GOFF global symbols

[philipc]

The spec says (note that it calls module scope "global"):

Binding Scope	5.4	Bit(4)	Requested binding or resolution-search scope of an external symbol.

B'0000'
    Unspecified
B'0001'
    Section scope ("local")
B'0010'
    Module scope ("global")
B'0011'
    Library scope.
B'0100'
    Import-Export scope

All other values are reserved.

GoffSymbol::is_global() treats module scope as local:
(Note that there's a separate issue where binding_scope() is reading the wrong nibble.)

object/src/read/goff/symbol.rs

Lines 285 to 290 in dce50f0

	// If binding scope is section or module symbol is local
	let scope = self.behavioral_flags().binding_scope();
	if is_section || is_local_name || scope == GOFF_SCOPE_SECTION || scope == GOFF_SCOPE_MODULE
	{
	return false;
	}

write_defined_symbol() uses module scope for SymbolScope::Linkage symbols:
(For ELF, SymbolScope::Linkage means it is global for static linking, but not visible for dynamic linking.)

object/src/write/goff.rs

Lines 716 to 723 in dce50f0

	// Determine binding scope based on symbol properties
	let scope = if symbol.is_local() {
	goff::GOFF_SCOPE_SECTION
	} else if symbol.scope == SymbolScope::Dynamic {
	goff::GOFF_SCOPE_IMPORT_EXPORT
	} else {
	goff::GOFF_SCOPE_MODULE
	};

LLVM reads module scope as local (does not set SF_GLOBAL):
https://github.com/llvm/llvm-project/blob/3306ae1c98d695d7d111bd8037380c7d09e818b6/llvm/lib/Object/GOFFObjectFile.cpp#L446-L458

LLVM writes library scope or import-export scope for globals, otherwise section scope (never module scope):
https://github.com/llvm/llvm-project/blob/3306ae1c98d695d7d111bd8037380c7d09e818b6/llvm/lib/CodeGen/TargetLoweringObjectFileImpl.cpp#L2910-L2914

@dalvescb Which ones need changing? Based on the LLVM behaviour we probably should switch write_defined_symbol() to library scope instead of module scope, but I'm a bit confused why the spec says module scope is global but LLVM treats it as local.

[dalvescb]

Which ones need changing? Based on the LLVM behaviour we probably should switch write_defined_symbol() to library scope instead of module scope, but I'm a bit confused why the spec says module scope is global but LLVM treats it as local.

We've had some debate internally on this ... The spec does not appear to correctly reflect our system tools on zOS (i.e nm / ld) in this respect and seems to be simply incorrect

We should be mirroring the LLVM behaviour. Reading module scope as local is correct according to the system nm and linker

Based on the LLVM behaviour we probably should switch write_defined_symbol() to library scope instead of module scope,

I would recommend doing this for now and I'll reach out about how/why we're writing globals in LLVM. If we update how we write globals in LLVM I'll open a PR here to reflect it