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 #include "X86Subtarget.h"
15 #include "X86GenSubtarget.inc"
16 #include "llvm/Module.h"
17 #include "llvm/Support/CommandLine.h"
18 #include "llvm/Target/TargetMachine.h"
19 #include "llvm/Target/TargetOptions.h"
22 cl::opt<X86Subtarget::AsmWriterFlavorTy>
23 AsmWriterFlavor("x86-asm-syntax", cl::init(X86Subtarget::Unset),
24 cl::desc("Choose style of code to emit from X86 backend:"),
26 clEnumValN(X86Subtarget::ATT, "att", " Emit AT&T-style assembly"),
27 clEnumValN(X86Subtarget::Intel, "intel", " Emit Intel-style assembly"),
31 /// True if accessing the GV requires an extra load. For Windows, dllimported
32 /// symbols are indirect, loading the value at address GV rather then the
33 /// value of GV itself. This means that the GlobalAddress must be in the base
34 /// or index register of the address, not the GV offset field.
35 bool X86Subtarget::GVRequiresExtraLoad(const GlobalValue* GV,
36 const TargetMachine& TM,
37 bool isDirectCall) const
40 if (TM.getRelocationModel() != Reloc::Static) {
41 if (isTargetDarwin()) {
42 return (!isDirectCall &&
43 (GV->hasWeakLinkage() || GV->hasLinkOnceLinkage() ||
44 (GV->isDeclaration() && !GV->hasNotBeenReadFromBitcode())));
45 } else if (isTargetELF()) {
46 // Extra load is needed for all non-statics.
47 return (!isDirectCall &&
48 (GV->isDeclaration() || !GV->hasInternalLinkage()));
49 } else if (isTargetCygMing() || isTargetWindows()) {
50 return (GV->hasDLLImportLinkage());
57 /// This function returns the name of a function which has an interface
58 /// like the non-standard bzero function, if such a function exists on
59 /// the current subtarget and it is considered prefereable over
60 /// memset with zero passed as the second argument. Otherwise it
62 const char *X86Subtarget::getBZeroEntry() const {
64 // Darwin 10 has a __bzero entry point for this purpose.
65 if (getDarwinVers() >= 10)
71 /// GetCpuIDAndInfo - Execute the specified cpuid and return the 4 values in the
72 /// specified arguments. If we can't run cpuid on the host, return true.
73 bool X86::GetCpuIDAndInfo(unsigned value, unsigned *rEAX, unsigned *rEBX,
74 unsigned *rECX, unsigned *rEDX) {
75 #if defined(__x86_64__)
76 // gcc doesn't know cpuid would clobber ebx/rbx. Preseve it manually.
77 asm ("movq\t%%rbx, %%rsi\n\t"
79 "xchgq\t%%rbx, %%rsi\n\t"
86 #elif defined(i386) || defined(__i386__) || defined(__x86__) || defined(_M_IX86)
88 asm ("movl\t%%ebx, %%esi\n\t"
90 "xchgl\t%%ebx, %%esi\n\t"
97 #elif defined(_MSC_VER)
102 mov dword ptr [esi],eax
104 mov dword ptr [esi],ebx
106 mov dword ptr [esi],ecx
108 mov dword ptr [esi],edx
116 void X86Subtarget::AutoDetectSubtargetFeatures() {
117 unsigned EAX = 0, EBX = 0, ECX = 0, EDX = 0;
123 if (X86::GetCpuIDAndInfo(0, &EAX, text.u+0, text.u+2, text.u+1))
126 X86::GetCpuIDAndInfo(0x1, &EAX, &EBX, &ECX, &EDX);
128 if ((EDX >> 23) & 0x1) X86SSELevel = MMX;
129 if ((EDX >> 25) & 0x1) X86SSELevel = SSE1;
130 if ((EDX >> 26) & 0x1) X86SSELevel = SSE2;
131 if (ECX & 0x1) X86SSELevel = SSE3;
132 if ((ECX >> 9) & 0x1) X86SSELevel = SSSE3;
133 if ((ECX >> 19) & 0x1) X86SSELevel = SSE41;
134 if ((ECX >> 20) & 0x1) X86SSELevel = SSE42;
136 if (memcmp(text.c, "GenuineIntel", 12) == 0 ||
137 memcmp(text.c, "AuthenticAMD", 12) == 0) {
138 X86::GetCpuIDAndInfo(0x80000001, &EAX, &EBX, &ECX, &EDX);
139 HasX86_64 = (EDX >> 29) & 0x1;
143 static const char *GetCurrentX86CPU() {
144 unsigned EAX = 0, EBX = 0, ECX = 0, EDX = 0;
145 if (X86::GetCpuIDAndInfo(0x1, &EAX, &EBX, &ECX, &EDX))
147 unsigned Family = (EAX >> 8) & 0xf; // Bits 8 - 11
148 unsigned Model = (EAX >> 4) & 0xf; // Bits 4 - 7
149 X86::GetCpuIDAndInfo(0x80000001, &EAX, &EBX, &ECX, &EDX);
150 bool Em64T = (EDX >> 29) & 0x1;
157 X86::GetCpuIDAndInfo(0, &EAX, text.u+0, text.u+2, text.u+1);
158 if (memcmp(text.c, "GenuineIntel", 12) == 0) {
166 case 4: return "pentium-mmx";
167 default: return "pentium";
171 case 1: return "pentiumpro";
174 case 6: return "pentium2";
178 case 11: return "pentium3";
180 case 13: return "pentium-m";
181 case 14: return "yonah";
182 case 15: return "core2";
183 default: return "i686";
189 return (Em64T) ? "nocona" : "prescott";
191 return (Em64T) ? "x86-64" : "pentium4";
198 } else if (memcmp(text.c, "AuthenticAMD", 12) == 0) {
199 // FIXME: this poorly matches the generated SubtargetFeatureKV table. There
200 // appears to be no way to generate the wide variety of AMD-specific targets
201 // from the information returned from CPUID.
209 case 8: return "k6-2";
211 case 13: return "k6-3";
212 default: return "pentium";
216 case 4: return "athlon-tbird";
219 case 8: return "athlon-mp";
220 case 10: return "athlon-xp";
221 default: return "athlon";
225 case 1: return "opteron";
226 case 5: return "athlon-fx"; // also opteron
227 default: return "athlon64";
237 X86Subtarget::X86Subtarget(const Module &M, const std::string &FS, bool is64Bit)
238 : AsmFlavor(AsmWriterFlavor)
239 , PICStyle(PICStyle::None)
240 , X86SSELevel(NoMMXSSE)
241 , X863DNowLevel(NoThreeDNow)
245 // FIXME: this is a known good value for Yonah. How about others?
246 , MaxInlineSizeThreshold(128)
248 , TargetType(isELF) { // Default to ELF unless otherwise specified.
250 // Determine default and user specified characteristics
252 // If feature string is not empty, parse features string.
253 std::string CPU = GetCurrentX86CPU();
254 ParseSubtargetFeatures(FS, CPU);
256 // Otherwise, use CPUID to auto-detect feature set.
257 AutoDetectSubtargetFeatures();
260 // If requesting codegen for X86-64, make sure that 64-bit and SSE2 features
261 // are enabled. These are available on all x86-64 CPUs.
264 if (X86SSELevel < SSE2)
268 // Set the boolean corresponding to the current target triple, or the default
269 // if one cannot be determined, to true.
270 const std::string& TT = M.getTargetTriple();
271 if (TT.length() > 5) {
273 if ((Pos = TT.find("-darwin")) != std::string::npos) {
274 TargetType = isDarwin;
276 // Compute the darwin version number.
277 if (isdigit(TT[Pos+7]))
278 DarwinVers = atoi(&TT[Pos+7]);
280 DarwinVers = 8; // Minimum supported darwin is Tiger.
281 } else if (TT.find("cygwin") != std::string::npos) {
282 TargetType = isCygwin;
283 } else if (TT.find("mingw") != std::string::npos) {
284 TargetType = isMingw;
285 } else if (TT.find("win32") != std::string::npos) {
286 TargetType = isWindows;
287 } else if (TT.find("windows") != std::string::npos) {
288 TargetType = isWindows;
290 } else if (TT.empty()) {
291 #if defined(__CYGWIN__)
292 TargetType = isCygwin;
293 #elif defined(__MINGW32__) || defined(__MINGW64__)
294 TargetType = isMingw;
295 #elif defined(__APPLE__)
296 TargetType = isDarwin;
297 #if __APPLE_CC__ > 5400
298 DarwinVers = 9; // GCC 5400+ is Leopard.
300 DarwinVers = 8; // Minimum supported darwin is Tiger.
303 #elif defined(_WIN32) || defined(_WIN64)
304 TargetType = isWindows;
308 // If the asm syntax hasn't been overridden on the command line, use whatever
310 if (AsmFlavor == X86Subtarget::Unset) {
311 AsmFlavor = (TargetType == isWindows)
312 ? X86Subtarget::Intel : X86Subtarget::ATT;
315 // Stack alignment is 16 bytes on Darwin (both 32 and 64 bit) and for all 64
317 if (TargetType == isDarwin || Is64Bit)
321 stackAlignment = StackAlignment;