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/DenseMap.h"
31 #include "llvm/ADT/SmallPtrSet.h"
32 #include "llvm/ADT/SmallVector.h"
33 #include "llvm/ADT/Statistic.h"
34 #include "llvm/ADT/STLExtras.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/Support/Debug.h"
40 #include "llvm/Support/ErrorHandling.h"
41 #include "llvm/Support/raw_ostream.h"
42 #include "llvm/Target/TargetInstrInfo.h"
43 #include "llvm/Target/TargetMachine.h"
47 STATISTIC(NumFXCH, "Number of fxch instructions inserted");
48 STATISTIC(NumFP , "Number of floating point instructions");
51 struct FPS : public MachineFunctionPass {
53 FPS() : MachineFunctionPass(&ID) {
54 // This is really only to keep valgrind quiet.
55 // The logic in isLive() is too much for it.
56 memset(Stack, 0, sizeof(Stack));
57 memset(RegMap, 0, sizeof(RegMap));
60 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
62 AU.addPreservedID(MachineLoopInfoID);
63 AU.addPreservedID(MachineDominatorsID);
64 MachineFunctionPass::getAnalysisUsage(AU);
67 virtual bool runOnMachineFunction(MachineFunction &MF);
69 virtual const char *getPassName() const { return "X86 FP Stackifier"; }
72 const TargetInstrInfo *TII; // Machine instruction info.
74 // Two CFG edges are related if they leave the same block, or enter the same
75 // block. The transitive closure of an edge under this relation is a
76 // LiveBundle. It represents a set of CFG edges where the live FP stack
77 // registers must be allocated identically in the x87 stack.
79 // A LiveBundle is usually all the edges leaving a block, or all the edges
80 // entering a block, but it can contain more edges if critical edges are
83 // The set of live FP registers in a LiveBundle is calculated by bundleCFG,
84 // but the exact mapping of FP registers to stack slots is fixed later.
86 // Bit mask of live FP registers. Bit 0 = FP0, bit 1 = FP1, &c.
89 // Number of pre-assigned live registers in FixStack. This is 0 when the
90 // stack order has not yet been fixed.
93 // Assigned stack order for live-in registers.
94 // FixStack[i] == getStackEntry(i) for all i < FixCount.
95 unsigned char FixStack[8];
97 LiveBundle(unsigned m = 0) : Mask(m), FixCount(0) {}
99 // Have the live registers been assigned a stack order yet?
100 bool isFixed() const { return !Mask || FixCount; }
103 // Numbered LiveBundle structs. LiveBundles[0] is used for all CFG edges
104 // with no live FP registers.
105 SmallVector<LiveBundle, 8> LiveBundles;
107 // Map each MBB in the current function to an (ingoing, outgoing) index into
108 // LiveBundles. Blocks with no FP registers live in or out map to (0, 0)
109 // and are not actually stored in the map.
110 DenseMap<MachineBasicBlock*, std::pair<unsigned, unsigned> > BlockBundle;
112 // Return a bitmask of FP registers in block's live-in list.
113 unsigned calcLiveInMask(MachineBasicBlock *MBB) {
115 for (MachineBasicBlock::livein_iterator I = MBB->livein_begin(),
116 E = MBB->livein_end(); I != E; ++I) {
117 unsigned Reg = *I - X86::FP0;
124 // Partition all the CFG edges into LiveBundles.
125 void bundleCFG(MachineFunction &MF);
127 MachineBasicBlock *MBB; // Current basic block
128 unsigned Stack[8]; // FP<n> Registers in each stack slot...
129 unsigned RegMap[8]; // Track which stack slot contains each register
130 unsigned StackTop; // The current top of the FP stack.
132 // Set up our stack model to match the incoming registers to MBB.
133 void setupBlockStack();
135 // Shuffle live registers to match the expectations of successor blocks.
136 void finishBlockStack();
138 void dumpStack() const {
139 dbgs() << "Stack contents:";
140 for (unsigned i = 0; i != StackTop; ++i) {
141 dbgs() << " FP" << Stack[i];
142 assert(RegMap[Stack[i]] == i && "Stack[] doesn't match RegMap[]!");
147 /// isStackEmpty - Return true if the FP stack is empty.
148 bool isStackEmpty() const {
149 return StackTop == 0;
152 /// getSlot - Return the stack slot number a particular register number is
154 unsigned getSlot(unsigned RegNo) const {
155 assert(RegNo < 8 && "Regno out of range!");
156 return RegMap[RegNo];
159 /// isLive - Is RegNo currently live in the stack?
160 bool isLive(unsigned RegNo) const {
161 unsigned Slot = getSlot(RegNo);
162 return Slot < StackTop && Stack[Slot] == RegNo;
165 /// getScratchReg - Return an FP register that is not currently in use.
166 unsigned getScratchReg() {
167 for (int i = 7; i >= 0; --i)
170 llvm_unreachable("Ran out of scratch FP registers");
173 /// getStackEntry - Return the X86::FP<n> register in register ST(i).
174 unsigned getStackEntry(unsigned STi) const {
175 assert(STi < StackTop && "Access past stack top!");
176 return Stack[StackTop-1-STi];
179 /// getSTReg - Return the X86::ST(i) register which contains the specified
180 /// FP<RegNo> register.
181 unsigned getSTReg(unsigned RegNo) const {
182 return StackTop - 1 - getSlot(RegNo) + llvm::X86::ST0;
185 // pushReg - Push the specified FP<n> register onto the stack.
186 void pushReg(unsigned Reg) {
187 assert(Reg < 8 && "Register number out of range!");
188 assert(StackTop < 8 && "Stack overflow!");
189 Stack[StackTop] = Reg;
190 RegMap[Reg] = StackTop++;
193 bool isAtTop(unsigned RegNo) const { return getSlot(RegNo) == StackTop-1; }
194 void moveToTop(unsigned RegNo, MachineBasicBlock::iterator I) {
195 DebugLoc dl = I == MBB->end() ? DebugLoc() : I->getDebugLoc();
196 if (isAtTop(RegNo)) return;
198 unsigned STReg = getSTReg(RegNo);
199 unsigned RegOnTop = getStackEntry(0);
201 // Swap the slots the regs are in.
202 std::swap(RegMap[RegNo], RegMap[RegOnTop]);
204 // Swap stack slot contents.
205 assert(RegMap[RegOnTop] < StackTop);
206 std::swap(Stack[RegMap[RegOnTop]], Stack[StackTop-1]);
208 // Emit an fxch to update the runtime processors version of the state.
209 BuildMI(*MBB, I, dl, TII->get(X86::XCH_F)).addReg(STReg);
213 void duplicateToTop(unsigned RegNo, unsigned AsReg, MachineInstr *I) {
214 DebugLoc dl = I == MBB->end() ? DebugLoc() : I->getDebugLoc();
215 unsigned STReg = getSTReg(RegNo);
216 pushReg(AsReg); // New register on top of stack
218 BuildMI(*MBB, I, dl, TII->get(X86::LD_Frr)).addReg(STReg);
221 /// popStackAfter - Pop the current value off of the top of the FP stack
222 /// after the specified instruction.
223 void popStackAfter(MachineBasicBlock::iterator &I);
225 /// freeStackSlotAfter - Free the specified register from the register
226 /// stack, so that it is no longer in a register. If the register is
227 /// currently at the top of the stack, we just pop the current instruction,
228 /// otherwise we store the current top-of-stack into the specified slot,
229 /// then pop the top of stack.
230 void freeStackSlotAfter(MachineBasicBlock::iterator &I, unsigned Reg);
232 /// freeStackSlotBefore - Just the pop, no folding. Return the inserted
234 MachineBasicBlock::iterator
235 freeStackSlotBefore(MachineBasicBlock::iterator I, unsigned FPRegNo);
237 /// Adjust the live registers to be the set in Mask.
238 void adjustLiveRegs(unsigned Mask, MachineBasicBlock::iterator I);
240 /// Shuffle the top FixCount stack entries susch that FP reg FixStack[0] is
241 /// st(0), FP reg FixStack[1] is st(1) etc.
242 void shuffleStackTop(const unsigned char *FixStack, unsigned FixCount,
243 MachineBasicBlock::iterator I);
245 bool processBasicBlock(MachineFunction &MF, MachineBasicBlock &MBB);
247 void handleZeroArgFP(MachineBasicBlock::iterator &I);
248 void handleOneArgFP(MachineBasicBlock::iterator &I);
249 void handleOneArgFPRW(MachineBasicBlock::iterator &I);
250 void handleTwoArgFP(MachineBasicBlock::iterator &I);
251 void handleCompareFP(MachineBasicBlock::iterator &I);
252 void handleCondMovFP(MachineBasicBlock::iterator &I);
253 void handleSpecialFP(MachineBasicBlock::iterator &I);
255 bool translateCopy(MachineInstr*);
260 FunctionPass *llvm::createX86FloatingPointStackifierPass() { return new FPS(); }
262 /// getFPReg - Return the X86::FPx register number for the specified operand.
263 /// For example, this returns 3 for X86::FP3.
264 static unsigned getFPReg(const MachineOperand &MO) {
265 assert(MO.isReg() && "Expected an FP register!");
266 unsigned Reg = MO.getReg();
267 assert(Reg >= X86::FP0 && Reg <= X86::FP6 && "Expected FP register!");
268 return Reg - X86::FP0;
271 /// runOnMachineFunction - Loop over all of the basic blocks, transforming FP
272 /// register references into FP stack references.
274 bool FPS::runOnMachineFunction(MachineFunction &MF) {
275 // We only need to run this pass if there are any FP registers used in this
276 // function. If it is all integer, there is nothing for us to do!
277 bool FPIsUsed = false;
279 assert(X86::FP6 == X86::FP0+6 && "Register enums aren't sorted right!");
280 for (unsigned i = 0; i <= 6; ++i)
281 if (MF.getRegInfo().isPhysRegUsed(X86::FP0+i)) {
287 if (!FPIsUsed) return false;
289 TII = MF.getTarget().getInstrInfo();
291 // Prepare cross-MBB liveness.
296 // Process the function in depth first order so that we process at least one
297 // of the predecessors for every reachable block in the function.
298 SmallPtrSet<MachineBasicBlock*, 8> Processed;
299 MachineBasicBlock *Entry = MF.begin();
301 bool Changed = false;
302 for (df_ext_iterator<MachineBasicBlock*, SmallPtrSet<MachineBasicBlock*, 8> >
303 I = df_ext_begin(Entry, Processed), E = df_ext_end(Entry, Processed);
305 Changed |= processBasicBlock(MF, **I);
307 // Process any unreachable blocks in arbitrary order now.
308 if (MF.size() != Processed.size())
309 for (MachineFunction::iterator BB = MF.begin(), E = MF.end(); BB != E; ++BB)
310 if (Processed.insert(BB))
311 Changed |= processBasicBlock(MF, *BB);
319 /// bundleCFG - Scan all the basic blocks to determine consistent live-in and
320 /// live-out sets for the FP registers. Consistent means that the set of
321 /// registers live-out from a block is identical to the live-in set of all
322 /// successors. This is not enforced by the normal live-in lists since
323 /// registers may be implicitly defined, or not used by all successors.
324 void FPS::bundleCFG(MachineFunction &MF) {
325 assert(LiveBundles.empty() && "Stale data in LiveBundles");
326 assert(BlockBundle.empty() && "Stale data in BlockBundle");
327 SmallPtrSet<MachineBasicBlock*, 8> PropDown, PropUp;
329 // LiveBundle[0] is the empty live-in set.
330 LiveBundles.resize(1);
332 // First gather the actual live-in masks for all MBBs.
333 for (MachineFunction::iterator I = MF.begin(), E = MF.end(); I != E; ++I) {
334 MachineBasicBlock *MBB = I;
335 const unsigned Mask = calcLiveInMask(MBB);
338 // Ingoing bundle index.
339 unsigned &Idx = BlockBundle[MBB].first;
340 // Already assigned an ingoing bundle?
343 // Allocate a new LiveBundle struct for this block's live-ins.
344 const unsigned BundleIdx = Idx = LiveBundles.size();
345 DEBUG(dbgs() << "Creating LB#" << BundleIdx << ": in:BB#"
346 << MBB->getNumber());
347 LiveBundles.push_back(Mask);
348 LiveBundle &Bundle = LiveBundles.back();
350 // Make sure all predecessors have the same live-out set.
353 // Keep pushing liveness up and down the CFG until convergence.
354 // Only critical edges cause iteration here, but when they do, multiple
355 // blocks can be assigned to the same LiveBundle index.
357 // Assign BundleIdx as liveout from predecessors in PropUp.
358 for (SmallPtrSet<MachineBasicBlock*, 16>::iterator I = PropUp.begin(),
359 E = PropUp.end(); I != E; ++I) {
360 MachineBasicBlock *MBB = *I;
361 for (MachineBasicBlock::const_pred_iterator LinkI = MBB->pred_begin(),
362 LinkE = MBB->pred_end(); LinkI != LinkE; ++LinkI) {
363 MachineBasicBlock *PredMBB = *LinkI;
364 // PredMBB's liveout bundle should be set to LIIdx.
365 unsigned &Idx = BlockBundle[PredMBB].second;
367 assert(Idx == BundleIdx && "Inconsistent CFG");
371 DEBUG(dbgs() << " out:BB#" << PredMBB->getNumber());
372 // Propagate to siblings.
373 if (PredMBB->succ_size() > 1)
374 PropDown.insert(PredMBB);
379 // Assign BundleIdx as livein to successors in PropDown.
380 for (SmallPtrSet<MachineBasicBlock*, 16>::iterator I = PropDown.begin(),
381 E = PropDown.end(); I != E; ++I) {
382 MachineBasicBlock *MBB = *I;
383 for (MachineBasicBlock::const_succ_iterator LinkI = MBB->succ_begin(),
384 LinkE = MBB->succ_end(); LinkI != LinkE; ++LinkI) {
385 MachineBasicBlock *SuccMBB = *LinkI;
386 // LinkMBB's livein bundle should be set to BundleIdx.
387 unsigned &Idx = BlockBundle[SuccMBB].first;
389 assert(Idx == BundleIdx && "Inconsistent CFG");
393 DEBUG(dbgs() << " in:BB#" << SuccMBB->getNumber());
394 // Propagate to siblings.
395 if (SuccMBB->pred_size() > 1)
396 PropUp.insert(SuccMBB);
397 // Also accumulate the bundle liveness mask from the liveins here.
398 Bundle.Mask |= calcLiveInMask(SuccMBB);
402 } while (!PropUp.empty());
405 for (unsigned i = 0; i < 8; ++i)
406 if (Bundle.Mask & (1<<i))
407 dbgs() << " %FP" << i;
413 /// processBasicBlock - Loop over all of the instructions in the basic block,
414 /// transforming FP instructions into their stack form.
416 bool FPS::processBasicBlock(MachineFunction &MF, MachineBasicBlock &BB) {
417 bool Changed = false;
422 for (MachineBasicBlock::iterator I = BB.begin(); I != BB.end(); ++I) {
423 MachineInstr *MI = I;
424 uint64_t Flags = MI->getDesc().TSFlags;
426 unsigned FPInstClass = Flags & X86II::FPTypeMask;
427 if (MI->isInlineAsm())
428 FPInstClass = X86II::SpecialFP;
430 if (MI->isCopy() && translateCopy(MI))
431 FPInstClass = X86II::SpecialFP;
433 if (FPInstClass == X86II::NotFP)
434 continue; // Efficiently ignore non-fp insts!
436 MachineInstr *PrevMI = 0;
440 ++NumFP; // Keep track of # of pseudo instrs
441 DEBUG(dbgs() << "\nFPInst:\t" << *MI);
443 // Get dead variables list now because the MI pointer may be deleted as part
445 SmallVector<unsigned, 8> DeadRegs;
446 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
447 const MachineOperand &MO = MI->getOperand(i);
448 if (MO.isReg() && MO.isDead())
449 DeadRegs.push_back(MO.getReg());
452 switch (FPInstClass) {
453 case X86II::ZeroArgFP: handleZeroArgFP(I); break;
454 case X86II::OneArgFP: handleOneArgFP(I); break; // fstp ST(0)
455 case X86II::OneArgFPRW: handleOneArgFPRW(I); break; // ST(0) = fsqrt(ST(0))
456 case X86II::TwoArgFP: handleTwoArgFP(I); break;
457 case X86II::CompareFP: handleCompareFP(I); break;
458 case X86II::CondMovFP: handleCondMovFP(I); break;
459 case X86II::SpecialFP: handleSpecialFP(I); break;
460 default: llvm_unreachable("Unknown FP Type!");
463 // Check to see if any of the values defined by this instruction are dead
464 // after definition. If so, pop them.
465 for (unsigned i = 0, e = DeadRegs.size(); i != e; ++i) {
466 unsigned Reg = DeadRegs[i];
467 if (Reg >= X86::FP0 && Reg <= X86::FP6) {
468 DEBUG(dbgs() << "Register FP#" << Reg-X86::FP0 << " is dead!\n");
469 freeStackSlotAfter(I, Reg-X86::FP0);
473 // Print out all of the instructions expanded to if -debug
475 MachineBasicBlock::iterator PrevI(PrevMI);
477 dbgs() << "Just deleted pseudo instruction\n";
479 MachineBasicBlock::iterator Start = I;
480 // Rewind to first instruction newly inserted.
481 while (Start != BB.begin() && prior(Start) != PrevI) --Start;
482 dbgs() << "Inserted instructions:\n\t";
483 Start->print(dbgs(), &MF.getTarget());
484 while (++Start != llvm::next(I)) {}
497 /// setupBlockStack - Use the BlockBundle map to set up our model of the stack
498 /// to match predecessors' live out stack.
499 void FPS::setupBlockStack() {
500 DEBUG(dbgs() << "\nSetting up live-ins for BB#" << MBB->getNumber()
501 << " derived from " << MBB->getName() << ".\n");
503 const LiveBundle &Bundle = LiveBundles[BlockBundle.lookup(MBB).first];
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 unsigned BundleIdx = BlockBundle.lookup(MBB).second;
541 LiveBundle &Bundle = LiveBundles[BundleIdx];
543 // We may need to kill and define some registers to match successors.
544 // FIXME: This can probably be combined with the shuffle below.
545 MachineBasicBlock::iterator Term = MBB->getFirstTerminator();
546 adjustLiveRegs(Bundle.Mask, Term);
549 DEBUG(dbgs() << "No live-outs.\n");
553 // Has the stack order been fixed yet?
554 DEBUG(dbgs() << "LB#" << BundleIdx << ": ");
555 if (Bundle.isFixed()) {
556 DEBUG(dbgs() << "Shuffling stack to match.\n");
557 shuffleStackTop(Bundle.FixStack, Bundle.FixCount, Term);
559 // Not fixed yet, we get to choose.
560 DEBUG(dbgs() << "Fixing stack order now.\n");
561 Bundle.FixCount = StackTop;
562 for (unsigned i = 0; i < StackTop; ++i)
563 Bundle.FixStack[i] = getStackEntry(i);
568 //===----------------------------------------------------------------------===//
569 // Efficient Lookup Table Support
570 //===----------------------------------------------------------------------===//
576 bool operator<(const TableEntry &TE) const { return from < TE.from; }
577 friend bool operator<(const TableEntry &TE, unsigned V) {
580 friend bool operator<(unsigned V, const TableEntry &TE) {
587 static bool TableIsSorted(const TableEntry *Table, unsigned NumEntries) {
588 for (unsigned i = 0; i != NumEntries-1; ++i)
589 if (!(Table[i] < Table[i+1])) return false;
594 static int Lookup(const TableEntry *Table, unsigned N, unsigned Opcode) {
595 const TableEntry *I = std::lower_bound(Table, Table+N, Opcode);
596 if (I != Table+N && I->from == Opcode)
602 #define ASSERT_SORTED(TABLE)
604 #define ASSERT_SORTED(TABLE) \
605 { static bool TABLE##Checked = false; \
606 if (!TABLE##Checked) { \
607 assert(TableIsSorted(TABLE, array_lengthof(TABLE)) && \
608 "All lookup tables must be sorted for efficient access!"); \
609 TABLE##Checked = true; \
614 //===----------------------------------------------------------------------===//
615 // Register File -> Register Stack Mapping Methods
616 //===----------------------------------------------------------------------===//
618 // OpcodeTable - Sorted map of register instructions to their stack version.
619 // The first element is an register file pseudo instruction, the second is the
620 // concrete X86 instruction which uses the register stack.
622 static const TableEntry OpcodeTable[] = {
623 { X86::ABS_Fp32 , X86::ABS_F },
624 { X86::ABS_Fp64 , X86::ABS_F },
625 { X86::ABS_Fp80 , X86::ABS_F },
626 { X86::ADD_Fp32m , X86::ADD_F32m },
627 { X86::ADD_Fp64m , X86::ADD_F64m },
628 { X86::ADD_Fp64m32 , X86::ADD_F32m },
629 { X86::ADD_Fp80m32 , X86::ADD_F32m },
630 { X86::ADD_Fp80m64 , X86::ADD_F64m },
631 { X86::ADD_FpI16m32 , X86::ADD_FI16m },
632 { X86::ADD_FpI16m64 , X86::ADD_FI16m },
633 { X86::ADD_FpI16m80 , X86::ADD_FI16m },
634 { X86::ADD_FpI32m32 , X86::ADD_FI32m },
635 { X86::ADD_FpI32m64 , X86::ADD_FI32m },
636 { X86::ADD_FpI32m80 , X86::ADD_FI32m },
637 { X86::CHS_Fp32 , X86::CHS_F },
638 { X86::CHS_Fp64 , X86::CHS_F },
639 { X86::CHS_Fp80 , X86::CHS_F },
640 { X86::CMOVBE_Fp32 , X86::CMOVBE_F },
641 { X86::CMOVBE_Fp64 , X86::CMOVBE_F },
642 { X86::CMOVBE_Fp80 , X86::CMOVBE_F },
643 { X86::CMOVB_Fp32 , X86::CMOVB_F },
644 { X86::CMOVB_Fp64 , X86::CMOVB_F },
645 { X86::CMOVB_Fp80 , X86::CMOVB_F },
646 { X86::CMOVE_Fp32 , X86::CMOVE_F },
647 { X86::CMOVE_Fp64 , X86::CMOVE_F },
648 { X86::CMOVE_Fp80 , X86::CMOVE_F },
649 { X86::CMOVNBE_Fp32 , X86::CMOVNBE_F },
650 { X86::CMOVNBE_Fp64 , X86::CMOVNBE_F },
651 { X86::CMOVNBE_Fp80 , X86::CMOVNBE_F },
652 { X86::CMOVNB_Fp32 , X86::CMOVNB_F },
653 { X86::CMOVNB_Fp64 , X86::CMOVNB_F },
654 { X86::CMOVNB_Fp80 , X86::CMOVNB_F },
655 { X86::CMOVNE_Fp32 , X86::CMOVNE_F },
656 { X86::CMOVNE_Fp64 , X86::CMOVNE_F },
657 { X86::CMOVNE_Fp80 , X86::CMOVNE_F },
658 { X86::CMOVNP_Fp32 , X86::CMOVNP_F },
659 { X86::CMOVNP_Fp64 , X86::CMOVNP_F },
660 { X86::CMOVNP_Fp80 , X86::CMOVNP_F },
661 { X86::CMOVP_Fp32 , X86::CMOVP_F },
662 { X86::CMOVP_Fp64 , X86::CMOVP_F },
663 { X86::CMOVP_Fp80 , X86::CMOVP_F },
664 { X86::COS_Fp32 , X86::COS_F },
665 { X86::COS_Fp64 , X86::COS_F },
666 { X86::COS_Fp80 , X86::COS_F },
667 { X86::DIVR_Fp32m , X86::DIVR_F32m },
668 { X86::DIVR_Fp64m , X86::DIVR_F64m },
669 { X86::DIVR_Fp64m32 , X86::DIVR_F32m },
670 { X86::DIVR_Fp80m32 , X86::DIVR_F32m },
671 { X86::DIVR_Fp80m64 , X86::DIVR_F64m },
672 { X86::DIVR_FpI16m32, X86::DIVR_FI16m},
673 { X86::DIVR_FpI16m64, X86::DIVR_FI16m},
674 { X86::DIVR_FpI16m80, X86::DIVR_FI16m},
675 { X86::DIVR_FpI32m32, X86::DIVR_FI32m},
676 { X86::DIVR_FpI32m64, X86::DIVR_FI32m},
677 { X86::DIVR_FpI32m80, X86::DIVR_FI32m},
678 { X86::DIV_Fp32m , X86::DIV_F32m },
679 { X86::DIV_Fp64m , X86::DIV_F64m },
680 { X86::DIV_Fp64m32 , X86::DIV_F32m },
681 { X86::DIV_Fp80m32 , X86::DIV_F32m },
682 { X86::DIV_Fp80m64 , X86::DIV_F64m },
683 { X86::DIV_FpI16m32 , X86::DIV_FI16m },
684 { X86::DIV_FpI16m64 , X86::DIV_FI16m },
685 { X86::DIV_FpI16m80 , X86::DIV_FI16m },
686 { X86::DIV_FpI32m32 , X86::DIV_FI32m },
687 { X86::DIV_FpI32m64 , X86::DIV_FI32m },
688 { X86::DIV_FpI32m80 , X86::DIV_FI32m },
689 { X86::ILD_Fp16m32 , X86::ILD_F16m },
690 { X86::ILD_Fp16m64 , X86::ILD_F16m },
691 { X86::ILD_Fp16m80 , X86::ILD_F16m },
692 { X86::ILD_Fp32m32 , X86::ILD_F32m },
693 { X86::ILD_Fp32m64 , X86::ILD_F32m },
694 { X86::ILD_Fp32m80 , X86::ILD_F32m },
695 { X86::ILD_Fp64m32 , X86::ILD_F64m },
696 { X86::ILD_Fp64m64 , X86::ILD_F64m },
697 { X86::ILD_Fp64m80 , X86::ILD_F64m },
698 { X86::ISTT_Fp16m32 , X86::ISTT_FP16m},
699 { X86::ISTT_Fp16m64 , X86::ISTT_FP16m},
700 { X86::ISTT_Fp16m80 , X86::ISTT_FP16m},
701 { X86::ISTT_Fp32m32 , X86::ISTT_FP32m},
702 { X86::ISTT_Fp32m64 , X86::ISTT_FP32m},
703 { X86::ISTT_Fp32m80 , X86::ISTT_FP32m},
704 { X86::ISTT_Fp64m32 , X86::ISTT_FP64m},
705 { X86::ISTT_Fp64m64 , X86::ISTT_FP64m},
706 { X86::ISTT_Fp64m80 , X86::ISTT_FP64m},
707 { X86::IST_Fp16m32 , X86::IST_F16m },
708 { X86::IST_Fp16m64 , X86::IST_F16m },
709 { X86::IST_Fp16m80 , X86::IST_F16m },
710 { X86::IST_Fp32m32 , X86::IST_F32m },
711 { X86::IST_Fp32m64 , X86::IST_F32m },
712 { X86::IST_Fp32m80 , X86::IST_F32m },
713 { X86::IST_Fp64m32 , X86::IST_FP64m },
714 { X86::IST_Fp64m64 , X86::IST_FP64m },
715 { X86::IST_Fp64m80 , X86::IST_FP64m },
716 { X86::LD_Fp032 , X86::LD_F0 },
717 { X86::LD_Fp064 , X86::LD_F0 },
718 { X86::LD_Fp080 , X86::LD_F0 },
719 { X86::LD_Fp132 , X86::LD_F1 },
720 { X86::LD_Fp164 , X86::LD_F1 },
721 { X86::LD_Fp180 , X86::LD_F1 },
722 { X86::LD_Fp32m , X86::LD_F32m },
723 { X86::LD_Fp32m64 , X86::LD_F32m },
724 { X86::LD_Fp32m80 , X86::LD_F32m },
725 { X86::LD_Fp64m , X86::LD_F64m },
726 { X86::LD_Fp64m80 , X86::LD_F64m },
727 { X86::LD_Fp80m , X86::LD_F80m },
728 { X86::MUL_Fp32m , X86::MUL_F32m },
729 { X86::MUL_Fp64m , X86::MUL_F64m },
730 { X86::MUL_Fp64m32 , X86::MUL_F32m },
731 { X86::MUL_Fp80m32 , X86::MUL_F32m },
732 { X86::MUL_Fp80m64 , X86::MUL_F64m },
733 { X86::MUL_FpI16m32 , X86::MUL_FI16m },
734 { X86::MUL_FpI16m64 , X86::MUL_FI16m },
735 { X86::MUL_FpI16m80 , X86::MUL_FI16m },
736 { X86::MUL_FpI32m32 , X86::MUL_FI32m },
737 { X86::MUL_FpI32m64 , X86::MUL_FI32m },
738 { X86::MUL_FpI32m80 , X86::MUL_FI32m },
739 { X86::SIN_Fp32 , X86::SIN_F },
740 { X86::SIN_Fp64 , X86::SIN_F },
741 { X86::SIN_Fp80 , X86::SIN_F },
742 { X86::SQRT_Fp32 , X86::SQRT_F },
743 { X86::SQRT_Fp64 , X86::SQRT_F },
744 { X86::SQRT_Fp80 , X86::SQRT_F },
745 { X86::ST_Fp32m , X86::ST_F32m },
746 { X86::ST_Fp64m , X86::ST_F64m },
747 { X86::ST_Fp64m32 , X86::ST_F32m },
748 { X86::ST_Fp80m32 , X86::ST_F32m },
749 { X86::ST_Fp80m64 , X86::ST_F64m },
750 { X86::ST_FpP80m , X86::ST_FP80m },
751 { X86::SUBR_Fp32m , X86::SUBR_F32m },
752 { X86::SUBR_Fp64m , X86::SUBR_F64m },
753 { X86::SUBR_Fp64m32 , X86::SUBR_F32m },
754 { X86::SUBR_Fp80m32 , X86::SUBR_F32m },
755 { X86::SUBR_Fp80m64 , X86::SUBR_F64m },
756 { X86::SUBR_FpI16m32, X86::SUBR_FI16m},
757 { X86::SUBR_FpI16m64, X86::SUBR_FI16m},
758 { X86::SUBR_FpI16m80, X86::SUBR_FI16m},
759 { X86::SUBR_FpI32m32, X86::SUBR_FI32m},
760 { X86::SUBR_FpI32m64, X86::SUBR_FI32m},
761 { X86::SUBR_FpI32m80, X86::SUBR_FI32m},
762 { X86::SUB_Fp32m , X86::SUB_F32m },
763 { X86::SUB_Fp64m , X86::SUB_F64m },
764 { X86::SUB_Fp64m32 , X86::SUB_F32m },
765 { X86::SUB_Fp80m32 , X86::SUB_F32m },
766 { X86::SUB_Fp80m64 , X86::SUB_F64m },
767 { X86::SUB_FpI16m32 , X86::SUB_FI16m },
768 { X86::SUB_FpI16m64 , X86::SUB_FI16m },
769 { X86::SUB_FpI16m80 , X86::SUB_FI16m },
770 { X86::SUB_FpI32m32 , X86::SUB_FI32m },
771 { X86::SUB_FpI32m64 , X86::SUB_FI32m },
772 { X86::SUB_FpI32m80 , X86::SUB_FI32m },
773 { X86::TST_Fp32 , X86::TST_F },
774 { X86::TST_Fp64 , X86::TST_F },
775 { X86::TST_Fp80 , X86::TST_F },
776 { X86::UCOM_FpIr32 , X86::UCOM_FIr },
777 { X86::UCOM_FpIr64 , X86::UCOM_FIr },
778 { X86::UCOM_FpIr80 , X86::UCOM_FIr },
779 { X86::UCOM_Fpr32 , X86::UCOM_Fr },
780 { X86::UCOM_Fpr64 , X86::UCOM_Fr },
781 { X86::UCOM_Fpr80 , X86::UCOM_Fr },
784 static unsigned getConcreteOpcode(unsigned Opcode) {
785 ASSERT_SORTED(OpcodeTable);
786 int Opc = Lookup(OpcodeTable, array_lengthof(OpcodeTable), Opcode);
787 assert(Opc != -1 && "FP Stack instruction not in OpcodeTable!");
791 //===----------------------------------------------------------------------===//
793 //===----------------------------------------------------------------------===//
795 // PopTable - Sorted map of instructions to their popping version. The first
796 // element is an instruction, the second is the version which pops.
798 static const TableEntry PopTable[] = {
799 { X86::ADD_FrST0 , X86::ADD_FPrST0 },
801 { X86::DIVR_FrST0, X86::DIVR_FPrST0 },
802 { X86::DIV_FrST0 , X86::DIV_FPrST0 },
804 { X86::IST_F16m , X86::IST_FP16m },
805 { X86::IST_F32m , X86::IST_FP32m },
807 { X86::MUL_FrST0 , X86::MUL_FPrST0 },
809 { X86::ST_F32m , X86::ST_FP32m },
810 { X86::ST_F64m , X86::ST_FP64m },
811 { X86::ST_Frr , X86::ST_FPrr },
813 { X86::SUBR_FrST0, X86::SUBR_FPrST0 },
814 { X86::SUB_FrST0 , X86::SUB_FPrST0 },
816 { X86::UCOM_FIr , X86::UCOM_FIPr },
818 { X86::UCOM_FPr , X86::UCOM_FPPr },
819 { X86::UCOM_Fr , X86::UCOM_FPr },
822 /// popStackAfter - Pop the current value off of the top of the FP stack after
823 /// the specified instruction. This attempts to be sneaky and combine the pop
824 /// into the instruction itself if possible. The iterator is left pointing to
825 /// the last instruction, be it a new pop instruction inserted, or the old
826 /// instruction if it was modified in place.
828 void FPS::popStackAfter(MachineBasicBlock::iterator &I) {
829 MachineInstr* MI = I;
830 DebugLoc dl = MI->getDebugLoc();
831 ASSERT_SORTED(PopTable);
832 assert(StackTop > 0 && "Cannot pop empty stack!");
833 RegMap[Stack[--StackTop]] = ~0; // Update state
835 // Check to see if there is a popping version of this instruction...
836 int Opcode = Lookup(PopTable, array_lengthof(PopTable), I->getOpcode());
838 I->setDesc(TII->get(Opcode));
839 if (Opcode == X86::UCOM_FPPr)
841 } else { // Insert an explicit pop
842 I = BuildMI(*MBB, ++I, dl, TII->get(X86::ST_FPrr)).addReg(X86::ST0);
846 /// freeStackSlotAfter - Free the specified register from the register stack, so
847 /// that it is no longer in a register. If the register is currently at the top
848 /// of the stack, we just pop the current instruction, otherwise we store the
849 /// current top-of-stack into the specified slot, then pop the top of stack.
850 void FPS::freeStackSlotAfter(MachineBasicBlock::iterator &I, unsigned FPRegNo) {
851 if (getStackEntry(0) == FPRegNo) { // already at the top of stack? easy.
856 // Otherwise, store the top of stack into the dead slot, killing the operand
857 // without having to add in an explicit xchg then pop.
859 I = freeStackSlotBefore(++I, FPRegNo);
862 /// freeStackSlotBefore - Free the specified register without trying any
864 MachineBasicBlock::iterator
865 FPS::freeStackSlotBefore(MachineBasicBlock::iterator I, unsigned FPRegNo) {
866 unsigned STReg = getSTReg(FPRegNo);
867 unsigned OldSlot = getSlot(FPRegNo);
868 unsigned TopReg = Stack[StackTop-1];
869 Stack[OldSlot] = TopReg;
870 RegMap[TopReg] = OldSlot;
871 RegMap[FPRegNo] = ~0;
872 Stack[--StackTop] = ~0;
873 return BuildMI(*MBB, I, DebugLoc(), TII->get(X86::ST_FPrr)).addReg(STReg);
876 /// adjustLiveRegs - Kill and revive registers such that exactly the FP
877 /// registers with a bit in Mask are live.
878 void FPS::adjustLiveRegs(unsigned Mask, MachineBasicBlock::iterator I) {
879 unsigned Defs = Mask;
881 for (unsigned i = 0; i < StackTop; ++i) {
882 unsigned RegNo = Stack[i];
883 if (!(Defs & (1 << RegNo)))
884 // This register is live, but we don't want it.
885 Kills |= (1 << RegNo);
887 // We don't need to imp-def this live register.
888 Defs &= ~(1 << RegNo);
890 assert((Kills & Defs) == 0 && "Register needs killing and def'ing?");
892 // Produce implicit-defs for free by using killed registers.
893 while (Kills && Defs) {
894 unsigned KReg = CountTrailingZeros_32(Kills);
895 unsigned DReg = CountTrailingZeros_32(Defs);
896 DEBUG(dbgs() << "Renaming %FP" << KReg << " as imp %FP" << DReg << "\n");
897 std::swap(Stack[getSlot(KReg)], Stack[getSlot(DReg)]);
898 std::swap(RegMap[KReg], RegMap[DReg]);
899 Kills &= ~(1 << KReg);
900 Defs &= ~(1 << DReg);
903 // Kill registers by popping.
904 if (Kills && I != MBB->begin()) {
905 MachineBasicBlock::iterator I2 = llvm::prior(I);
907 unsigned KReg = getStackEntry(0);
908 if (!(Kills & (1 << KReg)))
910 DEBUG(dbgs() << "Popping %FP" << KReg << "\n");
912 Kills &= ~(1 << KReg);
916 // Manually kill the rest.
918 unsigned KReg = CountTrailingZeros_32(Kills);
919 DEBUG(dbgs() << "Killing %FP" << KReg << "\n");
920 freeStackSlotBefore(I, KReg);
921 Kills &= ~(1 << KReg);
924 // Load zeros for all the imp-defs.
926 unsigned DReg = CountTrailingZeros_32(Defs);
927 DEBUG(dbgs() << "Defining %FP" << DReg << " as 0\n");
928 BuildMI(*MBB, I, DebugLoc(), TII->get(X86::LD_F0));
930 Defs &= ~(1 << DReg);
933 // Now we should have the correct registers live.
935 assert(StackTop == CountPopulation_32(Mask) && "Live count mismatch");
938 /// shuffleStackTop - emit fxch instructions before I to shuffle the top
939 /// FixCount entries into the order given by FixStack.
940 /// FIXME: Is there a better algorithm than insertion sort?
941 void FPS::shuffleStackTop(const unsigned char *FixStack,
943 MachineBasicBlock::iterator I) {
944 // Move items into place, starting from the desired stack bottom.
946 // Old register at position FixCount.
947 unsigned OldReg = getStackEntry(FixCount);
948 // Desired register at position FixCount.
949 unsigned Reg = FixStack[FixCount];
952 // (Reg st0) (OldReg st0) = (Reg OldReg st0)
954 moveToTop(OldReg, I);
960 //===----------------------------------------------------------------------===//
961 // Instruction transformation implementation
962 //===----------------------------------------------------------------------===//
964 /// handleZeroArgFP - ST(0) = fld0 ST(0) = flds <mem>
966 void FPS::handleZeroArgFP(MachineBasicBlock::iterator &I) {
967 MachineInstr *MI = I;
968 unsigned DestReg = getFPReg(MI->getOperand(0));
970 // Change from the pseudo instruction to the concrete instruction.
971 MI->RemoveOperand(0); // Remove the explicit ST(0) operand
972 MI->setDesc(TII->get(getConcreteOpcode(MI->getOpcode())));
974 // Result gets pushed on the stack.
978 /// handleOneArgFP - fst <mem>, ST(0)
980 void FPS::handleOneArgFP(MachineBasicBlock::iterator &I) {
981 MachineInstr *MI = I;
982 unsigned NumOps = MI->getDesc().getNumOperands();
983 assert((NumOps == X86::AddrNumOperands + 1 || NumOps == 1) &&
984 "Can only handle fst* & ftst instructions!");
986 // Is this the last use of the source register?
987 unsigned Reg = getFPReg(MI->getOperand(NumOps-1));
988 bool KillsSrc = MI->killsRegister(X86::FP0+Reg);
990 // FISTP64m is strange because there isn't a non-popping versions.
991 // If we have one _and_ we don't want to pop the operand, duplicate the value
992 // on the stack instead of moving it. This ensure that popping the value is
994 // Ditto FISTTP16m, FISTTP32m, FISTTP64m, ST_FpP80m.
997 (MI->getOpcode() == X86::IST_Fp64m32 ||
998 MI->getOpcode() == X86::ISTT_Fp16m32 ||
999 MI->getOpcode() == X86::ISTT_Fp32m32 ||
1000 MI->getOpcode() == X86::ISTT_Fp64m32 ||
1001 MI->getOpcode() == X86::IST_Fp64m64 ||
1002 MI->getOpcode() == X86::ISTT_Fp16m64 ||
1003 MI->getOpcode() == X86::ISTT_Fp32m64 ||
1004 MI->getOpcode() == X86::ISTT_Fp64m64 ||
1005 MI->getOpcode() == X86::IST_Fp64m80 ||
1006 MI->getOpcode() == X86::ISTT_Fp16m80 ||
1007 MI->getOpcode() == X86::ISTT_Fp32m80 ||
1008 MI->getOpcode() == X86::ISTT_Fp64m80 ||
1009 MI->getOpcode() == X86::ST_FpP80m)) {
1010 duplicateToTop(Reg, getScratchReg(), I);
1012 moveToTop(Reg, I); // Move to the top of the stack...
1015 // Convert from the pseudo instruction to the concrete instruction.
1016 MI->RemoveOperand(NumOps-1); // Remove explicit ST(0) operand
1017 MI->setDesc(TII->get(getConcreteOpcode(MI->getOpcode())));
1019 if (MI->getOpcode() == X86::IST_FP64m ||
1020 MI->getOpcode() == X86::ISTT_FP16m ||
1021 MI->getOpcode() == X86::ISTT_FP32m ||
1022 MI->getOpcode() == X86::ISTT_FP64m ||
1023 MI->getOpcode() == X86::ST_FP80m) {
1024 assert(StackTop > 0 && "Stack empty??");
1026 } else if (KillsSrc) { // Last use of operand?
1032 /// handleOneArgFPRW: Handle instructions that read from the top of stack and
1033 /// replace the value with a newly computed value. These instructions may have
1034 /// non-fp operands after their FP operands.
1038 /// R1 = fadd R2, [mem]
1040 void FPS::handleOneArgFPRW(MachineBasicBlock::iterator &I) {
1041 MachineInstr *MI = I;
1043 unsigned NumOps = MI->getDesc().getNumOperands();
1044 assert(NumOps >= 2 && "FPRW instructions must have 2 ops!!");
1047 // Is this the last use of the source register?
1048 unsigned Reg = getFPReg(MI->getOperand(1));
1049 bool KillsSrc = MI->killsRegister(X86::FP0+Reg);
1052 // If this is the last use of the source register, just make sure it's on
1053 // the top of the stack.
1055 assert(StackTop > 0 && "Stack cannot be empty!");
1057 pushReg(getFPReg(MI->getOperand(0)));
1059 // If this is not the last use of the source register, _copy_ it to the top
1061 duplicateToTop(Reg, getFPReg(MI->getOperand(0)), I);
1064 // Change from the pseudo instruction to the concrete instruction.
1065 MI->RemoveOperand(1); // Drop the source operand.
1066 MI->RemoveOperand(0); // Drop the destination operand.
1067 MI->setDesc(TII->get(getConcreteOpcode(MI->getOpcode())));
1071 //===----------------------------------------------------------------------===//
1072 // Define tables of various ways to map pseudo instructions
1075 // ForwardST0Table - Map: A = B op C into: ST(0) = ST(0) op ST(i)
1076 static const TableEntry ForwardST0Table[] = {
1077 { X86::ADD_Fp32 , X86::ADD_FST0r },
1078 { X86::ADD_Fp64 , X86::ADD_FST0r },
1079 { X86::ADD_Fp80 , X86::ADD_FST0r },
1080 { X86::DIV_Fp32 , X86::DIV_FST0r },
1081 { X86::DIV_Fp64 , X86::DIV_FST0r },
1082 { X86::DIV_Fp80 , X86::DIV_FST0r },
1083 { X86::MUL_Fp32 , X86::MUL_FST0r },
1084 { X86::MUL_Fp64 , X86::MUL_FST0r },
1085 { X86::MUL_Fp80 , X86::MUL_FST0r },
1086 { X86::SUB_Fp32 , X86::SUB_FST0r },
1087 { X86::SUB_Fp64 , X86::SUB_FST0r },
1088 { X86::SUB_Fp80 , X86::SUB_FST0r },
1091 // ReverseST0Table - Map: A = B op C into: ST(0) = ST(i) op ST(0)
1092 static const TableEntry ReverseST0Table[] = {
1093 { X86::ADD_Fp32 , X86::ADD_FST0r }, // commutative
1094 { X86::ADD_Fp64 , X86::ADD_FST0r }, // commutative
1095 { X86::ADD_Fp80 , X86::ADD_FST0r }, // commutative
1096 { X86::DIV_Fp32 , X86::DIVR_FST0r },
1097 { X86::DIV_Fp64 , X86::DIVR_FST0r },
1098 { X86::DIV_Fp80 , X86::DIVR_FST0r },
1099 { X86::MUL_Fp32 , X86::MUL_FST0r }, // commutative
1100 { X86::MUL_Fp64 , X86::MUL_FST0r }, // commutative
1101 { X86::MUL_Fp80 , X86::MUL_FST0r }, // commutative
1102 { X86::SUB_Fp32 , X86::SUBR_FST0r },
1103 { X86::SUB_Fp64 , X86::SUBR_FST0r },
1104 { X86::SUB_Fp80 , X86::SUBR_FST0r },
1107 // ForwardSTiTable - Map: A = B op C into: ST(i) = ST(0) op ST(i)
1108 static const TableEntry ForwardSTiTable[] = {
1109 { X86::ADD_Fp32 , X86::ADD_FrST0 }, // commutative
1110 { X86::ADD_Fp64 , X86::ADD_FrST0 }, // commutative
1111 { X86::ADD_Fp80 , X86::ADD_FrST0 }, // commutative
1112 { X86::DIV_Fp32 , X86::DIVR_FrST0 },
1113 { X86::DIV_Fp64 , X86::DIVR_FrST0 },
1114 { X86::DIV_Fp80 , X86::DIVR_FrST0 },
1115 { X86::MUL_Fp32 , X86::MUL_FrST0 }, // commutative
1116 { X86::MUL_Fp64 , X86::MUL_FrST0 }, // commutative
1117 { X86::MUL_Fp80 , X86::MUL_FrST0 }, // commutative
1118 { X86::SUB_Fp32 , X86::SUBR_FrST0 },
1119 { X86::SUB_Fp64 , X86::SUBR_FrST0 },
1120 { X86::SUB_Fp80 , X86::SUBR_FrST0 },
1123 // ReverseSTiTable - Map: A = B op C into: ST(i) = ST(i) op ST(0)
1124 static const TableEntry ReverseSTiTable[] = {
1125 { X86::ADD_Fp32 , X86::ADD_FrST0 },
1126 { X86::ADD_Fp64 , X86::ADD_FrST0 },
1127 { X86::ADD_Fp80 , X86::ADD_FrST0 },
1128 { X86::DIV_Fp32 , X86::DIV_FrST0 },
1129 { X86::DIV_Fp64 , X86::DIV_FrST0 },
1130 { X86::DIV_Fp80 , X86::DIV_FrST0 },
1131 { X86::MUL_Fp32 , X86::MUL_FrST0 },
1132 { X86::MUL_Fp64 , X86::MUL_FrST0 },
1133 { X86::MUL_Fp80 , X86::MUL_FrST0 },
1134 { X86::SUB_Fp32 , X86::SUB_FrST0 },
1135 { X86::SUB_Fp64 , X86::SUB_FrST0 },
1136 { X86::SUB_Fp80 , X86::SUB_FrST0 },
1140 /// handleTwoArgFP - Handle instructions like FADD and friends which are virtual
1141 /// instructions which need to be simplified and possibly transformed.
1143 /// Result: ST(0) = fsub ST(0), ST(i)
1144 /// ST(i) = fsub ST(0), ST(i)
1145 /// ST(0) = fsubr ST(0), ST(i)
1146 /// ST(i) = fsubr ST(0), ST(i)
1148 void FPS::handleTwoArgFP(MachineBasicBlock::iterator &I) {
1149 ASSERT_SORTED(ForwardST0Table); ASSERT_SORTED(ReverseST0Table);
1150 ASSERT_SORTED(ForwardSTiTable); ASSERT_SORTED(ReverseSTiTable);
1151 MachineInstr *MI = I;
1153 unsigned NumOperands = MI->getDesc().getNumOperands();
1154 assert(NumOperands == 3 && "Illegal TwoArgFP instruction!");
1155 unsigned Dest = getFPReg(MI->getOperand(0));
1156 unsigned Op0 = getFPReg(MI->getOperand(NumOperands-2));
1157 unsigned Op1 = getFPReg(MI->getOperand(NumOperands-1));
1158 bool KillsOp0 = MI->killsRegister(X86::FP0+Op0);
1159 bool KillsOp1 = MI->killsRegister(X86::FP0+Op1);
1160 DebugLoc dl = MI->getDebugLoc();
1162 unsigned TOS = getStackEntry(0);
1164 // One of our operands must be on the top of the stack. If neither is yet, we
1165 // need to move one.
1166 if (Op0 != TOS && Op1 != TOS) { // No operand at TOS?
1167 // We can choose to move either operand to the top of the stack. If one of
1168 // the operands is killed by this instruction, we want that one so that we
1169 // can update right on top of the old version.
1171 moveToTop(Op0, I); // Move dead operand to TOS.
1173 } else if (KillsOp1) {
1177 // All of the operands are live after this instruction executes, so we
1178 // cannot update on top of any operand. Because of this, we must
1179 // duplicate one of the stack elements to the top. It doesn't matter
1180 // which one we pick.
1182 duplicateToTop(Op0, Dest, I);
1186 } else if (!KillsOp0 && !KillsOp1) {
1187 // If we DO have one of our operands at the top of the stack, but we don't
1188 // have a dead operand, we must duplicate one of the operands to a new slot
1190 duplicateToTop(Op0, Dest, I);
1195 // Now we know that one of our operands is on the top of the stack, and at
1196 // least one of our operands is killed by this instruction.
1197 assert((TOS == Op0 || TOS == Op1) && (KillsOp0 || KillsOp1) &&
1198 "Stack conditions not set up right!");
1200 // We decide which form to use based on what is on the top of the stack, and
1201 // which operand is killed by this instruction.
1202 const TableEntry *InstTable;
1203 bool isForward = TOS == Op0;
1204 bool updateST0 = (TOS == Op0 && !KillsOp1) || (TOS == Op1 && !KillsOp0);
1207 InstTable = ForwardST0Table;
1209 InstTable = ReverseST0Table;
1212 InstTable = ForwardSTiTable;
1214 InstTable = ReverseSTiTable;
1217 int Opcode = Lookup(InstTable, array_lengthof(ForwardST0Table),
1219 assert(Opcode != -1 && "Unknown TwoArgFP pseudo instruction!");
1221 // NotTOS - The register which is not on the top of stack...
1222 unsigned NotTOS = (TOS == Op0) ? Op1 : Op0;
1224 // Replace the old instruction with a new instruction
1226 I = BuildMI(*MBB, I, dl, TII->get(Opcode)).addReg(getSTReg(NotTOS));
1228 // If both operands are killed, pop one off of the stack in addition to
1229 // overwriting the other one.
1230 if (KillsOp0 && KillsOp1 && Op0 != Op1) {
1231 assert(!updateST0 && "Should have updated other operand!");
1232 popStackAfter(I); // Pop the top of stack
1235 // Update stack information so that we know the destination register is now on
1237 unsigned UpdatedSlot = getSlot(updateST0 ? TOS : NotTOS);
1238 assert(UpdatedSlot < StackTop && Dest < 7);
1239 Stack[UpdatedSlot] = Dest;
1240 RegMap[Dest] = UpdatedSlot;
1241 MBB->getParent()->DeleteMachineInstr(MI); // Remove the old instruction
1244 /// handleCompareFP - Handle FUCOM and FUCOMI instructions, which have two FP
1245 /// register arguments and no explicit destinations.
1247 void FPS::handleCompareFP(MachineBasicBlock::iterator &I) {
1248 ASSERT_SORTED(ForwardST0Table); ASSERT_SORTED(ReverseST0Table);
1249 ASSERT_SORTED(ForwardSTiTable); ASSERT_SORTED(ReverseSTiTable);
1250 MachineInstr *MI = I;
1252 unsigned NumOperands = MI->getDesc().getNumOperands();
1253 assert(NumOperands == 2 && "Illegal FUCOM* instruction!");
1254 unsigned Op0 = getFPReg(MI->getOperand(NumOperands-2));
1255 unsigned Op1 = getFPReg(MI->getOperand(NumOperands-1));
1256 bool KillsOp0 = MI->killsRegister(X86::FP0+Op0);
1257 bool KillsOp1 = MI->killsRegister(X86::FP0+Op1);
1259 // Make sure the first operand is on the top of stack, the other one can be
1263 // Change from the pseudo instruction to the concrete instruction.
1264 MI->getOperand(0).setReg(getSTReg(Op1));
1265 MI->RemoveOperand(1);
1266 MI->setDesc(TII->get(getConcreteOpcode(MI->getOpcode())));
1268 // If any of the operands are killed by this instruction, free them.
1269 if (KillsOp0) freeStackSlotAfter(I, Op0);
1270 if (KillsOp1 && Op0 != Op1) freeStackSlotAfter(I, Op1);
1273 /// handleCondMovFP - Handle two address conditional move instructions. These
1274 /// instructions move a st(i) register to st(0) iff a condition is true. These
1275 /// instructions require that the first operand is at the top of the stack, but
1276 /// otherwise don't modify the stack at all.
1277 void FPS::handleCondMovFP(MachineBasicBlock::iterator &I) {
1278 MachineInstr *MI = I;
1280 unsigned Op0 = getFPReg(MI->getOperand(0));
1281 unsigned Op1 = getFPReg(MI->getOperand(2));
1282 bool KillsOp1 = MI->killsRegister(X86::FP0+Op1);
1284 // The first operand *must* be on the top of the stack.
1287 // Change the second operand to the stack register that the operand is in.
1288 // Change from the pseudo instruction to the concrete instruction.
1289 MI->RemoveOperand(0);
1290 MI->RemoveOperand(1);
1291 MI->getOperand(0).setReg(getSTReg(Op1));
1292 MI->setDesc(TII->get(getConcreteOpcode(MI->getOpcode())));
1294 // If we kill the second operand, make sure to pop it from the stack.
1295 if (Op0 != Op1 && KillsOp1) {
1296 // Get this value off of the register stack.
1297 freeStackSlotAfter(I, Op1);
1302 /// handleSpecialFP - Handle special instructions which behave unlike other
1303 /// floating point instructions. This is primarily intended for use by pseudo
1306 void FPS::handleSpecialFP(MachineBasicBlock::iterator &I) {
1307 MachineInstr *MI = I;
1308 DebugLoc dl = MI->getDebugLoc();
1309 switch (MI->getOpcode()) {
1310 default: llvm_unreachable("Unknown SpecialFP instruction!");
1311 case X86::FpGET_ST0_32:// Appears immediately after a call returning FP type!
1312 case X86::FpGET_ST0_64:// Appears immediately after a call returning FP type!
1313 case X86::FpGET_ST0_80:// Appears immediately after a call returning FP type!
1314 assert(StackTop == 0 && "Stack should be empty after a call!");
1315 pushReg(getFPReg(MI->getOperand(0)));
1317 case X86::FpGET_ST1_32:// Appears immediately after a call returning FP type!
1318 case X86::FpGET_ST1_64:// Appears immediately after a call returning FP type!
1319 case X86::FpGET_ST1_80:{// Appears immediately after a call returning FP type!
1320 // FpGET_ST1 should occur right after a FpGET_ST0 for a call or inline asm.
1321 // The pattern we expect is:
1326 // At this point, we've pushed FP1 on the top of stack, so it should be
1327 // present if it isn't dead. If it was dead, we already emitted a pop to
1328 // remove it from the stack and StackTop = 0.
1330 // Push FP4 as top of stack next.
1331 pushReg(getFPReg(MI->getOperand(0)));
1333 // If StackTop was 0 before we pushed our operand, then ST(0) must have been
1334 // dead. In this case, the ST(1) value is the only thing that is live, so
1335 // it should be on the TOS (after the pop that was emitted) and is. Just
1336 // continue in this case.
1340 // Because pushReg just pushed ST(1) as TOS, we now have to swap the two top
1341 // elements so that our accounting is correct.
1342 unsigned RegOnTop = getStackEntry(0);
1343 unsigned RegNo = getStackEntry(1);
1345 // Swap the slots the regs are in.
1346 std::swap(RegMap[RegNo], RegMap[RegOnTop]);
1348 // Swap stack slot contents.
1349 assert(RegMap[RegOnTop] < StackTop);
1350 std::swap(Stack[RegMap[RegOnTop]], Stack[StackTop-1]);
1353 case X86::FpSET_ST0_32:
1354 case X86::FpSET_ST0_64:
1355 case X86::FpSET_ST0_80: {
1356 // FpSET_ST0_80 is generated by copyRegToReg for setting up inline asm
1357 // arguments that use an st constraint. We expect a sequence of
1358 // instructions: Fp_SET_ST0 Fp_SET_ST1? INLINEASM
1359 unsigned Op0 = getFPReg(MI->getOperand(0));
1361 if (!MI->killsRegister(X86::FP0 + Op0)) {
1362 // Duplicate Op0 into a temporary on the stack top.
1363 duplicateToTop(Op0, getScratchReg(), I);
1365 // Op0 is killed, so just swap it into position.
1368 --StackTop; // "Forget" we have something on the top of stack!
1371 case X86::FpSET_ST1_32:
1372 case X86::FpSET_ST1_64:
1373 case X86::FpSET_ST1_80: {
1374 // Set up st(1) for inline asm. We are assuming that st(0) has already been
1375 // set up by FpSET_ST0, and our StackTop is off by one because of it.
1376 unsigned Op0 = getFPReg(MI->getOperand(0));
1377 // Restore the actual StackTop from before Fp_SET_ST0.
1378 // Note we can't handle Fp_SET_ST1 without a preceeding Fp_SET_ST0, and we
1379 // are not enforcing the constraint.
1381 unsigned RegOnTop = getStackEntry(0); // This reg must remain in st(0).
1382 if (!MI->killsRegister(X86::FP0 + Op0)) {
1383 duplicateToTop(Op0, getScratchReg(), I);
1384 moveToTop(RegOnTop, I);
1385 } else if (getSTReg(Op0) != X86::ST1) {
1386 // We have the wrong value at st(1). Shuffle! Untested!
1387 moveToTop(getStackEntry(1), I);
1389 moveToTop(RegOnTop, I);
1391 assert(StackTop >= 2 && "Too few live registers");
1392 StackTop -= 2; // "Forget" both st(0) and st(1).
1395 case X86::MOV_Fp3232:
1396 case X86::MOV_Fp3264:
1397 case X86::MOV_Fp6432:
1398 case X86::MOV_Fp6464:
1399 case X86::MOV_Fp3280:
1400 case X86::MOV_Fp6480:
1401 case X86::MOV_Fp8032:
1402 case X86::MOV_Fp8064:
1403 case X86::MOV_Fp8080: {
1404 const MachineOperand &MO1 = MI->getOperand(1);
1405 unsigned SrcReg = getFPReg(MO1);
1407 const MachineOperand &MO0 = MI->getOperand(0);
1408 unsigned DestReg = getFPReg(MO0);
1409 if (MI->killsRegister(X86::FP0+SrcReg)) {
1410 // If the input operand is killed, we can just change the owner of the
1411 // incoming stack slot into the result.
1412 unsigned Slot = getSlot(SrcReg);
1413 assert(Slot < 7 && DestReg < 7 && "FpMOV operands invalid!");
1414 Stack[Slot] = DestReg;
1415 RegMap[DestReg] = Slot;
1418 // For FMOV we just duplicate the specified value to a new stack slot.
1419 // This could be made better, but would require substantial changes.
1420 duplicateToTop(SrcReg, DestReg, I);
1424 case TargetOpcode::INLINEASM: {
1425 // The inline asm MachineInstr currently only *uses* FP registers for the
1426 // 'f' constraint. These should be turned into the current ST(x) register
1427 // in the machine instr. Also, any kills should be explicitly popped after
1430 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
1431 MachineOperand &Op = MI->getOperand(i);
1432 if (!Op.isReg() || Op.getReg() < X86::FP0 || Op.getReg() > X86::FP6)
1434 assert(Op.isUse() && "Only handle inline asm uses right now");
1436 unsigned FPReg = getFPReg(Op);
1437 Op.setReg(getSTReg(FPReg));
1439 // If we kill this operand, make sure to pop it from the stack after the
1440 // asm. We just remember it for now, and pop them all off at the end in
1443 Kills |= 1U << FPReg;
1446 // If this asm kills any FP registers (is the last use of them) we must
1447 // explicitly emit pop instructions for them. Do this now after the asm has
1448 // executed so that the ST(x) numbers are not off (which would happen if we
1449 // did this inline with operand rewriting).
1451 // Note: this might be a non-optimal pop sequence. We might be able to do
1452 // better by trying to pop in stack order or something.
1453 MachineBasicBlock::iterator InsertPt = MI;
1455 unsigned FPReg = CountTrailingZeros_32(Kills);
1456 freeStackSlotAfter(InsertPt, FPReg);
1457 Kills &= ~(1U << FPReg);
1459 // Don't delete the inline asm!
1465 // If RET has an FP register use operand, pass the first one in ST(0) and
1466 // the second one in ST(1).
1468 // Find the register operands.
1469 unsigned FirstFPRegOp = ~0U, SecondFPRegOp = ~0U;
1470 unsigned LiveMask = 0;
1472 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
1473 MachineOperand &Op = MI->getOperand(i);
1474 if (!Op.isReg() || Op.getReg() < X86::FP0 || Op.getReg() > X86::FP6)
1476 // FP Register uses must be kills unless there are two uses of the same
1477 // register, in which case only one will be a kill.
1478 assert(Op.isUse() &&
1479 (Op.isKill() || // Marked kill.
1480 getFPReg(Op) == FirstFPRegOp || // Second instance.
1481 MI->killsRegister(Op.getReg())) && // Later use is marked kill.
1482 "Ret only defs operands, and values aren't live beyond it");
1484 if (FirstFPRegOp == ~0U)
1485 FirstFPRegOp = getFPReg(Op);
1487 assert(SecondFPRegOp == ~0U && "More than two fp operands!");
1488 SecondFPRegOp = getFPReg(Op);
1490 LiveMask |= (1 << getFPReg(Op));
1492 // Remove the operand so that later passes don't see it.
1493 MI->RemoveOperand(i);
1497 // We may have been carrying spurious live-ins, so make sure only the returned
1498 // registers are left live.
1499 adjustLiveRegs(LiveMask, MI);
1500 if (!LiveMask) return; // Quick check to see if any are possible.
1502 // There are only four possibilities here:
1503 // 1) we are returning a single FP value. In this case, it has to be in
1504 // ST(0) already, so just declare success by removing the value from the
1506 if (SecondFPRegOp == ~0U) {
1507 // Assert that the top of stack contains the right FP register.
1508 assert(StackTop == 1 && FirstFPRegOp == getStackEntry(0) &&
1509 "Top of stack not the right register for RET!");
1511 // Ok, everything is good, mark the value as not being on the stack
1512 // anymore so that our assertion about the stack being empty at end of
1513 // block doesn't fire.
1518 // Otherwise, we are returning two values:
1519 // 2) If returning the same value for both, we only have one thing in the FP
1520 // stack. Consider: RET FP1, FP1
1521 if (StackTop == 1) {
1522 assert(FirstFPRegOp == SecondFPRegOp && FirstFPRegOp == getStackEntry(0)&&
1523 "Stack misconfiguration for RET!");
1525 // Duplicate the TOS so that we return it twice. Just pick some other FPx
1526 // register to hold it.
1527 unsigned NewReg = getScratchReg();
1528 duplicateToTop(FirstFPRegOp, NewReg, MI);
1529 FirstFPRegOp = NewReg;
1532 /// Okay we know we have two different FPx operands now:
1533 assert(StackTop == 2 && "Must have two values live!");
1535 /// 3) If SecondFPRegOp is currently in ST(0) and FirstFPRegOp is currently
1536 /// in ST(1). In this case, emit an fxch.
1537 if (getStackEntry(0) == SecondFPRegOp) {
1538 assert(getStackEntry(1) == FirstFPRegOp && "Unknown regs live");
1539 moveToTop(FirstFPRegOp, MI);
1542 /// 4) Finally, FirstFPRegOp must be in ST(0) and SecondFPRegOp must be in
1543 /// ST(1). Just remove both from our understanding of the stack and return.
1544 assert(getStackEntry(0) == FirstFPRegOp && "Unknown regs live");
1545 assert(getStackEntry(1) == SecondFPRegOp && "Unknown regs live");
1550 I = MBB->erase(I); // Remove the pseudo instruction
1552 // We want to leave I pointing to the previous instruction, but what if we
1553 // just erased the first instruction?
1554 if (I == MBB->begin()) {
1555 DEBUG(dbgs() << "Inserting dummy KILL\n");
1556 I = BuildMI(*MBB, I, DebugLoc(), TII->get(TargetOpcode::KILL));
1561 // Translate a COPY instruction to a pseudo-op that handleSpecialFP understands.
1562 bool FPS::translateCopy(MachineInstr *MI) {
1563 unsigned DstReg = MI->getOperand(0).getReg();
1564 unsigned SrcReg = MI->getOperand(1).getReg();
1566 if (DstReg == X86::ST0) {
1567 MI->setDesc(TII->get(X86::FpSET_ST0_80));
1568 MI->RemoveOperand(0);
1571 if (DstReg == X86::ST1) {
1572 MI->setDesc(TII->get(X86::FpSET_ST1_80));
1573 MI->RemoveOperand(0);
1576 if (SrcReg == X86::ST0) {
1577 MI->setDesc(TII->get(X86::FpGET_ST0_80));
1580 if (SrcReg == X86::ST1) {
1581 MI->setDesc(TII->get(X86::FpGET_ST1_80));
1584 if (X86::RFP80RegClass.contains(DstReg, SrcReg)) {
1585 MI->setDesc(TII->get(X86::MOV_Fp8080));