[LLVMdev] Operating on contents of virtual registers

[Eli Friedman]

On Wed, Mar 16, 2011 at 10:44 PM, Ashay Rane <ashay.rane at tacc.utexas.edu> wrote:
> Thanks Eli, that worked!
> For the record, I had to make one change though. I had to bitcast to i32*
> (instead of i8*) otherwise I was seeing errors like:
> Instruction referencing instruction not embedded in a basic block!
>   %retval = alloca i32
>   <badref> = bitcast i32* %retval to i8*
> and sometimes:
> Instruction does not dominate all uses!
> That makes me curious to ask, what does the choice of i32 or i8 depend on?
> Is it the architecture?

A bitcast from i32* to i32* gets optimized out. :)  I think you're
missing a call to IRBuilder::SetInsertPoint.

-Eli

> Ashay
>
> On Wed, Mar 16, 2011 at 11:40 PM, Eli Friedman <eli.friedman at gmail.com>
> wrote:
>>
>> On Wed, Mar 16, 2011 at 9:33 PM, Ashay Rane <ashay.rane at asu.edu> wrote:
>> > Hi Eli,
>> > Thanks for the reply. The problem is that getOperand() returns an
>> > llvm::Instruction (that refers to the definition of the operand). What I
>> > am
>> > trying to find out is how to get the value of the operand. When you
>> > refer to
>> > bitcasting to i8*, do you mean casting the return value from
>> > getOperand()
>> > itself?
>> > Ashay
>>
>> I mean calling IRBuilder::CreateBitCast with the return value from
>> getOperand().
>>
>> -Eli
>>
>> > On Wed, Mar 16, 2011 at 11:03 PM, Eli Friedman <eli.friedman at gmail.com>
>> > wrote:
>> >>
>> >> On Wed, Mar 16, 2011 at 6:00 PM, Ashay Rane <ashay.rane at asu.edu> wrote:
>> >> > Hello,
>> >> > I was facing some difficulty in implementing a transform and I was
>> >> > wondering
>> >> > if I could get some help please.
>> >> > The transform needs to operate on the operands of certain
>> >> > instructions.
>> >> > For
>> >> > example, given an instruction, say "%10 = load i32* %9, align 4", I
>> >> > have
>> >> > to
>> >> > record the value of %9 and process it. Of course, this is only
>> >> > possible
>> >> > at
>> >> > runtime and so I am instrumenting the code such that a particular
>> >> > function
>> >> > is invoked just before the instruction of interest is executed.
>> >> > The problem that I am facing is in getting the value of the operands.
>> >> > As
>> >> > I
>> >> > understand, the operands could either be program variables (e.g. in
>> >> > "%6
>> >> > =
>> >> > load i32* %old, align 4") or one of the virtual registers (as in the
>> >> > first
>> >> > load instruction). For both cases, is it possible to extract the
>> >> > value/address of the operand (%9 or %old)?
>> >>
>> >> Those are both the same case; names for instructions only exist for
>> >> the sake of readability.
>> >>
>> >> > If yes, what should be the best
>> >> > function type to use so that I can pass this value as an argument to
>> >> > my
>> >> > function? For now, I am only concerned with load and store
>> >> > instructions
>> >> > and
>> >> > so (I suppose) I have to deal with pointers only.
>> >>
>> >> If you only need pointers, just use i8* as the type of the argument,
>> >> and bitcast the value to i8*.
>> >>
>> >> > An easy way that I can think of is to directly insert the LLVM IR
>> >> > (e.g. call
>> >> > void @my_function(%old)) but because I am using
>> >> > Module::getOrInsertFunction(), I have to have a function type. So
>> >> > alternatively, is there a way to insert direct LLVM instructions
>> >> > (without
>> >> > going through the type hierarchy)?
>> >>
>> >> Your approach of inserting a call is fine; making your own instruction
>> >> or intrinsic wouldn't make things any simpler.  There aren't any
>> >> shortcuts here.
>> >>
>> >> -Eli
>
>