On 13-09-08 2:04 PM, Richard Sandiford wrote:
Yvan Roux <yvan.r...@linaro.org> writes:
@@ -5786,7 +5796,11 @@ get_index_scale (const struct address_info *info)
&& info->index_term == &XEXP (index, 0))
return INTVAL (XEXP (index, 1));
- if (GET_CODE (index) == ASHIFT
+ if ((GET_CODE (index) == ASHIFT
+ || GET_CODE (index) == ASHIFTRT
+ || GET_CODE (index) == LSHIFTRT
+ || GET_CODE (index) == ROTATE
+ || GET_CODE (index) == ROTATERT)
&& CONST_INT_P (XEXP (index, 1))
&& info->index_term == &XEXP (index, 0))
return (HOST_WIDE_INT) 1 << INTVAL (XEXP (index, 1));
This bit doesn't look right. The scale is only 1 << N for (ashift ... N).
I think we have to stick to returning zero for the other codes.
Thanks, Richard. I missed this.
I am agree it might be a problem although the probability of it is
negligible as index reg hardly can be equivalent to something + constant
displacement (except may be case reg+reg when each reg can be base or
index but in this case we don't use shifts). Still returning zero is a
safe solution as it prevents such equivalence substitution completely.
Yvan, could modify this part of code and resend it to Richard Henderson
for approval.
The other two (snipped) rtlanal.c hunks like fine to me FWIW. Maybe now
would be a good time to add an "is this a shift code?" predicate though,
if we don't have one already.
Yes, I think we could somehow to promote this info to LRA but I don't
think it is necessary or urgent. There is already an alternative
solution -- just doing nothing as prohibiting equivalence substitution
for complicated index part hardly worsens the generated code.
