Bug 51866 – Use __builtin_expect when compiling using RVCT in GNU mode

• Component: Web Template Framework
• Status: RESOLVED FIXED

[Daniel Bates]

When in GNU mode, RVCT supports the GNU language intrinsic __builtin_expect, which is used to provide the compiler with branch prediction information. We should favor using this intrinsic if we are compiling with RVCT in GNU mode (i.e. COMPILER(RVCT) && defined(__GNUC__)).

[Daniel Bates]

Created attachment 77867 [details]
Patch

[Darin Adler]

Comment on attachment 77867 [details]
Patch

OK

[Laszlo Gombos]

Looks good to me too. A simpler way to express this is to test for defined(__GNUC__), instead of "COMPILER(GCC) || (COMPILER(RVCT) && defined(__GNUC__))".

[Daniel Bates]

(In reply to comment #3)
> Looks good to me too. A simpler way to express this is to test for defined(__GNUC__), instead of "COMPILER(GCC) || (COMPILER(RVCT) && defined(__GNUC__))".

Will change before I land.

[Daniel Bates]

Committed r74973: <http://trac.webkit.org/changeset/74973>

[Siddharth Mathur]

Created attachment 78423 [details]
Exclude RVCT2.2 from using this intrinsic

According to ARM's documentation (http://infocenter.arm.com/help/index.jsp?topic=/com.arm.doc.dui0202h/Cjabddedbde.html), this GNU-style intrinsic is only available with RVCT 3 and later , so exclude previous major version from using it.

[Siddharth Mathur]

Created attachment 78425 [details]
Same patch as above, but with Changelog format issue corrected.

[Eric Seidel (no email)]

Comment on attachment 78425 [details]
Same patch as above, but with Changelog format issue corrected. 

rs=me.

[Eric Seidel (no email)]

Comment on attachment 77867 [details]
Patch

I'm confused by the status of this bug.  Did this get landed?

[Siddharth Mathur]

(In reply to comment #9)
> (From update of attachment 77867 [details])
> I'm confused by the status of this bug.  Did this get landed?

Sorry Eric for the confusion. I reopened a previously closed bug (manually landed by Daniel Bates) to add an amendment, which you just r+ed.

[Daniel Bates]

Comment on attachment 78425 [details]
Same patch as above, but with Changelog format issue corrected. 

Why are we hardcoding __ARMCC_2_2__ instead of checking __ARMCC_VERSION >= 300000?

[Daniel Bates]

(In reply to comment #11)
> (From update of attachment 78425 [details])
> Why are we hardcoding __ARMCC_2_2__ instead of checking __ARMCC_VERSION >= 300000?

I mean checking that __ARMCC_VERSION <= 300000

[Daniel Bates]

(In reply to comment #12)
> (In reply to comment #11)
> > (From update of attachment 78425 [details] [details])
> > Why are we hardcoding __ARMCC_2_2__ instead of checking __ARMCC_VERSION >= 300000?
> 
> I mean checking that __ARMCC_VERSION <= 300000

Disregard this comment. The inequality in the previous comment (comment 11) was correct. That is, we should change the check from:

#if defined(__GNUC__)

to

#if COMPILER(GCC) || (COMPILER(RVCT) && defined(__GNUC__) && __ARMCC_VERSION >= 300000)

so that we only use __builtin_expect if we are compiling with GCC or compiling with RVCT version 3.0 or higher in GNU mode.

[Daniel Bates]

Created attachment 78461 [details]
Patch

[WebKit Review Bot]

Attachment 78461 [details] did not pass style-queue:

Failed to run "['Tools/Scripts/check-webkit-style', '--diff-files', u'Source/JavaScriptCore/ChangeLog', u'Source..." exit_code: 1
Source/JavaScriptCore/wtf/Platform.h:1016:  Missing space after ,  [whitespace/comma] [3]
Source/JavaScriptCore/wtf/AlwaysInline.h:44:  Missing space after ,  [whitespace/comma] [3]
Source/JavaScriptCore/wtf/AlwaysInline.h:52:  Missing space after ,  [whitespace/comma] [3]
Total errors found: 3 in 3 files


If any of these errors are false positives, please file a bug against check-webkit-style.

[Darin Adler]

Comment on attachment 78461 [details]
Patch

View in context: https://bugs.webkit.org/attachment.cgi?id=78461&action=review

> Source/JavaScriptCore/wtf/Platform.h:79
> +/* Define this for !RVCT compilers, just so we can write things like COMPILER(RVCT) && RVCT_VERSION_AT_LEAST(3,0,0,0). */

It occurs to me that this means that you could always just use RVCT_VERSION_AT_LEAST(3,0,0,0) without COMPILER(RVCT); maybe we should.

> Source/JavaScriptCore/wtf/Platform.h:91
> -/* define this for !GCC compilers, just so we can write things like COMPILER(GCC) && GCC_VERSION_AT_LEAST(4,1,0) */
> +/* Define this for !GCC compilers, just so we can write things like COMPILER(GCC) && GCC_VERSION_AT_LEAST(4,1,0). */
>  #define GCC_VERSION_AT_LEAST(major, minor, patch) 0

It occurs to me that this means that you could always just use GCC_VERSION_AT_LEAST(3,0,0,0) without COMPILER(GCC); maybe we should.

[Daniel Bates]

(In reply to comment #16)
> (From update of attachment 78461 [details])
> View in context: https://bugs.webkit.org/attachment.cgi?id=78461&action=review
> 
> > Source/JavaScriptCore/wtf/Platform.h:79
> > +/* Define this for !RVCT compilers, just so we can write things like COMPILER(RVCT) && RVCT_VERSION_AT_LEAST(3,0,0,0). */
> 
> It occurs to me that this means that you could always just use RVCT_VERSION_AT_LEAST(3,0,0,0) without COMPILER(RVCT); maybe we should.

Correct. Will change both comment and callsites before landing.

> 
> > Source/JavaScriptCore/wtf/Platform.h:91
> > -/* define this for !GCC compilers, just so we can write things like COMPILER(GCC) && GCC_VERSION_AT_LEAST(4,1,0) */
> > +/* Define this for !GCC compilers, just so we can write things like COMPILER(GCC) && GCC_VERSION_AT_LEAST(4,1,0). */
> >  #define GCC_VERSION_AT_LEAST(major, minor, patch) 0
> 
> It occurs to me that this means that you could always just use GCC_VERSION_AT_LEAST(3,0,0,0) without COMPILER(GCC); maybe we should.

Correct. Will change comment before landing. I suggest that we update the callsites in a separate bug.

[Daniel Bates]

Committed r75447: <http://trac.webkit.org/changeset/75447>

[Daniel Bates]

(In reply to comment #17)
> (In reply to comment #16)
> > It occurs to me that this means that you could always just use GCC_VERSION_AT_LEAST(3,0,0,0) without COMPILER(GCC); maybe we should.
> 
> Correct. Will change comment before landing. I suggest that we update the callsites in a separate bug.

Filed bug #52178 for this and decided to make the comment change in the patch for this bug.