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 "X86InstrInfo.h"
16 #include "X86TargetMachine.h"
17 #include "llvm/IR/Attributes.h"
18 #include "llvm/IR/Function.h"
19 #include "llvm/IR/GlobalValue.h"
20 #include "llvm/Support/CommandLine.h"
21 #include "llvm/Support/Debug.h"
22 #include "llvm/Support/ErrorHandling.h"
23 #include "llvm/Support/Host.h"
24 #include "llvm/Support/raw_ostream.h"
25 #include "llvm/Target/TargetMachine.h"
26 #include "llvm/Target/TargetOptions.h"
27
28 #if defined(_MSC_VER)
29 #include <intrin.h>
30 #endif
31
32 using namespace llvm;
33
34 #define DEBUG_TYPE "subtarget"
35
36 #define GET_SUBTARGETINFO_TARGET_DESC
37 #define GET_SUBTARGETINFO_CTOR
38 #include "X86GenSubtargetInfo.inc"
39
40 // Temporary option to control early if-conversion for x86 while adding machine
41 // models.
42 static cl::opt<bool>
43 X86EarlyIfConv("x86-early-ifcvt", cl::Hidden,
44 cl::desc("Enable early if-conversion on X86"));
45
46
47 /// Classify a blockaddress reference for the current subtarget according to how
48 /// we should reference it in a non-pcrel context.
classifyBlockAddressReference() const49 unsigned char X86Subtarget::classifyBlockAddressReference() const {
50 return classifyLocalReference(nullptr);
51 }
52
53 /// Classify a global variable reference for the current subtarget according to
54 /// how we should reference it in a non-pcrel context.
55 unsigned char
classifyGlobalReference(const GlobalValue * GV) const56 X86Subtarget::classifyGlobalReference(const GlobalValue *GV) const {
57 return classifyGlobalReference(GV, *GV->getParent());
58 }
59
60 unsigned char
classifyLocalReference(const GlobalValue * GV) const61 X86Subtarget::classifyLocalReference(const GlobalValue *GV) const {
62 // 64 bits can use %rip addressing for anything local.
63 if (is64Bit())
64 return X86II::MO_NO_FLAG;
65
66 // If this is for a position dependent executable, the static linker can
67 // figure it out.
68 if (!isPositionIndependent())
69 return X86II::MO_NO_FLAG;
70
71 // The COFF dynamic linker just patches the executable sections.
72 if (isTargetCOFF())
73 return X86II::MO_NO_FLAG;
74
75 if (isTargetDarwin()) {
76 // 32 bit macho has no relocation for a-b if a is undefined, even if
77 // b is in the section that is being relocated.
78 // This means we have to use o load even for GVs that are known to be
79 // local to the dso.
80 if (GV && (GV->isDeclarationForLinker() || GV->hasCommonLinkage()))
81 return X86II::MO_DARWIN_NONLAZY_PIC_BASE;
82
83 return X86II::MO_PIC_BASE_OFFSET;
84 }
85
86 return X86II::MO_GOTOFF;
87 }
88
classifyGlobalReference(const GlobalValue * GV,const Module & M) const89 unsigned char X86Subtarget::classifyGlobalReference(const GlobalValue *GV,
90 const Module &M) const {
91 // Large model never uses stubs.
92 if (TM.getCodeModel() == CodeModel::Large)
93 return X86II::MO_NO_FLAG;
94
95 if (TM.shouldAssumeDSOLocal(M, GV))
96 return classifyLocalReference(GV);
97
98 if (isTargetCOFF())
99 return X86II::MO_DLLIMPORT;
100
101 if (is64Bit())
102 return X86II::MO_GOTPCREL;
103
104 if (isTargetDarwin()) {
105 if (!isPositionIndependent())
106 return X86II::MO_DARWIN_NONLAZY;
107 return X86II::MO_DARWIN_NONLAZY_PIC_BASE;
108 }
109
110 return X86II::MO_GOT;
111 }
112
113 unsigned char
classifyGlobalFunctionReference(const GlobalValue * GV) const114 X86Subtarget::classifyGlobalFunctionReference(const GlobalValue *GV) const {
115 return classifyGlobalFunctionReference(GV, *GV->getParent());
116 }
117
118 unsigned char
classifyGlobalFunctionReference(const GlobalValue * GV,const Module & M) const119 X86Subtarget::classifyGlobalFunctionReference(const GlobalValue *GV,
120 const Module &M) const {
121 if (TM.shouldAssumeDSOLocal(M, GV))
122 return X86II::MO_NO_FLAG;
123
124 assert(!isTargetCOFF());
125
126 if (isTargetELF())
127 return X86II::MO_PLT;
128
129 if (is64Bit()) {
130 auto *F = dyn_cast_or_null<Function>(GV);
131 if (F && F->hasFnAttribute(Attribute::NonLazyBind))
132 // If the function is marked as non-lazy, generate an indirect call
133 // which loads from the GOT directly. This avoids runtime overhead
134 // at the cost of eager binding (and one extra byte of encoding).
135 return X86II::MO_GOTPCREL;
136 return X86II::MO_NO_FLAG;
137 }
138
139 return X86II::MO_NO_FLAG;
140 }
141
142 /// This function returns the name of a function which has an interface like
143 /// the non-standard bzero function, if such a function exists on the
144 /// current subtarget and it is considered preferable over memset with zero
145 /// passed as the second argument. Otherwise it returns null.
getBZeroEntry() const146 const char *X86Subtarget::getBZeroEntry() const {
147 // Darwin 10 has a __bzero entry point for this purpose.
148 if (getTargetTriple().isMacOSX() &&
149 !getTargetTriple().isMacOSXVersionLT(10, 6))
150 return "__bzero";
151
152 return nullptr;
153 }
154
hasSinCos() const155 bool X86Subtarget::hasSinCos() const {
156 return getTargetTriple().isMacOSX() &&
157 !getTargetTriple().isMacOSXVersionLT(10, 9) &&
158 is64Bit();
159 }
160
161 /// Return true if the subtarget allows calls to immediate address.
isLegalToCallImmediateAddr() const162 bool X86Subtarget::isLegalToCallImmediateAddr() const {
163 // FIXME: I386 PE/COFF supports PC relative calls using IMAGE_REL_I386_REL32
164 // but WinCOFFObjectWriter::RecordRelocation cannot emit them. Once it does,
165 // the following check for Win32 should be removed.
166 if (In64BitMode || isTargetWin32())
167 return false;
168 return isTargetELF() || TM.getRelocationModel() == Reloc::Static;
169 }
170
initSubtargetFeatures(StringRef CPU,StringRef FS)171 void X86Subtarget::initSubtargetFeatures(StringRef CPU, StringRef FS) {
172 std::string CPUName = CPU;
173 if (CPUName.empty())
174 CPUName = "generic";
175
176 // Make sure 64-bit features are available in 64-bit mode. (But make sure
177 // SSE2 can be turned off explicitly.)
178 std::string FullFS = FS;
179 if (In64BitMode) {
180 if (!FullFS.empty())
181 FullFS = "+64bit,+sse2," + FullFS;
182 else
183 FullFS = "+64bit,+sse2";
184 }
185
186 // LAHF/SAHF are always supported in non-64-bit mode.
187 if (!In64BitMode) {
188 if (!FullFS.empty())
189 FullFS = "+sahf," + FullFS;
190 else
191 FullFS = "+sahf";
192 }
193
194
195 // Parse features string and set the CPU.
196 ParseSubtargetFeatures(CPUName, FullFS);
197
198 // All CPUs that implement SSE4.2 or SSE4A support unaligned accesses of
199 // 16-bytes and under that are reasonably fast. These features were
200 // introduced with Intel's Nehalem/Silvermont and AMD's Family10h
201 // micro-architectures respectively.
202 if (hasSSE42() || hasSSE4A())
203 IsUAMem16Slow = false;
204
205 InstrItins = getInstrItineraryForCPU(CPUName);
206
207 // It's important to keep the MCSubtargetInfo feature bits in sync with
208 // target data structure which is shared with MC code emitter, etc.
209 if (In64BitMode)
210 ToggleFeature(X86::Mode64Bit);
211 else if (In32BitMode)
212 ToggleFeature(X86::Mode32Bit);
213 else if (In16BitMode)
214 ToggleFeature(X86::Mode16Bit);
215 else
216 llvm_unreachable("Not 16-bit, 32-bit or 64-bit mode!");
217
218 DEBUG(dbgs() << "Subtarget features: SSELevel " << X86SSELevel
219 << ", 3DNowLevel " << X863DNowLevel
220 << ", 64bit " << HasX86_64 << "\n");
221 assert((!In64BitMode || HasX86_64) &&
222 "64-bit code requested on a subtarget that doesn't support it!");
223
224 // Stack alignment is 16 bytes on Darwin, Linux, kFreeBSD and Solaris (both
225 // 32 and 64 bit) and for all 64-bit targets.
226 if (StackAlignOverride)
227 stackAlignment = StackAlignOverride;
228 else if (isTargetDarwin() || isTargetLinux() || isTargetSolaris() ||
229 isTargetKFreeBSD() || In64BitMode)
230 stackAlignment = 16;
231 }
232
initializeEnvironment()233 void X86Subtarget::initializeEnvironment() {
234 X86SSELevel = NoSSE;
235 X863DNowLevel = NoThreeDNow;
236 HasX87 = false;
237 HasCMov = false;
238 HasX86_64 = false;
239 HasPOPCNT = false;
240 HasSSE4A = false;
241 HasAES = false;
242 HasFXSR = false;
243 HasXSAVE = false;
244 HasXSAVEOPT = false;
245 HasXSAVEC = false;
246 HasXSAVES = false;
247 HasPCLMUL = false;
248 HasFMA = false;
249 HasFMA4 = false;
250 HasXOP = false;
251 HasTBM = false;
252 HasMOVBE = false;
253 HasRDRAND = false;
254 HasF16C = false;
255 HasFSGSBase = false;
256 HasLZCNT = false;
257 HasBMI = false;
258 HasBMI2 = false;
259 HasVBMI = false;
260 HasIFMA = false;
261 HasRTM = false;
262 HasHLE = false;
263 HasERI = false;
264 HasCDI = false;
265 HasPFI = false;
266 HasDQI = false;
267 HasBWI = false;
268 HasVLX = false;
269 HasADX = false;
270 HasPKU = false;
271 HasSHA = false;
272 HasPRFCHW = false;
273 HasRDSEED = false;
274 HasLAHFSAHF = false;
275 HasMWAITX = false;
276 HasMPX = false;
277 IsBTMemSlow = false;
278 IsSHLDSlow = false;
279 IsUAMem16Slow = false;
280 IsUAMem32Slow = false;
281 HasSSEUnalignedMem = false;
282 HasCmpxchg16b = false;
283 UseLeaForSP = false;
284 HasFastPartialYMMWrite = false;
285 HasSlowDivide32 = false;
286 HasSlowDivide64 = false;
287 PadShortFunctions = false;
288 CallRegIndirect = false;
289 LEAUsesAG = false;
290 SlowLEA = false;
291 SlowIncDec = false;
292 stackAlignment = 4;
293 // FIXME: this is a known good value for Yonah. How about others?
294 MaxInlineSizeThreshold = 128;
295 UseSoftFloat = false;
296 }
297
initializeSubtargetDependencies(StringRef CPU,StringRef FS)298 X86Subtarget &X86Subtarget::initializeSubtargetDependencies(StringRef CPU,
299 StringRef FS) {
300 initializeEnvironment();
301 initSubtargetFeatures(CPU, FS);
302 return *this;
303 }
304
X86Subtarget(const Triple & TT,StringRef CPU,StringRef FS,const X86TargetMachine & TM,unsigned StackAlignOverride)305 X86Subtarget::X86Subtarget(const Triple &TT, StringRef CPU, StringRef FS,
306 const X86TargetMachine &TM,
307 unsigned StackAlignOverride)
308 : X86GenSubtargetInfo(TT, CPU, FS), X86ProcFamily(Others),
309 PICStyle(PICStyles::None), TM(TM), TargetTriple(TT),
310 StackAlignOverride(StackAlignOverride),
311 In64BitMode(TargetTriple.getArch() == Triple::x86_64),
312 In32BitMode(TargetTriple.getArch() == Triple::x86 &&
313 TargetTriple.getEnvironment() != Triple::CODE16),
314 In16BitMode(TargetTriple.getArch() == Triple::x86 &&
315 TargetTriple.getEnvironment() == Triple::CODE16),
316 TSInfo(), InstrInfo(initializeSubtargetDependencies(CPU, FS)),
317 TLInfo(TM, *this), FrameLowering(*this, getStackAlignment()) {
318 // Determine the PICStyle based on the target selected.
319 if (!isPositionIndependent())
320 setPICStyle(PICStyles::None);
321 else if (is64Bit())
322 setPICStyle(PICStyles::RIPRel);
323 else if (isTargetCOFF())
324 setPICStyle(PICStyles::None);
325 else if (isTargetDarwin())
326 setPICStyle(PICStyles::StubPIC);
327 else if (isTargetELF())
328 setPICStyle(PICStyles::GOT);
329 }
330
enableEarlyIfConversion() const331 bool X86Subtarget::enableEarlyIfConversion() const {
332 return hasCMov() && X86EarlyIfConv;
333 }
334
335