//
//===----------------------------------------------------------------------===//
+#define DEBUG_TYPE "jit"
+#ifndef _POSIX_MAPPED_FILES
+#define _POSIX_MAPPED_FILES
+#endif
#include "VM.h"
#include "llvm/CodeGen/MachineCodeEmitter.h"
#include "llvm/CodeGen/MachineFunction.h"
#include "llvm/CodeGen/MachineConstantPool.h"
#include "llvm/Target/TargetData.h"
-#include "llvm/Function.h"
+#include "llvm/Module.h"
+#include "Support/Debug.h"
#include "Support/Statistic.h"
+#include "Config/unistd.h"
+#include "Config/sys/mman.h"
namespace {
- Statistic<> NumBytes("jello", "Number of bytes of machine code compiled");
+ Statistic<> NumBytes("jit", "Number of bytes of machine code compiled");
+ VM *TheVM = 0;
+
+ /// JITMemoryManager - Manage memory for the JIT code generation in a logical,
+ /// sane way. This splits a large block of MAP_NORESERVE'd memory into two
+ /// sections, one for function stubs, one for the functions themselves. We
+ /// have to do this because we may need to emit a function stub while in the
+ /// middle of emitting a function, and we don't know how large the function we
+ /// are emitting is. This never bothers to release the memory, because when
+ /// we are ready to destroy the JIT, the program exits.
+ class JITMemoryManager {
+ unsigned char *MemBase; // Base of block of memory, start of stub mem
+ unsigned char *FunctionBase; // Start of the function body area
+ unsigned char *CurStubPtr, *CurFunctionPtr;
+ public:
+ JITMemoryManager();
+
+ inline unsigned char *allocateStub(unsigned StubSize);
+ inline unsigned char *startFunctionBody();
+ inline void endFunctionBody(unsigned char *FunctionEnd);
+ };
+}
+
+// getMemory - Return a pointer to the specified number of bytes, which is
+// mapped as executable readable and writable.
+static void *getMemory(unsigned NumBytes) {
+ if (NumBytes == 0) return 0;
+ static const long pageSize = sysconf(_SC_PAGESIZE);
+ unsigned NumPages = (NumBytes+pageSize-1)/pageSize;
+
+#if defined(i386) || defined(__i386__) || defined(__x86__)
+ /* Linux and *BSD tend to have these flags named differently. */
+#if defined(MAP_ANON) && !defined(MAP_ANONYMOUS)
+# define MAP_ANONYMOUS MAP_ANON
+#endif /* defined(MAP_ANON) && !defined(MAP_ANONYMOUS) */
+#elif defined(sparc) || defined(__sparc__) || defined(__sparcv9)
+/* nothing */
+#else
+ std::cerr << "This architecture is not supported by the JIT!\n";
+ abort();
+#endif
+
+#if defined(__linux__)
+#define fd 0
+#else
+#define fd -1
+#endif
+
+ unsigned mmapFlags = MAP_PRIVATE|MAP_ANONYMOUS;
+#ifdef MAP_NORESERVE
+ mmapFlags |= MAP_NORESERVE;
+#endif
+
+ void *pa = mmap(0, pageSize*NumPages, PROT_READ|PROT_WRITE|PROT_EXEC,
+ mmapFlags, fd, 0);
+ if (pa == MAP_FAILED) {
+ perror("mmap");
+ abort();
+ }
+ return pa;
+}
+
+JITMemoryManager::JITMemoryManager() {
+ // Allocate a 16M block of memory...
+ MemBase = (unsigned char*)getMemory(16 << 20);
+ FunctionBase = MemBase + 512*1024; // Use 512k for stubs
+
+ // Allocate stubs backwards from the function base, allocate functions forward
+ // from the function base.
+ CurStubPtr = CurFunctionPtr = FunctionBase;
+}
+unsigned char *JITMemoryManager::allocateStub(unsigned StubSize) {
+ CurStubPtr -= StubSize;
+ if (CurStubPtr < MemBase) {
+ std::cerr << "JIT ran out of memory for function stubs!\n";
+ abort();
+ }
+ return CurStubPtr;
+}
+
+unsigned char *JITMemoryManager::startFunctionBody() {
+ // Round up to an even multiple of 4 bytes, this should eventually be target
+ // specific.
+ return (unsigned char*)(((intptr_t)CurFunctionPtr + 3) & ~3);
+}
+
+void JITMemoryManager::endFunctionBody(unsigned char *FunctionEnd) {
+ assert(FunctionEnd > CurFunctionPtr);
+ CurFunctionPtr = FunctionEnd;
+}
+
+
+
+namespace {
+ /// Emitter - The JIT implementation of the MachineCodeEmitter, which is used
+ /// to output functions to memory for execution.
class Emitter : public MachineCodeEmitter {
- VM &TheVM;
+ JITMemoryManager MemMgr;
- unsigned char *CurBlock;
- unsigned char *CurByte;
-
- std::vector<std::pair<BasicBlock*, unsigned *> > BBRefs;
- std::map<BasicBlock*, unsigned> BBLocations;
+ // CurBlock - The start of the current block of memory. CurByte - The
+ // current byte being emitted to.
+ unsigned char *CurBlock, *CurByte;
+
+ // When outputting a function stub in the context of some other function, we
+ // save CurBlock and CurByte here.
+ unsigned char *SavedCurBlock, *SavedCurByte;
+
+ // ConstantPoolAddresses - Contains the location for each entry in the
+ // constant pool.
std::vector<void*> ConstantPoolAddresses;
public:
- Emitter(VM &vm) : TheVM(vm) {}
+ Emitter(VM &vm) { TheVM = &vm; }
virtual void startFunction(MachineFunction &F);
virtual void finishFunction(MachineFunction &F);
virtual void emitConstantPool(MachineConstantPool *MCP);
- virtual void startBasicBlock(MachineBasicBlock &BB);
+ virtual void startFunctionStub(const Function &F, unsigned StubSize);
+ virtual void* finishFunctionStub(const Function &F);
virtual void emitByte(unsigned char B);
- virtual void emitPCRelativeDisp(Value *V);
- virtual void emitGlobalAddress(GlobalValue *V, bool isPCRelative);
- virtual void emitGlobalAddress(const std::string &Name, bool isPCRelative);
- virtual void emitFunctionConstantValueAddress(unsigned ConstantNum,
- int Offset);
- private:
- void emitAddress(void *Addr, bool isPCRelative);
+ virtual void emitWord(unsigned W);
+
+ virtual uint64_t getGlobalValueAddress(GlobalValue *V);
+ virtual uint64_t getGlobalValueAddress(const std::string &Name);
+ virtual uint64_t getConstantPoolEntryAddress(unsigned Entry);
+ virtual uint64_t getCurrentPCValue();
+
+ // forceCompilationOf - Force the compilation of the specified function, and
+ // return its address, because we REALLY need the address now.
+ //
+ // FIXME: This is JIT specific!
+ //
+ virtual uint64_t forceCompilationOf(Function *F);
};
}
return new Emitter(V);
}
-
-#define _POSIX_MAPPED_FILES
-#include <unistd.h>
-#include <sys/mman.h>
-
-static void *getMemory() {
- return mmap(0, 4096*8, PROT_READ|PROT_WRITE|PROT_EXEC,
- MAP_PRIVATE|MAP_ANONYMOUS, 0, 0);
-}
-
-
void Emitter::startFunction(MachineFunction &F) {
- CurBlock = (unsigned char *)getMemory();
- CurByte = CurBlock; // Start writing at the beginning of the fn.
- TheVM.addGlobalMapping(F.getFunction(), CurBlock);
+ CurByte = CurBlock = MemMgr.startFunctionBody();
+ TheVM->addGlobalMapping(F.getFunction(), CurBlock);
}
void Emitter::finishFunction(MachineFunction &F) {
+ MemMgr.endFunctionBody(CurByte);
ConstantPoolAddresses.clear();
- for (unsigned i = 0, e = BBRefs.size(); i != e; ++i) {
- unsigned Location = BBLocations[BBRefs[i].first];
- unsigned *Ref = BBRefs[i].second;
- *Ref = Location-(unsigned)(intptr_t)Ref-4;
- }
- BBRefs.clear();
- BBLocations.clear();
-
NumBytes += CurByte-CurBlock;
- DEBUG(std::cerr << "Finished CodeGen of [0x" << std::hex
- << (unsigned)(intptr_t)CurBlock
- << std::dec << "] Function: " << F.getFunction()->getName()
+ DEBUG(std::cerr << "Finished CodeGen of [" << (void*)CurBlock
+ << "] Function: " << F.getFunction()->getName()
<< ": " << CurByte-CurBlock << " bytes of text\n");
}
// For now we just allocate some memory on the heap, this can be
// dramatically improved.
const Type *Ty = ((Value*)Constants[i])->getType();
- void *Addr = malloc(TheVM.getTargetData().getTypeSize(Ty));
- TheVM.InitializeMemory(Constants[i], Addr);
+ void *Addr = malloc(TheVM->getTargetData().getTypeSize(Ty));
+ TheVM->InitializeMemory(Constants[i], Addr);
ConstantPoolAddresses.push_back(Addr);
}
}
-
-void Emitter::startBasicBlock(MachineBasicBlock &BB) {
- BBLocations[BB.getBasicBlock()] = (unsigned)(intptr_t)CurByte;
+void Emitter::startFunctionStub(const Function &F, unsigned StubSize) {
+ SavedCurBlock = CurBlock; SavedCurByte = CurByte;
+ CurByte = CurBlock = MemMgr.allocateStub(StubSize);
}
+void *Emitter::finishFunctionStub(const Function &F) {
+ NumBytes += CurByte-CurBlock;
+ DEBUG(std::cerr << "Finished CodeGen of [0x" << std::hex
+ << (unsigned)(intptr_t)CurBlock
+ << std::dec << "] Function stub for: " << F.getName()
+ << ": " << CurByte-CurBlock << " bytes of text\n");
+ std::swap(CurBlock, SavedCurBlock);
+ CurByte = SavedCurByte;
+ return SavedCurBlock;
+}
void Emitter::emitByte(unsigned char B) {
*CurByte++ = B; // Write the byte to memory
}
+void Emitter::emitWord(unsigned W) {
+ // This won't work if the endianness of the host and target don't agree! (For
+ // a JIT this can't happen though. :)
+ *(unsigned*)CurByte = W;
+ CurByte += sizeof(unsigned);
+}
-// emitPCRelativeDisp - For functions, just output a displacement that will
-// cause a reference to the zero page, which will cause a seg-fault, causing
-// things to get resolved on demand. Keep track of these markers.
-//
-// For basic block references, keep track of where the references are so they
-// may be patched up when the basic block is defined.
-//
-void Emitter::emitPCRelativeDisp(Value *V) {
- BasicBlock *BB = cast<BasicBlock>(V); // Keep track of reference...
- BBRefs.push_back(std::make_pair(BB, (unsigned*)CurByte));
- CurByte += 4;
+uint64_t Emitter::getGlobalValueAddress(GlobalValue *V) {
+ // Try looking up the function to see if it is already compiled, if not return
+ // 0.
+ return (intptr_t)TheVM->getPointerToGlobalIfAvailable(V);
+}
+uint64_t Emitter::getGlobalValueAddress(const std::string &Name) {
+ return (intptr_t)TheVM->getPointerToNamedFunction(Name);
}
-// emitAddress - Emit an address in either direct or PCRelative form...
+// getConstantPoolEntryAddress - Return the address of the 'ConstantNum' entry
+// in the constant pool that was last emitted with the 'emitConstantPool'
+// method.
//
-void Emitter::emitAddress(void *Addr, bool isPCRelative) {
- if (isPCRelative) {
- *(intptr_t*)CurByte = (intptr_t)Addr - (intptr_t)CurByte-4;
- } else {
- *(void**)CurByte = Addr;
- }
- CurByte += 4;
+uint64_t Emitter::getConstantPoolEntryAddress(unsigned ConstantNum) {
+ assert(ConstantNum < ConstantPoolAddresses.size() &&
+ "Invalid ConstantPoolIndex!");
+ return (intptr_t)ConstantPoolAddresses[ConstantNum];
}
-void Emitter::emitGlobalAddress(GlobalValue *V, bool isPCRelative) {
- if (isPCRelative) { // must be a call, this is a major hack!
- // FIXME: Try looking up the function to see if it is already compiled!
- TheVM.addFunctionRef(CurByte, cast<Function>(V));
-
- // Delayed resolution...
- emitAddress((void*)VM::CompilationCallback, isPCRelative);
- } else {
- emitAddress(TheVM.getPointerToGlobal(V), isPCRelative);
- }
+// getCurrentPCValue - This returns the address that the next emitted byte
+// will be output to.
+//
+uint64_t Emitter::getCurrentPCValue() {
+ return (intptr_t)CurByte;
}
-void Emitter::emitGlobalAddress(const std::string &Name, bool isPCRelative) {
- emitAddress(TheVM.getPointerToNamedFunction(Name), isPCRelative);
+uint64_t Emitter::forceCompilationOf(Function *F) {
+ return (intptr_t)TheVM->getPointerToFunction(F);
}
-void Emitter::emitFunctionConstantValueAddress(unsigned ConstantNum,
- int Offset) {
- assert(ConstantNum < ConstantPoolAddresses.size() &&
- "Invalid ConstantPoolIndex!");
- *(void**)CurByte = (char*)ConstantPoolAddresses[ConstantNum]+Offset;
- CurByte += 4;
+// getPointerToNamedFunction - This function is used as a global wrapper to
+// VM::getPointerToNamedFunction for the purpose of resolving symbols when
+// bugpoint is debugging the JIT. In that scenario, we are loading an .so and
+// need to resolve function(s) that are being mis-codegenerated, so we need to
+// resolve their addresses at runtime, and this is the way to do it.
+extern "C" {
+ void *getPointerToNamedFunction(const char *Name) {
+ Module &M = TheVM->getModule();
+ if (Function *F = M.getNamedFunction(Name))
+ return TheVM->getPointerToFunction(F);
+ return TheVM->getPointerToNamedFunction(Name);
+ }
}