X86VZeroUpper.cpp revision 9d57cdfa172c1af3712ef52862a1976c57c679a5
1//===-- X86VZeroUpper.cpp - AVX vzeroupper instruction inserter -----------===//
2//
3//                     The LLVM Compiler Infrastructure
4//
5// This file is distributed under the University of Illinois Open Source
6// License. See LICENSE.TXT for details.
7//
8//===----------------------------------------------------------------------===//
9//
10// This file defines the pass which inserts x86 AVX vzeroupper instructions
11// before calls to SSE encoded functions. This avoids transition latency
12// penalty when tranfering control between AVX encoded instructions and old
13// SSE encoding mode.
14//
15//===----------------------------------------------------------------------===//
16
17#define DEBUG_TYPE "x86-vzeroupper"
18#include "X86.h"
19#include "X86InstrInfo.h"
20#include "llvm/ADT/Statistic.h"
21#include "llvm/CodeGen/MachineFunctionPass.h"
22#include "llvm/CodeGen/MachineInstrBuilder.h"
23#include "llvm/CodeGen/MachineRegisterInfo.h"
24#include "llvm/CodeGen/Passes.h"
25#include "llvm/Support/Debug.h"
26#include "llvm/Support/raw_ostream.h"
27#include "llvm/Target/TargetInstrInfo.h"
28using namespace llvm;
29
30STATISTIC(NumVZU, "Number of vzeroupper instructions inserted");
31
32namespace {
33  struct VZeroUpperInserter : public MachineFunctionPass {
34    static char ID;
35    VZeroUpperInserter() : MachineFunctionPass(ID) {}
36
37    virtual bool runOnMachineFunction(MachineFunction &MF);
38
39    bool processBasicBlock(MachineFunction &MF, MachineBasicBlock &MBB);
40
41    virtual const char *getPassName() const { return "X86 vzeroupper inserter";}
42
43  private:
44    const TargetInstrInfo *TII; // Machine instruction info.
45
46    // Any YMM register live-in to this function?
47    bool FnHasLiveInYmm;
48
49    // BBState - Contains the state of each MBB: unknown, clean, dirty
50    SmallVector<uint8_t, 8> BBState;
51
52    // BBSolved - Keep track of all MBB which had been already analyzed
53    // and there is no further processing required.
54    BitVector BBSolved;
55
56    // Machine Basic Blocks are classified according this pass:
57    //
58    //  ST_UNKNOWN - The MBB state is unknown, meaning from the entry state
59    //    until the MBB exit there isn't a instruction using YMM to change
60    //    the state to dirty, or one of the incoming predecessors is unknown
61    //    and there's not a dirty predecessor between them.
62    //
63    //  ST_CLEAN - No YMM usage in the end of the MBB. A MBB could have
64    //    instructions using YMM and be marked ST_CLEAN, as long as the state
65    //    is cleaned by a vzeroupper before any call.
66    //
67    //  ST_DIRTY - Any MBB ending with a YMM usage not cleaned up by a
68    //    vzeroupper instruction.
69    //
70    //  ST_INIT - Placeholder for an empty state set
71    //
72    enum {
73      ST_UNKNOWN = 0,
74      ST_CLEAN   = 1,
75      ST_DIRTY   = 2,
76      ST_INIT    = 3
77    };
78
79    // computeState - Given two states, compute the resulting state, in
80    // the following way
81    //
82    //  1) One dirty state yields another dirty state
83    //  2) All states must be clean for the result to be clean
84    //  3) If none above and one unknown, the result state is also unknown
85    //
86    static unsigned computeState(unsigned PrevState, unsigned CurState) {
87      if (PrevState == ST_INIT)
88        return CurState;
89
90      if (PrevState == ST_DIRTY || CurState == ST_DIRTY)
91        return ST_DIRTY;
92
93      if (PrevState == ST_CLEAN && CurState == ST_CLEAN)
94        return ST_CLEAN;
95
96      return ST_UNKNOWN;
97    }
98
99  };
100  char VZeroUpperInserter::ID = 0;
101}
102
103FunctionPass *llvm::createX86IssueVZeroUpperPass() {
104  return new VZeroUpperInserter();
105}
106
107static bool isYmmReg(unsigned Reg) {
108  if (Reg >= X86::YMM0 && Reg <= X86::YMM15)
109    return true;
110
111  return false;
112}
113
114static bool checkFnHasLiveInYmm(MachineRegisterInfo &MRI) {
115  for (MachineRegisterInfo::livein_iterator I = MRI.livein_begin(),
116       E = MRI.livein_end(); I != E; ++I)
117    if (isYmmReg(I->first))
118      return true;
119
120  return false;
121}
122
123static bool hasYmmReg(MachineInstr *MI) {
124  for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
125    const MachineOperand &MO = MI->getOperand(i);
126    if (!MO.isReg())
127      continue;
128    if (MO.isDebug())
129      continue;
130    if (isYmmReg(MO.getReg()))
131      return true;
132  }
133  return false;
134}
135
136/// runOnMachineFunction - Loop over all of the basic blocks, inserting
137/// vzero upper instructions before function calls.
138bool VZeroUpperInserter::runOnMachineFunction(MachineFunction &MF) {
139  TII = MF.getTarget().getInstrInfo();
140  MachineRegisterInfo &MRI = MF.getRegInfo();
141  bool EverMadeChange = false;
142
143  // Fast check: if the function doesn't use any ymm registers, we don't need
144  // to insert any VZEROUPPER instructions.  This is constant-time, so it is
145  // cheap in the common case of no ymm use.
146  bool YMMUsed = false;
147  const TargetRegisterClass *RC = &X86::VR256RegClass;
148  for (TargetRegisterClass::iterator i = RC->begin(), e = RC->end();
149       i != e; i++) {
150    if (!MRI.reg_nodbg_empty(*i)) {
151      YMMUsed = true;
152      break;
153    }
154  }
155  if (!YMMUsed)
156    return EverMadeChange;
157
158  // Pre-compute the existence of any live-in YMM registers to this function
159  FnHasLiveInYmm = checkFnHasLiveInYmm(MRI);
160
161  assert(BBState.empty());
162  BBState.resize(MF.getNumBlockIDs(), 0);
163  BBSolved.resize(MF.getNumBlockIDs(), 0);
164
165  // Each BB state depends on all predecessors, loop over until everything
166  // converges.  (Once we converge, we can implicitly mark everything that is
167  // still ST_UNKNOWN as ST_CLEAN.)
168  while (1) {
169    bool MadeChange = false;
170
171    // Process all basic blocks.
172    for (MachineFunction::iterator I = MF.begin(), E = MF.end(); I != E; ++I)
173      MadeChange |= processBasicBlock(MF, *I);
174
175    // If this iteration over the code changed anything, keep iterating.
176    if (!MadeChange) break;
177    EverMadeChange = true;
178  }
179
180  BBState.clear();
181  BBSolved.clear();
182  return EverMadeChange;
183}
184
185/// processBasicBlock - Loop over all of the instructions in the basic block,
186/// inserting vzero upper instructions before function calls.
187bool VZeroUpperInserter::processBasicBlock(MachineFunction &MF,
188                                           MachineBasicBlock &BB) {
189  bool Changed = false;
190  unsigned BBNum = BB.getNumber();
191
192  // Don't process already solved BBs
193  if (BBSolved[BBNum])
194    return false; // No changes
195
196  // Check the state of all predecessors
197  unsigned EntryState = ST_INIT;
198  for (MachineBasicBlock::const_pred_iterator PI = BB.pred_begin(),
199       PE = BB.pred_end(); PI != PE; ++PI) {
200    EntryState = computeState(EntryState, BBState[(*PI)->getNumber()]);
201    if (EntryState == ST_DIRTY)
202      break;
203  }
204
205
206  // The entry MBB for the function may set the initial state to dirty if
207  // the function receives any YMM incoming arguments
208  if (&BB == MF.begin()) {
209    EntryState = ST_CLEAN;
210    if (FnHasLiveInYmm)
211      EntryState = ST_DIRTY;
212  }
213
214  // The current state is initialized according to the predecessors
215  unsigned CurState = EntryState;
216  bool BBHasCall = false;
217
218  for (MachineBasicBlock::iterator I = BB.begin(); I != BB.end(); ++I) {
219    MachineInstr *MI = I;
220    DebugLoc dl = I->getDebugLoc();
221    bool isControlFlow = MI->isCall() || MI->isReturn();
222
223    // Shortcut: don't need to check regular instructions in dirty state.
224    if (!isControlFlow && CurState == ST_DIRTY)
225      continue;
226
227    if (hasYmmReg(MI)) {
228      // We found a ymm-using instruction; this could be an AVX instruction,
229      // or it could be control flow.
230      CurState = ST_DIRTY;
231      continue;
232    }
233
234    // Check for control-flow out of the current function (which might
235    // indirectly execute SSE instructions).
236    if (!isControlFlow)
237      continue;
238
239    BBHasCall = true;
240
241    // The VZEROUPPER instruction resets the upper 128 bits of all Intel AVX
242    // registers. This instruction has zero latency. In addition, the processor
243    // changes back to Clean state, after which execution of Intel SSE
244    // instructions or Intel AVX instructions has no transition penalty. Add
245    // the VZEROUPPER instruction before any function call/return that might
246    // execute SSE code.
247    // FIXME: In some cases, we may want to move the VZEROUPPER into a
248    // predecessor block.
249    if (CurState == ST_DIRTY) {
250      // Only insert the VZEROUPPER in case the entry state isn't unknown.
251      // When unknown, only compute the information within the block to have
252      // it available in the exit if possible, but don't change the block.
253      if (EntryState != ST_UNKNOWN) {
254        BuildMI(BB, I, dl, TII->get(X86::VZEROUPPER));
255        ++NumVZU;
256      }
257
258      // After the inserted VZEROUPPER the state becomes clean again, but
259      // other YMM may appear before other subsequent calls or even before
260      // the end of the BB.
261      CurState = ST_CLEAN;
262    }
263  }
264
265  DEBUG(dbgs() << "MBB #" << BBNum
266               << ", current state: " << CurState << '\n');
267
268  // A BB can only be considered solved when we both have done all the
269  // necessary transformations, and have computed the exit state.  This happens
270  // in two cases:
271  //  1) We know the entry state: this immediately implies the exit state and
272  //     all the necessary transformations.
273  //  2) There are no calls, and and a non-call instruction marks this block:
274  //     no transformations are necessary, and we know the exit state.
275  if (EntryState != ST_UNKNOWN || (!BBHasCall && CurState != ST_UNKNOWN))
276    BBSolved[BBNum] = true;
277
278  if (CurState != BBState[BBNum])
279    Changed = true;
280
281  BBState[BBNum] = CurState;
282  return Changed;
283}
284