-
Notifications
You must be signed in to change notification settings - Fork 165
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
On RISC-V, lazily bound functions must follow the ABI #66
Comments
I'd phrase it as "Functions exported from shared libraries must follow the standard calling convention. Alternate calling conventions are permitted only for functions with internal linkage." |
The glibc people were strict about restricting this to lazily bound functions, as those aren't the default for shared libraries on some systems. IDK if those actually matter, though, as they're probably pretty esoteric. |
Fair enough |
On Tue, Dec 19, 2017 at 6:02 PM, Palmer Dabbelt ***@***.***> wrote:
The glibc people were strict about restricting this to lazily bound
functions, as those aren't the default for shared libraries on some
systems. IDK if those actually matter, though, as they're probably pretty
esoteric.
There is a linker option -z now that turns off lazy binding at compile
time. There is also a glibc environment variable LD_BIND_NOW=1 that turns
off lazy binding at run time. Probably not many people other than
toolchain developers know about this stuff, though I do see some references
to LD_BIND_NOW in some FAQs found via a web search, so some people do find
it occasionally useful.
Jim
|
|
It had never occurred to me that people might do something other than use the standard calling convention for lazily bound functions. I'm not entirely sure where it should go. A note under "procedure calling convention" perhaps? Phrasing it as "must follow the standard calling convention" might be confusing - with RISC-V we of course have a range of standard calling conventions, and the requirement is really to conform to the specific standard calling convention used on that system. |
I'm not convinced it really makes sense to document this? We define the calling convention for each ABI, and compilers have to implement that. If they want to perform optimisations that result in a different calling convention then they can do whatever they like provided it is a sound optimisation, just like anything else. Calling a lazily-bound function with a non-standard ABI is not a sound optimisation if it uses any of the call-clobbered non-argument registers. If we end up extending the psABI to use veneers for anything like AAPCS32/64 then we will need to be careful there to ensure we don't end up with veneers clobbering non-standard argument registers, but we do not have such a thing today. |
IMO this is more of a relocation issue than a calling convention issue: regardless of whether a symbol is a function under the standard calling convention, the act of resolving a lazy call imposes the standard calling on that symbol. This is important for two reasons:
The worry was that this may come up in the context of the V extension, where users may want non-standard calling conventions (maybe to pass vector arguments in registers, wouldn't really be describable in a C calling convention) and dynamic loader developers may want to rely on the standard calling convention (do to a vector memcpy without saving the V state, for example). As it's likely that users end up with V binaries way before glibc contains any V instructions, the glibc change would break user code and without something concrete to point at in a spec we'd have to argue. |
In fact, we (SiFive) have idea for alternative (or you can say it's extension) calling convention for vector function, the motivation is normal function call[1] is treat all register are caller-save which means all vector registers are not preserved across calls, and we want another calling convention like SVE's AAVPCS(aarch64_vector_pcs), which will not all caller-save for vector registers. So I guess we might have that issue if we introduce another vector CC, Jim also point me ARM folks has hit this issue before, and seems that require bunch of fix for that[2], we might take more time for that. [1] #171 |
This was inspired by an x86 toolchain bug report. Looking at that bug report now, I see that they originally refused to support non-standard calling sequence in the resolver, and then 6 months later a patch was checked in to make this work by using the fxsave or xsave or xsavec instruction as appropriate to save/restore all registers in the resolver routine. Testing showed that there was no appreciable performance difference from the change, as this only happens once for each function called via the plt. The issue with V is similar to the x86 case. If we want an optimized C library that is allowed to use V instructions when it makes sense, and if we want to allow optimized user code that passes arguments in V registers when it makes sense, then the resolver needs to save/restore some (or all) of the V registers. Though the variable size of V registers makes this more complicated. It probably makes more sense to wait until we have a use case for the feature so we can better define what we need to make it work. The aarch64 solution Kito pointed at marks vector routines as having a variant ABI in the symbol table, and then only those routines need the resolver to save the vector registers. We probably should implement a similar solution. |
It seems very short sighted to look at the x86 ABI break and the arm64 ABI break and declare that we shouldn't write anything down. If anything we should at least codify what the X and F register clobbers are on lazily bound PLT calls, as those are already de-facto ABI since glibc has the weakest correct implementation for both. If we were to support non-standard X or F calling conventions we'd need some ABI extension (the arm64 decorated symbols seem reasonable) -- I doubt that's worth it, though, as the non-standard calling conventions for X and F registers probably won't be worth supporting on systems with dynamic linking. If you don't want to write anything down for the V ABIs then that's fine, I'll just re-open this when the bugs show up ;) |
I guess we should document down that those discussion, especially remind us that we should very carefully to handle new calling conversion with new register set like So keep this opened until we document that :) @jrtc27 or me will written down that into the policy of add new calling convention (in near future I hope). |
I have written a draft about the lazily bound function in #190. |
See this x86 ABI bug for more details: https://sourceware.org/bugzilla/show_bug.cgi?id=21265
I vote we solve this on RISC-V by just requiring that lazily bound functions follow the standard ABI. This lets us avoid saving the rest of the X registers, and assuming we can do the fixup without any extensions (which seems reasonable, and is what we currently do despite it not actually being enforced) we won't need to save F or V state here.
I don't know where this should go in the manual, @asb?
The text was updated successfully, but these errors were encountered: