[PATCH] D141481: [SCEV] Canonicalize ext(min/max(x, y)) to min/max(ext(x), ext(y))

[Eli Friedman via Phabricator via llvm-commits]

efriedma added a comment.

I'm seeing a regression on the following testcase:

  define void @foo() {
  entry:
    br label %for.body
  
  for.body:
    %indvars.iv696 = phi i64 [ 0, %entry ], [ %indvars.iv.next, %for.body ]
    %x3 = trunc i64 %indvars.iv696 to i32
    %x4 = add i32 %x3, 143
    %x5 = icmp ult i32 %x4, 302
    %cond355 = select i1 %x5, i32 %x4, i32 302
    %idxprom363 = zext i32 %cond355 to i64
    %indvars.iv.next = add nuw nsw i64 %indvars.iv696, 1
    %exitcond = icmp eq i64 %indvars.iv.next, 160
    br i1 %exitcond, label %exit, label %for.body
  
  exit:
    ret void
  }

`opt '-passes=print<scalar-evolution>'` used to print the following:

  %idxprom363 = zext i32 %cond355 to i64
  -->  {143,+,1}<nuw><%for.body> U: [143,303) S: [143,303)              Exits: 302              LoopDispositions: { %for.body: Computable }

Now it prints

  %idxprom363 = zext i32 %cond355 to i64
  -->  (zext i32 (302 umin {143,+,1}<%for.body>) to i64) U: [0,303) S: [0,303)          Exits: 302              LoopDispositions: { %for.body: Computable }

Is this expected?


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D141481/new/

https://reviews.llvm.org/D141481