253 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 5define void @shuffle_any_ext_2i8_to_2i64(ptr %ptr, ptr %dst) nounwind {6; CHECK-LABEL: shuffle_any_ext_2i8_to_2i64:7; CHECK: # %bb.0:8; CHECK-NEXT: ld.h $a0, $a0, 09; CHECK-NEXT: vinsgr2vr.h $vr0, $a0, 010; CHECK-NEXT: vilvl.b $vr0, $vr0, $vr011; CHECK-NEXT: vilvl.h $vr0, $vr0, $vr012; CHECK-NEXT: vilvl.w $vr0, $vr0, $vr013; CHECK-NEXT: vst $vr0, $a1, 014; CHECK-NEXT: ret15 %x = load <2 x i8>, ptr %ptr16 %y = shufflevector <2 x i8> %x, <2 x i8> poison, <16 x i32> <i32 0, i32 3, i32 3, i32 3, i32 3, i32 3, i32 3, i32 3, i32 1, i32 2, i32 2, i32 2,i32 2, i32 2, i32 2, i32 2>17 %r = bitcast <16 x i8> %y to <2 x i64>18 store <2 x i64> %r, ptr %dst19 ret void20}21 22define void @shuffle_any_ext_2i16_to_2i64(ptr %ptr, ptr %dst) nounwind {23; CHECK-LABEL: shuffle_any_ext_2i16_to_2i64:24; CHECK: # %bb.0:25; CHECK-NEXT: ld.w $a0, $a0, 026; CHECK-NEXT: vinsgr2vr.w $vr0, $a0, 027; CHECK-NEXT: vilvl.h $vr0, $vr0, $vr028; CHECK-NEXT: vilvl.w $vr0, $vr0, $vr029; CHECK-NEXT: vst $vr0, $a1, 030; CHECK-NEXT: ret31 %x = load <2 x i16>, ptr %ptr32 %y = shufflevector <2 x i16> %x, <2 x i16> poison, <8 x i32> <i32 0, i32 3, i32 3, i32 3, i32 1, i32 2, i32 2, i32 2>33 %r = bitcast <8 x i16> %y to <2 x i64>34 store <2 x i64> %r, ptr %dst35 ret void36}37 38define void @shuffle_any_ext_2i32_to_2i64(ptr %ptr, ptr %dst) nounwind {39; LA32-LABEL: shuffle_any_ext_2i32_to_2i64:40; LA32: # %bb.0:41; LA32-NEXT: ld.w $a2, $a0, 042; LA32-NEXT: ld.w $a0, $a0, 443; LA32-NEXT: vinsgr2vr.w $vr0, $a2, 044; LA32-NEXT: vinsgr2vr.w $vr0, $a0, 245; LA32-NEXT: vst $vr0, $a1, 046; LA32-NEXT: ret47;48; LA64-LABEL: shuffle_any_ext_2i32_to_2i64:49; LA64: # %bb.0:50; LA64-NEXT: ld.d $a0, $a0, 051; LA64-NEXT: vinsgr2vr.d $vr0, $a0, 052; LA64-NEXT: vshuf4i.w $vr0, $vr0, 1653; LA64-NEXT: vst $vr0, $a1, 054; LA64-NEXT: ret55 %x = load <2 x i32>, ptr %ptr56 %y = shufflevector <2 x i32> %x, <2 x i32> poison, <4 x i32> <i32 0, i32 3, i32 1, i32 2>57 %r = bitcast <4 x i32> %y to <2 x i64>58 store <2 x i64> %r, ptr %dst59 ret void60}61 62define void @shuffle_any_ext_4i8_to_4i32(ptr %ptr, ptr %dst) nounwind {63; CHECK-LABEL: shuffle_any_ext_4i8_to_4i32:64; CHECK: # %bb.0:65; CHECK-NEXT: ld.w $a0, $a0, 066; CHECK-NEXT: vinsgr2vr.w $vr0, $a0, 067; CHECK-NEXT: vilvl.b $vr0, $vr0, $vr068; CHECK-NEXT: vilvl.h $vr0, $vr0, $vr069; CHECK-NEXT: vst $vr0, $a1, 070; CHECK-NEXT: ret71 %x = load <4 x i8>, ptr %ptr72 %y = shufflevector <4 x i8> %x, <4 x i8> poison, <16 x i32> <i32 0, i32 7, i32 7, i32 7, i32 1, i32 6, i32 6, i32 6, i32 2, i32 5, i32 5, i32 5, i32 3, i32 4, i32 4, i32 4>73 %r = bitcast <16 x i8> %y to <4 x i32>74 store <4 x i32> %r, ptr %dst75 ret void76}77 78define void @shuffle_any_ext_4i16_to_4i32(ptr %ptr, ptr %dst) nounwind {79; LA32-LABEL: shuffle_any_ext_4i16_to_4i32:80; LA32: # %bb.0:81; LA32-NEXT: ld.w $a2, $a0, 082; LA32-NEXT: ld.w $a0, $a0, 483; LA32-NEXT: vinsgr2vr.w $vr0, $a2, 084; LA32-NEXT: vinsgr2vr.w $vr0, $a0, 185; LA32-NEXT: vilvl.h $vr0, $vr0, $vr086; LA32-NEXT: vst $vr0, $a1, 087; LA32-NEXT: ret88;89; LA64-LABEL: shuffle_any_ext_4i16_to_4i32:90; LA64: # %bb.0:91; LA64-NEXT: ld.d $a0, $a0, 092; LA64-NEXT: vinsgr2vr.d $vr0, $a0, 093; LA64-NEXT: vilvl.h $vr0, $vr0, $vr094; LA64-NEXT: vst $vr0, $a1, 095; LA64-NEXT: ret96 %x = load <4 x i16>, ptr %ptr97 %y = shufflevector <4 x i16> %x, <4 x i16> poison, <8 x i32> <i32 0, i32 7, i32 1, i32 6, i32 2, i32 5, i32 3, i32 4>98 %r = bitcast <8 x i16> %y to <4 x i32>99 store <4 x i32> %r, ptr %dst100 ret void101}102 103define void @shuffle_any_ext_8i8_to_8i16(ptr %ptr, ptr %dst) nounwind {104; LA32-LABEL: shuffle_any_ext_8i8_to_8i16:105; LA32: # %bb.0:106; LA32-NEXT: ld.w $a2, $a0, 0107; LA32-NEXT: ld.w $a0, $a0, 4108; LA32-NEXT: vinsgr2vr.w $vr0, $a2, 0109; LA32-NEXT: vinsgr2vr.w $vr0, $a0, 1110; LA32-NEXT: vilvl.b $vr0, $vr0, $vr0111; LA32-NEXT: vst $vr0, $a1, 0112; LA32-NEXT: ret113;114; LA64-LABEL: shuffle_any_ext_8i8_to_8i16:115; LA64: # %bb.0:116; LA64-NEXT: ld.d $a0, $a0, 0117; LA64-NEXT: vinsgr2vr.d $vr0, $a0, 0118; LA64-NEXT: vilvl.b $vr0, $vr0, $vr0119; LA64-NEXT: vst $vr0, $a1, 0120; LA64-NEXT: ret121 %x = load <8 x i8>, ptr %ptr122 %y = shufflevector <8 x i8> %x, <8 x i8> poison, <16 x i32> <i32 0, i32 15, i32 1, i32 14, i32 2, i32 13, i32 3, i32 12, i32 4, i32 11, i32 5, i32 10, i32 6, i32 9, i32 7, i32 8>123 %r = bitcast <16 x i8> %y to <8 x i16>124 store <8 x i16> %r, ptr %dst125 ret void126}127 128define void @shuffle_any_ext_4i32_to_4i64(ptr %ptr, ptr %dst) nounwind {129; CHECK-LABEL: shuffle_any_ext_4i32_to_4i64:130; CHECK: # %bb.0:131; CHECK-NEXT: vld $vr0, $a0, 0132; CHECK-NEXT: vshuf4i.w $vr1, $vr0, 16133; CHECK-NEXT: vshuf4i.w $vr0, $vr0, 50134; CHECK-NEXT: vst $vr0, $a1, 16135; CHECK-NEXT: vst $vr1, $a1, 0136; CHECK-NEXT: ret137 %x = load <4 x i32>, ptr %ptr138 %y = shufflevector <4 x i32> %x, <4 x i32> poison, <8 x i32> <i32 0, i32 7, i32 1, i32 6, i32 2, i32 5, i32 3, i32 4>139 %r = bitcast <8 x i32> %y to <4 x i64>140 store <4 x i64> %r, ptr %dst141 ret void142}143 144define void @shuffle_any_ext_8i16_to_8i32(ptr %ptr, ptr %dst) nounwind {145; CHECK-LABEL: shuffle_any_ext_8i16_to_8i32:146; CHECK: # %bb.0:147; CHECK-NEXT: vld $vr0, $a0, 0148; CHECK-NEXT: vilvl.h $vr1, $vr0, $vr0149; CHECK-NEXT: vilvh.h $vr0, $vr0, $vr0150; CHECK-NEXT: vst $vr0, $a1, 16151; CHECK-NEXT: vst $vr1, $a1, 0152; CHECK-NEXT: ret153 %x = load <8 x i16>, ptr %ptr154 %y = shufflevector <8 x i16> %x, <8 x i16> poison, <16 x i32> <i32 0, i32 15, i32 1, i32 15, i32 2, i32 13, i32 3, i32 12, i32 4, i32 11, i32 5, i32 10, i32 6, i32 9, i32 7, i32 8>155 %r = bitcast <16 x i16> %y to <8 x i32>156 store <8 x i32> %r, ptr %dst157 ret void158}159 160define void @shuffle_any_ext_8i16_to_8i64(ptr %ptr, ptr %dst) nounwind {161; CHECK-LABEL: shuffle_any_ext_8i16_to_8i64:162; CHECK: # %bb.0:163; CHECK-NEXT: vld $vr0, $a0, 0164; CHECK-NEXT: vilvl.h $vr1, $vr0, $vr0165; CHECK-NEXT: vilvl.w $vr2, $vr1, $vr1166; CHECK-NEXT: vilvh.w $vr1, $vr1, $vr1167; CHECK-NEXT: vilvh.h $vr0, $vr0, $vr0168; CHECK-NEXT: vilvl.w $vr3, $vr0, $vr0169; CHECK-NEXT: vilvh.w $vr0, $vr0, $vr0170; CHECK-NEXT: vst $vr0, $a1, 48171; CHECK-NEXT: vst $vr3, $a1, 32172; CHECK-NEXT: vst $vr1, $a1, 16173; CHECK-NEXT: vst $vr2, $a1, 0174; CHECK-NEXT: ret175 %x = load <8 x i16>, ptr %ptr176 %y = shufflevector <8 x i16> %x, <8 x i16> poison, <32 x i32> <i32 0, i32 15, i32 15, i32 15, i32 1, i32 14, i32 14, i32 14, i32 2, i32 13, i32 13, i32 13, i32 3, i32 12, i32 12, i32 12, i32 4, i32 11, i32 11, i32 11, i32 5, i32 10, i32 10, i32 10, i32 6, i32 9, i32 9, i32 9, i32 7, i32 8, i32 8, i32 8>177 %r = bitcast <32 x i16> %y to <8 x i64>178 store <8 x i64> %r, ptr %dst179 ret void180}181 182define void @shuffle_any_ext_16i8_to_16i16(ptr %ptr, ptr %dst) nounwind {183; CHECK-LABEL: shuffle_any_ext_16i8_to_16i16:184; CHECK: # %bb.0:185; CHECK-NEXT: vld $vr0, $a0, 0186; CHECK-NEXT: vilvl.b $vr1, $vr0, $vr0187; CHECK-NEXT: vilvh.b $vr0, $vr0, $vr0188; CHECK-NEXT: vst $vr0, $a1, 16189; CHECK-NEXT: vst $vr1, $a1, 0190; CHECK-NEXT: ret191 %x = load <16 x i8>, ptr %ptr192 %y = shufflevector <16 x i8> %x, <16 x i8> poison, <32 x i32> <i32 0, i32 31, i32 1, i32 31, i32 2, i32 29, i32 3, i32 28, i32 4, i32 27, i32 5, i32 26, i32 6, i32 25, i32 7, i32 24, i32 8, i32 23, i32 9, i32 22, i32 10, i32 21, i32 11, i32 20, i32 12, i32 19, i32 13, i32 18, i32 14, i32 17, i32 15, i32 16 >193 %r = bitcast <32 x i8> %y to <16 x i16>194 store <16 x i16> %r, ptr %dst195 ret void196}197 198define void @shuffle_any_ext_16i8_to_16i32(ptr %ptr, ptr %dst) nounwind {199; CHECK-LABEL: shuffle_any_ext_16i8_to_16i32:200; CHECK: # %bb.0:201; CHECK-NEXT: vld $vr0, $a0, 0202; CHECK-NEXT: vilvl.b $vr1, $vr0, $vr0203; CHECK-NEXT: vilvl.h $vr2, $vr1, $vr1204; CHECK-NEXT: vilvh.h $vr1, $vr1, $vr1205; CHECK-NEXT: vilvh.b $vr0, $vr0, $vr0206; CHECK-NEXT: vilvl.h $vr3, $vr0, $vr0207; CHECK-NEXT: vilvh.h $vr0, $vr0, $vr0208; CHECK-NEXT: vst $vr0, $a1, 48209; CHECK-NEXT: vst $vr3, $a1, 32210; CHECK-NEXT: vst $vr1, $a1, 16211; CHECK-NEXT: vst $vr2, $a1, 0212; CHECK-NEXT: ret213 %x = load <16 x i8>, ptr %ptr214 %y = shufflevector <16 x i8> %x, <16 x i8> poison, <64 x i32> <i32 0, i32 31, i32 31, i32 31, i32 1, i32 30, i32 30, i32 30, i32 2, i32 29, i32 29, i32 29, i32 3, i32 28, i32 28, i32 28, i32 4, i32 27, i32 27, i32 27, i32 5, i32 26, i32 26, i32 26, i32 6, i32 25, i32 25, i32 25, i32 7, i32 24, i32 24, i32 24, i32 8, i32 23, i32 23, i32 23, i32 9, i32 22, i32 22, i32 22, i32 10, i32 21, i32 21, i32 21, i32 11, i32 20, i32 20, i32 20, i32 12, i32 19, i32 19, i32 19, i32 13, i32 18, i32 18, i32 18, i32 14, i32 17, i32 17, i32 17, i32 15, i32 16, i32 16, i32 16>215 %r = bitcast <64 x i8> %y to <16 x i32>216 store <16 x i32> %r, ptr %dst217 ret void218}219 220define void @shuffle_any_ext_16i8_to_16i64(ptr %ptr, ptr %dst) nounwind {221; CHECK-LABEL: shuffle_any_ext_16i8_to_16i64:222; CHECK: # %bb.0:223; CHECK-NEXT: vld $vr0, $a0, 0224; CHECK-NEXT: vilvl.b $vr1, $vr0, $vr0225; CHECK-NEXT: vilvl.h $vr2, $vr1, $vr1226; CHECK-NEXT: vilvl.w $vr3, $vr2, $vr2227; CHECK-NEXT: vilvh.w $vr2, $vr2, $vr2228; CHECK-NEXT: vilvh.h $vr1, $vr1, $vr1229; CHECK-NEXT: vilvl.w $vr4, $vr1, $vr1230; CHECK-NEXT: vilvh.w $vr1, $vr1, $vr1231; CHECK-NEXT: vilvh.b $vr0, $vr0, $vr0232; CHECK-NEXT: vilvl.h $vr5, $vr0, $vr0233; CHECK-NEXT: vilvl.w $vr6, $vr5, $vr5234; CHECK-NEXT: vilvh.w $vr5, $vr5, $vr5235; CHECK-NEXT: vilvh.h $vr0, $vr0, $vr0236; CHECK-NEXT: vilvl.w $vr7, $vr0, $vr0237; CHECK-NEXT: vilvh.w $vr0, $vr0, $vr0238; CHECK-NEXT: vst $vr0, $a1, 112239; CHECK-NEXT: vst $vr7, $a1, 96240; CHECK-NEXT: vst $vr5, $a1, 80241; CHECK-NEXT: vst $vr6, $a1, 64242; CHECK-NEXT: vst $vr1, $a1, 48243; CHECK-NEXT: vst $vr4, $a1, 32244; CHECK-NEXT: vst $vr2, $a1, 16245; CHECK-NEXT: vst $vr3, $a1, 0246; CHECK-NEXT: ret247 %x = load <16 x i8>, ptr %ptr248 %y = shufflevector <16 x i8> %x, <16 x i8> poison, <128 x i32> <i32 0, i32 31, i32 31, i32 31, i32 31, i32 31, i32 31, i32 31, i32 1, i32 30, i32 30, i32 30, i32 30, i32 30, i32 30, i32 30, i32 2, i32 29, i32 29, i32 29, i32 29, i32 29, i32 29, i32 29, i32 3, i32 28, i32 28, i32 28, i32 28, i32 28, i32 28, i32 28, i32 4, i32 27, i32 27, i32 27, i32 27, i32 27, i32 27, i32 27, i32 5, i32 26, i32 26, i32 26, i32 26, i32 26, i32 26, i32 26, i32 6, i32 25, i32 25, i32 25, i32 25, i32 25, i32 25, i32 25, i32 7, i32 24, i32 24, i32 24, i32 24, i32 24, i32 24, i32 24, i32 8, i32 23, i32 23, i32 23, i32 23, i32 23, i32 23, i32 23, i32 9, i32 22, i32 22, i32 22, i32 22, i32 22, i32 22, i32 22, i32 10, i32 21, i32 21, i32 21, i32 21, i32 21, i32 21, i32 21, i32 11, i32 20, i32 20, i32 20, i32 20, i32 20, i32 20, i32 20, i32 12, i32 19, i32 19, i32 19, i32 19, i32 19, i32 19, i32 19, i32 13, i32 18, i32 18, i32 18, i32 18, i32 18, i32 18, i32 18, i32 14, i32 17, i32 17, i32 17, i32 17, i32 17, i32 17, i32 17, i32 15, i32 16, i32 16, i32 16, i32 16, i32 16, i32 16, i32 16>249 %r = bitcast <128 x i8> %y to <16 x i64>250 store <16 x i64> %r, ptr %dst251 ret void252}253