Add !noalias and !alias.scope metadata

The main idea here is that the TBAA domain is ill-equipped for
reasoning about regions (and, in particular, suffers total
precision less when merging disparate types in a `memcpy`).
Instead, `!noalias` should be used for region-based memory
information and `!tbaa` should be used exclusively for layout.

We use (5) regions corresponding to the top level of the TBAA tree:
  - gcframe
  - stack
  - data
  - constant
  - type_metadata

For now, this leaves the TBAA hierarchy in tact and only adds
additional `!noalias` metadata. `!tbaa` annotations should be
the same as before.

Author: topolarity

src/cgutils.cpp

@@ -890,8 +890,8 @@ static Value *data_pointer(jl_codectx_t &ctx, const jl_cgval_t &x)
     return data;
 }
 
-static void emit_memcpy_llvm(jl_codectx_t &ctx, Value *dst, MDNode *tbaa_dst, Value *src, MDNode *tbaa_src,
-                             uint64_t sz, unsigned align, bool is_volatile)
+static void emit_memcpy_llvm(jl_codectx_t &ctx, Value *dst, jl_aliasinfo_t const &dst_ai, Value *src,
+                             jl_aliasinfo_t const &src_ai, uint64_t sz, unsigned align, bool is_volatile)
 {
     if (sz == 0)
         return;
@@ -933,44 +933,73 @@ static void emit_memcpy_llvm(jl_codectx_t &ctx, Value *dst, MDNode *tbaa_dst, Va
             src = emit_bitcast(ctx, src, dstty);
         }
         if (directel) {
-            auto val = tbaa_decorate(tbaa_src, ctx.builder.CreateAlignedLoad(directel, src, Align(align), is_volatile));
-            tbaa_decorate(tbaa_dst, ctx.builder.CreateAlignedStore(val, dst, Align(align), is_volatile));
+            auto val = src_ai.decorateInst(ctx.builder.CreateAlignedLoad(directel, src, Align(align), is_volatile));
+            dst_ai.decorateInst(ctx.builder.CreateAlignedStore(val, dst, Align(align), is_volatile));
             ++SkippedMemcpys;
             return;
         }
     }
 #endif
+    ++EmittedMemcpys;
+
     // the memcpy intrinsic does not allow to specify different alias tags
     // for the load part (x.tbaa) and the store part (ctx.tbaa().tbaa_stack).
     // since the tbaa lattice has to be a tree we have unfortunately
     // x.tbaa ∪ ctx.tbaa().tbaa_stack = tbaa_root if x.tbaa != ctx.tbaa().tbaa_stack
-    ++EmittedMemcpys;
-    ctx.builder.CreateMemCpy(dst, MaybeAlign(align), src, MaybeAlign(0), sz, is_volatile, MDNode::getMostGenericTBAA(tbaa_dst, tbaa_src));
+
+    // Now that we use scoped aliases to label disparate regions of memory, the TBAA
+    // metadata should be revisited so that it only represents memory layouts. Once
+    // that's done, we can expect that in most cases tbaa(src) == tbaa(dst) and the
+    // above problem won't be as serious.
+
+    auto merged_ai = dst_ai.merge(src_ai);
+    ctx.builder.CreateMemCpy(dst, MaybeAlign(align), src, MaybeAlign(0), sz, is_volatile,
+                             merged_ai.tbaa, merged_ai.tbaa_struct, merged_ai.scope, merged_ai.noalias);
 }
 
-static void emit_memcpy_llvm(jl_codectx_t &ctx, Value *dst, MDNode *tbaa_dst, Value *src, MDNode *tbaa_src,
-                             Value *sz, unsigned align, bool is_volatile)
+static void emit_memcpy_llvm(jl_codectx_t &ctx, Value *dst, jl_aliasinfo_t const &dst_ai, Value *src,
+                             jl_aliasinfo_t const &src_ai, Value *sz, unsigned align, bool is_volatile)
 {
     if (auto const_sz = dyn_cast<ConstantInt>(sz)) {
-        emit_memcpy_llvm(ctx, dst, tbaa_dst, src, tbaa_src, const_sz->getZExtValue(), align, is_volatile);
+        emit_memcpy_llvm(ctx, dst, dst_ai, src, src_ai, const_sz->getZExtValue(), align, is_volatile);
         return;
     }
     ++EmittedMemcpys;
-    ctx.builder.CreateMemCpy(dst, MaybeAlign(align), src, MaybeAlign(0), sz, is_volatile, MDNode::getMostGenericTBAA(tbaa_dst, tbaa_src));
+
+    auto merged_ai = dst_ai.merge(src_ai);
+    ctx.builder.CreateMemCpy(dst, MaybeAlign(align), src, MaybeAlign(0), sz, is_volatile,
+                             merged_ai.tbaa, merged_ai.tbaa_struct, merged_ai.scope, merged_ai.noalias);
 }
 
 template<typename T1>
 static void emit_memcpy(jl_codectx_t &ctx, Value *dst, MDNode *tbaa_dst, Value *src, MDNode *tbaa_src,
                         T1 &&sz, unsigned align, bool is_volatile=false)
 {
-    emit_memcpy_llvm(ctx, dst, tbaa_dst, src, tbaa_src, sz, align, is_volatile);
+    emit_memcpy_llvm(ctx, dst, jl_aliasinfo_t::fromTBAA(ctx, tbaa_dst), src,
+                     jl_aliasinfo_t::fromTBAA(ctx, tbaa_src), sz, align, is_volatile);
 }
 
 template<typename T1>
 static void emit_memcpy(jl_codectx_t &ctx, Value *dst, MDNode *tbaa_dst, const jl_cgval_t &src,
                         T1 &&sz, unsigned align, bool is_volatile=false)
 {
-    emit_memcpy_llvm(ctx, dst, tbaa_dst, data_pointer(ctx, src), src.tbaa, sz, align, is_volatile);
+    emit_memcpy_llvm(ctx, dst, jl_aliasinfo_t::fromTBAA(ctx, tbaa_dst), data_pointer(ctx, src),
+                     jl_aliasinfo_t::fromTBAA(ctx, src.tbaa), sz, align, is_volatile);
+}
+
+template<typename T1>
+static void emit_memcpy(jl_codectx_t &ctx, Value *dst, jl_aliasinfo_t const &dst_ai, Value *src,
+                        jl_aliasinfo_t const &src_ai, T1 &&sz, unsigned align, bool is_volatile=false)
+{
+    emit_memcpy_llvm(ctx, dst, dst_ai, src, src_ai, sz, align, is_volatile);
+}
+
+template<typename T1>
+static void emit_memcpy(jl_codectx_t &ctx, Value *dst, jl_aliasinfo_t const &dst_ai, const jl_cgval_t &src,
+                        T1 &&sz, unsigned align, bool is_volatile=false)
+{
+    auto src_ai = jl_aliasinfo_t::fromTBAA(ctx, src.tbaa);
+    emit_memcpy_llvm(ctx, dst, dst_ai, data_pointer(ctx, src), src_ai, sz, align, is_volatile);
 }
 
 static LoadInst *emit_nthptr_recast(jl_codectx_t &ctx, Value *v, Value *idx, MDNode *tbaa, Type *type)
@@ -2699,7 +2728,9 @@ static Value *emit_arrayptr_internal(jl_codectx_t &ctx, const jl_cgval_t &tinfo,
     LoadInst *LI = ctx.builder.CreateAlignedLoad(LoadT, addr, Align(sizeof(char *)));
     LI->setOrdering(AtomicOrdering::NotAtomic);
     LI->setMetadata(LLVMContext::MD_nonnull, MDNode::get(ctx.builder.getContext(), None));
-    tbaa_decorate(arraytype_constshape(tinfo.typ) ? ctx.tbaa().tbaa_const : ctx.tbaa().tbaa_arrayptr, LI);
+    jl_aliasinfo_t aliasinfo = jl_aliasinfo_t::fromTBAA(ctx, arraytype_constshape(tinfo.typ) ? ctx.tbaa().tbaa_const : ctx.tbaa().tbaa_arrayptr);
+    aliasinfo.decorateInst(LI);
+
     return LI;
 }
 

src/codegen.cpp

@@ -356,6 +356,57 @@ struct jl_tbaacache_t {
     }
 };
 
+struct jl_noaliascache_t {
+    // Each domain operates completely independently.
+    // "No aliasing" is inferred if it is implied by any domain.
+
+    // memory regions domain
+    struct jl_regions_t {
+        MDNode *gcframe;        // GC frame
+        MDNode *stack;          // Stack slot
+        MDNode *data;           // Any user data that `pointerset/ref` are allowed to alias
+        MDNode *type_metadata;  // Non-user-accessible type metadata incl. size, union selectors, etc.
+        MDNode *constant;       // Memory that is immutable by the time LLVM can see it
+
+        jl_regions_t(): gcframe(nullptr), stack(nullptr), data(nullptr), type_metadata(nullptr), constant(nullptr) {}
+
+        void initialize(llvm::LLVMContext &context) {
+            MDBuilder mbuilder(context);
+            MDNode *domain = mbuilder.createAliasScopeDomain("jnoalias");
+
+            this->gcframe = mbuilder.createAliasScope("jnoalias_gcframe", domain);
+            this->stack = mbuilder.createAliasScope("jnoalias_stack", domain);
+            this->data = mbuilder.createAliasScope("jnoalias_data", domain);
+            this->type_metadata = mbuilder.createAliasScope("jnoalias_typemd", domain);
+            this->constant = mbuilder.createAliasScope("jnoalias_const", domain);
+        }
+    } regions;
+
+    // `@aliasscope` domain
+    struct jl_aliasscope_t {
+        MDNode *current;
+
+        jl_aliasscope_t(): current(nullptr) {}
+
+        // No init required, this->current is only used to store the currently active aliasscope
+        void initialize(llvm::LLVMContext &context) {}
+    } aliasscope;
+
+    bool initialized;
+
+    jl_noaliascache_t(): regions(), aliasscope(), initialized(false) {}
+
+    void initialize(llvm::LLVMContext &context) {
+        if (initialized) {
+            assert(&regions.constant->getContext() == &context);
+            return;
+        }
+        initialized = true;
+        regions.initialize(context);
+        aliasscope.initialize(context);
+    }
+};
+
 struct jl_debugcache_t {
     // Basic DITypes
     DIDerivedType *jl_pvalue_dillvmt;
@@ -1276,6 +1327,69 @@ static bool deserves_sret(jl_value_t *dt, Type *T)
     return (size_t)jl_datatype_size(dt) > sizeof(void*) && !T->isFloatingPointTy() && !T->isVectorTy();
 }
 
+// Alias Analysis Info (analogous to llvm::AAMDNodes)
+struct jl_aliasinfo_t {
+    MDNode *tbaa = nullptr;          // '!tbaa': Struct-path TBAA. TBAA graph forms a tree (indexed by offset). 
+                                     //          Two pointers do not alias if they are not transitive parents
+                                     //          (effectively, subfields) of each other or equal.
+    MDNode *tbaa_struct = nullptr;   // '!tbaa.struct': Describes memory layout of struct. 
+    MDNode *scope = nullptr;         // '!alias.scope': Generic "noalias" memory access sets.
+                                     //                 If alias.scope(inst_a) ⊆ noalias(inst_b) (in any "domain")
+                                     //                    => inst_a, inst_b do not alias.
+    MDNode *noalias = nullptr;       // '!noalias': See '!alias.scope' above.
+
+    enum class Region { unknown, gcframe, stack, data, constant, type_metadata }; // See jl_regions_t
+
+    explicit jl_aliasinfo_t() = default;
+    explicit jl_aliasinfo_t(jl_codectx_t &ctx, Region r, MDNode *tbaa);
+    explicit jl_aliasinfo_t(MDNode *tbaa, MDNode *tbaa_struct, MDNode *scope, MDNode *noalias)
+        : tbaa(tbaa), tbaa_struct(tbaa_struct), scope(scope), noalias(noalias) {}
+    jl_aliasinfo_t(const jl_aliasinfo_t &) = default;
+
+    // Add !tbaa, !tbaa.struct, !alias.scope, !noalias annotations to an instruction.
+    //
+    // Also adds `invariant.load` to load instructions in the constant !noalias scope.
+    Instruction *decorateInst(Instruction *inst) const {
+
+        if (this->tbaa)
+            inst->setMetadata(LLVMContext::MD_tbaa, this->tbaa);
+        if (this->tbaa_struct)
+            inst->setMetadata(LLVMContext::MD_tbaa_struct, this->tbaa_struct);
+        if (this->scope)
+            inst->setMetadata(LLVMContext::MD_alias_scope, this->scope);
+        if (this->noalias)
+            inst->setMetadata(LLVMContext::MD_noalias, this->noalias);
+
+        if (this->scope && isa<LoadInst>(inst)) {
+            // If this is in the read-only region, mark the load with "!invariant.load"
+            if (this->scope->getNumOperands() == 1) {
+                MDNode *operand = cast<MDNode>(this->scope->getOperand(0));
+                auto scope_name = cast<MDString>(operand->getOperand(0))->getString();
+                if (scope_name == "jnoalias_const")
+                    inst->setMetadata(LLVMContext::MD_invariant_load, MDNode::get(inst->getContext(), None));
+            }
+        }
+
+        return inst;
+    }
+
+    // Merge two sets of alias information.
+    jl_aliasinfo_t merge(const jl_aliasinfo_t &other) const {
+        jl_aliasinfo_t result;
+        result.tbaa = MDNode::getMostGenericTBAA(this->tbaa, other.tbaa);
+        result.tbaa_struct = nullptr;
+        result.scope = MDNode::getMostGenericAliasScope(this->scope, other.scope);
+        result.noalias = MDNode::intersect(this->noalias, other.noalias);
+        return result;
+    }
+
+    // Create alias information based on the provided TBAA metadata.
+    //
+    // This function only exists to help transition to using !noalias to encode
+    // memory region non-aliasing. It should be deleted once the TBAA metadata
+    // is improved to encode only memory layout and *not* memory regions.
+    static jl_aliasinfo_t fromTBAA(jl_codectx_t &ctx, MDNode *tbaa);
+};
 
 // metadata tracking for a llvm Value* during codegen
 struct jl_cgval_t {
@@ -1441,6 +1555,7 @@ class jl_codectx_t {
     jl_module_t *module = NULL;
     jl_typecache_t type_cache;
     jl_tbaacache_t tbaa_cache;
+    jl_noaliascache_t aliasscope_cache;
     jl_method_instance_t *linfo = NULL;
     jl_value_t *rettype = NULL;
     jl_code_info_t *source = NULL;
@@ -1452,7 +1567,6 @@ class jl_codectx_t {
     Value *spvals_ptr = NULL;
     Value *argArray = NULL;
     Value *argCount = NULL;
-    MDNode *aliasscope = NULL;
     std::string funcName;
     int vaSlot = -1;        // name of vararg argument
     int nReqArgs = 0;
@@ -1491,6 +1605,11 @@ class jl_codectx_t {
         return tbaa_cache;
     }
 
+    jl_noaliascache_t &noalias() {
+        aliasscope_cache.initialize(builder.getContext());
+        return aliasscope_cache;
+    }
+
     ~jl_codectx_t() {
         // Transfer local delayed calls to the global queue
         for (auto call_target : call_targets)
@@ -1502,6 +1621,77 @@ GlobalVariable *JuliaVariable::realize(jl_codectx_t &ctx) {
     return realize(jl_Module);
 }
 
+jl_aliasinfo_t::jl_aliasinfo_t(jl_codectx_t &ctx, Region r, MDNode *tbaa): tbaa(tbaa), tbaa_struct(nullptr) {
+    MDNode *alias_scope;
+    jl_noaliascache_t::jl_regions_t regions = ctx.noalias().regions;
+    switch (r) {
+        case Region::unknown:
+            alias_scope = nullptr;
+            break;
+        case Region::gcframe:
+            alias_scope = regions.gcframe;
+            break;
+        case Region::stack:
+            alias_scope = regions.stack;
+            break;
+        case Region::data:
+            alias_scope = regions.data;
+            break;
+        case Region::constant:
+            alias_scope = regions.constant;
+            break;
+        case Region::type_metadata:
+            alias_scope = regions.type_metadata;
+            break;
+    }
+
+    MDNode *all_scopes[5] = { regions.gcframe, regions.stack, regions.data, regions.type_metadata, regions.constant };
+    if (alias_scope) {
+        // The matching region is added to !alias.scope
+        // All other regions are added to !noalias
+
+        int i = 0;
+        Metadata *scopes[1] = { alias_scope };
+        Metadata *noaliases[4];
+        for (auto const &scope: all_scopes) {
+            if (scope == alias_scope) continue;
+            noaliases[i++] = scope;
+        }
+
+        this->scope = MDNode::get(ctx.builder.getContext(), ArrayRef<Metadata*>(scopes));
+        this->noalias = MDNode::get(ctx.builder.getContext(), ArrayRef<Metadata*>(noaliases));
+    }
+}
+
+jl_aliasinfo_t jl_aliasinfo_t::fromTBAA(jl_codectx_t &ctx, MDNode *tbaa) {
+    auto cache = ctx.tbaa();
+
+    // Each top-level TBAA node has a corresponding !alias.scope scope
+    MDNode *tbaa_srcs[5] = { cache.tbaa_gcframe, cache.tbaa_stack, cache.tbaa_data, cache.tbaa_array, cache.tbaa_const };
+    Region regions[5] = { Region::gcframe, Region::stack, Region::data, Region::type_metadata, Region::constant };
+
+    if (tbaa != nullptr) {
+        MDNode *node = cast<MDNode>(tbaa->getOperand(1));
+        if (cast<MDString>(node->getOperand(0))->getString() != "jtbaa") {
+
+            // Climb up to node just before root
+            MDNode *parent_node = cast<MDNode>(node->getOperand(1));
+            while (cast<MDString>(parent_node->getOperand(0))->getString() != "jtbaa") {
+                node = parent_node;
+                parent_node = cast<MDNode>(node->getOperand(1));
+            }
+
+            // Find the matching node's index
+            for (int i = 0; i < 5; i++) {
+                if (cast<MDNode>(tbaa_srcs[i]->getOperand(1)) == node)
+                    return jl_aliasinfo_t(ctx, regions[i], tbaa);
+            }
+        }
+    }
+
+    return jl_aliasinfo_t(ctx, Region::unknown, tbaa);
+}
+
 static Type *julia_type_to_llvm(jl_codectx_t &ctx, jl_value_t *jt, bool *isboxed = NULL);
 static jl_returninfo_t get_specsig_function(jl_codectx_t &ctx, Module *M, StringRef name, jl_value_t *sig, jl_value_t *jlrettype, bool is_opaque_closure);
 static jl_cgval_t emit_expr(jl_codectx_t &ctx, jl_value_t *expr, ssize_t ssaval = -1);
@@ -3254,7 +3444,7 @@ static bool emit_builtin_call(jl_codectx_t &ctx, jl_cgval_t *ret, jl_value_t *f,
                     *ret = emit_unionload(ctx, data, ptindex, ety, elsz, al, ctx.tbaa().tbaa_arraybuf, true, union_max, ctx.tbaa().tbaa_arrayselbyte);
                 }
                 else {
-                    MDNode *aliasscope = (f == jl_builtin_const_arrayref) ? ctx.aliasscope : nullptr;
+                    MDNode *aliasscope = (f == jl_builtin_const_arrayref) ? ctx.noalias().aliasscope.current : nullptr;
                     *ret = typed_load(ctx,
                             emit_arrayptr(ctx, ary, ary_ex),
                             idx, ety,
@@ -3369,7 +3559,7 @@ static bool emit_builtin_call(jl_codectx_t &ctx, jl_cgval_t *ret, jl_value_t *f,
                                     emit_arrayptr(ctx, ary, ary_ex, isboxed),
                                     idx, val, jl_cgval_t(), ety,
                                     isboxed ? ctx.tbaa().tbaa_ptrarraybuf : ctx.tbaa().tbaa_arraybuf,
-                                    ctx.aliasscope,
+                                    ctx.noalias().aliasscope.current,
                                     data_owner,
                                     isboxed,
                                     isboxed ? AtomicOrdering::Release : AtomicOrdering::NotAtomic, // TODO: we should do this for anything with CountTrackedPointers(elty).count > 0
@@ -7679,7 +7869,7 @@ static jl_llvm_functions_t
                 ctx.builder.SetCurrentDebugLocation(linetable.at(debuginfoloc).loc);
             coverageVisitStmt(debuginfoloc);
         }
-        ctx.aliasscope = aliasscopes[cursor];
+        ctx.noalias().aliasscope.current = aliasscopes[cursor];
         jl_value_t *stmt = jl_array_ptr_ref(stmts, cursor);
         jl_expr_t *expr = jl_is_expr(stmt) ? (jl_expr_t*)stmt : nullptr;
         if (jl_is_returnnode(stmt)) {