IRC session, #gcc on OFTC

gooberirc.log (text/plain), 2.43 KB, created by Eric Shattow on 2004-10-13 20:08:35 UTC

<Farfetch> pinskia
<Farfetch>     25da: 31 f6                 xor    %esi,%esi
<Farfetch>     25dc: 80 ba 29 01 00 00 00  cmpb   $0x0,0x129(%edx)
<Farfetch>     25e3: 0f 85 1f 02 00 00     jne    2808 <_ZN8OscilGen7prepareEv+0x388>
<Farfetch>     25e9: 8b 82 88 05 00 00     mov    0x588(%edx),%eax
<Farfetch>     25ef: 31 ff                 xor    %edi,%edi
<Farfetch>     25f1: f3 0f 10 05 00 00 00  movss  0x0,%xmm0
<Farfetch>     25f8: 00 
<Farfetch>     25f9: 0f 29 45 88           movaps %xmm0,0xffffff88(%ebp)
<Farfetch>     25fd: 89 45 84              mov    %eax,0xffffff84(%ebp)
<Farfetch>     2600: 89 85 7c ff ff ff     mov    %eax,0xffffff7c(%ebp)
<Farfetch>     2606: 8d 77 01              lea    0x1(%edi),%esi
<Farfetch>     2609: 8b 55 e0              mov    0xffffffe0(%ebp),%edx
<Farfetch>     260c: 0f 57 c0              xorps  %xmm0,%xmm0
<Farfetch> unaligned stack access?
<pinskia> movaps is the problem
<Farfetch> yup
<Farfetch> malloc()
<pinskia> there is already a bug about it
<Farfetch> glibc alignment

...

<pinskia> yes but we do it for 16byte
<Farfetch> jnc: a fix could be wrapping the global ::operator new and replace it with one which allocates more memory and align it
<Farfetch> jnc: there are glibc functions like memalignm
<Farfetch> jnc: or you can disable SSE stuff, but i'm not sure you want that

...

<pinskia> Farfetch, tm2: http://gcc.gnu.org/bugzilla/show_bug.cgi?id=15795
<gccbot> gcc bug #15795: No way to teach operator new anything about alignment requirements Product: gcc, Component: c++, Severity: enhancement, Assigned to: [e-mail], Status: RESOLVED, Resolution: WONTFIX
<pinskia> tm2: C99 7.20.3
<pinskia> so it is a glibc bug
<Farfetch> pinskia, hmm but this time the array is of floats
<Farfetch> pinskia, while the bug is about vector intrinsics

...

<Farfetch> uhm but there is a thing I still don't understand
<tm2> the embedded x86 people will have a herd of cows if you change alignment to 128
<Farfetch> the PR is about a __m128 being 16byte aligned
<Farfetch> which I can understand being a problem
<Farfetch> but in the code I pasted, I'm accessing array of floats
<Farfetch> can we demand such a strictier alignement to those arrays?
<pinskia> I don't know
<Farfetch> I don't think we can
<Farfetch> also movaps is definitely into the stack
<pinskia> xmm0,0xffffff88 is not aligned
<pinskia> there is a bug about it somewhere

...
---privmsg to jnc

<Farfetch> jnc, I think you can try filing a GCC bugreport