• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 //===- X86EvexToVex.cpp ---------------------------------------------------===//
2 // Compress EVEX instructions to VEX encoding when possible to reduce code size
3 //
4 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
5 // See https://llvm.org/LICENSE.txt for license information.
6 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
7 //
8 //===----------------------------------------------------------------------===//
9 //
10 /// \file
11 /// This file defines the pass that goes over all AVX-512 instructions which
12 /// are encoded using the EVEX prefix and if possible replaces them by their
13 /// corresponding VEX encoding which is usually shorter by 2 bytes.
14 /// EVEX instructions may be encoded via the VEX prefix when the AVX-512
15 /// instruction has a corresponding AVX/AVX2 opcode, when vector length
16 /// accessed by instruction is less than 512 bits and when it does not use
17 //  the xmm or the mask registers or xmm/ymm registers with indexes higher than 15.
18 /// The pass applies code reduction on the generated code for AVX-512 instrs.
19 //
20 //===----------------------------------------------------------------------===//
21 
22 #include "MCTargetDesc/X86BaseInfo.h"
23 #include "MCTargetDesc/X86InstComments.h"
24 #include "X86.h"
25 #include "X86InstrInfo.h"
26 #include "X86Subtarget.h"
27 #include "llvm/ADT/StringRef.h"
28 #include "llvm/CodeGen/MachineFunction.h"
29 #include "llvm/CodeGen/MachineFunctionPass.h"
30 #include "llvm/CodeGen/MachineInstr.h"
31 #include "llvm/CodeGen/MachineOperand.h"
32 #include "llvm/MC/MCInstrDesc.h"
33 #include "llvm/Pass.h"
34 #include <cassert>
35 #include <cstdint>
36 
37 using namespace llvm;
38 
39 // Including the generated EVEX2VEX tables.
40 struct X86EvexToVexCompressTableEntry {
41   uint16_t EvexOpcode;
42   uint16_t VexOpcode;
43 
operator <X86EvexToVexCompressTableEntry44   bool operator<(const X86EvexToVexCompressTableEntry &RHS) const {
45     return EvexOpcode < RHS.EvexOpcode;
46   }
47 
operator <(const X86EvexToVexCompressTableEntry & TE,unsigned Opc)48   friend bool operator<(const X86EvexToVexCompressTableEntry &TE,
49                         unsigned Opc) {
50     return TE.EvexOpcode < Opc;
51   }
52 };
53 #include "X86GenEVEX2VEXTables.inc"
54 
55 #define EVEX2VEX_DESC "Compressing EVEX instrs to VEX encoding when possible"
56 #define EVEX2VEX_NAME "x86-evex-to-vex-compress"
57 
58 #define DEBUG_TYPE EVEX2VEX_NAME
59 
60 namespace {
61 
62 class EvexToVexInstPass : public MachineFunctionPass {
63 
64   /// For EVEX instructions that can be encoded using VEX encoding, replace
65   /// them by the VEX encoding in order to reduce size.
66   bool CompressEvexToVexImpl(MachineInstr &MI) const;
67 
68 public:
69   static char ID;
70 
EvexToVexInstPass()71   EvexToVexInstPass() : MachineFunctionPass(ID) { }
72 
getPassName() const73   StringRef getPassName() const override { return EVEX2VEX_DESC; }
74 
75   /// Loop over all of the basic blocks, replacing EVEX instructions
76   /// by equivalent VEX instructions when possible for reducing code size.
77   bool runOnMachineFunction(MachineFunction &MF) override;
78 
79   // This pass runs after regalloc and doesn't support VReg operands.
getRequiredProperties() const80   MachineFunctionProperties getRequiredProperties() const override {
81     return MachineFunctionProperties().set(
82         MachineFunctionProperties::Property::NoVRegs);
83   }
84 
85 private:
86   /// Machine instruction info used throughout the class.
87   const X86InstrInfo *TII = nullptr;
88 
89   const X86Subtarget *ST = nullptr;
90 };
91 
92 } // end anonymous namespace
93 
94 char EvexToVexInstPass::ID = 0;
95 
runOnMachineFunction(MachineFunction & MF)96 bool EvexToVexInstPass::runOnMachineFunction(MachineFunction &MF) {
97   TII = MF.getSubtarget<X86Subtarget>().getInstrInfo();
98 
99   ST = &MF.getSubtarget<X86Subtarget>();
100   if (!ST->hasAVX512())
101     return false;
102 
103   bool Changed = false;
104 
105   /// Go over all basic blocks in function and replace
106   /// EVEX encoded instrs by VEX encoding when possible.
107   for (MachineBasicBlock &MBB : MF) {
108 
109     // Traverse the basic block.
110     for (MachineInstr &MI : MBB)
111       Changed |= CompressEvexToVexImpl(MI);
112   }
113 
114   return Changed;
115 }
116 
usesExtendedRegister(const MachineInstr & MI)117 static bool usesExtendedRegister(const MachineInstr &MI) {
118   auto isHiRegIdx = [](unsigned Reg) {
119     // Check for XMM register with indexes between 16 - 31.
120     if (Reg >= X86::XMM16 && Reg <= X86::XMM31)
121       return true;
122 
123     // Check for YMM register with indexes between 16 - 31.
124     if (Reg >= X86::YMM16 && Reg <= X86::YMM31)
125       return true;
126 
127     return false;
128   };
129 
130   // Check that operands are not ZMM regs or
131   // XMM/YMM regs with hi indexes between 16 - 31.
132   for (const MachineOperand &MO : MI.explicit_operands()) {
133     if (!MO.isReg())
134       continue;
135 
136     Register Reg = MO.getReg();
137 
138     assert(!(Reg >= X86::ZMM0 && Reg <= X86::ZMM31) &&
139            "ZMM instructions should not be in the EVEX->VEX tables");
140 
141     if (isHiRegIdx(Reg))
142       return true;
143   }
144 
145   return false;
146 }
147 
148 // Do any custom cleanup needed to finalize the conversion.
performCustomAdjustments(MachineInstr & MI,unsigned NewOpc,const X86Subtarget * ST)149 static bool performCustomAdjustments(MachineInstr &MI, unsigned NewOpc,
150                                      const X86Subtarget *ST) {
151   (void)NewOpc;
152   unsigned Opc = MI.getOpcode();
153   switch (Opc) {
154   case X86::VPDPBUSDSZ256m:
155   case X86::VPDPBUSDSZ256r:
156   case X86::VPDPBUSDSZ128m:
157   case X86::VPDPBUSDSZ128r:
158   case X86::VPDPBUSDZ256m:
159   case X86::VPDPBUSDZ256r:
160   case X86::VPDPBUSDZ128m:
161   case X86::VPDPBUSDZ128r:
162   case X86::VPDPWSSDSZ256m:
163   case X86::VPDPWSSDSZ256r:
164   case X86::VPDPWSSDSZ128m:
165   case X86::VPDPWSSDSZ128r:
166   case X86::VPDPWSSDZ256m:
167   case X86::VPDPWSSDZ256r:
168   case X86::VPDPWSSDZ128m:
169   case X86::VPDPWSSDZ128r:
170     // These can only VEX convert if AVXVNNI is enabled.
171     return ST->hasAVXVNNI();
172   case X86::VALIGNDZ128rri:
173   case X86::VALIGNDZ128rmi:
174   case X86::VALIGNQZ128rri:
175   case X86::VALIGNQZ128rmi: {
176     assert((NewOpc == X86::VPALIGNRrri || NewOpc == X86::VPALIGNRrmi) &&
177            "Unexpected new opcode!");
178     unsigned Scale = (Opc == X86::VALIGNQZ128rri ||
179                       Opc == X86::VALIGNQZ128rmi) ? 8 : 4;
180     MachineOperand &Imm = MI.getOperand(MI.getNumExplicitOperands()-1);
181     Imm.setImm(Imm.getImm() * Scale);
182     break;
183   }
184   case X86::VSHUFF32X4Z256rmi:
185   case X86::VSHUFF32X4Z256rri:
186   case X86::VSHUFF64X2Z256rmi:
187   case X86::VSHUFF64X2Z256rri:
188   case X86::VSHUFI32X4Z256rmi:
189   case X86::VSHUFI32X4Z256rri:
190   case X86::VSHUFI64X2Z256rmi:
191   case X86::VSHUFI64X2Z256rri: {
192     assert((NewOpc == X86::VPERM2F128rr || NewOpc == X86::VPERM2I128rr ||
193             NewOpc == X86::VPERM2F128rm || NewOpc == X86::VPERM2I128rm) &&
194            "Unexpected new opcode!");
195     MachineOperand &Imm = MI.getOperand(MI.getNumExplicitOperands()-1);
196     int64_t ImmVal = Imm.getImm();
197     // Set bit 5, move bit 1 to bit 4, copy bit 0.
198     Imm.setImm(0x20 | ((ImmVal & 2) << 3) | (ImmVal & 1));
199     break;
200   }
201   case X86::VRNDSCALEPDZ128rri:
202   case X86::VRNDSCALEPDZ128rmi:
203   case X86::VRNDSCALEPSZ128rri:
204   case X86::VRNDSCALEPSZ128rmi:
205   case X86::VRNDSCALEPDZ256rri:
206   case X86::VRNDSCALEPDZ256rmi:
207   case X86::VRNDSCALEPSZ256rri:
208   case X86::VRNDSCALEPSZ256rmi:
209   case X86::VRNDSCALESDZr:
210   case X86::VRNDSCALESDZm:
211   case X86::VRNDSCALESSZr:
212   case X86::VRNDSCALESSZm:
213   case X86::VRNDSCALESDZr_Int:
214   case X86::VRNDSCALESDZm_Int:
215   case X86::VRNDSCALESSZr_Int:
216   case X86::VRNDSCALESSZm_Int:
217     const MachineOperand &Imm = MI.getOperand(MI.getNumExplicitOperands()-1);
218     int64_t ImmVal = Imm.getImm();
219     // Ensure that only bits 3:0 of the immediate are used.
220     if ((ImmVal & 0xf) != ImmVal)
221       return false;
222     break;
223   }
224 
225   return true;
226 }
227 
228 
229 // For EVEX instructions that can be encoded using VEX encoding
230 // replace them by the VEX encoding in order to reduce size.
CompressEvexToVexImpl(MachineInstr & MI) const231 bool EvexToVexInstPass::CompressEvexToVexImpl(MachineInstr &MI) const {
232   // VEX format.
233   // # of bytes: 0,2,3  1      1      0,1   0,1,2,4  0,1
234   //  [Prefixes] [VEX]  OPCODE ModR/M [SIB] [DISP]  [IMM]
235   //
236   // EVEX format.
237   //  # of bytes: 4    1      1      1      4       / 1         1
238   //  [Prefixes]  EVEX Opcode ModR/M [SIB] [Disp32] / [Disp8*N] [Immediate]
239 
240   const MCInstrDesc &Desc = MI.getDesc();
241 
242   // Check for EVEX instructions only.
243   if ((Desc.TSFlags & X86II::EncodingMask) != X86II::EVEX)
244     return false;
245 
246   // Check for EVEX instructions with mask or broadcast as in these cases
247   // the EVEX prefix is needed in order to carry this information
248   // thus preventing the transformation to VEX encoding.
249   if (Desc.TSFlags & (X86II::EVEX_K | X86II::EVEX_B))
250     return false;
251 
252   // Check for EVEX instructions with L2 set. These instructions are 512-bits
253   // and can't be converted to VEX.
254   if (Desc.TSFlags & X86II::EVEX_L2)
255     return false;
256 
257 #ifndef NDEBUG
258   // Make sure the tables are sorted.
259   static std::atomic<bool> TableChecked(false);
260   if (!TableChecked.load(std::memory_order_relaxed)) {
261     assert(llvm::is_sorted(X86EvexToVex128CompressTable) &&
262            "X86EvexToVex128CompressTable is not sorted!");
263     assert(llvm::is_sorted(X86EvexToVex256CompressTable) &&
264            "X86EvexToVex256CompressTable is not sorted!");
265     TableChecked.store(true, std::memory_order_relaxed);
266   }
267 #endif
268 
269   // Use the VEX.L bit to select the 128 or 256-bit table.
270   ArrayRef<X86EvexToVexCompressTableEntry> Table =
271     (Desc.TSFlags & X86II::VEX_L) ? makeArrayRef(X86EvexToVex256CompressTable)
272                                   : makeArrayRef(X86EvexToVex128CompressTable);
273 
274   const auto *I = llvm::lower_bound(Table, MI.getOpcode());
275   if (I == Table.end() || I->EvexOpcode != MI.getOpcode())
276     return false;
277 
278   unsigned NewOpc = I->VexOpcode;
279 
280   if (usesExtendedRegister(MI))
281     return false;
282 
283   if (!performCustomAdjustments(MI, NewOpc, ST))
284     return false;
285 
286   MI.setDesc(TII->get(NewOpc));
287   MI.setAsmPrinterFlag(X86::AC_EVEX_2_VEX);
288   return true;
289 }
290 
INITIALIZE_PASS(EvexToVexInstPass,EVEX2VEX_NAME,EVEX2VEX_DESC,false,false)291 INITIALIZE_PASS(EvexToVexInstPass, EVEX2VEX_NAME, EVEX2VEX_DESC, false, false)
292 
293 FunctionPass *llvm::createX86EvexToVexInsts() {
294   return new EvexToVexInstPass();
295 }
296