1; RUN: opt < %s -loop-unroll -S | FileCheck %s 2 3target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128" 4target triple = "x86_64-unknown-linux-gnu" 5 6; loop4 contains a small loop which should be completely unrolled by 7; the default unrolling heuristics. It serves as a control for the 8; unroll(disable) pragma test loop4_with_disable. 9; 10; CHECK-LABEL: @loop4( 11; CHECK-NOT: br i1 12define void @loop4(i32* nocapture %a) { 13entry: 14 br label %for.body 15 16for.body: ; preds = %for.body, %entry 17 %indvars.iv = phi i64 [ 0, %entry ], [ %indvars.iv.next, %for.body ] 18 %arrayidx = getelementptr inbounds i32* %a, i64 %indvars.iv 19 %0 = load i32* %arrayidx, align 4 20 %inc = add nsw i32 %0, 1 21 store i32 %inc, i32* %arrayidx, align 4 22 %indvars.iv.next = add nuw nsw i64 %indvars.iv, 1 23 %exitcond = icmp eq i64 %indvars.iv.next, 4 24 br i1 %exitcond, label %for.end, label %for.body 25 26for.end: ; preds = %for.body 27 ret void 28} 29 30; #pragma clang loop unroll(disable) 31; 32; CHECK-LABEL: @loop4_with_disable( 33; CHECK: store i32 34; CHECK-NOT: store i32 35; CHECK: br i1 36define void @loop4_with_disable(i32* nocapture %a) { 37entry: 38 br label %for.body 39 40for.body: ; preds = %for.body, %entry 41 %indvars.iv = phi i64 [ 0, %entry ], [ %indvars.iv.next, %for.body ] 42 %arrayidx = getelementptr inbounds i32* %a, i64 %indvars.iv 43 %0 = load i32* %arrayidx, align 4 44 %inc = add nsw i32 %0, 1 45 store i32 %inc, i32* %arrayidx, align 4 46 %indvars.iv.next = add nuw nsw i64 %indvars.iv, 1 47 %exitcond = icmp eq i64 %indvars.iv.next, 4 48 br i1 %exitcond, label %for.end, label %for.body, !llvm.loop !1 49 50for.end: ; preds = %for.body 51 ret void 52} 53!1 = metadata !{metadata !1, metadata !2} 54!2 = metadata !{metadata !"llvm.loop.unroll.enable", i1 false} 55 56; loop64 has a high enough count that it should *not* be unrolled by 57; the default unrolling heuristic. It serves as the control for the 58; unroll(enable) pragma test loop64_with_.* tests below. 59; 60; CHECK-LABEL: @loop64( 61; CHECK: store i32 62; CHECK-NOT: store i32 63; CHECK: br i1 64define void @loop64(i32* nocapture %a) { 65entry: 66 br label %for.body 67 68for.body: ; preds = %for.body, %entry 69 %indvars.iv = phi i64 [ 0, %entry ], [ %indvars.iv.next, %for.body ] 70 %arrayidx = getelementptr inbounds i32* %a, i64 %indvars.iv 71 %0 = load i32* %arrayidx, align 4 72 %inc = add nsw i32 %0, 1 73 store i32 %inc, i32* %arrayidx, align 4 74 %indvars.iv.next = add nuw nsw i64 %indvars.iv, 1 75 %exitcond = icmp eq i64 %indvars.iv.next, 64 76 br i1 %exitcond, label %for.end, label %for.body 77 78for.end: ; preds = %for.body 79 ret void 80} 81 82; #pragma clang loop unroll(enable) 83; Loop should be fully unrolled. 84; 85; CHECK-LABEL: @loop64_with_enable( 86; CHECK-NOT: br i1 87define void @loop64_with_enable(i32* nocapture %a) { 88entry: 89 br label %for.body 90 91for.body: ; preds = %for.body, %entry 92 %indvars.iv = phi i64 [ 0, %entry ], [ %indvars.iv.next, %for.body ] 93 %arrayidx = getelementptr inbounds i32* %a, i64 %indvars.iv 94 %0 = load i32* %arrayidx, align 4 95 %inc = add nsw i32 %0, 1 96 store i32 %inc, i32* %arrayidx, align 4 97 %indvars.iv.next = add nuw nsw i64 %indvars.iv, 1 98 %exitcond = icmp eq i64 %indvars.iv.next, 64 99 br i1 %exitcond, label %for.end, label %for.body, !llvm.loop !3 100 101for.end: ; preds = %for.body 102 ret void 103} 104!3 = metadata !{metadata !3, metadata !4} 105!4 = metadata !{metadata !"llvm.loop.unroll.enable", i1 true} 106 107; #pragma clang loop unroll_count(4) 108; Loop should be unrolled 4 times. 109; 110; CHECK-LABEL: @loop64_with_count4( 111; CHECK: store i32 112; CHECK: store i32 113; CHECK: store i32 114; CHECK: store i32 115; CHECK-NOT: store i32 116; CHECK: br i1 117define void @loop64_with_count4(i32* nocapture %a) { 118entry: 119 br label %for.body 120 121for.body: ; preds = %for.body, %entry 122 %indvars.iv = phi i64 [ 0, %entry ], [ %indvars.iv.next, %for.body ] 123 %arrayidx = getelementptr inbounds i32* %a, i64 %indvars.iv 124 %0 = load i32* %arrayidx, align 4 125 %inc = add nsw i32 %0, 1 126 store i32 %inc, i32* %arrayidx, align 4 127 %indvars.iv.next = add nuw nsw i64 %indvars.iv, 1 128 %exitcond = icmp eq i64 %indvars.iv.next, 64 129 br i1 %exitcond, label %for.end, label %for.body, !llvm.loop !5 130 131for.end: ; preds = %for.body 132 ret void 133} 134!5 = metadata !{metadata !5, metadata !6} 135!6 = metadata !{metadata !"llvm.loop.unroll.count", i32 4} 136 137 138; #pragma clang loop unroll_count(enable) unroll_count(4) 139; Loop should be unrolled 4 times. 140; 141; CHECK-LABEL: @loop64_with_enable_and_count4( 142; CHECK: store i32 143; CHECK: store i32 144; CHECK: store i32 145; CHECK: store i32 146; CHECK-NOT: store i32 147; CHECK: br i1 148define void @loop64_with_enable_and_count4(i32* nocapture %a) { 149entry: 150 br label %for.body 151 152for.body: ; preds = %for.body, %entry 153 %indvars.iv = phi i64 [ 0, %entry ], [ %indvars.iv.next, %for.body ] 154 %arrayidx = getelementptr inbounds i32* %a, i64 %indvars.iv 155 %0 = load i32* %arrayidx, align 4 156 %inc = add nsw i32 %0, 1 157 store i32 %inc, i32* %arrayidx, align 4 158 %indvars.iv.next = add nuw nsw i64 %indvars.iv, 1 159 %exitcond = icmp eq i64 %indvars.iv.next, 64 160 br i1 %exitcond, label %for.end, label %for.body, !llvm.loop !7 161 162for.end: ; preds = %for.body 163 ret void 164} 165!7 = metadata !{metadata !7, metadata !6, metadata !4} 166 167; #pragma clang loop unroll_count(enable) 168; Full unrolling is requested, but loop has a dynamic trip count so 169; no unrolling should occur. 170; 171; CHECK-LABEL: @dynamic_loop_with_enable( 172; CHECK: store i32 173; CHECK-NOT: store i32 174; CHECK: br i1 175define void @dynamic_loop_with_enable(i32* nocapture %a, i32 %b) { 176entry: 177 %cmp3 = icmp sgt i32 %b, 0 178 br i1 %cmp3, label %for.body, label %for.end, !llvm.loop !8 179 180for.body: ; preds = %entry, %for.body 181 %indvars.iv = phi i64 [ %indvars.iv.next, %for.body ], [ 0, %entry ] 182 %arrayidx = getelementptr inbounds i32* %a, i64 %indvars.iv 183 %0 = load i32* %arrayidx, align 4 184 %inc = add nsw i32 %0, 1 185 store i32 %inc, i32* %arrayidx, align 4 186 %indvars.iv.next = add nuw nsw i64 %indvars.iv, 1 187 %lftr.wideiv = trunc i64 %indvars.iv.next to i32 188 %exitcond = icmp eq i32 %lftr.wideiv, %b 189 br i1 %exitcond, label %for.end, label %for.body, !llvm.loop !8 190 191for.end: ; preds = %for.body, %entry 192 ret void 193} 194!8 = metadata !{metadata !8, metadata !4} 195 196; #pragma clang loop unroll_count(4) 197; Loop has a dynamic trip count. Unrolling should occur, but no 198; conditional branches can be removed. 199; 200; CHECK-LABEL: @dynamic_loop_with_count4( 201; CHECK-NOT: store 202; CHECK: br i1 203; CHECK: store 204; CHECK: br i1 205; CHECK: store 206; CHECK: br i1 207; CHECK: store 208; CHECK: br i1 209; CHECK: store 210; CHECK: br i1 211; CHECK-NOT: br i1 212define void @dynamic_loop_with_count4(i32* nocapture %a, i32 %b) { 213entry: 214 %cmp3 = icmp sgt i32 %b, 0 215 br i1 %cmp3, label %for.body, label %for.end, !llvm.loop !9 216 217for.body: ; preds = %entry, %for.body 218 %indvars.iv = phi i64 [ %indvars.iv.next, %for.body ], [ 0, %entry ] 219 %arrayidx = getelementptr inbounds i32* %a, i64 %indvars.iv 220 %0 = load i32* %arrayidx, align 4 221 %inc = add nsw i32 %0, 1 222 store i32 %inc, i32* %arrayidx, align 4 223 %indvars.iv.next = add nuw nsw i64 %indvars.iv, 1 224 %lftr.wideiv = trunc i64 %indvars.iv.next to i32 225 %exitcond = icmp eq i32 %lftr.wideiv, %b 226 br i1 %exitcond, label %for.end, label %for.body, !llvm.loop !9 227 228for.end: ; preds = %for.body, %entry 229 ret void 230} 231!9 = metadata !{metadata !9, metadata !6} 232 233; #pragma clang loop unroll_count(1) 234; Loop should not be unrolled 235; 236; CHECK-LABEL: @unroll_1( 237; CHECK: store i32 238; CHECK-NOT: store i32 239; CHECK: br i1 240define void @unroll_1(i32* nocapture %a, i32 %b) { 241entry: 242 br label %for.body 243 244for.body: ; preds = %for.body, %entry 245 %indvars.iv = phi i64 [ 0, %entry ], [ %indvars.iv.next, %for.body ] 246 %arrayidx = getelementptr inbounds i32* %a, i64 %indvars.iv 247 %0 = load i32* %arrayidx, align 4 248 %inc = add nsw i32 %0, 1 249 store i32 %inc, i32* %arrayidx, align 4 250 %indvars.iv.next = add nuw nsw i64 %indvars.iv, 1 251 %exitcond = icmp eq i64 %indvars.iv.next, 4 252 br i1 %exitcond, label %for.end, label %for.body, !llvm.loop !10 253 254for.end: ; preds = %for.body 255 ret void 256} 257!10 = metadata !{metadata !10, metadata !11} 258!11 = metadata !{metadata !"llvm.loop.unroll.count", i32 1} 259 260; #pragma clang loop unroll(enable) 261; Loop has very high loop count (1 million) and full unrolling was requested. 262; Loop should unrolled up to the pragma threshold, but not completely. 263; 264; CHECK-LABEL: @unroll_1M( 265; CHECK: store i32 266; CHECK: store i32 267; CHECK: br i1 268define void @unroll_1M(i32* nocapture %a, i32 %b) { 269entry: 270 br label %for.body 271 272for.body: ; preds = %for.body, %entry 273 %indvars.iv = phi i64 [ 0, %entry ], [ %indvars.iv.next, %for.body ] 274 %arrayidx = getelementptr inbounds i32* %a, i64 %indvars.iv 275 %0 = load i32* %arrayidx, align 4 276 %inc = add nsw i32 %0, 1 277 store i32 %inc, i32* %arrayidx, align 4 278 %indvars.iv.next = add nuw nsw i64 %indvars.iv, 1 279 %exitcond = icmp eq i64 %indvars.iv.next, 1000000 280 br i1 %exitcond, label %for.end, label %for.body, !llvm.loop !12 281 282for.end: ; preds = %for.body 283 ret void 284} 285!12 = metadata !{metadata !12, metadata !4} 286