1; NOTE: Assertions have been autogenerated by utils/update_test_checks.py 2; RUN: opt -loop-vectorize -force-vector-width=4 -force-vector-interleave=1 -S %s | FileCheck %s 3 4 5@p = external local_unnamed_addr global [257 x i32], align 16 6@q = external local_unnamed_addr global [257 x i32], align 16 7 8; Test case for PR43398. 9 10define void @can_sink_after_store(i32 %x, i32* %ptr, i64 %tc) local_unnamed_addr #0 { 11; CHECK-LABEL: @can_sink_after_store( 12; CHECK-NEXT: entry: 13; CHECK-NEXT: br label [[PREHEADER:%.*]] 14; CHECK: preheader: 15; CHECK-NEXT: [[IDX_PHI_TRANS:%.*]] = getelementptr inbounds [257 x i32], [257 x i32]* @p, i64 0, i64 1 16; CHECK-NEXT: [[DOTPRE:%.*]] = load i32, i32* [[IDX_PHI_TRANS]], align 4 17; CHECK-NEXT: br i1 false, label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]] 18; CHECK: vector.ph: 19; CHECK-NEXT: [[VECTOR_RECUR_INIT:%.*]] = insertelement <4 x i32> poison, i32 [[DOTPRE]], i32 3 20; CHECK-NEXT: [[BROADCAST_SPLATINSERT:%.*]] = insertelement <4 x i32> poison, i32 [[X:%.*]], i32 0 21; CHECK-NEXT: [[BROADCAST_SPLAT:%.*]] = shufflevector <4 x i32> [[BROADCAST_SPLATINSERT]], <4 x i32> poison, <4 x i32> zeroinitializer 22; CHECK-NEXT: br label [[VECTOR_BODY:%.*]] 23; CHECK: vector.body: 24; CHECK-NEXT: [[INDEX:%.*]] = phi i64 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ] 25; CHECK-NEXT: [[VECTOR_RECUR:%.*]] = phi <4 x i32> [ [[VECTOR_RECUR_INIT]], [[VECTOR_PH]] ], [ [[WIDE_LOAD:%.*]], [[VECTOR_BODY]] ] 26; CHECK-NEXT: [[OFFSET_IDX:%.*]] = add i64 1, [[INDEX]] 27; CHECK-NEXT: [[TMP0:%.*]] = add i64 [[OFFSET_IDX]], 0 28; CHECK-NEXT: [[TMP1:%.*]] = getelementptr inbounds [257 x i32], [257 x i32]* @p, i64 0, i64 [[TMP0]] 29; CHECK-NEXT: [[TMP2:%.*]] = getelementptr inbounds i32, i32* [[TMP1]], i32 0 30; CHECK-NEXT: [[TMP3:%.*]] = bitcast i32* [[TMP2]] to <4 x i32>* 31; CHECK-NEXT: [[WIDE_LOAD]] = load <4 x i32>, <4 x i32>* [[TMP3]], align 4 32; CHECK-NEXT: [[TMP4:%.*]] = shufflevector <4 x i32> [[VECTOR_RECUR]], <4 x i32> [[WIDE_LOAD]], <4 x i32> <i32 3, i32 4, i32 5, i32 6> 33; CHECK-NEXT: [[TMP5:%.*]] = add <4 x i32> [[TMP4]], [[BROADCAST_SPLAT]] 34; CHECK-NEXT: [[TMP6:%.*]] = add <4 x i32> [[TMP5]], [[WIDE_LOAD]] 35; CHECK-NEXT: [[TMP7:%.*]] = getelementptr inbounds [257 x i32], [257 x i32]* @q, i64 0, i64 [[TMP0]] 36; CHECK-NEXT: [[TMP8:%.*]] = getelementptr inbounds i32, i32* [[TMP7]], i32 0 37; CHECK-NEXT: [[TMP9:%.*]] = bitcast i32* [[TMP8]] to <4 x i32>* 38; CHECK-NEXT: store <4 x i32> [[TMP6]], <4 x i32>* [[TMP9]], align 4 39; CHECK-NEXT: [[INDEX_NEXT]] = add nuw i64 [[INDEX]], 4 40; CHECK-NEXT: [[TMP10:%.*]] = icmp eq i64 [[INDEX_NEXT]], 1996 41; CHECK-NEXT: br i1 [[TMP10]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP0:![0-9]+]] 42; CHECK: middle.block: 43; CHECK-NEXT: [[CMP_N:%.*]] = icmp eq i64 1999, 1996 44; CHECK-NEXT: [[VECTOR_RECUR_EXTRACT:%.*]] = extractelement <4 x i32> [[WIDE_LOAD]], i32 3 45; CHECK-NEXT: [[VECTOR_RECUR_EXTRACT_FOR_PHI:%.*]] = extractelement <4 x i32> [[WIDE_LOAD]], i32 2 46; CHECK-NEXT: br i1 [[CMP_N]], label [[EXIT:%.*]], label [[SCALAR_PH]] 47; CHECK: scalar.ph: 48; CHECK-NEXT: [[SCALAR_RECUR_INIT:%.*]] = phi i32 [ [[DOTPRE]], [[PREHEADER]] ], [ [[VECTOR_RECUR_EXTRACT]], [[MIDDLE_BLOCK]] ] 49; CHECK-NEXT: [[BC_RESUME_VAL:%.*]] = phi i64 [ 1997, [[MIDDLE_BLOCK]] ], [ 1, [[PREHEADER]] ] 50; CHECK-NEXT: br label [[FOR:%.*]] 51; CHECK: for: 52; CHECK-NEXT: [[SCALAR_RECUR:%.*]] = phi i32 [ [[SCALAR_RECUR_INIT]], [[SCALAR_PH]] ], [ [[PRE_NEXT:%.*]], [[FOR]] ] 53; CHECK-NEXT: [[IV:%.*]] = phi i64 [ [[BC_RESUME_VAL]], [[SCALAR_PH]] ], [ [[IV_NEXT:%.*]], [[FOR]] ] 54; CHECK-NEXT: [[ADD_1:%.*]] = add i32 [[SCALAR_RECUR]], [[X]] 55; CHECK-NEXT: [[IDX_1:%.*]] = getelementptr inbounds [257 x i32], [257 x i32]* @p, i64 0, i64 [[IV]] 56; CHECK-NEXT: [[PRE_NEXT]] = load i32, i32* [[IDX_1]], align 4 57; CHECK-NEXT: [[ADD_2:%.*]] = add i32 [[ADD_1]], [[PRE_NEXT]] 58; CHECK-NEXT: [[IDX_2:%.*]] = getelementptr inbounds [257 x i32], [257 x i32]* @q, i64 0, i64 [[IV]] 59; CHECK-NEXT: store i32 [[ADD_2]], i32* [[IDX_2]], align 4 60; CHECK-NEXT: [[IV_NEXT]] = add nuw nsw i64 [[IV]], 1 61; CHECK-NEXT: [[EXITCOND:%.*]] = icmp eq i64 [[IV_NEXT]], 2000 62; CHECK-NEXT: br i1 [[EXITCOND]], label [[EXIT]], label [[FOR]], !llvm.loop [[LOOP2:![0-9]+]] 63; CHECK: exit: 64; CHECK-NEXT: ret void 65; 66 67entry: 68 br label %preheader 69 70preheader: 71 %idx.phi.trans = getelementptr inbounds [257 x i32], [257 x i32]* @p, i64 0, i64 1 72 %.pre = load i32, i32* %idx.phi.trans, align 4 73 br label %for 74 75for: 76 %pre.phi = phi i32 [ %.pre, %preheader ], [ %pre.next, %for ] 77 %iv = phi i64 [ 1, %preheader ], [ %iv.next, %for ] 78 %add.1 = add i32 %pre.phi, %x 79 %idx.1 = getelementptr inbounds [257 x i32], [257 x i32]* @p, i64 0, i64 %iv 80 %pre.next = load i32, i32* %idx.1, align 4 81 %add.2 = add i32 %add.1, %pre.next 82 %idx.2 = getelementptr inbounds [257 x i32], [257 x i32]* @q, i64 0, i64 %iv 83 store i32 %add.2, i32* %idx.2, align 4 84 %iv.next = add nuw nsw i64 %iv, 1 85 %exitcond = icmp eq i64 %iv.next, 2000 86 br i1 %exitcond, label %exit, label %for 87 88exit: 89 ret void 90} 91 92; We can sink potential trapping instructions, as this will only delay the trap 93; and not introduce traps on additional paths. 94define void @sink_sdiv(i32 %x, i32* %ptr, i64 %tc) local_unnamed_addr #0 { 95; CHECK-LABEL: @sink_sdiv( 96; CHECK-NEXT: entry: 97; CHECK-NEXT: br label [[PREHEADER:%.*]] 98; CHECK: preheader: 99; CHECK-NEXT: [[IDX_PHI_TRANS:%.*]] = getelementptr inbounds [257 x i32], [257 x i32]* @p, i64 0, i64 1 100; CHECK-NEXT: [[DOTPRE:%.*]] = load i32, i32* [[IDX_PHI_TRANS]], align 4 101; CHECK-NEXT: br i1 false, label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]] 102; CHECK: vector.ph: 103; CHECK-NEXT: [[VECTOR_RECUR_INIT:%.*]] = insertelement <4 x i32> poison, i32 [[DOTPRE]], i32 3 104; CHECK-NEXT: [[BROADCAST_SPLATINSERT:%.*]] = insertelement <4 x i32> poison, i32 [[X:%.*]], i32 0 105; CHECK-NEXT: [[BROADCAST_SPLAT:%.*]] = shufflevector <4 x i32> [[BROADCAST_SPLATINSERT]], <4 x i32> poison, <4 x i32> zeroinitializer 106; CHECK-NEXT: br label [[VECTOR_BODY:%.*]] 107; CHECK: vector.body: 108; CHECK-NEXT: [[INDEX:%.*]] = phi i64 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ] 109; CHECK-NEXT: [[VECTOR_RECUR:%.*]] = phi <4 x i32> [ [[VECTOR_RECUR_INIT]], [[VECTOR_PH]] ], [ [[WIDE_LOAD:%.*]], [[VECTOR_BODY]] ] 110; CHECK-NEXT: [[OFFSET_IDX:%.*]] = add i64 1, [[INDEX]] 111; CHECK-NEXT: [[TMP0:%.*]] = add i64 [[OFFSET_IDX]], 0 112; CHECK-NEXT: [[TMP1:%.*]] = getelementptr inbounds [257 x i32], [257 x i32]* @p, i64 0, i64 [[TMP0]] 113; CHECK-NEXT: [[TMP2:%.*]] = getelementptr inbounds i32, i32* [[TMP1]], i32 0 114; CHECK-NEXT: [[TMP3:%.*]] = bitcast i32* [[TMP2]] to <4 x i32>* 115; CHECK-NEXT: [[WIDE_LOAD]] = load <4 x i32>, <4 x i32>* [[TMP3]], align 4 116; CHECK-NEXT: [[TMP4:%.*]] = shufflevector <4 x i32> [[VECTOR_RECUR]], <4 x i32> [[WIDE_LOAD]], <4 x i32> <i32 3, i32 4, i32 5, i32 6> 117; CHECK-NEXT: [[TMP5:%.*]] = sdiv <4 x i32> [[TMP4]], [[BROADCAST_SPLAT]] 118; CHECK-NEXT: [[TMP6:%.*]] = add <4 x i32> [[TMP5]], [[WIDE_LOAD]] 119; CHECK-NEXT: [[TMP7:%.*]] = getelementptr inbounds [257 x i32], [257 x i32]* @q, i64 0, i64 [[TMP0]] 120; CHECK-NEXT: [[TMP8:%.*]] = getelementptr inbounds i32, i32* [[TMP7]], i32 0 121; CHECK-NEXT: [[TMP9:%.*]] = bitcast i32* [[TMP8]] to <4 x i32>* 122; CHECK-NEXT: store <4 x i32> [[TMP6]], <4 x i32>* [[TMP9]], align 4 123; CHECK-NEXT: [[INDEX_NEXT]] = add nuw i64 [[INDEX]], 4 124; CHECK-NEXT: [[TMP10:%.*]] = icmp eq i64 [[INDEX_NEXT]], 1996 125; CHECK-NEXT: br i1 [[TMP10]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP4:![0-9]+]] 126; CHECK: middle.block: 127; CHECK-NEXT: [[CMP_N:%.*]] = icmp eq i64 1999, 1996 128; CHECK-NEXT: [[VECTOR_RECUR_EXTRACT:%.*]] = extractelement <4 x i32> [[WIDE_LOAD]], i32 3 129; CHECK-NEXT: [[VECTOR_RECUR_EXTRACT_FOR_PHI:%.*]] = extractelement <4 x i32> [[WIDE_LOAD]], i32 2 130; CHECK-NEXT: br i1 [[CMP_N]], label [[EXIT:%.*]], label [[SCALAR_PH]] 131; CHECK: scalar.ph: 132; CHECK-NEXT: [[SCALAR_RECUR_INIT:%.*]] = phi i32 [ [[DOTPRE]], [[PREHEADER]] ], [ [[VECTOR_RECUR_EXTRACT]], [[MIDDLE_BLOCK]] ] 133; CHECK-NEXT: [[BC_RESUME_VAL:%.*]] = phi i64 [ 1997, [[MIDDLE_BLOCK]] ], [ 1, [[PREHEADER]] ] 134; CHECK-NEXT: br label [[FOR:%.*]] 135; CHECK: for: 136; CHECK-NEXT: [[SCALAR_RECUR:%.*]] = phi i32 [ [[SCALAR_RECUR_INIT]], [[SCALAR_PH]] ], [ [[PRE_NEXT:%.*]], [[FOR]] ] 137; CHECK-NEXT: [[IV:%.*]] = phi i64 [ [[BC_RESUME_VAL]], [[SCALAR_PH]] ], [ [[IV_NEXT:%.*]], [[FOR]] ] 138; CHECK-NEXT: [[DIV_1:%.*]] = sdiv i32 [[SCALAR_RECUR]], [[X]] 139; CHECK-NEXT: [[IDX_1:%.*]] = getelementptr inbounds [257 x i32], [257 x i32]* @p, i64 0, i64 [[IV]] 140; CHECK-NEXT: [[PRE_NEXT]] = load i32, i32* [[IDX_1]], align 4 141; CHECK-NEXT: [[ADD_2:%.*]] = add i32 [[DIV_1]], [[PRE_NEXT]] 142; CHECK-NEXT: [[IDX_2:%.*]] = getelementptr inbounds [257 x i32], [257 x i32]* @q, i64 0, i64 [[IV]] 143; CHECK-NEXT: store i32 [[ADD_2]], i32* [[IDX_2]], align 4 144; CHECK-NEXT: [[IV_NEXT]] = add nuw nsw i64 [[IV]], 1 145; CHECK-NEXT: [[EXITCOND:%.*]] = icmp eq i64 [[IV_NEXT]], 2000 146; CHECK-NEXT: br i1 [[EXITCOND]], label [[EXIT]], label [[FOR]], !llvm.loop [[LOOP5:![0-9]+]] 147; CHECK: exit: 148; CHECK-NEXT: ret void 149; 150 151entry: 152 br label %preheader 153 154preheader: 155 %idx.phi.trans = getelementptr inbounds [257 x i32], [257 x i32]* @p, i64 0, i64 1 156 %.pre = load i32, i32* %idx.phi.trans, align 4 157 br label %for 158 159for: 160 %pre.phi = phi i32 [ %.pre, %preheader ], [ %pre.next, %for ] 161 %iv = phi i64 [ 1, %preheader ], [ %iv.next, %for ] 162 %div.1 = sdiv i32 %pre.phi, %x 163 %idx.1 = getelementptr inbounds [257 x i32], [257 x i32]* @p, i64 0, i64 %iv 164 %pre.next = load i32, i32* %idx.1, align 4 165 %add.2 = add i32 %div.1, %pre.next 166 %idx.2 = getelementptr inbounds [257 x i32], [257 x i32]* @q, i64 0, i64 %iv 167 store i32 %add.2, i32* %idx.2, align 4 168 %iv.next = add nuw nsw i64 %iv, 1 169 %exitcond = icmp eq i64 %iv.next, 2000 170 br i1 %exitcond, label %exit, label %for 171 172exit: 173 ret void 174} 175 176; Sink users of %pre.phi recursively. 177define void @can_sink_with_additional_user(i32 %x, i32* %ptr, i64 %tc) { 178; CHECK-LABEL: @can_sink_with_additional_user( 179; CHECK-NEXT: entry: 180; CHECK-NEXT: br label [[PREHEADER:%.*]] 181; CHECK: preheader: 182; CHECK-NEXT: [[IDX_PHI_TRANS:%.*]] = getelementptr inbounds [257 x i32], [257 x i32]* @p, i64 0, i64 1 183; CHECK-NEXT: [[DOTPRE:%.*]] = load i32, i32* [[IDX_PHI_TRANS]], align 4 184; CHECK-NEXT: br i1 false, label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]] 185; CHECK: vector.ph: 186; CHECK-NEXT: [[VECTOR_RECUR_INIT:%.*]] = insertelement <4 x i32> poison, i32 [[DOTPRE]], i32 3 187; CHECK-NEXT: [[BROADCAST_SPLATINSERT:%.*]] = insertelement <4 x i32> poison, i32 [[X:%.*]], i32 0 188; CHECK-NEXT: [[BROADCAST_SPLAT:%.*]] = shufflevector <4 x i32> [[BROADCAST_SPLATINSERT]], <4 x i32> poison, <4 x i32> zeroinitializer 189; CHECK-NEXT: br label [[VECTOR_BODY:%.*]] 190; CHECK: vector.body: 191; CHECK-NEXT: [[INDEX:%.*]] = phi i64 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ] 192; CHECK-NEXT: [[VECTOR_RECUR:%.*]] = phi <4 x i32> [ [[VECTOR_RECUR_INIT]], [[VECTOR_PH]] ], [ [[WIDE_LOAD:%.*]], [[VECTOR_BODY]] ] 193; CHECK-NEXT: [[OFFSET_IDX:%.*]] = add i64 1, [[INDEX]] 194; CHECK-NEXT: [[TMP0:%.*]] = add i64 [[OFFSET_IDX]], 0 195; CHECK-NEXT: [[TMP1:%.*]] = getelementptr inbounds [257 x i32], [257 x i32]* @p, i64 0, i64 [[TMP0]] 196; CHECK-NEXT: [[TMP2:%.*]] = getelementptr inbounds i32, i32* [[TMP1]], i32 0 197; CHECK-NEXT: [[TMP3:%.*]] = bitcast i32* [[TMP2]] to <4 x i32>* 198; CHECK-NEXT: [[WIDE_LOAD]] = load <4 x i32>, <4 x i32>* [[TMP3]], align 4 199; CHECK-NEXT: [[TMP4:%.*]] = shufflevector <4 x i32> [[VECTOR_RECUR]], <4 x i32> [[WIDE_LOAD]], <4 x i32> <i32 3, i32 4, i32 5, i32 6> 200; CHECK-NEXT: [[TMP5:%.*]] = add <4 x i32> [[TMP4]], [[BROADCAST_SPLAT]] 201; CHECK-NEXT: [[TMP6:%.*]] = add <4 x i32> [[TMP5]], [[BROADCAST_SPLAT]] 202; CHECK-NEXT: [[TMP7:%.*]] = add <4 x i32> [[TMP5]], [[WIDE_LOAD]] 203; CHECK-NEXT: [[TMP8:%.*]] = add <4 x i32> [[TMP6]], [[TMP7]] 204; CHECK-NEXT: [[TMP9:%.*]] = getelementptr inbounds [257 x i32], [257 x i32]* @q, i64 0, i64 [[TMP0]] 205; CHECK-NEXT: [[TMP10:%.*]] = getelementptr inbounds i32, i32* [[TMP9]], i32 0 206; CHECK-NEXT: [[TMP11:%.*]] = bitcast i32* [[TMP10]] to <4 x i32>* 207; CHECK-NEXT: store <4 x i32> [[TMP8]], <4 x i32>* [[TMP11]], align 4 208; CHECK-NEXT: [[INDEX_NEXT]] = add nuw i64 [[INDEX]], 4 209; CHECK-NEXT: [[TMP12:%.*]] = icmp eq i64 [[INDEX_NEXT]], 1996 210; CHECK-NEXT: br i1 [[TMP12]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP6:![0-9]+]] 211; CHECK: middle.block: 212; CHECK-NEXT: [[CMP_N:%.*]] = icmp eq i64 1999, 1996 213; CHECK-NEXT: [[VECTOR_RECUR_EXTRACT:%.*]] = extractelement <4 x i32> [[WIDE_LOAD]], i32 3 214; CHECK-NEXT: [[VECTOR_RECUR_EXTRACT_FOR_PHI:%.*]] = extractelement <4 x i32> [[WIDE_LOAD]], i32 2 215; CHECK-NEXT: br i1 [[CMP_N]], label [[EXIT:%.*]], label [[SCALAR_PH]] 216; CHECK: scalar.ph: 217; CHECK-NEXT: [[SCALAR_RECUR_INIT:%.*]] = phi i32 [ [[DOTPRE]], [[PREHEADER]] ], [ [[VECTOR_RECUR_EXTRACT]], [[MIDDLE_BLOCK]] ] 218; CHECK-NEXT: [[BC_RESUME_VAL:%.*]] = phi i64 [ 1997, [[MIDDLE_BLOCK]] ], [ 1, [[PREHEADER]] ] 219; CHECK-NEXT: br label [[FOR:%.*]] 220; CHECK: for: 221; CHECK-NEXT: [[SCALAR_RECUR:%.*]] = phi i32 [ [[SCALAR_RECUR_INIT]], [[SCALAR_PH]] ], [ [[PRE_NEXT:%.*]], [[FOR]] ] 222; CHECK-NEXT: [[IV:%.*]] = phi i64 [ [[BC_RESUME_VAL]], [[SCALAR_PH]] ], [ [[IV_NEXT:%.*]], [[FOR]] ] 223; CHECK-NEXT: [[ADD_1:%.*]] = add i32 [[SCALAR_RECUR]], [[X]] 224; CHECK-NEXT: [[ADD_2:%.*]] = add i32 [[ADD_1]], [[X]] 225; CHECK-NEXT: [[IDX_1:%.*]] = getelementptr inbounds [257 x i32], [257 x i32]* @p, i64 0, i64 [[IV]] 226; CHECK-NEXT: [[PRE_NEXT]] = load i32, i32* [[IDX_1]], align 4 227; CHECK-NEXT: [[ADD_3:%.*]] = add i32 [[ADD_1]], [[PRE_NEXT]] 228; CHECK-NEXT: [[ADD_4:%.*]] = add i32 [[ADD_2]], [[ADD_3]] 229; CHECK-NEXT: [[IDX_2:%.*]] = getelementptr inbounds [257 x i32], [257 x i32]* @q, i64 0, i64 [[IV]] 230; CHECK-NEXT: store i32 [[ADD_4]], i32* [[IDX_2]], align 4 231; CHECK-NEXT: [[IV_NEXT]] = add nuw nsw i64 [[IV]], 1 232; CHECK-NEXT: [[EXITCOND:%.*]] = icmp eq i64 [[IV_NEXT]], 2000 233; CHECK-NEXT: br i1 [[EXITCOND]], label [[EXIT]], label [[FOR]], !llvm.loop [[LOOP7:![0-9]+]] 234; CHECK: exit: 235; CHECK-NEXT: ret void 236; 237 238 239 240entry: 241 br label %preheader 242 243preheader: 244 %idx.phi.trans = getelementptr inbounds [257 x i32], [257 x i32]* @p, i64 0, i64 1 245 %.pre = load i32, i32* %idx.phi.trans, align 4 246 br label %for 247 248for: 249 %pre.phi = phi i32 [ %.pre, %preheader ], [ %pre.next, %for ] 250 %iv = phi i64 [ 1, %preheader ], [ %iv.next, %for ] 251 %add.1 = add i32 %pre.phi, %x 252 %add.2 = add i32 %add.1, %x 253 %idx.1 = getelementptr inbounds [257 x i32], [257 x i32]* @p, i64 0, i64 %iv 254 %pre.next = load i32, i32* %idx.1, align 4 255 %add.3 = add i32 %add.1, %pre.next 256 %add.4 = add i32 %add.2, %add.3 257 %idx.2 = getelementptr inbounds [257 x i32], [257 x i32]* @q, i64 0, i64 %iv 258 store i32 %add.4, i32* %idx.2, align 4 259 %iv.next = add nuw nsw i64 %iv, 1 260 %exitcond = icmp eq i64 %iv.next, 2000 261 br i1 %exitcond, label %exit, label %for 262 263exit: 264 ret void 265} 266 267; FIXME: We can sink a store, if we can guarantee that it does not alias any 268; loads/stores in between. 269define void @cannot_sink_store(i32 %x, i32* %ptr, i64 %tc) { 270; CHECK-LABEL: @cannot_sink_store( 271; CHECK-NEXT: entry: 272; CHECK-NEXT: br label [[PREHEADER:%.*]] 273; CHECK: preheader: 274; CHECK-NEXT: [[IDX_PHI_TRANS:%.*]] = getelementptr inbounds [257 x i32], [257 x i32]* @p, i64 0, i64 1 275; CHECK-NEXT: [[DOTPRE:%.*]] = load i32, i32* [[IDX_PHI_TRANS]], align 4 276; CHECK-NEXT: br label [[FOR:%.*]] 277; CHECK: for: 278; CHECK-NEXT: [[PRE_PHI:%.*]] = phi i32 [ [[DOTPRE]], [[PREHEADER]] ], [ [[PRE_NEXT:%.*]], [[FOR]] ] 279; CHECK-NEXT: [[IV:%.*]] = phi i64 [ 1, [[PREHEADER]] ], [ [[IV_NEXT:%.*]], [[FOR]] ] 280; CHECK-NEXT: [[ADD_1:%.*]] = add i32 [[PRE_PHI]], [[X:%.*]] 281; CHECK-NEXT: store i32 [[ADD_1]], i32* [[PTR:%.*]], align 4 282; CHECK-NEXT: [[IDX_1:%.*]] = getelementptr inbounds [257 x i32], [257 x i32]* @p, i64 0, i64 [[IV]] 283; CHECK-NEXT: [[PRE_NEXT]] = load i32, i32* [[IDX_1]], align 4 284; CHECK-NEXT: [[ADD_2:%.*]] = add i32 [[ADD_1]], [[PRE_NEXT]] 285; CHECK-NEXT: [[IDX_2:%.*]] = getelementptr inbounds [257 x i32], [257 x i32]* @q, i64 0, i64 [[IV]] 286; CHECK-NEXT: store i32 [[ADD_2]], i32* [[IDX_2]], align 4 287; CHECK-NEXT: [[IV_NEXT]] = add nuw nsw i64 [[IV]], 1 288; CHECK-NEXT: [[EXITCOND:%.*]] = icmp eq i64 [[IV_NEXT]], 2000 289; CHECK-NEXT: br i1 [[EXITCOND]], label [[EXIT:%.*]], label [[FOR]] 290; CHECK: exit: 291; CHECK-NEXT: ret void 292; 293 294 295 296entry: 297 br label %preheader 298 299preheader: 300 %idx.phi.trans = getelementptr inbounds [257 x i32], [257 x i32]* @p, i64 0, i64 1 301 %.pre = load i32, i32* %idx.phi.trans, align 4 302 br label %for 303 304for: 305 %pre.phi = phi i32 [ %.pre, %preheader ], [ %pre.next, %for ] 306 %iv = phi i64 [ 1, %preheader ], [ %iv.next, %for ] 307 %add.1 = add i32 %pre.phi, %x 308 store i32 %add.1, i32* %ptr 309 %idx.1 = getelementptr inbounds [257 x i32], [257 x i32]* @p, i64 0, i64 %iv 310 %pre.next = load i32, i32* %idx.1, align 4 311 %add.2 = add i32 %add.1, %pre.next 312 %idx.2 = getelementptr inbounds [257 x i32], [257 x i32]* @q, i64 0, i64 %iv 313 store i32 %add.2, i32* %idx.2, align 4 314 %iv.next = add nuw nsw i64 %iv, 1 315 %exitcond = icmp eq i64 %iv.next, 2000 316 br i1 %exitcond, label %exit, label %for 317 318exit: 319 ret void 320} 321 322; Some kinds of reductions are not detected by IVDescriptors. If we have a 323; cycle, we cannot sink it. 324define void @cannot_sink_reduction(i32 %x, i32* %ptr, i64 %tc) { 325; CHECK-LABEL: @cannot_sink_reduction( 326; CHECK-NEXT: entry: 327; CHECK-NEXT: br label [[PREHEADER:%.*]] 328; CHECK: preheader: 329; CHECK-NEXT: [[IDX_PHI_TRANS:%.*]] = getelementptr inbounds [257 x i32], [257 x i32]* @p, i64 0, i64 1 330; CHECK-NEXT: [[DOTPRE:%.*]] = load i32, i32* [[IDX_PHI_TRANS]], align 4 331; CHECK-NEXT: br label [[FOR:%.*]] 332; CHECK: for: 333; CHECK-NEXT: [[PRE_PHI:%.*]] = phi i32 [ [[DOTPRE]], [[PREHEADER]] ], [ [[D:%.*]], [[FOR]] ] 334; CHECK-NEXT: [[IV:%.*]] = phi i64 [ 1, [[PREHEADER]] ], [ [[IV_NEXT:%.*]], [[FOR]] ] 335; CHECK-NEXT: [[D]] = sdiv i32 [[PRE_PHI]], [[X:%.*]] 336; CHECK-NEXT: [[IDX_1:%.*]] = getelementptr inbounds [257 x i32], [257 x i32]* @p, i64 0, i64 [[IV]] 337; CHECK-NEXT: [[PRE_NEXT:%.*]] = load i32, i32* [[IDX_1]], align 4 338; CHECK-NEXT: [[ADD_2:%.*]] = add i32 [[X]], [[PRE_NEXT]] 339; CHECK-NEXT: [[IDX_2:%.*]] = getelementptr inbounds [257 x i32], [257 x i32]* @q, i64 0, i64 [[IV]] 340; CHECK-NEXT: store i32 [[ADD_2]], i32* [[IDX_2]], align 4 341; CHECK-NEXT: [[IV_NEXT]] = add nuw nsw i64 [[IV]], 1 342; CHECK-NEXT: [[EXITCOND:%.*]] = icmp eq i64 [[IV_NEXT]], 2000 343; CHECK-NEXT: br i1 [[EXITCOND]], label [[EXIT:%.*]], label [[FOR]] 344; CHECK: exit: 345; CHECK-NEXT: ret void 346; 347 348 349 350; CHECK-NET: ret void 351entry: 352 br label %preheader 353 354preheader: 355 %idx.phi.trans = getelementptr inbounds [257 x i32], [257 x i32]* @p, i64 0, i64 1 356 %.pre = load i32, i32* %idx.phi.trans, align 4 357 br label %for 358 359for: 360 %pre.phi = phi i32 [ %.pre, %preheader ], [ %d, %for ] 361 %iv = phi i64 [ 1, %preheader ], [ %iv.next, %for ] 362 %d = sdiv i32 %pre.phi, %x 363 %idx.1 = getelementptr inbounds [257 x i32], [257 x i32]* @p, i64 0, i64 %iv 364 %pre.next = load i32, i32* %idx.1, align 4 365 %add.2 = add i32 %x, %pre.next 366 %idx.2 = getelementptr inbounds [257 x i32], [257 x i32]* @q, i64 0, i64 %iv 367 store i32 %add.2, i32* %idx.2, align 4 368 %iv.next = add nuw nsw i64 %iv, 1 369 %exitcond = icmp eq i64 %iv.next, 2000 370 br i1 %exitcond, label %exit, label %for 371 372exit: 373 ret void 374} 375 376; Sink %tmp38 after %tmp60, then it enable the loop vectorization. 377define void @instruction_with_2_FOR_operands() { 378; CHECK-LABEL: @instruction_with_2_FOR_operands( 379; CHECK-NEXT: bb: 380; CHECK-NEXT: [[SMAX:%.*]] = call i64 @llvm.smax.i64(i64 undef, i64 0) 381; CHECK-NEXT: [[TMP0:%.*]] = add nuw i64 [[SMAX]], 1 382; CHECK-NEXT: [[MIN_ITERS_CHECK:%.*]] = icmp ult i64 [[TMP0]], 4 383; CHECK-NEXT: br i1 [[MIN_ITERS_CHECK]], label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]] 384; CHECK: vector.ph: 385; CHECK-NEXT: [[N_MOD_VF:%.*]] = urem i64 [[TMP0]], 4 386; CHECK-NEXT: [[N_VEC:%.*]] = sub i64 [[TMP0]], [[N_MOD_VF]] 387; CHECK-NEXT: br label [[VECTOR_BODY:%.*]] 388; CHECK: vector.body: 389; CHECK-NEXT: [[INDEX:%.*]] = phi i64 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ] 390; CHECK-NEXT: [[VECTOR_RECUR:%.*]] = phi <4 x float> [ <float poison, float poison, float poison, float undef>, [[VECTOR_PH]] ], [ [[BROADCAST_SPLAT3:%.*]], [[VECTOR_BODY]] ] 391; CHECK-NEXT: [[VECTOR_RECUR1:%.*]] = phi <4 x float> [ <float poison, float poison, float poison, float undef>, [[VECTOR_PH]] ], [ [[BROADCAST_SPLAT:%.*]], [[VECTOR_BODY]] ] 392; CHECK-NEXT: [[TMP1:%.*]] = add i64 [[INDEX]], 0 393; CHECK-NEXT: [[TMP2:%.*]] = load float, float* undef, align 4 394; CHECK-NEXT: [[BROADCAST_SPLATINSERT:%.*]] = insertelement <4 x float> poison, float [[TMP2]], i32 0 395; CHECK-NEXT: [[BROADCAST_SPLAT]] = shufflevector <4 x float> [[BROADCAST_SPLATINSERT]], <4 x float> poison, <4 x i32> zeroinitializer 396; CHECK-NEXT: [[TMP3:%.*]] = shufflevector <4 x float> [[VECTOR_RECUR1]], <4 x float> [[BROADCAST_SPLAT]], <4 x i32> <i32 3, i32 4, i32 5, i32 6> 397; CHECK-NEXT: [[TMP4:%.*]] = load float, float* undef, align 4 398; CHECK-NEXT: [[BROADCAST_SPLATINSERT2:%.*]] = insertelement <4 x float> poison, float [[TMP4]], i32 0 399; CHECK-NEXT: [[BROADCAST_SPLAT3]] = shufflevector <4 x float> [[BROADCAST_SPLATINSERT2]], <4 x float> poison, <4 x i32> zeroinitializer 400; CHECK-NEXT: [[TMP5:%.*]] = shufflevector <4 x float> [[VECTOR_RECUR]], <4 x float> [[BROADCAST_SPLAT3]], <4 x i32> <i32 3, i32 4, i32 5, i32 6> 401; CHECK-NEXT: [[TMP6:%.*]] = fmul fast <4 x float> [[TMP5]], [[TMP3]] 402; CHECK-NEXT: [[INDEX_NEXT]] = add nuw i64 [[INDEX]], 4 403; CHECK-NEXT: [[TMP7:%.*]] = icmp eq i64 [[INDEX_NEXT]], [[N_VEC]] 404; CHECK-NEXT: br i1 [[TMP7]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP8:![0-9]+]] 405; CHECK: middle.block: 406; CHECK-NEXT: [[CMP_N:%.*]] = icmp eq i64 [[TMP0]], [[N_VEC]] 407; CHECK-NEXT: [[VECTOR_RECUR_EXTRACT:%.*]] = extractelement <4 x float> [[BROADCAST_SPLAT3]], i32 3 408; CHECK-NEXT: [[VECTOR_RECUR_EXTRACT_FOR_PHI:%.*]] = extractelement <4 x float> [[BROADCAST_SPLAT3]], i32 2 409; CHECK-NEXT: [[VECTOR_RECUR_EXTRACT4:%.*]] = extractelement <4 x float> [[BROADCAST_SPLAT]], i32 3 410; CHECK-NEXT: [[VECTOR_RECUR_EXTRACT_FOR_PHI5:%.*]] = extractelement <4 x float> [[BROADCAST_SPLAT]], i32 2 411; CHECK-NEXT: br i1 [[CMP_N]], label [[BB74:%.*]], label [[SCALAR_PH]] 412; CHECK: scalar.ph: 413; CHECK-NEXT: [[SCALAR_RECUR_INIT6:%.*]] = phi float [ undef, [[BB:%.*]] ], [ [[VECTOR_RECUR_EXTRACT4]], [[MIDDLE_BLOCK]] ] 414; CHECK-NEXT: [[SCALAR_RECUR_INIT:%.*]] = phi float [ undef, [[BB]] ], [ [[VECTOR_RECUR_EXTRACT]], [[MIDDLE_BLOCK]] ] 415; CHECK-NEXT: [[BC_RESUME_VAL:%.*]] = phi i64 [ [[N_VEC]], [[MIDDLE_BLOCK]] ], [ 0, [[BB]] ] 416; CHECK-NEXT: br label [[BB13:%.*]] 417; CHECK: bb13: 418; CHECK-NEXT: [[SCALAR_RECUR:%.*]] = phi float [ [[TMP60:%.*]], [[BB13]] ], [ [[SCALAR_RECUR_INIT]], [[SCALAR_PH]] ] 419; CHECK-NEXT: [[SCALAR_RECUR7:%.*]] = phi float [ [[TMP49:%.*]], [[BB13]] ], [ [[SCALAR_RECUR_INIT6]], [[SCALAR_PH]] ] 420; CHECK-NEXT: [[INDVARS_IV:%.*]] = phi i64 [ [[INDVARS_IV_NEXT:%.*]], [[BB13]] ], [ [[BC_RESUME_VAL]], [[SCALAR_PH]] ] 421; CHECK-NEXT: [[TMP38:%.*]] = fmul fast float [[SCALAR_RECUR]], [[SCALAR_RECUR7]] 422; CHECK-NEXT: [[INDVARS_IV_NEXT]] = add nuw nsw i64 [[INDVARS_IV]], 1 423; CHECK-NEXT: [[TMP49]] = load float, float* undef, align 4 424; CHECK-NEXT: [[TMP60]] = load float, float* undef, align 4 425; CHECK-NEXT: [[TMP12:%.*]] = icmp slt i64 [[INDVARS_IV]], undef 426; CHECK-NEXT: br i1 [[TMP12]], label [[BB13]], label [[BB74]], !llvm.loop [[LOOP9:![0-9]+]] 427; CHECK: bb74: 428; CHECK-NEXT: ret void 429; 430 431 432bb: 433 br label %bb13 434 435bb13: ; preds = %bb13, %bb 436 %tmp37 = phi float [ %tmp60, %bb13 ], [ undef, %bb ] 437 %tmp27 = phi float [ %tmp49, %bb13 ], [ undef, %bb ] 438 %indvars.iv = phi i64 [ %indvars.iv.next, %bb13 ], [ 0, %bb ] 439 %tmp38 = fmul fast float %tmp37, %tmp27 440 %indvars.iv.next = add nuw nsw i64 %indvars.iv, 1 441 %tmp49 = load float, float* undef, align 4 442 %tmp60 = load float, float* undef, align 4 443 %tmp12 = icmp slt i64 %indvars.iv, undef 444 br i1 %tmp12, label %bb13, label %bb74 445 446bb74: ; preds = %bb13 447 ret void 448} 449 450define void @instruction_with_2_FOR_operands_and_multiple_other_uses(float* noalias %dst.1, float* noalias %dst.2, float* noalias %dst.3, float* noalias %for.ptr.1, float* noalias %for.ptr.2) { 451; CHECK-LABEL: @instruction_with_2_FOR_operands_and_multiple_other_uses( 452; CHECK-NEXT: bb: 453; CHECK-NEXT: br label [[LOOP:%.*]] 454; CHECK: loop: 455; CHECK-NEXT: [[FOR_1:%.*]] = phi float [ 0.000000e+00, [[BB:%.*]] ], [ [[FOR_1_NEXT:%.*]], [[LOOP]] ] 456; CHECK-NEXT: [[FOR_2:%.*]] = phi float [ 0.000000e+00, [[BB]] ], [ [[FOR_2_NEXT:%.*]], [[LOOP]] ] 457; CHECK-NEXT: [[IV:%.*]] = phi i64 [ 0, [[BB]] ], [ [[IV_NEXT:%.*]], [[LOOP]] ] 458; CHECK-NEXT: [[FOR_1_USE_1:%.*]] = fmul fast float [[FOR_1]], 2.000000e+00 459; CHECK-NEXT: [[USED_BY_BOTH:%.*]] = fmul fast float [[FOR_1]], [[FOR_2]] 460; CHECK-NEXT: [[FOR_2_NEXT]] = load float, float* [[FOR_PTR_2:%.*]], align 4 461; CHECK-NEXT: [[FOR_1_USE_3:%.*]] = fadd fast float [[FOR_1]], 1.000000e+00 462; CHECK-NEXT: [[IV_NEXT]] = add nuw nsw i64 [[IV]], 1 463; CHECK-NEXT: [[FOR_1_NEXT]] = load float, float* [[FOR_PTR_1:%.*]], align 4 464; CHECK-NEXT: [[GEP_DST_1:%.*]] = getelementptr inbounds float, float* [[DST_1:%.*]], i64 [[IV]] 465; CHECK-NEXT: store float [[USED_BY_BOTH]], float* [[GEP_DST_1]], align 4 466; CHECK-NEXT: [[GEP_DST_2:%.*]] = getelementptr inbounds float, float* [[DST_2:%.*]], i64 [[IV]] 467; CHECK-NEXT: store float [[FOR_1_USE_1]], float* [[GEP_DST_2]], align 4 468; CHECK-NEXT: [[GEP_DST_3:%.*]] = getelementptr inbounds float, float* [[DST_3:%.*]], i64 [[IV]] 469; CHECK-NEXT: store float [[FOR_1_USE_3]], float* [[GEP_DST_3]], align 4 470; CHECK-NEXT: [[EC:%.*]] = icmp slt i64 [[IV]], 1000 471; CHECK-NEXT: br i1 [[EC]], label [[LOOP]], label [[EXIT:%.*]] 472; CHECK: exit: 473; CHECK-NEXT: ret void 474; 475bb: 476 br label %loop 477 478loop: 479 %for.1 = phi float [ 0.0, %bb ], [ %for.1.next, %loop] 480 %for.2 = phi float [ 0.0, %bb ], [ %for.2.next, %loop] 481 %iv = phi i64 [ 0, %bb ], [ %iv.next, %loop ] 482 %for.1.use.1 = fmul fast float %for.1, 2.0 483 %used.by.both = fmul fast float %for.1, %for.2 484 %for.2.next = load float, float* %for.ptr.2, align 4 485 %for.1.use.3 = fadd fast float %for.1, 1.0 486 %iv.next = add nuw nsw i64 %iv, 1 487 %for.1.next = load float, float* %for.ptr.1, align 4 488 %gep.dst.1 = getelementptr inbounds float, float* %dst.1, i64 %iv 489 store float %used.by.both, float* %gep.dst.1 490 %gep.dst.2 = getelementptr inbounds float, float* %dst.2, i64 %iv 491 store float %for.1.use.1, float* %gep.dst.2 492 %gep.dst.3 = getelementptr inbounds float, float* %dst.3, i64 %iv 493 store float %for.1.use.3, float* %gep.dst.3 494 %ec = icmp slt i64 %iv, 1000 495 br i1 %ec, label %loop, label %exit 496 497exit: 498 ret void 499} 500 501; The (first) reason `%first_time.1` cannot be sunk is because it appears outside 502; the header and is not dominated by Previous. The fact that it feeds Previous 503; is a second sinking-preventing reason. 504define void @cannot_sink_phi(i32* %ptr) { 505; CHECK-LABEL: @cannot_sink_phi( 506; CHECK-NEXT: entry: 507; CHECK-NEXT: br label [[LOOP_HEADER:%.*]] 508; CHECK: loop.header: 509; CHECK-NEXT: [[IV:%.*]] = phi i64 [ 1, [[ENTRY:%.*]] ], [ [[IV_NEXT:%.*]], [[LOOP_LATCH:%.*]] ] 510; CHECK-NEXT: [[FOR:%.*]] = phi i32 [ 0, [[ENTRY]] ], [ [[FOR_NEXT:%.*]], [[LOOP_LATCH]] ] 511; CHECK-NEXT: [[C_1:%.*]] = icmp ult i64 [[IV]], 500 512; CHECK-NEXT: br i1 [[C_1]], label [[IF_TRUEBB:%.*]], label [[IF_FALSEBB:%.*]] 513; CHECK: if.truebb: 514; CHECK-NEXT: br label [[LOOP_LATCH]] 515; CHECK: if.falsebb: 516; CHECK-NEXT: br label [[LOOP_LATCH]] 517; CHECK: loop.latch: 518; CHECK-NEXT: [[FIRST_TIME_1:%.*]] = phi i32 [ 20, [[IF_TRUEBB]] ], [ [[FOR]], [[IF_FALSEBB]] ] 519; CHECK-NEXT: [[C_2:%.*]] = icmp ult i64 [[IV]], 800 520; CHECK-NEXT: [[FOR_NEXT]] = select i1 [[C_2]], i32 30, i32 [[FIRST_TIME_1]] 521; CHECK-NEXT: [[PTR_IDX:%.*]] = getelementptr i32, i32* [[PTR:%.*]], i64 [[IV]] 522; CHECK-NEXT: store i32 [[FOR_NEXT]], i32* [[PTR_IDX]], align 4 523; CHECK-NEXT: [[IV_NEXT]] = add nuw nsw i64 [[IV]], 1 524; CHECK-NEXT: [[EXITCOND_NOT:%.*]] = icmp eq i64 [[IV_NEXT]], 1000 525; CHECK-NEXT: br i1 [[EXITCOND_NOT]], label [[EXIT:%.*]], label [[LOOP_HEADER]] 526; CHECK: exit: 527; CHECK-NEXT: ret void 528; 529entry: 530 br label %loop.header 531 532loop.header: 533 %iv = phi i64 [ 1, %entry ], [ %iv.next, %loop.latch ] 534 %for = phi i32 [ 0, %entry ], [ %for.next, %loop.latch ] 535 %c.1 = icmp ult i64 %iv, 500 536 br i1 %c.1, label %if.truebb, label %if.falsebb 537 538if.truebb: 539 br label %loop.latch 540 541if.falsebb: 542 br label %loop.latch 543 544loop.latch: 545 %first_time.1 = phi i32 [ 20, %if.truebb ], [ %for, %if.falsebb ] 546 %c.2 = icmp ult i64 %iv, 800 547 %for.next = select i1 %c.2, i32 30, i32 %first_time.1 548 %ptr.idx = getelementptr i32, i32* %ptr, i64 %iv 549 store i32 %for.next, i32* %ptr.idx 550 %iv.next = add nuw nsw i64 %iv, 1 551 %exitcond.not = icmp eq i64 %iv.next, 1000 552 br i1 %exitcond.not, label %exit, label %loop.header 553 554exit: 555 ret void 556} 557 558; A recurrence in a multiple exit loop. 559define i16 @multiple_exit(i16* %p, i32 %n) { 560; CHECK-LABEL: @multiple_exit( 561; CHECK-NEXT: entry: 562; CHECK-NEXT: [[SMAX:%.*]] = call i32 @llvm.smax.i32(i32 [[N:%.*]], i32 0) 563; CHECK-NEXT: [[UMIN:%.*]] = call i32 @llvm.umin.i32(i32 [[SMAX]], i32 2096) 564; CHECK-NEXT: [[TMP0:%.*]] = add nuw nsw i32 [[UMIN]], 1 565; CHECK-NEXT: [[MIN_ITERS_CHECK:%.*]] = icmp ule i32 [[TMP0]], 4 566; CHECK-NEXT: br i1 [[MIN_ITERS_CHECK]], label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]] 567; CHECK: vector.ph: 568; CHECK-NEXT: [[N_MOD_VF:%.*]] = urem i32 [[TMP0]], 4 569; CHECK-NEXT: [[TMP1:%.*]] = icmp eq i32 [[N_MOD_VF]], 0 570; CHECK-NEXT: [[TMP2:%.*]] = select i1 [[TMP1]], i32 4, i32 [[N_MOD_VF]] 571; CHECK-NEXT: [[N_VEC:%.*]] = sub i32 [[TMP0]], [[TMP2]] 572; CHECK-NEXT: br label [[VECTOR_BODY:%.*]] 573; CHECK: vector.body: 574; CHECK-NEXT: [[INDEX:%.*]] = phi i32 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ] 575; CHECK-NEXT: [[VEC_IND:%.*]] = phi <4 x i32> [ <i32 0, i32 1, i32 2, i32 3>, [[VECTOR_PH]] ], [ [[VEC_IND_NEXT:%.*]], [[VECTOR_BODY]] ] 576; CHECK-NEXT: [[VECTOR_RECUR:%.*]] = phi <4 x i16> [ <i16 poison, i16 poison, i16 poison, i16 0>, [[VECTOR_PH]] ], [ [[WIDE_LOAD:%.*]], [[VECTOR_BODY]] ] 577; CHECK-NEXT: [[TMP3:%.*]] = add i32 [[INDEX]], 0 578; CHECK-NEXT: [[TMP4:%.*]] = sext i32 [[TMP3]] to i64 579; CHECK-NEXT: [[TMP5:%.*]] = getelementptr inbounds i16, i16* [[P:%.*]], i64 [[TMP4]] 580; CHECK-NEXT: [[TMP6:%.*]] = getelementptr inbounds i16, i16* [[TMP5]], i32 0 581; CHECK-NEXT: [[TMP7:%.*]] = bitcast i16* [[TMP6]] to <4 x i16>* 582; CHECK-NEXT: [[WIDE_LOAD]] = load <4 x i16>, <4 x i16>* [[TMP7]], align 2 583; CHECK-NEXT: [[TMP8:%.*]] = shufflevector <4 x i16> [[VECTOR_RECUR]], <4 x i16> [[WIDE_LOAD]], <4 x i32> <i32 3, i32 4, i32 5, i32 6> 584; CHECK-NEXT: [[TMP9:%.*]] = bitcast i16* [[TMP6]] to <4 x i16>* 585; CHECK-NEXT: store <4 x i16> [[TMP8]], <4 x i16>* [[TMP9]], align 4 586; CHECK-NEXT: [[INDEX_NEXT]] = add nuw i32 [[INDEX]], 4 587; CHECK-NEXT: [[VEC_IND_NEXT]] = add <4 x i32> [[VEC_IND]], <i32 4, i32 4, i32 4, i32 4> 588; CHECK-NEXT: [[TMP10:%.*]] = icmp eq i32 [[INDEX_NEXT]], [[N_VEC]] 589; CHECK-NEXT: br i1 [[TMP10]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP8:![0-9]+]] 590; CHECK: middle.block: 591; CHECK-NEXT: [[VECTOR_RECUR_EXTRACT:%.*]] = extractelement <4 x i16> [[WIDE_LOAD]], i32 3 592; CHECK-NEXT: [[VECTOR_RECUR_EXTRACT_FOR_PHI:%.*]] = extractelement <4 x i16> [[WIDE_LOAD]], i32 2 593; CHECK-NEXT: br label [[SCALAR_PH]] 594; CHECK: scalar.ph: 595; CHECK-NEXT: [[SCALAR_RECUR_INIT:%.*]] = phi i16 [ 0, [[ENTRY:%.*]] ], [ [[VECTOR_RECUR_EXTRACT]], [[MIDDLE_BLOCK]] ] 596; CHECK-NEXT: [[BC_RESUME_VAL:%.*]] = phi i32 [ [[N_VEC]], [[MIDDLE_BLOCK]] ], [ 0, [[ENTRY]] ] 597; CHECK-NEXT: br label [[FOR_COND:%.*]] 598; CHECK: for.cond: 599; CHECK-NEXT: [[I:%.*]] = phi i32 [ [[BC_RESUME_VAL]], [[SCALAR_PH]] ], [ [[INC:%.*]], [[FOR_BODY:%.*]] ] 600; CHECK-NEXT: [[SCALAR_RECUR:%.*]] = phi i16 [ [[SCALAR_RECUR_INIT]], [[SCALAR_PH]] ], [ [[REC_NEXT:%.*]], [[FOR_BODY]] ] 601; CHECK-NEXT: [[IPROM:%.*]] = sext i32 [[I]] to i64 602; CHECK-NEXT: [[B:%.*]] = getelementptr inbounds i16, i16* [[P]], i64 [[IPROM]] 603; CHECK-NEXT: [[REC_NEXT]] = load i16, i16* [[B]], align 2 604; CHECK-NEXT: [[CMP:%.*]] = icmp slt i32 [[I]], [[N]] 605; CHECK-NEXT: br i1 [[CMP]], label [[FOR_BODY]], label [[IF_END:%.*]] 606; CHECK: for.body: 607; CHECK-NEXT: store i16 [[SCALAR_RECUR]], i16* [[B]], align 4 608; CHECK-NEXT: [[INC]] = add nsw i32 [[I]], 1 609; CHECK-NEXT: [[CMP2:%.*]] = icmp slt i32 [[I]], 2096 610; CHECK-NEXT: br i1 [[CMP2]], label [[FOR_COND]], label [[IF_END]], !llvm.loop [[LOOP9:![0-9]+]] 611; CHECK: if.end: 612; CHECK-NEXT: [[REC_LCSSA:%.*]] = phi i16 [ [[SCALAR_RECUR]], [[FOR_BODY]] ], [ [[SCALAR_RECUR]], [[FOR_COND]] ] 613; CHECK-NEXT: ret i16 [[REC_LCSSA]] 614; 615entry: 616 br label %for.cond 617 618for.cond: 619 %i = phi i32 [ 0, %entry ], [ %inc, %for.body ] 620 %rec = phi i16 [0, %entry], [ %rec.next, %for.body ] 621 %iprom = sext i32 %i to i64 622 %b = getelementptr inbounds i16, i16* %p, i64 %iprom 623 %rec.next = load i16, i16* %b 624 %cmp = icmp slt i32 %i, %n 625 br i1 %cmp, label %for.body, label %if.end 626 627for.body: 628 store i16 %rec , i16* %b, align 4 629 %inc = add nsw i32 %i, 1 630 %cmp2 = icmp slt i32 %i, 2096 631 br i1 %cmp2, label %for.cond, label %if.end 632 633if.end: 634 ret i16 %rec 635} 636 637 638; A multiple exit case where one of the exiting edges involves a value 639; from the recurrence and one does not. 640define i16 @multiple_exit2(i16* %p, i32 %n) { 641; CHECK-LABEL: @multiple_exit2( 642; CHECK-NEXT: entry: 643; CHECK-NEXT: [[SMAX:%.*]] = call i32 @llvm.smax.i32(i32 [[N:%.*]], i32 0) 644; CHECK-NEXT: [[UMIN:%.*]] = call i32 @llvm.umin.i32(i32 [[SMAX]], i32 2096) 645; CHECK-NEXT: [[TMP0:%.*]] = add nuw nsw i32 [[UMIN]], 1 646; CHECK-NEXT: [[MIN_ITERS_CHECK:%.*]] = icmp ule i32 [[TMP0]], 4 647; CHECK-NEXT: br i1 [[MIN_ITERS_CHECK]], label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]] 648; CHECK: vector.ph: 649; CHECK-NEXT: [[N_MOD_VF:%.*]] = urem i32 [[TMP0]], 4 650; CHECK-NEXT: [[TMP1:%.*]] = icmp eq i32 [[N_MOD_VF]], 0 651; CHECK-NEXT: [[TMP2:%.*]] = select i1 [[TMP1]], i32 4, i32 [[N_MOD_VF]] 652; CHECK-NEXT: [[N_VEC:%.*]] = sub i32 [[TMP0]], [[TMP2]] 653; CHECK-NEXT: br label [[VECTOR_BODY:%.*]] 654; CHECK: vector.body: 655; CHECK-NEXT: [[INDEX:%.*]] = phi i32 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ] 656; CHECK-NEXT: [[VEC_IND:%.*]] = phi <4 x i32> [ <i32 0, i32 1, i32 2, i32 3>, [[VECTOR_PH]] ], [ [[VEC_IND_NEXT:%.*]], [[VECTOR_BODY]] ] 657; CHECK-NEXT: [[VECTOR_RECUR:%.*]] = phi <4 x i16> [ <i16 poison, i16 poison, i16 poison, i16 0>, [[VECTOR_PH]] ], [ [[WIDE_LOAD:%.*]], [[VECTOR_BODY]] ] 658; CHECK-NEXT: [[TMP3:%.*]] = add i32 [[INDEX]], 0 659; CHECK-NEXT: [[TMP4:%.*]] = sext i32 [[TMP3]] to i64 660; CHECK-NEXT: [[TMP5:%.*]] = getelementptr inbounds i16, i16* [[P:%.*]], i64 [[TMP4]] 661; CHECK-NEXT: [[TMP6:%.*]] = getelementptr inbounds i16, i16* [[TMP5]], i32 0 662; CHECK-NEXT: [[TMP7:%.*]] = bitcast i16* [[TMP6]] to <4 x i16>* 663; CHECK-NEXT: [[WIDE_LOAD]] = load <4 x i16>, <4 x i16>* [[TMP7]], align 2 664; CHECK-NEXT: [[TMP8:%.*]] = shufflevector <4 x i16> [[VECTOR_RECUR]], <4 x i16> [[WIDE_LOAD]], <4 x i32> <i32 3, i32 4, i32 5, i32 6> 665; CHECK-NEXT: [[TMP9:%.*]] = bitcast i16* [[TMP6]] to <4 x i16>* 666; CHECK-NEXT: store <4 x i16> [[TMP8]], <4 x i16>* [[TMP9]], align 4 667; CHECK-NEXT: [[INDEX_NEXT]] = add nuw i32 [[INDEX]], 4 668; CHECK-NEXT: [[VEC_IND_NEXT]] = add <4 x i32> [[VEC_IND]], <i32 4, i32 4, i32 4, i32 4> 669; CHECK-NEXT: [[TMP10:%.*]] = icmp eq i32 [[INDEX_NEXT]], [[N_VEC]] 670; CHECK-NEXT: br i1 [[TMP10]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP10:![0-9]+]] 671; CHECK: middle.block: 672; CHECK-NEXT: [[VECTOR_RECUR_EXTRACT:%.*]] = extractelement <4 x i16> [[WIDE_LOAD]], i32 3 673; CHECK-NEXT: [[VECTOR_RECUR_EXTRACT_FOR_PHI:%.*]] = extractelement <4 x i16> [[WIDE_LOAD]], i32 2 674; CHECK-NEXT: br label [[SCALAR_PH]] 675; CHECK: scalar.ph: 676; CHECK-NEXT: [[SCALAR_RECUR_INIT:%.*]] = phi i16 [ 0, [[ENTRY:%.*]] ], [ [[VECTOR_RECUR_EXTRACT]], [[MIDDLE_BLOCK]] ] 677; CHECK-NEXT: [[BC_RESUME_VAL:%.*]] = phi i32 [ [[N_VEC]], [[MIDDLE_BLOCK]] ], [ 0, [[ENTRY]] ] 678; CHECK-NEXT: br label [[FOR_COND:%.*]] 679; CHECK: for.cond: 680; CHECK-NEXT: [[I:%.*]] = phi i32 [ [[BC_RESUME_VAL]], [[SCALAR_PH]] ], [ [[INC:%.*]], [[FOR_BODY:%.*]] ] 681; CHECK-NEXT: [[SCALAR_RECUR:%.*]] = phi i16 [ [[SCALAR_RECUR_INIT]], [[SCALAR_PH]] ], [ [[REC_NEXT:%.*]], [[FOR_BODY]] ] 682; CHECK-NEXT: [[IPROM:%.*]] = sext i32 [[I]] to i64 683; CHECK-NEXT: [[B:%.*]] = getelementptr inbounds i16, i16* [[P]], i64 [[IPROM]] 684; CHECK-NEXT: [[REC_NEXT]] = load i16, i16* [[B]], align 2 685; CHECK-NEXT: [[CMP:%.*]] = icmp slt i32 [[I]], [[N]] 686; CHECK-NEXT: br i1 [[CMP]], label [[FOR_BODY]], label [[IF_END:%.*]] 687; CHECK: for.body: 688; CHECK-NEXT: store i16 [[SCALAR_RECUR]], i16* [[B]], align 4 689; CHECK-NEXT: [[INC]] = add nsw i32 [[I]], 1 690; CHECK-NEXT: [[CMP2:%.*]] = icmp slt i32 [[I]], 2096 691; CHECK-NEXT: br i1 [[CMP2]], label [[FOR_COND]], label [[IF_END]], !llvm.loop [[LOOP11:![0-9]+]] 692; CHECK: if.end: 693; CHECK-NEXT: [[REC_LCSSA:%.*]] = phi i16 [ [[SCALAR_RECUR]], [[FOR_COND]] ], [ 10, [[FOR_BODY]] ] 694; CHECK-NEXT: ret i16 [[REC_LCSSA]] 695; 696entry: 697 br label %for.cond 698 699for.cond: 700 %i = phi i32 [ 0, %entry ], [ %inc, %for.body ] 701 %rec = phi i16 [0, %entry], [ %rec.next, %for.body ] 702 %iprom = sext i32 %i to i64 703 %b = getelementptr inbounds i16, i16* %p, i64 %iprom 704 %rec.next = load i16, i16* %b 705 %cmp = icmp slt i32 %i, %n 706 br i1 %cmp, label %for.body, label %if.end 707 708for.body: 709 store i16 %rec , i16* %b, align 4 710 %inc = add nsw i32 %i, 1 711 %cmp2 = icmp slt i32 %i, 2096 712 br i1 %cmp2, label %for.cond, label %if.end 713 714if.end: 715 %rec.lcssa = phi i16 [ %rec, %for.cond ], [ 10, %for.body ] 716 ret i16 %rec.lcssa 717} 718 719; A test where the instructions to sink may not be visited in dominance order. 720define void @sink_dominance(i32* %ptr, i32 %N) { 721; CHECK-LABEL: @sink_dominance( 722; CHECK-NEXT: entry: 723; CHECK-NEXT: [[UMAX1:%.*]] = call i32 @llvm.umax.i32(i32 [[N:%.*]], i32 1) 724; CHECK-NEXT: [[MIN_ITERS_CHECK:%.*]] = icmp ult i32 [[UMAX1]], 4 725; CHECK-NEXT: br i1 [[MIN_ITERS_CHECK]], label [[SCALAR_PH:%.*]], label [[VECTOR_SCEVCHECK:%.*]] 726; CHECK: vector.scevcheck: 727; CHECK-NEXT: [[UMAX:%.*]] = call i32 @llvm.umax.i32(i32 [[N]], i32 1) 728; CHECK-NEXT: [[TMP0:%.*]] = add i32 [[UMAX]], -1 729; CHECK-NEXT: [[TMP4:%.*]] = icmp slt i32 [[TMP0]], 0 730; CHECK-NEXT: br i1 [[TMP4]], label [[SCALAR_PH]], label [[VECTOR_PH:%.*]] 731; CHECK: vector.ph: 732; CHECK-NEXT: [[N_MOD_VF:%.*]] = urem i32 [[UMAX1]], 4 733; CHECK-NEXT: [[N_VEC:%.*]] = sub i32 [[UMAX1]], [[N_MOD_VF]] 734; CHECK-NEXT: br label [[VECTOR_BODY:%.*]] 735; CHECK: vector.body: 736; CHECK-NEXT: [[INDEX:%.*]] = phi i32 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ] 737; CHECK-NEXT: [[VECTOR_RECUR:%.*]] = phi <4 x i64> [ <i64 poison, i64 poison, i64 poison, i64 0>, [[VECTOR_PH]] ], [ [[TMP11:%.*]], [[VECTOR_BODY]] ] 738; CHECK-NEXT: [[TMP7:%.*]] = add i32 [[INDEX]], 0 739; CHECK-NEXT: [[TMP8:%.*]] = getelementptr inbounds i32, i32* [[PTR:%.*]], i32 [[TMP7]] 740; CHECK-NEXT: [[TMP9:%.*]] = getelementptr inbounds i32, i32* [[TMP8]], i32 0 741; CHECK-NEXT: [[TMP10:%.*]] = bitcast i32* [[TMP9]] to <4 x i32>* 742; CHECK-NEXT: [[WIDE_LOAD:%.*]] = load <4 x i32>, <4 x i32>* [[TMP10]], align 4 743; CHECK-NEXT: [[TMP11]] = zext <4 x i32> [[WIDE_LOAD]] to <4 x i64> 744; CHECK-NEXT: [[TMP12:%.*]] = shufflevector <4 x i64> [[VECTOR_RECUR]], <4 x i64> [[TMP11]], <4 x i32> <i32 3, i32 4, i32 5, i32 6> 745; CHECK-NEXT: [[TMP13:%.*]] = trunc <4 x i64> [[TMP12]] to <4 x i32> 746; CHECK-NEXT: [[TMP14:%.*]] = icmp slt <4 x i32> [[TMP13]], <i32 213, i32 213, i32 213, i32 213> 747; CHECK-NEXT: [[TMP15:%.*]] = select <4 x i1> [[TMP14]], <4 x i32> [[TMP13]], <4 x i32> <i32 22, i32 22, i32 22, i32 22> 748; CHECK-NEXT: [[TMP16:%.*]] = bitcast i32* [[TMP9]] to <4 x i32>* 749; CHECK-NEXT: store <4 x i32> [[TMP15]], <4 x i32>* [[TMP16]], align 4 750; CHECK-NEXT: [[INDEX_NEXT]] = add nuw i32 [[INDEX]], 4 751; CHECK-NEXT: [[TMP17:%.*]] = icmp eq i32 [[INDEX_NEXT]], [[N_VEC]] 752; CHECK-NEXT: br i1 [[TMP17]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP12:![0-9]+]] 753; CHECK: middle.block: 754; CHECK-NEXT: [[CMP_N:%.*]] = icmp eq i32 [[UMAX1]], [[N_VEC]] 755; CHECK-NEXT: [[VECTOR_RECUR_EXTRACT:%.*]] = extractelement <4 x i64> [[TMP11]], i32 3 756; CHECK-NEXT: [[VECTOR_RECUR_EXTRACT_FOR_PHI:%.*]] = extractelement <4 x i64> [[TMP11]], i32 2 757; CHECK-NEXT: br i1 [[CMP_N]], label [[EXIT:%.*]], label [[SCALAR_PH]] 758; CHECK: scalar.ph: 759; CHECK-NEXT: [[SCALAR_RECUR_INIT:%.*]] = phi i64 [ 0, [[VECTOR_SCEVCHECK]] ], [ 0, [[ENTRY:%.*]] ], [ [[VECTOR_RECUR_EXTRACT]], [[MIDDLE_BLOCK]] ] 760; CHECK-NEXT: [[BC_RESUME_VAL:%.*]] = phi i32 [ [[N_VEC]], [[MIDDLE_BLOCK]] ], [ 0, [[ENTRY]] ], [ 0, [[VECTOR_SCEVCHECK]] ] 761; CHECK-NEXT: br label [[LOOP:%.*]] 762; CHECK: loop: 763; CHECK-NEXT: [[SCALAR_RECUR:%.*]] = phi i64 [ [[SCALAR_RECUR_INIT]], [[SCALAR_PH]] ], [ [[FOR_NEXT:%.*]], [[LOOP]] ] 764; CHECK-NEXT: [[IV:%.*]] = phi i32 [ [[BC_RESUME_VAL]], [[SCALAR_PH]] ], [ [[IV_NEXT:%.*]], [[LOOP]] ] 765; CHECK-NEXT: [[FOR_TRUNC:%.*]] = trunc i64 [[SCALAR_RECUR]] to i32 766; CHECK-NEXT: [[CMP:%.*]] = icmp slt i32 [[FOR_TRUNC]], 213 767; CHECK-NEXT: [[SELECT:%.*]] = select i1 [[CMP]], i32 [[FOR_TRUNC]], i32 22 768; CHECK-NEXT: [[GEP:%.*]] = getelementptr inbounds i32, i32* [[PTR]], i32 [[IV]] 769; CHECK-NEXT: [[LV:%.*]] = load i32, i32* [[GEP]], align 4 770; CHECK-NEXT: [[FOR_NEXT]] = zext i32 [[LV]] to i64 771; CHECK-NEXT: store i32 [[SELECT]], i32* [[GEP]], align 4 772; CHECK-NEXT: [[IV_NEXT]] = add i32 [[IV]], 1 773; CHECK-NEXT: [[CMP73:%.*]] = icmp ugt i32 [[N]], [[IV_NEXT]] 774; CHECK-NEXT: br i1 [[CMP73]], label [[LOOP]], label [[EXIT]], !llvm.loop [[LOOP13:![0-9]+]] 775; CHECK: exit: 776; CHECK-NEXT: ret void 777; 778entry: 779 br label %loop 780 781loop: 782 %for = phi i64 [ 0, %entry ], [ %for.next, %loop ] 783 %iv = phi i32 [ 0, %entry ], [ %iv.next, %loop ] 784 785 %for.trunc = trunc i64 %for to i32 786 %cmp = icmp slt i32 %for.trunc, 213 787 %select = select i1 %cmp, i32 %for.trunc, i32 22 788 789 %gep = getelementptr inbounds i32, i32* %ptr, i32 %iv 790 %lv = load i32, i32* %gep, align 4 791 %for.next = zext i32 %lv to i64 792 store i32 %select, i32* %gep 793 794 %iv.next = add i32 %iv, 1 795 %cmp73 = icmp ugt i32 %N, %iv.next 796 br i1 %cmp73, label %loop, label %exit 797 798exit: 799 ret void 800} 801 802; Similar to @sink_dominance, but with 2 separate chains that merge at %select 803; with a different number of instructions in between. 804define void @sink_dominance_2(i32* %ptr, i32 %N) { 805; CHECK-LABEL: @sink_dominance_2( 806; CHECK-NEXT: entry: 807; CHECK-NEXT: [[UMAX1:%.*]] = call i32 @llvm.umax.i32(i32 [[N:%.*]], i32 1) 808; CHECK-NEXT: [[MIN_ITERS_CHECK:%.*]] = icmp ult i32 [[UMAX1]], 4 809; CHECK-NEXT: br i1 [[MIN_ITERS_CHECK]], label [[SCALAR_PH:%.*]], label [[VECTOR_SCEVCHECK:%.*]] 810; CHECK: vector.scevcheck: 811; CHECK-NEXT: [[UMAX:%.*]] = call i32 @llvm.umax.i32(i32 [[N]], i32 1) 812; CHECK-NEXT: [[TMP0:%.*]] = add i32 [[UMAX]], -1 813; CHECK-NEXT: [[TMP4:%.*]] = icmp slt i32 [[TMP0]], 0 814; CHECK-NEXT: br i1 [[TMP4]], label [[SCALAR_PH]], label [[VECTOR_PH:%.*]] 815; CHECK: vector.ph: 816; CHECK-NEXT: [[N_MOD_VF:%.*]] = urem i32 [[UMAX1]], 4 817; CHECK-NEXT: [[N_VEC:%.*]] = sub i32 [[UMAX1]], [[N_MOD_VF]] 818; CHECK-NEXT: br label [[VECTOR_BODY:%.*]] 819; CHECK: vector.body: 820; CHECK-NEXT: [[INDEX:%.*]] = phi i32 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ] 821; CHECK-NEXT: [[VECTOR_RECUR:%.*]] = phi <4 x i64> [ <i64 poison, i64 poison, i64 poison, i64 0>, [[VECTOR_PH]] ], [ [[TMP11:%.*]], [[VECTOR_BODY]] ] 822; CHECK-NEXT: [[TMP7:%.*]] = add i32 [[INDEX]], 0 823; CHECK-NEXT: [[TMP8:%.*]] = getelementptr inbounds i32, i32* [[PTR:%.*]], i32 [[TMP7]] 824; CHECK-NEXT: [[TMP9:%.*]] = getelementptr inbounds i32, i32* [[TMP8]], i32 0 825; CHECK-NEXT: [[TMP10:%.*]] = bitcast i32* [[TMP9]] to <4 x i32>* 826; CHECK-NEXT: [[WIDE_LOAD:%.*]] = load <4 x i32>, <4 x i32>* [[TMP10]], align 4 827; CHECK-NEXT: [[TMP11]] = zext <4 x i32> [[WIDE_LOAD]] to <4 x i64> 828; CHECK-NEXT: [[TMP12:%.*]] = shufflevector <4 x i64> [[VECTOR_RECUR]], <4 x i64> [[TMP11]], <4 x i32> <i32 3, i32 4, i32 5, i32 6> 829; CHECK-NEXT: [[TMP13:%.*]] = trunc <4 x i64> [[TMP12]] to <4 x i32> 830; CHECK-NEXT: [[TMP14:%.*]] = add <4 x i32> [[TMP13]], <i32 2, i32 2, i32 2, i32 2> 831; CHECK-NEXT: [[TMP15:%.*]] = mul <4 x i32> [[TMP14]], <i32 99, i32 99, i32 99, i32 99> 832; CHECK-NEXT: [[TMP16:%.*]] = icmp slt <4 x i32> [[TMP13]], <i32 213, i32 213, i32 213, i32 213> 833; CHECK-NEXT: [[TMP17:%.*]] = select <4 x i1> [[TMP16]], <4 x i32> [[TMP13]], <4 x i32> [[TMP15]] 834; CHECK-NEXT: [[TMP18:%.*]] = bitcast i32* [[TMP9]] to <4 x i32>* 835; CHECK-NEXT: store <4 x i32> [[TMP17]], <4 x i32>* [[TMP18]], align 4 836; CHECK-NEXT: [[INDEX_NEXT]] = add nuw i32 [[INDEX]], 4 837; CHECK-NEXT: [[TMP19:%.*]] = icmp eq i32 [[INDEX_NEXT]], [[N_VEC]] 838; CHECK-NEXT: br i1 [[TMP19]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP14:![0-9]+]] 839; CHECK: middle.block: 840; CHECK-NEXT: [[CMP_N:%.*]] = icmp eq i32 [[UMAX1]], [[N_VEC]] 841; CHECK-NEXT: [[VECTOR_RECUR_EXTRACT:%.*]] = extractelement <4 x i64> [[TMP11]], i32 3 842; CHECK-NEXT: [[VECTOR_RECUR_EXTRACT_FOR_PHI:%.*]] = extractelement <4 x i64> [[TMP11]], i32 2 843; CHECK-NEXT: br i1 [[CMP_N]], label [[EXIT:%.*]], label [[SCALAR_PH]] 844; CHECK: scalar.ph: 845; CHECK-NEXT: [[SCALAR_RECUR_INIT:%.*]] = phi i64 [ 0, [[VECTOR_SCEVCHECK]] ], [ 0, [[ENTRY:%.*]] ], [ [[VECTOR_RECUR_EXTRACT]], [[MIDDLE_BLOCK]] ] 846; CHECK-NEXT: [[BC_RESUME_VAL:%.*]] = phi i32 [ [[N_VEC]], [[MIDDLE_BLOCK]] ], [ 0, [[ENTRY]] ], [ 0, [[VECTOR_SCEVCHECK]] ] 847; CHECK-NEXT: br label [[LOOP:%.*]] 848; CHECK: loop: 849; CHECK-NEXT: [[SCALAR_RECUR:%.*]] = phi i64 [ [[SCALAR_RECUR_INIT]], [[SCALAR_PH]] ], [ [[FOR_NEXT:%.*]], [[LOOP]] ] 850; CHECK-NEXT: [[IV:%.*]] = phi i32 [ [[BC_RESUME_VAL]], [[SCALAR_PH]] ], [ [[IV_NEXT:%.*]], [[LOOP]] ] 851; CHECK-NEXT: [[FOR_TRUNC:%.*]] = trunc i64 [[SCALAR_RECUR]] to i32 852; CHECK-NEXT: [[STEP:%.*]] = add i32 [[FOR_TRUNC]], 2 853; CHECK-NEXT: [[STEP_2:%.*]] = mul i32 [[STEP]], 99 854; CHECK-NEXT: [[CMP:%.*]] = icmp slt i32 [[FOR_TRUNC]], 213 855; CHECK-NEXT: [[SELECT:%.*]] = select i1 [[CMP]], i32 [[FOR_TRUNC]], i32 [[STEP_2]] 856; CHECK-NEXT: [[GEP:%.*]] = getelementptr inbounds i32, i32* [[PTR]], i32 [[IV]] 857; CHECK-NEXT: [[LV:%.*]] = load i32, i32* [[GEP]], align 4 858; CHECK-NEXT: [[FOR_NEXT]] = zext i32 [[LV]] to i64 859; CHECK-NEXT: store i32 [[SELECT]], i32* [[GEP]], align 4 860; CHECK-NEXT: [[IV_NEXT]] = add i32 [[IV]], 1 861; CHECK-NEXT: [[CMP73:%.*]] = icmp ugt i32 [[N]], [[IV_NEXT]] 862; CHECK-NEXT: br i1 [[CMP73]], label [[LOOP]], label [[EXIT]], !llvm.loop [[LOOP15:![0-9]+]] 863; CHECK: exit: 864; CHECK-NEXT: ret void 865; 866entry: 867 br label %loop 868 869loop: 870 %for = phi i64 [ 0, %entry ], [ %for.next, %loop ] 871 %iv = phi i32 [ 0, %entry ], [ %iv.next, %loop ] 872 873 %for.trunc = trunc i64 %for to i32 874 %step = add i32 %for.trunc, 2 875 %step.2 = mul i32 %step, 99 876 877 %cmp = icmp slt i32 %for.trunc, 213 878 %select = select i1 %cmp, i32 %for.trunc, i32 %step.2 879 880 %gep = getelementptr inbounds i32, i32* %ptr, i32 %iv 881 %lv = load i32, i32* %gep, align 4 882 %for.next = zext i32 %lv to i64 883 store i32 %select, i32* %gep 884 885 %iv.next = add i32 %iv, 1 886 %cmp73 = icmp ugt i32 %N, %iv.next 887 br i1 %cmp73, label %loop, label %exit 888 889exit: 890 ret void 891} 892 893define void @cannot_sink_load_past_store(i32* %ptr, i32 %N) { 894; CHECK-LABEL: @cannot_sink_load_past_store( 895; CHECK-NEXT: entry: 896; CHECK-NEXT: br label [[LOOP:%.*]] 897; CHECK: loop: 898; CHECK-NEXT: [[FOR:%.*]] = phi i64 [ 0, [[ENTRY:%.*]] ], [ [[FOR_NEXT:%.*]], [[LOOP]] ] 899; CHECK-NEXT: [[IV:%.*]] = phi i32 [ 0, [[ENTRY]] ], [ [[IV_NEXT:%.*]], [[LOOP]] ] 900; CHECK-NEXT: [[GEP_FOR:%.*]] = getelementptr inbounds i32, i32* [[PTR:%.*]], i64 [[FOR]] 901; CHECK-NEXT: [[LV_FOR:%.*]] = load i32, i32* [[GEP_FOR]], align 4 902; CHECK-NEXT: [[FOR_TRUNC:%.*]] = trunc i64 [[FOR]] to i32 903; CHECK-NEXT: [[CMP:%.*]] = icmp slt i32 [[LV_FOR]], [[FOR_TRUNC]] 904; CHECK-NEXT: [[SELECT:%.*]] = select i1 [[CMP]], i32 [[LV_FOR]], i32 22 905; CHECK-NEXT: [[GEP_IV:%.*]] = getelementptr inbounds i32, i32* [[PTR]], i32 [[IV]] 906; CHECK-NEXT: store i32 0, i32* [[GEP_IV]], align 4 907; CHECK-NEXT: [[IV_NEXT]] = add i32 [[IV]], 1 908; CHECK-NEXT: [[FOR_NEXT]] = zext i32 [[IV]] to i64 909; CHECK-NEXT: [[CMP73:%.*]] = icmp ugt i32 [[N:%.*]], [[IV_NEXT]] 910; CHECK-NEXT: br i1 [[CMP73]], label [[LOOP]], label [[EXIT:%.*]] 911; CHECK: exit: 912; CHECK-NEXT: ret void 913; 914entry: 915 br label %loop 916 917loop: 918 %for = phi i64 [ 0, %entry ], [ %for.next, %loop ] 919 %iv = phi i32 [ 0, %entry ], [ %iv.next, %loop ] 920 921 %gep.for = getelementptr inbounds i32, i32* %ptr, i64 %for 922 %lv.for = load i32, i32* %gep.for, align 4 923 %for.trunc = trunc i64 %for to i32 924 %cmp = icmp slt i32 %lv.for, %for.trunc 925 %select = select i1 %cmp, i32 %lv.for, i32 22 926 927 %gep.iv = getelementptr inbounds i32, i32* %ptr, i32 %iv 928 store i32 0, i32* %gep.iv 929 %iv.next = add i32 %iv, 1 930 %for.next = zext i32 %iv to i64 931 932 %cmp73 = icmp ugt i32 %N, %iv.next 933 br i1 %cmp73, label %loop, label %exit 934 935exit: 936 ret void 937} 938 939define void @test_for_sink_instruction_after_same_incoming_1(double* %ptr) { 940; CHECK-LABEL: @test_for_sink_instruction_after_same_incoming_1 941; CHECK-NOT: vector.body: 942; 943entry: 944 br label %loop 945 946loop: 947 %for.1 = phi double [ 10.0, %entry ], [ %for.1.next, %loop ] 948 %for.2 = phi double [ 20.0, %entry ], [ %for.1.next, %loop ] 949 %iv = phi i64 [ 1, %entry ], [ %iv.next, %loop ] 950 %add.1 = fadd double 10.0, %for.2 951 %add.2 = fadd double %add.1, %for.1 952 %iv.next = add nuw nsw i64 %iv, 1 953 %gep.ptr = getelementptr inbounds double, double* %ptr, i64 %iv 954 %for.1.next = load double, double* %gep.ptr, align 8 955 store double %add.2, double* %gep.ptr 956 %exitcond.not = icmp eq i64 %iv.next, 1000 957 br i1 %exitcond.not, label %exit, label %loop 958 959exit: 960 ret void 961} 962 963 964define void @test_for_sink_instruction_after_same_incoming_2(double* %ptr) { 965; CHECK-LABEL: @test_for_sink_instruction_after_same_incoming_2 966; CHECK-NOT: vector.body: 967entry: 968 br label %loop 969 970loop: 971 %for.2 = phi double [ 20.0, %entry ], [ %for.1.next, %loop ] 972 %for.1 = phi double [ 10.0, %entry ], [ %for.1.next, %loop ] 973 %iv = phi i64 [ 1, %entry ], [ %iv.next, %loop ] 974 %add.1 = fadd double 10.0, %for.2 975 %add.2 = fadd double %add.1, %for.1 976 %iv.next = add nuw nsw i64 %iv, 1 977 %gep.ptr = getelementptr inbounds double, double* %ptr, i64 %iv 978 %for.1.next = load double, double* %gep.ptr, align 8 979 store double %add.2, double* %gep.ptr 980 %exitcond.not = icmp eq i64 %iv.next, 1000 981 br i1 %exitcond.not, label %exit, label %loop 982 983exit: 984 ret void 985} 986