register allocator thinks.
146 unsigned RegMap[NumFPRegs];
148 // Pending fixed registers - Inline assembly needs FP registers to appear
149 // in fixed stack slot positions. This is handled by copying FP registers
150 // to ST registers before the instruction, and copying back after the
153 // This is modeled with pending ST registers. NumPendingSTs is the number
154 // of ST registers (ST0-STn) we are tracking. PendingST[n] points to an FP
155 // register that holds the ST value. The ST registers are not moved into
156 // place until immediately before the instruction that needs them.
158 // It can happen that we need an ST register to be live when no FP register
161 // %ST0 = COPY %FP4<kill>
163 // When that happens, we allocate a scratch FP register to hold the ST
164 // value. That means every register in PendingST must be live.
166 unsigned NumPendingSTs;
167 unsigned char PendingST[8];
169 // Set up our stack model to match the incoming registers to MBB.
170 void setupBlockStack();
172 // Shuffle live registers to match the expectations of successor blocks.
173 void finishBlockStack();
175 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
176 void dumpStack() const {
177 dbgs() << "Stack contents:";
178 for (unsigned i = 0; i != StackTop; ++i) {
179 dbgs() << " FP" << Stack[i];
180 assert(RegMap[Stack[i]] == i && "Stack[] doesn't match RegMap[]!");
182 for (unsigned i = 0; i != NumPendingSTs; ++i)
183 dbgs() << ", ST" << i << " in FP" << unsigned(PendingST[i]);
188 /// getSlot - Return the stack slot number a particular register number is
190 unsigned getSlot(unsigned RegNo) const {
191 assert(RegNo < NumFPRegs && "Regno out of range!");
192 return RegMap[RegNo];
195 /// isLive - Is RegNo currently live in the stack?
196 bool isLive(unsigned RegNo) const {
197 unsigned Slot = getSlot(RegNo);
198 return Slot < StackTop && Stack[Slot] == RegNo;
201 /// getScratchReg - Return an FP register that is not currently in use.
202 unsigned getScratchReg() const {
203 for (int i = NumFPRegs - 1; i >= 8; --i)
206 llvm_unreachable("Ran out of scratch FP registers");
209 /// isScratchReg - Returns trus if RegNo is a scratch FP register.
210 static bool isScratchReg(unsigned RegNo) {
211 return RegNo > 8 && RegNo < NumFPRegs;
214 /// getStackEntry - Return the X86::FP<n> register in register ST(i).
215 unsigned getStackEntry(unsigned STi) const {
217 report_fatal_error("Access past stack top!");
218 return Stack[StackTop-1-STi];
221 /// getSTReg - Return the X86::ST(i) register which contains the specified
222 /// FP<RegNo> register.
223 unsigned getSTReg(unsigned RegNo) const {
224 return StackTop - 1 - getSlot(RegNo) + X86::ST0;
227 // pushReg - Push the specified FP<n> register onto the stack.
228 void pushReg(unsigned Reg) {
229 assert(Reg < NumFPRegs && "Register number out of range!");
231 report_fatal_error("Stack overflow!");
232 Stack[StackTop] = Reg;
233 RegMap[Reg] = StackTop++;
236 bool isAtTop(unsigned RegNo) const { return getSlot(RegNo) == StackTop-1; }
237 void moveToTop(unsigned RegNo, MachineBasicBlock::iterator I) {
238 DebugLoc dl = I == MBB->end() ? DebugLoc() : I->getDebugLoc();
239 if (isAtTop(RegNo)) return;
241 unsigned STReg = getSTReg(RegNo);
242 unsigned RegOnTop = getStackEntry(0);
244 // Swap the slots the regs are in.
245 std::swap(RegMap[RegNo], RegMap[RegOnTop]);
247 // Swap stack slot contents.
248 if (RegMap[RegOnTop] >= StackTop)
249 report_fatal_error("Access past stack top!");
250 std::swap(Stack[RegMap[RegOnTop]], Stack[StackTop-1]);
252 // Emit an fxch to update the runtime processors version of the state.
253 BuildMI(*MBB, I, dl, TII->get(X86::XCH_F)).addReg(STReg);
257 void duplicateToTop(unsigned RegNo, unsigned AsReg, MachineInstr *I) {
258 DebugLoc dl = I == MBB->end() ? DebugLoc() : I->getDebugLoc();
259 unsigned STReg = getSTReg(RegNo);
260 pushReg(AsReg); // New register on top of stack
262 BuildMI(*MBB, I, dl, TII->get(X86::LD_Frr)).addReg(STReg);
265 /// duplicatePendingSTBeforeKill - The instruction at I is about to kill
266 /// RegNo. If any PendingST registers still need the RegNo value, duplicate
267 /// them to new scratch registers.
268 void duplicatePendingSTBeforeKill(unsigned RegNo, MachineInstr *I) {
269 for (unsigned i = 0; i != NumPendingSTs; ++i) {
270 if (PendingST[i] != RegNo)
272 unsigned SR = getScratchReg();
273 DEBUG(dbgs() << "Duplicating pending ST" << i
274 << " in FP" << RegNo << " to FP" << SR << '\n');
275 duplicateToTop(RegNo, SR, I);
280 /// popStackAfter - Pop the current value off of the top of the FP stack
281 /// after the specified instruction.
282 void popStackAfter(MachineBasicBlock::iterator &I);
284 /// freeStackSlotAfter - Free the specified register from the register
285 /// stack, so that it is no longer in a register. If the register is
286 /// currently at the top of the stack, we just pop the current instruction,
287 /// otherwise we store the current top-of-stack into the specified slot,
288 /// then pop the top of stack.
289 void freeStackSlotAfter(MachineBasicBlock::iterator &I, unsigned Reg);
291 /// freeStackSlotBefore - Just the pop, no folding. Return the inserted
293 MachineBasicBlock::iterator
294 freeStackSlotBefore(MachineBasicBlock::iterator I, unsigned FPRegNo);
296 /// Adjust the live registers to be the set in Mask.
297 void adjustLiveRegs(unsigned Mask, MachineBasicBlock::iterator I);
299 /// Shuffle the top FixCount stack entries such that FP reg FixStack[0] is
300 /// st(0), FP reg FixStack[1] is st(1) etc.
301 void shuffleStackTop(const unsigned char *FixStack, unsigned FixCount,
302 MachineBasicBlock::iterator I);
304 bool processBasicBlock(MachineFunction &MF, MachineBasicBlock &MBB);
306 void handleZeroArgFP(MachineBasicBlock::iterator &I);
307 void handleOneArgFP(MachineBasicBlock::iterator &I);
308 void handleOneArgFPRW(MachineBasicBlock::iterator &I);
309 void handleTwoArgFP(MachineBasicBlock::iterator &I);
310 void handleCompareFP(MachineBasicBlock::iterator &I);
311 void handleCondMovFP(MachineBasicBlock::iterator &I);
312 void handleSpecialFP(MachineBasicBlock::iterator &I);
314 // Check if a COPY instruction is using FP registers.
315 static bool isFPCopy(MachineInstr *MI) {
316 unsigned DstReg = MI->getOperand(0).getReg();
317 unsigned SrcReg = MI->getOperand(1).getReg();
319 return X86::RFP80RegClass.contains(DstReg) ||
320 X86::RFP80RegClass.contains(SrcReg);
326 FunctionPass *llvm::createX86FloatingPointStackifierPass() { return new FPS(); }
328 /// getFPReg - Return the X86::FPx register number for the specified operand.
329 /// For example, this returns 3 for X86::FP3.
330 static unsigned getFPReg(const MachineOperand &MO) {
331 assert(MO.isReg() && "Expected an FP register!");
332 unsigned Reg = MO.getReg();
333 assert(Reg >= X86::FP0 && Reg <= X86::FP6 && "Expected FP register!");
334 return Reg - X86::FP0;
337 /// runOnMachineFunction - Loop over all of the basic blocks, transforming FP
338 /// register references into FP stack references.
340 bool FPS::runOnMachineFunction(MachineFunction &MF) {
341 // We only need to run this pass if there are any FP registers used in this
342 // function. If it is all integer, there is nothing for us to do!
343 bool FPIsUsed = false;
345 assert(X86::FP6 == X86::FP0+6 && "Register enums aren't sorted right!");
346 for (unsigned i = 0; i <= 6; ++i)
347 if (MF.getRegInfo().isPhysRegUsed(X86::FP0+i)) {
353 if (!FPIsUsed) return false;
355 Bundles = &getAnalysis<EdgeBundles>();
356 TII = MF.getTarget().getInstrInfo();
358 // Prepare cross-MBB liveness.
363 // Process the function in depth first order so that we process at least one
364 // of the predecessors for every reachable block in the function.
365 SmallPtrSet<MachineBasicBlock*, 8> Processed;
366 MachineBasicBlock *Entry = MF.begin();
368 bool Changed = false;
369 for (df_ext_iterator<MachineBasicBlock*, SmallPtrSet<MachineBasicBlock*, 8> >
370 I = df_ext_begin(Entry, Processed), E = df_ext_end(Entry, Processed);
372 Changed |= processBasicBlock(MF, **I);
374 // Process any unreachable blocks in arbitrary order now.
375 if (MF.size() != Processed.size())
376 for (MachineFunction::iterator BB = MF.begin(), E = MF.end(); BB != E; ++BB)
377 if (Processed.insert(BB))
378 Changed |= processBasicBlock(MF, *BB);
385 /// bundleCFG - Scan all the basic blocks to determine consistent live-in and
386 /// live-out sets for the FP registers. Consistent means that the set of
387 /// registers live-out from a block is identical to the live-in set of all
388 /// successors. This is not enforced by the normal live-in lists since
389 /// registers may be implicitly defined, or not used by all successors.
390 void FPS::bundleCFG(MachineFunction &MF) {
391 assert(LiveBundles.empty() && "Stale data in LiveBundles");
392 LiveBundles.resize(Bundles->getNumBundles());
394 // Gather the actual live-in masks for all MBBs.
395 for (MachineFunction::iterator I = MF.begin(), E = MF.end(); I != E; ++I) {
396 MachineBasicBlock *MBB = I;
397 const unsigned Mask = calcLiveInMask(MBB);
400 // Update MBB ingoing bundle mask.
401 LiveBundles[Bundles->getBundle(MBB->getNumber(), false)].Mask |= Mask;
405 /// processBasicBlock - Loop over all of the instructions in the basic block,
406 /// transforming FP instructions into their stack form.
408 bool FPS::processBasicBlock(MachineFunction &MF, MachineBasicBlock &BB) {
409 bool Changed = false;
415 for (MachineBasicBlock::iterator I = BB.begin(); I != BB.end(); ++I) {
416 MachineInstr *MI = I;
417 uint64_t Flags = MI->getDesc().TSFlags;
419 unsigned FPInstClass = Flags & X86II::FPTypeMask;
420 if (MI->isInlineAsm())
421 FPInstClass = X86II::SpecialFP;
423 if (MI->isCopy() && isFPCopy(MI))
424 FPInstClass = X86II::SpecialFP;
426 if (MI->isImplicitDef() &&
427 X86::RFP80RegClass.contains(MI->getOperand(0).getReg()))
428 FPInstClass = X86II::SpecialFP;
430 if (FPInstClass == X86II::NotFP)
431 continue; // Efficiently ignore non-fp insts!
433 MachineInstr *PrevMI = 0;
435 PrevMI = std::prev(I);
437 ++NumFP; // Keep track of # of pseudo instrs
438 DEBUG(dbgs() << "\nFPInst:\t" << *MI);
440 // Get dead variables list now because the MI pointer may be deleted as part
442 SmallVector<unsigned, 8> DeadRegs;
443 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
444 const MachineOperand &MO = MI->getOperand(i);
445 if (MO.isReg() && MO.isDead())
446 DeadRegs.push_back(MO.getReg());
449 switch (FPInstClass) {
450 case X86II::ZeroArgFP: handleZeroArgFP(I); break;
451 case X86II::OneArgFP: handleOneArgFP(I); break; // fstp ST(0)
452 case X86II::OneArgFPRW: handleOneArgFPRW(I); break; // ST(0) = fsqrt(ST(0))
453 case X86II::TwoArgFP: handleTwoArgFP(I); break;
454 case X86II::CompareFP: handleCompareFP(I); break;
455 case X86II::CondMovFP: handleCondMovFP(I); break;
456 case X86II::SpecialFP: handleSpecialFP(I); break;
457 default: llvm_unreachable("Unknown FP Type!");
460 // Check to see if any of the values defined by this instruction are dead
461 // after definition. If so, pop them.
462 for (unsigned i = 0, e = DeadRegs.size(); i != e; ++i) {
463 unsigned Reg = DeadRegs[i];
464 if (Reg >= X86::FP0 && Reg <= X86::FP6) {
465 DEBUG(dbgs() << "Register FP#" << Reg-X86::FP0 << " is dead!\n");
466 freeStackSlotAfter(I, Reg-X86::FP0);
470 // Print out all of the instructions expanded to if -debug
472 MachineBasicBlock::iterator PrevI(PrevMI);
474 dbgs() << "Just deleted pseudo instruction\n";
476 MachineBasicBlock::iterator Start = I;
477 // Rewind to first instruction newly inserted.
478 while (Start != BB.begin() && std::prev(Start) != PrevI) --Start;
479 dbgs() << "Inserted instructions:\n\t";
480 Start->print(dbgs(), &MF.getTarget());
481 while (++Start != std::next(I)) {}
495 /// setupBlockStack - Use the live bundles to set up our model of the stack
496 /// to match predecessors' live out stack.
497 void FPS::setupBlockStack() {
498 DEBUG(dbgs() << "\nSetting up live-ins for BB#" << MBB->getNumber()
499 << " derived from " << MBB->getName() << ".\n");
501 // Get the live-in bundle for MBB.
502 const LiveBundle &Bundle =
503 LiveBundles[Bundles->getBundle(MBB->getNumber(), false)];
506 DEBUG(dbgs() << "Block has no FP live-ins.\n");
510 // Depth-first iteration should ensure that we always have an assigned stack.
511 assert(Bundle.isFixed() && "Reached block before any predecessors");
513 // Push the fixed live-in registers.
514 for (unsigned i = Bundle.FixCount; i > 0; --i) {
515 MBB->addLiveIn(X86::ST0+i-1);
516 DEBUG(dbgs() << "Live-in st(" << (i-1) << "): %FP"
517 << unsigned(Bundle.FixStack[i-1]) << '\n');
518 pushReg(Bundle.FixStack[i-1]);
521 // Kill off unwanted live-ins. This can happen with a critical edge.
522 // FIXME: We could keep these live registers around as zombies. They may need
523 // to be revived at the end of a short block. It might save a few instrs.
524 adjustLiveRegs(calcLiveInMask(MBB), MBB->begin());
528 /// finishBlockStack - Revive live-outs that are implicitly defined out of
529 /// MBB. Shuffle live registers to match the expected fixed stack of any
530 /// predecessors, and ensure that all predecessors are expecting the same
532 void FPS::finishBlockStack() {
533 // The RET handling below takes care of return blocks for us.
534 if (MBB->succ_empty())
537 DEBUG(dbgs() << "Setting up live-outs for BB#" << MBB->getNumber()
538 << " derived from " << MBB->getName() << ".\n");
540 // Get MBB's live-out bundle.
541 unsigned BundleIdx = Bundles->getBundle(MBB->getNumber(), true);
542 LiveBundle &Bundle = LiveBundles[BundleIdx];
544 // We may need to kill and define some registers to match successors.
545 // FIXME: This can probably be combined with the shuffle below.
546 MachineBasicBlock::iterator Term = MBB->getFirstTerminator();
547 adjustLiveRegs(Bundle.Mask, Term);
550 DEBUG(dbgs() << "No live-outs.\n");
554 // Has the stack order been fixed yet?
555 DEBUG(dbgs() << "LB#" << BundleIdx << ": ");
556 if (Bundle.isFixed()) {
557 DEBUG(dbgs() << "Shuffling stack to match.\n");
558 shuffleStackTop(Bundle.FixStack, Bundle.FixCount, Term);
560 // Not fixed yet, we get to choose.
561 DEBUG(dbgs() << "Fixing stack order now.\n");
562 Bundle.FixCount = StackTop;
563 for (unsigned i = 0; i < StackTop; ++i)
564 Bundle.FixStack[i] = getStackEntry(i);
569 //===----------------------------------------------------------------------===//
570 // Efficient Lookup Table Support
571 //===----------------------------------------------------------------------===//
577 bool operator<(const TableEntry &TE) const { return from < TE.from; }
578 friend bool operator<(const TableEntry &TE, unsigned V) {
581 friend bool LLVM_ATTRIBUTE_UNUSED operator<(unsigned V,
582 const TableEntry &TE) {
589 static bool TableIsSorted(const TableEntry *Table, unsigned NumEntries) {
590 for (unsigned i = 0; i != NumEntries-1; ++i)
591 if (!(Table[i] < Table[i+1])) return false;
596 static int Lookup(const TableEntry *Table, unsigned N, unsigned Opcode) {
597 const TableEntry *I = std::lower_bound(Table, Table+N, Opcode);
598 if (I != Table+N && I->from == Opcode)
604 #define ASSERT_SORTED(TABLE)
606 #define ASSERT_SORTED(TABLE) \
607 { static bool TABLE##Checked = false; \
608 if (!TABLE##Checked) { \
609 assert(TableIsSorted(TABLE, array_lengthof(TABLE)) && \
610 "All lookup tables must be sorted for efficient access!"); \
611 TABLE##Checked = true; \
616 //===----------------------------------------------------------------------===//
617 // Register File -> Register Stack Mapping Methods
618 //===----------------------------------------------------------------------===//
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.
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::COS_Fp32 , X86::COS_F },
667 { X86::COS_Fp64 , X86::COS_F },
668 { X86::COS_Fp80 , X86::COS_F },
669 { X86::DIVR_Fp32m , X86::DIVR_F32m },
670 { X86::DIVR_Fp64m , X86::DIVR_F64m },
671 { X86::DIVR_Fp64m32 , X86::DIVR_F32m },
672 { X86::DIVR_Fp80m32 , X86::DIVR_F32m },
673 { X86::DIVR_Fp80m64 , X86::DIVR_F64m },
674 { X86::DIVR_FpI16m32, X86::DIVR_FI16m},
675 { X86::DIVR_FpI16m64, X86::DIVR_FI16m},
676 { X86::DIVR_FpI16m80, X86::DIVR_FI16m},
677 { X86::DIVR_FpI32m32, X86::DIVR_FI32m},
678 { X86::DIVR_FpI32m64, X86::DIVR_FI32m},
679 { X86::DIVR_FpI32m80, X86::DIVR_FI32m},
680 { X86::DIV_Fp32m , X86::DIV_F32m },
681 { X86::DIV_Fp64m , X86::DIV_F64m },
682 { X86::DIV_Fp64m32 , X86::DIV_F32m },
683 { X86::DIV_Fp80m32 , X86::DIV_F32m },
684 { X86::DIV_Fp80m64 , X86::DIV_F64m },
685 { X86::DIV_FpI16m32 , X86::DIV_FI16m },
686 { X86::DIV_FpI16m64 , X86::DIV_FI16m },
687 { X86::DIV_FpI16m80 , X86::DIV_FI16m },
688 { X86::DIV_FpI32m32 , X86::DIV_FI32m },
689 { X86::DIV_FpI32m64 , X86::DIV_FI32m },
690 { X86::DIV_FpI32m80 , X86::DIV_FI32m },
691 { X86::ILD_Fp16m32 , X86::ILD_F16m },
692 { X86::ILD_Fp16m64 , X86::ILD_F16m },
693 { X86::ILD_Fp16m80 , X86::ILD_F16m },
694 { X86::ILD_Fp32m32 , X86::ILD_F32m },
695 { X86::ILD_Fp32m64 , X86::ILD_F32m },
696 { X86::ILD_Fp32m80 , X86::ILD_F32m },
697 { X86::ILD_Fp64m32 , X86::ILD_F64m },
698 { X86::ILD_Fp64m64 , X86::ILD_F64m },
699 { X86::ILD_Fp64m80 , X86::ILD_F64m },
700 { X86::ISTT_Fp16m32 , X86::ISTT_FP16m},
701 { X86::ISTT_Fp16m64 , X86::ISTT_FP16m},
702 { X86::ISTT_Fp16m80 , X86::ISTT_FP16m},
703 { X86::ISTT_Fp32m32 , X86::ISTT_FP32m},
704 { X86::ISTT_Fp32m64 , X86::ISTT_FP32m},
705 { X86::ISTT_Fp32m80 , X86::ISTT_FP32m},
706 { X86::ISTT_Fp64m32 , X86::ISTT_FP64m},
707 { X86::ISTT_Fp64m64 , X86::ISTT_FP64m},
708 { X86::ISTT_Fp64m80 , X86::ISTT_FP64m},
709 { X86::IST_Fp16m32 , X86::IST_F16m },
710 { X86::IST_Fp16m64 , X86::IST_F16m },
711 { X86::IST_Fp16m80 , X86::IST_F16m },
712 { X86::IST_Fp32m32 , X86::IST_F32m },
713 { X86::IST_Fp32m64 , X86::IST_F32m },
714 { X86::IST_Fp32m80 , X86::IST_F32m },
715 { X86::IST_Fp64m32 , X86::IST_FP64m },
716 { X86::IST_Fp64m64 , X86::IST_FP64m },
717 { X86::IST_Fp64m80 , X86::IST_FP64m },
718 { X86::LD_Fp032 , X86::LD_F0 },
719 { X86::LD_Fp064 , X86::LD_F0 },
720 { X86::LD_Fp080 , X86::LD_F0 },
721 { X86::LD_Fp132 , X86::LD_F1 },
722 { X86::LD_Fp164 , X86::LD_F1 },
723 { X86::LD_Fp180 , X86::LD_F1 },
724 { X86::LD_Fp32m , X86::LD_F32m },
725 { X86::LD_Fp32m64 , X86::LD_F32m },
726 { X86::LD_Fp32m80 , X86::LD_F32m },
727 { X86::LD_Fp64m , X86::LD_F64m },
728 { X86::LD_Fp64m80 , X86::LD_F64m },
729 { X86::LD_Fp80m , X86::LD_F80m },
730 { X86::MUL_Fp32m , X86::MUL_F32m },
731 { X86::MUL_Fp64m , X86::MUL_F64m },
732 { X86::MUL_Fp64m32 , X86::MUL_F32m },
733 { X86::MUL_Fp80m32 , X86::MUL_F32m },
734 { X86::MUL_Fp80m64 , X86::MUL_F64m },
735 { X86::MUL_FpI16m32 , X86::MUL_FI16m },
736 { X86::MUL_FpI16m64 , X86::MUL_FI16m },
737 { X86::MUL_FpI16m80 , X86::MUL_FI16m },
738 { X86::MUL_FpI32m32 , X86::MUL_FI32m },
739 { X86::MUL_FpI32m64 , X86::MUL_FI32m },
740 { X86::MUL_FpI32m80 , X86::MUL_FI32m },
741 { X86::SIN_Fp32 , X86::SIN_F },
742 { X86::SIN_Fp64 , X86::SIN_F },
743 { X86::SIN_Fp80 , X86::SIN_F },
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 },
786 static unsigned getConcreteOpcode(unsigned Opcode) {
787 ASSERT_SORTED(OpcodeTable);
788 int Opc = Lookup(OpcodeTable, array_lengthof(OpcodeTable), Opcode);
789 assert(Opc != -1 && "FP Stack instruction not in OpcodeTable!");
793 //===----------------------------------------------------------------------===//
795 //===----------------------------------------------------------------------===//
797 // PopTable - Sorted map of instructions to their popping version. The first
798 // element is an instruction, the second is the version which pops.
800 static const TableEntry PopTable[] = {
801 { X86::ADD_FrST0 , X86::ADD_FPrST0 },
803 { X86::DIVR_FrST0, X86::DIVR_FPrST0 },
804 { X86::DIV_FrST0 , X86::DIV_FPrST0 },
806 { X86::IST_F16m , X86::IST_FP16m },
807 { X86::IST_F32m , X86::IST_FP32m },
809 { X86::MUL_FrST0 , X86::MUL_FPrST0 },
811 { X86::ST_F32m , X86::ST_FP32m },
812 { X86::ST_F64m , X86::ST_FP64m },
813 { X86::ST_Frr , X86::ST_FPrr },
815 { X86::SUBR_FrST0, X86::SUBR_FPrST0 },
816 { X86::SUB_FrST0 , X86::SUB_FPrST0 },
818 { X86::UCOM_FIr , X86::UCOM_FIPr },
820 { X86::UCOM_FPr , X86::UCOM_FPPr },
821 { X86::UCOM_Fr , X86::UCOM_FPr },
824 /// popStackAfter - Pop the current value off of the top of the FP stack after
825 /// the specified instruction. This attempts to be sneaky and combine the pop
826 /// into the instruction itself if possible. The iterator is left pointing to
827 /// the last instruction, be it a new pop instruction inserted, or the old
828 /// instruction if it was modified in place.
830 void FPS::popStackAfter(MachineBasicBlock::iterator &I) {
831 MachineInstr* MI = I;
832 DebugLoc dl = MI->getDebugLoc();
833 ASSERT_SORTED(PopTable);
835 report_fatal_error("Cannot pop empty stack!");
836 RegMap[Stack[--StackTop]] = ~0; // Update state
838 // Check to see if there is a popping version of this instruction...
839 int Opcode = Lookup(PopTable, array_lengthof(PopTable), I->getOpcode());
841 I->setDesc(TII->get(Opcode));
842 if (Opcode == X86::UCOM_FPPr)
844 } else { // Insert an explicit pop
845 I = BuildMI(*MBB, ++I, dl, TII->get(X86::ST_FPrr)).addReg(X86::ST0);
849 /// freeStackSlotAfter - Free the specified register from the register stack, so
850 /// that it is no longer in a register. If the register is currently at the top
851 /// of the stack, we just pop the current instruction, otherwise we store the
852 /// current top-of-stack into the specified slot, then pop the top of stack.
853 void FPS::freeStackSlotAfter(MachineBasicBlock::iterator &I, unsigned FPRegNo) {
854 if (getStackEntry(0) == FPRegNo) { // already at the top of stack? easy.
859 // Otherwise, store the top of stack into the dead slot, killing the operand
860 // without having to add in an explicit xchg then pop.
862 I = freeStackSlotBefore(++I, FPRegNo);
865 /// freeStackSlotBefore - Free the specified register without trying any
867 MachineBasicBlock::iterator
868 FPS::freeStackSlotBefore(MachineBasicBlock::iterator I, unsigned FPRegNo) {
869 unsigned STReg = getSTReg(FPRegNo);
870 unsigned OldSlot = getSlot(FPRegNo);
871 unsigned TopReg = Stack[StackTop-1];
872 Stack[OldSlot] = TopReg;
873 RegMap[TopReg] = OldSlot;
874 RegMap[FPRegNo] = ~0;
875 Stack[--StackTop] = ~0;
876 return BuildMI(*MBB, I, DebugLoc(), TII->get(X86::ST_FPrr)).addReg(STReg);
879 /// adjustLiveRegs - Kill and revive registers such that exactly the FP
880 /// registers with a bit in Mask are live.
881 void FPS::adjustLiveRegs(unsigned Mask, MachineBasicBlock::iterator I) {
882 unsigned Defs = Mask;
884 for (unsigned i = 0; i < StackTop; ++i) {
885 unsigned RegNo = Stack[i];
886 if (!(Defs & (1 << RegNo)))
887 // This register is live, but we don't want it.
888 Kills |= (1 << RegNo);
890 // We don't need to imp-def this live register.
891 Defs &= ~(1 << RegNo);
893 assert((Kills & Defs) == 0 && "Register needs killing and def'ing?");
895 // Produce implicit-defs for free by using killed registers.
896 while (Kills && Defs) {
897 unsigned KReg = countTrailingZeros(Kills);
898 unsigned DReg = countTrailingZeros(Defs);
899 DEBUG(dbgs() << "Renaming %FP" << KReg << " as imp %FP" << DReg << "\n");
900 std::swap(Stack[getSlot(KReg)], Stack[getSlot(DReg)]);
901 std::swap(RegMap[KReg], RegMap[DReg]);
902 Kills &= ~(1 << KReg);
903 Defs &= ~(1 << DReg);
906 // Kill registers by popping.
907 if (Kills && I != MBB->begin()) {
908 MachineBasicBlock::iterator I2 = std::prev(I);
910 unsigned KReg = getStackEntry(0);
911 if (!(Kills & (1 << KReg)))
913 DEBUG(dbgs() << "Popping %FP" << KReg << "\n");
915 Kills &= ~(1 << KReg);
919 // Manually kill the rest.
921 unsigned KReg = countTrailingZeros(Kills);
922 DEBUG(dbgs() << "Killing %FP" << KReg << "\n");
923 freeStackSlotBefore(I, KReg);
924 Kills &= ~(1 << KReg);
927 // Load zeros for all the imp-defs.
929 unsigned DReg = countTrailingZeros(Defs);
930 DEBUG(dbgs() << "Defining %FP" << DReg << " as 0\n");
931 BuildMI(*MBB, I, DebugLoc(), TII->get(X86::LD_F0));
933 Defs &= ~(1 << DReg);
936 // Now we should have the correct registers live.
938 assert(StackTop == CountPopulation_32(Mask) && "Live count mismatch");
941 /// shuffleStackTop - emit fxch instructions before I to shuffle the top
942 /// FixCount entries into the order given by FixStack.
943 /// FIXME: Is there a better algorithm than insertion sort?
944 void FPS::shuffleStackTop(const unsigned char *FixStack,
946 MachineBasicBlock::iterator I) {
947 // Move items into place, starting from the desired stack bottom.
949 // Old register at position FixCount.
950 unsigned OldReg = getStackEntry(FixCount);
951 // Desired register at position FixCount.
952 unsigned Reg = FixStack[FixCount];
955 // (Reg st0) (OldReg st0) = (Reg OldReg st0)
958 moveToTop(OldReg, I);
964 //===----------------------------------------------------------------------===//
965 // Instruction transformation implementation
966 //===----------------------------------------------------------------------===//
968 /// handleZeroArgFP - ST(0) = fld0 ST(0) = flds <mem>
970 void FPS::handleZeroArgFP(MachineBasicBlock::iterator &I) {
971 MachineInstr *MI = I;
972 unsigned DestReg = getFPReg(MI->getOperand(0));
974 // Change from the pseudo instruction to the concrete instruction.
975 MI->RemoveOperand(0); // Remove the explicit ST(0) operand
976 MI->setDesc(TII->get(getConcreteOpcode(MI->getOpcode())));
978 // Result gets pushed on the stack.
982 /// handleOneArgFP - fst <mem>, ST(0)
984 void FPS::handleOneArgFP(MachineBasicBlock::iterator &I) {
985 MachineInstr *MI = I;
986 unsigned NumOps = MI->getDesc().getNumOperands();
987 assert((NumOps == X86::AddrNumOperands + 1 || NumOps == 1) &&
988 "Can only handle fst* & ftst instructions!");
990 // Is this the last use of the source register?
991 unsigned Reg = getFPReg(MI->getOperand(NumOps-1));
992 bool KillsSrc = MI->killsRegister(X86::FP0+Reg);
995 duplicatePendingSTBeforeKill(Reg, I);
997 // FISTP64m is strange because there isn't a non-popping versions.
998 // If we have one _and_ we don't want to pop the operand, duplicate the value
999 // on the stack instead of moving it. This ensure that popping the value is
1001 // Ditto FISTTP16m, FISTTP32m, FISTTP64m, ST_FpP80m.
1004 (MI->getOpcode() == X86::IST_Fp64m32 ||
1005 MI->getOpcode() == X86::ISTT_Fp16m32 ||
1006 MI->getOpcode() == X86::ISTT_Fp32m32 ||
1007 MI->getOpcode() == X86::ISTT_Fp64m32 ||
1008 MI->getOpcode() == X86::IST_Fp64m64 ||
1009 MI->getOpcode() == X86::ISTT_Fp16m64 ||
1010 MI->getOpcode() == X86::ISTT_Fp32m64 ||
1011 MI->getOpcode() == X86::ISTT_Fp64m64 ||
1012 MI->getOpcode() == X86::IST_Fp64m80 ||
1013 MI->getOpcode() == X86::ISTT_Fp16m80 ||
1014 MI->getOpcode() == X86::ISTT_Fp32m80 ||
1015 MI->getOpcode() == X86::ISTT_Fp64m80 ||
1016 MI->getOpcode() == X86::ST_FpP80m)) {
1017 duplicateToTop(Reg, getScratchReg(), I);
1019 moveToTop(Reg, I); // Move to the top of the stack...
1022 // Convert from the pseudo instruction to the concrete instruction.
1023 MI->RemoveOperand(NumOps-1); // Remove explicit ST(0) operand
1024 MI->setDesc(TII->get(getConcreteOpcode(MI->getOpcode())));
1026 if (MI->getOpcode() == X86::IST_FP64m ||
1027 MI->getOpcode() == X86::ISTT_FP16m ||
1028 MI->getOpcode() == X86::ISTT_FP32m ||
1029 MI->getOpcode() == X86::ISTT_FP64m ||
1030 MI->getOpcode() == X86::ST_FP80m) {
1032 report_fatal_error("Stack empty??");
1034 } else if (KillsSrc) { // Last use of operand?
1040 /// handleOneArgFPRW: Handle instructions that read from the top of stack and
1041 /// replace the value with a newly computed value. These instructions may have
1042 /// non-fp operands after their FP operands.
1046 /// R1 = fadd R2, [mem]
1048 void FPS::handleOneArgFPRW(MachineBasicBlock::iterator &I) {
1049 MachineInstr *MI = I;
1051 unsigned NumOps = MI->getDesc().getNumOperands();
1052 assert(NumOps >= 2 && "FPRW instructions must have 2 ops!!");
1055 // Is this the last use of the source register?
1056 unsigned Reg = getFPReg(MI->getOperand(1));
1057 bool KillsSrc = MI->killsRegister(X86::FP0+Reg);
1060 duplicatePendingSTBeforeKill(Reg, I);
1061 // If this is the last use of the source register, just make sure it's on
1062 // the top of the stack.
1065 report_fatal_error("Stack cannot be empty!");
1067 pushReg(getFPReg(MI->getOperand(0)));
1069 // If this is not the last use of the source register, _copy_ it to the top
1071 duplicateToTop(Reg, getFPReg(MI->getOperand(0)), I);
1074 // Change from the pseudo instruction to the concrete instruction.
1075 MI->RemoveOperand(1); // Drop the source operand.
1076 MI->RemoveOperand(0); // Drop the destination operand.
1077 MI->setDesc(TII->get(getConcreteOpcode(MI->getOpcode())));
1081 //===----------------------------------------------------------------------===//
1082 // Define tables of various ways to map pseudo instructions
1085 // ForwardST0Table - Map: A = B op C into: ST(0) = ST(0) op ST(i)
1086 static const TableEntry ForwardST0Table[] = {
1087 { X86::ADD_Fp32 , X86::ADD_FST0r },
1088 { X86::ADD_Fp64 , X86::ADD_FST0r },
1089 { X86::ADD_Fp80 , X86::ADD_FST0r },
1090 { X86::DIV_Fp32 , X86::DIV_FST0r },
1091 { X86::DIV_Fp64 , X86::DIV_FST0r },
1092 { X86::DIV_Fp80 , X86::DIV_FST0r },
1093 { X86::MUL_Fp32 , X86::MUL_FST0r },
1094 { X86::MUL_Fp64 , X86::MUL_FST0r },
1095 { X86::MUL_Fp80 , X86::MUL_FST0r },
1096 { X86::SUB_Fp32 , X86::SUB_FST0r },
1097 { X86::SUB_Fp64 , X86::SUB_FST0r },
1098 { X86::SUB_Fp80 , X86::SUB_FST0r },
1101 // ReverseST0Table - Map: A = B op C into: ST(0) = ST(i) op ST(0)
1102 static const TableEntry ReverseST0Table[] = {
1103 { X86::ADD_Fp32 , X86::ADD_FST0r }, // commutative
1104 { X86::ADD_Fp64 , X86::ADD_FST0r }, // commutative
1105 { X86::ADD_Fp80 , X86::ADD_FST0r }, // commutative
1106 { X86::DIV_Fp32 , X86::DIVR_FST0r },
1107 { X86::DIV_Fp64 , X86::DIVR_FST0r },
1108 { X86::DIV_Fp80 , X86::DIVR_FST0r },
1109 { X86::MUL_Fp32 , X86::MUL_FST0r }, // commutative
1110 { X86::MUL_Fp64 , X86::MUL_FST0r }, // commutative
1111 { X86::MUL_Fp80 , X86::MUL_FST0r }, // commutative
1112 { X86::SUB_Fp32 , X86::SUBR_FST0r },
1113 { X86::SUB_Fp64 , X86::SUBR_FST0r },
1114 { X86::SUB_Fp80 , X86::SUBR_FST0r },
1117 // ForwardSTiTable - Map: A = B op C into: ST(i) = ST(0) op ST(i)
1118 static const TableEntry ForwardSTiTable[] = {
1119 { X86::ADD_Fp32 , X86::ADD_FrST0 }, // commutative
1120 { X86::ADD_Fp64 , X86::ADD_FrST0 }, // commutative
1121 { X86::ADD_Fp80 , X86::ADD_FrST0 }, // commutative
1122 { X86::DIV_Fp32 , X86::DIVR_FrST0 },
1123 { X86::DIV_Fp64 , X86::DIVR_FrST0 },
1124 { X86::DIV_Fp80 , X86::DIVR_FrST0 },
1125 { X86::MUL_Fp32 , X86::MUL_FrST0 }, // commutative
1126 { X86::MUL_Fp64 , X86::MUL_FrST0 }, // commutative
1127 { X86::MUL_Fp80 , X86::MUL_FrST0 }, // commutative
1128 { X86::SUB_Fp32 , X86::SUBR_FrST0 },
1129 { X86::SUB_Fp64 , X86::SUBR_FrST0 },
1130 { X86::SUB_Fp80 , X86::SUBR_FrST0 },
1133 // ReverseSTiTable - Map: A = B op C into: ST(i) = ST(i) op ST(0)
1134 static const TableEntry ReverseSTiTable[] = {
1135 { X86::ADD_Fp32 , X86::ADD_FrST0 },
1136 { X86::ADD_Fp64 , X86::ADD_FrST0 },
1137 { X86::ADD_Fp80 , X86::ADD_FrST0 },
1138 { X86::DIV_Fp32 , X86::DIV_FrST0 },
1139 { X86::DIV_Fp64 , X86::DIV_FrST0 },
1140 { X86::DIV_Fp80 , X86::DIV_FrST0 },
1141 { X86::MUL_Fp32 , X86::MUL_FrST0 },
1142 { X86::MUL_Fp64 , X86::MUL_FrST0 },
1143 { X86::MUL_Fp80 , X86::MUL_FrST0 },
1144 { X86::SUB_Fp32 , X86::SUB_FrST0 },
1145 { X86::SUB_Fp64 , X86::SUB_FrST0 },
1146 { X86::SUB_Fp80 , X86::SUB_FrST0 },
1150 /// handleTwoArgFP - Handle instructions like FADD and friends which are virtual
1151 /// instructions which need to be simplified and possibly transformed.
1153 /// Result: ST(0) = fsub ST(0), ST(i)
1154 /// ST(i) = fsub ST(0), ST(i)
1155 /// ST(0) = fsubr ST(0), ST(i)
1156 /// ST(i) = fsubr ST(0), ST(i)
1158 void FPS::handleTwoArgFP(MachineBasicBlock::iterator &I) {
1159 ASSERT_SORTED(ForwardST0Table); ASSERT_SORTED(ReverseST0Table);
1160 ASSERT_SORTED(ForwardSTiTable); ASSERT_SORTED(ReverseSTiTable);
1161 MachineInstr *MI = I;
1163 unsigned NumOperands = MI->getDesc().getNumOperands();
1164 assert(NumOperands == 3 && "Illegal TwoArgFP instruction!");
1165 unsigned Dest = getFPReg(MI->getOperand(0));
1166 unsigned Op0 = getFPReg(MI->getOperand(NumOperands-2));
1167 unsigned Op1 = getFPReg(MI->getOperand(NumOperands-1));
1168 bool KillsOp0 = MI->killsRegister(X86::FP0+Op0);
1169 bool KillsOp1 = MI->killsRegister(X86::FP0+Op1);
1170 DebugLoc dl = MI->getDebugLoc();
1172 unsigned TOS = getStackEntry(0);
1174 // One of our operands must be on the top of the stack. If neither is yet, we
1175 // need to move one.
1176 if (Op0 != TOS && Op1 != TOS) { // No operand at TOS?
1177 // We can choose to move either operand to the top of the stack. If one of
1178 // the operands is killed by this instruction, we want that one so that we
1179 // can update right on top of the old version.
1181 moveToTop(Op0, I); // Move dead operand to TOS.
1183 } else if (KillsOp1) {
1187 // All of the operands are live after this instruction executes, so we
1188 // cannot update on top of any operand. Because of this, we must
1189 // duplicate one of the stack elements to the top. It doesn't matter
1190 // which one we pick.
1192 duplicateToTop(Op0, Dest, I);
1196 } else if (!KillsOp0 && !KillsOp1) {
1197 // If we DO have one of our operands at the top of the stack, but we don't
1198 // have a dead operand, we must duplicate one of the operands to a new slot
1200 duplicateToTop(Op0, Dest, I);
1205 // Now we know that one of our operands is on the top of the stack, and at
1206 // least one of our operands is killed by this instruction.
1207 assert((TOS == Op0 || TOS == Op1) && (KillsOp0 || KillsOp1) &&
1208 "Stack conditions not set up right!");
1210 // We decide which form to use based on what is on the top of the stack, and
1211 // which operand is killed by this instruction.
1212 const TableEntry *InstTable;
1213 bool isForward = TOS == Op0;
1214 bool updateST0 = (TOS == Op0 && !KillsOp1) || (TOS == Op1 && !KillsOp0);
1217 InstTable = ForwardST0Table;
1219 InstTable = ReverseST0Table;
1222 InstTable = ForwardSTiTable;
1224 InstTable = ReverseSTiTable;
1227 int Opcode = Lookup(InstTable, array_lengthof(ForwardST0Table),
1229 assert(Opcode != -1 && "Unknown TwoArgFP pseudo instruction!");
1231 // NotTOS - The register which is not on the top of stack...
1232 unsigned NotTOS = (TOS == Op0) ? Op1 : Op0;
1234 // Replace the old instruction with a new instruction
1236 I = BuildMI(*MBB, I, dl, TII->get(Opcode)).addReg(getSTReg(NotTOS));
1238 // If both operands are killed, pop one off of the stack in addition to
1239 // overwriting the other one.
1240 if (KillsOp0 && KillsOp1 && Op0 != Op1) {
1241 assert(!updateST0 && "Should have updated other operand!");
1242 popStackAfter(I); // Pop the top of stack
1245 // Update stack information so that we know the destination register is now on
1247 unsigned UpdatedSlot = getSlot(updateST0 ? TOS : NotTOS);
1248 assert(UpdatedSlot < StackTop && Dest < 7);
1249 Stack[UpdatedSlot] = Dest;
1250 RegMap[Dest] = UpdatedSlot;
1251 MBB->getParent()->DeleteMachineInstr(MI); // Remove the old instruction
1254 /// handleCompareFP - Handle FUCOM and FUCOMI instructions, which have two FP
1255 /// register arguments and no explicit destinations.
1257 void FPS::handleCompareFP(MachineBasicBlock::iterator &I) {
1258 ASSERT_SORTED(ForwardST0Table); ASSERT_SORTED(ReverseST0Table);
1259 ASSERT_SORTED(ForwardSTiTable); ASSERT_SORTED(ReverseSTiTable);
1260 MachineInstr *MI = I;
1262 unsigned NumOperands = MI->getDesc().getNumOperands();
1263 assert(NumOperands == 2 && "Illegal FUCOM* instruction!");
1264 unsigned Op0 = getFPReg(MI->getOperand(NumOperands-2));
1265 unsigned Op1 = getFPReg(MI->getOperand(NumOperands-1));
1266 bool KillsOp0 = MI->killsRegister(X86::FP0+Op0);
1267 bool KillsOp1 = MI->killsRegister(X86::FP0+Op1);
1269 // Make sure the first operand is on the top of stack, the other one can be
1273 // Change from the pseudo instruction to the concrete instruction.
1274 MI->getOperand(0).setReg(getSTReg(Op1));
1275 MI->RemoveOperand(1);
1276 MI->setDesc(TII->get(getConcreteOpcode(MI->getOpcode())));
1278 // If any of the operands are killed by this instruction, free them.
1279 if (KillsOp0) freeStackSlotAfter(I, Op0);
1280 if (KillsOp1 && Op0 != Op1) freeStackSlotAfter(I, Op1);
1283 /// handleCondMovFP - Handle two address conditional move instructions. These
1284 /// instructions move a st(i) register to st(0) iff a condition is true. These
1285 /// instructions require that the first operand is at the top of the stack, but
1286 /// otherwise don't modify the stack at all.
1287 void FPS::handleCondMovFP(MachineBasicBlock::iterator &I) {
1288 MachineInstr *MI = I;
1290 unsigned Op0 = getFPReg(MI->getOperand(0));
1291 unsigned Op1 = getFPReg(MI->getOperand(2));
1292 bool KillsOp1 = MI->killsRegister(X86::FP0+Op1);
1294 // The first operand *must* be on the top of the stack.
1297 // Change the second operand to the stack register that the operand is in.
1298 // Change from the pseudo instruction to the concrete instruction.
1299 MI->RemoveOperand(0);
1300 MI->RemoveOperand(1);
1301 MI->getOperand(0).setReg(getSTReg(Op1));
1302 MI->setDesc(TII->get(getConcreteOpcode(MI->getOpcode())));
1304 // If we kill the second operand, make sure to pop it from the stack.
1305 if (Op0 != Op1 && KillsOp1) {
1306 // Get this value off of the register stack.
1307 freeStackSlotAfter(I, Op1);
1312 /// handleSpecialFP - Handle special instructions which behave unlike other
1313 /// floating point instructions. This is primarily intended for use by pseudo
1316 void FPS::handleSpecialFP(MachineBasicBlock::iterator &I) {
1317 MachineInstr *MI = I;
1318 switch (MI->getOpcode()) {
1319 default: llvm_unreachable("Unknown SpecialFP instruction!");
1320 case TargetOpcode::COPY: {
1321 // We handle three kinds of copies: FP <- FP, FP <- ST, and ST <- FP.
1322 const MachineOperand &MO1 = MI->getOperand(1);
1323 const MachineOperand &MO0 = MI->getOperand(0);
1324 unsigned DstST = MO0.getReg() - X86::ST0;
1325 unsigned SrcST = MO1.getReg() - X86::ST0;
1326 bool KillsSrc = MI->killsRegister(MO1.getReg());
1328 // ST = COPY FP. Set up a pending ST register.
1330 unsigned SrcFP = getFPReg(MO1);
1331 assert(isLive(SrcFP) && "Cannot copy dead register");
1332 assert(!MO0.isDead() && "Cannot copy to dead ST register");
1334 // Unallocated STs are marked as the nonexistent FP255.
1335 while (NumPendingSTs <= DstST)
1336 PendingST[NumPendingSTs++] = NumFPRegs;
1338 // STi could still be live from a previous inline asm.
1339 if (isScratchReg(PendingST[DstST])) {
1340 DEBUG(dbgs() << "Clobbering old ST in FP" << unsigned(PendingST[DstST])
1342 freeStackSlotBefore(MI, PendingST[DstST]);
1345 // When the source is killed, allocate a scratch FP register.
1347 duplicatePendingSTBeforeKill(SrcFP, I);
1348 unsigned Slot = getSlot(SrcFP);
1349 unsigned SR = getScratchReg();
1350 PendingST[DstST] = SR;
1354 PendingST[DstST] = SrcFP;
1358 // FP = COPY ST. Extract fixed stack value.
1359 // Any instruction defining ST registers must have assigned them to a
1360 // scratch register.
1362 unsigned DstFP = getFPReg(MO0);
1363 assert(!isLive(DstFP) && "Cannot copy ST to live FP register");
1364 assert(NumPendingSTs > SrcST && "Cannot copy from dead ST register");
1365 unsigned SrcFP = PendingST[SrcST];
1366 assert(isScratchReg(SrcFP) && "Expected ST in a scratch register");
1367 assert(isLive(SrcFP) && "Scratch holding ST is dead");
1369 // DstFP steals the stack slot from SrcFP.
1370 unsigned Slot = getSlot(SrcFP);
1371 Stack[Slot] = DstFP;
1372 RegMap[DstFP] = Slot;
1374 // Always treat the ST as killed.
1375 PendingST[SrcST] = NumFPRegs;
1376 while (NumPendingSTs && PendingST[NumPendingSTs - 1] == NumFPRegs)
1382 unsigned DstFP = getFPReg(MO0);
1383 unsigned SrcFP = getFPReg(MO1);
1384 assert(isLive(SrcFP) && "Cannot copy dead register");
1386 // If the input operand is killed, we can just change the owner of the
1387 // incoming stack slot into the result.
1388 unsigned Slot = getSlot(SrcFP);
1389 Stack[Slot] = DstFP;
1390 RegMap[DstFP] = Slot;
1392 // For COPY we just duplicate the specified value to a new stack slot.
1393 // This could be made better, but would require substantial changes.
1394 duplicateToTop(SrcFP, DstFP, I);
1399 case TargetOpcode::IMPLICIT_DEF: {
1400 // All FP registers must be explicitly defined, so load a 0 instead.
1401 unsigned Reg = MI->getOperand(0).getReg() - X86::FP0;
1402 DEBUG(dbgs() << "Emitting LD_F0 for implicit FP" << Reg << '\n');
1403 BuildMI(*MBB, I, MI->getDebugLoc(), TII->get(X86::LD_F0));
1408 case X86::FpPOP_RETVAL: {
1409 // The FpPOP_RETVAL instruction is used after calls that return a value on
1410 // the floating point stack. We cannot model this with ST defs since CALL
1411 // instructions have fixed clobber lists. This instruction is interpreted
1412 // to mean that there is one more live register on the stack than we
1415 // This means that StackTop does not match the hardware stack between a
1416 // call and the FpPOP_RETVAL instructions. We do tolerate FP instructions
1417 // between CALL and FpPOP_RETVAL as long as they don't overflow the
1419 unsigned DstFP = getFPReg(MI->getOperand(0));
1421 // Move existing stack elements up to reflect reality.
1422 assert(StackTop < 8 && "Stack overflowed before FpPOP_RETVAL");
1424 std::copy_backward(Stack, Stack + StackTop, Stack + StackTop + 1);
1425 for (unsigned i = 0; i != NumFPRegs; ++i)
1430 // DstFP is the new bottom of the stack.
1434 // DstFP will be killed by processBasicBlock if this was a dead def.
1438 case TargetOpcode::INLINEASM: {
1439 // The inline asm MachineInstr currently only *uses* FP registers for the
1440 // 'f' constraint. These should be turned into the current ST(x) register
1441 // in the machine instr.
1443 // There are special rules for x87 inline assembly. The compiler must know
1444 // exactly how many registers are popped and pushed implicitly by the asm.
1445 // Otherwise it is not possible to restore the stack state after the inline
1448 // There are 3 kinds of input operands:
1450 // 1. Popped inputs. These must appear at the stack top in ST0-STn. A
1451 // popped input operand must be in a fixed stack slot, and it is either
1452 // tied to an output operand, or in the clobber list. The MI has ST use
1453 // and def operands for these inputs.
1455 // 2. Fixed inputs. These inputs appear in fixed stack slots, but are
1456 // preserved by the inline asm. The fixed stack slots must be STn-STm
1457 // following the popped inputs. A fixed input operand cannot be tied to
1458 // an output or appear in the clobber list. The MI has ST use operands
1459 // and no defs for these inputs.
1461 // 3. Preserved inputs. These inputs use the "f" constraint which is
1462 // represented as an FP register. The inline asm won't change these
1465 // Outputs must be in ST registers, FP outputs are not allowed. Clobbered
1466 // registers do not count as output operands. The inline asm changes the
1467 // stack as if it popped all the popped inputs and then pushed all the
1470 // Scan the assembly for ST registers used, defined and clobbered. We can
1471 // only tell clobbers from defs by looking at the asm descriptor.
1472 unsigned STUses = 0, STDefs = 0, STClobbers = 0, STDeadDefs = 0;
1473 unsigned NumOps = 0;
1474 for (unsigned i = InlineAsm::MIOp_FirstOperand, e = MI->getNumOperands();
1475 i != e && MI->getOperand(i).isImm(); i += 1 + NumOps) {
1476 unsigned Flags = MI->getOperand(i).getImm();
1477 NumOps = InlineAsm::getNumOperandRegisters(Flags);
1480 const MachineOperand &MO = MI->getOperand(i + 1);
1483 unsigned STReg = MO.getReg() - X86::ST0;
1487 switch (InlineAsm::getKind(Flags)) {
1488 case InlineAsm::Kind_RegUse:
1489 STUses |= (1u << STReg);
1491 case InlineAsm::Kind_RegDef:
1492 case InlineAsm::Kind_RegDefEarlyClobber:
1493 STDefs |= (1u << STReg);
1495 STDeadDefs |= (1u << STReg);
1497 case InlineAsm::Kind_Clobber:
1498 STClobbers |= (1u << STReg);
1505 if (STUses && !isMask_32(STUses))
1506 MI->emitError("fixed input regs must be last on the x87 stack");
1507 unsigned NumSTUses = CountTrailingOnes_32(STUses);
1509 // Defs must be contiguous from the stack top. ST0-STn.
1510 if (STDefs && !isMask_32(STDefs)) {
1511 MI->emitError("output regs must be last on the x87 stack");
1512 STDefs = NextPowerOf2(STDefs) - 1;
1514 unsigned NumSTDefs = CountTrailingOnes_32(STDefs);
1516 // So must the clobbered stack slots. ST0-STm, m >= n.
1517 if (STClobbers && !isMask_32(STDefs | STClobbers))
1518 MI->emitError("clobbers must be last on the x87 stack");
1520 // Popped inputs are the ones that are also clobbered or defined.
1521 unsigned STPopped = STUses & (STDefs | STClobbers);
1522 if (STPopped && !isMask_32(STPopped))
1523 MI->emitError("implicitly popped regs must be last on the x87 stack");
1524 unsigned NumSTPopped = CountTrailingOnes_32(STPopped);
1526 DEBUG(dbgs() << "Asm uses " << NumSTUses << " fixed regs, pops "
1527 << NumSTPopped << ", and defines " << NumSTDefs << " regs.\n");
1529 // Scan the instruction for FP uses corresponding to "f" constraints.
1530 // Collect FP registers to kill afer the instruction.
1531 // Always kill all the scratch regs.
1532 unsigned FPKills = ((1u << NumFPRegs) - 1) & ~0xff;
1533 unsigned FPUsed = 0;
1534 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
1535 MachineOperand &Op = MI->getOperand(i);
1536 if (!Op.isReg() || Op.getReg() < X86::FP0 || Op.getReg() > X86::FP6)
1539 MI->emitError("illegal \"f\" output constraint");
1540 unsigned FPReg = getFPReg(Op);
1541 FPUsed |= 1U << FPReg;
1543 // If we kill this operand, make sure to pop it from the stack after the
1544 // asm. We just remember it for now, and pop them all off at the end in
1547 FPKills |= 1U << FPReg;
1550 // The popped inputs will be killed by the instruction, so duplicate them
1551 // if the FP register needs to be live after the instruction, or if it is
1552 // used in the instruction itself. We effectively treat the popped inputs
1553 // as early clobbers.
1554 for (unsigned i = 0; i < NumSTPopped; ++i) {
1555 if ((FPKills & ~FPUsed) & (1u << PendingST[i]))
1557 unsigned SR = getScratchReg();
1558 duplicateToTop(PendingST[i], SR, I);
1559 DEBUG(dbgs() << "Duplicating ST" << i << " in FP"
1560 << unsigned(PendingST[i]) << " to avoid clobbering it.\n");
1564 // Make sure we have a unique live register for every fixed use. Some of
1565 // them could be undef uses, and we need to emit LD_F0 instructions.
1566 for (unsigned i = 0; i < NumSTUses; ++i) {
1567 if (i < NumPendingSTs && PendingST[i] < NumFPRegs) {
1568 // Check for shared assignments.
1569 for (unsigned j = 0; j < i; ++j) {
1570 if (PendingST[j] != PendingST[i])
1572 // STi and STj are inn the same register, create a copy.
1573 unsigned SR = getScratchReg();
1574 duplicateToTop(PendingST[i], SR, I);
1575 DEBUG(dbgs() << "Duplicating ST" << i << " in FP"
1576 << unsigned(PendingST[i])
1577 << " to avoid collision with ST" << j << '\n');
1582 unsigned SR = getScratchReg();
1583 DEBUG(dbgs() << "Emitting LD_F0 for ST" << i << " in FP" << SR << '\n');
1584 BuildMI(*MBB, I, MI->getDebugLoc(), TII->get(X86::LD_F0));
1587 if (NumPendingSTs == i)
1590 assert(NumPendingSTs >= NumSTUses && "Fixed registers should be assigned");
1592 // Now we can rearrange the live registers to match what was requested.
1593 shuffleStackTop(PendingST, NumPendingSTs, I);
1594 DEBUG({dbgs() << "Before asm: "; dumpStack();});
1596 // With the stack layout fixed, rewrite the FP registers.
1597 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
1598 MachineOperand &Op = MI->getOperand(i);
1599 if (!Op.isReg() || Op.getReg() < X86::FP0 || Op.getReg() > X86::FP6)
1601 unsigned FPReg = getFPReg(Op);
1602 Op.setReg(getSTReg(FPReg));
1605 // Simulate the inline asm popping its inputs and pushing its outputs.
1606 StackTop -= NumSTPopped;
1608 // Hold the fixed output registers in scratch FP registers. They will be
1609 // transferred to real FP registers by copies.
1611 for (unsigned i = 0; i < NumSTDefs; ++i) {
1612 unsigned SR = getScratchReg();
1614 FPKills &= ~(1u << SR);
1616 for (unsigned i = 0; i < NumSTDefs; ++i)
1617 PendingST[NumPendingSTs++] = getStackEntry(i);
1618 DEBUG({dbgs() << "After asm: "; dumpStack();});
1620 // If any of the ST defs were dead, pop them immediately. Our caller only
1621 // handles dead FP defs.
1622 MachineBasicBlock::iterator InsertPt = MI;
1623 for (unsigned i = 0; STDefs & (1u << i); ++i) {
1624 if (!(STDeadDefs & (1u << i)))
1626 freeStackSlotAfter(InsertPt, PendingST[i]);
1627 PendingST[i] = NumFPRegs;
1629 while (NumPendingSTs && PendingST[NumPendingSTs - 1] == NumFPRegs)
1632 // If this asm kills any FP registers (is the last use of them) we must
1633 // explicitly emit pop instructions for them. Do this now after the asm has
1634 // executed so that the ST(x) numbers are not off (which would happen if we
1635 // did this inline with operand rewriting).
1637 // Note: this might be a non-optimal pop sequence. We might be able to do
1638 // better by trying to pop in stack order or something.
1640 unsigned FPReg = countTrailingZeros(FPKills);
1642 freeStackSlotAfter(InsertPt, FPReg);
1643 FPKills &= ~(1U << FPReg);
1645 // Don't delete the inline asm!
1649 case X86::WIN_FTOL_32:
1650 case X86::WIN_FTOL_64: {
1651 // Push the operand into ST0.
1652 MachineOperand &Op = MI->getOperand(0);
1653 assert(Op.isUse() && Op.isReg() &&
1654 Op.getReg() >= X86::FP0 && Op.getReg() <= X86::FP6);
1655 unsigned FPReg = getFPReg(Op);
1657 moveToTop(FPReg, I);
1659 duplicateToTop(FPReg, FPReg, I);
1661 // Emit the call. This will pop the operand.
1662 BuildMI(*MBB, I, MI->getDebugLoc(), TII->get(X86::CALLpcrel32))
1663 .addExternalSymbol("_ftol2")
1664 .addReg(X86::ST0, RegState::ImplicitKill)
1665 .addReg(X86::ECX, RegState::ImplicitDefine)
1666 .addReg(X86::EAX, RegState::Define | RegState::Implicit)
1667 .addReg(X86::EDX, RegState::Define | RegState::Implicit)
1668 .addReg(X86::EFLAGS, RegState::Define | RegState::Implicit);
1678 // If RET has an FP register use operand, pass the first one in ST(0) and
1679 // the second one in ST(1).
1681 // Find the register operands.
1682 unsigned FirstFPRegOp = ~0U, SecondFPRegOp = ~0U;
1683 unsigned LiveMask = 0;
1685 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
1686 MachineOperand &Op = MI->getOperand(i);
1687 if (!Op.isReg() || Op.getReg() < X86::FP0 || Op.getReg() > X86::FP6)
1689 // FP Register uses must be kills unless there are two uses of the same
1690 // register, in which case only one will be a kill.
1691 assert(Op.isUse() &&
1692 (Op.isKill() || // Marked kill.
1693 getFPReg(Op) == FirstFPRegOp || // Second instance.
1694 MI->killsRegister(Op.getReg())) && // Later use is marked kill.
1695 "Ret only defs operands, and values aren't live beyond it");
1697 if (FirstFPRegOp == ~0U)
1698 FirstFPRegOp = getFPReg(Op);
1700 assert(SecondFPRegOp == ~0U && "More than two fp operands!");
1701 SecondFPRegOp = getFPReg(Op);
1703 LiveMask |= (1 << getFPReg(Op));
1705 // Remove the operand so that later passes don't see it.
1706 MI->RemoveOperand(i);
1710 // We may have been carrying spurious live-ins, so make sure only the returned
1711 // registers are left live.
1712 adjustLiveRegs(LiveMask, MI);
1713 if (!LiveMask) return; // Quick check to see if any are possible.
1715 // There are only four possibilities here:
1716 // 1) we are returning a single FP value. In this case, it has to be in
1717 // ST(0) already, so just declare success by removing the value from the
1719 if (SecondFPRegOp == ~0U) {
1720 // Assert that the top of stack contains the right FP register.
1721 assert(StackTop == 1 && FirstFPRegOp == getStackEntry(0) &&
1722 "Top of stack not the right register for RET!");
1724 // Ok, everything is good, mark the value as not being on the stack
1725 // anymore so that our assertion about the stack being empty at end of
1726 // block doesn't fire.
1731 // Otherwise, we are returning two values:
1732 // 2) If returning the same value for both, we only have one thing in the FP
1733 // stack. Consider: RET FP1, FP1
1734 if (StackTop == 1) {
1735 assert(FirstFPRegOp == SecondFPRegOp && FirstFPRegOp == getStackEntry(0)&&
1736 "Stack misconfiguration for RET!");
1738 // Duplicate the TOS so that we return it twice. Just pick some other FPx
1739 // register to hold it.
1740 unsigned NewReg = getScratchReg();
1741 duplicateToTop(FirstFPRegOp, NewReg, MI);
1742 FirstFPRegOp = NewReg;
1745 /// Okay we know we have two different FPx operands now:
1746 assert(StackTop == 2 && "Must have two values live!");
1748 /// 3) If SecondFPRegOp is currently in ST(0) and FirstFPRegOp is currently
1749 /// in ST(1). In this case, emit an fxch.
1750 if (getStackEntry(0) == SecondFPRegOp) {
1751 assert(getStackEntry(1) == FirstFPRegOp && "Unknown regs live");
1752 moveToTop(FirstFPRegOp, MI);
1755 /// 4) Finally, FirstFPRegOp must be in ST(0) and SecondFPRegOp must be in
1756 /// ST(1). Just remove both from our understanding of the stack and return.
1757 assert(getStackEntry(0) == FirstFPRegOp && "Unknown regs live");
1758 assert(getStackEntry(1) == SecondFPRegOp && "Unknown regs live");
1763 I = MBB->erase(I); // Remove the pseudo instruction
1765 // We want to leave I pointing to the previous instruction, but what if we
1766 // just erased the first instruction?
1767 if (I == MBB->begin()) {
1768 DEBUG(dbgs() << "Inserting dummy KILL\n");
1769 I = BuildMI(*MBB, I, DebugLoc(), TII->get(TargetOpcode::KILL));