From: Basile Starynkevitch <basile@starynkevitch.net>
To: David Malcolm <dmalcolm@redhat.com>
Cc: jit@gcc.gnu.org
Subject: Re: GCCJIT and tail-recursive calls
Date: Thu, 01 Jan 2015 00:00:00 -0000 [thread overview]
Message-ID: <559FE5FB.7080503@starynkevitch.net> (raw)
In-Reply-To: <1436539272.31573.58.camel@surprise>
On 07/10/2015 04:41 PM, David Malcolm wrote:
> This is done by: gcc/tree-tailcall.c controlled by:
> -foptimize-sibling-calls which is on by default at -02 and above.
> Normally you'd do something like: gcc_jit_block_end_with_return (...
> gcc_jit_context_new_call (...)); or gcc_jit_block_add_eval (...
> gcc_jit_context_new_call (...)); gcc_jit_block_end_with_void return
> (...); The tailcall is obvious in both of these cases.
>>>> I was thinking of adding a construct where the GCCJIT client knows
>>>> for sure that the call should be tail-rec (and if GCCJIT was not able to
>>>> make a tailcall, that would be an error).
>>> It's not clear to me why that would be useful. Am I missing something?
>>
>> You are coding a JIT for your implementation of Scheme.
>> Scheme requires that tailcails are effectively ilmplemented as such.
> What do you mean by the word "requires" here?
>
> performance? that if enough tailcalls aren't optimized, then things
> will be painfully slow?
>
> legal? that if every tailcall isn't optimized, then some trademark
> owner will send a Cease-and-Desist letter?
No, technically only.
IIRC, Scheme specification has words that don't allow a stack overflow
for a very deep tail-recursive function call.
>> So your Scheme implementation has to detect tailcalls and want to be
>> sure that GCCJIT
>> is implementing them as needed.
> Can't you just trust the optimizer?
>
>> How would you ensure that?
> >From a testing perspective, I suppose you could pass in:
> -fdump-tree-tailr1-details -fdump-tree-tailr2-details
> and then use:
> https://gcc.gnu.org/onlinedocs/jit/topics/contexts.html#gcc_jit_context_enable_dump
>
> to capture these dumpfiles in memory and somehow analyze them. I use
> this approach in one of the jit testcase to verify that an optimization
> does take place (similar to how many of gcc's DejaGnu testcases work).
Ok.
> Another approach might be to add an attribute to the call, saying "must
> be handled as a tail call", and have that pass issue an error if it
> can't do it. Maybe via a builtin that can wrap calls and tags them as
> such? That would require some work from the gcc side.
Yes, I was thinking of something similar...
Thanks!
--
Basile STARYNKEVITCH http://starynkevitch.net/Basile/
email: basile<at>starynkevitch<dot>net mobile: +33 6 8501 2359
8, rue de la Faiencerie, 92340 Bourg La Reine, France
*** opinions {are only mine, sont seulement les miennes} ***
prev parent reply other threads:[~2015-07-10 15:34 UTC|newest]
Thread overview: 5+ messages / expand[flat|nested] mbox.gz Atom feed top
2015-01-01 0:00 Basile Starynkevitch
2015-01-01 0:00 ` David Malcolm
2015-01-01 0:00 ` Basile Starynkevitch
2015-01-01 0:00 ` David Malcolm
2015-01-01 0:00 ` Basile Starynkevitch [this message]
Reply instructions:
You may reply publicly to this message via plain-text email
using any one of the following methods:
* Save the following mbox file, import it into your mail client,
and reply-to-all from there: mbox
Avoid top-posting and favor interleaved quoting:
https://en.wikipedia.org/wiki/Posting_style#Interleaved_style
* Reply using the --to, --cc, and --in-reply-to
switches of git-send-email(1):
git send-email \
--in-reply-to=559FE5FB.7080503@starynkevitch.net \
--to=basile@starynkevitch.net \
--cc=dmalcolm@redhat.com \
--cc=jit@gcc.gnu.org \
/path/to/YOUR_REPLY
https://kernel.org/pub/software/scm/git/docs/git-send-email.html
* If your mail client supports setting the In-Reply-To header
via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line
before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for read-only IMAP folder(s) and NNTP newsgroup(s).