RFR (L): 7173584: Implement arraycopy as a macro node
Roland Westrelin
roland.westrelin at oracle.com
Fri Sep 28 07:51:37 PDT 2012
Vladimir,
Thanks for taking a look at this.
> I am not comfortable with this predicate change. The check come from generate_limit_guard() :
>
> if (a.length u< offset+copy_length) uncommon_trap; else arraycopy;
>
> which is reversed by igvn to
>
> if (offset+copy_length u< a.length) arraycopy; else uncommon_trap;
>
> So it should not be IfFalse projection. Can you explain the case which you hit?
This is the java code that triggers this:
for (int i=0; i<a1.length; i++) {
a2 = new …[a3.length+a1.length-i-1];
System.arraycopy(a3, 0, a2, 0, a3.length);
System.arraycopy(a1, i+1, a2, a3.length, a1.length-i-1);
}
For the first arraycopy, one of the check is:
if (a2.length < a3.length) uncommon_trap; else arraycopy;
if (a3.length + a1.length-i-1 < a3.length) uncommon_trap; else arraycopy;
a1.length is 5
if (a3.length + 4 - i < a3.length) uncommon_trap; else arraycopy;
a3.length is a loop invariant.
> Also instead of duplicating gen_subtype_check() in macroArrayCopy.cpp can you modify the original one to work for you (by passing additional flag)?
Ok I'll do that.
Roland.
More information about the hotspot-compiler-dev
mailing list