Merge branch 'develop' into update-install-docs

bin/llvm-kompile

@@ -42,6 +42,7 @@ Options:
                                     (immutable) that are enabled by default.
   --hidden-visibility               Set the visibility of all global symbols in generated code to
                                     "hidden"
+  --use-gcstrategy                  Use GC strategy defined for the LLVM backend.
   --profile-matching                Instrument interpeter to emit a profile of time spent in
                                     top-level rule matching on stderr.
   --verify-ir                       Verify result of IR generation.
@@ -197,6 +198,11 @@ while [[ $# -gt 0 ]]; do
       kompile_clang_flags+=("--hidden-visibility")
       shift
       ;;
+    --use-gcstrategy)
+      codegen_flags+=("--use-gcstrategy")
+      kompile_clang_flags+=("--use-gcstrategy")
+      shift
+      ;;
     --profile-matching)
       codegen_flags+=("--profile-matching")
       codegen_verify_flags+=("--profile-matching")

bin/llvm-kompile-clang

@@ -15,6 +15,7 @@ flags=()
 llc_flags=()
 llc_opt_flags="-O0"
 visibility_hidden=false
+use_gcstrategy=false
 link=true
 export verbose=false
 export profile=false
@@ -101,6 +102,10 @@ while [[ $# -gt 0 ]]; do
       visibility_hidden=true
       shift
       ;;
+    --use-gcstrategy)
+      use_gcstrategy=true
+      shift
+      ;;
     *)
       ;;
   esac
@@ -188,6 +193,9 @@ if [ "$main" != "python_ast" ]; then
         run @OPT@ "$modopt" -load-pass-plugin "$passes" -set-visibility-hidden -o "$modhidden"
         modopt="$modhidden"
       fi
+      if $use_gcstrategy; then
+        llc_flags+=("-load="$passes"")
+      fi
       run @LLC@ \
         "$modopt" -mtriple=@BACKEND_TARGET_TRIPLE@ \
         -filetype=obj "$llc_opt_flags" "${llc_flags[@]}" -o "$modasm"

include/kllvm/codegen/GCStrategy.h

@@ -0,0 +1,37 @@
+//===- Extend GCStrategy of llvm/CodeGen/GCStrategy.h ---------------------===//
+//
+// We extend the base GCStrategy as follows:
+// - use gc.safepoints instead of (default) gc.roots.
+// - specify that the RewriteStatepointsForGC pass should rewrite the calls of
+//   this function.
+// - pointers with address space != 0 are pointing to GC-managed memory.
+//===----------------------------------------------------------------------===//
+
+// NOLINTBEGIN
+
+#ifndef LLVM_BACKEND_GC_STRATEGY_H
+#define LLVM_BACKEND_GC_STRATEGY_H
+
+#include "llvm/IR/GCStrategy.h"
+#include "llvm/IR/Type.h"
+
+namespace kllvm {
+
+/// The GCStrategy for the LLVM Backend
+class LLVMBackendGCStrategy : public llvm::GCStrategy {
+public:
+  LLVMBackendGCStrategy();
+
+  // Override
+#if LLVM_VERSION_MAJOR == 15
+  llvm::Optional<bool> isGCManagedPointer(llvm::Type const *Ty) const override;
+#else
+  std::optional<bool> isGCManagedPointer(llvm::Type const *Ty) const override;
+#endif
+};
+
+} // namespace kllvm
+
+#endif // LLVM_BACKEND_GC_STRATEGY_H
+
+// NOLINTEND

include/kllvm/codegen/Options.h

@@ -10,6 +10,7 @@ extern llvm::cl::opt<bool> no_optimize;
 extern llvm::cl::opt<bool> emit_object;
 extern llvm::cl::opt<bool> binary_ir;
 extern llvm::cl::opt<bool> force_binary;
+extern llvm::cl::opt<bool> use_gcstrategy;
 extern llvm::cl::opt<bool> proof_hint_instrumentation;
 extern llvm::cl::opt<bool> proof_hint_instrumentation_slow;
 extern llvm::cl::opt<bool> keep_frame_pointer;

include/runtime/arena.h

@@ -1,38 +1,49 @@
 #ifndef ARENA_H
 #define ARENA_H
 
+#include <algorithm>
 #include <cstddef>
+#include <cstdint>
 #include <sys/types.h>
+#include <utility>
 
 #include "runtime/alloc.h"
 
 extern "C" {
 
+size_t const HYPERBLOCK_SIZE = (size_t)BLOCK_SIZE * 1024 * 1024;
+
 // An arena can be used to allocate objects that can then be deallocated all at
 // once.
 class arena {
 public:
   arena(char id)
-      : allocation_semispace_id(id) { }
+      : allocation_semispace_id(id) {
+    initialize_semispace();
+  }
 
   // Allocates the requested number of bytes as a contiguous region and returns a
   // pointer to the first allocated byte.
-  // If called with requested size greater than the maximun single allocation
-  // size, the space is allocated in a general (not garbage collected pool).
   void *kore_arena_alloc(size_t requested);
 
   // Returns the address of the first byte that belongs in the given arena.
   // Returns 0 if nothing has been allocated ever in that arena.
-  char *arena_start_ptr() const;
+  char *arena_start_ptr() const {
+    return current_addr_ptr ? current_addr_ptr + sizeof(memory_block_header)
+                            : nullptr;
+  }
 
   // Returns a pointer to a location holding the address of last allocated
   // byte in the given arena plus 1.
   // This address is 0 if nothing has been allocated ever in that arena.
-  char **arena_end_ptr();
+  char **arena_end_ptr() { return &allocation_ptr; }
 
   // return the total number of allocatable bytes currently in the arena in its
   // active semispace.
-  size_t arena_size() const;
+  size_t arena_size() const {
+    update_num_blocks();
+    return BLOCK_SIZE * std::max(num_blocks, num_collection_blocks);
+  }
 
   // Clears the current allocation space by setting its start back to its first
   // block. It is used during garbage collection to effectively collect all of the
@@ -41,15 +52,18 @@ class arena {
 
   // Resizes the last allocation as long as the resize does not require a new
   // block allocation.
-  // Returns the address of the byte following the last newlly allocated byte when
-  // the resize succeeds, returns 0 otherwise.
-  void *arena_resize_last_alloc(ssize_t increase);
+  // Returns the address of the byte following the last newlly allocated byte.
+  void *arena_resize_last_alloc(ssize_t increase) {
+    return (allocation_ptr += increase);
+  }
 
   // Returns the given arena's current collection semispace ID.
   // Each arena has 2 semispace IDs one equal to the arena ID and the other equal
   // to the 1's complement of the arena ID. At any time one of these semispaces
   // is used for allocation and the other is used for collection.
-  char get_arena_collection_semispace_id() const;
+  char get_arena_collection_semispace_id() const {
+    return ~allocation_semispace_id;
+  }
 
   // Exchanges the current allocation and collection semispaces and clears the new
   // current allocation semispace by setting its start back to its first block.
@@ -61,7 +75,7 @@ class arena {
   // by the blocks of that arena. This difference will include blocks containing
   // sentinel bytes. Undefined behavior will result if the pointers belong to
   // different arenas.
-  static ssize_t ptr_diff(char *ptr1, char *ptr2);
+  static ssize_t ptr_diff(char *ptr1, char *ptr2) { return ptr1 - ptr2; }
 
   // Given a starting pointer to an address allocated in an arena and a size in
   // bytes, this function returns a pointer to an address allocated in the
@@ -72,42 +86,70 @@ class arena {
   // 3rd argument: the address of last allocated byte in the arena plus 1
   // Return value: the address allocated in the arena after size bytes from the
   //               starting pointer, or 0 if this is equal to the 3rd argument.
-  static char *move_ptr(char *ptr, size_t size, char const *arena_end_ptr);
+  static char *move_ptr(char *ptr, size_t size, char const *arena_end_ptr) {
+    char *next_ptr = ptr + size;
+    return (next_ptr == arena_end_ptr) ? 0 : next_ptr;
+  }
 
   // Returns the ID of the semispace where the given address was allocated.
   // The behavior is undefined if called with an address that has not been
   // allocated within an arena.
   static char get_arena_semispace_id_of_object(void *ptr);
 
 private:
-  struct memory_block_header {
-    char *next_block;
+  union memory_block_header {
+    //
+    //	Currently the header just holds the semispace id. But we need it to be a
+    //	multiple of sizeof(char*) for alignment purposes so we add a dummy char*.
+    //
     char semispace;
+    char *alignment_dummy;
   };
 
-  void fresh_block();
-  static memory_block_header *mem_block_header(void *ptr);
+  //
+  //	We update the number of 1MB blocks actually written to, only when we need this value,
+  //	or before a garbage collection rather than trying to determine when we write to a fresh block.
+  //
+  void update_num_blocks() const {
+    //
+    //	Calculate how many 1M blocks of the current arena we used.
+    //
+    size_t num_used_blocks
+        = (allocation_ptr - current_addr_ptr - 1) / BLOCK_SIZE + 1;
+    if (num_used_blocks > num_blocks)
+      num_blocks = num_used_blocks;
+  }
+
+  void initialize_semispace();
 
-  // helper function for `kore_arena_alloc`. Do not call directly.
-  void *do_alloc_slow(size_t requested);
+  static memory_block_header *mem_block_header(void *ptr) {
+    uintptr_t address = reinterpret_cast<uintptr_t>(ptr);
+    return reinterpret_cast<arena::memory_block_header *>(
+        (address - 1) & ~(HYPERBLOCK_SIZE - 1));
+  }
 
-  char *first_block; // beginning of first block
-  char *block; // where allocations are being made in current block
-  char *block_start; // start of current block
-  char *block_end; // 1 past end of current block
-  char *first_collection_block; // beginning of other semispace
-  size_t num_blocks; // number of blocks in current semispace
-  size_t num_collection_blocks; // number of blocks in other semispace
+  //
+  //	Current semispace where allocations are being made.
+  //
+  char *current_addr_ptr; // pointer to start of current address space
+  char *allocation_ptr; // next available location in current semispace
+  char *tripwire; // allocating past this triggers slow allocation
+  mutable size_t
+      num_blocks; // notional number of BLOCK_SIZE blocks in current semispace
   char allocation_semispace_id; // id of current semispace
+  //
+  //	Semispace where allocations will be made during and after garbage collect.
+  //
+  char *collection_addr_ptr
+      = nullptr; // pointer to start of collection address space
+  size_t num_collection_blocks
+      = 0; // notional number of BLOCK_SIZE blocks in collection semispace
 };
 
 // Macro to define a new arena with the given ID. Supports IDs ranging from 0 to
 // 127.
 #define REGISTER_ARENA(name, id) static thread_local arena name(id)
 
-#define MEM_BLOCK_START(ptr)                                                   \
-  ((char *)(((uintptr_t)(ptr)-1) & ~(BLOCK_SIZE - 1)))
-
 #ifdef __MACH__
 //
 //	thread_local disabled for Apple
@@ -120,16 +162,51 @@ extern thread_local bool time_for_collection;
 size_t get_gc_threshold();
 
 inline void *arena::kore_arena_alloc(size_t requested) {
-  if (block + requested > block_end) {
-    return do_alloc_slow(requested);
+  if (allocation_ptr + requested >= tripwire) {
+    //
+    //	We got close to or past the last location accessed in this address range so far,
+    //	depending on the requested size and tripwire setting. This triggers a garbage
+    //	collect when allowed.
+    //
+    time_for_collection = true;
+    tripwire = current_addr_ptr
+               + HYPERBLOCK_SIZE; // won't trigger again until arena swap
   }
-  void *result = block;
-  block += requested;
+  void *result = allocation_ptr;
+  allocation_ptr += requested;
   MEM_LOG(
-      "Allocation at %p (size %zd), next alloc at %p (if it fits)\n", result,
-      requested, block);
+      "Allocation at %p (size %zd), next alloc at %p\n", result, requested,
+      block);
   return result;
 }
+
+inline void arena::arena_clear() {
+  //
+  //	We set the allocation pointer to the first available address.
+  //
+  allocation_ptr = arena_start_ptr();
+  //
+  //	If the number of blocks we've touched is >= threshold, we want to trigger
+  //	a garbage collection if we get within 1 block of the end of this area.
+  //	Otherwise we only want to generate a garbage collect if we allocate off the
+  //	end of this area.
+  //
+  tripwire = current_addr_ptr
+             + (num_blocks - (num_blocks >= get_gc_threshold())) * BLOCK_SIZE;
 }
 
+inline void arena::arena_swap_and_clear() {
+  update_num_blocks(); // so we save the correct number of touched blocks
+  std::swap(current_addr_ptr, collection_addr_ptr);
+  std::swap(num_blocks, num_collection_blocks);
+  allocation_semispace_id = ~allocation_semispace_id;
+  if (current_addr_ptr == nullptr) {
+    //
+    //	The other semispace hasn't be initialized yet.
+    //
+    initialize_semispace();
+  } else
+    arena_clear();
+}
+}
 #endif // ARENA_H

lib/codegen/CreateTerm.cpp

@@ -1,6 +1,7 @@
 #include "kllvm/codegen/CreateTerm.h"
 #include "kllvm/codegen/CreateStaticTerm.h"
 #include "kllvm/codegen/Debug.h"
+#include "kllvm/codegen/Options.h"
 #include "kllvm/codegen/ProofEvent.h"
 #include "kllvm/codegen/Util.h"
 
@@ -1224,6 +1225,9 @@ bool make_function(
       = llvm::FunctionType::get(return_type, param_types, false);
   llvm::Function *apply_rule = get_or_insert_function(module, name, func_type);
   apply_rule->setLinkage(llvm::GlobalValue::InternalLinkage);
+  if (use_gcstrategy) {
+    apply_rule->setGC("gcs-llvm-backend");
+  }
   init_debug_axiom(axiom->attributes());
   std::string debug_name = name;
   if (axiom->attributes().contains(attribute_set::key::Label)) {

lib/codegen/EmitConfigParser.cpp

@@ -389,7 +389,7 @@ static std::pair<llvm::Value *, llvm::BasicBlock *> get_eval(
     auto *malloc = create_malloc(
         creator.get_current_block(),
         llvm::ConstantExpr::getSizeOf(result->getType()),
-        get_or_insert_function(mod, "malloc", ptr_ty, ptr_ty));
+        get_or_insert_function(mod, "kore_alloc_always_gc", ptr_ty, ptr_ty));
     new llvm::StoreInst(result, malloc, creator.get_current_block());
     retval = malloc;
     break;
@@ -563,7 +563,8 @@ static void emit_get_token(kore_definition *definition, llvm::Module *module) {
           case_block);
       auto *malloc = create_malloc(
           case_block, llvm::ConstantExpr::getSizeOf(compare->getType()),
-          get_or_insert_function(module, "malloc", ptr_ty, ptr_ty));
+          get_or_insert_function(
+              module, "kore_alloc_always_gc", ptr_ty, ptr_ty));
       new llvm::StoreInst(compare, malloc, case_block);
       phi->addIncoming(malloc, case_block);
       llvm::BranchInst::Create(merge_block, case_block);

lib/codegen/Options.cpp

@@ -48,6 +48,10 @@ cl::opt<bool> force_binary(
     "f", cl::desc("Force binary bitcode output to stdout"), cl::Hidden,
     cl::cat(codegen_lib_cat));
 
+cl::opt<bool> use_gcstrategy(
+    "use-gcstrategy", cl::desc("Use GC strategy defined for the LLVM backend."),
+    cl::Hidden, cl::init(false), cl::cat(codegen_lib_cat));
+
 namespace kllvm {
 
 void validate_codegen_args(bool is_tty) {

lib/passes/CMakeLists.txt

@@ -2,13 +2,15 @@ add_library(KLLVMPassInternal
   SetVisibilityHidden.cpp
   RemoveDeadKFunctions.cpp
   MustTailDeadArgElimination.cpp
+  GCStrategy.cpp
   PluginInfo.cpp
 )
 
 add_library(KLLVMPass MODULE
   SetVisibilityHidden.cpp
   RemoveDeadKFunctions.cpp
   MustTailDeadArgElimination.cpp
+  GCStrategy.cpp
   PluginInfo.cpp
 )