[cfe-dev] --gcc-toolchain (GCC_INSTALL_PREFIX) and --prefix(-B)

Thu Mar 4 16:06:26 PST 2021

On Thu, Mar 4, 2021 at 3:24 PM Manoj Gupta <manojgupta at google.com> wrote:
>
> In the case of Chrome OS, we currently are using all 3 options:
> "-B", "--prefix" and "--gcc-toolchain". "--gcc-toolchain" is only used when the clang binary is not inside /usr/bin e.g. installed as /opt/llvm/bin/clang.
>
> I think the reason is the layout of gcc and binutils for cross compilation. Example for aarch64:
>
> binutils installation: /usr/libexec/gcc/aarch64-cros-linux-gnu  -> This is path for "--prefix" (--prefix='/usr/libexec/gcc/aarch64-cros-linux-gnu/aarch64-cros-linux-gnu-') and "-B" (-B'/usr/libexec/gcc/aarch64-cros-linux-gnu/')
> gcc binaries: /usr/x86_64-pc-linux-gnu/aarch64-cros-linux-gnu/gcc-bin/10.2.0/
> gcc libraries: /usr/lib/gcc/aarch64-cros-linux-gnu/10.2.0/
> libc headers + libs: /usr/aarch64-cros-linux-gnu/usr/{include,lib64}, /usr/aarch64-cros-linux-gnu/lib64
> sysroot: /build/device : Also has libc headers + libraries when building for the device, substitutes libc files from /usr/aarch64-cros-linux-gnu (--sysroot=/build/device).
>
> The final command line looks like:
> /opt/llvm/bin/clang --prefix='/usr/libexec/gcc/aarch64-cros-linux-gnu/aarch64-cros-linux-gnu-' -B'/usr/libexec/gcc/aarch64-cros-linux-gnu --gcc-toolchain='/usr' --sysroot=/build/device .
>
> My understanding is without passing "--gcc-toolchain=/usr", clang (when not installed in /usr/bin) is not able to find libgcc from /usr/lib/gcc/aarch64-cros-linux-gnu/10.2.0/. i.e. -B/--prefix options were insufficient for finding gcc library paths.
>
> Thanks,
> Manoj

--prefix is an alias for -B. I've added some documentation for
existing behaviors in https://reviews.llvm.org/D97902

For ChromeOS, I think -B/usr/libexec/gcc/aarch64-cros-linux-gnu may
not be needed. Having -B...aarch64-cros-linux-gnu-
(--prefix=.../aarch64-cros-linux-gnu-) is sufficient.

In Clang, we detect GCC installations under -B.
However, in GCC, -B is not used to detect installations. Executable
files and startup files are found via $prefix$file. $prefix/include is
added as an include directory.

--gcc-toolchain currently suppresses GCC detection under sysroot. It
probably makes sense to suppress detection under -B as well.
Unfortunately, some platforms may have relied on the GCC detection
behavior with -B (perhaps Android).

> On Thu, Mar 4, 2021 at 3:14 PM Fāng-ruì Sòng via cfe-dev <cfe-dev at lists.llvm.org> wrote:
>>
>> On Thu, Mar 4, 2021 at 2:31 PM Hubert Tong
>> <hubert.reinterpretcast at gmail.com> wrote:
>> >
>> > On Thu, Mar 4, 2021 at 1:54 PM Fāng-ruì Sòng <maskray at google.com> wrote:
>> >>
>> >> On 2021-03-04, Hubert Tong wrote:
>> >> >On Thu, Mar 4, 2021 at 3:19 AM Fāng-ruì Sòng via cfe-dev <
>> >> >cfe-dev at lists.llvm.org> wrote:
>> >> >
>> >> >> If --gcc-toolchain is specified, its value overrides the cmake variable
>> >> >> GCC_INSTALL_PREFIX.
>> >> >> When the value is non-empty: the value is appended to the --prefix list
>> >> >> and is used to detect GCC installations.
>> >> >> The GCC installation is used to provide include directories/library
>> >> >> directories and some startup files (e.g. crtbegin).
>> >> >>
>> >> >> Problem 1.
>> >> >>
>> >> >> --prefix(-B) does more than --gcc-toolchain:
>> >> >> clang::driver::Driver::GetProgramPath basically searches for
>> >> >> $prefix/$triple-$file and $prefix$file,
>> >> >> where $prefix is taken from the list of --prefix(-B). --gcc-toolchain does
>> >> >> not participate in the search. <have attached a summary of the algorithm at
>> >> >> the bottom>
>> >> >> The result is that 'ld' and 'as' may come from the system (more precisely,
>> >> >> sysroot):
>> >> >>
>> >> >> cd clang/test/Driver
>> >> >> # Make sure Inputs/opensuse_42.2_aarch64_tree/usr/bin/ld exists.
>> >> >> clang -target aarch64-suse-linux
>> >> >> --gcc-toolchain=Inputs/opensuse_42.2_aarch64_tree/usr '-###'
>> >> >> gcc-toolchain.cpp -v
>> >> >>
>> >> >>    # I have ld in my /usr/local/bin and it takes precedence over
>> >> >> /usr/bin/ld
>> >> >>  "/usr/local/bin/ld" "-EL" "--eh-frame-hdr" "-m" "aarch64linux"
>> >> >> "-dynamic-linker" "/lib/ld-linux-aarch64.so.1" "-o" "a.out"
>> >> >> "Inputs/opensuse_42.2_aarch64_tree/usr/lib64/gcc/aarch64-suse-linux/4.8/../../../../lib64/crt1.o"
>> >> >> "Inputs/opensuse_42.2_aarch64_tree/usr/lib64/gcc/aarch64-suse-linux/4.8/../../../../lib64/crti.o"
>> >> >> "Inputs/opensuse_42.2_aarch64_tree/usr/lib64/gcc/aarch64-suse-linux/4.8/crtbegin.o"
>> >> >> "-LInputs/opensuse_42.2_aarch64_tree/usr/lib64/gcc/aarch64-suse-linux/4.8"
>> >> >> "-LInputs/opensuse_42.2_aarch64_tree/usr/lib64/gcc/aarch64-suse-linux/4.8/../../../../lib64"
>> >> >> "-L/lib/aarch64-linux-gnu" "-L/lib/../lib64" "-L/usr/lib/aarch64-linux-gnu"
>> >> >> "-L/usr/lib/../lib64"
>> >> >> "-LInputs/opensuse_42.2_aarch64_tree/usr/lib64/gcc/aarch64-suse-linux/4.8/../../.."
>> >> >> "-L/tmp/RelA/bin/../lib" "-L/lib" "-L/usr/lib"
>> >> >> "/tmp/gcc-toolchain-f87f08.o" "-lgcc" "--as-needed" "-lgcc_s"
>> >> >> "--no-as-needed" "-lc" "-lgcc" "--as-needed" "-lgcc_s" "--no-as-needed"
>> >> >> "Inputs/opensuse_42.2_aarch64_tree/usr/lib64/gcc/aarch64-suse-linux/4.8/crtend.o"
>> >> >> "Inputs/opensuse_42.2_aarch64_tree/usr/lib64/gcc/aarch64-suse-linux/4.8/../../../../lib64/crtn.o"
>> >> >>
>> >> >> The -L and crt* files are indeed from
>> >> >> Inputs/opensuse_42.2_aarch64_tree/usr, but ld (and as if
>> >> >> -fno-integrated-as) is from the system.
>> >> >> On many Linux distributions you can normally assume that the system ld and
>> >> >> as only support the host architecture.
>> >> >> This means --gcc-toolchain can only be used to specify a GCC installation
>> >> >> with the same architecture.
>> >> >>
>> >> >... and specifying a GCC installation with the same architecture (e.g., to
>> >> >use a newer libstdc++) is something that people do:
>> >> >$ clang++ -fuse-ld=ld --gcc-toolchain=/opt/rh/devtoolset-7/root -xc -
>> >> >-fno-integrated-as <<<'#include <stdio.h>'$'\n''int main(void) {
>> >> >printf("Hello, world!\n"); }' && echo OKAY
>> >> >OKAY
>> >> >... but it does seem ld is coming from --gcc-toolchain for me anyway:
>> >> >$ ./build/bootstrap/stage1/build/bin/clang++ -fuse-ld=ld
>> >> >--gcc-toolchain=/opt/rh/devtoolset-7/root -xc - -fno-integrated-as
>> >> ><<<'#include <stdio.h>'$'\n''int main(void) { printf("Hello, world!\n"); }'
>> >> >-### 2>&1 | grep /ld
>> >> > "/opt/rh/devtoolset-7/root/lib/gcc/ppc64le-redhat-linux/7/../../../../bin/ld"
>> >> >"--hash-style=gnu" "--eh-frame-hdr" "-m" "elf64lppc" "-dynamic-linker"
>> >> >"/lib64/ld64.so.2" "-o" "a.out" "/lib/../lib64/crt1.o"
>> >> >"/lib/../lib64/crti.o"
>> >> >"/opt/rh/devtoolset-7/root/lib/gcc/ppc64le-redhat-linux/7/crtbegin.o"
>> >> >"-L/opt/rh/devtoolset-7/root/lib/gcc/ppc64le-redhat-linux/7"
>> >> >"-L/opt/rh/devtoolset-7/root/lib/gcc/ppc64le-redhat-linux/7/../../../../lib64"
>> >> >"-L/lib/powerpc64le-linux-gnu" "-L/lib/../lib64"
>> >> >"-L/usr/lib/powerpc64le-linux-gnu" "-L/usr/lib/../lib64"
>> >> >"-L/opt/rh/devtoolset-7/root/lib/gcc/ppc64le-redhat-linux/7/../../.."
>> >> >"-L/data1/hstong.local/wybuild/build/bootstrap/stage1/build/bin/../lib"
>> >> >"-L/lib" "-L/usr/lib" "/tmp/--8f8c6a.o" "-lstdc++" "-lm" "-lgcc_s" "-lgcc"
>> >> >"-lc" "-lgcc_s" "-lgcc"
>> >> >"/opt/rh/devtoolset-7/root/lib/gcc/ppc64le-redhat-linux/7/crtend.o"
>> >> >"/lib/../lib64/crtn.o"
>> >>
>> >> 'ld' path is computed by clang/lib/Driver/Driver.cpp Driver::GetProgramPath.
>> >> Did you set PATH to the devtoolset-7 bin directory?
>> >
>> > No, I did not.
>>
>> Do you mind debugging clang '-###' a bit: set a breakpoint on
>> clang/lib/Driver/Driver.cpp `GetProgramPath` and report how the 'ld'
>> under --gcc-toolchain is selected?
>>
>> >>
>> >>
>> >> If --gcc-toolchain and multiple -B are specified, the search order is:
>> >>
>> >> -B -B -B --gcc-toolchain
>> >>
>> >> Currently the GCC installation with the largest version wins.
>> >> Do you think we should stop search after one -B provides a GCC installation?
>> >
>> > Well, the ordering where --gcc-toolchain follows all of the -B seems okay in the context of looking for ld or as. For changing "largest version wins", it might make some sense to order --gcc-toolchain first for the resolution of the GCC installation (which allows "largest version wins" to continue with -B). I suppose it means people using builds preconfigured with a GCC install prefix will be somewhat inconvenienced when they don't want to use the preconfigured path; however, I think this retains most functionality (there is a way to limit the search for GCC installations but also a way to compose search paths for the latest GCC version).
>>
>> If cmake GCC_INSTALL_PREFIX was specified configuring llvm-project, I
>> think it makes sense that explicit options (-B, --gcc-toolchain) take
>> precedence over GCC_INSTALL_PREFIX.
>> When --gcc-toolchain is specified: yes, probably it should take
>> precedence over --prefix. I wonder whether it should suppress --prefix
>> for GCC installation.
>>
>> I played with gcc -B a bit: looks like it picks the first --prefix
>> where files can be found. So I am going to send a patch making the
>> "largest version wins" local to one -B (i.e. if multiple versions are
>> detected, pick the largest).
>>
>>
>> >>
>> >>
>> >> >> --prefix can make as and ld paths correct, but: if another --prefix is
>> >> >> needed, why do we use --gcc-toolchain?
>> >> >>
>> >> >Well, I'm not sure that another --prefix is needed. Also, --prefix doesn't
>> >> >help to pick up the GCC header and library paths and --sysroot overrides
>> >> >too much:
>> >> >$ clang++ -fuse-ld=ld --sysroot=/opt/rh/devtoolset-7/root -xc -
>> >> >-fno-integrated-as <<<'#include <stdio.h>'$'\n''int main(void) {
>> >> >printf("Hello, world!\n"); }'
>> >> ><stdin>:1:10: fatal error: 'stdio.h' file not found
>> >> >#include <stdio.h>
>> >> >         ^~~~~~~~~
>> >> >1 error generated.
>> >>
>> >> (The example used --sysroot instead of --prefix)
>> >
>> > Yes.
>> >
>> >>
>> >>
>> >> OK, your other reply clarified that --prefix does pick up include/library directories.
>> >
>> > Yes.
>>
>>
>>
>> --
>> 宋方睿
>> _______________________________________________
>> cfe-dev mailing list
>> cfe-dev at lists.llvm.org
>> https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-dev

-- 
宋方睿