1727 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py2; RUN: llc -mtriple=thumbv8.1m.main-none-none-eabi -mattr=+mve.fp -verify-machineinstrs %s -o - | FileCheck %s3 4define void @vaddq(ptr %x, ptr %y, i32 %n) {5; CHECK-LABEL: vaddq:6; CHECK: @ %bb.0: @ %entry7; CHECK-NEXT: .save {r7, lr}8; CHECK-NEXT: push {r7, lr}9; CHECK-NEXT: cmp r2, #110; CHECK-NEXT: it lt11; CHECK-NEXT: poplt {r7, pc}12; CHECK-NEXT: .LBB0_1: @ %for.body.preheader13; CHECK-NEXT: movs r3, #1014; CHECK-NEXT: dlstp.32 lr, r215; CHECK-NEXT: .LBB0_2: @ %for.body16; CHECK-NEXT: @ =>This Inner Loop Header: Depth=117; CHECK-NEXT: vldrw.u32 q0, [r0], #1618; CHECK-NEXT: vadd.i32 q0, q0, r319; CHECK-NEXT: vstrw.32 q0, [r1], #1620; CHECK-NEXT: letp lr, .LBB0_221; CHECK-NEXT: @ %bb.3: @ %for.cond.cleanup22; CHECK-NEXT: pop {r7, pc}23entry:24 %cmp11 = icmp sgt i32 %n, 025 br i1 %cmp11, label %for.body, label %for.cond.cleanup26 27for.cond.cleanup: ; preds = %for.body, %entry28 ret void29 30for.body: ; preds = %entry, %for.body31 %x.addr.014 = phi ptr [ %add.ptr, %for.body ], [ %x, %entry ]32 %y.addr.013 = phi ptr [ %add.ptr1, %for.body ], [ %y, %entry ]33 %i.012 = phi i32 [ %sub, %for.body ], [ %n, %entry ]34 %0 = tail call <4 x i1> @llvm.arm.mve.vctp32(i32 %i.012)35 %1 = bitcast ptr %x.addr.014 to ptr36 %2 = tail call <4 x i32> @llvm.masked.load.v4i32.p0(ptr %1, i32 4, <4 x i1> %0, <4 x i32> zeroinitializer)37 %add.ptr = getelementptr inbounds i32, ptr %x.addr.014, i32 438 %3 = add <4 x i32> %2, <i32 10, i32 10, i32 10, i32 10>39 %4 = bitcast ptr %y.addr.013 to ptr40 tail call void @llvm.masked.store.v4i32.p0(<4 x i32> %3, ptr %4, i32 4, <4 x i1> %0)41 %add.ptr1 = getelementptr inbounds i32, ptr %y.addr.013, i32 442 %sub = add nsw i32 %i.012, -443 %cmp = icmp sgt i32 %i.012, 444 br i1 %cmp, label %for.body, label %for.cond.cleanup45}46 47define void @vadd(ptr %s1, i32 %c0, i32 %N) {48; CHECK-LABEL: vadd:49; CHECK: @ %bb.0: @ %entry50; CHECK-NEXT: .save {r7, lr}51; CHECK-NEXT: push {r7, lr}52; CHECK-NEXT: cmp r2, #153; CHECK-NEXT: it lt54; CHECK-NEXT: poplt {r7, pc}55; CHECK-NEXT: .LBB1_1: @ %while.body.lr.ph56; CHECK-NEXT: dlstp.32 lr, r257; CHECK-NEXT: .LBB1_2: @ %while.body58; CHECK-NEXT: @ =>This Inner Loop Header: Depth=159; CHECK-NEXT: vldrw.u32 q0, [r0]60; CHECK-NEXT: vadd.i32 q0, q0, r161; CHECK-NEXT: vstrw.32 q0, [r0], #1662; CHECK-NEXT: letp lr, .LBB1_263; CHECK-NEXT: @ %bb.3: @ %while.end64; CHECK-NEXT: pop {r7, pc}65entry:66 %cmp11 = icmp sgt i32 %N, 067 br i1 %cmp11, label %while.body.lr.ph, label %while.end68 69while.body.lr.ph: ; preds = %entry70 %.splatinsert = insertelement <4 x i32> undef, i32 %c0, i32 071 %.splat = shufflevector <4 x i32> %.splatinsert, <4 x i32> undef, <4 x i32> zeroinitializer72 br label %while.body73 74while.body: ; preds = %while.body.lr.ph, %while.body75 %s1.addr.013 = phi ptr [ %s1, %while.body.lr.ph ], [ %add.ptr, %while.body ]76 %N.addr.012 = phi i32 [ %N, %while.body.lr.ph ], [ %sub, %while.body ]77 %0 = tail call <4 x i1> @llvm.arm.mve.vctp32(i32 %N.addr.012)78 %1 = bitcast ptr %s1.addr.013 to ptr79 %2 = tail call <4 x i32> @llvm.masked.load.v4i32.p0(ptr %1, i32 4, <4 x i1> %0, <4 x i32> zeroinitializer)80 %3 = tail call <4 x i32> @llvm.arm.mve.add.predicated.v4i32.v4i1(<4 x i32> %2, <4 x i32> %.splat, <4 x i1> %0, <4 x i32> %2)81 tail call void @llvm.masked.store.v4i32.p0(<4 x i32> %3, ptr %1, i32 4, <4 x i1> %0)82 %add.ptr = getelementptr inbounds i32, ptr %s1.addr.013, i32 483 %sub = add nsw i32 %N.addr.012, -484 %cmp = icmp sgt i32 %N.addr.012, 485 br i1 %cmp, label %while.body, label %while.end86 87while.end: ; preds = %while.body, %entry88 ret void89}90 91define void @vsubq(ptr %x, ptr %y, i32 %n) {92; CHECK-LABEL: vsubq:93; CHECK: @ %bb.0: @ %entry94; CHECK-NEXT: .save {r7, lr}95; CHECK-NEXT: push {r7, lr}96; CHECK-NEXT: cmp r2, #197; CHECK-NEXT: it lt98; CHECK-NEXT: poplt {r7, pc}99; CHECK-NEXT: .LBB2_1: @ %for.body.preheader100; CHECK-NEXT: movs r3, #10101; CHECK-NEXT: dlstp.32 lr, r2102; CHECK-NEXT: .LBB2_2: @ %for.body103; CHECK-NEXT: @ =>This Inner Loop Header: Depth=1104; CHECK-NEXT: vldrw.u32 q0, [r0], #16105; CHECK-NEXT: vsub.i32 q0, q0, r3106; CHECK-NEXT: vstrw.32 q0, [r1], #16107; CHECK-NEXT: letp lr, .LBB2_2108; CHECK-NEXT: @ %bb.3: @ %for.cond.cleanup109; CHECK-NEXT: pop {r7, pc}110entry:111 %cmp11 = icmp sgt i32 %n, 0112 br i1 %cmp11, label %for.body, label %for.cond.cleanup113 114for.cond.cleanup: ; preds = %for.body, %entry115 ret void116 117for.body: ; preds = %entry, %for.body118 %x.addr.014 = phi ptr [ %add.ptr, %for.body ], [ %x, %entry ]119 %y.addr.013 = phi ptr [ %add.ptr1, %for.body ], [ %y, %entry ]120 %i.012 = phi i32 [ %sub, %for.body ], [ %n, %entry ]121 %0 = tail call <4 x i1> @llvm.arm.mve.vctp32(i32 %i.012)122 %1 = bitcast ptr %x.addr.014 to ptr123 %2 = tail call <4 x i32> @llvm.masked.load.v4i32.p0(ptr %1, i32 4, <4 x i1> %0, <4 x i32> zeroinitializer)124 %add.ptr = getelementptr inbounds i32, ptr %x.addr.014, i32 4125 %3 = sub <4 x i32> %2, <i32 10, i32 10, i32 10, i32 10>126 %4 = bitcast ptr %y.addr.013 to ptr127 tail call void @llvm.masked.store.v4i32.p0(<4 x i32> %3, ptr %4, i32 4, <4 x i1> %0)128 %add.ptr1 = getelementptr inbounds i32, ptr %y.addr.013, i32 4129 %sub = add nsw i32 %i.012, -4130 %cmp = icmp sgt i32 %i.012, 4131 br i1 %cmp, label %for.body, label %for.cond.cleanup132}133 134define void @vsub(ptr %s1, i32 %N) {135; CHECK-LABEL: vsub:136; CHECK: @ %bb.0: @ %entry137; CHECK-NEXT: .save {r7, lr}138; CHECK-NEXT: push {r7, lr}139; CHECK-NEXT: cmp r1, #1140; CHECK-NEXT: it lt141; CHECK-NEXT: poplt {r7, pc}142; CHECK-NEXT: .LBB3_1: @ %while.body.preheader143; CHECK-NEXT: movs r2, #10144; CHECK-NEXT: dlstp.32 lr, r1145; CHECK-NEXT: .LBB3_2: @ %while.body146; CHECK-NEXT: @ =>This Inner Loop Header: Depth=1147; CHECK-NEXT: vldrw.u32 q0, [r0]148; CHECK-NEXT: vsub.i32 q0, q0, r2149; CHECK-NEXT: vstrw.32 q0, [r0], #16150; CHECK-NEXT: letp lr, .LBB3_2151; CHECK-NEXT: @ %bb.3: @ %while.end152; CHECK-NEXT: pop {r7, pc}153entry:154 %cmp11 = icmp sgt i32 %N, 0155 br i1 %cmp11, label %while.body.lr.ph, label %while.end156 157while.body.lr.ph: ; preds = %entry158 br label %while.body159 160while.body: ; preds = %while.body.lr.ph, %while.body161 %s1.addr.013 = phi ptr [ %s1, %while.body.lr.ph ], [ %add.ptr, %while.body ]162 %N.addr.012 = phi i32 [ %N, %while.body.lr.ph ], [ %sub, %while.body ]163 %0 = tail call <4 x i1> @llvm.arm.mve.vctp32(i32 %N.addr.012)164 %1 = bitcast ptr %s1.addr.013 to ptr165 %2 = tail call <4 x i32> @llvm.masked.load.v4i32.p0(ptr %1, i32 4, <4 x i1> %0, <4 x i32> zeroinitializer)166 %3 = tail call <4 x i32> @llvm.arm.mve.sub.predicated.v4i32.v4i1(<4 x i32> %2, <4 x i32> <i32 10, i32 10, i32 10, i32 10>, <4 x i1> %0, <4 x i32> %2)167 tail call void @llvm.masked.store.v4i32.p0(<4 x i32> %3, ptr %1, i32 4, <4 x i1> %0)168 %add.ptr = getelementptr inbounds i32, ptr %s1.addr.013, i32 4169 %sub = add nsw i32 %N.addr.012, -4170 %cmp = icmp sgt i32 %N.addr.012, 4171 br i1 %cmp, label %while.body, label %while.end172 173while.end: ; preds = %while.body, %entry174 ret void175}176 177define void @vmulq(ptr %x, ptr %y, i32 %n) {178; CHECK-LABEL: vmulq:179; CHECK: @ %bb.0: @ %entry180; CHECK-NEXT: .save {r7, lr}181; CHECK-NEXT: push {r7, lr}182; CHECK-NEXT: cmp r2, #1183; CHECK-NEXT: it lt184; CHECK-NEXT: poplt {r7, pc}185; CHECK-NEXT: .LBB4_1: @ %for.body.preheader186; CHECK-NEXT: movs r3, #10187; CHECK-NEXT: dlstp.32 lr, r2188; CHECK-NEXT: .LBB4_2: @ %for.body189; CHECK-NEXT: @ =>This Inner Loop Header: Depth=1190; CHECK-NEXT: vldrw.u32 q0, [r0], #16191; CHECK-NEXT: vmul.i32 q0, q0, r3192; CHECK-NEXT: vstrw.32 q0, [r1], #16193; CHECK-NEXT: letp lr, .LBB4_2194; CHECK-NEXT: @ %bb.3: @ %for.cond.cleanup195; CHECK-NEXT: pop {r7, pc}196entry:197 %cmp11 = icmp sgt i32 %n, 0198 br i1 %cmp11, label %for.body, label %for.cond.cleanup199 200for.cond.cleanup: ; preds = %for.body, %entry201 ret void202 203for.body: ; preds = %entry, %for.body204 %x.addr.014 = phi ptr [ %add.ptr, %for.body ], [ %x, %entry ]205 %y.addr.013 = phi ptr [ %add.ptr1, %for.body ], [ %y, %entry ]206 %i.012 = phi i32 [ %sub, %for.body ], [ %n, %entry ]207 %0 = tail call <4 x i1> @llvm.arm.mve.vctp32(i32 %i.012)208 %1 = bitcast ptr %x.addr.014 to ptr209 %2 = tail call <4 x i32> @llvm.masked.load.v4i32.p0(ptr %1, i32 4, <4 x i1> %0, <4 x i32> zeroinitializer)210 %add.ptr = getelementptr inbounds i32, ptr %x.addr.014, i32 4211 %3 = mul <4 x i32> %2, <i32 10, i32 10, i32 10, i32 10>212 %4 = bitcast ptr %y.addr.013 to ptr213 tail call void @llvm.masked.store.v4i32.p0(<4 x i32> %3, ptr %4, i32 4, <4 x i1> %0)214 %add.ptr1 = getelementptr inbounds i32, ptr %y.addr.013, i32 4215 %sub = add nsw i32 %i.012, -4216 %cmp = icmp sgt i32 %i.012, 4217 br i1 %cmp, label %for.body, label %for.cond.cleanup218}219 220define void @vmul(ptr %s1, i32 %N) {221; CHECK-LABEL: vmul:222; CHECK: @ %bb.0: @ %entry223; CHECK-NEXT: .save {r7, lr}224; CHECK-NEXT: push {r7, lr}225; CHECK-NEXT: cmp r1, #1226; CHECK-NEXT: it lt227; CHECK-NEXT: poplt {r7, pc}228; CHECK-NEXT: .LBB5_1: @ %while.body.preheader229; CHECK-NEXT: movs r2, #10230; CHECK-NEXT: dlstp.32 lr, r1231; CHECK-NEXT: .LBB5_2: @ %while.body232; CHECK-NEXT: @ =>This Inner Loop Header: Depth=1233; CHECK-NEXT: vldrw.u32 q0, [r0]234; CHECK-NEXT: vmul.i32 q0, q0, r2235; CHECK-NEXT: vstrw.32 q0, [r0], #16236; CHECK-NEXT: letp lr, .LBB5_2237; CHECK-NEXT: @ %bb.3: @ %while.end238; CHECK-NEXT: pop {r7, pc}239entry:240 %cmp11 = icmp sgt i32 %N, 0241 br i1 %cmp11, label %while.body.lr.ph, label %while.end242 243while.body.lr.ph: ; preds = %entry244 br label %while.body245 246while.body: ; preds = %while.body.lr.ph, %while.body247 %s1.addr.013 = phi ptr [ %s1, %while.body.lr.ph ], [ %add.ptr, %while.body ]248 %N.addr.012 = phi i32 [ %N, %while.body.lr.ph ], [ %sub, %while.body ]249 %0 = tail call <4 x i1> @llvm.arm.mve.vctp32(i32 %N.addr.012)250 %1 = bitcast ptr %s1.addr.013 to ptr251 %2 = tail call <4 x i32> @llvm.masked.load.v4i32.p0(ptr %1, i32 4, <4 x i1> %0, <4 x i32> zeroinitializer)252 %3 = tail call <4 x i32> @llvm.arm.mve.mul.predicated.v4i32.v4i1(<4 x i32> %2, <4 x i32> <i32 10, i32 10, i32 10, i32 10>, <4 x i1> %0, <4 x i32> %2)253 tail call void @llvm.masked.store.v4i32.p0(<4 x i32> %3, ptr %1, i32 4, <4 x i1> %0)254 %add.ptr = getelementptr inbounds i32, ptr %s1.addr.013, i32 4255 %sub = add nsw i32 %N.addr.012, -4256 %cmp = icmp sgt i32 %N.addr.012, 4257 br i1 %cmp, label %while.body, label %while.end258 259while.end: ; preds = %while.body, %entry260 ret void261}262 263define void @vqaddq(ptr %x, ptr %y, i32 %n) {264; CHECK-LABEL: vqaddq:265; CHECK: @ %bb.0: @ %entry266; CHECK-NEXT: .save {r7, lr}267; CHECK-NEXT: push {r7, lr}268; CHECK-NEXT: cmp r2, #1269; CHECK-NEXT: it lt270; CHECK-NEXT: poplt {r7, pc}271; CHECK-NEXT: .LBB6_1: @ %for.body.preheader272; CHECK-NEXT: movs r3, #10273; CHECK-NEXT: dlstp.32 lr, r2274; CHECK-NEXT: .LBB6_2: @ %for.body275; CHECK-NEXT: @ =>This Inner Loop Header: Depth=1276; CHECK-NEXT: vldrw.u32 q0, [r0], #16277; CHECK-NEXT: vqadd.s32 q0, q0, r3278; CHECK-NEXT: vstrw.32 q0, [r1], #16279; CHECK-NEXT: letp lr, .LBB6_2280; CHECK-NEXT: @ %bb.3: @ %for.cond.cleanup281; CHECK-NEXT: pop {r7, pc}282entry:283 %cmp11 = icmp sgt i32 %n, 0284 br i1 %cmp11, label %for.body, label %for.cond.cleanup285 286for.cond.cleanup: ; preds = %for.body, %entry287 ret void288 289for.body: ; preds = %entry, %for.body290 %x.addr.014 = phi ptr [ %add.ptr, %for.body ], [ %x, %entry ]291 %y.addr.013 = phi ptr [ %add.ptr1, %for.body ], [ %y, %entry ]292 %i.012 = phi i32 [ %sub, %for.body ], [ %n, %entry ]293 %0 = tail call <4 x i1> @llvm.arm.mve.vctp32(i32 %i.012)294 %1 = bitcast ptr %x.addr.014 to ptr295 %2 = tail call <4 x i32> @llvm.masked.load.v4i32.p0(ptr %1, i32 4, <4 x i1> %0, <4 x i32> zeroinitializer)296 %add.ptr = getelementptr inbounds i32, ptr %x.addr.014, i32 4297 %3 = tail call <4 x i32> @llvm.sadd.sat.v4i32(<4 x i32> %2, <4 x i32> <i32 10, i32 10, i32 10, i32 10>)298 %4 = bitcast ptr %y.addr.013 to ptr299 tail call void @llvm.masked.store.v4i32.p0(<4 x i32> %3, ptr %4, i32 4, <4 x i1> %0)300 %add.ptr1 = getelementptr inbounds i32, ptr %y.addr.013, i32 4301 %sub = add nsw i32 %i.012, -4302 %cmp = icmp sgt i32 %i.012, 4303 br i1 %cmp, label %for.body, label %for.cond.cleanup304}305 306define void @vqaddqu(ptr %x, ptr %y, i32 %n) {307; CHECK-LABEL: vqaddqu:308; CHECK: @ %bb.0: @ %entry309; CHECK-NEXT: .save {r7, lr}310; CHECK-NEXT: push {r7, lr}311; CHECK-NEXT: cmp r2, #1312; CHECK-NEXT: it lt313; CHECK-NEXT: poplt {r7, pc}314; CHECK-NEXT: .LBB7_1: @ %for.body.preheader315; CHECK-NEXT: movs r3, #10316; CHECK-NEXT: dlstp.32 lr, r2317; CHECK-NEXT: .LBB7_2: @ %for.body318; CHECK-NEXT: @ =>This Inner Loop Header: Depth=1319; CHECK-NEXT: vldrw.u32 q0, [r0], #16320; CHECK-NEXT: vqadd.u32 q0, q0, r3321; CHECK-NEXT: vstrw.32 q0, [r1], #16322; CHECK-NEXT: letp lr, .LBB7_2323; CHECK-NEXT: @ %bb.3: @ %for.cond.cleanup324; CHECK-NEXT: pop {r7, pc}325entry:326 %cmp11 = icmp sgt i32 %n, 0327 br i1 %cmp11, label %for.body, label %for.cond.cleanup328 329for.cond.cleanup: ; preds = %for.body, %entry330 ret void331 332for.body: ; preds = %entry, %for.body333 %x.addr.014 = phi ptr [ %add.ptr, %for.body ], [ %x, %entry ]334 %y.addr.013 = phi ptr [ %add.ptr1, %for.body ], [ %y, %entry ]335 %i.012 = phi i32 [ %sub, %for.body ], [ %n, %entry ]336 %0 = tail call <4 x i1> @llvm.arm.mve.vctp32(i32 %i.012)337 %1 = bitcast ptr %x.addr.014 to ptr338 %2 = tail call <4 x i32> @llvm.masked.load.v4i32.p0(ptr %1, i32 4, <4 x i1> %0, <4 x i32> zeroinitializer)339 %add.ptr = getelementptr inbounds i32, ptr %x.addr.014, i32 4340 %3 = tail call <4 x i32> @llvm.uadd.sat.v4i32(<4 x i32> %2, <4 x i32> <i32 10, i32 10, i32 10, i32 10>)341 %4 = bitcast ptr %y.addr.013 to ptr342 tail call void @llvm.masked.store.v4i32.p0(<4 x i32> %3, ptr %4, i32 4, <4 x i1> %0)343 %add.ptr1 = getelementptr inbounds i32, ptr %y.addr.013, i32 4344 %sub = add nsw i32 %i.012, -4345 %cmp = icmp sgt i32 %i.012, 4346 br i1 %cmp, label %for.body, label %for.cond.cleanup347}348 349define void @vqadd(ptr %s1, i32 %N) {350; CHECK-LABEL: vqadd:351; CHECK: @ %bb.0: @ %entry352; CHECK-NEXT: .save {r7, lr}353; CHECK-NEXT: push {r7, lr}354; CHECK-NEXT: cmp r1, #1355; CHECK-NEXT: it lt356; CHECK-NEXT: poplt {r7, pc}357; CHECK-NEXT: .LBB8_1: @ %while.body.preheader358; CHECK-NEXT: movs r2, #10359; CHECK-NEXT: dlstp.32 lr, r1360; CHECK-NEXT: .LBB8_2: @ %while.body361; CHECK-NEXT: @ =>This Inner Loop Header: Depth=1362; CHECK-NEXT: vldrw.u32 q0, [r0]363; CHECK-NEXT: vqadd.s32 q0, q0, r2364; CHECK-NEXT: vstrw.32 q0, [r0], #16365; CHECK-NEXT: letp lr, .LBB8_2366; CHECK-NEXT: @ %bb.3: @ %while.end367; CHECK-NEXT: pop {r7, pc}368entry:369 %cmp11 = icmp sgt i32 %N, 0370 br i1 %cmp11, label %while.body.lr.ph, label %while.end371 372while.body.lr.ph: ; preds = %entry373 br label %while.body374 375while.body: ; preds = %while.body.lr.ph, %while.body376 %s1.addr.013 = phi ptr [ %s1, %while.body.lr.ph ], [ %add.ptr, %while.body ]377 %N.addr.012 = phi i32 [ %N, %while.body.lr.ph ], [ %sub, %while.body ]378 %0 = tail call <4 x i1> @llvm.arm.mve.vctp32(i32 %N.addr.012)379 %1 = bitcast ptr %s1.addr.013 to ptr380 %2 = tail call <4 x i32> @llvm.masked.load.v4i32.p0(ptr %1, i32 4, <4 x i1> %0, <4 x i32> zeroinitializer)381 %3 = tail call <4 x i32> @llvm.arm.mve.qadd.predicated.v4i32.v4i1(<4 x i32> %2, <4 x i32> <i32 10, i32 10, i32 10, i32 10>, i32 0, <4 x i1> %0, <4 x i32> %2)382 tail call void @llvm.masked.store.v4i32.p0(<4 x i32> %3, ptr %1, i32 4, <4 x i1> %0)383 %add.ptr = getelementptr inbounds i32, ptr %s1.addr.013, i32 4384 %sub = add nsw i32 %N.addr.012, -4385 %cmp = icmp sgt i32 %N.addr.012, 4386 br i1 %cmp, label %while.body, label %while.end387 388while.end: ; preds = %while.body, %entry389 ret void390}391 392define void @vqsubq(ptr %x, ptr %y, i32 %n) {393; CHECK-LABEL: vqsubq:394; CHECK: @ %bb.0: @ %entry395; CHECK-NEXT: .save {r7, lr}396; CHECK-NEXT: push {r7, lr}397; CHECK-NEXT: cmp r2, #1398; CHECK-NEXT: it lt399; CHECK-NEXT: poplt {r7, pc}400; CHECK-NEXT: .LBB9_1: @ %for.body.preheader401; CHECK-NEXT: movs r3, #10402; CHECK-NEXT: dlstp.32 lr, r2403; CHECK-NEXT: .LBB9_2: @ %for.body404; CHECK-NEXT: @ =>This Inner Loop Header: Depth=1405; CHECK-NEXT: vldrw.u32 q0, [r0], #16406; CHECK-NEXT: vqsub.s32 q0, q0, r3407; CHECK-NEXT: vstrw.32 q0, [r1], #16408; CHECK-NEXT: letp lr, .LBB9_2409; CHECK-NEXT: @ %bb.3: @ %for.cond.cleanup410; CHECK-NEXT: pop {r7, pc}411entry:412 %cmp11 = icmp sgt i32 %n, 0413 br i1 %cmp11, label %for.body, label %for.cond.cleanup414 415for.cond.cleanup: ; preds = %for.body, %entry416 ret void417 418for.body: ; preds = %entry, %for.body419 %x.addr.014 = phi ptr [ %add.ptr, %for.body ], [ %x, %entry ]420 %y.addr.013 = phi ptr [ %add.ptr1, %for.body ], [ %y, %entry ]421 %i.012 = phi i32 [ %sub, %for.body ], [ %n, %entry ]422 %0 = tail call <4 x i1> @llvm.arm.mve.vctp32(i32 %i.012)423 %1 = bitcast ptr %x.addr.014 to ptr424 %2 = tail call <4 x i32> @llvm.masked.load.v4i32.p0(ptr %1, i32 4, <4 x i1> %0, <4 x i32> zeroinitializer)425 %add.ptr = getelementptr inbounds i32, ptr %x.addr.014, i32 4426 %3 = tail call <4 x i32> @llvm.ssub.sat.v4i32(<4 x i32> %2, <4 x i32> <i32 10, i32 10, i32 10, i32 10>)427 %4 = bitcast ptr %y.addr.013 to ptr428 tail call void @llvm.masked.store.v4i32.p0(<4 x i32> %3, ptr %4, i32 4, <4 x i1> %0)429 %add.ptr1 = getelementptr inbounds i32, ptr %y.addr.013, i32 4430 %sub = add nsw i32 %i.012, -4431 %cmp = icmp sgt i32 %i.012, 4432 br i1 %cmp, label %for.body, label %for.cond.cleanup433}434 435define void @vqsubqu(ptr %x, ptr %y, i32 %n) {436; CHECK-LABEL: vqsubqu:437; CHECK: @ %bb.0: @ %entry438; CHECK-NEXT: .save {r7, lr}439; CHECK-NEXT: push {r7, lr}440; CHECK-NEXT: cmp r2, #1441; CHECK-NEXT: it lt442; CHECK-NEXT: poplt {r7, pc}443; CHECK-NEXT: .LBB10_1: @ %for.body.preheader444; CHECK-NEXT: movs r3, #10445; CHECK-NEXT: dlstp.32 lr, r2446; CHECK-NEXT: .LBB10_2: @ %for.body447; CHECK-NEXT: @ =>This Inner Loop Header: Depth=1448; CHECK-NEXT: vldrw.u32 q0, [r0], #16449; CHECK-NEXT: vqsub.u32 q0, q0, r3450; CHECK-NEXT: vstrw.32 q0, [r1], #16451; CHECK-NEXT: letp lr, .LBB10_2452; CHECK-NEXT: @ %bb.3: @ %for.cond.cleanup453; CHECK-NEXT: pop {r7, pc}454entry:455 %cmp11 = icmp sgt i32 %n, 0456 br i1 %cmp11, label %for.body, label %for.cond.cleanup457 458for.cond.cleanup: ; preds = %for.body, %entry459 ret void460 461for.body: ; preds = %entry, %for.body462 %x.addr.014 = phi ptr [ %add.ptr, %for.body ], [ %x, %entry ]463 %y.addr.013 = phi ptr [ %add.ptr1, %for.body ], [ %y, %entry ]464 %i.012 = phi i32 [ %sub, %for.body ], [ %n, %entry ]465 %0 = tail call <4 x i1> @llvm.arm.mve.vctp32(i32 %i.012)466 %1 = bitcast ptr %x.addr.014 to ptr467 %2 = tail call <4 x i32> @llvm.masked.load.v4i32.p0(ptr %1, i32 4, <4 x i1> %0, <4 x i32> zeroinitializer)468 %add.ptr = getelementptr inbounds i32, ptr %x.addr.014, i32 4469 %3 = tail call <4 x i32> @llvm.usub.sat.v4i32(<4 x i32> %2, <4 x i32> <i32 10, i32 10, i32 10, i32 10>)470 %4 = bitcast ptr %y.addr.013 to ptr471 tail call void @llvm.masked.store.v4i32.p0(<4 x i32> %3, ptr %4, i32 4, <4 x i1> %0)472 %add.ptr1 = getelementptr inbounds i32, ptr %y.addr.013, i32 4473 %sub = add nsw i32 %i.012, -4474 %cmp = icmp sgt i32 %i.012, 4475 br i1 %cmp, label %for.body, label %for.cond.cleanup476}477 478define void @vqsub(ptr %s1, i32 %N) {479; CHECK-LABEL: vqsub:480; CHECK: @ %bb.0: @ %entry481; CHECK-NEXT: .save {r7, lr}482; CHECK-NEXT: push {r7, lr}483; CHECK-NEXT: cmp r1, #1484; CHECK-NEXT: it lt485; CHECK-NEXT: poplt {r7, pc}486; CHECK-NEXT: .LBB11_1: @ %while.body.preheader487; CHECK-NEXT: movs r2, #10488; CHECK-NEXT: dlstp.32 lr, r1489; CHECK-NEXT: .LBB11_2: @ %while.body490; CHECK-NEXT: @ =>This Inner Loop Header: Depth=1491; CHECK-NEXT: vldrw.u32 q0, [r0]492; CHECK-NEXT: vqsub.s32 q0, q0, r2493; CHECK-NEXT: vstrw.32 q0, [r0], #16494; CHECK-NEXT: letp lr, .LBB11_2495; CHECK-NEXT: @ %bb.3: @ %while.end496; CHECK-NEXT: pop {r7, pc}497entry:498 %cmp11 = icmp sgt i32 %N, 0499 br i1 %cmp11, label %while.body.lr.ph, label %while.end500 501while.body.lr.ph: ; preds = %entry502 br label %while.body503 504while.body: ; preds = %while.body.lr.ph, %while.body505 %s1.addr.013 = phi ptr [ %s1, %while.body.lr.ph ], [ %add.ptr, %while.body ]506 %N.addr.012 = phi i32 [ %N, %while.body.lr.ph ], [ %sub, %while.body ]507 %0 = tail call <4 x i1> @llvm.arm.mve.vctp32(i32 %N.addr.012)508 %1 = bitcast ptr %s1.addr.013 to ptr509 %2 = tail call <4 x i32> @llvm.masked.load.v4i32.p0(ptr %1, i32 4, <4 x i1> %0, <4 x i32> zeroinitializer)510 %3 = tail call <4 x i32> @llvm.arm.mve.qsub.predicated.v4i32.v4i1(<4 x i32> %2, <4 x i32> <i32 10, i32 10, i32 10, i32 10>, i32 0, <4 x i1> %0, <4 x i32> %2)511 tail call void @llvm.masked.store.v4i32.p0(<4 x i32> %3, ptr %1, i32 4, <4 x i1> %0)512 %add.ptr = getelementptr inbounds i32, ptr %s1.addr.013, i32 4513 %sub = add nsw i32 %N.addr.012, -4514 %cmp = icmp sgt i32 %N.addr.012, 4515 br i1 %cmp, label %while.body, label %while.end516 517while.end: ; preds = %while.body, %entry518 ret void519}520 521define void @vhaddq(ptr %x, ptr %y, i32 %n) {522; CHECK-LABEL: vhaddq:523; CHECK: @ %bb.0: @ %entry524; CHECK-NEXT: .save {r7, lr}525; CHECK-NEXT: push {r7, lr}526; CHECK-NEXT: cmp r2, #1527; CHECK-NEXT: it lt528; CHECK-NEXT: poplt {r7, pc}529; CHECK-NEXT: .LBB12_1: @ %for.body.preheader530; CHECK-NEXT: movs r3, #10531; CHECK-NEXT: dlstp.32 lr, r2532; CHECK-NEXT: .LBB12_2: @ %for.body533; CHECK-NEXT: @ =>This Inner Loop Header: Depth=1534; CHECK-NEXT: vldrw.u32 q0, [r0], #16535; CHECK-NEXT: vhadd.s32 q0, q0, r3536; CHECK-NEXT: vstrw.32 q0, [r1], #16537; CHECK-NEXT: letp lr, .LBB12_2538; CHECK-NEXT: @ %bb.3: @ %for.cond.cleanup539; CHECK-NEXT: pop {r7, pc}540entry:541 %cmp11 = icmp sgt i32 %n, 0542 br i1 %cmp11, label %for.body, label %for.cond.cleanup543 544for.cond.cleanup: ; preds = %for.body, %entry545 ret void546 547for.body: ; preds = %entry, %for.body548 %x.addr.014 = phi ptr [ %add.ptr, %for.body ], [ %x, %entry ]549 %y.addr.013 = phi ptr [ %add.ptr1, %for.body ], [ %y, %entry ]550 %i.012 = phi i32 [ %sub, %for.body ], [ %n, %entry ]551 %0 = tail call <4 x i1> @llvm.arm.mve.vctp32(i32 %i.012)552 %1 = bitcast ptr %x.addr.014 to ptr553 %2 = tail call <4 x i32> @llvm.masked.load.v4i32.p0(ptr %1, i32 4, <4 x i1> %0, <4 x i32> zeroinitializer)554 %add.ptr = getelementptr inbounds i32, ptr %x.addr.014, i32 4555 %3 = tail call <4 x i32> @llvm.arm.mve.vhadd.v4i32(<4 x i32> %2, <4 x i32> <i32 10, i32 10, i32 10, i32 10>, i32 0)556 %4 = bitcast ptr %y.addr.013 to ptr557 tail call void @llvm.masked.store.v4i32.p0(<4 x i32> %3, ptr %4, i32 4, <4 x i1> %0)558 %add.ptr1 = getelementptr inbounds i32, ptr %y.addr.013, i32 4559 %sub = add nsw i32 %i.012, -4560 %cmp = icmp sgt i32 %i.012, 4561 br i1 %cmp, label %for.body, label %for.cond.cleanup562}563 564define void @vhadd(ptr %s1, i32 %N) {565; CHECK-LABEL: vhadd:566; CHECK: @ %bb.0: @ %entry567; CHECK-NEXT: .save {r7, lr}568; CHECK-NEXT: push {r7, lr}569; CHECK-NEXT: cmp r1, #1570; CHECK-NEXT: it lt571; CHECK-NEXT: poplt {r7, pc}572; CHECK-NEXT: .LBB13_1: @ %while.body.preheader573; CHECK-NEXT: movs r2, #10574; CHECK-NEXT: dlstp.32 lr, r1575; CHECK-NEXT: .LBB13_2: @ %while.body576; CHECK-NEXT: @ =>This Inner Loop Header: Depth=1577; CHECK-NEXT: vldrw.u32 q0, [r0]578; CHECK-NEXT: vhadd.s32 q0, q0, r2579; CHECK-NEXT: vstrw.32 q0, [r0], #16580; CHECK-NEXT: letp lr, .LBB13_2581; CHECK-NEXT: @ %bb.3: @ %while.end582; CHECK-NEXT: pop {r7, pc}583entry:584 %cmp11 = icmp sgt i32 %N, 0585 br i1 %cmp11, label %while.body.lr.ph, label %while.end586 587while.body.lr.ph: ; preds = %entry588 br label %while.body589 590while.body: ; preds = %while.body.lr.ph, %while.body591 %s1.addr.013 = phi ptr [ %s1, %while.body.lr.ph ], [ %add.ptr, %while.body ]592 %N.addr.012 = phi i32 [ %N, %while.body.lr.ph ], [ %sub, %while.body ]593 %0 = tail call <4 x i1> @llvm.arm.mve.vctp32(i32 %N.addr.012)594 %1 = bitcast ptr %s1.addr.013 to ptr595 %2 = tail call <4 x i32> @llvm.masked.load.v4i32.p0(ptr %1, i32 4, <4 x i1> %0, <4 x i32> zeroinitializer)596 %3 = tail call <4 x i32> @llvm.arm.mve.hadd.predicated.v4i32.v4i1(<4 x i32> %2, <4 x i32> <i32 10, i32 10, i32 10, i32 10>, i32 0, <4 x i1> %0, <4 x i32> %2)597 tail call void @llvm.masked.store.v4i32.p0(<4 x i32> %3, ptr %1, i32 4, <4 x i1> %0)598 %add.ptr = getelementptr inbounds i32, ptr %s1.addr.013, i32 4599 %sub = add nsw i32 %N.addr.012, -4600 %cmp = icmp sgt i32 %N.addr.012, 4601 br i1 %cmp, label %while.body, label %while.end602 603while.end: ; preds = %while.body, %entry604 ret void605}606 607define void @vhsubq(ptr %x, ptr %y, i32 %n) {608; CHECK-LABEL: vhsubq:609; CHECK: @ %bb.0: @ %entry610; CHECK-NEXT: .save {r7, lr}611; CHECK-NEXT: push {r7, lr}612; CHECK-NEXT: cmp r2, #1613; CHECK-NEXT: it lt614; CHECK-NEXT: poplt {r7, pc}615; CHECK-NEXT: .LBB14_1: @ %for.body.preheader616; CHECK-NEXT: movs r3, #10617; CHECK-NEXT: dlstp.32 lr, r2618; CHECK-NEXT: .LBB14_2: @ %for.body619; CHECK-NEXT: @ =>This Inner Loop Header: Depth=1620; CHECK-NEXT: vldrw.u32 q0, [r0], #16621; CHECK-NEXT: vhsub.s32 q0, q0, r3622; CHECK-NEXT: vstrw.32 q0, [r1], #16623; CHECK-NEXT: letp lr, .LBB14_2624; CHECK-NEXT: @ %bb.3: @ %for.cond.cleanup625; CHECK-NEXT: pop {r7, pc}626entry:627 %cmp11 = icmp sgt i32 %n, 0628 br i1 %cmp11, label %for.body, label %for.cond.cleanup629 630for.cond.cleanup: ; preds = %for.body, %entry631 ret void632 633for.body: ; preds = %entry, %for.body634 %x.addr.014 = phi ptr [ %add.ptr, %for.body ], [ %x, %entry ]635 %y.addr.013 = phi ptr [ %add.ptr1, %for.body ], [ %y, %entry ]636 %i.012 = phi i32 [ %sub, %for.body ], [ %n, %entry ]637 %0 = tail call <4 x i1> @llvm.arm.mve.vctp32(i32 %i.012)638 %1 = bitcast ptr %x.addr.014 to ptr639 %2 = tail call <4 x i32> @llvm.masked.load.v4i32.p0(ptr %1, i32 4, <4 x i1> %0, <4 x i32> zeroinitializer)640 %add.ptr = getelementptr inbounds i32, ptr %x.addr.014, i32 4641 %3 = tail call <4 x i32> @llvm.arm.mve.vhsub.v4i32(<4 x i32> %2, <4 x i32> <i32 10, i32 10, i32 10, i32 10>, i32 0)642 %4 = bitcast ptr %y.addr.013 to ptr643 tail call void @llvm.masked.store.v4i32.p0(<4 x i32> %3, ptr %4, i32 4, <4 x i1> %0)644 %add.ptr1 = getelementptr inbounds i32, ptr %y.addr.013, i32 4645 %sub = add nsw i32 %i.012, -4646 %cmp = icmp sgt i32 %i.012, 4647 br i1 %cmp, label %for.body, label %for.cond.cleanup648}649 650define void @vhsub(ptr %s1, i32 %N) {651; CHECK-LABEL: vhsub:652; CHECK: @ %bb.0: @ %entry653; CHECK-NEXT: .save {r7, lr}654; CHECK-NEXT: push {r7, lr}655; CHECK-NEXT: cmp r1, #1656; CHECK-NEXT: it lt657; CHECK-NEXT: poplt {r7, pc}658; CHECK-NEXT: .LBB15_1: @ %while.body.preheader659; CHECK-NEXT: movs r2, #10660; CHECK-NEXT: dlstp.32 lr, r1661; CHECK-NEXT: .LBB15_2: @ %while.body662; CHECK-NEXT: @ =>This Inner Loop Header: Depth=1663; CHECK-NEXT: vldrw.u32 q0, [r0]664; CHECK-NEXT: vhsub.s32 q0, q0, r2665; CHECK-NEXT: vstrw.32 q0, [r0], #16666; CHECK-NEXT: letp lr, .LBB15_2667; CHECK-NEXT: @ %bb.3: @ %while.end668; CHECK-NEXT: pop {r7, pc}669entry:670 %cmp11 = icmp sgt i32 %N, 0671 br i1 %cmp11, label %while.body.lr.ph, label %while.end672 673while.body.lr.ph: ; preds = %entry674 br label %while.body675 676while.body: ; preds = %while.body.lr.ph, %while.body677 %s1.addr.013 = phi ptr [ %s1, %while.body.lr.ph ], [ %add.ptr, %while.body ]678 %N.addr.012 = phi i32 [ %N, %while.body.lr.ph ], [ %sub, %while.body ]679 %0 = tail call <4 x i1> @llvm.arm.mve.vctp32(i32 %N.addr.012)680 %1 = bitcast ptr %s1.addr.013 to ptr681 %2 = tail call <4 x i32> @llvm.masked.load.v4i32.p0(ptr %1, i32 4, <4 x i1> %0, <4 x i32> zeroinitializer)682 %3 = tail call <4 x i32> @llvm.arm.mve.hsub.predicated.v4i32.v4i1(<4 x i32> %2, <4 x i32> <i32 10, i32 10, i32 10, i32 10>, i32 0, <4 x i1> %0, <4 x i32> %2)683 tail call void @llvm.masked.store.v4i32.p0(<4 x i32> %3, ptr %1, i32 4, <4 x i1> %0)684 %add.ptr = getelementptr inbounds i32, ptr %s1.addr.013, i32 4685 %sub = add nsw i32 %N.addr.012, -4686 %cmp = icmp sgt i32 %N.addr.012, 4687 br i1 %cmp, label %while.body, label %while.end688 689while.end: ; preds = %while.body, %entry690 ret void691}692 693define void @vqdmullbq(ptr %x, ptr %y, i32 %n) {694; CHECK-LABEL: vqdmullbq:695; CHECK: @ %bb.0: @ %entry696; CHECK-NEXT: .save {r7, lr}697; CHECK-NEXT: push {r7, lr}698; CHECK-NEXT: cmp r2, #1699; CHECK-NEXT: it lt700; CHECK-NEXT: poplt {r7, pc}701; CHECK-NEXT: .LBB16_1: @ %for.body.preheader702; CHECK-NEXT: movs r3, #10703; CHECK-NEXT: dlstp.32 lr, r2704; CHECK-NEXT: .LBB16_2: @ %for.body705; CHECK-NEXT: @ =>This Inner Loop Header: Depth=1706; CHECK-NEXT: vldrw.u32 q0, [r0], #16707; CHECK-NEXT: vqdmullb.s32 q1, q0, r3708; CHECK-NEXT: vstrw.32 q1, [r1], #16709; CHECK-NEXT: letp lr, .LBB16_2710; CHECK-NEXT: @ %bb.3: @ %for.cond.cleanup711; CHECK-NEXT: pop {r7, pc}712entry:713 %cmp11 = icmp sgt i32 %n, 0714 br i1 %cmp11, label %for.body, label %for.cond.cleanup715 716for.cond.cleanup: ; preds = %for.body, %entry717 ret void718 719for.body: ; preds = %entry, %for.body720 %x.addr.014 = phi ptr [ %add.ptr, %for.body ], [ %x, %entry ]721 %y.addr.013 = phi ptr [ %add.ptr1, %for.body ], [ %y, %entry ]722 %i.012 = phi i32 [ %sub, %for.body ], [ %n, %entry ]723 %0 = tail call <4 x i1> @llvm.arm.mve.vctp32(i32 %i.012)724 %1 = bitcast ptr %x.addr.014 to ptr725 %2 = tail call <4 x i32> @llvm.masked.load.v4i32.p0(ptr %1, i32 4, <4 x i1> %0, <4 x i32> zeroinitializer)726 %add.ptr = getelementptr inbounds i32, ptr %x.addr.014, i32 4727 %3 = tail call <2 x i64> @llvm.arm.mve.vqdmull.v2i64.v4i32(<4 x i32> %2, <4 x i32> <i32 10, i32 10, i32 10, i32 10>, i32 0)728 %4 = bitcast <2 x i64> %3 to <4 x i32>729 %5 = bitcast ptr %y.addr.013 to ptr730 tail call void @llvm.masked.store.v4i32.p0(<4 x i32> %4, ptr %5, i32 4, <4 x i1> %0)731 %add.ptr1 = getelementptr inbounds i32, ptr %y.addr.013, i32 4732 %sub = add nsw i32 %i.012, -4733 %cmp = icmp sgt i32 %i.012, 4734 br i1 %cmp, label %for.body, label %for.cond.cleanup735}736 737 738define void @vqdmull(ptr %s1, i32 %N) {739; CHECK-LABEL: vqdmull:740; CHECK: @ %bb.0: @ %entry741; CHECK-NEXT: .save {r7, lr}742; CHECK-NEXT: push {r7, lr}743; CHECK-NEXT: cmp r1, #1744; CHECK-NEXT: it lt745; CHECK-NEXT: poplt {r7, pc}746; CHECK-NEXT: .LBB17_1: @ %while.body.preheader747; CHECK-NEXT: movs r2, #10748; CHECK-NEXT: dlstp.32 lr, r1749; CHECK-NEXT: .LBB17_2: @ %while.body750; CHECK-NEXT: @ =>This Inner Loop Header: Depth=1751; CHECK-NEXT: vldrh.s32 q0, [r0]752; CHECK-NEXT: vqdmullb.s16 q0, q0, r2753; CHECK-NEXT: vstrw.32 q0, [r0], #16754; CHECK-NEXT: letp lr, .LBB17_2755; CHECK-NEXT: @ %bb.3: @ %while.end756; CHECK-NEXT: pop {r7, pc}757entry:758 %cmp11 = icmp sgt i32 %N, 0759 br i1 %cmp11, label %while.body.lr.ph, label %while.end760 761while.body.lr.ph: ; preds = %entry762 br label %while.body763 764while.body: ; preds = %while.body.lr.ph, %while.body765 %s1.addr.013 = phi ptr [ %s1, %while.body.lr.ph ], [ %add.ptr, %while.body ]766 %N.addr.012 = phi i32 [ %N, %while.body.lr.ph ], [ %sub, %while.body ]767 %0 = tail call <4 x i1> @llvm.arm.mve.vctp32(i32 %N.addr.012)768 %1 = bitcast ptr %s1.addr.013 to ptr769 %2 = tail call <4 x i16> @llvm.masked.load.v4i16.p0(ptr %1, i32 2, <4 x i1> %0, <4 x i16> zeroinitializer)770 %3 = sext <4 x i16> %2 to <4 x i32>771 %4 = bitcast <4 x i32> %3 to <8 x i16>772 %5 = tail call <4 x i32> @llvm.arm.mve.vqdmull.predicated.v4i32.v8i16.v4i1(<8 x i16> %4, <8 x i16> <i16 10, i16 10, i16 10, i16 10, i16 10, i16 10, i16 10, i16 10>, i32 0, <4 x i1> %0, <4 x i32> %3)773 %6 = bitcast ptr %s1.addr.013 to ptr774 tail call void @llvm.masked.store.v4i32.p0(<4 x i32> %5, ptr %6, i32 4, <4 x i1> %0)775 %add.ptr = getelementptr inbounds i32, ptr %s1.addr.013, i32 4776 %sub = add nsw i32 %N.addr.012, -4777 %cmp = icmp sgt i32 %N.addr.012, 4778 br i1 %cmp, label %while.body, label %while.end779 780while.end: ; preds = %while.body, %entry781 ret void782}783 784define void @vqdmulhq(ptr %x, ptr %y, i32 %n) {785; CHECK-LABEL: vqdmulhq:786; CHECK: @ %bb.0: @ %entry787; CHECK-NEXT: .save {r7, lr}788; CHECK-NEXT: push {r7, lr}789; CHECK-NEXT: cmp r2, #1790; CHECK-NEXT: it lt791; CHECK-NEXT: poplt {r7, pc}792; CHECK-NEXT: .LBB18_1: @ %for.body.preheader793; CHECK-NEXT: movs r3, #10794; CHECK-NEXT: dlstp.32 lr, r2795; CHECK-NEXT: .LBB18_2: @ %for.body796; CHECK-NEXT: @ =>This Inner Loop Header: Depth=1797; CHECK-NEXT: vldrw.u32 q0, [r0], #16798; CHECK-NEXT: vqdmulh.s32 q0, q0, r3799; CHECK-NEXT: vstrw.32 q0, [r1], #16800; CHECK-NEXT: letp lr, .LBB18_2801; CHECK-NEXT: @ %bb.3: @ %for.cond.cleanup802; CHECK-NEXT: pop {r7, pc}803entry:804 %cmp11 = icmp sgt i32 %n, 0805 br i1 %cmp11, label %for.body, label %for.cond.cleanup806 807for.cond.cleanup: ; preds = %for.body, %entry808 ret void809 810for.body: ; preds = %entry, %for.body811 %x.addr.014 = phi ptr [ %add.ptr, %for.body ], [ %x, %entry ]812 %y.addr.013 = phi ptr [ %add.ptr1, %for.body ], [ %y, %entry ]813 %i.012 = phi i32 [ %sub, %for.body ], [ %n, %entry ]814 %0 = tail call <4 x i1> @llvm.arm.mve.vctp32(i32 %i.012)815 %1 = bitcast ptr %x.addr.014 to ptr816 %2 = tail call <4 x i32> @llvm.masked.load.v4i32.p0(ptr %1, i32 4, <4 x i1> %0, <4 x i32> zeroinitializer)817 %add.ptr = getelementptr inbounds i32, ptr %x.addr.014, i32 4818 %3 = tail call <4 x i32> @llvm.arm.mve.vqdmulh.v4i32(<4 x i32> %2, <4 x i32> <i32 10, i32 10, i32 10, i32 10>)819 %4 = bitcast ptr %y.addr.013 to ptr820 tail call void @llvm.masked.store.v4i32.p0(<4 x i32> %3, ptr %4, i32 4, <4 x i1> %0)821 %add.ptr1 = getelementptr inbounds i32, ptr %y.addr.013, i32 4822 %sub = add nsw i32 %i.012, -4823 %cmp = icmp sgt i32 %i.012, 4824 br i1 %cmp, label %for.body, label %for.cond.cleanup825}826 827define void @vqdmulh(ptr %s1, i32 %N) {828; CHECK-LABEL: vqdmulh:829; CHECK: @ %bb.0: @ %entry830; CHECK-NEXT: .save {r7, lr}831; CHECK-NEXT: push {r7, lr}832; CHECK-NEXT: cmp r1, #1833; CHECK-NEXT: it lt834; CHECK-NEXT: poplt {r7, pc}835; CHECK-NEXT: .LBB19_1: @ %while.body.preheader836; CHECK-NEXT: movs r2, #10837; CHECK-NEXT: dlstp.32 lr, r1838; CHECK-NEXT: .LBB19_2: @ %while.body839; CHECK-NEXT: @ =>This Inner Loop Header: Depth=1840; CHECK-NEXT: vldrw.u32 q0, [r0]841; CHECK-NEXT: vqdmulh.s32 q0, q0, r2842; CHECK-NEXT: vstrw.32 q0, [r0], #16843; CHECK-NEXT: letp lr, .LBB19_2844; CHECK-NEXT: @ %bb.3: @ %while.end845; CHECK-NEXT: pop {r7, pc}846entry:847 %cmp11 = icmp sgt i32 %N, 0848 br i1 %cmp11, label %while.body.lr.ph, label %while.end849 850while.body.lr.ph: ; preds = %entry851 br label %while.body852 853while.body: ; preds = %while.body.lr.ph, %while.body854 %s1.addr.013 = phi ptr [ %s1, %while.body.lr.ph ], [ %add.ptr, %while.body ]855 %N.addr.012 = phi i32 [ %N, %while.body.lr.ph ], [ %sub, %while.body ]856 %0 = tail call <4 x i1> @llvm.arm.mve.vctp32(i32 %N.addr.012)857 %1 = bitcast ptr %s1.addr.013 to ptr858 %2 = tail call <4 x i32> @llvm.masked.load.v4i32.p0(ptr %1, i32 4, <4 x i1> %0, <4 x i32> zeroinitializer)859 %3 = tail call <4 x i32> @llvm.arm.mve.qdmulh.predicated.v4i32.v4i1(<4 x i32> %2, <4 x i32> <i32 10, i32 10, i32 10, i32 10>, <4 x i1> %0, <4 x i32> %2)860 tail call void @llvm.masked.store.v4i32.p0(<4 x i32> %3, ptr %1, i32 4, <4 x i1> %0)861 %add.ptr = getelementptr inbounds i32, ptr %s1.addr.013, i32 4862 %sub = add nsw i32 %N.addr.012, -4863 %cmp = icmp sgt i32 %N.addr.012, 4864 br i1 %cmp, label %while.body, label %while.end865 866while.end: ; preds = %while.body, %entry867 ret void868}869 870define void @vqrdmulhq(ptr %x, ptr %y, i32 %n) {871; CHECK-LABEL: vqrdmulhq:872; CHECK: @ %bb.0: @ %entry873; CHECK-NEXT: .save {r7, lr}874; CHECK-NEXT: push {r7, lr}875; CHECK-NEXT: cmp r2, #1876; CHECK-NEXT: it lt877; CHECK-NEXT: poplt {r7, pc}878; CHECK-NEXT: .LBB20_1: @ %for.body.preheader879; CHECK-NEXT: movs r3, #10880; CHECK-NEXT: dlstp.32 lr, r2881; CHECK-NEXT: .LBB20_2: @ %for.body882; CHECK-NEXT: @ =>This Inner Loop Header: Depth=1883; CHECK-NEXT: vldrw.u32 q0, [r0], #16884; CHECK-NEXT: vqrdmulh.s32 q0, q0, r3885; CHECK-NEXT: vstrw.32 q0, [r1], #16886; CHECK-NEXT: letp lr, .LBB20_2887; CHECK-NEXT: @ %bb.3: @ %for.cond.cleanup888; CHECK-NEXT: pop {r7, pc}889entry:890 %cmp11 = icmp sgt i32 %n, 0891 br i1 %cmp11, label %for.body, label %for.cond.cleanup892 893for.cond.cleanup: ; preds = %for.body, %entry894 ret void895 896for.body: ; preds = %entry, %for.body897 %x.addr.014 = phi ptr [ %add.ptr, %for.body ], [ %x, %entry ]898 %y.addr.013 = phi ptr [ %add.ptr1, %for.body ], [ %y, %entry ]899 %i.012 = phi i32 [ %sub, %for.body ], [ %n, %entry ]900 %0 = tail call <4 x i1> @llvm.arm.mve.vctp32(i32 %i.012)901 %1 = bitcast ptr %x.addr.014 to ptr902 %2 = tail call <4 x i32> @llvm.masked.load.v4i32.p0(ptr %1, i32 4, <4 x i1> %0, <4 x i32> zeroinitializer)903 %add.ptr = getelementptr inbounds i32, ptr %x.addr.014, i32 4904 %3 = tail call <4 x i32> @llvm.arm.mve.vqrdmulh.v4i32(<4 x i32> %2, <4 x i32> <i32 10, i32 10, i32 10, i32 10>)905 %4 = bitcast ptr %y.addr.013 to ptr906 tail call void @llvm.masked.store.v4i32.p0(<4 x i32> %3, ptr %4, i32 4, <4 x i1> %0)907 %add.ptr1 = getelementptr inbounds i32, ptr %y.addr.013, i32 4908 %sub = add nsw i32 %i.012, -4909 %cmp = icmp sgt i32 %i.012, 4910 br i1 %cmp, label %for.body, label %for.cond.cleanup911}912 913define void @vqrdmulh(ptr %s1, i32 %N) {914; CHECK-LABEL: vqrdmulh:915; CHECK: @ %bb.0: @ %entry916; CHECK-NEXT: .save {r7, lr}917; CHECK-NEXT: push {r7, lr}918; CHECK-NEXT: cmp r1, #1919; CHECK-NEXT: it lt920; CHECK-NEXT: poplt {r7, pc}921; CHECK-NEXT: .LBB21_1: @ %while.body.preheader922; CHECK-NEXT: movs r2, #10923; CHECK-NEXT: dlstp.32 lr, r1924; CHECK-NEXT: .LBB21_2: @ %while.body925; CHECK-NEXT: @ =>This Inner Loop Header: Depth=1926; CHECK-NEXT: vldrw.u32 q0, [r0]927; CHECK-NEXT: vqrdmulh.s32 q0, q0, r2928; CHECK-NEXT: vstrw.32 q0, [r0], #16929; CHECK-NEXT: letp lr, .LBB21_2930; CHECK-NEXT: @ %bb.3: @ %while.end931; CHECK-NEXT: pop {r7, pc}932entry:933 %cmp11 = icmp sgt i32 %N, 0934 br i1 %cmp11, label %while.body.lr.ph, label %while.end935 936while.body.lr.ph: ; preds = %entry937 br label %while.body938 939while.body: ; preds = %while.body.lr.ph, %while.body940 %s1.addr.013 = phi ptr [ %s1, %while.body.lr.ph ], [ %add.ptr, %while.body ]941 %N.addr.012 = phi i32 [ %N, %while.body.lr.ph ], [ %sub, %while.body ]942 %0 = tail call <4 x i1> @llvm.arm.mve.vctp32(i32 %N.addr.012)943 %1 = bitcast ptr %s1.addr.013 to ptr944 %2 = tail call <4 x i32> @llvm.masked.load.v4i32.p0(ptr %1, i32 4, <4 x i1> %0, <4 x i32> zeroinitializer)945 %3 = tail call <4 x i32> @llvm.arm.mve.qrdmulh.predicated.v4i32.v4i1(<4 x i32> %2, <4 x i32> <i32 10, i32 10, i32 10, i32 10>, <4 x i1> %0, <4 x i32> %2)946 tail call void @llvm.masked.store.v4i32.p0(<4 x i32> %3, ptr %1, i32 4, <4 x i1> %0)947 %add.ptr = getelementptr inbounds i32, ptr %s1.addr.013, i32 4948 %sub = add nsw i32 %N.addr.012, -4949 %cmp = icmp sgt i32 %N.addr.012, 4950 br i1 %cmp, label %while.body, label %while.end951 952while.end: ; preds = %while.body, %entry953 ret void954}955 956define void @vmlaq(ptr %x, ptr %y, i32 %n) {957; CHECK-LABEL: vmlaq:958; CHECK: @ %bb.0: @ %entry959; CHECK-NEXT: .save {r7, lr}960; CHECK-NEXT: push {r7, lr}961; CHECK-NEXT: cmp r2, #1962; CHECK-NEXT: it lt963; CHECK-NEXT: poplt {r7, pc}964; CHECK-NEXT: .LBB22_1: @ %for.body.preheader965; CHECK-NEXT: movs r3, #10966; CHECK-NEXT: dlstp.32 lr, r2967; CHECK-NEXT: .LBB22_2: @ %for.body968; CHECK-NEXT: @ =>This Inner Loop Header: Depth=1969; CHECK-NEXT: vldrw.u32 q0, [r1]970; CHECK-NEXT: vldrw.u32 q1, [r0], #16971; CHECK-NEXT: vmla.i32 q1, q0, r3972; CHECK-NEXT: vstrw.32 q1, [r1], #16973; CHECK-NEXT: letp lr, .LBB22_2974; CHECK-NEXT: @ %bb.3: @ %for.cond.cleanup975; CHECK-NEXT: pop {r7, pc}976entry:977 %cmp14 = icmp sgt i32 %n, 0978 br i1 %cmp14, label %for.body, label %for.cond.cleanup979 980for.cond.cleanup: ; preds = %for.body, %entry981 ret void982 983for.body: ; preds = %entry, %for.body984 %x.addr.017 = phi ptr [ %add.ptr, %for.body ], [ %x, %entry ]985 %y.addr.016 = phi ptr [ %add.ptr1, %for.body ], [ %y, %entry ]986 %i.015 = phi i32 [ %sub, %for.body ], [ %n, %entry ]987 %0 = tail call <4 x i1> @llvm.arm.mve.vctp32(i32 %i.015)988 %1 = bitcast ptr %x.addr.017 to ptr989 %2 = tail call <4 x i32> @llvm.masked.load.v4i32.p0(ptr %1, i32 4, <4 x i1> %0, <4 x i32> zeroinitializer)990 %add.ptr = getelementptr inbounds i32, ptr %x.addr.017, i32 4991 %3 = bitcast ptr %y.addr.016 to ptr992 %4 = tail call <4 x i32> @llvm.masked.load.v4i32.p0(ptr %3, i32 4, <4 x i1> %0, <4 x i32> zeroinitializer)993 %5 = mul <4 x i32> %4, <i32 10, i32 10, i32 10, i32 10>994 %6 = add <4 x i32> %5, %2995 tail call void @llvm.masked.store.v4i32.p0(<4 x i32> %6, ptr %3, i32 4, <4 x i1> %0)996 %add.ptr1 = getelementptr inbounds i32, ptr %y.addr.016, i32 4997 %sub = add nsw i32 %i.015, -4998 %cmp = icmp sgt i32 %i.015, 4999 br i1 %cmp, label %for.body, label %for.cond.cleanup1000}1001 1002define void @vmlaqp(ptr %x, ptr %y, i32 %n) {1003; CHECK-LABEL: vmlaqp:1004; CHECK: @ %bb.0: @ %entry1005; CHECK-NEXT: .save {r7, lr}1006; CHECK-NEXT: push {r7, lr}1007; CHECK-NEXT: cmp r2, #11008; CHECK-NEXT: it lt1009; CHECK-NEXT: poplt {r7, pc}1010; CHECK-NEXT: .LBB23_1: @ %for.body.preheader1011; CHECK-NEXT: movs r3, #101012; CHECK-NEXT: dlstp.32 lr, r21013; CHECK-NEXT: .LBB23_2: @ %for.body1014; CHECK-NEXT: @ =>This Inner Loop Header: Depth=11015; CHECK-NEXT: vldrw.u32 q0, [r1]1016; CHECK-NEXT: vldrw.u32 q1, [r0], #161017; CHECK-NEXT: vmla.i32 q1, q0, r31018; CHECK-NEXT: vstrw.32 q1, [r1], #161019; CHECK-NEXT: letp lr, .LBB23_21020; CHECK-NEXT: @ %bb.3: @ %for.cond.cleanup1021; CHECK-NEXT: pop {r7, pc}1022entry:1023 %cmp15 = icmp sgt i32 %n, 01024 br i1 %cmp15, label %for.body, label %for.cond.cleanup1025 1026for.cond.cleanup: ; preds = %for.body, %entry1027 ret void1028 1029for.body: ; preds = %entry, %for.body1030 %x.addr.018 = phi ptr [ %add.ptr, %for.body ], [ %x, %entry ]1031 %y.addr.017 = phi ptr [ %add.ptr1, %for.body ], [ %y, %entry ]1032 %i.016 = phi i32 [ %sub, %for.body ], [ %n, %entry ]1033 %0 = tail call <4 x i1> @llvm.arm.mve.vctp32(i32 %i.016)1034 %1 = bitcast ptr %x.addr.018 to ptr1035 %2 = tail call <4 x i32> @llvm.masked.load.v4i32.p0(ptr %1, i32 4, <4 x i1> %0, <4 x i32> zeroinitializer)1036 %add.ptr = getelementptr inbounds i32, ptr %x.addr.018, i32 41037 %3 = bitcast ptr %y.addr.017 to ptr1038 %4 = tail call <4 x i32> @llvm.masked.load.v4i32.p0(ptr %3, i32 4, <4 x i1> %0, <4 x i32> zeroinitializer)1039 %5 = tail call <4 x i32> @llvm.arm.mve.vmla.n.predicated.v4i32.v4i1(<4 x i32> %2, <4 x i32> %4, i32 10, <4 x i1> %0)1040 tail call void @llvm.masked.store.v4i32.p0(<4 x i32> %5, ptr %3, i32 4, <4 x i1> %0)1041 %add.ptr1 = getelementptr inbounds i32, ptr %y.addr.017, i32 41042 %sub = add nsw i32 %i.016, -41043 %cmp = icmp sgt i32 %i.016, 41044 br i1 %cmp, label %for.body, label %for.cond.cleanup1045}1046 1047define void @vmlasq(ptr %x, ptr %y, i32 %n) {1048; CHECK-LABEL: vmlasq:1049; CHECK: @ %bb.0: @ %entry1050; CHECK-NEXT: .save {r7, lr}1051; CHECK-NEXT: push {r7, lr}1052; CHECK-NEXT: cmp r2, #11053; CHECK-NEXT: it lt1054; CHECK-NEXT: poplt {r7, pc}1055; CHECK-NEXT: .LBB24_1: @ %for.body.preheader1056; CHECK-NEXT: movs r3, #101057; CHECK-NEXT: dlstp.32 lr, r21058; CHECK-NEXT: .LBB24_2: @ %for.body1059; CHECK-NEXT: @ =>This Inner Loop Header: Depth=11060; CHECK-NEXT: vldrw.u32 q0, [r0], #161061; CHECK-NEXT: vldrw.u32 q1, [r1]1062; CHECK-NEXT: vmlas.i32 q1, q0, r31063; CHECK-NEXT: vstrw.32 q1, [r1], #161064; CHECK-NEXT: letp lr, .LBB24_21065; CHECK-NEXT: @ %bb.3: @ %for.cond.cleanup1066; CHECK-NEXT: pop {r7, pc}1067entry:1068 %cmp14 = icmp sgt i32 %n, 01069 br i1 %cmp14, label %for.body, label %for.cond.cleanup1070 1071for.cond.cleanup: ; preds = %for.body, %entry1072 ret void1073 1074for.body: ; preds = %entry, %for.body1075 %x.addr.017 = phi ptr [ %add.ptr, %for.body ], [ %x, %entry ]1076 %y.addr.016 = phi ptr [ %add.ptr1, %for.body ], [ %y, %entry ]1077 %i.015 = phi i32 [ %sub, %for.body ], [ %n, %entry ]1078 %0 = tail call <4 x i1> @llvm.arm.mve.vctp32(i32 %i.015)1079 %1 = bitcast ptr %x.addr.017 to ptr1080 %2 = tail call <4 x i32> @llvm.masked.load.v4i32.p0(ptr %1, i32 4, <4 x i1> %0, <4 x i32> zeroinitializer)1081 %add.ptr = getelementptr inbounds i32, ptr %x.addr.017, i32 41082 %3 = bitcast ptr %y.addr.016 to ptr1083 %4 = tail call <4 x i32> @llvm.masked.load.v4i32.p0(ptr %3, i32 4, <4 x i1> %0, <4 x i32> zeroinitializer)1084 %5 = mul <4 x i32> %4, %21085 %6 = add <4 x i32> %5, <i32 10, i32 10, i32 10, i32 10>1086 tail call void @llvm.masked.store.v4i32.p0(<4 x i32> %6, ptr %3, i32 4, <4 x i1> %0)1087 %add.ptr1 = getelementptr inbounds i32, ptr %y.addr.016, i32 41088 %sub = add nsw i32 %i.015, -41089 %cmp = icmp sgt i32 %i.015, 41090 br i1 %cmp, label %for.body, label %for.cond.cleanup1091}1092 1093define void @vmlasqp(ptr %x, ptr %y, i32 %n) {1094; CHECK-LABEL: vmlasqp:1095; CHECK: @ %bb.0: @ %entry1096; CHECK-NEXT: .save {r7, lr}1097; CHECK-NEXT: push {r7, lr}1098; CHECK-NEXT: cmp r2, #11099; CHECK-NEXT: it lt1100; CHECK-NEXT: poplt {r7, pc}1101; CHECK-NEXT: .LBB25_1: @ %for.body.preheader1102; CHECK-NEXT: movs r3, #101103; CHECK-NEXT: dlstp.32 lr, r21104; CHECK-NEXT: .LBB25_2: @ %for.body1105; CHECK-NEXT: @ =>This Inner Loop Header: Depth=11106; CHECK-NEXT: vldrw.u32 q0, [r1]1107; CHECK-NEXT: vldrw.u32 q1, [r0], #161108; CHECK-NEXT: vmlas.i32 q1, q0, r31109; CHECK-NEXT: vstrw.32 q1, [r1], #161110; CHECK-NEXT: letp lr, .LBB25_21111; CHECK-NEXT: @ %bb.3: @ %for.cond.cleanup1112; CHECK-NEXT: pop {r7, pc}1113entry:1114 %cmp15 = icmp sgt i32 %n, 01115 br i1 %cmp15, label %for.body, label %for.cond.cleanup1116 1117for.cond.cleanup: ; preds = %for.body, %entry1118 ret void1119 1120for.body: ; preds = %entry, %for.body1121 %x.addr.018 = phi ptr [ %add.ptr, %for.body ], [ %x, %entry ]1122 %y.addr.017 = phi ptr [ %add.ptr1, %for.body ], [ %y, %entry ]1123 %i.016 = phi i32 [ %sub, %for.body ], [ %n, %entry ]1124 %0 = tail call <4 x i1> @llvm.arm.mve.vctp32(i32 %i.016)1125 %1 = bitcast ptr %x.addr.018 to ptr1126 %2 = tail call <4 x i32> @llvm.masked.load.v4i32.p0(ptr %1, i32 4, <4 x i1> %0, <4 x i32> zeroinitializer)1127 %add.ptr = getelementptr inbounds i32, ptr %x.addr.018, i32 41128 %3 = bitcast ptr %y.addr.017 to ptr1129 %4 = tail call <4 x i32> @llvm.masked.load.v4i32.p0(ptr %3, i32 4, <4 x i1> %0, <4 x i32> zeroinitializer)1130 %5 = tail call <4 x i32> @llvm.arm.mve.vmlas.n.predicated.v4i32.v4i1(<4 x i32> %2, <4 x i32> %4, i32 10, <4 x i1> %0)1131 tail call void @llvm.masked.store.v4i32.p0(<4 x i32> %5, ptr %3, i32 4, <4 x i1> %0)1132 %add.ptr1 = getelementptr inbounds i32, ptr %y.addr.017, i32 41133 %sub = add nsw i32 %i.016, -41134 %cmp = icmp sgt i32 %i.016, 41135 br i1 %cmp, label %for.body, label %for.cond.cleanup1136}1137 1138define void @vaddqf(ptr %x, ptr %y, i32 %n) {1139; CHECK-LABEL: vaddqf:1140; CHECK: @ %bb.0: @ %entry1141; CHECK-NEXT: .save {r7, lr}1142; CHECK-NEXT: push {r7, lr}1143; CHECK-NEXT: cmp r2, #11144; CHECK-NEXT: it lt1145; CHECK-NEXT: poplt {r7, pc}1146; CHECK-NEXT: .LBB26_1: @ %for.body.preheader1147; CHECK-NEXT: vmov.f32 q0, #1.000000e+011148; CHECK-NEXT: dlstp.32 lr, r21149; CHECK-NEXT: .LBB26_2: @ %for.body1150; CHECK-NEXT: @ =>This Inner Loop Header: Depth=11151; CHECK-NEXT: vldrw.u32 q1, [r0], #161152; CHECK-NEXT: vadd.f32 q1, q1, q01153; CHECK-NEXT: vstrw.32 q1, [r1], #161154; CHECK-NEXT: letp lr, .LBB26_21155; CHECK-NEXT: @ %bb.3: @ %for.cond.cleanup1156; CHECK-NEXT: pop {r7, pc}1157entry:1158 %cmp11 = icmp sgt i32 %n, 01159 br i1 %cmp11, label %for.body, label %for.cond.cleanup1160 1161for.cond.cleanup: ; preds = %for.body, %entry1162 ret void1163 1164for.body: ; preds = %entry, %for.body1165 %x.addr.014 = phi ptr [ %add.ptr, %for.body ], [ %x, %entry ]1166 %y.addr.013 = phi ptr [ %add.ptr1, %for.body ], [ %y, %entry ]1167 %i.012 = phi i32 [ %sub, %for.body ], [ %n, %entry ]1168 %0 = tail call <4 x i1> @llvm.arm.mve.vctp32(i32 %i.012)1169 %1 = bitcast ptr %x.addr.014 to ptr1170 %2 = tail call fast <4 x float> @llvm.masked.load.v4f32.p0(ptr %1, i32 4, <4 x i1> %0, <4 x float> zeroinitializer)1171 %add.ptr = getelementptr inbounds float, ptr %x.addr.014, i32 41172 %3 = fadd fast <4 x float> %2, <float 10.0, float 10.0, float 10.0, float 10.0>1173 %4 = bitcast ptr %y.addr.013 to ptr1174 tail call void @llvm.masked.store.v4f32.p0(<4 x float> %3, ptr %4, i32 4, <4 x i1> %0)1175 %add.ptr1 = getelementptr inbounds float, ptr %y.addr.013, i32 41176 %sub = add nsw i32 %i.012, -41177 %cmp = icmp sgt i32 %i.012, 41178 br i1 %cmp, label %for.body, label %for.cond.cleanup1179}1180 1181define void @vaddf(ptr %s1, i32 %N) {1182; CHECK-LABEL: vaddf:1183; CHECK: @ %bb.0: @ %entry1184; CHECK-NEXT: .save {r7, lr}1185; CHECK-NEXT: push {r7, lr}1186; CHECK-NEXT: cmp r1, #11187; CHECK-NEXT: it lt1188; CHECK-NEXT: poplt {r7, pc}1189; CHECK-NEXT: .LBB27_1: @ %while.body.preheader1190; CHECK-NEXT: movs r2, #01191; CHECK-NEXT: movt r2, #166721192; CHECK-NEXT: dlstp.32 lr, r11193; CHECK-NEXT: .LBB27_2: @ %while.body1194; CHECK-NEXT: @ =>This Inner Loop Header: Depth=11195; CHECK-NEXT: vldrw.u32 q0, [r0]1196; CHECK-NEXT: vadd.f32 q0, q0, r21197; CHECK-NEXT: vstrw.32 q0, [r0], #161198; CHECK-NEXT: letp lr, .LBB27_21199; CHECK-NEXT: @ %bb.3: @ %while.end1200; CHECK-NEXT: pop {r7, pc}1201entry:1202 %cmp11 = icmp sgt i32 %N, 01203 br i1 %cmp11, label %while.body.lr.ph, label %while.end1204 1205while.body.lr.ph: ; preds = %entry1206 br label %while.body1207 1208while.body: ; preds = %while.body.lr.ph, %while.body1209 %s1.addr.013 = phi ptr [ %s1, %while.body.lr.ph ], [ %add.ptr, %while.body ]1210 %N.addr.012 = phi i32 [ %N, %while.body.lr.ph ], [ %sub, %while.body ]1211 %0 = tail call <4 x i1> @llvm.arm.mve.vctp32(i32 %N.addr.012)1212 %1 = bitcast ptr %s1.addr.013 to ptr1213 %2 = tail call fast <4 x float> @llvm.masked.load.v4f32.p0(ptr %1, i32 4, <4 x i1> %0, <4 x float> zeroinitializer)1214 %3 = tail call fast <4 x float> @llvm.arm.mve.add.predicated.v4f32.v4i1(<4 x float> %2, <4 x float> <float 10.0, float 10.0, float 10.0, float 10.0>, <4 x i1> %0, <4 x float> %2)1215 tail call void @llvm.masked.store.v4f32.p0(<4 x float> %3, ptr %1, i32 4, <4 x i1> %0)1216 %add.ptr = getelementptr inbounds float, ptr %s1.addr.013, i32 41217 %sub = add nsw i32 %N.addr.012, -41218 %cmp = icmp sgt i32 %N.addr.012, 41219 br i1 %cmp, label %while.body, label %while.end1220 1221while.end: ; preds = %while.body, %entry1222 ret void1223}1224 1225define void @vsubqf(ptr %x, ptr %y, i32 %n) {1226; CHECK-LABEL: vsubqf:1227; CHECK: @ %bb.0: @ %entry1228; CHECK-NEXT: .save {r7, lr}1229; CHECK-NEXT: push {r7, lr}1230; CHECK-NEXT: cmp r2, #11231; CHECK-NEXT: it lt1232; CHECK-NEXT: poplt {r7, pc}1233; CHECK-NEXT: .LBB28_1: @ %for.body.preheader1234; CHECK-NEXT: vmov.f32 q0, #-1.000000e+011235; CHECK-NEXT: dlstp.32 lr, r21236; CHECK-NEXT: .LBB28_2: @ %for.body1237; CHECK-NEXT: @ =>This Inner Loop Header: Depth=11238; CHECK-NEXT: vldrw.u32 q1, [r0], #161239; CHECK-NEXT: vadd.f32 q1, q1, q01240; CHECK-NEXT: vstrw.32 q1, [r1], #161241; CHECK-NEXT: letp lr, .LBB28_21242; CHECK-NEXT: @ %bb.3: @ %for.cond.cleanup1243; CHECK-NEXT: pop {r7, pc}1244entry:1245 %cmp11 = icmp sgt i32 %n, 01246 br i1 %cmp11, label %for.body, label %for.cond.cleanup1247 1248for.cond.cleanup: ; preds = %for.body, %entry1249 ret void1250 1251for.body: ; preds = %entry, %for.body1252 %x.addr.014 = phi ptr [ %add.ptr, %for.body ], [ %x, %entry ]1253 %y.addr.013 = phi ptr [ %add.ptr1, %for.body ], [ %y, %entry ]1254 %i.012 = phi i32 [ %sub, %for.body ], [ %n, %entry ]1255 %0 = tail call <4 x i1> @llvm.arm.mve.vctp32(i32 %i.012)1256 %1 = bitcast ptr %x.addr.014 to ptr1257 %2 = tail call fast <4 x float> @llvm.masked.load.v4f32.p0(ptr %1, i32 4, <4 x i1> %0, <4 x float> zeroinitializer)1258 %add.ptr = getelementptr inbounds float, ptr %x.addr.014, i32 41259 %3 = fsub fast <4 x float> %2, <float 10.0, float 10.0, float 10.0, float 10.0>1260 %4 = bitcast ptr %y.addr.013 to ptr1261 tail call void @llvm.masked.store.v4f32.p0(<4 x float> %3, ptr %4, i32 4, <4 x i1> %0)1262 %add.ptr1 = getelementptr inbounds float, ptr %y.addr.013, i32 41263 %sub = add nsw i32 %i.012, -41264 %cmp = icmp sgt i32 %i.012, 41265 br i1 %cmp, label %for.body, label %for.cond.cleanup1266}1267 1268define void @vsubf(ptr %s1, i32 %N) {1269; CHECK-LABEL: vsubf:1270; CHECK: @ %bb.0: @ %entry1271; CHECK-NEXT: .save {r7, lr}1272; CHECK-NEXT: push {r7, lr}1273; CHECK-NEXT: cmp r1, #11274; CHECK-NEXT: it lt1275; CHECK-NEXT: poplt {r7, pc}1276; CHECK-NEXT: .LBB29_1: @ %while.body.preheader1277; CHECK-NEXT: movs r2, #01278; CHECK-NEXT: movt r2, #166721279; CHECK-NEXT: dlstp.32 lr, r11280; CHECK-NEXT: .LBB29_2: @ %while.body1281; CHECK-NEXT: @ =>This Inner Loop Header: Depth=11282; CHECK-NEXT: vldrw.u32 q0, [r0]1283; CHECK-NEXT: vsub.f32 q0, q0, r21284; CHECK-NEXT: vstrw.32 q0, [r0], #161285; CHECK-NEXT: letp lr, .LBB29_21286; CHECK-NEXT: @ %bb.3: @ %while.end1287; CHECK-NEXT: pop {r7, pc}1288entry:1289 %cmp11 = icmp sgt i32 %N, 01290 br i1 %cmp11, label %while.body.lr.ph, label %while.end1291 1292while.body.lr.ph: ; preds = %entry1293 br label %while.body1294 1295while.body: ; preds = %while.body.lr.ph, %while.body1296 %s1.addr.013 = phi ptr [ %s1, %while.body.lr.ph ], [ %add.ptr, %while.body ]1297 %N.addr.012 = phi i32 [ %N, %while.body.lr.ph ], [ %sub, %while.body ]1298 %0 = tail call <4 x i1> @llvm.arm.mve.vctp32(i32 %N.addr.012)1299 %1 = bitcast ptr %s1.addr.013 to ptr1300 %2 = tail call fast <4 x float> @llvm.masked.load.v4f32.p0(ptr %1, i32 4, <4 x i1> %0, <4 x float> zeroinitializer)1301 %3 = tail call fast <4 x float> @llvm.arm.mve.sub.predicated.v4f32.v4i1(<4 x float> %2, <4 x float> <float 10.0, float 10.0, float 10.0, float 10.0>, <4 x i1> %0, <4 x float> %2)1302 tail call void @llvm.masked.store.v4f32.p0(<4 x float> %3, ptr %1, i32 4, <4 x i1> %0)1303 %add.ptr = getelementptr inbounds float, ptr %s1.addr.013, i32 41304 %sub = add nsw i32 %N.addr.012, -41305 %cmp = icmp sgt i32 %N.addr.012, 41306 br i1 %cmp, label %while.body, label %while.end1307 1308while.end: ; preds = %while.body, %entry1309 ret void1310}1311 1312define void @vmulqf(ptr %x, ptr %y, i32 %n) {1313; CHECK-LABEL: vmulqf:1314; CHECK: @ %bb.0: @ %entry1315; CHECK-NEXT: .save {r7, lr}1316; CHECK-NEXT: push {r7, lr}1317; CHECK-NEXT: cmp r2, #11318; CHECK-NEXT: it lt1319; CHECK-NEXT: poplt {r7, pc}1320; CHECK-NEXT: .LBB30_1: @ %for.body.preheader1321; CHECK-NEXT: vmov.f32 q0, #1.000000e+011322; CHECK-NEXT: dlstp.32 lr, r21323; CHECK-NEXT: .LBB30_2: @ %for.body1324; CHECK-NEXT: @ =>This Inner Loop Header: Depth=11325; CHECK-NEXT: vldrw.u32 q1, [r0], #161326; CHECK-NEXT: vmul.f32 q1, q1, q01327; CHECK-NEXT: vstrw.32 q1, [r1], #161328; CHECK-NEXT: letp lr, .LBB30_21329; CHECK-NEXT: @ %bb.3: @ %for.cond.cleanup1330; CHECK-NEXT: pop {r7, pc}1331entry:1332 %cmp11 = icmp sgt i32 %n, 01333 br i1 %cmp11, label %for.body, label %for.cond.cleanup1334 1335for.cond.cleanup: ; preds = %for.body, %entry1336 ret void1337 1338for.body: ; preds = %entry, %for.body1339 %x.addr.014 = phi ptr [ %add.ptr, %for.body ], [ %x, %entry ]1340 %y.addr.013 = phi ptr [ %add.ptr1, %for.body ], [ %y, %entry ]1341 %i.012 = phi i32 [ %sub, %for.body ], [ %n, %entry ]1342 %0 = tail call <4 x i1> @llvm.arm.mve.vctp32(i32 %i.012)1343 %1 = bitcast ptr %x.addr.014 to ptr1344 %2 = tail call fast <4 x float> @llvm.masked.load.v4f32.p0(ptr %1, i32 4, <4 x i1> %0, <4 x float> zeroinitializer)1345 %add.ptr = getelementptr inbounds float, ptr %x.addr.014, i32 41346 %3 = fmul fast <4 x float> %2, <float 10.0, float 10.0, float 10.0, float 10.0>1347 %4 = bitcast ptr %y.addr.013 to ptr1348 tail call void @llvm.masked.store.v4f32.p0(<4 x float> %3, ptr %4, i32 4, <4 x i1> %0)1349 %add.ptr1 = getelementptr inbounds float, ptr %y.addr.013, i32 41350 %sub = add nsw i32 %i.012, -41351 %cmp = icmp sgt i32 %i.012, 41352 br i1 %cmp, label %for.body, label %for.cond.cleanup1353}1354 1355define void @vmulf(ptr %s1, i32 %N) {1356; CHECK-LABEL: vmulf:1357; CHECK: @ %bb.0: @ %entry1358; CHECK-NEXT: .save {r7, lr}1359; CHECK-NEXT: push {r7, lr}1360; CHECK-NEXT: cmp r1, #11361; CHECK-NEXT: it lt1362; CHECK-NEXT: poplt {r7, pc}1363; CHECK-NEXT: .LBB31_1: @ %while.body.preheader1364; CHECK-NEXT: movs r2, #01365; CHECK-NEXT: movt r2, #166721366; CHECK-NEXT: dlstp.32 lr, r11367; CHECK-NEXT: .LBB31_2: @ %while.body1368; CHECK-NEXT: @ =>This Inner Loop Header: Depth=11369; CHECK-NEXT: vldrw.u32 q0, [r0]1370; CHECK-NEXT: vmul.f32 q0, q0, r21371; CHECK-NEXT: vstrw.32 q0, [r0], #161372; CHECK-NEXT: letp lr, .LBB31_21373; CHECK-NEXT: @ %bb.3: @ %while.end1374; CHECK-NEXT: pop {r7, pc}1375entry:1376 %cmp11 = icmp sgt i32 %N, 01377 br i1 %cmp11, label %while.body.lr.ph, label %while.end1378 1379while.body.lr.ph: ; preds = %entry1380 br label %while.body1381 1382while.body: ; preds = %while.body.lr.ph, %while.body1383 %s1.addr.013 = phi ptr [ %s1, %while.body.lr.ph ], [ %add.ptr, %while.body ]1384 %N.addr.012 = phi i32 [ %N, %while.body.lr.ph ], [ %sub, %while.body ]1385 %0 = tail call <4 x i1> @llvm.arm.mve.vctp32(i32 %N.addr.012)1386 %1 = bitcast ptr %s1.addr.013 to ptr1387 %2 = tail call fast <4 x float> @llvm.masked.load.v4f32.p0(ptr %1, i32 4, <4 x i1> %0, <4 x float> zeroinitializer)1388 %3 = tail call fast <4 x float> @llvm.arm.mve.mul.predicated.v4f32.v4i1(<4 x float> %2, <4 x float> <float 10.0, float 10.0, float 10.0, float 10.0>, <4 x i1> %0, <4 x float> %2)1389 tail call void @llvm.masked.store.v4f32.p0(<4 x float> %3, ptr %1, i32 4, <4 x i1> %0)1390 %add.ptr = getelementptr inbounds float, ptr %s1.addr.013, i32 41391 %sub = add nsw i32 %N.addr.012, -41392 %cmp = icmp sgt i32 %N.addr.012, 41393 br i1 %cmp, label %while.body, label %while.end1394 1395while.end: ; preds = %while.body, %entry1396 ret void1397}1398 1399define void @vfmaq(ptr %x, ptr %y, i32 %n) {1400; CHECK-LABEL: vfmaq:1401; CHECK: @ %bb.0: @ %entry1402; CHECK-NEXT: .save {r7, lr}1403; CHECK-NEXT: push {r7, lr}1404; CHECK-NEXT: cmp r2, #11405; CHECK-NEXT: it lt1406; CHECK-NEXT: poplt {r7, pc}1407; CHECK-NEXT: .LBB32_1: @ %for.body.preheader1408; CHECK-NEXT: vmov.f32 q0, #1.000000e+011409; CHECK-NEXT: dlstp.32 lr, r21410; CHECK-NEXT: .LBB32_2: @ %for.body1411; CHECK-NEXT: @ =>This Inner Loop Header: Depth=11412; CHECK-NEXT: vldrw.u32 q1, [r1]1413; CHECK-NEXT: vldrw.u32 q2, [r0], #161414; CHECK-NEXT: vfma.f32 q2, q1, q01415; CHECK-NEXT: vstrw.32 q2, [r1], #161416; CHECK-NEXT: letp lr, .LBB32_21417; CHECK-NEXT: @ %bb.3: @ %for.cond.cleanup1418; CHECK-NEXT: pop {r7, pc}1419entry:1420 %cmp14 = icmp sgt i32 %n, 01421 br i1 %cmp14, label %for.body, label %for.cond.cleanup1422 1423for.cond.cleanup: ; preds = %for.body, %entry1424 ret void1425 1426for.body: ; preds = %entry, %for.body1427 %x.addr.017 = phi ptr [ %add.ptr, %for.body ], [ %x, %entry ]1428 %y.addr.016 = phi ptr [ %add.ptr1, %for.body ], [ %y, %entry ]1429 %i.015 = phi i32 [ %sub, %for.body ], [ %n, %entry ]1430 %0 = tail call <4 x i1> @llvm.arm.mve.vctp32(i32 %i.015)1431 %1 = bitcast ptr %x.addr.017 to ptr1432 %2 = tail call fast <4 x float> @llvm.masked.load.v4f32.p0(ptr %1, i32 4, <4 x i1> %0, <4 x float> zeroinitializer)1433 %add.ptr = getelementptr inbounds float, ptr %x.addr.017, i32 41434 %3 = bitcast ptr %y.addr.016 to ptr1435 %4 = tail call fast <4 x float> @llvm.masked.load.v4f32.p0(ptr %3, i32 4, <4 x i1> %0, <4 x float> zeroinitializer)1436 %5 = tail call fast <4 x float> @llvm.fma.v4f32(<4 x float> %4, <4 x float> <float 10.0, float 10.0, float 10.0, float 10.0>, <4 x float> %2)1437 tail call void @llvm.masked.store.v4f32.p0(<4 x float> %5, ptr %3, i32 4, <4 x i1> %0)1438 %add.ptr1 = getelementptr inbounds float, ptr %y.addr.016, i32 41439 %sub = add nsw i32 %i.015, -41440 %cmp = icmp sgt i32 %i.015, 41441 br i1 %cmp, label %for.body, label %for.cond.cleanup1442}1443 1444define void @vfma(ptr %s1, ptr %s2, i32 %N) {1445; CHECK-LABEL: vfma:1446; CHECK: @ %bb.0: @ %entry1447; CHECK-NEXT: .save {r7, lr}1448; CHECK-NEXT: push {r7, lr}1449; CHECK-NEXT: cmp r2, #11450; CHECK-NEXT: it lt1451; CHECK-NEXT: poplt {r7, pc}1452; CHECK-NEXT: .LBB33_1: @ %while.body.lr.ph1453; CHECK-NEXT: vmov.f32 q0, #1.000000e+011454; CHECK-NEXT: dlstp.32 lr, r21455; CHECK-NEXT: .LBB33_2: @ %while.body1456; CHECK-NEXT: @ =>This Inner Loop Header: Depth=11457; CHECK-NEXT: vldrw.u32 q1, [r1]1458; CHECK-NEXT: vldrw.u32 q2, [r0]1459; CHECK-NEXT: vfma.f32 q2, q1, q01460; CHECK-NEXT: vstrw.32 q2, [r0], #161461; CHECK-NEXT: letp lr, .LBB33_21462; CHECK-NEXT: @ %bb.3: @ %while.end1463; CHECK-NEXT: pop {r7, pc}1464entry:1465 %cmp12 = icmp sgt i32 %N, 01466 br i1 %cmp12, label %while.body.lr.ph, label %while.end1467 1468while.body.lr.ph: ; preds = %entry1469 %0 = bitcast ptr %s2 to ptr1470 br label %while.body1471 1472while.body: ; preds = %while.body.lr.ph, %while.body1473 %s1.addr.014 = phi ptr [ %s1, %while.body.lr.ph ], [ %add.ptr, %while.body ]1474 %N.addr.013 = phi i32 [ %N, %while.body.lr.ph ], [ %sub, %while.body ]1475 %1 = tail call <4 x i1> @llvm.arm.mve.vctp32(i32 %N.addr.013)1476 %2 = bitcast ptr %s1.addr.014 to ptr1477 %3 = tail call fast <4 x float> @llvm.masked.load.v4f32.p0(ptr %2, i32 4, <4 x i1> %1, <4 x float> zeroinitializer)1478 %4 = tail call fast <4 x float> @llvm.masked.load.v4f32.p0(ptr %0, i32 4, <4 x i1> %1, <4 x float> zeroinitializer)1479 %5 = tail call fast <4 x float> @llvm.arm.mve.fma.predicated.v4f32.v4i1(<4 x float> %4, <4 x float> <float 10.0, float 10.0, float 10.0, float 10.0>, <4 x float> %3, <4 x i1> %1)1480 tail call void @llvm.masked.store.v4f32.p0(<4 x float> %5, ptr %2, i32 4, <4 x i1> %1)1481 %add.ptr = getelementptr inbounds float, ptr %s1.addr.014, i32 41482 %sub = add nsw i32 %N.addr.013, -41483 %cmp = icmp sgt i32 %N.addr.013, 41484 br i1 %cmp, label %while.body, label %while.end1485 1486while.end: ; preds = %while.body, %entry1487 ret void1488}1489 1490define void @vfmasq(ptr %x, ptr %y, i32 %n) {1491; CHECK-LABEL: vfmasq:1492; CHECK: @ %bb.0: @ %entry1493; CHECK-NEXT: .save {r7, lr}1494; CHECK-NEXT: push {r7, lr}1495; CHECK-NEXT: cmp r2, #11496; CHECK-NEXT: it lt1497; CHECK-NEXT: poplt {r7, pc}1498; CHECK-NEXT: .LBB34_1: @ %for.body.preheader1499; CHECK-NEXT: dlstp.32 lr, r21500; CHECK-NEXT: .LBB34_2: @ %for.body1501; CHECK-NEXT: @ =>This Inner Loop Header: Depth=11502; CHECK-NEXT: vmov.f32 q2, #1.000000e+011503; CHECK-NEXT: vldrw.u32 q0, [r1]1504; CHECK-NEXT: vldrw.u32 q1, [r0], #161505; CHECK-NEXT: vfma.f32 q2, q1, q01506; CHECK-NEXT: vstrw.32 q2, [r1], #161507; CHECK-NEXT: letp lr, .LBB34_21508; CHECK-NEXT: @ %bb.3: @ %for.cond.cleanup1509; CHECK-NEXT: pop {r7, pc}1510entry:1511 %cmp14 = icmp sgt i32 %n, 01512 br i1 %cmp14, label %for.body, label %for.cond.cleanup1513 1514for.cond.cleanup: ; preds = %for.body, %entry1515 ret void1516 1517for.body: ; preds = %entry, %for.body1518 %x.addr.017 = phi ptr [ %add.ptr, %for.body ], [ %x, %entry ]1519 %y.addr.016 = phi ptr [ %add.ptr1, %for.body ], [ %y, %entry ]1520 %i.015 = phi i32 [ %sub, %for.body ], [ %n, %entry ]1521 %0 = tail call <4 x i1> @llvm.arm.mve.vctp32(i32 %i.015)1522 %1 = bitcast ptr %x.addr.017 to ptr1523 %2 = tail call fast <4 x float> @llvm.masked.load.v4f32.p0(ptr %1, i32 4, <4 x i1> %0, <4 x float> zeroinitializer)1524 %add.ptr = getelementptr inbounds float, ptr %x.addr.017, i32 41525 %3 = bitcast ptr %y.addr.016 to ptr1526 %4 = tail call fast <4 x float> @llvm.masked.load.v4f32.p0(ptr %3, i32 4, <4 x i1> %0, <4 x float> zeroinitializer)1527 %5 = tail call fast <4 x float> @llvm.fma.v4f32(<4 x float> %2, <4 x float> %4, <4 x float> <float 10.0, float 10.0, float 10.0, float 10.0>)1528 tail call void @llvm.masked.store.v4f32.p0(<4 x float> %5, ptr %3, i32 4, <4 x i1> %0)1529 %add.ptr1 = getelementptr inbounds float, ptr %y.addr.016, i32 41530 %sub = add nsw i32 %i.015, -41531 %cmp = icmp sgt i32 %i.015, 41532 br i1 %cmp, label %for.body, label %for.cond.cleanup1533}1534 1535define void @vfmas(ptr %s1, ptr %s2, i32 %N) {1536; CHECK-LABEL: vfmas:1537; CHECK: @ %bb.0: @ %entry1538; CHECK-NEXT: .save {r7, lr}1539; CHECK-NEXT: push {r7, lr}1540; CHECK-NEXT: cmp r2, #11541; CHECK-NEXT: it lt1542; CHECK-NEXT: poplt {r7, pc}1543; CHECK-NEXT: .LBB35_1: @ %while.body.lr.ph1544; CHECK-NEXT: dlstp.32 lr, r21545; CHECK-NEXT: .LBB35_2: @ %while.body1546; CHECK-NEXT: @ =>This Inner Loop Header: Depth=11547; CHECK-NEXT: vmov.f32 q2, #1.000000e+011548; CHECK-NEXT: vldrw.u32 q0, [r1]1549; CHECK-NEXT: vldrw.u32 q1, [r0]1550; CHECK-NEXT: vfma.f32 q2, q1, q01551; CHECK-NEXT: vstrw.32 q2, [r0], #161552; CHECK-NEXT: letp lr, .LBB35_21553; CHECK-NEXT: @ %bb.3: @ %while.end1554; CHECK-NEXT: pop {r7, pc}1555entry:1556 %cmp12 = icmp sgt i32 %N, 01557 br i1 %cmp12, label %while.body.lr.ph, label %while.end1558 1559while.body.lr.ph: ; preds = %entry1560 %0 = bitcast ptr %s2 to ptr1561 br label %while.body1562 1563while.body: ; preds = %while.body.lr.ph, %while.body1564 %s1.addr.014 = phi ptr [ %s1, %while.body.lr.ph ], [ %add.ptr, %while.body ]1565 %N.addr.013 = phi i32 [ %N, %while.body.lr.ph ], [ %sub, %while.body ]1566 %1 = tail call <4 x i1> @llvm.arm.mve.vctp32(i32 %N.addr.013)1567 %2 = bitcast ptr %s1.addr.014 to ptr1568 %3 = tail call fast <4 x float> @llvm.masked.load.v4f32.p0(ptr %2, i32 4, <4 x i1> %1, <4 x float> zeroinitializer)1569 %4 = tail call fast <4 x float> @llvm.masked.load.v4f32.p0(ptr %0, i32 4, <4 x i1> %1, <4 x float> zeroinitializer)1570 %5 = tail call fast <4 x float> @llvm.arm.mve.fma.predicated.v4f32.v4i1(<4 x float> %3, <4 x float> %4, <4 x float> <float 10.0, float 10.0, float 10.0, float 10.0>, <4 x i1> %1)1571 tail call void @llvm.masked.store.v4f32.p0(<4 x float> %5, ptr %2, i32 4, <4 x i1> %1)1572 %add.ptr = getelementptr inbounds float, ptr %s1.addr.014, i32 41573 %sub = add nsw i32 %N.addr.013, -41574 %cmp = icmp sgt i32 %N.addr.013, 41575 br i1 %cmp, label %while.body, label %while.end1576 1577while.end: ; preds = %while.body, %entry1578 ret void1579}1580 1581define void @rgbconvert(ptr noalias %pwSourceBase, i16 signext %iSourceStride, ptr noalias %phwTargetBase, i16 signext %iTargetStride, i16 %iHeight, i16 %iWidth) {1582; CHECK-LABEL: rgbconvert:1583; CHECK: @ %bb.0: @ %entry1584; CHECK-NEXT: .save {r4, r5, r6, r7, r8, r9, r10, r11, lr}1585; CHECK-NEXT: push.w {r4, r5, r6, r7, r8, r9, r10, r11, lr}1586; CHECK-NEXT: .pad #41587; CHECK-NEXT: sub sp, #41588; CHECK-NEXT: .vsave {d8, d9, d10, d11}1589; CHECK-NEXT: vpush {d8, d9, d10, d11}1590; CHECK-NEXT: .pad #81591; CHECK-NEXT: sub sp, #81592; CHECK-NEXT: str r3, [sp, #4] @ 4-byte Spill1593; CHECK-NEXT: ldrsh.w r3, [sp, #80]1594; CHECK-NEXT: cmp r3, #11595; CHECK-NEXT: blt .LBB36_51596; CHECK-NEXT: @ %bb.1: @ %for.body.lr.ph1597; CHECK-NEXT: mov r9, r21598; CHECK-NEXT: ldr r2, [sp, #84]1599; CHECK-NEXT: mov.w r10, #01600; CHECK-NEXT: mov.w r11, #83886081601; CHECK-NEXT: mov.w r4, #671088641602; CHECK-NEXT: sxth.w r12, r21603; CHECK-NEXT: vmov.i32 q0, #0xf8001604; CHECK-NEXT: vmov.i32 q1, #0x1f1605; CHECK-NEXT: mov.w r2, #20161606; CHECK-NEXT: mov.w r7, #2684354561607; CHECK-NEXT: vdup.32 q2, r21608; CHECK-NEXT: .LBB36_2: @ %for.body1609; CHECK-NEXT: @ =>This Loop Header: Depth=11610; CHECK-NEXT: @ Child Loop BB36_3 Depth 21611; CHECK-NEXT: mov r2, r91612; CHECK-NEXT: mov r5, r01613; CHECK-NEXT: dlstp.32 lr, r121614; CHECK-NEXT: .LBB36_3: @ %do.body1615; CHECK-NEXT: @ Parent Loop BB36_2 Depth=11616; CHECK-NEXT: @ => This Inner Loop Header: Depth=21617; CHECK-NEXT: vldrw.u32 q3, [r5], #161618; CHECK-NEXT: vqdmulh.s32 q4, q3, r41619; CHECK-NEXT: vqdmulh.s32 q5, q3, r71620; CHECK-NEXT: vqdmulh.s32 q3, q3, r111621; CHECK-NEXT: vand q4, q4, q21622; CHECK-NEXT: vand q5, q5, q11623; CHECK-NEXT: vand q3, q3, q01624; CHECK-NEXT: vorr q4, q4, q51625; CHECK-NEXT: vorr q3, q4, q31626; CHECK-NEXT: vstrh.32 q3, [r2], #81627; CHECK-NEXT: letp lr, .LBB36_31628; CHECK-NEXT: @ %bb.4: @ %do.end1629; CHECK-NEXT: @ in Loop: Header=BB36_2 Depth=11630; CHECK-NEXT: ldr r2, [sp, #4] @ 4-byte Reload1631; CHECK-NEXT: add.w r10, r10, #11632; CHECK-NEXT: add.w r0, r0, r1, lsl #21633; CHECK-NEXT: cmp r10, r31634; CHECK-NEXT: add.w r9, r9, r2, lsl #11635; CHECK-NEXT: bne .LBB36_21636; CHECK-NEXT: .LBB36_5: @ %for.cond.cleanup1637; CHECK-NEXT: add sp, #81638; CHECK-NEXT: vpop {d8, d9, d10, d11}1639; CHECK-NEXT: add sp, #41640; CHECK-NEXT: pop.w {r4, r5, r6, r7, r8, r9, r10, r11, pc}1641entry:1642 %conv = sext i16 %iHeight to i321643 %conv9 = sext i16 %iSourceStride to i321644 %conv11 = sext i16 %iTargetStride to i321645 %cmp37 = icmp sgt i16 %iHeight, 01646 br i1 %cmp37, label %for.body.lr.ph, label %for.cond.cleanup1647 1648for.body.lr.ph: ; preds = %entry1649 %conv2 = sext i16 %iWidth to i321650 br label %for.body1651 1652for.cond.cleanup: ; preds = %do.end, %entry1653 ret void1654 1655for.body: ; preds = %for.body.lr.ph, %do.end1656 %pwSourceBase.addr.040 = phi ptr [ %pwSourceBase, %for.body.lr.ph ], [ %add.ptr10, %do.end ]1657 %phwTargetBase.addr.039 = phi ptr [ %phwTargetBase, %for.body.lr.ph ], [ %add.ptr12, %do.end ]1658 %y.038 = phi i32 [ 0, %for.body.lr.ph ], [ %inc, %do.end ]1659 br label %do.body1660 1661do.body: ; preds = %do.body, %for.body1662 %pTarget.0 = phi ptr [ %phwTargetBase.addr.039, %for.body ], [ %add.ptr6, %do.body ]1663 %pSource.0 = phi ptr [ %pwSourceBase.addr.040, %for.body ], [ %add.ptr, %do.body ]1664 %blkCnt.0 = phi i32 [ %conv2, %for.body ], [ %sub, %do.body ]1665 %l2 = tail call <4 x i1> @llvm.arm.mve.vctp32(i32 %blkCnt.0)1666 %l3 = bitcast ptr %pSource.0 to ptr1667 %l4 = tail call <4 x i32> @llvm.masked.load.v4i32.p0(ptr %l3, i32 4, <4 x i1> %l2, <4 x i32> zeroinitializer)1668 %l5 = tail call <4 x i32> @llvm.arm.mve.qdmulh.predicated.v4i32.v4i1(<4 x i32> %l4, <4 x i32> <i32 268435456, i32 268435456, i32 268435456, i32 268435456>, <4 x i1> %l2, <4 x i32> undef)1669 %and = and <4 x i32> %l5, <i32 31, i32 31, i32 31, i32 31>1670 %l6 = tail call <4 x i32> @llvm.arm.mve.qdmulh.predicated.v4i32.v4i1(<4 x i32> %l4, <4 x i32> <i32 67108864, i32 67108864, i32 67108864, i32 67108864>, <4 x i1> %l2, <4 x i32> undef)1671 %and3 = and <4 x i32> %l6, <i32 2016, i32 2016, i32 2016, i32 2016>1672 %l7 = tail call <4 x i32> @llvm.arm.mve.qdmulh.predicated.v4i32.v4i1(<4 x i32> %l4, <4 x i32> <i32 8388608, i32 8388608, i32 8388608, i32 8388608>, <4 x i1> %l2, <4 x i32> undef)1673 %and4 = and <4 x i32> %l7, <i32 63488, i32 63488, i32 63488, i32 63488>1674 %or = or <4 x i32> %and3, %and1675 %or5 = or <4 x i32> %or, %and41676 %l8 = trunc <4 x i32> %or5 to <4 x i16>1677 %l9 = bitcast ptr %pTarget.0 to ptr1678 tail call void @llvm.masked.store.v4i16.p0(<4 x i16> %l8, ptr %l9, i32 2, <4 x i1> %l2)1679 %add.ptr = getelementptr inbounds i32, ptr %pSource.0, i32 41680 %add.ptr6 = getelementptr inbounds i16, ptr %pTarget.0, i32 41681 %sub = add nsw i32 %blkCnt.0, -41682 %cmp7 = icmp sgt i32 %blkCnt.0, 41683 br i1 %cmp7, label %do.body, label %do.end1684 1685do.end: ; preds = %do.body1686 %add.ptr10 = getelementptr inbounds i32, ptr %pwSourceBase.addr.040, i32 %conv91687 %add.ptr12 = getelementptr inbounds i16, ptr %phwTargetBase.addr.039, i32 %conv111688 %inc = add nuw nsw i32 %y.038, 11689 %exitcond.not = icmp eq i32 %inc, %conv1690 br i1 %exitcond.not, label %for.cond.cleanup, label %for.body1691}1692 1693declare <4 x i1> @llvm.arm.mve.vctp32(i32)1694declare <4 x i16> @llvm.masked.load.v4i16.p0(ptr, i32 immarg, <4 x i1>, <4 x i16>)1695declare <4 x i32> @llvm.masked.load.v4i32.p0(ptr, i32 immarg, <4 x i1>, <4 x i32>)1696declare <4 x float> @llvm.masked.load.v4f32.p0(ptr, i32 immarg, <4 x i1>, <4 x float>)1697declare void @llvm.masked.store.v4i32.p0(<4 x i32>, ptr, i32 immarg, <4 x i1>)1698declare void @llvm.masked.store.v4f32.p0(<4 x float>, ptr, i32 immarg, <4 x i1>)1699declare void @llvm.masked.store.v4i16.p0(<4 x i16>, ptr, i32 immarg, <4 x i1>) #31700 1701declare <4 x i32> @llvm.arm.mve.add.predicated.v4i32.v4i1(<4 x i32>, <4 x i32>, <4 x i1>, <4 x i32>)1702declare <4 x i32> @llvm.arm.mve.sub.predicated.v4i32.v4i1(<4 x i32>, <4 x i32>, <4 x i1>, <4 x i32>)1703declare <4 x i32> @llvm.arm.mve.mul.predicated.v4i32.v4i1(<4 x i32>, <4 x i32>, <4 x i1>, <4 x i32>)1704declare <4 x i32> @llvm.arm.mve.qadd.predicated.v4i32.v4i1(<4 x i32>, <4 x i32>, i32, <4 x i1>, <4 x i32>)1705declare <4 x i32> @llvm.sadd.sat.v4i32(<4 x i32>, <4 x i32>)1706declare <4 x i32> @llvm.uadd.sat.v4i32(<4 x i32>, <4 x i32>)1707declare <4 x i32> @llvm.arm.mve.qsub.predicated.v4i32.v4i1(<4 x i32>, <4 x i32>, i32, <4 x i1>, <4 x i32>)1708declare <4 x i32> @llvm.ssub.sat.v4i32(<4 x i32>, <4 x i32>)1709declare <4 x i32> @llvm.usub.sat.v4i32(<4 x i32>, <4 x i32>)1710declare <4 x i32> @llvm.arm.mve.hadd.predicated.v4i32.v4i1(<4 x i32>, <4 x i32>, i32, <4 x i1>, <4 x i32>)1711declare <4 x i32> @llvm.arm.mve.vhadd.v4i32(<4 x i32>, <4 x i32>, i32)1712declare <4 x i32> @llvm.arm.mve.hsub.predicated.v4i32.v4i1(<4 x i32>, <4 x i32>, i32, <4 x i1>, <4 x i32>)1713declare <4 x i32> @llvm.arm.mve.vhsub.v4i32(<4 x i32>, <4 x i32>, i32)1714declare <2 x i64> @llvm.arm.mve.vqdmull.v2i64.v4i32(<4 x i32>, <4 x i32>, i32) #11715declare <4 x i32> @llvm.arm.mve.vqdmull.predicated.v4i32.v8i16.v4i1(<8 x i16>, <8 x i16>, i32, <4 x i1>, <4 x i32>)1716declare <4 x i32> @llvm.arm.mve.qdmulh.predicated.v4i32.v4i1(<4 x i32>, <4 x i32>, <4 x i1>, <4 x i32>)1717declare <4 x i32> @llvm.arm.mve.vqdmulh.v4i32(<4 x i32>, <4 x i32>)1718declare <4 x i32> @llvm.arm.mve.qrdmulh.predicated.v4i32.v4i1(<4 x i32>, <4 x i32>, <4 x i1>, <4 x i32>)1719declare <4 x i32> @llvm.arm.mve.vqrdmulh.v4i32(<4 x i32>, <4 x i32>)1720declare <4 x i32> @llvm.arm.mve.vmla.n.predicated.v4i32.v4i1(<4 x i32>, <4 x i32>, i32, <4 x i1>)1721declare <4 x i32> @llvm.arm.mve.vmlas.n.predicated.v4i32.v4i1(<4 x i32>, <4 x i32>, i32, <4 x i1>)1722declare <4 x float> @llvm.arm.mve.add.predicated.v4f32.v4i1(<4 x float>, <4 x float>, <4 x i1>, <4 x float>)1723declare <4 x float> @llvm.arm.mve.sub.predicated.v4f32.v4i1(<4 x float>, <4 x float>, <4 x i1>, <4 x float>)1724declare <4 x float> @llvm.arm.mve.mul.predicated.v4f32.v4i1(<4 x float>, <4 x float>, <4 x i1>, <4 x float>)1725declare <4 x float> @llvm.arm.mve.fma.predicated.v4f32.v4i1(<4 x float>, <4 x float>, <4 x float>, <4 x i1>)1726declare <4 x float> @llvm.fma.v4f32(<4 x float>, <4 x float>, <4 x float>)1727