[tree-ssa][java] fix some -O3 test cases

[Jeff Sturm]

On Mon, 1 Dec 2003, Richard Henderson wrote:
> -FAIL: TLtest -O3 compilation from source
> -FAIL: anfi -O3 compilation from source
> -FAIL: anon -O3 compilation from source
> -FAIL: anon2 -O3 compilation from source
> -FAIL: anon3 -O3 compilation from source
> -FAIL: anon4 -O3 compilation from source
> -FAIL: anon_ctor_itf_arg -O3 compilation from source
>
> All of these failures were for missing symbols.

Oddly enough, these didn't fail before 11-21.  I agree with your
assessment, but I'm still curious why this just started breaking then.

> It looks like this same fix may be required on mainline, but I don't
> know why these tests aren't failing there.  Would someone Java savvy
> please look into this?

Originally I had it the other way around on mainline too.  The problem
there was that the class data isn't complete until all functions are
expanded.  In particular, expanding NEW_ARRAY_INIT can add entries to the
constant pool.

Regrettably, the current situation is the result of a bit of
trial-and-error.  In hindsight all this cgraph stuff would've been easier
if we'd just lowered NEW_ARRAY_INIT earlier.  If you think it's important
enough on mainline I can pursue it.  My suspicion is that the hook that
forces an out-of-line copy of every method to be emitted is masking the
problem, though that won't prevent changing the calling conventions as you
say.

Jeff