optimizer.hpp

// Copyright (c) 2016 Google Inc.
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
//     http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.

#ifndef SPIRV_TOOLS_OPTIMIZER_HPP_
#define SPIRV_TOOLS_OPTIMIZER_HPP_

#include <memory>
#include <ostream>
#include <string>
#include <unordered_map>
#include <vector>

#include "libspirv.hpp"

namespace spvtools {

namespace opt {
class Pass;
}

// C++ interface for SPIR-V optimization functionalities. It wraps the context
// (including target environment and the corresponding SPIR-V grammar) and
// provides methods for registering optimization passes and optimizing.
//
// Instances of this class provides basic thread-safety guarantee.
class Optimizer {
 public:
  // The token for an optimization pass. It is returned via one of the
  // Create*Pass() standalone functions at the end of this header file and
  // consumed by the RegisterPass() method. Tokens are one-time objects that
  // only support move; copying is not allowed.
  struct PassToken {
    struct Impl;  // Opaque struct for holding inernal data.

    PassToken(std::unique_ptr<Impl>);

    // Tokens for built-in passes should be created using Create*Pass functions
    // below; for out-of-tree passes, use this constructor instead.
    // Note that this API isn't guaranteed to be stable and may change without
    // preserving source or binary compatibility in the future.
    PassToken(std::unique_ptr<opt::Pass>&& pass);

    // Tokens can only be moved. Copying is disabled.
    PassToken(const PassToken&) = delete;
    PassToken(PassToken&&);
    PassToken& operator=(const PassToken&) = delete;
    PassToken& operator=(PassToken&&);

    ~PassToken();

    std::unique_ptr<Impl> impl_;  // Unique pointer to internal data.
  };

  // Constructs an instance with the given target |env|, which is used to decode
  // the binaries to be optimized later.
  //
  // The constructed instance will have an empty message consumer, which just
  // ignores all messages from the library. Use SetMessageConsumer() to supply
  // one if messages are of concern.
  explicit Optimizer(spv_target_env env);

  // Disables copy/move constructor/assignment operations.
  Optimizer(const Optimizer&) = delete;
  Optimizer(Optimizer&&) = delete;
  Optimizer& operator=(const Optimizer&) = delete;
  Optimizer& operator=(Optimizer&&) = delete;

  // Destructs this instance.
  ~Optimizer();

  // Sets the message consumer to the given |consumer|. The |consumer| will be
  // invoked once for each message communicated from the library.
  void SetMessageConsumer(MessageConsumer consumer);

  // Registers the given |pass| to this optimizer. Passes will be run in the
  // exact order of registration. The token passed in will be consumed by this
  // method.
  Optimizer& RegisterPass(PassToken&& pass);

  // Registers passes that attempt to improve performance of generated code.
  // This sequence of passes is subject to constant review and will change
  // from time to time.
  Optimizer& RegisterPerformancePasses();

  // Registers passes that attempt to improve the size of generated code.
  // This sequence of passes is subject to constant review and will change
  // from time to time.
  Optimizer& RegisterSizePasses();

  // Registers passes that attempt to legalize the generated code.
  //
  // Note: this recipe is specially for legalizing SPIR-V. It should be used
  // by compilers after translating HLSL source code literally. It should
  // *not* be used by general workloads for performance or size improvement.
  //
  // This sequence of passes is subject to constant review and will change
  // from time to time.
  Optimizer& RegisterLegalizationPasses();

  // Optimizes the given SPIR-V module |original_binary| and writes the
  // optimized binary into |optimized_binary|.
  // Returns true on successful optimization, whether or not the module is
  // modified. Returns false if errors occur when processing |original_binary|
  // using any of the registered passes. In that case, no further passes are
  // executed and the contents in |optimized_binary| may be invalid.
  //
  // It's allowed to alias |original_binary| to the start of |optimized_binary|.
  bool Run(const uint32_t* original_binary, size_t original_binary_size,
           std::vector<uint32_t>* optimized_binary) const;

  // Returns a vector of strings with all the pass names added to this
  // optimizer's pass manager. These strings are valid until the associated
  // pass manager is destroyed.
  std::vector<const char*> GetPassNames() const;

  // Sets the option to print the disassembly before each pass and after the
  // last pass.  If |out| is null, then no output is generated.  Otherwise,
  // output is sent to the |out| output stream.
  Optimizer& SetPrintAll(std::ostream* out);

  // Sets the option to print the resource utilization of each pass. If |out|
  // is null, then no output is generated. Otherwise, output is sent to the
  // |out| output stream.
  Optimizer& SetTimeReport(std::ostream* out);

 private:
  struct Impl;                  // Opaque struct for holding internal data.
  std::unique_ptr<Impl> impl_;  // Unique pointer to internal data.
};

// Creates a null pass.
// A null pass does nothing to the SPIR-V module to be optimized.
Optimizer::PassToken CreateNullPass();

// Creates a strip-debug-info pass.
// A strip-debug-info pass removes all debug instructions (as documented in
// Section 3.32.2 of the SPIR-V spec) of the SPIR-V module to be optimized.
Optimizer::PassToken CreateStripDebugInfoPass();

// Creates a strip-reflect-info pass.
// A strip-reflect-info pass removes all reflections instructions.
// For now, this is limited to removing decorations defined in
// SPV_GOOGLE_hlsl_functionality1.  The coverage may expand in
// the future.
Optimizer::PassToken CreateStripReflectInfoPass();

// Creates an eliminate-dead-functions pass.
// An eliminate-dead-functions pass will remove all functions that are not in
// the call trees rooted at entry points and exported functions.  These
// functions are not needed because they will never be called.
Optimizer::PassToken CreateEliminateDeadFunctionsPass();

// Creates a set-spec-constant-default-value pass from a mapping from spec-ids
// to the default values in the form of string.
// A set-spec-constant-default-value pass sets the default values for the
// spec constants that have SpecId decorations (i.e., those defined by
// OpSpecConstant{|True|False} instructions).
Optimizer::PassToken CreateSetSpecConstantDefaultValuePass(
    const std::unordered_map<uint32_t, std::string>& id_value_map);

// Creates a set-spec-constant-default-value pass from a mapping from spec-ids
// to the default values in the form of bit pattern.
// A set-spec-constant-default-value pass sets the default values for the
// spec constants that have SpecId decorations (i.e., those defined by
// OpSpecConstant{|True|False} instructions).
Optimizer::PassToken CreateSetSpecConstantDefaultValuePass(
    const std::unordered_map<uint32_t, std::vector<uint32_t>>& id_value_map);

// Creates a flatten-decoration pass.
// A flatten-decoration pass replaces grouped decorations with equivalent
// ungrouped decorations.  That is, it replaces each OpDecorationGroup
// instruction and associated OpGroupDecorate and OpGroupMemberDecorate
// instructions with equivalent OpDecorate and OpMemberDecorate instructions.
// The pass does not attempt to preserve debug information for instructions
// it removes.
Optimizer::PassToken CreateFlattenDecorationPass();

// Creates a freeze-spec-constant-value pass.
// A freeze-spec-constant pass specializes the value of spec constants to
// their default values. This pass only processes the spec constants that have
// SpecId decorations (defined by OpSpecConstant, OpSpecConstantTrue, or
// OpSpecConstantFalse instructions) and replaces them with their normal
// counterparts (OpConstant, OpConstantTrue, or OpConstantFalse). The
// corresponding SpecId annotation instructions will also be removed. This
// pass does not fold the newly added normal constants and does not process
// other spec constants defined by OpSpecConstantComposite or
// OpSpecConstantOp.
Optimizer::PassToken CreateFreezeSpecConstantValuePass();

// Creates a fold-spec-constant-op-and-composite pass.
// A fold-spec-constant-op-and-composite pass folds spec constants defined by
// OpSpecConstantOp or OpSpecConstantComposite instruction, to normal Constants
// defined by OpConstantTrue, OpConstantFalse, OpConstant, OpConstantNull, or
// OpConstantComposite instructions. Note that spec constants defined with
// OpSpecConstant, OpSpecConstantTrue, or OpSpecConstantFalse instructions are
// not handled, as these instructions indicate their value are not determined
// and can be changed in future. A spec constant is foldable if all of its
// value(s) can be determined from the module. E.g., an integer spec constant
// defined with OpSpecConstantOp instruction can be folded if its value won't
// change later. This pass will replace the original OpSpecContantOp instruction
// with an OpConstant instruction. When folding composite spec constants,
// new instructions may be inserted to define the components of the composite
// constant first, then the original spec constants will be replaced by
// OpConstantComposite instructions.
//
// There are some operations not supported yet:
//   OpSConvert, OpFConvert, OpQuantizeToF16 and
//   all the operations under Kernel capability.
// TODO(qining): Add support for the operations listed above.
Optimizer::PassToken CreateFoldSpecConstantOpAndCompositePass();

// Creates a unify-constant pass.
// A unify-constant pass de-duplicates the constants. Constants with the exact
// same value and identical form will be unified and only one constant will
// be kept for each unique pair of type and value.
// There are several cases not handled by this pass:
//  1) Constants defined by OpConstantNull instructions (null constants) and
//  constants defined by OpConstantFalse, OpConstant or OpConstantComposite
//  with value 0 (zero-valued normal constants) are not considered equivalent.
//  So null constants won't be used to replace zero-valued normal constants,
//  vice versa.
//  2) Whenever there are decorations to the constant's result id id, the
//  constant won't be handled, which means, it won't be used to replace any
//  other constants, neither can other constants replace it.
//  3) NaN in float point format with different bit patterns are not unified.
Optimizer::PassToken CreateUnifyConstantPass();

// Creates a eliminate-dead-constant pass.
// A eliminate-dead-constant pass removes dead constants, including normal
// contants defined by OpConstant, OpConstantComposite, OpConstantTrue, or
// OpConstantFalse and spec constants defined by OpSpecConstant,
// OpSpecConstantComposite, OpSpecConstantTrue, OpSpecConstantFalse or
// OpSpecConstantOp.
Optimizer::PassToken CreateEliminateDeadConstantPass();

// Creates a strength-reduction pass.
// A strength-reduction pass will look for opportunities to replace an
// instruction with an equivalent and less expensive one.  For example,
// multiplying by a power of 2 can be replaced by a bit shift.
Optimizer::PassToken CreateStrengthReductionPass();

// Creates a block merge pass.
// This pass searches for blocks with a single Branch to a block with no
// other predecessors and merges the blocks into a single block. Continue
// blocks and Merge blocks are not candidates for the second block.
//
// The pass is most useful after Dead Branch Elimination, which can leave
// such sequences of blocks. Merging them makes subsequent passes more
// effective, such as single block local store-load elimination.
//
// While this pass reduces the number of occurrences of this sequence, at
// this time it does not guarantee all such sequences are eliminated.
//
// Presence of phi instructions can inhibit this optimization. Handling
// these is left for future improvements.
Optimizer::PassToken CreateBlockMergePass();

// Creates an exhaustive inline pass.
// An exhaustive inline pass attempts to exhaustively inline all function
// calls in all functions in an entry point call tree. The intent is to enable,
// albeit through brute force, analysis and optimization across function
// calls by subsequent optimization passes. As the inlining is exhaustive,
// there is no attempt to optimize for size or runtime performance. Functions
// that are not in the call tree of an entry point are not changed.
Optimizer::PassToken CreateInlineExhaustivePass();

// Creates an opaque inline pass.
// An opaque inline pass inlines all function calls in all functions in all
// entry point call trees where the called function contains an opaque type
// in either its parameter types or return type. An opaque type is currently
// defined as Image, Sampler or SampledImage. The intent is to enable, albeit
// through brute force, analysis and optimization across these function calls
// by subsequent passes in order to remove the storing of opaque types which is
// not legal in Vulkan. Functions that are not in the call tree of an entry
// point are not changed.
Optimizer::PassToken CreateInlineOpaquePass();

// Creates a single-block local variable load/store elimination pass.
// For every entry point function, do single block memory optimization of
// function variables referenced only with non-access-chain loads and stores.
// For each targeted variable load, if previous store to that variable in the
// block, replace the load's result id with the value id of the store.
// If previous load within the block, replace the current load's result id
// with the previous load's result id. In either case, delete the current
// load. Finally, check if any remaining stores are useless, and delete store
// and variable if possible.
//
// The presence of access chain references and function calls can inhibit
// the above optimization.
//
// Only modules with relaxed logical addressing (see opt/instruction.h) are
// currently processed.
//
// This pass is most effective if preceeded by Inlining and
// LocalAccessChainConvert. This pass will reduce the work needed to be done
// by LocalSingleStoreElim and LocalMultiStoreElim.
//
// Only functions in the call tree of an entry point are processed.
Optimizer::PassToken CreateLocalSingleBlockLoadStoreElimPass();

// Create dead branch elimination pass.
// For each entry point function, this pass will look for SelectionMerge
// BranchConditionals with constant condition and convert to a Branch to
// the indicated label. It will delete resulting dead blocks.
//
// For all phi functions in merge block, replace all uses with the id
// corresponding to the living predecessor.
//
// Note that some branches and blocks may be left to avoid creating invalid
// control flow. Improving this is left to future work.
//
// This pass is most effective when preceeded by passes which eliminate
// local loads and stores, effectively propagating constant values where
// possible.
Optimizer::PassToken CreateDeadBranchElimPass();

// Creates an SSA local variable load/store elimination pass.
// For every entry point function, eliminate all loads and stores of function
// scope variables only referenced with non-access-chain loads and stores.
// Eliminate the variables as well.
//
// The presence of access chain references and function calls can inhibit
// the above optimization.
//
// Only shader modules with relaxed logical addressing (see opt/instruction.h)
// are currently processed. Currently modules with any extensions enabled are
// not processed. This is left for future work.
//
// This pass is most effective if preceeded by Inlining and
// LocalAccessChainConvert. LocalSingleStoreElim and LocalSingleBlockElim
// will reduce the work that this pass has to do.
Optimizer::PassToken CreateLocalMultiStoreElimPass();

// Creates a local access chain conversion pass.
// A local access chain conversion pass identifies all function scope
// variables which are accessed only with loads, stores and access chains
// with constant indices. It then converts all loads and stores of such
// variables into equivalent sequences of loads, stores, extracts and inserts.
//
// This pass only processes entry point functions. It currently only converts
// non-nested, non-ptr access chains. It does not process modules with
// non-32-bit integer types present. Optional memory access options on loads
// and stores are ignored as we are only processing function scope variables.
//
// This pass unifies access to these variables to a single mode and simplifies
// subsequent analysis and elimination of these variables along with their
// loads and stores allowing values to propagate to their points of use where
// possible.
Optimizer::PassToken CreateLocalAccessChainConvertPass();

// Creates a local single store elimination pass.
// For each entry point function, this pass eliminates loads and stores for
// function scope variable that are stored to only once, where possible. Only
// whole variable loads and stores are eliminated; access-chain references are
// not optimized. Replace all loads of such variables with the value that is
// stored and eliminate any resulting dead code.
//
// Currently, the presence of access chains and function calls can inhibit this
// pass, however the Inlining and LocalAccessChainConvert passes can make it
// more effective. In additional, many non-load/store memory operations are
// not supported and will prohibit optimization of a function. Support of
// these operations are future work.
//
// Only shader modules with relaxed logical addressing (see opt/instruction.h)
// are currently processed.
//
// This pass will reduce the work needed to be done by LocalSingleBlockElim
// and LocalMultiStoreElim and can improve the effectiveness of other passes
// such as DeadBranchElimination which depend on values for their analysis.
Optimizer::PassToken CreateLocalSingleStoreElimPass();

// Creates an insert/extract elimination pass.
// This pass processes each entry point function in the module, searching for
// extracts on a sequence of inserts. It further searches the sequence for an
// insert with indices identical to the extract. If such an insert can be
// found before hitting a conflicting insert, the extract's result id is
// replaced with the id of the values from the insert.
//
// Besides removing extracts this pass enables subsequent dead code elimination
// passes to delete the inserts. This pass performs best after access chains are
// converted to inserts and extracts and local loads and stores are eliminated.
Optimizer::PassToken CreateInsertExtractElimPass();

// Creates a dead insert elimination pass.
// This pass processes each entry point function in the module, searching for
// unreferenced inserts into composite types. These are most often unused
// stores to vector components. They are unused because they are never
// referenced, or because there is another insert to the same component between
// the insert and the reference. After removing the inserts, dead code
// elimination is attempted on the inserted values.
//
// This pass performs best after access chains are converted to inserts and
// extracts and local loads and stores are eliminated. While executing this
// pass can be advantageous on its own, it is also advantageous to execute
// this pass after CreateInsertExtractPass() as it will remove any unused
// inserts created by that pass.
Optimizer::PassToken CreateDeadInsertElimPass();

// Creates a pass to consolidate uniform references.
// For each entry point function in the module, first change all constant index
// access chain loads into equivalent composite extracts. Then consolidate
// identical uniform loads into one uniform load. Finally, consolidate
// identical uniform extracts into one uniform extract. This may require
// moving a load or extract to a point which dominates all uses.
//
// This pass requires a module to have structured control flow ie shader
// capability. It also requires logical addressing ie Addresses capability
// is not enabled. It also currently does not support any extensions.
//
// This pass currently only optimizes loads with a single index.
Optimizer::PassToken CreateCommonUniformElimPass();

// Create aggressive dead code elimination pass
// This pass eliminates unused code from the module. In addition,
// it detects and eliminates code which may have spurious uses but which do
// not contribute to the output of the function. The most common cause of
// such code sequences is summations in loops whose result is no longer used
// due to dead code elimination. This optimization has additional compile
// time cost over standard dead code elimination.
//
// This pass only processes entry point functions. It also only processes
// shaders with relaxed logical addressing (see opt/instruction.h). It
// currently will not process functions with function calls. Unreachable
// functions are deleted.
//
// This pass will be made more effective by first running passes that remove
// dead control flow and inlines function calls.
//
// This pass can be especially useful after running Local Access Chain
// Conversion, which tends to cause cycles of dead code to be left after
// Store/Load elimination passes are completed. These cycles cannot be
// eliminated with standard dead code elimination.
Optimizer::PassToken CreateAggressiveDCEPass();

// Creates a compact ids pass.
// The pass remaps result ids to a compact and gapless range starting from %1.
Optimizer::PassToken CreateCompactIdsPass();

// Creates a remove duplicate pass.
// This pass removes various duplicates:
// * duplicate capabilities;
// * duplicate extended instruction imports;
// * duplicate types;
// * duplicate decorations.
Optimizer::PassToken CreateRemoveDuplicatesPass();

// Creates a CFG cleanup pass.
// This pass removes cruft from the control flow graph of functions that are
// reachable from entry points and exported functions. It currently includes the
// following functionality:
//
// - Removal of unreachable basic blocks.
Optimizer::PassToken CreateCFGCleanupPass();

// Create dead variable elimination pass.
// This pass will delete module scope variables, along with their decorations,
// that are not referenced.
Optimizer::PassToken CreateDeadVariableEliminationPass();

// create merge return pass.
// changes functions that have multiple return statements so they have a single
// return statement.
//
// for structured control flow it is assumed that the only unreachable blocks in
// the function are trivial merge and continue blocks.
//
// a trivial merge block contains the label and an opunreachable instructions,
// nothing else.  a trivial continue block contain a label and an opbranch to
// the header, nothing else.
//
// these conditions are guaranteed to be met after running dead-branch
// elimination.
Optimizer::PassToken CreateMergeReturnPass();

// Create value numbering pass.
// This pass will look for instructions in the same basic block that compute the
// same value, and remove the redundant ones.
Optimizer::PassToken CreateLocalRedundancyEliminationPass();

// Create LICM pass.
// This pass will look for invariant instructions inside loops and hoist them to
// the loops preheader.
Optimizer::PassToken CreateLoopInvariantCodeMotionPass();

// Creates a loop fission pass.
// This pass will split all top level loops whose register pressure exceedes the
// given |threshold|.
Optimizer::PassToken CreateLoopFissionPass(size_t threshold);

// Creates a loop fusion pass.
// This pass will look for adjacent loops that are compatible and legal to be
// fused. The fuse all such loops as long as the register usage for the fused
// loop stays under the threshold defined by |max_registers_per_loop|.
Optimizer::PassToken CreateLoopFusionPass(size_t max_registers_per_loop);

// Creates a loop peeling pass.
// This pass will look for conditions inside a loop that are true or false only
// for the N first or last iteration. For loop with such condition, those N
// iterations of the loop will be executed outside of the main loop.
// To limit code size explosion, the loop peeling can only happen if the code
// size growth for each loop is under |code_growth_threshold|.
Optimizer::PassToken CreateLoopPeelingPass();

// Creates a loop unswitch pass.
// This pass will look for loop independent branch conditions and move the
// condition out of the loop and version the loop based on the taken branch.
// Works best after LICM and local multi store elimination pass.
Optimizer::PassToken CreateLoopUnswitchPass();

// Create global value numbering pass.
// This pass will look for instructions where the same value is computed on all
// paths leading to the instruction.  Those instructions are deleted.
Optimizer::PassToken CreateRedundancyEliminationPass();

// Create scalar replacement pass.
// This pass replaces composite function scope variables with variables for each
// element if those elements are accessed individually.  The parameter is a
// limit on the number of members in the composite variable that the pass will
// consider replacing.
Optimizer::PassToken CreateScalarReplacementPass(uint32_t size_limit = 100);

// Create a private to local pass.
// This pass looks for variables delcared in the private storage class that are
// used in only one function.  Those variables are moved to the function storage
// class in the function that they are used.
Optimizer::PassToken CreatePrivateToLocalPass();

// Creates a conditional constant propagation (CCP) pass.
// This pass implements the SSA-CCP algorithm in
//
//      Constant propagation with conditional branches,
//      Wegman and Zadeck, ACM TOPLAS 13(2):181-210.
//
// Constant values in expressions and conditional jumps are folded and
// simplified. This may reduce code size by removing never executed jump targets
// and computations with constant operands.
Optimizer::PassToken CreateCCPPass();

// Creates a workaround driver bugs pass.  This pass attempts to work around
// a known driver bug (issue #1209) by identifying the bad code sequences and
// rewriting them.
//
// Current workaround: Avoid OpUnreachable instructions in loops.
Optimizer::PassToken CreateWorkaround1209Pass();

// Creates a pass that converts if-then-else like assignments into OpSelect.
Optimizer::PassToken CreateIfConversionPass();

// Creates a pass that will replace instructions that are not valid for the
// current shader stage by constants.  Has no effect on non-shader modules.
Optimizer::PassToken CreateReplaceInvalidOpcodePass();

// Creates a pass that simplifies instructions using the instruction folder.
Optimizer::PassToken CreateSimplificationPass();

// Create loop unroller pass.
// Creates a pass to unroll loops which have the "Unroll" loop control
// mask set. The loops must meet a specific criteria in order to be unrolled
// safely this criteria is checked before doing the unroll by the
// LoopUtils::CanPerformUnroll method. Any loop that does not meet the criteria
// won't be unrolled. See CanPerformUnroll LoopUtils.h for more information.
Optimizer::PassToken CreateLoopUnrollPass(bool fully_unroll, int factor = 0);

// Create the SSA rewrite pass.
// This pass converts load/store operations on function local variables into
// operations on SSA IDs.  This allows SSA optimizers to act on these variables.
// Only variables that are local to the function and of supported types are
// processed (see IsSSATargetVar for details).
Optimizer::PassToken CreateSSARewritePass();

// Create copy propagate arrays pass.
// This pass looks to copy propagate memory references for arrays.  It looks
// for specific code patterns to recognize array copies.
Optimizer::PassToken CreateCopyPropagateArraysPass();

// Create a vector dce pass.
// This pass looks for components of vectors that are unused, and removes them
// from the vector.  Note this would still leave around lots of dead code that
// a pass of ADCE will be able to remove.
Optimizer::PassToken CreateVectorDCEPass();

// Create a pass to reduce the size of loads.
// This pass looks for loads of structures where only a few of its members are
// used.  It replaces the loads feeding an OpExtract with an OpAccessChain and
// a load of the specific elements.
Optimizer::PassToken CreateReduceLoadSizePass();

}  // namespace spvtools

#endif  // SPIRV_TOOLS_OPTIMIZER_HPP_