Mauricio Fernandez <mfp@acm.org> writes:

> On Mon, Nov 02, 2009 at 05:11:27PM +0100, Goswin von Brederlow wrote:
>> Richard Jones <rich@annexia.org> writes:
>> 
>> > On Sun, Nov 01, 2009 at 04:11:52PM +0100, Goswin von Brederlow wrote:
>> >> But C calls are still 33% slower than direct access in ocaml (if one
>> >> doesn't use the polymorphic functions).
>> >
>> > Are you using noalloc calls?
>> >
>> > http://camltastic.blogspot.com/2008/08/tip-calling-c-functions-directly-with.html
>> 
>> Yes. And I looked at the bigarray module and couldn't figure out how
>> they differ from my own external function. Only difference I see is
>> the leading "%" on the external name. What does that do?
>
> That means that it is using a hardcoded OCaml primitive, whose code can be
> generated by the compiler via C--. See asmcomp/cmmgen.ml.
>
>> > I would love to see inline assembler supported by the compiler.
>
> It might be possible to hack support for C-- expressions in external
> declarations. That'd be a sort of portable assembler.

This brings me a lot closer to a fast buffer structure. I know have
this code:

(* buffer.ml: Buffer module for libaio-ocaml
 * Copyright (C) 2009 Goswin von Brederlow
 *
 * This program is free software: you can redistribute it and/or modify
 * it under the terms of the GNU Lesser General Public License as
 * published by the Free Software Foundation, either version 3 of the
 * License, or (at your option) any later version.
 *
 * This program is distributed in the hope that it will be useful,
 * but WITHOUT ANY WARRANTY; without even the implied warranty of
 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
 * GNU General Public License for more details.
 *
 * You should have received a copy of the GNU General Public License
 * along with this program.  If not, see <http://www.gnu.org/licenses/>.
 * Under Debian a copy can be found in /usr/share/common-licenses/LGPL-3.
 *)

open Bigarray

type buffer = (int, int8_unsigned_elt, c_layout) Array1.t

exception Unaligned

let create size = (Array1.create int8_unsigned c_layout size : buffer)

let unsafe_get_uint8 (buf : buffer) off = Array1.unsafe_get buf off

let unsafe_get_uint16 (buf : buffer) off =
  let off = off asr 1 in
  let buf = ((Obj.magic buf) : (int, int16_unsigned_elt, c_layout) Array1.t) 
  in
    Array1.unsafe_get buf off

let unsafe_get_int31 (buf : buffer) off =
  let off = off asr 2 in
  let buf = ((Obj.magic buf) : (int32, int32, c_layout) Array1.t) in
  let x = Array1.unsafe_get buf off
  in
    Int32.to_int x

let unsafe_get_int63 (buf : buffer) off =
  let off = off asr 3 in
  let buf = ((Obj.magic buf) : (int, int, c_layout) Array1.t)
  in
    Array1.unsafe_get buf off


Looking at the generated code I see that this works nicely for 8 and
16bit:

0000000000404a50 <camlBuffer__unsafe_get_uint8_131>:
  404a50:       48 d1 fb                sar    %rbx
  404a53:       48 8b 40 08             mov    0x8(%rax),%rax
  404a57:       48 0f b6 04 18          movzbq (%rax,%rbx,1),%rax
  404a5c:       48 8d 44 00 01          lea    0x1(%rax,%rax,1),%rax
  404a61:       c3                      retq   

0000000000404a90 <camlBuffer__unsafe_get_uint16_137>:
  404a90:       48 d1 fb                sar    %rbx
  404a93:       48 83 cb 01             or     $0x1,%rbx
  404a97:       48 d1 fb                sar    %rbx
  404a9a:       48 8b 40 08             mov    0x8(%rax),%rax
  404a9e:       48 0f b7 04 58          movzwq (%rax,%rbx,2),%rax
  404aa3:       48 8d 44 00 01          lea    0x1(%rax,%rax,1),%rax
  404aa8:       c3                      retq   

But for 31/63 bits I get:

0000000000404b90 <camlBuffer__unsafe_get_int31_145>:
  404b90:       48 83 ec 08             sub    $0x8,%rsp
  404b94:       48 c1 fb 02             sar    $0x2,%rbx
  404b98:       48 83 cb 01             or     $0x1,%rbx
  404b9c:       48 89 c7                mov    %rax,%rdi
  404b9f:       48 89 de                mov    %rbx,%rsi
  404ba2:       48 8b 05 5f bc 21 00    mov    0x21bc5f(%rip),%rax        # 620808 <_DYNAMIC+0x7e0>
  404ba9:       e8 92 2a 01 00          callq  417640 <caml_c_call>
  404bae:       48 63 40 08             movslq 0x8(%rax),%rax
  404bb2:       48 d1 e0                shl    %rax
  404bb5:       48 83 c8 01             or     $0x1,%rax
  404bb9:       48 83 c4 08             add    $0x8,%rsp
  404bbd:       c3                      retq   

0000000000404ca0 <camlBuffer__unsafe_get_int63_154>:
  404ca0:       48 83 ec 08             sub    $0x8,%rsp
  404ca4:       48 c1 fb 03             sar    $0x3,%rbx
  404ca8:       48 83 cb 01             or     $0x1,%rbx
  404cac:       48 89 c7                mov    %rax,%rdi
  404caf:       48 89 de                mov    %rbx,%rsi
  404cb2:       48 8b 05 4f bb 21 00    mov    0x21bb4f(%rip),%rax        # 620808 <_DYNAMIC+0x7e0>
  404cb9:       e8 82 29 01 00          callq  417640 <caml_c_call>
  404cbe:       48 83 c4 08             add    $0x8,%rsp
  404cc2:       c3                      retq   

At least in the int63 case I would have thought the compiler would
emit asm code to read the int instead of a function call. In the 31bit
case I would have hoped it would optimize the intermittend int32 away.

Is there something I can do better to get_int31? I was hoping for code
like this:

0000000000404a90 <camlBuffer__unsafe_get_uint31_137>:
  404c90:       48 c1 fb 03             sar    $0x3,%rbx
  404a94:       48 83 cb 01             or     $0x1,%rbx
  404a98:       48 d1 fb                sar    %rbx
  404a9b:       48 8b 40 08             mov    0x8(%rax),%rax
  404a9f:       xx xx xx xx xx          movzwq (%rax,%rbx,4),%rax
  404aa4:       48 8d 44 00 01          lea    0x1(%rax,%rax,1),%rax
  404aa9:       c3                      retq   

MfG
        Goswin