optimization/6028: libcall handling breaks 173.applu with -Os on ia64

[janis187@us.ibm.com]

>Number:         6028
>Category:       optimization
>Synopsis:       libcall handling breaks 173.applu with -Os on ia64
>Confidential:   no
>Severity:       serious
>Priority:       medium
>Responsible:    unassigned
>State:          open
>Class:          wrong-code
>Submitter-Id:   net
>Arrival-Date:   Thu Mar 21 18:06:01 PST 2002
>Closed-Date:
>Last-Modified:
>Originator:     Janis Johnson
>Release:        gcc version 3.1 20020321 (prerelease)
>Organization:
>Environment:
Itanium, Red Hat Linux release 7.1.94 (Roswell)
>Description:
The SPEC CPU2000 test 173.applu gets incorrect results when
compiled with -Os on ia64-unknown-linux-gnu with the 3.1
prerelease compiler.

This is a regression from GCC 3.0.4 that begins with this
change:

Sun Dec 23 00:49:37 CET 2001  Jan Hubicka  <jh@suse.cz>
        * calls.c (ECF_LIBCALL_BLOCK): New constant.
        (emit_call_1, initialize_argument_information,
        precompute_arguments, expand_call,
        emit_library_call_value_1): Use ECF_LIBCALL_BLOCK
        instead of ECF_PURE | ECF_CONST. Honnor LCT_CONST/LCT_PURE.

I have not yet put together a test case to include with
this PR, but I can continue to work on that if necessary.
The subroutine that is compiled incorrectly with -Os is
jaclf.f, which has three nested DO loops with lots of math,
four calls to __divdf3, and no other calls.  This test
passes with all other optimization options I've tried.
I used SPEC's test (small) input.
>How-To-Repeat:
Build and run 173.applu from SPEC CPU2000 with -Os.
>Fix:
Unknown.
>Release-Note:
>Audit-Trail:
>Unformatted: