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 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);
}
\f
/* Return true iff the condbranches at the end of BB1 and BB2 match. */
-static bool
-condjump_equiv_p (basic_block bb1, basic_block bb2)
+bool
+condjump_equiv_p (struct equiv_info *info, bool call_init)
{
- edge b1, f1, b2, f2;
+ 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;
-
- 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))
!= (XEXP (SET_SRC (set2), 1) == pc_rtx))
reverse = !reverse;
- cond1 = XEXP (SET_SRC (set1), 0);
- cond2 = XEXP (SET_SRC (set2), 0);
+ 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));
if (code2 == UNKNOWN)
return false;
+ if (call_init && !struct_equiv_init (STRUCT_EQUIV_START | info->mode, info))
+ gcc_unreachable ();
+ /* Make the sources of the pc sets unreadable so that when we call
+ insns_match_p it won't process them.
+ 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. */
- 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 (code1 == code2)
+ {
+ 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));
+
+ }
+ else if (code1 == swap_condition (code2))
+ {
+ 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));
+
+ }
+ else
+ match = false;
+ SET_SRC (set1) = src1;
+ SET_SRC (set2) = src2;
+ match &= verify_changes (0);
/* If we return true, we will join the blocks. Which means that
we will only have one branch prediction bit to work with. Thus
"Outcomes of branch in bb %i and %i differ too much (%i %i)\n",
bb1->index, bb2->index, b1->probability, prob2);
- return false;
+ match = false;
}
}
fprintf (dump_file, "Conditionals in bb %i and %i match.\n",
bb1->index, bb2->index);
+ 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.
+
+/* Return true iff outgoing edges of INFO->y_block and INFO->x_block 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 in *MODE,
+ and pass *MODE to struct_equiv_init or assign it to INFO->mode, as
+ appropriate.
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;
& (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 (bb2))
|| !onlyjump_p (BB_END (bb2)))
return false;
- return condjump_equiv_p (bb1, bb2);
+ info->mode = *mode;
+ return condjump_equiv_p (info, true);
}
/* Generic case - we are seeing a computed jump, table jump or trapping
identical = false;
}
- if (identical)
+ if (identical
+ && struct_equiv_init (STRUCT_EQUIV_START | *mode, info))
{
- replace_label_data rr;
bool match;
- /* Temporarily replace references to LABEL1 with LABEL2
+ /* Indicate that LABEL1 is to be replaced 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);
+ info->y_label = label1;
+ info->x_label = label2;
- 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",
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);
-
return match;
}
}
/* 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
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.
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.
/* 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;
/* 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);
Software Foundation, 51 Franklin Street, Fifth Floor, Boston, MA
02110-1301, USA. */
-/* This file contains helper functions for Cross jumping (tail merging). */
+/* Try to match two basic blocks - or their ends - for structural equivalence.
+ We scan the blocks from their ends backwards, and expect that insns are
+ identical, except for certain cases involving registers. A mismatch
+ We scan the blocks from their ends backwards, hoping to find a match, I.e.
+ insns are identical, except for certain cases involving registers. A
+ mismatch between register number RX (used in block X) and RY (used in the
+ same way in block Y) can be handled in one of the following cases:
+ 1. RX and RY are local to their respective blocks; they are set there and
+ die there. If so, they can effectively be ignored.
+ 2. RX and RY die in their blocks, but live at the start. If any path
+ gets redirected through X instead of Y, the caller must emit
+ compensation code to move RY to RX. If there are overlapping inputs,
+ the function resolve_input_conflict ensures that this can be done.
+ Information about these registers are tracked in the X_LOCAL, Y_LOCAL,
+ LOCAL_COUNT and LOCAL_RVALUE fields.
+ 3. RX and RY live throughout their blocks, including the start and the end.
+ Either RX and RY must be identical, or we have to replace all uses in
+ block X with a new pseudo, which is stored in the INPUT_REG field. The
+ caller can then use block X instead of block Y by copying RY to the new
+ pseudo.
+
+ 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.
+
+ 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 liveness
+ 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 "tm_p.h"
#include "target.h"
#include "emit-rtl.h"
+#include "reload.h"
static void merge_memattrs (rtx, rtx);
+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);
+
+/* 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
\f
return;
}
+/* 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)
+{
+ unsigned regno = REGNO (reg);
+ int nregs, i, old;
+
+ if (regno >= FIRST_PSEUDO_REGISTER)
+ {
+ gcc_assert (!reload_completed);
+ nregs = 1;
+ }
+ else
+ nregs = hard_regno_nregs[regno][GET_MODE (reg)];
+ for (old = 0, i = nregs; --i >= 0; regno++)
+ {
+ 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 old;
+}
+
+/* Record state about current inputs / local registers / liveness
+ in *P. */
+static inline void
+struct_equiv_make_checkpoint (struct struct_equiv_checkpoint *p,
+ struct equiv_info *info)
+{
+ *p = info->cur;
+}
+
+/* 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)
+ {
+ 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);
+ }
+}
+
+/* Restore state about current inputs / local registers / liveness
+ 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)
+ {
+ info->cur.local_count--;
+ info->cur.version--;
+ if (REGNO_REG_SET_P (info->x_local_live,
+ REGNO (info->x_local[info->cur.local_count])))
+ {
+ 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--;
+ }
+ }
+ if (info->cur.version != p->version)
+ info->need_rerun = true;
+}
+
+
+/* Update register liveness to reflect that X is now life (if rvalue is
+ nonzero) or dead (if rvalue is zero) in INFO->x_block, and likewise Y
+ in INFO->y_block. Return the number of registers the liveness of which
+ changed in each block (as a negative number if registers became dead). */
+static int
+note_local_live (struct equiv_info *info, rtx x, rtx y, int rvalue)
+{
+ int x_change = assign_reg_reg_set (info->x_local_live, x, rvalue);
+ int y_change = assign_reg_reg_set (info->y_local_live, y, rvalue);
+
+ gcc_assert (x_change == y_change);
+ if (y_change)
+ {
+ if (reload_completed)
+ {
+ unsigned x_regno ATTRIBUTE_UNUSED = REGNO (x);
+ unsigned y_regno = REGNO (y);
+ enum machine_mode x_mode = GET_MODE (x);
+
+ 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
+ )
+ y_change *= IMPOSSIBLE_MOVE_FACTOR;
+ }
+ info->cur.input_count += y_change;
+ info->cur.version++;
+ }
+ return x_change;
+}
+
+/* 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, int rvalue, struct equiv_info *info)
+{
+ rtx x = *xp;
+ enum rtx_code code;
+ int length;
+ const char *format;
+ int i;
+
+ if (!y || !x)
+ return x == y;
+ code = GET_CODE (y);
+ if (code != REG && x == y)
+ return true;
+ if (GET_CODE (x) != code
+ || GET_MODE (x) != GET_MODE (y))
+ return false;
+
+ /* ??? could extend to allow CONST_INT inputs. */
+ switch (code)
+ {
+ case SUBREG:
+ gcc_assert (!reload_completed
+ || !info->live_update);
+ break;
+ case REG:
+ {
+ unsigned x_regno = REGNO (x);
+ unsigned y_regno = REGNO (y);
+ int x_common_live, y_common_live;
+
+ if (reload_completed
+ && (x_regno >= FIRST_PSEUDO_REGISTER
+ || y_regno >= FIRST_PSEUDO_REGISTER))
+ {
+ /* We should only see this in REG_NOTEs. */
+ gcc_assert (!info->live_update);
+ /* Returning false will cause us to remove the notes. */
+ return false;
+ }
+#ifdef STACK_REGS
+ /* After reg-stack, can only accept literal matches of stack regs. */
+ if (info->mode & CLEANUP_POST_REGSTACK
+ && (IN_RANGE (x_regno, FIRST_STACK_REG, LAST_STACK_REG)
+ || IN_RANGE (y_regno, FIRST_STACK_REG, LAST_STACK_REG)))
+ return x_regno == y_regno;
+#endif
+
+ /* 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;
+
+ /* Update liveness information. */
+ if (info->live_update
+ && assign_reg_reg_set (info->common_live, x, rvalue))
+ info->cur.version++;
+
+ return true;
+ }
+
+ 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 local_count = info->cur.local_count;
+
+ /* 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]));
+
+ /* 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 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;
+ }
+
+ /* 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++;
+ }
+ note_local_live (info, x, y, rvalue);
+ }
+ return true;
+ }
+ case SET:
+ gcc_assert (rvalue < 0);
+ /* Ignore the destinations role as a destination. Still, we have
+ to consider input registers embedded in the addresses of a MEM.
+ N.B., we process the rvalue aspect of STRICT_LOW_PART /
+ ZERO_EXTEND / SIGN_EXTEND along with their lvalue aspect. */
+ if(!set_dest_addr_equiv_p (SET_DEST (x), SET_DEST (y), info))
+ return false;
+ /* Process source. */
+ return rtx_equiv_p (&SET_SRC (x), SET_SRC (y), 1, info);
+ case PRE_MODIFY:
+ /* Process destination. */
+ if (!rtx_equiv_p (&XEXP (x, 0), XEXP (y, 0), 0, info))
+ return false;
+ /* Process source. */
+ return rtx_equiv_p (&XEXP (x, 1), XEXP (y, 1), 1, info);
+ case POST_MODIFY:
+ {
+ rtx x_dest0, x_dest1;
+
+ /* Process destination. */
+ x_dest0 = XEXP (x, 0);
+ gcc_assert (REG_P (x_dest0));
+ if (!rtx_equiv_p (&XEXP (x, 0), XEXP (y, 0), 0, info))
+ return false;
+ x_dest1 = XEXP (x, 0);
+ /* validate_change might have changed the destination. Put it back
+ so that we can do a valid source match. */
+ XEXP (x, 0) = x_dest0;
+ if (!rtx_equiv_p (&XEXP (x, 1), XEXP (y, 1), 0, info))
+ return false;
+ gcc_assert (x_dest1 == XEXP (x, 0));
+ /* Process source. */
+ return rtx_equiv_p (&XEXP (x, 1), XEXP (y, 1), 1, info);
+ if (!rtx_equiv_p (&XEXP(x, 0), XEXP (y, 0), 0, info))
+ return false;
+ /* Process both subexpressions as inputs. */
+ break;
+ }
+ case CLOBBER:
+ gcc_assert (rvalue < 0);
+ return true;
+ /* Some special forms are also rvalues when they appear in lvalue
+ positions. However, we must ont try to match a register after we
+ have already altered it with validate_change, consider the rvalue
+ aspect while we process the lvalue. */
+ case STRICT_LOW_PART:
+ case ZERO_EXTEND:
+ case SIGN_EXTEND:
+ {
+ rtx x_inner, y_inner;
+ enum rtx_code code;
+ int change;
+
+ if (rvalue)
+ break;
+ x_inner = XEXP (x, 0);
+ y_inner = XEXP (y, 0);
+ if (GET_MODE (x_inner) != GET_MODE (y_inner))
+ return false;
+ code = GET_CODE (x_inner);
+ if (code != GET_CODE (y_inner))
+ return false;
+ /* The address of a MEM is an input that will be processed during
+ rvalue == -1 processing. */
+ if (code == SUBREG)
+ {
+ if (SUBREG_BYTE (x_inner) != SUBREG_BYTE (y_inner))
+ return false;
+ x = x_inner;
+ x_inner = SUBREG_REG (x_inner);
+ y_inner = SUBREG_REG (y_inner);
+ if (GET_MODE (x_inner) != GET_MODE (y_inner))
+ return false;
+ code = GET_CODE (x_inner);
+ if (code != GET_CODE (y_inner))
+ return false;
+ }
+ if (code == MEM)
+ return true;
+ gcc_assert (code == REG);
+ if (! rtx_equiv_p (&XEXP (x, 0), y_inner, rvalue, info))
+ return false;
+ if (REGNO (x_inner) == REGNO (y_inner))
+ {
+ change = assign_reg_reg_set (info->common_live, x_inner, 1);
+ info->cur.version++;
+ }
+ else
+ change = note_local_live (info, x_inner, y_inner, 1);
+ gcc_assert (change);
+ return true;
+ }
+ /* 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. */
+ case MEM:
+ return !rvalue || rtx_equiv_p (&XEXP (x, 0), XEXP (y, 0), rvalue, 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:
+ gcc_assert (rvalue < 0);
+ break;
+ case LABEL_REF:
+ /* Check special tablejump match case. */
+ if (XEXP (y, 0) == info->y_label)
+ return (XEXP (x, 0) == info->x_label);
+ /* 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);
+
+ /* 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;
+ }
+
+ /* 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)));
+
+ /* Process subexpressions - this is similar to rtx_equal_p. */
+ length = GET_RTX_LENGTH (code);
+ format = GET_RTX_FORMAT (code);
+
+ 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;
+}
+
+/* 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) != GET_CODE (y))
+ return false;
+ else 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)
+ {
+ int j;
+
+ if (XVECLEN (x, 0) != XVECLEN (y, 0))
+ return false;
+ for (j = 0; j < XVECLEN (x, 0); ++j)
+ {
+ rtx xe = XVECEXP (x, 0, j);
+ rtx ye = XVECEXP (y, 0, j);
+
+ 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;
+}
+
+/* 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;
+
+ if (code != GET_CODE (y))
+ return false;
+ if (code == MEM)
+ return rtx_equiv_p (&XEXP (x, 0), XEXP (y, 0), 1, info);
+
+ /* Process subexpressions. */
+ length = GET_RTX_LENGTH (code);
+ format = GET_RTX_FORMAT (code);
+
+ for (i = 0; i < length; ++i)
+ {
+ 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;
+ }
+ }
+ return true;
+}
+
/* 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. */
+ 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
death_notes_match_p (rtx i1 ATTRIBUTE_UNUSED, rtx i2 ATTRIBUTE_UNUSED,
- int mode ATTRIBUTE_UNUSED)
+ struct equiv_info *info ATTRIBUTE_UNUSED)
{
#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. */
+ indicates whether or not the insn contains any stack-like regs. */
- if ((mode & CLEANUP_POST_REGSTACK) && stack_regs_mentioned (i1))
+ if ((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
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)));
+ {
+ unsigned regno = REGNO (XEXP (note, 0));
+ int i;
+
+ 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);
+ }
GO_IF_HARD_REG_EQUAL (i1_regset, i2_regset, done);
/* Return true if I1 and I2 are equivalent and thus can be crossjumped. */
bool
-insns_match_p (int mode, rtx i1, rtx i2)
+insns_match_p (rtx i1, rtx i2, struct equiv_info *info)
{
- rtx p1, p2;
+ int rvalue_change_start;
+ struct struct_equiv_checkpoint before_rvalue_change;
/* 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;
+ info->cur.x_start = i1;
+ info->cur.y_start = i2;
/* If this is a CALL_INSN, compare register usage information.
If we don't check this on stack register machines, the two
??? 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;
-
- if (!death_notes_match_p (i1, i2, mode))
- return false;
-
- if (reload_completed
- ? rtx_renumbered_equal_p (p1, p2) : rtx_equal_p (p1, p2))
+ if (CALL_P (i1))
+ {
+ 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))
+ {
+ cancel_changes (0);
+ return false;
+ }
+ }
+ else if (INSN_P (i1))
+ {
+ if (! set_dest_equiv_p (PATTERN (i1), PATTERN (i2), info))
+ {
+ cancel_changes (0);
+ return false;
+ }
+ }
+ 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;
/* Do not do EQUIV substitution after reload. First, we're undoing the
targets to disallow the troublesome insns after splitting. */
if (!reload_completed)
{
- /* 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);
+ rtx equiv1, equiv2;
+ cancel_changes (rvalue_change_start);
+ struct_equiv_restore_checkpoint (&before_rvalue_change, info);
+
+ /* 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
{
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)))
+
+ 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);
- if (! rtx_renumbered_equal_p (p1, p2))
- cancel_changes (0);
- else if (apply_change_group ())
- return true;
+ /* 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;
+ }
}
}
}
+ cancel_changes (0);
return false;
}
-\f
-/* 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. */
+/* Set up mode and register information in INFO. Return true for success. */
+bool
+struct_equiv_init (int mode, struct equiv_info *info)
+{
+ if ((info->x_block->flags | info->y_block->flags) & BB_DIRTY)
+ update_life_info_in_dirty_blocks (UPDATE_LIFE_GLOBAL_RM_NOTES,
+ (PROP_DEATH_NOTES
+ | ((mode & CLEANUP_POST_REGSTACK)
+ ? PROP_POST_REGSTACK : 0)));
+ if (!REG_SET_EQUAL_P (info->x_block->il.rtl->global_live_at_end,
+ info->y_block->il.rtl->global_live_at_end))
+ {
+#ifdef STACK_REGS
+ unsigned rn;
+
+ if (!(mode & CLEANUP_POST_REGSTACK))
+ return false;
+ /* After reg-stack. Remove bogus live info about stack regs. N.B.
+ these regs are not necessarily all dead - we swap random bogosity
+ against constant bogosity. However, clearing these bits at
+ least makes the regsets comparable. */
+ for (rn = FIRST_STACK_REG; rn < LAST_STACK_REG; rn++)
+ {
+ CLEAR_REGNO_REG_SET (info->x_block->il.rtl->global_live_at_end, rn);
+ CLEAR_REGNO_REG_SET (info->y_block->il.rtl->global_live_at_end, rn);
+ }
+ if (!REG_SET_EQUAL_P (info->x_block->il.rtl->global_live_at_end,
+ info->y_block->il.rtl->global_live_at_end))
+#endif
+ return false;
+ }
+ 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->x_label = info->y_label = 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;
+}
+
+/* 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)
+{
+ rtx equiv1, equiv2;
+
+ merge_memattrs (xi, yi);
+
+ /* 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;
+}
+/* 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
-flow_find_cross_jump (int mode ATTRIBUTE_UNUSED, basic_block bb1,
- basic_block bb2, rtx *f1, rtx *f2)
+struct_equiv_block_eq (int mode, struct equiv_info *info)
{
- rtx i1, i2, last1, last2, afterlast1, afterlast2;
- int ninsns = 0;
+ rtx x_stop, y_stop;
+ rtx xi, yi;
+ int i;
+
+ 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))
+ gcc_unreachable ();
/* 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))))
+ xi = BB_END (info->x_block);
+ if (onlyjump_p (xi)
+ || (returnjump_p (xi) && !side_effects_p (PATTERN (xi))))
{
- last1 = i1;
- i1 = PREV_INSN (i1);
+ info->cur.x_start = xi;
+ xi = PREV_INSN (xi);
}
- i2 = BB_END (bb2);
- if (onlyjump_p (i2)
- || (returnjump_p (i2) && !side_effects_p (PATTERN (i2))))
+ yi = BB_END (info->y_block);
+ if (onlyjump_p (yi)
+ || (returnjump_p (yi) && !side_effects_p (PATTERN (yi))))
{
- last2 = i2;
+ info->cur.y_start = yi;
/* Count everything except for unconditional jump as insn. */
- if (!simplejump_p (i2) && !returnjump_p (i2) && last1)
- ninsns++;
- i2 = PREV_INSN (i2);
+ /* ??? 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);
}
- while (true)
+ if (mode & STRUCT_EQUIV_MATCH_JUMPS)
{
- /* Ignore notes. */
- while (!INSN_P (i1) && i1 != BB_HEAD (bb1))
- i1 = PREV_INSN (i1);
+ /* 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)
+ {
+ if (any_condjump_p (info->cur.x_start)
+ ? !condjump_equiv_p (info, false)
+ : !insns_match_p (info->cur.x_start, info->cur.y_start, info))
+ gcc_unreachable ();
+ }
+ else if (any_condjump_p (xi) && any_condjump_p (yi))
+ {
+ 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, false))
+ gcc_unreachable ();
+ }
+ if (info->cur.x_start && info->mode & STRUCT_EQUIV_FINAL)
+ struct_equiv_merge (info->cur.x_start, info->cur.y_start, info);
+ }
- while (!INSN_P (i2) && i2 != BB_HEAD (bb2))
- i2 = PREV_INSN (i2);
+ 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);
- if (i1 == BB_HEAD (bb1) || i2 == BB_HEAD (bb2))
- break;
+ while (!INSN_P (yi) && yi != y_stop)
+ yi = PREV_INSN (yi);
+
+ 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 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);
+
+ /* 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);
- if (!insns_match_p (mode, i1, i2))
- break;
+ while (yi != y_stop && !INSN_P (PREV_INSN (yi)))
+ yi = PREV_INSN (yi);
- merge_memattrs (i1, i2);
+ info->cur.x_start = xi;
+ info->cur.y_start = yi;
+ }
- /* Don't begin a cross-jump with a NOTE insn. */
- if (INSN_P (i1))
+ if (!info->cur.input_valid)
+ info->x_input = info->y_input = info->input_reg = NULL_RTX;
+ if (!info->need_rerun)
+ {
+ find_dying_inputs (info);
+ if (info->mode & STRUCT_EQUIV_FINAL)
+ {
+ if (info->check_input_conflict && ! resolve_input_conflict (info))
+ gcc_unreachable ();
+ }
+ else
{
- /* 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)))
+ bool input_conflict = info->had_input_conflict;
+
+ if (!input_conflict
+ && info->dying_inputs > 1
+ && bitmap_intersect_p (info->x_local_live, info->y_local_live))
{
- remove_note (i1, equiv1);
- remove_note (i2, equiv2);
+ regset_head clobbered_regs;
+
+ INIT_REG_SET (&clobbered_regs);
+ for (i = 0; i < info->cur.local_count; i++)
+ {
+ if (assign_reg_reg_set (&clobbered_regs, info->y_local[i], 0))
+ {
+ input_conflict = true;
+ break;
+ }
+ assign_reg_reg_set (&clobbered_regs, info->x_local[i], 1);
+ }
+ CLEAR_REG_SET (&clobbered_regs);
}
-
- afterlast1 = last1, afterlast2 = last2;
- last1 = i1, last2 = i2;
- ninsns++;
+ if (input_conflict && !info->check_input_conflict)
+ info->need_rerun = true;
+ info->check_input_conflict = input_conflict;
}
-
- 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
+ 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;
+}
- /* 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);
+/* 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)
+{
+ int i;
- if (last1 != BB_HEAD (bb1) && LABEL_P (PREV_INSN (last1)))
- last1 = PREV_INSN (last1);
+ info->dying_inputs = 0;
+ for (i = info->cur.local_count-1; i >=0; i--)
+ {
+ rtx x = info->x_local[i];
+ unsigned regno = REGNO (x);
+ int nregs = (regno >= FIRST_PSEUDO_REGISTER
+ ? 1 : hard_regno_nregs[regno][GET_MODE (x)]);
+
+ 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;
+ }
+ }
+}
- while (last2 != BB_HEAD (bb2) && !INSN_P (PREV_INSN (last2)))
- last2 = PREV_INSN (last2);
+/* 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)
+{
+ int i, j, end;
+ int nswaps = 0;
+ rtx save_x_local[STRUCT_EQUIV_MAX_LOCAL];
+ rtx save_y_local[STRUCT_EQUIV_MAX_LOCAL];
- if (last2 != BB_HEAD (bb2) && LABEL_P (PREV_INSN (last2)))
- last2 = PREV_INSN (last2);
+ 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++)
+ {
+ /* 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;
+
+ /* 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++)
+ {
+ rtx tmp;
- *f1 = last1;
- *f2 = last2;
+ 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;
+ }
}
-
- return ninsns;
+ 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;
}
OSDN Git Service
