[pro] Default safety level and specialized arrays
Hi everybody, After rummaging around in the HyperSpec and various implementation manuals I came to the following conclusion: 1. The standard doesn't prescribe such thing as a default safety level for CL implementations and there's no portable way to figure out the safety level of the global environment. (And the same holds for other optimize qualities.) The closest I could find in the CLHS is 3.3.1 where they talk about an "implementation that always processes code as if safety were high." From this I conclude that implementations are allowed NOT to do this or otherwise there would be no need to mention it. The LispWorks manual for example describes the defaults in section 9.5 about compiler control, but for SBCL I couldn't find them in the manual. 2. The standard doesn't prescribe what should happen if you try to write an element of a wrong type to a specialized array, even in safe code. Both LispWorks and SBCL will by default complain in this case but my understanding is that you can't rely on this in portable code. So, my question is whether the above is correct or if I missed something? Thanks, Edi.
On 10 Mar 2015, at 17:02, Edi Weitz <edi@weitz.de> wrote:
Hi everybody,
After rummaging around in the HyperSpec and various implementation manuals I came to the following conclusion:
1. The standard doesn't prescribe such thing as a default safety level for CL implementations and there's no portable way to figure out the safety level of the global environment. (And the same holds for other optimize qualities.) The closest I could find in the CLHS is 3.3.1 where they talk about an "implementation that always processes code as if safety were high." From this I conclude that implementations are allowed NOT to do this or otherwise there would be no need to mention it.
The LispWorks manual for example describes the defaults in section 9.5 about compiler control, but for SBCL I couldn't find them in the manual.
2. The standard doesn't prescribe what should happen if you try to write an element of a wrong type to a specialized array, even in safe code.
Both LispWorks and SBCL will by default complain in this case but my understanding is that you can't rely on this in portable code.
So, my question is whether the above is correct or if I missed something?
Yes. Notably, with (safety 0), Common Lisp implementation could be as good as common C implementations... This is why it is very bad to put optimization declarations in libraries (well, if they're not (safety 3) (speed 0)): it HAS and WILL make users lose days in debugging because YOUR, Edi, libraries, compiled on conforming implementations fail to detect errors. -- __Pascal J. Bourguignon__
On 2015-03-10 21:15:31+0100, Pascal J. Bourguignon wrote:
On 10 Mar 2015, at 17:02, Edi Weitz <edi@weitz.de> wrote:
After rummaging around in the HyperSpec and various implementation manuals I came to the following conclusion:
1. The standard doesn't prescribe such thing as a default safety level for CL implementations and there's no portable way to figure out the safety level of the global environment. (And the same holds for other optimize qualities.) The closest I could find in the CLHS is 3.3.1 where they talk about an "implementation that always processes code as if safety were high." From this I conclude that implementations are allowed NOT to do this or otherwise there would be no need to mention it.
The LispWorks manual for example describes the defaults in section 9.5 about compiler control, but for SBCL I couldn't find them in the manual.
2. The standard doesn't prescribe what should happen if you try to write an element of a wrong type to a specialized array, even in safe code.
Both LispWorks and SBCL will by default complain in this case but my understanding is that you can't rely on this in portable code.
So, my question is whether the above is correct or if I missed something?
Yes.
Notably, with (safety 0), Common Lisp implementation could be as good as common C implementations...
This is why it is very bad to put optimization declarations in libraries (well, if they're not (safety 3) (speed 0)): it HAS and WILL make users lose days in debugging because YOUR, Edi, libraries, compiled on conforming implementations fail to detect errors.
Hello! This begs the question: how do I get a speed optimized library then? Do I need to make a local fork just to insert declarations? Is there a way to tell e. g. ASDF to use certain declamations? Is making the declarations in the library dependent on documented entries to *features* the way to go? Where would I push the desired settings into *features*? Yours aye Svante -- Svante von Erichsen GPG fingerprint: A78A D4FB 762F A922 A495 57E8 2649 9081 6E61 20DE
On 11 Mar 2015, at 21:25, Svante v. Erichsen <Svante.v.Erichsen@web.de> wrote:
This begs the question: how do I get a speed optimized library then? Do I need to make a local fork just to insert declarations? Is there a way to tell
This is not up to the library writer to decde on this. This is up to the library final user. Declarations are global. When you are developping/debugging, you will set them to (optimize (speed 0) (space 0) (debug 3) (safety 3)). When generating the end-user executable, you will set them to (optimize (speed 3) (space 3) (debug 0) (safety 3)), if you follow Bertrand Meyer's advice.
e. g. ASDF to use certain declamations? Is making the declarations in the library dependent on documented entries to *features* the way to go? Where would I push the desired settings into *features*?
Even if you set them in ASDF, you will have to have a way to tell it when you are debugging code (always), and when you are generating a end-user application (once in a blue moon, and in a different, explicitely controlled process). And yes, since asdf doesn't cache the compiled files in separate directories for the various optimization settings (or the various *features* list), since that would be very unwiedly, you have to clear the cache when you switch between generating the end-user application, and normal developping work. rm -rf ~/.cache/common-lisp/ is found in all application generation makefiles. Optimize declaimations and declarations in libraries is the curse of the CL application developper. -- __Pascal J. Bourguignon__
On Thu, Mar 12, 2015 at 01:14:41PM +0100, Pascal J. Bourguignon wrote:
And yes, since asdf doesn't cache the compiled files in separate directories for the various optimization settings (or the various *features* list), since that would be very unwiedly, you have to clear the cache when you switch between generating the end-user application, and normal developping work. rm -rf ~/.cache/common-lisp/ is found in all application generation makefiles.
Hmm, I think ASDF_OUTPUT_TRANSLATIONS still works, so my makefile looks like: LISP := "/usr/local/bin/sbcl" ARGS := "--non-interactive" debug: export ASDF_OUTPUT_TRANSLATIONS = "/:/tmp/lisp-cache/$@/" debug: $(LISP) $(ARGS) --load $@.lisp --load build.lisp release: export ASDF_OUTPUT_TRANSLATIONS = "/:/tmp/lisp-cache/$@/" release: $(LISP) $(ARGS) --load $@.lisp --load build.lisp where debug.lisp / release.lisp would contain the wanted optimization settings. HTH Ralf Mattes
: Svante v. Erichsen <Svante.v.Erichsen@web.de> : Pascal J. Bourguignon <pjb@informatimago.com>
This begs the question: how do I get a speed optimized library then? Do I need to make a local fork just to insert declarations? Is there a way to tell
This is not up to the library writer to decde on this. This is up to the library final user.
Declarations are global. When you are developping/debugging, you will set them to (optimize (speed 0) (space 0) (debug 3) (safety 3)).
When generating the end-user executable, you will set them to (optimize (speed 3) (space 3) (debug 0) (safety 3)), if you follow Bertrand Meyer's advice.
e. g. ASDF to use certain declamations? Is making the declarations in the library dependent on documented entries to *features* the way to go? Where would I push the desired settings into *features*?
Even if you set them in ASDF, you will have to have a way to tell it when you are debugging code (always), and when you are generating a end-user application (once in a blue moon, and in a different, explicitely controlled process).
And yes, since asdf doesn't cache the compiled files in separate directories for the various optimization settings (or the various *features* list), since that would be very unwiedly, you have to clear the cache when you switch between generating the end-user application, and normal developping work. rm -rf ~/.cache/common-lisp/ is found in all application generation makefiles.
Optimize declaimations and declarations in libraries is the curse of the CL application developper.
For an involved way of solving this problem as part of a reproducible build of a company's software, see how we do it in the quux snapshot: https://common-lisp.net/project/qitab/ http://common-lisp.net/project/qitab/archives/quux-2013-09-24.tar.bz2 —♯ƒ • François-René ÐVB Rideau •Reflection&Cybernethics• http://fare.tunes.org The [classical] liberal, of course, does not deny that there are some superior people — he is not an egalitarian — but he denies that anyone has authority to decide who these superior people are. — F. A. Hayek, "Why I Am Not a Conservative"
participants (5)
-
Edi Weitz -
Faré -
Pascal J. Bourguignon -
Ralf Mattes -
Svante v. Erichsen