1; RUN: opt < %s -passes=loop-vectorize,dce,instcombine -force-vector-interleave=1 -force-vector-width=4 -S -enable-if-conversion | FileCheck %s 2 3target datalayout = "e-p:64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-f32:32:32-f64:64:64-v64:64:64-v128:128:128-a0:0:64-s0:64:64-f80:128:128-n8:16:32:64-S128" 4 5; This is kernel11 from "LivermoreLoops". We can't vectorize it because we 6; access both x[k] and x[k-1]. 7; 8; void kernel11(ptr x, ptr y, int n) { 9; for ( int k=1 ; k<n ; k++ ) 10; x[k] = x[k-1] + y[k]; 11; } 12 13; CHECK-LABEL: @kernel11( 14; CHECK-NOT: <4 x double> 15; CHECK: ret 16define i32 @kernel11(ptr %x, ptr %y, i32 %n) nounwind uwtable ssp { 17 %1 = alloca ptr, align 8 18 %2 = alloca ptr, align 8 19 %3 = alloca i32, align 4 20 %k = alloca i32, align 4 21 store ptr %x, ptr %1, align 8 22 store ptr %y, ptr %2, align 8 23 store i32 %n, ptr %3, align 4 24 store i32 1, ptr %k, align 4 25 br label %4 26 27; <label>:4 ; preds = %25, %0 28 %5 = load i32, ptr %k, align 4 29 %6 = load i32, ptr %3, align 4 30 %7 = icmp slt i32 %5, %6 31 br i1 %7, label %8, label %28 32 33; <label>:8 ; preds = %4 34 %9 = load i32, ptr %k, align 4 35 %10 = sub nsw i32 %9, 1 36 %11 = sext i32 %10 to i64 37 %12 = load ptr, ptr %1, align 8 38 %13 = getelementptr inbounds double, ptr %12, i64 %11 39 %14 = load double, ptr %13, align 8 40 %15 = load i32, ptr %k, align 4 41 %16 = sext i32 %15 to i64 42 %17 = load ptr, ptr %2, align 8 43 %18 = getelementptr inbounds double, ptr %17, i64 %16 44 %19 = load double, ptr %18, align 8 45 %20 = fadd double %14, %19 46 %21 = load i32, ptr %k, align 4 47 %22 = sext i32 %21 to i64 48 %23 = load ptr, ptr %1, align 8 49 %24 = getelementptr inbounds double, ptr %23, i64 %22 50 store double %20, ptr %24, align 8 51 br label %25 52 53; <label>:25 ; preds = %8 54 %26 = load i32, ptr %k, align 4 55 %27 = add nsw i32 %26, 1 56 store i32 %27, ptr %k, align 4 57 br label %4 58 59; <label>:28 ; preds = %4 60 ret i32 0 61} 62 63 64; A[i*7] is scalarized, and the different scalars can in theory wrap 65; around and overwrite other scalar elements. However we can still 66; vectorize because we can version the loop to avoid this case. 67; 68; void foo(int *a) { 69; for (int i=0; i<256; ++i) { 70; int x = a[i*7]; 71; if (x>3) 72; x = x*x+x*4; 73; a[i*7] = x+3; 74; } 75; } 76 77; CHECK-LABEL: @func2( 78; CHECK: <4 x i32> 79; CHECK: ret 80define i32 @func2(ptr nocapture %a) nounwind uwtable ssp { 81 br label %1 82 83; <label>:1 ; preds = %7, %0 84 %indvars.iv = phi i64 [ 0, %0 ], [ %indvars.iv.next, %7 ] 85 %2 = mul nsw i64 %indvars.iv, 7 86 %3 = getelementptr inbounds i32, ptr %a, i64 %2 87 %4 = load i32, ptr %3, align 4 88 %5 = icmp sgt i32 %4, 3 89 br i1 %5, label %6, label %7 90 91; <label>:6 ; preds = %1 92 %tmp = add i32 %4, 4 93 %tmp1 = mul i32 %tmp, %4 94 br label %7 95 96; <label>:7 ; preds = %6, %1 97 %x.0 = phi i32 [ %tmp1, %6 ], [ %4, %1 ] 98 %8 = add nsw i32 %x.0, 3 99 store i32 %8, ptr %3, align 4 100 %indvars.iv.next = add i64 %indvars.iv, 1 101 %lftr.wideiv = trunc i64 %indvars.iv.next to i32 102 %exitcond = icmp eq i32 %lftr.wideiv, 256 103 br i1 %exitcond, label %9, label %1 104 105; <label>:9 ; preds = %7 106 ret i32 0 107} 108