[PATCH] D113349: [SCEV] Fix and validate ValueExprMap/ExprValueMap consistency

[Nikita Popov via Phabricator via llvm-commits]

nikic added a comment.

@sberg Thanks for the report! Here's a reduction for `opt -passes='print<scalar-evolution>'`:

  define void @test(i32* %p) {
  entry:
    br label %loop.header
  
  loop.header:
    %iv = phi i32 [ 0, %entry ], [ %iv.next, %loop.latch ]
    br label %loop2
  
  loop2:
    %iv2 = phi i32 [ %iv, %loop.header ], [ %iv2.next, %loop2 ]
    %iv2.next = add i32 %iv2, 1
    %v = load i32, i32* %p
    %cmp = icmp slt i32 %iv2, %v
    br i1 %cmp, label %loop2, label %loop2.end
  
  loop2.end:
    %iv2.ext = sext i32 %iv2 to i64
    br label %loop3
  
  loop3:
    %iv3 = phi i64 [ %iv2.ext, %loop2.end ], [ %iv3.next, %loop3 ]
    %iv3.next = add nsw i64 %iv3, 1
    br i1 false, label %loop3, label %loop.latch
  
  loop.latch:
    %iv.next = trunc i64 %iv3 to i32
    br label %loop.header
  }

I'll probably just drop the assert -- while it should be the same in an ideal world, lazy nowrap flag updates mean that it can be a different (but equivalent) SCEV expression, as is the case here.


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D113349/new/

https://reviews.llvm.org/D113349