llvm dev - Oct 2007 - [LLVMdev] User-specified JIT passes

Greetings,

I have a plan to use LLVM to undertake runtime optimisation of
computational chemistry programs.  As part of this, I'd like to be able
to invoke lli and pass it a list of passes to run, in the manner of
opt.  For example:
> lli -load=llvm/mem_trace_pass/bin/lib/hello.so -hello ./my_bitcode.bc
The current implementation of lli (updated from SVN as of three nights
ago) doesn't seem to support this.  Too see if it would work, I
transplanted the required functionality from the opt tool, so the above
invocation now works - the pass 'hello' is run by the JIT.  This
has led to a couple of questions:

1. Did I miss something?  As far as I can see, the current JIT
implementation has a hardcoded list of pre-defined passes, which deal with
 bitcode to machine code transformations.  The interpreter doesn't seem
to have anything to do with passes.

2.  Is granting this ability to the JIT desirable?  If so, I'm happy to
clean up my temporary hack job and submit a patch. 

Cheers,
Warren Armstrong

On Oct 28, 2007, at 10:38 PM, Warren Armstrong wrote:
> Greetings,
>
> I have a plan to use LLVM to undertake runtime optimisation of
> computational chemistry programs.
Ok.
> As part of this, I'd like to be able
> to invoke lli and pass it a list of passes to run, in the manner of
> opt.  For example:
>
>> lli -load=llvm/mem_trace_pass/bin/lib/hello.so -hello ./my_bitcode.bc
>
> The current implementation of lli (updated from SVN as of three nights
> ago) doesn't seem to support this.  Too see if it would work, I
> transplanted the required functionality from the opt tool, so the  
> above
> invocation now works - the pass 'hello' is run by the JIT.  This
> has led to a couple of questions:
>
> 1. Did I miss something?
Nope.
> 2.  Is granting this ability to the JIT desirable?  If so, I'm  
> happy to
> clean up my temporary hack job and submit a patch.
Not so much.  "lli" is supposed to be a really simple driver around  
the JIT, it isn't supposed to be a maximally useful user program.

If you'd like a version of lli that runs optimizations, I'd suggest  
taking lli, copying it, and tuning it into the tool that you'd like  
it to be.  I'd like to keep lli simple and focused on what it does,  
which is provide primitive command line access to the jit/interpreter.

Note that you can potentially just use a shell script which does:

opt foo.bc -myopt | lli ...

-Chris

Chris Lattner wrote:
>> 2.  Is granting this ability to the JIT desirable?  If so, I'm  
>> happy to
>> clean up my temporary hack job and submit a patch.
>>     
>
> Not so much.  "lli" is supposed to be a really simple driver
around
> the JIT, it isn't supposed to be a maximally useful user program.
>
> If you'd like a version of lli that runs optimizations, I'd suggest
> taking lli, copying it, and tuning it into the tool that you'd like  
> it to be.  I'd like to keep lli simple and focused on what it does,  
> which is provide primitive command line access to the jit/interpreter.
>
> Note that you can potentially just use a shell script which does:
>
> opt foo.bc -myopt | lli ...
>
> -Chris
>   
Ah, ok.  A shell script wouldn't work, because I'm envisioning
optimisation happening at runtime. 
I'll use your idea of creating a customised version of lli.

Thanks,
Warren