1 | //===- X86CompressEVEX.cpp ------------------------------------------------===// |
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 pass compresses instructions from EVEX space to legacy/VEX/EVEX space |
10 | // when possible in order to reduce code size or facilitate HW decoding. |
11 | // |
12 | // Possible compression: |
13 | // a. AVX512 instruction (EVEX) -> AVX instruction (VEX) |
14 | // b. Promoted instruction (EVEX) -> pre-promotion instruction (legacy/VEX) |
15 | // c. NDD (EVEX) -> non-NDD (legacy) |
16 | // d. NF_ND (EVEX) -> NF (EVEX) |
17 | // e. NonNF (EVEX) -> NF (EVEX) |
18 | // |
19 | // Compression a, b and c can always reduce code size, with some exceptions |
20 | // such as promoted 16-bit CRC32 which is as long as the legacy version. |
21 | // |
22 | // legacy: |
23 | // crc32w %si, %eax ## encoding: [0x66,0xf2,0x0f,0x38,0xf1,0xc6] |
24 | // promoted: |
25 | // crc32w %si, %eax ## encoding: [0x62,0xf4,0x7d,0x08,0xf1,0xc6] |
26 | // |
27 | // From performance perspective, these should be same (same uops and same EXE |
28 | // ports). From a FMV perspective, an older legacy encoding is preferred b/c it |
29 | // can execute in more places (broader HW install base). So we will still do |
30 | // the compression. |
31 | // |
32 | // Compression d can help hardware decode (HW may skip reading the NDD |
33 | // register) although the instruction length remains unchanged. |
34 | // |
35 | // Compression e can help hardware skip updating EFLAGS although the instruction |
36 | // length remains unchanged. |
37 | //===----------------------------------------------------------------------===// |
38 | |
39 | #include "MCTargetDesc/X86BaseInfo.h" |
40 | #include "MCTargetDesc/X86InstComments.h" |
41 | #include "X86.h" |
42 | #include "X86InstrInfo.h" |
43 | #include "X86Subtarget.h" |
44 | #include "llvm/ADT/StringRef.h" |
45 | #include "llvm/CodeGen/MachineFunction.h" |
46 | #include "llvm/CodeGen/MachineFunctionPass.h" |
47 | #include "llvm/CodeGen/MachineInstr.h" |
48 | #include "llvm/CodeGen/MachineOperand.h" |
49 | #include "llvm/MC/MCInstrDesc.h" |
50 | #include "llvm/Pass.h" |
51 | #include <atomic> |
52 | #include <cassert> |
53 | #include <cstdint> |
54 | |
55 | using namespace llvm; |
56 | |
57 | #define COMP_EVEX_DESC "Compressing EVEX instrs when possible" |
58 | #define COMP_EVEX_NAME "x86-compress-evex" |
59 | |
60 | #define DEBUG_TYPE COMP_EVEX_NAME |
61 | |
62 | namespace { |
63 | // Including the generated EVEX compression tables. |
64 | #define GET_X86_COMPRESS_EVEX_TABLE |
65 | #include "X86GenInstrMapping.inc" |
66 | |
67 | class CompressEVEXPass : public MachineFunctionPass { |
68 | public: |
69 | static char ID; |
70 | CompressEVEXPass() : MachineFunctionPass(ID) {} |
71 | StringRef getPassName() const override { return COMP_EVEX_DESC; } |
72 | |
73 | bool runOnMachineFunction(MachineFunction &MF) override; |
74 | |
75 | // This pass runs after regalloc and doesn't support VReg operands. |
76 | MachineFunctionProperties getRequiredProperties() const override { |
77 | return MachineFunctionProperties().set( |
78 | MachineFunctionProperties::Property::NoVRegs); |
79 | } |
80 | }; |
81 | |
82 | } // end anonymous namespace |
83 | |
84 | char CompressEVEXPass::ID = 0; |
85 | |
86 | static bool usesExtendedRegister(const MachineInstr &MI) { |
87 | auto isHiRegIdx = [](unsigned Reg) { |
88 | // Check for XMM register with indexes between 16 - 31. |
89 | if (Reg >= X86::XMM16 && Reg <= X86::XMM31) |
90 | return true; |
91 | // Check for YMM register with indexes between 16 - 31. |
92 | if (Reg >= X86::YMM16 && Reg <= X86::YMM31) |
93 | return true; |
94 | // Check for GPR with indexes between 16 - 31. |
95 | if (X86II::isApxExtendedReg(RegNo: Reg)) |
96 | return true; |
97 | return false; |
98 | }; |
99 | |
100 | // Check that operands are not ZMM regs or |
101 | // XMM/YMM regs with hi indexes between 16 - 31. |
102 | for (const MachineOperand &MO : MI.explicit_operands()) { |
103 | if (!MO.isReg()) |
104 | continue; |
105 | |
106 | Register Reg = MO.getReg(); |
107 | assert(!X86II::isZMMReg(Reg) && |
108 | "ZMM instructions should not be in the EVEX->VEX tables" ); |
109 | if (isHiRegIdx(Reg)) |
110 | return true; |
111 | } |
112 | |
113 | return false; |
114 | } |
115 | |
116 | // Do any custom cleanup needed to finalize the conversion. |
117 | static bool performCustomAdjustments(MachineInstr &MI, unsigned NewOpc) { |
118 | (void)NewOpc; |
119 | unsigned Opc = MI.getOpcode(); |
120 | switch (Opc) { |
121 | case X86::VALIGNDZ128rri: |
122 | case X86::VALIGNDZ128rmi: |
123 | case X86::VALIGNQZ128rri: |
124 | case X86::VALIGNQZ128rmi: { |
125 | assert((NewOpc == X86::VPALIGNRrri || NewOpc == X86::VPALIGNRrmi) && |
126 | "Unexpected new opcode!" ); |
127 | unsigned Scale = |
128 | (Opc == X86::VALIGNQZ128rri || Opc == X86::VALIGNQZ128rmi) ? 8 : 4; |
129 | MachineOperand &Imm = MI.getOperand(i: MI.getNumExplicitOperands() - 1); |
130 | Imm.setImm(Imm.getImm() * Scale); |
131 | break; |
132 | } |
133 | case X86::VSHUFF32X4Z256rmi: |
134 | case X86::VSHUFF32X4Z256rri: |
135 | case X86::VSHUFF64X2Z256rmi: |
136 | case X86::VSHUFF64X2Z256rri: |
137 | case X86::VSHUFI32X4Z256rmi: |
138 | case X86::VSHUFI32X4Z256rri: |
139 | case X86::VSHUFI64X2Z256rmi: |
140 | case X86::VSHUFI64X2Z256rri: { |
141 | assert((NewOpc == X86::VPERM2F128rr || NewOpc == X86::VPERM2I128rr || |
142 | NewOpc == X86::VPERM2F128rm || NewOpc == X86::VPERM2I128rm) && |
143 | "Unexpected new opcode!" ); |
144 | MachineOperand &Imm = MI.getOperand(i: MI.getNumExplicitOperands() - 1); |
145 | int64_t ImmVal = Imm.getImm(); |
146 | // Set bit 5, move bit 1 to bit 4, copy bit 0. |
147 | Imm.setImm(0x20 | ((ImmVal & 2) << 3) | (ImmVal & 1)); |
148 | break; |
149 | } |
150 | case X86::VRNDSCALEPDZ128rri: |
151 | case X86::VRNDSCALEPDZ128rmi: |
152 | case X86::VRNDSCALEPSZ128rri: |
153 | case X86::VRNDSCALEPSZ128rmi: |
154 | case X86::VRNDSCALEPDZ256rri: |
155 | case X86::VRNDSCALEPDZ256rmi: |
156 | case X86::VRNDSCALEPSZ256rri: |
157 | case X86::VRNDSCALEPSZ256rmi: |
158 | case X86::VRNDSCALESDZr: |
159 | case X86::VRNDSCALESDZm: |
160 | case X86::VRNDSCALESSZr: |
161 | case X86::VRNDSCALESSZm: |
162 | case X86::VRNDSCALESDZr_Int: |
163 | case X86::VRNDSCALESDZm_Int: |
164 | case X86::VRNDSCALESSZr_Int: |
165 | case X86::VRNDSCALESSZm_Int: |
166 | const MachineOperand &Imm = MI.getOperand(i: MI.getNumExplicitOperands() - 1); |
167 | int64_t ImmVal = Imm.getImm(); |
168 | // Ensure that only bits 3:0 of the immediate are used. |
169 | if ((ImmVal & 0xf) != ImmVal) |
170 | return false; |
171 | break; |
172 | } |
173 | |
174 | return true; |
175 | } |
176 | |
177 | static bool CompressEVEXImpl(MachineInstr &MI, const X86Subtarget &ST) { |
178 | uint64_t TSFlags = MI.getDesc().TSFlags; |
179 | |
180 | // Check for EVEX instructions only. |
181 | if ((TSFlags & X86II::EncodingMask) != X86II::EVEX) |
182 | return false; |
183 | |
184 | // Instructions with mask or 512-bit vector can't be converted to VEX. |
185 | if (TSFlags & (X86II::EVEX_K | X86II::EVEX_L2)) |
186 | return false; |
187 | |
188 | auto IsRedundantNewDataDest = [&](unsigned &Opc) { |
189 | // $rbx = ADD64rr_ND $rbx, $rax / $rbx = ADD64rr_ND $rax, $rbx |
190 | // -> |
191 | // $rbx = ADD64rr $rbx, $rax |
192 | const MCInstrDesc &Desc = MI.getDesc(); |
193 | Register Reg0 = MI.getOperand(i: 0).getReg(); |
194 | const MachineOperand &Op1 = MI.getOperand(i: 1); |
195 | if (!Op1.isReg() || X86::getFirstAddrOperandIdx(MI) == 1 || |
196 | X86::isCFCMOVCC(Opcode: MI.getOpcode())) |
197 | return false; |
198 | Register Reg1 = Op1.getReg(); |
199 | if (Reg1 == Reg0) |
200 | return true; |
201 | |
202 | // Op1 and Op2 may be commutable for ND instructions. |
203 | if (!Desc.isCommutable() || Desc.getNumOperands() < 3 || |
204 | !MI.getOperand(i: 2).isReg() || MI.getOperand(i: 2).getReg() != Reg0) |
205 | return false; |
206 | // Opcode may change after commute, e.g. SHRD -> SHLD |
207 | ST.getInstrInfo()->commuteInstruction(MI, NewMI: false, OpIdx1: 1, OpIdx2: 2); |
208 | Opc = MI.getOpcode(); |
209 | return true; |
210 | }; |
211 | |
212 | // EVEX_B has several meanings. |
213 | // AVX512: |
214 | // register form: rounding control or SAE |
215 | // memory form: broadcast |
216 | // |
217 | // APX: |
218 | // MAP4: NDD |
219 | // |
220 | // For AVX512 cases, EVEX prefix is needed in order to carry this information |
221 | // thus preventing the transformation to VEX encoding. |
222 | bool IsND = X86II::hasNewDataDest(TSFlags); |
223 | if (TSFlags & X86II::EVEX_B && !IsND) |
224 | return false; |
225 | unsigned Opc = MI.getOpcode(); |
226 | // MOVBE*rr is special because it has semantic of NDD but not set EVEX_B. |
227 | bool IsNDLike = IsND || Opc == X86::MOVBE32rr || Opc == X86::MOVBE64rr; |
228 | bool IsRedundantNDD = IsNDLike ? IsRedundantNewDataDest(Opc) : false; |
229 | |
230 | auto GetCompressedOpc = [&](unsigned Opc) -> unsigned { |
231 | ArrayRef<X86TableEntry> Table = ArrayRef(X86CompressEVEXTable); |
232 | const auto I = llvm::lower_bound(Range&: Table, Value&: Opc); |
233 | if (I == Table.end() || I->OldOpc != Opc) |
234 | return 0; |
235 | |
236 | if (usesExtendedRegister(MI) || !checkPredicate(Opc: I->NewOpc, Subtarget: &ST) || |
237 | !performCustomAdjustments(MI, NewOpc: I->NewOpc)) |
238 | return 0; |
239 | return I->NewOpc; |
240 | }; |
241 | // NonNF -> NF only if it's not a compressible NDD instruction and eflags is |
242 | // dead. |
243 | unsigned NewOpc = IsRedundantNDD |
244 | ? X86::getNonNDVariant(Opc) |
245 | : ((IsNDLike && ST.hasNF() && |
246 | MI.registerDefIsDead(Reg: X86::EFLAGS, /*TRI=*/nullptr)) |
247 | ? X86::getNFVariant(Opc) |
248 | : GetCompressedOpc(Opc)); |
249 | |
250 | if (!NewOpc) |
251 | return false; |
252 | |
253 | const MCInstrDesc &NewDesc = ST.getInstrInfo()->get(Opcode: NewOpc); |
254 | MI.setDesc(NewDesc); |
255 | unsigned ; |
256 | switch (NewDesc.TSFlags & X86II::EncodingMask) { |
257 | case X86II::LEGACY: |
258 | AsmComment = X86::AC_EVEX_2_LEGACY; |
259 | break; |
260 | case X86II::VEX: |
261 | AsmComment = X86::AC_EVEX_2_VEX; |
262 | break; |
263 | case X86II::EVEX: |
264 | AsmComment = X86::AC_EVEX_2_EVEX; |
265 | assert(IsND && (NewDesc.TSFlags & X86II::EVEX_NF) && |
266 | "Unknown EVEX2EVEX compression" ); |
267 | break; |
268 | default: |
269 | llvm_unreachable("Unknown EVEX compression" ); |
270 | } |
271 | MI.setAsmPrinterFlag(AsmComment); |
272 | if (IsRedundantNDD) |
273 | MI.tieOperands(DefIdx: 0, UseIdx: 1); |
274 | |
275 | return true; |
276 | } |
277 | |
278 | bool CompressEVEXPass::runOnMachineFunction(MachineFunction &MF) { |
279 | #ifndef NDEBUG |
280 | // Make sure the tables are sorted. |
281 | static std::atomic<bool> TableChecked(false); |
282 | if (!TableChecked.load(std::memory_order_relaxed)) { |
283 | assert(llvm::is_sorted(X86CompressEVEXTable) && |
284 | "X86CompressEVEXTable is not sorted!" ); |
285 | TableChecked.store(true, std::memory_order_relaxed); |
286 | } |
287 | #endif |
288 | const X86Subtarget &ST = MF.getSubtarget<X86Subtarget>(); |
289 | if (!ST.hasAVX512() && !ST.hasEGPR() && !ST.hasNDD()) |
290 | return false; |
291 | |
292 | bool Changed = false; |
293 | |
294 | for (MachineBasicBlock &MBB : MF) { |
295 | // Traverse the basic block. |
296 | for (MachineInstr &MI : MBB) |
297 | Changed |= CompressEVEXImpl(MI, ST); |
298 | } |
299 | |
300 | return Changed; |
301 | } |
302 | |
303 | INITIALIZE_PASS(CompressEVEXPass, COMP_EVEX_NAME, COMP_EVEX_DESC, false, false) |
304 | |
305 | FunctionPass *llvm::createX86CompressEVEXPass() { |
306 | return new CompressEVEXPass(); |
307 | } |
308 | |