Add attribute hot judgement for INLINE_HINT_known_hot hint.

Message ID 20220920084810.25157-1-lili.cui@intel.com
State Accepted, archived
Headers
Series Add attribute hot judgement for INLINE_HINT_known_hot hint. |

Checks

Context Check Description
snail/gcc-patches-check success Github commit url

Commit Message

Li, Pan2 via Gcc-patches Sept. 20, 2022, 8:48 a.m. UTC
  Hi Honza,

This patch is to add attribute hot judgement for INLINE_HINT_known_hot hint.

We set up INLINE_HINT_known_hot hint only when we have profile feedback,
now add function attribute judgement for it, when both caller and callee
have __attribute__((hot)), we will also set up INLINE_HINT_known_hot hint
for it.

With this patch applied
                                 Ratio   Codesize
ADL Multi-copy:    538.imagic_r  16.7%    1.6%
SPR Multi-copy:    538.imagic_r  15%      1.7%
ICX Multi-copy:    538.imagic_r  15.2%    1.7%
CLX Multi-copy:    538.imagic_r  12.7%    1.7%
Znver3 Multi-copy: 538.imagic_r  10.6%    1.5%

Bootstrap and regtest pending on x86_64-unknown-linux-gnu.
OK for trunk?

Thanks,
Lili.

gcc/ChangeLog

  * ipa-inline-analysis.cc (do_estimate_edge_time): Add function attribute
  judgement for INLINE_HINT_known_hot hint.
---
 gcc/ipa-inline-analysis.cc | 13 +++++++++----
 1 file changed, 9 insertions(+), 4 deletions(-)
  

Comments

Jan Hubicka Sept. 20, 2022, 4:17 p.m. UTC | #1
> Hi Honza,
> 
> This patch is to add attribute hot judgement for INLINE_HINT_known_hot hint.
> 
> We set up INLINE_HINT_known_hot hint only when we have profile feedback,
> now add function attribute judgement for it, when both caller and callee
> have __attribute__((hot)), we will also set up INLINE_HINT_known_hot hint
> for it.
> 
> With this patch applied
>                                  Ratio   Codesize
> ADL Multi-copy:    538.imagic_r  16.7%    1.6%
> SPR Multi-copy:    538.imagic_r  15%      1.7%
> ICX Multi-copy:    538.imagic_r  15.2%    1.7%
> CLX Multi-copy:    538.imagic_r  12.7%    1.7%
> Znver3 Multi-copy: 538.imagic_r  10.6%    1.5%
> 
> Bootstrap and regtest pending on x86_64-unknown-linux-gnu.
> OK for trunk?
> 
> Thanks,
> Lili.
> 
> gcc/ChangeLog
> 
>   * ipa-inline-analysis.cc (do_estimate_edge_time): Add function attribute
>   judgement for INLINE_HINT_known_hot hint.

Thank you.  Can you please also add a testcase that tests for this.
So you modify imagemagick marking attribute hot on the specific inline?
I will try to also look again at your earlier patch - I had very busy
summer and unfortunately lost track on this one.

Honza
> ---
>  gcc/ipa-inline-analysis.cc | 13 +++++++++----
>  1 file changed, 9 insertions(+), 4 deletions(-)
> 
> diff --git a/gcc/ipa-inline-analysis.cc b/gcc/ipa-inline-analysis.cc
> index 1ca685d1b0e..7bd29c36590 100644
> --- a/gcc/ipa-inline-analysis.cc
> +++ b/gcc/ipa-inline-analysis.cc
> @@ -48,6 +48,7 @@ along with GCC; see the file COPYING3.  If not see
>  #include "ipa-utils.h"
>  #include "cfgexpand.h"
>  #include "gimplify.h"
> +#include "attribs.h"
>  
>  /* Cached node/edge growths.  */
>  fast_call_summary<edge_growth_cache_entry *, va_heap> *edge_growth_cache = NULL;
> @@ -249,15 +250,19 @@ do_estimate_edge_time (struct cgraph_edge *edge, sreal *ret_nonspec_time)
>        hints = estimates.hints;
>      }
>  
> -  /* When we have profile feedback, we can quite safely identify hot
> -     edges and for those we disable size limits.  Don't do that when
> -     probability that caller will call the callee is low however, since it
> +  /* When we have profile feedback or function attribute, we can quite safely
> +     identify hot edges and for those we disable size limits.  Don't do that
> +     when probability that caller will call the callee is low however, since it
>       may hurt optimization of the caller's hot path.  */
> -  if (edge->count.ipa ().initialized_p () && edge->maybe_hot_p ()
> +  if ((edge->count.ipa ().initialized_p () && edge->maybe_hot_p ()
>        && (edge->count.ipa () * 2
>  	  > (edge->caller->inlined_to
>  	     ? edge->caller->inlined_to->count.ipa ()
>  	     : edge->caller->count.ipa ())))
> +      || (lookup_attribute ("hot", DECL_ATTRIBUTES (edge->caller->decl))
> +	  != NULL
> +	 && lookup_attribute ("hot", DECL_ATTRIBUTES (edge->callee->decl))
> +	  != NULL))
>      hints |= INLINE_HINT_known_hot;
>  
>    gcc_checking_assert (size >= 0);
> -- 
> 2.17.1
>
  
Li, Pan2 via Gcc-patches Sept. 21, 2022, 9:21 a.m. UTC | #2
> Thank you.  Can you please also add a testcase that tests for this.
> So you modify imagemagick marking attribute hot on the specific inline?

Thanks Honza. Added the testcase. I didn't modify source code of 538.imagic_r, the original source code has attribute like:

#define magick_hot_spot  __attribute__((__hot__))
static Cache *SetPixelCacheNexusPixels( ... ) magick_hot_spot;

> I will try to also look again at your earlier patch - I had very busy summer and
> unfortunately lost track on this one.
>
NP, I guessed you were busy during that time, my earlier patch was partially duplicated with function "Elimination_by_inlining_prob", 
except "parameter points to caller local memory" part, maybe we can find a suitable place to add local memory part  to the IPA.

> Honza

gcc/ChangeLog

  * ipa-inline-analysis.cc (do_estimate_edge_time): Add function attribute
  judgement for INLINE_HINT_known_hot hint.

gcc/testsuite/ChangeLog:

  * gcc.dg/ipa/inlinehint-6.c: New test.
---
 gcc/ipa-inline-analysis.cc              | 13 ++++---
 gcc/testsuite/gcc.dg/ipa/inlinehint-6.c | 47 +++++++++++++++++++++++++
 2 files changed, 56 insertions(+), 4 deletions(-)
 create mode 100644 gcc/testsuite/gcc.dg/ipa/inlinehint-6.c

diff --git a/gcc/ipa-inline-analysis.cc b/gcc/ipa-inline-analysis.cc
index 1ca685d1b0e..7bd29c36590 100644
--- a/gcc/ipa-inline-analysis.cc
+++ b/gcc/ipa-inline-analysis.cc
@@ -48,6 +48,7 @@ along with GCC; see the file COPYING3.  If not see
 #include "ipa-utils.h"
 #include "cfgexpand.h"
 #include "gimplify.h"
+#include "attribs.h"
 
 /* Cached node/edge growths.  */
 fast_call_summary<edge_growth_cache_entry *, va_heap> *edge_growth_cache = NULL;
@@ -249,15 +250,19 @@ do_estimate_edge_time (struct cgraph_edge *edge, sreal *ret_nonspec_time)
       hints = estimates.hints;
     }
 
-  /* When we have profile feedback, we can quite safely identify hot
-     edges and for those we disable size limits.  Don't do that when
-     probability that caller will call the callee is low however, since it
+  /* When we have profile feedback or function attribute, we can quite safely
+     identify hot edges and for those we disable size limits.  Don't do that
+     when probability that caller will call the callee is low however, since it
      may hurt optimization of the caller's hot path.  */
-  if (edge->count.ipa ().initialized_p () && edge->maybe_hot_p ()
+  if ((edge->count.ipa ().initialized_p () && edge->maybe_hot_p ()
       && (edge->count.ipa () * 2
 	  > (edge->caller->inlined_to
 	     ? edge->caller->inlined_to->count.ipa ()
 	     : edge->caller->count.ipa ())))
+      || (lookup_attribute ("hot", DECL_ATTRIBUTES (edge->caller->decl))
+	  != NULL
+	 && lookup_attribute ("hot", DECL_ATTRIBUTES (edge->callee->decl))
+	  != NULL))
     hints |= INLINE_HINT_known_hot;
 
   gcc_checking_assert (size >= 0);
diff --git a/gcc/testsuite/gcc.dg/ipa/inlinehint-6.c b/gcc/testsuite/gcc.dg/ipa/inlinehint-6.c
new file mode 100644
index 00000000000..1f3be641c6d
--- /dev/null
+++ b/gcc/testsuite/gcc.dg/ipa/inlinehint-6.c
@@ -0,0 +1,47 @@
+/* { dg-options "-O3 -c -fdump-ipa-inline-details -fno-early-inlining -fno-ipa-cp"  } */
+/* { dg-add-options bind_pic_locally } */
+
+#define size_t long long int
+
+struct A
+{
+  size_t f1, f2, f3, f4;
+};
+struct C
+{
+  struct A a;
+  size_t b;
+};
+struct C x;
+
+__attribute__((hot)) struct C callee (struct A *a, struct C *c)
+{
+  c->a=(*a);
+
+  if((c->b + 7) & 17)
+   {
+      c->a.f1 = c->a.f2 + c->a.f1;
+      c->a.f2 = c->a.f3 - c->a.f2;
+      c->a.f3 = c->a.f2 + c->a.f3;
+      c->a.f4 = c->a.f2 - c->a.f4;
+      c->b = c->a.f2;
+
+    }
+  return *c;
+}
+
+__attribute__((hot)) struct C caller (size_t d, size_t e, size_t f, size_t g, struct C *c)
+{
+  struct A a;
+  a.f1 = 1 + d;
+  a.f2 = e;
+  a.f3 = 12 + f;
+  a.f4 = 68 + g;
+  if (c->b > 0)
+    return callee (&a, c);
+  else
+    return *c;
+}
+
+/* { dg-final { scan-ipa-dump "known_hot"  "inline"  } } */
+
  
Li, Pan2 via Gcc-patches Oct. 8, 2022, 12:33 a.m. UTC | #3
Hi Honza,

Gentle ping  https://gcc.gnu.org/pipermail/gcc-patches/2022-September/601934.html

Thanks,
Lili.

> -----Original Message-----
> From: Gcc-patches <gcc-patches-bounces+lili.cui=intel.com@gcc.gnu.org> On
> Behalf Of Cui, Lili via Gcc-patches
> Sent: Wednesday, September 21, 2022 5:22 PM
> To: Jan Hubicka <hubicka@ucw.cz>
> Cc: Lu, Hongjiu <hongjiu.lu@intel.com>; Liu, Hongtao
> <hongtao.liu@intel.com>; gcc-patches@gcc.gnu.org
> Subject: RE: [PATCH] Add attribute hot judgement for
> INLINE_HINT_known_hot hint.
> 
> > Thank you.  Can you please also add a testcase that tests for this.
> > So you modify imagemagick marking attribute hot on the specific inline?
> 
> Thanks Honza. Added the testcase. I didn't modify source code of
> 538.imagic_r, the original source code has attribute like:
> 
> #define magick_hot_spot  __attribute__((__hot__)) static Cache
> *SetPixelCacheNexusPixels( ... ) magick_hot_spot;
> 
> > I will try to also look again at your earlier patch - I had very busy
> > summer and unfortunately lost track on this one.
> >
> NP, I guessed you were busy during that time, my earlier patch was partially
> duplicated with function "Elimination_by_inlining_prob", except "parameter
> points to caller local memory" part, maybe we can find a suitable place to
> add local memory part  to the IPA.
> 
> > Honza
> 
> gcc/ChangeLog
> 
>   * ipa-inline-analysis.cc (do_estimate_edge_time): Add function attribute
>   judgement for INLINE_HINT_known_hot hint.
> 
> gcc/testsuite/ChangeLog:
> 
>   * gcc.dg/ipa/inlinehint-6.c: New test.
> ---
>  gcc/ipa-inline-analysis.cc              | 13 ++++---
>  gcc/testsuite/gcc.dg/ipa/inlinehint-6.c | 47 +++++++++++++++++++++++++
>  2 files changed, 56 insertions(+), 4 deletions(-)  create mode 100644
> gcc/testsuite/gcc.dg/ipa/inlinehint-6.c
> 
> diff --git a/gcc/ipa-inline-analysis.cc b/gcc/ipa-inline-analysis.cc index
> 1ca685d1b0e..7bd29c36590 100644
> --- a/gcc/ipa-inline-analysis.cc
> +++ b/gcc/ipa-inline-analysis.cc
> @@ -48,6 +48,7 @@ along with GCC; see the file COPYING3.  If not see
> #include "ipa-utils.h"
>  #include "cfgexpand.h"
>  #include "gimplify.h"
> +#include "attribs.h"
> 
>  /* Cached node/edge growths.  */
>  fast_call_summary<edge_growth_cache_entry *, va_heap>
> *edge_growth_cache = NULL; @@ -249,15 +250,19 @@
> do_estimate_edge_time (struct cgraph_edge *edge, sreal *ret_nonspec_time)
>        hints = estimates.hints;
>      }
> 
> -  /* When we have profile feedback, we can quite safely identify hot
> -     edges and for those we disable size limits.  Don't do that when
> -     probability that caller will call the callee is low however, since it
> +  /* When we have profile feedback or function attribute, we can quite
> safely
> +     identify hot edges and for those we disable size limits.  Don't do that
> +     when probability that caller will call the callee is low however,
> + since it
>       may hurt optimization of the caller's hot path.  */
> -  if (edge->count.ipa ().initialized_p () && edge->maybe_hot_p ()
> +  if ((edge->count.ipa ().initialized_p () && edge->maybe_hot_p ()
>        && (edge->count.ipa () * 2
>  	  > (edge->caller->inlined_to
>  	     ? edge->caller->inlined_to->count.ipa ()
>  	     : edge->caller->count.ipa ())))
> +      || (lookup_attribute ("hot", DECL_ATTRIBUTES (edge->caller->decl))
> +	  != NULL
> +	 && lookup_attribute ("hot", DECL_ATTRIBUTES (edge->callee->decl))
> +	  != NULL))
>      hints |= INLINE_HINT_known_hot;
> 
>    gcc_checking_assert (size >= 0);
> diff --git a/gcc/testsuite/gcc.dg/ipa/inlinehint-6.c
> b/gcc/testsuite/gcc.dg/ipa/inlinehint-6.c
> new file mode 100644
> index 00000000000..1f3be641c6d
> --- /dev/null
> +++ b/gcc/testsuite/gcc.dg/ipa/inlinehint-6.c
> @@ -0,0 +1,47 @@
> +/* { dg-options "-O3 -c -fdump-ipa-inline-details -fno-early-inlining
> +-fno-ipa-cp"  } */
> +/* { dg-add-options bind_pic_locally } */
> +
> +#define size_t long long int
> +
> +struct A
> +{
> +  size_t f1, f2, f3, f4;
> +};
> +struct C
> +{
> +  struct A a;
> +  size_t b;
> +};
> +struct C x;
> +
> +__attribute__((hot)) struct C callee (struct A *a, struct C *c) {
> +  c->a=(*a);
> +
> +  if((c->b + 7) & 17)
> +   {
> +      c->a.f1 = c->a.f2 + c->a.f1;
> +      c->a.f2 = c->a.f3 - c->a.f2;
> +      c->a.f3 = c->a.f2 + c->a.f3;
> +      c->a.f4 = c->a.f2 - c->a.f4;
> +      c->b = c->a.f2;
> +
> +    }
> +  return *c;
> +}
> +
> +__attribute__((hot)) struct C caller (size_t d, size_t e, size_t f,
> +size_t g, struct C *c) {
> +  struct A a;
> +  a.f1 = 1 + d;
> +  a.f2 = e;
> +  a.f3 = 12 + f;
> +  a.f4 = 68 + g;
> +  if (c->b > 0)
> +    return callee (&a, c);
> +  else
> +    return *c;
> +}
> +
> +/* { dg-final { scan-ipa-dump "known_hot"  "inline"  } } */
> +
> --
> 2.17.1
> 
> Thanks,
> Lili.
  
Li, Pan2 via Gcc-patches Oct. 14, 2022, 6:28 a.m. UTC | #4
Hi Honza,

 Gentle ping  https://gcc.gnu.org/pipermail/gcc-patches/2022-September/601934.html
 
 Thanks,
 Lili.

> -----Original Message-----
> From: Cui, Lili <lili.cui@intel.com>
> Sent: Saturday, October 8, 2022 8:33 AM
> To: Cui, Lili <lili.cui@intel.com>; Jan Hubicka <hubicka@ucw.cz>
> Cc: Lu, Hongjiu <hongjiu.lu@intel.com>; Liu, Hongtao
> <hongtao.liu@intel.com>; gcc-patches@gcc.gnu.org
> Subject: Ping^1 [PATCH] Add attribute hot judgement for
> INLINE_HINT_known_hot hint.
> 
> Hi Honza,
> 
> Gentle ping  https://gcc.gnu.org/pipermail/gcc-patches/2022-
> September/601934.html
> 
> Thanks,
> Lili.
> 
> > -----Original Message-----
> > From: Gcc-patches <gcc-patches-bounces+lili.cui=intel.com@gcc.gnu.org>
> > On Behalf Of Cui, Lili via Gcc-patches
> > Sent: Wednesday, September 21, 2022 5:22 PM
> > To: Jan Hubicka <hubicka@ucw.cz>
> > Cc: Lu, Hongjiu <hongjiu.lu@intel.com>; Liu, Hongtao
> > <hongtao.liu@intel.com>; gcc-patches@gcc.gnu.org
> > Subject: RE: [PATCH] Add attribute hot judgement for
> > INLINE_HINT_known_hot hint.
> >
> > > Thank you.  Can you please also add a testcase that tests for this.
> > > So you modify imagemagick marking attribute hot on the specific inline?
> >
> > Thanks Honza. Added the testcase. I didn't modify source code of
> > 538.imagic_r, the original source code has attribute like:
> >
> > #define magick_hot_spot  __attribute__((__hot__)) static Cache
> > *SetPixelCacheNexusPixels( ... ) magick_hot_spot;
> >
> > > I will try to also look again at your earlier patch - I had very
> > > busy summer and unfortunately lost track on this one.
> > >
> > NP, I guessed you were busy during that time, my earlier patch was
> > partially duplicated with function "Elimination_by_inlining_prob",
> > except "parameter points to caller local memory" part, maybe we can
> > find a suitable place to add local memory part  to the IPA.
> >
> > > Honza
> >
> > gcc/ChangeLog
> >
> >   * ipa-inline-analysis.cc (do_estimate_edge_time): Add function attribute
> >   judgement for INLINE_HINT_known_hot hint.
> >
> > gcc/testsuite/ChangeLog:
> >
> >   * gcc.dg/ipa/inlinehint-6.c: New test.
> > ---
> >  gcc/ipa-inline-analysis.cc              | 13 ++++---
> >  gcc/testsuite/gcc.dg/ipa/inlinehint-6.c | 47
> > +++++++++++++++++++++++++
> >  2 files changed, 56 insertions(+), 4 deletions(-)  create mode 100644
> > gcc/testsuite/gcc.dg/ipa/inlinehint-6.c
> >
> > diff --git a/gcc/ipa-inline-analysis.cc b/gcc/ipa-inline-analysis.cc
> > index
> > 1ca685d1b0e..7bd29c36590 100644
> > --- a/gcc/ipa-inline-analysis.cc
> > +++ b/gcc/ipa-inline-analysis.cc
> > @@ -48,6 +48,7 @@ along with GCC; see the file COPYING3.  If not see
> > #include "ipa-utils.h"
> >  #include "cfgexpand.h"
> >  #include "gimplify.h"
> > +#include "attribs.h"
> >
> >  /* Cached node/edge growths.  */
> >  fast_call_summary<edge_growth_cache_entry *, va_heap>
> > *edge_growth_cache = NULL; @@ -249,15 +250,19 @@
> do_estimate_edge_time
> > (struct cgraph_edge *edge, sreal *ret_nonspec_time)
> >        hints = estimates.hints;
> >      }
> >
> > -  /* When we have profile feedback, we can quite safely identify hot
> > -     edges and for those we disable size limits.  Don't do that when
> > -     probability that caller will call the callee is low however, since it
> > +  /* When we have profile feedback or function attribute, we can
> > + quite
> > safely
> > +     identify hot edges and for those we disable size limits.  Don't do that
> > +     when probability that caller will call the callee is low
> > + however, since it
> >       may hurt optimization of the caller's hot path.  */
> > -  if (edge->count.ipa ().initialized_p () && edge->maybe_hot_p ()
> > +  if ((edge->count.ipa ().initialized_p () && edge->maybe_hot_p ()
> >        && (edge->count.ipa () * 2
> >  	  > (edge->caller->inlined_to
> >  	     ? edge->caller->inlined_to->count.ipa ()
> >  	     : edge->caller->count.ipa ())))
> > +      || (lookup_attribute ("hot", DECL_ATTRIBUTES (edge->caller->decl))
> > +	  != NULL
> > +	 && lookup_attribute ("hot", DECL_ATTRIBUTES (edge->callee->decl))
> > +	  != NULL))
> >      hints |= INLINE_HINT_known_hot;
> >
> >    gcc_checking_assert (size >= 0);
> > diff --git a/gcc/testsuite/gcc.dg/ipa/inlinehint-6.c
> > b/gcc/testsuite/gcc.dg/ipa/inlinehint-6.c
> > new file mode 100644
> > index 00000000000..1f3be641c6d
> > --- /dev/null
> > +++ b/gcc/testsuite/gcc.dg/ipa/inlinehint-6.c
> > @@ -0,0 +1,47 @@
> > +/* { dg-options "-O3 -c -fdump-ipa-inline-details -fno-early-inlining
> > +-fno-ipa-cp"  } */
> > +/* { dg-add-options bind_pic_locally } */
> > +
> > +#define size_t long long int
> > +
> > +struct A
> > +{
> > +  size_t f1, f2, f3, f4;
> > +};
> > +struct C
> > +{
> > +  struct A a;
> > +  size_t b;
> > +};
> > +struct C x;
> > +
> > +__attribute__((hot)) struct C callee (struct A *a, struct C *c) {
> > +  c->a=(*a);
> > +
> > +  if((c->b + 7) & 17)
> > +   {
> > +      c->a.f1 = c->a.f2 + c->a.f1;
> > +      c->a.f2 = c->a.f3 - c->a.f2;
> > +      c->a.f3 = c->a.f2 + c->a.f3;
> > +      c->a.f4 = c->a.f2 - c->a.f4;
> > +      c->b = c->a.f2;
> > +
> > +    }
> > +  return *c;
> > +}
> > +
> > +__attribute__((hot)) struct C caller (size_t d, size_t e, size_t f,
> > +size_t g, struct C *c) {
> > +  struct A a;
> > +  a.f1 = 1 + d;
> > +  a.f2 = e;
> > +  a.f3 = 12 + f;
> > +  a.f4 = 68 + g;
> > +  if (c->b > 0)
> > +    return callee (&a, c);
> > +  else
> > +    return *c;
> > +}
> > +
> > +/* { dg-final { scan-ipa-dump "known_hot"  "inline"  } } */
> > +
> > --
> > 2.17.1
> >
> > Thanks,
> > Lili.
  

Patch

diff --git a/gcc/ipa-inline-analysis.cc b/gcc/ipa-inline-analysis.cc
index 1ca685d1b0e..7bd29c36590 100644
--- a/gcc/ipa-inline-analysis.cc
+++ b/gcc/ipa-inline-analysis.cc
@@ -48,6 +48,7 @@  along with GCC; see the file COPYING3.  If not see
 #include "ipa-utils.h"
 #include "cfgexpand.h"
 #include "gimplify.h"
+#include "attribs.h"
 
 /* Cached node/edge growths.  */
 fast_call_summary<edge_growth_cache_entry *, va_heap> *edge_growth_cache = NULL;
@@ -249,15 +250,19 @@  do_estimate_edge_time (struct cgraph_edge *edge, sreal *ret_nonspec_time)
       hints = estimates.hints;
     }
 
-  /* When we have profile feedback, we can quite safely identify hot
-     edges and for those we disable size limits.  Don't do that when
-     probability that caller will call the callee is low however, since it
+  /* When we have profile feedback or function attribute, we can quite safely
+     identify hot edges and for those we disable size limits.  Don't do that
+     when probability that caller will call the callee is low however, since it
      may hurt optimization of the caller's hot path.  */
-  if (edge->count.ipa ().initialized_p () && edge->maybe_hot_p ()
+  if ((edge->count.ipa ().initialized_p () && edge->maybe_hot_p ()
       && (edge->count.ipa () * 2
 	  > (edge->caller->inlined_to
 	     ? edge->caller->inlined_to->count.ipa ()
 	     : edge->caller->count.ipa ())))
+      || (lookup_attribute ("hot", DECL_ATTRIBUTES (edge->caller->decl))
+	  != NULL
+	 && lookup_attribute ("hot", DECL_ATTRIBUTES (edge->callee->decl))
+	  != NULL))
     hints |= INLINE_HINT_known_hot;
 
   gcc_checking_assert (size >= 0);