On Tue, Dec 4, 2018 at 4:40 PM Bin.Cheng <[email protected]> wrote:
>
> On Thu, Nov 29, 2018 at 12:20 AM Jan Hubicka <[email protected]> wrote:
> >
> > > On Tue, Nov 20, 2018 at 6:55 PM bin.cheng <[email protected]>
> > > wrote:
> > > >
> > > > Sender:Jan Hubicka <[email protected]>
> > > > Sent at:2018 Nov 5 (Mon) 22:21
> > > > To:Richard Biener <[email protected]>
> > > > Cc:bin.cheng <[email protected]>; GCC Patches
> > > > <[email protected]>
> > > > Subject:Re: [PATCH AutoFDO/2]Treat ZERO as common profile
> > > > probability/count
> > > >
> > > > >
> > > > > > On Wed, Oct 31, 2018 at 7:30 AM bin.cheng
> > > > > > <[email protected]> wrote:
> > > > > > >
> > > > > > > Hi,
> > > > > > > In new profile probability/count infra, we have different
> > > > > > > precision quality categories,
> > > > > > > and probabilities/counts of different categories are not supposed
> > > > > > > to be compared or
> > > > > > > calculated. Though in general is an improvement, it introduces
> > > > > > > unexpected behavior.
> > > > > > > Specifically, class profile_probablity and profile_count
> > > > > > > themselves are implemented
> > > > > > > by comparing probabilities/counts against profile_count::zero().
> > > > > > > while zero() is of
> > > > > > > profile_precision category, it's always compared different to
> > > > > > > zero of other precision
> > > > > > > categories including afdo.
> > > > > > >
> > > > > > > I can see two ways fixing this: 1) Treat zero as a common
> > > > > > > probability/count regardless
> > > > > > > of its category; 2) Provide an "is_zero" method rather than
> > > > > > > relying on "==" comparison
> > > > > > > against probability_count::zero(). 2) requires lots of code
> > > > > > > changes so I went with 1)
> > > > > > > in this patch set. This patch doesn't handle "always" but it
> > > > > > > might be.
> > > > > > >
> > > > > > > This patch also corrects a minor issue where we try to invert an
> > > > > > > uninitialized value.
> > > > > > >
> > > > > > > Bootstrap and test on x86_64 in patch set. Is it OK?
> > > > > >
> > > > > > I'll defer on the emit_store_flag_force change, likewise for the
> > > > > > zero
> > > > > > handling in
> > > > > > compares - I don't think zeros of different qualities should
> > > > > > compare equal.
> > > > > > Would compares against ::always() not have the very same issue?
> > > > > > Likewise ::even(),
> > > > > > ::likely(), etc.? Those always get guessed quality.
> > > > > >
> > > > > > The invert change looks OK to me. The related change to the
> > > > > > always() API would
> > > > > > suggest to replace guessed_always() with always (guessed) and also
> > > > > > do similar
> > > > > > changes throughout the whole API...
> > > > > >
> > > > > > Honza?
> > > > >
> > > > > The zeros are really differenct zeros. profile_count::zero makes us
> > > > > to
> > > > > drop the basic block into cold section because we know that it won't
> > > > > be
> > > > > executed in normal run of program (either we have accurate profile
> > > > > feedback or by proving that the program is on way to crash or user
> > > > > annotated cold section). Having guessed zero or auto-fdo zero won't
> > > > > make us to do such agressive size optimization.
> > > > > This is important since those zeros relatively commonly happens by
> > > > > accident and thus if we dropped all the code to cold section the cold
> > > > > section would be visited relativel often during execution of program
> > > > > which would eliminate its need.
> > > > >
> > > > > Most comparsion in profile-count.h which goes agains
> > > > > profile_count==zero
> > > > > are realy intended to pass only for this "aboslute zero". They bypass
> > > > > the precision adjusmtents which normally happen when you merge values
> > > > > of different precision.
> > > > >
> > > > > What kind of unexpected behaviour are you seeing?
> > > > > We already have nonzero_p which is what we use when we want to know
> > > > > that
> > > > > count is non-zero in some sense of precision.
> > > > Hi Honza,
> > > > Sorry for letting this slip away. So in case of AutoFDO, due to the
> > > > nature
> > > > of sampling, lots of funcs/bbs are annotated with zero profile_count in
> > > > afdo
> > > > precision, and we have checks against zero profile_count in precise
> > > > precision
> > > > All these checks end up with false and cause issues. Take the code in
> > > > update_profiling_info as an example:
> > > >
> > > > update_profiling_info (struct cgraph_node *orig_node,
> > > > struct cgraph_node *new_node)
> > > > {
> > > > struct cgraph_edge *cs;
> > > > struct caller_statistics stats;
> > > > profile_count new_sum, orig_sum;
> > > > profile_count remainder, orig_node_count = orig_node->count;
> > > >
> > > > if (!(orig_node_count.ipa () > profile_count::zero ()))
> > > > return;
> > > > //...
> > > > for (cs = new_node->callees; cs; cs = cs->next_callee)
> > > > cs->count = cs->count.apply_scale (new_sum, orig_node_count);
> > > >
> > > > Since we also have below code in profile_count::operator>,
> > > > if (other == profile_count::zero ())
> > > > return !(*this == profile_count::zero ());
> > > >
> > > > If orig_node_count is afdo zero, the above zero check for
> > > > orig_node_count
> > > > returns false, we end up with passing zero density to apply_scale issue
> > > > and
> > > > asserting.
> > > >
> > > > In this updated patch, I restrcited changes only to
> > > > profile_count::operator
> > > > <, >, <= and >=. Plus, I think there is a latent typo in operator>=
> > > > because
> > > > current code return TRUE if '*this' is precise zero and 'other' is
> > > > precise
> > > > non-zero.
> > > > @@ -879,7 +879,7 @@ public:
> > > > if (other == profile_count::zero ())
> > > > return true;
> > > > if (*this == profile_count::zero ())
> > > > - return !(other == profile_count::zero ());
> > > > + return !other.nonzero_p ();
> >
> > We already have
> >
> > True:
> > profile_count::zero < any other value
> > any other value > profile_count::zero
> > profile_count::zero <= any initialized value
> > profile_count::zero <= profile_count::zero
> > any initialized value >= profile_count::zero
> >
> > false
> > profile_count::zero > any other value
> > any other value < profile_count::zero
> >
> > You are right about typo in >=, it should be:
> >
> > Index: profile-count.h
> > ===================================================================
> > --- profile-count.h (revision 266450)
> > +++ profile-count.h (working copy)
> > @@ -879,7 +879,7 @@
> > if (other == profile_count::zero ())
> > return true;
> > if (*this == profile_count::zero ())
> > - return !(other == profile_count::zero ());
> > + return other == profile_count::zero ();
> > gcc_checking_assert (compatible_p (other));
> > return m_val >= other.m_val;
> > }
> >
> > With your patch we get false for:
> > profile_count::zero < guessed/auto_fdo/other 0
> > guessed/auto_fdo/other > profile_count::zero
> > guessed/auto_fdo/other <= profile_count::zero
> > profile_count::zero >= profile_count::zero
> >
> > The original idea was to intentionally make profile_count::zero smaller
> > than any toher types of initialized values, since it is more strict hint
> > that the path will not be taken.
> > For example in bb_reorder if you end up with "funny" profile with two
> > exit edges one having profile_count::zero and other being zero as result
> > of (unsucesfull) profile updates it is still better idea to pick the
> > profile_count::zero for taken edge. With your patch it will end up
> > picking either of the paths.
> >
> > How the patch helps to your situation?
> Hi Honza, thanks very much for elaborating. Issue in case of autofdo
> is as described in last message:
> Given update_profiling_info implemented as below:
>
> update_profiling_info (struct cgraph_node *orig_node,
> struct cgraph_node *new_node)
> {
> struct cgraph_edge *cs;
> struct caller_statistics stats;
> profile_count new_sum, orig_sum;
> profile_count remainder, orig_node_count = orig_node->count;
>
> //*****Operator ">" returns true if orig_node_count == autofdo.zero.
> if (!(orig_node_count.ipa () > profile_count::zero ()))
> return;
> //...
> for (cs = new_node->callees; cs; cs = cs->next_callee)
> //*****Result in apply_scale called with autofdo.zero as the 2nd
> argument.
> cs->count = cs->count.apply_scale (new_sum, orig_node_count);
>
> Also apply_scale is implemented as:
> profile_count apply_scale (profile_count num, profile_count den) const
> {
> if (*this == profile_count::zero ())
> return *this;
> if (num == profile_count::zero ())
> return num;
> if (!initialized_p () || !num.initialized_p () || !den.initialized_p ())
> return profile_count::uninitialized ();
> if (num == den)
> return *this;
> gcc_checking_assert (den.m_val);
>
> Here we have (num != zero && den == autofdo.zero), it triggers the
> gcc_checking_assert.
> According to your explanation, guess we need to call force_nonzero for
> orig_node_count before calling apply_scale, right?
Hi Honza,
I have committed the typo fix as revision 266885.
Also I followed your suggestion (IIUC) by calling
profile_count::adjust_for_ipa_scaling for zero den in function
update_profiling_info. It works and does make more sense than
changing the global zero check logic.
Patch tested as before, is it ok?
Thanks,
bin
diff --git a/gcc/ipa-cp.c b/gcc/ipa-cp.c
index 4471bae11c7..5074ef63da1 100644
--- a/gcc/ipa-cp.c
+++ b/gcc/ipa-cp.c
@@ -3715,9 +3715,11 @@ update_profiling_info (struct cgraph_node *orig_node,
new_sum = orig_node_count.combine_with_ipa_count (new_sum);
orig_node->count = remainder;
+ profile_count::adjust_for_ipa_scaling (&new_sum, &orig_node_count);
for (cs = new_node->callees; cs; cs = cs->next_callee)
cs->count = cs->count.apply_scale (new_sum, orig_node_count);
+ profile_count::adjust_for_ipa_scaling (&remainder, &orig_node_count);
for (cs = orig_node->callees; cs; cs = cs->next_callee)
cs->count = cs->count.apply_scale (remainder, orig_node_count);
2018-12-07 Bin Cheng <[email protected]>
* ipa-cp.c (update_profiling_info): Call adjust_for_ipa_scaling for
zero profile count.
