[wwwdocs] Add info about IPA optimization and LTO improvments

[wwwdocs] Add info about IPA optimization and LTO improvments

[Jan Hubicka]

Hi,
this patch adds some info on changes in IPA and LTO. Not sure if I missed
something important.  Martin, perhaps you could add some example on new
devirutalization stuff?
Suggestions for less lame explanations are welcome :)

Honza

Index: changes.html
===================================================================
RCS file: /cvs/gcc/wwwdocs/htdocs/gcc-4.7/changes.html,v
retrieving revision 1.28
diff -c -3 -r1.28 changes.html
*** changes.html	12 Sep 2011 07:46:05 -0000	1.28
--- changes.html	21 Sep 2011 09:51:12 -0000
***************
*** 58,63 ****
--- 58,122 ----
      was added to allow users to control the cutoff between doing switch statements
      as a series of if statements and using a jump table.
      </li>
+     <li>Link-time optimization improvements:
+     <ul>
+       <li>Improved scalability and reduced memory usage.  Link time optimization
+       of Firefox now require 3GB of RAM on 64bit system, while over 8GB was needed
+       previously. Linking time has been improved, too. The serial stage of linking
+       Firefox binary has been sped up approximately by factor of 10.</li>
+       <li>Reduced size of object files and temporary storage used during linktime.</li>
+       <li>Streaming performance has been improved.</li>
+       <li>Number of bug fixes, especially in symbol table handling and merging.</li>
+     </ul>
+     <li>Interprocedural optimization improvements:
+       <li>Inliner heuristic can now take into account fact that after inlining
+       code will be optimized out because of known values (or properties) of function parameters.
+       For example:
+       <pre>
+ void foo(int a)
+ {
+   if (a>10)
+     ... huge code ...
+ }
+ void bar (void)
+ {
+   foo (0);
+ }
+       </pre>
+       The call of <code>foo</code> will be inlined into <code>bar</code> even when
+       optimizing for code size. Constructs based on <code>__builtin_constant_p</code>
+       are now understood by inliner and code size estimates are evaulated a lot
+       more realistically.</li>
+       <li>Representation of C++ virtual thunks and aliases has been
+       re-engineered. The aliases no longer pose optimization barriers and calls to an
+       alias can be inlined and otherwise optimized.</li>
+       <li>Inter-procedural constant propagation pass has been rewritten.  It now performs
+       generic function specialization.  For example when compiling the following:
+       <pre>
+ void foo(bool flag)
+ {
+   if (flag)
+     ... do something ...
+   else
+     ... do something else ...
+ }
+ void bar (void)
+ {
+   foo (false);
+   foo (true);
+   foo (false);
+   foo (true);
+   foo (false);
+   foo (true);
+ }
+       </pre>
+       GCC will now produce two copies of <code>foo</code>. One with <code>flag</code> being
+       <code>true</code>, while other with <code>flag</code> being
+       <code>false</code>.  This leads to preformance improvements previusly
+       possibly only by inlining all calls.  Cloning cause a lot less code size
+       growth.
+     <ul>
+     </ul>
    </ul>
  
  <h2>New Languages and Language specific improvements</h2>

Re: [wwwdocs] Add info about IPA optimization and LTO improvments

[Gerald Pfeifer]

On Wed, 21 Sep 2011, Jan Hubicka wrote:
> Index: changes.html
> ===================================================================
> +     <li>Link-time optimization improvements:
> +     <ul>
> +       <li>Improved scalability and reduced memory usage.  Link time optimization
> +       of Firefox now require 3GB of RAM on 64bit system, while over 8GB was needed

"requires"
"on a 64-bit"

> +       previously. Linking time has been improved, too. The serial stage of linking
> +       Firefox binary has been sped up approximately by factor of 10.</li>
> +       <li>Reduced size of object files and temporary storage used during linktime.</li>

"linking"?

> +       <li>Streaming performance has been improved.</li>

Outbound, inbound, or both?

> +       <li>Number of bug fixes, especially in symbol table handling and merging.</li>

"Several bug fixes"

> +       <li>Inliner heuristic can now take into account fact that after inlining
> +       code will be optimized out because of known values (or properties) of function parameters.

Omit "fact".

> +       The call of <code>foo</code> will be inlined into <code>bar</code> even when
> +       optimizing for code size. Constructs based on <code>__builtin_constant_p</code>
> +       are now understood by inliner and code size estimates are evaulated a lot
> +       more realistically.</li>

"by the inliner"
Typo: "evaluated"

> +       <li>Representation of C++ virtual thunks and aliases has been

"The representation"

> +       re-engineered. The aliases no longer pose optimization barriers and calls to an

"C++ aliases" or just "Aliases" if you think the context is sufficiently
clear.

> +       <li>Inter-procedural constant propagation pass has been rewritten.  It now performsa

"The inter-procedural"

> +       GCC will now produce two copies of <code>foo</code>. One with <code>flag</code> being
> +       <code>true</code>, while other with <code>flag</code> being
> +       <code>false</code>. 

"...de>, the other with..."

> This leads to preformance improvements previusly

Typo: "performance"

Typo: "previously"

> + possibly only by inlining all calls.  Cloning cause a lot less code 

"causes"


Fine from my perspective modulo the changes above.

I guess I need to dust my old codebase and see how modern GCC fares
in terms of performance. :-)

Gerald

Re: [wwwdocs] Add info about IPA optimization and LTO improvments

[Andi Kleen]

Jan Hubicka <hubicka@ucw.cz> writes:
>       </li>
> +     <li>Link-time optimization improvements:
> +     <ul>
> +       <li>Improved scalability and reduced memory usage.  Link time optimization
> +       of Firefox now require 3GB of RAM on 64bit system, while over 8GB was needed
> +       previously. Linking time has been improved, too. The serial stage of linking
> +       Firefox binary has been sped up approximately by factor of 10.</li>
> +       <li>Reduced size of object files and temporary storage used during linktime.</li>
> +       <li>Streaming performance has been improved.</li>
> +       <li>Number of bug fixes, especially in symbol table handling
> and merging.</li>

<li>ld -r is now supported with LTO</li>


-Andi

-- 
ak@linux.intel.com -- Speaking for myself only

Re: [wwwdocs] Add info about IPA optimization and LTO improvments

[Jan Hubicka]

> Jan Hubicka <hubicka@ucw.cz> writes:
> >       </li>
> > +     <li>Link-time optimization improvements:
> > +     <ul>
> > +       <li>Improved scalability and reduced memory usage.  Link time optimization
> > +       of Firefox now require 3GB of RAM on 64bit system, while over 8GB was needed
> > +       previously. Linking time has been improved, too. The serial stage of linking
> > +       Firefox binary has been sped up approximately by factor of 10.</li>
> > +       <li>Reduced size of object files and temporary storage used during linktime.</li>
> > +       <li>Streaming performance has been improved.</li>
> > +       <li>Number of bug fixes, especially in symbol table handling
> > and merging.</li>
> 
> <li>ld -r is now supported with LTO</li>

Thanks, forgot about this one.  This also needs support at linker side, right?
Do you know minimal GNU ld/Gold versions that works fine?

Honza

Re: [wwwdocs] Add info about IPA optimization and LTO improvments

[Andi Kleen]

> > <li>ld -r is now supported with LTO</li>
> 
> Thanks, forgot about this one.  This also needs support at linker side, right?

Only if you include assembler or non LTO code.
Without that it should work with any linker.

> Do you know minimal GNU ld/Gold versions that works fine?

I believe the assembler code only works with HJ's version of binutils
and his BFD ld (not gold) currently. HJ can you supply the minimal version?
Non assembler/LTO should work always.

I hope both gold and normal binutils ld will eventually get the support
too.

-Andi
-- 
ak@linux.intel.com -- Speaking for myself only.

Re: [wwwdocs] Add info about IPA optimization and LTO improvments

[Jan Hubicka]

Gerald, Andi,
thanks for corrections.  This is what I've comitted now.

Index: changes.html
===================================================================
RCS file: /cvs/gcc/wwwdocs/htdocs/gcc-4.7/changes.html,v
retrieving revision 1.33
diff -u -r1.33 changes.html
--- changes.html	25 Sep 2011 22:49:42 -0000	1.33
+++ changes.html	27 Sep 2011 13:50:13 -0000
@@ -58,6 +58,68 @@
     was added to allow users to control the cutoff between doing switch statements
     as a series of if statements and using a jump table.
     </li>
+    <li>Link-time optimization improvements:
+    <ul>
+      <li>Improved scalability and reduced memory usage.  Link time optimization
+      of Firefox now requires 3GB of RAM on a 64bit system, while over 8GB was needed
+      previously. Linking time has been improved, too. The serial stage of linking
+      Firefox binary has been sped up approximately by factor of 10.</li>
+      <li>Reduced size of object files and temporary storage used during linking.</li>
+      <li>Streaming performance (both outbound and inbound) has been improved.</li>
+      <li><code>ld -r</code> is now supported with LTO.</li>
+      <li>Several bug fixes, especially in symbol table handling and merging.</li>
+    </ul>
+    <li>Interprocedural optimization improvements:
+      <li>Inliner heuristic can now take into account that after inlining
+      code will be optimized out because of known values (or properties) of function parameters.
+      For example:
+      <pre>
+void foo(int a)
+{
+  if (a>10)
+    ... huge code ...
+}
+void bar (void)
+{
+  foo (0);
+}
+      </pre>
+      The call of <code>foo</code> will be inlined into <code>bar</code> even when
+      optimizing for code size. Constructs based on <code>__builtin_constant_p</code>
+      are now understood by the inliner and code size estimates are evaluated a lot
+      more realistically.</li>
+      <li>The representation of C++ virtual thunks and aliases (both implicit and defined
+      via <code>alias</code>attribute) has been re-engineered. The aliases no
+      longer pose optimization barriers and calls to an alias can be inlined
+      and otherwise optimized.</li>
+      <li>The inter-procedural constant propagation pass has been rewritten.
+      It now performs generic function specialization.  For example when
+      compiling the following:
+      <pre>
+void foo(bool flag)
+{
+  if (flag)
+    ... do something ...
+  else
+    ... do something else ...
+}
+void bar (void)
+{
+  foo (false);
+  foo (true);
+  foo (false);
+  foo (true);
+  foo (false);
+  foo (true);
+}
+      </pre>
+      GCC will now produce two copies of <code>foo</code>. One with <code>flag</code> being
+      <code>true</code>, while other with <code>flag</code> being
+      <code>false</code>.  This leads to performance improvements previously
+      possibly only by inlining all calls.  Cloning causes a lot less code size
+      growth.
+    <ul>
+    </ul>
   </ul>
 
 <h2>New Languages and Language specific improvements</h2>

Re: [wwwdocs] Add info about IPA optimization and LTO improvments

[Gerald Pfeifer]

On Tue, 27 Sep 2011, Jan Hubicka wrote:
> thanks for corrections.  This is what I've comitted now.

And here are some markup fixes on top, that address complaints by
the validator.

Gerald


Index: changes.html
===================================================================
RCS file: /cvs/gcc/wwwdocs/htdocs/gcc-4.7/changes.html,v
retrieving revision 1.34
diff -u -r1.34 changes.html
--- changes.html	27 Sep 2011 13:50:27 -0000	1.34
+++ changes.html	27 Sep 2011 16:56:42 -0000
@@ -58,6 +58,7 @@
     was added to allow users to control the cutoff between doing switch statements
     as a series of if statements and using a jump table.
     </li>
+
     <li>Link-time optimization improvements:
     <ul>
       <li>Improved scalability and reduced memory usage.  Link time optimization
@@ -68,8 +69,10 @@
       <li>Streaming performance (both outbound and inbound) has been improved.</li>
       <li><code>ld -r</code> is now supported with LTO.</li>
       <li>Several bug fixes, especially in symbol table handling and merging.</li>
-    </ul>
+    </ul></li>
+
     <li>Interprocedural optimization improvements:
+    <ul>
       <li>Inliner heuristic can now take into account that after inlining
       code will be optimized out because of known values (or properties) of function parameters.
       For example:
@@ -117,9 +120,8 @@
       <code>true</code>, while other with <code>flag</code> being
       <code>false</code>.  This leads to performance improvements previously
       possibly only by inlining all calls.  Cloning causes a lot less code size
-      growth.
-    <ul>
-    </ul>
+      growth.</li>
+    </ul></li>
   </ul>
 
 <h2>New Languages and Language specific improvements</h2>

Re: [wwwdocs] Add info about IPA optimization and LTO improvments

[Gerald Pfeifer]

On Tue, 27 Sep 2011, Jan Hubicka wrote:
> thanks for corrections.  This is what I've comitted now.

Here is another small update I just crafted, that makes a stylistic
change here and there, introduces LTO as an abbreviation, adds some
missing articles and spaces,...

Installed.

Gerald

Index: changes.html
===================================================================
RCS file: /cvs/gcc/wwwdocs/htdocs/gcc-4.7/changes.html,v
retrieving revision 1.37
diff -u -r1.37 changes.html
--- changes.html	27 Sep 2011 16:57:02 -0000	1.37
+++ changes.html	27 Sep 2011 18:42:03 -0000
@@ -59,12 +59,13 @@
     as a series of if statements and using a jump table.
     </li>
 
-    <li>Link-time optimization improvements:
+    <li>Link-time optimization (LTO) improvements:
     <ul>
-      <li>Improved scalability and reduced memory usage.  Link time optimization
-      of Firefox now requires 3GB of RAM on a 64bit system, while over 8GB was needed
-      previously. Linking time has been improved, too. The serial stage of linking
-      Firefox binary has been sped up approximately by factor of 10.</li>
+      <li>Improved scalability and reduced memory usage.  Link time
+      optimization of Firefox now requires 3GB of RAM on a 64-bit system,
+      while over 8GB was needed previously. Linking time has been improved,
+      too. The serial stage of linking Firefox has been sped up by about a
+      factor of 10.</li>
       <li>Reduced size of object files and temporary storage used during linking.</li>
       <li>Streaming performance (both outbound and inbound) has been improved.</li>
       <li><code>ld -r</code> is now supported with LTO.</li>
@@ -73,8 +74,9 @@
 
     <li>Interprocedural optimization improvements:
     <ul>
-      <li>Inliner heuristic can now take into account that after inlining
-      code will be optimized out because of known values (or properties) of function parameters.
+      <li>Heuristics now take into account that after inlining code will
+      be optimized out because of known values (or properties) of function
+      parameters.
       For example:
       <pre>
 void foo(int a)
@@ -92,7 +94,7 @@
       are now understood by the inliner and code size estimates are evaluated a lot
       more realistically.</li>
       <li>The representation of C++ virtual thunks and aliases (both implicit and defined
-      via <code>alias</code>attribute) has been re-engineered. The aliases no
+      via the <code>alias</code> attribute) has been re-engineered. Aliases no
       longer pose optimization barriers and calls to an alias can be inlined
       and otherwise optimized.</li>
       <li>The inter-procedural constant propagation pass has been rewritten.

Re: [wwwdocs] Add info about IPA optimization and LTO improvments

[H.J. Lu]

On Mon, Sep 26, 2011 at 7:59 AM, Andi Kleen <andi@firstfloor.org> wrote:
>> > <li>ld -r is now supported with LTO</li>
>>
>> Thanks, forgot about this one.  This also needs support at linker side, right?
>
> Only if you include assembler or non LTO code.
> Without that it should work with any linker.
>
>> Do you know minimal GNU ld/Gold versions that works fine?
>
> I believe the assembler code only works with HJ's version of binutils
> and his BFD ld (not gold) currently. HJ can you supply the minimal version?
> Non assembler/LTO should work always.

The Linux binutils 2.21.51.0.3 is the first Linux binutils which
supports "ld -r" on mixed IR/non-IR objects.


-- 
H.J.

Re: [wwwdocs] Add info about IPA optimization and LTO improvments

[Andi Kleen]

> > I believe the assembler code only works with HJ's version of binutils
> > and his BFD ld (not gold) currently. HJ can you supply the minimal version?
> > Non assembler/LTO should work always.
> 
> The Linux binutils 2.21.51.0.3 is the first Linux binutils which
> supports "ld -r" on mixed IR/non-IR objects.

Okay updated line

<li>ld -r is now supported with LTO. When using assembler files
or non LTOed objects inside ld -r objects together with LTO then the Linux
binutils 2.21.51.0.3 or later are needed.</li>


-Andi
-- 
ak@linux.intel.com -- Speaking for myself only.

Re: [wwwdocs] Add info about IPA optimization and LTO improvments

[Gerald Pfeifer]

On Wed, 28 Sep 2011, Andi Kleen wrote:
> <li>ld -r is now supported with LTO. When using assembler files or non 
> LTOed objects inside ld -r objects together with LTO then the Linux 
> binutils 2.21.51.0.3 or later are needed.</li>

I think this should be GNU/Linux, if anything, but then I also think
we should not refer to this $notsurewhatthepropertermis variant of our 
official binutils project as part of GCC release notes.

Gerald

Re: [wwwdocs] Add info about IPA optimization and LTO improvments

[Andi Kleen]

On Sat, Oct 08, 2011 at 05:38:07PM -0400, Gerald Pfeifer wrote:
> On Wed, 28 Sep 2011, Andi Kleen wrote:
> > <li>ld -r is now supported with LTO. When using assembler files or non 
> > LTOed objects inside ld -r objects together with LTO then the Linux 
> > binutils 2.21.51.0.3 or later are needed.</li>
> 
> I think this should be GNU/Linux, if anything, but then I also think
> we should not refer to this $notsurewhatthepropertermis variant of our 
> official binutils project as part of GCC release notes.

Okay. The users will stay mystified then.  Hopefully they are good
at googling.

-Andi

-- 
ak@linux.intel.com -- Speaking for myself only.

Re: [wwwdocs] Add info about IPA optimization and LTO improvments

[Gerald Pfeifer]

On Sat, 8 Oct 2011, Andi Kleen wrote:
>> On Wed, 28 Sep 2011, Andi Kleen wrote:
>>> <li>ld -r is now supported with LTO. When using assembler files or non 
>>> LTOed objects inside ld -r objects together with LTO then the Linux 
>>> binutils 2.21.51.0.3 or later are needed.</li>
>> I think this should be GNU/Linux, if anything, but then I also think
>> we should not refer to this $notsurewhatthepropertermis variant of our 
>> official binutils project as part of GCC release notes.
> Okay. The users will stay mystified then.  Hopefully they are good
> at googling.

Are you saying no official version of binutils is able to address
this, and H.J.'s is required?  If so, please go ahead and apply the
patch (just avoid the "naked" instance of Linux).

Thanks,
Gerald