re PR tree-optimization/40436 (0.5% code size regression caused by r147852)

PR tree-optimization/40436
	* ipa-inline.c (leaf_node_p): Implement using is_inexpensive_builtin.
	* tree-inline.c (estimate_num_insns): Inexpensive builtins are like
	normal instructions; be sure bultin is not implemented in this file;
	compute non-zero return cost.
	(init_inline_once): Reduce builtin_call_cost to 1; set return cost.
	* tree-inline.h (eni_weights_d): Add return cost.

From-SVN: r166517
This commit is contained in:
Jan Hubicka 2010-11-10 03:35:19 +01:00 committed by Jan Hubicka
parent b8cbdff525
commit 9bb2f479f3
6 changed files with 36 additions and 9 deletions

View File

@ -1,3 +1,13 @@
2010-11-09 Jan Hubicka <jh@suse.cz>
PR tree-optimization/40436
* ipa-inline.c (leaf_node_p): Implement using is_inexpensive_builtin.
* tree-inline.c (estimate_num_insns): Inexpensive builtins are like
normal instructions; be sure bultin is not implemented in this file;
compute non-zero return cost.
(init_inline_once): Reduce builtin_call_cost to 1; set return cost.
* tree-inline.h (eni_weights_d): Add return cost.
2010-11-09 Joseph Myers <joseph@codesourcery.com>
* c-parser.c (c_parser_struct_declaration): Handle declaration

View File

@ -1578,16 +1578,15 @@ cgraph_decide_inlining (void)
return 0;
}
/* Return true when N is leaf function. Accept cheap (pure&const) builtins
/* Return true when N is leaf function. Accept cheap builtins
in leaf functions. */
static bool
leaf_node_p (struct cgraph_node *n)
{
struct cgraph_edge *e;
for (e = n->callees; e; e = e->next_callee)
if (!DECL_BUILT_IN (e->callee->decl)
|| (!TREE_READONLY (e->callee->decl)
|| DECL_PURE_P (e->callee->decl)))
if (!is_inexpensive_builtin (e->callee->decl))
return false;
return true;
}

View File

@ -1,3 +1,7 @@
2010-11-09 Jan Hubicka <jh@suse.cz>
* testsuite/gcc.target/i386/recip-vec-sqrtf-avx.c: Update for loop unrolling.
2010-11-09 Joseph Myers <joseph@codesourcery.com>
* gcc.dg/struct-semi-4.c: New test.

View File

@ -31,4 +31,5 @@ void t3(void)
r[i] = sqrtf (a[i]);
}
/* { dg-final { scan-assembler-times "vrsqrtps\[ \\t\]+\[^\n\]*%ymm" 3 } } */
/* Last loop is small enough to be fully unrolled. */
/* { dg-final { scan-assembler-times "vrsqrtps\[ \\t\]+\[^\n\]*%ymm" 4 } } */

View File

@ -3484,10 +3484,16 @@ estimate_num_insns (gimple stmt, eni_weights *weights)
if (POINTER_TYPE_P (funtype))
funtype = TREE_TYPE (funtype);
if (is_simple_builtin (decl))
/* Do not special case builtins where we see the body.
This just confuse inliner. */
if (!decl || cgraph_node (decl)->analyzed)
cost = weights->call_cost;
/* For buitins that are likely expanded to nothing or
inlined do not account operand costs. */
else if (is_simple_builtin (decl))
return 0;
else if (is_inexpensive_builtin (decl))
cost = weights->target_builtin_call_cost;
return weights->target_builtin_call_cost;
else
cost = weights->call_cost;
@ -3536,11 +3542,13 @@ estimate_num_insns (gimple stmt, eni_weights *weights)
break;
}
case GIMPLE_RETURN:
return weights->return_cost;
case GIMPLE_GOTO:
case GIMPLE_LABEL:
case GIMPLE_NOP:
case GIMPLE_PHI:
case GIMPLE_RETURN:
case GIMPLE_PREDICT:
case GIMPLE_DEBUG:
return 0;
@ -3640,16 +3648,18 @@ init_inline_once (void)
eni_size_weights.div_mod_cost = 1;
eni_size_weights.omp_cost = 40;
eni_size_weights.time_based = false;
eni_size_weights.return_cost = 1;
/* Estimating time for call is difficult, since we have no idea what the
called function does. In the current uses of eni_time_weights,
underestimating the cost does less harm than overestimating it, so
we choose a rather small value here. */
eni_time_weights.call_cost = 10;
eni_time_weights.target_builtin_call_cost = 10;
eni_time_weights.target_builtin_call_cost = 1;
eni_time_weights.div_mod_cost = 10;
eni_time_weights.omp_cost = 40;
eni_time_weights.time_based = true;
eni_time_weights.return_cost = 2;
}
/* Estimate the number of instructions in a gimple_seq. */

View File

@ -144,6 +144,9 @@ typedef struct eni_weights_d
/* Cost for omp construct. */
unsigned omp_cost;
/* Cost of return. */
unsigned return_cost;
/* True when time of statemnt should be estimated. Thus i.e
cost of switch statement is logarithmic rather than linear in number
of cases. */