clang-15 outputs better code than clang-16

[k-arrows]

Test case is the following trivial function.

int foo(int a, int b)
{
    return a - b ? b - a : b - a ;
}

https://godbolt.org/z/MexYMfeK1
https://alive2.llvm.org/ce/z/9Ma_ob

Clang 15.0.0

foo(int, int):                               # @foo(int, int)
        mov     eax, esi
        sub     eax, edi
        ret

Clang 16.0.0

foo(int, int):                               # @foo(int, int)
        xor     eax, eax
        sub     esi, edi
        cmovne  eax, esi
        ret

If this change is as expected, feel free to close this issue.

[junaire]

I believe we should blame InstCombine pass. In LLVM 15 the optimizer can merge all the same code of the 2 branches and then eliminate the phi node. However, in LLVM 16 it can't fully merge the code so the branch was kept.

Original IR:

define dso_local noundef i32 @foo(int, int)(i32 noundef %a, i32 noundef %b) local_unnamed_addr {
entry:
  %sub = sub nsw i32 %a, %b
  %tobool = icmp ne i32 %sub, 0
  br i1 %tobool, label %cond.true, label %cond.false

cond.true:                                        ; preds = %entry
  %sub1 = sub nsw i32 %b, %a
  br label %cond.end

cond.false:                                       ; preds = %entry
  %sub2 = sub nsw i32 %b, %a
  br label %cond.end

cond.end:                                         ; preds = %cond.false, %cond.true
  %cond = phi i32 [ %sub1, %cond.true ], [ %sub2, %cond.false ]
  ret i32 %cond
}

LLVM 15:

define dso_local noundef i32 @foo(int, int)(i32 noundef %a, i32 noundef %b) local_unnamed_addr {
entry:
  %tobool.not = icmp eq i32 %b, %a
  br i1 %tobool.not, label %cond.false, label %cond.true

cond.true:                                        ; preds = %entry
  br label %cond.end

cond.false:                                       ; preds = %entry
  br label %cond.end

cond.end:                                         ; preds = %cond.false, %cond.true
  %cond = sub nsw i32 %b, %a
  ret i32 %cond
}

LLVM 16:

define dso_local noundef i32 @foo(int, int)(i32 noundef %a, i32 noundef %b) local_unnamed_addr {
entry:
  %tobool.not = icmp eq i32 %b, %a
  br i1 %tobool.not, label %cond.false, label %cond.true

cond.true:                                        ; preds = %entry
  %sub1 = sub nsw i32 %b, %a       ; <===== This code is kept so we can't eliminate the phi node.
  br label %cond.end

cond.false:                                       ; preds = %entry
  br label %cond.end

cond.end:                                         ; preds = %cond.false, %cond.true
  %cond = phi i32 [ %sub1, %cond.true ], [ 0, %cond.false ]
  ret i32 %cond
}

[junaire]

My guess is that when the optimizer is in cond.false (%a == %b), it constant folds %sub2 = sub nsw i32 %b, %a to 0 then it loses further optimization opportunity.

[DianQK]

Hi @junaire, if you are interested in this issue, you can check out https://reviews.llvm.org/D148979. If you're not interested, I'll probably come back to this after I've addressed the other issues.

[junaire]

Candidate patch: https://reviews.llvm.org/D150378