[llvm-dev] missing simplification in ScalarEvolution?

[Chawla, Pankaj via llvm-dev]

Hi,

I have this small test case-

%struct1 = type { i32, i32 }

@glob_const = internal constant [4 x %struct1] [%struct1 { i32 4, i32 5 }, %struct1 { i32 8, i32 9 }, %struct1 { i32 16, i32 0 }, %struct1 { i32 32, i32 10 }], align 16

define void @foo() {
entry:
  br label %loop

loop:                                              ; preds = %loop, %entry
  %iv = phi %struct1* [ getelementptr inbounds ([4 x %struct1], [4 x %struct1]* @glob_const, i64 0, i64 0), %entry ], [ %iv.inc, %loop ]
  %gep = getelementptr inbounds %struct1, %struct1* %iv, i64 0, i32 0
  %ld = load i32, i32* %gep, align 8
  %iv.inc = getelementptr inbounds %struct1, %struct1* %iv, i64 1
  %cmp = icmp ult %struct1* %iv.inc, getelementptr inbounds ([4 x %struct1], [4 x %struct1]* @glob_const, i64 1, i64 0)
  br i1 %cmp, label %loop, label %exit

exit:
  ret void
}

I was expecting the backedge taken count of the loop to be evaluated to 3 by ScalarEvolution but instead it evaluates to this-

Loop %loop: backedge-taken count is ((-1 + (-1 * @glob_const) + ((8 + @glob_const)<nuw><nsw> umax (32 + @glob_const)<nsw>)) /u 8)
Loop %loop: max backedge-taken count is 3, actual taken count either this or zero. 


Can we deduce the final value of the IV (32 + @glob_const) to be greater than the initial value (8 + @glob_const) and simplify the backedge taken count to 3?

Thanks,
Pankaj