Make global/static variables thread_local; use mmap() for managing me…

…mory (#1161) Memory allocation is simplified by using mmap() to allocate a single 1TB block of addresses and relying on demand paging. Global and static variables are made thread_local and setThreadLocal(true) is used on such variables in the code generator. --------- Co-authored-by: Steven Eker <[email protected]> Co-authored-by: Dwight Guth <[email protected]>
runtimeverification · Nov 26, 2024 · 0f72ec9 · 0f72ec9
1 parent f18f71c
commit 0f72ec9
Show file tree

Hide file tree

Showing 13 changed files with 154 additions and 59 deletions.
diff --git a/config/llvm_header.inc b/config/llvm_header.inc
@@ -202,7 +202,7 @@ declare void @write_configuration_to_proof_trace(ptr, ptr, i1)
 @current_interval = thread_local global i64 0
 @GC_THRESHOLD = thread_local global i64 @GC_THRESHOLD@
 
-@gc_roots = global [256 x ptr] zeroinitializer
+@gc_roots = thread_local global [256 x ptr] zeroinitializer
 
 define i64 @get_gc_threshold() {
   %threshold = load i64, ptr @GC_THRESHOLD

diff --git a/include/runtime/arena.h b/include/runtime/arena.h
@@ -30,12 +30,19 @@ using memory_block_header = struct {
 // Macro to define a new arena with the given ID. Supports IDs ranging from 0 to
 // 127.
 #define REGISTER_ARENA(name, id)                                               \
-  static struct arena name = {.allocation_semispace_id = (id)}
+  static thread_local struct arena name = {.allocation_semispace_id = (id)}
 
 #define MEM_BLOCK_START(ptr)                                                   \
   ((char *)(((uintptr_t)(ptr)-1) & ~(BLOCK_SIZE - 1)))
 
+#ifdef __MACH__
+//
+//	thread_local disabled for Apple
+//
 extern bool time_for_collection;
+#else
+extern thread_local bool time_for_collection;
+#endif
 
 size_t get_gc_threshold();
 

diff --git a/include/runtime/collect.h b/include/runtime/collect.h
@@ -26,8 +26,8 @@ using set_node = set::iterator::node_t;
 using set_impl = set::iterator::tree_t;
 
 extern "C" {
-extern size_t numBytesLiveAtCollection[1 << AGE_WIDTH];
-extern bool collect_old;
+extern thread_local size_t numBytesLiveAtCollection[1 << AGE_WIDTH];
+extern thread_local bool collect_old;
 size_t get_size(uint64_t, uint16_t);
 void migrate_static_roots(void);
 void migrate(block **block_ptr);

diff --git a/include/runtime/header.h b/include/runtime/header.h
@@ -47,8 +47,14 @@ size_t hash_k(block *);
 void k_hash(block *, void *);
 bool hash_enter(void);
 void hash_exit(void);
-
+#ifdef __MACH__
+//
+//	thread_local disabled for Apple
+//
 extern bool gc_enabled;
+#else
+extern thread_local bool gc_enabled;
+#endif
 }
 
 class k_elem {

diff --git a/lib/codegen/CreateTerm.cpp b/lib/codegen/CreateTerm.cpp
@@ -782,18 +782,47 @@ llvm::Value *create_term::disable_gc() {
   llvm::Constant *global
       = module_->getOrInsertGlobal("gc_enabled", llvm::Type::getInt1Ty(ctx_));
   auto *global_var = llvm::cast<llvm::GlobalVariable>(global);
+#ifdef __MACH__
+  //
+  //	thread_local disabled for Apple
+  //
+  /*
+  global_var->setThreadLocal(true);
+  llvm::IRBuilder b(current_block_);
+  auto *global_var_address = b.CreateThreadLocalAddress(global_var);
+  */
+  auto *global_var_address = global_var;
+#else
+  global_var->setThreadLocal(true);
+  auto *global_var_address = global_var;
+#endif
   auto *old_val = new llvm::LoadInst(
-      llvm::Type::getInt1Ty(ctx_), global_var, "was_enabled", current_block_);
+      llvm::Type::getInt1Ty(ctx_), global_var_address, "was_enabled",
+      current_block_);
   new llvm::StoreInst(
-      llvm::ConstantInt::getFalse(ctx_), global_var, current_block_);
+      llvm::ConstantInt::getFalse(ctx_), global_var_address, current_block_);
   return old_val;
 }
 
 void create_term::enable_gc(llvm::Value *was_enabled) {
   llvm::Constant *global
       = module_->getOrInsertGlobal("gc_enabled", llvm::Type::getInt1Ty(ctx_));
   auto *global_var = llvm::cast<llvm::GlobalVariable>(global);
-  new llvm::StoreInst(was_enabled, global_var, current_block_);
+#ifdef __MACH__
+  //
+  //	thread_local disabled for Apple
+  //
+  /*
+  global_var->setThreadLocal(true);
+  llvm::IRBuilder b(current_block_);
+  auto *global_var_address = b.CreateThreadLocalAddress(global_var);
+  */
+  auto *global_var_address = global_var;
+#else
+  global_var->setThreadLocal(true);
+  auto *global_var_address = global_var;
+#endif
+  new llvm::StoreInst(was_enabled, global_var_address, current_block_);
 }
 
 // We use tailcc calling convention for apply_rule_* and eval_* functions to

diff --git a/lib/codegen/Decision.cpp b/lib/codegen/Decision.cpp
@@ -5,6 +5,7 @@
 #include "kllvm/codegen/ProofEvent.h"
 #include "kllvm/codegen/Util.h"
 
+#include "llvm/IR/IRBuilder.h"
 #include <llvm/ADT/APInt.h>
 #include <llvm/ADT/SmallString.h>
 #include <llvm/ADT/StringMap.h>
@@ -1006,9 +1007,25 @@ std::pair<std::vector<llvm::Value *>, llvm::BasicBlock *> step_function_header(
 
   auto *collection = module->getOrInsertGlobal(
       "time_for_collection", llvm::Type::getInt1Ty(module->getContext()));
+
+#ifdef __MACH__
+  //
+  //	thread_local disabled for Apple
+  //
+  /*
+  llvm::cast<llvm::GlobalVariable>(collection)->setThreadLocal(true);
+  llvm::IRBuilder b(check_collect);
+  auto *collection_address = b.CreateThreadLocalAddress(collection);
+  */
+  auto *collection_address = collection;
+#else
+  llvm::cast<llvm::GlobalVariable>(collection)->setThreadLocal(true);
+  auto *collection_address = collection;
+#endif
+
   auto *is_collection = new llvm::LoadInst(
-      llvm::Type::getInt1Ty(module->getContext()), collection, "is_collection",
-      check_collect);
+      llvm::Type::getInt1Ty(module->getContext()), collection_address,
+      "is_collection", check_collect);
   set_debug_loc(is_collection);
   auto *collect = llvm::BasicBlock::Create(
       module->getContext(), "isCollect", block->getParent());

diff --git a/runtime/alloc/arena.cpp b/runtime/alloc/arena.cpp
@@ -3,6 +3,8 @@
 #include <cstdio>
 #include <cstdlib>
 #include <cstring>
+#include <memory>
+#include <sys/mman.h>
 
 #include "runtime/alloc.h"
 #include "runtime/arena.h"
@@ -47,36 +49,68 @@ get_arena_semispace_id_of_object(void *ptr) {
   return mem_block_header(ptr)->semispace;
 }
 
-static void *first_superblock_ptr = nullptr;
-static void *superblock_ptr = nullptr;
-static char **next_superblock_ptr = nullptr;
-static unsigned blocks_left = 0;
+//
+//	We will reserve enough address space for 1 million 1MB blocks. Might want to increase this on a > 1TB server.
+//
+size_t const HYPERBLOCK_SIZE = (size_t)BLOCK_SIZE * 1024 * 1024;
+static thread_local void *hyperblock_ptr = nullptr; // only needed for munmap()
 
 static void *megabyte_malloc() {
-  if (blocks_left == 0) {
-    blocks_left = 15;
-    if (int result
-        = posix_memalign(&superblock_ptr, BLOCK_SIZE, BLOCK_SIZE * 15)) {
-      errno = result;
-      perror("posix_memalign");
-    }
-    if (!first_superblock_ptr) {
-      first_superblock_ptr = superblock_ptr;
-    }
-    if (next_superblock_ptr) {
-      *next_superblock_ptr = (char *)superblock_ptr;
+  //
+  //	Return pointer to a BLOCK_SIZE chunk of memory with BLOCK_SIZE alignment.
+  //
+  static thread_local char *currentblock_ptr
+      = nullptr; // char* rather than void* to permit pointer arithmetic
+  if (currentblock_ptr) {
+    //
+    //	We expect an page fault due to not being able to map physical memory to this block or the
+    //	process to be killed by the OOM killer long before we run off the end of our address space.
+    //
+    currentblock_ptr += BLOCK_SIZE;
+  } else {
+    //
+    //	First call - need to reserve the address space.
+    //
+    size_t request = HYPERBLOCK_SIZE;
+    void *addr = mmap(
+        nullptr, // let OS choose the address
+        request, // Linux and MacOS both allow up to 64TB
+        PROT_READ | PROT_WRITE, // read, write but not execute
+        MAP_ANONYMOUS | MAP_PRIVATE
+            | MAP_NORESERVE, // allocate address space only
+        -1, // no file backing
+        0); // no offset
+    if (addr == MAP_FAILED) {
+      perror("mmap()");
+      abort();
     }
-    auto *hdr = (memory_block_header *)superblock_ptr;
-    next_superblock_ptr = &hdr->next_superblock;
-    hdr->next_superblock = nullptr;
+    hyperblock_ptr = addr;
+    //
+    //	We ask for one block worth of address space less than we allocated so alignment will always succeed.
+    //	We don't worry about unused address space either side of our aligned address space because there will be no
+    //	memory mapped to it.
+    //
+    currentblock_ptr = reinterpret_cast<char *>(
+        std::align(BLOCK_SIZE, HYPERBLOCK_SIZE - BLOCK_SIZE, addr, request));
   }
-  blocks_left--;
-  void *result = superblock_ptr;
-  superblock_ptr = (char *)superblock_ptr + BLOCK_SIZE;
-  return result;
+  return currentblock_ptr;
 }
 
+void free_all_memory() {
+  //
+  //	Frees all memory that was demand paged into this address range.
+  //
+  munmap(hyperblock_ptr, HYPERBLOCK_SIZE);
+}
+
+#ifdef __MACH__
+//
+//	thread_local disabled for Apple
+//
 bool time_for_collection;
+#else
+thread_local bool time_for_collection;
+#endif
 
 static void fresh_block(struct arena *arena) {
   char *next_block = nullptr;
@@ -122,7 +156,14 @@ static void fresh_block(struct arena *arena) {
       BLOCK_SIZE - sizeof(memory_block_header));
 }
 
+#ifdef __MACH__
+//
+//	thread_local disabled for Apple
+//
 bool gc_enabled = true;
+#else
+thread_local bool gc_enabled = true;
+#endif
 
 __attribute__((noinline)) void *
 do_alloc_slow(size_t requested, struct arena *arena) {
@@ -229,16 +270,3 @@ size_t arena_size(const struct arena *arena) {
               : arena->num_collection_blocks)
          * (BLOCK_SIZE - sizeof(memory_block_header));
 }
-
-void free_all_memory() {
-  auto *superblock = (memory_block_header *)first_superblock_ptr;
-  while (superblock) {
-    auto *next_superblock = (memory_block_header *)superblock->next_superblock;
-    free(superblock);
-    superblock = next_superblock;
-  }
-  first_superblock_ptr = nullptr;
-  superblock_ptr = nullptr;
-  next_superblock_ptr = nullptr;
-  blocks_left = 0;
-}
diff --git a/runtime/alloc/register_gc_roots_enum.cpp b/runtime/alloc/register_gc_roots_enum.cpp
@@ -3,7 +3,7 @@
 #include "runtime/collect.h"
 #include "runtime/header.h"
 
-std::vector<BlockEnumerator> block_enumerators;
+thread_local std::vector<BlockEnumerator> block_enumerators;
 
 void register_gc_roots_enumerator(BlockEnumerator f) {
   block_enumerators.push_back(f);

diff --git a/runtime/arithmetic/int.cpp b/runtime/arithmetic/int.cpp
@@ -373,8 +373,8 @@ void int_hash(mpz_t i, void *hasher) {
   }
 }
 
-gmp_randstate_t kllvm_rand_state;
-bool kllvm_rand_state_initialized = false;
+thread_local gmp_randstate_t kllvm_rand_state;
+thread_local bool kllvm_rand_state_initialized = false;
 
 SortK hook_INT_srand(SortInt seed) {
   if (!kllvm_rand_state_initialized) {

diff --git a/runtime/collect/collect.cpp b/runtime/collect/collect.cpp
@@ -16,15 +16,15 @@ char **old_alloc_ptr(void);
 char *youngspace_ptr(void);
 char *oldspace_ptr(void);
 
-static bool is_gc = false;
-bool collect_old = false;
+static thread_local bool is_gc = false;
+bool thread_local collect_old = false;
 #ifndef GC_DBG
-static uint8_t num_collection_only_young = 0;
+static thread_local uint8_t num_collection_only_young = 0;
 #else
-static char *last_alloc_ptr;
+static thread_local char *last_alloc_ptr;
 #endif
 
-size_t numBytesLiveAtCollection[1 << AGE_WIDTH];
+size_t thread_local numBytesLiveAtCollection[1 << AGE_WIDTH];
 
 bool during_gc() {
   return is_gc;

diff --git a/runtime/collect/migrate_static_roots.cpp b/runtime/collect/migrate_static_roots.cpp
@@ -2,10 +2,10 @@
 
 #include "runtime/collect.h"
 
-extern std::vector<BlockEnumerator> block_enumerators;
+extern thread_local std::vector<BlockEnumerator> block_enumerators;
 
-extern gmp_randstate_t kllvm_rand_state;
-extern bool kllvm_rand_state_initialized;
+extern thread_local gmp_randstate_t kllvm_rand_state;
+extern thread_local bool kllvm_rand_state_initialized;
 
 extern "C" {
 

diff --git a/runtime/lto/alloc.cpp b/runtime/lto/alloc.cpp
@@ -159,8 +159,8 @@ static inline void *kore_alloc_collection(kllvm::sort_category cat) {
   void *mem
       = kore_alloc(sizeof(blockheader) + sizeof(collection) + sizeof(uint64_t));
   auto *hdr = (blockheader *)mem;
-  static std::string name = get_raw_symbol_name(cat) + "{}";
-  static blockheader hdr_val
+  static thread_local std::string name = get_raw_symbol_name(cat) + "{}";
+  static thread_local blockheader hdr_val
       = get_block_header_for_symbol(get_tag_for_symbol_name(name.c_str()));
   *hdr = hdr_val;
   auto *offset = (uint64_t *)(hdr + 1);

diff --git a/unittests/runtime-collections/lists.cpp b/unittests/runtime-collections/lists.cpp
@@ -62,7 +62,15 @@ block D1 = {{1}};
 block *DUMMY1 = &D1;
 }
 
+#ifdef __MACH__
+//
+//	thread_local disabled for Apple
+//
 bool gc_enabled;
+#else
+thread_local bool gc_enabled;
+#endif
+
 size_t get_gc_threshold() {
   return SIZE_MAX;
 }