[llvm] [VPlan] Implement VPWidenCastRecipe::computeCost(). (NFCI) (PR #111339)
Elvis Wang via llvm-commits
llvm-commits at lists.llvm.org
Mon Oct 21 18:40:46 PDT 2024
https://github.com/ElvisWang123 updated https://github.com/llvm/llvm-project/pull/111339
>From 8a8462e74b323e8cc1a7a8f5ea9bc9e2a7101e97 Mon Sep 17 00:00:00 2001
From: Elvis Wang <elvis.wang at sifive.com>
Date: Sun, 6 Oct 2024 22:16:14 -0700
Subject: [PATCH 1/7] [VPlan] Implement VPWidenCastRecipe::computeCost().
(NFCI)
This patch implement VPWidenCastRecipe::computeCost() and skip more
cast recipies in the in-loop reduction.
---
.../Transforms/Vectorize/LoopVectorize.cpp | 19 +++++++-
llvm/lib/Transforms/Vectorize/VPlan.h | 4 ++
.../lib/Transforms/Vectorize/VPlanRecipes.cpp | 43 +++++++++++++++++++
llvm/lib/Transforms/Vectorize/VPlanValue.h | 2 +-
4 files changed, 66 insertions(+), 2 deletions(-)
diff --git a/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp b/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp
index 027ee21527d228..0c8e2d1a2d9895 100644
--- a/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp
+++ b/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp
@@ -7258,12 +7258,29 @@ LoopVectorizationPlanner::precomputeCosts(VPlan &Plan, ElementCount VF,
const auto &ChainOps = RdxDesc.getReductionOpChain(RedPhi, OrigLoop);
SetVector<Instruction *> ChainOpsAndOperands(ChainOps.begin(),
ChainOps.end());
+ auto isZExtOrSExt = [](const unsigned Opcode) -> bool {
+ return Opcode == Instruction::ZExt || Opcode == Instruction::SExt;
+ };
// Also include the operands of instructions in the chain, as the cost-model
// may mark extends as free.
+ //
+ // For ARM, some of the instruction can folded into the reducion
+ // instruction. So we need to mark all folded instructions free.
+ // For example: We can fold reduce(mul(ext(A), ext(B))) into one
+ // instruction.
for (auto *ChainOp : ChainOps) {
for (Value *Op : ChainOp->operands()) {
- if (auto *I = dyn_cast<Instruction>(Op))
+ if (auto *I = dyn_cast<Instruction>(Op)) {
ChainOpsAndOperands.insert(I);
+ if (I->getOpcode() == Instruction::Mul) {
+ auto *Ext0 = dyn_cast<Instruction>(I->getOperand(0));
+ if (Ext0 && isZExtOrSExt(Ext0->getOpcode()))
+ ChainOpsAndOperands.insert(Ext0);
+ auto *Ext1 = dyn_cast<Instruction>(I->getOperand(1));
+ if (Ext1 && isZExtOrSExt(Ext1->getOpcode()))
+ ChainOpsAndOperands.insert(Ext1);
+ }
+ }
}
}
diff --git a/llvm/lib/Transforms/Vectorize/VPlan.h b/llvm/lib/Transforms/Vectorize/VPlan.h
index 68a62638b9d588..94b7ddefbbb961 100644
--- a/llvm/lib/Transforms/Vectorize/VPlan.h
+++ b/llvm/lib/Transforms/Vectorize/VPlan.h
@@ -1571,6 +1571,10 @@ class VPWidenCastRecipe : public VPRecipeWithIRFlags {
/// Produce widened copies of the cast.
void execute(VPTransformState &State) override;
+ /// Return the cost of this VPWidenCastRecipe.
+ InstructionCost computeCost(ElementCount VF,
+ VPCostContext &Ctx) const override;
+
#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
/// Print the recipe.
void print(raw_ostream &O, const Twine &Indent,
diff --git a/llvm/lib/Transforms/Vectorize/VPlanRecipes.cpp b/llvm/lib/Transforms/Vectorize/VPlanRecipes.cpp
index 2948ecc580edc0..92131e8382bb74 100644
--- a/llvm/lib/Transforms/Vectorize/VPlanRecipes.cpp
+++ b/llvm/lib/Transforms/Vectorize/VPlanRecipes.cpp
@@ -1462,6 +1462,49 @@ void VPWidenCastRecipe::execute(VPTransformState &State) {
State.addMetadata(Cast, cast_or_null<Instruction>(getUnderlyingValue()));
}
+InstructionCost VPWidenCastRecipe::computeCost(ElementCount VF,
+ VPCostContext &Ctx) const {
+ auto *SrcTy = cast<VectorType>(
+ ToVectorTy(Ctx.Types.inferScalarType(getOperand(0)), VF));
+ auto *DestTy = cast<VectorType>(ToVectorTy(getResultType(), VF));
+ // Computes the CastContextHint from a VPWidenMemoryRecipe instruction.
+ auto ComputeCCH = [&](VPWidenMemoryRecipe *R) -> TTI::CastContextHint {
+ assert((isa<VPWidenLoadRecipe>(R) || isa<VPWidenStoreRecipe>(R)) &&
+ "Expected a load or a store!");
+
+ if (VF.isScalar())
+ return TTI::CastContextHint::Normal;
+ if (!R->isConsecutive())
+ return TTI::CastContextHint::GatherScatter;
+ if (R->isReverse())
+ return TTI::CastContextHint::Reversed;
+ if (R->isMasked())
+ return TTI::CastContextHint::Masked;
+ return TTI::CastContextHint::Normal;
+ };
+
+ TTI::CastContextHint CCH = TTI::CastContextHint::None;
+ // For Trunc, the context is the only user, which must be a
+ // VPWidenStoreRecipe.
+ if (Opcode == Instruction::Trunc || Opcode == Instruction::FPTrunc) {
+ if (!cast<VPValue>(this)->hasMoreThanOneUniqueUser())
+ if (VPWidenMemoryRecipe *Store =
+ dyn_cast<VPWidenMemoryRecipe>(*this->user_begin()))
+ CCH = ComputeCCH(Store);
+ }
+ // For Z/Sext, the context is the operand, which must be a VPWidenLoadRecipe.
+ else if (Opcode == Instruction::ZExt || Opcode == Instruction::SExt ||
+ Opcode == Instruction::FPExt) {
+ if (VPWidenMemoryRecipe *Load = dyn_cast<VPWidenMemoryRecipe>(
+ this->getOperand(0)->getDefiningRecipe()))
+ CCH = ComputeCCH(Load);
+ }
+ // Arm TTI will use the underlying instruction to determine the cost.
+ return Ctx.TTI.getCastInstrCost(
+ Opcode, DestTy, SrcTy, CCH, TTI::TCK_RecipThroughput,
+ dyn_cast_if_present<Instruction>(getUnderlyingValue()));
+}
+
#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
void VPWidenCastRecipe::print(raw_ostream &O, const Twine &Indent,
VPSlotTracker &SlotTracker) const {
diff --git a/llvm/lib/Transforms/Vectorize/VPlanValue.h b/llvm/lib/Transforms/Vectorize/VPlanValue.h
index f2978b0a758b6a..1900182f76e071 100644
--- a/llvm/lib/Transforms/Vectorize/VPlanValue.h
+++ b/llvm/lib/Transforms/Vectorize/VPlanValue.h
@@ -135,7 +135,7 @@ class VPValue {
}
/// Returns true if the value has more than one unique user.
- bool hasMoreThanOneUniqueUser() {
+ bool hasMoreThanOneUniqueUser() const {
if (getNumUsers() == 0)
return false;
>From b43e6f9ca7bce726dcaced72e1024e5674787dee Mon Sep 17 00:00:00 2001
From: Elvis Wang <elvis.wang at sifive.com>
Date: Sun, 13 Oct 2024 23:59:06 -0700
Subject: [PATCH 2/7] Fixup! consider VPInterleaveRecipe, VPReplicateRecipe,
Live-in/out
---
.../lib/Transforms/Vectorize/VPlanRecipes.cpp | 37 ++++++++++++-------
1 file changed, 24 insertions(+), 13 deletions(-)
diff --git a/llvm/lib/Transforms/Vectorize/VPlanRecipes.cpp b/llvm/lib/Transforms/Vectorize/VPlanRecipes.cpp
index 92131e8382bb74..445f062773a1a7 100644
--- a/llvm/lib/Transforms/Vectorize/VPlanRecipes.cpp
+++ b/llvm/lib/Transforms/Vectorize/VPlanRecipes.cpp
@@ -1468,10 +1468,9 @@ InstructionCost VPWidenCastRecipe::computeCost(ElementCount VF,
ToVectorTy(Ctx.Types.inferScalarType(getOperand(0)), VF));
auto *DestTy = cast<VectorType>(ToVectorTy(getResultType(), VF));
// Computes the CastContextHint from a VPWidenMemoryRecipe instruction.
- auto ComputeCCH = [&](VPWidenMemoryRecipe *R) -> TTI::CastContextHint {
+ auto ComputeCCH = [&](const VPWidenMemoryRecipe *R) -> TTI::CastContextHint {
assert((isa<VPWidenLoadRecipe>(R) || isa<VPWidenStoreRecipe>(R)) &&
"Expected a load or a store!");
-
if (VF.isScalar())
return TTI::CastContextHint::Normal;
if (!R->isConsecutive())
@@ -1485,19 +1484,31 @@ InstructionCost VPWidenCastRecipe::computeCost(ElementCount VF,
TTI::CastContextHint CCH = TTI::CastContextHint::None;
// For Trunc, the context is the only user, which must be a
- // VPWidenStoreRecipe.
- if (Opcode == Instruction::Trunc || Opcode == Instruction::FPTrunc) {
- if (!cast<VPValue>(this)->hasMoreThanOneUniqueUser())
- if (VPWidenMemoryRecipe *Store =
- dyn_cast<VPWidenMemoryRecipe>(*this->user_begin()))
- CCH = ComputeCCH(Store);
- }
- // For Z/Sext, the context is the operand, which must be a VPWidenLoadRecipe.
+ // VPWidenStoreRecipe, a VPInterleaveRecipe ,a VPReplicateRecipe or a live-out
+ // value.
+ if ((Opcode == Instruction::Trunc || Opcode == Instruction::FPTrunc) &&
+ !hasMoreThanOneUniqueUser() && getNumUsers() > 0) {
+ auto *StoreRecipe = dyn_cast<VPRecipeBase>(*user_begin());
+ if (isa<VPLiveOut>(*user_begin()) || isa<VPReplicateRecipe>(StoreRecipe))
+ CCH = TTI::CastContextHint::Normal;
+ else if (const VPWidenMemoryRecipe *MemoryStoreRecipe =
+ dyn_cast<VPWidenMemoryRecipe>(StoreRecipe))
+ CCH = ComputeCCH(MemoryStoreRecipe);
+ else if (isa<VPInterleaveRecipe>(StoreRecipe))
+ CCH = TTI::CastContextHint::Interleave;
+ }
+ // For Z/Sext, the context is the operand, which must be a VPWidenLoadRecipe,
+ // a VPInterleaveRecipe, a VPReplicateRecipe or a live-in value.
else if (Opcode == Instruction::ZExt || Opcode == Instruction::SExt ||
Opcode == Instruction::FPExt) {
- if (VPWidenMemoryRecipe *Load = dyn_cast<VPWidenMemoryRecipe>(
- this->getOperand(0)->getDefiningRecipe()))
- CCH = ComputeCCH(Load);
+ const VPRecipeBase *LoadRecipe = getOperand(0)->getDefiningRecipe();
+ if (getOperand(0)->isLiveIn() || isa<VPReplicateRecipe>(LoadRecipe))
+ CCH = TTI::CastContextHint::Normal;
+ else if (const VPWidenMemoryRecipe *MemoryLoadRecipe =
+ dyn_cast<VPWidenMemoryRecipe>(LoadRecipe))
+ CCH = ComputeCCH(MemoryLoadRecipe);
+ else if (isa<VPInterleaveRecipe>(LoadRecipe))
+ CCH = TTI::CastContextHint::Interleave;
}
// Arm TTI will use the underlying instruction to determine the cost.
return Ctx.TTI.getCastInstrCost(
>From de00e4b44357d77873c117ad37d750adcff61611 Mon Sep 17 00:00:00 2001
From: Elvis Wang <elvis.wang at sifive.com>
Date: Mon, 14 Oct 2024 17:54:52 -0700
Subject: [PATCH 3/7] Fixup! merge ComputeCCH
---
.../lib/Transforms/Vectorize/VPlanRecipes.cpp | 38 +++++++++----------
1 file changed, 19 insertions(+), 19 deletions(-)
diff --git a/llvm/lib/Transforms/Vectorize/VPlanRecipes.cpp b/llvm/lib/Transforms/Vectorize/VPlanRecipes.cpp
index 445f062773a1a7..6283da51cc3fd9 100644
--- a/llvm/lib/Transforms/Vectorize/VPlanRecipes.cpp
+++ b/llvm/lib/Transforms/Vectorize/VPlanRecipes.cpp
@@ -1467,17 +1467,23 @@ InstructionCost VPWidenCastRecipe::computeCost(ElementCount VF,
auto *SrcTy = cast<VectorType>(
ToVectorTy(Ctx.Types.inferScalarType(getOperand(0)), VF));
auto *DestTy = cast<VectorType>(ToVectorTy(getResultType(), VF));
- // Computes the CastContextHint from a VPWidenMemoryRecipe instruction.
- auto ComputeCCH = [&](const VPWidenMemoryRecipe *R) -> TTI::CastContextHint {
- assert((isa<VPWidenLoadRecipe>(R) || isa<VPWidenStoreRecipe>(R)) &&
- "Expected a load or a store!");
+ // Computes the CastContextHint from a recipes that may access memory.
+ auto ComputeCCH = [&](const VPRecipeBase *R) -> TTI::CastContextHint {
if (VF.isScalar())
return TTI::CastContextHint::Normal;
- if (!R->isConsecutive())
+ if (isa<VPInterleaveRecipe>(R))
+ return TTI::CastContextHint::Interleave;
+ if (const auto *ReplicateRecipe = dyn_cast<VPReplicateRecipe>(R))
+ return ReplicateRecipe->isPredicated() ? TTI::CastContextHint::Masked
+ : TTI::CastContextHint::Normal;
+ const auto *WidenMemoryRecipe = dyn_cast<VPWidenMemoryRecipe>(R);
+ if (WidenMemoryRecipe == nullptr)
+ return TTI::CastContextHint::None;
+ if (!WidenMemoryRecipe->isConsecutive())
return TTI::CastContextHint::GatherScatter;
- if (R->isReverse())
+ if (WidenMemoryRecipe->isReverse())
return TTI::CastContextHint::Reversed;
- if (R->isMasked())
+ if (WidenMemoryRecipe->isMasked())
return TTI::CastContextHint::Masked;
return TTI::CastContextHint::Normal;
};
@@ -1489,26 +1495,20 @@ InstructionCost VPWidenCastRecipe::computeCost(ElementCount VF,
if ((Opcode == Instruction::Trunc || Opcode == Instruction::FPTrunc) &&
!hasMoreThanOneUniqueUser() && getNumUsers() > 0) {
auto *StoreRecipe = dyn_cast<VPRecipeBase>(*user_begin());
- if (isa<VPLiveOut>(*user_begin()) || isa<VPReplicateRecipe>(StoreRecipe))
+ if (isa<VPLiveOut>(*user_begin()))
CCH = TTI::CastContextHint::Normal;
- else if (const VPWidenMemoryRecipe *MemoryStoreRecipe =
- dyn_cast<VPWidenMemoryRecipe>(StoreRecipe))
- CCH = ComputeCCH(MemoryStoreRecipe);
- else if (isa<VPInterleaveRecipe>(StoreRecipe))
- CCH = TTI::CastContextHint::Interleave;
+ else if (StoreRecipe)
+ CCH = ComputeCCH(StoreRecipe);
}
// For Z/Sext, the context is the operand, which must be a VPWidenLoadRecipe,
// a VPInterleaveRecipe, a VPReplicateRecipe or a live-in value.
else if (Opcode == Instruction::ZExt || Opcode == Instruction::SExt ||
Opcode == Instruction::FPExt) {
const VPRecipeBase *LoadRecipe = getOperand(0)->getDefiningRecipe();
- if (getOperand(0)->isLiveIn() || isa<VPReplicateRecipe>(LoadRecipe))
+ if (getOperand(0)->isLiveIn())
CCH = TTI::CastContextHint::Normal;
- else if (const VPWidenMemoryRecipe *MemoryLoadRecipe =
- dyn_cast<VPWidenMemoryRecipe>(LoadRecipe))
- CCH = ComputeCCH(MemoryLoadRecipe);
- else if (isa<VPInterleaveRecipe>(LoadRecipe))
- CCH = TTI::CastContextHint::Interleave;
+ else if (LoadRecipe)
+ CCH = ComputeCCH(LoadRecipe);
}
// Arm TTI will use the underlying instruction to determine the cost.
return Ctx.TTI.getCastInstrCost(
>From 5eadd23601baa4c6261a226c71833281c6590121 Mon Sep 17 00:00:00 2001
From: Elvis Wang <elvis.wang at sifive.com>
Date: Mon, 14 Oct 2024 18:02:54 -0700
Subject: [PATCH 4/7] fixup! move def close to use.
---
llvm/lib/Transforms/Vectorize/VPlanRecipes.cpp | 7 ++++---
1 file changed, 4 insertions(+), 3 deletions(-)
diff --git a/llvm/lib/Transforms/Vectorize/VPlanRecipes.cpp b/llvm/lib/Transforms/Vectorize/VPlanRecipes.cpp
index 6283da51cc3fd9..f3e309402e91b2 100644
--- a/llvm/lib/Transforms/Vectorize/VPlanRecipes.cpp
+++ b/llvm/lib/Transforms/Vectorize/VPlanRecipes.cpp
@@ -1464,9 +1464,6 @@ void VPWidenCastRecipe::execute(VPTransformState &State) {
InstructionCost VPWidenCastRecipe::computeCost(ElementCount VF,
VPCostContext &Ctx) const {
- auto *SrcTy = cast<VectorType>(
- ToVectorTy(Ctx.Types.inferScalarType(getOperand(0)), VF));
- auto *DestTy = cast<VectorType>(ToVectorTy(getResultType(), VF));
// Computes the CastContextHint from a recipes that may access memory.
auto ComputeCCH = [&](const VPRecipeBase *R) -> TTI::CastContextHint {
if (VF.isScalar())
@@ -1510,6 +1507,10 @@ InstructionCost VPWidenCastRecipe::computeCost(ElementCount VF,
else if (LoadRecipe)
CCH = ComputeCCH(LoadRecipe);
}
+
+ auto *SrcTy = cast<VectorType>(
+ ToVectorTy(Ctx.Types.inferScalarType(getOperand(0)), VF));
+ auto *DestTy = cast<VectorType>(ToVectorTy(getResultType(), VF));
// Arm TTI will use the underlying instruction to determine the cost.
return Ctx.TTI.getCastInstrCost(
Opcode, DestTy, SrcTy, CCH, TTI::TCK_RecipThroughput,
>From 29c4d81a90ff250194b14313736178f9b6fb1d39 Mon Sep 17 00:00:00 2001
From: Elvis Wang <elvis.wang at sifive.com>
Date: Wed, 16 Oct 2024 20:18:43 -0700
Subject: [PATCH 5/7] Fixup! Remove live-out since it's in the loop.
---
llvm/lib/Transforms/Vectorize/VPlanRecipes.cpp | 13 ++++---------
1 file changed, 4 insertions(+), 9 deletions(-)
diff --git a/llvm/lib/Transforms/Vectorize/VPlanRecipes.cpp b/llvm/lib/Transforms/Vectorize/VPlanRecipes.cpp
index f3e309402e91b2..ff1b75e90f5a38 100644
--- a/llvm/lib/Transforms/Vectorize/VPlanRecipes.cpp
+++ b/llvm/lib/Transforms/Vectorize/VPlanRecipes.cpp
@@ -1487,25 +1487,20 @@ InstructionCost VPWidenCastRecipe::computeCost(ElementCount VF,
TTI::CastContextHint CCH = TTI::CastContextHint::None;
// For Trunc, the context is the only user, which must be a
- // VPWidenStoreRecipe, a VPInterleaveRecipe ,a VPReplicateRecipe or a live-out
- // value.
+ // VPWidenStoreRecipe, a VPInterleaveRecipe ,or a VPReplicateRecipe.
if ((Opcode == Instruction::Trunc || Opcode == Instruction::FPTrunc) &&
!hasMoreThanOneUniqueUser() && getNumUsers() > 0) {
- auto *StoreRecipe = dyn_cast<VPRecipeBase>(*user_begin());
- if (isa<VPLiveOut>(*user_begin()))
- CCH = TTI::CastContextHint::Normal;
- else if (StoreRecipe)
+ if (auto *StoreRecipe = dyn_cast<VPRecipeBase>(*user_begin()))
CCH = ComputeCCH(StoreRecipe);
}
// For Z/Sext, the context is the operand, which must be a VPWidenLoadRecipe,
// a VPInterleaveRecipe, a VPReplicateRecipe or a live-in value.
else if (Opcode == Instruction::ZExt || Opcode == Instruction::SExt ||
Opcode == Instruction::FPExt) {
- const VPRecipeBase *LoadRecipe = getOperand(0)->getDefiningRecipe();
if (getOperand(0)->isLiveIn())
CCH = TTI::CastContextHint::Normal;
- else if (LoadRecipe)
- CCH = ComputeCCH(LoadRecipe);
+ else if (getOperand(0)->getDefiningRecipe())
+ CCH = ComputeCCH(getOperand(0)->getDefiningRecipe());
}
auto *SrcTy = cast<VectorType>(
>From fe37f080bd44625cf9a406a324eefdd7ca2bec45 Mon Sep 17 00:00:00 2001
From: Elvis Wang <elvis.wang at sifive.com>
Date: Sun, 20 Oct 2024 22:16:48 -0700
Subject: [PATCH 6/7] Address comments.
---
llvm/lib/Transforms/Vectorize/LoopVectorize.cpp | 9 +++++----
llvm/lib/Transforms/Vectorize/VPlanRecipes.cpp | 11 ++++++-----
2 files changed, 11 insertions(+), 9 deletions(-)
diff --git a/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp b/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp
index 0c8e2d1a2d9895..c6158d6efa505a 100644
--- a/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp
+++ b/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp
@@ -7258,7 +7258,7 @@ LoopVectorizationPlanner::precomputeCosts(VPlan &Plan, ElementCount VF,
const auto &ChainOps = RdxDesc.getReductionOpChain(RedPhi, OrigLoop);
SetVector<Instruction *> ChainOpsAndOperands(ChainOps.begin(),
ChainOps.end());
- auto isZExtOrSExt = [](const unsigned Opcode) -> bool {
+ auto IsZExtOrSExt = [](const unsigned Opcode) -> bool {
return Opcode == Instruction::ZExt || Opcode == Instruction::SExt;
};
// Also include the operands of instructions in the chain, as the cost-model
@@ -7274,11 +7274,12 @@ LoopVectorizationPlanner::precomputeCosts(VPlan &Plan, ElementCount VF,
ChainOpsAndOperands.insert(I);
if (I->getOpcode() == Instruction::Mul) {
auto *Ext0 = dyn_cast<Instruction>(I->getOperand(0));
- if (Ext0 && isZExtOrSExt(Ext0->getOpcode()))
- ChainOpsAndOperands.insert(Ext0);
auto *Ext1 = dyn_cast<Instruction>(I->getOperand(1));
- if (Ext1 && isZExtOrSExt(Ext1->getOpcode()))
+ if (Ext0 && IsZExtOrSExt(Ext0->getOpcode()) && Ext1 &&
+ Ext0->getOpcode() == Ext1->getOpcode()) {
+ ChainOpsAndOperands.insert(Ext0);
ChainOpsAndOperands.insert(Ext1);
+ }
}
}
}
diff --git a/llvm/lib/Transforms/Vectorize/VPlanRecipes.cpp b/llvm/lib/Transforms/Vectorize/VPlanRecipes.cpp
index ff1b75e90f5a38..b3034ff74cb16e 100644
--- a/llvm/lib/Transforms/Vectorize/VPlanRecipes.cpp
+++ b/llvm/lib/Transforms/Vectorize/VPlanRecipes.cpp
@@ -1485,6 +1485,7 @@ InstructionCost VPWidenCastRecipe::computeCost(ElementCount VF,
return TTI::CastContextHint::Normal;
};
+ VPValue *Operand = getOperand(0);
TTI::CastContextHint CCH = TTI::CastContextHint::None;
// For Trunc, the context is the only user, which must be a
// VPWidenStoreRecipe, a VPInterleaveRecipe ,or a VPReplicateRecipe.
@@ -1497,14 +1498,14 @@ InstructionCost VPWidenCastRecipe::computeCost(ElementCount VF,
// a VPInterleaveRecipe, a VPReplicateRecipe or a live-in value.
else if (Opcode == Instruction::ZExt || Opcode == Instruction::SExt ||
Opcode == Instruction::FPExt) {
- if (getOperand(0)->isLiveIn())
+ if (Operand->isLiveIn())
CCH = TTI::CastContextHint::Normal;
- else if (getOperand(0)->getDefiningRecipe())
- CCH = ComputeCCH(getOperand(0)->getDefiningRecipe());
+ else if (Operand->getDefiningRecipe())
+ CCH = ComputeCCH(Operand->getDefiningRecipe());
}
- auto *SrcTy = cast<VectorType>(
- ToVectorTy(Ctx.Types.inferScalarType(getOperand(0)), VF));
+ auto *SrcTy =
+ cast<VectorType>(ToVectorTy(Ctx.Types.inferScalarType(Operand), VF));
auto *DestTy = cast<VectorType>(ToVectorTy(getResultType(), VF));
// Arm TTI will use the underlying instruction to determine the cost.
return Ctx.TTI.getCastInstrCost(
>From 230a0d2d250c9926776d0324a7f4f9805f6b09d1 Mon Sep 17 00:00:00 2001
From: Elvis Wang <elvis.wang at sifive.com>
Date: Mon, 21 Oct 2024 18:39:47 -0700
Subject: [PATCH 7/7] Fixup! Refine comments.
---
llvm/lib/Transforms/Vectorize/VPlanRecipes.cpp | 6 ++----
1 file changed, 2 insertions(+), 4 deletions(-)
diff --git a/llvm/lib/Transforms/Vectorize/VPlanRecipes.cpp b/llvm/lib/Transforms/Vectorize/VPlanRecipes.cpp
index b3034ff74cb16e..3acd7d5e3ca4cd 100644
--- a/llvm/lib/Transforms/Vectorize/VPlanRecipes.cpp
+++ b/llvm/lib/Transforms/Vectorize/VPlanRecipes.cpp
@@ -1487,15 +1487,13 @@ InstructionCost VPWidenCastRecipe::computeCost(ElementCount VF,
VPValue *Operand = getOperand(0);
TTI::CastContextHint CCH = TTI::CastContextHint::None;
- // For Trunc, the context is the only user, which must be a
- // VPWidenStoreRecipe, a VPInterleaveRecipe ,or a VPReplicateRecipe.
+ // For Trunc/FPTrunc, get the context from the only user.
if ((Opcode == Instruction::Trunc || Opcode == Instruction::FPTrunc) &&
!hasMoreThanOneUniqueUser() && getNumUsers() > 0) {
if (auto *StoreRecipe = dyn_cast<VPRecipeBase>(*user_begin()))
CCH = ComputeCCH(StoreRecipe);
}
- // For Z/Sext, the context is the operand, which must be a VPWidenLoadRecipe,
- // a VPInterleaveRecipe, a VPReplicateRecipe or a live-in value.
+ // For Z/Sext, get the context from the operand.
else if (Opcode == Instruction::ZExt || Opcode == Instruction::SExt ||
Opcode == Instruction::FPExt) {
if (Operand->isLiveIn())
More information about the llvm-commits
mailing list