571 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py2; RUN: llc -mtriple=hexagon -mattr=+hvxv60,+hvx-length128b,-packets < %s | FileCheck --check-prefix=V60 %s3; RUN: llc -mtriple=hexagon -mattr=+hvxv62,+hvx-length128b,-packets < %s | FileCheck --check-prefix=V62 %s4 5define <32 x i32> @f0(<32 x i32> %a0, <32 x i32> %a1) #0 {6; V60-LABEL: f0:7; V60: // %bb.0:8; V60-NEXT: {9; V60-NEXT: r0 = #1610; V60-NEXT: }11; V60-NEXT: {12; V60-NEXT: v2.w = vmpye(v1.w,v0.uh)13; V60-NEXT: }14; V60-NEXT: {15; V60-NEXT: v31.w = vasr(v0.w,r0)16; V60-NEXT: }17; V60-NEXT: {18; V60-NEXT: v3.w = vasr(v1.w,r0)19; V60-NEXT: }20; V60-NEXT: {21; V60-NEXT: v5:4.w = vmpy(v31.h,v1.uh)22; V60-NEXT: }23; V60-NEXT: {24; V60-NEXT: v31:30.w = vmpy(v31.h,v3.h)25; V60-NEXT: }26; V60-NEXT: {27; V60-NEXT: v7:6.w = vadd(v2.uh,v4.uh)28; V60-NEXT: }29; V60-NEXT: {30; V60-NEXT: v29:28.w = vadd(v2.h,v4.h)31; V60-NEXT: }32; V60-NEXT: {33; V60-NEXT: v29.w += vasr(v6.w,r0)34; V60-NEXT: }35; V60-NEXT: {36; V60-NEXT: v0.w = vadd(v29.w,v30.w)37; V60-NEXT: }38; V60-NEXT: {39; V60-NEXT: jumpr r3140; V60-NEXT: }41;42; V62-LABEL: f0:43; V62: // %bb.0:44; V62-NEXT: {45; V62-NEXT: v3:2 = vmpye(v0.w,v1.uh)46; V62-NEXT: }47; V62-NEXT: {48; V62-NEXT: v3:2 += vmpyo(v0.w,v1.h)49; V62-NEXT: }50; V62-NEXT: {51; V62-NEXT: v0 = v352; V62-NEXT: }53; V62-NEXT: {54; V62-NEXT: jumpr r3155; V62-NEXT: }56 %v0 = call {<32 x i32>, <32 x i32>} @llvm.hexagon.V6.vmpyss.parts.128B(<32 x i32> %a0, <32 x i32> %a1)57 %v1 = extractvalue {<32 x i32>, <32 x i32>} %v0, 058 ret <32 x i32> %v159}60 61define <32 x i32> @f1(<32 x i32> %a0, <32 x i32> %a1) #0 {62; V60-LABEL: f1:63; V60: // %bb.0:64; V60-NEXT: {65; V60-NEXT: v2.w = vmpyieo(v0.h,v1.h)66; V60-NEXT: }67; V60-NEXT: {68; V60-NEXT: v2.w += vmpyie(v0.w,v1.uh)69; V60-NEXT: }70; V60-NEXT: {71; V60-NEXT: v0 = v272; V60-NEXT: }73; V60-NEXT: {74; V60-NEXT: jumpr r3175; V60-NEXT: }76;77; V62-LABEL: f1:78; V62: // %bb.0:79; V62-NEXT: {80; V62-NEXT: v2.w = vmpyieo(v0.h,v1.h)81; V62-NEXT: }82; V62-NEXT: {83; V62-NEXT: v2.w += vmpyie(v0.w,v1.uh)84; V62-NEXT: }85; V62-NEXT: {86; V62-NEXT: v0 = v287; V62-NEXT: }88; V62-NEXT: {89; V62-NEXT: jumpr r3190; V62-NEXT: }91 %v0 = call {<32 x i32>, <32 x i32>} @llvm.hexagon.V6.vmpyss.parts.128B(<32 x i32> %a0, <32 x i32> %a1)92 %v1 = extractvalue {<32 x i32>, <32 x i32>} %v0, 193 ret <32 x i32> %v194}95 96define <32 x i32> @f2(<32 x i32> %a0, <32 x i32> %a1) #0 {97; V60-LABEL: f2:98; V60: // %bb.0:99; V60-NEXT: {100; V60-NEXT: r1:0 = combine(#16,##33686018)101; V60-NEXT: }102; V60-NEXT: {103; V60-NEXT: v3:2.uw = vmpy(v0.uh,v1.uh)104; V60-NEXT: }105; V60-NEXT: {106; V60-NEXT: v30 = vxor(v30,v30)107; V60-NEXT: }108; V60-NEXT: {109; V60-NEXT: v4 = vsplat(r0)110; V60-NEXT: }111; V60-NEXT: {112; V60-NEXT: q0 = vcmp.gt(v30.w,v1.w)113; V60-NEXT: }114; V60-NEXT: {115; V60-NEXT: v2.uw = vlsr(v2.uw,r1)116; V60-NEXT: }117; V60-NEXT: {118; V60-NEXT: v4 = vdelta(v1,v4)119; V60-NEXT: }120; V60-NEXT: {121; V60-NEXT: v5:4.uw = vmpy(v0.uh,v4.uh)122; V60-NEXT: }123; V60-NEXT: {124; V60-NEXT: v5:4.w = vadd(v5.uh,v4.uh)125; V60-NEXT: }126; V60-NEXT: {127; V60-NEXT: v2.w = vadd(v4.w,v2.w)128; V60-NEXT: }129; V60-NEXT: {130; V60-NEXT: v5.w += vasr(v2.w,r1)131; V60-NEXT: }132; V60-NEXT: {133; V60-NEXT: v31.w = vadd(v3.w,v5.w)134; V60-NEXT: }135; V60-NEXT: {136; V60-NEXT: if (q0) v31.w -= v0.w137; V60-NEXT: }138; V60-NEXT: {139; V60-NEXT: v0 = v31140; V60-NEXT: }141; V60-NEXT: {142; V60-NEXT: jumpr r31143; V60-NEXT: }144;145; V62-LABEL: f2:146; V62: // %bb.0:147; V62-NEXT: {148; V62-NEXT: v3:2 = vmpye(v0.w,v1.uh)149; V62-NEXT: }150; V62-NEXT: {151; V62-NEXT: v4 = vxor(v4,v4)152; V62-NEXT: }153; V62-NEXT: {154; V62-NEXT: q0 = vcmp.gt(v4.w,v0.w)155; V62-NEXT: }156; V62-NEXT: {157; V62-NEXT: v3:2 += vmpyo(v0.w,v1.h)158; V62-NEXT: }159; V62-NEXT: {160; V62-NEXT: if (q0) v3.w += v1.w161; V62-NEXT: }162; V62-NEXT: {163; V62-NEXT: v0 = v3164; V62-NEXT: }165; V62-NEXT: {166; V62-NEXT: jumpr r31167; V62-NEXT: }168 %v0 = call {<32 x i32>, <32 x i32>} @llvm.hexagon.V6.vmpyus.parts.128B(<32 x i32> %a0, <32 x i32> %a1)169 %v1 = extractvalue {<32 x i32>, <32 x i32>} %v0, 0170 ret <32 x i32> %v1171}172 173define <32 x i32> @f3(<32 x i32> %a0, <32 x i32> %a1) #0 {174; V60-LABEL: f3:175; V60: // %bb.0:176; V60-NEXT: {177; V60-NEXT: v2.w = vmpyieo(v0.h,v1.h)178; V60-NEXT: }179; V60-NEXT: {180; V60-NEXT: v2.w += vmpyie(v0.w,v1.uh)181; V60-NEXT: }182; V60-NEXT: {183; V60-NEXT: v0 = v2184; V60-NEXT: }185; V60-NEXT: {186; V60-NEXT: jumpr r31187; V60-NEXT: }188;189; V62-LABEL: f3:190; V62: // %bb.0:191; V62-NEXT: {192; V62-NEXT: v2.w = vmpyieo(v0.h,v1.h)193; V62-NEXT: }194; V62-NEXT: {195; V62-NEXT: v2.w += vmpyie(v0.w,v1.uh)196; V62-NEXT: }197; V62-NEXT: {198; V62-NEXT: v0 = v2199; V62-NEXT: }200; V62-NEXT: {201; V62-NEXT: jumpr r31202; V62-NEXT: }203 %v0 = call {<32 x i32>, <32 x i32>} @llvm.hexagon.V6.vmpyus.parts.128B(<32 x i32> %a0, <32 x i32> %a1)204 %v1 = extractvalue {<32 x i32>, <32 x i32>} %v0, 1205 ret <32 x i32> %v1206}207 208define <32 x i32> @f4(<32 x i32> %a0, <32 x i32> %a1) #0 {209; V60-LABEL: f4:210; V60: // %bb.0:211; V60-NEXT: {212; V60-NEXT: r0 = ##33686018213; V60-NEXT: }214; V60-NEXT: {215; V60-NEXT: v3:2.uw = vmpy(v0.uh,v1.uh)216; V60-NEXT: }217; V60-NEXT: {218; V60-NEXT: r2 = #16219; V60-NEXT: }220; V60-NEXT: {221; V60-NEXT: v4 = vsplat(r0)222; V60-NEXT: }223; V60-NEXT: {224; V60-NEXT: v2.uw = vlsr(v2.uw,r2)225; V60-NEXT: }226; V60-NEXT: {227; V60-NEXT: v31 = vdelta(v1,v4)228; V60-NEXT: }229; V60-NEXT: {230; V60-NEXT: v1:0.uw = vmpy(v0.uh,v31.uh)231; V60-NEXT: }232; V60-NEXT: {233; V60-NEXT: v1:0.w = vadd(v1.uh,v0.uh)234; V60-NEXT: }235; V60-NEXT: {236; V60-NEXT: v0.w = vadd(v0.w,v2.w)237; V60-NEXT: }238; V60-NEXT: {239; V60-NEXT: v1.w += vasr(v0.w,r2)240; V60-NEXT: }241; V60-NEXT: {242; V60-NEXT: v0.w = vadd(v3.w,v1.w)243; V60-NEXT: }244; V60-NEXT: {245; V60-NEXT: jumpr r31246; V60-NEXT: }247;248; V62-LABEL: f4:249; V62: // %bb.0:250; V62-NEXT: {251; V62-NEXT: v2 = vxor(v2,v2)252; V62-NEXT: }253; V62-NEXT: {254; V62-NEXT: v5:4 = vmpye(v0.w,v1.uh)255; V62-NEXT: }256; V62-NEXT: {257; V62-NEXT: q0 = vcmp.gt(v2.w,v0.w)258; V62-NEXT: }259; V62-NEXT: {260; V62-NEXT: q1 = vcmp.gt(v2.w,v1.w)261; V62-NEXT: }262; V62-NEXT: {263; V62-NEXT: v5:4 += vmpyo(v0.w,v1.h)264; V62-NEXT: }265; V62-NEXT: {266; V62-NEXT: v31 = vand(q0,v1)267; V62-NEXT: }268; V62-NEXT: {269; V62-NEXT: if (q1) v31.w += v0.w270; V62-NEXT: }271; V62-NEXT: {272; V62-NEXT: v0.w = vadd(v5.w,v31.w)273; V62-NEXT: }274; V62-NEXT: {275; V62-NEXT: jumpr r31276; V62-NEXT: }277 %v0 = call {<32 x i32>, <32 x i32>} @llvm.hexagon.V6.vmpyuu.parts.128B(<32 x i32> %a0, <32 x i32> %a1)278 %v1 = extractvalue {<32 x i32>, <32 x i32>} %v0, 0279 ret <32 x i32> %v1280}281 282define <32 x i32> @f5(<32 x i32> %a0, <32 x i32> %a1) #0 {283; V60-LABEL: f5:284; V60: // %bb.0:285; V60-NEXT: {286; V60-NEXT: v2.w = vmpyieo(v0.h,v1.h)287; V60-NEXT: }288; V60-NEXT: {289; V60-NEXT: v2.w += vmpyie(v0.w,v1.uh)290; V60-NEXT: }291; V60-NEXT: {292; V60-NEXT: v0 = v2293; V60-NEXT: }294; V60-NEXT: {295; V60-NEXT: jumpr r31296; V60-NEXT: }297;298; V62-LABEL: f5:299; V62: // %bb.0:300; V62-NEXT: {301; V62-NEXT: v2.w = vmpyieo(v0.h,v1.h)302; V62-NEXT: }303; V62-NEXT: {304; V62-NEXT: v2.w += vmpyie(v0.w,v1.uh)305; V62-NEXT: }306; V62-NEXT: {307; V62-NEXT: v0 = v2308; V62-NEXT: }309; V62-NEXT: {310; V62-NEXT: jumpr r31311; V62-NEXT: }312 %v0 = call {<32 x i32>, <32 x i32>} @llvm.hexagon.V6.vmpyuu.parts.128B(<32 x i32> %a0, <32 x i32> %a1)313 %v1 = extractvalue {<32 x i32>, <32 x i32>} %v0, 1314 ret <32 x i32> %v1315}316 317define <64 x i32> @f10(<32 x i32> %a0, <32 x i32> %a1) #0 {318; V60-LABEL: f10:319; V60: // %bb.0:320; V60-NEXT: {321; V60-NEXT: r0 = ##33686018322; V60-NEXT: }323; V60-NEXT: {324; V60-NEXT: v3:2 = vcombine(v0,v1)325; V60-NEXT: }326; V60-NEXT: {327; V60-NEXT: v1:0.uw = vmpy(v0.uh,v1.uh)328; V60-NEXT: }329; V60-NEXT: {330; V60-NEXT: r2 = #16331; V60-NEXT: }332; V60-NEXT: {333; V60-NEXT: v5 = vsplat(r0)334; V60-NEXT: }335; V60-NEXT: {336; V60-NEXT: v4 = vxor(v4,v4)337; V60-NEXT: }338; V60-NEXT: {339; V60-NEXT: v6.uw = vlsr(v0.uw,r2)340; V60-NEXT: }341; V60-NEXT: {342; V60-NEXT: q1 = vcmp.gt(v4.w,v3.w)343; V60-NEXT: }344; V60-NEXT: {345; V60-NEXT: q0 = vcmp.gt(v4.w,v2.w)346; V60-NEXT: }347; V60-NEXT: {348; V60-NEXT: v5 = vdelta(v2,v5)349; V60-NEXT: }350; V60-NEXT: {351; V60-NEXT: v2 = vmux(q1,v2,v4)352; V60-NEXT: }353; V60-NEXT: {354; V60-NEXT: if (q0) v2.w += v3.w355; V60-NEXT: }356; V60-NEXT: {357; V60-NEXT: v9:8.uw = vmpy(v3.uh,v5.uh)358; V60-NEXT: }359; V60-NEXT: {360; V60-NEXT: v9:8.w = vadd(v9.uh,v8.uh)361; V60-NEXT: }362; V60-NEXT: {363; V60-NEXT: v31.w = vadd(v8.w,v6.w)364; V60-NEXT: }365; V60-NEXT: {366; V60-NEXT: v0.w += vasl(v8.w,r2)367; V60-NEXT: }368; V60-NEXT: {369; V60-NEXT: v9.w += vasr(v31.w,r2)370; V60-NEXT: }371; V60-NEXT: {372; V60-NEXT: v1.w = vadd(v1.w,v9.w)373; V60-NEXT: }374; V60-NEXT: {375; V60-NEXT: v1.w = vsub(v1.w,v2.w)376; V60-NEXT: }377; V60-NEXT: {378; V60-NEXT: jumpr r31379; V60-NEXT: }380;381; V62-LABEL: f10:382; V62: // %bb.0:383; V62-NEXT: {384; V62-NEXT: v3:2 = vmpye(v0.w,v1.uh)385; V62-NEXT: }386; V62-NEXT: {387; V62-NEXT: v3:2 += vmpyo(v0.w,v1.h)388; V62-NEXT: }389; V62-NEXT: {390; V62-NEXT: v1:0 = vcombine(v3,v2)391; V62-NEXT: }392; V62-NEXT: {393; V62-NEXT: jumpr r31394; V62-NEXT: }395 %v0 = call {<32 x i32>, <32 x i32>} @llvm.hexagon.V6.vmpyss.parts.128B(<32 x i32> %a0, <32 x i32> %a1)396 %v1 = extractvalue {<32 x i32>, <32 x i32>} %v0, 0397 %v2 = extractvalue {<32 x i32>, <32 x i32>} %v0, 1398 %v3 = shufflevector <32 x i32> %v2, <32 x i32> %v1, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 32, i32 33, i32 34, i32 35, i32 36, i32 37, i32 38, i32 39, i32 40, i32 41, i32 42, i32 43, i32 44, i32 45, i32 46, i32 47, i32 48, i32 49, i32 50, i32 51, i32 52, i32 53, i32 54, i32 55, i32 56, i32 57, i32 58, i32 59, i32 60, i32 61, i32 62, i32 63>399 ret <64 x i32> %v3400}401 402define <64 x i32> @f11(<32 x i32> %a0, <32 x i32> %a1) #0 {403; V60-LABEL: f11:404; V60: // %bb.0:405; V60-NEXT: {406; V60-NEXT: r1:0 = combine(#16,##33686018)407; V60-NEXT: }408; V60-NEXT: {409; V60-NEXT: v3:2.uw = vmpy(v0.uh,v1.uh)410; V60-NEXT: }411; V60-NEXT: {412; V60-NEXT: v31 = vxor(v31,v31)413; V60-NEXT: }414; V60-NEXT: {415; V60-NEXT: v4 = vsplat(r0)416; V60-NEXT: }417; V60-NEXT: {418; V60-NEXT: q0 = vcmp.gt(v31.w,v1.w)419; V60-NEXT: }420; V60-NEXT: {421; V60-NEXT: v5.uw = vlsr(v2.uw,r1)422; V60-NEXT: }423; V60-NEXT: {424; V60-NEXT: v4 = vdelta(v1,v4)425; V60-NEXT: }426; V60-NEXT: {427; V60-NEXT: v7:6.uw = vmpy(v0.uh,v4.uh)428; V60-NEXT: }429; V60-NEXT: {430; V60-NEXT: v7:6.w = vadd(v7.uh,v6.uh)431; V60-NEXT: }432; V60-NEXT: {433; V60-NEXT: v30.w = vadd(v6.w,v5.w)434; V60-NEXT: }435; V60-NEXT: {436; V60-NEXT: v2.w += vasl(v6.w,r1)437; V60-NEXT: }438; V60-NEXT: {439; V60-NEXT: v7.w += vasr(v30.w,r1)440; V60-NEXT: }441; V60-NEXT: {442; V60-NEXT: v3.w = vadd(v3.w,v7.w)443; V60-NEXT: }444; V60-NEXT: {445; V60-NEXT: if (q0) v3.w -= v0.w446; V60-NEXT: }447; V60-NEXT: {448; V60-NEXT: v1:0 = vcombine(v3,v2)449; V60-NEXT: }450; V60-NEXT: {451; V60-NEXT: jumpr r31452; V60-NEXT: }453;454; V62-LABEL: f11:455; V62: // %bb.0:456; V62-NEXT: {457; V62-NEXT: v3:2 = vmpye(v0.w,v1.uh)458; V62-NEXT: }459; V62-NEXT: {460; V62-NEXT: v4 = vxor(v4,v4)461; V62-NEXT: }462; V62-NEXT: {463; V62-NEXT: q0 = vcmp.gt(v4.w,v0.w)464; V62-NEXT: }465; V62-NEXT: {466; V62-NEXT: v3:2 += vmpyo(v0.w,v1.h)467; V62-NEXT: }468; V62-NEXT: {469; V62-NEXT: if (q0) v3.w += v1.w470; V62-NEXT: }471; V62-NEXT: {472; V62-NEXT: v1:0 = vcombine(v3,v2)473; V62-NEXT: }474; V62-NEXT: {475; V62-NEXT: jumpr r31476; V62-NEXT: }477 %v0 = call {<32 x i32>, <32 x i32>} @llvm.hexagon.V6.vmpyus.parts.128B(<32 x i32> %a0, <32 x i32> %a1)478 %v1 = extractvalue {<32 x i32>, <32 x i32>} %v0, 0479 %v2 = extractvalue {<32 x i32>, <32 x i32>} %v0, 1480 %v3 = shufflevector <32 x i32> %v2, <32 x i32> %v1, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 32, i32 33, i32 34, i32 35, i32 36, i32 37, i32 38, i32 39, i32 40, i32 41, i32 42, i32 43, i32 44, i32 45, i32 46, i32 47, i32 48, i32 49, i32 50, i32 51, i32 52, i32 53, i32 54, i32 55, i32 56, i32 57, i32 58, i32 59, i32 60, i32 61, i32 62, i32 63>481 ret <64 x i32> %v3482}483 484define <64 x i32> @f12(<32 x i32> %a0, <32 x i32> %a1) #0 {485; V60-LABEL: f12:486; V60: // %bb.0:487; V60-NEXT: {488; V60-NEXT: r1:0 = combine(#16,##33686018)489; V60-NEXT: }490; V60-NEXT: {491; V60-NEXT: v3:2.uw = vmpy(v0.uh,v1.uh)492; V60-NEXT: }493; V60-NEXT: {494; V60-NEXT: v4 = vsplat(r0)495; V60-NEXT: }496; V60-NEXT: {497; V60-NEXT: v5.uw = vlsr(v2.uw,r1)498; V60-NEXT: }499; V60-NEXT: {500; V60-NEXT: v30 = vdelta(v1,v4)501; V60-NEXT: }502; V60-NEXT: {503; V60-NEXT: v1:0.uw = vmpy(v0.uh,v30.uh)504; V60-NEXT: }505; V60-NEXT: {506; V60-NEXT: v1:0.w = vadd(v1.uh,v0.uh)507; V60-NEXT: }508; V60-NEXT: {509; V60-NEXT: v31.w = vadd(v0.w,v5.w)510; V60-NEXT: }511; V60-NEXT: {512; V60-NEXT: v2.w += vasl(v0.w,r1)513; V60-NEXT: }514; V60-NEXT: {515; V60-NEXT: v1.w += vasr(v31.w,r1)516; V60-NEXT: }517; V60-NEXT: {518; V60-NEXT: v3.w = vadd(v3.w,v1.w)519; V60-NEXT: }520; V60-NEXT: {521; V60-NEXT: v1:0 = vcombine(v3,v2)522; V60-NEXT: }523; V60-NEXT: {524; V60-NEXT: jumpr r31525; V60-NEXT: }526;527; V62-LABEL: f12:528; V62: // %bb.0:529; V62-NEXT: {530; V62-NEXT: v4 = vxor(v4,v4)531; V62-NEXT: }532; V62-NEXT: {533; V62-NEXT: v3:2 = vmpye(v0.w,v1.uh)534; V62-NEXT: }535; V62-NEXT: {536; V62-NEXT: q0 = vcmp.gt(v4.w,v0.w)537; V62-NEXT: }538; V62-NEXT: {539; V62-NEXT: q1 = vcmp.gt(v4.w,v1.w)540; V62-NEXT: }541; V62-NEXT: {542; V62-NEXT: v3:2 += vmpyo(v0.w,v1.h)543; V62-NEXT: }544; V62-NEXT: {545; V62-NEXT: v31 = vand(q0,v1)546; V62-NEXT: }547; V62-NEXT: {548; V62-NEXT: if (q1) v31.w += v0.w549; V62-NEXT: }550; V62-NEXT: {551; V62-NEXT: v3.w = vadd(v3.w,v31.w)552; V62-NEXT: }553; V62-NEXT: {554; V62-NEXT: v1:0 = vcombine(v3,v2)555; V62-NEXT: }556; V62-NEXT: {557; V62-NEXT: jumpr r31558; V62-NEXT: }559 %v0 = call {<32 x i32>, <32 x i32>} @llvm.hexagon.V6.vmpyuu.parts.128B(<32 x i32> %a0, <32 x i32> %a1)560 %v1 = extractvalue {<32 x i32>, <32 x i32>} %v0, 0561 %v2 = extractvalue {<32 x i32>, <32 x i32>} %v0, 1562 %v3 = shufflevector <32 x i32> %v2, <32 x i32> %v1, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 32, i32 33, i32 34, i32 35, i32 36, i32 37, i32 38, i32 39, i32 40, i32 41, i32 42, i32 43, i32 44, i32 45, i32 46, i32 47, i32 48, i32 49, i32 50, i32 51, i32 52, i32 53, i32 54, i32 55, i32 56, i32 57, i32 58, i32 59, i32 60, i32 61, i32 62, i32 63>563 ret <64 x i32> %v3564}565 566declare {<32 x i32>, <32 x i32>} @llvm.hexagon.V6.vmpyss.parts.128B(<32 x i32>, <32 x i32>)567declare {<32 x i32>, <32 x i32>} @llvm.hexagon.V6.vmpyus.parts.128B(<32 x i32>, <32 x i32>)568declare {<32 x i32>, <32 x i32>} @llvm.hexagon.V6.vmpyuu.parts.128B(<32 x i32>, <32 x i32>)569 570attributes #0 = { nounwind }571