[Bug sanitizer/63927] New: AddressSanitizer painfully slow on ppc64

[Bug sanitizer/63927] New: AddressSanitizer painfully slow on ppc64

[trippels at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=63927

            Bug ID: 63927
           Summary: AddressSanitizer painfully slow on ppc64
           Product: gcc
           Version: unknown
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: sanitizer
          Assignee: unassigned at gcc dot gnu.org
          Reporter: trippels at gcc dot gnu.org
                CC: dodji at gcc dot gnu.org, dvyukov at gcc dot gnu.org,
                    jakub at gcc dot gnu.org, kcc at gcc dot gnu.org

Created attachment 34013
  --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=34013&action=edit
testcase

With the attached testcase, that measures the performance of different
C++ containers, AddressSanitizer is very slow on ppc64. x86_64 is fine.

For example:

trippels@gcc2-power8 ~ % g++ -fsanitize=address -g -O2 bench.cpp
trippels@gcc2-power8 ~ % ./a.out
size            array   vector_point    vector_iters            deque          
list            set             multiset
10              15.79           13.85           13.14           26.04          
232.16          132.72          266.89
^C

On x86_64:

markus@x4 ~ % ./a.out
size            array   vector_point    vector_iters            deque          
list            set             multiset
10              0.46            0.46            0.46            1.39           
4.55            2.36            3.74
...

[Bug sanitizer/63927] AddressSanitizer painfully slow on ppc64

[trippels at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=63927

--- Comment #1 from Markus Trippelsdorf <trippels at gcc dot gnu.org> ---
The numbers above are seconds.

Perf shows:
  29.44%  a.out  libgcc_s.so.1       [.] uw_update_context_1
  13.08%  a.out  libasan.so.2.0.0    [.] __sanitizer::mem_is_zero(char const*,
unsigned long)
  11.40%  a.out  libgcc_s.so.1       [.] execute_cfa_program
   5.61%  a.out  libpthread-2.18.so  [.] pthread_mutex_unlock
   5.26%  a.out  libc-2.18.so        [.] memcpy
   4.90%  a.out  libgcc_s.so.1       [.] _Unwind_IteratePhdrCallback
   4.07%  a.out  libasan.so.2.0.0    [.] __asan_region_is_poisoned
   3.33%  a.out  libpthread-2.18.so  [.] pthread_mutex_lock
   2.51%  a.out  libgcc_s.so.1       [.] uw_frame_state_for
   2.36%  a.out  libc-2.18.so        [.] memset
   1.53%  a.out  libasan.so.2.0.0    [.] __interceptor_strlen
   1.51%  a.out  libgcc_s.so.1       [.] read_encoded_value_with_base
   1.38%  a.out  libasan.so.2.0.0    [.] __asan_memcpy
   0.95%  a.out  libasan.so.2.0.0    [.] __asan_memset

[Bug sanitizer/63927] AddressSanitizer painfully slow on ppc64

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=63927

--- Comment #2 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
PowerPC is the only target with an abi that can backtrace correctly without a
frame pointer so it should be easy to implement that.

Even x86 is broken backtracking without a frame pointer.

[Bug sanitizer/63927] AddressSanitizer painfully slow on ppc64

[jakub at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=63927

--- Comment #3 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
Right, but it is a library change, so somebody has to code it up, test and push
upstream first, then we can cherry-pick it.

[Bug sanitizer/63927] AddressSanitizer painfully slow on ppc64

[kcc at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=63927

--- Comment #4 from Kostya Serebryany <kcc at gcc dot gnu.org> ---
Why can't we use frame pointers on PPC? 
So far I have not seen any implementation of unwinder not based on FPs
that was fast enough for ASan. 
ASan unwinds the stack on every malloc/free and thus it is very sensitive to
the unwinder performance.

[Bug sanitizer/63927] AddressSanitizer painfully slow on ppc64

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=63927

--- Comment #5 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
(In reply to Kostya Serebryany from comment #4)
> Why can't we use frame pointers on PPC? 


You don't need to use the frame pointer.  The ABI says the stack frame always
contains a reference to the previous frame at a specific location (I think
sp+0). 

> So far I have not seen any implementation of unwinder not based on FPs
> that was fast enough for ASan. 
> ASan unwinds the stack on every malloc/free and thus it is very sensitive to
> the unwinder performance.

[Bug sanitizer/63927] AddressSanitizer painfully slow on ppc64

[trippels at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=63927

--- Comment #6 from Markus Trippelsdorf <trippels at gcc dot gnu.org> ---
See: http://refspecs.linuxfoundation.org/ELF/ppc64/PPC-elf64abi-1.9.html#STACK

[Bug sanitizer/63927] AddressSanitizer painfully slow on ppc64

[kcc at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=63927

Kostya Serebryany <kcc at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |eugeni.stepanov at gmail dot com,
                   |                            |samsonov at google dot com

--- Comment #7 from Kostya Serebryany <kcc at gcc dot gnu.org> ---
(In reply to Markus Trippelsdorf from comment #6)
> See:
> http://refspecs.linuxfoundation.org/ELF/ppc64/PPC-elf64abi-1.9.html#STACK

Then someone needs to implement the appropriate unwinder 
in sanitizer_common/sanitizer_stacktrace.cc
and submit the patch for review upstream. 
https://code.google.com/p/address-sanitizer/wiki/HowToContribute

[Bug sanitizer/63927] AddressSanitizer painfully slow on ppc64

[wschmidt at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=63927

Bill Schmidt <wschmidt at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |NEW
   Last reconfirmed|                            |2015-07-28
     Ever confirmed|0                           |1

--- Comment #8 from Bill Schmidt <wschmidt at gcc dot gnu.org> ---
Patch submitted as http://reviews.llvm.org/D11552.

[Bug sanitizer/63927] AddressSanitizer painfully slow on ppc64

[trippels at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=63927

--- Comment #9 from Markus Trippelsdorf <trippels at gcc dot gnu.org> ---
(In reply to Bill Schmidt from comment #8)
> Patch submitted as http://reviews.llvm.org/D11552.

Wow. Very nice speedup for such a simple patch.
Would be great if could be cherry-picked directly.

[Bug sanitizer/63927] AddressSanitizer painfully slow on ppc64

[wschmidt at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=63927

--- Comment #10 from Bill Schmidt <wschmidt at gcc dot gnu.org> ---
The fix was accepted and committed upstream in the LLVM compiler-rt project. 
Jakub, is applying this patch to GCC's libsanitizer ok?

[Bug sanitizer/63927] AddressSanitizer painfully slow on ppc64

[jakub at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=63927

--- Comment #11 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
(In reply to Bill Schmidt from comment #10)
> The fix was accepted and committed upstream in the LLVM compiler-rt project.
> Jakub, is applying this patch to GCC's libsanitizer ok?

After proper testing it is preapproved, provided you write ChangeLog entry for
it.

[Bug sanitizer/63927] AddressSanitizer painfully slow on ppc64

[wschmidt at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=63927

Bill Schmidt <wschmidt at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|NEW                         |RESOLVED
         Resolution|---                         |FIXED

--- Comment #13 from Bill Schmidt <wschmidt at gcc dot gnu.org> ---
Work is complete.

[Bug sanitizer/63927] AddressSanitizer painfully slow on ppc64

[trippels at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=63927

Markus Trippelsdorf <trippels at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|RESOLVED                    |REOPENED
         Resolution|FIXED                       |---

--- Comment #14 from Markus Trippelsdorf <trippels at gcc dot gnu.org> ---
Sorry but the libsanitizer/sanitizer_common/sanitizer_stacktrace.h hunk is
missing
from your committed patch.
So SANITIZER_CAN_FAST_UNWIND is still 0 and the new code in
sanitizer_stacktrace.c isn't used at all.

[Bug sanitizer/63927] AddressSanitizer painfully slow on ppc64

[trippels at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=63927

--- Comment #15 from Markus Trippelsdorf <trippels at gcc dot gnu.org> ---
Author: trippels
Date: Wed Jul 29 06:32:09 2015
New Revision: 226338

URL: https://gcc.gnu.org/viewcvs?rev=226338&root=gcc&view=rev
Log:
Use fast unwinder for PowerPC

2015-07-29  Markus Trippelsdorf  <markus@trippelsdorf.de>

        PR sanitizer/63927
        * sanitizer_common/sanitizer_stacktrace.h: Use fast unwinder
        for PowerPC.

Modified:
    trunk/libsanitizer/ChangeLog
    trunk/libsanitizer/sanitizer_common/sanitizer_stacktrace.h

[Bug sanitizer/63927] AddressSanitizer painfully slow on ppc64

[trippels at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=63927

Markus Trippelsdorf <trippels at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|REOPENED                    |RESOLVED
         Resolution|---                         |FIXED

--- Comment #16 from Markus Trippelsdorf <trippels at gcc dot gnu.org> ---
Fixed as obvious.

[Bug sanitizer/63927] AddressSanitizer painfully slow on ppc64

[wschmidt at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=63927

--- Comment #17 from Bill Schmidt <wschmidt at gcc dot gnu.org> ---
Argh, sorry, Markus.  Thanks for fixing it up.