GCC Middle and Back End API Reference
|
Data Structures | |
struct | loop_size |
Enumerations | |
enum | unroll_level { UL_SINGLE_ITER, UL_NO_GROWTH, UL_ALL } |
Variables | |
static vec< loop_p > | loops_to_unloop |
static vec< int > | loops_to_unloop_nunroll |
enum unroll_level |
@verbatim
Induction variable canonicalization and loop peeling. Copyright (C) 2004-2013 Free Software Foundation, Inc.
This file is part of GCC.
GCC is free software; you can redistribute it and/or modify it under the terms of the GNU General Public License as published by the Free Software Foundation; either version 3, or (at your option) any later version.
GCC is distributed in the hope that it will be useful, but WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for more details.
You should have received a copy of the GNU General Public License along with GCC; see the file COPYING3. If not see http://www.gnu.org/licenses/.
This pass detects the loops that iterate a constant number of times, adds a canonical induction variable (step -1, tested against 0) and replaces the exit test. This enables the less powerful rtl level analysis to use this information. This might spoil the code in some cases (by increasing register pressure). Note that in the case the new variable is not needed, ivopts will get rid of it, so it might only be a problem when there are no other linear induction variables. In that case the created optimization possibilities are likely to pay up. Additionally in case we detect that it is beneficial to unroll the loop completely, we do it right here to expose the optimization possibilities to the following passes.
Specifies types of loops that may be unrolled.
unsigned int canonicalize_induction_variables | ( | ) |
The main entry point of the pass. Adds canonical induction variables to the suitable loops.
Clean up the information about numbers of iterations, since brute force evaluation could reveal new information.
|
static |
Adds a canonical induction variable to LOOP if suitable. CREATE_IV is true if we may create a new iv. UL determines which loops we are allowed to completely unroll. If TRY_EVAL is true, we try to determine the number of iterations of a loop by direct evaluation. Returns true if cfg is changed.
If the loop has more than one exit, try checking all of them for # of iterations determinable through scev.
Finally if everything else fails, try brute force evaluation.
We work exceptionally hard here to estimate the bound by find_loop_niter_by_eval. Be sure to keep it for future.
Force re-computation of loop bounds so we can remove redundant exits.
Remove exits that are known to be never taken based on loop bound. Needs to be called after compilation of max_loop_iterations_int that populates the loop bounds.
|
static |
Return true if OP in STMT will be constant after peeling LOOP.
We can still fold accesses to constant arrays when index is known.
First make fast look if we see constant array inside.
If so, see if we understand all the indices.
Induction variables are constants.
References handled_component_p().
|
static |
Adds a canonical induction variable to LOOP iterating NITER times. EXIT is the exit edge whose condition is replaced.
Note that we do not need to worry about overflows, since type of niter is always unsigned and all comparisons are just for equality/nonequality -- i.e. everything works with a modulo arithmetics.
References dump_file, loop::num, and print_generic_expr().
|
static |
Estimate number of insns of completely unrolled loop. It is (NUNROLL + 1) * size of loop body with taking into account the fact that in last copy everything after exit conditional is dead and that some instructions will be eliminated after peeling. Loop body is likely going to simplify further, this is difficult to guess, we just decrease the result by 1/3.
References edge_def::flags, get_loop_exit_edges(), loop::latch, basic_block_def::preds, edge_def::src, and basic_block_def::succs.
|
static |
References execute(), gate_tree_complete_unroll_inner(), and tree_complete_unroll_inner().
Referenced by gate_tree_ssa_loop_ivcanon().
|
static |
Referenced by gate_tree_complete_unroll().
|
static |
References gate_tree_complete_unroll().
|
static |
Loop LOOP is known to not loop. See if there is an edge in the loop body that can be remove to make the loop to always exit and at the same time it does not make any code potentially executed during the last iteration dead. After complette unrolling we still may get rid of the conditional on the exit in the last copy even if we have no idea what it does. This is quite common case for loops of form int a[5]; for (i=0;i<b;i++) a[i]=0; Here we prove the loop to iterate 5 times but we do not know it from induction variable. For now we handle only simple case where there is exit condition just before the latch block and the latch block contains no statements with side effect that may otherwise terminate the execution of loop (such as by EH or by terminating the program or longjmp). In the general case we may want to cancel the paths leading to statements loop-niter identified as having undefined effect in the last iteration. The other cases are hopefully rare and will be cleaned up later.
We want only one predecestor of the loop.
Find the other edge than the loop exit leaving the conditoinal.
We only can handle conditionals.
We should never have conditionals in the loop latch.
Check that it leads to loop latch.
Verify that the code in loop latch does nothing that may end program execution without really reaching the exit. This may include non-pure/const function calls, EH statements, volatile ASMs etc.
gimple_opt_pass* make_pass_complete_unroll | ( | ) |
gimple_opt_pass* make_pass_complete_unrolli | ( | ) |
gimple_opt_pass* make_pass_iv_canon | ( | ) |
|
static |
Propagate constant SSA_NAMEs defined in basic block BB.
Look for degenerate PHI nodes with constant argument.
Look for assignments to SSA names with constant RHS.
|
static |
Propagate VAL into all uses of SSA_NAME.
|
static |
Remove all tests for exits that are known to be taken after LOOP was peeled NPEELED times. Put gcc_unreachable before every statement known to not be executed.
If statement is known to be undefined after peeling, turn it into unreachable (or trap when debugging experience is supposed to be good).
If we know the exit will be taken after peeling, update.
References dump_file, print_gimple_stmt(), and nb_iter_bound::stmt.
|
static |
Remove all exits that are known to be never taken because of the loop bound discovered.
Exit is pointless if it won't be taken before loop reaches upper bound.
Only when we know the actual number of iterations, not just a bound, we can remove the exit.
|
static |
Complete unrolling of loops.
|
static |
Complete unrolling of inner loops.
Referenced by gate_tree_complete_unroll().
|
static |
Computes an estimated number of insns in LOOP. EXIT (if non-NULL) is an exite edge that will be eliminated in all but last iteration of the loop. EDGE_TO_CANCEL (if non-NULL) is an non-exit edge eliminated in the last iteration of loop. Return results in SIZE, estimate benefits for complete unrolling exiting by EXIT. Stop estimating after UPPER_BOUND is met. Return true in this case.
Look for reasons why we might optimize this stmt away.
Exit conditional.
Sets of IV variables
Assignments of IV variables.
Conditionals.
|
static |
Canonical induction variable creation pass.
unsigned int tree_unroll_loops_completely | ( | ) |
Unroll LOOPS completely if they iterate just few times. Unless MAY_INCREASE_SIZE is true, perform the unrolling only if the size of the code does not increase.
Be sure to skip unlooped loops while procesing father_stack array.
We can not use TODO_update_ssa_no_phi because VOPS gets confused.
Propagate the constants within the new basic blocks.
This will take care of removing completely unrolled loops from the loop structures so we can continue unrolling now innermost loops.
Clean up the information about numbers of iterations, since complete unrolling might have invalidated it.
|
static |
Process loops from innermost to outer, stopping at the innermost loop we unrolled.
Process inner loops first.
If we changed an inner loop we cannot process outer loops in this iteration because SSA form is not up-to-date. Continue with siblings of outer loops instead.
Don't unroll #pragma omp simd loops until the vectorizer attempts to vectorize those.
Try to unroll this loop.
Unroll outermost loops only if asked to do so or they do not cause code growth.
If we'll continue unrolling, we need to propagate constants within the new basic blocks to fold away induction variable computations; otherwise, the size might blow up before the iteration is complete and the IR eventually cleaned up.
|
static |
Tries to unroll LOOP completely, i.e. NITER times. UL determines which loops we are allowed to unroll. EXIT is the exit of the loop that should be eliminated. MAXITER specfy bound on number of iterations, -1 if it is not known or too large for HOST_WIDE_INT. The location LOCUS corresponding to the loop is used when emitting a summary of the unroll to the dump file.
See if we proved number of iterations to be low constant. EXIT is an edge that will be removed in all but last iteration of the loop. EDGE_TO_CACNEL is an edge that will be removed from the last iteration of the unrolled sequence and is expected to make the final loop not rolling. If the number of execution of loop is determined by standard induction variable test, then EXIT and EDGE_TO_CANCEL are the two edges leaving from the iv test.
We do not know the number of iterations and thus we can not eliminate the EXIT edge.
See if we can improve our estimate by using recorded loop bounds.
Loop terminates before the IV variable test, so we can not remove it in the last iteration.
If the code is going to shrink, we don't need to be extra cautious on guessing if the unrolling is going to be profitable.
If there is IV variable that will become constant, we save one instruction in the loop prologue we do not account otherwise.
We unroll only inner loops, because we do not consider it profitable otheriwse. We still can cancel loopback edge of not rolling loop; this is always a good idea.
Outer loops tend to be less interesting candidates for complette unrolling unless we can do a lot of propagation into the inner loop body. For now we disable outer loop unrolling when the code would grow.
If there is call on a hot path through the loop, then there is most probably not much to optimize.
If there is pure/const call in the function, then we can still optimize the unrolled loop body if it contains some other interesting code than the calls and code storing or cumulating the return value.
One IV increment, one test, one ivtmp store and one useful stmt. That is about minimal loop doing pure call.
Complette unrolling is major win when control flow is removed and one big basic block is created. If the loop contains control flow the optimization may still be a win because of eliminating the loop overhead but it also may blow the branch predictor tables. Limit number of branches on the hot path through the peeled sequence.
Remove the conditional from the last copy of the loop.
Do not remove the path. Doing so may remove outer loop and confuse bookkeeping code in tree_unroll_loops_completelly.
Store the loop for later unlooping and exit removal.
|
static |
Cancel all fully unrolled loops by putting __builtin_unreachable on the latch edge. We do it after all unrolling since unlooping moves basic blocks across loop boundaries trashing loop closed SSA form as well as SCEV info needed to be intact during unrolling. IRRED_INVALIDATED is used to bookkeep if information about irreducible regions may become invalid as a result of the transformation. LOOP_CLOSED_SSA_INVALIDATED is used to bookkepp the case when we need to go into loop closed SSA form.
Unloop destroys the latch edge.
Create new basic block for the latch edge destination and wire it in.
References host_integerp(), HOST_WIDE_INT, loop::ninsns, edge_def::src, and tree_low_cst().
Stores loops that will be unlooped after we process whole loop tree.
|
static |