Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

[FunctionAttrs] Add the "initializes" attribute inference #97373

Merged
merged 13 commits into from
Nov 19, 2024
331 changes: 328 additions & 3 deletions llvm/lib/Transforms/IPO/FunctionAttrs.cpp
Original file line number Diff line number Diff line change
Expand Up @@ -15,6 +15,7 @@
#include "llvm/Transforms/IPO/FunctionAttrs.h"
#include "llvm/ADT/ArrayRef.h"
#include "llvm/ADT/DenseMap.h"
#include "llvm/ADT/PostOrderIterator.h"
#include "llvm/ADT/SCCIterator.h"
#include "llvm/ADT/STLExtras.h"
#include "llvm/ADT/SetVector.h"
Expand All @@ -36,6 +37,7 @@
#include "llvm/IR/Attributes.h"
#include "llvm/IR/BasicBlock.h"
#include "llvm/IR/Constant.h"
#include "llvm/IR/ConstantRangeList.h"
#include "llvm/IR/Constants.h"
#include "llvm/IR/Function.h"
#include "llvm/IR/InstIterator.h"
Expand Down Expand Up @@ -580,6 +582,206 @@ struct ArgumentUsesTracker : public CaptureTracker {
const SCCNodeSet &SCCNodes;
};

// A struct of argument use: a Use and the offset it accesses. This struct
// is to track uses inside function via GEP. If GEP has a non-constant index,
// the Offset field is nullopt.
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Please use doxygen comments (///) for classes.

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Done.

struct ArgumentUse {
Copy link

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

missing documentation.

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Thanks for reminding! Done.

Use *U;
std::optional<int64_t> Offset;
};

// A struct of argument access info. "Unknown" accesses are the cases like
// unrecognized instructions, instructions that have more than one use of
// the argument, or volatile memory accesses. "Unknown" implies "IsClobber"
// and an empty access range.
// Write or Read accesses can be clobbers as well for example, a Load with
// scalable type.
struct ArgumentAccessInfo {
enum class AccessType : uint8_t { Write, Read, Unknown };
AccessType ArgAccessType;
bool IsClobber = false;
jvoung marked this conversation as resolved.
Show resolved Hide resolved
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

The "clobber" terminology here is pretty weird, because usually "clobber" means "write", while here it's more like "read".

Also, after looking at the code, I'm a bit unclear on whether we really need both AccessType::Unknown and IsClobber. Isn't one of them sufficient?

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Thanks for this comment. Cleaned up the code:

  • Removed "IsClobber".
  • Changed the AccessType enum to: Read/Write/WriteWithSideEffect/Unknown.
    • Read: simple load/memcpy.
    • Write: simple store/memstore/memcpy/call(write a parameter and the parameter is writeonly/readnone and nocapture).
    • WriteWithSideEffect: call instructions write a parameter but the parameter is not writeonly/readnone or nocapture.
    • Unknown: unrecognized accesses or non-simple load/store/memset/...

While inferring the initializes attribute in a block (from end inst to begin), we empty the initializes range list for WriteWithSideEffect/Unknown, union the range list of Write/WriteWithSideEffect, and subtract the range list of Read.

ConstantRangeList AccessRanges;
};

// A struct to wrap the argument use info per block.
struct UsesPerBlockInfo {
SmallDenseMap<Instruction *, ArgumentAccessInfo, 4> Insts;
bool HasWrites = false;
bool HasClobber = false;
};

// A struct to summarize the argument use info in a function.
struct ArgumentUsesSummary {
bool HasAnyWrite = false;
bool HasWriteOutsideEntryBB = false;
SmallDenseMap<const BasicBlock *, UsesPerBlockInfo, 16> UsesPerBlock;
};

ArgumentAccessInfo GetArgmentAccessInfo(const Instruction *I,
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Suggested change
ArgumentAccessInfo GetArgmentAccessInfo(const Instruction *I,
ArgumentAccessInfo getArgmentAccessInfo(const Instruction *I,

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Done.

const ArgumentUse &ArgUse,
const DataLayout &DL) {
auto GetTypeAccessRange =
[&DL](Type *Ty,
std::optional<int64_t> Offset) -> std::optional<ConstantRange> {
auto TypeSize = DL.getTypeStoreSize(Ty);
if (!TypeSize.isScalable() && Offset) {
int64_t Size = TypeSize.getFixedValue();
return ConstantRange(APInt(64, *Offset, true),
APInt(64, *Offset + Size, true));
}
return std::nullopt;
};
auto GetConstantIntRange =
[](Value *Length,
std::optional<int64_t> Offset) -> std::optional<ConstantRange> {
auto *ConstantLength = dyn_cast<ConstantInt>(Length);
if (ConstantLength && Offset)
return ConstantRange(
APInt(64, *Offset, true),
APInt(64, *Offset + ConstantLength->getSExtValue(), true));
return std::nullopt;
};
if (auto *SI = dyn_cast<StoreInst>(I)) {
aeubanks marked this conversation as resolved.
Show resolved Hide resolved
if (!SI->isVolatile() && &SI->getOperandUse(1) == ArgUse.U) {
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

These isVolatile() checks should probably be isSimple() checks? Presumably you don't want to handle atomics.

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Ah, just noticed isSimple(). Replaced with isSimple() checks.

// Get the fixed type size of "SI". Since the access range of a write
// will be unioned, if "SI" doesn't have a fixed type size, we just set
// the access range to empty.
ConstantRangeList AccessRanges;
if (auto TypeAccessRange =
GetTypeAccessRange(SI->getAccessType(), ArgUse.Offset))
AccessRanges.insert(*TypeAccessRange);
return {ArgumentAccessInfo::AccessType::Write,
/*IsClobber=*/false, AccessRanges};
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Suggested change
/*IsClobber=*/false, AccessRanges};
/*IsClobber=*/false, std::move(AccessRanges)};

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Done.

}
} else if (auto *LI = dyn_cast<LoadInst>(I)) {
if (!LI->isVolatile()) {
assert(&LI->getOperandUse(0) == ArgUse.U);
// Get the fixed type size of "LI". Different from Write, if "LI"
// doesn't have a fixed type size, we conservatively set as a clobber
// with an empty access range.
if (auto TypeAccessRange =
GetTypeAccessRange(LI->getAccessType(), ArgUse.Offset))
return {ArgumentAccessInfo::AccessType::Read,
/*IsClobber=*/false,
{*TypeAccessRange}};
return {ArgumentAccessInfo::AccessType::Read, /*IsClobber=*/true, {}};
}
} else if (auto *MemSet = dyn_cast<MemSetInst>(I)) {
if (!MemSet->isVolatile()) {
ConstantRangeList AccessRanges;
if (auto AccessRange =
GetConstantIntRange(MemSet->getLength(), ArgUse.Offset))
AccessRanges.insert(*AccessRange);
return {ArgumentAccessInfo::AccessType::Write,
/*IsClobber=*/false, AccessRanges};
}
} else if (auto *MTI = dyn_cast<MemTransferInst>(I)) {
if (!MTI->isVolatile()) {
if (&MTI->getOperandUse(0) == ArgUse.U) {
ConstantRangeList AccessRanges;
if (auto AccessRange =
GetConstantIntRange(MTI->getLength(), ArgUse.Offset))
AccessRanges.insert(*AccessRange);
return {ArgumentAccessInfo::AccessType::Write,
/*IsClobber=*/false, AccessRanges};
} else if (&MTI->getOperandUse(1) == ArgUse.U) {
if (auto AccessRange =
GetConstantIntRange(MTI->getLength(), ArgUse.Offset))
return {ArgumentAccessInfo::AccessType::Read,
/*IsClobber=*/false,
{*AccessRange}};
return {ArgumentAccessInfo::AccessType::Read, /*IsClobber=*/true, {}};
}
}
} else if (auto *CB = dyn_cast<CallBase>(I)) {
if (CB->isArgOperand(ArgUse.U)) {
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

we should add a test where the argument is passed as an operand bundle to a call, rather than an argument

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Done!

unsigned ArgNo = CB->getArgOperandNo(ArgUse.U);
bool IsInitialize = CB->paramHasAttr(ArgNo, Attribute::Initializes);
// Argument is only not clobbered when parameter is writeonly/readnone
// and nocapture.
bool IsClobber = !(CB->onlyWritesMemory(ArgNo) &&
CB->paramHasAttr(ArgNo, Attribute::NoCapture));
ConstantRangeList AccessRanges;
if (IsInitialize && ArgUse.Offset) {
Attribute Attr = CB->getParamAttr(ArgNo, Attribute::Initializes);
ConstantRangeList CBCRL = Attr.getValueAsConstantRangeList();
for (ConstantRange &CR : CBCRL)
AccessRanges.insert(ConstantRange(CR.getLower() + *ArgUse.Offset,
CR.getUpper() + *ArgUse.Offset));
return {ArgumentAccessInfo::AccessType::Write, IsClobber, AccessRanges};
}
}
}
// Unrecognized instructions are considered clobbers.
return {ArgumentAccessInfo::AccessType::Unknown, /*IsClobber=*/true, {}};
}

// Collect the uses of argument "A" in "F".
ArgumentUsesSummary CollectArgumentUsesPerBlock(Argument &A, Function &F) {
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Suggested change
ArgumentUsesSummary CollectArgumentUsesPerBlock(Argument &A, Function &F) {
ArgumentUsesSummary collectArgumentUsesPerBlock(Argument &A, Function &F) {

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Done.

auto &DL = F.getParent()->getDataLayout();
auto PointerSize =
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Suggested change
auto PointerSize =
unsigned PointerSize =

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Done.

DL.getIndexSizeInBits(A.getType()->getPointerAddressSpace());
ArgumentUsesSummary Result;

BasicBlock &EntryBB = F.getEntryBlock();
SmallVector<ArgumentUse, 4> Worklist;
for (Use &U : A.uses())
Worklist.push_back({&U, 0});

// Update "UsesPerBlock" with the block of "I" as key and "Info" as value.
// Return true if the block of "I" has write accesses after updating.
auto UpdateUseInfo = [&Result](Instruction *I, ArgumentAccessInfo Info) {
auto *BB = I->getParent();
auto &BBInfo = Result.UsesPerBlock.getOrInsertDefault(BB);
bool AlreadyVisitedInst = BBInfo.Insts.contains(I);
auto &IInfo = BBInfo.Insts[I];

// Instructions that have more than one use of the argument are considered
// as clobbers.
if (AlreadyVisitedInst) {
IInfo = {ArgumentAccessInfo::AccessType::Unknown, /*IsClobber=*/true, {}};
BBInfo.HasClobber = true;
return false;
}

IInfo = Info;
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Suggested change
IInfo = Info;
IInfo = std::move(Info);

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Done.

BBInfo.HasClobber |= IInfo.IsClobber;
bool InfoHasWrites =
IInfo.ArgAccessType == ArgumentAccessInfo::AccessType::Write &&
!IInfo.AccessRanges.empty();
BBInfo.HasWrites |= InfoHasWrites;
return InfoHasWrites;
};

// No need for a visited set because we don't look through phis, so there are
// no cycles.
while (!Worklist.empty()) {
ArgumentUse ArgUse = Worklist.pop_back_val();
User *U = ArgUse.U->getUser();
// Add GEP uses to worklist.
// If the GEP is not a constant GEP, set the ArgumentUse::Offset to nullopt.
if (auto *GEP = dyn_cast<GEPOperator>(U)) {
APInt Offset(PointerSize, 0, /*isSigned=*/true);
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Suggested change
APInt Offset(PointerSize, 0, /*isSigned=*/true);
APInt Offset(PointerSize, 0);

The flag is not meaningful for a zero value.

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Thanks, done.

bool IsConstGEP = GEP->accumulateConstantOffset(DL, Offset);
std::optional<int64_t> NewOffset = std::nullopt;
if (IsConstGEP && ArgUse.Offset)
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

If ArgUse.Offset is already nullopt, you can skip the accumulateConstantOffset() call.

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Good point! Refactored code to check if ArgUse.Offset is nullopt first.

NewOffset = *ArgUse.Offset + Offset.getSExtValue();
for (Use &U : GEP->uses())
Worklist.push_back({&U, NewOffset});
continue;
}

auto *I = cast<Instruction>(U);
bool HasWrite = UpdateUseInfo(I, GetArgmentAccessInfo(I, ArgUse, DL));

Result.HasAnyWrite |= HasWrite;

if (HasWrite && I->getParent() != &EntryBB)
Result.HasWriteOutsideEntryBB = true;
}
return Result;
}

} // end anonymous namespace

namespace llvm {
Expand Down Expand Up @@ -866,9 +1068,125 @@ static bool addAccessAttr(Argument *A, Attribute::AttrKind R) {
return true;
}

static bool inferInitializes(Argument &A, Function &F) {
auto ArgumentUses = CollectArgumentUsesPerBlock(A, F);
// No write anywhere in the function, bail.
if (!ArgumentUses.HasAnyWrite)
return false;

auto &UsesPerBlock = ArgumentUses.UsesPerBlock;
BasicBlock &EntryBB = F.getEntryBlock();
// A map to store the argument ranges initialized by a BasicBlock (including
// its successors).
DenseMap<const BasicBlock *, ConstantRangeList> Initialized;
// Visit the successors of "BB" block and the instructions in BB (post-order)
// to get the argument ranges initialized by "BB" (including its successors).
// The result will be cached in "Initialized".
auto VisitBlock = [&](const BasicBlock *BB) -> ConstantRangeList {
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

can you add a comment on what VisitBlock does

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Done and found a corner case :-)

"If this block has uses and none are writes, the argument is not initialized in this block."
Removed this early return. See the test.

auto UPB = UsesPerBlock.find(BB);
ConstantRangeList CRL;

// Start with intersection of successors.
// If this block has any clobbering use, we're going to clear out the
// ranges at some point in this block anyway, so don't bother looking at
// successors.
if (UPB == UsesPerBlock.end() || !UPB->second.HasClobber) {
bool HasAddedSuccessor = false;
for (auto *Succ : successors(BB)) {
if (auto SuccI = Initialized.find(Succ); SuccI != Initialized.end()) {
if (HasAddedSuccessor) {
CRL = CRL.intersectWith(SuccI->second);
} else {
CRL = SuccI->second;
HasAddedSuccessor = true;
}
} else {
CRL = ConstantRangeList();
break;
}
}
}

if (UPB != UsesPerBlock.end()) {
// Sort uses in this block by instruction order.
SmallVector<std::pair<Instruction *, ArgumentAccessInfo>, 2> Insts;
append_range(Insts, UPB->second.Insts);
sort(Insts, [](std::pair<Instruction *, ArgumentAccessInfo> &LHS,
std::pair<Instruction *, ArgumentAccessInfo> &RHS) {
return LHS.first->comesBefore(RHS.first);
});

// From the end of the block to the beginning of the block, set
// initializes ranges.
for (auto &[_, Info] : reverse(Insts)) {
if (Info.IsClobber)
CRL = ConstantRangeList();
if (!Info.AccessRanges.empty()) {
if (Info.ArgAccessType == ArgumentAccessInfo::AccessType::Write) {
CRL = CRL.unionWith(Info.AccessRanges);
} else {
assert(Info.ArgAccessType == ArgumentAccessInfo::AccessType::Read);
for (const auto &ReadRange : Info.AccessRanges)
CRL.subtract(ReadRange);
}
}
}
}
return CRL;
};

ConstantRangeList EntryCRL;
// If all write instructions are in the EntryBB, or if the EntryBB has
// a clobbering use, we only need to look at EntryBB.
bool OnlyScanEntryBlock = !ArgumentUses.HasWriteOutsideEntryBB;
if (!OnlyScanEntryBlock)
if (auto EntryUPB = UsesPerBlock.find(&EntryBB);
EntryUPB != UsesPerBlock.end())
OnlyScanEntryBlock = EntryUPB->second.HasClobber;
if (OnlyScanEntryBlock) {
EntryCRL = VisitBlock(&EntryBB);
if (EntryCRL.empty())
return false;
} else {
// Now we have to go through CFG to get the initialized argument ranges
// across blocks. With dominance and post-dominance, the initialized ranges
// by a block include both accesses inside this block and accesses in its
// (transitive) successors. So visit successors before predecessors with a
// post-order walk of the blocks and memorize the results in "Initialized".
for (const BasicBlock *BB : post_order(&F)) {
ConstantRangeList CRL = VisitBlock(BB);
if (!CRL.empty())
Initialized[BB] = CRL;
}

auto EntryCRLI = Initialized.find(&EntryBB);
if (EntryCRLI == Initialized.end())
return false;

EntryCRL = EntryCRLI->second;
}

assert(!EntryCRL.empty() &&
"should have bailed already if EntryCRL is empty");

if (A.hasAttribute(Attribute::Initializes)) {
ConstantRangeList PreviousCRL =
A.getAttribute(Attribute::Initializes).getValueAsConstantRangeList();
if (PreviousCRL == EntryCRL)
return false;
EntryCRL = EntryCRL.unionWith(PreviousCRL);
}

A.addAttr(Attribute::get(A.getContext(), Attribute::Initializes,
EntryCRL.rangesRef()));

return true;
}

/// Deduce nocapture attributes for the SCC.
static void addArgumentAttrs(const SCCNodeSet &SCCNodes,
SmallSet<Function *, 8> &Changed) {
SmallSet<Function *, 8> &Changed,
bool SkipInitializes) {
ArgumentGraph AG;

// Check each function in turn, determining which pointer arguments are not
Expand Down Expand Up @@ -936,6 +1254,10 @@ static void addArgumentAttrs(const SCCNodeSet &SCCNodes,
if (addAccessAttr(&A, R))
Changed.insert(F);
}
if (!SkipInitializes && !A.onlyReadsMemory()) {
if (inferInitializes(A, *F))
Changed.insert(F);
}
}
}

Expand Down Expand Up @@ -1844,13 +2166,16 @@ deriveAttrsInPostOrder(ArrayRef<Function *> Functions, AARGetterT &&AARGetter,

SmallSet<Function *, 8> Changed;
if (ArgAttrsOnly) {
addArgumentAttrs(Nodes.SCCNodes, Changed);
// ArgAttrsOnly means to only infer attributes that may aid optimizations
// on the *current* function. "initializes" attribute is to aid
// optimizations (like DSE) on the callers, so skip "initializes" here.
addArgumentAttrs(Nodes.SCCNodes, Changed, /*SkipInitializes=*/true);
jvoung marked this conversation as resolved.
Show resolved Hide resolved
return Changed;
}

addArgumentReturnedAttrs(Nodes.SCCNodes, Changed);
addMemoryAttrs(Nodes.SCCNodes, AARGetter, Changed);
addArgumentAttrs(Nodes.SCCNodes, Changed);
addArgumentAttrs(Nodes.SCCNodes, Changed, /*SkipInitializes=*/false);
inferConvergent(Nodes.SCCNodes, Changed);
addNoReturnAttrs(Nodes.SCCNodes, Changed);
addWillReturn(Nodes.SCCNodes, Changed);
Expand Down
2 changes: 1 addition & 1 deletion llvm/test/Analysis/TypeBasedAliasAnalysis/functionattrs.ll
Original file line number Diff line number Diff line change
Expand Up @@ -15,7 +15,7 @@ define void @test0_yes(ptr %p) nounwind {
ret void
}

; CHECK: define void @test0_no(ptr nocapture writeonly %p) #1 {
; CHECK: define void @test0_no(ptr nocapture writeonly initializes((0, 4)) %p) #1 {
define void @test0_no(ptr %p) nounwind {
store i32 0, ptr %p, !tbaa !2
ret void
Expand Down
Original file line number Diff line number Diff line change
Expand Up @@ -10,7 +10,7 @@
; Should have call to sincos declarations, not calls to the asm pseudo-libcalls
define protected amdgpu_kernel void @swdev456865(ptr addrspace(1) %out0, ptr addrspace(1) %out1, ptr addrspace(1) %out2, float noundef %x) #0 {
; CHECK-LABEL: define protected amdgpu_kernel void @swdev456865(
; CHECK-SAME: ptr addrspace(1) nocapture writeonly [[OUT0:%.*]], ptr addrspace(1) nocapture writeonly [[OUT1:%.*]], ptr addrspace(1) nocapture writeonly [[OUT2:%.*]], float noundef [[X:%.*]]) local_unnamed_addr #[[ATTR0:[0-9]+]] {
; CHECK-SAME: ptr addrspace(1) nocapture writeonly initializes((0, 8)) [[OUT0:%.*]], ptr addrspace(1) nocapture writeonly initializes((0, 8)) [[OUT1:%.*]], ptr addrspace(1) nocapture writeonly initializes((0, 8)) [[OUT2:%.*]], float noundef [[X:%.*]]) local_unnamed_addr #[[ATTR0:[0-9]+]] {
; CHECK-NEXT: entry:
; CHECK-NEXT: [[__SINCOS_:%.*]] = alloca float, align 4, addrspace(5)
; CHECK-NEXT: [[I_I:%.*]] = call float @_Z6sincosfPU3AS5f(float [[X]], ptr addrspace(5) [[__SINCOS_]]) #[[ATTR1:[0-9]+]]
Expand Down
Loading