On Fri, 19 Dec 2025 08:56:27 GMT, Tobias Hartmann <[email protected]> wrote:
>> I don't understand why it's not a problem on mainline, tho. And I agree it's
>> weird. I've looked, but it's hard to investigate why something works rather
>> than not (on a somewhat diverging codebase). I spent some time on that, but
>> eventually decided it's not that critical, as long as we have a way to make
>> it work for Valhalla too.
>
> I was suggesting that we should do what mainline does if
> `!InlineTypePassFieldsAsArgs`, would that make sense?
>
> I.e. do `store_check(masm, dst.base(), dst);`
That is the else-branch, right? Right now we have
if (tmp3 != noreg) {
__ mov(tmp3, dst.base());
store_check(masm, tmp3, dst);
} else {
// It's OK to corrupt the dst.base() register.
store_check(masm, dst.base(), dst);
}
If I understand well, you're suggesting to write
if (InlineTypePassFieldsAsArgs) {
if (tmp3 != noreg) {
__ mov(tmp3, dst.base());
store_check(masm, tmp3, dst);
} else {
// It's OK to corrupt the dst.base() register.
store_check(masm, dst.base(), dst);
}
} else {
// as mainline
store_check(masm, dst.base(), dst);
}
but to me, it looks equivalent to
if (InlineTypePassFieldsAsArgs && tmp3 != noreg) {
__ mov(tmp3, dst.base());
store_check(masm, tmp3, dst);
} else {
// It's OK to corrupt the dst.base() register.
store_check(masm, dst.base(), dst);
}
since both else-branch are identical.
And this, I experimentally found that it's failing quite a lot, with backtraces
as I mentioned. Is there an obvious mistake in my logic I'm missing?
-------------
PR Review Comment:
https://git.openjdk.org/valhalla/pull/1824#discussion_r2634271672
