1; NOTE: Assertions have been autogenerated by utils/update_test_checks.py 2; RUN: opt -S < %s -passes='loop(loop-rotate)' -loop-rotate-multi=true | FileCheck %s 3 4; Test loop rotation with multiple exits, some of them - deoptimizing. 5; We should end up with a latch which exit is non-deoptimizing, so we should rotate 6; more than once. 7 8declare i32 @llvm.experimental.deoptimize.i32(...) 9 10define i32 @test_cond_with_one_deopt_exit(ptr nonnull %a, i64 %x) { 11; Rotation done twice. 12; Latch should be at the 2nd condition (for.cond2), exiting to %return. 13; 14; CHECK-LABEL: @test_cond_with_one_deopt_exit( 15; CHECK-NEXT: entry: 16; CHECK-NEXT: [[VAL_A_IDX3:%.*]] = load i32, ptr %a, align 4 17; CHECK-NEXT: [[ZERO_CHECK4:%.*]] = icmp eq i32 [[VAL_A_IDX3]], 0 18; CHECK-NEXT: br i1 [[ZERO_CHECK4]], label %deopt.exit, label %for.cond2.lr.ph 19; CHECK: for.cond2.lr.ph: 20; CHECK-NEXT: [[FOR_CHECK8:%.*]] = icmp ult i64 0, %x 21; CHECK-NEXT: br i1 [[FOR_CHECK8]], label %for.body.lr.ph, label %return 22; CHECK: for.body.lr.ph: 23; CHECK-NEXT: br label %for.body 24; CHECK: for.cond2: 25; CHECK: [[FOR_CHECK:%.*]] = icmp ult i64 {{%.*}}, %x 26; CHECK-NEXT: br i1 [[FOR_CHECK]], label %for.body, label %for.cond2.return_crit_edge 27; CHECK: for.body: 28; CHECK: br label %for.tail 29; CHECK: for.tail: 30; CHECK: [[VAL_A_IDX:%.*]] = load i32, ptr 31; CHECK-NEXT: [[ZERO_CHECK:%.*]] = icmp eq i32 [[VAL_A_IDX]], 0 32; CHECK-NEXT: br i1 [[ZERO_CHECK]], label %for.cond1.deopt.exit_crit_edge, label %for.cond2 33; CHECK: for.cond2.return_crit_edge: 34; CHECK-NEXT: {{%.*}} = phi i32 35; CHECK-NEXT: br label %return 36; CHECK: return: 37; CHECK-NEXT: [[SUM_LCSSA2:%.*]] = phi i32 38; CHECK-NEXT: ret i32 [[SUM_LCSSA2]] 39; CHECK: for.cond1.deopt.exit_crit_edge: 40; CHECK-NEXT: {{%.*}} = phi i32 41; CHECK-NEXT: br label %deopt.exit 42; CHECK: deopt.exit: 43; CHECK: [[DEOPT_VAL:%.*]] = call i32 (...) @llvm.experimental.deoptimize.i32() [ "deopt"(i32 {{%.*}}) ] 44; CHECK-NEXT: ret i32 [[DEOPT_VAL]] 45; 46entry: 47 br label %for.cond1 48 49for.cond1: 50 %idx = phi i64 [ 0, %entry ], [ %idx.next, %for.tail ] 51 %sum = phi i32 [ 0, %entry ], [ %sum.next, %for.tail ] 52 %a.idx = getelementptr inbounds i32, ptr %a, i64 %idx 53 %val.a.idx = load i32, ptr %a.idx, align 4 54 %zero.check = icmp eq i32 %val.a.idx, 0 55 br i1 %zero.check, label %deopt.exit, label %for.cond2 56 57for.cond2: 58 %for.check = icmp ult i64 %idx, %x 59 br i1 %for.check, label %for.body, label %return 60 61for.body: 62 br label %for.tail 63 64for.tail: 65 %sum.next = add i32 %sum, %val.a.idx 66 %idx.next = add nuw nsw i64 %idx, 1 67 br label %for.cond1 68 69return: 70 ret i32 %sum 71 72deopt.exit: 73 %deopt.val = call i32(...) @llvm.experimental.deoptimize.i32() [ "deopt"(i32 %val.a.idx) ] 74 ret i32 %deopt.val 75} 76 77define i32 @test_cond_with_two_deopt_exits(ptr nonnull %a, i64 %x) { 78; Rotation done three times. 79; Latch should be at the 3rd condition (for.cond3), exiting to %return. 80; 81; CHECK-LABEL: @test_cond_with_two_deopt_exits( 82; CHECK-NEXT: entry: 83; CHECK-NEXT: [[A_IDX_DEREF4:%.*]] = load ptr, ptr %a 84; CHECK-NEXT: [[NULL_CHECK5:%.*]] = icmp eq ptr [[A_IDX_DEREF4]], null 85; CHECK-NEXT: br i1 [[NULL_CHECK5]], label %deopt.exit1, label %for.cond2.lr.ph 86; CHECK: for.cond2.lr.ph: 87; CHECK-NEXT: [[VAL_A_IDX9:%.*]] = load i32, ptr [[A_IDX_DEREF4]], align 4 88; CHECK-NEXT: [[ZERO_CHECK10:%.*]] = icmp eq i32 [[VAL_A_IDX9]], 0 89; CHECK-NEXT: br i1 [[ZERO_CHECK10]], label %deopt.exit2, label %for.cond3.lr.ph 90; CHECK: for.cond3.lr.ph: 91; CHECK-NEXT: [[FOR_CHECK14:%.*]] = icmp ult i64 0, %x 92; CHECK-NEXT: br i1 [[FOR_CHECK14]], label %for.body.lr.ph, label %return 93; CHECK: for.body.lr.ph: 94; CHECK-NEXT: br label %for.body 95; CHECK: for.cond2: 96; CHECK: [[VAL_A_IDX:%.*]] = load i32, ptr 97; CHECK-NEXT: [[ZERO_CHECK:%.*]] = icmp eq i32 [[VAL_A_IDX]], 0 98; CHECK-NEXT: br i1 [[ZERO_CHECK]], label %for.cond2.deopt.exit2_crit_edge, label %for.cond3 99; CHECK: for.cond3: 100; CHECK: [[FOR_CHECK:%.*]] = icmp ult i64 {{%.*}}, %x 101; CHECK-NEXT: br i1 [[FOR_CHECK]], label %for.body, label %for.cond3.return_crit_edge 102; CHECK: for.body: 103; CHECK: br label %for.tail 104; CHECK: for.tail: 105; CHECK: [[IDX_NEXT:%.*]] = add nuw nsw i64 {{%.*}}, 1 106; CHECK: [[NULL_CHECK:%.*]] = icmp eq ptr {{%.*}}, null 107; CHECK-NEXT: br i1 [[NULL_CHECK]], label %for.cond1.deopt.exit1_crit_edge, label %for.cond2 108; CHECK: for.cond3.return_crit_edge: 109; CHECK-NEXT: [[SPLIT18:%.*]] = phi i32 110; CHECK-NEXT: br label %return 111; CHECK: return: 112; CHECK-NEXT: [[SUM_LCSSA2:%.*]] = phi i32 113; CHECK-NEXT: ret i32 [[SUM_LCSSA2]] 114; CHECK: for.cond1.deopt.exit1_crit_edge: 115; CHECK-NEXT: br label %deopt.exit1 116; CHECK: deopt.exit1: 117; CHECK-NEXT: [[DEOPT_VAL1:%.*]] = call i32 (...) @llvm.experimental.deoptimize.i32() [ "deopt"(i32 0) ] 118; CHECK-NEXT: ret i32 [[DEOPT_VAL1]] 119; CHECK: for.cond2.deopt.exit2_crit_edge: 120; CHECK-NEXT: [[SPLIT:%.*]] = phi i32 121; CHECK-NEXT: br label %deopt.exit2 122; CHECK: deopt.exit2: 123; CHECK-NEXT: [[VAL_A_IDX_LCSSA:%.*]] = phi i32 124; CHECK-NEXT: [[DEOPT_VAL2:%.*]] = call i32 (...) @llvm.experimental.deoptimize.i32() [ "deopt"(i32 [[VAL_A_IDX_LCSSA]]) ] 125; CHECK-NEXT: ret i32 [[DEOPT_VAL2]] 126; 127entry: 128 br label %for.cond1 129 130for.cond1: 131 %idx = phi i64 [ 0, %entry ], [ %idx.next, %for.tail ] 132 %sum = phi i32 [ 0, %entry ], [ %sum.next, %for.tail ] 133 %a.idx = getelementptr inbounds ptr, ptr %a, i64 %idx 134 %a.idx.deref = load ptr, ptr %a.idx 135 %null.check = icmp eq ptr %a.idx.deref, null 136 br i1 %null.check, label %deopt.exit1, label %for.cond2 137 138for.cond2: 139 %val.a.idx = load i32, ptr %a.idx.deref, align 4 140 %zero.check = icmp eq i32 %val.a.idx, 0 141 br i1 %zero.check, label %deopt.exit2, label %for.cond3 142 143for.cond3: 144 %for.check = icmp ult i64 %idx, %x 145 br i1 %for.check, label %for.body, label %return 146 147for.body: 148 br label %for.tail 149 150for.tail: 151 %sum.next = add i32 %sum, %val.a.idx 152 %idx.next = add nuw nsw i64 %idx, 1 153 br label %for.cond1 154 155return: 156 ret i32 %sum 157 158deopt.exit1: 159 %deopt.val1 = call i32(...) @llvm.experimental.deoptimize.i32() [ "deopt"(i32 0) ] 160 ret i32 %deopt.val1 161deopt.exit2: 162 %deopt.val2 = call i32(...) @llvm.experimental.deoptimize.i32() [ "deopt"(i32 %val.a.idx) ] 163 ret i32 %deopt.val2 164} 165