TNorthover
commented
5 years ago UBSan reports a shift of 1024 (which is out of range and so undefined behaviour). Does the issue still happen if you fix that?
It would also be useful to describe why the IR is wrong, or put some diagnostic printf in to demonstrate the issue.
Extended Description
In the experiment, when clang uses O2 and O3 optimization options, the pass of Loop unrolling will cause an result error.The procedure is as follows： typedef int A attribute((vector _size(64),aligned(64))); A Aa,Ab,Ai; static inline A attribute((always_inline__)) SET_A (int A,int B,int C,int D,int E,int F,int G,int H,int I,int J,int K,int L,int M,int N,int O,int P){ union{ int a[16] attribute((aligned(64))); A v; }u; u.a[0]=A; u.a[1]=B; u.a[2]=C; u.a[3]=D; u.a[4]=E; u.a[5]=F; u.a[6]=G; u.a[7]=H; u.a[8]=I; u.a[9]=J; u.a[10]=K; u.a[11]=L; u.a[12]=M; u.a[13]=N; u.a[14]=O; u.a[15]=P; return u.__v; } int main(){ int r[16]={2047,2047,2047,2047,2047,2047,2047,2047,2047,2047,2047,2047,2047,2047,2047,2047}; int i; for(i=1024;i>=1;i>>=1){ Ai=SET_A(i,i,i,i,i,i,i,i,i,i,i,i,i,i,i,i); Aa^=Aa>>i; Ab+=Aa+Ai; } } You can reproduce the error with the following command： clang -arch=x86 -static -O2 -emit-llvm -S a.c Loop unrolling optimization produces incorrect values in IR.