1 //===-- X86Subtarget.cpp - X86 Subtarget Information ------------*- C++ -*-===//
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 TargetSubtarget.
12 //===----------------------------------------------------------------------===//
14 #define DEBUG_TYPE "subtarget"
15 #include "X86Subtarget.h"
16 #include "X86InstrInfo.h"
17 #include "X86GenSubtarget.inc"
18 #include "llvm/GlobalValue.h"
19 #include "llvm/Support/Debug.h"
20 #include "llvm/Support/raw_ostream.h"
21 #include "llvm/System/Host.h"
22 #include "llvm/Target/TargetMachine.h"
23 #include "llvm/Target/TargetOptions.h"
24 #include "llvm/ADT/SmallVector.h"
31 /// ClassifyBlockAddressReference - Classify a blockaddress reference for the
32 /// current subtarget according to how we should reference it in a non-pcrel
34 unsigned char X86Subtarget::
35 ClassifyBlockAddressReference() const {
36 if (isPICStyleGOT()) // 32-bit ELF targets.
37 return X86II::MO_GOTOFF;
39 if (isPICStyleStubPIC()) // Darwin/32 in PIC mode.
40 return X86II::MO_PIC_BASE_OFFSET;
42 // Direct static reference to label.
43 return X86II::MO_NO_FLAG;
46 /// ClassifyGlobalReference - Classify a global variable reference for the
47 /// current subtarget according to how we should reference it in a non-pcrel
49 unsigned char X86Subtarget::
50 ClassifyGlobalReference(const GlobalValue *GV, const TargetMachine &TM) const {
51 // DLLImport only exists on windows, it is implemented as a load from a
53 if (GV->hasDLLImportLinkage())
54 return X86II::MO_DLLIMPORT;
56 // Materializable GVs (in JIT lazy compilation mode) do not require an
57 // extra load from stub.
58 bool isDecl = GV->isDeclaration() && !GV->isMaterializable();
60 // X86-64 in PIC mode.
61 if (isPICStyleRIPRel()) {
62 // Large model never uses stubs.
63 if (TM.getCodeModel() == CodeModel::Large)
64 return X86II::MO_NO_FLAG;
66 if (isTargetDarwin()) {
67 // If symbol visibility is hidden, the extra load is not needed if
68 // target is x86-64 or the symbol is definitely defined in the current
70 if (GV->hasDefaultVisibility() &&
71 (isDecl || GV->isWeakForLinker()))
72 return X86II::MO_GOTPCREL;
74 assert(isTargetELF() && "Unknown rip-relative target");
76 // Extra load is needed for all externally visible.
77 if (!GV->hasLocalLinkage() && GV->hasDefaultVisibility())
78 return X86II::MO_GOTPCREL;
81 return X86II::MO_NO_FLAG;
84 if (isPICStyleGOT()) { // 32-bit ELF targets.
85 // Extra load is needed for all externally visible.
86 if (GV->hasLocalLinkage() || GV->hasHiddenVisibility())
87 return X86II::MO_GOTOFF;
91 if (isPICStyleStubPIC()) { // Darwin/32 in PIC mode.
92 // Determine whether we have a stub reference and/or whether the reference
93 // is relative to the PIC base or not.
95 // If this is a strong reference to a definition, it is definitely not
97 if (!isDecl && !GV->isWeakForLinker())
98 return X86II::MO_PIC_BASE_OFFSET;
100 // Unless we have a symbol with hidden visibility, we have to go through a
101 // normal $non_lazy_ptr stub because this symbol might be resolved late.
102 if (!GV->hasHiddenVisibility()) // Non-hidden $non_lazy_ptr reference.
103 return X86II::MO_DARWIN_NONLAZY_PIC_BASE;
105 // If symbol visibility is hidden, we have a stub for common symbol
106 // references and external declarations.
107 if (isDecl || GV->hasCommonLinkage()) {
108 // Hidden $non_lazy_ptr reference.
109 return X86II::MO_DARWIN_HIDDEN_NONLAZY_PIC_BASE;
112 // Otherwise, no stub.
113 return X86II::MO_PIC_BASE_OFFSET;
116 if (isPICStyleStubNoDynamic()) { // Darwin/32 in -mdynamic-no-pic mode.
117 // Determine whether we have a stub reference.
119 // If this is a strong reference to a definition, it is definitely not
121 if (!isDecl && !GV->isWeakForLinker())
122 return X86II::MO_NO_FLAG;
124 // Unless we have a symbol with hidden visibility, we have to go through a
125 // normal $non_lazy_ptr stub because this symbol might be resolved late.
126 if (!GV->hasHiddenVisibility()) // Non-hidden $non_lazy_ptr reference.
127 return X86II::MO_DARWIN_NONLAZY;
129 // Otherwise, no stub.
130 return X86II::MO_NO_FLAG;
133 // Direct static reference to global.
134 return X86II::MO_NO_FLAG;
138 /// getBZeroEntry - This function returns the name of a function which has an
139 /// interface like the non-standard bzero function, if such a function exists on
140 /// the current subtarget and it is considered prefereable over memset with zero
141 /// passed as the second argument. Otherwise it returns null.
142 const char *X86Subtarget::getBZeroEntry() const {
143 // Darwin 10 has a __bzero entry point for this purpose.
144 if (getDarwinVers() >= 10)
150 /// IsLegalToCallImmediateAddr - Return true if the subtarget allows calls
151 /// to immediate address.
152 bool X86Subtarget::IsLegalToCallImmediateAddr(const TargetMachine &TM) const {
155 return isTargetELF() || TM.getRelocationModel() == Reloc::Static;
158 /// getSpecialAddressLatency - For targets where it is beneficial to
159 /// backschedule instructions that compute addresses, return a value
160 /// indicating the number of scheduling cycles of backscheduling that
161 /// should be attempted.
162 unsigned X86Subtarget::getSpecialAddressLatency() const {
163 // For x86 out-of-order targets, back-schedule address computations so
164 // that loads and stores aren't blocked.
165 // This value was chosen arbitrarily.
169 /// GetCpuIDAndInfo - Execute the specified cpuid and return the 4 values in the
170 /// specified arguments. If we can't run cpuid on the host, return true.
171 static bool GetCpuIDAndInfo(unsigned value, unsigned *rEAX,
172 unsigned *rEBX, unsigned *rECX, unsigned *rEDX) {
173 #if defined(__x86_64__) || defined(_M_AMD64) || defined (_M_X64)
174 #if defined(__GNUC__)
175 // gcc doesn't know cpuid would clobber ebx/rbx. Preseve it manually.
176 asm ("movq\t%%rbx, %%rsi\n\t"
178 "xchgq\t%%rbx, %%rsi\n\t"
185 #elif defined(_MSC_VER)
187 __cpuid(registers, value);
188 *rEAX = registers[0];
189 *rEBX = registers[1];
190 *rECX = registers[2];
191 *rEDX = registers[3];
194 #elif defined(i386) || defined(__i386__) || defined(__x86__) || defined(_M_IX86)
195 #if defined(__GNUC__)
196 asm ("movl\t%%ebx, %%esi\n\t"
198 "xchgl\t%%ebx, %%esi\n\t"
205 #elif defined(_MSC_VER)
210 mov dword ptr [esi],eax
212 mov dword ptr [esi],ebx
214 mov dword ptr [esi],ecx
216 mov dword ptr [esi],edx
224 static void DetectFamilyModel(unsigned EAX, unsigned &Family, unsigned &Model) {
225 Family = (EAX >> 8) & 0xf; // Bits 8 - 11
226 Model = (EAX >> 4) & 0xf; // Bits 4 - 7
227 if (Family == 6 || Family == 0xf) {
229 // Examine extended family ID if family ID is F.
230 Family += (EAX >> 20) & 0xff; // Bits 20 - 27
231 // Examine extended model ID if family ID is 6 or F.
232 Model += ((EAX >> 16) & 0xf) << 4; // Bits 16 - 19
236 void X86Subtarget::AutoDetectSubtargetFeatures() {
237 unsigned EAX = 0, EBX = 0, ECX = 0, EDX = 0;
243 if (GetCpuIDAndInfo(0, &EAX, text.u+0, text.u+2, text.u+1))
246 GetCpuIDAndInfo(0x1, &EAX, &EBX, &ECX, &EDX);
248 if ((EDX >> 15) & 1) HasCMov = true;
249 if ((EDX >> 23) & 1) X86SSELevel = MMX;
250 if ((EDX >> 25) & 1) X86SSELevel = SSE1;
251 if ((EDX >> 26) & 1) X86SSELevel = SSE2;
252 if (ECX & 0x1) X86SSELevel = SSE3;
253 if ((ECX >> 9) & 1) X86SSELevel = SSSE3;
254 if ((ECX >> 19) & 1) X86SSELevel = SSE41;
255 if ((ECX >> 20) & 1) X86SSELevel = SSE42;
257 bool IsIntel = memcmp(text.c, "GenuineIntel", 12) == 0;
258 bool IsAMD = !IsIntel && memcmp(text.c, "AuthenticAMD", 12) == 0;
260 HasFMA3 = IsIntel && ((ECX >> 12) & 0x1);
261 HasAVX = ((ECX >> 28) & 0x1);
263 if (IsIntel || IsAMD) {
264 // Determine if bit test memory instructions are slow.
267 DetectFamilyModel(EAX, Family, Model);
268 IsBTMemSlow = IsAMD || (Family == 6 && Model >= 13);
270 GetCpuIDAndInfo(0x80000001, &EAX, &EBX, &ECX, &EDX);
271 HasX86_64 = (EDX >> 29) & 0x1;
272 HasSSE4A = IsAMD && ((ECX >> 6) & 0x1);
273 HasFMA4 = IsAMD && ((ECX >> 16) & 0x1);
277 X86Subtarget::X86Subtarget(const std::string &TT, const std::string &FS,
279 : PICStyle(PICStyles::None)
280 , X86SSELevel(NoMMXSSE)
281 , X863DNowLevel(NoThreeDNow)
289 , HasVectorUAMem(false)
292 // FIXME: this is a known good value for Yonah. How about others?
293 , MaxInlineSizeThreshold(128)
295 , TargetType(isELF) { // Default to ELF unless otherwise specified.
297 // default to hard float ABI
298 if (FloatABIType == FloatABI::Default)
299 FloatABIType = FloatABI::Hard;
301 // Determine default and user specified characteristics
303 // If feature string is not empty, parse features string.
304 std::string CPU = sys::getHostCPUName();
305 ParseSubtargetFeatures(FS, CPU);
306 // All X86-64 CPUs also have SSE2, however user might request no SSE via
307 // -mattr, so don't force SSELevel here.
309 // Otherwise, use CPUID to auto-detect feature set.
310 AutoDetectSubtargetFeatures();
311 // Make sure SSE2 is enabled; it is available on all X86-64 CPUs.
312 if (Is64Bit && X86SSELevel < SSE2)
316 // If requesting codegen for X86-64, make sure that 64-bit features
321 DEBUG(dbgs() << "Subtarget features: SSELevel " << X86SSELevel
322 << ", 3DNowLevel " << X863DNowLevel
323 << ", 64bit " << HasX86_64 << "\n");
324 assert((!Is64Bit || HasX86_64) &&
325 "64-bit code requested on a subtarget that doesn't support it!");
327 // Set the boolean corresponding to the current target triple, or the default
328 // if one cannot be determined, to true.
329 if (TT.length() > 5) {
331 if ((Pos = TT.find("-darwin")) != std::string::npos) {
332 TargetType = isDarwin;
334 // Compute the darwin version number.
335 if (isdigit(TT[Pos+7]))
336 DarwinVers = atoi(&TT[Pos+7]);
338 DarwinVers = 8; // Minimum supported darwin is Tiger.
339 } else if (TT.find("linux") != std::string::npos) {
340 // Linux doesn't imply ELF, but we don't currently support anything else.
342 } else if (TT.find("cygwin") != std::string::npos) {
343 TargetType = isCygwin;
344 } else if (TT.find("mingw") != std::string::npos) {
345 TargetType = isMingw;
346 } else if (TT.find("win32") != std::string::npos) {
347 TargetType = isWindows;
348 } else if (TT.find("windows") != std::string::npos) {
349 TargetType = isWindows;
350 } else if (TT.find("-cl") != std::string::npos) {
351 TargetType = isDarwin;
356 // Stack alignment is 16 bytes on Darwin (both 32 and 64 bit) and for all 64
358 if (TargetType == isDarwin || Is64Bit)
362 stackAlignment = StackAlignment;
365 bool X86Subtarget::enablePostRAScheduler(
366 CodeGenOpt::Level OptLevel,
367 TargetSubtarget::AntiDepBreakMode& Mode,
368 RegClassVector& CriticalPathRCs) const {
369 Mode = TargetSubtarget::ANTIDEP_CRITICAL;
370 CriticalPathRCs.clear();
371 return OptLevel >= CodeGenOpt::Aggressive;