clang -O3 for循环的LLVM IR

O3都是怪物，这里分析的是CLANG怪物，示例程序遍历数组每个元素然后放大。

void foreach_scale(int arr[],int elem){
	for(int i=0;i<elem;i++){
		arr[i] += (elem*1024);
	}
}

这里删去了用处不大的内容，只保留了关键的LLVM IR。通过分析可以看到，如果循环小于8 LLVM IR会使用vector，vector使用SIMD指令高效进行计算，如果大于8则是普通的for形式。

; Function Attrs: norecurse nounwind
define void @"\01?foreach_scale@@YAXQAHH@Z"(i32* nocapture %arr, i32 %elem) local_unnamed_addr #0 {
entry:
  ;elem>0则进入循环，否则整个函数结束
  %cmp5 = icmp sgt i32 %elem, 0
  br i1 %cmp5, label %for.body.lr.ph, label %for.cond.cleanup

for.body.lr.ph:
  %mul = shl i32 %elem, 10;
  ; elem和8进行比较（utl表示unsigned less than）
  ; elem<8则跳到正常循环%for.body.preheader,否则跳到%vector.ph
  %min.iters.check = icmp ult i32 %elem, 8
  br i1 %min.iters.check, label %for.body.preheader, label %vector.ph

for.body.preheader:
  ;phi表示SSA里面的φ函数，详细参见LLVM DOC
  %i.06.ph = phi i32 [ 0, %for.body.lr.ph ], [ %n.vec, %middle.block ]
  br label %for.body

vector.ph:
  %n.vec = and i32 %elem, -8
  ;首先构造<%mul val val val>，然后shufflevector构造<%mul %mul %mul %mul>
  %broadcast.splatinsert9 = insertelement <4 x i32> undef, i32 %mul, i32 0
  %broadcast.splat10 = shufflevector <4 x i32> %broadcast.splatinsert9, <4 x i32> undef, <4 x i32> zeroinitializer
  ;ditto， %broadcast.splatinsert9 == <%mul %mul %mul %mul>
  %broadcast.splatinsert11 = insertelement <4 x i32> undef, i32 %mul, i32 0
  %broadcast.splat12 = shufflevector <4 x i32> %broadcast.splatinsert11, <4 x i32> undef, <4 x i32> zeroinitializer
  br label %vector.body

vector.body:
  %index = phi i32 [ 0, %vector.ph ], [ %index.next, %vector.body ]

  ;从arr指向的内存加载数据
  ;%0现在表示<arr[0] arr[1] arr[2] arr[3]>
  %0 = getelementoptr inbounds i32, i32* %arr, i32 %index
  %1 = bitcast i32* %0 to <4 x i32>*
  %wide.load = load <4 x i32>, <4 x i32>* %1, align 4, !tbaa !3

  ;%2表示<arr[4] arr[5] arr[6] arr[7]>
  %2 = getelementptr i32, i32* %0, i32 4
  %3 = bitcast i32* %2 to <4 x i32>*
  %wide.load8 = load <4 x i32>, <4 x i32>* %3, align 4, !tbaa !3

  ;<arr[0] arr[1] arr[2] arr[3]>与<%mul %mul %mul %mul>相加，得到vector:%4
  ;<arr[4] arr[5] arr[6] arr[7]>与<%mul %mul %mul %mul>相加，得到vector:%5
  %4 = add nsw <4 x i32> %wide.load, %broadcast.splat10
  %5 = add nsw <4 x i32> %wide.load8, %broadcast.splat12

  ;%4,%5写回内存
  %6 = bitcast i32* %0 to <4 x i32>*
  store <4 x i32> %4, <4 x i32>* %6, align 4, !tbaa !3
  %7 = bitcast i32* %2 to <4 x i32>*
  store <4 x i32> %5, <4 x i32>* %7, align 4, !tbaa !3

  %index.next = add i32 %index, 8
  %8 = icmp eq i32 %index.next, %n.vec
  br i1 %8, label %middle.block, label %vector.body, !llvm.loop !7

middle.block:
  %cmp.n = icmp eq i32 %n.vec, %elem
  br i1 %cmp.n, label %for.cond.cleanup, label %for.body.preheader

;函数返回
for.cond.cleanup:
  ret void

for.body:
  %i.06 = phi i32 [ %inc, %for.body ], [ %i.06.ph, %for.body.preheader ]‘
  ; arr[i] = arr[i]+ (elem*1024)，其中%mul=(elem*1024)
  %arrayidx = getelementptr inbounds i32, i32* %arr, i32 %i.06
  %9 = load i32, i32* %arrayidx, align 4, !tbaa !3
  %add = add nsw i32 %9, %mul
  store i32 %add, i32* %arrayidx, align 4, !tbaa !3
  ; i++
  %inc = add nuw nsw i32 %i.06, 1
  ; 循环条件i<elem判断
  %exitcond = icmp eq i32 %inc, %elem
  br i1 %exitcond, label %for.cond.cleanup, label %for.body, !llvm.loop !10
}