Old Java PRs

Mon May 12 18:11:00 GMT 2003

On 2003年5月12日, Andrew Haley wrote:
> > Unless I misunderstand this PR, it is fixed by the new inliner (except for
> > the bytecode frontend, which still does not emit whole functions as
> > trees).
>> It isn't. The new inliner still will not inline functions until they
> have been expanded.

Unfortunately the PR doesn't include a test case. This is what I tried:
$ cat A.java
class A {
 int _y;
 final int x() {
 return y() + 1;
 }
 final int y() {
 return _y;
 }
}
$ gcj -O3 -S -g0 -fno-exceptions -fomit-frame-pointer A.java
$ more A.s
...
_ZN1A1xEv:
.LFB2:
 subl 12,ドル %esp
.LCFI0:
 movl 16(%esp), %eax
 movl %eax, (%esp)
 call _ZN1A1yEv
 incl %eax
 addl 12,ドル %esp
 ret
Reversing the declaration order of x and y, I get:
_ZN1A1xEv:
.LFB3:
 movl 4(%esp), %eax
 movl 4(%eax), %eax
 incl %eax
 ret
So it looks as though you're right, though I could swear I've seen this
work before. My understanding of the tree inliner is that it does not
require expanding the inlined bodies to RTL, but I don't remember the
details.
Jeff