Skip to content

[opt] Teach mem2reg about begin_access. #30812

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.

Sign up for GitHub

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

Closed
wants to merge 1 commit into from
Closed

[opt] Teach mem2reg about begin_access. #30812

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
73 changes: 48 additions & 25 deletions lib/SILOptimizer/Transforms/SILMem2Reg.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -156,7 +156,7 @@ class MemoryToRegisters {
SILBuilder B;

/// Check if the AllocStackInst \p ASI is only written into.
bool isWriteOnlyAllocation(AllocStackInst *ASI);
bool isWriteOnlyAllocation(SingleValueInstruction *addr);

/// Promote all of the AllocStacks in a single basic block in one
/// linear scan. Note: This function deletes all of the users of the
Expand Down Expand Up @@ -227,7 +227,7 @@ static bool isAddressForLoad(SILInstruction *I, SILBasicBlock *&singleBlock,
/// Returns true if \p I is a dead struct_element_addr or tuple_element_addr.
static bool isDeadAddrProjection(SILInstruction *I) {
if (!isa<UncheckedAddrCastInst>(I) && !isa<StructElementAddrInst>(I) &&
!isa<TupleElementAddrInst>(I))
!isa<TupleElementAddrInst>(I) && !isa<BeginAccessInst>(I))
return false;

// Recursively search for uses which are dead themselves.
Expand All @@ -241,8 +241,8 @@ static bool isDeadAddrProjection(SILInstruction *I) {

/// Returns true if this AllocStacks is captured.
/// Sets \p inSingleBlock to true if all uses of \p ASI are in a single block.
static bool isCaptured(AllocStackInst *ASI, bool &inSingleBlock) {
static bool isCaptured(SingleValueInstruction *ASI, bool &inSingleBlock) {

SILBasicBlock *singleBlock = ASI->getParent();

// For all users of the AllocStack instruction.
Expand All @@ -262,9 +262,10 @@ static bool isCaptured(AllocStackInst *ASI, bool &inSingleBlock) {
if (SI->getDest() == ASI)
continue;

// Deallocation is also okay, as are DebugValueAddr. We will turn
// the latter into DebugValue.
if (isa<DeallocStackInst>(II) || isa<DebugValueAddrInst>(II))
// Deallocation is also okay, as are EndAccess and DebugValueAddr. We will
// turn the latter into DebugValue.
if (isa<DeallocStackInst>(II) || isa<DebugValueAddrInst>(II) ||
isa<EndAccessInst>(II))
continue;

// Destroys of loadable types can be rewritten as releases, so
Expand All @@ -273,6 +274,13 @@ static bool isCaptured(AllocStackInst *ASI, bool &inSingleBlock) {
if (DAI->getOperand()->getType().isLoadable(*DAI->getFunction()))
continue;

// Begin access is OK if it isn't captured.
if (auto *beginAccess = dyn_cast<BeginAccessInst>(II)) {
if (isCaptured(beginAccess, inSingleBlock))
return true;
continue;
}

// Other instructions are assumed to capture the AllocStack.
LLVM_DEBUG(llvm::dbgs() << "*** AllocStack is captured by: " << *II);
return true;
Expand All @@ -284,9 +292,9 @@ static bool isCaptured(AllocStackInst *ASI, bool &inSingleBlock) {
}

/// Returns true if the AllocStack is only stored into.
bool MemoryToRegisters::isWriteOnlyAllocation(AllocStackInst *ASI) {
bool MemoryToRegisters::isWriteOnlyAllocation(SingleValueInstruction *addr) {
// For all users of the AllocStack:
for (auto UI = ASI->use_begin(), E = ASI->use_end(); UI != E; ++UI) {
for (auto UI = addr->use_begin(), E = addr->use_end(); UI != E; ++UI) {
SILInstruction *II = UI->getUser();

// It is okay to store into this AllocStack.
Expand All @@ -300,11 +308,16 @@ bool MemoryToRegisters::isWriteOnlyAllocation(AllocStackInst *ASI) {

// If we haven't already promoted the AllocStack, we may see
// DebugValueAddr uses.
if (isa<DebugValueAddrInst>(II))
if (isa<DebugValueAddrInst>(II) || isDeadAddrProjection(II) ||
isa<EndAccessInst>(II))
continue;

if (isDeadAddrProjection(II))
continue;
// Begin access is OK if it is write only.
if (auto *beginAccess = dyn_cast<BeginAccessInst>(II)) {
if (isWriteOnlyAllocation(beginAccess))
continue;
return false;
}

// Can't do anything else with it.
LLVM_DEBUG(llvm::dbgs() << "*** AllocStack has non-write use: " << *II);
Expand Down Expand Up @@ -343,7 +356,7 @@ static bool isLoadFromStack(SILInstruction *I, AllocStackInst *ASI) {
ValueBase *op = I->getOperand(0);
while (op != ASI) {
if (!isa<UncheckedAddrCastInst>(op) && !isa<StructElementAddrInst>(op) &&
!isa<TupleElementAddrInst>(op))
!isa<TupleElementAddrInst>(op) && !isa<BeginAccessInst>(op))
return false;

op = cast<SingleValueInstruction>(op)->getOperand(0);
Expand All @@ -358,7 +371,7 @@ static void collectLoads(SILInstruction *I, SmallVectorImpl<LoadInst *> &Loads)
return;
}
if (!isa<UncheckedAddrCastInst>(I) && !isa<StructElementAddrInst>(I) &&
!isa<TupleElementAddrInst>(I))
!isa<TupleElementAddrInst>(I) && !isa<BeginAccessInst>(I))
return;

// Recursively search for other loads in the instruction's uses.
Expand All @@ -375,9 +388,13 @@ static void replaceLoad(LoadInst *LI, SILValue val, AllocStackInst *ASI) {

while (op != ASI) {
assert(isa<UncheckedAddrCastInst>(op) || isa<StructElementAddrInst>(op) ||
isa<TupleElementAddrInst>(op));
isa<TupleElementAddrInst>(op) || isa<BeginAccessInst>(op));
auto *Inst = cast<SingleValueInstruction>(op);
projections.push_back(Projection(Inst));

// We don't want to project these instructions but we do want to follow
// their operand.
if (!isa<BeginAccessInst>(op))
projections.push_back(Projection(Inst));
op = Inst->getOperand(0);
}

Expand Down Expand Up @@ -421,11 +438,17 @@ static void replaceLoad(LoadInst *LI, SILValue val, AllocStackInst *ASI) {

while (op != ASI && op->use_empty()) {
assert(isa<UncheckedAddrCastInst>(op) || isa<StructElementAddrInst>(op) ||
isa<TupleElementAddrInst>(op));
isa<TupleElementAddrInst>(op) || isa<BeginAccessInst>(op));
auto *Inst = cast<SingleValueInstruction>(op);
SILValue next = Inst->getOperand(0);
Inst->eraseFromParent();
op = next;
op = Inst->getOperand(0);
if (Inst->use_empty())
Inst->eraseFromParent();
if (auto singleUse = Inst->getSingleUse()) {
if (isa<EndAccessInst>(singleUse->getUser())) {
singleUse->getUser()->eraseFromParent();
Inst->eraseFromParent();
}
}
}
}

Expand Down Expand Up @@ -596,10 +619,10 @@ void MemoryToRegisters::removeSingleBlockAllocation(AllocStackInst *ASI) {
SILValue RunningVal = SILValue();

// For all instructions in the block.
for (auto BBI = BB->begin(), E = BB->end(); BBI != E;) {
SILInstruction *Inst = &*BBI;
++BBI;

SmallVector<SILInstruction *, 64> allInstInBlock;
for (auto &i : *BB)
allInstInBlock.push_back(&i);
for (auto *Inst : allInstInBlock) {
// Remove instructions that we are loading from. Replace the loaded value
// with our running value.
if (isLoadFromStack(Inst, ASI)) {
Expand Down Expand Up @@ -665,7 +688,7 @@ void MemoryToRegisters::removeSingleBlockAllocation(AllocStackInst *ASI) {
SILNode *Node = Inst;
while (isa<StructElementAddrInst>(Node) ||
isa<TupleElementAddrInst>(Node) ||
isa<UncheckedAddrCastInst>(Node)) {
isa<UncheckedAddrCastInst>(Node) || isa<BeginAccessInst>(Node)) {
auto *I = cast<SingleValueInstruction>(Node);
if (!I->use_empty()) break;
Node = I->getOperand(0);
Expand Down
93 changes: 93 additions & 0 deletions test/SILOptimizer/mem2reg.sil
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -492,3 +492,96 @@ bb3:
%11 = tuple ()
return %11 : $()
}

// CHECK-LABEL: sil @single_block_with_access
// CHECK: bb0
// CHECK: [[IL:%.*]] = integer_literal
// CHECK: [[X:%.*]] = struct $Int ([[IL]]
// CHECK: return [[X]]
// CHECK-LABEL: end sil function 'single_block_with_access'
sil @single_block_with_access : $@convention(thin) () -> Int {
bb0:
%1 = integer_literal $Builtin.Int64, 0 // user: %2
%2 = struct $Int (%1 : $Builtin.Int64) // user: %4
%3 = alloc_stack $Int // users: %8, %5, %4
store %2 to %3 : $*Int // id: %4
%5 = begin_access [read] [dynamic] [no_nested_conflict] %3 : $*Int // users: %6, %7
%6 = load %5 : $*Int // user: %9
end_access %5 : $*Int // id: %7
dealloc_stack %3 : $*Int // id: %8
return %6 : $Int // id: %9
}

// CHECK-LABEL: sil @multi_block_with_access
// CHECK-NOT: alloc_stack
// CHECK-NOT: begin_access

// CHECK: bb0(%0 : $Bool):
// CHECK: [[IL:%.*]] = integer_literal
// CHECK-NEXT: [[X:%.*]] = struct $Int ([[IL]]
// CHECK-NEXT: [[COND:%.*]] = struct_extract %0
// CHECK-NEXT: cond_br [[COND]]

// CHECK: bb2:
// CHECK-NEXT: [[IL2:%.*]] = integer_literal
// CHECK-NEXT: [[XVAL:%.*]] = struct_extract [[X]]
// CHECK-NEXT: integer_literal
// CHECK-NEXT: [[ADD1:%.*]] = builtin "sadd_with_overflow_Int64"([[XVAL]] : {{.*}}, [[IL2]]
// CHECK-NEXT: tuple_extract
// CHECK-NEXT: cond_fail
// CHECK-NEXT: br

// CHECK: bb3:
// CHECK-NEXT: [[XVAL2:%.*]] = struct_extract [[X]]
// CHECK-NEXT: [[IL3:%.*]] = integer_literal
// CHECK-NEXT: integer_literal
// CHECK-NEXT: [[ADD2:%.*]] = builtin "sadd_with_overflow_Int64"([[XVAL2]] : {{.*}}, [[IL3]]
// CHECK-NEXT: [[SUM_VAL:%.*]] = tuple_extract [[ADD2]]
// CHECK-NEXT: tuple_extract
// CHECK-NEXT: cond_fail
// CHECK-NEXT: [[OUT:%.*]] = struct $Int ([[SUM_VAL]]
// CHECK-NEXT: return [[OUT]]

// CHECK: } // end sil function 'multi_block_with_access'
sil @multi_block_with_access : $@convention(thin) (Bool) -> Int {
bb0(%0 : $Bool):
%2 = integer_literal $Builtin.Int64, 0 // user: %3
%3 = struct $Int (%2 : $Builtin.Int64) // user: %5
%4 = alloc_stack $Int // users: %26, %10, %22, %5
store %3 to %4 : $*Int // id: %5
%6 = struct_extract %0 : $Bool, #Bool._value // user: %7
cond_br %6, bb2, bb1 // id: %7

bb1: // Preds: bb0
br bb3 // id: %8

bb2: // Preds: bb0
%9 = integer_literal $Builtin.Int64, 1 // user: %14
%10 = begin_access [modify] [static] [no_nested_conflict] %4 : $*Int // users: %20, %19, %11
%11 = struct_element_addr %10 : $*Int, #Int._value // user: %12
%12 = load %11 : $*Builtin.Int64 // user: %14
%13 = integer_literal $Builtin.Int1, -1 // user: %14
%14 = builtin "sadd_with_overflow_Int64"(%12 : $Builtin.Int64, %9 : $Builtin.Int64, %13 : $Builtin.Int1) : $(Builtin.Int64, Builtin.Int1) // users: %16, %15
%15 = tuple_extract %14 : $(Builtin.Int64, Builtin.Int1), 0 // user: %18
%16 = tuple_extract %14 : $(Builtin.Int64, Builtin.Int1), 1 // user: %17
cond_fail %16 : $Builtin.Int1, "arithmetic overflow" // id: %17
%18 = struct $Int (%15 : $Builtin.Int64) // user: %19
store %18 to %10 : $*Int // id: %19
end_access %10 : $*Int // id: %20
br bb3 // id: %21

bb3: // Preds: bb1 bb2
%22 = begin_access [read] [static] [no_nested_conflict] %4 : $*Int // users: %25, %23
%23 = struct_element_addr %22 : $*Int, #Int._value // user: %24
%24 = load %23 : $*Builtin.Int64 // user: %29
end_access %22 : $*Int // id: %25
dealloc_stack %4 : $*Int // id: %26
%27 = integer_literal $Builtin.Int64, 2 // user: %29
%28 = integer_literal $Builtin.Int1, -1 // user: %29
%29 = builtin "sadd_with_overflow_Int64"(%24 : $Builtin.Int64, %27 : $Builtin.Int64, %28 : $Builtin.Int1) : $(Builtin.Int64, Builtin.Int1) // users: %31, %30
%30 = tuple_extract %29 : $(Builtin.Int64, Builtin.Int1), 0 // user: %33
%31 = tuple_extract %29 : $(Builtin.Int64, Builtin.Int1), 1 // user: %32
cond_fail %31 : $Builtin.Int1, "arithmetic overflow" // id: %32
%33 = struct $Int (%30 : $Builtin.Int64) // user: %34
return %33 : $Int // id: %34
}