790 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_test_checks.py2; RUN: opt -matrix-allow-contract=false -passes='lower-matrix-intrinsics' -S < %s | FileCheck %s3 4define void @matrix_phi_loop(ptr %in1, ptr %in2, i32 %count, ptr %out) {5; CHECK-LABEL: @matrix_phi_loop(6; CHECK-NEXT: entry:7; CHECK-NEXT: [[COL_LOAD:%.*]] = load <3 x double>, ptr [[IN1:%.*]], align 88; CHECK-NEXT: [[VEC_GEP:%.*]] = getelementptr double, ptr [[IN1]], i64 39; CHECK-NEXT: [[COL_LOAD1:%.*]] = load <3 x double>, ptr [[VEC_GEP]], align 810; CHECK-NEXT: [[VEC_GEP2:%.*]] = getelementptr double, ptr [[IN1]], i64 611; CHECK-NEXT: [[COL_LOAD3:%.*]] = load <3 x double>, ptr [[VEC_GEP2]], align 812; CHECK-NEXT: br label [[LOOP:%.*]]13; CHECK: loop:14; CHECK-NEXT: [[PHI9:%.*]] = phi <3 x double> [ [[COL_LOAD]], [[ENTRY:%.*]] ], [ [[TMP0:%.*]], [[LOOP]] ]15; CHECK-NEXT: [[PHI10:%.*]] = phi <3 x double> [ [[COL_LOAD1]], [[ENTRY]] ], [ [[TMP1:%.*]], [[LOOP]] ]16; CHECK-NEXT: [[PHI11:%.*]] = phi <3 x double> [ [[COL_LOAD3]], [[ENTRY]] ], [ [[TMP2:%.*]], [[LOOP]] ]17; CHECK-NEXT: [[CTR:%.*]] = phi i32 [ [[COUNT:%.*]], [[ENTRY]] ], [ [[DEC:%.*]], [[LOOP]] ]18; CHECK-NEXT: [[COL_LOAD4:%.*]] = load <3 x double>, ptr [[IN2:%.*]], align 819; CHECK-NEXT: [[VEC_GEP5:%.*]] = getelementptr double, ptr [[IN2]], i64 320; CHECK-NEXT: [[COL_LOAD6:%.*]] = load <3 x double>, ptr [[VEC_GEP5]], align 821; CHECK-NEXT: [[VEC_GEP7:%.*]] = getelementptr double, ptr [[IN2]], i64 622; CHECK-NEXT: [[COL_LOAD8:%.*]] = load <3 x double>, ptr [[VEC_GEP7]], align 823; CHECK-NEXT: [[TMP6:%.*]] = fadd <3 x double> [[PHI9]], [[COL_LOAD4]]24; CHECK-NEXT: [[TMP7:%.*]] = fadd <3 x double> [[PHI10]], [[COL_LOAD6]]25; CHECK-NEXT: [[TMP8:%.*]] = fadd <3 x double> [[PHI11]], [[COL_LOAD8]]26; CHECK-NEXT: [[TMP3:%.*]] = shufflevector <3 x double> [[TMP6]], <3 x double> [[TMP7]], <6 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5>27; CHECK-NEXT: [[TMP4:%.*]] = shufflevector <3 x double> [[TMP8]], <3 x double> poison, <6 x i32> <i32 0, i32 1, i32 2, i32 poison, i32 poison, i32 poison>28; CHECK-NEXT: [[TMP5:%.*]] = shufflevector <6 x double> [[TMP3]], <6 x double> [[TMP4]], <9 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8>29; CHECK-NEXT: [[TMP0]] = shufflevector <9 x double> [[TMP5]], <9 x double> poison, <3 x i32> <i32 0, i32 1, i32 2>30; CHECK-NEXT: [[TMP1]] = shufflevector <9 x double> [[TMP5]], <9 x double> poison, <3 x i32> <i32 3, i32 4, i32 5>31; CHECK-NEXT: [[TMP2]] = shufflevector <9 x double> [[TMP5]], <9 x double> poison, <3 x i32> <i32 6, i32 7, i32 8>32; CHECK-NEXT: [[DEC]] = sub i32 [[CTR]], 133; CHECK-NEXT: [[CMP:%.*]] = icmp eq i32 [[DEC]], 034; CHECK-NEXT: br i1 [[CMP]], label [[EXIT:%.*]], label [[LOOP]]35; CHECK: exit:36; CHECK-NEXT: store <3 x double> [[TMP6]], ptr [[OUT:%.*]], align 12837; CHECK-NEXT: [[VEC_GEP12:%.*]] = getelementptr double, ptr [[OUT]], i64 338; CHECK-NEXT: store <3 x double> [[TMP7]], ptr [[VEC_GEP12]], align 839; CHECK-NEXT: [[VEC_GEP13:%.*]] = getelementptr double, ptr [[OUT]], i64 640; CHECK-NEXT: store <3 x double> [[TMP8]], ptr [[VEC_GEP13]], align 1641; CHECK-NEXT: ret void42;43entry:44 %in1v = call <9 x double> @llvm.matrix.column.major.load(ptr %in1, i64 3, i1 false, i32 3, i32 3)45 br label %loop46 47loop:48 %phi = phi <9 x double> [%in1v, %entry], [%sum, %loop]49 %ctr = phi i32 [%count, %entry], [%dec, %loop]50 51 %in2v = call <9 x double> @llvm.matrix.column.major.load(ptr %in2, i64 3, i1 false, i32 3, i32 3)52 53 %sum = fadd <9 x double> %phi, %in2v54 55 %dec = sub i32 %ctr, 156 %cmp = icmp eq i32 %dec, 057 br i1 %cmp, label %exit, label %loop58 59exit:60 store <9 x double> %sum, ptr %out61 ret void62}63 64define void @matrix_phi_loop_zeroinitializer(ptr %in1, ptr %in2, i32 %count, ptr %out) {65; CHECK-LABEL: @matrix_phi_loop_zeroinitializer(66; CHECK-NEXT: entry:67; CHECK-NEXT: br label [[LOOP:%.*]]68; CHECK: loop:69; CHECK-NEXT: [[PHI4:%.*]] = phi <3 x double> [ zeroinitializer, [[ENTRY:%.*]] ], [ [[TMP0:%.*]], [[LOOP]] ]70; CHECK-NEXT: [[PHI5:%.*]] = phi <3 x double> [ zeroinitializer, [[ENTRY]] ], [ [[TMP1:%.*]], [[LOOP]] ]71; CHECK-NEXT: [[PHI6:%.*]] = phi <3 x double> [ zeroinitializer, [[ENTRY]] ], [ [[TMP2:%.*]], [[LOOP]] ]72; CHECK-NEXT: [[CTR:%.*]] = phi i32 [ [[COUNT:%.*]], [[ENTRY]] ], [ [[DEC:%.*]], [[LOOP]] ]73; CHECK-NEXT: [[COL_LOAD:%.*]] = load <3 x double>, ptr [[IN2:%.*]], align 874; CHECK-NEXT: [[VEC_GEP:%.*]] = getelementptr double, ptr [[IN2]], i64 375; CHECK-NEXT: [[COL_LOAD1:%.*]] = load <3 x double>, ptr [[VEC_GEP]], align 876; CHECK-NEXT: [[VEC_GEP2:%.*]] = getelementptr double, ptr [[IN2]], i64 677; CHECK-NEXT: [[COL_LOAD3:%.*]] = load <3 x double>, ptr [[VEC_GEP2]], align 878; CHECK-NEXT: [[TMP6:%.*]] = fadd <3 x double> [[PHI4]], [[COL_LOAD]]79; CHECK-NEXT: [[TMP7:%.*]] = fadd <3 x double> [[PHI5]], [[COL_LOAD1]]80; CHECK-NEXT: [[TMP8:%.*]] = fadd <3 x double> [[PHI6]], [[COL_LOAD3]]81; CHECK-NEXT: [[TMP3:%.*]] = shufflevector <3 x double> [[TMP6]], <3 x double> [[TMP7]], <6 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5>82; CHECK-NEXT: [[TMP4:%.*]] = shufflevector <3 x double> [[TMP8]], <3 x double> poison, <6 x i32> <i32 0, i32 1, i32 2, i32 poison, i32 poison, i32 poison>83; CHECK-NEXT: [[TMP5:%.*]] = shufflevector <6 x double> [[TMP3]], <6 x double> [[TMP4]], <9 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8>84; CHECK-NEXT: [[TMP0]] = shufflevector <9 x double> [[TMP5]], <9 x double> poison, <3 x i32> <i32 0, i32 1, i32 2>85; CHECK-NEXT: [[TMP1]] = shufflevector <9 x double> [[TMP5]], <9 x double> poison, <3 x i32> <i32 3, i32 4, i32 5>86; CHECK-NEXT: [[TMP2]] = shufflevector <9 x double> [[TMP5]], <9 x double> poison, <3 x i32> <i32 6, i32 7, i32 8>87; CHECK-NEXT: [[DEC]] = sub i32 [[CTR]], 188; CHECK-NEXT: [[CMP:%.*]] = icmp eq i32 [[DEC]], 089; CHECK-NEXT: br i1 [[CMP]], label [[EXIT:%.*]], label [[LOOP]]90; CHECK: exit:91; CHECK-NEXT: store <3 x double> [[TMP6]], ptr [[OUT:%.*]], align 12892; CHECK-NEXT: [[VEC_GEP7:%.*]] = getelementptr double, ptr [[OUT]], i64 393; CHECK-NEXT: store <3 x double> [[TMP7]], ptr [[VEC_GEP7]], align 894; CHECK-NEXT: [[VEC_GEP8:%.*]] = getelementptr double, ptr [[OUT]], i64 695; CHECK-NEXT: store <3 x double> [[TMP8]], ptr [[VEC_GEP8]], align 1696; CHECK-NEXT: ret void97;98entry:99 br label %loop100 101loop:102 %phi = phi <9 x double> [zeroinitializer, %entry], [%sum, %loop]103 %ctr = phi i32 [%count, %entry], [%dec, %loop]104 105 %inv = call <9 x double> @llvm.matrix.column.major.load(ptr %in2, i64 3, i1 false, i32 3, i32 3)106 107 %sum = fadd <9 x double> %phi, %inv108 109 %dec = sub i32 %ctr, 1110 %cmp = icmp eq i32 %dec, 0111 br i1 %cmp, label %exit, label %loop112 113exit:114 store <9 x double> %sum, ptr %out115 ret void116}117 118define void @matrix_phi_loop_undef(ptr %in, i32 %count, ptr %out) {119; CHECK-LABEL: @matrix_phi_loop_undef(120; CHECK-NEXT: entry:121; CHECK-NEXT: br label [[LOOP:%.*]]122; CHECK: loop:123; CHECK-NEXT: [[PHI:%.*]] = phi <9 x double> [ undef, [[ENTRY:%.*]] ], [ [[SUM:%.*]], [[LOOP]] ]124; CHECK-NEXT: [[CTR:%.*]] = phi i32 [ [[COUNT:%.*]], [[ENTRY]] ], [ [[DEC:%.*]], [[LOOP]] ]125; CHECK-NEXT: [[INV:%.*]] = load <9 x double>, ptr [[IN:%.*]], align 128126; CHECK-NEXT: [[SUM]] = fadd <9 x double> [[PHI]], [[INV]]127; CHECK-NEXT: [[DEC]] = sub i32 [[CTR]], 1128; CHECK-NEXT: [[CMP:%.*]] = icmp eq i32 [[DEC]], 0129; CHECK-NEXT: br i1 [[CMP]], label [[EXIT:%.*]], label [[LOOP]]130; CHECK: exit:131; CHECK-NEXT: store <9 x double> [[SUM]], ptr [[OUT:%.*]], align 128132; CHECK-NEXT: ret void133;134entry:135 br label %loop136 137loop:138 %phi = phi <9 x double> [undef, %entry], [%sum, %loop]139 %ctr = phi i32 [%count, %entry], [%dec, %loop]140 141 %inv = load <9 x double>, ptr %in142 143 %sum = fadd <9 x double> %phi, %inv144 145 %dec = sub i32 %ctr, 1146 %cmp = icmp eq i32 %dec, 0147 br i1 %cmp, label %exit, label %loop148 149exit:150 store <9 x double> %sum, ptr %out151 ret void152}153 154define void @matrix_phi_loop_poison(ptr %in, i32 %count, ptr %out) {155; CHECK-LABEL: @matrix_phi_loop_poison(156; CHECK-NEXT: entry:157; CHECK-NEXT: br label [[LOOP:%.*]]158; CHECK: loop:159; CHECK-NEXT: [[PHI4:%.*]] = phi <3 x double> [ poison, [[ENTRY:%.*]] ], [ [[TMP0:%.*]], [[LOOP]] ]160; CHECK-NEXT: [[PHI5:%.*]] = phi <3 x double> [ poison, [[ENTRY]] ], [ [[TMP1:%.*]], [[LOOP]] ]161; CHECK-NEXT: [[PHI6:%.*]] = phi <3 x double> [ poison, [[ENTRY]] ], [ [[TMP2:%.*]], [[LOOP]] ]162; CHECK-NEXT: [[CTR:%.*]] = phi i32 [ [[COUNT:%.*]], [[ENTRY]] ], [ [[DEC:%.*]], [[LOOP]] ]163; CHECK-NEXT: [[COL_LOAD:%.*]] = load <3 x double>, ptr [[IN2:%.*]], align 8164; CHECK-NEXT: [[VEC_GEP:%.*]] = getelementptr double, ptr [[IN2]], i64 3165; CHECK-NEXT: [[COL_LOAD1:%.*]] = load <3 x double>, ptr [[VEC_GEP]], align 8166; CHECK-NEXT: [[VEC_GEP2:%.*]] = getelementptr double, ptr [[IN2]], i64 6167; CHECK-NEXT: [[COL_LOAD3:%.*]] = load <3 x double>, ptr [[VEC_GEP2]], align 8168; CHECK-NEXT: [[TMP6:%.*]] = fadd <3 x double> [[PHI4]], [[COL_LOAD]]169; CHECK-NEXT: [[TMP7:%.*]] = fadd <3 x double> [[PHI5]], [[COL_LOAD1]]170; CHECK-NEXT: [[TMP8:%.*]] = fadd <3 x double> [[PHI6]], [[COL_LOAD3]]171; CHECK-NEXT: [[TMP3:%.*]] = shufflevector <3 x double> [[TMP6]], <3 x double> [[TMP7]], <6 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5>172; CHECK-NEXT: [[TMP4:%.*]] = shufflevector <3 x double> [[TMP8]], <3 x double> poison, <6 x i32> <i32 0, i32 1, i32 2, i32 poison, i32 poison, i32 poison>173; CHECK-NEXT: [[TMP5:%.*]] = shufflevector <6 x double> [[TMP3]], <6 x double> [[TMP4]], <9 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8>174; CHECK-NEXT: [[TMP0]] = shufflevector <9 x double> [[TMP5]], <9 x double> poison, <3 x i32> <i32 0, i32 1, i32 2>175; CHECK-NEXT: [[TMP1]] = shufflevector <9 x double> [[TMP5]], <9 x double> poison, <3 x i32> <i32 3, i32 4, i32 5>176; CHECK-NEXT: [[TMP2]] = shufflevector <9 x double> [[TMP5]], <9 x double> poison, <3 x i32> <i32 6, i32 7, i32 8>177; CHECK-NEXT: [[DEC]] = sub i32 [[CTR]], 1178; CHECK-NEXT: [[CMP:%.*]] = icmp eq i32 [[DEC]], 0179; CHECK-NEXT: br i1 [[CMP]], label [[EXIT:%.*]], label [[LOOP]]180; CHECK: exit:181; CHECK-NEXT: store <3 x double> [[TMP6]], ptr [[OUT:%.*]], align 128182; CHECK-NEXT: [[VEC_GEP7:%.*]] = getelementptr double, ptr [[OUT]], i64 3183; CHECK-NEXT: store <3 x double> [[TMP7]], ptr [[VEC_GEP7]], align 8184; CHECK-NEXT: [[VEC_GEP8:%.*]] = getelementptr double, ptr [[OUT]], i64 6185; CHECK-NEXT: store <3 x double> [[TMP8]], ptr [[VEC_GEP8]], align 16186; CHECK-NEXT: ret void187;188entry:189 br label %loop190 191loop:192 %phi = phi <9 x double> [poison, %entry], [%sum, %loop]193 %ctr = phi i32 [%count, %entry], [%dec, %loop]194 195 %inv = call <9 x double> @llvm.matrix.column.major.load(ptr %in, i64 3, i1 false, i32 3, i32 3)196 197 %sum = fadd <9 x double> %phi, %inv198 199 %dec = sub i32 %ctr, 1200 %cmp = icmp eq i32 %dec, 0201 br i1 %cmp, label %exit, label %loop202 203exit:204 store <9 x double> %sum, ptr %out205 ret void206}207 208define void @matrix_phi_loop_cdv(ptr %in, i32 %count, ptr %out) {209; CHECK-LABEL: @matrix_phi_loop_cdv(210; CHECK-NEXT: entry:211; CHECK-NEXT: br label [[LOOP:%.*]]212; CHECK: loop:213; CHECK-NEXT: [[PHI4:%.*]] = phi <3 x double> [ <double 0.000000e+00, double 1.000000e+00, double 2.000000e+00>, [[ENTRY:%.*]] ], [ [[TMP0:%.*]], [[LOOP]] ]214; CHECK-NEXT: [[PHI5:%.*]] = phi <3 x double> [ <double 3.000000e+00, double 4.000000e+00, double 5.000000e+00>, [[ENTRY]] ], [ [[TMP1:%.*]], [[LOOP]] ]215; CHECK-NEXT: [[PHI6:%.*]] = phi <3 x double> [ <double 6.000000e+00, double 7.000000e+00, double 8.000000e+00>, [[ENTRY]] ], [ [[TMP2:%.*]], [[LOOP]] ]216; CHECK-NEXT: [[CTR:%.*]] = phi i32 [ [[COUNT:%.*]], [[ENTRY]] ], [ [[DEC:%.*]], [[LOOP]] ]217; CHECK-NEXT: [[COL_LOAD:%.*]] = load <3 x double>, ptr [[IN2:%.*]], align 8218; CHECK-NEXT: [[VEC_GEP:%.*]] = getelementptr double, ptr [[IN2]], i64 3219; CHECK-NEXT: [[COL_LOAD1:%.*]] = load <3 x double>, ptr [[VEC_GEP]], align 8220; CHECK-NEXT: [[VEC_GEP2:%.*]] = getelementptr double, ptr [[IN2]], i64 6221; CHECK-NEXT: [[COL_LOAD3:%.*]] = load <3 x double>, ptr [[VEC_GEP2]], align 8222; CHECK-NEXT: [[TMP6:%.*]] = fadd <3 x double> [[PHI4]], [[COL_LOAD]]223; CHECK-NEXT: [[TMP7:%.*]] = fadd <3 x double> [[PHI5]], [[COL_LOAD1]]224; CHECK-NEXT: [[TMP8:%.*]] = fadd <3 x double> [[PHI6]], [[COL_LOAD3]]225; CHECK-NEXT: [[TMP3:%.*]] = shufflevector <3 x double> [[TMP6]], <3 x double> [[TMP7]], <6 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5>226; CHECK-NEXT: [[TMP4:%.*]] = shufflevector <3 x double> [[TMP8]], <3 x double> poison, <6 x i32> <i32 0, i32 1, i32 2, i32 poison, i32 poison, i32 poison>227; CHECK-NEXT: [[TMP5:%.*]] = shufflevector <6 x double> [[TMP3]], <6 x double> [[TMP4]], <9 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8>228; CHECK-NEXT: [[TMP0]] = shufflevector <9 x double> [[TMP5]], <9 x double> poison, <3 x i32> <i32 0, i32 1, i32 2>229; CHECK-NEXT: [[TMP1]] = shufflevector <9 x double> [[TMP5]], <9 x double> poison, <3 x i32> <i32 3, i32 4, i32 5>230; CHECK-NEXT: [[TMP2]] = shufflevector <9 x double> [[TMP5]], <9 x double> poison, <3 x i32> <i32 6, i32 7, i32 8>231; CHECK-NEXT: [[DEC]] = sub i32 [[CTR]], 1232; CHECK-NEXT: [[CMP:%.*]] = icmp eq i32 [[DEC]], 0233; CHECK-NEXT: br i1 [[CMP]], label [[EXIT:%.*]], label [[LOOP]]234; CHECK: exit:235; CHECK-NEXT: store <3 x double> [[TMP6]], ptr [[OUT:%.*]], align 128236; CHECK-NEXT: [[VEC_GEP7:%.*]] = getelementptr double, ptr [[OUT]], i64 3237; CHECK-NEXT: store <3 x double> [[TMP7]], ptr [[VEC_GEP7]], align 8238; CHECK-NEXT: [[VEC_GEP8:%.*]] = getelementptr double, ptr [[OUT]], i64 6239; CHECK-NEXT: store <3 x double> [[TMP8]], ptr [[VEC_GEP8]], align 16240; CHECK-NEXT: ret void241;242entry:243 br label %loop244 245loop:246 %phi = phi <9 x double> [<double 0., double 1., double 2., double 3., double 4., double 5., double 6., double 7., double 8.>, %entry], [%sum, %loop]247 %ctr = phi i32 [%count, %entry], [%dec, %loop]248 249 %inv = call <9 x double> @llvm.matrix.column.major.load(ptr %in, i64 3, i1 false, i32 3, i32 3)250 251 %sum = fadd <9 x double> %phi, %inv252 253 %dec = sub i32 %ctr, 1254 %cmp = icmp eq i32 %dec, 0255 br i1 %cmp, label %exit, label %loop256 257exit:258 store <9 x double> %sum, ptr %out259 ret void260}261 262define void @matrix_phi_loop_delay(ptr %in1, ptr %in2, i32 %count, ptr %out) {263; CHECK-LABEL: @matrix_phi_loop_delay(264; CHECK-NEXT: entry:265; CHECK-NEXT: [[COL_LOAD:%.*]] = load <3 x double>, ptr [[IN1:%.*]], align 8266; CHECK-NEXT: [[VEC_GEP:%.*]] = getelementptr double, ptr [[IN1]], i64 3267; CHECK-NEXT: [[COL_LOAD1:%.*]] = load <3 x double>, ptr [[VEC_GEP]], align 8268; CHECK-NEXT: [[VEC_GEP2:%.*]] = getelementptr double, ptr [[IN1]], i64 6269; CHECK-NEXT: [[COL_LOAD3:%.*]] = load <3 x double>, ptr [[VEC_GEP2]], align 8270; CHECK-NEXT: br label [[LOOP:%.*]]271; CHECK: loop:272; CHECK-NEXT: [[PHI14:%.*]] = phi <3 x double> [ [[COL_LOAD]], [[ENTRY:%.*]] ], [ [[TMP0:%.*]], [[LOOP]] ]273; CHECK-NEXT: [[PHI15:%.*]] = phi <3 x double> [ [[COL_LOAD1]], [[ENTRY]] ], [ [[TMP1:%.*]], [[LOOP]] ]274; CHECK-NEXT: [[PHI16:%.*]] = phi <3 x double> [ [[COL_LOAD3]], [[ENTRY]] ], [ [[TMP2:%.*]], [[LOOP]] ]275; CHECK-NEXT: [[TMP0]] = phi <3 x double> [ [[COL_LOAD]], [[ENTRY]] ], [ [[SPLIT:%.*]], [[LOOP]] ]276; CHECK-NEXT: [[TMP1]] = phi <3 x double> [ [[COL_LOAD1]], [[ENTRY]] ], [ [[SPLIT10:%.*]], [[LOOP]] ]277; CHECK-NEXT: [[TMP2]] = phi <3 x double> [ [[COL_LOAD3]], [[ENTRY]] ], [ [[SPLIT11:%.*]], [[LOOP]] ]278; CHECK-NEXT: [[CTR:%.*]] = phi i32 [ [[COUNT:%.*]], [[ENTRY]] ], [ [[DEC:%.*]], [[LOOP]] ]279; CHECK-NEXT: [[COL_LOAD4:%.*]] = load <3 x double>, ptr [[IN2:%.*]], align 8280; CHECK-NEXT: [[VEC_GEP5:%.*]] = getelementptr double, ptr [[IN2]], i64 3281; CHECK-NEXT: [[COL_LOAD6:%.*]] = load <3 x double>, ptr [[VEC_GEP5]], align 8282; CHECK-NEXT: [[VEC_GEP7:%.*]] = getelementptr double, ptr [[IN2]], i64 6283; CHECK-NEXT: [[COL_LOAD8:%.*]] = load <3 x double>, ptr [[VEC_GEP7]], align 8284; CHECK-NEXT: [[TMP6:%.*]] = fadd <3 x double> [[PHI14]], [[COL_LOAD4]]285; CHECK-NEXT: [[TMP7:%.*]] = fadd <3 x double> [[PHI15]], [[COL_LOAD6]]286; CHECK-NEXT: [[TMP8:%.*]] = fadd <3 x double> [[PHI16]], [[COL_LOAD8]]287; CHECK-NEXT: [[TMP9:%.*]] = shufflevector <3 x double> [[TMP6]], <3 x double> [[TMP7]], <6 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5>288; CHECK-NEXT: [[TMP10:%.*]] = shufflevector <3 x double> [[TMP8]], <3 x double> poison, <6 x i32> <i32 0, i32 1, i32 2, i32 poison, i32 poison, i32 poison>289; CHECK-NEXT: [[TMP11:%.*]] = shufflevector <6 x double> [[TMP9]], <6 x double> [[TMP10]], <9 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8>290; CHECK-NEXT: [[SPLIT]] = shufflevector <9 x double> [[TMP11]], <9 x double> poison, <3 x i32> <i32 0, i32 1, i32 2>291; CHECK-NEXT: [[SPLIT10]] = shufflevector <9 x double> [[TMP11]], <9 x double> poison, <3 x i32> <i32 3, i32 4, i32 5>292; CHECK-NEXT: [[SPLIT11]] = shufflevector <9 x double> [[TMP11]], <9 x double> poison, <3 x i32> <i32 6, i32 7, i32 8>293; CHECK-NEXT: [[TMP12:%.*]] = fadd <3 x double> [[TMP6]], [[TMP0]]294; CHECK-NEXT: [[TMP13:%.*]] = fadd <3 x double> [[TMP7]], [[TMP1]]295; CHECK-NEXT: [[TMP14:%.*]] = fadd <3 x double> [[TMP8]], [[TMP2]]296; CHECK-NEXT: [[DEC]] = sub i32 [[CTR]], 1297; CHECK-NEXT: [[CMP:%.*]] = icmp eq i32 [[DEC]], 0298; CHECK-NEXT: br i1 [[CMP]], label [[EXIT:%.*]], label [[LOOP]]299; CHECK: exit:300; CHECK-NEXT: store <3 x double> [[TMP12]], ptr [[OUT:%.*]], align 128301; CHECK-NEXT: [[VEC_GEP12:%.*]] = getelementptr double, ptr [[OUT]], i64 3302; CHECK-NEXT: store <3 x double> [[TMP13]], ptr [[VEC_GEP12]], align 8303; CHECK-NEXT: [[VEC_GEP13:%.*]] = getelementptr double, ptr [[OUT]], i64 6304; CHECK-NEXT: store <3 x double> [[TMP14]], ptr [[VEC_GEP13]], align 16305; CHECK-NEXT: ret void306;307entry:308 %in1v = call <9 x double> @llvm.matrix.column.major.load(ptr %in1, i64 3, i1 false, i32 3, i32 3)309 br label %loop310 311loop:312 %phi2 = phi <9 x double> [%in1v, %entry], [%phi, %loop]313 %phi = phi <9 x double> [%in1v, %entry], [%sum, %loop]314 %ctr = phi i32 [%count, %entry], [%dec, %loop]315 316 %in2v = call <9 x double> @llvm.matrix.column.major.load(ptr %in2, i64 3, i1 false, i32 3, i32 3)317 318 %sum = fadd <9 x double> %phi2, %in2v319 %sum2 = fadd <9 x double> %sum, %phi320 321 %dec = sub i32 %ctr, 1322 %cmp = icmp eq i32 %dec, 0323 br i1 %cmp, label %exit, label %loop324 325exit:326 store <9 x double> %sum2, ptr %out327 ret void328}329 330define void @matrix_phi_loop_delay_reshape(ptr %in1, ptr %in2, ptr %in3, i32 %count, ptr %out) {331; CHECK-LABEL: @matrix_phi_loop_delay_reshape(332; CHECK-NEXT: entry:333; CHECK-NEXT: [[COL_LOAD1:%.*]] = load <2 x double>, ptr [[IN3:%.*]], align 8334; CHECK-NEXT: [[VEC_GEP2:%.*]] = getelementptr double, ptr [[IN3]], i64 2335; CHECK-NEXT: [[COL_LOAD8:%.*]] = load <2 x double>, ptr [[VEC_GEP2]], align 8336; CHECK-NEXT: [[VEC_GEP1:%.*]] = getelementptr double, ptr [[IN3]], i64 4337; CHECK-NEXT: [[COL_LOAD12:%.*]] = load <2 x double>, ptr [[VEC_GEP1]], align 8338; CHECK-NEXT: [[TMP0:%.*]] = shufflevector <2 x double> [[COL_LOAD1]], <2 x double> [[COL_LOAD8]], <4 x i32> <i32 0, i32 1, i32 2, i32 3>339; CHECK-NEXT: [[TMP1:%.*]] = shufflevector <2 x double> [[COL_LOAD12]], <2 x double> poison, <4 x i32> <i32 0, i32 1, i32 poison, i32 poison>340; CHECK-NEXT: [[TMP2:%.*]] = shufflevector <4 x double> [[TMP0]], <4 x double> [[TMP1]], <6 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5>341; CHECK-NEXT: [[SPLIT:%.*]] = shufflevector <6 x double> [[TMP2]], <6 x double> poison, <3 x i32> <i32 0, i32 1, i32 2>342; CHECK-NEXT: [[COL_LOAD10:%.*]] = shufflevector <6 x double> [[TMP2]], <6 x double> poison, <3 x i32> <i32 3, i32 4, i32 5>343; CHECK-NEXT: [[COL_LOAD11:%.*]] = load <6 x double>, ptr [[IN2:%.*]], align 8344; CHECK-NEXT: [[COL_LOAD:%.*]] = load <3 x double>, ptr [[IN1:%.*]], align 8345; CHECK-NEXT: [[VEC_GEP:%.*]] = getelementptr double, ptr [[IN1]], i64 3346; CHECK-NEXT: [[COL_LOAD14:%.*]] = load <3 x double>, ptr [[VEC_GEP]], align 8347; CHECK-NEXT: br label [[LOOP:%.*]]348; CHECK: loop:349; CHECK-NEXT: [[PHI2:%.*]] = phi <3 x double> [ [[SPLIT]], [[ENTRY:%.*]] ], [ [[PHI1:%.*]], [[LOOP]] ]350; CHECK-NEXT: [[PHI39:%.*]] = phi <3 x double> [ [[COL_LOAD10]], [[ENTRY]] ], [ [[PHI4:%.*]], [[LOOP]] ]351; CHECK-NEXT: [[PHI25:%.*]] = phi <6 x double> [ [[COL_LOAD11]], [[ENTRY]] ], [ [[PHI25]], [[LOOP]] ]352; CHECK-NEXT: [[PHI1]] = phi <3 x double> [ [[COL_LOAD]], [[ENTRY]] ], [ [[PHI2]], [[LOOP]] ]353; CHECK-NEXT: [[PHI4]] = phi <3 x double> [ [[COL_LOAD14]], [[ENTRY]] ], [ [[PHI39]], [[LOOP]] ]354; CHECK-NEXT: [[CTR:%.*]] = phi i32 [ [[COUNT:%.*]], [[ENTRY]] ], [ [[DEC:%.*]], [[LOOP]] ]355; CHECK-NEXT: [[TMP3:%.*]] = shufflevector <6 x double> [[PHI25]], <6 x double> poison, <3 x i32> <i32 0, i32 1, i32 2>356; CHECK-NEXT: [[TMP4:%.*]] = shufflevector <6 x double> [[PHI25]], <6 x double> poison, <3 x i32> <i32 3, i32 4, i32 5>357; CHECK-NEXT: [[TMP7:%.*]] = fadd <3 x double> [[TMP3]], [[PHI1]]358; CHECK-NEXT: [[TMP8:%.*]] = fadd <3 x double> [[TMP4]], [[PHI4]]359; CHECK-NEXT: [[TMP5:%.*]] = fadd <3 x double> [[TMP7]], [[PHI2]]360; CHECK-NEXT: [[TMP6:%.*]] = fadd <3 x double> [[TMP8]], [[PHI39]]361; CHECK-NEXT: [[DEC]] = sub i32 [[CTR]], 1362; CHECK-NEXT: [[CMP:%.*]] = icmp eq i32 [[DEC]], 0363; CHECK-NEXT: br i1 [[CMP]], label [[EXIT:%.*]], label [[LOOP]]364; CHECK: exit:365; CHECK-NEXT: store <3 x double> [[TMP5]], ptr [[OUT:%.*]], align 64366; CHECK-NEXT: [[VEC_GEP30:%.*]] = getelementptr double, ptr [[OUT]], i64 3367; CHECK-NEXT: store <3 x double> [[TMP6]], ptr [[VEC_GEP30]], align 8368; CHECK-NEXT: ret void369;370entry:371 %in1v = call <6 x double> @llvm.matrix.column.major.load(ptr %in3, i64 2, i1 false, i32 2, i32 3)372 %in2v = call <6 x double> @llvm.matrix.column.major.load(ptr %in2, i64 6, i1 false, i32 6, i32 1)373 %in3v = call <6 x double> @llvm.matrix.column.major.load(ptr %in1, i64 3, i1 false, i32 3, i32 2)374 br label %loop375 376loop:377 %phi = phi <6 x double> [%in1v, %entry], [%phi3, %loop]378 %phi2 = phi <6 x double> [%in2v, %entry], [%phi2, %loop]379 %phi3 = phi <6 x double> [%in3v, %entry], [%phi, %loop]380 %ctr = phi i32 [%count, %entry], [%dec, %loop]381 382 %sum = fadd <6 x double> %phi2, %phi3383 %sum2 = fadd <6 x double> %sum, %phi384 385 %dec = sub i32 %ctr, 1386 %cmp = icmp eq i32 %dec, 0387 br i1 %cmp, label %exit, label %loop388 389exit:390 store <6 x double> %sum2, ptr %out391 ret void392}393 394define void @matrix_phi_three_preds(i1 %cond1, i1 %cond2, ptr %a, ptr %b, ptr %c, ptr %out) {395; CHECK-LABEL: @matrix_phi_three_preds(396; CHECK-NEXT: entry:397; CHECK-NEXT: br i1 [[COND1:%.*]], label [[BB1:%.*]], label [[BBA:%.*]]398; CHECK: bb1:399; CHECK-NEXT: br i1 [[COND2:%.*]], label [[BBB:%.*]], label [[BBC:%.*]]400; CHECK: bba:401; CHECK-NEXT: [[COL_LOAD:%.*]] = load <3 x double>, ptr [[A:%.*]], align 8402; CHECK-NEXT: [[VEC_GEP:%.*]] = getelementptr double, ptr [[A]], i64 3403; CHECK-NEXT: [[COL_LOAD1:%.*]] = load <3 x double>, ptr [[VEC_GEP]], align 8404; CHECK-NEXT: [[VEC_GEP2:%.*]] = getelementptr double, ptr [[A]], i64 6405; CHECK-NEXT: [[COL_LOAD3:%.*]] = load <3 x double>, ptr [[VEC_GEP2]], align 8406; CHECK-NEXT: br label [[EXIT:%.*]]407; CHECK: bbb:408; CHECK-NEXT: [[COL_LOAD9:%.*]] = load <3 x double>, ptr [[B:%.*]], align 8409; CHECK-NEXT: [[VEC_GEP10:%.*]] = getelementptr double, ptr [[B]], i64 3410; CHECK-NEXT: [[COL_LOAD11:%.*]] = load <3 x double>, ptr [[VEC_GEP10]], align 8411; CHECK-NEXT: [[VEC_GEP12:%.*]] = getelementptr double, ptr [[B]], i64 6412; CHECK-NEXT: [[COL_LOAD13:%.*]] = load <3 x double>, ptr [[VEC_GEP12]], align 8413; CHECK-NEXT: br label [[EXIT]]414; CHECK: bbc:415; CHECK-NEXT: [[COL_LOAD4:%.*]] = load <3 x double>, ptr [[C:%.*]], align 8416; CHECK-NEXT: [[VEC_GEP5:%.*]] = getelementptr double, ptr [[C]], i64 3417; CHECK-NEXT: [[COL_LOAD6:%.*]] = load <3 x double>, ptr [[VEC_GEP5]], align 8418; CHECK-NEXT: [[VEC_GEP7:%.*]] = getelementptr double, ptr [[C]], i64 6419; CHECK-NEXT: [[COL_LOAD8:%.*]] = load <3 x double>, ptr [[VEC_GEP7]], align 8420; CHECK-NEXT: br label [[EXIT]]421; CHECK: exit:422; CHECK-NEXT: [[PHI14:%.*]] = phi <3 x double> [ [[COL_LOAD]], [[BBA]] ], [ [[COL_LOAD9]], [[BBB]] ], [ [[COL_LOAD4]], [[BBC]] ]423; CHECK-NEXT: [[PHI15:%.*]] = phi <3 x double> [ [[COL_LOAD1]], [[BBA]] ], [ [[COL_LOAD11]], [[BBB]] ], [ [[COL_LOAD6]], [[BBC]] ]424; CHECK-NEXT: [[PHI16:%.*]] = phi <3 x double> [ [[COL_LOAD3]], [[BBA]] ], [ [[COL_LOAD13]], [[BBB]] ], [ [[COL_LOAD8]], [[BBC]] ]425; CHECK-NEXT: store <3 x double> [[PHI14]], ptr [[OUT:%.*]], align 8426; CHECK-NEXT: [[VEC_GEP17:%.*]] = getelementptr double, ptr [[OUT]], i64 3427; CHECK-NEXT: store <3 x double> [[PHI15]], ptr [[VEC_GEP17]], align 8428; CHECK-NEXT: [[VEC_GEP18:%.*]] = getelementptr double, ptr [[OUT]], i64 6429; CHECK-NEXT: store <3 x double> [[PHI16]], ptr [[VEC_GEP18]], align 8430; CHECK-NEXT: ret void431;432entry:433 br i1 %cond1, label %bb1, label %bba434 435bb1:436 br i1 %cond2, label %bbb, label %bbc437 438bba:439 %va = call <9 x double> @llvm.matrix.column.major.load(ptr %a, i64 3, i1 false, i32 3, i32 3)440 br label %exit441 442bbb:443 %vb = call <9 x double> @llvm.matrix.column.major.load(ptr %b, i64 3, i1 false, i32 3, i32 3)444 br label %exit445 446bbc:447 %vc = call <9 x double> @llvm.matrix.column.major.load(ptr %c, i64 3, i1 false, i32 3, i32 3)448 br label %exit449 450exit:451 %phi = phi <9 x double> [%va, %bba], [%vb, %bbb], [%vc, %bbc]452 call void @llvm.matrix.column.major.store(<9 x double> %phi, ptr %out, i64 3, i1 false, i32 3, i32 3)453 ret void454}455 456define void @matrix_phi_two_preds_shape_mismatch1(i1 %cond1, ptr %a, ptr %b, ptr %out) {457; CHECK-LABEL: @matrix_phi_two_preds_shape_mismatch1(458; CHECK-NEXT: entry:459; CHECK-NEXT: br i1 [[COND1:%.*]], label [[BBA:%.*]], label [[BBB:%.*]]460; CHECK: bba:461; CHECK-NEXT: [[COL_LOAD1:%.*]] = load <3 x double>, ptr [[A:%.*]], align 8462; CHECK-NEXT: [[VEC_GEP:%.*]] = getelementptr double, ptr [[A]], i64 3463; CHECK-NEXT: [[COL_LOAD2:%.*]] = load <3 x double>, ptr [[VEC_GEP]], align 8464; CHECK-NEXT: [[VEC_GEP3:%.*]] = getelementptr double, ptr [[A]], i64 6465; CHECK-NEXT: [[COL_LOAD4:%.*]] = load <3 x double>, ptr [[VEC_GEP3]], align 8466; CHECK-NEXT: [[TMP0:%.*]] = shufflevector <3 x double> [[COL_LOAD1]], <3 x double> [[COL_LOAD2]], <6 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5>467; CHECK-NEXT: [[TMP1:%.*]] = shufflevector <3 x double> [[COL_LOAD4]], <3 x double> poison, <6 x i32> <i32 0, i32 1, i32 2, i32 poison, i32 poison, i32 poison>468; CHECK-NEXT: [[TMP2:%.*]] = shufflevector <6 x double> [[TMP0]], <6 x double> [[TMP1]], <9 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8>469; CHECK-NEXT: [[SPLIT:%.*]] = shufflevector <9 x double> [[TMP2]], <9 x double> poison, <9 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8>470; CHECK-NEXT: br label [[EXIT:%.*]]471; CHECK: bbb:472; CHECK-NEXT: [[COL_LOAD:%.*]] = load <9 x double>, ptr [[B:%.*]], align 8473; CHECK-NEXT: br label [[EXIT]]474; CHECK: exit:475; CHECK-NEXT: [[PHI5:%.*]] = phi <9 x double> [ [[SPLIT]], [[BBA]] ], [ [[COL_LOAD]], [[BBB]] ]476; CHECK-NEXT: store <9 x double> [[PHI5]], ptr [[OUT:%.*]], align 128477; CHECK-NEXT: ret void478;479entry:480 br i1 %cond1, label %bba, label %bbb481 482bba:483 %va = call <9 x double> @llvm.matrix.column.major.load(ptr %a, i64 3, i1 false, i32 3, i32 3)484 br label %exit485 486bbb:487 %vb = call <9 x double> @llvm.matrix.column.major.load(ptr %b, i64 9, i1 false, i32 9, i32 1)488 br label %exit489 490exit:491 %phi = phi <9 x double> [%va, %bba], [%vb, %bbb]492 store <9 x double> %phi, ptr %out493 ret void494}495 496define void @matrix_phi_two_preds_shape_mismatch2(i1 %cond1, ptr %a, ptr %b, ptr %out) {497; CHECK-LABEL: @matrix_phi_two_preds_shape_mismatch2(498; CHECK-NEXT: entry:499; CHECK-NEXT: br i1 [[COND1:%.*]], label [[BBA:%.*]], label [[BBB:%.*]]500; CHECK: bba:501; CHECK-NEXT: [[COL_LOAD4:%.*]] = load <9 x double>, ptr [[A:%.*]], align 8502; CHECK-NEXT: [[SPLIT:%.*]] = shufflevector <9 x double> [[COL_LOAD4]], <9 x double> poison, <3 x i32> <i32 0, i32 1, i32 2>503; CHECK-NEXT: [[SPLIT8:%.*]] = shufflevector <9 x double> [[COL_LOAD4]], <9 x double> poison, <3 x i32> <i32 3, i32 4, i32 5>504; CHECK-NEXT: [[SPLIT9:%.*]] = shufflevector <9 x double> [[COL_LOAD4]], <9 x double> poison, <3 x i32> <i32 6, i32 7, i32 8>505; CHECK-NEXT: br label [[EXIT:%.*]]506; CHECK: bbb:507; CHECK-NEXT: [[COL_LOAD:%.*]] = load <3 x double>, ptr [[B:%.*]], align 8508; CHECK-NEXT: [[VEC_GEP:%.*]] = getelementptr double, ptr [[B]], i64 3509; CHECK-NEXT: [[COL_LOAD1:%.*]] = load <3 x double>, ptr [[VEC_GEP]], align 8510; CHECK-NEXT: [[VEC_GEP2:%.*]] = getelementptr double, ptr [[B]], i64 6511; CHECK-NEXT: [[COL_LOAD3:%.*]] = load <3 x double>, ptr [[VEC_GEP2]], align 8512; CHECK-NEXT: br label [[EXIT]]513; CHECK: exit:514; CHECK-NEXT: [[PHI5:%.*]] = phi <3 x double> [ [[SPLIT]], [[BBA]] ], [ [[COL_LOAD]], [[BBB]] ]515; CHECK-NEXT: [[PHI6:%.*]] = phi <3 x double> [ [[SPLIT8]], [[BBA]] ], [ [[COL_LOAD1]], [[BBB]] ]516; CHECK-NEXT: [[PHI7:%.*]] = phi <3 x double> [ [[SPLIT9]], [[BBA]] ], [ [[COL_LOAD3]], [[BBB]] ]517; CHECK-NEXT: store <3 x double> [[PHI5]], ptr [[OUT:%.*]], align 128518; CHECK-NEXT: [[VEC_GEP10:%.*]] = getelementptr double, ptr [[OUT]], i64 3519; CHECK-NEXT: store <3 x double> [[PHI6]], ptr [[VEC_GEP10]], align 8520; CHECK-NEXT: [[VEC_GEP11:%.*]] = getelementptr double, ptr [[OUT]], i64 6521; CHECK-NEXT: store <3 x double> [[PHI7]], ptr [[VEC_GEP11]], align 16522; CHECK-NEXT: ret void523;524entry:525 br i1 %cond1, label %bba, label %bbb526 527bba:528 %va = call <9 x double> @llvm.matrix.column.major.load(ptr %a, i64 9, i1 false, i32 9, i32 1)529 br label %exit530 531bbb:532 %vb = call <9 x double> @llvm.matrix.column.major.load(ptr %b, i64 3, i1 false, i32 3, i32 3)533 br label %exit534 535exit:536 %phi = phi <9 x double> [%va, %bba], [%vb, %bbb]537 store <9 x double> %phi, ptr %out538 ret void539}540 541define <9 x double> @matrix_phi_ifthenelse(i1 %cond, <9 x double> %A, <9 x double> %B, <9 x double> %C) {542; CHECK-LABEL: @matrix_phi_ifthenelse(543; CHECK-NEXT: entry:544; CHECK-NEXT: br i1 [[COND:%.*]], label [[IF_THEN:%.*]], label [[IF_ELSE:%.*]]545; CHECK: if.then:546; CHECK-NEXT: [[SPLIT3:%.*]] = shufflevector <9 x double> [[A:%.*]], <9 x double> poison, <3 x i32> <i32 0, i32 1, i32 2>547; CHECK-NEXT: [[SPLIT4:%.*]] = shufflevector <9 x double> [[A]], <9 x double> poison, <3 x i32> <i32 3, i32 4, i32 5>548; CHECK-NEXT: [[SPLIT5:%.*]] = shufflevector <9 x double> [[A]], <9 x double> poison, <3 x i32> <i32 6, i32 7, i32 8>549; CHECK-NEXT: [[TMP0:%.*]] = extractelement <3 x double> [[SPLIT3]], i64 0550; CHECK-NEXT: [[TMP1:%.*]] = insertelement <3 x double> poison, double [[TMP0]], i64 0551; CHECK-NEXT: [[TMP2:%.*]] = extractelement <3 x double> [[SPLIT4]], i64 0552; CHECK-NEXT: [[TMP3:%.*]] = insertelement <3 x double> [[TMP1]], double [[TMP2]], i64 1553; CHECK-NEXT: [[TMP4:%.*]] = extractelement <3 x double> [[SPLIT5]], i64 0554; CHECK-NEXT: [[TMP5:%.*]] = insertelement <3 x double> [[TMP3]], double [[TMP4]], i64 2555; CHECK-NEXT: [[TMP6:%.*]] = extractelement <3 x double> [[SPLIT3]], i64 1556; CHECK-NEXT: [[TMP7:%.*]] = insertelement <3 x double> poison, double [[TMP6]], i64 0557; CHECK-NEXT: [[TMP8:%.*]] = extractelement <3 x double> [[SPLIT4]], i64 1558; CHECK-NEXT: [[TMP9:%.*]] = insertelement <3 x double> [[TMP7]], double [[TMP8]], i64 1559; CHECK-NEXT: [[TMP10:%.*]] = extractelement <3 x double> [[SPLIT5]], i64 1560; CHECK-NEXT: [[TMP11:%.*]] = insertelement <3 x double> [[TMP9]], double [[TMP10]], i64 2561; CHECK-NEXT: [[TMP12:%.*]] = extractelement <3 x double> [[SPLIT3]], i64 2562; CHECK-NEXT: [[TMP13:%.*]] = insertelement <3 x double> poison, double [[TMP12]], i64 0563; CHECK-NEXT: [[TMP14:%.*]] = extractelement <3 x double> [[SPLIT4]], i64 2564; CHECK-NEXT: [[TMP15:%.*]] = insertelement <3 x double> [[TMP13]], double [[TMP14]], i64 1565; CHECK-NEXT: [[TMP16:%.*]] = extractelement <3 x double> [[SPLIT5]], i64 2566; CHECK-NEXT: [[TMP17:%.*]] = insertelement <3 x double> [[TMP15]], double [[TMP16]], i64 2567; CHECK-NEXT: br label [[IF_END:%.*]]568; CHECK: if.else:569; CHECK-NEXT: [[SPLIT:%.*]] = shufflevector <9 x double> [[B:%.*]], <9 x double> poison, <3 x i32> <i32 0, i32 1, i32 2>570; CHECK-NEXT: [[SPLIT1:%.*]] = shufflevector <9 x double> [[B]], <9 x double> poison, <3 x i32> <i32 3, i32 4, i32 5>571; CHECK-NEXT: [[SPLIT2:%.*]] = shufflevector <9 x double> [[B]], <9 x double> poison, <3 x i32> <i32 6, i32 7, i32 8>572; CHECK-NEXT: [[TMP21:%.*]] = extractelement <3 x double> [[SPLIT]], i64 0573; CHECK-NEXT: [[TMP22:%.*]] = insertelement <3 x double> poison, double [[TMP21]], i64 0574; CHECK-NEXT: [[TMP23:%.*]] = extractelement <3 x double> [[SPLIT1]], i64 0575; CHECK-NEXT: [[TMP24:%.*]] = insertelement <3 x double> [[TMP22]], double [[TMP23]], i64 1576; CHECK-NEXT: [[TMP25:%.*]] = extractelement <3 x double> [[SPLIT2]], i64 0577; CHECK-NEXT: [[TMP26:%.*]] = insertelement <3 x double> [[TMP24]], double [[TMP25]], i64 2578; CHECK-NEXT: [[TMP27:%.*]] = extractelement <3 x double> [[SPLIT]], i64 1579; CHECK-NEXT: [[TMP28:%.*]] = insertelement <3 x double> poison, double [[TMP27]], i64 0580; CHECK-NEXT: [[TMP29:%.*]] = extractelement <3 x double> [[SPLIT1]], i64 1581; CHECK-NEXT: [[TMP30:%.*]] = insertelement <3 x double> [[TMP28]], double [[TMP29]], i64 1582; CHECK-NEXT: [[TMP31:%.*]] = extractelement <3 x double> [[SPLIT2]], i64 1583; CHECK-NEXT: [[TMP32:%.*]] = insertelement <3 x double> [[TMP30]], double [[TMP31]], i64 2584; CHECK-NEXT: [[TMP33:%.*]] = extractelement <3 x double> [[SPLIT]], i64 2585; CHECK-NEXT: [[TMP34:%.*]] = insertelement <3 x double> poison, double [[TMP33]], i64 0586; CHECK-NEXT: [[TMP35:%.*]] = extractelement <3 x double> [[SPLIT1]], i64 2587; CHECK-NEXT: [[TMP36:%.*]] = insertelement <3 x double> [[TMP34]], double [[TMP35]], i64 1588; CHECK-NEXT: [[TMP37:%.*]] = extractelement <3 x double> [[SPLIT2]], i64 2589; CHECK-NEXT: [[TMP38:%.*]] = insertelement <3 x double> [[TMP36]], double [[TMP37]], i64 2590; CHECK-NEXT: br label [[IF_END]]591; CHECK: if.end:592; CHECK-NEXT: [[MERGE9:%.*]] = phi <3 x double> [ [[TMP5]], [[IF_THEN]] ], [ [[TMP26]], [[IF_ELSE]] ]593; CHECK-NEXT: [[MERGE10:%.*]] = phi <3 x double> [ [[TMP11]], [[IF_THEN]] ], [ [[TMP32]], [[IF_ELSE]] ]594; CHECK-NEXT: [[MERGE11:%.*]] = phi <3 x double> [ [[TMP17]], [[IF_THEN]] ], [ [[TMP38]], [[IF_ELSE]] ]595; CHECK-NEXT: [[SPLIT9:%.*]] = shufflevector <9 x double> [[MERGE:%.*]], <9 x double> poison, <3 x i32> <i32 0, i32 1, i32 2>596; CHECK-NEXT: [[SPLIT10:%.*]] = shufflevector <9 x double> [[MERGE]], <9 x double> poison, <3 x i32> <i32 3, i32 4, i32 5>597; CHECK-NEXT: [[SPLIT11:%.*]] = shufflevector <9 x double> [[MERGE]], <9 x double> poison, <3 x i32> <i32 6, i32 7, i32 8>598; CHECK-NEXT: [[BLOCK:%.*]] = shufflevector <3 x double> [[SPLIT9]], <3 x double> poison, <1 x i32> zeroinitializer599; CHECK-NEXT: [[TMP42:%.*]] = extractelement <3 x double> [[MERGE9]], i64 0600; CHECK-NEXT: [[SPLAT_SPLATINSERT:%.*]] = insertelement <1 x double> poison, double [[TMP42]], i64 0601; CHECK-NEXT: [[SPLAT_SPLAT:%.*]] = shufflevector <1 x double> [[SPLAT_SPLATINSERT]], <1 x double> poison, <1 x i32> zeroinitializer602; CHECK-NEXT: [[TMP43:%.*]] = fmul <1 x double> [[BLOCK]], [[SPLAT_SPLAT]]603; CHECK-NEXT: [[BLOCK12:%.*]] = shufflevector <3 x double> [[SPLIT10]], <3 x double> poison, <1 x i32> zeroinitializer604; CHECK-NEXT: [[TMP44:%.*]] = extractelement <3 x double> [[MERGE9]], i64 1605; CHECK-NEXT: [[SPLAT_SPLATINSERT13:%.*]] = insertelement <1 x double> poison, double [[TMP44]], i64 0606; CHECK-NEXT: [[SPLAT_SPLAT14:%.*]] = shufflevector <1 x double> [[SPLAT_SPLATINSERT13]], <1 x double> poison, <1 x i32> zeroinitializer607; CHECK-NEXT: [[TMP45:%.*]] = fmul <1 x double> [[BLOCK12]], [[SPLAT_SPLAT14]]608; CHECK-NEXT: [[TMP46:%.*]] = fadd <1 x double> [[TMP43]], [[TMP45]]609; CHECK-NEXT: [[BLOCK15:%.*]] = shufflevector <3 x double> [[SPLIT11]], <3 x double> poison, <1 x i32> zeroinitializer610; CHECK-NEXT: [[TMP47:%.*]] = extractelement <3 x double> [[MERGE9]], i64 2611; CHECK-NEXT: [[SPLAT_SPLATINSERT16:%.*]] = insertelement <1 x double> poison, double [[TMP47]], i64 0612; CHECK-NEXT: [[SPLAT_SPLAT17:%.*]] = shufflevector <1 x double> [[SPLAT_SPLATINSERT16]], <1 x double> poison, <1 x i32> zeroinitializer613; CHECK-NEXT: [[TMP48:%.*]] = fmul <1 x double> [[BLOCK15]], [[SPLAT_SPLAT17]]614; CHECK-NEXT: [[TMP49:%.*]] = fadd <1 x double> [[TMP46]], [[TMP48]]615; CHECK-NEXT: [[TMP50:%.*]] = shufflevector <1 x double> [[TMP49]], <1 x double> poison, <3 x i32> <i32 0, i32 poison, i32 poison>616; CHECK-NEXT: [[TMP51:%.*]] = shufflevector <3 x double> poison, <3 x double> [[TMP50]], <3 x i32> <i32 3, i32 1, i32 2>617; CHECK-NEXT: [[BLOCK18:%.*]] = shufflevector <3 x double> [[SPLIT9]], <3 x double> poison, <1 x i32> <i32 1>618; CHECK-NEXT: [[TMP52:%.*]] = extractelement <3 x double> [[MERGE9]], i64 0619; CHECK-NEXT: [[SPLAT_SPLATINSERT19:%.*]] = insertelement <1 x double> poison, double [[TMP52]], i64 0620; CHECK-NEXT: [[SPLAT_SPLAT20:%.*]] = shufflevector <1 x double> [[SPLAT_SPLATINSERT19]], <1 x double> poison, <1 x i32> zeroinitializer621; CHECK-NEXT: [[TMP53:%.*]] = fmul <1 x double> [[BLOCK18]], [[SPLAT_SPLAT20]]622; CHECK-NEXT: [[BLOCK21:%.*]] = shufflevector <3 x double> [[SPLIT10]], <3 x double> poison, <1 x i32> <i32 1>623; CHECK-NEXT: [[TMP54:%.*]] = extractelement <3 x double> [[MERGE9]], i64 1624; CHECK-NEXT: [[SPLAT_SPLATINSERT22:%.*]] = insertelement <1 x double> poison, double [[TMP54]], i64 0625; CHECK-NEXT: [[SPLAT_SPLAT23:%.*]] = shufflevector <1 x double> [[SPLAT_SPLATINSERT22]], <1 x double> poison, <1 x i32> zeroinitializer626; CHECK-NEXT: [[TMP55:%.*]] = fmul <1 x double> [[BLOCK21]], [[SPLAT_SPLAT23]]627; CHECK-NEXT: [[TMP56:%.*]] = fadd <1 x double> [[TMP53]], [[TMP55]]628; CHECK-NEXT: [[BLOCK24:%.*]] = shufflevector <3 x double> [[SPLIT11]], <3 x double> poison, <1 x i32> <i32 1>629; CHECK-NEXT: [[TMP57:%.*]] = extractelement <3 x double> [[MERGE9]], i64 2630; CHECK-NEXT: [[SPLAT_SPLATINSERT25:%.*]] = insertelement <1 x double> poison, double [[TMP57]], i64 0631; CHECK-NEXT: [[SPLAT_SPLAT26:%.*]] = shufflevector <1 x double> [[SPLAT_SPLATINSERT25]], <1 x double> poison, <1 x i32> zeroinitializer632; CHECK-NEXT: [[TMP58:%.*]] = fmul <1 x double> [[BLOCK24]], [[SPLAT_SPLAT26]]633; CHECK-NEXT: [[TMP59:%.*]] = fadd <1 x double> [[TMP56]], [[TMP58]]634; CHECK-NEXT: [[TMP60:%.*]] = shufflevector <1 x double> [[TMP59]], <1 x double> poison, <3 x i32> <i32 0, i32 poison, i32 poison>635; CHECK-NEXT: [[TMP61:%.*]] = shufflevector <3 x double> [[TMP51]], <3 x double> [[TMP60]], <3 x i32> <i32 0, i32 3, i32 2>636; CHECK-NEXT: [[BLOCK27:%.*]] = shufflevector <3 x double> [[SPLIT9]], <3 x double> poison, <1 x i32> <i32 2>637; CHECK-NEXT: [[TMP62:%.*]] = extractelement <3 x double> [[MERGE9]], i64 0638; CHECK-NEXT: [[SPLAT_SPLATINSERT28:%.*]] = insertelement <1 x double> poison, double [[TMP62]], i64 0639; CHECK-NEXT: [[SPLAT_SPLAT29:%.*]] = shufflevector <1 x double> [[SPLAT_SPLATINSERT28]], <1 x double> poison, <1 x i32> zeroinitializer640; CHECK-NEXT: [[TMP63:%.*]] = fmul <1 x double> [[BLOCK27]], [[SPLAT_SPLAT29]]641; CHECK-NEXT: [[BLOCK30:%.*]] = shufflevector <3 x double> [[SPLIT10]], <3 x double> poison, <1 x i32> <i32 2>642; CHECK-NEXT: [[TMP64:%.*]] = extractelement <3 x double> [[MERGE9]], i64 1643; CHECK-NEXT: [[SPLAT_SPLATINSERT31:%.*]] = insertelement <1 x double> poison, double [[TMP64]], i64 0644; CHECK-NEXT: [[SPLAT_SPLAT32:%.*]] = shufflevector <1 x double> [[SPLAT_SPLATINSERT31]], <1 x double> poison, <1 x i32> zeroinitializer645; CHECK-NEXT: [[TMP65:%.*]] = fmul <1 x double> [[BLOCK30]], [[SPLAT_SPLAT32]]646; CHECK-NEXT: [[TMP66:%.*]] = fadd <1 x double> [[TMP63]], [[TMP65]]647; CHECK-NEXT: [[BLOCK33:%.*]] = shufflevector <3 x double> [[SPLIT11]], <3 x double> poison, <1 x i32> <i32 2>648; CHECK-NEXT: [[TMP67:%.*]] = extractelement <3 x double> [[MERGE9]], i64 2649; CHECK-NEXT: [[SPLAT_SPLATINSERT34:%.*]] = insertelement <1 x double> poison, double [[TMP67]], i64 0650; CHECK-NEXT: [[SPLAT_SPLAT35:%.*]] = shufflevector <1 x double> [[SPLAT_SPLATINSERT34]], <1 x double> poison, <1 x i32> zeroinitializer651; CHECK-NEXT: [[TMP68:%.*]] = fmul <1 x double> [[BLOCK33]], [[SPLAT_SPLAT35]]652; CHECK-NEXT: [[TMP69:%.*]] = fadd <1 x double> [[TMP66]], [[TMP68]]653; CHECK-NEXT: [[TMP70:%.*]] = shufflevector <1 x double> [[TMP69]], <1 x double> poison, <3 x i32> <i32 0, i32 poison, i32 poison>654; CHECK-NEXT: [[TMP71:%.*]] = shufflevector <3 x double> [[TMP61]], <3 x double> [[TMP70]], <3 x i32> <i32 0, i32 1, i32 3>655; CHECK-NEXT: [[BLOCK36:%.*]] = shufflevector <3 x double> [[SPLIT9]], <3 x double> poison, <1 x i32> zeroinitializer656; CHECK-NEXT: [[TMP72:%.*]] = extractelement <3 x double> [[MERGE10]], i64 0657; CHECK-NEXT: [[SPLAT_SPLATINSERT37:%.*]] = insertelement <1 x double> poison, double [[TMP72]], i64 0658; CHECK-NEXT: [[SPLAT_SPLAT38:%.*]] = shufflevector <1 x double> [[SPLAT_SPLATINSERT37]], <1 x double> poison, <1 x i32> zeroinitializer659; CHECK-NEXT: [[TMP73:%.*]] = fmul <1 x double> [[BLOCK36]], [[SPLAT_SPLAT38]]660; CHECK-NEXT: [[BLOCK39:%.*]] = shufflevector <3 x double> [[SPLIT10]], <3 x double> poison, <1 x i32> zeroinitializer661; CHECK-NEXT: [[TMP74:%.*]] = extractelement <3 x double> [[MERGE10]], i64 1662; CHECK-NEXT: [[SPLAT_SPLATINSERT40:%.*]] = insertelement <1 x double> poison, double [[TMP74]], i64 0663; CHECK-NEXT: [[SPLAT_SPLAT41:%.*]] = shufflevector <1 x double> [[SPLAT_SPLATINSERT40]], <1 x double> poison, <1 x i32> zeroinitializer664; CHECK-NEXT: [[TMP75:%.*]] = fmul <1 x double> [[BLOCK39]], [[SPLAT_SPLAT41]]665; CHECK-NEXT: [[TMP76:%.*]] = fadd <1 x double> [[TMP73]], [[TMP75]]666; CHECK-NEXT: [[BLOCK42:%.*]] = shufflevector <3 x double> [[SPLIT11]], <3 x double> poison, <1 x i32> zeroinitializer667; CHECK-NEXT: [[TMP77:%.*]] = extractelement <3 x double> [[MERGE10]], i64 2668; CHECK-NEXT: [[SPLAT_SPLATINSERT43:%.*]] = insertelement <1 x double> poison, double [[TMP77]], i64 0669; CHECK-NEXT: [[SPLAT_SPLAT44:%.*]] = shufflevector <1 x double> [[SPLAT_SPLATINSERT43]], <1 x double> poison, <1 x i32> zeroinitializer670; CHECK-NEXT: [[TMP78:%.*]] = fmul <1 x double> [[BLOCK42]], [[SPLAT_SPLAT44]]671; CHECK-NEXT: [[TMP79:%.*]] = fadd <1 x double> [[TMP76]], [[TMP78]]672; CHECK-NEXT: [[TMP80:%.*]] = shufflevector <1 x double> [[TMP79]], <1 x double> poison, <3 x i32> <i32 0, i32 poison, i32 poison>673; CHECK-NEXT: [[TMP81:%.*]] = shufflevector <3 x double> poison, <3 x double> [[TMP80]], <3 x i32> <i32 3, i32 1, i32 2>674; CHECK-NEXT: [[BLOCK45:%.*]] = shufflevector <3 x double> [[SPLIT9]], <3 x double> poison, <1 x i32> <i32 1>675; CHECK-NEXT: [[TMP82:%.*]] = extractelement <3 x double> [[MERGE10]], i64 0676; CHECK-NEXT: [[SPLAT_SPLATINSERT46:%.*]] = insertelement <1 x double> poison, double [[TMP82]], i64 0677; CHECK-NEXT: [[SPLAT_SPLAT47:%.*]] = shufflevector <1 x double> [[SPLAT_SPLATINSERT46]], <1 x double> poison, <1 x i32> zeroinitializer678; CHECK-NEXT: [[TMP83:%.*]] = fmul <1 x double> [[BLOCK45]], [[SPLAT_SPLAT47]]679; CHECK-NEXT: [[BLOCK48:%.*]] = shufflevector <3 x double> [[SPLIT10]], <3 x double> poison, <1 x i32> <i32 1>680; CHECK-NEXT: [[TMP84:%.*]] = extractelement <3 x double> [[MERGE10]], i64 1681; CHECK-NEXT: [[SPLAT_SPLATINSERT49:%.*]] = insertelement <1 x double> poison, double [[TMP84]], i64 0682; CHECK-NEXT: [[SPLAT_SPLAT50:%.*]] = shufflevector <1 x double> [[SPLAT_SPLATINSERT49]], <1 x double> poison, <1 x i32> zeroinitializer683; CHECK-NEXT: [[TMP85:%.*]] = fmul <1 x double> [[BLOCK48]], [[SPLAT_SPLAT50]]684; CHECK-NEXT: [[TMP86:%.*]] = fadd <1 x double> [[TMP83]], [[TMP85]]685; CHECK-NEXT: [[BLOCK51:%.*]] = shufflevector <3 x double> [[SPLIT11]], <3 x double> poison, <1 x i32> <i32 1>686; CHECK-NEXT: [[TMP87:%.*]] = extractelement <3 x double> [[MERGE10]], i64 2687; CHECK-NEXT: [[SPLAT_SPLATINSERT52:%.*]] = insertelement <1 x double> poison, double [[TMP87]], i64 0688; CHECK-NEXT: [[SPLAT_SPLAT53:%.*]] = shufflevector <1 x double> [[SPLAT_SPLATINSERT52]], <1 x double> poison, <1 x i32> zeroinitializer689; CHECK-NEXT: [[TMP88:%.*]] = fmul <1 x double> [[BLOCK51]], [[SPLAT_SPLAT53]]690; CHECK-NEXT: [[TMP89:%.*]] = fadd <1 x double> [[TMP86]], [[TMP88]]691; CHECK-NEXT: [[TMP90:%.*]] = shufflevector <1 x double> [[TMP89]], <1 x double> poison, <3 x i32> <i32 0, i32 poison, i32 poison>692; CHECK-NEXT: [[TMP91:%.*]] = shufflevector <3 x double> [[TMP81]], <3 x double> [[TMP90]], <3 x i32> <i32 0, i32 3, i32 2>693; CHECK-NEXT: [[BLOCK54:%.*]] = shufflevector <3 x double> [[SPLIT9]], <3 x double> poison, <1 x i32> <i32 2>694; CHECK-NEXT: [[TMP92:%.*]] = extractelement <3 x double> [[MERGE10]], i64 0695; CHECK-NEXT: [[SPLAT_SPLATINSERT55:%.*]] = insertelement <1 x double> poison, double [[TMP92]], i64 0696; CHECK-NEXT: [[SPLAT_SPLAT56:%.*]] = shufflevector <1 x double> [[SPLAT_SPLATINSERT55]], <1 x double> poison, <1 x i32> zeroinitializer697; CHECK-NEXT: [[TMP93:%.*]] = fmul <1 x double> [[BLOCK54]], [[SPLAT_SPLAT56]]698; CHECK-NEXT: [[BLOCK57:%.*]] = shufflevector <3 x double> [[SPLIT10]], <3 x double> poison, <1 x i32> <i32 2>699; CHECK-NEXT: [[TMP94:%.*]] = extractelement <3 x double> [[MERGE10]], i64 1700; CHECK-NEXT: [[SPLAT_SPLATINSERT58:%.*]] = insertelement <1 x double> poison, double [[TMP94]], i64 0701; CHECK-NEXT: [[SPLAT_SPLAT59:%.*]] = shufflevector <1 x double> [[SPLAT_SPLATINSERT58]], <1 x double> poison, <1 x i32> zeroinitializer702; CHECK-NEXT: [[TMP95:%.*]] = fmul <1 x double> [[BLOCK57]], [[SPLAT_SPLAT59]]703; CHECK-NEXT: [[TMP96:%.*]] = fadd <1 x double> [[TMP93]], [[TMP95]]704; CHECK-NEXT: [[BLOCK60:%.*]] = shufflevector <3 x double> [[SPLIT11]], <3 x double> poison, <1 x i32> <i32 2>705; CHECK-NEXT: [[TMP97:%.*]] = extractelement <3 x double> [[MERGE10]], i64 2706; CHECK-NEXT: [[SPLAT_SPLATINSERT61:%.*]] = insertelement <1 x double> poison, double [[TMP97]], i64 0707; CHECK-NEXT: [[SPLAT_SPLAT62:%.*]] = shufflevector <1 x double> [[SPLAT_SPLATINSERT61]], <1 x double> poison, <1 x i32> zeroinitializer708; CHECK-NEXT: [[TMP98:%.*]] = fmul <1 x double> [[BLOCK60]], [[SPLAT_SPLAT62]]709; CHECK-NEXT: [[TMP99:%.*]] = fadd <1 x double> [[TMP96]], [[TMP98]]710; CHECK-NEXT: [[TMP100:%.*]] = shufflevector <1 x double> [[TMP99]], <1 x double> poison, <3 x i32> <i32 0, i32 poison, i32 poison>711; CHECK-NEXT: [[TMP101:%.*]] = shufflevector <3 x double> [[TMP91]], <3 x double> [[TMP100]], <3 x i32> <i32 0, i32 1, i32 3>712; CHECK-NEXT: [[BLOCK63:%.*]] = shufflevector <3 x double> [[SPLIT9]], <3 x double> poison, <1 x i32> zeroinitializer713; CHECK-NEXT: [[TMP102:%.*]] = extractelement <3 x double> [[MERGE11]], i64 0714; CHECK-NEXT: [[SPLAT_SPLATINSERT64:%.*]] = insertelement <1 x double> poison, double [[TMP102]], i64 0715; CHECK-NEXT: [[SPLAT_SPLAT65:%.*]] = shufflevector <1 x double> [[SPLAT_SPLATINSERT64]], <1 x double> poison, <1 x i32> zeroinitializer716; CHECK-NEXT: [[TMP103:%.*]] = fmul <1 x double> [[BLOCK63]], [[SPLAT_SPLAT65]]717; CHECK-NEXT: [[BLOCK66:%.*]] = shufflevector <3 x double> [[SPLIT10]], <3 x double> poison, <1 x i32> zeroinitializer718; CHECK-NEXT: [[TMP104:%.*]] = extractelement <3 x double> [[MERGE11]], i64 1719; CHECK-NEXT: [[SPLAT_SPLATINSERT67:%.*]] = insertelement <1 x double> poison, double [[TMP104]], i64 0720; CHECK-NEXT: [[SPLAT_SPLAT68:%.*]] = shufflevector <1 x double> [[SPLAT_SPLATINSERT67]], <1 x double> poison, <1 x i32> zeroinitializer721; CHECK-NEXT: [[TMP105:%.*]] = fmul <1 x double> [[BLOCK66]], [[SPLAT_SPLAT68]]722; CHECK-NEXT: [[TMP106:%.*]] = fadd <1 x double> [[TMP103]], [[TMP105]]723; CHECK-NEXT: [[BLOCK69:%.*]] = shufflevector <3 x double> [[SPLIT11]], <3 x double> poison, <1 x i32> zeroinitializer724; CHECK-NEXT: [[TMP107:%.*]] = extractelement <3 x double> [[MERGE11]], i64 2725; CHECK-NEXT: [[SPLAT_SPLATINSERT70:%.*]] = insertelement <1 x double> poison, double [[TMP107]], i64 0726; CHECK-NEXT: [[SPLAT_SPLAT71:%.*]] = shufflevector <1 x double> [[SPLAT_SPLATINSERT70]], <1 x double> poison, <1 x i32> zeroinitializer727; CHECK-NEXT: [[TMP108:%.*]] = fmul <1 x double> [[BLOCK69]], [[SPLAT_SPLAT71]]728; CHECK-NEXT: [[TMP109:%.*]] = fadd <1 x double> [[TMP106]], [[TMP108]]729; CHECK-NEXT: [[TMP110:%.*]] = shufflevector <1 x double> [[TMP109]], <1 x double> poison, <3 x i32> <i32 0, i32 poison, i32 poison>730; CHECK-NEXT: [[TMP111:%.*]] = shufflevector <3 x double> poison, <3 x double> [[TMP110]], <3 x i32> <i32 3, i32 1, i32 2>731; CHECK-NEXT: [[BLOCK72:%.*]] = shufflevector <3 x double> [[SPLIT9]], <3 x double> poison, <1 x i32> <i32 1>732; CHECK-NEXT: [[TMP112:%.*]] = extractelement <3 x double> [[MERGE11]], i64 0733; CHECK-NEXT: [[SPLAT_SPLATINSERT73:%.*]] = insertelement <1 x double> poison, double [[TMP112]], i64 0734; CHECK-NEXT: [[SPLAT_SPLAT74:%.*]] = shufflevector <1 x double> [[SPLAT_SPLATINSERT73]], <1 x double> poison, <1 x i32> zeroinitializer735; CHECK-NEXT: [[TMP113:%.*]] = fmul <1 x double> [[BLOCK72]], [[SPLAT_SPLAT74]]736; CHECK-NEXT: [[BLOCK75:%.*]] = shufflevector <3 x double> [[SPLIT10]], <3 x double> poison, <1 x i32> <i32 1>737; CHECK-NEXT: [[TMP114:%.*]] = extractelement <3 x double> [[MERGE11]], i64 1738; CHECK-NEXT: [[SPLAT_SPLATINSERT76:%.*]] = insertelement <1 x double> poison, double [[TMP114]], i64 0739; CHECK-NEXT: [[SPLAT_SPLAT77:%.*]] = shufflevector <1 x double> [[SPLAT_SPLATINSERT76]], <1 x double> poison, <1 x i32> zeroinitializer740; CHECK-NEXT: [[TMP115:%.*]] = fmul <1 x double> [[BLOCK75]], [[SPLAT_SPLAT77]]741; CHECK-NEXT: [[TMP116:%.*]] = fadd <1 x double> [[TMP113]], [[TMP115]]742; CHECK-NEXT: [[BLOCK78:%.*]] = shufflevector <3 x double> [[SPLIT11]], <3 x double> poison, <1 x i32> <i32 1>743; CHECK-NEXT: [[TMP117:%.*]] = extractelement <3 x double> [[MERGE11]], i64 2744; CHECK-NEXT: [[SPLAT_SPLATINSERT79:%.*]] = insertelement <1 x double> poison, double [[TMP117]], i64 0745; CHECK-NEXT: [[SPLAT_SPLAT80:%.*]] = shufflevector <1 x double> [[SPLAT_SPLATINSERT79]], <1 x double> poison, <1 x i32> zeroinitializer746; CHECK-NEXT: [[TMP118:%.*]] = fmul <1 x double> [[BLOCK78]], [[SPLAT_SPLAT80]]747; CHECK-NEXT: [[TMP119:%.*]] = fadd <1 x double> [[TMP116]], [[TMP118]]748; CHECK-NEXT: [[TMP120:%.*]] = shufflevector <1 x double> [[TMP119]], <1 x double> poison, <3 x i32> <i32 0, i32 poison, i32 poison>749; CHECK-NEXT: [[TMP121:%.*]] = shufflevector <3 x double> [[TMP111]], <3 x double> [[TMP120]], <3 x i32> <i32 0, i32 3, i32 2>750; CHECK-NEXT: [[BLOCK81:%.*]] = shufflevector <3 x double> [[SPLIT9]], <3 x double> poison, <1 x i32> <i32 2>751; CHECK-NEXT: [[TMP122:%.*]] = extractelement <3 x double> [[MERGE11]], i64 0752; CHECK-NEXT: [[SPLAT_SPLATINSERT82:%.*]] = insertelement <1 x double> poison, double [[TMP122]], i64 0753; CHECK-NEXT: [[SPLAT_SPLAT83:%.*]] = shufflevector <1 x double> [[SPLAT_SPLATINSERT82]], <1 x double> poison, <1 x i32> zeroinitializer754; CHECK-NEXT: [[TMP123:%.*]] = fmul <1 x double> [[BLOCK81]], [[SPLAT_SPLAT83]]755; CHECK-NEXT: [[BLOCK84:%.*]] = shufflevector <3 x double> [[SPLIT10]], <3 x double> poison, <1 x i32> <i32 2>756; CHECK-NEXT: [[TMP124:%.*]] = extractelement <3 x double> [[MERGE11]], i64 1757; CHECK-NEXT: [[SPLAT_SPLATINSERT85:%.*]] = insertelement <1 x double> poison, double [[TMP124]], i64 0758; CHECK-NEXT: [[SPLAT_SPLAT86:%.*]] = shufflevector <1 x double> [[SPLAT_SPLATINSERT85]], <1 x double> poison, <1 x i32> zeroinitializer759; CHECK-NEXT: [[TMP125:%.*]] = fmul <1 x double> [[BLOCK84]], [[SPLAT_SPLAT86]]760; CHECK-NEXT: [[TMP126:%.*]] = fadd <1 x double> [[TMP123]], [[TMP125]]761; CHECK-NEXT: [[BLOCK87:%.*]] = shufflevector <3 x double> [[SPLIT11]], <3 x double> poison, <1 x i32> <i32 2>762; CHECK-NEXT: [[TMP127:%.*]] = extractelement <3 x double> [[MERGE11]], i64 2763; CHECK-NEXT: [[SPLAT_SPLATINSERT88:%.*]] = insertelement <1 x double> poison, double [[TMP127]], i64 0764; CHECK-NEXT: [[SPLAT_SPLAT89:%.*]] = shufflevector <1 x double> [[SPLAT_SPLATINSERT88]], <1 x double> poison, <1 x i32> zeroinitializer765; CHECK-NEXT: [[TMP128:%.*]] = fmul <1 x double> [[BLOCK87]], [[SPLAT_SPLAT89]]766; CHECK-NEXT: [[TMP129:%.*]] = fadd <1 x double> [[TMP126]], [[TMP128]]767; CHECK-NEXT: [[TMP130:%.*]] = shufflevector <1 x double> [[TMP129]], <1 x double> poison, <3 x i32> <i32 0, i32 poison, i32 poison>768; CHECK-NEXT: [[TMP131:%.*]] = shufflevector <3 x double> [[TMP121]], <3 x double> [[TMP130]], <3 x i32> <i32 0, i32 1, i32 3>769; CHECK-NEXT: [[TMP132:%.*]] = shufflevector <3 x double> [[TMP71]], <3 x double> [[TMP101]], <6 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5>770; CHECK-NEXT: [[TMP133:%.*]] = shufflevector <3 x double> [[TMP131]], <3 x double> poison, <6 x i32> <i32 0, i32 1, i32 2, i32 poison, i32 poison, i32 poison>771; CHECK-NEXT: [[TMP134:%.*]] = shufflevector <6 x double> [[TMP132]], <6 x double> [[TMP133]], <9 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8>772; CHECK-NEXT: ret <9 x double> [[TMP134]]773;774entry:775 br i1 %cond, label %if.then, label %if.else776 777if.then: ; preds = %entry778 %A.trans = tail call <9 x double> @llvm.matrix.transpose.v9f64(<9 x double> %A, i32 3, i32 3)779 br label %if.end780 781if.else: ; preds = %entry782 %B.trans = tail call <9 x double> @llvm.matrix.transpose.v9f64(<9 x double> %B, i32 3, i32 3)783 br label %if.end784 785if.end: ; preds = %if.then, %if.else786 %merge = phi <9 x double> [ %A.trans, %if.then], [ %B.trans, %if.else ]787 %res = tail call <9 x double> @llvm.matrix.multiply.v9f64.v9f64.v9f64(<9 x double> %C, <9 x double> %merge, i32 3, i32 3, i32 3)788 ret <9 x double> %res789}790