Consider doloop IV in IVOPTs

For the targets supporting low-overhead loops, IVOPTs has to take care of
   the loops which will probably be transformed in RTL doloop optimization,
   to try to make selected IV candidate set optimal.  The process of doloop
   support includes:

   1) Analyze the current loop will be transformed to doloop or not, find and
      mark its compare type IV use as doloop use (iv_group field doloop_p), and
      set flag doloop_use_p of ivopts_data to notify subsequent processings on
      doloop.  See analyze_and_mark_doloop_use and its callees for the details.
      The target hook predict_doloop_p can be used for target specific checks.

   2) Add one doloop dedicated IV cand {(may_be_zero ? 1 : (niter + 1)), +, -1},
      set flag doloop_p of iv_cand, step cost is set as zero and no extra cost
      like biv.  For cost determination between doloop IV cand and IV use, the
      target hooks doloop_cost_for_generic and doloop_cost_for_address are
      provided to add on extra costs for generic type and address type IV use.
      Zero cost is assigned to the pair between doloop IV cand and doloop IV
      use, and bound zero is set for IV elimination.

   3) With the cost setting in step 2), the current cost model based IV
      selection algorithm will process as usual, pick up doloop dedicated IV if
      profitable.

gcc/ChangeLog

2019-09-14  Kewen Lin  <linkw@gcc.gnu.org>

    PR middle-end/80791
    * config/rs6000/rs6000.c (TARGET_HAVE_COUNT_REG_DECR_P): New macro.
    (TARGET_DOLOOP_COST_FOR_GENERIC): Likewise.
    (TARGET_DOLOOP_COST_FOR_ADDRESS): Likewise.
    * target.def (have_count_reg_decr_p): New hook.
    (doloop_cost_for_generic): Likewise.
    (doloop_cost_for_address): Likewise.
    * doc/tm.texi.in (TARGET_HAVE_COUNT_REG_DECR_P): Likewise.
    (TARGET_DOLOOP_COST_FOR_GENERIC): Likewise.
    (TARGET_DOLOOP_COST_FOR_ADDRESS): Likewise.
    * doc/tm.texi: Regenerate.
    * tree-ssa-loop-ivopts.c (comp_cost::operator+=): Consider infinite cost
    addend.
    (record_group): Init doloop_p.
    (add_candidate_1): Add optional argument doloop, change the handlings
    accordingly.
    (add_candidate): Likewise.
    (generic_predict_doloop_p): Update attribute.
    (force_expr_to_var_cost): Add costing for expressions COND_EXPR/LT_EXPR/
    LE_EXPR/GT_EXPR/GE_EXPR/EQ_EXPR/NE_EXPR/UNORDERED_EXPR/ORDERED_EXPR/
    UNLT_EXPR/UNLE_EXPR/UNGT_EXPR/UNGE_EXPR/UNEQ_EXPR/LTGT_EXPR/MAX_EXPR/
    MIN_EXPR.
    (get_computation_cost): Update for doloop IV cand extra cost.    
    (determine_group_iv_cost_cond): Update for doloop IV cand.
    (determine_iv_cost): Likewise.
    (ivopts_estimate_reg_pressure): Likewise.
    (may_eliminate_iv): Update handlings for doloop IV cand.
    (add_iv_candidate_for_doloop): New function.
    (find_iv_candidates): Call function add_iv_candidate_for_doloop.
    (iv_ca_set_no_cp): Update for doloop IV cand.
    (iv_ca_set_cp): Likewise.
    (iv_ca_dump): Dump register cost.
    (find_doloop_use): New function.
    (analyze_and_mark_doloop_use): Likewise.
    (tree_ssa_iv_optimize_loop): Call function analyze_and_mark_doloop_use.

gcc/testsuite/ChangeLog

2019-09-14  Kewen Lin  <linkw@gcc.gnu.org>

    PR middle-end/80791
    * gcc.dg/tree-ssa/ivopts-3.c: Adjust for doloop change.
    * gcc.dg/tree-ssa/ivopts-lt.c: Likewise.
    * gcc.dg/tree-ssa/pr32044.c: Likewise.

From-SVN: r275713
This commit is contained in:
Kewen Lin 2019-09-14 09:01:21 +00:00
parent f389864485
commit 603a4ad499
10 changed files with 400 additions and 26 deletions

View File

@ -1,3 +1,41 @@
2019-09-14 Kewen Lin <linkw@gcc.gnu.org>
PR middle-end/80791
* config/rs6000/rs6000.c (TARGET_HAVE_COUNT_REG_DECR_P): New macro.
(TARGET_DOLOOP_COST_FOR_GENERIC): Likewise.
(TARGET_DOLOOP_COST_FOR_ADDRESS): Likewise.
* target.def (have_count_reg_decr_p): New hook.
(doloop_cost_for_generic): Likewise.
(doloop_cost_for_address): Likewise.
* doc/tm.texi.in (TARGET_HAVE_COUNT_REG_DECR_P): Likewise.
(TARGET_DOLOOP_COST_FOR_GENERIC): Likewise.
(TARGET_DOLOOP_COST_FOR_ADDRESS): Likewise.
* doc/tm.texi: Regenerate.
* tree-ssa-loop-ivopts.c (comp_cost::operator+=): Consider infinite cost
addend.
(record_group): Init doloop_p.
(add_candidate_1): Add optional argument doloop, change the handlings
accordingly.
(add_candidate): Likewise.
(generic_predict_doloop_p): Update attribute.
(force_expr_to_var_cost): Add costing for expressions COND_EXPR/LT_EXPR/
LE_EXPR/GT_EXPR/GE_EXPR/EQ_EXPR/NE_EXPR/UNORDERED_EXPR/ORDERED_EXPR/
UNLT_EXPR/UNLE_EXPR/UNGT_EXPR/UNGE_EXPR/UNEQ_EXPR/LTGT_EXPR/MAX_EXPR/
MIN_EXPR.
(get_computation_cost): Update for doloop IV cand extra cost.
(determine_group_iv_cost_cond): Update for doloop IV cand.
(determine_iv_cost): Likewise.
(ivopts_estimate_reg_pressure): Likewise.
(may_eliminate_iv): Update handlings for doloop IV cand.
(add_iv_candidate_for_doloop): New function.
(find_iv_candidates): Call function add_iv_candidate_for_doloop.
(iv_ca_set_no_cp): Update for doloop IV cand.
(iv_ca_set_cp): Likewise.
(iv_ca_dump): Dump register cost.
(find_doloop_use): New function.
(analyze_and_mark_doloop_use): Likewise.
(tree_ssa_iv_optimize_loop): Call function analyze_and_mark_doloop_use.
2019-09-13 Bernd Edlinger <bernd.edlinger@hotmail.de>
PR middle-end/91708

View File

@ -1651,6 +1651,16 @@ static const struct attribute_spec rs6000_attribute_table[] =
#undef TARGET_PREDICT_DOLOOP_P
#define TARGET_PREDICT_DOLOOP_P rs6000_predict_doloop_p
#undef TARGET_HAVE_COUNT_REG_DECR_P
#define TARGET_HAVE_COUNT_REG_DECR_P true
/* 1000000000 is infinite cost in IVOPTs. */
#undef TARGET_DOLOOP_COST_FOR_GENERIC
#define TARGET_DOLOOP_COST_FOR_GENERIC 1000000000
#undef TARGET_DOLOOP_COST_FOR_ADDRESS
#define TARGET_DOLOOP_COST_FOR_ADDRESS 1000000000
#undef TARGET_ATOMIC_ASSIGN_EXPAND_FENV
#define TARGET_ATOMIC_ASSIGN_EXPAND_FENV rs6000_atomic_assign_expand_fenv

View File

@ -11615,6 +11615,36 @@ loops, and will help ivopts to make some decisions.
The default version of this hook returns false.
@end deftypefn
@deftypevr {Target Hook} bool TARGET_HAVE_COUNT_REG_DECR_P
Return true if the target supports hardware count register for decrement
and branch.
The default value is false.
@end deftypevr
@deftypevr {Target Hook} int64_t TARGET_DOLOOP_COST_FOR_GENERIC
One IV candidate dedicated for doloop is introduced in IVOPTs, we can
calculate the computation cost of adopting it to any generic IV use by
function get_computation_cost as before. But for targets which have
hardware count register support for decrement and branch, it may have to
move IV value from hardware count register to general purpose register
while doloop IV candidate is used for generic IV uses. It probably takes
expensive penalty. This hook allows target owners to define the cost for
this especially for generic IV uses.
The default value is zero.
@end deftypevr
@deftypevr {Target Hook} int64_t TARGET_DOLOOP_COST_FOR_ADDRESS
One IV candidate dedicated for doloop is introduced in IVOPTs, we can
calculate the computation cost of adopting it to any address IV use by
function get_computation_cost as before. But for targets which have
hardware count register support for decrement and branch, it may have to
move IV value from hardware count register to general purpose register
while doloop IV candidate is used for address IV uses. It probably takes
expensive penalty. This hook allows target owners to define the cost for
this escpecially for address IV uses.
The default value is zero.
@end deftypevr
@deftypefn {Target Hook} bool TARGET_CAN_USE_DOLOOP_P (const widest_int @var{&iterations}, const widest_int @var{&iterations_max}, unsigned int @var{loop_depth}, bool @var{entered_at_top})
Return true if it is possible to use low-overhead loops (@code{doloop_end}
and @code{doloop_begin}) for a particular loop. @var{iterations} gives the

View File

@ -7950,6 +7950,12 @@ to by @var{ce_info}.
@hook TARGET_PREDICT_DOLOOP_P
@hook TARGET_HAVE_COUNT_REG_DECR_P
@hook TARGET_DOLOOP_COST_FOR_GENERIC
@hook TARGET_DOLOOP_COST_FOR_ADDRESS
@hook TARGET_CAN_USE_DOLOOP_P
@hook TARGET_INVALID_WITHIN_DOLOOP

View File

@ -4225,6 +4225,39 @@ The default version of this hook returns false.",
bool, (class loop *loop),
default_predict_doloop_p)
DEFHOOKPOD
(have_count_reg_decr_p,
"Return true if the target supports hardware count register for decrement\n\
and branch.\n\
The default value is false.",
bool, false)
DEFHOOKPOD
(doloop_cost_for_generic,
"One IV candidate dedicated for doloop is introduced in IVOPTs, we can\n\
calculate the computation cost of adopting it to any generic IV use by\n\
function get_computation_cost as before. But for targets which have\n\
hardware count register support for decrement and branch, it may have to\n\
move IV value from hardware count register to general purpose register\n\
while doloop IV candidate is used for generic IV uses. It probably takes\n\
expensive penalty. This hook allows target owners to define the cost for\n\
this especially for generic IV uses.\n\
The default value is zero.",
int64_t, 0)
DEFHOOKPOD
(doloop_cost_for_address,
"One IV candidate dedicated for doloop is introduced in IVOPTs, we can\n\
calculate the computation cost of adopting it to any address IV use by\n\
function get_computation_cost as before. But for targets which have\n\
hardware count register support for decrement and branch, it may have to\n\
move IV value from hardware count register to general purpose register\n\
while doloop IV candidate is used for address IV uses. It probably takes\n\
expensive penalty. This hook allows target owners to define the cost for\n\
this escpecially for address IV uses.\n\
The default value is zero.",
int64_t, 0)
DEFHOOK
(can_use_doloop_p,
"Return true if it is possible to use low-overhead loops (@code{doloop_end}\n\

View File

@ -1,3 +1,10 @@
2019-09-14 Kewen Lin <linkw@gcc.gnu.org>
PR middle-end/80791
* gcc.dg/tree-ssa/ivopts-3.c: Adjust for doloop change.
* gcc.dg/tree-ssa/ivopts-lt.c: Likewise.
* gcc.dg/tree-ssa/pr32044.c: Likewise.
2019-09-13 Steven G. Kargl <kargl@gcc.gnu.org>
PR fortran/91566

View File

@ -10,4 +10,6 @@ int main (void)
f2 ();
}
/* { dg-final { scan-tree-dump-times "!= 0" 5 "ivopts" } } */
/* { dg-final { scan-tree-dump-times "!= 0" 5 "ivopts" { target { ! powerpc*-*-* } } } } */
/* More debug information emitted for doloop on powerpc. */
/* { dg-final { scan-tree-dump-times "!= 0" 6 "ivopts" { target { powerpc*-*-* } } } } */

View File

@ -17,6 +17,7 @@ f1 (char *p, uintptr_t i, uintptr_t n)
while (i < n);
}
/* { dg-final { scan-tree-dump-times "PHI" 1 "ivopts" } } */
/* { dg-final { scan-tree-dump-times "PHI <p_" 1 "ivopts"} } */
/* { dg-final { scan-tree-dump-times "p_\[0-9\]* <" 1 "ivopts" } } */
/* { dg-final { scan-tree-dump-times "PHI" 1 "ivopts" { target { ! powerpc*-*-* } } } } */
/* { dg-final { scan-tree-dump-times "PHI" 2 "ivopts" { target { powerpc*-*-* } } } } */
/* { dg-final { scan-tree-dump-times "PHI <p_" 1 "ivopts" { target { ! powerpc*-*-* } } } } */
/* { dg-final { scan-tree-dump-times "p_\[0-9\]* <" 1 "ivopts" { target { ! powerpc*-*-* } } } } */

View File

@ -1,6 +1,10 @@
/* { dg-do compile } */
/* { dg-options "-O2 -fdump-tree-optimized" } */
/* For powerpc, disable doloop IV cand generation in IVOPTs to avoid unexpected
division operation for its base setup. */
/* { dg-additional-options "-fno-branch-count-reg" { target { powerpc*-*-* } } } */
int foo (int n)
{
while (n >= 45)

View File

@ -64,7 +64,30 @@ along with GCC; see the file COPYING3. If not see
All of this is done loop by loop. Doing it globally is theoretically
possible, it might give a better performance and it might enable us
to decide costs more precisely, but getting all the interactions right
would be complicated. */
would be complicated.
For the targets supporting low-overhead loops, IVOPTs has to take care of
the loops which will probably be transformed in RTL doloop optimization,
to try to make selected IV candidate set optimal. The process of doloop
support includes:
1) Analyze the current loop will be transformed to doloop or not, find and
mark its compare type IV use as doloop use (iv_group field doloop_p), and
set flag doloop_use_p of ivopts_data to notify subsequent processings on
doloop. See analyze_and_mark_doloop_use and its callees for the details.
The target hook predict_doloop_p can be used for target specific checks.
2) Add one doloop dedicated IV cand {(may_be_zero ? 1 : (niter + 1)), +, -1},
set flag doloop_p of iv_cand, step cost is set as zero and no extra cost
like biv. For cost determination between doloop IV cand and IV use, the
target hooks doloop_cost_for_generic and doloop_cost_for_address are
provided to add on extra costs for generic type and address type IV use.
Zero cost is assigned to the pair between doloop IV cand and doloop IV
use, and bound zero is set for IV elimination.
3) With the cost setting in step 2), the current cost model based IV
selection algorithm will process as usual, pick up doloop dedicated IV if
profitable. */
#include "config.h"
#include "system.h"
@ -276,6 +299,9 @@ comp_cost::operator+= (comp_cost cost)
comp_cost
comp_cost::operator+= (HOST_WIDE_INT c)
{
if (c >= INFTY)
this->cost = INFTY;
if (infinite_cost_p ())
return *this;
@ -401,6 +427,8 @@ struct iv_group
class cost_pair *cost_map;
/* The selected candidate for the group. */
struct iv_cand *selected;
/* To indicate this is a doloop use group. */
bool doloop_p;
/* Uses in the group. */
vec<struct iv_use *> vuses;
};
@ -441,6 +469,7 @@ struct iv_cand
be hoisted out of loop. */
struct iv *orig_iv; /* The original iv if this cand is added from biv with
smaller type. */
bool doloop_p; /* Whether this is a doloop candidate. */
};
/* Hashtable entry for common candidate derived from iv uses. */
@ -618,6 +647,9 @@ struct ivopts_data
/* Whether the loop body can only be exited via single exit. */
bool loop_single_exit_p;
/* Whether the loop has doloop comparison use. */
bool doloop_use_p;
};
/* An assignment of iv candidates to uses. */
@ -1542,6 +1574,7 @@ record_group (struct ivopts_data *data, enum use_type type)
group->type = type;
group->related_cands = BITMAP_ALLOC (NULL);
group->vuses.create (1);
group->doloop_p = false;
data->vgroups.safe_push (group);
return group;
@ -3033,10 +3066,10 @@ get_loop_invariant_expr (struct ivopts_data *data, tree inv_expr)
replacement of the final value of the iv by a direct computation. */
static struct iv_cand *
add_candidate_1 (struct ivopts_data *data,
tree base, tree step, bool important, enum iv_position pos,
struct iv_use *use, gimple *incremented_at,
struct iv *orig_iv = NULL)
add_candidate_1 (struct ivopts_data *data, tree base, tree step, bool important,
enum iv_position pos, struct iv_use *use,
gimple *incremented_at, struct iv *orig_iv = NULL,
bool doloop = false)
{
unsigned i;
struct iv_cand *cand = NULL;
@ -3095,11 +3128,15 @@ add_candidate_1 (struct ivopts_data *data,
cand->pos = pos;
if (pos != IP_ORIGINAL)
{
cand->var_before = create_tmp_var_raw (TREE_TYPE (base), "ivtmp");
if (doloop)
cand->var_before = create_tmp_var_raw (TREE_TYPE (base), "doloop");
else
cand->var_before = create_tmp_var_raw (TREE_TYPE (base), "ivtmp");
cand->var_after = cand->var_before;
}
cand->important = important;
cand->incremented_at = incremented_at;
cand->doloop_p = doloop;
data->vcands.safe_push (cand);
if (!poly_int_tree_p (step))
@ -3132,6 +3169,7 @@ add_candidate_1 (struct ivopts_data *data,
}
cand->important |= important;
cand->doloop_p |= doloop;
/* Relate candidate to the group for which it is added. */
if (use)
@ -3225,14 +3263,16 @@ add_autoinc_candidates (struct ivopts_data *data, tree base, tree step,
the end of loop. */
static void
add_candidate (struct ivopts_data *data,
tree base, tree step, bool important, struct iv_use *use,
struct iv *orig_iv = NULL)
add_candidate (struct ivopts_data *data, tree base, tree step, bool important,
struct iv_use *use, struct iv *orig_iv = NULL,
bool doloop = false)
{
if (ip_normal_pos (data->current_loop))
add_candidate_1 (data, base, step, important,
IP_NORMAL, use, NULL, orig_iv);
if (ip_end_pos (data->current_loop)
add_candidate_1 (data, base, step, important, IP_NORMAL, use, NULL, orig_iv,
doloop);
/* Exclude doloop candidate here since it requires decrement then comparison
and jump, the IP_END position doesn't match. */
if (!doloop && ip_end_pos (data->current_loop)
&& allow_ip_end_pos_p (data->current_loop))
add_candidate_1 (data, base, step, important, IP_END, use, NULL, orig_iv);
}
@ -3760,7 +3800,7 @@ prepare_decl_rtl (tree *expr_p, int *ws, void *data)
Some RTL specific checks seems unable to be checked in gimple, if any new
checks or easy checks _are_ missing here, please add them. */
static bool ATTRIBUTE_UNUSED
static bool
generic_predict_doloop_p (struct ivopts_data *data)
{
class loop *loop = data->current_loop;
@ -4213,6 +4253,36 @@ force_expr_to_var_cost (tree expr, bool speed)
STRIP_NOPS (op0);
op1 = NULL_TREE;
break;
/* See add_iv_candidate_for_doloop, for doloop may_be_zero case, we
introduce COND_EXPR for IV base, need to support better cost estimation
for this COND_EXPR and tcc_comparison. */
case COND_EXPR:
op0 = TREE_OPERAND (expr, 1);
STRIP_NOPS (op0);
op1 = TREE_OPERAND (expr, 2);
STRIP_NOPS (op1);
break;
case LT_EXPR:
case LE_EXPR:
case GT_EXPR:
case GE_EXPR:
case EQ_EXPR:
case NE_EXPR:
case UNORDERED_EXPR:
case ORDERED_EXPR:
case UNLT_EXPR:
case UNLE_EXPR:
case UNGT_EXPR:
case UNGE_EXPR:
case UNEQ_EXPR:
case LTGT_EXPR:
case MAX_EXPR:
case MIN_EXPR:
op0 = TREE_OPERAND (expr, 0);
STRIP_NOPS (op0);
op1 = TREE_OPERAND (expr, 1);
STRIP_NOPS (op1);
break;
default:
/* Just an arbitrary value, FIXME. */
@ -4294,6 +4364,35 @@ force_expr_to_var_cost (tree expr, bool speed)
case RSHIFT_EXPR:
cost = comp_cost (add_cost (speed, mode), 0);
break;
case COND_EXPR:
op0 = TREE_OPERAND (expr, 0);
STRIP_NOPS (op0);
if (op0 == NULL_TREE || TREE_CODE (op0) == SSA_NAME
|| CONSTANT_CLASS_P (op0))
cost = no_cost;
else
cost = force_expr_to_var_cost (op0, speed);
break;
case LT_EXPR:
case LE_EXPR:
case GT_EXPR:
case GE_EXPR:
case EQ_EXPR:
case NE_EXPR:
case UNORDERED_EXPR:
case ORDERED_EXPR:
case UNLT_EXPR:
case UNLE_EXPR:
case UNGT_EXPR:
case UNGE_EXPR:
case UNEQ_EXPR:
case LTGT_EXPR:
case MAX_EXPR:
case MIN_EXPR:
/* Simply use add cost for now, FIXME if there is some more accurate cost
evaluation way. */
cost = comp_cost (add_cost (speed, mode), 0);
break;
default:
gcc_unreachable ();
@ -4670,7 +4769,10 @@ get_computation_cost (struct ivopts_data *data, struct iv_use *use,
{
cost = get_address_cost (data, use, cand, &aff_inv, &aff_var, ratio,
inv_vars, inv_expr, can_autoinc, speed);
return get_scaled_computation_cost_at (data, at, cost);
cost = get_scaled_computation_cost_at (data, at, cost);
/* For doloop IV cand, add on the extra cost. */
cost += cand->doloop_p ? targetm.doloop_cost_for_address : 0;
return cost;
}
bool simple_inv = (aff_combination_const_p (&aff_inv)
@ -4720,7 +4822,13 @@ get_computation_cost (struct ivopts_data *data, struct iv_use *use,
if (comp_inv && !integer_zerop (comp_inv))
cost += add_cost (speed, TYPE_MODE (utype));
return get_scaled_computation_cost_at (data, at, cost);
cost = get_scaled_computation_cost_at (data, at, cost);
/* For doloop IV cand, add on the extra cost. */
if (cand->doloop_p && use->type == USE_NONLINEAR_EXPR)
cost += targetm.doloop_cost_for_generic;
return cost;
}
/* Determines cost of computing the use in GROUP with CAND in a generic
@ -5178,6 +5286,15 @@ may_eliminate_iv (struct ivopts_data *data,
}
}
/* For doloop IV cand, the bound would be zero. It's safe whether
may_be_zero set or not. */
if (cand->doloop_p)
{
*bound = build_int_cst (TREE_TYPE (cand->iv->base), 0);
*comp = iv_elimination_compare (data, use);
return true;
}
cand_value_at (loop, cand, use->stmt, desc->niter, &bnd);
*bound = fold_convert (TREE_TYPE (cand->iv->base),
@ -5300,6 +5417,9 @@ determine_group_iv_cost_cond (struct ivopts_data *data,
inv_vars = inv_vars_elim;
inv_vars_elim = NULL;
inv_expr = inv_expr_elim;
/* For doloop candidate/use pair, adjust to zero cost. */
if (group->doloop_p && cand->doloop_p && elim_cost.cost > no_cost.cost)
cost = no_cost;
}
else
{
@ -5426,6 +5546,42 @@ relate_compare_use_with_all_cands (struct ivopts_data *data)
}
}
/* Add one doloop dedicated IV candidate:
- Base is (may_be_zero ? 1 : (niter + 1)).
- Step is -1. */
static void
add_iv_candidate_for_doloop (struct ivopts_data *data)
{
tree_niter_desc *niter_desc = niter_for_single_dom_exit (data);
gcc_assert (niter_desc && niter_desc->assumptions);
tree niter = niter_desc->niter;
tree ntype = TREE_TYPE (niter);
gcc_assert (TREE_CODE (ntype) == INTEGER_TYPE);
tree may_be_zero = niter_desc->may_be_zero;
if (may_be_zero && integer_zerop (may_be_zero))
may_be_zero = NULL_TREE;
if (may_be_zero)
{
if (COMPARISON_CLASS_P (may_be_zero))
{
niter = fold_build3 (COND_EXPR, ntype, may_be_zero,
build_int_cst (ntype, 0),
rewrite_to_non_trapping_overflow (niter));
}
/* Don't try to obtain the iteration count expression when may_be_zero is
integer_nonzerop (actually iteration count is one) or else. */
else
return;
}
tree base = fold_build2 (PLUS_EXPR, ntype, unshare_expr (niter),
build_int_cst (ntype, 1));
add_candidate (data, base, build_int_cst (ntype, -1), true, NULL, NULL, true);
}
/* Finds the candidates for the induction variables. */
static void
@ -5434,6 +5590,10 @@ find_iv_candidates (struct ivopts_data *data)
/* Add commonly used ivs. */
add_standard_iv_candidates (data);
/* Add doloop dedicated ivs. */
if (data->doloop_use_p)
add_iv_candidate_for_doloop (data);
/* Add old induction variables. */
add_iv_candidate_for_bivs (data);
@ -5614,16 +5774,21 @@ determine_iv_cost (struct ivopts_data *data, struct iv_cand *cand)
or a const set. */
if (cost_base.cost == 0)
cost_base.cost = COSTS_N_INSNS (1);
cost_step = add_cost (data->speed, TYPE_MODE (TREE_TYPE (base)));
/* Doloop decrement should be considered as zero cost. */
if (cand->doloop_p)
cost_step = 0;
else
cost_step = add_cost (data->speed, TYPE_MODE (TREE_TYPE (base)));
cost = cost_step + adjust_setup_cost (data, cost_base.cost);
/* Prefer the original ivs unless we may gain something by replacing it.
The reason is to make debugging simpler; so this is not relevant for
artificial ivs created by other optimization passes. */
if (cand->pos != IP_ORIGINAL
|| !SSA_NAME_VAR (cand->var_before)
|| DECL_ARTIFICIAL (SSA_NAME_VAR (cand->var_before)))
if ((cand->pos != IP_ORIGINAL
|| !SSA_NAME_VAR (cand->var_before)
|| DECL_ARTIFICIAL (SSA_NAME_VAR (cand->var_before)))
/* Prefer doloop as well. */
&& !cand->doloop_p)
cost++;
/* Prefer not to insert statements into latch unless there are some
@ -5868,7 +6033,8 @@ iv_ca_set_no_cp (struct ivopts_data *data, class iv_ca *ivs,
if (ivs->n_cand_uses[cid] == 0)
{
bitmap_clear_bit (ivs->cands, cid);
ivs->n_cands--;
if (!cp->cand->doloop_p || !targetm.have_count_reg_decr_p)
ivs->n_cands--;
ivs->cand_cost -= cp->cand->cost;
iv_ca_set_remove_invs (ivs, cp->cand->inv_vars, ivs->n_inv_var_uses);
iv_ca_set_remove_invs (ivs, cp->cand->inv_exprs, ivs->n_inv_expr_uses);
@ -5925,7 +6091,8 @@ iv_ca_set_cp (struct ivopts_data *data, class iv_ca *ivs,
if (ivs->n_cand_uses[cid] == 1)
{
bitmap_set_bit (ivs->cands, cid);
ivs->n_cands++;
if (!cp->cand->doloop_p || !targetm.have_count_reg_decr_p)
ivs->n_cands++;
ivs->cand_cost += cp->cand->cost;
iv_ca_set_add_invs (ivs, cp->cand->inv_vars, ivs->n_inv_var_uses);
iv_ca_set_add_invs (ivs, cp->cand->inv_exprs, ivs->n_inv_expr_uses);
@ -6170,6 +6337,8 @@ iv_ca_dump (struct ivopts_data *data, FILE *file, class iv_ca *ivs)
fprintf (file, " cost: %" PRId64 " (complexity %d)\n", cost.cost,
cost.complexity);
fprintf (file, " reg_cost: %d\n",
ivopts_estimate_reg_pressure (data, ivs->n_invs, ivs->n_cands));
fprintf (file, " cand_cost: %" PRId64 "\n cand_group_cost: "
"%" PRId64 " (complexity %d)\n", ivs->cand_cost,
ivs->cand_use_cost.cost, ivs->cand_use_cost.complexity);
@ -7608,6 +7777,77 @@ determine_scaling_factor (struct ivopts_data *data, basic_block *body)
}
}
/* Find doloop comparison use and set its doloop_p on if found. */
static bool
find_doloop_use (struct ivopts_data *data)
{
struct loop *loop = data->current_loop;
for (unsigned i = 0; i < data->vgroups.length (); i++)
{
struct iv_group *group = data->vgroups[i];
if (group->type == USE_COMPARE)
{
gcc_assert (group->vuses.length () == 1);
struct iv_use *use = group->vuses[0];
gimple *stmt = use->stmt;
if (gimple_code (stmt) == GIMPLE_COND)
{
basic_block bb = gimple_bb (stmt);
edge true_edge, false_edge;
extract_true_false_edges_from_block (bb, &true_edge, &false_edge);
/* This comparison is used for loop latch. Require latch is empty
for now. */
if ((loop->latch == true_edge->dest
|| loop->latch == false_edge->dest)
&& empty_block_p (loop->latch))
{
group->doloop_p = true;
if (dump_file && (dump_flags & TDF_DETAILS))
{
fprintf (dump_file, "Doloop cmp iv use: ");
print_gimple_stmt (dump_file, stmt, TDF_DETAILS);
}
return true;
}
}
}
}
return false;
}
/* For the targets which support doloop, to predict whether later RTL doloop
transformation will perform on this loop, further detect the doloop use and
mark the flag doloop_use_p if predicted. */
void
analyze_and_mark_doloop_use (struct ivopts_data *data)
{
data->doloop_use_p = false;
if (!flag_branch_on_count_reg)
return;
if (!generic_predict_doloop_p (data))
return;
if (find_doloop_use (data))
{
data->doloop_use_p = true;
if (dump_file && (dump_flags & TDF_DETAILS))
{
struct loop *loop = data->current_loop;
fprintf (dump_file,
"Predict loop %d can perform"
" doloop optimization later.\n",
loop->num);
flow_loop_dump (loop, dump_file, NULL, 1);
}
}
}
/* Optimizes the LOOP. Returns true if anything changed. */
static bool
@ -7662,6 +7902,9 @@ tree_ssa_iv_optimize_loop (struct ivopts_data *data, class loop *loop,
/* Determine cost scaling factor for basic blocks in loop. */
determine_scaling_factor (data, body);
/* Analyze doloop possibility and mark the doloop use if predicted. */
analyze_and_mark_doloop_use (data);
/* Finds candidates for the induction variables (item 2). */
find_iv_candidates (data);