[flang-commits] [clang] [compiler-rt] [flang] [llvm] [mlir] [DebugMetadata][DwarfDebug][CodeView] Support function-local static variables in lexical block scopes (6/7) (PR #187927)

[Vladislav Dzhidzhoev via flang-commits]

dzhidzhoev wrote:

> > I've modified DebugImportant locally, so that, as I believe, it should be able to cache DIGlobalVariableExpressions (it's here https://github.com/dzhidzhoev/llvm-project/commits/debuginfo/rfc-krisb/6/rebased/). Then I ran the test from the gist mentioned above (https://gist.github.com/dzhidzhoev/34364710e5478a30ab1023239026e04d). I get the following output: https://gist.github.com/dzhidzhoev/4da6642e9a02817a2ca738d38d55bd65. di_global_variable_expression's and di_global_variable's for "a" and "b" are duplicated.
> 
> Thanks for sharing the output. The variables are indeed not uniqued due to the different subprograms. It seems like we need to make either DIGlobalVariableAttr or DIGlobalVariableExpressionsAttr self recursive. Apparently, the first variable that has cyclic dependencies in traversal order needs to be self recursive.
> 
> Changing DIGlobalVariableAttr seems easier since it is a DINodeAttr. There is a recent PR that shows the necessary steps to make a node self recursive #190808. It is somewhat involved which is why I wanted to understand that this step is necessary.

@chbessonova did some investigation, and it turns out that the same issue also affects DILocalVariables, which already live in `retainedNodes`. 
See https://gist.github.com/chbessonova/67fc7460ee6a148dd0f9734a25b37564. When 
`bin/mlir-translate -import-llvm -mlir-print-debuginfo dilocalvariable-duplication.ll`
is run on it, the output contains duplicate `#di_local_variable`s.
Would it make more sense to create a separate issue for the duplication problem involving DILocalVariables/DIGlobalVariables living in `retainedNodes`, and to keep the fix for that separate from this PR?

https://github.com/llvm/llvm-project/pull/187927