Is profiledbootstrap supposed to work currently?

Is profiledbootstrap supposed to work currently?

[Richard Guenther]

Trying profiledbootstrap today on ia32-linux leads to

stage1/xgcc -Bstage1/
-B/home/rguenth/ix86/gcc3.5-051004/i686-pc-linux-gnu/bin/ -c   -O2 -g
-fomit-frame-pointer -fprofile-use -DIN_GCC   -W -Wall -Wwrite-strings
-Wstrict-prototypes -Wmissing-prototypes -pedantic -Wno-long-long
-Wno-variadic-macros -Wold-style-definition -Werror    -DHAVE_CONFIG_H
-I. -I. -I/home/rguenth/src/gcc/gcc3.5/gcc
-I/home/rguenth/src/gcc/gcc3.5/gcc/.
-I/home/rguenth/src/gcc/gcc3.5/gcc/../include
-I/home/rguenth/src/gcc/gcc3.5/gcc/../libcpp/include
/home/rguenth/src/gcc/gcc3.5/gcc/attribs.c -o attribs.o
/home/rguenth/src/gcc/gcc3.5/gcc/attribs.c: In function 'decl_attributes':
/home/rguenth/src/gcc/gcc3.5/gcc/attribs.c:337: error: coverage mismatch
for function 'decl_attributes' while reading counter 'arcs'.
/home/rguenth/src/gcc/gcc3.5/gcc/attribs.c:337: error: number of counters
is 67 instead of 69
make[2]: *** [attribs.o] Error 1
make[2]: Leaving directory `/tmp/gcc-obj/gcc'
make[1]: *** [stagefeedback_build] Error 2
make[1]: Leaving directory `/tmp/gcc-obj/gcc'
make: *** [profiledbootstrap] Error 2

is this supposed to happen at the moment, i.e. are there pending fixes, or
is this a genuine bug?

Thanks,
Richard.

--
Richard Guenther <richard dot guenther at uni-tuebingen dot de>
WWW: http://www.tat.physik.uni-tuebingen.de/~rguenth/

Re: Is profiledbootstrap supposed to work currently?

[Andrew Pinski]

On Oct 5, 2004, at 10:22 AM, Richard Guenther wrote:

> Trying profiledbootstrap today on ia32-linux leads to
>
> is this supposed to happen at the moment, i.e. are there pending 
> fixes, or
> is this a genuine bug?

This was caused by the fact -funroll-loops is supplied only when
-fprofile-use is supplied and some unrolling is done on the tree level
but profiling is done on the RTL level so there is a mismatch.

This is PR 14684.

A workaround is to change -fprofile-generate to imply -funroll-loops 
too.

The correct fix is most likely not going to happen fully until 4.1 with
tree based profiling.

Thanks,
Andrew Pinski

Re: Is profiledbootstrap supposed to work currently? [PR bootstrap/14684]

[Kelley Cook]

[-- Attachment #1: Type: text/plain, Size: 1471 bytes --]

On Oct 5, 2004, at 10:22 AM, Richard Guenther wrote:

    Trying profiledbootstrap today on ia32-linux leads to

    is this supposed to happen at the moment, i.e. are there pending
    fixes, or
    is this a genuine bug? 

> This was caused by the fact -funroll-loops is supplied only when
> -fprofile-use is supplied and some unrolling is done on the tree level
> but profiling is done on the RTL level so there is a mismatch.

 >
 > This is PR 14684.
 >
 > A workaround is to change -fprofile-generate to imply -funroll-loops too.

I've attached the seemingly obvious cut-and-paste patch which causes 
-fprofile-generate to match the same options implied by -fprofile-use.

It does survive make bootstrap.  Unfortunately, it leads to a new 
problem when attempting a make profiledbootstrap:

stage1/xgcc.exe -Bstage1/ -B/usr/local/i686-pc-cygwin/bin/ -c   -g -O2 
-fprofile-generate -DIN_GCC   -W -Wall -Wwrite-strings 
-Wstrict-prototypes -Wmissing-prototypes -fno-common   -DHAVE_CONFIG_H 
-DGENERATOR_FILE    -I. -Ibuild -I../../gcc-snapshot/gcc 
-I../../gcc-snapshot/gcc/build -I../../gcc-snapshot/gcc/../include 
-I../../gcc-snapshot/gcc/../libcpp/include -o build/gengtype.o 
../../gcc-snapshot/gcc/gengtype.c
../../gcc-snapshot/gcc/gengtype.c: In function 'get_base_file_bitmap':
../../gcc-snapshot/gcc/gengtype.c:1194: internal compiler error: 
Segmentation fault

I have a feeling that this patch is nevertheless correct and should be 
applied.

Kelley Cook

[-- Attachment #2: pb.patch --]
[-- Type: text/plain, Size: 955 bytes --]

2004-10-05  Kelley Cook  <kcook@gcc.gnu.org>

	PR middle-end/14684
	* opts.c (OPT_fprofile_generate): Default to -funroll-loops,
	-fpeel-loops and -ftracer to match -fprofile-use.

--- gcc-orig/gcc/opts.c	2004-09-10 07:02:25.000000000 -0400
+++ gcc-snapshot/gcc/opts.c	2004-10-05 13:04:09.383317700 -0400
@@ -850,10 +850,16 @@ common_handle_option (size_t scode, cons
     case OPT_fprofile_generate:
       if (!profile_arc_flag_set)
         profile_arc_flag = value;
       if (!flag_profile_values_set)
         flag_profile_values = value;
+      if (!flag_unroll_loops_set)
+        flag_unroll_loops = value;
+      if (!flag_peel_loops_set)
+        flag_peel_loops = value;
+      if (!flag_tracer_set)
+        flag_tracer = value;
       if (!flag_value_profile_transformations_set)
         flag_value_profile_transformations = value;
 #ifdef HAVE_prefetch
       if (!flag_speculative_prefetching_set)
 	flag_speculative_prefetching = value;

Re: Is profiledbootstrap supposed to work currently? [PR bootstrap/14684]

[Jan Hubicka]

> On Oct 5, 2004, at 10:22 AM, Richard Guenther wrote:
> 
>    Trying profiledbootstrap today on ia32-linux leads to
> 
>    is this supposed to happen at the moment, i.e. are there pending
>    fixes, or
>    is this a genuine bug? 
> 
> >This was caused by the fact -funroll-loops is supplied only when
> >-fprofile-use is supplied and some unrolling is done on the tree level
> >but profiling is done on the RTL level so there is a mismatch.
> 
> >
> > This is PR 14684.
> >
> > A workaround is to change -fprofile-generate to imply -funroll-loops too.
> 
> I've attached the seemingly obvious cut-and-paste patch which causes 
> -fprofile-generate to match the same options implied by -fprofile-use.
> 
> It does survive make bootstrap.  Unfortunately, it leads to a new 
> problem when attempting a make profiledbootstrap:
> 
> stage1/xgcc.exe -Bstage1/ -B/usr/local/i686-pc-cygwin/bin/ -c   -g -O2 
> -fprofile-generate -DIN_GCC   -W -Wall -Wwrite-strings 
> -Wstrict-prototypes -Wmissing-prototypes -fno-common   -DHAVE_CONFIG_H 
> -DGENERATOR_FILE    -I. -Ibuild -I../../gcc-snapshot/gcc 
> -I../../gcc-snapshot/gcc/build -I../../gcc-snapshot/gcc/../include 
> -I../../gcc-snapshot/gcc/../libcpp/include -o build/gengtype.o 
> ../../gcc-snapshot/gcc/gengtype.c
> ../../gcc-snapshot/gcc/gengtype.c: In function 'get_base_file_bitmap':
> ../../gcc-snapshot/gcc/gengtype.c:1194: internal compiler error: 
> Segmentation fault
> 
> I have a feeling that this patch is nevertheless correct and should be 
> applied.

I would approve the patch with -funroll-loops only.  There is no need to
enable loop peeling and tracer when they are still done at RTL level.
Without profile feedback these can bring quite noticeable bloat that is
unnecesary at profiling stage.

Honza