Adding support for MMTk (non-moving Immix)

Author: udesou

Make.inc

@@ -80,6 +80,9 @@ HAVE_SSP := 0
 WITH_GC_VERIFY := 0
 WITH_GC_DEBUG_ENV := 0
 
+# Use stock if MMTK_PLAN hasn't been defined
+MMTK_PLAN ?= None
+
 # Enable DTrace support
 WITH_DTRACE := 0
 
@@ -833,6 +836,41 @@ JCXXFLAGS += -DGC_DEBUG_ENV
 JCFLAGS += -DGC_DEBUG_ENV
 endif
 
+ifneq (${MMTK_PLAN},None)
+ifeq (${MMTK_JULIA_DIR},)
+$(error MMTK_JULIA_DIR must be set to use MMTk)
+endif
+JCXXFLAGS += -DMMTK_GC
+JCFLAGS += -DMMTK_GC
+ifeq (${MMTK_BUILD},)
+ifeq (debug,$(findstring debug,$(MAKECMDGOALS)))
+MMTK_BUILD = debug
+else
+MMTK_BUILD = release
+endif
+endif
+ifeq (${MMTK_PLAN},Immix)
+JCXXFLAGS += -DMMTK_PLAN_IMMIX
+JCFLAGS += -DMMTK_PLAN_IMMIX
+else
+$(error "Unsupported MMTk plan: $(MMTK_PLAN)")
+endif
+MMTK_DIR = ${MMTK_JULIA_DIR}/mmtk
+MMTK_API_INC = $(MMTK_DIR)/api
+ifeq ($(OS),Linux)
+MMTK_LIB_NAME := libmmtk_julia.so
+else
+$(error "Unsupported OS for MMTk")
+endif
+MMTK_LIB_SRC := $(MMTK_DIR)/target/$(MMTK_BUILD)/$(MMTK_LIB_NAME)
+MMTK_LIB_DST := $(BUILDROOT)/usr/lib/$(MMTK_LIB_NAME)
+MMTK_LIB := -lmmtk_julia
+LDFLAGS += -Wl,-rpath=$(MMTK_DIR)/target/$(MMTK_BUILD)/
+else
+MMTK_JULIA_INC :=
+MMTK_LIB :=
+endif
+
 ifeq ($(WITH_DTRACE), 1)
 JCXXFLAGS += -DUSE_DTRACE
 JCFLAGS += -DUSE_DTRACE
@@ -1827,6 +1865,9 @@ PRINT_PERL = printf '    %b %b\n' $(PERLCOLOR)PERL$(ENDCOLOR) $(BINCOLOR)$(GOAL)
 PRINT_FLISP = printf '    %b %b\n' $(FLISPCOLOR)FLISP$(ENDCOLOR) $(BINCOLOR)$(GOAL)$(ENDCOLOR); $(1)
 PRINT_JULIA = printf '    %b %b\n' $(JULIACOLOR)JULIA$(ENDCOLOR) $(BINCOLOR)$(GOAL)$(ENDCOLOR); $(1)
 PRINT_DTRACE = printf '    %b %b\n' $(DTRACECOLOR)DTRACE$(ENDCOLOR) $(BINCOLOR)$(GOAL)$(ENDCOLOR); $(1)
+ifneq (${MMTK_PLAN},None)
+PRINT_MMTK = printf '    %b %b\n' $(LINKCOLOR)MMTK$(ENDCOLOR) $(BINCOLOR)$(GOAL)$(ENDCOLOR); $(1)
+endif
 
 else
 QUIET_MAKE =
@@ -1837,6 +1878,9 @@ PRINT_PERL = echo '$(subst ','\'',$(1))'; $(1)
 PRINT_FLISP = echo '$(subst ','\'',$(1))'; $(1)
 PRINT_JULIA = echo '$(subst ','\'',$(1))'; $(1)
 PRINT_DTRACE = echo '$(subst ','\'',$(1))'; $(1)
+ifneq (${MMTK_PLAN},None)
+PRINT_MMTK = echo '$(subst ','\'',$(1))'; $(1)
+endif
 
 endif # VERBOSE
 

base/timing.jl

@@ -106,9 +106,14 @@ function gc_page_utilization_data()
     return Base.unsafe_wrap(Array, page_utilization_raw, JL_GC_N_MAX_POOLS, own=false)
 end
 
+
+const USING_STOCK_GC = occursin("stock", unsafe_string(ccall(:jl_gc_active_impl, Ptr{UInt8}, ())))
+# Full sweep reasons are currently only available for the stock GC
+@static if USING_STOCK_GC
 # must be kept in sync with `src/gc-stock.h``
 const FULL_SWEEP_REASONS = [:FULL_SWEEP_REASON_SWEEP_ALWAYS_FULL, :FULL_SWEEP_REASON_FORCED_FULL_SWEEP,
                             :FULL_SWEEP_REASON_USER_MAX_EXCEEDED, :FULL_SWEEP_REASON_LARGE_PROMOTION_RATE]
+end
 
 """
     Base.full_sweep_reasons()
@@ -124,11 +129,15 @@ The reasons are:
 Note that the set of reasons is not guaranteed to be stable across minor versions of Julia.
 """
 function full_sweep_reasons()
-    reason = cglobal(:jl_full_sweep_reasons, UInt64)
-    reasons_as_array = Base.unsafe_wrap(Vector{UInt64}, reason, length(FULL_SWEEP_REASONS), own=false)
     d = Dict{Symbol, Int64}()
-    for (i, r) in enumerate(FULL_SWEEP_REASONS)
-        d[r] = reasons_as_array[i]
+    # populate the dictionary according to the reasons above for the stock GC
+    # otherwise return an empty dictionary for now
+    @static if USING_STOCK_GC
+        reason = cglobal(:jl_full_sweep_reasons, UInt64)
+        reasons_as_array = Base.unsafe_wrap(Vector{UInt64}, reason, length(FULL_SWEEP_REASONS), own=false)
+        for (i, r) in enumerate(FULL_SWEEP_REASONS)
+            d[r] = reasons_as_array[i]
+        end
     end
     return d
 end

src/Makefile

@@ -29,6 +29,10 @@ ifeq ($(USECLANG),1)
 FLAGS += -Wno-return-type-c-linkage -Wno-atomic-alignment
 endif
 
+ifeq ($(WITH_MMTK), 1)
+FLAGS += -I$(MMTK_API_INC)
+endif
+
 FLAGS += -DJL_BUILD_ARCH='"$(ARCH)"'
 ifeq ($(OS),WINNT)
 FLAGS += -DJL_BUILD_UNAME='"NT"'
@@ -40,23 +44,41 @@ ifeq ($(OS),FreeBSD)
 FLAGS += -I$(LOCALBASE)/include
 endif
 
+# GC source code. It depends on which GC implementation to use.
+GC_SRCS := gc-common gc-stacks gc-alloc-profiler gc-heap-snapshot
+ifneq (${MMTK_PLAN},None)
+GC_SRCS += gc-mmtk
+else
+GC_SRCS += gc-stock gc-debug gc-pages gc-page-profiler
+endif
+
 SRCS := \
 	jltypes gf typemap smallintset ast builtins module interpreter symbol \
 	dlload sys init task array genericmemory staticdata toplevel jl_uv datatype \
 	simplevector runtime_intrinsics precompile jloptions mtarraylist \
-	threading scheduler stackwalk gc-common gc-stock gc-debug gc-pages gc-stacks gc-alloc-profiler gc-page-profiler method \
-	jlapi signal-handling safepoint timing subtype rtutils gc-heap-snapshot \
-	crc32c APInt-C processor ircode opaque_closure codegen-stubs coverage runtime_ccall engine
+	threading scheduler stackwalk \
+	method jlapi signal-handling safepoint timing subtype rtutils \
+	crc32c APInt-C processor ircode opaque_closure codegen-stubs coverage runtime_ccall engine \
+	$(GC_SRCS)
 
 RT_LLVMLINK :=
 CG_LLVMLINK :=
 
 ifeq ($(JULIACODEGEN),LLVM)
+# Currently these files are used by both GCs. But we should make the list specific to stock, and MMTk should have its own implementation.
+GC_CODEGEN_SRCS := llvm-final-gc-lowering llvm-late-gc-lowering llvm-gc-invariant-verifier
+ifneq (${MMTK_PLAN},None)
+FLAGS += -I$(MMTK_API_INC)
+GC_CODEGEN_SRCS += llvm-late-gc-lowering-mmtk
+else
+GC_CODEGEN_SRCS += llvm-late-gc-lowering-stock
+endif
 CODEGEN_SRCS := codegen jitlayers aotcompile debuginfo disasm llvm-simdloop \
-	llvm-final-gc-lowering llvm-pass-helpers llvm-late-gc-lowering llvm-ptls \
-	llvm-lower-handlers llvm-gc-invariant-verifier llvm-propagate-addrspaces \
+	llvm-pass-helpers llvm-ptls \
+	llvm-lower-handlers llvm-propagate-addrspaces \
 	llvm-multiversioning llvm-alloc-opt llvm-alloc-helpers cgmemmgr llvm-remove-addrspaces \
-	llvm-remove-ni llvm-julia-licm llvm-demote-float16 llvm-cpufeatures pipeline llvm_api
+	llvm-remove-ni llvm-julia-licm llvm-demote-float16 llvm-cpufeatures pipeline llvm_api \
+	$(GC_CODEGEN_SRCS)
 FLAGS += -I$(shell $(LLVM_CONFIG_HOST) --includedir)
 CG_LLVM_LIBS := all
 ifeq ($(USE_POLLY),1)
@@ -99,7 +121,12 @@ ifeq ($(USE_SYSTEM_LIBUV),0)
 UV_HEADERS += uv.h
 UV_HEADERS += uv/*.h
 endif
-PUBLIC_HEADERS := $(BUILDDIR)/julia_version.h $(wildcard $(SRCDIR)/support/*.h) $(addprefix $(SRCDIR)/,work-stealing-queue.h gc-interface.h gc-tls.h gc-tls-common.h julia.h julia_assert.h julia_threads.h julia_fasttls.h julia_locks.h julia_atomics.h jloptions.h)
+PUBLIC_HEADERS := $(BUILDDIR)/julia_version.h $(wildcard $(SRCDIR)/support/*.h) $(addprefix $(SRCDIR)/,work-stealing-queue.h gc-interface.h gc-tls-common.h julia.h julia_assert.h julia_threads.h julia_fasttls.h julia_locks.h julia_atomics.h jloptions.h)
+ifneq (${MMTK_PLAN},None)
+	PUBLIC_HEADERS += $(addprefix $(SRCDIR)/,gc-tls-mmtk.h)
+else
+	PUBLIC_HEADERS += $(addprefix $(SRCDIR)/,gc-tls-stock.h)
+endif
 ifeq ($(OS),WINNT)
 PUBLIC_HEADERS += $(addprefix $(SRCDIR)/,win32_ucontext.h)
 endif
@@ -164,8 +191,8 @@ LIBJULIA_PATH_REL := libjulia
 endif
 
 COMMON_LIBPATHS := -L$(build_libdir) -L$(build_shlibdir)
-RT_LIBS := $(WHOLE_ARCHIVE) $(LIBUV) $(WHOLE_ARCHIVE) $(LIBUTF8PROC) $(NO_WHOLE_ARCHIVE) $(LIBUNWIND) $(RT_LLVMLINK) $(OSLIBS) $(LIBTRACYCLIENT) $(LIBITTAPI)
-CG_LIBS := $(LIBUNWIND) $(CG_LLVMLINK) $(OSLIBS) $(LIBTRACYCLIENT) $(LIBITTAPI)
+RT_LIBS := $(WHOLE_ARCHIVE) $(LIBUV) $(WHOLE_ARCHIVE) $(LIBUTF8PROC) $(NO_WHOLE_ARCHIVE) $(LIBUNWIND) $(RT_LLVMLINK) $(OSLIBS) $(LIBTRACYCLIENT) $(LIBITTAPI) $(MMTK_LIB)
+CG_LIBS := $(LIBUNWIND) $(CG_LLVMLINK) $(OSLIBS) $(LIBTRACYCLIENT) $(LIBITTAPI) $(MMTK_LIB)
 RT_DEBUG_LIBS := $(COMMON_LIBPATHS) $(WHOLE_ARCHIVE) $(BUILDDIR)/flisp/libflisp-debug.a $(WHOLE_ARCHIVE) $(BUILDDIR)/support/libsupport-debug.a -ljulia-debug $(RT_LIBS)
 CG_DEBUG_LIBS := $(COMMON_LIBPATHS) $(CG_LIBS) -ljulia-debug -ljulia-internal-debug
 RT_RELEASE_LIBS := $(COMMON_LIBPATHS) $(WHOLE_ARCHIVE) $(BUILDDIR)/flisp/libflisp.a $(WHOLE_ARCHIVE) $(BUILDDIR)/support/libsupport.a -ljulia $(RT_LIBS)
@@ -222,6 +249,12 @@ $(BUILDDIR)/%.h.gen : $(SRCDIR)/%.d
 	sed 's/JULIA_/JL_PROBE_/' $@ > $@.tmp
 	mv $@.tmp $@
 
+# Compile files from the binding side and copy so file into lib folder
+ifneq (${MMTK_PLAN},None)
+$(MMTK_LIB_DST): $(MMTK_LIB_SRC)
+	@$(call PRINT_MMTK, cp $< $@)
+endif
+
 $(BUILDDIR)/jl_internal_funcs.inc: $(SRCDIR)/jl_exported_funcs.inc
 	# Generate `.inc` file that contains a list of `#define` macros to rename functions defined in `libjulia-internal`
 	# to have a `ijl_` prefix instead of `jl_`, to denote that they are coming from `libjulia-internal`.  This avoids
@@ -314,6 +347,7 @@ $(BUILDDIR)/debuginfo.o $(BUILDDIR)/debuginfo.dbg.obj: $(addprefix $(SRCDIR)/,de
 $(BUILDDIR)/disasm.o $(BUILDDIR)/disasm.dbg.obj: $(SRCDIR)/debuginfo.h $(SRCDIR)/processor.h
 $(BUILDDIR)/gc-debug.o $(BUILDDIR)/gc-debug.dbg.obj: $(SRCDIR)/gc-common.h $(SRCDIR)/gc-stock.h
 $(BUILDDIR)/gc-pages.o $(BUILDDIR)/gc-pages.dbg.obj: $(SRCDIR)/gc-common.h $(SRCDIR)/gc-stock.h
+$(BUILDDIR)/gc-mmtk.o $(BUILDDIR)/gc-mmtk.dbg.obj: $(SRCDIR)/gc-common.h $(SRCDIR)/gc-heap-snapshot.h $(SRCDIR)/gc-alloc-profiler.h
 $(BUILDDIR)/gc-stacks.o $(BUILDDIR)/gc-stacks.dbg.obj: $(SRCDIR)/gc-common.h $(SRCDIR)/gc-stock.h
 $(BUILDDIR)/gc-stock.o $(BUILDDIR)/gc.dbg.obj: $(SRCDIR)/gc-common.h $(SRCDIR)/gc-stock.h $(SRCDIR)/gc-heap-snapshot.h $(SRCDIR)/gc-alloc-profiler.h $(SRCDIR)/gc-page-profiler.h
 $(BUILDDIR)/gc-heap-snapshot.o $(BUILDDIR)/gc-heap-snapshot.dbg.obj: $(SRCDIR)/gc-heap-snapshot.h
@@ -386,13 +420,13 @@ $(BUILDDIR)/julia.expmap: $(SRCDIR)/julia.expmap.in $(JULIAHOME)/VERSION $(LLVM_
 	sed <'$<' >'$@' -e "s/@JULIA_SHLIB_SYMBOL_VERSION@/JL_LIBJULIA_$(SOMAJOR)/" \
 		        -e "s/@LLVM_SHLIB_SYMBOL_VERSION@/$(LLVM_SHLIB_SYMBOL_VERSION)/"
 
-$(build_shlibdir)/libjulia-internal.$(JL_MAJOR_MINOR_SHLIB_EXT): $(BUILDDIR)/julia.expmap $(OBJS) $(BUILDDIR)/flisp/libflisp.a $(BUILDDIR)/support/libsupport.a $(LIBUV)
+$(build_shlibdir)/libjulia-internal.$(JL_MAJOR_MINOR_SHLIB_EXT): $(BUILDDIR)/julia.expmap $(OBJS) $(MMTK_LIB_DST) $(BUILDDIR)/flisp/libflisp.a $(BUILDDIR)/support/libsupport.a $(LIBUV)
 	@$(call PRINT_LINK, $(CXXLD) $(call IMPLIB_FLAGS,$@) $(JCXXFLAGS) $(JL_CXXFLAGS) $(CXXLDFLAGS) $(SHIPFLAGS) $(OBJS) $(RPATH_LIB) -o $@ \
 		$(JLDFLAGS) $(BOLT_LDFLAGS) $(JLIBLDFLAGS) $(RT_RELEASE_LIBS) $(call SONAME_FLAGS,libjulia-internal.$(JL_MAJOR_SHLIB_EXT)))
 	@$(INSTALL_NAME_CMD)libjulia-internal.$(SHLIB_EXT) $@
 	$(DSYMUTIL) $@
 
-$(build_shlibdir)/libjulia-internal-debug.$(JL_MAJOR_MINOR_SHLIB_EXT): $(BUILDDIR)/julia.expmap $(DOBJS) $(BUILDDIR)/flisp/libflisp-debug.a $(BUILDDIR)/support/libsupport-debug.a $(LIBUV)
+$(build_shlibdir)/libjulia-internal-debug.$(JL_MAJOR_MINOR_SHLIB_EXT): $(BUILDDIR)/julia.expmap $(DOBJS) $(MMTK_LIB_DST) $(BUILDDIR)/flisp/libflisp-debug.a $(BUILDDIR)/support/libsupport-debug.a $(LIBUV)
 	@$(call PRINT_LINK, $(CXXLD) $(call IMPLIB_FLAGS,$@) $(JCXXFLAGS) $(JL_CXXFLAGS) $(CXXLDFLAGS) $(DEBUGFLAGS) $(DOBJS) $(RPATH_LIB) -o $@ \
 		$(JLDFLAGS) $(JLIBLDFLAGS) $(RT_DEBUG_LIBS) $(call SONAME_FLAGS,libjulia-internal-debug.$(JL_MAJOR_SHLIB_EXT)))
 	@$(INSTALL_NAME_CMD)libjulia-internal-debug.$(SHLIB_EXT) $@

src/gc-common.c

@@ -540,6 +540,38 @@ JL_DLLEXPORT jl_value_t *(jl_gc_alloc)(jl_ptls_t ptls, size_t sz, void *ty)
     return jl_gc_alloc_(ptls, sz, ty);
 }
 
+JL_DLLEXPORT void *jl_malloc(size_t sz)
+{
+    return jl_gc_counted_malloc(sz);
+}
+
+//_unchecked_calloc does not check for potential overflow of nm*sz
+STATIC_INLINE void *_unchecked_calloc(size_t nm, size_t sz) {
+    size_t nmsz = nm*sz;
+    return jl_gc_counted_calloc(nmsz, 1);
+}
+
+JL_DLLEXPORT void *jl_calloc(size_t nm, size_t sz)
+{
+    if (nm > SSIZE_MAX/sz)
+        return NULL;
+    return _unchecked_calloc(nm, sz);
+}
+
+JL_DLLEXPORT void jl_free(void *p)
+{
+    if (p != NULL) {
+        size_t sz = memory_block_usable_size(p, 0);
+        return jl_gc_counted_free_with_size(p, sz);
+    }
+}
+
+JL_DLLEXPORT void *jl_realloc(void *p, size_t sz)
+{
+    size_t old = p ? memory_block_usable_size(p, 0) : 0;
+    return jl_gc_counted_realloc_with_old_size(p, old, sz);
+}
+
 // =========================================================================== //
 // Generic Memory
 // =========================================================================== //
@@ -668,6 +700,24 @@ JL_DLLEXPORT void jl_throw_out_of_memory_error(void)
     jl_throw(jl_memory_exception);
 }
 
+// Sweeping mtarraylist_buffers:
+// These buffers are made unreachable via `mtarraylist_resizeto` from mtarraylist.c
+// and are freed at the end of GC via jl_gc_sweep_stack_pools_and_mtarraylist_buffers
+void sweep_mtarraylist_buffers(void) JL_NOTSAFEPOINT
+{
+    for (int i = 0; i < gc_n_threads; i++) {
+        jl_ptls_t ptls = gc_all_tls_states[i];
+        if (ptls == NULL) {
+            continue;
+        }
+        small_arraylist_t *buffers = &ptls->lazily_freed_mtarraylist_buffers;
+        void *buf;
+        while ((buf = small_arraylist_pop(buffers)) != NULL) {
+            free(buf);
+        }
+    }
+}
+
 #ifdef __cplusplus
 }
 #endif

src/gc-common.h

@@ -24,6 +24,31 @@
 extern "C" {
 #endif
 
+// =========================================================================== //
+// GC Big objects
+// =========================================================================== //
+
+JL_EXTENSION typedef struct _bigval_t {
+    struct _bigval_t *next;
+    struct _bigval_t *prev;
+    size_t sz;
+#ifdef _P64 // Add padding so that the value is 64-byte aligned
+    // (8 pointers of 8 bytes each) - (4 other pointers in struct)
+    void *_padding[8 - 4];
+#else
+    // (16 pointers of 4 bytes each) - (4 other pointers in struct)
+    void *_padding[16 - 4];
+#endif
+    //struct jl_taggedvalue_t <>;
+    union {
+        uintptr_t header;
+        struct {
+            uintptr_t gc:2;
+        } bits;
+    };
+    // must be 64-byte aligned here, in 32 & 64 bit modes
+} bigval_t;
+
 // =========================================================================== //
 // GC Callbacks
 // =========================================================================== //
@@ -187,4 +212,14 @@ extern jl_ptls_t* gc_all_tls_states;
 
 extern int gc_logging_enabled;
 
+// =========================================================================== //
+// MISC
+// =========================================================================== //
+
+// number of stacks to always keep available per pool
+#define MIN_STACK_MAPPINGS_PER_POOL 5
+
+void _jl_free_stack(jl_ptls_t ptls, void *stkbuf, size_t bufsz) JL_NOTSAFEPOINT;
+void sweep_mtarraylist_buffers(void) JL_NOTSAFEPOINT;
+
 #endif // JL_GC_COMMON_H

src/gc-interface.h

@@ -98,6 +98,13 @@ JL_DLLEXPORT void jl_gc_set_max_memory(uint64_t max_mem);
 JL_DLLEXPORT void jl_gc_collect(jl_gc_collection_t collection);
 // Returns whether the thread with `tid` is a collector thread
 JL_DLLEXPORT int gc_is_collector_thread(int tid) JL_NOTSAFEPOINT;
+// Returns which GC implementation is being used and possibly its version according to the list of supported GCs
+// NB: it should clearly identify the GC by including e.g. ‘stock’ or ‘mmtk’ as a substring.
+JL_DLLEXPORT const char* jl_gc_active_impl(void);
+// Sweep Julia's stack pools and mtarray buffers. Note that this function has been added to the interface as
+// each GC should implement it but it will most likely not be used by other code in the runtime.
+// It still needs to be annotated with JL_DLLEXPORT since it is called from Rust by MMTk.
+JL_DLLEXPORT void jl_gc_sweep_stack_pools_and_mtarraylist_buffers(jl_ptls_t ptls) JL_NOTSAFEPOINT;
 
 // ========================================================================= //
 // Metrics
@@ -138,7 +145,6 @@ JL_DLLEXPORT uint64_t jl_gc_total_hrtime(void);
 // **must** also set the type of the returning object to be `ty`. The type `ty` may also be used to record
 // an allocation of that type in the allocation profiler.
 struct _jl_value_t *jl_gc_alloc_(struct _jl_tls_states_t * ptls, size_t sz, void *ty);
-
 // Allocates small objects and increments Julia allocation counterst. Size of the object
 // header must be included in the object size. The (possibly unused in some implementations)
 // offset to the arena in which we're allocating is passed in the second parameter, and the
@@ -198,6 +204,10 @@ JL_DLLEXPORT void *jl_gc_perm_alloc(size_t sz, int zero, unsigned align,
 //              the allocated object. All objects stored in fields of this object
 //              must be either permanently allocated or have other roots.
 struct _jl_value_t *jl_gc_permobj(size_t sz, void *ty) JL_NOTSAFEPOINT;
+// This function notifies the GC about memory addresses that are set when loading the boot image.
+// The GC may use that information to, for instance, determine that such objects should
+// be treated as marked and belonged to the old generation in nursery collections.
+void jl_gc_notify_image_load(const char* img_data, size_t len);
 
 // ========================================================================= //
 // Runtime Write-Barriers