1; NOTE: Assertions have been autogenerated by utils/update_test_checks.py 2; RUN: opt -S < %s -loop-rotate -loop-rotate-multi=true | FileCheck %s 3; RUN: opt -S < %s -passes='loop(loop-rotate)' -loop-rotate-multi=true | FileCheck %s 4 5; Test loop rotation with multiple exits, some of them - deoptimizing. 6; We should end up with a latch which exit is non-deoptimizing, so we should rotate 7; more than once. 8 9declare i32 @llvm.experimental.deoptimize.i32(...) 10 11define i32 @test_cond_with_one_deopt_exit(i32 * nonnull %a, i64 %x) { 12; Rotation done twice. 13; Latch should be at the 2nd condition (for.cond2), exiting to %return. 14; 15; CHECK-LABEL: @test_cond_with_one_deopt_exit( 16; CHECK-NEXT: entry: 17; CHECK-NEXT: [[VAL_A_IDX3:%.*]] = load i32, i32* %a, align 4 18; CHECK-NEXT: [[ZERO_CHECK4:%.*]] = icmp eq i32 [[VAL_A_IDX3]], 0 19; CHECK-NEXT: br i1 [[ZERO_CHECK4]], label %deopt.exit, label %for.cond2.lr.ph 20; CHECK: for.cond2.lr.ph: 21; CHECK-NEXT: [[FOR_CHECK8:%.*]] = icmp ult i64 0, %x 22; CHECK-NEXT: br i1 [[FOR_CHECK8]], label %for.body.lr.ph, label %return 23; CHECK: for.body.lr.ph: 24; CHECK-NEXT: br label %for.body 25; CHECK: for.cond2: 26; CHECK: [[FOR_CHECK:%.*]] = icmp ult i64 {{%.*}}, %x 27; CHECK-NEXT: br i1 [[FOR_CHECK]], label %for.body, label %for.cond2.return_crit_edge 28; CHECK: for.body: 29; CHECK: br label %for.tail 30; CHECK: for.tail: 31; CHECK: [[VAL_A_IDX:%.*]] = load i32, i32* 32; CHECK-NEXT: [[ZERO_CHECK:%.*]] = icmp eq i32 [[VAL_A_IDX]], 0 33; CHECK-NEXT: br i1 [[ZERO_CHECK]], label %for.cond1.deopt.exit_crit_edge, label %for.cond2 34; CHECK: for.cond2.return_crit_edge: 35; CHECK-NEXT: {{%.*}} = phi i32 36; CHECK-NEXT: br label %return 37; CHECK: return: 38; CHECK-NEXT: [[SUM_LCSSA2:%.*]] = phi i32 39; CHECK-NEXT: ret i32 [[SUM_LCSSA2]] 40; CHECK: for.cond1.deopt.exit_crit_edge: 41; CHECK-NEXT: {{%.*}} = phi i32 42; CHECK-NEXT: br label %deopt.exit 43; CHECK: deopt.exit: 44; CHECK: [[DEOPT_VAL:%.*]] = call i32 (...) @llvm.experimental.deoptimize.i32() [ "deopt"(i32 {{%.*}}) ] 45; CHECK-NEXT: ret i32 [[DEOPT_VAL]] 46; 47entry: 48 br label %for.cond1 49 50for.cond1: 51 %idx = phi i64 [ 0, %entry ], [ %idx.next, %for.tail ] 52 %sum = phi i32 [ 0, %entry ], [ %sum.next, %for.tail ] 53 %a.idx = getelementptr inbounds i32, i32 *%a, i64 %idx 54 %val.a.idx = load i32, i32* %a.idx, align 4 55 %zero.check = icmp eq i32 %val.a.idx, 0 56 br i1 %zero.check, label %deopt.exit, label %for.cond2 57 58for.cond2: 59 %for.check = icmp ult i64 %idx, %x 60 br i1 %for.check, label %for.body, label %return 61 62for.body: 63 br label %for.tail 64 65for.tail: 66 %sum.next = add i32 %sum, %val.a.idx 67 %idx.next = add nuw nsw i64 %idx, 1 68 br label %for.cond1 69 70return: 71 ret i32 %sum 72 73deopt.exit: 74 %deopt.val = call i32(...) @llvm.experimental.deoptimize.i32() [ "deopt"(i32 %val.a.idx) ] 75 ret i32 %deopt.val 76} 77 78define i32 @test_cond_with_two_deopt_exits(i32 ** nonnull %a, i64 %x) { 79; Rotation done three times. 80; Latch should be at the 3rd condition (for.cond3), exiting to %return. 81; 82; CHECK-LABEL: @test_cond_with_two_deopt_exits( 83; CHECK-NEXT: entry: 84; CHECK-NEXT: [[A_IDX_DEREF4:%.*]] = load i32*, i32** %a 85; CHECK-NEXT: [[NULL_CHECK5:%.*]] = icmp eq i32* [[A_IDX_DEREF4]], null 86; CHECK-NEXT: br i1 [[NULL_CHECK5]], label %deopt.exit1, label %for.cond2.lr.ph 87; CHECK: for.cond2.lr.ph: 88; CHECK-NEXT: [[VAL_A_IDX9:%.*]] = load i32, i32* [[A_IDX_DEREF4]], align 4 89; CHECK-NEXT: [[ZERO_CHECK10:%.*]] = icmp eq i32 [[VAL_A_IDX9]], 0 90; CHECK-NEXT: br i1 [[ZERO_CHECK10]], label %deopt.exit2, label %for.cond3.lr.ph 91; CHECK: for.cond3.lr.ph: 92; CHECK-NEXT: [[FOR_CHECK14:%.*]] = icmp ult i64 0, %x 93; CHECK-NEXT: br i1 [[FOR_CHECK14]], label %for.body.lr.ph, label %return 94; CHECK: for.body.lr.ph: 95; CHECK-NEXT: br label %for.body 96; CHECK: for.cond2: 97; CHECK: [[VAL_A_IDX:%.*]] = load i32, i32* 98; CHECK-NEXT: [[ZERO_CHECK:%.*]] = icmp eq i32 [[VAL_A_IDX]], 0 99; CHECK-NEXT: br i1 [[ZERO_CHECK]], label %for.cond2.deopt.exit2_crit_edge, label %for.cond3 100; CHECK: for.cond3: 101; CHECK: [[FOR_CHECK:%.*]] = icmp ult i64 {{%.*}}, %x 102; CHECK-NEXT: br i1 [[FOR_CHECK]], label %for.body, label %for.cond3.return_crit_edge 103; CHECK: for.body: 104; CHECK: br label %for.tail 105; CHECK: for.tail: 106; CHECK: [[IDX_NEXT:%.*]] = add nuw nsw i64 {{%.*}}, 1 107; CHECK: [[NULL_CHECK:%.*]] = icmp eq i32* {{%.*}}, null 108; CHECK-NEXT: br i1 [[NULL_CHECK]], label %for.cond1.deopt.exit1_crit_edge, label %for.cond2 109; CHECK: for.cond3.return_crit_edge: 110; CHECK-NEXT: [[SPLIT18:%.*]] = phi i32 111; CHECK-NEXT: br label %return 112; CHECK: return: 113; CHECK-NEXT: [[SUM_LCSSA2:%.*]] = phi i32 114; CHECK-NEXT: ret i32 [[SUM_LCSSA2]] 115; CHECK: for.cond1.deopt.exit1_crit_edge: 116; CHECK-NEXT: br label %deopt.exit1 117; CHECK: deopt.exit1: 118; CHECK-NEXT: [[DEOPT_VAL1:%.*]] = call i32 (...) @llvm.experimental.deoptimize.i32() [ "deopt"(i32 0) ] 119; CHECK-NEXT: ret i32 [[DEOPT_VAL1]] 120; CHECK: for.cond2.deopt.exit2_crit_edge: 121; CHECK-NEXT: [[SPLIT:%.*]] = phi i32 122; CHECK-NEXT: br label %deopt.exit2 123; CHECK: deopt.exit2: 124; CHECK-NEXT: [[VAL_A_IDX_LCSSA:%.*]] = phi i32 125; CHECK-NEXT: [[DEOPT_VAL2:%.*]] = call i32 (...) @llvm.experimental.deoptimize.i32() [ "deopt"(i32 [[VAL_A_IDX_LCSSA]]) ] 126; CHECK-NEXT: ret i32 [[DEOPT_VAL2]] 127; 128entry: 129 br label %for.cond1 130 131for.cond1: 132 %idx = phi i64 [ 0, %entry ], [ %idx.next, %for.tail ] 133 %sum = phi i32 [ 0, %entry ], [ %sum.next, %for.tail ] 134 %a.idx = getelementptr inbounds i32*, i32 **%a, i64 %idx 135 %a.idx.deref = load i32*, i32** %a.idx 136 %null.check = icmp eq i32* %a.idx.deref, null 137 br i1 %null.check, label %deopt.exit1, label %for.cond2 138 139for.cond2: 140 %val.a.idx = load i32, i32* %a.idx.deref, align 4 141 %zero.check = icmp eq i32 %val.a.idx, 0 142 br i1 %zero.check, label %deopt.exit2, label %for.cond3 143 144for.cond3: 145 %for.check = icmp ult i64 %idx, %x 146 br i1 %for.check, label %for.body, label %return 147 148for.body: 149 br label %for.tail 150 151for.tail: 152 %sum.next = add i32 %sum, %val.a.idx 153 %idx.next = add nuw nsw i64 %idx, 1 154 br label %for.cond1 155 156return: 157 ret i32 %sum 158 159deopt.exit1: 160 %deopt.val1 = call i32(...) @llvm.experimental.deoptimize.i32() [ "deopt"(i32 0) ] 161 ret i32 %deopt.val1 162deopt.exit2: 163 %deopt.val2 = call i32(...) @llvm.experimental.deoptimize.i32() [ "deopt"(i32 %val.a.idx) ] 164 ret i32 %deopt.val2 165} 166