[LLVMdev] Questions before moving the new debug info hierarchy into place

[Duncan P. N. Exon Smith]

> On 2015-Feb-20, at 09:00, David Blaikie <dblaikie at gmail.com> wrote:
> 
> 
> 
> On Thu, Feb 19, 2015 at 7:49 PM, Duncan P. N. Exon Smith <dexonsmith at apple.com> wrote:
> I'm getting close to executing the transition to the new debug info
> hierarchy.  For reference, I've attached two WIP patches (which would be
> squashed before commit) and the WIP upgrade script I'm using.
> 
>     - transition-code.patch: Change the `DIDescriptor` hierarchy to
>       lightweight wrappers around the subclasses of `DebugNode` (instead
>       of rather heavier wrappers around `MDTuple`), and update
>       `DIBuilder` to generate the right things.
>     - transition-testcases.patch: Upgrade of testcases, entirely
>       generated from the upgrade script (so far).
>     - upgrade-specialized-nodes.sh: Script to upgrade LLVM assembly.
> 
> (Feel free to have a look, but I haven't updated LangRef, I don't quite
> have `make check` passing, and I haven't even started on `make
> check-clang` (I imagine that'll all be done by hand).)
> 
> There are two outstanding issues I'd like some feedback on.
> 
> Pretty-printing the flags
> =========================
> 
> I've noticed the `flags:` field is *harder* to read in the new assembly:
> 
>     !MDDerivedType(flags: 16384, ...)
> 
> than the pretty-printed comments in the old:
> 
>     !{!"...\\0016384", ...} ; ... [public] [rvalue reference]
> 
> I don't want to regress here.
> 
> In `DIDescriptor`, the flags are described in an enum bitfield:
> 
>     FlagAccessibility     = 1 << 0 | 1 << 1,
>     FlagPrivate           = 1,
>     FlagProtected         = 2,
>     FlagPublic            = 3,
>     FlagFwdDecl           = 1 << 2,
>     FlagAppleBlock        = 1 << 3,
>     FlagBlockByrefStruct  = 1 << 4,
>     FlagVirtual           = 1 << 5,
>     FlagArtificial        = 1 << 6,
>     FlagExplicit          = 1 << 7,
>     FlagPrototyped        = 1 << 8,
>     FlagObjcClassComplete = 1 << 9,
>     FlagObjectPointer     = 1 << 10,
>     FlagVector            = 1 << 11,
>     FlagStaticMember      = 1 << 12,
>     FlagLValueReference   = 1 << 13,
>     FlagRValueReference   = 1 << 14
> 
> I think the right short-term solution is to use these names directly in
> assembly, giving us:
> 
>     !MDDerivedType(flags: FlagPublic | FlagRValueReference, ...)
> 
> This is easy to implement and easy to CHECK against.
> 
> Sound good?
> 
> (Eventually, I'd like to use the `DW_AT` symbols that each of these
> corresponds to, but `FlagStaticMember` doesn't seem to line up with any
> such `DW_AT` so that will take some refactoring (and I don't think it
> makes sense for that to block moving the hierarchy into place).)
> 
> Merging the two types of files
> ==============================
> 
> In the old format, we have two types of files -- an untagged file node,
> and a DIFile/DW_TAG_file_type/0x29 which references the untagged node.
> 
>     !0 = !{!"path/to/file", !"/path/to/dir"}
>     !1 = !{!"0x29", !0}
> 
> In the actual metadata graph, most file references use !0, but in
> DIBuilder !1 is always passed in and the !0 is extracted from it.
> 
> I've been planning to merge these into:
> 
>     !1 = !MDFile(filename: "path/to/file", directory: "/path/to/dir")
> 
> Anyone see a problem with that?  Why?
> 
> If the strings are deduplicated elsewhere, that should be fine. (I think I made the change originally to pull out the names because of the duplication I saw in many constructsn needing to reference the file/directory and they all contained the same file/directory)

Yup, they'll be de-duped via `MDFile`.

> I take it you'll have other constructs (I think all scopes have a file/directory) reference the MDFile directly?

I'll make it so all `file:` references point to an `MDFile`; so
far this is only true for `MDCompileUnit` (the others point to
the untagged node).  There are some places (at least in
testcases) where `MDFile` is being used in a `scope:` reference;
I'll leave those unchanged (although IIUC, Adrian thinks a file
should never be used as a scope -- you always have a better
option (compile unit, namespace, etc.) -- so maybe that'll be
changing eventually... we just need be mindful of type uniquing).

>  
> 
> If not, I'd like to roll that change into the same patch in order to
> reduce testcase churn.  (I've discovered that it won't complicate the
> code patch at all.)