501 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py2; RUN: llc -mtriple=thumbv8.1m.main -mattr=+mve.fp,+fp-armv8d16sp,+fp16,+fullfp16 -tail-predication=enabled %s -o - | FileCheck %s3 4define arm_aapcs_vfpcc void @fast_float_mul(ptr nocapture %a, ptr nocapture readonly %b, ptr nocapture readonly %c, i32 %N) {5; CHECK-LABEL: fast_float_mul:6; CHECK: @ %bb.0: @ %entry7; CHECK-NEXT: push {r4, r5, r6, r7, lr}8; CHECK-NEXT: cmp r3, #09; CHECK-NEXT: beq.w .LBB0_1110; CHECK-NEXT: @ %bb.1: @ %vector.memcheck11; CHECK-NEXT: add.w r4, r2, r3, lsl #212; CHECK-NEXT: cmp r4, r013; CHECK-NEXT: add.w r4, r0, r3, lsl #214; CHECK-NEXT: cset r12, hi15; CHECK-NEXT: cmp r4, r216; CHECK-NEXT: csel r12, zr, r12, ls17; CHECK-NEXT: cmp r4, r118; CHECK-NEXT: add.w r4, r1, r3, lsl #219; CHECK-NEXT: cset lr, hi20; CHECK-NEXT: cmp r4, r021; CHECK-NEXT: cset r4, hi22; CHECK-NEXT: tst.w r4, lr23; CHECK-NEXT: it eq24; CHECK-NEXT: cmpeq.w r12, #025; CHECK-NEXT: beq .LBB0_426; CHECK-NEXT: @ %bb.2: @ %for.body.preheader27; CHECK-NEXT: and r12, r3, #328; CHECK-NEXT: subs r4, r3, #129; CHECK-NEXT: cmp r4, #330; CHECK-NEXT: bhs .LBB0_631; CHECK-NEXT: @ %bb.3:32; CHECK-NEXT: movs r3, #033; CHECK-NEXT: b .LBB0_834; CHECK-NEXT: .LBB0_4: @ %vector.ph35; CHECK-NEXT: dlstp.32 lr, r336; CHECK-NEXT: .LBB0_5: @ %vector.body37; CHECK-NEXT: @ =>This Inner Loop Header: Depth=138; CHECK-NEXT: vldrw.u32 q0, [r1], #1639; CHECK-NEXT: vldrw.u32 q1, [r2], #1640; CHECK-NEXT: vmul.f32 q0, q1, q041; CHECK-NEXT: vstrw.32 q0, [r0], #1642; CHECK-NEXT: letp lr, .LBB0_543; CHECK-NEXT: b .LBB0_1144; CHECK-NEXT: .LBB0_6: @ %for.body.preheader.new45; CHECK-NEXT: bic r3, r3, #346; CHECK-NEXT: movs r4, #147; CHECK-NEXT: subs r3, #448; CHECK-NEXT: add.w lr, r4, r3, lsr #249; CHECK-NEXT: movs r4, #050; CHECK-NEXT: movs r3, #051; CHECK-NEXT: .LBB0_7: @ %for.body52; CHECK-NEXT: @ =>This Inner Loop Header: Depth=153; CHECK-NEXT: adds r5, r1, r454; CHECK-NEXT: adds r6, r2, r455; CHECK-NEXT: adds r7, r0, r456; CHECK-NEXT: adds r4, #1657; CHECK-NEXT: vldr s0, [r5]58; CHECK-NEXT: adds r3, #459; CHECK-NEXT: vldr s2, [r6]60; CHECK-NEXT: vmul.f32 s0, s2, s061; CHECK-NEXT: vstr s0, [r7]62; CHECK-NEXT: vldr s0, [r5, #4]63; CHECK-NEXT: vldr s2, [r6, #4]64; CHECK-NEXT: vmul.f32 s0, s2, s065; CHECK-NEXT: vstr s0, [r7, #4]66; CHECK-NEXT: vldr s0, [r5, #8]67; CHECK-NEXT: vldr s2, [r6, #8]68; CHECK-NEXT: vmul.f32 s0, s2, s069; CHECK-NEXT: vstr s0, [r7, #8]70; CHECK-NEXT: vldr s0, [r5, #12]71; CHECK-NEXT: vldr s2, [r6, #12]72; CHECK-NEXT: vmul.f32 s0, s2, s073; CHECK-NEXT: vstr s0, [r7, #12]74; CHECK-NEXT: le lr, .LBB0_775; CHECK-NEXT: .LBB0_8: @ %for.cond.cleanup.loopexit.unr-lcssa76; CHECK-NEXT: wls lr, r12, .LBB0_1177; CHECK-NEXT: @ %bb.9: @ %for.body.epil.preheader78; CHECK-NEXT: add.w r1, r1, r3, lsl #279; CHECK-NEXT: add.w r2, r2, r3, lsl #280; CHECK-NEXT: add.w r0, r0, r3, lsl #281; CHECK-NEXT: .LBB0_10: @ %for.body.epil82; CHECK-NEXT: @ =>This Inner Loop Header: Depth=183; CHECK-NEXT: vldmia r1!, {s0}84; CHECK-NEXT: vldmia r2!, {s2}85; CHECK-NEXT: vmul.f32 s0, s2, s086; CHECK-NEXT: vstmia r0!, {s0}87; CHECK-NEXT: le lr, .LBB0_1088; CHECK-NEXT: .LBB0_11: @ %for.cond.cleanup89; CHECK-NEXT: pop {r4, r5, r6, r7, pc}90entry:91 %cmp8 = icmp eq i32 %N, 092 br i1 %cmp8, label %for.cond.cleanup, label %vector.memcheck93 94vector.memcheck: ; preds = %entry95 %scevgep = getelementptr float, ptr %a, i32 %N96 %scevgep13 = getelementptr float, ptr %b, i32 %N97 %scevgep16 = getelementptr float, ptr %c, i32 %N98 %bound0 = icmp ugt ptr %scevgep13, %a99 %bound1 = icmp ugt ptr %scevgep, %b100 %found.conflict = and i1 %bound0, %bound1101 %bound018 = icmp ugt ptr %scevgep16, %a102 %bound119 = icmp ugt ptr %scevgep, %c103 %found.conflict20 = and i1 %bound018, %bound119104 %conflict.rdx = or i1 %found.conflict, %found.conflict20105 br i1 %conflict.rdx, label %for.body.preheader, label %vector.ph106 107for.body.preheader: ; preds = %vector.memcheck108 %i = add i32 %N, -1109 %xtraiter = and i32 %N, 3110 %i1 = icmp ult i32 %i, 3111 br i1 %i1, label %for.cond.cleanup.loopexit.unr-lcssa, label %for.body.preheader.new112 113for.body.preheader.new: ; preds = %for.body.preheader114 %unroll_iter = sub i32 %N, %xtraiter115 br label %for.body116 117vector.ph: ; preds = %vector.memcheck118 %n.rnd.up = add i32 %N, 3119 %n.vec = and i32 %n.rnd.up, -4120 br label %vector.body121 122vector.body: ; preds = %vector.body, %vector.ph123 %index = phi i32 [ 0, %vector.ph ], [ %index.next, %vector.body ]124 %i2 = getelementptr inbounds float, ptr %b, i32 %index125 %i3 = call <4 x i1> @llvm.get.active.lane.mask.v4i1.i32(i32 %index, i32 %N)126 %wide.masked.load = call <4 x float> @llvm.masked.load.v4f32.p0(ptr %i2, i32 4, <4 x i1> %i3, <4 x float> undef)127 %i5 = getelementptr inbounds float, ptr %c, i32 %index128 %wide.masked.load23 = call <4 x float> @llvm.masked.load.v4f32.p0(ptr %i5, i32 4, <4 x i1> %i3, <4 x float> undef)129 %i7 = fmul fast <4 x float> %wide.masked.load23, %wide.masked.load130 %i8 = getelementptr inbounds float, ptr %a, i32 %index131 call void @llvm.masked.store.v4f32.p0(<4 x float> %i7, ptr %i8, i32 4, <4 x i1> %i3)132 %index.next = add i32 %index, 4133 %i10 = icmp eq i32 %index.next, %n.vec134 br i1 %i10, label %for.cond.cleanup, label %vector.body135 136for.cond.cleanup.loopexit.unr-lcssa: ; preds = %for.body, %for.body.preheader137 %i.09.unr = phi i32 [ 0, %for.body.preheader ], [ %inc.3, %for.body ]138 %lcmp.mod = icmp eq i32 %xtraiter, 0139 br i1 %lcmp.mod, label %for.cond.cleanup, label %for.body.epil140 141for.body.epil: ; preds = %for.body.epil, %for.cond.cleanup.loopexit.unr-lcssa142 %i.09.epil = phi i32 [ %inc.epil, %for.body.epil ], [ %i.09.unr, %for.cond.cleanup.loopexit.unr-lcssa ]143 %epil.iter = phi i32 [ %epil.iter.sub, %for.body.epil ], [ %xtraiter, %for.cond.cleanup.loopexit.unr-lcssa ]144 %arrayidx.epil = getelementptr inbounds float, ptr %b, i32 %i.09.epil145 %i11 = load float, ptr %arrayidx.epil, align 4146 %arrayidx1.epil = getelementptr inbounds float, ptr %c, i32 %i.09.epil147 %i12 = load float, ptr %arrayidx1.epil, align 4148 %mul.epil = fmul fast float %i12, %i11149 %arrayidx2.epil = getelementptr inbounds float, ptr %a, i32 %i.09.epil150 store float %mul.epil, ptr %arrayidx2.epil, align 4151 %inc.epil = add nuw i32 %i.09.epil, 1152 %epil.iter.sub = add i32 %epil.iter, -1153 %epil.iter.cmp = icmp eq i32 %epil.iter.sub, 0154 br i1 %epil.iter.cmp, label %for.cond.cleanup, label %for.body.epil155 156for.cond.cleanup: ; preds = %for.body.epil, %for.cond.cleanup.loopexit.unr-lcssa, %vector.body, %entry157 ret void158 159for.body: ; preds = %for.body, %for.body.preheader.new160 %i.09 = phi i32 [ 0, %for.body.preheader.new ], [ %inc.3, %for.body ]161 %niter = phi i32 [ %unroll_iter, %for.body.preheader.new ], [ %niter.nsub.3, %for.body ]162 %arrayidx = getelementptr inbounds float, ptr %b, i32 %i.09163 %i13 = load float, ptr %arrayidx, align 4164 %arrayidx1 = getelementptr inbounds float, ptr %c, i32 %i.09165 %i14 = load float, ptr %arrayidx1, align 4166 %mul = fmul fast float %i14, %i13167 %arrayidx2 = getelementptr inbounds float, ptr %a, i32 %i.09168 store float %mul, ptr %arrayidx2, align 4169 %inc = or disjoint i32 %i.09, 1170 %arrayidx.1 = getelementptr inbounds float, ptr %b, i32 %inc171 %i15 = load float, ptr %arrayidx.1, align 4172 %arrayidx1.1 = getelementptr inbounds float, ptr %c, i32 %inc173 %i16 = load float, ptr %arrayidx1.1, align 4174 %mul.1 = fmul fast float %i16, %i15175 %arrayidx2.1 = getelementptr inbounds float, ptr %a, i32 %inc176 store float %mul.1, ptr %arrayidx2.1, align 4177 %inc.1 = or disjoint i32 %i.09, 2178 %arrayidx.2 = getelementptr inbounds float, ptr %b, i32 %inc.1179 %i17 = load float, ptr %arrayidx.2, align 4180 %arrayidx1.2 = getelementptr inbounds float, ptr %c, i32 %inc.1181 %i18 = load float, ptr %arrayidx1.2, align 4182 %mul.2 = fmul fast float %i18, %i17183 %arrayidx2.2 = getelementptr inbounds float, ptr %a, i32 %inc.1184 store float %mul.2, ptr %arrayidx2.2, align 4185 %inc.2 = or disjoint i32 %i.09, 3186 %arrayidx.3 = getelementptr inbounds float, ptr %b, i32 %inc.2187 %i19 = load float, ptr %arrayidx.3, align 4188 %arrayidx1.3 = getelementptr inbounds float, ptr %c, i32 %inc.2189 %i20 = load float, ptr %arrayidx1.3, align 4190 %mul.3 = fmul fast float %i20, %i19191 %arrayidx2.3 = getelementptr inbounds float, ptr %a, i32 %inc.2192 store float %mul.3, ptr %arrayidx2.3, align 4193 %inc.3 = add nuw i32 %i.09, 4194 %niter.nsub.3 = add i32 %niter, -4195 %niter.ncmp.3 = icmp eq i32 %niter.nsub.3, 0196 br i1 %niter.ncmp.3, label %for.cond.cleanup.loopexit.unr-lcssa, label %for.body197}198 199define arm_aapcs_vfpcc float @fast_float_mac(ptr nocapture readonly %b, ptr nocapture readonly %c, i32 %N) {200; CHECK-LABEL: fast_float_mac:201; CHECK: @ %bb.0: @ %entry202; CHECK-NEXT: cmp r2, #0203; CHECK-NEXT: itt eq204; CHECK-NEXT: vldreq s0, .LCPI1_0205; CHECK-NEXT: bxeq lr206; CHECK-NEXT: .LBB1_1: @ %vector.ph207; CHECK-NEXT: push {r7, lr}208; CHECK-NEXT: adds r3, r2, #3209; CHECK-NEXT: mov.w r12, #1210; CHECK-NEXT: bic r3, r3, #3211; CHECK-NEXT: vmov.i32 q0, #0x0212; CHECK-NEXT: subs r3, #4213; CHECK-NEXT: add.w r3, r12, r3, lsr #2214; CHECK-NEXT: dls lr, r3215; CHECK-NEXT: .LBB1_2: @ %vector.body216; CHECK-NEXT: @ =>This Inner Loop Header: Depth=1217; CHECK-NEXT: vctp.32 r2218; CHECK-NEXT: subs r2, #4219; CHECK-NEXT: vpstt220; CHECK-NEXT: vldrwt.u32 q2, [r0], #16221; CHECK-NEXT: vldrwt.u32 q3, [r1], #16222; CHECK-NEXT: vmov q1, q0223; CHECK-NEXT: vfma.f32 q0, q3, q2224; CHECK-NEXT: le lr, .LBB1_2225; CHECK-NEXT: @ %bb.3: @ %middle.block226; CHECK-NEXT: vpsel q0, q0, q1227; CHECK-NEXT: vmov.f32 s4, s2228; CHECK-NEXT: vmov.f32 s5, s3229; CHECK-NEXT: vadd.f32 q0, q0, q1230; CHECK-NEXT: vmov r0, s1231; CHECK-NEXT: vadd.f32 q0, q0, r0232; CHECK-NEXT: pop {r7, pc}233; CHECK-NEXT: .p2align 2234; CHECK-NEXT: @ %bb.4:235; CHECK-NEXT: .LCPI1_0:236; CHECK-NEXT: .long 0x00000000 @ float 0237entry:238 %cmp8 = icmp eq i32 %N, 0239 br i1 %cmp8, label %for.cond.cleanup, label %vector.ph240 241vector.ph: ; preds = %entry242 %n.rnd.up = add i32 %N, 3243 %n.vec = and i32 %n.rnd.up, -4244 br label %vector.body245 246vector.body: ; preds = %vector.body, %vector.ph247 %index = phi i32 [ 0, %vector.ph ], [ %index.next, %vector.body ]248 %vec.phi = phi <4 x float> [ zeroinitializer, %vector.ph ], [ %i6, %vector.body ]249 %i = getelementptr inbounds float, ptr %b, i32 %index250 %i1 = call <4 x i1> @llvm.get.active.lane.mask.v4i1.i32(i32 %index, i32 %N)251 %wide.masked.load = call <4 x float> @llvm.masked.load.v4f32.p0(ptr %i, i32 4, <4 x i1> %i1, <4 x float> undef)252 %i3 = getelementptr inbounds float, ptr %c, i32 %index253 %wide.masked.load13 = call <4 x float> @llvm.masked.load.v4f32.p0(ptr %i3, i32 4, <4 x i1> %i1, <4 x float> undef)254 %i5 = fmul fast <4 x float> %wide.masked.load13, %wide.masked.load255 %i6 = fadd fast <4 x float> %i5, %vec.phi256 %index.next = add i32 %index, 4257 %i7 = icmp eq i32 %index.next, %n.vec258 br i1 %i7, label %middle.block, label %vector.body259 260middle.block: ; preds = %vector.body261 %i8 = select <4 x i1> %i1, <4 x float> %i6, <4 x float> %vec.phi262 %rdx.shuf = shufflevector <4 x float> %i8, <4 x float> undef, <4 x i32> <i32 2, i32 3, i32 undef, i32 undef>263 %bin.rdx = fadd fast <4 x float> %i8, %rdx.shuf264 %rdx.shuf14 = shufflevector <4 x float> %bin.rdx, <4 x float> undef, <4 x i32> <i32 1, i32 undef, i32 undef, i32 undef>265 %bin.rdx15 = fadd fast <4 x float> %bin.rdx, %rdx.shuf14266 %i9 = extractelement <4 x float> %bin.rdx15, i32 0267 br label %for.cond.cleanup268 269for.cond.cleanup: ; preds = %middle.block, %entry270 %a.0.lcssa = phi float [ 0.000000e+00, %entry ], [ %i9, %middle.block ]271 ret float %a.0.lcssa272}273 274define arm_aapcs_vfpcc float @fast_float_half_mac(ptr nocapture readonly %b, ptr nocapture readonly %c, i32 %N) {275; CHECK-LABEL: fast_float_half_mac:276; CHECK: @ %bb.0: @ %entry277; CHECK-NEXT: cmp r2, #0278; CHECK-NEXT: itt eq279; CHECK-NEXT: vldreq s0, .LCPI2_0280; CHECK-NEXT: bxeq lr281; CHECK-NEXT: .LBB2_1: @ %vector.ph282; CHECK-NEXT: push {r4, r5, r7, lr}283; CHECK-NEXT: vpush {d8, d9, d10, d11, d12, d13}284; CHECK-NEXT: sub sp, #8285; CHECK-NEXT: adds r3, r2, #3286; CHECK-NEXT: vmov.i32 q5, #0x0287; CHECK-NEXT: bic r3, r3, #3288; CHECK-NEXT: sub.w r12, r3, #4289; CHECK-NEXT: movs r3, #1290; CHECK-NEXT: add.w lr, r3, r12, lsr #2291; CHECK-NEXT: sub.w r12, r2, #1292; CHECK-NEXT: adr r2, .LCPI2_1293; CHECK-NEXT: mov lr, lr294; CHECK-NEXT: vldrw.u32 q0, [r2]295; CHECK-NEXT: movs r3, #0296; CHECK-NEXT: vdup.32 q1, r12297; CHECK-NEXT: vdup.32 q2, r12298; CHECK-NEXT: b .LBB2_3299; CHECK-NEXT: .LBB2_2: @ %else24300; CHECK-NEXT: @ in Loop: Header=BB2_3 Depth=1301; CHECK-NEXT: vmul.f16 q5, q6, q5302; CHECK-NEXT: adds r0, #8303; CHECK-NEXT: vcvtt.f32.f16 s23, s21304; CHECK-NEXT: vcvtb.f32.f16 s22, s21305; CHECK-NEXT: vcvtt.f32.f16 s21, s20306; CHECK-NEXT: vcvtb.f32.f16 s20, s20307; CHECK-NEXT: adds r1, #8308; CHECK-NEXT: adds r3, #4309; CHECK-NEXT: vadd.f32 q5, q3, q5310; CHECK-NEXT: subs.w lr, lr, #1311; CHECK-NEXT: bne .LBB2_3312; CHECK-NEXT: b .LBB2_19313; CHECK-NEXT: .LBB2_3: @ %vector.body314; CHECK-NEXT: @ =>This Inner Loop Header: Depth=1315; CHECK-NEXT: vadd.i32 q4, q0, r3316; CHECK-NEXT: vmov q3, q5317; CHECK-NEXT: vcmp.u32 cs, q1, q4318; CHECK-NEXT: @ implicit-def: $q5319; CHECK-NEXT: vmrs r4, p0320; CHECK-NEXT: and r2, r4, #1321; CHECK-NEXT: rsbs r5, r2, #0322; CHECK-NEXT: movs r2, #0323; CHECK-NEXT: bfi r2, r5, #0, #1324; CHECK-NEXT: ubfx r5, r4, #4, #1325; CHECK-NEXT: rsbs r5, r5, #0326; CHECK-NEXT: bfi r2, r5, #1, #1327; CHECK-NEXT: ubfx r5, r4, #8, #1328; CHECK-NEXT: ubfx r4, r4, #12, #1329; CHECK-NEXT: rsbs r5, r5, #0330; CHECK-NEXT: bfi r2, r5, #2, #1331; CHECK-NEXT: rsbs r4, r4, #0332; CHECK-NEXT: bfi r2, r4, #3, #1333; CHECK-NEXT: lsls r4, r2, #31334; CHECK-NEXT: bne .LBB2_12335; CHECK-NEXT: @ %bb.4: @ %else336; CHECK-NEXT: @ in Loop: Header=BB2_3 Depth=1337; CHECK-NEXT: lsls r4, r2, #30338; CHECK-NEXT: bmi .LBB2_13339; CHECK-NEXT: .LBB2_5: @ %else5340; CHECK-NEXT: @ in Loop: Header=BB2_3 Depth=1341; CHECK-NEXT: lsls r4, r2, #29342; CHECK-NEXT: bmi .LBB2_14343; CHECK-NEXT: .LBB2_6: @ %else8344; CHECK-NEXT: @ in Loop: Header=BB2_3 Depth=1345; CHECK-NEXT: lsls r2, r2, #28346; CHECK-NEXT: bpl .LBB2_8347; CHECK-NEXT: .LBB2_7: @ %cond.load10348; CHECK-NEXT: @ in Loop: Header=BB2_3 Depth=1349; CHECK-NEXT: vldr.16 s22, [r0, #6]350; CHECK-NEXT: vins.f16 s21, s22351; CHECK-NEXT: .LBB2_8: @ %else11352; CHECK-NEXT: @ in Loop: Header=BB2_3 Depth=1353; CHECK-NEXT: vcmp.u32 cs, q2, q4354; CHECK-NEXT: @ implicit-def: $q6355; CHECK-NEXT: vmrs r4, p0356; CHECK-NEXT: and r2, r4, #1357; CHECK-NEXT: rsbs r5, r2, #0358; CHECK-NEXT: movs r2, #0359; CHECK-NEXT: bfi r2, r5, #0, #1360; CHECK-NEXT: ubfx r5, r4, #4, #1361; CHECK-NEXT: rsbs r5, r5, #0362; CHECK-NEXT: bfi r2, r5, #1, #1363; CHECK-NEXT: ubfx r5, r4, #8, #1364; CHECK-NEXT: ubfx r4, r4, #12, #1365; CHECK-NEXT: rsbs r5, r5, #0366; CHECK-NEXT: bfi r2, r5, #2, #1367; CHECK-NEXT: rsbs r4, r4, #0368; CHECK-NEXT: bfi r2, r4, #3, #1369; CHECK-NEXT: lsls r4, r2, #31370; CHECK-NEXT: bne .LBB2_15371; CHECK-NEXT: @ %bb.9: @ %else15372; CHECK-NEXT: @ in Loop: Header=BB2_3 Depth=1373; CHECK-NEXT: lsls r4, r2, #30374; CHECK-NEXT: bmi .LBB2_16375; CHECK-NEXT: .LBB2_10: @ %else18376; CHECK-NEXT: @ in Loop: Header=BB2_3 Depth=1377; CHECK-NEXT: lsls r4, r2, #29378; CHECK-NEXT: bmi .LBB2_17379; CHECK-NEXT: .LBB2_11: @ %else21380; CHECK-NEXT: @ in Loop: Header=BB2_3 Depth=1381; CHECK-NEXT: lsls r2, r2, #28382; CHECK-NEXT: bpl .LBB2_2383; CHECK-NEXT: b .LBB2_18384; CHECK-NEXT: .LBB2_12: @ %cond.load385; CHECK-NEXT: @ in Loop: Header=BB2_3 Depth=1386; CHECK-NEXT: vldr.16 s20, [r0]387; CHECK-NEXT: lsls r4, r2, #30388; CHECK-NEXT: bpl .LBB2_5389; CHECK-NEXT: .LBB2_13: @ %cond.load4390; CHECK-NEXT: @ in Loop: Header=BB2_3 Depth=1391; CHECK-NEXT: vldr.16 s22, [r0, #2]392; CHECK-NEXT: vins.f16 s20, s22393; CHECK-NEXT: lsls r4, r2, #29394; CHECK-NEXT: bpl .LBB2_6395; CHECK-NEXT: .LBB2_14: @ %cond.load7396; CHECK-NEXT: @ in Loop: Header=BB2_3 Depth=1397; CHECK-NEXT: vldr.16 s21, [r0, #4]398; CHECK-NEXT: vmovx.f16 s22, s0399; CHECK-NEXT: vins.f16 s21, s22400; CHECK-NEXT: lsls r2, r2, #28401; CHECK-NEXT: bmi .LBB2_7402; CHECK-NEXT: b .LBB2_8403; CHECK-NEXT: .LBB2_15: @ %cond.load14404; CHECK-NEXT: @ in Loop: Header=BB2_3 Depth=1405; CHECK-NEXT: vldr.16 s24, [r1]406; CHECK-NEXT: lsls r4, r2, #30407; CHECK-NEXT: bpl .LBB2_10408; CHECK-NEXT: .LBB2_16: @ %cond.load17409; CHECK-NEXT: @ in Loop: Header=BB2_3 Depth=1410; CHECK-NEXT: vldr.16 s26, [r1, #2]411; CHECK-NEXT: vins.f16 s24, s26412; CHECK-NEXT: lsls r4, r2, #29413; CHECK-NEXT: bpl .LBB2_11414; CHECK-NEXT: .LBB2_17: @ %cond.load20415; CHECK-NEXT: @ in Loop: Header=BB2_3 Depth=1416; CHECK-NEXT: vldr.16 s25, [r1, #4]417; CHECK-NEXT: vmovx.f16 s26, s0418; CHECK-NEXT: vins.f16 s25, s26419; CHECK-NEXT: lsls r2, r2, #28420; CHECK-NEXT: bpl.w .LBB2_2421; CHECK-NEXT: .LBB2_18: @ %cond.load23422; CHECK-NEXT: @ in Loop: Header=BB2_3 Depth=1423; CHECK-NEXT: vldr.16 s26, [r1, #6]424; CHECK-NEXT: vins.f16 s25, s26425; CHECK-NEXT: b .LBB2_2426; CHECK-NEXT: .LBB2_19: @ %middle.block427; CHECK-NEXT: vdup.32 q0, r12428; CHECK-NEXT: vcmp.u32 cs, q0, q4429; CHECK-NEXT: vpsel q0, q5, q3430; CHECK-NEXT: vmov.f32 s4, s2431; CHECK-NEXT: vmov.f32 s5, s3432; CHECK-NEXT: vadd.f32 q0, q0, q1433; CHECK-NEXT: vmov r0, s1434; CHECK-NEXT: vadd.f32 q0, q0, r0435; CHECK-NEXT: add sp, #8436; CHECK-NEXT: vpop {d8, d9, d10, d11, d12, d13}437; CHECK-NEXT: pop {r4, r5, r7, pc}438; CHECK-NEXT: .p2align 4439; CHECK-NEXT: @ %bb.20:440; CHECK-NEXT: .LCPI2_1:441; CHECK-NEXT: .long 0 @ 0x0442; CHECK-NEXT: .long 1 @ 0x1443; CHECK-NEXT: .long 2 @ 0x2444; CHECK-NEXT: .long 3 @ 0x3445; CHECK-NEXT: .LCPI2_0:446; CHECK-NEXT: .long 0x00000000 @ float 0447entry:448 %cmp8 = icmp eq i32 %N, 0449 br i1 %cmp8, label %for.cond.cleanup, label %vector.ph450 451vector.ph: ; preds = %entry452 %n.rnd.up = add i32 %N, 3453 %n.vec = and i32 %n.rnd.up, -4454 %trip.count.minus.1 = add i32 %N, -1455 %broadcast.splatinsert11 = insertelement <4 x i32> undef, i32 %trip.count.minus.1, i32 0456 %broadcast.splat12 = shufflevector <4 x i32> %broadcast.splatinsert11, <4 x i32> undef, <4 x i32> zeroinitializer457 br label %vector.body458 459vector.body: ; preds = %vector.body, %vector.ph460 %index = phi i32 [ 0, %vector.ph ], [ %index.next, %vector.body ]461 %vec.phi = phi <4 x float> [ zeroinitializer, %vector.ph ], [ %i7, %vector.body ]462 %broadcast.splatinsert = insertelement <4 x i32> undef, i32 %index, i32 0463 %broadcast.splat = shufflevector <4 x i32> %broadcast.splatinsert, <4 x i32> undef, <4 x i32> zeroinitializer464 %induction = add <4 x i32> %broadcast.splat, <i32 0, i32 1, i32 2, i32 3>465 %i = getelementptr inbounds half, ptr %b, i32 %index466 %i1 = icmp ule <4 x i32> %induction, %broadcast.splat12467 %wide.masked.load = call <4 x half> @llvm.masked.load.v4f16.p0(ptr %i, i32 2, <4 x i1> %i1, <4 x half> undef)468 %i3 = getelementptr inbounds half, ptr %c, i32 %index469 %wide.masked.load13 = call <4 x half> @llvm.masked.load.v4f16.p0(ptr %i3, i32 2, <4 x i1> %i1, <4 x half> undef)470 %i5 = fmul fast <4 x half> %wide.masked.load13, %wide.masked.load471 %i6 = fpext <4 x half> %i5 to <4 x float>472 %i7 = fadd fast <4 x float> %vec.phi, %i6473 %index.next = add i32 %index, 4474 %i8 = icmp eq i32 %index.next, %n.vec475 br i1 %i8, label %middle.block, label %vector.body476 477middle.block: ; preds = %vector.body478 %i9 = select <4 x i1> %i1, <4 x float> %i7, <4 x float> %vec.phi479 %rdx.shuf = shufflevector <4 x float> %i9, <4 x float> undef, <4 x i32> <i32 2, i32 3, i32 undef, i32 undef>480 %bin.rdx = fadd fast <4 x float> %i9, %rdx.shuf481 %rdx.shuf14 = shufflevector <4 x float> %bin.rdx, <4 x float> undef, <4 x i32> <i32 1, i32 undef, i32 undef, i32 undef>482 %bin.rdx15 = fadd fast <4 x float> %bin.rdx, %rdx.shuf14483 %i10 = extractelement <4 x float> %bin.rdx15, i32 0484 br label %for.cond.cleanup485 486for.cond.cleanup: ; preds = %middle.block, %entry487 %a.0.lcssa = phi float [ 0.000000e+00, %entry ], [ %i10, %middle.block ]488 ret float %a.0.lcssa489}490 491; Function Attrs: argmemonly nounwind readonly willreturn492declare <4 x float> @llvm.masked.load.v4f32.p0(ptr, i32 immarg, <4 x i1>, <4 x float>)493 494; Function Attrs: argmemonly nounwind willreturn495declare void @llvm.masked.store.v4f32.p0(<4 x float>, ptr, i32 immarg, <4 x i1>)496 497; Function Attrs: argmemonly nounwind readonly willreturn498declare <4 x half> @llvm.masked.load.v4f16.p0(ptr, i32 immarg, <4 x i1>, <4 x half>)499 500declare <4 x i1> @llvm.get.active.lane.mask.v4i1.i32(i32, i32)501