On 11/10/16 8:26 AM, Dicebot wrote:
On 11/03/2016 05:51 PM, Steven Schveighoffer wrote:
On 11/3/16 10:49 AM, Johan Engelen wrote:
On Wednesday, 2 November 2016 at 15:13:43 UTC, anonymous wrote:
On Wednesday, 2 November 2016 at 15:08:26 UTC, anonymous wrote:
I confirm, dmd 2.072 can't build dmd 2.071.2, same error, but boot
since master straping works there's probably something that's been
fixed in one or two of these ddmd modules, likely a static ctor...
Maybe after this:
https://github.com/dlang/dmd/commit/1d0ab8b9c136e46bf449c506ca25d2c8a784f7b9#diff-b4674e7b5d3a44178526afdefc9aa368
ddmd.attribs was removed
https://github.com/dlang/dmd/commit/1d0ab8b9c136e46bf449c506ca25d2c8a784f7b9#diff-b4674e7b5d3a44178526afdefc9aa368L15
and it was also part of the cycle.
Thanks for the detective work.
I wonder where the bug is: in 2.071 or in 2.072 :)
Any cycles that are "newly discovered" were already there. What was
happening is that the runtime did not detect the cycle, and was
arbitrarily choosing an order for initializing these modules.
It does not justify immediate breaking change.
What should have been
done instead:
- Keep old behavior by default by default but print non-fatal runtime
message that cycles are detected.
This is not possible. Old behavior DID detect some cycles. The new
algorithm detects ALL cycles, and handles them all in the same way. What
you are asking is for cycles detected by old algorithm to fail, but ones
that wouldn't have been detected to pass. Not to mention that the order
of ctor execution is not guaranteed to be the same (i.e. some latent bug
may be hiding there).
Only possibility is just to ignore ALL cycles, and print them if any are
detected.
I presume I have to look for commits that implement
http://dlang.org/changelog/2.072.0.html#drt-oncycle to fixup this?
The PR for this was here: https://github.com/dlang/druntime/pull/1668
-Steve
