Commit 2466ba97 authored by Matt Arsenault's avatar Matt Arsenault
Browse files

LoopDistribute/LAA: Respect convergent 

This case is slightly tricky, because loop distribution should be
allowed in some cases, and not others. As long as runtime dependency
checks don't need to be introduced, this should be OK. This is further
complicated by the fact that LoopDistribute partially ignores if LAA
says that vectorization is safe, and then does its own runtime pointer
legality checks.

Note this pass still does not handle noduplicate correctly, as this
should always be forbidden with it. I'm not going to bother trying to
fix it, as it would require more effort and I think noduplicate should
be removed.

https://reviews.llvm.org/D62607

llvm-svn: 363160
parent 4c69c56b

llvm/include/llvm/Analysis/LoopAccessAnalysis.h

+6 −0
Original line number Diff line number Diff line
@@ -522,6 +522,11 @@ public: 
  /// no memory dependence cycles.

  bool canVectorizeMemory() const { return CanVecMem; }



  /// Return true if there is a convergent operation in the loop. There may

  /// still be reported runtime pointer checks that would be required, but it is

  /// not legal to insert them.

  bool hasConvergentOp() const { return HasConvergentOp; }



  const RuntimePointerChecking *getRuntimePointerChecking() const {

    return PtrRtChecking.get();

  }
@@ -642,6 +647,7 @@ private: 


  /// Cache the result of analyzeLoop.

  bool CanVecMem;

  bool HasConvergentOp;



  /// Indicator that there are non vectorizable stores to a uniform address.

  bool HasDependenceInvolvingLoopInvariantAddress;

llvm/lib/Analysis/LoopAccessAnalysis.cpp

+55 −9
Original line number Diff line number Diff line
@@ -1778,6 +1778,11 @@ void LoopAccessInfo::analyzeLoop(AliasAnalysis *AA, LoopInfo *LI, 
  unsigned NumReads = 0;

  unsigned NumReadWrites = 0;



  bool HasComplexMemInst = false;



  // A runtime check is only legal to insert if there are no convergent calls.

  HasConvergentOp = false;



  PtrRtChecking->Pointers.clear();

  PtrRtChecking->Need = false;
@@ -1785,8 +1790,25 @@ void LoopAccessInfo::analyzeLoop(AliasAnalysis *AA, LoopInfo *LI, 


  // For each block.

  for (BasicBlock *BB : TheLoop->blocks()) {

    // Scan the BB and collect legal loads and stores.

    // Scan the BB and collect legal loads and stores. Also detect any

    // convergent instructions.

    for (Instruction &I : *BB) {

      if (auto *Call = dyn_cast<CallBase>(&I)) {

        if (Call->isConvergent())

          HasConvergentOp = true;

      }



      // With both a non-vectorizable memory instruction and a convergent

      // operation, found in this loop, no reason to continue the search.

      if (HasComplexMemInst && HasConvergentOp) {

        CanVecMem = false;

        return;

      }



      // Avoid hitting recordAnalysis multiple times.

      if (HasComplexMemInst)

        continue;



      // If this is a load, save it. If this instruction can read from memory

      // but is not a load, then we quit. Notice that we don't handle function

      // calls that read or write.
@@ -1805,12 +1827,18 @@ void LoopAccessInfo::analyzeLoop(AliasAnalysis *AA, LoopInfo *LI, 
          continue;



        auto *Ld = dyn_cast<LoadInst>(&I);

        if (!Ld || (!Ld->isSimple() && !IsAnnotatedParallel)) {

        if (!Ld) {

          recordAnalysis("CantVectorizeInstruction", Ld)

            << "instruction cannot be vectorized";

          HasComplexMemInst = true;

          continue;

        }

        if (!Ld->isSimple() && !IsAnnotatedParallel) {

          recordAnalysis("NonSimpleLoad", Ld)

              << "read with atomic ordering or volatile read";

          LLVM_DEBUG(dbgs() << "LAA: Found a non-simple load.\n");

          CanVecMem = false;

          return;

          HasComplexMemInst = true;

          continue;

        }

        NumLoads++;

        Loads.push_back(Ld);
@@ -1826,15 +1854,15 @@ void LoopAccessInfo::analyzeLoop(AliasAnalysis *AA, LoopInfo *LI, 
        if (!St) {

          recordAnalysis("CantVectorizeInstruction", St)

              << "instruction cannot be vectorized";

          CanVecMem = false;

          return;

          HasComplexMemInst = true;

          continue;

        }

        if (!St->isSimple() && !IsAnnotatedParallel) {

          recordAnalysis("NonSimpleStore", St)

              << "write with atomic ordering or volatile write";

          LLVM_DEBUG(dbgs() << "LAA: Found a non-simple store.\n");

          CanVecMem = false;

          return;

          HasComplexMemInst = true;

          continue;

        }

        NumStores++;

        Stores.push_back(St);
@@ -1845,6 +1873,11 @@ void LoopAccessInfo::analyzeLoop(AliasAnalysis *AA, LoopInfo *LI, 
    } // Next instr.

  } // Next block.



  if (HasComplexMemInst) {

    CanVecMem = false;

    return;

  }



  // Now we have two lists that hold the loads and the stores.

  // Next, we find the pointers that they use.
@@ -1962,7 +1995,7 @@ void LoopAccessInfo::analyzeLoop(AliasAnalysis *AA, LoopInfo *LI, 
  }



  LLVM_DEBUG(

      dbgs() << "LAA: We can perform a memory runtime check if needed.\n");

    dbgs() << "LAA: May be able to perform a memory runtime check if needed.\n");



  CanVecMem = true;

  if (Accesses.isDependencyCheckNeeded()) {
@@ -1997,6 +2030,15 @@ void LoopAccessInfo::analyzeLoop(AliasAnalysis *AA, LoopInfo *LI, 
    }

  }



  if (HasConvergentOp) {

    recordAnalysis("CantInsertRuntimeCheckWithConvergent")

      << "cannot add control dependency to convergent operation";

    LLVM_DEBUG(dbgs() << "LAA: We can't vectorize because a runtime check "

                         "would be needed with a convergent operation\n");

    CanVecMem = false;

    return;

  }



  if (CanVecMem)

    LLVM_DEBUG(

        dbgs() << "LAA: No unsafe dependent memory operations in loop.  We"
@@ -2285,6 +2327,7 @@ LoopAccessInfo::LoopAccessInfo(Loop *L, ScalarEvolution *SE, 
      PtrRtChecking(llvm::make_unique<RuntimePointerChecking>(SE)),

      DepChecker(llvm::make_unique<MemoryDepChecker>(*PSE, L)), TheLoop(L),

      NumLoads(0), NumStores(0), MaxSafeDepDistBytes(-1), CanVecMem(false),

      HasConvergentOp(false),

      HasDependenceInvolvingLoopInvariantAddress(false) {

  if (canAnalyzeLoop())

    analyzeLoop(AA, LI, TLI, DT);
@@ -2301,6 +2344,9 @@ void LoopAccessInfo::print(raw_ostream &OS, unsigned Depth) const { 
    OS << "\n";

  }



  if (HasConvergentOp)

    OS.indent(Depth) << "Has convergent operation in loop\n";



  if (Report)

    OS.indent(Depth) << "Report: " << Report->getMsg() << "\n";

llvm/lib/Transforms/Scalar/LoopDistribute.cpp

+14 −1
Original line number Diff line number Diff line
@@ -766,8 +766,14 @@ public: 
                    "cannot isolate unsafe dependencies");

    }



    // Don't distribute the loop if we need too many SCEV run-time checks.

    // Don't distribute the loop if we need too many SCEV run-time checks, or

    // any if it's illegal.

    const SCEVUnionPredicate &Pred = LAI->getPSE().getUnionPredicate();

    if (LAI->hasConvergentOp() && !Pred.isAlwaysTrue()) {

      return fail("RuntimeCheckWithConvergent",

                  "may not insert runtime check with convergent operation");

    }



    if (Pred.getComplexity() > (IsForced.getValueOr(false)

                                    ? PragmaDistributeSCEVCheckThreshold

                                    : DistributeSCEVCheckThreshold))
@@ -795,7 +801,14 @@ public: 
    auto Checks = includeOnlyCrossPartitionChecks(AllChecks, PtrToPartition,

                                                  RtPtrChecking);



    if (LAI->hasConvergentOp() && !Checks.empty()) {

      return fail("RuntimeCheckWithConvergent",

                  "may not insert runtime check with convergent operation");

    }



    if (!Pred.isAlwaysTrue() || !Checks.empty()) {

      assert(!LAI->hasConvergentOp() && "inserting illegal loop versioning");



      MDNode *OrigLoopID = L->getLoopID();



      LLVM_DEBUG(dbgs() << "\nPointers:\n");

llvm/test/Analysis/LoopAccessAnalysis/unsafe-and-rt-checks-convergent.ll

0 → 100644
+73 −0
Original line number Diff line number Diff line 
; RUN: opt -loop-accesses -analyze < %s | FileCheck %s

; RUN: opt -passes='require<scalar-evolution>,require<aa>,loop(print-access-info)' -disable-output  < %s 2>&1 | FileCheck %s



; Analyze this loop:

;   for (i = 0; i < n; i++)

;    A[i + 1] = A[i] * B[i] * C[i];



target datalayout = "e-m:o-i64:64-f80:128-n8:16:32:64-S128"



; CHECK: for.body:

; CHECK: Has convergent operation in loop

; CHECK: Report: cannot add control dependency to convergent operation

; CHECK-NEXT: Dependences:

; CHECK-NEXT:   Backward:

; CHECK-NEXT:     %loadA = load i16, i16* %arrayidxA, align 2 ->

; CHECK-NEXT:     store i16 %mul1, i16* %arrayidxA_plus_2, align 2

; CHECK: Run-time memory checks:

; CHECK-NEXT: 0:

; CHECK-NEXT: Comparing group

; CHECK-NEXT:   %arrayidxA = getelementptr inbounds i16, i16* %a, i64 %storemerge3

; CHECK-NEXT:   %arrayidxA_plus_2 = getelementptr inbounds i16, i16* %a, i64 %add

; CHECK-NEXT: Against group

; CHECK-NEXT:   %arrayidxB = getelementptr inbounds i16, i16* %b, i64 %storemerge3

; CHECK-NEXT: 1:

; CHECK-NEXT: Comparing group

; CHECK-NEXT:   %arrayidxA = getelementptr inbounds i16, i16* %a, i64 %storemerge3

; CHECK-NEXT:   %arrayidxA_plus_2 = getelementptr inbounds i16, i16* %a, i64 %add

; CHECK-NEXT: Against group

; CHECK-NEXT:   %arrayidxC = getelementptr inbounds i16, i16* %c, i64 %storemerge3



@B = common global i16* null, align 8

@A = common global i16* null, align 8

@C = common global i16* null, align 8



define void @f() #1 {

entry:

  %a = load i16*, i16** @A, align 8

  %b = load i16*, i16** @B, align 8

  %c = load i16*, i16** @C, align 8

  br label %for.body



for.body:                                         ; preds = %for.body, %entry

  %storemerge3 = phi i64 [ 0, %entry ], [ %add, %for.body ]



  %arrayidxA = getelementptr inbounds i16, i16* %a, i64 %storemerge3

  %loadA = load i16, i16* %arrayidxA, align 2



  %arrayidxB = getelementptr inbounds i16, i16* %b, i64 %storemerge3

  %loadB = load i16, i16* %arrayidxB, align 2



  %arrayidxC = getelementptr inbounds i16, i16* %c, i64 %storemerge3

  %loadC = load i16, i16* %arrayidxC, align 2



  call void @llvm.convergent()



  %mul = mul i16 %loadB, %loadA

  %mul1 = mul i16 %mul, %loadC



  %add = add nuw nsw i64 %storemerge3, 1

  %arrayidxA_plus_2 = getelementptr inbounds i16, i16* %a, i64 %add

  store i16 %mul1, i16* %arrayidxA_plus_2, align 2



  %exitcond = icmp eq i64 %add, 20

  br i1 %exitcond, label %for.end, label %for.body



for.end:                                          ; preds = %for.body

  ret void

}



declare void @llvm.convergent() #0



attributes #0 = { nounwind readnone convergent }

attributes #1 = { nounwind convergent }

llvm/test/Transforms/LoopDistribute/basic-with-memchecks.ll

+113 −0
Original line number Diff line number Diff line
@@ -5,6 +5,9 @@ 
; RUN:   -verify-loop-info -verify-dom-info -S < %s | \

; RUN:   FileCheck --check-prefix=VECTORIZE %s



; RUN: opt -basicaa -loop-distribute -enable-loop-distribute -verify-loop-info -verify-dom-info \

; RUN:   -loop-accesses -analyze < %s | FileCheck %s --check-prefix=ANALYSIS



; The memcheck version of basic.ll.  We should distribute and vectorize the

; second part of this loop with 5 memchecks (A+1 x {C, D, E} + C x {A, B})

;
@@ -173,3 +176,113 @@ for.body: 
for.end:

  ret void

}



declare i32 @llvm.convergent(i32) #0



; This is the same as f, and would require the same bounds

; check. However, it is not OK to introduce new control dependencies

; on the convergent call.



; CHECK-LABEL: @f_with_convergent(

; CHECK: call i32 @llvm.convergent

; CHECK-NOT: call i32 @llvm.convergent



; ANALYSIS: for.body:

; ANALYSIS: Report: cannot add control dependency to convergent operation

define void @f_with_convergent() #1 {

entry:

  %a = load i32*, i32** @A, align 8

  %b = load i32*, i32** @B, align 8

  %c = load i32*, i32** @C, align 8

  %d = load i32*, i32** @D, align 8

  %e = load i32*, i32** @E, align 8

  br label %for.body



for.body:                                         ; preds = %for.body, %entry

  %ind = phi i64 [ 0, %entry ], [ %add, %for.body ]



  %arrayidxA = getelementptr inbounds i32, i32* %a, i64 %ind

  %loadA = load i32, i32* %arrayidxA, align 4



  %arrayidxB = getelementptr inbounds i32, i32* %b, i64 %ind

  %loadB = load i32, i32* %arrayidxB, align 4



  %mulA = mul i32 %loadB, %loadA



  %add = add nuw nsw i64 %ind, 1

  %arrayidxA_plus_4 = getelementptr inbounds i32, i32* %a, i64 %add

  store i32 %mulA, i32* %arrayidxA_plus_4, align 4



  %arrayidxD = getelementptr inbounds i32, i32* %d, i64 %ind

  %loadD = load i32, i32* %arrayidxD, align 4



  %arrayidxE = getelementptr inbounds i32, i32* %e, i64 %ind

  %loadE = load i32, i32* %arrayidxE, align 4



  %convergentD = call i32 @llvm.convergent(i32 %loadD)

  %mulC = mul i32 %convergentD, %loadE



  %arrayidxC = getelementptr inbounds i32, i32* %c, i64 %ind

  store i32 %mulC, i32* %arrayidxC, align 4



  %exitcond = icmp eq i64 %add, 20

  br i1 %exitcond, label %for.end, label %for.body



for.end:                                          ; preds = %for.body

  ret void

}



; Make sure an explicit request for distribution is ignored if it

; requires possibly illegal checks.



; CHECK-LABEL: @f_with_convergent_forced_distribute(

; CHECK: call i32 @llvm.convergent

; CHECK-NOT: call i32 @llvm.convergent

define void @f_with_convergent_forced_distribute() #1 {

entry:

  %a = load i32*, i32** @A, align 8

  %b = load i32*, i32** @B, align 8

  %c = load i32*, i32** @C, align 8

  %d = load i32*, i32** @D, align 8

  %e = load i32*, i32** @E, align 8

  br label %for.body



for.body:                                         ; preds = %for.body, %entry

  %ind = phi i64 [ 0, %entry ], [ %add, %for.body ]



  %arrayidxA = getelementptr inbounds i32, i32* %a, i64 %ind

  %loadA = load i32, i32* %arrayidxA, align 4



  %arrayidxB = getelementptr inbounds i32, i32* %b, i64 %ind

  %loadB = load i32, i32* %arrayidxB, align 4



  %mulA = mul i32 %loadB, %loadA



  %add = add nuw nsw i64 %ind, 1

  %arrayidxA_plus_4 = getelementptr inbounds i32, i32* %a, i64 %add

  store i32 %mulA, i32* %arrayidxA_plus_4, align 4



  %arrayidxD = getelementptr inbounds i32, i32* %d, i64 %ind

  %loadD = load i32, i32* %arrayidxD, align 4



  %arrayidxE = getelementptr inbounds i32, i32* %e, i64 %ind

  %loadE = load i32, i32* %arrayidxE, align 4



  %convergentD = call i32 @llvm.convergent(i32 %loadD)

  %mulC = mul i32 %convergentD, %loadE



  %arrayidxC = getelementptr inbounds i32, i32* %c, i64 %ind

  store i32 %mulC, i32* %arrayidxC, align 4



  %exitcond = icmp eq i64 %add, 20

  br i1 %exitcond, label %for.end, label %for.body, !llvm.loop !0



for.end:                                          ; preds = %for.body

  ret void

}



attributes #0 = { nounwind readnone convergent }

attributes #1 = { nounwind convergent }



!0 = distinct !{!0, !1}

!1 = !{!"llvm.loop.distribute.enable", i1 true}