[llvm] r296770 - The patch turns on epilogue unroll for loops with constant recurency start.

[Evgeny Stupachenko via llvm-commits]

Ok. I'll take a look.

On Fri, Jul 21, 2017 at 6:04 PM, Davide Italiano <davide at freebsd.org> wrote:
> On Fri, Jul 21, 2017 at 6:00 PM, Davide Italiano <davide at freebsd.org> wrote:
>> On Thu, Mar 2, 2017 at 9:38 AM, Evgeny Stupachenko via llvm-commits
>> <llvm-commits at lists.llvm.org> wrote:
>>> Author: evstupac
>>> Date: Thu Mar  2 11:38:46 2017
>>> New Revision: 296770
>>>
>>
>> Hi Evgeny,
>> sorry for being late, but this patch triggers a misoptimization at LTO
>> time in some of our internal code.
>>
>> The optimized bitcode reverting the change is:
>>
>> define i32 @main() local_unnamed_addr #0 {
>> fib.exit17:
>>   %0 = tail call i32 (i8*, ...) @printf(i8* getelementptr inbounds ([5
>> x i8], [5 x i8]* @.str, i64
>> 0, i64 0), double 0x3FF9E4129E4129E4)
>>   ret i32 0
>> }
>>
>>
>> while ToT looks like:
>>
>> define i32 @main() local_unnamed_addr #0 {
>>   br label %1
>>
>> ; <label>:1:                                      ; preds = %1, %0
>>   %2 = phi i32 [ %5, %1 ], [ 34, %0 ]
>>   %3 = phi i32 [ %2, %1 ], [ 21, %0 ]
>>   %4 = phi i32 [ %6, %1 ], [ 2, %0 ]
>>   %5 = add nsw i32 %2, %3
>>   %6 = add nsw i32 %4, -1
>>   %7 = icmp eq i32 %6, 0
>>   br i1 %7, label %fib.exit, label %1, !llvm.loop !2
>>
>> fib.exit:                                         ; preds = %1
>>   %8 = sitofp i32 %5 to double
>>   %9 = fdiv double %8, 5.500000e+01
>>   %10 = tail call i32 (i8*, ...) @printf(i8* getelementptr inbounds
>> ([5 x i8], [5 x i8]* @.str, i64
>> 0, i64 0), double %9)
>>   ret i32 0
>> }
>>
>> I'll open a PR with the details on how to reproduce.
>>
>> Thanks,
>>
>> --
>> Davide
>
> We can continue the discussion here https://bugs.llvm.org/show_bug.cgi?id=33885
>
> --
> Davide
>
> "There are no solved problems; there are only problems that are more
> or less solved" -- Henri Poincare