266 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_test_checks.py2; RUN: opt -passes=slp-vectorizer -mtriple=arm64-apple-ios -S %s | FileCheck %s3; RUN: opt -aa-pipeline='basic-aa,scoped-noalias-aa' -passes=slp-vectorizer -mtriple=arm64-apple-darwin -S %s | FileCheck %s4 5define void @loop1(ptr %A, ptr %B, i64 %N) {6; CHECK-LABEL: @loop1(7; CHECK-NEXT: entry:8; CHECK-NEXT: br label [[LOOP:%.*]]9; CHECK: loop:10; CHECK-NEXT: [[IV:%.*]] = phi i64 [ 0, [[ENTRY:%.*]] ], [ [[IV_NEXT:%.*]], [[LOOP]] ]11; CHECK-NEXT: [[B_GEP_0:%.*]] = getelementptr inbounds i32, ptr [[B:%.*]], i64 [[IV]]12; CHECK-NEXT: [[B_0:%.*]] = load i32, ptr [[B_GEP_0]], align 413; CHECK-NEXT: [[A_GEP_0:%.*]] = getelementptr inbounds i32, ptr [[A:%.*]], i64 [[IV]]14; CHECK-NEXT: [[A_0:%.*]] = load i32, ptr [[A_GEP_0]], align 415; CHECK-NEXT: [[ADD_0:%.*]] = add i32 [[A_0]], 2016; CHECK-NEXT: [[XOR_0:%.*]] = xor i32 [[ADD_0]], [[B_0]]17; CHECK-NEXT: store i32 [[XOR_0]], ptr [[A_GEP_0]], align 418; CHECK-NEXT: [[IV_1:%.*]] = or i64 [[IV]], 119; CHECK-NEXT: [[B_GEP_1:%.*]] = getelementptr inbounds i32, ptr [[B]], i64 [[IV_1]]20; CHECK-NEXT: [[B_1:%.*]] = load i32, ptr [[B_GEP_1]], align 421; CHECK-NEXT: [[A_GEP_1:%.*]] = getelementptr inbounds i32, ptr [[A]], i64 [[IV_1]]22; CHECK-NEXT: [[A_1:%.*]] = load i32, ptr [[A_GEP_1]], align 423; CHECK-NEXT: [[ADD_1:%.*]] = add i32 [[A_1]], 2024; CHECK-NEXT: [[XOR_1:%.*]] = xor i32 [[ADD_1]], [[B_1]]25; CHECK-NEXT: store i32 [[XOR_1]], ptr [[A_GEP_1]], align 426; CHECK-NEXT: [[IV_2:%.*]] = or i64 [[IV]], 227; CHECK-NEXT: [[B_GEP_2:%.*]] = getelementptr inbounds i32, ptr [[B]], i64 [[IV_2]]28; CHECK-NEXT: [[B_2:%.*]] = load i32, ptr [[B_GEP_2]], align 429; CHECK-NEXT: [[A_GEP_2:%.*]] = getelementptr inbounds i32, ptr [[A]], i64 [[IV_2]]30; CHECK-NEXT: [[A_2:%.*]] = load i32, ptr [[A_GEP_2]], align 431; CHECK-NEXT: [[ADD_2:%.*]] = add i32 [[A_2]], 2032; CHECK-NEXT: [[XOR_2:%.*]] = xor i32 [[ADD_2]], [[B_2]]33; CHECK-NEXT: store i32 [[XOR_2]], ptr [[A_GEP_2]], align 434; CHECK-NEXT: [[IV_3:%.*]] = or i64 [[IV]], 335; CHECK-NEXT: [[B_GEP_3:%.*]] = getelementptr inbounds i32, ptr [[B]], i64 [[IV_3]]36; CHECK-NEXT: [[B_3:%.*]] = load i32, ptr [[B_GEP_3]], align 437; CHECK-NEXT: [[A_GEP_3:%.*]] = getelementptr inbounds i32, ptr [[A]], i64 [[IV_3]]38; CHECK-NEXT: [[A_3:%.*]] = load i32, ptr [[A_GEP_3]], align 439; CHECK-NEXT: [[ADD_3:%.*]] = add i32 [[A_3]], 2040; CHECK-NEXT: [[XOR_3:%.*]] = xor i32 [[ADD_3]], [[B_3]]41; CHECK-NEXT: store i32 [[XOR_3]], ptr [[A_GEP_3]], align 442; CHECK-NEXT: [[IV_NEXT]] = add nuw nsw i64 [[IV]], 1643; CHECK-NEXT: [[COND:%.*]] = icmp ult i64 [[IV_NEXT]], [[N:%.*]]44; CHECK-NEXT: br i1 [[COND]], label [[LOOP]], label [[EXIT:%.*]]45; CHECK: exit:46; CHECK-NEXT: ret void47;48entry:49 br label %loop50 51loop:52 %iv = phi i64 [ 0, %entry ], [ %iv.next, %loop ]53 %B.gep.0 = getelementptr inbounds i32, ptr %B, i64 %iv54 %B.0 = load i32, ptr %B.gep.0, align 455 %A.gep.0 = getelementptr inbounds i32, ptr %A, i64 %iv56 %A.0 = load i32, ptr %A.gep.0, align 457 %add.0 = add i32 %A.0, 2058 %xor.0 = xor i32 %add.0, %B.059 store i32 %xor.0, ptr %A.gep.0, align 460 %iv.1 = or i64 %iv, 161 %B.gep.1 = getelementptr inbounds i32, ptr %B, i64 %iv.162 %B.1 = load i32, ptr %B.gep.1, align 463 %A.gep.1 = getelementptr inbounds i32, ptr %A, i64 %iv.164 %A.1 = load i32, ptr %A.gep.1, align 465 %add.1 = add i32 %A.1, 2066 %xor.1 = xor i32 %add.1, %B.167 store i32 %xor.1, ptr %A.gep.1, align 468 %iv.2 = or i64 %iv, 269 %B.gep.2 = getelementptr inbounds i32, ptr %B, i64 %iv.270 %B.2 = load i32, ptr %B.gep.2, align 471 %A.gep.2 = getelementptr inbounds i32, ptr %A, i64 %iv.272 %A.2 = load i32, ptr %A.gep.2, align 473 %add.2 = add i32 %A.2, 2074 %xor.2 = xor i32 %add.2, %B.275 store i32 %xor.2, ptr %A.gep.2, align 476 %iv.3 = or i64 %iv, 377 %B.gep.3 = getelementptr inbounds i32, ptr %B, i64 %iv.378 %B.3 = load i32, ptr %B.gep.3, align 479 %A.gep.3 = getelementptr inbounds i32, ptr %A, i64 %iv.380 %A.3 = load i32, ptr %A.gep.3, align 481 %add.3 = add i32 %A.3, 2082 %xor.3 = xor i32 %add.3, %B.383 store i32 %xor.3, ptr %A.gep.3, align 484 %iv.next = add nuw nsw i64 %iv, 1685 %cond = icmp ult i64 %iv.next, %N86 br i1 %cond, label %loop, label %exit87 88exit:89 ret void90}91 92define void @loop_iv_update_at_start(ptr %src, ptr %dst) #0 {93; CHECK-LABEL: @loop_iv_update_at_start(94; CHECK-NEXT: entry:95; CHECK-NEXT: br label [[LOOP:%.*]]96; CHECK: loop:97; CHECK-NEXT: [[IV:%.*]] = phi i32 [ 0, [[ENTRY:%.*]] ], [ [[IV_NEXT:%.*]], [[LOOP]] ]98; CHECK-NEXT: [[IV_NEXT]] = add i32 [[IV]], 199; CHECK-NEXT: [[COND:%.*]] = icmp ult i32 [[IV]], 2000100; CHECK-NEXT: [[SRC_0:%.*]] = load float, ptr [[SRC:%.*]], align 8101; CHECK-NEXT: [[ADD_0:%.*]] = fadd float [[SRC_0]], 1.000000e+00102; CHECK-NEXT: [[MUL_0:%.*]] = fmul float [[ADD_0]], [[SRC_0]]103; CHECK-NEXT: store float [[MUL_0]], ptr [[DST:%.*]], align 8104; CHECK-NEXT: [[SRC_GEP_1:%.*]] = getelementptr inbounds float, ptr [[SRC]], i64 1105; CHECK-NEXT: [[SRC_1:%.*]] = load float, ptr [[SRC_GEP_1]], align 8106; CHECK-NEXT: [[ADD_1:%.*]] = fadd float [[SRC_1]], 1.000000e+00107; CHECK-NEXT: [[MUL_1:%.*]] = fmul float [[ADD_1]], [[SRC_1]]108; CHECK-NEXT: [[DST_GEP_1:%.*]] = getelementptr inbounds float, ptr [[DST]], i64 1109; CHECK-NEXT: store float [[MUL_1]], ptr [[DST_GEP_1]], align 8110; CHECK-NEXT: [[SRC_GEP_2:%.*]] = getelementptr inbounds float, ptr [[SRC]], i64 2111; CHECK-NEXT: [[SRC_2:%.*]] = load float, ptr [[SRC_GEP_2]], align 8112; CHECK-NEXT: [[ADD_2:%.*]] = fadd float [[SRC_2]], 1.000000e+00113; CHECK-NEXT: [[MUL_2:%.*]] = fmul float [[ADD_2]], [[SRC_2]]114; CHECK-NEXT: [[DST_GEP_2:%.*]] = getelementptr inbounds float, ptr [[DST]], i64 2115; CHECK-NEXT: store float [[MUL_2]], ptr [[DST_GEP_2]], align 8116; CHECK-NEXT: [[SRC_GEP_3:%.*]] = getelementptr inbounds float, ptr [[SRC]], i64 3117; CHECK-NEXT: [[SRC_3:%.*]] = load float, ptr [[SRC_GEP_3]], align 8118; CHECK-NEXT: [[ADD_3:%.*]] = fadd float [[SRC_3]], 1.000000e+00119; CHECK-NEXT: [[MUL_3:%.*]] = fmul float [[ADD_3]], [[SRC_3]]120; CHECK-NEXT: [[DST_GEP_3:%.*]] = getelementptr inbounds float, ptr [[DST]], i64 3121; CHECK-NEXT: store float [[MUL_3]], ptr [[DST_GEP_3]], align 8122; CHECK-NEXT: [[SRC_GEP_4:%.*]] = getelementptr inbounds float, ptr [[SRC]], i64 4123; CHECK-NEXT: [[SRC_4:%.*]] = load float, ptr [[SRC_GEP_4]], align 8124; CHECK-NEXT: [[ADD_4:%.*]] = fadd float [[SRC_4]], 1.000000e+00125; CHECK-NEXT: [[MUL_4:%.*]] = fmul float [[ADD_4]], [[SRC_4]]126; CHECK-NEXT: [[DST_GEP_4:%.*]] = getelementptr inbounds float, ptr [[DST]], i64 4127; CHECK-NEXT: store float [[MUL_4]], ptr [[DST_GEP_4]], align 8128; CHECK-NEXT: br i1 [[COND]], label [[LOOP]], label [[EXIT:%.*]]129; CHECK: exit:130; CHECK-NEXT: ret void131;132entry:133 br label %loop134 135loop:136 %iv = phi i32 [ 0, %entry ], [ %iv.next, %loop ]137 %iv.next = add i32 %iv, 1138 %cond = icmp ult i32 %iv, 2000139 140 %src.0 = load float, ptr %src, align 8141 %add.0 = fadd float %src.0, 1.0142 %mul.0 = fmul float %add.0, %src.0143 store float %mul.0, ptr %dst, align 8144 145 %src.gep.1 = getelementptr inbounds float, ptr %src, i64 1146 %src.1 = load float, ptr %src.gep.1, align 8147 %add.1 = fadd float %src.1, 1.0148 %mul.1 = fmul float %add.1, %src.1149 %dst.gep.1 = getelementptr inbounds float, ptr %dst, i64 1150 store float %mul.1, ptr %dst.gep.1, align 8151 %src.gep.2 = getelementptr inbounds float, ptr %src, i64 2152 %src.2 = load float, ptr %src.gep.2, align 8153 %add.2 = fadd float %src.2, 1.0154 %mul.2 = fmul float %add.2, %src.2155 %dst.gep.2 = getelementptr inbounds float, ptr %dst, i64 2156 store float %mul.2, ptr %dst.gep.2, align 8157 %src.gep.3 = getelementptr inbounds float, ptr %src, i64 3158 %src.3 = load float, ptr %src.gep.3, align 8159 %add.3 = fadd float %src.3, 1.0160 %mul.3 = fmul float %add.3, %src.3161 %dst.gep.3 = getelementptr inbounds float, ptr %dst, i64 3162 store float %mul.3, ptr %dst.gep.3, align 8163 %src.gep.4 = getelementptr inbounds float, ptr %src, i64 4164 %src.4 = load float, ptr %src.gep.4, align 8165 %add.4 = fadd float %src.4, 1.0166 %mul.4 = fmul float %add.4, %src.4167 %dst.gep.4 = getelementptr inbounds float, ptr %dst, i64 4168 store float %mul.4, ptr %dst.gep.4, align 8169 br i1 %cond, label %loop, label %exit170 171exit:172 ret void173}174 175; Similar to @loop1, but a load is used in a phi in the same basic block.176define i32 @value_used_in_phi(ptr %A, ptr %B, i64 %N) {177; CHECK-LABEL: @value_used_in_phi(178; CHECK-NEXT: entry:179; CHECK-NEXT: br label [[LOOP:%.*]]180; CHECK: loop:181; CHECK-NEXT: [[IV:%.*]] = phi i64 [ 0, [[ENTRY:%.*]] ], [ [[IV_NEXT:%.*]], [[LOOP]] ]182; CHECK-NEXT: [[P:%.*]] = phi i32 [ 0, [[ENTRY]] ], [ [[A_3:%.*]], [[LOOP]] ]183; CHECK-NEXT: [[B_GEP_0:%.*]] = getelementptr inbounds i32, ptr [[B:%.*]], i64 [[IV]]184; CHECK-NEXT: [[B_0:%.*]] = load i32, ptr [[B_GEP_0]], align 4185; CHECK-NEXT: [[A_GEP_0:%.*]] = getelementptr inbounds i32, ptr [[A:%.*]], i64 [[IV]]186; CHECK-NEXT: [[A_0:%.*]] = load i32, ptr [[A_GEP_0]], align 4187; CHECK-NEXT: [[ADD_0:%.*]] = add i32 [[A_0]], 20188; CHECK-NEXT: [[XOR_0:%.*]] = xor i32 [[ADD_0]], [[B_0]]189; CHECK-NEXT: store i32 [[XOR_0]], ptr [[A_GEP_0]], align 4190; CHECK-NEXT: [[IV_1:%.*]] = or i64 [[IV]], 1191; CHECK-NEXT: [[B_GEP_1:%.*]] = getelementptr inbounds i32, ptr [[B]], i64 [[IV_1]]192; CHECK-NEXT: [[B_1:%.*]] = load i32, ptr [[B_GEP_1]], align 4193; CHECK-NEXT: [[A_GEP_1:%.*]] = getelementptr inbounds i32, ptr [[A]], i64 [[IV_1]]194; CHECK-NEXT: [[A_1:%.*]] = load i32, ptr [[A_GEP_1]], align 4195; CHECK-NEXT: [[ADD_1:%.*]] = add i32 [[A_1]], 20196; CHECK-NEXT: [[XOR_1:%.*]] = xor i32 [[ADD_1]], [[B_1]]197; CHECK-NEXT: store i32 [[XOR_1]], ptr [[A_GEP_1]], align 4198; CHECK-NEXT: [[IV_2:%.*]] = or i64 [[IV]], 2199; CHECK-NEXT: [[B_GEP_2:%.*]] = getelementptr inbounds i32, ptr [[B]], i64 [[IV_2]]200; CHECK-NEXT: [[B_2:%.*]] = load i32, ptr [[B_GEP_2]], align 4201; CHECK-NEXT: [[A_GEP_2:%.*]] = getelementptr inbounds i32, ptr [[A]], i64 [[IV_2]]202; CHECK-NEXT: [[A_2:%.*]] = load i32, ptr [[A_GEP_2]], align 4203; CHECK-NEXT: [[ADD_2:%.*]] = add i32 [[A_2]], 20204; CHECK-NEXT: [[XOR_2:%.*]] = xor i32 [[ADD_2]], [[B_2]]205; CHECK-NEXT: store i32 [[XOR_2]], ptr [[A_GEP_2]], align 4206; CHECK-NEXT: [[IV_3:%.*]] = or i64 [[IV]], 3207; CHECK-NEXT: [[B_GEP_3:%.*]] = getelementptr inbounds i32, ptr [[B]], i64 [[IV_3]]208; CHECK-NEXT: [[B_3:%.*]] = load i32, ptr [[B_GEP_3]], align 4209; CHECK-NEXT: [[A_GEP_3:%.*]] = getelementptr inbounds i32, ptr [[A]], i64 [[IV_3]]210; CHECK-NEXT: [[A_3]] = load i32, ptr [[A_GEP_3]], align 4211; CHECK-NEXT: [[ADD_3:%.*]] = add i32 [[A_3]], 20212; CHECK-NEXT: [[XOR_3:%.*]] = xor i32 [[ADD_3]], [[B_3]]213; CHECK-NEXT: store i32 [[XOR_3]], ptr [[A_GEP_3]], align 4214; CHECK-NEXT: [[IV_NEXT]] = add nuw nsw i64 [[IV]], 16215; CHECK-NEXT: [[COND:%.*]] = icmp ult i64 [[IV_NEXT]], [[N:%.*]]216; CHECK-NEXT: br i1 [[COND]], label [[LOOP]], label [[EXIT:%.*]]217; CHECK: exit:218; CHECK-NEXT: ret i32 [[P]]219;220entry:221 br label %loop222 223loop:224 %iv = phi i64 [ 0, %entry ], [ %iv.next, %loop ]225 %p = phi i32 [ 0, %entry ], [ %A.3, %loop ]226 %B.gep.0 = getelementptr inbounds i32, ptr %B, i64 %iv227 %B.0 = load i32, ptr %B.gep.0, align 4228 %A.gep.0 = getelementptr inbounds i32, ptr %A, i64 %iv229 %A.0 = load i32, ptr %A.gep.0, align 4230 %add.0 = add i32 %A.0, 20231 %xor.0 = xor i32 %add.0, %B.0232 store i32 %xor.0, ptr %A.gep.0, align 4233 %iv.1 = or i64 %iv, 1234 %B.gep.1 = getelementptr inbounds i32, ptr %B, i64 %iv.1235 %B.1 = load i32, ptr %B.gep.1, align 4236 %A.gep.1 = getelementptr inbounds i32, ptr %A, i64 %iv.1237 %A.1 = load i32, ptr %A.gep.1, align 4238 %add.1 = add i32 %A.1, 20239 %xor.1 = xor i32 %add.1, %B.1240 store i32 %xor.1, ptr %A.gep.1, align 4241 %iv.2 = or i64 %iv, 2242 %B.gep.2 = getelementptr inbounds i32, ptr %B, i64 %iv.2243 %B.2 = load i32, ptr %B.gep.2, align 4244 %A.gep.2 = getelementptr inbounds i32, ptr %A, i64 %iv.2245 %A.2 = load i32, ptr %A.gep.2, align 4246 %add.2 = add i32 %A.2, 20247 %xor.2 = xor i32 %add.2, %B.2248 store i32 %xor.2, ptr %A.gep.2, align 4249 %iv.3 = or i64 %iv, 3250 %B.gep.3 = getelementptr inbounds i32, ptr %B, i64 %iv.3251 %B.3 = load i32, ptr %B.gep.3, align 4252 %A.gep.3 = getelementptr inbounds i32, ptr %A, i64 %iv.3253 %A.3 = load i32, ptr %A.gep.3, align 4254 %add.3 = add i32 %A.3, 20255 %xor.3 = xor i32 %add.3, %B.3256 store i32 %xor.3, ptr %A.gep.3, align 4257 %iv.next = add nuw nsw i64 %iv, 16258 %cond = icmp ult i64 %iv.next, %N259 br i1 %cond, label %loop, label %exit260 261exit:262 ret i32 %p263}264 265 266