3.6.2.3.3. 处理器特征数据
TD文件里通过SubTargetFeature定义来描述处理器所支持的指令集。反映到LLVM上，就体现为类型X86Subtarget（是下面输出X86GenSubtargetInfo类型的派生类）里的一系列开关。对每个处理器，显然会有一个对应的X86Subtarget实例。
SubtargetEmitter::run（续）
1468 OS << "\n#ifdef GET_SUBTARGETINFO_TARGET_DESC\n";
1469 OS << "#undef GET_SUBTARGETINFO_TARGET_DESC\n";
1470
1471 OS << "#include \"llvm/Support/Debug.h\"\n";
1472 OS << "#include \"llvm/Support/raw_ostream.h\"\n";
1473 ParseFeaturesFunction(OS, NumFeatures, NumProcs);
1474
1475 OS << "#endif // GET_SUBTARGETINFO_TARGET_DESC\n\n";
1476
1477 // Create a TargetSubtargetInfo subclass to hide the MC layer initialization.
1478 OS << "\n#ifdef GET_SUBTARGETINFO_HEADER\n";
1479 OS << "#undef GET_SUBTARGETINFO_HEADER\n";
1480
1481 std::string ClassName = Target + "GenSubtargetInfo";
1482 OS << "namespace llvm {\n";
1483 OS << "class DFAPacketizer;\n";
1484 OS << "struct " << ClassName << " : public TargetSubtargetInfo {\n"
1485 << " explicit " << ClassName << "(const Triple &TT, StringRef CPU, "
1486 << "StringRef FS);\n"
1487 << "public:\n"
1488 << " unsigned resolveSchedClass(unsigned SchedClass, "
1489 << " const MachineInstr *DefMI,"
1490 << " const TargetSchedModel *SchedModel) const override;\n"
1491 << " DFAPacketizer *createDFAPacketizer(const InstrItineraryData *IID)"
1492 << " const;\n"
1493 << "};\n";
1494 OS << "} // End llvm namespace \n";
1495
1496 OS << "#endif // GET_SUBTARGETINFO_HEADER\n\n";
1497
1498 OS << "\n#ifdef GET_SUBTARGETINFO_CTOR\n";
1499 OS << "#undef GET_SUBTARGETINFO_CTOR\n";
1500
1501 OS << "#include \"llvm/CodeGen/TargetSchedule.h\"\n";
1502 OS << "namespace llvm {\n";
1503 OS << "extern const llvm::SubtargetFeatureKV " << Target << "FeatureKV[];\n";
1504 OS << "extern const llvm::SubtargetFeatureKV " << Target << "SubTypeKV[];\n";
1505 OS << "extern const llvm::SubtargetInfoKV " << Target << "ProcSchedKV[];\n";
1506 OS << "extern const llvm::MCWriteProcResEntry "
1507 << Target << "WriteProcResTable[];\n";
1508 OS << "extern const llvm::MCWriteLatencyEntry "
1509 << Target << "WriteLatencyTable[];\n";
1510 OS << "extern const llvm::MCReadAdvanceEntry "
1511 << Target << "ReadAdvanceTable[];\n";
1512
1513 if (SchedModels.hasItineraries()) {
1514 OS << "extern const llvm::InstrStage " << Target << "Stages[];\n";
1515 OS << "extern const unsigned " << Target << "OperandCycles[];\n";
1516 OS << "extern const unsigned " << Target << "ForwardingPaths[];\n";
1517 }
1518
1519 OS << ClassName << "::" << ClassName << "(const Triple &TT, StringRef CPU, "
1520 << "StringRef FS)\n"
1521 << " : TargetSubtargetInfo() {\n"
1522 << " InitMCSubtargetInfo(TT, CPU, FS, ";
1523 if (NumFeatures)
1524 OS << "makeArrayRef(" << Target << "FeatureKV, " << NumFeatures << "), ";
1525 else
1526 OS << "None, ";
1527 if (NumProcs)
1528 OS << "makeArrayRef(" << Target << "SubTypeKV, " << NumProcs << "), ";
1529 else
1530 OS << "None, ";
1531 OS << '\n'; OS.indent(22);
1532 OS << Target << "ProcSchedKV, "
1533 << Target << "WriteProcResTable, "
1534 << Target << "WriteLatencyTable, "
1535 << Target << "ReadAdvanceTable, ";
1536 OS << '\n'; OS.indent(22);
1537 if (SchedModels.hasItineraries()) {
1538 OS << Target << "Stages, "
1539 << Target << "OperandCycles, "
1540 << Target << "ForwardingPaths";
1541 } else
1542 OS << "0, 0, 0";
1543 OS << ");\n}\n\n";
1544
1545 EmitSchedModelHelpers(ClassName, OS);
1546
1547 OS << "} // End llvm namespace \n";
1548
1549 OS << "#endif // GET_SUBTARGETINFO_CTOR\n\n";
1550 }
毫无疑问，可以根据.td文件里SubTargetFeature定义与处理器定义间的关系，自动生成一个X86Subtarget的方法来根据处理器类型自动设置这些开关。
1362 void SubtargetEmitter::ParseFeaturesFunction(raw_ostream &OS,
1363 unsigned NumFeatures,
1364 unsigned NumProcs) {
1365 std::vector<Record*> Features =
1366 Records.getAllDerivedDefinitions("SubtargetFeature");
1367 std::sort(Features.begin(), Features.end(), LessRecord());
1368
1369 OS << "// ParseSubtargetFeatures - Parses features string setting specified\n"
1370 << "// subtarget options.\n"
1371 << "void llvm::";
1372 OS << Target;
1373 OS << "Subtarget::ParseSubtargetFeatures(StringRef CPU, StringRef FS) {\n"
1374 << " DEBUG(dbgs() << \"\\nFeatures:\" << FS);\n"
1375 << " DEBUG(dbgs() << \"\\nCPU:\" << CPU << \"\\n\\n\");\n";
1376
1377 if (Features.empty()) {
1378 OS << "}\n";
1379 return;
1380 }
1381
1382 OS << " InitMCProcessorInfo(CPU, FS);\n"
1383 << " const FeatureBitset& Bits = getFeatureBits();\n";
1384
1385 for (unsigned i = 0; i < Features.size(); i++) {
1386 // Next record
1387 Record *R = Features[i];
1388 const std::string &Instance = R->getName();
1389 const std::string &Value = R->getValueAsString("Value");
1390 const std::string &Attribute = R->getValueAsString("Attribute");
1391
1392 if (Value=="true" || Value=="false")
1393 OS << " if (Bits[" << Target << "::"
1394 << Instance << "]) "
1395 << Attribute << " = " << Value << ";\n";
1396 else
1397 OS << " if (Bits[" << Target << "::"
1398 << Instance << "] && "
1399 << Attribute << " < " << Value << ") "
1400 << Attribute << " = " << Value << ";\n";
1401 }
1402
1403 OS << "}\n";
1404 }
方法ParseFeaturesFunction没有太特别的地方。所生成的X86Subtarget::ParseSubtargetFeatures方法定义如下：
#ifdef GET_SUBTARGETINFO_TARGET_DESC
#undef GET_SUBTARGETINFO_TARGET_DESC
#include "llvm/Support/Debug.h"
#include "llvm/Support/raw_ostream.h"
// ParseSubtargetFeatures - Parses features string setting specified
// subtarget options.
void llvm::X86Subtarget::ParseSubtargetFeatures(StringRef CPU, StringRef FS) {
DEBUG(dbgs() << "\nFeatures:" << FS);
DEBUG(dbgs() << "\nCPU:" << CPU << "\n\n");
InitMCProcessorInfo(CPU, FS);
const FeatureBitset& Bits = getFeatureBits();
if (Bits[X86::Feature3DNow] && X863DNowLevel < ThreeDNow) X863DNowLevel = ThreeDNow;
if (Bits[X86::Feature3DNowA] && X863DNowLevel < ThreeDNowA) X863DNowLevel = ThreeDNowA;
if (Bits[X86::Feature64Bit]) HasX86_64 = true;
if (Bits[X86::FeatureADX]) HasADX = true;
if (Bits[X86::FeatureAES]) HasAES = true;
if (Bits[X86::FeatureAVX] && X86SSELevel < AVX) X86SSELevel = AVX;
if (Bits[X86::FeatureAVX2] && X86SSELevel < AVX2) X86SSELevel = AVX2;
if (Bits[X86::FeatureAVX512] && X86SSELevel < AVX512F) X86SSELevel = AVX512F;
if (Bits[X86::FeatureBMI]) HasBMI = true;
if (Bits[X86::FeatureBMI2]) HasBMI2 = true;
if (Bits[X86::FeatureBWI]) HasBWI = true;
if (Bits[X86::FeatureCDI]) HasCDI = true;
if (Bits[X86::FeatureCMOV]) HasCMov = true;
if (Bits[X86::FeatureCMPXCHG16B]) HasCmpxchg16b = true;
if (Bits[X86::FeatureCallRegIndirect]) CallRegIndirect = true;
if (Bits[X86::FeatureDQI]) HasDQI = true;
if (Bits[X86::FeatureERI]) HasERI = true;
if (Bits[X86::FeatureF16C]) HasF16C = true;
if (Bits[X86::FeatureFMA]) HasFMA = true;
if (Bits[X86::FeatureFMA4]) HasFMA4 = true;
if (Bits[X86::FeatureFSGSBase]) HasFSGSBase = true;
if (Bits[X86::FeatureFastUAMem]) IsUAMemFast = true;
if (Bits[X86::FeatureHLE]) HasHLE = true;
if (Bits[X86::FeatureLEAUsesAG]) LEAUsesAG = true;
if (Bits[X86::FeatureLZCNT]) HasLZCNT = true;
if (Bits[X86::FeatureLeaForSP]) UseLeaForSP = true;
if (Bits[X86::FeatureMMX] && X86SSELevel < MMX) X86SSELevel = MMX;
if (Bits[X86::FeatureMOVBE]) HasMOVBE = true;
if (Bits[X86::FeatureMPX]) HasMPX = true;
if (Bits[X86::FeaturePCLMUL]) HasPCLMUL = true;
if (Bits[X86::FeaturePFI]) HasPFI = true;
if (Bits[X86::FeaturePOPCNT]) HasPOPCNT = true;
if (Bits[X86::FeaturePRFCHW]) HasPRFCHW = true;
if (Bits[X86::FeaturePadShortFunctions]) PadShortFunctions = true;
if (Bits[X86::FeatureRDRAND]) HasRDRAND = true;
if (Bits[X86::FeatureRDSEED]) HasRDSEED = true;
if (Bits[X86::FeatureRTM]) HasRTM = true;
if (Bits[X86::FeatureSHA]) HasSHA = true;
if (Bits[X86::FeatureSSE1] && X86SSELevel < SSE1) X86SSELevel = SSE1;
if (Bits[X86::FeatureSSE2] && X86SSELevel < SSE2) X86SSELevel = SSE2;
if (Bits[X86::FeatureSSE3] && X86SSELevel < SSE3) X86SSELevel = SSE3;
if (Bits[X86::FeatureSSE4A]) HasSSE4A = true;
if (Bits[X86::FeatureSSE41] && X86SSELevel < SSE41) X86SSELevel = SSE41;
if (Bits[X86::FeatureSSE42] && X86SSELevel < SSE42) X86SSELevel = SSE42;
if (Bits[X86::FeatureSSEUnalignedMem]) HasSSEUnalignedMem = true;
if (Bits[X86::FeatureSSSE3] && X86SSELevel < SSSE3) X86SSELevel = SSSE3;
if (Bits[X86::FeatureSlowBTMem]) IsBTMemSlow = true;
if (Bits[X86::FeatureSlowDivide32]) HasSlowDivide32 = true;
if (Bits[X86::FeatureSlowDivide64]) HasSlowDivide64 = true;
if (Bits[X86::FeatureSlowIncDec]) SlowIncDec = true;
if (Bits[X86::FeatureSlowLEA]) SlowLEA = true;
if (Bits[X86::FeatureSlowSHLD]) IsSHLDSlow = true;
if (Bits[X86::FeatureSlowUAMem32]) IsUAMem32Slow = true;
if (Bits[X86::FeatureSoftFloat]) UseSoftFloat = true;
if (Bits[X86::FeatureTBM]) HasTBM = true;
if (Bits[X86::FeatureVLX]) HasVLX = true;
if (Bits[X86::FeatureXOP]) HasXOP = true;
if (Bits[X86::Mode16Bit]) In16BitMode = true;
if (Bits[X86::Mode32Bit]) In32BitMode = true;
if (Bits[X86::Mode64Bit]) In64BitMode = true;
if (Bits[X86::ProcIntelAtom] && X86ProcFamily < IntelAtom) X86ProcFamily = IntelAtom;
if (Bits[X86::ProcIntelSLM] && X86ProcFamily < IntelSLM) X86ProcFamily = IntelSLM;
}
#endif // GET_SUBTARGETINFO_TARGET_DESC
X86Subtarget是MCSubtargetInfo的间接派生类，在调用ParseSubtargetFeatures方法之前，必须先通过InitMCSubtargetInfo方法（由上面生成的InitX86MCSubtargetInfo方法所调用）记录下上面的X86FeatureKV与X86SubTypeKV，接着通过InitMCProcessorInfo与getFeatureBits方法获取一个比特集，然后根据比特集里设置的比特位设置这些布尔变量。另外，InitMCProcessorInfo还会调用方法InitCPUSchedModel来设置目标机器使用的调度模型：
1478~1496行就是输出这个X86GenSubtargetInfo类型的声明：
#ifdef GET_SUBTARGETINFO_HEADER
#undef GET_SUBTARGETINFO_HEADER
namespace llvm {
class DFAPacketizer;
struct X86GenSubtargetInfo : public TargetSubtargetInfo {
explicit X86GenSubtargetInfo(const Triple &TT, StringRef CPU, StringRef FS);
public:
unsigned resolveSchedClass(unsigned SchedClass, const MachineInstr *DefMI, const TargetSchedModel *SchedModel) const override;
DFAPacketizer *createDFAPacketizer(const InstrItineraryData *IID) const;
};
} // End llvm namespace
#endif // GET_SUBTARGETINFO_HEADER
1499~1543行生成以下的代码片段：
#ifdef GET_SUBTARGETINFO_CTOR
#undef GET_SUBTARGETINFO_CTOR
#include "llvm/CodeGen/TargetSchedule.h"
namespace llvm {
extern const llvm::SubtargetFeatureKV X86FeatureKV[];
extern const llvm::SubtargetFeatureKV X86SubTypeKV[];
extern const llvm::SubtargetInfoKV X86ProcSchedKV[];
extern const llvm::MCWriteProcResEntry X86WriteProcResTable[];
extern const llvm::MCWriteLatencyEntry X86WriteLatencyTable[];
extern const llvm::MCReadAdvanceEntry X86ReadAdvanceTable[];
extern const llvm::InstrStage X86Stages[];
extern const unsigned X86OperandCycles[];
extern const unsigned X86ForwardingPaths[];
X86GenSubtargetInfo::X86GenSubtargetInfo(const Triple &TT, StringRef CPU, StringRef FS)
: TargetSubtargetInfo() {
InitMCSubtargetInfo(TT, CPU, FS, makeArrayRef(X86FeatureKV, 62), makeArrayRef(X86SubTypeKV, 65),
X86ProcSchedKV, X86WriteProcResTable, X86WriteLatencyTable, X86ReadAdvanceTable,
X86Stages, X86OperandCycles, X86ForwardingPaths);
}
因为这段代码被宏GET_SUBTARGETINFO_CTOR所选中，这时它所援引的X86FeatureKV数组等，在当前上下文不可见，因此需要extern声明。
前面我们看过，对于差异比较大的处理器，可以使用instRW与ItinRW定义将特定的指令关联到新的SchedReadWrite。这种情况下会产生新的调度类型，而原有调度类型CodeGenSchedClass对象的Transitions容器会记录下到这个新类型的迁移。
在LLVM里，这个迁移是动态发生的。前面在生成CodeGenSchedClass对应的MCSchedClassDesc实例时，如果针对特定的处理器，该调度类型应该被新调度类型替代，这个MCSchedClassDesc的NumMicroOps被设置为MCSchedClassDesc::VariantNumMicroOps。在执行是一旦发现这种形式的MCSchedClassDesc对象，就会执行动态的解析。具体实现在下面这个方法（TargetSchedule.cpp）：
101 const MCSchedClassDesc *TargetSchedModel::
102 resolveSchedClass(const MachineInstr *MI) const {
103
104 // Get the definition's scheduling class descriptor from this machine model.
105 unsigned SchedClass = MI->getDesc().getSchedClass();
106 const MCSchedClassDesc *SCDesc = SchedModel.getSchedClassDesc(SchedClass);
107 if (!SCDesc->isValid())
108 return SCDesc;
109
110 #ifndef NDEBUG
111 unsigned NIter = 0;
112 #endif
113 while (SCDesc->isVariant()) {
114 assert(++NIter < 6 && "Variants are nested deeper than the magic number");
115
116 SchedClass = STI->resolveSchedClass(SchedClass, MI, this);
117 SCDesc = SchedModel.getSchedClassDesc(SchedClass);
118 }
119 return SCDesc;
120 }
113行的isVariant方法就是检查NumMicroOps是否为MCSchedClassDesc::VariantNumMicroOps。显然调度类型之间如何替换，只能由具体的处理器给出，因此116行调用的resolveSchedClass方法就是由下面的EmitSchedModelHelpers方法生成的。
1284 void SubtargetEmitter::EmitSchedModelHelpers(std::string ClassName,
1285 raw_ostream &OS) {
1286 OS << "unsigned " << ClassName
1287 << "\n::resolveSchedClass(unsigned SchedClass, const MachineInstr *MI,"
1288 << " const TargetSchedModel *SchedModel) const {\n";
1289
1290 std::vector<Record*> Prologs = Records.getAllDerivedDefinitions("PredicateProlog");
1291 std::sort(Prologs.begin(), Prologs.end(), LessRecord());
1292 for (std::vector<Record*>::const_iterator
1293 PI = Prologs.begin(), PE = Prologs.end(); PI != PE; ++PI) {
1294 OS << (*PI)->getValueAsString("Code") << '\n';
1295 }
1296 IdxVec VariantClasses;
1297 for (CodeGenSchedModels::SchedClassIter SCI = SchedModels.schedClassBegin(),
1298 SCE = SchedModels.schedClassEnd(); SCI != SCE; ++SCI) {
1299 if (SCI->Transitions.empty())
1300 continue;
1301 VariantClasses.push_back(SCI->Index);
1302 }
1303 if (!VariantClasses.empty()) {
1304 OS << " switch (SchedClass) {\n";
1305 for (IdxIter VCI = VariantClasses.begin(), VCE = VariantClasses.end();
1306 VCI != VCE; ++VCI) {
1307 const CodeGenSchedClass &SC = SchedModels.getSchedClass(*VCI);
1308 OS << " case " << *VCI << ": // " << SC.Name << '\n';
1309 IdxVec ProcIndices;
1310 for (std::vector<CodeGenSchedTransition>::const_iterator
1311 TI = SC.Transitions.begin(), TE = SC.Transitions.end();
1312 TI != TE; ++TI) {
1313 IdxVec PI;
1314 std::set_union(TI->ProcIndices.begin(), TI->ProcIndices.end(),
1315 ProcIndices.begin(), ProcIndices.end(),
1316 std::back_inserter(PI));
1317 ProcIndices.swap(PI);
1318 }
1319 for (IdxIter PI = ProcIndices.begin(), PE = ProcIndices.end();
1320 PI != PE; ++PI) {
1321 OS << " ";
1322 if (*PI != 0)
1323 OS << "if (SchedModel->getProcessorID() == " << *PI << ") ";
1324 OS << "{ // " << (SchedModels.procModelBegin() + *PI)->ModelName
1325 << '\n';
1326 for (std::vector<CodeGenSchedTransition>::const_iterator
1327 TI = SC.Transitions.begin(), TE = SC.Transitions.end();
1328 TI != TE; ++TI) {
1329 if (*PI != 0 && !std::count(TI->ProcIndices.begin(),
1330 TI->ProcIndices.end(), *PI)) {
1331 continue;
1332 }
1333 OS << " if (";
1334 for (RecIter RI = TI->PredTerm.begin(), RE = TI->PredTerm.end();
1335 RI != RE; ++RI) {
1336 if (RI != TI->PredTerm.begin())
1337 OS << "\n && ";
1338 OS << "(" << (*RI)->getValueAsString("Predicate") << ")";
1339 }
1340 OS << ")\n"
1341 << " return " << TI->ToClassIdx << "; // "
1342 << SchedModels.getSchedClass(TI->ToClassIdx).Name << '\n';
1343 }
1344 OS << " }\n";
1345 if (*PI == 0)
1346 break;
1347 }
1348 if (SC.isInferred())
1349 OS << " return " << SC.Index << ";\n";
1350 OS << " break;\n";
1351 }
1352 OS << " };\n";
1353 }
1354 OS << " report_fatal_error(\"Expected a variant SchedClass\");\n"
1355 << "} // " << ClassName << "::resolveSchedClass\n";
1356 }
首先，处理器的resolveSchedClass方法最后参数的类型是const TargetSchedModel*，但是判定替换能否发生的谓词代码（来自SchedVar的Predicate，而SchedVar嵌入在SchedVariant中）却是来自XXXInstrInfo类的方法。为此，TableGen提供了一个PredicateProlog定义，允许目标机器提供从const TargetSchedModel*获取XXXInstrInfo实例的代码片段。
对于X86目标机器，resolveSchedClass是一个不应该调用的方法，因为X86不使用SchedVar，SchedWriteVariant以及SchedReadVariant定义
unsigned X86GenSubtargetInfo
::resolveSchedClass(unsigned SchedClass, const MachineInstr *MI, const TargetSchedModel *SchedModel) const {
report_fatal_error("Expected a variant SchedClass");
} // X86GenSubtargetInfo::resolveSchedClass
下面是ARM这个方法的一个片段（ARM使用了大量的SchedVar，SchedWriteVariant以及SchedReadVariant定义，其resolveSchedClass超过1300行）：
unsigned ARMGenSubtargetInfo
::resolveSchedClass(unsigned SchedClass, const MachineInstr *MI, const TargetSchedModel *SchedModel) const {
const ARMBaseInstrInfo *TII =
static_cast<const ARMBaseInstrInfo*>(SchedModel->getInstrInfo());
(void)TII;
switch (SchedClass) {
case 3: // IIC_iALUsr_WriteALUsi_ReadALU
if (SchedModel->getProcessorID() == 4) { // SwiftModel
if ((TII->isSwiftFastImmShift(MI)))
return 591; // SwiftWriteP01TwoCycle_ReadALU
if ((true))
return 592; // WriteALU_ReadALU
}
break;
case 4: // IIC_iALUsr_WriteALUsr_ReadALUsr
if (SchedModel->getProcessorID() == 4) { // SwiftModel
if ((TII->isPredicated(MI)))
return 593; // SwiftWriteP01ThreeCycleTwoUops_anonymous_3760
if ((true))
return 594; // SwiftWriteP01TwoCycle_NoReadAdvance
}
break;
…
};
report_fatal_error("Expected a variant SchedClass");
} // ARMGenSubtargetInfo::resolveSchedClass
在case语句中包含在“if (SchedModel->getProcessorID() == NUM”块里的if语句是对应SchedVar的中选谓词（即Predicate成员）。从上面输出的代码结构，我们就能明白SchedVariant一定需要是这个格式，最后一项的谓词一定是NoSchedPred：
def A57WriteISReg : SchedWriteVariant<[
SchedVar<RegShiftedPred, [WriteISReg]>,
SchedVar<NoSchedPred, [WriteI]>]>;
从方法EmitSchedModelHelpers返回，SubtargetEmitter::run给X86GenSubTargetInfo.inc输出最后两行代码也就结束了。