aboutsummaryrefslogtreecommitdiffstats
path: root/contrib/libs/llvm12/lib/Target/X86/X86EvexToVex.cpp
blob: 97f843fa24eb4b8a1f6db5262355828ef7621f71 (plain) (blame)
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
//===- X86EvexToVex.cpp ---------------------------------------------------===//
// Compress EVEX instructions to VEX encoding when possible to reduce code size
//
// 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
//
//===----------------------------------------------------------------------===//
//
/// \file
/// This file defines the pass that goes over all AVX-512 instructions which
/// are encoded using the EVEX prefix and if possible replaces them by their
/// corresponding VEX encoding which is usually shorter by 2 bytes.
/// EVEX instructions may be encoded via the VEX prefix when the AVX-512
/// instruction has a corresponding AVX/AVX2 opcode, when vector length 
/// accessed by instruction is less than 512 bits and when it does not use 
//  the xmm or the mask registers or xmm/ymm registers with indexes higher than 15.
/// The pass applies code reduction on the generated code for AVX-512 instrs.
//
//===----------------------------------------------------------------------===//

#include "MCTargetDesc/X86BaseInfo.h"
#include "MCTargetDesc/X86InstComments.h"
#include "X86.h"
#include "X86InstrInfo.h"
#include "X86Subtarget.h"
#include "llvm/ADT/StringRef.h"
#include "llvm/CodeGen/MachineFunction.h"
#include "llvm/CodeGen/MachineFunctionPass.h"
#include "llvm/CodeGen/MachineInstr.h"
#include "llvm/CodeGen/MachineOperand.h"
#include "llvm/MC/MCInstrDesc.h"
#include "llvm/Pass.h"
#include <cassert>
#include <cstdint>

using namespace llvm;

// Including the generated EVEX2VEX tables.
struct X86EvexToVexCompressTableEntry {
  uint16_t EvexOpcode;
  uint16_t VexOpcode;

  bool operator<(const X86EvexToVexCompressTableEntry &RHS) const {
    return EvexOpcode < RHS.EvexOpcode;
  }

  friend bool operator<(const X86EvexToVexCompressTableEntry &TE,
                        unsigned Opc) {
    return TE.EvexOpcode < Opc;
  }
};
#include "X86GenEVEX2VEXTables.inc"

#define EVEX2VEX_DESC "Compressing EVEX instrs to VEX encoding when possible"
#define EVEX2VEX_NAME "x86-evex-to-vex-compress"

#define DEBUG_TYPE EVEX2VEX_NAME

namespace {

class EvexToVexInstPass : public MachineFunctionPass {

  /// For EVEX instructions that can be encoded using VEX encoding, replace
  /// them by the VEX encoding in order to reduce size.
  bool CompressEvexToVexImpl(MachineInstr &MI) const;

public:
  static char ID;

  EvexToVexInstPass() : MachineFunctionPass(ID) { }

  StringRef getPassName() const override { return EVEX2VEX_DESC; }

  /// Loop over all of the basic blocks, replacing EVEX instructions
  /// by equivalent VEX instructions when possible for reducing code size.
  bool runOnMachineFunction(MachineFunction &MF) override;

  // This pass runs after regalloc and doesn't support VReg operands.
  MachineFunctionProperties getRequiredProperties() const override {
    return MachineFunctionProperties().set(
        MachineFunctionProperties::Property::NoVRegs);
  }

private:
  /// Machine instruction info used throughout the class.
  const X86InstrInfo *TII = nullptr;

  const X86Subtarget *ST = nullptr;
};

} // end anonymous namespace

char EvexToVexInstPass::ID = 0;

bool EvexToVexInstPass::runOnMachineFunction(MachineFunction &MF) {
  TII = MF.getSubtarget<X86Subtarget>().getInstrInfo();

  ST = &MF.getSubtarget<X86Subtarget>();
  if (!ST->hasAVX512())
    return false;

  bool Changed = false;

  /// Go over all basic blocks in function and replace
  /// EVEX encoded instrs by VEX encoding when possible.
  for (MachineBasicBlock &MBB : MF) {

    // Traverse the basic block.
    for (MachineInstr &MI : MBB)
      Changed |= CompressEvexToVexImpl(MI);
  }

  return Changed;
}

static bool usesExtendedRegister(const MachineInstr &MI) {
  auto isHiRegIdx = [](unsigned Reg) {
    // Check for XMM register with indexes between 16 - 31.
    if (Reg >= X86::XMM16 && Reg <= X86::XMM31)
      return true;

    // Check for YMM register with indexes between 16 - 31.
    if (Reg >= X86::YMM16 && Reg <= X86::YMM31)
      return true;

    return false;
  };

  // Check that operands are not ZMM regs or
  // XMM/YMM regs with hi indexes between 16 - 31.
  for (const MachineOperand &MO : MI.explicit_operands()) {
    if (!MO.isReg())
      continue;

    Register Reg = MO.getReg();

    assert(!(Reg >= X86::ZMM0 && Reg <= X86::ZMM31) &&
           "ZMM instructions should not be in the EVEX->VEX tables");

    if (isHiRegIdx(Reg))
      return true;
  }

  return false;
}

// Do any custom cleanup needed to finalize the conversion.
static bool performCustomAdjustments(MachineInstr &MI, unsigned NewOpc,
                                     const X86Subtarget *ST) {
  (void)NewOpc;
  unsigned Opc = MI.getOpcode();
  switch (Opc) {
  case X86::VPDPBUSDSZ256m:
  case X86::VPDPBUSDSZ256r:
  case X86::VPDPBUSDSZ128m:
  case X86::VPDPBUSDSZ128r:
  case X86::VPDPBUSDZ256m:
  case X86::VPDPBUSDZ256r:
  case X86::VPDPBUSDZ128m:
  case X86::VPDPBUSDZ128r:
  case X86::VPDPWSSDSZ256m:
  case X86::VPDPWSSDSZ256r:
  case X86::VPDPWSSDSZ128m:
  case X86::VPDPWSSDSZ128r:
  case X86::VPDPWSSDZ256m:
  case X86::VPDPWSSDZ256r:
  case X86::VPDPWSSDZ128m:
  case X86::VPDPWSSDZ128r:
    // These can only VEX convert if AVXVNNI is enabled.
    return ST->hasAVXVNNI();
  case X86::VALIGNDZ128rri:
  case X86::VALIGNDZ128rmi:
  case X86::VALIGNQZ128rri:
  case X86::VALIGNQZ128rmi: {
    assert((NewOpc == X86::VPALIGNRrri || NewOpc == X86::VPALIGNRrmi) &&
           "Unexpected new opcode!");
    unsigned Scale = (Opc == X86::VALIGNQZ128rri ||
                      Opc == X86::VALIGNQZ128rmi) ? 8 : 4;
    MachineOperand &Imm = MI.getOperand(MI.getNumExplicitOperands()-1);
    Imm.setImm(Imm.getImm() * Scale);
    break;
  }
  case X86::VSHUFF32X4Z256rmi:
  case X86::VSHUFF32X4Z256rri:
  case X86::VSHUFF64X2Z256rmi:
  case X86::VSHUFF64X2Z256rri:
  case X86::VSHUFI32X4Z256rmi:
  case X86::VSHUFI32X4Z256rri:
  case X86::VSHUFI64X2Z256rmi:
  case X86::VSHUFI64X2Z256rri: {
    assert((NewOpc == X86::VPERM2F128rr || NewOpc == X86::VPERM2I128rr ||
            NewOpc == X86::VPERM2F128rm || NewOpc == X86::VPERM2I128rm) &&
           "Unexpected new opcode!");
    MachineOperand &Imm = MI.getOperand(MI.getNumExplicitOperands()-1);
    int64_t ImmVal = Imm.getImm();
    // Set bit 5, move bit 1 to bit 4, copy bit 0.
    Imm.setImm(0x20 | ((ImmVal & 2) << 3) | (ImmVal & 1));
    break;
  }
  case X86::VRNDSCALEPDZ128rri:
  case X86::VRNDSCALEPDZ128rmi:
  case X86::VRNDSCALEPSZ128rri:
  case X86::VRNDSCALEPSZ128rmi:
  case X86::VRNDSCALEPDZ256rri:
  case X86::VRNDSCALEPDZ256rmi:
  case X86::VRNDSCALEPSZ256rri:
  case X86::VRNDSCALEPSZ256rmi:
  case X86::VRNDSCALESDZr:
  case X86::VRNDSCALESDZm:
  case X86::VRNDSCALESSZr:
  case X86::VRNDSCALESSZm:
  case X86::VRNDSCALESDZr_Int:
  case X86::VRNDSCALESDZm_Int:
  case X86::VRNDSCALESSZr_Int:
  case X86::VRNDSCALESSZm_Int:
    const MachineOperand &Imm = MI.getOperand(MI.getNumExplicitOperands()-1);
    int64_t ImmVal = Imm.getImm();
    // Ensure that only bits 3:0 of the immediate are used.
    if ((ImmVal & 0xf) != ImmVal)
      return false;
    break;
  }

  return true;
}


// For EVEX instructions that can be encoded using VEX encoding
// replace them by the VEX encoding in order to reduce size.
bool EvexToVexInstPass::CompressEvexToVexImpl(MachineInstr &MI) const {
  // VEX format.
  // # of bytes: 0,2,3  1      1      0,1   0,1,2,4  0,1
  //  [Prefixes] [VEX]  OPCODE ModR/M [SIB] [DISP]  [IMM]
  //
  // EVEX format.
  //  # of bytes: 4    1      1      1      4       / 1         1
  //  [Prefixes]  EVEX Opcode ModR/M [SIB] [Disp32] / [Disp8*N] [Immediate]

  const MCInstrDesc &Desc = MI.getDesc();

  // Check for EVEX instructions only.
  if ((Desc.TSFlags & X86II::EncodingMask) != X86II::EVEX)
    return false;

  // Check for EVEX instructions with mask or broadcast as in these cases
  // the EVEX prefix is needed in order to carry this information
  // thus preventing the transformation to VEX encoding.
  if (Desc.TSFlags & (X86II::EVEX_K | X86II::EVEX_B))
    return false;

  // Check for EVEX instructions with L2 set. These instructions are 512-bits
  // and can't be converted to VEX.
  if (Desc.TSFlags & X86II::EVEX_L2)
    return false;

#ifndef NDEBUG
  // Make sure the tables are sorted.
  static std::atomic<bool> TableChecked(false);
  if (!TableChecked.load(std::memory_order_relaxed)) {
    assert(llvm::is_sorted(X86EvexToVex128CompressTable) &&
           "X86EvexToVex128CompressTable is not sorted!");
    assert(llvm::is_sorted(X86EvexToVex256CompressTable) &&
           "X86EvexToVex256CompressTable is not sorted!");
    TableChecked.store(true, std::memory_order_relaxed);
  }
#endif

  // Use the VEX.L bit to select the 128 or 256-bit table.
  ArrayRef<X86EvexToVexCompressTableEntry> Table =
    (Desc.TSFlags & X86II::VEX_L) ? makeArrayRef(X86EvexToVex256CompressTable)
                                  : makeArrayRef(X86EvexToVex128CompressTable);

  const auto *I = llvm::lower_bound(Table, MI.getOpcode());
  if (I == Table.end() || I->EvexOpcode != MI.getOpcode())
    return false;

  unsigned NewOpc = I->VexOpcode;

  if (usesExtendedRegister(MI))
    return false;

  if (!performCustomAdjustments(MI, NewOpc, ST))
    return false;

  MI.setDesc(TII->get(NewOpc));
  MI.setAsmPrinterFlag(X86::AC_EVEX_2_VEX);
  return true;
}

INITIALIZE_PASS(EvexToVexInstPass, EVEX2VEX_NAME, EVEX2VEX_DESC, false, false)

FunctionPass *llvm::createX86EvexToVexInsts() {
  return new EvexToVexInstPass();
}