third_party_spirv-tools/source/opt/fold.cpp
Alan Baker 802cf053c7 Merge arithmetic with non-trivial constant operands
Adding basis of arithmetic merging

* Refactored constant collection in ConstantManager
* New rules:
 * consecutive negates
 * negate of arithmetic op with a constant
 * consecutive muls
 * reciprocal of div

* Removed IRContext::CanFoldFloatingPoint
 * replaced by Instruction::IsFloatingPointFoldingAllowed
* Fixed some bad tests
* added some header comments

Added PerformIntegerOperation

* minor fixes to constants and tests
* fixed IntMultiplyBy1 to work with 64 bit ints
* added tests for integer mul merging

Adding test for vector integer multiply merging

Adding support for merging integer add and sub through negate

* Added tests

Adding rules to merge mult with preceding divide

* Has a couple tests, but needs more
* Added more comments

Fixed bug in integer division folding

* Will no longer merge through integer division if there would be a
remainder in the division
* Added a bunch more tests

Adding rules to merge divide and multiply through divide

* Improved comments
* Added tests

Adding rules to handle mul or div of a negation

* Added tests

Changes for review

* Early exit if no constants are involved in more functions
* fixed some comments
* removed unused declaration
* clarified some logic

Adding new rules for add and subtract

* Fold adds of adds, subtracts or negates
* Fold subtracts of adds, subtracts or negates
* Added tests
2018-02-27 13:02:13 -05:00

684 lines
22 KiB
C++

// Copyright (c) 2017 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.
#include "fold.h"
#include <cassert>
#include <cstdint>
#include <vector>
#include "const_folding_rules.h"
#include "def_use_manager.h"
#include "folding_rules.h"
#include "ir_builder.h"
#include "ir_context.h"
namespace spvtools {
namespace opt {
namespace {
#ifndef INT32_MIN
#define INT32_MIN (-2147483648)
#endif
#ifndef INT32_MAX
#define INT32_MAX 2147483647
#endif
#ifndef UINT32_MAX
#define UINT32_MAX 0xffffffff /* 4294967295U */
#endif
// Returns the single-word result from performing the given unary operation on
// the operand value which is passed in as a 32-bit word.
uint32_t UnaryOperate(SpvOp opcode, uint32_t operand) {
switch (opcode) {
// Arthimetics
case SpvOp::SpvOpSNegate:
return -static_cast<int32_t>(operand);
case SpvOp::SpvOpNot:
return ~operand;
case SpvOp::SpvOpLogicalNot:
return !static_cast<bool>(operand);
default:
assert(false &&
"Unsupported unary operation for OpSpecConstantOp instruction");
return 0u;
}
}
// Returns the single-word result from performing the given binary operation on
// the operand values which are passed in as two 32-bit word.
uint32_t BinaryOperate(SpvOp opcode, uint32_t a, uint32_t b) {
switch (opcode) {
// Arthimetics
case SpvOp::SpvOpIAdd:
return a + b;
case SpvOp::SpvOpISub:
return a - b;
case SpvOp::SpvOpIMul:
return a * b;
case SpvOp::SpvOpUDiv:
assert(b != 0);
return a / b;
case SpvOp::SpvOpSDiv:
assert(b != 0u);
return (static_cast<int32_t>(a)) / (static_cast<int32_t>(b));
case SpvOp::SpvOpSRem: {
// The sign of non-zero result comes from the first operand: a. This is
// guaranteed by C++11 rules for integer division operator. The division
// result is rounded toward zero, so the result of '%' has the sign of
// the first operand.
assert(b != 0u);
return static_cast<int32_t>(a) % static_cast<int32_t>(b);
}
case SpvOp::SpvOpSMod: {
// The sign of non-zero result comes from the second operand: b
assert(b != 0u);
int32_t rem = BinaryOperate(SpvOp::SpvOpSRem, a, b);
int32_t b_prim = static_cast<int32_t>(b);
return (rem + b_prim) % b_prim;
}
case SpvOp::SpvOpUMod:
assert(b != 0u);
return (a % b);
// Shifting
case SpvOp::SpvOpShiftRightLogical: {
return a >> b;
}
case SpvOp::SpvOpShiftRightArithmetic:
return (static_cast<int32_t>(a)) >> b;
case SpvOp::SpvOpShiftLeftLogical:
return a << b;
// Bitwise operations
case SpvOp::SpvOpBitwiseOr:
return a | b;
case SpvOp::SpvOpBitwiseAnd:
return a & b;
case SpvOp::SpvOpBitwiseXor:
return a ^ b;
// Logical
case SpvOp::SpvOpLogicalEqual:
return (static_cast<bool>(a)) == (static_cast<bool>(b));
case SpvOp::SpvOpLogicalNotEqual:
return (static_cast<bool>(a)) != (static_cast<bool>(b));
case SpvOp::SpvOpLogicalOr:
return (static_cast<bool>(a)) || (static_cast<bool>(b));
case SpvOp::SpvOpLogicalAnd:
return (static_cast<bool>(a)) && (static_cast<bool>(b));
// Comparison
case SpvOp::SpvOpIEqual:
return a == b;
case SpvOp::SpvOpINotEqual:
return a != b;
case SpvOp::SpvOpULessThan:
return a < b;
case SpvOp::SpvOpSLessThan:
return (static_cast<int32_t>(a)) < (static_cast<int32_t>(b));
case SpvOp::SpvOpUGreaterThan:
return a > b;
case SpvOp::SpvOpSGreaterThan:
return (static_cast<int32_t>(a)) > (static_cast<int32_t>(b));
case SpvOp::SpvOpULessThanEqual:
return a <= b;
case SpvOp::SpvOpSLessThanEqual:
return (static_cast<int32_t>(a)) <= (static_cast<int32_t>(b));
case SpvOp::SpvOpUGreaterThanEqual:
return a >= b;
case SpvOp::SpvOpSGreaterThanEqual:
return (static_cast<int32_t>(a)) >= (static_cast<int32_t>(b));
default:
assert(false &&
"Unsupported binary operation for OpSpecConstantOp instruction");
return 0u;
}
}
// Returns the single-word result from performing the given ternary operation
// on the operand values which are passed in as three 32-bit word.
uint32_t TernaryOperate(SpvOp opcode, uint32_t a, uint32_t b, uint32_t c) {
switch (opcode) {
case SpvOp::SpvOpSelect:
return (static_cast<bool>(a)) ? b : c;
default:
assert(false &&
"Unsupported ternary operation for OpSpecConstantOp instruction");
return 0u;
}
}
// Returns the single-word result from performing the given operation on the
// operand words. This only works with 32-bit operations and uses boolean
// convention that 0u is false, and anything else is boolean true.
// TODO(qining): Support operands other than 32-bit wide.
uint32_t OperateWords(SpvOp opcode,
const std::vector<uint32_t>& operand_words) {
switch (operand_words.size()) {
case 1:
return UnaryOperate(opcode, operand_words.front());
case 2:
return BinaryOperate(opcode, operand_words.front(), operand_words.back());
case 3:
return TernaryOperate(opcode, operand_words[0], operand_words[1],
operand_words[2]);
default:
assert(false && "Invalid number of operands");
return 0;
}
}
bool FoldInstructionInternal(ir::Instruction* inst) {
ir::IRContext* context = inst->context();
auto identity_map = [](uint32_t id) { return id; };
ir::Instruction* folded_inst = FoldInstructionToConstant(inst, identity_map);
if (folded_inst != nullptr) {
inst->SetOpcode(SpvOpCopyObject);
inst->SetInOperands({{SPV_OPERAND_TYPE_ID, {folded_inst->result_id()}}});
return true;
}
SpvOp opcode = inst->opcode();
analysis::ConstantManager* const_manager = context->get_constant_mgr();
std::vector<const analysis::Constant*> constants =
const_manager->GetOperandConstants(inst);
static FoldingRules* rules = new FoldingRules();
for (FoldingRule rule : rules->GetRulesForOpcode(opcode)) {
if (rule(inst, constants)) {
return true;
}
}
return false;
}
} // namespace
const ConstantFoldingRules& GetConstantFoldingRules() {
static ConstantFoldingRules* rules = new ConstantFoldingRules();
return *rules;
}
// Returns the result of performing an operation on scalar constant operands.
// This function extracts the operand values as 32 bit words and returns the
// result in 32 bit word. Scalar constants with longer than 32-bit width are
// not accepted in this function.
uint32_t FoldScalars(SpvOp opcode,
const std::vector<const analysis::Constant*>& operands) {
assert(IsFoldableOpcode(opcode) &&
"Unhandled instruction opcode in FoldScalars");
std::vector<uint32_t> operand_values_in_raw_words;
for (const auto& operand : operands) {
if (const analysis::ScalarConstant* scalar = operand->AsScalarConstant()) {
const auto& scalar_words = scalar->words();
assert(scalar_words.size() == 1 &&
"Scalar constants with longer than 32-bit width are not allowed "
"in FoldScalars()");
operand_values_in_raw_words.push_back(scalar_words.front());
} else if (operand->AsNullConstant()) {
operand_values_in_raw_words.push_back(0u);
} else {
assert(false &&
"FoldScalars() only accepts ScalarConst or NullConst type of "
"constant");
}
}
return OperateWords(opcode, operand_values_in_raw_words);
}
// Returns true if |inst| is a binary operation that takes two integers as
// parameters and folds to a constant that can be represented as an unsigned
// 32-bit value when the ids have been replaced by |id_map|. If |inst| can be
// folded, the resulting value is returned in |*result|. Valid result types for
// the instruction are any integer (signed or unsigned) with 32-bits or less, or
// a boolean value.
bool FoldBinaryIntegerOpToConstant(ir::Instruction* inst,
std::function<uint32_t(uint32_t)> id_map,
uint32_t* result) {
SpvOp opcode = inst->opcode();
ir::IRContext* context = inst->context();
analysis::ConstantManager* const_manger = context->get_constant_mgr();
uint32_t ids[2];
const analysis::IntConstant* constants[2];
for (uint32_t i = 0; i < 2; i++) {
const ir::Operand* operand = &inst->GetInOperand(i);
if (operand->type != SPV_OPERAND_TYPE_ID) {
return false;
}
ids[i] = id_map(operand->words[0]);
const analysis::Constant* constant =
const_manger->FindDeclaredConstant(ids[i]);
constants[i] = (constant != nullptr ? constant->AsIntConstant() : nullptr);
}
switch (opcode) {
// Arthimetics
case SpvOp::SpvOpIMul:
for (uint32_t i = 0; i < 2; i++) {
if (constants[i] != nullptr && constants[i]->IsZero()) {
*result = 0;
return true;
}
}
break;
case SpvOp::SpvOpUDiv:
case SpvOp::SpvOpSDiv:
case SpvOp::SpvOpSRem:
case SpvOp::SpvOpSMod:
case SpvOp::SpvOpUMod:
// This changes undefined behaviour (ie divide by 0) into a 0.
for (uint32_t i = 0; i < 2; i++) {
if (constants[i] != nullptr && constants[i]->IsZero()) {
*result = 0;
return true;
}
}
break;
// Shifting
case SpvOp::SpvOpShiftRightLogical:
case SpvOp::SpvOpShiftLeftLogical:
if (constants[1] != nullptr) {
// When shifting by a value larger than the size of the result, the
// result is undefined. We are setting the undefined behaviour to a
// result of 0.
uint32_t shift_amount = constants[1]->GetU32BitValue();
if (shift_amount >= 32) {
*result = 0;
return true;
}
}
break;
// Bitwise operations
case SpvOp::SpvOpBitwiseOr:
for (uint32_t i = 0; i < 2; i++) {
if (constants[i] != nullptr) {
// TODO: Change the mask against a value based on the bit width of the
// instruction result type. This way we can handle say 16-bit values
// as well.
uint32_t mask = constants[i]->GetU32BitValue();
if (mask == 0xFFFFFFFF) {
*result = 0xFFFFFFFF;
return true;
}
}
}
break;
case SpvOp::SpvOpBitwiseAnd:
for (uint32_t i = 0; i < 2; i++) {
if (constants[i] != nullptr) {
if (constants[i]->IsZero()) {
*result = 0;
return true;
}
}
}
break;
// Comparison
case SpvOp::SpvOpULessThan:
if (constants[0] != nullptr &&
constants[0]->GetU32BitValue() == UINT32_MAX) {
*result = false;
return true;
}
if (constants[1] != nullptr && constants[1]->GetU32BitValue() == 0) {
*result = false;
return true;
}
break;
case SpvOp::SpvOpSLessThan:
if (constants[0] != nullptr &&
constants[0]->GetS32BitValue() == INT32_MAX) {
*result = false;
return true;
}
if (constants[1] != nullptr &&
constants[1]->GetS32BitValue() == INT32_MIN) {
*result = false;
return true;
}
break;
case SpvOp::SpvOpUGreaterThan:
if (constants[0] != nullptr && constants[0]->IsZero()) {
*result = false;
return true;
}
if (constants[1] != nullptr &&
constants[1]->GetU32BitValue() == UINT32_MAX) {
*result = false;
return true;
}
break;
case SpvOp::SpvOpSGreaterThan:
if (constants[0] != nullptr &&
constants[0]->GetS32BitValue() == INT32_MIN) {
*result = false;
return true;
}
if (constants[1] != nullptr &&
constants[1]->GetS32BitValue() == INT32_MAX) {
*result = false;
return true;
}
break;
case SpvOp::SpvOpULessThanEqual:
if (constants[0] != nullptr && constants[0]->IsZero()) {
*result = true;
return true;
}
if (constants[1] != nullptr &&
constants[1]->GetU32BitValue() == UINT32_MAX) {
*result = true;
return true;
}
break;
case SpvOp::SpvOpSLessThanEqual:
if (constants[0] != nullptr &&
constants[0]->GetS32BitValue() == INT32_MIN) {
*result = true;
return true;
}
if (constants[1] != nullptr &&
constants[1]->GetS32BitValue() == INT32_MAX) {
*result = true;
return true;
}
break;
case SpvOp::SpvOpUGreaterThanEqual:
if (constants[0] != nullptr &&
constants[0]->GetU32BitValue() == UINT32_MAX) {
*result = true;
return true;
}
if (constants[1] != nullptr && constants[1]->GetU32BitValue() == 0) {
*result = true;
return true;
}
break;
case SpvOp::SpvOpSGreaterThanEqual:
if (constants[0] != nullptr &&
constants[0]->GetS32BitValue() == INT32_MAX) {
*result = true;
return true;
}
if (constants[1] != nullptr &&
constants[1]->GetS32BitValue() == INT32_MIN) {
*result = true;
return true;
}
break;
default:
break;
}
return false;
}
// Returns true if |inst| is a binary operation on two boolean values, and folds
// to a constant boolean value when the ids have been replaced using |id_map|.
// If |inst| can be folded, the result value is returned in |*result|.
bool FoldBinaryBooleanOpToConstant(ir::Instruction* inst,
std::function<uint32_t(uint32_t)> id_map,
uint32_t* result) {
SpvOp opcode = inst->opcode();
ir::IRContext* context = inst->context();
analysis::ConstantManager* const_manger = context->get_constant_mgr();
uint32_t ids[2];
const analysis::BoolConstant* constants[2];
for (uint32_t i = 0; i < 2; i++) {
const ir::Operand* operand = &inst->GetInOperand(i);
if (operand->type != SPV_OPERAND_TYPE_ID) {
return false;
}
ids[i] = id_map(operand->words[0]);
const analysis::Constant* constant =
const_manger->FindDeclaredConstant(ids[i]);
constants[i] = (constant != nullptr ? constant->AsBoolConstant() : nullptr);
}
switch (opcode) {
// Logical
case SpvOp::SpvOpLogicalOr:
for (uint32_t i = 0; i < 2; i++) {
if (constants[i] != nullptr) {
if (constants[i]->value()) {
*result = true;
return true;
}
}
}
break;
case SpvOp::SpvOpLogicalAnd:
for (uint32_t i = 0; i < 2; i++) {
if (constants[i] != nullptr) {
if (!constants[i]->value()) {
*result = false;
return true;
}
}
}
break;
default:
break;
}
return false;
}
// Returns true if |inst| can be folded to an constant when the ids have been
// substituted using id_map. If it can, the value is returned in |result|. If
// not, |result| is unchanged. It is assumed that not all operands are
// constant. Those cases are handled by |FoldScalar|.
bool FoldIntegerOpToConstant(ir::Instruction* inst,
std::function<uint32_t(uint32_t)> id_map,
uint32_t* result) {
assert(IsFoldableOpcode(inst->opcode()) &&
"Unhandled instruction opcode in FoldScalars");
switch (inst->NumInOperands()) {
case 2:
return FoldBinaryIntegerOpToConstant(inst, id_map, result) ||
FoldBinaryBooleanOpToConstant(inst, id_map, result);
default:
return false;
}
}
std::vector<uint32_t> FoldVectors(
SpvOp opcode, uint32_t num_dims,
const std::vector<const analysis::Constant*>& operands) {
assert(IsFoldableOpcode(opcode) &&
"Unhandled instruction opcode in FoldVectors");
std::vector<uint32_t> result;
for (uint32_t d = 0; d < num_dims; d++) {
std::vector<uint32_t> operand_values_for_one_dimension;
for (const auto& operand : operands) {
if (const analysis::VectorConstant* vector_operand =
operand->AsVectorConstant()) {
// Extract the raw value of the scalar component constants
// in 32-bit words here. The reason of not using FoldScalars() here
// is that we do not create temporary null constants as components
// when the vector operand is a NullConstant because Constant creation
// may need extra checks for the validity and that is not manageed in
// here.
if (const analysis::ScalarConstant* scalar_component =
vector_operand->GetComponents().at(d)->AsScalarConstant()) {
const auto& scalar_words = scalar_component->words();
assert(
scalar_words.size() == 1 &&
"Vector components with longer than 32-bit width are not allowed "
"in FoldVectors()");
operand_values_for_one_dimension.push_back(scalar_words.front());
} else if (operand->AsNullConstant()) {
operand_values_for_one_dimension.push_back(0u);
} else {
assert(false &&
"VectorConst should only has ScalarConst or NullConst as "
"components");
}
} else if (operand->AsNullConstant()) {
operand_values_for_one_dimension.push_back(0u);
} else {
assert(false &&
"FoldVectors() only accepts VectorConst or NullConst type of "
"constant");
}
}
result.push_back(OperateWords(opcode, operand_values_for_one_dimension));
}
return result;
}
bool IsFoldableOpcode(SpvOp opcode) {
// NOTE: Extend to more opcodes as new cases are handled in the folder
// functions.
switch (opcode) {
case SpvOp::SpvOpBitwiseAnd:
case SpvOp::SpvOpBitwiseOr:
case SpvOp::SpvOpBitwiseXor:
case SpvOp::SpvOpIAdd:
case SpvOp::SpvOpIEqual:
case SpvOp::SpvOpIMul:
case SpvOp::SpvOpINotEqual:
case SpvOp::SpvOpISub:
case SpvOp::SpvOpLogicalAnd:
case SpvOp::SpvOpLogicalEqual:
case SpvOp::SpvOpLogicalNot:
case SpvOp::SpvOpLogicalNotEqual:
case SpvOp::SpvOpLogicalOr:
case SpvOp::SpvOpNot:
case SpvOp::SpvOpSDiv:
case SpvOp::SpvOpSelect:
case SpvOp::SpvOpSGreaterThan:
case SpvOp::SpvOpSGreaterThanEqual:
case SpvOp::SpvOpShiftLeftLogical:
case SpvOp::SpvOpShiftRightArithmetic:
case SpvOp::SpvOpShiftRightLogical:
case SpvOp::SpvOpSLessThan:
case SpvOp::SpvOpSLessThanEqual:
case SpvOp::SpvOpSMod:
case SpvOp::SpvOpSNegate:
case SpvOp::SpvOpSRem:
case SpvOp::SpvOpUDiv:
case SpvOp::SpvOpUGreaterThan:
case SpvOp::SpvOpUGreaterThanEqual:
case SpvOp::SpvOpULessThan:
case SpvOp::SpvOpULessThanEqual:
case SpvOp::SpvOpUMod:
return true;
default:
return false;
}
}
bool IsFoldableConstant(const analysis::Constant* cst) {
// Currently supported constants are 32-bit values or null constants.
if (const analysis::ScalarConstant* scalar = cst->AsScalarConstant())
return scalar->words().size() == 1;
else
return cst->AsNullConstant() != nullptr;
}
ir::Instruction* FoldInstructionToConstant(
ir::Instruction* inst, std::function<uint32_t(uint32_t)> id_map) {
ir::IRContext* context = inst->context();
analysis::ConstantManager* const_mgr = context->get_constant_mgr();
if (!inst->IsFoldableByFoldScalar() &&
!GetConstantFoldingRules().HasFoldingRule(inst->opcode())) {
return nullptr;
}
// Collect the values of the constant parameters.
std::vector<const analysis::Constant*> constants;
bool missing_constants = false;
inst->ForEachInId([&constants, &missing_constants, const_mgr,
&id_map](uint32_t* op_id) {
uint32_t id = id_map(*op_id);
const analysis::Constant* const_op = const_mgr->FindDeclaredConstant(id);
if (!const_op) {
constants.push_back(nullptr);
missing_constants = true;
} else {
constants.push_back(const_op);
}
});
if (GetConstantFoldingRules().HasFoldingRule(inst->opcode())) {
const analysis::Constant* folded_const = nullptr;
for (auto rule :
GetConstantFoldingRules().GetRulesForOpcode(inst->opcode())) {
folded_const = rule(inst, constants);
if (folded_const != nullptr) {
ir::Instruction* const_inst =
const_mgr->GetDefiningInstruction(folded_const);
// May be a new instruction that needs to be analysed.
context->UpdateDefUse(const_inst);
return const_inst;
}
}
}
uint32_t result_val = 0;
bool successful = false;
// If all parameters are constant, fold the instruction to a constant.
if (!missing_constants && inst->IsFoldableByFoldScalar()) {
result_val = FoldScalars(inst->opcode(), constants);
successful = true;
}
if (!successful && inst->IsFoldableByFoldScalar()) {
successful = FoldIntegerOpToConstant(inst, id_map, &result_val);
}
if (successful) {
const analysis::Constant* result_const =
const_mgr->GetConstant(const_mgr->GetType(inst), {result_val});
return const_mgr->GetDefiningInstruction(result_const);
}
return nullptr;
}
bool IsFoldableType(ir::Instruction* type_inst) {
// Support 32-bit integers.
if (type_inst->opcode() == SpvOpTypeInt) {
return type_inst->GetSingleWordInOperand(0) == 32;
}
// Support booleans.
if (type_inst->opcode() == SpvOpTypeBool) {
return true;
}
// Nothing else yet.
return false;
}
bool FoldInstruction(ir::Instruction* inst) {
bool modified = false;
ir::Instruction* folded_inst(inst);
while (folded_inst->opcode() != SpvOpCopyObject &&
FoldInstructionInternal(&*folded_inst)) {
modified = true;
}
return modified;
}
} // namespace opt
} // namespace spvtools