Bug #7556: test error on refinement

• Status: Closed
• Assignee: ko1 (Koichi Sasada)
• Target version: 2.0.0
• ruby -v: ruby -v: ruby 2.0.0dev (2012-12-13 trunk 38354) [x64-mswin64_100]

[usa (Usaku NAKAMURA)]

1. Error:
   test_refine_recursion(TestRefinement):
   NoMethodError: undefined method recursive_length' for "oo":String C:/Users/usa/ruby/test/ruby/test_refinement.rb:567:in recursive_length'
   :in <main>' C:/Users/usa/ruby/test/ruby/test_refinement.rb:806:in eval'
   C:/Users/usa/ruby/test/ruby/test_refinement.rb:806:in eval_using' C:/Users/usa/ruby/test/ruby/test_refinement.rb:574:in test_refine_recursion'

On my box this error is 100% reproducible, but I also know that RubyCI and
RubyInstaller CI don't report this error.
I've heard that nobu reproduced this bug on x86_64-dawrin, but I don't know
the detail of his environment.

Once I wrote the detail of my debuggin, but it is lost by accidenal reboot

of my PC.

I have no energy to rewrite it, because writing long English sentences

irritates me, especially after seeing mails which reproach our native

language.

[duerst (Martin Dürst)]

On 2012/12/13 20:13, usa (Usaku NAKAMURA) wrote:

Issue #7556 has been reported by usa (Usaku NAKAMURA).

Once I wrote the detail of my debuggin, but it is lost by accidenal reboot

of my PC.

I have no energy to rewrite it, because writing long English sentences

irritates me, especially after seeing mails which reproach our native

language.

Two solutions:

1. Write short English sentences.
2. Write in Japanese. (If somebody thinks they really need to know what
   you wrote, they can ask on the list.)

But I know that for creative people, it is much harder to do the same
work again than to do it the first time. I once worked on Devanagari
rendering, but lost that work in a reboot. I didn't want to redo it, so
I worked on Tamil instead. (I just wanted to see whether my architecture
was able to handle Indic rendering issues.)

Regards, Martin.

[shugo (Shugo Maeda)]

• Assignee changed from shugo (Shugo Maeda) to ko1 (Koichi Sasada)

usa (Usaku NAKAMURA) wrote:

1. Error:
   test_refine_recursion(TestRefinement):
   NoMethodError: undefined method recursive_length' for "oo":String C:/Users/usa/ruby/test/ruby/test_refinement.rb:567:in recursive_length' :in `' C:/Users/usa/ruby/test/ruby/test_refinement.rb:806:in `eval' C:/Users/usa/ruby/test/ruby/test_refinement.rb:806:in `eval_using' C:/Users/usa/ruby/test/ruby/test_refinement.rb:574:in `test_refine_recursion'

To avoid an infinite loop by super in a refinement, ci->call is used to distinguish super calls from normal calls,
and if it's a super call, skip the same method.
However, VC++ optimizes vm_call_general and vm_call_super_method into the same method because they have the same definition, so ci->call cannot be used to distinguish super calls from normal calls. How intelligent VC++ is!

I've found that the following hack fixes the problem:

static VALUE
vm_call_super_method(rb_thread_t *th, rb_control_frame_t *reg_cfp, rb_call_info_t ci)
{
#ifdef _WIN32
volatile int x = 0; / to avoid VC++ optimization which makes
vm_call_super_method as an alias of
vm_call_general! */
#endif
return vm_call_method(th, reg_cfp, ci);
}

Sasada-san, do you accept this ugly hack, or do you come up with a neat solution?

[ko1 (Koichi Sasada)]

(2012/12/13 22:32), shugo (Shugo Maeda) wrote:

Sasada-san, do you accept this ugly hack, or do you come up with a neat solution?

To answer your question, I need to learn how refinement is implemented.
Please wait a moment.
(or please commit it ahead and left this ticket open)

--
// SASADA Koichi at atdot dot net

[shugo (Shugo Maeda)]

Hi,

2012/12/13 SASADA Koichi [e-mail]:

Sasada-san, do you accept this ugly hack, or do you come up with a neat solution?

To answer your question, I need to learn how refinement is implemented.
Please wait a moment.
(or please commit it ahead and left this ticket open)

Thanks for your quick response.
I've commited the workaround, and have left the ticket open.

--
Shugo Maeda

[phasis68 (Heesob Park)]

Here is another workaround:

#ifdef _MSC_VER
#pragma optimize( "", off )
#endif
static VALUE
vm_call_super_method(rb_thread_t *th, rb_control_frame_t *reg_cfp, rb_call_info_t *ci)
{
return vm_call_method(th, reg_cfp, ci);
}
#ifdef _MSC_VER
#pragma optimize( "", on )
#endif

[shugo (Shugo Maeda)]

Hello,

phasis68 (Heesob Park) wrote:

Here is another workaround:

#ifdef _MSC_VER
#pragma optimize( "", off )
#endif

Thanks for your suggestion.

But it seems that the #pragma optimize( "", off ) version is slightly slower than the volatile int x = 0 version.

with volatile int x = 0:

C:\Users\shugo\Documents\Source\ruby>\ruby\bin\ruby bm_vm2_super.rb
Rehearsal -----------------------------------------
super 1.778000 0.000000 1.778000 ( 1.783227)
-------------------------------- total: 1.778000sec

        user     system      total        real

super 1.810000 0.000000 1.810000 ( 1.806230)

C:\Users\shugo\Documents\Source\ruby>\ruby\bin\ruby bm_vm2_super.rb
Rehearsal -----------------------------------------
super 1.747000 0.000000 1.747000 ( 1.789727)
-------------------------------- total: 1.747000sec

        user     system      total        real

super 1.763000 0.000000 1.763000 ( 1.760224)

C:\Users\shugo\Documents\Source\ruby>\ruby\bin\ruby bm_vm2_super.rb
Rehearsal -----------------------------------------
super 1.763000 0.000000 1.763000 ( 1.804229)
-------------------------------- total: 1.763000sec

        user     system      total        real

super 1.809000 0.000000 1.809000 ( 1.841734)

with #pragma optimize( "", off ):

C:\Users\shugo\Documents\Source\ruby>\ruby\bin\ruby bm_vm2_super.rb
Rehearsal -----------------------------------------
super 1.825000 0.000000 1.825000 ( 1.859236)
-------------------------------- total: 1.825000sec

        user     system      total        real

super 1.872000 0.000000 1.872000 ( 1.864237)

C:\Users\shugo\Documents\Source\ruby>\ruby\bin\ruby bm_vm2_super.rb
Rehearsal -----------------------------------------
super 1.919000 0.000000 1.919000 ( 1.920744)
-------------------------------- total: 1.919000sec

        user     system      total        real

super 1.794000 0.000000 1.794000 ( 1.820232)

C:\Users\shugo\Documents\Source\ruby>\ruby\bin\ruby bm_vm2_super.rb
Rehearsal -----------------------------------------
super 1.872000 0.000000 1.872000 ( 1.913243)
-------------------------------- total: 1.872000sec

        user     system      total        real

super 1.810000 0.000000 1.810000 ( 1.817231)

I guess #pragma optimize( "", off ) disables function inlining of vm_call_method.

[phasis68 (Heesob Park)]

Here is a modified version which does not disable function inline expansion.

#ifdef _MSC_VER
#pragma optimize("g",off)
#endif
static VALUE
vm_call_super_method(rb_thread_t *th, rb_control_frame_t *reg_cfp, rb_call_info_t *ci)
{
return vm_call_method(th, reg_cfp, ci);
}
#ifdef _MSC_VER
#pragma optimize("g",on)
#endif

[shugo (Shugo Maeda)]

phasis68 (Heesob Park) wrote:

Here is a modified version which does not disable function inline expansion.

#ifdef _MSC_VER
#pragma optimize("g",off)
#endif

I also tried it, but couldn't see improvement:

C:\Users\shugo\Documents\Source\ruby>\ruby\bin\ruby bm_vm2_super.rb
Rehearsal -----------------------------------------
super 1.856000 0.000000 1.856000 ( 1.868237)
-------------------------------- total: 1.856000sec

        user     system      total        real

super 1.872000 0.000000 1.872000 ( 1.910743)

C:\Users\shugo\Documents\Source\ruby>\ruby\bin\ruby bm_vm2_super.rb
Rehearsal -----------------------------------------
super 1.841000 0.000000 1.841000 ( 1.860736)
-------------------------------- total: 1.841000sec

        user     system      total        real

super 1.903000 0.016000 1.919000 ( 1.956249)

C:\Users\shugo\Documents\Source\ruby>\ruby\bin\ruby bm_vm2_super.rb
Rehearsal -----------------------------------------
super 1.872000 0.000000 1.872000 ( 1.927244)
-------------------------------- total: 1.872000sec

        user     system      total        real

super 1.841000 0.000000 1.841000 ( 1.868237)

And, other options of the optimize pragma such as "p" don't fix the problem.

[phasis68 (Heesob Park)]

Here is a different workaround using __forceinline on vm_call_method function.

static
#ifdef _MSC_VER
__forceinline
#else
inline
#endif
VALUE
vm_call_method(rb_thread_t *th, rb_control_frame_t *cfp, rb_call_info_t *ci)
{
...
}

[shugo (Shugo Maeda)]

phasis68 (Heesob Park) wrote:

Here is a different workaround using __forceinline on vm_call_method function.

static
#ifdef _MSC_VER
__forceinline
#else
inline
#endif
VALUE
vm_call_method(rb_thread_t *th, rb_control_frame_t *cfp, rb_call_info_t *ci)

Thanks, it solves the problem without the optimize progma for rb_call_super_method, and performance decrease hasn't been observed.

Could you tell me why __forceinline for vm_call_method prevent VC++ to make vm_call_general and vm_call_super_method as the same function?
I couldn't find the reason at URL:http://msdn.microsoft.com/library/vstudio/z8y1yy88.

[shugo (Shugo Maeda)]

• Status changed from Assigned to Closed
• % Done changed from 0 to 100

This issue was solved with changeset r38377.
Usaku, thank you for reporting this issue.
Your contribution to Ruby is greatly appreciated.
May Ruby be with you.

---

• vm_insnhelper.c (vm_call_super_method): remove volatile introduced
  in r38365.

• vm_insnhelper.c (vm_call_method): use __forceinline to prevent
  VC to make vm_call_general and vm_call_super_method as the same
  method. Thanks, Heesob Park. [Bug #7556] [ruby-core:50867]

[phasis68 (Heesob Park)]

It seems that inline or __inline is not respected by the compiler (ignored by compiler cost/benefit analyzer)

Refer to http://en.wikipedia.org/wiki/Inline_function

[shugo (Shugo Maeda)]

phasis68 (Heesob Park) wrote:

It seems that inline or __inline is not respected by the compiler (ignored by compiler cost/benefit analyzer)

Refer to http://en.wikipedia.org/wiki/Inline_function

I know it, but I don't know why __forceinline prevent user functions (in this case, vm_call_general and vm_call_super_method) to be optimized into a single function.

[phasis68 (Heesob Park)]

2012/12/14 shugo (Shugo Maeda) [e-mail]

Issue #7556 has been updated by shugo (Shugo Maeda).

phasis68 (Heesob Park) wrote:

It seems that inline or __inline is not respected by the compiler
(ignored by compiler cost/benefit analyzer)

Refer to http://en.wikipedia.org/wiki/Inline_function

I know it, but I don't know why __forceinline prevent user functions (in
this case, vm_call_general and vm_call_super_method) to be optimized into a
single function.

The __forceinline keyword overrides the cost/benefit analysis and
relies on the judgment of the programmer instead.
Using __forceinline insures that all functions which call
vm_call_method function have the inline expanded code instead of
vm_call_method function call.
Thus, vm_call_general and vm_call_super_method are separated function.