[lldb-dap] incorrect offsets in assembly view

[vogelsgesang]

In the disassemble call, the instructionOffset is treated as if it would be measuring the offset in bytes instead of as instructions (source code). This leads to an corrupted disassembly view in VS Code when scrolling

[llvmbot]

@llvm/issue-subscribers-lldb

Author: Adrian Vogelsgesang (vogelsgesang)

In the `disassemble` call, the `instructionOffset` is treated as if it would be measuring the offset in bytes instead of as instructions ([source code](https://github.com/llvm/llvm-project/blob/2913e71865dfc063a47ddfaf1e2ce07763f69614/lldb/tools/lldb-dap/lldb-dap.cpp#L3913)) this leads to an corrupted disassembly view in VS Code when scrolling

[santhoshe447]

Hi @vogelsgesang

Could you pls provide more details to understand the issue?
We noticed a crash in lldb-dap when scrolling in the disassembly view window. This was caused by incorrect handling of "instructionOffset" and "Offset" for the Hexagon arch (As it does instruction packetization).
We have addressed this issue by handling all possible causes related to "instructionOffset" and "instructionCount".

"instructionOffset" applied before disassembling, as it will tell us from where to begin the disassemble.
If the "instructionOffset" is positive - start disassembling after memoryReference.
If the "instructionOffset" is negative - start disassembling before memoryReference.

If you know the significance of the "Offset" value in the disassemble request command, could you explain how it differs from "instructionOffset"?

Thanks,

[vogelsgesang]

If you know the significance of the "Offset" value in the disassemble request command, could you explain how it differs from "instructionOffset"?

The offset is measured in bytes, while the instruction offset is measured in instructions. This is particularly important for variable-length encoded instruction sets (such as Intel assembly), where there is no direct way to map from "instruction count" to "byte count"

VS-Code uses instructionOffset for lazy-loading / scrolling in the UI. I don't know what offset is actually useful for and if there are any users of it.

I just uploaded #105446 which shows my current progress at fixing this issue. The Pull Request is not completely finished, yet, though. In particular, the test cases are still broken.

[fr43nk]

Hi,

I wanted to use the disassembly view and the current line was not correct. I think this is because if the inline code beside the instructions. What also would be nice: If the disassembly view is active, then the step over and step into commands in the debug toolbar, should also relate to the ni and si commands.
@vogelsgesang is that what you are working on with this issue?

Best regards.

[vogelsgesang]

If the disassembly view is active, then the step over and step into commands in the debug toolbar, should also relate to the ni and si commands.

This should already be working on main. However, the disassembly view itself is broken currently, so the single-stepping behavior is not correctly reflected in the UI currently

This defect is about fixing the broken disassembly view

[da-viper]

Fixed in #140486