egcs-1.0.3a and egcs-19980727 on RH51/alpha

[Stefan Schroepfer <scr at iis dot fhg dot de>]

I built and installed egcs-1.0.3a successfully on my Linux/Alpha
system, output of srcdir/config.guess is (for both versions):
alphaev5-unknown-linux-gnu.

Commands issued (most simple approach for both egcs versions):
% <srcdir>/configure
% make bootstrap
% make install       # impossible for egcs-19980727

Hardware / Software (as far as considered possibly relevant):
PC164 / 21164a
RH 5.1 from PowerTools CD set
glibc*-2.0.7-19.alpha.rpm from RH updates directory

This message is longer than a mere installation success report for
two reasons:
 - Failure building current egcs version on my system (using
   egcs 1.0.2).
 - Results of my private (heavily floating point arithmetics
   dominated) benchmark suite are appended.

Note: I tried to build a current version because of an error in
21164a optimisation (concerning egcs versions 1.0.2 and 1.0.3a).
This is described in a seperate mail I sent to egcs-bugs (subject
"deref_puc_assign,21164a on RH51/alpha").

Attempting to build egcs-19980727 leads to error messages during
compilation of libf2c/libI77/backspace.c.
First four lines of these messages:
------------------------------------------------------------
make[3]: Entering directory
`/mnt/sdb6/sw/misc/egcs-19980727-obj/alphaev5-unknown-linux-gnu/libf2c/libI77'
/mnt/sdb6/sw/misc/egcs-19980727-obj/gcc/xgcc
-B/mnt/sdb6/sw/misc/egcs-19980727-obj/gcc/ -c -DSkip_f2c_Undefs
-DAllow_TYQUAD -I. -I/userdisk2/sw/misc/egcs-19980727/libf2c/libI77 -I..
-I/userdisk2/sw/misc/egcs-19980727/libf2c/libI77/..  -DSTDC_HEADERS=1
-D_POSIX_SOURCE=1 -DHAVE_TEMPNAM=1 -DNON_ANSI_RW_MODES=1
-DNO_EOF_CHAR_CHECK=1 -DSkip_f2c_Undefs=1  -g -O2
/userdisk2/sw/misc/egcs-19980727/libf2c/libI77/backspace.c
In file included from
/userdisk2/sw/misc/egcs-19980727/libf2c/libI77/backspace.c:3:
/userdisk2/sw/misc/egcs-19980727/libf2c/libI77/fio.h:38: parse error
before `flag'
------------------------------------------------------------

Results of my private benchmark suite, comparing of egcs-1.0.3a
generated code to other compilers, all results in seconds using
clock (3) in all software environments:

Source Fragment  ENV1   ENV2   ENV3   ENV4
  "drwPh"        10.54  10.13  10.26   7.39
  "crPrj"         2.73   7.95   8.29   3.37
  "doART"        30.55  30.27  32.07  24.91

ENV1: RH 4.1 (#), gcc 2.7.2.1 (#), "-O2"
ENV2: RH 5.1 (+), egcs 1.0.2 (+), "-O2"
ENV3: RH 5.1 (+), egcs 1.0.3a, "-O2"
ENV4: NT4 (SP3), MSVC5 (SP3), "Maximize Speed", "/QA21164a"

(#):   from RH 4.1 PowerTools CD set, no performance relevant
       changes made IIRC
(+):   from RH 5.1 PowerTools CD set, updated glibc
"-O2": using -O2 as sole option affecting code generation,
       especially no -mieee* stuff needed for my benchmarks

Because of a runtime error in generated code using optimisation
for 21164a (-mcpu=21164a -Wa,-m21164a) I was not able to use this
feature of egcs.

If desired I could try to isolate those inner loops in "crPrj"
sources, which give such disappointing results.

Regards,
Stefan Schroepfer
scr@iis.fhg.de