This is the mail archive of the
gcc-patches@gcc.gnu.org
mailing list for the GCC project.
Re: Repost: RFA [4.1]: improvement to if-conversion and cross-jumping (PR20070)
Jan Hubicka wrote:
Could you try how it behaves pre-reload then? (if you can send me
patch, I can give it a SPEC run too) I saw pretty common case where code
duplication prevented ifconvert, so it might be interesting to see how
things are affected by this.
I added some sanity checks and they trigger on i686-pc-linux-gnu. It
appears that
we don't have consistent register liveness information to start with.
I've started
to add some more debug code (not in the ChangeLog), but I can't finish
this today.
I have attached what I have so far.
2005-12-01 J"orn Rennecke <joern.rennecke@st.com>
PR rtl-optimization/20070 / part1
* basic-block.h (STRUCT_EQUIV_START, STRUCT_EQUIV_RERUN): Define.
(STRUCT_EQUIV_FINAL, STRUCT_EQUIV_MAX_LOCAL): Likewise.
(STRUCT_EQUIV_NEED_FULL_BLOCK): Likewise.
(struct struct_equiv_checkpoint, struct equiv_info): Likewise.
(struct_equiv_block_eq, rtx_equiv_p): Declare.
* cfgcleanup.c (merge_memattrs): Move from here into..
* struct-equiv.c: New file.
(IMPOSSIBLE_MOVE_FACTOR): Define.
(assign_reg_reg_set, rtx_equiv_p, struct_equiv_set): New functions.
(struct_equiv_dst_mem, struct_equiv_make_checkpoint): Likewise.
(struct_equiv_improve_checkpoint): Likewise.
(struct_equiv_restore_checkpoint, struct_equiv_death): Likewise.
(struct_equiv_block_eq): Likewise.
(find_dying_inputs, resolve_input_conflict): Likewise.
* cfgcleanup.c (condjump_equiv_p): New function, broken out of
outgoing_edges_match. Use rtx_equiv_p.
(outgoing_edges_match): Replace bb1/bb2 arguments with info argument.
Change caller. Use condjump_equiv_p.
(try_crossjump_to_edge): Use struct_equiv_block_eq
instead of flow_find_cross_jump.
(flow_find_cross_jump): Remove.
Back out this change:
2005-03-07 Kazu Hirata <kazu@cs.umass.edu>
* recog.c (verify_changes): Make it static.
* recog.h: Remove the corresponding prototype.
PR rtl-optimization/20070 / part2
* ifcvt.c (noce_try_complex_cmove): New function.
(noce_process_if_block): Call it.
For flag_expensive_optimizations, update cond_exec_changed_p on
success.
(rest_of_handle_if_conversion): For flag_expensive_optimizations,
provide if_convert with register lifeness info.
Index: flow.c
===================================================================
/usr/bin/diff -dup -F'^(' -u -L flow.c (revision 107723) -L flow.c (working copy) .svn/text-base/flow.c.svn-base flow.c
--- flow.c (revision 107723)
+++ flow.c (working copy)
@@ -4614,6 +4614,17 @@ struct tree_opt_pass pass_life =
static void
rest_of_handle_flow2 (void)
{
+ basic_block bb;
+
+ /* We got lots of pseudo registers sitting around in the live information.
+ This contains irrelevant insonsistencies that confuse the sanity checks,
+ constitute data that is unnecessarily duplicated as we call
+ find_many_sub_basic_blocks, and also slow down the convergence of the
+ global live update. */
+ if (optimize)
+ FOR_EACH_BB (bb)
+FIXME: implement using bitmap_elt_clear_from.
+ bitmap_clear_from (bb->global_live_at_start, FIRST_PSEUDO_REGISTER);
/* If optimizing, then go ahead and split insns now. */
#ifndef STACK_REGS
if (optimize > 0)
Index: global.c
===================================================================
/usr/bin/diff -dup -F'^(' -u -L global.c (revision 107723) -L global.c (working copy) .svn/text-base/global.c.svn-base global.c
--- global.c (revision 107723)
+++ global.c (working copy)
@@ -1741,6 +1741,18 @@ mark_elimination (int from, int to)
CLEAR_REGNO_REG_SET (r, from);
SET_REGNO_REG_SET (r, to);
}
+ /* Partial life updates can cause blocks with the same successors
+ to have different global_live_at_end regsets. Also, it's hard
+ to sanity check global_live_at_end aagainst the
+ global_live_at_end of the successor if they don't match. */
+#ifdef ENABLE_CHECKING
+ r = bb->il.rtl->global_live_at_end;
+ if (REGNO_REG_SET_P (r, from))
+ {
+ CLEAR_REGNO_REG_SET (r, from);
+ SET_REGNO_REG_SET (r, to);
+ }
+#endif
}
}
Index: ifcvt.c
===================================================================
/usr/bin/diff -dup -F'^(' -u -L ifcvt.c (revision 107723) -L ifcvt.c (working copy) .svn/text-base/ifcvt.c.svn-base ifcvt.c
--- ifcvt.c (revision 107723)
+++ ifcvt.c (working copy)
@@ -616,6 +616,8 @@ static rtx noce_get_alt_condition (struc
static int noce_try_minmax (struct noce_if_info *);
static int noce_try_abs (struct noce_if_info *);
static int noce_try_sign_mask (struct noce_if_info *);
+static bool noce_try_complex_cmove (struct ce_if_block *,
+ struct noce_if_info *);
/* Helper function for noce_try_store_flag*. */
@@ -2141,10 +2143,17 @@ noce_process_if_block (struct ce_if_bloc
/* Look for one of the potential sets. */
insn_a = first_active_insn (then_bb);
+ /* Set up the info block for noce_try_complex_cmove. */
+ if_info.test_bb = test_bb;
+ if_info.cond = cond;
+ if_info.jump = jump;
+ if_info.insn_a = insn_a;
if (! insn_a
|| insn_a != last_active_insn (then_bb, FALSE)
|| (set_a = single_set (insn_a)) == NULL_RTX)
- return FALSE;
+ return ((else_bb && HAVE_conditional_move)
+ ? noce_try_complex_cmove (ce_info, &if_info)
+ : FALSE);
x = SET_DEST (set_a);
a = SET_SRC (set_a);
@@ -2215,11 +2224,7 @@ noce_process_if_block (struct ce_if_bloc
if (! noce_operand_ok (a) || ! noce_operand_ok (b))
return FALSE;
- /* Set up the info block for our subroutines. */
- if_info.test_bb = test_bb;
- if_info.cond = cond;
- if_info.jump = jump;
- if_info.insn_a = insn_a;
+ /* Set up the rest of the info block for our subroutines. */
if_info.insn_b = insn_b;
if_info.x = x;
if_info.a = a;
@@ -2337,6 +2342,8 @@ noce_process_if_block (struct ce_if_bloc
/* Merge the blocks! */
merge_if_block (ce_info);
+ if (flag_expensive_optimizations)
+ cond_exec_changed_p = TRUE;
return TRUE;
}
@@ -2846,6 +2853,85 @@ find_if_block (struct ce_if_block * ce_i
return process_if_block (ce_info);
}
+/* Try to use a cmove where if end else blocks are structurally equivalent
+ blocks that differ only by an input register, and possible some local
+ registers. */
+static bool
+noce_try_complex_cmove (struct ce_if_block * ce_info,
+ struct noce_if_info *if_info)
+{
+ basic_block then_bb = ce_info->then_bb; /* THEN */
+ basic_block else_bb = ce_info->else_bb; /* ELSE or NULL */
+ rtx yi;
+ struct equiv_info info;
+ int i;
+ rtx temp;
+ rtx next;
+ int n_matched;
+
+ if (!life_data_ok)
+ return false;
+ info.x_block = else_bb;
+ info.y_block = then_bb;
+ info.input_cost = 0;
+ n_matched = struct_equiv_block_eq (STRUCT_EQUIV_START
+ | STRUCT_EQUIV_NEED_FULL_BLOCK, &info);
+ if (! n_matched)
+ return false;
+ while (info.need_rerun)
+ {
+ n_matched = struct_equiv_block_eq (STRUCT_EQUIV_RERUN
+ | STRUCT_EQUIV_NEED_FULL_BLOCK, &info);
+ if (! n_matched)
+ return false;
+ }
+ /* We want exactly one input.
+ ??? We might allow more inputs if BRANCH_COST is high, or no inputs
+ as a means of commoning basically identical code. */
+ if (info.cur.input_valid && ! info.dying_inputs)
+ {
+ temp = info.input_reg;
+ if_info->a = info.y_input;
+ if_info->b = info.x_input;
+ }
+ else
+ {
+ if (info.dying_inputs != 1)
+ return false;
+ for (i = info.cur.local_count-1; ! info.local_rvalue[i]; ) i--;
+ temp = info.x_local[i];
+ if_info->a = info.y_local[i];
+ if_info->b = info.x_local[i];
+ }
+ if_info->x = temp;
+ if (! noce_try_cmove (if_info))
+ return false;
+ if (dump_file)
+ {
+ fprintf (dump_file, "Using temp ");
+ print_simple_rtl (dump_file, temp);
+ fprintf (dump_file, " for inputs ");
+ print_simple_rtl (dump_file, if_info->a);
+ fprintf (dump_file, " and ");
+ print_simple_rtl (dump_file, if_info->b);
+ fprintf (dump_file, ".\n%d local registers.\n",
+ info.cur.local_count - info.dying_inputs);
+ }
+ struct_equiv_block_eq (STRUCT_EQUIV_FINAL | STRUCT_EQUIV_NEED_FULL_BLOCK,
+ &info);
+ if (flag_expensive_optimizations)
+ cond_exec_changed_p = TRUE;
+ for (yi = BB_HEAD (then_bb); yi != BB_END (then_bb); yi = next)
+ {
+ next = NEXT_INSN (yi);
+ if (INSN_P (yi))
+ next = delete_insn (yi);
+ }
+ delete_insn (if_info->jump);
+ merge_if_block (ce_info);
+ return true;
+}
+
/* Convert a branch over a trap, or a branch
to a trap, into a conditional trap. */
@@ -3651,7 +3737,24 @@ rest_of_handle_if_conversion (void)
dump_flow_info (dump_file);
cleanup_cfg (CLEANUP_EXPENSIVE);
reg_scan (get_insns (), max_reg_num ());
- if_convert (0);
+ if (flag_expensive_optimizations)
+ {
+ basic_block bb;
+
+ life_analysis (dump_file, PROP_REG_INFO);
+ verify_flow_info ();
+ if_convert (1);
+ count_or_remove_death_notes (NULL, 1);
+ FOR_EACH_BB (bb)
+ {
+ bb->il.rtl->global_live_at_start = 0;
+ bb->il.rtl->global_live_at_end = 0;
+ }
+ ENTRY_BLOCK_PTR->il.rtl->global_live_at_start = 0;
+ EXIT_BLOCK_PTR->il.rtl->global_live_at_start = 0;
+ }
+ else
+ if_convert (0);
}
timevar_push (TV_JUMP);
Index: recog.c
===================================================================
/usr/bin/diff -dup -F'^(' -u -L recog.c (revision 107723) -L recog.c (working copy) .svn/text-base/recog.c.svn-base recog.c
--- recog.c (revision 107723)
+++ recog.c (working copy)
@@ -339,7 +339,7 @@ num_changes_pending (void)
/* Tentatively apply the changes numbered NUM and up.
Return 1 if all changes are valid, zero otherwise. */
-static int
+int
verify_changes (int num)
{
int i;
Index: recog.h
===================================================================
/usr/bin/diff -dup -F'^(' -u -L recog.h (revision 107723) -L recog.h (working copy) .svn/text-base/recog.h.svn-base recog.h
--- recog.h (revision 107723)
+++ recog.h (working copy)
@@ -76,6 +76,7 @@ extern int asm_operand_ok (rtx, const ch
extern int validate_change (rtx, rtx *, rtx, int);
extern int validate_change_maybe_volatile (rtx, rtx *, rtx);
extern int insn_invalid_p (rtx);
+extern int verify_changes (int);
extern void confirm_change_group (void);
extern int apply_change_group (void);
extern int num_validated_changes (void);
Index: profile.c
===================================================================
/usr/bin/diff -dup -F'^(' -u -L profile.c (revision 107723) -L profile.c (working copy) .svn/text-base/profile.c.svn-base profile.c
--- profile.c (revision 107723)
+++ profile.c (working copy)
@@ -1348,6 +1348,7 @@ rest_of_handle_branch_prob (void)
{
struct loops loops;
+ verify_flow_info ();
/* Discover and record the loop depth at the head of each basic
block. The loop infrastructure does the real job for us. */
flow_loops_find (&loops);
@@ -1362,6 +1363,7 @@ rest_of_handle_branch_prob (void)
flow_loops_free (&loops);
free_dominance_info (CDI_DOMINATORS);
+ verify_flow_info ();
}
struct tree_opt_pass pass_branch_prob =
Index: cfgcleanup.c
===================================================================
/usr/bin/diff -dup -F'^(' -u -L cfgcleanup.c (revision 107723) -L cfgcleanup.c (working copy) .svn/text-base/cfgcleanup.c.svn-base cfgcleanup.c
--- cfgcleanup.c (revision 107723)
+++ cfgcleanup.c (working copy)
@@ -60,9 +60,7 @@ Software Foundation, 51 Franklin Street,
static bool first_pass;
static bool try_crossjump_to_edge (int, edge, edge);
static bool try_crossjump_bb (int, basic_block);
-static bool outgoing_edges_match (int, basic_block, basic_block);
-static int flow_find_cross_jump (int, basic_block, basic_block, rtx *, rtx *);
-static bool insns_match_p (int, rtx, rtx);
+static bool outgoing_edges_match (int *, struct equiv_info *);
static void merge_blocks_move_predecessor_nojumps (basic_block, basic_block);
static void merge_blocks_move_successor_nojumps (basic_block, basic_block);
@@ -74,7 +72,6 @@ static bool mark_effect (rtx, bitmap);
static void notice_new_block (basic_block);
static void update_forwarder_flag (basic_block);
static int mentions_nonequal_regs (rtx *, void *);
-static void merge_memattrs (rtx, rtx);
/* Set flags for newly created block. */
@@ -881,328 +878,144 @@ merge_blocks_move (edge e, basic_block b
return NULL;
}
-
-/* Removes the memory attributes of MEM expression
- if they are not equal. */
-
-void
-merge_memattrs (rtx x, rtx y)
+bool
+condjump_equiv_p (struct equiv_info *info)
{
- int i;
- int j;
- enum rtx_code code;
- const char *fmt;
-
- if (x == y)
- return;
- if (x == 0 || y == 0)
- return;
-
- code = GET_CODE (x);
-
- if (code != GET_CODE (y))
- return;
-
- if (GET_MODE (x) != GET_MODE (y))
- return;
-
- if (code == MEM && MEM_ATTRS (x) != MEM_ATTRS (y))
- {
- if (! MEM_ATTRS (x))
- MEM_ATTRS (y) = 0;
- else if (! MEM_ATTRS (y))
- MEM_ATTRS (x) = 0;
- else
- {
- rtx mem_size;
-
- if (MEM_ALIAS_SET (x) != MEM_ALIAS_SET (y))
- {
- set_mem_alias_set (x, 0);
- set_mem_alias_set (y, 0);
- }
-
- if (! mem_expr_equal_p (MEM_EXPR (x), MEM_EXPR (y)))
- {
- set_mem_expr (x, 0);
- set_mem_expr (y, 0);
- set_mem_offset (x, 0);
- set_mem_offset (y, 0);
- }
- else if (MEM_OFFSET (x) != MEM_OFFSET (y))
- {
- set_mem_offset (x, 0);
- set_mem_offset (y, 0);
- }
-
- if (!MEM_SIZE (x))
- mem_size = NULL_RTX;
- else if (!MEM_SIZE (y))
- mem_size = NULL_RTX;
- else
- mem_size = GEN_INT (MAX (INTVAL (MEM_SIZE (x)),
- INTVAL (MEM_SIZE (y))));
- set_mem_size (x, mem_size);
- set_mem_size (y, mem_size);
-
- set_mem_align (x, MIN (MEM_ALIGN (x), MEM_ALIGN (y)));
- set_mem_align (y, MEM_ALIGN (x));
- }
- }
-
- fmt = GET_RTX_FORMAT (code);
- for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
- {
- switch (fmt[i])
- {
- case 'E':
- /* Two vectors must have the same length. */
- if (XVECLEN (x, i) != XVECLEN (y, i))
- return;
-
- for (j = 0; j < XVECLEN (x, i); j++)
- merge_memattrs (XVECEXP (x, i, j), XVECEXP (y, i, j));
-
- break;
-
- case 'e':
- merge_memattrs (XEXP (x, i), XEXP (y, i));
- }
- }
- return;
-}
-
+ basic_block bb1 = info->x_block;
+ basic_block bb2 = info->y_block;
+ edge b1 = BRANCH_EDGE (bb1);
+ edge b2 = BRANCH_EDGE (bb2);
+ edge f1 = FALLTHRU_EDGE (bb1);
+ edge f2 = FALLTHRU_EDGE (bb2);
+ bool reverse, match;
+ rtx set1, set2, cond1, cond2;
+ rtx src1, src2;
+ enum rtx_code code1, code2;
-/* Return true if I1 and I2 are equivalent and thus can be crossjumped. */
+ /* Get around possible forwarders on fallthru edges. Other cases
+ should be optimized out already. */
+ if (FORWARDER_BLOCK_P (f1->dest))
+ f1 = single_succ_edge (f1->dest);
-static bool
-insns_match_p (int mode ATTRIBUTE_UNUSED, rtx i1, rtx i2)
-{
- rtx p1, p2;
+ if (FORWARDER_BLOCK_P (f2->dest))
+ f2 = single_succ_edge (f2->dest);
- /* Verify that I1 and I2 are equivalent. */
- if (GET_CODE (i1) != GET_CODE (i2))
+ /* To simplify use of this function, return false if there are
+ unneeded forwarder blocks. These will get eliminated later
+ during cleanup_cfg. */
+ if (FORWARDER_BLOCK_P (f1->dest)
+ || FORWARDER_BLOCK_P (f2->dest)
+ || FORWARDER_BLOCK_P (b1->dest)
+ || FORWARDER_BLOCK_P (b2->dest))
return false;
- p1 = PATTERN (i1);
- p2 = PATTERN (i2);
-
- if (GET_CODE (p1) != GET_CODE (p2))
+ if (f1->dest == f2->dest && b1->dest == b2->dest)
+ reverse = false;
+ else if (f1->dest == b2->dest && b1->dest == f2->dest)
+ reverse = true;
+ else
return false;
- /* If this is a CALL_INSN, compare register usage information.
- If we don't check this on stack register machines, the two
- CALL_INSNs might be merged leaving reg-stack.c with mismatching
- numbers of stack registers in the same basic block.
- If we don't check this on machines with delay slots, a delay slot may
- be filled that clobbers a parameter expected by the subroutine.
+ set1 = pc_set (BB_END (bb1));
+ set2 = pc_set (BB_END (bb2));
+ if ((XEXP (SET_SRC (set1), 1) == pc_rtx)
+ != (XEXP (SET_SRC (set2), 1) == pc_rtx))
+ reverse = !reverse;
- ??? We take the simple route for now and assume that if they're
- equal, they were constructed identically. */
+ src1 = SET_SRC (set1);
+ src2 = SET_SRC (set2);
+ cond1 = XEXP (src1, 0);
+ cond2 = XEXP (src2, 0);
+ code1 = GET_CODE (cond1);
+ if (reverse)
+ code2 = reversed_comparison_code (cond2, BB_END (bb2));
+ else
+ code2 = GET_CODE (cond2);
- if (CALL_P (i1)
- && (!rtx_equal_p (CALL_INSN_FUNCTION_USAGE (i1),
- CALL_INSN_FUNCTION_USAGE (i2))
- || SIBLING_CALL_P (i1) != SIBLING_CALL_P (i2)))
+ if (code2 == UNKNOWN)
return false;
-#ifdef STACK_REGS
- /* If cross_jump_death_matters is not 0, the insn's mode
- indicates whether or not the insn contains any stack-like
- regs. */
-
- if ((mode & CLEANUP_POST_REGSTACK) && stack_regs_mentioned (i1))
- {
- /* If register stack conversion has already been done, then
- death notes must also be compared before it is certain that
- the two instruction streams match. */
-
- rtx note;
- HARD_REG_SET i1_regset, i2_regset;
-
- CLEAR_HARD_REG_SET (i1_regset);
- CLEAR_HARD_REG_SET (i2_regset);
-
- for (note = REG_NOTES (i1); note; note = XEXP (note, 1))
- if (REG_NOTE_KIND (note) == REG_DEAD && STACK_REG_P (XEXP (note, 0)))
- SET_HARD_REG_BIT (i1_regset, REGNO (XEXP (note, 0)));
-
- for (note = REG_NOTES (i2); note; note = XEXP (note, 1))
- if (REG_NOTE_KIND (note) == REG_DEAD && STACK_REG_P (XEXP (note, 0)))
- SET_HARD_REG_BIT (i2_regset, REGNO (XEXP (note, 0)));
-
- GO_IF_HARD_REG_EQUAL (i1_regset, i2_regset, done);
-
- return false;
-
- done:
- ;
- }
-#endif
-
- if (reload_completed
- ? rtx_renumbered_equal_p (p1, p2) : rtx_equal_p (p1, p2))
- return true;
-
- /* Do not do EQUIV substitution after reload. First, we're undoing the
- work of reload_cse. Second, we may be undoing the work of the post-
- reload splitting pass. */
- /* ??? Possibly add a new phase switch variable that can be used by
- targets to disallow the troublesome insns after splitting. */
- if (!reload_completed)
+ /* Make the source of the pc sets unreadable so that we can insns_match_p
+ won't process it.
+ The death_notes_match_p from insns_match_p won't see the local registers
+ used for the pc set, but that could only cause missed optimizations when
+ there are actually condjumps that use stack registers. */
+ SET_SRC (set1) = pc_rtx;
+ SET_SRC (set2) = pc_rtx;
+ /* Verify codes and operands match. */
+ if (code1 == code2)
{
- /* The following code helps take care of G++ cleanups. */
- rtx equiv1 = find_reg_equal_equiv_note (i1);
- rtx equiv2 = find_reg_equal_equiv_note (i2);
+ match = (insns_match_p (BB_END (bb1), BB_END (bb2), info)
+ && rtx_equiv_p (&XEXP (cond1, 0), XEXP (cond2, 0), 1, info)
+ && rtx_equiv_p (&XEXP (cond1, 1), XEXP (cond2, 1), 1, info));
- if (equiv1 && equiv2
- /* If the equivalences are not to a constant, they may
- reference pseudos that no longer exist, so we can't
- use them. */
- && (! reload_completed
- || (CONSTANT_P (XEXP (equiv1, 0))
- && rtx_equal_p (XEXP (equiv1, 0), XEXP (equiv2, 0)))))
- {
- rtx s1 = single_set (i1);
- rtx s2 = single_set (i2);
- if (s1 != 0 && s2 != 0
- && rtx_renumbered_equal_p (SET_DEST (s1), SET_DEST (s2)))
- {
- validate_change (i1, &SET_SRC (s1), XEXP (equiv1, 0), 1);
- validate_change (i2, &SET_SRC (s2), XEXP (equiv2, 0), 1);
- if (! rtx_renumbered_equal_p (p1, p2))
- cancel_changes (0);
- else if (apply_change_group ())
- return true;
- }
- }
}
-
- return false;
-}
-
-/* Look through the insns at the end of BB1 and BB2 and find the longest
- sequence that are equivalent. Store the first insns for that sequence
- in *F1 and *F2 and return the sequence length.
-
- To simplify callers of this function, if the blocks match exactly,
- store the head of the blocks in *F1 and *F2. */
-
-static int
-flow_find_cross_jump (int mode ATTRIBUTE_UNUSED, basic_block bb1,
- basic_block bb2, rtx *f1, rtx *f2)
-{
- rtx i1, i2, last1, last2, afterlast1, afterlast2;
- int ninsns = 0;
-
- /* Skip simple jumps at the end of the blocks. Complex jumps still
- need to be compared for equivalence, which we'll do below. */
-
- i1 = BB_END (bb1);
- last1 = afterlast1 = last2 = afterlast2 = NULL_RTX;
- if (onlyjump_p (i1)
- || (returnjump_p (i1) && !side_effects_p (PATTERN (i1))))
+ else if (code1 == swap_condition (code2))
{
- last1 = i1;
- i1 = PREV_INSN (i1);
- }
+ match = (insns_match_p (BB_END (bb1), BB_END (bb2), info)
+ && rtx_equiv_p (&XEXP (cond1, 1), XEXP (cond2, 0), 1, info)
+ && rtx_equiv_p (&XEXP (cond1, 0), XEXP (cond2, 1), 1, info));
- i2 = BB_END (bb2);
- if (onlyjump_p (i2)
- || (returnjump_p (i2) && !side_effects_p (PATTERN (i2))))
- {
- last2 = i2;
- /* Count everything except for unconditional jump as insn. */
- if (!simplejump_p (i2) && !returnjump_p (i2) && last1)
- ninsns++;
- i2 = PREV_INSN (i2);
}
+ else
+ match = false;
+ SET_SRC (set1) = src1;
+ SET_SRC (set2) = src2;
+ match &= verify_changes (0);
- while (true)
+ /* If we return true, we will join the blocks. Which means that
+ we will only have one branch prediction bit to work with. Thus
+ we require the existing branches to have probabilities that are
+ roughly similar. */
+ if (match
+ && !optimize_size
+ && maybe_hot_bb_p (bb1)
+ && maybe_hot_bb_p (bb2))
{
- /* Ignore notes. */
- while (!INSN_P (i1) && i1 != BB_HEAD (bb1))
- i1 = PREV_INSN (i1);
-
- while (!INSN_P (i2) && i2 != BB_HEAD (bb2))
- i2 = PREV_INSN (i2);
-
- if (i1 == BB_HEAD (bb1) || i2 == BB_HEAD (bb2))
- break;
-
- if (!insns_match_p (mode, i1, i2))
- break;
+ int prob2;
- merge_memattrs (i1, i2);
+ if (b1->dest == b2->dest)
+ prob2 = b2->probability;
+ else
+ /* Do not use f2 probability as f2 may be forwarded. */
+ prob2 = REG_BR_PROB_BASE - b2->probability;
- /* Don't begin a cross-jump with a NOTE insn. */
- if (INSN_P (i1))
+ /* Fail if the difference in probabilities is greater than 50%.
+ This rules out two well-predicted branches with opposite
+ outcomes. */
+ if (abs (b1->probability - prob2) > REG_BR_PROB_BASE / 2)
{
- /* If the merged insns have different REG_EQUAL notes, then
- remove them. */
- rtx equiv1 = find_reg_equal_equiv_note (i1);
- rtx equiv2 = find_reg_equal_equiv_note (i2);
-
- if (equiv1 && !equiv2)
- remove_note (i1, equiv1);
- else if (!equiv1 && equiv2)
- remove_note (i2, equiv2);
- else if (equiv1 && equiv2
- && !rtx_equal_p (XEXP (equiv1, 0), XEXP (equiv2, 0)))
- {
- remove_note (i1, equiv1);
- remove_note (i2, equiv2);
- }
+ if (dump_file)
+ fprintf (dump_file,
+ "Outcomes of branch in bb %i and %i differ too much (%i %i)\n",
+ bb1->index, bb2->index, b1->probability, prob2);
- afterlast1 = last1, afterlast2 = last2;
- last1 = i1, last2 = i2;
- ninsns++;
+ match = false;
}
-
- i1 = PREV_INSN (i1);
- i2 = PREV_INSN (i2);
}
-#ifdef HAVE_cc0
- /* Don't allow the insn after a compare to be shared by
- cross-jumping unless the compare is also shared. */
- if (ninsns && reg_mentioned_p (cc0_rtx, last1) && ! sets_cc0_p (last1))
- last1 = afterlast1, last2 = afterlast2, ninsns--;
-#endif
-
- /* Include preceding notes and labels in the cross-jump. One,
- this may bring us to the head of the blocks as requested above.
- Two, it keeps line number notes as matched as may be. */
- if (ninsns)
- {
- while (last1 != BB_HEAD (bb1) && !INSN_P (PREV_INSN (last1)))
- last1 = PREV_INSN (last1);
-
- if (last1 != BB_HEAD (bb1) && LABEL_P (PREV_INSN (last1)))
- last1 = PREV_INSN (last1);
-
- while (last2 != BB_HEAD (bb2) && !INSN_P (PREV_INSN (last2)))
- last2 = PREV_INSN (last2);
-
- if (last2 != BB_HEAD (bb2) && LABEL_P (PREV_INSN (last2)))
- last2 = PREV_INSN (last2);
-
- *f1 = last1;
- *f2 = last2;
- }
+ if (dump_file && match)
+ fprintf (dump_file, "Conditionals in bb %i and %i match.\n",
+ bb1->index, bb2->index);
- return ninsns;
+ if (!match)
+ cancel_changes (0);
+ return match;
}
/* Return true iff outgoing edges of BB1 and BB2 match, together with
the branch instruction. This means that if we commonize the control
flow before end of the basic block, the semantic remains unchanged.
+ If we need to compare jumps, we set STRUCT_EQUIV_MATCH_JUMPS is *MODE,
+ and pass it to struct_equiv_init.
We may assume that there exists one edge with a common destination. */
static bool
-outgoing_edges_match (int mode, basic_block bb1, basic_block bb2)
+outgoing_edges_match (int *mode, struct equiv_info *info)
{
+ basic_block bb1 = info->y_block;
+ basic_block bb2 = info->x_block;
int nehedges1 = 0, nehedges2 = 0;
edge fallthru1 = 0, fallthru2 = 0;
edge e1, e2;
@@ -1218,114 +1031,19 @@ outgoing_edges_match (int mode, basic_bl
& (EDGE_COMPLEX | EDGE_FAKE)) == 0
&& (!JUMP_P (BB_END (bb2)) || simplejump_p (BB_END (bb2))));
+ *mode |= STRUCT_EQUIV_MATCH_JUMPS;
/* Match conditional jumps - this may get tricky when fallthru and branch
edges are crossed. */
if (EDGE_COUNT (bb1->succs) == 2
&& any_condjump_p (BB_END (bb1))
&& onlyjump_p (BB_END (bb1)))
{
- edge b1, f1, b2, f2;
- bool reverse, match;
- rtx set1, set2, cond1, cond2;
- enum rtx_code code1, code2;
-
if (EDGE_COUNT (bb2->succs) != 2
|| !any_condjump_p (BB_END (bb2))
- || !onlyjump_p (BB_END (bb2)))
- return false;
-
- b1 = BRANCH_EDGE (bb1);
- b2 = BRANCH_EDGE (bb2);
- f1 = FALLTHRU_EDGE (bb1);
- f2 = FALLTHRU_EDGE (bb2);
-
- /* Get around possible forwarders on fallthru edges. Other cases
- should be optimized out already. */
- if (FORWARDER_BLOCK_P (f1->dest))
- f1 = single_succ_edge (f1->dest);
-
- if (FORWARDER_BLOCK_P (f2->dest))
- f2 = single_succ_edge (f2->dest);
-
- /* To simplify use of this function, return false if there are
- unneeded forwarder blocks. These will get eliminated later
- during cleanup_cfg. */
- if (FORWARDER_BLOCK_P (f1->dest)
- || FORWARDER_BLOCK_P (f2->dest)
- || FORWARDER_BLOCK_P (b1->dest)
- || FORWARDER_BLOCK_P (b2->dest))
- return false;
-
- if (f1->dest == f2->dest && b1->dest == b2->dest)
- reverse = false;
- else if (f1->dest == b2->dest && b1->dest == f2->dest)
- reverse = true;
- else
- return false;
-
- set1 = pc_set (BB_END (bb1));
- set2 = pc_set (BB_END (bb2));
- if ((XEXP (SET_SRC (set1), 1) == pc_rtx)
- != (XEXP (SET_SRC (set2), 1) == pc_rtx))
- reverse = !reverse;
-
- cond1 = XEXP (SET_SRC (set1), 0);
- cond2 = XEXP (SET_SRC (set2), 0);
- code1 = GET_CODE (cond1);
- if (reverse)
- code2 = reversed_comparison_code (cond2, BB_END (bb2));
- else
- code2 = GET_CODE (cond2);
-
- if (code2 == UNKNOWN)
+ || !onlyjump_p (BB_END (bb2))
+ || !struct_equiv_init (STRUCT_EQUIV_START | *mode, info))
return false;
-
- /* Verify codes and operands match. */
- match = ((code1 == code2
- && rtx_renumbered_equal_p (XEXP (cond1, 0), XEXP (cond2, 0))
- && rtx_renumbered_equal_p (XEXP (cond1, 1), XEXP (cond2, 1)))
- || (code1 == swap_condition (code2)
- && rtx_renumbered_equal_p (XEXP (cond1, 1),
- XEXP (cond2, 0))
- && rtx_renumbered_equal_p (XEXP (cond1, 0),
- XEXP (cond2, 1))));
-
- /* If we return true, we will join the blocks. Which means that
- we will only have one branch prediction bit to work with. Thus
- we require the existing branches to have probabilities that are
- roughly similar. */
- if (match
- && !optimize_size
- && maybe_hot_bb_p (bb1)
- && maybe_hot_bb_p (bb2))
- {
- int prob2;
-
- if (b1->dest == b2->dest)
- prob2 = b2->probability;
- else
- /* Do not use f2 probability as f2 may be forwarded. */
- prob2 = REG_BR_PROB_BASE - b2->probability;
-
- /* Fail if the difference in probabilities is greater than 50%.
- This rules out two well-predicted branches with opposite
- outcomes. */
- if (abs (b1->probability - prob2) > REG_BR_PROB_BASE / 2)
- {
- if (dump_file)
- fprintf (dump_file,
- "Outcomes of branch in bb %i and %i differ too much (%i %i)\n",
- bb1->index, bb2->index, b1->probability, prob2);
-
- return false;
- }
- }
-
- if (dump_file && match)
- fprintf (dump_file, "Conditionals in bb %i and %i match.\n",
- bb1->index, bb2->index);
-
- return match;
+ return condjump_equiv_p (info);
}
/* Generic case - we are seeing a computed jump, table jump or trapping
@@ -1385,7 +1103,7 @@ outgoing_edges_match (int mode, basic_bl
rr.update_label_nuses = false;
for_each_rtx (&BB_END (bb1), replace_label, &rr);
- match = insns_match_p (mode, BB_END (bb1), BB_END (bb2));
+ match = insns_match_p (BB_END (bb1), BB_END (bb2), info);
if (dump_file && match)
fprintf (dump_file,
"Tablejumps in bb %i and %i match.\n",
@@ -1407,7 +1125,8 @@ outgoing_edges_match (int mode, basic_bl
/* First ensure that the instructions match. There may be many outgoing
edges so this test is generally cheaper. */
- if (!insns_match_p (mode, BB_END (bb1), BB_END (bb2)))
+ if (!struct_equiv_init (STRUCT_EQUIV_START | *mode, info)
+ || !insns_match_p (BB_END (bb1), BB_END (bb2), info))
return false;
/* Search the outgoing edges, ensure that the counts do match, find possible
@@ -1473,14 +1192,13 @@ outgoing_edges_match (int mode, basic_bl
static bool
try_crossjump_to_edge (int mode, edge e1, edge e2)
{
- int nmatch;
+ int nmatch, i;
basic_block src1 = e1->src, src2 = e2->src;
basic_block redirect_to, redirect_from, to_remove;
- rtx newpos1, newpos2;
edge s;
edge_iterator ei;
-
- newpos1 = newpos2 = NULL_RTX;
+ struct equiv_info info;
+ rtx x_active, y_active;
/* If we have partitioned hot/cold basic blocks, it is a bad idea
to try this optimization.
@@ -1527,20 +1245,67 @@ try_crossjump_to_edge (int mode, edge e1
return false;
/* Look for the common insn sequence, part the first ... */
- if (!outgoing_edges_match (mode, src1, src2))
+ info.x_block = src2;
+ info.y_block = src1;
+ if (!outgoing_edges_match (&mode, &info))
return false;
/* ... and part the second. */
- nmatch = flow_find_cross_jump (mode, src1, src2, &newpos1, &newpos2);
+ info.input_cost = optimize_size ? COSTS_N_INSNS (1) : -1;
+ nmatch = struct_equiv_block_eq (STRUCT_EQUIV_START | mode, &info);
/* Don't proceed with the crossjump unless we found a sufficient number
of matching instructions or the 'from' block was totally matched
(such that its predecessors will hopefully be redirected and the
block removed). */
- if ((nmatch < PARAM_VALUE (PARAM_MIN_CROSSJUMP_INSNS))
- && (newpos1 != BB_HEAD (src1)))
+ if (!nmatch)
+ return false;
+ if ((nmatch -info.cur.input_count < PARAM_VALUE (PARAM_MIN_CROSSJUMP_INSNS))
+ && (info.cur.y_start != BB_HEAD (src1)))
+ return false;
+ while (info.need_rerun)
+ {
+ nmatch = struct_equiv_block_eq (STRUCT_EQUIV_RERUN | mode, &info);
+ if (!nmatch)
+ return false;
+ if ((nmatch -info.cur.input_count < PARAM_VALUE (PARAM_MIN_CROSSJUMP_INSNS))
+ && (info.cur.y_start != BB_HEAD (src1)))
+ return false;
+ }
+ nmatch = struct_equiv_block_eq (STRUCT_EQUIV_FINAL | mode, &info);
+ if ((nmatch -info.cur.input_count < PARAM_VALUE (PARAM_MIN_CROSSJUMP_INSNS))
+ && (info.cur.y_start != BB_HEAD (src1)))
return false;
+ /* Skip possible basic block header. */
+ x_active = info.cur.x_start;
+ if (LABEL_P (x_active))
+ x_active = NEXT_INSN (x_active);
+ if (NOTE_P (x_active))
+ x_active = NEXT_INSN (x_active);
+
+ y_active = info.cur.y_start;
+ if (LABEL_P (y_active))
+ y_active = NEXT_INSN (y_active);
+ if (NOTE_P (y_active))
+ y_active = NEXT_INSN (y_active);
+
+ /* In order for this code to become active, either we have to be called
+ before reload, or struct_equiv_block_eq needs to add register scavenging
+ code to allocate input_reg after reload. */
+ if (info.input_reg)
+ {
+ emit_insn_before (gen_move_insn (info.input_reg, info.x_input),
+ x_active);
+ emit_insn_before (gen_move_insn (info.input_reg, info.y_input),
+ y_active);
+ }
+
+ for (i = 0; i < info.cur.local_count; i++)
+ if (info.local_rvalue[i])
+ emit_insn_before (gen_move_insn (info.x_local[i], info.y_local[i]),
+ y_active);
+
/* Here we know that the insns in the end of SRC1 which are common with SRC2
will be deleted.
If we have tablejumps in the end of SRC1 and SRC2
@@ -1575,30 +1340,36 @@ try_crossjump_to_edge (int mode, edge e1
/* Avoid splitting if possible. We must always split when SRC2 has
EH predecessor edges, or we may end up with basic blocks with both
normal and EH predecessor edges. */
- if (newpos2 == BB_HEAD (src2)
+ if (info.cur.x_start == BB_HEAD (src2)
&& !(EDGE_PRED (src2, 0)->flags & EDGE_EH))
redirect_to = src2;
else
{
- if (newpos2 == BB_HEAD (src2))
+ if (info.cur.x_start == BB_HEAD (src2))
{
/* Skip possible basic block header. */
- if (LABEL_P (newpos2))
- newpos2 = NEXT_INSN (newpos2);
- if (NOTE_P (newpos2))
- newpos2 = NEXT_INSN (newpos2);
+ if (LABEL_P (info.cur.x_start))
+ info.cur.x_start = NEXT_INSN (info.cur.x_start);
+ if (NOTE_P (info.cur.x_start))
+ info.cur.x_start = NEXT_INSN (info.cur.x_start);
}
if (dump_file)
fprintf (dump_file, "Splitting bb %i before %i insns\n",
src2->index, nmatch);
- redirect_to = split_block (src2, PREV_INSN (newpos2))->dest;
+ redirect_to = split_block (src2, PREV_INSN (info.cur.x_start))->dest;
+ COPY_REG_SET (info.y_block->il.rtl->global_live_at_end,
+ info.x_block->il.rtl->global_live_at_end);
}
if (dump_file)
- fprintf (dump_file,
- "Cross jumping from bb %i to bb %i; %i common insns\n",
- src1->index, src2->index, nmatch);
+ {
+ fprintf (dump_file, "Cross jumping from bb %i to bb %i; %i common insns",
+ src1->index, src2->index, nmatch);
+ if (info.cur.local_count)
+ fprintf (dump_file, ", %i local registers", info.cur.local_count);
+ fprintf (dump_file, "\n");
+ }
redirect_to->count += src1->count;
redirect_to->frequency += src1->frequency;
@@ -1662,14 +1433,7 @@ try_crossjump_to_edge (int mode, edge e1
/* Edit SRC1 to go to REDIRECT_TO at NEWPOS1. */
- /* Skip possible basic block header. */
- if (LABEL_P (newpos1))
- newpos1 = NEXT_INSN (newpos1);
-
- if (NOTE_P (newpos1))
- newpos1 = NEXT_INSN (newpos1);
-
- redirect_from = split_block (src1, PREV_INSN (newpos1))->src;
+ redirect_from = split_block (src1, PREV_INSN (y_active))->src;
to_remove = single_succ (redirect_from);
redirect_edge_and_branch_force (single_succ_edge (redirect_from), redirect_to);
Index: Makefile.in
===================================================================
/usr/bin/diff -dup -F'^(' -u -L Makefile.in (revision 107723) -L Makefile.in (working copy) .svn/text-base/Makefile.in.svn-base Makefile.in
--- Makefile.in (revision 107723)
+++ Makefile.in (working copy)
@@ -978,7 +978,7 @@ OBJS-common = \
reload.o reload1.o reorg.o resource.o rtl.o rtlanal.o rtl-error.o \
sbitmap.o sched-deps.o sched-ebb.o sched-rgn.o sched-vis.o sdbout.o \
simplify-rtx.o sreal.o stmt.o stor-layout.o stringpool.o \
- targhooks.o timevar.o toplev.o tracer.o tree.o tree-dump.o \
+ struct-equiv.o targhooks.o timevar.o toplev.o tracer.o tree.o tree-dump.o \
varasm.o varray.o vec.o version.o vmsdbgout.o xcoffout.o alloc-pool.o \
et-forest.o cfghooks.o bt-load.o pretty-print.o $(GGC) web.o passes.o \
tree-profile.o rtlhooks.o cfgexpand.o lambda-mat.o \
@@ -2316,6 +2316,10 @@ cfgloop.o : cfgloop.c $(CONFIG_H) $(SYST
cfgloopanal.o : cfgloopanal.c $(CONFIG_H) $(SYSTEM_H) $(RTL_H) \
$(BASIC_BLOCK_H) hard-reg-set.h $(CFGLOOP_H) $(EXPR_H) coretypes.h $(TM_H) \
$(OBSTACK_H) output.h
+struct-equiv.o : struct-equiv.c $(CONFIG_H) $(SYSTEM_H) coretypes.h $(TM_H) \
+ $(RTL_H) $(TIMEVAR_H) hard-reg-set.h output.h $(FLAGS_H) $(RECOG_H) \
+ toplev.h insn-config.h cselib.h $(TARGET_H) $(TM_P_H) $(PARAMS_H) \
+ $(REGS_H) $(EMIT_RTL_H) $(CFGLAYOUT_H) tree-pass.h cfgloop.h expr.h
loop-iv.o : loop-iv.c $(CONFIG_H) $(SYSTEM_H) $(RTL_H) $(BASIC_BLOCK_H) \
hard-reg-set.h $(CFGLOOP_H) $(EXPR_H) coretypes.h $(TM_H) $(OBSTACK_H) \
output.h intl.h
Index: basic-block.h
===================================================================
/usr/bin/diff -dup -F'^(' -u -L basic-block.h (revision 107723) -L basic-block.h (working copy) .svn/text-base/basic-block.h.svn-base basic-block.h
--- basic-block.h (revision 107723)
+++ basic-block.h (working copy)
@@ -831,6 +831,16 @@ enum update_life_extent
#define CLEANUP_CFGLAYOUT 128 /* Do cleanup in cfglayout mode. */
#define CLEANUP_LOG_LINKS 256 /* Update log links. */
+/* The following are ORed in on top of the CLEANUP* flags in calls to
+ struct_equiv_block_eq. */
+#define STRUCT_EQUIV_START 512 /* Initializes the search range. */
+#define STRUCT_EQUIV_RERUN 1024 /* Rerun to find register use in
+ found equivalence. */
+#define STRUCT_EQUIV_FINAL 2048 /* Make any changes necessary to get
+ actual equivalence. */
+#define STRUCT_EQUIV_NEED_FULL_BLOCK 4096 /* struct_equiv_block_eq is required
+ to match only full blocks */
+#define STRUCT_EQUIV_MATCH_JUMPS 8192 /* Also include the jumps at the end of the block in the comparison. */
extern void life_analysis (FILE *, int);
extern int update_life_info (sbitmap, enum update_life_extent, int);
extern int update_life_info_in_dirty_blocks (enum update_life_extent, int);
@@ -989,6 +999,173 @@ extern basic_block get_bb_original (basi
extern void set_bb_copy (basic_block, basic_block);
extern basic_block get_bb_copy (basic_block);
+/* In struct-equiv.c */
+
#include "cfghooks.h"
+/* Constants used to size arrays in struct equiv_info (currently only one).
+ When these limits are exceeded, struct_equiv returns zero.
+ The maximum number of pseudo registers that are different in the two blocks,
+ but appear in equivalent places and are dead at the end (or where one of
+ a pair is dead at the end). */
+#define STRUCT_EQUIV_MAX_LOCAL 16
+/* The maximum number of references to an input register that struct_equiv
+ can handle. */
+
+/* Structure used to track state during struct_equiv that can be rolled
+ back when we find we can't match an insn, or if we want to match part
+ of it in a different way.
+ This information pertains to the pair of partial blocks that has been
+ matched so far. Since this pair is structurally equivalent, this is
+ conceptually just one partial block expressed in two potentially
+ different ways. */
+struct struct_equiv_checkpoint
+{
+ int ninsns; /* Insns are matched so far. */
+ int local_count; /* Number of block-local registers. */
+ int input_count; /* Number of inputs to the block. */
+
+ /* X_START and Y_START are the first insns (in insn stream order)
+ of the partial blocks that have been considered for matching so far.
+ Since we are scanning backwards, they are also the instructions that
+ are currently considered - or the last ones that have been considered -
+ for matching (Unless we tracked back to these because a preceding
+ instruction failed to match). */
+ rtx x_start, y_start;
+
+ /* INPUT_VALID indicates if we have actually set up X_INPUT / Y_INPUT
+ during the current pass; we keep X_INPUT / Y_INPUT around between passes
+ so that we can match REG_EQUAL / REG_EQUIV notes referring to these. */
+ bool input_valid;
+
+ /* Some information would be expensive to exactly checkpoint, so we
+ merely increment VERSION any time information about local
+ registers, inputs and/or register liveness changes. When backtracking,
+ it is decremented for changes that can be undone, and if a discrepancy
+ remains, NEED_RERUN in the relevant struct equiv_info is set to indicate
+ that a new pass should be made over the entire block match to get
+ accurate register information. */
+ int version;
+};
+
+/* A struct equiv_info is used to pass information to struct_equiv and
+ to gather state while two basic blocks are checked for structural
+ equivalence. */
+
+struct equiv_info
+{
+ /* Fields set up by the caller to struct_equiv_block_eq */
+
+ basic_block x_block, y_block; /* The two blocks being matched. */
+
+ /* MODE carries the mode bits from cleanup_cfg if we are called from
+ try_crossjump_to_edge, and additionally it carries the
+ STRUCT_EQUIV_* bits described above. */
+ int mode;
+
+ /* INPUT_COST is the cost that adding an extra input to the matched blocks
+ is supposed to have, and is taken into account when considering if the
+ matched sequence should be extended backwards. input_cost < 0 means
+ don't accept any inputs at all. */
+ int input_cost;
+
+
+ /* Fields to track state inside of struct_equiv_block_eq. Some of these
+ are also outputs. */
+
+ /* X_INPUT and Y_INPUT are used by struct_equiv to record a register that
+ is used as an input parameter, i.e. where different registers are used
+ as sources. This is only used for a register that is live at the end
+ of the blocks, or in some identical code at the end of the blocks;
+ Inputs that are dead at the end go into X_LOCAL / Y_LOCAL. */
+ rtx x_input, y_input;
+ /* When a previous pass has identified a valid input, INPUT_REG is set
+ by struct_equiv_block_eq, and it is henceforth replaced in X_BLOCK
+ for the input. */
+ rtx input_reg;
+
+ /* COMMON_LIVE keeps track of the registers which are currently live
+ (as we scan backwards from the end) and have the same numbers in both
+ blocks. N.B. a register that is in common_live is unsuitable to become
+ a local reg. */
+ regset common_live;
+ /* Likewise, X_LOCAL_LIVE / Y_LOCAL_LIVE keep track of registers that are
+ local to one of the blocks; these registers must not be accepted as
+ identical when encountered in both blocks. */
+ regset x_local_live, y_local_live;
+
+ /* EQUIV_USED indicates for which insns a REG_EQUAL or REG_EQUIV note is
+ being used, to avoid having to backtrack in the next pass, so that we
+ get accurate life info for this insn then. For each such insn,
+ the bit with the number corresponding to the CUR.NINSNS value at the
+ time of scanning is set. */
+ bitmap equiv_used;
+
+ /* Current state that can be saved & restored easily. */
+ struct struct_equiv_checkpoint cur;
+ /* BEST_MATCH is used to store the best match so far, weighing the
+ cost of matched insns COSTS_N_INSNS (CUR.NINSNS) against the cost
+ CUR.INPUT_COUNT * INPUT_COST of setting up the inputs. */
+ struct struct_equiv_checkpoint best_match;
+ /* If a checkpoint restore failed, or an input conflict newly arises,
+ NEED_RERUN is set. This has to be tested by the caller to re-run
+ the comparison if the match appears otherwise sound. The state kept in
+ x_start, y_start, equiv_used and check_input_conflict ensures that
+ we won't loop indefinetly. */
+ bool need_rerun;
+ /* If there is indication of an input conflict at the end,
+ CHECK_INPUT_CONFLICT is set so that we'll check for input conflicts
+ for each insn in the next pass. This is needed so that we won't discard
+ a partial match if there is a longer match that has to be abandoned due
+ to an input conflict. */
+ bool check_input_conflict;
+ /* HAD_INPUT_CONFLICT is set if CHECK_INPUT_CONFLICT was already set and we
+ have passed a point where there were multiple dying inputs. This helps
+ us decide if we should set check_input_conflict for the next pass. */
+ bool had_input_conflict;
+
+ /* LIVE_UPDATE controls if we want to change any life info at all. We
+ set it to false during REG_EQUAL / REG_EUQIV note comparison of the final
+ pass so that we don't introduce new registers just for the note; if we
+ can't match the notes without the current register information, we drop
+ them. */
+ bool live_update;
+
+ /* X_LOCAL and Y_LOCAL are used to gather register numbers of register pairs
+ that are local to X_BLOCK and Y_BLOCK, with CUR.LOCAL_COUNT being the index
+ to the next free entry. */
+ rtx x_local[STRUCT_EQUIV_MAX_LOCAL], y_local[STRUCT_EQUIV_MAX_LOCAL];
+ /* LOCAL_RVALUE is nonzero if the corresponding X_LOCAL / Y_LOCAL entry
+ was a source operand (including STRICT_LOW_PART) for the last invocation
+ of struct_equiv mentioning it, zero if it was a destination-only operand.
+ Since we are scanning backwards, this means the register is input/local
+ for the (partial) block scanned so far. */
+ bool local_rvalue[STRUCT_EQUIV_MAX_LOCAL];
+
+
+ /* Additional fields that are computed for the convenience of the caller. */
+
+ /* DYING_INPUTS is set to the number of local registers that turn out
+ to be inputs to the (possibly partial) block. */
+ int dying_inputs;
+ /* X_END and Y_END are the last insns in X_BLOCK and Y_BLOCK, respectively,
+ that are being compared. A final jump insn will not be included. */
+ rtx x_end, y_end;
+
+};
+
+enum equiv_rvalue
+{ dst = 0, /* destination */
+ src = 1, /* source */
+ ignore_dst = -1 /* source, ignore SET_DEST of SET / clobber. */
+};
+
+extern bool insns_match_p (rtx, rtx, struct equiv_info *);
+extern int struct_equiv_block_eq (int, struct equiv_info *);
+extern bool struct_equiv_init (int, struct equiv_info *);
+extern bool rtx_equiv_p (rtx *, rtx, enum equiv_rvalue, struct equiv_info *);
+
+/* In cfgrtl.c */
+extern bool condjump_equiv_p (struct equiv_info *);
+
#endif /* GCC_BASIC_BLOCK_H */
Index: struct-equiv.c
===================================================================
/usr/bin/diff -dup -F'^(' -u -L struct-equiv.c (revision 107723) -L struct-equiv.c (working copy) .svn/text-base/struct-equiv.c.svn-base struct-equiv.c
--- struct-equiv.c (revision 107723)
+++ struct-equiv.c (working copy)
@@ -19,867 +19,78 @@ along with GCC; see the file COPYING. I
Software Foundation, 51 Franklin Street, Fifth Floor, Boston, MA
02110-1301, USA. */
-/* This file contains optimizer of the control flow. The main entry point is
- cleanup_cfg. Following optimizations are performed:
+/* Try to match two basic blocks - or their ends - for structural equivalence.
+ We scan the blocks from their ends backwards. When we see mismatching
+ registers, we try to figure out if these could be registers local to these
+ (partial) blocks, or input values.
+ The main entry point to this file is struct_equiv_block_eq. This function
+ uses a struct equiv_info to accept some of its inputs, to keep track of its
+ internal state, to pass down to its helper functions, and to communicate
+ some of the results back to the caller.
- - Unreachable blocks removal
- - Edge forwarding (edge to the forwarder block is forwarded to its
- successor. Simplification of the branch instruction is performed by
- underlying infrastructure so branch can be converted to simplejump or
- eliminated).
- - Cross jumping (tail merging)
- - Conditional jump-around-simplejump simplification
- - Basic block merging. */
+ Most scans will result in a failure to match a sufficient number of insns
+ to make any optimization worth while, therefore the process is geared more
+ to quick scanning rather than the ability to exactly backtrack when we
+ find a mismatch. The information gathered is still meaningful to make a
+ preliminary decision if we want to do an optimization, we might only
+ slightly overestimate the number of matchable insns, and underestimate
+ the number of inputs an miss an input conflict. Sufficient information
+ is gathered so that when we make another pass, we won't have to backtrack
+ at the same point.
+ Another issue is that information in memory atttributes and/or REG_NOTES
+ might have to be merged or discarded to make a valid match. We don't want
+ to discard such information when we are not certain that we want to merge
+ the two (partial) blocks.
+ For these reasons, struct_equiv_block_eq has to be called first with the
+ STRUCT_EQUIV_START bit set in the mode parameter. This will calculate the
+ number of matched insns and the number and types of inputs. If the
+ need_rerun field is set, the results are only tentative, and the caller
+ has to call again with STRUCT_EQUIV_RERUN till need_rerun is false in
+ order to get a reliable match.
+ To install the changes necessary for the match, the function has to be
+ called again with STRUCT_EQUIV_FINAL.
+
+ While scanning an insn, we process first all the SET_DESTs, then the
+ SET_SRCes, then the REG_NOTES, in order to keep the register lifeness
+ information consistent.
+ If we were to mix up the order for sources / destinations in an insn where
+ a source is also a destination, we'd end up being mistaken to think that
+ the register is not live in the preceding insn. */
#include "config.h"
#include "system.h"
#include "coretypes.h"
#include "tm.h"
#include "rtl.h"
-#include "hard-reg-set.h"
#include "regs.h"
-#include "timevar.h"
#include "output.h"
#include "insn-config.h"
#include "flags.h"
#include "recog.h"
-#include "toplev.h"
-#include "cselib.h"
-#include "params.h"
#include "tm_p.h"
#include "target.h"
-#include "cfglayout.h"
#include "emit-rtl.h"
-#include "tree-pass.h"
-#include "cfgloop.h"
-#include "expr.h"
-
-#define FORWARDER_BLOCK_P(BB) ((BB)->flags & BB_FORWARDER_BLOCK)
-
-/* Set to true when we are running first pass of try_optimize_cfg loop. */
-static bool first_pass;
-static bool try_crossjump_to_edge (int, edge, edge);
-static bool try_crossjump_bb (int, basic_block);
-static bool outgoing_edges_match (int, basic_block, basic_block);
-static int flow_find_cross_jump (int, basic_block, basic_block, rtx *, rtx *);
-static bool insns_match_p (int, rtx, rtx);
+#include "reload.h"
-static void merge_blocks_move_predecessor_nojumps (basic_block, basic_block);
-static void merge_blocks_move_successor_nojumps (basic_block, basic_block);
-static bool try_optimize_cfg (int);
-static bool try_simplify_condjump (basic_block);
-static bool try_forward_edges (int, basic_block);
-static edge thread_jump (int, edge, basic_block);
-static bool mark_effect (rtx, bitmap);
-static void notice_new_block (basic_block);
-static void update_forwarder_flag (basic_block);
-static int mentions_nonequal_regs (rtx *, void *);
static void merge_memattrs (rtx, rtx);
-
-/* Set flags for newly created block. */
-
-static void
-notice_new_block (basic_block bb)
-{
- if (!bb)
- return;
-
- if (forwarder_block_p (bb))
- bb->flags |= BB_FORWARDER_BLOCK;
-}
-
-/* Recompute forwarder flag after block has been modified. */
-
-static void
-update_forwarder_flag (basic_block bb)
-{
- if (forwarder_block_p (bb))
- bb->flags |= BB_FORWARDER_BLOCK;
- else
- bb->flags &= ~BB_FORWARDER_BLOCK;
-}
-
-/* Simplify a conditional jump around an unconditional jump.
- Return true if something changed. */
-
-static bool
-try_simplify_condjump (basic_block cbranch_block)
-{
- basic_block jump_block, jump_dest_block, cbranch_dest_block;
- edge cbranch_jump_edge, cbranch_fallthru_edge;
- rtx cbranch_insn;
-
- /* Verify that there are exactly two successors. */
- if (EDGE_COUNT (cbranch_block->succs) != 2)
- return false;
-
- /* Verify that we've got a normal conditional branch at the end
- of the block. */
- cbranch_insn = BB_END (cbranch_block);
- if (!any_condjump_p (cbranch_insn))
- return false;
-
- cbranch_fallthru_edge = FALLTHRU_EDGE (cbranch_block);
- cbranch_jump_edge = BRANCH_EDGE (cbranch_block);
-
- /* The next block must not have multiple predecessors, must not
- be the last block in the function, and must contain just the
- unconditional jump. */
- jump_block = cbranch_fallthru_edge->dest;
- if (!single_pred_p (jump_block)
- || jump_block->next_bb == EXIT_BLOCK_PTR
- || !FORWARDER_BLOCK_P (jump_block))
- return false;
- jump_dest_block = single_succ (jump_block);
-
- /* If we are partitioning hot/cold basic blocks, we don't want to
- mess up unconditional or indirect jumps that cross between hot
- and cold sections.
-
- Basic block partitioning may result in some jumps that appear to
- be optimizable (or blocks that appear to be mergeable), but which really
- must be left untouched (they are required to make it safely across
- partition boundaries). See the comments at the top of
- bb-reorder.c:partition_hot_cold_basic_blocks for complete details. */
-
- if (BB_PARTITION (jump_block) != BB_PARTITION (jump_dest_block)
- || (cbranch_jump_edge->flags & EDGE_CROSSING))
- return false;
-
- /* The conditional branch must target the block after the
- unconditional branch. */
- cbranch_dest_block = cbranch_jump_edge->dest;
-
- if (cbranch_dest_block == EXIT_BLOCK_PTR
- || !can_fallthru (jump_block, cbranch_dest_block))
- return false;
-
- /* Invert the conditional branch. */
- if (!invert_jump (cbranch_insn, block_label (jump_dest_block), 0))
- return false;
-
- if (dump_file)
- fprintf (dump_file, "Simplifying condjump %i around jump %i\n",
- INSN_UID (cbranch_insn), INSN_UID (BB_END (jump_block)));
-
- /* Success. Update the CFG to match. Note that after this point
- the edge variable names appear backwards; the redirection is done
- this way to preserve edge profile data. */
- cbranch_jump_edge = redirect_edge_succ_nodup (cbranch_jump_edge,
- cbranch_dest_block);
- cbranch_fallthru_edge = redirect_edge_succ_nodup (cbranch_fallthru_edge,
- jump_dest_block);
- cbranch_jump_edge->flags |= EDGE_FALLTHRU;
- cbranch_fallthru_edge->flags &= ~EDGE_FALLTHRU;
- update_br_prob_note (cbranch_block);
-
- /* Delete the block with the unconditional jump, and clean up the mess. */
- delete_basic_block (jump_block);
- tidy_fallthru_edge (cbranch_jump_edge);
- update_forwarder_flag (cbranch_block);
-
- return true;
-}
-
-/* Attempt to prove that operation is NOOP using CSElib or mark the effect
- on register. Used by jump threading. */
-
-static bool
-mark_effect (rtx exp, regset nonequal)
-{
- int regno;
- rtx dest;
- switch (GET_CODE (exp))
- {
- /* In case we do clobber the register, mark it as equal, as we know the
- value is dead so it don't have to match. */
- case CLOBBER:
- if (REG_P (XEXP (exp, 0)))
- {
- dest = XEXP (exp, 0);
- regno = REGNO (dest);
- CLEAR_REGNO_REG_SET (nonequal, regno);
- if (regno < FIRST_PSEUDO_REGISTER)
- {
- int n = hard_regno_nregs[regno][GET_MODE (dest)];
- while (--n > 0)
- CLEAR_REGNO_REG_SET (nonequal, regno + n);
- }
- }
- return false;
-
- case SET:
- if (rtx_equal_for_cselib_p (SET_DEST (exp), SET_SRC (exp)))
- return false;
- dest = SET_DEST (exp);
- if (dest == pc_rtx)
- return false;
- if (!REG_P (dest))
- return true;
- regno = REGNO (dest);
- SET_REGNO_REG_SET (nonequal, regno);
- if (regno < FIRST_PSEUDO_REGISTER)
- {
- int n = hard_regno_nregs[regno][GET_MODE (dest)];
- while (--n > 0)
- SET_REGNO_REG_SET (nonequal, regno + n);
- }
- return false;
-
- default:
- return false;
- }
-}
-
-/* Return nonzero if X is a register set in regset DATA.
- Called via for_each_rtx. */
-static int
-mentions_nonequal_regs (rtx *x, void *data)
-{
- regset nonequal = (regset) data;
- if (REG_P (*x))
- {
- int regno;
-
- regno = REGNO (*x);
- if (REGNO_REG_SET_P (nonequal, regno))
- return 1;
- if (regno < FIRST_PSEUDO_REGISTER)
- {
- int n = hard_regno_nregs[regno][GET_MODE (*x)];
- while (--n > 0)
- if (REGNO_REG_SET_P (nonequal, regno + n))
- return 1;
- }
- }
- return 0;
-}
-/* Attempt to prove that the basic block B will have no side effects and
- always continues in the same edge if reached via E. Return the edge
- if exist, NULL otherwise. */
-
-static edge
-thread_jump (int mode, edge e, basic_block b)
-{
- rtx set1, set2, cond1, cond2, insn;
- enum rtx_code code1, code2, reversed_code2;
- bool reverse1 = false;
- unsigned i;
- regset nonequal;
- bool failed = false;
- reg_set_iterator rsi;
-
- if (b->flags & BB_NONTHREADABLE_BLOCK)
- return NULL;
-
- /* At the moment, we do handle only conditional jumps, but later we may
- want to extend this code to tablejumps and others. */
- if (EDGE_COUNT (e->src->succs) != 2)
- return NULL;
- if (EDGE_COUNT (b->succs) != 2)
- {
- b->flags |= BB_NONTHREADABLE_BLOCK;
- return NULL;
- }
-
- /* Second branch must end with onlyjump, as we will eliminate the jump. */
- if (!any_condjump_p (BB_END (e->src)))
- return NULL;
-
- if (!any_condjump_p (BB_END (b)) || !onlyjump_p (BB_END (b)))
- {
- b->flags |= BB_NONTHREADABLE_BLOCK;
- return NULL;
- }
-
- set1 = pc_set (BB_END (e->src));
- set2 = pc_set (BB_END (b));
- if (((e->flags & EDGE_FALLTHRU) != 0)
- != (XEXP (SET_SRC (set1), 1) == pc_rtx))
- reverse1 = true;
-
- cond1 = XEXP (SET_SRC (set1), 0);
- cond2 = XEXP (SET_SRC (set2), 0);
- if (reverse1)
- code1 = reversed_comparison_code (cond1, BB_END (e->src));
- else
- code1 = GET_CODE (cond1);
-
- code2 = GET_CODE (cond2);
- reversed_code2 = reversed_comparison_code (cond2, BB_END (b));
-
- if (!comparison_dominates_p (code1, code2)
- && !comparison_dominates_p (code1, reversed_code2))
- return NULL;
-
- /* Ensure that the comparison operators are equivalent.
- ??? This is far too pessimistic. We should allow swapped operands,
- different CCmodes, or for example comparisons for interval, that
- dominate even when operands are not equivalent. */
- if (!rtx_equal_p (XEXP (cond1, 0), XEXP (cond2, 0))
- || !rtx_equal_p (XEXP (cond1, 1), XEXP (cond2, 1)))
- return NULL;
-
- /* Short circuit cases where block B contains some side effects, as we can't
- safely bypass it. */
- for (insn = NEXT_INSN (BB_HEAD (b)); insn != NEXT_INSN (BB_END (b));
- insn = NEXT_INSN (insn))
- if (INSN_P (insn) && side_effects_p (PATTERN (insn)))
- {
- b->flags |= BB_NONTHREADABLE_BLOCK;
- return NULL;
- }
-
- cselib_init (false);
-
- /* First process all values computed in the source basic block. */
- for (insn = NEXT_INSN (BB_HEAD (e->src));
- insn != NEXT_INSN (BB_END (e->src));
- insn = NEXT_INSN (insn))
- if (INSN_P (insn))
- cselib_process_insn (insn);
-
- nonequal = BITMAP_ALLOC (NULL);
- CLEAR_REG_SET (nonequal);
-
- /* Now assume that we've continued by the edge E to B and continue
- processing as if it were same basic block.
- Our goal is to prove that whole block is an NOOP. */
-
- for (insn = NEXT_INSN (BB_HEAD (b));
- insn != NEXT_INSN (BB_END (b)) && !failed;
- insn = NEXT_INSN (insn))
- {
- if (INSN_P (insn))
- {
- rtx pat = PATTERN (insn);
-
- if (GET_CODE (pat) == PARALLEL)
- {
- for (i = 0; i < (unsigned)XVECLEN (pat, 0); i++)
- failed |= mark_effect (XVECEXP (pat, 0, i), nonequal);
- }
- else
- failed |= mark_effect (pat, nonequal);
- }
-
- cselib_process_insn (insn);
- }
-
- /* Later we should clear nonequal of dead registers. So far we don't
- have life information in cfg_cleanup. */
- if (failed)
- {
- b->flags |= BB_NONTHREADABLE_BLOCK;
- goto failed_exit;
- }
-
- /* cond2 must not mention any register that is not equal to the
- former block. */
- if (for_each_rtx (&cond2, mentions_nonequal_regs, nonequal))
- goto failed_exit;
-
- /* In case liveness information is available, we need to prove equivalence
- only of the live values. */
- if (mode & CLEANUP_UPDATE_LIFE)
- AND_REG_SET (nonequal, b->il.rtl->global_live_at_end);
-
- EXECUTE_IF_SET_IN_REG_SET (nonequal, 0, i, rsi)
- goto failed_exit;
-
- BITMAP_FREE (nonequal);
- cselib_finish ();
- if ((comparison_dominates_p (code1, code2) != 0)
- != (XEXP (SET_SRC (set2), 1) == pc_rtx))
- return BRANCH_EDGE (b);
- else
- return FALLTHRU_EDGE (b);
-
-failed_exit:
- BITMAP_FREE (nonequal);
- cselib_finish ();
- return NULL;
-}
-
-/* Attempt to forward edges leaving basic block B.
- Return true if successful. */
-
-static bool
-try_forward_edges (int mode, basic_block b)
-{
- bool changed = false;
- edge_iterator ei;
- edge e, *threaded_edges = NULL;
-
- /* If we are partitioning hot/cold basic blocks, we don't want to
- mess up unconditional or indirect jumps that cross between hot
- and cold sections.
-
- Basic block partitioning may result in some jumps that appear to
- be optimizable (or blocks that appear to be mergeable), but which really m
- ust be left untouched (they are required to make it safely across
- partition boundaries). See the comments at the top of
- bb-reorder.c:partition_hot_cold_basic_blocks for complete details. */
-
- if (find_reg_note (BB_END (b), REG_CROSSING_JUMP, NULL_RTX))
- return false;
-
- for (ei = ei_start (b->succs); (e = ei_safe_edge (ei)); )
- {
- basic_block target, first;
- int counter;
- bool threaded = false;
- int nthreaded_edges = 0;
- bool may_thread = first_pass | (b->flags & BB_DIRTY);
-
- /* Skip complex edges because we don't know how to update them.
-
- Still handle fallthru edges, as we can succeed to forward fallthru
- edge to the same place as the branch edge of conditional branch
- and turn conditional branch to an unconditional branch. */
- if (e->flags & EDGE_COMPLEX)
- {
- ei_next (&ei);
- continue;
- }
-
- target = first = e->dest;
- counter = 0;
-
- /* If we are partitioning hot/cold basic_blocks, we don't want to mess
- up jumps that cross between hot/cold sections.
-
- Basic block partitioning may result in some jumps that appear
- to be optimizable (or blocks that appear to be mergeable), but which
- really must be left untouched (they are required to make it safely
- across partition boundaries). See the comments at the top of
- bb-reorder.c:partition_hot_cold_basic_blocks for complete
- details. */
-
- if (first != EXIT_BLOCK_PTR
- && find_reg_note (BB_END (first), REG_CROSSING_JUMP, NULL_RTX))
- return false;
-
- while (counter < n_basic_blocks)
- {
- basic_block new_target = NULL;
- bool new_target_threaded = false;
- may_thread |= target->flags & BB_DIRTY;
-
- if (FORWARDER_BLOCK_P (target)
- && !(single_succ_edge (target)->flags & EDGE_CROSSING)
- && single_succ (target) != EXIT_BLOCK_PTR)
- {
- /* Bypass trivial infinite loops. */
- new_target = single_succ (target);
- if (target == new_target)
- counter = n_basic_blocks;
- }
-
- /* Allow to thread only over one edge at time to simplify updating
- of probabilities. */
- else if ((mode & CLEANUP_THREADING) && may_thread)
- {
- edge t = thread_jump (mode, e, target);
- if (t)
- {
- if (!threaded_edges)
- threaded_edges = xmalloc (sizeof (*threaded_edges)
- * n_basic_blocks);
- else
- {
- int i;
-
- /* Detect an infinite loop across blocks not
- including the start block. */
- for (i = 0; i < nthreaded_edges; ++i)
- if (threaded_edges[i] == t)
- break;
- if (i < nthreaded_edges)
- {
- counter = n_basic_blocks;
- break;
- }
- }
-
- /* Detect an infinite loop across the start block. */
- if (t->dest == b)
- break;
-
- gcc_assert (nthreaded_edges < n_basic_blocks);
- threaded_edges[nthreaded_edges++] = t;
-
- new_target = t->dest;
- new_target_threaded = true;
- }
- }
-
- if (!new_target)
- break;
-
- /* Avoid killing of loop pre-headers, as it is the place loop
- optimizer wants to hoist code to.
-
- For fallthru forwarders, the LOOP_BEG note must appear between
- the header of block and CODE_LABEL of the loop, for non forwarders
- it must appear before the JUMP_INSN. */
- if ((mode & CLEANUP_PRE_LOOP) && optimize && flag_loop_optimize)
- {
- rtx insn = (EDGE_SUCC (target, 0)->flags & EDGE_FALLTHRU
- ? BB_HEAD (target) : prev_nonnote_insn (BB_END (target)));
-
- if (!NOTE_P (insn))
- insn = NEXT_INSN (insn);
-
- for (; insn && !LABEL_P (insn) && !INSN_P (insn);
- insn = NEXT_INSN (insn))
- if (NOTE_P (insn)
- && NOTE_LINE_NUMBER (insn) == NOTE_INSN_LOOP_BEG)
- break;
-
- if (insn && NOTE_P (insn))
- break;
-
- /* Do not clean up branches to just past the end of a loop
- at this time; it can mess up the loop optimizer's
- recognition of some patterns. */
-
- insn = PREV_INSN (BB_HEAD (target));
- if (insn && NOTE_P (insn)
- && NOTE_LINE_NUMBER (insn) == NOTE_INSN_LOOP_END)
- break;
- }
-
- counter++;
- target = new_target;
- threaded |= new_target_threaded;
- }
-
- if (counter >= n_basic_blocks)
- {
- if (dump_file)
- fprintf (dump_file, "Infinite loop in BB %i.\n",
- target->index);
- }
- else if (target == first)
- ; /* We didn't do anything. */
- else
- {
- /* Save the values now, as the edge may get removed. */
- gcov_type edge_count = e->count;
- int edge_probability = e->probability;
- int edge_frequency;
- int n = 0;
-
- /* Don't force if target is exit block. */
- if (threaded && target != EXIT_BLOCK_PTR)
- {
- notice_new_block (redirect_edge_and_branch_force (e, target));
- if (dump_file)
- fprintf (dump_file, "Conditionals threaded.\n");
- }
- else if (!redirect_edge_and_branch (e, target))
- {
- if (dump_file)
- fprintf (dump_file,
- "Forwarding edge %i->%i to %i failed.\n",
- b->index, e->dest->index, target->index);
- ei_next (&ei);
- continue;
- }
-
- /* We successfully forwarded the edge. Now update profile
- data: for each edge we traversed in the chain, remove
- the original edge's execution count. */
- edge_frequency = ((edge_probability * b->frequency
- + REG_BR_PROB_BASE / 2)
- / REG_BR_PROB_BASE);
-
- if (!FORWARDER_BLOCK_P (b) && forwarder_block_p (b))
- b->flags |= BB_FORWARDER_BLOCK;
-
- do
- {
- edge t;
-
- if (!single_succ_p (first))
- {
- gcc_assert (n < nthreaded_edges);
- t = threaded_edges [n++];
- gcc_assert (t->src == first);
- update_bb_profile_for_threading (first, edge_frequency,
- edge_count, t);
- update_br_prob_note (first);
- }
- else
- {
- first->count -= edge_count;
- if (first->count < 0)
- first->count = 0;
- first->frequency -= edge_frequency;
- if (first->frequency < 0)
- first->frequency = 0;
- /* It is possible that as the result of
- threading we've removed edge as it is
- threaded to the fallthru edge. Avoid
- getting out of sync. */
- if (n < nthreaded_edges
- && first == threaded_edges [n]->src)
- n++;
- t = single_succ_edge (first);
- }
-
- t->count -= edge_count;
- if (t->count < 0)
- t->count = 0;
- first = t->dest;
- }
- while (first != target);
-
- changed = true;
- continue;
- }
- ei_next (&ei);
- }
-
- if (threaded_edges)
- free (threaded_edges);
- return changed;
-}
-
-
-/* Blocks A and B are to be merged into a single block. A has no incoming
- fallthru edge, so it can be moved before B without adding or modifying
- any jumps (aside from the jump from A to B). */
-
-static void
-merge_blocks_move_predecessor_nojumps (basic_block a, basic_block b)
-{
- rtx barrier;
- bool only_notes;
-
- /* If we are partitioning hot/cold basic blocks, we don't want to
- mess up unconditional or indirect jumps that cross between hot
- and cold sections.
-
- Basic block partitioning may result in some jumps that appear to
- be optimizable (or blocks that appear to be mergeable), but which really
- must be left untouched (they are required to make it safely across
- partition boundaries). See the comments at the top of
- bb-reorder.c:partition_hot_cold_basic_blocks for complete details. */
-
- if (BB_PARTITION (a) != BB_PARTITION (b))
- return;
-
- barrier = next_nonnote_insn (BB_END (a));
- gcc_assert (BARRIER_P (barrier));
- delete_insn (barrier);
-
- /* Move block and loop notes out of the chain so that we do not
- disturb their order.
-
- ??? A better solution would be to squeeze out all the non-nested notes
- and adjust the block trees appropriately. Even better would be to have
- a tighter connection between block trees and rtl so that this is not
- necessary. */
- only_notes = squeeze_notes (&BB_HEAD (a), &BB_END (a));
- gcc_assert (!only_notes);
-
- /* Scramble the insn chain. */
- if (BB_END (a) != PREV_INSN (BB_HEAD (b)))
- reorder_insns_nobb (BB_HEAD (a), BB_END (a), PREV_INSN (BB_HEAD (b)));
- a->flags |= BB_DIRTY;
-
- if (dump_file)
- fprintf (dump_file, "Moved block %d before %d and merged.\n",
- a->index, b->index);
-
- /* Swap the records for the two blocks around. */
-
- unlink_block (a);
- link_block (a, b->prev_bb);
-
- /* Now blocks A and B are contiguous. Merge them. */
- merge_blocks (a, b);
-}
-
-/* Blocks A and B are to be merged into a single block. B has no outgoing
- fallthru edge, so it can be moved after A without adding or modifying
- any jumps (aside from the jump from A to B). */
-
-static void
-merge_blocks_move_successor_nojumps (basic_block a, basic_block b)
-{
- rtx barrier, real_b_end;
- rtx label, table;
- bool only_notes;
-
- /* If we are partitioning hot/cold basic blocks, we don't want to
- mess up unconditional or indirect jumps that cross between hot
- and cold sections.
-
- Basic block partitioning may result in some jumps that appear to
- be optimizable (or blocks that appear to be mergeable), but which really
- must be left untouched (they are required to make it safely across
- partition boundaries). See the comments at the top of
- bb-reorder.c:partition_hot_cold_basic_blocks for complete details. */
-
- if (BB_PARTITION (a) != BB_PARTITION (b))
- return;
-
- real_b_end = BB_END (b);
-
- /* If there is a jump table following block B temporarily add the jump table
- to block B so that it will also be moved to the correct location. */
- if (tablejump_p (BB_END (b), &label, &table)
- && prev_active_insn (label) == BB_END (b))
- {
- BB_END (b) = table;
- }
-
- /* There had better have been a barrier there. Delete it. */
- barrier = NEXT_INSN (BB_END (b));
- if (barrier && BARRIER_P (barrier))
- delete_insn (barrier);
-
- /* Move block and loop notes out of the chain so that we do not
- disturb their order.
-
- ??? A better solution would be to squeeze out all the non-nested notes
- and adjust the block trees appropriately. Even better would be to have
- a tighter connection between block trees and rtl so that this is not
- necessary. */
- only_notes = squeeze_notes (&BB_HEAD (b), &BB_END (b));
- gcc_assert (!only_notes);
-
-
- /* Scramble the insn chain. */
- reorder_insns_nobb (BB_HEAD (b), BB_END (b), BB_END (a));
-
- /* Restore the real end of b. */
- BB_END (b) = real_b_end;
-
- if (dump_file)
- fprintf (dump_file, "Moved block %d after %d and merged.\n",
- b->index, a->index);
-
- /* Now blocks A and B are contiguous. Merge them. */
- merge_blocks (a, b);
-}
-
-/* Attempt to merge basic blocks that are potentially non-adjacent.
- Return NULL iff the attempt failed, otherwise return basic block
- where cleanup_cfg should continue. Because the merging commonly
- moves basic block away or introduces another optimization
- possibility, return basic block just before B so cleanup_cfg don't
- need to iterate.
-
- It may be good idea to return basic block before C in the case
- C has been moved after B and originally appeared earlier in the
- insn sequence, but we have no information available about the
- relative ordering of these two. Hopefully it is not too common. */
-
-static basic_block
-merge_blocks_move (edge e, basic_block b, basic_block c, int mode)
-{
- basic_block next;
-
- /* If we are partitioning hot/cold basic blocks, we don't want to
- mess up unconditional or indirect jumps that cross between hot
- and cold sections.
-
- Basic block partitioning may result in some jumps that appear to
- be optimizable (or blocks that appear to be mergeable), but which really
- must be left untouched (they are required to make it safely across
- partition boundaries). See the comments at the top of
- bb-reorder.c:partition_hot_cold_basic_blocks for complete details. */
-
- if (BB_PARTITION (b) != BB_PARTITION (c))
- return NULL;
-
-
-
- /* If B has a fallthru edge to C, no need to move anything. */
- if (e->flags & EDGE_FALLTHRU)
- {
- int b_index = b->index, c_index = c->index;
- merge_blocks (b, c);
- update_forwarder_flag (b);
-
- if (dump_file)
- fprintf (dump_file, "Merged %d and %d without moving.\n",
- b_index, c_index);
-
- return b->prev_bb == ENTRY_BLOCK_PTR ? b : b->prev_bb;
- }
-
- /* Otherwise we will need to move code around. Do that only if expensive
- transformations are allowed. */
- else if (mode & CLEANUP_EXPENSIVE)
- {
- edge tmp_edge, b_fallthru_edge;
- bool c_has_outgoing_fallthru;
- bool b_has_incoming_fallthru;
- edge_iterator ei;
-
- /* Avoid overactive code motion, as the forwarder blocks should be
- eliminated by edge redirection instead. One exception might have
- been if B is a forwarder block and C has no fallthru edge, but
- that should be cleaned up by bb-reorder instead. */
- if (FORWARDER_BLOCK_P (b) || FORWARDER_BLOCK_P (c))
- return NULL;
-
- /* We must make sure to not munge nesting of lexical blocks,
- and loop notes. This is done by squeezing out all the notes
- and leaving them there to lie. Not ideal, but functional. */
-
- FOR_EACH_EDGE (tmp_edge, ei, c->succs)
- if (tmp_edge->flags & EDGE_FALLTHRU)
- break;
-
- c_has_outgoing_fallthru = (tmp_edge != NULL);
-
- FOR_EACH_EDGE (tmp_edge, ei, b->preds)
- if (tmp_edge->flags & EDGE_FALLTHRU)
- break;
-
- b_has_incoming_fallthru = (tmp_edge != NULL);
- b_fallthru_edge = tmp_edge;
- next = b->prev_bb;
- if (next == c)
- next = next->prev_bb;
-
- /* Otherwise, we're going to try to move C after B. If C does
- not have an outgoing fallthru, then it can be moved
- immediately after B without introducing or modifying jumps. */
- if (! c_has_outgoing_fallthru)
- {
- merge_blocks_move_successor_nojumps (b, c);
- return next == ENTRY_BLOCK_PTR ? next->next_bb : next;
- }
-
- /* If B does not have an incoming fallthru, then it can be moved
- immediately before C without introducing or modifying jumps.
- C cannot be the first block, so we do not have to worry about
- accessing a non-existent block. */
-
- if (b_has_incoming_fallthru)
- {
- basic_block bb;
-
- if (b_fallthru_edge->src == ENTRY_BLOCK_PTR)
- return NULL;
- bb = force_nonfallthru (b_fallthru_edge);
- if (bb)
- notice_new_block (bb);
- }
+static bool set_dest_equiv_p (rtx x, rtx y, struct equiv_info *info);
+static bool set_dest_addr_equiv_p (rtx x, rtx y, struct equiv_info *info);
+static void find_dying_inputs (struct equiv_info *info);
+static bool resolve_input_conflict (struct equiv_info *info);
- merge_blocks_move_predecessor_nojumps (b, c);
- return next == ENTRY_BLOCK_PTR ? next->next_bb : next;
- }
+/* After reload, some moves, as indicated by SECONDARY_RELOAD_CLASS and
+ SECONDARY_MEMORY_NEEDED, cannot be done directly. For our purposes, we
+ consider them impossible to generate after reload (even though some
+ might be synthesized when you throw enough code at them).
+ Since we don't know while procesing a cross-jump if a local register
+ that is currently live will eventually be live and thus be an input,
+ we keep track of potential inputs that would require an impossible move
+ by using a prohibitively high cost for them.
+ This number, multiplied with the larger of STRUCT_EQUIV_MAX_LOCAL and
+ FIRST_PSEUDO_REGISTER, must fit in the input_cost field of
+ struct equiv_info. */
+#define IMPOSSIBLE_MOVE_FACTOR 20000
- return NULL;
-}
/* Removes the memory attributes of MEM expression
@@ -972,1245 +183,1040 @@ merge_memattrs (rtx x, rtx y)
return;
}
-
-/* Return true if I1 and I2 are equivalent and thus can be crossjumped. */
-
-static bool
-insns_match_p (int mode ATTRIBUTE_UNUSED, rtx i1, rtx i2)
+/* In SET, assign the bit for the register number of REG the value VALUE.
+ If REG is a hard register, do so for all its consituent registers.
+ Return the number of registers that have become included (as a positive
+ number) or excluded (as a negative number). */
+static int
+assign_reg_reg_set (regset set, rtx reg, int value)
{
- rtx p1, p2;
-
- /* Verify that I1 and I2 are equivalent. */
- if (GET_CODE (i1) != GET_CODE (i2))
- return false;
-
- p1 = PATTERN (i1);
- p2 = PATTERN (i2);
-
- if (GET_CODE (p1) != GET_CODE (p2))
- return false;
-
- /* If this is a CALL_INSN, compare register usage information.
- If we don't check this on stack register machines, the two
- CALL_INSNs might be merged leaving reg-stack.c with mismatching
- numbers of stack registers in the same basic block.
- If we don't check this on machines with delay slots, a delay slot may
- be filled that clobbers a parameter expected by the subroutine.
-
- ??? We take the simple route for now and assume that if they're
- equal, they were constructed identically. */
-
- if (CALL_P (i1)
- && (!rtx_equal_p (CALL_INSN_FUNCTION_USAGE (i1),
- CALL_INSN_FUNCTION_USAGE (i2))
- || SIBLING_CALL_P (i1) != SIBLING_CALL_P (i2)))
- return false;
-
-#ifdef STACK_REGS
- /* If cross_jump_death_matters is not 0, the insn's mode
- indicates whether or not the insn contains any stack-like
- regs. */
+ unsigned regno = REGNO (reg);
+ int nregs, i, old;
- if ((mode & CLEANUP_POST_REGSTACK) && stack_regs_mentioned (i1))
+ if (regno >= FIRST_PSEUDO_REGISTER)
{
- /* If register stack conversion has already been done, then
- death notes must also be compared before it is certain that
- the two instruction streams match. */
-
- rtx note;
- HARD_REG_SET i1_regset, i2_regset;
-
- CLEAR_HARD_REG_SET (i1_regset);
- CLEAR_HARD_REG_SET (i2_regset);
-
- for (note = REG_NOTES (i1); note; note = XEXP (note, 1))
- if (REG_NOTE_KIND (note) == REG_DEAD && STACK_REG_P (XEXP (note, 0)))
- SET_HARD_REG_BIT (i1_regset, REGNO (XEXP (note, 0)));
-
- for (note = REG_NOTES (i2); note; note = XEXP (note, 1))
- if (REG_NOTE_KIND (note) == REG_DEAD && STACK_REG_P (XEXP (note, 0)))
- SET_HARD_REG_BIT (i2_regset, REGNO (XEXP (note, 0)));
-
- GO_IF_HARD_REG_EQUAL (i1_regset, i2_regset, done);
-
- return false;
-
- done:
- ;
+ gcc_assert (!reload_completed);
+ nregs = 1;
}
-#endif
-
- if (reload_completed
- ? rtx_renumbered_equal_p (p1, p2) : rtx_equal_p (p1, p2))
- return true;
-
- /* Do not do EQUIV substitution after reload. First, we're undoing the
- work of reload_cse. Second, we may be undoing the work of the post-
- reload splitting pass. */
- /* ??? Possibly add a new phase switch variable that can be used by
- targets to disallow the troublesome insns after splitting. */
- if (!reload_completed)
+ else
+ nregs = hard_regno_nregs[regno][GET_MODE (reg)];
+ for (old = 0, i = nregs; --i >= 0; regno++)
{
- /* The following code helps take care of G++ cleanups. */
- rtx equiv1 = find_reg_equal_equiv_note (i1);
- rtx equiv2 = find_reg_equal_equiv_note (i2);
-
- if (equiv1 && equiv2
- /* If the equivalences are not to a constant, they may
- reference pseudos that no longer exist, so we can't
- use them. */
- && (! reload_completed
- || (CONSTANT_P (XEXP (equiv1, 0))
- && rtx_equal_p (XEXP (equiv1, 0), XEXP (equiv2, 0)))))
- {
- rtx s1 = single_set (i1);
- rtx s2 = single_set (i2);
- if (s1 != 0 && s2 != 0
- && rtx_renumbered_equal_p (SET_DEST (s1), SET_DEST (s2)))
- {
- validate_change (i1, &SET_SRC (s1), XEXP (equiv1, 0), 1);
- validate_change (i2, &SET_SRC (s2), XEXP (equiv2, 0), 1);
- if (! rtx_renumbered_equal_p (p1, p2))
- cancel_changes (0);
- else if (apply_change_group ())
- return true;
- }
- }
+ if ((value != 0) == REGNO_REG_SET_P (set, regno))
+ continue;
+ if (value)
+ old++, SET_REGNO_REG_SET (set, regno);
+ else
+ old--, CLEAR_REGNO_REG_SET (set, regno);
}
-
- return false;
+ return old;
}
-
-/* Look through the insns at the end of BB1 and BB2 and find the longest
- sequence that are equivalent. Store the first insns for that sequence
- in *F1 and *F2 and return the sequence length.
-
- To simplify callers of this function, if the blocks match exactly,
- store the head of the blocks in *F1 and *F2. */
-static int
-flow_find_cross_jump (int mode ATTRIBUTE_UNUSED, basic_block bb1,
- basic_block bb2, rtx *f1, rtx *f2)
+/* Record state about current inputs / local registers / lifeness
+ in *P. */
+static inline void
+struct_equiv_make_checkpoint (struct struct_equiv_checkpoint *p,
+ struct equiv_info *info)
{
- rtx i1, i2, last1, last2, afterlast1, afterlast2;
- int ninsns = 0;
-
- /* Skip simple jumps at the end of the blocks. Complex jumps still
- need to be compared for equivalence, which we'll do below. */
-
- i1 = BB_END (bb1);
- last1 = afterlast1 = last2 = afterlast2 = NULL_RTX;
- if (onlyjump_p (i1)
- || (returnjump_p (i1) && !side_effects_p (PATTERN (i1))))
- {
- last1 = i1;
- i1 = PREV_INSN (i1);
- }
+ *p = info->cur;
+}
- i2 = BB_END (bb2);
- if (onlyjump_p (i2)
- || (returnjump_p (i2) && !side_effects_p (PATTERN (i2))))
+/* Call struct_equiv_make_checkpoint (P, INFO) if the current partial block
+ is suitable to split off - i.e. there is no dangling cc0 user - and
+ if the current cost of the common instructions, minus the cost for
+ setting up the inputs, is higher than what has been recorded before
+ in CHECKPOINT[N]. Also, if we do so, confirm or cancel any pending
+ changes. */
+static void
+struct_equiv_improve_checkpoint (struct struct_equiv_checkpoint *p,
+ struct equiv_info *info)
+{
+#ifdef HAVE_cc0
+ if (reg_mentioned_p (cc0_rtx, info->x_start) && !sets_cc0_p (info->x_start))
+ return;
+#endif
+ if (info->cur.input_count >= IMPOSSIBLE_MOVE_FACTOR)
+ return;
+ if (info->input_cost >= 0
+ ? (COSTS_N_INSNS(info->cur.ninsns - p->ninsns)
+ > info->input_cost * (info->cur.input_count - p->input_count))
+ : info->cur.ninsns > p->ninsns && !info->cur.input_count)
{
- last2 = i2;
- /* Count everything except for unconditional jump as insn. */
- if (!simplejump_p (i2) && !returnjump_p (i2) && last1)
- ninsns++;
- i2 = PREV_INSN (i2);
+ if (info->check_input_conflict && ! resolve_input_conflict (info))
+ return;
+ /* We have a profitable set of changes. If this is the final pass,
+ commit them now. Otherwise, we don't know yet if we can make any
+ change, so put the old code back for now. */
+ if (info->mode & STRUCT_EQUIV_FINAL)
+ confirm_change_group ();
+ else
+ cancel_changes (0);
+ struct_equiv_make_checkpoint (p, info);
}
+}
- while (true)
+/* Restore state about current inputs / local registers / lifeness
+ from P. */
+static void
+struct_equiv_restore_checkpoint (struct struct_equiv_checkpoint *p,
+ struct equiv_info *info)
+{
+ info->cur.ninsns = p->ninsns;
+ info->cur.x_start = p->x_start;
+ info->cur.y_start = p->y_start;
+ info->cur.input_count = p->input_count;
+ info->cur.input_valid = p->input_valid;
+ while (info->cur.local_count > p->local_count)
{
- /* Ignore notes. */
- while (!INSN_P (i1) && i1 != BB_HEAD (bb1))
- i1 = PREV_INSN (i1);
-
- while (!INSN_P (i2) && i2 != BB_HEAD (bb2))
- i2 = PREV_INSN (i2);
-
- if (i1 == BB_HEAD (bb1) || i2 == BB_HEAD (bb2))
- break;
-
- if (!insns_match_p (mode, i1, i2))
- break;
-
- merge_memattrs (i1, i2);
-
- /* Don't begin a cross-jump with a NOTE insn. */
- if (INSN_P (i1))
+ info->cur.local_count--;
+ info->cur.version--;
+ if (REGNO_REG_SET_P (info->x_local_live,
+ REGNO (info->x_local[info->cur.local_count])))
{
- /* If the merged insns have different REG_EQUAL notes, then
- remove them. */
- rtx equiv1 = find_reg_equal_equiv_note (i1);
- rtx equiv2 = find_reg_equal_equiv_note (i2);
-
- if (equiv1 && !equiv2)
- remove_note (i1, equiv1);
- else if (!equiv1 && equiv2)
- remove_note (i2, equiv2);
- else if (equiv1 && equiv2
- && !rtx_equal_p (XEXP (equiv1, 0), XEXP (equiv2, 0)))
- {
- remove_note (i1, equiv1);
- remove_note (i2, equiv2);
- }
-
- afterlast1 = last1, afterlast2 = last2;
- last1 = i1, last2 = i2;
- ninsns++;
+ assign_reg_reg_set (info->x_local_live,
+ info->x_local[info->cur.local_count], 0);
+ assign_reg_reg_set (info->y_local_live,
+ info->y_local[info->cur.local_count], 0);
+ info->cur.version--;
}
-
- i1 = PREV_INSN (i1);
- i2 = PREV_INSN (i2);
- }
-
-#ifdef HAVE_cc0
- /* Don't allow the insn after a compare to be shared by
- cross-jumping unless the compare is also shared. */
- if (ninsns && reg_mentioned_p (cc0_rtx, last1) && ! sets_cc0_p (last1))
- last1 = afterlast1, last2 = afterlast2, ninsns--;
-#endif
-
- /* Include preceding notes and labels in the cross-jump. One,
- this may bring us to the head of the blocks as requested above.
- Two, it keeps line number notes as matched as may be. */
- if (ninsns)
- {
- while (last1 != BB_HEAD (bb1) && !INSN_P (PREV_INSN (last1)))
- last1 = PREV_INSN (last1);
-
- if (last1 != BB_HEAD (bb1) && LABEL_P (PREV_INSN (last1)))
- last1 = PREV_INSN (last1);
-
- while (last2 != BB_HEAD (bb2) && !INSN_P (PREV_INSN (last2)))
- last2 = PREV_INSN (last2);
-
- if (last2 != BB_HEAD (bb2) && LABEL_P (PREV_INSN (last2)))
- last2 = PREV_INSN (last2);
-
- *f1 = last1;
- *f2 = last2;
}
-
- return ninsns;
+ if (info->cur.version != p->version)
+ info->need_rerun = true;
}
-/* Return true iff outgoing edges of BB1 and BB2 match, together with
- the branch instruction. This means that if we commonize the control
- flow before end of the basic block, the semantic remains unchanged.
-
- We may assume that there exists one edge with a common destination. */
-static bool
-outgoing_edges_match (int mode, basic_block bb1, basic_block bb2)
+/* Check if *xp is equivalent to Y. Until an an unreconcilable difference is
+ found, use in-group changes with validate_change on *xp to make register
+ assignments agree. It is the (not necessarily direct) callers
+ responsibility to verify / confirm / cancel these changes, as appropriate.
+ RVALUE indicates if the processed piece of rtl is used as a destination, in
+ which case we can't have different registers being an input. Returns
+ nonzero if the two blocks have been identified as equivalent, zero otherwise.
+ RVALUE == 0: destination
+ RVALUE == 1: source
+ RVALUE == -1: source, ignore SET_DEST of SET / clobber. */
+bool
+rtx_equiv_p (rtx *xp, rtx y, enum equiv_rvalue rvalue, struct equiv_info *info)
{
- int nehedges1 = 0, nehedges2 = 0;
- edge fallthru1 = 0, fallthru2 = 0;
- edge e1, e2;
- edge_iterator ei;
-
- /* If BB1 has only one successor, we may be looking at either an
- unconditional jump, or a fake edge to exit. */
- if (single_succ_p (bb1)
- && (single_succ_edge (bb1)->flags & (EDGE_COMPLEX | EDGE_FAKE)) == 0
- && (!JUMP_P (BB_END (bb1)) || simplejump_p (BB_END (bb1))))
- return (single_succ_p (bb2)
- && (single_succ_edge (bb2)->flags
- & (EDGE_COMPLEX | EDGE_FAKE)) == 0
- && (!JUMP_P (BB_END (bb2)) || simplejump_p (BB_END (bb2))));
+ rtx x = *xp;
+ enum rtx_code code;
+ int length;
+ const char *format;
+ int i;
- /* Match conditional jumps - this may get tricky when fallthru and branch
- edges are crossed. */
- if (EDGE_COUNT (bb1->succs) == 2
- && any_condjump_p (BB_END (bb1))
- && onlyjump_p (BB_END (bb1)))
+ if (!y || !x)
+ return x == y;
+ code = GET_CODE (y);
+ if (code != REG && x == y)
+ return true;
+ if (reload_completed)
{
- edge b1, f1, b2, f2;
- bool reverse, match;
- rtx set1, set2, cond1, cond2;
- enum rtx_code code1, code2;
-
- if (EDGE_COUNT (bb2->succs) != 2
- || !any_condjump_p (BB_END (bb2))
- || !onlyjump_p (BB_END (bb2)))
- return false;
+ gcc_assert (GET_CODE (x) != SUBREG || !REG_P (SUBREG_REG (x)));
+ gcc_assert (code != SUBREG || !REG_P (SUBREG_REG (y)));
+ gcc_assert (GET_CODE (x) != REG || REGNO (x) < FIRST_PSEUDO_REGISTER);
+ gcc_assert (code != REG || REGNO (y) < FIRST_PSEUDO_REGISTER);
+ }
+ if (GET_CODE (x) != code
+ || GET_MODE (x) != GET_MODE (y))
+ return false;
+ /* ??? could extend to allow CONST_INT inputs. */
+ switch (code)
+ {
+ case REG:
+ {
+ unsigned x_regno = REGNO (x);
+ unsigned y_regno = REGNO (y);
+ int x_common_live, y_common_live;
- b1 = BRANCH_EDGE (bb1);
- b2 = BRANCH_EDGE (bb2);
- f1 = FALLTHRU_EDGE (bb1);
- f2 = FALLTHRU_EDGE (bb2);
+ if (reload_completed)
+ gcc_assert (x_regno < FIRST_PSEUDO_REGISTER
+ && y_regno < FIRST_PSEUDO_REGISTER);
- /* Get around possible forwarders on fallthru edges. Other cases
- should be optimized out already. */
- if (FORWARDER_BLOCK_P (f1->dest))
- f1 = single_succ_edge (f1->dest);
+ /* If the register is a locally live one in one block, the
+ corresponding one must be locally live in the other, too, and
+ match of identical regnos doesn't apply. */
+ if (REGNO_REG_SET_P (info->x_local_live, x_regno))
+ {
+ if (!REGNO_REG_SET_P (info->y_local_live, y_regno))
+ return false;
+ }
+ else if (REGNO_REG_SET_P (info->y_local_live, y_regno))
+ return false;
+ else if (x_regno == y_regno)
+ {
+ if (!rvalue && info->cur.input_valid
+ && (reg_overlap_mentioned_p (x, info->x_input)
+ || reg_overlap_mentioned_p (x, info->y_input)))
+ return false;
- if (FORWARDER_BLOCK_P (f2->dest))
- f2 = single_succ_edge (f2->dest);
+ /* Update liveness information. */
+ if (info->live_update
+ && assign_reg_reg_set (info->common_live, x, rvalue))
+ info->cur.version++;
- /* To simplify use of this function, return false if there are
- unneeded forwarder blocks. These will get eliminated later
- during cleanup_cfg. */
- if (FORWARDER_BLOCK_P (f1->dest)
- || FORWARDER_BLOCK_P (f2->dest)
- || FORWARDER_BLOCK_P (b1->dest)
- || FORWARDER_BLOCK_P (b2->dest))
- return false;
+ return true;
+ }
- if (f1->dest == f2->dest && b1->dest == b2->dest)
- reverse = false;
- else if (f1->dest == b2->dest && b1->dest == f2->dest)
- reverse = true;
- else
- return false;
+ x_common_live = REGNO_REG_SET_P (info->common_live, x_regno);
+ y_common_live = REGNO_REG_SET_P (info->common_live, y_regno);
+ if (x_common_live != y_common_live)
+ return false;
+ else if (x_common_live)
+ {
+ if (! rvalue || info->input_cost < 0 || no_new_pseudos)
+ return false;
+ /* If info->live_update is not set, we are processing notes.
+ We then allow a match with x_input / y_input found in a
+ previous pass. */
+ if (info->live_update && !info->cur.input_valid)
+ {
+ info->cur.input_valid = true;
+ info->x_input = x;
+ info->y_input = y;
+ info->cur.input_count += optimize_size ? 2 : 1;
+ if (info->input_reg
+ && GET_MODE (info->input_reg) != GET_MODE (info->x_input))
+ info->input_reg = NULL_RTX;
+ if (!info->input_reg)
+ info->input_reg = gen_reg_rtx (GET_MODE (info->x_input));
+ }
+ else if ((info->live_update
+ ? ! info->cur.input_valid : ! info->x_input)
+ || ! rtx_equal_p (x, info->x_input)
+ || ! rtx_equal_p (y, info->y_input))
+ return false;
+ validate_change (info->cur.x_start, xp, info->input_reg, 1);
+ }
+ else
+ {
+ int x_nregs = (x_regno >= FIRST_PSEUDO_REGISTER
+ ? 1 : hard_regno_nregs[x_regno][GET_MODE (x)]);
+ int y_nregs = (y_regno >= FIRST_PSEUDO_REGISTER
+ ? 1 : hard_regno_nregs[y_regno][GET_MODE (y)]);
+ int size = GET_MODE_SIZE (GET_MODE (x));
+ enum machine_mode x_mode = GET_MODE (x);
+ unsigned x_regno_i, y_regno_i;
+ int x_nregs_i, y_nregs_i, size_i;
+ int change;
+ int local_count = info->cur.local_count;
- set1 = pc_set (BB_END (bb1));
- set2 = pc_set (BB_END (bb2));
- if ((XEXP (SET_SRC (set1), 1) == pc_rtx)
- != (XEXP (SET_SRC (set2), 1) == pc_rtx))
- reverse = !reverse;
+ /* This might be a register local to each block. See if we have
+ it already registered. */
+ for (i = local_count - 1; i >= 0; i--)
+ {
+ x_regno_i = REGNO (info->x_local[i]);
+ x_nregs_i = (x_regno_i >= FIRST_PSEUDO_REGISTER
+ ? 1 : hard_regno_nregs[x_regno_i][GET_MODE (x)]);
+ y_regno_i = REGNO (info->y_local[i]);
+ y_nregs_i = (y_regno_i >= FIRST_PSEUDO_REGISTER
+ ? 1 : hard_regno_nregs[y_regno_i][GET_MODE (y)]);
+ size_i = GET_MODE_SIZE (GET_MODE (info->x_local[i]));
- cond1 = XEXP (SET_SRC (set1), 0);
- cond2 = XEXP (SET_SRC (set2), 0);
- code1 = GET_CODE (cond1);
- if (reverse)
- code2 = reversed_comparison_code (cond2, BB_END (bb2));
- else
- code2 = GET_CODE (cond2);
+ /* If we have a new pair of registers that is wider than an
+ old pair and enclosing it with matching offsets,
+ remove the old pair. If we find a matching, wider, old
+ pair, use the old one. If the width is the same, use the
+ old one if the modes match, but the new if they don't.
+ We don't want to get too fancy with subreg_regno_offset
+ here, so we just test two straightforwad cases each. */
+ if (info->live_update
+ && (x_mode != GET_MODE (info->x_local[i])
+ ? size >= size_i : size > size_i))
+ {
+ /* If the new pair is fully enclosing a matching
+ existing pair, remove the old one. N.B. because
+ we are removing one entry here, the check below
+ if we have space for a new entry will succeed. */
+ if ((x_regno <= x_regno_i
+ && x_regno + x_nregs >= x_regno_i + x_nregs_i
+ && x_nregs == y_nregs && x_nregs_i == y_nregs_i
+ && x_regno - x_regno_i == y_regno - y_regno_i)
+ || (x_regno == x_regno_i && y_regno == y_regno_i
+ && x_nregs >= x_nregs_i && y_nregs >= y_nregs_i))
+ {
+ info->cur.local_count = --local_count;
+ info->x_local[i] = info->x_local[local_count];
+ info->y_local[i] = info->y_local[local_count];
+ continue;
+ }
+ }
+ else
+ {
- if (code2 == UNKNOWN)
- return false;
+ /* If the new pair is fully enclosed within a matching
+ existing pair, succeed. */
+ if (x_regno >= x_regno_i
+ && x_regno + x_nregs <= x_regno_i + x_nregs_i
+ && x_nregs == y_nregs && x_nregs_i == y_nregs_i
+ && x_regno - x_regno_i == y_regno - y_regno_i)
+ break;
+ if (x_regno == x_regno_i && y_regno == y_regno_i
+ && x_nregs <= x_nregs_i && y_nregs <= y_nregs_i)
+ break;
+ }
- /* Verify codes and operands match. */
- match = ((code1 == code2
- && rtx_renumbered_equal_p (XEXP (cond1, 0), XEXP (cond2, 0))
- && rtx_renumbered_equal_p (XEXP (cond1, 1), XEXP (cond2, 1)))
- || (code1 == swap_condition (code2)
- && rtx_renumbered_equal_p (XEXP (cond1, 1),
- XEXP (cond2, 0))
- && rtx_renumbered_equal_p (XEXP (cond1, 0),
- XEXP (cond2, 1))));
+ /* Any other overlap causes a match failure. */
+ if (x_regno + x_nregs > x_regno_i
+ && x_regno_i + x_nregs_i > x_regno)
+ return false;
+ if (y_regno + y_nregs > y_regno_i
+ && y_regno_i + y_nregs_i > y_regno)
+ return false;
+ }
+ if (i < 0)
+ {
+ /* Not found. Create a new entry if possible. */
+ if (!info->live_update
+ || info->cur.local_count >= STRUCT_EQUIV_MAX_LOCAL)
+ return false;
+ info->x_local[info->cur.local_count] = x;
+ info->y_local[info->cur.local_count] = y;
+ info->cur.local_count++;
+ info->cur.version++;
+ }
+ assign_reg_reg_set (info->x_local_live, x, rvalue);
+ change = assign_reg_reg_set (info->y_local_live, y, rvalue);
+ if (change)
+ {
+ if (reload_completed)
+ {
+ if (secondary_reload_class
+ (0, REGNO_REG_CLASS (y_regno), x_mode, x) != NO_REGS
+#ifdef SECONDARY_MEMORY_NEEDED
+ || SECONDARY_MEMORY_NEEDED (REGNO_REG_CLASS (y_regno),
+ REGNO_REG_CLASS (x_regno),
+ x_mode)
+#endif
+ )
+ change *= IMPOSSIBLE_MOVE_FACTOR;
+ }
+ info->cur.input_count += change;
+ info->cur.version++;
+ }
+ }
+ return true;
+ }
+ case SET:
+ {
+ /* Process destination. */
+ if (rvalue < 0)
+ {
+ /* Ignore the destinations role as a destination. Still,
+ we have to consider input registers embedded in the addresses
+ of a MEM. Some special forms also make the entire destination
+ a source. */
+ enum rtx_code dst_code = GET_CODE (SET_DEST (x));
- /* If we return true, we will join the blocks. Which means that
- we will only have one branch prediction bit to work with. Thus
- we require the existing branches to have probabilities that are
- roughly similar. */
- if (match
- && !optimize_size
- && maybe_hot_bb_p (bb1)
- && maybe_hot_bb_p (bb2))
- {
- int prob2;
+ if ((dst_code != STRICT_LOW_PART
+ && dst_code != ZERO_EXTEND
+ && dst_code != SIGN_EXTEND)
+ ? !set_dest_addr_equiv_p (SET_DEST (x), SET_DEST (y), info)
+ : !rtx_equiv_p (&SET_DEST (x), SET_DEST (y), 1, info))
+ return false;
+ }
+ else if (!rtx_equiv_p (&SET_DEST (x), SET_DEST (y), 0, info))
+ return false;
+ /* Process source. */
+ return rtx_equiv_p (&SET_SRC (x), SET_SRC (y), 1, info);
+ }
+ case CLOBBER:
+ return rvalue < 0 || rtx_equiv_p (&XEXP (x, 0), XEXP (y, 0), 0, info);
+ /* The AUTO_INC / POST_MODIFY / PRE_MODIFY sets are modelled to take
+ place during input processing, however, that is benign, since they
+ are paired with reads.
+ We force 1 for rvalue in the MEM recursion to make sure that
+ PRE_MODIFY / POST_MODIFY SET_DESTS are seen at all. */
+ case MEM:
+ return !rvalue || rtx_equiv_p (&XEXP (x, 0), XEXP (y, 0), 1, info);
+ case POST_INC: case POST_DEC: case PRE_INC: case PRE_DEC:
+ return (rtx_equiv_p (&XEXP (x, 0), XEXP (y, 0), 0, info)
+ && rtx_equiv_p (&XEXP (x, 0), XEXP (y, 0), 1, info));
+ case PARALLEL:
+ {
+ int j;
- if (b1->dest == b2->dest)
- prob2 = b2->probability;
- else
- /* Do not use f2 probability as f2 may be forwarded. */
- prob2 = REG_BR_PROB_BASE - b2->probability;
+ if (!set_dest_equiv_p (x, y, info))
+ return false;
+ for (j = 0; j < XVECLEN (x, 0); ++j)
+ if (! rtx_equiv_p (&XVECEXP (x, 0, j), XVECEXP (y, 0, j), -1, info))
+ return false;
+ return true;
+ }
+ case LABEL_REF:
+ /* We can't assume nonlocal labels have their following insns yet. */
+ if (LABEL_REF_NONLOCAL_P (x) || LABEL_REF_NONLOCAL_P (y))
+ return XEXP (x, 0) == XEXP (y, 0);
- /* Fail if the difference in probabilities is greater than 50%.
- This rules out two well-predicted branches with opposite
- outcomes. */
- if (abs (b1->probability - prob2) > REG_BR_PROB_BASE / 2)
- {
- if (dump_file)
- fprintf (dump_file,
- "Outcomes of branch in bb %i and %i differ too much (%i %i)\n",
- bb1->index, bb2->index, b1->probability, prob2);
+ /* Two label-refs are equivalent if they point at labels
+ in the same position in the instruction stream. */
+ return (next_real_insn (XEXP (x, 0))
+ == next_real_insn (XEXP (y, 0)));
+ case SYMBOL_REF:
+ return XSTR (x, 0) == XSTR (y, 0);
+ /* Some rtl is guaranteed to be shared, or unique; If we didn't match
+ EQ equality above, they aren't the same. */
+ case CONST_INT:
+ case CODE_LABEL:
+ return false;
+ default:
+ break;
+ }
- return false;
- }
- }
+ /* For commutative operations, the RTX match if the operands match in any
+ order. */
+ if (targetm.commutative_p (x, UNKNOWN))
+ return ((rtx_equiv_p (&XEXP (x, 0), XEXP (y, 0), rvalue, info)
+ && rtx_equiv_p (&XEXP (x, 1), XEXP (y, 1), rvalue, info))
+ || (rtx_equiv_p (&XEXP (x, 0), XEXP (y, 1), rvalue, info)
+ && rtx_equiv_p (&XEXP (x, 1), XEXP (y, 0), rvalue, info)));
- if (dump_file && match)
- fprintf (dump_file, "Conditionals in bb %i and %i match.\n",
- bb1->index, bb2->index);
+ /* Process subexpressions - this is similar to rtx_equal_p. */
+ length = GET_RTX_LENGTH (code);
+ format = GET_RTX_FORMAT (code);
- return match;
+ for (i = 0; i < length; ++i)
+ {
+ switch (format[i])
+ {
+ case 'w':
+ if (XWINT (x, i) != XWINT (y, i))
+ return false;
+ break;
+ case 'n':
+ case 'i':
+ if (XINT (x, i) != XINT (y, i))
+ return false;
+ break;
+ case 'V':
+ case 'E':
+ if (XVECLEN (x, i) != XVECLEN (y, i))
+ return false;
+ if (XVEC (x, i) != 0)
+ {
+ int j;
+ for (j = 0; j < XVECLEN (x, i); ++j)
+ {
+ if (! rtx_equiv_p (&XVECEXP (x, i, j), XVECEXP (y, i, j),
+ rvalue, info))
+ return false;
+ }
+ }
+ break;
+ case 'e':
+ if (! rtx_equiv_p (&XEXP (x, i), XEXP (y, i), rvalue, info))
+ return false;
+ break;
+ case 'S':
+ case 's':
+ if ((XSTR (x, i) || XSTR (y, i))
+ && (! XSTR (x, i) || ! XSTR (y, i)
+ || strcmp (XSTR (x, i), XSTR (y, i))))
+ return false;
+ break;
+ case 'u':
+ /* These are just backpointers, so they don't matter. */
+ break;
+ case '0':
+ case 't':
+ break;
+ /* It is believed that rtx's at this level will never
+ contain anything but integers and other rtx's,
+ except for within LABEL_REFs and SYMBOL_REFs. */
+ default:
+ gcc_unreachable ();
+ }
}
+ return true;
+}
- /* Generic case - we are seeing a computed jump, table jump or trapping
- instruction. */
-
- /* Check whether there are tablejumps in the end of BB1 and BB2.
- Return true if they are identical. */
+/* Do only the rtx_equiv_p SET_DEST processing for SETs and CLOBBERs.
+ Since we are scanning backwards, this the first step in processing each
+ insn. Return true for success. */
+static bool
+set_dest_equiv_p (rtx x, rtx y, struct equiv_info *info)
+{
+ if (!x || !y)
+ return x == y;
+ if (GET_CODE (x) == SET || GET_CODE (x) == CLOBBER)
+ return rtx_equiv_p (&XEXP (x, 0), XEXP (y, 0), 0, info);
+ else if (GET_CODE (x) == PARALLEL)
{
- rtx label1, label2;
- rtx table1, table2;
+ int j;
- if (tablejump_p (BB_END (bb1), &label1, &table1)
- && tablejump_p (BB_END (bb2), &label2, &table2)
- && GET_CODE (PATTERN (table1)) == GET_CODE (PATTERN (table2)))
+ if (XVECLEN (x, 0) != XVECLEN (y, 0))
+ return false;
+ for (j = 0; j < XVECLEN (x, 0); ++j)
{
- /* The labels should never be the same rtx. If they really are same
- the jump tables are same too. So disable crossjumping of blocks BB1
- and BB2 because when deleting the common insns in the end of BB1
- by delete_basic_block () the jump table would be deleted too. */
- /* If LABEL2 is referenced in BB1->END do not do anything
- because we would loose information when replacing
- LABEL1 by LABEL2 and then LABEL2 by LABEL1 in BB1->END. */
- if (label1 != label2 && !rtx_referenced_p (label2, BB_END (bb1)))
- {
- /* Set IDENTICAL to true when the tables are identical. */
- bool identical = false;
- rtx p1, p2;
-
- p1 = PATTERN (table1);
- p2 = PATTERN (table2);
- if (GET_CODE (p1) == ADDR_VEC && rtx_equal_p (p1, p2))
- {
- identical = true;
- }
- else if (GET_CODE (p1) == ADDR_DIFF_VEC
- && (XVECLEN (p1, 1) == XVECLEN (p2, 1))
- && rtx_equal_p (XEXP (p1, 2), XEXP (p2, 2))
- && rtx_equal_p (XEXP (p1, 3), XEXP (p2, 3)))
- {
- int i;
-
- identical = true;
- for (i = XVECLEN (p1, 1) - 1; i >= 0 && identical; i--)
- if (!rtx_equal_p (XVECEXP (p1, 1, i), XVECEXP (p2, 1, i)))
- identical = false;
- }
-
- if (identical)
- {
- replace_label_data rr;
- bool match;
-
- /* Temporarily replace references to LABEL1 with LABEL2
- in BB1->END so that we could compare the instructions. */
- rr.r1 = label1;
- rr.r2 = label2;
- rr.update_label_nuses = false;
- for_each_rtx (&BB_END (bb1), replace_label, &rr);
-
- match = insns_match_p (mode, BB_END (bb1), BB_END (bb2));
- if (dump_file && match)
- fprintf (dump_file,
- "Tablejumps in bb %i and %i match.\n",
- bb1->index, bb2->index);
-
- /* Set the original label in BB1->END because when deleting
- a block whose end is a tablejump, the tablejump referenced
- from the instruction is deleted too. */
- rr.r1 = label2;
- rr.r2 = label1;
- for_each_rtx (&BB_END (bb1), replace_label, &rr);
+ rtx xe = XVECEXP (x, 0, j);
+ rtx ye = XVECEXP (y, 0, j);
- return match;
- }
- }
- return false;
+ if (GET_CODE (xe) != GET_CODE (ye))
+ return false;
+ if ((GET_CODE (xe) == SET || GET_CODE (xe) == CLOBBER)
+ && ! rtx_equiv_p (&XEXP (xe, 0), XEXP (ye, 0), 0, info))
+ return false;
}
}
+ return true;
+}
- /* First ensure that the instructions match. There may be many outgoing
- edges so this test is generally cheaper. */
- if (!insns_match_p (mode, BB_END (bb1), BB_END (bb2)))
- return false;
+/* Process MEMs in SET_DEST destinations. We must not process this together
+ with REG SET_DESTs, but must do it separately, lest when we see
+ [(set (reg:SI foo) (bar))
+ (set (mem:SI (reg:SI foo) (baz)))]
+ struct_equiv_block_eq could get confused to assume that (reg:SI foo)
+ is not live before this instruction. */
+static bool
+set_dest_addr_equiv_p (rtx x, rtx y, struct equiv_info *info)
+{
+ enum rtx_code code = GET_CODE (x);
+ int length;
+ const char *format;
+ int i;
- /* Search the outgoing edges, ensure that the counts do match, find possible
- fallthru and exception handling edges since these needs more
- validation. */
- if (EDGE_COUNT (bb1->succs) != EDGE_COUNT (bb2->succs))
+ if (code != GET_CODE (y))
return false;
+ if (code == MEM)
+ return rtx_equiv_p (&XEXP (x, 0), XEXP (y, 0), 1, info);
- FOR_EACH_EDGE (e1, ei, bb1->succs)
- {
- e2 = EDGE_SUCC (bb2, ei.index);
-
- if (e1->flags & EDGE_EH)
- nehedges1++;
-
- if (e2->flags & EDGE_EH)
- nehedges2++;
-
- if (e1->flags & EDGE_FALLTHRU)
- fallthru1 = e1;
- if (e2->flags & EDGE_FALLTHRU)
- fallthru2 = e2;
- }
-
- /* If number of edges of various types does not match, fail. */
- if (nehedges1 != nehedges2
- || (fallthru1 != 0) != (fallthru2 != 0))
- return false;
+ /* Process subexpressions. */
+ length = GET_RTX_LENGTH (code);
+ format = GET_RTX_FORMAT (code);
- /* fallthru edges must be forwarded to the same destination. */
- if (fallthru1)
+ for (i = 0; i < length; ++i)
{
- basic_block d1 = (forwarder_block_p (fallthru1->dest)
- ? single_succ (fallthru1->dest): fallthru1->dest);
- basic_block d2 = (forwarder_block_p (fallthru2->dest)
- ? single_succ (fallthru2->dest): fallthru2->dest);
-
- if (d1 != d2)
- return false;
+ switch (format[i])
+ {
+ case 'V':
+ case 'E':
+ if (XVECLEN (x, i) != XVECLEN (y, i))
+ return false;
+ if (XVEC (x, i) != 0)
+ {
+ int j;
+ for (j = 0; j < XVECLEN (x, i); ++j)
+ {
+ if (! set_dest_addr_equiv_p (XVECEXP (x, i, j),
+ XVECEXP (y, i, j), info))
+ return false;
+ }
+ }
+ break;
+ case 'e':
+ if (! set_dest_addr_equiv_p (XEXP (x, i), XEXP (y, i), info))
+ return false;
+ break;
+ default:
+ break;
+ }
}
-
- /* Ensure the same EH region. */
- {
- rtx n1 = find_reg_note (BB_END (bb1), REG_EH_REGION, 0);
- rtx n2 = find_reg_note (BB_END (bb2), REG_EH_REGION, 0);
-
- if (!n1 && n2)
- return false;
-
- if (n1 && (!n2 || XEXP (n1, 0) != XEXP (n2, 0)))
- return false;
- }
-
- /* We don't need to match the rest of edges as above checks should be enough
- to ensure that they are equivalent. */
return true;
}
-/* E1 and E2 are edges with the same destination block. Search their
- predecessors for common code. If found, redirect control flow from
- (maybe the middle of) E1->SRC to (maybe the middle of) E2->SRC. */
-
+/* Check if the set of REG_DEAD notes attached to I1 and I2 allows us to
+ go ahead with merging I1 and I2, which otherwise look fine.
+ Inputs / local registers for the inputs of I1 and I2 have already been
+ set up. */
static bool
-try_crossjump_to_edge (int mode, edge e1, edge e2)
+death_notes_match_p (rtx i1 ATTRIBUTE_UNUSED, rtx i2 ATTRIBUTE_UNUSED,
+ struct equiv_info *info ATTRIBUTE_UNUSED)
{
- int nmatch;
- basic_block src1 = e1->src, src2 = e2->src;
- basic_block redirect_to, redirect_from, to_remove;
- rtx newpos1, newpos2;
- edge s;
- edge_iterator ei;
-
- newpos1 = newpos2 = NULL_RTX;
+#ifdef STACK_REGS
+ /* If cross_jump_death_matters is not 0, the insn's mode
+ indicates whether or not the insn contains any stack-like regs. */
+
+ if (INSN_P (i1)
+ && (info->mode & CLEANUP_POST_REGSTACK) && stack_regs_mentioned (i1))
+ {
+ /* If register stack conversion has already been done, then
+ death notes must also be compared before it is certain that
+ the two instruction streams match. */
+
+ rtx note;
+ HARD_REG_SET i1_regset, i2_regset;
+
+ CLEAR_HARD_REG_SET (i1_regset);
+ CLEAR_HARD_REG_SET (i2_regset);
- /* If we have partitioned hot/cold basic blocks, it is a bad idea
- to try this optimization.
+ for (note = REG_NOTES (i1); note; note = XEXP (note, 1))
+ if (REG_NOTE_KIND (note) == REG_DEAD
+ && STACK_REG_P (XEXP (note, 0)))
+ SET_HARD_REG_BIT (i1_regset, REGNO (XEXP (note, 0)));
- Basic block partitioning may result in some jumps that appear to
- be optimizable (or blocks that appear to be mergeable), but which really
- must be left untouched (they are required to make it safely across
- partition boundaries). See the comments at the top of
- bb-reorder.c:partition_hot_cold_basic_blocks for complete details. */
+ for (note = REG_NOTES (i2); note; note = XEXP (note, 1))
+ if (REG_NOTE_KIND (note) == REG_DEAD
+ && STACK_REG_P (XEXP (note, 0)))
+ {
+ unsigned regno = REGNO (XEXP (note, 0));
+ int i;
- if (flag_reorder_blocks_and_partition && no_new_pseudos)
- return false;
+ for (i = info->cur.local_count - 1; i >= 0; i--)
+ if (regno == REGNO (info->y_local[i]))
+ {
+ regno = REGNO (info->x_local[i]);
+ break;
+ }
+ SET_HARD_REG_BIT (i2_regset, regno);
+ }
- /* Search backward through forwarder blocks. We don't need to worry
- about multiple entry or chained forwarders, as they will be optimized
- away. We do this to look past the unconditional jump following a
- conditional jump that is required due to the current CFG shape. */
- if (single_pred_p (src1)
- && FORWARDER_BLOCK_P (src1))
- e1 = single_pred_edge (src1), src1 = e1->src;
+ GO_IF_HARD_REG_EQUAL (i1_regset, i2_regset, done);
- if (single_pred_p (src2)
- && FORWARDER_BLOCK_P (src2))
- e2 = single_pred_edge (src2), src2 = e2->src;
+ return false;
- /* Nothing to do if we reach ENTRY, or a common source block. */
- if (src1 == ENTRY_BLOCK_PTR || src2 == ENTRY_BLOCK_PTR)
- return false;
- if (src1 == src2)
- return false;
+ done:
+ ;
+ }
+#endif
+ return true;
+}
- /* Seeing more than 1 forwarder blocks would confuse us later... */
- if (FORWARDER_BLOCK_P (e1->dest)
- && FORWARDER_BLOCK_P (single_succ (e1->dest)))
- return false;
+/* Return true if I1 and I2 are equivalent and thus can be crossjumped. */
- if (FORWARDER_BLOCK_P (e2->dest)
- && FORWARDER_BLOCK_P (single_succ (e2->dest)))
- return false;
+bool
+insns_match_p (rtx i1, rtx i2, struct equiv_info *info)
+{
+ int rvalue_change_start;
+ struct struct_equiv_checkpoint before_rvalue_change;
- /* Likewise with dead code (possibly newly created by the other optimizations
- of cfg_cleanup). */
- if (EDGE_COUNT (src1->preds) == 0 || EDGE_COUNT (src2->preds) == 0)
+ /* Verify that I1 and I2 are equivalent. */
+ if (GET_CODE (i1) != GET_CODE (i2))
return false;
- /* Look for the common insn sequence, part the first ... */
- if (!outgoing_edges_match (mode, src1, src2))
- return false;
+ info->cur.x_start = i1;
+ info->cur.y_start = i2;
- /* ... and part the second. */
- nmatch = flow_find_cross_jump (mode, src1, src2, &newpos1, &newpos2);
+ /* If this is a CALL_INSN, compare register usage information.
+ If we don't check this on stack register machines, the two
+ CALL_INSNs might be merged leaving reg-stack.c with mismatching
+ numbers of stack registers in the same basic block.
+ If we don't check this on machines with delay slots, a delay slot may
+ be filled that clobbers a parameter expected by the subroutine.
- /* Don't proceed with the crossjump unless we found a sufficient number
- of matching instructions or the 'from' block was totally matched
- (such that its predecessors will hopefully be redirected and the
- block removed). */
- if ((nmatch < PARAM_VALUE (PARAM_MIN_CROSSJUMP_INSNS))
- && (newpos1 != BB_HEAD (src1)))
- return false;
+ ??? We take the simple route for now and assume that if they're
+ equal, they were constructed identically. */
- /* Here we know that the insns in the end of SRC1 which are common with SRC2
- will be deleted.
- If we have tablejumps in the end of SRC1 and SRC2
- they have been already compared for equivalence in outgoing_edges_match ()
- so replace the references to TABLE1 by references to TABLE2. */
+ if (CALL_P (i1))
{
- rtx label1, label2;
- rtx table1, table2;
-
- if (tablejump_p (BB_END (src1), &label1, &table1)
- && tablejump_p (BB_END (src2), &label2, &table2)
- && label1 != label2)
+ if (SIBLING_CALL_P (i1) != SIBLING_CALL_P (i2)
+ || ! set_dest_equiv_p (PATTERN (i1), PATTERN (i2), info)
+ || ! set_dest_equiv_p (CALL_INSN_FUNCTION_USAGE (i1),
+ CALL_INSN_FUNCTION_USAGE (i2), info)
+ || ! rtx_equiv_p (&CALL_INSN_FUNCTION_USAGE (i1),
+ CALL_INSN_FUNCTION_USAGE (i2), -1, info))
{
- replace_label_data rr;
- rtx insn;
-
- /* Replace references to LABEL1 with LABEL2. */
- rr.r1 = label1;
- rr.r2 = label2;
- rr.update_label_nuses = true;
- for (insn = get_insns (); insn; insn = NEXT_INSN (insn))
- {
- /* Do not replace the label in SRC1->END because when deleting
- a block whose end is a tablejump, the tablejump referenced
- from the instruction is deleted too. */
- if (insn != BB_END (src1))
- for_each_rtx (&insn, replace_label, &rr);
- }
+ cancel_changes (0);
+ return false;
}
}
-
- /* Avoid splitting if possible. We must always split when SRC2 has
- EH predecessor edges, or we may end up with basic blocks with both
- normal and EH predecessor edges. */
- if (newpos2 == BB_HEAD (src2)
- && !(EDGE_PRED (src2, 0)->flags & EDGE_EH))
- redirect_to = src2;
- else
+ else if (INSN_P (i1))
{
- if (newpos2 == BB_HEAD (src2))
+ if (! set_dest_equiv_p (PATTERN (i1), PATTERN (i2), info))
{
- /* Skip possible basic block header. */
- if (LABEL_P (newpos2))
- newpos2 = NEXT_INSN (newpos2);
- if (NOTE_P (newpos2))
- newpos2 = NEXT_INSN (newpos2);
+ cancel_changes (0);
+ return false;
}
-
- if (dump_file)
- fprintf (dump_file, "Splitting bb %i before %i insns\n",
- src2->index, nmatch);
- redirect_to = split_block (src2, PREV_INSN (newpos2))->dest;
}
+ rvalue_change_start = num_validated_changes ();
+ struct_equiv_make_checkpoint (&before_rvalue_change, info);
+ /* Check death_notes_match_p *after* the inputs have been processed,
+ so that local inputs will already have been set up. */
+ if (! INSN_P (i1)
+ || (!bitmap_bit_p (info->equiv_used, info->cur.ninsns)
+ && rtx_equiv_p (&PATTERN (i1), PATTERN (i2), -1, info)
+ && death_notes_match_p (i1, i2, info)
+ && verify_changes (0)))
+ return true;
- if (dump_file)
- fprintf (dump_file,
- "Cross jumping from bb %i to bb %i; %i common insns\n",
- src1->index, src2->index, nmatch);
-
- redirect_to->count += src1->count;
- redirect_to->frequency += src1->frequency;
- /* We may have some registers visible trought the block. */
- redirect_to->flags |= BB_DIRTY;
-
- /* Recompute the frequencies and counts of outgoing edges. */
- FOR_EACH_EDGE (s, ei, redirect_to->succs)
+ /* Do not do EQUIV substitution after reload. First, we're undoing the
+ work of reload_cse. Second, we may be undoing the work of the post-
+ reload splitting pass. */
+ /* ??? Possibly add a new phase switch variable that can be used by
+ targets to disallow the troublesome insns after splitting. */
+ if (!reload_completed)
{
- edge s2;
- edge_iterator ei;
- basic_block d = s->dest;
-
- if (FORWARDER_BLOCK_P (d))
- d = single_succ (d);
-
- FOR_EACH_EDGE (s2, ei, src1->succs)
- {
- basic_block d2 = s2->dest;
- if (FORWARDER_BLOCK_P (d2))
- d2 = single_succ (d2);
- if (d == d2)
- break;
- }
+ rtx equiv1, equiv2;
- s->count += s2->count;
+ cancel_changes (rvalue_change_start);
+ struct_equiv_restore_checkpoint (&before_rvalue_change, info);
- /* Take care to update possible forwarder blocks. We verified
- that there is no more than one in the chain, so we can't run
- into infinite loop. */
- if (FORWARDER_BLOCK_P (s->dest))
+ /* The following code helps take care of G++ cleanups. */
+ equiv1 = find_reg_equal_equiv_note (i1);
+ equiv2 = find_reg_equal_equiv_note (i2);
+ if (equiv1 && equiv2
+ /* If the equivalences are not to a constant, they may
+ reference pseudos that no longer exist, so we can't
+ use them. */
+ && (! reload_completed
+ || (CONSTANT_P (XEXP (equiv1, 0))
+ && rtx_equal_p (XEXP (equiv1, 0), XEXP (equiv2, 0)))))
{
- single_succ_edge (s->dest)->count += s2->count;
- s->dest->count += s2->count;
- s->dest->frequency += EDGE_FREQUENCY (s);
- }
+ rtx s1 = single_set (i1);
+ rtx s2 = single_set (i2);
- if (FORWARDER_BLOCK_P (s2->dest))
- {
- single_succ_edge (s2->dest)->count -= s2->count;
- if (single_succ_edge (s2->dest)->count < 0)
- single_succ_edge (s2->dest)->count = 0;
- s2->dest->count -= s2->count;
- s2->dest->frequency -= EDGE_FREQUENCY (s);
- if (s2->dest->frequency < 0)
- s2->dest->frequency = 0;
- if (s2->dest->count < 0)
- s2->dest->count = 0;
+ if (s1 != 0 && s2 != 0)
+ {
+ validate_change (i1, &SET_SRC (s1), XEXP (equiv1, 0), 1);
+ validate_change (i2, &SET_SRC (s2), XEXP (equiv2, 0), 1);
+ /* Only inspecting the new SET_SRC is not good enough,
+ because there may also be bare USEs in a single_set
+ PARALLEL. */
+ if (rtx_equiv_p (&PATTERN (i1), PATTERN (i2), -1, info)
+ && death_notes_match_p (i1, i2, info)
+ && verify_changes (0))
+ {
+ /* Mark this insn so that we'll use the equivalence in
+ all subsequent passes. */
+ bitmap_set_bit (info->equiv_used, info->cur.ninsns);
+ return true;
+ }
+ }
}
-
- if (!redirect_to->frequency && !src1->frequency)
- s->probability = (s->probability + s2->probability) / 2;
- else
- s->probability
- = ((s->probability * redirect_to->frequency +
- s2->probability * src1->frequency)
- / (redirect_to->frequency + src1->frequency));
}
- update_br_prob_note (redirect_to);
-
- /* Edit SRC1 to go to REDIRECT_TO at NEWPOS1. */
-
- /* Skip possible basic block header. */
- if (LABEL_P (newpos1))
- newpos1 = NEXT_INSN (newpos1);
-
- if (NOTE_P (newpos1))
- newpos1 = NEXT_INSN (newpos1);
-
- redirect_from = split_block (src1, PREV_INSN (newpos1))->src;
- to_remove = single_succ (redirect_from);
-
- redirect_edge_and_branch_force (single_succ_edge (redirect_from), redirect_to);
- delete_basic_block (to_remove);
-
- update_forwarder_flag (redirect_from);
- if (redirect_to != src2)
- update_forwarder_flag (src2);
+ cancel_changes (0);
+ return false;
+}
+/* Set up mode and register information in INFO. Return true for success. */
+bool
+struct_equiv_init (int mode, struct equiv_info *info)
+{
+ info->mode = mode;
+ if (mode & STRUCT_EQUIV_START)
+ {
+ info->x_input = info->y_input = info->input_reg = NULL_RTX;
+ info->equiv_used = ALLOC_REG_SET (®_obstack);
+ info->check_input_conflict = false;
+ }
+ info->had_input_conflict = false;
+ info->cur.ninsns = info->cur.version = 0;
+ info->cur.local_count = info->cur.input_count = 0;
+ info->cur.x_start = info->cur.y_start = NULL_RTX;
+ info->need_rerun = false;
+ info->live_update = true;
+ info->cur.input_valid = false;
+ info->common_live = ALLOC_REG_SET (®_obstack);
+ info->x_local_live = ALLOC_REG_SET (®_obstack);
+ info->y_local_live = ALLOC_REG_SET (®_obstack);
+ COPY_REG_SET (info->common_live, info->x_block->il.rtl->global_live_at_end);
+ struct_equiv_make_checkpoint (&info->best_match, info);
return true;
}
-/* Search the predecessors of BB for common insn sequences. When found,
- share code between them by redirecting control flow. Return true if
- any changes made. */
-
-static bool
-try_crossjump_bb (int mode, basic_block bb)
+/* Insns XI and YI have been matched. Merge memory attributes and reg
+ notes. */
+static void
+struct_equiv_merge (rtx xi, rtx yi, struct equiv_info *info)
{
- edge e, e2, fallthru;
- bool changed;
- unsigned max, ix, ix2;
- basic_block ev, ev2;
- edge_iterator ei;
-
- /* Nothing to do if there is not at least two incoming edges. */
- if (EDGE_COUNT (bb->preds) < 2)
- return false;
+ rtx equiv1, equiv2;
- /* Don't crossjump if this block ends in a computed jump,
- unless we are optimizing for size. */
- if (!optimize_size
- && bb != EXIT_BLOCK_PTR
- && computed_jump_p (BB_END (bb)))
- return false;
+ merge_memattrs (xi, yi);
- /* If we are partitioning hot/cold basic blocks, we don't want to
- mess up unconditional or indirect jumps that cross between hot
- and cold sections.
-
- Basic block partitioning may result in some jumps that appear to
- be optimizable (or blocks that appear to be mergeable), but which really
- must be left untouched (they are required to make it safely across
- partition boundaries). See the comments at the top of
- bb-reorder.c:partition_hot_cold_basic_blocks for complete details. */
+ /* If the merged insns have different REG_EQUAL notes, then
+ remove them. */
+ info->live_update = false;
+ equiv1 = find_reg_equal_equiv_note (xi);
+ equiv2 = find_reg_equal_equiv_note (yi);
+ if (equiv1 && !equiv2)
+ remove_note (xi, equiv1);
+ else if (!equiv1 && equiv2)
+ remove_note (yi, equiv2);
+ else if (equiv1 && equiv2
+ && !rtx_equiv_p (&XEXP (equiv1, 0), XEXP (equiv2, 0),
+ 1, info))
+ {
+ remove_note (xi, equiv1);
+ remove_note (yi, equiv2);
+ }
+ info->live_update = true;
+}
- if (BB_PARTITION (EDGE_PRED (bb, 0)->src) !=
- BB_PARTITION (EDGE_PRED (bb, 1)->src)
- || (EDGE_PRED (bb, 0)->flags & EDGE_CROSSING))
- return false;
+/* Return number of matched insns.
+ This function must be called up to three times for a successful cross-jump
+ match:
+ first to find out which instructions do match. While trying to match
+ another instruction that doesn't match, we destroy information in info
+ about the actual inputs. So if there have been any before the last
+ match attempt, we need to call this function again to recompute the
+ actual inputs up to the actual start of the matching sequence.
+ When we are then satisfied that the cross-jump is worthwhile, we
+ call this function a third time to make any changes needed to make the
+ sequences match: apply equivalences, remove non-matching
+ notes and merge memory attributes. */
+int
+struct_equiv_block_eq (int mode, struct equiv_info *info)
+{
+ rtx x_stop, y_stop;
+ rtx xi, yi;
+ int i;
- /* It is always cheapest to redirect a block that ends in a branch to
- a block that falls through into BB, as that adds no branches to the
- program. We'll try that combination first. */
- fallthru = NULL;
- max = PARAM_VALUE (PARAM_MAX_CROSSJUMP_EDGES);
+ if (! REG_SET_EQUAL_P (info->x_block->il.rtl->global_live_at_end,
+ info->y_block->il.rtl->global_live_at_end))
+ {
+ gcc_unreachable ();
+ /* Cross-jumping should already have verified that the destinations
+ are the same. */
+ gcc_assert (mode & STRUCT_EQUIV_NEED_FULL_BLOCK);
+ return 0;
+ }
+ if (mode & STRUCT_EQUIV_START)
+ {
+ x_stop = BB_HEAD (info->x_block);
+ y_stop = BB_HEAD (info->y_block);
+ if (!x_stop || !y_stop)
+ return 0;
+ }
+ else
+ {
+ x_stop = info->cur.x_start;
+ y_stop = info->cur.y_start;
+ }
+ if (!struct_equiv_init (mode, info))
+ return 0;
- if (EDGE_COUNT (bb->preds) > max)
- return false;
+ /* Skip simple jumps at the end of the blocks. Complex jumps still
+ need to be compared for equivalence, which we'll do below. */
- FOR_EACH_EDGE (e, ei, bb->preds)
+ xi = BB_END (info->x_block);
+ if (onlyjump_p (xi)
+ || (returnjump_p (xi) && !side_effects_p (PATTERN (xi))))
{
- if (e->flags & EDGE_FALLTHRU)
- fallthru = e;
+ info->cur.x_start = xi;
+ xi = PREV_INSN (xi);
}
- changed = false;
- for (ix = 0, ev = bb; ix < EDGE_COUNT (ev->preds); )
+ yi = BB_END (info->y_block);
+ if (onlyjump_p (yi)
+ || (returnjump_p (yi) && !side_effects_p (PATTERN (yi))))
{
- e = EDGE_PRED (ev, ix);
- ix++;
+ info->cur.y_start = yi;
+ /* Count everything except for unconditional jump as insn. */
+ /* ??? Is it right to count unconditional jumps with a clobber?
+ Should we count conditional returns? */
+ if (!simplejump_p (yi) && !returnjump_p (yi) && info->cur.x_start)
+ info->cur.ninsns++;
+ yi = PREV_INSN (yi);
+ }
- /* As noted above, first try with the fallthru predecessor. */
- if (fallthru)
+ if (mode & STRUCT_EQUIV_MATCH_JUMPS)
+ {
+ /* The caller is expected to have comapred the jumps already, but we
+ need to match them again to get any local registers and inputs. */
+ gcc_assert (!info->cur.x_start == !info->cur.y_start);
+ if (info->cur.x_start)
{
- /* Don't combine the fallthru edge into anything else.
- If there is a match, we'll do it the other way around. */
- if (e == fallthru)
- continue;
- /* If nothing changed since the last attempt, there is nothing
- we can do. */
- if (!first_pass
- && (!(e->src->flags & BB_DIRTY)
- && !(fallthru->src->flags & BB_DIRTY)))
- continue;
-
- if (try_crossjump_to_edge (mode, e, fallthru))
- {
- changed = true;
- ix = 0;
- ev = bb;
- continue;
- }
+ if (any_condjump_p (info->cur.x_start)
+ ? !condjump_equiv_p (info)
+ : !insns_match_p (info->cur.x_start, info->cur.y_start, info))
+ gcc_unreachable ();
}
-
- /* Non-obvious work limiting check: Recognize that we're going
- to call try_crossjump_bb on every basic block. So if we have
- two blocks with lots of outgoing edges (a switch) and they
- share lots of common destinations, then we would do the
- cross-jump check once for each common destination.
-
- Now, if the blocks actually are cross-jump candidates, then
- all of their destinations will be shared. Which means that
- we only need check them for cross-jump candidacy once. We
- can eliminate redundant checks of crossjump(A,B) by arbitrarily
- choosing to do the check from the block for which the edge
- in question is the first successor of A. */
- if (EDGE_SUCC (e->src, 0) != e)
- continue;
-
- for (ix2 = 0, ev2 = bb; ix2 < EDGE_COUNT (ev2->preds); )
+ else if (any_condjump_p (xi) && any_condjump_p (yi))
{
- e2 = EDGE_PRED (ev2, ix2);
- ix2++;
-
- if (e2 == e)
- continue;
-
- /* We've already checked the fallthru edge above. */
- if (e2 == fallthru)
- continue;
-
- /* The "first successor" check above only prevents multiple
- checks of crossjump(A,B). In order to prevent redundant
- checks of crossjump(B,A), require that A be the block
- with the lowest index. */
- if (e->src->index > e2->src->index)
- continue;
-
- /* If nothing changed since the last attempt, there is nothing
- we can do. */
- if (!first_pass
- && (!(e->src->flags & BB_DIRTY)
- && !(e2->src->flags & BB_DIRTY)))
- continue;
-
- if (try_crossjump_to_edge (mode, e, e2))
- {
- changed = true;
- ev2 = bb;
- ix = 0;
- break;
- }
+ info->cur.x_start = xi;
+ info->cur.y_start = yi;
+ xi = PREV_INSN (xi);
+ yi = PREV_INSN (yi);
+ info->cur.ninsns++;
+ if (!condjump_equiv_p (info))
+ gcc_unreachable ();
}
+ if (info->cur.x_start && info->mode & STRUCT_EQUIV_FINAL)
+ struct_equiv_merge (info->cur.x_start, info->cur.y_start, info);
}
- return changed;
-}
+ struct_equiv_improve_checkpoint (&info->best_match, info);
+ info->x_end = xi;
+ info->y_end = yi;
+ if (info->cur.x_start != x_stop)
+ for (;;)
+ {
+ /* Ignore notes. */
+ while (!INSN_P (xi) && xi != x_stop)
+ xi = PREV_INSN (xi);
-/* Do simple CFG optimizations - basic block merging, simplifying of jump
- instructions etc. Return nonzero if changes were made. */
+ while (!INSN_P (yi) && yi != y_stop)
+ yi = PREV_INSN (yi);
-static bool
-try_optimize_cfg (int mode)
-{
- bool changed_overall = false;
- bool changed;
- int iterations = 0;
- basic_block bb, b, next;
+ if (!insns_match_p (xi, yi, info))
+ break;
+ if (INSN_P (xi))
+ {
+ if (info->mode & STRUCT_EQUIV_FINAL)
+ struct_equiv_merge (xi, yi, info);
+ info->cur.ninsns++;
+ struct_equiv_improve_checkpoint (&info->best_match, info);
+ }
+ if (xi == x_stop || yi == y_stop)
+ {
+ /* If we reached the start of at least one of the blocks, but
+ best_match hasn't been advanced back to the first valid insn
+ yet, represent the increased benefit of completing the block
+ as an increased instruction count. */
+ if (info->best_match.x_start != info->cur.x_start
+ && (xi == BB_HEAD (info->x_block)
+ || yi == BB_HEAD (info->y_block)))
+ {
+ info->cur.ninsns++;
+ struct_equiv_improve_checkpoint (&info->best_match, info);
+ info->cur.ninsns--;
+ if (info->best_match.ninsns > info->cur.ninsns)
+ info->best_match.ninsns = info->cur.ninsns;
+ }
+ break;
+ }
+ xi = PREV_INSN (xi);
+ yi = PREV_INSN (yi);
+ }
- if (mode & CLEANUP_CROSSJUMP)
- add_noreturn_fake_exit_edges ();
+ /* If we failed to match an insn, but had some changes registered from
+ trying to make the insns match, we need to cancel these changes now. */
+ cancel_changes (0);
+ /* Restore to best_match to get the sequence with the best known-so-far
+ cost-benefit difference. */
+ struct_equiv_restore_checkpoint (&info->best_match, info);
- if (mode & (CLEANUP_UPDATE_LIFE | CLEANUP_CROSSJUMP | CLEANUP_THREADING))
- clear_bb_flags ();
+ /* Include preceding notes and labels in the cross-jump / if-conversion.
+ One, this may bring us to the head of the blocks.
+ Two, it keeps line number notes as matched as may be. */
+ if (info->cur.ninsns)
+ {
+ xi = info->cur.x_start;
+ yi = info->cur.y_start;
+ while (xi != x_stop && !INSN_P (PREV_INSN (xi)))
+ xi = PREV_INSN (xi);
- FOR_EACH_BB (bb)
- update_forwarder_flag (bb);
+ while (yi != y_stop && !INSN_P (PREV_INSN (yi)))
+ yi = PREV_INSN (yi);
- if (! targetm.cannot_modify_jumps_p ())
+ info->cur.x_start = xi;
+ info->cur.y_start = yi;
+ }
+
+ if (!info->cur.input_valid)
+ info->x_input = info->y_input = info->input_reg = NULL_RTX;
+ if (!info->need_rerun)
{
- first_pass = true;
- /* Attempt to merge blocks as made possible by edge removal. If
- a block has only one successor, and the successor has only
- one predecessor, they may be combined. */
- do
+ find_dying_inputs (info);
+ if (info->mode & STRUCT_EQUIV_FINAL)
{
- changed = false;
- iterations++;
-
- if (dump_file)
- fprintf (dump_file,
- "\n\ntry_optimize_cfg iteration %i\n\n",
- iterations);
+ if (info->check_input_conflict && ! resolve_input_conflict (info))
+ gcc_unreachable ();
+ }
+ else
+ {
+ bool input_conflict = info->had_input_conflict;
- for (b = ENTRY_BLOCK_PTR->next_bb; b != EXIT_BLOCK_PTR;)
+ if (!input_conflict
+ && info->dying_inputs > 1
+ && bitmap_intersect_p (info->x_local_live, info->y_local_live))
{
- basic_block c;
- edge s;
- bool changed_here = false;
-
- /* Delete trivially dead basic blocks. */
- while (EDGE_COUNT (b->preds) == 0)
- {
- c = b->prev_bb;
- if (dump_file)
- fprintf (dump_file, "Deleting block %i.\n",
- b->index);
-
- delete_basic_block (b);
- if (!(mode & CLEANUP_CFGLAYOUT))
- changed = true;
- b = c;
- }
-
- /* Remove code labels no longer used. */
- if (single_pred_p (b)
- && (single_pred_edge (b)->flags & EDGE_FALLTHRU)
- && !(single_pred_edge (b)->flags & EDGE_COMPLEX)
- && LABEL_P (BB_HEAD (b))
- /* If the previous block ends with a branch to this
- block, we can't delete the label. Normally this
- is a condjump that is yet to be simplified, but
- if CASE_DROPS_THRU, this can be a tablejump with
- some element going to the same place as the
- default (fallthru). */
- && (single_pred (b) == ENTRY_BLOCK_PTR
- || !JUMP_P (BB_END (single_pred (b)))
- || ! label_is_jump_target_p (BB_HEAD (b),
- BB_END (single_pred (b)))))
- {
- rtx label = BB_HEAD (b);
-
- delete_insn_chain (label, label);
- /* In the case label is undeletable, move it after the
- BASIC_BLOCK note. */
- if (NOTE_LINE_NUMBER (BB_HEAD (b)) == NOTE_INSN_DELETED_LABEL)
- {
- rtx bb_note = NEXT_INSN (BB_HEAD (b));
-
- reorder_insns_nobb (label, label, bb_note);
- BB_HEAD (b) = bb_note;
- }
- if (dump_file)
- fprintf (dump_file, "Deleted label in block %i.\n",
- b->index);
- }
-
- /* If we fall through an empty block, we can remove it. */
- if (!(mode & CLEANUP_CFGLAYOUT)
- && single_pred_p (b)
- && (single_pred_edge (b)->flags & EDGE_FALLTHRU)
- && !LABEL_P (BB_HEAD (b))
- && FORWARDER_BLOCK_P (b)
- /* Note that forwarder_block_p true ensures that
- there is a successor for this block. */
- && (single_succ_edge (b)->flags & EDGE_FALLTHRU)
- && n_basic_blocks > 1)
- {
- if (dump_file)
- fprintf (dump_file,
- "Deleting fallthru block %i.\n",
- b->index);
-
- c = b->prev_bb == ENTRY_BLOCK_PTR ? b->next_bb : b->prev_bb;
- redirect_edge_succ_nodup (single_pred_edge (b),
- single_succ (b));
- delete_basic_block (b);
- changed = true;
- b = c;
- }
+ regset_head clobbered_regs;
- if (single_succ_p (b)
- && (s = single_succ_edge (b))
- && !(s->flags & EDGE_COMPLEX)
- && (c = s->dest) != EXIT_BLOCK_PTR
- && single_pred_p (c)
- && b != c)
+ INIT_REG_SET (&clobbered_regs);
+ for (i = 0; i < info->cur.local_count; i++)
{
- /* When not in cfg_layout mode use code aware of reordering
- INSN. This code possibly creates new basic blocks so it
- does not fit merge_blocks interface and is kept here in
- hope that it will become useless once more of compiler
- is transformed to use cfg_layout mode. */
-
- if ((mode & CLEANUP_CFGLAYOUT)
- && can_merge_blocks_p (b, c))
+ if (assign_reg_reg_set (&clobbered_regs, info->y_local[i], 0))
{
- merge_blocks (b, c);
- update_forwarder_flag (b);
- changed_here = true;
+ input_conflict = true;
+ break;
}
- else if (!(mode & CLEANUP_CFGLAYOUT)
- /* If the jump insn has side effects,
- we can't kill the edge. */
- && (!JUMP_P (BB_END (b))
- || (reload_completed
- ? simplejump_p (BB_END (b))
- : (onlyjump_p (BB_END (b))
- && !tablejump_p (BB_END (b),
- NULL, NULL))))
- && (next = merge_blocks_move (s, b, c, mode)))
- {
- b = next;
- changed_here = true;
- }
- }
-
- /* Simplify branch over branch. */
- if ((mode & CLEANUP_EXPENSIVE)
- && !(mode & CLEANUP_CFGLAYOUT)
- && try_simplify_condjump (b))
- changed_here = true;
-
- /* If B has a single outgoing edge, but uses a
- non-trivial jump instruction without side-effects, we
- can either delete the jump entirely, or replace it
- with a simple unconditional jump. */
- if (single_succ_p (b)
- && single_succ (b) != EXIT_BLOCK_PTR
- && onlyjump_p (BB_END (b))
- && !find_reg_note (BB_END (b), REG_CROSSING_JUMP, NULL_RTX)
- && try_redirect_by_replacing_jump (single_succ_edge (b),
- single_succ (b),
- (mode & CLEANUP_CFGLAYOUT) != 0))
- {
- update_forwarder_flag (b);
- changed_here = true;
+ assign_reg_reg_set (&clobbered_regs, info->x_local[i], 1);
}
-
- /* Simplify branch to branch. */
- if (try_forward_edges (mode, b))
- changed_here = true;
-
- /* Look for shared code between blocks. */
- if ((mode & CLEANUP_CROSSJUMP)
- && try_crossjump_bb (mode, b))
- changed_here = true;
-
- /* Don't get confused by the index shift caused by
- deleting blocks. */
- if (!changed_here)
- b = b->next_bb;
- else
- changed = true;
+ CLEAR_REG_SET (&clobbered_regs);
}
-
- if ((mode & CLEANUP_CROSSJUMP)
- && try_crossjump_bb (mode, EXIT_BLOCK_PTR))
- changed = true;
-
-#ifdef ENABLE_CHECKING
- if (changed)
- verify_flow_info ();
-#endif
-
- changed_overall |= changed;
- first_pass = false;
- }
- while (changed);
- }
-
- if (mode & CLEANUP_CROSSJUMP)
- remove_fake_exit_edges ();
-
- FOR_ALL_BB (b)
- b->flags &= ~(BB_FORWARDER_BLOCK | BB_NONTHREADABLE_BLOCK);
-
- return changed_overall;
-}
-
-/* Delete all unreachable basic blocks. */
-
-bool
-delete_unreachable_blocks (void)
-{
- bool changed = false;
- basic_block b, next_bb;
-
- find_unreachable_blocks ();
-
- /* Delete all unreachable basic blocks. */
-
- for (b = ENTRY_BLOCK_PTR->next_bb; b != EXIT_BLOCK_PTR; b = next_bb)
- {
- next_bb = b->next_bb;
-
- if (!(b->flags & BB_REACHABLE))
- {
- delete_basic_block (b);
- changed = true;
+ if (input_conflict && !info->check_input_conflict)
+ info->need_rerun = true;
+ info->check_input_conflict = input_conflict;
}
}
- if (changed)
- tidy_fallthru_edges ();
- return changed;
+ if (info->mode & STRUCT_EQUIV_NEED_FULL_BLOCK
+ && (info->cur.x_start != x_stop || info->cur.y_start != y_stop))
+ return 0;
+ return info->cur.ninsns;
}
-/* Merges sequential blocks if possible. */
-
-bool
-merge_seq_blocks (void)
+/* For each local register, set info->local_rvalue to true iff the register
+ is a dying input. Store the total number of these in info->dying_inputs. */
+static void
+find_dying_inputs (struct equiv_info *info)
{
- basic_block bb;
- bool changed = false;
+ int i;
- for (bb = ENTRY_BLOCK_PTR->next_bb; bb != EXIT_BLOCK_PTR; )
+ info->dying_inputs = 0;
+ for (i = info->cur.local_count-1; i >=0; i--)
{
- if (single_succ_p (bb)
- && can_merge_blocks_p (bb, single_succ (bb)))
- {
- /* Merge the blocks and retry. */
- merge_blocks (bb, single_succ (bb));
- changed = true;
- continue;
- }
+ rtx x = info->x_local[i];
+ unsigned regno = REGNO (x);
+ int nregs = (regno >= FIRST_PSEUDO_REGISTER
+ ? 1 : hard_regno_nregs[regno][GET_MODE (x)]);
- bb = bb->next_bb;
+ for (info->local_rvalue[i] = false; nregs >= 0; regno++, --nregs)
+ if (REGNO_REG_SET_P (info->x_local_live, regno))
+ {
+ info->dying_inputs++;
+ info->local_rvalue[i] = true;
+ break;
+ }
}
-
- return changed;
}
-
-/* Tidy the CFG by deleting unreachable code and whatnot. */
-bool
-cleanup_cfg (int mode)
+/* For each local register that is a dying input, y_local[i] will be
+ copied to x_local[i]. We'll do this in ascending order. Try to
+ re-order the locals to avoid conflicts like r3 = r2; r4 = r3; .
+ Return true iff the re-ordering is successful, or not necessary. */
+static bool
+resolve_input_conflict (struct equiv_info *info)
{
- bool changed = false;
+ int i, j, end;
+ int nswaps = 0;
+ rtx save_x_local[STRUCT_EQUIV_MAX_LOCAL];
+ rtx save_y_local[STRUCT_EQUIV_MAX_LOCAL];
- timevar_push (TV_CLEANUP_CFG);
- if (delete_unreachable_blocks ())
+ find_dying_inputs (info);
+ if (info->dying_inputs <= 1)
+ return true;
+ memcpy (save_x_local, info->x_local, sizeof save_x_local);
+ memcpy (save_y_local, info->y_local, sizeof save_y_local);
+ end = info->cur.local_count - 1;
+ for (i = 0; i <= end; i++)
{
- changed = true;
- /* We've possibly created trivially dead code. Cleanup it right
- now to introduce more opportunities for try_optimize_cfg. */
- if (!(mode & (CLEANUP_NO_INSN_DEL | CLEANUP_UPDATE_LIFE))
- && !reload_completed)
- delete_trivially_dead_insns (get_insns(), max_reg_num ());
- }
-
- compact_blocks ();
+ /* Cycle detection with regsets is expensive, so we just check that
+ we don't exceed the maximum number of swaps needed in the acyclic
+ case. */
+ int max_swaps = end - i;
- while (try_optimize_cfg (mode))
- {
- delete_unreachable_blocks (), changed = true;
- if (mode & CLEANUP_UPDATE_LIFE)
- {
- /* Cleaning up CFG introduces more opportunities for dead code
- removal that in turn may introduce more opportunities for
- cleaning up the CFG. */
- if (!update_life_info_in_dirty_blocks (UPDATE_LIFE_GLOBAL_RM_NOTES,
- PROP_DEATH_NOTES
- | PROP_SCAN_DEAD_CODE
- | PROP_KILL_DEAD_CODE
- | ((mode & CLEANUP_LOG_LINKS)
- ? PROP_LOG_LINKS : 0)))
- break;
- }
- else if (!(mode & CLEANUP_NO_INSN_DEL)
- && (mode & CLEANUP_EXPENSIVE)
- && !reload_completed)
+ /* Check if x_local[i] will be clobbered. */
+ if (!info->local_rvalue[i])
+ continue;
+ /* Check if any later value needs to be copied earlier. */
+ for (j = i + 1; j <= end; j++)
{
- if (!delete_trivially_dead_insns (get_insns(), max_reg_num ()))
- break;
+ rtx tmp;
+
+ if (!info->local_rvalue[j])
+ continue;
+ if (!reg_overlap_mentioned_p (info->x_local[i], info->y_local[j]))
+ continue;
+ if (--max_swaps < 0)
+ {
+ memcpy (info->x_local, save_x_local, sizeof save_x_local);
+ memcpy (info->y_local, save_y_local, sizeof save_y_local);
+ return false;
+ }
+ nswaps++;
+ tmp = info->x_local[i];
+ info->x_local[i] = info->x_local[j];
+ info->x_local[j] = tmp;
+ tmp = info->y_local[i];
+ info->y_local[i] = info->y_local[j];
+ info->y_local[j] = tmp;
+ j = i;
}
- else
- break;
- delete_dead_jumptables ();
}
-
- timevar_pop (TV_CLEANUP_CFG);
-
- return changed;
-}
-
-static void
-rest_of_handle_jump (void)
-{
- delete_unreachable_blocks ();
-
- if (cfun->tail_call_emit)
- fixup_tail_calls ();
-}
-
-struct tree_opt_pass pass_jump =
-{
- "sibling", /* name */
- NULL, /* gate */
- rest_of_handle_jump, /* execute */
- NULL, /* sub */
- NULL, /* next */
- 0, /* static_pass_number */
- TV_JUMP, /* tv_id */
- 0, /* properties_required */
- 0, /* properties_provided */
- 0, /* properties_destroyed */
- TODO_ggc_collect, /* todo_flags_start */
- TODO_dump_func |
- TODO_verify_flow, /* todo_flags_finish */
- 'i' /* letter */
-};
-
-
-static void
-rest_of_handle_jump2 (void)
-{
- /* Turn NOTE_INSN_EXPECTED_VALUE into REG_BR_PROB. Do this
- before jump optimization switches branch directions. */
- if (flag_guess_branch_prob)
- expected_value_to_br_prob ();
-
- delete_trivially_dead_insns (get_insns (), max_reg_num ());
- reg_scan (get_insns (), max_reg_num ());
- if (dump_file)
- dump_flow_info (dump_file);
- cleanup_cfg ((optimize ? CLEANUP_EXPENSIVE : 0) | CLEANUP_PRE_LOOP
- | (flag_thread_jumps ? CLEANUP_THREADING : 0));
-
- create_loop_notes ();
-
- purge_line_number_notes ();
-
- if (optimize)
- cleanup_cfg (CLEANUP_EXPENSIVE | CLEANUP_PRE_LOOP);
-
- /* Jump optimization, and the removal of NULL pointer checks, may
- have reduced the number of instructions substantially. CSE, and
- future passes, allocate arrays whose dimensions involve the
- maximum instruction UID, so if we can reduce the maximum UID
- we'll save big on memory. */
- renumber_insns (dump_file);
+ info->had_input_conflict = true;
+ if (dump_file && nswaps)
+ fprintf (dump_file, "Resolved input conflict, %d %s.\n",
+ nswaps, nswaps == 1 ? "swap" : "swaps");
+ return true;
}
-
-
-struct tree_opt_pass pass_jump2 =
-{
- "jump", /* name */
- NULL, /* gate */
- rest_of_handle_jump2, /* execute */
- NULL, /* sub */
- NULL, /* next */
- 0, /* static_pass_number */
- TV_JUMP, /* tv_id */
- 0, /* properties_required */
- 0, /* properties_provided */
- 0, /* properties_destroyed */
- TODO_ggc_collect, /* todo_flags_start */
- TODO_dump_func, /* todo_flags_finish */
- 'j' /* letter */
-};
-
-
Index: cfgrtl.c
===================================================================
/usr/bin/diff -dup -F'^(' -u -L cfgrtl.c (revision 107723) -L cfgrtl.c (working copy) .svn/text-base/cfgrtl.c.svn-base cfgrtl.c
--- cfgrtl.c (revision 107723)
+++ cfgrtl.c (working copy)
@@ -2130,6 +2130,18 @@ rtl_verify_flow_info_1 (void)
}
}
}
+ /* Register livenes sanity checks. */
+ FOR_EACH_BB_REVERSE (bb)
+ {
+ if (single_succ_p (bb) && bb->il.rtl->global_live_at_end)
+ gcc_assert (REG_SET_EQUAL_P (bb->il.rtl->global_live_at_end,
+ single_succ (bb)->il.rtl->global_live_at_start)
+ /* Before reload, calculate_global_regs_live sets some
+ registers -like the arg pointer - live in each basic
+ block, but not in the exit block. They still linger
+ in the life information after reload. */
+ || single_succ (bb) == EXIT_BLOCK_PTR);
+ }
/* Clean up. */
free (bb_info);
Index: reload1.c
===================================================================
/usr/bin/diff -dup -F'^(' -u -L reload1.c (revision 107723) -L reload1.c (working copy) .svn/text-base/reload1.c.svn-base reload1.c
--- reload1.c (revision 107723)
+++ reload1.c (working copy)
@@ -8287,6 +8287,7 @@ fixup_abnormal_edges (void)
bool inserted = false;
basic_block bb;
+ verify_flow_info ();
FOR_EACH_BB (bb)
{
edge e;