1 //===-- X86Subtarget.cpp - X86 Subtarget Information ----------------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file implements the X86 specific subclass of TargetSubtargetInfo.
12 //===----------------------------------------------------------------------===//
14 #define DEBUG_TYPE "subtarget"
15 #include "X86Subtarget.h"
16 #include "X86InstrInfo.h"
17 #include "llvm/GlobalValue.h"
18 #include "llvm/Support/Debug.h"
19 #include "llvm/Support/ErrorHandling.h"
20 #include "llvm/Support/raw_ostream.h"
21 #include "llvm/Support/Host.h"
22 #include "llvm/Target/TargetMachine.h"
23 #include "llvm/Target/TargetOptions.h"
25 #define GET_SUBTARGETINFO_TARGET_DESC
26 #define GET_SUBTARGETINFO_CTOR
27 #include "X86GenSubtargetInfo.inc"
35 /// ClassifyBlockAddressReference - Classify a blockaddress reference for the
36 /// current subtarget according to how we should reference it in a non-pcrel
38 unsigned char X86Subtarget::
39 ClassifyBlockAddressReference() const {
40 if (isPICStyleGOT()) // 32-bit ELF targets.
41 return X86II::MO_GOTOFF;
43 if (isPICStyleStubPIC()) // Darwin/32 in PIC mode.
44 return X86II::MO_PIC_BASE_OFFSET;
46 // Direct static reference to label.
47 return X86II::MO_NO_FLAG;
50 /// ClassifyGlobalReference - Classify a global variable reference for the
51 /// current subtarget according to how we should reference it in a non-pcrel
53 unsigned char X86Subtarget::
54 ClassifyGlobalReference(const GlobalValue *GV, const TargetMachine &TM) const {
55 // DLLImport only exists on windows, it is implemented as a load from a
57 if (GV->hasDLLImportLinkage())
58 return X86II::MO_DLLIMPORT;
60 // Determine whether this is a reference to a definition or a declaration.
61 // Materializable GVs (in JIT lazy compilation mode) do not require an extra
63 bool isDecl = GV->hasAvailableExternallyLinkage();
64 if (GV->isDeclaration() && !GV->isMaterializable())
67 // X86-64 in PIC mode.
68 if (isPICStyleRIPRel()) {
69 // Large model never uses stubs.
70 if (TM.getCodeModel() == CodeModel::Large)
71 return X86II::MO_NO_FLAG;
73 if (isTargetDarwin()) {
74 // If symbol visibility is hidden, the extra load is not needed if
75 // target is x86-64 or the symbol is definitely defined in the current
77 if (GV->hasDefaultVisibility() &&
78 (isDecl || GV->isWeakForLinker()))
79 return X86II::MO_GOTPCREL;
80 } else if (!isTargetWin64()) {
81 assert(isTargetELF() && "Unknown rip-relative target");
83 // Extra load is needed for all externally visible.
84 if (!GV->hasLocalLinkage() && GV->hasDefaultVisibility())
85 return X86II::MO_GOTPCREL;
88 return X86II::MO_NO_FLAG;
91 if (isPICStyleGOT()) { // 32-bit ELF targets.
92 // Extra load is needed for all externally visible.
93 if (GV->hasLocalLinkage() || GV->hasHiddenVisibility())
94 return X86II::MO_GOTOFF;
98 if (isPICStyleStubPIC()) { // Darwin/32 in PIC mode.
99 // Determine whether we have a stub reference and/or whether the reference
100 // is relative to the PIC base or not.
102 // If this is a strong reference to a definition, it is definitely not
104 if (!isDecl && !GV->isWeakForLinker())
105 return X86II::MO_PIC_BASE_OFFSET;
107 // Unless we have a symbol with hidden visibility, we have to go through a
108 // normal $non_lazy_ptr stub because this symbol might be resolved late.
109 if (!GV->hasHiddenVisibility()) // Non-hidden $non_lazy_ptr reference.
110 return X86II::MO_DARWIN_NONLAZY_PIC_BASE;
112 // If symbol visibility is hidden, we have a stub for common symbol
113 // references and external declarations.
114 if (isDecl || GV->hasCommonLinkage()) {
115 // Hidden $non_lazy_ptr reference.
116 return X86II::MO_DARWIN_HIDDEN_NONLAZY_PIC_BASE;
119 // Otherwise, no stub.
120 return X86II::MO_PIC_BASE_OFFSET;
123 if (isPICStyleStubNoDynamic()) { // Darwin/32 in -mdynamic-no-pic mode.
124 // Determine whether we have a stub reference.
126 // If this is a strong reference to a definition, it is definitely not
128 if (!isDecl && !GV->isWeakForLinker())
129 return X86II::MO_NO_FLAG;
131 // Unless we have a symbol with hidden visibility, we have to go through a
132 // normal $non_lazy_ptr stub because this symbol might be resolved late.
133 if (!GV->hasHiddenVisibility()) // Non-hidden $non_lazy_ptr reference.
134 return X86II::MO_DARWIN_NONLAZY;
136 // Otherwise, no stub.
137 return X86II::MO_NO_FLAG;
140 // Direct static reference to global.
141 return X86II::MO_NO_FLAG;
145 /// getBZeroEntry - This function returns the name of a function which has an
146 /// interface like the non-standard bzero function, if such a function exists on
147 /// the current subtarget and it is considered prefereable over memset with zero
148 /// passed as the second argument. Otherwise it returns null.
149 const char *X86Subtarget::getBZeroEntry() const {
150 // Darwin 10 has a __bzero entry point for this purpose.
151 if (getTargetTriple().isMacOSX() &&
152 !getTargetTriple().isMacOSXVersionLT(10, 6))
158 /// IsLegalToCallImmediateAddr - Return true if the subtarget allows calls
159 /// to immediate address.
160 bool X86Subtarget::IsLegalToCallImmediateAddr(const TargetMachine &TM) const {
163 return isTargetELF() || TM.getRelocationModel() == Reloc::Static;
166 /// getSpecialAddressLatency - For targets where it is beneficial to
167 /// backschedule instructions that compute addresses, return a value
168 /// indicating the number of scheduling cycles of backscheduling that
169 /// should be attempted.
170 unsigned X86Subtarget::getSpecialAddressLatency() const {
171 // For x86 out-of-order targets, back-schedule address computations so
172 // that loads and stores aren't blocked.
173 // This value was chosen arbitrarily.
177 void X86Subtarget::AutoDetectSubtargetFeatures() {
178 unsigned EAX = 0, EBX = 0, ECX = 0, EDX = 0;
185 if (X86_MC::GetCpuIDAndInfo(0, &MaxLevel, text.u+0, text.u+2, text.u+1) ||
189 X86_MC::GetCpuIDAndInfo(0x1, &EAX, &EBX, &ECX, &EDX);
191 if ((EDX >> 15) & 1) { HasCMov = true; ToggleFeature(X86::FeatureCMOV); }
192 if ((EDX >> 23) & 1) { X86SSELevel = MMX; ToggleFeature(X86::FeatureMMX); }
193 if ((EDX >> 25) & 1) { X86SSELevel = SSE1; ToggleFeature(X86::FeatureSSE1); }
194 if ((EDX >> 26) & 1) { X86SSELevel = SSE2; ToggleFeature(X86::FeatureSSE2); }
195 if (ECX & 0x1) { X86SSELevel = SSE3; ToggleFeature(X86::FeatureSSE3); }
196 if ((ECX >> 9) & 1) { X86SSELevel = SSSE3; ToggleFeature(X86::FeatureSSSE3);}
197 if ((ECX >> 19) & 1) { X86SSELevel = SSE41; ToggleFeature(X86::FeatureSSE41);}
198 if ((ECX >> 20) & 1) { X86SSELevel = SSE42; ToggleFeature(X86::FeatureSSE42);}
199 if ((ECX >> 28) & 1) { X86SSELevel = AVX; ToggleFeature(X86::FeatureAVX); }
201 bool IsIntel = memcmp(text.c, "GenuineIntel", 12) == 0;
202 bool IsAMD = !IsIntel && memcmp(text.c, "AuthenticAMD", 12) == 0;
204 if ((ECX >> 1) & 0x1) {
206 ToggleFeature(X86::FeaturePCLMUL);
208 if ((ECX >> 12) & 0x1) {
210 ToggleFeature(X86::FeatureFMA);
212 if (IsIntel && ((ECX >> 22) & 0x1)) {
214 ToggleFeature(X86::FeatureMOVBE);
216 if ((ECX >> 23) & 0x1) {
218 ToggleFeature(X86::FeaturePOPCNT);
220 if ((ECX >> 25) & 0x1) {
222 ToggleFeature(X86::FeatureAES);
224 if ((ECX >> 29) & 0x1) {
226 ToggleFeature(X86::FeatureF16C);
228 if (IsIntel && ((ECX >> 30) & 0x1)) {
230 ToggleFeature(X86::FeatureRDRAND);
233 if ((ECX >> 13) & 0x1) {
234 HasCmpxchg16b = true;
235 ToggleFeature(X86::FeatureCMPXCHG16B);
238 if (IsIntel || IsAMD) {
239 // Determine if bit test memory instructions are slow.
242 X86_MC::DetectFamilyModel(EAX, Family, Model);
243 if (IsAMD || (Family == 6 && Model >= 13)) {
245 ToggleFeature(X86::FeatureSlowBTMem);
248 // If it's Nehalem, unaligned memory access is fast.
249 // Include Westmere and Sandy Bridge as well.
250 // FIXME: add later processors.
251 if (IsIntel && ((Family == 6 && Model == 26) ||
252 (Family == 6 && Model == 44) ||
253 (Family == 6 && Model == 42))) {
255 ToggleFeature(X86::FeatureFastUAMem);
258 // Set processor type. Currently only Atom is detected.
260 (Model == 28 || Model == 38 || Model == 39
261 || Model == 53 || Model == 54)) {
262 X86ProcFamily = IntelAtom;
265 ToggleFeature(X86::FeatureLeaForSP);
268 unsigned MaxExtLevel;
269 X86_MC::GetCpuIDAndInfo(0x80000000, &MaxExtLevel, &EBX, &ECX, &EDX);
271 if (MaxExtLevel >= 0x80000001) {
272 X86_MC::GetCpuIDAndInfo(0x80000001, &EAX, &EBX, &ECX, &EDX);
273 if ((EDX >> 29) & 0x1) {
275 ToggleFeature(X86::Feature64Bit);
277 if ((ECX >> 5) & 0x1) {
279 ToggleFeature(X86::FeatureLZCNT);
282 if ((ECX >> 6) & 0x1) {
284 ToggleFeature(X86::FeatureSSE4A);
286 if ((ECX >> 11) & 0x1) {
288 ToggleFeature(X86::FeatureXOP);
290 if ((ECX >> 16) & 0x1) {
292 ToggleFeature(X86::FeatureFMA4);
299 if (!X86_MC::GetCpuIDAndInfoEx(0x7, 0x0, &EAX, &EBX, &ECX, &EDX)) {
300 if (IsIntel && (EBX & 0x1)) {
302 ToggleFeature(X86::FeatureFSGSBase);
304 if ((EBX >> 3) & 0x1) {
306 ToggleFeature(X86::FeatureBMI);
308 if (IsIntel && ((EBX >> 5) & 0x1)) {
310 ToggleFeature(X86::FeatureAVX2);
312 if (IsIntel && ((EBX >> 8) & 0x1)) {
314 ToggleFeature(X86::FeatureBMI2);
320 X86Subtarget::X86Subtarget(const std::string &TT, const std::string &CPU,
321 const std::string &FS,
322 unsigned StackAlignOverride, bool is64Bit)
323 : X86GenSubtargetInfo(TT, CPU, FS)
324 , X86ProcFamily(Others)
325 , PICStyle(PICStyles::None)
326 , X86SSELevel(NoMMXSSE)
327 , X863DNowLevel(NoThreeDNow)
346 , HasVectorUAMem(false)
347 , HasCmpxchg16b(false)
349 , PostRAScheduler(false)
351 // FIXME: this is a known good value for Yonah. How about others?
352 , MaxInlineSizeThreshold(128)
354 , In64BitMode(is64Bit) {
355 // Determine default and user specified characteristics
356 std::string CPUName = CPU;
357 if (!FS.empty() || !CPU.empty()) {
358 if (CPUName.empty()) {
359 #if defined(i386) || defined(__i386__) || defined(__x86__) || defined(_M_IX86)\
360 || defined(__x86_64__) || defined(_M_AMD64) || defined (_M_X64)
361 CPUName = sys::getHostCPUName();
367 // Make sure 64-bit features are available in 64-bit mode. (But make sure
368 // SSE2 can be turned off explicitly.)
369 std::string FullFS = FS;
372 FullFS = "+64bit,+sse2," + FullFS;
374 FullFS = "+64bit,+sse2";
377 // If feature string is not empty, parse features string.
378 ParseSubtargetFeatures(CPUName, FullFS);
380 if (CPUName.empty()) {
381 #if defined (__x86_64__) || defined(__i386__)
382 CPUName = sys::getHostCPUName();
387 // Otherwise, use CPUID to auto-detect feature set.
388 AutoDetectSubtargetFeatures();
390 // Make sure 64-bit features are available in 64-bit mode.
392 HasX86_64 = true; ToggleFeature(X86::Feature64Bit);
393 HasCMov = true; ToggleFeature(X86::FeatureCMOV);
395 if (X86SSELevel < SSE2) {
397 ToggleFeature(X86::FeatureSSE1);
398 ToggleFeature(X86::FeatureSSE2);
403 if (X86ProcFamily == IntelAtom)
404 PostRAScheduler = true;
406 InstrItins = getInstrItineraryForCPU(CPUName);
408 // It's important to keep the MCSubtargetInfo feature bits in sync with
409 // target data structure which is shared with MC code emitter, etc.
411 ToggleFeature(X86::Mode64Bit);
413 DEBUG(dbgs() << "Subtarget features: SSELevel " << X86SSELevel
414 << ", 3DNowLevel " << X863DNowLevel
415 << ", 64bit " << HasX86_64 << "\n");
416 assert((!In64BitMode || HasX86_64) &&
417 "64-bit code requested on a subtarget that doesn't support it!");
419 // Stack alignment is 16 bytes on Darwin, FreeBSD, Linux and Solaris (both
420 // 32 and 64 bit) and for all 64-bit targets.
421 if (StackAlignOverride)
422 stackAlignment = StackAlignOverride;
423 else if (isTargetDarwin() || isTargetFreeBSD() || isTargetLinux() ||
424 isTargetSolaris() || In64BitMode)
428 bool X86Subtarget::enablePostRAScheduler(
429 CodeGenOpt::Level OptLevel,
430 TargetSubtargetInfo::AntiDepBreakMode& Mode,
431 RegClassVector& CriticalPathRCs) const {
432 Mode = TargetSubtargetInfo::ANTIDEP_CRITICAL;
433 CriticalPathRCs.clear();
434 return PostRAScheduler && OptLevel >= CodeGenOpt::Default;