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/System/Disassembler.h"
32 #include "llvm/ADT/Statistic.h"
33 #include "llvm/System/Memory.h"
37 STATISTIC(NumBytes, "Number of bytes of machine code compiled");
38 STATISTIC(NumRelos, "Number of relocations applied");
39 static JIT *TheJIT = 0;
41 //===----------------------------------------------------------------------===//
42 // JITMemoryManager code.
45 /// MemoryRangeHeader - For a range of memory, this is the header that we put
46 /// on the block of memory. It is carefully crafted to be one word of memory.
47 /// Allocated blocks have just this header, free'd blocks have FreeRangeHeader
48 /// which starts with this.
49 struct FreeRangeHeader;
50 struct MemoryRangeHeader {
51 /// ThisAllocated - This is true if this block is currently allocated. If
52 /// not, this can be converted to a FreeRangeHeader.
53 intptr_t ThisAllocated : 1;
55 /// PrevAllocated - Keep track of whether the block immediately before us is
56 /// allocated. If not, the word immediately before this header is the size
57 /// of the previous block.
58 intptr_t PrevAllocated : 1;
60 /// BlockSize - This is the size in bytes of this memory block,
61 /// including this header.
62 uintptr_t BlockSize : (sizeof(intptr_t)*8 - 2);
65 /// getBlockAfter - Return the memory block immediately after this one.
67 MemoryRangeHeader &getBlockAfter() const {
68 return *(MemoryRangeHeader*)((char*)this+BlockSize);
71 /// getFreeBlockBefore - If the block before this one is free, return it,
72 /// otherwise return null.
73 FreeRangeHeader *getFreeBlockBefore() const {
74 if (PrevAllocated) return 0;
75 intptr_t PrevSize = ((intptr_t *)this)[-1];
76 return (FreeRangeHeader*)((char*)this-PrevSize);
79 /// FreeBlock - Turn an allocated block into a free block, adjusting
80 /// bits in the object headers, and adding an end of region memory block.
81 FreeRangeHeader *FreeBlock(FreeRangeHeader *FreeList);
83 /// TrimAllocationToSize - If this allocated block is significantly larger
84 /// than NewSize, split it into two pieces (where the former is NewSize
85 /// bytes, including the header), and add the new block to the free list.
86 FreeRangeHeader *TrimAllocationToSize(FreeRangeHeader *FreeList,
90 /// FreeRangeHeader - For a memory block that isn't already allocated, this
91 /// keeps track of the current block and has a pointer to the next free block.
92 /// Free blocks are kept on a circularly linked list.
93 struct FreeRangeHeader : public MemoryRangeHeader {
94 FreeRangeHeader *Prev;
95 FreeRangeHeader *Next;
97 /// getMinBlockSize - Get the minimum size for a memory block. Blocks
98 /// smaller than this size cannot be created.
99 static unsigned getMinBlockSize() {
100 return sizeof(FreeRangeHeader)+sizeof(intptr_t);
103 /// SetEndOfBlockSizeMarker - The word at the end of every free block is
104 /// known to be the size of the free block. Set it for this block.
105 void SetEndOfBlockSizeMarker() {
106 void *EndOfBlock = (char*)this + BlockSize;
107 ((intptr_t *)EndOfBlock)[-1] = BlockSize;
110 FreeRangeHeader *RemoveFromFreeList() {
111 assert(Next->Prev == this && Prev->Next == this && "Freelist broken!");
113 return Prev->Next = Next;
116 void AddToFreeList(FreeRangeHeader *FreeList) {
118 Prev = FreeList->Prev;
123 /// GrowBlock - The block after this block just got deallocated. Merge it
124 /// into the current block.
125 void GrowBlock(uintptr_t NewSize);
127 /// AllocateBlock - Mark this entire block allocated, updating freelists
128 /// etc. This returns a pointer to the circular free-list.
129 FreeRangeHeader *AllocateBlock();
134 /// AllocateBlock - Mark this entire block allocated, updating freelists
135 /// etc. This returns a pointer to the circular free-list.
136 FreeRangeHeader *FreeRangeHeader::AllocateBlock() {
137 assert(!ThisAllocated && !getBlockAfter().PrevAllocated &&
138 "Cannot allocate an allocated block!");
139 // Mark this block allocated.
141 getBlockAfter().PrevAllocated = 1;
143 // Remove it from the free list.
144 return RemoveFromFreeList();
147 /// FreeBlock - Turn an allocated block into a free block, adjusting
148 /// bits in the object headers, and adding an end of region memory block.
149 /// If possible, coallesce this block with neighboring blocks. Return the
150 /// FreeRangeHeader to allocate from.
151 FreeRangeHeader *MemoryRangeHeader::FreeBlock(FreeRangeHeader *FreeList) {
152 MemoryRangeHeader *FollowingBlock = &getBlockAfter();
153 assert(ThisAllocated && "This block is already allocated!");
154 assert(FollowingBlock->PrevAllocated && "Flags out of sync!");
156 FreeRangeHeader *FreeListToReturn = FreeList;
158 // If the block after this one is free, merge it into this block.
159 if (!FollowingBlock->ThisAllocated) {
160 FreeRangeHeader &FollowingFreeBlock = *(FreeRangeHeader *)FollowingBlock;
161 // "FreeList" always needs to be a valid free block. If we're about to
162 // coallesce with it, update our notion of what the free list is.
163 if (&FollowingFreeBlock == FreeList) {
164 FreeList = FollowingFreeBlock.Next;
165 FreeListToReturn = 0;
166 assert(&FollowingFreeBlock != FreeList && "No tombstone block?");
168 FollowingFreeBlock.RemoveFromFreeList();
170 // Include the following block into this one.
171 BlockSize += FollowingFreeBlock.BlockSize;
172 FollowingBlock = &FollowingFreeBlock.getBlockAfter();
174 // Tell the block after the block we are coallescing that this block is
176 FollowingBlock->PrevAllocated = 1;
179 assert(FollowingBlock->ThisAllocated && "Missed coallescing?");
181 if (FreeRangeHeader *PrevFreeBlock = getFreeBlockBefore()) {
182 PrevFreeBlock->GrowBlock(PrevFreeBlock->BlockSize + BlockSize);
183 return FreeListToReturn ? FreeListToReturn : PrevFreeBlock;
186 // Otherwise, mark this block free.
187 FreeRangeHeader &FreeBlock = *(FreeRangeHeader*)this;
188 FollowingBlock->PrevAllocated = 0;
189 FreeBlock.ThisAllocated = 0;
191 // Link this into the linked list of free blocks.
192 FreeBlock.AddToFreeList(FreeList);
194 // Add a marker at the end of the block, indicating the size of this free
196 FreeBlock.SetEndOfBlockSizeMarker();
197 return FreeListToReturn ? FreeListToReturn : &FreeBlock;
200 /// GrowBlock - The block after this block just got deallocated. Merge it
201 /// into the current block.
202 void FreeRangeHeader::GrowBlock(uintptr_t NewSize) {
203 assert(NewSize > BlockSize && "Not growing block?");
205 SetEndOfBlockSizeMarker();
206 getBlockAfter().PrevAllocated = 0;
209 /// TrimAllocationToSize - If this allocated block is significantly larger
210 /// than NewSize, split it into two pieces (where the former is NewSize
211 /// bytes, including the header), and add the new block to the free list.
212 FreeRangeHeader *MemoryRangeHeader::
213 TrimAllocationToSize(FreeRangeHeader *FreeList, uint64_t NewSize) {
214 assert(ThisAllocated && getBlockAfter().PrevAllocated &&
215 "Cannot deallocate part of an allocated block!");
217 // Round up size for alignment of header.
218 unsigned HeaderAlign = __alignof(FreeRangeHeader);
219 NewSize = (NewSize+ (HeaderAlign-1)) & ~(HeaderAlign-1);
221 // Size is now the size of the block we will remove from the start of the
223 assert(NewSize <= BlockSize &&
224 "Allocating more space from this block than exists!");
226 // If splitting this block will cause the remainder to be too small, do not
228 if (BlockSize <= NewSize+FreeRangeHeader::getMinBlockSize())
231 // Otherwise, we splice the required number of bytes out of this block, form
232 // a new block immediately after it, then mark this block allocated.
233 MemoryRangeHeader &FormerNextBlock = getBlockAfter();
235 // Change the size of this block.
238 // Get the new block we just sliced out and turn it into a free block.
239 FreeRangeHeader &NewNextBlock = (FreeRangeHeader &)getBlockAfter();
240 NewNextBlock.BlockSize = (char*)&FormerNextBlock - (char*)&NewNextBlock;
241 NewNextBlock.ThisAllocated = 0;
242 NewNextBlock.PrevAllocated = 1;
243 NewNextBlock.SetEndOfBlockSizeMarker();
244 FormerNextBlock.PrevAllocated = 0;
245 NewNextBlock.AddToFreeList(FreeList);
246 return &NewNextBlock;
251 /// JITMemoryManager - Manage memory for the JIT code generation in a logical,
252 /// sane way. This splits a large block of MAP_NORESERVE'd memory into two
253 /// sections, one for function stubs, one for the functions themselves. We
254 /// have to do this because we may need to emit a function stub while in the
255 /// middle of emitting a function, and we don't know how large the function we
256 /// are emitting is. This never bothers to release the memory, because when
257 /// we are ready to destroy the JIT, the program exits.
258 class JITMemoryManager {
259 std::vector<sys::MemoryBlock> Blocks; // Memory blocks allocated by the JIT
260 FreeRangeHeader *FreeMemoryList; // Circular list of free blocks.
262 // When emitting code into a memory block, this is the block.
263 MemoryRangeHeader *CurBlock;
265 unsigned char *CurStubPtr, *StubBase;
266 unsigned char *GOTBase; // Target Specific reserved memory
268 // Centralize memory block allocation.
269 sys::MemoryBlock getNewMemoryBlock(unsigned size);
271 std::map<const Function*, MemoryRangeHeader*> FunctionBlocks;
273 JITMemoryManager(bool useGOT);
276 inline unsigned char *allocateStub(unsigned StubSize, unsigned Alignment);
278 /// startFunctionBody - When a function starts, allocate a block of free
279 /// executable memory, returning a pointer to it and its actual size.
280 unsigned char *startFunctionBody(uintptr_t &ActualSize) {
281 CurBlock = FreeMemoryList;
283 // Allocate the entire memory block.
284 FreeMemoryList = FreeMemoryList->AllocateBlock();
285 ActualSize = CurBlock->BlockSize-sizeof(MemoryRangeHeader);
286 return (unsigned char *)(CurBlock+1);
289 /// endFunctionBody - The function F is now allocated, and takes the memory
290 /// in the range [FunctionStart,FunctionEnd).
291 void endFunctionBody(const Function *F, unsigned char *FunctionStart,
292 unsigned char *FunctionEnd) {
293 assert(FunctionEnd > FunctionStart);
294 assert(FunctionStart == (unsigned char *)(CurBlock+1) &&
295 "Mismatched function start/end!");
297 uintptr_t BlockSize = FunctionEnd - (unsigned char *)CurBlock;
298 FunctionBlocks[F] = CurBlock;
300 // Release the memory at the end of this block that isn't needed.
301 FreeMemoryList =CurBlock->TrimAllocationToSize(FreeMemoryList, BlockSize);
304 unsigned char *getGOTBase() const {
307 bool isManagingGOT() const {
308 return GOTBase != NULL;
311 /// deallocateMemForFunction - Deallocate all memory for the specified
313 void deallocateMemForFunction(const Function *F) {
314 std::map<const Function*, MemoryRangeHeader*>::iterator
315 I = FunctionBlocks.find(F);
316 if (I == FunctionBlocks.end()) return;
318 // Find the block that is allocated for this function.
319 MemoryRangeHeader *MemRange = I->second;
320 assert(MemRange->ThisAllocated && "Block isn't allocated!");
322 // Fill the buffer with garbage!
323 DEBUG(memset(MemRange+1, 0xCD, MemRange->BlockSize-sizeof(*MemRange)));
326 FreeMemoryList = MemRange->FreeBlock(FreeMemoryList);
328 // Finally, remove this entry from FunctionBlocks.
329 FunctionBlocks.erase(I);
334 JITMemoryManager::JITMemoryManager(bool useGOT) {
335 // Allocate a 16M block of memory for functions.
336 sys::MemoryBlock MemBlock = getNewMemoryBlock(16 << 20);
338 unsigned char *MemBase = reinterpret_cast<unsigned char*>(MemBlock.base());
340 // Allocate stubs backwards from the base, allocate functions forward
343 CurStubPtr = MemBase + 512*1024; // Use 512k for stubs, working backwards.
345 // We set up the memory chunk with 4 mem regions, like this:
347 // [ Free #0 ] -> Large space to allocate functions from.
348 // [ Allocated #1 ] -> Tiny space to separate regions.
349 // [ Free #2 ] -> Tiny space so there is always at least 1 free block.
350 // [ Allocated #3 ] -> Tiny space to prevent looking past end of block.
353 // The last three blocks are never deallocated or touched.
355 // Add MemoryRangeHeader to the end of the memory region, indicating that
356 // the space after the block of memory is allocated. This is block #3.
357 MemoryRangeHeader *Mem3 = (MemoryRangeHeader*)(MemBase+MemBlock.size())-1;
358 Mem3->ThisAllocated = 1;
359 Mem3->PrevAllocated = 0;
362 /// Add a tiny free region so that the free list always has one entry.
363 FreeRangeHeader *Mem2 =
364 (FreeRangeHeader *)(((char*)Mem3)-FreeRangeHeader::getMinBlockSize());
365 Mem2->ThisAllocated = 0;
366 Mem2->PrevAllocated = 1;
367 Mem2->BlockSize = FreeRangeHeader::getMinBlockSize();
368 Mem2->SetEndOfBlockSizeMarker();
369 Mem2->Prev = Mem2; // Mem2 *is* the free list for now.
372 /// Add a tiny allocated region so that Mem2 is never coallesced away.
373 MemoryRangeHeader *Mem1 = (MemoryRangeHeader*)Mem2-1;
374 Mem1->ThisAllocated = 1;
375 Mem1->PrevAllocated = 0;
376 Mem1->BlockSize = (char*)Mem2 - (char*)Mem1;
378 // Add a FreeRangeHeader to the start of the function body region, indicating
379 // that the space is free. Mark the previous block allocated so we never look
381 FreeRangeHeader *Mem0 = (FreeRangeHeader*)CurStubPtr;
382 Mem0->ThisAllocated = 0;
383 Mem0->PrevAllocated = 1;
384 Mem0->BlockSize = (char*)Mem1-(char*)Mem0;
385 Mem0->SetEndOfBlockSizeMarker();
386 Mem0->AddToFreeList(Mem2);
388 // Start out with the freelist pointing to Mem0.
389 FreeMemoryList = Mem0;
393 if (useGOT) GOTBase = new unsigned char[sizeof(void*) * 8192];
396 JITMemoryManager::~JITMemoryManager() {
397 for (unsigned i = 0, e = Blocks.size(); i != e; ++i)
398 sys::Memory::ReleaseRWX(Blocks[i]);
404 unsigned char *JITMemoryManager::allocateStub(unsigned StubSize,
405 unsigned Alignment) {
406 CurStubPtr -= StubSize;
407 CurStubPtr = (unsigned char*)(((intptr_t)CurStubPtr) &
408 ~(intptr_t)(Alignment-1));
409 if (CurStubPtr < StubBase) {
410 // FIXME: allocate a new block
411 cerr << "JIT ran out of memory for function stubs!\n";
417 sys::MemoryBlock JITMemoryManager::getNewMemoryBlock(unsigned size) {
418 // Allocate a new block close to the last one.
419 const sys::MemoryBlock *BOld = Blocks.empty() ? 0 : &Blocks.front();
421 sys::MemoryBlock B = sys::Memory::AllocateRWX(size, BOld, &ErrMsg);
423 cerr << "Allocation failed when allocating new memory in the JIT\n";
424 cerr << ErrMsg << "\n";
431 //===----------------------------------------------------------------------===//
432 // JIT lazy compilation code.
435 class JITResolverState {
437 /// FunctionToStubMap - Keep track of the stub created for a particular
438 /// function so that we can reuse them if necessary.
439 std::map<Function*, void*> FunctionToStubMap;
441 /// StubToFunctionMap - Keep track of the function that each stub
443 std::map<void*, Function*> StubToFunctionMap;
446 std::map<Function*, void*>& getFunctionToStubMap(const MutexGuard& locked) {
447 assert(locked.holds(TheJIT->lock));
448 return FunctionToStubMap;
451 std::map<void*, Function*>& getStubToFunctionMap(const MutexGuard& locked) {
452 assert(locked.holds(TheJIT->lock));
453 return StubToFunctionMap;
457 /// JITResolver - Keep track of, and resolve, call sites for functions that
458 /// have not yet been compiled.
460 /// MCE - The MachineCodeEmitter to use to emit stubs with.
461 MachineCodeEmitter &MCE;
463 /// LazyResolverFn - The target lazy resolver function that we actually
464 /// rewrite instructions to use.
465 TargetJITInfo::LazyResolverFn LazyResolverFn;
467 JITResolverState state;
469 /// ExternalFnToStubMap - This is the equivalent of FunctionToStubMap for
470 /// external functions.
471 std::map<void*, void*> ExternalFnToStubMap;
473 //map addresses to indexes in the GOT
474 std::map<void*, unsigned> revGOTMap;
475 unsigned nextGOTIndex;
478 JITResolver(MachineCodeEmitter &mce) : MCE(mce), nextGOTIndex(0) {
480 TheJIT->getJITInfo().getLazyResolverFunction(JITCompilerFn);
483 /// getFunctionStub - This returns a pointer to a function stub, creating
484 /// one on demand as needed.
485 void *getFunctionStub(Function *F);
487 /// getExternalFunctionStub - Return a stub for the function at the
488 /// specified address, created lazily on demand.
489 void *getExternalFunctionStub(void *FnAddr);
491 /// AddCallbackAtLocation - If the target is capable of rewriting an
492 /// instruction without the use of a stub, record the location of the use so
493 /// we know which function is being used at the location.
494 void *AddCallbackAtLocation(Function *F, void *Location) {
495 MutexGuard locked(TheJIT->lock);
496 /// Get the target-specific JIT resolver function.
497 state.getStubToFunctionMap(locked)[Location] = F;
498 return (void*)(intptr_t)LazyResolverFn;
501 /// getGOTIndexForAddress - Return a new or existing index in the GOT for
502 /// and address. This function only manages slots, it does not manage the
503 /// contents of the slots or the memory associated with the GOT.
504 unsigned getGOTIndexForAddr(void* addr);
506 /// JITCompilerFn - This function is called to resolve a stub to a compiled
507 /// address. If the LLVM Function corresponding to the stub has not yet
508 /// been compiled, this function compiles it first.
509 static void *JITCompilerFn(void *Stub);
513 /// getJITResolver - This function returns the one instance of the JIT resolver.
515 static JITResolver &getJITResolver(MachineCodeEmitter *MCE = 0) {
516 static JITResolver TheJITResolver(*MCE);
517 return TheJITResolver;
520 #if (defined(__POWERPC__) || defined (__ppc__) || defined(_POWER)) && \
522 extern "C" void sys_icache_invalidate(const void *Addr, size_t len);
525 /// synchronizeICache - On some targets, the JIT emitted code must be
526 /// explicitly refetched to ensure correct execution.
527 static void synchronizeICache(const void *Addr, size_t len) {
528 #if (defined(__POWERPC__) || defined (__ppc__) || defined(_POWER)) && \
530 sys_icache_invalidate(Addr, len);
534 /// getFunctionStub - This returns a pointer to a function stub, creating
535 /// one on demand as needed.
536 void *JITResolver::getFunctionStub(Function *F) {
537 MutexGuard locked(TheJIT->lock);
539 // If we already have a stub for this function, recycle it.
540 void *&Stub = state.getFunctionToStubMap(locked)[F];
541 if (Stub) return Stub;
543 // Call the lazy resolver function unless we already KNOW it is an external
544 // function, in which case we just skip the lazy resolution step.
545 void *Actual = (void*)(intptr_t)LazyResolverFn;
546 if (F->isExternal() && !F->hasNotBeenReadFromBytecode())
547 Actual = TheJIT->getPointerToFunction(F);
549 // Otherwise, codegen a new stub. For now, the stub will call the lazy
550 // resolver function.
551 Stub = TheJIT->getJITInfo().emitFunctionStub(Actual, MCE);
553 if (Actual != (void*)(intptr_t)LazyResolverFn) {
554 // If we are getting the stub for an external function, we really want the
555 // address of the stub in the GlobalAddressMap for the JIT, not the address
556 // of the external function.
557 TheJIT->updateGlobalMapping(F, Stub);
560 // Invalidate the icache if necessary.
561 synchronizeICache(Stub, MCE.getCurrentPCValue()-(intptr_t)Stub);
563 DOUT << "JIT: Stub emitted at [" << Stub << "] for function '"
564 << F->getName() << "'\n";
566 // Finally, keep track of the stub-to-Function mapping so that the
567 // JITCompilerFn knows which function to compile!
568 state.getStubToFunctionMap(locked)[Stub] = F;
572 /// getExternalFunctionStub - Return a stub for the function at the
573 /// specified address, created lazily on demand.
574 void *JITResolver::getExternalFunctionStub(void *FnAddr) {
575 // If we already have a stub for this function, recycle it.
576 void *&Stub = ExternalFnToStubMap[FnAddr];
577 if (Stub) return Stub;
579 Stub = TheJIT->getJITInfo().emitFunctionStub(FnAddr, MCE);
581 // Invalidate the icache if necessary.
582 synchronizeICache(Stub, MCE.getCurrentPCValue()-(intptr_t)Stub);
584 DOUT << "JIT: Stub emitted at [" << Stub
585 << "] for external function at '" << FnAddr << "'\n";
589 unsigned JITResolver::getGOTIndexForAddr(void* addr) {
590 unsigned idx = revGOTMap[addr];
592 idx = ++nextGOTIndex;
593 revGOTMap[addr] = idx;
594 DOUT << "Adding GOT entry " << idx
595 << " for addr " << addr << "\n";
596 // ((void**)MemMgr.getGOTBase())[idx] = addr;
601 /// JITCompilerFn - This function is called when a lazy compilation stub has
602 /// been entered. It looks up which function this stub corresponds to, compiles
603 /// it if necessary, then returns the resultant function pointer.
604 void *JITResolver::JITCompilerFn(void *Stub) {
605 JITResolver &JR = getJITResolver();
607 MutexGuard locked(TheJIT->lock);
609 // The address given to us for the stub may not be exactly right, it might be
610 // a little bit after the stub. As such, use upper_bound to find it.
611 std::map<void*, Function*>::iterator I =
612 JR.state.getStubToFunctionMap(locked).upper_bound(Stub);
613 assert(I != JR.state.getStubToFunctionMap(locked).begin() &&
614 "This is not a known stub!");
615 Function *F = (--I)->second;
617 // If disabled, emit a useful error message and abort.
618 if (TheJIT->isLazyCompilationDisabled()) {
619 cerr << "LLVM JIT requested to do lazy compilation of function '"
620 << F->getName() << "' when lazy compiles are disabled!\n";
624 // We might like to remove the stub from the StubToFunction map.
625 // We can't do that! Multiple threads could be stuck, waiting to acquire the
626 // lock above. As soon as the 1st function finishes compiling the function,
627 // the next one will be released, and needs to be able to find the function it
629 //JR.state.getStubToFunctionMap(locked).erase(I);
631 DOUT << "JIT: Lazily resolving function '" << F->getName()
632 << "' In stub ptr = " << Stub << " actual ptr = "
635 void *Result = TheJIT->getPointerToFunction(F);
637 // We don't need to reuse this stub in the future, as F is now compiled.
638 JR.state.getFunctionToStubMap(locked).erase(F);
640 // FIXME: We could rewrite all references to this stub if we knew them.
642 // What we will do is set the compiled function address to map to the
643 // same GOT entry as the stub so that later clients may update the GOT
644 // if they see it still using the stub address.
645 // Note: this is done so the Resolver doesn't have to manage GOT memory
646 // Do this without allocating map space if the target isn't using a GOT
647 if(JR.revGOTMap.find(Stub) != JR.revGOTMap.end())
648 JR.revGOTMap[Result] = JR.revGOTMap[Stub];
654 //===----------------------------------------------------------------------===//
658 /// JITEmitter - The JIT implementation of the MachineCodeEmitter, which is
659 /// used to output functions to memory for execution.
660 class JITEmitter : public MachineCodeEmitter {
661 JITMemoryManager MemMgr;
663 // When outputting a function stub in the context of some other function, we
664 // save BufferBegin/BufferEnd/CurBufferPtr here.
665 unsigned char *SavedBufferBegin, *SavedBufferEnd, *SavedCurBufferPtr;
667 /// Relocations - These are the relocations that the function needs, as
669 std::vector<MachineRelocation> Relocations;
671 /// MBBLocations - This vector is a mapping from MBB ID's to their address.
672 /// It is filled in by the StartMachineBasicBlock callback and queried by
673 /// the getMachineBasicBlockAddress callback.
674 std::vector<intptr_t> MBBLocations;
676 /// ConstantPool - The constant pool for the current function.
678 MachineConstantPool *ConstantPool;
680 /// ConstantPoolBase - A pointer to the first entry in the constant pool.
682 void *ConstantPoolBase;
684 /// JumpTable - The jump tables for the current function.
686 MachineJumpTableInfo *JumpTable;
688 /// JumpTableBase - A pointer to the first entry in the jump table.
692 JITEmitter(JIT &jit) : MemMgr(jit.getJITInfo().needsGOT()) {
694 if (MemMgr.isManagingGOT()) DOUT << "JIT is managing a GOT\n";
697 virtual void startFunction(MachineFunction &F);
698 virtual bool finishFunction(MachineFunction &F);
700 void emitConstantPool(MachineConstantPool *MCP);
701 void initJumpTableInfo(MachineJumpTableInfo *MJTI);
702 void emitJumpTableInfo(MachineJumpTableInfo *MJTI);
704 virtual void startFunctionStub(unsigned StubSize, unsigned Alignment = 1);
705 virtual void* finishFunctionStub(const Function *F);
707 virtual void addRelocation(const MachineRelocation &MR) {
708 Relocations.push_back(MR);
711 virtual void StartMachineBasicBlock(MachineBasicBlock *MBB) {
712 if (MBBLocations.size() <= (unsigned)MBB->getNumber())
713 MBBLocations.resize((MBB->getNumber()+1)*2);
714 MBBLocations[MBB->getNumber()] = getCurrentPCValue();
717 virtual intptr_t getConstantPoolEntryAddress(unsigned Entry) const;
718 virtual intptr_t getJumpTableEntryAddress(unsigned Entry) const;
720 virtual intptr_t getMachineBasicBlockAddress(MachineBasicBlock *MBB) const {
721 assert(MBBLocations.size() > (unsigned)MBB->getNumber() &&
722 MBBLocations[MBB->getNumber()] && "MBB not emitted!");
723 return MBBLocations[MBB->getNumber()];
726 /// deallocateMemForFunction - Deallocate all memory for the specified
728 void deallocateMemForFunction(Function *F) {
729 MemMgr.deallocateMemForFunction(F);
732 void *getPointerToGlobal(GlobalValue *GV, void *Reference, bool NoNeedStub);
736 void *JITEmitter::getPointerToGlobal(GlobalValue *V, void *Reference,
737 bool DoesntNeedStub) {
738 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(V)) {
739 /// FIXME: If we straightened things out, this could actually emit the
740 /// global immediately instead of queuing it for codegen later!
741 return TheJIT->getOrEmitGlobalVariable(GV);
744 // If we have already compiled the function, return a pointer to its body.
745 Function *F = cast<Function>(V);
746 void *ResultPtr = TheJIT->getPointerToGlobalIfAvailable(F);
747 if (ResultPtr) return ResultPtr;
749 if (F->isExternal() && !F->hasNotBeenReadFromBytecode()) {
750 // If this is an external function pointer, we can force the JIT to
751 // 'compile' it, which really just adds it to the map.
753 return TheJIT->getPointerToFunction(F);
755 return getJITResolver(this).getFunctionStub(F);
758 // Okay, the function has not been compiled yet, if the target callback
759 // mechanism is capable of rewriting the instruction directly, prefer to do
760 // that instead of emitting a stub.
762 return getJITResolver(this).AddCallbackAtLocation(F, Reference);
764 // Otherwise, we have to emit a lazy resolving stub.
765 return getJITResolver(this).getFunctionStub(F);
768 void JITEmitter::startFunction(MachineFunction &F) {
769 uintptr_t ActualSize;
770 BufferBegin = CurBufferPtr = MemMgr.startFunctionBody(ActualSize);
771 BufferEnd = BufferBegin+ActualSize;
773 // Ensure the constant pool/jump table info is at least 4-byte aligned.
776 emitConstantPool(F.getConstantPool());
777 initJumpTableInfo(F.getJumpTableInfo());
779 // About to start emitting the machine code for the function.
780 emitAlignment(std::max(F.getFunction()->getAlignment(), 8U));
781 TheJIT->updateGlobalMapping(F.getFunction(), CurBufferPtr);
783 MBBLocations.clear();
786 bool JITEmitter::finishFunction(MachineFunction &F) {
787 if (CurBufferPtr == BufferEnd) {
788 // FIXME: Allocate more space, then try again.
789 cerr << "JIT: Ran out of space for generated machine code!\n";
793 emitJumpTableInfo(F.getJumpTableInfo());
795 // FnStart is the start of the text, not the start of the constant pool and
796 // other per-function data.
797 unsigned char *FnStart =
798 (unsigned char *)TheJIT->getPointerToGlobalIfAvailable(F.getFunction());
799 unsigned char *FnEnd = CurBufferPtr;
801 MemMgr.endFunctionBody(F.getFunction(), BufferBegin, FnEnd);
802 NumBytes += FnEnd-FnStart;
804 if (!Relocations.empty()) {
805 NumRelos += Relocations.size();
807 // Resolve the relocations to concrete pointers.
808 for (unsigned i = 0, e = Relocations.size(); i != e; ++i) {
809 MachineRelocation &MR = Relocations[i];
812 ResultPtr = TheJIT->getPointerToNamedFunction(MR.getString());
814 // If the target REALLY wants a stub for this function, emit it now.
815 if (!MR.doesntNeedFunctionStub())
816 ResultPtr = getJITResolver(this).getExternalFunctionStub(ResultPtr);
817 } else if (MR.isGlobalValue()) {
818 ResultPtr = getPointerToGlobal(MR.getGlobalValue(),
819 BufferBegin+MR.getMachineCodeOffset(),
820 MR.doesntNeedFunctionStub());
821 } else if (MR.isBasicBlock()) {
822 ResultPtr = (void*)getMachineBasicBlockAddress(MR.getBasicBlock());
823 } else if (MR.isConstantPoolIndex()) {
824 ResultPtr=(void*)getConstantPoolEntryAddress(MR.getConstantPoolIndex());
826 assert(MR.isJumpTableIndex());
827 ResultPtr=(void*)getJumpTableEntryAddress(MR.getJumpTableIndex());
830 MR.setResultPointer(ResultPtr);
832 // if we are managing the GOT and the relocation wants an index,
834 if (MemMgr.isManagingGOT() && MR.isGOTRelative()) {
835 unsigned idx = getJITResolver(this).getGOTIndexForAddr(ResultPtr);
837 if (((void**)MemMgr.getGOTBase())[idx] != ResultPtr) {
838 DOUT << "GOT was out of date for " << ResultPtr
839 << " pointing at " << ((void**)MemMgr.getGOTBase())[idx]
841 ((void**)MemMgr.getGOTBase())[idx] = ResultPtr;
846 TheJIT->getJITInfo().relocate(BufferBegin, &Relocations[0],
847 Relocations.size(), MemMgr.getGOTBase());
850 // Update the GOT entry for F to point to the new code.
851 if (MemMgr.isManagingGOT()) {
852 unsigned idx = getJITResolver(this).getGOTIndexForAddr((void*)BufferBegin);
853 if (((void**)MemMgr.getGOTBase())[idx] != (void*)BufferBegin) {
854 DOUT << "GOT was out of date for " << (void*)BufferBegin
855 << " pointing at " << ((void**)MemMgr.getGOTBase())[idx] << "\n";
856 ((void**)MemMgr.getGOTBase())[idx] = (void*)BufferBegin;
860 // Invalidate the icache if necessary.
861 synchronizeICache(FnStart, FnEnd-FnStart);
863 DOUT << "JIT: Finished CodeGen of [" << (void*)FnStart
864 << "] Function: " << F.getFunction()->getName()
865 << ": " << (FnEnd-FnStart) << " bytes of text, "
866 << Relocations.size() << " relocations\n";
870 DOUT << "Disassembled code:\n"
871 << sys::disassembleBuffer(FnStart, FnEnd-FnStart, (uintptr_t)FnStart);
877 void JITEmitter::emitConstantPool(MachineConstantPool *MCP) {
878 const std::vector<MachineConstantPoolEntry> &Constants = MCP->getConstants();
879 if (Constants.empty()) return;
881 MachineConstantPoolEntry CPE = Constants.back();
882 unsigned Size = CPE.Offset;
883 const Type *Ty = CPE.isMachineConstantPoolEntry()
884 ? CPE.Val.MachineCPVal->getType() : CPE.Val.ConstVal->getType();
885 Size += TheJIT->getTargetData()->getTypeSize(Ty);
887 ConstantPoolBase = allocateSpace(Size, 1 << MCP->getConstantPoolAlignment());
890 if (ConstantPoolBase == 0) return; // Buffer overflow.
892 // Initialize the memory for all of the constant pool entries.
893 for (unsigned i = 0, e = Constants.size(); i != e; ++i) {
894 void *CAddr = (char*)ConstantPoolBase+Constants[i].Offset;
895 if (Constants[i].isMachineConstantPoolEntry()) {
896 // FIXME: add support to lower machine constant pool values into bytes!
897 cerr << "Initialize memory with machine specific constant pool entry"
898 << " has not been implemented!\n";
901 TheJIT->InitializeMemory(Constants[i].Val.ConstVal, CAddr);
905 void JITEmitter::initJumpTableInfo(MachineJumpTableInfo *MJTI) {
906 const std::vector<MachineJumpTableEntry> &JT = MJTI->getJumpTables();
907 if (JT.empty()) return;
909 unsigned NumEntries = 0;
910 for (unsigned i = 0, e = JT.size(); i != e; ++i)
911 NumEntries += JT[i].MBBs.size();
913 unsigned EntrySize = MJTI->getEntrySize();
915 // Just allocate space for all the jump tables now. We will fix up the actual
916 // MBB entries in the tables after we emit the code for each block, since then
917 // we will know the final locations of the MBBs in memory.
919 JumpTableBase = allocateSpace(NumEntries * EntrySize, MJTI->getAlignment());
922 void JITEmitter::emitJumpTableInfo(MachineJumpTableInfo *MJTI) {
923 const std::vector<MachineJumpTableEntry> &JT = MJTI->getJumpTables();
924 if (JT.empty() || JumpTableBase == 0) return;
926 if (TargetMachine::getRelocationModel() == Reloc::PIC_) {
927 assert(MJTI->getEntrySize() == 4 && "Cross JIT'ing?");
928 // For each jump table, place the offset from the beginning of the table
929 // to the target address.
930 int *SlotPtr = (int*)JumpTableBase;
932 for (unsigned i = 0, e = JT.size(); i != e; ++i) {
933 const std::vector<MachineBasicBlock*> &MBBs = JT[i].MBBs;
934 // Store the offset of the basic block for this jump table slot in the
935 // memory we allocated for the jump table in 'initJumpTableInfo'
936 intptr_t Base = (intptr_t)SlotPtr;
937 for (unsigned mi = 0, me = MBBs.size(); mi != me; ++mi)
938 *SlotPtr++ = (intptr_t)getMachineBasicBlockAddress(MBBs[mi]) - Base;
941 assert(MJTI->getEntrySize() == sizeof(void*) && "Cross JIT'ing?");
943 // For each jump table, map each target in the jump table to the address of
944 // an emitted MachineBasicBlock.
945 intptr_t *SlotPtr = (intptr_t*)JumpTableBase;
947 for (unsigned i = 0, e = JT.size(); i != e; ++i) {
948 const std::vector<MachineBasicBlock*> &MBBs = JT[i].MBBs;
949 // Store the address of the basic block for this jump table slot in the
950 // memory we allocated for the jump table in 'initJumpTableInfo'
951 for (unsigned mi = 0, me = MBBs.size(); mi != me; ++mi)
952 *SlotPtr++ = getMachineBasicBlockAddress(MBBs[mi]);
957 void JITEmitter::startFunctionStub(unsigned StubSize, unsigned Alignment) {
958 SavedBufferBegin = BufferBegin;
959 SavedBufferEnd = BufferEnd;
960 SavedCurBufferPtr = CurBufferPtr;
962 BufferBegin = CurBufferPtr = MemMgr.allocateStub(StubSize, Alignment);
963 BufferEnd = BufferBegin+StubSize+1;
966 void *JITEmitter::finishFunctionStub(const Function *F) {
967 NumBytes += getCurrentPCOffset();
968 std::swap(SavedBufferBegin, BufferBegin);
969 BufferEnd = SavedBufferEnd;
970 CurBufferPtr = SavedCurBufferPtr;
971 return SavedBufferBegin;
974 // getConstantPoolEntryAddress - Return the address of the 'ConstantNum' entry
975 // in the constant pool that was last emitted with the 'emitConstantPool'
978 intptr_t JITEmitter::getConstantPoolEntryAddress(unsigned ConstantNum) const {
979 assert(ConstantNum < ConstantPool->getConstants().size() &&
980 "Invalid ConstantPoolIndex!");
981 return (intptr_t)ConstantPoolBase +
982 ConstantPool->getConstants()[ConstantNum].Offset;
985 // getJumpTableEntryAddress - Return the address of the JumpTable with index
986 // 'Index' in the jumpp table that was last initialized with 'initJumpTableInfo'
988 intptr_t JITEmitter::getJumpTableEntryAddress(unsigned Index) const {
989 const std::vector<MachineJumpTableEntry> &JT = JumpTable->getJumpTables();
990 assert(Index < JT.size() && "Invalid jump table index!");
993 unsigned EntrySize = JumpTable->getEntrySize();
995 for (unsigned i = 0; i < Index; ++i)
996 Offset += JT[i].MBBs.size();
1000 return (intptr_t)((char *)JumpTableBase + Offset);
1003 //===----------------------------------------------------------------------===//
1004 // Public interface to this file
1005 //===----------------------------------------------------------------------===//
1007 MachineCodeEmitter *JIT::createEmitter(JIT &jit) {
1008 return new JITEmitter(jit);
1011 // getPointerToNamedFunction - This function is used as a global wrapper to
1012 // JIT::getPointerToNamedFunction for the purpose of resolving symbols when
1013 // bugpoint is debugging the JIT. In that scenario, we are loading an .so and
1014 // need to resolve function(s) that are being mis-codegenerated, so we need to
1015 // resolve their addresses at runtime, and this is the way to do it.
1017 void *getPointerToNamedFunction(const char *Name) {
1018 if (Function *F = TheJIT->FindFunctionNamed(Name))
1019 return TheJIT->getPointerToFunction(F);
1020 return TheJIT->getPointerToNamedFunction(Name);
1024 // getPointerToFunctionOrStub - If the specified function has been
1025 // code-gen'd, return a pointer to the function. If not, compile it, or use
1026 // a stub to implement lazy compilation if available.
1028 void *JIT::getPointerToFunctionOrStub(Function *F) {
1029 // If we have already code generated the function, just return the address.
1030 if (void *Addr = getPointerToGlobalIfAvailable(F))
1033 // Get a stub if the target supports it
1034 return getJITResolver(MCE).getFunctionStub(F);
1037 /// freeMachineCodeForFunction - release machine code memory for given Function.
1039 void JIT::freeMachineCodeForFunction(Function *F) {
1040 // Delete translation for this from the ExecutionEngine, so it will get
1041 // retranslated next time it is used.
1042 updateGlobalMapping(F, 0);
1044 // Free the actual memory for the function body and related stuff.
1045 assert(dynamic_cast<JITEmitter*>(MCE) && "Unexpected MCE?");
1046 dynamic_cast<JITEmitter*>(MCE)->deallocateMemForFunction(F);