Another conceptual bug in ASDF.
Checking the plans generated for load-fasl-op, I find that with my new ASDF 2.26.x (currently x=45), it systematically loads all the individual .fasl's before it loads the .system.fasl, defeating the purpose.
This was not the case before: in 2.26.7 and earlier, the (separately distributed before 2.26.7) asdf-bundle has load-fasl-op depend-on fasl-op which depends-on compile-op, which doesn't depend-on load-op, but do-first's on it — so that the dependency is only considered but if the immediate timestamp of the fasl is older than the timestamp of its lisp source. If all fasl's satisfy this local timestamp check, nothing is recompiled, none of the do-first is triggered, and the load-op's are not attempted. But the old design also fails to propagate timestamp information from the other files that are depended-on, let alone do-firsted. Therefore, there is some notion in the old do-first that must be preserved in the new ASDF, even though the previous design was conceptually flawed overall.
In the end, what must be preserved is that some dependencies need to have been performed in the current image, whereas other dependencies only need to have been performed in some previous (or current) image. This means a dependency can be walked over twice: once in without "image-necessity", and a second time with it. Without image-necessity, only the timestamp matters; with image-necessity, whether it was performed in current image matters, too. Presumably, a node with output-files clears the image-necessity flag before checking timestamps, but re-traverses its dependencies with the flag if the timestamp check fails. A node with no output-files on the other hand is done for in-image effect, and heeds and propagates the image-necessity flag to its subnodes. I believe that it's possible to continue the current strategy of traversal without an explicit graph, though nodes may now be traversed twice: once without image-necessity, once with. Sigh.
—♯ƒ • François-René ÐVB Rideau •Reflection&Cybernethics• http://fare.tunes.org Success is getting what you want. Happiness is wanting what you get. — Dale Carnegie
Dear CL hackers,
this issue with needed-in-image vs not dependencies is solved in 2.26.51; a new refactoring of traverse was required, which led to new simplifications afterwards. It's simpler than ever despite doing more (and more correctly) than ever — see function traverse-action for the substance. POIU was updated, and once again, making POIU work helped make the ASDF code much better, and once again, some of antifuchs' code that I had once deleted anticipated things I ended up needing.
I *really* hope 2.27 will be the last deep refactoring of ASDF internals that's ever needed.
Now I'll go back to fixing any ASDF client that depends on obsolete behavior so I may release at long last. I hope this release can then make it to ECL, killing asdf-bundle and asdf-ecl, as well as to other implementations.
PS: For very bad reasons, I resurrected the GCL 2.6 port, which sadly is lacking support for output-translations (lack for pathname-match-p).
—♯ƒ • François-René ÐVB Rideau •Reflection&Cybernethics• http://fare.tunes.org When you've seen one nuclear war, you've seen them all.
On Mon, Dec 31, 2012 at 12:57 PM, Faré fahree@gmail.com wrote:
Another conceptual bug in ASDF.
Checking the plans generated for load-fasl-op, I find that with my new ASDF 2.26.x (currently x=45), it systematically loads all the individual .fasl's before it loads the .system.fasl, defeating the purpose.
This was not the case before: in 2.26.7 and earlier, the (separately distributed before 2.26.7) asdf-bundle has load-fasl-op depend-on fasl-op which depends-on compile-op, which doesn't depend-on load-op, but do-first's on it — so that the dependency is only considered but if the immediate timestamp of the fasl is older than the timestamp of its lisp source. If all fasl's satisfy this local timestamp check, nothing is recompiled, none of the do-first is triggered, and the load-op's are not attempted. But the old design also fails to propagate timestamp information from the other files that are depended-on, let alone do-firsted. Therefore, there is some notion in the old do-first that must be preserved in the new ASDF, even though the previous design was conceptually flawed overall.
In the end, what must be preserved is that some dependencies need to have been performed in the current image, whereas other dependencies only need to have been performed in some previous (or current) image. This means a dependency can be walked over twice: once in without "image-necessity", and a second time with it. Without image-necessity, only the timestamp matters; with image-necessity, whether it was performed in current image matters, too. Presumably, a node with output-files clears the image-necessity flag before checking timestamps, but re-traverses its dependencies with the flag if the timestamp check fails. A node with no output-files on the other hand is done for in-image effect, and heeds and propagates the image-necessity flag to its subnodes. I believe that it's possible to continue the current strategy of traversal without an explicit graph, though nodes may now be traversed twice: once without image-necessity, once with. Sigh.
—♯ƒ • François-René ÐVB Rideau •Reflection&Cybernethics• http://fare.tunes.org Success is getting what you want. Happiness is wanting what you get. — Dale Carnegie
-
Faré