1; NOTE: Assertions have been autogenerated by utils/update_test_checks.py 2; RUN: opt -passes=loop-vectorize,dce,instcombine -mtriple aarch64-linux-gnu -mattr=+sve \ 3; RUN: -prefer-predicate-over-epilogue=scalar-epilogue -S %s -force-target-instruction-cost=1 -o - | FileCheck %s 4 5define void @gather_nxv4i32_ind64(ptr noalias nocapture readonly %a, ptr noalias nocapture readonly %b, ptr noalias nocapture %c, i64 %n) #0 { 6; CHECK-LABEL: @gather_nxv4i32_ind64( 7; CHECK-NEXT: entry: 8; CHECK-NEXT: [[TMP0:%.*]] = call i64 @llvm.vscale.i64() 9; CHECK-NEXT: [[TMP1:%.*]] = shl nuw nsw i64 [[TMP0]], 2 10; CHECK-NEXT: [[MIN_ITERS_CHECK:%.*]] = icmp ult i64 [[N:%.*]], [[TMP1]] 11; CHECK-NEXT: br i1 [[MIN_ITERS_CHECK]], label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]] 12; CHECK: vector.ph: 13; CHECK-NEXT: [[TMP2:%.*]] = call i64 @llvm.vscale.i64() 14; CHECK-NEXT: [[DOTNEG:%.*]] = mul nsw i64 [[TMP2]], -4 15; CHECK-NEXT: [[N_VEC:%.*]] = and i64 [[N]], [[DOTNEG]] 16; CHECK-NEXT: [[TMP3:%.*]] = call i64 @llvm.vscale.i64() 17; CHECK-NEXT: [[TMP4:%.*]] = shl nuw nsw i64 [[TMP3]], 2 18; CHECK-NEXT: br label [[VECTOR_BODY:%.*]] 19; CHECK: vector.body: 20; CHECK-NEXT: [[INDEX:%.*]] = phi i64 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ] 21; CHECK-NEXT: [[TMP5:%.*]] = getelementptr inbounds i64, ptr [[B:%.*]], i64 [[INDEX]] 22; CHECK-NEXT: [[WIDE_LOAD:%.*]] = load <vscale x 4 x i64>, ptr [[TMP5]], align 8 23; CHECK-NEXT: [[TMP6:%.*]] = getelementptr inbounds float, ptr [[A:%.*]], <vscale x 4 x i64> [[WIDE_LOAD]] 24; CHECK-NEXT: [[WIDE_MASKED_GATHER:%.*]] = call <vscale x 4 x float> @llvm.masked.gather.nxv4f32.nxv4p0(<vscale x 4 x ptr> [[TMP6]], i32 4, <vscale x 4 x i1> splat (i1 true), <vscale x 4 x float> poison) 25; CHECK-NEXT: [[TMP7:%.*]] = getelementptr inbounds float, ptr [[C:%.*]], i64 [[INDEX]] 26; CHECK-NEXT: store <vscale x 4 x float> [[WIDE_MASKED_GATHER]], ptr [[TMP7]], align 4 27; CHECK-NEXT: [[INDEX_NEXT]] = add nuw i64 [[INDEX]], [[TMP4]] 28; CHECK-NEXT: [[TMP8:%.*]] = icmp eq i64 [[INDEX_NEXT]], [[N_VEC]] 29; CHECK-NEXT: br i1 [[TMP8]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP0:![0-9]+]] 30; CHECK: middle.block: 31; CHECK-NEXT: [[CMP_N:%.*]] = icmp eq i64 [[N]], [[N_VEC]] 32; CHECK-NEXT: br i1 [[CMP_N]], label [[FOR_COND_CLEANUP:%.*]], label [[SCALAR_PH]] 33; CHECK: scalar.ph: 34; CHECK-NEXT: [[BC_RESUME_VAL:%.*]] = phi i64 [ [[N_VEC]], [[MIDDLE_BLOCK]] ], [ 0, [[ENTRY:%.*]] ] 35; CHECK-NEXT: br label [[FOR_BODY:%.*]] 36; CHECK: for.body: 37; CHECK-NEXT: [[INDVARS_IV:%.*]] = phi i64 [ [[INDVARS_IV_NEXT:%.*]], [[FOR_BODY]] ], [ [[BC_RESUME_VAL]], [[SCALAR_PH]] ] 38; CHECK-NEXT: [[ARRAYIDX:%.*]] = getelementptr inbounds i64, ptr [[B]], i64 [[INDVARS_IV]] 39; CHECK-NEXT: [[TMP9:%.*]] = load i64, ptr [[ARRAYIDX]], align 8 40; CHECK-NEXT: [[ARRAYIDX3:%.*]] = getelementptr inbounds float, ptr [[A]], i64 [[TMP9]] 41; CHECK-NEXT: [[TMP10:%.*]] = load float, ptr [[ARRAYIDX3]], align 4 42; CHECK-NEXT: [[ARRAYIDX5:%.*]] = getelementptr inbounds float, ptr [[C]], i64 [[INDVARS_IV]] 43; CHECK-NEXT: store float [[TMP10]], ptr [[ARRAYIDX5]], align 4 44; CHECK-NEXT: [[INDVARS_IV_NEXT]] = add nuw nsw i64 [[INDVARS_IV]], 1 45; CHECK-NEXT: [[EXITCOND_NOT:%.*]] = icmp eq i64 [[INDVARS_IV_NEXT]], [[N]] 46; CHECK-NEXT: br i1 [[EXITCOND_NOT]], label [[FOR_COND_CLEANUP]], label [[FOR_BODY]], !llvm.loop [[LOOP4:![0-9]+]] 47; CHECK: for.cond.cleanup: 48; CHECK-NEXT: ret void 49; 50entry: 51 br label %for.body 52 53for.body: ; preds = %entry, %for.body 54 %indvars.iv = phi i64 [ %indvars.iv.next, %for.body ], [ 0, %entry ] 55 %arrayidx = getelementptr inbounds i64, ptr %b, i64 %indvars.iv 56 %0 = load i64, ptr %arrayidx, align 8 57 %arrayidx3 = getelementptr inbounds float, ptr %a, i64 %0 58 %1 = load float, ptr %arrayidx3, align 4 59 %arrayidx5 = getelementptr inbounds float, ptr %c, i64 %indvars.iv 60 store float %1, ptr %arrayidx5, align 4 61 %indvars.iv.next = add nuw nsw i64 %indvars.iv, 1 62 %exitcond.not = icmp eq i64 %indvars.iv.next, %n 63 br i1 %exitcond.not, label %for.cond.cleanup, label %for.body, !llvm.loop !0 64 65for.cond.cleanup: ; preds = %for.cond.cleanup.loopexit, %entry 66 ret void 67} 68 69; NOTE: I deliberately chose '%b' as an array of i32 indices, since the 70; additional 'sext' in the for.body loop exposes additional code paths 71; during vectorisation. 72define void @scatter_nxv4i32_ind32(ptr noalias nocapture %a, ptr noalias nocapture readonly %b, ptr noalias nocapture readonly %c, i64 %n) #0 { 73; CHECK-LABEL: @scatter_nxv4i32_ind32( 74; CHECK-NEXT: entry: 75; CHECK-NEXT: [[TMP0:%.*]] = call i64 @llvm.vscale.i64() 76; CHECK-NEXT: [[TMP1:%.*]] = shl nuw nsw i64 [[TMP0]], 2 77; CHECK-NEXT: [[MIN_ITERS_CHECK:%.*]] = icmp ult i64 [[N:%.*]], [[TMP1]] 78; CHECK-NEXT: br i1 [[MIN_ITERS_CHECK]], label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]] 79; CHECK: vector.ph: 80; CHECK-NEXT: [[TMP2:%.*]] = call i64 @llvm.vscale.i64() 81; CHECK-NEXT: [[DOTNEG:%.*]] = mul nsw i64 [[TMP2]], -4 82; CHECK-NEXT: [[N_VEC:%.*]] = and i64 [[N]], [[DOTNEG]] 83; CHECK-NEXT: [[TMP3:%.*]] = call i64 @llvm.vscale.i64() 84; CHECK-NEXT: [[TMP4:%.*]] = shl nuw nsw i64 [[TMP3]], 2 85; CHECK-NEXT: br label [[VECTOR_BODY:%.*]] 86; CHECK: vector.body: 87; CHECK-NEXT: [[INDEX:%.*]] = phi i64 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ] 88; CHECK-NEXT: [[TMP5:%.*]] = getelementptr inbounds float, ptr [[C:%.*]], i64 [[INDEX]] 89; CHECK-NEXT: [[WIDE_LOAD:%.*]] = load <vscale x 4 x float>, ptr [[TMP5]], align 4 90; CHECK-NEXT: [[TMP6:%.*]] = getelementptr inbounds i32, ptr [[B:%.*]], i64 [[INDEX]] 91; CHECK-NEXT: [[WIDE_LOAD1:%.*]] = load <vscale x 4 x i32>, ptr [[TMP6]], align 4 92; CHECK-NEXT: [[TMP7:%.*]] = sext <vscale x 4 x i32> [[WIDE_LOAD1]] to <vscale x 4 x i64> 93; CHECK-NEXT: [[TMP8:%.*]] = getelementptr inbounds float, ptr [[A:%.*]], <vscale x 4 x i64> [[TMP7]] 94; CHECK-NEXT: call void @llvm.masked.scatter.nxv4f32.nxv4p0(<vscale x 4 x float> [[WIDE_LOAD]], <vscale x 4 x ptr> [[TMP8]], i32 4, <vscale x 4 x i1> splat (i1 true)) 95; CHECK-NEXT: [[INDEX_NEXT]] = add nuw i64 [[INDEX]], [[TMP4]] 96; CHECK-NEXT: [[TMP9:%.*]] = icmp eq i64 [[INDEX_NEXT]], [[N_VEC]] 97; CHECK-NEXT: br i1 [[TMP9]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP5:![0-9]+]] 98; CHECK: middle.block: 99; CHECK-NEXT: [[CMP_N:%.*]] = icmp eq i64 [[N]], [[N_VEC]] 100; CHECK-NEXT: br i1 [[CMP_N]], label [[FOR_COND_CLEANUP:%.*]], label [[SCALAR_PH]] 101; CHECK: scalar.ph: 102; CHECK-NEXT: [[BC_RESUME_VAL:%.*]] = phi i64 [ [[N_VEC]], [[MIDDLE_BLOCK]] ], [ 0, [[ENTRY:%.*]] ] 103; CHECK-NEXT: br label [[FOR_BODY:%.*]] 104; CHECK: for.body: 105; CHECK-NEXT: [[INDVARS_IV:%.*]] = phi i64 [ [[INDVARS_IV_NEXT:%.*]], [[FOR_BODY]] ], [ [[BC_RESUME_VAL]], [[SCALAR_PH]] ] 106; CHECK-NEXT: [[ARRAYIDX:%.*]] = getelementptr inbounds float, ptr [[C]], i64 [[INDVARS_IV]] 107; CHECK-NEXT: [[TMP10:%.*]] = load float, ptr [[ARRAYIDX]], align 4 108; CHECK-NEXT: [[ARRAYIDX3:%.*]] = getelementptr inbounds i32, ptr [[B]], i64 [[INDVARS_IV]] 109; CHECK-NEXT: [[TMP11:%.*]] = load i32, ptr [[ARRAYIDX3]], align 4 110; CHECK-NEXT: [[IDXPROM4:%.*]] = sext i32 [[TMP11]] to i64 111; CHECK-NEXT: [[ARRAYIDX5:%.*]] = getelementptr inbounds float, ptr [[A]], i64 [[IDXPROM4]] 112; CHECK-NEXT: store float [[TMP10]], ptr [[ARRAYIDX5]], align 4 113; CHECK-NEXT: [[INDVARS_IV_NEXT]] = add nuw nsw i64 [[INDVARS_IV]], 1 114; CHECK-NEXT: [[EXITCOND_NOT:%.*]] = icmp eq i64 [[INDVARS_IV_NEXT]], [[N]] 115; CHECK-NEXT: br i1 [[EXITCOND_NOT]], label [[FOR_COND_CLEANUP]], label [[FOR_BODY]], !llvm.loop [[LOOP6:![0-9]+]] 116; CHECK: for.cond.cleanup: 117; CHECK-NEXT: ret void 118; 119entry: 120 br label %for.body 121 122for.body: ; preds = %entry, %for.body 123 %indvars.iv = phi i64 [ %indvars.iv.next, %for.body ], [ 0, %entry ] 124 %arrayidx = getelementptr inbounds float, ptr %c, i64 %indvars.iv 125 %0 = load float, ptr %arrayidx, align 4 126 %arrayidx3 = getelementptr inbounds i32, ptr %b, i64 %indvars.iv 127 %1 = load i32, ptr %arrayidx3, align 4 128 %idxprom4 = sext i32 %1 to i64 129 %arrayidx5 = getelementptr inbounds float, ptr %a, i64 %idxprom4 130 store float %0, ptr %arrayidx5, align 4 131 %indvars.iv.next = add nuw nsw i64 %indvars.iv, 1 132 %exitcond.not = icmp eq i64 %indvars.iv.next, %n 133 br i1 %exitcond.not, label %for.cond.cleanup, label %for.body, !llvm.loop !0 134 135for.cond.cleanup: ; preds = %for.body, %entry 136 ret void 137} 138 139define void @scatter_inv_nxv4i32(ptr noalias nocapture %inv, ptr noalias nocapture readonly %b, i64 %n) #0 { 140; CHECK-LABEL: @scatter_inv_nxv4i32( 141; CHECK-NEXT: entry: 142; CHECK-NEXT: [[TMP0:%.*]] = call i64 @llvm.vscale.i64() 143; CHECK-NEXT: [[TMP1:%.*]] = shl nuw nsw i64 [[TMP0]], 2 144; CHECK-NEXT: [[MIN_ITERS_CHECK:%.*]] = icmp ult i64 [[N:%.*]], [[TMP1]] 145; CHECK-NEXT: br i1 [[MIN_ITERS_CHECK]], label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]] 146; CHECK: vector.ph: 147; CHECK-NEXT: [[TMP2:%.*]] = call i64 @llvm.vscale.i64() 148; CHECK-NEXT: [[DOTNEG:%.*]] = mul nsw i64 [[TMP2]], -4 149; CHECK-NEXT: [[N_VEC:%.*]] = and i64 [[N]], [[DOTNEG]] 150; CHECK-NEXT: [[TMP3:%.*]] = call i64 @llvm.vscale.i64() 151; CHECK-NEXT: [[TMP4:%.*]] = shl nuw nsw i64 [[TMP3]], 2 152; CHECK-NEXT: [[BROADCAST_SPLATINSERT:%.*]] = insertelement <vscale x 4 x ptr> poison, ptr [[INV:%.*]], i64 0 153; CHECK-NEXT: [[BROADCAST_SPLAT:%.*]] = shufflevector <vscale x 4 x ptr> [[BROADCAST_SPLATINSERT]], <vscale x 4 x ptr> poison, <vscale x 4 x i32> zeroinitializer 154; CHECK-NEXT: br label [[VECTOR_BODY:%.*]] 155; CHECK: vector.body: 156; CHECK-NEXT: [[INDEX:%.*]] = phi i64 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ] 157; CHECK-NEXT: [[TMP5:%.*]] = getelementptr inbounds i32, ptr [[B:%.*]], i64 [[INDEX]] 158; CHECK-NEXT: [[WIDE_LOAD:%.*]] = load <vscale x 4 x i32>, ptr [[TMP5]], align 4 159; CHECK-NEXT: [[TMP6:%.*]] = icmp ne <vscale x 4 x i32> [[WIDE_LOAD]], zeroinitializer 160; CHECK-NEXT: call void @llvm.masked.scatter.nxv4i32.nxv4p0(<vscale x 4 x i32> splat (i32 3), <vscale x 4 x ptr> [[BROADCAST_SPLAT]], i32 4, <vscale x 4 x i1> [[TMP6]]) 161; CHECK-NEXT: [[INDEX_NEXT]] = add nuw i64 [[INDEX]], [[TMP4]] 162; CHECK-NEXT: [[TMP7:%.*]] = icmp eq i64 [[INDEX_NEXT]], [[N_VEC]] 163; CHECK-NEXT: br i1 [[TMP7]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP7:![0-9]+]] 164; CHECK: middle.block: 165; CHECK-NEXT: [[CMP_N:%.*]] = icmp eq i64 [[N]], [[N_VEC]] 166; CHECK-NEXT: br i1 [[CMP_N]], label [[FOR_COND_CLEANUP:%.*]], label [[SCALAR_PH]] 167; CHECK: scalar.ph: 168; CHECK-NEXT: [[BC_RESUME_VAL:%.*]] = phi i64 [ [[N_VEC]], [[MIDDLE_BLOCK]] ], [ 0, [[ENTRY:%.*]] ] 169; CHECK-NEXT: br label [[FOR_BODY:%.*]] 170; CHECK: for.body: 171; CHECK-NEXT: [[INDVARS_IV:%.*]] = phi i64 [ [[INDVARS_IV_NEXT:%.*]], [[FOR_INC:%.*]] ], [ [[BC_RESUME_VAL]], [[SCALAR_PH]] ] 172; CHECK-NEXT: [[ARRAYIDX:%.*]] = getelementptr inbounds i32, ptr [[B]], i64 [[INDVARS_IV]] 173; CHECK-NEXT: [[TMP8:%.*]] = load i32, ptr [[ARRAYIDX]], align 4 174; CHECK-NEXT: [[TOBOOL_NOT:%.*]] = icmp eq i32 [[TMP8]], 0 175; CHECK-NEXT: br i1 [[TOBOOL_NOT]], label [[FOR_INC]], label [[IF_THEN:%.*]] 176; CHECK: if.then: 177; CHECK-NEXT: store i32 3, ptr [[INV]], align 4 178; CHECK-NEXT: br label [[FOR_INC]] 179; CHECK: for.inc: 180; CHECK-NEXT: [[INDVARS_IV_NEXT]] = add nuw nsw i64 [[INDVARS_IV]], 1 181; CHECK-NEXT: [[EXITCOND_NOT:%.*]] = icmp eq i64 [[INDVARS_IV_NEXT]], [[N]] 182; CHECK-NEXT: br i1 [[EXITCOND_NOT]], label [[FOR_COND_CLEANUP]], label [[FOR_BODY]], !llvm.loop [[LOOP8:![0-9]+]] 183; CHECK: for.cond.cleanup: 184; CHECK-NEXT: ret void 185; 186entry: 187 br label %for.body 188 189for.body: ; preds = %entry, %for.inc 190 %indvars.iv = phi i64 [ %indvars.iv.next, %for.inc ], [ 0, %entry ] 191 %arrayidx = getelementptr inbounds i32, ptr %b, i64 %indvars.iv 192 %0 = load i32, ptr %arrayidx, align 4 193 %tobool.not = icmp eq i32 %0, 0 194 br i1 %tobool.not, label %for.inc, label %if.then 195 196if.then: ; preds = %for.body 197 store i32 3, ptr %inv, align 4 198 br label %for.inc 199 200for.inc: ; preds = %for.body, %if.then 201 %indvars.iv.next = add nuw nsw i64 %indvars.iv, 1 202 %exitcond.not = icmp eq i64 %indvars.iv.next, %n 203 br i1 %exitcond.not, label %for.cond.cleanup, label %for.body, !llvm.loop !0 204 205for.cond.cleanup: ; preds = %for.cond.cleanup.loopexit, %entry 206 ret void 207} 208 209define void @gather_inv_nxv4i32(ptr noalias nocapture %a, ptr noalias nocapture readonly %inv, i64 %n) #0 { 210; CHECK-LABEL: @gather_inv_nxv4i32( 211; CHECK-NEXT: entry: 212; CHECK-NEXT: [[TMP0:%.*]] = call i64 @llvm.vscale.i64() 213; CHECK-NEXT: [[TMP1:%.*]] = shl nuw nsw i64 [[TMP0]], 2 214; CHECK-NEXT: [[MIN_ITERS_CHECK:%.*]] = icmp ult i64 [[N:%.*]], [[TMP1]] 215; CHECK-NEXT: br i1 [[MIN_ITERS_CHECK]], label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]] 216; CHECK: vector.ph: 217; CHECK-NEXT: [[TMP2:%.*]] = call i64 @llvm.vscale.i64() 218; CHECK-NEXT: [[DOTNEG:%.*]] = mul nsw i64 [[TMP2]], -4 219; CHECK-NEXT: [[N_VEC:%.*]] = and i64 [[N]], [[DOTNEG]] 220; CHECK-NEXT: [[TMP3:%.*]] = call i64 @llvm.vscale.i64() 221; CHECK-NEXT: [[TMP4:%.*]] = shl nuw nsw i64 [[TMP3]], 2 222; CHECK-NEXT: [[BROADCAST_SPLATINSERT:%.*]] = insertelement <vscale x 4 x ptr> poison, ptr [[INV:%.*]], i64 0 223; CHECK-NEXT: [[BROADCAST_SPLAT:%.*]] = shufflevector <vscale x 4 x ptr> [[BROADCAST_SPLATINSERT]], <vscale x 4 x ptr> poison, <vscale x 4 x i32> zeroinitializer 224; CHECK-NEXT: br label [[VECTOR_BODY:%.*]] 225; CHECK: vector.body: 226; CHECK-NEXT: [[INDEX:%.*]] = phi i64 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ] 227; CHECK-NEXT: [[TMP5:%.*]] = getelementptr i32, ptr [[A:%.*]], i64 [[INDEX]] 228; CHECK-NEXT: [[WIDE_LOAD:%.*]] = load <vscale x 4 x i32>, ptr [[TMP5]], align 4 229; CHECK-NEXT: [[TMP6:%.*]] = icmp sgt <vscale x 4 x i32> [[WIDE_LOAD]], splat (i32 3) 230; CHECK-NEXT: [[WIDE_MASKED_GATHER:%.*]] = call <vscale x 4 x i32> @llvm.masked.gather.nxv4i32.nxv4p0(<vscale x 4 x ptr> [[BROADCAST_SPLAT]], i32 4, <vscale x 4 x i1> [[TMP6]], <vscale x 4 x i32> poison) 231; CHECK-NEXT: call void @llvm.masked.store.nxv4i32.p0(<vscale x 4 x i32> [[WIDE_MASKED_GATHER]], ptr [[TMP5]], i32 4, <vscale x 4 x i1> [[TMP6]]) 232; CHECK-NEXT: [[INDEX_NEXT]] = add nuw i64 [[INDEX]], [[TMP4]] 233; CHECK-NEXT: [[TMP7:%.*]] = icmp eq i64 [[INDEX_NEXT]], [[N_VEC]] 234; CHECK-NEXT: br i1 [[TMP7]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP9:![0-9]+]] 235; CHECK: middle.block: 236; CHECK-NEXT: [[CMP_N:%.*]] = icmp eq i64 [[N]], [[N_VEC]] 237; CHECK-NEXT: br i1 [[CMP_N]], label [[FOR_COND_CLEANUP:%.*]], label [[SCALAR_PH]] 238; CHECK: scalar.ph: 239; CHECK-NEXT: [[BC_RESUME_VAL:%.*]] = phi i64 [ [[N_VEC]], [[MIDDLE_BLOCK]] ], [ 0, [[ENTRY:%.*]] ] 240; CHECK-NEXT: br label [[FOR_BODY:%.*]] 241; CHECK: for.body: 242; CHECK-NEXT: [[INDVARS_IV:%.*]] = phi i64 [ [[INDVARS_IV_NEXT:%.*]], [[FOR_INC:%.*]] ], [ [[BC_RESUME_VAL]], [[SCALAR_PH]] ] 243; CHECK-NEXT: [[ARRAYIDX:%.*]] = getelementptr inbounds i32, ptr [[A]], i64 [[INDVARS_IV]] 244; CHECK-NEXT: [[TMP8:%.*]] = load i32, ptr [[ARRAYIDX]], align 4 245; CHECK-NEXT: [[CMP2:%.*]] = icmp sgt i32 [[TMP8]], 3 246; CHECK-NEXT: br i1 [[CMP2]], label [[IF_THEN:%.*]], label [[FOR_INC]] 247; CHECK: if.then: 248; CHECK-NEXT: [[TMP9:%.*]] = load i32, ptr [[INV]], align 4 249; CHECK-NEXT: store i32 [[TMP9]], ptr [[ARRAYIDX]], align 4 250; CHECK-NEXT: br label [[FOR_INC]] 251; CHECK: for.inc: 252; CHECK-NEXT: [[INDVARS_IV_NEXT]] = add nuw nsw i64 [[INDVARS_IV]], 1 253; CHECK-NEXT: [[EXITCOND_NOT:%.*]] = icmp eq i64 [[INDVARS_IV_NEXT]], [[N]] 254; CHECK-NEXT: br i1 [[EXITCOND_NOT]], label [[FOR_COND_CLEANUP]], label [[FOR_BODY]], !llvm.loop [[LOOP10:![0-9]+]] 255; CHECK: for.cond.cleanup: 256; CHECK-NEXT: ret void 257; 258entry: 259 br label %for.body 260 261for.body: ; preds = %entry, %for.inc 262 %indvars.iv = phi i64 [ %indvars.iv.next, %for.inc ], [ 0, %entry ] 263 %arrayidx = getelementptr inbounds i32, ptr %a, i64 %indvars.iv 264 %0 = load i32, ptr %arrayidx, align 4 265 %cmp2 = icmp sgt i32 %0, 3 266 br i1 %cmp2, label %if.then, label %for.inc 267 268if.then: ; preds = %for.body 269 %1 = load i32, ptr %inv, align 4 270 store i32 %1, ptr %arrayidx, align 4 271 br label %for.inc 272 273for.inc: ; preds = %for.body, %if.then 274 %indvars.iv.next = add nuw nsw i64 %indvars.iv, 1 275 %exitcond.not = icmp eq i64 %indvars.iv.next, %n 276 br i1 %exitcond.not, label %for.cond.cleanup, label %for.body, !llvm.loop !0 277 278for.cond.cleanup: ; preds = %for.inc, %entry 279 ret void 280} 281 282define void @gather_nxv4i32_ind64_stride2(ptr noalias nocapture %a, ptr noalias nocapture readonly %b, i64 %n) #0 { 283; CHECK-LABEL: @gather_nxv4i32_ind64_stride2( 284; CHECK-NEXT: entry: 285; CHECK-NEXT: [[TMP0:%.*]] = call i64 @llvm.vscale.i64() 286; CHECK-NEXT: [[TMP1:%.*]] = shl nuw nsw i64 [[TMP0]], 3 287; CHECK-NEXT: [[MIN_ITERS_CHECK:%.*]] = icmp ugt i64 [[N:%.*]], [[TMP1]] 288; CHECK-NEXT: br i1 [[MIN_ITERS_CHECK]], label [[VECTOR_PH:%.*]], label [[SCALAR_PH:%.*]] 289; CHECK: vector.ph: 290; CHECK-NEXT: [[TMP2:%.*]] = call i64 @llvm.vscale.i64() 291; CHECK-NEXT: [[TMP7:%.*]] = shl nuw nsw i64 [[TMP2]], 3 292; CHECK-NEXT: [[DOTNEG:%.*]] = add nsw i64 [[TMP7]], -1 293; CHECK-NEXT: [[N_VEC:%.*]] = and i64 [[N]], [[DOTNEG]] 294; CHECK-NEXT: [[TMP5:%.*]] = icmp eq i64 [[N_VEC]], 0 295; CHECK-NEXT: [[TMP6:%.*]] = select i1 [[TMP5]], i64 [[TMP7]], i64 [[N_VEC]] 296; CHECK-NEXT: [[N_VEC1:%.*]] = sub i64 [[N]], [[TMP6]] 297; CHECK-NEXT: [[TMP3:%.*]] = call i64 @llvm.vscale.i64() 298; CHECK-NEXT: [[TMP4:%.*]] = shl nuw nsw i64 [[TMP3]], 3 299; CHECK-NEXT: br label [[VECTOR_BODY:%.*]] 300; CHECK: vector.body: 301; CHECK-NEXT: [[INDEX:%.*]] = phi i64 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ] 302; CHECK-NEXT: [[TMP9:%.*]] = call i64 @llvm.vscale.i64() 303; CHECK-NEXT: [[DOTIDX1:%.*]] = shl i64 [[INDEX]], 3 304; CHECK-NEXT: [[TMP10:%.*]] = getelementptr inbounds i8, ptr [[B:%.*]], i64 [[DOTIDX1]] 305; CHECK-NEXT: [[DOTIDX3:%.*]] = shl nuw nsw i64 [[TMP9]], 5 306; CHECK-NEXT: [[TMP11:%.*]] = getelementptr i8, ptr [[B]], i64 [[DOTIDX3]] 307; CHECK-NEXT: [[DOTIDX4:%.*]] = shl i64 [[INDEX]], 3 308; CHECK-NEXT: [[TMP15:%.*]] = getelementptr i8, ptr [[TMP11]], i64 [[DOTIDX4]] 309; CHECK-NEXT: [[WIDE_VEC:%.*]] = load <vscale x 8 x float>, ptr [[TMP10]], align 4 310; CHECK-NEXT: [[STRIDED_VEC:%.*]] = call { <vscale x 4 x float>, <vscale x 4 x float> } @llvm.vector.deinterleave2.nxv8f32(<vscale x 8 x float> [[WIDE_VEC]]) 311; CHECK-NEXT: [[WIDE_MASKED_GATHER:%.*]] = extractvalue { <vscale x 4 x float>, <vscale x 4 x float> } [[STRIDED_VEC]], 0 312; CHECK-NEXT: [[WIDE_VEC1:%.*]] = load <vscale x 8 x float>, ptr [[TMP15]], align 4 313; CHECK-NEXT: [[STRIDED_VEC2:%.*]] = call { <vscale x 4 x float>, <vscale x 4 x float> } @llvm.vector.deinterleave2.nxv8f32(<vscale x 8 x float> [[WIDE_VEC1]]) 314; CHECK-NEXT: [[WIDE_MASKED_GATHER2:%.*]] = extractvalue { <vscale x 4 x float>, <vscale x 4 x float> } [[STRIDED_VEC2]], 0 315; CHECK-NEXT: [[TMP12:%.*]] = getelementptr inbounds float, ptr [[A:%.*]], i64 [[INDEX]] 316; CHECK-NEXT: [[TMP13:%.*]] = call i64 @llvm.vscale.i64() 317; CHECK-NEXT: [[DOTIDX:%.*]] = shl nuw nsw i64 [[TMP13]], 4 318; CHECK-NEXT: [[TMP14:%.*]] = getelementptr inbounds nuw i8, ptr [[TMP12]], i64 [[DOTIDX]] 319; CHECK-NEXT: store <vscale x 4 x float> [[WIDE_MASKED_GATHER]], ptr [[TMP12]], align 4 320; CHECK-NEXT: store <vscale x 4 x float> [[WIDE_MASKED_GATHER2]], ptr [[TMP14]], align 4 321; CHECK-NEXT: [[INDEX_NEXT]] = add nuw i64 [[INDEX]], [[TMP4]] 322; CHECK-NEXT: [[TMP18:%.*]] = icmp eq i64 [[INDEX_NEXT]], [[N_VEC1]] 323; CHECK-NEXT: br i1 [[TMP18]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP11:![0-9]+]] 324; CHECK: middle.block: 325; CHECK-NEXT: br label [[SCALAR_PH]] 326; CHECK: scalar.ph: 327; CHECK-NEXT: [[BC_RESUME_VAL:%.*]] = phi i64 [ [[N_VEC1]], [[MIDDLE_BLOCK]] ], [ 0, [[ENTRY:%.*]] ] 328; CHECK-NEXT: br label [[FOR_BODY:%.*]] 329; CHECK: for.body: 330; CHECK-NEXT: [[INDVARS_IV:%.*]] = phi i64 [ [[INDVARS_IV_NEXT:%.*]], [[FOR_BODY]] ], [ [[BC_RESUME_VAL]], [[SCALAR_PH]] ] 331; CHECK-NEXT: [[ARRAYIDX_IDX:%.*]] = shl i64 [[INDVARS_IV]], 3 332; CHECK-NEXT: [[ARRAYIDX:%.*]] = getelementptr inbounds i8, ptr [[B]], i64 [[ARRAYIDX_IDX]] 333; CHECK-NEXT: [[TMP16:%.*]] = load float, ptr [[ARRAYIDX]], align 4 334; CHECK-NEXT: [[ARRAYIDX2:%.*]] = getelementptr inbounds float, ptr [[A]], i64 [[INDVARS_IV]] 335; CHECK-NEXT: store float [[TMP16]], ptr [[ARRAYIDX2]], align 4 336; CHECK-NEXT: [[INDVARS_IV_NEXT]] = add nuw nsw i64 [[INDVARS_IV]], 1 337; CHECK-NEXT: [[EXITCOND_NOT:%.*]] = icmp eq i64 [[INDVARS_IV_NEXT]], [[N]] 338; CHECK-NEXT: br i1 [[EXITCOND_NOT]], label [[FOR_COND_CLEANUP:%.*]], label [[FOR_BODY]], !llvm.loop [[LOOP12:![0-9]+]] 339; CHECK: for.cond.cleanup: 340; CHECK-NEXT: ret void 341; 342entry: 343 br label %for.body 344 345for.body: ; preds = %entry, %for.body 346 %indvars.iv = phi i64 [ %indvars.iv.next, %for.body ], [ 0, %entry ] 347 %indvars.iv.stride2 = mul i64 %indvars.iv, 2 348 %arrayidx = getelementptr inbounds float, ptr %b, i64 %indvars.iv.stride2 349 %0 = load float, ptr %arrayidx, align 4 350 %arrayidx2 = getelementptr inbounds float, ptr %a, i64 %indvars.iv 351 store float %0, ptr %arrayidx2, align 4 352 %indvars.iv.next = add nuw nsw i64 %indvars.iv, 1 353 %exitcond.not = icmp eq i64 %indvars.iv.next, %n 354 br i1 %exitcond.not, label %for.cond.cleanup, label %for.body 355 356for.cond.cleanup: ; preds = %for.cond.cleanup.loopexit, %entry 357 ret void 358} 359 360attributes #0 = { vscale_range(1, 16) } 361 362!0 = distinct !{!0, !1, !2, !3, !4, !5} 363!1 = !{!"llvm.loop.mustprogress"} 364!2 = !{!"llvm.loop.vectorize.width", i32 4} 365!3 = !{!"llvm.loop.vectorize.scalable.enable", i1 true} 366!4 = !{!"llvm.loop.interleave.count", i32 1} 367!5 = !{!"llvm.loop.vectorize.enable", i1 true} 368