remove a dead bool.
[llvm/avr.git] / lib / Target / X86 / X86Subtarget.cpp
blobfb76aeb05556a17d4fe5bfc36abbf74a80be3150
1 //===-- X86Subtarget.cpp - X86 Subtarget Information ------------*- C++ -*-===//
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 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/Target/TargetMachine.h"
22 #include "llvm/Target/TargetOptions.h"
23 using namespace llvm;
25 #if defined(_MSC_VER)
26 #include <intrin.h>
27 #endif
29 /// ClassifyGlobalReference - Classify a global variable reference for the
30 /// current subtarget according to how we should reference it in a non-pcrel
31 /// context.
32 unsigned char X86Subtarget::
33 ClassifyGlobalReference(const GlobalValue *GV, const TargetMachine &TM) const {
34 // DLLImport only exists on windows, it is implemented as a load from a
35 // DLLIMPORT stub.
36 if (GV->hasDLLImportLinkage())
37 return X86II::MO_DLLIMPORT;
39 // GV with ghost linkage (in JIT lazy compilation mode) do not require an
40 // extra load from stub.
41 bool isDecl = GV->isDeclaration() && !GV->hasNotBeenReadFromBitcode();
43 // X86-64 in PIC mode.
44 if (isPICStyleRIPRel()) {
45 // Large model never uses stubs.
46 if (TM.getCodeModel() == CodeModel::Large)
47 return X86II::MO_NO_FLAG;
49 if (isTargetDarwin()) {
50 // If symbol visibility is hidden, the extra load is not needed if
51 // target is x86-64 or the symbol is definitely defined in the current
52 // translation unit.
53 if (GV->hasDefaultVisibility() &&
54 (isDecl || GV->isWeakForLinker()))
55 return X86II::MO_GOTPCREL;
56 } else {
57 assert(isTargetELF() && "Unknown rip-relative target");
59 // Extra load is needed for all externally visible.
60 if (!GV->hasLocalLinkage() && GV->hasDefaultVisibility())
61 return X86II::MO_GOTPCREL;
64 return X86II::MO_NO_FLAG;
67 if (isPICStyleGOT()) { // 32-bit ELF targets.
68 // Extra load is needed for all externally visible.
69 if (GV->hasLocalLinkage() || GV->hasHiddenVisibility())
70 return X86II::MO_GOTOFF;
71 return X86II::MO_GOT;
74 if (isPICStyleStubPIC()) { // Darwin/32 in PIC mode.
75 // Determine whether we have a stub reference and/or whether the reference
76 // is relative to the PIC base or not.
78 // If this is a strong reference to a definition, it is definitely not
79 // through a stub.
80 if (!isDecl && !GV->isWeakForLinker())
81 return X86II::MO_PIC_BASE_OFFSET;
83 // Unless we have a symbol with hidden visibility, we have to go through a
84 // normal $non_lazy_ptr stub because this symbol might be resolved late.
85 if (!GV->hasHiddenVisibility()) // Non-hidden $non_lazy_ptr reference.
86 return X86II::MO_DARWIN_NONLAZY_PIC_BASE;
88 // If symbol visibility is hidden, we have a stub for common symbol
89 // references and external declarations.
90 if (isDecl || GV->hasCommonLinkage()) {
91 // Hidden $non_lazy_ptr reference.
92 return X86II::MO_DARWIN_HIDDEN_NONLAZY_PIC_BASE;
95 // Otherwise, no stub.
96 return X86II::MO_PIC_BASE_OFFSET;
99 if (isPICStyleStubNoDynamic()) { // Darwin/32 in -mdynamic-no-pic mode.
100 // Determine whether we have a stub reference.
102 // If this is a strong reference to a definition, it is definitely not
103 // through a stub.
104 if (!isDecl && !GV->isWeakForLinker())
105 return X86II::MO_NO_FLAG;
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;
112 // Otherwise, no stub.
113 return X86II::MO_NO_FLAG;
116 // Direct static reference to global.
117 return X86II::MO_NO_FLAG;
121 /// getBZeroEntry - This function returns the name of a function which has an
122 /// interface like the non-standard bzero function, if such a function exists on
123 /// the current subtarget and it is considered prefereable over memset with zero
124 /// passed as the second argument. Otherwise it returns null.
125 const char *X86Subtarget::getBZeroEntry() const {
126 // Darwin 10 has a __bzero entry point for this purpose.
127 if (getDarwinVers() >= 10)
128 return "__bzero";
130 return 0;
133 /// IsLegalToCallImmediateAddr - Return true if the subtarget allows calls
134 /// to immediate address.
135 bool X86Subtarget::IsLegalToCallImmediateAddr(const TargetMachine &TM) const {
136 if (Is64Bit)
137 return false;
138 return isTargetELF() || TM.getRelocationModel() == Reloc::Static;
141 /// getSpecialAddressLatency - For targets where it is beneficial to
142 /// backschedule instructions that compute addresses, return a value
143 /// indicating the number of scheduling cycles of backscheduling that
144 /// should be attempted.
145 unsigned X86Subtarget::getSpecialAddressLatency() const {
146 // For x86 out-of-order targets, back-schedule address computations so
147 // that loads and stores aren't blocked.
148 // This value was chosen arbitrarily.
149 return 200;
152 /// GetCpuIDAndInfo - Execute the specified cpuid and return the 4 values in the
153 /// specified arguments. If we can't run cpuid on the host, return true.
154 static bool GetCpuIDAndInfo(unsigned value, unsigned *rEAX,
155 unsigned *rEBX, unsigned *rECX, unsigned *rEDX) {
156 #if defined(__x86_64__) || defined(_M_AMD64) || defined (_M_X64)
157 #if defined(__GNUC__)
158 // gcc doesn't know cpuid would clobber ebx/rbx. Preseve it manually.
159 asm ("movq\t%%rbx, %%rsi\n\t"
160 "cpuid\n\t"
161 "xchgq\t%%rbx, %%rsi\n\t"
162 : "=a" (*rEAX),
163 "=S" (*rEBX),
164 "=c" (*rECX),
165 "=d" (*rEDX)
166 : "a" (value));
167 return false;
168 #elif defined(_MSC_VER)
169 int registers[4];
170 __cpuid(registers, value);
171 *rEAX = registers[0];
172 *rEBX = registers[1];
173 *rECX = registers[2];
174 *rEDX = registers[3];
175 return false;
176 #endif
177 #elif defined(i386) || defined(__i386__) || defined(__x86__) || defined(_M_IX86)
178 #if defined(__GNUC__)
179 asm ("movl\t%%ebx, %%esi\n\t"
180 "cpuid\n\t"
181 "xchgl\t%%ebx, %%esi\n\t"
182 : "=a" (*rEAX),
183 "=S" (*rEBX),
184 "=c" (*rECX),
185 "=d" (*rEDX)
186 : "a" (value));
187 return false;
188 #elif defined(_MSC_VER)
189 __asm {
190 mov eax,value
191 cpuid
192 mov esi,rEAX
193 mov dword ptr [esi],eax
194 mov esi,rEBX
195 mov dword ptr [esi],ebx
196 mov esi,rECX
197 mov dword ptr [esi],ecx
198 mov esi,rEDX
199 mov dword ptr [esi],edx
201 return false;
202 #endif
203 #endif
204 return true;
207 static void DetectFamilyModel(unsigned EAX, unsigned &Family, unsigned &Model) {
208 Family = (EAX >> 8) & 0xf; // Bits 8 - 11
209 Model = (EAX >> 4) & 0xf; // Bits 4 - 7
210 if (Family == 6 || Family == 0xf) {
211 if (Family == 0xf)
212 // Examine extended family ID if family ID is F.
213 Family += (EAX >> 20) & 0xff; // Bits 20 - 27
214 // Examine extended model ID if family ID is 6 or F.
215 Model += ((EAX >> 16) & 0xf) << 4; // Bits 16 - 19
219 void X86Subtarget::AutoDetectSubtargetFeatures() {
220 unsigned EAX = 0, EBX = 0, ECX = 0, EDX = 0;
221 union {
222 unsigned u[3];
223 char c[12];
224 } text;
226 if (GetCpuIDAndInfo(0, &EAX, text.u+0, text.u+2, text.u+1))
227 return;
229 GetCpuIDAndInfo(0x1, &EAX, &EBX, &ECX, &EDX);
231 if ((EDX >> 15) & 1) HasCMov = true;
232 if ((EDX >> 23) & 1) X86SSELevel = MMX;
233 if ((EDX >> 25) & 1) X86SSELevel = SSE1;
234 if ((EDX >> 26) & 1) X86SSELevel = SSE2;
235 if (ECX & 0x1) X86SSELevel = SSE3;
236 if ((ECX >> 9) & 1) X86SSELevel = SSSE3;
237 if ((ECX >> 19) & 1) X86SSELevel = SSE41;
238 if ((ECX >> 20) & 1) X86SSELevel = SSE42;
240 bool IsIntel = memcmp(text.c, "GenuineIntel", 12) == 0;
241 bool IsAMD = !IsIntel && memcmp(text.c, "AuthenticAMD", 12) == 0;
243 HasFMA3 = IsIntel && ((ECX >> 12) & 0x1);
244 HasAVX = ((ECX >> 28) & 0x1);
246 if (IsIntel || IsAMD) {
247 // Determine if bit test memory instructions are slow.
248 unsigned Family = 0;
249 unsigned Model = 0;
250 DetectFamilyModel(EAX, Family, Model);
251 IsBTMemSlow = IsAMD || (Family == 6 && Model >= 13);
253 GetCpuIDAndInfo(0x80000001, &EAX, &EBX, &ECX, &EDX);
254 HasX86_64 = (EDX >> 29) & 0x1;
255 HasSSE4A = IsAMD && ((ECX >> 6) & 0x1);
256 HasFMA4 = IsAMD && ((ECX >> 16) & 0x1);
260 static const char *GetCurrentX86CPU() {
261 unsigned EAX = 0, EBX = 0, ECX = 0, EDX = 0;
262 if (GetCpuIDAndInfo(0x1, &EAX, &EBX, &ECX, &EDX))
263 return "generic";
264 unsigned Family = 0;
265 unsigned Model = 0;
266 DetectFamilyModel(EAX, Family, Model);
268 GetCpuIDAndInfo(0x80000001, &EAX, &EBX, &ECX, &EDX);
269 bool Em64T = (EDX >> 29) & 0x1;
270 bool HasSSE3 = (ECX & 0x1);
272 union {
273 unsigned u[3];
274 char c[12];
275 } text;
277 GetCpuIDAndInfo(0, &EAX, text.u+0, text.u+2, text.u+1);
278 if (memcmp(text.c, "GenuineIntel", 12) == 0) {
279 switch (Family) {
280 case 3:
281 return "i386";
282 case 4:
283 return "i486";
284 case 5:
285 switch (Model) {
286 case 4: return "pentium-mmx";
287 default: return "pentium";
289 case 6:
290 switch (Model) {
291 case 1: return "pentiumpro";
292 case 3:
293 case 5:
294 case 6: return "pentium2";
295 case 7:
296 case 8:
297 case 10:
298 case 11: return "pentium3";
299 case 9:
300 case 13: return "pentium-m";
301 case 14: return "yonah";
302 case 15:
303 case 22: // Celeron M 540
304 return "core2";
305 case 23: // 45nm: Penryn , Wolfdale, Yorkfield (XE)
306 return "penryn";
307 default: return "i686";
309 case 15: {
310 switch (Model) {
311 case 3:
312 case 4:
313 case 6: // same as 4, but 65nm
314 return (Em64T) ? "nocona" : "prescott";
315 case 26:
316 return "corei7";
317 case 28:
318 return "atom";
319 default:
320 return (Em64T) ? "x86-64" : "pentium4";
324 default:
325 return "generic";
327 } else if (memcmp(text.c, "AuthenticAMD", 12) == 0) {
328 // FIXME: this poorly matches the generated SubtargetFeatureKV table. There
329 // appears to be no way to generate the wide variety of AMD-specific targets
330 // from the information returned from CPUID.
331 switch (Family) {
332 case 4:
333 return "i486";
334 case 5:
335 switch (Model) {
336 case 6:
337 case 7: return "k6";
338 case 8: return "k6-2";
339 case 9:
340 case 13: return "k6-3";
341 default: return "pentium";
343 case 6:
344 switch (Model) {
345 case 4: return "athlon-tbird";
346 case 6:
347 case 7:
348 case 8: return "athlon-mp";
349 case 10: return "athlon-xp";
350 default: return "athlon";
352 case 15:
353 if (HasSSE3) {
354 return "k8-sse3";
355 } else {
356 switch (Model) {
357 case 1: return "opteron";
358 case 5: return "athlon-fx"; // also opteron
359 default: return "athlon64";
362 case 16:
363 return "amdfam10";
364 default:
365 return "generic";
367 } else {
368 return "generic";
372 X86Subtarget::X86Subtarget(const std::string &TT, const std::string &FS,
373 bool is64Bit)
374 : PICStyle(PICStyles::None)
375 , X86SSELevel(NoMMXSSE)
376 , X863DNowLevel(NoThreeDNow)
377 , HasCMov(false)
378 , HasX86_64(false)
379 , HasSSE4A(false)
380 , HasAVX(false)
381 , HasFMA3(false)
382 , HasFMA4(false)
383 , IsBTMemSlow(false)
384 , DarwinVers(0)
385 , IsLinux(false)
386 , stackAlignment(8)
387 // FIXME: this is a known good value for Yonah. How about others?
388 , MaxInlineSizeThreshold(128)
389 , Is64Bit(is64Bit)
390 , TargetType(isELF) { // Default to ELF unless otherwise specified.
392 // default to hard float ABI
393 if (FloatABIType == FloatABI::Default)
394 FloatABIType = FloatABI::Hard;
396 // Determine default and user specified characteristics
397 if (!FS.empty()) {
398 // If feature string is not empty, parse features string.
399 std::string CPU = GetCurrentX86CPU();
400 ParseSubtargetFeatures(FS, CPU);
401 // All X86-64 CPUs also have SSE2, however user might request no SSE via
402 // -mattr, so don't force SSELevel here.
403 } else {
404 // Otherwise, use CPUID to auto-detect feature set.
405 AutoDetectSubtargetFeatures();
406 // Make sure SSE2 is enabled; it is available on all X86-64 CPUs.
407 if (Is64Bit && X86SSELevel < SSE2)
408 X86SSELevel = SSE2;
411 // If requesting codegen for X86-64, make sure that 64-bit features
412 // are enabled.
413 if (Is64Bit)
414 HasX86_64 = true;
416 DEBUG(errs() << "Subtarget features: SSELevel " << X86SSELevel
417 << ", 3DNowLevel " << X863DNowLevel
418 << ", 64bit " << HasX86_64 << "\n");
419 assert((!Is64Bit || HasX86_64) &&
420 "64-bit code requested on a subtarget that doesn't support it!");
422 // Set the boolean corresponding to the current target triple, or the default
423 // if one cannot be determined, to true.
424 if (TT.length() > 5) {
425 size_t Pos;
426 if ((Pos = TT.find("-darwin")) != std::string::npos) {
427 TargetType = isDarwin;
429 // Compute the darwin version number.
430 if (isdigit(TT[Pos+7]))
431 DarwinVers = atoi(&TT[Pos+7]);
432 else
433 DarwinVers = 8; // Minimum supported darwin is Tiger.
434 } else if (TT.find("linux") != std::string::npos) {
435 // Linux doesn't imply ELF, but we don't currently support anything else.
436 TargetType = isELF;
437 IsLinux = true;
438 } else if (TT.find("cygwin") != std::string::npos) {
439 TargetType = isCygwin;
440 } else if (TT.find("mingw") != std::string::npos) {
441 TargetType = isMingw;
442 } else if (TT.find("win32") != std::string::npos) {
443 TargetType = isWindows;
444 } else if (TT.find("windows") != std::string::npos) {
445 TargetType = isWindows;
446 } else if (TT.find("-cl") != std::string::npos) {
447 TargetType = isDarwin;
448 DarwinVers = 9;
452 // Stack alignment is 16 bytes on Darwin (both 32 and 64 bit) and for all 64
453 // bit targets.
454 if (TargetType == isDarwin || Is64Bit)
455 stackAlignment = 16;
457 if (StackAlignment)
458 stackAlignment = StackAlignment;