1 //===-- Emitter.cpp - Write machine code to executable memory -------------===//
3 // This file defines a MachineCodeEmitter object that is used by Jello to write
4 // machine code to memory and remember where relocatable values lie.
6 //===----------------------------------------------------------------------===//
8 #define DEBUG_TYPE "jit"
9 #ifndef _POSIX_MAPPED_FILES
10 #define _POSIX_MAPPED_FILES
13 #include "llvm/CodeGen/MachineCodeEmitter.h"
14 #include "llvm/CodeGen/MachineFunction.h"
15 #include "llvm/CodeGen/MachineConstantPool.h"
16 #include "llvm/Target/TargetData.h"
17 #include "llvm/Module.h"
18 #include "Support/Debug.h"
19 #include "Support/Statistic.h"
20 #include "Config/unistd.h"
21 #include "Config/sys/mman.h"
24 Statistic<> NumBytes("jello", "Number of bytes of machine code compiled");
27 /// JITMemoryManager - Manage memory for the JIT code generation in a logical,
28 /// sane way. This splits a large block of MAP_NORESERVE'd memory into two
29 /// sections, one for function stubs, one for the functions themselves. We
30 /// have to do this because we may need to emit a function stub while in the
31 /// middle of emitting a function, and we don't know how large the function we
32 /// are emitting is. This never bothers to release the memory, because when
33 /// we are ready to destroy the JIT, the program exits.
34 class JITMemoryManager {
35 unsigned char *MemBase; // Base of block of memory, start of stub mem
36 unsigned char *FunctionBase; // Start of the function body area
37 unsigned char *CurStubPtr, *CurFunctionPtr;
41 inline unsigned char *allocateStub(unsigned StubSize);
42 inline unsigned char *startFunctionBody();
43 inline void endFunctionBody(unsigned char *FunctionEnd);
47 // getMemory - Return a pointer to the specified number of bytes, which is
48 // mapped as executable readable and writable.
49 static void *getMemory(unsigned NumBytes) {
50 if (NumBytes == 0) return 0;
51 static const long pageSize = sysconf(_SC_PAGESIZE);
52 unsigned NumPages = (NumBytes+pageSize-1)/pageSize;
54 #if defined(i386) || defined(__i386__) || defined(__x86__)
55 /* Linux and *BSD tend to have these flags named differently. */
56 #if defined(MAP_ANON) && !defined(MAP_ANONYMOUS)
57 # define MAP_ANONYMOUS MAP_ANON
58 #endif /* defined(MAP_ANON) && !defined(MAP_ANONYMOUS) */
60 #elif defined(sparc) || defined(__sparc__) || defined(__sparcv9)
63 std::cerr << "This architecture is not supported by the JIT!\n";
67 unsigned mmapFlags = MAP_PRIVATE|MAP_ANONYMOUS;
69 mmapFlags |= MAP_NORESERVE;
72 void *pa = mmap(0, pageSize*NumPages, PROT_READ|PROT_WRITE|PROT_EXEC,
74 if (pa == MAP_FAILED) {
81 JITMemoryManager::JITMemoryManager() {
82 // Allocate a 16M block of memory...
83 MemBase = (unsigned char*)getMemory(16 << 20);
84 FunctionBase = MemBase + 512*1024; // Use 512k for stubs
86 // Allocate stubs backwards from the function base, allocate functions forward
87 // from the function base.
88 CurStubPtr = CurFunctionPtr = FunctionBase;
91 unsigned char *JITMemoryManager::allocateStub(unsigned StubSize) {
92 CurStubPtr -= StubSize;
93 if (CurStubPtr < MemBase) {
94 std::cerr << "JIT ran out of memory for function stubs!\n";
100 unsigned char *JITMemoryManager::startFunctionBody() {
101 // Round up to an even multiple of 4 bytes, this should eventually be target
103 return (unsigned char*)(((intptr_t)CurFunctionPtr + 3) & ~3);
106 void JITMemoryManager::endFunctionBody(unsigned char *FunctionEnd) {
107 assert(FunctionEnd > CurFunctionPtr);
108 CurFunctionPtr = FunctionEnd;
114 /// Emitter - The JIT implementation of the MachineCodeEmiter, which is used
115 /// to output functions to memory for execution.
116 class Emitter : public MachineCodeEmitter {
117 JITMemoryManager MemMgr;
119 // CurBlock - The start of the current block of memory. CurByte - The
120 // current byte being emitted to.
121 unsigned char *CurBlock, *CurByte;
123 // When outputting a function stub in the context of some other function, we
124 // save CurBlock and CurByte here.
125 unsigned char *SavedCurBlock, *SavedCurByte;
127 // ConstantPoolAddresses - Contains the location for each entry in the
129 std::vector<void*> ConstantPoolAddresses;
131 Emitter(VM &vm) { TheVM = &vm; }
133 virtual void startFunction(MachineFunction &F);
134 virtual void finishFunction(MachineFunction &F);
135 virtual void emitConstantPool(MachineConstantPool *MCP);
136 virtual void startFunctionStub(const Function &F, unsigned StubSize);
137 virtual void* finishFunctionStub(const Function &F);
138 virtual void emitByte(unsigned char B);
139 virtual void emitWord(unsigned W);
141 virtual uint64_t getGlobalValueAddress(GlobalValue *V);
142 virtual uint64_t getGlobalValueAddress(const std::string &Name);
143 virtual uint64_t getConstantPoolEntryAddress(unsigned Entry);
144 virtual uint64_t getCurrentPCValue();
146 // forceCompilationOf - Force the compilation of the specified function, and
147 // return its address, because we REALLY need the address now.
149 // FIXME: This is JIT specific!
151 virtual uint64_t forceCompilationOf(Function *F);
155 MachineCodeEmitter *VM::createEmitter(VM &V) {
156 return new Emitter(V);
159 void Emitter::startFunction(MachineFunction &F) {
160 CurByte = CurBlock = MemMgr.startFunctionBody();
161 TheVM->addGlobalMapping(F.getFunction(), CurBlock);
164 void Emitter::finishFunction(MachineFunction &F) {
165 MemMgr.endFunctionBody(CurByte);
166 ConstantPoolAddresses.clear();
167 NumBytes += CurByte-CurBlock;
169 DEBUG(std::cerr << "Finished CodeGen of [" << (void*)CurBlock
170 << "] Function: " << F.getFunction()->getName()
171 << ": " << CurByte-CurBlock << " bytes of text\n");
174 void Emitter::emitConstantPool(MachineConstantPool *MCP) {
175 const std::vector<Constant*> &Constants = MCP->getConstants();
176 for (unsigned i = 0, e = Constants.size(); i != e; ++i) {
177 // For now we just allocate some memory on the heap, this can be
178 // dramatically improved.
179 const Type *Ty = ((Value*)Constants[i])->getType();
180 void *Addr = malloc(TheVM->getTargetData().getTypeSize(Ty));
181 TheVM->InitializeMemory(Constants[i], Addr);
182 ConstantPoolAddresses.push_back(Addr);
186 void Emitter::startFunctionStub(const Function &F, unsigned StubSize) {
187 SavedCurBlock = CurBlock; SavedCurByte = CurByte;
188 CurByte = CurBlock = MemMgr.allocateStub(StubSize);
191 void *Emitter::finishFunctionStub(const Function &F) {
192 NumBytes += CurByte-CurBlock;
193 DEBUG(std::cerr << "Finished CodeGen of [0x" << std::hex
194 << (unsigned)(intptr_t)CurBlock
195 << std::dec << "] Function stub for: " << F.getName()
196 << ": " << CurByte-CurBlock << " bytes of text\n");
197 std::swap(CurBlock, SavedCurBlock);
198 CurByte = SavedCurByte;
199 return SavedCurBlock;
202 void Emitter::emitByte(unsigned char B) {
203 *CurByte++ = B; // Write the byte to memory
206 void Emitter::emitWord(unsigned W) {
207 // This won't work if the endianness of the host and target don't agree! (For
208 // a JIT this can't happen though. :)
209 *(unsigned*)CurByte = W;
210 CurByte += sizeof(unsigned);
214 uint64_t Emitter::getGlobalValueAddress(GlobalValue *V) {
215 // Try looking up the function to see if it is already compiled, if not return
217 return (intptr_t)TheVM->getPointerToGlobalIfAvailable(V);
219 uint64_t Emitter::getGlobalValueAddress(const std::string &Name) {
220 return (intptr_t)TheVM->getPointerToNamedFunction(Name);
223 // getConstantPoolEntryAddress - Return the address of the 'ConstantNum' entry
224 // in the constant pool that was last emitted with the 'emitConstantPool'
227 uint64_t Emitter::getConstantPoolEntryAddress(unsigned ConstantNum) {
228 assert(ConstantNum < ConstantPoolAddresses.size() &&
229 "Invalid ConstantPoolIndex!");
230 return (intptr_t)ConstantPoolAddresses[ConstantNum];
233 // getCurrentPCValue - This returns the address that the next emitted byte
234 // will be output to.
236 uint64_t Emitter::getCurrentPCValue() {
237 return (intptr_t)CurByte;
240 uint64_t Emitter::forceCompilationOf(Function *F) {
241 return (intptr_t)TheVM->getPointerToFunction(F);
244 // getPointerToNamedFunction - This function is used as a global wrapper to
245 // VM::getPointerToNamedFunction for the purpose of resolving symbols when
246 // bugpoint is debugging the JIT. In that scenario, we are loading an .so and
247 // need to resolve function(s) that are being mis-codegenerated, so we need to
248 // resolve their addresses at runtime, and this is the way to do it.
250 void *getPointerToNamedFunction(const char *Name) {
251 Module &M = TheVM->getModule();
252 if (Function *F = M.getNamedFunction(Name))
253 return TheVM->getPointerToFunction(F);
254 return TheVM->getPointerToNamedFunction(Name);