[llvm][patch] Adjust behavior of FDE cross-section relocs for targets that don't support abs-differences.
iain at codesourcery.com
Tue Dec 17 14:50:02 PST 2013
On 17 Dec 2013, at 22:28, Nick Kledzik wrote:
> On Dec 12, 2013, at 3:57 PM, Iain Sandoe <iain at codesourcery.com> wrote:
>> @jim - although this was reported against an old target, I believe it potentially effects any target using non-trivial relocs for Start-PC and LSDAs in FDEs.
>> On 3 Dec 2013, at 15:59, Iain Sandoe wrote:
>>> (I hope I have directed this to the correct reviewers)
>>> At the dev. conf. David reported to me that EH was not working on ppc-darwin8. Once we checked a few more platforms, it turns out it's not working with any ld64 < 97.17 (so OSX < 10.6 on both x86/ppc).
>>> The reason is that the older ld/ld64 versions require proper scattered relocs for expressions in CIE/FDE symbol refs that cross section boundaries.
>>> Later versions of the ld64 allow that these may be replaced by an absolute difference***
> I can tell you the motivation for the linker change. ld64 takes the “smart linker” approach. If fully breaks up the __eh_frame section and associates each FDE with its owning function. Since it knows how to parse an FDE, a relocation is redundant and only slows down the linker. On the other hand, a “dumb linker” which does not know anything special about the __eh_frame section would need relocations.
indeed, but (as far as I am aware) the only such smart linker is ld64 >= 97.1 …
… so should the patch be applied to allow other linkers to function?
>>> AFAIK, this behaviour is specific to Darwin/OSX (and ld64 >= 97.17), I suspect that any other target that elected to use a non-trivial FDE symbol encoding would likely be surprised by the abs values. However, these are hardwired in MCDwarf.cpp, as of now.
>>> NOTE1: in general, this would not cause a linkage error for the target - 'just' a failure to unwind at runtime.
>>> NOTE2: it also seems that the abs-ification is not done even for x86-64-darwin12 for compact unwind.
>>> The solution I am proposing is to introduce a "DwarfFDESymbolsUseAbsDiff" boolean value in MCAsmInfo and to use that to control this behaviour. In the patch, this is only switched on for X86 Darwin/OSX & for OSX >= 10.6. It should really be contingent on -target-linker-version >= 97.17, but this information is not yet passed to the target.
>>> With the patch EH works on *-darwin9 with the native linker (and still emits the same set of relocs on darwin12).
>>> Is this a reasonable solution?
>>> *** P.S.
>>> I assume that these extra hoops and inconsistencies (and presumably matching ones in ld64) are really worthwhile?
>>> I.E. that there is some measurable performance gain?
>>> (if not, we could just elect to emit the relocs and drop this special casing).
>>> llvm-commits mailing list
>>> llvm-commits at cs.uiuc.edu
>> llvm-commits mailing list
>> llvm-commits at cs.uiuc.edu
> llvm-commits mailing list
> llvm-commits at cs.uiuc.edu
More information about the llvm-commits