-//===-- X86Subtarget.cpp - X86 Subtarget Information ------------*- C++ -*-===//
+//===-- X86Subtarget.cpp - X86 Subtarget Information ----------------------===//
//
// The LLVM Compiler Infrastructure
//
#include "X86InstrInfo.h"
#include "X86GenSubtarget.inc"
#include "llvm/GlobalValue.h"
-#include "llvm/Support/CommandLine.h"
#include "llvm/Support/Debug.h"
#include "llvm/Support/raw_ostream.h"
+#include "llvm/Support/Host.h"
#include "llvm/Target/TargetMachine.h"
#include "llvm/Target/TargetOptions.h"
+#include "llvm/ADT/SmallVector.h"
using namespace llvm;
#if defined(_MSC_VER)
#include <intrin.h>
#endif
-static cl::opt<X86Subtarget::AsmWriterFlavorTy>
-AsmWriterFlavor("x86-asm-syntax", cl::init(X86Subtarget::Unset),
- cl::desc("Choose style of code to emit from X86 backend:"),
- cl::values(
- clEnumValN(X86Subtarget::ATT, "att", "Emit AT&T-style assembly"),
- clEnumValN(X86Subtarget::Intel, "intel", "Emit Intel-style assembly"),
- clEnumValEnd));
+/// ClassifyBlockAddressReference - Classify a blockaddress reference for the
+/// current subtarget according to how we should reference it in a non-pcrel
+/// context.
+unsigned char X86Subtarget::
+ClassifyBlockAddressReference() const {
+ if (isPICStyleGOT()) // 32-bit ELF targets.
+ return X86II::MO_GOTOFF;
+
+ if (isPICStyleStubPIC()) // Darwin/32 in PIC mode.
+ return X86II::MO_PIC_BASE_OFFSET;
+
+ // Direct static reference to label.
+ return X86II::MO_NO_FLAG;
+}
/// ClassifyGlobalReference - Classify a global variable reference for the
/// current subtarget according to how we should reference it in a non-pcrel
if (GV->hasDLLImportLinkage())
return X86II::MO_DLLIMPORT;
- // GV with ghost linkage (in JIT lazy compilation mode) do not require an
- // extra load from stub.
- bool isDecl = GV->isDeclaration() && !GV->hasNotBeenReadFromBitcode();
+ // Determine whether this is a reference to a definition or a declaration.
+ // Materializable GVs (in JIT lazy compilation mode) do not require an extra
+ // load from stub.
+ bool isDecl = GV->hasAvailableExternallyLinkage();
+ if (GV->isDeclaration() && !GV->isMaterializable())
+ isDecl = true;
// X86-64 in PIC mode.
if (isPICStyleRIPRel()) {
if (GV->hasDefaultVisibility() &&
(isDecl || GV->isWeakForLinker()))
return X86II::MO_GOTPCREL;
- } else {
+ } else if (!isTargetWin64()) {
assert(isTargetELF() && "Unknown rip-relative target");
// Extra load is needed for all externally visible.
// normal $non_lazy_ptr stub because this symbol might be resolved late.
if (!GV->hasHiddenVisibility()) // Non-hidden $non_lazy_ptr reference.
return X86II::MO_DARWIN_NONLAZY;
-
- // If symbol visibility is hidden, we have a stub for common symbol
- // references and external declarations.
- if (isDecl || GV->hasCommonLinkage()) {
- // Hidden $non_lazy_ptr reference.
- return X86II::MO_DARWIN_HIDDEN_NONLAZY;
- }
-
+
// Otherwise, no stub.
return X86II::MO_NO_FLAG;
}
/// GetCpuIDAndInfo - Execute the specified cpuid and return the 4 values in the
/// specified arguments. If we can't run cpuid on the host, return true.
-bool X86::GetCpuIDAndInfo(unsigned value, unsigned *rEAX, unsigned *rEBX,
- unsigned *rECX, unsigned *rEDX) {
-#if defined(__x86_64__) || defined(_M_AMD64)
+static bool GetCpuIDAndInfo(unsigned value, unsigned *rEAX,
+ unsigned *rEBX, unsigned *rECX, unsigned *rEDX) {
+#if defined(__x86_64__) || defined(_M_AMD64) || defined (_M_X64)
#if defined(__GNUC__)
// gcc doesn't know cpuid would clobber ebx/rbx. Preseve it manually.
asm ("movq\t%%rbx, %%rsi\n\t"
char c[12];
} text;
- if (X86::GetCpuIDAndInfo(0, &EAX, text.u+0, text.u+2, text.u+1))
+ if (GetCpuIDAndInfo(0, &EAX, text.u+0, text.u+2, text.u+1))
return;
- X86::GetCpuIDAndInfo(0x1, &EAX, &EBX, &ECX, &EDX);
+ GetCpuIDAndInfo(0x1, &EAX, &EBX, &ECX, &EDX);
- if ((EDX >> 23) & 0x1) X86SSELevel = MMX;
- if ((EDX >> 25) & 0x1) X86SSELevel = SSE1;
- if ((EDX >> 26) & 0x1) X86SSELevel = SSE2;
- if (ECX & 0x1) X86SSELevel = SSE3;
- if ((ECX >> 9) & 0x1) X86SSELevel = SSSE3;
- if ((ECX >> 19) & 0x1) X86SSELevel = SSE41;
- if ((ECX >> 20) & 0x1) X86SSELevel = SSE42;
+ if ((EDX >> 15) & 1) HasCMov = true;
+ if ((EDX >> 23) & 1) X86SSELevel = MMX;
+ if ((EDX >> 25) & 1) X86SSELevel = SSE1;
+ if ((EDX >> 26) & 1) X86SSELevel = SSE2;
+ if (ECX & 0x1) X86SSELevel = SSE3;
+ if ((ECX >> 9) & 1) X86SSELevel = SSSE3;
+ if ((ECX >> 19) & 1) X86SSELevel = SSE41;
+ if ((ECX >> 20) & 1) X86SSELevel = SSE42;
+ // FIXME: AVX codegen support is not ready.
+ //if ((ECX >> 28) & 1) { HasAVX = true; X86SSELevel = NoMMXSSE; }
bool IsIntel = memcmp(text.c, "GenuineIntel", 12) == 0;
bool IsAMD = !IsIntel && memcmp(text.c, "AuthenticAMD", 12) == 0;
- HasFMA3 = IsIntel && ((ECX >> 12) & 0x1);
- HasAVX = ((ECX >> 28) & 0x1);
+ HasCLMUL = IsIntel && ((ECX >> 1) & 0x1);
+ HasFMA3 = IsIntel && ((ECX >> 12) & 0x1);
+ HasAES = IsIntel && ((ECX >> 25) & 0x1);
if (IsIntel || IsAMD) {
// Determine if bit test memory instructions are slow.
unsigned Model = 0;
DetectFamilyModel(EAX, Family, Model);
IsBTMemSlow = IsAMD || (Family == 6 && Model >= 13);
+ // If it's Nehalem, unaligned memory access is fast.
+ if (Family == 15 && Model == 26)
+ IsUAMemFast = true;
- X86::GetCpuIDAndInfo(0x80000001, &EAX, &EBX, &ECX, &EDX);
+ GetCpuIDAndInfo(0x80000001, &EAX, &EBX, &ECX, &EDX);
HasX86_64 = (EDX >> 29) & 0x1;
HasSSE4A = IsAMD && ((ECX >> 6) & 0x1);
HasFMA4 = IsAMD && ((ECX >> 16) & 0x1);
}
}
-static const char *GetCurrentX86CPU() {
- unsigned EAX = 0, EBX = 0, ECX = 0, EDX = 0;
- if (X86::GetCpuIDAndInfo(0x1, &EAX, &EBX, &ECX, &EDX))
- return "generic";
- unsigned Family = 0;
- unsigned Model = 0;
- DetectFamilyModel(EAX, Family, Model);
-
- X86::GetCpuIDAndInfo(0x80000001, &EAX, &EBX, &ECX, &EDX);
- bool Em64T = (EDX >> 29) & 0x1;
- bool HasSSE3 = (ECX & 0x1);
-
- union {
- unsigned u[3];
- char c[12];
- } text;
-
- X86::GetCpuIDAndInfo(0, &EAX, text.u+0, text.u+2, text.u+1);
- if (memcmp(text.c, "GenuineIntel", 12) == 0) {
- switch (Family) {
- case 3:
- return "i386";
- case 4:
- return "i486";
- case 5:
- switch (Model) {
- case 4: return "pentium-mmx";
- default: return "pentium";
- }
- case 6:
- switch (Model) {
- case 1: return "pentiumpro";
- case 3:
- case 5:
- case 6: return "pentium2";
- case 7:
- case 8:
- case 10:
- case 11: return "pentium3";
- case 9:
- case 13: return "pentium-m";
- case 14: return "yonah";
- case 15:
- case 22: // Celeron M 540
- return "core2";
- case 23: // 45nm: Penryn , Wolfdale, Yorkfield (XE)
- return "penryn";
- default: return "i686";
- }
- case 15: {
- switch (Model) {
- case 3:
- case 4:
- case 6: // same as 4, but 65nm
- return (Em64T) ? "nocona" : "prescott";
- case 26:
- return "corei7";
- case 28:
- return "atom";
- default:
- return (Em64T) ? "x86-64" : "pentium4";
- }
- }
-
- default:
- return "generic";
- }
- } else if (memcmp(text.c, "AuthenticAMD", 12) == 0) {
- // FIXME: this poorly matches the generated SubtargetFeatureKV table. There
- // appears to be no way to generate the wide variety of AMD-specific targets
- // from the information returned from CPUID.
- switch (Family) {
- case 4:
- return "i486";
- case 5:
- switch (Model) {
- case 6:
- case 7: return "k6";
- case 8: return "k6-2";
- case 9:
- case 13: return "k6-3";
- default: return "pentium";
- }
- case 6:
- switch (Model) {
- case 4: return "athlon-tbird";
- case 6:
- case 7:
- case 8: return "athlon-mp";
- case 10: return "athlon-xp";
- default: return "athlon";
- }
- case 15:
- if (HasSSE3) {
- return "k8-sse3";
- } else {
- switch (Model) {
- case 1: return "opteron";
- case 5: return "athlon-fx"; // also opteron
- default: return "athlon64";
- }
- }
- case 16:
- return "amdfam10";
- default:
- return "generic";
- }
- } else {
- return "generic";
- }
-}
-
X86Subtarget::X86Subtarget(const std::string &TT, const std::string &FS,
bool is64Bit)
- : AsmFlavor(AsmWriterFlavor)
- , PICStyle(PICStyles::None)
+ : PICStyle(PICStyles::None)
, X86SSELevel(NoMMXSSE)
, X863DNowLevel(NoThreeDNow)
+ , HasCMov(false)
, HasX86_64(false)
+ , HasPOPCNT(false)
, HasSSE4A(false)
, HasAVX(false)
+ , HasAES(false)
+ , HasCLMUL(false)
, HasFMA3(false)
, HasFMA4(false)
, IsBTMemSlow(false)
- , DarwinVers(0)
- , IsLinux(false)
+ , IsUAMemFast(false)
+ , HasVectorUAMem(false)
, stackAlignment(8)
// FIXME: this is a known good value for Yonah. How about others?
, MaxInlineSizeThreshold(128)
- , Is64Bit(is64Bit)
- , TargetType(isELF) { // Default to ELF unless otherwise specified.
+ , TargetTriple(TT)
+ , Is64Bit(is64Bit) {
// default to hard float ABI
if (FloatABIType == FloatABI::Default)
// Determine default and user specified characteristics
if (!FS.empty()) {
// If feature string is not empty, parse features string.
- std::string CPU = GetCurrentX86CPU();
+ std::string CPU = sys::getHostCPUName();
ParseSubtargetFeatures(FS, CPU);
// All X86-64 CPUs also have SSE2, however user might request no SSE via
// -mattr, so don't force SSELevel here.
+ if (HasAVX)
+ X86SSELevel = NoMMXSSE;
} else {
// Otherwise, use CPUID to auto-detect feature set.
AutoDetectSubtargetFeatures();
// Make sure SSE2 is enabled; it is available on all X86-64 CPUs.
- if (Is64Bit && X86SSELevel < SSE2)
+ if (Is64Bit && !HasAVX && X86SSELevel < SSE2)
X86SSELevel = SSE2;
}
// If requesting codegen for X86-64, make sure that 64-bit features
// are enabled.
- if (Is64Bit)
+ if (Is64Bit) {
HasX86_64 = true;
- DEBUG(errs() << "Subtarget features: SSELevel " << X86SSELevel
+ // All 64-bit cpus have cmov support.
+ HasCMov = true;
+ }
+
+ DEBUG(dbgs() << "Subtarget features: SSELevel " << X86SSELevel
<< ", 3DNowLevel " << X863DNowLevel
<< ", 64bit " << HasX86_64 << "\n");
assert((!Is64Bit || HasX86_64) &&
"64-bit code requested on a subtarget that doesn't support it!");
- // Set the boolean corresponding to the current target triple, or the default
- // if one cannot be determined, to true.
- if (TT.length() > 5) {
- size_t Pos;
- if ((Pos = TT.find("-darwin")) != std::string::npos) {
- TargetType = isDarwin;
-
- // Compute the darwin version number.
- if (isdigit(TT[Pos+7]))
- DarwinVers = atoi(&TT[Pos+7]);
- else
- DarwinVers = 8; // Minimum supported darwin is Tiger.
- } else if (TT.find("linux") != std::string::npos) {
- // Linux doesn't imply ELF, but we don't currently support anything else.
- TargetType = isELF;
- IsLinux = true;
- } else if (TT.find("cygwin") != std::string::npos) {
- TargetType = isCygwin;
- } else if (TT.find("mingw") != std::string::npos) {
- TargetType = isMingw;
- } else if (TT.find("win32") != std::string::npos) {
- TargetType = isWindows;
- } else if (TT.find("windows") != std::string::npos) {
- TargetType = isWindows;
- }
- else if (TT.find("-cl") != std::string::npos) {
- TargetType = isDarwin;
- DarwinVers = 9;
- }
- } else if (TT.empty()) {
-#if defined(__CYGWIN__)
- TargetType = isCygwin;
-#elif defined(__MINGW32__) || defined(__MINGW64__)
- TargetType = isMingw;
-#elif defined(__APPLE__)
- TargetType = isDarwin;
-#if __APPLE_CC__ > 5400
- DarwinVers = 9; // GCC 5400+ is Leopard.
-#else
- DarwinVers = 8; // Minimum supported darwin is Tiger.
-#endif
-
-#elif defined(_WIN32) || defined(_WIN64)
- TargetType = isWindows;
-#elif defined(__linux__)
- // Linux doesn't imply ELF, but we don't currently support anything else.
- TargetType = isELF;
- IsLinux = true;
-#endif
- }
-
- // If the asm syntax hasn't been overridden on the command line, use whatever
- // the target wants.
- if (AsmFlavor == X86Subtarget::Unset) {
- AsmFlavor = (TargetType == isWindows)
- ? X86Subtarget::Intel : X86Subtarget::ATT;
- }
-
// Stack alignment is 16 bytes on Darwin (both 32 and 64 bit) and for all 64
// bit targets.
- if (TargetType == isDarwin || Is64Bit)
+ if (isTargetDarwin() || Is64Bit)
stackAlignment = 16;
if (StackAlignment)
stackAlignment = StackAlignment;
}
+
+/// IsCalleePop - Determines whether the callee is required to pop its
+/// own arguments. Callee pop is necessary to support tail calls.
+bool X86Subtarget::IsCalleePop(bool IsVarArg,
+ CallingConv::ID CallingConv) const {
+ if (IsVarArg)
+ return false;
+
+ switch (CallingConv) {
+ default:
+ return false;
+ case CallingConv::X86_StdCall:
+ return !is64Bit();
+ case CallingConv::X86_FastCall:
+ return !is64Bit();
+ case CallingConv::X86_ThisCall:
+ return !is64Bit();
+ case CallingConv::Fast:
+ return GuaranteedTailCallOpt;
+ case CallingConv::GHC:
+ return GuaranteedTailCallOpt;
+ }
+}