illumos / libdwarf Goto Github PK

We should have a README that explains how this repository works.

It appears the 20200612 branch has a LOT of commits in it, vs. master. So are we keeping master as the origin?

"The branch contains a base commit, which is the release tarball. From there, changes that we have applied to build on illumos are committed as individual patches."

it looks like maybe README needs to be updated to instead of "base commit as tarball" it is instead "base pull from upstream"?

When building a 32-bit relocatable object, dwarfdump and libdwarf misidentify the strings that are used for clang. DWARF has two different ways to emit strings. One, DW_FORM_string basically emits the ascii string into the dwarf sections. The other, DW_FORM_strp emits a four byte offset into the .debug_str section.

A 32-bit clang will emit references of the form DW_FORM_strp; however, it will leave the offset in the dwarf as zero because it sets up a relocation of the type R_386_32 whose symbol is to a specific spot in the .debug_str section. libdwarf doesn't process these relocations today because it assumes that they'll be done in a way that gcc does it, where by it sets the symbol for the relocation to be the start of the .debug_str section and instead leaves an addend in the symbol location which is its offset into the .debug_str section. The side effect of this is that it means that every string reference ends up always being to the first string in the .debug_str section, which humerously in our case means that every type, member, and path is clang version 9.0.0 (tags/RELEASE_900/final).

The fix to this is to have libdwarf actually process the SHT_REL sections and not just the SHT_RELA sections. I've put together a change that's good enough for 32-bit x86.