patch for latest lra changes.
Vladimir Makarov
vmakarov@redhat.com
Wed Apr 24 18:07:00 GMT 2013
The following patch incorporates some LRA changes on lra and mike-lra
branches. I am committing them to get a better testing of them for
x86/x86-64 on trunk.
The patch was successfully tested and bootstrapped on x86/x86-64.
Committed as rev. 198252.
2013-04-24 Vladimir Makarov <vmakarov@redhat.com>
* rtl.h (struct rtx_def): Add comment for field jump.
(LRA_SUBREG_P): New macro.
* recog.c (register_operand): Check LRA_SUBREG_P.
* lra.c (lra): Add note at the end of RTL code. Align non-empty
stack frame.
* lra-spills.c (lra_spill): Align stack after spilling pseudos.
(lra_final_code_change): Skip subreg change for operators.
* lra-eliminations.c (eliminate_regs_in_insn): Make return earlier
if there are no operand changes.
* lra-constraints.c (curr_insn_set): New.
(match_reload): Set LRA_SUBREG_P.
(emit_spill_move): Ditto.
(check_and_process_move): Use curr_insn_set. Process only single
set insns. Don't initialize sec_mem_p and change_p.
(simplify_operand_subreg): Use LRA_SUBREG_P.
(reg_in_class_p): New function.
(process_alt_operands): Use it. Use #if HAVE_ATTR_enabled instead
of #ifdef. Add code to remove cycling.
(process_address): Check EXTRA_CONSTRAINT_STR. Process even if
non-null disp. Reload inner instead of disp when base and index
are null. Try to put lo_sum into register.
(EBB_PROBABILITY_CUTOFF): Redefine probability in percents.
(check_and_process_move): Move code for move cost check to
simple_move_p. Remove equiv_substitution.
(simple_move_p): New function.
(curr_insn_transform): Initialize sec_mem_p and change_p. Set up
curr_insn_set. Call check_and_process_move only for single set
insns. Use the new function. Move call of check_and_process_move
after operand equiv substitution and address process.
-------------- next part --------------
Index: rtl.h
===================================================================
--- rtl.h (revision 198251)
+++ rtl.h (working copy)
@@ -265,7 +265,8 @@ struct GTY((chain_next ("RTX_NEXT (&%h)"
1 in a SET that is for a return.
In a CODE_LABEL, part of the two-bit alternate entry field.
1 in a CONCAT is VAL_EXPR_IS_COPIED in var-tracking.c.
- 1 in a VALUE is SP_BASED_VALUE_P in cselib.c. */
+ 1 in a VALUE is SP_BASED_VALUE_P in cselib.c.
+ 1 in a SUBREG generated by LRA for reload insns. */
unsigned int jump : 1;
/* In a CODE_LABEL, part of the two-bit alternate entry field.
1 in a MEM if it cannot trap.
@@ -1411,6 +1412,11 @@ do { \
((RTL_FLAG_CHECK1("SUBREG_PROMOTED_UNSIGNED_P", (RTX), SUBREG)->volatil) \
? -1 : (int) (RTX)->unchanging)
+/* True if the subreg was generated by LRA for reload insns. Such
+ subregs are valid only during LRA. */
+#define LRA_SUBREG_P(RTX) \
+ (RTL_FLAG_CHECK1("LRA_SUBREG_P", (RTX), SUBREG)->jump)
+
/* Access various components of an ASM_OPERANDS rtx. */
#define ASM_OPERANDS_TEMPLATE(RTX) XCSTR (RTX, 0, ASM_OPERANDS)
Index: recog.c
===================================================================
--- recog.c (revision 198251)
+++ recog.c (working copy)
@@ -1065,7 +1065,11 @@ register_operand (rtx op, enum machine_m
&& REGNO (sub) < FIRST_PSEUDO_REGISTER
&& REG_CANNOT_CHANGE_MODE_P (REGNO (sub), GET_MODE (sub), mode)
&& GET_MODE_CLASS (GET_MODE (sub)) != MODE_COMPLEX_INT
- && GET_MODE_CLASS (GET_MODE (sub)) != MODE_COMPLEX_FLOAT)
+ && GET_MODE_CLASS (GET_MODE (sub)) != MODE_COMPLEX_FLOAT
+ /* LRA can generate some invalid SUBREGS just for matched
+ operand reload presentation. LRA needs to treat them as
+ valid. */
+ && ! LRA_SUBREG_P (op))
return 0;
#endif
Index: lra.c
===================================================================
--- lra.c (revision 198251)
+++ lra.c (working copy)
@@ -2203,6 +2203,10 @@ lra (FILE *f)
timevar_push (TV_LRA);
+ /* Make sure that the last insn is a note. Some subsequent passes
+ need it. */
+ emit_note (NOTE_INSN_DELETED);
+
COPY_HARD_REG_SET (lra_no_alloc_regs, ira_no_alloc_regs);
init_reg_info ();
@@ -2259,6 +2263,11 @@ lra (FILE *f)
bitmap_initialize (&lra_split_regs, ®_obstack);
bitmap_initialize (&lra_optional_reload_pseudos, ®_obstack);
live_p = false;
+ if (get_frame_size () != 0 && crtl->stack_alignment_needed)
+ /* If we have a stack frame, we must align it now. The stack size
+ may be a part of the offset computation for register
+ elimination. */
+ assign_stack_local (BLKmode, 0, crtl->stack_alignment_needed);
for (;;)
{
for (;;)
Index: lra-spills.c
===================================================================
--- lra-spills.c (revision 198251)
+++ lra-spills.c (working copy)
@@ -548,6 +548,11 @@ lra_spill (void)
for (i = 0; i < n; i++)
if (pseudo_slots[pseudo_regnos[i]].mem == NULL_RTX)
assign_mem_slot (pseudo_regnos[i]);
+ if (n > 0 && crtl->stack_alignment_needed)
+ /* If we have a stack frame, we must align it now. The stack size
+ may be a part of the offset computation for register
+ elimination. */
+ assign_stack_local (BLKmode, 0, crtl->stack_alignment_needed);
if (lra_dump_file != NULL)
{
for (i = 0; i < slots_num; i++)
@@ -644,10 +649,12 @@ lra_final_code_change (void)
}
lra_insn_recog_data_t id = lra_get_insn_recog_data (insn);
+ struct lra_static_insn_data *static_id = id->insn_static_data;
bool insn_change_p = false;
for (i = id->insn_static_data->n_operands - 1; i >= 0; i--)
- if (alter_subregs (id->operand_loc[i], ! DEBUG_INSN_P (insn)))
+ if ((DEBUG_INSN_P (insn) || ! static_id->operand[i].is_operator)
+ && alter_subregs (id->operand_loc[i], ! DEBUG_INSN_P (insn)))
{
lra_update_dup (id, i);
insn_change_p = true;
Index: lra-eliminations.c
===================================================================
--- lra-eliminations.c (revision 198251)
+++ lra-eliminations.c (working copy)
@@ -977,6 +977,9 @@ eliminate_regs_in_insn (rtx insn, bool r
}
}
+ if (! validate_p)
+ return;
+
/* Substitute the operands; the new values are in the substed_operand
array. */
for (i = 0; i < static_id->n_operands; i++)
@@ -984,16 +987,13 @@ eliminate_regs_in_insn (rtx insn, bool r
for (i = 0; i < static_id->n_dups; i++)
*id->dup_loc[i] = substed_operand[(int) static_id->dup_num[i]];
- if (validate_p)
- {
- /* If we had a move insn but now we don't, re-recognize it.
- This will cause spurious re-recognition if the old move had a
- PARALLEL since the new one still will, but we can't call
- single_set without having put new body into the insn and the
- re-recognition won't hurt in this rare case. */
- id = lra_update_insn_recog_data (insn);
- static_id = id->insn_static_data;
- }
+ /* If we had a move insn but now we don't, re-recognize it.
+ This will cause spurious re-recognition if the old move had a
+ PARALLEL since the new one still will, but we can't call
+ single_set without having put new body into the insn and the
+ re-recognition won't hurt in this rare case. */
+ id = lra_update_insn_recog_data (insn);
+ static_id = id->insn_static_data;
}
/* Spill pseudos which are assigned to hard registers in SET. Add
Index: lra-constraints.c
===================================================================
--- lra-constraints.c (revision 198251)
+++ lra-constraints.c (working copy)
@@ -135,10 +135,11 @@
reload insns. */
static int bb_reload_num;
-/* The current insn being processed and corresponding its data (basic
- block, the insn data, the insn static data, and the mode of each
- operand). */
+/* The current insn being processed and corresponding its single set
+ (NULL otherwise), its data (basic block, the insn data, the insn
+ static data, and the mode of each operand). */
static rtx curr_insn;
+static rtx curr_insn_set;
static basic_block curr_bb;
static lra_insn_recog_data_t curr_id;
static struct lra_static_insn_data *curr_static_id;
@@ -698,6 +699,7 @@ match_reload (signed char out, signed ch
new_out_reg = gen_lowpart_SUBREG (outmode, reg);
else
new_out_reg = gen_rtx_SUBREG (outmode, reg, 0);
+ LRA_SUBREG_P (new_out_reg) = 1;
/* If the input reg is dying here, we can use the same hard
register for REG and IN_RTX. We do it only for original
pseudos as reload pseudos can die although original
@@ -721,6 +723,7 @@ match_reload (signed char out, signed ch
it at the end of LRA work. */
clobber = emit_clobber (new_out_reg);
LRA_TEMP_CLOBBER_P (PATTERN (clobber)) = 1;
+ LRA_SUBREG_P (new_in_reg) = 1;
if (GET_CODE (in_rtx) == SUBREG)
{
rtx subreg_reg = SUBREG_REG (in_rtx);
@@ -855,40 +858,34 @@ static rtx
emit_spill_move (bool to_p, rtx mem_pseudo, rtx val)
{
if (GET_MODE (mem_pseudo) != GET_MODE (val))
- val = gen_rtx_SUBREG (GET_MODE (mem_pseudo),
- GET_CODE (val) == SUBREG ? SUBREG_REG (val) : val,
- 0);
+ {
+ val = gen_rtx_SUBREG (GET_MODE (mem_pseudo),
+ GET_CODE (val) == SUBREG ? SUBREG_REG (val) : val,
+ 0);
+ LRA_SUBREG_P (val) = 1;
+ }
return (to_p
- ? gen_move_insn (mem_pseudo, val)
- : gen_move_insn (val, mem_pseudo));
+ ? gen_move_insn (mem_pseudo, val)
+ : gen_move_insn (val, mem_pseudo));
}
/* Process a special case insn (register move), return true if we
- don't need to process it anymore. Return that RTL was changed
- through CHANGE_P and macro SECONDARY_MEMORY_NEEDED says to use
- secondary memory through SEC_MEM_P. */
+ don't need to process it anymore. INSN should be a single set
+ insn. Set up that RTL was changed through CHANGE_P and macro
+ SECONDARY_MEMORY_NEEDED says to use secondary memory through
+ SEC_MEM_P. */
static bool
-check_and_process_move (bool *change_p, bool *sec_mem_p)
+check_and_process_move (bool *change_p, bool *sec_mem_p ATTRIBUTE_UNUSED)
{
int sregno, dregno;
- rtx set, dest, src, dreg, sreg, old_sreg, new_reg, before, scratch_reg;
+ rtx dest, src, dreg, sreg, old_sreg, new_reg, before, scratch_reg;
enum reg_class dclass, sclass, secondary_class;
enum machine_mode sreg_mode;
secondary_reload_info sri;
- *sec_mem_p = *change_p = false;
- if ((set = single_set (curr_insn)) == NULL)
- return false;
- dreg = dest = SET_DEST (set);
- sreg = src = SET_SRC (set);
- /* Quick check on the right move insn which does not need
- reloads. */
- if ((dclass = get_op_class (dest)) != NO_REGS
- && (sclass = get_op_class (src)) != NO_REGS
- /* The backend guarantees that register moves of cost 2 never
- need reloads. */
- && targetm.register_move_cost (GET_MODE (src), dclass, sclass) == 2)
- return true;
+ lra_assert (curr_insn_set != NULL_RTX);
+ dreg = dest = SET_DEST (curr_insn_set);
+ sreg = src = SET_SRC (curr_insn_set);
if (GET_CODE (dest) == SUBREG)
dreg = SUBREG_REG (dest);
if (GET_CODE (src) == SUBREG)
@@ -896,7 +893,6 @@ check_and_process_move (bool *change_p,
if (! REG_P (dreg) || ! REG_P (sreg))
return false;
sclass = dclass = NO_REGS;
- dreg = get_equiv_substitution (dreg);
if (REG_P (dreg))
dclass = get_reg_class (REGNO (dreg));
if (dclass == ALL_REGS)
@@ -910,7 +906,6 @@ check_and_process_move (bool *change_p,
return false;
sreg_mode = GET_MODE (sreg);
old_sreg = sreg;
- sreg = get_equiv_substitution (sreg);
if (REG_P (sreg))
sclass = get_reg_class (REGNO (sreg));
if (sclass == ALL_REGS)
@@ -1007,7 +1002,7 @@ check_and_process_move (bool *change_p,
if (GET_CODE (src) == SUBREG)
SUBREG_REG (src) = new_reg;
else
- SET_SRC (set) = new_reg;
+ SET_SRC (curr_insn_set) = new_reg;
}
else
{
@@ -1204,7 +1199,10 @@ simplify_operand_subreg (int nop, enum m
&& (hard_regno_nregs[hard_regno][GET_MODE (reg)]
>= hard_regno_nregs[hard_regno][mode])
&& simplify_subreg_regno (hard_regno, GET_MODE (reg),
- SUBREG_BYTE (operand), mode) < 0)
+ SUBREG_BYTE (operand), mode) < 0
+ /* Don't reload subreg for matching reload. It is actually
+ valid subreg in LRA. */
+ && ! LRA_SUBREG_P (operand))
|| CONSTANT_P (reg) || GET_CODE (reg) == PLUS || MEM_P (reg))
{
enum op_type type = curr_static_id->operand[nop].type;
@@ -1311,6 +1309,14 @@ general_constant_p (rtx x)
return CONSTANT_P (x) && (! flag_pic || LEGITIMATE_PIC_OPERAND_P (x));
}
+static bool
+reg_in_class_p (rtx reg, enum reg_class cl)
+{
+ if (cl == NO_REGS)
+ return get_reg_class (REGNO (reg)) == NO_REGS;
+ return in_class_p (reg, cl, NULL);
+}
+
/* Major function to choose the current insn alternative and what
operands should be reloaded and how. If ONLY_ALTERNATIVE is not
negative we should consider only this alternative. Return false if
@@ -1390,7 +1396,7 @@ process_alt_operands (int only_alternati
for (nalt = 0; nalt < n_alternatives; nalt++)
{
/* Loop over operands for one constraint alternative. */
-#ifdef HAVE_ATTR_enabled
+#if HAVE_ATTR_enabled
if (curr_id->alternative_enabled_p != NULL
&& ! curr_id->alternative_enabled_p[nalt])
continue;
@@ -2054,6 +2060,31 @@ process_alt_operands (int only_alternati
if (early_clobber_p && operand_reg[nop] != NULL_RTX)
early_clobbered_nops[early_clobbered_regs_num++] = nop;
}
+ if (curr_insn_set != NULL_RTX && n_operands == 2
+ && ((! curr_alt_win[0] && ! curr_alt_win[1]
+ && REG_P (no_subreg_reg_operand[0])
+ && REG_P (no_subreg_reg_operand[1])
+ && (reg_in_class_p (no_subreg_reg_operand[0], curr_alt[1])
+ || reg_in_class_p (no_subreg_reg_operand[1], curr_alt[0])))
+ || (! curr_alt_win[0] && curr_alt_win[1]
+ && REG_P (no_subreg_reg_operand[1])
+ && reg_in_class_p (no_subreg_reg_operand[1], curr_alt[0]))
+ || (curr_alt_win[0] && ! curr_alt_win[1]
+ && REG_P (no_subreg_reg_operand[0])
+ && reg_in_class_p (no_subreg_reg_operand[0], curr_alt[1])
+ && (! CONST_POOL_OK_P (curr_operand_mode[1],
+ no_subreg_reg_operand[1])
+ || (targetm.preferred_reload_class
+ (no_subreg_reg_operand[1],
+ (enum reg_class) curr_alt[1]) != NO_REGS))
+ /* If it is a result of recent elimination in move
+ insn we can transform it into an add still by
+ using this alternative. */
+ && GET_CODE (no_subreg_reg_operand[1]) != PLUS)))
+ /* We have a move insn and a new reload insn will be similar
+ to the current insn. We should avoid such situation as it
+ results in LRA cycling. */
+ overall += LRA_MAX_REJECT;
ok_p = true;
curr_alt_dont_inherit_ops_num = 0;
for (nop = 0; nop < early_clobbered_regs_num; nop++)
@@ -2425,27 +2456,35 @@ process_address (int nop, rtx *before, r
&& process_addr_reg (ad.index_term, before, NULL, INDEX_REG_CLASS))
change_p = true;
+#ifdef EXTRA_CONSTRAINT_STR
+ /* Target hooks sometimes reject extra constraint addresses -- use
+ EXTRA_CONSTRAINT_STR for the validation. */
+ if (constraint[0] != 'p'
+ && EXTRA_ADDRESS_CONSTRAINT (constraint[0], constraint)
+ && EXTRA_CONSTRAINT_STR (op, constraint[0], constraint))
+ return change_p;
+#endif
+
/* There are three cases where the shape of *AD.INNER may now be invalid:
1) the original address was valid, but either elimination or
- equiv_address_substitution applied a displacement that made
- it invalid.
+ equiv_address_substitution was applied and that made
+ the address invalid.
2) the address is an invalid symbolic address created by
force_const_to_mem.
3) the address is a frame address with an invalid offset.
- All these cases involve a displacement and a non-autoinc address,
- so there is no point revalidating other types. */
- if (ad.disp == NULL || ad.autoinc_p || valid_address_p (&ad))
+ All these cases involve a non-autoinc address, so there is no
+ point revalidating other types. */
+ if (ad.autoinc_p || valid_address_p (&ad))
return change_p;
/* Any index existed before LRA started, so we can assume that the
presence and shape of the index is valid. */
push_to_sequence (*before);
- gcc_assert (ad.segment == NULL);
- gcc_assert (ad.disp == ad.disp_term);
+ lra_assert (ad.disp == ad.disp_term);
if (ad.base == NULL)
{
if (ad.index == NULL)
@@ -2453,26 +2492,39 @@ process_address (int nop, rtx *before, r
int code = -1;
enum reg_class cl = base_reg_class (ad.mode, ad.as,
SCRATCH, SCRATCH);
- rtx disp = *ad.disp;
+ rtx addr = *ad.inner;
- new_reg = lra_create_new_reg (Pmode, NULL_RTX, cl, "disp");
+ new_reg = lra_create_new_reg (Pmode, NULL_RTX, cl, "addr");
#ifdef HAVE_lo_sum
{
rtx insn;
rtx last = get_last_insn ();
- /* disp => lo_sum (new_base, disp), case (2) above. */
+ /* addr => lo_sum (new_base, addr), case (2) above. */
insn = emit_insn (gen_rtx_SET
(VOIDmode, new_reg,
- gen_rtx_HIGH (Pmode, copy_rtx (disp))));
+ gen_rtx_HIGH (Pmode, copy_rtx (addr))));
code = recog_memoized (insn);
if (code >= 0)
{
- *ad.disp = gen_rtx_LO_SUM (Pmode, new_reg, disp);
+ *ad.inner = gen_rtx_LO_SUM (Pmode, new_reg, addr);
if (! valid_address_p (ad.mode, *ad.outer, ad.as))
{
- *ad.disp = disp;
- code = -1;
+ /* Try to put lo_sum into register. */
+ insn = emit_insn (gen_rtx_SET
+ (VOIDmode, new_reg,
+ gen_rtx_LO_SUM (Pmode, new_reg, addr)));
+ code = recog_memoized (insn);
+ if (code >= 0)
+ {
+ *ad.inner = new_reg;
+ if (! valid_address_p (ad.mode, *ad.outer, ad.as))
+ {
+ *ad.inner = addr;
+ code = -1;
+ }
+ }
+
}
}
if (code < 0)
@@ -2481,9 +2533,9 @@ process_address (int nop, rtx *before, r
#endif
if (code < 0)
{
- /* disp => new_base, case (2) above. */
- lra_emit_move (new_reg, disp);
- *ad.disp = new_reg;
+ /* addr => new_base, case (2) above. */
+ lra_emit_move (new_reg, addr);
+ *ad.inner = new_reg;
}
}
else
@@ -2650,6 +2702,24 @@ emit_inc (enum reg_class new_rclass, rtx
return result;
}
+/* Return true if the current move insn does not need processing as we
+ already know that it satisfies its constraints. */
+static bool
+simple_move_p (void)
+{
+ rtx dest, src;
+ enum reg_class dclass, sclass;
+
+ lra_assert (curr_insn_set != NULL_RTX);
+ dest = SET_DEST (curr_insn_set);
+ src = SET_SRC (curr_insn_set);
+ return ((dclass = get_op_class (dest)) != NO_REGS
+ && (sclass = get_op_class (src)) != NO_REGS
+ /* The backend guarantees that register moves of cost 2
+ never need reloads. */
+ && targetm.register_move_cost (GET_MODE (src), dclass, sclass) == 2);
+ }
+
/* Swap operands NOP and NOP + 1. */
static inline void
swap_operands (int nop)
@@ -2693,12 +2763,13 @@ curr_insn_transform (void)
int max_regno_before;
int reused_alternative_num;
+ curr_insn_set = single_set (curr_insn);
+ if (curr_insn_set != NULL_RTX && simple_move_p ())
+ return false;
+
no_input_reloads_p = no_output_reloads_p = false;
goal_alt_number = -1;
-
- if (check_and_process_move (&change_p, &sec_mem_p))
- return change_p;
-
+ change_p = sec_mem_p = false;
/* JUMP_INSNs and CALL_INSNs are not allowed to have any output
reloads; neither are insns that SET cc0. Insns that use CC0 are
not allowed to have any input reloads. */
@@ -2793,6 +2864,10 @@ curr_insn_transform (void)
we chose previously may no longer be valid. */
lra_set_used_insn_alternative (curr_insn, -1);
+ if (curr_insn_set != NULL_RTX
+ && check_and_process_move (&change_p, &sec_mem_p))
+ return change_p;
+
try_swapped:
reused_alternative_num = curr_id->used_insn_alternative;
@@ -4812,7 +4887,7 @@ inherit_in_ebb (rtx head, rtx tail)
/* This value affects EBB forming. If probability of edge from EBB to
a BB is not greater than the following value, we don't add the BB
to EBB. */
-#define EBB_PROBABILITY_CUTOFF (REG_BR_PROB_BASE / 2)
+#define EBB_PROBABILITY_CUTOFF ((REG_BR_PROB_BASE * 50) / 100)
/* Current number of inheritance/split iteration. */
int lra_inheritance_iter;
More information about the Gcc-patches
mailing list