arm CPU support

[vicuna]

Original bug ID: 3746
Reporter: administrator
Assigned to: @xavierleroy
Status: closed (set by @xavierleroy on 2011-05-29T10:14:22Z)
Resolution: fixed
Priority: normal
Severity: feature
Category: ~DO NOT USE (was: OCaml general)
Monitored by: @glondu bacam karl n8gray gerd

Bug description

Full_Name: Robert Nagy
Version: 3.08.3
OS: OpenBSD
Submission from: amazonas-2734.adsl.datanet.hu (195.56.242.194)

I am trying to build ocaml on OpenBSD/Zaurus.
cpu0 at mainbus0: PXA27x rev 4 (XScale core)
cpu0: DC enabled IC enabled WB enabled LABT branch prediction enabled
cpu0: 32KB(32b/l,32way) I-cache, 32KB(32b/l,32way) wr-back-lock D-cache
hw.machine=zaurus
hw.model=PXA27x rev 4 (XScale core)
hw.ncpu=1
hw.byteorder=1234
hw.physmem=67108864
hw.usermem=66850816
hw.pagesize=4096
hw.disknames=wd0
hw.diskcount=1

But it seems that there are some problems in arm.S

gcc -c -DSYS_openbsd -pthread -o arm.o arm.S ||
{ echo "If your assembler produced syntax errors, it is probably unhappy with
the"; echo "preprocessor. Check your assembler, or try producing arm.o by
hand."; exit 2; }
arm.S: Assembler messages:
arm.S:21: Warning: ignoring redefinition of register alias 'sp'
arm.S:22: Warning: ignoring redefinition of register alias 'lr'
arm.S:23: Warning: ignoring redefinition of register alias 'pc'
arm.S:112: Error: selected processor does not support stfd f0,[sp,#-8]!' arm.S:113: Error: selected processor does not support stfd f1,[sp,#-8]!'
arm.S:114: Error: selected processor does not support stfd f2,[sp,#-8]!' arm.S:115: Error: selected processor does not support stfd f3,[sp,#-8]!'
arm.S:125: Error: selected processor does not support ldfd f4,[sp],#8' arm.S:126: Error: selected processor does not support ldfd f5,[sp],#8'
arm.S:127: Error: selected processor does not support ldfd f6,[sp],#8' arm.S:128: Error: selected processor does not support ldfd f7,[sp],#8'
arm.S:184: Error: selected processor does not support stfd f7,[sp,#-8]!' arm.S:185: Error: selected processor does not support stfd f6,[sp,#-8]!'
arm.S:186: Error: selected processor does not support stfd f5,[sp,#-8]!' arm.S:187: Error: selected processor does not support stfd f4,[sp,#-8]!'
arm.S:240: Error: selected processor does not support ldfd f4,[sp],#8' arm.S:241: Error: selected processor does not support ldfd f5,[sp],#8'
arm.S:242: Error: selected processor does not support ldfd f6,[sp],#8' arm.S:243: Error: selected processor does not support ldfd f7,[sp],#8'
If your assembler produced syntax errors, it is probably unhappy with the
preprocessor. Check your assembler, or try producing arm.o by hand.
gmake: *** [arm.o] Error 2

[vicuna]

Comment author: administrator

I am trying to build ocaml on OpenBSD/Zaurus.
But it seems that there are some problems in arm.S
gcc -c -DSYS_openbsd -pthread -o arm.o arm.S ||
{ echo "If your assembler produced syntax errors, it is probably unhappy with
the"; echo "preprocessor. Check your assembler, or try producing arm.o by
hand."; exit 2; }
arm.S: Assembler messages:
arm.S:112: Error: selected processor does not support `stfd f0,[sp,#-8]!'

Apparently, your gcc toolchain is configured for "soft
floating-point", i.e. no float instructions such as stfd and ldfd.
The ARM port of ocamlopt uses "hard floating-point", with float
instructions, possibly emulated by the kernel. Hence the error.

You could try to add -mhard-float to ASPPFLAGS in config/Makefile,
but it is likely you'll run into errors later while trying to mix
softfloat-compiled code with hardfloat-compiled code.

If you really want to run ocamlopt on this platform, you should
investigate the availability of a gcc toolchain and standard libraries
built in "hard floating-point" mode.

Regards,

• Xavier Leroy

[vicuna]

Comment author: administrator

wish: ARM port with soft floating-point

[vicuna]

Comment author: mindbender

i can only second that wish.

we at www.linkstationwiki.net tried to compile ocaml as we needed it to compile mldonkey for the LS Pro...and marvel-based board with arm9-CPU (like MANY nas devices the get released the last months!..currently i count 6 that are based on the same marvel board)....

this board does not have support for hardware-floating-point...
http://forum.linkstationwiki.net/index.php?action=vthread&forum=16&topic=2151

we need support for soft-floating-point on arm also.....

[vicuna]

Comment author: gerd

I've also tried to build ocamlopt for my Nokia N800, and ran into the mentioned fp problem. As I see it the current arm status of ocaml is "out of date" (and useless):

• No hardware is available that supports fpa natively. Kernel emulation
  is slow (and I couldn't get it working, maybe broken)
• The way to go is to use EABI with softfp, and optionally vfp (the new
  hardware feature that speeds fp up). EABI is the new standard by ARM.
• The toolchains for the current devices base on EABI, and it is difficult
  to build alternate ones
• The EABI makes it easy to support soft fp for compilers, as there are
  standardized functions for fp (no OS-specific libraries etc.)
• The EABI provides IEEE-compliant fp (didn't we want that?)

So I think it is now time to switch to EABI. Here are a few links that might be helpful:

• http://www.arm.com/products/DevTools/ABI.html
• http://wiki.debian.org/ArmEabiPort
• http://en.wikipedia.org/wiki/ARM_architecture (lists ARM-based products)

For my Nokia there is an SDK that can be installed on x86 Debian systems. It includes the toolchain, and transparent CPU emulation (using qemu). I could successfully compile ocaml on it (except fp for native code). It might be good enough as development platform for an up-to-date ARM port:

• http://maemo.org/development/sdks/maemo-3-x--bora-.html
  Probably you need only scratchbox (scratchbox.org)

If you need to run code on a real arm device, please tell me - I've sshd on the N800, and can leave it on now and then.

---

Btw. I saw that the "as" invocation is hardcoded in asmrun/proc.ml. By adding
-mfloat-abi=hard -mfpu=fpa to the "as" call I could at least build native code programs, although fp operations do not work (wrong results, but no "illegal instruction"). I guess this is an error in the Linux kernel, since mixing hardfp and softfp is said to work under EABI.

[vicuna]

Comment author: @xavierleroy

Thanks for the background information. After reviewing it, I think there are two possible approaches.

One is to use VFD registers and instructions, which should fit very well the OCaml back-end architecture. For calling external C functions, softfp conventions would be used, for compatibility with EABI. This approach relies crucially on the existence of a good VFP emulation in the kernel, though.

The other is to go all the way to softfp in OCaml-generated code. It violates a number of basic assumptions of the back-end, e.g. where the back-end expects one float pseudo-register, there must be two integer pseudo-registers instead. I'm not sure this can be made to work at all.

[vicuna]

Comment author: @glondu

The latter seems more elegant. Would a patch be accepted?

[vicuna]

Comment author: @xavierleroy

It took a long time, but I finally managed to update the ARM OCaml port to EABI and soft floating-point. Everything seems to work well under Debian Lenny ARM-EABI running under QEMU emulation. The new port is in the CVS trunk and will be released in 3.12, replacing the previous (old-ABI) ARM port.