codegen: fix memory type issue causing lost gc roots

When using memory to contain GC roots, you must be VERY careful never to
access that typed-memory through a different type of pointer (in this
case, accessing it as an i8 in the memcpy).

src/ccall.cpp

@@ -513,12 +513,7 @@ static Value *julia_to_native(
     // pass the address of an alloca'd thing, not a box
     // since those are immutable.
     Value *slot = emit_static_alloca(ctx, to);
-    if (!jvinfo.ispointer()) {
-        tbaa_decorate(jvinfo.tbaa, ctx.builder.CreateStore(emit_unbox(ctx, to, jvinfo, jlto), slot));
-    }
-    else {
-        emit_memcpy(ctx, slot, jvinfo.tbaa, jvinfo, jl_datatype_size(jlto), julia_alignment(jlto));
-    }
+    emit_unbox(ctx, to, jvinfo, jlto, slot, jvinfo.tbaa, julia_alignment(jlto));
     return slot;
 }
 

src/cgutils.cpp

@@ -2091,13 +2091,7 @@ static void init_bits_value(jl_codectx_t &ctx, Value *newv, Value *v, MDNode *tb
 
 static void init_bits_cgval(jl_codectx_t &ctx, Value *newv, const jl_cgval_t& v, MDNode *tbaa)
 {
-    // newv should already be tagged
-    if (v.ispointer()) {
-        emit_memcpy(ctx, newv, tbaa, v, jl_datatype_size(v.typ), sizeof(void*));
-    }
-    else {
-        init_bits_value(ctx, newv, v.V, tbaa);
-    }
+    emit_unbox(ctx, julia_type_to_llvm(v.typ), v, v.typ, newv, tbaa);
 }
 
 static jl_value_t *static_constant_instance(Constant *constant, jl_value_t *jt)

src/intrinsics.cpp

@@ -367,7 +367,7 @@ static Value *emit_unbox(jl_codectx_t &ctx, Type *to, const jl_cgval_t &x, jl_va
 
     unsigned alignment = julia_alignment(jt);
     Type *ptype = to->getPointerTo();
-    if (dest) {
+    if (dest && CountTrackedPointers(to).count == 0) {
         emit_memcpy(ctx, dest, tbaa_dest, p, x.tbaa, jl_datatype_size(jt), alignment, false);
         return NULL;
     }
@@ -385,12 +385,22 @@ static Value *emit_unbox(jl_codectx_t &ctx, Type *to, const jl_cgval_t &x, jl_va
                     (to->isFloatingPointTy() || to->isIntegerTy() || to->isPointerTy()) &&
                     DL.getTypeSizeInBits(AllocType) == DL.getTypeSizeInBits(to)) {
                 Instruction *load = ctx.builder.CreateAlignedLoad(p, alignment);
-                return emit_unboxed_coercion(ctx, to, tbaa_decorate(x.tbaa, load));
+                Value *unboxed = emit_unboxed_coercion(ctx, to, tbaa_decorate(x.tbaa, load));
+                if (dest) {
+                    tbaa_decorate(tbaa_dest, ctx.builder.CreateStore(unboxed, dest));
+                    return NULL;
+                }
+                return unboxed;
             }
         }
         p = maybe_bitcast(ctx, p, ptype);
         Instruction *load = ctx.builder.CreateAlignedLoad(p, alignment);
-        return tbaa_decorate(x.tbaa, load);
+        Value *unboxed = tbaa_decorate(x.tbaa, load);
+        if (dest) {
+            tbaa_decorate(tbaa_dest, ctx.builder.CreateStore(unboxed, dest));
+            return NULL;
+        }
+        return unboxed;
     }
 }