[Bug c/43987] New: [4.5 regression] type-punning causes broken binaries unless -O0 is used
I noticed this when building busybox 1.16.1, but I also saw this in other code: When building with -O2 I get warnings like: 'warning: dereferencing type-punned pointer will break strict-aliasing rules' The produced binary is broken, in the case of busybox, the 'sed' applet segfaults. When adding -fno-strict-aliasing, the warnings disappear, but the resulting binary is still broken (with no warning or error indicating that something is broken). The same also happens with -O1 or -Os, only a -O0 binary will work. This used to work fine in gcc 4.4. I don't know how a test case would look like, as I don't know the code that breaks that closely. In my opinion, one of the following should happen: a) gcc 4.5 should produce code that works b) gcc 4.5 should throw a warning or error indicating that the code is incorrect -- Summary: [4.5 regression] type-punning causes broken binaries unless -O0 is used Product: gcc Version: 4.5.0 Status: UNCONFIRMED Severity: critical Priority: P3 Component: c AssignedTo: unassigned at gcc dot gnu dot org ReportedBy: thomas at archlinux dot org GCC build triplet: x86_64-unknown-linux-gnu / i686-pc-linux-gnu GCC host triplet: x86_64-unknown-linux-gnu / i686-pc-linux-gnu GCC target triplet: x86_64-unknown-linux-gnu / i686-pc-linux-gnu http://gcc.gnu.org/bugzilla/show_bug.cgi?id=43987
[Bug middle-end/43987] [4.5 regression] type-punning causes broken binaries unless -O0 is used
--- Comment #3 from thomas at archlinux dot org 2010-05-04 23:06 --- What is -fwrapv supposed to do? I don't see any changes in behaviour or compiler output. As I said, I don't know exactly what to put into a test case (I didn't write any of the code that breaks), but I can try my best to find one. I first encountered this problem in gap (http://www.gap-system.org), and it was fixed by this patch: http://www.math.leidenuniv.nl/~wpalenst/sage/gap_gcc45_strict_aliasing.patch (look at the sysinfo.c part, that was what caused the problems apparently). I now encountered it again in busybox's 'sed' applet. I will try to take that code apart until I find out what's going wrong and hopefully produce a test case. -- http://gcc.gnu.org/bugzilla/show_bug.cgi?id=43987
[Bug middle-end/43987] [4.5 regression] type-punning causes broken binaries unless -O0 is used
--- Comment #5 from thomas at archlinux dot org 2010-05-04 23:49 --- I am still not able to produce a short test case, but maybe this will help: http://git.busybox.net/busybox/tree/editors/sed.c?h=1_16_stable&id=1_16_1#n738 In the function add_input_file(), the 'file' argument is not a NULL pointer, but at the end of the function, G.input_file_list[G.input_file_count-1]==NULL (and !=file in particular), which ultimately leads to the segfault. -- http://gcc.gnu.org/bugzilla/show_bug.cgi?id=43987
[Bug middle-end/43987] [4.5 regression] type-punning causes broken binaries unless -O0 is used
--- Comment #7 from thomas at archlinux dot org 2010-05-05 07:23 --- It has its own file: http://git.busybox.net/busybox/tree/libbb/xrealloc_vector.c?h=1_16_stable&id=1_16_1 -- http://gcc.gnu.org/bugzilla/show_bug.cgi?id=43987
[Bug middle-end/43987] [4.5 regression] type-punning causes broken binaries unless -O0 is used
--- Comment #9 from thomas at archlinux dot org 2010-05-05 10:53 --- Created an attachment (id=20561) --> (http://gcc.gnu.org/bugzilla/attachment.cgi?id=20561&action=view) Preprocessor output of sed.c -- http://gcc.gnu.org/bugzilla/show_bug.cgi?id=43987
[Bug middle-end/43987] [4.5 regression] type-punning causes broken binaries unless -O0 is used
--- Comment #10 from thomas at archlinux dot org 2010-05-05 10:58 --- (In reply to comment #8) > Waiting for a testcase (which could be just preprocessed source of > sed.c). Attached. If I find a better (shorter) test case, I'll add that too. > Can you reproduce the issue when you add -fno-inline? Yes. > Can you reproduce the issue with the current 4.5 branch? We've scheduled to build the latest snapshot next week. Maybe I can do it before then, I'll let you know as soon as I get to try. > As you can reproduce with -O1 I doubt this is an issue with type-punning. That was my thought too, but it seems to happen only in places where type-punning occurs. -- http://gcc.gnu.org/bugzilla/show_bug.cgi?id=43987
[Bug middle-end/43987] [4.5 regression] type-punning causes broken binaries unless -O0 is used
--- Comment #11 from thomas at archlinux dot org 2010-05-05 11:57 --- The problem persists with the 4.5-20100429 snapshot. -- http://gcc.gnu.org/bugzilla/show_bug.cgi?id=43987
[Bug middle-end/43987] [4.5/4.6 Regression] type-punning causes broken binaries unless -O0 is used
--- Comment #17 from thomas at archlinux dot org 2010-05-05 14:10 --- > I have a patch. Workaround for 4.5.0: -fno-pta. There is no -fno-pta option in 4.5.0, but I can confirm that -fno-tree-pta fixes the problem in the testcase and in busybox. -- http://gcc.gnu.org/bugzilla/show_bug.cgi?id=43987
[Bug middle-end/43987] [4.5 Regression] type-punning causes broken binaries unless -O0 is used
--- Comment #24 from thomas at archlinux dot org 2010-05-21 08:16 --- Preliminary tests with a recent gcc snapshot from the 4.5 branch shows that the problem is entirely solved by this patch (maybe in combination with other fixes). Thanks for the quick responses and fixes. -- http://gcc.gnu.org/bugzilla/show_bug.cgi?id=43987
