ldc/gen/pgo_ASTbased.cpp
liushuyu fc4d884d2d
Multiple changes ...
* Use getVoidPtrType helper
* Replace #define with constexpr
2024-03-26 14:48:36 -06:00

1144 lines
40 KiB
C++
Raw Blame History

This file contains ambiguous Unicode characters

This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.

//===-- gen/pgo_ASTbased.cpp ------------------------------------*- C++ -*-===//
//
// LDC the LLVM D compiler
//
// This file is adapted from CodeGenPGO.cpp (Clang, LLVM). Therefore,
// this file is distributed under the LLVM license.
// See the LICENSE file for details.
//
//===----------------------------------------------------------------------===//
//
// Instrumentation-based profile-guided optimization. This is AST-based PGO, in
// contrast to LLVM's IR-based PGO.
//
//===----------------------------------------------------------------------===//
#include "gen/pgo_ASTbased.h"
#include "dmd/errors.h"
#include "dmd/expression.h"
#include "dmd/globals.h"
#include "dmd/init.h"
#include "dmd/statement.h"
#include "driver/cl_options_instrumentation.h"
#include "gen/irstate.h"
#include "gen/llvm.h"
#include "gen/logger.h"
#include "gen/recursivevisitor.h"
#include "gen/tollvm.h"
#include "llvm/IR/Intrinsics.h"
#include "llvm/IR/MDBuilder.h"
#include "llvm/ProfileData/InstrProfReader.h"
#include "llvm/Support/Endian.h"
#include "llvm/Support/FileSystem.h"
#include "llvm/Support/MD5.h"
namespace {
llvm::cl::opt<bool, false, opts::FlagParser<bool>> enablePGOIndirectCalls(
"pgo-indirect-calls", llvm::cl::ZeroOrMore, llvm::cl::Hidden,
llvm::cl::desc("(*) Enable PGO of indirect calls"),
llvm::cl::init(true));
}
#if LDC_LLVM_VER >= 1800
namespace llvm::support {
const auto little = llvm::endianness::little;
const auto big = llvm::endianness::big;
}
#endif
/// \brief Stable hasher for PGO region counters.
///
/// PGOHash produces a stable hash of a given function's control flow.
/// It is used to detect whether the function has changed from the function with
/// the same name for which profile information is available.
/// Because only control flow is input to the hasher, other changes are not
/// detected and possibly wrong profiling data will be used. An example of
/// an undetected change is:
/// - if (x==0) {...}
/// + if (y==0) {...}
/// This can obviously result in very wrong branch weights. It's up for debate
/// whether these kind of changes should be detected or not; it is probably
/// difficult to distinguish such changes from simple changes in a variables
/// name.
///
/// Changing the output of this hash will invalidate all previously generated
/// profiles -- i.e., do it only with very strong arguments.
///
/// \note When this hash does eventually change (years?), we still need to
/// support old hashes. We'll need to pull in the version number from the
/// profile data format and use the matching hash function.
class PGOHash {
uint64_t Working;
unsigned Count;
llvm::MD5 MD5;
static const int NumBitsPerType = 6;
static const unsigned NumTypesPerWord = sizeof(uint64_t) * 8 / NumBitsPerType;
static const unsigned TooBig = 1u << NumBitsPerType;
public:
// TODO: When this format changes, take in a version number here, and use the
// old hash calculation for file formats that used the old hash.
PGOHash() : Working(0), Count(0) {}
/// \brief Hash values for AST nodes.
///
/// Distinct values for AST nodes that have region counters attached.
///
/// These values must be stable. All new members must be added at the end,
/// and no members should be removed. Changing the enumeration value for an
/// AST node will affect the hash of every function that contains that node.
enum HashType : unsigned char {
None = 0,
LabelStmt = 1,
WhileStmt,
DoStmt,
ForStmt,
ForeachStmt,
ForeachRangeStmt,
SwitchStmt,
CaseStmt,
DefaultStmt,
CaseGoto,
IfStmt,
TryCatchStmt,
TryCatchCatch,
TryFinallyStmt,
ConditionalExpr,
AndAndExpr,
OrOrExpr,
UnrolledLoopIterationScope,
// Keep this last. It's for the static assert that follows.
LastHashType
};
static_assert(LastHashType <= TooBig, "Too many types in HashType");
void combine(HashType Type) {
// Check that we never combine 0 and only have six bits.
assert(Type && "Hash is invalid: unexpected type 0");
assert(unsigned(Type) < TooBig && "Hash is invalid: too many types");
// Pass through MD5 if enough work has built up.
if (Count && Count % NumTypesPerWord == 0) {
using namespace llvm::support;
uint64_t Swapped = endian::byte_swap<uint64_t, little>(Working);
MD5.update(llvm::ArrayRef<uint8_t>((uint8_t *)&Swapped, sizeof(Swapped)));
Working = 0;
}
// Accumulate the current type.
++Count;
Working = Working << NumBitsPerType | Type;
}
uint64_t finalize() {
// Use Working as the hash directly if we never used MD5.
if (Count <= NumTypesPerWord)
// No need to byte swap here, since none of the math was endian-dependent.
// This number will be byte-swapped as required on endianness transitions,
// so we will see the same value on the other side.
return Working;
// Check for remaining work in Working.
if (Working) {
using namespace llvm::support;
uint64_t Swapped = endian::byte_swap<uint64_t, little>(Working);
MD5.update(llvm::ArrayRef<uint8_t>((uint8_t *)&Swapped, sizeof(Swapped)));
}
// Finalize the MD5 and return the hash.
llvm::MD5::MD5Result Result;
MD5.final(Result);
return Result.low();
}
};
///////////////////////////////////////////////////////////////////////////////
///////////////////////////////////////////////////////////////////////////////
/// An ASTVisitor that fills a map of (statements -> PGO counter numbers).
struct MapRegionCounters : public StoppableVisitor {
/// The next counter value to assign.
unsigned NextCounter;
/// The function hash.
PGOHash Hash;
/// The map of statements to counters.
llvm::DenseMap<const RootObject *, unsigned> &CounterMap;
MapRegionCounters(llvm::DenseMap<const RootObject *, unsigned> &CounterMap)
: NextCounter(0), CounterMap(CounterMap) {}
using StoppableVisitor::visit;
// FIXME: this macro should also stop deeper traversal at duplicate nodes, using
// "stop=false;"
// However, the regexp microbench by David breaks in that case. I feel there
// is a bug lingering somewhere: needs further investigation!
#define SKIP_VISITED(Stmt) \
do { \
if (CounterMap.count(Stmt)) { \
return; \
} \
} while (0)
void visit(Statement *stmt) override {
// If this assert fails, a new statement type was added to the frontend, and
// thus we need to decide on how to handle PGO calculations for that, both
// in MapRegionCounters and in ComputeRegionCounts.
assert(0 && "All statement types should be explicitly handled to avoid "
"missing new statement types in MapRegionCounters and "
"ComputeRegionCounts");
}
void visit(CompoundStatement *) override {}
void visit(ExpStatement *) override {}
void visit(ImportStatement *) override {}
void visit(ScopeStatement *) override {}
void visit(ReturnStatement *) override {}
void visit(StaticAssertStatement *) override {}
void visit(MixinStatement *) override {}
void visit(ScopeGuardStatement *) override {}
void visit(ConditionalStatement *) override {}
void visit(StaticForeachStatement *) override {}
void visit(PragmaStatement *) override {}
void visit(BreakStatement *) override {}
void visit(ContinueStatement *) override {}
void visit(GotoDefaultStatement *) override {}
void visit(GotoCaseStatement *) override {}
void visit(GotoStatement *) override {}
void visit(SynchronizedStatement *) override {}
void visit(WithStatement *) override {}
void visit(ThrowStatement *) override {}
void visit(AsmStatement *) override {}
void visit(Expression *) override {}
void visit(Declaration *) override {}
void visit(Initializer *) override {}
void visit(Dsymbol *) override {}
void visit(FuncDeclaration *fd) override {
if (NextCounter) {
// This is a nested function declaration. Don't add counters for it, as it
// is treated as a separate function elsewhere in the AST.
// Stop recursion at this depth.
stop = true;
} else {
CounterMap[fd->fbody] = NextCounter++;
}
}
void visit(IfStatement *stmt) override {
SKIP_VISITED(stmt);
CounterMap[stmt] = NextCounter++;
Hash.combine(PGOHash::IfStmt);
}
void visit(WhileStatement *stmt) override {
SKIP_VISITED(stmt);
CounterMap[stmt] = NextCounter++;
Hash.combine(PGOHash::WhileStmt);
}
void visit(DoStatement *stmt) override {
SKIP_VISITED(stmt);
CounterMap[stmt] = NextCounter++;
Hash.combine(PGOHash::DoStmt);
}
void visit(ForStatement *stmt) override {
SKIP_VISITED(stmt);
CounterMap[stmt] = NextCounter++;
Hash.combine(PGOHash::ForStmt);
}
void visit(ForeachStatement *stmt) override {
SKIP_VISITED(stmt);
CounterMap[stmt] = NextCounter++;
Hash.combine(PGOHash::ForeachStmt);
}
void visit(ForeachRangeStatement *stmt) override {
SKIP_VISITED(stmt);
CounterMap[stmt] = NextCounter++;
Hash.combine(PGOHash::ForeachRangeStmt);
}
void visit(UnrolledLoopStatement *stmt) override {
// Continues and breaks in the scopes of UnrolledLoop iterations can "goto"
// to "labels" at the start of the next iteration or end of the loop. We
// should therefore treat each unrolled loop iteration the same as
// LabelStatements (with redundant counting of the first iteration which is
// always executed).
// The counter for the UnrolledLoopStatement itself counts the
// exit block of the 'loop'.
SKIP_VISITED(stmt);
CounterMap[stmt] = NextCounter++;
Hash.combine(PGOHash::UnrolledLoopIterationScope);
for (auto s : *stmt->statements) {
CounterMap[s] = NextCounter++;
Hash.combine(PGOHash::UnrolledLoopIterationScope);
}
}
void visit(LabelStatement *stmt) override {
SKIP_VISITED(stmt);
CounterMap[stmt] = NextCounter++;
Hash.combine(PGOHash::LabelStmt);
}
void visit(SwitchStatement *stmt) override {
SKIP_VISITED(stmt);
CounterMap[stmt] = NextCounter++;
Hash.combine(PGOHash::SwitchStmt);
}
void visit(CaseStatement *stmt) override {
SKIP_VISITED(stmt);
CounterMap[stmt] = NextCounter++;
Hash.combine(PGOHash::CaseStmt);
// Iff this statement is the target of a goto case statement, add an extra
// counter for this case (as if it is a label statement).
if (stmt->gototarget) {
CounterMap[CodeGenPGO::getCounterPtr(stmt, 1)] = NextCounter++;
Hash.combine(PGOHash::CaseGoto);
}
}
void visit(CaseRangeStatement *stmt) override {
assert(0 &&
"Case range statement should be lowered to regular case statements");
}
void visit(DefaultStatement *stmt) override {
SKIP_VISITED(stmt);
CounterMap[stmt] = NextCounter++;
Hash.combine(PGOHash::DefaultStmt);
// Iff this statement is the target of a goto case statement, add an extra
// counter for this case (as if it is a label statement).
if (stmt->gototarget) {
CounterMap[CodeGenPGO::getCounterPtr(stmt, 1)] = NextCounter++;
Hash.combine(PGOHash::CaseGoto);
}
}
void visit(TryCatchStatement *stmt) override {
SKIP_VISITED(stmt);
CounterMap[stmt] = NextCounter++;
Hash.combine(PGOHash::TryCatchStmt);
// Note that this results in the exception counters obtaining their counter
// numbers before recursing into the counter handlers:
for (auto c : *stmt->catches) {
CounterMap[c] = NextCounter++;
Hash.combine(PGOHash::TryCatchCatch);
}
}
void visit(TryFinallyStatement *stmt) override {
// If there is nothing to "try" or no cleanup, do nothing:
if (!stmt->_body || !stmt->finalbody)
return;
SKIP_VISITED(stmt);
CounterMap[stmt] = NextCounter++;
Hash.combine(PGOHash::TryFinallyStmt);
}
void visit(CondExp *expr) override {
SKIP_VISITED(expr);
CounterMap[expr] = NextCounter++;
Hash.combine(PGOHash::ConditionalExpr);
}
void visit(LogicalExp *expr) override {
SKIP_VISITED(expr);
CounterMap[expr] = NextCounter++;
Hash.combine(expr->op == EXP::andAnd ? PGOHash::AndAndExpr
: PGOHash::OrOrExpr);
}
#undef SKIP_VISITED
};
///////////////////////////////////////////////////////////////////////////////
///////////////////////////////////////////////////////////////////////////////
/// An Recursive AST Visitor that propagates the raw counts through the AST and
/// records the count at statements where the value may change.
struct ComputeRegionCounts : public RecursiveVisitor {
/// PGO state.
CodeGenPGO &PGO;
/// A flag that is set when the current count should be recorded on the
/// next statement, such as at the exit of a loop.
bool RecordNextStmtCount;
/// The count at the current location in the traversal.
uint64_t CurrentCount;
/// The map of statements to count values.
llvm::DenseMap<const RootObject *, uint64_t> &CountMap;
/// BreakContinueStack - Keep counts of breaks and continues inside loops.
struct BreakContinue {
uint64_t BreakCount;
uint64_t ContinueCount;
BreakContinue() : BreakCount(0), ContinueCount(0) {}
};
llvm::SmallVector<BreakContinue, 8> BreakContinueStack;
struct LoopLabel {
// If a label is used as break/continue target, this struct stores the
// BreakContinue stack index at the label point
LabelStatement *label;
size_t stackindex;
LoopLabel(LabelStatement *_label, size_t index)
: label(_label), stackindex(index) {}
};
llvm::SmallVector<LoopLabel, 8> LoopLabels;
ComputeRegionCounts(llvm::DenseMap<const RootObject *, uint64_t> &CountMap,
CodeGenPGO &PGO)
: PGO(PGO), RecordNextStmtCount(false), CountMap(CountMap) {}
void RecordStmtCount(const RootObject *S) {
if (RecordNextStmtCount) {
CountMap[S] = CurrentCount;
RecordNextStmtCount = false;
}
}
/// Set and return the current count.
uint64_t setCount(uint64_t Count) {
CurrentCount = Count;
return Count;
}
using RecursiveVisitor::visit;
void visit(FuncDeclaration *fd) override {
// Counter tracks entry to the function body.
uint64_t BodyCount = setCount(PGO.getRegionCount(fd->fbody));
CountMap[fd->fbody] = BodyCount;
recurse(fd->fbody);
}
void visit(Statement *S) override { RecordStmtCount(S); }
void visit(ReturnStatement *S) override {
RecordStmtCount(S);
recurse(S->exp);
CurrentCount = 0;
RecordNextStmtCount = true;
}
void visit(ThrowStatement *S) override {
RecordStmtCount(S);
recurse(S->exp);
CurrentCount = 0;
RecordNextStmtCount = true;
}
void visit(GotoStatement *S) override {
RecordStmtCount(S);
CurrentCount = 0;
RecordNextStmtCount = true;
}
void visit(LabelStatement *S) override {
RecordNextStmtCount = false;
// Counter tracks the block following the label.
uint64_t BlockCount = setCount(PGO.getRegionCount(S));
CountMap[S] = BlockCount;
// For each label pointing to a loop, store the current index of
// BreakContinueStack. This is needed for `break label;` and `continue
// label;` statements in loops.
// Assume all labels point to loops. (TODO: find predicate to filter which
// labels to add)
LoopLabels.push_back(LoopLabel(S, BreakContinueStack.size()));
recurse(S->statement);
}
void visit(BreakStatement *S) override {
RecordStmtCount(S);
assert(!BreakContinueStack.empty() && "break not in a loop or switch!");
if (S->target) {
auto it = std::find_if(
LoopLabels.begin(), LoopLabels.end(),
[S](const LoopLabel &LL) { return LL.label == S->target; });
assert(it != LoopLabels.end() && "It is not possible to break to a label "
"that has not been visited yet");
auto LL = *it;
assert(LL.stackindex < BreakContinueStack.size());
BreakContinueStack[LL.stackindex].BreakCount += CurrentCount;
} else {
BreakContinueStack.back().BreakCount += CurrentCount;
}
CurrentCount = 0;
RecordNextStmtCount = true;
}
void visit(ContinueStatement *S) override {
RecordStmtCount(S);
assert(!BreakContinueStack.empty() && "continue stmt not in a loop!");
if (S->target) {
auto it = std::find_if(
LoopLabels.begin(), LoopLabels.end(),
[S](const LoopLabel &LL) { return LL.label == S->target; });
assert(it != LoopLabels.end() &&
"It is not possible to continue to a label "
"that has not been visited yet");
auto LL = *it;
assert(LL.stackindex < BreakContinueStack.size());
BreakContinueStack[LL.stackindex].ContinueCount += CurrentCount;
} else {
BreakContinueStack.back().ContinueCount += CurrentCount;
}
CurrentCount = 0;
RecordNextStmtCount = true;
}
void visit(UnrolledLoopStatement *S) override {
RecordStmtCount(S);
// The iterations can still have `break` and `continue` even though we are
// no longer in a loop. We need to provide this BreakContinue struct for
// those breaks&continues to refer to, but we do not use it otherwise.
BreakContinueStack.push_back(BreakContinue());
// Iteration statement counters track the entry block of each iteration
// (redundant for first iteration)
for (auto iteration_stmt : *S->statements) {
setCount(PGO.getRegionCount(iteration_stmt));
RecordNextStmtCount = true;
recurse(iteration_stmt);
}
BreakContinueStack.pop_back();
// UnrolledLoopStatement counter tracks the continuation after the statement.
setCount(PGO.getRegionCount(S));
RecordNextStmtCount = true;
}
void visit(WhileStatement *S) override {
RecordStmtCount(S);
uint64_t ParentCount = CurrentCount;
BreakContinueStack.push_back(BreakContinue());
// Visit the body region first so the break/continue adjustments can be
// included when visiting the condition.
uint64_t BodyCount = setCount(PGO.getRegionCount(S));
CountMap[S->_body] = CurrentCount;
recurse(S->_body);
uint64_t BackedgeCount = CurrentCount;
// ...then go back and propagate counts through the condition. The count
// at the start of the condition is the sum of the incoming edges,
// the backedge from the end of the loop body, and the edges from
// continue statements.
BreakContinue BC = BreakContinueStack.pop_back_val();
uint64_t CondCount =
setCount(ParentCount + BackedgeCount + BC.ContinueCount);
CountMap[S->condition] = CondCount;
recurse(S->condition);
setCount(BC.BreakCount + CondCount - BodyCount);
RecordNextStmtCount = true;
}
void visit(DoStatement *S) override {
RecordStmtCount(S);
uint64_t FallThroughCount = CurrentCount;
// The instr count includes the fallthrough from the parent scope.
BreakContinueStack.push_back(BreakContinue());
uint64_t BodyCount = setCount(PGO.getRegionCount(S));
CountMap[S->_body] = BodyCount;
recurse(S->_body);
uint64_t BackedgeCount = CurrentCount;
BreakContinue BC = BreakContinueStack.pop_back_val();
// The count at the start of the condition is equal to the count at the
// end of the body, plus any continues.
uint64_t CondCount = setCount(BackedgeCount + BC.ContinueCount);
CountMap[S->condition] = CondCount;
recurse(S->condition);
uint64_t LoopCount = BodyCount - FallThroughCount;
setCount(BC.BreakCount + CondCount - LoopCount);
RecordNextStmtCount = true;
}
void visit(ForStatement *S) override {
RecordStmtCount(S);
recurse(S->_init);
uint64_t ParentCount = CurrentCount;
BreakContinueStack.push_back(BreakContinue());
// Visit the body region first. (This is basically the same as a while
// loop; see further comments in VisitWhileStmt.)
uint64_t BodyCount = setCount(PGO.getRegionCount(S));
CountMap[S->_body] = BodyCount;
recurse(S->_body);
uint64_t BackedgeCount = CurrentCount;
BreakContinue BC = BreakContinueStack.pop_back_val();
// The increment is essentially part of the body but it needs to include
// the count for all the continue statements.
if (S->increment) {
uint64_t IncCount = setCount(BackedgeCount + BC.ContinueCount);
CountMap[S->increment] = IncCount;
recurse(S->increment);
}
// ...then go back and propagate counts through the condition.
uint64_t CondCount =
setCount(ParentCount + BackedgeCount + BC.ContinueCount);
// If condition is nullptr, store CondCount in a derived ptr
CountMap[S->condition ? S->condition : PGO.getCounterPtr(S, 1)] = CondCount;
recurse(S->condition);
setCount(BC.BreakCount + CondCount - BodyCount);
RecordNextStmtCount = true;
}
void visit(ForeachStatement *S) override {
RecordStmtCount(S);
recurse(S->aggr);
uint64_t ParentCount = CurrentCount;
BreakContinueStack.push_back(BreakContinue());
// Visit the body region first. (This is basically the same as a while
// loop; see further comments in VisitWhileStmt.)
uint64_t BodyCount = setCount(PGO.getRegionCount(S));
CountMap[S->_body] = BodyCount;
recurse(S->_body);
uint64_t BackedgeCount = CurrentCount;
BreakContinue BC = BreakContinueStack.pop_back_val();
uint64_t CondCount = ParentCount + BackedgeCount + BC.ContinueCount;
// save the condition count as the second counter for the foreach statement
// (there is no explicit condition statement).
CountMap[PGO.getCounterPtr(S, 1)] = CondCount;
setCount(BC.BreakCount + CondCount - BodyCount);
RecordNextStmtCount = true;
}
void visit(ForeachRangeStatement *S) override {
RecordStmtCount(S);
recurse(S->lwr);
recurse(S->upr);
uint64_t ParentCount = CurrentCount;
BreakContinueStack.push_back(BreakContinue());
// Visit the body region first. (This is basically the same as a while
// loop; see further comments in VisitWhileStmt.)
uint64_t BodyCount = setCount(PGO.getRegionCount(S));
CountMap[S->_body] = BodyCount;
recurse(S->_body);
uint64_t BackedgeCount = CurrentCount;
BreakContinue BC = BreakContinueStack.pop_back_val();
uint64_t CondCount = ParentCount + BackedgeCount + BC.ContinueCount;
// save the condition count as the second counter for the foreach statement
// (there is no explicit condition statement).
CountMap[PGO.getCounterPtr(S, 1)] = CondCount;
setCount(BC.BreakCount + CondCount - BodyCount);
RecordNextStmtCount = true;
}
void visit(SwitchStatement *S) override {
RecordStmtCount(S);
recurse(S->condition);
CurrentCount = 0;
BreakContinueStack.push_back(BreakContinue());
recurse(S->_body);
// If the switch is inside a loop, add the continue counts.
BreakContinue BC = BreakContinueStack.pop_back_val();
if (!BreakContinueStack.empty())
BreakContinueStack.back().ContinueCount += BC.ContinueCount;
// Counter tracks the exit block of the switch.
setCount(PGO.getRegionCount(S));
RecordNextStmtCount = true;
}
void visit(CaseStatement *S) override {
// Counter for this particular case. This counts only jumps from the
// switch header and does not include fallthrough from the case before
// this one. We need the count without fallthrough in the mapping, so it's
// more useful for branch probabilities.
uint64_t CaseCount = PGO.getRegionCount(S);
CountMap[S] = CaseCount;
// If this Case is the target of a goto case, it will have its own extra
// counter and behaves like a LabelStatement.
if (S->gototarget) {
RootObject *cntr = PGO.getCounterPtr(S, 1);
CountMap[cntr] = setCount(PGO.getRegionCount(cntr));
} else {
setCount(CurrentCount + CaseCount);
}
RecordNextStmtCount = true;
recurse(S->statement);
}
void visit(DefaultStatement *S) override {
// Identical to CaseStatement handler.
uint64_t CaseCount = PGO.getRegionCount(S);
CountMap[S] = CaseCount;
if (S->gototarget) {
RootObject *cntr = PGO.getCounterPtr(S, 1);
CountMap[cntr] = setCount(PGO.getRegionCount(cntr));
} else {
setCount(CurrentCount + CaseCount);
}
RecordNextStmtCount = true;
recurse(S->statement);
}
void visit(GotoDefaultStatement *S) override {
// Identical to GotoStatement
RecordStmtCount(S);
CurrentCount = 0;
RecordNextStmtCount = true;
}
void visit(GotoCaseStatement *S) override {
// Identical to GotoStatement
RecordStmtCount(S);
CurrentCount = 0;
RecordNextStmtCount = true;
}
void visit(IfStatement *S) override {
RecordStmtCount(S);
uint64_t ParentCount = CurrentCount;
recurse(S->condition);
// Counter tracks the "then" part of an if statement. The count for
// the "else" part, if it exists, will be calculated from this counter.
uint64_t ThenCount = setCount(PGO.getRegionCount(S));
CountMap[S->ifbody] = ThenCount;
recurse(S->ifbody);
uint64_t OutCount = CurrentCount;
uint64_t ElseCount = ParentCount - ThenCount;
if (S->elsebody) {
setCount(ElseCount);
CountMap[S->elsebody] = ElseCount;
recurse(S->elsebody);
OutCount += CurrentCount;
} else {
OutCount += ElseCount;
}
setCount(OutCount);
RecordNextStmtCount = true;
}
void visit(TryCatchStatement *S) override {
RecordStmtCount(S);
// Because the order of codegen, the body is generated after the catch
// handlers and the current count (from the try statement) will be wrong
// going into codegen for the body. Safest to store the current count in the
// body too.
RecordNextStmtCount = true;
recurse(S->_body);
for (auto c : *S->catches) {
// Catch counter tracks the entry block of catch handler
setCount(PGO.getRegionCount(c));
RecordNextStmtCount = true;
recurse(c->handler);
}
// Try counter tracks the continuation block of the try statement.
setCount(PGO.getRegionCount(S));
RecordNextStmtCount = true;
}
void visit(TryFinallyStatement *S) override {
RecordStmtCount(S);
uint64_t ParentCount = CurrentCount;
// Because the order of codegen, the body is generated after the catch
// handlers and the current count (from the try statement) will be wrong
// going into codegen for the body. Safest to store the current count in the
// body too.
RecordNextStmtCount = true;
recurse(S->_body);
// Finally is always executed, so has same incoming count as the parent
// count of the try statement.
setCount(ParentCount);
RecordNextStmtCount = true;
recurse(S->finalbody);
// The TryFinally counter tracks the continuation block of the try
// statement.
setCount(PGO.getRegionCount(S));
RecordNextStmtCount = true;
}
void visit(CondExp *E) override {
RecordStmtCount(E);
uint64_t ParentCount = CurrentCount;
recurse(E->econd);
// Counter tracks the "true" part of a conditional operator. The
// count in the "false" part will be calculated from this counter.
uint64_t TrueCount = setCount(PGO.getRegionCount(E));
CountMap[E->e1] = TrueCount;
recurse(E->e1);
uint64_t OutCount = CurrentCount;
uint64_t FalseCount = setCount(ParentCount - TrueCount);
CountMap[E->e2] = FalseCount;
recurse(E->e2);
OutCount += CurrentCount;
setCount(OutCount);
RecordNextStmtCount = true;
}
void visit(LogicalExp *E) override {
RecordStmtCount(E);
uint64_t ParentCount = CurrentCount;
recurse(E->e1);
// Counter tracks the right hand side of a logical operator.
uint64_t RHSCount = setCount(PGO.getRegionCount(E));
CountMap[E->e2] = RHSCount;
recurse(E->e2);
setCount(ParentCount + RHSCount - CurrentCount);
RecordNextStmtCount = true;
}
};
///////////////////////////////////////////////////////////////////////////////
///////////////////////////////////////////////////////////////////////////////
// Pointer math to add an extra counter for one statement/expression.
// Increasing (the size_t value of) the pointer by counter results in a new
// "pointer" that will never clash with the other RootObject pointers (the size
// of a statement/expression object is much larger).
RootObject *CodeGenPGO::getCounterPtr(const RootObject *ptr,
unsigned counter_idx) {
return reinterpret_cast<RootObject *>(reinterpret_cast<size_t>(ptr) +
counter_idx);
}
void CodeGenPGO::setFuncName(llvm::StringRef Name,
llvm::GlobalValue::LinkageTypes Linkage) {
llvm::IndexedInstrProfReader *PGOReader = gIR->getPGOReader();
FuncName = llvm::getPGOFuncName(Name, Linkage, "",
PGOReader ? PGOReader->getVersion()
: llvm::IndexedInstrProf::Version);
// If we're generating a profile, create a variable for the name.
if (opts::isInstrumentingForASTBasedPGO() && emitInstrumentation) {
FuncNameVar = llvm::createPGOFuncNameVar(gIR->module, Linkage, FuncName);
// If Linkage is private, and the function is in a comdat "any" group, set
// the linkage to internal to prevent LLVM from erroring with "comdat global
// value has private linkage".
if (needsCOMDAT() &&
FuncNameVar->getLinkage() == llvm::GlobalValue::PrivateLinkage) {
FuncNameVar->setLinkage(llvm::GlobalValue::InternalLinkage);
}
}
}
void CodeGenPGO::setFuncName(llvm::Function *fn) {
setFuncName(fn->getName(), fn->getLinkage());
}
void CodeGenPGO::assignRegionCounters(const FuncDeclaration *D,
llvm::Function *fn) {
llvm::IndexedInstrProfReader *PGOReader = gIR->getPGOReader();
if (!opts::isInstrumentingForASTBasedPGO() && !PGOReader)
return;
emitInstrumentation = D->emitInstrumentation;
setFuncName(fn);
mapRegionCounters(D);
if (PGOReader) {
loadRegionCounts(PGOReader, D);
computeRegionCounts(D);
applyFunctionAttributes(fn);
}
}
void CodeGenPGO::mapRegionCounters(const FuncDeclaration *D) {
RegionCounterMap.reset(new llvm::DenseMap<const RootObject *, unsigned>);
MapRegionCounters regioncounter(*RegionCounterMap);
RecursiveWalker walker(&regioncounter);
walker.visit(const_cast<FuncDeclaration *>(D));
assert(regioncounter.NextCounter > 0 && "no entry counter mapped for decl");
assert(regioncounter.NextCounter == RegionCounterMap->size());
NumRegionCounters = regioncounter.NextCounter;
FunctionHash = regioncounter.Hash.finalize();
}
void CodeGenPGO::computeRegionCounts(const FuncDeclaration *FD) {
StmtCountMap.reset(new llvm::DenseMap<const RootObject *, uint64_t>);
ComputeRegionCounts Walker(*StmtCountMap, *this);
Walker.visit(const_cast<FuncDeclaration *>(FD));
}
/// Apply attributes to llvm::Function based on profiling data.
void CodeGenPGO::applyFunctionAttributes(llvm::Function *Fn) {
if (!haveRegionCounts())
return;
uint64_t FunctionCount = getRegionCount(nullptr);
Fn->setEntryCount(FunctionCount);
}
void CodeGenPGO::emitCounterIncrement(const RootObject *S) const {
if (!opts::isInstrumentingForASTBasedPGO() || !RegionCounterMap ||
!emitInstrumentation)
return;
auto counter_it = (*RegionCounterMap).find(S);
assert(counter_it != (*RegionCounterMap).end() &&
"Statement not found in PGO counter map!");
unsigned counter = counter_it->second;
auto *I8PtrTy = getVoidPtrType();
gIR->ir->CreateCall(GET_INTRINSIC_DECL(instrprof_increment),
{llvm::ConstantExpr::getBitCast(FuncNameVar, I8PtrTy),
gIR->ir->getInt64(FunctionHash),
gIR->ir->getInt32(NumRegionCounters),
gIR->ir->getInt32(counter)});
}
void CodeGenPGO::loadRegionCounts(llvm::IndexedInstrProfReader *PGOReader,
const FuncDeclaration *fd) {
RegionCounts.clear();
llvm::Expected<llvm::InstrProfRecord> RecordExpected =
PGOReader->getInstrProfRecord(FuncName, FunctionHash);
auto EC = RecordExpected.takeError();
if (EC) {
#if LDC_LLVM_VER >= 1700
auto IPE = std::get<0>(llvm::InstrProfError::take(std::move(EC)));
#else
auto IPE = llvm::InstrProfError::take(std::move(EC));
#endif
if (IPE == llvm::instrprof_error::unknown_function) {
IF_LOG Logger::println("No profile data for function: %s",
FuncName.c_str());
// Don't output a compiler warning when profile data is missing for a
// function, because it could be intentional.
} else if (IPE == llvm::instrprof_error::hash_mismatch) {
IF_LOG Logger::println(
"Ignoring profile data: hash mismatch for function: %s",
FuncName.c_str());
warning(fd->loc,
"Ignoring profile data for function `%s` (`%s`): "
"control-flow hash mismatch",
const_cast<FuncDeclaration *>(fd)->toPrettyChars(),
FuncName.c_str());
} else if (IPE == llvm::instrprof_error::malformed) {
IF_LOG Logger::println("Profile data is malformed for function: %s",
FuncName.c_str());
warning(fd->loc,
"Ignoring profile data for function `%s` (`%s`): "
"control-flow hash mismatch",
const_cast<FuncDeclaration *>(fd)->toPrettyChars(),
FuncName.c_str());
} else {
IF_LOG Logger::println("Error loading profile counts for function: %s",
FuncName.c_str());
warning(fd->loc, "Error loading profile data for function `%s` (`%s`)",
const_cast<FuncDeclaration *>(fd)->toPrettyChars(),
FuncName.c_str());
}
RegionCounts.clear();
return;
}
ProfRecord =
std::make_unique<llvm::InstrProfRecord>(std::move(RecordExpected.get()));
RegionCounts = ProfRecord->Counts;
IF_LOG Logger::println("Loaded profile data for function: %s",
FuncName.c_str());
}
/// \brief Calculate what to divide by to scale weights.
///
/// Given the maximum weight, calculate a divisor that will scale all the
/// weights to strictly less than UINT32_MAX.
static uint64_t calculateWeightScale(uint64_t MaxWeight) {
return MaxWeight < UINT32_MAX ? 1 : MaxWeight / UINT32_MAX + 1;
}
/// \brief Scale an individual branch weight (and add 1).
///
/// Scale a 64-bit weight down to 32-bits using \c Scale.
///
/// According to Laplace's Rule of Succession, it is better to compute the
/// weight based on the count plus 1, so universally add 1 to the value.
///
/// \pre \c Scale was calculated by \a calculateWeightScale() with a weight no
/// greater than \c Weight.
static uint32_t scaleBranchWeight(uint64_t Weight, uint64_t Scale) {
assert(Scale && "scale by 0?");
uint64_t Scaled = Weight / Scale + 1;
assert(Scaled <= UINT32_MAX && "overflow 32-bits");
return Scaled;
}
llvm::MDNode *CodeGenPGO::createProfileWeights(uint64_t TrueCount,
uint64_t FalseCount) const {
// Check for empty weights.
if (!TrueCount && !FalseCount)
return nullptr;
// Calculate how to scale down to 32-bits.
uint64_t Scale = calculateWeightScale(std::max(TrueCount, FalseCount));
llvm::MDBuilder MDHelper(gIR->context());
return MDHelper.createBranchWeights(scaleBranchWeight(TrueCount, Scale),
scaleBranchWeight(FalseCount, Scale));
}
llvm::MDNode *
CodeGenPGO::createProfileWeights(llvm::ArrayRef<uint64_t> Weights) const {
// We need at least two elements to create meaningful weights.
if (Weights.size() < 2)
return nullptr;
// Check for empty weights.
uint64_t MaxWeight = *std::max_element(Weights.begin(), Weights.end());
if (MaxWeight == 0)
return nullptr;
// Calculate how to scale down to 32-bits.
uint64_t Scale = calculateWeightScale(MaxWeight);
llvm::SmallVector<uint32_t, 16> ScaledWeights;
ScaledWeights.reserve(Weights.size());
for (uint64_t W : Weights) {
ScaledWeights.push_back(scaleBranchWeight(W, Scale));
}
llvm::MDBuilder MDHelper(gIR->context());
return MDHelper.createBranchWeights(ScaledWeights);
}
llvm::MDNode *
CodeGenPGO::createProfileWeightsWhileLoop(const RootObject *Cond,
uint64_t LoopCount) const {
if (!haveRegionCounts())
return nullptr;
auto StmtCount = getStmtCount(Cond);
assert(StmtCount.first && "missing expected while loop condition count");
auto CondCount = StmtCount.second;
if (CondCount == 0)
return nullptr;
return createProfileWeights(LoopCount,
std::max(CondCount, LoopCount) - LoopCount);
}
llvm::MDNode *
CodeGenPGO::createProfileWeightsForLoop(const ForStatement *stmt) const {
if (!haveRegionCounts())
return nullptr;
auto LoopCount = getRegionCount(stmt);
auto StmtCount =
getStmtCount(stmt->condition ? stmt->condition : getCounterPtr(stmt, 1));
assert(StmtCount.first && "missing expected for loop condition count");
auto CondCount = StmtCount.second;
if (CondCount == 0)
return nullptr;
return createProfileWeights(LoopCount,
std::max(CondCount, LoopCount) - LoopCount);
}
llvm::MDNode *
CodeGenPGO::createProfileWeightsForeach(const ForeachStatement *stmt) const {
if (!haveRegionCounts())
return nullptr;
auto LoopCount = getRegionCount(stmt);
auto StmtCount = getStmtCount(getCounterPtr(stmt, 1));
assert(StmtCount.first && "missing expected foreach loop condition count");
auto CondCount = StmtCount.second;
if (CondCount == 0)
return nullptr;
return createProfileWeights(LoopCount,
std::max(CondCount, LoopCount) - LoopCount);
}
llvm::MDNode *CodeGenPGO::createProfileWeightsForeachRange(
const ForeachRangeStatement *stmt) const {
if (!haveRegionCounts())
return nullptr;
auto LoopCount = getRegionCount(stmt);
auto StmtCount = getStmtCount(getCounterPtr(stmt, 1));
assert(StmtCount.first &&
"missing expected foreachrange loop condition count");
auto CondCount = StmtCount.second;
if (CondCount == 0)
return nullptr;
return createProfileWeights(LoopCount,
std::max(CondCount, LoopCount) - LoopCount);
}
void CodeGenPGO::emitIndirectCallPGO(llvm::Instruction *callSite,
llvm::Value *funcPtr) {
if (enablePGOIndirectCalls)
valueProfile(llvm::IPVK_IndirectCallTarget, callSite, funcPtr, true);
}
void CodeGenPGO::valueProfile(uint32_t valueKind, llvm::Instruction *valueSite,
llvm::Value *value, bool ptrCastNeeded) {
if (!value || !valueSite)
return;
bool instrumentValueSites =
opts::isInstrumentingForASTBasedPGO() && emitInstrumentation;
if (instrumentValueSites && RegionCounterMap) {
// Instrumentation must be inserted just before the valueSite instruction.
// Save the current insertion point to be able to restore it later.
auto savedInsertPoint = gIR->ir->saveIP();
gIR->ir->SetInsertPoint(valueSite);
if (ptrCastNeeded)
value = gIR->ir->CreatePtrToInt(value, gIR->ir->getInt64Ty());
auto *i8PtrTy = getVoidPtrType();
llvm::Value *Args[5] = {
llvm::ConstantExpr::getBitCast(FuncNameVar, i8PtrTy),
gIR->ir->getInt64(FunctionHash), value, gIR->ir->getInt32(valueKind),
gIR->ir->getInt32(NumValueSites[valueKind])};
gIR->ir->CreateCall(GET_INTRINSIC_DECL(instrprof_value_profile), Args);
gIR->ir->restoreIP(savedInsertPoint);
NumValueSites[valueKind]++;
return;
}
if (ProfRecord) {
if (NumValueSites[valueKind] >= ProfRecord->getNumValueSites(valueKind))
return;
llvm::annotateValueSite(gIR->module, *valueSite, *ProfRecord,
static_cast<llvm::InstrProfValueKind>(valueKind),
NumValueSites[valueKind]);
NumValueSites[valueKind]++;
}
}