The current implementation of llvm::Mangler does not check if a symbol name exists before mangling. This can cause problems if the mangled name aliases another symbol in the IR. For example, consider the following:
@"a(b" = addrspace(1) global float 1.0
@"a_28_b" = addrspace(1) global float 1.0
If I run this through the x86 backend, I get:
llc: /scratch/jholewinski/llvm/src/llvm/lib/MC/MCAsmStreamer.cpp:351: virtual void <anonymous namespace>::MCAsmStreamer::EmitLabel(llvm::MCSymbol *): Assertion `Symbol->isUndefined() && "Cannot define a symbol twice!"' failed.
The backend tries to define the mangled symbol name twice, since "a(b" mangles to "a_28_b".
Was it a conscious design choice to do it this way? Perhaps to save the time of verifying the uniqueness of the symbol? Or is the idea that you can always de-mangle by searching for "_XX_" sequences and replacing those with the ASCII value of XX?