1; NOTE: Assertions have been autogenerated by utils/update_test_checks.py 2; RUN: opt < %s -passes=slp-vectorizer -S -mtriple=x86_64-apple-macosx10.8.0 -mcpu=corei7-avx | FileCheck %s 3 4define void @tiny_tree_fully_vectorizable(ptr noalias nocapture %dst, ptr noalias nocapture readonly %src, i64 %count) #0 { 5; CHECK-LABEL: @tiny_tree_fully_vectorizable( 6; CHECK-NEXT: entry: 7; CHECK-NEXT: [[CMP12:%.*]] = icmp eq i64 [[COUNT:%.*]], 0 8; CHECK-NEXT: br i1 [[CMP12]], label [[FOR_END:%.*]], label [[FOR_BODY:%.*]] 9; CHECK: for.body: 10; CHECK-NEXT: [[I_015:%.*]] = phi i64 [ [[INC:%.*]], [[FOR_BODY]] ], [ 0, [[ENTRY:%.*]] ] 11; CHECK-NEXT: [[DST_ADDR_014:%.*]] = phi ptr [ [[ADD_PTR4:%.*]], [[FOR_BODY]] ], [ [[DST:%.*]], [[ENTRY]] ] 12; CHECK-NEXT: [[SRC_ADDR_013:%.*]] = phi ptr [ [[ADD_PTR:%.*]], [[FOR_BODY]] ], [ [[SRC:%.*]], [[ENTRY]] ] 13; CHECK-NEXT: [[TMP0:%.*]] = load <2 x double>, ptr [[SRC_ADDR_013]], align 8 14; CHECK-NEXT: store <2 x double> [[TMP0]], ptr [[DST_ADDR_014]], align 8 15; CHECK-NEXT: [[ADD_PTR]] = getelementptr inbounds double, ptr [[SRC_ADDR_013]], i64 [[I_015]] 16; CHECK-NEXT: [[ADD_PTR4]] = getelementptr inbounds double, ptr [[DST_ADDR_014]], i64 [[I_015]] 17; CHECK-NEXT: [[INC]] = add i64 [[I_015]], 1 18; CHECK-NEXT: [[EXITCOND:%.*]] = icmp eq i64 [[INC]], [[COUNT]] 19; CHECK-NEXT: br i1 [[EXITCOND]], label [[FOR_END]], label [[FOR_BODY]] 20; CHECK: for.end: 21; CHECK-NEXT: ret void 22; 23entry: 24 %cmp12 = icmp eq i64 %count, 0 25 br i1 %cmp12, label %for.end, label %for.body 26 27for.body: ; preds = %entry, %for.body 28 %i.015 = phi i64 [ %inc, %for.body ], [ 0, %entry ] 29 %dst.addr.014 = phi ptr [ %add.ptr4, %for.body ], [ %dst, %entry ] 30 %src.addr.013 = phi ptr [ %add.ptr, %for.body ], [ %src, %entry ] 31 %0 = load double, ptr %src.addr.013, align 8 32 store double %0, ptr %dst.addr.014, align 8 33 %arrayidx2 = getelementptr inbounds double, ptr %src.addr.013, i64 1 34 %1 = load double, ptr %arrayidx2, align 8 35 %arrayidx3 = getelementptr inbounds double, ptr %dst.addr.014, i64 1 36 store double %1, ptr %arrayidx3, align 8 37 %add.ptr = getelementptr inbounds double, ptr %src.addr.013, i64 %i.015 38 %add.ptr4 = getelementptr inbounds double, ptr %dst.addr.014, i64 %i.015 39 %inc = add i64 %i.015, 1 40 %exitcond = icmp eq i64 %inc, %count 41 br i1 %exitcond, label %for.end, label %for.body 42 43for.end: ; preds = %for.body, %entry 44 ret void 45} 46 47define void @tiny_tree_fully_vectorizable2(ptr noalias nocapture %dst, ptr noalias nocapture readonly %src, i64 %count) #0 { 48; CHECK-LABEL: @tiny_tree_fully_vectorizable2( 49; CHECK-NEXT: entry: 50; CHECK-NEXT: [[CMP20:%.*]] = icmp eq i64 [[COUNT:%.*]], 0 51; CHECK-NEXT: br i1 [[CMP20]], label [[FOR_END:%.*]], label [[FOR_BODY:%.*]] 52; CHECK: for.body: 53; CHECK-NEXT: [[I_023:%.*]] = phi i64 [ [[INC:%.*]], [[FOR_BODY]] ], [ 0, [[ENTRY:%.*]] ] 54; CHECK-NEXT: [[DST_ADDR_022:%.*]] = phi ptr [ [[ADD_PTR8:%.*]], [[FOR_BODY]] ], [ [[DST:%.*]], [[ENTRY]] ] 55; CHECK-NEXT: [[SRC_ADDR_021:%.*]] = phi ptr [ [[ADD_PTR:%.*]], [[FOR_BODY]] ], [ [[SRC:%.*]], [[ENTRY]] ] 56; CHECK-NEXT: [[TMP0:%.*]] = load <4 x float>, ptr [[SRC_ADDR_021]], align 4 57; CHECK-NEXT: store <4 x float> [[TMP0]], ptr [[DST_ADDR_022]], align 4 58; CHECK-NEXT: [[ADD_PTR]] = getelementptr inbounds float, ptr [[SRC_ADDR_021]], i64 [[I_023]] 59; CHECK-NEXT: [[ADD_PTR8]] = getelementptr inbounds float, ptr [[DST_ADDR_022]], i64 [[I_023]] 60; CHECK-NEXT: [[INC]] = add i64 [[I_023]], 1 61; CHECK-NEXT: [[EXITCOND:%.*]] = icmp eq i64 [[INC]], [[COUNT]] 62; CHECK-NEXT: br i1 [[EXITCOND]], label [[FOR_END]], label [[FOR_BODY]] 63; CHECK: for.end: 64; CHECK-NEXT: ret void 65; 66entry: 67 %cmp20 = icmp eq i64 %count, 0 68 br i1 %cmp20, label %for.end, label %for.body 69 70for.body: ; preds = %entry, %for.body 71 %i.023 = phi i64 [ %inc, %for.body ], [ 0, %entry ] 72 %dst.addr.022 = phi ptr [ %add.ptr8, %for.body ], [ %dst, %entry ] 73 %src.addr.021 = phi ptr [ %add.ptr, %for.body ], [ %src, %entry ] 74 %0 = load float, ptr %src.addr.021, align 4 75 store float %0, ptr %dst.addr.022, align 4 76 %arrayidx2 = getelementptr inbounds float, ptr %src.addr.021, i64 1 77 %1 = load float, ptr %arrayidx2, align 4 78 %arrayidx3 = getelementptr inbounds float, ptr %dst.addr.022, i64 1 79 store float %1, ptr %arrayidx3, align 4 80 %arrayidx4 = getelementptr inbounds float, ptr %src.addr.021, i64 2 81 %2 = load float, ptr %arrayidx4, align 4 82 %arrayidx5 = getelementptr inbounds float, ptr %dst.addr.022, i64 2 83 store float %2, ptr %arrayidx5, align 4 84 %arrayidx6 = getelementptr inbounds float, ptr %src.addr.021, i64 3 85 %3 = load float, ptr %arrayidx6, align 4 86 %arrayidx7 = getelementptr inbounds float, ptr %dst.addr.022, i64 3 87 store float %3, ptr %arrayidx7, align 4 88 %add.ptr = getelementptr inbounds float, ptr %src.addr.021, i64 %i.023 89 %add.ptr8 = getelementptr inbounds float, ptr %dst.addr.022, i64 %i.023 90 %inc = add i64 %i.023, 1 91 %exitcond = icmp eq i64 %inc, %count 92 br i1 %exitcond, label %for.end, label %for.body 93 94for.end: ; preds = %for.body, %entry 95 ret void 96} 97 98; We do not vectorize the tiny tree which is not fully vectorizable. 99 100define void @tiny_tree_not_fully_vectorizable(ptr noalias nocapture %dst, ptr noalias nocapture readonly %src, i64 %count) #0 { 101; CHECK-LABEL: @tiny_tree_not_fully_vectorizable( 102; CHECK-NEXT: entry: 103; CHECK-NEXT: [[CMP12:%.*]] = icmp eq i64 [[COUNT:%.*]], 0 104; CHECK-NEXT: br i1 [[CMP12]], label [[FOR_END:%.*]], label [[FOR_BODY:%.*]] 105; CHECK: for.body: 106; CHECK-NEXT: [[I_015:%.*]] = phi i64 [ [[INC:%.*]], [[FOR_BODY]] ], [ 0, [[ENTRY:%.*]] ] 107; CHECK-NEXT: [[DST_ADDR_014:%.*]] = phi ptr [ [[ADD_PTR4:%.*]], [[FOR_BODY]] ], [ [[DST:%.*]], [[ENTRY]] ] 108; CHECK-NEXT: [[SRC_ADDR_013:%.*]] = phi ptr [ [[ADD_PTR:%.*]], [[FOR_BODY]] ], [ [[SRC:%.*]], [[ENTRY]] ] 109; CHECK-NEXT: [[TMP0:%.*]] = load double, ptr [[SRC_ADDR_013]], align 8 110; CHECK-NEXT: store double [[TMP0]], ptr [[DST_ADDR_014]], align 8 111; CHECK-NEXT: [[ARRAYIDX2:%.*]] = getelementptr inbounds double, ptr [[SRC_ADDR_013]], i64 2 112; CHECK-NEXT: [[TMP1:%.*]] = load double, ptr [[ARRAYIDX2]], align 8 113; CHECK-NEXT: [[ARRAYIDX3:%.*]] = getelementptr inbounds double, ptr [[DST_ADDR_014]], i64 1 114; CHECK-NEXT: store double [[TMP1]], ptr [[ARRAYIDX3]], align 8 115; CHECK-NEXT: [[ADD_PTR]] = getelementptr inbounds double, ptr [[SRC_ADDR_013]], i64 [[I_015]] 116; CHECK-NEXT: [[ADD_PTR4]] = getelementptr inbounds double, ptr [[DST_ADDR_014]], i64 [[I_015]] 117; CHECK-NEXT: [[INC]] = add i64 [[I_015]], 1 118; CHECK-NEXT: [[EXITCOND:%.*]] = icmp eq i64 [[INC]], [[COUNT]] 119; CHECK-NEXT: br i1 [[EXITCOND]], label [[FOR_END]], label [[FOR_BODY]] 120; CHECK: for.end: 121; CHECK-NEXT: ret void 122; 123entry: 124 %cmp12 = icmp eq i64 %count, 0 125 br i1 %cmp12, label %for.end, label %for.body 126 127for.body: ; preds = %entry, %for.body 128 %i.015 = phi i64 [ %inc, %for.body ], [ 0, %entry ] 129 %dst.addr.014 = phi ptr [ %add.ptr4, %for.body ], [ %dst, %entry ] 130 %src.addr.013 = phi ptr [ %add.ptr, %for.body ], [ %src, %entry ] 131 %0 = load double, ptr %src.addr.013, align 8 132 store double %0, ptr %dst.addr.014, align 8 133 %arrayidx2 = getelementptr inbounds double, ptr %src.addr.013, i64 2 134 %1 = load double, ptr %arrayidx2, align 8 135 %arrayidx3 = getelementptr inbounds double, ptr %dst.addr.014, i64 1 136 store double %1, ptr %arrayidx3, align 8 137 %add.ptr = getelementptr inbounds double, ptr %src.addr.013, i64 %i.015 138 %add.ptr4 = getelementptr inbounds double, ptr %dst.addr.014, i64 %i.015 139 %inc = add i64 %i.015, 1 140 %exitcond = icmp eq i64 %inc, %count 141 br i1 %exitcond, label %for.end, label %for.body 142 143for.end: ; preds = %for.body, %entry 144 ret void 145} 146 147define void @tiny_tree_not_fully_vectorizable2(ptr noalias nocapture %dst, ptr noalias nocapture readonly %src, i64 %count) #0 { 148; CHECK-LABEL: @tiny_tree_not_fully_vectorizable2( 149; CHECK-NEXT: entry: 150; CHECK-NEXT: [[CMP20:%.*]] = icmp eq i64 [[COUNT:%.*]], 0 151; CHECK-NEXT: br i1 [[CMP20]], label [[FOR_END:%.*]], label [[FOR_BODY:%.*]] 152; CHECK: for.body: 153; CHECK-NEXT: [[I_023:%.*]] = phi i64 [ [[INC:%.*]], [[FOR_BODY]] ], [ 0, [[ENTRY:%.*]] ] 154; CHECK-NEXT: [[DST_ADDR_022:%.*]] = phi ptr [ [[ADD_PTR8:%.*]], [[FOR_BODY]] ], [ [[DST:%.*]], [[ENTRY]] ] 155; CHECK-NEXT: [[SRC_ADDR_021:%.*]] = phi ptr [ [[ADD_PTR:%.*]], [[FOR_BODY]] ], [ [[SRC:%.*]], [[ENTRY]] ] 156; CHECK-NEXT: [[TMP0:%.*]] = load float, ptr [[SRC_ADDR_021]], align 4 157; CHECK-NEXT: [[ARRAYIDX2:%.*]] = getelementptr inbounds float, ptr [[SRC_ADDR_021]], i64 4 158; CHECK-NEXT: [[TMP1:%.*]] = load float, ptr [[ARRAYIDX2]], align 4 159; CHECK-NEXT: [[ARRAYIDX4:%.*]] = getelementptr inbounds float, ptr [[SRC_ADDR_021]], i64 2 160; CHECK-NEXT: [[TMP2:%.*]] = load <2 x float>, ptr [[ARRAYIDX4]], align 4 161; CHECK-NEXT: [[TMP3:%.*]] = insertelement <4 x float> poison, float [[TMP0]], i32 0 162; CHECK-NEXT: [[TMP4:%.*]] = insertelement <4 x float> [[TMP3]], float [[TMP1]], i32 1 163; CHECK-NEXT: [[TMP5:%.*]] = call <4 x float> @llvm.vector.insert.v4f32.v2f32(<4 x float> [[TMP4]], <2 x float> [[TMP2]], i64 2) 164; CHECK-NEXT: store <4 x float> [[TMP5]], ptr [[DST_ADDR_022]], align 4 165; CHECK-NEXT: [[ADD_PTR]] = getelementptr inbounds float, ptr [[SRC_ADDR_021]], i64 [[I_023]] 166; CHECK-NEXT: [[ADD_PTR8]] = getelementptr inbounds float, ptr [[DST_ADDR_022]], i64 [[I_023]] 167; CHECK-NEXT: [[INC]] = add i64 [[I_023]], 1 168; CHECK-NEXT: [[EXITCOND:%.*]] = icmp eq i64 [[INC]], [[COUNT]] 169; CHECK-NEXT: br i1 [[EXITCOND]], label [[FOR_END]], label [[FOR_BODY]] 170; CHECK: for.end: 171; CHECK-NEXT: ret void 172; 173entry: 174 %cmp20 = icmp eq i64 %count, 0 175 br i1 %cmp20, label %for.end, label %for.body 176 177for.body: ; preds = %entry, %for.body 178 %i.023 = phi i64 [ %inc, %for.body ], [ 0, %entry ] 179 %dst.addr.022 = phi ptr [ %add.ptr8, %for.body ], [ %dst, %entry ] 180 %src.addr.021 = phi ptr [ %add.ptr, %for.body ], [ %src, %entry ] 181 %0 = load float, ptr %src.addr.021, align 4 182 store float %0, ptr %dst.addr.022, align 4 183 %arrayidx2 = getelementptr inbounds float, ptr %src.addr.021, i64 4 184 %1 = load float, ptr %arrayidx2, align 4 185 %arrayidx3 = getelementptr inbounds float, ptr %dst.addr.022, i64 1 186 store float %1, ptr %arrayidx3, align 4 187 %arrayidx4 = getelementptr inbounds float, ptr %src.addr.021, i64 2 188 %2 = load float, ptr %arrayidx4, align 4 189 %arrayidx5 = getelementptr inbounds float, ptr %dst.addr.022, i64 2 190 store float %2, ptr %arrayidx5, align 4 191 %arrayidx6 = getelementptr inbounds float, ptr %src.addr.021, i64 3 192 %3 = load float, ptr %arrayidx6, align 4 193 %arrayidx7 = getelementptr inbounds float, ptr %dst.addr.022, i64 3 194 store float %3, ptr %arrayidx7, align 4 195 %add.ptr = getelementptr inbounds float, ptr %src.addr.021, i64 %i.023 196 %add.ptr8 = getelementptr inbounds float, ptr %dst.addr.022, i64 %i.023 197 %inc = add i64 %i.023, 1 198 %exitcond = icmp eq i64 %inc, %count 199 br i1 %exitcond, label %for.end, label %for.body 200 201for.end: ; preds = %for.body, %entry 202 ret void 203} 204 205define void @store_splat(ptr, float) { 206; CHECK-LABEL: @store_splat( 207; CHECK-NEXT: [[TMP3:%.*]] = insertelement <4 x float> poison, float [[TMP1:%.*]], i32 0 208; CHECK-NEXT: [[TMP4:%.*]] = shufflevector <4 x float> [[TMP3]], <4 x float> poison, <4 x i32> zeroinitializer 209; CHECK-NEXT: store <4 x float> [[TMP4]], ptr [[TMP0:%.*]], align 4 210; CHECK-NEXT: ret void 211; 212 store float %1, ptr %0, align 4 213 %3 = getelementptr inbounds float, ptr %0, i64 1 214 store float %1, ptr %3, align 4 215 %4 = getelementptr inbounds float, ptr %0, i64 2 216 store float %1, ptr %4, align 4 217 %5 = getelementptr inbounds float, ptr %0, i64 3 218 store float %1, ptr %5, align 4 219 ret void 220} 221 222define void @store_const(ptr %a) { 223; CHECK-LABEL: @store_const( 224; CHECK-NEXT: entry: 225; CHECK-NEXT: store <4 x i32> <i32 10, i32 30, i32 20, i32 40>, ptr [[A:%.*]], align 4 226; CHECK-NEXT: ret void 227; 228entry: 229 store i32 10, ptr %a, align 4 230 %ptr1 = getelementptr inbounds i32, ptr %a, i64 1 231 store i32 30, ptr %ptr1, align 4 232 %ptr2 = getelementptr inbounds i32, ptr %a, i64 2 233 store i32 20, ptr %ptr2, align 4 234 %ptr3 = getelementptr inbounds i32, ptr %a, i64 3 235 store i32 40, ptr %ptr3, align 4 236 ret void 237} 238 239define void @tiny_vector_gather(ptr %a, ptr %v1, ptr %v2) { 240; CHECK-LABEL: @tiny_vector_gather( 241; CHECK-NEXT: [[TMP1:%.*]] = load i32, ptr [[V1:%.*]], align 4 242; CHECK-NEXT: [[TMP2:%.*]] = load i32, ptr [[V2:%.*]], align 4 243; CHECK-NEXT: [[TMP3:%.*]] = insertelement <8 x i32> poison, i32 [[TMP1]], i32 0 244; CHECK-NEXT: [[TMP4:%.*]] = insertelement <8 x i32> [[TMP3]], i32 [[TMP2]], i32 1 245; CHECK-NEXT: [[TMP5:%.*]] = shufflevector <8 x i32> [[TMP4]], <8 x i32> poison, <8 x i32> <i32 0, i32 1, i32 0, i32 1, i32 0, i32 1, i32 0, i32 1> 246; CHECK-NEXT: store <8 x i32> [[TMP5]], ptr [[A:%.*]], align 16 247; CHECK-NEXT: ret void 248; 249 %1 = load i32, ptr %v1, align 4 250 %2 = load i32, ptr %v2, align 4 251 store i32 %1, ptr %a, align 16 252 %ptr1 = getelementptr inbounds i32, ptr %a, i64 1 253 store i32 %2, ptr %ptr1, align 4 254 %ptr2 = getelementptr inbounds i32, ptr %a, i64 2 255 store i32 %1, ptr %ptr2, align 8 256 %ptr3 = getelementptr inbounds i32, ptr %a, i64 3 257 store i32 %2, ptr %ptr3, align 4 258 %ptr4 = getelementptr inbounds i32, ptr %a, i64 4 259 store i32 %1, ptr %ptr4, align 16 260 %ptr5 = getelementptr inbounds i32, ptr %a, i64 5 261 store i32 %2, ptr %ptr5, align 4 262 %ptr6 = getelementptr inbounds i32, ptr %a, i64 6 263 store i32 %1, ptr %ptr6, align 8 264 %ptr7 = getelementptr inbounds i32, ptr %a, i64 7 265 store i32 %2, ptr %ptr7, align 4 266 ret void 267} 268 269define void @tiny_vector_with_diff_opcode(ptr %a, ptr %v1) { 270; CHECK-LABEL: @tiny_vector_with_diff_opcode( 271; CHECK-NEXT: [[TMP1:%.*]] = load i16, ptr [[V1:%.*]], align 4 272; CHECK-NEXT: [[TMP2:%.*]] = trunc i64 undef to i16 273; CHECK-NEXT: [[TMP3:%.*]] = insertelement <8 x i16> poison, i16 [[TMP1]], i32 0 274; CHECK-NEXT: [[TMP4:%.*]] = insertelement <8 x i16> [[TMP3]], i16 [[TMP2]], i32 1 275; CHECK-NEXT: [[TMP5:%.*]] = shufflevector <8 x i16> [[TMP4]], <8 x i16> poison, <8 x i32> <i32 0, i32 1, i32 0, i32 1, i32 0, i32 1, i32 0, i32 1> 276; CHECK-NEXT: store <8 x i16> [[TMP5]], ptr [[A:%.*]], align 16 277; CHECK-NEXT: ret void 278; 279 %1 = load i16, ptr %v1, align 4 280 %2 = trunc i64 undef to i16 281 store i16 %1, ptr %a, align 16 282 %ptr1 = getelementptr inbounds i16, ptr %a, i64 1 283 store i16 %2, ptr %ptr1, align 4 284 %ptr2 = getelementptr inbounds i16, ptr %a, i64 2 285 store i16 %1, ptr %ptr2, align 8 286 %ptr3 = getelementptr inbounds i16, ptr %a, i64 3 287 store i16 %2, ptr %ptr3, align 4 288 %ptr4 = getelementptr inbounds i16, ptr %a, i64 4 289 store i16 %1, ptr %ptr4, align 16 290 %ptr5 = getelementptr inbounds i16, ptr %a, i64 5 291 store i16 %2, ptr %ptr5, align 4 292 %ptr6 = getelementptr inbounds i16, ptr %a, i64 6 293 store i16 %1, ptr %ptr6, align 8 294 %ptr7 = getelementptr inbounds i16, ptr %a, i64 7 295 store i16 %2, ptr %ptr7, align 4 296 ret void 297} 298