On Thu, Jun 15, 2017 at 7:17 PM, H.J. Lu <hjl.to...@gmail.com> wrote: > On Thu, Jun 15, 2017 at 4:28 PM, Andy Lutomirski <l...@kernel.org> wrote: >> On Thu, Jun 15, 2017 at 4:11 PM, H.J. Lu <hjl.to...@gmail.com> wrote: >>> On Thu, Jun 15, 2017 at 3:45 PM, Andy Lutomirski <l...@kernel.org> wrote: >>>> On Thu, Jun 15, 2017 at 3:40 PM, H.J. Lu <hjl.to...@gmail.com> wrote: >>>>> On Thu, Jun 15, 2017 at 3:18 PM, Andy Lutomirski <l...@kernel.org> wrote: >>>>>> On Thu, Jun 15, 2017 at 7:33 AM, Dave Hansen <dave.han...@intel.com> >>>>>> wrote: >>>>>>> On 06/14/2017 10:18 PM, Andy Lutomirski wrote: >>>>>>>> Dave, why is XINUSE exposed at all to userspace? >>>>>>> >>>>>>> You need it for XSAVEOPT when it is using the init optimization to be >>>>>>> able to tell which state was written and which state in the XSAVE buffer >>>>>>> is potentially stale with respect to what's in the registers. I guess >>>>>>> you can just use XSAVE instead of XSAVEOPT, though. >>>>>>> >>>>>>> As you pointed out, if you are using XSAVEC's compaction features by >>>>>>> leaving bits unset in the requested feature bitmap registers, you have >>>>>>> no idea how much data XSAVEC will write, unless you read XINUSE with >>>>>>> XGETBV. But, you can get around *that* by just presizing the XSAVE >>>>>>> buffer to be big. >>>>>> >>>>>> I imagine that, if you're going to save, do something quick, and >>>>>> restore, you'd be better off allocating a big buffer rather than >>>>>> trying to find the smallest buffer you can get away with by reading >>>>>> XINUSE. Also, what happens if XINUSE nondeterministically changes out >>>>>> from under you before you do XSAVEC? I assume you can avoid this >>>>>> becoming a problem by using RFBM carefully. >>>>>> >>>>>>> >>>>>>> So, I guess that leaves its use to just figuring out how much XSAVEOPT >>>>>>> (and friends) are going to write. >>>>>>> >>>>>>>> To be fair, glibc uses this new XGETBV feature, but I suspect its >>>>>>>> usage is rather dubious. Shouldn't it just do XSAVEC directly rather >>>>>>>> than rolling its own code? >>>>>>> >>>>>>> A quick grep through my glibc source only shows XGETBV(0) used which >>>>>>> reads XCR0. I don't see any XGETBV(1) which reads XINUSE. Did I miss >>>>>>> it. >>>>>> >>>>>> Take a look at sysdeps/x86_64/dl-trampoline.h in a new enough version. >>>>> >>>>> I wrote a test to compare latency against different approaches. This >>>>> is on Skylake: >>>>> >>>>> [hjl@gnu-skl-1 glibc-test]$ make >>>>> ./test >>>>> move : 47212 >>>>> fxsave : 719440 >>>>> xsave : 925146 >>>>> xsavec : 811036 >>>>> xsave_state_size: 1088 >>>>> xsave_state_comp_size: 896 >>>>> >>>>> load/store is about 17X faster than xsavec. >>>>> >>>>> I put my hjl/pr21265/xsavec branch at >>>>> >>>>> https://sourceware.org/git/?p=glibc.git;a=summary >>>>> >>>>> It uses xsave/xsave/xsavec in _dl_runtime_resolve. >>>> >>>> What is this used for? Is it just to avoid clobbering argument regs >>>> when resolving a symbol that uses an ifunc, or is there more to it? >>> >>> It is used for lazy binding the first time when an external function is >>> called. >>> >> >> Maybe I'm just being dense, but why? What does ld.so need to do to >> resolve a symbol and update the GOT that requires using extended >> state? > > Since the first 8 vector registers are used to pass function parameters > and ld.so uses vector registers, _dl_runtime_resolve needs to preserve > the first 8 vector registers when transferring control to ld.so. >
Wouldn't it be faster and more future-proof to recompile the relevant parts of ld.so to avoid using extended state? --Andy