x86_64: pass more tests #17788
x86_64: pass more tests #17788
Conversation
jacobly0
force-pushed
the
x86_64
branch
2 times, most recently
from
a4e7be8 to
b4c1579
Compare
| self.rdata_section_index = try self.allocateSection(".rdata", file_size, .{ | ||
| .CNT_INITIALIZED_DATA = 1, | ||
| .MEM_READ = 1, | ||
| .MEM_WRITE = 1, |
There was a problem hiding this comment.
rdata is supposed to be read-only isn't it?
There was a problem hiding this comment.
Due to a linker bug mutable globals are ending up here. I don't understand this enough to open an issue, but @kubkon said to just do this for now.
For reference, the global I ran into this with is lib.test_runner.cmdline_buffer, where the test program would crash on first write to this buffer, and only when the emitted code reaches a certain size (such as enabling n tests, or making unrelated changes to what code is emitted).
There was a problem hiding this comment.
It's not rdata, it's .data.rel.ro since we may have pointers inside a struct we lower in that section. The linker code path responsible for selecting output section doesn't do a recursive check for embedded pointers in any linker (macho, elf or coff), and this is important for PIC/PIE targets where the dynamic linker has to slide the address of each pointer within the binary - hence .data.rel.ro which is only writeable during dynamic relocation startup phase - well, in theory at least.
There was a problem hiding this comment.
Anyhow, as @jacobly0 mentioned, that was my suggestion to make it writeable to unblock his progress. I am aware of this nuance and will provide a fix in ELF first, and then backport it to other linkers as I start working on them.
This reduces the regression from 0.11.0 by 95%. Closes ziglang#17678
|
right-o, well, feel free to merge and save the stuff for follow-up changes if you want |
3 participants
Depends on #17802
Closes #17678
Closes #17618