| 1 |
// Copyright (c) 2018 Google LLC. |
| 2 |
// |
| 3 |
// Licensed under the Apache License, Version 2.0 (the "License"); |
| 4 |
// you may not use this file except in compliance with the License. |
| 5 |
// You may obtain a copy of the License at |
| 6 |
// |
| 7 |
// http://www.apache.org/licenses/LICENSE-2.0 |
| 8 |
// |
| 9 |
// Unless required by applicable law or agreed to in writing, software |
| 10 |
// distributed under the License is distributed on an "AS IS" BASIS, |
| 11 |
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. |
| 12 |
// See the License for the specific language governing permissions and |
| 13 |
// limitations under the License. |
| 14 |
|
| 15 |
#ifndef SOURCE_OPT_LOOP_PEELING_H_ |
| 16 |
#define SOURCE_OPT_LOOP_PEELING_H_ |
| 17 |
|
| 18 |
#include <algorithm> |
| 19 |
#include <limits> |
| 20 |
#include <memory> |
| 21 |
#include <tuple> |
| 22 |
#include <unordered_map> |
| 23 |
#include <unordered_set> |
| 24 |
#include <utility> |
| 25 |
#include <vector> |
| 26 |
|
| 27 |
#include "source/opt/ir_context.h" |
| 28 |
#include "source/opt/loop_descriptor.h" |
| 29 |
#include "source/opt/loop_utils.h" |
| 30 |
#include "source/opt/pass.h" |
| 31 |
#include "source/opt/scalar_analysis.h" |
| 32 |
|
| 33 |
namespace spvtools { |
| 34 |
namespace opt { |
| 35 |
|
| 36 |
// Utility class to perform the peeling of a given loop. |
| 37 |
// The loop peeling transformation make a certain amount of a loop iterations to |
| 38 |
// be executed either before (peel before) or after (peel after) the transformed |
| 39 |
// loop. |
| 40 |
// |
| 41 |
// For peeling cases the transformation does the following steps: |
| 42 |
// - It clones the loop and inserts the cloned loop before the original loop; |
| 43 |
// - It connects all iterating values of the cloned loop with the |
| 44 |
// corresponding original loop values so that the second loop starts with |
| 45 |
// the appropriate values. |
| 46 |
// - It inserts a new induction variable "i" is inserted into the cloned that |
| 47 |
// starts with the value 0 and increment by step of one. |
| 48 |
// |
| 49 |
// The last step is specific to each case: |
| 50 |
// - Peel before: the transformation is to peel the "N" first iterations. |
| 51 |
// The exit condition of the cloned loop is changed so that the loop |
| 52 |
// exits when "i < N" becomes false. The original loop is then protected to |
| 53 |
// only execute if there is any iteration left to do. |
| 54 |
// - Peel after: the transformation is to peel the "N" last iterations, |
| 55 |
// then the exit condition of the cloned loop is changed so that the loop |
| 56 |
// exits when "i + N < max_iteration" becomes false, where "max_iteration" |
| 57 |
// is the upper bound of the loop. The cloned loop is then protected to |
| 58 |
// only execute if there is any iteration left to do no covered by the |
| 59 |
// second. |
| 60 |
// |
| 61 |
// To be peelable: |
| 62 |
// - The loop must be in LCSSA form; |
| 63 |
// - The loop must not contain any breaks; |
| 64 |
// - The loop must not have any ambiguous iterators updates (see |
| 65 |
// "CanPeelLoop"). |
| 66 |
// The method "CanPeelLoop" checks that those constrained are met. |
| 67 |
class LoopPeeling { |
| 68 |
public: |
| 69 |
// LoopPeeling constructor. |
| 70 |
// |loop| is the loop to peel. |
| 71 |
// |loop_iteration_count| is the instruction holding the |loop| iteration |
| 72 |
// count, must be invariant for |loop| and must be of an int 32 type (signed |
| 73 |
// or unsigned). |
| 74 |
// |canonical_induction_variable| is an induction variable that can be used to |
| 75 |
// count the number of iterations, must be of the same type as |
| 76 |
// |loop_iteration_count| and start at 0 and increase by step of one at each |
| 77 |
// iteration. The value nullptr is interpreted as no suitable variable exists |
| 78 |
// and one will be created. |
| 79 |
LoopPeeling(Loop* loop, Instruction* loop_iteration_count, |
| 80 |
Instruction* canonical_induction_variable = nullptr) |
| 81 |
: context_(loop->GetContext()), |
| 82 |
loop_utils_(loop->GetContext(), loop), |
| 83 |
loop_(loop), |
| 84 |
loop_iteration_count_(!loop->IsInsideLoop(loop_iteration_count) |
| 85 |
? loop_iteration_count |
| 86 |
: nullptr), |
| 87 |
int_type_(nullptr), |
| 88 |
original_loop_canonical_induction_variable_( |
| 89 |
canonical_induction_variable), |
| 90 |
canonical_induction_variable_(nullptr) { |
| 91 |
if (loop_iteration_count_) { |
| 92 |
int_type_ = context_->get_type_mgr() |
| 93 |
->GetType(loop_iteration_count_->type_id()) |
| 94 |
->AsInteger(); |
| 95 |
if (canonical_induction_variable_) { |
| 96 |
assert(canonical_induction_variable_->type_id() == |
| 97 |
loop_iteration_count_->type_id() && |
| 98 |
"loop_iteration_count and canonical_induction_variable do not " |
| 99 |
"have the same type"); |
| 100 |
} |
| 101 |
} |
| 102 |
GetIteratingExitValues(); |
| 103 |
} |
| 104 |
|
| 105 |
// Returns true if the loop can be peeled. |
| 106 |
// To be peelable, all operation involved in the update of the loop iterators |
| 107 |
// must not dominates the exit condition. This restriction is a work around to |
| 108 |
// not miss compile code like: |
| 109 |
// |
| 110 |
// for (int i = 0; i + 1 < N; i++) {} |
| 111 |
// for (int i = 0; ++i < N; i++) {} |
| 112 |
// |
| 113 |
// The increment will happen before the test on the exit condition leading to |
| 114 |
// very look-a-like code. |
| 115 |
// |
| 116 |
// This restriction will not apply if a loop rotate is applied before (i.e. |
| 117 |
// becomes a do-while loop). |
| 118 |
bool CanPeelLoop() const { |
| 119 |
CFG& cfg = *context_->cfg(); |
| 120 |
|
| 121 |
if (!loop_iteration_count_) { |
| 122 |
return false; |
| 123 |
} |
| 124 |
if (!int_type_) { |
| 125 |
return false; |
| 126 |
} |
| 127 |
if (int_type_->width() != 32) { |
| 128 |
return false; |
| 129 |
} |
| 130 |
if (!loop_->IsLCSSA()) { |
| 131 |
return false; |
| 132 |
} |
| 133 |
if (!loop_->GetMergeBlock()) { |
| 134 |
return false; |
| 135 |
} |
| 136 |
if (cfg.preds(loop_->GetMergeBlock()->id()).size() != 1) { |
| 137 |
return false; |
| 138 |
} |
| 139 |
if (!IsConditionCheckSideEffectFree()) { |
| 140 |
return false; |
| 141 |
} |
| 142 |
|
| 143 |
return !std::any_of(exit_value_.cbegin(), exit_value_.cend(), |
| 144 |
[](std::pair<uint32_t, Instruction*> it) { |
| 145 |
return it.second == nullptr; |
| 146 |
}); |
| 147 |
} |
| 148 |
|
| 149 |
// Moves the execution of the |factor| first iterations of the loop into a |
| 150 |
// dedicated loop. |
| 151 |
void PeelBefore(uint32_t factor); |
| 152 |
|
| 153 |
// Moves the execution of the |factor| last iterations of the loop into a |
| 154 |
// dedicated loop. |
| 155 |
void PeelAfter(uint32_t factor); |
| 156 |
|
| 157 |
// Returns the cloned loop. |
| 158 |
Loop* GetClonedLoop() { return cloned_loop_; } |
| 159 |
// Returns the original loop. |
| 160 |
Loop* GetOriginalLoop() { return loop_; } |
| 161 |
|
| 162 |
private: |
| 163 |
IRContext* context_; |
| 164 |
LoopUtils loop_utils_; |
| 165 |
// The original loop. |
| 166 |
Loop* loop_; |
| 167 |
// The initial |loop_| upper bound. |
| 168 |
Instruction* loop_iteration_count_; |
| 169 |
// The int type to use for the canonical_induction_variable_. |
| 170 |
analysis::Integer* int_type_; |
| 171 |
// The cloned loop. |
| 172 |
Loop* cloned_loop_; |
| 173 |
// This is set to true when the exit and back-edge branch instruction is the |
| 174 |
// same. |
| 175 |
bool do_while_form_; |
| 176 |
// The canonical induction variable from the original loop if it exists. |
| 177 |
Instruction* original_loop_canonical_induction_variable_; |
| 178 |
// The canonical induction variable of the cloned loop. The induction variable |
| 179 |
// is initialized to 0 and incremented by step of 1. |
| 180 |
Instruction* canonical_induction_variable_; |
| 181 |
// Map between loop iterators and exit values. Loop iterators |
| 182 |
std::unordered_map<uint32_t, Instruction*> exit_value_; |
| 183 |
|
| 184 |
// Duplicate |loop_| and place the new loop before the cloned loop. Iterating |
| 185 |
// values from the cloned loop are then connected to the original loop as |
| 186 |
// initializer. |
| 187 |
void DuplicateAndConnectLoop(LoopUtils::LoopCloningResult* clone_results); |
| 188 |
|
| 189 |
// Insert the canonical induction variable into the first loop as a simplified |
| 190 |
// counter. |
| 191 |
void InsertCanonicalInductionVariable( |
| 192 |
LoopUtils::LoopCloningResult* clone_results); |
| 193 |
|
| 194 |
// Fixes the exit condition of the before loop. The function calls |
| 195 |
// |condition_builder| to get the condition to use in the conditional branch |
| 196 |
// of the loop exit. The loop will be exited if the condition evaluate to |
| 197 |
// true. |condition_builder| takes an Instruction* that represent the |
| 198 |
// insertion point. |
| 199 |
void FixExitCondition( |
| 200 |
const std::function<uint32_t(Instruction*)>& condition_builder); |
| 201 |
|
| 202 |
// Gathers all operations involved in the update of |iterator| into |
| 203 |
// |operations|. |
| 204 |
void GetIteratorUpdateOperations( |
| 205 |
const Loop* loop, Instruction* iterator, |
| 206 |
std::unordered_set<Instruction*>* operations); |
| 207 |
|
| 208 |
// Gathers exiting iterator values. The function builds a map between each |
| 209 |
// iterating value in the loop (a phi instruction in the loop header) and its |
| 210 |
// SSA value when it exit the loop. If no exit value can be accurately found, |
| 211 |
// it is map to nullptr (see comment on CanPeelLoop). |
| 212 |
void GetIteratingExitValues(); |
| 213 |
|
| 214 |
// Returns true if a for-loop has no instruction with effects before the |
| 215 |
// condition check. |
| 216 |
bool IsConditionCheckSideEffectFree() const; |
| 217 |
|
| 218 |
// Creates a new basic block and insert it between |bb| and the predecessor of |
| 219 |
// |bb|. |
| 220 |
BasicBlock* CreateBlockBefore(BasicBlock* bb); |
| 221 |
|
| 222 |
// Inserts code to only execute |loop| only if the given |condition| is true. |
| 223 |
// |if_merge| is a suitable basic block to be used by the if condition as |
| 224 |
// merge block. |
| 225 |
// The function returns the if block protecting the loop. |
| 226 |
BasicBlock* ProtectLoop(Loop* loop, Instruction* condition, |
| 227 |
BasicBlock* if_merge); |
| 228 |
}; |
| 229 |
|
| 230 |
// Implements a loop peeling optimization. |
| 231 |
// For each loop, the pass will try to peel it if there is conditions that |
| 232 |
// are true for the "N" first or last iterations of the loop. |
| 233 |
// To avoid code size explosion, too large loops will not be peeled. |
| 234 |
class LoopPeelingPass : public Pass { |
| 235 |
public: |
| 236 |
// Describes the peeling direction. |
| 237 |
enum class PeelDirection { |
| 238 |
kNone, // Cannot peel |
| 239 |
kBefore, // Can peel before |
| 240 |
kAfter // Can peel last |
| 241 |
}; |
| 242 |
|
| 243 |
// Holds some statistics about peeled function. |
| 244 |
struct LoopPeelingStats { |
| 245 |
std::vector<std::tuple<const Loop*, PeelDirection, uint32_t>> peeled_loops_; |
| 246 |
}; |
| 247 |
|
| 248 |
LoopPeelingPass(LoopPeelingStats* stats = nullptr) : stats_(stats) {} |
| 249 |
|
| 250 |
// Sets the loop peeling growth threshold. If the code size increase is above |
| 251 |
// |code_grow_threshold|, the loop will not be peeled. The code size is |
| 252 |
// measured in terms of SPIR-V instructions. |
| 253 |
static void SetLoopPeelingThreshold(size_t code_grow_threshold) { |
| 254 |
code_grow_threshold_ = code_grow_threshold; |
| 255 |
} |
| 256 |
|
| 257 |
// Returns the loop peeling code growth threshold. |
| 258 |
static size_t GetLoopPeelingThreshold() { return code_grow_threshold_; } |
| 259 |
|
| 260 |
const char* name() const override { return "loop-peeling"; } |
| 261 |
|
| 262 |
// Processes the given |module|. Returns Status::Failure if errors occur when |
| 263 |
// processing. Returns the corresponding Status::Success if processing is |
| 264 |
// succesful to indicate whether changes have been made to the modue. |
| 265 |
Pass::Status Process() override; |
| 266 |
|
| 267 |
private: |
| 268 |
// Describes the peeling direction. |
| 269 |
enum class CmpOperator { |
| 270 |
kLT, // less than |
| 271 |
kGT, // greater than |
| 272 |
kLE, // less than or equal |
| 273 |
kGE, // greater than or equal |
| 274 |
}; |
| 275 |
|
| 276 |
class LoopPeelingInfo { |
| 277 |
public: |
| 278 |
using Direction = std::pair<PeelDirection, uint32_t>; |
| 279 |
|
| 280 |
LoopPeelingInfo(Loop* loop, size_t loop_max_iterations, |
| 281 |
ScalarEvolutionAnalysis* scev_analysis) |
| 282 |
: context_(loop->GetContext()), |
| 283 |
loop_(loop), |
| 284 |
scev_analysis_(scev_analysis), |
| 285 |
loop_max_iterations_(loop_max_iterations) {} |
| 286 |
|
| 287 |
// Returns by how much and to which direction a loop should be peeled to |
| 288 |
// make the conditional branch of the basic block |bb| an unconditional |
| 289 |
// branch. If |bb|'s terminator is not a conditional branch or the condition |
| 290 |
// is not workable then it returns PeelDirection::kNone and a 0 factor. |
| 291 |
Direction GetPeelingInfo(BasicBlock* bb) const; |
| 292 |
|
| 293 |
private: |
| 294 |
// Returns the id of the loop invariant operand of the conditional |
| 295 |
// expression |condition|. It returns if no operand is invariant. |
| 296 |
uint32_t GetFirstLoopInvariantOperand(Instruction* condition) const; |
| 297 |
// Returns the id of the non loop invariant operand of the conditional |
| 298 |
// expression |condition|. It returns if all operands are invariant. |
| 299 |
uint32_t GetFirstNonLoopInvariantOperand(Instruction* condition) const; |
| 300 |
|
| 301 |
// Returns the value of |rec| at the first loop iteration. |
| 302 |
SExpression GetValueAtFirstIteration(SERecurrentNode* rec) const; |
| 303 |
// Returns the value of |rec| at the given |iteration|. |
| 304 |
SExpression GetValueAtIteration(SERecurrentNode* rec, |
| 305 |
int64_t iteration) const; |
| 306 |
// Returns the value of |rec| at the last loop iteration. |
| 307 |
SExpression GetValueAtLastIteration(SERecurrentNode* rec) const; |
| 308 |
|
| 309 |
bool EvalOperator(CmpOperator cmp_op, SExpression lhs, SExpression rhs, |
| 310 |
bool* result) const; |
| 311 |
|
| 312 |
Direction HandleEquality(SExpression lhs, SExpression rhs) const; |
| 313 |
Direction HandleInequality(CmpOperator cmp_op, SExpression lhs, |
| 314 |
SERecurrentNode* rhs) const; |
| 315 |
|
| 316 |
static Direction GetNoneDirection() { |
| 317 |
return Direction{LoopPeelingPass::PeelDirection::kNone, 0}; |
| 318 |
} |
| 319 |
IRContext* context_; |
| 320 |
Loop* loop_; |
| 321 |
ScalarEvolutionAnalysis* scev_analysis_; |
| 322 |
size_t loop_max_iterations_; |
| 323 |
}; |
| 324 |
// Peel profitable loops in |f|. |
| 325 |
bool ProcessFunction(Function* f); |
| 326 |
// Peel |loop| if profitable. |
| 327 |
std::pair<bool, Loop*> ProcessLoop(Loop* loop, CodeMetrics* loop_size); |
| 328 |
|
| 329 |
static size_t code_grow_threshold_; |
| 330 |
LoopPeelingStats* stats_; |
| 331 |
}; |
| 332 |
|
| 333 |
} // namespace opt |
| 334 |
} // namespace spvtools |
| 335 |
|
| 336 |
#endif // SOURCE_OPT_LOOP_PEELING_H_ |