1 //===-- X86FloatingPoint.cpp - Floating point Reg -> Stack converter ------===// 2 // 3 // The LLVM Compiler Infrastructure 4 // 5 // This file is distributed under the University of Illinois Open Source 6 // License. See LICENSE.TXT for details. 7 // 8 //===----------------------------------------------------------------------===// 9 // 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 13 // lifetimes. 14 // 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. 20 // 21 // This approach avoids splitting critical edges at the potential cost of more 22 // live register shuffling instructions when critical edges are present. 23 // 24 //===----------------------------------------------------------------------===// 25 26 #define DEBUG_TYPE "x86-codegen" 27 #include "X86.h" 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" 45 #include <algorithm> 46 using namespace llvm; 47 48 STATISTIC(NumFXCH, "Number of fxch instructions inserted"); 49 STATISTIC(NumFP , "Number of floating point instructions"); 50 51 namespace { 52 struct FPS : public MachineFunctionPass { 53 static char ID; 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)); 60 } 61 62 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 63 AU.setPreservesCFG(); 64 AU.addRequired<EdgeBundles>(); 65 AU.addPreservedID(MachineLoopInfoID); 66 AU.addPreservedID(MachineDominatorsID); 67 MachineFunctionPass::getAnalysisUsage(AU); 68 } 69 70 virtual bool runOnMachineFunction(MachineFunction &MF); 71 72 virtual const char *getPassName() const { return "X86 FP Stackifier"; } 73 74 private: 75 const TargetInstrInfo *TII; // Machine instruction info. 76 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. 81 // 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 84 // present. 85 // 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. 88 struct LiveBundle { 89 // Bit mask of live FP registers. Bit 0 = FP0, bit 1 = FP1, &c. 90 unsigned Mask; 91 92 // Number of pre-assigned live registers in FixStack. This is 0 when the 93 // stack order has not yet been fixed. 94 unsigned FixCount; 95 96 // Assigned stack order for live-in registers. 97 // FixStack[i] == getStackEntry(i) for all i < FixCount. 98 unsigned char FixStack[8]; 99 100 LiveBundle() : Mask(0), FixCount(0) {} 101 102 // Have the live registers been assigned a stack order yet? 103 bool isFixed() const { return !Mask || FixCount; } 104 }; 105 106 // Numbered LiveBundle structs. LiveBundles[0] is used for all CFG edges 107 // with no live FP registers. 108 SmallVector<LiveBundle, 8> LiveBundles; 109 110 // The edge bundle analysis provides indices into the LiveBundles vector. 111 EdgeBundles *Bundles; 112 113 // Return a bitmask of FP registers in block's live-in list. 114 static unsigned calcLiveInMask(MachineBasicBlock *MBB) { 115 unsigned Mask = 0; 116 for (MachineBasicBlock::livein_iterator I = MBB->livein_begin(), 117 E = MBB->livein_end(); I != E; ++I) { 118 unsigned Reg = *I - X86::FP0; 119 if (Reg < 8) 120 Mask |= 1 << Reg; 121 } 122 return Mask; 123 } 124 125 // Partition all the CFG edges into LiveBundles. 126 void bundleCFG(MachineFunction &MF); 127 128 MachineBasicBlock *MBB; // Current basic block 129 130 // The hardware keeps track of how many FP registers are live, so we have 131 // to model that exactly. Usually, each live register corresponds to an 132 // FP<n> register, but when dealing with calls, returns, and inline 133 // assembly, it is sometimes necessary to have live scratch registers. 134 unsigned Stack[8]; // FP<n> Registers in each stack slot... 135 unsigned StackTop; // The current top of the FP stack. 136 137 enum { 138 NumFPRegs = 16 // Including scratch pseudo-registers. 139 }; 140 141 // For each live FP<n> register, point to its Stack[] entry. 142 // The first entries correspond to FP0-FP6, the rest are scratch registers 143 // used when we need slightly different live registers than what the 144 // register allocator thinks. 145 unsigned RegMap[NumFPRegs]; 146 147 // Pending fixed registers - Inline assembly needs FP registers to appear 148 // in fixed stack slot positions. This is handled by copying FP registers 149 // to ST registers before the instruction, and copying back after the 150 // instruction. 151 // 152 // This is modeled with pending ST registers. NumPendingSTs is the number 153 // of ST registers (ST0-STn) we are tracking. PendingST[n] points to an FP 154 // register that holds the ST value. The ST registers are not moved into 155 // place until immediately before the instruction that needs them. 156 // 157 // It can happen that we need an ST register to be live when no FP register 158 // holds the value: 159 // 160 // %ST0 = COPY %FP4<kill> 161 // 162 // When that happens, we allocate a scratch FP register to hold the ST 163 // value. That means every register in PendingST must be live. 164 165 unsigned NumPendingSTs; 166 unsigned char PendingST[8]; 167 168 // Set up our stack model to match the incoming registers to MBB. 169 void setupBlockStack(); 170 171 // Shuffle live registers to match the expectations of successor blocks. 172 void finishBlockStack(); 173 174 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) 175 void dumpStack() const { 176 dbgs() << "Stack contents:"; 177 for (unsigned i = 0; i != StackTop; ++i) { 178 dbgs() << " FP" << Stack[i]; 179 assert(RegMap[Stack[i]] == i && "Stack[] doesn't match RegMap[]!"); 180 } 181 for (unsigned i = 0; i != NumPendingSTs; ++i) 182 dbgs() << ", ST" << i << " in FP" << unsigned(PendingST[i]); 183 dbgs() << "\n"; 184 } 185 #endif 186 187 /// getSlot - Return the stack slot number a particular register number is 188 /// in. 189 unsigned getSlot(unsigned RegNo) const { 190 assert(RegNo < NumFPRegs && "Regno out of range!"); 191 return RegMap[RegNo]; 192 } 193 194 /// isLive - Is RegNo currently live in the stack? 195 bool isLive(unsigned RegNo) const { 196 unsigned Slot = getSlot(RegNo); 197 return Slot < StackTop && Stack[Slot] == RegNo; 198 } 199 200 /// getScratchReg - Return an FP register that is not currently in use. 201 unsigned getScratchReg() const { 202 for (int i = NumFPRegs - 1; i >= 8; --i) 203 if (!isLive(i)) 204 return i; 205 llvm_unreachable("Ran out of scratch FP registers"); 206 } 207 208 /// isScratchReg - Returns trus if RegNo is a scratch FP register. 209 static bool isScratchReg(unsigned RegNo) { 210 return RegNo > 8 && RegNo < NumFPRegs; 211 } 212 213 /// getStackEntry - Return the X86::FP<n> register in register ST(i). 214 unsigned getStackEntry(unsigned STi) const { 215 if (STi >= StackTop) 216 report_fatal_error("Access past stack top!"); 217 return Stack[StackTop-1-STi]; 218 } 219 220 /// getSTReg - Return the X86::ST(i) register which contains the specified 221 /// FP<RegNo> register. 222 unsigned getSTReg(unsigned RegNo) const { 223 return StackTop - 1 - getSlot(RegNo) + X86::ST0; 224 } 225 226 // pushReg - Push the specified FP<n> register onto the stack. 227 void pushReg(unsigned Reg) { 228 assert(Reg < NumFPRegs && "Register number out of range!"); 229 if (StackTop >= 8) 230 report_fatal_error("Stack overflow!"); 231 Stack[StackTop] = Reg; 232 RegMap[Reg] = StackTop++; 233 } 234 235 bool isAtTop(unsigned RegNo) const { return getSlot(RegNo) == StackTop-1; } 236 void moveToTop(unsigned RegNo, MachineBasicBlock::iterator I) { 237 DebugLoc dl = I == MBB->end() ? DebugLoc() : I->getDebugLoc(); 238 if (isAtTop(RegNo)) return; 239 240 unsigned STReg = getSTReg(RegNo); 241 unsigned RegOnTop = getStackEntry(0); 242 243 // Swap the slots the regs are in. 244 std::swap(RegMap[RegNo], RegMap[RegOnTop]); 245 246 // Swap stack slot contents. 247 if (RegMap[RegOnTop] >= StackTop) 248 report_fatal_error("Access past stack top!"); 249 std::swap(Stack[RegMap[RegOnTop]], Stack[StackTop-1]); 250 251 // Emit an fxch to update the runtime processors version of the state. 252 BuildMI(*MBB, I, dl, TII->get(X86::XCH_F)).addReg(STReg); 253 ++NumFXCH; 254 } 255 256 void duplicateToTop(unsigned RegNo, unsigned AsReg, MachineInstr *I) { 257 DebugLoc dl = I == MBB->end() ? DebugLoc() : I->getDebugLoc(); 258 unsigned STReg = getSTReg(RegNo); 259 pushReg(AsReg); // New register on top of stack 260 261 BuildMI(*MBB, I, dl, TII->get(X86::LD_Frr)).addReg(STReg); 262 } 263 264 /// duplicatePendingSTBeforeKill - The instruction at I is about to kill 265 /// RegNo. If any PendingST registers still need the RegNo value, duplicate 266 /// them to new scratch registers. 267 void duplicatePendingSTBeforeKill(unsigned RegNo, MachineInstr *I) { 268 for (unsigned i = 0; i != NumPendingSTs; ++i) { 269 if (PendingST[i] != RegNo) 270 continue; 271 unsigned SR = getScratchReg(); 272 DEBUG(dbgs() << "Duplicating pending ST" << i 273 << " in FP" << RegNo << " to FP" << SR << '\n'); 274 duplicateToTop(RegNo, SR, I); 275 PendingST[i] = SR; 276 } 277 } 278 279 /// popStackAfter - Pop the current value off of the top of the FP stack 280 /// after the specified instruction. 281 void popStackAfter(MachineBasicBlock::iterator &I); 282 283 /// freeStackSlotAfter - Free the specified register from the register 284 /// stack, so that it is no longer in a register. If the register is 285 /// currently at the top of the stack, we just pop the current instruction, 286 /// otherwise we store the current top-of-stack into the specified slot, 287 /// then pop the top of stack. 288 void freeStackSlotAfter(MachineBasicBlock::iterator &I, unsigned Reg); 289 290 /// freeStackSlotBefore - Just the pop, no folding. Return the inserted 291 /// instruction. 292 MachineBasicBlock::iterator 293 freeStackSlotBefore(MachineBasicBlock::iterator I, unsigned FPRegNo); 294 295 /// Adjust the live registers to be the set in Mask. 296 void adjustLiveRegs(unsigned Mask, MachineBasicBlock::iterator I); 297 298 /// Shuffle the top FixCount stack entries such that FP reg FixStack[0] is 299 /// st(0), FP reg FixStack[1] is st(1) etc. 300 void shuffleStackTop(const unsigned char *FixStack, unsigned FixCount, 301 MachineBasicBlock::iterator I); 302 303 bool processBasicBlock(MachineFunction &MF, MachineBasicBlock &MBB); 304 305 void handleZeroArgFP(MachineBasicBlock::iterator &I); 306 void handleOneArgFP(MachineBasicBlock::iterator &I); 307 void handleOneArgFPRW(MachineBasicBlock::iterator &I); 308 void handleTwoArgFP(MachineBasicBlock::iterator &I); 309 void handleCompareFP(MachineBasicBlock::iterator &I); 310 void handleCondMovFP(MachineBasicBlock::iterator &I); 311 void handleSpecialFP(MachineBasicBlock::iterator &I); 312 313 // Check if a COPY instruction is using FP registers. 314 static bool isFPCopy(MachineInstr *MI) { 315 unsigned DstReg = MI->getOperand(0).getReg(); 316 unsigned SrcReg = MI->getOperand(1).getReg(); 317 318 return X86::RFP80RegClass.contains(DstReg) || 319 X86::RFP80RegClass.contains(SrcReg); 320 } 321 }; 322 char FPS::ID = 0; 323 } 324 325 FunctionPass *llvm::createX86FloatingPointStackifierPass() { return new FPS(); } 326 327 /// getFPReg - Return the X86::FPx register number for the specified operand. 328 /// For example, this returns 3 for X86::FP3. 329 static unsigned getFPReg(const MachineOperand &MO) { 330 assert(MO.isReg() && "Expected an FP register!"); 331 unsigned Reg = MO.getReg(); 332 assert(Reg >= X86::FP0 && Reg <= X86::FP6 && "Expected FP register!"); 333 return Reg - X86::FP0; 334 } 335 336 /// runOnMachineFunction - Loop over all of the basic blocks, transforming FP 337 /// register references into FP stack references. 338 /// 339 bool FPS::runOnMachineFunction(MachineFunction &MF) { 340 // We only need to run this pass if there are any FP registers used in this 341 // function. If it is all integer, there is nothing for us to do! 342 bool FPIsUsed = false; 343 344 assert(X86::FP6 == X86::FP0+6 && "Register enums aren't sorted right!"); 345 for (unsigned i = 0; i <= 6; ++i) 346 if (MF.getRegInfo().isPhysRegUsed(X86::FP0+i)) { 347 FPIsUsed = true; 348 break; 349 } 350 351 // Early exit. 352 if (!FPIsUsed) return false; 353 354 Bundles = &getAnalysis<EdgeBundles>(); 355 TII = MF.getTarget().getInstrInfo(); 356 357 // Prepare cross-MBB liveness. 358 bundleCFG(MF); 359 360 StackTop = 0; 361 362 // Process the function in depth first order so that we process at least one 363 // of the predecessors for every reachable block in the function. 364 SmallPtrSet<MachineBasicBlock*, 8> Processed; 365 MachineBasicBlock *Entry = MF.begin(); 366 367 bool Changed = false; 368 for (df_ext_iterator<MachineBasicBlock*, SmallPtrSet<MachineBasicBlock*, 8> > 369 I = df_ext_begin(Entry, Processed), E = df_ext_end(Entry, Processed); 370 I != E; ++I) 371 Changed |= processBasicBlock(MF, **I); 372 373 // Process any unreachable blocks in arbitrary order now. 374 if (MF.size() != Processed.size()) 375 for (MachineFunction::iterator BB = MF.begin(), E = MF.end(); BB != E; ++BB) 376 if (Processed.insert(BB)) 377 Changed |= processBasicBlock(MF, *BB); 378 379 LiveBundles.clear(); 380 381 return Changed; 382 } 383 384 /// bundleCFG - Scan all the basic blocks to determine consistent live-in and 385 /// live-out sets for the FP registers. Consistent means that the set of 386 /// registers live-out from a block is identical to the live-in set of all 387 /// successors. This is not enforced by the normal live-in lists since 388 /// registers may be implicitly defined, or not used by all successors. 389 void FPS::bundleCFG(MachineFunction &MF) { 390 assert(LiveBundles.empty() && "Stale data in LiveBundles"); 391 LiveBundles.resize(Bundles->getNumBundles()); 392 393 // Gather the actual live-in masks for all MBBs. 394 for (MachineFunction::iterator I = MF.begin(), E = MF.end(); I != E; ++I) { 395 MachineBasicBlock *MBB = I; 396 const unsigned Mask = calcLiveInMask(MBB); 397 if (!Mask) 398 continue; 399 // Update MBB ingoing bundle mask. 400 LiveBundles[Bundles->getBundle(MBB->getNumber(), false)].Mask |= Mask; 401 } 402 } 403 404 /// processBasicBlock - Loop over all of the instructions in the basic block, 405 /// transforming FP instructions into their stack form. 406 /// 407 bool FPS::processBasicBlock(MachineFunction &MF, MachineBasicBlock &BB) { 408 bool Changed = false; 409 MBB = &BB; 410 NumPendingSTs = 0; 411 412 setupBlockStack(); 413 414 for (MachineBasicBlock::iterator I = BB.begin(); I != BB.end(); ++I) { 415 MachineInstr *MI = I; 416 uint64_t Flags = MI->getDesc().TSFlags; 417 418 unsigned FPInstClass = Flags & X86II::FPTypeMask; 419 if (MI->isInlineAsm()) 420 FPInstClass = X86II::SpecialFP; 421 422 if (MI->isCopy() && isFPCopy(MI)) 423 FPInstClass = X86II::SpecialFP; 424 425 if (MI->isImplicitDef() && 426 X86::RFP80RegClass.contains(MI->getOperand(0).getReg())) 427 FPInstClass = X86II::SpecialFP; 428 429 if (FPInstClass == X86II::NotFP) 430 continue; // Efficiently ignore non-fp insts! 431 432 MachineInstr *PrevMI = 0; 433 if (I != BB.begin()) 434 PrevMI = prior(I); 435 436 ++NumFP; // Keep track of # of pseudo instrs 437 DEBUG(dbgs() << "\nFPInst:\t" << *MI); 438 439 // Get dead variables list now because the MI pointer may be deleted as part 440 // of processing! 441 SmallVector<unsigned, 8> DeadRegs; 442 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) { 443 const MachineOperand &MO = MI->getOperand(i); 444 if (MO.isReg() && MO.isDead()) 445 DeadRegs.push_back(MO.getReg()); 446 } 447 448 switch (FPInstClass) { 449 case X86II::ZeroArgFP: handleZeroArgFP(I); break; 450 case X86II::OneArgFP: handleOneArgFP(I); break; // fstp ST(0) 451 case X86II::OneArgFPRW: handleOneArgFPRW(I); break; // ST(0) = fsqrt(ST(0)) 452 case X86II::TwoArgFP: handleTwoArgFP(I); break; 453 case X86II::CompareFP: handleCompareFP(I); break; 454 case X86II::CondMovFP: handleCondMovFP(I); break; 455 case X86II::SpecialFP: handleSpecialFP(I); break; 456 default: llvm_unreachable("Unknown FP Type!"); 457 } 458 459 // Check to see if any of the values defined by this instruction are dead 460 // after definition. If so, pop them. 461 for (unsigned i = 0, e = DeadRegs.size(); i != e; ++i) { 462 unsigned Reg = DeadRegs[i]; 463 if (Reg >= X86::FP0 && Reg <= X86::FP6) { 464 DEBUG(dbgs() << "Register FP#" << Reg-X86::FP0 << " is dead!\n"); 465 freeStackSlotAfter(I, Reg-X86::FP0); 466 } 467 } 468 469 // Print out all of the instructions expanded to if -debug 470 DEBUG( 471 MachineBasicBlock::iterator PrevI(PrevMI); 472 if (I == PrevI) { 473 dbgs() << "Just deleted pseudo instruction\n"; 474 } else { 475 MachineBasicBlock::iterator Start = I; 476 // Rewind to first instruction newly inserted. 477 while (Start != BB.begin() && prior(Start) != PrevI) --Start; 478 dbgs() << "Inserted instructions:\n\t"; 479 Start->print(dbgs(), &MF.getTarget()); 480 while (++Start != llvm::next(I)) {} 481 } 482 dumpStack(); 483 ); 484 (void)PrevMI; 485 486 Changed = true; 487 } 488 489 finishBlockStack(); 490 491 return Changed; 492 } 493 494 /// setupBlockStack - Use the live bundles to set up our model of the stack 495 /// to match predecessors' live out stack. 496 void FPS::setupBlockStack() { 497 DEBUG(dbgs() << "\nSetting up live-ins for BB#" << MBB->getNumber() 498 << " derived from " << MBB->getName() << ".\n"); 499 StackTop = 0; 500 // Get the live-in bundle for MBB. 501 const LiveBundle &Bundle = 502 LiveBundles[Bundles->getBundle(MBB->getNumber(), false)]; 503 504 if (!Bundle.Mask) { 505 DEBUG(dbgs() << "Block has no FP live-ins.\n"); 506 return; 507 } 508 509 // Depth-first iteration should ensure that we always have an assigned stack. 510 assert(Bundle.isFixed() && "Reached block before any predecessors"); 511 512 // Push the fixed live-in registers. 513 for (unsigned i = Bundle.FixCount; i > 0; --i) { 514 MBB->addLiveIn(X86::ST0+i-1); 515 DEBUG(dbgs() << "Live-in st(" << (i-1) << "): %FP" 516 << unsigned(Bundle.FixStack[i-1]) << '\n'); 517 pushReg(Bundle.FixStack[i-1]); 518 } 519 520 // Kill off unwanted live-ins. This can happen with a critical edge. 521 // FIXME: We could keep these live registers around as zombies. They may need 522 // to be revived at the end of a short block. It might save a few instrs. 523 adjustLiveRegs(calcLiveInMask(MBB), MBB->begin()); 524 DEBUG(MBB->dump()); 525 } 526 527 /// finishBlockStack - Revive live-outs that are implicitly defined out of 528 /// MBB. Shuffle live registers to match the expected fixed stack of any 529 /// predecessors, and ensure that all predecessors are expecting the same 530 /// stack. 531 void FPS::finishBlockStack() { 532 // The RET handling below takes care of return blocks for us. 533 if (MBB->succ_empty()) 534 return; 535 536 DEBUG(dbgs() << "Setting up live-outs for BB#" << MBB->getNumber() 537 << " derived from " << MBB->getName() << ".\n"); 538 539 // Get MBB's live-out bundle. 540 unsigned BundleIdx = Bundles->getBundle(MBB->getNumber(), true); 541 LiveBundle &Bundle = LiveBundles[BundleIdx]; 542 543 // 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); 547 548 if (!Bundle.Mask) { 549 DEBUG(dbgs() << "No live-outs.\n"); 550 return; 551 } 552 553 // 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); 558 } else { 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); 564 } 565 } 566 567 568 //===----------------------------------------------------------------------===// 569 // Efficient Lookup Table Support 570 //===----------------------------------------------------------------------===// 571 572 namespace { 573 struct TableEntry { 574 uint16_t from; 575 uint16_t to; 576 bool operator<(const TableEntry &TE) const { return from < TE.from; } 577 friend bool operator<(const TableEntry &TE, unsigned V) { 578 return TE.from < V; 579 } 580 friend bool LLVM_ATTRIBUTE_UNUSED operator<(unsigned V, 581 const TableEntry &TE) { 582 return V < TE.from; 583 } 584 }; 585 } 586 587 #ifndef NDEBUG 588 static bool TableIsSorted(const TableEntry *Table, unsigned NumEntries) { 589 for (unsigned i = 0; i != NumEntries-1; ++i) 590 if (!(Table[i] < Table[i+1])) return false; 591 return true; 592 } 593 #endif 594 595 static int Lookup(const TableEntry *Table, unsigned N, unsigned Opcode) { 596 const TableEntry *I = std::lower_bound(Table, Table+N, Opcode); 597 if (I != Table+N && I->from == Opcode) 598 return I->to; 599 return -1; 600 } 601 602 #ifdef NDEBUG 603 #define ASSERT_SORTED(TABLE) 604 #else 605 #define ASSERT_SORTED(TABLE) \ 606 { static bool TABLE##Checked = false; \ 607 if (!TABLE##Checked) { \ 608 assert(TableIsSorted(TABLE, array_lengthof(TABLE)) && \ 609 "All lookup tables must be sorted for efficient access!"); \ 610 TABLE##Checked = true; \ 611 } \ 612 } 613 #endif 614 615 //===----------------------------------------------------------------------===// 616 // Register File -> Register Stack Mapping Methods 617 //===----------------------------------------------------------------------===// 618 619 // OpcodeTable - Sorted map of register instructions to their stack version. 620 // The first element is an register file pseudo instruction, the second is the 621 // concrete X86 instruction which uses the register stack. 622 // 623 static const TableEntry OpcodeTable[] = { 624 { X86::ABS_Fp32 , X86::ABS_F }, 625 { X86::ABS_Fp64 , X86::ABS_F }, 626 { X86::ABS_Fp80 , X86::ABS_F }, 627 { X86::ADD_Fp32m , X86::ADD_F32m }, 628 { X86::ADD_Fp64m , X86::ADD_F64m }, 629 { X86::ADD_Fp64m32 , X86::ADD_F32m }, 630 { X86::ADD_Fp80m32 , X86::ADD_F32m }, 631 { X86::ADD_Fp80m64 , X86::ADD_F64m }, 632 { X86::ADD_FpI16m32 , X86::ADD_FI16m }, 633 { X86::ADD_FpI16m64 , X86::ADD_FI16m }, 634 { X86::ADD_FpI16m80 , X86::ADD_FI16m }, 635 { X86::ADD_FpI32m32 , X86::ADD_FI32m }, 636 { X86::ADD_FpI32m64 , X86::ADD_FI32m }, 637 { X86::ADD_FpI32m80 , X86::ADD_FI32m }, 638 { X86::CHS_Fp32 , X86::CHS_F }, 639 { X86::CHS_Fp64 , X86::CHS_F }, 640 { X86::CHS_Fp80 , X86::CHS_F }, 641 { X86::CMOVBE_Fp32 , X86::CMOVBE_F }, 642 { X86::CMOVBE_Fp64 , X86::CMOVBE_F }, 643 { X86::CMOVBE_Fp80 , X86::CMOVBE_F }, 644 { X86::CMOVB_Fp32 , X86::CMOVB_F }, 645 { X86::CMOVB_Fp64 , X86::CMOVB_F }, 646 { X86::CMOVB_Fp80 , X86::CMOVB_F }, 647 { X86::CMOVE_Fp32 , X86::CMOVE_F }, 648 { X86::CMOVE_Fp64 , X86::CMOVE_F }, 649 { X86::CMOVE_Fp80 , X86::CMOVE_F }, 650 { X86::CMOVNBE_Fp32 , X86::CMOVNBE_F }, 651 { X86::CMOVNBE_Fp64 , X86::CMOVNBE_F }, 652 { X86::CMOVNBE_Fp80 , X86::CMOVNBE_F }, 653 { X86::CMOVNB_Fp32 , X86::CMOVNB_F }, 654 { X86::CMOVNB_Fp64 , X86::CMOVNB_F }, 655 { X86::CMOVNB_Fp80 , X86::CMOVNB_F }, 656 { X86::CMOVNE_Fp32 , X86::CMOVNE_F }, 657 { X86::CMOVNE_Fp64 , X86::CMOVNE_F }, 658 { X86::CMOVNE_Fp80 , X86::CMOVNE_F }, 659 { X86::CMOVNP_Fp32 , X86::CMOVNP_F }, 660 { X86::CMOVNP_Fp64 , X86::CMOVNP_F }, 661 { X86::CMOVNP_Fp80 , X86::CMOVNP_F }, 662 { X86::CMOVP_Fp32 , X86::CMOVP_F }, 663 { X86::CMOVP_Fp64 , X86::CMOVP_F }, 664 { X86::CMOVP_Fp80 , X86::CMOVP_F }, 665 { X86::COS_Fp32 , X86::COS_F }, 666 { X86::COS_Fp64 , X86::COS_F }, 667 { X86::COS_Fp80 , X86::COS_F }, 668 { X86::DIVR_Fp32m , X86::DIVR_F32m }, 669 { X86::DIVR_Fp64m , X86::DIVR_F64m }, 670 { X86::DIVR_Fp64m32 , X86::DIVR_F32m }, 671 { X86::DIVR_Fp80m32 , X86::DIVR_F32m }, 672 { X86::DIVR_Fp80m64 , X86::DIVR_F64m }, 673 { X86::DIVR_FpI16m32, X86::DIVR_FI16m}, 674 { X86::DIVR_FpI16m64, X86::DIVR_FI16m}, 675 { X86::DIVR_FpI16m80, X86::DIVR_FI16m}, 676 { X86::DIVR_FpI32m32, X86::DIVR_FI32m}, 677 { X86::DIVR_FpI32m64, X86::DIVR_FI32m}, 678 { X86::DIVR_FpI32m80, X86::DIVR_FI32m}, 679 { X86::DIV_Fp32m , X86::DIV_F32m }, 680 { X86::DIV_Fp64m , X86::DIV_F64m }, 681 { X86::DIV_Fp64m32 , X86::DIV_F32m }, 682 { X86::DIV_Fp80m32 , X86::DIV_F32m }, 683 { X86::DIV_Fp80m64 , X86::DIV_F64m }, 684 { X86::DIV_FpI16m32 , X86::DIV_FI16m }, 685 { X86::DIV_FpI16m64 , X86::DIV_FI16m }, 686 { X86::DIV_FpI16m80 , X86::DIV_FI16m }, 687 { X86::DIV_FpI32m32 , X86::DIV_FI32m }, 688 { X86::DIV_FpI32m64 , X86::DIV_FI32m }, 689 { X86::DIV_FpI32m80 , X86::DIV_FI32m }, 690 { X86::ILD_Fp16m32 , X86::ILD_F16m }, 691 { X86::ILD_Fp16m64 , X86::ILD_F16m }, 692 { X86::ILD_Fp16m80 , X86::ILD_F16m }, 693 { X86::ILD_Fp32m32 , X86::ILD_F32m }, 694 { X86::ILD_Fp32m64 , X86::ILD_F32m }, 695 { X86::ILD_Fp32m80 , X86::ILD_F32m }, 696 { X86::ILD_Fp64m32 , X86::ILD_F64m }, 697 { X86::ILD_Fp64m64 , X86::ILD_F64m }, 698 { X86::ILD_Fp64m80 , X86::ILD_F64m }, 699 { X86::ISTT_Fp16m32 , X86::ISTT_FP16m}, 700 { X86::ISTT_Fp16m64 , X86::ISTT_FP16m}, 701 { X86::ISTT_Fp16m80 , X86::ISTT_FP16m}, 702 { X86::ISTT_Fp32m32 , X86::ISTT_FP32m}, 703 { X86::ISTT_Fp32m64 , X86::ISTT_FP32m}, 704 { X86::ISTT_Fp32m80 , X86::ISTT_FP32m}, 705 { X86::ISTT_Fp64m32 , X86::ISTT_FP64m}, 706 { X86::ISTT_Fp64m64 , X86::ISTT_FP64m}, 707 { X86::ISTT_Fp64m80 , X86::ISTT_FP64m}, 708 { X86::IST_Fp16m32 , X86::IST_F16m }, 709 { X86::IST_Fp16m64 , X86::IST_F16m }, 710 { X86::IST_Fp16m80 , X86::IST_F16m }, 711 { X86::IST_Fp32m32 , X86::IST_F32m }, 712 { X86::IST_Fp32m64 , X86::IST_F32m }, 713 { X86::IST_Fp32m80 , X86::IST_F32m }, 714 { X86::IST_Fp64m32 , X86::IST_FP64m }, 715 { X86::IST_Fp64m64 , X86::IST_FP64m }, 716 { X86::IST_Fp64m80 , X86::IST_FP64m }, 717 { X86::LD_Fp032 , X86::LD_F0 }, 718 { X86::LD_Fp064 , X86::LD_F0 }, 719 { X86::LD_Fp080 , X86::LD_F0 }, 720 { X86::LD_Fp132 , X86::LD_F1 }, 721 { X86::LD_Fp164 , X86::LD_F1 }, 722 { X86::LD_Fp180 , X86::LD_F1 }, 723 { X86::LD_Fp32m , X86::LD_F32m }, 724 { X86::LD_Fp32m64 , X86::LD_F32m }, 725 { X86::LD_Fp32m80 , X86::LD_F32m }, 726 { X86::LD_Fp64m , X86::LD_F64m }, 727 { X86::LD_Fp64m80 , X86::LD_F64m }, 728 { X86::LD_Fp80m , X86::LD_F80m }, 729 { X86::MUL_Fp32m , X86::MUL_F32m }, 730 { X86::MUL_Fp64m , X86::MUL_F64m }, 731 { X86::MUL_Fp64m32 , X86::MUL_F32m }, 732 { X86::MUL_Fp80m32 , X86::MUL_F32m }, 733 { X86::MUL_Fp80m64 , X86::MUL_F64m }, 734 { X86::MUL_FpI16m32 , X86::MUL_FI16m }, 735 { X86::MUL_FpI16m64 , X86::MUL_FI16m }, 736 { X86::MUL_FpI16m80 , X86::MUL_FI16m }, 737 { X86::MUL_FpI32m32 , X86::MUL_FI32m }, 738 { X86::MUL_FpI32m64 , X86::MUL_FI32m }, 739 { X86::MUL_FpI32m80 , X86::MUL_FI32m }, 740 { X86::SIN_Fp32 , X86::SIN_F }, 741 { X86::SIN_Fp64 , X86::SIN_F }, 742 { X86::SIN_Fp80 , X86::SIN_F }, 743 { X86::SQRT_Fp32 , X86::SQRT_F }, 744 { X86::SQRT_Fp64 , X86::SQRT_F }, 745 { X86::SQRT_Fp80 , X86::SQRT_F }, 746 { X86::ST_Fp32m , X86::ST_F32m }, 747 { X86::ST_Fp64m , X86::ST_F64m }, 748 { X86::ST_Fp64m32 , X86::ST_F32m }, 749 { X86::ST_Fp80m32 , X86::ST_F32m }, 750 { X86::ST_Fp80m64 , X86::ST_F64m }, 751 { X86::ST_FpP80m , X86::ST_FP80m }, 752 { X86::SUBR_Fp32m , X86::SUBR_F32m }, 753 { X86::SUBR_Fp64m , X86::SUBR_F64m }, 754 { X86::SUBR_Fp64m32 , X86::SUBR_F32m }, 755 { X86::SUBR_Fp80m32 , X86::SUBR_F32m }, 756 { X86::SUBR_Fp80m64 , X86::SUBR_F64m }, 757 { X86::SUBR_FpI16m32, X86::SUBR_FI16m}, 758 { X86::SUBR_FpI16m64, X86::SUBR_FI16m}, 759 { X86::SUBR_FpI16m80, X86::SUBR_FI16m}, 760 { X86::SUBR_FpI32m32, X86::SUBR_FI32m}, 761 { X86::SUBR_FpI32m64, X86::SUBR_FI32m}, 762 { X86::SUBR_FpI32m80, X86::SUBR_FI32m}, 763 { X86::SUB_Fp32m , X86::SUB_F32m }, 764 { X86::SUB_Fp64m , X86::SUB_F64m }, 765 { X86::SUB_Fp64m32 , X86::SUB_F32m }, 766 { X86::SUB_Fp80m32 , X86::SUB_F32m }, 767 { X86::SUB_Fp80m64 , X86::SUB_F64m }, 768 { X86::SUB_FpI16m32 , X86::SUB_FI16m }, 769 { X86::SUB_FpI16m64 , X86::SUB_FI16m }, 770 { X86::SUB_FpI16m80 , X86::SUB_FI16m }, 771 { X86::SUB_FpI32m32 , X86::SUB_FI32m }, 772 { X86::SUB_FpI32m64 , X86::SUB_FI32m }, 773 { X86::SUB_FpI32m80 , X86::SUB_FI32m }, 774 { X86::TST_Fp32 , X86::TST_F }, 775 { X86::TST_Fp64 , X86::TST_F }, 776 { X86::TST_Fp80 , X86::TST_F }, 777 { X86::UCOM_FpIr32 , X86::UCOM_FIr }, 778 { X86::UCOM_FpIr64 , X86::UCOM_FIr }, 779 { X86::UCOM_FpIr80 , X86::UCOM_FIr }, 780 { X86::UCOM_Fpr32 , X86::UCOM_Fr }, 781 { X86::UCOM_Fpr64 , X86::UCOM_Fr }, 782 { X86::UCOM_Fpr80 , X86::UCOM_Fr }, 783 }; 784 785 static unsigned getConcreteOpcode(unsigned Opcode) { 786 ASSERT_SORTED(OpcodeTable); 787 int Opc = Lookup(OpcodeTable, array_lengthof(OpcodeTable), Opcode); 788 assert(Opc != -1 && "FP Stack instruction not in OpcodeTable!"); 789 return Opc; 790 } 791 792 //===----------------------------------------------------------------------===// 793 // Helper Methods 794 //===----------------------------------------------------------------------===// 795 796 // PopTable - Sorted map of instructions to their popping version. The first 797 // element is an instruction, the second is the version which pops. 798 // 799 static const TableEntry PopTable[] = { 800 { X86::ADD_FrST0 , X86::ADD_FPrST0 }, 801 802 { X86::DIVR_FrST0, X86::DIVR_FPrST0 }, 803 { X86::DIV_FrST0 , X86::DIV_FPrST0 }, 804 805 { X86::IST_F16m , X86::IST_FP16m }, 806 { X86::IST_F32m , X86::IST_FP32m }, 807 808 { X86::MUL_FrST0 , X86::MUL_FPrST0 }, 809 810 { X86::ST_F32m , X86::ST_FP32m }, 811 { X86::ST_F64m , X86::ST_FP64m }, 812 { X86::ST_Frr , X86::ST_FPrr }, 813 814 { X86::SUBR_FrST0, X86::SUBR_FPrST0 }, 815 { X86::SUB_FrST0 , X86::SUB_FPrST0 }, 816 817 { X86::UCOM_FIr , X86::UCOM_FIPr }, 818 819 { X86::UCOM_FPr , X86::UCOM_FPPr }, 820 { X86::UCOM_Fr , X86::UCOM_FPr }, 821 }; 822 823 /// popStackAfter - Pop the current value off of the top of the FP stack after 824 /// the specified instruction. This attempts to be sneaky and combine the pop 825 /// into the instruction itself if possible. The iterator is left pointing to 826 /// the last instruction, be it a new pop instruction inserted, or the old 827 /// instruction if it was modified in place. 828 /// 829 void FPS::popStackAfter(MachineBasicBlock::iterator &I) { 830 MachineInstr* MI = I; 831 DebugLoc dl = MI->getDebugLoc(); 832 ASSERT_SORTED(PopTable); 833 if (StackTop == 0) 834 report_fatal_error("Cannot pop empty stack!"); 835 RegMap[Stack[--StackTop]] = ~0; // Update state 836 837 // Check to see if there is a popping version of this instruction... 838 int Opcode = Lookup(PopTable, array_lengthof(PopTable), I->getOpcode()); 839 if (Opcode != -1) { 840 I->setDesc(TII->get(Opcode)); 841 if (Opcode == X86::UCOM_FPPr) 842 I->RemoveOperand(0); 843 } else { // Insert an explicit pop 844 I = BuildMI(*MBB, ++I, dl, TII->get(X86::ST_FPrr)).addReg(X86::ST0); 845 } 846 } 847 848 /// freeStackSlotAfter - Free the specified register from the register stack, so 849 /// that it is no longer in a register. If the register is currently at the top 850 /// of the stack, we just pop the current instruction, otherwise we store the 851 /// current top-of-stack into the specified slot, then pop the top of stack. 852 void FPS::freeStackSlotAfter(MachineBasicBlock::iterator &I, unsigned FPRegNo) { 853 if (getStackEntry(0) == FPRegNo) { // already at the top of stack? easy. 854 popStackAfter(I); 855 return; 856 } 857 858 // Otherwise, store the top of stack into the dead slot, killing the operand 859 // without having to add in an explicit xchg then pop. 860 // 861 I = freeStackSlotBefore(++I, FPRegNo); 862 } 863 864 /// freeStackSlotBefore - Free the specified register without trying any 865 /// folding. 866 MachineBasicBlock::iterator 867 FPS::freeStackSlotBefore(MachineBasicBlock::iterator I, unsigned FPRegNo) { 868 unsigned STReg = getSTReg(FPRegNo); 869 unsigned OldSlot = getSlot(FPRegNo); 870 unsigned TopReg = Stack[StackTop-1]; 871 Stack[OldSlot] = TopReg; 872 RegMap[TopReg] = OldSlot; 873 RegMap[FPRegNo] = ~0; 874 Stack[--StackTop] = ~0; 875 return BuildMI(*MBB, I, DebugLoc(), TII->get(X86::ST_FPrr)).addReg(STReg); 876 } 877 878 /// adjustLiveRegs - Kill and revive registers such that exactly the FP 879 /// registers with a bit in Mask are live. 880 void FPS::adjustLiveRegs(unsigned Mask, MachineBasicBlock::iterator I) { 881 unsigned Defs = Mask; 882 unsigned Kills = 0; 883 for (unsigned i = 0; i < StackTop; ++i) { 884 unsigned RegNo = Stack[i]; 885 if (!(Defs & (1 << RegNo))) 886 // This register is live, but we don't want it. 887 Kills |= (1 << RegNo); 888 else 889 // We don't need to imp-def this live register. 890 Defs &= ~(1 << RegNo); 891 } 892 assert((Kills & Defs) == 0 && "Register needs killing and def'ing?"); 893 894 // Produce implicit-defs for free by using killed registers. 895 while (Kills && Defs) { 896 unsigned KReg = CountTrailingZeros_32(Kills); 897 unsigned DReg = CountTrailingZeros_32(Defs); 898 DEBUG(dbgs() << "Renaming %FP" << KReg << " as imp %FP" << DReg << "\n"); 899 std::swap(Stack[getSlot(KReg)], Stack[getSlot(DReg)]); 900 std::swap(RegMap[KReg], RegMap[DReg]); 901 Kills &= ~(1 << KReg); 902 Defs &= ~(1 << DReg); 903 } 904 905 // Kill registers by popping. 906 if (Kills && I != MBB->begin()) { 907 MachineBasicBlock::iterator I2 = llvm::prior(I); 908 while (StackTop) { 909 unsigned KReg = getStackEntry(0); 910 if (!(Kills & (1 << KReg))) 911 break; 912 DEBUG(dbgs() << "Popping %FP" << KReg << "\n"); 913 popStackAfter(I2); 914 Kills &= ~(1 << KReg); 915 } 916 } 917 918 // Manually kill the rest. 919 while (Kills) { 920 unsigned KReg = CountTrailingZeros_32(Kills); 921 DEBUG(dbgs() << "Killing %FP" << KReg << "\n"); 922 freeStackSlotBefore(I, KReg); 923 Kills &= ~(1 << KReg); 924 } 925 926 // Load zeros for all the imp-defs. 927 while(Defs) { 928 unsigned DReg = CountTrailingZeros_32(Defs); 929 DEBUG(dbgs() << "Defining %FP" << DReg << " as 0\n"); 930 BuildMI(*MBB, I, DebugLoc(), TII->get(X86::LD_F0)); 931 pushReg(DReg); 932 Defs &= ~(1 << DReg); 933 } 934 935 // Now we should have the correct registers live. 936 DEBUG(dumpStack()); 937 assert(StackTop == CountPopulation_32(Mask) && "Live count mismatch"); 938 } 939 940 /// shuffleStackTop - emit fxch instructions before I to shuffle the top 941 /// FixCount entries into the order given by FixStack. 942 /// FIXME: Is there a better algorithm than insertion sort? 943 void FPS::shuffleStackTop(const unsigned char *FixStack, 944 unsigned FixCount, 945 MachineBasicBlock::iterator I) { 946 // Move items into place, starting from the desired stack bottom. 947 while (FixCount--) { 948 // Old register at position FixCount. 949 unsigned OldReg = getStackEntry(FixCount); 950 // Desired register at position FixCount. 951 unsigned Reg = FixStack[FixCount]; 952 if (Reg == OldReg) 953 continue; 954 // (Reg st0) (OldReg st0) = (Reg OldReg st0) 955 moveToTop(Reg, I); 956 if (FixCount > 0) 957 moveToTop(OldReg, I); 958 } 959 DEBUG(dumpStack()); 960 } 961 962 963 //===----------------------------------------------------------------------===// 964 // Instruction transformation implementation 965 //===----------------------------------------------------------------------===// 966 967 /// handleZeroArgFP - ST(0) = fld0 ST(0) = flds <mem> 968 /// 969 void FPS::handleZeroArgFP(MachineBasicBlock::iterator &I) { 970 MachineInstr *MI = I; 971 unsigned DestReg = getFPReg(MI->getOperand(0)); 972 973 // Change from the pseudo instruction to the concrete instruction. 974 MI->RemoveOperand(0); // Remove the explicit ST(0) operand 975 MI->setDesc(TII->get(getConcreteOpcode(MI->getOpcode()))); 976 977 // Result gets pushed on the stack. 978 pushReg(DestReg); 979 } 980 981 /// handleOneArgFP - fst <mem>, ST(0) 982 /// 983 void FPS::handleOneArgFP(MachineBasicBlock::iterator &I) { 984 MachineInstr *MI = I; 985 unsigned NumOps = MI->getDesc().getNumOperands(); 986 assert((NumOps == X86::AddrNumOperands + 1 || NumOps == 1) && 987 "Can only handle fst* & ftst instructions!"); 988 989 // Is this the last use of the source register? 990 unsigned Reg = getFPReg(MI->getOperand(NumOps-1)); 991 bool KillsSrc = MI->killsRegister(X86::FP0+Reg); 992 993 if (KillsSrc) 994 duplicatePendingSTBeforeKill(Reg, I); 995 996 // FISTP64m is strange because there isn't a non-popping versions. 997 // If we have one _and_ we don't want to pop the operand, duplicate the value 998 // on the stack instead of moving it. This ensure that popping the value is 999 // always ok. 1000 // Ditto FISTTP16m, FISTTP32m, FISTTP64m, ST_FpP80m. 1001 // 1002 if (!KillsSrc && 1003 (MI->getOpcode() == X86::IST_Fp64m32 || 1004 MI->getOpcode() == X86::ISTT_Fp16m32 || 1005 MI->getOpcode() == X86::ISTT_Fp32m32 || 1006 MI->getOpcode() == X86::ISTT_Fp64m32 || 1007 MI->getOpcode() == X86::IST_Fp64m64 || 1008 MI->getOpcode() == X86::ISTT_Fp16m64 || 1009 MI->getOpcode() == X86::ISTT_Fp32m64 || 1010 MI->getOpcode() == X86::ISTT_Fp64m64 || 1011 MI->getOpcode() == X86::IST_Fp64m80 || 1012 MI->getOpcode() == X86::ISTT_Fp16m80 || 1013 MI->getOpcode() == X86::ISTT_Fp32m80 || 1014 MI->getOpcode() == X86::ISTT_Fp64m80 || 1015 MI->getOpcode() == X86::ST_FpP80m)) { 1016 duplicateToTop(Reg, getScratchReg(), I); 1017 } else { 1018 moveToTop(Reg, I); // Move to the top of the stack... 1019 } 1020 1021 // Convert from the pseudo instruction to the concrete instruction. 1022 MI->RemoveOperand(NumOps-1); // Remove explicit ST(0) operand 1023 MI->setDesc(TII->get(getConcreteOpcode(MI->getOpcode()))); 1024 1025 if (MI->getOpcode() == X86::IST_FP64m || 1026 MI->getOpcode() == X86::ISTT_FP16m || 1027 MI->getOpcode() == X86::ISTT_FP32m || 1028 MI->getOpcode() == X86::ISTT_FP64m || 1029 MI->getOpcode() == X86::ST_FP80m) { 1030 if (StackTop == 0) 1031 report_fatal_error("Stack empty??"); 1032 --StackTop; 1033 } else if (KillsSrc) { // Last use of operand? 1034 popStackAfter(I); 1035 } 1036 } 1037 1038 1039 /// handleOneArgFPRW: Handle instructions that read from the top of stack and 1040 /// replace the value with a newly computed value. These instructions may have 1041 /// non-fp operands after their FP operands. 1042 /// 1043 /// Examples: 1044 /// R1 = fchs R2 1045 /// R1 = fadd R2, [mem] 1046 /// 1047 void FPS::handleOneArgFPRW(MachineBasicBlock::iterator &I) { 1048 MachineInstr *MI = I; 1049 #ifndef NDEBUG 1050 unsigned NumOps = MI->getDesc().getNumOperands(); 1051 assert(NumOps >= 2 && "FPRW instructions must have 2 ops!!"); 1052 #endif 1053 1054 // Is this the last use of the source register? 1055 unsigned Reg = getFPReg(MI->getOperand(1)); 1056 bool KillsSrc = MI->killsRegister(X86::FP0+Reg); 1057 1058 if (KillsSrc) { 1059 duplicatePendingSTBeforeKill(Reg, I); 1060 // If this is the last use of the source register, just make sure it's on 1061 // the top of the stack. 1062 moveToTop(Reg, I); 1063 if (StackTop == 0) 1064 report_fatal_error("Stack cannot be empty!"); 1065 --StackTop; 1066 pushReg(getFPReg(MI->getOperand(0))); 1067 } else { 1068 // If this is not the last use of the source register, _copy_ it to the top 1069 // of the stack. 1070 duplicateToTop(Reg, getFPReg(MI->getOperand(0)), I); 1071 } 1072 1073 // Change from the pseudo instruction to the concrete instruction. 1074 MI->RemoveOperand(1); // Drop the source operand. 1075 MI->RemoveOperand(0); // Drop the destination operand. 1076 MI->setDesc(TII->get(getConcreteOpcode(MI->getOpcode()))); 1077 } 1078 1079 1080 //===----------------------------------------------------------------------===// 1081 // Define tables of various ways to map pseudo instructions 1082 // 1083 1084 // ForwardST0Table - Map: A = B op C into: ST(0) = ST(0) op ST(i) 1085 static const TableEntry ForwardST0Table[] = { 1086 { X86::ADD_Fp32 , X86::ADD_FST0r }, 1087 { X86::ADD_Fp64 , X86::ADD_FST0r }, 1088 { X86::ADD_Fp80 , X86::ADD_FST0r }, 1089 { X86::DIV_Fp32 , X86::DIV_FST0r }, 1090 { X86::DIV_Fp64 , X86::DIV_FST0r }, 1091 { X86::DIV_Fp80 , X86::DIV_FST0r }, 1092 { X86::MUL_Fp32 , X86::MUL_FST0r }, 1093 { X86::MUL_Fp64 , X86::MUL_FST0r }, 1094 { X86::MUL_Fp80 , X86::MUL_FST0r }, 1095 { X86::SUB_Fp32 , X86::SUB_FST0r }, 1096 { X86::SUB_Fp64 , X86::SUB_FST0r }, 1097 { X86::SUB_Fp80 , X86::SUB_FST0r }, 1098 }; 1099 1100 // ReverseST0Table - Map: A = B op C into: ST(0) = ST(i) op ST(0) 1101 static const TableEntry ReverseST0Table[] = { 1102 { X86::ADD_Fp32 , X86::ADD_FST0r }, // commutative 1103 { X86::ADD_Fp64 , X86::ADD_FST0r }, // commutative 1104 { X86::ADD_Fp80 , X86::ADD_FST0r }, // commutative 1105 { X86::DIV_Fp32 , X86::DIVR_FST0r }, 1106 { X86::DIV_Fp64 , X86::DIVR_FST0r }, 1107 { X86::DIV_Fp80 , X86::DIVR_FST0r }, 1108 { X86::MUL_Fp32 , X86::MUL_FST0r }, // commutative 1109 { X86::MUL_Fp64 , X86::MUL_FST0r }, // commutative 1110 { X86::MUL_Fp80 , X86::MUL_FST0r }, // commutative 1111 { X86::SUB_Fp32 , X86::SUBR_FST0r }, 1112 { X86::SUB_Fp64 , X86::SUBR_FST0r }, 1113 { X86::SUB_Fp80 , X86::SUBR_FST0r }, 1114 }; 1115 1116 // ForwardSTiTable - Map: A = B op C into: ST(i) = ST(0) op ST(i) 1117 static const TableEntry ForwardSTiTable[] = { 1118 { X86::ADD_Fp32 , X86::ADD_FrST0 }, // commutative 1119 { X86::ADD_Fp64 , X86::ADD_FrST0 }, // commutative 1120 { X86::ADD_Fp80 , X86::ADD_FrST0 }, // commutative 1121 { X86::DIV_Fp32 , X86::DIVR_FrST0 }, 1122 { X86::DIV_Fp64 , X86::DIVR_FrST0 }, 1123 { X86::DIV_Fp80 , X86::DIVR_FrST0 }, 1124 { X86::MUL_Fp32 , X86::MUL_FrST0 }, // commutative 1125 { X86::MUL_Fp64 , X86::MUL_FrST0 }, // commutative 1126 { X86::MUL_Fp80 , X86::MUL_FrST0 }, // commutative 1127 { X86::SUB_Fp32 , X86::SUBR_FrST0 }, 1128 { X86::SUB_Fp64 , X86::SUBR_FrST0 }, 1129 { X86::SUB_Fp80 , X86::SUBR_FrST0 }, 1130 }; 1131 1132 // ReverseSTiTable - Map: A = B op C into: ST(i) = ST(i) op ST(0) 1133 static const TableEntry ReverseSTiTable[] = { 1134 { X86::ADD_Fp32 , X86::ADD_FrST0 }, 1135 { X86::ADD_Fp64 , X86::ADD_FrST0 }, 1136 { X86::ADD_Fp80 , X86::ADD_FrST0 }, 1137 { X86::DIV_Fp32 , X86::DIV_FrST0 }, 1138 { X86::DIV_Fp64 , X86::DIV_FrST0 }, 1139 { X86::DIV_Fp80 , X86::DIV_FrST0 }, 1140 { X86::MUL_Fp32 , X86::MUL_FrST0 }, 1141 { X86::MUL_Fp64 , X86::MUL_FrST0 }, 1142 { X86::MUL_Fp80 , X86::MUL_FrST0 }, 1143 { X86::SUB_Fp32 , X86::SUB_FrST0 }, 1144 { X86::SUB_Fp64 , X86::SUB_FrST0 }, 1145 { X86::SUB_Fp80 , X86::SUB_FrST0 }, 1146 }; 1147 1148 1149 /// handleTwoArgFP - Handle instructions like FADD and friends which are virtual 1150 /// instructions which need to be simplified and possibly transformed. 1151 /// 1152 /// Result: ST(0) = fsub ST(0), ST(i) 1153 /// ST(i) = fsub ST(0), ST(i) 1154 /// ST(0) = fsubr ST(0), ST(i) 1155 /// ST(i) = fsubr ST(0), ST(i) 1156 /// 1157 void FPS::handleTwoArgFP(MachineBasicBlock::iterator &I) { 1158 ASSERT_SORTED(ForwardST0Table); ASSERT_SORTED(ReverseST0Table); 1159 ASSERT_SORTED(ForwardSTiTable); ASSERT_SORTED(ReverseSTiTable); 1160 MachineInstr *MI = I; 1161 1162 unsigned NumOperands = MI->getDesc().getNumOperands(); 1163 assert(NumOperands == 3 && "Illegal TwoArgFP instruction!"); 1164 unsigned Dest = getFPReg(MI->getOperand(0)); 1165 unsigned Op0 = getFPReg(MI->getOperand(NumOperands-2)); 1166 unsigned Op1 = getFPReg(MI->getOperand(NumOperands-1)); 1167 bool KillsOp0 = MI->killsRegister(X86::FP0+Op0); 1168 bool KillsOp1 = MI->killsRegister(X86::FP0+Op1); 1169 DebugLoc dl = MI->getDebugLoc(); 1170 1171 unsigned TOS = getStackEntry(0); 1172 1173 // One of our operands must be on the top of the stack. If neither is yet, we 1174 // need to move one. 1175 if (Op0 != TOS && Op1 != TOS) { // No operand at TOS? 1176 // We can choose to move either operand to the top of the stack. If one of 1177 // the operands is killed by this instruction, we want that one so that we 1178 // can update right on top of the old version. 1179 if (KillsOp0) { 1180 moveToTop(Op0, I); // Move dead operand to TOS. 1181 TOS = Op0; 1182 } else if (KillsOp1) { 1183 moveToTop(Op1, I); 1184 TOS = Op1; 1185 } else { 1186 // All of the operands are live after this instruction executes, so we 1187 // cannot update on top of any operand. Because of this, we must 1188 // duplicate one of the stack elements to the top. It doesn't matter 1189 // which one we pick. 1190 // 1191 duplicateToTop(Op0, Dest, I); 1192 Op0 = TOS = Dest; 1193 KillsOp0 = true; 1194 } 1195 } else if (!KillsOp0 && !KillsOp1) { 1196 // If we DO have one of our operands at the top of the stack, but we don't 1197 // have a dead operand, we must duplicate one of the operands to a new slot 1198 // on the stack. 1199 duplicateToTop(Op0, Dest, I); 1200 Op0 = TOS = Dest; 1201 KillsOp0 = true; 1202 } 1203 1204 // Now we know that one of our operands is on the top of the stack, and at 1205 // least one of our operands is killed by this instruction. 1206 assert((TOS == Op0 || TOS == Op1) && (KillsOp0 || KillsOp1) && 1207 "Stack conditions not set up right!"); 1208 1209 // We decide which form to use based on what is on the top of the stack, and 1210 // which operand is killed by this instruction. 1211 const TableEntry *InstTable; 1212 bool isForward = TOS == Op0; 1213 bool updateST0 = (TOS == Op0 && !KillsOp1) || (TOS == Op1 && !KillsOp0); 1214 if (updateST0) { 1215 if (isForward) 1216 InstTable = ForwardST0Table; 1217 else 1218 InstTable = ReverseST0Table; 1219 } else { 1220 if (isForward) 1221 InstTable = ForwardSTiTable; 1222 else 1223 InstTable = ReverseSTiTable; 1224 } 1225 1226 int Opcode = Lookup(InstTable, array_lengthof(ForwardST0Table), 1227 MI->getOpcode()); 1228 assert(Opcode != -1 && "Unknown TwoArgFP pseudo instruction!"); 1229 1230 // NotTOS - The register which is not on the top of stack... 1231 unsigned NotTOS = (TOS == Op0) ? Op1 : Op0; 1232 1233 // Replace the old instruction with a new instruction 1234 MBB->remove(I++); 1235 I = BuildMI(*MBB, I, dl, TII->get(Opcode)).addReg(getSTReg(NotTOS)); 1236 1237 // If both operands are killed, pop one off of the stack in addition to 1238 // overwriting the other one. 1239 if (KillsOp0 && KillsOp1 && Op0 != Op1) { 1240 assert(!updateST0 && "Should have updated other operand!"); 1241 popStackAfter(I); // Pop the top of stack 1242 } 1243 1244 // Update stack information so that we know the destination register is now on 1245 // the stack. 1246 unsigned UpdatedSlot = getSlot(updateST0 ? TOS : NotTOS); 1247 assert(UpdatedSlot < StackTop && Dest < 7); 1248 Stack[UpdatedSlot] = Dest; 1249 RegMap[Dest] = UpdatedSlot; 1250 MBB->getParent()->DeleteMachineInstr(MI); // Remove the old instruction 1251 } 1252 1253 /// handleCompareFP - Handle FUCOM and FUCOMI instructions, which have two FP 1254 /// register arguments and no explicit destinations. 1255 /// 1256 void FPS::handleCompareFP(MachineBasicBlock::iterator &I) { 1257 ASSERT_SORTED(ForwardST0Table); ASSERT_SORTED(ReverseST0Table); 1258 ASSERT_SORTED(ForwardSTiTable); ASSERT_SORTED(ReverseSTiTable); 1259 MachineInstr *MI = I; 1260 1261 unsigned NumOperands = MI->getDesc().getNumOperands(); 1262 assert(NumOperands == 2 && "Illegal FUCOM* instruction!"); 1263 unsigned Op0 = getFPReg(MI->getOperand(NumOperands-2)); 1264 unsigned Op1 = getFPReg(MI->getOperand(NumOperands-1)); 1265 bool KillsOp0 = MI->killsRegister(X86::FP0+Op0); 1266 bool KillsOp1 = MI->killsRegister(X86::FP0+Op1); 1267 1268 // Make sure the first operand is on the top of stack, the other one can be 1269 // anywhere. 1270 moveToTop(Op0, I); 1271 1272 // Change from the pseudo instruction to the concrete instruction. 1273 MI->getOperand(0).setReg(getSTReg(Op1)); 1274 MI->RemoveOperand(1); 1275 MI->setDesc(TII->get(getConcreteOpcode(MI->getOpcode()))); 1276 1277 // If any of the operands are killed by this instruction, free them. 1278 if (KillsOp0) freeStackSlotAfter(I, Op0); 1279 if (KillsOp1 && Op0 != Op1) freeStackSlotAfter(I, Op1); 1280 } 1281 1282 /// handleCondMovFP - Handle two address conditional move instructions. These 1283 /// instructions move a st(i) register to st(0) iff a condition is true. These 1284 /// instructions require that the first operand is at the top of the stack, but 1285 /// otherwise don't modify the stack at all. 1286 void FPS::handleCondMovFP(MachineBasicBlock::iterator &I) { 1287 MachineInstr *MI = I; 1288 1289 unsigned Op0 = getFPReg(MI->getOperand(0)); 1290 unsigned Op1 = getFPReg(MI->getOperand(2)); 1291 bool KillsOp1 = MI->killsRegister(X86::FP0+Op1); 1292 1293 // The first operand *must* be on the top of the stack. 1294 moveToTop(Op0, I); 1295 1296 // Change the second operand to the stack register that the operand is in. 1297 // Change from the pseudo instruction to the concrete instruction. 1298 MI->RemoveOperand(0); 1299 MI->RemoveOperand(1); 1300 MI->getOperand(0).setReg(getSTReg(Op1)); 1301 MI->setDesc(TII->get(getConcreteOpcode(MI->getOpcode()))); 1302 1303 // If we kill the second operand, make sure to pop it from the stack. 1304 if (Op0 != Op1 && KillsOp1) { 1305 // Get this value off of the register stack. 1306 freeStackSlotAfter(I, Op1); 1307 } 1308 } 1309 1310 1311 /// handleSpecialFP - Handle special instructions which behave unlike other 1312 /// floating point instructions. This is primarily intended for use by pseudo 1313 /// instructions. 1314 /// 1315 void FPS::handleSpecialFP(MachineBasicBlock::iterator &I) { 1316 MachineInstr *MI = I; 1317 switch (MI->getOpcode()) { 1318 default: llvm_unreachable("Unknown SpecialFP instruction!"); 1319 case TargetOpcode::COPY: { 1320 // We handle three kinds of copies: FP <- FP, FP <- ST, and ST <- FP. 1321 const MachineOperand &MO1 = MI->getOperand(1); 1322 const MachineOperand &MO0 = MI->getOperand(0); 1323 unsigned DstST = MO0.getReg() - X86::ST0; 1324 unsigned SrcST = MO1.getReg() - X86::ST0; 1325 bool KillsSrc = MI->killsRegister(MO1.getReg()); 1326 1327 // ST = COPY FP. Set up a pending ST register. 1328 if (DstST < 8) { 1329 unsigned SrcFP = getFPReg(MO1); 1330 assert(isLive(SrcFP) && "Cannot copy dead register"); 1331 assert(!MO0.isDead() && "Cannot copy to dead ST register"); 1332 1333 // Unallocated STs are marked as the nonexistent FP255. 1334 while (NumPendingSTs <= DstST) 1335 PendingST[NumPendingSTs++] = NumFPRegs; 1336 1337 // STi could still be live from a previous inline asm. 1338 if (isScratchReg(PendingST[DstST])) { 1339 DEBUG(dbgs() << "Clobbering old ST in FP" << unsigned(PendingST[DstST]) 1340 << '\n'); 1341 freeStackSlotBefore(MI, PendingST[DstST]); 1342 } 1343 1344 // When the source is killed, allocate a scratch FP register. 1345 if (KillsSrc) { 1346 duplicatePendingSTBeforeKill(SrcFP, I); 1347 unsigned Slot = getSlot(SrcFP); 1348 unsigned SR = getScratchReg(); 1349 PendingST[DstST] = SR; 1350 Stack[Slot] = SR; 1351 RegMap[SR] = Slot; 1352 } else 1353 PendingST[DstST] = SrcFP; 1354 break; 1355 } 1356 1357 // FP = COPY ST. Extract fixed stack value. 1358 // Any instruction defining ST registers must have assigned them to a 1359 // scratch register. 1360 if (SrcST < 8) { 1361 unsigned DstFP = getFPReg(MO0); 1362 assert(!isLive(DstFP) && "Cannot copy ST to live FP register"); 1363 assert(NumPendingSTs > SrcST && "Cannot copy from dead ST register"); 1364 unsigned SrcFP = PendingST[SrcST]; 1365 assert(isScratchReg(SrcFP) && "Expected ST in a scratch register"); 1366 assert(isLive(SrcFP) && "Scratch holding ST is dead"); 1367 1368 // DstFP steals the stack slot from SrcFP. 1369 unsigned Slot = getSlot(SrcFP); 1370 Stack[Slot] = DstFP; 1371 RegMap[DstFP] = Slot; 1372 1373 // Always treat the ST as killed. 1374 PendingST[SrcST] = NumFPRegs; 1375 while (NumPendingSTs && PendingST[NumPendingSTs - 1] == NumFPRegs) 1376 --NumPendingSTs; 1377 break; 1378 } 1379 1380 // FP <- FP copy. 1381 unsigned DstFP = getFPReg(MO0); 1382 unsigned SrcFP = getFPReg(MO1); 1383 assert(isLive(SrcFP) && "Cannot copy dead register"); 1384 if (KillsSrc) { 1385 // If the input operand is killed, we can just change the owner of the 1386 // incoming stack slot into the result. 1387 unsigned Slot = getSlot(SrcFP); 1388 Stack[Slot] = DstFP; 1389 RegMap[DstFP] = Slot; 1390 } else { 1391 // For COPY we just duplicate the specified value to a new stack slot. 1392 // This could be made better, but would require substantial changes. 1393 duplicateToTop(SrcFP, DstFP, I); 1394 } 1395 break; 1396 } 1397 1398 case TargetOpcode::IMPLICIT_DEF: { 1399 // All FP registers must be explicitly defined, so load a 0 instead. 1400 unsigned Reg = MI->getOperand(0).getReg() - X86::FP0; 1401 DEBUG(dbgs() << "Emitting LD_F0 for implicit FP" << Reg << '\n'); 1402 BuildMI(*MBB, I, MI->getDebugLoc(), TII->get(X86::LD_F0)); 1403 pushReg(Reg); 1404 break; 1405 } 1406 1407 case X86::FpPOP_RETVAL: { 1408 // The FpPOP_RETVAL instruction is used after calls that return a value on 1409 // the floating point stack. We cannot model this with ST defs since CALL 1410 // instructions have fixed clobber lists. This instruction is interpreted 1411 // to mean that there is one more live register on the stack than we 1412 // thought. 1413 // 1414 // This means that StackTop does not match the hardware stack between a 1415 // call and the FpPOP_RETVAL instructions. We do tolerate FP instructions 1416 // between CALL and FpPOP_RETVAL as long as they don't overflow the 1417 // hardware stack. 1418 unsigned DstFP = getFPReg(MI->getOperand(0)); 1419 1420 // Move existing stack elements up to reflect reality. 1421 assert(StackTop < 8 && "Stack overflowed before FpPOP_RETVAL"); 1422 if (StackTop) { 1423 std::copy_backward(Stack, Stack + StackTop, Stack + StackTop + 1); 1424 for (unsigned i = 0; i != NumFPRegs; ++i) 1425 ++RegMap[i]; 1426 } 1427 ++StackTop; 1428 1429 // DstFP is the new bottom of the stack. 1430 Stack[0] = DstFP; 1431 RegMap[DstFP] = 0; 1432 1433 // DstFP will be killed by processBasicBlock if this was a dead def. 1434 break; 1435 } 1436 1437 case TargetOpcode::INLINEASM: { 1438 // The inline asm MachineInstr currently only *uses* FP registers for the 1439 // 'f' constraint. These should be turned into the current ST(x) register 1440 // in the machine instr. 1441 // 1442 // There are special rules for x87 inline assembly. The compiler must know 1443 // exactly how many registers are popped and pushed implicitly by the asm. 1444 // Otherwise it is not possible to restore the stack state after the inline 1445 // asm. 1446 // 1447 // There are 3 kinds of input operands: 1448 // 1449 // 1. Popped inputs. These must appear at the stack top in ST0-STn. A 1450 // popped input operand must be in a fixed stack slot, and it is either 1451 // tied to an output operand, or in the clobber list. The MI has ST use 1452 // and def operands for these inputs. 1453 // 1454 // 2. Fixed inputs. These inputs appear in fixed stack slots, but are 1455 // preserved by the inline asm. The fixed stack slots must be STn-STm 1456 // following the popped inputs. A fixed input operand cannot be tied to 1457 // an output or appear in the clobber list. The MI has ST use operands 1458 // and no defs for these inputs. 1459 // 1460 // 3. Preserved inputs. These inputs use the "f" constraint which is 1461 // represented as an FP register. The inline asm won't change these 1462 // stack slots. 1463 // 1464 // Outputs must be in ST registers, FP outputs are not allowed. Clobbered 1465 // registers do not count as output operands. The inline asm changes the 1466 // stack as if it popped all the popped inputs and then pushed all the 1467 // output operands. 1468 1469 // Scan the assembly for ST registers used, defined and clobbered. We can 1470 // only tell clobbers from defs by looking at the asm descriptor. 1471 unsigned STUses = 0, STDefs = 0, STClobbers = 0, STDeadDefs = 0; 1472 unsigned NumOps = 0; 1473 for (unsigned i = InlineAsm::MIOp_FirstOperand, e = MI->getNumOperands(); 1474 i != e && MI->getOperand(i).isImm(); i += 1 + NumOps) { 1475 unsigned Flags = MI->getOperand(i).getImm(); 1476 NumOps = InlineAsm::getNumOperandRegisters(Flags); 1477 if (NumOps != 1) 1478 continue; 1479 const MachineOperand &MO = MI->getOperand(i + 1); 1480 if (!MO.isReg()) 1481 continue; 1482 unsigned STReg = MO.getReg() - X86::ST0; 1483 if (STReg >= 8) 1484 continue; 1485 1486 switch (InlineAsm::getKind(Flags)) { 1487 case InlineAsm::Kind_RegUse: 1488 STUses |= (1u << STReg); 1489 break; 1490 case InlineAsm::Kind_RegDef: 1491 case InlineAsm::Kind_RegDefEarlyClobber: 1492 STDefs |= (1u << STReg); 1493 if (MO.isDead()) 1494 STDeadDefs |= (1u << STReg); 1495 break; 1496 case InlineAsm::Kind_Clobber: 1497 STClobbers |= (1u << STReg); 1498 break; 1499 default: 1500 break; 1501 } 1502 } 1503 1504 if (STUses && !isMask_32(STUses)) 1505 MI->emitError("fixed input regs must be last on the x87 stack"); 1506 unsigned NumSTUses = CountTrailingOnes_32(STUses); 1507 1508 // Defs must be contiguous from the stack top. ST0-STn. 1509 if (STDefs && !isMask_32(STDefs)) { 1510 MI->emitError("output regs must be last on the x87 stack"); 1511 STDefs = NextPowerOf2(STDefs) - 1; 1512 } 1513 unsigned NumSTDefs = CountTrailingOnes_32(STDefs); 1514 1515 // So must the clobbered stack slots. ST0-STm, m >= n. 1516 if (STClobbers && !isMask_32(STDefs | STClobbers)) 1517 MI->emitError("clobbers must be last on the x87 stack"); 1518 1519 // Popped inputs are the ones that are also clobbered or defined. 1520 unsigned STPopped = STUses & (STDefs | STClobbers); 1521 if (STPopped && !isMask_32(STPopped)) 1522 MI->emitError("implicitly popped regs must be last on the x87 stack"); 1523 unsigned NumSTPopped = CountTrailingOnes_32(STPopped); 1524 1525 DEBUG(dbgs() << "Asm uses " << NumSTUses << " fixed regs, pops " 1526 << NumSTPopped << ", and defines " << NumSTDefs << " regs.\n"); 1527 1528 // Scan the instruction for FP uses corresponding to "f" constraints. 1529 // Collect FP registers to kill afer the instruction. 1530 // Always kill all the scratch regs. 1531 unsigned FPKills = ((1u << NumFPRegs) - 1) & ~0xff; 1532 unsigned FPUsed = 0; 1533 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) { 1534 MachineOperand &Op = MI->getOperand(i); 1535 if (!Op.isReg() || Op.getReg() < X86::FP0 || Op.getReg() > X86::FP6) 1536 continue; 1537 if (!Op.isUse()) 1538 MI->emitError("illegal \"f\" output constraint"); 1539 unsigned FPReg = getFPReg(Op); 1540 FPUsed |= 1U << FPReg; 1541 1542 // If we kill this operand, make sure to pop it from the stack after the 1543 // asm. We just remember it for now, and pop them all off at the end in 1544 // a batch. 1545 if (Op.isKill()) 1546 FPKills |= 1U << FPReg; 1547 } 1548 1549 // The popped inputs will be killed by the instruction, so duplicate them 1550 // if the FP register needs to be live after the instruction, or if it is 1551 // used in the instruction itself. We effectively treat the popped inputs 1552 // as early clobbers. 1553 for (unsigned i = 0; i < NumSTPopped; ++i) { 1554 if ((FPKills & ~FPUsed) & (1u << PendingST[i])) 1555 continue; 1556 unsigned SR = getScratchReg(); 1557 duplicateToTop(PendingST[i], SR, I); 1558 DEBUG(dbgs() << "Duplicating ST" << i << " in FP" 1559 << unsigned(PendingST[i]) << " to avoid clobbering it.\n"); 1560 PendingST[i] = SR; 1561 } 1562 1563 // Make sure we have a unique live register for every fixed use. Some of 1564 // them could be undef uses, and we need to emit LD_F0 instructions. 1565 for (unsigned i = 0; i < NumSTUses; ++i) { 1566 if (i < NumPendingSTs && PendingST[i] < NumFPRegs) { 1567 // Check for shared assignments. 1568 for (unsigned j = 0; j < i; ++j) { 1569 if (PendingST[j] != PendingST[i]) 1570 continue; 1571 // STi and STj are inn the same register, create a copy. 1572 unsigned SR = getScratchReg(); 1573 duplicateToTop(PendingST[i], SR, I); 1574 DEBUG(dbgs() << "Duplicating ST" << i << " in FP" 1575 << unsigned(PendingST[i]) 1576 << " to avoid collision with ST" << j << '\n'); 1577 PendingST[i] = SR; 1578 } 1579 continue; 1580 } 1581 unsigned SR = getScratchReg(); 1582 DEBUG(dbgs() << "Emitting LD_F0 for ST" << i << " in FP" << SR << '\n'); 1583 BuildMI(*MBB, I, MI->getDebugLoc(), TII->get(X86::LD_F0)); 1584 pushReg(SR); 1585 PendingST[i] = SR; 1586 if (NumPendingSTs == i) 1587 ++NumPendingSTs; 1588 } 1589 assert(NumPendingSTs >= NumSTUses && "Fixed registers should be assigned"); 1590 1591 // Now we can rearrange the live registers to match what was requested. 1592 shuffleStackTop(PendingST, NumPendingSTs, I); 1593 DEBUG({dbgs() << "Before asm: "; dumpStack();}); 1594 1595 // With the stack layout fixed, rewrite the FP registers. 1596 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) { 1597 MachineOperand &Op = MI->getOperand(i); 1598 if (!Op.isReg() || Op.getReg() < X86::FP0 || Op.getReg() > X86::FP6) 1599 continue; 1600 unsigned FPReg = getFPReg(Op); 1601 Op.setReg(getSTReg(FPReg)); 1602 } 1603 1604 // Simulate the inline asm popping its inputs and pushing its outputs. 1605 StackTop -= NumSTPopped; 1606 1607 // Hold the fixed output registers in scratch FP registers. They will be 1608 // transferred to real FP registers by copies. 1609 NumPendingSTs = 0; 1610 for (unsigned i = 0; i < NumSTDefs; ++i) { 1611 unsigned SR = getScratchReg(); 1612 pushReg(SR); 1613 FPKills &= ~(1u << SR); 1614 } 1615 for (unsigned i = 0; i < NumSTDefs; ++i) 1616 PendingST[NumPendingSTs++] = getStackEntry(i); 1617 DEBUG({dbgs() << "After asm: "; dumpStack();}); 1618 1619 // If any of the ST defs were dead, pop them immediately. Our caller only 1620 // handles dead FP defs. 1621 MachineBasicBlock::iterator InsertPt = MI; 1622 for (unsigned i = 0; STDefs & (1u << i); ++i) { 1623 if (!(STDeadDefs & (1u << i))) 1624 continue; 1625 freeStackSlotAfter(InsertPt, PendingST[i]); 1626 PendingST[i] = NumFPRegs; 1627 } 1628 while (NumPendingSTs && PendingST[NumPendingSTs - 1] == NumFPRegs) 1629 --NumPendingSTs; 1630 1631 // If this asm kills any FP registers (is the last use of them) we must 1632 // explicitly emit pop instructions for them. Do this now after the asm has 1633 // executed so that the ST(x) numbers are not off (which would happen if we 1634 // did this inline with operand rewriting). 1635 // 1636 // Note: this might be a non-optimal pop sequence. We might be able to do 1637 // better by trying to pop in stack order or something. 1638 while (FPKills) { 1639 unsigned FPReg = CountTrailingZeros_32(FPKills); 1640 if (isLive(FPReg)) 1641 freeStackSlotAfter(InsertPt, FPReg); 1642 FPKills &= ~(1U << FPReg); 1643 } 1644 // Don't delete the inline asm! 1645 return; 1646 } 1647 1648 case X86::WIN_FTOL_32: 1649 case X86::WIN_FTOL_64: { 1650 // Push the operand into ST0. 1651 MachineOperand &Op = MI->getOperand(0); 1652 assert(Op.isUse() && Op.isReg() && 1653 Op.getReg() >= X86::FP0 && Op.getReg() <= X86::FP6); 1654 unsigned FPReg = getFPReg(Op); 1655 if (Op.isKill()) 1656 moveToTop(FPReg, I); 1657 else 1658 duplicateToTop(FPReg, FPReg, I); 1659 1660 // Emit the call. This will pop the operand. 1661 BuildMI(*MBB, I, MI->getDebugLoc(), TII->get(X86::CALLpcrel32)) 1662 .addExternalSymbol("_ftol2") 1663 .addReg(X86::ST0, RegState::ImplicitKill) 1664 .addReg(X86::EAX, RegState::Define | RegState::Implicit) 1665 .addReg(X86::EDX, RegState::Define | RegState::Implicit) 1666 .addReg(X86::EFLAGS, RegState::Define | RegState::Implicit); 1667 --StackTop; 1668 1669 break; 1670 } 1671 1672 case X86::RET: 1673 case X86::RETI: 1674 // If RET has an FP register use operand, pass the first one in ST(0) and 1675 // the second one in ST(1). 1676 1677 // Find the register operands. 1678 unsigned FirstFPRegOp = ~0U, SecondFPRegOp = ~0U; 1679 unsigned LiveMask = 0; 1680 1681 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) { 1682 MachineOperand &Op = MI->getOperand(i); 1683 if (!Op.isReg() || Op.getReg() < X86::FP0 || Op.getReg() > X86::FP6) 1684 continue; 1685 // FP Register uses must be kills unless there are two uses of the same 1686 // register, in which case only one will be a kill. 1687 assert(Op.isUse() && 1688 (Op.isKill() || // Marked kill. 1689 getFPReg(Op) == FirstFPRegOp || // Second instance. 1690 MI->killsRegister(Op.getReg())) && // Later use is marked kill. 1691 "Ret only defs operands, and values aren't live beyond it"); 1692 1693 if (FirstFPRegOp == ~0U) 1694 FirstFPRegOp = getFPReg(Op); 1695 else { 1696 assert(SecondFPRegOp == ~0U && "More than two fp operands!"); 1697 SecondFPRegOp = getFPReg(Op); 1698 } 1699 LiveMask |= (1 << getFPReg(Op)); 1700 1701 // Remove the operand so that later passes don't see it. 1702 MI->RemoveOperand(i); 1703 --i, --e; 1704 } 1705 1706 // We may have been carrying spurious live-ins, so make sure only the returned 1707 // registers are left live. 1708 adjustLiveRegs(LiveMask, MI); 1709 if (!LiveMask) return; // Quick check to see if any are possible. 1710 1711 // There are only four possibilities here: 1712 // 1) we are returning a single FP value. In this case, it has to be in 1713 // ST(0) already, so just declare success by removing the value from the 1714 // FP Stack. 1715 if (SecondFPRegOp == ~0U) { 1716 // Assert that the top of stack contains the right FP register. 1717 assert(StackTop == 1 && FirstFPRegOp == getStackEntry(0) && 1718 "Top of stack not the right register for RET!"); 1719 1720 // Ok, everything is good, mark the value as not being on the stack 1721 // anymore so that our assertion about the stack being empty at end of 1722 // block doesn't fire. 1723 StackTop = 0; 1724 return; 1725 } 1726 1727 // Otherwise, we are returning two values: 1728 // 2) If returning the same value for both, we only have one thing in the FP 1729 // stack. Consider: RET FP1, FP1 1730 if (StackTop == 1) { 1731 assert(FirstFPRegOp == SecondFPRegOp && FirstFPRegOp == getStackEntry(0)&& 1732 "Stack misconfiguration for RET!"); 1733 1734 // Duplicate the TOS so that we return it twice. Just pick some other FPx 1735 // register to hold it. 1736 unsigned NewReg = getScratchReg(); 1737 duplicateToTop(FirstFPRegOp, NewReg, MI); 1738 FirstFPRegOp = NewReg; 1739 } 1740 1741 /// Okay we know we have two different FPx operands now: 1742 assert(StackTop == 2 && "Must have two values live!"); 1743 1744 /// 3) If SecondFPRegOp is currently in ST(0) and FirstFPRegOp is currently 1745 /// in ST(1). In this case, emit an fxch. 1746 if (getStackEntry(0) == SecondFPRegOp) { 1747 assert(getStackEntry(1) == FirstFPRegOp && "Unknown regs live"); 1748 moveToTop(FirstFPRegOp, MI); 1749 } 1750 1751 /// 4) Finally, FirstFPRegOp must be in ST(0) and SecondFPRegOp must be in 1752 /// ST(1). Just remove both from our understanding of the stack and return. 1753 assert(getStackEntry(0) == FirstFPRegOp && "Unknown regs live"); 1754 assert(getStackEntry(1) == SecondFPRegOp && "Unknown regs live"); 1755 StackTop = 0; 1756 return; 1757 } 1758 1759 I = MBB->erase(I); // Remove the pseudo instruction 1760 1761 // We want to leave I pointing to the previous instruction, but what if we 1762 // just erased the first instruction? 1763 if (I == MBB->begin()) { 1764 DEBUG(dbgs() << "Inserting dummy KILL\n"); 1765 I = BuildMI(*MBB, I, DebugLoc(), TII->get(TargetOpcode::KILL)); 1766 } else 1767 --I; 1768 } 1769