From: "Юрий Соколов" <funny.falcon@...>
Date: 2014-03-18T13:43:24+04:00
Subject: [ruby-core:61576] Re: [ruby-trunk - Feature #9425] [PATCH] st: use power-of-two sizes to avoid slow modulo ops

--047d7beb9b800f5ac304f4de5edb
Content-Type: text/plain; charset=ISO-8859-1

Can you test this
 + if (SYMBOL_P(a)) a = (a >> RUBY_SPECIAL_SHIFT) ^ (a >>
(RUBY_SPECIAL_SHIFT + 3)); /* 3=ID_SCOPE_SHIFT */


2014-03-18 13:02 GMT+04:00 Eric Wong <normalperson@yhbt.net>:

> normalperson@yhbt.net wrote:
> >       hash_aref_sym   1.000
>
> Lack of improvement here was disappointing since symbol keys are common,
> and this showed a regression on my x86 (32-bit) VMs.
>
> I tweaked rb_any_hash to be symbol-aware:
>
>         http://bogomips.org/ruby.git/patch?id=497ed6355
>
> 12-30% improvement on this test from trunk depending on CPU so far \o/
> (Phenom II X4 got 30%, newer/faster x86-64 CPUs show less speedup).
>
> I'm comfortable with improvements of this series on x86 VMs running on
> x86-64 (and of course native x86-64).
>
> Can anybody with real 32-bit hardware verify this series?  Not sure I
> can trust VM results; my remaining x86 hardware is on its last legs
> and showing occasional HW errors.
>
> git://80x24.org/ruby.git st-noprime-v4
>
>       st: use power-of-two sizes to avoid slow modulo ops
>       add hash benchmarks
>       st.c: tweak numhash to match common Ruby use cases
>       hash.c: improve symbol hash distribution
>

--047d7beb9b800f5ac304f4de5edb
Content-Type: text/html; charset=ISO-8859-1
Content-Transfer-Encoding: quoted-printable

<div dir=3D"ltr">Can you test this<div>=A0+<span class=3D"" style=3D"white-=
space:pre">	</span>if (SYMBOL_P(a)) a =3D (a &gt;&gt; RUBY_SPECIAL_SHIFT) ^=
 (a &gt;&gt; (RUBY_SPECIAL_SHIFT + 3)); /* 3=3DID_SCOPE_SHIFT */</div></div=
><div class=3D"gmail_extra">
<br><br><div class=3D"gmail_quote">2014-03-18 13:02 GMT+04:00 Eric Wong <sp=
an dir=3D"ltr">&lt;<a href=3D"mailto:normalperson@yhbt.net" target=3D"_blan=
k">normalperson@yhbt.net</a>&gt;</span>:<br><blockquote class=3D"gmail_quot=
e" style=3D"margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex">
<a href=3D"mailto:normalperson@yhbt.net">normalperson@yhbt.net</a> wrote:<b=
r>
&gt; =A0 =A0 =A0 hash_aref_sym =A0 1.000<br>
<br>
Lack of improvement here was disappointing since symbol keys are common,<br=
>
and this showed a regression on my x86 (32-bit) VMs.<br>
<br>
I tweaked rb_any_hash to be symbol-aware:<br>
<br>
=A0 =A0 =A0 =A0 <a href=3D"http://bogomips.org/ruby.git/patch?id=3D497ed635=
5" target=3D"_blank">http://bogomips.org/ruby.git/patch?id=3D497ed6355</a><=
br>
<br>
12-30% improvement on this test from trunk depending on CPU so far \o/<br>
(Phenom II X4 got 30%, newer/faster x86-64 CPUs show less speedup).<br>
<br>
I&#39;m comfortable with improvements of this series on x86 VMs running on<=
br>
x86-64 (and of course native x86-64).<br>
<br>
Can anybody with real 32-bit hardware verify this series? =A0Not sure I<br>
can trust VM results; my remaining x86 hardware is on its last legs<br>
and showing occasional HW errors.<br>
<br>
git://<a href=3D"http://80x24.org/ruby.git" target=3D"_blank">80x24.org/rub=
y.git</a> st-noprime-v4<br>
<div class=3D""><br>
=A0 =A0 =A0 st: use power-of-two sizes to avoid slow modulo ops<br>
</div>=A0 =A0 =A0 add hash benchmarks<br>
=A0 =A0 =A0 st.c: tweak numhash to match common Ruby use cases<br>
=A0 =A0 =A0 hash.c: improve symbol hash distribution<br>
</blockquote></div><br></div>

--047d7beb9b800f5ac304f4de5edb--