This is the mail archive of the
gcc@gcc.gnu.org
mailing list for the GCC project.
Re: performance regression on ia64
There is a regression on ia32, too. 70% from Jun04 to Jul21.
optimized tree dumps of the ia64 case don't show something obvious - but
there's a lot of noise due to differing numbers. Anyone with a tool for a
more meaningful diff?
Richard.
On Wed, 21 Jul 2004, Richard Guenther wrote:
> The (non-leafified) tramp3d-v3 testcase got a runtime performance hit of
> about 15% between Jul16 and Jul18 (unfortunately bootstrap failed on
> Jul16). Last patch that could have caused this is
>
> gcc/ChangeLog
> 2004-07-17 Eric Botcazou <ebotcazou@act-europe.fr>
>
> * libgcc2.c (__enable_execute_stack): New symbol.
> * libgcc-std.ver (GCC_3.4.2): New version. Inherit from GCC_3.4
> and declare __enable_execute_stack.
> * mklibgcc.in (lib2funcs): Add _enable_execute_stack.
> * config/netbsd.h (NETBSD_ENABLE_EXECUTE_STACK): ANSIfy.
> * config/sol2.h (TRANSFER_FROM_TRAMPOLINE): Rename into
> ENABLE_EXECUTE_STACK.
> * config/alpha/alpha.c (alpha_initialize_trampoline): Conditionalize
> [...]
> gcc/cp/ChangeLog
> 2004-07-17 Joseph S. Myers <jsm@polyomino.org.uk>
>
> * cp-tree.h (builtin_function): Declare.
>
> and first one to maybe blame would have been
>
> gcc/ChangeLog
> 2004-07-15 Nathan Sidwell <nathan@codesourcery.com>
>
> * vec.h (VEC_address): New function.
> gcc/cp/ChangeLog
> 2004-07-15 Mark Mitchell <mark@codesourcery.com>
>
> * cp-tree.h (lang_type_class): Remove has_real_assign_ref and
> has_abstract_assign_ref. Make methods a VEC(tree) *.
> (TYPE_HAS_CONST_ASSIGN_REF): Add documentation.
> (CLASSTYPE_CONSTRUCTORS): Adjust for changes to
> CLASSTYPE_METHOD_VEC.
> (CLASSTYPE_DESTRUCTORS): Likewise.
> (TYPE_HAS_REAL_ASSIGN_REF): Remove.
> (TYPE_HAS_ABSTRACT_ASSIGN_REF): Likewise.
> (add_method): Change prototoype.
> [...]
>
> I don't know if cp/ patches can cause runtime-regressions (I suspect not).
> I still have both compilers before and after the regression available (and
> I'm just building an ia32 one, just to check if the regression is
> architecture specific).
>
> Richard.
>
> --
> Richard Guenther <richard dot guenther at uni-tuebingen dot de>
> WWW: http://www.tat.physik.uni-tuebingen.de/~rguenth/
>
--
Richard Guenther <richard dot guenther at uni-tuebingen dot de>
WWW: http://www.tat.physik.uni-tuebingen.de/~rguenth/