1//===-- X86FloatingPoint.cpp - Floating point Reg -> Stack converter ------===//
2//
3// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4// See https://llvm.org/LICENSE.txt for license information.
5// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6//
7//===----------------------------------------------------------------------===//
8//
9// This file defines the pass which converts floating point instructions from
10// pseudo registers into register stack instructions. This pass uses live
11// variable information to indicate where the FPn registers are used and their
12// lifetimes.
13//
14// The x87 hardware tracks liveness of the stack registers, so it is necessary
15// to implement exact liveness tracking between basic blocks. The CFG edges are
16// partitioned into bundles where the same FP registers must be live in
17// identical stack positions. Instructions are inserted at the end of each basic
18// block to rearrange the live registers to match the outgoing bundle.
19//
20// This approach avoids splitting critical edges at the potential cost of more
21// live register shuffling instructions when critical edges are present.
22//
23//===----------------------------------------------------------------------===//
24
25#include "X86.h"
26#include "X86InstrInfo.h"
27#include "llvm/ADT/DepthFirstIterator.h"
28#include "llvm/ADT/STLExtras.h"
29#include "llvm/ADT/SmallSet.h"
30#include "llvm/ADT/SmallVector.h"
31#include "llvm/ADT/Statistic.h"
32#include "llvm/CodeGen/EdgeBundles.h"
33#include "llvm/CodeGen/LiveRegUnits.h"
34#include "llvm/CodeGen/MachineFunctionPass.h"
35#include "llvm/CodeGen/MachineInstrBuilder.h"
36#include "llvm/CodeGen/MachineRegisterInfo.h"
37#include "llvm/CodeGen/Passes.h"
38#include "llvm/CodeGen/TargetInstrInfo.h"
39#include "llvm/CodeGen/TargetSubtargetInfo.h"
40#include "llvm/Config/llvm-config.h"
41#include "llvm/IR/InlineAsm.h"
42#include "llvm/InitializePasses.h"
43#include "llvm/Support/Debug.h"
44#include "llvm/Support/ErrorHandling.h"
45#include "llvm/Support/raw_ostream.h"
46#include "llvm/Target/TargetMachine.h"
47#include <algorithm>
48#include <bitset>
49using namespace llvm;
50
51#define DEBUG_TYPE "x86-codegen"
52
53STATISTIC(NumFXCH, "Number of fxch instructions inserted");
54STATISTIC(NumFP , "Number of floating point instructions");
55
56namespace {
57 const unsigned ScratchFPReg = 7;
58
59 struct FPS : public MachineFunctionPass {
60 static char ID;
61 FPS() : MachineFunctionPass(ID) {
62 // This is really only to keep valgrind quiet.
63 // The logic in isLive() is too much for it.
64 memset(s: Stack, c: 0, n: sizeof(Stack));
65 memset(s: RegMap, c: 0, n: sizeof(RegMap));
66 }
67
68 void getAnalysisUsage(AnalysisUsage &AU) const override {
69 AU.setPreservesCFG();
70 AU.addRequired<EdgeBundlesWrapperLegacy>();
71 AU.addPreservedID(ID&: MachineLoopInfoID);
72 AU.addPreservedID(ID&: MachineDominatorsID);
73 MachineFunctionPass::getAnalysisUsage(AU);
74 }
75
76 bool runOnMachineFunction(MachineFunction &MF) override;
77
78 MachineFunctionProperties getRequiredProperties() const override {
79 return MachineFunctionProperties().setNoVRegs();
80 }
81
82 StringRef getPassName() const override { return "X86 FP Stackifier"; }
83
84 private:
85 const TargetInstrInfo *TII = nullptr; // Machine instruction info.
86
87 // Two CFG edges are related if they leave the same block, or enter the same
88 // block. The transitive closure of an edge under this relation is a
89 // LiveBundle. It represents a set of CFG edges where the live FP stack
90 // registers must be allocated identically in the x87 stack.
91 //
92 // A LiveBundle is usually all the edges leaving a block, or all the edges
93 // entering a block, but it can contain more edges if critical edges are
94 // present.
95 //
96 // The set of live FP registers in a LiveBundle is calculated by bundleCFG,
97 // but the exact mapping of FP registers to stack slots is fixed later.
98 struct LiveBundle {
99 // Bit mask of live FP registers. Bit 0 = FP0, bit 1 = FP1, &c.
100 unsigned Mask = 0;
101
102 // Number of pre-assigned live registers in FixStack. This is 0 when the
103 // stack order has not yet been fixed.
104 unsigned FixCount = 0;
105
106 // Assigned stack order for live-in registers.
107 // FixStack[i] == getStackEntry(i) for all i < FixCount.
108 unsigned char FixStack[8];
109
110 LiveBundle() = default;
111
112 // Have the live registers been assigned a stack order yet?
113 bool isFixed() const { return !Mask || FixCount; }
114 };
115
116 // Numbered LiveBundle structs. LiveBundles[0] is used for all CFG edges
117 // with no live FP registers.
118 SmallVector<LiveBundle, 8> LiveBundles;
119
120 // The edge bundle analysis provides indices into the LiveBundles vector.
121 EdgeBundles *Bundles = nullptr;
122
123 // Return a bitmask of FP registers in block's live-in list.
124 static unsigned calcLiveInMask(MachineBasicBlock *MBB, bool RemoveFPs) {
125 unsigned Mask = 0;
126 for (MachineBasicBlock::livein_iterator I = MBB->livein_begin();
127 I != MBB->livein_end(); ) {
128 MCPhysReg Reg = I->PhysReg;
129 static_assert(X86::FP6 - X86::FP0 == 6, "sequential regnums");
130 if (Reg >= X86::FP0 && Reg <= X86::FP6) {
131 Mask |= 1 << (Reg - X86::FP0);
132 if (RemoveFPs) {
133 I = MBB->removeLiveIn(I);
134 continue;
135 }
136 }
137 ++I;
138 }
139 return Mask;
140 }
141
142 // Partition all the CFG edges into LiveBundles.
143 void bundleCFGRecomputeKillFlags(MachineFunction &MF);
144
145 MachineBasicBlock *MBB = nullptr; // Current basic block
146
147 // The hardware keeps track of how many FP registers are live, so we have
148 // to model that exactly. Usually, each live register corresponds to an
149 // FP<n> register, but when dealing with calls, returns, and inline
150 // assembly, it is sometimes necessary to have live scratch registers.
151 unsigned Stack[8]; // FP<n> Registers in each stack slot...
152 unsigned StackTop = 0; // The current top of the FP stack.
153
154 enum {
155 NumFPRegs = 8 // Including scratch pseudo-registers.
156 };
157
158 // For each live FP<n> register, point to its Stack[] entry.
159 // The first entries correspond to FP0-FP6, the rest are scratch registers
160 // used when we need slightly different live registers than what the
161 // register allocator thinks.
162 unsigned RegMap[NumFPRegs];
163
164 // Set up our stack model to match the incoming registers to MBB.
165 void setupBlockStack();
166
167 // Shuffle live registers to match the expectations of successor blocks.
168 void finishBlockStack();
169
170#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
171 void dumpStack() const {
172 dbgs() << "Stack contents:";
173 for (unsigned i = 0; i != StackTop; ++i) {
174 dbgs() << " FP" << Stack[i];
175 assert(RegMap[Stack[i]] == i && "Stack[] doesn't match RegMap[]!");
176 }
177 }
178#endif
179
180 /// getSlot - Return the stack slot number a particular register number is
181 /// in.
182 unsigned getSlot(unsigned RegNo) const {
183 assert(RegNo < NumFPRegs && "Regno out of range!");
184 return RegMap[RegNo];
185 }
186
187 /// isLive - Is RegNo currently live in the stack?
188 bool isLive(unsigned RegNo) const {
189 unsigned Slot = getSlot(RegNo);
190 return Slot < StackTop && Stack[Slot] == RegNo;
191 }
192
193 /// getStackEntry - Return the X86::FP<n> register in register ST(i).
194 unsigned getStackEntry(unsigned STi) const {
195 if (STi >= StackTop)
196 report_fatal_error(reason: "Access past stack top!");
197 return Stack[StackTop-1-STi];
198 }
199
200 /// getSTReg - Return the X86::ST(i) register which contains the specified
201 /// FP<RegNo> register.
202 unsigned getSTReg(unsigned RegNo) const {
203 return StackTop - 1 - getSlot(RegNo) + X86::ST0;
204 }
205
206 // pushReg - Push the specified FP<n> register onto the stack.
207 void pushReg(unsigned Reg) {
208 assert(Reg < NumFPRegs && "Register number out of range!");
209 if (StackTop >= 8)
210 report_fatal_error(reason: "Stack overflow!");
211 Stack[StackTop] = Reg;
212 RegMap[Reg] = StackTop++;
213 }
214
215 // popReg - Pop a register from the stack.
216 void popReg() {
217 if (StackTop == 0)
218 report_fatal_error(reason: "Cannot pop empty stack!");
219 RegMap[Stack[--StackTop]] = ~0; // Update state
220 }
221
222 bool isAtTop(unsigned RegNo) const { return getSlot(RegNo) == StackTop-1; }
223 void moveToTop(unsigned RegNo, MachineBasicBlock::iterator I) {
224 DebugLoc dl = I == MBB->end() ? DebugLoc() : I->getDebugLoc();
225 if (isAtTop(RegNo)) return;
226
227 unsigned STReg = getSTReg(RegNo);
228 unsigned RegOnTop = getStackEntry(STi: 0);
229
230 // Swap the slots the regs are in.
231 std::swap(a&: RegMap[RegNo], b&: RegMap[RegOnTop]);
232
233 // Swap stack slot contents.
234 if (RegMap[RegOnTop] >= StackTop)
235 report_fatal_error(reason: "Access past stack top!");
236 std::swap(a&: Stack[RegMap[RegOnTop]], b&: Stack[StackTop-1]);
237
238 // Emit an fxch to update the runtime processors version of the state.
239 BuildMI(BB&: *MBB, I, MIMD: dl, MCID: TII->get(Opcode: X86::XCH_F)).addReg(RegNo: STReg);
240 ++NumFXCH;
241 }
242
243 void duplicateToTop(unsigned RegNo, unsigned AsReg,
244 MachineBasicBlock::iterator I) {
245 DebugLoc dl = I == MBB->end() ? DebugLoc() : I->getDebugLoc();
246 unsigned STReg = getSTReg(RegNo);
247 pushReg(Reg: AsReg); // New register on top of stack
248
249 BuildMI(BB&: *MBB, I, MIMD: dl, MCID: TII->get(Opcode: X86::LD_Frr)).addReg(RegNo: STReg);
250 }
251
252 /// popStackAfter - Pop the current value off of the top of the FP stack
253 /// after the specified instruction.
254 void popStackAfter(MachineBasicBlock::iterator &I);
255
256 /// freeStackSlotAfter - Free the specified register from the register
257 /// stack, so that it is no longer in a register. If the register is
258 /// currently at the top of the stack, we just pop the current instruction,
259 /// otherwise we store the current top-of-stack into the specified slot,
260 /// then pop the top of stack.
261 void freeStackSlotAfter(MachineBasicBlock::iterator &I, unsigned Reg);
262
263 /// freeStackSlotBefore - Just the pop, no folding. Return the inserted
264 /// instruction.
265 MachineBasicBlock::iterator
266 freeStackSlotBefore(MachineBasicBlock::iterator I, unsigned FPRegNo);
267
268 /// Adjust the live registers to be the set in Mask.
269 void adjustLiveRegs(unsigned Mask, MachineBasicBlock::iterator I);
270
271 /// Shuffle the top FixCount stack entries such that FP reg FixStack[0] is
272 /// st(0), FP reg FixStack[1] is st(1) etc.
273 void shuffleStackTop(const unsigned char *FixStack, unsigned FixCount,
274 MachineBasicBlock::iterator I);
275
276 bool processBasicBlock(MachineFunction &MF, MachineBasicBlock &MBB);
277
278 void handleCall(MachineBasicBlock::iterator &I);
279 void handleReturn(MachineBasicBlock::iterator &I);
280 void handleZeroArgFP(MachineBasicBlock::iterator &I);
281 void handleOneArgFP(MachineBasicBlock::iterator &I);
282 void handleOneArgFPRW(MachineBasicBlock::iterator &I);
283 void handleTwoArgFP(MachineBasicBlock::iterator &I);
284 void handleCompareFP(MachineBasicBlock::iterator &I);
285 void handleCondMovFP(MachineBasicBlock::iterator &I);
286 void handleSpecialFP(MachineBasicBlock::iterator &I);
287
288 // Check if a COPY instruction is using FP registers.
289 static bool isFPCopy(MachineInstr &MI) {
290 Register DstReg = MI.getOperand(i: 0).getReg();
291 Register SrcReg = MI.getOperand(i: 1).getReg();
292
293 return X86::RFP80RegClass.contains(Reg: DstReg) ||
294 X86::RFP80RegClass.contains(Reg: SrcReg);
295 }
296
297 void setKillFlags(MachineBasicBlock &MBB) const;
298 };
299}
300
301char FPS::ID = 0;
302
303INITIALIZE_PASS_BEGIN(FPS, DEBUG_TYPE, "X86 FP Stackifier",
304 false, false)
305INITIALIZE_PASS_DEPENDENCY(EdgeBundlesWrapperLegacy)
306INITIALIZE_PASS_END(FPS, DEBUG_TYPE, "X86 FP Stackifier",
307 false, false)
308
309FunctionPass *llvm::createX86FloatingPointStackifierPass() { return new FPS(); }
310
311/// getFPReg - Return the X86::FPx register number for the specified operand.
312/// For example, this returns 3 for X86::FP3.
313static unsigned getFPReg(const MachineOperand &MO) {
314 assert(MO.isReg() && "Expected an FP register!");
315 Register Reg = MO.getReg();
316 assert(Reg >= X86::FP0 && Reg <= X86::FP6 && "Expected FP register!");
317 return Reg - X86::FP0;
318}
319
320/// runOnMachineFunction - Loop over all of the basic blocks, transforming FP
321/// register references into FP stack references.
322///
323bool FPS::runOnMachineFunction(MachineFunction &MF) {
324 // We only need to run this pass if there are any FP registers used in this
325 // function. If it is all integer, there is nothing for us to do!
326 bool FPIsUsed = false;
327
328 static_assert(X86::FP6 == X86::FP0+6, "Register enums aren't sorted right!");
329 const MachineRegisterInfo &MRI = MF.getRegInfo();
330 for (unsigned i = 0; i <= 6; ++i)
331 if (!MRI.reg_nodbg_empty(RegNo: X86::FP0 + i)) {
332 FPIsUsed = true;
333 break;
334 }
335
336 // Early exit.
337 if (!FPIsUsed) return false;
338
339 Bundles = &getAnalysis<EdgeBundlesWrapperLegacy>().getEdgeBundles();
340 TII = MF.getSubtarget().getInstrInfo();
341
342 // Prepare cross-MBB liveness.
343 bundleCFGRecomputeKillFlags(MF);
344
345 StackTop = 0;
346
347 // Process the function in depth first order so that we process at least one
348 // of the predecessors for every reachable block in the function.
349 df_iterator_default_set<MachineBasicBlock*> Processed;
350 MachineBasicBlock *Entry = &MF.front();
351
352 LiveBundle &Bundle =
353 LiveBundles[Bundles->getBundle(N: Entry->getNumber(), Out: false)];
354
355 // In regcall convention, some FP registers may not be passed through
356 // the stack, so they will need to be assigned to the stack first
357 if ((Entry->getParent()->getFunction().getCallingConv() ==
358 CallingConv::X86_RegCall) && (Bundle.Mask && !Bundle.FixCount)) {
359 // In the register calling convention, up to one FP argument could be
360 // saved in the first FP register.
361 // If bundle.mask is non-zero and Bundle.FixCount is zero, it means
362 // that the FP registers contain arguments.
363 // The actual value is passed in FP0.
364 // Here we fix the stack and mark FP0 as pre-assigned register.
365 assert((Bundle.Mask & 0xFE) == 0 &&
366 "Only FP0 could be passed as an argument");
367 Bundle.FixCount = 1;
368 Bundle.FixStack[0] = 0;
369 }
370
371 bool Changed = false;
372 for (MachineBasicBlock *BB : depth_first_ext(G: Entry, S&: Processed))
373 Changed |= processBasicBlock(MF, MBB&: *BB);
374
375 // Process any unreachable blocks in arbitrary order now.
376 if (MF.size() != Processed.size())
377 for (MachineBasicBlock &BB : MF)
378 if (Processed.insert(N: &BB).second)
379 Changed |= processBasicBlock(MF, MBB&: BB);
380
381 LiveBundles.clear();
382
383 return Changed;
384}
385
386/// bundleCFG - Scan all the basic blocks to determine consistent live-in and
387/// live-out sets for the FP registers. Consistent means that the set of
388/// registers live-out from a block is identical to the live-in set of all
389/// successors. This is not enforced by the normal live-in lists since
390/// registers may be implicitly defined, or not used by all successors.
391void FPS::bundleCFGRecomputeKillFlags(MachineFunction &MF) {
392 assert(LiveBundles.empty() && "Stale data in LiveBundles");
393 LiveBundles.resize(N: Bundles->getNumBundles());
394
395 // Gather the actual live-in masks for all MBBs.
396 for (MachineBasicBlock &MBB : MF) {
397 setKillFlags(MBB);
398
399 const unsigned Mask = calcLiveInMask(MBB: &MBB, RemoveFPs: false);
400 if (!Mask)
401 continue;
402 // Update MBB ingoing bundle mask.
403 LiveBundles[Bundles->getBundle(N: MBB.getNumber(), Out: false)].Mask |= Mask;
404 }
405}
406
407/// processBasicBlock - Loop over all of the instructions in the basic block,
408/// transforming FP instructions into their stack form.
409///
410bool FPS::processBasicBlock(MachineFunction &MF, MachineBasicBlock &BB) {
411 bool Changed = false;
412 MBB = &BB;
413
414 setupBlockStack();
415
416 for (MachineBasicBlock::iterator I = BB.begin(); I != BB.end(); ++I) {
417 MachineInstr &MI = *I;
418 uint64_t Flags = MI.getDesc().TSFlags;
419
420 unsigned FPInstClass = Flags & X86II::FPTypeMask;
421 if (MI.isInlineAsm())
422 FPInstClass = X86II::SpecialFP;
423
424 if (MI.isCopy() && isFPCopy(MI))
425 FPInstClass = X86II::SpecialFP;
426
427 if (MI.isImplicitDef() &&
428 X86::RFP80RegClass.contains(Reg: MI.getOperand(i: 0).getReg()))
429 FPInstClass = X86II::SpecialFP;
430
431 if (MI.isCall())
432 FPInstClass = X86II::SpecialFP;
433
434 // A fake_use with a floating point pseudo register argument that is
435 // killed must behave like any other floating point operation and pop
436 // the floating point stack (this is done in handleSpecialFP()).
437 // Fake_use is, however, unusual, in that sometimes its operand is not
438 // killed because a later instruction (probably a return) will use it.
439 // It is this instruction that will pop the stack.
440 // In this scenario we can safely remove the fake_use's operand
441 // (it is live anyway).
442 if (MI.isFakeUse()) {
443 const MachineOperand &MO = MI.getOperand(i: 0);
444 if (MO.isReg() && X86::RFP80RegClass.contains(Reg: MO.getReg())) {
445 if (MO.isKill())
446 FPInstClass = X86II::SpecialFP;
447 else
448 MI.removeOperand(OpNo: 0);
449 }
450 }
451
452 if (FPInstClass == X86II::NotFP)
453 continue; // Efficiently ignore non-fp insts!
454
455 MachineInstr *PrevMI = nullptr;
456 if (I != BB.begin())
457 PrevMI = &*std::prev(x: I);
458
459 ++NumFP; // Keep track of # of pseudo instrs
460 LLVM_DEBUG(dbgs() << "\nFPInst:\t" << MI);
461
462 // Get dead variables list now because the MI pointer may be deleted as part
463 // of processing!
464 SmallVector<Register, 8> DeadRegs;
465 for (const MachineOperand &MO : MI.operands())
466 if (MO.isReg() && MO.isDead())
467 DeadRegs.push_back(Elt: MO.getReg());
468
469 switch (FPInstClass) {
470 case X86II::ZeroArgFP: handleZeroArgFP(I); break;
471 case X86II::OneArgFP: handleOneArgFP(I); break; // fstp ST(0)
472 case X86II::OneArgFPRW: handleOneArgFPRW(I); break; // ST(0) = fsqrt(ST(0))
473 case X86II::TwoArgFP: handleTwoArgFP(I); break;
474 case X86II::CompareFP: handleCompareFP(I); break;
475 case X86II::CondMovFP: handleCondMovFP(I); break;
476 case X86II::SpecialFP: handleSpecialFP(I); break;
477 default: llvm_unreachable("Unknown FP Type!");
478 }
479
480 // Check to see if any of the values defined by this instruction are dead
481 // after definition. If so, pop them.
482 for (Register Reg : DeadRegs) {
483 // Check if Reg is live on the stack. An inline-asm register operand that
484 // is in the clobber list and marked dead might not be live on the stack.
485 static_assert(X86::FP7 - X86::FP0 == 7, "sequential FP regnumbers");
486 if (Reg >= X86::FP0 && Reg <= X86::FP6 && isLive(RegNo: Reg-X86::FP0)) {
487 LLVM_DEBUG(dbgs() << "Register FP#" << Reg - X86::FP0 << " is dead!\n");
488 freeStackSlotAfter(I, Reg: Reg-X86::FP0);
489 }
490 }
491
492 // Print out all of the instructions expanded to if -debug
493 LLVM_DEBUG({
494 MachineBasicBlock::iterator PrevI = PrevMI;
495 if (I == PrevI) {
496 dbgs() << "Just deleted pseudo instruction\n";
497 } else {
498 MachineBasicBlock::iterator Start = I;
499 // Rewind to first instruction newly inserted.
500 while (Start != BB.begin() && std::prev(Start) != PrevI)
501 --Start;
502 dbgs() << "Inserted instructions:\n\t";
503 Start->print(dbgs());
504 while (++Start != std::next(I)) {
505 }
506 }
507 dumpStack();
508 });
509 (void)PrevMI;
510
511 Changed = true;
512 }
513
514 finishBlockStack();
515
516 return Changed;
517}
518
519/// setupBlockStack - Use the live bundles to set up our model of the stack
520/// to match predecessors' live out stack.
521void FPS::setupBlockStack() {
522 LLVM_DEBUG(dbgs() << "\nSetting up live-ins for " << printMBBReference(*MBB)
523 << " derived from " << MBB->getName() << ".\n");
524 StackTop = 0;
525 // Get the live-in bundle for MBB.
526 const LiveBundle &Bundle =
527 LiveBundles[Bundles->getBundle(N: MBB->getNumber(), Out: false)];
528
529 if (!Bundle.Mask) {
530 LLVM_DEBUG(dbgs() << "Block has no FP live-ins.\n");
531 return;
532 }
533
534 // Depth-first iteration should ensure that we always have an assigned stack.
535 assert(Bundle.isFixed() && "Reached block before any predecessors");
536
537 // Push the fixed live-in registers.
538 for (unsigned i = Bundle.FixCount; i > 0; --i) {
539 LLVM_DEBUG(dbgs() << "Live-in st(" << (i - 1) << "): %fp"
540 << unsigned(Bundle.FixStack[i - 1]) << '\n');
541 pushReg(Reg: Bundle.FixStack[i-1]);
542 }
543
544 // Kill off unwanted live-ins. This can happen with a critical edge.
545 // FIXME: We could keep these live registers around as zombies. They may need
546 // to be revived at the end of a short block. It might save a few instrs.
547 unsigned Mask = calcLiveInMask(MBB, /*RemoveFPs=*/true);
548 adjustLiveRegs(Mask, I: MBB->begin());
549 LLVM_DEBUG(MBB->dump());
550}
551
552/// finishBlockStack - Revive live-outs that are implicitly defined out of
553/// MBB. Shuffle live registers to match the expected fixed stack of any
554/// predecessors, and ensure that all predecessors are expecting the same
555/// stack.
556void FPS::finishBlockStack() {
557 // The RET handling below takes care of return blocks for us.
558 if (MBB->succ_empty())
559 return;
560
561 LLVM_DEBUG(dbgs() << "Setting up live-outs for " << printMBBReference(*MBB)
562 << " derived from " << MBB->getName() << ".\n");
563
564 // Get MBB's live-out bundle.
565 unsigned BundleIdx = Bundles->getBundle(N: MBB->getNumber(), Out: true);
566 LiveBundle &Bundle = LiveBundles[BundleIdx];
567
568 // We may need to kill and define some registers to match successors.
569 // FIXME: This can probably be combined with the shuffle below.
570 MachineBasicBlock::iterator Term = MBB->getFirstTerminator();
571 adjustLiveRegs(Mask: Bundle.Mask, I: Term);
572
573 if (!Bundle.Mask) {
574 LLVM_DEBUG(dbgs() << "No live-outs.\n");
575 return;
576 }
577
578 // Has the stack order been fixed yet?
579 LLVM_DEBUG(dbgs() << "LB#" << BundleIdx << ": ");
580 if (Bundle.isFixed()) {
581 LLVM_DEBUG(dbgs() << "Shuffling stack to match.\n");
582 shuffleStackTop(FixStack: Bundle.FixStack, FixCount: Bundle.FixCount, I: Term);
583 } else {
584 // Not fixed yet, we get to choose.
585 LLVM_DEBUG(dbgs() << "Fixing stack order now.\n");
586 Bundle.FixCount = StackTop;
587 for (unsigned i = 0; i < StackTop; ++i)
588 Bundle.FixStack[i] = getStackEntry(STi: i);
589 }
590}
591
592
593//===----------------------------------------------------------------------===//
594// Efficient Lookup Table Support
595//===----------------------------------------------------------------------===//
596
597namespace {
598 struct TableEntry {
599 uint16_t from;
600 uint16_t to;
601 bool operator<(const TableEntry &TE) const { return from < TE.from; }
602 friend bool operator<(const TableEntry &TE, unsigned V) {
603 return TE.from < V;
604 }
605 friend bool LLVM_ATTRIBUTE_UNUSED operator<(unsigned V,
606 const TableEntry &TE) {
607 return V < TE.from;
608 }
609 };
610}
611
612static int Lookup(ArrayRef<TableEntry> Table, unsigned Opcode) {
613 const TableEntry *I = llvm::lower_bound(Range&: Table, Value&: Opcode);
614 if (I != Table.end() && I->from == Opcode)
615 return I->to;
616 return -1;
617}
618
619#ifdef NDEBUG
620#define ASSERT_SORTED(TABLE)
621#else
622#define ASSERT_SORTED(TABLE) \
623 { \
624 static std::atomic<bool> TABLE##Checked(false); \
625 if (!TABLE##Checked.load(std::memory_order_relaxed)) { \
626 assert(is_sorted(TABLE) && \
627 "All lookup tables must be sorted for efficient access!"); \
628 TABLE##Checked.store(true, std::memory_order_relaxed); \
629 } \
630 }
631#endif
632
633//===----------------------------------------------------------------------===//
634// Register File -> Register Stack Mapping Methods
635//===----------------------------------------------------------------------===//
636
637// OpcodeTable - Sorted map of register instructions to their stack version.
638// The first element is an register file pseudo instruction, the second is the
639// concrete X86 instruction which uses the register stack.
640//
641static const TableEntry OpcodeTable[] = {
642 { .from: X86::ABS_Fp32 , .to: X86::ABS_F },
643 { .from: X86::ABS_Fp64 , .to: X86::ABS_F },
644 { .from: X86::ABS_Fp80 , .to: X86::ABS_F },
645 { .from: X86::ADD_Fp32m , .to: X86::ADD_F32m },
646 { .from: X86::ADD_Fp64m , .to: X86::ADD_F64m },
647 { .from: X86::ADD_Fp64m32 , .to: X86::ADD_F32m },
648 { .from: X86::ADD_Fp80m32 , .to: X86::ADD_F32m },
649 { .from: X86::ADD_Fp80m64 , .to: X86::ADD_F64m },
650 { .from: X86::ADD_FpI16m32 , .to: X86::ADD_FI16m },
651 { .from: X86::ADD_FpI16m64 , .to: X86::ADD_FI16m },
652 { .from: X86::ADD_FpI16m80 , .to: X86::ADD_FI16m },
653 { .from: X86::ADD_FpI32m32 , .to: X86::ADD_FI32m },
654 { .from: X86::ADD_FpI32m64 , .to: X86::ADD_FI32m },
655 { .from: X86::ADD_FpI32m80 , .to: X86::ADD_FI32m },
656 { .from: X86::CHS_Fp32 , .to: X86::CHS_F },
657 { .from: X86::CHS_Fp64 , .to: X86::CHS_F },
658 { .from: X86::CHS_Fp80 , .to: X86::CHS_F },
659 { .from: X86::CMOVBE_Fp32 , .to: X86::CMOVBE_F },
660 { .from: X86::CMOVBE_Fp64 , .to: X86::CMOVBE_F },
661 { .from: X86::CMOVBE_Fp80 , .to: X86::CMOVBE_F },
662 { .from: X86::CMOVB_Fp32 , .to: X86::CMOVB_F },
663 { .from: X86::CMOVB_Fp64 , .to: X86::CMOVB_F },
664 { .from: X86::CMOVB_Fp80 , .to: X86::CMOVB_F },
665 { .from: X86::CMOVE_Fp32 , .to: X86::CMOVE_F },
666 { .from: X86::CMOVE_Fp64 , .to: X86::CMOVE_F },
667 { .from: X86::CMOVE_Fp80 , .to: X86::CMOVE_F },
668 { .from: X86::CMOVNBE_Fp32 , .to: X86::CMOVNBE_F },
669 { .from: X86::CMOVNBE_Fp64 , .to: X86::CMOVNBE_F },
670 { .from: X86::CMOVNBE_Fp80 , .to: X86::CMOVNBE_F },
671 { .from: X86::CMOVNB_Fp32 , .to: X86::CMOVNB_F },
672 { .from: X86::CMOVNB_Fp64 , .to: X86::CMOVNB_F },
673 { .from: X86::CMOVNB_Fp80 , .to: X86::CMOVNB_F },
674 { .from: X86::CMOVNE_Fp32 , .to: X86::CMOVNE_F },
675 { .from: X86::CMOVNE_Fp64 , .to: X86::CMOVNE_F },
676 { .from: X86::CMOVNE_Fp80 , .to: X86::CMOVNE_F },
677 { .from: X86::CMOVNP_Fp32 , .to: X86::CMOVNP_F },
678 { .from: X86::CMOVNP_Fp64 , .to: X86::CMOVNP_F },
679 { .from: X86::CMOVNP_Fp80 , .to: X86::CMOVNP_F },
680 { .from: X86::CMOVP_Fp32 , .to: X86::CMOVP_F },
681 { .from: X86::CMOVP_Fp64 , .to: X86::CMOVP_F },
682 { .from: X86::CMOVP_Fp80 , .to: X86::CMOVP_F },
683 { .from: X86::COM_FpIr32 , .to: X86::COM_FIr },
684 { .from: X86::COM_FpIr64 , .to: X86::COM_FIr },
685 { .from: X86::COM_FpIr80 , .to: X86::COM_FIr },
686 { .from: X86::COM_Fpr32 , .to: X86::COM_FST0r },
687 { .from: X86::COM_Fpr64 , .to: X86::COM_FST0r },
688 { .from: X86::COM_Fpr80 , .to: X86::COM_FST0r },
689 { .from: X86::DIVR_Fp32m , .to: X86::DIVR_F32m },
690 { .from: X86::DIVR_Fp64m , .to: X86::DIVR_F64m },
691 { .from: X86::DIVR_Fp64m32 , .to: X86::DIVR_F32m },
692 { .from: X86::DIVR_Fp80m32 , .to: X86::DIVR_F32m },
693 { .from: X86::DIVR_Fp80m64 , .to: X86::DIVR_F64m },
694 { .from: X86::DIVR_FpI16m32, .to: X86::DIVR_FI16m},
695 { .from: X86::DIVR_FpI16m64, .to: X86::DIVR_FI16m},
696 { .from: X86::DIVR_FpI16m80, .to: X86::DIVR_FI16m},
697 { .from: X86::DIVR_FpI32m32, .to: X86::DIVR_FI32m},
698 { .from: X86::DIVR_FpI32m64, .to: X86::DIVR_FI32m},
699 { .from: X86::DIVR_FpI32m80, .to: X86::DIVR_FI32m},
700 { .from: X86::DIV_Fp32m , .to: X86::DIV_F32m },
701 { .from: X86::DIV_Fp64m , .to: X86::DIV_F64m },
702 { .from: X86::DIV_Fp64m32 , .to: X86::DIV_F32m },
703 { .from: X86::DIV_Fp80m32 , .to: X86::DIV_F32m },
704 { .from: X86::DIV_Fp80m64 , .to: X86::DIV_F64m },
705 { .from: X86::DIV_FpI16m32 , .to: X86::DIV_FI16m },
706 { .from: X86::DIV_FpI16m64 , .to: X86::DIV_FI16m },
707 { .from: X86::DIV_FpI16m80 , .to: X86::DIV_FI16m },
708 { .from: X86::DIV_FpI32m32 , .to: X86::DIV_FI32m },
709 { .from: X86::DIV_FpI32m64 , .to: X86::DIV_FI32m },
710 { .from: X86::DIV_FpI32m80 , .to: X86::DIV_FI32m },
711 { .from: X86::ILD_Fp16m32 , .to: X86::ILD_F16m },
712 { .from: X86::ILD_Fp16m64 , .to: X86::ILD_F16m },
713 { .from: X86::ILD_Fp16m80 , .to: X86::ILD_F16m },
714 { .from: X86::ILD_Fp32m32 , .to: X86::ILD_F32m },
715 { .from: X86::ILD_Fp32m64 , .to: X86::ILD_F32m },
716 { .from: X86::ILD_Fp32m80 , .to: X86::ILD_F32m },
717 { .from: X86::ILD_Fp64m32 , .to: X86::ILD_F64m },
718 { .from: X86::ILD_Fp64m64 , .to: X86::ILD_F64m },
719 { .from: X86::ILD_Fp64m80 , .to: X86::ILD_F64m },
720 { .from: X86::ISTT_Fp16m32 , .to: X86::ISTT_FP16m},
721 { .from: X86::ISTT_Fp16m64 , .to: X86::ISTT_FP16m},
722 { .from: X86::ISTT_Fp16m80 , .to: X86::ISTT_FP16m},
723 { .from: X86::ISTT_Fp32m32 , .to: X86::ISTT_FP32m},
724 { .from: X86::ISTT_Fp32m64 , .to: X86::ISTT_FP32m},
725 { .from: X86::ISTT_Fp32m80 , .to: X86::ISTT_FP32m},
726 { .from: X86::ISTT_Fp64m32 , .to: X86::ISTT_FP64m},
727 { .from: X86::ISTT_Fp64m64 , .to: X86::ISTT_FP64m},
728 { .from: X86::ISTT_Fp64m80 , .to: X86::ISTT_FP64m},
729 { .from: X86::IST_Fp16m32 , .to: X86::IST_F16m },
730 { .from: X86::IST_Fp16m64 , .to: X86::IST_F16m },
731 { .from: X86::IST_Fp16m80 , .to: X86::IST_F16m },
732 { .from: X86::IST_Fp32m32 , .to: X86::IST_F32m },
733 { .from: X86::IST_Fp32m64 , .to: X86::IST_F32m },
734 { .from: X86::IST_Fp32m80 , .to: X86::IST_F32m },
735 { .from: X86::IST_Fp64m32 , .to: X86::IST_FP64m },
736 { .from: X86::IST_Fp64m64 , .to: X86::IST_FP64m },
737 { .from: X86::IST_Fp64m80 , .to: X86::IST_FP64m },
738 { .from: X86::LD_Fp032 , .to: X86::LD_F0 },
739 { .from: X86::LD_Fp064 , .to: X86::LD_F0 },
740 { .from: X86::LD_Fp080 , .to: X86::LD_F0 },
741 { .from: X86::LD_Fp132 , .to: X86::LD_F1 },
742 { .from: X86::LD_Fp164 , .to: X86::LD_F1 },
743 { .from: X86::LD_Fp180 , .to: X86::LD_F1 },
744 { .from: X86::LD_Fp32m , .to: X86::LD_F32m },
745 { .from: X86::LD_Fp32m64 , .to: X86::LD_F32m },
746 { .from: X86::LD_Fp32m80 , .to: X86::LD_F32m },
747 { .from: X86::LD_Fp64m , .to: X86::LD_F64m },
748 { .from: X86::LD_Fp64m80 , .to: X86::LD_F64m },
749 { .from: X86::LD_Fp80m , .to: X86::LD_F80m },
750 { .from: X86::MUL_Fp32m , .to: X86::MUL_F32m },
751 { .from: X86::MUL_Fp64m , .to: X86::MUL_F64m },
752 { .from: X86::MUL_Fp64m32 , .to: X86::MUL_F32m },
753 { .from: X86::MUL_Fp80m32 , .to: X86::MUL_F32m },
754 { .from: X86::MUL_Fp80m64 , .to: X86::MUL_F64m },
755 { .from: X86::MUL_FpI16m32 , .to: X86::MUL_FI16m },
756 { .from: X86::MUL_FpI16m64 , .to: X86::MUL_FI16m },
757 { .from: X86::MUL_FpI16m80 , .to: X86::MUL_FI16m },
758 { .from: X86::MUL_FpI32m32 , .to: X86::MUL_FI32m },
759 { .from: X86::MUL_FpI32m64 , .to: X86::MUL_FI32m },
760 { .from: X86::MUL_FpI32m80 , .to: X86::MUL_FI32m },
761 { .from: X86::SQRT_Fp32 , .to: X86::SQRT_F },
762 { .from: X86::SQRT_Fp64 , .to: X86::SQRT_F },
763 { .from: X86::SQRT_Fp80 , .to: X86::SQRT_F },
764 { .from: X86::ST_Fp32m , .to: X86::ST_F32m },
765 { .from: X86::ST_Fp64m , .to: X86::ST_F64m },
766 { .from: X86::ST_Fp64m32 , .to: X86::ST_F32m },
767 { .from: X86::ST_Fp80m32 , .to: X86::ST_F32m },
768 { .from: X86::ST_Fp80m64 , .to: X86::ST_F64m },
769 { .from: X86::ST_FpP80m , .to: X86::ST_FP80m },
770 { .from: X86::SUBR_Fp32m , .to: X86::SUBR_F32m },
771 { .from: X86::SUBR_Fp64m , .to: X86::SUBR_F64m },
772 { .from: X86::SUBR_Fp64m32 , .to: X86::SUBR_F32m },
773 { .from: X86::SUBR_Fp80m32 , .to: X86::SUBR_F32m },
774 { .from: X86::SUBR_Fp80m64 , .to: X86::SUBR_F64m },
775 { .from: X86::SUBR_FpI16m32, .to: X86::SUBR_FI16m},
776 { .from: X86::SUBR_FpI16m64, .to: X86::SUBR_FI16m},
777 { .from: X86::SUBR_FpI16m80, .to: X86::SUBR_FI16m},
778 { .from: X86::SUBR_FpI32m32, .to: X86::SUBR_FI32m},
779 { .from: X86::SUBR_FpI32m64, .to: X86::SUBR_FI32m},
780 { .from: X86::SUBR_FpI32m80, .to: X86::SUBR_FI32m},
781 { .from: X86::SUB_Fp32m , .to: X86::SUB_F32m },
782 { .from: X86::SUB_Fp64m , .to: X86::SUB_F64m },
783 { .from: X86::SUB_Fp64m32 , .to: X86::SUB_F32m },
784 { .from: X86::SUB_Fp80m32 , .to: X86::SUB_F32m },
785 { .from: X86::SUB_Fp80m64 , .to: X86::SUB_F64m },
786 { .from: X86::SUB_FpI16m32 , .to: X86::SUB_FI16m },
787 { .from: X86::SUB_FpI16m64 , .to: X86::SUB_FI16m },
788 { .from: X86::SUB_FpI16m80 , .to: X86::SUB_FI16m },
789 { .from: X86::SUB_FpI32m32 , .to: X86::SUB_FI32m },
790 { .from: X86::SUB_FpI32m64 , .to: X86::SUB_FI32m },
791 { .from: X86::SUB_FpI32m80 , .to: X86::SUB_FI32m },
792 { .from: X86::TST_Fp32 , .to: X86::TST_F },
793 { .from: X86::TST_Fp64 , .to: X86::TST_F },
794 { .from: X86::TST_Fp80 , .to: X86::TST_F },
795 { .from: X86::UCOM_FpIr32 , .to: X86::UCOM_FIr },
796 { .from: X86::UCOM_FpIr64 , .to: X86::UCOM_FIr },
797 { .from: X86::UCOM_FpIr80 , .to: X86::UCOM_FIr },
798 { .from: X86::UCOM_Fpr32 , .to: X86::UCOM_Fr },
799 { .from: X86::UCOM_Fpr64 , .to: X86::UCOM_Fr },
800 { .from: X86::UCOM_Fpr80 , .to: X86::UCOM_Fr },
801 { .from: X86::XAM_Fp32 , .to: X86::XAM_F },
802 { .from: X86::XAM_Fp64 , .to: X86::XAM_F },
803 { .from: X86::XAM_Fp80 , .to: X86::XAM_F },
804};
805
806static unsigned getConcreteOpcode(unsigned Opcode) {
807 ASSERT_SORTED(OpcodeTable);
808 int Opc = Lookup(Table: OpcodeTable, Opcode);
809 assert(Opc != -1 && "FP Stack instruction not in OpcodeTable!");
810 return Opc;
811}
812
813//===----------------------------------------------------------------------===//
814// Helper Methods
815//===----------------------------------------------------------------------===//
816
817// PopTable - Sorted map of instructions to their popping version. The first
818// element is an instruction, the second is the version which pops.
819//
820static const TableEntry PopTable[] = {
821 { .from: X86::ADD_FrST0 , .to: X86::ADD_FPrST0 },
822
823 { .from: X86::COMP_FST0r, .to: X86::FCOMPP },
824 { .from: X86::COM_FIr , .to: X86::COM_FIPr },
825 { .from: X86::COM_FST0r , .to: X86::COMP_FST0r },
826
827 { .from: X86::DIVR_FrST0, .to: X86::DIVR_FPrST0 },
828 { .from: X86::DIV_FrST0 , .to: X86::DIV_FPrST0 },
829
830 { .from: X86::IST_F16m , .to: X86::IST_FP16m },
831 { .from: X86::IST_F32m , .to: X86::IST_FP32m },
832
833 { .from: X86::MUL_FrST0 , .to: X86::MUL_FPrST0 },
834
835 { .from: X86::ST_F32m , .to: X86::ST_FP32m },
836 { .from: X86::ST_F64m , .to: X86::ST_FP64m },
837 { .from: X86::ST_Frr , .to: X86::ST_FPrr },
838
839 { .from: X86::SUBR_FrST0, .to: X86::SUBR_FPrST0 },
840 { .from: X86::SUB_FrST0 , .to: X86::SUB_FPrST0 },
841
842 { .from: X86::UCOM_FIr , .to: X86::UCOM_FIPr },
843
844 { .from: X86::UCOM_FPr , .to: X86::UCOM_FPPr },
845 { .from: X86::UCOM_Fr , .to: X86::UCOM_FPr },
846};
847
848static bool doesInstructionSetFPSW(MachineInstr &MI) {
849 if (const MachineOperand *MO =
850 MI.findRegisterDefOperand(Reg: X86::FPSW, /*TRI=*/nullptr))
851 if (!MO->isDead())
852 return true;
853 return false;
854}
855
856static MachineBasicBlock::iterator
857getNextFPInstruction(MachineBasicBlock::iterator I) {
858 MachineBasicBlock &MBB = *I->getParent();
859 while (++I != MBB.end()) {
860 MachineInstr &MI = *I;
861 if (X86::isX87Instruction(MI))
862 return I;
863 }
864 return MBB.end();
865}
866
867/// popStackAfter - Pop the current value off of the top of the FP stack after
868/// the specified instruction. This attempts to be sneaky and combine the pop
869/// into the instruction itself if possible. The iterator is left pointing to
870/// the last instruction, be it a new pop instruction inserted, or the old
871/// instruction if it was modified in place.
872///
873void FPS::popStackAfter(MachineBasicBlock::iterator &I) {
874 MachineInstr &MI = *I;
875 const DebugLoc &dl = MI.getDebugLoc();
876 ASSERT_SORTED(PopTable);
877
878 popReg();
879
880 // Check to see if there is a popping version of this instruction...
881 int Opcode = Lookup(Table: PopTable, Opcode: I->getOpcode());
882 if (Opcode != -1) {
883 I->setDesc(TII->get(Opcode));
884 if (Opcode == X86::FCOMPP || Opcode == X86::UCOM_FPPr)
885 I->removeOperand(OpNo: 0);
886 MI.dropDebugNumber();
887 } else { // Insert an explicit pop
888 // If this instruction sets FPSW, which is read in following instruction,
889 // insert pop after that reader.
890 if (doesInstructionSetFPSW(MI)) {
891 MachineBasicBlock &MBB = *MI.getParent();
892 MachineBasicBlock::iterator Next = getNextFPInstruction(I);
893 if (Next != MBB.end() && Next->readsRegister(Reg: X86::FPSW, /*TRI=*/nullptr))
894 I = Next;
895 }
896 I = BuildMI(BB&: *MBB, I: ++I, MIMD: dl, MCID: TII->get(Opcode: X86::ST_FPrr)).addReg(RegNo: X86::ST0);
897 }
898}
899
900/// freeStackSlotAfter - Free the specified register from the register stack, so
901/// that it is no longer in a register. If the register is currently at the top
902/// of the stack, we just pop the current instruction, otherwise we store the
903/// current top-of-stack into the specified slot, then pop the top of stack.
904void FPS::freeStackSlotAfter(MachineBasicBlock::iterator &I, unsigned FPRegNo) {
905 if (getStackEntry(STi: 0) == FPRegNo) { // already at the top of stack? easy.
906 popStackAfter(I);
907 return;
908 }
909
910 // Otherwise, store the top of stack into the dead slot, killing the operand
911 // without having to add in an explicit xchg then pop.
912 //
913 I = freeStackSlotBefore(I: ++I, FPRegNo);
914}
915
916/// freeStackSlotBefore - Free the specified register without trying any
917/// folding.
918MachineBasicBlock::iterator
919FPS::freeStackSlotBefore(MachineBasicBlock::iterator I, unsigned FPRegNo) {
920 unsigned STReg = getSTReg(RegNo: FPRegNo);
921 unsigned OldSlot = getSlot(RegNo: FPRegNo);
922 unsigned TopReg = Stack[StackTop-1];
923 Stack[OldSlot] = TopReg;
924 RegMap[TopReg] = OldSlot;
925 RegMap[FPRegNo] = ~0;
926 Stack[--StackTop] = ~0;
927 return BuildMI(BB&: *MBB, I, MIMD: DebugLoc(), MCID: TII->get(Opcode: X86::ST_FPrr))
928 .addReg(RegNo: STReg)
929 .getInstr();
930}
931
932/// adjustLiveRegs - Kill and revive registers such that exactly the FP
933/// registers with a bit in Mask are live.
934void FPS::adjustLiveRegs(unsigned Mask, MachineBasicBlock::iterator I) {
935 unsigned Defs = Mask;
936 unsigned Kills = 0;
937 for (unsigned i = 0; i < StackTop; ++i) {
938 unsigned RegNo = Stack[i];
939 if (!(Defs & (1 << RegNo)))
940 // This register is live, but we don't want it.
941 Kills |= (1 << RegNo);
942 else
943 // We don't need to imp-def this live register.
944 Defs &= ~(1 << RegNo);
945 }
946 assert((Kills & Defs) == 0 && "Register needs killing and def'ing?");
947
948 // Produce implicit-defs for free by using killed registers.
949 while (Kills && Defs) {
950 unsigned KReg = llvm::countr_zero(Val: Kills);
951 unsigned DReg = llvm::countr_zero(Val: Defs);
952 LLVM_DEBUG(dbgs() << "Renaming %fp" << KReg << " as imp %fp" << DReg
953 << "\n");
954 std::swap(a&: Stack[getSlot(RegNo: KReg)], b&: Stack[getSlot(RegNo: DReg)]);
955 std::swap(a&: RegMap[KReg], b&: RegMap[DReg]);
956 Kills &= ~(1 << KReg);
957 Defs &= ~(1 << DReg);
958 }
959
960 // Kill registers by popping.
961 if (Kills && I != MBB->begin()) {
962 MachineBasicBlock::iterator I2 = std::prev(x: I);
963 while (StackTop) {
964 unsigned KReg = getStackEntry(STi: 0);
965 if (!(Kills & (1 << KReg)))
966 break;
967 LLVM_DEBUG(dbgs() << "Popping %fp" << KReg << "\n");
968 popStackAfter(I&: I2);
969 Kills &= ~(1 << KReg);
970 }
971 }
972
973 // Manually kill the rest.
974 while (Kills) {
975 unsigned KReg = llvm::countr_zero(Val: Kills);
976 LLVM_DEBUG(dbgs() << "Killing %fp" << KReg << "\n");
977 freeStackSlotBefore(I, FPRegNo: KReg);
978 Kills &= ~(1 << KReg);
979 }
980
981 // Load zeros for all the imp-defs.
982 while(Defs) {
983 unsigned DReg = llvm::countr_zero(Val: Defs);
984 LLVM_DEBUG(dbgs() << "Defining %fp" << DReg << " as 0\n");
985 BuildMI(BB&: *MBB, I, MIMD: DebugLoc(), MCID: TII->get(Opcode: X86::LD_F0));
986 pushReg(Reg: DReg);
987 Defs &= ~(1 << DReg);
988 }
989
990 // Now we should have the correct registers live.
991 LLVM_DEBUG(dumpStack());
992 assert(StackTop == (unsigned)llvm::popcount(Mask) && "Live count mismatch");
993}
994
995/// shuffleStackTop - emit fxch instructions before I to shuffle the top
996/// FixCount entries into the order given by FixStack.
997/// FIXME: Is there a better algorithm than insertion sort?
998void FPS::shuffleStackTop(const unsigned char *FixStack,
999 unsigned FixCount,
1000 MachineBasicBlock::iterator I) {
1001 // Move items into place, starting from the desired stack bottom.
1002 while (FixCount--) {
1003 // Old register at position FixCount.
1004 unsigned OldReg = getStackEntry(STi: FixCount);
1005 // Desired register at position FixCount.
1006 unsigned Reg = FixStack[FixCount];
1007 if (Reg == OldReg)
1008 continue;
1009 // (Reg st0) (OldReg st0) = (Reg OldReg st0)
1010 moveToTop(RegNo: Reg, I);
1011 if (FixCount > 0)
1012 moveToTop(RegNo: OldReg, I);
1013 }
1014 LLVM_DEBUG(dumpStack());
1015}
1016
1017
1018//===----------------------------------------------------------------------===//
1019// Instruction transformation implementation
1020//===----------------------------------------------------------------------===//
1021
1022void FPS::handleCall(MachineBasicBlock::iterator &I) {
1023 MachineInstr &MI = *I;
1024 unsigned STReturns = 0;
1025
1026 bool ClobbersFPStack = false;
1027 for (unsigned i = 0, e = MI.getNumOperands(); i != e; ++i) {
1028 MachineOperand &Op = MI.getOperand(i);
1029 // Check if this call clobbers the FP stack.
1030 // is sufficient.
1031 if (Op.isRegMask()) {
1032 bool ClobbersFP0 = Op.clobbersPhysReg(PhysReg: X86::FP0);
1033#ifndef NDEBUG
1034 static_assert(X86::FP7 - X86::FP0 == 7, "sequential FP regnumbers");
1035 for (unsigned i = 1; i != 8; ++i)
1036 assert(Op.clobbersPhysReg(X86::FP0 + i) == ClobbersFP0 &&
1037 "Inconsistent FP register clobber");
1038#endif
1039
1040 if (ClobbersFP0)
1041 ClobbersFPStack = true;
1042 }
1043
1044 if (!Op.isReg() || Op.getReg() < X86::FP0 || Op.getReg() > X86::FP6)
1045 continue;
1046
1047 assert(Op.isImplicit() && "Expected implicit def/use");
1048
1049 if (Op.isDef())
1050 STReturns |= 1 << getFPReg(MO: Op);
1051
1052 // Remove the operand so that later passes don't see it.
1053 MI.removeOperand(OpNo: i);
1054 --i;
1055 --e;
1056 }
1057
1058 // Most calls should have a regmask that clobbers the FP registers. If it
1059 // isn't present then the register allocator didn't spill the FP registers
1060 // so they are still on the stack.
1061 assert((ClobbersFPStack || STReturns == 0) &&
1062 "ST returns without FP stack clobber");
1063 if (!ClobbersFPStack)
1064 return;
1065
1066 unsigned N = llvm::countr_one(Value: STReturns);
1067
1068 // FP registers used for function return must be consecutive starting at
1069 // FP0
1070 assert(STReturns == 0 || (isMask_32(STReturns) && N <= 2));
1071
1072 // Reset the FP Stack - It is required because of possible leftovers from
1073 // passed arguments. The caller should assume that the FP stack is
1074 // returned empty (unless the callee returns values on FP stack).
1075 while (StackTop > 0)
1076 popReg();
1077
1078 for (unsigned I = 0; I < N; ++I)
1079 pushReg(Reg: N - I - 1);
1080
1081 // If this call has been modified, drop all variable values defined by it.
1082 // We can't track them once they've been stackified.
1083 if (STReturns)
1084 I->dropDebugNumber();
1085}
1086
1087/// If RET has an FP register use operand, pass the first one in ST(0) and
1088/// the second one in ST(1).
1089void FPS::handleReturn(MachineBasicBlock::iterator &I) {
1090 MachineInstr &MI = *I;
1091
1092 // Find the register operands.
1093 unsigned FirstFPRegOp = ~0U, SecondFPRegOp = ~0U;
1094 unsigned LiveMask = 0;
1095
1096 for (unsigned i = 0, e = MI.getNumOperands(); i != e; ++i) {
1097 MachineOperand &Op = MI.getOperand(i);
1098 if (!Op.isReg() || Op.getReg() < X86::FP0 || Op.getReg() > X86::FP6)
1099 continue;
1100 // FP Register uses must be kills unless there are two uses of the same
1101 // register, in which case only one will be a kill.
1102 assert(Op.isUse() &&
1103 (Op.isKill() || // Marked kill.
1104 getFPReg(Op) == FirstFPRegOp || // Second instance.
1105 MI.killsRegister(Op.getReg(),
1106 /*TRI=*/nullptr)) && // Later use is marked kill.
1107 "Ret only defs operands, and values aren't live beyond it");
1108
1109 if (FirstFPRegOp == ~0U)
1110 FirstFPRegOp = getFPReg(MO: Op);
1111 else {
1112 assert(SecondFPRegOp == ~0U && "More than two fp operands!");
1113 SecondFPRegOp = getFPReg(MO: Op);
1114 }
1115 LiveMask |= (1 << getFPReg(MO: Op));
1116
1117 // Remove the operand so that later passes don't see it.
1118 MI.removeOperand(OpNo: i);
1119 --i;
1120 --e;
1121 }
1122
1123 // We may have been carrying spurious live-ins, so make sure only the
1124 // returned registers are left live.
1125 adjustLiveRegs(Mask: LiveMask, I: MI);
1126 if (!LiveMask) return; // Quick check to see if any are possible.
1127
1128 // There are only four possibilities here:
1129 // 1) we are returning a single FP value. In this case, it has to be in
1130 // ST(0) already, so just declare success by removing the value from the
1131 // FP Stack.
1132 if (SecondFPRegOp == ~0U) {
1133 // Assert that the top of stack contains the right FP register.
1134 assert(StackTop == 1 && FirstFPRegOp == getStackEntry(0) &&
1135 "Top of stack not the right register for RET!");
1136
1137 // Ok, everything is good, mark the value as not being on the stack
1138 // anymore so that our assertion about the stack being empty at end of
1139 // block doesn't fire.
1140 StackTop = 0;
1141 return;
1142 }
1143
1144 // Otherwise, we are returning two values:
1145 // 2) If returning the same value for both, we only have one thing in the FP
1146 // stack. Consider: RET FP1, FP1
1147 if (StackTop == 1) {
1148 assert(FirstFPRegOp == SecondFPRegOp && FirstFPRegOp == getStackEntry(0)&&
1149 "Stack misconfiguration for RET!");
1150
1151 // Duplicate the TOS so that we return it twice. Just pick some other FPx
1152 // register to hold it.
1153 unsigned NewReg = ScratchFPReg;
1154 duplicateToTop(RegNo: FirstFPRegOp, AsReg: NewReg, I: MI);
1155 FirstFPRegOp = NewReg;
1156 }
1157
1158 /// Okay we know we have two different FPx operands now:
1159 assert(StackTop == 2 && "Must have two values live!");
1160
1161 /// 3) If SecondFPRegOp is currently in ST(0) and FirstFPRegOp is currently
1162 /// in ST(1). In this case, emit an fxch.
1163 if (getStackEntry(STi: 0) == SecondFPRegOp) {
1164 assert(getStackEntry(1) == FirstFPRegOp && "Unknown regs live");
1165 moveToTop(RegNo: FirstFPRegOp, I: MI);
1166 }
1167
1168 /// 4) Finally, FirstFPRegOp must be in ST(0) and SecondFPRegOp must be in
1169 /// ST(1). Just remove both from our understanding of the stack and return.
1170 assert(getStackEntry(0) == FirstFPRegOp && "Unknown regs live");
1171 assert(getStackEntry(1) == SecondFPRegOp && "Unknown regs live");
1172 StackTop = 0;
1173}
1174
1175/// handleZeroArgFP - ST(0) = fld0 ST(0) = flds <mem>
1176///
1177void FPS::handleZeroArgFP(MachineBasicBlock::iterator &I) {
1178 MachineInstr &MI = *I;
1179 unsigned DestReg = getFPReg(MO: MI.getOperand(i: 0));
1180
1181 // Change from the pseudo instruction to the concrete instruction.
1182 MI.removeOperand(OpNo: 0); // Remove the explicit ST(0) operand
1183 MI.setDesc(TII->get(Opcode: getConcreteOpcode(Opcode: MI.getOpcode())));
1184 MI.addOperand(
1185 Op: MachineOperand::CreateReg(Reg: X86::ST0, /*isDef*/ true, /*isImp*/ true));
1186
1187 // Result gets pushed on the stack.
1188 pushReg(Reg: DestReg);
1189
1190 MI.dropDebugNumber();
1191}
1192
1193/// handleOneArgFP - fst <mem>, ST(0)
1194///
1195void FPS::handleOneArgFP(MachineBasicBlock::iterator &I) {
1196 MachineInstr &MI = *I;
1197 unsigned NumOps = MI.getDesc().getNumOperands();
1198 assert((NumOps == X86::AddrNumOperands + 1 || NumOps == 1) &&
1199 "Can only handle fst* & ftst instructions!");
1200
1201 // Is this the last use of the source register?
1202 unsigned Reg = getFPReg(MO: MI.getOperand(i: NumOps - 1));
1203 bool KillsSrc = MI.killsRegister(Reg: X86::FP0 + Reg, /*TRI=*/nullptr);
1204
1205 // FISTP64m is strange because there isn't a non-popping versions.
1206 // If we have one _and_ we don't want to pop the operand, duplicate the value
1207 // on the stack instead of moving it. This ensure that popping the value is
1208 // always ok.
1209 // Ditto FISTTP16m, FISTTP32m, FISTTP64m, ST_FpP80m.
1210 //
1211 if (!KillsSrc && (MI.getOpcode() == X86::IST_Fp64m32 ||
1212 MI.getOpcode() == X86::ISTT_Fp16m32 ||
1213 MI.getOpcode() == X86::ISTT_Fp32m32 ||
1214 MI.getOpcode() == X86::ISTT_Fp64m32 ||
1215 MI.getOpcode() == X86::IST_Fp64m64 ||
1216 MI.getOpcode() == X86::ISTT_Fp16m64 ||
1217 MI.getOpcode() == X86::ISTT_Fp32m64 ||
1218 MI.getOpcode() == X86::ISTT_Fp64m64 ||
1219 MI.getOpcode() == X86::IST_Fp64m80 ||
1220 MI.getOpcode() == X86::ISTT_Fp16m80 ||
1221 MI.getOpcode() == X86::ISTT_Fp32m80 ||
1222 MI.getOpcode() == X86::ISTT_Fp64m80 ||
1223 MI.getOpcode() == X86::ST_FpP80m)) {
1224 duplicateToTop(RegNo: Reg, AsReg: ScratchFPReg, I);
1225 } else {
1226 moveToTop(RegNo: Reg, I); // Move to the top of the stack...
1227 }
1228
1229 // Convert from the pseudo instruction to the concrete instruction.
1230 MI.removeOperand(OpNo: NumOps - 1); // Remove explicit ST(0) operand
1231 MI.setDesc(TII->get(Opcode: getConcreteOpcode(Opcode: MI.getOpcode())));
1232 MI.addOperand(
1233 Op: MachineOperand::CreateReg(Reg: X86::ST0, /*isDef*/ false, /*isImp*/ true));
1234
1235 if (MI.getOpcode() == X86::IST_FP64m || MI.getOpcode() == X86::ISTT_FP16m ||
1236 MI.getOpcode() == X86::ISTT_FP32m || MI.getOpcode() == X86::ISTT_FP64m ||
1237 MI.getOpcode() == X86::ST_FP80m) {
1238 if (StackTop == 0)
1239 report_fatal_error(reason: "Stack empty??");
1240 --StackTop;
1241 } else if (KillsSrc) { // Last use of operand?
1242 popStackAfter(I);
1243 }
1244
1245 MI.dropDebugNumber();
1246}
1247
1248
1249/// handleOneArgFPRW: Handle instructions that read from the top of stack and
1250/// replace the value with a newly computed value. These instructions may have
1251/// non-fp operands after their FP operands.
1252///
1253/// Examples:
1254/// R1 = fchs R2
1255/// R1 = fadd R2, [mem]
1256///
1257void FPS::handleOneArgFPRW(MachineBasicBlock::iterator &I) {
1258 MachineInstr &MI = *I;
1259#ifndef NDEBUG
1260 unsigned NumOps = MI.getDesc().getNumOperands();
1261 assert(NumOps >= 2 && "FPRW instructions must have 2 ops!!");
1262#endif
1263
1264 // Is this the last use of the source register?
1265 unsigned Reg = getFPReg(MO: MI.getOperand(i: 1));
1266 bool KillsSrc = MI.killsRegister(Reg: X86::FP0 + Reg, /*TRI=*/nullptr);
1267
1268 if (KillsSrc) {
1269 // If this is the last use of the source register, just make sure it's on
1270 // the top of the stack.
1271 moveToTop(RegNo: Reg, I);
1272 if (StackTop == 0)
1273 report_fatal_error(reason: "Stack cannot be empty!");
1274 --StackTop;
1275 pushReg(Reg: getFPReg(MO: MI.getOperand(i: 0)));
1276 } else {
1277 // If this is not the last use of the source register, _copy_ it to the top
1278 // of the stack.
1279 duplicateToTop(RegNo: Reg, AsReg: getFPReg(MO: MI.getOperand(i: 0)), I);
1280 }
1281
1282 // Change from the pseudo instruction to the concrete instruction.
1283 MI.removeOperand(OpNo: 1); // Drop the source operand.
1284 MI.removeOperand(OpNo: 0); // Drop the destination operand.
1285 MI.setDesc(TII->get(Opcode: getConcreteOpcode(Opcode: MI.getOpcode())));
1286 MI.dropDebugNumber();
1287}
1288
1289
1290//===----------------------------------------------------------------------===//
1291// Define tables of various ways to map pseudo instructions
1292//
1293
1294// ForwardST0Table - Map: A = B op C into: ST(0) = ST(0) op ST(i)
1295static const TableEntry ForwardST0Table[] = {
1296 { .from: X86::ADD_Fp32 , .to: X86::ADD_FST0r },
1297 { .from: X86::ADD_Fp64 , .to: X86::ADD_FST0r },
1298 { .from: X86::ADD_Fp80 , .to: X86::ADD_FST0r },
1299 { .from: X86::DIV_Fp32 , .to: X86::DIV_FST0r },
1300 { .from: X86::DIV_Fp64 , .to: X86::DIV_FST0r },
1301 { .from: X86::DIV_Fp80 , .to: X86::DIV_FST0r },
1302 { .from: X86::MUL_Fp32 , .to: X86::MUL_FST0r },
1303 { .from: X86::MUL_Fp64 , .to: X86::MUL_FST0r },
1304 { .from: X86::MUL_Fp80 , .to: X86::MUL_FST0r },
1305 { .from: X86::SUB_Fp32 , .to: X86::SUB_FST0r },
1306 { .from: X86::SUB_Fp64 , .to: X86::SUB_FST0r },
1307 { .from: X86::SUB_Fp80 , .to: X86::SUB_FST0r },
1308};
1309
1310// ReverseST0Table - Map: A = B op C into: ST(0) = ST(i) op ST(0)
1311static const TableEntry ReverseST0Table[] = {
1312 { .from: X86::ADD_Fp32 , .to: X86::ADD_FST0r }, // commutative
1313 { .from: X86::ADD_Fp64 , .to: X86::ADD_FST0r }, // commutative
1314 { .from: X86::ADD_Fp80 , .to: X86::ADD_FST0r }, // commutative
1315 { .from: X86::DIV_Fp32 , .to: X86::DIVR_FST0r },
1316 { .from: X86::DIV_Fp64 , .to: X86::DIVR_FST0r },
1317 { .from: X86::DIV_Fp80 , .to: X86::DIVR_FST0r },
1318 { .from: X86::MUL_Fp32 , .to: X86::MUL_FST0r }, // commutative
1319 { .from: X86::MUL_Fp64 , .to: X86::MUL_FST0r }, // commutative
1320 { .from: X86::MUL_Fp80 , .to: X86::MUL_FST0r }, // commutative
1321 { .from: X86::SUB_Fp32 , .to: X86::SUBR_FST0r },
1322 { .from: X86::SUB_Fp64 , .to: X86::SUBR_FST0r },
1323 { .from: X86::SUB_Fp80 , .to: X86::SUBR_FST0r },
1324};
1325
1326// ForwardSTiTable - Map: A = B op C into: ST(i) = ST(0) op ST(i)
1327static const TableEntry ForwardSTiTable[] = {
1328 { .from: X86::ADD_Fp32 , .to: X86::ADD_FrST0 }, // commutative
1329 { .from: X86::ADD_Fp64 , .to: X86::ADD_FrST0 }, // commutative
1330 { .from: X86::ADD_Fp80 , .to: X86::ADD_FrST0 }, // commutative
1331 { .from: X86::DIV_Fp32 , .to: X86::DIVR_FrST0 },
1332 { .from: X86::DIV_Fp64 , .to: X86::DIVR_FrST0 },
1333 { .from: X86::DIV_Fp80 , .to: X86::DIVR_FrST0 },
1334 { .from: X86::MUL_Fp32 , .to: X86::MUL_FrST0 }, // commutative
1335 { .from: X86::MUL_Fp64 , .to: X86::MUL_FrST0 }, // commutative
1336 { .from: X86::MUL_Fp80 , .to: X86::MUL_FrST0 }, // commutative
1337 { .from: X86::SUB_Fp32 , .to: X86::SUBR_FrST0 },
1338 { .from: X86::SUB_Fp64 , .to: X86::SUBR_FrST0 },
1339 { .from: X86::SUB_Fp80 , .to: X86::SUBR_FrST0 },
1340};
1341
1342// ReverseSTiTable - Map: A = B op C into: ST(i) = ST(i) op ST(0)
1343static const TableEntry ReverseSTiTable[] = {
1344 { .from: X86::ADD_Fp32 , .to: X86::ADD_FrST0 },
1345 { .from: X86::ADD_Fp64 , .to: X86::ADD_FrST0 },
1346 { .from: X86::ADD_Fp80 , .to: X86::ADD_FrST0 },
1347 { .from: X86::DIV_Fp32 , .to: X86::DIV_FrST0 },
1348 { .from: X86::DIV_Fp64 , .to: X86::DIV_FrST0 },
1349 { .from: X86::DIV_Fp80 , .to: X86::DIV_FrST0 },
1350 { .from: X86::MUL_Fp32 , .to: X86::MUL_FrST0 },
1351 { .from: X86::MUL_Fp64 , .to: X86::MUL_FrST0 },
1352 { .from: X86::MUL_Fp80 , .to: X86::MUL_FrST0 },
1353 { .from: X86::SUB_Fp32 , .to: X86::SUB_FrST0 },
1354 { .from: X86::SUB_Fp64 , .to: X86::SUB_FrST0 },
1355 { .from: X86::SUB_Fp80 , .to: X86::SUB_FrST0 },
1356};
1357
1358
1359/// handleTwoArgFP - Handle instructions like FADD and friends which are virtual
1360/// instructions which need to be simplified and possibly transformed.
1361///
1362/// Result: ST(0) = fsub ST(0), ST(i)
1363/// ST(i) = fsub ST(0), ST(i)
1364/// ST(0) = fsubr ST(0), ST(i)
1365/// ST(i) = fsubr ST(0), ST(i)
1366///
1367void FPS::handleTwoArgFP(MachineBasicBlock::iterator &I) {
1368 ASSERT_SORTED(ForwardST0Table); ASSERT_SORTED(ReverseST0Table);
1369 ASSERT_SORTED(ForwardSTiTable); ASSERT_SORTED(ReverseSTiTable);
1370 MachineInstr &MI = *I;
1371
1372 unsigned NumOperands = MI.getDesc().getNumOperands();
1373 assert(NumOperands == 3 && "Illegal TwoArgFP instruction!");
1374 unsigned Dest = getFPReg(MO: MI.getOperand(i: 0));
1375 unsigned Op0 = getFPReg(MO: MI.getOperand(i: NumOperands - 2));
1376 unsigned Op1 = getFPReg(MO: MI.getOperand(i: NumOperands - 1));
1377 bool KillsOp0 = MI.killsRegister(Reg: X86::FP0 + Op0, /*TRI=*/nullptr);
1378 bool KillsOp1 = MI.killsRegister(Reg: X86::FP0 + Op1, /*TRI=*/nullptr);
1379 const DebugLoc &dl = MI.getDebugLoc();
1380
1381 unsigned TOS = getStackEntry(STi: 0);
1382
1383 // One of our operands must be on the top of the stack. If neither is yet, we
1384 // need to move one.
1385 if (Op0 != TOS && Op1 != TOS) { // No operand at TOS?
1386 // We can choose to move either operand to the top of the stack. If one of
1387 // the operands is killed by this instruction, we want that one so that we
1388 // can update right on top of the old version.
1389 if (KillsOp0) {
1390 moveToTop(RegNo: Op0, I); // Move dead operand to TOS.
1391 TOS = Op0;
1392 } else if (KillsOp1) {
1393 moveToTop(RegNo: Op1, I);
1394 TOS = Op1;
1395 } else {
1396 // All of the operands are live after this instruction executes, so we
1397 // cannot update on top of any operand. Because of this, we must
1398 // duplicate one of the stack elements to the top. It doesn't matter
1399 // which one we pick.
1400 //
1401 duplicateToTop(RegNo: Op0, AsReg: Dest, I);
1402 Op0 = TOS = Dest;
1403 KillsOp0 = true;
1404 }
1405 } else if (!KillsOp0 && !KillsOp1) {
1406 // If we DO have one of our operands at the top of the stack, but we don't
1407 // have a dead operand, we must duplicate one of the operands to a new slot
1408 // on the stack.
1409 duplicateToTop(RegNo: Op0, AsReg: Dest, I);
1410 Op0 = TOS = Dest;
1411 KillsOp0 = true;
1412 }
1413
1414 // Now we know that one of our operands is on the top of the stack, and at
1415 // least one of our operands is killed by this instruction.
1416 assert((TOS == Op0 || TOS == Op1) && (KillsOp0 || KillsOp1) &&
1417 "Stack conditions not set up right!");
1418
1419 // We decide which form to use based on what is on the top of the stack, and
1420 // which operand is killed by this instruction.
1421 ArrayRef<TableEntry> InstTable;
1422 bool isForward = TOS == Op0;
1423 bool updateST0 = (TOS == Op0 && !KillsOp1) || (TOS == Op1 && !KillsOp0);
1424 if (updateST0) {
1425 if (isForward)
1426 InstTable = ForwardST0Table;
1427 else
1428 InstTable = ReverseST0Table;
1429 } else {
1430 if (isForward)
1431 InstTable = ForwardSTiTable;
1432 else
1433 InstTable = ReverseSTiTable;
1434 }
1435
1436 int Opcode = Lookup(Table: InstTable, Opcode: MI.getOpcode());
1437 assert(Opcode != -1 && "Unknown TwoArgFP pseudo instruction!");
1438
1439 // NotTOS - The register which is not on the top of stack...
1440 unsigned NotTOS = (TOS == Op0) ? Op1 : Op0;
1441
1442 // Replace the old instruction with a new instruction
1443 MBB->remove(I: &*I++);
1444 I = BuildMI(BB&: *MBB, I, MIMD: dl, MCID: TII->get(Opcode)).addReg(RegNo: getSTReg(RegNo: NotTOS));
1445
1446 if (!MI.mayRaiseFPException())
1447 I->setFlag(MachineInstr::MIFlag::NoFPExcept);
1448
1449 // If both operands are killed, pop one off of the stack in addition to
1450 // overwriting the other one.
1451 if (KillsOp0 && KillsOp1 && Op0 != Op1) {
1452 assert(!updateST0 && "Should have updated other operand!");
1453 popStackAfter(I); // Pop the top of stack
1454 }
1455
1456 // Update stack information so that we know the destination register is now on
1457 // the stack.
1458 unsigned UpdatedSlot = getSlot(RegNo: updateST0 ? TOS : NotTOS);
1459 assert(UpdatedSlot < StackTop && Dest < 7);
1460 Stack[UpdatedSlot] = Dest;
1461 RegMap[Dest] = UpdatedSlot;
1462 MBB->getParent()->deleteMachineInstr(MI: &MI); // Remove the old instruction
1463}
1464
1465/// handleCompareFP - Handle FUCOM and FUCOMI instructions, which have two FP
1466/// register arguments and no explicit destinations.
1467///
1468void FPS::handleCompareFP(MachineBasicBlock::iterator &I) {
1469 MachineInstr &MI = *I;
1470
1471 unsigned NumOperands = MI.getDesc().getNumOperands();
1472 assert(NumOperands == 2 && "Illegal FUCOM* instruction!");
1473 unsigned Op0 = getFPReg(MO: MI.getOperand(i: NumOperands - 2));
1474 unsigned Op1 = getFPReg(MO: MI.getOperand(i: NumOperands - 1));
1475 bool KillsOp0 = MI.killsRegister(Reg: X86::FP0 + Op0, /*TRI=*/nullptr);
1476 bool KillsOp1 = MI.killsRegister(Reg: X86::FP0 + Op1, /*TRI=*/nullptr);
1477
1478 // Make sure the first operand is on the top of stack, the other one can be
1479 // anywhere.
1480 moveToTop(RegNo: Op0, I);
1481
1482 // Change from the pseudo instruction to the concrete instruction.
1483 MI.getOperand(i: 0).setReg(getSTReg(RegNo: Op1));
1484 MI.removeOperand(OpNo: 1);
1485 MI.setDesc(TII->get(Opcode: getConcreteOpcode(Opcode: MI.getOpcode())));
1486 MI.dropDebugNumber();
1487
1488 // If any of the operands are killed by this instruction, free them.
1489 if (KillsOp0) freeStackSlotAfter(I, FPRegNo: Op0);
1490 if (KillsOp1 && Op0 != Op1) freeStackSlotAfter(I, FPRegNo: Op1);
1491}
1492
1493/// handleCondMovFP - Handle two address conditional move instructions. These
1494/// instructions move a st(i) register to st(0) iff a condition is true. These
1495/// instructions require that the first operand is at the top of the stack, but
1496/// otherwise don't modify the stack at all.
1497void FPS::handleCondMovFP(MachineBasicBlock::iterator &I) {
1498 MachineInstr &MI = *I;
1499
1500 unsigned Op0 = getFPReg(MO: MI.getOperand(i: 0));
1501 unsigned Op1 = getFPReg(MO: MI.getOperand(i: 2));
1502 bool KillsOp1 = MI.killsRegister(Reg: X86::FP0 + Op1, /*TRI=*/nullptr);
1503
1504 // The first operand *must* be on the top of the stack.
1505 moveToTop(RegNo: Op0, I);
1506
1507 // Change the second operand to the stack register that the operand is in.
1508 // Change from the pseudo instruction to the concrete instruction.
1509 MI.removeOperand(OpNo: 0);
1510 MI.removeOperand(OpNo: 1);
1511 MI.getOperand(i: 0).setReg(getSTReg(RegNo: Op1));
1512 MI.setDesc(TII->get(Opcode: getConcreteOpcode(Opcode: MI.getOpcode())));
1513 MI.dropDebugNumber();
1514
1515 // If we kill the second operand, make sure to pop it from the stack.
1516 if (Op0 != Op1 && KillsOp1) {
1517 // Get this value off of the register stack.
1518 freeStackSlotAfter(I, FPRegNo: Op1);
1519 }
1520}
1521
1522
1523/// handleSpecialFP - Handle special instructions which behave unlike other
1524/// floating point instructions. This is primarily intended for use by pseudo
1525/// instructions.
1526///
1527void FPS::handleSpecialFP(MachineBasicBlock::iterator &Inst) {
1528 MachineInstr &MI = *Inst;
1529
1530 if (MI.isCall()) {
1531 handleCall(I&: Inst);
1532 return;
1533 }
1534
1535 if (MI.isReturn()) {
1536 handleReturn(I&: Inst);
1537 return;
1538 }
1539
1540 switch (MI.getOpcode()) {
1541 default: llvm_unreachable("Unknown SpecialFP instruction!");
1542 case TargetOpcode::COPY: {
1543 // We handle three kinds of copies: FP <- FP, FP <- ST, and ST <- FP.
1544 const MachineOperand &MO1 = MI.getOperand(i: 1);
1545 const MachineOperand &MO0 = MI.getOperand(i: 0);
1546 bool KillsSrc = MI.killsRegister(Reg: MO1.getReg(), /*TRI=*/nullptr);
1547
1548 // FP <- FP copy.
1549 unsigned DstFP = getFPReg(MO: MO0);
1550 unsigned SrcFP = getFPReg(MO: MO1);
1551 assert(isLive(SrcFP) && "Cannot copy dead register");
1552 if (KillsSrc) {
1553 // If the input operand is killed, we can just change the owner of the
1554 // incoming stack slot into the result.
1555 unsigned Slot = getSlot(RegNo: SrcFP);
1556 Stack[Slot] = DstFP;
1557 RegMap[DstFP] = Slot;
1558 } else {
1559 // For COPY we just duplicate the specified value to a new stack slot.
1560 // This could be made better, but would require substantial changes.
1561 duplicateToTop(RegNo: SrcFP, AsReg: DstFP, I: Inst);
1562 }
1563 break;
1564 }
1565
1566 case TargetOpcode::IMPLICIT_DEF: {
1567 // All FP registers must be explicitly defined, so load a 0 instead.
1568 unsigned Reg = MI.getOperand(i: 0).getReg() - X86::FP0;
1569 LLVM_DEBUG(dbgs() << "Emitting LD_F0 for implicit FP" << Reg << '\n');
1570 BuildMI(BB&: *MBB, I: Inst, MIMD: MI.getDebugLoc(), MCID: TII->get(Opcode: X86::LD_F0));
1571 pushReg(Reg);
1572 break;
1573 }
1574
1575 case TargetOpcode::INLINEASM:
1576 case TargetOpcode::INLINEASM_BR: {
1577 // The inline asm MachineInstr currently only *uses* FP registers for the
1578 // 'f' constraint. These should be turned into the current ST(x) register
1579 // in the machine instr.
1580 //
1581 // There are special rules for x87 inline assembly. The compiler must know
1582 // exactly how many registers are popped and pushed implicitly by the asm.
1583 // Otherwise it is not possible to restore the stack state after the inline
1584 // asm.
1585 //
1586 // There are 3 kinds of input operands:
1587 //
1588 // 1. Popped inputs. These must appear at the stack top in ST0-STn. A
1589 // popped input operand must be in a fixed stack slot, and it is either
1590 // tied to an output operand, or in the clobber list. The MI has ST use
1591 // and def operands for these inputs.
1592 //
1593 // 2. Fixed inputs. These inputs appear in fixed stack slots, but are
1594 // preserved by the inline asm. The fixed stack slots must be STn-STm
1595 // following the popped inputs. A fixed input operand cannot be tied to
1596 // an output or appear in the clobber list. The MI has ST use operands
1597 // and no defs for these inputs.
1598 //
1599 // 3. Preserved inputs. These inputs use the "f" constraint which is
1600 // represented as an FP register. The inline asm won't change these
1601 // stack slots.
1602 //
1603 // Outputs must be in ST registers, FP outputs are not allowed. Clobbered
1604 // registers do not count as output operands. The inline asm changes the
1605 // stack as if it popped all the popped inputs and then pushed all the
1606 // output operands.
1607
1608 // Scan the assembly for ST registers used, defined and clobbered. We can
1609 // only tell clobbers from defs by looking at the asm descriptor.
1610 unsigned STUses = 0, STDefs = 0, STClobbers = 0;
1611 unsigned NumOps = 0;
1612 SmallSet<unsigned, 1> FRegIdx;
1613 unsigned RCID;
1614
1615 for (unsigned i = InlineAsm::MIOp_FirstOperand, e = MI.getNumOperands();
1616 i != e && MI.getOperand(i).isImm(); i += 1 + NumOps) {
1617 unsigned Flags = MI.getOperand(i).getImm();
1618 const InlineAsm::Flag F(Flags);
1619
1620 NumOps = F.getNumOperandRegisters();
1621 if (NumOps != 1)
1622 continue;
1623 const MachineOperand &MO = MI.getOperand(i: i + 1);
1624 if (!MO.isReg())
1625 continue;
1626 unsigned STReg = MO.getReg() - X86::FP0;
1627 if (STReg >= 8)
1628 continue;
1629
1630 // If the flag has a register class constraint, this must be an operand
1631 // with constraint "f". Record its index and continue.
1632 if (F.hasRegClassConstraint(RC&: RCID)) {
1633 FRegIdx.insert(V: i + 1);
1634 continue;
1635 }
1636
1637 switch (F.getKind()) {
1638 case InlineAsm::Kind::RegUse:
1639 STUses |= (1u << STReg);
1640 break;
1641 case InlineAsm::Kind::RegDef:
1642 case InlineAsm::Kind::RegDefEarlyClobber:
1643 STDefs |= (1u << STReg);
1644 break;
1645 case InlineAsm::Kind::Clobber:
1646 STClobbers |= (1u << STReg);
1647 break;
1648 default:
1649 break;
1650 }
1651 }
1652
1653 if (STUses && !isMask_32(Value: STUses))
1654 MI.emitGenericError(ErrMsg: "fixed input regs must be last on the x87 stack");
1655 unsigned NumSTUses = llvm::countr_one(Value: STUses);
1656
1657 // Defs must be contiguous from the stack top. ST0-STn.
1658 if (STDefs && !isMask_32(Value: STDefs)) {
1659 MI.emitGenericError(ErrMsg: "output regs must be last on the x87 stack");
1660 STDefs = NextPowerOf2(A: STDefs) - 1;
1661 }
1662 unsigned NumSTDefs = llvm::countr_one(Value: STDefs);
1663
1664 // So must the clobbered stack slots. ST0-STm, m >= n.
1665 if (STClobbers && !isMask_32(Value: STDefs | STClobbers))
1666 MI.emitGenericError(ErrMsg: "clobbers must be last on the x87 stack");
1667
1668 // Popped inputs are the ones that are also clobbered or defined.
1669 unsigned STPopped = STUses & (STDefs | STClobbers);
1670 if (STPopped && !isMask_32(Value: STPopped))
1671 MI.emitGenericError(
1672 ErrMsg: "implicitly popped regs must be last on the x87 stack");
1673 unsigned NumSTPopped = llvm::countr_one(Value: STPopped);
1674
1675 LLVM_DEBUG(dbgs() << "Asm uses " << NumSTUses << " fixed regs, pops "
1676 << NumSTPopped << ", and defines " << NumSTDefs
1677 << " regs.\n");
1678
1679#ifndef NDEBUG
1680 // If any input operand uses constraint "f", all output register
1681 // constraints must be early-clobber defs.
1682 for (unsigned I = 0, E = MI.getNumOperands(); I < E; ++I)
1683 if (FRegIdx.count(I)) {
1684 assert((1 << getFPReg(MI.getOperand(I)) & STDefs) == 0 &&
1685 "Operands with constraint \"f\" cannot overlap with defs");
1686 }
1687#endif
1688
1689 // Collect all FP registers (register operands with constraints "t", "u",
1690 // and "f") to kill afer the instruction.
1691 unsigned FPKills = ((1u << NumFPRegs) - 1) & ~0xff;
1692 for (const MachineOperand &Op : MI.operands()) {
1693 if (!Op.isReg() || Op.getReg() < X86::FP0 || Op.getReg() > X86::FP6)
1694 continue;
1695 unsigned FPReg = getFPReg(MO: Op);
1696
1697 // If we kill this operand, make sure to pop it from the stack after the
1698 // asm. We just remember it for now, and pop them all off at the end in
1699 // a batch.
1700 if (Op.isUse() && Op.isKill())
1701 FPKills |= 1U << FPReg;
1702 }
1703
1704 // Do not include registers that are implicitly popped by defs/clobbers.
1705 FPKills &= ~(STDefs | STClobbers);
1706
1707 // Now we can rearrange the live registers to match what was requested.
1708 unsigned char STUsesArray[8];
1709
1710 for (unsigned I = 0; I < NumSTUses; ++I)
1711 STUsesArray[I] = I;
1712
1713 shuffleStackTop(FixStack: STUsesArray, FixCount: NumSTUses, I: Inst);
1714 LLVM_DEBUG({
1715 dbgs() << "Before asm: ";
1716 dumpStack();
1717 });
1718
1719 // With the stack layout fixed, rewrite the FP registers.
1720 for (unsigned i = 0, e = MI.getNumOperands(); i != e; ++i) {
1721 MachineOperand &Op = MI.getOperand(i);
1722 if (!Op.isReg() || Op.getReg() < X86::FP0 || Op.getReg() > X86::FP6)
1723 continue;
1724
1725 unsigned FPReg = getFPReg(MO: Op);
1726
1727 if (FRegIdx.count(V: i))
1728 // Operand with constraint "f".
1729 Op.setReg(getSTReg(RegNo: FPReg));
1730 else
1731 // Operand with a single register class constraint ("t" or "u").
1732 Op.setReg(X86::ST0 + FPReg);
1733 }
1734
1735 // Simulate the inline asm popping its inputs and pushing its outputs.
1736 StackTop -= NumSTPopped;
1737
1738 for (unsigned i = 0; i < NumSTDefs; ++i)
1739 pushReg(Reg: NumSTDefs - i - 1);
1740
1741 // If this asm kills any FP registers (is the last use of them) we must
1742 // explicitly emit pop instructions for them. Do this now after the asm has
1743 // executed so that the ST(x) numbers are not off (which would happen if we
1744 // did this inline with operand rewriting).
1745 //
1746 // Note: this might be a non-optimal pop sequence. We might be able to do
1747 // better by trying to pop in stack order or something.
1748 while (FPKills) {
1749 unsigned FPReg = llvm::countr_zero(Val: FPKills);
1750 if (isLive(RegNo: FPReg))
1751 freeStackSlotAfter(I&: Inst, FPRegNo: FPReg);
1752 FPKills &= ~(1U << FPReg);
1753 }
1754
1755 // Don't delete the inline asm!
1756 return;
1757 }
1758
1759 // FAKE_USE must pop its register operand off the stack if it is killed,
1760 // because this constitutes the register's last use. If the operand
1761 // is not killed, it will have its last use later, so we leave it alone.
1762 // In either case we remove the operand so later passes don't see it.
1763 case TargetOpcode::FAKE_USE: {
1764 assert(MI.getNumExplicitOperands() == 1 &&
1765 "FAKE_USE must have exactly one operand");
1766 if (MI.getOperand(i: 0).isKill()) {
1767 freeStackSlotBefore(I: Inst, FPRegNo: getFPReg(MO: MI.getOperand(i: 0)));
1768 }
1769 MI.removeOperand(OpNo: 0);
1770 return;
1771 }
1772 }
1773
1774 Inst = MBB->erase(I: Inst); // Remove the pseudo instruction
1775
1776 // We want to leave I pointing to the previous instruction, but what if we
1777 // just erased the first instruction?
1778 if (Inst == MBB->begin()) {
1779 LLVM_DEBUG(dbgs() << "Inserting dummy KILL\n");
1780 Inst = BuildMI(BB&: *MBB, I: Inst, MIMD: DebugLoc(), MCID: TII->get(Opcode: TargetOpcode::KILL));
1781 } else
1782 --Inst;
1783}
1784
1785void FPS::setKillFlags(MachineBasicBlock &MBB) const {
1786 const TargetRegisterInfo &TRI =
1787 *MBB.getParent()->getSubtarget().getRegisterInfo();
1788 LiveRegUnits LPR(TRI);
1789
1790 LPR.addLiveOuts(MBB);
1791
1792 for (MachineInstr &MI : llvm::reverse(C&: MBB)) {
1793 if (MI.isDebugInstr())
1794 continue;
1795
1796 std::bitset<8> Defs;
1797 SmallVector<MachineOperand *, 2> Uses;
1798
1799 for (auto &MO : MI.operands()) {
1800 if (!MO.isReg())
1801 continue;
1802
1803 unsigned Reg = MO.getReg() - X86::FP0;
1804
1805 if (Reg >= 8)
1806 continue;
1807
1808 if (MO.isDef()) {
1809 Defs.set(position: Reg);
1810 if (LPR.available(Reg: MO.getReg()))
1811 MO.setIsDead();
1812 } else
1813 Uses.push_back(Elt: &MO);
1814 }
1815
1816 for (auto *MO : Uses)
1817 if (Defs.test(position: getFPReg(MO: *MO)) || LPR.available(Reg: MO->getReg()))
1818 MO->setIsKill();
1819
1820 LPR.stepBackward(MI);
1821 }
1822}
1823