266 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_test_checks.py2; RUN: opt -mcpu=bdver1 -passes=loop-vectorize -S %s | FileCheck %s --check-prefix=CHECK3; RUN: opt -mcpu=bdver1 -passes=loop-vectorize -debug-only=loop-vectorize -disable-output %s 2>&1 | FileCheck %s --check-prefix=CHECK-COST4; REQUIRES: asserts5 6target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"7target triple = "x86_64-unknown-linux-gnu"8 9; CHECK-COST-LABEL: uaddsat10 11define void @uaddsat(ptr nocapture readonly %pSrc, i16 signext %offset, ptr nocapture noalias %pDst, i32 %blockSize) #0 {12; CHECK-LABEL: @uaddsat(13; CHECK-NEXT: entry:14; CHECK-NEXT: [[CMP_NOT6:%.*]] = icmp eq i32 [[BLOCKSIZE:%.*]], 015; CHECK-NEXT: br i1 [[CMP_NOT6]], label [[WHILE_END:%.*]], label [[ITER_CHECK:%.*]]16; CHECK: iter.check:17; CHECK-NEXT: [[TMP0:%.*]] = zext i32 [[BLOCKSIZE]] to i6418; CHECK-NEXT: [[MIN_ITERS_CHECK:%.*]] = icmp ult i64 [[TMP0]], 819; CHECK-NEXT: br i1 [[MIN_ITERS_CHECK]], label [[VEC_EPILOG_SCALAR_PH:%.*]], label [[VECTOR_MAIN_LOOP_ITER_CHECK:%.*]]20; CHECK: vector.main.loop.iter.check:21; CHECK-NEXT: [[MIN_ITERS_CHECK1:%.*]] = icmp ult i64 [[TMP0]], 6422; CHECK-NEXT: br i1 [[MIN_ITERS_CHECK1]], label [[VEC_EPILOG_PH:%.*]], label [[VECTOR_PH:%.*]]23; CHECK: vector.ph:24; CHECK-NEXT: [[N_MOD_VF:%.*]] = urem i64 [[TMP0]], 6425; CHECK-NEXT: [[N_VEC:%.*]] = sub i64 [[TMP0]], [[N_MOD_VF]]26; CHECK-NEXT: [[BROADCAST_SPLATINSERT:%.*]] = insertelement <16 x i16> poison, i16 [[OFFSET:%.*]], i64 027; CHECK-NEXT: [[BROADCAST_SPLAT:%.*]] = shufflevector <16 x i16> [[BROADCAST_SPLATINSERT]], <16 x i16> poison, <16 x i32> zeroinitializer28; CHECK-NEXT: br label [[VECTOR_BODY:%.*]]29; CHECK: vector.body:30; CHECK-NEXT: [[INDEX:%.*]] = phi i64 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]31; CHECK-NEXT: [[OFFSET_IDX:%.*]] = mul i64 [[INDEX]], 232; CHECK-NEXT: [[NEXT_GEP:%.*]] = getelementptr i8, ptr [[PSRC:%.*]], i64 [[OFFSET_IDX]]33; CHECK-NEXT: [[OFFSET_IDX2:%.*]] = mul i64 [[INDEX]], 234; CHECK-NEXT: [[NEXT_GEP3:%.*]] = getelementptr i8, ptr [[PDST:%.*]], i64 [[OFFSET_IDX2]]35; CHECK-NEXT: [[TMP1:%.*]] = getelementptr i16, ptr [[NEXT_GEP]], i64 1636; CHECK-NEXT: [[TMP2:%.*]] = getelementptr i16, ptr [[NEXT_GEP]], i64 3237; CHECK-NEXT: [[TMP3:%.*]] = getelementptr i16, ptr [[NEXT_GEP]], i64 4838; CHECK-NEXT: [[WIDE_LOAD:%.*]] = load <16 x i16>, ptr [[NEXT_GEP]], align 239; CHECK-NEXT: [[WIDE_LOAD4:%.*]] = load <16 x i16>, ptr [[TMP1]], align 240; CHECK-NEXT: [[WIDE_LOAD5:%.*]] = load <16 x i16>, ptr [[TMP2]], align 241; CHECK-NEXT: [[WIDE_LOAD6:%.*]] = load <16 x i16>, ptr [[TMP3]], align 242; CHECK-NEXT: [[TMP4:%.*]] = call <16 x i16> @llvm.uadd.sat.v16i16(<16 x i16> [[WIDE_LOAD]], <16 x i16> [[BROADCAST_SPLAT]])43; CHECK-NEXT: [[TMP5:%.*]] = call <16 x i16> @llvm.uadd.sat.v16i16(<16 x i16> [[WIDE_LOAD4]], <16 x i16> [[BROADCAST_SPLAT]])44; CHECK-NEXT: [[TMP6:%.*]] = call <16 x i16> @llvm.uadd.sat.v16i16(<16 x i16> [[WIDE_LOAD5]], <16 x i16> [[BROADCAST_SPLAT]])45; CHECK-NEXT: [[TMP7:%.*]] = call <16 x i16> @llvm.uadd.sat.v16i16(<16 x i16> [[WIDE_LOAD6]], <16 x i16> [[BROADCAST_SPLAT]])46; CHECK-NEXT: [[TMP8:%.*]] = getelementptr i16, ptr [[NEXT_GEP3]], i64 1647; CHECK-NEXT: [[TMP9:%.*]] = getelementptr i16, ptr [[NEXT_GEP3]], i64 3248; CHECK-NEXT: [[TMP10:%.*]] = getelementptr i16, ptr [[NEXT_GEP3]], i64 4849; CHECK-NEXT: store <16 x i16> [[TMP4]], ptr [[NEXT_GEP3]], align 250; CHECK-NEXT: store <16 x i16> [[TMP5]], ptr [[TMP8]], align 251; CHECK-NEXT: store <16 x i16> [[TMP6]], ptr [[TMP9]], align 252; CHECK-NEXT: store <16 x i16> [[TMP7]], ptr [[TMP10]], align 253; CHECK-NEXT: [[INDEX_NEXT]] = add nuw i64 [[INDEX]], 6454; CHECK-NEXT: [[TMP11:%.*]] = icmp eq i64 [[INDEX_NEXT]], [[N_VEC]]55; CHECK-NEXT: br i1 [[TMP11]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP0:![0-9]+]]56; CHECK: middle.block:57; CHECK-NEXT: [[CMP_N:%.*]] = icmp eq i64 [[TMP0]], [[N_VEC]]58; CHECK-NEXT: br i1 [[CMP_N]], label [[WHILE_END_LOOPEXIT:%.*]], label [[VEC_EPILOG_ITER_CHECK:%.*]]59; CHECK: vec.epilog.iter.check:60; CHECK-NEXT: [[DOTCAST1:%.*]] = trunc i64 [[N_VEC]] to i3261; CHECK-NEXT: [[IND_END10:%.*]] = sub i32 [[BLOCKSIZE]], [[DOTCAST1]]62; CHECK-NEXT: [[TMP12:%.*]] = mul i64 [[N_VEC]], 263; CHECK-NEXT: [[IND_END12:%.*]] = getelementptr i8, ptr [[PSRC]], i64 [[TMP12]]64; CHECK-NEXT: [[TMP13:%.*]] = mul i64 [[N_VEC]], 265; CHECK-NEXT: [[IND_END15:%.*]] = getelementptr i8, ptr [[PDST]], i64 [[TMP13]]66; CHECK-NEXT: [[MIN_EPILOG_ITERS_CHECK:%.*]] = icmp ult i64 [[N_MOD_VF]], 867; CHECK-NEXT: br i1 [[MIN_EPILOG_ITERS_CHECK]], label [[VEC_EPILOG_SCALAR_PH]], label [[VEC_EPILOG_PH]], !prof [[PROF3:![0-9]+]]68; CHECK: vec.epilog.ph:69; CHECK-NEXT: [[VEC_EPILOG_RESUME_VAL:%.*]] = phi i64 [ [[N_VEC]], [[VEC_EPILOG_ITER_CHECK]] ], [ 0, [[VECTOR_MAIN_LOOP_ITER_CHECK]] ]70; CHECK-NEXT: [[N_MOD_VF6:%.*]] = urem i64 [[TMP0]], 871; CHECK-NEXT: [[N_VEC8:%.*]] = sub i64 [[TMP0]], [[N_MOD_VF6]]72; CHECK-NEXT: [[DOTCAST:%.*]] = trunc i64 [[N_VEC8]] to i3273; CHECK-NEXT: [[IND_END:%.*]] = sub i32 [[BLOCKSIZE]], [[DOTCAST]]74; CHECK-NEXT: [[TMP14:%.*]] = mul i64 [[N_VEC8]], 275; CHECK-NEXT: [[IND_END11:%.*]] = getelementptr i8, ptr [[PSRC]], i64 [[TMP14]]76; CHECK-NEXT: [[TMP15:%.*]] = mul i64 [[N_VEC8]], 277; CHECK-NEXT: [[IND_END14:%.*]] = getelementptr i8, ptr [[PDST]], i64 [[TMP15]]78; CHECK-NEXT: [[BROADCAST_SPLATINSERT23:%.*]] = insertelement <8 x i16> poison, i16 [[OFFSET]], i64 079; CHECK-NEXT: [[BROADCAST_SPLAT24:%.*]] = shufflevector <8 x i16> [[BROADCAST_SPLATINSERT23]], <8 x i16> poison, <8 x i32> zeroinitializer80; CHECK-NEXT: br label [[VEC_EPILOG_VECTOR_BODY:%.*]]81; CHECK: vec.epilog.vector.body:82; CHECK-NEXT: [[INDEX17:%.*]] = phi i64 [ [[VEC_EPILOG_RESUME_VAL]], [[VEC_EPILOG_PH]] ], [ [[INDEX_NEXT25:%.*]], [[VEC_EPILOG_VECTOR_BODY]] ]83; CHECK-NEXT: [[OFFSET_IDX18:%.*]] = mul i64 [[INDEX17]], 284; CHECK-NEXT: [[NEXT_GEP19:%.*]] = getelementptr i8, ptr [[PSRC]], i64 [[OFFSET_IDX18]]85; CHECK-NEXT: [[OFFSET_IDX20:%.*]] = mul i64 [[INDEX17]], 286; CHECK-NEXT: [[NEXT_GEP21:%.*]] = getelementptr i8, ptr [[PDST]], i64 [[OFFSET_IDX20]]87; CHECK-NEXT: [[WIDE_LOAD22:%.*]] = load <8 x i16>, ptr [[NEXT_GEP19]], align 288; CHECK-NEXT: [[TMP16:%.*]] = call <8 x i16> @llvm.uadd.sat.v8i16(<8 x i16> [[WIDE_LOAD22]], <8 x i16> [[BROADCAST_SPLAT24]])89; CHECK-NEXT: store <8 x i16> [[TMP16]], ptr [[NEXT_GEP21]], align 290; CHECK-NEXT: [[INDEX_NEXT25]] = add nuw i64 [[INDEX17]], 891; CHECK-NEXT: [[TMP17:%.*]] = icmp eq i64 [[INDEX_NEXT25]], [[N_VEC8]]92; CHECK-NEXT: br i1 [[TMP17]], label [[VEC_EPILOG_MIDDLE_BLOCK:%.*]], label [[VEC_EPILOG_VECTOR_BODY]], !llvm.loop [[LOOP4:![0-9]+]]93; CHECK: vec.epilog.middle.block:94; CHECK-NEXT: [[CMP_N16:%.*]] = icmp eq i64 [[TMP0]], [[N_VEC8]]95; CHECK-NEXT: br i1 [[CMP_N16]], label [[WHILE_END_LOOPEXIT]], label [[VEC_EPILOG_SCALAR_PH]]96; CHECK: vec.epilog.scalar.ph:97; CHECK-NEXT: [[BC_RESUME_VAL:%.*]] = phi i32 [ [[IND_END]], [[VEC_EPILOG_MIDDLE_BLOCK]] ], [ [[IND_END10]], [[VEC_EPILOG_ITER_CHECK]] ], [ [[BLOCKSIZE]], [[ITER_CHECK]] ]98; CHECK-NEXT: [[BC_RESUME_VAL13:%.*]] = phi ptr [ [[IND_END11]], [[VEC_EPILOG_MIDDLE_BLOCK]] ], [ [[IND_END12]], [[VEC_EPILOG_ITER_CHECK]] ], [ [[PSRC]], [[ITER_CHECK]] ]99; CHECK-NEXT: [[BC_RESUME_VAL16:%.*]] = phi ptr [ [[IND_END14]], [[VEC_EPILOG_MIDDLE_BLOCK]] ], [ [[IND_END15]], [[VEC_EPILOG_ITER_CHECK]] ], [ [[PDST]], [[ITER_CHECK]] ]100; CHECK-NEXT: br label [[WHILE_BODY:%.*]]101; CHECK: while.body:102; CHECK-NEXT: [[BLKCNT_09:%.*]] = phi i32 [ [[DEC:%.*]], [[WHILE_BODY]] ], [ [[BC_RESUME_VAL]], [[VEC_EPILOG_SCALAR_PH]] ]103; CHECK-NEXT: [[PSRC_ADDR_08:%.*]] = phi ptr [ [[INCDEC_PTR:%.*]], [[WHILE_BODY]] ], [ [[BC_RESUME_VAL13]], [[VEC_EPILOG_SCALAR_PH]] ]104; CHECK-NEXT: [[PDST_ADDR_07:%.*]] = phi ptr [ [[INCDEC_PTR3:%.*]], [[WHILE_BODY]] ], [ [[BC_RESUME_VAL16]], [[VEC_EPILOG_SCALAR_PH]] ]105; CHECK-NEXT: [[INCDEC_PTR]] = getelementptr inbounds i16, ptr [[PSRC_ADDR_08]], i32 1106; CHECK-NEXT: [[TMP18:%.*]] = load i16, ptr [[PSRC_ADDR_08]], align 2107; CHECK-NEXT: [[TMP19:%.*]] = tail call i16 @llvm.uadd.sat.i16(i16 [[TMP18]], i16 [[OFFSET]])108; CHECK-NEXT: [[INCDEC_PTR3]] = getelementptr inbounds i16, ptr [[PDST_ADDR_07]], i32 1109; CHECK-NEXT: store i16 [[TMP19]], ptr [[PDST_ADDR_07]], align 2110; CHECK-NEXT: [[DEC]] = add i32 [[BLKCNT_09]], -1111; CHECK-NEXT: [[CMP_NOT:%.*]] = icmp eq i32 [[DEC]], 0112; CHECK-NEXT: br i1 [[CMP_NOT]], label [[WHILE_END_LOOPEXIT]], label [[WHILE_BODY]], !llvm.loop [[LOOP5:![0-9]+]]113; CHECK: while.end.loopexit:114; CHECK-NEXT: br label [[WHILE_END]]115; CHECK: while.end:116; CHECK-NEXT: ret void117;118entry:119 %cmp.not6 = icmp eq i32 %blockSize, 0120 br i1 %cmp.not6, label %while.end, label %while.body121 122while.body: ; preds = %entry, %while.body123 %blkCnt.09 = phi i32 [ %dec, %while.body ], [ %blockSize, %entry ]124 %pSrc.addr.08 = phi ptr [ %incdec.ptr, %while.body ], [ %pSrc, %entry ]125 %pDst.addr.07 = phi ptr [ %incdec.ptr3, %while.body ], [ %pDst, %entry ]126 %incdec.ptr = getelementptr inbounds i16, ptr %pSrc.addr.08, i32 1127 %0 = load i16, ptr %pSrc.addr.08, align 2128 %1 = tail call i16 @llvm.uadd.sat.i16(i16 %0, i16 %offset)129 %incdec.ptr3 = getelementptr inbounds i16, ptr %pDst.addr.07, i32 1130 store i16 %1, ptr %pDst.addr.07, align 2131 %dec = add i32 %blkCnt.09, -1132 %cmp.not = icmp eq i32 %dec, 0133 br i1 %cmp.not, label %while.end, label %while.body134 135while.end: ; preds = %while.body, %entry136 ret void137}138 139; CHECK-COST-LABEL: fshl140 141define void @fshl(ptr nocapture readonly %pSrc, i8 signext %offset, ptr nocapture noalias %pDst, i32 %blockSize) #0 {142; CHECK-LABEL: @fshl(143; CHECK-NEXT: entry:144; CHECK-NEXT: [[CMP_NOT6:%.*]] = icmp eq i32 [[BLOCKSIZE:%.*]], 0145; CHECK-NEXT: br i1 [[CMP_NOT6]], label [[WHILE_END:%.*]], label [[ITER_CHECK:%.*]]146; CHECK: iter.check:147; CHECK-NEXT: [[TMP0:%.*]] = zext i32 [[BLOCKSIZE]] to i64148; CHECK-NEXT: [[MIN_ITERS_CHECK:%.*]] = icmp ult i64 [[TMP0]], 8149; CHECK-NEXT: br i1 [[MIN_ITERS_CHECK]], label [[VEC_EPILOG_SCALAR_PH:%.*]], label [[VECTOR_MAIN_LOOP_ITER_CHECK:%.*]]150; CHECK: vector.main.loop.iter.check:151; CHECK-NEXT: [[MIN_ITERS_CHECK1:%.*]] = icmp ult i64 [[TMP0]], 128152; CHECK-NEXT: br i1 [[MIN_ITERS_CHECK1]], label [[VEC_EPILOG_PH:%.*]], label [[VECTOR_PH:%.*]]153; CHECK: vector.ph:154; CHECK-NEXT: [[N_MOD_VF:%.*]] = urem i64 [[TMP0]], 128155; CHECK-NEXT: [[N_VEC:%.*]] = sub i64 [[TMP0]], [[N_MOD_VF]]156; CHECK-NEXT: [[BROADCAST_SPLATINSERT:%.*]] = insertelement <32 x i8> poison, i8 [[OFFSET:%.*]], i64 0157; CHECK-NEXT: [[BROADCAST_SPLAT:%.*]] = shufflevector <32 x i8> [[BROADCAST_SPLATINSERT]], <32 x i8> poison, <32 x i32> zeroinitializer158; CHECK-NEXT: br label [[VECTOR_BODY:%.*]]159; CHECK: vector.body:160; CHECK-NEXT: [[INDEX:%.*]] = phi i64 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]161; CHECK-NEXT: [[NEXT_GEP:%.*]] = getelementptr i8, ptr [[PSRC:%.*]], i64 [[INDEX]]162; CHECK-NEXT: [[NEXT_GEP2:%.*]] = getelementptr i8, ptr [[PDST:%.*]], i64 [[INDEX]]163; CHECK-NEXT: [[TMP1:%.*]] = getelementptr i8, ptr [[NEXT_GEP]], i64 32164; CHECK-NEXT: [[TMP2:%.*]] = getelementptr i8, ptr [[NEXT_GEP]], i64 64165; CHECK-NEXT: [[TMP3:%.*]] = getelementptr i8, ptr [[NEXT_GEP]], i64 96166; CHECK-NEXT: [[WIDE_LOAD:%.*]] = load <32 x i8>, ptr [[NEXT_GEP]], align 2167; CHECK-NEXT: [[WIDE_LOAD3:%.*]] = load <32 x i8>, ptr [[TMP1]], align 2168; CHECK-NEXT: [[WIDE_LOAD4:%.*]] = load <32 x i8>, ptr [[TMP2]], align 2169; CHECK-NEXT: [[WIDE_LOAD5:%.*]] = load <32 x i8>, ptr [[TMP3]], align 2170; CHECK-NEXT: [[TMP4:%.*]] = call <32 x i8> @llvm.fshl.v32i8(<32 x i8> [[WIDE_LOAD]], <32 x i8> [[WIDE_LOAD]], <32 x i8> [[BROADCAST_SPLAT]])171; CHECK-NEXT: [[TMP5:%.*]] = call <32 x i8> @llvm.fshl.v32i8(<32 x i8> [[WIDE_LOAD3]], <32 x i8> [[WIDE_LOAD3]], <32 x i8> [[BROADCAST_SPLAT]])172; CHECK-NEXT: [[TMP6:%.*]] = call <32 x i8> @llvm.fshl.v32i8(<32 x i8> [[WIDE_LOAD4]], <32 x i8> [[WIDE_LOAD4]], <32 x i8> [[BROADCAST_SPLAT]])173; CHECK-NEXT: [[TMP7:%.*]] = call <32 x i8> @llvm.fshl.v32i8(<32 x i8> [[WIDE_LOAD5]], <32 x i8> [[WIDE_LOAD5]], <32 x i8> [[BROADCAST_SPLAT]])174; CHECK-NEXT: [[TMP8:%.*]] = getelementptr i8, ptr [[NEXT_GEP2]], i64 32175; CHECK-NEXT: [[TMP9:%.*]] = getelementptr i8, ptr [[NEXT_GEP2]], i64 64176; CHECK-NEXT: [[TMP10:%.*]] = getelementptr i8, ptr [[NEXT_GEP2]], i64 96177; CHECK-NEXT: store <32 x i8> [[TMP4]], ptr [[NEXT_GEP2]], align 2178; CHECK-NEXT: store <32 x i8> [[TMP5]], ptr [[TMP8]], align 2179; CHECK-NEXT: store <32 x i8> [[TMP6]], ptr [[TMP9]], align 2180; CHECK-NEXT: store <32 x i8> [[TMP7]], ptr [[TMP10]], align 2181; CHECK-NEXT: [[INDEX_NEXT]] = add nuw i64 [[INDEX]], 128182; CHECK-NEXT: [[TMP11:%.*]] = icmp eq i64 [[INDEX_NEXT]], [[N_VEC]]183; CHECK-NEXT: br i1 [[TMP11]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP6:![0-9]+]]184; CHECK: middle.block:185; CHECK-NEXT: [[CMP_N:%.*]] = icmp eq i64 [[TMP0]], [[N_VEC]]186; CHECK-NEXT: br i1 [[CMP_N]], label [[WHILE_END_LOOPEXIT:%.*]], label [[VEC_EPILOG_ITER_CHECK:%.*]]187; CHECK: vec.epilog.iter.check:188; CHECK-NEXT: [[DOTCAST8:%.*]] = trunc i64 [[N_VEC]] to i32189; CHECK-NEXT: [[IND_END9:%.*]] = sub i32 [[BLOCKSIZE]], [[DOTCAST8]]190; CHECK-NEXT: [[IND_END11:%.*]] = getelementptr i8, ptr [[PSRC]], i64 [[N_VEC]]191; CHECK-NEXT: [[IND_END14:%.*]] = getelementptr i8, ptr [[PDST]], i64 [[N_VEC]]192; CHECK-NEXT: [[MIN_EPILOG_ITERS_CHECK:%.*]] = icmp ult i64 [[N_MOD_VF]], 8193; CHECK-NEXT: br i1 [[MIN_EPILOG_ITERS_CHECK]], label [[VEC_EPILOG_SCALAR_PH]], label [[VEC_EPILOG_PH]], !prof [[PROF7:![0-9]+]]194; CHECK: vec.epilog.ph:195; CHECK-NEXT: [[VEC_EPILOG_RESUME_VAL:%.*]] = phi i64 [ [[N_VEC]], [[VEC_EPILOG_ITER_CHECK]] ], [ 0, [[VECTOR_MAIN_LOOP_ITER_CHECK]] ]196; CHECK-NEXT: [[N_MOD_VF6:%.*]] = urem i64 [[TMP0]], 8197; CHECK-NEXT: [[N_VEC7:%.*]] = sub i64 [[TMP0]], [[N_MOD_VF6]]198; CHECK-NEXT: [[DOTCAST:%.*]] = trunc i64 [[N_VEC7]] to i32199; CHECK-NEXT: [[IND_END:%.*]] = sub i32 [[BLOCKSIZE]], [[DOTCAST]]200; CHECK-NEXT: [[IND_END10:%.*]] = getelementptr i8, ptr [[PSRC]], i64 [[N_VEC7]]201; CHECK-NEXT: [[IND_END13:%.*]] = getelementptr i8, ptr [[PDST]], i64 [[N_VEC7]]202; CHECK-NEXT: [[BROADCAST_SPLATINSERT20:%.*]] = insertelement <8 x i8> poison, i8 [[OFFSET]], i64 0203; CHECK-NEXT: [[BROADCAST_SPLAT21:%.*]] = shufflevector <8 x i8> [[BROADCAST_SPLATINSERT20]], <8 x i8> poison, <8 x i32> zeroinitializer204; CHECK-NEXT: br label [[VEC_EPILOG_VECTOR_BODY:%.*]]205; CHECK: vec.epilog.vector.body:206; CHECK-NEXT: [[INDEX16:%.*]] = phi i64 [ [[VEC_EPILOG_RESUME_VAL]], [[VEC_EPILOG_PH]] ], [ [[INDEX_NEXT22:%.*]], [[VEC_EPILOG_VECTOR_BODY]] ]207; CHECK-NEXT: [[NEXT_GEP17:%.*]] = getelementptr i8, ptr [[PSRC]], i64 [[INDEX16]]208; CHECK-NEXT: [[NEXT_GEP18:%.*]] = getelementptr i8, ptr [[PDST]], i64 [[INDEX16]]209; CHECK-NEXT: [[WIDE_LOAD19:%.*]] = load <8 x i8>, ptr [[NEXT_GEP17]], align 2210; CHECK-NEXT: [[TMP12:%.*]] = call <8 x i8> @llvm.fshl.v8i8(<8 x i8> [[WIDE_LOAD19]], <8 x i8> [[WIDE_LOAD19]], <8 x i8> [[BROADCAST_SPLAT21]])211; CHECK-NEXT: store <8 x i8> [[TMP12]], ptr [[NEXT_GEP18]], align 2212; CHECK-NEXT: [[INDEX_NEXT22]] = add nuw i64 [[INDEX16]], 8213; CHECK-NEXT: [[TMP13:%.*]] = icmp eq i64 [[INDEX_NEXT22]], [[N_VEC7]]214; CHECK-NEXT: br i1 [[TMP13]], label [[VEC_EPILOG_MIDDLE_BLOCK:%.*]], label [[VEC_EPILOG_VECTOR_BODY]], !llvm.loop [[LOOP8:![0-9]+]]215; CHECK: vec.epilog.middle.block:216; CHECK-NEXT: [[CMP_N15:%.*]] = icmp eq i64 [[TMP0]], [[N_VEC7]]217; CHECK-NEXT: br i1 [[CMP_N15]], label [[WHILE_END_LOOPEXIT]], label [[VEC_EPILOG_SCALAR_PH]]218; CHECK: vec.epilog.scalar.ph:219; CHECK-NEXT: [[BC_RESUME_VAL:%.*]] = phi i32 [ [[IND_END]], [[VEC_EPILOG_MIDDLE_BLOCK]] ], [ [[IND_END9]], [[VEC_EPILOG_ITER_CHECK]] ], [ [[BLOCKSIZE]], [[ITER_CHECK]] ]220; CHECK-NEXT: [[BC_RESUME_VAL12:%.*]] = phi ptr [ [[IND_END10]], [[VEC_EPILOG_MIDDLE_BLOCK]] ], [ [[IND_END11]], [[VEC_EPILOG_ITER_CHECK]] ], [ [[PSRC]], [[ITER_CHECK]] ]221; CHECK-NEXT: [[BC_RESUME_VAL15:%.*]] = phi ptr [ [[IND_END13]], [[VEC_EPILOG_MIDDLE_BLOCK]] ], [ [[IND_END14]], [[VEC_EPILOG_ITER_CHECK]] ], [ [[PDST]], [[ITER_CHECK]] ]222; CHECK-NEXT: br label [[WHILE_BODY:%.*]]223; CHECK: while.body:224; CHECK-NEXT: [[BLKCNT_09:%.*]] = phi i32 [ [[DEC:%.*]], [[WHILE_BODY]] ], [ [[BC_RESUME_VAL]], [[VEC_EPILOG_SCALAR_PH]] ]225; CHECK-NEXT: [[PSRC_ADDR_08:%.*]] = phi ptr [ [[INCDEC_PTR:%.*]], [[WHILE_BODY]] ], [ [[BC_RESUME_VAL12]], [[VEC_EPILOG_SCALAR_PH]] ]226; CHECK-NEXT: [[PDST_ADDR_07:%.*]] = phi ptr [ [[INCDEC_PTR3:%.*]], [[WHILE_BODY]] ], [ [[BC_RESUME_VAL15]], [[VEC_EPILOG_SCALAR_PH]] ]227; CHECK-NEXT: [[INCDEC_PTR]] = getelementptr inbounds i8, ptr [[PSRC_ADDR_08]], i32 1228; CHECK-NEXT: [[TMP14:%.*]] = load i8, ptr [[PSRC_ADDR_08]], align 2229; CHECK-NEXT: [[TMP15:%.*]] = tail call i8 @llvm.fshl.i8(i8 [[TMP14]], i8 [[TMP14]], i8 [[OFFSET]])230; CHECK-NEXT: [[INCDEC_PTR3]] = getelementptr inbounds i8, ptr [[PDST_ADDR_07]], i32 1231; CHECK-NEXT: store i8 [[TMP15]], ptr [[PDST_ADDR_07]], align 2232; CHECK-NEXT: [[DEC]] = add i32 [[BLKCNT_09]], -1233; CHECK-NEXT: [[CMP_NOT:%.*]] = icmp eq i32 [[DEC]], 0234; CHECK-NEXT: br i1 [[CMP_NOT]], label [[WHILE_END_LOOPEXIT]], label [[WHILE_BODY]], !llvm.loop [[LOOP9:![0-9]+]]235; CHECK: while.end.loopexit:236; CHECK-NEXT: br label [[WHILE_END]]237; CHECK: while.end:238; CHECK-NEXT: ret void239;240entry:241 %cmp.not6 = icmp eq i32 %blockSize, 0242 br i1 %cmp.not6, label %while.end, label %while.body243 244while.body: ; preds = %entry, %while.body245 %blkCnt.09 = phi i32 [ %dec, %while.body ], [ %blockSize, %entry ]246 %pSrc.addr.08 = phi ptr [ %incdec.ptr, %while.body ], [ %pSrc, %entry ]247 %pDst.addr.07 = phi ptr [ %incdec.ptr3, %while.body ], [ %pDst, %entry ]248 %incdec.ptr = getelementptr inbounds i8, ptr %pSrc.addr.08, i32 1249 %0 = load i8, ptr %pSrc.addr.08, align 2250 %1 = tail call i8 @llvm.fshl.i8(i8 %0, i8 %0, i8 %offset)251 %incdec.ptr3 = getelementptr inbounds i8, ptr %pDst.addr.07, i32 1252 store i8 %1, ptr %pDst.addr.07, align 2253 %dec = add i32 %blkCnt.09, -1254 %cmp.not = icmp eq i32 %dec, 0255 br i1 %cmp.not, label %while.end, label %while.body256 257while.end: ; preds = %while.body, %entry258 ret void259}260 261declare i16 @llvm.uadd.sat.i16(i16, i16)262declare i8 @llvm.fshl.i8(i8, i8, i8)263 264;; NOTE: These prefixes are unused and the list is autogenerated. Do not add tests below this line:265; CHECK-COST: {{.*}}266