299 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_test_checks.py2; RUN: opt < %s -passes=slp-vectorizer -S -mtriple=x86_64-apple-macosx10.8.0 -mcpu=corei7-avx | FileCheck %s3 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:%.*]], 08; 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 814; CHECK-NEXT: store <2 x double> [[TMP0]], ptr [[DST_ADDR_014]], align 815; 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]], 118; 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 void22;23entry:24 %cmp12 = icmp eq i64 %count, 025 br i1 %cmp12, label %for.end, label %for.body26 27for.body: ; preds = %entry, %for.body28 %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 832 store double %0, ptr %dst.addr.014, align 833 %arrayidx2 = getelementptr inbounds double, ptr %src.addr.013, i64 134 %1 = load double, ptr %arrayidx2, align 835 %arrayidx3 = getelementptr inbounds double, ptr %dst.addr.014, i64 136 store double %1, ptr %arrayidx3, align 837 %add.ptr = getelementptr inbounds double, ptr %src.addr.013, i64 %i.01538 %add.ptr4 = getelementptr inbounds double, ptr %dst.addr.014, i64 %i.01539 %inc = add i64 %i.015, 140 %exitcond = icmp eq i64 %inc, %count41 br i1 %exitcond, label %for.end, label %for.body42 43for.end: ; preds = %for.body, %entry44 ret void45}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:%.*]], 051; 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 457; CHECK-NEXT: store <4 x float> [[TMP0]], ptr [[DST_ADDR_022]], align 458; 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]], 161; 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 void65;66entry:67 %cmp20 = icmp eq i64 %count, 068 br i1 %cmp20, label %for.end, label %for.body69 70for.body: ; preds = %entry, %for.body71 %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 475 store float %0, ptr %dst.addr.022, align 476 %arrayidx2 = getelementptr inbounds float, ptr %src.addr.021, i64 177 %1 = load float, ptr %arrayidx2, align 478 %arrayidx3 = getelementptr inbounds float, ptr %dst.addr.022, i64 179 store float %1, ptr %arrayidx3, align 480 %arrayidx4 = getelementptr inbounds float, ptr %src.addr.021, i64 281 %2 = load float, ptr %arrayidx4, align 482 %arrayidx5 = getelementptr inbounds float, ptr %dst.addr.022, i64 283 store float %2, ptr %arrayidx5, align 484 %arrayidx6 = getelementptr inbounds float, ptr %src.addr.021, i64 385 %3 = load float, ptr %arrayidx6, align 486 %arrayidx7 = getelementptr inbounds float, ptr %dst.addr.022, i64 387 store float %3, ptr %arrayidx7, align 488 %add.ptr = getelementptr inbounds float, ptr %src.addr.021, i64 %i.02389 %add.ptr8 = getelementptr inbounds float, ptr %dst.addr.022, i64 %i.02390 %inc = add i64 %i.023, 191 %exitcond = icmp eq i64 %inc, %count92 br i1 %exitcond, label %for.end, label %for.body93 94for.end: ; preds = %for.body, %entry95 ret void96}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:%.*]], 0104; 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 8110; CHECK-NEXT: store double [[TMP0]], ptr [[DST_ADDR_014]], align 8111; CHECK-NEXT: [[ARRAYIDX2:%.*]] = getelementptr inbounds double, ptr [[SRC_ADDR_013]], i64 2112; CHECK-NEXT: [[TMP1:%.*]] = load double, ptr [[ARRAYIDX2]], align 8113; CHECK-NEXT: [[ARRAYIDX3:%.*]] = getelementptr inbounds double, ptr [[DST_ADDR_014]], i64 1114; CHECK-NEXT: store double [[TMP1]], ptr [[ARRAYIDX3]], align 8115; 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]], 1118; 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 void122;123entry:124 %cmp12 = icmp eq i64 %count, 0125 br i1 %cmp12, label %for.end, label %for.body126 127for.body: ; preds = %entry, %for.body128 %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 8132 store double %0, ptr %dst.addr.014, align 8133 %arrayidx2 = getelementptr inbounds double, ptr %src.addr.013, i64 2134 %1 = load double, ptr %arrayidx2, align 8135 %arrayidx3 = getelementptr inbounds double, ptr %dst.addr.014, i64 1136 store double %1, ptr %arrayidx3, align 8137 %add.ptr = getelementptr inbounds double, ptr %src.addr.013, i64 %i.015138 %add.ptr4 = getelementptr inbounds double, ptr %dst.addr.014, i64 %i.015139 %inc = add i64 %i.015, 1140 %exitcond = icmp eq i64 %inc, %count141 br i1 %exitcond, label %for.end, label %for.body142 143for.end: ; preds = %for.body, %entry144 ret void145}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:%.*]], 0151; 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 4157; CHECK-NEXT: [[ARRAYIDX2:%.*]] = getelementptr inbounds float, ptr [[SRC_ADDR_021]], i64 4158; CHECK-NEXT: [[TMP1:%.*]] = load float, ptr [[ARRAYIDX2]], align 4159; CHECK-NEXT: [[ARRAYIDX4:%.*]] = getelementptr inbounds float, ptr [[SRC_ADDR_021]], i64 2160; CHECK-NEXT: [[TMP2:%.*]] = load <2 x float>, ptr [[ARRAYIDX4]], align 4161; CHECK-NEXT: [[TMP3:%.*]] = insertelement <4 x float> poison, float [[TMP0]], i32 0162; CHECK-NEXT: [[TMP4:%.*]] = insertelement <4 x float> [[TMP3]], float [[TMP1]], i32 1163; CHECK-NEXT: [[TMP6:%.*]] = shufflevector <2 x float> [[TMP2]], <2 x float> poison, <4 x i32> <i32 0, i32 1, i32 poison, i32 poison>164; CHECK-NEXT: [[TMP5:%.*]] = shufflevector <4 x float> [[TMP4]], <4 x float> [[TMP6]], <4 x i32> <i32 0, i32 1, i32 4, i32 5>165; CHECK-NEXT: store <4 x float> [[TMP5]], ptr [[DST_ADDR_022]], align 4166; CHECK-NEXT: [[ADD_PTR]] = getelementptr inbounds float, ptr [[SRC_ADDR_021]], i64 [[I_023]]167; CHECK-NEXT: [[ADD_PTR8]] = getelementptr inbounds float, ptr [[DST_ADDR_022]], i64 [[I_023]]168; CHECK-NEXT: [[INC]] = add i64 [[I_023]], 1169; CHECK-NEXT: [[EXITCOND:%.*]] = icmp eq i64 [[INC]], [[COUNT]]170; CHECK-NEXT: br i1 [[EXITCOND]], label [[FOR_END]], label [[FOR_BODY]]171; CHECK: for.end:172; CHECK-NEXT: ret void173;174entry:175 %cmp20 = icmp eq i64 %count, 0176 br i1 %cmp20, label %for.end, label %for.body177 178for.body: ; preds = %entry, %for.body179 %i.023 = phi i64 [ %inc, %for.body ], [ 0, %entry ]180 %dst.addr.022 = phi ptr [ %add.ptr8, %for.body ], [ %dst, %entry ]181 %src.addr.021 = phi ptr [ %add.ptr, %for.body ], [ %src, %entry ]182 %0 = load float, ptr %src.addr.021, align 4183 store float %0, ptr %dst.addr.022, align 4184 %arrayidx2 = getelementptr inbounds float, ptr %src.addr.021, i64 4185 %1 = load float, ptr %arrayidx2, align 4186 %arrayidx3 = getelementptr inbounds float, ptr %dst.addr.022, i64 1187 store float %1, ptr %arrayidx3, align 4188 %arrayidx4 = getelementptr inbounds float, ptr %src.addr.021, i64 2189 %2 = load float, ptr %arrayidx4, align 4190 %arrayidx5 = getelementptr inbounds float, ptr %dst.addr.022, i64 2191 store float %2, ptr %arrayidx5, align 4192 %arrayidx6 = getelementptr inbounds float, ptr %src.addr.021, i64 3193 %3 = load float, ptr %arrayidx6, align 4194 %arrayidx7 = getelementptr inbounds float, ptr %dst.addr.022, i64 3195 store float %3, ptr %arrayidx7, align 4196 %add.ptr = getelementptr inbounds float, ptr %src.addr.021, i64 %i.023197 %add.ptr8 = getelementptr inbounds float, ptr %dst.addr.022, i64 %i.023198 %inc = add i64 %i.023, 1199 %exitcond = icmp eq i64 %inc, %count200 br i1 %exitcond, label %for.end, label %for.body201 202for.end: ; preds = %for.body, %entry203 ret void204}205 206define void @store_splat(ptr, float) {207; CHECK-LABEL: @store_splat(208; CHECK-NEXT: [[TMP3:%.*]] = insertelement <4 x float> poison, float [[TMP1:%.*]], i32 0209; CHECK-NEXT: [[TMP4:%.*]] = shufflevector <4 x float> [[TMP3]], <4 x float> poison, <4 x i32> zeroinitializer210; CHECK-NEXT: store <4 x float> [[TMP4]], ptr [[TMP0:%.*]], align 4211; CHECK-NEXT: ret void212;213 store float %1, ptr %0, align 4214 %3 = getelementptr inbounds float, ptr %0, i64 1215 store float %1, ptr %3, align 4216 %4 = getelementptr inbounds float, ptr %0, i64 2217 store float %1, ptr %4, align 4218 %5 = getelementptr inbounds float, ptr %0, i64 3219 store float %1, ptr %5, align 4220 ret void221}222 223define void @store_const(ptr %a) {224; CHECK-LABEL: @store_const(225; CHECK-NEXT: entry:226; CHECK-NEXT: store <4 x i32> <i32 10, i32 30, i32 20, i32 40>, ptr [[A:%.*]], align 4227; CHECK-NEXT: ret void228;229entry:230 store i32 10, ptr %a, align 4231 %ptr1 = getelementptr inbounds i32, ptr %a, i64 1232 store i32 30, ptr %ptr1, align 4233 %ptr2 = getelementptr inbounds i32, ptr %a, i64 2234 store i32 20, ptr %ptr2, align 4235 %ptr3 = getelementptr inbounds i32, ptr %a, i64 3236 store i32 40, ptr %ptr3, align 4237 ret void238}239 240define void @tiny_vector_gather(ptr %a, ptr %v1, ptr %v2) {241; CHECK-LABEL: @tiny_vector_gather(242; CHECK-NEXT: [[TMP1:%.*]] = load i32, ptr [[V1:%.*]], align 4243; CHECK-NEXT: [[TMP2:%.*]] = load i32, ptr [[V2:%.*]], align 4244; CHECK-NEXT: [[TMP3:%.*]] = insertelement <8 x i32> poison, i32 [[TMP1]], i32 0245; CHECK-NEXT: [[TMP4:%.*]] = insertelement <8 x i32> [[TMP3]], i32 [[TMP2]], i32 1246; 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>247; CHECK-NEXT: store <8 x i32> [[TMP5]], ptr [[A:%.*]], align 16248; CHECK-NEXT: ret void249;250 %1 = load i32, ptr %v1, align 4251 %2 = load i32, ptr %v2, align 4252 store i32 %1, ptr %a, align 16253 %ptr1 = getelementptr inbounds i32, ptr %a, i64 1254 store i32 %2, ptr %ptr1, align 4255 %ptr2 = getelementptr inbounds i32, ptr %a, i64 2256 store i32 %1, ptr %ptr2, align 8257 %ptr3 = getelementptr inbounds i32, ptr %a, i64 3258 store i32 %2, ptr %ptr3, align 4259 %ptr4 = getelementptr inbounds i32, ptr %a, i64 4260 store i32 %1, ptr %ptr4, align 16261 %ptr5 = getelementptr inbounds i32, ptr %a, i64 5262 store i32 %2, ptr %ptr5, align 4263 %ptr6 = getelementptr inbounds i32, ptr %a, i64 6264 store i32 %1, ptr %ptr6, align 8265 %ptr7 = getelementptr inbounds i32, ptr %a, i64 7266 store i32 %2, ptr %ptr7, align 4267 ret void268}269 270define void @tiny_vector_with_diff_opcode(ptr %a, ptr %v1) {271; CHECK-LABEL: @tiny_vector_with_diff_opcode(272; CHECK-NEXT: [[TMP1:%.*]] = load i16, ptr [[V1:%.*]], align 4273; CHECK-NEXT: [[TMP2:%.*]] = trunc i64 undef to i16274; CHECK-NEXT: [[TMP3:%.*]] = insertelement <8 x i16> poison, i16 [[TMP1]], i32 0275; CHECK-NEXT: [[TMP4:%.*]] = insertelement <8 x i16> [[TMP3]], i16 [[TMP2]], i32 1276; 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>277; CHECK-NEXT: store <8 x i16> [[TMP5]], ptr [[A:%.*]], align 16278; CHECK-NEXT: ret void279;280 %1 = load i16, ptr %v1, align 4281 %2 = trunc i64 undef to i16282 store i16 %1, ptr %a, align 16283 %ptr1 = getelementptr inbounds i16, ptr %a, i64 1284 store i16 %2, ptr %ptr1, align 4285 %ptr2 = getelementptr inbounds i16, ptr %a, i64 2286 store i16 %1, ptr %ptr2, align 8287 %ptr3 = getelementptr inbounds i16, ptr %a, i64 3288 store i16 %2, ptr %ptr3, align 4289 %ptr4 = getelementptr inbounds i16, ptr %a, i64 4290 store i16 %1, ptr %ptr4, align 16291 %ptr5 = getelementptr inbounds i16, ptr %a, i64 5292 store i16 %2, ptr %ptr5, align 4293 %ptr6 = getelementptr inbounds i16, ptr %a, i64 6294 store i16 %1, ptr %ptr6, align 8295 %ptr7 = getelementptr inbounds i16, ptr %a, i64 7296 store i16 %2, ptr %ptr7, align 4297 ret void298}299