This patch modifies SelectionDAG and FastISel to produce DBG_INSTR_REFs with variadic expressions, and produce DBG_INSTR_REFs for debug values with variadic location expressions. The former essentially means just prepending DW_OP_LLVM_arg, 0 to the existing expression. The latter is achieved in MachineFunction::finalizeDebugInstrRefs and InstrEmitter::EmitDbgInstrRef.
Details
Diff Detail
- Repository
- rG LLVM Github Monorepo
Event Timeline
Generally looks good -- needs tests though.
llvm/lib/CodeGen/SelectionDAG/InstrEmitter.cpp | ||
---|---|---|
833–835 | Haven't these all been excluded due to IsInvalidOp? |
The "arg zero" form of DIExpression is going to be a very common case -- is it uniqued, and would it be worth storing a copy somewhere and just using that?
DIExpressions should be uniqued based on their expression elements, and the get should just be a map lookup underneath, so I don't think there should be any issues with duplicate storage.
Rebased, fixed some errors when emitting undef values and update some tests accordingly.
Added some test coverage for SelectionDAG+FastISel emission of variadic instruction references.
This LGTM with a few inline questions. The addition of DW_OP_arg, 0 to all INSTR_REF expressions does add some visual noise but I personally prefer this. Given how convoluted debug instruction handling can get, I like the idea of continuing in this direction and minimising the number of special cases.
llvm/lib/CodeGen/MachineFunction.cpp | ||
---|---|---|
1201 | 👍 | |
llvm/lib/CodeGen/SelectionDAG/InstrEmitter.cpp | ||
698 | Possibly worth expanding the comment to mention that these will have been emitted as INSTR_REF by the code above if instr ref is enabled. | |
856–857 | Looks like this one slipped through the cracks - can this TODO be done? IMO eliminating duplicated code is especially important around these fiddly areas. | |
922 | Is there a benefit to preserving the number of operands (rather than just using DBG_VALUE $noreg) - does it strip assertions about DW_OP_arg usage or something if we don't? |
llvm/lib/CodeGen/SelectionDAG/InstrEmitter.cpp | ||
---|---|---|
922 | The only important part about this is that it doesn't invalidate the DIExpression - even for an undef debug value, I believe there are still some points at which having a mismatched number of debug operands with the DIExpression's arg operands is a problem. This could be resolved by also changing the expression, making it empty while any fragment info or other expression elements that apply to undef values - which sounds like something that should be done for all explicitly killed debug values, but happy to start the ball rolling here. |
Address comments (extracted MO selection for const debug operands, expanded comments in EmitDbgValue), add a function for setting a DIExpression "undef".
Rebased onto latest main, added unit tests for new convertToUndefExpression, and converted some existing tests to test for the new fully-undef debug values that SelectionDAG produces.
👍