127 lines · plain
1; RUN: llc -mtriple=hexagon -mv67t -debug-only=pipeliner < %s 2>&1 | FileCheck %s2; REQUIRES: asserts3 4; Test that the artificial dependencies have been created.5 6; CHECK: Ord Latency=0 Artificial7 8define void @bkfir(ptr nocapture readonly %in, ptr nocapture readonly %coefs, i32 %tap, i32 %length, ptr nocapture %out) local_unnamed_addr #0 {9entry:10 %cmp141 = icmp sgt i32 %length, 011 br i1 %cmp141, label %for.body.lr.ph, label %for.end5212 13for.body.lr.ph:14 %cmp8127 = icmp sgt i32 %tap, 015 br i1 %cmp8127, label %for.body.us.preheader, label %for.body.lr.ph.split16 17for.body.us.preheader:18 br label %for.body.us19 20for.body.us:21 %add.ptr.us.phi = phi ptr [ %add.ptr.us.inc, %for.cond7.for.end_crit_edge.us ], [ %in, %for.body.us.preheader ]22 %i.0143.us = phi i32 [ %add51.us, %for.cond7.for.end_crit_edge.us ], [ 0, %for.body.us.preheader ]23 %optr.0142.us = phi ptr [ %incdec.ptr49.us, %for.cond7.for.end_crit_edge.us ], [ %out, %for.body.us.preheader ]24 %incdec.ptr.us = getelementptr inbounds i32, ptr %add.ptr.us.phi, i32 225 %0 = load i64, ptr %add.ptr.us.phi, align 826 %incdec.ptr1.us = getelementptr inbounds i32, ptr %add.ptr.us.phi, i32 427 %1 = load i64, ptr %incdec.ptr.us, align 828 %_Q6V64_internal_union.sroa.0.0.extract.trunc.us = trunc i64 %1 to i3229 %_Q6V64_internal_union2.sroa.3.0.extract.shift.us = lshr i64 %0, 3230 %_Q6V64_internal_union2.sroa.3.0.extract.trunc.us = trunc i64 %_Q6V64_internal_union2.sroa.3.0.extract.shift.us to i3231 %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)32 %add.ptr.us.inc = getelementptr i32, ptr %add.ptr.us.phi, i32 433 br label %for.body9.us34 35for.body9.us:36 %j.0137.us = phi i32 [ 0, %for.body.us ], [ %add.us, %for.body9.us ]37 %x0x1.0136.us = phi i64 [ %0, %for.body.us ], [ %5, %for.body9.us ]38 %x2x3.0135.us = phi i64 [ %1, %for.body.us ], [ %6, %for.body9.us ]39 %x1x2.0134.us = phi i64 [ %2, %for.body.us ], [ %8, %for.body9.us ]40 %iptrD.0133.us = phi ptr [ %incdec.ptr1.us, %for.body.us ], [ %incdec.ptr13.us, %for.body9.us ]41 %iptrC.0132.us = phi ptr [ %coefs, %for.body.us ], [ %incdec.ptr11.us, %for.body9.us ]42 %sum0.0131.us = phi i64 [ 0, %for.body.us ], [ %13, %for.body9.us ]43 %sum1.0130.us = phi i64 [ 0, %for.body.us ], [ %14, %for.body9.us ]44 %sum2.0129.us = phi i64 [ 0, %for.body.us ], [ %15, %for.body9.us ]45 %sum3.0128.us = phi i64 [ 0, %for.body.us ], [ %16, %for.body9.us ]46 %incdec.ptr10.us = getelementptr inbounds i64, ptr %iptrC.0132.us, i32 147 %3 = load i64, ptr %iptrC.0132.us, align 848 %incdec.ptr11.us = getelementptr inbounds i64, ptr %iptrC.0132.us, i32 249 %4 = load i64, ptr %incdec.ptr10.us, align 850 %incdec.ptr12.us = getelementptr inbounds i64, ptr %iptrD.0133.us, i32 151 %5 = load i64, ptr %iptrD.0133.us, align 852 %incdec.ptr13.us = getelementptr inbounds i64, ptr %iptrD.0133.us, i32 253 %6 = load i64, ptr %incdec.ptr12.us, align 854 %_Q6V64_internal_union14.sroa.0.0.extract.trunc.us = trunc i64 %5 to i3255 %_Q6V64_internal_union14.sroa.4.0.extract.shift.us = lshr i64 %5, 3256 %_Q6V64_internal_union19.sroa.3.0.extract.shift.us = lshr i64 %x2x3.0135.us, 3257 %_Q6V64_internal_union19.sroa.3.0.extract.trunc.us = trunc i64 %_Q6V64_internal_union19.sroa.3.0.extract.shift.us to i3258 %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)59 %_Q6V64_internal_union24.sroa.0.0.extract.trunc.us = trunc i64 %6 to i3260 %_Q6V64_internal_union29.sroa.3.0.extract.trunc.us = trunc i64 %_Q6V64_internal_union14.sroa.4.0.extract.shift.us to i3261 %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)62 %9 = tail call i64 @llvm.hexagon.M7.dcmpyrwc.acc(i64 %sum0.0131.us, i64 %x0x1.0136.us, i64 %3)63 %10 = tail call i64 @llvm.hexagon.M7.dcmpyrwc.acc(i64 %sum1.0130.us, i64 %x1x2.0134.us, i64 %3)64 %11 = tail call i64 @llvm.hexagon.M7.dcmpyrwc.acc(i64 %sum2.0129.us, i64 %x2x3.0135.us, i64 %3)65 %12 = tail call i64 @llvm.hexagon.M7.dcmpyrwc.acc(i64 %sum3.0128.us, i64 %7, i64 %3)66 %13 = tail call i64 @llvm.hexagon.M7.dcmpyrwc.acc(i64 %9, i64 %x2x3.0135.us, i64 %4)67 %14 = tail call i64 @llvm.hexagon.M7.dcmpyrwc.acc(i64 %10, i64 %7, i64 %4)68 %15 = tail call i64 @llvm.hexagon.M7.dcmpyrwc.acc(i64 %11, i64 %5, i64 %4)69 %16 = tail call i64 @llvm.hexagon.M7.dcmpyrwc.acc(i64 %12, i64 %8, i64 %4)70 %add.us = add nuw nsw i32 %j.0137.us, 471 %cmp8.us = icmp slt i32 %add.us, %tap72 br i1 %cmp8.us, label %for.body9.us, label %for.cond7.for.end_crit_edge.us73 74for.cond7.for.end_crit_edge.us:75 %17 = ashr i64 %13, 3976 %18 = ashr i64 %14, 3977 %19 = ashr i64 %15, 3978 %20 = ashr i64 %16, 3979 %21 = tail call i32 @llvm.hexagon.A2.sat(i64 %17)80 %22 = tail call i32 @llvm.hexagon.A2.sat(i64 %18)81 %23 = tail call i32 @llvm.hexagon.A2.sat(i64 %19)82 %24 = tail call i32 @llvm.hexagon.A2.sat(i64 %20)83 %_Q6V64_internal_union34.sroa.4.0.insert.ext.us = zext i32 %22 to i6484 %_Q6V64_internal_union34.sroa.4.0.insert.shift.us = shl nuw i64 %_Q6V64_internal_union34.sroa.4.0.insert.ext.us, 3285 %_Q6V64_internal_union34.sroa.0.0.insert.ext.us = zext i32 %21 to i6486 %_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.us87 %incdec.ptr41.us = getelementptr inbounds i64, ptr %optr.0142.us, i32 188 store i64 %_Q6V64_internal_union34.sroa.0.0.insert.insert.us, ptr %optr.0142.us, align 889 %_Q6V64_internal_union42.sroa.4.0.insert.ext.us = zext i32 %24 to i6490 %_Q6V64_internal_union42.sroa.4.0.insert.shift.us = shl nuw i64 %_Q6V64_internal_union42.sroa.4.0.insert.ext.us, 3291 %_Q6V64_internal_union42.sroa.0.0.insert.ext.us = zext i32 %23 to i6492 %_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.us93 %incdec.ptr49.us = getelementptr inbounds i64, ptr %optr.0142.us, i32 294 store i64 %_Q6V64_internal_union42.sroa.0.0.insert.insert.us, ptr %incdec.ptr41.us, align 895 %add51.us = add nuw nsw i32 %i.0143.us, 496 %cmp.us = icmp slt i32 %add51.us, %length97 br i1 %cmp.us, label %for.body.us, label %for.end5298 99for.body.lr.ph.split:100 %25 = tail call i32 @llvm.hexagon.A2.sat(i64 0)101 %_Q6V64_internal_union34.sroa.4.0.insert.ext = zext i32 %25 to i64102 %_Q6V64_internal_union34.sroa.4.0.insert.shift = shl nuw i64 %_Q6V64_internal_union34.sroa.4.0.insert.ext, 32103 %_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.ext104 br label %for.body105 106for.body:107 %i.0143 = phi i32 [ 0, %for.body.lr.ph.split ], [ %add51, %for.body ]108 %optr.0142 = phi ptr [ %out, %for.body.lr.ph.split ], [ %incdec.ptr49, %for.body ]109 %incdec.ptr41 = getelementptr inbounds i64, ptr %optr.0142, i32 1110 store i64 %_Q6V64_internal_union34.sroa.0.0.insert.insert, ptr %optr.0142, align 8111 %incdec.ptr49 = getelementptr inbounds i64, ptr %optr.0142, i32 2112 store i64 %_Q6V64_internal_union34.sroa.0.0.insert.insert, ptr %incdec.ptr41, align 8113 %add51 = add nuw nsw i32 %i.0143, 4114 %cmp = icmp slt i32 %add51, %length115 br i1 %cmp, label %for.body, label %for.end52116 117for.end52:118 ret void119}120 121declare i64 @llvm.hexagon.A2.combinew(i32, i32) #1122declare i64 @llvm.hexagon.M7.dcmpyrwc.acc(i64, i64, i64) #1123declare i32 @llvm.hexagon.A2.sat(i64) #1124 125attributes #0 = { nounwind "target-cpu"="hexagonv67t" "target-features"="+audio" }126attributes #1 = { nounwind readnone }127