Re: [Emacs-diffs] feature/byte-switch 086c4ea: * src/bytecode.c: (exec_byte_code) Use hash_lookup for Bswitch

Re: [Emacs-diffs] feature/byte-switch 086c4ea: * src/bytecode.c: (exec_byte_code) Use hash_lookup for Bswitch

[Stefan Monnier]

>     * src/bytecode.c: (exec_byte_code) Use hash_lookup for Bswitch
>     Fgethash type checks the provided table object, which is unnecessary
>     for compiled bytecode.

While it's true that we can cause a core dump of Emacs if we feed it an
invalid .elc file, that's a "feature" I'd rather shrink rather
than generalize.


        Stefan

Re: [Emacs-diffs] feature/byte-switch 086c4ea: * src/bytecode.c: (exec_byte_code) Use hash_lookup for Bswitch

[Vibhav Pant]

[-- Attachment #1: Type: text/plain, Size: 472 bytes --]

Will push a fix soon, thanks.

On 18 Jan 2017 11:18 p.m., "Stefan Monnier" <monnier@iro.umontreal.ca>
wrote:

> >     * src/bytecode.c: (exec_byte_code) Use hash_lookup for Bswitch
> >     Fgethash type checks the provided table object, which is unnecessary
> >     for compiled bytecode.
>
> While it's true that we can cause a core dump of Emacs if we feed it an
> invalid .elc file, that's a "feature" I'd rather shrink rather
> than generalize.
>
>
>         Stefan
>

[-- Attachment #2: Type: text/html, Size: 833 bytes --]

Re: [Emacs-diffs] feature/byte-switch 086c4ea: * src/bytecode.c: (exec_byte_code) Use hash_lookup for Bswitch

[Vibhav Pant]

On another thought, `byte-switch` is used while compiling certain
`cond` forms. It
replaces the traditional goto-if-nil bytecode by using a hash table mapping
values to addresses/tags to be jumped to. Since `byte-switch` is essentially a
"dynamic" goto (in the sense that the address/tag cannot be known at
compile time),
wouldn't doing a runtime type check in the bytecode VM for what is  a
hash table lookup + goto have a significant performance penalty?

On Wed, Jan 18, 2017 at 11:18 PM, Stefan Monnier
<monnier@iro.umontreal.ca> wrote:
>>     * src/bytecode.c: (exec_byte_code) Use hash_lookup for Bswitch
>>     Fgethash type checks the provided table object, which is unnecessary
>>     for compiled bytecode.
>
> While it's true that we can cause a core dump of Emacs if we feed it an
> invalid .elc file, that's a "feature" I'd rather shrink rather
> than generalize.
>
>
>         Stefan



-- 
Vibhav Pant
vibhavp@gmail.com

Re: [Emacs-diffs] feature/byte-switch 086c4ea: * src/bytecode.c: (exec_byte_code) Use hash_lookup for Bswitch

[Stefan Monnier]

> On another thought, `byte-switch` is used while compiling certain
> `cond` forms. It
> replaces the traditional goto-if-nil bytecode by using a hash table mapping
> values to addresses/tags to be jumped to. Since `byte-switch` is essentially a
> "dynamic" goto (in the sense that the address/tag cannot be known at
> compile time),
> wouldn't doing a runtime type check in the bytecode VM for what is  a
> hash table lookup + goto have a significant performance penalty?

I don't know.  Only measurement can tell.  My guess is that hash_lookup
already takes a significant amount of time, so a HASH_TABLE_P test would be
negligible in comparison.


        Stefan


> On Wed, Jan 18, 2017 at 11:18 PM, Stefan Monnier
> <monnier@iro.umontreal.ca> wrote:
>>> * src/bytecode.c: (exec_byte_code) Use hash_lookup for Bswitch
>>> Fgethash type checks the provided table object, which is unnecessary
>>> for compiled bytecode.
>> 
>> While it's true that we can cause a core dump of Emacs if we feed it an
>> invalid .elc file, that's a "feature" I'd rather shrink rather
>> than generalize.
>> 
>> 
>> Stefan



> -- 
> Vibhav Pant
> vibhavp@gmail.com

Re: [Emacs-diffs] feature/byte-switch 086c4ea: * src/bytecode.c: (exec_byte_code) Use hash_lookup for Bswitch

[Vibhav Pant]

[-- Attachment #1: Type: text/plain, Size: 1516 bytes --]

I see. For now, I've made the type check only work when BYTE_CODE_SAFE is
defined at compile time. I'll make it a part of the regular code once the
byte-switch is implemented and there are no performance issues with doing
so.

On Thu, 19 Jan 2017, 22:04 Stefan Monnier, <monnier@iro.umontreal.ca> wrote:

> > On another thought, `byte-switch` is used while compiling certain
> > `cond` forms. It
> > replaces the traditional goto-if-nil bytecode by using a hash table
> mapping
> > values to addresses/tags to be jumped to. Since `byte-switch` is
> essentially a
> > "dynamic" goto (in the sense that the address/tag cannot be known at
> > compile time),
> > wouldn't doing a runtime type check in the bytecode VM for what is  a
> > hash table lookup + goto have a significant performance penalty?
>
> I don't know.  Only measurement can tell.  My guess is that hash_lookup
> already takes a significant amount of time, so a HASH_TABLE_P test would be
> negligible in comparison.
>
>
>         Stefan
>
>
> > On Wed, Jan 18, 2017 at 11:18 PM, Stefan Monnier
> > <monnier@iro.umontreal.ca> wrote:
> >>> * src/bytecode.c: (exec_byte_code) Use hash_lookup for Bswitch
> >>> Fgethash type checks the provided table object, which is unnecessary
> >>> for compiled bytecode.
> >>
> >> While it's true that we can cause a core dump of Emacs if we feed it an
> >> invalid .elc file, that's a "feature" I'd rather shrink rather
> >> than generalize.
> >>
> >>
> >> Stefan
>
>
>
> > --
> > Vibhav Pant
> > vibhavp@gmail.com
>

[-- Attachment #2: Type: text/html, Size: 2819 bytes --]

Re: [Emacs-diffs] feature/byte-switch 086c4ea: * src/bytecode.c: (exec_byte_code) Use hash_lookup for Bswitch

[Daniel Colascione]

On Wed, Jan 18 2017, Stefan Monnier wrote:
>>     * src/bytecode.c: (exec_byte_code) Use hash_lookup for Bswitch
>>     Fgethash type checks the provided table object, which is unnecessary
>>     for compiled bytecode.
>
> While it's true that we can cause a core dump of Emacs if we feed it an
> invalid .elc file, that's a "feature" I'd rather shrink rather
> than generalize.

Why? If anything, the opposite direction is the right one. elc files
should be tightly coupled to the Emacs that built them so that they can
be specialized for specific machines.  You could imagine them
containing, say, saved JIT traces.

Re: [Emacs-diffs] feature/byte-switch 086c4ea: * src/bytecode.c: (exec_byte_code) Use hash_lookup for Bswitch

[Stefan Monnier]

> Why? If anything, the opposite direction is the right one. elc files
> should be tightly coupled to the Emacs that built them so that they can
> be specialized for specific machines.  You could imagine them
> containing, say, saved JIT traces.

But that basically means auto-generating and managing the compiled
files, so it's very different from what happens with our .elc files.

I would also like to move in the direction you suggest, tho (and
get rid of the Elisp interpreter altogether, while we're at it), and
we'd then indeed be more free to change the format of byte-code.


        Stefan