This is the mail archive of the
mailing list for the GCC project.
Re: [PATCH] Redesign jump threading profile updates
- From: Christophe Lyon <christophe dot lyon at linaro dot org>
- To: Teresa Johnson <tejohnson at google dot com>
- Cc: Jeff Law <law at redhat dot com>, "gcc-patches at gcc dot gnu dot org" <gcc-patches at gcc dot gnu dot org>, Jan Hubicka <hubicka at ucw dot cz>, David Li <davidxl at google dot com>
- Date: Wed, 1 Oct 2014 09:03:02 +0200
- Subject: Re: [PATCH] Redesign jump threading profile updates
- Authentication-results: sourceware.org; auth=none
- References: <CAAe5K+XMKuBvrA2zbSdA38nq+KiezseL6z_4KayvV322VmJtZQ at mail dot gmail dot com> <53CF1DFD dot 7080805 at redhat dot com> <CAAe5K+UM+gXeSuunYbMz-Ln14kbb5-v0=Ysox1CUQ2esrfwU-A at mail dot gmail dot com> <CAAe5K+UFGUo2YbNXUPnkTh0P7L+UrsKRTLFJdWggQ3i_y0cyQw at mail dot gmail dot com> <CAAe5K+VD=aWnLq3XAUBbRqk9SR1MHiq-VSyd4sxgNmcs8_E2MA at mail dot gmail dot com> <542A32AB dot 1040708 at redhat dot com> <CAAe5K+XX+QVNgW0iY3EfF26nC02Z2_OLMXyaseB7jfbNJ=dFUA at mail dot gmail dot com>
On 30 September 2014 20:20, Teresa Johnson <firstname.lastname@example.org> wrote:
> On Mon, Sep 29, 2014 at 9:33 PM, Jeff Law <email@example.com> wrote:
>> On 09/29/14 08:19, Teresa Johnson wrote:
>>>> Just an update - I found some good test cases by compiling the
>>>> c-torture tests with profile feedback with and without my patch. But
>>>> in the cases I pulled out I saw that there were still a couple profile
>>>> or probability insanities introduced by jump threading (albeit far
>>>> less than before), so I wanted to investigate before I commit. I ran
>>>> out of time this week and will not get to this until I get back from
>>>> vacation the week after next.
>>> Hi Jeff,
>>> I finally had a chance to get back to this and look at the remaining
>>> insanities in the new test cases I created. It turns out that there
>>> were still a few issues in the case where there were guessed
>>> frequencies and no profile counts. The two test cases I created do use
>>> FDO, and the insanities in the routines with profile counts went away
>>> with my patch. But the outlined copies of routines that were also
>>> inlined into the main routine still had estimated frequencies, and
>>> these still had a few issues.
>>> The problem is that the profile updates are done incrementally as we
>>> walk and update the paths in ssa_fix_duplicate_block_edges, including
>>> the block and edge counts, the block frequencies and the
>>> probabilities. This is very difficult to do when only operating on
>>> frequencies since the edge frequencies are derived from the source
>>> block frequency and the probability. Therefore, once the source block
>>> frequency is updated, the edge frequency is also affected, and it is
>>> really difficult to figure out what the update to the edge frequency
>>> (essentially the probability) is using the same incremental update
>>> approach. I was attempting to handle this with the routine
>>> deduce_freq, for example, but this turned out to have issues for
>>> certain types of paths. I tried a few other approaches, but they start
>>> looking really ugly and I didn't want to add a parallel but different
>>> algorithm in the case of no profile counts.
>>> So by far the simplest approach was simply to take a snapshot of the
>>> existing block and edge frequencies along the path before we start the
>>> updates in ssa_fix_duplicate_block_edges, by copying them into the
>>> profile count fields of those blocks and edges. Then the existing
>>> algorithm operates the same as when we do have counts, and can
>>> essentially operate incrementally on the edge frequencies since they
>>> live in the count field of the edge and are no longer affected anytime
>>> the source block is updated. Since the algorithm does update block
>>> frequencies and probabilities as well (based on the count updates
>>> performed), we can simply clear out these fake count fields at the end
>>> of ssa_fix_duplicate_block_edges. This takes care of the remaining
>>> insanities introduced by jump threading from these test cases. During
>>> testing I also added in some checking to ensure that the count fields
>>> for the whole routine were cleared properly to make sure the new
>>> clear_counts_path was not missing anything (checking is a little too
>>> heavyweight to add in normally).
>>> New patch below (also attached since my mailer sometimes eats spaces).
>>> The differences between the old patch and the new one:
>>> - removed deduce_freq (which was my least favorite part of the patch
>>> anyway!), and its call from recompute_probabilities, since it is no
>>> longer necessary.
>>> - two new routines freqs_to_counts_path and clear_counts_path, invoked
>>> from ssa_fix_duplicate_block_edges.
>>> - two new tests
>>> Bootstrapped and tested on x86_64-unknown-linux-gnu, ok for trunk?
>>> 2014-09-29 Teresa Johnson <firstname.lastname@example.org>
>>> * tree-ssa-threadupdate.c (struct ssa_local_info_t): New
>>> duplicate_blocks bitmap.
>>> (remove_ctrl_stmt_and_useless_edges): Ditto.
>>> (create_block_for_threading): Ditto.
>>> (compute_path_counts): New function.
>>> (update_profile): Ditto.
>>> (recompute_probabilities): Ditto.
>>> (update_joiner_offpath_counts): Ditto.
>>> (freqs_to_counts_path): Ditto.
>>> (clear_counts_path): Ditto.
>>> (ssa_fix_duplicate_block_edges): Update profile info.
>>> (ssa_create_duplicates): Pass new parameter.
>>> (ssa_redirect_edges): Remove old profile update.
>>> (thread_block_1): New duplicate_blocks bitmap,
>>> remove old profile update.
>>> (thread_single_edge): Pass new parameter.
>>> 2014-09-29 Teresa Johnson <email@example.com>
>>> * testsuite/gcc.dg/tree-prof/20050826-2.c: New test.
>>> * testsuite/gcc.dg/tree-prof/cmpsf-1.c: Ditto.
>> Given I'd already been through this pretty thoroughly, I just gave this a
>> cursory review.
>> clear_counts_path needs a function comment. It's pretty obvious what it's
>> doing, but for completeness let's go ahead and get the obvious comment in
> Done and committed as r215739.
Since this commit, I can see all my builds for arm*linux* and
aarch64*linux* fail while building glibc:
iso-2022-cn.c -c -std=gnu99 -fgnu89-inline -O2 -Wall -Win
line -Wundef -Wwrite-strings -fmerge-all-constants -frounding-math -g
-Wstrict-prototypes -fPIC -I../include
u -I../sysdeps/unix/inet -I../nptl/sysdeps/unix/sysv
-I../sysdeps/unix/sysv -I../nptl/sysdeps/unix -I../sysdeps/unix
eps/aarch64/fpu -I../sysdeps/aarch64/nptl -I../sysdeps/aarch64
ordsize-64 -I../sysdeps/ieee754/dbl-64 -I../sysdeps/ieee754/flt-32
l -I.. -I../libio -I. -nostdinc -isystem
-D_LIBC_REENTRANT -include ../include/libc-symbols.h -DPIC -DSHARED
-MD -MP -MF /tmp/3
In file included from iso-2022-cn.c:407:0:
../iconv/skeleton.c: In function 'gconv':
../iconv/skeleton.c:800:1: internal compiler error: in
check_probability, at basic-block.h:959
Please submit a full bug report,
with preprocessed source if appropriate.
Please include the complete backtrace with any bug report.
See <http://gcc.gnu.org/bugs.html> for instructions.
Can you have a look?
>> With that fix, approved for the trunk. Thanks for taking the time to sort
>> out all these issues.
> Teresa Johnson | Software Engineer | firstname.lastname@example.org | 408-460-2413