[ARM][MVE] Clear MaskedInsts vector

[samparker]

Summary

In MVETailPredication, clear the vector before running on a new loop.

[SjoerdMeijer]

Looks like a straightforward fix to me.

Nit: perhaps describe somewhere the loop structure causing a problem, perhaps the test name.

[dmgreen]

llvm/lib/Target/ARM/MVETailPredication.cpp

151

ClonedVCTPInExitBlock too?

[SjoerdMeijer]

llvm/lib/Target/ARM/MVETailPredication.cpp

455

I guess we're fine for that one because that is (re)initialised here for each call to Cleanup

llvm/lib/Target/ARM/MVETailPredication.cpp

void MVETailPredication::RematerializeIterCount() {
   rewriteLoopExitValues(L, LI, TLI, SE, Rewriter, DT, ReplaceExitValue,
                         DeadInsts);
 }
 
 bool MVETailPredication::runOnLoop(Loop *L, LPPassManager&) {
   if (skipLoop(L) || DisableTailPredication)
     return false;
 
+  MaskedInsts.clear();
   Function &F = *L->getHeader()->getParent();
   auto &TPC = getAnalysis<TargetPassConfig>();
   auto &TM = TPC.getTM<TargetMachine>();
   auto *ST = &TM.getSubtarget<ARMSubtarget>(F);
   DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
   LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
   TTI = &getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
   SE = &getAnalysis<ScalarEvolutionWrapperPass>().getSE();

 static bool Cleanup(DenseMap<Instruction*, Instruction*> &NewPredicates,
                     SetVector<Instruction*> &MaybeDead, Loop *L) {
   BasicBlock *Exit = L->getUniqueExitBlock();
   if (!Exit) {
     LLVM_DEBUG(dbgs() << "ARM TP: can't find loop exit block\n");
     return false;
   }
 
   bool ClonedVCTPInExitBlock = false;
 
   for (auto &Pair : NewPredicates) {
     Instruction *OldPred = Pair.first;
     Instruction *NewPred = Pair.second;
 
     for (auto &I : *Exit) {
       if (I.isSameOperationAs(OldPred)) {
         Instruction *PredClone = NewPred->clone();

llvm/test/CodeGen/Thumb2/LowOverheadLoops/clear-maskedinsts.ll

+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
+; RUN: opt -mtriple=thumbv8.1m.main -mattr=+mve.fp -mve-tail-predication -disable-mve-tail-predication=false %s -S -o - | FileCheck %s
+
+define hidden i32 @_Z4loopPiPjiS0_i(i32* noalias nocapture readonly %s1, i32* noalias nocapture readonly %s2, i32 %x, i32* noalias nocapture %d, i32 %n) {
+; CHECK-LABEL: @_Z4loopPiPjiS0_i(
+; CHECK-NEXT:  entry:
+; CHECK-NEXT:    [[CMP63:%.*]] = icmp sgt i32 [[N:%.*]], 0
+; CHECK-NEXT:    br i1 [[CMP63]], label [[FOR_BODY_LR_PH:%.*]], label [[FOR_COND_CLEANUP:%.*]]
+; CHECK:       for.body.lr.ph:
+; CHECK-NEXT:    [[TOBOOL:%.*]] = icmp eq i32 [[X:%.*]], 0
+; CHECK-NEXT:    [[N_RND_UP77:%.*]] = add nuw i32 [[N]], 3
+; CHECK-NEXT:    [[N_VEC79:%.*]] = and i32 [[N_RND_UP77]], -4
+; CHECK-NEXT:    [[TRIP_COUNT_MINUS_183:%.*]] = add nsw i32 [[N]], -1
+; CHECK-NEXT:    [[TMP0:%.*]] = add i32 [[N_VEC79]], -4
+; CHECK-NEXT:    [[TMP1:%.*]] = lshr i32 [[TMP0]], 2
+; CHECK-NEXT:    [[TMP2:%.*]] = add nuw nsw i32 [[TMP1]], 1
+; CHECK-NEXT:    [[TMP3:%.*]] = add nuw nsw i32 [[TMP1]], 1
+; CHECK-NEXT:    br i1 [[TOBOOL]], label [[VECTOR_BODY75_PREHEADER:%.*]], label [[VECTOR_PH:%.*]]
+; CHECK:       vector.body75.preheader:
+; CHECK-NEXT:    call void @llvm.set.loop.iterations.i32(i32 [[TMP2]])
+; CHECK-NEXT:    br label [[VECTOR_BODY75:%.*]]
+; CHECK:       vector.ph:
+; CHECK-NEXT:    [[BROADCAST_SPLATINSERT71:%.*]] = insertelement <4 x i32> undef, i32 [[X]], i32 0
+; CHECK-NEXT:    [[BROADCAST_SPLAT72:%.*]] = shufflevector <4 x i32> [[BROADCAST_SPLATINSERT71]], <4 x i32> undef, <4 x i32> zeroinitializer
+; CHECK-NEXT:    call void @llvm.set.loop.iterations.i32(i32 [[TMP3]])
+; CHECK-NEXT:    br label [[VECTOR_BODY:%.*]]
+; CHECK:       vector.body:
+; CHECK-NEXT:    [[LSR_IV9:%.*]] = phi i32* [ [[SCEVGEP10:%.*]], [[VECTOR_BODY]] ], [ [[D:%.*]], [[VECTOR_PH]] ]
+; CHECK-NEXT:    [[TMP4:%.*]] = phi i32 [ [[TMP3]], [[VECTOR_PH]] ], [ [[TMP8:%.*]], [[VECTOR_BODY]] ]
+; CHECK-NEXT:    [[TMP5:%.*]] = phi i32 [ [[N]], [[VECTOR_PH]] ], [ [[TMP7:%.*]], [[VECTOR_BODY]] ]
+; CHECK-NEXT:    [[LSR_IV911:%.*]] = bitcast i32* [[LSR_IV9]] to <4 x i32>*
+; CHECK-NEXT:    [[TMP6:%.*]] = call <4 x i1> @llvm.arm.mve.vctp32(i32 [[TMP5]])
+; CHECK-NEXT:    [[TMP7]] = sub i32 [[TMP5]], 4
+; CHECK-NEXT:    call void @llvm.masked.store.v4i32.p0v4i32(<4 x i32> [[BROADCAST_SPLAT72]], <4 x i32>* [[LSR_IV911]], i32 4, <4 x i1> [[TMP6]])
+; CHECK-NEXT:    [[SCEVGEP10]] = getelementptr i32, i32* [[LSR_IV9]], i32 4
+; CHECK-NEXT:    [[TMP8]] = call i32 @llvm.loop.decrement.reg.i32.i32.i32(i32 [[TMP4]], i32 1)
+; CHECK-NEXT:    [[TMP9:%.*]] = icmp ne i32 [[TMP8]], 0
+; CHECK-NEXT:    br i1 [[TMP9]], label [[VECTOR_BODY]], label [[FOR_COND_CLEANUP]]
+; CHECK:       vector.body75:
+; CHECK-NEXT:    [[LSR_IV6:%.*]] = phi i32* [ [[S1:%.*]], [[VECTOR_BODY75_PREHEADER]] ], [ [[SCEVGEP7:%.*]], [[VECTOR_BODY75]] ]
+; CHECK-NEXT:    [[LSR_IV3:%.*]] = phi i32* [ [[S2:%.*]], [[VECTOR_BODY75_PREHEADER]] ], [ [[SCEVGEP4:%.*]], [[VECTOR_BODY75]] ]
+; CHECK-NEXT:    [[LSR_IV:%.*]] = phi i32* [ [[D]], [[VECTOR_BODY75_PREHEADER]] ], [ [[SCEVGEP:%.*]], [[VECTOR_BODY75]] ]
+; CHECK-NEXT:    [[INDEX80:%.*]] = phi i32 [ [[INDEX_NEXT81:%.*]], [[VECTOR_BODY75]] ], [ 0, [[VECTOR_BODY75_PREHEADER]] ]
+; CHECK-NEXT:    [[TMP10:%.*]] = phi i32 [ [[TMP2]], [[VECTOR_BODY75_PREHEADER]] ], [ [[TMP15:%.*]], [[VECTOR_BODY75]] ]
+; CHECK-NEXT:    [[LSR_IV68:%.*]] = bitcast i32* [[LSR_IV6]] to <4 x i32>*
+; CHECK-NEXT:    [[LSR_IV35:%.*]] = bitcast i32* [[LSR_IV3]] to <4 x i32>*
+; CHECK-NEXT:    [[LSR_IV2:%.*]] = bitcast i32* [[LSR_IV]] to <4 x i32>*
+; CHECK-NEXT:    [[BROADCAST_SPLATINSERT84:%.*]] = insertelement <4 x i32> undef, i32 [[INDEX80]], i32 0
+; CHECK-NEXT:    [[BROADCAST_SPLAT85:%.*]] = shufflevector <4 x i32> [[BROADCAST_SPLATINSERT84]], <4 x i32> undef, <4 x i32> zeroinitializer
+; CHECK-NEXT:    [[INDUCTION86:%.*]] = add <4 x i32> [[BROADCAST_SPLAT85]], <i32 0, i32 1, i32 2, i32 3>
+; CHECK-NEXT:    [[TMP11:%.*]] = insertelement <4 x i32> undef, i32 [[TRIP_COUNT_MINUS_183]], i32 0
+; CHECK-NEXT:    [[TMP12:%.*]] = shufflevector <4 x i32> [[TMP11]], <4 x i32> undef, <4 x i32> zeroinitializer
+; CHECK-NEXT:    [[TMP13:%.*]] = icmp ule <4 x i32> [[INDUCTION86]], [[TMP12]]
+; CHECK-NEXT:    [[WIDE_MASKED_LOAD:%.*]] = call <4 x i32> @llvm.masked.load.v4i32.p0v4i32(<4 x i32>* [[LSR_IV68]], i32 4, <4 x i1> [[TMP13]], <4 x i32> undef)
+; CHECK-NEXT:    [[WIDE_MASKED_LOAD89:%.*]] = call <4 x i32> @llvm.masked.load.v4i32.p0v4i32(<4 x i32>* [[LSR_IV35]], i32 4, <4 x i1> [[TMP13]], <4 x i32> undef)
+; CHECK-NEXT:    [[TMP14:%.*]] = call <4 x i32> @llvm.usub.sat.v4i32(<4 x i32> [[WIDE_MASKED_LOAD89]], <4 x i32> [[WIDE_MASKED_LOAD]])
+; CHECK-NEXT:    call void @llvm.masked.store.v4i32.p0v4i32(<4 x i32> [[TMP14]], <4 x i32>* [[LSR_IV2]], i32 4, <4 x i1> [[TMP13]])
+; CHECK-NEXT:    [[INDEX_NEXT81]] = add i32 [[INDEX80]], 4
+; CHECK-NEXT:    [[SCEVGEP]] = getelementptr i32, i32* [[LSR_IV]], i32 4
+; CHECK-NEXT:    [[SCEVGEP4]] = getelementptr i32, i32* [[LSR_IV3]], i32 4
+; CHECK-NEXT:    [[SCEVGEP7]] = getelementptr i32, i32* [[LSR_IV6]], i32 4
+; CHECK-NEXT:    [[TMP15]] = call i32 @llvm.loop.decrement.reg.i32.i32.i32(i32 [[TMP10]], i32 1)
+; CHECK-NEXT:    [[TMP16:%.*]] = icmp ne i32 [[TMP15]], 0
+; CHECK-NEXT:    br i1 [[TMP16]], label [[VECTOR_BODY75]], label [[FOR_COND_CLEANUP]]
+; CHECK:       for.cond.cleanup:
+; CHECK-NEXT:    ret i32 0
+;
+entry:
+  %cmp63 = icmp sgt i32 %n, 0
+  br i1 %cmp63, label %for.body.lr.ph, label %for.cond.cleanup
+
+for.body.lr.ph:                                   ; preds = %entry
+  %tobool = icmp eq i32 %x, 0
+  %n.rnd.up77 = add nuw i32 %n, 3
+  %n.vec79 = and i32 %n.rnd.up77, -4
+  %trip.count.minus.183 = add nsw i32 %n, -1
+  %0 = add i32 %n.vec79, -4
+  %1 = lshr i32 %0, 2
+  %2 = add nuw nsw i32 %1, 1
+  %3 = add nuw nsw i32 %1, 1
+  br i1 %tobool, label %vector.body75.preheader, label %vector.ph
+
+vector.body75.preheader:                          ; preds = %for.body.lr.ph
+  call void @llvm.set.loop.iterations.i32(i32 %2)
+  br label %vector.body75
+
+vector.ph:                                        ; preds = %for.body.lr.ph
+  %broadcast.splatinsert71 = insertelement <4 x i32> undef, i32 %x, i32 0
+  %broadcast.splat72 = shufflevector <4 x i32> %broadcast.splatinsert71, <4 x i32> undef, <4 x i32> zeroinitializer
+  call void @llvm.set.loop.iterations.i32(i32 %3)
+  br label %vector.body
+
+vector.body:                                      ; preds = %vector.body, %vector.ph
+  %lsr.iv9 = phi i32* [ %scevgep10, %vector.body ], [ %d, %vector.ph ]
+  %index = phi i32 [ 0, %vector.ph ], [ %index.next, %vector.body ]
+  %4 = phi i32 [ %3, %vector.ph ], [ %8, %vector.body ]
+  %lsr.iv911 = bitcast i32* %lsr.iv9 to <4 x i32>*
+  %broadcast.splatinsert = insertelement <4 x i32> undef, i32 %index, i32 0
+  %broadcast.splat = shufflevector <4 x i32> %broadcast.splatinsert, <4 x i32> undef, <4 x i32> zeroinitializer
+  %induction = add <4 x i32> %broadcast.splat, <i32 0, i32 1, i32 2, i32 3>
+  %5 = insertelement <4 x i32> undef, i32 %trip.count.minus.183, i32 0
+  %6 = shufflevector <4 x i32> %5, <4 x i32> undef, <4 x i32> zeroinitializer
+  %7 = icmp ule <4 x i32> %induction, %6
+  call void @llvm.masked.store.v4i32.p0v4i32(<4 x i32> %broadcast.splat72, <4 x i32>* %lsr.iv911, i32 4, <4 x i1> %7)
+  %index.next = add i32 %index, 4
+  %scevgep10 = getelementptr i32, i32* %lsr.iv9, i32 4
+  %8 = call i32 @llvm.loop.decrement.reg.i32.i32.i32(i32 %4, i32 1)
+  %9 = icmp ne i32 %8, 0
+  br i1 %9, label %vector.body, label %for.cond.cleanup
+
+vector.body75:                                    ; preds = %vector.body75, %vector.body75.preheader
+  %lsr.iv6 = phi i32* [ %s1, %vector.body75.preheader ], [ %scevgep7, %vector.body75 ]
+  %lsr.iv3 = phi i32* [ %s2, %vector.body75.preheader ], [ %scevgep4, %vector.body75 ]
+  %lsr.iv = phi i32* [ %d, %vector.body75.preheader ], [ %scevgep, %vector.body75 ]
+  %index80 = phi i32 [ %index.next81, %vector.body75 ], [ 0, %vector.body75.preheader ]
+  %10 = phi i32 [ %2, %vector.body75.preheader ], [ %15, %vector.body75 ]
+  %lsr.iv68 = bitcast i32* %lsr.iv6 to <4 x i32>*
+  %lsr.iv35 = bitcast i32* %lsr.iv3 to <4 x i32>*
+  %lsr.iv2 = bitcast i32* %lsr.iv to <4 x i32>*
+  %broadcast.splatinsert84 = insertelement <4 x i32> undef, i32 %index80, i32 0
+  %broadcast.splat85 = shufflevector <4 x i32> %broadcast.splatinsert84, <4 x i32> undef, <4 x i32> zeroinitializer
+  %induction86 = add <4 x i32> %broadcast.splat85, <i32 0, i32 1, i32 2, i32 3>
+  %11 = insertelement <4 x i32> undef, i32 %trip.count.minus.183, i32 0
+  %12 = shufflevector <4 x i32> %11, <4 x i32> undef, <4 x i32> zeroinitializer
+  %13 = icmp ule <4 x i32> %induction86, %12
+  %wide.masked.load = call <4 x i32> @llvm.masked.load.v4i32.p0v4i32(<4 x i32>* %lsr.iv68, i32 4, <4 x i1> %13, <4 x i32> undef)
+  %wide.masked.load89 = call <4 x i32> @llvm.masked.load.v4i32.p0v4i32(<4 x i32>* %lsr.iv35, i32 4, <4 x i1> %13, <4 x i32> undef)
+  %14 = call <4 x i32> @llvm.usub.sat.v4i32(<4 x i32> %wide.masked.load89, <4 x i32> %wide.masked.load)
+  call void @llvm.masked.store.v4i32.p0v4i32(<4 x i32> %14, <4 x i32>* %lsr.iv2, i32 4, <4 x i1> %13)
+  %index.next81 = add i32 %index80, 4
+  %scevgep = getelementptr i32, i32* %lsr.iv, i32 4
+  %scevgep4 = getelementptr i32, i32* %lsr.iv3, i32 4
+  %scevgep7 = getelementptr i32, i32* %lsr.iv6, i32 4
+  %15 = call i32 @llvm.loop.decrement.reg.i32.i32.i32(i32 %10, i32 1)
+  %16 = icmp ne i32 %15, 0
+  br i1 %16, label %vector.body75, label %for.cond.cleanup
+
+for.cond.cleanup:                                 ; preds = %vector.body, %vector.body75, %entry
+  ret i32 0
+}
+declare void @llvm.masked.store.v4i32.p0v4i32(<4 x i32>, <4 x i32>*, i32 immarg, <4 x i1>)
+declare <4 x i32> @llvm.masked.load.v4i32.p0v4i32(<4 x i32>*, i32 immarg, <4 x i1>, <4 x i32>)
+declare <4 x i32> @llvm.usub.sat.v4i32(<4 x i32>, <4 x i32>)
+declare void @llvm.set.loop.iterations.i32(i32)
+declare i32 @llvm.loop.decrement.reg.i32.i32.i32(i32, i32)