379 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 %s -o - | FileCheck %s3 4; Check some loop postinc's for properly distributed post-incs5 6define i32 @vaddv(ptr nocapture readonly %data, i32 %N) {7; CHECK-LABEL: vaddv:8; CHECK: @ %bb.0: @ %entry9; CHECK-NEXT: .save {r7, lr}10; CHECK-NEXT: push {r7, lr}11; CHECK-NEXT: mov lr, r112; CHECK-NEXT: cmp r1, #113; CHECK-NEXT: blt .LBB0_414; CHECK-NEXT: @ %bb.1: @ %for.body.preheader15; CHECK-NEXT: mov r1, r016; CHECK-NEXT: movs r0, #017; CHECK-NEXT: .LBB0_2: @ %for.body18; CHECK-NEXT: @ =>This Inner Loop Header: Depth=119; CHECK-NEXT: vldrw.u32 q0, [r1], #3220; CHECK-NEXT: vaddva.s32 r0, q021; CHECK-NEXT: vldrw.u32 q0, [r1, #-16]22; CHECK-NEXT: vaddva.s32 r0, q023; CHECK-NEXT: le lr, .LBB0_224; CHECK-NEXT: @ %bb.3: @ %for.cond.cleanup25; CHECK-NEXT: pop {r7, pc}26; CHECK-NEXT: .LBB0_4:27; CHECK-NEXT: movs r0, #028; CHECK-NEXT: pop {r7, pc}29entry:30 %cmp11 = icmp sgt i32 %N, 031 br i1 %cmp11, label %for.body, label %for.cond.cleanup32 33for.cond.cleanup: ; preds = %for.body, %entry34 %x.0.lcssa = phi i32 [ 0, %entry ], [ %7, %for.body ]35 ret i32 %x.0.lcssa36 37for.body: ; preds = %entry, %for.body38 %data.addr.014 = phi ptr [ %add.ptr1, %for.body ], [ %data, %entry ]39 %i.013 = phi i32 [ %inc, %for.body ], [ 0, %entry ]40 %x.012 = phi i32 [ %7, %for.body ], [ 0, %entry ]41 %0 = bitcast ptr %data.addr.014 to ptr42 %1 = load <4 x i32>, ptr %0, align 443 %2 = tail call i32 @llvm.arm.mve.addv.v4i32(<4 x i32> %1, i32 0)44 %3 = add i32 %2, %x.01245 %add.ptr = getelementptr inbounds i32, ptr %data.addr.014, i32 446 %4 = bitcast ptr %add.ptr to ptr47 %5 = load <4 x i32>, ptr %4, align 448 %6 = tail call i32 @llvm.arm.mve.addv.v4i32(<4 x i32> %5, i32 0)49 %7 = add i32 %3, %650 %add.ptr1 = getelementptr inbounds i32, ptr %data.addr.014, i32 851 %inc = add nuw nsw i32 %i.013, 152 %exitcond = icmp eq i32 %inc, %N53 br i1 %exitcond, label %for.cond.cleanup, label %for.body54}55 56define void @arm_cmplx_dot_prod_q15(ptr nocapture readonly %pSrcA, ptr nocapture readonly %pSrcB, i32 %numSamples, ptr nocapture %realResult, ptr nocapture %imagResult) {57; CHECK-LABEL: arm_cmplx_dot_prod_q15:58; CHECK: @ %bb.0: @ %entry59; CHECK-NEXT: .save {r4, r5, r6, r7, r8, r9, r10, r11, lr}60; CHECK-NEXT: push.w {r4, r5, r6, r7, r8, r9, r10, r11, lr}61; CHECK-NEXT: mvn r7, #762; CHECK-NEXT: mov.w r12, #063; CHECK-NEXT: add.w r7, r7, r2, lsl #164; CHECK-NEXT: vldrh.u16 q0, [r0]65; CHECK-NEXT: vldrh.u16 q1, [r1]66; CHECK-NEXT: movs r4, #067; CHECK-NEXT: lsr.w r9, r7, #368; CHECK-NEXT: mov r7, r1269; CHECK-NEXT: mov r11, r1270; CHECK-NEXT: wls lr, r9, .LBB1_471; CHECK-NEXT: @ %bb.1: @ %while.body.preheader72; CHECK-NEXT: add.w r8, r0, r9, lsl #573; CHECK-NEXT: mov.w r11, #074; CHECK-NEXT: adds r0, #3275; CHECK-NEXT: add.w r6, r1, #3276; CHECK-NEXT: lsl.w r9, r9, #477; CHECK-NEXT: mov r4, r1178; CHECK-NEXT: movs r7, #079; CHECK-NEXT: mov r12, r1180; CHECK-NEXT: .LBB1_2: @ %while.body81; CHECK-NEXT: @ =>This Inner Loop Header: Depth=182; CHECK-NEXT: vldrh.u16 q2, [r6, #-16]83; CHECK-NEXT: vldrh.u16 q3, [r0, #-16]84; CHECK-NEXT: vmlaldavax.s16 r4, r11, q0, q185; CHECK-NEXT: vmlsldava.s16 r12, r7, q0, q186; CHECK-NEXT: vldrh.u16 q0, [r0], #3287; CHECK-NEXT: vldrh.u16 q1, [r6], #3288; CHECK-NEXT: vmlaldavax.s16 r4, r11, q3, q289; CHECK-NEXT: vmlsldava.s16 r12, r7, q3, q290; CHECK-NEXT: le lr, .LBB1_291; CHECK-NEXT: @ %bb.3: @ %while.cond.while.end_crit_edge92; CHECK-NEXT: add.w r1, r1, r9, lsl #193; CHECK-NEXT: mov r0, r894; CHECK-NEXT: .LBB1_4: @ %while.end95; CHECK-NEXT: vmlaldavax.s16 r4, r11, q0, q196; CHECK-NEXT: vmlsldava.s16 r12, r7, q0, q197; CHECK-NEXT: mov r10, r498; CHECK-NEXT: mov r5, r1199; CHECK-NEXT: lsrl r10, r5, #6100; CHECK-NEXT: ldr.w r8, [sp, #36]101; CHECK-NEXT: mov r6, r12102; CHECK-NEXT: mov r5, r7103; CHECK-NEXT: lsrl r6, r5, #6104; CHECK-NEXT: and r2, r2, #3105; CHECK-NEXT: wls lr, r2, .LBB1_7106; CHECK-NEXT: .LBB1_5: @ %while.body11107; CHECK-NEXT: @ =>This Inner Loop Header: Depth=1108; CHECK-NEXT: ldrsh r9, [r0], #4109; CHECK-NEXT: ldrsh r6, [r1], #4110; CHECK-NEXT: ldrsh r5, [r0, #-2]111; CHECK-NEXT: ldrsh r2, [r1, #-2]112; CHECK-NEXT: smlalbb r12, r7, r6, r9113; CHECK-NEXT: smlalbb r4, r11, r6, r5114; CHECK-NEXT: muls r5, r2, r5115; CHECK-NEXT: smlalbb r4, r11, r2, r9116; CHECK-NEXT: subs.w r12, r12, r5117; CHECK-NEXT: sbc.w r7, r7, r5, asr #31118; CHECK-NEXT: le lr, .LBB1_5119; CHECK-NEXT: @ %bb.6: @ %while.end34.loopexit120; CHECK-NEXT: lsrl r12, r7, #6121; CHECK-NEXT: lsrl r4, r11, #6122; CHECK-NEXT: mov r6, r12123; CHECK-NEXT: mov r10, r4124; CHECK-NEXT: .LBB1_7: @ %while.end34125; CHECK-NEXT: str r6, [r3]126; CHECK-NEXT: str.w r10, [r8]127; CHECK-NEXT: pop.w {r4, r5, r6, r7, r8, r9, r10, r11, pc}128entry:129 %mul = shl i32 %numSamples, 1130 %sub = add i32 %mul, -8131 %shr = lshr i32 %sub, 3132 %vecSrcB.0.in102 = bitcast ptr %pSrcB to ptr133 %vecSrcB.0103 = load <8 x i16>, ptr %vecSrcB.0.in102, align 2134 %vecSrcA.0.in104 = bitcast ptr %pSrcA to ptr135 %vecSrcA.0105 = load <8 x i16>, ptr %vecSrcA.0.in104, align 2136 %cmp106 = icmp eq i32 %shr, 0137 br i1 %cmp106, label %while.end, label %while.body.preheader138 139while.body.preheader: ; preds = %entry140 %0 = shl i32 %shr, 4141 %scevgep = getelementptr i16, ptr %pSrcA, i32 %0142 br label %while.body143 144while.body: ; preds = %while.body.preheader, %while.body145 %vecSrcA.0115 = phi <8 x i16> [ %vecSrcA.0, %while.body ], [ %vecSrcA.0105, %while.body.preheader ]146 %vecSrcB.0114 = phi <8 x i16> [ %vecSrcB.0, %while.body ], [ %vecSrcB.0103, %while.body.preheader ]147 %vecSrcB.0.in.in113 = phi ptr [ %add.ptr3, %while.body ], [ %pSrcB, %while.body.preheader ]148 %vecSrcA.0.in.in112 = phi ptr [ %add.ptr2, %while.body ], [ %pSrcA, %while.body.preheader ]149 %accImag.0.off32111 = phi i32 [ %15, %while.body ], [ 0, %while.body.preheader ]150 %accImag.0.off0110 = phi i32 [ %16, %while.body ], [ 0, %while.body.preheader ]151 %accReal.0.off32109 = phi i32 [ %12, %while.body ], [ 0, %while.body.preheader ]152 %accReal.0.off0108 = phi i32 [ %13, %while.body ], [ 0, %while.body.preheader ]153 %blkCnt.0107 = phi i32 [ %dec, %while.body ], [ %shr, %while.body.preheader ]154 %pSrcB.addr.0 = getelementptr inbounds i16, ptr %vecSrcB.0.in.in113, i32 8155 %pSrcA.addr.0 = getelementptr inbounds i16, ptr %vecSrcA.0.in.in112, i32 8156 %1 = tail call { i32, i32 } @llvm.arm.mve.vmlldava.v8i16(i32 0, i32 1, i32 0, i32 %accReal.0.off0108, i32 %accReal.0.off32109, <8 x i16> %vecSrcA.0115, <8 x i16> %vecSrcB.0114)157 %2 = extractvalue { i32, i32 } %1, 1158 %3 = extractvalue { i32, i32 } %1, 0159 %4 = bitcast ptr %pSrcA.addr.0 to ptr160 %5 = load <8 x i16>, ptr %4, align 2161 %add.ptr2 = getelementptr inbounds i16, ptr %vecSrcA.0.in.in112, i32 16162 %6 = tail call { i32, i32 } @llvm.arm.mve.vmlldava.v8i16(i32 0, i32 0, i32 1, i32 %accImag.0.off0110, i32 %accImag.0.off32111, <8 x i16> %vecSrcA.0115, <8 x i16> %vecSrcB.0114)163 %7 = extractvalue { i32, i32 } %6, 1164 %8 = extractvalue { i32, i32 } %6, 0165 %9 = bitcast ptr %pSrcB.addr.0 to ptr166 %10 = load <8 x i16>, ptr %9, align 2167 %add.ptr3 = getelementptr inbounds i16, ptr %vecSrcB.0.in.in113, i32 16168 %11 = tail call { i32, i32 } @llvm.arm.mve.vmlldava.v8i16(i32 0, i32 1, i32 0, i32 %3, i32 %2, <8 x i16> %5, <8 x i16> %10)169 %12 = extractvalue { i32, i32 } %11, 1170 %13 = extractvalue { i32, i32 } %11, 0171 %14 = tail call { i32, i32 } @llvm.arm.mve.vmlldava.v8i16(i32 0, i32 0, i32 1, i32 %8, i32 %7, <8 x i16> %5, <8 x i16> %10)172 %15 = extractvalue { i32, i32 } %14, 1173 %16 = extractvalue { i32, i32 } %14, 0174 %dec = add nsw i32 %blkCnt.0107, -1175 %vecSrcB.0.in = bitcast ptr %add.ptr3 to ptr176 %vecSrcB.0 = load <8 x i16>, ptr %vecSrcB.0.in, align 2177 %vecSrcA.0.in = bitcast ptr %add.ptr2 to ptr178 %vecSrcA.0 = load <8 x i16>, ptr %vecSrcA.0.in, align 2179 %cmp = icmp eq i32 %dec, 0180 br i1 %cmp, label %while.cond.while.end_crit_edge, label %while.body181 182while.cond.while.end_crit_edge: ; preds = %while.body183 %scevgep136 = getelementptr i16, ptr %pSrcB, i32 %0184 br label %while.end185 186while.end: ; preds = %while.cond.while.end_crit_edge, %entry187 %accReal.0.off0.lcssa = phi i32 [ %13, %while.cond.while.end_crit_edge ], [ 0, %entry ]188 %accReal.0.off32.lcssa = phi i32 [ %12, %while.cond.while.end_crit_edge ], [ 0, %entry ]189 %accImag.0.off0.lcssa = phi i32 [ %16, %while.cond.while.end_crit_edge ], [ 0, %entry ]190 %accImag.0.off32.lcssa = phi i32 [ %15, %while.cond.while.end_crit_edge ], [ 0, %entry ]191 %vecSrcA.0.in.in.lcssa = phi ptr [ %scevgep, %while.cond.while.end_crit_edge ], [ %pSrcA, %entry ]192 %vecSrcB.0.in.in.lcssa = phi ptr [ %scevgep136, %while.cond.while.end_crit_edge ], [ %pSrcB, %entry ]193 %vecSrcB.0.lcssa = phi <8 x i16> [ %vecSrcB.0, %while.cond.while.end_crit_edge ], [ %vecSrcB.0103, %entry ]194 %vecSrcA.0.lcssa = phi <8 x i16> [ %vecSrcA.0, %while.cond.while.end_crit_edge ], [ %vecSrcA.0105, %entry ]195 %17 = tail call { i32, i32 } @llvm.arm.mve.vmlldava.v8i16(i32 0, i32 1, i32 0, i32 %accReal.0.off0.lcssa, i32 %accReal.0.off32.lcssa, <8 x i16> %vecSrcA.0.lcssa, <8 x i16> %vecSrcB.0.lcssa)196 %18 = extractvalue { i32, i32 } %17, 1197 %19 = zext i32 %18 to i64198 %20 = shl nuw i64 %19, 32199 %21 = extractvalue { i32, i32 } %17, 0200 %22 = zext i32 %21 to i64201 %23 = or i64 %20, %22202 %24 = tail call { i32, i32 } @llvm.arm.mve.vmlldava.v8i16(i32 0, i32 0, i32 1, i32 %accImag.0.off0.lcssa, i32 %accImag.0.off32.lcssa, <8 x i16> %vecSrcA.0.lcssa, <8 x i16> %vecSrcB.0.lcssa)203 %25 = extractvalue { i32, i32 } %24, 1204 %26 = zext i32 %25 to i64205 %27 = shl nuw i64 %26, 32206 %28 = extractvalue { i32, i32 } %24, 0207 %29 = zext i32 %28 to i64208 %30 = or i64 %27, %29209 %shr8 = and i32 %numSamples, 3210 %cmp1095 = icmp eq i32 %shr8, 0211 %extract = lshr i64 %23, 6212 %extract.t = trunc i64 %extract to i32213 %extract129 = lshr i64 %30, 6214 %extract.t130 = trunc i64 %extract129 to i32215 br i1 %cmp1095, label %while.end34, label %while.body11216 217while.body11: ; preds = %while.end, %while.body11218 %pSrcA.addr.1100 = phi ptr [ %incdec.ptr12, %while.body11 ], [ %vecSrcA.0.in.in.lcssa, %while.end ]219 %pSrcB.addr.199 = phi ptr [ %incdec.ptr14, %while.body11 ], [ %vecSrcB.0.in.in.lcssa, %while.end ]220 %accImag.198 = phi i64 [ %add32, %while.body11 ], [ %30, %while.end ]221 %accReal.197 = phi i64 [ %sub27, %while.body11 ], [ %23, %while.end ]222 %blkCnt.196 = phi i32 [ %dec33, %while.body11 ], [ %shr8, %while.end ]223 %incdec.ptr = getelementptr inbounds i16, ptr %pSrcA.addr.1100, i32 1224 %31 = load i16, ptr %pSrcA.addr.1100, align 2225 %incdec.ptr12 = getelementptr inbounds i16, ptr %pSrcA.addr.1100, i32 2226 %32 = load i16, ptr %incdec.ptr, align 2227 %incdec.ptr13 = getelementptr inbounds i16, ptr %pSrcB.addr.199, i32 1228 %33 = load i16, ptr %pSrcB.addr.199, align 2229 %incdec.ptr14 = getelementptr inbounds i16, ptr %pSrcB.addr.199, i32 2230 %34 = load i16, ptr %incdec.ptr13, align 2231 %conv = sext i16 %31 to i32232 %conv15 = sext i16 %33 to i32233 %mul16 = mul nsw i32 %conv15, %conv234 %conv17 = sext i32 %mul16 to i64235 %add = add nsw i64 %accReal.197, %conv17236 %conv19 = sext i16 %34 to i32237 %mul20 = mul nsw i32 %conv19, %conv238 %conv21 = sext i32 %mul20 to i64239 %conv23 = sext i16 %32 to i32240 %mul25 = mul nsw i32 %conv19, %conv23241 %conv26 = sext i32 %mul25 to i64242 %sub27 = sub i64 %add, %conv26243 %mul30 = mul nsw i32 %conv15, %conv23244 %conv31 = sext i32 %mul30 to i64245 %add22 = add i64 %accImag.198, %conv31246 %add32 = add i64 %add22, %conv21247 %dec33 = add nsw i32 %blkCnt.196, -1248 %cmp10 = icmp eq i32 %dec33, 0249 br i1 %cmp10, label %while.end34.loopexit, label %while.body11250 251while.end34.loopexit: ; preds = %while.body11252 %extract131 = lshr i64 %add32, 6253 %extract.t132 = trunc i64 %extract131 to i32254 %extract127 = lshr i64 %sub27, 6255 %extract.t128 = trunc i64 %extract127 to i32256 br label %while.end34257 258while.end34: ; preds = %while.end34.loopexit, %while.end259 %accReal.1.lcssa.off6 = phi i32 [ %extract.t, %while.end ], [ %extract.t128, %while.end34.loopexit ]260 %accImag.1.lcssa.off6 = phi i32 [ %extract.t130, %while.end ], [ %extract.t132, %while.end34.loopexit ]261 store i32 %accReal.1.lcssa.off6, ptr %realResult, align 4262 store i32 %accImag.1.lcssa.off6, ptr %imagResult, align 4263 ret void264}265 266 267define void @fma8(ptr noalias nocapture readonly %A, ptr noalias nocapture readonly %B, ptr noalias nocapture %C, i32 %n) {268; CHECK-LABEL: fma8:269; CHECK: @ %bb.0: @ %entry270; CHECK-NEXT: .save {r4, r5, r6, lr}271; CHECK-NEXT: push {r4, r5, r6, lr}272; CHECK-NEXT: cmp r3, #1273; CHECK-NEXT: blt .LBB2_8274; CHECK-NEXT: @ %bb.1: @ %for.body.preheader275; CHECK-NEXT: cmp r3, #7276; CHECK-NEXT: bhi .LBB2_3277; CHECK-NEXT: @ %bb.2:278; CHECK-NEXT: mov.w r12, #0279; CHECK-NEXT: b .LBB2_6280; CHECK-NEXT: .LBB2_3: @ %vector.ph281; CHECK-NEXT: bic r12, r3, #7282; CHECK-NEXT: movs r5, #1283; CHECK-NEXT: sub.w r6, r12, #8284; CHECK-NEXT: mov r4, r0285; CHECK-NEXT: add.w lr, r5, r6, lsr #3286; CHECK-NEXT: mov r5, r1287; CHECK-NEXT: mov r6, r2288; CHECK-NEXT: .LBB2_4: @ %vector.body289; CHECK-NEXT: @ =>This Inner Loop Header: Depth=1290; CHECK-NEXT: vldrw.u32 q0, [r4, #16]291; CHECK-NEXT: vldrw.u32 q1, [r5, #16]292; CHECK-NEXT: vldrw.u32 q2, [r6, #16]293; CHECK-NEXT: vldrw.u32 q3, [r6]294; CHECK-NEXT: vfma.f32 q2, q1, q0295; CHECK-NEXT: vldrw.u32 q0, [r4], #32296; CHECK-NEXT: vldrw.u32 q1, [r5], #32297; CHECK-NEXT: vfma.f32 q3, q1, q0298; CHECK-NEXT: vstrw.32 q3, [r6], #32299; CHECK-NEXT: vstrw.32 q2, [r6, #-16]300; CHECK-NEXT: le lr, .LBB2_4301; CHECK-NEXT: @ %bb.5: @ %middle.block302; CHECK-NEXT: cmp r12, r3303; CHECK-NEXT: it eq304; CHECK-NEXT: popeq {r4, r5, r6, pc}305; CHECK-NEXT: .LBB2_6: @ %for.body.preheader12306; CHECK-NEXT: sub.w lr, r3, r12307; CHECK-NEXT: add.w r0, r0, r12, lsl #2308; CHECK-NEXT: add.w r1, r1, r12, lsl #2309; CHECK-NEXT: add.w r2, r2, r12, lsl #2310; CHECK-NEXT: .LBB2_7: @ %for.body311; CHECK-NEXT: @ =>This Inner Loop Header: Depth=1312; CHECK-NEXT: vldmia r0!, {s0}313; CHECK-NEXT: vldmia r1!, {s2}314; CHECK-NEXT: vldr s4, [r2]315; CHECK-NEXT: vfma.f32 s4, s2, s0316; CHECK-NEXT: vstmia r2!, {s4}317; CHECK-NEXT: le lr, .LBB2_7318; CHECK-NEXT: .LBB2_8: @ %for.cond.cleanup319; CHECK-NEXT: pop {r4, r5, r6, pc}320entry:321 %cmp8 = icmp sgt i32 %n, 0322 br i1 %cmp8, label %for.body.preheader, label %for.cond.cleanup323 324for.body.preheader: ; preds = %entry325 %min.iters.check = icmp ult i32 %n, 8326 br i1 %min.iters.check, label %for.body.preheader12, label %vector.ph327 328for.body.preheader12: ; preds = %middle.block, %for.body.preheader329 %i.09.ph = phi i32 [ 0, %for.body.preheader ], [ %n.vec, %middle.block ]330 br label %for.body331 332vector.ph: ; preds = %for.body.preheader333 %n.vec = and i32 %n, -8334 br label %vector.body335 336vector.body: ; preds = %vector.body, %vector.ph337 %index = phi i32 [ 0, %vector.ph ], [ %index.next, %vector.body ]338 %0 = getelementptr inbounds float, ptr %A, i32 %index339 %1 = bitcast ptr %0 to ptr340 %wide.load = load <8 x float>, ptr %1, align 4341 %2 = getelementptr inbounds float, ptr %B, i32 %index342 %3 = bitcast ptr %2 to ptr343 %wide.load10 = load <8 x float>, ptr %3, align 4344 %4 = fmul fast <8 x float> %wide.load10, %wide.load345 %5 = getelementptr inbounds float, ptr %C, i32 %index346 %6 = bitcast ptr %5 to ptr347 %wide.load11 = load <8 x float>, ptr %6, align 4348 %7 = fadd fast <8 x float> %wide.load11, %4349 store <8 x float> %7, ptr %6, align 4350 %index.next = add i32 %index, 8351 %8 = icmp eq i32 %index.next, %n.vec352 br i1 %8, label %middle.block, label %vector.body353 354middle.block: ; preds = %vector.body355 %cmp.n = icmp eq i32 %n.vec, %n356 br i1 %cmp.n, label %for.cond.cleanup, label %for.body.preheader12357 358for.cond.cleanup: ; preds = %for.body, %middle.block, %entry359 ret void360 361for.body: ; preds = %for.body.preheader12, %for.body362 %i.09 = phi i32 [ %inc, %for.body ], [ %i.09.ph, %for.body.preheader12 ]363 %arrayidx = getelementptr inbounds float, ptr %A, i32 %i.09364 %9 = load float, ptr %arrayidx, align 4365 %arrayidx1 = getelementptr inbounds float, ptr %B, i32 %i.09366 %10 = load float, ptr %arrayidx1, align 4367 %mul = fmul fast float %10, %9368 %arrayidx2 = getelementptr inbounds float, ptr %C, i32 %i.09369 %11 = load float, ptr %arrayidx2, align 4370 %add = fadd fast float %11, %mul371 store float %add, ptr %arrayidx2, align 4372 %inc = add nuw nsw i32 %i.09, 1373 %exitcond = icmp eq i32 %inc, %n374 br i1 %exitcond, label %for.cond.cleanup, label %for.body375}376 377declare i32 @llvm.arm.mve.addv.v4i32(<4 x i32>, i32)378declare { i32, i32 } @llvm.arm.mve.vmlldava.v8i16(i32, i32, i32, i32, i32, <8 x i16>, <8 x i16>)379