tnn+nbsd@nygren.pp.se wrote:

> >> +    ldq_u    t1, UVM_PAGE_IDLE_ZERO(t0)
> >> +    lda    t0, UVM_PAGE_IDLE_ZERO(t0)    /* should we zero some 
> >> pages? */
> >> +    extbl    t1, t0, t0
> >> +    and    t0, 0xff, t0
> >>      beq    t0, 3f                /* nope. */
> >>      CALL(uvm_pageidlezero)
> >>  3:    ldl    t0, sched_whichqs        /* look for non-empty queue */

> > AARM section (I) 4-53 says that's the intended sequence for loading
> > a byte from unaligned storage, so it should be ok.

> Actually that final "and" is not required. Did gcc emit that?

Hmm. I added the following code at the end of uvm_init() in uvm_init.c:
---
	if (uvm.page_idle_zero)
		printf("foo");
---

then gcc -O0 emitted:
---
 15c:   15 00 41 2c     ldq_u   t1,21(t0)
 160:   15 00 21 20     lda     t0,21(t0)
 164:   c1 00 41 48     extbl   t1,t0,t0
 168:   01 f0 3f 44     and     t0,0xff,t0
 16c:   06 00 20 e4     beq     t0,188 <uvm_init+0x188>
---

but now gcc -O1 says:
---
 144:   15 00 29 2c     ldq_u   t0,21(s0)
 148:   15 00 29 21     lda     s0,21(s0)
 14c:   c1 00 29 48     extbl   t0,s0,t0
 150:   06 00 20 e4     beq     t0,16c <uvm_init+0x16c>
---

and -O2:
---
 134:   15 00 29 2c     ldq_u   t0,21(s0)
 138:   15 00 29 21     lda     s0,21(s0)
 13c:   c1 00 29 48     extbl   t0,s0,t0
 140:   1f 04 ff 47     nop     
 144:   06 00 20 e4     beq     t0,160 <uvm_init+0x160>
---

I'm not sure why gcc -O2 puts an extra nop, but I think
you are correct.
---
Izumi Tsutsui