Bug 2206 – unnamed template mixin of class inside function or class has incorrect classinfo and mangleof

Test: -- template T() { class C {} } void main() { mixin T!(); // using a named mixin here fixes it pragma(msg, T!().C.mangleof); pragma(msg, C.mangleof); // incorrectly outputs the same as above assert(T!().C.classinfo !is C.classinfo); // fails assert(T!().C.mangleof != C.mangleof); // fails } -- The types of T!().C and main.C are clearly different as the nested one carries an additional context pointer. This means that allocation relying on classinfo.init.length should be wrong too. In the frontend the types are initially different and only merged into a single type during semantic. To be specific, Type::mangle unifies them as toDecoBuffer yields the same result for both. Maybe this could be fixed by adding a TemplateMixin::mangle overload. This frontend bug causes some template-mixin related bugs in LLVMDC that seem to be worked around somehow in DMD.

To illustrate the resulting problem with class allocation: -- import std.stdio; class D {} template T() { class C { this() { } } } void main() { mixin T!(); // all print 8 writefln(T!().C.classinfo.init.length); writefln(C.classinfo.init.length); writefln(D.classinfo.init.length); auto c = new C; // segfault in _d_newclass }

Adding the following override for TemplateMixin::mangle seems to have fixed the issue in LLVMDC: -- char *TemplateMixin::mangle() { OutBuffer buf; char *id; #if 0 printf("TemplateMixin::mangle() %s", toChars()); if (parent) printf(" parent = %s %s", parent->kind(), parent->toChars()); printf("\n"); #endif id = ident ? ident->toChars() : toChars(); // use parent instead of tempdecl->parent here if (parent) { char *p = parent->mangle(); if (p[0] == '_' && p[1] == 'D') p += 2; buf.writestring(p); } buf.printf("%"PRIuSIZE"%s", strlen(id), id); id = buf.toChars(); buf.data = NULL; //printf("TemplateMixin::mangle() %s = %s\n", toChars(), id); return id; } --