565 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> zeroinitializer, 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 42define <8 x float> @simple_select2(<4 x float> %a, <4 x float> %b, <4 x i32> %c) {43; CHECK-LABEL: @simple_select2(44; CHECK-NEXT: [[TMP1:%.*]] = icmp ne <4 x i32> [[C:%.*]], zeroinitializer45; CHECK-NEXT: [[TMP2:%.*]] = select <4 x i1> [[TMP1]], <4 x float> [[A:%.*]], <4 x float> [[B:%.*]]46; CHECK-NEXT: [[TMP4:%.*]] = shufflevector <4 x float> [[TMP2]], <4 x float> poison, <8 x i32> <i32 0, i32 poison, i32 1, i32 poison, i32 2, i32 poison, i32 poison, i32 3>47; CHECK-NEXT: [[TMP3:%.*]] = shufflevector <8 x float> zeroinitializer, <8 x float> [[TMP4]], <8 x i32> <i32 8, i32 1, i32 10, i32 3, i32 12, i32 5, i32 6, i32 15>48; CHECK-NEXT: ret <8 x float> [[TMP3]]49;50 %c0 = extractelement <4 x i32> %c, i32 051 %c1 = extractelement <4 x i32> %c, i32 152 %c2 = extractelement <4 x i32> %c, i32 253 %c3 = extractelement <4 x i32> %c, i32 354 %a0 = extractelement <4 x float> %a, i32 055 %a1 = extractelement <4 x float> %a, i32 156 %a2 = extractelement <4 x float> %a, i32 257 %a3 = extractelement <4 x float> %a, i32 358 %b0 = extractelement <4 x float> %b, i32 059 %b1 = extractelement <4 x float> %b, i32 160 %b2 = extractelement <4 x float> %b, i32 261 %b3 = extractelement <4 x float> %b, i32 362 %cmp0 = icmp ne i32 %c0, 063 %cmp1 = icmp ne i32 %c1, 064 %cmp2 = icmp ne i32 %c2, 065 %cmp3 = icmp ne i32 %c3, 066 %s0 = select i1 %cmp0, float %a0, float %b067 %s1 = select i1 %cmp1, float %a1, float %b168 %s2 = select i1 %cmp2, float %a2, float %b269 %s3 = select i1 %cmp3, float %a3, float %b370 %ra = insertelement <8 x float> zeroinitializer, float %s0, i32 071 %rb = insertelement <8 x float> %ra, float %s1, i32 272 %rc = insertelement <8 x float> %rb, float %s2, i32 473 %rd = insertelement <8 x float> %rc, float %s3, i32 774 ret <8 x float> %rd75}76 77declare void @llvm.assume(i1) nounwind78 79; This entire tree is ephemeral, don't vectorize any of it.80define <4 x float> @simple_select_eph(<4 x float> %a, <4 x float> %b, <4 x i32> %c) {81; THRESHOLD-LABEL: @simple_select_eph(82; THRESHOLD-NEXT: [[C0:%.*]] = extractelement <4 x i32> [[C:%.*]], i32 083; THRESHOLD-NEXT: [[C1:%.*]] = extractelement <4 x i32> [[C]], i32 184; THRESHOLD-NEXT: [[C2:%.*]] = extractelement <4 x i32> [[C]], i32 285; THRESHOLD-NEXT: [[C3:%.*]] = extractelement <4 x i32> [[C]], i32 386; THRESHOLD-NEXT: [[A0:%.*]] = extractelement <4 x float> [[A:%.*]], i32 087; THRESHOLD-NEXT: [[A1:%.*]] = extractelement <4 x float> [[A]], i32 188; THRESHOLD-NEXT: [[A2:%.*]] = extractelement <4 x float> [[A]], i32 289; THRESHOLD-NEXT: [[A3:%.*]] = extractelement <4 x float> [[A]], i32 390; THRESHOLD-NEXT: [[B0:%.*]] = extractelement <4 x float> [[B:%.*]], i32 091; THRESHOLD-NEXT: [[B1:%.*]] = extractelement <4 x float> [[B]], i32 192; THRESHOLD-NEXT: [[B2:%.*]] = extractelement <4 x float> [[B]], i32 293; THRESHOLD-NEXT: [[B3:%.*]] = extractelement <4 x float> [[B]], i32 394; THRESHOLD-NEXT: [[CMP0:%.*]] = icmp ne i32 [[C0]], 095; THRESHOLD-NEXT: [[CMP1:%.*]] = icmp ne i32 [[C1]], 096; THRESHOLD-NEXT: [[CMP2:%.*]] = icmp ne i32 [[C2]], 097; THRESHOLD-NEXT: [[CMP3:%.*]] = icmp ne i32 [[C3]], 098; THRESHOLD-NEXT: [[S0:%.*]] = select i1 [[CMP0]], float [[A0]], float [[B0]]99; THRESHOLD-NEXT: [[S1:%.*]] = select i1 [[CMP1]], float [[A1]], float [[B1]]100; THRESHOLD-NEXT: [[S2:%.*]] = select i1 [[CMP2]], float [[A2]], float [[B2]]101; THRESHOLD-NEXT: [[S3:%.*]] = select i1 [[CMP3]], float [[A3]], float [[B3]]102; THRESHOLD-NEXT: [[RA:%.*]] = insertelement <4 x float> zeroinitializer, float [[S0]], i32 0103; THRESHOLD-NEXT: [[RB:%.*]] = insertelement <4 x float> [[RA]], float [[S1]], i32 1104; THRESHOLD-NEXT: [[RC:%.*]] = insertelement <4 x float> [[RB]], float [[S2]], i32 2105; THRESHOLD-NEXT: [[RD:%.*]] = insertelement <4 x float> [[RC]], float [[S3]], i32 3106; THRESHOLD-NEXT: [[Q0:%.*]] = extractelement <4 x float> [[RD]], i32 0107; THRESHOLD-NEXT: [[Q1:%.*]] = extractelement <4 x float> [[RD]], i32 1108; THRESHOLD-NEXT: [[TMP1:%.*]] = shufflevector <4 x float> [[RD]], <4 x float> poison, <2 x i32> <i32 0, i32 1>109; THRESHOLD-NEXT: [[Q2:%.*]] = extractelement <4 x float> [[RD]], i32 2110; THRESHOLD-NEXT: [[Q3:%.*]] = extractelement <4 x float> [[RD]], i32 3111; THRESHOLD-NEXT: [[TMP2:%.*]] = shufflevector <4 x float> [[RD]], <4 x float> poison, <2 x i32> <i32 2, i32 3>112; THRESHOLD-NEXT: [[Q4:%.*]] = fadd float [[Q0]], [[Q1]]113; THRESHOLD-NEXT: [[Q5:%.*]] = fadd float [[Q2]], [[Q3]]114; THRESHOLD-NEXT: [[Q6:%.*]] = fadd float [[Q4]], [[Q5]]115; THRESHOLD-NEXT: [[QI:%.*]] = fcmp olt float [[Q6]], [[Q5]]116; THRESHOLD-NEXT: call void @llvm.assume(i1 [[QI]])117; THRESHOLD-NEXT: ret <4 x float> zeroinitializer118;119; NOTHRESHOLD-LABEL: @simple_select_eph(120; NOTHRESHOLD-NEXT: [[C0:%.*]] = extractelement <4 x i32> [[C:%.*]], i32 0121; NOTHRESHOLD-NEXT: [[C1:%.*]] = extractelement <4 x i32> [[C]], i32 1122; NOTHRESHOLD-NEXT: [[C2:%.*]] = extractelement <4 x i32> [[C]], i32 2123; NOTHRESHOLD-NEXT: [[C3:%.*]] = extractelement <4 x i32> [[C]], i32 3124; NOTHRESHOLD-NEXT: [[A0:%.*]] = extractelement <4 x float> [[A:%.*]], i32 0125; NOTHRESHOLD-NEXT: [[A1:%.*]] = extractelement <4 x float> [[A]], i32 1126; NOTHRESHOLD-NEXT: [[A2:%.*]] = extractelement <4 x float> [[A]], i32 2127; NOTHRESHOLD-NEXT: [[A3:%.*]] = extractelement <4 x float> [[A]], i32 3128; NOTHRESHOLD-NEXT: [[B0:%.*]] = extractelement <4 x float> [[B:%.*]], i32 0129; NOTHRESHOLD-NEXT: [[B1:%.*]] = extractelement <4 x float> [[B]], i32 1130; NOTHRESHOLD-NEXT: [[B2:%.*]] = extractelement <4 x float> [[B]], i32 2131; NOTHRESHOLD-NEXT: [[B3:%.*]] = extractelement <4 x float> [[B]], i32 3132; NOTHRESHOLD-NEXT: [[CMP0:%.*]] = icmp ne i32 [[C0]], 0133; NOTHRESHOLD-NEXT: [[CMP1:%.*]] = icmp ne i32 [[C1]], 0134; NOTHRESHOLD-NEXT: [[CMP2:%.*]] = icmp ne i32 [[C2]], 0135; NOTHRESHOLD-NEXT: [[CMP3:%.*]] = icmp ne i32 [[C3]], 0136; NOTHRESHOLD-NEXT: [[S0:%.*]] = select i1 [[CMP0]], float [[A0]], float [[B0]]137; NOTHRESHOLD-NEXT: [[S1:%.*]] = select i1 [[CMP1]], float [[A1]], float [[B1]]138; NOTHRESHOLD-NEXT: [[S2:%.*]] = select i1 [[CMP2]], float [[A2]], float [[B2]]139; NOTHRESHOLD-NEXT: [[S3:%.*]] = select i1 [[CMP3]], float [[A3]], float [[B3]]140; NOTHRESHOLD-NEXT: [[RA:%.*]] = insertelement <4 x float> zeroinitializer, float [[S0]], i32 0141; NOTHRESHOLD-NEXT: [[RB:%.*]] = insertelement <4 x float> [[RA]], float [[S1]], i32 1142; NOTHRESHOLD-NEXT: [[RC:%.*]] = insertelement <4 x float> [[RB]], float [[S2]], i32 2143; NOTHRESHOLD-NEXT: [[RD:%.*]] = insertelement <4 x float> [[RC]], float [[S3]], i32 3144; NOTHRESHOLD-NEXT: [[Q0:%.*]] = extractelement <4 x float> [[RD]], i32 0145; NOTHRESHOLD-NEXT: [[Q1:%.*]] = extractelement <4 x float> [[RD]], i32 1146; NOTHRESHOLD-NEXT: [[Q2:%.*]] = extractelement <4 x float> [[RD]], i32 2147; NOTHRESHOLD-NEXT: [[Q3:%.*]] = extractelement <4 x float> [[RD]], i32 3148; NOTHRESHOLD-NEXT: [[Q4:%.*]] = fadd float [[Q0]], [[Q1]]149; NOTHRESHOLD-NEXT: [[Q5:%.*]] = fadd float [[Q2]], [[Q3]]150; NOTHRESHOLD-NEXT: [[Q6:%.*]] = fadd float [[Q4]], [[Q5]]151; NOTHRESHOLD-NEXT: [[QI:%.*]] = fcmp olt float [[Q6]], [[Q5]]152; NOTHRESHOLD-NEXT: call void @llvm.assume(i1 [[QI]])153; NOTHRESHOLD-NEXT: ret <4 x float> zeroinitializer154;155; MINTREESIZE-LABEL: @simple_select_eph(156; MINTREESIZE-NEXT: [[C0:%.*]] = extractelement <4 x i32> [[C:%.*]], i32 0157; MINTREESIZE-NEXT: [[C1:%.*]] = extractelement <4 x i32> [[C]], i32 1158; MINTREESIZE-NEXT: [[C2:%.*]] = extractelement <4 x i32> [[C]], i32 2159; MINTREESIZE-NEXT: [[C3:%.*]] = extractelement <4 x i32> [[C]], i32 3160; MINTREESIZE-NEXT: [[A0:%.*]] = extractelement <4 x float> [[A:%.*]], i32 0161; MINTREESIZE-NEXT: [[A1:%.*]] = extractelement <4 x float> [[A]], i32 1162; MINTREESIZE-NEXT: [[A2:%.*]] = extractelement <4 x float> [[A]], i32 2163; MINTREESIZE-NEXT: [[A3:%.*]] = extractelement <4 x float> [[A]], i32 3164; MINTREESIZE-NEXT: [[B0:%.*]] = extractelement <4 x float> [[B:%.*]], i32 0165; MINTREESIZE-NEXT: [[B1:%.*]] = extractelement <4 x float> [[B]], i32 1166; MINTREESIZE-NEXT: [[B2:%.*]] = extractelement <4 x float> [[B]], i32 2167; MINTREESIZE-NEXT: [[B3:%.*]] = extractelement <4 x float> [[B]], i32 3168; MINTREESIZE-NEXT: [[CMP0:%.*]] = icmp ne i32 [[C0]], 0169; MINTREESIZE-NEXT: [[CMP1:%.*]] = icmp ne i32 [[C1]], 0170; MINTREESIZE-NEXT: [[CMP2:%.*]] = icmp ne i32 [[C2]], 0171; MINTREESIZE-NEXT: [[CMP3:%.*]] = icmp ne i32 [[C3]], 0172; MINTREESIZE-NEXT: [[TMP1:%.*]] = insertelement <4 x i1> poison, i1 [[CMP3]], i32 0173; MINTREESIZE-NEXT: [[TMP2:%.*]] = insertelement <4 x i1> [[TMP1]], i1 [[CMP2]], i32 1174; MINTREESIZE-NEXT: [[TMP3:%.*]] = insertelement <4 x i1> [[TMP2]], i1 [[CMP1]], i32 2175; MINTREESIZE-NEXT: [[TMP4:%.*]] = insertelement <4 x i1> [[TMP3]], i1 [[CMP0]], i32 3176; MINTREESIZE-NEXT: [[S0:%.*]] = select i1 [[CMP0]], float [[A0]], float [[B0]]177; MINTREESIZE-NEXT: [[S1:%.*]] = select i1 [[CMP1]], float [[A1]], float [[B1]]178; MINTREESIZE-NEXT: [[S2:%.*]] = select i1 [[CMP2]], float [[A2]], float [[B2]]179; MINTREESIZE-NEXT: [[S3:%.*]] = select i1 [[CMP3]], float [[A3]], float [[B3]]180; MINTREESIZE-NEXT: [[RA:%.*]] = insertelement <4 x float> zeroinitializer, float [[S0]], i32 0181; MINTREESIZE-NEXT: [[RB:%.*]] = insertelement <4 x float> [[RA]], float [[S1]], i32 1182; MINTREESIZE-NEXT: [[RC:%.*]] = insertelement <4 x float> [[RB]], float [[S2]], i32 2183; MINTREESIZE-NEXT: [[RD:%.*]] = insertelement <4 x float> [[RC]], float [[S3]], i32 3184; MINTREESIZE-NEXT: [[Q0:%.*]] = extractelement <4 x float> [[RD]], i32 0185; MINTREESIZE-NEXT: [[Q1:%.*]] = extractelement <4 x float> [[RD]], i32 1186; MINTREESIZE-NEXT: [[TMP5:%.*]] = shufflevector <4 x float> [[RD]], <4 x float> poison, <2 x i32> <i32 0, i32 1>187; MINTREESIZE-NEXT: [[Q2:%.*]] = extractelement <4 x float> [[RD]], i32 2188; MINTREESIZE-NEXT: [[Q3:%.*]] = extractelement <4 x float> [[RD]], i32 3189; MINTREESIZE-NEXT: [[TMP6:%.*]] = shufflevector <4 x float> [[RD]], <4 x float> poison, <2 x i32> <i32 2, i32 3>190; MINTREESIZE-NEXT: [[Q4:%.*]] = fadd float [[Q0]], [[Q1]]191; MINTREESIZE-NEXT: [[Q5:%.*]] = fadd float [[Q2]], [[Q3]]192; MINTREESIZE-NEXT: [[TMP7:%.*]] = insertelement <2 x float> poison, float [[Q4]], i32 0193; MINTREESIZE-NEXT: [[TMP8:%.*]] = insertelement <2 x float> [[TMP7]], float [[Q5]], i32 1194; MINTREESIZE-NEXT: [[Q6:%.*]] = fadd float [[Q4]], [[Q5]]195; MINTREESIZE-NEXT: [[QI:%.*]] = fcmp olt float [[Q6]], [[Q5]]196; MINTREESIZE-NEXT: call void @llvm.assume(i1 [[QI]])197; MINTREESIZE-NEXT: ret <4 x float> zeroinitializer198;199 %c0 = extractelement <4 x i32> %c, i32 0200 %c1 = extractelement <4 x i32> %c, i32 1201 %c2 = extractelement <4 x i32> %c, i32 2202 %c3 = extractelement <4 x i32> %c, i32 3203 %a0 = extractelement <4 x float> %a, i32 0204 %a1 = extractelement <4 x float> %a, i32 1205 %a2 = extractelement <4 x float> %a, i32 2206 %a3 = extractelement <4 x float> %a, i32 3207 %b0 = extractelement <4 x float> %b, i32 0208 %b1 = extractelement <4 x float> %b, i32 1209 %b2 = extractelement <4 x float> %b, i32 2210 %b3 = extractelement <4 x float> %b, i32 3211 %cmp0 = icmp ne i32 %c0, 0212 %cmp1 = icmp ne i32 %c1, 0213 %cmp2 = icmp ne i32 %c2, 0214 %cmp3 = icmp ne i32 %c3, 0215 %s0 = select i1 %cmp0, float %a0, float %b0216 %s1 = select i1 %cmp1, float %a1, float %b1217 %s2 = select i1 %cmp2, float %a2, float %b2218 %s3 = select i1 %cmp3, float %a3, float %b3219 %ra = insertelement <4 x float> zeroinitializer, float %s0, i32 0220 %rb = insertelement <4 x float> %ra, float %s1, i32 1221 %rc = insertelement <4 x float> %rb, float %s2, i32 2222 %rd = insertelement <4 x float> %rc, float %s3, i32 3223 %q0 = extractelement <4 x float> %rd, i32 0224 %q1 = extractelement <4 x float> %rd, i32 1225 %q2 = extractelement <4 x float> %rd, i32 2226 %q3 = extractelement <4 x float> %rd, i32 3227 %q4 = fadd float %q0, %q1228 %q5 = fadd float %q2, %q3229 %q6 = fadd float %q4, %q5230 %qi = fcmp olt float %q6, %q5231 call void @llvm.assume(i1 %qi)232 ret <4 x float> zeroinitializer233}234 235; Insert in an order different from the vector indices to make sure it236; doesn't matter237define <4 x float> @simple_select_insert_out_of_order(<4 x float> %a, <4 x float> %b, <4 x i32> %c) {238; CHECK-LABEL: @simple_select_insert_out_of_order(239; CHECK-NEXT: [[TMP1:%.*]] = icmp ne <4 x i32> [[C:%.*]], zeroinitializer240; CHECK-NEXT: [[TMP2:%.*]] = select <4 x i1> [[TMP1]], <4 x float> [[A:%.*]], <4 x float> [[B:%.*]]241; CHECK-NEXT: [[TMP3:%.*]] = shufflevector <4 x float> [[TMP2]], <4 x float> poison, <4 x i32> <i32 2, i32 1, i32 0, i32 3>242; CHECK-NEXT: ret <4 x float> [[TMP3]]243;244 %c0 = extractelement <4 x i32> %c, i32 0245 %c1 = extractelement <4 x i32> %c, i32 1246 %c2 = extractelement <4 x i32> %c, i32 2247 %c3 = extractelement <4 x i32> %c, i32 3248 %a0 = extractelement <4 x float> %a, i32 0249 %a1 = extractelement <4 x float> %a, i32 1250 %a2 = extractelement <4 x float> %a, i32 2251 %a3 = extractelement <4 x float> %a, i32 3252 %b0 = extractelement <4 x float> %b, i32 0253 %b1 = extractelement <4 x float> %b, i32 1254 %b2 = extractelement <4 x float> %b, i32 2255 %b3 = extractelement <4 x float> %b, i32 3256 %cmp0 = icmp ne i32 %c0, 0257 %cmp1 = icmp ne i32 %c1, 0258 %cmp2 = icmp ne i32 %c2, 0259 %cmp3 = icmp ne i32 %c3, 0260 %s0 = select i1 %cmp0, float %a0, float %b0261 %s1 = select i1 %cmp1, float %a1, float %b1262 %s2 = select i1 %cmp2, float %a2, float %b2263 %s3 = select i1 %cmp3, float %a3, float %b3264 %ra = insertelement <4 x float> zeroinitializer, float %s0, i32 2265 %rb = insertelement <4 x float> %ra, float %s1, i32 1266 %rc = insertelement <4 x float> %rb, float %s2, i32 0267 %rd = insertelement <4 x float> %rc, float %s3, i32 3268 ret <4 x float> %rd269}270 271declare void @v4f32_user(<4 x float>)272declare void @f32_user(float)273 274; Multiple users of the final constructed vector275define <4 x float> @simple_select_users(<4 x float> %a, <4 x float> %b, <4 x i32> %c) {276; CHECK-LABEL: @simple_select_users(277; CHECK-NEXT: [[TMP1:%.*]] = icmp ne <4 x i32> [[C:%.*]], zeroinitializer278; CHECK-NEXT: [[TMP2:%.*]] = select <4 x i1> [[TMP1]], <4 x float> [[A:%.*]], <4 x float> [[B:%.*]]279; CHECK-NEXT: call void @v4f32_user(<4 x float> [[TMP2]])280; CHECK-NEXT: ret <4 x float> [[TMP2]]281;282 %c0 = extractelement <4 x i32> %c, i32 0283 %c1 = extractelement <4 x i32> %c, i32 1284 %c2 = extractelement <4 x i32> %c, i32 2285 %c3 = extractelement <4 x i32> %c, i32 3286 %a0 = extractelement <4 x float> %a, i32 0287 %a1 = extractelement <4 x float> %a, i32 1288 %a2 = extractelement <4 x float> %a, i32 2289 %a3 = extractelement <4 x float> %a, i32 3290 %b0 = extractelement <4 x float> %b, i32 0291 %b1 = extractelement <4 x float> %b, i32 1292 %b2 = extractelement <4 x float> %b, i32 2293 %b3 = extractelement <4 x float> %b, i32 3294 %cmp0 = icmp ne i32 %c0, 0295 %cmp1 = icmp ne i32 %c1, 0296 %cmp2 = icmp ne i32 %c2, 0297 %cmp3 = icmp ne i32 %c3, 0298 %s0 = select i1 %cmp0, float %a0, float %b0299 %s1 = select i1 %cmp1, float %a1, float %b1300 %s2 = select i1 %cmp2, float %a2, float %b2301 %s3 = select i1 %cmp3, float %a3, float %b3302 %ra = insertelement <4 x float> zeroinitializer, float %s0, i32 0303 %rb = insertelement <4 x float> %ra, float %s1, i32 1304 %rc = insertelement <4 x float> %rb, float %s2, i32 2305 %rd = insertelement <4 x float> %rc, float %s3, i32 3306 call void @v4f32_user(<4 x float> %rd)307 ret <4 x float> %rd308}309 310; Unused insertelement311define <4 x float> @simple_select_no_users(<4 x float> %a, <4 x float> %b, <4 x i32> %c) {312; CHECK-LABEL: @simple_select_no_users(313; CHECK-NEXT: [[TMP1:%.*]] = shufflevector <4 x i32> [[C:%.*]], <4 x i32> poison, <2 x i32> <i32 0, i32 1>314; CHECK-NEXT: [[TMP2:%.*]] = icmp ne <2 x i32> [[TMP1]], zeroinitializer315; CHECK-NEXT: [[TMP3:%.*]] = shufflevector <4 x float> [[A:%.*]], <4 x float> poison, <2 x i32> <i32 0, i32 1>316; CHECK-NEXT: [[TMP4:%.*]] = shufflevector <4 x float> [[B:%.*]], <4 x float> poison, <2 x i32> <i32 0, i32 1>317; CHECK-NEXT: [[TMP5:%.*]] = select <2 x i1> [[TMP2]], <2 x float> [[TMP3]], <2 x float> [[TMP4]]318; CHECK-NEXT: [[TMP6:%.*]] = shufflevector <4 x i32> [[C]], <4 x i32> poison, <2 x i32> <i32 2, i32 3>319; CHECK-NEXT: [[TMP7:%.*]] = icmp ne <2 x i32> [[TMP6]], zeroinitializer320; CHECK-NEXT: [[TMP8:%.*]] = shufflevector <4 x float> [[A]], <4 x float> poison, <2 x i32> <i32 2, i32 3>321; CHECK-NEXT: [[TMP9:%.*]] = shufflevector <4 x float> [[B]], <4 x float> poison, <2 x i32> <i32 2, i32 3>322; CHECK-NEXT: [[TMP10:%.*]] = select <2 x i1> [[TMP7]], <2 x float> [[TMP8]], <2 x float> [[TMP9]]323; CHECK-NEXT: [[TMP11:%.*]] = shufflevector <2 x float> [[TMP5]], <2 x float> poison, <4 x i32> <i32 0, i32 1, i32 poison, i32 poison>324; CHECK-NEXT: [[RB2:%.*]] = shufflevector <4 x float> zeroinitializer, <4 x float> [[TMP11]], <4 x i32> <i32 4, i32 5, i32 2, i32 3>325; CHECK-NEXT: [[TMP12:%.*]] = shufflevector <2 x float> [[TMP10]], <2 x float> poison, <4 x i32> <i32 0, i32 1, i32 poison, i32 poison>326; CHECK-NEXT: [[RD1:%.*]] = shufflevector <4 x float> zeroinitializer, <4 x float> [[TMP12]], <4 x i32> <i32 0, i32 1, i32 4, i32 5>327; CHECK-NEXT: ret <4 x float> [[RD1]]328;329 %c0 = extractelement <4 x i32> %c, i32 0330 %c1 = extractelement <4 x i32> %c, i32 1331 %c2 = extractelement <4 x i32> %c, i32 2332 %c3 = extractelement <4 x i32> %c, i32 3333 %a0 = extractelement <4 x float> %a, i32 0334 %a1 = extractelement <4 x float> %a, i32 1335 %a2 = extractelement <4 x float> %a, i32 2336 %a3 = extractelement <4 x float> %a, i32 3337 %b0 = extractelement <4 x float> %b, i32 0338 %b1 = extractelement <4 x float> %b, i32 1339 %b2 = extractelement <4 x float> %b, i32 2340 %b3 = extractelement <4 x float> %b, i32 3341 %cmp0 = icmp ne i32 %c0, 0342 %cmp1 = icmp ne i32 %c1, 0343 %cmp2 = icmp ne i32 %c2, 0344 %cmp3 = icmp ne i32 %c3, 0345 %s0 = select i1 %cmp0, float %a0, float %b0346 %s1 = select i1 %cmp1, float %a1, float %b1347 %s2 = select i1 %cmp2, float %a2, float %b2348 %s3 = select i1 %cmp3, float %a3, float %b3349 %ra = insertelement <4 x float> zeroinitializer, float %s0, i32 0350 %rb = insertelement <4 x float> %ra, float %s1, i32 1351 %rc = insertelement <4 x float> zeroinitializer, float %s2, i32 2352 %rd = insertelement <4 x float> %rc, float %s3, i32 3353 ret <4 x float> %rd354}355 356; Make sure infinite loop doesn't happen which I ran into when trying357; to do this backwards this backwards358define <4 x i32> @reconstruct(<4 x i32> %c) {359; CHECK-LABEL: @reconstruct(360; CHECK-NEXT: [[C0:%.*]] = extractelement <4 x i32> [[C:%.*]], i32 0361; CHECK-NEXT: [[C1:%.*]] = extractelement <4 x i32> [[C]], i32 1362; CHECK-NEXT: [[C2:%.*]] = extractelement <4 x i32> [[C]], i32 2363; CHECK-NEXT: [[C3:%.*]] = extractelement <4 x i32> [[C]], i32 3364; CHECK-NEXT: [[RA:%.*]] = insertelement <4 x i32> zeroinitializer, i32 [[C0]], i32 0365; CHECK-NEXT: [[RB:%.*]] = insertelement <4 x i32> [[RA]], i32 [[C1]], i32 1366; CHECK-NEXT: [[RC:%.*]] = insertelement <4 x i32> [[RB]], i32 [[C2]], i32 2367; CHECK-NEXT: [[RD:%.*]] = insertelement <4 x i32> [[RC]], i32 [[C3]], i32 3368; CHECK-NEXT: ret <4 x i32> [[RD]]369;370 %c0 = extractelement <4 x i32> %c, i32 0371 %c1 = extractelement <4 x i32> %c, i32 1372 %c2 = extractelement <4 x i32> %c, i32 2373 %c3 = extractelement <4 x i32> %c, i32 3374 %ra = insertelement <4 x i32> zeroinitializer, i32 %c0, i32 0375 %rb = insertelement <4 x i32> %ra, i32 %c1, i32 1376 %rc = insertelement <4 x i32> %rb, i32 %c2, i32 2377 %rd = insertelement <4 x i32> %rc, i32 %c3, i32 3378 ret <4 x i32> %rd379}380 381define <2 x float> @simple_select_v2(<2 x float> %a, <2 x float> %b, <2 x i32> %c) {382; CHECK-LABEL: @simple_select_v2(383; CHECK-NEXT: [[TMP1:%.*]] = icmp ne <2 x i32> [[C:%.*]], zeroinitializer384; CHECK-NEXT: [[TMP2:%.*]] = select <2 x i1> [[TMP1]], <2 x float> [[A:%.*]], <2 x float> [[B:%.*]]385; CHECK-NEXT: ret <2 x float> [[TMP2]]386;387 %c0 = extractelement <2 x i32> %c, i32 0388 %c1 = extractelement <2 x i32> %c, i32 1389 %a0 = extractelement <2 x float> %a, i32 0390 %a1 = extractelement <2 x float> %a, i32 1391 %b0 = extractelement <2 x float> %b, i32 0392 %b1 = extractelement <2 x float> %b, i32 1393 %cmp0 = icmp ne i32 %c0, 0394 %cmp1 = icmp ne i32 %c1, 0395 %s0 = select i1 %cmp0, float %a0, float %b0396 %s1 = select i1 %cmp1, float %a1, float %b1397 %ra = insertelement <2 x float> zeroinitializer, float %s0, i32 0398 %rb = insertelement <2 x float> %ra, float %s1, i32 1399 ret <2 x float> %rb400}401 402; Make sure when we construct partial vectors, we don't keep403; re-visiting the insertelement chains starting with zeroinitializer404; (low cost threshold needed to force this to happen)405define <4 x float> @simple_select_partial_vector(<4 x float> %a, <4 x float> %b, <4 x i32> %c) {406; CHECK-LABEL: @simple_select_partial_vector(407; CHECK-NEXT: [[C0:%.*]] = extractelement <4 x i32> [[C:%.*]], i32 0408; CHECK-NEXT: [[C1:%.*]] = extractelement <4 x i32> [[C]], i32 1409; CHECK-NEXT: [[A0:%.*]] = extractelement <4 x float> [[A:%.*]], i32 0410; CHECK-NEXT: [[A1:%.*]] = extractelement <4 x float> [[A]], i32 1411; CHECK-NEXT: [[B0:%.*]] = extractelement <4 x float> [[B:%.*]], i32 0412; CHECK-NEXT: [[B1:%.*]] = extractelement <4 x float> [[B]], i32 1413; CHECK-NEXT: [[TMP1:%.*]] = insertelement <2 x i32> zeroinitializer, i32 [[C0]], i32 0414; CHECK-NEXT: [[TMP2:%.*]] = insertelement <2 x i32> [[TMP1]], i32 [[C1]], i32 1415; CHECK-NEXT: [[TMP3:%.*]] = icmp ne <2 x i32> [[TMP2]], zeroinitializer416; CHECK-NEXT: [[TMP4:%.*]] = insertelement <2 x float> zeroinitializer, float [[A0]], i32 0417; CHECK-NEXT: [[TMP5:%.*]] = insertelement <2 x float> [[TMP4]], float [[A1]], i32 1418; CHECK-NEXT: [[TMP6:%.*]] = insertelement <2 x float> zeroinitializer, float [[B0]], i32 0419; CHECK-NEXT: [[TMP7:%.*]] = insertelement <2 x float> [[TMP6]], float [[B1]], i32 1420; CHECK-NEXT: [[TMP8:%.*]] = select <2 x i1> [[TMP3]], <2 x float> [[TMP5]], <2 x float> [[TMP7]]421; CHECK-NEXT: [[TMP9:%.*]] = extractelement <2 x float> [[TMP8]], i32 0422; CHECK-NEXT: [[RA:%.*]] = insertelement <4 x float> zeroinitializer, float [[TMP9]], i32 0423; CHECK-NEXT: [[TMP10:%.*]] = extractelement <2 x float> [[TMP8]], i32 1424; CHECK-NEXT: [[RB:%.*]] = insertelement <4 x float> [[RA]], float [[TMP10]], i32 1425; CHECK-NEXT: ret <4 x float> [[RB]]426;427 %c0 = extractelement <4 x i32> %c, i32 0428 %c1 = extractelement <4 x i32> %c, i32 1429 %a0 = extractelement <4 x float> %a, i32 0430 %a1 = extractelement <4 x float> %a, i32 1431 %b0 = extractelement <4 x float> %b, i32 0432 %b1 = extractelement <4 x float> %b, i32 1433 %1 = insertelement <2 x i32> zeroinitializer, i32 %c0, i32 0434 %2 = insertelement <2 x i32> %1, i32 %c1, i32 1435 %3 = icmp ne <2 x i32> %2, zeroinitializer436 %4 = insertelement <2 x float> zeroinitializer, float %a0, i32 0437 %5 = insertelement <2 x float> %4, float %a1, i32 1438 %6 = insertelement <2 x float> zeroinitializer, float %b0, i32 0439 %7 = insertelement <2 x float> %6, float %b1, i32 1440 %8 = select <2 x i1> %3, <2 x float> %5, <2 x float> %7441 %9 = extractelement <2 x float> %8, i32 0442 %ra = insertelement <4 x float> zeroinitializer, float %9, i32 0443 %10 = extractelement <2 x float> %8, i32 1444 %rb = insertelement <4 x float> %ra, float %10, i32 1445 ret <4 x float> %rb446}447 448; Make sure that vectorization happens even if insertelements operations449; must be rescheduled. The case here is from compiling Julia.450define <4 x float> @reschedule_extract(<4 x float> %a, <4 x float> %b) {451; CHECK-LABEL: @reschedule_extract(452; CHECK-NEXT: [[TMP1:%.*]] = fadd <4 x float> [[A:%.*]], [[B:%.*]]453; CHECK-NEXT: ret <4 x float> [[TMP1]]454;455 %a0 = extractelement <4 x float> %a, i32 0456 %b0 = extractelement <4 x float> %b, i32 0457 %c0 = fadd float %a0, %b0458 %v0 = insertelement <4 x float> zeroinitializer, float %c0, i32 0459 %a1 = extractelement <4 x float> %a, i32 1460 %b1 = extractelement <4 x float> %b, i32 1461 %c1 = fadd float %a1, %b1462 %v1 = insertelement <4 x float> %v0, float %c1, i32 1463 %a2 = extractelement <4 x float> %a, i32 2464 %b2 = extractelement <4 x float> %b, i32 2465 %c2 = fadd float %a2, %b2466 %v2 = insertelement <4 x float> %v1, float %c2, i32 2467 %a3 = extractelement <4 x float> %a, i32 3468 %b3 = extractelement <4 x float> %b, i32 3469 %c3 = fadd float %a3, %b3470 %v3 = insertelement <4 x float> %v2, float %c3, i32 3471 ret <4 x float> %v3472}473 474; Check that cost model for vectorization takes credit for475; instructions that are erased.476define <4 x float> @take_credit(<4 x float> %a, <4 x float> %b) {477; CHECK-LABEL: @take_credit(478; CHECK-NEXT: [[TMP1:%.*]] = fadd <4 x float> [[A:%.*]], [[B:%.*]]479; CHECK-NEXT: ret <4 x float> [[TMP1]]480;481 %a0 = extractelement <4 x float> %a, i32 0482 %b0 = extractelement <4 x float> %b, i32 0483 %c0 = fadd float %a0, %b0484 %a1 = extractelement <4 x float> %a, i32 1485 %b1 = extractelement <4 x float> %b, i32 1486 %c1 = fadd float %a1, %b1487 %a2 = extractelement <4 x float> %a, i32 2488 %b2 = extractelement <4 x float> %b, i32 2489 %c2 = fadd float %a2, %b2490 %a3 = extractelement <4 x float> %a, i32 3491 %b3 = extractelement <4 x float> %b, i32 3492 %c3 = fadd float %a3, %b3493 %v0 = insertelement <4 x float> zeroinitializer, float %c0, i32 0494 %v1 = insertelement <4 x float> %v0, float %c1, i32 1495 %v2 = insertelement <4 x float> %v1, float %c2, i32 2496 %v3 = insertelement <4 x float> %v2, float %c3, i32 3497 ret <4 x float> %v3498}499 500; Make sure we handle multiple trees that feed one build vector correctly.501define <4 x double> @multi_tree(double %w, double %x, double %y, double %z) {502; CHECK-LABEL: @multi_tree(503; CHECK-NEXT: [[TMP1:%.*]] = insertelement <4 x double> poison, double [[Z:%.*]], i32 0504; CHECK-NEXT: [[TMP2:%.*]] = insertelement <4 x double> [[TMP1]], double [[Y:%.*]], i32 1505; CHECK-NEXT: [[TMP3:%.*]] = insertelement <4 x double> [[TMP2]], double [[X:%.*]], i32 2506; CHECK-NEXT: [[TMP4:%.*]] = insertelement <4 x double> [[TMP3]], double [[W:%.*]], i32 3507; CHECK-NEXT: [[TMP5:%.*]] = fadd <4 x double> [[TMP4]], <double 3.000000e+00, double 2.000000e+00, double 1.000000e+00, double 0.000000e+00>508; CHECK-NEXT: [[TMP6:%.*]] = fmul <4 x double> [[TMP5]], splat (double 1.000000e+00)509; CHECK-NEXT: ret <4 x double> [[TMP6]]510;511 %t0 = fadd double %w , 0.000000e+00512 %t1 = fadd double %x , 1.000000e+00513 %t2 = fadd double %y , 2.000000e+00514 %t3 = fadd double %z , 3.000000e+00515 %t4 = fmul double %t0, 1.000000e+00516 %i1 = insertelement <4 x double> zeroinitializer, double %t4, i32 3517 %t5 = fmul double %t1, 1.000000e+00518 %i2 = insertelement <4 x double> %i1, double %t5, i32 2519 %t6 = fmul double %t2, 1.000000e+00520 %i3 = insertelement <4 x double> %i2, double %t6, i32 1521 %t7 = fmul double %t3, 1.000000e+00522 %i4 = insertelement <4 x double> %i3, double %t7, i32 0523 ret <4 x double> %i4524}525 526define <8 x float> @_vadd256(<8 x float> %a, <8 x float> %b) local_unnamed_addr {527; CHECK-LABEL: @_vadd256(528; CHECK-NEXT: [[TMP1:%.*]] = fadd <8 x float> [[A:%.*]], [[B:%.*]]529; CHECK-NEXT: ret <8 x float> [[TMP1]]530;531 %vecext = extractelement <8 x float> %a, i32 0532 %vecext1 = extractelement <8 x float> %b, i32 0533 %add = fadd float %vecext, %vecext1534 %vecext2 = extractelement <8 x float> %a, i32 1535 %vecext3 = extractelement <8 x float> %b, i32 1536 %add4 = fadd float %vecext2, %vecext3537 %vecext5 = extractelement <8 x float> %a, i32 2538 %vecext6 = extractelement <8 x float> %b, i32 2539 %add7 = fadd float %vecext5, %vecext6540 %vecext8 = extractelement <8 x float> %a, i32 3541 %vecext9 = extractelement <8 x float> %b, i32 3542 %add10 = fadd float %vecext8, %vecext9543 %vecext11 = extractelement <8 x float> %a, i32 4544 %vecext12 = extractelement <8 x float> %b, i32 4545 %add13 = fadd float %vecext11, %vecext12546 %vecext14 = extractelement <8 x float> %a, i32 5547 %vecext15 = extractelement <8 x float> %b, i32 5548 %add16 = fadd float %vecext14, %vecext15549 %vecext17 = extractelement <8 x float> %a, i32 6550 %vecext18 = extractelement <8 x float> %b, i32 6551 %add19 = fadd float %vecext17, %vecext18552 %vecext20 = extractelement <8 x float> %a, i32 7553 %vecext21 = extractelement <8 x float> %b, i32 7554 %add22 = fadd float %vecext20, %vecext21555 %vecinit.i = insertelement <8 x float> zeroinitializer, float %add, i32 0556 %vecinit1.i = insertelement <8 x float> %vecinit.i, float %add4, i32 1557 %vecinit2.i = insertelement <8 x float> %vecinit1.i, float %add7, i32 2558 %vecinit3.i = insertelement <8 x float> %vecinit2.i, float %add10, i32 3559 %vecinit4.i = insertelement <8 x float> %vecinit3.i, float %add13, i32 4560 %vecinit5.i = insertelement <8 x float> %vecinit4.i, float %add16, i32 5561 %vecinit6.i = insertelement <8 x float> %vecinit5.i, float %add19, i32 6562 %vecinit7.i = insertelement <8 x float> %vecinit6.i, float %add22, i32 7563 ret <8 x float> %vecinit7.i564}565