fix: Generate debug metadata for methods #1384
Conversation
...specifically in the context of function variables. Given
```
FUNCTION_BLOCK foo
// ...
METHOD bar
// ...
END_METHOD
END_FUNCTION_BLOCK
```
the compiler will generate two types for `foo` and `bar` and furthermore handle method calls by passing both
these generated types to some function. That is we get the following
```
%foo = type {}
%foo.bar = type {}
// ...
define void @foo.bar(%foo* %0, %foo.bar* %1) !dbg !<...> {
call void @llvm.dbg.declare(metadata %foo* %0, metadata !<...>, metadata !DIExpression()), !dbg !<...>
call void @llvm.dbg.declare(metadata %foo.bar* %1, metadata !<...>, metadata !DIExpression()), !dbg !<...>
// ^^^^^^
// was previously ignored
}
```
Appparently when dealing with `%foo` and `%foo.bar` in the `@foo.bar` function context, we intentionally
didn't create any debug information for `foo.bar`, resulting in the `metadata !<...>` field pointing to an
empty member (e.g. `!<num> = {}`). As a result LLVM would segfault when dealing with this and therefore any
`rusty <file> --debug` call.
This commit fixes this issue, such that the metadata field points to a correct `DILocalVariable` definition
for `%foo.bar`.
volsa
changed the title
src/codegen/tests/debug_tests.rs
Outdated
| END_METHOD | ||
| END_FUNCTION_BLOCK | ||
| ", | ||
| 4, |
There was a problem hiding this comment.
is there any particular reason to override the debug version here?
There was a problem hiding this comment.
None, I just copy-pasted the previous test and modified it. Good catch though, it probably should be a codegen call
src/codegen/tests/debug_tests.rs
Outdated
| assert!(codegen | ||
| .contains(r#"!17 = !DILocalVariable(name: "fb.foo", scope: !15, file: !2, line: 3, type: !18)"#)); | ||
|
|
||
| // ...for reference, it would previously point to `!4` which is empty (`!4 = {}`). `!4` shouldn't exist in |
There was a problem hiding this comment.
I think !4 is not wrong here - it just points out that the struct fb does not have any members and therefore no local nodes need to be retained.
For comparison, this is an example by LLVM. Here, the function foo does not have any members, !DISubprogram's retainedNodes metadata points to !2 which is !2 = !{}.
There was a problem hiding this comment.
Yes, you're right. I was under the impression that the !4 wasn't being used in the IR at all but that's not true.
...specifically in the context of function variables. Given
the compiler will generate two types for
fooandbarand furthermore handle method calls by passing both these generated types to some function. That is we get the followingAppparently when dealing with
%fooand%foo.barin the@foo.barfunction context, we intentionally didn't create any debug information forfoo.bar, resulting in themetadata !<...>field pointing to an empty member (e.g.!<num> = {}). As a result LLVM would segfault when dealing with this and therefore anyrusty <file> --debugcall.This commit fixes this issue, such that the metadata field points to a correct
DILocalVariabledefinition for%foo.bar.