code review 4650072: hash/crc32: add SSE4.2 support

hash/crc32: add SSE4.2 support

Using the CRC32 instruction speeds up the Castagnoli computation by
about 20x on a modern Intel CPU.

[agl1, 2011-07-06 20:07:07]

*** Abandoned ***

[agl1, 2011-07-06 20:12:28]

Hello rsc@golang.org (cc: golang-dev@googlegroups.com),

I'd like you to review this change to
https://go.googlecode.com/hg/

[rsc, 2011-07-11 19:50:02]

Only substantive thing is the handling of REX

http://codereview.appspot.com/4650072/diff/4023/src/cmd/6a/lex.c
File src/cmd/6a/lex.c (right):

http://codereview.appspot.com/4650072/diff/4023/src/cmd/6a/lex.c#newcode172
src/cmd/6a/lex.c:172: /* type is the lexical type to return.  It dictates what
kind of
please make the new comments start with /* and end with */ on lines
by themselves

http://codereview.appspot.com/4650072/diff/4023/src/cmd/6l/optab.c
File src/cmd/6l/optab.c (right):

http://codereview.appspot.com/4650072/diff/4023/src/cmd/6l/optab.c#newcode537
src/cmd/6l/optab.c:537: /* You are doasm, holding in your hand a Prog* with
p->as set to, say, ACRC32,
same

http://codereview.appspot.com/4650072/diff/4023/src/cmd/6l/optab.c#newcode1265
src/cmd/6l/optab.c:1265: { ACRC32Q,       ycrc32l,Px,
0xf2,0x48,0x0f,0x38,0xf1,0},
You should be able to write

{ ACRC32L, ycrc32l, Px, 0xf2, 0x0f, 0x38, 0xf1, 0},
{ ACRC32Q, ycrc32l, Pw, 0xf2, 0x0f, 0x38, 0xf1, 0},

(note the Pw) and then 6l will put the REX prefix in for you
as appropriate, without loss of registers.  That is assuming
there is a CRC32L form of the instruction.  (My point is
that if you use Pw 6l will do the right thing with the REX byte.)

[agl1, 2011-07-12 13:30:49]

*** Submitted as http://code.google.com/p/go/source/detail?r=cc56eaa003b6 ***

hash/crc32: add SSE4.2 support

Using the CRC32 instruction speeds up the Castagnoli computation by
about 20x on a modern Intel CPU.

R=rsc
CC=golang-dev
http://codereview.appspot.com/4650072

http://codereview.appspot.com/4650072/diff/4023/src/cmd/6a/lex.c
File src/cmd/6a/lex.c (right):

http://codereview.appspot.com/4650072/diff/4023/src/cmd/6a/lex.c#newcode172
src/cmd/6a/lex.c:172: /* type is the lexical type to return.  It dictates what
kind of
On 2011/07/11 19:50:02, rsc wrote:
> please make the new comments start with /* and end with */ on lines
> by themselves

Done.

http://codereview.appspot.com/4650072/diff/4023/src/cmd/6l/optab.c
File src/cmd/6l/optab.c (right):

http://codereview.appspot.com/4650072/diff/4023/src/cmd/6l/optab.c#newcode537
src/cmd/6l/optab.c:537: /* You are doasm, holding in your hand a Prog* with
p->as set to, say, ACRC32,
On 2011/07/11 19:50:02, rsc wrote:
> same

Done.

http://codereview.appspot.com/4650072/diff/4023/src/cmd/6l/optab.c#newcode1265
src/cmd/6l/optab.c:1265: { ACRC32Q,       ycrc32l,Px,
0xf2,0x48,0x0f,0x38,0xf1,0},
On 2011/07/11 19:50:02, rsc wrote:
> You should be able to write
> 
> { ACRC32L, ycrc32l, Px, 0xf2, 0x0f, 0x38, 0xf1, 0},
> { ACRC32Q, ycrc32l, Pw, 0xf2, 0x0f, 0x38, 0xf1, 0},
> 
> (note the Pw) and then 6l will put the REX prefix in for you
> as appropriate, without loss of registers.  That is assuming
> there is a CRC32L form of the instruction.  (My point is
> that if you use Pw 6l will do the right thing with the REX byte.)

Done.

[rsc, 2011-07-12 14:27:59]

just for my own curiosity,
is there a 32-bit argument form of the CRC32 instruction?

[agl1, 2011-07-12 14:29:04]

On Tue, Jul 12, 2011 at 10:27 AM, Russ Cox <rsc@golang.org> wrote:
> just for my own curiosity,
> is there a 32-bit argument form of the CRC32 instruction?

Yes, and a 16-bit one. I didn't add them because I wasn't using them
and so had not testing coverage, but I can put them in if you like.


Cheers

AGL

[rsc, 2011-07-12 14:32:45]

On Tue, Jul 12, 2011 at 10:29, Adam Langley <agl@golang.org> wrote:
> On Tue, Jul 12, 2011 at 10:27 AM, Russ Cox <rsc@golang.org> wrote:
>> just for my own curiosity,
>> is there a 32-bit argument form of the CRC32 instruction?
>
> Yes, and a 16-bit one. I didn't add them because I wasn't using them
> and so had not testing coverage, but I can put them in if you like.

that'd be great.  even if untested they're
better than nothing.   thanks.