[LLVMdev] replacing a global variable by a constant

[Ralf Karrenberg]

Thanks a lot for your help Matthijs! :)

basically this does the job quite nicely I think:
for (llvm::GlobalVariable::use_iterator U = gv->use_begin(); U !=
gv->use_end(); ++U) {
    llvm::Instruction *I = llvm::cast<llvm::Instruction>(U);
    I->replaceAllUsesWith(constPtr);
    I->eraseFromParent();
}

Cheers,
Ralf

Matthijs Kooijman wrote:
> Hi Ralf,
>
>   
>> I am trying to replace a global variable with a constant.
>>     
> You should be able to replace the initializer for the global with your
> constant, mark it as having internal linkage and then run some optimization
> pass over it to do the rest (not sure which one it was again, probably some
> constant propagation or global value optimization).
>
> That pass would then probably do some similar iterating as you've described,
> but at least you won't have to maintain that part of code :-)
>
>   
>> I did manage to do it, but somehow it appears to be fairly bruteforce by
>> just iterating over all functions/bblocks/instructions and filtering for
>>  those that load the variable and replacing the instruction with
>> Instruction::replaceAllUsesWith().
>> The more intuitive way of iterating over the uses of the variable did
>> not work out as I have to replace the whole instruction instead of just
>> the operand (as GlobalVariable::replaceAllUsesWith() does) in order to
>> prevent loading from the memory address of the constant instead of
>> loading the constant :p. Basically what seems to be missing is some
>> method that returns the parent-instruction of a Use.
>>     
> You can use Use::getUser(), which returns a User (from which Instruction and
> LoadInst derive). Note that you will have to accoutn for other uses than Loads
> (such as the initializers of other globals), but in those cases your
> transformation can probably not proceed anyway.
>
>   
>> Does an invokation of a JIT through runFunction() perform any
>> optimization(s) by itself? 
>>     
> I don't thinks so, not sure.
>
>   
>> Is there a way to display the generated code?
>>     
> No clue :-)
>
> Gr.
>
> Matthijs
>   
> ------------------------------------------------------------------------
>
> _______________________________________________
> LLVM Developers mailing list
> LLVMdev at cs.uiuc.edu         http://llvm.cs.uiuc.edu
> http://lists.cs.uiuc.edu/mailman/listinfo/llvmdev
>