Adding support for MMTk (non-moving immix) (#56288)

This PR adds the possibility of building/running Julia using MMTk,
running non-moving immix.
The binding code (Rust) associated with it is in
https://github.com/mmtk/mmtk-julia/tree/upstream-ready/immix.
Instructions on how to build/run Julia with MMTk are described in the
[`README`](https://github.com/mmtk/mmtk-julia/tree/upstream-ready/immix?tab=readme-ov-file#an-mmtk-binding-for-the-julia-programming-language)
file inside the binding repo.

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 +836,31 @@ JCXXFLAGS += -DGC_DEBUG_ENV
 JCFLAGS += -DGC_DEBUG_ENV
 endif
 
+ifneq (${MMTK_PLAN},None)
+JCXXFLAGS += -DMMTK_GC
+JCFLAGS += -DMMTK_GC
+# Do a release build on the binding by default
+MMTK_BUILD ?= release
+ifeq (${MMTK_PLAN},Immix)
+JCXXFLAGS += -DMMTK_PLAN_IMMIX
+JCFLAGS += -DMMTK_PLAN_IMMIX
+else
+$(error "Unsupported MMTk plan: $(MMTK_PLAN)")
+endif
+
+# Location of mmtk-julia binding
+# (needed for api/*.h and .so file)
+MMTK_JULIA_DIR ?= $(BUILDROOT)/usr/lib/mmtk_julia
+
+MMTK_DIR = ${MMTK_JULIA_DIR}/mmtk
+MMTK_API_INC = ${MMTK_DIR}/api
+
+MMTK_LIB := -lmmtk_julia
+else
+MMTK_JULIA_INC :=
+MMTK_LIB :=
+endif
+
 ifeq ($(WITH_DTRACE), 1)
 JCXXFLAGS += -DUSE_DTRACE
 JCFLAGS += -DUSE_DTRACE
@@ -933,6 +961,21 @@ ARCH := $(BUILD_OS)
 endif
 endif
 
+# MMTk is only available on x86_64 Linux for now
+ifneq (${MMTK_PLAN},None)
+
+ifeq ($(OS),Linux)
+MMTK_LIB_NAME := libmmtk_julia.so
+else
+$(error "Unsupported OS for MMTk")
+endif
+
+ifneq ($(ARCH),x86_64)
+$(error "Unsupported build architecture for MMTk")
+endif
+
+endif
+
 # Detect common pre-SSE2 JULIA_CPU_TARGET values known not to work (#7185)
 ifeq ($(MARCH),)
 ifneq ($(findstring $(ARCH),i386 i486 i586 i686),)

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

deps/Makefile

@@ -26,7 +26,7 @@ BUILDDIR := $(BUILDDIR)$(MAYBE_HOST)
 # custom Makefile rules: openlibm dsfmt libsuitesparse lapack blastrampoline openblas utf8proc objconv libwhich
 # CMake libs: llvm llvmunwind libgit2 libssh2 libtracyclient
 #
-# downloadable via git: llvm-svn, libuv, libopenlibm, utf8proc, libgit2, libssh2, libtracyclient
+# downloadable via git: llvm-svn, libuv, libopenlibm, utf8proc, libgit2, libssh2, libtracyclient, mmtk_julia
 #
 # to debug 'define' rules, replace eval at the usage site with info or error
 
@@ -195,14 +195,18 @@ DEP_LIBS += libwhich
 endif
 endif
 
+ifneq (${MMTK_PLAN},None)
+DEP_LIBS += mmtk_julia
+endif
+
 DEP_LIBS_STAGED := $(DEP_LIBS)
 
 # list all targets
 DEP_LIBS_STAGED_ALL := llvm llvm-tools clang llvmunwind unwind libuv pcre \
 	openlibm dsfmt blastrampoline openblas lapack gmp mpfr patchelf utf8proc \
 	objconv openssl libssh2 nghttp2 curl libgit2 libwhich zlib p7zip csl \
 	sanitizers libsuitesparse lld libtracyclient ittapi nvtx JuliaSyntax \
-	terminfo
+	terminfo mmtk_julia
 DEP_LIBS_ALL := $(DEP_LIBS_STAGED_ALL)
 
 ifneq ($(USE_BINARYBUILDER_OPENBLAS),0)
@@ -266,6 +270,9 @@ include $(SRCDIR)/p7zip.mk
 include $(SRCDIR)/libtracyclient.mk
 include $(SRCDIR)/terminfo.mk
 
+# MMTk
+include $(SRCDIR)/mmtk_julia.mk
+
 # vendored Julia libs
 include $(SRCDIR)/JuliaSyntax.mk
 

deps/checksums/mmtk_julia-b69acf5af7a7dd97c1cc6fd99f7c2d51b477f214.tar.gz/md5

@@ -0,0 +1 @@
+1911cf084d26c48e2ed58af3d268b4b6

deps/checksums/mmtk_julia-b69acf5af7a7dd97c1cc6fd99f7c2d51b477f214.tar.gz/sha512

@@ -0,0 +1 @@
+75beab54398989c46b62e714b242cf6705d88d220f40c21e494e0f29161437f5fbe9ba05b543d2353a1ad76f4239ac4025b476be0be864649f310f14935289fe

deps/mmtk_julia.mk

@@ -0,0 +1,71 @@
+## MMTK ##
+
+# Both MMTK_MOVING and MMTK_PLAN should be specified in the Make.user file.
+# At this point, since we only support non-moving this is always set to 0
+# FIXME: change it to `?:` when introducing moving plans
+MMTK_MOVING := 0
+MMTK_VARS := MMTK_PLAN=$(MMTK_PLAN) MMTK_MOVING=$(MMTK_MOVING)
+
+# Download the binding, build it from source
+ifeq (${MMTK_JULIA_DIR},$(BUILDROOT)/usr/lib/mmtk_julia)
+$(eval $(call git-external,mmtk_julia,MMTK_JULIA,,,$(BUILDDIR)))
+
+MMTK_JULIA_DIR=$(BUILDROOT)/deps/$(BUILDDIR)/$(MMTK_JULIA_SRC_DIR)
+MMTK_JULIA_LIB_PATH=$(MMTK_JULIA_DIR)/mmtk/target/$(MMTK_BUILD)
+PROJECT_DIRS := JULIA_PATH=$(JULIAHOME) JULIA_BUILDROOT=$(BUILDROOT) MMTK_JULIA_DIR=$(MMTK_JULIA_DIR)
+
+$(BUILDDIR)/$(MMTK_JULIA_SRC_DIR)/build-compiled: $(BUILDROOT)/usr/lib/libmmtk_julia.so
+	@echo 1 > $@
+
+# NB: use the absolute dir when creating the symlink
+$(BUILDROOT)/usr/lib/libmmtk_julia.so: $(MMTK_JULIA_LIB_PATH)/libmmtk_julia.so
+	@ln -sf $(MMTK_JULIA_LIB_PATH)/libmmtk_julia.so $@
+
+$(MMTK_JULIA_LIB_PATH)/libmmtk_julia.so: $(BUILDDIR)/$(MMTK_JULIA_SRC_DIR)/source-extracted
+	@$(PROJECT_DIRS) $(MMTK_VARS) $(MAKE) -C $(MMTK_JULIA_DIR) $(MMTK_BUILD)
+
+get-mmtk_julia: $(MMTK_JULIA_SRC_FILE)
+extract-mmtk_julia: $(BUILDDIR)/$(MMTK_JULIA_SRC_DIR)/source-extracted
+configure-mmtk_julia: extract-mmtk_julia
+compile-mmtk_julia: $(BUILDROOT)/usr/lib/libmmtk_julia.so
+fastcheck-mmtk_julia: #none
+check-mmtk_julia: compile-mmtk_julia
+
+$(eval $(call symlink_install,mmtk_julia,$$(MMTK_JULIA_SRC_DIR),$$(BUILDROOT)/usr/lib))
+
+# In this case, there is a custom version of the binding in MMTK_JULIA_DIR
+# Build it and symlink libmmtk_julia.so file into $(BUILDROOT)/usr/lib
+else
+
+PROJECT_DIRS := JULIA_PATH=$(JULIAHOME) JULIA_BUILDROOT=$(BUILDROOT) MMTK_JULIA_DIR=$(MMTK_JULIA_DIR)
+MMTK_JULIA_LIB_PATH=$(MMTK_JULIA_DIR)/mmtk/target/$(MMTK_BUILD)
+
+install-mmtk_julia: compile-mmtk_julia $(build_prefix)/manifest/mmtk_julia
+
+compile-mmtk_julia: $(BUILDROOT)/usr/lib/libmmtk_julia.so
+
+version-check-mmtk_julia: $(MMTK_JULIA_DIR)/mmtk/target/$(MMTK_BUILD)/libmmtk_julia.so
+
+# NB: This will NOT run `cargo build` if there are changes in the Rust source files
+# inside the binding repo. However the target below should remake the symlink if there
+# are changes in the libmmtk_julia.so from the custom MMTK_JULIA_DIR folder
+$(BUILDROOT)/usr/lib/libmmtk_julia.so: $(MMTK_JULIA_DIR)/mmtk/target/$(MMTK_BUILD)/libmmtk_julia.so
+	@ln -sf $(MMTK_JULIA_DIR)/mmtk/target/$(MMTK_BUILD)/libmmtk_julia.so $@
+
+$(MMTK_JULIA_DIR)/mmtk/target/$(MMTK_BUILD)/libmmtk_julia.so:
+	@$(PROJECT_DIRS) $(MMTK_VARS) $(MAKE) -C $(MMTK_JULIA_DIR) $(MMTK_BUILD)
+
+MMTK_JULIA_VER := mmtk_julia_custom
+
+UNINSTALL_mmtk_julia := $(MMTK_JULIA_VER) manual_mmtk_julia
+
+define manual_mmtk_julia
+uninstall-mmtk_julia:
+	-rm -f $(build_prefix)/manifest/mmtk_julia
+	-rm -f $(BUILDROOT)/usr/lib/libmmtk_julia.so
+endef
+
+$(build_prefix)/manifest/mmtk_julia: $(BUILDROOT)/usr/lib/libmmtk_julia.so
+	@echo $(UNINSTALL_mmtk_julia) > $@
+
+endif # MMTK_JULIA_DIR

deps/mmtk_julia.version

@@ -0,0 +1,4 @@
+MMTK_JULIA_BRANCH = master
+MMTK_JULIA_SHA1 = b69acf5af7a7dd97c1cc6fd99f7c2d51b477f214
+MMTK_JULIA_GIT_URL := https://github.com/mmtk/mmtk-julia.git
+MMTK_JULIA_TAR_URL = https://github.com/mmtk/mmtk-julia/archive/refs/tags/v0.30.2.tar.gz

src/Makefile

@@ -29,6 +29,10 @@ ifeq ($(USECLANG),1)
 FLAGS += -Wno-return-type-c-linkage -Wno-atomic-alignment
 endif
 
+ifneq (${MMTK_PLAN},None)
+FLAGS += -I$(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)
@@ -314,6 +341,7 @@ $(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

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