[PATCH] D151952: [clang] adds `__type_pack_index` so we can get a type's parameter pack index

[Christopher Di Bella via Phabricator via cfe-commits]

cjdb added inline comments.


================
Comment at: clang/include/clang/AST/Stmt.h:796-802
+    /// If this expression is not value-dependent, this stores the value.
+    unsigned Value : 8;
 
     /// The number of arguments to this type trait. According to [implimits]
     /// 8 bits would be enough, but we require (and test for) at least 16 bits
     /// to mirror FunctionType.
+    unsigned NumArgs : 16;
----------------
erichkeane wrote:
> dblaikie wrote:
> > cjdb wrote:
> > > These numbers feel very low for how this modification is using them. Perhaps they should both be bumped to 32?
> > clang does, I think, have specific implementation limits for this sort of thing - 
> > 
> > Well, maybe we don't actually enforce a limit on number of template arguments... https://godbolt.org/z/accncYson compiles successfully, and has 2^18 parameters, beyond the 2^16 suggested here.
> > 
> > But maybe 2^16 is just what we test for/somewhat guarantee.
> > 
> > But if the `Value` is going to be the index into the args, shouldn't it be the same size as `NumArgs`, at least? (& the comment says 16, so 16 for both Value and NumArgs would seem suitably symmetric, and no larger than the current situation - since it'd just be splitting NumArgs in half, while still meeting the comment's suggestion/requirements?)
> > 
> > An assert, at least, when populating NumArgs to ensure it's no larger might not hurt... (which might be reachable from code/not checked prior - so it could be a hard compilation error, I guess, but I wouldn't feel super strongly about it)
> Can you explain the math here of how you chose to change this to 16?  I see that you removed 7 bits, but took away 16.  I'm not sure what NumExprBits is doing though, I got lost a little in that, so if you can calculate that to make sure this needs to be done, it would be appreciated.
> 
> Additionally, a static_assert after this to ensure the size isn't changing would also be appreciated.
> 
> But if the Value is going to be the index into the args, shouldn't it be the same size as NumArgs, at least? (& the comment says 16, so 16 for both Value and NumArgs would seem suitably symmetric, and no larger than the current situation - since it'd just be splitting NumArgs in half, while still meeting the comment's suggestion/requirements?)

Someone independently confirmed that you can have 200k types in a template, so we should probably be able to count at least that high (or alternatively, we should possibly consider not allowing more than 2^16 template parameters).

> Can you explain the math here of how you chose to change this to 16? I see that you removed 7 bits, but took away 16. I'm not sure what NumExprBits is doing though, I got lost a little in that, so if you can calculate that to make sure this needs to be done, it would be appreciated.

The value of `NumArgs` hasn't changed: I've just codified it. I can't work out why we need `NumExprBits`, but it's apparently required padding (when I removed it, a bunch of tests failed). Its value is 18, courtesy of clangd in VS Code.

> Additionally, a static_assert after this to ensure the size isn't changing would also be appreciated.

There's a static_assert in one of the `Stmt` constructors, which doesn't want more than eight bytes (we apparently need 16 if we're to change this).


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D151952/new/

https://reviews.llvm.org/D151952