1 | //===- utils/TableGen/X86FoldTablesEmitter.cpp - X86 backend-*- C++ -*-===// |
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 tablegen backend is responsible for emitting the memory fold tables of |
10 | // the X86 backend instructions. |
11 | // |
12 | //===----------------------------------------------------------------------===// |
13 | |
14 | #include "Common/CodeGenInstruction.h" |
15 | #include "Common/CodeGenTarget.h" |
16 | #include "X86RecognizableInstr.h" |
17 | #include "llvm/ADT/StringSwitch.h" |
18 | #include "llvm/Support/X86FoldTablesUtils.h" |
19 | #include "llvm/TableGen/Record.h" |
20 | #include "llvm/TableGen/TableGenBackend.h" |
21 | #include <set> |
22 | |
23 | using namespace llvm; |
24 | using namespace X86Disassembler; |
25 | |
26 | namespace { |
27 | // Represents an entry in the manual mapped instructions set. |
28 | struct ManualMapEntry { |
29 | const char *RegInstStr; |
30 | const char *MemInstStr; |
31 | uint16_t Strategy; |
32 | }; |
33 | |
34 | // List of instructions requiring explicitly aligned memory. |
35 | const char *ExplicitAlign[] = {"MOVDQA" , "MOVAPS" , "MOVAPD" , "MOVNTPS" , |
36 | "MOVNTPD" , "MOVNTDQ" , "MOVNTDQA" }; |
37 | |
38 | // List of instructions NOT requiring explicit memory alignment. |
39 | const char *ExplicitUnalign[] = {"MOVDQU" , "MOVUPS" , "MOVUPD" , |
40 | "PCMPESTRM" , "PCMPESTRI" , "PCMPISTRM" , |
41 | "PCMPISTRI" }; |
42 | |
43 | const ManualMapEntry ManualMapSet[] = { |
44 | #define ENTRY(REG, MEM, FLAGS) {#REG, #MEM, FLAGS}, |
45 | #include "X86ManualFoldTables.def" |
46 | }; |
47 | |
48 | const std::set<StringRef> NoFoldSet = { |
49 | #define NOFOLD(INSN) #INSN, |
50 | #include "X86ManualFoldTables.def" |
51 | }; |
52 | |
53 | static bool isExplicitAlign(const CodeGenInstruction *Inst) { |
54 | return any_of(Range&: ExplicitAlign, P: [Inst](const char *InstStr) { |
55 | return Inst->TheDef->getName().contains(Other: InstStr); |
56 | }); |
57 | } |
58 | |
59 | static bool isExplicitUnalign(const CodeGenInstruction *Inst) { |
60 | return any_of(Range&: ExplicitUnalign, P: [Inst](const char *InstStr) { |
61 | return Inst->TheDef->getName().contains(Other: InstStr); |
62 | }); |
63 | } |
64 | |
65 | class X86FoldTablesEmitter { |
66 | RecordKeeper &Records; |
67 | CodeGenTarget Target; |
68 | |
69 | // Represents an entry in the folding table |
70 | class X86FoldTableEntry { |
71 | const CodeGenInstruction *RegInst; |
72 | const CodeGenInstruction *MemInst; |
73 | |
74 | public: |
75 | bool NoReverse = false; |
76 | bool NoForward = false; |
77 | bool FoldLoad = false; |
78 | bool FoldStore = false; |
79 | enum BcastType { |
80 | BCAST_NONE, |
81 | BCAST_W, |
82 | BCAST_D, |
83 | BCAST_Q, |
84 | BCAST_SS, |
85 | BCAST_SD, |
86 | BCAST_SH, |
87 | }; |
88 | BcastType BroadcastKind = BCAST_NONE; |
89 | |
90 | Align Alignment; |
91 | |
92 | X86FoldTableEntry() = default; |
93 | X86FoldTableEntry(const CodeGenInstruction *RegInst, |
94 | const CodeGenInstruction *MemInst) |
95 | : RegInst(RegInst), MemInst(MemInst) {} |
96 | |
97 | void print(raw_ostream &OS) const { |
98 | OS.indent(NumSpaces: 2); |
99 | OS << "{X86::" << RegInst->TheDef->getName() << ", " ; |
100 | OS << "X86::" << MemInst->TheDef->getName() << ", " ; |
101 | |
102 | std::string Attrs; |
103 | if (FoldLoad) |
104 | Attrs += "TB_FOLDED_LOAD|" ; |
105 | if (FoldStore) |
106 | Attrs += "TB_FOLDED_STORE|" ; |
107 | if (NoReverse) |
108 | Attrs += "TB_NO_REVERSE|" ; |
109 | if (NoForward) |
110 | Attrs += "TB_NO_FORWARD|" ; |
111 | if (Alignment != Align(1)) |
112 | Attrs += "TB_ALIGN_" + std::to_string(val: Alignment.value()) + "|" ; |
113 | switch (BroadcastKind) { |
114 | case BCAST_NONE: |
115 | break; |
116 | case BCAST_W: |
117 | Attrs += "TB_BCAST_W|" ; |
118 | break; |
119 | case BCAST_D: |
120 | Attrs += "TB_BCAST_D|" ; |
121 | break; |
122 | case BCAST_Q: |
123 | Attrs += "TB_BCAST_Q|" ; |
124 | break; |
125 | case BCAST_SS: |
126 | Attrs += "TB_BCAST_SS|" ; |
127 | break; |
128 | case BCAST_SD: |
129 | Attrs += "TB_BCAST_SD|" ; |
130 | break; |
131 | case BCAST_SH: |
132 | Attrs += "TB_BCAST_SH|" ; |
133 | break; |
134 | } |
135 | |
136 | StringRef SimplifiedAttrs = StringRef(Attrs).rtrim(Chars: "|" ); |
137 | if (SimplifiedAttrs.empty()) |
138 | SimplifiedAttrs = "0" ; |
139 | |
140 | OS << SimplifiedAttrs << "},\n" ; |
141 | } |
142 | |
143 | #ifndef NDEBUG |
144 | // Check that Uses and Defs are same after memory fold. |
145 | void checkCorrectness() const { |
146 | auto &RegInstRec = *RegInst->TheDef; |
147 | auto &MemInstRec = *MemInst->TheDef; |
148 | auto ListOfUsesReg = RegInstRec.getValueAsListOfDefs("Uses" ); |
149 | auto ListOfUsesMem = MemInstRec.getValueAsListOfDefs("Uses" ); |
150 | auto ListOfDefsReg = RegInstRec.getValueAsListOfDefs("Defs" ); |
151 | auto ListOfDefsMem = MemInstRec.getValueAsListOfDefs("Defs" ); |
152 | if (ListOfUsesReg != ListOfUsesMem || ListOfDefsReg != ListOfDefsMem) |
153 | report_fatal_error("Uses/Defs couldn't be changed after folding " + |
154 | RegInstRec.getName() + " to " + |
155 | MemInstRec.getName()); |
156 | } |
157 | #endif |
158 | }; |
159 | |
160 | // NOTE: We check the fold tables are sorted in X86InstrFoldTables.cpp by the |
161 | // enum of the instruction, which is computed in |
162 | // CodeGenTarget::ComputeInstrsByEnum. So we should use the same comparator |
163 | // here. |
164 | // FIXME: Could we share the code with CodeGenTarget::ComputeInstrsByEnum? |
165 | struct CompareInstrsByEnum { |
166 | bool operator()(const CodeGenInstruction *LHS, |
167 | const CodeGenInstruction *RHS) const { |
168 | assert(LHS && RHS && "LHS and RHS shouldn't be nullptr" ); |
169 | const auto &D1 = *LHS->TheDef; |
170 | const auto &D2 = *RHS->TheDef; |
171 | return std::tuple(!D1.getValueAsBit(FieldName: "isPseudo" ), D1.getName()) < |
172 | std::tuple(!D2.getValueAsBit(FieldName: "isPseudo" ), D2.getName()); |
173 | } |
174 | }; |
175 | |
176 | typedef std::map<const CodeGenInstruction *, X86FoldTableEntry, |
177 | CompareInstrsByEnum> |
178 | FoldTable; |
179 | // Table2Addr - Holds instructions which their memory form performs |
180 | // load+store. |
181 | // |
182 | // Table#i - Holds instructions which the their memory form |
183 | // performs a load OR a store, and their #i'th operand is folded. |
184 | // |
185 | // BroadcastTable#i - Holds instructions which the their memory form performs |
186 | // a broadcast load and their #i'th operand is folded. |
187 | FoldTable Table2Addr; |
188 | FoldTable Table0; |
189 | FoldTable Table1; |
190 | FoldTable Table2; |
191 | FoldTable Table3; |
192 | FoldTable Table4; |
193 | FoldTable BroadcastTable1; |
194 | FoldTable BroadcastTable2; |
195 | FoldTable BroadcastTable3; |
196 | FoldTable BroadcastTable4; |
197 | |
198 | public: |
199 | X86FoldTablesEmitter(RecordKeeper &R) : Records(R), Target(R) {} |
200 | |
201 | // run - Generate the 6 X86 memory fold tables. |
202 | void run(raw_ostream &OS); |
203 | |
204 | private: |
205 | // Decides to which table to add the entry with the given instructions. |
206 | // S sets the strategy of adding the TB_NO_REVERSE flag. |
207 | void updateTables(const CodeGenInstruction *RegInst, |
208 | const CodeGenInstruction *MemInst, uint16_t S = 0, |
209 | bool IsManual = false, bool IsBroadcast = false); |
210 | |
211 | // Generates X86FoldTableEntry with the given instructions and fill it with |
212 | // the appropriate flags, then adds it to a memory fold table. |
213 | void addEntryWithFlags(FoldTable &Table, const CodeGenInstruction *RegInst, |
214 | const CodeGenInstruction *MemInst, uint16_t S, |
215 | unsigned FoldedIdx, bool IsManual); |
216 | // Generates X86FoldTableEntry with the given instructions and adds it to a |
217 | // broadcast table. |
218 | void addBroadcastEntry(FoldTable &Table, const CodeGenInstruction *RegInst, |
219 | const CodeGenInstruction *MemInst); |
220 | |
221 | // Print the given table as a static const C++ array of type |
222 | // X86FoldTableEntry. |
223 | void printTable(const FoldTable &Table, StringRef TableName, |
224 | raw_ostream &OS) { |
225 | OS << "static const X86FoldTableEntry " << TableName << "[] = {\n" ; |
226 | |
227 | for (auto &E : Table) |
228 | E.second.print(OS); |
229 | |
230 | OS << "};\n\n" ; |
231 | } |
232 | }; |
233 | |
234 | // Return true if one of the instruction's operands is a RST register class |
235 | static bool hasRSTRegClass(const CodeGenInstruction *Inst) { |
236 | return any_of(Range: Inst->Operands, P: [](const CGIOperandList::OperandInfo &OpIn) { |
237 | return OpIn.Rec->getName() == "RST" || OpIn.Rec->getName() == "RSTi" ; |
238 | }); |
239 | } |
240 | |
241 | // Return true if one of the instruction's operands is a ptr_rc_tailcall |
242 | static bool hasPtrTailcallRegClass(const CodeGenInstruction *Inst) { |
243 | return any_of(Range: Inst->Operands, P: [](const CGIOperandList::OperandInfo &OpIn) { |
244 | return OpIn.Rec->getName() == "ptr_rc_tailcall" ; |
245 | }); |
246 | } |
247 | |
248 | static uint8_t byteFromBitsInit(const BitsInit *B) { |
249 | unsigned N = B->getNumBits(); |
250 | assert(N <= 8 && "Field is too large for uint8_t!" ); |
251 | |
252 | uint8_t Value = 0; |
253 | for (unsigned I = 0; I != N; ++I) { |
254 | BitInit *Bit = cast<BitInit>(Val: B->getBit(Bit: I)); |
255 | Value |= Bit->getValue() << I; |
256 | } |
257 | return Value; |
258 | } |
259 | |
260 | static bool mayFoldFromForm(uint8_t Form) { |
261 | switch (Form) { |
262 | default: |
263 | return Form >= X86Local::MRM0r && Form <= X86Local::MRM7r; |
264 | case X86Local::MRMXr: |
265 | case X86Local::MRMXrCC: |
266 | case X86Local::MRMDestReg: |
267 | case X86Local::MRMSrcReg: |
268 | case X86Local::MRMSrcReg4VOp3: |
269 | case X86Local::MRMSrcRegOp4: |
270 | case X86Local::MRMSrcRegCC: |
271 | return true; |
272 | } |
273 | } |
274 | |
275 | static bool mayFoldToForm(uint8_t Form) { |
276 | switch (Form) { |
277 | default: |
278 | return Form >= X86Local::MRM0m && Form <= X86Local::MRM7m; |
279 | case X86Local::MRMXm: |
280 | case X86Local::MRMXmCC: |
281 | case X86Local::MRMDestMem: |
282 | case X86Local::MRMSrcMem: |
283 | case X86Local::MRMSrcMem4VOp3: |
284 | case X86Local::MRMSrcMemOp4: |
285 | case X86Local::MRMSrcMemCC: |
286 | return true; |
287 | } |
288 | } |
289 | |
290 | static bool mayFoldFromLeftToRight(uint8_t LHS, uint8_t RHS) { |
291 | switch (LHS) { |
292 | default: |
293 | llvm_unreachable("Unexpected Form!" ); |
294 | case X86Local::MRM0r: |
295 | return RHS == X86Local::MRM0m; |
296 | case X86Local::MRM1r: |
297 | return RHS == X86Local::MRM1m; |
298 | case X86Local::MRM2r: |
299 | return RHS == X86Local::MRM2m; |
300 | case X86Local::MRM3r: |
301 | return RHS == X86Local::MRM3m; |
302 | case X86Local::MRM4r: |
303 | return RHS == X86Local::MRM4m; |
304 | case X86Local::MRM5r: |
305 | return RHS == X86Local::MRM5m; |
306 | case X86Local::MRM6r: |
307 | return RHS == X86Local::MRM6m; |
308 | case X86Local::MRM7r: |
309 | return RHS == X86Local::MRM7m; |
310 | case X86Local::MRMXr: |
311 | return RHS == X86Local::MRMXm; |
312 | case X86Local::MRMXrCC: |
313 | return RHS == X86Local::MRMXmCC; |
314 | case X86Local::MRMDestReg: |
315 | return RHS == X86Local::MRMDestMem; |
316 | case X86Local::MRMSrcReg: |
317 | return RHS == X86Local::MRMSrcMem; |
318 | case X86Local::MRMSrcReg4VOp3: |
319 | return RHS == X86Local::MRMSrcMem4VOp3; |
320 | case X86Local::MRMSrcRegOp4: |
321 | return RHS == X86Local::MRMSrcMemOp4; |
322 | case X86Local::MRMSrcRegCC: |
323 | return RHS == X86Local::MRMSrcMemCC; |
324 | } |
325 | } |
326 | |
327 | static bool isNOREXRegClass(const Record *Op) { |
328 | return Op->getName().contains(Other: "_NOREX" ); |
329 | } |
330 | |
331 | // Function object - Operator() returns true if the given Reg instruction |
332 | // matches the Mem instruction of this object. |
333 | class IsMatch { |
334 | const CodeGenInstruction *MemInst; |
335 | const X86Disassembler::RecognizableInstrBase MemRI; |
336 | bool IsBroadcast; |
337 | const unsigned Variant; |
338 | |
339 | public: |
340 | IsMatch(const CodeGenInstruction *Inst, bool IsBroadcast, unsigned V) |
341 | : MemInst(Inst), MemRI(*MemInst), IsBroadcast(IsBroadcast), Variant(V) {} |
342 | |
343 | bool operator()(const CodeGenInstruction *RegInst) { |
344 | X86Disassembler::RecognizableInstrBase RegRI(*RegInst); |
345 | const Record *RegRec = RegInst->TheDef; |
346 | const Record *MemRec = MemInst->TheDef; |
347 | |
348 | // EVEX_B means different things for memory and register forms. |
349 | // register form: rounding control or SAE |
350 | // memory form: broadcast |
351 | if (IsBroadcast && (RegRI.HasEVEX_B || !MemRI.HasEVEX_B)) |
352 | return false; |
353 | // EVEX_B indicates NDD for MAP4 instructions |
354 | if (!IsBroadcast && (RegRI.HasEVEX_B || MemRI.HasEVEX_B) && |
355 | RegRI.OpMap != X86Local::T_MAP4) |
356 | return false; |
357 | |
358 | if (!mayFoldFromLeftToRight(LHS: RegRI.Form, RHS: MemRI.Form)) |
359 | return false; |
360 | |
361 | // X86 encoding is crazy, e.g |
362 | // |
363 | // f3 0f c7 30 vmxon (%rax) |
364 | // f3 0f c7 f0 senduipi %rax |
365 | // |
366 | // This two instruction have similiar encoding fields but are unrelated |
367 | if (X86Disassembler::getMnemonic(I: MemInst, Variant) != |
368 | X86Disassembler::getMnemonic(I: RegInst, Variant)) |
369 | return false; |
370 | |
371 | // Return false if any of the following fields of does not match. |
372 | if (std::tuple(RegRI.Encoding, RegRI.Opcode, RegRI.OpPrefix, RegRI.OpMap, |
373 | RegRI.OpSize, RegRI.AdSize, RegRI.HasREX_W, RegRI.HasVEX_4V, |
374 | RegRI.HasVEX_L, RegRI.IgnoresVEX_L, RegRI.IgnoresW, |
375 | RegRI.HasEVEX_K, RegRI.HasEVEX_KZ, RegRI.HasEVEX_L2, |
376 | RegRI.HasEVEX_NF, RegRec->getValueAsBit(FieldName: "hasEVEX_RC" ), |
377 | RegRec->getValueAsBit(FieldName: "hasLockPrefix" ), |
378 | RegRec->getValueAsBit(FieldName: "hasNoTrackPrefix" )) != |
379 | std::tuple(MemRI.Encoding, MemRI.Opcode, MemRI.OpPrefix, MemRI.OpMap, |
380 | MemRI.OpSize, MemRI.AdSize, MemRI.HasREX_W, MemRI.HasVEX_4V, |
381 | MemRI.HasVEX_L, MemRI.IgnoresVEX_L, MemRI.IgnoresW, |
382 | MemRI.HasEVEX_K, MemRI.HasEVEX_KZ, MemRI.HasEVEX_L2, |
383 | MemRI.HasEVEX_NF, MemRec->getValueAsBit(FieldName: "hasEVEX_RC" ), |
384 | MemRec->getValueAsBit(FieldName: "hasLockPrefix" ), |
385 | MemRec->getValueAsBit(FieldName: "hasNoTrackPrefix" ))) |
386 | return false; |
387 | |
388 | // Make sure the sizes of the operands of both instructions suit each other. |
389 | // This is needed for instructions with intrinsic version (_Int). |
390 | // Where the only difference is the size of the operands. |
391 | // For example: VUCOMISDZrm and VUCOMISDrm_Int |
392 | // Also for instructions that their EVEX version was upgraded to work with |
393 | // k-registers. For example VPCMPEQBrm (xmm output register) and |
394 | // VPCMPEQBZ128rm (k register output register). |
395 | unsigned MemOutSize = MemRec->getValueAsDag(FieldName: "OutOperandList" )->getNumArgs(); |
396 | unsigned RegOutSize = RegRec->getValueAsDag(FieldName: "OutOperandList" )->getNumArgs(); |
397 | unsigned MemInSize = MemRec->getValueAsDag(FieldName: "InOperandList" )->getNumArgs(); |
398 | unsigned RegInSize = RegRec->getValueAsDag(FieldName: "InOperandList" )->getNumArgs(); |
399 | |
400 | // Instructions with one output in their memory form use the memory folded |
401 | // operand as source and destination (Read-Modify-Write). |
402 | unsigned RegStartIdx = |
403 | (MemOutSize + 1 == RegOutSize) && (MemInSize == RegInSize) ? 1 : 0; |
404 | |
405 | bool FoundFoldedOp = false; |
406 | for (unsigned I = 0, E = MemInst->Operands.size(); I != E; I++) { |
407 | Record *MemOpRec = MemInst->Operands[I].Rec; |
408 | Record *RegOpRec = RegInst->Operands[I + RegStartIdx].Rec; |
409 | |
410 | if (MemOpRec == RegOpRec) |
411 | continue; |
412 | |
413 | if (isRegisterOperand(Rec: MemOpRec) && isRegisterOperand(Rec: RegOpRec) && |
414 | ((getRegOperandSize(RegRec: MemOpRec) != getRegOperandSize(RegRec: RegOpRec)) || |
415 | (isNOREXRegClass(Op: MemOpRec) != isNOREXRegClass(Op: RegOpRec)))) |
416 | return false; |
417 | |
418 | if (isMemoryOperand(Rec: MemOpRec) && isMemoryOperand(Rec: RegOpRec) && |
419 | (getMemOperandSize(MemRec: MemOpRec) != getMemOperandSize(MemRec: RegOpRec))) |
420 | return false; |
421 | |
422 | if (isImmediateOperand(Rec: MemOpRec) && isImmediateOperand(Rec: RegOpRec) && |
423 | (MemOpRec->getValueAsDef(FieldName: "Type" ) != RegOpRec->getValueAsDef(FieldName: "Type" ))) |
424 | return false; |
425 | |
426 | // Only one operand can be folded. |
427 | if (FoundFoldedOp) |
428 | return false; |
429 | |
430 | assert(isRegisterOperand(RegOpRec) && isMemoryOperand(MemOpRec)); |
431 | FoundFoldedOp = true; |
432 | } |
433 | |
434 | return FoundFoldedOp; |
435 | } |
436 | }; |
437 | |
438 | } // end anonymous namespace |
439 | |
440 | void X86FoldTablesEmitter::addEntryWithFlags(FoldTable &Table, |
441 | const CodeGenInstruction *RegInst, |
442 | const CodeGenInstruction *MemInst, |
443 | uint16_t S, unsigned FoldedIdx, |
444 | bool IsManual) { |
445 | |
446 | assert((IsManual || Table.find(RegInst) == Table.end()) && |
447 | "Override entry unexpectedly" ); |
448 | X86FoldTableEntry Result = X86FoldTableEntry(RegInst, MemInst); |
449 | Record *RegRec = RegInst->TheDef; |
450 | Result.NoReverse = S & TB_NO_REVERSE; |
451 | Result.NoForward = S & TB_NO_FORWARD; |
452 | Result.FoldLoad = S & TB_FOLDED_LOAD; |
453 | Result.FoldStore = S & TB_FOLDED_STORE; |
454 | Result.Alignment = Align(1ULL << ((S & TB_ALIGN_MASK) >> TB_ALIGN_SHIFT)); |
455 | if (IsManual) { |
456 | Table[RegInst] = Result; |
457 | return; |
458 | } |
459 | |
460 | Record *RegOpRec = RegInst->Operands[FoldedIdx].Rec; |
461 | Record *MemOpRec = MemInst->Operands[FoldedIdx].Rec; |
462 | |
463 | // Unfolding code generates a load/store instruction according to the size of |
464 | // the register in the register form instruction. |
465 | // If the register's size is greater than the memory's operand size, do not |
466 | // allow unfolding. |
467 | |
468 | // the unfolded load size will be based on the register size. If that’s bigger |
469 | // than the memory operand size, the unfolded load will load more memory and |
470 | // potentially cause a memory fault. |
471 | if (getRegOperandSize(RegRec: RegOpRec) > getMemOperandSize(MemRec: MemOpRec)) |
472 | Result.NoReverse = true; |
473 | |
474 | // Check no-kz version's isMoveReg |
475 | StringRef RegInstName = RegRec->getName(); |
476 | unsigned DropLen = |
477 | RegInstName.ends_with(Suffix: "rkz" ) ? 2 : (RegInstName.ends_with(Suffix: "rk" ) ? 1 : 0); |
478 | Record *BaseDef = |
479 | DropLen ? Records.getDef(Name: RegInstName.drop_back(N: DropLen)) : nullptr; |
480 | bool IsMoveReg = |
481 | BaseDef ? Target.getInstruction(InstRec: BaseDef).isMoveReg : RegInst->isMoveReg; |
482 | // A masked load can not be unfolded to a full load, otherwise it would access |
483 | // unexpected memory. A simple store can not be unfolded. |
484 | if (IsMoveReg && (BaseDef || Result.FoldStore)) |
485 | Result.NoReverse = true; |
486 | |
487 | uint8_t Enc = byteFromBitsInit(B: RegRec->getValueAsBitsInit(FieldName: "OpEncBits" )); |
488 | if (isExplicitAlign(Inst: RegInst)) { |
489 | // The instruction require explicitly aligned memory. |
490 | BitsInit *VectSize = RegRec->getValueAsBitsInit(FieldName: "VectSize" ); |
491 | Result.Alignment = Align(byteFromBitsInit(B: VectSize)); |
492 | } else if (!Enc && !isExplicitUnalign(Inst: RegInst) && |
493 | getMemOperandSize(MemRec: MemOpRec) > 64) { |
494 | // Instructions with XOP/VEX/EVEX encoding do not require alignment while |
495 | // SSE packed vector instructions require a 16 byte alignment. |
496 | Result.Alignment = Align(16); |
497 | } |
498 | // Expand is only ever created as a masked instruction. It is not safe to |
499 | // unfold a masked expand because we don't know if it came from an expand load |
500 | // intrinsic or folding a plain load. If it is from a expand load intrinsic, |
501 | // Unfolding to plain load would read more elements and could trigger a fault. |
502 | if (RegRec->getName().contains(Other: "EXPAND" )) |
503 | Result.NoReverse = true; |
504 | |
505 | Table[RegInst] = Result; |
506 | } |
507 | |
508 | void X86FoldTablesEmitter::addBroadcastEntry( |
509 | FoldTable &Table, const CodeGenInstruction *RegInst, |
510 | const CodeGenInstruction *MemInst) { |
511 | |
512 | assert(Table.find(RegInst) == Table.end() && "Override entry unexpectedly" ); |
513 | X86FoldTableEntry Result = X86FoldTableEntry(RegInst, MemInst); |
514 | |
515 | DagInit *In = MemInst->TheDef->getValueAsDag(FieldName: "InOperandList" ); |
516 | for (unsigned I = 0, E = In->getNumArgs(); I != E; ++I) { |
517 | Result.BroadcastKind = |
518 | StringSwitch<X86FoldTableEntry::BcastType>(In->getArg(Num: I)->getAsString()) |
519 | .Case(S: "i16mem" , Value: X86FoldTableEntry::BCAST_W) |
520 | .Case(S: "i32mem" , Value: X86FoldTableEntry::BCAST_D) |
521 | .Case(S: "i64mem" , Value: X86FoldTableEntry::BCAST_Q) |
522 | .Case(S: "f16mem" , Value: X86FoldTableEntry::BCAST_SH) |
523 | .Case(S: "f32mem" , Value: X86FoldTableEntry::BCAST_SS) |
524 | .Case(S: "f64mem" , Value: X86FoldTableEntry::BCAST_SD) |
525 | .Default(Value: X86FoldTableEntry::BCAST_NONE); |
526 | if (Result.BroadcastKind != X86FoldTableEntry::BCAST_NONE) |
527 | break; |
528 | } |
529 | assert(Result.BroadcastKind != X86FoldTableEntry::BCAST_NONE && |
530 | "Unknown memory operand for broadcast" ); |
531 | |
532 | Table[RegInst] = Result; |
533 | } |
534 | |
535 | void X86FoldTablesEmitter::updateTables(const CodeGenInstruction *RegInst, |
536 | const CodeGenInstruction *MemInst, |
537 | uint16_t S, bool IsManual, |
538 | bool IsBroadcast) { |
539 | |
540 | Record *RegRec = RegInst->TheDef; |
541 | Record *MemRec = MemInst->TheDef; |
542 | unsigned MemOutSize = MemRec->getValueAsDag(FieldName: "OutOperandList" )->getNumArgs(); |
543 | unsigned RegOutSize = RegRec->getValueAsDag(FieldName: "OutOperandList" )->getNumArgs(); |
544 | unsigned MemInSize = MemRec->getValueAsDag(FieldName: "InOperandList" )->getNumArgs(); |
545 | unsigned RegInSize = RegRec->getValueAsDag(FieldName: "InOperandList" )->getNumArgs(); |
546 | |
547 | // Instructions which Read-Modify-Write should be added to Table2Addr. |
548 | if (!MemOutSize && RegOutSize == 1 && MemInSize == RegInSize) { |
549 | assert(!IsBroadcast && "Read-Modify-Write can not be broadcast" ); |
550 | // X86 would not unfold Read-Modify-Write instructions so add TB_NO_REVERSE. |
551 | addEntryWithFlags(Table&: Table2Addr, RegInst, MemInst, S: S | TB_NO_REVERSE, FoldedIdx: 0, |
552 | IsManual); |
553 | return; |
554 | } |
555 | |
556 | // Only table0 entries should explicitly specify a load or store flag. |
557 | // If the instruction writes to the folded operand, it will appear as |
558 | // an output in the register form instruction and as an input in the |
559 | // memory form instruction. If the instruction reads from the folded |
560 | // operand, it will appear as in input in both forms. |
561 | if (MemInSize == RegInSize && MemOutSize == RegOutSize) { |
562 | // Load-Folding cases. |
563 | // If the i'th register form operand is a register and the i'th memory form |
564 | // operand is a memory operand, add instructions to Table#i. |
565 | for (unsigned I = RegOutSize, E = RegInst->Operands.size(); I < E; I++) { |
566 | Record *RegOpRec = RegInst->Operands[I].Rec; |
567 | Record *MemOpRec = MemInst->Operands[I].Rec; |
568 | // PointerLikeRegClass: For instructions like TAILJMPr, TAILJMPr64, |
569 | // TAILJMPr64_REX |
570 | if ((isRegisterOperand(Rec: RegOpRec) || |
571 | RegOpRec->isSubClassOf(Name: "PointerLikeRegClass" )) && |
572 | isMemoryOperand(Rec: MemOpRec)) { |
573 | switch (I) { |
574 | case 0: |
575 | assert(!IsBroadcast && "BroadcastTable0 needs to be added" ); |
576 | addEntryWithFlags(Table&: Table0, RegInst, MemInst, S: S | TB_FOLDED_LOAD, FoldedIdx: 0, |
577 | IsManual); |
578 | return; |
579 | case 1: |
580 | IsBroadcast |
581 | ? addBroadcastEntry(Table&: BroadcastTable1, RegInst, MemInst) |
582 | : addEntryWithFlags(Table&: Table1, RegInst, MemInst, S, FoldedIdx: 1, IsManual); |
583 | return; |
584 | case 2: |
585 | IsBroadcast |
586 | ? addBroadcastEntry(Table&: BroadcastTable2, RegInst, MemInst) |
587 | : addEntryWithFlags(Table&: Table2, RegInst, MemInst, S, FoldedIdx: 2, IsManual); |
588 | return; |
589 | case 3: |
590 | IsBroadcast |
591 | ? addBroadcastEntry(Table&: BroadcastTable3, RegInst, MemInst) |
592 | : addEntryWithFlags(Table&: Table3, RegInst, MemInst, S, FoldedIdx: 3, IsManual); |
593 | return; |
594 | case 4: |
595 | IsBroadcast |
596 | ? addBroadcastEntry(Table&: BroadcastTable4, RegInst, MemInst) |
597 | : addEntryWithFlags(Table&: Table4, RegInst, MemInst, S, FoldedIdx: 4, IsManual); |
598 | return; |
599 | } |
600 | } |
601 | } |
602 | } else if (MemInSize == RegInSize + 1 && MemOutSize + 1 == RegOutSize) { |
603 | // Store-Folding cases. |
604 | // If the memory form instruction performs a store, the *output* |
605 | // register of the register form instructions disappear and instead a |
606 | // memory *input* operand appears in the memory form instruction. |
607 | // For example: |
608 | // MOVAPSrr => (outs VR128:$dst), (ins VR128:$src) |
609 | // MOVAPSmr => (outs), (ins f128mem:$dst, VR128:$src) |
610 | Record *RegOpRec = RegInst->Operands[RegOutSize - 1].Rec; |
611 | Record *MemOpRec = MemInst->Operands[RegOutSize - 1].Rec; |
612 | if (isRegisterOperand(Rec: RegOpRec) && isMemoryOperand(Rec: MemOpRec) && |
613 | getRegOperandSize(RegRec: RegOpRec) == getMemOperandSize(MemRec: MemOpRec)) { |
614 | assert(!IsBroadcast && "Store can not be broadcast" ); |
615 | addEntryWithFlags(Table&: Table0, RegInst, MemInst, S: S | TB_FOLDED_STORE, FoldedIdx: 0, |
616 | IsManual); |
617 | } |
618 | } |
619 | } |
620 | |
621 | void X86FoldTablesEmitter::run(raw_ostream &OS) { |
622 | // Holds all memory instructions |
623 | std::vector<const CodeGenInstruction *> MemInsts; |
624 | // Holds all register instructions - divided according to opcode. |
625 | std::map<uint8_t, std::vector<const CodeGenInstruction *>> RegInsts; |
626 | |
627 | ArrayRef<const CodeGenInstruction *> NumberedInstructions = |
628 | Target.getInstructionsByEnumValue(); |
629 | |
630 | for (const CodeGenInstruction *Inst : NumberedInstructions) { |
631 | const Record *Rec = Inst->TheDef; |
632 | if (!Rec->isSubClassOf(Name: "X86Inst" ) || Rec->getValueAsBit(FieldName: "isAsmParserOnly" )) |
633 | continue; |
634 | |
635 | if (NoFoldSet.find(x: Rec->getName()) != NoFoldSet.end()) |
636 | continue; |
637 | |
638 | // Promoted legacy instruction is in EVEX space, and has REX2-encoding |
639 | // alternative. It's added due to HW design and never emitted by compiler. |
640 | if (byteFromBitsInit(B: Rec->getValueAsBitsInit(FieldName: "OpMapBits" )) == |
641 | X86Local::T_MAP4 && |
642 | byteFromBitsInit(B: Rec->getValueAsBitsInit(FieldName: "explicitOpPrefixBits" )) == |
643 | X86Local::ExplicitEVEX) |
644 | continue; |
645 | |
646 | // - Instructions including RST register class operands are not relevant |
647 | // for memory folding (for further details check the explanation in |
648 | // lib/Target/X86/X86InstrFPStack.td file). |
649 | // - Some instructions (listed in the manual map above) use the register |
650 | // class ptr_rc_tailcall, which can be of a size 32 or 64, to ensure |
651 | // safe mapping of these instruction we manually map them and exclude |
652 | // them from the automation. |
653 | if (hasRSTRegClass(Inst) || hasPtrTailcallRegClass(Inst)) |
654 | continue; |
655 | |
656 | // Add all the memory form instructions to MemInsts, and all the register |
657 | // form instructions to RegInsts[Opc], where Opc is the opcode of each |
658 | // instructions. this helps reducing the runtime of the backend. |
659 | const BitsInit *FormBits = Rec->getValueAsBitsInit(FieldName: "FormBits" ); |
660 | uint8_t Form = byteFromBitsInit(B: FormBits); |
661 | if (mayFoldToForm(Form)) |
662 | MemInsts.push_back(x: Inst); |
663 | else if (mayFoldFromForm(Form)) { |
664 | uint8_t Opc = byteFromBitsInit(B: Rec->getValueAsBitsInit(FieldName: "Opcode" )); |
665 | RegInsts[Opc].push_back(x: Inst); |
666 | } |
667 | } |
668 | |
669 | // Create a copy b/c the register instruction will removed when a new entry is |
670 | // added into memory fold tables. |
671 | auto RegInstsForBroadcast = RegInsts; |
672 | |
673 | Record *AsmWriter = Target.getAsmWriter(); |
674 | unsigned Variant = AsmWriter->getValueAsInt(FieldName: "Variant" ); |
675 | auto FixUp = [&](const CodeGenInstruction *RegInst) { |
676 | StringRef RegInstName = RegInst->TheDef->getName(); |
677 | if (RegInstName.ends_with(Suffix: "_REV" ) || RegInstName.ends_with(Suffix: "_alt" )) |
678 | if (auto *RegAltRec = Records.getDef(Name: RegInstName.drop_back(N: 4))) |
679 | RegInst = &Target.getInstruction(InstRec: RegAltRec); |
680 | return RegInst; |
681 | }; |
682 | // For each memory form instruction, try to find its register form |
683 | // instruction. |
684 | for (const CodeGenInstruction *MemInst : MemInsts) { |
685 | uint8_t Opc = |
686 | byteFromBitsInit(B: MemInst->TheDef->getValueAsBitsInit(FieldName: "Opcode" )); |
687 | |
688 | auto RegInstsIt = RegInsts.find(x: Opc); |
689 | if (RegInstsIt == RegInsts.end()) |
690 | continue; |
691 | |
692 | // Two forms (memory & register) of the same instruction must have the same |
693 | // opcode. |
694 | std::vector<const CodeGenInstruction *> &OpcRegInsts = RegInstsIt->second; |
695 | |
696 | // Memory fold tables |
697 | auto Match = |
698 | find_if(Range&: OpcRegInsts, P: IsMatch(MemInst, /*IsBroadcast=*/false, Variant)); |
699 | if (Match != OpcRegInsts.end()) { |
700 | updateTables(RegInst: FixUp(*Match), MemInst); |
701 | OpcRegInsts.erase(position: Match); |
702 | } |
703 | |
704 | // Broadcast tables |
705 | StringRef MemInstName = MemInst->TheDef->getName(); |
706 | if (!MemInstName.contains(Other: "mb" ) && !MemInstName.contains(Other: "mib" )) |
707 | continue; |
708 | RegInstsIt = RegInstsForBroadcast.find(x: Opc); |
709 | assert(RegInstsIt != RegInstsForBroadcast.end() && |
710 | "Unexpected control flow" ); |
711 | std::vector<const CodeGenInstruction *> &OpcRegInstsForBroadcast = |
712 | RegInstsIt->second; |
713 | Match = find_if(Range&: OpcRegInstsForBroadcast, |
714 | P: IsMatch(MemInst, /*IsBroadcast=*/true, Variant)); |
715 | if (Match != OpcRegInstsForBroadcast.end()) { |
716 | updateTables(RegInst: FixUp(*Match), MemInst, S: 0, /*IsManual=*/false, |
717 | /*IsBroadcast=*/true); |
718 | OpcRegInstsForBroadcast.erase(position: Match); |
719 | } |
720 | } |
721 | |
722 | // Add the manually mapped instructions listed above. |
723 | for (const ManualMapEntry &Entry : ManualMapSet) { |
724 | Record *RegInstIter = Records.getDef(Name: Entry.RegInstStr); |
725 | Record *MemInstIter = Records.getDef(Name: Entry.MemInstStr); |
726 | |
727 | updateTables(RegInst: &(Target.getInstruction(InstRec: RegInstIter)), |
728 | MemInst: &(Target.getInstruction(InstRec: MemInstIter)), S: Entry.Strategy, IsManual: true); |
729 | } |
730 | |
731 | #ifndef NDEBUG |
732 | auto CheckMemFoldTable = [](const FoldTable &Table) -> void { |
733 | for (const auto &Record : Table) { |
734 | auto &FoldEntry = Record.second; |
735 | FoldEntry.checkCorrectness(); |
736 | } |
737 | }; |
738 | CheckMemFoldTable(Table2Addr); |
739 | CheckMemFoldTable(Table0); |
740 | CheckMemFoldTable(Table1); |
741 | CheckMemFoldTable(Table2); |
742 | CheckMemFoldTable(Table3); |
743 | CheckMemFoldTable(Table4); |
744 | CheckMemFoldTable(BroadcastTable1); |
745 | CheckMemFoldTable(BroadcastTable2); |
746 | CheckMemFoldTable(BroadcastTable3); |
747 | CheckMemFoldTable(BroadcastTable4); |
748 | #endif |
749 | #define PRINT_TABLE(TABLE) printTable(TABLE, #TABLE, OS); |
750 | // Print all tables. |
751 | PRINT_TABLE(Table2Addr) |
752 | PRINT_TABLE(Table0) |
753 | PRINT_TABLE(Table1) |
754 | PRINT_TABLE(Table2) |
755 | PRINT_TABLE(Table3) |
756 | PRINT_TABLE(Table4) |
757 | PRINT_TABLE(BroadcastTable1) |
758 | PRINT_TABLE(BroadcastTable2) |
759 | PRINT_TABLE(BroadcastTable3) |
760 | PRINT_TABLE(BroadcastTable4) |
761 | } |
762 | |
763 | static TableGen::Emitter::OptClass<X86FoldTablesEmitter> |
764 | X("gen-x86-fold-tables" , "Generate X86 fold tables" ); |
765 | |