1017 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,+lasx --verify-machineinstrs < %s | FileCheck %s --check-prefixes=CHECK,LA323; RUN: llc --mtriple=loongarch64 --mattr=+lasx --verify-machineinstrs < %s | FileCheck %s --check-prefixes=CHECK,LA644 5define i32 @xmsk_eq_allzeros_i8(<32 x i8 > %a) {6; LA32-LABEL: xmsk_eq_allzeros_i8:7; LA32: # %bb.0: # %entry8; LA32-NEXT: xvmsknz.b $xr0, $xr09; LA32-NEXT: xvnor.v $xr0, $xr0, $xr010; LA32-NEXT: xvpickve2gr.wu $a0, $xr0, 011; LA32-NEXT: xvpickve2gr.wu $a1, $xr0, 412; LA32-NEXT: bstrins.w $a0, $a1, 31, 1613; LA32-NEXT: ret14;15; LA64-LABEL: xmsk_eq_allzeros_i8:16; LA64: # %bb.0: # %entry17; LA64-NEXT: xvmsknz.b $xr0, $xr018; LA64-NEXT: xvnor.v $xr0, $xr0, $xr019; LA64-NEXT: xvpickve2gr.wu $a0, $xr0, 020; LA64-NEXT: xvpickve2gr.wu $a1, $xr0, 421; LA64-NEXT: bstrins.d $a0, $a1, 31, 1622; LA64-NEXT: ret23entry:24 %1 = icmp eq <32 x i8> %a, splat (i8 0)25 %2 = bitcast <32 x i1> %1 to i3226 ret i32 %227}28 29define i32 @xmsk_sgt_allzeros_i8(<32 x i8 > %a) {30; LA32-LABEL: xmsk_sgt_allzeros_i8:31; LA32: # %bb.0: # %entry32; LA32-NEXT: xvrepli.b $xr1, 033; LA32-NEXT: xvslt.b $xr0, $xr1, $xr034; LA32-NEXT: xvmskltz.b $xr0, $xr035; LA32-NEXT: xvpickve2gr.wu $a0, $xr0, 036; LA32-NEXT: xvpickve2gr.wu $a1, $xr0, 437; LA32-NEXT: bstrins.w $a0, $a1, 31, 1638; LA32-NEXT: ret39;40; LA64-LABEL: xmsk_sgt_allzeros_i8:41; LA64: # %bb.0: # %entry42; LA64-NEXT: xvrepli.b $xr1, 043; LA64-NEXT: xvslt.b $xr0, $xr1, $xr044; LA64-NEXT: xvmskltz.b $xr0, $xr045; LA64-NEXT: xvpickve2gr.wu $a0, $xr0, 046; LA64-NEXT: xvpickve2gr.wu $a1, $xr0, 447; LA64-NEXT: bstrins.d $a0, $a1, 31, 1648; LA64-NEXT: ret49entry:50 %1 = icmp sgt <32 x i8> %a, splat (i8 0)51 %2 = bitcast <32 x i1> %1 to i3252 ret i32 %253}54 55define i32 @xmsk_sgt_allones_i8(<32 x i8 > %a) {56; LA32-LABEL: xmsk_sgt_allones_i8:57; LA32: # %bb.0: # %entry58; LA32-NEXT: xvmskgez.b $xr0, $xr059; LA32-NEXT: xvpickve2gr.wu $a0, $xr0, 060; LA32-NEXT: xvpickve2gr.wu $a1, $xr0, 461; LA32-NEXT: bstrins.w $a0, $a1, 31, 1662; LA32-NEXT: ret63;64; LA64-LABEL: xmsk_sgt_allones_i8:65; LA64: # %bb.0: # %entry66; LA64-NEXT: xvmskgez.b $xr0, $xr067; LA64-NEXT: xvpickve2gr.wu $a0, $xr0, 068; LA64-NEXT: xvpickve2gr.wu $a1, $xr0, 469; LA64-NEXT: bstrins.d $a0, $a1, 31, 1670; LA64-NEXT: ret71entry:72 %1 = icmp sgt <32 x i8> %a, splat (i8 -1)73 %2 = bitcast <32 x i1> %1 to i3274 ret i32 %275}76 77define i32 @xmsk_sge_allzeros_i8(<32 x i8 > %a) {78; LA32-LABEL: xmsk_sge_allzeros_i8:79; LA32: # %bb.0: # %entry80; LA32-NEXT: xvmskgez.b $xr0, $xr081; LA32-NEXT: xvpickve2gr.wu $a0, $xr0, 082; LA32-NEXT: xvpickve2gr.wu $a1, $xr0, 483; LA32-NEXT: bstrins.w $a0, $a1, 31, 1684; LA32-NEXT: ret85;86; LA64-LABEL: xmsk_sge_allzeros_i8:87; LA64: # %bb.0: # %entry88; LA64-NEXT: xvmskgez.b $xr0, $xr089; LA64-NEXT: xvpickve2gr.wu $a0, $xr0, 090; LA64-NEXT: xvpickve2gr.wu $a1, $xr0, 491; LA64-NEXT: bstrins.d $a0, $a1, 31, 1692; LA64-NEXT: ret93entry:94 %1 = icmp sge <32 x i8> %a, splat (i8 0)95 %2 = bitcast <32 x i1> %1 to i3296 ret i32 %297}98 99define i32 @xmsk_slt_allzeros_i8(<32 x i8 > %a) {100; LA32-LABEL: xmsk_slt_allzeros_i8:101; LA32: # %bb.0: # %entry102; LA32-NEXT: xvmskltz.b $xr0, $xr0103; LA32-NEXT: xvpickve2gr.wu $a0, $xr0, 0104; LA32-NEXT: xvpickve2gr.wu $a1, $xr0, 4105; LA32-NEXT: bstrins.w $a0, $a1, 31, 16106; LA32-NEXT: ret107;108; LA64-LABEL: xmsk_slt_allzeros_i8:109; LA64: # %bb.0: # %entry110; LA64-NEXT: xvmskltz.b $xr0, $xr0111; LA64-NEXT: xvpickve2gr.wu $a0, $xr0, 0112; LA64-NEXT: xvpickve2gr.wu $a1, $xr0, 4113; LA64-NEXT: bstrins.d $a0, $a1, 31, 16114; LA64-NEXT: ret115entry:116 %1 = icmp slt <32 x i8> %a, splat (i8 0)117 %2 = bitcast <32 x i1> %1 to i32118 ret i32 %2119}120 121define i16 @xmsk_slt_allzeros_i16(<16 x i16 > %a) {122; LA32-LABEL: xmsk_slt_allzeros_i16:123; LA32: # %bb.0: # %entry124; LA32-NEXT: xvmskltz.h $xr0, $xr0125; LA32-NEXT: xvpickve2gr.wu $a0, $xr0, 0126; LA32-NEXT: xvpickve2gr.wu $a1, $xr0, 4127; LA32-NEXT: bstrins.w $a0, $a1, 15, 8128; LA32-NEXT: ret129;130; LA64-LABEL: xmsk_slt_allzeros_i16:131; LA64: # %bb.0: # %entry132; LA64-NEXT: xvmskltz.h $xr0, $xr0133; LA64-NEXT: xvpickve2gr.wu $a0, $xr0, 0134; LA64-NEXT: xvpickve2gr.wu $a1, $xr0, 4135; LA64-NEXT: bstrins.d $a0, $a1, 15, 8136; LA64-NEXT: ret137entry:138 %1 = icmp slt <16 x i16> %a, splat (i16 0)139 %2 = bitcast <16 x i1> %1 to i16140 ret i16 %2141}142 143define i8 @xmsk_slt_allzeros_i32(<8 x i32 > %a) {144; LA32-LABEL: xmsk_slt_allzeros_i32:145; LA32: # %bb.0: # %entry146; LA32-NEXT: xvmskltz.w $xr0, $xr0147; LA32-NEXT: xvpickve2gr.wu $a0, $xr0, 0148; LA32-NEXT: xvpickve2gr.wu $a1, $xr0, 4149; LA32-NEXT: bstrins.w $a0, $a1, 7, 4150; LA32-NEXT: ret151;152; LA64-LABEL: xmsk_slt_allzeros_i32:153; LA64: # %bb.0: # %entry154; LA64-NEXT: xvmskltz.w $xr0, $xr0155; LA64-NEXT: xvpickve2gr.wu $a0, $xr0, 0156; LA64-NEXT: xvpickve2gr.wu $a1, $xr0, 4157; LA64-NEXT: bstrins.d $a0, $a1, 7, 4158; LA64-NEXT: ret159entry:160 %1 = icmp slt <8 x i32> %a, splat (i32 0)161 %2 = bitcast <8 x i1> %1 to i8162 ret i8 %2163}164 165define i4 @xmsk_slt_allzeros_i64(<4 x i64 > %a) {166; LA32-LABEL: xmsk_slt_allzeros_i64:167; LA32: # %bb.0: # %entry168; LA32-NEXT: xvmskltz.d $xr0, $xr0169; LA32-NEXT: xvpickve2gr.wu $a0, $xr0, 0170; LA32-NEXT: xvpickve2gr.wu $a1, $xr0, 4171; LA32-NEXT: bstrins.w $a0, $a1, 3, 2172; LA32-NEXT: ret173;174; LA64-LABEL: xmsk_slt_allzeros_i64:175; LA64: # %bb.0: # %entry176; LA64-NEXT: xvmskltz.d $xr0, $xr0177; LA64-NEXT: xvpickve2gr.wu $a0, $xr0, 0178; LA64-NEXT: xvpickve2gr.wu $a1, $xr0, 4179; LA64-NEXT: bstrins.d $a0, $a1, 3, 2180; LA64-NEXT: ret181entry:182 %1 = icmp slt <4 x i64> %a, splat (i64 0)183 %2 = bitcast <4 x i1> %1 to i4184 ret i4 %2185}186 187define i32 @xmsk_sle_allzeros_i8(<32 x i8 > %a) {188; LA32-LABEL: xmsk_sle_allzeros_i8:189; LA32: # %bb.0: # %entry190; LA32-NEXT: xvslei.b $xr0, $xr0, 0191; LA32-NEXT: xvmskltz.b $xr0, $xr0192; LA32-NEXT: xvpickve2gr.wu $a0, $xr0, 0193; LA32-NEXT: xvpickve2gr.wu $a1, $xr0, 4194; LA32-NEXT: bstrins.w $a0, $a1, 31, 16195; LA32-NEXT: ret196;197; LA64-LABEL: xmsk_sle_allzeros_i8:198; LA64: # %bb.0: # %entry199; LA64-NEXT: xvslei.b $xr0, $xr0, 0200; LA64-NEXT: xvmskltz.b $xr0, $xr0201; LA64-NEXT: xvpickve2gr.wu $a0, $xr0, 0202; LA64-NEXT: xvpickve2gr.wu $a1, $xr0, 4203; LA64-NEXT: bstrins.d $a0, $a1, 31, 16204; LA64-NEXT: ret205entry:206 %1 = icmp sle <32 x i8> %a, splat (i8 0)207 %2 = bitcast <32 x i1> %1 to i32208 ret i32 %2209}210 211define i32 @xmsk_sle_allones_i8(<32 x i8 > %a) {212; LA32-LABEL: xmsk_sle_allones_i8:213; LA32: # %bb.0: # %entry214; LA32-NEXT: xvmskltz.b $xr0, $xr0215; LA32-NEXT: xvpickve2gr.wu $a0, $xr0, 0216; LA32-NEXT: xvpickve2gr.wu $a1, $xr0, 4217; LA32-NEXT: bstrins.w $a0, $a1, 31, 16218; LA32-NEXT: ret219;220; LA64-LABEL: xmsk_sle_allones_i8:221; LA64: # %bb.0: # %entry222; LA64-NEXT: xvmskltz.b $xr0, $xr0223; LA64-NEXT: xvpickve2gr.wu $a0, $xr0, 0224; LA64-NEXT: xvpickve2gr.wu $a1, $xr0, 4225; LA64-NEXT: bstrins.d $a0, $a1, 31, 16226; LA64-NEXT: ret227entry:228 %1 = icmp sle <32 x i8> %a, splat (i8 -1)229 %2 = bitcast <32 x i1> %1 to i32230 ret i32 %2231}232 233define i16 @xmsk_sle_allones_i32(<16 x i16 > %a) {234; LA32-LABEL: xmsk_sle_allones_i32:235; LA32: # %bb.0: # %entry236; LA32-NEXT: xvmskltz.h $xr0, $xr0237; LA32-NEXT: xvpickve2gr.wu $a0, $xr0, 0238; LA32-NEXT: xvpickve2gr.wu $a1, $xr0, 4239; LA32-NEXT: bstrins.w $a0, $a1, 15, 8240; LA32-NEXT: ret241;242; LA64-LABEL: xmsk_sle_allones_i32:243; LA64: # %bb.0: # %entry244; LA64-NEXT: xvmskltz.h $xr0, $xr0245; LA64-NEXT: xvpickve2gr.wu $a0, $xr0, 0246; LA64-NEXT: xvpickve2gr.wu $a1, $xr0, 4247; LA64-NEXT: bstrins.d $a0, $a1, 15, 8248; LA64-NEXT: ret249entry:250 %1 = icmp sle <16 x i16> %a, splat (i16 -1)251 %2 = bitcast <16 x i1> %1 to i16252 ret i16 %2253}254 255define i8 @xmsk_sle_allones_i16(<8 x i32 > %a) {256; LA32-LABEL: xmsk_sle_allones_i16:257; LA32: # %bb.0: # %entry258; LA32-NEXT: xvmskltz.w $xr0, $xr0259; LA32-NEXT: xvpickve2gr.wu $a0, $xr0, 0260; LA32-NEXT: xvpickve2gr.wu $a1, $xr0, 4261; LA32-NEXT: bstrins.w $a0, $a1, 7, 4262; LA32-NEXT: ret263;264; LA64-LABEL: xmsk_sle_allones_i16:265; LA64: # %bb.0: # %entry266; LA64-NEXT: xvmskltz.w $xr0, $xr0267; LA64-NEXT: xvpickve2gr.wu $a0, $xr0, 0268; LA64-NEXT: xvpickve2gr.wu $a1, $xr0, 4269; LA64-NEXT: bstrins.d $a0, $a1, 7, 4270; LA64-NEXT: ret271entry:272 %1 = icmp sle <8 x i32> %a, splat (i32 -1)273 %2 = bitcast <8 x i1> %1 to i8274 ret i8 %2275}276 277define i4 @xmsk_sle_allones_i64(<4 x i64 > %a) {278; LA32-LABEL: xmsk_sle_allones_i64:279; LA32: # %bb.0: # %entry280; LA32-NEXT: xvmskltz.d $xr0, $xr0281; LA32-NEXT: xvpickve2gr.wu $a0, $xr0, 0282; LA32-NEXT: xvpickve2gr.wu $a1, $xr0, 4283; LA32-NEXT: bstrins.w $a0, $a1, 3, 2284; LA32-NEXT: ret285;286; LA64-LABEL: xmsk_sle_allones_i64:287; LA64: # %bb.0: # %entry288; LA64-NEXT: xvmskltz.d $xr0, $xr0289; LA64-NEXT: xvpickve2gr.wu $a0, $xr0, 0290; LA64-NEXT: xvpickve2gr.wu $a1, $xr0, 4291; LA64-NEXT: bstrins.d $a0, $a1, 3, 2292; LA64-NEXT: ret293entry:294 %1 = icmp sle <4 x i64> %a, splat (i64 -1)295 %2 = bitcast <4 x i1> %1 to i4296 ret i4 %2297}298 299define i32 @xmsk_ne_allzeros_i8(<32 x i8 > %a) {300; LA32-LABEL: xmsk_ne_allzeros_i8:301; LA32: # %bb.0: # %entry302; LA32-NEXT: xvmsknz.b $xr0, $xr0303; LA32-NEXT: xvpickve2gr.wu $a0, $xr0, 0304; LA32-NEXT: xvpickve2gr.wu $a1, $xr0, 4305; LA32-NEXT: bstrins.w $a0, $a1, 31, 16306; LA32-NEXT: ret307;308; LA64-LABEL: xmsk_ne_allzeros_i8:309; LA64: # %bb.0: # %entry310; LA64-NEXT: xvmsknz.b $xr0, $xr0311; LA64-NEXT: xvpickve2gr.wu $a0, $xr0, 0312; LA64-NEXT: xvpickve2gr.wu $a1, $xr0, 4313; LA64-NEXT: bstrins.d $a0, $a1, 31, 16314; LA64-NEXT: ret315entry:316 %1 = icmp ne <32 x i8> %a, splat (i8 0)317 %2 = bitcast <32 x i1> %1 to i32318 ret i32 %2319}320 321define i4 @xvmsk_sgt_v4i64(<4 x i64> %a, <4 x i64> %b) {322; LA32-LABEL: xvmsk_sgt_v4i64:323; LA32: # %bb.0:324; LA32-NEXT: xvslt.d $xr0, $xr1, $xr0325; LA32-NEXT: xvmskltz.d $xr0, $xr0326; LA32-NEXT: xvpickve2gr.wu $a0, $xr0, 0327; LA32-NEXT: xvpickve2gr.wu $a1, $xr0, 4328; LA32-NEXT: bstrins.w $a0, $a1, 3, 2329; LA32-NEXT: ret330;331; LA64-LABEL: xvmsk_sgt_v4i64:332; LA64: # %bb.0:333; LA64-NEXT: xvslt.d $xr0, $xr1, $xr0334; LA64-NEXT: xvmskltz.d $xr0, $xr0335; LA64-NEXT: xvpickve2gr.wu $a0, $xr0, 0336; LA64-NEXT: xvpickve2gr.wu $a1, $xr0, 4337; LA64-NEXT: bstrins.d $a0, $a1, 3, 2338; LA64-NEXT: ret339 %x = icmp sgt <4 x i64> %a, %b340 %res = bitcast <4 x i1> %x to i4341 ret i4 %res342}343 344define i4 @xvmsk_ogt_v4f64(<4 x double> %a, <4 x double> %b) {345; LA32-LABEL: xvmsk_ogt_v4f64:346; LA32: # %bb.0:347; LA32-NEXT: xvfcmp.clt.d $xr0, $xr1, $xr0348; LA32-NEXT: xvmskltz.d $xr0, $xr0349; LA32-NEXT: xvpickve2gr.wu $a0, $xr0, 0350; LA32-NEXT: xvpickve2gr.wu $a1, $xr0, 4351; LA32-NEXT: bstrins.w $a0, $a1, 3, 2352; LA32-NEXT: ret353;354; LA64-LABEL: xvmsk_ogt_v4f64:355; LA64: # %bb.0:356; LA64-NEXT: xvfcmp.clt.d $xr0, $xr1, $xr0357; LA64-NEXT: xvmskltz.d $xr0, $xr0358; LA64-NEXT: xvpickve2gr.wu $a0, $xr0, 0359; LA64-NEXT: xvpickve2gr.wu $a1, $xr0, 4360; LA64-NEXT: bstrins.d $a0, $a1, 3, 2361; LA64-NEXT: ret362 %x = fcmp ogt <4 x double> %a, %b363 %res = bitcast <4 x i1> %x to i4364 ret i4 %res365}366 367define i8 @xvmsk_sgt_v8i32(<8 x i32> %a, <8 x i32> %b) {368; LA32-LABEL: xvmsk_sgt_v8i32:369; LA32: # %bb.0:370; LA32-NEXT: xvslt.w $xr0, $xr1, $xr0371; LA32-NEXT: xvmskltz.w $xr0, $xr0372; LA32-NEXT: xvpickve2gr.wu $a0, $xr0, 0373; LA32-NEXT: xvpickve2gr.wu $a1, $xr0, 4374; LA32-NEXT: bstrins.w $a0, $a1, 7, 4375; LA32-NEXT: ret376;377; LA64-LABEL: xvmsk_sgt_v8i32:378; LA64: # %bb.0:379; LA64-NEXT: xvslt.w $xr0, $xr1, $xr0380; LA64-NEXT: xvmskltz.w $xr0, $xr0381; LA64-NEXT: xvpickve2gr.wu $a0, $xr0, 0382; LA64-NEXT: xvpickve2gr.wu $a1, $xr0, 4383; LA64-NEXT: bstrins.d $a0, $a1, 7, 4384; LA64-NEXT: ret385 %x = icmp sgt <8 x i32> %a, %b386 %res = bitcast <8 x i1> %x to i8387 ret i8 %res388}389 390define i8 @xvmsk_ogt_v8f32(<8 x float> %a, <8 x float> %b) {391; LA32-LABEL: xvmsk_ogt_v8f32:392; LA32: # %bb.0:393; LA32-NEXT: xvfcmp.clt.s $xr0, $xr1, $xr0394; LA32-NEXT: xvmskltz.w $xr0, $xr0395; LA32-NEXT: xvpickve2gr.wu $a0, $xr0, 0396; LA32-NEXT: xvpickve2gr.wu $a1, $xr0, 4397; LA32-NEXT: bstrins.w $a0, $a1, 7, 4398; LA32-NEXT: ret399;400; LA64-LABEL: xvmsk_ogt_v8f32:401; LA64: # %bb.0:402; LA64-NEXT: xvfcmp.clt.s $xr0, $xr1, $xr0403; LA64-NEXT: xvmskltz.w $xr0, $xr0404; LA64-NEXT: xvpickve2gr.wu $a0, $xr0, 0405; LA64-NEXT: xvpickve2gr.wu $a1, $xr0, 4406; LA64-NEXT: bstrins.d $a0, $a1, 7, 4407; LA64-NEXT: ret408 %x = fcmp ogt <8 x float> %a, %b409 %res = bitcast <8 x i1> %x to i8410 ret i8 %res411}412 413define i16 @xvmsk_sgt_v16i16(<16 x i16> %a, <16 x i16> %b) {414; LA32-LABEL: xvmsk_sgt_v16i16:415; LA32: # %bb.0:416; LA32-NEXT: xvslt.h $xr0, $xr1, $xr0417; LA32-NEXT: xvmskltz.h $xr0, $xr0418; LA32-NEXT: xvpickve2gr.wu $a0, $xr0, 0419; LA32-NEXT: xvpickve2gr.wu $a1, $xr0, 4420; LA32-NEXT: bstrins.w $a0, $a1, 15, 8421; LA32-NEXT: ret422;423; LA64-LABEL: xvmsk_sgt_v16i16:424; LA64: # %bb.0:425; LA64-NEXT: xvslt.h $xr0, $xr1, $xr0426; LA64-NEXT: xvmskltz.h $xr0, $xr0427; LA64-NEXT: xvpickve2gr.wu $a0, $xr0, 0428; LA64-NEXT: xvpickve2gr.wu $a1, $xr0, 4429; LA64-NEXT: bstrins.d $a0, $a1, 15, 8430; LA64-NEXT: ret431 %x = icmp sgt <16 x i16> %a, %b432 %res = bitcast <16 x i1> %x to i16433 ret i16 %res434}435 436define i32 @xvmsk_sgt_v32i8(<32 x i8> %a, <32 x i8> %b) {437; LA32-LABEL: xvmsk_sgt_v32i8:438; LA32: # %bb.0:439; LA32-NEXT: xvslt.b $xr0, $xr1, $xr0440; LA32-NEXT: xvmskltz.b $xr0, $xr0441; LA32-NEXT: xvpickve2gr.wu $a0, $xr0, 0442; LA32-NEXT: xvpickve2gr.wu $a1, $xr0, 4443; LA32-NEXT: bstrins.w $a0, $a1, 31, 16444; LA32-NEXT: ret445;446; LA64-LABEL: xvmsk_sgt_v32i8:447; LA64: # %bb.0:448; LA64-NEXT: xvslt.b $xr0, $xr1, $xr0449; LA64-NEXT: xvmskltz.b $xr0, $xr0450; LA64-NEXT: xvpickve2gr.wu $a0, $xr0, 0451; LA64-NEXT: xvpickve2gr.wu $a1, $xr0, 4452; LA64-NEXT: bstrins.d $a0, $a1, 31, 16453; LA64-NEXT: ret454 %x = icmp sgt <32 x i8> %a, %b455 %res = bitcast <32 x i1> %x to i32456 ret i32 %res457}458 459define i4 @xvmsk_sgt_and_sgt_v4i64(<4 x i64> %a, <4 x i64> %b, <4 x i64> %c, <4 x i64> %d) {460; LA32-LABEL: xvmsk_sgt_and_sgt_v4i64:461; LA32: # %bb.0:462; LA32-NEXT: xvslt.d $xr2, $xr3, $xr2463; LA32-NEXT: xvslt.d $xr0, $xr1, $xr0464; LA32-NEXT: xvand.v $xr0, $xr0, $xr2465; LA32-NEXT: xvmskltz.d $xr0, $xr0466; LA32-NEXT: xvpickve2gr.wu $a0, $xr0, 0467; LA32-NEXT: xvpickve2gr.wu $a1, $xr0, 4468; LA32-NEXT: bstrins.w $a0, $a1, 3, 2469; LA32-NEXT: ret470;471; LA64-LABEL: xvmsk_sgt_and_sgt_v4i64:472; LA64: # %bb.0:473; LA64-NEXT: xvslt.d $xr2, $xr3, $xr2474; LA64-NEXT: xvslt.d $xr0, $xr1, $xr0475; LA64-NEXT: xvand.v $xr0, $xr0, $xr2476; LA64-NEXT: xvmskltz.d $xr0, $xr0477; LA64-NEXT: xvpickve2gr.wu $a0, $xr0, 0478; LA64-NEXT: xvpickve2gr.wu $a1, $xr0, 4479; LA64-NEXT: bstrins.d $a0, $a1, 3, 2480; LA64-NEXT: ret481 %x0 = icmp sgt <4 x i64> %a, %b482 %x1 = icmp sgt <4 x i64> %c, %d483 %y = and <4 x i1> %x0, %x1484 %res = bitcast <4 x i1> %y to i4485 ret i4 %res486}487 488define i4 @xvmsk_ogt_and_ogt_v4f64(<4 x double> %a, <4 x double> %b, <4 x double> %c, <4 x double> %d) {489; LA32-LABEL: xvmsk_ogt_and_ogt_v4f64:490; LA32: # %bb.0:491; LA32-NEXT: xvfcmp.clt.d $xr2, $xr3, $xr2492; LA32-NEXT: xvfcmp.clt.d $xr0, $xr1, $xr0493; LA32-NEXT: xvand.v $xr0, $xr0, $xr2494; LA32-NEXT: xvmskltz.d $xr0, $xr0495; LA32-NEXT: xvpickve2gr.wu $a0, $xr0, 0496; LA32-NEXT: xvpickve2gr.wu $a1, $xr0, 4497; LA32-NEXT: bstrins.w $a0, $a1, 3, 2498; LA32-NEXT: ret499;500; LA64-LABEL: xvmsk_ogt_and_ogt_v4f64:501; LA64: # %bb.0:502; LA64-NEXT: xvfcmp.clt.d $xr2, $xr3, $xr2503; LA64-NEXT: xvfcmp.clt.d $xr0, $xr1, $xr0504; LA64-NEXT: xvand.v $xr0, $xr0, $xr2505; LA64-NEXT: xvmskltz.d $xr0, $xr0506; LA64-NEXT: xvpickve2gr.wu $a0, $xr0, 0507; LA64-NEXT: xvpickve2gr.wu $a1, $xr0, 4508; LA64-NEXT: bstrins.d $a0, $a1, 3, 2509; LA64-NEXT: ret510 %x0 = fcmp ogt <4 x double> %a, %b511 %x1 = fcmp ogt <4 x double> %c, %d512 %y = and <4 x i1> %x0, %x1513 %res = bitcast <4 x i1> %y to i4514 ret i4 %res515}516 517define i8 @xvmsk_sgt_and_sgt_v8i32(<8 x i32> %a, <8 x i32> %b, <8 x i32> %c, <8 x i32> %d) {518; LA32-LABEL: xvmsk_sgt_and_sgt_v8i32:519; LA32: # %bb.0:520; LA32-NEXT: xvslt.w $xr2, $xr3, $xr2521; LA32-NEXT: xvslt.w $xr0, $xr1, $xr0522; LA32-NEXT: xvand.v $xr0, $xr0, $xr2523; LA32-NEXT: xvmskltz.w $xr0, $xr0524; LA32-NEXT: xvpickve2gr.wu $a0, $xr0, 0525; LA32-NEXT: xvpickve2gr.wu $a1, $xr0, 4526; LA32-NEXT: bstrins.w $a0, $a1, 7, 4527; LA32-NEXT: ret528;529; LA64-LABEL: xvmsk_sgt_and_sgt_v8i32:530; LA64: # %bb.0:531; LA64-NEXT: xvslt.w $xr2, $xr3, $xr2532; LA64-NEXT: xvslt.w $xr0, $xr1, $xr0533; LA64-NEXT: xvand.v $xr0, $xr0, $xr2534; LA64-NEXT: xvmskltz.w $xr0, $xr0535; LA64-NEXT: xvpickve2gr.wu $a0, $xr0, 0536; LA64-NEXT: xvpickve2gr.wu $a1, $xr0, 4537; LA64-NEXT: bstrins.d $a0, $a1, 7, 4538; LA64-NEXT: ret539 %x0 = icmp sgt <8 x i32> %a, %b540 %x1 = icmp sgt <8 x i32> %c, %d541 %y = and <8 x i1> %x0, %x1542 %res = bitcast <8 x i1> %y to i8543 ret i8 %res544}545 546define i8 @xvmsk_sgt_or_sgt_v8i32(<8 x i32> %a, <8 x i32> %b, <8 x i32> %c, <8 x i32> %d) {547; LA32-LABEL: xvmsk_sgt_or_sgt_v8i32:548; LA32: # %bb.0:549; LA32-NEXT: xvslt.w $xr2, $xr3, $xr2550; LA32-NEXT: xvslt.w $xr0, $xr1, $xr0551; LA32-NEXT: xvor.v $xr0, $xr0, $xr2552; LA32-NEXT: xvmskltz.w $xr0, $xr0553; LA32-NEXT: xvpickve2gr.wu $a0, $xr0, 0554; LA32-NEXT: xvpickve2gr.wu $a1, $xr0, 4555; LA32-NEXT: bstrins.w $a0, $a1, 7, 4556; LA32-NEXT: ret557;558; LA64-LABEL: xvmsk_sgt_or_sgt_v8i32:559; LA64: # %bb.0:560; LA64-NEXT: xvslt.w $xr2, $xr3, $xr2561; LA64-NEXT: xvslt.w $xr0, $xr1, $xr0562; LA64-NEXT: xvor.v $xr0, $xr0, $xr2563; LA64-NEXT: xvmskltz.w $xr0, $xr0564; LA64-NEXT: xvpickve2gr.wu $a0, $xr0, 0565; LA64-NEXT: xvpickve2gr.wu $a1, $xr0, 4566; LA64-NEXT: bstrins.d $a0, $a1, 7, 4567; LA64-NEXT: ret568 %x0 = icmp sgt <8 x i32> %a, %b569 %x1 = icmp sgt <8 x i32> %c, %d570 %y = or <8 x i1> %x0, %x1571 %res = bitcast <8 x i1> %y to i8572 ret i8 %res573}574 575define i8 @xvmsk_sgt_or_slt_and_eq_v8i32(<8 x i32> %a, <8 x i32> %b, <8 x i32> %c, <8 x i32> %d, <8 x i32> %e, <8 x i32> %f) {576; LA32-LABEL: xvmsk_sgt_or_slt_and_eq_v8i32:577; LA32: # %bb.0:578; LA32-NEXT: xvslt.w $xr2, $xr2, $xr3579; LA32-NEXT: xvslt.w $xr0, $xr1, $xr0580; LA32-NEXT: xvor.v $xr0, $xr0, $xr2581; LA32-NEXT: xvseq.w $xr1, $xr4, $xr5582; LA32-NEXT: xvand.v $xr0, $xr0, $xr1583; LA32-NEXT: xvmskltz.w $xr0, $xr0584; LA32-NEXT: xvpickve2gr.wu $a0, $xr0, 0585; LA32-NEXT: xvpickve2gr.wu $a1, $xr0, 4586; LA32-NEXT: bstrins.w $a0, $a1, 7, 4587; LA32-NEXT: ret588;589; LA64-LABEL: xvmsk_sgt_or_slt_and_eq_v8i32:590; LA64: # %bb.0:591; LA64-NEXT: xvslt.w $xr2, $xr2, $xr3592; LA64-NEXT: xvslt.w $xr0, $xr1, $xr0593; LA64-NEXT: xvor.v $xr0, $xr0, $xr2594; LA64-NEXT: xvseq.w $xr1, $xr4, $xr5595; LA64-NEXT: xvand.v $xr0, $xr0, $xr1596; LA64-NEXT: xvmskltz.w $xr0, $xr0597; LA64-NEXT: xvpickve2gr.wu $a0, $xr0, 0598; LA64-NEXT: xvpickve2gr.wu $a1, $xr0, 4599; LA64-NEXT: bstrins.d $a0, $a1, 7, 4600; LA64-NEXT: ret601 %x0 = icmp sgt <8 x i32> %a, %b602 %x1 = icmp slt <8 x i32> %c, %d603 %x2 = icmp eq <8 x i32> %e, %f604 %y = or <8 x i1> %x0, %x1605 %z = and <8 x i1> %y, %x2606 %res = bitcast <8 x i1> %z to i8607 ret i8 %res608}609 610define i8 @xvmsk_eq_vsel_slt_v8i32(<8 x i32> %a0, <8 x i32> %a1, <8 x i32> %a2) {611; LA32-LABEL: xvmsk_eq_vsel_slt_v8i32:612; LA32: # %bb.0:613; LA32-NEXT: xvseq.w $xr0, $xr0, $xr1614; LA32-NEXT: xvor.v $xr0, $xr0, $xr2615; LA32-NEXT: xvmskltz.w $xr0, $xr0616; LA32-NEXT: xvpickve2gr.wu $a0, $xr0, 0617; LA32-NEXT: xvpickve2gr.wu $a1, $xr0, 4618; LA32-NEXT: bstrins.w $a0, $a1, 7, 4619; LA32-NEXT: ret620;621; LA64-LABEL: xvmsk_eq_vsel_slt_v8i32:622; LA64: # %bb.0:623; LA64-NEXT: xvseq.w $xr0, $xr0, $xr1624; LA64-NEXT: xvor.v $xr0, $xr0, $xr2625; LA64-NEXT: xvmskltz.w $xr0, $xr0626; LA64-NEXT: xvpickve2gr.wu $a0, $xr0, 0627; LA64-NEXT: xvpickve2gr.wu $a1, $xr0, 4628; LA64-NEXT: bstrins.d $a0, $a1, 7, 4629; LA64-NEXT: ret630 %cmp = icmp eq <8 x i32> %a0, %a1631 %slt = icmp slt <8 x i32> %a2, zeroinitializer632 %sel = select <8 x i1> %cmp, <8 x i1> <i1 true, i1 true, i1 true, i1 true, i1 true, i1 true, i1 true, i1 true>, <8 x i1> %slt633 %res = bitcast <8 x i1> %sel to i8634 ret i8 %res635}636 637define i8 @xvmsk_sel_eq_or_eq_or_slt_v8i32(<8 x i32> %a0, <8 x i32> %a1, <8 x i32> %a2, <8 x i32> %a3, i1 %a4) {638; LA32-LABEL: xvmsk_sel_eq_or_eq_or_slt_v8i32:639; LA32: # %bb.0:640; LA32-NEXT: andi $a0, $a0, 1641; LA32-NEXT: xvseq.w $xr2, $xr0, $xr2642; LA32-NEXT: addi.w $a1, $zero, -1643; LA32-NEXT: maskeqz $a0, $a1, $a0644; LA32-NEXT: xvreplgr2vr.w $xr4, $a0645; LA32-NEXT: xvand.v $xr2, $xr2, $xr4646; LA32-NEXT: xvseq.w $xr0, $xr0, $xr1647; LA32-NEXT: xvor.v $xr0, $xr3, $xr0648; LA32-NEXT: xvor.v $xr0, $xr0, $xr2649; LA32-NEXT: xvmskltz.w $xr0, $xr0650; LA32-NEXT: xvpickve2gr.wu $a0, $xr0, 0651; LA32-NEXT: xvpickve2gr.wu $a1, $xr0, 4652; LA32-NEXT: bstrins.w $a0, $a1, 7, 4653; LA32-NEXT: ret654;655; LA64-LABEL: xvmsk_sel_eq_or_eq_or_slt_v8i32:656; LA64: # %bb.0:657; LA64-NEXT: andi $a0, $a0, 1658; LA64-NEXT: xvseq.w $xr2, $xr0, $xr2659; LA64-NEXT: addi.d $a1, $zero, -1660; LA64-NEXT: maskeqz $a0, $a1, $a0661; LA64-NEXT: xvreplgr2vr.w $xr4, $a0662; LA64-NEXT: xvand.v $xr2, $xr2, $xr4663; LA64-NEXT: xvseq.w $xr0, $xr0, $xr1664; LA64-NEXT: xvor.v $xr0, $xr3, $xr0665; LA64-NEXT: xvor.v $xr0, $xr0, $xr2666; LA64-NEXT: xvmskltz.w $xr0, $xr0667; LA64-NEXT: xvpickve2gr.wu $a0, $xr0, 0668; LA64-NEXT: xvpickve2gr.wu $a1, $xr0, 4669; LA64-NEXT: bstrins.d $a0, $a1, 7, 4670; LA64-NEXT: ret671 %cmp0 = icmp eq <8 x i32> %a0, %a1672 %cmp1 = icmp eq <8 x i32> %a0, %a2673 %cmp2 = icmp slt <8 x i32> %a3, zeroinitializer674 %sel = select i1 %a4, <8 x i1> %cmp1, <8 x i1> zeroinitializer675 %or0 = or <8 x i1> %cmp2, %cmp0676 %or1 = or <8 x i1> %or0, %sel677 %res = bitcast <8 x i1> %or1 to i8678 ret i8 %res679}680 681define i8 @xvmsk_ogt_and_ogt_v8f32(<8 x float> %a, <8 x float> %b, <8 x float> %c, <8 x float> %d) {682; LA32-LABEL: xvmsk_ogt_and_ogt_v8f32:683; LA32: # %bb.0:684; LA32-NEXT: xvfcmp.clt.s $xr2, $xr3, $xr2685; LA32-NEXT: xvfcmp.clt.s $xr0, $xr1, $xr0686; LA32-NEXT: xvand.v $xr0, $xr0, $xr2687; LA32-NEXT: xvmskltz.w $xr0, $xr0688; LA32-NEXT: xvpickve2gr.wu $a0, $xr0, 0689; LA32-NEXT: xvpickve2gr.wu $a1, $xr0, 4690; LA32-NEXT: bstrins.w $a0, $a1, 7, 4691; LA32-NEXT: ret692;693; LA64-LABEL: xvmsk_ogt_and_ogt_v8f32:694; LA64: # %bb.0:695; LA64-NEXT: xvfcmp.clt.s $xr2, $xr3, $xr2696; LA64-NEXT: xvfcmp.clt.s $xr0, $xr1, $xr0697; LA64-NEXT: xvand.v $xr0, $xr0, $xr2698; LA64-NEXT: xvmskltz.w $xr0, $xr0699; LA64-NEXT: xvpickve2gr.wu $a0, $xr0, 0700; LA64-NEXT: xvpickve2gr.wu $a1, $xr0, 4701; LA64-NEXT: bstrins.d $a0, $a1, 7, 4702; LA64-NEXT: ret703 %x0 = fcmp ogt <8 x float> %a, %b704 %x1 = fcmp ogt <8 x float> %c, %d705 %y = and <8 x i1> %x0, %x1706 %res = bitcast <8 x i1> %y to i8707 ret i8 %res708}709 710define i8 @xvmsk_sgt_xor_sgt_v8f32(<8 x float> %a, <8 x float> %b, <8 x float> %c, <8 x float> %d) {711; LA32-LABEL: xvmsk_sgt_xor_sgt_v8f32:712; LA32: # %bb.0:713; LA32-NEXT: xvfcmp.clt.s $xr2, $xr3, $xr2714; LA32-NEXT: xvfcmp.clt.s $xr0, $xr1, $xr0715; LA32-NEXT: xvxor.v $xr0, $xr0, $xr2716; LA32-NEXT: xvmskltz.w $xr0, $xr0717; LA32-NEXT: xvpickve2gr.wu $a0, $xr0, 0718; LA32-NEXT: xvpickve2gr.wu $a1, $xr0, 4719; LA32-NEXT: bstrins.w $a0, $a1, 7, 4720; LA32-NEXT: ret721;722; LA64-LABEL: xvmsk_sgt_xor_sgt_v8f32:723; LA64: # %bb.0:724; LA64-NEXT: xvfcmp.clt.s $xr2, $xr3, $xr2725; LA64-NEXT: xvfcmp.clt.s $xr0, $xr1, $xr0726; LA64-NEXT: xvxor.v $xr0, $xr0, $xr2727; LA64-NEXT: xvmskltz.w $xr0, $xr0728; LA64-NEXT: xvpickve2gr.wu $a0, $xr0, 0729; LA64-NEXT: xvpickve2gr.wu $a1, $xr0, 4730; LA64-NEXT: bstrins.d $a0, $a1, 7, 4731; LA64-NEXT: ret732 %x0 = fcmp ogt <8 x float> %a, %b733 %x1 = fcmp ogt <8 x float> %c, %d734 %y = xor <8 x i1> %x0, %x1735 %res = bitcast <8 x i1> %y to i8736 ret i8 %res737}738 739define i8 @xvmsk_ugt_xor_ueq_and_ogt_v8f32(<8 x float> %a, <8 x float> %b, <8 x float> %c, <8 x float> %d, <8 x float> %e, <8 x float> %f) {740; LA32-LABEL: xvmsk_ugt_xor_ueq_and_ogt_v8f32:741; LA32: # %bb.0:742; LA32-NEXT: xvfcmp.cueq.s $xr2, $xr2, $xr3743; LA32-NEXT: xvfcmp.cult.s $xr0, $xr1, $xr0744; LA32-NEXT: xvxor.v $xr0, $xr0, $xr2745; LA32-NEXT: xvfcmp.clt.s $xr1, $xr5, $xr4746; LA32-NEXT: xvand.v $xr0, $xr0, $xr1747; LA32-NEXT: xvmskltz.w $xr0, $xr0748; LA32-NEXT: xvpickve2gr.wu $a0, $xr0, 0749; LA32-NEXT: xvpickve2gr.wu $a1, $xr0, 4750; LA32-NEXT: bstrins.w $a0, $a1, 7, 4751; LA32-NEXT: ret752;753; LA64-LABEL: xvmsk_ugt_xor_ueq_and_ogt_v8f32:754; LA64: # %bb.0:755; LA64-NEXT: xvfcmp.cueq.s $xr2, $xr2, $xr3756; LA64-NEXT: xvfcmp.cult.s $xr0, $xr1, $xr0757; LA64-NEXT: xvxor.v $xr0, $xr0, $xr2758; LA64-NEXT: xvfcmp.clt.s $xr1, $xr5, $xr4759; LA64-NEXT: xvand.v $xr0, $xr0, $xr1760; LA64-NEXT: xvmskltz.w $xr0, $xr0761; LA64-NEXT: xvpickve2gr.wu $a0, $xr0, 0762; LA64-NEXT: xvpickve2gr.wu $a1, $xr0, 4763; LA64-NEXT: bstrins.d $a0, $a1, 7, 4764; LA64-NEXT: ret765 %x0 = fcmp ugt <8 x float> %a, %b766 %x1 = fcmp ueq <8 x float> %c, %d767 %x2 = fcmp ogt <8 x float> %e, %f768 %y = xor <8 x i1> %x0, %x1769 %z = and <8 x i1> %y, %x2770 %res = bitcast <8 x i1> %z to i8771 ret i8 %res772}773 774define i16 @xvmsk_sgt_and_sgt_v16i16(<16 x i16> %a, <16 x i16> %b, <16 x i16> %c, <16 x i16> %d) {775; LA32-LABEL: xvmsk_sgt_and_sgt_v16i16:776; LA32: # %bb.0:777; LA32-NEXT: xvslt.h $xr2, $xr3, $xr2778; LA32-NEXT: xvslt.h $xr0, $xr1, $xr0779; LA32-NEXT: xvand.v $xr0, $xr0, $xr2780; LA32-NEXT: xvmskltz.h $xr0, $xr0781; LA32-NEXT: xvpickve2gr.wu $a0, $xr0, 0782; LA32-NEXT: xvpickve2gr.wu $a1, $xr0, 4783; LA32-NEXT: bstrins.w $a0, $a1, 15, 8784; LA32-NEXT: ret785;786; LA64-LABEL: xvmsk_sgt_and_sgt_v16i16:787; LA64: # %bb.0:788; LA64-NEXT: xvslt.h $xr2, $xr3, $xr2789; LA64-NEXT: xvslt.h $xr0, $xr1, $xr0790; LA64-NEXT: xvand.v $xr0, $xr0, $xr2791; LA64-NEXT: xvmskltz.h $xr0, $xr0792; LA64-NEXT: xvpickve2gr.wu $a0, $xr0, 0793; LA64-NEXT: xvpickve2gr.wu $a1, $xr0, 4794; LA64-NEXT: bstrins.d $a0, $a1, 15, 8795; LA64-NEXT: ret796 %x0 = icmp sgt <16 x i16> %a, %b797 %x1 = icmp sgt <16 x i16> %c, %d798 %y = and <16 x i1> %x0, %x1799 %res = bitcast <16 x i1> %y to i16800 ret i16 %res801}802 803define i32 @xvmsk_sgt_and_sgt_v32i8(<32 x i8> %a, <32 x i8> %b, <32 x i8> %c, <32 x i8> %d) {804; LA32-LABEL: xvmsk_sgt_and_sgt_v32i8:805; LA32: # %bb.0:806; LA32-NEXT: xvslt.b $xr0, $xr1, $xr0807; LA32-NEXT: xvslt.b $xr1, $xr3, $xr2808; LA32-NEXT: xvand.v $xr0, $xr0, $xr1809; LA32-NEXT: xvmskltz.b $xr0, $xr0810; LA32-NEXT: xvpickve2gr.wu $a0, $xr0, 0811; LA32-NEXT: xvpickve2gr.wu $a1, $xr0, 4812; LA32-NEXT: bstrins.w $a0, $a1, 31, 16813; LA32-NEXT: ret814;815; LA64-LABEL: xvmsk_sgt_and_sgt_v32i8:816; LA64: # %bb.0:817; LA64-NEXT: xvslt.b $xr0, $xr1, $xr0818; LA64-NEXT: xvslt.b $xr1, $xr3, $xr2819; LA64-NEXT: xvand.v $xr0, $xr0, $xr1820; LA64-NEXT: xvmskltz.b $xr0, $xr0821; LA64-NEXT: xvpickve2gr.wu $a0, $xr0, 0822; LA64-NEXT: xvpickve2gr.wu $a1, $xr0, 4823; LA64-NEXT: bstrins.d $a0, $a1, 31, 16824; LA64-NEXT: ret825 %x0 = icmp sgt <32 x i8> %a, %b826 %x1 = icmp sgt <32 x i8> %c, %d827 %y = and <32 x i1> %x0, %x1828 %res = bitcast <32 x i1> %y to i32829 ret i32 %res830}831 832define i8 @xvmsk_eq_v2i64_concat_poison(<2 x i64> %vec) {833; LA32-LABEL: xvmsk_eq_v2i64_concat_poison:834; LA32: # %bb.0:835; LA32-NEXT: vseqi.d $vr0, $vr0, 0836; LA32-NEXT: vpickve2gr.w $a0, $vr0, 0837; LA32-NEXT: vinsgr2vr.h $vr1, $a0, 0838; LA32-NEXT: vpickve2gr.w $a0, $vr0, 2839; LA32-NEXT: vinsgr2vr.h $vr1, $a0, 1840; LA32-NEXT: vslli.h $vr0, $vr1, 15841; LA32-NEXT: vmskltz.h $vr0, $vr0842; LA32-NEXT: vpickve2gr.hu $a0, $vr0, 0843; LA32-NEXT: ret844;845; LA64-LABEL: xvmsk_eq_v2i64_concat_poison:846; LA64: # %bb.0:847; LA64-NEXT: vseqi.d $vr0, $vr0, 0848; LA64-NEXT: vpickve2gr.d $a0, $vr0, 0849; LA64-NEXT: vinsgr2vr.h $vr1, $a0, 0850; LA64-NEXT: vpickve2gr.d $a0, $vr0, 1851; LA64-NEXT: vinsgr2vr.h $vr1, $a0, 1852; LA64-NEXT: vslli.h $vr0, $vr1, 15853; LA64-NEXT: vmskltz.h $vr0, $vr0854; LA64-NEXT: vpickve2gr.hu $a0, $vr0, 0855; LA64-NEXT: ret856 %tobool = icmp eq <2 x i64> %vec, zeroinitializer857 %insertvec = shufflevector <2 x i1> %tobool, <2 x i1> poison, <8 x i32> <i32 0, i32 1, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison>858 %res = bitcast <8 x i1> %insertvec to i8859 ret i8 %res860}861 862define i8 @xvmsk_ne_v4i32_concat_poison(<4 x i32> %vec) {863; CHECK-LABEL: xvmsk_ne_v4i32_concat_poison:864; CHECK: # %bb.0:865; CHECK-NEXT: vseqi.w $vr0, $vr0, 0866; CHECK-NEXT: vrepli.b $vr1, -1867; CHECK-NEXT: vxor.v $vr0, $vr0, $vr1868; CHECK-NEXT: vpickve2gr.w $a0, $vr0, 0869; CHECK-NEXT: vinsgr2vr.h $vr1, $a0, 0870; CHECK-NEXT: vpickve2gr.w $a0, $vr0, 1871; CHECK-NEXT: vinsgr2vr.h $vr1, $a0, 1872; CHECK-NEXT: vpickve2gr.w $a0, $vr0, 2873; CHECK-NEXT: vinsgr2vr.h $vr1, $a0, 2874; CHECK-NEXT: vpickve2gr.w $a0, $vr0, 3875; CHECK-NEXT: vinsgr2vr.h $vr1, $a0, 3876; CHECK-NEXT: vslli.h $vr0, $vr1, 15877; CHECK-NEXT: vmskltz.h $vr0, $vr0878; CHECK-NEXT: vpickve2gr.hu $a0, $vr0, 0879; CHECK-NEXT: ret880 %tobool = icmp ne <4 x i32> %vec, zeroinitializer881 %insertvec = shufflevector <4 x i1> %tobool, <4 x i1> poison, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 poison, i32 poison, i32 poison, i32 poison>882 %res = bitcast <8 x i1> %insertvec to i8883 ret i8 %res884}885 886define i8 @xvmsk_ogt_v4f64_concat_poison(<4 x double> %vec) {887; LA32-LABEL: xvmsk_ogt_v4f64_concat_poison:888; LA32: # %bb.0:889; LA32-NEXT: xvrepli.b $xr1, 0890; LA32-NEXT: xvfcmp.clt.d $xr0, $xr1, $xr0891; LA32-NEXT: xvpickve2gr.w $a0, $xr0, 6892; LA32-NEXT: xvpickve2gr.w $a1, $xr0, 4893; LA32-NEXT: xvpickve2gr.w $a2, $xr0, 2894; LA32-NEXT: xvpickve2gr.w $a3, $xr0, 0895; LA32-NEXT: vinsgr2vr.h $vr0, $a3, 0896; LA32-NEXT: vinsgr2vr.h $vr0, $a2, 1897; LA32-NEXT: vinsgr2vr.h $vr0, $a1, 2898; LA32-NEXT: vinsgr2vr.h $vr0, $a0, 3899; LA32-NEXT: vslli.h $vr0, $vr0, 15900; LA32-NEXT: vmskltz.h $vr0, $vr0901; LA32-NEXT: vpickve2gr.hu $a0, $vr0, 0902; LA32-NEXT: ret903;904; LA64-LABEL: xvmsk_ogt_v4f64_concat_poison:905; LA64: # %bb.0:906; LA64-NEXT: xvrepli.b $xr1, 0907; LA64-NEXT: xvfcmp.clt.d $xr0, $xr1, $xr0908; LA64-NEXT: xvpickve2gr.d $a0, $xr0, 3909; LA64-NEXT: xvpickve2gr.d $a1, $xr0, 2910; LA64-NEXT: xvpickve2gr.d $a2, $xr0, 1911; LA64-NEXT: xvpickve2gr.d $a3, $xr0, 0912; LA64-NEXT: vinsgr2vr.h $vr0, $a3, 0913; LA64-NEXT: vinsgr2vr.h $vr0, $a2, 1914; LA64-NEXT: vinsgr2vr.h $vr0, $a1, 2915; LA64-NEXT: vinsgr2vr.h $vr0, $a0, 3916; LA64-NEXT: vslli.h $vr0, $vr0, 15917; LA64-NEXT: vmskltz.h $vr0, $vr0918; LA64-NEXT: vpickve2gr.hu $a0, $vr0, 0919; LA64-NEXT: ret920 %tobool = fcmp ogt <4 x double> %vec, zeroinitializer921 %insertvec = shufflevector <4 x i1> %tobool, <4 x i1> poison, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 poison, i32 poison, i32 poison, i32 poison>922 %res = bitcast <8 x i1> %insertvec to i8923 ret i8 %res924}925 926define i32 @xvmsk_trunc_i8(<32 x i8> %a) {927; LA32-LABEL: xvmsk_trunc_i8:928; LA32: # %bb.0:929; LA32-NEXT: xvslli.b $xr0, $xr0, 7930; LA32-NEXT: xvmskltz.b $xr0, $xr0931; LA32-NEXT: xvpickve2gr.wu $a0, $xr0, 0932; LA32-NEXT: xvpickve2gr.wu $a1, $xr0, 4933; LA32-NEXT: bstrins.w $a0, $a1, 31, 16934; LA32-NEXT: ret935;936; LA64-LABEL: xvmsk_trunc_i8:937; LA64: # %bb.0:938; LA64-NEXT: xvslli.b $xr0, $xr0, 7939; LA64-NEXT: xvmskltz.b $xr0, $xr0940; LA64-NEXT: xvpickve2gr.wu $a0, $xr0, 0941; LA64-NEXT: xvpickve2gr.wu $a1, $xr0, 4942; LA64-NEXT: bstrins.d $a0, $a1, 31, 16943; LA64-NEXT: ret944 %y = trunc <32 x i8> %a to <32 x i1>945 %res = bitcast <32 x i1> %y to i32946 ret i32 %res947}948 949define i16 @xvmsk_trunc_i16(<16 x i16> %a) {950; LA32-LABEL: xvmsk_trunc_i16:951; LA32: # %bb.0:952; LA32-NEXT: xvslli.h $xr0, $xr0, 15953; LA32-NEXT: xvmskltz.h $xr0, $xr0954; LA32-NEXT: xvpickve2gr.wu $a0, $xr0, 0955; LA32-NEXT: xvpickve2gr.wu $a1, $xr0, 4956; LA32-NEXT: bstrins.w $a0, $a1, 15, 8957; LA32-NEXT: ret958;959; LA64-LABEL: xvmsk_trunc_i16:960; LA64: # %bb.0:961; LA64-NEXT: xvslli.h $xr0, $xr0, 15962; LA64-NEXT: xvmskltz.h $xr0, $xr0963; LA64-NEXT: xvpickve2gr.wu $a0, $xr0, 0964; LA64-NEXT: xvpickve2gr.wu $a1, $xr0, 4965; LA64-NEXT: bstrins.d $a0, $a1, 15, 8966; LA64-NEXT: ret967 %y = trunc <16 x i16> %a to <16 x i1>968 %res = bitcast <16 x i1> %y to i16969 ret i16 %res970}971 972define i8 @xvmsk_trunc_i32(<8 x i32> %a) {973; LA32-LABEL: xvmsk_trunc_i32:974; LA32: # %bb.0:975; LA32-NEXT: xvslli.w $xr0, $xr0, 31976; LA32-NEXT: xvmskltz.w $xr0, $xr0977; LA32-NEXT: xvpickve2gr.wu $a0, $xr0, 0978; LA32-NEXT: xvpickve2gr.wu $a1, $xr0, 4979; LA32-NEXT: bstrins.w $a0, $a1, 7, 4980; LA32-NEXT: ret981;982; LA64-LABEL: xvmsk_trunc_i32:983; LA64: # %bb.0:984; LA64-NEXT: xvslli.w $xr0, $xr0, 31985; LA64-NEXT: xvmskltz.w $xr0, $xr0986; LA64-NEXT: xvpickve2gr.wu $a0, $xr0, 0987; LA64-NEXT: xvpickve2gr.wu $a1, $xr0, 4988; LA64-NEXT: bstrins.d $a0, $a1, 7, 4989; LA64-NEXT: ret990 %y = trunc <8 x i32> %a to <8 x i1>991 %res = bitcast <8 x i1> %y to i8992 ret i8 %res993}994 995define i4 @xvmsk_trunc_i64(<4 x i64> %a) {996; LA32-LABEL: xvmsk_trunc_i64:997; LA32: # %bb.0:998; LA32-NEXT: xvslli.d $xr0, $xr0, 63999; LA32-NEXT: xvmskltz.d $xr0, $xr01000; LA32-NEXT: xvpickve2gr.wu $a0, $xr0, 01001; LA32-NEXT: xvpickve2gr.wu $a1, $xr0, 41002; LA32-NEXT: bstrins.w $a0, $a1, 3, 21003; LA32-NEXT: ret1004;1005; LA64-LABEL: xvmsk_trunc_i64:1006; LA64: # %bb.0:1007; LA64-NEXT: xvslli.d $xr0, $xr0, 631008; LA64-NEXT: xvmskltz.d $xr0, $xr01009; LA64-NEXT: xvpickve2gr.wu $a0, $xr0, 01010; LA64-NEXT: xvpickve2gr.wu $a1, $xr0, 41011; LA64-NEXT: bstrins.d $a0, $a1, 3, 21012; LA64-NEXT: ret1013 %y = trunc <4 x i64> %a to <4 x i1>1014 %res = bitcast <4 x i1> %y to i41015 ret i4 %res1016}1017