[LLVMdev] How to change the type of an Instruction?
Nick Lewycky
nicholas at mxc.ca
Fri Jan 28 18:15:27 PST 2011
On 01/28/2011 03:01 PM, Douglas do Couto Teixeira wrote:
> Hi, guys,
>
> Thanks a lot for your help. As you know, I am trying to implement
> something to change the types of the instructions. And I chose the
> trunc's approach because it seems be simple. But I still have some
> problems and questions. Would be great if you can help me.
>
> I have used the results of my range analysis implementation to
> change the intermediate representation. I am using the very simple
> expedient of inserting a trunc right after the definition of a small
> variable, and inverse truncation right after the uses of this variable.
> I am doing this for correctness. So, if I have something like:
>
> a = b + c
> ...
> x = a + y
> ...
> w = a - z
>
> Then I change the program into:
>
> a = b + c
> a_small = trunc(a)
> ...
> a1 = truncInv(a_small)
> x = a1 + y
> ...
> a2 = truncInv(a_small)
> w = a2 - z
>
> Again, I did it like this just to check for correctness. But, even if I
> optimize the insertion of truncs, it seems that the end code will not be
> good, and I am afraid I will miss many optimization opportunities.
You won't, just run llvm's optimizer. For example, there's something
called SimplifyDemandedBits which asserts that the only bits needed to
perform a given computation are in a particular bitmask, and will look
at the chain of instructions which feed into it and simplify them as
much as possible. If you just insert your truncates, you should be able
to run opt -O2 over it and get all the optimizations we can legally do.
Nick
Like,
> one of the first things that would pay off would be small peephole
> optmizations, such as removing unnecessary truncation, or using small
> multiplication whenever possible, and eliminating instructions such as x
> = y & 0xFF, whenever I know that y < 0xFF. So, what I would like to ask
> you is this: is it right to do bitwidth analysis at the intermediate
> representation level, or should I re-implement my analysis at the
> machine level? The type system seems to be getting in my way at the
> intermediate level. If I go to the machine level, can I avoid this types
> of problems?
>
> Thank you a lot,
>
> Douglas
>
>
> On Mon, Jan 24, 2011 at 4:24 PM, John Criswell <criswell at illinois.edu
> <mailto:criswell at illinois.edu>> wrote:
>
> On 1/24/11 12:05 PM, Douglas do Couto Teixeira wrote:
>>
>>
>> On Mon, Jan 24, 2011 at 3:01 PM, Nick Lewycky <nicholas at mxc.ca
>> <mailto:nicholas at mxc.ca>> wrote:
>>
>> On 01/24/2011 04:41 AM, Douglas do Couto Teixeira wrote:
>>
>> Hi,
>>
>> Nick, thanks for the reply.
>> I still have a problem: I only need to "clone" an
>> Instruction, changing
>> its type. That is, I would like to keep all
>> characteristics of the old
>> Instruction and create a new one only with a different type.
>>
>>
>> Sure, but what about its operands? An "add" instruction must
>> have the same type as its operands, what do you want to do
>> with them?
>>
>>
>> I also need to convert the type of the operands. But I want to do
>> this when they are created instead of inserting "trunc"
>> instructions before performing an operation. But it seems hard to me.
>
> Actually, I don't think it will be that difficult. You basically
> need to do the following:
>
> 1) Take the backwards, intra-procedural slice of the instruction
> (i.e., find the instructions, operands, the operands' operands, the
> operands' operands' operands, etc.).
>
> 2) Visit all of the instructions in the slice and convert them. You
> want to visit definitions before uses. To do that,
> a) Make new phi instructions for all phis in the slice. The
> operands of the phis should be the Undef value.
> b) Use the dominator tree analysis and traverse basic blocks in
> dominator tree order (i.e., start at the top of the dominator tree
> and process each node breadth first). Convert all of the
> instructions in each basic block (except phis).
> c) Revisit all the phis and plug in their new operands.
>
> 3) Delete all the old instructions.
>
> This is a variation of the SSA construction algorithm in Zadeck et.
> al.'s paper (Effeciently Computing Single Static Assignment Form and
> the Control Dependence Graph).
>
> The only tricky part is handling non-instruction operands (e.g.,
> function arguments, global variables, etc.). Some might be trivial
> to convert. Others may be difficult. You should look over all the
> classes derived from llvm::Value and decide how difficult it would
> be to convert them.
>
> -- John T.
>
>
>
>> Suppose you're going from a 32-bit add to a 64-bit add, do the
>> old operands get zero extended? Sign extended? You'll need to
>> insert instructions for that (unless they're constants in
>> which case you can use constant expressions). Similarly, what
>> if the old type is a float and the new one is an int? float to
>> signed int, float to unsigned int, or bitcast (only legal
>> sometimes)?
>>
>>
>> I believe I don't need worry about it because I only create
>> smaller instructions. So I never convert a 32-bit instruction in a
>> 64-bit instruction.
>>
>>
>>
>> I am trying
>>
>> create a new Instruction thus:
>>
>> %3 = add nsw i32 %1, %2 ; <i16> [#uses=2] //Old Instruction
>>
>> Value* Op0 = I->getOperand(0);
>> Value* Op1 = I->getOperand(1);
>> Value* V0 = new Value(Type::getInt16Ty(Op0->getContext()),
>> Op0->getValueID());
>>
>>
>> Hunh, Value's constructor is protected.
>>
>> In any event, Value is pure base. Constructing one this way
>> will never get you what you want. If the ValueID indicates an
>> Instruction, go through Instruction to create one.
>>
>>
>> Value* V1 = new Value(Type::getInt16Ty(Op1->getContext()),
>> Op1->getValueID());
>> Instruction* newInst = BinaryOperator::CreateNSWAdd(V0,
>> V1, "test");
>> errs() << "NewInst:\n" << *newInst << "\n";
>>
>>
>> But I get something like this:
>>
>> %test = add nsw i16 <badref>, <badref> ; <i16> [#uses=0]
>>
>>
>> The two instructions V0 and V1 you created were never inserted
>> into the BasicBlock so they can't be numbered, and also they
>> don't have names.
>>
>>
>> What I am doing wrong?
>>
>>
>> Suppose that you're going from i32 to i16. Your only choice
>> with that particular pair of types is a truncate. So:
>>
>> IRBuilder builder(OldInst);
>> Value *V0 = builder.CreateTrunc(Op0, Type::getInt16Ty());
>> Value *V1 = builder.CreateTrunc(Op1, Type::getInt16Ty());
>> Value *Add = builder.CreateNSWAdd(V0, V1, "test");
>>
>> The IRBuilder will take care of the distinction between
>> instructions and constants for you. Note that I have not
>> tested the above code, it may need some fixing before it compiles.
>>
>> Nick
>>
>>
>>
>> Best,
>>
>> Douglas
>>
>> On Fri, Jan 21, 2011 at 8:25 PM, Nick Lewycky
>> <nlewycky at google.com <mailto:nlewycky at google.com>
>> <mailto:nlewycky at google.com <mailto:nlewycky at google.com>>>
>> wrote:
>>
>> On 21 January 2011 12:56, Douglas do Couto Teixeira
>> <douglasdocouto at gmail.com
>> <mailto:douglasdocouto at gmail.com>
>> <mailto:douglasdocouto at gmail.com
>> <mailto:douglasdocouto at gmail.com>>> wrote:
>>
>> Hello guys,
>>
>> I wonder how I can change the type of an integer
>> variable. For
>> instance, given the instruction "%3 = add i32 %1,
>> %2" I would
>> like to alter the instruction to "%3 = add i16 %1,
>> %2". Is there
>> any way to do this?
>>
>>
>> No. Instead you create a new Instruction, in this case with
>> BinaryOperator::CreateAdd, then
>> OldInst->replaceAllUsesWith(NewInst)
>> to update all the users, then
>> OldInst->eraseFromParent() since it's
>> now dead code.
>>
>> Also, all values have types immutably assigned at
>> creation, so
>> you'll need to insert casts (trunc instructions in your
>> case) to
>> cast %1 and %2 from i32 to i16 for the smaller add.
>>
>> Nick
>>
>>
>>
>>
>> _______________________________________________
>> LLVM Developers mailing list
>> LLVMdev at cs.uiuc.edu <mailto:LLVMdev at cs.uiuc.edu>
>> http://llvm.cs.uiuc.edu
>> http://lists.cs.uiuc.edu/mailman/listinfo/llvmdev
>>
>>
>>
>
>
>
>
> _______________________________________________
> LLVM Developers mailing list
> LLVMdev at cs.uiuc.edu http://llvm.cs.uiuc.edu
> http://lists.cs.uiuc.edu/mailman/listinfo/llvmdev
More information about the llvm-dev
mailing list