bjit: free codeblocks after tiering up to the LLVM jit

we free now the code blocks after a recompile in the LLVM tier (except for OSR frames) because
it is likely that we will not use the code anymore.
- we have to make sure we are not currently executing any code we will delete that's why I added bjit_num_inside
- there were some cases where we forgot to deregister stuff
- when profiling we don't actually unmap the code in order to not brake profiling

src/asm_writing/icinfo.cpp

@@ -326,6 +326,7 @@ ICSlotInfo* ICInfo::pickEntryForRewrite(const char* debug_name) {
 }
 
 static llvm::DenseMap<void*, ICInfo*> ics_by_return_addr;
+static llvm::DenseMap<AST*, ICInfo*> ics_by_ast_node;
 
 ICInfo::ICInfo(void* start_addr, void* slowpath_rtn_addr, void* continue_addr, StackInfo stack_info, int size,
                llvm::CallingConv::ID calling_conv, LiveOutSet _live_outs, assembler::GenericRegister return_register,
@@ -342,6 +343,7 @@ ICInfo::ICInfo(void* start_addr, void* slowpath_rtn_addr, void* continue_addr, S
       times_rewritten(0),
       allocatable_registers(allocatable_registers),
       ic_global_decref_locations(std::move(ic_global_decref_locations)),
+      node(NULL),
       start_addr(start_addr),
       slowpath_rtn_addr(slowpath_rtn_addr),
       continue_addr(continue_addr) {
@@ -353,6 +355,8 @@ ICInfo::ICInfo(void* start_addr, void* slowpath_rtn_addr, void* continue_addr, S
 ICInfo::~ICInfo() {
     // if this ICInfo got created with registerCompiledPatchpoint we have to unregister this
     ics_by_return_addr.erase(slowpath_rtn_addr);
+    if (node)
+        ics_by_ast_node.erase(node);
     deregisterGCTrackedICInfo(this);
 
     for (auto& slot : slots) {
@@ -469,8 +473,6 @@ bool ICInfo::isMegamorphic() {
     return times_rewritten >= IC_MEGAMORPHIC_THRESHOLD;
 }
 
-static llvm::DenseMap<AST*, ICInfo*> ics_by_ast_node;
-
 ICInfo* ICInfo::getICInfoForNode(AST* node) {
     auto&& it = ics_by_ast_node.find(node);
     if (it != ics_by_ast_node.end())
@@ -478,6 +480,8 @@ ICInfo* ICInfo::getICInfoForNode(AST* node) {
     return NULL;
 }
 void ICInfo::associateNodeWithICInfo(AST* node) {
+    assert(!this->node);
+    this->node = node;
     ics_by_ast_node[node] = this;
 }
 void ICInfo::appendDecrefInfosTo(std::vector<DecrefInfo>& dest_decref_infos) {

src/asm_writing/icinfo.h

@@ -102,6 +102,9 @@ private:
     // global ones.
     std::vector<Location> ic_global_decref_locations;
 
+    // associated AST node for this IC
+    AST* node;
+
     // for ICSlotRewrite:
     ICSlotInfo* pickEntryForRewrite(const char* debug_name);
 

src/codegen/ast_interpreter.cpp

@@ -306,7 +306,7 @@ void ASTInterpreter::startJITing(CFGBlock* block, int exit_offset, llvm::DenseSe
         code_block = code_blocks[code_blocks.size() - 1].get();
 
     if (!code_block || code_block->shouldCreateNewBlock()) {
-        code_blocks.push_back(llvm::make_unique<JitCodeBlock>(source_info->getName()->s()));
+        code_blocks.push_back(llvm::make_unique<JitCodeBlock>(getMD(), source_info->getName()->s()));
         code_block = code_blocks[code_blocks.size() - 1].get();
         exit_offset = 0;
     }
@@ -354,12 +354,16 @@ void ASTInterpreter::finishJITing(CFGBlock* continue_block) {
 }
 
 Box* ASTInterpreter::execJITedBlock(CFGBlock* b) {
+    auto& num_inside = getMD()->bjit_num_inside;
     try {
         UNAVOIDABLE_STAT_TIMER(t0, "us_timer_in_baseline_jitted_code");
+        ++num_inside;
         std::pair<CFGBlock*, Box*> rtn = b->entry_code(this, b, vregs);
+        --num_inside;
         next_block = rtn.first;
         return rtn.second;
     } catch (ExcInfo e) {
+        --num_inside;
         AST_stmt* stmt = getCurrentStatement();
         if (stmt->type != AST_TYPE::Invoke)
             throw e;

src/codegen/baseline_jit.cpp

@@ -73,14 +73,19 @@ JitCodeBlock::MemoryManager::MemoryManager() {
 }
 
 JitCodeBlock::MemoryManager::~MemoryManager() {
+    // unfortunately we can't free the memory when profiling otherwise we would reuse the same addresses which makes
+    // profiling impossible
+    if (!PROFILE)
         munmap(addr, JitCodeBlock::memory_size);
     addr = NULL;
-
-    RELEASE_ASSERT(0, "we have to unregister this block from g.func_addr_registry");
 }
 
-JitCodeBlock::JitCodeBlock(llvm::StringRef name)
-    : entry_offset(0), a(memory.get() + sizeof(eh_info), code_size), is_currently_writing(false), asm_failed(false) {
+JitCodeBlock::JitCodeBlock(FunctionMetadata* md, llvm::StringRef name)
+    : md(md),
+      entry_offset(0),
+      a(memory.get() + sizeof(eh_info), code_size),
+      is_currently_writing(false),
+      asm_failed(false) {
     static StatCounter num_jit_code_blocks("num_baselinejit_code_blocks");
     num_jit_code_blocks.log();
     static StatCounter num_jit_total_bytes("num_baselinejit_total_bytes");
@@ -116,6 +121,17 @@ JitCodeBlock::JitCodeBlock(llvm::StringRef name)
     g.func_addr_registry.registerFunction(unique_name, code, code_size, NULL);
 }
 
+JitCodeBlock::~JitCodeBlock() {
+    // we should not deregister the function in profiling mode because otherwise the profiler can't show it
+    if (!PROFILE)
+        g.func_addr_registry.deregisterFunction(a.getStartAddr());
+    register_eh_info.deregisterFrame();
+
+    for (auto&& block : md->source->cfg->blocks) {
+        block_patch_locations.erase(block);
+    }
+}
+
 std::unique_ptr<JitFragmentWriter> JitCodeBlock::newFragment(CFGBlock* block, int patch_jump_offset,
                                                              llvm::DenseSet<int> known_non_null_vregs) {
     if (is_currently_writing || blocks_aborted.count(block))
@@ -137,7 +153,7 @@ std::unique_ptr<JitFragmentWriter> JitCodeBlock::newFragment(CFGBlock* block, in
     std::unique_ptr<ICSlotRewrite> rewrite = ic_info->startRewrite("");
 
     return std::unique_ptr<JitFragmentWriter>(
-        new JitFragmentWriter(block, std::move(ic_info), std::move(rewrite), fragment_offset, patch_jump_offset,
+        new JitFragmentWriter(md, block, std::move(ic_info), std::move(rewrite), fragment_offset, patch_jump_offset,
                               a.getStartAddr(), *this, std::move(known_non_null_vregs)));
 }
 
@@ -147,7 +163,7 @@ void JitCodeBlock::fragmentAbort(bool not_enough_space) {
 }
 
 void JitCodeBlock::fragmentFinished(int bytes_written, int num_bytes_overlapping, void* next_fragment_start,
-                                    ICInfo& ic_info) {
+                                    std::vector<std::unique_ptr<ICInfo>>&& pp_ic_infos, ICInfo& ic_info) {
     assert(next_fragment_start == bytes_written + a.curInstPointer() - num_bytes_overlapping);
     a.setCurInstPointer((uint8_t*)next_fragment_start);
 
@@ -155,14 +171,17 @@ void JitCodeBlock::fragmentFinished(int bytes_written, int num_bytes_overlapping
     is_currently_writing = false;
 
     ic_info.appendDecrefInfosTo(decref_infos);
+    std::move(std::begin(pp_ic_infos), std::end(pp_ic_infos), std::back_inserter(this->pp_ic_infos));
+    pp_ic_infos.clear();
 }
 
-JitFragmentWriter::JitFragmentWriter(CFGBlock* block, std::unique_ptr<ICInfo> ic_info,
+JitFragmentWriter::JitFragmentWriter(FunctionMetadata* md, CFGBlock* block, std::unique_ptr<ICInfo> ic_info,
                                      std::unique_ptr<ICSlotRewrite> rewrite, int code_offset, int num_bytes_overlapping,
                                      void* entry_code, JitCodeBlock& code_block,
                                      llvm::DenseSet<int> known_non_null_vregs)
     : ICInfoManager(std::move(ic_info)),
       Rewriter(std::move(rewrite), 0, {}, /* needs_invalidation_support = */ false),
+      md(md),
       block(block),
       code_offset(code_offset),
       exit_info(),
@@ -185,6 +204,13 @@ JitFragmentWriter::JitFragmentWriter(CFGBlock* block, std::unique_ptr<ICInfo> ic
     addAction([=]() { vregs_array->bumpUse(); }, vregs_array, ActionType::NORMAL);
     if (LOG_BJIT_ASSEMBLY)
         comment("BJIT: JitFragmentWriter() end");
+
+    // this makes sure that we can't delete the code blocks while we are inside JitFragmentWriter
+    ++md->bjit_num_inside;
+}
+
+JitFragmentWriter::~JitFragmentWriter() {
+    --md->bjit_num_inside;
 }
 
 RewriterVar* JitFragmentWriter::getInterp() {
@@ -823,6 +849,7 @@ std::pair<int, llvm::DenseSet<int>> JitFragmentWriter::finishCompilation() {
         block_patch_locations[side_exit_patch_location.first].push_back(patch_location);
     }
 
+    std::vector<std::unique_ptr<ICInfo>> ic_infos;
     for (auto&& pp_info : pp_infos) {
         SpillMap _spill_map;
         uint8_t* start_addr = pp_info.start_addr;
@@ -837,7 +864,7 @@ std::pair<int, llvm::DenseSet<int>> JitFragmentWriter::finishCompilation() {
             start_addr, slowpath_start, initialization_info.continue_addr, slowpath_rtn_addr, pp_info.ic.get(),
             pp_info.stack_info, LiveOutSet(), std::move(pp_info.decref_infos));
         pp->associateNodeWithICInfo(pp_info.node);
-        pp.release();
+        ic_infos.push_back(std::move(pp));
     }
 
 #ifndef NDEBUG
@@ -852,7 +879,8 @@ std::pair<int, llvm::DenseSet<int>> JitFragmentWriter::finishCompilation() {
         ASSERT(assembler->curInstPointer() == (uint8_t*)exit_info.exit_start + exit_info.num_bytes,
                "Error! wrote more bytes out after the 'retq' that we thought was going to be the end of the assembly.  "
                "We will end up overwriting those instructions.");
-    code_block.fragmentFinished(assembler->bytesWritten(), num_bytes_overlapping, next_fragment_start, *ic_info);
+    code_block.fragmentFinished(assembler->bytesWritten(), num_bytes_overlapping, next_fragment_start,
+                                std::move(ic_infos), *ic_info);
 
 #if MOVING_GC
     // If JitFragmentWriter is destroyed, we don't necessarily want the ICInfo to be destroyed also,

src/codegen/baseline_jit.h

@@ -168,6 +168,7 @@ private:
         uint8_t* get() { return addr; }
     };
 
+    FunctionMetadata* md;
     // the memory block contains the EH frame directly followed by the generated machine code.
     MemoryManager memory;
     int entry_offset;
@@ -178,15 +179,19 @@ private:
     // this allows us to deregister them when we release the code
     std::vector<DecrefInfo> decref_infos;
     RegisterEHFrame register_eh_info;
+    std::vector<std::unique_ptr<ICInfo>> pp_ic_infos;
+
 
 public:
-    JitCodeBlock(llvm::StringRef name);
+    JitCodeBlock(FunctionMetadata* md, llvm::StringRef name);
+    ~JitCodeBlock();
 
     std::unique_ptr<JitFragmentWriter> newFragment(CFGBlock* block, int patch_jump_offset,
                                                    llvm::DenseSet<int> known_non_null_vregs);
     bool shouldCreateNewBlock() const { return asm_failed || a.bytesLeft() < 128; }
     void fragmentAbort(bool not_enough_space);
-    void fragmentFinished(int bytes_witten, int num_bytes_overlapping, void* next_fragment_start, ICInfo& ic_info);
+    void fragmentFinished(int bytes_witten, int num_bytes_overlapping, void* next_fragment_start,
+                          std::vector<std::unique_ptr<ICInfo>>&& pp_ic_infos, ICInfo& ic_info);
 };
 
 // Hold the ICInfo of the JitFragmentWriter in a separate class from which JitFragmentWriter derives.
@@ -209,6 +214,7 @@ private:
 
     static constexpr int min_patch_size = 13;
 
+    FunctionMetadata* md;
     CFGBlock* block;
     int code_offset; // offset inside the JitCodeBlock to the start of this block
 
@@ -250,9 +256,10 @@ private:
     llvm::SmallVector<PPInfo, 8> pp_infos;
 
 public:
-    JitFragmentWriter(CFGBlock* block, std::unique_ptr<ICInfo> ic_info, std::unique_ptr<ICSlotRewrite> rewrite,
-                      int code_offset, int num_bytes_overlapping, void* entry_code, JitCodeBlock& code_block,
-                      llvm::DenseSet<int> known_non_null_vregs);
+    JitFragmentWriter(FunctionMetadata* md, CFGBlock* block, std::unique_ptr<ICInfo> ic_info,
+                      std::unique_ptr<ICSlotRewrite> rewrite, int code_offset, int num_bytes_overlapping,
+                      void* entry_code, JitCodeBlock& code_block, llvm::DenseSet<int> known_non_null_vregs);
+    ~JitFragmentWriter();
 
     RewriterVar* getInterp();
     RewriterVar* imm(uint64_t val);

src/codegen/codegen.h

@@ -50,6 +50,7 @@ public:
     std::string getFuncNameAtAddress(void* addr, bool demangle, bool* out_success = NULL);
     llvm::Function* getLLVMFuncAtAddress(void* addr);
     void registerFunction(const std::string& name, void* addr, int length, llvm::Function* llvm_func);
+    void deregisterFunction(void* addr) { functions.erase(addr); }
     void dumpPerfMap();
 };
 

src/codegen/irgen/hooks.cpp

@@ -331,6 +331,10 @@ CompiledFunction* compileFunction(FunctionMetadata* f, FunctionSpecialization* s
             RELEASE_ASSERT(0, "%d", effort);
     }
 
+    // free the bjit code if this is not a OSR compilation
+    if (!entry_descriptor)
+        f->tryDeallocatingTheBJitCode();
+
     return cf;
 }
 
@@ -841,4 +845,24 @@ void FunctionMetadata::addVersion(void* f, ConcreteCompilerType* rtn_type,
     FunctionSpecialization* spec = new FunctionSpecialization(processType(rtn_type), arg_types);
     addVersion(new CompiledFunction(this, spec, f, EffortLevel::MAXIMAL, exception_style, NULL));
 }
+
+bool FunctionMetadata::tryDeallocatingTheBJitCode() {
+    // we can only delete the code object if we are not executing it currently
+    assert(bjit_num_inside >= 0);
+    if (bjit_num_inside != 0) {
+        // TODO: we could check later on again
+        static StatCounter num_baselinejit_blocks_failed_to_free("num_baselinejit_code_blocks_cant_free");
+        num_baselinejit_blocks_failed_to_free.log(code_blocks.size());
+        return false;
+    }
+
+    static StatCounter num_baselinejit_blocks_freed("num_baselinejit_code_blocks_freed");
+    num_baselinejit_blocks_freed.log(code_blocks.size());
+    code_blocks.clear();
+    for (CFGBlock* block : source->cfg->blocks) {
+        block->code = NULL;
+        block->entry_code = NULL;
+    }
+    return true;
+}
 }

src/core/cfg.h

@@ -67,7 +67,7 @@ public:
 
 class CFGBlock {
 public:
-    CFG* cfg;
+    CFG* const cfg;
 
     // Baseline JIT helper fields:
     // contains address to the start of the code of this basic block

src/core/types.h

@@ -477,6 +477,7 @@ public:
 
     // For use by the interpreter/baseline jit:
     int times_interpreted;
+    long bjit_num_inside = 0;
     std::vector<std::unique_ptr<JitCodeBlock>> code_blocks;
     ICInvalidator dependent_interp_callsites;
 
@@ -530,6 +531,9 @@ public:
                                     ExceptionStyle exception_style = CXX) {
         return create(f, rtn_type, nargs, false, false, param_names, exception_style);
     }
+
+    // tries to free the bjit allocated code. returns true on success
+    bool tryDeallocatingTheBJitCode();
 };