[llvm-commits] [llvm] r124724 - in /llvm/trunk: lib/Target/README.txt lib/Transforms/Utils/SimplifyCFG.cpp test/Transforms/SimplifyCFG/switch_create.ll test/Transforms/SimplifyCFG/switch_formation.dbg.ll

[Benjamin Kramer]

On 02.02.2011, at 21:11, Frits van Bommel wrote:

> On Wed, Feb 2, 2011 at 4:56 PM, Benjamin Kramer
> <benny.kra at googlemail.com> wrote:
>> +/// TurnSwitchRangeIntoICmp - Turns a switch with that contains only a
>> +/// integer range comparison into a sub, an icmp and a branch.
>> +static bool TurnSwitchRangeIntoICmp(SwitchInst *SI) {
>> +  assert(SI->getNumCases() > 2 && "Degenerate switch?");
>> +  // We can do this transform if the switch consists of an ascending series
>> +  // and all cases point to the same destination.
>> +  for (unsigned I = 2, E = SI->getNumCases(); I != E; ++I)
>> +    if (SI->getSuccessor(I-1) != SI->getSuccessor(I) ||
>> +        SI->getCaseValue(I-1)->getValue()+1 != SI->getCaseValue(I)->getValue())
>> +      return false;
> 
> What about out-of-order case values? For example:
> 
> define zeroext i1 @quux(i32 %x) nounwind readnone ssp noredzone {
> entry:
>  switch i32 %x, label %lor.rhs [
>    i32 1, label %lor.end
>    i32 3, label %lor.end
>    i32 2, label %lor.end
>  ]
> 
> lor.rhs:                                          ; preds = %entry
>  br label %lor.end
> 
> lor.end:                                          ; preds = %lor.rhs,
> %entry, %entry, %entry
>  %0 = phi i1 [ true, %entry ], [ false, %lor.rhs ], [ true, %entry ],
> [ true, %entry ]
>  ret i1 %0
> }
> 
> This is supposed to be equivalent to using ordered case values, but
> your code doesn't touch this switch while it works fine if the case
> values are sorted.

I thought transforming unsorted switches wouldn't have much impact on usual C code but testing shows that it does. I extended the transform to catch unsorted cases too in r124826.