Bug 33374 - GCC 4.1.2 produce wrong assembler code with -O2 option enabled.
Summary: GCC 4.1.2 produce wrong assembler code with -O2 option enabled.
Status: RESOLVED INVALID
Alias: None
Product: gcc
Classification: Unclassified
Component: target (show other bugs)
Version: 4.1.2
: P3 normal
Target Milestone: ---
Assignee: Not yet assigned to anyone
URL:
Keywords:
Depends on:
Blocks:
 
Reported: 2007-09-10 07:42 UTC by Yury Kirpichev
Modified: 2007-09-11 21:25 UTC (History)
3 users (show)

See Also:
Host:
Target:
Build:
Known to work:
Known to fail:
Last reconfirmed:


Attachments
test programm (913 bytes, text/plain)
2007-09-10 07:45 UTC, Yury Kirpichev
Details
Correct and incorrect assembler dumps. (585 bytes, text/plain)
2007-09-10 07:53 UTC, Yury Kirpichev
Details

Note You need to log in before you can comment on or make changes to this bug.
Description Yury Kirpichev 2007-09-10 07:42:10 UTC
I use gcc 4.1.2 under solaris and linux and compile code using next commands:
Linux:
gcc -pipe -Wall -Wextra -Winvalid-pch -ansi -pedantic -Wundef -Wc++-compat -Wfloat-equal -Wredundant-decls -Waggregate-return -Wbad-function-cast -Wmissing-declarations -Wmissing-prototypes -Wstrict-prototypes -Wcast-align -Wshadow -Wsign-compare -Wnested-externs -Wcast-qual -Wpointer-arith -Wdeclaration-after-statement -Werror -O2 -finline-functions -Wdisabled-optimization -DNDEBUG test.c -o mtest_l

Solaris:
gcc -pipe -Wall -Wextra -Winvalid-pch -ansi -pedantic -Wundef -Wc++-compat -Wfloat-equal -Wredundant-decls -Waggregate-return -Wbad-function-cast -Wstrict-prototypes -Wcast-align -Wshadow -Wsign-compare -Wnested-externs -Wcast-qual -Wpointer-arith -Wdeclaration-after-statement -Werror -O2 -finline-functions -Wdisabled-optimization -DNDEBUG -g -o mtest test.c

Segmentation fault happens when I try to run test program (see test.c attached).
Comment 1 Yury Kirpichev 2007-09-10 07:45:07 UTC
Created attachment 14182 [details]
test programm

This example is compiled incorrectly by gcc 4.1.2 with -O2 flag.
But all works fine if we change line 93 in this example on next:
XML_Remove(currNode_sp);
Comment 2 Yury Kirpichev 2007-09-10 07:53:40 UTC
Created attachment 14183 [details]
Correct and incorrect assembler dumps.

From incorrect dump we can see where segmentation fault heppens:

0x0804839e <testFunction+30>:   mov    DWORD PTR [ecx+4],0x0
0x080483a5 <testFunction+37>:   mov    DWORD PTR [eax+4],edx
0x080483a8 <testFunction+40>:   mov    edx,DWORD PTR [ecx+4]
0x080483ab <testFunction+43>:   mov    ecx,ebx
0x080483ad <testFunction+45>:   mov    DWORD PTR [edx],eax

Instruction 0x080483ad tries to dereferencing pointer in edx, but
in edx zero was put by 0x080483a8.
Comment 3 Andrew Pinski 2007-09-10 08:34:05 UTC
From the look at the source, it looks like you are violating C/C++ aliasing rules.  Try with -fno-strict-aliasing.
Comment 4 Yury Kirpichev 2007-09-10 08:53:54 UTC
There is no core dump with -fno-strict-aliasing.
But, our code is used by others who can not set this option.
Should code compiled without -fno-strict-aliasing produce core dump?
In my opinion, the best way is produce only compiliation warnings in this case.

Should we rewrite our code if it is not possible to set -fno-strict-aliasing option, or it is gcc issue that fixed in other versions?
Comment 5 Wolfgang Bangerth 2007-09-11 20:28:33 UTC
This code is certainly undefined and violates all sorts of type
correctness rules. In essence you write code that violates the language
standards -- using -fno-strict-aliasing is only a workaround and you should
fix your code.

You may want to try to rewrite your code so that you don't need all those
type casts. It isn't particularly hard to write a doubly-linked list without
any type casts at all.

W.
Comment 6 Rask Ingemann Lambertsen 2007-09-11 21:25:06 UTC
That code looks very similar to the doubly linked lists that AmigaOS exec.library uses. Just remove the XL and XLM prefixes from the names and it's the same. That code indeed plays nasty type casting tricks to reuse a NULL pointer in the list head, so as to save four bytes of memory. I guess aliasing violations were not a major concern when that code was written back in about 1985 or so. :-) Maybe the AROS project has revised code which survives alias analysis.