Re: [racket-dev] SGC as default
What about ./configure --enable-racket=`which racket` ? Already needed and used for cross-compilation. Tobias On Tue, 12 Aug 2014 18:00:21 +0200, Matthew Flatt mfl...@cs.utah.edu wrote: I'm not sure how difficult it will be. It's tedious enough that the last time I thought about it, I just left a note next to no-cgc-needed in racket/src/racket/Makefile.in, but maybe it's worth pursuing now. At Tue, 12 Aug 2014 04:39:55 -0700, Sam Tobin-Hochstadt wrote: How difficult would it be to allow the bootstrap process to use a preexisting Racket installation? This would alleviate some of the performance loss, for example in rebuilds by developers or in continuous integration. Sam On Aug 11, 2014 11:16 PM, Matthew Flatt mfl...@cs.utah.edu wrote: I've changed the Racket CGC implementation --- which is mostly used only to build the normal Racket variant --- to use SGC by default, instead of the Boehm GC. The intent of the switch is to make the more portable GC the default. If you have an existing build in a repo checkout, then `make` is likely to fail, because the makefile dependencies are not precise enough to deal with the switch. You can discard your old build directory, or it might work to simply delete builddir/racket/libmzgc.a If you're using CGC and want to continue using the Boehm GC, then provide `--disable-sgc` to `configure`. I've tuned SGC to bring its performance closer to the Boehm GC, but it's still slower. _ Racket Developers list: http://lists.racket-lang.org/dev _ Racket Developers list: http://lists.racket-lang.org/dev -- Tobias Hammer DLR / Robotics and Mechatronics Center (RMC) Muenchner Str. 20, D-82234 Wessling Tel.: 08153/28-1487 Mail: tobias.ham...@dlr.de _ Racket Developers list: http://lists.racket-lang.org/dev
Re: [racket-dev] SGC as default
Good point! I assumed (based on earlier conversations?) that Sam knows about `--enable-racket=...`. I thought that he was commenting on how the build process would create things that are needed only for `racketcgc`, even when `racketcgc` itself isn't needed. I've now pushed changes that should skip all `racketcgc`-specific pieces when `--enable-racket=...` is provided. At Wed, 13 Aug 2014 11:08:22 +0200, Tobias Hammer wrote: What about ./configure --enable-racket=`which racket` ? Already needed and used for cross-compilation. Tobias On Tue, 12 Aug 2014 18:00:21 +0200, Matthew Flatt mfl...@cs.utah.edu wrote: I'm not sure how difficult it will be. It's tedious enough that the last time I thought about it, I just left a note next to no-cgc-needed in racket/src/racket/Makefile.in, but maybe it's worth pursuing now. At Tue, 12 Aug 2014 04:39:55 -0700, Sam Tobin-Hochstadt wrote: How difficult would it be to allow the bootstrap process to use a preexisting Racket installation? This would alleviate some of the performance loss, for example in rebuilds by developers or in continuous integration. Sam On Aug 11, 2014 11:16 PM, Matthew Flatt mfl...@cs.utah.edu wrote: I've changed the Racket CGC implementation --- which is mostly used only to build the normal Racket variant --- to use SGC by default, instead of the Boehm GC. The intent of the switch is to make the more portable GC the default. If you have an existing build in a repo checkout, then `make` is likely to fail, because the makefile dependencies are not precise enough to deal with the switch. You can discard your old build directory, or it might work to simply delete builddir/racket/libmzgc.a If you're using CGC and want to continue using the Boehm GC, then provide `--disable-sgc` to `configure`. I've tuned SGC to bring its performance closer to the Boehm GC, but it's still slower. _ Racket Developers list: http://lists.racket-lang.org/dev _ Racket Developers list: http://lists.racket-lang.org/dev -- Tobias Hammer DLR / Robotics and Mechatronics Center (RMC) Muenchner Str. 20, D-82234 Wessling Tel.: 08153/28-1487 Mail: tobias.ham...@dlr.de _ Racket Developers list: http://lists.racket-lang.org/dev
Re: [racket-dev] SGC as default
Apparently, when gcc 4.9.1 sees memcpy(x, y, n); if (y) then it believes `y` can be assumed to be NULL --- even if `n` turns out to be zero --- and so the conditional can be optimized away. I'm surprised by that rule, but it's easy enough to move the test before the memcpy(). At Tue, 12 Aug 2014 02:00:04 -0400, Asumu Takikawa wrote: On 2014-08-12 06:06:51 +0100, Matthew Flatt wrote: What platform are you using? I imagine that running `./racketcgc` within the racket subdirectory of your build directory will similarly crash. Can you get any information from running `gdb racketcgc`? This is on Linux. Running the built executable immediately produces a segfault. Here's a backtrace from gdb: (gdb) bt #0 0x005e78db in free_managed (s=s@entry=0x0) at ../../../racket/sgc/sgc.c:1535 #1 0x005e7f63 in GC_add_roots (start=start@entry=0x881b70 collects_path, end=end@entry=0x881b79 run_cmd+1) at ../../../racket/sgc/sgc.c:1657 #2 0x00576f9a in scheme_register_static (ptr=ptr@entry=0x881b70 collects_path, size=size@entry=8) at ../../../racket/src/salloc.c:720 #3 0x0045b87f in scheme_set_collects_path (p=p@entry=0x77f701a0) at ../../../racket/src/file.c:6841 #4 0x00436673 in run_from_cmd_line (mk_basic_env=optimized out, cont_run=0x435a10 cont_run, _argv=optimized out, argc=0) at ../../racket/cmdline.inc:1444 #5 main_after_stack (data=data@entry=0x7fffdc30) at ../../racket/main.c:450 #6 0x0057694d in do_main_stack_setup (data=0x7fffdc30, _main=0x435a80 main_after_stack, no_auto_statics=1) at ../../../racket/src/salloc.c:198 #7 scheme_main_stack_setup (no_auto_statics=no_auto_statics@entry=1, _main=_main@entry=0x435a80 main_after_stack, data=data@entry=0x7fffdc30) at ../../../racket/src/salloc.c:310 #8 0x004346ee in main_after_dlls (argv=optimized out, argc=optimized out) at ../../racket/main.c:381 #9 main (argc=optimized out, argv=optimized out) at ../../racket/main.c:341 Cheers, Asumu _ Racket Developers list: http://lists.racket-lang.org/dev
Re: [racket-dev] SGC as default
The gcc 4.9 release notes warn about this optimization: https://gcc.gnu.org/gcc-4.9/porting_to.html I'm surprised that this change hasn't caused more trouble for us. At Tue, 12 Aug 2014 08:39:01 +0100, Matthew Flatt wrote: Apparently, when gcc 4.9.1 sees memcpy(x, y, n); if (y) then it believes `y` can be assumed to be NULL --- even if `n` turns out to be zero --- and so the conditional can be optimized away. I'm surprised by that rule, but it's easy enough to move the test before the memcpy(). At Tue, 12 Aug 2014 02:00:04 -0400, Asumu Takikawa wrote: On 2014-08-12 06:06:51 +0100, Matthew Flatt wrote: What platform are you using? I imagine that running `./racketcgc` within the racket subdirectory of your build directory will similarly crash. Can you get any information from running `gdb racketcgc`? This is on Linux. Running the built executable immediately produces a segfault. Here's a backtrace from gdb: (gdb) bt #0 0x005e78db in free_managed (s=s@entry=0x0) at ../../../racket/sgc/sgc.c:1535 #1 0x005e7f63 in GC_add_roots (start=start@entry=0x881b70 collects_path, end=end@entry=0x881b79 run_cmd+1) at ../../../racket/sgc/sgc.c:1657 #2 0x00576f9a in scheme_register_static (ptr=ptr@entry=0x881b70 collects_path, size=size@entry=8) at ../../../racket/src/salloc.c:720 #3 0x0045b87f in scheme_set_collects_path (p=p@entry=0x77f701a0) at ../../../racket/src/file.c:6841 #4 0x00436673 in run_from_cmd_line (mk_basic_env=optimized out, cont_run=0x435a10 cont_run, _argv=optimized out, argc=0) at ../../racket/cmdline.inc:1444 #5 main_after_stack (data=data@entry=0x7fffdc30) at ../../racket/main.c:450 #6 0x0057694d in do_main_stack_setup (data=0x7fffdc30, _main=0x435a80 main_after_stack, no_auto_statics=1) at ../../../racket/src/salloc.c:198 #7 scheme_main_stack_setup (no_auto_statics=no_auto_statics@entry=1, _main=_main@entry=0x435a80 main_after_stack, data=data@entry=0x7fffdc30) at ../../../racket/src/salloc.c:310 #8 0x004346ee in main_after_dlls (argv=optimized out, argc=optimized out) at ../../racket/main.c:381 #9 main (argc=optimized out, argv=optimized out) at ../../racket/main.c:341 Cheers, Asumu _ Racket Developers list: http://lists.racket-lang.org/dev _ Racket Developers list: http://lists.racket-lang.org/dev
Re: [racket-dev] SGC as default
How difficult would it be to allow the bootstrap process to use a preexisting Racket installation? This would alleviate some of the performance loss, for example in rebuilds by developers or in continuous integration. Sam On Aug 11, 2014 11:16 PM, Matthew Flatt mfl...@cs.utah.edu wrote: I've changed the Racket CGC implementation --- which is mostly used only to build the normal Racket variant --- to use SGC by default, instead of the Boehm GC. The intent of the switch is to make the more portable GC the default. If you have an existing build in a repo checkout, then `make` is likely to fail, because the makefile dependencies are not precise enough to deal with the switch. You can discard your old build directory, or it might work to simply delete builddir/racket/libmzgc.a If you're using CGC and want to continue using the Boehm GC, then provide `--disable-sgc` to `configure`. I've tuned SGC to bring its performance closer to the Boehm GC, but it's still slower. _ Racket Developers list: http://lists.racket-lang.org/dev _ Racket Developers list: http://lists.racket-lang.org/dev
Re: [racket-dev] SGC as default
I'm not sure how difficult it will be. It's tedious enough that the last time I thought about it, I just left a note next to no-cgc-needed in racket/src/racket/Makefile.in, but maybe it's worth pursuing now. At Tue, 12 Aug 2014 04:39:55 -0700, Sam Tobin-Hochstadt wrote: How difficult would it be to allow the bootstrap process to use a preexisting Racket installation? This would alleviate some of the performance loss, for example in rebuilds by developers or in continuous integration. Sam On Aug 11, 2014 11:16 PM, Matthew Flatt mfl...@cs.utah.edu wrote: I've changed the Racket CGC implementation --- which is mostly used only to build the normal Racket variant --- to use SGC by default, instead of the Boehm GC. The intent of the switch is to make the more portable GC the default. If you have an existing build in a repo checkout, then `make` is likely to fail, because the makefile dependencies are not precise enough to deal with the switch. You can discard your old build directory, or it might work to simply delete builddir/racket/libmzgc.a If you're using CGC and want to continue using the Boehm GC, then provide `--disable-sgc` to `configure`. I've tuned SGC to bring its performance closer to the Boehm GC, but it's still slower. _ Racket Developers list: http://lists.racket-lang.org/dev _ Racket Developers list: http://lists.racket-lang.org/dev
Re: [racket-dev] SGC as default
On 2014-08-12 05:16:21 +0100, Matthew Flatt wrote: If you have an existing build in a repo checkout, then `make` is likely to fail, because the makefile dependencies are not precise enough to deal with the switch. You can discard your old build directory, or it might work to simply delete builddir/racket/libmzgc.a The build didn't work for me in an existing checkout, so I tried making a fresh one and still got this failure: make xsrc/precomp.h make[7]: Entering directory '/home/asumu/plt/racket-fresh/racket/src/build/racket/gc2' env XFORM_PRECOMP=yes ../racketcgc -G /home/asumu/plt/racket-fresh/build/config -cqu ../../../racket/gc2/xform.rkt --setup . --cpp gcc -E -I./.. -I../../../racket/gc2/../include -pthread -DUSE_SENORA_GC --keep-lines -o xsrc/precomp.h ../../../racket/gc2/precomp.c Segmentation fault Makefile:202: recipe for target 'xsrc/precomp.h' failed make[7]: *** [xsrc/precomp.h] Error 139 Anything I should try to debug this? Cheers, Asumu _ Racket Developers list: http://lists.racket-lang.org/dev
Re: [racket-dev] SGC as default
At Tue, 12 Aug 2014 00:43:04 -0400, Asumu Takikawa wrote: On 2014-08-12 05:16:21 +0100, Matthew Flatt wrote: If you have an existing build in a repo checkout, then `make` is likely to fail, because the makefile dependencies are not precise enough to deal with the switch. You can discard your old build directory, or it might work to simply delete builddir/racket/libmzgc.a The build didn't work for me in an existing checkout, so I tried making a fresh one and still got this failure: make xsrc/precomp.h make[7]: Entering directory '/home/asumu/plt/racket-fresh/racket/src/build/racket/gc2' env XFORM_PRECOMP=yes ../racketcgc -G /home/asumu/plt/racket-fresh/build/config -cqu ../../../racket/gc2/xform.rkt --setup . --cpp gcc -E -I./.. -I../../../racket/gc2/../include -pthread -DUSE_SENORA_GC --keep-lines -o xsrc/precomp.h ../../../racket/gc2/precomp.c Segmentation fault Makefile:202: recipe for target 'xsrc/precomp.h' failed make[7]: *** [xsrc/precomp.h] Error 139 Anything I should try to debug this? What platform are you using? I imagine that running `./racketcgc` within the racket subdirectory of your build directory will similarly crash. Can you get any information from running `gdb racketcgc`? _ Racket Developers list: http://lists.racket-lang.org/dev
