[llvm-dev] Defining the DIExpression operator DW_OP_LLVM_arg

[Adrian Prantl via llvm-dev]

Thanks for bringing this up!

To summarize my understanding: Neither of these operators is strictly necessary, since the same effect can be achieved by implicitly pushing all operands of a DBG_VALUE to the stack, followed by a combination of DW_OP_dup, DW_OP_pick, DW_OP_swap, DW_OP_rot, and DW_OP_over. However, the resulting expressions can get very long and unwieldy and it is easier to generate efficient DWARF from DIExpressions that explicitly refer to arguments. DW_OP_LLVM_argN has the advantage of using less memory, at the cost of limiting the number of arguments to a hardcoded value.

By restricting their use to new intrinsics only (derefval and DBG_VALUE_VAR) we can avoid any confusion about whether argument 0 is implicitly pushed to the stack or not — in those new intrinsics, arguments always have to be referred to explicitly. Personally, I'm leaning more towards DW_OP_LLVM_argN, because of the space saving aspect.

Finally, to completely derail this discussion — once we have multi-arg support, we may be able to simplify parts of the debug intrinsic handling by combining multiple DW_OP_LLVM_fragment intrinsics describing the same variable into one multi-argument expression with multiple DW_OP_LLVM_fragments.

-- adrian