En utilisant l'exemple de mukunda ci-dessus, dans la catégorie 3.4 avec -O2, le problème semble être en phase de vectorisation. Le code vectorisé saute à l'entrée pour dépasser le code vectorisé:
br i1 true, label %middle.block, label %vector.ph
Ainsi, la valeur de count
reste inchangée depuis son initialisation.
*** IR Dump Before Combine redundant instructions ***
; Function Attrs: nounwind readnone ssp uwtable
define i32 @main() #0 {
entry:
br i1 true, label %middle.block, label %vector.ph
vector.ph: ; preds = %entry
br label %vector.body
vector.body: ; preds = %vector.body, %vector.ph
%index = phi i32 [ 0, %vector.ph ], [ %index.next, %vector.body ]
%vec.phi = phi <4 x i32> [ <i32 1, i32 0, i32 0, i32 0>, %vector.ph ], [ %4, %vector.body ]
%vec.phi8 = phi <4 x i32> [ zeroinitializer, %vector.ph ], [ %5, %vector.body ]
%broadcast.splatinsert = insertelement <4 x i32> undef, i32 %index, i32 0
%broadcast.splat = shufflevector <4 x i32> %broadcast.splatinsert, <4 x i32> undef, <4 x i32> zeroinitializer
%induction = add <4 x i32> %broadcast.splat, <i32 0, i32 1, i32 2, i32 3>
%induction7 = add <4 x i32> %broadcast.splat, <i32 4, i32 5, i32 6, i32 7>
%0 = icmp ult <4 x i32> %induction, <i32 5, i32 5, i32 5, i32 5>
%1 = icmp ult <4 x i32> %induction7, <i32 5, i32 5, i32 5, i32 5>
%2 = zext <4 x i1> %0 to <4 x i32>
%3 = zext <4 x i1> %1 to <4 x i32>
%4 = add <4 x i32> %2, %vec.phi
%5 = add <4 x i32> %3, %vec.phi8
%6 = icmp eq <4 x i32> %induction, <i32 -1, i32 -1, i32 -1, i32 -1>
%7 = icmp eq <4 x i32> %induction7, <i32 -1, i32 -1, i32 -1, i32 -1>
%8 = add <4 x i32> %induction, <i32 1, i32 1, i32 1, i32 1>
%9 = add <4 x i32> %induction7, <i32 1, i32 1, i32 1, i32 1>
%index.next = add i32 %index, 8
%10 = icmp eq i32 %index.next, 0
br i1 %10, label %middle.block, label %vector.body, !llvm.loop !1
middle.block: ; preds = %vector.body, %entry
%resume.val = phi i32 [ 0, %entry ], [ 0, %vector.body ]
%trunc.resume.val = phi i32 [ 0, %entry ], [ 0, %vector.body ]
%rdx.vec.exit.phi = phi <4 x i32> [ <i32 1, i32 0, i32 0, i32 0>, %entry ], [ %4, %vector.body ]
%rdx.vec.exit.phi9 = phi <4 x i32> [ zeroinitializer, %entry ], [ %5, %vector.body ]
%bin.rdx = add <4 x i32> %rdx.vec.exit.phi9, %rdx.vec.exit.phi
%rdx.shuf = shufflevector <4 x i32> %bin.rdx, <4 x i32> undef, <4 x i32> <i32 2, i32 3, i32 undef, i32 undef>
%bin.rdx10 = add <4 x i32> %bin.rdx, %rdx.shuf
%rdx.shuf11 = shufflevector <4 x i32> %bin.rdx10, <4 x i32> undef, <4 x i32> <i32 1, i32 undef, i32 undef, i32 undef>
%bin.rdx12 = add <4 x i32> %bin.rdx10, %rdx.shuf11
%11 = extractelement <4 x i32> %bin.rdx12, i32 0
%cmp.n = icmp eq i32 0, %resume.val
br i1 %cmp.n, label %while.end, label %scalar.ph
scalar.ph: ; preds = %middle.block
br label %while.body
while.body: ; preds = %while.body, %scalar.ph
%i.0 = phi i32 [ %trunc.resume.val, %scalar.ph ], [ %inc, %while.body ]
%count.0 = phi i32 [ %11, %scalar.ph ], [ %add.count.0, %while.body ]
%cmp = icmp ult i32 %i.0, 5
%add = zext i1 %cmp to i32
%add.count.0 = add i32 %add, %count.0
%cmp1 = icmp eq i32 %i.0, -1
%inc = add i32 %i.0, 1
br i1 %cmp1, label %while.end, label %while.body, !llvm.loop !4
while.end: ; preds = %middle.block, %while.body
%add.count.0.lcssa = phi i32 [ %add.count.0, %while.body ], [ %11, %middle.block ]
ret i32 %add.count.0.lcssa
}
L'optimiseur efface ensuite le code inaccessible et inefficace, qui représente presque le corps de la fonction.