From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: (qmail 13629 invoked by alias); 6 Jan 2004 18:10:30 -0000 Mailing-List: contact gcc-bugs-help@gcc.gnu.org; run by ezmlm Precedence: bulk List-Archive: List-Post: List-Help: Sender: gcc-bugs-owner@gcc.gnu.org Received: (qmail 13620 invoked by uid 48); 6 Jan 2004 18:10:30 -0000 Date: Tue, 06 Jan 2004 18:10:00 -0000 Message-ID: <20040106181030.13619.qmail@sources.redhat.com> From: "danglin at gcc dot gnu dot org" To: gcc-bugs@gcc.gnu.org In-Reply-To: <20031216031240.13408.danglin@gcc.gnu.org> References: <20031216031240.13408.danglin@gcc.gnu.org> Reply-To: gcc-bugzilla@gcc.gnu.org Subject: [Bug ada/13408] acats numeric tests cxg* fail on pa/hpux X-Bugzilla-Reason: CC X-SW-Source: 2004-01/txt/msg00539.txt.bz2 List-Id: ------- Additional Comments From danglin at gcc dot gnu dot org 2004-01-06 18:10 ------- There's potentially two separate issues involved in this PR and 13528: 1) The float to integer conversion of an Inf which causes a SIGFPU and the termination of the acats test sequence. This is definitely a testsuite/ada problem. I don't have a copy of the IEEE spec handy but the PA-RISC 2.0 Architecture book indicates that an unimplemented exception is always signaled for this condition. 2) The failure of various cxg* tests (and some others) due to the error limit for the test being exceeded by a small amount. I suspect that the cxg* failures have been present on the hppa2.0w-hp-hpux11 port since the acats testsuite was installed. This report represented the first successful build of ada since the testsuite was installed. I could do a build just after the testsuite was installed to see if there was any regression from that time. I would have to backport Roger Sayle's loop.c patch. It resolved the miscompilation of gnat1 which prevented builds for a long time. The numeric failures don't occur on the hppa-linux port. Thus, I doubt the problem is directly in the GCC code. The primary differences between the two ports are: a) HP libc/libm versus glibc b) Long double on hppa-linux port is the same as double. Long double on hppa-hpux port uses HP "quadlib" routines in libc. c) Linux port has flat memory model and doesn't need to be built with -mdisable-indexing. `c' usually results in segmentation faults, so I doubt it is the problem. -- http://gcc.gnu.org/bugzilla/show_bug.cgi?id=13408