146 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py2; RUN: llc -verify-machineinstrs -mtriple=powerpc64le-unknown-linux-gnu -mcpu=pwr9 < %s | FileCheck -check-prefix=CHECK-P9 %s3; RUN: llc -verify-machineinstrs -mtriple=powerpc64le-unknown-linux-gnu -mcpu=pwr9 -ppc-postra-bias-addi=false < %s |\4; RUN: FileCheck -check-prefix=CHECK-P9-NO-HEURISTIC %s5 6%_type_of_scalars = type <{ [16 x i8], double, [152 x i8] }>7%_elem_type_of_x = type <{ double }>8%_elem_type_of_a = type <{ double }>9 10@scalars = common dso_local local_unnamed_addr global %_type_of_scalars zeroinitializer, align 1611 12define dso_local void @test(ptr noalias %.x, ptr %.a, ptr noalias %.n) {13; CHECK-P9-LABEL: test:14; CHECK-P9: # %bb.0: # %entry15; CHECK-P9-NEXT: ld 5, 0(5)16; CHECK-P9-NEXT: addis 6, 2, scalars@toc@ha17; CHECK-P9-NEXT: addi 6, 6, scalars@toc@l18; CHECK-P9-NEXT: rldicr 5, 5, 0, 5819; CHECK-P9-NEXT: addi 6, 6, 1620; CHECK-P9-NEXT: addi 5, 5, -3221; CHECK-P9-NEXT: lxvdsx 0, 0, 622; CHECK-P9-NEXT: rldicl 5, 5, 59, 523; CHECK-P9-NEXT: addi 5, 5, 124; CHECK-P9-NEXT: mtctr 525; CHECK-P9-NEXT: .p2align 426; CHECK-P9-NEXT: .LBB0_1: # %vector.body27; CHECK-P9-NEXT: #28; CHECK-P9-NEXT: lxv 1, 16(4)29; CHECK-P9-NEXT: lxv 2, 0(4)30; CHECK-P9-NEXT: lxv 3, 48(4)31; CHECK-P9-NEXT: lxv 4, 32(4)32; CHECK-P9-NEXT: xvmuldp 2, 2, 033; CHECK-P9-NEXT: lxv 5, 240(4)34; CHECK-P9-NEXT: lxv 6, 224(4)35; CHECK-P9-NEXT: xvmuldp 1, 1, 036; CHECK-P9-NEXT: xvmuldp 4, 4, 037; CHECK-P9-NEXT: xvmuldp 3, 3, 038; CHECK-P9-NEXT: xvmuldp 6, 6, 039; CHECK-P9-NEXT: xvmuldp 5, 5, 040; CHECK-P9-NEXT: addi 4, 4, 25641; CHECK-P9-NEXT: stxv 1, 16(3)42; CHECK-P9-NEXT: stxv 2, 0(3)43; CHECK-P9-NEXT: stxv 3, 48(3)44; CHECK-P9-NEXT: stxv 4, 32(3)45; CHECK-P9-NEXT: stxv 5, 240(3)46; CHECK-P9-NEXT: stxv 6, 224(3)47; CHECK-P9-NEXT: addi 3, 3, 25648; CHECK-P9-NEXT: bdnz .LBB0_149; CHECK-P9-NEXT: # %bb.2: # %return.block50; CHECK-P9-NEXT: blr51;52; CHECK-P9-NO-HEURISTIC-LABEL: test:53; CHECK-P9-NO-HEURISTIC: # %bb.0: # %entry54; CHECK-P9-NO-HEURISTIC-NEXT: ld 5, 0(5)55; CHECK-P9-NO-HEURISTIC-NEXT: addis 6, 2, scalars@toc@ha56; CHECK-P9-NO-HEURISTIC-NEXT: addi 6, 6, scalars@toc@l57; CHECK-P9-NO-HEURISTIC-NEXT: rldicr 5, 5, 0, 5858; CHECK-P9-NO-HEURISTIC-NEXT: addi 6, 6, 1659; CHECK-P9-NO-HEURISTIC-NEXT: addi 5, 5, -3260; CHECK-P9-NO-HEURISTIC-NEXT: lxvdsx 0, 0, 661; CHECK-P9-NO-HEURISTIC-NEXT: rldicl 5, 5, 59, 562; CHECK-P9-NO-HEURISTIC-NEXT: addi 5, 5, 163; CHECK-P9-NO-HEURISTIC-NEXT: mtctr 564; CHECK-P9-NO-HEURISTIC-NEXT: .p2align 465; CHECK-P9-NO-HEURISTIC-NEXT: .LBB0_1: # %vector.body66; CHECK-P9-NO-HEURISTIC-NEXT: #67; CHECK-P9-NO-HEURISTIC-NEXT: lxv 1, 16(4)68; CHECK-P9-NO-HEURISTIC-NEXT: lxv 2, 0(4)69; CHECK-P9-NO-HEURISTIC-NEXT: lxv 3, 48(4)70; CHECK-P9-NO-HEURISTIC-NEXT: lxv 4, 32(4)71; CHECK-P9-NO-HEURISTIC-NEXT: xvmuldp 2, 2, 072; CHECK-P9-NO-HEURISTIC-NEXT: lxv 5, 240(4)73; CHECK-P9-NO-HEURISTIC-NEXT: lxv 6, 224(4)74; CHECK-P9-NO-HEURISTIC-NEXT: xvmuldp 1, 1, 075; CHECK-P9-NO-HEURISTIC-NEXT: xvmuldp 4, 4, 076; CHECK-P9-NO-HEURISTIC-NEXT: xvmuldp 3, 3, 077; CHECK-P9-NO-HEURISTIC-NEXT: xvmuldp 6, 6, 078; CHECK-P9-NO-HEURISTIC-NEXT: xvmuldp 5, 5, 079; CHECK-P9-NO-HEURISTIC-NEXT: addi 4, 4, 25680; CHECK-P9-NO-HEURISTIC-NEXT: stxv 1, 16(3)81; CHECK-P9-NO-HEURISTIC-NEXT: stxv 2, 0(3)82; CHECK-P9-NO-HEURISTIC-NEXT: stxv 3, 48(3)83; CHECK-P9-NO-HEURISTIC-NEXT: stxv 4, 32(3)84; CHECK-P9-NO-HEURISTIC-NEXT: stxv 5, 240(3)85; CHECK-P9-NO-HEURISTIC-NEXT: stxv 6, 224(3)86; CHECK-P9-NO-HEURISTIC-NEXT: addi 3, 3, 25687; CHECK-P9-NO-HEURISTIC-NEXT: bdnz .LBB0_188; CHECK-P9-NO-HEURISTIC-NEXT: # %bb.2: # %return.block89; CHECK-P9-NO-HEURISTIC-NEXT: blr90entry:91 %x_rvo_based_addr_3 = getelementptr inbounds [0 x %_elem_type_of_x], ptr %.x, i64 0, i64 -192 %a_rvo_based_addr_5 = getelementptr inbounds [0 x %_elem_type_of_a], ptr %.a, i64 0, i64 -193 %_val_n_ = load i64, ptr %.n, align 894 %_val_c1_ = load double, ptr getelementptr inbounds (%_type_of_scalars, ptr @scalars, i64 0, i32 1), align 1695 %n.vec = and i64 %_val_n_, -3296 %broadcast.splatinsert26 = insertelement <4 x double> undef, double %_val_c1_, i32 097 %broadcast.splat27 = shufflevector <4 x double> %broadcast.splatinsert26, <4 x double> undef, <4 x i32> zeroinitializer98 br label %vector.body99 100vector.body:101 %index = phi i64 [ 0, %entry ], [ %index.next, %vector.body ]102 %offset.idx = or disjoint i64 %index, 1103 %0 = getelementptr %_elem_type_of_x, ptr %x_rvo_based_addr_3, i64 %offset.idx, i32 0104 %1 = getelementptr %_elem_type_of_a, ptr %a_rvo_based_addr_5, i64 %offset.idx, i32 0105 %wide.load = load <4 x double>, ptr %1, align 8106 %2 = getelementptr double, ptr %1, i64 4107 %wide.load19 = load <4 x double>, ptr %2, align 8108 %3 = getelementptr double, ptr %1, i64 8109 %wide.load20 = load <4 x double>, ptr %3, align 8110 %4 = getelementptr double, ptr %1, i64 12111 %wide.load21 = load <4 x double>, ptr %4, align 8112 %5 = getelementptr double, ptr %1, i64 16113 %wide.load22 = load <4 x double>, ptr %5, align 8114 %6 = getelementptr double, ptr %1, i64 20115 %wide.load23 = load <4 x double>, ptr %6, align 8116 %7 = getelementptr double, ptr %1, i64 24117 %wide.load24 = load <4 x double>, ptr %7, align 8118 %8 = getelementptr double, ptr %1, i64 28119 %wide.load25 = load <4 x double>, ptr %8, align 8120 %9 = fmul fast <4 x double> %wide.load, %broadcast.splat27121 %10 = fmul fast <4 x double> %wide.load19, %broadcast.splat27122 %11 = fmul fast <4 x double> %wide.load20, %broadcast.splat27123 %12 = fmul fast <4 x double> %wide.load21, %broadcast.splat27124 %13 = fmul fast <4 x double> %wide.load22, %broadcast.splat27125 %14 = fmul fast <4 x double> %wide.load23, %broadcast.splat27126 %15 = fmul fast <4 x double> %wide.load24, %broadcast.splat27127 %16 = fmul fast <4 x double> %wide.load25, %broadcast.splat27128 store <4 x double> %9, ptr %0, align 8129 %17 = getelementptr double, ptr %0, i64 4130 store <4 x double> %10, ptr %17, align 8131 %18 = getelementptr double, ptr %0, i64 8132 %19 = getelementptr double, ptr %0, i64 12133 %20 = getelementptr double, ptr %0, i64 16134 %21 = getelementptr double, ptr %0, i64 20135 %22 = getelementptr double, ptr %0, i64 24136 %23 = getelementptr double, ptr %0, i64 28137 store <4 x double> %16, ptr %23, align 8138 %index.next = add i64 %index, 32139 %cm = icmp eq i64 %index.next, %n.vec140 br i1 %cm, label %return.block, label %vector.body141 142return.block:143 ret void144}145 146