[Pythonmac-SIG] No arbitrary precison math on Mac-tel say it ain't so!

[Alex Martelli]

On Apr 22, 2006, at 1:29 PM, Ronald Oussoren wrote:
    ...
> On other platforms I'd say that gmp should be compiled with -fPIC  
> in the compiler flags. On OSX however this

That's one of the things I tried (adding CFLAGS=-fPIC to the ./ 
configure invocation), uselessly.

> should be the default. It could be a problem in an assembly file.  
> Given the .asm files and configure script I'd guess that it might  
> help to add '--with-pic' to the configure line.

OK, tried this - and now the assembly is failing:

/bin/sh ../libtool --mode=compile --tag=CC ../mpn/m4-ccas --m4="m4"  
gcc -c -DHAVE_CONFIG_H -I. -I. -I.. -D__GMP_WITHIN_GMP -I.. - 
DOPERATION_`echo dive_1 | sed 's/_$//'`    -fPIC `test -f  
'dive_1.asm' || echo './'`dive_1.asm
../mpn/m4-ccas --m4=m4 gcc -c -DHAVE_CONFIG_H -I. -I. -I.. - 
D__GMP_WITHIN_GMP -I.. -DOPERATION_dive_1 -fPIC dive_1.asm  -fno- 
common -DPIC -o dive_1.o
m4  -DHAVE_CONFIG_H -D__GMP_WITHIN_GMP -DOPERATION_dive_1 -DPIC  
dive_1.asm >tmp-dive_1.s
gcc -c -DHAVE_CONFIG_H -I. -I. -I.. -D__GMP_WITHIN_GMP -I.. - 
DOPERATION_dive_1 -fPIC tmp-dive_1.s -fno-common -DPIC -o dive_1.o
tmp-dive_1.s:108:invalid character '@' in first operand
make[2]: *** [dive_1.lo] Error 1
make[1]: *** [all-recursive] Error 1
make: *** [all] Error 2


The relevant fragment of that tmp-dive_1.s is:

     106         call    Lmovl_eip_ebp
     107         addl    $_GLOBAL_OFFSET_TABLE_, %ebp
     108         movl    ___gmp_modlimb_invert_table at GOT(%ebp), %ebp
     109         movzbl  (%eax,%ebp), %ebp

and the corresponding fragment of dive_1.s:

      79 ifdef(`PIC',`
      80         call    L(movl_eip_ebp)
      81         addl    $_GLOBAL_OFFSET_TABLE_, %ebp
      82         movl    modlimb_invert_table at GOT(%ebp), %ebp
      83         movzbl  (%eax,%ebp), %ebp                       C  
inv 8 bits
      84
      85 ',`
      86 dnl non-PIC
      87         movzbl  modlimb_invert_table(%eax), %ebp        C  
inv 8 bits
      88 ')

So, m4 is doing its job (so, your suggestion was no doubt right,  
thanks!), picking the PIC version rather the non-PIC one... and  
unfortunately the PIC version does hit some assembler-syntax issue,  
it would seem.


One note in the Makefile in that directory may be relevant...:

# Libtool adds -DPIC when building a shared object and the .asm files  
look
# for that.  But it should be noted that the other PIC flags are on  
occasion
# important too, in particular FreeBSD 2.2.8 gas 1.92.3 requires -k  
before
# it accepts PIC constructs like @GOT, and gcc adds that flag only under
# -fPIC.  (Later versions of gas are happy to accept PIC stuff any  
time.)


I can't find any docs for the "@GOT" syntax (presumably of GNU as),  
though I'm sure it's somehow referring to a global-offset-table --  
the only occurrence of '@' I've found in GNU as's Reference manual is  
about .symver (and it does appear relevant to what ld is later meant  
to do), but there are no occurrences of that directive in these  
source and include files anyway.

Unfortunately, no web search engine that I know of has any notion of  
letting you search for special characters such as '@', and the word  
'GOT' is so common that it is absolutely no use whatsoever - I've  
searched for 'global offset table, 'pic' and 'position independent  
code', and so on and so forth, but all to no avail.

Assuming there's some parallel between Apple assembler syntax for x86  
and PPC I've pored over the Powerpc assembler parts of GMP 4.2, and  
there are some very cryptic notes about assembly PIC and non-PIC for  
Linux AIX and Darwin, but nothing that seems to apply or help in any  
way here.


I'm rather stalled, so, for now, I've built GMP 4.2 to NOT use any  
assembly on my Mac-Intel and proceeded with other work needed to make  
gmpy compatible with GMP 4.2 (mostly tweaks on unittests, but one  
somewhat major rework to the random-number facility, which changed  
drastically in 4.2).


Anyway, thanks for the help so far -- my plans now include checking  
the current gmpy (with GMP 4.2 _and_ 4.1) on PPC, measuring speed (to  
guess how much I lose by having to disable assembly on Mac-Intel),  
further checking and measuring on Linux and Windows, etc.  I hope,  
before I release gmpy 1.02 at some point, to get back to the  
"assembly on Mac-Intel" issues... can anybody perchance suggest  
suitable newsgroups, mailing lists, or sites, where somebody might,  
upon seeing the issue with "whatever at GOT(%bx)" etc, immediately  
suggest some solution?-)


Alex