llvm dev - Oct 2010 - [LLVMdev] Splitting basic blocks while preserving machine code

Hi,

I'm writing a pass that does some static analysis on C code. A  
critical requirement of the pass is that it must preserve not only the  
correctness of the code but also its timing -- It cannot add or remove  
even one CPU cycle from the resulting machine code. To put it another  
way, the binary executable that is produced when the pass is present  
must be identical to what is produced when the pass is not present.

That said, I have run into a situation where it would help to split  
certain basic blocks, probably by calling BasicBlock::splitBasicBlock.  
But this method adds an unconditional branch instruction. Would this  
new instruction alter the machine code that is later generated, or  
would the additional branch somehow be optimized away in subsequent  
passes?

Thanks,

Trevor

On Oct 20, 2010, at 5:16 PM, Trevor Harmon wrote:
> I'm writing a pass that does some static analysis on C code. A  
> critical requirement of the pass is that it must preserve not only the  
> correctness of the code but also its timing -- It cannot add or remove  
> even one CPU cycle from the resulting machine code. To put it another  
> way, the binary executable that is produced when the pass is present  
> must be identical to what is produced when the pass is not present.
> 
> That said, I have run into a situation where it would help to split  
> certain basic blocks, probably by calling BasicBlock::splitBasicBlock.  
> But this method adds an unconditional branch instruction. Would this  
> new instruction alter the machine code that is later generated, or  
> would the additional branch somehow be optimized away in subsequent  
> passes?
It probably would be optimized away, yes, but I would recommend against counting
on it with requirements as strict as yours.

-Jim

On Wed, Oct 20, 2010 at 7:19 PM, Jim Grosbach <grosbach at apple.com>
wrote:
>
> On Oct 20, 2010, at 5:16 PM, Trevor Harmon wrote:
>> I'm writing a pass that does some static analysis on C code. A
>> critical requirement of the pass is that it must preserve not only the
>> correctness of the code but also its timing -- It cannot add or remove
>> even one CPU cycle from the resulting machine code. To put it another
>> way, the binary executable that is produced when the pass is present
>> must be identical to what is produced when the pass is not present.
>>
>> That said, I have run into a situation where it would help to split
>> certain basic blocks, probably by calling BasicBlock::splitBasicBlock.
>> But this method adds an unconditional branch instruction. Would this
>> new instruction alter the machine code that is later generated, or
>> would the additional branch somehow be optimized away in subsequent
>> passes?
>
> It probably would be optimized away, yes, but I would recommend against
counting on it with requirements as strict as yours.
>
> -Jim
> _______________________________________________
> LLVM Developers mailing list
> LLVMdev at cs.uiuc.edu         http://llvm.cs.uiuc.edu
> http://lists.cs.uiuc.edu/mailman/listinfo/llvmdev
>
You could always undo the split right before your pass finishes.