1 ; RUN: opt -loop-vectorize -force-vector-width=4 -S < %s | FileCheck %s 2 3 ; This is the test case from PR26314. 4 ; When we were retrying dependence checking with memchecks only, 5 ; the loop-invariant access in the inner loop was incorrectly determined to be wrapping 6 ; because it was not strided in the inner loop. 7 ; Improved wrapping detection allows vectorization in the following case. 8 9 ; #define Z 32 10 ; typedef struct s { 11 ; int v1[Z]; 12 ; int v2[Z]; 13 ; int v3[Z][Z]; 14 ; } s; 15 ; 16 ; void slow_function (s* const obj) { 17 ; for (int j=0; j<Z; j++) { 18 ; for (int k=0; k<Z; k++) { 19 ; int x = obj->v1[k] + obj->v2[j]; 20 ; obj->v3[j][k] += x; 21 ; } 22 ; } 23 ; } 24 25 ; CHECK-LABEL: Test 26 ; CHECK: <4 x i64> 27 ; CHECK: <4 x i32>, <4 x i32> 28 ; CHECK: llvm.loop.vectorize.width 29 30 target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128" 31 target triple = "x86_64-unknown-linux-gnu" 32 33 %struct.s = type { [32 x i32], [32 x i32], [32 x [32 x i32]] } 34 35 define void @Test(%struct.s* nocapture %obj) #0 { 36 br label %.outer.preheader 37 38 39 .outer.preheader: 40 %i = phi i64 [ 0, %0 ], [ %i.next, %.outer ] 41 %1 = getelementptr inbounds %struct.s, %struct.s* %obj, i64 0, i32 1, i64 %i 42 br label %.inner 43 44 .exit: 45 ret void 46 47 .outer: 48 %i.next = add nuw nsw i64 %i, 1 49 %exitcond.outer = icmp eq i64 %i.next, 32 50 br i1 %exitcond.outer, label %.exit, label %.outer.preheader 51 52 .inner: 53 %j = phi i64 [ 0, %.outer.preheader ], [ %j.next, %.inner ] 54 %2 = getelementptr inbounds %struct.s, %struct.s* %obj, i64 0, i32 0, i64 %j 55 %3 = load i32, i32* %2 56 %4 = load i32, i32* %1 57 %5 = add nsw i32 %4, %3 58 %6 = getelementptr inbounds %struct.s, %struct.s* %obj, i64 0, i32 2, i64 %i, i64 %j 59 %7 = load i32, i32* %6 60 %8 = add nsw i32 %5, %7 61 store i32 %8, i32* %6 62 %j.next = add nuw nsw i64 %j, 1 63 %exitcond.inner = icmp eq i64 %j.next, 32 64 br i1 %exitcond.inner, label %.outer, label %.inner 65 } 66
