[Lldb-commits] [lldb] [lldb] Fix `ObjectFileMachO` object format when missing version load commands (PR #144177)

[Jason Molenda via lldb-commits]

jasonmolenda wrote:

I don't want to go with this approach, I don't want to see `-macho` in a triple that we're creating for any of these modules; this is going to break things down the line with firmware environments, I'm sure of it.  It's also ugly.

This is an undesired behavior from llvm's Triple, because we create a Triple in ArchSpec::SetArchitecture with only a cpu specified, and llvm's Triple `getDefaultFormat()` is called when constructing the Triple, and assumes that an OS is defined.  If no OS is specified, it sets the object file format to Triple::ELF by default.  This is clearly wrong, it should be UnknownObjectFormat.  If we later call Triple::setObjectFormat to specify the correct object file format, Triple adds "-macho" to the triple unconditionally.

My guess is that other parts of llvm depend on the object file format being set to Triple::ELF when there is an unspecified OS.  I wouldn't be surprised to see regressions if `getDefaultFormat()` changed to return  UnknownObjectFormat when no OS is specified, as it really should be.  And then if they got "-elf" added to triple strings when they called Triple::setObjectFormat they would also not want that behavior.

I don't think any of these approaches to fixing this in ObjectFileMachO / ArchSpec are the right approach.  llvm's Triple needs to have a way to set the object file format *without* changing the triple string, it needs to not set the object file format when no OS is specified, or we need to live with a Triple having an incorrect ObjectFormat in lldb.

https://github.com/llvm/llvm-project/pull/144177