/* IRA hard register and memory cost calculation for allocnos or pseudos.
- Copyright (C) 2006, 2007, 2008, 2009, 2010, 2011
- Free Software Foundation, Inc.
+ Copyright (C) 2006-2014 Free Software Foundation, Inc.
Contributed by Vladimir Makarov <vmakarov@redhat.com>.
This file is part of GCC.
#include "system.h"
#include "coretypes.h"
#include "tm.h"
+#include "hash-table.h"
#include "hard-reg-set.h"
#include "rtl.h"
#include "expr.h"
#include "tm_p.h"
#include "flags.h"
+#include "predict.h"
+#include "vec.h"
+#include "hashtab.h"
+#include "hash-set.h"
+#include "machmode.h"
+#include "input.h"
+#include "function.h"
+#include "dominance.h"
+#include "cfg.h"
#include "basic-block.h"
#include "regs.h"
#include "addresses.h"
/* Container of the cost classes. */
enum reg_class classes[N_REG_CLASSES];
/* Map reg class -> index of the reg class in the previous array.
- -1 if it is not a cost classe. */
+ -1 if it is not a cost class. */
int index[N_REG_CLASSES];
/* Map hard regno index of first class in array CLASSES containing
the hard regno, -1 otherwise. */
/* Info about cost classes for each pseudo. */
static cost_classes_t *regno_cost_classes;
-/* Returns hash value for cost classes info V. */
-static hashval_t
-cost_classes_hash (const void *v)
+/* Helper for cost_classes hashing. */
+
+struct cost_classes_hasher
{
- const_cost_classes_t hv = (const_cost_classes_t) v;
+ typedef cost_classes value_type;
+ typedef cost_classes compare_type;
+ static inline hashval_t hash (const value_type *);
+ static inline bool equal (const value_type *, const compare_type *);
+ static inline void remove (value_type *);
+};
+/* Returns hash value for cost classes info HV. */
+inline hashval_t
+cost_classes_hasher::hash (const value_type *hv)
+{
return iterative_hash (&hv->classes, sizeof (enum reg_class) * hv->num, 0);
}
-/* Compares cost classes info V1 and V2. */
-static int
-cost_classes_eq (const void *v1, const void *v2)
+/* Compares cost classes info HV1 and HV2. */
+inline bool
+cost_classes_hasher::equal (const value_type *hv1, const compare_type *hv2)
{
- const_cost_classes_t hv1 = (const_cost_classes_t) v1;
- const_cost_classes_t hv2 = (const_cost_classes_t) v2;
-
- return hv1->num == hv2->num && memcmp (hv1->classes, hv2->classes,
- sizeof (enum reg_class) * hv1->num);
+ return (hv1->num == hv2->num
+ && memcmp (hv1->classes, hv2->classes,
+ sizeof (enum reg_class) * hv1->num) == 0);
}
/* Delete cost classes info V from the hash table. */
-static void
-cost_classes_del (void *v)
+inline void
+cost_classes_hasher::remove (value_type *v)
{
ira_free (v);
}
/* Hash table of unique cost classes. */
-static htab_t cost_classes_htab;
+static hash_table<cost_classes_hasher> *cost_classes_htab;
/* Map allocno class -> cost classes for pseudo of given allocno
class. */
/* Map mode -> cost classes for pseudo of give mode. */
static cost_classes_t cost_classes_mode_cache[MAX_MACHINE_MODE];
+/* Cost classes that include all classes in ira_important_classes. */
+static cost_classes all_cost_classes;
+
+/* Use the array of classes in CLASSES_PTR to fill out the rest of
+ the structure. */
+static void
+complete_cost_classes (cost_classes_t classes_ptr)
+{
+ for (int i = 0; i < N_REG_CLASSES; i++)
+ classes_ptr->index[i] = -1;
+ for (int i = 0; i < FIRST_PSEUDO_REGISTER; i++)
+ classes_ptr->hard_regno_index[i] = -1;
+ for (int i = 0; i < classes_ptr->num; i++)
+ {
+ enum reg_class cl = classes_ptr->classes[i];
+ classes_ptr->index[cl] = i;
+ for (int j = ira_class_hard_regs_num[cl] - 1; j >= 0; j--)
+ {
+ unsigned int hard_regno = ira_class_hard_regs[cl][j];
+ if (classes_ptr->hard_regno_index[hard_regno] < 0)
+ classes_ptr->hard_regno_index[hard_regno] = i;
+ }
+ }
+}
+
/* Initialize info about the cost classes for each pseudo. */
static void
initiate_regno_cost_classes (void)
sizeof (cost_classes_t) * N_REG_CLASSES);
memset (cost_classes_mode_cache, 0,
sizeof (cost_classes_t) * MAX_MACHINE_MODE);
- cost_classes_htab
- = htab_create (200, cost_classes_hash, cost_classes_eq, cost_classes_del);
+ cost_classes_htab = new hash_table<cost_classes_hasher> (200);
+ all_cost_classes.num = ira_important_classes_num;
+ for (int i = 0; i < ira_important_classes_num; i++)
+ all_cost_classes.classes[i] = ira_important_classes[i];
+ complete_cost_classes (&all_cost_classes);
}
/* Create new cost classes from cost classes FROM and set up members
setup_cost_classes (cost_classes_t from)
{
cost_classes_t classes_ptr;
- enum reg_class cl;
- int i, j, hard_regno;
classes_ptr = (cost_classes_t) ira_allocate (sizeof (struct cost_classes));
classes_ptr->num = from->num;
- for (i = 0; i < N_REG_CLASSES; i++)
- classes_ptr->index[i] = -1;
- for (i = 0; i < FIRST_PSEUDO_REGISTER; i++)
- classes_ptr->hard_regno_index[i] = -1;
- for (i = 0; i < from->num; i++)
+ for (int i = 0; i < from->num; i++)
+ classes_ptr->classes[i] = from->classes[i];
+ complete_cost_classes (classes_ptr);
+ return classes_ptr;
+}
+
+/* Return a version of FULL that only considers registers in REGS that are
+ valid for mode MODE. Both FULL and the returned class are globally
+ allocated. */
+static cost_classes_t
+restrict_cost_classes (cost_classes_t full, machine_mode mode,
+ const HARD_REG_SET ®s)
+{
+ static struct cost_classes narrow;
+ int map[N_REG_CLASSES];
+ narrow.num = 0;
+ for (int i = 0; i < full->num; i++)
{
- cl = classes_ptr->classes[i] = from->classes[i];
- classes_ptr->index[cl] = i;
- for (j = ira_class_hard_regs_num[cl] - 1; j >= 0; j--)
+ /* Assume that we'll drop the class. */
+ map[i] = -1;
+
+ /* Ignore classes that are too small for the mode. */
+ enum reg_class cl = full->classes[i];
+ if (!contains_reg_of_mode[cl][mode])
+ continue;
+
+ /* Calculate the set of registers in CL that belong to REGS and
+ are valid for MODE. */
+ HARD_REG_SET valid_for_cl;
+ COPY_HARD_REG_SET (valid_for_cl, reg_class_contents[cl]);
+ AND_HARD_REG_SET (valid_for_cl, regs);
+ AND_COMPL_HARD_REG_SET (valid_for_cl,
+ ira_prohibited_class_mode_regs[cl][mode]);
+ AND_COMPL_HARD_REG_SET (valid_for_cl, ira_no_alloc_regs);
+ if (hard_reg_set_empty_p (valid_for_cl))
+ continue;
+
+ /* Don't use this class if the set of valid registers is a subset
+ of an existing class. For example, suppose we have two classes
+ GR_REGS and FR_REGS and a union class GR_AND_FR_REGS. Suppose
+ that the mode changes allowed by FR_REGS are not as general as
+ the mode changes allowed by GR_REGS.
+
+ In this situation, the mode changes for GR_AND_FR_REGS could
+ either be seen as the union or the intersection of the mode
+ changes allowed by the two subclasses. The justification for
+ the union-based definition would be that, if you want a mode
+ change that's only allowed by GR_REGS, you can pick a register
+ from the GR_REGS subclass. The justification for the
+ intersection-based definition would be that every register
+ from the class would allow the mode change.
+
+ However, if we have a register that needs to be in GR_REGS,
+ using GR_AND_FR_REGS with the intersection-based definition
+ would be too pessimistic, since it would bring in restrictions
+ that only apply to FR_REGS. Conversely, if we have a register
+ that needs to be in FR_REGS, using GR_AND_FR_REGS with the
+ union-based definition would lose the extra restrictions
+ placed on FR_REGS. GR_AND_FR_REGS is therefore only useful
+ for cases where GR_REGS and FP_REGS are both valid. */
+ int pos;
+ for (pos = 0; pos < narrow.num; ++pos)
{
- hard_regno = ira_class_hard_regs[cl][j];
- if (classes_ptr->hard_regno_index[hard_regno] < 0)
- classes_ptr->hard_regno_index[hard_regno] = i;
+ enum reg_class cl2 = narrow.classes[pos];
+ if (hard_reg_set_subset_p (valid_for_cl, reg_class_contents[cl2]))
+ break;
+ }
+ map[i] = pos;
+ if (pos == narrow.num)
+ {
+ /* If several classes are equivalent, prefer to use the one
+ that was chosen as the allocno class. */
+ enum reg_class cl2 = ira_allocno_class_translate[cl];
+ if (ira_class_hard_regs_num[cl] == ira_class_hard_regs_num[cl2])
+ cl = cl2;
+ narrow.classes[narrow.num++] = cl;
}
}
- return classes_ptr;
+ if (narrow.num == full->num)
+ return full;
+
+ cost_classes **slot = cost_classes_htab->find_slot (&narrow, INSERT);
+ if (*slot == NULL)
+ {
+ cost_classes_t classes = setup_cost_classes (&narrow);
+ /* Map equivalent classes to the representative that we chose above. */
+ for (int i = 0; i < ira_important_classes_num; i++)
+ {
+ enum reg_class cl = ira_important_classes[i];
+ int index = full->index[cl];
+ if (index >= 0)
+ classes->index[cl] = map[index];
+ }
+ *slot = classes;
+ }
+ return *slot;
}
/* Setup cost classes for pseudo REGNO whose allocno class is ACLASS.
cost_classes_t classes_ptr;
enum reg_class cl;
int i;
- PTR *slot;
+ cost_classes **slot;
HARD_REG_SET temp, temp2;
bool exclude_p;
COPY_HARD_REG_SET (temp, reg_class_contents[aclass]);
AND_COMPL_HARD_REG_SET (temp, ira_no_alloc_regs);
/* We exclude classes from consideration which are subsets of
- ACLASS only if ACLASS is a pressure class or subset of a
- pressure class. It means by the definition of pressure classes
- that cost of moving between susbets of ACLASS is cheaper than
- load or store. */
- for (i = 0; i < ira_pressure_classes_num; i++)
- {
- cl = ira_pressure_classes[i];
- if (cl == aclass)
- break;
- COPY_HARD_REG_SET (temp2, reg_class_contents[cl]);
- AND_COMPL_HARD_REG_SET (temp2, ira_no_alloc_regs);
- if (hard_reg_set_subset_p (temp, temp2))
- break;
- }
- exclude_p = i < ira_pressure_classes_num;
+ ACLASS only if ACLASS is an uniform class. */
+ exclude_p = ira_uniform_class_p[aclass];
classes.num = 0;
for (i = 0; i < ira_important_classes_num; i++)
{
cl = ira_important_classes[i];
if (exclude_p)
{
- /* Exclude no-pressure classes which are subsets of
+ /* Exclude non-uniform classes which are subsets of
ACLASS. */
COPY_HARD_REG_SET (temp2, reg_class_contents[cl]);
AND_COMPL_HARD_REG_SET (temp2, ira_no_alloc_regs);
- if (! ira_reg_pressure_class_p[cl]
- && hard_reg_set_subset_p (temp2, temp) && cl != aclass)
+ if (hard_reg_set_subset_p (temp2, temp) && cl != aclass)
continue;
}
classes.classes[classes.num++] = cl;
}
- slot = htab_find_slot (cost_classes_htab, &classes, INSERT);
+ slot = cost_classes_htab->find_slot (&classes, INSERT);
if (*slot == NULL)
{
classes_ptr = setup_cost_classes (&classes);
}
classes_ptr = cost_classes_aclass_cache[aclass] = (cost_classes_t) *slot;
}
+ if (regno_reg_rtx[regno] != NULL_RTX)
+ {
+ /* Restrict the classes to those that are valid for REGNO's mode
+ (which might for example exclude singleton classes if the mode
+ requires two registers). Also restrict the classes to those that
+ are valid for subregs of REGNO. */
+ const HARD_REG_SET *valid_regs = valid_mode_changes_for_regno (regno);
+ if (!valid_regs)
+ valid_regs = ®_class_contents[ALL_REGS];
+ classes_ptr = restrict_cost_classes (classes_ptr,
+ PSEUDO_REGNO_MODE (regno),
+ *valid_regs);
+ }
regno_cost_classes[regno] = classes_ptr;
}
decrease number of cost classes for the pseudo, if hard registers
of some important classes can not hold a value of MODE. So the
pseudo can not get hard register of some important classes and cost
- calculation for such important classes is only waisting CPU
+ calculation for such important classes is only wasting CPU
time. */
static void
-setup_regno_cost_classes_by_mode (int regno, enum machine_mode mode)
+setup_regno_cost_classes_by_mode (int regno, machine_mode mode)
{
- static struct cost_classes classes;
- cost_classes_t classes_ptr;
- enum reg_class cl;
- int i;
- PTR *slot;
- HARD_REG_SET temp;
-
- if ((classes_ptr = cost_classes_mode_cache[mode]) == NULL)
+ if (const HARD_REG_SET *valid_regs = valid_mode_changes_for_regno (regno))
+ regno_cost_classes[regno] = restrict_cost_classes (&all_cost_classes,
+ mode, *valid_regs);
+ else
{
- classes.num = 0;
- for (i = 0; i < ira_important_classes_num; i++)
- {
- cl = ira_important_classes[i];
- COPY_HARD_REG_SET (temp, ira_prohibited_class_mode_regs[cl][mode]);
- IOR_HARD_REG_SET (temp, ira_no_alloc_regs);
- if (hard_reg_set_subset_p (reg_class_contents[cl], temp))
- continue;
- classes.classes[classes.num++] = cl;
- }
- slot = htab_find_slot (cost_classes_htab, &classes, INSERT);
- if (*slot == NULL)
- {
- classes_ptr = setup_cost_classes (&classes);
- *slot = classes_ptr;
- }
- else
- classes_ptr = (cost_classes_t) *slot;
- cost_classes_mode_cache[mode] = (cost_classes_t) *slot;
+ if (cost_classes_mode_cache[mode] == NULL)
+ cost_classes_mode_cache[mode]
+ = restrict_cost_classes (&all_cost_classes, mode,
+ reg_class_contents[ALL_REGS]);
+ regno_cost_classes[regno] = cost_classes_mode_cache[mode];
}
- regno_cost_classes[regno] = classes_ptr;
}
-/* Finilize info about the cost classes for each pseudo. */
+/* Finalize info about the cost classes for each pseudo. */
static void
finish_regno_cost_classes (void)
{
ira_free (regno_cost_classes);
- htab_delete (cost_classes_htab);
+ delete cost_classes_htab;
+ cost_classes_htab = NULL;
}
\f
TO_P is FALSE) a register of class RCLASS in mode MODE. X must not
be a pseudo register. */
static int
-copy_cost (rtx x, enum machine_mode mode, reg_class_t rclass, bool to_p,
+copy_cost (rtx x, machine_mode mode, reg_class_t rclass, bool to_p,
secondary_reload_info *prev_sri)
{
secondary_reload_info sri;
if (secondary_class != NO_REGS)
{
- if (!move_cost[mode])
- init_move_cost (mode);
- return (move_cost[mode][(int) secondary_class][(int) rclass]
+ ira_init_register_move_cost_if_necessary (mode);
+ return (ira_register_move_cost[mode][(int) secondary_class][(int) rclass]
+ sri.extra_cost
+ copy_cost (x, mode, secondary_class, to_p, &sri));
}
+ ira_memory_move_cost[mode][(int) rclass][to_p != 0];
else if (REG_P (x))
{
- if (!move_cost[mode])
- init_move_cost (mode);
+ reg_class_t x_class = REGNO_REG_CLASS (REGNO (x));
+
+ ira_init_register_move_cost_if_necessary (mode);
return (sri.extra_cost
- + move_cost[mode][REGNO_REG_CLASS (REGNO (x))][(int) rclass]);
+ + ira_register_move_cost[mode][(int) x_class][(int) rclass]);
}
else
/* If this is a constant, we may eventually want to call rtx_cost
the alternatives. */
static void
record_reg_classes (int n_alts, int n_ops, rtx *ops,
- enum machine_mode *modes, const char **constraints,
- rtx insn, enum reg_class *pref)
+ machine_mode *modes, const char **constraints,
+ rtx_insn *insn, enum reg_class *pref)
{
int alt;
int i, j, k;
- rtx set;
int insn_allows_mem[MAX_RECOG_OPERANDS];
+ move_table *move_in_cost, *move_out_cost;
+ short (*mem_cost)[2];
for (i = 0; i < n_ops; i++)
insn_allows_mem[i] = 0;
/* Process each alternative, each time minimizing an operand's cost
with the cost for each operand in that alternative. */
+ alternative_mask preferred = get_preferred_alternatives (insn);
for (alt = 0; alt < n_alts; alt++)
{
enum reg_class classes[MAX_RECOG_OPERANDS];
int alt_fail = 0;
int alt_cost = 0, op_cost_add;
- if (!recog_data.alternative_enabled_p[alt])
+ if (!TEST_BIT (preferred, alt))
{
for (i = 0; i < recog_data.n_operands; i++)
constraints[i] = skip_alternative (constraints[i]);
unsigned char c;
const char *p = constraints[i];
rtx op = ops[i];
- enum machine_mode mode = modes[i];
+ machine_mode mode = modes[i];
int allows_addr = 0;
int win = 0;
bool in_p = recog_data.operand_type[i] != OP_OUT;
bool out_p = recog_data.operand_type[i] != OP_IN;
enum reg_class op_class = classes[i];
- move_table *move_in_cost, *move_out_cost;
ira_init_register_move_cost_if_necessary (mode);
if (! in_p)
{
ira_assert (out_p);
- move_out_cost = ira_may_move_out_cost[mode];
- for (k = cost_classes_ptr->num - 1; k >= 0; k--)
+ if (op_class == NO_REGS)
{
- rclass = cost_classes[k];
- pp_costs[k]
- = move_out_cost[op_class][rclass] * frequency;
+ mem_cost = ira_memory_move_cost[mode];
+ for (k = cost_classes_ptr->num - 1; k >= 0; k--)
+ {
+ rclass = cost_classes[k];
+ pp_costs[k] = mem_cost[rclass][0] * frequency;
+ }
+ }
+ else
+ {
+ move_out_cost = ira_may_move_out_cost[mode];
+ for (k = cost_classes_ptr->num - 1; k >= 0; k--)
+ {
+ rclass = cost_classes[k];
+ pp_costs[k]
+ = move_out_cost[op_class][rclass] * frequency;
+ }
}
}
else if (! out_p)
{
ira_assert (in_p);
- move_in_cost = ira_may_move_in_cost[mode];
- for (k = cost_classes_ptr->num - 1; k >= 0; k--)
+ if (op_class == NO_REGS)
{
- rclass = cost_classes[k];
- pp_costs[k]
- = move_in_cost[rclass][op_class] * frequency;
+ mem_cost = ira_memory_move_cost[mode];
+ for (k = cost_classes_ptr->num - 1; k >= 0; k--)
+ {
+ rclass = cost_classes[k];
+ pp_costs[k] = mem_cost[rclass][1] * frequency;
+ }
+ }
+ else
+ {
+ move_in_cost = ira_may_move_in_cost[mode];
+ for (k = cost_classes_ptr->num - 1; k >= 0; k--)
+ {
+ rclass = cost_classes[k];
+ pp_costs[k]
+ = move_in_cost[rclass][op_class] * frequency;
+ }
}
}
else
{
- move_in_cost = ira_may_move_in_cost[mode];
- move_out_cost = ira_may_move_out_cost[mode];
- for (k = cost_classes_ptr->num - 1; k >= 0; k--)
+ if (op_class == NO_REGS)
{
- rclass = cost_classes[k];
- pp_costs[k] = ((move_in_cost[rclass][op_class]
- + move_out_cost[op_class][rclass])
- * frequency);
+ mem_cost = ira_memory_move_cost[mode];
+ for (k = cost_classes_ptr->num - 1; k >= 0; k--)
+ {
+ rclass = cost_classes[k];
+ pp_costs[k] = ((mem_cost[rclass][0]
+ + mem_cost[rclass][1])
+ * frequency);
+ }
+ }
+ else
+ {
+ move_in_cost = ira_may_move_in_cost[mode];
+ move_out_cost = ira_may_move_out_cost[mode];
+ for (k = cost_classes_ptr->num - 1; k >= 0; k--)
+ {
+ rclass = cost_classes[k];
+ pp_costs[k] = ((move_in_cost[rclass][op_class]
+ + move_out_cost[op_class][rclass])
+ * frequency);
+ }
}
}
{
switch (c)
{
- case ',':
- break;
case '*':
/* Ignore the next letter for this pass. */
c = *++p;
case '?':
alt_cost += 2;
- case '!': case '#': case '&':
- case '0': case '1': case '2': case '3': case '4':
- case '5': case '6': case '7': case '8': case '9':
- break;
-
- case 'p':
- allows_addr = 1;
- win = address_operand (op, GET_MODE (op));
- /* We know this operand is an address, so we want it
- to be allocated to a register that can be the
- base of an address, i.e. BASE_REG_CLASS. */
- classes[i]
- = ira_reg_class_subunion[classes[i]]
- [base_reg_class (VOIDmode, ADDRESS, SCRATCH)];
- break;
-
- case 'm': case 'o': case 'V':
- /* It doesn't seem worth distinguishing between
- offsettable and non-offsettable addresses
- here. */
- insn_allows_mem[i] = allows_mem[i] = 1;
- if (MEM_P (op))
- win = 1;
- break;
-
- case '<':
- if (MEM_P (op)
- && (GET_CODE (XEXP (op, 0)) == PRE_DEC
- || GET_CODE (XEXP (op, 0)) == POST_DEC))
- win = 1;
- break;
-
- case '>':
- if (MEM_P (op)
- && (GET_CODE (XEXP (op, 0)) == PRE_INC
- || GET_CODE (XEXP (op, 0)) == POST_INC))
- win = 1;
- break;
-
- case 'E':
- case 'F':
- if (GET_CODE (op) == CONST_DOUBLE
- || (GET_CODE (op) == CONST_VECTOR
- && (GET_MODE_CLASS (GET_MODE (op))
- == MODE_VECTOR_FLOAT)))
- win = 1;
- break;
-
- case 'G':
- case 'H':
- if (GET_CODE (op) == CONST_DOUBLE
- && CONST_DOUBLE_OK_FOR_CONSTRAINT_P (op, c, p))
- win = 1;
- break;
-
- case 's':
- if (CONST_INT_P (op)
- || (GET_CODE (op) == CONST_DOUBLE
- && GET_MODE (op) == VOIDmode))
- break;
-
- case 'i':
- if (CONSTANT_P (op)
- && (! flag_pic || LEGITIMATE_PIC_OPERAND_P (op)))
- win = 1;
- break;
-
- case 'n':
- if (CONST_INT_P (op)
- || (GET_CODE (op) == CONST_DOUBLE
- && GET_MODE (op) == VOIDmode))
- win = 1;
- break;
-
- case 'I':
- case 'J':
- case 'K':
- case 'L':
- case 'M':
- case 'N':
- case 'O':
- case 'P':
- if (CONST_INT_P (op)
- && CONST_OK_FOR_CONSTRAINT_P (INTVAL (op), c, p))
- win = 1;
- break;
-
- case 'X':
- win = 1;
break;
case 'g':
&& (! flag_pic || LEGITIMATE_PIC_OPERAND_P (op))))
win = 1;
insn_allows_mem[i] = allows_mem[i] = 1;
- case 'r':
classes[i] = ira_reg_class_subunion[classes[i]][GENERAL_REGS];
break;
default:
- if (REG_CLASS_FROM_CONSTRAINT (c, p) != NO_REGS)
- classes[i] = ira_reg_class_subunion[classes[i]]
- [REG_CLASS_FROM_CONSTRAINT (c, p)];
-#ifdef EXTRA_CONSTRAINT_STR
- else if (EXTRA_CONSTRAINT_STR (op, c, p))
- win = 1;
-
- if (EXTRA_MEMORY_CONSTRAINT (c, p))
+ enum constraint_num cn = lookup_constraint (p);
+ enum reg_class cl;
+ switch (get_constraint_type (cn))
{
+ case CT_REGISTER:
+ cl = reg_class_for_constraint (cn);
+ if (cl != NO_REGS)
+ classes[i] = ira_reg_class_subunion[classes[i]][cl];
+ break;
+
+ case CT_CONST_INT:
+ if (CONST_INT_P (op)
+ && insn_const_int_ok_for_constraint (INTVAL (op), cn))
+ win = 1;
+ break;
+
+ case CT_MEMORY:
/* Every MEM can be reloaded to fit. */
insn_allows_mem[i] = allows_mem[i] = 1;
if (MEM_P (op))
win = 1;
- }
- if (EXTRA_ADDRESS_CONSTRAINT (c, p))
- {
+ break;
+
+ case CT_ADDRESS:
/* Every address can be reloaded to fit. */
allows_addr = 1;
- if (address_operand (op, GET_MODE (op)))
+ if (address_operand (op, GET_MODE (op))
+ || constraint_satisfied_p (op, cn))
win = 1;
/* We know this operand is an address, so we
want it to be allocated to a hard register
i.e. BASE_REG_CLASS. */
classes[i]
= ira_reg_class_subunion[classes[i]]
- [base_reg_class (VOIDmode, ADDRESS, SCRATCH)];
+ [base_reg_class (VOIDmode, ADDR_SPACE_GENERIC,
+ ADDRESS, SCRATCH)];
+ break;
+
+ case CT_FIXED_FORM:
+ if (constraint_satisfied_p (op, cn))
+ win = 1;
+ break;
}
-#endif
break;
}
p += CONSTRAINT_LEN (c, p);
into that class. */
if (REG_P (op) && REGNO (op) >= FIRST_PSEUDO_REGISTER)
{
- if (classes[i] == NO_REGS)
+ if (classes[i] == NO_REGS && ! allows_mem[i])
{
/* We must always fail if the operand is a REG, but
- we did not find a suitable class.
+ we did not find a suitable class and memory is
+ not allowed.
Otherwise we may perform an uninitialized read
from this_op_costs after the `continue' statement
bool in_p = recog_data.operand_type[i] != OP_OUT;
bool out_p = recog_data.operand_type[i] != OP_IN;
enum reg_class op_class = classes[i];
- move_table *move_in_cost, *move_out_cost;
ira_init_register_move_cost_if_necessary (mode);
if (! in_p)
{
ira_assert (out_p);
- move_out_cost = ira_may_move_out_cost[mode];
- for (k = cost_classes_ptr->num - 1; k >= 0; k--)
+ if (op_class == NO_REGS)
+ {
+ mem_cost = ira_memory_move_cost[mode];
+ for (k = cost_classes_ptr->num - 1; k >= 0; k--)
+ {
+ rclass = cost_classes[k];
+ pp_costs[k] = mem_cost[rclass][0] * frequency;
+ }
+ }
+ else
{
- rclass = cost_classes[k];
- pp_costs[k]
- = move_out_cost[op_class][rclass] * frequency;
+ move_out_cost = ira_may_move_out_cost[mode];
+ for (k = cost_classes_ptr->num - 1; k >= 0; k--)
+ {
+ rclass = cost_classes[k];
+ pp_costs[k]
+ = move_out_cost[op_class][rclass] * frequency;
+ }
}
}
else if (! out_p)
{
ira_assert (in_p);
- move_in_cost = ira_may_move_in_cost[mode];
- for (k = cost_classes_ptr->num - 1; k >= 0; k--)
+ if (op_class == NO_REGS)
{
- rclass = cost_classes[k];
- pp_costs[k]
- = move_in_cost[rclass][op_class] * frequency;
+ mem_cost = ira_memory_move_cost[mode];
+ for (k = cost_classes_ptr->num - 1; k >= 0; k--)
+ {
+ rclass = cost_classes[k];
+ pp_costs[k] = mem_cost[rclass][1] * frequency;
+ }
+ }
+ else
+ {
+ move_in_cost = ira_may_move_in_cost[mode];
+ for (k = cost_classes_ptr->num - 1; k >= 0; k--)
+ {
+ rclass = cost_classes[k];
+ pp_costs[k]
+ = move_in_cost[rclass][op_class] * frequency;
+ }
}
}
else
{
- move_in_cost = ira_may_move_in_cost[mode];
- move_out_cost = ira_may_move_out_cost[mode];
- for (k = cost_classes_ptr->num - 1; k >= 0; k--)
+ if (op_class == NO_REGS)
{
- rclass = cost_classes[k];
- pp_costs[k] = ((move_in_cost[rclass][op_class]
- + move_out_cost[op_class][rclass])
- * frequency);
+ mem_cost = ira_memory_move_cost[mode];
+ for (k = cost_classes_ptr->num - 1; k >= 0; k--)
+ {
+ rclass = cost_classes[k];
+ pp_costs[k] = ((mem_cost[rclass][0]
+ + mem_cost[rclass][1])
+ * frequency);
+ }
+ }
+ else
+ {
+ move_in_cost = ira_may_move_in_cost[mode];
+ move_out_cost = ira_may_move_out_cost[mode];
+ for (k = cost_classes_ptr->num - 1; k >= 0; k--)
+ {
+ rclass = cost_classes[k];
+ pp_costs[k] = ((move_in_cost[rclass][op_class]
+ + move_out_cost[op_class][rclass])
+ * frequency);
+ }
}
}
- /* If the alternative actually allows memory, make
- things a bit cheaper since we won't need an extra
- insn to load it. */
- pp->mem_cost
- = ((out_p ? ira_memory_move_cost[mode][op_class][0] : 0)
- + (in_p ? ira_memory_move_cost[mode][op_class][1] : 0)
- - allows_mem[i]) * frequency;
+ if (op_class == NO_REGS)
+ /* Although we don't need insn to reload from
+ memory, still accessing memory is usually more
+ expensive than a register. */
+ pp->mem_cost = frequency;
+ else
+ /* If the alternative actually allows memory, make
+ things a bit cheaper since we won't need an
+ extra insn to load it. */
+ pp->mem_cost
+ = ((out_p ? ira_memory_move_cost[mode][op_class][0] : 0)
+ + (in_p ? ira_memory_move_cost[mode][op_class][1] : 0)
+ - allows_mem[i]) * frequency;
/* If we have assigned a class to this allocno in
our first pass, add a cost to this alternative
corresponding to what we would add if this
enum reg_class pref_class = pref[COST_INDEX (REGNO (op))];
if (pref_class == NO_REGS)
+ {
+ if (op_class != NO_REGS)
+ alt_cost
+ += ((out_p
+ ? ira_memory_move_cost[mode][op_class][0]
+ : 0)
+ + (in_p
+ ? ira_memory_move_cost[mode][op_class][1]
+ : 0));
+ }
+ else if (op_class == NO_REGS)
alt_cost
+= ((out_p
- ? ira_memory_move_cost[mode][op_class][0] : 0)
+ ? ira_memory_move_cost[mode][pref_class][1]
+ : 0)
+ (in_p
- ? ira_memory_move_cost[mode][op_class][1]
+ ? ira_memory_move_cost[mode][pref_class][0]
: 0));
else if (ira_reg_class_intersect[pref_class][op_class]
== NO_REGS)
- alt_cost += ira_register_move_cost[mode][pref_class][op_class];
+ alt_cost += (ira_register_move_cost
+ [mode][pref_class][op_class]);
}
}
}
ALLOCNO_BAD_SPILL_P (a) = true;
}
- /* If this insn is a single set copying operand 1 to operand 0 and
- one operand is an allocno with the other a hard reg or an allocno
- that prefers a hard register that is in its own register class
- then we may want to adjust the cost of that register class to -1.
-
- Avoid the adjustment if the source does not die to avoid
- stressing of register allocator by preferrencing two colliding
- registers into single class.
-
- Also avoid the adjustment if a copy between hard registers of the
- class is expensive (ten times the cost of a default copy is
- considered arbitrarily expensive). This avoids losing when the
- preferred class is very expensive as the source of a copy
- instruction. */
- if ((set = single_set (insn)) != 0
- && ops[0] == SET_DEST (set) && ops[1] == SET_SRC (set)
- && REG_P (ops[0]) && REG_P (ops[1])
- && find_regno_note (insn, REG_DEAD, REGNO (ops[1])))
- for (i = 0; i <= 1; i++)
- if (REGNO (ops[i]) >= FIRST_PSEUDO_REGISTER
- && REGNO (ops[!i]) < FIRST_PSEUDO_REGISTER)
- {
- unsigned int regno = REGNO (ops[i]);
- unsigned int other_regno = REGNO (ops[!i]);
- enum machine_mode mode = GET_MODE (ops[!i]);
- cost_classes_t cost_classes_ptr = regno_cost_classes[regno];
- enum reg_class *cost_classes = cost_classes_ptr->classes;
- reg_class_t rclass;
- int nr;
-
- for (k = cost_classes_ptr->num - 1; k >= 0; k--)
- {
- rclass = cost_classes[k];
- if (TEST_HARD_REG_BIT (reg_class_contents[rclass], other_regno)
- && (reg_class_size[(int) rclass]
- == ira_reg_class_max_nregs [(int) rclass][(int) mode]))
- {
- if (reg_class_size[rclass] == 1)
- op_costs[i]->cost[k] = -frequency;
- else
- {
- for (nr = 0;
- nr < hard_regno_nregs[other_regno][mode];
- nr++)
- if (! TEST_HARD_REG_BIT (reg_class_contents[rclass],
- other_regno + nr))
- break;
-
- if (nr == hard_regno_nregs[other_regno][mode])
- op_costs[i]->cost[k] = -frequency;
- }
- }
- }
- }
}
\f
pseudo-registers should count as OK. Arguments as for
regno_ok_for_base_p. */
static inline bool
-ok_for_base_p_nonstrict (rtx reg, enum machine_mode mode,
+ok_for_base_p_nonstrict (rtx reg, machine_mode mode, addr_space_t as,
enum rtx_code outer_code, enum rtx_code index_code)
{
unsigned regno = REGNO (reg);
if (regno >= FIRST_PSEUDO_REGISTER)
return true;
- return ok_for_base_p_1 (regno, mode, outer_code, index_code);
+ return ok_for_base_p_1 (regno, mode, as, outer_code, index_code);
}
/* Record the pseudo registers we must reload into hard registers in a
If CONTEXT is 0, we are looking at the base part of an address,
otherwise we are looking at the index part.
- MODE is the mode of the memory reference; OUTER_CODE and INDEX_CODE
- give the context that the rtx appears in. These three arguments
- are passed down to base_reg_class.
+ MODE and AS are the mode and address space of the memory reference;
+ OUTER_CODE and INDEX_CODE give the context that the rtx appears in.
+ These four arguments are passed down to base_reg_class.
SCALE is twice the amount to multiply the cost by (it is twice so
we can represent half-cost adjustments). */
static void
-record_address_regs (enum machine_mode mode, rtx x, int context,
- enum rtx_code outer_code, enum rtx_code index_code,
- int scale)
+record_address_regs (machine_mode mode, addr_space_t as, rtx x,
+ int context, enum rtx_code outer_code,
+ enum rtx_code index_code, int scale)
{
enum rtx_code code = GET_CODE (x);
enum reg_class rclass;
if (context == 1)
rclass = INDEX_REG_CLASS;
else
- rclass = base_reg_class (mode, outer_code, index_code);
+ rclass = base_reg_class (mode, as, outer_code, index_code);
switch (code)
{
/* If this machine only allows one register per address, it
must be in the first operand. */
if (MAX_REGS_PER_ADDRESS == 1)
- record_address_regs (mode, arg0, 0, PLUS, code1, scale);
+ record_address_regs (mode, as, arg0, 0, PLUS, code1, scale);
/* If index and base registers are the same on this machine,
just record registers in any non-constant operands. We
assume here, as well as in the tests below, that all
addresses are in canonical form. */
- else if (INDEX_REG_CLASS == base_reg_class (VOIDmode, PLUS, SCRATCH))
+ else if (INDEX_REG_CLASS
+ == base_reg_class (VOIDmode, as, PLUS, SCRATCH))
{
- record_address_regs (mode, arg0, context, PLUS, code1, scale);
+ record_address_regs (mode, as, arg0, context, PLUS, code1, scale);
if (! CONSTANT_P (arg1))
- record_address_regs (mode, arg1, context, PLUS, code0, scale);
+ record_address_regs (mode, as, arg1, context, PLUS, code0, scale);
}
/* If the second operand is a constant integer, it doesn't
change what class the first operand must be. */
- else if (code1 == CONST_INT || code1 == CONST_DOUBLE)
- record_address_regs (mode, arg0, context, PLUS, code1, scale);
+ else if (CONST_SCALAR_INT_P (arg1))
+ record_address_regs (mode, as, arg0, context, PLUS, code1, scale);
/* If the second operand is a symbolic constant, the first
operand must be an index register. */
else if (code1 == SYMBOL_REF || code1 == CONST || code1 == LABEL_REF)
- record_address_regs (mode, arg0, 1, PLUS, code1, scale);
+ record_address_regs (mode, as, arg0, 1, PLUS, code1, scale);
/* If both operands are registers but one is already a hard
register of index or reg-base class, give the other the
class that the hard register is not. */
else if (code0 == REG && code1 == REG
&& REGNO (arg0) < FIRST_PSEUDO_REGISTER
- && (ok_for_base_p_nonstrict (arg0, mode, PLUS, REG)
+ && (ok_for_base_p_nonstrict (arg0, mode, as, PLUS, REG)
|| ok_for_index_p_nonstrict (arg0)))
- record_address_regs (mode, arg1,
- ok_for_base_p_nonstrict (arg0, mode, PLUS, REG)
- ? 1 : 0,
+ record_address_regs (mode, as, arg1,
+ ok_for_base_p_nonstrict (arg0, mode, as,
+ PLUS, REG) ? 1 : 0,
PLUS, REG, scale);
else if (code0 == REG && code1 == REG
&& REGNO (arg1) < FIRST_PSEUDO_REGISTER
- && (ok_for_base_p_nonstrict (arg1, mode, PLUS, REG)
+ && (ok_for_base_p_nonstrict (arg1, mode, as, PLUS, REG)
|| ok_for_index_p_nonstrict (arg1)))
- record_address_regs (mode, arg0,
- ok_for_base_p_nonstrict (arg1, mode, PLUS, REG)
- ? 1 : 0,
+ record_address_regs (mode, as, arg0,
+ ok_for_base_p_nonstrict (arg1, mode, as,
+ PLUS, REG) ? 1 : 0,
PLUS, REG, scale);
/* If one operand is known to be a pointer, it must be the
base with the other operand the index. Likewise if the
other operand is a MULT. */
else if ((code0 == REG && REG_POINTER (arg0)) || code1 == MULT)
{
- record_address_regs (mode, arg0, 0, PLUS, code1, scale);
- record_address_regs (mode, arg1, 1, PLUS, code0, scale);
+ record_address_regs (mode, as, arg0, 0, PLUS, code1, scale);
+ record_address_regs (mode, as, arg1, 1, PLUS, code0, scale);
}
else if ((code1 == REG && REG_POINTER (arg1)) || code0 == MULT)
{
- record_address_regs (mode, arg0, 1, PLUS, code1, scale);
- record_address_regs (mode, arg1, 0, PLUS, code0, scale);
+ record_address_regs (mode, as, arg0, 1, PLUS, code1, scale);
+ record_address_regs (mode, as, arg1, 0, PLUS, code0, scale);
}
/* Otherwise, count equal chances that each might be a base or
index register. This case should be rare. */
else
{
- record_address_regs (mode, arg0, 0, PLUS, code1, scale / 2);
- record_address_regs (mode, arg0, 1, PLUS, code1, scale / 2);
- record_address_regs (mode, arg1, 0, PLUS, code0, scale / 2);
- record_address_regs (mode, arg1, 1, PLUS, code0, scale / 2);
+ record_address_regs (mode, as, arg0, 0, PLUS, code1, scale / 2);
+ record_address_regs (mode, as, arg0, 1, PLUS, code1, scale / 2);
+ record_address_regs (mode, as, arg1, 0, PLUS, code0, scale / 2);
+ record_address_regs (mode, as, arg1, 1, PLUS, code0, scale / 2);
}
}
break;
up in the wrong place. */
case POST_MODIFY:
case PRE_MODIFY:
- record_address_regs (mode, XEXP (x, 0), 0, code,
+ record_address_regs (mode, as, XEXP (x, 0), 0, code,
GET_CODE (XEXP (XEXP (x, 1), 1)), 2 * scale);
if (REG_P (XEXP (XEXP (x, 1), 1)))
- record_address_regs (mode, XEXP (XEXP (x, 1), 1), 1, code, REG,
+ record_address_regs (mode, as, XEXP (XEXP (x, 1), 1), 1, code, REG,
2 * scale);
break;
/* Double the importance of an allocno that is incremented or
decremented, since it would take two extra insns if it ends
up in the wrong place. */
- record_address_regs (mode, XEXP (x, 0), 0, code, SCRATCH, 2 * scale);
+ record_address_regs (mode, as, XEXP (x, 0), 0, code, SCRATCH, 2 * scale);
break;
case REG:
int i;
for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
if (fmt[i] == 'e')
- record_address_regs (mode, XEXP (x, i), context, code, SCRATCH,
+ record_address_regs (mode, as, XEXP (x, i), context, code, SCRATCH,
scale);
}
}
/* Calculate the costs of insn operands. */
static void
-record_operand_costs (rtx insn, enum reg_class *pref)
+record_operand_costs (rtx_insn *insn, enum reg_class *pref)
{
const char *constraints[MAX_RECOG_OPERANDS];
- enum machine_mode modes[MAX_RECOG_OPERANDS];
+ machine_mode modes[MAX_RECOG_OPERANDS];
+ rtx ops[MAX_RECOG_OPERANDS];
+ rtx set;
int i;
for (i = 0; i < recog_data.n_operands; i++)
{
memcpy (op_costs[i], init_cost, struct_costs_size);
+ ops[i] = recog_data.operand[i];
if (GET_CODE (recog_data.operand[i]) == SUBREG)
recog_data.operand[i] = SUBREG_REG (recog_data.operand[i]);
if (MEM_P (recog_data.operand[i]))
record_address_regs (GET_MODE (recog_data.operand[i]),
+ MEM_ADDR_SPACE (recog_data.operand[i]),
XEXP (recog_data.operand[i], 0),
0, MEM, SCRATCH, frequency * 2);
else if (constraints[i][0] == 'p'
- || EXTRA_ADDRESS_CONSTRAINT (constraints[i][0],
- constraints[i]))
- record_address_regs (VOIDmode, recog_data.operand[i], 0, ADDRESS,
- SCRATCH, frequency * 2);
+ || (insn_extra_address_constraint
+ (lookup_constraint (constraints[i]))))
+ record_address_regs (VOIDmode, ADDR_SPACE_GENERIC,
+ recog_data.operand[i], 0, ADDRESS, SCRATCH,
+ frequency * 2);
}
/* Check for commutative in a separate loop so everything will have
record_reg_classes (recog_data.n_alternatives, recog_data.n_operands,
recog_data.operand, modes,
constraints, insn, pref);
+
+ /* If this insn is a single set copying operand 1 to operand 0 and
+ one operand is an allocno with the other a hard reg or an allocno
+ that prefers a hard register that is in its own register class
+ then we may want to adjust the cost of that register class to -1.
+
+ Avoid the adjustment if the source does not die to avoid
+ stressing of register allocator by preferencing two colliding
+ registers into single class.
+
+ Also avoid the adjustment if a copy between hard registers of the
+ class is expensive (ten times the cost of a default copy is
+ considered arbitrarily expensive). This avoids losing when the
+ preferred class is very expensive as the source of a copy
+ instruction. */
+ if ((set = single_set (insn)) != NULL_RTX
+ /* In rare cases the single set insn might have less 2 operands
+ as the source can be a fixed special reg. */
+ && recog_data.n_operands > 1
+ && ops[0] == SET_DEST (set) && ops[1] == SET_SRC (set))
+ {
+ int regno, other_regno;
+ rtx dest = SET_DEST (set);
+ rtx src = SET_SRC (set);
+
+ dest = SET_DEST (set);
+ src = SET_SRC (set);
+ if (GET_CODE (dest) == SUBREG
+ && (GET_MODE_SIZE (GET_MODE (dest))
+ == GET_MODE_SIZE (GET_MODE (SUBREG_REG (dest)))))
+ dest = SUBREG_REG (dest);
+ if (GET_CODE (src) == SUBREG
+ && (GET_MODE_SIZE (GET_MODE (src))
+ == GET_MODE_SIZE (GET_MODE (SUBREG_REG (src)))))
+ src = SUBREG_REG (src);
+ if (REG_P (src) && REG_P (dest)
+ && find_regno_note (insn, REG_DEAD, REGNO (src))
+ && (((regno = REGNO (src)) >= FIRST_PSEUDO_REGISTER
+ && (other_regno = REGNO (dest)) < FIRST_PSEUDO_REGISTER)
+ || ((regno = REGNO (dest)) >= FIRST_PSEUDO_REGISTER
+ && (other_regno = REGNO (src)) < FIRST_PSEUDO_REGISTER)))
+ {
+ machine_mode mode = GET_MODE (src);
+ cost_classes_t cost_classes_ptr = regno_cost_classes[regno];
+ enum reg_class *cost_classes = cost_classes_ptr->classes;
+ reg_class_t rclass;
+ int k, nr;
+
+ i = regno == (int) REGNO (src) ? 1 : 0;
+ for (k = cost_classes_ptr->num - 1; k >= 0; k--)
+ {
+ rclass = cost_classes[k];
+ if (TEST_HARD_REG_BIT (reg_class_contents[rclass], other_regno)
+ && (reg_class_size[(int) rclass]
+ == ira_reg_class_max_nregs [(int) rclass][(int) mode]))
+ {
+ if (reg_class_size[rclass] == 1)
+ op_costs[i]->cost[k] = -frequency;
+ else
+ {
+ for (nr = 0;
+ nr < hard_regno_nregs[other_regno][mode];
+ nr++)
+ if (! TEST_HARD_REG_BIT (reg_class_contents[rclass],
+ other_regno + nr))
+ break;
+
+ if (nr == hard_regno_nregs[other_regno][mode])
+ op_costs[i]->cost[k] = -frequency;
+ }
+ }
+ }
+ }
+ }
}
\f
/* Process one insn INSN. Scan it and record each time it would save
code to put a certain allocnos in a certain class. Return the last
insn processed, so that the scan can be continued from there. */
-static rtx
-scan_one_insn (rtx insn)
+static rtx_insn *
+scan_one_insn (rtx_insn *insn)
{
enum rtx_code pat_code;
rtx set, note;
return insn;
pat_code = GET_CODE (PATTERN (insn));
- if (pat_code == USE || pat_code == CLOBBER || pat_code == ASM_INPUT
- || pat_code == ADDR_VEC || pat_code == ADDR_DIFF_VEC)
+ if (pat_code == USE || pat_code == CLOBBER || pat_code == ASM_INPUT)
return insn;
counted_mem = false;
Similarly if we're loading other constants from memory (constant
pool, TOC references, small data areas, etc) and this is the only
- assignment to the destination pseudo. */
+ assignment to the destination pseudo.
+
+ Don't do this if SET_SRC (set) isn't a general operand, if it is
+ a memory requiring special instructions to load it, decreasing
+ mem_cost might result in it being loaded using the specialized
+ instruction into a register, then stored into stack and loaded
+ again from the stack. See PR52208.
+
+ Don't do this if SET_SRC (set) has side effect. See PR56124. */
if (set != 0 && REG_P (SET_DEST (set)) && MEM_P (SET_SRC (set))
&& (note = find_reg_note (insn, REG_EQUIV, NULL_RTX)) != NULL_RTX
- && ((MEM_P (XEXP (note, 0)))
+ && ((MEM_P (XEXP (note, 0))
+ && !side_effects_p (SET_SRC (set)))
|| (CONSTANT_P (XEXP (note, 0))
&& targetm.legitimate_constant_p (GET_MODE (SET_DEST (set)),
XEXP (note, 0))
- && REG_N_SETS (REGNO (SET_DEST (set))) == 1)))
+ && REG_N_SETS (REGNO (SET_DEST (set))) == 1))
+ && general_operand (SET_SRC (set), GET_MODE (SET_SRC (set))))
{
enum reg_class cl = GENERAL_REGS;
rtx reg = SET_DEST (set);
COSTS (costs, num)->mem_cost
-= ira_memory_move_cost[GET_MODE (reg)][cl][1] * frequency;
- record_address_regs (GET_MODE (SET_SRC (set)), XEXP (SET_SRC (set), 0),
- 0, MEM, SCRATCH, frequency * 2);
+ record_address_regs (GET_MODE (SET_SRC (set)),
+ MEM_ADDR_SPACE (SET_SRC (set)),
+ XEXP (SET_SRC (set), 0), 0, MEM, SCRATCH,
+ frequency * 2);
counted_mem = true;
}
if ((bb = ALLOCNO_LOOP_TREE_NODE (a)->bb) != NULL)
fprintf (f, "b%d", bb->index);
else
- fprintf (f, "l%d", ALLOCNO_LOOP_TREE_NODE (a)->loop->num);
+ fprintf (f, "l%d", ALLOCNO_LOOP_TREE_NODE (a)->loop_num);
fprintf (f, ") costs:");
for (k = 0; k < cost_classes_ptr->num; k++)
{
rclass = cost_classes[k];
- if (contains_reg_of_mode[rclass][PSEUDO_REGNO_MODE (regno)]
-#ifdef CANNOT_CHANGE_MODE_CLASS
- && ! invalid_mode_change_p (regno, (enum reg_class) rclass)
-#endif
- )
- {
- fprintf (f, " %s:%d", reg_class_names[rclass],
- COSTS (costs, i)->cost[k]);
- if (flag_ira_region == IRA_REGION_ALL
- || flag_ira_region == IRA_REGION_MIXED)
- fprintf (f, ",%d", COSTS (total_allocno_costs, i)->cost[k]);
- }
+ fprintf (f, " %s:%d", reg_class_names[rclass],
+ COSTS (costs, i)->cost[k]);
+ if (flag_ira_region == IRA_REGION_ALL
+ || flag_ira_region == IRA_REGION_MIXED)
+ fprintf (f, ",%d", COSTS (total_allocno_costs, i)->cost[k]);
}
fprintf (f, " MEM:%i", COSTS (costs, i)->mem_cost);
if (flag_ira_region == IRA_REGION_ALL
for (k = 0; k < cost_classes_ptr->num; k++)
{
rclass = cost_classes[k];
- if (contains_reg_of_mode[rclass][PSEUDO_REGNO_MODE (regno)]
-#ifdef CANNOT_CHANGE_MODE_CLASS
- && ! invalid_mode_change_p (regno, (enum reg_class) rclass)
-#endif
- )
- fprintf (f, " %s:%d", reg_class_names[rclass],
- COSTS (costs, regno)->cost[k]);
+ fprintf (f, " %s:%d", reg_class_names[rclass],
+ COSTS (costs, regno)->cost[k]);
}
fprintf (f, " MEM:%i\n", COSTS (costs, regno)->mem_cost);
}
static void
process_bb_for_costs (basic_block bb)
{
- rtx insn;
+ rtx_insn *insn;
frequency = REG_FREQ_FROM_BB (bb);
if (frequency == 0)
{
basic_block bb;
- FOR_EACH_BB (bb)
+ FOR_EACH_BB_FN (bb, cfun)
process_bb_for_costs (bb);
}
COSTS (total_allocno_costs, parent_a_num)->mem_cost
+= add_cost;
+ if (i >= first_moveable_pseudo && i < last_moveable_pseudo)
+ COSTS (total_allocno_costs, parent_a_num)->mem_cost = 0;
}
a_costs = COSTS (costs, a_num)->cost;
for (k = cost_classes_ptr->num - 1; k >= 0; k--)
i_mem_cost += add_cost;
}
}
- if (equiv_savings < 0)
+ if (i >= first_moveable_pseudo && i < last_moveable_pseudo)
+ i_mem_cost = 0;
+ else if (equiv_savings < 0)
i_mem_cost = -equiv_savings;
else if (equiv_savings > 0)
{
for (k = 0; k < cost_classes_ptr->num; k++)
{
rclass = cost_classes[k];
- /* Ignore classes that are too small or invalid for this
- operand. */
- if (! contains_reg_of_mode[rclass][PSEUDO_REGNO_MODE (i)]
-#ifdef CANNOT_CHANGE_MODE_CLASS
- || invalid_mode_change_p (i, (enum reg_class) rclass)
-#endif
- )
- continue;
if (i_costs[k] < best_cost)
{
best_cost = i_costs[k];
else if (i_costs[k] == best_cost)
best = ira_reg_class_subunion[best][rclass];
if (pass == flag_expensive_optimizations
- && i_costs[k] < i_mem_cost
+ /* We still prefer registers to memory even at this
+ stage if their costs are the same. We will make
+ a final decision during assigning hard registers
+ when we have all info including more accurate
+ costs which might be affected by assigning hard
+ registers to other pseudos because the pseudos
+ involved in moves can be coalesced. */
+ && i_costs[k] <= i_mem_cost
&& (reg_class_size[reg_class_subunion[alt_class][rclass]]
> reg_class_size[alt_class]))
alt_class = reg_class_subunion[alt_class][rclass];
alt_class = ira_allocno_class_translate[alt_class];
if (best_cost > i_mem_cost)
regno_aclass[i] = NO_REGS;
+ else if (!optimize && !targetm.class_likely_spilled_p (best))
+ /* Registers in the alternative class are likely to need
+ longer or slower sequences than registers in the best class.
+ When optimizing we make some effort to use the best class
+ over the alternative class where possible, but at -O0 we
+ effectively give the alternative class equal weight.
+ We then run the risk of using slower alternative registers
+ when plenty of registers from the best class are still free.
+ This is especially true because live ranges tend to be very
+ short in -O0 code and so register pressure tends to be low.
+
+ Avoid that by ignoring the alternative class if the best
+ class has plenty of registers. */
+ regno_aclass[i] = best;
else
{
/* Make the common class the biggest class of best and
a != NULL;
a = ALLOCNO_NEXT_REGNO_ALLOCNO (a))
{
- a_num = ALLOCNO_NUM (a);
- if (regno_aclass[i] == NO_REGS)
+ enum reg_class aclass = regno_aclass[i];
+ int a_num = ALLOCNO_NUM (a);
+ int *total_a_costs = COSTS (total_allocno_costs, a_num)->cost;
+ int *a_costs = COSTS (costs, a_num)->cost;
+
+ if (aclass == NO_REGS)
best = NO_REGS;
else
{
- int *total_a_costs = COSTS (total_allocno_costs, a_num)->cost;
- int *a_costs = COSTS (costs, a_num)->cost;
-
/* Finding best class which is subset of the common
class. */
best_cost = (1 << (HOST_BITS_PER_INT - 2)) - 1;
for (k = 0; k < cost_classes_ptr->num; k++)
{
rclass = cost_classes[k];
- if (! ira_class_subset_p[rclass][regno_aclass[i]])
+ if (! ira_class_subset_p[rclass][aclass])
continue;
- /* Ignore classes that are too small or invalid
- for this operand. */
- if (! contains_reg_of_mode[rclass][PSEUDO_REGNO_MODE (i)]
-#ifdef CANNOT_CHANGE_MODE_CLASS
- || invalid_mode_change_p (i, (enum reg_class) rclass)
-#endif
- )
- ;
- else if (total_a_costs[k] < best_cost)
+ if (total_a_costs[k] < best_cost)
{
best_cost = total_a_costs[k];
allocno_cost = a_costs[k];
fprintf (dump_file, "b%d", bb->index);
else
fprintf (dump_file, "l%d",
- ALLOCNO_LOOP_TREE_NODE (a)->loop->num);
+ ALLOCNO_LOOP_TREE_NODE (a)->loop_num);
fprintf (dump_file, ") best %s, allocno %s\n",
reg_class_names[best],
- reg_class_names[regno_aclass[i]]);
+ reg_class_names[aclass]);
}
pref[a_num] = best;
+ if (pass == flag_expensive_optimizations && best != aclass
+ && ira_class_hard_regs_num[best] > 0
+ && (ira_reg_class_max_nregs[best][ALLOCNO_MODE (a)]
+ >= ira_class_hard_regs_num[best]))
+ {
+ int ind = cost_classes_ptr->index[aclass];
+
+ ira_assert (ind >= 0);
+ ira_init_register_move_cost_if_necessary (ALLOCNO_MODE (a));
+ ira_add_allocno_pref (a, ira_class_hard_regs[best][0],
+ (a_costs[ind] - ALLOCNO_CLASS_COST (a))
+ / (ira_register_move_cost
+ [ALLOCNO_MODE (a)][best][aclass]));
+ for (k = 0; k < cost_classes_ptr->num; k++)
+ if (ira_class_subset_p[cost_classes[k]][best])
+ a_costs[k] = a_costs[ind];
+ }
}
}
/* Process moves involving hard regs to modify allocno hard register
costs. We can do this only after determining allocno class. If a
- hard register forms a register class, than moves with the hard
+ hard register forms a register class, then moves with the hard
register are already taken into account in class costs for the
allocno. */
static void
process_bb_node_for_hard_reg_moves (ira_loop_tree_node_t loop_tree_node)
{
- int i, freq, cost, src_regno, dst_regno, hard_regno;
+ int i, freq, src_regno, dst_regno, hard_regno, a_regno;
bool to_p;
- ira_allocno_t a;
- enum reg_class rclass, hard_reg_class;
- enum machine_mode mode;
+ ira_allocno_t a, curr_a;
+ ira_loop_tree_node_t curr_loop_tree_node;
+ enum reg_class rclass;
basic_block bb;
- rtx insn, set, src, dst;
+ rtx_insn *insn;
+ rtx set, src, dst;
bb = loop_tree_node->bb;
if (bb == NULL)
&& src_regno < FIRST_PSEUDO_REGISTER)
{
hard_regno = src_regno;
- to_p = true;
a = ira_curr_regno_allocno_map[dst_regno];
+ to_p = true;
}
else if (src_regno >= FIRST_PSEUDO_REGISTER
&& dst_regno < FIRST_PSEUDO_REGISTER)
{
hard_regno = dst_regno;
- to_p = false;
a = ira_curr_regno_allocno_map[src_regno];
+ to_p = false;
}
else
continue;
i = ira_class_hard_reg_index[rclass][hard_regno];
if (i < 0)
continue;
- mode = ALLOCNO_MODE (a);
- hard_reg_class = REGNO_REG_CLASS (hard_regno);
- ira_init_register_move_cost_if_necessary (mode);
- cost
- = (to_p ? ira_register_move_cost[mode][hard_reg_class][rclass]
- : ira_register_move_cost[mode][rclass][hard_reg_class]) * freq;
- ira_allocate_and_set_costs (&ALLOCNO_HARD_REG_COSTS (a), rclass,
- ALLOCNO_CLASS_COST (a));
- ira_allocate_and_set_costs (&ALLOCNO_CONFLICT_HARD_REG_COSTS (a),
- rclass, 0);
- ALLOCNO_HARD_REG_COSTS (a)[i] -= cost;
- ALLOCNO_CONFLICT_HARD_REG_COSTS (a)[i] -= cost;
- ALLOCNO_CLASS_COST (a) = MIN (ALLOCNO_CLASS_COST (a),
- ALLOCNO_HARD_REG_COSTS (a)[i]);
+ a_regno = ALLOCNO_REGNO (a);
+ for (curr_loop_tree_node = ALLOCNO_LOOP_TREE_NODE (a);
+ curr_loop_tree_node != NULL;
+ curr_loop_tree_node = curr_loop_tree_node->parent)
+ if ((curr_a = curr_loop_tree_node->regno_allocno_map[a_regno]) != NULL)
+ ira_add_allocno_pref (curr_a, hard_regno, freq);
+ {
+ int cost;
+ enum reg_class hard_reg_class;
+ machine_mode mode;
+
+ mode = ALLOCNO_MODE (a);
+ hard_reg_class = REGNO_REG_CLASS (hard_regno);
+ ira_init_register_move_cost_if_necessary (mode);
+ cost = (to_p ? ira_register_move_cost[mode][hard_reg_class][rclass]
+ : ira_register_move_cost[mode][rclass][hard_reg_class]) * freq;
+ ira_allocate_and_set_costs (&ALLOCNO_HARD_REG_COSTS (a), rclass,
+ ALLOCNO_CLASS_COST (a));
+ ira_allocate_and_set_costs (&ALLOCNO_CONFLICT_HARD_REG_COSTS (a),
+ rclass, 0);
+ ALLOCNO_HARD_REG_COSTS (a)[i] -= cost;
+ ALLOCNO_CONFLICT_HARD_REG_COSTS (a)[i] -= cost;
+ ALLOCNO_CLASS_COST (a) = MIN (ALLOCNO_CLASS_COST (a),
+ ALLOCNO_HARD_REG_COSTS (a)[i]);
+ }
}
}
}
/* Free allocated temporary cost vectors. */
-static void
-free_ira_costs (void)
+void
+target_ira_int::free_ira_costs ()
{
int i;
- free (init_cost);
- init_cost = NULL;
+ free (x_init_cost);
+ x_init_cost = NULL;
for (i = 0; i < MAX_RECOG_OPERANDS; i++)
{
- free (op_costs[i]);
- free (this_op_costs[i]);
- op_costs[i] = this_op_costs[i] = NULL;
+ free (x_op_costs[i]);
+ free (x_this_op_costs[i]);
+ x_op_costs[i] = x_this_op_costs[i] = NULL;
}
- free (temp_costs);
- temp_costs = NULL;
+ free (x_temp_costs);
+ x_temp_costs = NULL;
}
/* This is called each time register related information is
{
int i;
- free_ira_costs ();
+ this_target_ira_int->free_ira_costs ();
max_struct_costs_size
= sizeof (struct costs) + sizeof (int) * (ira_important_classes_num - 1);
/* Don't use ira_allocate because vectors live through several IRA
temp_costs = (struct costs *) xmalloc (max_struct_costs_size);
}
-/* Function called once at the end of compiler work. */
-void
-ira_finish_costs_once (void)
-{
- free_ira_costs ();
-}
-
\f
/* Common initialization function for ira_costs and
ira_free (total_allocno_costs);
}
-/* Entry function which defines classes for pseudos. */
+/* Entry function which defines classes for pseudos.
+ Set pseudo_classes_defined_p only if DEFINE_PSEUDO_CLASSES is true. */
void
-ira_set_pseudo_classes (FILE *dump_file)
+ira_set_pseudo_classes (bool define_pseudo_classes, FILE *dump_file)
{
allocno_p = false;
internal_flag_ira_verbose = flag_ira_verbose;
initiate_regno_cost_classes ();
find_costs_and_classes (dump_file);
finish_regno_cost_classes ();
- pseudo_classes_defined_p = true;
+ if (define_pseudo_classes)
+ pseudo_classes_defined_p = true;
+
finish_costs ();
}
int j, n, regno;
int cost, min_cost, *reg_costs;
enum reg_class aclass, rclass;
- enum machine_mode mode;
+ machine_mode mode;
ira_allocno_t a;
ira_allocno_iterator ai;
ira_allocno_object_iterator oi;
ira_object_t obj;
bool skip_p;
+ HARD_REG_SET *crossed_calls_clobber_regs;
FOR_EACH_ALLOCNO (a, ai)
{
mode = ALLOCNO_MODE (a);
n = ira_class_hard_regs_num[aclass];
min_cost = INT_MAX;
- if (ALLOCNO_CALLS_CROSSED_NUM (a) != 0)
+ if (ALLOCNO_CALLS_CROSSED_NUM (a)
+ != ALLOCNO_CHEAP_CALLS_CROSSED_NUM (a))
{
ira_allocate_and_set_costs
(&ALLOCNO_HARD_REG_COSTS (a), aclass,
continue;
rclass = REGNO_REG_CLASS (regno);
cost = 0;
- if (ira_hard_reg_set_intersection_p (regno, mode, call_used_reg_set)
- || HARD_REGNO_CALL_PART_CLOBBERED (regno, mode))
+ crossed_calls_clobber_regs
+ = &(ALLOCNO_CROSSED_CALLS_CLOBBERED_REGS (a));
+ if (ira_hard_reg_set_intersection_p (regno, mode,
+ *crossed_calls_clobber_regs)
+ && (ira_hard_reg_set_intersection_p (regno, mode,
+ call_used_reg_set)
+ || HARD_REGNO_CALL_PART_CLOBBERED (regno, mode)))
cost += (ALLOCNO_CALL_FREQ (a)
* (ira_memory_move_cost[mode][rclass][0]
+ ira_memory_move_cost[mode][rclass][1]));