1; RUN: opt < %s -slsr -gvn -S | FileCheck %s 2 3target datalayout = "e-i64:64-v16:16-v32:32-n16:32:64" 4 5; foo(input[0]); 6; foo(input[s]); 7; foo(input[s * 2]); 8; => 9; p0 = &input[0]; 10; foo(*p); 11; p1 = p0 + s; 12; foo(*p1); 13; p2 = p1 + s; 14; foo(*p2); 15define void @slsr_gep(i32* %input, i64 %s) { 16; CHECK-LABEL: @slsr_gep( 17 ; v0 = input[0]; 18 %p0 = getelementptr inbounds i32, i32* %input, i64 0 19 %v0 = load i32, i32* %p0 20 call void @foo(i32 %v0) 21 22 ; v1 = input[s]; 23 %p1 = getelementptr inbounds i32, i32* %input, i64 %s 24; CHECK: %p1 = getelementptr inbounds i32, i32* %input, i64 %s 25 %v1 = load i32, i32* %p1 26 call void @foo(i32 %v1) 27 28 ; v2 = input[s * 2]; 29 %s2 = shl nsw i64 %s, 1 30 %p2 = getelementptr inbounds i32, i32* %input, i64 %s2 31; CHECK: %p2 = getelementptr inbounds i32, i32* %p1, i64 %s 32 %v2 = load i32, i32* %p2 33 call void @foo(i32 %v2) 34 35 ret void 36} 37 38; foo(input[0]); 39; foo(input[(long)s]); 40; foo(input[(long)(s * 2)]); 41; => 42; p0 = &input[0]; 43; foo(*p); 44; p1 = p0 + (long)s; 45; foo(*p1); 46; p2 = p1 + (long)s; 47; foo(*p2); 48define void @slsr_gep_sext(i32* %input, i32 %s) { 49; CHECK-LABEL: @slsr_gep_sext( 50 ; v0 = input[0]; 51 %p0 = getelementptr inbounds i32, i32* %input, i64 0 52 %v0 = load i32, i32* %p0 53 call void @foo(i32 %v0) 54 55 ; v1 = input[s]; 56 %t = sext i32 %s to i64 57 %p1 = getelementptr inbounds i32, i32* %input, i64 %t 58; CHECK: %p1 = getelementptr inbounds i32, i32* %input, i64 %t 59 %v1 = load i32, i32* %p1 60 call void @foo(i32 %v1) 61 62 ; v2 = input[s * 2]; 63 %s2 = shl nsw i32 %s, 1 64 %t2 = sext i32 %s2 to i64 65 %p2 = getelementptr inbounds i32, i32* %input, i64 %t2 66; CHECK: %p2 = getelementptr inbounds i32, i32* %p1, i64 %t 67 %v2 = load i32, i32* %p2 68 call void @foo(i32 %v2) 69 70 ret void 71} 72 73; int input[10][5]; 74; foo(input[s][t]); 75; foo(input[s * 2][t]); 76; foo(input[s * 3][t]); 77; => 78; p0 = &input[s][t]; 79; foo(*p0); 80; p1 = p0 + 5s; 81; foo(*p1); 82; p2 = p1 + 5s; 83; foo(*p2); 84define void @slsr_gep_2d([10 x [5 x i32]]* %input, i64 %s, i64 %t) { 85; CHECK-LABEL: @slsr_gep_2d( 86 ; v0 = input[s][t]; 87 %p0 = getelementptr inbounds [10 x [5 x i32]], [10 x [5 x i32]]* %input, i64 0, i64 %s, i64 %t 88 %v0 = load i32, i32* %p0 89 call void @foo(i32 %v0) 90 91 ; v1 = input[s * 2][t]; 92 %s2 = shl nsw i64 %s, 1 93; CHECK: [[BUMP:%[a-zA-Z0-9]+]] = mul i64 %s, 5 94 %p1 = getelementptr inbounds [10 x [5 x i32]], [10 x [5 x i32]]* %input, i64 0, i64 %s2, i64 %t 95; CHECK: %p1 = getelementptr inbounds i32, i32* %p0, i64 [[BUMP]] 96 %v1 = load i32, i32* %p1 97 call void @foo(i32 %v1) 98 99 ; v3 = input[s * 3][t]; 100 %s3 = mul nsw i64 %s, 3 101 %p2 = getelementptr inbounds [10 x [5 x i32]], [10 x [5 x i32]]* %input, i64 0, i64 %s3, i64 %t 102; CHECK: %p2 = getelementptr inbounds i32, i32* %p1, i64 [[BUMP]] 103 %v2 = load i32, i32* %p2 104 call void @foo(i32 %v2) 105 106 ret void 107} 108 109%struct.S = type <{ i64, i32 }> 110 111; In this case, the bump 112; = (char *)&input[s * 2][t].f1 - (char *)&input[s][t].f1 113; = 60 * s 114; which may not be divisible by typeof(input[s][t].f1) = 8. Therefore, we 115; rewrite the candidates using byte offset instead of index offset as in 116; @slsr_gep_2d. 117define void @slsr_gep_uglygep([10 x [5 x %struct.S]]* %input, i64 %s, i64 %t) { 118; CHECK-LABEL: @slsr_gep_uglygep( 119 ; v0 = input[s][t].f1; 120 %p0 = getelementptr inbounds [10 x [5 x %struct.S]], [10 x [5 x %struct.S]]* %input, i64 0, i64 %s, i64 %t, i32 0 121 %v0 = load i64, i64* %p0 122 call void @bar(i64 %v0) 123 124 ; v1 = input[s * 2][t].f1; 125 %s2 = shl nsw i64 %s, 1 126; CHECK: [[BUMP:%[a-zA-Z0-9]+]] = mul i64 %s, 60 127 %p1 = getelementptr inbounds [10 x [5 x %struct.S]], [10 x [5 x %struct.S]]* %input, i64 0, i64 %s2, i64 %t, i32 0 128; CHECK: getelementptr inbounds i8, i8* %{{[0-9]+}}, i64 [[BUMP]] 129 %v1 = load i64, i64* %p1 130 call void @bar(i64 %v1) 131 132 ; v2 = input[s * 3][t].f1; 133 %s3 = mul nsw i64 %s, 3 134 %p2 = getelementptr inbounds [10 x [5 x %struct.S]], [10 x [5 x %struct.S]]* %input, i64 0, i64 %s3, i64 %t, i32 0 135; CHECK: getelementptr inbounds i8, i8* %{{[0-9]+}}, i64 [[BUMP]] 136 %v2 = load i64, i64* %p2 137 call void @bar(i64 %v2) 138 139 ret void 140} 141 142define void @slsr_out_of_bounds_gep(i32* %input, i32 %s) { 143; CHECK-LABEL: @slsr_out_of_bounds_gep( 144 ; v0 = input[0]; 145 %p0 = getelementptr i32, i32* %input, i64 0 146 %v0 = load i32, i32* %p0 147 call void @foo(i32 %v0) 148 149 ; v1 = input[(long)s]; 150 %t = sext i32 %s to i64 151 %p1 = getelementptr i32, i32* %input, i64 %t 152; CHECK: %p1 = getelementptr i32, i32* %input, i64 %t 153 %v1 = load i32, i32* %p1 154 call void @foo(i32 %v1) 155 156 ; v2 = input[(long)(s * 2)]; 157 %s2 = shl nsw i32 %s, 1 158 %t2 = sext i32 %s2 to i64 159 %p2 = getelementptr i32, i32* %input, i64 %t2 160; CHECK: %p2 = getelementptr i32, i32* %p1, i64 %t 161 %v2 = load i32, i32* %p2 162 call void @foo(i32 %v2) 163 164 ret void 165} 166 167declare void @foo(i32) 168declare void @bar(i64) 169