1 //===-- X86FloatingPoint.cpp - Floating point Reg -> Stack converter ------===//
3 // The LLVM Compiler Infrastructure
5 // This file was developed by the LLVM research group and is distributed under
6 // the University of Illinois Open Source License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file defines the pass which converts floating point instructions from
11 // virtual registers into register stack instructions. This pass uses live
12 // variable information to indicate where the FPn registers are used and their
15 // This pass is hampered by the lack of decent CFG manipulation routines for
16 // machine code. In particular, this wants to be able to split critical edges
17 // as necessary, traverse the machine basic block CFG in depth-first order, and
18 // allow there to be multiple machine basic blocks for each LLVM basicblock
19 // (needed for critical edge splitting).
21 // In particular, this pass currently barfs on critical edges. Because of this,
22 // it requires the instruction selector to insert FP_REG_KILL instructions on
23 // the exits of any basic block that has critical edges going from it, or which
24 // branch to a critical basic block.
26 // FIXME: this is not implemented yet. The stackifier pass only works on local
29 //===----------------------------------------------------------------------===//
31 #define DEBUG_TYPE "x86-codegen"
33 #include "X86InstrInfo.h"
34 #include "llvm/CodeGen/MachineFunctionPass.h"
35 #include "llvm/CodeGen/MachineInstrBuilder.h"
36 #include "llvm/CodeGen/LiveVariables.h"
37 #include "llvm/CodeGen/Passes.h"
38 #include "llvm/Target/TargetInstrInfo.h"
39 #include "llvm/Target/TargetMachine.h"
40 #include "llvm/Support/Debug.h"
41 #include "llvm/Support/Compiler.h"
42 #include "llvm/ADT/DepthFirstIterator.h"
43 #include "llvm/ADT/SmallVector.h"
44 #include "llvm/ADT/Statistic.h"
45 #include "llvm/ADT/STLExtras.h"
50 STATISTIC(NumFXCH, "Number of fxch instructions inserted");
51 STATISTIC(NumFP , "Number of floating point instructions");
54 struct VISIBILITY_HIDDEN FPS : public MachineFunctionPass {
56 FPS() : MachineFunctionPass((intptr_t)&ID) {}
58 virtual bool runOnMachineFunction(MachineFunction &MF);
60 virtual const char *getPassName() const { return "X86 FP Stackifier"; }
62 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
63 AU.addRequired<LiveVariables>();
64 MachineFunctionPass::getAnalysisUsage(AU);
67 const TargetInstrInfo *TII; // Machine instruction info.
68 LiveVariables *LV; // Live variable info for current function...
69 MachineBasicBlock *MBB; // Current basic block
70 unsigned Stack[8]; // FP<n> Registers in each stack slot...
71 unsigned RegMap[8]; // Track which stack slot contains each register
72 unsigned StackTop; // The current top of the FP stack.
74 void dumpStack() const {
75 cerr << "Stack contents:";
76 for (unsigned i = 0; i != StackTop; ++i) {
77 cerr << " FP" << Stack[i];
78 assert(RegMap[Stack[i]] == i && "Stack[] doesn't match RegMap[]!");
83 // getSlot - Return the stack slot number a particular register number is
85 unsigned getSlot(unsigned RegNo) const {
86 assert(RegNo < 8 && "Regno out of range!");
90 // getStackEntry - Return the X86::FP<n> register in register ST(i)
91 unsigned getStackEntry(unsigned STi) const {
92 assert(STi < StackTop && "Access past stack top!");
93 return Stack[StackTop-1-STi];
96 // getSTReg - Return the X86::ST(i) register which contains the specified
98 unsigned getSTReg(unsigned RegNo) const {
99 return StackTop - 1 - getSlot(RegNo) + llvm::X86::ST0;
102 // pushReg - Push the specified FP<n> register onto the stack
103 void pushReg(unsigned Reg) {
104 assert(Reg < 8 && "Register number out of range!");
105 assert(StackTop < 8 && "Stack overflow!");
106 Stack[StackTop] = Reg;
107 RegMap[Reg] = StackTop++;
110 bool isAtTop(unsigned RegNo) const { return getSlot(RegNo) == StackTop-1; }
111 void moveToTop(unsigned RegNo, MachineBasicBlock::iterator &I) {
112 if (!isAtTop(RegNo)) {
113 unsigned STReg = getSTReg(RegNo);
114 unsigned RegOnTop = getStackEntry(0);
116 // Swap the slots the regs are in
117 std::swap(RegMap[RegNo], RegMap[RegOnTop]);
119 // Swap stack slot contents
120 assert(RegMap[RegOnTop] < StackTop);
121 std::swap(Stack[RegMap[RegOnTop]], Stack[StackTop-1]);
123 // Emit an fxch to update the runtime processors version of the state
124 BuildMI(*MBB, I, TII->get(X86::XCH_F)).addReg(STReg);
129 void duplicateToTop(unsigned RegNo, unsigned AsReg, MachineInstr *I) {
130 unsigned STReg = getSTReg(RegNo);
131 pushReg(AsReg); // New register on top of stack
133 BuildMI(*MBB, I, TII->get(X86::LD_Frr)).addReg(STReg);
136 // popStackAfter - Pop the current value off of the top of the FP stack
137 // after the specified instruction.
138 void popStackAfter(MachineBasicBlock::iterator &I);
140 // freeStackSlotAfter - Free the specified register from the register stack,
141 // so that it is no longer in a register. If the register is currently at
142 // the top of the stack, we just pop the current instruction, otherwise we
143 // store the current top-of-stack into the specified slot, then pop the top
145 void freeStackSlotAfter(MachineBasicBlock::iterator &I, unsigned Reg);
147 bool processBasicBlock(MachineFunction &MF, MachineBasicBlock &MBB);
149 void handleZeroArgFP(MachineBasicBlock::iterator &I);
150 void handleOneArgFP(MachineBasicBlock::iterator &I);
151 void handleOneArgFPRW(MachineBasicBlock::iterator &I);
152 void handleTwoArgFP(MachineBasicBlock::iterator &I);
153 void handleCompareFP(MachineBasicBlock::iterator &I);
154 void handleCondMovFP(MachineBasicBlock::iterator &I);
155 void handleSpecialFP(MachineBasicBlock::iterator &I);
160 FunctionPass *llvm::createX86FloatingPointStackifierPass() { return new FPS(); }
162 /// runOnMachineFunction - Loop over all of the basic blocks, transforming FP
163 /// register references into FP stack references.
165 bool FPS::runOnMachineFunction(MachineFunction &MF) {
166 // We only need to run this pass if there are any FP registers used in this
167 // function. If it is all integer, there is nothing for us to do!
168 bool FPIsUsed = false;
170 assert(X86::FP6 == X86::FP0+6 && "Register enums aren't sorted right!");
171 for (unsigned i = 0; i <= 6; ++i)
172 if (MF.isPhysRegUsed(X86::FP0+i)) {
178 if (!FPIsUsed) return false;
180 TII = MF.getTarget().getInstrInfo();
181 LV = &getAnalysis<LiveVariables>();
184 // Process the function in depth first order so that we process at least one
185 // of the predecessors for every reachable block in the function.
186 std::set<MachineBasicBlock*> Processed;
187 MachineBasicBlock *Entry = MF.begin();
189 bool Changed = false;
190 for (df_ext_iterator<MachineBasicBlock*, std::set<MachineBasicBlock*> >
191 I = df_ext_begin(Entry, Processed), E = df_ext_end(Entry, Processed);
193 Changed |= processBasicBlock(MF, **I);
198 /// processBasicBlock - Loop over all of the instructions in the basic block,
199 /// transforming FP instructions into their stack form.
201 bool FPS::processBasicBlock(MachineFunction &MF, MachineBasicBlock &BB) {
202 bool Changed = false;
205 for (MachineBasicBlock::iterator I = BB.begin(); I != BB.end(); ++I) {
206 MachineInstr *MI = I;
207 unsigned Flags = MI->getInstrDescriptor()->TSFlags;
208 if ((Flags & X86II::FPTypeMask) == X86II::NotFP)
209 continue; // Efficiently ignore non-fp insts!
211 MachineInstr *PrevMI = 0;
215 ++NumFP; // Keep track of # of pseudo instrs
216 DOUT << "\nFPInst:\t" << *MI;
218 // Get dead variables list now because the MI pointer may be deleted as part
220 SmallVector<unsigned, 8> DeadRegs;
221 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
222 const MachineOperand &MO = MI->getOperand(i);
223 if (MO.isReg() && MO.isDead())
224 DeadRegs.push_back(MO.getReg());
227 switch (Flags & X86II::FPTypeMask) {
228 case X86II::ZeroArgFP: handleZeroArgFP(I); break;
229 case X86II::OneArgFP: handleOneArgFP(I); break; // fstp ST(0)
230 case X86II::OneArgFPRW: handleOneArgFPRW(I); break; // ST(0) = fsqrt(ST(0))
231 case X86II::TwoArgFP: handleTwoArgFP(I); break;
232 case X86II::CompareFP: handleCompareFP(I); break;
233 case X86II::CondMovFP: handleCondMovFP(I); break;
234 case X86II::SpecialFP: handleSpecialFP(I); break;
235 default: assert(0 && "Unknown FP Type!");
238 // Check to see if any of the values defined by this instruction are dead
239 // after definition. If so, pop them.
240 for (unsigned i = 0, e = DeadRegs.size(); i != e; ++i) {
241 unsigned Reg = DeadRegs[i];
242 if (Reg >= X86::FP0 && Reg <= X86::FP6) {
243 DOUT << "Register FP#" << Reg-X86::FP0 << " is dead!\n";
244 freeStackSlotAfter(I, Reg-X86::FP0);
248 // Print out all of the instructions expanded to if -debug
250 MachineBasicBlock::iterator PrevI(PrevMI);
252 cerr << "Just deleted pseudo instruction\n";
254 MachineBasicBlock::iterator Start = I;
255 // Rewind to first instruction newly inserted.
256 while (Start != BB.begin() && prior(Start) != PrevI) --Start;
257 cerr << "Inserted instructions:\n\t";
258 Start->print(*cerr.stream(), &MF.getTarget());
259 while (++Start != next(I));
267 assert(StackTop == 0 && "Stack not empty at end of basic block?");
271 //===----------------------------------------------------------------------===//
272 // Efficient Lookup Table Support
273 //===----------------------------------------------------------------------===//
279 bool operator<(const TableEntry &TE) const { return from < TE.from; }
280 friend bool operator<(const TableEntry &TE, unsigned V) {
283 friend bool operator<(unsigned V, const TableEntry &TE) {
289 static bool TableIsSorted(const TableEntry *Table, unsigned NumEntries) {
290 for (unsigned i = 0; i != NumEntries-1; ++i)
291 if (!(Table[i] < Table[i+1])) return false;
295 static int Lookup(const TableEntry *Table, unsigned N, unsigned Opcode) {
296 const TableEntry *I = std::lower_bound(Table, Table+N, Opcode);
297 if (I != Table+N && I->from == Opcode)
302 #define ARRAY_SIZE(TABLE) \
303 (sizeof(TABLE)/sizeof(TABLE[0]))
306 #define ASSERT_SORTED(TABLE)
308 #define ASSERT_SORTED(TABLE) \
309 { static bool TABLE##Checked = false; \
310 if (!TABLE##Checked) { \
311 assert(TableIsSorted(TABLE, ARRAY_SIZE(TABLE)) && \
312 "All lookup tables must be sorted for efficient access!"); \
313 TABLE##Checked = true; \
318 //===----------------------------------------------------------------------===//
319 // Register File -> Register Stack Mapping Methods
320 //===----------------------------------------------------------------------===//
322 // OpcodeTable - Sorted map of register instructions to their stack version.
323 // The first element is an register file pseudo instruction, the second is the
324 // concrete X86 instruction which uses the register stack.
326 static const TableEntry OpcodeTable[] = {
327 { X86::ABS_Fp32 , X86::ABS_F },
328 { X86::ABS_Fp64 , X86::ABS_F },
329 { X86::ABS_Fp80 , X86::ABS_F },
330 { X86::ADD_Fp32m , X86::ADD_F32m },
331 { X86::ADD_Fp64m , X86::ADD_F64m },
332 { X86::ADD_Fp64m32 , X86::ADD_F32m },
333 { X86::ADD_Fp80m32 , X86::ADD_F32m },
334 { X86::ADD_Fp80m64 , X86::ADD_F64m },
335 { X86::ADD_FpI16m32 , X86::ADD_FI16m },
336 { X86::ADD_FpI16m64 , X86::ADD_FI16m },
337 { X86::ADD_FpI16m80 , X86::ADD_FI16m },
338 { X86::ADD_FpI32m32 , X86::ADD_FI32m },
339 { X86::ADD_FpI32m64 , X86::ADD_FI32m },
340 { X86::ADD_FpI32m80 , X86::ADD_FI32m },
341 { X86::CHS_Fp32 , X86::CHS_F },
342 { X86::CHS_Fp64 , X86::CHS_F },
343 { X86::CHS_Fp80 , X86::CHS_F },
344 { X86::CMOVBE_Fp32 , X86::CMOVBE_F },
345 { X86::CMOVBE_Fp64 , X86::CMOVBE_F },
346 { X86::CMOVBE_Fp80 , X86::CMOVBE_F },
347 { X86::CMOVB_Fp32 , X86::CMOVB_F },
348 { X86::CMOVB_Fp64 , X86::CMOVB_F },
349 { X86::CMOVB_Fp80 , X86::CMOVB_F },
350 { X86::CMOVE_Fp32 , X86::CMOVE_F },
351 { X86::CMOVE_Fp64 , X86::CMOVE_F },
352 { X86::CMOVE_Fp80 , X86::CMOVE_F },
353 { X86::CMOVNBE_Fp32 , X86::CMOVNBE_F },
354 { X86::CMOVNBE_Fp64 , X86::CMOVNBE_F },
355 { X86::CMOVNBE_Fp80 , X86::CMOVNBE_F },
356 { X86::CMOVNB_Fp32 , X86::CMOVNB_F },
357 { X86::CMOVNB_Fp64 , X86::CMOVNB_F },
358 { X86::CMOVNB_Fp80 , X86::CMOVNB_F },
359 { X86::CMOVNE_Fp32 , X86::CMOVNE_F },
360 { X86::CMOVNE_Fp64 , X86::CMOVNE_F },
361 { X86::CMOVNE_Fp80 , X86::CMOVNE_F },
362 { X86::CMOVNP_Fp32 , X86::CMOVNP_F },
363 { X86::CMOVNP_Fp64 , X86::CMOVNP_F },
364 { X86::CMOVNP_Fp80 , X86::CMOVNP_F },
365 { X86::CMOVP_Fp32 , X86::CMOVP_F },
366 { X86::CMOVP_Fp64 , X86::CMOVP_F },
367 { X86::CMOVP_Fp80 , X86::CMOVP_F },
368 { X86::COS_Fp32 , X86::COS_F },
369 { X86::COS_Fp64 , X86::COS_F },
370 { X86::COS_Fp80 , X86::COS_F },
371 { X86::DIVR_Fp32m , X86::DIVR_F32m },
372 { X86::DIVR_Fp64m , X86::DIVR_F64m },
373 { X86::DIVR_Fp64m32 , X86::DIVR_F32m },
374 { X86::DIVR_Fp80m32 , X86::DIVR_F32m },
375 { X86::DIVR_Fp80m64 , X86::DIVR_F64m },
376 { X86::DIVR_FpI16m32, X86::DIVR_FI16m},
377 { X86::DIVR_FpI16m64, X86::DIVR_FI16m},
378 { X86::DIVR_FpI16m80, X86::DIVR_FI16m},
379 { X86::DIVR_FpI32m32, X86::DIVR_FI32m},
380 { X86::DIVR_FpI32m64, X86::DIVR_FI32m},
381 { X86::DIVR_FpI32m80, X86::DIVR_FI32m},
382 { X86::DIV_Fp32m , X86::DIV_F32m },
383 { X86::DIV_Fp64m , X86::DIV_F64m },
384 { X86::DIV_Fp64m32 , X86::DIV_F32m },
385 { X86::DIV_Fp80m32 , X86::DIV_F32m },
386 { X86::DIV_Fp80m64 , X86::DIV_F64m },
387 { X86::DIV_FpI16m32 , X86::DIV_FI16m },
388 { X86::DIV_FpI16m64 , X86::DIV_FI16m },
389 { X86::DIV_FpI16m80 , X86::DIV_FI16m },
390 { X86::DIV_FpI32m32 , X86::DIV_FI32m },
391 { X86::DIV_FpI32m64 , X86::DIV_FI32m },
392 { X86::DIV_FpI32m80 , X86::DIV_FI32m },
393 { X86::ILD_Fp16m32 , X86::ILD_F16m },
394 { X86::ILD_Fp16m64 , X86::ILD_F16m },
395 { X86::ILD_Fp16m80 , X86::ILD_F16m },
396 { X86::ILD_Fp32m32 , X86::ILD_F32m },
397 { X86::ILD_Fp32m64 , X86::ILD_F32m },
398 { X86::ILD_Fp32m80 , X86::ILD_F32m },
399 { X86::ILD_Fp64m32 , X86::ILD_F64m },
400 { X86::ILD_Fp64m64 , X86::ILD_F64m },
401 { X86::ILD_Fp64m80 , X86::ILD_F64m },
402 { X86::ISTT_Fp16m32 , X86::ISTT_FP16m},
403 { X86::ISTT_Fp16m64 , X86::ISTT_FP16m},
404 { X86::ISTT_Fp32m32 , X86::ISTT_FP32m},
405 { X86::ISTT_Fp32m64 , X86::ISTT_FP32m},
406 { X86::ISTT_Fp64m32 , X86::ISTT_FP64m},
407 { X86::ISTT_Fp64m64 , X86::ISTT_FP64m},
408 { X86::IST_Fp16m32 , X86::IST_F16m },
409 { X86::IST_Fp16m64 , X86::IST_F16m },
410 { X86::IST_Fp16m80 , X86::IST_F16m },
411 { X86::IST_Fp32m32 , X86::IST_F32m },
412 { X86::IST_Fp32m64 , X86::IST_F32m },
413 { X86::IST_Fp32m80 , X86::IST_F32m },
414 { X86::IST_Fp64m32 , X86::IST_FP64m },
415 { X86::IST_Fp64m64 , X86::IST_FP64m },
416 { X86::IST_Fp64m80 , X86::IST_FP64m },
417 { X86::LD_Fp032 , X86::LD_F0 },
418 { X86::LD_Fp064 , X86::LD_F0 },
419 { X86::LD_Fp080 , X86::LD_F0 },
420 { X86::LD_Fp132 , X86::LD_F1 },
421 { X86::LD_Fp164 , X86::LD_F1 },
422 { X86::LD_Fp180 , X86::LD_F1 },
423 { X86::LD_Fp32m , X86::LD_F32m },
424 { X86::LD_Fp64m , X86::LD_F64m },
425 { X86::LD_Fp80m , X86::LD_F80m },
426 { X86::MUL_Fp32m , X86::MUL_F32m },
427 { X86::MUL_Fp64m , X86::MUL_F64m },
428 { X86::MUL_Fp64m32 , X86::MUL_F32m },
429 { X86::MUL_Fp80m32 , X86::MUL_F32m },
430 { X86::MUL_Fp80m64 , X86::MUL_F64m },
431 { X86::MUL_FpI16m32 , X86::MUL_FI16m },
432 { X86::MUL_FpI16m64 , X86::MUL_FI16m },
433 { X86::MUL_FpI16m80 , X86::MUL_FI16m },
434 { X86::MUL_FpI32m32 , X86::MUL_FI32m },
435 { X86::MUL_FpI32m64 , X86::MUL_FI32m },
436 { X86::MUL_FpI32m80 , X86::MUL_FI32m },
437 { X86::SIN_Fp32 , X86::SIN_F },
438 { X86::SIN_Fp64 , X86::SIN_F },
439 { X86::SIN_Fp80 , X86::SIN_F },
440 { X86::SQRT_Fp32 , X86::SQRT_F },
441 { X86::SQRT_Fp64 , X86::SQRT_F },
442 { X86::SQRT_Fp80 , X86::SQRT_F },
443 { X86::ST_Fp32m , X86::ST_F32m },
444 { X86::ST_Fp64m , X86::ST_F64m },
445 { X86::ST_Fp64m32 , X86::ST_F32m },
446 { X86::ST_Fp80m32 , X86::ST_F32m },
447 { X86::ST_Fp80m64 , X86::ST_F64m },
448 { X86::ST_FpP80m , X86::ST_FP80m },
449 { X86::SUBR_Fp32m , X86::SUBR_F32m },
450 { X86::SUBR_Fp64m , X86::SUBR_F64m },
451 { X86::SUBR_Fp64m32 , X86::SUBR_F32m },
452 { X86::SUBR_Fp80m32 , X86::SUBR_F32m },
453 { X86::SUBR_Fp80m64 , X86::SUBR_F64m },
454 { X86::SUBR_FpI16m32, X86::SUBR_FI16m},
455 { X86::SUBR_FpI16m64, X86::SUBR_FI16m},
456 { X86::SUBR_FpI16m80, X86::SUBR_FI16m},
457 { X86::SUBR_FpI32m32, X86::SUBR_FI32m},
458 { X86::SUBR_FpI32m64, X86::SUBR_FI32m},
459 { X86::SUBR_FpI32m80, X86::SUBR_FI32m},
460 { X86::SUB_Fp32m , X86::SUB_F32m },
461 { X86::SUB_Fp64m , X86::SUB_F64m },
462 { X86::SUB_Fp64m32 , X86::SUB_F32m },
463 { X86::SUB_Fp80m32 , X86::SUB_F32m },
464 { X86::SUB_Fp80m64 , X86::SUB_F64m },
465 { X86::SUB_FpI16m32 , X86::SUB_FI16m },
466 { X86::SUB_FpI16m64 , X86::SUB_FI16m },
467 { X86::SUB_FpI16m80 , X86::SUB_FI16m },
468 { X86::SUB_FpI32m32 , X86::SUB_FI32m },
469 { X86::SUB_FpI32m64 , X86::SUB_FI32m },
470 { X86::SUB_FpI32m80 , X86::SUB_FI32m },
471 { X86::TST_Fp32 , X86::TST_F },
472 { X86::TST_Fp64 , X86::TST_F },
473 { X86::TST_Fp80 , X86::TST_F },
474 { X86::UCOM_FpIr32 , X86::UCOM_FIr },
475 { X86::UCOM_FpIr64 , X86::UCOM_FIr },
476 { X86::UCOM_FpIr80 , X86::UCOM_FIr },
477 { X86::UCOM_Fpr32 , X86::UCOM_Fr },
478 { X86::UCOM_Fpr64 , X86::UCOM_Fr },
479 { X86::UCOM_Fpr80 , X86::UCOM_Fr },
482 static unsigned getConcreteOpcode(unsigned Opcode) {
483 ASSERT_SORTED(OpcodeTable);
484 int Opc = Lookup(OpcodeTable, ARRAY_SIZE(OpcodeTable), Opcode);
485 assert(Opc != -1 && "FP Stack instruction not in OpcodeTable!");
489 //===----------------------------------------------------------------------===//
491 //===----------------------------------------------------------------------===//
493 // PopTable - Sorted map of instructions to their popping version. The first
494 // element is an instruction, the second is the version which pops.
496 static const TableEntry PopTable[] = {
497 { X86::ADD_FrST0 , X86::ADD_FPrST0 },
499 { X86::DIVR_FrST0, X86::DIVR_FPrST0 },
500 { X86::DIV_FrST0 , X86::DIV_FPrST0 },
502 { X86::IST_F16m , X86::IST_FP16m },
503 { X86::IST_F32m , X86::IST_FP32m },
505 { X86::MUL_FrST0 , X86::MUL_FPrST0 },
507 { X86::ST_F32m , X86::ST_FP32m },
508 { X86::ST_F64m , X86::ST_FP64m },
509 { X86::ST_Frr , X86::ST_FPrr },
511 { X86::SUBR_FrST0, X86::SUBR_FPrST0 },
512 { X86::SUB_FrST0 , X86::SUB_FPrST0 },
514 { X86::UCOM_FIr , X86::UCOM_FIPr },
516 { X86::UCOM_FPr , X86::UCOM_FPPr },
517 { X86::UCOM_Fr , X86::UCOM_FPr },
520 /// popStackAfter - Pop the current value off of the top of the FP stack after
521 /// the specified instruction. This attempts to be sneaky and combine the pop
522 /// into the instruction itself if possible. The iterator is left pointing to
523 /// the last instruction, be it a new pop instruction inserted, or the old
524 /// instruction if it was modified in place.
526 void FPS::popStackAfter(MachineBasicBlock::iterator &I) {
527 ASSERT_SORTED(PopTable);
528 assert(StackTop > 0 && "Cannot pop empty stack!");
529 RegMap[Stack[--StackTop]] = ~0; // Update state
531 // Check to see if there is a popping version of this instruction...
532 int Opcode = Lookup(PopTable, ARRAY_SIZE(PopTable), I->getOpcode());
534 I->setInstrDescriptor(TII->get(Opcode));
535 if (Opcode == X86::UCOM_FPPr)
537 } else { // Insert an explicit pop
538 I = BuildMI(*MBB, ++I, TII->get(X86::ST_FPrr)).addReg(X86::ST0);
542 /// freeStackSlotAfter - Free the specified register from the register stack, so
543 /// that it is no longer in a register. If the register is currently at the top
544 /// of the stack, we just pop the current instruction, otherwise we store the
545 /// current top-of-stack into the specified slot, then pop the top of stack.
546 void FPS::freeStackSlotAfter(MachineBasicBlock::iterator &I, unsigned FPRegNo) {
547 if (getStackEntry(0) == FPRegNo) { // already at the top of stack? easy.
552 // Otherwise, store the top of stack into the dead slot, killing the operand
553 // without having to add in an explicit xchg then pop.
555 unsigned STReg = getSTReg(FPRegNo);
556 unsigned OldSlot = getSlot(FPRegNo);
557 unsigned TopReg = Stack[StackTop-1];
558 Stack[OldSlot] = TopReg;
559 RegMap[TopReg] = OldSlot;
560 RegMap[FPRegNo] = ~0;
561 Stack[--StackTop] = ~0;
562 I = BuildMI(*MBB, ++I, TII->get(X86::ST_FPrr)).addReg(STReg);
566 static unsigned getFPReg(const MachineOperand &MO) {
567 assert(MO.isRegister() && "Expected an FP register!");
568 unsigned Reg = MO.getReg();
569 assert(Reg >= X86::FP0 && Reg <= X86::FP6 && "Expected FP register!");
570 return Reg - X86::FP0;
574 //===----------------------------------------------------------------------===//
575 // Instruction transformation implementation
576 //===----------------------------------------------------------------------===//
578 /// handleZeroArgFP - ST(0) = fld0 ST(0) = flds <mem>
580 void FPS::handleZeroArgFP(MachineBasicBlock::iterator &I) {
581 MachineInstr *MI = I;
582 unsigned DestReg = getFPReg(MI->getOperand(0));
584 // Change from the pseudo instruction to the concrete instruction.
585 MI->RemoveOperand(0); // Remove the explicit ST(0) operand
586 MI->setInstrDescriptor(TII->get(getConcreteOpcode(MI->getOpcode())));
588 // Result gets pushed on the stack.
592 /// handleOneArgFP - fst <mem>, ST(0)
594 void FPS::handleOneArgFP(MachineBasicBlock::iterator &I) {
595 MachineInstr *MI = I;
596 unsigned NumOps = MI->getInstrDescriptor()->numOperands;
597 assert((NumOps == 5 || NumOps == 1) &&
598 "Can only handle fst* & ftst instructions!");
600 // Is this the last use of the source register?
601 unsigned Reg = getFPReg(MI->getOperand(NumOps-1));
602 bool KillsSrc = LV->KillsRegister(MI, X86::FP0+Reg);
604 // FISTP64m is strange because there isn't a non-popping versions.
605 // If we have one _and_ we don't want to pop the operand, duplicate the value
606 // on the stack instead of moving it. This ensure that popping the value is
608 // Ditto FISTTP16m, FISTTP32m, FISTTP64m.
611 (MI->getOpcode() == X86::IST_Fp64m32 ||
612 MI->getOpcode() == X86::ISTT_Fp16m32 ||
613 MI->getOpcode() == X86::ISTT_Fp32m32 ||
614 MI->getOpcode() == X86::ISTT_Fp64m32 ||
615 MI->getOpcode() == X86::IST_Fp64m64 ||
616 MI->getOpcode() == X86::ISTT_Fp16m64 ||
617 MI->getOpcode() == X86::ISTT_Fp32m64 ||
618 MI->getOpcode() == X86::ISTT_Fp64m64 ||
619 MI->getOpcode() == X86::ST_FpP80m)) {
620 duplicateToTop(Reg, 7 /*temp register*/, I);
622 moveToTop(Reg, I); // Move to the top of the stack...
625 // Convert from the pseudo instruction to the concrete instruction.
626 MI->RemoveOperand(NumOps-1); // Remove explicit ST(0) operand
627 MI->setInstrDescriptor(TII->get(getConcreteOpcode(MI->getOpcode())));
629 if (MI->getOpcode() == X86::IST_FP64m ||
630 MI->getOpcode() == X86::ISTT_FP16m ||
631 MI->getOpcode() == X86::ISTT_FP32m ||
632 MI->getOpcode() == X86::ISTT_FP64m ||
633 MI->getOpcode() == X86::ST_FP80m) {
634 assert(StackTop > 0 && "Stack empty??");
636 } else if (KillsSrc) { // Last use of operand?
642 /// handleOneArgFPRW: Handle instructions that read from the top of stack and
643 /// replace the value with a newly computed value. These instructions may have
644 /// non-fp operands after their FP operands.
648 /// R1 = fadd R2, [mem]
650 void FPS::handleOneArgFPRW(MachineBasicBlock::iterator &I) {
651 MachineInstr *MI = I;
652 unsigned NumOps = MI->getInstrDescriptor()->numOperands;
653 assert(NumOps >= 2 && "FPRW instructions must have 2 ops!!");
655 // Is this the last use of the source register?
656 unsigned Reg = getFPReg(MI->getOperand(1));
657 bool KillsSrc = LV->KillsRegister(MI, X86::FP0+Reg);
660 // If this is the last use of the source register, just make sure it's on
661 // the top of the stack.
663 assert(StackTop > 0 && "Stack cannot be empty!");
665 pushReg(getFPReg(MI->getOperand(0)));
667 // If this is not the last use of the source register, _copy_ it to the top
669 duplicateToTop(Reg, getFPReg(MI->getOperand(0)), I);
672 // Change from the pseudo instruction to the concrete instruction.
673 MI->RemoveOperand(1); // Drop the source operand.
674 MI->RemoveOperand(0); // Drop the destination operand.
675 MI->setInstrDescriptor(TII->get(getConcreteOpcode(MI->getOpcode())));
679 //===----------------------------------------------------------------------===//
680 // Define tables of various ways to map pseudo instructions
683 // ForwardST0Table - Map: A = B op C into: ST(0) = ST(0) op ST(i)
684 static const TableEntry ForwardST0Table[] = {
685 { X86::ADD_Fp32 , X86::ADD_FST0r },
686 { X86::ADD_Fp64 , X86::ADD_FST0r },
687 { X86::ADD_Fp80 , X86::ADD_FST0r },
688 { X86::DIV_Fp32 , X86::DIV_FST0r },
689 { X86::DIV_Fp64 , X86::DIV_FST0r },
690 { X86::DIV_Fp80 , X86::DIV_FST0r },
691 { X86::MUL_Fp32 , X86::MUL_FST0r },
692 { X86::MUL_Fp64 , X86::MUL_FST0r },
693 { X86::MUL_Fp80 , X86::MUL_FST0r },
694 { X86::SUB_Fp32 , X86::SUB_FST0r },
695 { X86::SUB_Fp64 , X86::SUB_FST0r },
696 { X86::SUB_Fp80 , X86::SUB_FST0r },
699 // ReverseST0Table - Map: A = B op C into: ST(0) = ST(i) op ST(0)
700 static const TableEntry ReverseST0Table[] = {
701 { X86::ADD_Fp32 , X86::ADD_FST0r }, // commutative
702 { X86::ADD_Fp64 , X86::ADD_FST0r }, // commutative
703 { X86::ADD_Fp80 , X86::ADD_FST0r }, // commutative
704 { X86::DIV_Fp32 , X86::DIVR_FST0r },
705 { X86::DIV_Fp64 , X86::DIVR_FST0r },
706 { X86::DIV_Fp80 , X86::DIVR_FST0r },
707 { X86::MUL_Fp32 , X86::MUL_FST0r }, // commutative
708 { X86::MUL_Fp64 , X86::MUL_FST0r }, // commutative
709 { X86::MUL_Fp80 , X86::MUL_FST0r }, // commutative
710 { X86::SUB_Fp32 , X86::SUBR_FST0r },
711 { X86::SUB_Fp64 , X86::SUBR_FST0r },
712 { X86::SUB_Fp80 , X86::SUBR_FST0r },
715 // ForwardSTiTable - Map: A = B op C into: ST(i) = ST(0) op ST(i)
716 static const TableEntry ForwardSTiTable[] = {
717 { X86::ADD_Fp32 , X86::ADD_FrST0 }, // commutative
718 { X86::ADD_Fp64 , X86::ADD_FrST0 }, // commutative
719 { X86::ADD_Fp80 , X86::ADD_FrST0 }, // commutative
720 { X86::DIV_Fp32 , X86::DIVR_FrST0 },
721 { X86::DIV_Fp64 , X86::DIVR_FrST0 },
722 { X86::DIV_Fp80 , X86::DIVR_FrST0 },
723 { X86::MUL_Fp32 , X86::MUL_FrST0 }, // commutative
724 { X86::MUL_Fp64 , X86::MUL_FrST0 }, // commutative
725 { X86::MUL_Fp80 , X86::MUL_FrST0 }, // commutative
726 { X86::SUB_Fp32 , X86::SUBR_FrST0 },
727 { X86::SUB_Fp64 , X86::SUBR_FrST0 },
728 { X86::SUB_Fp80 , X86::SUBR_FrST0 },
731 // ReverseSTiTable - Map: A = B op C into: ST(i) = ST(i) op ST(0)
732 static const TableEntry ReverseSTiTable[] = {
733 { X86::ADD_Fp32 , X86::ADD_FrST0 },
734 { X86::ADD_Fp64 , X86::ADD_FrST0 },
735 { X86::ADD_Fp80 , X86::ADD_FrST0 },
736 { X86::DIV_Fp32 , X86::DIV_FrST0 },
737 { X86::DIV_Fp64 , X86::DIV_FrST0 },
738 { X86::DIV_Fp80 , X86::DIV_FrST0 },
739 { X86::MUL_Fp32 , X86::MUL_FrST0 },
740 { X86::MUL_Fp64 , X86::MUL_FrST0 },
741 { X86::MUL_Fp80 , X86::MUL_FrST0 },
742 { X86::SUB_Fp32 , X86::SUB_FrST0 },
743 { X86::SUB_Fp64 , X86::SUB_FrST0 },
744 { X86::SUB_Fp80 , X86::SUB_FrST0 },
748 /// handleTwoArgFP - Handle instructions like FADD and friends which are virtual
749 /// instructions which need to be simplified and possibly transformed.
751 /// Result: ST(0) = fsub ST(0), ST(i)
752 /// ST(i) = fsub ST(0), ST(i)
753 /// ST(0) = fsubr ST(0), ST(i)
754 /// ST(i) = fsubr ST(0), ST(i)
756 void FPS::handleTwoArgFP(MachineBasicBlock::iterator &I) {
757 ASSERT_SORTED(ForwardST0Table); ASSERT_SORTED(ReverseST0Table);
758 ASSERT_SORTED(ForwardSTiTable); ASSERT_SORTED(ReverseSTiTable);
759 MachineInstr *MI = I;
761 unsigned NumOperands = MI->getInstrDescriptor()->numOperands;
762 assert(NumOperands == 3 && "Illegal TwoArgFP instruction!");
763 unsigned Dest = getFPReg(MI->getOperand(0));
764 unsigned Op0 = getFPReg(MI->getOperand(NumOperands-2));
765 unsigned Op1 = getFPReg(MI->getOperand(NumOperands-1));
766 bool KillsOp0 = LV->KillsRegister(MI, X86::FP0+Op0);
767 bool KillsOp1 = LV->KillsRegister(MI, X86::FP0+Op1);
769 unsigned TOS = getStackEntry(0);
771 // One of our operands must be on the top of the stack. If neither is yet, we
773 if (Op0 != TOS && Op1 != TOS) { // No operand at TOS?
774 // We can choose to move either operand to the top of the stack. If one of
775 // the operands is killed by this instruction, we want that one so that we
776 // can update right on top of the old version.
778 moveToTop(Op0, I); // Move dead operand to TOS.
780 } else if (KillsOp1) {
784 // All of the operands are live after this instruction executes, so we
785 // cannot update on top of any operand. Because of this, we must
786 // duplicate one of the stack elements to the top. It doesn't matter
787 // which one we pick.
789 duplicateToTop(Op0, Dest, I);
793 } else if (!KillsOp0 && !KillsOp1) {
794 // If we DO have one of our operands at the top of the stack, but we don't
795 // have a dead operand, we must duplicate one of the operands to a new slot
797 duplicateToTop(Op0, Dest, I);
802 // Now we know that one of our operands is on the top of the stack, and at
803 // least one of our operands is killed by this instruction.
804 assert((TOS == Op0 || TOS == Op1) && (KillsOp0 || KillsOp1) &&
805 "Stack conditions not set up right!");
807 // We decide which form to use based on what is on the top of the stack, and
808 // which operand is killed by this instruction.
809 const TableEntry *InstTable;
810 bool isForward = TOS == Op0;
811 bool updateST0 = (TOS == Op0 && !KillsOp1) || (TOS == Op1 && !KillsOp0);
814 InstTable = ForwardST0Table;
816 InstTable = ReverseST0Table;
819 InstTable = ForwardSTiTable;
821 InstTable = ReverseSTiTable;
824 int Opcode = Lookup(InstTable, ARRAY_SIZE(ForwardST0Table), MI->getOpcode());
825 assert(Opcode != -1 && "Unknown TwoArgFP pseudo instruction!");
827 // NotTOS - The register which is not on the top of stack...
828 unsigned NotTOS = (TOS == Op0) ? Op1 : Op0;
830 // Replace the old instruction with a new instruction
832 I = BuildMI(*MBB, I, TII->get(Opcode)).addReg(getSTReg(NotTOS));
834 // If both operands are killed, pop one off of the stack in addition to
835 // overwriting the other one.
836 if (KillsOp0 && KillsOp1 && Op0 != Op1) {
837 assert(!updateST0 && "Should have updated other operand!");
838 popStackAfter(I); // Pop the top of stack
841 // Update stack information so that we know the destination register is now on
843 unsigned UpdatedSlot = getSlot(updateST0 ? TOS : NotTOS);
844 assert(UpdatedSlot < StackTop && Dest < 7);
845 Stack[UpdatedSlot] = Dest;
846 RegMap[Dest] = UpdatedSlot;
847 delete MI; // Remove the old instruction
850 /// handleCompareFP - Handle FUCOM and FUCOMI instructions, which have two FP
851 /// register arguments and no explicit destinations.
853 void FPS::handleCompareFP(MachineBasicBlock::iterator &I) {
854 ASSERT_SORTED(ForwardST0Table); ASSERT_SORTED(ReverseST0Table);
855 ASSERT_SORTED(ForwardSTiTable); ASSERT_SORTED(ReverseSTiTable);
856 MachineInstr *MI = I;
858 unsigned NumOperands = MI->getInstrDescriptor()->numOperands;
859 assert(NumOperands == 2 && "Illegal FUCOM* instruction!");
860 unsigned Op0 = getFPReg(MI->getOperand(NumOperands-2));
861 unsigned Op1 = getFPReg(MI->getOperand(NumOperands-1));
862 bool KillsOp0 = LV->KillsRegister(MI, X86::FP0+Op0);
863 bool KillsOp1 = LV->KillsRegister(MI, X86::FP0+Op1);
865 // Make sure the first operand is on the top of stack, the other one can be
869 // Change from the pseudo instruction to the concrete instruction.
870 MI->getOperand(0).setReg(getSTReg(Op1));
871 MI->RemoveOperand(1);
872 MI->setInstrDescriptor(TII->get(getConcreteOpcode(MI->getOpcode())));
874 // If any of the operands are killed by this instruction, free them.
875 if (KillsOp0) freeStackSlotAfter(I, Op0);
876 if (KillsOp1 && Op0 != Op1) freeStackSlotAfter(I, Op1);
879 /// handleCondMovFP - Handle two address conditional move instructions. These
880 /// instructions move a st(i) register to st(0) iff a condition is true. These
881 /// instructions require that the first operand is at the top of the stack, but
882 /// otherwise don't modify the stack at all.
883 void FPS::handleCondMovFP(MachineBasicBlock::iterator &I) {
884 MachineInstr *MI = I;
886 unsigned Op0 = getFPReg(MI->getOperand(0));
887 unsigned Op1 = getFPReg(MI->getOperand(2));
888 bool KillsOp1 = LV->KillsRegister(MI, X86::FP0+Op1);
890 // The first operand *must* be on the top of the stack.
893 // Change the second operand to the stack register that the operand is in.
894 // Change from the pseudo instruction to the concrete instruction.
895 MI->RemoveOperand(0);
896 MI->RemoveOperand(1);
897 MI->getOperand(0).setReg(getSTReg(Op1));
898 MI->setInstrDescriptor(TII->get(getConcreteOpcode(MI->getOpcode())));
900 // If we kill the second operand, make sure to pop it from the stack.
901 if (Op0 != Op1 && KillsOp1) {
902 // Get this value off of the register stack.
903 freeStackSlotAfter(I, Op1);
908 /// handleSpecialFP - Handle special instructions which behave unlike other
909 /// floating point instructions. This is primarily intended for use by pseudo
912 void FPS::handleSpecialFP(MachineBasicBlock::iterator &I) {
913 MachineInstr *MI = I;
914 switch (MI->getOpcode()) {
915 default: assert(0 && "Unknown SpecialFP instruction!");
916 case X86::FpGETRESULT32: // Appears immediately after a call returning FP type!
917 case X86::FpGETRESULT64: // Appears immediately after a call returning FP type!
918 case X86::FpGETRESULT80:
919 assert(StackTop == 0 && "Stack should be empty after a call!");
920 pushReg(getFPReg(MI->getOperand(0)));
922 case X86::FpSETRESULT32:
923 case X86::FpSETRESULT64:
924 case X86::FpSETRESULT80:
925 assert(StackTop == 1 && "Stack should have one element on it to return!");
926 --StackTop; // "Forget" we have something on the top of stack!
928 case X86::MOV_Fp3232:
929 case X86::MOV_Fp3264:
930 case X86::MOV_Fp6432:
931 case X86::MOV_Fp6464:
932 case X86::MOV_Fp3280:
933 case X86::MOV_Fp6480:
934 case X86::MOV_Fp8032:
935 case X86::MOV_Fp8064:
936 case X86::MOV_Fp8080: {
937 unsigned SrcReg = getFPReg(MI->getOperand(1));
938 unsigned DestReg = getFPReg(MI->getOperand(0));
940 if (LV->KillsRegister(MI, X86::FP0+SrcReg)) {
941 // If the input operand is killed, we can just change the owner of the
942 // incoming stack slot into the result.
943 unsigned Slot = getSlot(SrcReg);
944 assert(Slot < 7 && DestReg < 7 && "FpMOV operands invalid!");
945 Stack[Slot] = DestReg;
946 RegMap[DestReg] = Slot;
949 // For FMOV we just duplicate the specified value to a new stack slot.
950 // This could be made better, but would require substantial changes.
951 duplicateToTop(SrcReg, DestReg, I);
957 I = MBB->erase(I); // Remove the pseudo instruction