529 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_test_checks.py2; RUN: %if x86-registered-target %{ opt -S -passes=slp-vectorizer -slp-threshold=-10000 -mtriple=x86_64-unknown-unknown < %s | FileCheck %s --check-prefixes=CHECK,THRESHOLD %}3; RUN: %if x86-registered-target %{ opt -S -passes=slp-vectorizer -slp-threshold=0 -mtriple=x86_64-unknown-unknown < %s | FileCheck %s --check-prefixes=CHECK,NOTHRESHOLD %}4; RUN: %if x86-registered-target %{ opt -S -passes=slp-vectorizer -slp-threshold=-10000 -slp-min-tree-size=0 -mtriple=x86_64-unknown-unknown < %s | FileCheck %s --check-prefixes=CHECK,MINTREESIZE %}5; RUN: %if aarch64-registered-target %{ opt -S -passes=slp-vectorizer -slp-threshold=-10000 -mtriple=aarch64-unknown-linux-gnu < %s | FileCheck %s --check-prefixes=CHECK,THRESHOLD %}6; RUN: %if aarch64-registered-target %{ opt -S -passes=slp-vectorizer -slp-threshold=0 -mtriple=aarch64-unknown-linux-gnu < %s | FileCheck %s --check-prefixes=CHECK,NOTHRESHOLD %}7; RUN: %if aarch64-registered-target %{ opt -S -passes=slp-vectorizer -slp-threshold=-10000 -slp-min-tree-size=0 -mtriple=aarch64-unknown-linux-gnu < %s | FileCheck %s --check-prefixes=CHECK,MINTREESIZE %}8 9define <4 x float> @simple_select(<4 x float> %a, <4 x float> %b, <4 x i32> %c) {10; CHECK-LABEL: @simple_select(11; CHECK-NEXT: [[TMP1:%.*]] = icmp ne <4 x i32> [[C:%.*]], zeroinitializer12; CHECK-NEXT: [[TMP2:%.*]] = select <4 x i1> [[TMP1]], <4 x float> [[A:%.*]], <4 x float> [[B:%.*]]13; CHECK-NEXT: ret <4 x float> [[TMP2]]14;15 %c0 = extractelement <4 x i32> %c, i32 016 %c1 = extractelement <4 x i32> %c, i32 117 %c2 = extractelement <4 x i32> %c, i32 218 %c3 = extractelement <4 x i32> %c, i32 319 %a0 = extractelement <4 x float> %a, i32 020 %a1 = extractelement <4 x float> %a, i32 121 %a2 = extractelement <4 x float> %a, i32 222 %a3 = extractelement <4 x float> %a, i32 323 %b0 = extractelement <4 x float> %b, i32 024 %b1 = extractelement <4 x float> %b, i32 125 %b2 = extractelement <4 x float> %b, i32 226 %b3 = extractelement <4 x float> %b, i32 327 %cmp0 = icmp ne i32 %c0, 028 %cmp1 = icmp ne i32 %c1, 029 %cmp2 = icmp ne i32 %c2, 030 %cmp3 = icmp ne i32 %c3, 031 %s0 = select i1 %cmp0, float %a0, float %b032 %s1 = select i1 %cmp1, float %a1, float %b133 %s2 = select i1 %cmp2, float %a2, float %b234 %s3 = select i1 %cmp3, float %a3, float %b335 %ra = insertelement <4 x float> poison, float %s0, i32 036 %rb = insertelement <4 x float> %ra, float %s1, i32 137 %rc = insertelement <4 x float> %rb, float %s2, i32 238 %rd = insertelement <4 x float> %rc, float %s3, i32 339 ret <4 x float> %rd40}41 42declare void @llvm.assume(i1) nounwind43 44; This entire tree is ephemeral, don't vectorize any of it.45define <4 x float> @simple_select_eph(<4 x float> %a, <4 x float> %b, <4 x i32> %c) {46; THRESHOLD-LABEL: @simple_select_eph(47; THRESHOLD-NEXT: [[C0:%.*]] = extractelement <4 x i32> [[C:%.*]], i32 048; THRESHOLD-NEXT: [[C1:%.*]] = extractelement <4 x i32> [[C]], i32 149; THRESHOLD-NEXT: [[C2:%.*]] = extractelement <4 x i32> [[C]], i32 250; THRESHOLD-NEXT: [[C3:%.*]] = extractelement <4 x i32> [[C]], i32 351; THRESHOLD-NEXT: [[A0:%.*]] = extractelement <4 x float> [[A:%.*]], i32 052; THRESHOLD-NEXT: [[A1:%.*]] = extractelement <4 x float> [[A]], i32 153; THRESHOLD-NEXT: [[A2:%.*]] = extractelement <4 x float> [[A]], i32 254; THRESHOLD-NEXT: [[A3:%.*]] = extractelement <4 x float> [[A]], i32 355; THRESHOLD-NEXT: [[B0:%.*]] = extractelement <4 x float> [[B:%.*]], i32 056; THRESHOLD-NEXT: [[B1:%.*]] = extractelement <4 x float> [[B]], i32 157; THRESHOLD-NEXT: [[B2:%.*]] = extractelement <4 x float> [[B]], i32 258; THRESHOLD-NEXT: [[B3:%.*]] = extractelement <4 x float> [[B]], i32 359; THRESHOLD-NEXT: [[CMP0:%.*]] = icmp ne i32 [[C0]], 060; THRESHOLD-NEXT: [[CMP1:%.*]] = icmp ne i32 [[C1]], 061; THRESHOLD-NEXT: [[CMP2:%.*]] = icmp ne i32 [[C2]], 062; THRESHOLD-NEXT: [[CMP3:%.*]] = icmp ne i32 [[C3]], 063; THRESHOLD-NEXT: [[S0:%.*]] = select i1 [[CMP0]], float [[A0]], float [[B0]]64; THRESHOLD-NEXT: [[S1:%.*]] = select i1 [[CMP1]], float [[A1]], float [[B1]]65; THRESHOLD-NEXT: [[S2:%.*]] = select i1 [[CMP2]], float [[A2]], float [[B2]]66; THRESHOLD-NEXT: [[S3:%.*]] = select i1 [[CMP3]], float [[A3]], float [[B3]]67; THRESHOLD-NEXT: [[RA:%.*]] = insertelement <4 x float> poison, float [[S0]], i32 068; THRESHOLD-NEXT: [[RB:%.*]] = insertelement <4 x float> [[RA]], float [[S1]], i32 169; THRESHOLD-NEXT: [[RC:%.*]] = insertelement <4 x float> [[RB]], float [[S2]], i32 270; THRESHOLD-NEXT: [[RD:%.*]] = insertelement <4 x float> [[RC]], float [[S3]], i32 371; THRESHOLD-NEXT: [[Q0:%.*]] = extractelement <4 x float> [[RD]], i32 072; THRESHOLD-NEXT: [[Q1:%.*]] = extractelement <4 x float> [[RD]], i32 173; THRESHOLD-NEXT: [[TMP1:%.*]] = shufflevector <4 x float> [[RD]], <4 x float> poison, <2 x i32> <i32 0, i32 1>74; THRESHOLD-NEXT: [[Q2:%.*]] = extractelement <4 x float> [[RD]], i32 275; THRESHOLD-NEXT: [[Q3:%.*]] = extractelement <4 x float> [[RD]], i32 376; THRESHOLD-NEXT: [[TMP2:%.*]] = shufflevector <4 x float> [[RD]], <4 x float> poison, <2 x i32> <i32 2, i32 3>77; THRESHOLD-NEXT: [[Q4:%.*]] = fadd float [[Q0]], [[Q1]]78; THRESHOLD-NEXT: [[Q5:%.*]] = fadd float [[Q2]], [[Q3]]79; THRESHOLD-NEXT: [[Q6:%.*]] = fadd float [[Q4]], [[Q5]]80; THRESHOLD-NEXT: [[QI:%.*]] = fcmp olt float [[Q6]], [[Q5]]81; THRESHOLD-NEXT: call void @llvm.assume(i1 [[QI]])82; THRESHOLD-NEXT: ret <4 x float> undef83;84; NOTHRESHOLD-LABEL: @simple_select_eph(85; NOTHRESHOLD-NEXT: [[C0:%.*]] = extractelement <4 x i32> [[C:%.*]], i32 086; NOTHRESHOLD-NEXT: [[C1:%.*]] = extractelement <4 x i32> [[C]], i32 187; NOTHRESHOLD-NEXT: [[C2:%.*]] = extractelement <4 x i32> [[C]], i32 288; NOTHRESHOLD-NEXT: [[C3:%.*]] = extractelement <4 x i32> [[C]], i32 389; NOTHRESHOLD-NEXT: [[A0:%.*]] = extractelement <4 x float> [[A:%.*]], i32 090; NOTHRESHOLD-NEXT: [[A1:%.*]] = extractelement <4 x float> [[A]], i32 191; NOTHRESHOLD-NEXT: [[A2:%.*]] = extractelement <4 x float> [[A]], i32 292; NOTHRESHOLD-NEXT: [[A3:%.*]] = extractelement <4 x float> [[A]], i32 393; NOTHRESHOLD-NEXT: [[B0:%.*]] = extractelement <4 x float> [[B:%.*]], i32 094; NOTHRESHOLD-NEXT: [[B1:%.*]] = extractelement <4 x float> [[B]], i32 195; NOTHRESHOLD-NEXT: [[B2:%.*]] = extractelement <4 x float> [[B]], i32 296; NOTHRESHOLD-NEXT: [[B3:%.*]] = extractelement <4 x float> [[B]], i32 397; NOTHRESHOLD-NEXT: [[CMP0:%.*]] = icmp ne i32 [[C0]], 098; NOTHRESHOLD-NEXT: [[CMP1:%.*]] = icmp ne i32 [[C1]], 099; NOTHRESHOLD-NEXT: [[CMP2:%.*]] = icmp ne i32 [[C2]], 0100; NOTHRESHOLD-NEXT: [[CMP3:%.*]] = icmp ne i32 [[C3]], 0101; NOTHRESHOLD-NEXT: [[S0:%.*]] = select i1 [[CMP0]], float [[A0]], float [[B0]]102; NOTHRESHOLD-NEXT: [[S1:%.*]] = select i1 [[CMP1]], float [[A1]], float [[B1]]103; NOTHRESHOLD-NEXT: [[S2:%.*]] = select i1 [[CMP2]], float [[A2]], float [[B2]]104; NOTHRESHOLD-NEXT: [[S3:%.*]] = select i1 [[CMP3]], float [[A3]], float [[B3]]105; NOTHRESHOLD-NEXT: [[RA:%.*]] = insertelement <4 x float> poison, float [[S0]], i32 0106; NOTHRESHOLD-NEXT: [[RB:%.*]] = insertelement <4 x float> [[RA]], float [[S1]], i32 1107; NOTHRESHOLD-NEXT: [[RC:%.*]] = insertelement <4 x float> [[RB]], float [[S2]], i32 2108; NOTHRESHOLD-NEXT: [[RD:%.*]] = insertelement <4 x float> [[RC]], float [[S3]], i32 3109; NOTHRESHOLD-NEXT: [[Q0:%.*]] = extractelement <4 x float> [[RD]], i32 0110; NOTHRESHOLD-NEXT: [[Q1:%.*]] = extractelement <4 x float> [[RD]], i32 1111; NOTHRESHOLD-NEXT: [[Q2:%.*]] = extractelement <4 x float> [[RD]], i32 2112; NOTHRESHOLD-NEXT: [[Q3:%.*]] = extractelement <4 x float> [[RD]], i32 3113; NOTHRESHOLD-NEXT: [[Q4:%.*]] = fadd float [[Q0]], [[Q1]]114; NOTHRESHOLD-NEXT: [[Q5:%.*]] = fadd float [[Q2]], [[Q3]]115; NOTHRESHOLD-NEXT: [[Q6:%.*]] = fadd float [[Q4]], [[Q5]]116; NOTHRESHOLD-NEXT: [[QI:%.*]] = fcmp olt float [[Q6]], [[Q5]]117; NOTHRESHOLD-NEXT: call void @llvm.assume(i1 [[QI]])118; NOTHRESHOLD-NEXT: ret <4 x float> undef119;120; MINTREESIZE-LABEL: @simple_select_eph(121; MINTREESIZE-NEXT: [[C0:%.*]] = extractelement <4 x i32> [[C:%.*]], i32 0122; MINTREESIZE-NEXT: [[C1:%.*]] = extractelement <4 x i32> [[C]], i32 1123; MINTREESIZE-NEXT: [[C2:%.*]] = extractelement <4 x i32> [[C]], i32 2124; MINTREESIZE-NEXT: [[C3:%.*]] = extractelement <4 x i32> [[C]], i32 3125; MINTREESIZE-NEXT: [[A0:%.*]] = extractelement <4 x float> [[A:%.*]], i32 0126; MINTREESIZE-NEXT: [[A1:%.*]] = extractelement <4 x float> [[A]], i32 1127; MINTREESIZE-NEXT: [[A2:%.*]] = extractelement <4 x float> [[A]], i32 2128; MINTREESIZE-NEXT: [[A3:%.*]] = extractelement <4 x float> [[A]], i32 3129; MINTREESIZE-NEXT: [[B0:%.*]] = extractelement <4 x float> [[B:%.*]], i32 0130; MINTREESIZE-NEXT: [[B1:%.*]] = extractelement <4 x float> [[B]], i32 1131; MINTREESIZE-NEXT: [[B2:%.*]] = extractelement <4 x float> [[B]], i32 2132; MINTREESIZE-NEXT: [[B3:%.*]] = extractelement <4 x float> [[B]], i32 3133; MINTREESIZE-NEXT: [[CMP0:%.*]] = icmp ne i32 [[C0]], 0134; MINTREESIZE-NEXT: [[CMP1:%.*]] = icmp ne i32 [[C1]], 0135; MINTREESIZE-NEXT: [[CMP2:%.*]] = icmp ne i32 [[C2]], 0136; MINTREESIZE-NEXT: [[CMP3:%.*]] = icmp ne i32 [[C3]], 0137; MINTREESIZE-NEXT: [[TMP1:%.*]] = insertelement <4 x i1> poison, i1 [[CMP3]], i32 0138; MINTREESIZE-NEXT: [[TMP2:%.*]] = insertelement <4 x i1> [[TMP1]], i1 [[CMP2]], i32 1139; MINTREESIZE-NEXT: [[TMP3:%.*]] = insertelement <4 x i1> [[TMP2]], i1 [[CMP1]], i32 2140; MINTREESIZE-NEXT: [[TMP4:%.*]] = insertelement <4 x i1> [[TMP3]], i1 [[CMP0]], i32 3141; MINTREESIZE-NEXT: [[S0:%.*]] = select i1 [[CMP0]], float [[A0]], float [[B0]]142; MINTREESIZE-NEXT: [[S1:%.*]] = select i1 [[CMP1]], float [[A1]], float [[B1]]143; MINTREESIZE-NEXT: [[S2:%.*]] = select i1 [[CMP2]], float [[A2]], float [[B2]]144; MINTREESIZE-NEXT: [[S3:%.*]] = select i1 [[CMP3]], float [[A3]], float [[B3]]145; MINTREESIZE-NEXT: [[RA:%.*]] = insertelement <4 x float> poison, float [[S0]], i32 0146; MINTREESIZE-NEXT: [[RB:%.*]] = insertelement <4 x float> [[RA]], float [[S1]], i32 1147; MINTREESIZE-NEXT: [[RC:%.*]] = insertelement <4 x float> [[RB]], float [[S2]], i32 2148; MINTREESIZE-NEXT: [[RD:%.*]] = insertelement <4 x float> [[RC]], float [[S3]], i32 3149; MINTREESIZE-NEXT: [[Q0:%.*]] = extractelement <4 x float> [[RD]], i32 0150; MINTREESIZE-NEXT: [[Q1:%.*]] = extractelement <4 x float> [[RD]], i32 1151; MINTREESIZE-NEXT: [[TMP5:%.*]] = shufflevector <4 x float> [[RD]], <4 x float> poison, <2 x i32> <i32 0, i32 1>152; MINTREESIZE-NEXT: [[Q2:%.*]] = extractelement <4 x float> [[RD]], i32 2153; MINTREESIZE-NEXT: [[Q3:%.*]] = extractelement <4 x float> [[RD]], i32 3154; MINTREESIZE-NEXT: [[TMP6:%.*]] = shufflevector <4 x float> [[RD]], <4 x float> poison, <2 x i32> <i32 2, i32 3>155; MINTREESIZE-NEXT: [[Q4:%.*]] = fadd float [[Q0]], [[Q1]]156; MINTREESIZE-NEXT: [[Q5:%.*]] = fadd float [[Q2]], [[Q3]]157; MINTREESIZE-NEXT: [[TMP7:%.*]] = insertelement <2 x float> poison, float [[Q4]], i32 0158; MINTREESIZE-NEXT: [[TMP8:%.*]] = insertelement <2 x float> [[TMP7]], float [[Q5]], i32 1159; MINTREESIZE-NEXT: [[Q6:%.*]] = fadd float [[Q4]], [[Q5]]160; MINTREESIZE-NEXT: [[QI:%.*]] = fcmp olt float [[Q6]], [[Q5]]161; MINTREESIZE-NEXT: call void @llvm.assume(i1 [[QI]])162; MINTREESIZE-NEXT: ret <4 x float> undef163;164 %c0 = extractelement <4 x i32> %c, i32 0165 %c1 = extractelement <4 x i32> %c, i32 1166 %c2 = extractelement <4 x i32> %c, i32 2167 %c3 = extractelement <4 x i32> %c, i32 3168 %a0 = extractelement <4 x float> %a, i32 0169 %a1 = extractelement <4 x float> %a, i32 1170 %a2 = extractelement <4 x float> %a, i32 2171 %a3 = extractelement <4 x float> %a, i32 3172 %b0 = extractelement <4 x float> %b, i32 0173 %b1 = extractelement <4 x float> %b, i32 1174 %b2 = extractelement <4 x float> %b, i32 2175 %b3 = extractelement <4 x float> %b, i32 3176 %cmp0 = icmp ne i32 %c0, 0177 %cmp1 = icmp ne i32 %c1, 0178 %cmp2 = icmp ne i32 %c2, 0179 %cmp3 = icmp ne i32 %c3, 0180 %s0 = select i1 %cmp0, float %a0, float %b0181 %s1 = select i1 %cmp1, float %a1, float %b1182 %s2 = select i1 %cmp2, float %a2, float %b2183 %s3 = select i1 %cmp3, float %a3, float %b3184 %ra = insertelement <4 x float> poison, float %s0, i32 0185 %rb = insertelement <4 x float> %ra, float %s1, i32 1186 %rc = insertelement <4 x float> %rb, float %s2, i32 2187 %rd = insertelement <4 x float> %rc, float %s3, i32 3188 %q0 = extractelement <4 x float> %rd, i32 0189 %q1 = extractelement <4 x float> %rd, i32 1190 %q2 = extractelement <4 x float> %rd, i32 2191 %q3 = extractelement <4 x float> %rd, i32 3192 %q4 = fadd float %q0, %q1193 %q5 = fadd float %q2, %q3194 %q6 = fadd float %q4, %q5195 %qi = fcmp olt float %q6, %q5196 call void @llvm.assume(i1 %qi)197 ret <4 x float> undef198}199 200; Insert in an order different from the vector indices to make sure it201; doesn't matter202define <4 x float> @simple_select_insert_out_of_order(<4 x float> %a, <4 x float> %b, <4 x i32> %c) {203; CHECK-LABEL: @simple_select_insert_out_of_order(204; CHECK-NEXT: [[TMP1:%.*]] = icmp ne <4 x i32> [[C:%.*]], zeroinitializer205; CHECK-NEXT: [[TMP2:%.*]] = select <4 x i1> [[TMP1]], <4 x float> [[A:%.*]], <4 x float> [[B:%.*]]206; CHECK-NEXT: [[TMP3:%.*]] = shufflevector <4 x float> [[TMP2]], <4 x float> poison, <4 x i32> <i32 2, i32 1, i32 0, i32 3>207; CHECK-NEXT: ret <4 x float> [[TMP3]]208;209 %c0 = extractelement <4 x i32> %c, i32 0210 %c1 = extractelement <4 x i32> %c, i32 1211 %c2 = extractelement <4 x i32> %c, i32 2212 %c3 = extractelement <4 x i32> %c, i32 3213 %a0 = extractelement <4 x float> %a, i32 0214 %a1 = extractelement <4 x float> %a, i32 1215 %a2 = extractelement <4 x float> %a, i32 2216 %a3 = extractelement <4 x float> %a, i32 3217 %b0 = extractelement <4 x float> %b, i32 0218 %b1 = extractelement <4 x float> %b, i32 1219 %b2 = extractelement <4 x float> %b, i32 2220 %b3 = extractelement <4 x float> %b, i32 3221 %cmp0 = icmp ne i32 %c0, 0222 %cmp1 = icmp ne i32 %c1, 0223 %cmp2 = icmp ne i32 %c2, 0224 %cmp3 = icmp ne i32 %c3, 0225 %s0 = select i1 %cmp0, float %a0, float %b0226 %s1 = select i1 %cmp1, float %a1, float %b1227 %s2 = select i1 %cmp2, float %a2, float %b2228 %s3 = select i1 %cmp3, float %a3, float %b3229 %ra = insertelement <4 x float> poison, float %s0, i32 2230 %rb = insertelement <4 x float> %ra, float %s1, i32 1231 %rc = insertelement <4 x float> %rb, float %s2, i32 0232 %rd = insertelement <4 x float> %rc, float %s3, i32 3233 ret <4 x float> %rd234}235 236declare void @v4f32_user(<4 x float>)237declare void @f32_user(float)238 239; Multiple users of the final constructed vector240define <4 x float> @simple_select_users(<4 x float> %a, <4 x float> %b, <4 x i32> %c) {241; CHECK-LABEL: @simple_select_users(242; CHECK-NEXT: [[TMP1:%.*]] = icmp ne <4 x i32> [[C:%.*]], zeroinitializer243; CHECK-NEXT: [[TMP2:%.*]] = select <4 x i1> [[TMP1]], <4 x float> [[A:%.*]], <4 x float> [[B:%.*]]244; CHECK-NEXT: call void @v4f32_user(<4 x float> [[TMP2]])245; CHECK-NEXT: ret <4 x float> [[TMP2]]246;247 %c0 = extractelement <4 x i32> %c, i32 0248 %c1 = extractelement <4 x i32> %c, i32 1249 %c2 = extractelement <4 x i32> %c, i32 2250 %c3 = extractelement <4 x i32> %c, i32 3251 %a0 = extractelement <4 x float> %a, i32 0252 %a1 = extractelement <4 x float> %a, i32 1253 %a2 = extractelement <4 x float> %a, i32 2254 %a3 = extractelement <4 x float> %a, i32 3255 %b0 = extractelement <4 x float> %b, i32 0256 %b1 = extractelement <4 x float> %b, i32 1257 %b2 = extractelement <4 x float> %b, i32 2258 %b3 = extractelement <4 x float> %b, i32 3259 %cmp0 = icmp ne i32 %c0, 0260 %cmp1 = icmp ne i32 %c1, 0261 %cmp2 = icmp ne i32 %c2, 0262 %cmp3 = icmp ne i32 %c3, 0263 %s0 = select i1 %cmp0, float %a0, float %b0264 %s1 = select i1 %cmp1, float %a1, float %b1265 %s2 = select i1 %cmp2, float %a2, float %b2266 %s3 = select i1 %cmp3, float %a3, float %b3267 %ra = insertelement <4 x float> poison, float %s0, i32 0268 %rb = insertelement <4 x float> %ra, float %s1, i32 1269 %rc = insertelement <4 x float> %rb, float %s2, i32 2270 %rd = insertelement <4 x float> %rc, float %s3, i32 3271 call void @v4f32_user(<4 x float> %rd)272 ret <4 x float> %rd273}274 275; Unused insertelement276define <4 x float> @simple_select_no_users(<4 x float> %a, <4 x float> %b, <4 x i32> %c) {277; CHECK-LABEL: @simple_select_no_users(278; CHECK-NEXT: [[TMP1:%.*]] = shufflevector <4 x i32> [[C:%.*]], <4 x i32> poison, <2 x i32> <i32 0, i32 1>279; CHECK-NEXT: [[TMP2:%.*]] = icmp ne <2 x i32> [[TMP1]], zeroinitializer280; CHECK-NEXT: [[TMP3:%.*]] = shufflevector <4 x float> [[A:%.*]], <4 x float> poison, <2 x i32> <i32 0, i32 1>281; CHECK-NEXT: [[TMP4:%.*]] = shufflevector <4 x float> [[B:%.*]], <4 x float> poison, <2 x i32> <i32 0, i32 1>282; CHECK-NEXT: [[TMP5:%.*]] = select <2 x i1> [[TMP2]], <2 x float> [[TMP3]], <2 x float> [[TMP4]]283; CHECK-NEXT: [[TMP6:%.*]] = shufflevector <4 x i32> [[C]], <4 x i32> poison, <2 x i32> <i32 2, i32 3>284; CHECK-NEXT: [[TMP7:%.*]] = icmp ne <2 x i32> [[TMP6]], zeroinitializer285; CHECK-NEXT: [[TMP8:%.*]] = shufflevector <4 x float> [[A]], <4 x float> poison, <2 x i32> <i32 2, i32 3>286; CHECK-NEXT: [[TMP9:%.*]] = shufflevector <4 x float> [[B]], <4 x float> poison, <2 x i32> <i32 2, i32 3>287; CHECK-NEXT: [[TMP10:%.*]] = select <2 x i1> [[TMP7]], <2 x float> [[TMP8]], <2 x float> [[TMP9]]288; CHECK-NEXT: [[TMP11:%.*]] = shufflevector <2 x float> [[TMP5]], <2 x float> poison, <4 x i32> <i32 0, i32 1, i32 poison, i32 poison>289; CHECK-NEXT: [[TMP12:%.*]] = shufflevector <2 x float> [[TMP10]], <2 x float> poison, <4 x i32> <i32 0, i32 1, i32 poison, i32 poison>290; CHECK-NEXT: [[RD1:%.*]] = shufflevector <4 x float> [[TMP12]], <4 x float> poison, <4 x i32> <i32 poison, i32 poison, i32 0, i32 1>291; CHECK-NEXT: ret <4 x float> [[RD1]]292;293 %c0 = extractelement <4 x i32> %c, i32 0294 %c1 = extractelement <4 x i32> %c, i32 1295 %c2 = extractelement <4 x i32> %c, i32 2296 %c3 = extractelement <4 x i32> %c, i32 3297 %a0 = extractelement <4 x float> %a, i32 0298 %a1 = extractelement <4 x float> %a, i32 1299 %a2 = extractelement <4 x float> %a, i32 2300 %a3 = extractelement <4 x float> %a, i32 3301 %b0 = extractelement <4 x float> %b, i32 0302 %b1 = extractelement <4 x float> %b, i32 1303 %b2 = extractelement <4 x float> %b, i32 2304 %b3 = extractelement <4 x float> %b, i32 3305 %cmp0 = icmp ne i32 %c0, 0306 %cmp1 = icmp ne i32 %c1, 0307 %cmp2 = icmp ne i32 %c2, 0308 %cmp3 = icmp ne i32 %c3, 0309 %s0 = select i1 %cmp0, float %a0, float %b0310 %s1 = select i1 %cmp1, float %a1, float %b1311 %s2 = select i1 %cmp2, float %a2, float %b2312 %s3 = select i1 %cmp3, float %a3, float %b3313 %ra = insertelement <4 x float> poison, float %s0, i32 0314 %rb = insertelement <4 x float> %ra, float %s1, i32 1315 %rc = insertelement <4 x float> poison, float %s2, i32 2316 %rd = insertelement <4 x float> %rc, float %s3, i32 3317 ret <4 x float> %rd318}319 320; Make sure infinite loop doesn't happen which I ran into when trying321; to do this backwards this backwards322define <4 x i32> @reconstruct(<4 x i32> %c) {323; CHECK-LABEL: @reconstruct(324; CHECK-NEXT: [[C0:%.*]] = extractelement <4 x i32> [[C:%.*]], i32 0325; CHECK-NEXT: [[C1:%.*]] = extractelement <4 x i32> [[C]], i32 1326; CHECK-NEXT: [[C2:%.*]] = extractelement <4 x i32> [[C]], i32 2327; CHECK-NEXT: [[C3:%.*]] = extractelement <4 x i32> [[C]], i32 3328; CHECK-NEXT: [[RA:%.*]] = insertelement <4 x i32> poison, i32 [[C0]], i32 0329; CHECK-NEXT: [[RB:%.*]] = insertelement <4 x i32> [[RA]], i32 [[C1]], i32 1330; CHECK-NEXT: [[RC:%.*]] = insertelement <4 x i32> [[RB]], i32 [[C2]], i32 2331; CHECK-NEXT: [[RD:%.*]] = insertelement <4 x i32> [[RC]], i32 [[C3]], i32 3332; CHECK-NEXT: ret <4 x i32> [[RD]]333;334 %c0 = extractelement <4 x i32> %c, i32 0335 %c1 = extractelement <4 x i32> %c, i32 1336 %c2 = extractelement <4 x i32> %c, i32 2337 %c3 = extractelement <4 x i32> %c, i32 3338 %ra = insertelement <4 x i32> poison, i32 %c0, i32 0339 %rb = insertelement <4 x i32> %ra, i32 %c1, i32 1340 %rc = insertelement <4 x i32> %rb, i32 %c2, i32 2341 %rd = insertelement <4 x i32> %rc, i32 %c3, i32 3342 ret <4 x i32> %rd343}344 345define <2 x float> @simple_select_v2(<2 x float> %a, <2 x float> %b, <2 x i32> %c) {346; CHECK-LABEL: @simple_select_v2(347; CHECK-NEXT: [[TMP1:%.*]] = icmp ne <2 x i32> [[C:%.*]], zeroinitializer348; CHECK-NEXT: [[TMP2:%.*]] = select <2 x i1> [[TMP1]], <2 x float> [[A:%.*]], <2 x float> [[B:%.*]]349; CHECK-NEXT: ret <2 x float> [[TMP2]]350;351 %c0 = extractelement <2 x i32> %c, i32 0352 %c1 = extractelement <2 x i32> %c, i32 1353 %a0 = extractelement <2 x float> %a, i32 0354 %a1 = extractelement <2 x float> %a, i32 1355 %b0 = extractelement <2 x float> %b, i32 0356 %b1 = extractelement <2 x float> %b, i32 1357 %cmp0 = icmp ne i32 %c0, 0358 %cmp1 = icmp ne i32 %c1, 0359 %s0 = select i1 %cmp0, float %a0, float %b0360 %s1 = select i1 %cmp1, float %a1, float %b1361 %ra = insertelement <2 x float> poison, float %s0, i32 0362 %rb = insertelement <2 x float> %ra, float %s1, i32 1363 ret <2 x float> %rb364}365 366; Make sure when we construct partial vectors, we don't keep367; re-visiting the insertelement chains starting with undef368; (low cost threshold needed to force this to happen)369define <4 x float> @simple_select_partial_vector(<4 x float> %a, <4 x float> %b, <4 x i32> %c) {370; CHECK-LABEL: @simple_select_partial_vector(371; CHECK-NEXT: [[C0:%.*]] = extractelement <4 x i32> [[C:%.*]], i32 0372; CHECK-NEXT: [[C1:%.*]] = extractelement <4 x i32> [[C]], i32 1373; CHECK-NEXT: [[A0:%.*]] = extractelement <4 x float> [[A:%.*]], i32 0374; CHECK-NEXT: [[A1:%.*]] = extractelement <4 x float> [[A]], i32 1375; CHECK-NEXT: [[B0:%.*]] = extractelement <4 x float> [[B:%.*]], i32 0376; CHECK-NEXT: [[B1:%.*]] = extractelement <4 x float> [[B]], i32 1377; CHECK-NEXT: [[TMP1:%.*]] = insertelement <2 x i32> poison, i32 [[C0]], i32 0378; CHECK-NEXT: [[TMP2:%.*]] = insertelement <2 x i32> [[TMP1]], i32 [[C1]], i32 1379; CHECK-NEXT: [[TMP3:%.*]] = icmp ne <2 x i32> [[TMP2]], zeroinitializer380; CHECK-NEXT: [[TMP4:%.*]] = insertelement <2 x float> poison, float [[A0]], i32 0381; CHECK-NEXT: [[TMP5:%.*]] = insertelement <2 x float> [[TMP4]], float [[A1]], i32 1382; CHECK-NEXT: [[TMP6:%.*]] = insertelement <2 x float> poison, float [[B0]], i32 0383; CHECK-NEXT: [[TMP7:%.*]] = insertelement <2 x float> [[TMP6]], float [[B1]], i32 1384; CHECK-NEXT: [[TMP8:%.*]] = select <2 x i1> [[TMP3]], <2 x float> [[TMP5]], <2 x float> [[TMP7]]385; CHECK-NEXT: [[TMP9:%.*]] = extractelement <2 x float> [[TMP8]], i32 0386; CHECK-NEXT: [[RA:%.*]] = insertelement <4 x float> poison, float [[TMP9]], i32 0387; CHECK-NEXT: [[TMP10:%.*]] = extractelement <2 x float> [[TMP8]], i32 1388; CHECK-NEXT: [[RB:%.*]] = insertelement <4 x float> [[RA]], float [[TMP10]], i32 1389; CHECK-NEXT: ret <4 x float> [[RB]]390;391 %c0 = extractelement <4 x i32> %c, i32 0392 %c1 = extractelement <4 x i32> %c, i32 1393 %a0 = extractelement <4 x float> %a, i32 0394 %a1 = extractelement <4 x float> %a, i32 1395 %b0 = extractelement <4 x float> %b, i32 0396 %b1 = extractelement <4 x float> %b, i32 1397 %1 = insertelement <2 x i32> poison, i32 %c0, i32 0398 %2 = insertelement <2 x i32> %1, i32 %c1, i32 1399 %3 = icmp ne <2 x i32> %2, zeroinitializer400 %4 = insertelement <2 x float> poison, float %a0, i32 0401 %5 = insertelement <2 x float> %4, float %a1, i32 1402 %6 = insertelement <2 x float> poison, float %b0, i32 0403 %7 = insertelement <2 x float> %6, float %b1, i32 1404 %8 = select <2 x i1> %3, <2 x float> %5, <2 x float> %7405 %9 = extractelement <2 x float> %8, i32 0406 %ra = insertelement <4 x float> poison, float %9, i32 0407 %10 = extractelement <2 x float> %8, i32 1408 %rb = insertelement <4 x float> %ra, float %10, i32 1409 ret <4 x float> %rb410}411 412; Make sure that vectorization happens even if insertelements operations413; must be rescheduled. The case here is from compiling Julia.414define <4 x float> @reschedule_extract(<4 x float> %a, <4 x float> %b) {415; CHECK-LABEL: @reschedule_extract(416; CHECK-NEXT: [[TMP1:%.*]] = fadd <4 x float> [[A:%.*]], [[B:%.*]]417; CHECK-NEXT: ret <4 x float> [[TMP1]]418;419 %a0 = extractelement <4 x float> %a, i32 0420 %b0 = extractelement <4 x float> %b, i32 0421 %c0 = fadd float %a0, %b0422 %v0 = insertelement <4 x float> poison, float %c0, i32 0423 %a1 = extractelement <4 x float> %a, i32 1424 %b1 = extractelement <4 x float> %b, i32 1425 %c1 = fadd float %a1, %b1426 %v1 = insertelement <4 x float> %v0, float %c1, i32 1427 %a2 = extractelement <4 x float> %a, i32 2428 %b2 = extractelement <4 x float> %b, i32 2429 %c2 = fadd float %a2, %b2430 %v2 = insertelement <4 x float> %v1, float %c2, i32 2431 %a3 = extractelement <4 x float> %a, i32 3432 %b3 = extractelement <4 x float> %b, i32 3433 %c3 = fadd float %a3, %b3434 %v3 = insertelement <4 x float> %v2, float %c3, i32 3435 ret <4 x float> %v3436}437 438; Check that cost model for vectorization takes credit for439; instructions that are erased.440define <4 x float> @take_credit(<4 x float> %a, <4 x float> %b) {441; CHECK-LABEL: @take_credit(442; CHECK-NEXT: [[TMP1:%.*]] = fadd <4 x float> [[A:%.*]], [[B:%.*]]443; CHECK-NEXT: ret <4 x float> [[TMP1]]444;445 %a0 = extractelement <4 x float> %a, i32 0446 %b0 = extractelement <4 x float> %b, i32 0447 %c0 = fadd float %a0, %b0448 %a1 = extractelement <4 x float> %a, i32 1449 %b1 = extractelement <4 x float> %b, i32 1450 %c1 = fadd float %a1, %b1451 %a2 = extractelement <4 x float> %a, i32 2452 %b2 = extractelement <4 x float> %b, i32 2453 %c2 = fadd float %a2, %b2454 %a3 = extractelement <4 x float> %a, i32 3455 %b3 = extractelement <4 x float> %b, i32 3456 %c3 = fadd float %a3, %b3457 %v0 = insertelement <4 x float> poison, float %c0, i32 0458 %v1 = insertelement <4 x float> %v0, float %c1, i32 1459 %v2 = insertelement <4 x float> %v1, float %c2, i32 2460 %v3 = insertelement <4 x float> %v2, float %c3, i32 3461 ret <4 x float> %v3462}463 464; Make sure we handle multiple trees that feed one build vector correctly.465define <4 x double> @multi_tree(double %w, double %x, double %y, double %z) {466; CHECK-LABEL: @multi_tree(467; CHECK-NEXT: [[TMP1:%.*]] = insertelement <4 x double> poison, double [[Z:%.*]], i32 0468; CHECK-NEXT: [[TMP2:%.*]] = insertelement <4 x double> [[TMP1]], double [[Y:%.*]], i32 1469; CHECK-NEXT: [[TMP3:%.*]] = insertelement <4 x double> [[TMP2]], double [[X:%.*]], i32 2470; CHECK-NEXT: [[TMP4:%.*]] = insertelement <4 x double> [[TMP3]], double [[W:%.*]], i32 3471; CHECK-NEXT: [[TMP5:%.*]] = fadd <4 x double> [[TMP4]], <double 3.000000e+00, double 2.000000e+00, double 1.000000e+00, double 0.000000e+00>472; CHECK-NEXT: [[TMP6:%.*]] = fmul <4 x double> [[TMP5]], splat (double 1.000000e+00)473; CHECK-NEXT: ret <4 x double> [[TMP6]]474;475 %t0 = fadd double %w , 0.000000e+00476 %t1 = fadd double %x , 1.000000e+00477 %t2 = fadd double %y , 2.000000e+00478 %t3 = fadd double %z , 3.000000e+00479 %t4 = fmul double %t0, 1.000000e+00480 %i1 = insertelement <4 x double> poison, double %t4, i32 3481 %t5 = fmul double %t1, 1.000000e+00482 %i2 = insertelement <4 x double> %i1, double %t5, i32 2483 %t6 = fmul double %t2, 1.000000e+00484 %i3 = insertelement <4 x double> %i2, double %t6, i32 1485 %t7 = fmul double %t3, 1.000000e+00486 %i4 = insertelement <4 x double> %i3, double %t7, i32 0487 ret <4 x double> %i4488}489 490define <8 x float> @_vadd256(<8 x float> %a, <8 x float> %b) local_unnamed_addr {491; CHECK-LABEL: @_vadd256(492; CHECK-NEXT: [[TMP1:%.*]] = fadd <8 x float> [[A:%.*]], [[B:%.*]]493; CHECK-NEXT: ret <8 x float> [[TMP1]]494;495 %vecext = extractelement <8 x float> %a, i32 0496 %vecext1 = extractelement <8 x float> %b, i32 0497 %add = fadd float %vecext, %vecext1498 %vecext2 = extractelement <8 x float> %a, i32 1499 %vecext3 = extractelement <8 x float> %b, i32 1500 %add4 = fadd float %vecext2, %vecext3501 %vecext5 = extractelement <8 x float> %a, i32 2502 %vecext6 = extractelement <8 x float> %b, i32 2503 %add7 = fadd float %vecext5, %vecext6504 %vecext8 = extractelement <8 x float> %a, i32 3505 %vecext9 = extractelement <8 x float> %b, i32 3506 %add10 = fadd float %vecext8, %vecext9507 %vecext11 = extractelement <8 x float> %a, i32 4508 %vecext12 = extractelement <8 x float> %b, i32 4509 %add13 = fadd float %vecext11, %vecext12510 %vecext14 = extractelement <8 x float> %a, i32 5511 %vecext15 = extractelement <8 x float> %b, i32 5512 %add16 = fadd float %vecext14, %vecext15513 %vecext17 = extractelement <8 x float> %a, i32 6514 %vecext18 = extractelement <8 x float> %b, i32 6515 %add19 = fadd float %vecext17, %vecext18516 %vecext20 = extractelement <8 x float> %a, i32 7517 %vecext21 = extractelement <8 x float> %b, i32 7518 %add22 = fadd float %vecext20, %vecext21519 %vecinit.i = insertelement <8 x float> poison, float %add, i32 0520 %vecinit1.i = insertelement <8 x float> %vecinit.i, float %add4, i32 1521 %vecinit2.i = insertelement <8 x float> %vecinit1.i, float %add7, i32 2522 %vecinit3.i = insertelement <8 x float> %vecinit2.i, float %add10, i32 3523 %vecinit4.i = insertelement <8 x float> %vecinit3.i, float %add13, i32 4524 %vecinit5.i = insertelement <8 x float> %vecinit4.i, float %add16, i32 5525 %vecinit6.i = insertelement <8 x float> %vecinit5.i, float %add19, i32 6526 %vecinit7.i = insertelement <8 x float> %vecinit6.i, float %add22, i32 7527 ret <8 x float> %vecinit7.i528}529