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 "CodeGenTarget.h"
15 #include "X86RecognizableInstr.h"
16 #include "llvm/Support/FormattedStream.h"
17 #include "llvm/TableGen/Error.h"
18 #include "llvm/TableGen/TableGenBackend.h"
19 
20 using namespace llvm;
21 using namespace X86Disassembler;
22 
23 namespace {
24 
25 // 3 possible strategies for the unfolding flag (TB_NO_REVERSE) of the
26 // manual added entries.
27 enum UnfoldStrategy {
28   UNFOLD,     // Allow unfolding
29   NO_UNFOLD,  // Prevent unfolding
30   NO_STRATEGY // Make decision according to operands' sizes
31 };
32 
33 // Represents an entry in the manual mapped instructions set.
34 struct ManualMapEntry {
35   const char *RegInstStr;
36   const char *MemInstStr;
37   UnfoldStrategy Strategy;
38 
39   ManualMapEntry(const char *RegInstStr, const char *MemInstStr,
40                  UnfoldStrategy Strategy = NO_STRATEGY)
41       : RegInstStr(RegInstStr), MemInstStr(MemInstStr), Strategy(Strategy) {}
42 };
43 
44 // List of instructions requiring explicitly aligned memory.
45 const char *ExplicitAlign[] = {"MOVDQA",  "MOVAPS",  "MOVAPD",  "MOVNTPS",
46                                "MOVNTPD", "MOVNTDQ", "MOVNTDQA"};
47 
48 // List of instructions NOT requiring explicit memory alignment.
49 const char *ExplicitUnalign[] = {"MOVDQU", "MOVUPS", "MOVUPD",
50                                  "PCMPESTRM", "PCMPESTRI",
51                                  "PCMPISTRM", "PCMPISTRI" };
52 
53 // For manually mapping instructions that do not match by their encoding.
54 const ManualMapEntry ManualMapSet[] = {
55     { "ADD16ri_DB",         "ADD16mi",         NO_UNFOLD  },
56     { "ADD16ri8_DB",        "ADD16mi8",        NO_UNFOLD  },
57     { "ADD16rr_DB",         "ADD16mr",         NO_UNFOLD  },
58     { "ADD32ri_DB",         "ADD32mi",         NO_UNFOLD  },
59     { "ADD32ri8_DB",        "ADD32mi8",        NO_UNFOLD  },
60     { "ADD32rr_DB",         "ADD32mr",         NO_UNFOLD  },
61     { "ADD64ri32_DB",       "ADD64mi32",       NO_UNFOLD  },
62     { "ADD64ri8_DB",        "ADD64mi8",        NO_UNFOLD  },
63     { "ADD64rr_DB",         "ADD64mr",         NO_UNFOLD  },
64     { "ADD8ri_DB",          "ADD8mi",          NO_UNFOLD  },
65     { "ADD8rr_DB",          "ADD8mr",          NO_UNFOLD  },
66     { "ADD16rr_DB",         "ADD16rm",         NO_UNFOLD  },
67     { "ADD32rr_DB",         "ADD32rm",         NO_UNFOLD  },
68     { "ADD64rr_DB",         "ADD64rm",         NO_UNFOLD  },
69     { "ADD8rr_DB",          "ADD8rm",          NO_UNFOLD  },
70     { "MMX_MOVD64from64rr", "MMX_MOVQ64mr",    UNFOLD },
71     { "MMX_MOVD64grr",      "MMX_MOVD64mr",    UNFOLD },
72     { "MOVLHPSrr",          "MOVHPSrm",        NO_UNFOLD  },
73     { "PUSH16r",            "PUSH16rmm",       UNFOLD },
74     { "PUSH32r",            "PUSH32rmm",       UNFOLD },
75     { "PUSH64r",            "PUSH64rmm",       UNFOLD },
76     { "TAILJMPr",           "TAILJMPm",        UNFOLD },
77     { "TAILJMPr64",         "TAILJMPm64",      UNFOLD },
78     { "TAILJMPr64_REX",     "TAILJMPm64_REX",  UNFOLD },
79     { "VMOVLHPSZrr",        "VMOVHPSZ128rm",   NO_UNFOLD  },
80     { "VMOVLHPSrr",         "VMOVHPSrm",       NO_UNFOLD  },
81 };
82 
83 
84 static bool isExplicitAlign(const CodeGenInstruction *Inst) {
85   return any_of(ExplicitAlign, [Inst](const char *InstStr) {
86     return Inst->TheDef->getName().contains(InstStr);
87   });
88 }
89 
90 static bool isExplicitUnalign(const CodeGenInstruction *Inst) {
91   return any_of(ExplicitUnalign, [Inst](const char *InstStr) {
92     return Inst->TheDef->getName().contains(InstStr);
93   });
94 }
95 
96 class X86FoldTablesEmitter {
97   RecordKeeper &Records;
98   CodeGenTarget Target;
99 
100   // Represents an entry in the folding table
101   class X86FoldTableEntry {
102     const CodeGenInstruction *RegInst;
103     const CodeGenInstruction *MemInst;
104 
105   public:
106     bool CannotUnfold = false;
107     bool IsLoad = false;
108     bool IsStore = false;
109     bool IsAligned = false;
110     unsigned int Alignment = 0;
111 
112     X86FoldTableEntry(const CodeGenInstruction *RegInst,
113                       const CodeGenInstruction *MemInst)
114         : RegInst(RegInst), MemInst(MemInst) {}
115 
116     void print(formatted_raw_ostream &OS) const {
117       OS.indent(2);
118       OS << "{ X86::" << RegInst->TheDef->getName() << ",";
119       OS.PadToColumn(40);
120       OS  << "X86::" << MemInst->TheDef->getName() << ",";
121       OS.PadToColumn(75);
122 
123       std::string Attrs;
124       if (IsLoad)
125         Attrs += "TB_FOLDED_LOAD | ";
126       if (IsStore)
127         Attrs += "TB_FOLDED_STORE | ";
128       if (CannotUnfold)
129         Attrs += "TB_NO_REVERSE | ";
130       if (IsAligned)
131         Attrs += "TB_ALIGN_" + std::to_string(Alignment) + " | ";
132 
133       StringRef SimplifiedAttrs = StringRef(Attrs).rtrim("| ");
134       if (SimplifiedAttrs.empty())
135         SimplifiedAttrs = "0";
136 
137       OS << SimplifiedAttrs << " },\n";
138     }
139 
140     bool operator<(const X86FoldTableEntry &RHS) const {
141       bool LHSpseudo = RegInst->TheDef->getValueAsBit("isPseudo");
142       bool RHSpseudo = RHS.RegInst->TheDef->getValueAsBit("isPseudo");
143       if (LHSpseudo != RHSpseudo)
144         return LHSpseudo;
145 
146       return RegInst->TheDef->getName() < RHS.RegInst->TheDef->getName();
147     }
148   };
149 
150   typedef std::vector<X86FoldTableEntry> FoldTable;
151   // std::vector for each folding table.
152   // Table2Addr - Holds instructions which their memory form performs load+store
153   // Table#i - Holds instructions which the their memory form perform a load OR
154   //           a store,  and their #i'th operand is folded.
155   FoldTable Table2Addr;
156   FoldTable Table0;
157   FoldTable Table1;
158   FoldTable Table2;
159   FoldTable Table3;
160   FoldTable Table4;
161 
162 public:
163   X86FoldTablesEmitter(RecordKeeper &R) : Records(R), Target(R) {}
164 
165   // run - Generate the 6 X86 memory fold tables.
166   void run(formatted_raw_ostream &OS);
167 
168 private:
169   // Decides to which table to add the entry with the given instructions.
170   // S sets the strategy of adding the TB_NO_REVERSE flag.
171   void updateTables(const CodeGenInstruction *RegInstr,
172                     const CodeGenInstruction *MemInstr,
173                     const UnfoldStrategy S = NO_STRATEGY);
174 
175   // Generates X86FoldTableEntry with the given instructions and fill it with
176   // the appropriate flags - then adds it to Table.
177   void addEntryWithFlags(FoldTable &Table, const CodeGenInstruction *RegInstr,
178                          const CodeGenInstruction *MemInstr,
179                          const UnfoldStrategy S, const unsigned int FoldedInd);
180 
181   // Print the given table as a static const C++ array of type
182   // X86MemoryFoldTableEntry.
183   void printTable(const FoldTable &Table, StringRef TableName,
184                   formatted_raw_ostream &OS) {
185     OS << "static const X86MemoryFoldTableEntry MemoryFold" << TableName
186        << "[] = {\n";
187 
188     for (const X86FoldTableEntry &E : Table)
189       E.print(OS);
190 
191     OS << "};\n\n";
192   }
193 };
194 
195 // Return true if one of the instruction's operands is a RST register class
196 static bool hasRSTRegClass(const CodeGenInstruction *Inst) {
197   return any_of(Inst->Operands, [](const CGIOperandList::OperandInfo &OpIn) {
198     return OpIn.Rec->getName() == "RST" || OpIn.Rec->getName() == "RSTi";
199   });
200 }
201 
202 // Return true if one of the instruction's operands is a ptr_rc_tailcall
203 static bool hasPtrTailcallRegClass(const CodeGenInstruction *Inst) {
204   return any_of(Inst->Operands, [](const CGIOperandList::OperandInfo &OpIn) {
205     return OpIn.Rec->getName() == "ptr_rc_tailcall";
206   });
207 }
208 
209 // Calculates the integer value representing the BitsInit object
210 static inline uint64_t getValueFromBitsInit(const BitsInit *B) {
211   assert(B->getNumBits() <= sizeof(uint64_t) * 8 && "BitInits' too long!");
212 
213   uint64_t Value = 0;
214   for (unsigned i = 0, e = B->getNumBits(); i != e; ++i) {
215     BitInit *Bit = cast<BitInit>(B->getBit(i));
216     Value |= uint64_t(Bit->getValue()) << i;
217   }
218   return Value;
219 }
220 
221 // Return true if the instruction defined as a register flavor.
222 static inline bool hasRegisterFormat(const Record *Inst) {
223   const BitsInit *FormBits = Inst->getValueAsBitsInit("FormBits");
224   uint64_t FormBitsNum = getValueFromBitsInit(FormBits);
225 
226   // Values from X86Local namespace defined in X86RecognizableInstr.cpp
227   return FormBitsNum >= X86Local::MRMDestReg && FormBitsNum <= X86Local::MRM7r;
228 }
229 
230 // Return true if the instruction defined as a memory flavor.
231 static inline bool hasMemoryFormat(const Record *Inst) {
232   const BitsInit *FormBits = Inst->getValueAsBitsInit("FormBits");
233   uint64_t FormBitsNum = getValueFromBitsInit(FormBits);
234 
235   // Values from X86Local namespace defined in X86RecognizableInstr.cpp
236   return FormBitsNum >= X86Local::MRMDestMem && FormBitsNum <= X86Local::MRM7m;
237 }
238 
239 static inline bool isNOREXRegClass(const Record *Op) {
240   return Op->getName().contains("_NOREX");
241 }
242 
243 // Get the alternative instruction pointed by "FoldGenRegForm" field.
244 static inline const CodeGenInstruction *
245 getAltRegInst(const CodeGenInstruction *I, const RecordKeeper &Records,
246               const CodeGenTarget &Target) {
247 
248   StringRef AltRegInstStr = I->TheDef->getValueAsString("FoldGenRegForm");
249   Record *AltRegInstRec = Records.getDef(AltRegInstStr);
250   assert(AltRegInstRec &&
251          "Alternative register form instruction def not found");
252   CodeGenInstruction &AltRegInst = Target.getInstruction(AltRegInstRec);
253   return &AltRegInst;
254 }
255 
256 // Function object - Operator() returns true if the given VEX instruction
257 // matches the EVEX instruction of this object.
258 class IsMatch {
259   const CodeGenInstruction *MemInst;
260   unsigned Variant;
261 
262 public:
263   IsMatch(const CodeGenInstruction *Inst, unsigned V)
264       : MemInst(Inst), Variant(V) {}
265 
266   bool operator()(const CodeGenInstruction *RegInst) {
267     X86Disassembler::RecognizableInstrBase RegRI(*RegInst);
268     X86Disassembler::RecognizableInstrBase MemRI(*MemInst);
269     const Record *RegRec = RegInst->TheDef;
270     const Record *MemRec = MemInst->TheDef;
271 
272     // EVEX_B means different things for memory and register forms.
273     if (RegRI.HasEVEX_B != 0 || MemRI.HasEVEX_B != 0)
274       return false;
275 
276     // Instruction's format - The register form's "Form" field should be
277     // the opposite of the memory form's "Form" field.
278     if (!areOppositeForms(RegRI.Form, MemRI.Form))
279       return false;
280 
281     // X86 encoding is crazy, e.g
282     //
283     // f3 0f c7 30       vmxon   (%rax)
284     // f3 0f c7 f0       senduipi        %rax
285     //
286     // This two instruction have similiar encoding fields but are unrelated
287     if (X86Disassembler::getMnemonic(MemInst, Variant) !=
288         X86Disassembler::getMnemonic(RegInst, Variant))
289       return false;
290 
291     // Return false if one (at least) of the encoding fields of both
292     // instructions do not match.
293     if (RegRI.Encoding != MemRI.Encoding || RegRI.Opcode != MemRI.Opcode ||
294         RegRI.OpPrefix != MemRI.OpPrefix || RegRI.OpMap != MemRI.OpMap ||
295         RegRI.OpSize != MemRI.OpSize || RegRI.AdSize != MemRI.AdSize ||
296         RegRI.HasREX_W != MemRI.HasREX_W ||
297         RegRI.HasVEX_4V != MemRI.HasVEX_4V ||
298         RegRI.HasVEX_L != MemRI.HasVEX_L ||
299         RegRI.HasVEX_W != MemRI.HasVEX_W ||
300         RegRI.IgnoresVEX_L != MemRI.IgnoresVEX_L ||
301         RegRI.IgnoresVEX_W != MemRI.IgnoresVEX_W ||
302         RegRI.HasEVEX_K != MemRI.HasEVEX_K ||
303         RegRI.HasEVEX_KZ != MemRI.HasEVEX_KZ ||
304         RegRI.HasEVEX_L2 != MemRI.HasEVEX_L2 ||
305         RegRec->getValueAsBit("hasEVEX_RC") !=
306             MemRec->getValueAsBit("hasEVEX_RC") ||
307         RegRec->getValueAsBit("hasLockPrefix") !=
308             MemRec->getValueAsBit("hasLockPrefix") ||
309         RegRec->getValueAsBit("hasNoTrackPrefix") !=
310             MemRec->getValueAsBit("hasNoTrackPrefix") ||
311         RegRec->getValueAsBit("EVEX_W1_VEX_W0") !=
312             MemRec->getValueAsBit("EVEX_W1_VEX_W0"))
313       return false;
314 
315     // Make sure the sizes of the operands of both instructions suit each other.
316     // This is needed for instructions with intrinsic version (_Int).
317     // Where the only difference is the size of the operands.
318     // For example: VUCOMISDZrm and Int_VUCOMISDrm
319     // Also for instructions that their EVEX version was upgraded to work with
320     // k-registers. For example VPCMPEQBrm (xmm output register) and
321     // VPCMPEQBZ128rm (k register output register).
322     bool ArgFolded = false;
323     unsigned MemOutSize = MemRec->getValueAsDag("OutOperandList")->getNumArgs();
324     unsigned RegOutSize = RegRec->getValueAsDag("OutOperandList")->getNumArgs();
325     unsigned MemInSize = MemRec->getValueAsDag("InOperandList")->getNumArgs();
326     unsigned RegInSize = RegRec->getValueAsDag("InOperandList")->getNumArgs();
327 
328     // Instructions with one output in their memory form use the memory folded
329     // operand as source and destination (Read-Modify-Write).
330     unsigned RegStartIdx =
331         (MemOutSize + 1 == RegOutSize) && (MemInSize == RegInSize) ? 1 : 0;
332 
333     for (unsigned i = 0, e = MemInst->Operands.size(); i < e; i++) {
334       Record *MemOpRec = MemInst->Operands[i].Rec;
335       Record *RegOpRec = RegInst->Operands[i + RegStartIdx].Rec;
336 
337       if (MemOpRec == RegOpRec)
338         continue;
339 
340       if (isRegisterOperand(MemOpRec) && isRegisterOperand(RegOpRec)) {
341         if (getRegOperandSize(MemOpRec) != getRegOperandSize(RegOpRec) ||
342             isNOREXRegClass(MemOpRec) != isNOREXRegClass(RegOpRec))
343           return false;
344       } else if (isMemoryOperand(MemOpRec) && isMemoryOperand(RegOpRec)) {
345         if (getMemOperandSize(MemOpRec) != getMemOperandSize(RegOpRec))
346           return false;
347       } else if (isImmediateOperand(MemOpRec) && isImmediateOperand(RegOpRec)) {
348         if (MemOpRec->getValueAsDef("Type") != RegOpRec->getValueAsDef("Type"))
349           return false;
350       } else {
351         // Only one operand can be folded.
352         if (ArgFolded)
353           return false;
354 
355         assert(isRegisterOperand(RegOpRec) && isMemoryOperand(MemOpRec));
356         ArgFolded = true;
357       }
358     }
359 
360     return true;
361   }
362 
363 private:
364   // Return true of the 2 given forms are the opposite of each other.
365   bool areOppositeForms(unsigned RegForm, unsigned MemForm) {
366     if ((MemForm == X86Local::MRM0m && RegForm == X86Local::MRM0r) ||
367         (MemForm == X86Local::MRM1m && RegForm == X86Local::MRM1r) ||
368         (MemForm == X86Local::MRM2m && RegForm == X86Local::MRM2r) ||
369         (MemForm == X86Local::MRM3m && RegForm == X86Local::MRM3r) ||
370         (MemForm == X86Local::MRM4m && RegForm == X86Local::MRM4r) ||
371         (MemForm == X86Local::MRM5m && RegForm == X86Local::MRM5r) ||
372         (MemForm == X86Local::MRM6m && RegForm == X86Local::MRM6r) ||
373         (MemForm == X86Local::MRM7m && RegForm == X86Local::MRM7r) ||
374         (MemForm == X86Local::MRMXm && RegForm == X86Local::MRMXr) ||
375         (MemForm == X86Local::MRMXmCC && RegForm == X86Local::MRMXrCC) ||
376         (MemForm == X86Local::MRMDestMem && RegForm == X86Local::MRMDestReg) ||
377         (MemForm == X86Local::MRMSrcMem && RegForm == X86Local::MRMSrcReg) ||
378         (MemForm == X86Local::MRMSrcMem4VOp3 &&
379          RegForm == X86Local::MRMSrcReg4VOp3) ||
380         (MemForm == X86Local::MRMSrcMemOp4 &&
381          RegForm == X86Local::MRMSrcRegOp4) ||
382         (MemForm == X86Local::MRMSrcMemCC && RegForm == X86Local::MRMSrcRegCC))
383       return true;
384 
385     return false;
386   }
387 };
388 
389 } // end anonymous namespace
390 
391 void X86FoldTablesEmitter::addEntryWithFlags(FoldTable &Table,
392                                              const CodeGenInstruction *RegInstr,
393                                              const CodeGenInstruction *MemInstr,
394                                              const UnfoldStrategy S,
395                                              const unsigned int FoldedInd) {
396 
397   X86FoldTableEntry Result = X86FoldTableEntry(RegInstr, MemInstr);
398   Record *RegRec = RegInstr->TheDef;
399   Record *MemRec = MemInstr->TheDef;
400 
401   // Only table0 entries should explicitly specify a load or store flag.
402   if (&Table == &Table0) {
403     unsigned MemInOpsNum = MemRec->getValueAsDag("InOperandList")->getNumArgs();
404     unsigned RegInOpsNum = RegRec->getValueAsDag("InOperandList")->getNumArgs();
405     // If the instruction writes to the folded operand, it will appear as an
406     // output in the register form instruction and as an input in the memory
407     // form instruction.
408     // If the instruction reads from the folded operand, it well appear as in
409     // input in both forms.
410     if (MemInOpsNum == RegInOpsNum)
411       Result.IsLoad = true;
412     else
413       Result.IsStore = true;
414   }
415 
416   Record *RegOpRec = RegInstr->Operands[FoldedInd].Rec;
417   Record *MemOpRec = MemInstr->Operands[FoldedInd].Rec;
418 
419   // Unfolding code generates a load/store instruction according to the size of
420   // the register in the register form instruction.
421   // If the register's size is greater than the memory's operand size, do not
422   // allow unfolding.
423   if (S == UNFOLD)
424     Result.CannotUnfold = false;
425   else if (S == NO_UNFOLD)
426     Result.CannotUnfold = true;
427   else if (getRegOperandSize(RegOpRec) > getMemOperandSize(MemOpRec))
428     Result.CannotUnfold = true; // S == NO_STRATEGY
429 
430   uint64_t Enc = getValueFromBitsInit(RegRec->getValueAsBitsInit("OpEncBits"));
431   if (isExplicitAlign(RegInstr)) {
432     // The instruction require explicitly aligned memory.
433     BitsInit *VectSize = RegRec->getValueAsBitsInit("VectSize");
434     uint64_t Value = getValueFromBitsInit(VectSize);
435     Result.IsAligned = true;
436     Result.Alignment = Value;
437   } else if (Enc != X86Local::XOP && Enc != X86Local::VEX &&
438              Enc != X86Local::EVEX) {
439     // Instructions with VEX encoding do not require alignment.
440     if (!isExplicitUnalign(RegInstr) && getMemOperandSize(MemOpRec) > 64) {
441       // SSE packed vector instructions require a 16 byte alignment.
442       Result.IsAligned = true;
443       Result.Alignment = 16;
444     }
445   }
446 
447   Table.push_back(Result);
448 }
449 
450 void X86FoldTablesEmitter::updateTables(const CodeGenInstruction *RegInstr,
451                                         const CodeGenInstruction *MemInstr,
452                                         const UnfoldStrategy S) {
453 
454   Record *RegRec = RegInstr->TheDef;
455   Record *MemRec = MemInstr->TheDef;
456   unsigned MemOutSize = MemRec->getValueAsDag("OutOperandList")->getNumArgs();
457   unsigned RegOutSize = RegRec->getValueAsDag("OutOperandList")->getNumArgs();
458   unsigned MemInSize = MemRec->getValueAsDag("InOperandList")->getNumArgs();
459   unsigned RegInSize = RegRec->getValueAsDag("InOperandList")->getNumArgs();
460 
461   // Instructions which Read-Modify-Write should be added to Table2Addr.
462   if (MemOutSize != RegOutSize && MemInSize == RegInSize) {
463     addEntryWithFlags(Table2Addr, RegInstr, MemInstr, S, 0);
464     return;
465   }
466 
467   if (MemInSize == RegInSize && MemOutSize == RegOutSize) {
468     // Load-Folding cases.
469     // If the i'th register form operand is a register and the i'th memory form
470     // operand is a memory operand, add instructions to Table#i.
471     for (unsigned i = RegOutSize, e = RegInstr->Operands.size(); i < e; i++) {
472       Record *RegOpRec = RegInstr->Operands[i].Rec;
473       Record *MemOpRec = MemInstr->Operands[i].Rec;
474       // PointerLikeRegClass: For instructions like TAILJMPr, TAILJMPr64, TAILJMPr64_REX
475       if ((isRegisterOperand(RegOpRec) ||
476            RegOpRec->isSubClassOf("PointerLikeRegClass")) &&
477           isMemoryOperand(MemOpRec)) {
478         switch (i) {
479         case 0:
480           addEntryWithFlags(Table0, RegInstr, MemInstr, S, 0);
481           return;
482         case 1:
483           addEntryWithFlags(Table1, RegInstr, MemInstr, S, 1);
484           return;
485         case 2:
486           addEntryWithFlags(Table2, RegInstr, MemInstr, S, 2);
487           return;
488         case 3:
489           addEntryWithFlags(Table3, RegInstr, MemInstr, S, 3);
490           return;
491         case 4:
492           addEntryWithFlags(Table4, RegInstr, MemInstr, S, 4);
493           return;
494         }
495       }
496     }
497   } else if (MemInSize == RegInSize + 1 && MemOutSize + 1 == RegOutSize) {
498     // Store-Folding cases.
499     // If the memory form instruction performs a store, the *output*
500     // register of the register form instructions disappear and instead a
501     // memory *input* operand appears in the memory form instruction.
502     // For example:
503     //   MOVAPSrr => (outs VR128:$dst), (ins VR128:$src)
504     //   MOVAPSmr => (outs), (ins f128mem:$dst, VR128:$src)
505     Record *RegOpRec = RegInstr->Operands[RegOutSize - 1].Rec;
506     Record *MemOpRec = MemInstr->Operands[RegOutSize - 1].Rec;
507     if (isRegisterOperand(RegOpRec) && isMemoryOperand(MemOpRec) &&
508         getRegOperandSize(RegOpRec) == getMemOperandSize(MemOpRec))
509       addEntryWithFlags(Table0, RegInstr, MemInstr, S, 0);
510   }
511 }
512 
513 void X86FoldTablesEmitter::run(formatted_raw_ostream &OS) {
514   emitSourceFileHeader("X86 fold tables", OS);
515 
516   // Holds all memory instructions
517   std::vector<const CodeGenInstruction *> MemInsts;
518   // Holds all register instructions - divided according to opcode.
519   std::map<uint8_t, std::vector<const CodeGenInstruction *>> RegInsts;
520 
521   ArrayRef<const CodeGenInstruction *> NumberedInstructions =
522       Target.getInstructionsByEnumValue();
523 
524   for (const CodeGenInstruction *Inst : NumberedInstructions) {
525     const Record *Rec = Inst->TheDef;
526     if (!Rec->isSubClassOf("X86Inst") || Rec->getValueAsBit("isAsmParserOnly"))
527       continue;
528 
529     // - Do not proceed if the instruction is marked as notMemoryFoldable.
530     // - Instructions including RST register class operands are not relevant
531     //   for memory folding (for further details check the explanation in
532     //   lib/Target/X86/X86InstrFPStack.td file).
533     // - Some instructions (listed in the manual map above) use the register
534     //   class ptr_rc_tailcall, which can be of a size 32 or 64, to ensure
535     //   safe mapping of these instruction we manually map them and exclude
536     //   them from the automation.
537     if (Rec->getValueAsBit("isMemoryFoldable") == false ||
538         hasRSTRegClass(Inst) || hasPtrTailcallRegClass(Inst))
539       continue;
540 
541     // Add all the memory form instructions to MemInsts, and all the register
542     // form instructions to RegInsts[Opc], where Opc in the opcode of each
543     // instructions. this helps reducing the runtime of the backend.
544     if (hasMemoryFormat(Rec))
545       MemInsts.push_back(Inst);
546     else if (hasRegisterFormat(Rec)) {
547       uint8_t Opc = getValueFromBitsInit(Rec->getValueAsBitsInit("Opcode"));
548       RegInsts[Opc].push_back(Inst);
549     }
550   }
551 
552   Record *AsmWriter = Target.getAsmWriter();
553   unsigned Variant = AsmWriter->getValueAsInt("Variant");
554   // For each memory form instruction, try to find its register form
555   // instruction.
556   for (const CodeGenInstruction *MemInst : MemInsts) {
557     uint8_t Opc =
558         getValueFromBitsInit(MemInst->TheDef->getValueAsBitsInit("Opcode"));
559 
560     auto RegInstsIt = RegInsts.find(Opc);
561     if (RegInstsIt == RegInsts.end())
562       continue;
563 
564     // Two forms (memory & register) of the same instruction must have the same
565     // opcode. try matching only with register form instructions with the same
566     // opcode.
567     std::vector<const CodeGenInstruction *> &OpcRegInsts = RegInstsIt->second;
568 
569     auto Match = find_if(OpcRegInsts, IsMatch(MemInst, Variant));
570     if (Match != OpcRegInsts.end()) {
571       const CodeGenInstruction *RegInst = *Match;
572       // If the matched instruction has it's "FoldGenRegForm" set, map the
573       // memory form instruction to the register form instruction pointed by
574       // this field
575       if (RegInst->TheDef->isValueUnset("FoldGenRegForm")) {
576         updateTables(RegInst, MemInst);
577       } else {
578         const CodeGenInstruction *AltRegInst =
579             getAltRegInst(RegInst, Records, Target);
580         updateTables(AltRegInst, MemInst);
581       }
582       OpcRegInsts.erase(Match);
583     }
584   }
585 
586   // Add the manually mapped instructions listed above.
587   for (const ManualMapEntry &Entry : ManualMapSet) {
588     Record *RegInstIter = Records.getDef(Entry.RegInstStr);
589     Record *MemInstIter = Records.getDef(Entry.MemInstStr);
590 
591     updateTables(&(Target.getInstruction(RegInstIter)),
592                  &(Target.getInstruction(MemInstIter)), Entry.Strategy);
593   }
594 
595   // Sort the tables before printing.
596   llvm::sort(Table2Addr);
597   llvm::sort(Table0);
598   llvm::sort(Table1);
599   llvm::sort(Table2);
600   llvm::sort(Table3);
601   llvm::sort(Table4);
602 
603   // Print all tables.
604   printTable(Table2Addr, "Table2Addr", OS);
605   printTable(Table0, "Table0", OS);
606   printTable(Table1, "Table1", OS);
607   printTable(Table2, "Table2", OS);
608   printTable(Table3, "Table3", OS);
609   printTable(Table4, "Table4", OS);
610 }
611 
612 namespace llvm {
613 
614 void EmitX86FoldTables(RecordKeeper &RK, raw_ostream &o) {
615   formatted_raw_ostream OS(o);
616   X86FoldTablesEmitter(RK).run(OS);
617 }
618 } // namespace llvm
619