1 //===-- JITEmitter.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/Type.h"
20 #include "llvm/CodeGen/MachineCodeEmitter.h"
21 #include "llvm/CodeGen/MachineFunction.h"
22 #include "llvm/CodeGen/MachineConstantPool.h"
23 #include "llvm/CodeGen/MachineJumpTableInfo.h"
24 #include "llvm/CodeGen/MachineRelocation.h"
25 #include "llvm/ExecutionEngine/GenericValue.h"
26 #include "llvm/Target/TargetData.h"
27 #include "llvm/Target/TargetJITInfo.h"
28 #include "llvm/Target/TargetMachine.h"
29 #include "llvm/Support/Debug.h"
30 #include "llvm/Support/MutexGuard.h"
31 #include "llvm/ADT/Statistic.h"
32 #include "llvm/System/Memory.h"
36 STATISTIC(NumBytes, "Number of bytes of machine code compiled");
37 STATISTIC(NumRelos, "Number of relocations applied");
38 static JIT *TheJIT = 0;
40 //===----------------------------------------------------------------------===//
41 // JITMemoryManager code.
44 /// MemoryRangeHeader - For a range of memory, this is the header that we put
45 /// on the block of memory. It is carefully crafted to be one word of memory.
46 /// Allocated blocks have just this header, free'd blocks have FreeRangeHeader
47 /// which starts with this.
48 struct FreeRangeHeader;
49 struct MemoryRangeHeader {
50 /// ThisAllocated - This is true if this block is currently allocated. If
51 /// not, this can be converted to a FreeRangeHeader.
52 intptr_t ThisAllocated : 1;
54 /// PrevAllocated - Keep track of whether the block immediately before us is
55 /// allocated. If not, the word immediately before this header is the size
56 /// of the previous block.
57 intptr_t PrevAllocated : 1;
59 /// BlockSize - This is the size in bytes of this memory block,
60 /// including this header.
61 uintptr_t BlockSize : (sizeof(intptr_t)*8 - 2);
64 /// getBlockAfter - Return the memory block immediately after this one.
66 MemoryRangeHeader &getBlockAfter() const {
67 return *(MemoryRangeHeader*)((char*)this+BlockSize);
70 /// getFreeBlockBefore - If the block before this one is free, return it,
71 /// otherwise return null.
72 FreeRangeHeader *getFreeBlockBefore() const {
73 if (PrevAllocated) return 0;
74 intptr_t PrevSize = ((intptr_t *)this)[-1];
75 return (FreeRangeHeader*)((char*)this-PrevSize);
78 /// FreeBlock - Turn an allocated block into a free block, adjusting
79 /// bits in the object headers, and adding an end of region memory block.
80 FreeRangeHeader *FreeBlock(FreeRangeHeader *FreeList);
82 /// TrimAllocationToSize - If this allocated block is significantly larger
83 /// than NewSize, split it into two pieces (where the former is NewSize
84 /// bytes, including the header), and add the new block to the free list.
85 FreeRangeHeader *TrimAllocationToSize(FreeRangeHeader *FreeList,
89 /// FreeRangeHeader - For a memory block that isn't already allocated, this
90 /// keeps track of the current block and has a pointer to the next free block.
91 /// Free blocks are kept on a circularly linked list.
92 struct FreeRangeHeader : public MemoryRangeHeader {
93 FreeRangeHeader *Prev;
94 FreeRangeHeader *Next;
96 /// getMinBlockSize - Get the minimum size for a memory block. Blocks
97 /// smaller than this size cannot be created.
98 static unsigned getMinBlockSize() {
99 return sizeof(FreeRangeHeader)+sizeof(intptr_t);
102 /// SetEndOfBlockSizeMarker - The word at the end of every free block is
103 /// known to be the size of the free block. Set it for this block.
104 void SetEndOfBlockSizeMarker() {
105 void *EndOfBlock = (char*)this + BlockSize;
106 ((intptr_t *)EndOfBlock)[-1] = BlockSize;
109 FreeRangeHeader *RemoveFromFreeList() {
110 assert(Next->Prev == this && Prev->Next == this && "Freelist broken!");
112 return Prev->Next = Next;
115 void AddToFreeList(FreeRangeHeader *FreeList) {
117 Prev = FreeList->Prev;
122 /// GrowBlock - The block after this block just got deallocated. Merge it
123 /// into the current block.
124 void GrowBlock(uintptr_t NewSize);
126 /// AllocateBlock - Mark this entire block allocated, updating freelists
127 /// etc. This returns a pointer to the circular free-list.
128 FreeRangeHeader *AllocateBlock();
133 /// AllocateBlock - Mark this entire block allocated, updating freelists
134 /// etc. This returns a pointer to the circular free-list.
135 FreeRangeHeader *FreeRangeHeader::AllocateBlock() {
136 assert(!ThisAllocated && !getBlockAfter().PrevAllocated &&
137 "Cannot allocate an allocated block!");
138 // Mark this block allocated.
140 getBlockAfter().PrevAllocated = 1;
142 // Remove it from the free list.
143 return RemoveFromFreeList();
146 /// FreeBlock - Turn an allocated block into a free block, adjusting
147 /// bits in the object headers, and adding an end of region memory block.
148 /// If possible, coallesce this block with neighboring blocks. Return the
149 /// FreeRangeHeader to allocate from.
150 FreeRangeHeader *MemoryRangeHeader::FreeBlock(FreeRangeHeader *FreeList) {
151 MemoryRangeHeader *FollowingBlock = &getBlockAfter();
152 assert(ThisAllocated && "This block is already allocated!");
153 assert(FollowingBlock->PrevAllocated && "Flags out of sync!");
155 FreeRangeHeader *FreeListToReturn = FreeList;
157 // If the block after this one is free, merge it into this block.
158 if (!FollowingBlock->ThisAllocated) {
159 FreeRangeHeader &FollowingFreeBlock = *(FreeRangeHeader *)FollowingBlock;
160 // "FreeList" always needs to be a valid free block. If we're about to
161 // coallesce with it, update our notion of what the free list is.
162 if (&FollowingFreeBlock == FreeList) {
163 FreeList = FollowingFreeBlock.Next;
164 FreeListToReturn = 0;
165 assert(&FollowingFreeBlock != FreeList && "No tombstone block?");
167 FollowingFreeBlock.RemoveFromFreeList();
169 // Include the following block into this one.
170 BlockSize += FollowingFreeBlock.BlockSize;
171 FollowingBlock = &FollowingFreeBlock.getBlockAfter();
173 // Tell the block after the block we are coallescing that this block is
175 FollowingBlock->PrevAllocated = 1;
178 assert(FollowingBlock->ThisAllocated && "Missed coallescing?");
180 if (FreeRangeHeader *PrevFreeBlock = getFreeBlockBefore()) {
181 PrevFreeBlock->GrowBlock(PrevFreeBlock->BlockSize + BlockSize);
182 return FreeListToReturn ? FreeListToReturn : PrevFreeBlock;
185 // Otherwise, mark this block free.
186 FreeRangeHeader &FreeBlock = *(FreeRangeHeader*)this;
187 FollowingBlock->PrevAllocated = 0;
188 FreeBlock.ThisAllocated = 0;
190 // Link this into the linked list of free blocks.
191 FreeBlock.AddToFreeList(FreeList);
193 // Add a marker at the end of the block, indicating the size of this free
195 FreeBlock.SetEndOfBlockSizeMarker();
196 return FreeListToReturn ? FreeListToReturn : &FreeBlock;
199 /// GrowBlock - The block after this block just got deallocated. Merge it
200 /// into the current block.
201 void FreeRangeHeader::GrowBlock(uintptr_t NewSize) {
202 assert(NewSize > BlockSize && "Not growing block?");
204 SetEndOfBlockSizeMarker();
205 getBlockAfter().PrevAllocated = 0;
208 /// TrimAllocationToSize - If this allocated block is significantly larger
209 /// than NewSize, split it into two pieces (where the former is NewSize
210 /// bytes, including the header), and add the new block to the free list.
211 FreeRangeHeader *MemoryRangeHeader::
212 TrimAllocationToSize(FreeRangeHeader *FreeList, uint64_t NewSize) {
213 assert(ThisAllocated && getBlockAfter().PrevAllocated &&
214 "Cannot deallocate part of an allocated block!");
216 // Round up size for alignment of header.
217 unsigned HeaderAlign = __alignof(FreeRangeHeader);
218 NewSize = (NewSize+ (HeaderAlign-1)) & ~(HeaderAlign-1);
220 // Size is now the size of the block we will remove from the start of the
222 assert(NewSize <= BlockSize &&
223 "Allocating more space from this block than exists!");
225 // If splitting this block will cause the remainder to be too small, do not
227 if (BlockSize <= NewSize+FreeRangeHeader::getMinBlockSize())
230 // Otherwise, we splice the required number of bytes out of this block, form
231 // a new block immediately after it, then mark this block allocated.
232 MemoryRangeHeader &FormerNextBlock = getBlockAfter();
234 // Change the size of this block.
237 // Get the new block we just sliced out and turn it into a free block.
238 FreeRangeHeader &NewNextBlock = (FreeRangeHeader &)getBlockAfter();
239 NewNextBlock.BlockSize = (char*)&FormerNextBlock - (char*)&NewNextBlock;
240 NewNextBlock.ThisAllocated = 0;
241 NewNextBlock.PrevAllocated = 1;
242 NewNextBlock.SetEndOfBlockSizeMarker();
243 FormerNextBlock.PrevAllocated = 0;
244 NewNextBlock.AddToFreeList(FreeList);
245 return &NewNextBlock;
250 /// JITMemoryManager - Manage memory for the JIT code generation in a logical,
251 /// sane way. This splits a large block of MAP_NORESERVE'd memory into two
252 /// sections, one for function stubs, one for the functions themselves. We
253 /// have to do this because we may need to emit a function stub while in the
254 /// middle of emitting a function, and we don't know how large the function we
255 /// are emitting is. This never bothers to release the memory, because when
256 /// we are ready to destroy the JIT, the program exits.
257 class JITMemoryManager {
258 std::vector<sys::MemoryBlock> Blocks; // Memory blocks allocated by the JIT
259 FreeRangeHeader *FreeMemoryList; // Circular list of free blocks.
261 // When emitting code into a memory block, this is the block.
262 MemoryRangeHeader *CurBlock;
264 unsigned char *CurStubPtr, *StubBase;
265 unsigned char *GOTBase; // Target Specific reserved memory
267 // Centralize memory block allocation.
268 sys::MemoryBlock getNewMemoryBlock(unsigned size);
270 std::map<const Function*, MemoryRangeHeader*> FunctionBlocks;
272 JITMemoryManager(bool useGOT);
275 inline unsigned char *allocateStub(unsigned StubSize, unsigned Alignment);
277 /// startFunctionBody - When a function starts, allocate a block of free
278 /// executable memory, returning a pointer to it and its actual size.
279 unsigned char *startFunctionBody(uintptr_t &ActualSize) {
280 CurBlock = FreeMemoryList;
282 // Allocate the entire memory block.
283 FreeMemoryList = FreeMemoryList->AllocateBlock();
284 ActualSize = CurBlock->BlockSize-sizeof(MemoryRangeHeader);
285 return (unsigned char *)(CurBlock+1);
288 /// endFunctionBody - The function F is now allocated, and takes the memory
289 /// in the range [FunctionStart,FunctionEnd).
290 void endFunctionBody(const Function *F, unsigned char *FunctionStart,
291 unsigned char *FunctionEnd) {
292 assert(FunctionEnd > FunctionStart);
293 assert(FunctionStart == (unsigned char *)(CurBlock+1) &&
294 "Mismatched function start/end!");
296 uintptr_t BlockSize = FunctionEnd - (unsigned char *)CurBlock;
297 FunctionBlocks[F] = CurBlock;
299 // Release the memory at the end of this block that isn't needed.
300 FreeMemoryList =CurBlock->TrimAllocationToSize(FreeMemoryList, BlockSize);
303 unsigned char *getGOTBase() const {
306 bool isManagingGOT() const {
307 return GOTBase != NULL;
310 /// deallocateMemForFunction - Deallocate all memory for the specified
312 void deallocateMemForFunction(const Function *F) {
313 std::map<const Function*, MemoryRangeHeader*>::iterator
314 I = FunctionBlocks.find(F);
315 if (I == FunctionBlocks.end()) return;
317 // Find the block that is allocated for this function.
318 MemoryRangeHeader *MemRange = I->second;
319 assert(MemRange->ThisAllocated && "Block isn't allocated!");
321 // Fill the buffer with garbage!
322 DEBUG(memset(MemRange+1, 0xCD, MemRange->BlockSize-sizeof(*MemRange)));
325 FreeMemoryList = MemRange->FreeBlock(FreeMemoryList);
327 // Finally, remove this entry from FunctionBlocks.
328 FunctionBlocks.erase(I);
333 JITMemoryManager::JITMemoryManager(bool useGOT) {
334 // Allocate a 16M block of memory for functions.
335 sys::MemoryBlock MemBlock = getNewMemoryBlock(16 << 20);
337 unsigned char *MemBase = reinterpret_cast<unsigned char*>(MemBlock.base());
339 // Allocate stubs backwards from the base, allocate functions forward
342 CurStubPtr = MemBase + 512*1024; // Use 512k for stubs, working backwards.
344 // We set up the memory chunk with 4 mem regions, like this:
346 // [ Free #0 ] -> Large space to allocate functions from.
347 // [ Allocated #1 ] -> Tiny space to separate regions.
348 // [ Free #2 ] -> Tiny space so there is always at least 1 free block.
349 // [ Allocated #3 ] -> Tiny space to prevent looking past end of block.
352 // The last three blocks are never deallocated or touched.
354 // Add MemoryRangeHeader to the end of the memory region, indicating that
355 // the space after the block of memory is allocated. This is block #3.
356 MemoryRangeHeader *Mem3 = (MemoryRangeHeader*)(MemBase+MemBlock.size())-1;
357 Mem3->ThisAllocated = 1;
358 Mem3->PrevAllocated = 0;
361 /// Add a tiny free region so that the free list always has one entry.
362 FreeRangeHeader *Mem2 =
363 (FreeRangeHeader *)(((char*)Mem3)-FreeRangeHeader::getMinBlockSize());
364 Mem2->ThisAllocated = 0;
365 Mem2->PrevAllocated = 1;
366 Mem2->BlockSize = FreeRangeHeader::getMinBlockSize();
367 Mem2->SetEndOfBlockSizeMarker();
368 Mem2->Prev = Mem2; // Mem2 *is* the free list for now.
371 /// Add a tiny allocated region so that Mem2 is never coallesced away.
372 MemoryRangeHeader *Mem1 = (MemoryRangeHeader*)Mem2-1;
373 Mem1->ThisAllocated = 1;
374 Mem1->PrevAllocated = 0;
375 Mem1->BlockSize = (char*)Mem2 - (char*)Mem1;
377 // Add a FreeRangeHeader to the start of the function body region, indicating
378 // that the space is free. Mark the previous block allocated so we never look
380 FreeRangeHeader *Mem0 = (FreeRangeHeader*)CurStubPtr;
381 Mem0->ThisAllocated = 0;
382 Mem0->PrevAllocated = 1;
383 Mem0->BlockSize = (char*)Mem1-(char*)Mem0;
384 Mem0->SetEndOfBlockSizeMarker();
385 Mem0->AddToFreeList(Mem2);
387 // Start out with the freelist pointing to Mem0.
388 FreeMemoryList = Mem0;
392 if (useGOT) GOTBase = new unsigned char[sizeof(void*) * 8192];
395 JITMemoryManager::~JITMemoryManager() {
396 for (unsigned i = 0, e = Blocks.size(); i != e; ++i)
397 sys::Memory::ReleaseRWX(Blocks[i]);
403 unsigned char *JITMemoryManager::allocateStub(unsigned StubSize,
404 unsigned Alignment) {
405 CurStubPtr -= StubSize;
406 CurStubPtr = (unsigned char*)(((intptr_t)CurStubPtr) &
407 ~(intptr_t)(Alignment-1));
408 if (CurStubPtr < StubBase) {
409 // FIXME: allocate a new block
410 cerr << "JIT ran out of memory for function stubs!\n";
416 sys::MemoryBlock JITMemoryManager::getNewMemoryBlock(unsigned size) {
417 // Allocate a new block close to the last one.
418 const sys::MemoryBlock *BOld = Blocks.empty() ? 0 : &Blocks.front();
420 sys::MemoryBlock B = sys::Memory::AllocateRWX(size, BOld, &ErrMsg);
422 cerr << "Allocation failed when allocating new memory in the JIT\n";
423 cerr << ErrMsg << "\n";
430 //===----------------------------------------------------------------------===//
431 // JIT lazy compilation code.
434 class JITResolverState {
436 /// FunctionToStubMap - Keep track of the stub created for a particular
437 /// function so that we can reuse them if necessary.
438 std::map<Function*, void*> FunctionToStubMap;
440 /// StubToFunctionMap - Keep track of the function that each stub
442 std::map<void*, Function*> StubToFunctionMap;
445 std::map<Function*, void*>& getFunctionToStubMap(const MutexGuard& locked) {
446 assert(locked.holds(TheJIT->lock));
447 return FunctionToStubMap;
450 std::map<void*, Function*>& getStubToFunctionMap(const MutexGuard& locked) {
451 assert(locked.holds(TheJIT->lock));
452 return StubToFunctionMap;
456 /// JITResolver - Keep track of, and resolve, call sites for functions that
457 /// have not yet been compiled.
459 /// MCE - The MachineCodeEmitter to use to emit stubs with.
460 MachineCodeEmitter &MCE;
462 /// LazyResolverFn - The target lazy resolver function that we actually
463 /// rewrite instructions to use.
464 TargetJITInfo::LazyResolverFn LazyResolverFn;
466 JITResolverState state;
468 /// ExternalFnToStubMap - This is the equivalent of FunctionToStubMap for
469 /// external functions.
470 std::map<void*, void*> ExternalFnToStubMap;
472 //map addresses to indexes in the GOT
473 std::map<void*, unsigned> revGOTMap;
474 unsigned nextGOTIndex;
477 JITResolver(MachineCodeEmitter &mce) : MCE(mce), nextGOTIndex(0) {
479 TheJIT->getJITInfo().getLazyResolverFunction(JITCompilerFn);
482 /// getFunctionStub - This returns a pointer to a function stub, creating
483 /// one on demand as needed.
484 void *getFunctionStub(Function *F);
486 /// getExternalFunctionStub - Return a stub for the function at the
487 /// specified address, created lazily on demand.
488 void *getExternalFunctionStub(void *FnAddr);
490 /// AddCallbackAtLocation - If the target is capable of rewriting an
491 /// instruction without the use of a stub, record the location of the use so
492 /// we know which function is being used at the location.
493 void *AddCallbackAtLocation(Function *F, void *Location) {
494 MutexGuard locked(TheJIT->lock);
495 /// Get the target-specific JIT resolver function.
496 state.getStubToFunctionMap(locked)[Location] = F;
497 return (void*)(intptr_t)LazyResolverFn;
500 /// getGOTIndexForAddress - Return a new or existing index in the GOT for
501 /// and address. This function only manages slots, it does not manage the
502 /// contents of the slots or the memory associated with the GOT.
503 unsigned getGOTIndexForAddr(void* addr);
505 /// JITCompilerFn - This function is called to resolve a stub to a compiled
506 /// address. If the LLVM Function corresponding to the stub has not yet
507 /// been compiled, this function compiles it first.
508 static void *JITCompilerFn(void *Stub);
512 /// getJITResolver - This function returns the one instance of the JIT resolver.
514 static JITResolver &getJITResolver(MachineCodeEmitter *MCE = 0) {
515 static JITResolver TheJITResolver(*MCE);
516 return TheJITResolver;
519 #if (defined(__POWERPC__) || defined (__ppc__) || defined(_POWER)) && \
521 extern "C" void sys_icache_invalidate(const void *Addr, size_t len);
524 /// synchronizeICache - On some targets, the JIT emitted code must be
525 /// explicitly refetched to ensure correct execution.
526 static void synchronizeICache(const void *Addr, size_t len) {
527 #if (defined(__POWERPC__) || defined (__ppc__) || defined(_POWER)) && \
529 sys_icache_invalidate(Addr, len);
533 /// getFunctionStub - This returns a pointer to a function stub, creating
534 /// one on demand as needed.
535 void *JITResolver::getFunctionStub(Function *F) {
536 MutexGuard locked(TheJIT->lock);
538 // If we already have a stub for this function, recycle it.
539 void *&Stub = state.getFunctionToStubMap(locked)[F];
540 if (Stub) return Stub;
542 // Call the lazy resolver function unless we already KNOW it is an external
543 // function, in which case we just skip the lazy resolution step.
544 void *Actual = (void*)(intptr_t)LazyResolverFn;
545 if (F->isExternal() && !F->hasNotBeenReadFromBytecode())
546 Actual = TheJIT->getPointerToFunction(F);
548 // Otherwise, codegen a new stub. For now, the stub will call the lazy
549 // resolver function.
550 Stub = TheJIT->getJITInfo().emitFunctionStub(Actual, MCE);
552 if (Actual != (void*)(intptr_t)LazyResolverFn) {
553 // If we are getting the stub for an external function, we really want the
554 // address of the stub in the GlobalAddressMap for the JIT, not the address
555 // of the external function.
556 TheJIT->updateGlobalMapping(F, Stub);
559 // Invalidate the icache if necessary.
560 synchronizeICache(Stub, MCE.getCurrentPCValue()-(intptr_t)Stub);
562 DOUT << "JIT: Stub emitted at [" << Stub << "] for function '"
563 << F->getName() << "'\n";
565 // Finally, keep track of the stub-to-Function mapping so that the
566 // JITCompilerFn knows which function to compile!
567 state.getStubToFunctionMap(locked)[Stub] = F;
571 /// getExternalFunctionStub - Return a stub for the function at the
572 /// specified address, created lazily on demand.
573 void *JITResolver::getExternalFunctionStub(void *FnAddr) {
574 // If we already have a stub for this function, recycle it.
575 void *&Stub = ExternalFnToStubMap[FnAddr];
576 if (Stub) return Stub;
578 Stub = TheJIT->getJITInfo().emitFunctionStub(FnAddr, MCE);
580 // Invalidate the icache if necessary.
581 synchronizeICache(Stub, MCE.getCurrentPCValue()-(intptr_t)Stub);
583 DOUT << "JIT: Stub emitted at [" << Stub
584 << "] for external function at '" << FnAddr << "'\n";
588 unsigned JITResolver::getGOTIndexForAddr(void* addr) {
589 unsigned idx = revGOTMap[addr];
591 idx = ++nextGOTIndex;
592 revGOTMap[addr] = idx;
593 DOUT << "Adding GOT entry " << idx
594 << " for addr " << addr << "\n";
595 // ((void**)MemMgr.getGOTBase())[idx] = addr;
600 /// JITCompilerFn - This function is called when a lazy compilation stub has
601 /// been entered. It looks up which function this stub corresponds to, compiles
602 /// it if necessary, then returns the resultant function pointer.
603 void *JITResolver::JITCompilerFn(void *Stub) {
604 JITResolver &JR = getJITResolver();
606 MutexGuard locked(TheJIT->lock);
608 // The address given to us for the stub may not be exactly right, it might be
609 // a little bit after the stub. As such, use upper_bound to find it.
610 std::map<void*, Function*>::iterator I =
611 JR.state.getStubToFunctionMap(locked).upper_bound(Stub);
612 assert(I != JR.state.getStubToFunctionMap(locked).begin() &&
613 "This is not a known stub!");
614 Function *F = (--I)->second;
616 // If disabled, emit a useful error message and abort.
617 if (TheJIT->isLazyCompilationDisabled()) {
618 cerr << "LLVM JIT requested to do lazy compilation of function '"
619 << F->getName() << "' when lazy compiles are disabled!\n";
623 // We might like to remove the stub from the StubToFunction map.
624 // We can't do that! Multiple threads could be stuck, waiting to acquire the
625 // lock above. As soon as the 1st function finishes compiling the function,
626 // the next one will be released, and needs to be able to find the function it
628 //JR.state.getStubToFunctionMap(locked).erase(I);
630 DOUT << "JIT: Lazily resolving function '" << F->getName()
631 << "' In stub ptr = " << Stub << " actual ptr = "
634 void *Result = TheJIT->getPointerToFunction(F);
636 // We don't need to reuse this stub in the future, as F is now compiled.
637 JR.state.getFunctionToStubMap(locked).erase(F);
639 // FIXME: We could rewrite all references to this stub if we knew them.
641 // What we will do is set the compiled function address to map to the
642 // same GOT entry as the stub so that later clients may update the GOT
643 // if they see it still using the stub address.
644 // Note: this is done so the Resolver doesn't have to manage GOT memory
645 // Do this without allocating map space if the target isn't using a GOT
646 if(JR.revGOTMap.find(Stub) != JR.revGOTMap.end())
647 JR.revGOTMap[Result] = JR.revGOTMap[Stub];
653 //===----------------------------------------------------------------------===//
657 /// JITEmitter - The JIT implementation of the MachineCodeEmitter, which is
658 /// used to output functions to memory for execution.
659 class JITEmitter : public MachineCodeEmitter {
660 JITMemoryManager MemMgr;
662 // When outputting a function stub in the context of some other function, we
663 // save BufferBegin/BufferEnd/CurBufferPtr here.
664 unsigned char *SavedBufferBegin, *SavedBufferEnd, *SavedCurBufferPtr;
666 /// Relocations - These are the relocations that the function needs, as
668 std::vector<MachineRelocation> Relocations;
670 /// MBBLocations - This vector is a mapping from MBB ID's to their address.
671 /// It is filled in by the StartMachineBasicBlock callback and queried by
672 /// the getMachineBasicBlockAddress callback.
673 std::vector<intptr_t> MBBLocations;
675 /// ConstantPool - The constant pool for the current function.
677 MachineConstantPool *ConstantPool;
679 /// ConstantPoolBase - A pointer to the first entry in the constant pool.
681 void *ConstantPoolBase;
683 /// JumpTable - The jump tables for the current function.
685 MachineJumpTableInfo *JumpTable;
687 /// JumpTableBase - A pointer to the first entry in the jump table.
691 JITEmitter(JIT &jit) : MemMgr(jit.getJITInfo().needsGOT()) {
693 if (MemMgr.isManagingGOT()) DOUT << "JIT is managing a GOT\n";
696 virtual void startFunction(MachineFunction &F);
697 virtual bool finishFunction(MachineFunction &F);
699 void emitConstantPool(MachineConstantPool *MCP);
700 void initJumpTableInfo(MachineJumpTableInfo *MJTI);
701 void emitJumpTableInfo(MachineJumpTableInfo *MJTI);
703 virtual void startFunctionStub(unsigned StubSize, unsigned Alignment = 1);
704 virtual void* finishFunctionStub(const Function *F);
706 virtual void addRelocation(const MachineRelocation &MR) {
707 Relocations.push_back(MR);
710 virtual void StartMachineBasicBlock(MachineBasicBlock *MBB) {
711 if (MBBLocations.size() <= (unsigned)MBB->getNumber())
712 MBBLocations.resize((MBB->getNumber()+1)*2);
713 MBBLocations[MBB->getNumber()] = getCurrentPCValue();
716 virtual intptr_t getConstantPoolEntryAddress(unsigned Entry) const;
717 virtual intptr_t getJumpTableEntryAddress(unsigned Entry) const;
719 virtual intptr_t getMachineBasicBlockAddress(MachineBasicBlock *MBB) const {
720 assert(MBBLocations.size() > (unsigned)MBB->getNumber() &&
721 MBBLocations[MBB->getNumber()] && "MBB not emitted!");
722 return MBBLocations[MBB->getNumber()];
725 /// deallocateMemForFunction - Deallocate all memory for the specified
727 void deallocateMemForFunction(Function *F) {
728 MemMgr.deallocateMemForFunction(F);
731 void *getPointerToGlobal(GlobalValue *GV, void *Reference, bool NoNeedStub);
735 void *JITEmitter::getPointerToGlobal(GlobalValue *V, void *Reference,
736 bool DoesntNeedStub) {
737 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(V)) {
738 /// FIXME: If we straightened things out, this could actually emit the
739 /// global immediately instead of queuing it for codegen later!
740 return TheJIT->getOrEmitGlobalVariable(GV);
743 // If we have already compiled the function, return a pointer to its body.
744 Function *F = cast<Function>(V);
745 void *ResultPtr = TheJIT->getPointerToGlobalIfAvailable(F);
746 if (ResultPtr) return ResultPtr;
748 if (F->isExternal() && !F->hasNotBeenReadFromBytecode()) {
749 // If this is an external function pointer, we can force the JIT to
750 // 'compile' it, which really just adds it to the map.
752 return TheJIT->getPointerToFunction(F);
754 return getJITResolver(this).getFunctionStub(F);
757 // Okay, the function has not been compiled yet, if the target callback
758 // mechanism is capable of rewriting the instruction directly, prefer to do
759 // that instead of emitting a stub.
761 return getJITResolver(this).AddCallbackAtLocation(F, Reference);
763 // Otherwise, we have to emit a lazy resolving stub.
764 return getJITResolver(this).getFunctionStub(F);
767 void JITEmitter::startFunction(MachineFunction &F) {
768 uintptr_t ActualSize;
769 BufferBegin = CurBufferPtr = MemMgr.startFunctionBody(ActualSize);
770 BufferEnd = BufferBegin+ActualSize;
772 // Ensure the constant pool/jump table info is at least 4-byte aligned.
775 emitConstantPool(F.getConstantPool());
776 initJumpTableInfo(F.getJumpTableInfo());
778 // About to start emitting the machine code for the function.
779 emitAlignment(std::max(F.getFunction()->getAlignment(), 8U));
780 TheJIT->updateGlobalMapping(F.getFunction(), CurBufferPtr);
782 MBBLocations.clear();
785 bool JITEmitter::finishFunction(MachineFunction &F) {
786 if (CurBufferPtr == BufferEnd) {
787 // FIXME: Allocate more space, then try again.
788 cerr << "JIT: Ran out of space for generated machine code!\n";
792 emitJumpTableInfo(F.getJumpTableInfo());
794 // FnStart is the start of the text, not the start of the constant pool and
795 // other per-function data.
796 unsigned char *FnStart =
797 (unsigned char *)TheJIT->getPointerToGlobalIfAvailable(F.getFunction());
798 unsigned char *FnEnd = CurBufferPtr;
800 MemMgr.endFunctionBody(F.getFunction(), BufferBegin, FnEnd);
801 NumBytes += FnEnd-FnStart;
803 if (!Relocations.empty()) {
804 NumRelos += Relocations.size();
806 // Resolve the relocations to concrete pointers.
807 for (unsigned i = 0, e = Relocations.size(); i != e; ++i) {
808 MachineRelocation &MR = Relocations[i];
811 ResultPtr = TheJIT->getPointerToNamedFunction(MR.getString());
813 // If the target REALLY wants a stub for this function, emit it now.
814 if (!MR.doesntNeedFunctionStub())
815 ResultPtr = getJITResolver(this).getExternalFunctionStub(ResultPtr);
816 } else if (MR.isGlobalValue()) {
817 ResultPtr = getPointerToGlobal(MR.getGlobalValue(),
818 BufferBegin+MR.getMachineCodeOffset(),
819 MR.doesntNeedFunctionStub());
820 } else if (MR.isBasicBlock()) {
821 ResultPtr = (void*)getMachineBasicBlockAddress(MR.getBasicBlock());
822 } else if (MR.isConstantPoolIndex()) {
823 ResultPtr=(void*)getConstantPoolEntryAddress(MR.getConstantPoolIndex());
825 assert(MR.isJumpTableIndex());
826 ResultPtr=(void*)getJumpTableEntryAddress(MR.getJumpTableIndex());
829 MR.setResultPointer(ResultPtr);
831 // if we are managing the GOT and the relocation wants an index,
833 if (MemMgr.isManagingGOT() && MR.isGOTRelative()) {
834 unsigned idx = getJITResolver(this).getGOTIndexForAddr(ResultPtr);
836 if (((void**)MemMgr.getGOTBase())[idx] != ResultPtr) {
837 DOUT << "GOT was out of date for " << ResultPtr
838 << " pointing at " << ((void**)MemMgr.getGOTBase())[idx]
840 ((void**)MemMgr.getGOTBase())[idx] = ResultPtr;
845 TheJIT->getJITInfo().relocate(BufferBegin, &Relocations[0],
846 Relocations.size(), MemMgr.getGOTBase());
849 // Update the GOT entry for F to point to the new code.
850 if(MemMgr.isManagingGOT()) {
851 unsigned idx = getJITResolver(this).getGOTIndexForAddr((void*)BufferBegin);
852 if (((void**)MemMgr.getGOTBase())[idx] != (void*)BufferBegin) {
853 DOUT << "GOT was out of date for " << (void*)BufferBegin
854 << " pointing at " << ((void**)MemMgr.getGOTBase())[idx] << "\n";
855 ((void**)MemMgr.getGOTBase())[idx] = (void*)BufferBegin;
859 // Invalidate the icache if necessary.
860 synchronizeICache(FnStart, FnEnd-FnStart);
862 DOUT << "JIT: Finished CodeGen of [" << (void*)FnStart
863 << "] Function: " << F.getFunction()->getName()
864 << ": " << (FnEnd-FnStart) << " bytes of text, "
865 << Relocations.size() << " relocations\n";
870 void JITEmitter::emitConstantPool(MachineConstantPool *MCP) {
871 const std::vector<MachineConstantPoolEntry> &Constants = MCP->getConstants();
872 if (Constants.empty()) return;
874 MachineConstantPoolEntry CPE = Constants.back();
875 unsigned Size = CPE.Offset;
876 const Type *Ty = CPE.isMachineConstantPoolEntry()
877 ? CPE.Val.MachineCPVal->getType() : CPE.Val.ConstVal->getType();
878 Size += TheJIT->getTargetData()->getTypeSize(Ty);
880 ConstantPoolBase = allocateSpace(Size, 1 << MCP->getConstantPoolAlignment());
883 if (ConstantPoolBase == 0) return; // Buffer overflow.
885 // Initialize the memory for all of the constant pool entries.
886 for (unsigned i = 0, e = Constants.size(); i != e; ++i) {
887 void *CAddr = (char*)ConstantPoolBase+Constants[i].Offset;
888 if (Constants[i].isMachineConstantPoolEntry()) {
889 // FIXME: add support to lower machine constant pool values into bytes!
890 cerr << "Initialize memory with machine specific constant pool entry"
891 << " has not been implemented!\n";
894 TheJIT->InitializeMemory(Constants[i].Val.ConstVal, CAddr);
898 void JITEmitter::initJumpTableInfo(MachineJumpTableInfo *MJTI) {
899 const std::vector<MachineJumpTableEntry> &JT = MJTI->getJumpTables();
900 if (JT.empty()) return;
902 unsigned NumEntries = 0;
903 for (unsigned i = 0, e = JT.size(); i != e; ++i)
904 NumEntries += JT[i].MBBs.size();
906 unsigned EntrySize = MJTI->getEntrySize();
908 // Just allocate space for all the jump tables now. We will fix up the actual
909 // MBB entries in the tables after we emit the code for each block, since then
910 // we will know the final locations of the MBBs in memory.
912 JumpTableBase = allocateSpace(NumEntries * EntrySize, MJTI->getAlignment());
915 void JITEmitter::emitJumpTableInfo(MachineJumpTableInfo *MJTI) {
916 const std::vector<MachineJumpTableEntry> &JT = MJTI->getJumpTables();
917 if (JT.empty() || JumpTableBase == 0) return;
919 if (TargetMachine::getRelocationModel() == Reloc::PIC_) {
920 assert(MJTI->getEntrySize() == 4 && "Cross JIT'ing?");
921 // For each jump table, place the offset from the beginning of the table
922 // to the target address.
923 int *SlotPtr = (int*)JumpTableBase;
925 for (unsigned i = 0, e = JT.size(); i != e; ++i) {
926 const std::vector<MachineBasicBlock*> &MBBs = JT[i].MBBs;
927 // Store the offset of the basic block for this jump table slot in the
928 // memory we allocated for the jump table in 'initJumpTableInfo'
929 intptr_t Base = (intptr_t)SlotPtr;
930 for (unsigned mi = 0, me = MBBs.size(); mi != me; ++mi)
931 *SlotPtr++ = (intptr_t)getMachineBasicBlockAddress(MBBs[mi]) - Base;
934 assert(MJTI->getEntrySize() == sizeof(void*) && "Cross JIT'ing?");
936 // For each jump table, map each target in the jump table to the address of
937 // an emitted MachineBasicBlock.
938 intptr_t *SlotPtr = (intptr_t*)JumpTableBase;
940 for (unsigned i = 0, e = JT.size(); i != e; ++i) {
941 const std::vector<MachineBasicBlock*> &MBBs = JT[i].MBBs;
942 // Store the address of the basic block for this jump table slot in the
943 // memory we allocated for the jump table in 'initJumpTableInfo'
944 for (unsigned mi = 0, me = MBBs.size(); mi != me; ++mi)
945 *SlotPtr++ = getMachineBasicBlockAddress(MBBs[mi]);
950 void JITEmitter::startFunctionStub(unsigned StubSize, unsigned Alignment) {
951 SavedBufferBegin = BufferBegin;
952 SavedBufferEnd = BufferEnd;
953 SavedCurBufferPtr = CurBufferPtr;
955 BufferBegin = CurBufferPtr = MemMgr.allocateStub(StubSize, Alignment);
956 BufferEnd = BufferBegin+StubSize+1;
959 void *JITEmitter::finishFunctionStub(const Function *F) {
960 NumBytes += getCurrentPCOffset();
961 std::swap(SavedBufferBegin, BufferBegin);
962 BufferEnd = SavedBufferEnd;
963 CurBufferPtr = SavedCurBufferPtr;
964 return SavedBufferBegin;
967 // getConstantPoolEntryAddress - Return the address of the 'ConstantNum' entry
968 // in the constant pool that was last emitted with the 'emitConstantPool'
971 intptr_t JITEmitter::getConstantPoolEntryAddress(unsigned ConstantNum) const {
972 assert(ConstantNum < ConstantPool->getConstants().size() &&
973 "Invalid ConstantPoolIndex!");
974 return (intptr_t)ConstantPoolBase +
975 ConstantPool->getConstants()[ConstantNum].Offset;
978 // getJumpTableEntryAddress - Return the address of the JumpTable with index
979 // 'Index' in the jumpp table that was last initialized with 'initJumpTableInfo'
981 intptr_t JITEmitter::getJumpTableEntryAddress(unsigned Index) const {
982 const std::vector<MachineJumpTableEntry> &JT = JumpTable->getJumpTables();
983 assert(Index < JT.size() && "Invalid jump table index!");
986 unsigned EntrySize = JumpTable->getEntrySize();
988 for (unsigned i = 0; i < Index; ++i)
989 Offset += JT[i].MBBs.size();
993 return (intptr_t)((char *)JumpTableBase + Offset);
996 //===----------------------------------------------------------------------===//
997 // Public interface to this file
998 //===----------------------------------------------------------------------===//
1000 MachineCodeEmitter *JIT::createEmitter(JIT &jit) {
1001 return new JITEmitter(jit);
1004 // getPointerToNamedFunction - This function is used as a global wrapper to
1005 // JIT::getPointerToNamedFunction for the purpose of resolving symbols when
1006 // bugpoint is debugging the JIT. In that scenario, we are loading an .so and
1007 // need to resolve function(s) that are being mis-codegenerated, so we need to
1008 // resolve their addresses at runtime, and this is the way to do it.
1010 void *getPointerToNamedFunction(const char *Name) {
1011 if (Function *F = TheJIT->FindFunctionNamed(Name))
1012 return TheJIT->getPointerToFunction(F);
1013 return TheJIT->getPointerToNamedFunction(Name);
1017 // getPointerToFunctionOrStub - If the specified function has been
1018 // code-gen'd, return a pointer to the function. If not, compile it, or use
1019 // a stub to implement lazy compilation if available.
1021 void *JIT::getPointerToFunctionOrStub(Function *F) {
1022 // If we have already code generated the function, just return the address.
1023 if (void *Addr = getPointerToGlobalIfAvailable(F))
1026 // Get a stub if the target supports it
1027 return getJITResolver(MCE).getFunctionStub(F);
1030 /// freeMachineCodeForFunction - release machine code memory for given Function.
1032 void JIT::freeMachineCodeForFunction(Function *F) {
1033 // Delete translation for this from the ExecutionEngine, so it will get
1034 // retranslated next time it is used.
1035 updateGlobalMapping(F, 0);
1037 // Free the actual memory for the function body and related stuff.
1038 assert(dynamic_cast<JITEmitter*>(MCE) && "Unexpected MCE?");
1039 dynamic_cast<JITEmitter*>(MCE)->deallocateMemForFunction(F);