Re: target/6624: SSE misalignment with -O0

Re: target/6624: SSE misalignment with -O0

[Tim Prince]

The following reply was made to PR target/6624; it has been noted by GNATS.

From: Tim Prince <tprince@computer.org>
To: jdhall@uiuc.edu, gcc-gnats@gcc.gnu.org
Cc:  
Subject: Re: target/6624: SSE misalignment with -O0
Date: Fri, 10 May 2002 10:49:20 -0700

 On Friday 10 May 2002 09:12, Jesse Hall wrote:
 > >Number:         6624
 > >Category:       target
 > >Synopsis:       SSE mis-alignment with -O0
 > >Confidential:   no
 > >Severity:       serious
 > >Priority:       medium
 > >Responsible:    unassigned
 > >State:          open
 > >Class:          wrong-code
 > >Submitter-Id:   net
 > >Arrival-Date:   Fri May 10 09:16:01 PDT 2002
 > >Closed-Date:
 > >Last-Modified:
 > >Originator:     Jesse Hall
 > >Release:        3.1 20020426 (prerelease)
 > >Organization:
 > >Environment:
 >
 > System: Linux espresso 2.4.18-xfs #1 SMP Tue Mar 12 20:03:55 CST 2002 i686
 > unknown Architecture: i686 (more precisely, dual AthlonMP)
 >
 >
 > host: i386-pc-linux-gnu
 > build: i386-pc-linux-gnu
 > target: i386-pc-linux-gnu
 > configured with: ../src/configure -v
 > --enable-languages=c,c++,java,f77,proto,objc --prefix=/usr/lib/gcc-snapshot
 > --infodir=/share/info --mandir=/share/man --enable-shared --with-gnu-as
 > --with-gnu-ld --with-system-zlib --enable-long-long --enable-nls
 > --without-included-gettext --disable-checking --enable-threads=posix
 > --enable-java-gc=boehm --with-cpp-install-dir=bin --enable-objc-gc
 > i386-linux
 >
 > >Description:
 >
 >     The SSE intrinsic _mm_loadu_ps can do an aligned SSE load (LOADAPS)
 > from an unaligned address when compiling with -O0. The problem seems to go
 > away with -O1 and higher. This causes a segfault when the program is run.
 > The assembly generated looks like:
 >
 >         call   804834a <_mm_loadu_ps>
 >         add    $0x10,%esp
 >         movaps %xmm0,0xffffffe8(%ebp)
 >
 >     On my machine, the movaps address is consistently not on a 16-byte
 >     boundary like it should be.
 >
 > >How-To-Repeat:
 >
 >     This simple program exhibits the problem on my machine:
 >
 >     #include <xmmintrin.h>
 >     int main(int argc, char** argv) {
 >         __m128 x;
 >         float a[4] = {1.0f, 1.0f, 1.0f, 1.0f};
 >         x = _mm_loadu_ps(a);
 >         return 0;
 >     }
 >
 >     Compiled with "gcc -g -O0 -msse foo.c -o foo".
 >
 > >Fix:
 >
 >     Compiling with -O1 or higher makes the problem go away, at least in the
 >     simple example above.
 >
 > >Release-Note:
 > >Audit-Trail:
 > >Unformatted:
 According to my understanding, alignments aren't supported by gcc in main(), 
 as the OS doesn't necessarily present an aligned stack. The necessary 
 alignments are supported in functions which are called by a function built 
 with -mpreferred-stack-boundary=4.  The option -Os is intended to present a 
 lesser alignment; -O2 usually (but not always?) passes a fully aligned stack 
 to its callees.  I'm not sure of all this, and better documentation seems 
 needed.
 -- 
 Tim Prince

target/6624: SSE misalignment with -O0

[Jesse Hall]

>Number:         6624
>Category:       target
>Synopsis:       SSE mis-alignment with -O0
>Confidential:   no
>Severity:       serious
>Priority:       medium
>Responsible:    unassigned
>State:          open
>Class:          wrong-code
>Submitter-Id:   net
>Arrival-Date:   Fri May 10 09:16:01 PDT 2002
>Closed-Date:
>Last-Modified:
>Originator:     Jesse Hall
>Release:        3.1 20020426 (prerelease)
>Organization:
>Environment:
System: Linux espresso 2.4.18-xfs #1 SMP Tue Mar 12 20:03:55 CST 2002 i686 unknown
Architecture: i686 (more precisely, dual AthlonMP)

        
host: i386-pc-linux-gnu
build: i386-pc-linux-gnu
target: i386-pc-linux-gnu
configured with: ../src/configure -v --enable-languages=c,c++,java,f77,proto,objc --prefix=/usr/lib/gcc-snapshot --infodir=/share/info --mandir=/share/man --enable-shared --with-gnu-as --with-gnu-ld --with-system-zlib --enable-long-long --enable-nls --without-included-gettext --disable-checking --enable-threads=posix --enable-java-gc=boehm --with-cpp-install-dir=bin --enable-objc-gc i386-linux
>Description:
    The SSE intrinsic _mm_loadu_ps can do an aligned SSE load (LOADAPS) from
    an unaligned address when compiling with -O0. The problem seems to go away
    with -O1 and higher. This causes a segfault when the program is run. The
    assembly generated looks like:

        call   804834a <_mm_loadu_ps>
        add    $0x10,%esp
        movaps %xmm0,0xffffffe8(%ebp)

    On my machine, the movaps address is consistently not on a 16-byte
    boundary like it should be.

>How-To-Repeat:
    This simple program exhibits the problem on my machine:

    #include <xmmintrin.h>
    int main(int argc, char** argv) {
        __m128 x;
        float a[4] = {1.0f, 1.0f, 1.0f, 1.0f};
        x = _mm_loadu_ps(a);
        return 0;
    }

    Compiled with "gcc -g -O0 -msse foo.c -o foo".
    
>Fix:
    Compiling with -O1 or higher makes the problem go away, at least in the
    simple example above.
>Release-Note:
>Audit-Trail:
>Unformatted: