#include "config.h"#include "system.h"#include "coretypes.h"#include "tm.h"#include "tree.h"#include "tm_p.h"#include "basic-block.h"#include "gimple.h"#include "gimple-ssa.h"#include "tree-cfg.h"#include "tree-phinodes.h"#include "ssa-iterators.h"#include "tree-ssanames.h"#include "tree-ssa-loop-ivopts.h"#include "tree-ssa-loop-manip.h"#include "tree-ssa-loop-niter.h"#include "tree-ssa-loop.h"#include "tree-into-ssa.h"#include "tree-ssa.h"#include "dumpfile.h"#include "gimple-pretty-print.h"#include "cfgloop.h"#include "tree-pass.h"#include "tree-scalar-evolution.h"#include "params.h"#include "tree-inline.h"#include "langhooks.h"
Macros | |
| #define | PROB_UNROLLED_LOOP_ENTERED 90 |
Variables | |
| static bitmap_obstack | loop_renamer_obstack |
Macro Definition Documentation
| #define PROB_UNROLLED_LOOP_ENTERED 90 |
Unroll LOOP FACTOR times. DESC describes number of iterations of LOOP. EXIT is the exit of the loop to that DESC corresponds.
If N is number of iterations of the loop and MAY_BE_ZERO is the condition under that loop exits in the first iteration even if N != 0,
while (1) { x = phi (init, next);
pre; if (st) break; post; }
becomes (with possibly the exit conditions formulated a bit differently, avoiding the need to create a new iv):
if (MAY_BE_ZERO || N < FACTOR) goto rest;
do { x = phi (init, next);
pre; post; pre; post; ... pre; post; N -= FACTOR;
} while (N >= FACTOR);
rest: init' = phi (init, x);
while (1) { x = phi (init', next);
pre; if (st) break; post; }
Before the loop is unrolled, TRANSFORM is called for it (only for the unrolled loop, but not for its versioned copy). DATA is passed to TRANSFORM. Probability in % that the unrolled loop is entered. Just a guess.
Referenced by scale_dominated_blocks_in_loop().
Function Documentation
|
static |
Add a loop-closing PHI for VAR in basic block EXIT.
|
static |
Add exit phis for the names marked in NAMES_TO_RENAME. Exits of the loops are stored in EXITS. Sets of blocks where the ssa names are used are stored in USE_BLOCKS.
References BITMAP_ALLOC, bitmap_set_bit, edge_def::dest, FOR_EACH_VEC_ELT, get_loop_exit_edges(), basic_block_def::index, and loop::num.
|
static |
Add exit phis for VAR that is used in LIVEIN. Exits of the loops are stored in LOOP_EXITS.
References EXECUTE_IF_SET_IN_BITMAP, and ssa_name.
| bool can_unroll_loop_p | ( | struct loop * | loop, |
| unsigned | factor, | ||
| struct tree_niter_desc * | niter | ||
| ) |
Returns true if we can unroll LOOP FACTOR times. Number of iterations of the loop is returned in NITER.
Check whether unrolling is possible. We only want to unroll loops for that we are able to determine number of iterations. We also want to split the extra iterations of the loop from its end, therefore we require that the loop has precisely one exit.
Scalar evolutions analysis might have copy propagated
the abnormal ssa names into these expressions, hence
emitting the computations based on them during loop
unrolling might create overlapping life ranges for
them, and failures in out-of-ssa. And of course, we must be able to duplicate the loop.
The final loop should be small enough.
References affine_iv_d::base, boolean_false_node, boolean_true_node, tree_niter_desc::bound, tree_niter_desc::cmp, tree_niter_desc::control, fold_convert, gcc_assert, gcc_unreachable, lower_bound_in_type(), NULL_TREE, affine_iv_d::step, tree_int_cst_sign_bit(), TREE_TYPE, and upper_bound_in_type().
| tree canonicalize_loop_ivs | ( | ) |
Bases all the induction variables in LOOP on a single induction variable (unsigned with base 0 and step 1), whose final value is compared with *NIT. When the IV type precision has to be larger than *NIT type precision, *NIT is converted to the larger type, the conversion code is inserted before the loop, and *NIT is updated to the new definition. When BUMP_IN_LATCH is true, the induction variable is incremented in the loop latch, otherwise it is incremented in the loop header. Return the induction variable that was created.
Make the loop exit if the control condition is not satisfied.
|
static |
Checks invariants of loop closed ssa form in statement STMT in BB.
|
static |
Check invariants of the loop closed ssa form for the USE in BB.
|
static |
DEF_BB is a basic block containing a DEF that needs rewriting into loop-closed SSA form. USE_BLOCKS is the set of basic blocks containing uses of DEF that "escape" from the loop containing DEF_BB (i.e. blocks in USE_BLOCKS are dominated by DEF_BB but not in the loop father of DEF_B). ALL_EXITS[I] is the set of all basic blocks that exit loop I.
Compute the subset of LOOP_EXITS that exit the loop containing DEF_BB or one of its loop fathers, in which DEF is live. This set is returned in the bitmap LIVE_EXITS.
Instead of computing the complete livein set of the def, we use the loop nesting tree as a form of poor man's structure analysis. This greatly speeds up the analysis, which is important because this function may be called on all SSA names that need rewriting, one at a time.
Normally the work list size is bounded by the number of basic blocks in the largest loop. We don't know this number, but we can be fairly sure that it will be relatively small.
Iterate until the worklist is empty.
Pull a block off the worklist.
Make sure we have at least enough room in the work list
for all predecessors of this block. For each predecessor block.
We should have met DEF_BB along the way.
If we've reached DEF_LOOP, our train ends here.
Add PRED to the LIVEIN set. PRED_VISITED is true if
we had already added PRED to LIVEIN before. If we have visited PRED before, don't add it to the worklist.
If BB dominates PRED, then we're probably looking at a loop.
We're only interested in looking up in the dominance tree
because DEF_BB dominates all the uses.
|
static |
Copies phi node arguments for duplicated blocks. The index of the first duplicated block is FIRST_NEW_BLOCK.
| void create_iv | ( | tree | base, |
| tree | step, | ||
| tree | var, | ||
| struct loop * | loop, | ||
| gimple_stmt_iterator * | incr_pos, | ||
| bool | after, | ||
| tree * | var_before, | ||
| tree * | var_after | ||
| ) |
Creates an induction variable with value BASE + STEP * iteration in LOOP. It is expected that neither BASE nor STEP are shared with other expressions (unless the sharing rules allow this). Use VAR as a base var_decl for it (if NULL, a new temporary will be created). The increment will occur at INCR_POS (after it if AFTER is true, before it otherwise). INCR_POS and AFTER can be computed using standard_iv_increment_position. The ssa versions of the variable before and after increment will be stored in VAR_BEFORE and VAR_AFTER (unless they are NULL).
For easier readability of the created code, produce MINUS_EXPRs when suitable.
Gimplify the step if necessary. We put the computations in front of the loop (i.e. the step should be loop invariant).
References add_phi_arg(), convert_to_ptrofftype, create_phi_node(), fold_build1, force_gimple_operand(), gimple_build_assign_with_ops(), gsi_insert_after(), gsi_insert_before(), gsi_insert_seq_on_edge_immediate(), GSI_NEW_STMT, loop::header, loop_latch_edge(), loop_preheader_edge(), make_ssa_name(), make_temp_ssa_name(), mark_addressable(), may_negate_without_overflow_p(), NULL, NULL_TREE, POINTER_TYPE_P, TREE_CODE, tree_expr_nonnegative_warnv_p(), tree_int_cst_lt(), TREE_OPERAND, TREE_TYPE, TYPE_UNSIGNED, and UNKNOWN_LOCATION.
|
static |
Determines the conditions that control execution of LOOP unrolled FACTOR times. DESC is number of iterations of LOOP. ENTER_COND is set to condition that must be true if the main loop can be entered. EXIT_BASE, EXIT_STEP, EXIT_CMP and EXIT_BOUND are set to values describing how the exit from the unrolled loop should be controlled.
For pointers, do the arithmetics in the type of step.
We only need to be correct when we answer question "Do at least FACTOR more iterations remain?" in the unrolled loop. Thus, transforming BASE + STEP * i <> BOUND to BASE + STEP * i < BOUND is ok.
The main body of the loop may be entered iff:
1) desc->may_be_zero is false.
2) it is possible to check that there are at least FACTOR iterations
of the loop, i.e., BOUND - step * FACTOR does not overflow.
3) # of iterations is at least FACTOR cond now may be a gimple comparison, which would be OK, but also any other gimple rhs (say a && b). In this case we need to force it to operand.
Referenced by scale_dominated_blocks_in_loop().
|
staticread |
Return the innermost superloop LOOP of USE_LOOP that is a superloop of both DEF_LOOP and USE_LOOP.
|
static |
Marks names that are used outside of the loop they are defined in for rewrite. Records the set of blocks in that the ssa names are defined to USE_BLOCKS. If CHANGED_BBS is not NULL, scan only blocks in this set.
References BITMAP_ALLOC, bitmap_empty_p(), bitmap_obstack_initialize(), cfun, LOOP_CLOSED_SSA, loops_state_set(), num_ssa_names, number_of_loops(), and update_ssa().
|
static |
Marks names that are used in BB and outside of the loop they are defined in for rewrite. Records the set of blocks in that the ssa names are defined to USE_BLOCKS. Record the SSA names that will need exit PHIs in NEED_PHIS.
|
static |
For uses in STMT, mark names that are used outside of the loop they are defined to rewrite. Record the set of blocks in that the ssa names are defined to USE_BLOCKS and the ssa names themselves to NEED_PHIS.
References find_uses_to_rename_use(), gimple_phi_result(), gsi_stmt(), PHI_ARG_DEF_FROM_EDGE, and virtual_operand_p().
|
static |
For USE in BB, if it is used outside of the loop it is defined in, mark it for rewrite. Record basic block BB where it is used to USE_BLOCKS. Record the ssa name index to NEED_PHIS bitmap.
If the definition is not inside a loop, it is not interesting.
If the use is not outside of the loop it is defined in, it is not interesting.
If we're seeing VER for the first time, we still have to allocate a bitmap for its uses.
Referenced by find_uses_to_rename_stmt().
|
static |
Fill the array of bitmaps LOOP_EXITS with all loop exit edge targets.
References gimple_bb(), basic_block_def::loop_father, loop_outer(), SSA_NAME_DEF_STMT, SSA_NAME_VERSION, and TREE_CODE.
| bool gimple_duplicate_loop_to_header_edge | ( | struct loop * | loop, |
| edge | e, | ||
| unsigned int | ndupl, | ||
| sbitmap | wont_exit, | ||
| edge | orig, | ||
| vec< edge > * | to_remove, | ||
| int | flags | ||
| ) |
The same as cfgloopmanip.c:duplicate_loop_to_header_edge, but also updates the PHI nodes at start of the copied region. In order to achieve this, only loops whose exits all lead to the same location are handled.
Notice that we do not completely update the SSA web after duplication. The caller is responsible for calling update_ssa after the loop has been duplicated.
Readd the removed phi args for e.
Copy the phi node arguments.
| basic_block ip_end_pos | ( | ) |
Returns the basic block in that statements should be emitted for induction variables incremented at the end of the LOOP.
References gsi_last_bb(), ip_normal_pos(), last_stmt(), and loop::latch.
| basic_block ip_normal_pos | ( | ) |
Returns the basic block in that statements should be emitted for induction variables incremented just before exit condition of a LOOP.
References gsi_last_bb().
Referenced by ip_end_pos().
|
static |
Rewrite all the phi nodes of LOOP in function of the main induction variable MAIN_IV.
| void rewrite_into_loop_closed_ssa | ( | ) |
Rewrites the program into a loop closed ssa form – i.e. inserts extra phi nodes to ensure that no variable is used outside the loop it is defined in.
This strengthening of the basic ssa form has several advantages:
1) Updating it during unrolling/peeling/versioning is trivial, since we do not need to care about the uses outside of the loop. The same applies to virtual operands which are also rewritten into loop closed SSA form. Note that virtual operands are always live until function exit. 2) The behavior of all uses of an induction variable is the same. Without this, you need to distinguish the case when the variable is used outside of the loop it is defined in, for example
for (i = 0; i < 100; i++) { for (j = 0; j < 100; j++) { k = i + j; use1 (k); } use2 (k); }
Looking from the outer loop with the normal SSA form, the first use of k is not well-behaved, while the second one is an induction variable with base 99 and step 1.
If CHANGED_BBS is not NULL, we look for uses outside loops only in the basic blocks in this set.
UPDATE_FLAG is used in the call to update_ssa. See TODO_update_ssa* for documentation.
If the pass has caused the SSA form to be out-of-date, update it now.
Uses of names to rename. We don't have to initialize this array, because we know that we will only have entries for the SSA names in NAMES_TO_RENAME.
Find the uses outside loops.
An array of bitmaps where LOOP_EXITS[I] is the set of basic blocks
that are the destination of an edge exiting loop number I. Add the PHI nodes on exits of the loops for the names we need to
rewrite. Fix up all the names found to be used outside their original
loops.
|
static |
Rewrite the phi node at position PSI in function of the main induction variable MAIN_IV and insert the generated code at GSI.
|
static |
Scales the frequencies of all basic blocks in LOOP that are strictly dominated by BB by NUM/DEN.
References determine_exit_conditions(), expected_loop_iterations(), edge_def::flags, gcc_assert, integer_nonzerop(), loop_preheader_edge(), loop_version(), loop::next, NULL, PROB_UNROLLED_LOOP_ENTERED, REG_BR_PROB_BASE, TODO_update_ssa, update_ssa(), and vNULL.
| basic_block split_loop_exit_edge | ( | ) |
Split loop exit edge EXIT. The things are a bit complicated by a need to preserve the loop closed ssa form. The newly created block is returned.
If the argument of the PHI node is a constant, we do not need to keep it inside loop.
Otherwise create an auxiliary phi node that will copy the value
of the SSA name out of the loop.
References loop::latch.
| void standard_iv_increment_position | ( | struct loop * | loop, |
| gimple_stmt_iterator * | bsi, | ||
| bool * | insert_after | ||
| ) |
Stores the standard position for induction variable increment in LOOP (just before the exit condition if it is available and latch block is empty, end of the latch block otherwise) to BSI. INSERT_AFTER is set to true if the increment should be inserted after *BSI.
References cfun, LOOP_CLOSED_SSA, LOOPS_HAVE_PREHEADERS, LOOPS_HAVE_SIMPLE_LATCHES, loops_state_satisfies_p(), and need_ssa_update_p().
| void tree_transform_and_unroll_loop | ( | struct loop * | loop, |
| unsigned | factor, | ||
| edge | exit, | ||
| struct tree_niter_desc * | desc, | ||
| transform_callback | transform, | ||
| void * | data | ||
| ) |
Let us assume that the unrolled loop is quite likely to be entered.
The values for scales should keep profile consistent, and somewhat close to correct. TODO: The current value of SCALE_REST makes it appear that the loop that is created by splitting the remaining iterations of the unrolled loop is executed the same number of times as the original loop, and with the same frequencies, which is obviously wrong. This does not appear to cause problems, so we do not bother with fixing it for now. To make the profile correct, we would need to change the probability of the exit edge of the loop, and recompute the distribution of frequencies in its body because of this change (scale the frequencies of blocks before and after the exit by appropriate factors).
Determine the probability of the exit edge of the unrolled loop.
Without profile feedback, loops for that we do not know a better estimate are assumed to roll 10 times. When we unroll such loop, it appears to roll too little, and it may even seem to be cold. To avoid this, we ensure that the created loop appears to roll at least 5 times (but at most as many times as before unrolling).
Prepare the cfg and update the phi nodes. Move the loop exit to the loop latch (and make its condition dummy, for the moment).
Since the exit edge will be removed, the frequency of all the blocks in the loop that are dominated by it must be scaled by 1 / (1 - exit->probability).
Set the probability of new exit to the same of the old one. Fix the frequency of the latch block, by scaling it back by 1 - exit->probability.
Prefer using original variable as a base for the new ssa name.
This is necessary for virtual ops, and useful in order to avoid
losing debug info for real ops. Transform the loop.
Unroll the loop and remove the exits in all iterations except for the last one.
Ensure that the frequencies in the loop match the new estimated number of iterations, and change the probability of the new exit edge.
Finally create the new counter for number of iterations and add the new exit instruction.
| void tree_unroll_loop | ( | struct loop * | loop, |
| unsigned | factor, | ||
| edge | exit, | ||
| struct tree_niter_desc * | desc | ||
| ) |
Wrapper over tree_transform_and_unroll_loop for case we do not want to transform the loop before unrolling. The meaning of the arguments is the same as for tree_transform_and_unroll_loop.
References TREE_TYPE, and TYPE_PRECISION.
Referenced by insn_to_prefetch_ratio_too_small_p().
| DEBUG_FUNCTION void verify_loop_closed_ssa | ( | ) |
Checks that invariants of the loop closed ssa form are preserved. Call verify_ssa when VERIFY_SSA_P is true.
References add_phi_arg(), create_phi_node(), duplicate_ssa_name(), gimple_phi_arg_location_from_edge(), gsi_end_p(), gsi_next(), gsi_start_phis(), gsi_stmt(), NULL, PHI_ARG_DEF_PTR_FROM_EDGE, SET_USE, single_succ_edge(), split_edge(), TREE_CODE, and USE_FROM_PTR.
Variable Documentation
|
static |
High-level loop manipulation functions. 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/. All bitmaps for rewriting into loop-closed SSA go on this obstack, so that we can free them all at once.
