crosstools in buildroot build fails when hard fpu is selected

crosstools in buildroot build fails when hard fpu is selected

[Thierry Bultel]

Hello,

Using Buildroot-2012-08, I am building for arm-cortexa9-eabi,
crosstools+glibc, kernel base 2.6.38.8

I want to activate hardware floating point support, thus I have
unchecked the  "Use software floating point by default"

My final aim is to use Neon, but the defaut Hardware FPU is fine at this
time.

The build (from scratch) fails when at glibc:

[INFO ]  Installing C library headers & start files
[EXTRA]    Configuring C library
[CFG  ]    configure: loading cache
/bigdisk/shared/Buildroot/output.imx6hardfpu/build/build-toolchain/.build/arm-unknown-linux-gnueabi/build/build-libc-startfiles/config.cache
[CFG  ]    checking build system type... i686-build_pc-linux-gnu
[CFG  ]    checking host system type... arm-unknown-linux-gnueabi
[CFG  ]    checking for arm-unknown-linux-gnueabi-gcc...
arm-unknown-linux-gnueabi-gcc
[ERROR]    checking for suffix of object files... configure: error: in
`/bigdisk/shared/Buildroot/output.imx6hardfpu/build/build-toolchain/.build/arm-unknown-linux-gnueabi/build/build-libc-startfiles':
[ERROR]    configure: error: cannot compute suffix of object files:
cannot compile


The config.log shows:

configure:2477: arm-unknown-linux-gnueabi-gcc     -c  -U_FORTIFY_SOURCE
 -mlittle-endian -march=armv7-a   -mtune=cortex-a9 -mfpu=neon
-mhard-float  -O2   conftest.c >&5
conftest.c:1: sorry, unimplemented: -mfloat-abi=hard and VFP


GCC has not been generated with the harware FPU support.

I have attempted several hacks into
toolchain/toolchain-crosstool-ng/crosstool-ng.config-glibc

... with no better success

Any ideas ?

Thanks and cheers
Thierry



--
For unsubscribe information see http://sourceware.org/lists.html#faq

Re: crosstools in buildroot build fails when hard fpu is selected

[Yann E. MORIN]

Thierry, All,

On Wednesday 21 November 2012 Thierry Bultel wrote:
> Using Buildroot-2012-08, I am building for arm-cortexa9-eabi,
> crosstools+glibc, kernel base 2.6.38.8

'crosstool-NG', not 'crosstools' ;-)

> I want to activate hardware floating point support, thus I have
> unchecked the  "Use software floating point by default"
> 
> My final aim is to use Neon, but the defaut Hardware FPU is fine at this
> time.
> 
> The build (from scratch) fails when at glibc:
[--SNIP--]
> The config.log shows:
> 
> configure:2477: arm-unknown-linux-gnueabi-gcc     -c  -U_FORTIFY_SOURCE
>  -mlittle-endian -march=armv7-a   -mtune=cortex-a9 -mfpu=neon
> -mhard-float  -O2   conftest.c >&5
> conftest.c:1: sorry, unimplemented: -mfloat-abi=hard and VFP
> 
> GCC has not been generated with the harware FPU support.

No, gcc has indeed be configured with hard-float. It's just that gcc does
*not* support emitting VFP instructions for the hard-float ABI. See:
  gcc-X.Y.Z/gcc/config/arm/arm.c
  1850        if (arm_float_abi == ARM_FLOAT_ABI_HARD && TARGET_VFP)
  1851          sorry ("-mfloat-abi=hard and VFP");

So, no matter what, it is not possible, with stock gcc, to use the VFP.

However, it is possible to use neon. There is a sample in crosstool-NG
that builds a NEON-enabled toolchain. You can look at how it is configured
and replicate that in your setup:
  ct-ng list-samples
  ct-ng arm-cortex_a15-linux-gnueabi    <- this sample enables neon
  ct-ng menuconfig

Then, to reproduce those options with buildroot:
    make ctng-menuconfig

Also, you could use 'softfp' instead of 'hard'. The name is (IMHO) a
msi-nomer, as this ABI does emit hardware floating point instructions,
but uses the software ABI (the way floats are represented). Look at
the help entries for 'soft', 'hard' and 'softfp' for more info.

Note: Unfortunately, the ct-ng integration in buildroot does not allow
building a 'softfp' toolchain; it's either 'soft' or 'hard', because BR
does not differentiate between 'hard' and 'softfp'.

> I have attempted several hacks into
> toolchain/toolchain-crosstool-ng/crosstool-ng.config-glibc
> ... with no better success

The problem is *not* in glibc, so you can't fix anything by hacking the
glibc build script.

Regards,
Yann E. MORIN.

-- 
.-----------------.--------------------.------------------.--------------------.
|  Yann E. MORIN  | Real-Time Embedded | /"\ ASCII RIBBON | Erics' conspiracy: |
| +33 662 376 056 | Software  Designer | \ / CAMPAIGN     |  ___               |
| +33 223 225 172 `------------.-------:  X  AGAINST      |  \e/  There is no  |
| http://ymorin.is-a-geek.org/ | _/*\_ | / \ HTML MAIL    |   v   conspiracy.  |
'------------------------------^-------^------------------^--------------------'

--
For unsubscribe information see http://sourceware.org/lists.html#faq

Re: [Buildroot] crosstools build fails when hard fpu is selected

[Thierry Bultel]

Thanks Thomas,

thanks to Yann's advice,
I have been able to generate an ARM gcc with Neon support,
and performed some performance testing on my target.
( the root filesystem is for now the one with soft float).

But this is an external crosstool-ng (1.17.0) build.

Thus I have attempted to configure Buildroot to use this
external toolchain.

The build fails at ncurses (see below)

I thought that is was due to the fact that the "Use soft float by 
default" option was not available when using an external toolchain (why 
?) but I have the same issue when crosstol-ng is built within buildroot 
(and the option deselected of course)

I am still investigating, but if you have a clue I would appreciate

Thanks
Thierry

 >>> ncurses 5.9 Building
/usr/bin/make -j1 -C 
/shared/Buildroot/output.imx6hardfpu/build/ncurses-5.9 
DESTDIR=/shared/Buildroot/output.imx6hardfpu/host/usr/arm-unknown-linux-gnueabi/sysroot
make[1]: entrant dans le répertoire « 
/bigdisk/shared/Buildroot/output.imx6hardfpu/build/ncurses-5.9 »
cd man && /usr/bin/make -w -- 
DESTDIR=/shared/Buildroot/output.imx6hardfpu/host/usr/arm-unknown-linux-gnueabi/sysroot 
DESTDIR="/shared/Buildroot/output.imx6hardfpu/host/usr/arm-unknown-linux-gnueabi/sysroot" 
all
make[2]: entrant dans le répertoire « 
/bigdisk/shared/Buildroot/output.imx6hardfpu/build/ncurses-5.9/man »
make[2]: Rien à faire pour « all ».
make[2]: quittant le répertoire « 
/bigdisk/shared/Buildroot/output.imx6hardfpu/build/ncurses-5.9/man »
cd include && /usr/bin/make -w -- 
DESTDIR=/shared/Buildroot/output.imx6hardfpu/host/usr/arm-unknown-linux-gnueabi/sysroot 
DESTDIR="/shared/Buildroot/output.imx6hardfpu/host/usr/arm-unknown-linux-gnueabi/sysroot" 
all
make[2]: entrant dans le répertoire « 
/bigdisk/shared/Buildroot/output.imx6hardfpu/build/ncurses-5.9/include »
make[2]: Rien à faire pour « all ».
make[2]: quittant le répertoire « 
/bigdisk/shared/Buildroot/output.imx6hardfpu/build/ncurses-5.9/include »
cd ncurses && /usr/bin/make -w -- 
DESTDIR=/shared/Buildroot/output.imx6hardfpu/host/usr/arm-unknown-linux-gnueabi/sysroot 
DESTDIR="/shared/Buildroot/output.imx6hardfpu/host/usr/arm-unknown-linux-gnueabi/sysroot" 
all
make[2]: entrant dans le répertoire « 
/bigdisk/shared/Buildroot/output.imx6hardfpu/build/ncurses-5.9/ncurses »
linking ../lib/libncurses.so.5.9
/shared/Buildroot/output.imx6hardfpu/host/usr/bin/ccache 
/shared/Buildroot/output.imx6hardfpu/host/usr/bin/arm-unknown-linux-gnueabi-gcc 
-pipe -Os  -D_LARGEFILE_SOURCE -D_LARGEFILE64_SOURCE 
-D_FILE_OFFSET_BITS=64  --param max-inline-insns-single=1200 -shared 
-Wl,-soname,`basename ../lib/libncurses.so.5.9 .5.9`.5,-stats,-lc -o 
../lib/libncurses.so.5.9 ../obj_s/hardscroll.o ../obj_s/hashmap.o 
../obj_s/lib_addch.o ../obj_s/lib_addstr.o ../obj_s/lib_beep.o 
../obj_s/lib_bkgd.o ../obj_s/lib_box.o ../obj_s/lib_chgat.o 
../obj_s/lib_clear.o ../obj_s/lib_clearok.o ../obj_s/lib_clrbot.o 
../obj_s/lib_clreol.o ../obj_s/lib_color.o ../obj_s/lib_colorset.o 
../obj_s/lib_delch.o ../obj_s/lib_delwin.o ../obj_s/lib_echo.o 
../obj_s/lib_endwin.o ../obj_s/lib_erase.o ../obj_s/lib_flash.o 
../obj_s/lib_gen.o ../obj_s/lib_getch.o ../obj_s/lib_getstr.o 
../obj_s/lib_hline.o ../obj_s/lib_immedok.o ../obj_s/lib_inchstr.o 
../obj_s/lib_initscr.o ../obj_s/lib_insch.o ../obj_s/lib_insdel.o 
../obj_s/lib_insnstr.o ../obj_s/lib_instr.o ../obj_s/lib_isendwin.o 
../obj_s/lib_leaveok.o ../obj_s/lib_mouse.o ../obj_s/lib_move.o 
../obj_s/lib_mvcur.o ../obj_s/lib_mvwin.o ../obj_s/lib_newterm.o 
../obj_s/lib_newwin.o ../obj_s/lib_nl.o ../obj_s/lib_overlay.o 
../obj_s/lib_pad.o ../obj_s/lib_printw.o ../obj_s/lib_redrawln.o 
../obj_s/lib_refresh.o ../obj_s/lib_restart.o ../obj_s/lib_scanw.o 
../obj_s/lib_screen.o ../obj_s/lib_scroll.o ../obj_s/lib_scrollok.o 
../obj_s/lib_scrreg.o ../obj_s/lib_set_term.o ../obj_s/lib_slk.o 
../obj_s/lib_slkatr_set.o ../obj_s/lib_slkatrof.o 
../obj_s/lib_slkatron.o ../obj_s/lib_slkatrset.o ../obj_s/lib_slkattr.o 
../obj_s/lib_slkclear.o ../obj_s/lib_slkcolor.o ../obj_s/lib_slkinit.o 
../obj_s/lib_slklab.o ../obj_s/lib_slkrefr.o ../obj_s/lib_slkset.o 
../obj_s/lib_slktouch.o ../obj_s/lib_touch.o ../obj_s/lib_tstp.o 
../obj_s/lib_ungetch.o ../obj_s/lib_vidattr.o ../obj_s/lib_vline.o 
../obj_s/lib_wattroff.o ../obj_s/lib_wattron.o ../obj_s/lib_winch.o 
../obj_s/lib_window.o ../obj_s/nc_panel.o ../obj_s/safe_sprintf.o 
../obj_s/tty_update.o ../obj_s/varargs.o ../obj_s/memmove.o 
../obj_s/vsscanf.o ../obj_s/lib_freeall.o ../obj_s/expanded.o 
../obj_s/legacy_coding.o ../obj_s/lib_dft_fgbg.o ../obj_s/lib_print.o 
../obj_s/resizeterm.o ../obj_s/use_screen.o ../obj_s/use_window.o 
../obj_s/wresize.o ../obj_s/access.o ../obj_s/add_tries.o 
../obj_s/alloc_ttype.o ../obj_s/codes.o ../obj_s/comp_captab.o 
../obj_s/comp_error.o ../obj_s/comp_hash.o ../obj_s/db_iterator.o 
../obj_s/doalloc.o ../obj_s/entries.o ../obj_s/fallback.o 
../obj_s/free_ttype.o ../obj_s/getenv_num.o ../obj_s/home_terminfo.o 
../obj_s/init_keytry.o ../obj_s/lib_acs.o ../obj_s/lib_baudrate.o 
../obj_s/lib_cur_term.o ../obj_s/lib_data.o ../obj_s/lib_has_cap.o 
../obj_s/lib_kernel.o ../obj_s/lib_keyname.o ../obj_s/lib_longname.o 
../obj_s/lib_napms.o ../obj_s/lib_options.o ../obj_s/lib_raw.o 
../obj_s/lib_setup.o ../obj_s/lib_termcap.o ../obj_s/lib_termname.o 
../obj_s/lib_tgoto.o ../obj_s/lib_ti.o ../obj_s/lib_tparm.o 
../obj_s/lib_tputs.o ../obj_s/lib_trace.o ../obj_s/lib_ttyflags.o 
../obj_s/lib_twait.o ../obj_s/name_match.o ../obj_s/names.o 
../obj_s/read_entry.o ../obj_s/read_termcap.o ../obj_s/setbuf.o 
../obj_s/strings.o ../obj_s/tries.o ../obj_s/trim_sgr0.o 
../obj_s/unctrl.o ../obj_s/visbuf.o ../obj_s/define_key.o 
../obj_s/hashed_db.o ../obj_s/key_defined.o ../obj_s/keybound.o 
../obj_s/keyok.o ../obj_s/version.o -L../lib
/home/tbultel/x-tools/arm-unknown-linux-gnueabi/lib/gcc/arm-unknown-linux-gnueabi/4.6.3/../../../../arm-unknown-linux-gnueabi/bin/ld: 
error: ../lib/libncurses.so.5.9 uses VFP register arguments, 
../obj_s/hardscroll.o does not


Le 21/11/2012 13:25, Thomas Petazzoni a écrit :
> Dear Thierry Bultel,
>
> On Wed, 21 Nov 2012 11:23:20 +0100, Thierry Bultel wrote:
>
>> GCC has not been generated with the harware FPU support.
>>
>> I have attempted several hacks into
>> toolchain/toolchain-crosstool-ng/crosstool-ng.config-glibc
>>
>> ... with no better success
>>
>> Any ideas ?
>
> Yes, raise this problem on the crosstool-ng mailing list. I know Yann
> is aware of this, and he will most likely have recommendations on how
> to fix this problem.
>
> Thanks,
>
> Thomas
>


--
For unsubscribe information see http://sourceware.org/lists.html#faq

Re: [Buildroot] crosstools build fails when hard fpu is selected

[Thomas Petazzoni]

Dear Thierry Bultel,

On Thu, 22 Nov 2012 09:46:24 +0100, Thierry Bultel wrote:

> I thought that is was due to the fact that the "Use soft float by 
> default" option was not available when using an external toolchain (why 
> ?) but I have the same issue when crosstol-ng is built within buildroot 
> (and the option deselected of course)
> 
> I am still investigating, but if you have a clue I would appreciate

> /home/tbultel/x-tools/arm-unknown-linux-gnueabi/lib/gcc/arm-unknown-linux-gnueabi/4.6.3/../../../../arm-unknown-linux-gnueabi/bin/ld: 
> error: ../lib/libncurses.so.5.9 uses VFP register arguments, 
> ../obj_s/hardscroll.o does not

From a quick look, it smells like not everybody has been built with the
same CFLAGS or something like that. Like part of it has been built with
hardfp, and some other with softfp. I would think "use VFP register
arguments" means hardfp (floating point values are passed to functions
through the floating-point registers). So libncurses.so.5.9 would be
hardfp, and hardscroll.o would be softfp or soft-float and that cannot
be linked together.

Are you sure you disabled BR2_SOFT_FLOAT in Buildroot? Check the entire
build log of ncurses to see which compiler flags are being passed.

Best regards,

Thomas
-- 
Thomas Petazzoni, Free Electrons
Kernel, drivers, real-time and embedded Linux
development, consulting, training and support.
http://free-electrons.com

--
For unsubscribe information see http://sourceware.org/lists.html#faq

Re: [Buildroot] crosstools build fails when hard fpu is selected

[Thierry Bultel]

Le 22/11/2012 09:52, Thomas Petazzoni a écrit :
> Dear Thierry Bultel,
>
> On Thu, 22 Nov 2012 09:46:24 +0100, Thierry Bultel wrote:
>
>> I thought that is was due to the fact that the "Use soft float by
>> default" option was not available when using an external toolchain (why
>> ?) but I have the same issue when crosstol-ng is built within buildroot
>> (and the option deselected of course)
>>
>> I am still investigating, but if you have a clue I would appreciate
>
>> /home/tbultel/x-tools/arm-unknown-linux-gnueabi/lib/gcc/arm-unknown-linux-gnueabi/4.6.3/../../../../arm-unknown-linux-gnueabi/bin/ld:
>> error: ../lib/libncurses.so.5.9 uses VFP register arguments,
>> ../obj_s/hardscroll.o does not
>
>  From a quick look, it smells like not everybody has been built with the
> same CFLAGS or something like that. Like part of it has been built with
> hardfp, and some other with softfp. I would think "use VFP register
> arguments" means hardfp (floating point values are passed to functions
> through the floating-point registers). So libncurses.so.5.9 would be
> hardfp, and hardscroll.o would be softfp or soft-float and that cannot
> be linked together.
>
> Are you sure you disabled BR2_SOFT_FLOAT in Buildroot? Check the entire
> build log of ncurses to see which compiler flags are being passed.
>

Yes, I am definitively sure that BR2_SOFT_FLOAT is disabled in my .config
However I have BR2_PREFER_SOFT_FLOAT=y which is automatically set.

Looking at the build log, there are no '-mfpu=' flags at all when 
invoking crosstool's gcc

I do not reproduce the issue with the CodeSourcery as external 
toolchain, letting me think that my crosstool configuration is
wrong.

best regards
Thierry

> Best regards,
>
> Thomas
>


--
For unsubscribe information see http://sourceware.org/lists.html#faq