On 16.02.2012, at 18:28, Scott Wood wrote:

On 02/16/2012 11:18 AM, Alexander Graf wrote:

quoted

On 16.02.2012, at 17:58, Scott Wood wrote:

quoted

On 02/16/2012 04:24 AM, Alexander Graf wrote:

quoted

On 16.02.2012, at 10:26, Liu Yu [off-list ref] wrote:

quoted

+_GLOBAL(epapr_ev_idle)
+epapr_ev_idle:
+    rlwinm    r3,r1,0,0,31-THREAD_SHIFT    /* current thread_info */
+    lwz    r4,TI_LOCAL_FLAGS(r3)    /* set napping bit */
+    ori    r4,r4,_TLF_NAPPING    /* so when we take an exception */
+    stw    r4,TI_LOCAL_FLAGS(r3)    /* it will return to our caller */
+
+    wrteei    1
+
+idle_loop:
+    LOAD_REG_IMMEDIATE(r11, HC_VENDOR_EPAPR | HC_EV_IDLE)
+
+.global epapr_ev_idle_start
+epapr_ev_idle_start:
+    li    r3, -1
+    nop
+    nop
+    nop

Can't you just bl into epapr_hypercall_start? You don't even have to save the old lr. because we never return anyways :)

The interrupt will branch to LR, so no, we can't trash it or put it
anywhere else.

Hrm. But we can clobber ctr, right? So how about we make the generic version do a bctr and then just do a small C wrapper that takes lr, moves it to ctr and branches to the generic one?

If it's just for this, I would say don't mess with the normal hcall path
for the sake of idle.  If using CTR would let us get away without
creating a stack frame in call sites, maybe that would be worthwhile,
depending on what sort of hcalls we end up having.

quoted

Then we don't have to replicate the hypercall code all over again for every invocation.

We shouldn't need to do it for every invocation.  Idle is special due to
the TLF_NAPPING hack.

Famous last words. If it's the only case, duplication should be ok. Let's hope there are no others.


Alex