[LLVMdev] sign extensions, SCEVs, and wrap flags

[Dan Gohman]

On Sep 20, 2012, at 1:28 AM, Duncan Sands <baldrick at free.fr> wrote:

> Hi,
> 
>>    Sorry, I probably led you astray. No-self-wrap is useful for determining
>>    trip count, but does not mean that sign/zero extension can be hoisted.
>> 
>>    But if you run your analysis after -indvars, the sign-extension should be
>>    removed if possible. The algorithm walks the derived induction variables
>>    specifically looking for add nsw/nuw and replacing only those IR users with
>>    a wide induction variable. See GetExtendedOperandRecurrence.
>> 
>>    If you have a situation where the sign extends cannot removed, then you may
>>    need your own IR-level analysis to determine whether you can ignore them.
>>    Otherwise, you may run into trouble operating on SCEVs that contain
>>    sext/zext/truncs. Are all your array indices uniformly sign-extended? I
>>    don't know if this is a good idea, but why can't you consider the sext
>>    operand the array index rather than the gep operand? If you prove that the
>>    narrow indices are disjoint, then the extended indices must be disjoint.
>>    SCEV operations should work fine on the narrow indices which shouldn't have
>>    any impure type casting operations.
>> 
>>    This can all be avoided by limiting your optimization to code that uses
>>    pointer-size loop counters!
>> 
>>    -Andy
>> 
>> 
>> The array indices are out of my control; depends on the code people write. If
>> they'd use long ints for everything, life would be good; but ints happen.

Random aside: I realize that as a practical engineer, there is value in being
able to just accept certain obstacles as facts of life. However, valuable
engineering insights sometimes depend on recognizing the source of obstacles,
rather than just the ability to cleverly work around them, so there is value
in keeping in mind the fact that this whole situation is caused by a quirk of C
in common usage.

> I thought array indices were promoted to a larger size when possible...  This
> came up with Ada where loops with i8 counters are quite common.  Dan added
> logic to boost the size of the loop counters, but perhaps it doesn't promote
> them beyond i32, or doesn't apply here for some reason?

The pass that does this is -indvars, and it does promote beyond i32 as appropriate
in many common cases, and it sounds like it's working here, from subsequent
messages.

> 
> Ciao, Duncan.
> 
>> 
>> I'll play with -indvars.


Dan