A modest proposal: eliminate template code bloat

[Dmitry Olshansky]

On 09.04.2012 5:11, Daniel Murphy wrote:
> "Dmitry Olshansky"<dmitry.olsh at gmail.com>  wrote in message
> news:jlsmka$22ce$1 at digitalmars.com...
>>
>> The refinement is merging prefixes and suffixes of course.
>> And for that one needs to calculate hashes for all of prefixes and all of
>> suffixes. I will define _all_ later on.
>>
>
> I think you'll find that this is better done in the compiler instead of the
> linker.  Merging prefixes is problematic because at some point you will need
> to work out which tail to execute, so you'll always need to modify the
> generated code.

"Easy": just add a hidden pointer argument to functions that have merged 
prefix (call it dispatch). The prefix part of code is followed by an 
indirect jump to this pointer.
Compiler arranges so that every time function is called the correct 
dispatch address is passed behind the scenes.

BTW there are no extra checks and such it's one naked indirect jump, and 
it's totally predictable unlike say switch jump.
(well unless you use a few copy-paste-susceptible functions in the same 
loop that turn out to have prefixes merged)

It still implies that prefix merging should be applied with more care 
then suffix merging. Once this tested and working even merging arbitrary 
parts of functions is doable with this approach.

>
> Merging suffixes is easier, you can merge all returning blocks with
> identical code, and then merge all blocks that always jump to the same
> blocks, etc.
> This will need to happen after code generation if you want to merge int/uint
> code, which would be difficult in dmd's back end.
>

Any chance to fit this into IR-->CodeGen step? Like use alternative 
comparison then memcmp. Or better do basically the same algorithm but on 
map!(....)(IR) that morphs things so that some identical ops (e.g. 
uint/int == and !=) are considered the same. In fact this might be even 
faster then generating useless machine code!

> Merging functions with identical bodies is of course much easier, and can be
> done in the linker without needing to modify any code (just the
> relocations).
>
>


-- 
Dmitry Olshansky