Zdenek Dvorak - [patch] Allow analyzing loops without modifying cfg (original) (raw)

This is the mail archive of the gcc-patches@gcc.gnu.orgmailing list for the GCC project.

Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]
Other format: [Raw text]

Hello,

I was asked several times about the possibility to analyze loops without changing cfg. After about a hundred cleanups, this finally became possible to do in a satisfactory way. The patch makes flow_loops_find to only find loops (without modifying cfg), and splits disambiguation of loops with multiple latches to a separate function (called by loop_optimizer_init).

In addition to cleaning up the code and allowing usage in contexts where cfg cannot be modified, this split makes it possible to do better analysis of loops with multiple latches during their splitting. We use this to take structure of induction variables into account (on trees) when determining whether a latch edge corresponds to a subloop, or just a branch inside a single loop (loop-25.c testcase demonstrates several loops for that we now can disambiguate the latch edges correctly).

I think there are a few more places that could use AVOID_CFG_MODIFICATIONS for loop_optimizer_init, but that I did not dare to change. One of them is in fwprop, and IIRC I had seen some trivial loop analysis code that should be replaced by it in one of the backends (but I do not remember where).

Bootstrapped & regtested on i686, together with the fix for bug (http://gcc.gnu.org/ml/gcc-patches/2007-02/msg00138.html) that gets exposed by it. I will wait with commiting this patch till the fix gets approved, which should also give time to comment on this patch.

Zdenek

* doc/loop.texi: Document possibility not to perform disambiguation
of loops with multiple latches.
* cfgloopmanip.c (alp_enum_p): Removed.
(add_loop): Handle subloops.  Use get_loop_body_with_size.
(create_preheader): Do not allow ENTRY_BLOCK_PTR to be preheader.
* cfghooks.c (redirect_edge_and_branch_force): Set dominator for
the new forwarder block.
(make_forwarder_block): Only call new_bb_cbk if it is not NULL.
Handle the case latch is NULL.
* tree-ssa-dom.c (tree_ssa_dominator_optimize): Avoid cfg modifications
when marking loop exits.
* ifcvt.c (if_convert): Ditto.  Mark loop exits even if cfg cannot
be modified.
* loop-init.c (loop_optimizer_init): Do not modify cfg.  Call
disambiguate_loops_with_multiple_latches.
* tree-cfgcleanup.c (cleanup_tree_cfg_loop): Calculate dominators
before fix_loop_structure.
* cfgloop.c: Include pointer-set.h and output.h.
(canonicalize_loop_headers, HEADER_BLOCK, LATCH_EDGE,
update_latch_info, mfb_keep_just, mfb_keep_nonlatch): Removed.
(get_loop_latch_edges, find_subloop_latch_edge_by_profile,
find_subloop_latch_edge_by_ivs, find_subloop_latch_edge,
mfb_redirect_edges_in_set, form_subloop, merge_latch_edges,
disambiguate_multiple_latches, get_loop_body_with_size,
disambiguate_loops_with_multiple_latches): New functions.
(flow_loop_dump): Dump multiple latch edges.
(flow_loop_nodes_find): Handle loops with multiple latches.
(flow_loops_find): Ditto. Do not call canonicalize_loop_headers.
(glb_enum_p): Modified.
(get_loop_body): Use get_loop_body_with_size.
* cfgloop.h (LOOPS_HAVE_RECORDED_EXITS): New flag.
(AVOID_CFG_MODIFICATIONS): New constant.
(disambiguate_loops_with_multiple_latches, add_loop,
get_loop_body_with_size): Declare.
* Makefile.in (cfgloop.o): Add pointer-set.h and output.h.

* gcc.dg/tree-ssa/loop-25.c: New test.

Index: doc/loop.texi

*** doc/loop.texi (revision 121474) --- doc/loop.texi (working copy) *************** a single header, or if there is a branch *** 48,58 **** The representation of loops in GCC however allows only loops with a single latch. During loop analysis, headers of such loops are split and forwarder blocks are created in order to disambiguate their structures. ! A heuristic based on profile information is used to determine whether ! the latches correspond to sub-loops or to control flow in a single loop. ! This means that the analysis sometimes changes the CFG, and if you run ! it in the middle of an optimization pass, you must be able to deal with ! the new blocks.
Body of the loop is the set of blocks that are dominated by its header, and reachable from its latch against the direction of edges in CFG. The --- 48,62 ---- The representation of loops in GCC however allows only loops with a single latch. During loop analysis, headers of such loops are split and forwarder blocks are created in order to disambiguate their structures. ! Heuristic based on profile information and structure of the induction ! variables in the loops is used to determine whether the latches ! correspond to sub-loops or to control flow in a single loop. This means ! that the analysis sometimes changes the CFG, and if you run it in the ! middle of an optimization pass, you must be able to deal with the new ! blocks. You may avoid CFG changes by passing ! @code{LOOPS_MAY_HAVE_MULTIPLE_LATCHES} flag to the loop discovery, ! note however that most other loop manipulation functions will not work ! correctly for loops with multiple latch edges.
Body of the loop is the set of blocks that are dominated by its header, and reachable from its latch against the direction of edges in CFG. The *************** bitmask. These flags specify what other *** 96,101 **** --- 100,112 ---- structures should be calculated/enforced and preserved later:
@itemize + @item @code{LOOPS_MAY_HAVE_MULTIPLE_LATCHES}: If this flag is set, no + changes to CFG will be performed in the loop analysis, in particular, + loops with multiple latch edges will not be disambiguated. If a loop + has multiple latches, its latch block is set to NULL. Most of + the loop manipulation functions will not work for loops in this shape. + No other flags that require CFG changes can be passed to + loop_optimizer_init. @item @code{LOOPS_HAVE_PREHEADERS}: Forwarder blocks are created in such a way that each loop has only one entry edge, and additionally, the source block of this entry edge has only one successor. This creates a Index: cfgloopmanip.c

*** cfgloopmanip.c (revision 121474) --- cfgloopmanip.c (working copy) *************** static bool loop_delete_branch_edge (edg *** 39,49 **** static void remove_bbs (basic_block *, int); static bool rpe_enum_p (basic_block, void *); static int find_path (edge, basic_block **); - static bool alp_enum_p (basic_block, void *); static void fix_loop_placements (struct loop *, bool *); static bool fix_bb_placement (basic_block); static void fix_bb_placements (basic_block, bool *); - static void place_new_loop (struct loop *); static basic_block create_preheader (struct loop *, int); static void unloop (struct loop *, bool );
--- 39,47 ---- *************** remove_path (edge e) *** 397,435 **** return true; }
! / Predicate for enumeration in add_loop. */ ! static bool ! alp_enum_p (basic_block bb, void alp_header) { ! return bb != (basic_block) alp_header; }
/ Given LOOP structure with filled header and latch, find the body of the corresponding loop and add it to loops tree. Insert the LOOP as a son of outer. */
! static void add_loop (struct loop *loop, struct loop *outer) { basic_block bbs; int i, n;
/ Add it to loop structure. / place_new_loop (loop); flow_loop_tree_node_add (outer, loop);
/ Find its nodes. / ! bbs = XCNEWVEC (basic_block, n_basic_blocks); ! n = dfs_enumerate_from (loop->latch, 1, alp_enum_p, ! bbs, n_basic_blocks, loop->header);
for (i = 0; i < n; i++) { ! remove_bb_from_loops (bbs[i]); ! add_bb_to_loop (bbs[i], loop); } - remove_bb_from_loops (loop->header); - add_bb_to_loop (loop->header, loop);
free (bbs); } --- 395,447 ---- return true; }
! / Creates place for a new LOOP in loops structure. */ ! static void ! place_new_loop (struct loop loop) { ! loop->num = number_of_loops (); ! VEC_safe_push (loop_p, heap, current_loops->larray, loop); }
/ Given LOOP structure with filled header and latch, find the body of the corresponding loop and add it to loops tree. Insert the LOOP as a son of outer. */
! void add_loop (struct loop *loop, struct loop *outer) { basic_block *bbs; int i, n; + struct loop subloop;
/ Add it to loop structure. / place_new_loop (loop); flow_loop_tree_node_add (outer, loop);
/ Find its nodes. / ! bbs = XNEWVEC (basic_block, n_basic_blocks); ! n = get_loop_body_with_size (loop, bbs, n_basic_blocks);
for (i = 0; i < n; i++) { ! if (bbs[i]->loop_father == outer) ! { ! remove_bb_from_loops (bbs[i]); ! add_bb_to_loop (bbs[i], loop); ! continue; ! } ! ! loop->num_nodes++; ! ! / If we find a direct subloop of OUTER, move it to LOOP. */ ! subloop = bbs[i]->loop_father; ! if (subloop->outer == outer ! && subloop->header == bbs[i]) ! { ! flow_loop_tree_node_remove (subloop); ! flow_loop_tree_node_add (loop, subloop); ! } }
free (bbs); } *************** fix_loop_placements (struct loop loop, *** 633,646 **** } }
- / Creates place for a new LOOP in loops structure. */ - static void - place_new_loop (struct loop *loop) - { - loop->num = number_of_loops (); - VEC_safe_push (loop_p, heap, current_loops->larray, loop); - }

/* Copies copy of LOOP as subloop of TARGET loop, placing newly created loop into loops structure. */ struct loop * --- 645,650 ---- *************** create_preheader (struct loop loop, int *** 1156,1167 **** gcc_assert (nentry); if (nentry == 1) { ! / Get an edge that is different from the one from loop->latch ! to loop->header. / ! e = EDGE_PRED (loop->header, ! EDGE_PRED (loop->header, 0)->src == loop->latch);
! if (!(flags & CP_SIMPLE_PREHEADERS) || single_succ_p (e->src)) return NULL; }
--- 1160,1174 ---- gcc_assert (nentry); if (nentry == 1) { ! e = loop_preheader_edge (loop);
! if (/ We do not allow entry block to be the loop preheader, since we ! cannot emit code there. / ! e->src != ENTRY_BLOCK_PTR ! / If we want simple preheaders, also force the preheader to have ! just a single successor. */ ! && !((flags & CP_SIMPLE_PREHEADERS) ! && !single_succ_p (e->src))) return NULL; }
Index: cfghooks.c

*** cfghooks.c (revision 121474) --- cfghooks.c (working copy) *************** redirect_edge_and_branch_force (edge e, *** 336,341 **** --- 336,345 ---- rescan_loop_exit (e, false, true);
ret = cfg_hooks->redirect_edge_and_branch_force (e, dest); + if (ret != NULL + && dom_info_available_p (CDI_DOMINATORS)) + set_immediate_dominator (CDI_DOMINATORS, ret, src); + if (current_loops != NULL) { if (ret != NULL) *************** make_forwarder_block (basic_block bb, bo *** 682,688 **** fallthru->count = 0;
jump = redirect_edge_and_branch_force (e, bb); ! if (jump) new_bb_cbk (jump); }
--- 686,693 ---- fallthru->count = 0;
jump = redirect_edge_and_branch_force (e, bb); ! if (jump != NULL ! && new_bb_cbk != NULL) new_bb_cbk (jump); }
*************** make_forwarder_block (basic_block bb, bo *** 700,708 **** /* If we do not split a loop header, then both blocks belong to the same loop. In case we split loop header and do not redirect the latch edge to DUMMY, then DUMMY belongs to the outer loop, and ! BB becomes the new header. / loop = dummy->loop_father; if (loop->header == dummy && find_edge (loop->latch, dummy) == NULL) { remove_bb_from_loops (dummy); --- 705,716 ---- / If we do not split a loop header, then both blocks belong to the same loop. In case we split loop header and do not redirect the latch edge to DUMMY, then DUMMY belongs to the outer loop, and ! BB becomes the new header. If latch is not recorded for the loop, ! we leave this updating on the caller (this may only happen during ! loop analysis). */ loop = dummy->loop_father; if (loop->header == dummy + && loop->latch != NULL && find_edge (loop->latch, dummy) == NULL) { remove_bb_from_loops (dummy); Index: testsuite/gcc.dg/tree-ssa/loop-25.c

*** testsuite/gcc.dg/tree-ssa/loop-25.c (revision 0) --- testsuite/gcc.dg/tree-ssa/loop-25.c (revision 0) *************** *** 0 **** --- 1,128 ---- + /* { dg-do compile } / + / { dg-options "-O2 -fdump-tree-vrp1" } / + + int foo(void); + void bla(void); + void bar(void); + + void test1 (void) + { + unsigned i; + + / Only one loop should be found here. / + i = 0; + while (1) + { + i++; + if (i == 100) + break; + + if (foo ()) + bla (); + else + bar (); + } + } + + void test2 (void) + { + unsigned i, j; + + / Two loops should be found, in this case. / + i = j = 0; + while (1) + { + j++; + foo (); + if (j < 100) + continue; + + i++; + j = 0; + if (i == 100) + break; + } + } + + void test3 (void) + { + unsigned i, j, k; + + / Three loops. / + i = j = k = 0; + while (1) + { + j++; + foo (); + if (j < 100) + continue; + + j = 0; + k++; + if (k < 100) + continue; + + k = 0; + i++; + if (i == 100) + break; + } + } + + void test4 (void) + { + unsigned i, j, k; + + / Two loops with a nested subloop. / + i = j = 0; + while (1) + { + j++; + foo (); + for (k = 0; k < 100; k++) + foo (); + + if (j < 100) + continue; + + i++; + j = 0; + if (i == 100) + break; + } + } + + + void test5 (void) + { + unsigned i, j; + + / Both subloop and non-subloop back edges. / + i = j = 0; + while (1) + { + j++; + foo (); + if (j < 100) + continue; + j = 0; + + i++; + if (i == 100) + break; + + if (foo ()) + bla (); + else + bar (); + } + } + + / { dg-final { scan-tree-dump-times "Disambiguating loop" 5 "vrp1" } } / + / { dg-final { scan-tree-dump-times "Found latch edge" 5 "vrp1" } } / + / { dg-final { scan-tree-dump-times "Merged latch edges" 2 "vrp1" } } / + / { dg-final { scan-tree-dump-times "4 loops found" 2 "vrp1" } } / + / { dg-final { scan-tree-dump-times "3 loops found" 2 "vrp1" } } / + / { dg-final { scan-tree-dump-times "2 loops found" 1 "vrp1" } } / + + / { dg-final { cleanup-tree-dump "vrp1" } } */ Index: tree-ssa-dom.c

*** tree-ssa-dom.c (revision 121474) --- tree-ssa-dom.c (working copy) *************** tree_ssa_dominator_optimize (void) *** 279,285 **** /* We need to know which edges exit loops so that we can aggressively thread through loop headers to an exit edge. / ! loop_optimizer_init (0); if (current_loops) { mark_loop_exit_edges (); --- 279,285 ---- / We need to know which edges exit loops so that we can aggressively thread through loop headers to an exit edge. */ ! loop_optimizer_init (AVOID_CFG_MODIFICATIONS); if (current_loops) { mark_loop_exit_edges (); Index: loop-init.c

*** loop-init.c (revision 121474) --- loop-init.c (working copy) *************** Software Foundation, 51 Franklin Street, *** 40,61 **** void loop_optimizer_init (unsigned flags) { - edge e; - edge_iterator ei; struct loops loops;
gcc_assert (!current_loops); loops = XCNEW (struct loops);
- / Avoid annoying special cases of edges going to exit - block. */

--- 40,50 ---- *************** loop_optimizer_init (unsigned flags) *** 69,74 **** --- 58,76 ---- return; }
+ if (flags & LOOPS_MAY_HAVE_MULTIPLE_LATCHES) + { + /* If the loops may have multiple latches, we cannot canonicalize + them further (and most of the loop manipulation functions will + not work). However, we avoid modifying cfg, which some + passes may want. / + gcc_assert ((flags & ~(LOOPS_MAY_HAVE_MULTIPLE_LATCHES + | LOOPS_HAVE_RECORDED_EXITS)) == 0); + current_loops->state = LOOPS_MAY_HAVE_MULTIPLE_LATCHES; + } + else + disambiguate_loops_with_multiple_latches (); + / Create pre-headers. */ if (flags & LOOPS_HAVE_PREHEADERS) create_preheaders (CP_SIMPLE_PREHEADERS); Index: ifcvt.c

*** ifcvt.c (revision 121474) --- ifcvt.c (working copy) *************** if_convert (int x_life_data_ok) *** 3908,3925 **** num_true_changes = 0; life_data_ok = (x_life_data_ok != 0);
! if ((! targetm.cannot_modify_jumps_p ()) ! && (!flag_reorder_blocks_and_partition || !no_new_pseudos ! || !targetm.have_named_sections)) { ! loop_optimizer_init (0); ! if (current_loops) ! { ! mark_loop_exit_edges (); ! loop_optimizer_finalize (); ! } ! free_dominance_info (CDI_DOMINATORS); }
/* Compute postdominators if we think we'll use them. / if (HAVE_conditional_execution || life_data_ok) --- 3908,3920 ---- num_true_changes = 0; life_data_ok = (x_life_data_ok != 0);
! loop_optimizer_init (AVOID_CFG_MODIFICATIONS); ! if (current_loops) { ! mark_loop_exit_edges (); ! loop_optimizer_finalize (); } + free_dominance_info (CDI_DOMINATORS);
/ Compute postdominators if we think we'll use them. */ if (HAVE_conditional_execution || life_data_ok) Index: tree-cfgcleanup.c

*** tree-cfgcleanup.c (revision 121474) --- tree-cfgcleanup.c (working copy) *************** cleanup_tree_cfg_loop (void) *** 582,589 **** if (changed) { bitmap changed_bbs = BITMAP_ALLOC (NULL); - fix_loop_structure (changed_bbs); calculate_dominance_info (CDI_DOMINATORS);
/* This usually does nothing. But sometimes parts of cfg that originally were inside a loop get out of it due to edge removal (since they --- 582,589 ---- if (changed) { bitmap changed_bbs = BITMAP_ALLOC (NULL); calculate_dominance_info (CDI_DOMINATORS); + fix_loop_structure (changed_bbs);
/* This usually does nothing. But sometimes parts of cfg that originally were inside a loop get out of it due to edge removal (since they Index: cfgloop.c

*** cfgloop.c (revision 121474) --- cfgloop.c (working copy) *************** Software Foundation, 51 Franklin Street, *** 32,49 **** #include "flags.h" #include "tree.h" #include "tree-flow.h" ! ! /* Ratio of frequencies of edges so that one of more latch edges is ! considered to belong to inner loop with same header. / ! #define HEAVY_EDGE_RATIO 8 ! ! #define HEADER_BLOCK(B) ( (int ) (B)->aux) ! #define LATCH_EDGE(E) ((int *) (E)->aux)
static void flow_loops_cfg_dump (FILE *); static void establish_preds (struct loop *); - static void canonicalize_loop_headers (void); - static bool glb_enum_p (basic_block, void ); / Dump loop related CFG information. */
--- 32,42 ---- #include "flags.h" #include "tree.h" #include "tree-flow.h" ! #include "pointer-set.h" ! #include "output.h"
static void flow_loops_cfg_dump (FILE *); static void establish_preds (struct loop ); / Dump loop related CFG information. */
*************** superloop_at_depth (struct loop loop, u *** 90,95 **** --- 83,106 ---- return loop->pred[depth]; }
+ / Returns the list of the latch edges of LOOP. */ + + static VEC (edge, heap) * + get_loop_latch_edges (const struct loop *loop) + { + edge_iterator ei; + edge e; + VEC (edge, heap) ret = NULL; + + FOR_EACH_EDGE (e, ei, loop->header->preds) + { + if (dominated_by_p (CDI_DOMINATORS, e->src, loop->header)) + VEC_safe_push (edge, heap, ret, e); + } + + return ret; + } + / Dump the loop information specified by LOOP to the stream FILE using auxiliary dump callback function LOOP_DUMP_AUX if non null. */
*************** flow_loop_dump (const struct loop *loop, *** 100,113 **** { basic_block *bbs; unsigned i;
if (! loop || ! loop->header) return;
fprintf (file, ";;\n;; Loop %d\n", loop->num);
! fprintf (file, ";; header %d, latch %d\n", ! loop->header->index, loop->latch->index); fprintf (file, ";; depth %d, outer %ld\n", loop->depth, (long) (loop->outer ? loop->outer->num : -1));
--- 111,137 ---- { basic_block *bbs; unsigned i; + VEC (edge, heap) *latches; + edge e;
if (! loop || ! loop->header) return;
fprintf (file, ";;\n;; Loop %d\n", loop->num);
! fprintf (file, ";; header %d, ", loop->header->index); ! if (loop->latch) ! fprintf (file, "latch %d\n", loop->latch->index); ! else ! { ! fprintf (file, "multiple latches:"); ! latches = get_loop_latch_edges (loop); ! for (i = 0; VEC_iterate (edge, latches, i, e); i++) ! fprintf (file, " %d", e->src->index); ! VEC_free (edge, heap, latches); ! fprintf (file, "\n"); ! } ! fprintf (file, ";; depth %d, outer %ld\n", loop->depth, (long) (loop->outer ? loop->outer->num : -1));
*************** flow_loops_free (struct loops *loops) *** 198,243 **** int flow_loop_nodes_find (basic_block header, struct loop *loop) { ! basic_block *stack; ! int sp; int num_nodes = 1;
header->loop_father = loop; header->loop_depth = loop->depth;
! if (loop->latch->loop_father != loop) { ! stack = XNEWVEC (basic_block, n_basic_blocks); ! sp = 0; num_nodes++; ! stack[sp++] = loop->latch; ! loop->latch->loop_father = loop; ! loop->latch->loop_depth = loop->depth;
! while (sp) { basic_block node; edge e; edge_iterator ei;
! node = stack[--sp];
FOR_EACH_EDGE (e, ei, node->preds) { basic_block ancestor = e->src;
! if (ancestor != ENTRY_BLOCK_PTR ! && ancestor->loop_father != loop) { ancestor->loop_father = loop; ancestor->loop_depth = loop->depth; num_nodes++; ! stack[sp++] = ancestor; } } } - free (stack); } return num_nodes; }
--- 222,270 ---- int flow_loop_nodes_find (basic_block header, struct loop *loop) { ! VEC (basic_block, heap) stack = NULL; int num_nodes = 1; + edge latch; + edge_iterator latch_ei;
header->loop_father = loop; header->loop_depth = loop->depth;
! FOR_EACH_EDGE (latch, latch_ei, loop->header->preds) { ! if (latch->src->loop_father == loop ! || !dominated_by_p (CDI_DOMINATORS, latch->src, loop->header)) ! continue; ! num_nodes++; ! VEC_safe_push (basic_block, heap, stack, latch->src); ! latch->src->loop_father = loop; ! latch->src->loop_depth = loop->depth;
! while (!VEC_empty (basic_block, stack)) { basic_block node; edge e; edge_iterator ei;
! node = VEC_pop (basic_block, stack);
FOR_EACH_EDGE (e, ei, node->preds) { basic_block ancestor = e->src;
! if (ancestor->loop_father != loop) { ancestor->loop_father = loop; ancestor->loop_depth = loop->depth; num_nodes++; ! VEC_safe_push (basic_block, heap, stack, ancestor); } } } } + VEC_free (basic_block, heap, stack); + return num_nodes; }
*************** flow_loop_tree_node_remove (struct loop *** 299,454 **** loop->pred = NULL; }
- / A callback to update latch and header info for basic block JUMP created - by redirecting an edge. */

--- 326,331 ---- *************** flow_loops_find (struct loops loops) *** 494,502 **** / Ensure that the dominators are computed. / calculate_dominance_info (CDI_DOMINATORS);
- / Join loops with shared headers. */ - canonicalize_loop_headers ();

/* Count the number of loop headers.  This should be the
   same as the number of natural loops.  */
headers = sbitmap_alloc (last_basic_block);

--- 371,376 ---- *************** flow_loops_find (struct loops *loops) *** 506,512 **** FOR_EACH_BB (header) { edge_iterator ei; - int more_latches = 0;
header->loop_depth = 0;
--- 380,385 ---- *************** flow_loops_find (struct loops loops) *** 533,540 **** && dominated_by_p (CDI_DOMINATORS, latch, header)) { / Shared headers should be eliminated by now. */ - gcc_assert (!more_latches); - more_latches = 1; SET_BIT (headers, header->index); num_loops++; } --- 406,411 ---- *************** flow_loops_find (struct loops loops) *** 589,609 **** loop->num = num_loops; num_loops++;
! / Look for the latch for this header block. */ FOR_EACH_EDGE (e, ei, header->preds) { basic_block latch = e->src;
! if (latch != ENTRY_BLOCK_PTR ! && dominated_by_p (CDI_DOMINATORS, latch, header)) { loop->latch = latch; - break; } }

--- 460,485 ---- loop->num = num_loops; num_loops++;
! flow_loop_tree_node_add (header->loop_father, loop); ! loop->num_nodes = flow_loop_nodes_find (loop->header, loop); ! ! /* Look for the latch for this header block, if it has just a ! single one. / FOR_EACH_EDGE (e, ei, header->preds) { basic_block latch = e->src;
! if (flow_bb_inside_loop_p (loop, latch)) { + if (loop->latch != NULL) + { + / More than one latch edge. */ + loop->latch = NULL; + break; + } loop->latch = latch; } } }
free (dfs_order); *************** flow_loops_find (struct loops loops) *** 617,622 **** --- 493,756 ---- return VEC_length (loop_p, loops->larray); }
+ / Ratio of frequencies of edges so that one of more latch edges is + considered to belong to inner loop with same header. / + #define HEAVY_EDGE_RATIO 8 + + / Minimum number of samples for that we apply + find_subloop_latch_edge_by_profile heuristics. / + #define HEAVY_EDGE_MIN_SAMPLES 10 + + / If the profile info is available, finds an edge in LATCHES that much more + frequent than the remaining edges. Returns such an edge, or NULL if we do + not find one. + + We do not use guessed profile here, only the measured one. The guessed + profile is usually too flat and unreliable for this (and it is mostly based + on the loop structure of the program, so it does not make much sense to + derive the loop structure from it. */ +
+ static edge + find_subloop_latch_edge_by_profile (VEC (edge, heap) latches) + { + unsigned i; + edge e, me = NULL; + gcov_type mcount = 0, tcount = 0; + + for (i = 0; VEC_iterate (edge, latches, i, e); i++) + { + if (e->count > mcount) + { + me = e; + mcount = e->count; + } + tcount += e->count; + } + + if (tcount < HEAVY_EDGE_MIN_SAMPLES + || (tcount - mcount) * HEAVY_EDGE_RATIO > tcount) + return NULL; + + if (dump_file) + fprintf (dump_file, + "Found latch edge %d -> %d using profile information.\n", + me->src->index, me->dest->index); + return me; + } + + / Among LATCHES, guesses a latch edge of LOOP corresponding to subloop, based + on the structure of induction variables. Returns this edge, or NULL if we + do not find any. + + We are quite conservative, and look just for an obvious simple innermost + loop (which is the case where we would lose the most performance by not + disambiguating the loop). More precisely, we look for the following + situation: The source of the chosen latch edge dominates sources of all + the other latch edges. Additionally, the header does not contain a phi node + such that the argument from the chosen edge is equal to the argument from + another edge. */ + + static edge + find_subloop_latch_edge_by_ivs (struct loop *loop, VEC (edge, heap) latches) + { + edge e, latch = VEC_index (edge, latches, 0); + unsigned i; + tree phi, lop; + basic_block bb; + + / Find the candidate for the latch edge. / + for (i = 1; VEC_iterate (edge, latches, i, e); i++) + if (dominated_by_p (CDI_DOMINATORS, latch->src, e->src)) + latch = e; + + / Verify that it dominates all the latch edges. / + for (i = 0; VEC_iterate (edge, latches, i, e); i++) + if (!dominated_by_p (CDI_DOMINATORS, e->src, latch->src)) + return NULL; + + / Check for a phi node that would deny that this is a latch edge of + a subloop. / + for (phi = phi_nodes (loop->header); phi; phi = PHI_CHAIN (phi)) + { + lop = PHI_ARG_DEF_FROM_EDGE (phi, latch); + + / Ignore the values that are not changed inside the subloop. / + if (TREE_CODE (lop) != SSA_NAME + || SSA_NAME_DEF_STMT (lop) == phi) + continue; + bb = bb_for_stmt (SSA_NAME_DEF_STMT (lop)); + if (!bb || !flow_bb_inside_loop_p (loop, bb)) + continue; + + for (i = 0; VEC_iterate (edge, latches, i, e); i++) + if (e != latch + && PHI_ARG_DEF_FROM_EDGE (phi, e) == lop) + return NULL; + } + + if (dump_file) + fprintf (dump_file, + "Found latch edge %d -> %d using iv structure.\n", + latch->src->index, latch->dest->index); + return latch; + } + + / If we can determine that one of the several latch edges of LOOP behaves + as a latch edge of a separate subloop, returns this edge. Otherwise + returns NULL. */ + + static edge + find_subloop_latch_edge (struct loop *loop) + { + VEC (edge, heap) latches = get_loop_latch_edges (loop); + edge latch = NULL; + + if (VEC_length (edge, latches) > 1) + { + latch = find_subloop_latch_edge_by_profile (latches); + + if (!latch + / We consider ivs to guess the latch edge only in SSA. Perhaps we + should use cfghook for this, but it is hard to imagine it would + be useful elsewhere. / + && current_ir_type () == IR_GIMPLE) + latch = find_subloop_latch_edge_by_ivs (loop, latches); + } + + VEC_free (edge, heap, latches); + return latch; + } + + / Callback for make_forwarder_block. Returns true if the edge E is marked + in the set MFB_REIS_SET. */ + + static struct pointer_set_t mfb_reis_set; + static bool + mfb_redirect_edges_in_set (edge e) + { + return pointer_set_contains (mfb_reis_set, e); + } + + / Creates a subloop of LOOP with latch edge LATCH. */ + + static void + form_subloop (struct loop *loop, edge latch) + { + edge_iterator ei; + edge e, new_entry; + struct loop new_loop; +
+ mfb_reis_set = pointer_set_create (); + FOR_EACH_EDGE (e, ei, loop->header->preds) + { + if (e != latch) + pointer_set_insert (mfb_reis_set, e); + } + new_entry = make_forwarder_block (loop->header, mfb_redirect_edges_in_set, + NULL); + pointer_set_destroy (mfb_reis_set); + + loop->header = new_entry->src; + + / Find the blocks and subloops that belong to the new loop, and add it to + the appropriate place in the loop tree. / + new_loop = alloc_loop (); + new_loop->header = new_entry->dest; + new_loop->latch = latch->src; + add_loop (new_loop, loop); + } + + / Make all the latch edges of LOOP to go to a single forwarder block -- + a new latch of LOOP. */ + + static void + merge_latch_edges (struct loop *loop) + { + VEC (edge, heap) latches = get_loop_latch_edges (loop); + edge latch, e; + unsigned i; + + gcc_assert (VEC_length (edge, latches) > 0); + + if (VEC_length (edge, latches) == 1) + loop->latch = VEC_index (edge, latches, 0)->src; + else + { + if (dump_file) + fprintf (dump_file, "Merged latch edges of loop %d\n", loop->num); + + mfb_reis_set = pointer_set_create (); + for (i = 0; VEC_iterate (edge, latches, i, e); i++) + pointer_set_insert (mfb_reis_set, e); + latch = make_forwarder_block (loop->header, mfb_redirect_edges_in_set, + NULL); + pointer_set_destroy (mfb_reis_set); + + loop->header = latch->dest; + loop->latch = latch->src; + } + + VEC_free (edge, heap, latches); + } + + / LOOP may have several latch edges. Transform it into (possibly several) + loops with single latch edge. */ + + static void + disambiguate_multiple_latches (struct loop loop) + { + edge e; + + / We eliminate the mutiple latches by splitting the header to the forwarder + block F and the rest R, and redirecting the edges. There are two cases: + + 1) If there is a latch edge E that corresponds to a subloop (we guess + that based on profile -- if it is taken much more often than the + remaining edges; and on trees, using the information about induction + variables of the loops), we redirect E to R, all the remaining edges to + F, then rescan the loops and try again for the outer loop. + 2) If there is no such edge, we redirect all latch edges to F, and the + entry edges to R, thus making F the single latch of the loop. / + + if (dump_file) + fprintf (dump_file, "Disambiguating loop %d with multiple latches\n", + loop->num); + + / During latch merging, we may need to redirect the entry edges to a new + block. This would cause problems if the entry edge was the one from the + entry block. To avoid having to handle this case specially, split + such entry edge. / + e = find_edge (ENTRY_BLOCK_PTR, loop->header); + if (e) + split_edge (e); + + while (1) + { + e = find_subloop_latch_edge (loop); + if (!e) + break; + + form_subloop (loop, e); + } + + merge_latch_edges (loop); + } + + / Split loops with multiple latch edges. */ + + void + disambiguate_loops_with_multiple_latches (void) + { + loop_iterator li; + struct loop loop; + + FOR_EACH_LOOP (li, loop, 0) + { + if (!loop->latch) + disambiguate_multiple_latches (loop); + } + } + / Return nonzero if basic block BB belongs to LOOP. */ bool flow_bb_inside_loop_p (const struct loop loop, const basic_block bb) *************** flow_bb_inside_loop_p (const struct loop *** 630,673 **** return loop == source_loop || flow_loop_nested_p (loop, source_loop); }
! / Enumeration predicate for get_loop_body. */ static bool ! glb_enum_p (basic_block bb, void glb_header) { ! return bb != (basic_block) glb_header; }
/ Gets basic blocks of a LOOP. Header is the 0-th block, rest is in dfs order against direction of edges from latch. Specially, if header != latch, latch is the 1-st block. */ basic_block * get_loop_body (const struct loop loop) { ! basic_block tovisit, bb; unsigned tv = 0;
gcc_assert (loop->num_nodes);
! tovisit = XCNEWVEC (basic_block, loop->num_nodes); ! tovisit[tv++] = loop->header;
if (loop->latch == EXIT_BLOCK_PTR) { ! / There may be blocks unreachable from EXIT_BLOCK. / gcc_assert (loop->num_nodes == (unsigned) n_basic_blocks); FOR_EACH_BB (bb) ! tovisit[tv++] = bb; ! tovisit[tv++] = EXIT_BLOCK_PTR; ! } ! else if (loop->latch != loop->header) ! { ! tv = dfs_enumerate_from (loop->latch, 1, glb_enum_p, ! tovisit + 1, loop->num_nodes - 1, ! loop->header) + 1; }
gcc_assert (tv == loop->num_nodes); ! return tovisit; }
/ Fills dominance descendants inside LOOP of the basic block BB into --- 764,822 ---- return loop == source_loop || flow_loop_nested_p (loop, source_loop); }
! / Enumeration predicate for get_loop_body_with_size. */ static bool ! glb_enum_p (basic_block bb, void *glb_loop) ! { ! struct loop loop = glb_loop; ! return (bb != loop->header ! && dominated_by_p (CDI_DOMINATORS, bb, loop->header)); ! } ! ! / Gets basic blocks of a LOOP. Header is the 0-th block, rest is in dfs ! order against direction of edges from latch. Specially, if ! header != latch, latch is the 1-st block. LOOP cannot be the fake ! loop tree root, and its size must be at most MAX_SIZE. The blocks ! in the LOOP body are stored to BODY, and the size of the LOOP is ! returned. */ ! ! unsigned ! get_loop_body_with_size (const struct loop *loop, basic_block *body, ! unsigned max_size) { ! return dfs_enumerate_from (loop->header, 1, glb_enum_p, ! body, max_size, (void ) loop); }
/ Gets basic blocks of a LOOP. Header is the 0-th block, rest is in dfs order against direction of edges from latch. Specially, if header != latch, latch is the 1-st block. */ + basic_block * get_loop_body (const struct loop *loop) { ! basic_block body, bb; unsigned tv = 0;
gcc_assert (loop->num_nodes);
! body = XCNEWVEC (basic_block, loop->num_nodes);
if (loop->latch == EXIT_BLOCK_PTR) { ! / There may be blocks unreachable from EXIT_BLOCK, hence we need to ! special-case the fake loop that contains the whole function. / gcc_assert (loop->num_nodes == (unsigned) n_basic_blocks); + body[tv++] = loop->header; + body[tv++] = EXIT_BLOCK_PTR; FOR_EACH_BB (bb) ! body[tv++] = bb; } + else + tv = get_loop_body_with_size (loop, body, loop->num_nodes);
gcc_assert (tv == loop->num_nodes); ! return body; }
/ Fills dominance descendants inside LOOP of the basic block BB into Index: cfgloop.h

*** cfgloop.h (revision 121474) --- cfgloop.h (working copy) *************** enum *** 168,178 **** LOOPS_HAVE_PREHEADERS = 1, LOOPS_HAVE_SIMPLE_LATCHES = 2, LOOPS_HAVE_MARKED_IRREDUCIBLE_REGIONS = 4, ! LOOPS_HAVE_RECORDED_EXITS = 8 };
#define LOOPS_NORMAL (LOOPS_HAVE_PREHEADERS | LOOPS_HAVE_SIMPLE_LATCHES
| LOOPS_HAVE_MARKED_IRREDUCIBLE_REGIONS)
typedef struct loop *loop_p; DEF_VEC_P (loop_p); --- 168,180 ---- LOOPS_HAVE_PREHEADERS = 1, LOOPS_HAVE_SIMPLE_LATCHES = 2, LOOPS_HAVE_MARKED_IRREDUCIBLE_REGIONS = 4, ! LOOPS_HAVE_RECORDED_EXITS = 8, ! LOOPS_MAY_HAVE_MULTIPLE_LATCHES = 16 };
#define LOOPS_NORMAL (LOOPS_HAVE_PREHEADERS | LOOPS_HAVE_SIMPLE_LATCHES
| LOOPS_HAVE_MARKED_IRREDUCIBLE_REGIONS) + #define AVOID_CFG_MODIFICATIONS (LOOPS_MAY_HAVE_MULTIPLE_LATCHES)
typedef struct loop loop_p; DEF_VEC_P (loop_p); *************** struct loops *** 198,203 **** --- 200,206 ----
/ Loop recognition. */ extern int flow_loops_find (struct loops *); + extern void disambiguate_loops_with_multiple_latches (void); extern void flow_loops_free (struct loops *); extern void flow_loops_dump (FILE , void ()(const struct loop *, FILE , int), int); *************** void rescan_loop_exit (edge, bool, bool) *** 215,220 **** --- 218,224 ---- / Loop data structure manipulation/querying. */ extern void flow_loop_tree_node_add (struct loop *, struct loop *); extern void flow_loop_tree_node_remove (struct loop *); + extern void add_loop (struct loop *, struct loop *); extern bool flow_loop_nested_p (const struct loop *, const struct loop *); extern bool flow_bb_inside_loop_p (const struct loop *, const basic_block); extern struct loop * find_common_loop (struct loop *, struct loop ); *************** extern void mark_loop_exit_edges (void); *** 229,234 **** --- 233,240 ----
/ Loops & cfg manipulation. */ extern basic_block *get_loop_body (const struct loop *); + extern unsigned get_loop_body_with_size (const struct loop *, basic_block *, + unsigned); extern basic_block *get_loop_body_in_dom_order (const struct loop *); extern basic_block *get_loop_body_in_bfs_order (const struct loop *); extern VEC (edge, heap) *get_loop_exit_edges (const struct loop *); Index: Makefile.in

*** Makefile.in (revision 121474) --- Makefile.in (working copy) *************** cfgcleanup.o : cfgcleanup.c $(CONFIG_H) *** 2585,2591 **** (REGSH)(REGS_H) (REGSH​)(EMIT_RTL_H) (CFGLAYOUTH)tree−pass.h(CFGLAYOUT_H) tree-pass.h (CFGLAYOUTH​)tree−pass.h(CFGLOOP_H) $(EXPR_H) cfgloop.o : cfgloop.c (CONFIGH)(CONFIG_H) (CONFIGH​)(SYSTEM_H) (RTLH)coretypes.h(RTL_H) coretypes.h (RTLH​)coretypes.h(TM_H)
(BASICBLOCKH)hard−reg−set.h(BASIC_BLOCK_H) hard-reg-set.h (BASICB​LOCKH​)hard−reg−set.h(CFGLOOP_H) (FLAGSH)(FLAGS_H) (FLAGSH​)(FUNCTION_H)
! (OBSTACKH)toplev.h(OBSTACK_H) toplev.h (OBSTACKH​)toplev.h(TREE_FLOW_H) $(TREE_H) cfgloopanal.o : cfgloopanal.c (CONFIGH)(CONFIG_H) (CONFIGH​)(SYSTEM_H) $(RTL_H)
(BASICBLOCKH)hard−reg−set.h(BASIC_BLOCK_H) hard-reg-set.h (BASICB​LOCKH​)hard−reg−set.h(CFGLOOP_H) (EXPRH)coretypes.h(EXPR_H) coretypes.h (EXPRH​)coretypes.h(TM_H)
$(OBSTACK_H) output.h --- 2585,2591 ---- (REGSH)(REGS_H) (REGSH​)(EMIT_RTL_H) (CFGLAYOUTH)tree−pass.h(CFGLAYOUT_H) tree-pass.h (CFGLAYOUTH​)tree−pass.h(CFGLOOP_H) $(EXPR_H) cfgloop.o : cfgloop.c (CONFIGH)(CONFIG_H) (CONFIGH​)(SYSTEM_H) (RTLH)coretypes.h(RTL_H) coretypes.h (RTLH​)coretypes.h(TM_H)
(BASICBLOCKH)hard−reg−set.h(BASIC_BLOCK_H) hard-reg-set.h (BASICB​LOCKH​)hard−reg−set.h(CFGLOOP_H) (FLAGSH)(FLAGS_H) (FLAGSH​)(FUNCTION_H)
! (OBSTACKH)toplev.h(OBSTACK_H) toplev.h (OBSTACKH​)toplev.h(TREE_FLOW_H) $(TREE_H) pointer-set.h output.h cfgloopanal.o : cfgloopanal.c (CONFIGH)(CONFIG_H) (CONFIGH​)(SYSTEM_H) $(RTL_H)
(BASICBLOCKH)hard−reg−set.h(BASIC_BLOCK_H) hard-reg-set.h (BASICB​LOCKH​)hard−reg−set.h(CFGLOOP_H) (EXPRH)coretypes.h(EXPR_H) coretypes.h (EXPRH​)coretypes.h(TM_H)
$(OBSTACK_H) output.h

Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]