On 2/19/19 1:32 AM, Xi Ruoyao wrote:
> On 2019-02-19 00:58 -0500, Dennis Clarke wrote:
>>
>> After many many many experiments I am seeing that host and target data
>> is not being passed down into the build of gmp after configure. I took
>> a look into the gmp sources and multiple Makefile.in and see :
>>
>> host = @host@
>> host_alias = @host_alias@
>> host_cpu = @host_cpu@
>> host_os = @host_os@
>> host_vendor = @host_vendor@
>>
>> Well what should those really be?
>>
>> I am trying to force this :
>>
>> vesta_$ grep 'triplet' ./gmp-6.1.2/Makefile.in
>> build_triplet = x86_64-unknown-freebsd12.0
>> host_triplet = riscv64-unknown-freebsd12.0
>> vesta_$
>>
>> No amount of finangle of configure options seems to be able to tell the
>> gmp build that the local host is x86_64-unknown-freebsd12.0 and that the
>> target should be a RISC-V rv64imafdc architecture. The actual triplet
>> for the target should be riscv64-unknown-freebsd12.0 however what I am
>> actually seeing is :
>>
>>
>> configure: summary of build options:
>>
>> Version: GNU MP 6.1.2
>> Host type: none-unknown-freebsd12.0
>> ABI: standard
>> Install prefix: /usr/local/gcc8
>> Compiler: riscv64-unknown-freebsd12.0-gcc
>> Static libraries: yes
>> Shared libraries: no
>>
>>
>> That is wrong.
>
> I think that is correct. GMP uses some machine-dependent code to run faster. A
> "none-*-*" triplet just tell GMP to use portable code, instead of any machine-
> dependent code.
OKay this is good input. I simply am confused as to why the "none" is
there and I guess you explain it well.
>
> Please note that GMP's triplet is not GNU triplet. For example GMP said my
> machine is "ivybridge-pc-linux-gnu", but its GNU triplet should be "x86_64-pc-
> linux-gnu" or "i686-pc-linux-gnu".
>
> I configured a GMP with --host=none-linux-gnu on my machine and it built fine.
>
> And it seems "none" is hard-coded by GCC building system (deliberately) to make
> GCC executables portable. See makefile.in:12464 in gcc-8.2.0.
>
>> What I should see is this :
>>
>> configure: summary of build options:
>>
>> Version: GNU MP 6.1.2
>> Host type: riscv64-unknown-freebsd12.0
>> ABI: standard
>> Install prefix: /usr/local/gcc8
>> Compiler: /opt/tools/bin/riscv64-unknown-freebsd12.0-gcc
>> Static libraries: yes
>> Shared libraries: yes
>
> That's useless. GMP doesn't know what <del>the hell</del> is a RISCV so it will
> still use portable code only.
>
OKay!
Good to know.
So the real question then is why the mpfr stage fails with a complaint
about the ABI or "machine type" of libgmp? The actual archive has been
completed in the gmp stage and now mpfr will need that as a dependency
but the process breaks down and fails with :
checking for gmp.h... yes
checking whether gcc __attribute__ ((mode (XX))) works... yes
checking for recent GMP... yes
checking usable gmp.h at link time... yes
checking for GMP_NUMB_BITS and sizeof(mp_limb_t) consistency... cannot test
checking for __gmpz_init in -lgmp... no
configure: error: libgmp not found or uses a different ABI (including
static vs shared).
Please read the INSTALL file -- see "In case of problem".
gmake[1]: *** [Makefile:5139: configure-mpfr] Error 1
gmake[1]: Leaving directory
'/usr/home/dclarke/rv64g_local/build/gcc-8.2.0_rv64imafdc.007'
gmake: *** [Makefile:906: all] Error 2
vesta_$
All of this is fully laid out in :
https://gcc.gnu.org/ml/gcc-help/2019-02/msg00068.html
So I am quite stuck at this point and wondering how to build a native
RISC-V compiler for a native RISC-V target system? There must be some
configure "magic" but I have yet to discover it.
Dennis
