130 lines · plain
1; RUN: llc -O3 -mv67t -mtriple=hexagon < %s | FileCheck %s2 3; Test that the inner loop in the tiny core version of bkfir has the assembler4; directive "p2align 4".5 6; CHECK: loop0(.LBB0_[[LOOP:.]],7; CHECK-NOT: falign8; CHECK: p2align 49; CHECK: } :endloop010 11define void @bkfir(ptr nocapture readonly %in, ptr nocapture readonly %coefs, i32 %tap, i32 %length, ptr nocapture %out) local_unnamed_addr #0 {12entry:13 %cmp141 = icmp sgt i32 %length, 014 br i1 %cmp141, label %for.body.lr.ph, label %for.end5215 16for.body.lr.ph:17 %cmp8127 = icmp sgt i32 %tap, 018 br i1 %cmp8127, label %for.body.us.preheader, label %for.body.lr.ph.split19 20for.body.us.preheader:21 br label %for.body.us22 23for.body.us:24 %add.ptr.us.phi = phi ptr [ %add.ptr.us.inc, %for.cond7.for.end_crit_edge.us ], [ %in, %for.body.us.preheader ]25 %i.0143.us = phi i32 [ %add51.us, %for.cond7.for.end_crit_edge.us ], [ 0, %for.body.us.preheader ]26 %optr.0142.us = phi ptr [ %incdec.ptr49.us, %for.cond7.for.end_crit_edge.us ], [ %out, %for.body.us.preheader ]27 %incdec.ptr.us = getelementptr inbounds i32, ptr %add.ptr.us.phi, i32 228 %0 = load i64, ptr %add.ptr.us.phi, align 829 %incdec.ptr1.us = getelementptr inbounds i32, ptr %add.ptr.us.phi, i32 430 %1 = load i64, ptr %incdec.ptr.us, align 831 %_Q6V64_internal_union.sroa.0.0.extract.trunc.us = trunc i64 %1 to i3232 %_Q6V64_internal_union2.sroa.3.0.extract.shift.us = lshr i64 %0, 3233 %_Q6V64_internal_union2.sroa.3.0.extract.trunc.us = trunc i64 %_Q6V64_internal_union2.sroa.3.0.extract.shift.us to i3234 %2 = tail call i64 @llvm.hexagon.A2.combinew(i32 %_Q6V64_internal_union.sroa.0.0.extract.trunc.us, i32 %_Q6V64_internal_union2.sroa.3.0.extract.trunc.us)35 %add.ptr.us.inc = getelementptr i32, ptr %add.ptr.us.phi, i32 436 br label %for.body9.us37 38for.body9.us:39 %j.0137.us = phi i32 [ 0, %for.body.us ], [ %add.us, %for.body9.us ]40 %x0x1.0136.us = phi i64 [ %0, %for.body.us ], [ %5, %for.body9.us ]41 %x2x3.0135.us = phi i64 [ %1, %for.body.us ], [ %6, %for.body9.us ]42 %x1x2.0134.us = phi i64 [ %2, %for.body.us ], [ %8, %for.body9.us ]43 %iptrD.0133.us = phi ptr [ %incdec.ptr1.us, %for.body.us ], [ %incdec.ptr13.us, %for.body9.us ]44 %iptrC.0132.us = phi ptr [ %coefs, %for.body.us ], [ %incdec.ptr11.us, %for.body9.us ]45 %sum0.0131.us = phi i64 [ 0, %for.body.us ], [ %13, %for.body9.us ]46 %sum1.0130.us = phi i64 [ 0, %for.body.us ], [ %14, %for.body9.us ]47 %sum2.0129.us = phi i64 [ 0, %for.body.us ], [ %15, %for.body9.us ]48 %sum3.0128.us = phi i64 [ 0, %for.body.us ], [ %16, %for.body9.us ]49 %incdec.ptr10.us = getelementptr inbounds i64, ptr %iptrC.0132.us, i32 150 %3 = load i64, ptr %iptrC.0132.us, align 851 %incdec.ptr11.us = getelementptr inbounds i64, ptr %iptrC.0132.us, i32 252 %4 = load i64, ptr %incdec.ptr10.us, align 853 %incdec.ptr12.us = getelementptr inbounds i64, ptr %iptrD.0133.us, i32 154 %5 = load i64, ptr %iptrD.0133.us, align 855 %incdec.ptr13.us = getelementptr inbounds i64, ptr %iptrD.0133.us, i32 256 %6 = load i64, ptr %incdec.ptr12.us, align 857 %_Q6V64_internal_union14.sroa.0.0.extract.trunc.us = trunc i64 %5 to i3258 %_Q6V64_internal_union14.sroa.4.0.extract.shift.us = lshr i64 %5, 3259 %_Q6V64_internal_union19.sroa.3.0.extract.shift.us = lshr i64 %x2x3.0135.us, 3260 %_Q6V64_internal_union19.sroa.3.0.extract.trunc.us = trunc i64 %_Q6V64_internal_union19.sroa.3.0.extract.shift.us to i3261 %7 = tail call i64 @llvm.hexagon.A2.combinew(i32 %_Q6V64_internal_union14.sroa.0.0.extract.trunc.us, i32 %_Q6V64_internal_union19.sroa.3.0.extract.trunc.us)62 %_Q6V64_internal_union24.sroa.0.0.extract.trunc.us = trunc i64 %6 to i3263 %_Q6V64_internal_union29.sroa.3.0.extract.trunc.us = trunc i64 %_Q6V64_internal_union14.sroa.4.0.extract.shift.us to i3264 %8 = tail call i64 @llvm.hexagon.A2.combinew(i32 %_Q6V64_internal_union24.sroa.0.0.extract.trunc.us, i32 %_Q6V64_internal_union29.sroa.3.0.extract.trunc.us)65 %9 = tail call i64 @llvm.hexagon.M7.dcmpyrwc.acc(i64 %sum0.0131.us, i64 %x0x1.0136.us, i64 %3)66 %10 = tail call i64 @llvm.hexagon.M7.dcmpyrwc.acc(i64 %sum1.0130.us, i64 %x1x2.0134.us, i64 %3)67 %11 = tail call i64 @llvm.hexagon.M7.dcmpyrwc.acc(i64 %sum2.0129.us, i64 %x2x3.0135.us, i64 %3)68 %12 = tail call i64 @llvm.hexagon.M7.dcmpyrwc.acc(i64 %sum3.0128.us, i64 %7, i64 %3)69 %13 = tail call i64 @llvm.hexagon.M7.dcmpyrwc.acc(i64 %9, i64 %x2x3.0135.us, i64 %4)70 %14 = tail call i64 @llvm.hexagon.M7.dcmpyrwc.acc(i64 %10, i64 %7, i64 %4)71 %15 = tail call i64 @llvm.hexagon.M7.dcmpyrwc.acc(i64 %11, i64 %5, i64 %4)72 %16 = tail call i64 @llvm.hexagon.M7.dcmpyrwc.acc(i64 %12, i64 %8, i64 %4)73 %add.us = add nuw nsw i32 %j.0137.us, 474 %cmp8.us = icmp slt i32 %add.us, %tap75 br i1 %cmp8.us, label %for.body9.us, label %for.cond7.for.end_crit_edge.us76 77for.cond7.for.end_crit_edge.us:78 %17 = ashr i64 %13, 3979 %18 = ashr i64 %14, 3980 %19 = ashr i64 %15, 3981 %20 = ashr i64 %16, 3982 %21 = tail call i32 @llvm.hexagon.A2.sat(i64 %17)83 %22 = tail call i32 @llvm.hexagon.A2.sat(i64 %18)84 %23 = tail call i32 @llvm.hexagon.A2.sat(i64 %19)85 %24 = tail call i32 @llvm.hexagon.A2.sat(i64 %20)86 %_Q6V64_internal_union34.sroa.4.0.insert.ext.us = zext i32 %22 to i6487 %_Q6V64_internal_union34.sroa.4.0.insert.shift.us = shl nuw i64 %_Q6V64_internal_union34.sroa.4.0.insert.ext.us, 3288 %_Q6V64_internal_union34.sroa.0.0.insert.ext.us = zext i32 %21 to i6489 %_Q6V64_internal_union34.sroa.0.0.insert.insert.us = or i64 %_Q6V64_internal_union34.sroa.4.0.insert.shift.us, %_Q6V64_internal_union34.sroa.0.0.insert.ext.us90 %incdec.ptr41.us = getelementptr inbounds i64, ptr %optr.0142.us, i32 191 store i64 %_Q6V64_internal_union34.sroa.0.0.insert.insert.us, ptr %optr.0142.us, align 892 %_Q6V64_internal_union42.sroa.4.0.insert.ext.us = zext i32 %24 to i6493 %_Q6V64_internal_union42.sroa.4.0.insert.shift.us = shl nuw i64 %_Q6V64_internal_union42.sroa.4.0.insert.ext.us, 3294 %_Q6V64_internal_union42.sroa.0.0.insert.ext.us = zext i32 %23 to i6495 %_Q6V64_internal_union42.sroa.0.0.insert.insert.us = or i64 %_Q6V64_internal_union42.sroa.4.0.insert.shift.us, %_Q6V64_internal_union42.sroa.0.0.insert.ext.us96 %incdec.ptr49.us = getelementptr inbounds i64, ptr %optr.0142.us, i32 297 store i64 %_Q6V64_internal_union42.sroa.0.0.insert.insert.us, ptr %incdec.ptr41.us, align 898 %add51.us = add nuw nsw i32 %i.0143.us, 499 %cmp.us = icmp slt i32 %add51.us, %length100 br i1 %cmp.us, label %for.body.us, label %for.end52101 102for.body.lr.ph.split:103 %25 = tail call i32 @llvm.hexagon.A2.sat(i64 0)104 %_Q6V64_internal_union34.sroa.4.0.insert.ext = zext i32 %25 to i64105 %_Q6V64_internal_union34.sroa.4.0.insert.shift = shl nuw i64 %_Q6V64_internal_union34.sroa.4.0.insert.ext, 32106 %_Q6V64_internal_union34.sroa.0.0.insert.insert = or i64 %_Q6V64_internal_union34.sroa.4.0.insert.shift, %_Q6V64_internal_union34.sroa.4.0.insert.ext107 br label %for.body108 109for.body:110 %i.0143 = phi i32 [ 0, %for.body.lr.ph.split ], [ %add51, %for.body ]111 %optr.0142 = phi ptr [ %out, %for.body.lr.ph.split ], [ %incdec.ptr49, %for.body ]112 %incdec.ptr41 = getelementptr inbounds i64, ptr %optr.0142, i32 1113 store i64 %_Q6V64_internal_union34.sroa.0.0.insert.insert, ptr %optr.0142, align 8114 %incdec.ptr49 = getelementptr inbounds i64, ptr %optr.0142, i32 2115 store i64 %_Q6V64_internal_union34.sroa.0.0.insert.insert, ptr %incdec.ptr41, align 8116 %add51 = add nuw nsw i32 %i.0143, 4117 %cmp = icmp slt i32 %add51, %length118 br i1 %cmp, label %for.body, label %for.end52119 120for.end52:121 ret void122}123 124declare i64 @llvm.hexagon.A2.combinew(i32, i32) #1125declare i64 @llvm.hexagon.M7.dcmpyrwc.acc(i64, i64, i64) #1126declare i32 @llvm.hexagon.A2.sat(i64) #1127 128attributes #0 = { nounwind "target-cpu"="hexagonv67t" "target-features"="+audio" }129attributes #1 = { nounwind readnone }130