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/raw_ostream.h"
20 #include "llvm/Support/Host.h"
21 #include "llvm/Target/TargetMachine.h"
22 #include "llvm/ADT/SmallVector.h"
24 #define GET_SUBTARGETINFO_ENUM
25 #define GET_SUBTARGETINFO_MC_DESC
26 #define GET_SUBTARGETINFO_TARGET_DESC
27 #define GET_SUBTARGETINFO_CTOR
28 #include "X86GenSubtargetInfo.inc"
36 /// ClassifyBlockAddressReference - Classify a blockaddress reference for the
37 /// current subtarget according to how we should reference it in a non-pcrel
39 unsigned char X86Subtarget::
40 ClassifyBlockAddressReference() const {
41 if (isPICStyleGOT()) // 32-bit ELF targets.
42 return X86II::MO_GOTOFF;
44 if (isPICStyleStubPIC()) // Darwin/32 in PIC mode.
45 return X86II::MO_PIC_BASE_OFFSET;
47 // Direct static reference to label.
48 return X86II::MO_NO_FLAG;
51 /// ClassifyGlobalReference - Classify a global variable reference for the
52 /// current subtarget according to how we should reference it in a non-pcrel
54 unsigned char X86Subtarget::
55 ClassifyGlobalReference(const GlobalValue *GV, const TargetMachine &TM) const {
56 // DLLImport only exists on windows, it is implemented as a load from a
58 if (GV->hasDLLImportLinkage())
59 return X86II::MO_DLLIMPORT;
61 // Determine whether this is a reference to a definition or a declaration.
62 // Materializable GVs (in JIT lazy compilation mode) do not require an extra
64 bool isDecl = GV->hasAvailableExternallyLinkage();
65 if (GV->isDeclaration() && !GV->isMaterializable())
68 // X86-64 in PIC mode.
69 if (isPICStyleRIPRel()) {
70 // Large model never uses stubs.
71 if (TM.getCodeModel() == CodeModel::Large)
72 return X86II::MO_NO_FLAG;
74 if (isTargetDarwin()) {
75 // If symbol visibility is hidden, the extra load is not needed if
76 // target is x86-64 or the symbol is definitely defined in the current
78 if (GV->hasDefaultVisibility() &&
79 (isDecl || GV->isWeakForLinker()))
80 return X86II::MO_GOTPCREL;
81 } else if (!isTargetWin64()) {
82 assert(isTargetELF() && "Unknown rip-relative target");
84 // Extra load is needed for all externally visible.
85 if (!GV->hasLocalLinkage() && GV->hasDefaultVisibility())
86 return X86II::MO_GOTPCREL;
89 return X86II::MO_NO_FLAG;
92 if (isPICStyleGOT()) { // 32-bit ELF targets.
93 // Extra load is needed for all externally visible.
94 if (GV->hasLocalLinkage() || GV->hasHiddenVisibility())
95 return X86II::MO_GOTOFF;
99 if (isPICStyleStubPIC()) { // Darwin/32 in PIC mode.
100 // Determine whether we have a stub reference and/or whether the reference
101 // is relative to the PIC base or not.
103 // If this is a strong reference to a definition, it is definitely not
105 if (!isDecl && !GV->isWeakForLinker())
106 return X86II::MO_PIC_BASE_OFFSET;
108 // Unless we have a symbol with hidden visibility, we have to go through a
109 // normal $non_lazy_ptr stub because this symbol might be resolved late.
110 if (!GV->hasHiddenVisibility()) // Non-hidden $non_lazy_ptr reference.
111 return X86II::MO_DARWIN_NONLAZY_PIC_BASE;
113 // If symbol visibility is hidden, we have a stub for common symbol
114 // references and external declarations.
115 if (isDecl || GV->hasCommonLinkage()) {
116 // Hidden $non_lazy_ptr reference.
117 return X86II::MO_DARWIN_HIDDEN_NONLAZY_PIC_BASE;
120 // Otherwise, no stub.
121 return X86II::MO_PIC_BASE_OFFSET;
124 if (isPICStyleStubNoDynamic()) { // Darwin/32 in -mdynamic-no-pic mode.
125 // Determine whether we have a stub reference.
127 // If this is a strong reference to a definition, it is definitely not
129 if (!isDecl && !GV->isWeakForLinker())
130 return X86II::MO_NO_FLAG;
132 // Unless we have a symbol with hidden visibility, we have to go through a
133 // normal $non_lazy_ptr stub because this symbol might be resolved late.
134 if (!GV->hasHiddenVisibility()) // Non-hidden $non_lazy_ptr reference.
135 return X86II::MO_DARWIN_NONLAZY;
137 // Otherwise, no stub.
138 return X86II::MO_NO_FLAG;
141 // Direct static reference to global.
142 return X86II::MO_NO_FLAG;
146 /// getBZeroEntry - This function returns the name of a function which has an
147 /// interface like the non-standard bzero function, if such a function exists on
148 /// the current subtarget and it is considered prefereable over memset with zero
149 /// passed as the second argument. Otherwise it returns null.
150 const char *X86Subtarget::getBZeroEntry() const {
151 // Darwin 10 has a __bzero entry point for this purpose.
152 if (getTargetTriple().isMacOSX() &&
153 !getTargetTriple().isMacOSXVersionLT(10, 6))
159 /// IsLegalToCallImmediateAddr - Return true if the subtarget allows calls
160 /// to immediate address.
161 bool X86Subtarget::IsLegalToCallImmediateAddr(const TargetMachine &TM) const {
164 return isTargetELF() || TM.getRelocationModel() == Reloc::Static;
167 /// getSpecialAddressLatency - For targets where it is beneficial to
168 /// backschedule instructions that compute addresses, return a value
169 /// indicating the number of scheduling cycles of backscheduling that
170 /// should be attempted.
171 unsigned X86Subtarget::getSpecialAddressLatency() const {
172 // For x86 out-of-order targets, back-schedule address computations so
173 // that loads and stores aren't blocked.
174 // This value was chosen arbitrarily.
178 void X86Subtarget::AutoDetectSubtargetFeatures() {
179 unsigned EAX = 0, EBX = 0, ECX = 0, EDX = 0;
185 if (X86_MC::GetCpuIDAndInfo(0, &EAX, text.u+0, text.u+2, text.u+1))
188 X86_MC::GetCpuIDAndInfo(0x1, &EAX, &EBX, &ECX, &EDX);
190 if ((EDX >> 15) & 1) HasCMov = true;
191 if ((EDX >> 23) & 1) X86SSELevel = MMX;
192 if ((EDX >> 25) & 1) X86SSELevel = SSE1;
193 if ((EDX >> 26) & 1) X86SSELevel = SSE2;
194 if (ECX & 0x1) X86SSELevel = SSE3;
195 if ((ECX >> 9) & 1) X86SSELevel = SSSE3;
196 if ((ECX >> 19) & 1) X86SSELevel = SSE41;
197 if ((ECX >> 20) & 1) X86SSELevel = SSE42;
198 // FIXME: AVX codegen support is not ready.
199 //if ((ECX >> 28) & 1) { HasAVX = true; X86SSELevel = NoMMXSSE; }
201 bool IsIntel = memcmp(text.c, "GenuineIntel", 12) == 0;
202 bool IsAMD = !IsIntel && memcmp(text.c, "AuthenticAMD", 12) == 0;
204 HasCLMUL = IsIntel && ((ECX >> 1) & 0x1);
205 HasFMA3 = IsIntel && ((ECX >> 12) & 0x1);
206 HasPOPCNT = IsIntel && ((ECX >> 23) & 0x1);
207 HasAES = IsIntel && ((ECX >> 25) & 0x1);
209 if (IsIntel || IsAMD) {
210 // Determine if bit test memory instructions are slow.
213 X86_MC::DetectFamilyModel(EAX, Family, Model);
214 IsBTMemSlow = IsAMD || (Family == 6 && Model >= 13);
215 // If it's Nehalem, unaligned memory access is fast.
216 if (Family == 15 && Model == 26)
219 X86_MC::GetCpuIDAndInfo(0x80000001, &EAX, &EBX, &ECX, &EDX);
220 HasX86_64 = (EDX >> 29) & 0x1;
221 HasSSE4A = IsAMD && ((ECX >> 6) & 0x1);
222 HasFMA4 = IsAMD && ((ECX >> 16) & 0x1);
226 X86Subtarget::X86Subtarget(const std::string &TT, const std::string &CPU,
227 const std::string &FS,
228 unsigned StackAlignOverride, bool is64Bit)
229 : X86GenSubtargetInfo(TT, CPU, FS)
230 , PICStyle(PICStyles::None)
231 , X86SSELevel(NoMMXSSE)
232 , X863DNowLevel(NoThreeDNow)
244 , HasVectorUAMem(false)
246 // FIXME: this is a known good value for Yonah. How about others?
247 , MaxInlineSizeThreshold(128)
249 , In64BitMode(is64Bit) {
250 // Determine default and user specified characteristics
251 if (!FS.empty() || !CPU.empty()) {
252 std::string CPUName = CPU;
253 if (CPUName.empty()) {
254 #if defined (__x86_64__) || defined(__i386__)
255 CPUName = sys::getHostCPUName();
261 // If feature string is not empty, parse features string.
262 ParseSubtargetFeatures(CPUName, FS);
264 // All X86-64 CPUs also have SSE2, however user might request no SSE via
265 // -mattr, so don't force SSELevel here.
267 X86SSELevel = NoMMXSSE;
269 // Otherwise, use CPUID to auto-detect feature set.
270 AutoDetectSubtargetFeatures();
272 // Make sure SSE2 is enabled; it is available on all X86-64 CPUs.
273 if (In64BitMode && !HasAVX && X86SSELevel < SSE2)
277 DEBUG(dbgs() << "Subtarget features: SSELevel " << X86SSELevel
278 << ", 3DNowLevel " << X863DNowLevel
279 << ", 64bit " << HasX86_64 << "\n");
280 assert((!In64BitMode || HasX86_64) &&
281 "64-bit code requested on a subtarget that doesn't support it!");
283 // Stack alignment is 16 bytes on Darwin, FreeBSD, Linux and Solaris (both
284 // 32 and 64 bit) and for all 64-bit targets.
285 if (StackAlignOverride)
286 stackAlignment = StackAlignOverride;
287 else if (isTargetDarwin() || isTargetFreeBSD() || isTargetLinux() ||
288 isTargetSolaris() || In64BitMode)