213 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 22; RUN: llc < %s -mtriple=arm64-eabi -aarch64-enable-stp-suppress=false -verify-machineinstrs -mcpu=cyclone -aarch64-enable-sink-fold=true | FileCheck %s3 4define void @stp_int(i32 %a, i32 %b, ptr nocapture %p) nounwind {5; CHECK-LABEL: stp_int:6; CHECK: // %bb.0:7; CHECK-NEXT: stp w0, w1, [x2]8; CHECK-NEXT: ret9 store i32 %a, ptr %p, align 410 %add.ptr = getelementptr inbounds i32, ptr %p, i64 111 store i32 %b, ptr %add.ptr, align 412 ret void13}14 15define void @stp_long(i64 %a, i64 %b, ptr nocapture %p) nounwind {16; CHECK-LABEL: stp_long:17; CHECK: // %bb.0:18; CHECK-NEXT: stp x0, x1, [x2]19; CHECK-NEXT: ret20 store i64 %a, ptr %p, align 821 %add.ptr = getelementptr inbounds i64, ptr %p, i64 122 store i64 %b, ptr %add.ptr, align 823 ret void24}25 26define void @stp_float(float %a, float %b, ptr nocapture %p) nounwind {27; CHECK-LABEL: stp_float:28; CHECK: // %bb.0:29; CHECK-NEXT: stp s0, s1, [x0]30; CHECK-NEXT: ret31 store float %a, ptr %p, align 432 %add.ptr = getelementptr inbounds float, ptr %p, i64 133 store float %b, ptr %add.ptr, align 434 ret void35}36 37define void @stp_double(double %a, double %b, ptr nocapture %p) nounwind {38; CHECK-LABEL: stp_double:39; CHECK: // %bb.0:40; CHECK-NEXT: stp d0, d1, [x0]41; CHECK-NEXT: ret42 store double %a, ptr %p, align 843 %add.ptr = getelementptr inbounds double, ptr %p, i64 144 store double %b, ptr %add.ptr, align 845 ret void46}47 48define void @stp_doublex2(<2 x double> %a, <2 x double> %b, ptr nocapture %p) nounwind {49; CHECK-LABEL: stp_doublex2:50; CHECK: // %bb.0:51; CHECK-NEXT: stp q0, q1, [x0]52; CHECK-NEXT: ret53 store <2 x double> %a, ptr %p, align 1654 %add.ptr = getelementptr inbounds <2 x double>, ptr %p, i64 155 store <2 x double> %b, ptr %add.ptr, align 1656 ret void57}58 59; Test the load/store optimizer---combine ldurs into a ldp, if appropriate60define void @stur_int(i32 %a, i32 %b, ptr nocapture %p) nounwind {61; CHECK-LABEL: stur_int:62; CHECK: // %bb.0:63; CHECK-NEXT: stp w1, w0, [x2, #-8]64; CHECK-NEXT: ret65 %p1 = getelementptr inbounds i32, ptr %p, i32 -166 store i32 %a, ptr %p1, align 267 %p2 = getelementptr inbounds i32, ptr %p, i32 -268 store i32 %b, ptr %p2, align 269 ret void70}71 72define void @stur_long(i64 %a, i64 %b, ptr nocapture %p) nounwind {73; CHECK-LABEL: stur_long:74; CHECK: // %bb.0:75; CHECK-NEXT: stp x1, x0, [x2, #-16]76; CHECK-NEXT: ret77 %p1 = getelementptr inbounds i64, ptr %p, i32 -178 store i64 %a, ptr %p1, align 279 %p2 = getelementptr inbounds i64, ptr %p, i32 -280 store i64 %b, ptr %p2, align 281 ret void82}83 84define void @stur_float(float %a, float %b, ptr nocapture %p) nounwind {85; CHECK-LABEL: stur_float:86; CHECK: // %bb.0:87; CHECK-NEXT: stp s1, s0, [x0, #-8]88; CHECK-NEXT: ret89 %p1 = getelementptr inbounds float, ptr %p, i32 -190 store float %a, ptr %p1, align 291 %p2 = getelementptr inbounds float, ptr %p, i32 -292 store float %b, ptr %p2, align 293 ret void94}95 96define void @stur_double(double %a, double %b, ptr nocapture %p) nounwind {97; CHECK-LABEL: stur_double:98; CHECK: // %bb.0:99; CHECK-NEXT: stp d1, d0, [x0, #-16]100; CHECK-NEXT: ret101 %p1 = getelementptr inbounds double, ptr %p, i32 -1102 store double %a, ptr %p1, align 2103 %p2 = getelementptr inbounds double, ptr %p, i32 -2104 store double %b, ptr %p2, align 2105 ret void106}107 108define void @stur_doublex2(<2 x double> %a, <2 x double> %b, ptr nocapture %p) nounwind {109; CHECK-LABEL: stur_doublex2:110; CHECK: // %bb.0:111; CHECK-NEXT: stp q1, q0, [x0, #-32]112; CHECK-NEXT: ret113 %p1 = getelementptr inbounds <2 x double>, ptr %p, i32 -1114 store <2 x double> %a, ptr %p1, align 2115 %p2 = getelementptr inbounds <2 x double>, ptr %p, i32 -2116 store <2 x double> %b, ptr %p2, align 2117 ret void118}119 120define void @splat_v4i32(i32 %v, ptr %p) {121; CHECK-LABEL: splat_v4i32:122; CHECK: // %bb.0: // %entry123; CHECK-NEXT: dup v0.4s, w0124; CHECK-NEXT: str q0, [x1]125; CHECK-NEXT: ret126entry:127 %p17 = insertelement <4 x i32> undef, i32 %v, i32 0128 %p18 = insertelement <4 x i32> %p17, i32 %v, i32 1129 %p19 = insertelement <4 x i32> %p18, i32 %v, i32 2130 %p20 = insertelement <4 x i32> %p19, i32 %v, i32 3131 store <4 x i32> %p20, ptr %p, align 4132 ret void133}134 135; Check that a non-splat store that is storing a vector created by 4136; insertelements that is not a splat vector does not get split.137define void @nosplat_v4i32(i32 %v, ptr %p) {138; CHECK-LABEL: nosplat_v4i32:139; CHECK: // %bb.0: // %entry140; CHECK-NEXT: sub sp, sp, #16141; CHECK-NEXT: .cfi_def_cfa_offset 16142; CHECK-NEXT: // kill: def $w0 killed $w0 def $x0143; CHECK-NEXT: mov x8, sp144; CHECK-NEXT: bfi x8, x0, #2, #2145; CHECK-NEXT: str w0, [x8]146; CHECK-NEXT: ldr q0, [sp]147; CHECK-NEXT: mov v0.s[1], w0148; CHECK-NEXT: mov v0.s[2], w0149; CHECK-NEXT: mov v0.s[3], w0150; CHECK-NEXT: str q0, [x1]151; CHECK-NEXT: add sp, sp, #16152; CHECK-NEXT: ret153entry:154 %p17 = insertelement <4 x i32> undef, i32 %v, i32 %v155 %p18 = insertelement <4 x i32> %p17, i32 %v, i32 1156 %p19 = insertelement <4 x i32> %p18, i32 %v, i32 2157 %p20 = insertelement <4 x i32> %p19, i32 %v, i32 3158 store <4 x i32> %p20, ptr %p, align 4159 ret void160}161 162; Check that a non-splat store that is storing a vector created by 4163; insertelements that is not a splat vector does not get split.164define void @nosplat2_v4i32(i32 %v, ptr %p, <4 x i32> %vin) {165; CHECK-LABEL: nosplat2_v4i32:166; CHECK: // %bb.0: // %entry167; CHECK-NEXT: mov v0.s[1], w0168; CHECK-NEXT: mov v0.s[2], w0169; CHECK-NEXT: mov v0.s[3], w0170; CHECK-NEXT: str q0, [x1]171; CHECK-NEXT: ret172entry:173 %p18 = insertelement <4 x i32> %vin, i32 %v, i32 1174 %p19 = insertelement <4 x i32> %p18, i32 %v, i32 2175 %p20 = insertelement <4 x i32> %p19, i32 %v, i32 3176 store <4 x i32> %p20, ptr %p, align 4177 ret void178}179 180; Read of %b to compute %tmp2 shouldn't prevent formation of stp181define i32 @stp_int_rar_hazard(i32 %a, i32 %b, ptr nocapture %p) nounwind {182; CHECK-LABEL: stp_int_rar_hazard:183; CHECK: // %bb.0:184; CHECK-NEXT: ldr w8, [x2, #8]185; CHECK-NEXT: stp w0, w1, [x2]186; CHECK-NEXT: add w0, w8, w1187; CHECK-NEXT: ret188 store i32 %a, ptr %p, align 4189 %ld.ptr = getelementptr inbounds i32, ptr %p, i64 2190 %tmp = load i32, ptr %ld.ptr, align 4191 %tmp2 = add i32 %tmp, %b192 %add.ptr = getelementptr inbounds i32, ptr %p, i64 1193 store i32 %b, ptr %add.ptr, align 4194 ret i32 %tmp2195}196 197; Read of %b to compute %tmp2 shouldn't prevent formation of stp198define i32 @stp_int_rar_hazard_after(i32 %w0, i32 %a, i32 %b, ptr nocapture %p) nounwind {199; CHECK-LABEL: stp_int_rar_hazard_after:200; CHECK: // %bb.0:201; CHECK-NEXT: ldr w8, [x3, #4]202; CHECK-NEXT: stp w1, w2, [x3]203; CHECK-NEXT: add w0, w8, w2204; CHECK-NEXT: ret205 store i32 %a, ptr %p, align 4206 %ld.ptr = getelementptr inbounds i32, ptr %p, i64 1207 %tmp = load i32, ptr %ld.ptr, align 4208 %tmp2 = add i32 %tmp, %b209 %add.ptr = getelementptr inbounds i32, ptr %p, i64 1210 store i32 %b, ptr %add.ptr, align 4211 ret i32 %tmp2212}213