[IRGen] Don't use GOTPCREL relocations for x86 ELF. #82176
Conversation
Unforunately, x86 ELF linkers like to optimize GOTPCREL relocations by replacing `mov` instructions that go via the GOT with `lea` instructions that do not. That would be fine, but they aren't very selective and will happily perform this transformation in non-code sections if they think that the bytes before a relocation look like a `mov` instruction. This corrupts our metadata. rdar://148168098
|
I already know the PR testing will fail because we've currently got half of the changes for the new OS version numbering. However: @swift-ci Please smoke test |
|
I mean, I hadn't expected it to fail quite that quickly. Let's try again. |
|
@swift-ci Please smoke test |
|
Please test with the following PR: @swift-ci Please smoke test |
| if (!IGM.getOptions().UseJIT && | ||
| (!IGM.Triple.isOSDarwin() || IGM.Triple.getArch() != llvm::Triple::x86)) { | ||
| (!IGM.Triple.isOSDarwin() || IGM.Triple.getArch() != llvm::Triple::x86) && | ||
| (!IGM.Triple.isOSBinFormatELF() || !IGM.Triple.isX86())) { |
There was a problem hiding this comment.
Do you happen to know offhand if the linker on other ELF platforms (like FreeBSD) is also affected by this bug?
There was a problem hiding this comment.
FreeBSD uses lld, which does have this problem (and, unlike the BFD linker, while it understands -no-relax and won't complain about it, it ignores it for the purposes of this particular optimisation).
Basically, lld, gold and the BFD linker all do this; only the BFD linker has a way to disable it, and none of them test whether they're looking at a code section first. They do all have a handful of things they exempt from this behaviour (though the exact set differs; BFD and lld have more exceptions than gold does), but sadly none of them check whether they're in a code section first.
I'm aware that in principle there might be a platform using some other ELF linker that doesn't do this, but even in that case, someone might choose to use one of the three linkers above that all do.
(It looks to me like mold may be such a linker, just at a quick glance.)
|
Please test with the following PR: @swift-ci Please smoke test Windows platform |
|
@swift-ci Please smoke test Windows platform |
Unfortunately, x86 ELF linkers like to optimize GOTPCREL relocations by replacing
movinstructions that go via the GOT withleainstructions that do not.That would be fine, but they aren't very selective and will happily perform this transformation in non-code sections if they think that the bytes before a relocation look like a
movinstruction.This corrupts our metadata.
rdar://148168098