1; RUN: opt -mtriple aarch64-linux-gnu -mattr=+sve -loop-vectorize -dce -instcombine -S < %s | FileCheck %s
2
3; Ensure that we can vectorize loops such as:
4;   int *ptr = c;
5;   for (long long i = 0; i < n; i++) {
6;     int X1 = *ptr++;
7;     int X2 = *ptr++;
8;     a[i] = X1 + 1;
9;     b[i] = X2 + 1;
10;   }
11; with scalable vectors, including unrolling. The test below makes sure
12; that we can use gather instructions with the correct offsets, taking
13; vscale into account.
14
15define void @widen_ptr_phi_unrolled(i32* noalias nocapture %a, i32* noalias nocapture %b, i32* nocapture readonly %c, i64 %n) {
16; CHECK-LABEL: @widen_ptr_phi_unrolled(
17; CHECK:       vector.body:
18; CHECK-NEXT:    [[POINTER_PHI:%.*]] = phi i32* [ %c, %vector.ph ], [ %[[PTR_IND:.*]], %vector.body ]
19; CHECK:         [[TMP5:%.*]] = call i64 @llvm.vscale.i64()
20; CHECK-NEXT:    [[TMP6:%.*]] = shl i64 [[TMP5]], 2
21; CHECK-NEXT:    [[TMP7:%.*]] = shl i64 [[TMP5]], 4
22; CHECK-NEXT:    [[TMP8:%.*]] = call <vscale x 4 x i64> @llvm.experimental.stepvector.nxv4i64()
23; CHECK-NEXT:    [[VECTOR_GEP:%.*]] = shl <vscale x 4 x i64> [[TMP8]], shufflevector (<vscale x 4 x i64> insertelement (<vscale x 4 x i64> undef, i64 1, i32 0), <vscale x 4 x i64> undef, <vscale x 4 x i32> zeroinitializer)
24; CHECK-NEXT:    [[TMP9:%.*]] = getelementptr i32, i32* [[POINTER_PHI]], <vscale x 4 x i64> [[VECTOR_GEP]]
25; CHECK-NEXT:    [[DOTSPLATINSERT2:%.*]] = insertelement <vscale x 4 x i64> poison, i64 [[TMP6]], i32 0
26; CHECK-NEXT:    [[DOTSPLAT3:%.*]] = shufflevector <vscale x 4 x i64> [[DOTSPLATINSERT2]], <vscale x 4 x i64> poison, <vscale x 4 x i32> zeroinitializer
27; CHECK-NEXT:    [[TMP10:%.*]] = call <vscale x 4 x i64> @llvm.experimental.stepvector.nxv4i64()
28; CHECK-NEXT:    [[TMP11:%.*]] = add <vscale x 4 x i64> [[DOTSPLAT3]], [[TMP10]]
29; CHECK-NEXT:    [[VECTOR_GEP4:%.*]] = shl <vscale x 4 x i64> [[TMP11]], shufflevector (<vscale x 4 x i64> insertelement (<vscale x 4 x i64> undef, i64 1, i32 0), <vscale x 4 x i64> undef, <vscale x 4 x i32> zeroinitializer)
30; CHECK-NEXT:    [[TMP12:%.*]] = getelementptr i32, i32* [[POINTER_PHI]], <vscale x 4 x i64> [[VECTOR_GEP4]]
31; CHECK-NEXT:    [[TMP13:%.*]] = getelementptr inbounds i32, <vscale x 4 x i32*> [[TMP9]], i64 1
32; CHECK-NEXT:    [[TMP14:%.*]] = getelementptr inbounds i32, <vscale x 4 x i32*> [[TMP12]], i64 1
33; CHECK-NEXT:    {{%.*}} = call <vscale x 4 x i32> @llvm.masked.gather.nxv4i32.nxv4p0i32(<vscale x 4 x i32*> [[TMP9]],
34; CHECK-NEXT:    {{%.*}} = call <vscale x 4 x i32> @llvm.masked.gather.nxv4i32.nxv4p0i32(<vscale x 4 x i32*> [[TMP12]],
35; CHECK-NEXT:    {{%.*}} = call <vscale x 4 x i32> @llvm.masked.gather.nxv4i32.nxv4p0i32(<vscale x 4 x i32*> [[TMP13]],
36; CHECK-NEXT:    {{%.*}} = call <vscale x 4 x i32> @llvm.masked.gather.nxv4i32.nxv4p0i32(<vscale x 4 x i32*> [[TMP14]],
37; CHECK:         [[PTR_IND]] = getelementptr i32, i32* [[POINTER_PHI]], i64 [[TMP7]]
38entry:
39  br label %for.body
40
41for.body:                                         ; preds = %entry, %for.body
42  %ptr.014 = phi i32* [ %incdec.ptr1, %for.body ], [ %c, %entry ]
43  %i.013 = phi i64 [ %inc, %for.body ], [ 0, %entry ]
44  %incdec.ptr = getelementptr inbounds i32, i32* %ptr.014, i64 1
45  %0 = load i32, i32* %ptr.014, align 4
46  %incdec.ptr1 = getelementptr inbounds i32, i32* %ptr.014, i64 2
47  %1 = load i32, i32* %incdec.ptr, align 4
48  %add = add nsw i32 %0, 1
49  %arrayidx = getelementptr inbounds i32, i32* %a, i64 %i.013
50  store i32 %add, i32* %arrayidx, align 4
51  %add2 = add nsw i32 %1, 1
52  %arrayidx3 = getelementptr inbounds i32, i32* %b, i64 %i.013
53  store i32 %add2, i32* %arrayidx3, align 4
54  %inc = add nuw nsw i64 %i.013, 1
55  %exitcond.not = icmp eq i64 %inc, %n
56  br i1 %exitcond.not, label %for.exit, label %for.body, !llvm.loop !0
57
58for.exit:                                 ; preds = %for.body
59  ret void
60}
61
62
63; Ensure we can vectorise loops without interleaving, e.g.:
64;   int *D = dst;
65;   int *S = src;
66;   for (long long i = 0; i < n; i++) {
67;     *D = *S * 2;
68;     D++;
69;     S++;
70;   }
71; This takes us down a different codepath to the test above, where
72; here we treat the PHIs as being uniform.
73
74define void @widen_2ptrs_phi_unrolled(i32* noalias nocapture %dst, i32* noalias nocapture readonly %src, i64 %n) #0 {
75; CHECK-LABEL: @widen_2ptrs_phi_unrolled(
76; CHECK:       vector.body:
77; CHECK-NEXT:    %[[IDX:.*]] = phi i64 [ 0, %vector.ph ], [ %{{.*}}, %vector.body ]
78; CHECK-NEXT:    %[[LGEP1:.*]] = getelementptr i32, i32* %src, i64 %[[IDX]]
79; CHECK-NEXT:    %[[SGEP1:.*]] = getelementptr i32, i32* %dst, i64 %[[IDX]]
80; CHECK-NEXT:    %[[LPTR1:.*]] = bitcast i32* %[[LGEP1]] to <vscale x 4 x i32>*
81; CHECK-NEXT:    %{{.*}} = load <vscale x 4 x i32>, <vscale x 4 x i32>* %[[LPTR1]], align 4
82; CHECK-NEXT:    %[[VSCALE1:.*]] = call i32 @llvm.vscale.i32()
83; CHECK-NEXT:    %[[TMP1:.*]] = shl i32 %[[VSCALE1]], 2
84; CHECK-NEXT:    %[[TMP2:.*]] = sext i32 %[[TMP1]] to i64
85; CHECK-NEXT:    %[[LGEP2:.*]] = getelementptr i32, i32* %[[LGEP1]], i64 %[[TMP2]]
86; CHECK-NEXT:    %[[LPTR2:.*]] = bitcast i32* %[[LGEP2]] to <vscale x 4 x i32>*
87; CHECK-NEXT:    %{{.*}} = load <vscale x 4 x i32>, <vscale x 4 x i32>* %[[LPTR2]], align 4
88; CHECK:         %[[SPTR1:.*]] = bitcast i32* %[[SGEP1]] to <vscale x 4 x i32>*
89; CHECK-NEXT:    store <vscale x 4 x i32> %{{.*}}, <vscale x 4 x i32>* %[[SPTR1]], align 4
90; CHECK-NEXT:    %[[VSCALE2:.*]] = call i32 @llvm.vscale.i32()
91; CHECK-NEXT:    %[[TMP3:.*]] = shl i32 %[[VSCALE2]], 2
92; CHECK-NEXT:    %[[TMP4:.*]] = sext i32 %[[TMP3]] to i64
93; CHECK-NEXT:    %[[SGEP2:.*]] = getelementptr i32, i32* %[[SGEP1]], i64 %[[TMP4]]
94; CHECK-NEXT:    %[[SPTR2:.*]] = bitcast i32* %[[SGEP2]] to <vscale x 4 x i32>*
95; CHECK-NEXT:    store <vscale x 4 x i32> %{{.*}}, <vscale x 4 x i32>* %[[SPTR2]], align 4
96
97entry:
98  br label %for.body
99
100for.body:                                         ; preds = %entry, %for.body
101  %i.011 = phi i64 [ %inc, %for.body ], [ 0, %entry ]
102  %S.010 = phi i32* [ %incdec.ptr1, %for.body ], [ %src, %entry ]
103  %D.09 = phi i32* [ %incdec.ptr, %for.body ], [ %dst, %entry ]
104  %0 = load i32, i32* %S.010, align 4
105  %mul = shl nsw i32 %0, 1
106  store i32 %mul, i32* %D.09, align 4
107  %incdec.ptr = getelementptr inbounds i32, i32* %D.09, i64 1
108  %incdec.ptr1 = getelementptr inbounds i32, i32* %S.010, i64 1
109  %inc = add nuw nsw i64 %i.011, 1
110  %exitcond.not = icmp eq i64 %inc, %n
111  br i1 %exitcond.not, label %for.cond.cleanup, label %for.body, !llvm.loop !0
112
113for.cond.cleanup:                                 ; preds = %for.body
114  ret void
115}
116
117!0 = distinct !{!0, !1, !2, !3, !4, !5}
118!1 = !{!"llvm.loop.mustprogress"}
119!2 = !{!"llvm.loop.vectorize.width", i32 4}
120!3 = !{!"llvm.loop.vectorize.scalable.enable", i1 true}
121!4 = !{!"llvm.loop.vectorize.enable", i1 true}
122!5 = !{!"llvm.loop.interleave.count", i32 2}
123