Fixed floating point mathematic speed degradation (Christian)

Zend/Makefile.am

@@ -17,7 +17,7 @@ libZend_la_SOURCES=\
 	zend_objects_API.c zend_ts_hash.c zend_stream.c \
 	zend_default_classes.c \
 	zend_iterators.c zend_interfaces.c zend_exceptions.c \
-	zend_strtod.c zend_closures.c
+	zend_strtod.c zend_closures.c zend_float.c
 
 libZend_la_LDFLAGS =
 libZend_la_LIBADD = @ZEND_EXTRA_LIBS@

Zend/Zend.dsp

@@ -155,6 +155,10 @@ SOURCE=.\zend_extensions.c
 # End Source File
 # Begin Source File
 
+SOURCE=.\zend_float.c
+# End Source File
+# Begin Source File
+
 SOURCE=.\zend_hash.c
 # End Source File
 # Begin Source File

Zend/zend.c

@@ -549,6 +549,7 @@ static void executor_globals_ctor(zend_executor_globals *executor_globals TSRMLS
 	EG(current_execute_data) = NULL;
 	EG(current_module) = NULL;
 	EG(exit_status) = 0;
+	EG(saved_fpu_cw) = NULL;
 	EG(active) = 0;
 }
 /* }}} */

Zend/zend_execute_API.c

@@ -123,6 +123,8 @@ static int clean_non_persistent_class_full(zend_class_entry **ce TSRMLS_DC) /* {
 
 void init_executor(TSRMLS_D) /* {{{ */
 {
+	zend_init_fpu(TSRMLS_C);
+
 	INIT_ZVAL(EG(uninitialized_zval));
 	/* trick to make uninitialized_zval never be modified, passed by ref, etc. */
 	Z_ADDREF(EG(uninitialized_zval));
@@ -331,6 +333,9 @@ void shutdown_executor(TSRMLS_D) /* {{{ */
 			FREE_HASHTABLE(EG(in_autoload));
 		}
 	} zend_end_try();
+
+	zend_shutdown_fpu(TSRMLS_C);
+
 	EG(active) = 0;
 }
 /* }}} */

Zend/zend_float.h

@@ -21,10 +21,12 @@
 #ifndef ZEND_FLOAT_H
 #define ZEND_FLOAT_H
 
-#define ZEND_FLOAT_DECLARE				XPFPA_DECLARE
+/*
-#define ZEND_FLOAT_ENSURE()				XPFPA_SWITCH_DOUBLE()
+  Define functions for FP initialization and de-initialization.
-#define ZEND_FLOAT_RESTORE()			XPFPA_RESTORE()
+*/
-#define ZEND_FLOAT_RETURN(val)			XPFPA_RETURN_DOUBLE(val)
+extern ZEND_API void zend_init_fpu(TSRMLS_D);
+extern ZEND_API void zend_shutdown_fpu(TSRMLS_D);
+extern ZEND_API void zend_ensure_fpu_mode(TSRMLS_D);
 
 /* Copy of the contents of xpfpa.h (which is under public domain)
    See http://wiki.php.net/rfc/rounding for details.
@@ -45,7 +47,7 @@
    For further details, please visit:
    http://www.christian-seiler.de/projekte/fpmath/
 
-   Version: 20081026 */
+   Version: 20090317 */
 
 /*
  Implementation notes:
@@ -69,11 +71,8 @@
 /* MSVC detection (MSVC people usually don't use autoconf) */
 #ifdef _MSC_VER
 # if _MSC_VER >= 1500
-   /* Disable it, it slowdowns the floating operation more than
-      anything else, by a factor 3 (using Bench.php (mandel and 
-      mandel2 for example)*/
    /* Visual C++ 2008 or higher, supports _controlfp_s */
-   /*#  define HAVE__CONTROLFP_S */
+#  define HAVE__CONTROLFP_S
 # else
    /* Visual C++ (up to 2005), supports _controlfp */
 #  define HAVE__CONTROLFP
@@ -87,6 +86,19 @@
 /* float.h defines _controlfp_s */
 # include <float.h>
 
+# define XPFPA_HAVE_CW 1
+# define XPFPA_CW_DATATYPE \
+            unsigned int
+
+# define XPFPA_STORE_CW(vptr) do { \
+            _controlfp_s((unsigned int *)(vptr), 0, 0); \
+        } while (0)
+
+# define XPFPA_RESTORE_CW(vptr) do { \
+            unsigned int _xpfpa_fpu_cw; \
+            _controlfp_s(&_xpfpa_fpu_cw, *((unsigned int *)(vptr)), _MCW_PC); \
+        } while (0)
+
 # define XPFPA_DECLARE \
             unsigned int _xpfpa_fpu_oldcw, _xpfpa_fpu_cw;
 
@@ -141,6 +153,18 @@
 # define XPFPA_DECLARE \
             unsigned int _xpfpa_fpu_oldcw;
 
+# define XPFPA_HAVE_CW 1
+# define XPFPA_CW_DATATYPE \
+            unsigned int
+
+# define XPFPA_STORE_CW(vptr) do { \
+            *((unsigned int *)(vptr)) = _controlfp(0, 0); \
+        } while (0)
+
+# define XPFPA_RESTORE_CW(vptr) do { \
+            _controlfp(*((unsigned int *)(vptr)), _MCW_PC); \
+        } while (0)
+
 # define XPFPA_SWITCH_DOUBLE() do { \
             _xpfpa_fpu_oldcw = _controlfp(0, 0); \
             _controlfp(_PC_53, _MCW_PC); \
@@ -188,6 +212,18 @@
 # define XPFPA_DECLARE \
             fpu_control_t _xpfpa_fpu_oldcw, _xpfpa_fpu_cw;
 
+# define XPFPA_HAVE_CW 1
+# define XPFPA_CW_DATATYPE \
+            fpu_control_t
+
+# define XPFPA_STORE_CW(vptr) do { \
+            _FPU_GETCW((*((fpu_control_t *)(vptr)))); \
+        } while (0)
+
+# define XPFPA_RESTORE_CW(vptr) do { \
+            _FPU_SETCW((*((fpu_control_t *)(vptr)))); \
+        } while (0)
+
 # define XPFPA_SWITCH_DOUBLE() do { \
             _FPU_GETCW(_xpfpa_fpu_oldcw); \
             _xpfpa_fpu_cw = (_xpfpa_fpu_oldcw & ~_FPU_EXTENDED & ~_FPU_SINGLE) | _FPU_DOUBLE; \
@@ -235,6 +271,18 @@
 # define XPFPA_DECLARE \
             fp_prec_t _xpfpa_fpu_oldprec;
 
+# define XPFPA_HAVE_CW 1
+# define XPFPA_CW_DATATYPE \
+            fp_prec_t
+
+# define XPFPA_STORE_CW(vptr) do { \
+            *((fp_prec_t *)(vptr)) = fpgetprec(); \
+        } while (0)
+
+# define XPFPA_RESTORE_CW(vptr) do { \
+            fpsetprec(*((fp_prec_t *)(vptr))); \
+        } while (0)
+
 # define XPFPA_SWITCH_DOUBLE() do { \
             _xpfpa_fpu_oldprec = fpgetprec(); \
             fpsetprec(FP_PD); \
@@ -298,6 +346,18 @@
 # define XPFPA_DECLARE \
             unsigned int _xpfpa_fpu_oldcw, _xpfpa_fpu_cw;
 
+# define XPFPA_HAVE_CW 1
+# define XPFPA_CW_DATATYPE \
+            unsigned int
+
+# define XPFPA_STORE_CW(vptr) do { \
+            __asm__ __volatile__ ("fnstcw %0" : "=m" (*((unsigned int *)(vptr)))); \
+        } while (0)
+
+# define XPFPA_RESTORE_CW(vptr) do { \
+            __asm__ __volatile__ ("fldcw %0" : : "m" (*((unsigned int *)(vptr)))); \
+        } while (0)
+
 # define XPFPA_SWITCH_DOUBLE() do { \
             __asm__ __volatile__ ("fnstcw %0" : "=m" (*&_xpfpa_fpu_oldcw)); \
             _xpfpa_fpu_cw = (_xpfpa_fpu_oldcw & ~0x100) | 0x200; \
@@ -345,6 +405,10 @@
   generated code will behave as planned.
 */
 # define XPFPA_DECLARE                      /* NOP */
+# define XPFPA_HAVE_CW                      0
+# define XPFPA_CW_DATATYPE                  unsigned int
+# define XPFPA_STORE_CW(variable)           /* NOP */
+# define XPFPA_RESTORE_CW(variable)         /* NOP */
 # define XPFPA_SWITCH_DOUBLE()              /* NOP */
 # define XPFPA_SWITCH_SINGLE()              /* NOP */
 # define XPFPA_SWITCH_DOUBLE_EXTENDED()     /* NOP */

Zend/zend_globals.h

@@ -254,6 +254,8 @@ struct _zend_executor_globals {
 
 	zend_bool active; 
 
+	void *saved_fpu_cw;
+
 	void *reserved[ZEND_MAX_RESERVED_RESOURCES];
 };
 

Zend/zend_operators.c

@@ -30,7 +30,6 @@
 #include "zend_multiply.h"
 #include "zend_strtod.h"
 #include "zend_exceptions.h"
-#include "zend_float.h"
 
 #define LONG_SIGN_MASK (1L << (8*sizeof(long)-1))
 
@@ -742,7 +741,6 @@ ZEND_API void multi_convert_to_string_ex(int argc, ...)
 
 ZEND_API int add_function(zval *result, zval *op1, zval *op2 TSRMLS_DC)
 {
-	ZEND_FLOAT_DECLARE
 	zval op1_copy, op2_copy;
 	int converted = 0;
 
@@ -755,9 +753,7 @@ ZEND_API int add_function(zval *result, zval *op1, zval *op2 TSRMLS_DC)
 				if ((Z_LVAL_P(op1) & LONG_SIGN_MASK) == (Z_LVAL_P(op2) & LONG_SIGN_MASK)
 					&& (Z_LVAL_P(op1) & LONG_SIGN_MASK) != (lval & LONG_SIGN_MASK)) {
 
-					ZEND_FLOAT_ENSURE();
 					ZVAL_DOUBLE(result, (double) Z_LVAL_P(op1) + (double) Z_LVAL_P(op2));
-					ZEND_FLOAT_RESTORE();
 				} else {
 					ZVAL_LONG(result, lval);
 				}
@@ -765,21 +761,15 @@ ZEND_API int add_function(zval *result, zval *op1, zval *op2 TSRMLS_DC)
 			}
 
 			case TYPE_PAIR(IS_LONG, IS_DOUBLE):
-				ZEND_FLOAT_ENSURE();
 				ZVAL_DOUBLE(result, ((double)Z_LVAL_P(op1)) + Z_DVAL_P(op2));
-				ZEND_FLOAT_RESTORE();
 				return SUCCESS;
 
 			case TYPE_PAIR(IS_DOUBLE, IS_LONG):
-				ZEND_FLOAT_ENSURE();
 				ZVAL_DOUBLE(result, Z_DVAL_P(op1) + ((double)Z_LVAL_P(op2)));
-				ZEND_FLOAT_RESTORE();
 				return SUCCESS;
 
 			case TYPE_PAIR(IS_DOUBLE, IS_DOUBLE):
-				ZEND_FLOAT_ENSURE();
 				ZVAL_DOUBLE(result, Z_DVAL_P(op1) + Z_DVAL_P(op2));
-				ZEND_FLOAT_RESTORE();
 				return SUCCESS;
 
 			case TYPE_PAIR(IS_ARRAY, IS_ARRAY): {
@@ -813,7 +803,6 @@ ZEND_API int add_function(zval *result, zval *op1, zval *op2 TSRMLS_DC)
 
 ZEND_API int sub_function(zval *result, zval *op1, zval *op2 TSRMLS_DC)
 {
-	ZEND_FLOAT_DECLARE
 	zval op1_copy, op2_copy;
 	int converted = 0;
 
@@ -826,9 +815,7 @@ ZEND_API int sub_function(zval *result, zval *op1, zval *op2 TSRMLS_DC)
 				if ((Z_LVAL_P(op1) & LONG_SIGN_MASK) != (Z_LVAL_P(op2) & LONG_SIGN_MASK)
 					&& (Z_LVAL_P(op1) & LONG_SIGN_MASK) != (lval & LONG_SIGN_MASK)) {
 
-					ZEND_FLOAT_ENSURE();
 					ZVAL_DOUBLE(result, (double) Z_LVAL_P(op1) - (double) Z_LVAL_P(op2));
-					ZEND_FLOAT_RESTORE();
 				} else {
 					ZVAL_LONG(result, lval);
 				}
@@ -836,21 +823,15 @@ ZEND_API int sub_function(zval *result, zval *op1, zval *op2 TSRMLS_DC)
 
 			}
 			case TYPE_PAIR(IS_LONG, IS_DOUBLE):
-				ZEND_FLOAT_ENSURE();
 				ZVAL_DOUBLE(result, ((double)Z_LVAL_P(op1)) - Z_DVAL_P(op2));
-				ZEND_FLOAT_RESTORE();
 				return SUCCESS;
 
 			case TYPE_PAIR(IS_DOUBLE, IS_LONG):
-				ZEND_FLOAT_ENSURE();
 				ZVAL_DOUBLE(result, Z_DVAL_P(op1) - ((double)Z_LVAL_P(op2)));
-				ZEND_FLOAT_RESTORE();
 				return SUCCESS;
 
 			case TYPE_PAIR(IS_DOUBLE, IS_DOUBLE):
-				ZEND_FLOAT_ENSURE();
 				ZVAL_DOUBLE(result, Z_DVAL_P(op1) - Z_DVAL_P(op2));
-				ZEND_FLOAT_RESTORE();
 				return SUCCESS;
 
 			default:
@@ -869,7 +850,6 @@ ZEND_API int sub_function(zval *result, zval *op1, zval *op2 TSRMLS_DC)
 
 ZEND_API int mul_function(zval *result, zval *op1, zval *op2 TSRMLS_DC)
 {
-	ZEND_FLOAT_DECLARE
 	zval op1_copy, op2_copy;
 	int converted = 0;
 
@@ -878,29 +858,21 @@ ZEND_API int mul_function(zval *result, zval *op1, zval *op2 TSRMLS_DC)
 			case TYPE_PAIR(IS_LONG, IS_LONG): {
 				long overflow;
 
-				ZEND_FLOAT_ENSURE();
 				ZEND_SIGNED_MULTIPLY_LONG(Z_LVAL_P(op1),Z_LVAL_P(op2), Z_LVAL_P(result),Z_DVAL_P(result),overflow);
-				ZEND_FLOAT_RESTORE();
 				Z_TYPE_P(result) = overflow ? IS_DOUBLE : IS_LONG;
 				return SUCCESS;
 
 			}
 			case TYPE_PAIR(IS_LONG, IS_DOUBLE):
-				ZEND_FLOAT_ENSURE();
 				ZVAL_DOUBLE(result, ((double)Z_LVAL_P(op1)) * Z_DVAL_P(op2));
-				ZEND_FLOAT_RESTORE();
 				return SUCCESS;
 
 			case TYPE_PAIR(IS_DOUBLE, IS_LONG):
-				ZEND_FLOAT_ENSURE();
 				ZVAL_DOUBLE(result, Z_DVAL_P(op1) * ((double)Z_LVAL_P(op2)));
-				ZEND_FLOAT_RESTORE();
 				return SUCCESS;
 
 			case TYPE_PAIR(IS_DOUBLE, IS_DOUBLE):
-				ZEND_FLOAT_ENSURE();
 				ZVAL_DOUBLE(result, Z_DVAL_P(op1) * Z_DVAL_P(op2));
-				ZEND_FLOAT_RESTORE();
 				return SUCCESS;
 
 			default:
@@ -918,7 +890,6 @@ ZEND_API int mul_function(zval *result, zval *op1, zval *op2 TSRMLS_DC)
 
 ZEND_API int div_function(zval *result, zval *op1, zval *op2 TSRMLS_DC)
 {
-	ZEND_FLOAT_DECLARE
 	zval op1_copy, op2_copy;
 	int converted = 0;
 
@@ -931,17 +902,13 @@ ZEND_API int div_function(zval *result, zval *op1, zval *op2 TSRMLS_DC)
 					return FAILURE;			/* division by zero */
 				} else if (Z_LVAL_P(op2) == -1 && Z_LVAL_P(op1) == LONG_MIN) {
 					/* Prevent overflow error/crash */
-					ZEND_FLOAT_ENSURE();
 					ZVAL_DOUBLE(result, (double) LONG_MIN / -1);
-					ZEND_FLOAT_RESTORE();
 					return SUCCESS;
 				}
 				if (Z_LVAL_P(op1) % Z_LVAL_P(op2) == 0) { /* integer */
 					ZVAL_LONG(result, Z_LVAL_P(op1) / Z_LVAL_P(op2));
 				} else {
-					ZEND_FLOAT_ENSURE();
 					ZVAL_DOUBLE(result, ((double) Z_LVAL_P(op1)) / Z_LVAL_P(op2));
-					ZEND_FLOAT_RESTORE();
 				}
 				return SUCCESS;
 
@@ -951,9 +918,7 @@ ZEND_API int div_function(zval *result, zval *op1, zval *op2 TSRMLS_DC)
 					ZVAL_BOOL(result, 0);
 					return FAILURE;			/* division by zero */
 				}
-				ZEND_FLOAT_ENSURE();
 				ZVAL_DOUBLE(result, Z_DVAL_P(op1) / (double)Z_LVAL_P(op2));
-				ZEND_FLOAT_RESTORE();
 				return SUCCESS;
 
 			case TYPE_PAIR(IS_LONG, IS_DOUBLE):
@@ -962,9 +927,7 @@ ZEND_API int div_function(zval *result, zval *op1, zval *op2 TSRMLS_DC)
 					ZVAL_BOOL(result, 0);
 					return FAILURE;			/* division by zero */
 				}
-				ZEND_FLOAT_ENSURE();
 				ZVAL_DOUBLE(result, (double)Z_LVAL_P(op1) / Z_DVAL_P(op2));
-				ZEND_FLOAT_RESTORE();
 				return SUCCESS;
 
 			case TYPE_PAIR(IS_DOUBLE, IS_DOUBLE):
@@ -973,9 +936,7 @@ ZEND_API int div_function(zval *result, zval *op1, zval *op2 TSRMLS_DC)
 					ZVAL_BOOL(result, 0);
 					return FAILURE;			/* division by zero */
 				}
-				ZEND_FLOAT_ENSURE();
 				ZVAL_DOUBLE(result, Z_DVAL_P(op1) / Z_DVAL_P(op2));
-				ZEND_FLOAT_RESTORE();
 				return SUCCESS;
 
 			default:

Zend/zend_strtod.c

@@ -93,7 +93,6 @@
 
 #include <zend_operators.h>
 #include <zend_strtod.h>
-#include <zend_float.h>
 
 #ifdef ZTS
 #include <TSRM.h>
@@ -2033,7 +2032,6 @@ ret1:
 
 ZEND_API double zend_strtod (CONST char *s00, char **se)
 {
-	ZEND_FLOAT_DECLARE
 	int bb2, bb5, bbe, bd2, bd5, bbbits, bs2, c, dsign,
 		e, e1, esign, i, j, k, nd, nd0, nf, nz, nz0, sign;
 	CONST char *s, *s0, *s1;
@@ -2046,8 +2044,6 @@ ZEND_API double zend_strtod (CONST char *s00, char **se)
 
 	CONST char decimal_point = '.';
 
-	ZEND_FLOAT_ENSURE();
-
 	sign = nz0 = nz = 0;
 	value(rv) = 0.;
 
@@ -2578,7 +2574,7 @@ ret:
 	}
 	_THREAD_PRIVATE_MUTEX_UNLOCK(pow5mult_mutex);
 
-	ZEND_FLOAT_RETURN(result);
+	return result;
 }
 
 ZEND_API double zend_hex_strtod(const char *str, char **endptr)

configure.in

@@ -1452,7 +1452,7 @@ PHP_ADD_SOURCES(Zend, \
     zend_list.c zend_indent.c zend_builtin_functions.c zend_sprintf.c \
     zend_ini.c zend_qsort.c zend_multibyte.c zend_ts_hash.c zend_stream.c \
     zend_iterators.c zend_interfaces.c zend_exceptions.c zend_strtod.c zend_gc.c \
-    zend_closures.c)
+    zend_closures.c zend_float.c)
 
 if test -r "$abs_srcdir/Zend/zend_objects.c"; then
   PHP_ADD_SOURCES(Zend, zend_objects.c zend_object_handlers.c zend_objects_API.c zend_default_classes.c)

ext/standard/math.c

@@ -24,7 +24,6 @@
 #include "php.h"
 #include "php_math.h"
 #include "zend_multiply.h"
-#include "zend_float.h"
 
 #include <math.h>
 #include <float.h>
@@ -94,10 +93,8 @@ static inline double php_intpow10(int power) {
 /* {{{ php_round_helper
        Actually performs the rounding of a value to integer in a certain mode */
 static inline double php_round_helper(double value, int mode) {
-	ZEND_FLOAT_DECLARE
 	double tmp_value;
 
-	ZEND_FLOAT_ENSURE();
 	if (value >= 0.0) {
 		tmp_value = floor(value + 0.5);
 		if ((mode == PHP_ROUND_HALF_DOWN && value == (-0.5 + tmp_value)) ||
@@ -116,7 +113,7 @@ static inline double php_round_helper(double value, int mode) {
 		}
 	}
 
-	ZEND_FLOAT_RETURN(tmp_value);
+	return tmp_value;
 }
 /* }}} */
 
@@ -126,13 +123,10 @@ static inline double php_round_helper(double value, int mode) {
  * mode. For the specifics of the algorithm, see http://wiki.php.net/rfc/rounding
  */
 PHPAPI double _php_math_round(double value, int places, int mode) {
-	ZEND_FLOAT_DECLARE
 	double f1, f2;
 	double tmp_value;
 	int precision_places;
 
-	ZEND_FLOAT_ENSURE();
-
 	precision_places = 14 - php_intlog10abs(value);
 
 	f1 = php_intpow10(abs(places));
@@ -163,7 +157,7 @@ PHPAPI double _php_math_round(double value, int places, int mode) {
 		}
 		/* This value is beyond our precision, so rounding it is pointless */
 		if (fabs(tmp_value) >= 1e15) {
-			ZEND_FLOAT_RETURN(value);
+			return value;
 		}
 	}
 
@@ -196,7 +190,7 @@ PHPAPI double _php_math_round(double value, int places, int mode) {
 		}
 	}
 
-	ZEND_FLOAT_RETURN(tmp_value);
+	return tmp_value;
 }
 /* }}} */
 

win32/build/config.w32

@@ -311,7 +311,8 @@ ADD_SOURCES("Zend", "zend_language_parser.c zend_language_scanner.c \
 	zend_sprintf.c zend_ini.c zend_qsort.c zend_multibyte.c zend_ts_hash.c \
 	zend_stream.c zend_iterators.c zend_interfaces.c zend_objects.c \
 	zend_object_handlers.c zend_objects_API.c \
-	zend_default_classes.c zend_execute.c zend_strtod.c zend_gc.c zend_closures.c");
+	zend_default_classes.c zend_execute.c zend_strtod.c zend_gc.c zend_closures.c \
+	zend_float.c");
 
 ADD_SOURCES("main", "main.c snprintf.c spprintf.c safe_mode.c getopt.c fopen_wrappers.c \
 	php_scandir.c php_ini.c SAPI.c rfc1867.c php_content_types.c strlcpy.c \