235 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 52; RUN: llc --mtriple=loongarch32 -mattr=+32s,+lsx < %s | FileCheck %s --check-prefixes=CHECK,LA323; RUN: llc --mtriple=loongarch64 -mattr=+lsx < %s | FileCheck %s --check-prefixes=CHECK,LA644 5; A load has more than one user shouldn't be lowered to vldrepl6define <2 x i64> @should_not_be_optimized(ptr %ptr, ptr %dst){7; LA32-LABEL: should_not_be_optimized:8; LA32: # %bb.0:9; LA32-NEXT: ld.w $a2, $a0, 010; LA32-NEXT: ld.w $a0, $a0, 411; LA32-NEXT: st.w $a2, $a1, 012; LA32-NEXT: vinsgr2vr.w $vr0, $a2, 013; LA32-NEXT: vinsgr2vr.w $vr0, $a0, 114; LA32-NEXT: vreplvei.d $vr0, $vr0, 015; LA32-NEXT: st.w $a0, $a1, 416; LA32-NEXT: ret17;18; LA64-LABEL: should_not_be_optimized:19; LA64: # %bb.0:20; LA64-NEXT: ld.d $a0, $a0, 021; LA64-NEXT: vreplgr2vr.d $vr0, $a022; LA64-NEXT: st.d $a0, $a1, 023; LA64-NEXT: ret24 %tmp = load i64, ptr %ptr25 store i64 %tmp, ptr %dst26 %tmp1 = insertelement <2 x i64> zeroinitializer, i64 %tmp, i32 027 %tmp2 = shufflevector <2 x i64> %tmp1, <2 x i64> poison, <2 x i32> zeroinitializer28 ret <2 x i64> %tmp229}30 31define <8 x i16> @should_not_be_optimized_sext_load(ptr %ptr) {32; CHECK-LABEL: should_not_be_optimized_sext_load:33; CHECK: # %bb.0:34; CHECK-NEXT: ld.b $a0, $a0, 035; CHECK-NEXT: vreplgr2vr.h $vr0, $a036; CHECK-NEXT: ret37 %tmp = load i8, ptr %ptr38 %tmp1 = sext i8 %tmp to i1639 %tmp2 = insertelement <8 x i16> zeroinitializer, i16 %tmp1, i32 040 %tmp3 = shufflevector <8 x i16> %tmp2, <8 x i16> poison, <8 x i32> zeroinitializer41 ret <8 x i16> %tmp342}43 44define <8 x i16> @should_not_be_optimized_zext_load(ptr %ptr) {45; CHECK-LABEL: should_not_be_optimized_zext_load:46; CHECK: # %bb.0:47; CHECK-NEXT: ld.bu $a0, $a0, 048; CHECK-NEXT: vreplgr2vr.h $vr0, $a049; CHECK-NEXT: ret50 %tmp = load i8, ptr %ptr51 %tmp1 = zext i8 %tmp to i1652 %tmp2 = insertelement <8 x i16> zeroinitializer, i16 %tmp1, i32 053 %tmp3 = shufflevector <8 x i16> %tmp2, <8 x i16> poison, <8 x i32> zeroinitializer54 ret <8 x i16> %tmp355}56 57define <2 x i64> @vldrepl_d_unaligned_offset(ptr %ptr) {58; LA32-LABEL: vldrepl_d_unaligned_offset:59; LA32: # %bb.0:60; LA32-NEXT: ld.w $a1, $a0, 461; LA32-NEXT: ld.w $a0, $a0, 862; LA32-NEXT: vinsgr2vr.w $vr0, $a1, 063; LA32-NEXT: vinsgr2vr.w $vr0, $a0, 164; LA32-NEXT: vreplvei.d $vr0, $vr0, 065; LA32-NEXT: ret66;67; LA64-LABEL: vldrepl_d_unaligned_offset:68; LA64: # %bb.0:69; LA64-NEXT: addi.d $a0, $a0, 470; LA64-NEXT: vldrepl.d $vr0, $a0, 071; LA64-NEXT: ret72 %p = getelementptr i32, ptr %ptr, i32 173 %tmp = load i64, ptr %p74 %tmp1 = insertelement <2 x i64> zeroinitializer, i64 %tmp, i32 075 %tmp2 = shufflevector <2 x i64> %tmp1, <2 x i64> poison, <2 x i32> zeroinitializer76 ret <2 x i64> %tmp277}78 79define <16 x i8> @vldrepl_b(ptr %ptr) {80; CHECK-LABEL: vldrepl_b:81; CHECK: # %bb.0:82; CHECK-NEXT: vldrepl.b $vr0, $a0, 083; CHECK-NEXT: ret84 %tmp = load i8, ptr %ptr85 %tmp1 = insertelement <16 x i8> zeroinitializer, i8 %tmp, i32 086 %tmp2 = shufflevector <16 x i8> %tmp1, <16 x i8> poison, <16 x i32> zeroinitializer87 ret <16 x i8> %tmp288}89 90define <16 x i8> @vldrepl_b_offset(ptr %ptr) {91; CHECK-LABEL: vldrepl_b_offset:92; CHECK: # %bb.0:93; CHECK-NEXT: vldrepl.b $vr0, $a0, 3394; CHECK-NEXT: ret95 %p = getelementptr i8, ptr %ptr, i64 3396 %tmp = load i8, ptr %p97 %tmp1 = insertelement <16 x i8> zeroinitializer, i8 %tmp, i32 098 %tmp2 = shufflevector <16 x i8> %tmp1, <16 x i8> poison, <16 x i32> zeroinitializer99 ret <16 x i8> %tmp2100}101 102 103define <8 x i16> @vldrepl_h(ptr %ptr) {104; CHECK-LABEL: vldrepl_h:105; CHECK: # %bb.0:106; CHECK-NEXT: vldrepl.h $vr0, $a0, 0107; CHECK-NEXT: ret108 %tmp = load i16, ptr %ptr109 %tmp1 = insertelement <8 x i16> zeroinitializer, i16 %tmp, i32 0110 %tmp2 = shufflevector <8 x i16> %tmp1, <8 x i16> poison, <8 x i32> zeroinitializer111 ret <8 x i16> %tmp2112}113 114define <8 x i16> @vldrepl_h_offset(ptr %ptr) {115; CHECK-LABEL: vldrepl_h_offset:116; CHECK: # %bb.0:117; CHECK-NEXT: vldrepl.h $vr0, $a0, 66118; CHECK-NEXT: ret119 %p = getelementptr i16, ptr %ptr, i64 33120 %tmp = load i16, ptr %p121 %tmp1 = insertelement <8 x i16> zeroinitializer, i16 %tmp, i32 0122 %tmp2 = shufflevector <8 x i16> %tmp1, <8 x i16> poison, <8 x i32> zeroinitializer123 ret <8 x i16> %tmp2124}125 126define <4 x i32> @vldrepl_w(ptr %ptr) {127; CHECK-LABEL: vldrepl_w:128; CHECK: # %bb.0:129; CHECK-NEXT: vldrepl.w $vr0, $a0, 0130; CHECK-NEXT: ret131 %tmp = load i32, ptr %ptr132 %tmp1 = insertelement <4 x i32> zeroinitializer, i32 %tmp, i32 0133 %tmp2 = shufflevector <4 x i32> %tmp1, <4 x i32> poison, <4 x i32> zeroinitializer134 ret <4 x i32> %tmp2135}136 137define <4 x i32> @vldrepl_w_offset(ptr %ptr) {138; CHECK-LABEL: vldrepl_w_offset:139; CHECK: # %bb.0:140; CHECK-NEXT: vldrepl.w $vr0, $a0, 132141; CHECK-NEXT: ret142 %p = getelementptr i32, ptr %ptr, i64 33143 %tmp = load i32, ptr %p144 %tmp1 = insertelement <4 x i32> zeroinitializer, i32 %tmp, i32 0145 %tmp2 = shufflevector <4 x i32> %tmp1, <4 x i32> poison, <4 x i32> zeroinitializer146 ret <4 x i32> %tmp2147}148 149define <2 x i64> @vldrepl_d(ptr %ptr) {150; LA32-LABEL: vldrepl_d:151; LA32: # %bb.0:152; LA32-NEXT: ld.w $a1, $a0, 0153; LA32-NEXT: ld.w $a0, $a0, 4154; LA32-NEXT: vinsgr2vr.w $vr0, $a1, 0155; LA32-NEXT: vinsgr2vr.w $vr0, $a0, 1156; LA32-NEXT: vreplvei.d $vr0, $vr0, 0157; LA32-NEXT: ret158;159; LA64-LABEL: vldrepl_d:160; LA64: # %bb.0:161; LA64-NEXT: vldrepl.d $vr0, $a0, 0162; LA64-NEXT: ret163 %tmp = load i64, ptr %ptr164 %tmp1 = insertelement <2 x i64> zeroinitializer, i64 %tmp, i32 0165 %tmp2 = shufflevector <2 x i64> %tmp1, <2 x i64> poison, <2 x i32> zeroinitializer166 ret <2 x i64> %tmp2167}168 169define <2 x i64> @vldrepl_d_offset(ptr %ptr) {170; LA32-LABEL: vldrepl_d_offset:171; LA32: # %bb.0:172; LA32-NEXT: ld.w $a1, $a0, 264173; LA32-NEXT: ld.w $a0, $a0, 268174; LA32-NEXT: vinsgr2vr.w $vr0, $a1, 0175; LA32-NEXT: vinsgr2vr.w $vr0, $a0, 1176; LA32-NEXT: vreplvei.d $vr0, $vr0, 0177; LA32-NEXT: ret178;179; LA64-LABEL: vldrepl_d_offset:180; LA64: # %bb.0:181; LA64-NEXT: vldrepl.d $vr0, $a0, 264182; LA64-NEXT: ret183 %p = getelementptr i64, ptr %ptr, i64 33184 %tmp = load i64, ptr %p185 %tmp1 = insertelement <2 x i64> zeroinitializer, i64 %tmp, i32 0186 %tmp2 = shufflevector <2 x i64> %tmp1, <2 x i64> poison, <2 x i32> zeroinitializer187 ret <2 x i64> %tmp2188}189 190define <4 x float> @vldrepl_w_flt(ptr %ptr) {191; CHECK-LABEL: vldrepl_w_flt:192; CHECK: # %bb.0:193; CHECK-NEXT: vldrepl.w $vr0, $a0, 0194; CHECK-NEXT: ret195 %tmp = load float, ptr %ptr196 %tmp1 = insertelement <4 x float> zeroinitializer, float %tmp, i32 0197 %tmp2 = shufflevector <4 x float> %tmp1, <4 x float> poison, <4 x i32> zeroinitializer198 ret <4 x float> %tmp2199}200 201define <4 x float> @vldrepl_w_flt_offset(ptr %ptr) {202; CHECK-LABEL: vldrepl_w_flt_offset:203; CHECK: # %bb.0:204; CHECK-NEXT: vldrepl.w $vr0, $a0, 264205; CHECK-NEXT: ret206 %p = getelementptr i64, ptr %ptr, i64 33207 %tmp = load float, ptr %p208 %tmp1 = insertelement <4 x float> zeroinitializer, float %tmp, i32 0209 %tmp2 = shufflevector <4 x float> %tmp1, <4 x float> poison, <4 x i32> zeroinitializer210 ret <4 x float> %tmp2211}212 213define <2 x double> @vldrepl_d_dbl(ptr %ptr) {214; CHECK-LABEL: vldrepl_d_dbl:215; CHECK: # %bb.0:216; CHECK-NEXT: vldrepl.d $vr0, $a0, 0217; CHECK-NEXT: ret218 %tmp = load double, ptr %ptr219 %tmp1 = insertelement <2 x double> zeroinitializer, double %tmp, i32 0220 %tmp2 = shufflevector <2 x double> %tmp1, <2 x double> poison, <2 x i32> zeroinitializer221 ret <2 x double> %tmp2222}223 224define <2 x double> @vldrepl_d_dbl_offset(ptr %ptr) {225; CHECK-LABEL: vldrepl_d_dbl_offset:226; CHECK: # %bb.0:227; CHECK-NEXT: vldrepl.d $vr0, $a0, 264228; CHECK-NEXT: ret229 %p = getelementptr i64, ptr %ptr, i64 33230 %tmp = load double, ptr %p231 %tmp1 = insertelement <2 x double> zeroinitializer, double %tmp, i32 0232 %tmp2 = shufflevector <2 x double> %tmp1, <2 x double> poison, <2 x i32> zeroinitializer233 ret <2 x double> %tmp2234}235