[Bug c/45204] New: gcc generates incorrect code

[Bug c/45204] New: gcc generates incorrect code

[contact at philipashmore dot com]

Apologies in advance for not being able to create a small code sample.
The code is in SourceForge GIT repository for treedb.
To build treedb you will need v3c - the GIT version.

First download v3c and "make prefix=[install-location] && [sudo] make install".
cd [treedb-dir]
"make debug check" works but "make git branch=master check" fails - hangs.

The release build ("make release") is definately incorrect.

The code has some hacks for strict aliasing but contrary to the documentation,
these aren't caught with "-fstrict-aliasing -Wstrict-aliasing".

The makefile I used to build gcc and it's required libraries follows.

all:
        cd build/gmp && ../../gmp-4.3.2/configure --prefix=/opt/gcc
--enable-cxx && \
        make && sudo make install
        cd build/mpfr && ../../mpfr-3.0.0/configure --prefix=/opt/gcc
--with-gmp=/opt/gcc && \
        make && sudo make install
        cd build/mpc && ../../mpc-0.8.2/configure --prefix=/opt/gcc
--with-mpfr=/opt/gcc && \
        make && sudo make install
        cd build/ppl && ../../ppl-0.10.2/configure --prefix=/opt/gcc
--with-libgmp-prefix=/opt/gcc --with-libgmpxx-prefix=/opt/gcc --enable-cxx
--no-create --no-recursion && \
        make && sudo make install
        cd build/cloog-ppl && ../../cloog-ppl-0.15.9/configure
--with-ppl=/opt/gcc --with-gmp=/opt/gcc --prefix=/opt/gcc && \
        make && sudo make install
        cd build/gcc-4.5.1 && ../../gcc-4.5.1/configure -v
--with-pkgversion="Philip Ashmore's build 4.5.1-1"
--with-bugurl=file:///opt/gcc/share/doc/gcc-4.5/README.Bugs
--enable-languages=c,c++ --prefix=/opt/gcc --enable-shared
--enable-linker-build-id --with-system-zlib --libexecdir=/opt/gcc/lib
--without-included-gettext --enable-threads=posix
--with-gxx-include-dir=/opt/gcc/include/c++/4.5 --program-suffix=-4.5
--enable-nls --enable-clocale=gnu --enable-libstdcxx-debug --with-arch-32=i586
--with-tune=core2 --enable-checking=release --build=x86_64-linux-gnu
--host=x86_64-linux-gnu --target=x86_64-linux-gnu --with-mpc=/opt/gcc
--with-mpfr=/opt/gcc --with-gmp=/opt/gcc --with-ppl=/opt/gcc
--with-cloog=/opt/gcc && \
        make && sudo make install


-- 
           Summary: gcc generates incorrect code
           Product: gcc
           Version: 4.5.1
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: c
        AssignedTo: unassigned at gcc dot gnu dot org
        ReportedBy: contact at philipashmore dot com
 GCC build triplet: x86_64-linux-gnu
  GCC host triplet: x86_64-linux-gnu
GCC target triplet: x86_64-linux-gnu


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=45204

[Bug c/45204] gcc generates incorrect code

[rguenth at gcc dot gnu dot org]

------- Comment #1 from rguenth at gcc dot gnu dot org  2010-08-06 10:28 -------
We need a testcase.  Also please try -fno-strict-aliasing if you know the
code is bogus.


-- 

rguenth at gcc dot gnu dot org changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |WAITING


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=45204

[Bug c/45204] gcc generates incorrect code

[contact at philipashmore dot com]

------- Comment #2 from contact at philipashmore dot com  2010-08-06 13:37 -------
It's exactly what I tried first - I know there are obvious cases as per your
"frequently reported bugs" section.
I wanted the compiler to tell me where the aliasing was occurring in these and
in
any less obvious places so I could disable aliasing optimizations in these
cases.

Currently it's missing even the ones the docs claims it should spot - pointer->
uintptr_t -> pointer.

Previous release of treedb didn't need -fno-strict-aliasing, or did they?

If the compiler can't spot aliasing should -fno-strict-aliasing be the default?


-- 

contact at philipashmore dot com changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |contact at philipashmore dot
                   |                            |com
             Status|WAITING                     |UNCONFIRMED


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=45204

[Bug c/45204] gcc generates incorrect code

[contact at philipashmore dot com]

------- Comment #3 from contact at philipashmore dot com  2010-08-06 13:52 -------
Maybe I should add that the 0.6.0-beta1 release in GIT passed uintptr_t - sized
structures by value and the compiler spotted the aliasing, which is why I
introduced the pointer -> uintptr_t -> pointer hacks to begin with.

Without passing structs by value the compiler doesn't report the aliasing
problems.

I'm being sincere when I say that if the compiler really can't spot aliasing
problems then -fno-strict-aliasing should be on by default.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=45204

[Bug c/45204] gcc generates incorrect code

[rguenth at gcc dot gnu dot org]

------- Comment #4 from rguenth at gcc dot gnu dot org  2010-08-06 13:55 -------
(In reply to comment #3)
> Maybe I should add that the 0.6.0-beta1 release in GIT passed uintptr_t - sized
> structures by value and the compiler spotted the aliasing, which is why I
> introduced the pointer -> uintptr_t -> pointer hacks to begin with.
> 
> Without passing structs by value the compiler doesn't report the aliasing
> problems.
> 
> I'm being sincere when I say that if the compiler really can't spot aliasing
> problems then -fno-strict-aliasing should be on by default.

The compiler can spot all aliasing that is permitted by the language standard.
If you are writing non-C code then you need to tell that to the compiler
via -fno-strict-aliasing.  Merely adding hacks to avoid the diagnostic
(which can't be perfect, if we could detect the aliasing we wouldn't
optimize based on the fact that it isn't there).


-- 

rguenth at gcc dot gnu dot org changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |RESOLVED
         Resolution|                            |INVALID


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=45204

[Bug c/45204] gcc generates incorrect code

[contact at philipashmore dot com]

[-- Warning: decoded text below may be mangled, UTF-8 assumed --]
[-- Attachment #1: Type: text/plain, Size: 1899 bytes --]



------- Comment #5 from contact at philipashmore dot com  2010-08-07 02:01 -------
In my projects I add "-isystem ${top_srcdir}", the top source directory being
v3c.
This way, the examples can
    #include <v3c/avl.h>
Just like code you might develop yourself using treedb.

I do a lot of
   gcc -E -c a.c ... -o a.cc
followed by
   gcc -x c -c a.cc -o a.o ...
to track down problems in the code generated by macro expansion, and as a
last ditch effort, I decided to try to compile an object file from the
"cc" file, and debug the resulting program.

There they were - the aliasing violations!

It turns out that using -isystem for local include files was effectively
telling gcc not to report problems in header files included in this way.

You can check this yourself with these "snippets"

1: my-alias.c

  #include <alias-header.h>

2: alias-header.h

  #include <stdio.h>
  void test()
  {
    short a[2];
    a[0]=0x1111;
    a[1]=0x1111;
    *(int *)a = 0x22222222;

    printf("%x %x\n", a[0], a[1]);
  }

Run this:

  $ gcc -c -isystem. -fstrict-aliasing -Wstrict-aliasing my-alias.c

and you will get no warnings or errors.

3: my-alias.cc

  #include <stdio.h>
  void test()
  {
    short a[2];
    a[0]=0x1111;
    a[1]=0x1111;
    *(int *)a = 0x22222222;

    printf("%x %x\n", a[0], a[1]);
  }

Run this:

  $ gcc -x c -c -isystem. -fstrict-aliasing -Wstrict-aliasing my-alias.cc

and you will get

  my-alias.cc: In function ‘test’:
  my-alias.cc:7: warning: dereferencing type-punned pointer will break
strict-aliasing rules


-- 

contact at philipashmore dot com changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|RESOLVED                    |UNCONFIRMED
         Resolution|INVALID                     |


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=45204

[Bug c/45204] gcc doesn't report aliasing problems in -isystem includes

[contact at philipashmore dot com]

------- Comment #6 from contact at philipashmore dot com  2010-08-07 04:03 -------
I think the rename better describes this problem - please change it if you can
think up a better title.


-- 

contact at philipashmore dot com changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
            Summary|gcc generates incorrect code|gcc doesn't report aliasing
                   |                            |problems in -isystem
                   |                            |includes


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=45204

[Bug c/45204] gcc doesn't report aliasing problems in -isystem includes

[pinskia at gcc dot gnu dot org]

------- Comment #7 from pinskia at gcc dot gnu dot org  2010-08-07 04:09 -------
Use -Wsystem-headers.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=45204

[Bug c/45204] gcc doesn't report aliasing problems in -isystem includes

[rguenth at gcc dot gnu dot org]

------- Comment #8 from rguenth at gcc dot gnu dot org  2010-08-07 10:24 -------
See comment #7, this is a feature, not a bug.


-- 

rguenth at gcc dot gnu dot org changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |RESOLVED
         Resolution|                            |INVALID


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=45204