re PR fortran/25620 (Missed optimization with power)

2006-11-27  Richard Guenther  <rguenther@suse.de>

	PR middle-end/25620
	* builtins.c (expand_builtin_pow): Optimize non integer valued
	constant exponents using sqrt or cbrt if possible.  Always fall back
	to expanding via optabs.

	* gcc.target/i386/pow-1.c: New testcase.
	* gcc.dg/builtins-58.c: Likewise.

From-SVN: r119248
This commit is contained in:
Richard Guenther 2006-11-27 11:38:42 +00:00 committed by Richard Biener
parent 4090db01e8
commit 9fca6f972d
5 changed files with 171 additions and 24 deletions

View File

@ -1,3 +1,10 @@
2006-11-27 Richard Guenther <rguenther@suse.de>
PR middle-end/25620
* builtins.c (expand_builtin_pow): Optimize non integer valued
constant exponents using sqrt or cbrt if possible. Always fall back
to expanding via optabs.
2006-11-27 Ira Rosen <irar@il.ibm.com>
PR tree-optimization/22372

View File

@ -2601,8 +2601,13 @@ expand_powi (rtx x, enum machine_mode mode, HOST_WIDE_INT n)
static rtx
expand_builtin_pow (tree exp, rtx target, rtx subtarget)
{
tree arg0, arg1, fn, narg0, narglist;
tree arglist = TREE_OPERAND (exp, 1);
tree arg0, arg1;
tree type = TREE_TYPE (exp);
REAL_VALUE_TYPE cint, c, c2;
HOST_WIDE_INT n;
rtx op, op2;
enum machine_mode mode = TYPE_MODE (type);
if (! validate_arglist (arglist, REAL_TYPE, REAL_TYPE, VOID_TYPE))
return 0;
@ -2610,36 +2615,108 @@ expand_builtin_pow (tree exp, rtx target, rtx subtarget)
arg0 = TREE_VALUE (arglist);
arg1 = TREE_VALUE (TREE_CHAIN (arglist));
if (TREE_CODE (arg1) == REAL_CST
&& ! TREE_CONSTANT_OVERFLOW (arg1))
{
REAL_VALUE_TYPE cint;
REAL_VALUE_TYPE c;
HOST_WIDE_INT n;
if (TREE_CODE (arg1) != REAL_CST
|| TREE_CONSTANT_OVERFLOW (arg1))
return expand_builtin_mathfn_2 (exp, target, subtarget);
c = TREE_REAL_CST (arg1);
n = real_to_integer (&c);
real_from_integer (&cint, VOIDmode, n, n < 0 ? -1 : 0, 0);
if (real_identical (&c, &cint))
/* Handle constant exponents. */
/* For integer valued exponents we can expand to an optimal multiplication
sequence using expand_powi. */
c = TREE_REAL_CST (arg1);
n = real_to_integer (&c);
real_from_integer (&cint, VOIDmode, n, n < 0 ? -1 : 0, 0);
if (real_identical (&c, &cint)
&& ((n >= -1 && n <= 2)
|| (flag_unsafe_math_optimizations
&& !optimize_size
&& powi_cost (n) <= POWI_MAX_MULTS)))
{
op = expand_expr (arg0, subtarget, VOIDmode, 0);
if (n != 1)
{
/* If the exponent is -1, 0, 1 or 2, then expand_powi is exact.
Otherwise, check the number of multiplications required.
Note that pow never sets errno for an integer exponent. */
if ((n >= -1 && n <= 2)
|| (flag_unsafe_math_optimizations
&& ! optimize_size
&& powi_cost (n) <= POWI_MAX_MULTS))
op = force_reg (mode, op);
op = expand_powi (op, mode, n);
}
return op;
}
narg0 = builtin_save_expr (arg0);
narglist = build_tree_list (NULL_TREE, narg0);
/* If the exponent is not integer valued, check if it is half of an integer.
In this case we can expand to sqrt (x) * x**(n/2). */
fn = mathfn_built_in (type, BUILT_IN_SQRT);
if (fn != NULL_TREE)
{
real_arithmetic (&c2, MULT_EXPR, &c, &dconst2);
n = real_to_integer (&c2);
real_from_integer (&cint, VOIDmode, n, n < 0 ? -1 : 0, 0);
if (real_identical (&c2, &cint)
&& ((flag_unsafe_math_optimizations
&& !optimize_size
&& powi_cost (n/2) <= POWI_MAX_MULTS)
|| n == 1))
{
tree call_expr = build_function_call_expr (fn, narglist);
op = expand_builtin (call_expr, NULL_RTX, subtarget, mode, 0);
if (n != 1)
{
enum machine_mode mode = TYPE_MODE (TREE_TYPE (exp));
rtx op = expand_expr (arg0, subtarget, VOIDmode, 0);
op = force_reg (mode, op);
return expand_powi (op, mode, n);
op2 = expand_expr (narg0, subtarget, VOIDmode, 0);
op2 = force_reg (mode, op2);
op2 = expand_powi (op2, mode, abs (n / 2));
op = expand_simple_binop (mode, MULT, op, op2, NULL_RTX,
0, OPTAB_LIB_WIDEN);
/* If the original exponent was negative, reciprocate the
result. */
if (n < 0)
op = expand_binop (mode, sdiv_optab, CONST1_RTX (mode),
op, NULL_RTX, 0, OPTAB_LIB_WIDEN);
}
return op;
}
}
if (! flag_unsafe_math_optimizations)
return NULL_RTX;
/* Try if the exponent is a third of an integer. In this case
we can expand to x**(n/3) * cbrt(x)**(n%3). */
fn = mathfn_built_in (type, BUILT_IN_CBRT);
if (fn != NULL_TREE)
{
real_arithmetic (&c2, MULT_EXPR, &c, &dconst3);
real_round (&c2, mode, &c2);
n = real_to_integer (&c2);
real_from_integer (&cint, VOIDmode, n, n < 0 ? -1 : 0, 0);
real_arithmetic (&c2, RDIV_EXPR, &cint, &dconst3);
real_convert (&c2, mode, &c2);
if (real_identical (&c2, &c)
&& ((!optimize_size
&& flag_unsafe_math_optimizations
&& powi_cost (n/3) <= POWI_MAX_MULTS)
|| n == 1))
{
tree call_expr = build_function_call_expr (fn, narglist);
op = expand_builtin (call_expr, NULL_RTX, subtarget, mode, 0);
if (abs (n) % 3 == 2)
op = expand_simple_binop (mode, MULT, op, op, op,
0, OPTAB_LIB_WIDEN);
if (n != 1)
{
op2 = expand_expr (narg0, subtarget, VOIDmode, 0);
op2 = force_reg (mode, op2);
op2 = expand_powi (op2, mode, abs (n / 3));
op = expand_simple_binop (mode, MULT, op, op2, NULL_RTX,
0, OPTAB_LIB_WIDEN);
/* If the original exponent was negative, reciprocate the
result. */
if (n < 0)
op = expand_binop (mode, sdiv_optab, CONST1_RTX (mode),
op, NULL_RTX, 0, OPTAB_LIB_WIDEN);
}
return op;
}
}
/* Fall back to optab expansion. */
return expand_builtin_mathfn_2 (exp, target, subtarget);
}

View File

@ -1,3 +1,9 @@
2006-11-27 Richard Guenther <rguenther@suse.de>
PR middle-end/25620
* gcc.target/i386/pow-1.c: New testcase.
* gcc.dg/builtins-58.c: Likewise.
2006-11-26 Mark Mitchell <mark@codesourcery.com>
PR c++/29886

View File

@ -0,0 +1,33 @@
/* { dg-do compile } */
/* { dg-options "-O -ffast-math -std=c99" } */
#include "builtins-config.h"
#ifdef HAVE_C99_RUNTIME
double test1 (double x)
{
return __builtin_pow (x, 1./3.);
}
double test2 (double x)
{
return __builtin_pow (x, 4./3.);
}
double test3a (double x)
{
return __builtin_pow (x, 5./3.);
}
double test3b (double x)
{
return __builtin_pow (x, -5./3.);
}
double test4 (double x)
{
return __builtin_pow (x, 7./3.);
}
#endif
/* { dg-final { scan-assembler-not "pow" } } */

View File

@ -0,0 +1,24 @@
/* { dg-do compile } */
/* { dg-options "-O -ffast-math" } */
double test1 (double x)
{
return __builtin_pow (x, 1./2.);
}
double test2 (double x)
{
return __builtin_pow (x, 3./2.);
}
double test3 (double x)
{
return __builtin_pow (x, 5./2.);
}
double test4 (double x)
{
return __builtin_pow (x, -5./2.);
}
/* { dg-final { scan-assembler-not "call" } } */