[llvm-commits] [llvm] r132732 - /llvm/trunk/utils/TableGen/FastISelEmitter.cpp

[Chad Rosier]

On Jun 10, 2011, at 12:05 PM, David Blaikie wrote:

> Ahh.. yes, I misunderstood your example.  Here are the source comments, which best explain the use of llvm_unreachable:
> 
> /// Marks that the current location is not supposed to be reachable.
> /// In !NDEBUG builds, prints the message and location info to stderr.
> /// In NDEBUG builds, becomes an optimizer hint that the current location
> /// is not supposed to be reachable.  On compilers that don't support
> /// such hints, prints a reduced message instead.
> ///
> /// Use this instead of assert(0).  It conveys intent more clearly and
> /// allows compilers to omit some unnecessary code.
> In the case of a release build an unreachable hint is emitted.  The compiler then uses this hint to transform:
> 
>       } else if (Operands[i].isFP()) {
>         OS << "ConstantFP *f" << i;
>       } else {
>         llvm_unreachable("Unknown operand kind!");
>       }
> 
> to:
> 
>       } else if (Operands[i].isFP()) {
>         OS << "ConstantFP *f" << i;
>       }
> 
> I hope this helps answer your question.  In the end the compiler is being smart about removing the unreachable code.  For more details you could also checkout: http://www.nondot.org/sabre/LLVMNotes/UnreachableInstruction.txt
> 
> Ah, ok - thanks for the explanation. I didn't realize llvm_unreachable was that advanced/involved (though I suppose I Should've expected as much). 
> 
> So both versions (the assert & the unreachable) seem to have similar semantics (performance & correctness) given a sufficiently advanced compiler (that uses the unreachable to remove the isFP() call) - makes sense.
> 
> [is there any policy/recommendations on which way to go in this case? I guess it's not a problem to leave this up to the author of any particular piece of LLVM code about which expression they prefer/find more clear in a given situation]

Like the comment says, llvm_unreachable() should be used in lieu of assert (0).  Otherwise, assert (cond && "msg") should be used for code that is expected to execute.

> Thanks again for the help,
> - David 
> 

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20110610/e0aba5c5/attachment.html>