[llvm-commits] [llvm] r155884 - in /llvm/trunk: lib/Transforms/Scalar/LICM.cpp test/Transforms/LICM/speculate.ll

[Nick Lewycky]

Eli Friedman wrote:
> On Mon, Apr 30, 2012 at 9:03 PM, Nick Lewycky<nicholas at mxc.ca>  wrote:
>> Author: nicholas
>> Date: Mon Apr 30 23:03:01 2012
>> New Revision: 155884
>>
>> URL: http://llvm.org/viewvc/llvm-project?rev=155884&view=rev
>> Log:
>> An instruction in a loop is not guaranteed to be executed just because the loop
>> has no exit blocks. Fixes PR12706!
>>
>> Modified:
>>     llvm/trunk/lib/Transforms/Scalar/LICM.cpp
>>     llvm/trunk/test/Transforms/LICM/speculate.ll
>>
>> Modified: llvm/trunk/lib/Transforms/Scalar/LICM.cpp
>> URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Transforms/Scalar/LICM.cpp?rev=155884&r1=155883&r2=155884&view=diff
>> ==============================================================================
>> --- llvm/trunk/lib/Transforms/Scalar/LICM.cpp (original)
>> +++ llvm/trunk/lib/Transforms/Scalar/LICM.cpp Mon Apr 30 23:03:01 2012
>> @@ -618,6 +618,11 @@
>>      if (!DT->dominates(Inst.getParent(), ExitBlocks[i]))
>>        return false;
>>
>> +  // As a degenerate case, if the loop is statically infinite then we haven't
>> +  // proven anything since there are no exit blocks.
>> +  if (ExitBlocks.empty())
>> +    return false;
>
> Even if there are exit blocks, you still haven't proven anything; take
> something like the following, assuming the compiler doesn't know
> anything about exit():
>
> int f(int x, int y) {
>    int n = 0;
>    for (int i = 0; i<  10; i++) {
>      exit(0);
>      n += x / (y+i);
>    }
>    return n;
> }

Great point Eli. I was struggling with a different aspect of its design 
(why use exit blocks vs. exiting blocks) but missed this bug. I don't 
have the time to fix it, so I'd appreciate it if you want to take over!

Nick