llvm-project/mlir/lib/IR/AsmPrinter.cpp

Ignoring revisions in .git-blame-ignore-revs. Click here to bypass and see the normal blame view.

2704 lines
92 KiB
C++
Raw Normal View History

//===- AsmPrinter.cpp - MLIR Assembly Printer Implementation --------------===//
//
// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
// See https://llvm.org/LICENSE.txt for license information.
// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
//
//===----------------------------------------------------------------------===//
//
// This file implements the MLIR AsmPrinter class, which is used to implement
// the various print() methods on the core IR objects.
//
//===----------------------------------------------------------------------===//
#include "mlir/IR/AffineExpr.h"
#include "mlir/IR/AffineMap.h"
#include "mlir/IR/AsmState.h"
#include "mlir/IR/Attributes.h"
#include "mlir/IR/BuiltinTypes.h"
#include "mlir/IR/Dialect.h"
#include "mlir/IR/DialectImplementation.h"
#include "mlir/IR/IntegerSet.h"
#include "mlir/IR/MLIRContext.h"
#include "mlir/IR/OpImplementation.h"
#include "mlir/IR/Operation.h"
#include "llvm/ADT/APFloat.h"
#include "llvm/ADT/DenseMap.h"
#include "llvm/ADT/MapVector.h"
#include "llvm/ADT/STLExtras.h"
#include "llvm/ADT/ScopedHashTable.h"
#include "llvm/ADT/SetVector.h"
#include "llvm/ADT/SmallString.h"
#include "llvm/ADT/StringExtras.h"
#include "llvm/ADT/StringSet.h"
#include "llvm/ADT/TypeSwitch.h"
#include "llvm/Support/CommandLine.h"
#include "llvm/Support/Regex.h"
#include "llvm/Support/SaveAndRestore.h"
using namespace mlir;
using namespace mlir::detail;
void Identifier::print(raw_ostream &os) const { os << str(); }
void Identifier::dump() const { print(llvm::errs()); }
void OperationName::print(raw_ostream &os) const { os << getStringRef(); }
void OperationName::dump() const { print(llvm::errs()); }
DialectAsmPrinter::~DialectAsmPrinter() {}
//===--------------------------------------------------------------------===//
// OpAsmPrinter
//===--------------------------------------------------------------------===//
OpAsmPrinter::~OpAsmPrinter() {}
void OpAsmPrinter::printFunctionalType(Operation *op) {
auto &os = getStream();
os << '(';
llvm::interleaveComma(op->getOperands(), os, [&](Value operand) {
// Print the types of null values as <<NULL TYPE>>.
*this << (operand ? operand.getType() : Type());
});
os << ") -> ";
// Print the result list. We don't parenthesize single result types unless
// it is a function (avoiding a grammar ambiguity).
bool wrapped = op->getNumResults() != 1;
if (!wrapped && op->getResult(0).getType() &&
op->getResult(0).getType().isa<FunctionType>())
wrapped = true;
if (wrapped)
os << '(';
llvm::interleaveComma(op->getResults(), os, [&](const OpResult &result) {
// Print the types of null values as <<NULL TYPE>>.
*this << (result ? result.getType() : Type());
});
if (wrapped)
os << ')';
}
//===--------------------------------------------------------------------===//
// Operation OpAsm interface.
//===--------------------------------------------------------------------===//
/// The OpAsmOpInterface, see OpAsmInterface.td for more details.
#include "mlir/IR/OpAsmInterface.cpp.inc"
//===----------------------------------------------------------------------===//
// OpPrintingFlags
//===----------------------------------------------------------------------===//
namespace {
/// This struct contains command line options that can be used to initialize
/// various bits of the AsmPrinter. This uses a struct wrapper to avoid the need
/// for global command line options.
struct AsmPrinterOptions {
llvm::cl::opt<int64_t> printElementsAttrWithHexIfLarger{
"mlir-print-elementsattrs-with-hex-if-larger",
llvm::cl::desc(
"Print DenseElementsAttrs with a hex string that have "
"more elements than the given upper limit (use -1 to disable)")};
llvm::cl::opt<unsigned> elideElementsAttrIfLarger{
"mlir-elide-elementsattrs-if-larger",
llvm::cl::desc("Elide ElementsAttrs with \"...\" that have "
"more elements than the given upper limit")};
llvm::cl::opt<bool> printDebugInfoOpt{
"mlir-print-debuginfo", llvm::cl::init(false),
llvm::cl::desc("Print debug info in MLIR output")};
llvm::cl::opt<bool> printPrettyDebugInfoOpt{
"mlir-pretty-debuginfo", llvm::cl::init(false),
llvm::cl::desc("Print pretty debug info in MLIR output")};
// Use the generic op output form in the operation printer even if the custom
// form is defined.
llvm::cl::opt<bool> printGenericOpFormOpt{
"mlir-print-op-generic", llvm::cl::init(false),
llvm::cl::desc("Print the generic op form"), llvm::cl::Hidden};
llvm::cl::opt<bool> printLocalScopeOpt{
"mlir-print-local-scope", llvm::cl::init(false),
llvm::cl::desc("Print assuming in local scope by default"),
llvm::cl::Hidden};
};
} // end anonymous namespace
static llvm::ManagedStatic<AsmPrinterOptions> clOptions;
/// Register a set of useful command-line options that can be used to configure
/// various flags within the AsmPrinter.
void mlir::registerAsmPrinterCLOptions() {
// Make sure that the options struct has been initialized.
*clOptions;
}
/// Initialize the printing flags with default supplied by the cl::opts above.
OpPrintingFlags::OpPrintingFlags()
: printDebugInfoFlag(false), printDebugInfoPrettyFormFlag(false),
printGenericOpFormFlag(false), printLocalScope(false) {
// Initialize based upon command line options, if they are available.
if (!clOptions.isConstructed())
return;
if (clOptions->elideElementsAttrIfLarger.getNumOccurrences())
elementsAttrElementLimit = clOptions->elideElementsAttrIfLarger;
printDebugInfoFlag = clOptions->printDebugInfoOpt;
printDebugInfoPrettyFormFlag = clOptions->printPrettyDebugInfoOpt;
printGenericOpFormFlag = clOptions->printGenericOpFormOpt;
printLocalScope = clOptions->printLocalScopeOpt;
}
/// Enable the elision of large elements attributes, by printing a '...'
/// instead of the element data, when the number of elements is greater than
/// `largeElementLimit`. Note: The IR generated with this option is not
/// parsable.
OpPrintingFlags &
OpPrintingFlags::elideLargeElementsAttrs(int64_t largeElementLimit) {
elementsAttrElementLimit = largeElementLimit;
return *this;
}
/// Enable printing of debug information. If 'prettyForm' is set to true,
/// debug information is printed in a more readable 'pretty' form.
OpPrintingFlags &OpPrintingFlags::enableDebugInfo(bool prettyForm) {
printDebugInfoFlag = true;
printDebugInfoPrettyFormFlag = prettyForm;
return *this;
}
/// Always print operations in the generic form.
OpPrintingFlags &OpPrintingFlags::printGenericOpForm() {
printGenericOpFormFlag = true;
return *this;
}
/// Use local scope when printing the operation. This allows for using the
/// printer in a more localized and thread-safe setting, but may not necessarily
/// be identical of what the IR will look like when dumping the full module.
OpPrintingFlags &OpPrintingFlags::useLocalScope() {
printLocalScope = true;
return *this;
}
/// Return if the given ElementsAttr should be elided.
bool OpPrintingFlags::shouldElideElementsAttr(ElementsAttr attr) const {
return elementsAttrElementLimit.hasValue() &&
*elementsAttrElementLimit < int64_t(attr.getNumElements()) &&
!attr.isa<SplatElementsAttr>();
}
/// Return the size limit for printing large ElementsAttr.
Optional<int64_t> OpPrintingFlags::getLargeElementsAttrLimit() const {
return elementsAttrElementLimit;
}
/// Return if debug information should be printed.
bool OpPrintingFlags::shouldPrintDebugInfo() const {
return printDebugInfoFlag;
}
/// Return if debug information should be printed in the pretty form.
bool OpPrintingFlags::shouldPrintDebugInfoPrettyForm() const {
return printDebugInfoPrettyFormFlag;
}
/// Return if operations should be printed in the generic form.
bool OpPrintingFlags::shouldPrintGenericOpForm() const {
return printGenericOpFormFlag;
}
/// Return if the printer should use local scope when dumping the IR.
bool OpPrintingFlags::shouldUseLocalScope() const { return printLocalScope; }
/// Returns true if an ElementsAttr with the given number of elements should be
/// printed with hex.
static bool shouldPrintElementsAttrWithHex(int64_t numElements) {
// Check to see if a command line option was provided for the limit.
if (clOptions.isConstructed()) {
if (clOptions->printElementsAttrWithHexIfLarger.getNumOccurrences()) {
// -1 is used to disable hex printing.
if (clOptions->printElementsAttrWithHexIfLarger == -1)
return false;
return numElements > clOptions->printElementsAttrWithHexIfLarger;
}
}
// Otherwise, default to printing with hex if the number of elements is >100.
return numElements > 100;
}
//===----------------------------------------------------------------------===//
// NewLineCounter
//===----------------------------------------------------------------------===//
namespace {
/// This class is a simple formatter that emits a new line when inputted into a
/// stream, that enables counting the number of newlines emitted. This class
/// should be used whenever emitting newlines in the printer.
struct NewLineCounter {
unsigned curLine = 1;
};
} // end anonymous namespace
static raw_ostream &operator<<(raw_ostream &os, NewLineCounter &newLine) {
++newLine.curLine;
return os << '\n';
}
//===----------------------------------------------------------------------===//
[mlir][AsmPrinter] Refactor printing to only print aliases for attributes/types that will exist in the output. This revision refactors the way that attributes/types are considered when generating aliases. Instead of considering all of the attributes/types of every operation, we perform a "fake" print step that prints the operations using a dummy printer to collect the attributes and types that would actually be printed during the real process. This removes a lot of attributes/types from consideration that generally won't end up in the final output, e.g. affine map attributes in an `affine.apply`/`affine.for`. This resolves a long standing TODO w.r.t aliases, and helps to have a much cleaner textual output format. As a datapoint to the latter, as part of this change several tests were identified as testing for the presence of attributes aliases that weren't actually referenced by the custom form of any operation. To ensure that this wouldn't cause a large degradation in compile time due to the second full print, I benchmarked this change on a very large module with a lot of operations(The file is ~673M/~4.7 million lines long). This file before this change take ~6.9 seconds to print in the custom form, and ~7 seconds after this change. In the custom assembly case, this added an average of a little over ~100 miliseconds to the compile time. This increase was due to the way that argument attributes on functions are structured and how they get printed; i.e. with a better representation the negative impact here can be greatly decreased. When printing in the generic form, this revision had no observable impact on the compile time. This benchmarking leads me to believe that the impact of this change on compile time w.r.t printing is closely related to `print` methods that perform a lot of additional/complex processing outside of the OpAsmPrinter. Differential Revision: https://reviews.llvm.org/D90512
2020-11-10 13:50:31 +08:00
// AliasInitializer
//===----------------------------------------------------------------------===//
namespace {
/// This class represents a specific instance of a symbol Alias.
class SymbolAlias {
public:
SymbolAlias(StringRef name, bool isDeferrable)
: name(name), suffixIndex(0), hasSuffixIndex(false),
isDeferrable(isDeferrable) {}
SymbolAlias(StringRef name, uint32_t suffixIndex, bool isDeferrable)
: name(name), suffixIndex(suffixIndex), hasSuffixIndex(true),
isDeferrable(isDeferrable) {}
/// Print this alias to the given stream.
void print(raw_ostream &os) const {
os << name;
if (hasSuffixIndex)
os << suffixIndex;
}
/// Returns true if this alias supports deferred resolution when parsing.
bool canBeDeferred() const { return isDeferrable; }
private:
/// The main name of the alias.
StringRef name;
/// The optional suffix index of the alias, if multiple aliases had the same
/// name.
uint32_t suffixIndex : 30;
/// A flag indicating whether this alias has a suffix or not.
bool hasSuffixIndex : 1;
/// A flag indicating whether this alias may be deferred or not.
bool isDeferrable : 1;
};
[mlir][AsmPrinter] Refactor printing to only print aliases for attributes/types that will exist in the output. This revision refactors the way that attributes/types are considered when generating aliases. Instead of considering all of the attributes/types of every operation, we perform a "fake" print step that prints the operations using a dummy printer to collect the attributes and types that would actually be printed during the real process. This removes a lot of attributes/types from consideration that generally won't end up in the final output, e.g. affine map attributes in an `affine.apply`/`affine.for`. This resolves a long standing TODO w.r.t aliases, and helps to have a much cleaner textual output format. As a datapoint to the latter, as part of this change several tests were identified as testing for the presence of attributes aliases that weren't actually referenced by the custom form of any operation. To ensure that this wouldn't cause a large degradation in compile time due to the second full print, I benchmarked this change on a very large module with a lot of operations(The file is ~673M/~4.7 million lines long). This file before this change take ~6.9 seconds to print in the custom form, and ~7 seconds after this change. In the custom assembly case, this added an average of a little over ~100 miliseconds to the compile time. This increase was due to the way that argument attributes on functions are structured and how they get printed; i.e. with a better representation the negative impact here can be greatly decreased. When printing in the generic form, this revision had no observable impact on the compile time. This benchmarking leads me to believe that the impact of this change on compile time w.r.t printing is closely related to `print` methods that perform a lot of additional/complex processing outside of the OpAsmPrinter. Differential Revision: https://reviews.llvm.org/D90512
2020-11-10 13:50:31 +08:00
/// This class represents a utility that initializes the set of attribute and
/// type aliases, without the need to store the extra information within the
/// main AliasState class or pass it around via function arguments.
class AliasInitializer {
public:
[mlir][AsmPrinter] Refactor printing to only print aliases for attributes/types that will exist in the output. This revision refactors the way that attributes/types are considered when generating aliases. Instead of considering all of the attributes/types of every operation, we perform a "fake" print step that prints the operations using a dummy printer to collect the attributes and types that would actually be printed during the real process. This removes a lot of attributes/types from consideration that generally won't end up in the final output, e.g. affine map attributes in an `affine.apply`/`affine.for`. This resolves a long standing TODO w.r.t aliases, and helps to have a much cleaner textual output format. As a datapoint to the latter, as part of this change several tests were identified as testing for the presence of attributes aliases that weren't actually referenced by the custom form of any operation. To ensure that this wouldn't cause a large degradation in compile time due to the second full print, I benchmarked this change on a very large module with a lot of operations(The file is ~673M/~4.7 million lines long). This file before this change take ~6.9 seconds to print in the custom form, and ~7 seconds after this change. In the custom assembly case, this added an average of a little over ~100 miliseconds to the compile time. This increase was due to the way that argument attributes on functions are structured and how they get printed; i.e. with a better representation the negative impact here can be greatly decreased. When printing in the generic form, this revision had no observable impact on the compile time. This benchmarking leads me to believe that the impact of this change on compile time w.r.t printing is closely related to `print` methods that perform a lot of additional/complex processing outside of the OpAsmPrinter. Differential Revision: https://reviews.llvm.org/D90512
2020-11-10 13:50:31 +08:00
AliasInitializer(
DialectInterfaceCollection<OpAsmDialectInterface> &interfaces,
llvm::BumpPtrAllocator &aliasAllocator)
: interfaces(interfaces), aliasAllocator(aliasAllocator),
aliasOS(aliasBuffer) {}
void initialize(Operation *op, const OpPrintingFlags &printerFlags,
llvm::MapVector<Attribute, SymbolAlias> &attrToAlias,
llvm::MapVector<Type, SymbolAlias> &typeToAlias);
/// Visit the given attribute to see if it has an alias. `canBeDeferred` is
/// set to true if the originator of this attribute can resolve the alias
/// after parsing has completed (e.g. in the case of operation locations).
void visit(Attribute attr, bool canBeDeferred = false);
[mlir][AsmPrinter] Refactor printing to only print aliases for attributes/types that will exist in the output. This revision refactors the way that attributes/types are considered when generating aliases. Instead of considering all of the attributes/types of every operation, we perform a "fake" print step that prints the operations using a dummy printer to collect the attributes and types that would actually be printed during the real process. This removes a lot of attributes/types from consideration that generally won't end up in the final output, e.g. affine map attributes in an `affine.apply`/`affine.for`. This resolves a long standing TODO w.r.t aliases, and helps to have a much cleaner textual output format. As a datapoint to the latter, as part of this change several tests were identified as testing for the presence of attributes aliases that weren't actually referenced by the custom form of any operation. To ensure that this wouldn't cause a large degradation in compile time due to the second full print, I benchmarked this change on a very large module with a lot of operations(The file is ~673M/~4.7 million lines long). This file before this change take ~6.9 seconds to print in the custom form, and ~7 seconds after this change. In the custom assembly case, this added an average of a little over ~100 miliseconds to the compile time. This increase was due to the way that argument attributes on functions are structured and how they get printed; i.e. with a better representation the negative impact here can be greatly decreased. When printing in the generic form, this revision had no observable impact on the compile time. This benchmarking leads me to believe that the impact of this change on compile time w.r.t printing is closely related to `print` methods that perform a lot of additional/complex processing outside of the OpAsmPrinter. Differential Revision: https://reviews.llvm.org/D90512
2020-11-10 13:50:31 +08:00
/// Visit the given type to see if it has an alias.
void visit(Type type);
private:
[mlir][AsmPrinter] Refactor printing to only print aliases for attributes/types that will exist in the output. This revision refactors the way that attributes/types are considered when generating aliases. Instead of considering all of the attributes/types of every operation, we perform a "fake" print step that prints the operations using a dummy printer to collect the attributes and types that would actually be printed during the real process. This removes a lot of attributes/types from consideration that generally won't end up in the final output, e.g. affine map attributes in an `affine.apply`/`affine.for`. This resolves a long standing TODO w.r.t aliases, and helps to have a much cleaner textual output format. As a datapoint to the latter, as part of this change several tests were identified as testing for the presence of attributes aliases that weren't actually referenced by the custom form of any operation. To ensure that this wouldn't cause a large degradation in compile time due to the second full print, I benchmarked this change on a very large module with a lot of operations(The file is ~673M/~4.7 million lines long). This file before this change take ~6.9 seconds to print in the custom form, and ~7 seconds after this change. In the custom assembly case, this added an average of a little over ~100 miliseconds to the compile time. This increase was due to the way that argument attributes on functions are structured and how they get printed; i.e. with a better representation the negative impact here can be greatly decreased. When printing in the generic form, this revision had no observable impact on the compile time. This benchmarking leads me to believe that the impact of this change on compile time w.r.t printing is closely related to `print` methods that perform a lot of additional/complex processing outside of the OpAsmPrinter. Differential Revision: https://reviews.llvm.org/D90512
2020-11-10 13:50:31 +08:00
/// Try to generate an alias for the provided symbol. If an alias is
/// generated, the provided alias mapping and reverse mapping are updated.
/// Returns success if an alias was generated, failure otherwise.
[mlir][AsmPrinter] Refactor printing to only print aliases for attributes/types that will exist in the output. This revision refactors the way that attributes/types are considered when generating aliases. Instead of considering all of the attributes/types of every operation, we perform a "fake" print step that prints the operations using a dummy printer to collect the attributes and types that would actually be printed during the real process. This removes a lot of attributes/types from consideration that generally won't end up in the final output, e.g. affine map attributes in an `affine.apply`/`affine.for`. This resolves a long standing TODO w.r.t aliases, and helps to have a much cleaner textual output format. As a datapoint to the latter, as part of this change several tests were identified as testing for the presence of attributes aliases that weren't actually referenced by the custom form of any operation. To ensure that this wouldn't cause a large degradation in compile time due to the second full print, I benchmarked this change on a very large module with a lot of operations(The file is ~673M/~4.7 million lines long). This file before this change take ~6.9 seconds to print in the custom form, and ~7 seconds after this change. In the custom assembly case, this added an average of a little over ~100 miliseconds to the compile time. This increase was due to the way that argument attributes on functions are structured and how they get printed; i.e. with a better representation the negative impact here can be greatly decreased. When printing in the generic form, this revision had no observable impact on the compile time. This benchmarking leads me to believe that the impact of this change on compile time w.r.t printing is closely related to `print` methods that perform a lot of additional/complex processing outside of the OpAsmPrinter. Differential Revision: https://reviews.llvm.org/D90512
2020-11-10 13:50:31 +08:00
template <typename T>
LogicalResult
generateAlias(T symbol,
llvm::MapVector<StringRef, std::vector<T>> &aliasToSymbol);
[mlir][AsmPrinter] Refactor printing to only print aliases for attributes/types that will exist in the output. This revision refactors the way that attributes/types are considered when generating aliases. Instead of considering all of the attributes/types of every operation, we perform a "fake" print step that prints the operations using a dummy printer to collect the attributes and types that would actually be printed during the real process. This removes a lot of attributes/types from consideration that generally won't end up in the final output, e.g. affine map attributes in an `affine.apply`/`affine.for`. This resolves a long standing TODO w.r.t aliases, and helps to have a much cleaner textual output format. As a datapoint to the latter, as part of this change several tests were identified as testing for the presence of attributes aliases that weren't actually referenced by the custom form of any operation. To ensure that this wouldn't cause a large degradation in compile time due to the second full print, I benchmarked this change on a very large module with a lot of operations(The file is ~673M/~4.7 million lines long). This file before this change take ~6.9 seconds to print in the custom form, and ~7 seconds after this change. In the custom assembly case, this added an average of a little over ~100 miliseconds to the compile time. This increase was due to the way that argument attributes on functions are structured and how they get printed; i.e. with a better representation the negative impact here can be greatly decreased. When printing in the generic form, this revision had no observable impact on the compile time. This benchmarking leads me to believe that the impact of this change on compile time w.r.t printing is closely related to `print` methods that perform a lot of additional/complex processing outside of the OpAsmPrinter. Differential Revision: https://reviews.llvm.org/D90512
2020-11-10 13:50:31 +08:00
/// The set of asm interfaces within the context.
DialectInterfaceCollection<OpAsmDialectInterface> &interfaces;
/// Mapping between an alias and the set of symbols mapped to it.
llvm::MapVector<StringRef, std::vector<Attribute>> aliasToAttr;
llvm::MapVector<StringRef, std::vector<Type>> aliasToType;
/// An allocator used for alias names.
[mlir][AsmPrinter] Refactor printing to only print aliases for attributes/types that will exist in the output. This revision refactors the way that attributes/types are considered when generating aliases. Instead of considering all of the attributes/types of every operation, we perform a "fake" print step that prints the operations using a dummy printer to collect the attributes and types that would actually be printed during the real process. This removes a lot of attributes/types from consideration that generally won't end up in the final output, e.g. affine map attributes in an `affine.apply`/`affine.for`. This resolves a long standing TODO w.r.t aliases, and helps to have a much cleaner textual output format. As a datapoint to the latter, as part of this change several tests were identified as testing for the presence of attributes aliases that weren't actually referenced by the custom form of any operation. To ensure that this wouldn't cause a large degradation in compile time due to the second full print, I benchmarked this change on a very large module with a lot of operations(The file is ~673M/~4.7 million lines long). This file before this change take ~6.9 seconds to print in the custom form, and ~7 seconds after this change. In the custom assembly case, this added an average of a little over ~100 miliseconds to the compile time. This increase was due to the way that argument attributes on functions are structured and how they get printed; i.e. with a better representation the negative impact here can be greatly decreased. When printing in the generic form, this revision had no observable impact on the compile time. This benchmarking leads me to believe that the impact of this change on compile time w.r.t printing is closely related to `print` methods that perform a lot of additional/complex processing outside of the OpAsmPrinter. Differential Revision: https://reviews.llvm.org/D90512
2020-11-10 13:50:31 +08:00
llvm::BumpPtrAllocator &aliasAllocator;
/// The set of visited attributes.
DenseSet<Attribute> visitedAttributes;
/// The set of attributes that have aliases *and* can be deferred.
DenseSet<Attribute> deferrableAttributes;
[mlir][AsmPrinter] Refactor printing to only print aliases for attributes/types that will exist in the output. This revision refactors the way that attributes/types are considered when generating aliases. Instead of considering all of the attributes/types of every operation, we perform a "fake" print step that prints the operations using a dummy printer to collect the attributes and types that would actually be printed during the real process. This removes a lot of attributes/types from consideration that generally won't end up in the final output, e.g. affine map attributes in an `affine.apply`/`affine.for`. This resolves a long standing TODO w.r.t aliases, and helps to have a much cleaner textual output format. As a datapoint to the latter, as part of this change several tests were identified as testing for the presence of attributes aliases that weren't actually referenced by the custom form of any operation. To ensure that this wouldn't cause a large degradation in compile time due to the second full print, I benchmarked this change on a very large module with a lot of operations(The file is ~673M/~4.7 million lines long). This file before this change take ~6.9 seconds to print in the custom form, and ~7 seconds after this change. In the custom assembly case, this added an average of a little over ~100 miliseconds to the compile time. This increase was due to the way that argument attributes on functions are structured and how they get printed; i.e. with a better representation the negative impact here can be greatly decreased. When printing in the generic form, this revision had no observable impact on the compile time. This benchmarking leads me to believe that the impact of this change on compile time w.r.t printing is closely related to `print` methods that perform a lot of additional/complex processing outside of the OpAsmPrinter. Differential Revision: https://reviews.llvm.org/D90512
2020-11-10 13:50:31 +08:00
/// The set of visited types.
DenseSet<Type> visitedTypes;
/// Storage and stream used when generating an alias.
SmallString<32> aliasBuffer;
llvm::raw_svector_ostream aliasOS;
};
/// This class implements a dummy OpAsmPrinter that doesn't print any output,
/// and merely collects the attributes and types that *would* be printed in a
/// normal print invocation so that we can generate proper aliases. This allows
/// for us to generate aliases only for the attributes and types that would be
/// in the output, and trims down unnecessary output.
class DummyAliasOperationPrinter : private OpAsmPrinter {
public:
explicit DummyAliasOperationPrinter(const OpPrintingFlags &flags,
AliasInitializer &initializer)
: printerFlags(flags), initializer(initializer) {}
/// Print the given operation.
void print(Operation *op) {
// Visit the operation location.
if (printerFlags.shouldPrintDebugInfo())
initializer.visit(op->getLoc(), /*canBeDeferred=*/true);
[mlir][AsmPrinter] Refactor printing to only print aliases for attributes/types that will exist in the output. This revision refactors the way that attributes/types are considered when generating aliases. Instead of considering all of the attributes/types of every operation, we perform a "fake" print step that prints the operations using a dummy printer to collect the attributes and types that would actually be printed during the real process. This removes a lot of attributes/types from consideration that generally won't end up in the final output, e.g. affine map attributes in an `affine.apply`/`affine.for`. This resolves a long standing TODO w.r.t aliases, and helps to have a much cleaner textual output format. As a datapoint to the latter, as part of this change several tests were identified as testing for the presence of attributes aliases that weren't actually referenced by the custom form of any operation. To ensure that this wouldn't cause a large degradation in compile time due to the second full print, I benchmarked this change on a very large module with a lot of operations(The file is ~673M/~4.7 million lines long). This file before this change take ~6.9 seconds to print in the custom form, and ~7 seconds after this change. In the custom assembly case, this added an average of a little over ~100 miliseconds to the compile time. This increase was due to the way that argument attributes on functions are structured and how they get printed; i.e. with a better representation the negative impact here can be greatly decreased. When printing in the generic form, this revision had no observable impact on the compile time. This benchmarking leads me to believe that the impact of this change on compile time w.r.t printing is closely related to `print` methods that perform a lot of additional/complex processing outside of the OpAsmPrinter. Differential Revision: https://reviews.llvm.org/D90512
2020-11-10 13:50:31 +08:00
// If requested, always print the generic form.
if (!printerFlags.shouldPrintGenericOpForm()) {
// Check to see if this is a known operation. If so, use the registered
// custom printer hook.
if (auto *opInfo = op->getAbstractOperation()) {
opInfo->printAssembly(op, *this);
return;
}
}
// Otherwise print with the generic assembly form.
printGenericOp(op);
}
private:
/// Print the given operation in the generic form.
void printGenericOp(Operation *op) override {
// Consider nested opertions for aliases.
if (op->getNumRegions() != 0) {
for (Region &region : op->getRegions())
printRegion(region, /*printEntryBlockArgs=*/true,
/*printBlockTerminators=*/true);
}
// Visit all the types used in the operation.
for (Type type : op->getOperandTypes())
printType(type);
for (Type type : op->getResultTypes())
printType(type);
// Consider the attributes of the operation for aliases.
for (const NamedAttribute &attr : op->getAttrs())
printAttribute(attr.second);
}
/// Print the given block. If 'printBlockArgs' is false, the arguments of the
/// block are not printed. If 'printBlockTerminator' is false, the terminator
/// operation of the block is not printed.
void print(Block *block, bool printBlockArgs = true,
bool printBlockTerminator = true) {
// Consider the types of the block arguments for aliases if 'printBlockArgs'
// is set to true.
if (printBlockArgs) {
for (Type type : block->getArgumentTypes())
printType(type);
}
// Consider the operations within this block, ignoring the terminator if
// requested.
auto range = llvm::make_range(
block->begin(), std::prev(block->end(), printBlockTerminator ? 0 : 1));
for (Operation &op : range)
print(&op);
}
/// Print the given region.
void printRegion(Region &region, bool printEntryBlockArgs,
bool printBlockTerminators) override {
if (region.empty())
return;
auto *entryBlock = &region.front();
print(entryBlock, printEntryBlockArgs, printBlockTerminators);
for (Block &b : llvm::drop_begin(region, 1))
print(&b);
}
/// Consider the given type to be printed for an alias.
void printType(Type type) override { initializer.visit(type); }
/// Consider the given attribute to be printed for an alias.
void printAttribute(Attribute attr) override { initializer.visit(attr); }
void printAttributeWithoutType(Attribute attr) override {
printAttribute(attr);
}
/// Print the given set of attributes with names not included within
/// 'elidedAttrs'.
void printOptionalAttrDict(ArrayRef<NamedAttribute> attrs,
ArrayRef<StringRef> elidedAttrs = {}) override {
// Filter out any attributes that shouldn't be included.
SmallVector<NamedAttribute, 8> filteredAttrs(
llvm::make_filter_range(attrs, [&](NamedAttribute attr) {
return !llvm::is_contained(elidedAttrs, attr.first.strref());
}));
for (const NamedAttribute &attr : filteredAttrs)
printAttribute(attr.second);
}
void printOptionalAttrDictWithKeyword(
ArrayRef<NamedAttribute> attrs,
ArrayRef<StringRef> elidedAttrs = {}) override {
printOptionalAttrDict(attrs, elidedAttrs);
}
/// Return 'nulls' as the output stream, this will ignore any data fed to it.
raw_ostream &getStream() const override { return llvm::nulls(); }
/// The following are hooks of `OpAsmPrinter` that are not necessary for
/// determining potential aliases.
void printAffineMapOfSSAIds(AffineMapAttr, ValueRange) override {}
void printNewline() override {}
[mlir][AsmPrinter] Refactor printing to only print aliases for attributes/types that will exist in the output. This revision refactors the way that attributes/types are considered when generating aliases. Instead of considering all of the attributes/types of every operation, we perform a "fake" print step that prints the operations using a dummy printer to collect the attributes and types that would actually be printed during the real process. This removes a lot of attributes/types from consideration that generally won't end up in the final output, e.g. affine map attributes in an `affine.apply`/`affine.for`. This resolves a long standing TODO w.r.t aliases, and helps to have a much cleaner textual output format. As a datapoint to the latter, as part of this change several tests were identified as testing for the presence of attributes aliases that weren't actually referenced by the custom form of any operation. To ensure that this wouldn't cause a large degradation in compile time due to the second full print, I benchmarked this change on a very large module with a lot of operations(The file is ~673M/~4.7 million lines long). This file before this change take ~6.9 seconds to print in the custom form, and ~7 seconds after this change. In the custom assembly case, this added an average of a little over ~100 miliseconds to the compile time. This increase was due to the way that argument attributes on functions are structured and how they get printed; i.e. with a better representation the negative impact here can be greatly decreased. When printing in the generic form, this revision had no observable impact on the compile time. This benchmarking leads me to believe that the impact of this change on compile time w.r.t printing is closely related to `print` methods that perform a lot of additional/complex processing outside of the OpAsmPrinter. Differential Revision: https://reviews.llvm.org/D90512
2020-11-10 13:50:31 +08:00
void printOperand(Value) override {}
void printOperand(Value, raw_ostream &os) override {
// Users expect the output string to have at least the prefixed % to signal
// a value name. To maintain this invariant, emit a name even if it is
// guaranteed to go unused.
os << "%";
}
void printSymbolName(StringRef) override {}
void printSuccessor(Block *) override {}
void printSuccessorAndUseList(Block *, ValueRange) override {}
void shadowRegionArgs(Region &, ValueRange) override {}
/// The printer flags to use when determining potential aliases.
const OpPrintingFlags &printerFlags;
/// The initializer to use when identifying aliases.
AliasInitializer &initializer;
};
} // end anonymous namespace
/// Sanitize the given name such that it can be used as a valid identifier. If
/// the string needs to be modified in any way, the provided buffer is used to
/// store the new copy,
static StringRef sanitizeIdentifier(StringRef name, SmallString<16> &buffer,
StringRef allowedPunctChars = "$._-",
bool allowTrailingDigit = true) {
assert(!name.empty() && "Shouldn't have an empty name here");
auto copyNameToBuffer = [&] {
for (char ch : name) {
if (llvm::isAlnum(ch) || allowedPunctChars.contains(ch))
buffer.push_back(ch);
else if (ch == ' ')
buffer.push_back('_');
else
buffer.append(llvm::utohexstr((unsigned char)ch));
}
};
// Check to see if this name is valid. If it starts with a digit, then it
// could conflict with the autogenerated numeric ID's, so add an underscore
// prefix to avoid problems.
if (isdigit(name[0])) {
buffer.push_back('_');
copyNameToBuffer();
return buffer;
}
// If the name ends with a trailing digit, add a '_' to avoid potential
// conflicts with autogenerated ID's.
if (!allowTrailingDigit && isdigit(name.back())) {
copyNameToBuffer();
buffer.push_back('_');
return buffer;
}
// Check to see that the name consists of only valid identifier characters.
for (char ch : name) {
if (!llvm::isAlnum(ch) && !allowedPunctChars.contains(ch)) {
copyNameToBuffer();
return buffer;
}
}
// If there are no invalid characters, return the original name.
return name;
}
/// Given a collection of aliases and symbols, initialize a mapping from a
/// symbol to a given alias.
template <typename T>
static void
initializeAliases(llvm::MapVector<StringRef, std::vector<T>> &aliasToSymbol,
llvm::MapVector<T, SymbolAlias> &symbolToAlias,
DenseSet<T> *deferrableAliases = nullptr) {
std::vector<std::pair<StringRef, std::vector<T>>> aliases =
aliasToSymbol.takeVector();
llvm::array_pod_sort(aliases.begin(), aliases.end(),
[](const auto *lhs, const auto *rhs) {
return lhs->first.compare(rhs->first);
});
for (auto &it : aliases) {
// If there is only one instance for this alias, use the name directly.
if (it.second.size() == 1) {
T symbol = it.second.front();
bool isDeferrable = deferrableAliases && deferrableAliases->count(symbol);
symbolToAlias.insert({symbol, SymbolAlias(it.first, isDeferrable)});
continue;
}
// Otherwise, add the index to the name.
for (int i = 0, e = it.second.size(); i < e; ++i) {
T symbol = it.second[i];
bool isDeferrable = deferrableAliases && deferrableAliases->count(symbol);
symbolToAlias.insert({symbol, SymbolAlias(it.first, i, isDeferrable)});
}
}
}
[mlir][AsmPrinter] Refactor printing to only print aliases for attributes/types that will exist in the output. This revision refactors the way that attributes/types are considered when generating aliases. Instead of considering all of the attributes/types of every operation, we perform a "fake" print step that prints the operations using a dummy printer to collect the attributes and types that would actually be printed during the real process. This removes a lot of attributes/types from consideration that generally won't end up in the final output, e.g. affine map attributes in an `affine.apply`/`affine.for`. This resolves a long standing TODO w.r.t aliases, and helps to have a much cleaner textual output format. As a datapoint to the latter, as part of this change several tests were identified as testing for the presence of attributes aliases that weren't actually referenced by the custom form of any operation. To ensure that this wouldn't cause a large degradation in compile time due to the second full print, I benchmarked this change on a very large module with a lot of operations(The file is ~673M/~4.7 million lines long). This file before this change take ~6.9 seconds to print in the custom form, and ~7 seconds after this change. In the custom assembly case, this added an average of a little over ~100 miliseconds to the compile time. This increase was due to the way that argument attributes on functions are structured and how they get printed; i.e. with a better representation the negative impact here can be greatly decreased. When printing in the generic form, this revision had no observable impact on the compile time. This benchmarking leads me to believe that the impact of this change on compile time w.r.t printing is closely related to `print` methods that perform a lot of additional/complex processing outside of the OpAsmPrinter. Differential Revision: https://reviews.llvm.org/D90512
2020-11-10 13:50:31 +08:00
void AliasInitializer::initialize(
Operation *op, const OpPrintingFlags &printerFlags,
llvm::MapVector<Attribute, SymbolAlias> &attrToAlias,
llvm::MapVector<Type, SymbolAlias> &typeToAlias) {
[mlir][AsmPrinter] Refactor printing to only print aliases for attributes/types that will exist in the output. This revision refactors the way that attributes/types are considered when generating aliases. Instead of considering all of the attributes/types of every operation, we perform a "fake" print step that prints the operations using a dummy printer to collect the attributes and types that would actually be printed during the real process. This removes a lot of attributes/types from consideration that generally won't end up in the final output, e.g. affine map attributes in an `affine.apply`/`affine.for`. This resolves a long standing TODO w.r.t aliases, and helps to have a much cleaner textual output format. As a datapoint to the latter, as part of this change several tests were identified as testing for the presence of attributes aliases that weren't actually referenced by the custom form of any operation. To ensure that this wouldn't cause a large degradation in compile time due to the second full print, I benchmarked this change on a very large module with a lot of operations(The file is ~673M/~4.7 million lines long). This file before this change take ~6.9 seconds to print in the custom form, and ~7 seconds after this change. In the custom assembly case, this added an average of a little over ~100 miliseconds to the compile time. This increase was due to the way that argument attributes on functions are structured and how they get printed; i.e. with a better representation the negative impact here can be greatly decreased. When printing in the generic form, this revision had no observable impact on the compile time. This benchmarking leads me to believe that the impact of this change on compile time w.r.t printing is closely related to `print` methods that perform a lot of additional/complex processing outside of the OpAsmPrinter. Differential Revision: https://reviews.llvm.org/D90512
2020-11-10 13:50:31 +08:00
// Use a dummy printer when walking the IR so that we can collect the
// attributes/types that will actually be used during printing when
// considering aliases.
DummyAliasOperationPrinter aliasPrinter(printerFlags, *this);
aliasPrinter.print(op);
// Initialize the aliases sorted by name.
initializeAliases(aliasToAttr, attrToAlias, &deferrableAttributes);
initializeAliases(aliasToType, typeToAlias);
}
void AliasInitializer::visit(Attribute attr, bool canBeDeferred) {
if (!visitedAttributes.insert(attr).second) {
// If this attribute already has an alias and this instance can't be
// deferred, make sure that the alias isn't deferred.
if (!canBeDeferred)
deferrableAttributes.erase(attr);
return;
}
// Try to generate an alias for this attribute.
if (succeeded(generateAlias(attr, aliasToAttr))) {
if (canBeDeferred)
deferrableAttributes.insert(attr);
return;
}
if (auto arrayAttr = attr.dyn_cast<ArrayAttr>()) {
for (Attribute element : arrayAttr.getValue())
visit(element);
[mlir][AsmPrinter] Refactor printing to only print aliases for attributes/types that will exist in the output. This revision refactors the way that attributes/types are considered when generating aliases. Instead of considering all of the attributes/types of every operation, we perform a "fake" print step that prints the operations using a dummy printer to collect the attributes and types that would actually be printed during the real process. This removes a lot of attributes/types from consideration that generally won't end up in the final output, e.g. affine map attributes in an `affine.apply`/`affine.for`. This resolves a long standing TODO w.r.t aliases, and helps to have a much cleaner textual output format. As a datapoint to the latter, as part of this change several tests were identified as testing for the presence of attributes aliases that weren't actually referenced by the custom form of any operation. To ensure that this wouldn't cause a large degradation in compile time due to the second full print, I benchmarked this change on a very large module with a lot of operations(The file is ~673M/~4.7 million lines long). This file before this change take ~6.9 seconds to print in the custom form, and ~7 seconds after this change. In the custom assembly case, this added an average of a little over ~100 miliseconds to the compile time. This increase was due to the way that argument attributes on functions are structured and how they get printed; i.e. with a better representation the negative impact here can be greatly decreased. When printing in the generic form, this revision had no observable impact on the compile time. This benchmarking leads me to believe that the impact of this change on compile time w.r.t printing is closely related to `print` methods that perform a lot of additional/complex processing outside of the OpAsmPrinter. Differential Revision: https://reviews.llvm.org/D90512
2020-11-10 13:50:31 +08:00
} else if (auto dictAttr = attr.dyn_cast<DictionaryAttr>()) {
for (const NamedAttribute &attr : dictAttr)
visit(attr.second);
} else if (auto typeAttr = attr.dyn_cast<TypeAttr>()) {
visit(typeAttr.getValue());
}
}
[mlir][AsmPrinter] Refactor printing to only print aliases for attributes/types that will exist in the output. This revision refactors the way that attributes/types are considered when generating aliases. Instead of considering all of the attributes/types of every operation, we perform a "fake" print step that prints the operations using a dummy printer to collect the attributes and types that would actually be printed during the real process. This removes a lot of attributes/types from consideration that generally won't end up in the final output, e.g. affine map attributes in an `affine.apply`/`affine.for`. This resolves a long standing TODO w.r.t aliases, and helps to have a much cleaner textual output format. As a datapoint to the latter, as part of this change several tests were identified as testing for the presence of attributes aliases that weren't actually referenced by the custom form of any operation. To ensure that this wouldn't cause a large degradation in compile time due to the second full print, I benchmarked this change on a very large module with a lot of operations(The file is ~673M/~4.7 million lines long). This file before this change take ~6.9 seconds to print in the custom form, and ~7 seconds after this change. In the custom assembly case, this added an average of a little over ~100 miliseconds to the compile time. This increase was due to the way that argument attributes on functions are structured and how they get printed; i.e. with a better representation the negative impact here can be greatly decreased. When printing in the generic form, this revision had no observable impact on the compile time. This benchmarking leads me to believe that the impact of this change on compile time w.r.t printing is closely related to `print` methods that perform a lot of additional/complex processing outside of the OpAsmPrinter. Differential Revision: https://reviews.llvm.org/D90512
2020-11-10 13:50:31 +08:00
void AliasInitializer::visit(Type type) {
if (!visitedTypes.insert(type).second)
return;
// Try to generate an alias for this type.
if (succeeded(generateAlias(type, aliasToType)))
return;
// Visit several subtypes that contain types or atttributes.
if (auto funcType = type.dyn_cast<FunctionType>()) {
// Visit input and result types for functions.
for (auto input : funcType.getInputs())
visit(input);
for (auto result : funcType.getResults())
visit(result);
} else if (auto shapedType = type.dyn_cast<ShapedType>()) {
visit(shapedType.getElementType());
// Visit affine maps in memref type.
if (auto memref = type.dyn_cast<MemRefType>())
for (auto map : memref.getAffineMaps())
visit(AffineMapAttr::get(map));
}
}
template <typename T>
LogicalResult AliasInitializer::generateAlias(
T symbol, llvm::MapVector<StringRef, std::vector<T>> &aliasToSymbol) {
SmallString<16> tempBuffer;
for (const auto &interface : interfaces) {
interface.getAlias(symbol, aliasOS);
StringRef name = aliasOS.str();
if (name.empty())
continue;
name = sanitizeIdentifier(name, tempBuffer, /*allowedPunctChars=*/"$_-",
/*allowTrailingDigit=*/false);
name = name.copy(aliasAllocator);
aliasToSymbol[name].push_back(symbol);
aliasBuffer.clear();
return success();
}
return failure();
}
[mlir][AsmPrinter] Refactor printing to only print aliases for attributes/types that will exist in the output. This revision refactors the way that attributes/types are considered when generating aliases. Instead of considering all of the attributes/types of every operation, we perform a "fake" print step that prints the operations using a dummy printer to collect the attributes and types that would actually be printed during the real process. This removes a lot of attributes/types from consideration that generally won't end up in the final output, e.g. affine map attributes in an `affine.apply`/`affine.for`. This resolves a long standing TODO w.r.t aliases, and helps to have a much cleaner textual output format. As a datapoint to the latter, as part of this change several tests were identified as testing for the presence of attributes aliases that weren't actually referenced by the custom form of any operation. To ensure that this wouldn't cause a large degradation in compile time due to the second full print, I benchmarked this change on a very large module with a lot of operations(The file is ~673M/~4.7 million lines long). This file before this change take ~6.9 seconds to print in the custom form, and ~7 seconds after this change. In the custom assembly case, this added an average of a little over ~100 miliseconds to the compile time. This increase was due to the way that argument attributes on functions are structured and how they get printed; i.e. with a better representation the negative impact here can be greatly decreased. When printing in the generic form, this revision had no observable impact on the compile time. This benchmarking leads me to believe that the impact of this change on compile time w.r.t printing is closely related to `print` methods that perform a lot of additional/complex processing outside of the OpAsmPrinter. Differential Revision: https://reviews.llvm.org/D90512
2020-11-10 13:50:31 +08:00
//===----------------------------------------------------------------------===//
// AliasState
//===----------------------------------------------------------------------===//
namespace {
/// This class manages the state for type and attribute aliases.
class AliasState {
public:
// Initialize the internal aliases.
void
initialize(Operation *op, const OpPrintingFlags &printerFlags,
DialectInterfaceCollection<OpAsmDialectInterface> &interfaces);
/// Get an alias for the given attribute if it has one and print it in `os`.
/// Returns success if an alias was printed, failure otherwise.
LogicalResult getAlias(Attribute attr, raw_ostream &os) const;
/// Get an alias for the given type if it has one and print it in `os`.
/// Returns success if an alias was printed, failure otherwise.
LogicalResult getAlias(Type ty, raw_ostream &os) const;
/// Print all of the referenced aliases that can not be resolved in a deferred
/// manner.
void printNonDeferredAliases(raw_ostream &os, NewLineCounter &newLine) const {
printAliases(os, newLine, /*isDeferred=*/false);
}
/// Print all of the referenced aliases that support deferred resolution.
void printDeferredAliases(raw_ostream &os, NewLineCounter &newLine) const {
printAliases(os, newLine, /*isDeferred=*/true);
}
[mlir][AsmPrinter] Refactor printing to only print aliases for attributes/types that will exist in the output. This revision refactors the way that attributes/types are considered when generating aliases. Instead of considering all of the attributes/types of every operation, we perform a "fake" print step that prints the operations using a dummy printer to collect the attributes and types that would actually be printed during the real process. This removes a lot of attributes/types from consideration that generally won't end up in the final output, e.g. affine map attributes in an `affine.apply`/`affine.for`. This resolves a long standing TODO w.r.t aliases, and helps to have a much cleaner textual output format. As a datapoint to the latter, as part of this change several tests were identified as testing for the presence of attributes aliases that weren't actually referenced by the custom form of any operation. To ensure that this wouldn't cause a large degradation in compile time due to the second full print, I benchmarked this change on a very large module with a lot of operations(The file is ~673M/~4.7 million lines long). This file before this change take ~6.9 seconds to print in the custom form, and ~7 seconds after this change. In the custom assembly case, this added an average of a little over ~100 miliseconds to the compile time. This increase was due to the way that argument attributes on functions are structured and how they get printed; i.e. with a better representation the negative impact here can be greatly decreased. When printing in the generic form, this revision had no observable impact on the compile time. This benchmarking leads me to believe that the impact of this change on compile time w.r.t printing is closely related to `print` methods that perform a lot of additional/complex processing outside of the OpAsmPrinter. Differential Revision: https://reviews.llvm.org/D90512
2020-11-10 13:50:31 +08:00
private:
/// Print all of the referenced aliases that support the provided resolution
/// behavior.
void printAliases(raw_ostream &os, NewLineCounter &newLine,
bool isDeferred) const;
/// Mapping between attribute and alias.
llvm::MapVector<Attribute, SymbolAlias> attrToAlias;
/// Mapping between type and alias.
llvm::MapVector<Type, SymbolAlias> typeToAlias;
[mlir][AsmPrinter] Refactor printing to only print aliases for attributes/types that will exist in the output. This revision refactors the way that attributes/types are considered when generating aliases. Instead of considering all of the attributes/types of every operation, we perform a "fake" print step that prints the operations using a dummy printer to collect the attributes and types that would actually be printed during the real process. This removes a lot of attributes/types from consideration that generally won't end up in the final output, e.g. affine map attributes in an `affine.apply`/`affine.for`. This resolves a long standing TODO w.r.t aliases, and helps to have a much cleaner textual output format. As a datapoint to the latter, as part of this change several tests were identified as testing for the presence of attributes aliases that weren't actually referenced by the custom form of any operation. To ensure that this wouldn't cause a large degradation in compile time due to the second full print, I benchmarked this change on a very large module with a lot of operations(The file is ~673M/~4.7 million lines long). This file before this change take ~6.9 seconds to print in the custom form, and ~7 seconds after this change. In the custom assembly case, this added an average of a little over ~100 miliseconds to the compile time. This increase was due to the way that argument attributes on functions are structured and how they get printed; i.e. with a better representation the negative impact here can be greatly decreased. When printing in the generic form, this revision had no observable impact on the compile time. This benchmarking leads me to believe that the impact of this change on compile time w.r.t printing is closely related to `print` methods that perform a lot of additional/complex processing outside of the OpAsmPrinter. Differential Revision: https://reviews.llvm.org/D90512
2020-11-10 13:50:31 +08:00
/// An allocator used for alias names.
llvm::BumpPtrAllocator aliasAllocator;
};
} // end anonymous namespace
void AliasState::initialize(
[mlir][AsmPrinter] Refactor printing to only print aliases for attributes/types that will exist in the output. This revision refactors the way that attributes/types are considered when generating aliases. Instead of considering all of the attributes/types of every operation, we perform a "fake" print step that prints the operations using a dummy printer to collect the attributes and types that would actually be printed during the real process. This removes a lot of attributes/types from consideration that generally won't end up in the final output, e.g. affine map attributes in an `affine.apply`/`affine.for`. This resolves a long standing TODO w.r.t aliases, and helps to have a much cleaner textual output format. As a datapoint to the latter, as part of this change several tests were identified as testing for the presence of attributes aliases that weren't actually referenced by the custom form of any operation. To ensure that this wouldn't cause a large degradation in compile time due to the second full print, I benchmarked this change on a very large module with a lot of operations(The file is ~673M/~4.7 million lines long). This file before this change take ~6.9 seconds to print in the custom form, and ~7 seconds after this change. In the custom assembly case, this added an average of a little over ~100 miliseconds to the compile time. This increase was due to the way that argument attributes on functions are structured and how they get printed; i.e. with a better representation the negative impact here can be greatly decreased. When printing in the generic form, this revision had no observable impact on the compile time. This benchmarking leads me to believe that the impact of this change on compile time w.r.t printing is closely related to `print` methods that perform a lot of additional/complex processing outside of the OpAsmPrinter. Differential Revision: https://reviews.llvm.org/D90512
2020-11-10 13:50:31 +08:00
Operation *op, const OpPrintingFlags &printerFlags,
DialectInterfaceCollection<OpAsmDialectInterface> &interfaces) {
AliasInitializer initializer(interfaces, aliasAllocator);
[mlir][AsmPrinter] Refactor printing to only print aliases for attributes/types that will exist in the output. This revision refactors the way that attributes/types are considered when generating aliases. Instead of considering all of the attributes/types of every operation, we perform a "fake" print step that prints the operations using a dummy printer to collect the attributes and types that would actually be printed during the real process. This removes a lot of attributes/types from consideration that generally won't end up in the final output, e.g. affine map attributes in an `affine.apply`/`affine.for`. This resolves a long standing TODO w.r.t aliases, and helps to have a much cleaner textual output format. As a datapoint to the latter, as part of this change several tests were identified as testing for the presence of attributes aliases that weren't actually referenced by the custom form of any operation. To ensure that this wouldn't cause a large degradation in compile time due to the second full print, I benchmarked this change on a very large module with a lot of operations(The file is ~673M/~4.7 million lines long). This file before this change take ~6.9 seconds to print in the custom form, and ~7 seconds after this change. In the custom assembly case, this added an average of a little over ~100 miliseconds to the compile time. This increase was due to the way that argument attributes on functions are structured and how they get printed; i.e. with a better representation the negative impact here can be greatly decreased. When printing in the generic form, this revision had no observable impact on the compile time. This benchmarking leads me to believe that the impact of this change on compile time w.r.t printing is closely related to `print` methods that perform a lot of additional/complex processing outside of the OpAsmPrinter. Differential Revision: https://reviews.llvm.org/D90512
2020-11-10 13:50:31 +08:00
initializer.initialize(op, printerFlags, attrToAlias, typeToAlias);
}
LogicalResult AliasState::getAlias(Attribute attr, raw_ostream &os) const {
auto it = attrToAlias.find(attr);
if (it == attrToAlias.end())
return failure();
it->second.print(os << '#');
return success();
}
LogicalResult AliasState::getAlias(Type ty, raw_ostream &os) const {
auto it = typeToAlias.find(ty);
if (it == typeToAlias.end())
return failure();
it->second.print(os << '!');
return success();
}
void AliasState::printAliases(raw_ostream &os, NewLineCounter &newLine,
bool isDeferred) const {
auto filterFn = [=](const auto &aliasIt) {
return aliasIt.second.canBeDeferred() == isDeferred;
};
for (const auto &it : llvm::make_filter_range(attrToAlias, filterFn)) {
it.second.print(os << '#');
os << " = " << it.first << newLine;
}
for (const auto &it : llvm::make_filter_range(typeToAlias, filterFn)) {
it.second.print(os << '!');
os << " = " << it.first << newLine;
}
}
//===----------------------------------------------------------------------===//
// SSANameState
//===----------------------------------------------------------------------===//
namespace {
/// This class manages the state of SSA value names.
class SSANameState {
public:
/// A sentinel value used for values with names set.
enum : unsigned { NameSentinel = ~0U };
SSANameState(Operation *op,
DialectInterfaceCollection<OpAsmDialectInterface> &interfaces);
/// Print the SSA identifier for the given value to 'stream'. If
/// 'printResultNo' is true, it also presents the result number ('#' number)
/// of this value.
void printValueID(Value value, bool printResultNo, raw_ostream &stream) const;
/// Return the result indices for each of the result groups registered by this
/// operation, or empty if none exist.
ArrayRef<int> getOpResultGroups(Operation *op);
/// Get the ID for the given block.
unsigned getBlockID(Block *block);
/// Renumber the arguments for the specified region to the same names as the
/// SSA values in namesToUse. See OperationPrinter::shadowRegionArgs for
/// details.
void shadowRegionArgs(Region &region, ValueRange namesToUse);
private:
/// Number the SSA values within the given IR unit.
void numberValuesInRegion(
Region &region,
DialectInterfaceCollection<OpAsmDialectInterface> &interfaces);
void numberValuesInBlock(
Block &block,
DialectInterfaceCollection<OpAsmDialectInterface> &interfaces);
void numberValuesInOp(
Operation &op,
DialectInterfaceCollection<OpAsmDialectInterface> &interfaces);
/// Given a result of an operation 'result', find the result group head
/// 'lookupValue' and the result of 'result' within that group in
/// 'lookupResultNo'. 'lookupResultNo' is only filled in if the result group
/// has more than 1 result.
void getResultIDAndNumber(OpResult result, Value &lookupValue,
Optional<int> &lookupResultNo) const;
/// Set a special value name for the given value.
void setValueName(Value value, StringRef name);
/// Uniques the given value name within the printer. If the given name
/// conflicts, it is automatically renamed.
StringRef uniqueValueName(StringRef name);
/// This is the value ID for each SSA value. If this returns NameSentinel,
/// then the valueID has an entry in valueNames.
DenseMap<Value, unsigned> valueIDs;
DenseMap<Value, StringRef> valueNames;
/// This is a map of operations that contain multiple named result groups,
/// i.e. there may be multiple names for the results of the operation. The
/// value of this map are the result numbers that start a result group.
DenseMap<Operation *, SmallVector<int, 1>> opResultGroups;
/// This is the block ID for each block in the current.
DenseMap<Block *, unsigned> blockIDs;
/// This keeps track of all of the non-numeric names that are in flight,
/// allowing us to check for duplicates.
/// Note: the value of the map is unused.
llvm::ScopedHashTable<StringRef, char> usedNames;
llvm::BumpPtrAllocator usedNameAllocator;
/// This is the next value ID to assign in numbering.
unsigned nextValueID = 0;
/// This is the next ID to assign to a region entry block argument.
unsigned nextArgumentID = 0;
/// This is the next ID to assign when a name conflict is detected.
unsigned nextConflictID = 0;
};
} // end anonymous namespace
SSANameState::SSANameState(
Operation *op,
DialectInterfaceCollection<OpAsmDialectInterface> &interfaces) {
llvm::ScopedHashTable<StringRef, char>::ScopeTy usedNamesScope(usedNames);
numberValuesInOp(*op, interfaces);
for (auto &region : op->getRegions())
numberValuesInRegion(region, interfaces);
}
void SSANameState::printValueID(Value value, bool printResultNo,
raw_ostream &stream) const {
if (!value) {
stream << "<<NULL>>";
return;
}
Optional<int> resultNo;
auto lookupValue = value;
// If this is an operation result, collect the head lookup value of the result
// group and the result number of 'result' within that group.
if (OpResult result = value.dyn_cast<OpResult>())
getResultIDAndNumber(result, lookupValue, resultNo);
auto it = valueIDs.find(lookupValue);
if (it == valueIDs.end()) {
stream << "<<UNKNOWN SSA VALUE>>";
return;
}
stream << '%';
if (it->second != NameSentinel) {
stream << it->second;
} else {
auto nameIt = valueNames.find(lookupValue);
assert(nameIt != valueNames.end() && "Didn't have a name entry?");
stream << nameIt->second;
}
if (resultNo.hasValue() && printResultNo)
stream << '#' << resultNo;
}
ArrayRef<int> SSANameState::getOpResultGroups(Operation *op) {
auto it = opResultGroups.find(op);
return it == opResultGroups.end() ? ArrayRef<int>() : it->second;
}
unsigned SSANameState::getBlockID(Block *block) {
auto it = blockIDs.find(block);
return it != blockIDs.end() ? it->second : NameSentinel;
}
void SSANameState::shadowRegionArgs(Region &region, ValueRange namesToUse) {
assert(!region.empty() && "cannot shadow arguments of an empty region");
assert(region.getNumArguments() == namesToUse.size() &&
"incorrect number of names passed in");
assert(region.getParentOp()->isKnownIsolatedFromAbove() &&
"only KnownIsolatedFromAbove ops can shadow names");
SmallVector<char, 16> nameStr;
for (unsigned i = 0, e = namesToUse.size(); i != e; ++i) {
auto nameToUse = namesToUse[i];
if (nameToUse == nullptr)
continue;
auto nameToReplace = region.getArgument(i);
nameStr.clear();
llvm::raw_svector_ostream nameStream(nameStr);
printValueID(nameToUse, /*printResultNo=*/true, nameStream);
// Entry block arguments should already have a pretty "arg" name.
assert(valueIDs[nameToReplace] == NameSentinel);
// Use the name without the leading %.
auto name = StringRef(nameStream.str()).drop_front();
// Overwrite the name.
valueNames[nameToReplace] = name.copy(usedNameAllocator);
}
}
void SSANameState::numberValuesInRegion(
Region &region,
DialectInterfaceCollection<OpAsmDialectInterface> &interfaces) {
// Save the current value ids to allow for numbering values in sibling regions
// the same.
llvm::SaveAndRestore<unsigned> valueIDSaver(nextValueID);
llvm::SaveAndRestore<unsigned> argumentIDSaver(nextArgumentID);
llvm::SaveAndRestore<unsigned> conflictIDSaver(nextConflictID);
// Push a new used names scope.
llvm::ScopedHashTable<StringRef, char>::ScopeTy usedNamesScope(usedNames);
// Number the values within this region in a breadth-first order.
unsigned nextBlockID = 0;
for (auto &block : region) {
// Each block gets a unique ID, and all of the operations within it get
// numbered as well.
blockIDs[&block] = nextBlockID++;
numberValuesInBlock(block, interfaces);
}
// After that we traverse the nested regions.
// TODO: Rework this loop to not use recursion.
for (auto &block : region) {
for (auto &op : block)
for (auto &nestedRegion : op.getRegions())
numberValuesInRegion(nestedRegion, interfaces);
}
}
void SSANameState::numberValuesInBlock(
Block &block,
DialectInterfaceCollection<OpAsmDialectInterface> &interfaces) {
auto setArgNameFn = [&](Value arg, StringRef name) {
assert(!valueIDs.count(arg) && "arg numbered multiple times");
assert(arg.cast<BlockArgument>().getOwner() == &block &&
"arg not defined in 'block'");
setValueName(arg, name);
};
bool isEntryBlock = block.isEntryBlock();
if (isEntryBlock) {
if (auto *op = block.getParentOp()) {
if (auto asmInterface = interfaces.getInterfaceFor(op->getDialect()))
asmInterface->getAsmBlockArgumentNames(&block, setArgNameFn);
}
}
// Number the block arguments. We give entry block arguments a special name
// 'arg'.
SmallString<32> specialNameBuffer(isEntryBlock ? "arg" : "");
llvm::raw_svector_ostream specialName(specialNameBuffer);
for (auto arg : block.getArguments()) {
if (valueIDs.count(arg))
continue;
if (isEntryBlock) {
specialNameBuffer.resize(strlen("arg"));
specialName << nextArgumentID++;
}
setValueName(arg, specialName.str());
}
// Number the operations in this block.
for (auto &op : block)
numberValuesInOp(op, interfaces);
}
void SSANameState::numberValuesInOp(
Operation &op,
DialectInterfaceCollection<OpAsmDialectInterface> &interfaces) {
unsigned numResults = op.getNumResults();
if (numResults == 0)
return;
Value resultBegin = op.getResult(0);
// Function used to set the special result names for the operation.
SmallVector<int, 2> resultGroups(/*Size=*/1, /*Value=*/0);
auto setResultNameFn = [&](Value result, StringRef name) {
assert(!valueIDs.count(result) && "result numbered multiple times");
assert(result.getDefiningOp() == &op && "result not defined by 'op'");
setValueName(result, name);
// Record the result number for groups not anchored at 0.
if (int resultNo = result.cast<OpResult>().getResultNumber())
resultGroups.push_back(resultNo);
};
if (OpAsmOpInterface asmInterface = dyn_cast<OpAsmOpInterface>(&op))
asmInterface.getAsmResultNames(setResultNameFn);
else if (auto *asmInterface = interfaces.getInterfaceFor(op.getDialect()))
asmInterface->getAsmResultNames(&op, setResultNameFn);
// If the first result wasn't numbered, give it a default number.
if (valueIDs.try_emplace(resultBegin, nextValueID).second)
++nextValueID;
// If this operation has multiple result groups, mark it.
if (resultGroups.size() != 1) {
llvm::array_pod_sort(resultGroups.begin(), resultGroups.end());
opResultGroups.try_emplace(&op, std::move(resultGroups));
}
}
void SSANameState::getResultIDAndNumber(OpResult result, Value &lookupValue,
Optional<int> &lookupResultNo) const {
Operation *owner = result.getOwner();
if (owner->getNumResults() == 1)
return;
int resultNo = result.getResultNumber();
// If this operation has multiple result groups, we will need to find the
// one corresponding to this result.
auto resultGroupIt = opResultGroups.find(owner);
if (resultGroupIt == opResultGroups.end()) {
// If not, just use the first result.
lookupResultNo = resultNo;
lookupValue = owner->getResult(0);
return;
}
// Find the correct index using a binary search, as the groups are ordered.
ArrayRef<int> resultGroups = resultGroupIt->second;
auto it = llvm::upper_bound(resultGroups, resultNo);
int groupResultNo = 0, groupSize = 0;
// If there are no smaller elements, the last result group is the lookup.
if (it == resultGroups.end()) {
groupResultNo = resultGroups.back();
groupSize = static_cast<int>(owner->getNumResults()) - resultGroups.back();
} else {
// Otherwise, the previous element is the lookup.
groupResultNo = *std::prev(it);
groupSize = *it - groupResultNo;
}
// We only record the result number for a group of size greater than 1.
if (groupSize != 1)
lookupResultNo = resultNo - groupResultNo;
lookupValue = owner->getResult(groupResultNo);
}
void SSANameState::setValueName(Value value, StringRef name) {
// If the name is empty, the value uses the default numbering.
if (name.empty()) {
valueIDs[value] = nextValueID++;
return;
}
valueIDs[value] = NameSentinel;
valueNames[value] = uniqueValueName(name);
}
StringRef SSANameState::uniqueValueName(StringRef name) {
SmallString<16> tmpBuffer;
name = sanitizeIdentifier(name, tmpBuffer);
// Check to see if this name is already unique.
if (!usedNames.count(name)) {
name = name.copy(usedNameAllocator);
} else {
// Otherwise, we had a conflict - probe until we find a unique name. This
// is guaranteed to terminate (and usually in a single iteration) because it
// generates new names by incrementing nextConflictID.
SmallString<64> probeName(name);
probeName.push_back('_');
while (true) {
probeName += llvm::utostr(nextConflictID++);
if (!usedNames.count(probeName)) {
name = StringRef(probeName).copy(usedNameAllocator);
break;
}
probeName.resize(name.size() + 1);
}
}
usedNames.insert(name, char());
return name;
}
//===----------------------------------------------------------------------===//
// AsmState
//===----------------------------------------------------------------------===//
namespace mlir {
namespace detail {
class AsmStateImpl {
public:
explicit AsmStateImpl(Operation *op, AsmState::LocationMap *locationMap)
: interfaces(op->getContext()), nameState(op, interfaces),
locationMap(locationMap) {}
/// Initialize the alias state to enable the printing of aliases.
[mlir][AsmPrinter] Refactor printing to only print aliases for attributes/types that will exist in the output. This revision refactors the way that attributes/types are considered when generating aliases. Instead of considering all of the attributes/types of every operation, we perform a "fake" print step that prints the operations using a dummy printer to collect the attributes and types that would actually be printed during the real process. This removes a lot of attributes/types from consideration that generally won't end up in the final output, e.g. affine map attributes in an `affine.apply`/`affine.for`. This resolves a long standing TODO w.r.t aliases, and helps to have a much cleaner textual output format. As a datapoint to the latter, as part of this change several tests were identified as testing for the presence of attributes aliases that weren't actually referenced by the custom form of any operation. To ensure that this wouldn't cause a large degradation in compile time due to the second full print, I benchmarked this change on a very large module with a lot of operations(The file is ~673M/~4.7 million lines long). This file before this change take ~6.9 seconds to print in the custom form, and ~7 seconds after this change. In the custom assembly case, this added an average of a little over ~100 miliseconds to the compile time. This increase was due to the way that argument attributes on functions are structured and how they get printed; i.e. with a better representation the negative impact here can be greatly decreased. When printing in the generic form, this revision had no observable impact on the compile time. This benchmarking leads me to believe that the impact of this change on compile time w.r.t printing is closely related to `print` methods that perform a lot of additional/complex processing outside of the OpAsmPrinter. Differential Revision: https://reviews.llvm.org/D90512
2020-11-10 13:50:31 +08:00
void initializeAliases(Operation *op, const OpPrintingFlags &printerFlags) {
aliasState.initialize(op, printerFlags, interfaces);
}
/// Get an instance of the OpAsmDialectInterface for the given dialect, or
/// null if one wasn't registered.
const OpAsmDialectInterface *getOpAsmInterface(Dialect *dialect) {
return interfaces.getInterfaceFor(dialect);
}
/// Get the state used for aliases.
AliasState &getAliasState() { return aliasState; }
/// Get the state used for SSA names.
SSANameState &getSSANameState() { return nameState; }
/// Register the location, line and column, within the buffer that the given
/// operation was printed at.
void registerOperationLocation(Operation *op, unsigned line, unsigned col) {
if (locationMap)
(*locationMap)[op] = std::make_pair(line, col);
}
private:
/// Collection of OpAsm interfaces implemented in the context.
DialectInterfaceCollection<OpAsmDialectInterface> interfaces;
/// The state used for attribute and type aliases.
AliasState aliasState;
/// The state used for SSA value names.
SSANameState nameState;
/// An optional location map to be populated.
AsmState::LocationMap *locationMap;
};
} // end namespace detail
} // end namespace mlir
AsmState::AsmState(Operation *op, LocationMap *locationMap)
: impl(std::make_unique<AsmStateImpl>(op, locationMap)) {}
AsmState::~AsmState() {}
//===----------------------------------------------------------------------===//
// ModulePrinter
//===----------------------------------------------------------------------===//
namespace {
class ModulePrinter {
public:
ModulePrinter(raw_ostream &os, OpPrintingFlags flags = llvm::None,
AsmStateImpl *state = nullptr)
: os(os), printerFlags(flags), state(state) {}
explicit ModulePrinter(ModulePrinter &printer)
: os(printer.os), printerFlags(printer.printerFlags),
state(printer.state) {}
/// Returns the output stream of the printer.
raw_ostream &getStream() { return os; }
template <typename Container, typename UnaryFunctor>
inline void interleaveComma(const Container &c, UnaryFunctor each_fn) const {
llvm::interleaveComma(c, os, each_fn);
}
/// This enum describes the different kinds of elision for the type of an
/// attribute when printing it.
enum class AttrTypeElision {
/// The type must not be elided,
Never,
/// The type may be elided when it matches the default used in the parser
/// (for example i64 is the default for integer attributes).
May,
/// The type must be elided.
Must
};
/// Print the given attribute.
void printAttribute(Attribute attr,
AttrTypeElision typeElision = AttrTypeElision::Never);
void printType(Type type);
/// Print the given location to the stream. If `allowAlias` is true, this
/// allows for the internal location to use an attribute alias.
void printLocation(LocationAttr loc, bool allowAlias = false);
void printAffineMap(AffineMap map);
void
printAffineExpr(AffineExpr expr,
function_ref<void(unsigned, bool)> printValueName = nullptr);
void printAffineConstraint(AffineExpr expr, bool isEq);
void printIntegerSet(IntegerSet set);
protected:
void printOptionalAttrDict(ArrayRef<NamedAttribute> attrs,
ArrayRef<StringRef> elidedAttrs = {},
bool withKeyword = false);
void printNamedAttribute(NamedAttribute attr);
void printTrailingLocation(Location loc);
void printLocationInternal(LocationAttr loc, bool pretty = false);
/// Print a dense elements attribute. If 'allowHex' is true, a hex string is
/// used instead of individual elements when the elements attr is large.
void printDenseElementsAttr(DenseElementsAttr attr, bool allowHex);
/// Print a dense string elements attribute.
void printDenseStringElementsAttr(DenseStringElementsAttr attr);
/// Print a dense elements attribute. If 'allowHex' is true, a hex string is
/// used instead of individual elements when the elements attr is large.
void printDenseIntOrFPElementsAttr(DenseIntOrFPElementsAttr attr,
bool allowHex);
void printDialectAttribute(Attribute attr);
void printDialectType(Type type);
/// This enum is used to represent the binding strength of the enclosing
/// context that an AffineExprStorage is being printed in, so we can
/// intelligently produce parens.
enum class BindingStrength {
Weak, // + and -
Strong, // All other binary operators.
};
void printAffineExprInternal(
AffineExpr expr, BindingStrength enclosingTightness,
function_ref<void(unsigned, bool)> printValueName = nullptr);
/// The output stream for the printer.
raw_ostream &os;
/// A set of flags to control the printer's behavior.
OpPrintingFlags printerFlags;
/// An optional printer state for the module.
AsmStateImpl *state;
/// A tracker for the number of new lines emitted during printing.
NewLineCounter newLine;
};
} // end anonymous namespace
void ModulePrinter::printTrailingLocation(Location loc) {
// Check to see if we are printing debug information.
if (!printerFlags.shouldPrintDebugInfo())
return;
os << " ";
printLocation(loc, /*allowAlias=*/true);
}
void ModulePrinter::printLocationInternal(LocationAttr loc, bool pretty) {
TypeSwitch<LocationAttr>(loc)
.Case<OpaqueLoc>([&](OpaqueLoc loc) {
printLocationInternal(loc.getFallbackLocation(), pretty);
})
.Case<UnknownLoc>([&](UnknownLoc loc) {
if (pretty)
os << "[unknown]";
else
os << "unknown";
})
.Case<FileLineColLoc>([&](FileLineColLoc loc) {
StringRef mayQuote = pretty ? "" : "\"";
os << mayQuote << loc.getFilename() << mayQuote << ':' << loc.getLine()
<< ':' << loc.getColumn();
})
.Case<NameLoc>([&](NameLoc loc) {
os << '\"' << loc.getName() << '\"';
// Print the child if it isn't unknown.
auto childLoc = loc.getChildLoc();
if (!childLoc.isa<UnknownLoc>()) {
os << '(';
printLocationInternal(childLoc, pretty);
os << ')';
}
})
.Case<CallSiteLoc>([&](CallSiteLoc loc) {
Location caller = loc.getCaller();
Location callee = loc.getCallee();
if (!pretty)
os << "callsite(";
printLocationInternal(callee, pretty);
if (pretty) {
if (callee.isa<NameLoc>()) {
if (caller.isa<FileLineColLoc>()) {
os << " at ";
} else {
os << newLine << " at ";
}
} else {
os << newLine << " at ";
}
} else {
os << " at ";
}
printLocationInternal(caller, pretty);
if (!pretty)
os << ")";
})
.Case<FusedLoc>([&](FusedLoc loc) {
if (!pretty)
os << "fused";
if (Attribute metadata = loc.getMetadata())
os << '<' << metadata << '>';
os << '[';
interleave(
loc.getLocations(),
[&](Location loc) { printLocationInternal(loc, pretty); },
[&]() { os << ", "; });
os << ']';
});
}
/// Print a floating point value in a way that the parser will be able to
/// round-trip losslessly.
static void printFloatValue(const APFloat &apValue, raw_ostream &os) {
// We would like to output the FP constant value in exponential notation,
// but we cannot do this if doing so will lose precision. Check here to
// make sure that we only output it in exponential format if we can parse
// the value back and get the same value.
bool isInf = apValue.isInfinity();
bool isNaN = apValue.isNaN();
if (!isInf && !isNaN) {
SmallString<128> strValue;
apValue.toString(strValue, /*FormatPrecision=*/6, /*FormatMaxPadding=*/0,
/*TruncateZero=*/false);
// Check to make sure that the stringized number is not some string like
// "Inf" or NaN, that atof will accept, but the lexer will not. Check
// that the string matches the "[-+]?[0-9]" regex.
assert(((strValue[0] >= '0' && strValue[0] <= '9') ||
((strValue[0] == '-' || strValue[0] == '+') &&
(strValue[1] >= '0' && strValue[1] <= '9'))) &&
"[-+]?[0-9] regex does not match!");
// Parse back the stringized version and check that the value is equal
// (i.e., there is no precision loss).
if (APFloat(apValue.getSemantics(), strValue).bitwiseIsEqual(apValue)) {
os << strValue;
return;
}
// If it is not, use the default format of APFloat instead of the
// exponential notation.
strValue.clear();
apValue.toString(strValue);
// Make sure that we can parse the default form as a float.
if (StringRef(strValue).contains('.')) {
os << strValue;
return;
}
}
// Print special values in hexadecimal format. The sign bit should be included
// in the literal.
SmallVector<char, 16> str;
APInt apInt = apValue.bitcastToAPInt();
apInt.toString(str, /*Radix=*/16, /*Signed=*/false,
/*formatAsCLiteral=*/true);
os << str;
}
void ModulePrinter::printLocation(LocationAttr loc, bool allowAlias) {
if (printerFlags.shouldPrintDebugInfoPrettyForm())
return printLocationInternal(loc, /*pretty=*/true);
os << "loc(";
if (!allowAlias || !state || failed(state->getAliasState().getAlias(loc, os)))
printLocationInternal(loc);
os << ')';
}
/// Returns true if the given dialect symbol data is simple enough to print in
/// the pretty form, i.e. without the enclosing "".
static bool isDialectSymbolSimpleEnoughForPrettyForm(StringRef symName) {
// The name must start with an identifier.
if (symName.empty() || !isalpha(symName.front()))
return false;
// Ignore all the characters that are valid in an identifier in the symbol
// name.
symName = symName.drop_while(
[](char c) { return llvm::isAlnum(c) || c == '.' || c == '_'; });
if (symName.empty())
return true;
// If we got to an unexpected character, then it must be a <>. Check those
// recursively.
if (symName.front() != '<' || symName.back() != '>')
return false;
SmallVector<char, 8> nestedPunctuation;
do {
// If we ran out of characters, then we had a punctuation mismatch.
if (symName.empty())
return false;
auto c = symName.front();
symName = symName.drop_front();
switch (c) {
// We never allow null characters. This is an EOF indicator for the lexer
// which we could handle, but isn't important for any known dialect.
case '\0':
return false;
case '<':
case '[':
case '(':
case '{':
nestedPunctuation.push_back(c);
continue;
case '-':
// Treat `->` as a special token.
if (!symName.empty() && symName.front() == '>') {
symName = symName.drop_front();
continue;
}
break;
// Reject types with mismatched brackets.
case '>':
if (nestedPunctuation.pop_back_val() != '<')
return false;
break;
case ']':
if (nestedPunctuation.pop_back_val() != '[')
return false;
break;
case ')':
if (nestedPunctuation.pop_back_val() != '(')
return false;
break;
case '}':
if (nestedPunctuation.pop_back_val() != '{')
return false;
break;
default:
continue;
}
// We're done when the punctuation is fully matched.
} while (!nestedPunctuation.empty());
// If there were extra characters, then we failed.
return symName.empty();
}
/// Print the given dialect symbol to the stream.
static void printDialectSymbol(raw_ostream &os, StringRef symPrefix,
StringRef dialectName, StringRef symString) {
os << symPrefix << dialectName;
// If this symbol name is simple enough, print it directly in pretty form,
// otherwise, we print it as an escaped string.
if (isDialectSymbolSimpleEnoughForPrettyForm(symString)) {
os << '.' << symString;
return;
}
// TODO: escape the symbol name, it could contain " characters.
os << "<\"" << symString << "\">";
}
/// Returns true if the given string can be represented as a bare identifier.
static bool isBareIdentifier(StringRef name) {
assert(!name.empty() && "invalid name");
// By making this unsigned, the value passed in to isalnum will always be
// in the range 0-255. This is important when building with MSVC because
// its implementation will assert. This situation can arise when dealing
// with UTF-8 multibyte characters.
unsigned char firstChar = static_cast<unsigned char>(name[0]);
if (!isalpha(firstChar) && firstChar != '_')
return false;
return llvm::all_of(name.drop_front(), [](unsigned char c) {
return isalnum(c) || c == '_' || c == '$' || c == '.';
});
}
/// Print the given string as a symbol reference. A symbol reference is
/// represented as a string prefixed with '@'. The reference is surrounded with
/// ""'s and escaped if it has any special or non-printable characters in it.
static void printSymbolReference(StringRef symbolRef, raw_ostream &os) {
assert(!symbolRef.empty() && "expected valid symbol reference");
// If the symbol can be represented as a bare identifier, write it directly.
if (isBareIdentifier(symbolRef)) {
os << '@' << symbolRef;
return;
}
// Otherwise, output the reference wrapped in quotes with proper escaping.
os << "@\"";
printEscapedString(symbolRef, os);
os << '"';
}
// Print out a valid ElementsAttr that is succinct and can represent any
// potential shape/type, for use when eliding a large ElementsAttr.
//
// We choose to use an opaque ElementsAttr literal with conspicuous content to
// hopefully alert readers to the fact that this has been elided.
//
// Unfortunately, neither of the strings of an opaque ElementsAttr literal will
// accept the string "elided". The first string must be a registered dialect
// name and the latter must be a hex constant.
static void printElidedElementsAttr(raw_ostream &os) {
os << R"(opaque<"", "0xDEADBEEF">)";
}
void ModulePrinter::printAttribute(Attribute attr,
AttrTypeElision typeElision) {
if (!attr) {
os << "<<NULL ATTRIBUTE>>";
return;
}
// Try to print an alias for this attribute.
if (state && succeeded(state->getAliasState().getAlias(attr, os)))
return;
auto attrType = attr.getType();
if (auto opaqueAttr = attr.dyn_cast<OpaqueAttr>()) {
printDialectSymbol(os, "#", opaqueAttr.getDialectNamespace(),
opaqueAttr.getAttrData());
} else if (attr.isa<UnitAttr>()) {
os << "unit";
return;
} else if (auto dictAttr = attr.dyn_cast<DictionaryAttr>()) {
os << '{';
interleaveComma(dictAttr.getValue(),
[&](NamedAttribute attr) { printNamedAttribute(attr); });
os << '}';
} else if (auto intAttr = attr.dyn_cast<IntegerAttr>()) {
if (attrType.isSignlessInteger(1)) {
os << (intAttr.getValue().getBoolValue() ? "true" : "false");
// Boolean integer attributes always elides the type.
return;
}
// Only print attributes as unsigned if they are explicitly unsigned or are
// signless 1-bit values. Indexes, signed values, and multi-bit signless
// values print as signed.
bool isUnsigned =
attrType.isUnsignedInteger() || attrType.isSignlessInteger(1);
intAttr.getValue().print(os, !isUnsigned);
// IntegerAttr elides the type if I64.
[mlir] Add a signedness semantics bit to IntegerType Thus far IntegerType has been signless: a value of IntegerType does not have a sign intrinsically and it's up to the specific operation to decide how to interpret those bits. For example, std.addi does two's complement arithmetic, and std.divis/std.diviu treats the first bit as a sign. This design choice was made some time ago when we did't have lots of dialects and dialects were more rigid. Today we have much more extensible infrastructure and different dialect may want different modelling over integer signedness. So while we can say we want signless integers in the standard dialect, we cannot dictate for others. Requiring each dialect to model the signedness semantics with another set of custom types is duplicating the functionality everywhere, considering the fundamental role integer types play. This CL extends the IntegerType with a signedness semantics bit. This gives each dialect an option to opt in signedness semantics if that's what they want and helps code sharing. The parser is modified to recognize `si[1-9][0-9]*` and `ui[1-9][0-9]*` as signed and unsigned integer types, respectively, leaving the original `i[1-9][0-9]*` to continue to mean no indication over signedness semantics. All existing dialects are not affected (yet) as this is a feature to opt in. More discussions can be found at: https://groups.google.com/a/tensorflow.org/d/msg/mlir/XmkV8HOPWpo/7O4X0Nb_AQAJ Differential Revision: https://reviews.llvm.org/D72533
2020-01-11 03:48:24 +08:00
if (typeElision == AttrTypeElision::May && attrType.isSignlessInteger(64))
return;
} else if (auto floatAttr = attr.dyn_cast<FloatAttr>()) {
printFloatValue(floatAttr.getValue(), os);
// FloatAttr elides the type if F64.
if (typeElision == AttrTypeElision::May && attrType.isF64())
return;
} else if (auto strAttr = attr.dyn_cast<StringAttr>()) {
os << '"';
printEscapedString(strAttr.getValue(), os);
os << '"';
} else if (auto arrayAttr = attr.dyn_cast<ArrayAttr>()) {
os << '[';
interleaveComma(arrayAttr.getValue(), [&](Attribute attr) {
printAttribute(attr, AttrTypeElision::May);
});
os << ']';
} else if (auto affineMapAttr = attr.dyn_cast<AffineMapAttr>()) {
os << "affine_map<";
affineMapAttr.getValue().print(os);
os << '>';
// AffineMap always elides the type.
return;
} else if (auto integerSetAttr = attr.dyn_cast<IntegerSetAttr>()) {
os << "affine_set<";
integerSetAttr.getValue().print(os);
os << '>';
// IntegerSet always elides the type.
return;
} else if (auto typeAttr = attr.dyn_cast<TypeAttr>()) {
printType(typeAttr.getValue());
} else if (auto refAttr = attr.dyn_cast<SymbolRefAttr>()) {
printSymbolReference(refAttr.getRootReference(), os);
for (FlatSymbolRefAttr nestedRef : refAttr.getNestedReferences()) {
os << "::";
printSymbolReference(nestedRef.getValue(), os);
}
} else if (auto opaqueAttr = attr.dyn_cast<OpaqueElementsAttr>()) {
if (printerFlags.shouldElideElementsAttr(opaqueAttr)) {
printElidedElementsAttr(os);
} else {
os << "opaque<\"" << opaqueAttr.getDialect()->getNamespace() << "\", ";
os << '"' << "0x" << llvm::toHex(opaqueAttr.getValue()) << "\">";
}
} else if (auto intOrFpEltAttr = attr.dyn_cast<DenseIntOrFPElementsAttr>()) {
if (printerFlags.shouldElideElementsAttr(intOrFpEltAttr)) {
printElidedElementsAttr(os);
} else {
os << "dense<";
printDenseIntOrFPElementsAttr(intOrFpEltAttr, /*allowHex=*/true);
os << '>';
}
} else if (auto strEltAttr = attr.dyn_cast<DenseStringElementsAttr>()) {
if (printerFlags.shouldElideElementsAttr(strEltAttr)) {
printElidedElementsAttr(os);
} else {
os << "dense<";
printDenseStringElementsAttr(strEltAttr);
os << '>';
}
} else if (auto sparseEltAttr = attr.dyn_cast<SparseElementsAttr>()) {
if (printerFlags.shouldElideElementsAttr(sparseEltAttr.getIndices()) ||
printerFlags.shouldElideElementsAttr(sparseEltAttr.getValues())) {
printElidedElementsAttr(os);
} else {
os << "sparse<";
DenseIntElementsAttr indices = sparseEltAttr.getIndices();
if (indices.getNumElements() != 0) {
printDenseIntOrFPElementsAttr(indices, /*allowHex=*/false);
os << ", ";
printDenseElementsAttr(sparseEltAttr.getValues(), /*allowHex=*/true);
}
os << '>';
}
} else if (auto locAttr = attr.dyn_cast<LocationAttr>()) {
printLocation(locAttr);
} else {
return printDialectAttribute(attr);
}
// Don't print the type if we must elide it, or if it is a None type.
if (typeElision != AttrTypeElision::Must && !attrType.isa<NoneType>()) {
os << " : ";
printType(attrType);
}
}
/// Print the integer element of a DenseElementsAttr.
static void printDenseIntElement(const APInt &value, raw_ostream &os,
bool isSigned) {
if (value.getBitWidth() == 1)
os << (value.getBoolValue() ? "true" : "false");
else
value.print(os, isSigned);
}
static void
printDenseElementsAttrImpl(bool isSplat, ShapedType type, raw_ostream &os,
function_ref<void(unsigned)> printEltFn) {
// Special case for 0-d and splat tensors.
if (isSplat)
return printEltFn(0);
// Special case for degenerate tensors.
auto numElements = type.getNumElements();
if (numElements == 0)
return;
// We use a mixed-radix counter to iterate through the shape. When we bump a
// non-least-significant digit, we emit a close bracket. When we next emit an
// element we re-open all closed brackets.
// The mixed-radix counter, with radices in 'shape'.
int64_t rank = type.getRank();
SmallVector<unsigned, 4> counter(rank, 0);
// The number of brackets that have been opened and not closed.
unsigned openBrackets = 0;
auto shape = type.getShape();
auto bumpCounter = [&] {
// Bump the least significant digit.
++counter[rank - 1];
// Iterate backwards bubbling back the increment.
for (unsigned i = rank - 1; i > 0; --i)
if (counter[i] >= shape[i]) {
// Index 'i' is rolled over. Bump (i-1) and close a bracket.
counter[i] = 0;
++counter[i - 1];
--openBrackets;
os << ']';
}
};
for (unsigned idx = 0, e = numElements; idx != e; ++idx) {
if (idx != 0)
os << ", ";
while (openBrackets++ < rank)
os << '[';
openBrackets = rank;
printEltFn(idx);
bumpCounter();
}
while (openBrackets-- > 0)
os << ']';
}
void ModulePrinter::printDenseElementsAttr(DenseElementsAttr attr,
bool allowHex) {
if (auto stringAttr = attr.dyn_cast<DenseStringElementsAttr>())
return printDenseStringElementsAttr(stringAttr);
printDenseIntOrFPElementsAttr(attr.cast<DenseIntOrFPElementsAttr>(),
allowHex);
}
void ModulePrinter::printDenseIntOrFPElementsAttr(DenseIntOrFPElementsAttr attr,
bool allowHex) {
auto type = attr.getType();
auto elementType = type.getElementType();
// Check to see if we should format this attribute as a hex string.
auto numElements = type.getNumElements();
if (!attr.isSplat() && allowHex &&
shouldPrintElementsAttrWithHex(numElements)) {
ArrayRef<char> rawData = attr.getRawData();
if (llvm::support::endian::system_endianness() ==
llvm::support::endianness::big) {
// Convert endianess in big-endian(BE) machines. `rawData` is BE in BE
// machines. It is converted here to print in LE format.
SmallVector<char, 64> outDataVec(rawData.size());
MutableArrayRef<char> convRawData(outDataVec);
DenseIntOrFPElementsAttr::convertEndianOfArrayRefForBEmachine(
rawData, convRawData, type);
os << '"' << "0x"
<< llvm::toHex(StringRef(convRawData.data(), convRawData.size()))
<< "\"";
} else {
os << '"' << "0x"
<< llvm::toHex(StringRef(rawData.data(), rawData.size())) << "\"";
}
return;
}
if (ComplexType complexTy = elementType.dyn_cast<ComplexType>()) {
Type complexElementType = complexTy.getElementType();
// Note: The if and else below had a common lambda function which invoked
// printDenseElementsAttrImpl. This lambda was hitting a bug in gcc 9.1,9.2
// and hence was replaced.
if (complexElementType.isa<IntegerType>()) {
bool isSigned = !complexElementType.isUnsignedInteger();
printDenseElementsAttrImpl(attr.isSplat(), type, os, [&](unsigned index) {
auto complexValue = *(attr.getComplexIntValues().begin() + index);
os << "(";
printDenseIntElement(complexValue.real(), os, isSigned);
os << ",";
printDenseIntElement(complexValue.imag(), os, isSigned);
os << ")";
});
} else {
printDenseElementsAttrImpl(attr.isSplat(), type, os, [&](unsigned index) {
auto complexValue = *(attr.getComplexFloatValues().begin() + index);
os << "(";
printFloatValue(complexValue.real(), os);
os << ",";
printFloatValue(complexValue.imag(), os);
os << ")";
});
}
} else if (elementType.isIntOrIndex()) {
bool isSigned = !elementType.isUnsignedInteger();
auto intValues = attr.getIntValues();
printDenseElementsAttrImpl(attr.isSplat(), type, os, [&](unsigned index) {
printDenseIntElement(*(intValues.begin() + index), os, isSigned);
});
} else {
assert(elementType.isa<FloatType>() && "unexpected element type");
auto floatValues = attr.getFloatValues();
printDenseElementsAttrImpl(attr.isSplat(), type, os, [&](unsigned index) {
printFloatValue(*(floatValues.begin() + index), os);
});
}
}
void ModulePrinter::printDenseStringElementsAttr(DenseStringElementsAttr attr) {
ArrayRef<StringRef> data = attr.getRawStringData();
auto printFn = [&](unsigned index) {
os << "\"";
printEscapedString(data[index], os);
os << "\"";
};
printDenseElementsAttrImpl(attr.isSplat(), attr.getType(), os, printFn);
}
void ModulePrinter::printType(Type type) {
if (!type) {
os << "<<NULL TYPE>>";
return;
}
// Try to print an alias for this type.
if (state && succeeded(state->getAliasState().getAlias(type, os)))
return;
TypeSwitch<Type>(type)
.Case<OpaqueType>([&](OpaqueType opaqueTy) {
printDialectSymbol(os, "!", opaqueTy.getDialectNamespace(),
opaqueTy.getTypeData());
})
.Case<IndexType>([&](Type) { os << "index"; })
.Case<BFloat16Type>([&](Type) { os << "bf16"; })
.Case<Float16Type>([&](Type) { os << "f16"; })
.Case<Float32Type>([&](Type) { os << "f32"; })
.Case<Float64Type>([&](Type) { os << "f64"; })
.Case<IntegerType>([&](IntegerType integerTy) {
if (integerTy.isSigned())
os << 's';
else if (integerTy.isUnsigned())
os << 'u';
os << 'i' << integerTy.getWidth();
})
.Case<FunctionType>([&](FunctionType funcTy) {
os << '(';
interleaveComma(funcTy.getInputs(), [&](Type ty) { printType(ty); });
os << ") -> ";
ArrayRef<Type> results = funcTy.getResults();
if (results.size() == 1 && !results[0].isa<FunctionType>()) {
os << results[0];
} else {
os << '(';
interleaveComma(results, [&](Type ty) { printType(ty); });
os << ')';
}
})
.Case<VectorType>([&](VectorType vectorTy) {
os << "vector<";
for (int64_t dim : vectorTy.getShape())
os << dim << 'x';
os << vectorTy.getElementType() << '>';
})
.Case<RankedTensorType>([&](RankedTensorType tensorTy) {
os << "tensor<";
for (int64_t dim : tensorTy.getShape()) {
if (ShapedType::isDynamic(dim))
os << '?';
else
os << dim;
os << 'x';
}
os << tensorTy.getElementType() << '>';
})
.Case<UnrankedTensorType>([&](UnrankedTensorType tensorTy) {
os << "tensor<*x";
printType(tensorTy.getElementType());
os << '>';
})
.Case<MemRefType>([&](MemRefType memrefTy) {
os << "memref<";
for (int64_t dim : memrefTy.getShape()) {
if (ShapedType::isDynamic(dim))
os << '?';
else
os << dim;
os << 'x';
}
printType(memrefTy.getElementType());
for (auto map : memrefTy.getAffineMaps()) {
os << ", ";
printAttribute(AffineMapAttr::get(map));
}
// Only print the memory space if it is the non-default one.
if (memrefTy.getMemorySpace())
os << ", " << memrefTy.getMemorySpace();
os << '>';
})
.Case<UnrankedMemRefType>([&](UnrankedMemRefType memrefTy) {
os << "memref<*x";
printType(memrefTy.getElementType());
// Only print the memory space if it is the non-default one.
if (memrefTy.getMemorySpace())
os << ", " << memrefTy.getMemorySpace();
os << '>';
})
.Case<ComplexType>([&](ComplexType complexTy) {
os << "complex<";
printType(complexTy.getElementType());
os << '>';
})
.Case<TupleType>([&](TupleType tupleTy) {
os << "tuple<";
interleaveComma(tupleTy.getTypes(),
[&](Type type) { printType(type); });
os << '>';
})
.Case<NoneType>([&](Type) { os << "none"; })
.Default([&](Type type) { return printDialectType(type); });
}
void ModulePrinter::printOptionalAttrDict(ArrayRef<NamedAttribute> attrs,
ArrayRef<StringRef> elidedAttrs,
bool withKeyword) {
// If there are no attributes, then there is nothing to be done.
if (attrs.empty())
return;
// Filter out any attributes that shouldn't be included.
SmallVector<NamedAttribute, 8> filteredAttrs(
llvm::make_filter_range(attrs, [&](NamedAttribute attr) {
return !llvm::is_contained(elidedAttrs, attr.first.strref());
}));
// If there are no attributes left to print after filtering, then we're done.
if (filteredAttrs.empty())
return;
// Print the 'attributes' keyword if necessary.
if (withKeyword)
os << " attributes";
// Otherwise, print them all out in braces.
os << " {";
interleaveComma(filteredAttrs,
[&](NamedAttribute attr) { printNamedAttribute(attr); });
os << '}';
}
void ModulePrinter::printNamedAttribute(NamedAttribute attr) {
if (isBareIdentifier(attr.first)) {
os << attr.first;
} else {
os << '"';
printEscapedString(attr.first.strref(), os);
os << '"';
}
// Pretty printing elides the attribute value for unit attributes.
if (attr.second.isa<UnitAttr>())
return;
os << " = ";
printAttribute(attr.second);
}
//===----------------------------------------------------------------------===//
// CustomDialectAsmPrinter
//===----------------------------------------------------------------------===//
namespace {
/// This class provides the main specialization of the DialectAsmPrinter that is
/// used to provide support for print attributes and types. This hooks allows
/// for dialects to hook into the main ModulePrinter.
struct CustomDialectAsmPrinter : public DialectAsmPrinter {
public:
CustomDialectAsmPrinter(ModulePrinter &printer) : printer(printer) {}
~CustomDialectAsmPrinter() override {}
raw_ostream &getStream() const override { return printer.getStream(); }
/// Print the given attribute to the stream.
void printAttribute(Attribute attr) override { printer.printAttribute(attr); }
/// Print the given floating point value in a stablized form.
void printFloat(const APFloat &value) override {
printFloatValue(value, getStream());
}
/// Print the given type to the stream.
void printType(Type type) override { printer.printType(type); }
/// The main module printer.
ModulePrinter &printer;
};
} // end anonymous namespace
void ModulePrinter::printDialectAttribute(Attribute attr) {
auto &dialect = attr.getDialect();
// Ask the dialect to serialize the attribute to a string.
std::string attrName;
{
llvm::raw_string_ostream attrNameStr(attrName);
ModulePrinter subPrinter(attrNameStr, printerFlags, state);
CustomDialectAsmPrinter printer(subPrinter);
dialect.printAttribute(attr, printer);
}
printDialectSymbol(os, "#", dialect.getNamespace(), attrName);
}
void ModulePrinter::printDialectType(Type type) {
auto &dialect = type.getDialect();
// Ask the dialect to serialize the type to a string.
std::string typeName;
{
llvm::raw_string_ostream typeNameStr(typeName);
ModulePrinter subPrinter(typeNameStr, printerFlags, state);
CustomDialectAsmPrinter printer(subPrinter);
dialect.printType(type, printer);
}
printDialectSymbol(os, "!", dialect.getNamespace(), typeName);
}
//===----------------------------------------------------------------------===//
// Affine expressions and maps
//===----------------------------------------------------------------------===//
void ModulePrinter::printAffineExpr(
AffineExpr expr, function_ref<void(unsigned, bool)> printValueName) {
printAffineExprInternal(expr, BindingStrength::Weak, printValueName);
}
void ModulePrinter::printAffineExprInternal(
AffineExpr expr, BindingStrength enclosingTightness,
function_ref<void(unsigned, bool)> printValueName) {
const char *binopSpelling = nullptr;
switch (expr.getKind()) {
case AffineExprKind::SymbolId: {
unsigned pos = expr.cast<AffineSymbolExpr>().getPosition();
if (printValueName)
printValueName(pos, /*isSymbol=*/true);
else
os << 's' << pos;
return;
}
case AffineExprKind::DimId: {
unsigned pos = expr.cast<AffineDimExpr>().getPosition();
if (printValueName)
printValueName(pos, /*isSymbol=*/false);
else
os << 'd' << pos;
return;
}
case AffineExprKind::Constant:
os << expr.cast<AffineConstantExpr>().getValue();
return;
case AffineExprKind::Add:
binopSpelling = " + ";
break;
case AffineExprKind::Mul:
binopSpelling = " * ";
break;
case AffineExprKind::FloorDiv:
binopSpelling = " floordiv ";
break;
case AffineExprKind::CeilDiv:
binopSpelling = " ceildiv ";
break;
case AffineExprKind::Mod:
binopSpelling = " mod ";
break;
}
auto binOp = expr.cast<AffineBinaryOpExpr>();
AffineExpr lhsExpr = binOp.getLHS();
AffineExpr rhsExpr = binOp.getRHS();
// Handle tightly binding binary operators.
if (binOp.getKind() != AffineExprKind::Add) {
if (enclosingTightness == BindingStrength::Strong)
os << '(';
// Pretty print multiplication with -1.
auto rhsConst = rhsExpr.dyn_cast<AffineConstantExpr>();
if (rhsConst && binOp.getKind() == AffineExprKind::Mul &&
rhsConst.getValue() == -1) {
os << "-";
printAffineExprInternal(lhsExpr, BindingStrength::Strong, printValueName);
if (enclosingTightness == BindingStrength::Strong)
os << ')';
return;
}
printAffineExprInternal(lhsExpr, BindingStrength::Strong, printValueName);
os << binopSpelling;
printAffineExprInternal(rhsExpr, BindingStrength::Strong, printValueName);
if (enclosingTightness == BindingStrength::Strong)
os << ')';
return;
}
// Print out special "pretty" forms for add.
if (enclosingTightness == BindingStrength::Strong)
os << '(';
// Pretty print addition to a product that has a negative operand as a
// subtraction.
if (auto rhs = rhsExpr.dyn_cast<AffineBinaryOpExpr>()) {
if (rhs.getKind() == AffineExprKind::Mul) {
AffineExpr rrhsExpr = rhs.getRHS();
if (auto rrhs = rrhsExpr.dyn_cast<AffineConstantExpr>()) {
if (rrhs.getValue() == -1) {
printAffineExprInternal(lhsExpr, BindingStrength::Weak,
printValueName);
os << " - ";
if (rhs.getLHS().getKind() == AffineExprKind::Add) {
printAffineExprInternal(rhs.getLHS(), BindingStrength::Strong,
printValueName);
} else {
printAffineExprInternal(rhs.getLHS(), BindingStrength::Weak,
printValueName);
}
if (enclosingTightness == BindingStrength::Strong)
os << ')';
return;
}
if (rrhs.getValue() < -1) {
printAffineExprInternal(lhsExpr, BindingStrength::Weak,
printValueName);
os << " - ";
printAffineExprInternal(rhs.getLHS(), BindingStrength::Strong,
printValueName);
os << " * " << -rrhs.getValue();
if (enclosingTightness == BindingStrength::Strong)
os << ')';
return;
}
}
}
}
// Pretty print addition to a negative number as a subtraction.
if (auto rhsConst = rhsExpr.dyn_cast<AffineConstantExpr>()) {
if (rhsConst.getValue() < 0) {
printAffineExprInternal(lhsExpr, BindingStrength::Weak, printValueName);
os << " - " << -rhsConst.getValue();
if (enclosingTightness == BindingStrength::Strong)
os << ')';
return;
}
}
printAffineExprInternal(lhsExpr, BindingStrength::Weak, printValueName);
os << " + ";
printAffineExprInternal(rhsExpr, BindingStrength::Weak, printValueName);
if (enclosingTightness == BindingStrength::Strong)
os << ')';
}
void ModulePrinter::printAffineConstraint(AffineExpr expr, bool isEq) {
printAffineExprInternal(expr, BindingStrength::Weak);
isEq ? os << " == 0" : os << " >= 0";
}
void ModulePrinter::printAffineMap(AffineMap map) {
// Dimension identifiers.
os << '(';
for (int i = 0; i < (int)map.getNumDims() - 1; ++i)
os << 'd' << i << ", ";
if (map.getNumDims() >= 1)
os << 'd' << map.getNumDims() - 1;
os << ')';
// Symbolic identifiers.
if (map.getNumSymbols() != 0) {
os << '[';
for (unsigned i = 0; i < map.getNumSymbols() - 1; ++i)
os << 's' << i << ", ";
if (map.getNumSymbols() >= 1)
os << 's' << map.getNumSymbols() - 1;
os << ']';
}
// Result affine expressions.
os << " -> (";
interleaveComma(map.getResults(),
[&](AffineExpr expr) { printAffineExpr(expr); });
os << ')';
}
void ModulePrinter::printIntegerSet(IntegerSet set) {
// Dimension identifiers.
os << '(';
for (unsigned i = 1; i < set.getNumDims(); ++i)
os << 'd' << i - 1 << ", ";
if (set.getNumDims() >= 1)
os << 'd' << set.getNumDims() - 1;
os << ')';
// Symbolic identifiers.
if (set.getNumSymbols() != 0) {
os << '[';
for (unsigned i = 0; i < set.getNumSymbols() - 1; ++i)
os << 's' << i << ", ";
if (set.getNumSymbols() >= 1)
os << 's' << set.getNumSymbols() - 1;
os << ']';
}
// Print constraints.
os << " : (";
int numConstraints = set.getNumConstraints();
for (int i = 1; i < numConstraints; ++i) {
printAffineConstraint(set.getConstraint(i - 1), set.isEq(i - 1));
os << ", ";
}
if (numConstraints >= 1)
printAffineConstraint(set.getConstraint(numConstraints - 1),
set.isEq(numConstraints - 1));
os << ')';
}
//===----------------------------------------------------------------------===//
// OperationPrinter
//===----------------------------------------------------------------------===//
namespace {
/// This class contains the logic for printing operations, regions, and blocks.
class OperationPrinter : public ModulePrinter, private OpAsmPrinter {
public:
explicit OperationPrinter(raw_ostream &os, OpPrintingFlags flags,
AsmStateImpl &state)
: ModulePrinter(os, flags, &state) {}
/// Print the given top-level operation.
void printTopLevelOperation(Operation *op);
/// Print the given operation with its indent and location.
void print(Operation *op);
/// Print the bare location, not including indentation/location/etc.
void printOperation(Operation *op);
/// Print the given operation in the generic form.
void printGenericOp(Operation *op) override;
/// Print the name of the given block.
void printBlockName(Block *block);
/// Print the given block. If 'printBlockArgs' is false, the arguments of the
/// block are not printed. If 'printBlockTerminator' is false, the terminator
/// operation of the block is not printed.
void print(Block *block, bool printBlockArgs = true,
bool printBlockTerminator = true);
/// Print the ID of the given value, optionally with its result number.
void printValueID(Value value, bool printResultNo = true,
raw_ostream *streamOverride = nullptr) const;
//===--------------------------------------------------------------------===//
// OpAsmPrinter methods
//===--------------------------------------------------------------------===//
/// Return the current stream of the printer.
raw_ostream &getStream() const override { return os; }
/// Print a newline and indent the printer to the start of the current
/// operation.
void printNewline() override {
os << newLine;
os.indent(currentIndent);
}
/// Print the given type.
void printType(Type type) override { ModulePrinter::printType(type); }
/// Print the given attribute.
void printAttribute(Attribute attr) override {
ModulePrinter::printAttribute(attr);
}
/// Print the given attribute without its type. The corresponding parser must
/// provide a valid type for the attribute.
void printAttributeWithoutType(Attribute attr) override {
ModulePrinter::printAttribute(attr, AttrTypeElision::Must);
}
/// Print the ID for the given value.
void printOperand(Value value) override { printValueID(value); }
void printOperand(Value value, raw_ostream &os) override {
printValueID(value, /*printResultNo=*/true, &os);
}
/// Print an optional attribute dictionary with a given set of elided values.
void printOptionalAttrDict(ArrayRef<NamedAttribute> attrs,
ArrayRef<StringRef> elidedAttrs = {}) override {
ModulePrinter::printOptionalAttrDict(attrs, elidedAttrs);
}
void printOptionalAttrDictWithKeyword(
ArrayRef<NamedAttribute> attrs,
ArrayRef<StringRef> elidedAttrs = {}) override {
ModulePrinter::printOptionalAttrDict(attrs, elidedAttrs,
/*withKeyword=*/true);
}
/// Print the given successor.
void printSuccessor(Block *successor) override;
/// Print an operation successor with the operands used for the block
/// arguments.
void printSuccessorAndUseList(Block *successor,
ValueRange succOperands) override;
/// Print the given region.
void printRegion(Region &region, bool printEntryBlockArgs,
bool printBlockTerminators) override;
/// Renumber the arguments for the specified region to the same names as the
/// SSA values in namesToUse. This may only be used for IsolatedFromAbove
/// operations. If any entry in namesToUse is null, the corresponding
/// argument name is left alone.
void shadowRegionArgs(Region &region, ValueRange namesToUse) override {
state->getSSANameState().shadowRegionArgs(region, namesToUse);
}
/// Print the given affine map with the symbol and dimension operands printed
/// inline with the map.
void printAffineMapOfSSAIds(AffineMapAttr mapAttr,
ValueRange operands) override;
/// Print the given string as a symbol reference.
void printSymbolName(StringRef symbolRef) override {
::printSymbolReference(symbolRef, os);
}
private:
/// The number of spaces used for indenting nested operations.
const static unsigned indentWidth = 2;
// This is the current indentation level for nested structures.
unsigned currentIndent = 0;
};
} // end anonymous namespace
void OperationPrinter::printTopLevelOperation(Operation *op) {
// Output the aliases at the top level that can't be deferred.
state->getAliasState().printNonDeferredAliases(os, newLine);
// Print the module.
print(op);
os << newLine;
// Output the aliases at the top level that can be deferred.
state->getAliasState().printDeferredAliases(os, newLine);
}
void OperationPrinter::print(Operation *op) {
// Track the location of this operation.
state->registerOperationLocation(op, newLine.curLine, currentIndent);
os.indent(currentIndent);
printOperation(op);
printTrailingLocation(op->getLoc());
}
void OperationPrinter::printOperation(Operation *op) {
if (size_t numResults = op->getNumResults()) {
auto printResultGroup = [&](size_t resultNo, size_t resultCount) {
printValueID(op->getResult(resultNo), /*printResultNo=*/false);
if (resultCount > 1)
os << ':' << resultCount;
};
// Check to see if this operation has multiple result groups.
ArrayRef<int> resultGroups = state->getSSANameState().getOpResultGroups(op);
if (!resultGroups.empty()) {
// Interleave the groups excluding the last one, this one will be handled
// separately.
interleaveComma(llvm::seq<int>(0, resultGroups.size() - 1), [&](int i) {
printResultGroup(resultGroups[i],
resultGroups[i + 1] - resultGroups[i]);
});
os << ", ";
printResultGroup(resultGroups.back(), numResults - resultGroups.back());
} else {
printResultGroup(/*resultNo=*/0, /*resultCount=*/numResults);
}
os << " = ";
}
// If requested, always print the generic form.
if (!printerFlags.shouldPrintGenericOpForm()) {
// Check to see if this is a known operation. If so, use the registered
// custom printer hook.
if (auto *opInfo = op->getAbstractOperation()) {
opInfo->printAssembly(op, *this);
return;
}
}
// Otherwise print with the generic assembly form.
printGenericOp(op);
}
void OperationPrinter::printGenericOp(Operation *op) {
os << '"';
printEscapedString(op->getName().getStringRef(), os);
os << "\"(";
interleaveComma(op->getOperands(), [&](Value value) { printValueID(value); });
os << ')';
// For terminators, print the list of successors and their operands.
if (op->getNumSuccessors() != 0) {
os << '[';
interleaveComma(op->getSuccessors(),
[&](Block *successor) { printBlockName(successor); });
os << ']';
}
// Print regions.
if (op->getNumRegions() != 0) {
os << " (";
interleaveComma(op->getRegions(), [&](Region &region) {
printRegion(region, /*printEntryBlockArgs=*/true,
/*printBlockTerminators=*/true);
});
os << ')';
}
auto attrs = op->getAttrs();
printOptionalAttrDict(attrs);
// Print the type signature of the operation.
os << " : ";
printFunctionalType(op);
}
void OperationPrinter::printBlockName(Block *block) {
auto id = state->getSSANameState().getBlockID(block);
if (id != SSANameState::NameSentinel)
os << "^bb" << id;
else
os << "^INVALIDBLOCK";
}
void OperationPrinter::print(Block *block, bool printBlockArgs,
bool printBlockTerminator) {
// Print the block label and argument list if requested.
if (printBlockArgs) {
os.indent(currentIndent);
printBlockName(block);
// Print the argument list if non-empty.
if (!block->args_empty()) {
os << '(';
interleaveComma(block->getArguments(), [&](BlockArgument arg) {
printValueID(arg);
os << ": ";
printType(arg.getType());
});
os << ')';
}
os << ':';
// Print out some context information about the predecessors of this block.
if (!block->getParent()) {
os << " // block is not in a region!";
} else if (block->hasNoPredecessors()) {
os << " // no predecessors";
} else if (auto *pred = block->getSinglePredecessor()) {
os << " // pred: ";
printBlockName(pred);
} else {
// We want to print the predecessors in increasing numeric order, not in
// whatever order the use-list is in, so gather and sort them.
SmallVector<std::pair<unsigned, Block *>, 4> predIDs;
for (auto *pred : block->getPredecessors())
predIDs.push_back({state->getSSANameState().getBlockID(pred), pred});
llvm::array_pod_sort(predIDs.begin(), predIDs.end());
os << " // " << predIDs.size() << " preds: ";
interleaveComma(predIDs, [&](std::pair<unsigned, Block *> pred) {
printBlockName(pred.second);
});
}
os << newLine;
}
currentIndent += indentWidth;
auto range = llvm::make_range(
[mlir][AsmPrinter] Refactor printing to only print aliases for attributes/types that will exist in the output. This revision refactors the way that attributes/types are considered when generating aliases. Instead of considering all of the attributes/types of every operation, we perform a "fake" print step that prints the operations using a dummy printer to collect the attributes and types that would actually be printed during the real process. This removes a lot of attributes/types from consideration that generally won't end up in the final output, e.g. affine map attributes in an `affine.apply`/`affine.for`. This resolves a long standing TODO w.r.t aliases, and helps to have a much cleaner textual output format. As a datapoint to the latter, as part of this change several tests were identified as testing for the presence of attributes aliases that weren't actually referenced by the custom form of any operation. To ensure that this wouldn't cause a large degradation in compile time due to the second full print, I benchmarked this change on a very large module with a lot of operations(The file is ~673M/~4.7 million lines long). This file before this change take ~6.9 seconds to print in the custom form, and ~7 seconds after this change. In the custom assembly case, this added an average of a little over ~100 miliseconds to the compile time. This increase was due to the way that argument attributes on functions are structured and how they get printed; i.e. with a better representation the negative impact here can be greatly decreased. When printing in the generic form, this revision had no observable impact on the compile time. This benchmarking leads me to believe that the impact of this change on compile time w.r.t printing is closely related to `print` methods that perform a lot of additional/complex processing outside of the OpAsmPrinter. Differential Revision: https://reviews.llvm.org/D90512
2020-11-10 13:50:31 +08:00
block->begin(), std::prev(block->end(), printBlockTerminator ? 0 : 1));
for (auto &op : range) {
print(&op);
os << newLine;
}
currentIndent -= indentWidth;
}
void OperationPrinter::printValueID(Value value, bool printResultNo,
raw_ostream *streamOverride) const {
state->getSSANameState().printValueID(value, printResultNo,
streamOverride ? *streamOverride : os);
}
void OperationPrinter::printSuccessor(Block *successor) {
printBlockName(successor);
}
void OperationPrinter::printSuccessorAndUseList(Block *successor,
ValueRange succOperands) {
printBlockName(successor);
if (succOperands.empty())
return;
os << '(';
interleaveComma(succOperands,
[this](Value operand) { printValueID(operand); });
os << " : ";
interleaveComma(succOperands,
[this](Value operand) { printType(operand.getType()); });
os << ')';
}
void OperationPrinter::printRegion(Region &region, bool printEntryBlockArgs,
bool printBlockTerminators) {
os << " {" << newLine;
if (!region.empty()) {
auto *entryBlock = &region.front();
print(entryBlock, printEntryBlockArgs && entryBlock->getNumArguments() != 0,
printBlockTerminators);
for (auto &b : llvm::drop_begin(region.getBlocks(), 1))
print(&b);
}
os.indent(currentIndent) << "}";
}
void OperationPrinter::printAffineMapOfSSAIds(AffineMapAttr mapAttr,
ValueRange operands) {
AffineMap map = mapAttr.getValue();
unsigned numDims = map.getNumDims();
auto printValueName = [&](unsigned pos, bool isSymbol) {
unsigned index = isSymbol ? numDims + pos : pos;
assert(index < operands.size());
if (isSymbol)
os << "symbol(";
printValueID(operands[index]);
if (isSymbol)
os << ')';
};
interleaveComma(map.getResults(), [&](AffineExpr expr) {
printAffineExpr(expr, printValueName);
});
}
//===----------------------------------------------------------------------===//
// print and dump methods
//===----------------------------------------------------------------------===//
void Attribute::print(raw_ostream &os) const {
ModulePrinter(os).printAttribute(*this);
}
void Attribute::dump() const {
print(llvm::errs());
llvm::errs() << "\n";
}
void Type::print(raw_ostream &os) { ModulePrinter(os).printType(*this); }
void Type::dump() { print(llvm::errs()); }
void AffineMap::dump() const {
print(llvm::errs());
llvm::errs() << "\n";
}
void IntegerSet::dump() const {
print(llvm::errs());
llvm::errs() << "\n";
}
void AffineExpr::print(raw_ostream &os) const {
if (!expr) {
os << "<<NULL AFFINE EXPR>>";
return;
}
ModulePrinter(os).printAffineExpr(*this);
}
void AffineExpr::dump() const {
print(llvm::errs());
llvm::errs() << "\n";
}
void AffineMap::print(raw_ostream &os) const {
if (!map) {
os << "<<NULL AFFINE MAP>>";
return;
}
ModulePrinter(os).printAffineMap(*this);
}
void IntegerSet::print(raw_ostream &os) const {
ModulePrinter(os).printIntegerSet(*this);
}
void Value::print(raw_ostream &os) {
if (auto *op = getDefiningOp())
return op->print(os);
// TODO: Improve this.
BlockArgument arg = this->cast<BlockArgument>();
os << "<block argument> of type '" << arg.getType()
<< "' at index: " << arg.getArgNumber() << '\n';
}
void Value::print(raw_ostream &os, AsmState &state) {
if (auto *op = getDefiningOp())
return op->print(os, state);
// TODO: Improve this.
BlockArgument arg = this->cast<BlockArgument>();
os << "<block argument> of type '" << arg.getType()
<< "' at index: " << arg.getArgNumber() << '\n';
}
void Value::dump() {
print(llvm::errs());
llvm::errs() << "\n";
}
void Value::printAsOperand(raw_ostream &os, AsmState &state) {
// TODO: This doesn't necessarily capture all potential cases.
// Currently, region arguments can be shadowed when printing the main
// operation. If the IR hasn't been printed, this will produce the old SSA
// name and not the shadowed name.
state.getImpl().getSSANameState().printValueID(*this, /*printResultNo=*/true,
os);
}
void Operation::print(raw_ostream &os, OpPrintingFlags flags) {
// If this is a top level operation, we also print aliases.
if (!getParent() && !flags.shouldUseLocalScope()) {
AsmState state(this);
state.getImpl().initializeAliases(this, flags);
print(os, state, flags);
return;
}
// Find the operation to number from based upon the provided flags.
Operation *printedOp = this;
bool shouldUseLocalScope = flags.shouldUseLocalScope();
do {
// If we are printing local scope, stop at the first operation that is
// isolated from above.
if (shouldUseLocalScope && printedOp->isKnownIsolatedFromAbove())
break;
// Otherwise, traverse up to the next parent.
Operation *parentOp = printedOp->getParentOp();
if (!parentOp)
break;
printedOp = parentOp;
} while (true);
AsmState state(printedOp);
print(os, state, flags);
}
void Operation::print(raw_ostream &os, AsmState &state, OpPrintingFlags flags) {
OperationPrinter printer(os, flags, state.getImpl());
if (!getParent() && !flags.shouldUseLocalScope())
printer.printTopLevelOperation(this);
else
printer.print(this);
}
void Operation::dump() {
print(llvm::errs(), OpPrintingFlags().useLocalScope());
llvm::errs() << "\n";
}
void Block::print(raw_ostream &os) {
Operation *parentOp = getParentOp();
if (!parentOp) {
os << "<<UNLINKED BLOCK>>\n";
return;
}
// Get the top-level op.
while (auto *nextOp = parentOp->getParentOp())
parentOp = nextOp;
AsmState state(parentOp);
print(os, state);
}
void Block::print(raw_ostream &os, AsmState &state) {
OperationPrinter(os, /*flags=*/llvm::None, state.getImpl()).print(this);
}
void Block::dump() { print(llvm::errs()); }
/// Print out the name of the block without printing its body.
void Block::printAsOperand(raw_ostream &os, bool printType) {
Operation *parentOp = getParentOp();
if (!parentOp) {
os << "<<UNLINKED BLOCK>>\n";
return;
}
AsmState state(parentOp);
printAsOperand(os, state);
}
void Block::printAsOperand(raw_ostream &os, AsmState &state) {
OperationPrinter printer(os, /*flags=*/llvm::None, state.getImpl());
printer.printBlockName(this);
}