On Fri, Oct 13, 2017 at 12:17:19PM +0100, Catalin Marinas wrote:
> On Thu, Oct 12, 2017 at 05:11:57PM +0100, Dave P Martin wrote:
> > On Wed, Oct 11, 2017 at 05:40:52PM +0100, Catalin Marinas wrote:
> > > On Tue, Oct 10, 2017 at 07:38:30PM +0100, Dave P Martin wrote:
> > > > diff --git a/arch/arm64/kernel/fpsimd.c b/arch/arm64/kernel/fpsimd.c
> > > > index aabeaee..fa4ed34 100644
> > > > --- a/arch/arm64/kernel/fpsimd.c
> > > > +++ b/arch/arm64/kernel/fpsimd.c
> > > > @@ -310,6 +310,32 @@ static void fpsimd_to_sve(struct task_struct *task)
> > > > sizeof(fst->vregs[i]));
> > > > }
> > > >
> > > > +/*
> > > > + * Transfer the SVE state in task->thread.sve_state to
> > > > + * task->thread.fpsimd_state.
> > > > + *
> > > > + * Task can be a non-runnable task, or current. In the latter case,
> > > > + * softirqs (and preemption) must be disabled.
> > > > + * task->thread.sve_state must point to at least sve_state_size(task)
> > > > + * bytes of allocated kernel memory.
> > > > + * task->thread.sve_state must be up to date before calling this function.
> > > > + */
> > > > +static void sve_to_fpsimd(struct task_struct *task)
> > > > +{
> > > > + unsigned int vq;
> > > > + void const *sst = task->thread.sve_state;
> > > > + struct fpsimd_state *fst = &task->thread.fpsimd_state;
> > > > + unsigned int i;
> > > > +
> > > > + if (!system_supports_sve())
> > > > + return;
> > > > +
> > > > + vq = sve_vq_from_vl(task->thread.sve_vl);
> > > > + for (i = 0; i < 32; ++i)
> > > > + memcpy(&fst->vregs[i], ZREG(sst, vq, i),
> > > > + sizeof(fst->vregs[i]));
> > > > +}
> > >
> > > Nit: could we actually just do an assignment with some pointer casting?
> > > It looks like we invoke memcpy for every 16 bytes (same for
> > > fpsimd_to_sve).
> >
> > I was uneasy about what the type of ZREG(sst, vq, i) ought to be.
> > In any case, memest() is magic: my oldskool GCC (5.3.0) generates:
> >
> > ffff000008084c70 <sve_to_fpsimd>:
> > ffff000008084c70: 14000004 b ffff000008084c80 <sve_to_fpsimd+0x10>
> > ffff000008084c74: d503201f nop
> > ffff000008084c78: d65f03c0 ret
> > ffff000008084c7c: d503201f nop
> > ffff000008084c80: f0007d61 adrp x1, ffff000009033000 <reset_devices>
> > ffff000008084c84: f942a021 ldr x1, [x1,#1344]
> > ffff000008084c88: 36b001c1 tbz w1, #22, ffff000008084cc0 <sve_to_fpsimd+0x50>
> > ffff000008084c8c: b94ca805 ldr w5, [x0,#3240]
> > ffff000008084c90: 912a0001 add x1, x0, #0xa80
> > ffff000008084c94: 91320004 add x4, x0, #0xc80
> > ffff000008084c98: f9465006 ldr x6, [x0,#3232]
> > ffff000008084c9c: 121c6ca5 and w5, w5, #0xfffffff0
> > ffff000008084ca0: 52800000 mov w0, #0x0 // #0
> > ffff000008084ca4: 8b2040c2 add x2, x6, w0, uxtw
> > ffff000008084ca8: 0b050000 add w0, w0, w5
> > ffff000008084cac: a9400c42 ldp x2, x3, [x2]
> > ffff000008084cb0: a8810c22 stp x2, x3, [x1],#16
> > ffff000008084cb4: eb01009f cmp x4, x1
> > ffff000008084cb8: 54ffff61 b.ne ffff000008084ca4 <sve_to_fpsimd+0x34>
> > ffff000008084cbc: d65f03c0 ret
> > ffff000008084cc0: d65f03c0 ret
> > ffff000008084cc4: d503201f nop
> >
> >
> > Without volatile, I think assigning a single object and doing a memcpy()
> > are equivalent to the compiler: which it actually uses depends solely on
> > optimisation considerations.
> >
> > (But then I'm not a language lawyer ... not a professional one anyway).
> >
> > Are you concerned compilers may mess this up?
>
> That's fine, please ignore my comment then. I was worried that gcc would
> always generate a call to the memcpy implementation rather than inlining
> it.
OK. I'll keep an eye on this, but in any case it won't impact the
FPSIMD-only case.
Cheers
---Dave
