w2yehia wrote:
I don't like that we have two user interfaces of specifying a copyright string
(a pragma and a flag), but I can live with that for user's convenience and
given the two interfaces converge as early as possible (i.e. in clang). However
having seen this PR now, I see that we have two IRs:
1) pragma => generates an MD in clang => LowerCommentStringPass lowers to a
global variable that is referenced by every function with `!implicit.ref`
2) -mloadtime-comments-vars=foo,bar => generate a different MD in clang
`!loadtime_comment.vars = !{!{!"foo"}, !{!"bar"}}` => LowerCommentStringPass
processes the list of names on the MD and generates `!implicit.ref` to each
named global.
I think it's better to have a single representation in IR for the two FE
interfaces, and have the lowering pass handle one representation. Since we
don't want to mutate existing variables referenced via the flag, annotating the
variables (to be later consumed by the lowering pass) is fine. But we don't
just list their names on an MD, instead you can list the variables directly on
the MD (`!loadtime_comment.vars = !{ptr @sccsid, ptr @version}`) or better,
decorate the variables themselves with an MD:
```
@sccsid = internal global ptr @.str, align 8, !loadtime_comment
@version = internal global [27 x i8] c"@(#) Copyright Version 2.0\00", align 1,
!loadtime_comment
```
Then, for the pragma comment, you can generate a variable per comment (or
concatenate them into one, if that works for you) in clang and decorate the
variable(s) with `!loadtime_comment`.
Adding the variables to `@llvm.compiler.used` will be done in clang.
https://github.com/llvm/llvm-project/pull/187986
_______________________________________________
llvm-branch-commits mailing list
[email protected]
https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-branch-commits
