[llvm] [DebugInfo] Merge partially matching chains of textual inclusions (PR #125780)

[Vladislav Dzhidzhoev via llvm-commits]

dzhidzhoev wrote:

> Hi!
> 
> > In the case of textual inclusion, we don't produce any inline stack frames - the frame is always just the current subroutine, albeit with a different file. Because of this, partial matching chains of textual inclusion wouldn't give an output that would be visible in a debugger - the only thing that would matter is whether the top "frames" match. I think it makes sense to treat the file more like the outermost part of the line+column matching, i.e. if the files don't match then we use a line 0 in the nearest common scope, and if the files do match then we use that matching file (creating a new DILexicalBlockFile at the nearest common scope if necessary).
> 
> I have not dug into the details in this patch yet, but for my understanding, I tried creating some reproducers for @SLTozer's suggestion above:
> 
> ```
> #include <stdlib.h>
> int g1 = 4;
> 
> // Here we have one abort() call originating from two identical text inclusions.
> //
> // We expect the location for the call to be foo.inc:6, and the scope for
> // the DILexicalBlockFile to be the same_inclusion_scope() Subprogram scope
> // (since that is the nearest common scope for the two blocks in the function),
> // i.e. something like this (?):
> //
> // !20 = !DIFile(filename: "foo.inc", directory: "[...]")
> // !21 = !DILexicalBlockFile(scope: [same_inclusion_scope()], file: !20)
> // !22 = DILocation(line: 5, scope: !21)
> int same_inclusion_scope() {
>   int v = g1;
>   if (v > 5)
> # 5 "foo.inc"
>     abort();
>   if (v > 5)
> # 5 "foo.inc"
>     abort();
>   return 0;
> }
> 
> // Here we have one abort() call originating from two different included files.
> //
> // We expect the location for the call to be the nearest common scope from the
> // two merged calls, which is src.c:0 in the different_inclusion_scope
> // Subprogram scope, i.e. something like this (?):
> //
> // !24 = DILocation(line: 0, scope: [different_inclusion_scope()])
> int different_inclusion_scope() {
>   int v = g1;
>   if (v > 5)
> # 5 "qux.inc"
>     abort();
>   if (v > 5)
> # 10 "corge.inc"
>     abort();
>   return 0;
> }
> ```
> 
> AFAICT it looks like the patch does not have any effect on these examples (compiled using `-O3 -g`), but perhaps that has something to do with the nearest common scope being the subprogram.

Thank you! Indeed, for this reproducer, we can't get another common location but src.c:0 for different_inclusion_scope

https://github.com/llvm/llvm-project/pull/125780