1; RUN: opt -S -loop-predication -loop-predication-enable-count-down-loop=true < %s 2>&1 | FileCheck %s 2; RUN: opt -S -passes='require<scalar-evolution>,loop(loop-predication)' -loop-predication-enable-count-down-loop=true < %s 2>&1 | FileCheck %s 3 4declare void @llvm.experimental.guard(i1, ...) 5 6define i32 @signed_reverse_loop_n_to_lower_limit(i32* %array, i32 %length, i32 %n, i32 %lowerlimit) { 7; CHECK-LABEL: @signed_reverse_loop_n_to_lower_limit( 8entry: 9 %tmp5 = icmp eq i32 %n, 0 10 br i1 %tmp5, label %exit, label %loop.preheader 11 12; CHECK: loop.preheader: 13; CHECK-NEXT: [[range_start:%.*]] = add i32 %n, -1 14; CHECK-NEXT: [[first_iteration_check:%.*]] = icmp ult i32 [[range_start]], %length 15; CHECK-NEXT: [[no_wrap_check:%.*]] = icmp sge i32 %lowerlimit, 1 16; CHECK-NEXT: [[wide_cond:%.*]] = and i1 [[first_iteration_check]], [[no_wrap_check]] 17loop.preheader: 18 br label %loop 19 20; CHECK: loop: 21; CHECK: call void (i1, ...) @llvm.experimental.guard(i1 [[wide_cond]], i32 9) [ "deopt"() ] 22loop: 23 %loop.acc = phi i32 [ %loop.acc.next, %loop ], [ 0, %loop.preheader ] 24 %i = phi i32 [ %i.next, %loop ], [ %n, %loop.preheader ] 25 %i.next = add nsw i32 %i, -1 26 %within.bounds = icmp ult i32 %i.next, %length 27 call void (i1, ...) @llvm.experimental.guard(i1 %within.bounds, i32 9) [ "deopt"() ] 28 %i.i64 = zext i32 %i.next to i64 29 %array.i.ptr = getelementptr inbounds i32, i32* %array, i64 %i.i64 30 %array.i = load i32, i32* %array.i.ptr, align 4 31 %loop.acc.next = add i32 %loop.acc, %array.i 32 %continue = icmp sgt i32 %i, %lowerlimit 33 br i1 %continue, label %loop, label %exit 34 35exit: 36 %result = phi i32 [ 0, %entry ], [ %loop.acc.next, %loop ] 37 ret i32 %result 38} 39 40define i32 @unsigned_reverse_loop_n_to_lower_limit(i32* %array, i32 %length, i32 %n, i32 %lowerlimit) { 41; CHECK-LABEL: @unsigned_reverse_loop_n_to_lower_limit( 42entry: 43 %tmp5 = icmp eq i32 %n, 0 44 br i1 %tmp5, label %exit, label %loop.preheader 45 46; CHECK: loop.preheader: 47; CHECK-NEXT: [[range_start:%.*]] = add i32 %n, -1 48; CHECK-NEXT: [[first_iteration_check:%.*]] = icmp ult i32 [[range_start]], %length 49; CHECK-NEXT: [[no_wrap_check:%.*]] = icmp uge i32 %lowerlimit, 1 50; CHECK-NEXT: [[wide_cond:%.*]] = and i1 [[first_iteration_check]], [[no_wrap_check]] 51loop.preheader: 52 br label %loop 53 54; CHECK: loop: 55; CHECK: call void (i1, ...) @llvm.experimental.guard(i1 [[wide_cond]], i32 9) [ "deopt"() ] 56loop: 57 %loop.acc = phi i32 [ %loop.acc.next, %loop ], [ 0, %loop.preheader ] 58 %i = phi i32 [ %i.next, %loop ], [ %n, %loop.preheader ] 59 %i.next = add nsw i32 %i, -1 60 %within.bounds = icmp ult i32 %i.next, %length 61 call void (i1, ...) @llvm.experimental.guard(i1 %within.bounds, i32 9) [ "deopt"() ] 62 %i.i64 = zext i32 %i.next to i64 63 %array.i.ptr = getelementptr inbounds i32, i32* %array, i64 %i.i64 64 %array.i = load i32, i32* %array.i.ptr, align 4 65 %loop.acc.next = add i32 %loop.acc, %array.i 66 %continue = icmp ugt i32 %i, %lowerlimit 67 br i1 %continue, label %loop, label %exit 68 69exit: 70 %result = phi i32 [ 0, %entry ], [ %loop.acc.next, %loop ] 71 ret i32 %result 72} 73 74 75; if we predicated the loop, the guard will definitely fail and we will 76; deoptimize early on. 77define i32 @unsigned_reverse_loop_n_to_0(i32* %array, i32 %length, i32 %n, i32 %lowerlimit) { 78; CHECK-LABEL: @unsigned_reverse_loop_n_to_0( 79entry: 80 %tmp5 = icmp eq i32 %n, 0 81 br i1 %tmp5, label %exit, label %loop.preheader 82 83; CHECK: loop.preheader: 84; CHECK-NEXT: [[range_start:%.*]] = add i32 %n, -1 85; CHECK-NEXT: [[first_iteration_check:%.*]] = icmp ult i32 [[range_start]], %length 86; CHECK-NEXT: [[wide_cond:%.*]] = and i1 [[first_iteration_check]], false 87loop.preheader: 88 br label %loop 89 90; CHECK: loop: 91; CHECK: call void (i1, ...) @llvm.experimental.guard(i1 [[wide_cond]], i32 9) [ "deopt"() ] 92loop: 93 %loop.acc = phi i32 [ %loop.acc.next, %loop ], [ 0, %loop.preheader ] 94 %i = phi i32 [ %i.next, %loop ], [ %n, %loop.preheader ] 95 %i.next = add nsw i32 %i, -1 96 %within.bounds = icmp ult i32 %i.next, %length 97 call void (i1, ...) @llvm.experimental.guard(i1 %within.bounds, i32 9) [ "deopt"() ] 98 %i.i64 = zext i32 %i.next to i64 99 %array.i.ptr = getelementptr inbounds i32, i32* %array, i64 %i.i64 100 %array.i = load i32, i32* %array.i.ptr, align 4 101 %loop.acc.next = add i32 %loop.acc, %array.i 102 %continue = icmp ugt i32 %i, 0 103 br i1 %continue, label %loop, label %exit 104 105exit: 106 %result = phi i32 [ 0, %entry ], [ %loop.acc.next, %loop ] 107 ret i32 %result 108} 109 110; do not loop predicate when the range has step -1 and latch has step 1. 111define i32 @reverse_loop_range_step_increment(i32 %n, i32* %array, i32 %length) { 112; CHECK-LABEL: @reverse_loop_range_step_increment( 113entry: 114 %tmp5 = icmp eq i32 %n, 0 115 br i1 %tmp5, label %exit, label %loop.preheader 116 117loop.preheader: 118 br label %loop 119 120; CHECK: loop: 121; CHECK: llvm.experimental.guard(i1 %within.bounds, i32 9) 122loop: 123 %loop.acc = phi i32 [ %loop.acc.next, %loop ], [ 0, %loop.preheader ] 124 %i = phi i32 [ %i.next, %loop ], [ %n, %loop.preheader ] 125 %irc = phi i32 [ %i.inc, %loop ], [ 1, %loop.preheader ] 126 %i.inc = add nuw nsw i32 %irc, 1 127 %within.bounds = icmp ult i32 %irc, %length 128 call void (i1, ...) @llvm.experimental.guard(i1 %within.bounds, i32 9) [ "deopt"() ] 129 %i.i64 = zext i32 %irc to i64 130 %array.i.ptr = getelementptr inbounds i32, i32* %array, i64 %i.i64 131 %array.i = load i32, i32* %array.i.ptr, align 4 132 %i.next = add nsw i32 %i, -1 133 %loop.acc.next = add i32 %loop.acc, %array.i 134 %continue = icmp ugt i32 %i, 65534 135 br i1 %continue, label %loop, label %exit 136 137exit: 138 %result = phi i32 [ 0, %entry ], [ %loop.acc.next, %loop ] 139 ret i32 %result 140} 141 142define i32 @signed_reverse_loop_n_to_lower_limit_equal(i32* %array, i32 %length, i32 %n, i32 %lowerlimit) { 143; CHECK-LABEL: @signed_reverse_loop_n_to_lower_limit_equal( 144entry: 145 %tmp5 = icmp eq i32 %n, 0 146 br i1 %tmp5, label %exit, label %loop.preheader 147 148; CHECK: loop.preheader: 149; CHECK-NEXT: [[range_start:%.*]] = add i32 %n, -1 150; CHECK-NEXT: [[first_iteration_check:%.*]] = icmp ult i32 [[range_start]], %length 151; CHECK-NEXT: [[no_wrap_check:%.*]] = icmp sgt i32 %lowerlimit, 1 152; CHECK-NEXT: [[wide_cond:%.*]] = and i1 [[first_iteration_check]], [[no_wrap_check]] 153loop.preheader: 154 br label %loop 155 156; CHECK: loop: 157; CHECK: call void (i1, ...) @llvm.experimental.guard(i1 [[wide_cond]], i32 9) [ "deopt"() ] 158loop: 159 %loop.acc = phi i32 [ %loop.acc.next, %loop ], [ 0, %loop.preheader ] 160 %i = phi i32 [ %i.next, %loop ], [ %n, %loop.preheader ] 161 %i.next = add nsw i32 %i, -1 162 %within.bounds = icmp ult i32 %i.next, %length 163 call void (i1, ...) @llvm.experimental.guard(i1 %within.bounds, i32 9) [ "deopt"() ] 164 %i.i64 = zext i32 %i.next to i64 165 %array.i.ptr = getelementptr inbounds i32, i32* %array, i64 %i.i64 166 %array.i = load i32, i32* %array.i.ptr, align 4 167 %loop.acc.next = add i32 %loop.acc, %array.i 168 %continue = icmp sge i32 %i, %lowerlimit 169 br i1 %continue, label %loop, label %exit 170 171exit: 172 %result = phi i32 [ 0, %entry ], [ %loop.acc.next, %loop ] 173 ret i32 %result 174} 175 176define i32 @unsigned_reverse_loop_n_to_lower_limit_equal(i32* %array, i32 %length, i32 %n, i32 %lowerlimit) { 177; CHECK-LABEL: @unsigned_reverse_loop_n_to_lower_limit_equal( 178entry: 179 %tmp5 = icmp eq i32 %n, 0 180 br i1 %tmp5, label %exit, label %loop.preheader 181 182; CHECK: loop.preheader: 183; CHECK-NEXT: [[range_start:%.*]] = add i32 %n, -1 184; CHECK-NEXT: [[first_iteration_check:%.*]] = icmp ult i32 [[range_start]], %length 185; CHECK-NEXT: [[no_wrap_check:%.*]] = icmp ugt i32 %lowerlimit, 1 186; CHECK-NEXT: [[wide_cond:%.*]] = and i1 [[first_iteration_check]], [[no_wrap_check]] 187loop.preheader: 188 br label %loop 189 190; CHECK: loop: 191; CHECK: call void (i1, ...) @llvm.experimental.guard(i1 [[wide_cond]], i32 9) [ "deopt"() ] 192loop: 193 %loop.acc = phi i32 [ %loop.acc.next, %loop ], [ 0, %loop.preheader ] 194 %i = phi i32 [ %i.next, %loop ], [ %n, %loop.preheader ] 195 %i.next = add nsw i32 %i, -1 196 %within.bounds = icmp ult i32 %i.next, %length 197 call void (i1, ...) @llvm.experimental.guard(i1 %within.bounds, i32 9) [ "deopt"() ] 198 %i.i64 = zext i32 %i.next to i64 199 %array.i.ptr = getelementptr inbounds i32, i32* %array, i64 %i.i64 200 %array.i = load i32, i32* %array.i.ptr, align 4 201 %loop.acc.next = add i32 %loop.acc, %array.i 202 %continue = icmp uge i32 %i, %lowerlimit 203 br i1 %continue, label %loop, label %exit 204 205exit: 206 %result = phi i32 [ 0, %entry ], [ %loop.acc.next, %loop ] 207 ret i32 %result 208} 209 210 211; if we predicated the loop, the guard will definitely fail and we will 212; deoptimize early on. 213define i32 @unsigned_reverse_loop_n_to_1(i32* %array, i32 %length, i32 %n, i32 %lowerlimit) { 214; CHECK-LABEL: @unsigned_reverse_loop_n_to_1( 215entry: 216 %tmp5 = icmp eq i32 %n, 0 217 br i1 %tmp5, label %exit, label %loop.preheader 218 219; CHECK: loop.preheader: 220; CHECK-NEXT: [[range_start:%.*]] = add i32 %n, -1 221; CHECK-NEXT: [[first_iteration_check:%.*]] = icmp ult i32 [[range_start]], %length 222; CHECK-NEXT: [[wide_cond:%.*]] = and i1 [[first_iteration_check]], false 223loop.preheader: 224 br label %loop 225 226; CHECK: loop: 227; CHECK: call void (i1, ...) @llvm.experimental.guard(i1 [[wide_cond]], i32 9) [ "deopt"() ] 228loop: 229 %loop.acc = phi i32 [ %loop.acc.next, %loop ], [ 0, %loop.preheader ] 230 %i = phi i32 [ %i.next, %loop ], [ %n, %loop.preheader ] 231 %i.next = add nsw i32 %i, -1 232 %within.bounds = icmp ult i32 %i.next, %length 233 call void (i1, ...) @llvm.experimental.guard(i1 %within.bounds, i32 9) [ "deopt"() ] 234 %i.i64 = zext i32 %i.next to i64 235 %array.i.ptr = getelementptr inbounds i32, i32* %array, i64 %i.i64 236 %array.i = load i32, i32* %array.i.ptr, align 4 237 %loop.acc.next = add i32 %loop.acc, %array.i 238 %continue = icmp uge i32 %i, 1 239 br i1 %continue, label %loop, label %exit 240 241exit: 242 %result = phi i32 [ 0, %entry ], [ %loop.acc.next, %loop ] 243 ret i32 %result 244} 245 246