[LV][RFC] Generating conditional VPBB that will be skip when the mask is inactive in VPlan. #141900
+350
−0
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This patch add a need attribut in TTI to let LV knows which is better. Default value of preferFlattenControlFlow is true to match current TTI implementation. If preferFlattenControlFlow() return false, LV will try to generate conditional VPBB if possible.
This patch add the transformation that convert flatten control flow
with conditional vector basic block.
This transformation can help program skip masked operations without any
active lane.
First, this transformation will collect all masked stores and operands
bottom-up. And put these msaked operations into a new vector basic
block.
Second, this transformation will split original vector loop and insert
the new basic block between split blocks. And update the conditional
branch in the orignal blocks.
E.g.
Before: {
vector.loop:
...
BranchOnCount %IV, %TC
Successors middle.block, vector.loop
}
After: {
vector.loop:
...
%any.active.mask = any-of(%mask)
BranchOnCount %any.active.mask, 0
Successors vector.loop.split, vector.if.bb
vector.if.bb:
... (Masked operations)
Successors vector.loop.split
vector.loop.split:
...
BranchOnCount %IV, %TC
Successors middle.block, vector.loop
}
ElvisWang123
requested review from
fhahn,
preames,
npanchen,
alexey-bataev and
david-arm
|
@llvm/pr-subscribers-vectorizers @llvm/pr-subscribers-llvm-analysis Author: Elvis Wang (ElvisWang123) ChangesThis patch create an overview of VPlan transformation of creating conditional VPBB and will be split off into multiple patches later. This patch add the transformation that convert flatten control flow First, this transformation will collect all masked stores and operands Second, this transformation will split original vector loop and insert E.g. After: { vector.if.bb: vector.loop.split: Co-authored-by: nikolaypanchenko <kolya.panchenko@sifive.com> Patch is 20.32 KiB, truncated to 20.00 KiB below, full version: https://github.com/llvm/llvm-project/pull/141900.diff 9 Files Affected:
diff --git a/llvm/include/llvm/Analysis/TargetTransformInfo.h b/llvm/include/llvm/Analysis/TargetTransformInfo.h
index 8f4ce80ada5ed..20c4b52098770 100644
--- a/llvm/include/llvm/Analysis/TargetTransformInfo.h
+++ b/llvm/include/llvm/Analysis/TargetTransformInfo.h
@@ -1822,6 +1822,10 @@ class TargetTransformInfo {
/// otherwise scalar epilogue loop.
LLVM_ABI bool preferEpilogueVectorization() const;
+ /// Return true if the loop vectorizer shoud consider vectorizing with
+ /// flattern control flow, otherwise create conditional vector basic block.
+ bool preferFlattenControlFlow() const;
+
/// \returns True if the target wants to expand the given reduction intrinsic
/// into a shuffle sequence.
LLVM_ABI bool shouldExpandReduction(const IntrinsicInst *II) const;
diff --git a/llvm/include/llvm/Analysis/TargetTransformInfoImpl.h b/llvm/include/llvm/Analysis/TargetTransformInfoImpl.h
index a80b4c5179bad..1211c80f8ff51 100644
--- a/llvm/include/llvm/Analysis/TargetTransformInfoImpl.h
+++ b/llvm/include/llvm/Analysis/TargetTransformInfoImpl.h
@@ -1091,6 +1091,8 @@ class TargetTransformInfoImplBase {
virtual bool preferEpilogueVectorization() const { return true; }
+ virtual bool preferFlattenControlFlow() const { return true; }
+
virtual bool shouldExpandReduction(const IntrinsicInst *II) const {
return true;
}
diff --git a/llvm/include/llvm/CodeGen/BasicTTIImpl.h b/llvm/include/llvm/CodeGen/BasicTTIImpl.h
index ff8778168686d..5000bef968213 100644
--- a/llvm/include/llvm/CodeGen/BasicTTIImpl.h
+++ b/llvm/include/llvm/CodeGen/BasicTTIImpl.h
@@ -777,6 +777,10 @@ class BasicTTIImplBase : public TargetTransformInfoImplCRTPBase<T> {
return BaseT::preferPredicateOverEpilogue(TFI);
}
+ bool preferFlattenControlFlow() const override {
+ return thisT()->preferFlattenControlFlow();
+ }
+
TailFoldingStyle
getPreferredTailFoldingStyle(bool IVUpdateMayOverflow = true) const override {
return BaseT::getPreferredTailFoldingStyle(IVUpdateMayOverflow);
diff --git a/llvm/lib/Analysis/TargetTransformInfo.cpp b/llvm/lib/Analysis/TargetTransformInfo.cpp
index 0f857399660fe..d455e1dc63c13 100644
--- a/llvm/lib/Analysis/TargetTransformInfo.cpp
+++ b/llvm/lib/Analysis/TargetTransformInfo.cpp
@@ -371,6 +371,10 @@ bool TargetTransformInfo::preferPredicateOverEpilogue(
return TTIImpl->preferPredicateOverEpilogue(TFI);
}
+bool TargetTransformInfo::preferFlattenControlFlow() const {
+ return TTIImpl->preferFlattenControlFlow();
+}
+
TailFoldingStyle TargetTransformInfo::getPreferredTailFoldingStyle(
bool IVUpdateMayOverflow) const {
return TTIImpl->getPreferredTailFoldingStyle(IVUpdateMayOverflow);
diff --git a/llvm/lib/Target/RISCV/RISCVTargetTransformInfo.h b/llvm/lib/Target/RISCV/RISCVTargetTransformInfo.h
index 0a784461d67bf..412dd8ae2b991 100644
--- a/llvm/lib/Target/RISCV/RISCVTargetTransformInfo.h
+++ b/llvm/lib/Target/RISCV/RISCVTargetTransformInfo.h
@@ -146,6 +146,8 @@ class RISCVTTIImpl : public BasicTTIImplBase<RISCVTTIImpl> {
return false;
}
+ bool preferFlattenControlFlow() const { return false; }
+
InstructionCost
getMaskedMemoryOpCost(unsigned Opcode, Type *Src, Align Alignment,
unsigned AddressSpace,
diff --git a/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp b/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp
index 2fe59a464457f..a834b6106f028 100644
--- a/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp
+++ b/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp
@@ -351,6 +351,11 @@ static cl::opt<bool> PreferPredicatedReductionSelect(
cl::desc(
"Prefer predicating a reduction operation over an after loop select."));
+static cl::opt<bool>
+ PreferFlattenControlFlow("prefer-flatten-control-flow", cl::init(true),
+ cl::Hidden,
+ cl::desc("Prefer flatten control flow."));
+
cl::opt<bool> llvm::EnableVPlanNativePath(
"enable-vplan-native-path", cl::Hidden,
cl::desc("Enable VPlan-native vectorization path with "
@@ -9287,6 +9292,9 @@ VPlanPtr LoopVectorizationPlanner::tryToBuildVPlanWithVPRecipes(
}
VPlanTransforms::optimizeInductionExitUsers(*Plan, IVEndValues);
+ if (!PreferFlattenControlFlow && !TTI.preferFlattenControlFlow())
+ VPlanTransforms::optimizeConditionalVPBB(*Plan);
+
assert(verifyVPlanIsValid(*Plan) && "VPlan is invalid");
return Plan;
}
diff --git a/llvm/lib/Transforms/Vectorize/VPlanTransforms.cpp b/llvm/lib/Transforms/Vectorize/VPlanTransforms.cpp
index dc5be520505eb..e1c8690986599 100644
--- a/llvm/lib/Transforms/Vectorize/VPlanTransforms.cpp
+++ b/llvm/lib/Transforms/Vectorize/VPlanTransforms.cpp
@@ -2087,6 +2087,23 @@ void VPlanTransforms::addActiveLaneMask(
HeaderMask->replaceAllUsesWith(LaneMask);
}
+static bool replaceHeaderMaskToEVL(VPValue *HeaderMask, VPRecipeBase *R) {
+ using namespace llvm::VPlanPatternMatch;
+ VPValue *EdgeMask;
+ if (!R)
+ return false;
+ if (match(R, m_Binary<VPInstruction::BranchOnCount>(
+ m_VPInstruction<VPInstruction::AnyOf>(
+ m_Binary<VPInstruction::LogicalAnd>(
+ m_Specific(HeaderMask), m_VPValue(EdgeMask))),
+ m_VPValue()))) {
+
+ cast<VPInstruction>(R->getOperand(0))->setOperand(0, EdgeMask);
+ return true;
+ }
+ return false;
+}
+
/// Try to convert \p CurRecipe to a corresponding EVL-based recipe. Returns
/// nullptr if no EVL-based recipe could be created.
/// \p HeaderMask Header Mask.
@@ -2202,6 +2219,8 @@ static void transformRecipestoEVLRecipes(VPlan &Plan, VPValue &EVL) {
for (VPValue *HeaderMask : collectAllHeaderMasks(Plan)) {
for (VPUser *U : collectUsersRecursively(HeaderMask)) {
auto *CurRecipe = cast<VPRecipeBase>(U);
+ if (replaceHeaderMaskToEVL(HeaderMask, CurRecipe))
+ continue;
VPRecipeBase *EVLRecipe = createEVLRecipe(
HeaderMask, *CurRecipe, TypeInfo, *AllOneMask, EVL, PrevEVL);
if (!EVLRecipe)
@@ -3202,3 +3221,160 @@ void VPlanTransforms::narrowInterleaveGroups(VPlan &Plan, ElementCount VF,
Plan.getOrAddLiveIn(ConstantInt::get(CanIV->getScalarType(), 1)));
removeDeadRecipes(Plan);
}
+
+void VPlanTransforms::optimizeConditionalVPBB(VPlan &Plan) {
+
+ VPDominatorTree VPDT;
+ VPDT.recalculate(Plan);
+
+ SmallVector<VPValue *> HeaderMasks = collectAllHeaderMasks(Plan);
+
+ // Get the mask from the store recipes.
+ auto GetMask = [&HeaderMasks](VPRecipeBase &R) -> VPValue * {
+ using namespace llvm::VPlanPatternMatch;
+ if (isa<VPWidenStoreRecipe, VPWidenStoreEVLRecipe>(R)) {
+ VPValue *OrigMask = cast<VPWidenMemoryRecipe>(R).getMask();
+ if (!OrigMask)
+ return OrigMask;
+
+ if (any_of(HeaderMasks, [OrigMask](VPValue *HeaderMask) {
+ return OrigMask == HeaderMask;
+ }))
+ return nullptr;
+
+ // Match active.lane.mask.
+ if (match(OrigMask, m_VPInstruction<VPInstruction::ActiveLaneMask>(
+ m_VPValue(), m_VPValue())))
+ return nullptr;
+
+ return OrigMask;
+ }
+ return nullptr;
+ };
+
+ // First, collect all masked stores.
+ SmallVector<std::pair<VPRecipeBase *, VPValue *>> MaskedStores;
+ ReversePostOrderTraversal<VPBlockDeepTraversalWrapper<VPBlockBase *>> RPOT(
+ Plan.getEntry());
+ for (VPBasicBlock *VPBB : VPBlockUtils::blocksOnly<VPBasicBlock>(RPOT)) {
+ for (VPRecipeBase &R : *VPBB) {
+ if (VPValue *Mask = GetMask(R))
+ MaskedStores.emplace_back(&R, Mask);
+ }
+ }
+
+ DenseSet<VPRecipeBase *> Candidates;
+ auto AddOperandsToCandidates = [&Candidates](VPRecipeBase *R) {
+ for (VPValue *Op : R->operands())
+ if (VPRecipeBase *OpR = Op->getDefiningRecipe())
+ Candidates.insert(OpR);
+ };
+
+ SmallVector<SetVector<VPRecipeBase *>> Tries;
+ while (!MaskedStores.empty()) {
+ auto [LR, M] = MaskedStores.pop_back_val();
+ Candidates.clear();
+ AddOperandsToCandidates(LR);
+
+ SetVector<VPRecipeBase *> CurrentTree;
+ CurrentTree.insert(LR);
+
+ VPBasicBlock *MaskBlock =
+ M->hasDefiningRecipe() ? M->getDefiningRecipe()->getParent() : nullptr;
+ auto End = MaskBlock == LR->getParent()
+ ? M->getDefiningRecipe()->getReverseIterator()
+ : LR->getParent()->getFirstNonPhi()->getReverseIterator();
+ // Greedily add all recipes that are used to compute stored value to the
+ // tree. All users of the added recipe must dominate the store
+ // recipe.
+ for (VPRecipeBase &R : make_range(LR->getReverseIterator(), End)) {
+ // Recipe is not a part of the tree
+ if (!Candidates.contains(&R))
+ continue;
+
+ if (any_of(R.definedValues(), [&LR = LR, &VPDT](VPValue *Def) {
+ for (VPUser *U : Def->users()) {
+ if (auto *UR = dyn_cast<VPRecipeBase>(U)) {
+ if (UR == LR || VPDT.properlyDominates(UR, LR))
+ continue;
+ }
+ return true;
+ }
+ return false;
+ }))
+ continue;
+
+ CurrentTree.insert(&R);
+ AddOperandsToCandidates(&R);
+ }
+ // Previous traversal could add recipes that are used by non-added recipes,
+ // thus need to be removed from the list.
+ DenseSet<VPRecipeBase *> ToRemove;
+ bool Changed;
+ do {
+ Changed = false;
+ for (VPRecipeBase *R : CurrentTree) {
+ if (ToRemove.contains(R))
+ continue;
+ if (any_of(R->definedValues(), [&](VPValue *Def) {
+ for (VPUser *U : Def->users()) {
+ if (auto *UR = dyn_cast<VPRecipeBase>(U))
+ if (!CurrentTree.contains(UR) || ToRemove.contains(UR))
+ return true;
+ }
+ return false;
+ })) {
+ Changed = true;
+ ToRemove.insert(R);
+ }
+ }
+ } while (Changed);
+
+ for (VPRecipeBase *R : ToRemove)
+ CurrentTree.remove(R);
+
+ if (CurrentTree.size() > 1)
+ Tries.push_back(CurrentTree);
+ }
+ for (const auto &List : Tries) {
+ VPRecipeBase *LR = List.front();
+ VPValue *M = cast<VPWidenMemoryRecipe>(LR)->getMask();
+ assert(M && "Mask VPValue must exist at this point");
+ auto Recipes = reverse(List.getArrayRef());
+
+ // Split current basic block at LR point so that VPConditionalRegionBlock
+ // can be added inbetween.
+ VPBasicBlock *ParentBB = LR->getParent();
+ VPBasicBlock *ContBB = ParentBB->splitAt(LR->getIterator());
+
+ // Create VPBB and insert it between ParentBB and ContBB.
+ VPBasicBlock *IfBB = Plan.createVPBasicBlock("vector.if.bb");
+ VPBlockUtils::insertBlockAfter(IfBB, ParentBB);
+ if (ContBB->getNumSuccessors() == 0)
+ ParentBB->getEnclosingLoopRegion()->setExiting(ContBB);
+
+ // Copy recipes into conditional block.
+ for (VPRecipeBase *R : Recipes)
+ R->moveBefore(*IfBB, IfBB->end());
+
+ // Add the condition and brach in the parent block.
+ auto *ActiveLane =
+ new VPInstruction(VPInstruction::AnyOf, {M}, nullptr, "any.of.mask");
+
+ Type *CanonicalIVType = Plan.getCanonicalIV()->getScalarType();
+ LLVMContext &Ctx = CanonicalIVType->getContext();
+ VPValue *Zero =
+ Plan.getOrAddLiveIn(ConstantInt::get(Type::getInt1Ty(Ctx), 0));
+
+ auto *BranchOnCount =
+ new VPInstruction(VPInstruction::BranchOnCount, {ActiveLane, Zero});
+ ParentBB->appendRecipe(ActiveLane);
+ ParentBB->appendRecipe(BranchOnCount);
+
+ // Set proper predecessor and successors for modifed basicblocks.
+ ParentBB->clearSuccessors();
+ ParentBB->setTwoSuccessors(ContBB, IfBB);
+ ContBB->clearPredecessors();
+ ContBB->setPredecessors({ParentBB, IfBB});
+ }
+}
diff --git a/llvm/lib/Transforms/Vectorize/VPlanTransforms.h b/llvm/lib/Transforms/Vectorize/VPlanTransforms.h
index 34e2de4eb3b74..0e535d6bf7c36 100644
--- a/llvm/lib/Transforms/Vectorize/VPlanTransforms.h
+++ b/llvm/lib/Transforms/Vectorize/VPlanTransforms.h
@@ -234,6 +234,29 @@ struct VPlanTransforms {
/// removed in the future.
static DenseMap<VPBasicBlock *, VPValue *>
introduceMasksAndLinearize(VPlan &Plan, bool FoldTail);
+
+ /// Try to convet flatten control flow to the conditional vector basic block.
+ /// If no active bits in the mask, will skip all the masked operations.
+ /// This transformation will collect all masked operations bottom-up from the
+ /// masked stores and put all of masked operations in a new vector basic
+ /// block. This original vector.loop will be split and the new created basic
+ /// block will inserted in between.
+ ///
+ /// After transformation the vplan will looks like.
+ /// vector.loop:
+ /// ...
+ /// %any.active.mask = any-of(%Mask)
+ /// Branch-On-Count %any.active.mask, 0
+ /// successors vector.loop.split, vector.if.bb
+ ///
+ /// vector.if.bb:
+ /// (Masked operations)
+ /// successors vector.loop.split
+ ///
+ /// vector.loop.split:
+ /// ...
+ /// successors middle.block, vector.loop
+ static void optimizeConditionalVPBB(VPlan &Plan);
};
} // namespace llvm
diff --git a/llvm/test/Transforms/LoopVectorize/RISCV/vplan-conditional-basic-block.ll b/llvm/test/Transforms/LoopVectorize/RISCV/vplan-conditional-basic-block.ll
new file mode 100644
index 0000000000000..2b8f2542ab544
--- /dev/null
+++ b/llvm/test/Transforms/LoopVectorize/RISCV/vplan-conditional-basic-block.ll
@@ -0,0 +1,127 @@
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 5
+; RUN: opt -p loop-vectorize -force-vector-width=4 -S -mtriple=riscv64 -mattr=+v -prefer-flatten-control-flow=false %s | FileCheck %s
+
+define void @test(i32 %control1, i32 %control2, i32 %target, i32 %reg.4.val, ptr %reg.24.val) {
+; CHECK-LABEL: define void @test(
+; CHECK-SAME: i32 [[CONTROL1:%.*]], i32 [[CONTROL2:%.*]], i32 [[TARGET:%.*]], i32 [[REG_4_VAL:%.*]], ptr [[REG_24_VAL:%.*]]) #[[ATTR0:[0-9]+]] {
+; CHECK-NEXT: [[ENTRY:.*:]]
+; CHECK-NEXT: [[CMP1:%.*]] = icmp sgt i32 [[REG_4_VAL]], 0
+; CHECK-NEXT: br i1 [[CMP1]], label %[[FOR_BODY_LR_PH:.*]], label %[[FOR_END:.*]]
+; CHECK: [[FOR_BODY_LR_PH]]:
+; CHECK-NEXT: [[SH_PROM:%.*]] = zext nneg i32 [[CONTROL1]] to i64
+; CHECK-NEXT: [[SHL:%.*]] = shl nuw i64 1, [[SH_PROM]]
+; CHECK-NEXT: [[SH_PROM5:%.*]] = zext nneg i32 [[CONTROL2]] to i64
+; CHECK-NEXT: [[SHL6:%.*]] = shl nuw i64 1, [[SH_PROM5]]
+; CHECK-NEXT: [[SH_PROM10:%.*]] = zext nneg i32 [[TARGET]] to i64
+; CHECK-NEXT: [[SHL11:%.*]] = shl nuw nsw i64 1, [[SH_PROM10]]
+; CHECK-NEXT: [[WIDE_TRIP_COUNT:%.*]] = zext nneg i32 [[REG_4_VAL]] to i64
+; CHECK-NEXT: [[TMP0:%.*]] = freeze i64 [[SHL6]]
+; CHECK-NEXT: [[TMP1:%.*]] = or i64 [[SHL]], [[TMP0]]
+; CHECK-NEXT: [[MIN_ITERS_CHECK:%.*]] = icmp ult i64 [[WIDE_TRIP_COUNT]], 8
+; CHECK-NEXT: br i1 [[MIN_ITERS_CHECK]], label %[[SCALAR_PH:.*]], label %[[VECTOR_PH:.*]]
+; CHECK: [[VECTOR_PH]]:
+; CHECK-NEXT: [[N_MOD_VF:%.*]] = urem i64 [[WIDE_TRIP_COUNT]], 8
+; CHECK-NEXT: [[N_VEC:%.*]] = sub i64 [[WIDE_TRIP_COUNT]], [[N_MOD_VF]]
+; CHECK-NEXT: [[BROADCAST_SPLATINSERT:%.*]] = insertelement <4 x i64> poison, i64 [[SHL11]], i64 0
+; CHECK-NEXT: [[BROADCAST_SPLAT:%.*]] = shufflevector <4 x i64> [[BROADCAST_SPLATINSERT]], <4 x i64> poison, <4 x i32> zeroinitializer
+; CHECK-NEXT: [[BROADCAST_SPLATINSERT1:%.*]] = insertelement <4 x i64> poison, i64 [[TMP1]], i64 0
+; CHECK-NEXT: [[BROADCAST_SPLAT2:%.*]] = shufflevector <4 x i64> [[BROADCAST_SPLATINSERT1]], <4 x i64> poison, <4 x i32> zeroinitializer
+; CHECK-NEXT: br label %[[VECTOR_BODY:.*]]
+; CHECK: [[VECTOR_BODY]]:
+; CHECK-NEXT: [[INDEX:%.*]] = phi i64 [ 0, %[[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], %[[IF_THEN9_SPLIT:.*]] ]
+; CHECK-NEXT: [[TMP2:%.*]] = getelementptr i64, ptr [[REG_24_VAL]], i64 [[INDEX]]
+; CHECK-NEXT: [[TMP3:%.*]] = getelementptr inbounds i64, ptr [[TMP2]], i32 0
+; CHECK-NEXT: [[TMP4:%.*]] = getelementptr inbounds i64, ptr [[TMP2]], i32 4
+; CHECK-NEXT: [[WIDE_LOAD:%.*]] = load <4 x i64>, ptr [[TMP3]], align 8
+; CHECK-NEXT: [[WIDE_LOAD3:%.*]] = load <4 x i64>, ptr [[TMP4]], align 8
+; CHECK-NEXT: [[TMP5:%.*]] = and <4 x i64> [[WIDE_LOAD]], [[BROADCAST_SPLAT2]]
+; CHECK-NEXT: [[TMP6:%.*]] = and <4 x i64> [[WIDE_LOAD3]], [[BROADCAST_SPLAT2]]
+; CHECK-NEXT: [[TMP7:%.*]] = icmp eq <4 x i64> [[TMP5]], [[BROADCAST_SPLAT2]]
+; CHECK-NEXT: [[TMP8:%.*]] = icmp eq <4 x i64> [[TMP6]], [[BROADCAST_SPLAT2]]
+; CHECK-NEXT: [[TMP9:%.*]] = xor <4 x i64> [[WIDE_LOAD]], [[BROADCAST_SPLAT]]
+; CHECK-NEXT: [[TMP10:%.*]] = xor <4 x i64> [[WIDE_LOAD3]], [[BROADCAST_SPLAT]]
+; CHECK-NEXT: [[TMP13:%.*]] = call i1 @llvm.vector.reduce.or.v4i1(<4 x i1> [[TMP7]])
+; CHECK-NEXT: [[TMP14:%.*]] = icmp eq i1 [[TMP13]], false
+; CHECK-NEXT: br i1 [[TMP14]], label %[[IF_THEN9_SPLIT]], label %[[VECTOR_IF_BB:.*]]
+; CHECK: [[VECTOR_IF_BB]]:
+; CHECK-NEXT: [[TMP11:%.*]] = getelementptr i64, ptr [[TMP2]], i32 0
+; CHECK-NEXT: [[TMP12:%.*]] = getelementptr i64, ptr [[TMP2]], i32 4
+; CHECK-NEXT: call void @llvm.masked.store.v4i64.p0(<4 x i64> [[TMP9]], ptr [[TMP11]], i32 8, <4 x i1> [[TMP7]])
+; CHECK-NEXT: call void @llvm.masked.store.v4i64.p0(<4 x i64> [[TMP10]], ptr [[TMP12]], i32 8, <4 x i1> [[TMP8]])
+; CHECK-NEXT: br label %[[IF_THEN9_SPLIT]]
+; CHECK: [[IF_THEN9_SPLIT]]:
+; CHECK-NEXT: [[INDEX_NEXT]] = add nuw i64 [[INDEX]], 8
+; CHECK-NEXT: [[TMP26:%.*]] = icmp eq i64 [[INDEX_NEXT]], [[N_VEC]]
+; CHECK-NEXT: br i1 [[TMP26]], label %[[MIDDLE_BLOCK:.*]], label %[[VECTOR_BODY]], !llvm.loop [[LOOP0:![0-9]+]]
+; CHECK: [[MIDDLE_BLOCK]]:
+; CHECK-NEXT: [[CMP_N:%.*]] = icmp eq i64 [[WIDE_TRIP_COUNT]], [[N_VEC]]
+; CHECK-NEXT: br i1 [[CMP_N]], label %[[FOR_END_LOOPEXIT:.*]], label %[[SCALAR_PH]]
+; CHECK: [[SCALAR_PH]]:
+; CHECK-NEXT: [[BC_RESUME_VAL:%.*]] = phi i64 [ [[N_VEC]], %[[MIDDLE_BLOCK]] ], [ 0, %[[FOR_BODY_LR_PH]] ]
+; CHECK-NEXT: br label %[[FOR_BODY:.*]]
+; CHECK: [[FOR_BODY]]:
+; CHECK-NEXT: [[INDVARS_IV:%.*]] = phi i64 [ [[BC_RESUME_VAL]], %[[SCALAR_PH]] ], [ [[INDVARS_IV_NEXT:%.*]], %[[FOR_INC:.*]] ]
+; CHECK-NEXT: [[ARRAYIDX:%.*]] = getelementptr inbounds i64, ptr [[REG_24_VAL]], i64 [[INDVARS_IV]]
+; CHECK-NEXT: [[TMP27:%.*]] = load i64, ptr [[ARRAYIDX]], align 8
+; CHECK-NEXT: [[TMP28:%.*]] = and i64 [[TMP27]], [[TMP1]]
+; CHECK-NEXT: [[OR_COND_NOT:%.*]] = icmp eq i64 [[TMP28]], [[TMP1]]
+; CHECK-NEXT: br i1 [[OR_COND_NOT]], label %[[IF_THEN9:.*]], label %[[FOR_INC]]
+; CHECK: [[IF_THEN9]]:
+; CHECK-NEXT: [[XOR:%.*]] = xor i64 [[TMP27]], [[SHL11]]
+; CHECK-NEXT: store i64 [[XOR]], ptr [[ARRAYIDX]], align 8
+; CHECK-NEXT: br label %[[FOR_INC]]
+; CHECK: [[FOR_INC]]:
+; CHECK-NEXT: [[INDVARS_IV_NEXT]] = add nuw nsw i64 [[INDVARS_IV]], 1
+; CHECK-NEXT: [[EXITCOND_NOT:%.*]] = icmp eq i64 [[INDVARS_IV_NEXT]], [[WIDE_TRIP_COUNT]]
+; CHECK-NEXT: br i1 [[EXITCOND_NOT]], label %[[FOR_END_LOOPEXIT]], label %[[FOR_BODY]], !llvm.loop [[LOOP3:![0-9]+]]
+; CHECK: [[FOR_END_LOOPEXIT]]:
+; CHECK-NEXT: br label %[[FOR_END]]
+; CHECK: [[FOR_END]]:
+; CHECK-NEXT: ret void
+;
+entry:
+ %cmp1 = icmp sgt i32 %reg.4.val, 0
+ br i1 %cmp1, label %for.body.lr.ph, label %for.end
+
+for.body.lr.ph:
+ %sh_prom = zext nneg i32 %control1 to i64
+ %shl = shl nuw i64 1, %sh_prom
+ %sh_prom5 = zext nneg i32 %control2 to i64
+ %shl6 = shl nuw i64 1, %sh_prom5
+ %sh_prom10 = zext nneg i32 %target to i64
+ %shl11 = shl nuw nsw i64 1, %sh_prom10
+ %wide.trip.count = zext nneg i32 %reg.4.val to i64
+ %0 = freeze i64 %shl6
+ %1 = or i64 %shl, %0
+ br label %for.body
+
+for.body:
+ %indvars.iv = phi i64 [ 0, %for.body.lr.ph ], [ %indvars.iv.next, %for.inc ]
+ %arrayidx = getelementptr inbounds i64, ptr %reg.24.val, i64 %indvars.iv
+ %2 = load i64, ptr %arrayidx, align 8
+ %3 = and i64 %2, %1
+ %or.cond.not = icmp eq i64 %3, %1
+ br i1 %or.cond.not, label %if.then9, label %for.inc
+
+if.then9:
+ %xor = xor i64 %2, %shl11
+ store i64 %xor, ptr %arrayidx, align 8
+ br label %for.inc
+
+for.inc:
+ %indvars.iv.next = add nuw nsw i64 %indvars.iv, 1
+ %exitcond.not = icmp eq i64 %indvars.iv.next, %wide.trip.count
+ br i1 %exitcond.not, label %for.end.loopexit, label %for.body
+
+for.end.loope...
[truncated]
|
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
This patch create an overview of VPlan transformation of creating conditional VPBB and will be split off into multiple patches later.
RFC: https://discourse.llvm.org/t/rfc-lv-generating-conditional-vpbb-that-will-be-skip-when-the-mask-is-inactive-in-vplan/86591
This patch add the transformation that convert flatten control flow
with conditional vector basic block.
This transformation can help program skip masked operations without any
active lane.
First, this transformation will collect all masked stores and operands
bottom-up. And put these masked operations into a new vector basic
block.
Second, this transformation will split original vector loop and insert
the new basic block between split blocks. And update the conditional
branch in the original blocks.
E.g.
Before: {
vector.loop:
...
BranchOnCount %IV, %TC
Successors middle.block, vector.loop
}
After: {
vector.loop:
...
%any.active.mask = any-of(%mask)
BranchOnCount %any.active.mask, 0
Successors vector.loop.split, vector.if.bb
vector.if.bb:
... (Masked operations)
Successors vector.loop.split
vector.loop.split:
...
BranchOnCount %IV, %TC
Successors middle.block, vector.loop
}
Co-authored-by: nikolaypanchenko kolya.panchenko@sifive.com