--- Begin Message ---
- From: Steven Munroe <munroesj at us dot ibm dot com>
- To: David Edelsohn <dje at watson dot ibm dot com>, Roland McGrath <roland at redhat dot com>,Jakub Jelinek <jakub at redhat dot com>, Peter Bergner <bergner at us dot ibm dot com>
- Cc: libc-alpha at sources dot redhat dot com
- Date: Wed, 31 May 2006 15:04:30 -0500
- Subject: Status of long double soft-fp support for powerpc32
- Organization: IBM LTC
I am making progress on fixing soft-fp for powerpc32. I am using a patch
from David that provides code gen for -msoft-float -mlong-double-128
and I have implemented the IBM extented varients of addtf3, divtf3,
eqtf2, extenddftf2, extendsftf2, getf2, letf2, multf3, negtf2, sqrttf2,
subtf3, unordtf2. I also found come comflicts with the existing
ldbl-128ibm implementation, that required some minor changes (I'll be
submitting patches for these soon). With this work I am able to build
and run make check.
This raised a few questions. Where should this code go in the hierarcy?
and which naming (/libc soft-fp or gcc) should we use?
There are existing ieee-128 implementation in ./soft-fp and again in
./sysdeps/powerpc/soft-fp. For example there ./soft-fp/addtf3.c
implementing __addtf3() and ./sysdeps/powerpc/soft-fp /q_add.c
implementing q_add(). These are effectively the same except for the type
(one uses TFtype, the other long double). My implementation is currently
in ./ports/sysdeps/unix/sysv/linux/powerpc/nofpu/addtf3.c.
Clearly for powerpc32 linux implementations I need to override the IEEE
quad implementation. I believe that there are some other (not linux)
powerpc32 implementations that still use IEEE. But so far my soft-fp
IBM extended implementations are generic. So should we add a
./soft-fp/ldbl-128ibm directory? In trunk or ports? Should the existing
soft-fp quad implementation be moved to ./soft-fp/ldbl-128? These
changes would require some carefully placed Implies files to get the
desired soft-fp implementation.
At the moment we are not actually using the quad soft-fp implements from
glibc. GCC is already calling out of line for most IBM extented
operations. For example gcc already generates calls to __gcc_qadd,
__gcc_qmul etc for the hardware double implementation of IBM extended.
With the current patch this continues with generated libm code using the
implementations of __gcc_g??? implementations from libgcc. The same gcc
uses the libc implementations (__adddf3, __muldf3, etc) for soft-fp double.
Not sure if this a problem, but we should decide if this is the best
approach before we release this code. One immediate problem is that I am
seeing (only) ~200 failures in test-ldouble and some debuging is
required. For me this would be easier if the code used the glibc/soft-fp
implementation. As an experiment I tried using strong_alias to override
override the gcc_quad implementation with the libc/soft-fp code, but
this just resulted in the link failure below.
strong_alias (__addtf3, __gcc_qadd)
strong_alias (__subtf3, __gcc_qsub)
strong_alias (__multf3, __gcc_qmul)
strong_alias (__divtf3, __gcc_qdiv)
strong_alias (__extendsftf2, __gcc_stoq)
strong_alias (__extenddftf2, __gcc_dtoq)
strong_alias (__eqtf2, __gcc_qeq)
strong_alias (__eqtf2, __gcc_qne)
strong_alias (__letf2, __gcc_qle)
strong_alias (__letf2, __gcc_qlt)
strong_alias (__getf2, __gcc_qge)
strong_alias (__getf2, __gcc_qqt)
/opt/biarch/gcc41-softld/lib/gcc/powerpc64-linux/4.1.1/nof/libgcc.a(darwin-ldouble.o):
In function `__gcc_qdiv':
/home/sjmunroe/toolchain/src/gcc/gcc/config/rs6000/darwin-ldouble.c:183:
multiple definition of `__gcc_qdiv'
/home/sjmunroe/work/build32-nofpu/libc_pic.os:../ports/sysdeps/unix/sysv/linux/powerpc/nofpu/divtf3.c:51:
first defined here
/opt/biarch/gcc41-softld/bin/ld: Warning: size of symbol `__gcc_qdiv'
changed from 696 in /home/sjmunroe/work/build32-nofpu/libc_pic.os to 716
in
/opt/biarch/gcc41-softld/lib/gcc/powerpc64-linux/4.1.1/nof/libgcc.a(darwin-ldouble.o)
--- End Message ---