All the mail mirrored from lore.kernel.org
 help / color / mirror / Atom feed

• * Re: Building an existing .config (provided for gcc) with clang?
         [not found] <26f81a61-e2eb-47e9-b6e0-3989582bb955n@googlegroups.com>
  @ 2021-06-02 18:31 ` Nick Desaulniers
    2021-06-02 18:48   ` Arnd Bergmann
    0 siblings, 1 reply; 2+ messages in thread
  From: Nick Desaulniers @ 2021-06-02 18:31 UTC (permalink / raw
    To: Mathieu Acher
    Cc: Clang Built Linux, Masahiro Yamada, Arnd Bergmann,
  	Linux Kbuild mailing list
  
  On Wed, Jun 2, 2021 at 3:13 AM Mathieu Acher <mathieu.acher@gmail.com> wrote:
  >
  > Hi,
  >
  > tldr; some "fixes" seem needed to build an arbitrary .config with gcc *and* clang
  >
  > I'm investigating a build scenario where I already have a .config typically generated with randconfig and originally set up for building with gcc...
  > And I want to use clang instead to build it.
  > My experience so far is that the --syncconfig Kconfig asks me to fix some options' values
  > first, to choose between INIT_STACK_NONE and INIT_STACK_ALL... and then to choose KCSAN: dynamic data race detector (KCSAN)  or not...
  
  Hi Professor,
  It's a good question, one I don't really have a satisfactory answer for.
  
  I think it's impossible (but would love to be corrected) to reuse
  exactly the same .config with two different compilers (even different
  versions of the same compiler can have differences).  This is because
  KConfig also encodes not just various kernel drivers and features, but
  whether the toolchain supports certain features.  So you can't force
  one compiler to claim to support some compiler flag, then actually use
  it.  If at Kconfig time we detect the compiler has support for say a
  -fsanitize= flag, we encode that in the .config, then during the
  actual build just look that up in the .config and use the -fsanitze=
  flag.
  
  Probably the closest to what you're trying to do is the `olddefconfig`
  make target.
  
  >
  > Two examples below, on Linux kernel 5.8.
  >
  > First, with tinyconfig
  >
  > make CC=gcc tinyconfig
  > gcc --version
  > gcc (Debian 10.2.1-6) 10.2.1 20210110
  >
  > make CC=clang-11
  > scripts/kconfig/conf  --syncconfig Kconfig
  > *
  > * Restart config...
  > *
  > *
  > * Memory initialization
  > *
  > Initialize kernel stack variables at function entry
  > > 1. no automatic initialization (weakest) (INIT_STACK_NONE)
  >   2. 0xAA-init everything on the stack (strongest) (INIT_STACK_ALL) (NEW)
  > choice[1-2?]: 2
  > Enable heap memory zeroing on allocation by default (INIT_ON_ALLOC_DEFAULT_ON) [N/y/?] n
  > Enable heap memory zeroing on free by default (INIT_ON_FREE_DEFAULT_ON) [N/y/?] n
  >
  > and then the build proceeds...
  >
  > Maybe it's my setup/build environment, but I got errors:
  > CC      arch/x86/events/amd/core.o
  > In file included from arch/x86/events/amd/core.c:12:
  > arch/x86/events/amd/../perf_event.h:854:21: error: invalid output size for constraint '=q'
  >         u64 disable_mask = __this_cpu_read(cpu_hw_events.perf_ctr_virt_mask);
  >                            ^
  > ./include/linux/percpu-defs.h:446:2: note: expanded from macro '__this_cpu_read'
  >         raw_cpu_read(pcp);                                              \
  >         ^
  > ./include/linux/percpu-defs.h:420:28: note: expanded from macro 'raw_cpu_read'
  > #define raw_cpu_read(pcp)               __pcpu_size_call_return(raw_cpu_read_, pcp)
  >                                         ^
  > ./include/linux/percpu-defs.h:321:23: note: expanded from macro '__pcpu_size_call_return'
  >         case 1: pscr_ret__ = stem##1(variable); break;                  \
  >                              ^
  > <scratch space>:169:1: note: expanded from here
  > raw_cpu_read_1
  > ^
  > ./arch/x86/include/asm/percpu.h:394:30: note: expanded from macro 'raw_cpu_read_1'
  > #define raw_cpu_read_1(pcp)             percpu_from_op(, "mov", pcp)
  >                                         ^
  > ./arch/x86/include/asm/percpu.h:189:15: note: expanded from macro 'percpu_from_op'
  >                     : "=q" (pfo_ret__)                  \
  >                             ^
  > In file included from arch/x86/events/amd/core.c:12:
  > arch/x86/events/amd/../perf_event.h:854:21: error: invalid output size for constraint '=q'
  > ./include/linux/percpu-defs.h:446:2: note: expanded from macro '__this_cpu_read'
  >         raw_cpu_read(pcp);                                              \
  >         ^
  > ./include/linux/percpu-defs.h:420:28: note: expanded from macro 'raw_cpu_read'
  > #define raw_cpu_read(pcp)               __pcpu_size_call_return(raw_cpu_read_, pcp)
  >                                         ^
  > ./include/linux/percpu-defs.h:322:23: note: expanded from macro '__pcpu_size_call_return'
  >         case 2: pscr_ret__ = stem##2(variable); break;                  \
  >                              ^
  > <scratch space>:174:1: note: expanded from here
  > raw_cpu_read_2
  > ^
  > ./arch/x86/include/asm/percpu.h:395:30: note: expanded from macro 'raw_cpu_read_2'
  > #define raw_cpu_read_2(pcp)             percpu_from_op(, "mov", pcp)
  >                                         ^
  > ./arch/x86/include/asm/percpu.h:189:15: note: expanded from macro 'percpu_from_op'
  >                     : "=q" (pfo_ret__)                  \
  >                             ^
  > In file included from arch/x86/events/amd/core.c:12:
  > arch/x86/events/amd/../perf_event.h:854:21: error: invalid output size for constraint '=q'
  > ./include/linux/percpu-defs.h:446:2: note: expanded from macro '__this_cpu_read'
  >         raw_cpu_read(pcp);                                              \
  >         ^
  > ./include/linux/percpu-defs.h:420:28: note: expanded from macro 'raw_cpu_read'
  > #define raw_cpu_read(pcp)               __pcpu_size_call_return(raw_cpu_read_, pcp)
  >                                         ^
  > ./include/linux/percpu-defs.h:323:23: note: expanded from macro '__pcpu_size_call_return'
  >         case 4: pscr_ret__ = stem##4(variable); break;                  \
  >                              ^
  > <scratch space>:179:1: note: expanded from here
  > raw_cpu_read_4
  > ^
  > ./arch/x86/include/asm/percpu.h:396:30: note: expanded from macro 'raw_cpu_read_4'
  > #define raw_cpu_read_4(pcp)             percpu_from_op(, "mov", pcp)
  >                                         ^
  > ./arch/x86/include/asm/percpu.h:189:15: note: expanded from macro 'percpu_from_op'
  >                     : "=q" (pfo_ret__)                  \
  >                             ^
  > 3 errors generated.
  > make[3]: *** [scripts/Makefile.build:281: arch/x86/events/amd/core.o] Error 1
  > make[2]: *** [scripts/Makefile.build:497: arch/x86/events/amd] Error 2
  > make[1]: *** [scripts/Makefile.build:497: arch/x86/events] Error 2
  > make: *** [Makefile:1756: arch/x86] Error 2
  >
  > The difference between the .config (.config with gcc vs manual "fix" of .config with clang) is as follows:
  > scripts/diffconfig .config .config.old
  > -CC_HAS_ASM_GOTO_OUTPUT y
  > -CC_HAS_AUTO_VAR_INIT y
  > -CC_HAS_KASAN_SW_TAGS y
  > -CC_IS_CLANG y
  > -HAVE_KCSAN_COMPILER y
  > -INIT_STACK_ALL y
  >  CC_VERSION_TEXT "Debian clang version 11.0.1-2" -> "gcc (Debian 10.2.1-6) 10.2.1 20210110"
  >  CLANG_VERSION 110001 -> 0
  >  GCC_VERSION 0 -> 100201
  >  INIT_STACK_NONE n -> y
  >  KASAN_STACK 0 -> 1
  > +CC_IS_GCC y
  >
  >
  > Interestingly, I've then tried:
  > make CC=gcc tinyconfig
  > make CC=clang-11 tinyconfig
  > scripts/diffconfig .config .config.old
  > -X86_PAE n
  > +ARCH_ENABLE_MEMORY_HOTPLUG y
  > +CRASH_DUMP n
  > +DEBUG_HIGHMEM n
  > +FRAME_POINTER y
  > +HARDENED_USERCOPY n
  > +HAVE_ALIGNED_STRUCT_PAGE y
  > +HAVE_HARDENED_USERCOPY_ALLOCATOR y
  > +HIGHMEM y
  > +HIGHPTE n
  > +SLAB_FREELIST_HARDENED n
  > +SLAB_FREELIST_RANDOM n
  > make
  >
  > and it builds nicely!
  > It should be noted that the "diff" between .config is different.
  >
  > ==
  >
  > Second example, with a random configuration...
  > (I can upload the .config if needs be)
  >
  > make randconfig # with gcc
  > make CC=clang-11 # build was OK this time, manual fix seems effective
  > scripts/diffconfig .config .config.old
  > -CC_HAS_ASM_GOTO_OUTPUT y
  > -CC_HAS_AUTO_VAR_INIT y
  > -CC_HAS_KASAN_SW_TAGS y
  > -CC_IS_CLANG y
  > -HAVE_KCSAN_COMPILER y
  > -INIT_STACK_ALL y
  > -KCSAN n
  >  CC_VERSION_TEXT "Debian clang version 11.0.1-2" -> "gcc (Debian 10.2.1-6) 10.2.1 20210110"
  >  CLANG_VERSION 110001 -> 0
  >  GCC_VERSION 0 -> 100201
  >  INIT_STACK_NONE n -> y
  >  KASAN_STACK 0 -> 1
  >  LD_VERSION 235020000 -> 235010000
  > +CC_IS_GCC y
  >
  > ===
  >
  > So my questions...
  >
  > Can we envision to have a generic procedure that would "fix" a .config (gcc oriented) in such a way clang can be used?
  > if it's "just" taking a .config and fixing INIT_STACK and KCSAN, I'm fine :)
  
  I think if you run `make CC=clang olddefconfig` that will reuse as
  much as possible from the previous .config, no matter which toolchain
  it was built with.
  
  >
  > It seems that some options are supported only by clang (and vice versa)...
  > It could be useful for the fixing procedure...
  > is it possible to list all such options?
  >
  > More generally, do you have some experiences of building an arbitrary .config with gcc and then clang?
  >
  > The goal is to investigate how clang may differ from gcc given a configuration (partly discussed at Plumbers https://youtu.be/FFjV9f_Ub9o?t=10761)
  >
  > Thanks for your help!
  >
  > Mathieu Acher
  >
  > Associate Professor
  > Univ Rennes, Inria, CNRS, IRISA, France (DiverSE team)
  >
  > --
  > You received this message because you are subscribed to the Google Groups "Clang Built Linux" group.
  > To unsubscribe from this group and stop receiving emails from it, send an email to clang-built-linux+unsubscribe@googlegroups.com.
  > To view this discussion on the web visit https://groups.google.com/d/msgid/clang-built-linux/26f81a61-e2eb-47e9-b6e0-3989582bb955n%40googlegroups.com.
  
  
  
  -- 
  Thanks,
  ~Nick Desaulniers
  
  ^ permalink raw reply	[flat|nested] 2+ messages in thread

• • * Re: Building an existing .config (provided for gcc) with clang?
      2021-06-02 18:31 ` Building an existing .config (provided for gcc) with clang? Nick Desaulniers
    @ 2021-06-02 18:48   ` Arnd Bergmann
      0 siblings, 0 replies; 2+ messages in thread
    From: Arnd Bergmann @ 2021-06-02 18:48 UTC (permalink / raw
      To: Nick Desaulniers
      Cc: Mathieu Acher, Clang Built Linux, Masahiro Yamada,
    	Linux Kbuild mailing list
    
    On Wed, Jun 2, 2021 at 8:31 PM 'Nick Desaulniers' via Clang Built
    Linux <clang-built-linux@googlegroups.com> wrote:
    > On Wed, Jun 2, 2021 at 3:13 AM Mathieu Acher <mathieu.acher@gmail.com> wrote:
    > >
    > > tldr; some "fixes" seem needed to build an arbitrary .config with gcc *and* clang
    > >
    > > I'm investigating a build scenario where I already have a .config typically generated with randconfig and originally set up for building with gcc...
    > > And I want to use clang instead to build it.
    > > My experience so far is that the --syncconfig Kconfig asks me to fix some options' values
    > > first, to choose between INIT_STACK_NONE and INIT_STACK_ALL... and then to choose KCSAN: dynamic data race detector (KCSAN)  or not...
    >
    > Hi Professor,
    > It's a good question, one I don't really have a satisfactory answer for.
    >
    > I think it's impossible (but would love to be corrected) to reuse
    > exactly the same .config with two different compilers (even different
    > versions of the same compiler can have differences).  This is because
    > KConfig also encodes not just various kernel drivers and features, but
    > whether the toolchain supports certain features.  So you can't force
    > one compiler to claim to support some compiler flag, then actually use
    > it.  If at Kconfig time we detect the compiler has support for say a
    > -fsanitize= flag, we encode that in the .config, then during the
    > actual build just look that up in the .config and use the -fsanitze=
    > flag.
    >
    > Probably the closest to what you're trying to do is the `olddefconfig`
    > make target.
    
    Agreed. In particular, there is no guarantee of ABI compatibility between
    kernel modules that are built in a different environment, whether that is
    a different compiler (version) or different configuration.
    
    For the purpose of build testing, it is better to run the same commands for
    configuring the kernel (defconfig, tinyconfig, allmodconfig, randconfig, ...)
    rather than relying on an imported .config file.
    
    > > Two examples below, on Linux kernel 5.8.
    > >
    > > First, with tinyconfig
    > >
    > > make CC=gcc tinyconfig
    > > gcc --version
    > > gcc (Debian 10.2.1-6) 10.2.1 20210110
    > >
    > > make CC=clang-11
    > > scripts/kconfig/conf  --syncconfig Kconfig
    
    side note: when build testing, I would recommend using the '-s' flag to make
    in order to have more readable output (no output on success), and using
    separate object directories per target/compiler/config tuple.
    
    > > * Restart config...
    > > *
    > > *
    > > * Memory initialization
    > > *
    > > Initialize kernel stack variables at function entry
    > > > 1. no automatic initialization (weakest) (INIT_STACK_NONE)
    > >   2. 0xAA-init everything on the stack (strongest) (INIT_STACK_ALL) (NEW)
    > > choice[1-2?]: 2
    > > Enable heap memory zeroing on allocation by default (INIT_ON_ALLOC_DEFAULT_ON) [N/y/?] n
    > > Enable heap memory zeroing on free by default (INIT_ON_FREE_DEFAULT_ON) [N/y/?] n
    > >
    > > and then the build proceeds...
    > >
    > > Maybe it's my setup/build environment, but I got errors:
    > > CC      arch/x86/events/amd/core.o
    > > In file included from arch/x86/events/amd/core.c:12:
    > > arch/x86/events/amd/../perf_event.h:854:21: error: invalid output size for constraint '=q'
    > >         u64 disable_mask = __this_cpu_read(cpu_hw_events.perf_ctr_virt_mask);
    > >                            ^
    > > ./include/linux/percpu-defs.h:446:2: note: expanded from macro '__this_cpu_read'
    > >         raw_cpu_read(pcp);                                              \
    > >         ^
    > > ./include/linux/percpu-defs.h:420:28: note: expanded from macro 'raw_cpu_read'
    > > #define raw_cpu_read(pcp)               __pcpu_size_call_return(raw_cpu_read_, pcp)
    > >                                         ^
    > > ./include/linux/percpu-defs.h:321:23: note: expanded from macro '__pcpu_size_call_return'
    > >         case 1: pscr_ret__ = stem##1(variable); break;                  \
    > >                              ^
    
    This is a known bug that has since been fixed.
    
    > > Can we envision to have a generic procedure that would "fix" a .config (gcc oriented) in such a way clang can be used?
    > > if it's "just" taking a .config and fixing INIT_STACK and KCSAN, I'm fine :)
    >
    > I think if you run `make CC=clang olddefconfig` that will reuse as
    > much as possible from the previous .config, no matter which toolchain
    > it was built with.
    
    Yes, this should always result in a clean kernel build, if it does
    not, that is a bug
    we have to fix.
    
    However, this does not necessarily result in the intended kernel: a 'tinyconfig'
    may be larger, and an 'allmodconfig' or 'randconfig' may produce less build
    coverage because it builds only the common subset of the available options.
    
    
    > -CC_HAS_ASM_GOTO_OUTPUT y
    
    This was added in gcc-11
    
    > -CC_HAS_AUTO_VAR_INIT y
    > -CC_HAS_KASAN_SW_TAGS y
    > -INIT_STACK_ALL y
    >  INIT_STACK_NONE n -> y
    > -HAVE_KCSAN_COMPILER y
    > -KCSAN n
    These depend on the compiler version as well as the target architecture
    
    > -CC_IS_CLANG y
    > +CC_IS_GCC y
    >  CC_VERSION_TEXT "Debian clang version 11.0.1-2" -> "gcc (Debian 10.2.1-6) 10.2.1 20210110"
    >  CLANG_VERSION 110001 -> 0
    >  GCC_VERSION 0 -> 100201
    
    These are obviously unavoidable
    
    >  LD_VERSION 235020000 -> 235010000
    
    This one could be avoided by using the same binutils
    
    >  KASAN_STACK 0 -> 1
    
    This is a workaround for known problems in some compilers
    
           Arnd
    
    ^ permalink raw reply	[flat|nested] 2+ messages in thread

end of thread, other threads:[~2021-06-02 18:50 UTC | newest]

Thread overview: 2+ messages (download: mbox.gz follow: Atom feed
-- links below jump to the message on this page --
     [not found] <26f81a61-e2eb-47e9-b6e0-3989582bb955n@googlegroups.com>
2021-06-02 18:31 ` Building an existing .config (provided for gcc) with clang? Nick Desaulniers
2021-06-02 18:48   ` Arnd Bergmann

This is an external index of several public inboxes,
see mirroring instructions on how to clone and mirror
all data and code used by this external index.