Reconstruct "erased" tags from DWARF debugging info

[cadmic]

The DWARF info for the OOT GC emulator has useful info that has been partially overwritten by padding tags. Specifically, the overwritten info seems to be tags from things stripped by the linker, mainly unused functions or functions that have been inlined away. So far this has been incredibly helpful so far for figuring out potential inline functions when decomping.

I learned about this because https://github.com/LuigiBlood/dwarfone attempts to include these tags too. It requires some heuristics to parse because the format is not standard, and the first tag's type and length has been overwritten by the padding tag. It's also little-endian unlike the rest of the tags, and there seems to be some MetroWerks-specific stuff too.

I tried to improve the heuristics a little when porting this to dtk. I have no idea if it works for other games, other compiler versions, or C++, so erased tag parsing is hidden behind an --include-erased flag. Currently only erased functions are output, it seems like erased local variables are things like static double _half$localstatic0$sqrtf__Ff which are more confusing than useful IMO.

Is "erased" a good name for this? Maybe "stripped" or something is better.

Example:

$ dtk dwarf dump --include-erased oot-gc/SIM_original.elf
...
// Erased
static int xlFileLoad(char * szFileName /* r30 */, void * ppBuffer /* r31 */) {
    // Local variables
    int nSize; // r1+0x18
    struct tXL_FILE * pFile; // r1+0x14

    // References
    // -> static int (* gpfOpen)(char *, struct DVDFileInfo *);
    // -> struct _XL_OBJECTTYPE gTypeFile;
}
...

[cadmic]

hmm, I think we might have to pass in both the "default endianness" and the "current endianness" for this

[encounter]

I think this will break PS2 DWARF, which is little endian. Can we fix this?

[cadmic]

I fixed this by plumbing both "data" and "address" endianness through but I worry it's getting a bit messy. I tested this on Sonic Heroes and the output is identical (and unfortunately it did not discover any erased tags).

[encounter]

Cool stuff! Recovering partially overwritten data is a bit complicated, so I have a few questions while trying to follow the logic.

[encounter]

I think this will break PS2 DWARF, which is little endian. Can we fix this?

[encounter]

Interesting, can the user types not appear here?

[cadmic]

User types can still appear here, this is a hack to parse these values:

FundType::parse_int: 0101
FundType::parse_int: 0103
FundType::parse_int: 0105
FundType::parse_int: 0106
FundType::parse_int: 0107
FundType::parse_int: 0108
FundType::parse_int: 0109
FundType::parse_int: 010E
FundType::parse_int: 010F

maybe it would be more robust to add enums for all these explicitly, like MwSignedShort = 0x0105?

[encounter]

How does this happen?

[cadmic]

There are 3 values we need to handle in the OOT elf:

• 00: used for all pointers except signed int (not sure what to call this)
• 80: used for pointers to signed int (maybe should be called MwPointerTo?)
• 82: not sure, it's only used twice, for a local variable signed int & pDL in two different functions (maybe should be called MwReferenceTo? I'm still confused about pointer vs reference here)

[encounter]

Looks good. Thanks for the PR!