On 10/31/2016 10:21 AM, Artyom Tarasenko wrote:
On Mon, Oct 31, 2016 at 5:06 PM, Richard Henderson <r...@twiddle.net> wrote:
On 10/31/2016 09:40 AM, Artyom Tarasenko wrote:
* The "Remove asi helper code handled inline" patch retains the code
within
ldda to handle asis that must be handled out of line. This fixes the
FreeBSD 10.3 boot problem. While the UA2007 spec (and thus sun4v?)
doesn't
allow for such, it would seem that US2 hardware does.
What ASI was failing? It may still be a part of sun4v CPUs if it's
described in UST1/UST2 supplements.
0x00000000c08e5680: clr %g4
0x00000000c08e5684: mov 1, %g1
0x00000000c08e5688: sllx %g1, 0x24, %o5
0x00000000c08e568c: mov -1, %g1
0x00000000c08e5690: sllx %g1, 0x23, %g1
0x00000000c08e5694: xor %g1, -128, %o4
0x00000000c08e5698: ldda [ %g4 ] #ASI_IC_TAG, %g0 ****
0x00000000c08e569c: mov %g1, %g2
0x00000000c08e56a0: and %g1, %o5, %g1
0x00000000c08e56a4: brz,a %g1, 0xc08e56dc
So indeed it would appear that FreeBSD does in fact want the half of the
result that would get stored into %g1. As far as I can see they could have
achieved the same result with
ldxa [%g4] #ASI_IC_TAG, %g1
srl %g1, 0, %g1
to stay within the nominal spec.
While your code is indeed more readable, I think the current code is
still within the spec:
UA2005 chap 7.59:
"Note: Execution of an LDTWA instruction with rd = 0 modifies only
R[1]."
What I mean is found in the description of LDTWA (looking at ua2011 here):
# Nontranslating ASIs (see page 395) should only be accessed
# using LDXA (not LDTWA) instructions. If an LDTWA
# referencing a nontranslating ASI is executed, per the above
# table, it generates a DAE_invalid_asi exception
# (impl. dep. #300-U4-Cs10).
ASI_IC_TAG is certainly nontranslating ASI.
All of the translating ASIs supported by LDTXA are already handled by
GET_ASI_DTWINX; all of the translating ASIs supported by LDTWA are already
handled by GET_ASI_DIRECT. By that reading the "default:" case should simply
raise DAE_invalid_asi.
r~