With revision 192137 this test passed. From revision 192147 and on, this test has failed as follows: Running /tmp/hpautotest-gcc1/gcc/libstdc++-v3/testsuite/libstdc++-dg/conformance.exp ... FAIL: 23_containers/bitset/45713.cc (test for excess errors) With the message in the logfile being: /tmp/hpautotest-gcc1/gcc/libstdc++-v3/testsuite/23_containers/bitset/45713.cc:24:55: error: size of array 'test' is not\ an integral constant-expression^M int test[sizeof(std::bitset<0xffffffff>) != 1 ? 1 : -1];^M ^^M compiler exited with status 1 FWIW, the cris-axis-elf ABI is effectively __attribute__ ((__packed__)). Author of the only suspect patch in this revision range CC:ed. (Did this patch change the ABI for some targets, Jakub?) I'll attach a preprocessed file.
The problem is that IMHO the testcase is just invalid on 32-bit HWI targets. On 32-bit targets with 64-bit HWI, bitsizetype is 64-bit and thus std::bitset<0xffffffff> has size 0x20000000, and bitsize 0x100000000. But 32-bit HWI targets use 32-bit bitsizetype, thereforestd::bitset<0xffffffff> is larger than the largest supportable type, bitsize of the array is 0 overflow, and both bitsize and unit size of the bitset are therefore set to 0 overflow. Before my changes, sizeof(std::bitset<0xffffffff>) yielded 0 overflow, so the test magically passed even when sizeof 0 is as wrong as sizeof 1. With the patch, maybe_constant_size is used to fold the SIZEOF_EXPR, but as it is overflown, it is not considered a valid integer constant expression, thus we end up with size unfolded and therefore the error (or considering the array a VLA).
See if changing the test to the following (unconditional) works for you: int test[sizeof(std::bitset<__SIZE_MAX__>) != 1 ? 1 : -1]; It passes on x86_64-linux, -m32 and -m64. In case, you are welcome to commit the change. I think it does still test for the original issue on the affected targets. Anyway, it was only a quick try of mine at adapting the testcase provided by submitter, the actual issue was just a thinko, IMO we shouldn't spend too much time on this.
Actually, I was wrong about 32-bit HWI, the actual problem is MAX_FIXED_MODE_SIZE. bprecision = MIN (precision + BITS_PER_UNIT_LOG + 1, MAX_FIXED_MODE_SIZE); bprecision = GET_MODE_PRECISION (smallest_mode_for_size (bprecision, MODE_INT)); if (bprecision > HOST_BITS_PER_DOUBLE_INT) bprecision = HOST_BITS_PER_DOUBLE_INT; In cris case that is MIN (32 + 3 + 1, 32), while i?86/x86_64 have 64 resp. 128 (-m32 resp. -m64). So, such change isn't going to improve anything for cris, where simply any variables etc. of size 0x20000000 and larger can't be supported. Targets with similar issues: avr, h8300, mcore, moxie. I'd say the testcase should just be skipped for those targets.
Ah! I'm Ok with xfailing - I'm leaving that to you - or we can just remove the test, isn't a big deal.
I'd go with --- libstdc++-v3/testsuite/23_containers/bitset/45713.cc 2010-09-22 17:15:42.000000000 +0200 +++ libstdc++-v3/testsuite/23_containers/bitset/45713.cc 2012-10-11 12:28:49.865370623 +0200 @@ -1,4 +1,4 @@ -// Copyright (C) 2010 Free Software Foundation, Inc. +// Copyright (C) 2010, 2012 Free Software Foundation, Inc. // // This file is part of the GNU ISO C++ Library. This library is free // software; you can redistribute it and/or modify it under the @@ -15,7 +15,10 @@ // with this library; see the file COPYING3. If not see // <http://www.gnu.org/licenses/>. -// { dg-do compile } +// The testcase requires bitsizetype to be wider than sizetype, +// otherwise types/vars with 0x20000000 bytes or larger can't be used. +// See http://gcc.gnu.org/PR54897 +// { dg-do compile { target { ! { avr*-* cris*-* h8300*-* mcore*-* moxie*-* } } } } #include <bitset> but don't have time to test it on those targets...
(In reply to comment #3) > Actually, I was wrong about 32-bit HWI, the actual problem is > In cris case that is MIN (32 + 3 + 1, 32), while i?86/x86_64 have 64 resp. 128 > (-m32 resp. -m64). So, such change isn't going to improve anything for cris, > where simply any variables etc. of size 0x20000000 and larger can't be > supported. I hope you mean the actual size, not holding that value and larger. > Targets with similar issues: avr, h8300, mcore, moxie. > I'd say the testcase should just be skipped for those targets. Ok: I'm happy with a skipping patch until I find time to investigate fallout (and if it'd make sense) changing MAX_FIXED_MODE_SIZE for CRIS. The docs are a bit misleading, DImode works fine, for instance. And of course, it would be right for remaining targets anyway. If there are more tests like this, we can consider an effective-target. I'll test and commit. Thanks.
Correcting component as per conclusion.
Author: hp Date: Thu Oct 11 11:36:39 2012 New Revision: 192354 URL: http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=192354 Log: PR testsuite/54897 * testsuite/23_containers/bitset/45713.cc: Skip for avr*-*-* cris*-*-* h8300*-*-* mcore*-*-* moxie*-*-*. Modified: trunk/libstdc++-v3/ChangeLog trunk/libstdc++-v3/testsuite/23_containers/bitset/45713.cc
...and closing.