[Bug target/64580] New: very high rs6000_stack_info() usage during LTO Firefox build on ppc64

[Bug target/64580] New: very high rs6000_stack_info() usage during LTO Firefox build on ppc64

[trippels at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=64580

            Bug ID: 64580
           Summary: very high rs6000_stack_info() usage during LTO Firefox
                    build on ppc64
           Product: gcc
           Version: 5.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: target
          Assignee: unassigned at gcc dot gnu.org
          Reporter: trippels at gcc dot gnu.org
                CC: segher at gcc dot gnu.org
              Host: powerpc64-unknown-linux-gnu
            Target: powerpc64-unknown-linux-gnu
             Build: powerpc64-unknown-linux-gnu

"perf top" shows very high rs6000_stack_info() usage during final LTO link of
libxul (Firefox) on ppc64, e.g.:

  31.34%  lto1  [.]  rs6000_stack_info()
                 rs6000_stack_info()
                 rs6000_stack_info()
                 rs6000_output_function_prologue(_IO_FILE*, long)
                 final_start_function(rtx_insn*, _IO_FILE*, int)
                 rs6000_output_mi_thunk(_IO_FILE*, tree_node*, long, long,
tree_node*)
                 cgraph_node::expand_thunk(bool, bool)
                 cgraph_node::assemble_thunks_and_aliases()
                 cgraph_node::assemble_thunks_and_aliases()
                 cgraph_node::expand()
                 symbol_table::compile()
                 lto_main()
                 compile_file()
  1.36%  cc1plus [.]  _cpp_lex_token
...

powerpc64-unknown-linux-gnu

[Bug target/64580] very high rs6000_stack_info() usage during LTO Firefox build on ppc64

[segher at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=64580

--- Comment #1 from Segher Boessenkool <segher at gcc dot gnu.org> ---
Hi Markus,

How often is rs6000_stack_info called there?  Are there any hotspots
in the function?

Do you have a standalone testcase?

[Bug target/64580] very high rs6000_stack_info() usage during LTO Firefox build on ppc64

[trippels at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=64580

--- Comment #2 from Markus Trippelsdorf <trippels at gcc dot gnu.org> ---
Hi Segher,

on gcc112 you can use the following command to reproduce the issue:

 % g++ -xlto -c -mcpu=power8 -O3 -fPIC -fno-exceptions -fltrans -o /dev/null
/var/tmp/libxul.so.ltrans8.o

[Bug target/64580] very high rs6000_stack_info() usage during LTO Firefox build on ppc64

[segher at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=64580

--- Comment #3 from Segher Boessenkool <segher at gcc dot gnu.org> ---
Hi again,

Unfortunately it seems I am unable to build a version of GCC that is
compatible with that (every version I tried either complains or ICEs).

Maybe you can answer the above questions yourself?  How often
is rs6000_output_mi_thunk called (+ some sense of scale, e.g.
how many "real" functions are emitted)?  Is there a real hotspot
in rs6000_stack_info itself?  I haven't been able to get it above
0.16% of cpu time, and no hotspots.

Can perf be used to analyze results on a different machine than
used to generate the data?

[Bug target/64580] very high rs6000_stack_info() usage during LTO Firefox build on ppc64

[trippels at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=64580

--- Comment #4 from Markus Trippelsdorf <trippels at gcc dot gnu.org> ---
(In reply to Segher Boessenkool from comment #3)
> Hi again,
> 
> Unfortunately it seems I am unable to build a version of GCC that is
> compatible with that (every version I tried either complains or ICEs).

Yes, it's unfortunate that there is no stable interface at all. Every
patch in this area potentially breaks things and there were a couple of
them in the last few days.

> Maybe you can answer the above questions yourself?  How often
> is rs6000_output_mi_thunk called (+ some sense of scale, e.g.
> how many "real" functions are emitted)?  Is there a real hotspot
> in rs6000_stack_info itself?  I haven't been able to get it above
> 0.16% of cpu time, and no hotspots.

I will try to collect some callgrind data later today.

> Can perf be used to analyze results on a different machine than
> used to generate the data?

If the version of perf is the same then yes.

[Bug target/64580] very high rs6000_stack_info() usage during LTO Firefox build on ppc64

[trippels at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=64580

--- Comment #5 from Markus Trippelsdorf <trippels at gcc dot gnu.org> ---
Here is an example "callgrind_control -e -b" output (program is still running):

PID 53944:
/home/trippels/gcc_test/usr/local/bin/../libexec/gcc/powerpc64-unknown-linux-gnu/5.0.0/lto1
-quiet -dumpbase libxul.so.ltrans8.o -mcpu=power8 -auxbase-strip /dev/null -O3
-version -fPIC -fno-exceptions -fltrans /var/tmp/libxul.so.ltrans8.o -o
/home/trippels/tmp/ccxhYq0O.s
sending command status internal to pid 53944

  Totals:             Ir 
   Th 1  163,422,428,166 

  Frame:              Ir Backtrace for Thread 1
   [ 0]  127,387,164,848 rs6000_stack_info() (932 x)
   [ 1]  127,387,195,580 rs6000_output_function_prologue(_IO_FILE*, long) (932
x)
   [ 2]  127,387,263,387 final_start_function(rtx_insn*, _IO_FILE*, int) (132
x)
   [ 3]  127,390,072,883 rs6000_output_mi_thunk(_IO_FILE*, tree_node*, long,
long, tree_node*) (132 x)
   [ 4]  127,490,182,450 cgraph_node::expand_thunk(bool, bool) (132 x)
   [ 5]  127,490,272,999 cgraph_node::assemble_thunks_and_aliases() (190 x)
   [ 6]  127,491,063,296 cgraph_node::assemble_thunks_and_aliases() (800 x)
   [ 7]  156,624,357,942 cgraph_node::expand() (800 x)
   [ 8]  157,695,780,192 symbol_table::compile() [clone .part.53] (1 x)
   [ 9]  157,695,780,213 symbol_table::compile() (1 x)
   [10]  163,407,638,386 lto_main() (1 x)
   [11]  163,407,638,419 compile_file() (1 x)
   [12]  163,421,626,146 toplev::main(int, char**) (1 x)
   [13]  163,421,626,164 main (1 x)
   [14]  163,421,630,096 generic_start_main.isra.0 (1 x)
   [15]  163,421,630,244 (below main) (1 x)
   [16]  163,421,632,780 0x0000000010f23930 (1 x)
   [17]  163,421,632,786 0x000000001012f5a0 (1 x)
   [18]  163,421,632,794 _start (1 x)
   [19]  163,421,634,426 _dl_start_user (1 x)
   [20]                . _start

[Bug target/64580] very high rs6000_stack_info() usage during LTO Firefox build on ppc64

[dje at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=64580

David Edelsohn <dje at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |NEW
   Last reconfirmed|                            |2015-01-22
                 CC|                            |dje at gcc dot gnu.org
     Ever confirmed|0                           |1

--- Comment #7 from David Edelsohn <dje at gcc dot gnu.org> ---
Confirmed.

[Bug target/64580] very high rs6000_stack_info() usage during LTO Firefox build on ppc64

[trippels at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=64580

--- Comment #8 from Markus Trippelsdorf <trippels at gcc dot gnu.org> ---
The issue is also reproducible with an --enable-checking=release compiler.

The following command reproduces the issue using r220030 on gcc110:

/home/trippels/gcc_5/usr/local/bin/../libexec/gcc/powerpc64-unknown-linux-gnu/5.0.0/lto1
-quiet -dumpbase libxul.so.ltrans7.o -mcpu=power7 -auxbase-strip /dev/null -O3
-version -fPIC -fno-exceptions -fltrans /var/tmp/libxul.so.ltrans7.o -o
/dev/null

[Bug target/64580] very high rs6000_stack_info() usage during LTO Firefox build on ppc64

[trippels at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=64580

--- Comment #9 from Markus Trippelsdorf <trippels at gcc dot gnu.org> ---
(gdb) bt
#0  0x0000000010a64660 in compute_vrsave_mask () at
../../gcc/gcc/config/rs6000/rs6000.c:21149
#1  rs6000_stack_info () at ../../gcc/gcc/config/rs6000/rs6000.c:21686
#2  0x0000000010a65694 in rs6000_output_function_prologue (file=0x114a5800,
size=<optimized out>) at ../../gcc/gcc/config/rs6000/rs6000.c:24284
#3  0x0000000010320300 in final_start_function (first=<optimized out>,
file=0x114a5800, optimize_p=<optimized out>) at ../../gcc/gcc/final.c:1881
#4  0x0000000010a3cc88 in rs6000_output_mi_thunk (file=0x114a5800,
thunk_fndecl=<optimized out>, delta=<optimized out>, vcall_offset=<optimized
out>, 
    function=<optimized out>) at ../../gcc/gcc/config/rs6000/rs6000.c:25807
#5  0x000000001020f5b4 in cgraph_node::expand_thunk (this=this@entry=0xcbb3390,
output_asm_thunks=output_asm_thunks@entry=true, 
    force_gimple_thunk=force_gimple_thunk@entry=false) at
../../gcc/gcc/cgraphunit.c:1528
#6  0x0000000010210814 in cgraph_node::assemble_thunks_and_aliases
(this=this@entry=0xcbb3200) at ../../gcc/gcc/cgraphunit.c:1742
#7  0x000000001021079c in cgraph_node::assemble_thunks_and_aliases
(this=this@entry=0xcbb23f0) at ../../gcc/gcc/cgraphunit.c:1758
#8  0x0000000010210b04 in cgraph_node::expand (this=this@entry=0xcbb23f0) at
../../gcc/gcc/cgraphunit.c:1867
#9  0x00000000102128d8 in expand_all_functions () at
../../gcc/gcc/cgraphunit.c:1940
#10 symbol_table::compile (this=0xa8100a8) at ../../gcc/gcc/cgraphunit.c:2293
#11 0x000000001015b9a0 in lto_main () at ../../gcc/gcc/lto/lto.c:3457
#12 0x00000000106bc94c in compile_file () at ../../gcc/gcc/toplev.c:594
#13 0x00000000101262a8 in do_compile () at ../../gcc/gcc/toplev.c:2047
#14 toplev::main (this=<optimized out>, argc=15, argv=0xfff00ead8) at
../../gcc/gcc/toplev.c:2144
#15 0x0000000010126c7c in main (argc=<optimized out>, argv=0xfff00ead8) at
../../gcc/gcc/main.c:38

[Bug target/64580] very high rs6000_stack_info() usage during LTO Firefox build on ppc64

[trippels at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=64580

--- Comment #10 from Markus Trippelsdorf <trippels at gcc dot gnu.org> ---
Created attachment 34552
  --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=34552&action=edit
callgrind_annotate output

Output of "callgrind_annotate --tree=both callgrind.out.47690 >| out" is
attached.

[Bug target/64580] very high rs6000_stack_info() usage during LTO Firefox build on ppc64

[segher at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=64580

--- Comment #11 from Segher Boessenkool <segher at gcc dot gnu.org> ---
Author: segher
Date: Fri Jan 30 05:35:52 2015
New Revision: 220272

URL: https://gcc.gnu.org/viewcvs?rev=220272&root=gcc&view=rev
Log:
    PR target/64580
    * config.rs6000/rs6000.c (compute_vrsave_mask): Reverse loop order.
    (rs6000_stack_info): Add assert.
    (rs6000_output_savres_externs): New function, split off from...
    (rs6000_output_function_prologue): ... here.  Do not call it for
    thunks.

Modified:
    trunk/gcc/ChangeLog
    trunk/gcc/config/rs6000/rs6000.c

[Bug target/64580] very high rs6000_stack_info() usage during LTO Firefox build on ppc64

[trippels at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=64580

--- Comment #12 from Markus Trippelsdorf <trippels at gcc dot gnu.org> ---
Fixed for gcc-5. Many thanks.
I'll leave this bug open for possible backports.

[Bug target/64580] very high rs6000_stack_info() usage during LTO Firefox build on ppc64

[segher at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=64580

--- Comment #13 from Segher Boessenkool <segher at gcc dot gnu.org> ---
Author: segher
Date: Thu Feb  5 15:01:49 2015
New Revision: 220453

URL: https://gcc.gnu.org/viewcvs?rev=220453&root=gcc&view=rev
Log:
    PR target/64580
    Backport from mainline
    * config.rs6000/rs6000.c (compute_vrsave_mask): Reverse loop order.
    (rs6000_stack_info): Add assert.
    (rs6000_output_savres_externs): New function, split off from...
    (rs6000_output_function_prologue): ... here.  Do not call it for
    thunks.

Modified:
    branches/gcc-4_9-branch/gcc/ChangeLog
    branches/gcc-4_9-branch/gcc/config/rs6000/rs6000.c

[Bug target/64580] very high rs6000_stack_info() usage during LTO Firefox build on ppc64

[segher at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=64580

--- Comment #14 from Segher Boessenkool <segher at gcc dot gnu.org> ---
Author: segher
Date: Thu Feb  5 15:06:40 2015
New Revision: 220454

URL: https://gcc.gnu.org/viewcvs?rev=220454&root=gcc&view=rev
Log:
    PR target/64580
    Backport from mainline
    * config.rs6000/rs6000.c (compute_vrsave_mask): Reverse loop order.
    (rs6000_stack_info): Add assert.
    (rs6000_output_savres_externs): New function, split off from...
    (rs6000_output_function_prologue): ... here.  Do not call it for
    thunks.

Modified:
    branches/gcc-4_8-branch/gcc/ChangeLog
    branches/gcc-4_8-branch/gcc/config/rs6000/rs6000.c