[WIP] Handle x86 MSVC jumptables#195
Conversation
| # If a line begins with ".dword", it is a jumptable | ||
| was_previous_jumptable_entry = True | ||
|
|
||
| target_line_num: Union[int, None] |
There was a problem hiding this comment.
| target_line_num: Union[int, None] | |
| target_line_num: Optional[int] |
|
|
||
| orig_jump_table_target: str = line.original.split(".dword")[1].strip() | ||
|
|
||
| # Remove "-0x..." from older objdump relocations |
There was a problem hiding this comment.
What does older mean, are we talking about different compiler or objdump versions, or is it older in the sense of us not yet having processed them?
There was a problem hiding this comment.
Older objdump versions such as 2.39 emit -0x... in relocations relative to the section start
The binutils that decomp.me uses is GNU objdump (GNU Binutils) 2.41.90.20240122 which does not emit the -0x....
| if target_line_num: | ||
| line.original = line.original.replace( | ||
| orig_jump_table_target, hex(target_line_num).replace("0x", "") | ||
| ) |
There was a problem hiding this comment.
Can we set line.original = ".dword\t" + hex(...) instead? Search-replace always feels fragile to me
| # This will have errors for multiple text sections as objdump (at least 2.38) doesn't emit | ||
| # the correct labels for sections that are not the first | ||
|
|
||
| jump_table_targets: List[str] = [] # target_label |
There was a problem hiding this comment.
Looks like it. I have the targets stored to it but they are never read from it.
| # if the previous line was a jumptable and the current line is "add\t%al,(%eax)" | ||
| # this is the 2nd half of the previous jumptable entry | ||
| if line.original == "add\t%al,(%eax)": | ||
| lines.remove(line) |
There was a problem hiding this comment.
.remove(...) loops over the entire lines list and checks dataclass equality, which is slow and can have false positives. We could instead e.g. do
indices_to_delete = []
for index, ... in enumerate(lines):
if ...:
indices_to_delete.append(index)
for index in reversed(indices_to_delete):
del lines[index]| elif line.mnemonic == "<label>" and line.original.startswith("$L"): | ||
| # If this is a label that is an MSVC jumptable label "$L..." | ||
| # add it to the labels list | ||
| labels[line.original] = lines.index(line) |
There was a problem hiding this comment.
... and then this would let us use index instead of finding the line anew because lines has not yet been re-indexed
|
(Should [WIP] be removed from the title?) |
Co-authored-by: Simon Lindholm <simon.lindholm10@gmail.com>
Co-authored-by: Simon Lindholm <simon.lindholm10@gmail.com>
Co-authored-by: Simon Lindholm <simon.lindholm10@gmail.com>
2 participants
This is a draft pr that implements handling of MSVC jumptables for x86. It currently requires diff function symbols to work properly as that is needed for the labels to be processed but it should not require that in the long run.
A small issue with this is that if there are multiple text sections, arrows will be shown in each text section. There was also an issue with objdump not emitting proper information for jumptables in a 2nd text section if the first section had a jump table.
objdump output of the jumptable:
objdump output of the assembly that is referenced from the jumptable:
An image of the final result:
We need <$Lnum> in the 2nd image to get the location.