[PATCH] [PATCH] Fix a bug about memory dependence checking in loop vectorization

[Hao Liu]

Hi Jiangning,

I think the test case can be simplified so that we can understand the problem easily. I've simplified it into

  define void @test_loop_novect(double** %Array_.i, i64 %indvars.iv1438) {
  for.body209.lr.ph:
    %t275 = load double** %Array_.i, align 8
    br label %for.body209
  
  for.body209:                                      ; preds = %for.body209, %for.body209.lr.ph
    %indvars.iv1436 = phi i64 [ 0, %for.body209.lr.ph ], [ %indvars.iv.next1437, %for.body209 ]
    %arrayidx.i954 = getelementptr inbounds double* %t275, i64 %indvars.iv1436
    %indvars.iv.next1437 = add nuw nsw i64 %indvars.iv1436, 1
    %arrayidx.i997 = getelementptr inbounds double* %t275, i64 %indvars.iv.next1437
    %t282 = load double* %arrayidx.i954, align 8
    %t284 = load double* %arrayidx.i997, align 8
    store double %t282, double* %arrayidx.i997, align 8
    store double %t284, double* %arrayidx.i954, align 8
    %exitcond1441 = icmp eq i64 %indvars.iv1436, %indvars.iv1438
    br i1 %exitcond1441, label %invoke.cont239, label %for.body209
  
  invoke.cont239:                                   ; preds = %for.body209
    ret void
  }

Then it's easier to know that the case is like

  tmp1 = a[i]
  tmp2 = a[i+1]
  a[i+1] = tmp1
  a[i] = tmp2

Besides that, I think the fix itself is reasonable and LGTM.


REPOSITORY
  rL LLVM

http://reviews.llvm.org/D6735

EMAIL PREFERENCES
  http://reviews.llvm.org/settings/panel/emailpreferences/