1 //===-- X86Subtarget.cpp - X86 Subtarget Information ----------------------===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 //
9 // This file implements the X86 specific subclass of TargetSubtargetInfo.
10 //
11 //===----------------------------------------------------------------------===//
12 
13 #include "X86Subtarget.h"
14 #include "MCTargetDesc/X86BaseInfo.h"
15 #include "X86.h"
16 #include "X86CallLowering.h"
17 #include "X86LegalizerInfo.h"
18 #include "X86MacroFusion.h"
19 #include "X86RegisterBankInfo.h"
20 #include "X86TargetMachine.h"
21 #include "llvm/ADT/Triple.h"
22 #include "llvm/CodeGen/GlobalISel/CallLowering.h"
23 #include "llvm/CodeGen/GlobalISel/InstructionSelect.h"
24 #include "llvm/CodeGen/GlobalISel/InstructionSelector.h"
25 #include "llvm/CodeGen/ScheduleDAGMutation.h"
26 #include "llvm/IR/Attributes.h"
27 #include "llvm/IR/ConstantRange.h"
28 #include "llvm/IR/Function.h"
29 #include "llvm/IR/GlobalValue.h"
30 #include "llvm/Support/Casting.h"
31 #include "llvm/Support/CodeGen.h"
32 #include "llvm/Support/CommandLine.h"
33 #include "llvm/Support/Debug.h"
34 #include "llvm/Support/ErrorHandling.h"
35 #include "llvm/Support/raw_ostream.h"
36 #include "llvm/Target/TargetMachine.h"
37 
38 #if defined(_MSC_VER)
39 #include <intrin.h>
40 #endif
41 
42 using namespace llvm;
43 
44 #define DEBUG_TYPE "subtarget"
45 
46 #define GET_SUBTARGETINFO_TARGET_DESC
47 #define GET_SUBTARGETINFO_CTOR
48 #include "X86GenSubtargetInfo.inc"
49 
50 // Temporary option to control early if-conversion for x86 while adding machine
51 // models.
52 static cl::opt<bool>
53 X86EarlyIfConv("x86-early-ifcvt", cl::Hidden,
54                cl::desc("Enable early if-conversion on X86"));
55 
56 
57 /// Classify a blockaddress reference for the current subtarget according to how
58 /// we should reference it in a non-pcrel context.
59 unsigned char X86Subtarget::classifyBlockAddressReference() const {
60   return classifyLocalReference(nullptr);
61 }
62 
63 /// Classify a global variable reference for the current subtarget according to
64 /// how we should reference it in a non-pcrel context.
65 unsigned char
66 X86Subtarget::classifyGlobalReference(const GlobalValue *GV) const {
67   return classifyGlobalReference(GV, *GV->getParent());
68 }
69 
70 unsigned char
71 X86Subtarget::classifyLocalReference(const GlobalValue *GV) const {
72   // Tagged globals have non-zero upper bits, which makes direct references
73   // require a 64-bit immediate.  On the small code model this causes relocation
74   // errors, so we go through the GOT instead.
75   if (AllowTaggedGlobals && TM.getCodeModel() == CodeModel::Small && GV &&
76       !isa<Function>(GV))
77     return X86II::MO_GOTPCREL_NORELAX;
78 
79   // If we're not PIC, it's not very interesting.
80   if (!isPositionIndependent())
81     return X86II::MO_NO_FLAG;
82 
83   if (is64Bit()) {
84     // 64-bit ELF PIC local references may use GOTOFF relocations.
85     if (isTargetELF()) {
86       switch (TM.getCodeModel()) {
87       // 64-bit small code model is simple: All rip-relative.
88       case CodeModel::Tiny:
89         llvm_unreachable("Tiny codesize model not supported on X86");
90       case CodeModel::Small:
91       case CodeModel::Kernel:
92         return X86II::MO_NO_FLAG;
93 
94       // The large PIC code model uses GOTOFF.
95       case CodeModel::Large:
96         return X86II::MO_GOTOFF;
97 
98       // Medium is a hybrid: RIP-rel for code, GOTOFF for DSO local data.
99       case CodeModel::Medium:
100         // Constant pool and jump table handling pass a nullptr to this
101         // function so we need to use isa_and_nonnull.
102         if (isa_and_nonnull<Function>(GV))
103           return X86II::MO_NO_FLAG; // All code is RIP-relative
104         return X86II::MO_GOTOFF;    // Local symbols use GOTOFF.
105       }
106       llvm_unreachable("invalid code model");
107     }
108 
109     // Otherwise, this is either a RIP-relative reference or a 64-bit movabsq,
110     // both of which use MO_NO_FLAG.
111     return X86II::MO_NO_FLAG;
112   }
113 
114   // The COFF dynamic linker just patches the executable sections.
115   if (isTargetCOFF())
116     return X86II::MO_NO_FLAG;
117 
118   if (isTargetDarwin()) {
119     // 32 bit macho has no relocation for a-b if a is undefined, even if
120     // b is in the section that is being relocated.
121     // This means we have to use o load even for GVs that are known to be
122     // local to the dso.
123     if (GV && (GV->isDeclarationForLinker() || GV->hasCommonLinkage()))
124       return X86II::MO_DARWIN_NONLAZY_PIC_BASE;
125 
126     return X86II::MO_PIC_BASE_OFFSET;
127   }
128 
129   return X86II::MO_GOTOFF;
130 }
131 
132 unsigned char X86Subtarget::classifyGlobalReference(const GlobalValue *GV,
133                                                     const Module &M) const {
134   // The static large model never uses stubs.
135   if (TM.getCodeModel() == CodeModel::Large && !isPositionIndependent())
136     return X86II::MO_NO_FLAG;
137 
138   // Absolute symbols can be referenced directly.
139   if (GV) {
140     if (std::optional<ConstantRange> CR = GV->getAbsoluteSymbolRange()) {
141       // See if we can use the 8-bit immediate form. Note that some instructions
142       // will sign extend the immediate operand, so to be conservative we only
143       // accept the range [0,128).
144       if (CR->getUnsignedMax().ult(128))
145         return X86II::MO_ABS8;
146       else
147         return X86II::MO_NO_FLAG;
148     }
149   }
150 
151   if (TM.shouldAssumeDSOLocal(M, GV))
152     return classifyLocalReference(GV);
153 
154   if (isTargetCOFF()) {
155     // ExternalSymbolSDNode like _tls_index.
156     if (!GV)
157       return X86II::MO_NO_FLAG;
158     if (GV->hasDLLImportStorageClass())
159       return X86II::MO_DLLIMPORT;
160     return X86II::MO_COFFSTUB;
161   }
162   // Some JIT users use *-win32-elf triples; these shouldn't use GOT tables.
163   if (isOSWindows())
164     return X86II::MO_NO_FLAG;
165 
166   if (is64Bit()) {
167     // ELF supports a large, truly PIC code model with non-PC relative GOT
168     // references. Other object file formats do not. Use the no-flag, 64-bit
169     // reference for them.
170     if (TM.getCodeModel() == CodeModel::Large)
171       return isTargetELF() ? X86II::MO_GOT : X86II::MO_NO_FLAG;
172     // Tagged globals have non-zero upper bits, which makes direct references
173     // require a 64-bit immediate. So we can't let the linker relax the
174     // relocation to a 32-bit RIP-relative direct reference.
175     if (AllowTaggedGlobals && GV && !isa<Function>(GV))
176       return X86II::MO_GOTPCREL_NORELAX;
177     return X86II::MO_GOTPCREL;
178   }
179 
180   if (isTargetDarwin()) {
181     if (!isPositionIndependent())
182       return X86II::MO_DARWIN_NONLAZY;
183     return X86II::MO_DARWIN_NONLAZY_PIC_BASE;
184   }
185 
186   // 32-bit ELF references GlobalAddress directly in static relocation model.
187   // We cannot use MO_GOT because EBX may not be set up.
188   if (TM.getRelocationModel() == Reloc::Static)
189     return X86II::MO_NO_FLAG;
190   return X86II::MO_GOT;
191 }
192 
193 unsigned char
194 X86Subtarget::classifyGlobalFunctionReference(const GlobalValue *GV) const {
195   return classifyGlobalFunctionReference(GV, *GV->getParent());
196 }
197 
198 unsigned char
199 X86Subtarget::classifyGlobalFunctionReference(const GlobalValue *GV,
200                                               const Module &M) const {
201   if (TM.shouldAssumeDSOLocal(M, GV))
202     return X86II::MO_NO_FLAG;
203 
204   // Functions on COFF can be non-DSO local for three reasons:
205   // - They are intrinsic functions (!GV)
206   // - They are marked dllimport
207   // - They are extern_weak, and a stub is needed
208   if (isTargetCOFF()) {
209     if (!GV)
210       return X86II::MO_NO_FLAG;
211     if (GV->hasDLLImportStorageClass())
212       return X86II::MO_DLLIMPORT;
213     return X86II::MO_COFFSTUB;
214   }
215 
216   const Function *F = dyn_cast_or_null<Function>(GV);
217 
218   if (isTargetELF()) {
219     if (is64Bit() && F && (CallingConv::X86_RegCall == F->getCallingConv()))
220       // According to psABI, PLT stub clobbers XMM8-XMM15.
221       // In Regcall calling convention those registers are used for passing
222       // parameters. Thus we need to prevent lazy binding in Regcall.
223       return X86II::MO_GOTPCREL;
224     // If PLT must be avoided then the call should be via GOTPCREL.
225     if (((F && F->hasFnAttribute(Attribute::NonLazyBind)) ||
226          (!F && M.getRtLibUseGOT())) &&
227         is64Bit())
228        return X86II::MO_GOTPCREL;
229     // Reference ExternalSymbol directly in static relocation model.
230     if (!is64Bit() && !GV && TM.getRelocationModel() == Reloc::Static)
231       return X86II::MO_NO_FLAG;
232     return X86II::MO_PLT;
233   }
234 
235   if (is64Bit()) {
236     if (F && F->hasFnAttribute(Attribute::NonLazyBind))
237       // If the function is marked as non-lazy, generate an indirect call
238       // which loads from the GOT directly. This avoids runtime overhead
239       // at the cost of eager binding (and one extra byte of encoding).
240       return X86II::MO_GOTPCREL;
241     return X86II::MO_NO_FLAG;
242   }
243 
244   return X86II::MO_NO_FLAG;
245 }
246 
247 /// Return true if the subtarget allows calls to immediate address.
248 bool X86Subtarget::isLegalToCallImmediateAddr() const {
249   // FIXME: I386 PE/COFF supports PC relative calls using IMAGE_REL_I386_REL32
250   // but WinCOFFObjectWriter::RecordRelocation cannot emit them.  Once it does,
251   // the following check for Win32 should be removed.
252   if (Is64Bit || isTargetWin32())
253     return false;
254   return isTargetELF() || TM.getRelocationModel() == Reloc::Static;
255 }
256 
257 void X86Subtarget::initSubtargetFeatures(StringRef CPU, StringRef TuneCPU,
258                                          StringRef FS) {
259   if (CPU.empty())
260     CPU = "generic";
261 
262   if (TuneCPU.empty())
263     TuneCPU = "i586"; // FIXME: "generic" is more modern than llc tests expect.
264 
265   std::string FullFS = X86_MC::ParseX86Triple(TargetTriple);
266   assert(!FullFS.empty() && "Failed to parse X86 triple");
267 
268   if (!FS.empty())
269     FullFS = (Twine(FullFS) + "," + FS).str();
270 
271   // Parse features string and set the CPU.
272   ParseSubtargetFeatures(CPU, TuneCPU, FullFS);
273 
274   // All CPUs that implement SSE4.2 or SSE4A support unaligned accesses of
275   // 16-bytes and under that are reasonably fast. These features were
276   // introduced with Intel's Nehalem/Silvermont and AMD's Family10h
277   // micro-architectures respectively.
278   if (hasSSE42() || hasSSE4A())
279     IsUnalignedMem16Slow = false;
280 
281   LLVM_DEBUG(dbgs() << "Subtarget features: SSELevel " << X86SSELevel
282                     << ", 3DNowLevel " << X863DNowLevel << ", 64bit "
283                     << HasX86_64 << "\n");
284   if (Is64Bit && !HasX86_64)
285     report_fatal_error("64-bit code requested on a subtarget that doesn't "
286                        "support it!");
287 
288   // Stack alignment is 16 bytes on Darwin, Linux, kFreeBSD, NaCl, and for all
289   // 64-bit targets.  On Solaris (32-bit), stack alignment is 4 bytes
290   // following the i386 psABI, while on Illumos it is always 16 bytes.
291   if (StackAlignOverride)
292     stackAlignment = *StackAlignOverride;
293   else if (isTargetDarwin() || isTargetLinux() || isTargetKFreeBSD() ||
294            isTargetNaCl() || Is64Bit)
295     stackAlignment = Align(16);
296 
297   // Consume the vector width attribute or apply any target specific limit.
298   if (PreferVectorWidthOverride)
299     PreferVectorWidth = PreferVectorWidthOverride;
300   else if (Prefer128Bit)
301     PreferVectorWidth = 128;
302   else if (Prefer256Bit)
303     PreferVectorWidth = 256;
304 }
305 
306 X86Subtarget &X86Subtarget::initializeSubtargetDependencies(StringRef CPU,
307                                                             StringRef TuneCPU,
308                                                             StringRef FS) {
309   initSubtargetFeatures(CPU, TuneCPU, FS);
310   return *this;
311 }
312 
313 X86Subtarget::X86Subtarget(const Triple &TT, StringRef CPU, StringRef TuneCPU,
314                            StringRef FS, const X86TargetMachine &TM,
315                            MaybeAlign StackAlignOverride,
316                            unsigned PreferVectorWidthOverride,
317                            unsigned RequiredVectorWidth)
318     : X86GenSubtargetInfo(TT, CPU, TuneCPU, FS),
319       PICStyle(PICStyles::Style::None), TM(TM), TargetTriple(TT),
320       StackAlignOverride(StackAlignOverride),
321       PreferVectorWidthOverride(PreferVectorWidthOverride),
322       RequiredVectorWidth(RequiredVectorWidth),
323       InstrInfo(initializeSubtargetDependencies(CPU, TuneCPU, FS)),
324       TLInfo(TM, *this), FrameLowering(*this, getStackAlignment()) {
325   // Determine the PICStyle based on the target selected.
326   if (!isPositionIndependent())
327     setPICStyle(PICStyles::Style::None);
328   else if (is64Bit())
329     setPICStyle(PICStyles::Style::RIPRel);
330   else if (isTargetCOFF())
331     setPICStyle(PICStyles::Style::None);
332   else if (isTargetDarwin())
333     setPICStyle(PICStyles::Style::StubPIC);
334   else if (isTargetELF())
335     setPICStyle(PICStyles::Style::GOT);
336 
337   CallLoweringInfo.reset(new X86CallLowering(*getTargetLowering()));
338   Legalizer.reset(new X86LegalizerInfo(*this, TM));
339 
340   auto *RBI = new X86RegisterBankInfo(*getRegisterInfo());
341   RegBankInfo.reset(RBI);
342   InstSelector.reset(createX86InstructionSelector(TM, *this, *RBI));
343 }
344 
345 const CallLowering *X86Subtarget::getCallLowering() const {
346   return CallLoweringInfo.get();
347 }
348 
349 InstructionSelector *X86Subtarget::getInstructionSelector() const {
350   return InstSelector.get();
351 }
352 
353 const LegalizerInfo *X86Subtarget::getLegalizerInfo() const {
354   return Legalizer.get();
355 }
356 
357 const RegisterBankInfo *X86Subtarget::getRegBankInfo() const {
358   return RegBankInfo.get();
359 }
360 
361 bool X86Subtarget::enableEarlyIfConversion() const {
362   return canUseCMOV() && X86EarlyIfConv;
363 }
364 
365 void X86Subtarget::getPostRAMutations(
366     std::vector<std::unique_ptr<ScheduleDAGMutation>> &Mutations) const {
367   Mutations.push_back(createX86MacroFusionDAGMutation());
368 }
369 
370 bool X86Subtarget::isPositionIndependent() const {
371   return TM.isPositionIndependent();
372 }
373