1; RUN: opt < %s -nary-reassociate -early-cse -S | FileCheck %s 2 3target datalayout = "e-i64:64-v16:16-v32:32-n16:32:64" 4target triple = "nvptx64-unknown-unknown" 5 6declare void @foo(float*) 7 8; foo(&a[i]); 9; foo(&a[i + j]); 10; => 11; t = &a[i]; 12; foo(t); 13; foo(t + j); 14define void @reassociate_gep(float* %a, i64 %i, i64 %j) { 15; CHECK-LABEL: @reassociate_gep( 16 %1 = add i64 %i, %j 17 %2 = getelementptr float, float* %a, i64 %i 18; CHECK: [[t1:[^ ]+]] = getelementptr float, float* %a, i64 %i 19 call void @foo(float* %2) 20; CHECK: call void @foo(float* [[t1]]) 21 %3 = getelementptr float, float* %a, i64 %1 22; CHECK: [[t2:[^ ]+]] = getelementptr float, float* [[t1]], i64 %j 23 call void @foo(float* %3) 24; CHECK: call void @foo(float* [[t2]]) 25 ret void 26} 27 28; foo(&a[sext(j)]); 29; foo(&a[sext(i +nsw j)]); 30; foo(&a[sext((i +nsw j) +nsw i)]); 31; => 32; t1 = &a[sext(j)]; 33; foo(t1); 34; t2 = t1 + sext(i); 35; foo(t2); 36; t3 = t2 + sext(i); // sext(i) should be GVN'ed. 37; foo(t3); 38define void @reassociate_gep_nsw(float* %a, i32 %i, i32 %j) { 39; CHECK-LABEL: @reassociate_gep_nsw( 40 %idxprom.j = sext i32 %j to i64 41 %1 = getelementptr float, float* %a, i64 %idxprom.j 42; CHECK: [[t1:[^ ]+]] = getelementptr float, float* %a, i64 %idxprom.j 43 call void @foo(float* %1) 44; CHECK: call void @foo(float* [[t1]]) 45 46 %2 = add nsw i32 %i, %j 47 %idxprom.2 = sext i32 %2 to i64 48 %3 = getelementptr float, float* %a, i64 %idxprom.2 49; CHECK: [[sexti:[^ ]+]] = sext i32 %i to i64 50; CHECK: [[t2:[^ ]+]] = getelementptr float, float* [[t1]], i64 [[sexti]] 51 call void @foo(float* %3) 52; CHECK: call void @foo(float* [[t2]]) 53 54 %4 = add nsw i32 %2, %i 55 %idxprom.4 = sext i32 %4 to i64 56 %5 = getelementptr float, float* %a, i64 %idxprom.4 57; CHECK: [[t3:[^ ]+]] = getelementptr float, float* [[t2]], i64 [[sexti]] 58 call void @foo(float* %5) 59; CHECK: call void @foo(float* [[t3]]) 60 61 ret void 62} 63 64; assume(j >= 0); 65; foo(&a[zext(j)]); 66; assume(i + j >= 0); 67; foo(&a[zext(i + j)]); 68; => 69; t1 = &a[zext(j)]; 70; foo(t1); 71; t2 = t1 + sext(i); 72; foo(t2); 73define void @reassociate_gep_assume(float* %a, i32 %i, i32 %j) { 74; CHECK-LABEL: @reassociate_gep_assume( 75 ; assume(j >= 0) 76 %cmp = icmp sgt i32 %j, -1 77 call void @llvm.assume(i1 %cmp) 78 %1 = add i32 %i, %j 79 %cmp2 = icmp sgt i32 %1, -1 80 call void @llvm.assume(i1 %cmp2) 81 82 %idxprom.j = zext i32 %j to i64 83 %2 = getelementptr float, float* %a, i64 %idxprom.j 84; CHECK: [[t1:[^ ]+]] = getelementptr float, float* %a, i64 %idxprom.j 85 call void @foo(float* %2) 86; CHECK: call void @foo(float* [[t1]]) 87 88 %idxprom.1 = zext i32 %1 to i64 89 %3 = getelementptr float, float* %a, i64 %idxprom.1 90; CHECK: [[sexti:[^ ]+]] = sext i32 %i to i64 91; CHECK: [[t2:[^ ]+]] = getelementptr float, float* [[t1]], i64 [[sexti]] 92 call void @foo(float* %3) 93; CHECK: call void @foo(float* [[t2]]) 94 95 ret void 96} 97 98; Do not split the second GEP because sext(i + j) != sext(i) + sext(j). 99define void @reassociate_gep_no_nsw(float* %a, i32 %i, i32 %j) { 100; CHECK-LABEL: @reassociate_gep_no_nsw( 101 %1 = add i32 %i, %j 102 %2 = getelementptr float, float* %a, i32 %j 103; CHECK: getelementptr float, float* %a, i32 %j 104 call void @foo(float* %2) 105 %3 = getelementptr float, float* %a, i32 %1 106; CHECK: getelementptr float, float* %a, i32 %1 107 call void @foo(float* %3) 108 ret void 109} 110 111define void @reassociate_gep_128(float* %a, i128 %i, i128 %j) { 112; CHECK-LABEL: @reassociate_gep_128( 113 %1 = add i128 %i, %j 114 %2 = getelementptr float, float* %a, i128 %i 115; CHECK: [[t1:[^ ]+]] = getelementptr float, float* %a, i128 %i 116 call void @foo(float* %2) 117; CHECK: call void @foo(float* [[t1]]) 118 %3 = getelementptr float, float* %a, i128 %1 119; CHECK: [[truncj:[^ ]+]] = trunc i128 %j to i64 120; CHECK: [[t2:[^ ]+]] = getelementptr float, float* [[t1]], i64 [[truncj]] 121 call void @foo(float* %3) 122; CHECK: call void @foo(float* [[t2]]) 123 ret void 124} 125 126%struct.complex = type { float, float } 127 128declare void @bar(%struct.complex*) 129 130define void @different_types(%struct.complex* %input, i64 %i) { 131; CHECK-LABEL: @different_types( 132 %t1 = getelementptr %struct.complex, %struct.complex* %input, i64 %i 133 call void @bar(%struct.complex* %t1) 134 %j = add i64 %i, 5 135 %t2 = getelementptr %struct.complex, %struct.complex* %input, i64 %j, i32 0 136; CHECK: [[cast:[^ ]+]] = bitcast %struct.complex* %t1 to float* 137; CHECK-NEXT: %t2 = getelementptr float, float* [[cast]], i64 10 138; CHECK-NEXT: call void @foo(float* %t2) 139 call void @foo(float* %t2) 140 ret void 141} 142 143declare void @llvm.assume(i1) 144