X86VZeroUpper.cpp
12.6 KB
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
//===- X86VZeroUpper.cpp - AVX vzeroupper instruction inserter ------------===//
//
// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
// See https://llvm.org/LICENSE.txt for license information.
// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
//
//===----------------------------------------------------------------------===//
//
// This file defines the pass which inserts x86 AVX vzeroupper instructions
// before calls to SSE encoded functions. This avoids transition latency
// penalty when transferring control between AVX encoded instructions and old
// SSE encoding mode.
//
//===----------------------------------------------------------------------===//
#include "X86.h"
#include "X86InstrInfo.h"
#include "X86Subtarget.h"
#include "llvm/ADT/SmallVector.h"
#include "llvm/ADT/Statistic.h"
#include "llvm/CodeGen/MachineBasicBlock.h"
#include "llvm/CodeGen/MachineFunction.h"
#include "llvm/CodeGen/MachineFunctionPass.h"
#include "llvm/CodeGen/MachineInstr.h"
#include "llvm/CodeGen/MachineInstrBuilder.h"
#include "llvm/CodeGen/MachineOperand.h"
#include "llvm/CodeGen/MachineRegisterInfo.h"
#include "llvm/CodeGen/TargetInstrInfo.h"
#include "llvm/CodeGen/TargetRegisterInfo.h"
#include "llvm/IR/CallingConv.h"
#include "llvm/IR/DebugLoc.h"
#include "llvm/IR/Function.h"
#include "llvm/Support/Debug.h"
#include "llvm/Support/ErrorHandling.h"
#include "llvm/Support/raw_ostream.h"
#include <cassert>
using namespace llvm;
#define DEBUG_TYPE "x86-vzeroupper"
static cl::opt<bool>
UseVZeroUpper("x86-use-vzeroupper", cl::Hidden,
cl::desc("Minimize AVX to SSE transition penalty"),
cl::init(true));
STATISTIC(NumVZU, "Number of vzeroupper instructions inserted");
namespace {
class VZeroUpperInserter : public MachineFunctionPass {
public:
VZeroUpperInserter() : MachineFunctionPass(ID) {}
bool runOnMachineFunction(MachineFunction &MF) override;
MachineFunctionProperties getRequiredProperties() const override {
return MachineFunctionProperties().set(
MachineFunctionProperties::Property::NoVRegs);
}
StringRef getPassName() const override { return "X86 vzeroupper inserter"; }
private:
void processBasicBlock(MachineBasicBlock &MBB);
void insertVZeroUpper(MachineBasicBlock::iterator I,
MachineBasicBlock &MBB);
void addDirtySuccessor(MachineBasicBlock &MBB);
using BlockExitState = enum { PASS_THROUGH, EXITS_CLEAN, EXITS_DIRTY };
static const char* getBlockExitStateName(BlockExitState ST);
// Core algorithm state:
// BlockState - Each block is either:
// - PASS_THROUGH: There are neither YMM/ZMM dirtying instructions nor
// vzeroupper instructions in this block.
// - EXITS_CLEAN: There is (or will be) a vzeroupper instruction in this
// block that will ensure that YMM/ZMM is clean on exit.
// - EXITS_DIRTY: An instruction in the block dirties YMM/ZMM and no
// subsequent vzeroupper in the block clears it.
//
// AddedToDirtySuccessors - This flag is raised when a block is added to the
// DirtySuccessors list to ensure that it's not
// added multiple times.
//
// FirstUnguardedCall - Records the location of the first unguarded call in
// each basic block that may need to be guarded by a
// vzeroupper. We won't know whether it actually needs
// to be guarded until we discover a predecessor that
// is DIRTY_OUT.
struct BlockState {
BlockExitState ExitState = PASS_THROUGH;
bool AddedToDirtySuccessors = false;
MachineBasicBlock::iterator FirstUnguardedCall;
BlockState() = default;
};
using BlockStateMap = SmallVector<BlockState, 8>;
using DirtySuccessorsWorkList = SmallVector<MachineBasicBlock *, 8>;
BlockStateMap BlockStates;
DirtySuccessorsWorkList DirtySuccessors;
bool EverMadeChange;
bool IsX86INTR;
const TargetInstrInfo *TII;
static char ID;
};
} // end anonymous namespace
char VZeroUpperInserter::ID = 0;
FunctionPass *llvm::createX86IssueVZeroUpperPass() {
return new VZeroUpperInserter();
}
#ifndef NDEBUG
const char* VZeroUpperInserter::getBlockExitStateName(BlockExitState ST) {
switch (ST) {
case PASS_THROUGH: return "Pass-through";
case EXITS_DIRTY: return "Exits-dirty";
case EXITS_CLEAN: return "Exits-clean";
}
llvm_unreachable("Invalid block exit state.");
}
#endif
/// VZEROUPPER cleans state that is related to Y/ZMM0-15 only.
/// Thus, there is no need to check for Y/ZMM16 and above.
static bool isYmmOrZmmReg(unsigned Reg) {
return (Reg >= X86::YMM0 && Reg <= X86::YMM15) ||
(Reg >= X86::ZMM0 && Reg <= X86::ZMM15);
}
static bool checkFnHasLiveInYmmOrZmm(MachineRegisterInfo &MRI) {
for (std::pair<unsigned, unsigned> LI : MRI.liveins())
if (isYmmOrZmmReg(LI.first))
return true;
return false;
}
static bool clobbersAllYmmAndZmmRegs(const MachineOperand &MO) {
for (unsigned reg = X86::YMM0; reg <= X86::YMM15; ++reg) {
if (!MO.clobbersPhysReg(reg))
return false;
}
for (unsigned reg = X86::ZMM0; reg <= X86::ZMM15; ++reg) {
if (!MO.clobbersPhysReg(reg))
return false;
}
return true;
}
static bool hasYmmOrZmmReg(MachineInstr &MI) {
for (const MachineOperand &MO : MI.operands()) {
if (MI.isCall() && MO.isRegMask() && !clobbersAllYmmAndZmmRegs(MO))
return true;
if (!MO.isReg())
continue;
if (MO.isDebug())
continue;
if (isYmmOrZmmReg(MO.getReg()))
return true;
}
return false;
}
/// Check if given call instruction has a RegMask operand.
static bool callHasRegMask(MachineInstr &MI) {
assert(MI.isCall() && "Can only be called on call instructions.");
for (const MachineOperand &MO : MI.operands()) {
if (MO.isRegMask())
return true;
}
return false;
}
/// Insert a vzeroupper instruction before I.
void VZeroUpperInserter::insertVZeroUpper(MachineBasicBlock::iterator I,
MachineBasicBlock &MBB) {
DebugLoc dl = I->getDebugLoc();
BuildMI(MBB, I, dl, TII->get(X86::VZEROUPPER));
++NumVZU;
EverMadeChange = true;
}
/// Add MBB to the DirtySuccessors list if it hasn't already been added.
void VZeroUpperInserter::addDirtySuccessor(MachineBasicBlock &MBB) {
if (!BlockStates[MBB.getNumber()].AddedToDirtySuccessors) {
DirtySuccessors.push_back(&MBB);
BlockStates[MBB.getNumber()].AddedToDirtySuccessors = true;
}
}
/// Loop over all of the instructions in the basic block, inserting vzeroupper
/// instructions before function calls.
void VZeroUpperInserter::processBasicBlock(MachineBasicBlock &MBB) {
// Start by assuming that the block is PASS_THROUGH which implies no unguarded
// calls.
BlockExitState CurState = PASS_THROUGH;
BlockStates[MBB.getNumber()].FirstUnguardedCall = MBB.end();
for (MachineInstr &MI : MBB) {
bool IsCall = MI.isCall();
bool IsReturn = MI.isReturn();
bool IsControlFlow = IsCall || IsReturn;
// No need for vzeroupper before iret in interrupt handler function,
// epilogue will restore YMM/ZMM registers if needed.
if (IsX86INTR && IsReturn)
continue;
// An existing VZERO* instruction resets the state.
if (MI.getOpcode() == X86::VZEROALL || MI.getOpcode() == X86::VZEROUPPER) {
CurState = EXITS_CLEAN;
continue;
}
// Shortcut: don't need to check regular instructions in dirty state.
if (!IsControlFlow && CurState == EXITS_DIRTY)
continue;
if (hasYmmOrZmmReg(MI)) {
// We found a ymm/zmm-using instruction; this could be an AVX/AVX512
// instruction, or it could be control flow.
CurState = EXITS_DIRTY;
continue;
}
// Check for control-flow out of the current function (which might
// indirectly execute SSE instructions).
if (!IsControlFlow)
continue;
// If the call has no RegMask, skip it as well. It usually happens on
// helper function calls (such as '_chkstk', '_ftol2') where standard
// calling convention is not used (RegMask is not used to mark register
// clobbered and register usage (def/implicit-def/use) is well-defined and
// explicitly specified.
if (IsCall && !callHasRegMask(MI))
continue;
// The VZEROUPPER instruction resets the upper 128 bits of YMM0-YMM15
// registers. In addition, the processor changes back to Clean state, after
// which execution of SSE instructions or AVX instructions has no transition
// penalty. Add the VZEROUPPER instruction before any function call/return
// that might execute SSE code.
// FIXME: In some cases, we may want to move the VZEROUPPER into a
// predecessor block.
if (CurState == EXITS_DIRTY) {
// After the inserted VZEROUPPER the state becomes clean again, but
// other YMM/ZMM may appear before other subsequent calls or even before
// the end of the BB.
insertVZeroUpper(MI, MBB);
CurState = EXITS_CLEAN;
} else if (CurState == PASS_THROUGH) {
// If this block is currently in pass-through state and we encounter a
// call then whether we need a vzeroupper or not depends on whether this
// block has successors that exit dirty. Record the location of the call,
// and set the state to EXITS_CLEAN, but do not insert the vzeroupper yet.
// It will be inserted later if necessary.
BlockStates[MBB.getNumber()].FirstUnguardedCall = MI;
CurState = EXITS_CLEAN;
}
}
LLVM_DEBUG(dbgs() << "MBB #" << MBB.getNumber() << " exit state: "
<< getBlockExitStateName(CurState) << '\n');
if (CurState == EXITS_DIRTY)
for (MachineBasicBlock::succ_iterator SI = MBB.succ_begin(),
SE = MBB.succ_end();
SI != SE; ++SI)
addDirtySuccessor(**SI);
BlockStates[MBB.getNumber()].ExitState = CurState;
}
/// Loop over all of the basic blocks, inserting vzeroupper instructions before
/// function calls.
bool VZeroUpperInserter::runOnMachineFunction(MachineFunction &MF) {
if (!UseVZeroUpper)
return false;
const X86Subtarget &ST = MF.getSubtarget<X86Subtarget>();
if (!ST.hasAVX() || !ST.insertVZEROUPPER())
return false;
TII = ST.getInstrInfo();
MachineRegisterInfo &MRI = MF.getRegInfo();
EverMadeChange = false;
IsX86INTR = MF.getFunction().getCallingConv() == CallingConv::X86_INTR;
bool FnHasLiveInYmmOrZmm = checkFnHasLiveInYmmOrZmm(MRI);
// Fast check: if the function doesn't use any ymm/zmm registers, we don't
// need to insert any VZEROUPPER instructions. This is constant-time, so it
// is cheap in the common case of no ymm/zmm use.
bool YmmOrZmmUsed = FnHasLiveInYmmOrZmm;
for (auto *RC : {&X86::VR256RegClass, &X86::VR512_0_15RegClass}) {
if (!YmmOrZmmUsed) {
for (TargetRegisterClass::iterator i = RC->begin(), e = RC->end(); i != e;
i++) {
if (!MRI.reg_nodbg_empty(*i)) {
YmmOrZmmUsed = true;
break;
}
}
}
}
if (!YmmOrZmmUsed)
return false;
assert(BlockStates.empty() && DirtySuccessors.empty() &&
"X86VZeroUpper state should be clear");
BlockStates.resize(MF.getNumBlockIDs());
// Process all blocks. This will compute block exit states, record the first
// unguarded call in each block, and add successors of dirty blocks to the
// DirtySuccessors list.
for (MachineBasicBlock &MBB : MF)
processBasicBlock(MBB);
// If any YMM/ZMM regs are live-in to this function, add the entry block to
// the DirtySuccessors list
if (FnHasLiveInYmmOrZmm)
addDirtySuccessor(MF.front());
// Re-visit all blocks that are successors of EXITS_DIRTY blocks. Add
// vzeroupper instructions to unguarded calls, and propagate EXITS_DIRTY
// through PASS_THROUGH blocks.
while (!DirtySuccessors.empty()) {
MachineBasicBlock &MBB = *DirtySuccessors.back();
DirtySuccessors.pop_back();
BlockState &BBState = BlockStates[MBB.getNumber()];
// MBB is a successor of a dirty block, so its first call needs to be
// guarded.
if (BBState.FirstUnguardedCall != MBB.end())
insertVZeroUpper(BBState.FirstUnguardedCall, MBB);
// If this successor was a pass-through block, then it is now dirty. Its
// successors need to be added to the worklist (if they haven't been
// already).
if (BBState.ExitState == PASS_THROUGH) {
LLVM_DEBUG(dbgs() << "MBB #" << MBB.getNumber()
<< " was Pass-through, is now Dirty-out.\n");
for (MachineBasicBlock *Succ : MBB.successors())
addDirtySuccessor(*Succ);
}
}
BlockStates.clear();
return EverMadeChange;
}