Valid debug information being deleted by DAGCombiner

I am working on some debug problems that we are seeing with the x86 backend and OpenCL.

The input is linked.bc. There are three debug values, ip, tid and gid.

llc -march=x86 linked.bc -o linked-x86.s ← assembly file has all three values in the debug info section

However, if I attempt to optimize the bitcode with the following command:

opt -disable-opt -inline-all -mem2reg linked.bc -o optimized.bc

llc -march=x86 optimized.bc -o optimized-x86.s ← Assembly file only has ip and gid in the debug info section.

Any idea on how to get this to work? I’ve attached the output, which is missing the ‘tid’ debug variable.

Thanks,

Micah

linked.bc (1.68 KB)

optimized.s (16.1 KB)

Found another bitcode file where a debug symbol is being dropped.

In the attached bitcode file, the variable gid is not in the debug output.

x86_optimized.bc (1.38 KB)

Found another bitcode file where a debug symbol is being dropped.

In the attached bitcode file, the variable gid is not in the debug output.

John,

Mem2reg actually modifies debug information in a way that at least suggests it’s trying to maintain debugability. Specifically, it changes llvm.dbg.declare() calls (appropriate for variables that permanently reside in a single place) to llvm.dbg.value() calls (specifying that at this instant, this variable is in this register).

In fact, the .bc after inlining and mem2reg optimizations seems correct. Llvm.dbg.value() refers to a chain of single-input phi nodes left behind by inlining that ultimately resolve to the value in the inlined function that corresponds to the return value. But when code gen does its thing, DAGCombiner thinks some of the llvm.dbg.value() calls are associated with dead assignments and removes them. Only we haven’t been able to figure out why DAGCombiner is sometimes confused. For example (from a slightly different test case than Micah posted), here’s a block of post-optimized IR for a dbg.value() call that the DAGCombiner ends up deleting…

%8 = extractelement <4 x i32> %7, i32 0 ; [#uses=1]

br label %9

; :9 ; preds = %get_local_id.exit

%10 = phi i32 [ %8, %get_local_id.exit ] ; [#uses=1]

br label %11

; :11 ; preds = %9

%12 = phi i32 [ %10, %9 ] ; [#uses=1]

br label %get_global_id.exit

get_global_id.exit: ; preds = %11

%13 = phi i32 [ %12, %11 ] ; [#uses=4]

call void @llvm.dbg.value(metadata !{i32 %13}, i64 0, metadata !27), !dbg !28

However, a similar sequence of IR ‘works’, IF that sequence’s analog to %13 is used in the SAME basic block as the assignment to %13. It’s only if %13 is not used in the same basic block that the dbg.value gets removed. Or at least that’s the feeling we have at this point.

Thanks,

Richard

John/Richard,

I think I have found the problem to why the debug information is getting destroyed. The problem is in SelectionDAG and how it interacts with the SDDbgValue nodes and custom SDNodes.

When the dbg_value intrinsic is encountered, it adds the debug value to a specific SD Node in SelectionDAGBuilder.cpp::visitIntrinsicCall(). In one of my cases, it is vector_extract_elt.

During ISelLowering, the vector_extract_elt instruction is lowered to a custom backend SDNode and all of the operands are transferred over. The SDNode that stored vector_extract_elt is then deleted as it has no more uses. This then invalidates the debug information, causing it to no longer be printed. As the new SDNode has a debug loc, but not a corresponding SDDbgValue, it correctly shows dbg::line:col when printing the DAG, but not during assembly printing.

So, now that I’ve figured out what the problem is, anyone have an idea on how to fix it? Should getNode create a new SDDbgValue from a DebugLoc everytime? Should the DAG update the SDDbgValue to point to a new node when getNode is created?

Any other ideas?

Thanks,

Micah

John/Richard,
I think I have found the problem to why the debug information is getting destroyed. The problem is in SelectionDAG and how it interacts with the SDDbgValue nodes and custom SDNodes.

When the dbg_value intrinsic is encountered, it adds the debug value to a specific SD Node in SelectionDAGBuilder.cpp::visitIntrinsicCall(). In one of my cases, it is vector_extract_elt.

During ISelLowering, the vector_extract_elt instruction is lowered to a custom backend SDNode and all of the operands are transferred over. The SDNode that stored vector_extract_elt is then deleted as it has no more uses. This then invalidates the debug information, causing it to no longer be printed. As the new SDNode has a debug loc, but not a corresponding SDDbgValue, it correctly shows dbg:<filename>:line:col when printing the DAG, but not during assembly printing.

So, now that I’ve figured out what the problem is, anyone have an idea on how to fix it? Should getNode create a new SDDbgValue from a DebugLoc everytime? Should the DAG update the SDDbgValue to point to a new node when getNode is created?

You want to transfer SDDbgValue from old node to new node using SelectionDag::TransferDbgValues(From, To).

Thanks Devang.

Is there any plan on making this implicit in the node creation? Or instead of using DebugLoc in SelectionDAG, use SDDbgValue’s instead, or embedding the SDDbgValue in the debug loc itself somehow? This seems like something that shouldn’t have to be handled on a case by case basis in the backends.

Your thoughts?

Thanks,

Micah

DebugLoc and DbgValue serve different purposes. DebugLoc identifies source location and line number. It is part of llvm::Instruction, MachineInstr and SDNode. If these information is not generated by FE then it does not impact generated code. DebugLoc is more or less implicit in node creation.

DbgValue refer to value (or node) for a variable. We do not want to represent them as a node in DAG, because debugging information should not impact generated code.