[llvm-dev] Query in implementation of Simple GVN algorithm

Dimpal Gurabani via llvm-dev llvm-dev at lists.llvm.org
Mon Jan 8 21:07:04 PST 2018


This is Dimpal from NIT Calicut. I am currently implementing Simple GVN
algorithm given by  Nabizath Saleena and Vineeth Paleri [
https://arxiv.org/abs/1303.1880] in LLVM.
Our aim is to preserve semantics of the program by taking any conservative
assumptions.
For example, in case of call instruction we made a conservative assumption
that after the call Expression Pool is made empty. So, we may miss some
redundancies but there will be no harm to the semantics of program.
But we found that in case of pointers semantics is not preserved. If we
assign pointer to some variable than it is putting pointer and variable in
the same class so, whenever change in variable occurs using pointer it is
not behaving as per semantics.
For example, if we have this c program,

int *p, c=2;
p = &c;
*p = 1;

then my code will find the Pools given below after every instruction.

store i32 2, i32* %c, align 4
{ [ c : c , 2  ] }

store i32* %c, i32** %p, align 8
{ [ c : c , 2 , p  ] }

%tmp = load i32*, i32** %p, align 8
{ [ c : c , 2 , p , tmp  ] }

store i32 1, i32* %tmp, align 4
{ [ c : c , 2 , p  ] ,[ tmp : tmp , 1  ] }


As, we can see in above code that, we wanted to change the value of c but
it's not behaving in that way. It is treating pointer and variable in the
same way. So, Could you please help me in finding the "easiest way" to deal
with pointers with consideration of conservativeness?

Your response will be appreciated.

-- 
Thanks and Regards,
Dimpal Gurabani
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20180109/abaff721/attachment.html>


More information about the llvm-dev mailing list