dashboard/app: coverage precision

[tarasmadan]

For LLVM builds detailed information about basic blocks may be added with KCFLAGS="-mllvm -enable-bb-addr-map".
The same section is indirectly added by experimental CONFIG_AUTOFDO_CLANG.

Having the information about BB ends we can make coverage visualization much more precise (text ranges instead of lines).

Plan

1. dashboard/config/linux: enable BB descriptors #6679 .
2. Use bb_map data to improve local coverage reports on clang build.
3. Propagate bb_map details to BigQuery database. We already have field for bb ends.
4. Merge covered regions instead of coverage lines.

[tarasmadan]

Working with KCFLAGS wasn't the easy way.
The easy way is to enable 2 configs, CONFIG_AUTOFDO_CLANG and CONFIG_PROPELLER_CLANG.

Let me quote gemini:

The combination of these two configs gives you everything needed to map a PC to a specific Start Line:Col - End Line:Col range.

Here is the workflow enabled by these configs:

1. Find the Basic Block Boundaries (PC Range):

• Use the Address Map (from CONFIG_PROPELLER_CLANG / vmlinux section .llvm_bb_addr_map).
• Find the entry containing your PC.
• Get Block Start PC and Block Size.
• Calculate Block End PC = Start PC + Size - 1.

2. Map to Source Code (Line:Col Range):

• Start Line:Col: Run llvm-symbolizer on the Block Start PC.
  • (The Discriminators from CONFIG_AUTOFDO_CLANG ensure this is precise even for macros/multiple blocks on one line).
• End Line:Col: Run llvm-symbolizer on the Block End PC.

Proof of Concept: I tested this on the panic function in your built kernel:

• Block: Start 0xFFFFFFFF81200010, Size 0x4A.
• Start PC (...10) maps to kernel/panic.c:622:0.
• End PC (...59) maps to kernel/panic.c:626:2.

This confirms you have the full Line:Col -> Line:Col block mapping.

[tarasmadan]

AutoFDO is used to get line:col information. DWARF Line Table is used for the same purpose.
The difference is the so-called discriminators. The same line:col value with different discriminators is used to represent different "macro" or "inline" cases. It looks useful for the coverage visualization.