[Unroll/UnrollAndJam/Vectorizer/Distribute] Add followup loop attributes.

When multiple loop transformation are defined in a loop's metadata, their order of execution is defined by the order of their respective passes in the pass pipeline. For instance, e.g.

    #pragma clang loop unroll_and_jam(enable)
    #pragma clang loop distribute(enable)

is the same as

    #pragma clang loop distribute(enable)
    #pragma clang loop unroll_and_jam(enable)

and will try to loop-distribute before Unroll-And-Jam because the LoopDistribute pass is scheduled after UnrollAndJam pass. UnrollAndJamPass only supports one inner loop, i.e. it will necessarily fail after loop distribution. It is not possible to specify another execution order. Also,t the order of passes in the pipeline is subject to change between versions of LLVM, optimization options and which pass manager is used.

This patch adds 'followup' attributes to various loop transformation passes. These attributes define which attributes the resulting loop of a transformation should have. For instance,

    !0 = !{!0, !1, !2}
    !1 = !{!"llvm.loop.unroll_and_jam.enable"}
    !2 = !{!"llvm.loop.unroll_and_jam.followup_inner", !3}
    !3 = !{!"llvm.loop.distribute.enable"}

defines a loop ID (!0) to be unrolled-and-jammed (!1) and then the attribute !3 to be added to the jammed inner loop, which contains the instruction to distribute the inner loop.

Currently, in both pass managers, pass execution is in a fixed order and UnrollAndJamPass will not execute again after LoopDistribute. We hope to fix this in the future by allowing pass managers to run passes until a fixpoint is reached, use Polly to perform these transformations, or add a loop transformation pass which takes the order issue into account.

For mandatory/forced transformations (e.g. by having been declared by #pragma omp simd), the user must be notified when a transformation could not be performed. It is not possible that the responsible pass emits such a warning because the transformation might be 'hidden' in a followup attribute when it is executed, or it is not present in the pipeline at all. For this reason, this patche introduces a WarnMissedTransformations pass, to warn about orphaned transformations.

Since this changes the user-visible diagnostic message when a transformation is applied, two test cases in the clang repository need to be updated.

To ensure that no other transformation is executed before the intended one, the attribute `llvm.loop.disable_nonforced` can be added which should disable transformation heuristics before the intended transformation is applied. E.g. it would be surprising if a loop is distributed before a #pragma unroll_and_jam is applied.

With more supported code transformations (loop fusion, interchange, stripmining, offloading, etc.), transformations can be used as building blocks for more complex transformations (e.g. stripmining+stripmining+interchange -> tiling).

Reviewed By: hfinkel, dmgreen

Differential Revision: https://reviews.llvm.org/D49281
Differential Revision: https://reviews.llvm.org/D55288

llvm-svn: 348944

4 files changed, 286 insertions, 34 deletions

diff --git a/llvm/lib/Transforms/Utils/LoopUnroll.cpp b/llvm/lib/Transforms/Utils/LoopUnroll.cpp
index 877e0e4dcf9..efd8b92e814 100644
--- a/llvm/lib/Transforms/Utils/LoopUnroll.cpp
+++ b/llvm/lib/Transforms/Utils/LoopUnroll.cpp
@@ -329,12 +329,15 @@ void llvm::simplifyLoopAfterUnroll(Loop *L, bool SimplifyIVs, LoopInfo *LI,
 ///
 /// This utility preserves LoopInfo. It will also preserve ScalarEvolution and
 /// DominatorTree if they are non-null.
+///
+/// If RemainderLoop is non-null, it will receive the remainder loop (if
+/// required and not fully unrolled).
 LoopUnrollResult llvm::UnrollLoop(
     Loop *L, unsigned Count, unsigned TripCount, bool Force, bool AllowRuntime,
     bool AllowExpensiveTripCount, bool PreserveCondBr, bool PreserveOnlyFirst,
     unsigned TripMultiple, unsigned PeelCount, bool UnrollRemainder,
     LoopInfo *LI, ScalarEvolution *SE, DominatorTree *DT, AssumptionCache *AC,
-    OptimizationRemarkEmitter *ORE, bool PreserveLCSSA) {
+    OptimizationRemarkEmitter *ORE, bool PreserveLCSSA, Loop **RemainderLoop) {
 
   BasicBlock *Preheader = L->getLoopPreheader();
   if (!Preheader) {
@@ -468,7 +471,7 @@ LoopUnrollResult llvm::UnrollLoop(
   if (RuntimeTripCount && TripMultiple % Count != 0 &&
       !UnrollRuntimeLoopRemainder(L, Count, AllowExpensiveTripCount,
                                   EpilogProfitability, UnrollRemainder, LI, SE,
-                                  DT, AC, PreserveLCSSA)) {
+                                  DT, AC, PreserveLCSSA, RemainderLoop)) {
     if (Force)
       RuntimeTripCount = false;
     else {
diff --git a/llvm/lib/Transforms/Utils/LoopUnrollAndJam.cpp b/llvm/lib/Transforms/Utils/LoopUnrollAndJam.cpp
index 8949c603a84..b5d80f669fb 100644
--- a/llvm/lib/Transforms/Utils/LoopUnrollAndJam.cpp
+++ b/llvm/lib/Transforms/Utils/LoopUnrollAndJam.cpp
@@ -167,12 +167,14 @@ static void moveHeaderPhiOperandsToForeBlocks(BasicBlock *Header,
 
   isSafeToUnrollAndJam should be used prior to calling this to make sure the
   unrolling will be valid. Checking profitablility is also advisable.
+
+  If EpilogueLoop is non-null, it receives the epilogue loop (if it was
+  necessary to create one and not fully unrolled).
 */
-LoopUnrollResult
-llvm::UnrollAndJamLoop(Loop *L, unsigned Count, unsigned TripCount,
-                       unsigned TripMultiple, bool UnrollRemainder,
-                       LoopInfo *LI, ScalarEvolution *SE, DominatorTree *DT,
-                       AssumptionCache *AC, OptimizationRemarkEmitter *ORE) {
+LoopUnrollResult llvm::UnrollAndJamLoop(
+    Loop *L, unsigned Count, unsigned TripCount, unsigned TripMultiple,
+    bool UnrollRemainder, LoopInfo *LI, ScalarEvolution *SE, DominatorTree *DT,
+    AssumptionCache *AC, OptimizationRemarkEmitter *ORE, Loop **EpilogueLoop) {
 
   // When we enter here we should have already checked that it is safe
   BasicBlock *Header = L->getHeader();
@@ -196,7 +198,8 @@ llvm::UnrollAndJamLoop(Loop *L, unsigned Count, unsigned TripCount,
   if (TripMultiple == 1 || TripMultiple % Count != 0) {
     if (!UnrollRuntimeLoopRemainder(L, Count, /*AllowExpensiveTripCount*/ false,
                                     /*UseEpilogRemainder*/ true,
-                                    UnrollRemainder, LI, SE, DT, AC, true)) {
+                                    UnrollRemainder, LI, SE, DT, AC, true,
+                                    EpilogueLoop)) {
       LLVM_DEBUG(dbgs() << "Won't unroll-and-jam; remainder loop could not be "
                            "generated when assuming runtime trip count\n");
       return LoopUnrollResult::Unmodified;
diff --git a/llvm/lib/Transforms/Utils/LoopUnrollRuntime.cpp b/llvm/lib/Transforms/Utils/LoopUnrollRuntime.cpp
index 3361883acd0..3606ec4b9fc 100644
--- a/llvm/lib/Transforms/Utils/LoopUnrollRuntime.cpp
+++ b/llvm/lib/Transforms/Utils/LoopUnrollRuntime.cpp
@@ -380,6 +380,7 @@ CloneLoopBlocks(Loop *L, Value *NewIter, const bool CreateRemainderLoop,
   }
   if (CreateRemainderLoop) {
     Loop *NewLoop = NewLoops[L];
+    MDNode *LoopID = NewLoop->getLoopID();
     assert(NewLoop && "L should have been cloned");
 
     // Only add loop metadata if the loop is not going to be completely
@@ -387,6 +388,16 @@ CloneLoopBlocks(Loop *L, Value *NewIter, const bool CreateRemainderLoop,
     if (UnrollRemainder)
       return NewLoop;
 
+    Optional<MDNode *> NewLoopID = makeFollowupLoopID(
+        LoopID, {LLVMLoopUnrollFollowupAll, LLVMLoopUnrollFollowupRemainder});
+    if (NewLoopID.hasValue()) {
+      NewLoop->setLoopID(NewLoopID.getValue());
+
+      // Do not setLoopAlreadyUnrolled if loop attributes have been defined
+      // explicitly.
+      return NewLoop;
+    }
+
     // Add unroll disable metadata to disable future unrolling for this loop.
     NewLoop->setLoopAlreadyUnrolled();
     return NewLoop;
@@ -525,10 +536,10 @@ static bool canProfitablyUnrollMultiExitLoop(
 bool llvm::UnrollRuntimeLoopRemainder(Loop *L, unsigned Count,
                                       bool AllowExpensiveTripCount,
                                       bool UseEpilogRemainder,
-                                      bool UnrollRemainder,
-                                      LoopInfo *LI, ScalarEvolution *SE,
-                                      DominatorTree *DT, AssumptionCache *AC,
-                                      bool PreserveLCSSA) {
+                                      bool UnrollRemainder, LoopInfo *LI,
+                                      ScalarEvolution *SE, DominatorTree *DT,
+                                      AssumptionCache *AC, bool PreserveLCSSA,
+                                      Loop **ResultLoop) {
   LLVM_DEBUG(dbgs() << "Trying runtime unrolling on Loop: \n");
   LLVM_DEBUG(L->dump());
   LLVM_DEBUG(UseEpilogRemainder ? dbgs() << "Using epilog remainder.\n"
@@ -911,16 +922,20 @@ bool llvm::UnrollRuntimeLoopRemainder(Loop *L, unsigned Count,
       formDedicatedExitBlocks(remainderLoop, DT, LI, PreserveLCSSA);
   }
 
+  auto UnrollResult = LoopUnrollResult::Unmodified;
   if (remainderLoop && UnrollRemainder) {
     LLVM_DEBUG(dbgs() << "Unrolling remainder loop\n");
-    UnrollLoop(remainderLoop, /*Count*/ Count - 1, /*TripCount*/ Count - 1,
-               /*Force*/ false, /*AllowRuntime*/ false,
-               /*AllowExpensiveTripCount*/ false, /*PreserveCondBr*/ true,
-               /*PreserveOnlyFirst*/ false, /*TripMultiple*/ 1,
-               /*PeelCount*/ 0, /*UnrollRemainder*/ false, LI, SE, DT, AC,
-               /*ORE*/ nullptr, PreserveLCSSA);
+    UnrollResult =
+        UnrollLoop(remainderLoop, /*Count*/ Count - 1, /*TripCount*/ Count - 1,
+                   /*Force*/ false, /*AllowRuntime*/ false,
+                   /*AllowExpensiveTripCount*/ false, /*PreserveCondBr*/ true,
+                   /*PreserveOnlyFirst*/ false, /*TripMultiple*/ 1,
+                   /*PeelCount*/ 0, /*UnrollRemainder*/ false, LI, SE, DT, AC,
+                   /*ORE*/ nullptr, PreserveLCSSA);
   }
 
+  if (ResultLoop && UnrollResult != LoopUnrollResult::FullyUnrolled)
+    *ResultLoop = remainderLoop;
   NumRuntimeUnrolled++;
   return true;
 }
diff --git a/llvm/lib/Transforms/Utils/LoopUtils.cpp b/llvm/lib/Transforms/Utils/LoopUtils.cpp
index 249869e1bde..388553b1783 100644
--- a/llvm/lib/Transforms/Utils/LoopUtils.cpp
+++ b/llvm/lib/Transforms/Utils/LoopUtils.cpp
@@ -42,6 +42,8 @@ using namespace llvm::PatternMatch;
 
 #define DEBUG_TYPE "loop-utils"
 
+static const char *LLVMLoopDisableNonforced = "llvm.loop.disable_nonforced";
+
 bool llvm::formDedicatedExitBlocks(Loop *L, DominatorTree *DT, LoopInfo *LI,
                                    bool PreserveLCSSA) {
   bool Changed = false;
@@ -183,14 +185,8 @@ void llvm::initializeLoopPassPass(PassRegistry &Registry) {
   INITIALIZE_PASS_DEPENDENCY(ScalarEvolutionWrapperPass)
 }
 
-/// Find string metadata for loop
-///
-/// If it has a value (e.g. {"llvm.distribute", 1} return the value as an
-/// operand or null otherwise.  If the string metadata is not found return
-/// Optional's not-a-value.
-Optional<const MDOperand *> llvm::findStringMetadataForLoop(Loop *TheLoop,
-                                                            StringRef Name) {
-  MDNode *LoopID = TheLoop->getLoopID();
+static Optional<MDNode *> findOptionMDForLoopID(MDNode *LoopID,
+                                                StringRef Name) {
   // Return none if LoopID is false.
   if (!LoopID)
     return None;
@@ -209,18 +205,253 @@ Optional<const MDOperand *> llvm::findStringMetadataForLoop(Loop *TheLoop,
       continue;
     // Return true if MDString holds expected MetaData.
     if (Name.equals(S->getString()))
-      switch (MD->getNumOperands()) {
-      case 1:
-        return nullptr;
-      case 2:
-        return &MD->getOperand(1);
-      default:
-        llvm_unreachable("loop metadata has 0 or 1 operand");
-      }
+      return MD;
   }
   return None;
 }
 
+static Optional<MDNode *> findOptionMDForLoop(const Loop *TheLoop,
+                                              StringRef Name) {
+  return findOptionMDForLoopID(TheLoop->getLoopID(), Name);
+}
+
+/// Find string metadata for loop
+///
+/// If it has a value (e.g. {"llvm.distribute", 1} return the value as an
+/// operand or null otherwise.  If the string metadata is not found return
+/// Optional's not-a-value.
+Optional<const MDOperand *> llvm::findStringMetadataForLoop(Loop *TheLoop,
+                                                            StringRef Name) {
+  auto MD = findOptionMDForLoop(TheLoop, Name).getValueOr(nullptr);
+  if (!MD)
+    return None;
+  switch (MD->getNumOperands()) {
+  case 1:
+    return nullptr;
+  case 2:
+    return &MD->getOperand(1);
+  default:
+    llvm_unreachable("loop metadata has 0 or 1 operand");
+  }
+}
+
+static Optional<bool> getOptionalBoolLoopAttribute(const Loop *TheLoop,
+                                                   StringRef Name) {
+  Optional<MDNode *> MD = findOptionMDForLoop(TheLoop, Name);
+  if (!MD.hasValue())
+    return None;
+  MDNode *OptionNode = MD.getValue();
+  if (OptionNode == nullptr)
+    return None;
+  switch (OptionNode->getNumOperands()) {
+  case 1:
+    // When the value is absent it is interpreted as 'attribute set'.
+    return true;
+  case 2:
+    return mdconst::extract_or_null<ConstantInt>(
+        OptionNode->getOperand(1).get());
+  }
+  llvm_unreachable("unexpected number of options");
+}
+
+static bool getBooleanLoopAttribute(const Loop *TheLoop, StringRef Name) {
+  return getOptionalBoolLoopAttribute(TheLoop, Name).getValueOr(false);
+}
+
+llvm::Optional<int> llvm::getOptionalIntLoopAttribute(Loop *TheLoop,
+                                                      StringRef Name) {
+  const MDOperand *AttrMD =
+      findStringMetadataForLoop(TheLoop, Name).getValueOr(nullptr);
+  if (!AttrMD)
+    return None;
+
+  ConstantInt *IntMD = mdconst::extract_or_null<ConstantInt>(AttrMD->get());
+  if (!IntMD)
+    return None;
+
+  return IntMD->getSExtValue();
+}
+
+Optional<MDNode *> llvm::makeFollowupLoopID(
+    MDNode *OrigLoopID, ArrayRef<StringRef> FollowupOptions,
+    const char *InheritOptionsExceptPrefix, bool AlwaysNew) {
+  if (!OrigLoopID) {
+    if (AlwaysNew)
+      return nullptr;
+    return None;
+  }
+
+  assert(OrigLoopID->getOperand(0) == OrigLoopID);
+
+  bool InheritAllAttrs = !InheritOptionsExceptPrefix;
+  bool InheritSomeAttrs =
+      InheritOptionsExceptPrefix && InheritOptionsExceptPrefix[0] != '\0';
+  SmallVector<Metadata *, 8> MDs;
+  MDs.push_back(nullptr);
+
+  bool Changed = false;
+  if (InheritAllAttrs || InheritSomeAttrs) {
+    for (const MDOperand &Existing : drop_begin(OrigLoopID->operands(), 1)) {
+      MDNode *Op = cast<MDNode>(Existing.get());
+
+      auto InheritThisAttribute = [InheritSomeAttrs,
+                                   InheritOptionsExceptPrefix](MDNode *Op) {
+        if (!InheritSomeAttrs)
+          return false;
+
+        // Skip malformatted attribute metadata nodes.
+        if (Op->getNumOperands() == 0)
+          return true;
+        Metadata *NameMD = Op->getOperand(0).get();
+        if (!isa<MDString>(NameMD))
+          return true;
+        StringRef AttrName = cast<MDString>(NameMD)->getString();
+
+        // Do not inherit excluded attributes.
+        return !AttrName.startswith(InheritOptionsExceptPrefix);
+      };
+
+      if (InheritThisAttribute(Op))
+        MDs.push_back(Op);
+      else
+        Changed = true;
+    }
+  } else {
+    // Modified if we dropped at least one attribute.
+    Changed = OrigLoopID->getNumOperands() > 1;
+  }
+
+  bool HasAnyFollowup = false;
+  for (StringRef OptionName : FollowupOptions) {
+    MDNode *FollowupNode =
+        findOptionMDForLoopID(OrigLoopID, OptionName).getValueOr(nullptr);
+    if (!FollowupNode)
+      continue;
+
+    HasAnyFollowup = true;
+    for (const MDOperand &Option : drop_begin(FollowupNode->operands(), 1)) {
+      MDs.push_back(Option.get());
+      Changed = true;
+    }
+  }
+
+  // Attributes of the followup loop not specified explicity, so signal to the
+  // transformation pass to add suitable attributes.
+  if (!AlwaysNew && !HasAnyFollowup)
+    return None;
+
+  // If no attributes were added or remove, the previous loop Id can be reused.
+  if (!AlwaysNew && !Changed)
+    return OrigLoopID;
+
+  // No attributes is equivalent to having no !llvm.loop metadata at all.
+  if (MDs.size() == 1)
+    return nullptr;
+
+  // Build the new loop ID.
+  MDTuple *FollowupLoopID = MDNode::get(OrigLoopID->getContext(), MDs);
+  FollowupLoopID->replaceOperandWith(0, FollowupLoopID);
+  return FollowupLoopID;
+}
+
+bool llvm::hasDisableAllTransformsHint(const Loop *L) {
+  return getBooleanLoopAttribute(L, LLVMLoopDisableNonforced);
+}
+
+TransformationMode llvm::hasUnrollTransformation(Loop *L) {
+  if (getBooleanLoopAttribute(L, "llvm.loop.unroll.disable"))
+    return TM_SuppressedByUser;
+
+  Optional<int> Count =
+      getOptionalIntLoopAttribute(L, "llvm.loop.unroll.count");
+  if (Count.hasValue())
+    return Count.getValue() == 1 ? TM_SuppressedByUser : TM_ForcedByUser;
+
+  if (getBooleanLoopAttribute(L, "llvm.loop.unroll.enable"))
+    return TM_ForcedByUser;
+
+  if (getBooleanLoopAttribute(L, "llvm.loop.unroll.full"))
+    return TM_ForcedByUser;
+
+  if (hasDisableAllTransformsHint(L))
+    return TM_Disable;
+
+  return TM_Unspecified;
+}
+
+TransformationMode llvm::hasUnrollAndJamTransformation(Loop *L) {
+  if (getBooleanLoopAttribute(L, "llvm.loop.unroll_and_jam.disable"))
+    return TM_SuppressedByUser;
+
+  Optional<int> Count =
+      getOptionalIntLoopAttribute(L, "llvm.loop.unroll_and_jam.count");
+  if (Count.hasValue())
+    return Count.getValue() == 1 ? TM_SuppressedByUser : TM_ForcedByUser;
+
+  if (getBooleanLoopAttribute(L, "llvm.loop.unroll_and_jam.enable"))
+    return TM_ForcedByUser;
+
+  if (hasDisableAllTransformsHint(L))
+    return TM_Disable;
+
+  return TM_Unspecified;
+}
+
+TransformationMode llvm::hasVectorizeTransformation(Loop *L) {
+  Optional<bool> Enable =
+      getOptionalBoolLoopAttribute(L, "llvm.loop.vectorize.enable");
+
+  if (Enable == false)
+    return TM_SuppressedByUser;
+
+  Optional<int> VectorizeWidth =
+      getOptionalIntLoopAttribute(L, "llvm.loop.vectorize.width");
+  Optional<int> InterleaveCount =
+      getOptionalIntLoopAttribute(L, "llvm.loop.interleave.count");
+
+  if (Enable == true) {
+    // 'Forcing' vector width and interleave count to one effectively disables
+    // this tranformation.
+    if (VectorizeWidth == 1 && InterleaveCount == 1)
+      return TM_SuppressedByUser;
+    return TM_ForcedByUser;
+  }
+
+  if (getBooleanLoopAttribute(L, "llvm.loop.isvectorized"))
+    return TM_Disable;
+
+  if (VectorizeWidth == 1 && InterleaveCount == 1)
+    return TM_Disable;
+
+  if (VectorizeWidth > 1 || InterleaveCount > 1)
+    return TM_Enable;
+
+  if (hasDisableAllTransformsHint(L))
+    return TM_Disable;
+
+  return TM_Unspecified;
+}
+
+TransformationMode llvm::hasDistributeTransformation(Loop *L) {
+  if (getBooleanLoopAttribute(L, "llvm.loop.distribute.enable"))
+    return TM_ForcedByUser;
+
+  if (hasDisableAllTransformsHint(L))
+    return TM_Disable;
+
+  return TM_Unspecified;
+}
+
+TransformationMode llvm::hasLICMVersioningTransformation(Loop *L) {
+  if (getBooleanLoopAttribute(L, "llvm.loop.licm_versioning.disable"))
+    return TM_SuppressedByUser;
+
+  if (hasDisableAllTransformsHint(L))
+    return TM_Disable;
+
+  return TM_Unspecified;
+}
+
 /// Does a BFS from a given node to all of its children inside a given loop.
 /// The returned vector of nodes includes the starting point.
 SmallVector<DomTreeNode *, 16>