SPIRV-Tools/source/opt/scalar_analysis.h
Toomas Remmelg 1dc2458060 Add a loop fusion pass.
This pass will look for adjacent loops that are compatible and legal to
be fused.

Loops are compatible if:

- they both have one induction variable
- they have the same upper and lower bounds
    - same initial value
    - same condition
- they have the same update step
- they are adjacent
- there are no break/continue in either of them

Fusion is legal if:

- fused loops do not have any dependencies with dependence distance
  greater than 0 that did not exist in the original loops.
- there are no function calls in the loops (could have side-effects)
- there are no barriers in the loops

It will fuse all such loops as long as the number of registers used for
the fused loop stays under the threshold defined by
max_registers_per_loop.
2018-05-01 15:40:37 -04:00

312 lines
12 KiB
C++

// Copyright (c) 2018 Google LLC.
//
// 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 SOURCE_OPT_SCALAR_ANALYSIS_H_
#define SOURCE_OPT_SCALAR_ANALYSIS_H_
#include <algorithm>
#include <cstdint>
#include <map>
#include <memory>
#include <unordered_set>
#include <vector>
#include "opt/basic_block.h"
#include "opt/instruction.h"
#include "opt/scalar_analysis_nodes.h"
namespace spvtools {
namespace ir {
class IRContext;
class Loop;
} // namespace ir
namespace opt {
// Manager for the Scalar Evolution analysis. Creates and maintains a DAG of
// scalar operations generated from analysing the use def graph from incoming
// instructions. Each node is hashed as it is added so like node (for instance,
// two induction variables i=0,i++ and j=0,j++) become the same node. After
// creating a DAG with AnalyzeInstruction it can the be simplified into a more
// usable form with SimplifyExpression.
class ScalarEvolutionAnalysis {
public:
explicit ScalarEvolutionAnalysis(ir::IRContext* context);
// Create a unary negative node on |operand|.
SENode* CreateNegation(SENode* operand);
// Creates a subtraction between the two operands by adding |operand_1| to the
// negation of |operand_2|.
SENode* CreateSubtraction(SENode* operand_1, SENode* operand_2);
// Create an addition node between two operands. The |simplify| when set will
// allow the function to return an SEConstant instead of an addition if the
// two input operands are also constant.
SENode* CreateAddNode(SENode* operand_1, SENode* operand_2);
// Create a multiply node between two operands.
SENode* CreateMultiplyNode(SENode* operand_1, SENode* operand_2);
// Create a node representing a constant integer.
SENode* CreateConstant(int64_t integer);
// Create a value unknown node, such as a load.
SENode* CreateValueUnknownNode(const ir::Instruction* inst);
// Create a CantComputeNode. Used to exit out of analysis.
SENode* CreateCantComputeNode();
// Create a new recurrent node with |offset| and |coefficient|, with respect
// to |loop|.
SENode* CreateRecurrentExpression(const ir::Loop* loop, SENode* offset,
SENode* coefficient);
// Construct the DAG by traversing use def chain of |inst|.
SENode* AnalyzeInstruction(const ir::Instruction* inst);
// Simplify the |node| by grouping like terms or if contains a recurrent
// expression, rewrite the graph so the whole DAG (from |node| down) is in
// terms of that recurrent expression.
//
// For example.
// Induction variable i=0, i++ would produce Rec(0,1) so i+1 could be
// transformed into Rec(1,1).
//
// X+X*2+Y-Y+34-17 would be transformed into 3*X + 17, where X and Y are
// ValueUnknown nodes (such as a load instruction).
SENode* SimplifyExpression(SENode* node);
// Add |prospective_node| into the cache and return a raw pointer to it. If
// |prospective_node| is already in the cache just return the raw pointer.
SENode* GetCachedOrAdd(std::unique_ptr<SENode> prospective_node);
// Checks that the graph starting from |node| is invariant to the |loop|.
bool IsLoopInvariant(const ir::Loop* loop, const SENode* node) const;
// Sets |is_gt_zero| to true if |node| represent a value always strictly
// greater than 0. The result of |is_gt_zero| is valid only if the function
// returns true.
bool IsAlwaysGreaterThanZero(SENode* node, bool* is_gt_zero) const;
// Sets |is_ge_zero| to true if |node| represent a value greater or equals to
// 0. The result of |is_ge_zero| is valid only if the function returns true.
bool IsAlwaysGreaterOrEqualToZero(SENode* node, bool* is_ge_zero) const;
// Find the recurrent term belonging to |loop| in the graph starting from
// |node| and return the coefficient of that recurrent term. Constant zero
// will be returned if no recurrent could be found. |node| should be in
// simplest form.
SENode* GetCoefficientFromRecurrentTerm(SENode* node, const ir::Loop* loop);
// Return a rebuilt graph starting from |node| with the recurrent expression
// belonging to |loop| being zeroed out. Returned node will be simplified.
SENode* BuildGraphWithoutRecurrentTerm(SENode* node, const ir::Loop* loop);
// Return the recurrent term belonging to |loop| if it appears in the graph
// starting at |node| or null if it doesn't.
SERecurrentNode* GetRecurrentTerm(SENode* node, const ir::Loop* loop);
SENode* UpdateChildNode(SENode* parent, SENode* child, SENode* new_child);
// The loops in |loop_pair| will be considered the same when constructing
// SERecurrentNode objects. This enables analysing dependencies that will be
// created during loop fusion.
void AddLoopsToPretendAreTheSame(
const std::pair<const ir::Loop*, const ir::Loop*>& loop_pair) {
pretend_equal_[std::get<1>(loop_pair)] = std::get<0>(loop_pair);
}
private:
SENode* AnalyzeConstant(const ir::Instruction* inst);
// Handles both addition and subtraction. If the |instruction| is OpISub
// then the resulting node will be op1+(-op2) otherwise if it is OpIAdd then
// the result will be op1+op2. |instruction| must be OpIAdd or OpISub.
SENode* AnalyzeAddOp(const ir::Instruction* instruction);
SENode* AnalyzeMultiplyOp(const ir::Instruction* multiply);
SENode* AnalyzePhiInstruction(const ir::Instruction* phi);
ir::IRContext* context_;
// A map of instructions to SENodes. This is used to track recurrent
// expressions as they are added when analyzing instructions. Recurrent
// expressions come from phi nodes which by nature can include recursion so we
// check if nodes have already been built when analyzing instructions.
std::map<const ir::Instruction*, SENode*> recurrent_node_map_;
// On creation we create and cache the CantCompute node so we not need to
// perform a needless create step.
SENode* cached_cant_compute_;
// Helper functor to allow two unique_ptr to nodes to be compare. Only
// needed
// for the unordered_set implementation.
struct NodePointersEquality {
bool operator()(const std::unique_ptr<SENode>& lhs,
const std::unique_ptr<SENode>& rhs) const {
return *lhs == *rhs;
}
};
// Cache of nodes. All pointers to the nodes are references to the memory
// managed by they set.
std::unordered_set<std::unique_ptr<SENode>, SENodeHash, NodePointersEquality>
node_cache_;
// Loops that should be considered the same for performing analysis for loop
// fusion.
std::map<const ir::Loop*, const ir::Loop*> pretend_equal_;
};
// Wrapping class to manipulate SENode pointer using + - * / operators.
class SExpression {
public:
// Implicit on purpose !
SExpression(SENode* node)
: node_(node->GetParentAnalysis()->SimplifyExpression(node)),
scev_(node->GetParentAnalysis()) {}
inline operator SENode*() const { return node_; }
inline SENode* operator->() const { return node_; }
const SENode& operator*() const { return *node_; }
inline ScalarEvolutionAnalysis* GetScalarEvolutionAnalysis() const {
return scev_;
}
inline SExpression operator+(SENode* rhs) const;
template <typename T,
typename std::enable_if<std::is_integral<T>::value, int>::type = 0>
inline SExpression operator+(T integer) const;
inline SExpression operator+(SExpression rhs) const;
inline SExpression operator-() const;
inline SExpression operator-(SENode* rhs) const;
template <typename T,
typename std::enable_if<std::is_integral<T>::value, int>::type = 0>
inline SExpression operator-(T integer) const;
inline SExpression operator-(SExpression rhs) const;
inline SExpression operator*(SENode* rhs) const;
template <typename T,
typename std::enable_if<std::is_integral<T>::value, int>::type = 0>
inline SExpression operator*(T integer) const;
inline SExpression operator*(SExpression rhs) const;
template <typename T,
typename std::enable_if<std::is_integral<T>::value, int>::type = 0>
inline std::pair<SExpression, int64_t> operator/(T integer) const;
// Try to perform a division. Returns the pair <this.node_ / rhs, division
// remainder>. If it fails to simplify it, the function returns a
// CanNotCompute node.
std::pair<SExpression, int64_t> operator/(SExpression rhs) const;
private:
SENode* node_;
ScalarEvolutionAnalysis* scev_;
};
inline SExpression SExpression::operator+(SENode* rhs) const {
return scev_->CreateAddNode(node_, rhs);
}
template <typename T,
typename std::enable_if<std::is_integral<T>::value, int>::type>
inline SExpression SExpression::operator+(T integer) const {
return *this + scev_->CreateConstant(integer);
}
inline SExpression SExpression::operator+(SExpression rhs) const {
return *this + rhs.node_;
}
inline SExpression SExpression::operator-() const {
return scev_->CreateNegation(node_);
}
inline SExpression SExpression::operator-(SENode* rhs) const {
return *this + scev_->CreateNegation(rhs);
}
template <typename T,
typename std::enable_if<std::is_integral<T>::value, int>::type>
inline SExpression SExpression::operator-(T integer) const {
return *this - scev_->CreateConstant(integer);
}
inline SExpression SExpression::operator-(SExpression rhs) const {
return *this - rhs.node_;
}
inline SExpression SExpression::operator*(SENode* rhs) const {
return scev_->CreateMultiplyNode(node_, rhs);
}
template <typename T,
typename std::enable_if<std::is_integral<T>::value, int>::type>
inline SExpression SExpression::operator*(T integer) const {
return *this * scev_->CreateConstant(integer);
}
inline SExpression SExpression::operator*(SExpression rhs) const {
return *this * rhs.node_;
}
template <typename T,
typename std::enable_if<std::is_integral<T>::value, int>::type>
inline std::pair<SExpression, int64_t> SExpression::operator/(T integer) const {
return *this / scev_->CreateConstant(integer);
}
template <typename T,
typename std::enable_if<std::is_integral<T>::value, int>::type>
inline SExpression operator+(T lhs, SExpression rhs) {
return rhs + lhs;
}
inline SExpression operator+(SENode* lhs, SExpression rhs) { return rhs + lhs; }
template <typename T,
typename std::enable_if<std::is_integral<T>::value, int>::type>
inline SExpression operator-(T lhs, SExpression rhs) {
return SExpression{rhs.GetScalarEvolutionAnalysis()->CreateConstant(lhs)} -
rhs;
}
inline SExpression operator-(SENode* lhs, SExpression rhs) {
return SExpression{lhs} - rhs;
}
template <typename T,
typename std::enable_if<std::is_integral<T>::value, int>::type>
inline SExpression operator*(T lhs, SExpression rhs) {
return rhs * lhs;
}
inline SExpression operator*(SENode* lhs, SExpression rhs) { return rhs * lhs; }
template <typename T,
typename std::enable_if<std::is_integral<T>::value, int>::type>
inline std::pair<SExpression, int64_t> operator/(T lhs, SExpression rhs) {
return SExpression{rhs.GetScalarEvolutionAnalysis()->CreateConstant(lhs)} /
rhs;
}
inline std::pair<SExpression, int64_t> operator/(SENode* lhs, SExpression rhs) {
return SExpression{lhs} / rhs;
}
} // namespace opt
} // namespace spvtools
#endif // SOURCE_OPT_SCALAR_ANALYSIS_H__