1 //===-- X86Subtarget.cpp - X86 Subtarget Information ----------------------===//
2 //
3 //                     The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 //
10 // This file implements the X86 specific subclass of TargetSubtargetInfo.
11 //
12 //===----------------------------------------------------------------------===//
13 
14 #include "X86Subtarget.h"
15 #include "MCTargetDesc/X86BaseInfo.h"
16 #include "X86TargetMachine.h"
17 #include "llvm/ADT/Triple.h"
18 #include "llvm/IR/Attributes.h"
19 #include "llvm/IR/ConstantRange.h"
20 #include "llvm/IR/Function.h"
21 #include "llvm/IR/GlobalValue.h"
22 #include "llvm/Support/Casting.h"
23 #include "llvm/Support/CodeGen.h"
24 #include "llvm/Support/CommandLine.h"
25 #include "llvm/Support/Debug.h"
26 #include "llvm/Support/ErrorHandling.h"
27 #include "llvm/Support/raw_ostream.h"
28 #include "llvm/Target/TargetMachine.h"
29 #include <cassert>
30 #include <string>
31 
32 #if defined(_MSC_VER)
33 #include <intrin.h>
34 #endif
35 
36 using namespace llvm;
37 
38 #define DEBUG_TYPE "subtarget"
39 
40 #define GET_SUBTARGETINFO_TARGET_DESC
41 #define GET_SUBTARGETINFO_CTOR
42 #include "X86GenSubtargetInfo.inc"
43 
44 // Temporary option to control early if-conversion for x86 while adding machine
45 // models.
46 static cl::opt<bool>
47 X86EarlyIfConv("x86-early-ifcvt", cl::Hidden,
48                cl::desc("Enable early if-conversion on X86"));
49 
50 
51 /// Classify a blockaddress reference for the current subtarget according to how
52 /// we should reference it in a non-pcrel context.
53 unsigned char X86Subtarget::classifyBlockAddressReference() const {
54   return classifyLocalReference(nullptr);
55 }
56 
57 /// Classify a global variable reference for the current subtarget according to
58 /// how we should reference it in a non-pcrel context.
59 unsigned char
60 X86Subtarget::classifyGlobalReference(const GlobalValue *GV) const {
61   return classifyGlobalReference(GV, *GV->getParent());
62 }
63 
64 unsigned char
65 X86Subtarget::classifyLocalReference(const GlobalValue *GV) const {
66   // 64 bits can use %rip addressing for anything local.
67   if (is64Bit())
68     return X86II::MO_NO_FLAG;
69 
70   // If this is for a position dependent executable, the static linker can
71   // figure it out.
72   if (!isPositionIndependent())
73     return X86II::MO_NO_FLAG;
74 
75   // The COFF dynamic linker just patches the executable sections.
76   if (isTargetCOFF())
77     return X86II::MO_NO_FLAG;
78 
79   if (isTargetDarwin()) {
80     // 32 bit macho has no relocation for a-b if a is undefined, even if
81     // b is in the section that is being relocated.
82     // This means we have to use o load even for GVs that are known to be
83     // local to the dso.
84     if (GV && (GV->isDeclarationForLinker() || GV->hasCommonLinkage()))
85       return X86II::MO_DARWIN_NONLAZY_PIC_BASE;
86 
87     return X86II::MO_PIC_BASE_OFFSET;
88   }
89 
90   return X86II::MO_GOTOFF;
91 }
92 
93 unsigned char X86Subtarget::classifyGlobalReference(const GlobalValue *GV,
94                                                     const Module &M) const {
95   // Large model never uses stubs.
96   if (TM.getCodeModel() == CodeModel::Large)
97     return X86II::MO_NO_FLAG;
98 
99   // Absolute symbols can be referenced directly.
100   if (GV) {
101     if (Optional<ConstantRange> CR = GV->getAbsoluteSymbolRange()) {
102       // See if we can use the 8-bit immediate form. Note that some instructions
103       // will sign extend the immediate operand, so to be conservative we only
104       // accept the range [0,128).
105       if (CR->getUnsignedMax().ult(128))
106         return X86II::MO_ABS8;
107       else
108         return X86II::MO_NO_FLAG;
109     }
110   }
111 
112   if (TM.shouldAssumeDSOLocal(M, GV))
113     return classifyLocalReference(GV);
114 
115   if (isTargetCOFF())
116     return X86II::MO_DLLIMPORT;
117 
118   if (is64Bit())
119     return X86II::MO_GOTPCREL;
120 
121   if (isTargetDarwin()) {
122     if (!isPositionIndependent())
123       return X86II::MO_DARWIN_NONLAZY;
124     return X86II::MO_DARWIN_NONLAZY_PIC_BASE;
125   }
126 
127   return X86II::MO_GOT;
128 }
129 
130 unsigned char
131 X86Subtarget::classifyGlobalFunctionReference(const GlobalValue *GV) const {
132   return classifyGlobalFunctionReference(GV, *GV->getParent());
133 }
134 
135 unsigned char
136 X86Subtarget::classifyGlobalFunctionReference(const GlobalValue *GV,
137                                               const Module &M) const {
138   if (TM.shouldAssumeDSOLocal(M, GV))
139     return X86II::MO_NO_FLAG;
140 
141   assert(!isTargetCOFF());
142   const Function *F = dyn_cast_or_null<Function>(GV);
143 
144   if (isTargetELF()) {
145     if (is64Bit() && F && (CallingConv::X86_RegCall == F->getCallingConv()))
146       // According to psABI, PLT stub clobbers XMM8-XMM15.
147       // In Regcall calling convention those registers are used for passing
148       // parameters. Thus we need to prevent lazy binding in Regcall.
149       return X86II::MO_GOTPCREL;
150     return X86II::MO_PLT;
151   }
152 
153   if (is64Bit()) {
154     if (F && F->hasFnAttribute(Attribute::NonLazyBind))
155       // If the function is marked as non-lazy, generate an indirect call
156       // which loads from the GOT directly. This avoids runtime overhead
157       // at the cost of eager binding (and one extra byte of encoding).
158       return X86II::MO_GOTPCREL;
159     return X86II::MO_NO_FLAG;
160   }
161 
162   return X86II::MO_NO_FLAG;
163 }
164 
165 /// This function returns the name of a function which has an interface like
166 /// the non-standard bzero function, if such a function exists on the
167 /// current subtarget and it is considered preferable over memset with zero
168 /// passed as the second argument. Otherwise it returns null.
169 const char *X86Subtarget::getBZeroEntry() const {
170   // Darwin 10 has a __bzero entry point for this purpose.
171   if (getTargetTriple().isMacOSX() &&
172       !getTargetTriple().isMacOSXVersionLT(10, 6))
173     return "__bzero";
174 
175   return nullptr;
176 }
177 
178 bool X86Subtarget::hasSinCos() const {
179   return getTargetTriple().isMacOSX() &&
180     !getTargetTriple().isMacOSXVersionLT(10, 9) &&
181     is64Bit();
182 }
183 
184 /// Return true if the subtarget allows calls to immediate address.
185 bool X86Subtarget::isLegalToCallImmediateAddr() const {
186   // FIXME: I386 PE/COFF supports PC relative calls using IMAGE_REL_I386_REL32
187   // but WinCOFFObjectWriter::RecordRelocation cannot emit them.  Once it does,
188   // the following check for Win32 should be removed.
189   if (In64BitMode || isTargetWin32())
190     return false;
191   return isTargetELF() || TM.getRelocationModel() == Reloc::Static;
192 }
193 
194 void X86Subtarget::initSubtargetFeatures(StringRef CPU, StringRef FS) {
195   std::string CPUName = CPU;
196   if (CPUName.empty())
197     CPUName = "generic";
198 
199   // Make sure 64-bit features are available in 64-bit mode. (But make sure
200   // SSE2 can be turned off explicitly.)
201   std::string FullFS = FS;
202   if (In64BitMode) {
203     if (!FullFS.empty())
204       FullFS = "+64bit,+sse2," + FullFS;
205     else
206       FullFS = "+64bit,+sse2";
207   }
208 
209   // LAHF/SAHF are always supported in non-64-bit mode.
210   if (!In64BitMode) {
211     if (!FullFS.empty())
212       FullFS = "+sahf," + FullFS;
213     else
214       FullFS = "+sahf";
215   }
216 
217   // Parse features string and set the CPU.
218   ParseSubtargetFeatures(CPUName, FullFS);
219 
220   // All CPUs that implement SSE4.2 or SSE4A support unaligned accesses of
221   // 16-bytes and under that are reasonably fast. These features were
222   // introduced with Intel's Nehalem/Silvermont and AMD's Family10h
223   // micro-architectures respectively.
224   if (hasSSE42() || hasSSE4A())
225     IsUAMem16Slow = false;
226 
227   InstrItins = getInstrItineraryForCPU(CPUName);
228 
229   // It's important to keep the MCSubtargetInfo feature bits in sync with
230   // target data structure which is shared with MC code emitter, etc.
231   if (In64BitMode)
232     ToggleFeature(X86::Mode64Bit);
233   else if (In32BitMode)
234     ToggleFeature(X86::Mode32Bit);
235   else if (In16BitMode)
236     ToggleFeature(X86::Mode16Bit);
237   else
238     llvm_unreachable("Not 16-bit, 32-bit or 64-bit mode!");
239 
240   DEBUG(dbgs() << "Subtarget features: SSELevel " << X86SSELevel
241                << ", 3DNowLevel " << X863DNowLevel
242                << ", 64bit " << HasX86_64 << "\n");
243   assert((!In64BitMode || HasX86_64) &&
244          "64-bit code requested on a subtarget that doesn't support it!");
245 
246   // Stack alignment is 16 bytes on Darwin, Linux, kFreeBSD and Solaris (both
247   // 32 and 64 bit) and for all 64-bit targets.
248   if (StackAlignOverride)
249     stackAlignment = StackAlignOverride;
250   else if (isTargetDarwin() || isTargetLinux() || isTargetSolaris() ||
251            isTargetKFreeBSD() || In64BitMode)
252     stackAlignment = 16;
253 }
254 
255 void X86Subtarget::initializeEnvironment() {
256   X86SSELevel = NoSSE;
257   X863DNowLevel = NoThreeDNow;
258   HasX87 = false;
259   HasCMov = false;
260   HasX86_64 = false;
261   HasPOPCNT = false;
262   HasSSE4A = false;
263   HasAES = false;
264   HasFXSR = false;
265   HasXSAVE = false;
266   HasXSAVEOPT = false;
267   HasXSAVEC = false;
268   HasXSAVES = false;
269   HasPCLMUL = false;
270   HasFMA = false;
271   HasFMA4 = false;
272   HasXOP = false;
273   HasTBM = false;
274   HasLWP = false;
275   HasMOVBE = false;
276   HasRDRAND = false;
277   HasF16C = false;
278   HasFSGSBase = false;
279   HasLZCNT = false;
280   HasBMI = false;
281   HasBMI2 = false;
282   HasVBMI = false;
283   HasIFMA = false;
284   HasRTM = false;
285   HasERI = false;
286   HasCDI = false;
287   HasPFI = false;
288   HasDQI = false;
289   HasVPOPCNTDQ = false;
290   HasBWI = false;
291   HasVLX = false;
292   HasADX = false;
293   HasPKU = false;
294   HasSHA = false;
295   HasPRFCHW = false;
296   HasRDSEED = false;
297   HasLAHFSAHF = false;
298   HasMWAITX = false;
299   HasCLZERO = false;
300   HasMPX = false;
301   HasSGX = false;
302   HasCLFLUSHOPT = false;
303   HasCLWB = false;
304   IsBTMemSlow = false;
305   IsPMULLDSlow = false;
306   IsSHLDSlow = false;
307   IsUAMem16Slow = false;
308   IsUAMem32Slow = false;
309   HasSSEUnalignedMem = false;
310   HasCmpxchg16b = false;
311   UseLeaForSP = false;
312   HasFastPartialYMMorZMMWrite = false;
313   HasFastScalarFSQRT = false;
314   HasFastVectorFSQRT = false;
315   HasFastLZCNT = false;
316   HasFastSHLDRotate = false;
317   HasERMSB = false;
318   HasSlowDivide32 = false;
319   HasSlowDivide64 = false;
320   PadShortFunctions = false;
321   CallRegIndirect = false;
322   LEAUsesAG = false;
323   SlowLEA = false;
324   Slow3OpsLEA = false;
325   SlowIncDec = false;
326   stackAlignment = 4;
327   // FIXME: this is a known good value for Yonah. How about others?
328   MaxInlineSizeThreshold = 128;
329   UseSoftFloat = false;
330 }
331 
332 X86Subtarget &X86Subtarget::initializeSubtargetDependencies(StringRef CPU,
333                                                             StringRef FS) {
334   initializeEnvironment();
335   initSubtargetFeatures(CPU, FS);
336   return *this;
337 }
338 
339 X86Subtarget::X86Subtarget(const Triple &TT, StringRef CPU, StringRef FS,
340                            const X86TargetMachine &TM,
341                            unsigned StackAlignOverride)
342     : X86GenSubtargetInfo(TT, CPU, FS), X86ProcFamily(Others),
343       PICStyle(PICStyles::None), TM(TM), TargetTriple(TT),
344       StackAlignOverride(StackAlignOverride),
345       In64BitMode(TargetTriple.getArch() == Triple::x86_64),
346       In32BitMode(TargetTriple.getArch() == Triple::x86 &&
347                   TargetTriple.getEnvironment() != Triple::CODE16),
348       In16BitMode(TargetTriple.getArch() == Triple::x86 &&
349                   TargetTriple.getEnvironment() == Triple::CODE16),
350       InstrInfo(initializeSubtargetDependencies(CPU, FS)), TLInfo(TM, *this),
351       FrameLowering(*this, getStackAlignment()) {
352   // Determine the PICStyle based on the target selected.
353   if (!isPositionIndependent())
354     setPICStyle(PICStyles::None);
355   else if (is64Bit())
356     setPICStyle(PICStyles::RIPRel);
357   else if (isTargetCOFF())
358     setPICStyle(PICStyles::None);
359   else if (isTargetDarwin())
360     setPICStyle(PICStyles::StubPIC);
361   else if (isTargetELF())
362     setPICStyle(PICStyles::GOT);
363 }
364 
365 const CallLowering *X86Subtarget::getCallLowering() const {
366   assert(GISel && "Access to GlobalISel APIs not set");
367   return GISel->getCallLowering();
368 }
369 
370 const InstructionSelector *X86Subtarget::getInstructionSelector() const {
371   assert(GISel && "Access to GlobalISel APIs not set");
372   return GISel->getInstructionSelector();
373 }
374 
375 const LegalizerInfo *X86Subtarget::getLegalizerInfo() const {
376   assert(GISel && "Access to GlobalISel APIs not set");
377   return GISel->getLegalizerInfo();
378 }
379 
380 const RegisterBankInfo *X86Subtarget::getRegBankInfo() const {
381   assert(GISel && "Access to GlobalISel APIs not set");
382   return GISel->getRegBankInfo();
383 }
384 
385 bool X86Subtarget::enableEarlyIfConversion() const {
386   return hasCMov() && X86EarlyIfConv;
387 }
388