1 //===-- Emitter.cpp - Write machine code to executable memory -------------===//
3 // The LLVM Compiler Infrastructure
5 // This file was developed by the LLVM research group and is distributed under
6 // the University of Illinois Open Source License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file defines a MachineCodeEmitter object that is used by the JIT to
11 // write machine code to memory and remember where relocatable values are.
13 //===----------------------------------------------------------------------===//
15 #define DEBUG_TYPE "jit"
17 #include "llvm/Constant.h"
18 #include "llvm/Module.h"
19 #include "llvm/CodeGen/MachineCodeEmitter.h"
20 #include "llvm/CodeGen/MachineFunction.h"
21 #include "llvm/CodeGen/MachineConstantPool.h"
22 #include "llvm/CodeGen/MachineRelocation.h"
23 #include "llvm/Target/TargetData.h"
24 #include "llvm/Target/TargetJITInfo.h"
25 #include "llvm/Support/Debug.h"
26 #include "llvm/ADT/Statistic.h"
27 #include "llvm/System/Memory.h"
31 Statistic<> NumBytes("jit", "Number of bytes of machine code compiled");
36 //===----------------------------------------------------------------------===//
37 // JITMemoryManager code.
40 /// JITMemoryManager - Manage memory for the JIT code generation in a logical,
41 /// sane way. This splits a large block of MAP_NORESERVE'd memory into two
42 /// sections, one for function stubs, one for the functions themselves. We
43 /// have to do this because we may need to emit a function stub while in the
44 /// middle of emitting a function, and we don't know how large the function we
45 /// are emitting is. This never bothers to release the memory, because when
46 /// we are ready to destroy the JIT, the program exits.
47 class JITMemoryManager {
48 sys::MemoryBlock MemBlock; // Virtual memory block allocated RWX
49 unsigned char *MemBase; // Base of block of memory, start of stub mem
50 unsigned char *FunctionBase; // Start of the function body area
51 unsigned char *CurStubPtr, *CurFunctionPtr;
55 inline unsigned char *allocateStub(unsigned StubSize);
56 inline unsigned char *startFunctionBody();
57 inline void endFunctionBody(unsigned char *FunctionEnd);
61 JITMemoryManager::JITMemoryManager() {
62 // Allocate a 16M block of memory...
63 MemBlock = sys::Memory::AllocateRWX((16 << 20));
64 MemBase = reinterpret_cast<unsigned char*>(MemBlock.base());
65 FunctionBase = MemBase + 512*1024; // Use 512k for stubs
67 // Allocate stubs backwards from the function base, allocate functions forward
68 // from the function base.
69 CurStubPtr = CurFunctionPtr = FunctionBase;
72 unsigned char *JITMemoryManager::allocateStub(unsigned StubSize) {
73 CurStubPtr -= StubSize;
74 if (CurStubPtr < MemBase) {
75 std::cerr << "JIT ran out of memory for function stubs!\n";
81 unsigned char *JITMemoryManager::startFunctionBody() {
82 // Round up to an even multiple of 8 bytes, this should eventually be target
84 return (unsigned char*)(((intptr_t)CurFunctionPtr + 7) & ~7);
87 void JITMemoryManager::endFunctionBody(unsigned char *FunctionEnd) {
88 assert(FunctionEnd > CurFunctionPtr);
89 CurFunctionPtr = FunctionEnd;
92 //===----------------------------------------------------------------------===//
93 // JIT lazy compilation code.
96 /// JITResolver - Keep track of, and resolve, call sites for functions that
97 /// have not yet been compiled.
99 /// The MCE to use to emit stubs with.
100 MachineCodeEmitter &MCE;
102 // FunctionToStubMap - Keep track of the stub created for a particular
103 // function so that we can reuse them if necessary.
104 std::map<Function*, void*> FunctionToStubMap;
106 // StubToFunctionMap - Keep track of the function that each stub corresponds
108 std::map<void*, Function*> StubToFunctionMap;
111 JITResolver(MachineCodeEmitter &mce) : MCE(mce) {}
113 /// getFunctionStub - This returns a pointer to a function stub, creating
114 /// one on demand as needed.
115 void *getFunctionStub(Function *F);
117 /// JITCompilerFn - This function is called to resolve a stub to a compiled
118 /// address. If the LLVM Function corresponding to the stub has not yet
119 /// been compiled, this function compiles it first.
120 static void *JITCompilerFn(void *Stub);
124 /// getJITResolver - This function returns the one instance of the JIT resolver.
126 static JITResolver &getJITResolver(MachineCodeEmitter *MCE = 0) {
127 static JITResolver TheJITResolver(*MCE);
128 return TheJITResolver;
131 /// getFunctionStub - This returns a pointer to a function stub, creating
132 /// one on demand as needed.
133 void *JITResolver::getFunctionStub(Function *F) {
134 /// Get the target-specific JIT resolver function.
135 static TargetJITInfo::LazyResolverFn LazyResolverFn =
136 TheJIT->getJITInfo().getLazyResolverFunction(JITResolver::JITCompilerFn);
138 // If we already have a stub for this function, recycle it.
139 void *&Stub = FunctionToStubMap[F];
140 if (Stub) return Stub;
142 // Otherwise, codegen a new stub. For now, the stub will call the lazy
143 // resolver function.
144 Stub = TheJIT->getJITInfo().emitFunctionStub((void*)LazyResolverFn, MCE);
146 // Finally, keep track of the stub-to-Function mapping so that the
147 // JITCompilerFn knows which function to compile!
148 StubToFunctionMap[Stub] = F;
152 /// JITCompilerFn - This function is called when a lazy compilation stub has
153 /// been entered. It looks up which function this stub corresponds to, compiles
154 /// it if necessary, then returns the resultant function pointer.
155 void *JITResolver::JITCompilerFn(void *Stub) {
156 JITResolver &JR = getJITResolver();
158 // The address given to us for the stub may not be exactly right, it might be
159 // a little bit after the stub. As such, use upper_bound to find it.
160 std::map<void*, Function*>::iterator I =
161 JR.StubToFunctionMap.upper_bound(Stub);
162 assert(I != JR.StubToFunctionMap.begin() && "This is not a known stub!");
163 Function *F = (--I)->second;
165 // The target function will rewrite the stub so that the compilation callback
166 // function is no longer called from this stub.
167 JR.StubToFunctionMap.erase(I);
169 DEBUG(std::cerr << "Lazily resolving function '" << F->getName()
170 << "' In stub ptr = " << Stub << " actual ptr = "
171 << I->first << "\n");
173 void *Result = TheJIT->getPointerToFunction(F);
175 // We don't need to reuse this stub in the future, as F is now compiled.
176 JR.FunctionToStubMap.erase(F);
178 // FIXME: We could rewrite all references to this stub if we knew them.
183 //===----------------------------------------------------------------------===//
184 // JIT MachineCodeEmitter code.
187 /// Emitter - The JIT implementation of the MachineCodeEmitter, which is used
188 /// to output functions to memory for execution.
189 class Emitter : public MachineCodeEmitter {
190 JITMemoryManager MemMgr;
192 // CurBlock - The start of the current block of memory. CurByte - The
193 // current byte being emitted to.
194 unsigned char *CurBlock, *CurByte;
196 // When outputting a function stub in the context of some other function, we
197 // save CurBlock and CurByte here.
198 unsigned char *SavedCurBlock, *SavedCurByte;
200 // ConstantPoolAddresses - Contains the location for each entry in the
202 std::vector<void*> ConstantPoolAddresses;
204 /// Relocations - These are the relocations that the function needs, as
206 std::vector<MachineRelocation> Relocations;
208 Emitter(JIT &jit) { TheJIT = &jit; }
210 virtual void startFunction(MachineFunction &F);
211 virtual void finishFunction(MachineFunction &F);
212 virtual void emitConstantPool(MachineConstantPool *MCP);
213 virtual void startFunctionStub(unsigned StubSize);
214 virtual void* finishFunctionStub(const Function *F);
215 virtual void emitByte(unsigned char B);
216 virtual void emitWord(unsigned W);
217 virtual void emitWordAt(unsigned W, unsigned *Ptr);
219 virtual void addRelocation(const MachineRelocation &MR) {
220 Relocations.push_back(MR);
223 virtual uint64_t getCurrentPCValue();
224 virtual uint64_t getCurrentPCOffset();
225 virtual uint64_t getGlobalValueAddress(GlobalValue *V);
226 virtual uint64_t getGlobalValueAddress(const char *Name);
227 virtual uint64_t getConstantPoolEntryAddress(unsigned Entry);
229 // forceCompilationOf - Force the compilation of the specified function, and
230 // return its address, because we REALLY need the address now.
232 // FIXME: This is JIT specific!
234 virtual uint64_t forceCompilationOf(Function *F);
237 void *getPointerToGlobal(GlobalValue *GV);
241 MachineCodeEmitter *JIT::createEmitter(JIT &jit) {
242 return new Emitter(jit);
245 void *Emitter::getPointerToGlobal(GlobalValue *V) {
246 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(V)) {
247 /// FIXME: If we straightened things out, this could actually emit the
248 /// global immediately instead of queuing it for codegen later!
249 GlobalVariable *GV = cast<GlobalVariable>(V);
250 return TheJIT->getOrEmitGlobalVariable(GV);
253 // If we have already compiled the function, return a pointer to its body.
254 Function *F = cast<Function>(V);
255 void *ResultPtr = TheJIT->getPointerToGlobalIfAvailable(F);
256 if (ResultPtr) return ResultPtr;
258 if (F->hasExternalLinkage()) {
259 // If this is an external function pointer, we can force the JIT to
260 // 'compile' it, which really just adds it to the map.
261 return TheJIT->getPointerToFunction(F);
264 // Otherwise, we have to emit a lazy resolving stub.
265 return getJITResolver(this).getFunctionStub(F);
268 void Emitter::startFunction(MachineFunction &F) {
269 CurByte = CurBlock = MemMgr.startFunctionBody();
270 TheJIT->addGlobalMapping(F.getFunction(), CurBlock);
273 void Emitter::finishFunction(MachineFunction &F) {
274 MemMgr.endFunctionBody(CurByte);
275 ConstantPoolAddresses.clear();
276 NumBytes += CurByte-CurBlock;
278 if (!Relocations.empty()) {
279 // Resolve the relocations to concrete pointers.
280 for (unsigned i = 0, e = Relocations.size(); i != e; ++i) {
281 MachineRelocation &MR = Relocations[i];
284 ResultPtr = TheJIT->getPointerToNamedFunction(MR.getString());
286 ResultPtr = getPointerToGlobal(MR.getGlobalValue());
287 MR.setResultPointer(ResultPtr);
290 TheJIT->getJITInfo().relocate(CurBlock, &Relocations[0],
294 DEBUG(std::cerr << "Finished CodeGen of [" << (void*)CurBlock
295 << "] Function: " << F.getFunction()->getName()
296 << ": " << CurByte-CurBlock << " bytes of text, "
297 << Relocations.size() << " relocations\n");
301 void Emitter::emitConstantPool(MachineConstantPool *MCP) {
302 const std::vector<Constant*> &Constants = MCP->getConstants();
303 if (Constants.empty()) return;
305 std::vector<unsigned> ConstantOffset;
306 ConstantOffset.reserve(Constants.size());
308 // Calculate how much space we will need for all the constants, and the offset
309 // each one will live in.
310 unsigned TotalSize = 0;
311 for (unsigned i = 0, e = Constants.size(); i != e; ++i) {
312 const Type *Ty = Constants[i]->getType();
313 unsigned Size = TheJIT->getTargetData().getTypeSize(Ty);
314 unsigned Alignment = TheJIT->getTargetData().getTypeAlignment(Ty);
315 // Make sure to take into account the alignment requirements of the type.
316 TotalSize = (TotalSize + Alignment-1) & ~(Alignment-1);
318 // Remember the offset this element lives at.
319 ConstantOffset.push_back(TotalSize);
320 TotalSize += Size; // Reserve space for the constant.
323 // Now that we know how much memory to allocate, do so.
324 char *Pool = new char[TotalSize];
326 // Actually output all of the constants, and remember their addresses.
327 for (unsigned i = 0, e = Constants.size(); i != e; ++i) {
328 void *Addr = Pool + ConstantOffset[i];
329 TheJIT->InitializeMemory(Constants[i], Addr);
330 ConstantPoolAddresses.push_back(Addr);
334 void Emitter::startFunctionStub(unsigned StubSize) {
335 SavedCurBlock = CurBlock; SavedCurByte = CurByte;
336 CurByte = CurBlock = MemMgr.allocateStub(StubSize);
339 void *Emitter::finishFunctionStub(const Function *F) {
340 NumBytes += CurByte-CurBlock;
341 DEBUG(std::cerr << "Finished CodeGen of [0x" << (void*)CurBlock
342 << "] Function stub for: " << (F ? F->getName() : "")
343 << ": " << CurByte-CurBlock << " bytes of text\n");
344 std::swap(CurBlock, SavedCurBlock);
345 CurByte = SavedCurByte;
346 return SavedCurBlock;
349 void Emitter::emitByte(unsigned char B) {
350 *CurByte++ = B; // Write the byte to memory
353 void Emitter::emitWord(unsigned W) {
354 // This won't work if the endianness of the host and target don't agree! (For
355 // a JIT this can't happen though. :)
356 *(unsigned*)CurByte = W;
357 CurByte += sizeof(unsigned);
360 void Emitter::emitWordAt(unsigned W, unsigned *Ptr) {
364 uint64_t Emitter::getGlobalValueAddress(GlobalValue *V) {
365 // Try looking up the function to see if it is already compiled, if not return
367 if (Function *F = dyn_cast<Function>(V)) {
368 void *Addr = TheJIT->getPointerToGlobalIfAvailable(F);
369 if (Addr == 0 && F->hasExternalLinkage()) {
370 // Do not output stubs for external functions.
371 Addr = TheJIT->getPointerToFunction(F);
373 return (intptr_t)Addr;
375 return (intptr_t)TheJIT->getOrEmitGlobalVariable(cast<GlobalVariable>(V));
378 uint64_t Emitter::getGlobalValueAddress(const char *Name) {
379 return (intptr_t)TheJIT->getPointerToNamedFunction(Name);
382 // getConstantPoolEntryAddress - Return the address of the 'ConstantNum' entry
383 // in the constant pool that was last emitted with the 'emitConstantPool'
386 uint64_t Emitter::getConstantPoolEntryAddress(unsigned ConstantNum) {
387 assert(ConstantNum < ConstantPoolAddresses.size() &&
388 "Invalid ConstantPoolIndex!");
389 return (intptr_t)ConstantPoolAddresses[ConstantNum];
392 // getCurrentPCValue - This returns the address that the next emitted byte
393 // will be output to.
395 uint64_t Emitter::getCurrentPCValue() {
396 return (intptr_t)CurByte;
399 uint64_t Emitter::getCurrentPCOffset() {
400 return (intptr_t)CurByte-(intptr_t)CurBlock;
403 uint64_t Emitter::forceCompilationOf(Function *F) {
404 return (intptr_t)TheJIT->getPointerToFunction(F);
407 // getPointerToNamedFunction - This function is used as a global wrapper to
408 // JIT::getPointerToNamedFunction for the purpose of resolving symbols when
409 // bugpoint is debugging the JIT. In that scenario, we are loading an .so and
410 // need to resolve function(s) that are being mis-codegenerated, so we need to
411 // resolve their addresses at runtime, and this is the way to do it.
413 void *getPointerToNamedFunction(const char *Name) {
414 Module &M = TheJIT->getModule();
415 if (Function *F = M.getNamedFunction(Name))
416 return TheJIT->getPointerToFunction(F);
417 return TheJIT->getPointerToNamedFunction(Name);