[PATCH] D99249: [PassManager] Run additional LICM before LoopRotate

[Roman Lebedev via Phabricator via llvm-commits]

lebedev.ri added a comment.

I'm not sure how to write proper interestingness test for this, but e.g. given this input F16057899: reduced.ll <https://reviews.llvm.org/F16057899> (bad example, in the end everything optimizes away)
currently (`-looprotate -licm`):

  *** IR Dump After LICMPass ***
  ; Preheader:
  for.end.lr.ph:                                    ; preds = %entry
    %hue = getelementptr inbounds %"class.rawspeed::Cr2sRawInterpolator", %"class.rawspeed::Cr2sRawInterpolator"* %this, i64 0, i32 3
    %1 = load i32, i32* %hue, align 4, !tbaa !8
    %Cb.i.i = getelementptr inbounds %"struct.std::array.30", %"struct.std::array.30"* %MCUs, i64 0, i32 0, i64 undef, i32 0, i64 undef, i32 1
    %sub.i.i = add i32 %1, -16384
    %Cb.i.i.promoted = load i32, i32* %Cb.i.i, align 4, !tbaa !9
    br label %for.end
  
  ; Loop:
  for.end:                                          ; preds = %for.end.lr.ph, %for.end
    %2 = phi i32 [ %Cb.i.i.promoted, %for.end.lr.ph ], [ %add.i.i, %for.end ]
    %MCUIdx.03 = phi i32 [ undef, %for.end.lr.ph ], [ %inc21, %for.end ]
    %add.i.i = add i32 %sub.i.i, %2
    %inc21 = add nsw i32 %MCUIdx.03, 1
    %cmp = icmp slt i32 %inc21, %sub
    br i1 %cmp, label %for.end, label %for.cond.for.end22_crit_edge, !llvm.loop !11
  
  ; Exit blocks
  for.cond.for.end22_crit_edge:                     ; preds = %for.end
    %add.i.i.lcssa = phi i32 [ %add.i.i, %for.end ]
    store i32 %add.i.i.lcssa, i32* %Cb.i.i, align 4, !tbaa !9
    br label %for.end22

with `-licm -loop-rotate`:

  *** IR Dump After LoopRotatePass ***
  ; Preheader:
  for.end.lr.ph:                                    ; preds = %entry
    br label %for.end
  
  ; Loop:
  for.end:                                          ; preds = %for.end.lr.ph, %for.end
    %MCUIdx.03 = phi i32 [ undef, %for.end.lr.ph ], [ %inc21, %for.end ]
    %1 = load i32, i32* %hue, align 4, !tbaa !8
    %2 = load i32, i32* %Cb.i.i, align 4, !tbaa !9
    %sub.i.i = add i32 %1, -16384
    %add.i.i = add i32 %sub.i.i, %2
    store i32 %add.i.i, i32* %Cb.i.i, align 4, !tbaa !9
    %inc21 = add nsw i32 %MCUIdx.03, 1
    %cmp = icmp slt i32 %inc21, %sub
    br i1 %cmp, label %for.end, label %for.cond.for.end22_crit_edge, !llvm.loop !11
  
  ; Exit blocks
  for.cond.for.end22_crit_edge:                     ; preds = %for.end
    br label %for.end22

and with `-licm -loop-rotate -licm`:

  *** IR Dump After LICMPass ***
  ; Preheader:
  for.end.lr.ph:                                    ; preds = %entry
    %1 = load i32, i32* %hue, align 4, !tbaa !8
    %sub.i.i = add i32 %1, -16384
    %Cb.i.i.promoted = load i32, i32* %Cb.i.i, align 4, !tbaa !9
    br label %for.end
  
  ; Loop:
  for.end:                                          ; preds = %for.end.lr.ph, %for.end
    %2 = phi i32 [ %Cb.i.i.promoted, %for.end.lr.ph ], [ %add.i.i, %for.end ]
    %MCUIdx.03 = phi i32 [ undef, %for.end.lr.ph ], [ %inc21, %for.end ]
    %add.i.i = add i32 %sub.i.i, %2
    %inc21 = add nsw i32 %MCUIdx.03, 1
    %cmp = icmp slt i32 %inc21, %sub
    br i1 %cmp, label %for.end, label %for.cond.for.end22_crit_edge, !llvm.loop !11
  
  ; Exit blocks
  for.cond.for.end22_crit_edge:                     ; preds = %for.end
    %add.i.i.lcssa = phi i32 [ %add.i.i, %for.end ]
    store i32 %add.i.i.lcssa, i32* %Cb.i.i, align 4, !tbaa !9
    br label %for.end22

Note how the `-licm -loop-rotate` causes `%sub.i.i` to be computed in-loop.
Do tell if that isn't sufficient of an answer.


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D99249/new/

https://reviews.llvm.org/D99249