The proper way to do it would be moving the ZEND_VM_NEXT_OPCODE_EX (which does not check EG(exception) itself in non-debug, just EX(opline) vs %r15) before the exception check. (As that avoids the branch completely on that path.)

I don't see how that's better

Seems less expensive to me to just load EX(opline) than checking if(EG(exception))? Or am I mistaken?

Seems less expensive to me to just load EX(opline) than checking if(EG(exception))? Or am I mistaken?

I agree with this statement.
However, if I move the if (EXPECTED(opline->extended_value == 0 && (opline+1)->opcode == ZEND_DO_FCALL)) { check upwards then I still need the EG(exception) check anyway for the stack push...

Yes sure, but only for the case where there are actually args for the object construction.

Yes sure, but only for the case where there are actually args for the object construction.

It clicked. Thanks for pointing it out

I don't see how the new code is better.
For the case with arguments and no exception it performs exactly the same checks just in different order.
Case with exception becomes even more expensive...
May be I miss something?

No, I had originally changed ZEND_VM_NEXT_OPCODE_EX so that check_exception is 0. And then undid that and swapped the checks. I suppose the first idea was better as it kept the exception code path the same cost and changed an EX(opline) access in just an online access.

I suppose the first idea was better

I think so.

@dstogov I'm confused. ZEND_VM_NEXT_OPCODE_EX returns. The case with exception and arguments is slightly more expensive, yes. The case without arguments should be less expensive. And the case with arguments and no exceptions should be unchanged.

Correct?

@dstogov I'm confused. ZEND_VM_NEXT_OPCODE_EX returns. The case with exception and arguments is slightly more expensive, yes. The case without arguments should be less expensive. And the case with arguments and no exceptions should be unchanged.

Correct?

No. The case without arguments becomes more expensive.

A single ZEND_VM_NEXT_OPCODE_EX(0...) would be better than single EG(exception) check, but in this case ZEND_VM_NEXT_OPCODE_EX() is under two additional checks and instead of a single check we perform two.

The case without arguments and an exception. The case without arguments without exception is cheaper, right?

And IIRC, we optimize for the happy path, not the exception?

The case without arguments and an exception. The case without arguments without exception is cheaper, right?

right.

And IIRC, we optimize for the happy path, not the exception?

right. I was wrong.