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/Module.h"
19 #include "llvm/Support/CommandLine.h"
20 #include "llvm/Support/Debug.h"
21 #include "llvm/Target/TargetMachine.h"
22 #include "llvm/Target/TargetOptions.h"
29 static cl::opt<X86Subtarget::AsmWriterFlavorTy>
30 AsmWriterFlavor("x86-asm-syntax", cl::init(X86Subtarget::Unset),
31 cl::desc("Choose style of code to emit from X86 backend:"),
33 clEnumValN(X86Subtarget::ATT, "att", "Emit AT&T-style assembly"),
34 clEnumValN(X86Subtarget::Intel, "intel", "Emit Intel-style assembly"),
37 /// ClassifyGlobalReference - Classify a global variable reference for the
38 /// current subtarget according to how we should reference it in a non-pcrel
40 unsigned char X86Subtarget::
41 ClassifyGlobalReference(const GlobalValue *GV, const TargetMachine &TM) const {
42 // DLLImport only exists on windows, it is implemented as a load from a
44 if (GV->hasDLLImportLinkage())
45 return X86II::MO_DLLIMPORT;
47 // X86-64 in PIC mode.
48 if (isPICStyleRIPRel()) {
49 // Large model never uses stubs.
50 if (TM.getCodeModel() == CodeModel::Large)
51 return X86II::MO_NO_FLAG;
53 if (isTargetDarwin()) {
54 // If symbol visibility is hidden, the extra load is not needed if
55 // target is x86-64 or the symbol is definitely defined in the current
57 if (GV->hasDefaultVisibility() &&
58 (GV->isDeclaration() || GV->isWeakForLinker()))
59 return X86II::MO_GOTPCREL;
61 assert(isTargetELF() && "Unknown rip-relative target");
63 // Extra load is needed for all externally visible.
64 if (!GV->hasLocalLinkage() && GV->hasDefaultVisibility())
65 return X86II::MO_GOTPCREL;
68 return X86II::MO_NO_FLAG;
71 if (isPICStyleGOT()) { // 32-bit ELF targets.
72 // Extra load is needed for all externally visible.
73 if (GV->hasLocalLinkage() || GV->hasHiddenVisibility())
74 return X86II::MO_GOTOFF;
78 if (isPICStyleStubPIC()) { // Darwin/32 in PIC mode.
79 // Determine whether we have a stub reference and/or whether the reference
80 // is relative to the PIC base or not.
82 // If this is a strong reference to a definition, it is definitely not
84 if (!GV->isDeclaration() && !GV->isWeakForLinker())
85 return X86II::MO_PIC_BASE_OFFSET;
87 // Unless we have a symbol with hidden visibility, we have to go through a
88 // normal $non_lazy_ptr stub because this symbol might be resolved late.
89 if (!GV->hasHiddenVisibility()) // Non-hidden $non_lazy_ptr reference.
90 return X86II::MO_DARWIN_NONLAZY_PIC_BASE;
92 // If symbol visibility is hidden, we have a stub for common symbol
93 // references and external declarations.
94 if (GV->isDeclaration() || GV->hasCommonLinkage()) {
95 // Hidden $non_lazy_ptr reference.
96 return X86II::MO_DARWIN_HIDDEN_NONLAZY_PIC_BASE;
99 // Otherwise, no stub.
100 return X86II::MO_PIC_BASE_OFFSET;
103 if (isPICStyleStubNoDynamic()) { // Darwin/32 in -mdynamic-no-pic mode.
104 // Determine whether we have a stub reference.
106 // If this is a strong reference to a definition, it is definitely not
108 if (!GV->isDeclaration() && !GV->isWeakForLinker())
109 return X86II::MO_NO_FLAG;
111 // Unless we have a symbol with hidden visibility, we have to go through a
112 // normal $non_lazy_ptr stub because this symbol might be resolved late.
113 if (!GV->hasHiddenVisibility()) // Non-hidden $non_lazy_ptr reference.
114 return X86II::MO_DARWIN_NONLAZY;
116 // If symbol visibility is hidden, we have a stub for common symbol
117 // references and external declarations.
118 if (GV->isDeclaration() || GV->hasCommonLinkage()) {
119 // Hidden $non_lazy_ptr reference.
120 return X86II::MO_DARWIN_HIDDEN_NONLAZY;
123 // Otherwise, no stub.
124 return X86II::MO_NO_FLAG;
127 // Direct static reference to global.
128 return X86II::MO_NO_FLAG;
132 /// getBZeroEntry - This function returns the name of a function which has an
133 /// interface like the non-standard bzero function, if such a function exists on
134 /// the current subtarget and it is considered prefereable over memset with zero
135 /// passed as the second argument. Otherwise it returns null.
136 const char *X86Subtarget::getBZeroEntry() const {
137 // Darwin 10 has a __bzero entry point for this purpose.
138 if (getDarwinVers() >= 10)
144 /// IsLegalToCallImmediateAddr - Return true if the subtarget allows calls
145 /// to immediate address.
146 bool X86Subtarget::IsLegalToCallImmediateAddr(const TargetMachine &TM) const {
149 return isTargetELF() || TM.getRelocationModel() == Reloc::Static;
152 /// getSpecialAddressLatency - For targets where it is beneficial to
153 /// backschedule instructions that compute addresses, return a value
154 /// indicating the number of scheduling cycles of backscheduling that
155 /// should be attempted.
156 unsigned X86Subtarget::getSpecialAddressLatency() const {
157 // For x86 out-of-order targets, back-schedule address computations so
158 // that loads and stores aren't blocked.
159 // This value was chosen arbitrarily.
163 /// GetCpuIDAndInfo - Execute the specified cpuid and return the 4 values in the
164 /// specified arguments. If we can't run cpuid on the host, return true.
165 bool X86::GetCpuIDAndInfo(unsigned value, unsigned *rEAX, unsigned *rEBX,
166 unsigned *rECX, unsigned *rEDX) {
167 #if defined(__x86_64__) || defined(_M_AMD64)
168 #if defined(__GNUC__)
169 // gcc doesn't know cpuid would clobber ebx/rbx. Preseve it manually.
170 asm ("movq\t%%rbx, %%rsi\n\t"
172 "xchgq\t%%rbx, %%rsi\n\t"
179 #elif defined(_MSC_VER)
181 __cpuid(registers, value);
182 *rEAX = registers[0];
183 *rEBX = registers[1];
184 *rECX = registers[2];
185 *rEDX = registers[3];
188 #elif defined(i386) || defined(__i386__) || defined(__x86__) || defined(_M_IX86)
189 #if defined(__GNUC__)
190 asm ("movl\t%%ebx, %%esi\n\t"
192 "xchgl\t%%ebx, %%esi\n\t"
199 #elif defined(_MSC_VER)
204 mov dword ptr [esi],eax
206 mov dword ptr [esi],ebx
208 mov dword ptr [esi],ecx
210 mov dword ptr [esi],edx
218 static void DetectFamilyModel(unsigned EAX, unsigned &Family, unsigned &Model) {
219 Family = (EAX >> 8) & 0xf; // Bits 8 - 11
220 Model = (EAX >> 4) & 0xf; // Bits 4 - 7
221 if (Family == 6 || Family == 0xf) {
223 // Examine extended family ID if family ID is F.
224 Family += (EAX >> 20) & 0xff; // Bits 20 - 27
225 // Examine extended model ID if family ID is 6 or F.
226 Model += ((EAX >> 16) & 0xf) << 4; // Bits 16 - 19
230 void X86Subtarget::AutoDetectSubtargetFeatures() {
231 unsigned EAX = 0, EBX = 0, ECX = 0, EDX = 0;
237 if (X86::GetCpuIDAndInfo(0, &EAX, text.u+0, text.u+2, text.u+1))
240 X86::GetCpuIDAndInfo(0x1, &EAX, &EBX, &ECX, &EDX);
242 if ((EDX >> 23) & 0x1) X86SSELevel = MMX;
243 if ((EDX >> 25) & 0x1) X86SSELevel = SSE1;
244 if ((EDX >> 26) & 0x1) X86SSELevel = SSE2;
245 if (ECX & 0x1) X86SSELevel = SSE3;
246 if ((ECX >> 9) & 0x1) X86SSELevel = SSSE3;
247 if ((ECX >> 19) & 0x1) X86SSELevel = SSE41;
248 if ((ECX >> 20) & 0x1) X86SSELevel = SSE42;
250 bool IsIntel = memcmp(text.c, "GenuineIntel", 12) == 0;
251 bool IsAMD = !IsIntel && memcmp(text.c, "AuthenticAMD", 12) == 0;
253 HasFMA3 = IsIntel && ((ECX >> 12) & 0x1);
254 HasAVX = ((ECX >> 28) & 0x1);
256 if (IsIntel || IsAMD) {
257 // Determine if bit test memory instructions are slow.
260 DetectFamilyModel(EAX, Family, Model);
261 IsBTMemSlow = IsAMD || (Family == 6 && Model >= 13);
263 X86::GetCpuIDAndInfo(0x80000001, &EAX, &EBX, &ECX, &EDX);
264 HasX86_64 = (EDX >> 29) & 0x1;
265 HasSSE4A = IsAMD && ((ECX >> 6) & 0x1);
266 HasFMA4 = IsAMD && ((ECX >> 16) & 0x1);
270 static const char *GetCurrentX86CPU() {
271 unsigned EAX = 0, EBX = 0, ECX = 0, EDX = 0;
272 if (X86::GetCpuIDAndInfo(0x1, &EAX, &EBX, &ECX, &EDX))
276 DetectFamilyModel(EAX, Family, Model);
278 X86::GetCpuIDAndInfo(0x80000001, &EAX, &EBX, &ECX, &EDX);
279 bool Em64T = (EDX >> 29) & 0x1;
280 bool HasSSE3 = (ECX & 0x1);
287 X86::GetCpuIDAndInfo(0, &EAX, text.u+0, text.u+2, text.u+1);
288 if (memcmp(text.c, "GenuineIntel", 12) == 0) {
296 case 4: return "pentium-mmx";
297 default: return "pentium";
301 case 1: return "pentiumpro";
304 case 6: return "pentium2";
308 case 11: return "pentium3";
310 case 13: return "pentium-m";
311 case 14: return "yonah";
313 case 22: // Celeron M 540
315 case 23: // 45nm: Penryn , Wolfdale, Yorkfield (XE)
317 default: return "i686";
323 case 6: // same as 4, but 65nm
324 return (Em64T) ? "nocona" : "prescott";
330 return (Em64T) ? "x86-64" : "pentium4";
337 } else if (memcmp(text.c, "AuthenticAMD", 12) == 0) {
338 // FIXME: this poorly matches the generated SubtargetFeatureKV table. There
339 // appears to be no way to generate the wide variety of AMD-specific targets
340 // from the information returned from CPUID.
348 case 8: return "k6-2";
350 case 13: return "k6-3";
351 default: return "pentium";
355 case 4: return "athlon-tbird";
358 case 8: return "athlon-mp";
359 case 10: return "athlon-xp";
360 default: return "athlon";
365 default: return "k8-sse3";
369 case 1: return "opteron";
370 case 5: return "athlon-fx"; // also opteron
371 default: return "athlon64";
376 default: return "amdfam10";
386 X86Subtarget::X86Subtarget(const Module &M, const std::string &FS, bool is64Bit)
387 : AsmFlavor(AsmWriterFlavor)
388 , PICStyle(PICStyles::None)
389 , X86SSELevel(NoMMXSSE)
390 , X863DNowLevel(NoThreeDNow)
400 // FIXME: this is a known good value for Yonah. How about others?
401 , MaxInlineSizeThreshold(128)
403 , TargetType(isELF) { // Default to ELF unless otherwise specified.
405 // default to hard float ABI
406 if (FloatABIType == FloatABI::Default)
407 FloatABIType = FloatABI::Hard;
409 // Determine default and user specified characteristics
411 // If feature string is not empty, parse features string.
412 std::string CPU = GetCurrentX86CPU();
413 ParseSubtargetFeatures(FS, CPU);
414 // All X86-64 CPUs also have SSE2, however user might request no SSE via
415 // -mattr, so don't force SSELevel here.
417 // Otherwise, use CPUID to auto-detect feature set.
418 AutoDetectSubtargetFeatures();
419 // Make sure SSE2 is enabled; it is available on all X86-64 CPUs.
420 if (Is64Bit && X86SSELevel < SSE2)
424 // If requesting codegen for X86-64, make sure that 64-bit features
429 DOUT << "Subtarget features: SSELevel " << X86SSELevel
430 << ", 3DNowLevel " << X863DNowLevel
431 << ", 64bit " << HasX86_64 << "\n";
432 assert((!Is64Bit || HasX86_64) &&
433 "64-bit code requested on a subtarget that doesn't support it!");
435 // Set the boolean corresponding to the current target triple, or the default
436 // if one cannot be determined, to true.
437 const std::string& TT = M.getTargetTriple();
438 if (TT.length() > 5) {
440 if ((Pos = TT.find("-darwin")) != std::string::npos) {
441 TargetType = isDarwin;
443 // Compute the darwin version number.
444 if (isdigit(TT[Pos+7]))
445 DarwinVers = atoi(&TT[Pos+7]);
447 DarwinVers = 8; // Minimum supported darwin is Tiger.
448 } else if (TT.find("linux") != std::string::npos) {
449 // Linux doesn't imply ELF, but we don't currently support anything else.
452 } else if (TT.find("cygwin") != std::string::npos) {
453 TargetType = isCygwin;
454 } else if (TT.find("mingw") != std::string::npos) {
455 TargetType = isMingw;
456 } else if (TT.find("win32") != std::string::npos) {
457 TargetType = isWindows;
458 } else if (TT.find("windows") != std::string::npos) {
459 TargetType = isWindows;
461 else if (TT.find("-cl") != std::string::npos) {
462 TargetType = isDarwin;
465 } else if (TT.empty()) {
466 #if defined(__CYGWIN__)
467 TargetType = isCygwin;
468 #elif defined(__MINGW32__) || defined(__MINGW64__)
469 TargetType = isMingw;
470 #elif defined(__APPLE__)
471 TargetType = isDarwin;
472 #if __APPLE_CC__ > 5400
473 DarwinVers = 9; // GCC 5400+ is Leopard.
475 DarwinVers = 8; // Minimum supported darwin is Tiger.
478 #elif defined(_WIN32) || defined(_WIN64)
479 TargetType = isWindows;
480 #elif defined(__linux__)
481 // Linux doesn't imply ELF, but we don't currently support anything else.
487 // If the asm syntax hasn't been overridden on the command line, use whatever
489 if (AsmFlavor == X86Subtarget::Unset) {
490 AsmFlavor = (TargetType == isWindows)
491 ? X86Subtarget::Intel : X86Subtarget::ATT;
494 // Stack alignment is 16 bytes on Darwin (both 32 and 64 bit) and for all 64
496 if (TargetType == isDarwin || Is64Bit)
500 stackAlignment = StackAlignment;