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 "GISel/X86CallLowering.h" |
15 | #include "GISel/X86LegalizerInfo.h" |
16 | #include "GISel/X86RegisterBankInfo.h" |
17 | #include "MCTargetDesc/X86BaseInfo.h" |
18 | #include "X86.h" |
19 | #include "X86MacroFusion.h" |
20 | #include "X86TargetMachine.h" |
21 | #include "llvm/CodeGen/GlobalISel/CallLowering.h" |
22 | #include "llvm/CodeGen/GlobalISel/InstructionSelect.h" |
23 | #include "llvm/CodeGen/GlobalISel/InstructionSelector.h" |
24 | #include "llvm/CodeGen/ScheduleDAGMutation.h" |
25 | #include "llvm/IR/Attributes.h" |
26 | #include "llvm/IR/ConstantRange.h" |
27 | #include "llvm/IR/Function.h" |
28 | #include "llvm/IR/GlobalValue.h" |
29 | #include "llvm/IR/Module.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 | #include "llvm/TargetParser/Triple.h" |
38 | |
39 | #if defined(_MSC_VER) |
40 | #include <intrin.h> |
41 | #endif |
42 | |
43 | using namespace llvm; |
44 | |
45 | #define DEBUG_TYPE "subtarget" |
46 | |
47 | #define GET_SUBTARGETINFO_TARGET_DESC |
48 | #define GET_SUBTARGETINFO_CTOR |
49 | #include "X86GenSubtargetInfo.inc" |
50 | |
51 | // Temporary option to control early if-conversion for x86 while adding machine |
52 | // models. |
53 | static cl::opt<bool> |
54 | X86EarlyIfConv("x86-early-ifcvt" , cl::Hidden, |
55 | cl::desc("Enable early if-conversion on X86" )); |
56 | |
57 | |
58 | /// Classify a blockaddress reference for the current subtarget according to how |
59 | /// we should reference it in a non-pcrel context. |
60 | unsigned char X86Subtarget::classifyBlockAddressReference() const { |
61 | return classifyLocalReference(GV: nullptr); |
62 | } |
63 | |
64 | /// Classify a global variable reference for the current subtarget according to |
65 | /// how we should reference it in a non-pcrel context. |
66 | unsigned char |
67 | X86Subtarget::classifyGlobalReference(const GlobalValue *GV) const { |
68 | return classifyGlobalReference(GV, M: *GV->getParent()); |
69 | } |
70 | |
71 | unsigned char |
72 | X86Subtarget::classifyLocalReference(const GlobalValue *GV) const { |
73 | CodeModel::Model CM = TM.getCodeModel(); |
74 | // Tagged globals have non-zero upper bits, which makes direct references |
75 | // require a 64-bit immediate. With the small/medium code models this causes |
76 | // relocation errors, so we go through the GOT instead. |
77 | if (AllowTaggedGlobals && CM != CodeModel::Large && GV && !isa<Function>(Val: GV)) |
78 | return X86II::MO_GOTPCREL_NORELAX; |
79 | |
80 | // If we're not PIC, it's not very interesting. |
81 | if (!isPositionIndependent()) |
82 | return X86II::MO_NO_FLAG; |
83 | |
84 | if (is64Bit()) { |
85 | // 64-bit ELF PIC local references may use GOTOFF relocations. |
86 | if (isTargetELF()) { |
87 | assert(CM != CodeModel::Tiny && |
88 | "Tiny codesize model not supported on X86" ); |
89 | // In the large code model, all text is far from any global data, so we |
90 | // use GOTOFF. |
91 | if (CM == CodeModel::Large) |
92 | return X86II::MO_GOTOFF; |
93 | // Large GlobalValues use GOTOFF, otherwise use RIP-rel access. |
94 | if (GV) |
95 | return TM.isLargeGlobalValue(GV) ? X86II::MO_GOTOFF : X86II::MO_NO_FLAG; |
96 | // GV == nullptr is for all other non-GlobalValue global data like the |
97 | // constant pool, jump tables, labels, etc. The small and medium code |
98 | // models treat these as accessible with a RIP-rel access. |
99 | return X86II::MO_NO_FLAG; |
100 | } |
101 | |
102 | // Otherwise, this is either a RIP-relative reference or a 64-bit movabsq, |
103 | // both of which use MO_NO_FLAG. |
104 | return X86II::MO_NO_FLAG; |
105 | } |
106 | |
107 | // The COFF dynamic linker just patches the executable sections. |
108 | if (isTargetCOFF()) |
109 | return X86II::MO_NO_FLAG; |
110 | |
111 | if (isTargetDarwin()) { |
112 | // 32 bit macho has no relocation for a-b if a is undefined, even if |
113 | // b is in the section that is being relocated. |
114 | // This means we have to use o load even for GVs that are known to be |
115 | // local to the dso. |
116 | if (GV && (GV->isDeclarationForLinker() || GV->hasCommonLinkage())) |
117 | return X86II::MO_DARWIN_NONLAZY_PIC_BASE; |
118 | |
119 | return X86II::MO_PIC_BASE_OFFSET; |
120 | } |
121 | |
122 | return X86II::MO_GOTOFF; |
123 | } |
124 | |
125 | unsigned char X86Subtarget::classifyGlobalReference(const GlobalValue *GV, |
126 | const Module &M) const { |
127 | // The static large model never uses stubs. |
128 | if (TM.getCodeModel() == CodeModel::Large && !isPositionIndependent()) |
129 | return X86II::MO_NO_FLAG; |
130 | |
131 | // Absolute symbols can be referenced directly. |
132 | if (GV) { |
133 | if (std::optional<ConstantRange> CR = GV->getAbsoluteSymbolRange()) { |
134 | // See if we can use the 8-bit immediate form. Note that some instructions |
135 | // will sign extend the immediate operand, so to be conservative we only |
136 | // accept the range [0,128). |
137 | if (CR->getUnsignedMax().ult(RHS: 128)) |
138 | return X86II::MO_ABS8; |
139 | else |
140 | return X86II::MO_NO_FLAG; |
141 | } |
142 | } |
143 | |
144 | if (TM.shouldAssumeDSOLocal(GV)) |
145 | return classifyLocalReference(GV); |
146 | |
147 | if (isTargetCOFF()) { |
148 | // ExternalSymbolSDNode like _tls_index. |
149 | if (!GV) |
150 | return X86II::MO_NO_FLAG; |
151 | if (GV->hasDLLImportStorageClass()) |
152 | return X86II::MO_DLLIMPORT; |
153 | return X86II::MO_COFFSTUB; |
154 | } |
155 | // Some JIT users use *-win32-elf triples; these shouldn't use GOT tables. |
156 | if (isOSWindows()) |
157 | return X86II::MO_NO_FLAG; |
158 | |
159 | if (is64Bit()) { |
160 | // ELF supports a large, truly PIC code model with non-PC relative GOT |
161 | // references. Other object file formats do not. Use the no-flag, 64-bit |
162 | // reference for them. |
163 | if (TM.getCodeModel() == CodeModel::Large) |
164 | return isTargetELF() ? X86II::MO_GOT : X86II::MO_NO_FLAG; |
165 | // Tagged globals have non-zero upper bits, which makes direct references |
166 | // require a 64-bit immediate. So we can't let the linker relax the |
167 | // relocation to a 32-bit RIP-relative direct reference. |
168 | if (AllowTaggedGlobals && GV && !isa<Function>(Val: GV)) |
169 | return X86II::MO_GOTPCREL_NORELAX; |
170 | return X86II::MO_GOTPCREL; |
171 | } |
172 | |
173 | if (isTargetDarwin()) { |
174 | if (!isPositionIndependent()) |
175 | return X86II::MO_DARWIN_NONLAZY; |
176 | return X86II::MO_DARWIN_NONLAZY_PIC_BASE; |
177 | } |
178 | |
179 | // 32-bit ELF references GlobalAddress directly in static relocation model. |
180 | // We cannot use MO_GOT because EBX may not be set up. |
181 | if (TM.getRelocationModel() == Reloc::Static) |
182 | return X86II::MO_NO_FLAG; |
183 | return X86II::MO_GOT; |
184 | } |
185 | |
186 | unsigned char |
187 | X86Subtarget::classifyGlobalFunctionReference(const GlobalValue *GV) const { |
188 | return classifyGlobalFunctionReference(GV, M: *GV->getParent()); |
189 | } |
190 | |
191 | unsigned char |
192 | X86Subtarget::classifyGlobalFunctionReference(const GlobalValue *GV, |
193 | const Module &M) const { |
194 | if (TM.shouldAssumeDSOLocal(GV)) |
195 | return X86II::MO_NO_FLAG; |
196 | |
197 | // Functions on COFF can be non-DSO local for three reasons: |
198 | // - They are intrinsic functions (!GV) |
199 | // - They are marked dllimport |
200 | // - They are extern_weak, and a stub is needed |
201 | if (isTargetCOFF()) { |
202 | if (!GV) |
203 | return X86II::MO_NO_FLAG; |
204 | if (GV->hasDLLImportStorageClass()) |
205 | return X86II::MO_DLLIMPORT; |
206 | return X86II::MO_COFFSTUB; |
207 | } |
208 | |
209 | const Function *F = dyn_cast_or_null<Function>(Val: GV); |
210 | |
211 | if (isTargetELF()) { |
212 | if (is64Bit() && F && (CallingConv::X86_RegCall == F->getCallingConv())) |
213 | // According to psABI, PLT stub clobbers XMM8-XMM15. |
214 | // In Regcall calling convention those registers are used for passing |
215 | // parameters. Thus we need to prevent lazy binding in Regcall. |
216 | return X86II::MO_GOTPCREL; |
217 | // If PLT must be avoided then the call should be via GOTPCREL. |
218 | if (((F && F->hasFnAttribute(Kind: Attribute::NonLazyBind)) || |
219 | (!F && M.getRtLibUseGOT())) && |
220 | is64Bit()) |
221 | return X86II::MO_GOTPCREL; |
222 | // Reference ExternalSymbol directly in static relocation model. |
223 | if (!is64Bit() && !GV && TM.getRelocationModel() == Reloc::Static) |
224 | return X86II::MO_NO_FLAG; |
225 | return X86II::MO_PLT; |
226 | } |
227 | |
228 | if (is64Bit()) { |
229 | if (F && F->hasFnAttribute(Kind: Attribute::NonLazyBind)) |
230 | // If the function is marked as non-lazy, generate an indirect call |
231 | // which loads from the GOT directly. This avoids runtime overhead |
232 | // at the cost of eager binding (and one extra byte of encoding). |
233 | return X86II::MO_GOTPCREL; |
234 | return X86II::MO_NO_FLAG; |
235 | } |
236 | |
237 | return X86II::MO_NO_FLAG; |
238 | } |
239 | |
240 | /// Return true if the subtarget allows calls to immediate address. |
241 | bool X86Subtarget::isLegalToCallImmediateAddr() const { |
242 | // FIXME: I386 PE/COFF supports PC relative calls using IMAGE_REL_I386_REL32 |
243 | // but WinCOFFObjectWriter::RecordRelocation cannot emit them. Once it does, |
244 | // the following check for Win32 should be removed. |
245 | if (Is64Bit || isTargetWin32()) |
246 | return false; |
247 | return isTargetELF() || TM.getRelocationModel() == Reloc::Static; |
248 | } |
249 | |
250 | void X86Subtarget::initSubtargetFeatures(StringRef CPU, StringRef TuneCPU, |
251 | StringRef FS) { |
252 | if (CPU.empty()) |
253 | CPU = "generic" ; |
254 | |
255 | if (TuneCPU.empty()) |
256 | TuneCPU = "i586" ; // FIXME: "generic" is more modern than llc tests expect. |
257 | |
258 | std::string FullFS = X86_MC::ParseX86Triple(TT: TargetTriple); |
259 | assert(!FullFS.empty() && "Failed to parse X86 triple" ); |
260 | |
261 | if (!FS.empty()) |
262 | FullFS = (Twine(FullFS) + "," + FS).str(); |
263 | |
264 | // Attach EVEX512 feature when we have AVX512 features with a default CPU. |
265 | // "pentium4" is default CPU for 32-bit targets. |
266 | // "x86-64" is default CPU for 64-bit targets. |
267 | if (CPU == "generic" || CPU == "pentium4" || CPU == "x86-64" ) { |
268 | size_t posNoEVEX512 = FS.rfind(Str: "-evex512" ); |
269 | // Make sure we won't be cheated by "-avx512fp16". |
270 | size_t posNoAVX512F = |
271 | FS.ends_with(Suffix: "-avx512f" ) ? FS.size() - 8 : FS.rfind(Str: "-avx512f," ); |
272 | size_t posEVEX512 = FS.rfind(Str: "+evex512" ); |
273 | // Any AVX512XXX will enable AVX512F. |
274 | size_t posAVX512F = FS.rfind(Str: "+avx512" ); |
275 | |
276 | if (posAVX512F != StringRef::npos && |
277 | (posNoAVX512F == StringRef::npos || posNoAVX512F < posAVX512F)) |
278 | if (posEVEX512 == StringRef::npos && posNoEVEX512 == StringRef::npos) |
279 | FullFS += ",+evex512" ; |
280 | } |
281 | |
282 | // Parse features string and set the CPU. |
283 | ParseSubtargetFeatures(CPU, TuneCPU, FS: FullFS); |
284 | |
285 | // All CPUs that implement SSE4.2 or SSE4A support unaligned accesses of |
286 | // 16-bytes and under that are reasonably fast. These features were |
287 | // introduced with Intel's Nehalem/Silvermont and AMD's Family10h |
288 | // micro-architectures respectively. |
289 | if (hasSSE42() || hasSSE4A()) |
290 | IsUnalignedMem16Slow = false; |
291 | |
292 | LLVM_DEBUG(dbgs() << "Subtarget features: SSELevel " << X86SSELevel |
293 | << ", MMX " << HasMMX << ", 64bit " << HasX86_64 << "\n" ); |
294 | if (Is64Bit && !HasX86_64) |
295 | report_fatal_error(reason: "64-bit code requested on a subtarget that doesn't " |
296 | "support it!" ); |
297 | |
298 | // Stack alignment is 16 bytes on Darwin, Linux, kFreeBSD, NaCl, and for all |
299 | // 64-bit targets. On Solaris (32-bit), stack alignment is 4 bytes |
300 | // following the i386 psABI, while on Illumos it is always 16 bytes. |
301 | if (StackAlignOverride) |
302 | stackAlignment = *StackAlignOverride; |
303 | else if (isTargetDarwin() || isTargetLinux() || isTargetKFreeBSD() || |
304 | isTargetNaCl() || Is64Bit) |
305 | stackAlignment = Align(16); |
306 | |
307 | // Consume the vector width attribute or apply any target specific limit. |
308 | if (PreferVectorWidthOverride) |
309 | PreferVectorWidth = PreferVectorWidthOverride; |
310 | else if (Prefer128Bit) |
311 | PreferVectorWidth = 128; |
312 | else if (Prefer256Bit) |
313 | PreferVectorWidth = 256; |
314 | } |
315 | |
316 | X86Subtarget &X86Subtarget::initializeSubtargetDependencies(StringRef CPU, |
317 | StringRef TuneCPU, |
318 | StringRef FS) { |
319 | initSubtargetFeatures(CPU, TuneCPU, FS); |
320 | return *this; |
321 | } |
322 | |
323 | X86Subtarget::X86Subtarget(const Triple &TT, StringRef CPU, StringRef TuneCPU, |
324 | StringRef FS, const X86TargetMachine &TM, |
325 | MaybeAlign StackAlignOverride, |
326 | unsigned PreferVectorWidthOverride, |
327 | unsigned RequiredVectorWidth) |
328 | : X86GenSubtargetInfo(TT, CPU, TuneCPU, FS), |
329 | PICStyle(PICStyles::Style::None), TM(TM), TargetTriple(TT), |
330 | StackAlignOverride(StackAlignOverride), |
331 | PreferVectorWidthOverride(PreferVectorWidthOverride), |
332 | RequiredVectorWidth(RequiredVectorWidth), |
333 | InstrInfo(initializeSubtargetDependencies(CPU, TuneCPU, FS)), |
334 | TLInfo(TM, *this), FrameLowering(*this, getStackAlignment()) { |
335 | // Determine the PICStyle based on the target selected. |
336 | if (!isPositionIndependent() || TM.getCodeModel() == CodeModel::Large) |
337 | // With the large code model, None forces all memory accesses to be indirect |
338 | // rather than RIP-relative. |
339 | setPICStyle(PICStyles::Style::None); |
340 | else if (is64Bit()) |
341 | setPICStyle(PICStyles::Style::RIPRel); |
342 | else if (isTargetCOFF()) |
343 | setPICStyle(PICStyles::Style::None); |
344 | else if (isTargetDarwin()) |
345 | setPICStyle(PICStyles::Style::StubPIC); |
346 | else if (isTargetELF()) |
347 | setPICStyle(PICStyles::Style::GOT); |
348 | |
349 | CallLoweringInfo.reset(p: new X86CallLowering(*getTargetLowering())); |
350 | Legalizer.reset(p: new X86LegalizerInfo(*this, TM)); |
351 | |
352 | auto *RBI = new X86RegisterBankInfo(*getRegisterInfo()); |
353 | RegBankInfo.reset(p: RBI); |
354 | InstSelector.reset(p: createX86InstructionSelector(TM, *this, *RBI)); |
355 | } |
356 | |
357 | const CallLowering *X86Subtarget::getCallLowering() const { |
358 | return CallLoweringInfo.get(); |
359 | } |
360 | |
361 | InstructionSelector *X86Subtarget::getInstructionSelector() const { |
362 | return InstSelector.get(); |
363 | } |
364 | |
365 | const LegalizerInfo *X86Subtarget::getLegalizerInfo() const { |
366 | return Legalizer.get(); |
367 | } |
368 | |
369 | const RegisterBankInfo *X86Subtarget::getRegBankInfo() const { |
370 | return RegBankInfo.get(); |
371 | } |
372 | |
373 | bool X86Subtarget::enableEarlyIfConversion() const { |
374 | return canUseCMOV() && X86EarlyIfConv; |
375 | } |
376 | |
377 | void X86Subtarget::getPostRAMutations( |
378 | std::vector<std::unique_ptr<ScheduleDAGMutation>> &Mutations) const { |
379 | Mutations.push_back(x: createX86MacroFusionDAGMutation()); |
380 | } |
381 | |
382 | bool X86Subtarget::isPositionIndependent() const { |
383 | return TM.isPositionIndependent(); |
384 | } |
385 | |