add edges metadata field to CodeInfo/CodeInstance, prepare for using

This records all invoke targets as edges as a functionality test, before
finishing the implementation of recording the edges accurately during
inference (via backedges + inference).

Author: vtjnash

base/boot.jl

@@ -510,11 +510,11 @@ function CodeInstance(
     mi::MethodInstance, owner, @nospecialize(rettype), @nospecialize(exctype), @nospecialize(inferred_const),
     @nospecialize(inferred), const_flags::Int32, min_world::UInt, max_world::UInt,
     effects::UInt32, @nospecialize(analysis_results),
-    relocatability::UInt8, edges::Union{DebugInfo,Nothing})
+    relocatability::UInt8, di::Union{DebugInfo,Nothing}, edges::SimpleVector)
     return ccall(:jl_new_codeinst, Ref{CodeInstance},
-        (Any, Any, Any, Any, Any, Any, Int32, UInt, UInt, UInt32, Any, UInt8, Any),
+        (Any, Any, Any, Any, Any, Any, Int32, UInt, UInt, UInt32, Any, UInt8, Any, Any),
         mi, owner, rettype, exctype, inferred_const, inferred, const_flags, min_world, max_world,
-        effects, analysis_results, relocatability, edges)
+        effects, analysis_results, relocatability, di, edges)
 end
 GlobalRef(m::Module, s::Symbol) = ccall(:jl_module_globalref, Ref{GlobalRef}, (Any, Any), m, s)
 Module(name::Symbol=:anonymous, std_imports::Bool=true, default_names::Bool=true) = ccall(:jl_f_new_module, Ref{Module}, (Any, Bool, Bool), name, std_imports, default_names)

base/compiler/typeinfer.jl

@@ -224,13 +224,16 @@ function finish!(interp::AbstractInterpreter, caller::InferenceState;
     if isdefined(result, :ci)
         ci = result.ci
         inferred_result = nothing
+        edges = Core.svec() # This should be a computed input, for now it is approximated (badly) here
         relocatability = 0x1
         const_flag = is_result_constabi_eligible(result)
         if !can_discard_trees || (is_cached(caller) && !const_flag)
             inferred_result = transform_result_for_cache(interp, result.linfo, result.valid_worlds, result, can_discard_trees)
             relocatability = 0x0
             if inferred_result isa CodeInfo
-                edges = inferred_result.debuginfo
+                edges = ccall(:jl_ir_edges_legacy, Any, (Any,), inferred_result.code)
+                inferred_result.edges = edges
+                di = inferred_result.debuginfo
                 uncompressed = inferred_result
                 inferred_result = maybe_compress_codeinfo(interp, result.linfo, inferred_result, can_discard_trees)
                 result.is_src_volatile |= uncompressed !== inferred_result
@@ -245,14 +248,14 @@ function finish!(interp::AbstractInterpreter, caller::InferenceState;
             end
         end
         # n.b. relocatability = isa(inferred_result, String) && inferred_result[end]
-        if !@isdefined edges
-            edges = DebugInfo(result.linfo)
+        if !@isdefined di
+            di = DebugInfo(result.linfo)
         end
-        ccall(:jl_update_codeinst, Cvoid, (Any, Any, Int32, UInt, UInt, UInt32, Any, UInt8, Any),
+        ccall(:jl_update_codeinst, Cvoid, (Any, Any, Int32, UInt, UInt, UInt32, Any, UInt8, Any, Any),
                 ci, inferred_result, const_flag,
                 first(result.valid_worlds), last(result.valid_worlds),
                 encode_effects(result.ipo_effects), result.analysis_results,
-                relocatability, edges)
+                relocatability, di, edges)
         engine_reject(interp, ci)
     end
     return nothing
@@ -334,7 +337,6 @@ function is_result_constabi_eligible(result::InferenceResult)
     return isa(result_type, Const) && is_foldable_nothrow(result.ipo_effects) && is_inlineable_constant(result_type.val)
 end
 
-
 function transform_result_for_cache(interp::AbstractInterpreter,
         ::MethodInstance, valid_worlds::WorldRange, result::InferenceResult,
         can_discard_trees::Bool=may_discard_trees(interp))
@@ -528,7 +530,7 @@ function finishinfer!(me::InferenceState, interp::AbstractInterpreter)
         append!(s_edges, edges)
         empty!(edges)
     end
-    if me.src.edges !== nothing
+    if me.src.edges !== nothing && me.src.edges !== Core.svec()
         append!(s_edges, me.src.edges::Vector)
     end
     # inspect whether our inference had a limited result accuracy,
@@ -613,11 +615,12 @@ function finishinfer!(me::InferenceState, interp::AbstractInterpreter)
             const_flags = 0x00
         end
         relocatability = 0x0
-        edges = nothing
-        ccall(:jl_fill_codeinst, Cvoid, (Any, Any, Any, Any, Int32, UInt, UInt, UInt32, Any, Any),
+        di = nothing
+        edges = Core.svec()
+        ccall(:jl_fill_codeinst, Cvoid, (Any, Any, Any, Any, Int32, UInt, UInt, UInt32, Any, Any, Any),
                 result.ci, widenconst(result_type), widenconst(result.exc_result), rettype_const, const_flags,
                 first(result.valid_worlds), last(result.valid_worlds),
-                encode_effects(result.ipo_effects), result.analysis_results, edges)
+                encode_effects(result.ipo_effects), result.analysis_results, di, edges)
         if is_cached(me)
             cached_results = cache_result!(me.interp, me.result)
             if !cached_results
@@ -991,6 +994,7 @@ function codeinfo_for_const(interp::AbstractInterpreter, mi::MethodInstance, @no
     tree.debuginfo = DebugInfo(mi)
     tree.ssaflags = UInt32[0]
     tree.rettype = Core.Typeof(val)
+    tree.edges = Core.svec()
     set_inlineable!(tree, true)
     tree.parent = mi
     return tree
@@ -1008,7 +1012,7 @@ function codeinstance_for_const_with_code(interp::AbstractInterpreter, code::Cod
     return CodeInstance(code.def, cache_owner(interp), code.rettype, code.exctype, code.rettype_const, src,
         Int32(0x3), code.min_world, code.max_world,
         code.ipo_purity_bits, code.analysis_results,
-        code.relocatability, src.debuginfo)
+        code.relocatability, src.debuginfo, src.edges)
 end
 
 result_is_constabi(interp::AbstractInterpreter, result::InferenceResult,
@@ -1170,7 +1174,7 @@ function typeinf_ext(interp::AbstractInterpreter, mi::MethodInstance, source_mod
             src isa CodeInfo || return nothing
             return CodeInstance(mi, cache_owner(interp), Any, Any, nothing, src, Int32(0),
                 get_inference_world(interp), get_inference_world(interp),
-                UInt32(0), nothing, UInt8(0), src.debuginfo)
+                UInt32(0), nothing, UInt8(0), src.debuginfo, src.edges)
         end
     end
     ci = engine_reserve(interp, mi)

base/expr.jl

@@ -79,7 +79,7 @@ function copy(c::CodeInfo)
         cnew.slottypes = copy(cnew.slottypes::Vector{Any})
     end
     cnew.ssaflags  = copy(cnew.ssaflags)
-    cnew.edges     = cnew.edges === nothing ? nothing : copy(cnew.edges::Vector)
+    cnew.edges     = cnew.edges === nothing || cnew.edges isa Core.SimpleVector ? cnew.edges : copy(cnew.edges::Vector)
     ssavaluetypes  = cnew.ssavaluetypes
     ssavaluetypes isa Vector{Any} && (cnew.ssavaluetypes = copy(ssavaluetypes))
     return cnew

src/common_symbols1.inc

@@ -88,5 +88,3 @@ jl_symbol("ifelse"),
 jl_symbol("Array"),
 jl_symbol("eq_int"),
 jl_symbol("throw_inexacterror"),
-jl_symbol("|"),
-jl_symbol("setproperty!"),

src/common_symbols2.inc

@@ -1,3 +1,5 @@
+jl_symbol("|"),
+jl_symbol("setproperty!"),
 jl_symbol("sext_int"),
 jl_symbol("String"),
 jl_symbol("Int"),
@@ -244,5 +246,3 @@ jl_symbol("invokelatest"),
 jl_symbol("jl_array_del_end"),
 jl_symbol("_mod64"),
 jl_symbol("parameters"),
-jl_symbol("monotonic"),
-jl_symbol("regex.jl"),

src/gf.c

@@ -322,7 +322,7 @@ jl_datatype_t *jl_mk_builtin_func(jl_datatype_t *dt, const char *name, jl_fptr_a
 
     jl_code_instance_t *codeinst = jl_new_codeinst(mi, jl_nothing,
         (jl_value_t*)jl_any_type, (jl_value_t*)jl_any_type, jl_nothing, jl_nothing,
-        0, 1, ~(size_t)0, 0, jl_nothing, 0, NULL);
+        0, 1, ~(size_t)0, 0, jl_nothing, 0, NULL, NULL);
     jl_mi_cache_insert(mi, codeinst);
     jl_atomic_store_relaxed(&codeinst->specptr.fptr1, fptr);
     jl_atomic_store_relaxed(&codeinst->invoke, jl_fptr_args);
@@ -480,7 +480,7 @@ JL_DLLEXPORT jl_value_t *jl_call_in_typeinf_world(jl_value_t **args, int nargs)
 
 JL_DLLEXPORT jl_code_instance_t *jl_get_method_inferred(
         jl_method_instance_t *mi JL_PROPAGATES_ROOT, jl_value_t *rettype,
-        size_t min_world, size_t max_world, jl_debuginfo_t *edges)
+        size_t min_world, size_t max_world, jl_debuginfo_t *di, jl_svec_t *edges)
 {
     jl_value_t *owner = jl_nothing; // TODO: owner should be arg
     jl_code_instance_t *codeinst = jl_atomic_load_relaxed(&mi->cache);
@@ -489,27 +489,30 @@ JL_DLLEXPORT jl_code_instance_t *jl_get_method_inferred(
             jl_atomic_load_relaxed(&codeinst->max_world) == max_world &&
             jl_egal(codeinst->owner, owner) &&
             jl_egal(codeinst->rettype, rettype)) {
-            if (edges == NULL)
+            if (di == NULL)
                 return codeinst;
             jl_debuginfo_t *debuginfo = jl_atomic_load_relaxed(&codeinst->debuginfo);
-            if (edges == debuginfo)
-                return codeinst;
-            if (debuginfo == NULL && jl_atomic_cmpswap_relaxed(&codeinst->debuginfo, &debuginfo, edges))
-                return codeinst;
-            if (debuginfo && jl_egal((jl_value_t*)debuginfo, (jl_value_t*)edges))
+            if (di != debuginfo) {
+                if (!(debuginfo == NULL && jl_atomic_cmpswap_relaxed(&codeinst->debuginfo, &debuginfo, di)))
+                    if (!(debuginfo && jl_egal((jl_value_t*)debuginfo, (jl_value_t*)di)))
+                        continue;
+            }
+            // TODO: this is implied by the matching worlds, since it is intrinsic, so do we really need to verify it?
+            jl_svec_t *e = jl_atomic_load_relaxed(&codeinst->edges);
+            if (e && jl_egal((jl_value_t*)e, (jl_value_t*)edges))
                 return codeinst;
         }
         codeinst = jl_atomic_load_relaxed(&codeinst->next);
     }
     codeinst = jl_new_codeinst(
         mi, owner, rettype, (jl_value_t*)jl_any_type, NULL, NULL,
-        0, min_world, max_world, 0, jl_nothing, 0, edges);
+        0, min_world, max_world, 0, jl_nothing, 0, di, edges);
     jl_mi_cache_insert(mi, codeinst);
     return codeinst;
 }
 
 JL_DLLEXPORT int jl_mi_cache_has_ci(jl_method_instance_t *mi,
-                                     jl_code_instance_t *ci)
+                                    jl_code_instance_t *ci)
 {
     jl_code_instance_t *codeinst = jl_atomic_load_relaxed(&mi->cache);
     while (codeinst) {
@@ -527,14 +530,15 @@ JL_DLLEXPORT jl_code_instance_t *jl_new_codeinst(
         int32_t const_flags, size_t min_world, size_t max_world,
         uint32_t effects, jl_value_t *analysis_results,
         uint8_t relocatability,
-        jl_debuginfo_t *edges /* , int absolute_max*/)
+        jl_debuginfo_t *di, jl_svec_t *edges /*, int absolute_max*/)
 {
     jl_task_t *ct = jl_current_task;
     assert(min_world <= max_world && "attempting to set invalid world constraints");
     jl_code_instance_t *codeinst = (jl_code_instance_t*)jl_gc_alloc(ct->ptls, sizeof(jl_code_instance_t),
             jl_code_instance_type);
     codeinst->def = mi;
     codeinst->owner = owner;
+    jl_atomic_store_relaxed(&codeinst->edges, edges);
     jl_atomic_store_relaxed(&codeinst->min_world, min_world);
     jl_atomic_store_relaxed(&codeinst->max_world, max_world);
     codeinst->rettype = rettype;
@@ -543,7 +547,7 @@ JL_DLLEXPORT jl_code_instance_t *jl_new_codeinst(
     if ((const_flags & 2) == 0)
         inferred_const = NULL;
     codeinst->rettype_const = inferred_const;
-    jl_atomic_store_relaxed(&codeinst->debuginfo, (jl_value_t*)edges == jl_nothing ? NULL : edges);
+    jl_atomic_store_relaxed(&codeinst->debuginfo, (jl_value_t*)di == jl_nothing ? NULL : di);
     jl_atomic_store_relaxed(&codeinst->specptr.fptr, NULL);
     jl_atomic_store_relaxed(&codeinst->invoke, NULL);
     if ((const_flags & 1) != 0) {
@@ -563,13 +567,15 @@ JL_DLLEXPORT void jl_update_codeinst(
         jl_code_instance_t *codeinst, jl_value_t *inferred,
         int32_t const_flags, size_t min_world, size_t max_world,
         uint32_t effects, jl_value_t *analysis_results,
-        uint8_t relocatability, jl_debuginfo_t *edges /* , int absolute_max*/)
+        uint8_t relocatability, jl_debuginfo_t *di, jl_svec_t *edges /* , int absolute_max*/)
 {
     codeinst->relocatability = relocatability;
     codeinst->analysis_results = analysis_results;
     jl_gc_wb(codeinst, analysis_results);
     jl_atomic_store_relaxed(&codeinst->ipo_purity_bits, effects);
-    jl_atomic_store_relaxed(&codeinst->debuginfo, edges);
+    jl_atomic_store_relaxed(&codeinst->debuginfo, di);
+    jl_gc_wb(codeinst, di);
+    jl_atomic_store_relaxed(&codeinst->edges, edges);
     jl_gc_wb(codeinst, edges);
     if ((const_flags & 1) != 0) {
         assert(codeinst->rettype_const);
@@ -587,7 +593,7 @@ JL_DLLEXPORT void jl_fill_codeinst(
         jl_value_t *inferred_const,
         int32_t const_flags, size_t min_world, size_t max_world,
         uint32_t effects, jl_value_t *analysis_results,
-        jl_debuginfo_t *edges /* , int absolute_max*/)
+        jl_debuginfo_t *di, jl_svec_t *edges /* , int absolute_max*/)
 {
     assert(min_world <= max_world && "attempting to set invalid world constraints");
     codeinst->rettype = rettype;
@@ -598,8 +604,12 @@ JL_DLLEXPORT void jl_fill_codeinst(
         codeinst->rettype_const = inferred_const;
         jl_gc_wb(codeinst, inferred_const);
     }
-    jl_atomic_store_relaxed(&codeinst->debuginfo, (jl_value_t*)edges == jl_nothing ? NULL : edges);
+    jl_atomic_store_relaxed(&codeinst->edges, edges);
     jl_gc_wb(codeinst, edges);
+    if ((jl_value_t*)di != jl_nothing) {
+        jl_atomic_store_relaxed(&codeinst->debuginfo, di);
+        jl_gc_wb(codeinst, di);
+    }
     if ((const_flags & 1) != 0) {
         // TODO: may want to follow ordering restrictions here (see jitlayers.cpp)
         assert(const_flags & 2);
@@ -615,7 +625,7 @@ JL_DLLEXPORT void jl_fill_codeinst(
 
 JL_DLLEXPORT jl_code_instance_t *jl_new_codeinst_uninit(jl_method_instance_t *mi, jl_value_t *owner)
 {
-    jl_code_instance_t *codeinst = jl_new_codeinst(mi, owner, NULL, NULL, NULL, NULL, 0, 0, 0, 0, NULL, 0, NULL);
+    jl_code_instance_t *codeinst = jl_new_codeinst(mi, owner, NULL, NULL, NULL, NULL, 0, 0, 0, 0, NULL, 0, NULL, NULL);
     jl_atomic_store_relaxed(&codeinst->min_world, 1); // make temporarily invalid before returning, so that jl_fill_codeinst is valid later
     return codeinst;
 }
@@ -2563,8 +2573,10 @@ jl_code_instance_t *jl_compile_method_internal(jl_method_instance_t *mi, size_t
         jl_code_instance_t *codeinst2 = jl_compile_method_internal(mi2, world);
         jl_code_instance_t *codeinst = jl_get_method_inferred(
                 mi, codeinst2->rettype,
-                jl_atomic_load_relaxed(&codeinst2->min_world), jl_atomic_load_relaxed(&codeinst2->max_world),
-                jl_atomic_load_relaxed(&codeinst2->debuginfo));
+                jl_atomic_load_relaxed(&codeinst2->min_world),
+                jl_atomic_load_relaxed(&codeinst2->max_world),
+                jl_atomic_load_relaxed(&codeinst2->debuginfo),
+                jl_atomic_load_relaxed(&codeinst2->edges));
         if (jl_atomic_load_relaxed(&codeinst->invoke) == NULL) {
             codeinst->rettype_const = codeinst2->rettype_const;
             jl_gc_wb(codeinst, codeinst->rettype_const);
@@ -2623,7 +2635,7 @@ jl_code_instance_t *jl_compile_method_internal(jl_method_instance_t *mi, size_t
                 if (unspec && (unspec_invoke = jl_atomic_load_acquire(&unspec->invoke))) {
                     jl_code_instance_t *codeinst = jl_new_codeinst(mi, jl_nothing,
                         (jl_value_t*)jl_any_type, (jl_value_t*)jl_any_type, NULL, NULL,
-                        0, 1, ~(size_t)0, 0, jl_nothing, 0, NULL);
+                        0, 1, ~(size_t)0, 0, jl_nothing, 0, NULL, NULL);
                     void *unspec_fptr = jl_atomic_load_relaxed(&unspec->specptr.fptr);
                     if (unspec_fptr) {
                         // wait until invoke and specsigflags are properly set
@@ -2650,7 +2662,7 @@ jl_code_instance_t *jl_compile_method_internal(jl_method_instance_t *mi, size_t
         if (!jl_code_requires_compiler(src, 0)) {
             jl_code_instance_t *codeinst = jl_new_codeinst(mi, jl_nothing,
                 (jl_value_t*)jl_any_type, (jl_value_t*)jl_any_type, NULL, NULL,
-                0, 1, ~(size_t)0, 0, jl_nothing, 0, NULL);
+                0, 1, ~(size_t)0, 0, jl_nothing, 0, NULL, NULL);
             jl_atomic_store_release(&codeinst->invoke, jl_fptr_interpret_call);
             jl_mi_cache_insert(mi, codeinst);
             record_precompile_statement(mi, 0);
@@ -2710,7 +2722,7 @@ jl_code_instance_t *jl_compile_method_internal(jl_method_instance_t *mi, size_t
         jl_method_instance_t *unspec = jl_get_unspecialized(def);
         if (unspec == NULL)
             unspec = mi;
-        jl_code_instance_t *ucache = jl_get_method_inferred(unspec, (jl_value_t*)jl_any_type, 1, ~(size_t)0, NULL);
+        jl_code_instance_t *ucache = jl_get_method_inferred(unspec, (jl_value_t*)jl_any_type, 1, ~(size_t)0, NULL, NULL);
         // ask codegen to make the fptr for unspec
         jl_callptr_t ucache_invoke = jl_atomic_load_acquire(&ucache->invoke);
         if (ucache_invoke == NULL) {
@@ -2730,7 +2742,7 @@ jl_code_instance_t *jl_compile_method_internal(jl_method_instance_t *mi, size_t
         }
         codeinst = jl_new_codeinst(mi, jl_nothing,
             (jl_value_t*)jl_any_type, (jl_value_t*)jl_any_type, NULL, NULL,
-            0, 1, ~(size_t)0, 0, jl_nothing, 0, NULL);
+            0, 1, ~(size_t)0, 0, jl_nothing, 0, NULL, NULL);
         void *unspec_fptr = jl_atomic_load_relaxed(&ucache->specptr.fptr);
         if (unspec_fptr) {
             // wait until invoke and specsigflags are properly set