1 /* numeric.c 2 * 3 * Copyright (C) 1993, 1994, 1995, 1996, 1997, 1998, 1999, 2000, 2001, 4 * 2002, 2003, 2004, 2005, 2006, 2007, 2008 by Larry Wall and others 5 * 6 * You may distribute under the terms of either the GNU General Public 7 * License or the Artistic License, as specified in the README file. 8 * 9 */ 10 11 /* 12 * "That only makes eleven (plus one mislaid) and not fourteen, 13 * unless wizards count differently to other people." --Beorn 14 * 15 * [p.115 of _The Hobbit_: "Queer Lodgings"] 16 */ 17 18 /* 19 =head1 Numeric functions 20 21 =cut 22 23 This file contains all the stuff needed by perl for manipulating numeric 24 values, including such things as replacements for the OS's atof() function 25 26 */ 27 28 #include "EXTERN.h" 29 #define PERL_IN_NUMERIC_C 30 #include "perl.h" 31 32 U32 33 Perl_cast_ulong(NV f) 34 { 35 if (f < 0.0) 36 return f < I32_MIN ? (U32) I32_MIN : (U32)(I32) f; 37 if (f < U32_MAX_P1) { 38 #if CASTFLAGS & 2 39 if (f < U32_MAX_P1_HALF) 40 return (U32) f; 41 f -= U32_MAX_P1_HALF; 42 return ((U32) f) | (1 + (U32_MAX >> 1)); 43 #else 44 return (U32) f; 45 #endif 46 } 47 return f > 0 ? U32_MAX : 0 /* NaN */; 48 } 49 50 I32 51 Perl_cast_i32(NV f) 52 { 53 if (f < I32_MAX_P1) 54 return f < I32_MIN ? I32_MIN : (I32) f; 55 if (f < U32_MAX_P1) { 56 #if CASTFLAGS & 2 57 if (f < U32_MAX_P1_HALF) 58 return (I32)(U32) f; 59 f -= U32_MAX_P1_HALF; 60 return (I32)(((U32) f) | (1 + (U32_MAX >> 1))); 61 #else 62 return (I32)(U32) f; 63 #endif 64 } 65 return f > 0 ? (I32)U32_MAX : 0 /* NaN */; 66 } 67 68 IV 69 Perl_cast_iv(NV f) 70 { 71 if (f < IV_MAX_P1) 72 return f < IV_MIN ? IV_MIN : (IV) f; 73 if (f < UV_MAX_P1) { 74 #if CASTFLAGS & 2 75 /* For future flexibility allowing for sizeof(UV) >= sizeof(IV) */ 76 if (f < UV_MAX_P1_HALF) 77 return (IV)(UV) f; 78 f -= UV_MAX_P1_HALF; 79 return (IV)(((UV) f) | (1 + (UV_MAX >> 1))); 80 #else 81 return (IV)(UV) f; 82 #endif 83 } 84 return f > 0 ? (IV)UV_MAX : 0 /* NaN */; 85 } 86 87 UV 88 Perl_cast_uv(NV f) 89 { 90 if (f < 0.0) 91 return f < IV_MIN ? (UV) IV_MIN : (UV)(IV) f; 92 if (f < UV_MAX_P1) { 93 #if CASTFLAGS & 2 94 if (f < UV_MAX_P1_HALF) 95 return (UV) f; 96 f -= UV_MAX_P1_HALF; 97 return ((UV) f) | (1 + (UV_MAX >> 1)); 98 #else 99 return (UV) f; 100 #endif 101 } 102 return f > 0 ? UV_MAX : 0 /* NaN */; 103 } 104 105 /* 106 =for apidoc grok_bin 107 108 converts a string representing a binary number to numeric form. 109 110 On entry C<start> and C<*len> give the string to scan, C<*flags> gives 111 conversion flags, and C<result> should be C<NULL> or a pointer to an NV. 112 The scan stops at the end of the string, or the first invalid character. 113 Unless C<PERL_SCAN_SILENT_ILLDIGIT> is set in C<*flags>, encountering an 114 invalid character will also trigger a warning. 115 On return C<*len> is set to the length of the scanned string, 116 and C<*flags> gives output flags. 117 118 If the value is <= C<UV_MAX> it is returned as a UV, the output flags are clear, 119 and nothing is written to C<*result>. If the value is > C<UV_MAX>, C<grok_bin> 120 returns C<UV_MAX>, sets C<PERL_SCAN_GREATER_THAN_UV_MAX> in the output flags, 121 and writes the value to C<*result> (or the value is discarded if C<result> 122 is NULL). 123 124 The binary number may optionally be prefixed with C<"0b"> or C<"b"> unless 125 C<PERL_SCAN_DISALLOW_PREFIX> is set in C<*flags> on entry. If 126 C<PERL_SCAN_ALLOW_UNDERSCORES> is set in C<*flags> then the binary 127 number may use C<"_"> characters to separate digits. 128 129 =cut 130 131 Not documented yet because experimental is C<PERL_SCAN_SILENT_NON_PORTABLE 132 which suppresses any message for non-portable numbers that are still valid 133 on this platform. 134 */ 135 136 UV 137 Perl_grok_bin(pTHX_ const char *start, STRLEN *len_p, I32 *flags, NV *result) 138 { 139 const char *s = start; 140 STRLEN len = *len_p; 141 UV value = 0; 142 NV value_nv = 0; 143 144 const UV max_div_2 = UV_MAX / 2; 145 const bool allow_underscores = cBOOL(*flags & PERL_SCAN_ALLOW_UNDERSCORES); 146 bool overflowed = FALSE; 147 char bit; 148 149 PERL_ARGS_ASSERT_GROK_BIN; 150 151 if (!(*flags & PERL_SCAN_DISALLOW_PREFIX)) { 152 /* strip off leading b or 0b. 153 for compatibility silently suffer "b" and "0b" as valid binary 154 numbers. */ 155 if (len >= 1) { 156 if (isALPHA_FOLD_EQ(s[0], 'b')) { 157 s++; 158 len--; 159 } 160 else if (len >= 2 && s[0] == '0' && (isALPHA_FOLD_EQ(s[1], 'b'))) { 161 s+=2; 162 len-=2; 163 } 164 } 165 } 166 167 for (; len-- && (bit = *s); s++) { 168 if (bit == '0' || bit == '1') { 169 /* Write it in this wonky order with a goto to attempt to get the 170 compiler to make the common case integer-only loop pretty tight. 171 With gcc seems to be much straighter code than old scan_bin. */ 172 redo: 173 if (!overflowed) { 174 if (value <= max_div_2) { 175 value = (value << 1) | (bit - '0'); 176 continue; 177 } 178 /* Bah. We're just overflowed. */ 179 /* diag_listed_as: Integer overflow in %s number */ 180 Perl_ck_warner_d(aTHX_ packWARN(WARN_OVERFLOW), 181 "Integer overflow in binary number"); 182 overflowed = TRUE; 183 value_nv = (NV) value; 184 } 185 value_nv *= 2.0; 186 /* If an NV has not enough bits in its mantissa to 187 * represent a UV this summing of small low-order numbers 188 * is a waste of time (because the NV cannot preserve 189 * the low-order bits anyway): we could just remember when 190 * did we overflow and in the end just multiply value_nv by the 191 * right amount. */ 192 value_nv += (NV)(bit - '0'); 193 continue; 194 } 195 if (bit == '_' && len && allow_underscores && (bit = s[1]) 196 && (bit == '0' || bit == '1')) 197 { 198 --len; 199 ++s; 200 goto redo; 201 } 202 if (!(*flags & PERL_SCAN_SILENT_ILLDIGIT)) 203 Perl_ck_warner(aTHX_ packWARN(WARN_DIGIT), 204 "Illegal binary digit '%c' ignored", *s); 205 break; 206 } 207 208 if ( ( overflowed && value_nv > 4294967295.0) 209 #if UVSIZE > 4 210 || (!overflowed && value > 0xffffffff 211 && ! (*flags & PERL_SCAN_SILENT_NON_PORTABLE)) 212 #endif 213 ) { 214 Perl_ck_warner(aTHX_ packWARN(WARN_PORTABLE), 215 "Binary number > 0b11111111111111111111111111111111 non-portable"); 216 } 217 *len_p = s - start; 218 if (!overflowed) { 219 *flags = 0; 220 return value; 221 } 222 *flags = PERL_SCAN_GREATER_THAN_UV_MAX; 223 if (result) 224 *result = value_nv; 225 return UV_MAX; 226 } 227 228 /* 229 =for apidoc grok_hex 230 231 converts a string representing a hex number to numeric form. 232 233 On entry C<start> and C<*len_p> give the string to scan, C<*flags> gives 234 conversion flags, and C<result> should be C<NULL> or a pointer to an NV. 235 The scan stops at the end of the string, or the first invalid character. 236 Unless C<PERL_SCAN_SILENT_ILLDIGIT> is set in C<*flags>, encountering an 237 invalid character will also trigger a warning. 238 On return C<*len> is set to the length of the scanned string, 239 and C<*flags> gives output flags. 240 241 If the value is <= C<UV_MAX> it is returned as a UV, the output flags are clear, 242 and nothing is written to C<*result>. If the value is > C<UV_MAX>, C<grok_hex> 243 returns C<UV_MAX>, sets C<PERL_SCAN_GREATER_THAN_UV_MAX> in the output flags, 244 and writes the value to C<*result> (or the value is discarded if C<result> 245 is C<NULL>). 246 247 The hex number may optionally be prefixed with C<"0x"> or C<"x"> unless 248 C<PERL_SCAN_DISALLOW_PREFIX> is set in C<*flags> on entry. If 249 C<PERL_SCAN_ALLOW_UNDERSCORES> is set in C<*flags> then the hex 250 number may use C<"_"> characters to separate digits. 251 252 =cut 253 254 Not documented yet because experimental is C<PERL_SCAN_SILENT_NON_PORTABLE 255 which suppresses any message for non-portable numbers, but which are valid 256 on this platform. 257 */ 258 259 UV 260 Perl_grok_hex(pTHX_ const char *start, STRLEN *len_p, I32 *flags, NV *result) 261 { 262 const char *s = start; 263 STRLEN len = *len_p; 264 UV value = 0; 265 NV value_nv = 0; 266 const UV max_div_16 = UV_MAX / 16; 267 const bool allow_underscores = cBOOL(*flags & PERL_SCAN_ALLOW_UNDERSCORES); 268 bool overflowed = FALSE; 269 270 PERL_ARGS_ASSERT_GROK_HEX; 271 272 if (!(*flags & PERL_SCAN_DISALLOW_PREFIX)) { 273 /* strip off leading x or 0x. 274 for compatibility silently suffer "x" and "0x" as valid hex numbers. 275 */ 276 if (len >= 1) { 277 if (isALPHA_FOLD_EQ(s[0], 'x')) { 278 s++; 279 len--; 280 } 281 else if (len >= 2 && s[0] == '0' && (isALPHA_FOLD_EQ(s[1], 'x'))) { 282 s+=2; 283 len-=2; 284 } 285 } 286 } 287 288 for (; len-- && *s; s++) { 289 if (isXDIGIT(*s)) { 290 /* Write it in this wonky order with a goto to attempt to get the 291 compiler to make the common case integer-only loop pretty tight. 292 With gcc seems to be much straighter code than old scan_hex. */ 293 redo: 294 if (!overflowed) { 295 if (value <= max_div_16) { 296 value = (value << 4) | XDIGIT_VALUE(*s); 297 continue; 298 } 299 /* Bah. We're just overflowed. */ 300 /* diag_listed_as: Integer overflow in %s number */ 301 Perl_ck_warner_d(aTHX_ packWARN(WARN_OVERFLOW), 302 "Integer overflow in hexadecimal number"); 303 overflowed = TRUE; 304 value_nv = (NV) value; 305 } 306 value_nv *= 16.0; 307 /* If an NV has not enough bits in its mantissa to 308 * represent a UV this summing of small low-order numbers 309 * is a waste of time (because the NV cannot preserve 310 * the low-order bits anyway): we could just remember when 311 * did we overflow and in the end just multiply value_nv by the 312 * right amount of 16-tuples. */ 313 value_nv += (NV) XDIGIT_VALUE(*s); 314 continue; 315 } 316 if (*s == '_' && len && allow_underscores && s[1] 317 && isXDIGIT(s[1])) 318 { 319 --len; 320 ++s; 321 goto redo; 322 } 323 if (!(*flags & PERL_SCAN_SILENT_ILLDIGIT)) 324 Perl_ck_warner(aTHX_ packWARN(WARN_DIGIT), 325 "Illegal hexadecimal digit '%c' ignored", *s); 326 break; 327 } 328 329 if ( ( overflowed && value_nv > 4294967295.0) 330 #if UVSIZE > 4 331 || (!overflowed && value > 0xffffffff 332 && ! (*flags & PERL_SCAN_SILENT_NON_PORTABLE)) 333 #endif 334 ) { 335 Perl_ck_warner(aTHX_ packWARN(WARN_PORTABLE), 336 "Hexadecimal number > 0xffffffff non-portable"); 337 } 338 *len_p = s - start; 339 if (!overflowed) { 340 *flags = 0; 341 return value; 342 } 343 *flags = PERL_SCAN_GREATER_THAN_UV_MAX; 344 if (result) 345 *result = value_nv; 346 return UV_MAX; 347 } 348 349 /* 350 =for apidoc grok_oct 351 352 converts a string representing an octal number to numeric form. 353 354 On entry C<start> and C<*len> give the string to scan, C<*flags> gives 355 conversion flags, and C<result> should be C<NULL> or a pointer to an NV. 356 The scan stops at the end of the string, or the first invalid character. 357 Unless C<PERL_SCAN_SILENT_ILLDIGIT> is set in C<*flags>, encountering an 358 8 or 9 will also trigger a warning. 359 On return C<*len> is set to the length of the scanned string, 360 and C<*flags> gives output flags. 361 362 If the value is <= C<UV_MAX> it is returned as a UV, the output flags are clear, 363 and nothing is written to C<*result>. If the value is > C<UV_MAX>, C<grok_oct> 364 returns C<UV_MAX>, sets C<PERL_SCAN_GREATER_THAN_UV_MAX> in the output flags, 365 and writes the value to C<*result> (or the value is discarded if C<result> 366 is C<NULL>). 367 368 If C<PERL_SCAN_ALLOW_UNDERSCORES> is set in C<*flags> then the octal 369 number may use C<"_"> characters to separate digits. 370 371 =cut 372 373 Not documented yet because experimental is C<PERL_SCAN_SILENT_NON_PORTABLE> 374 which suppresses any message for non-portable numbers, but which are valid 375 on this platform. 376 */ 377 378 UV 379 Perl_grok_oct(pTHX_ const char *start, STRLEN *len_p, I32 *flags, NV *result) 380 { 381 const char *s = start; 382 STRLEN len = *len_p; 383 UV value = 0; 384 NV value_nv = 0; 385 const UV max_div_8 = UV_MAX / 8; 386 const bool allow_underscores = cBOOL(*flags & PERL_SCAN_ALLOW_UNDERSCORES); 387 bool overflowed = FALSE; 388 389 PERL_ARGS_ASSERT_GROK_OCT; 390 391 for (; len-- && *s; s++) { 392 if (isOCTAL(*s)) { 393 /* Write it in this wonky order with a goto to attempt to get the 394 compiler to make the common case integer-only loop pretty tight. 395 */ 396 redo: 397 if (!overflowed) { 398 if (value <= max_div_8) { 399 value = (value << 3) | OCTAL_VALUE(*s); 400 continue; 401 } 402 /* Bah. We're just overflowed. */ 403 /* diag_listed_as: Integer overflow in %s number */ 404 Perl_ck_warner_d(aTHX_ packWARN(WARN_OVERFLOW), 405 "Integer overflow in octal number"); 406 overflowed = TRUE; 407 value_nv = (NV) value; 408 } 409 value_nv *= 8.0; 410 /* If an NV has not enough bits in its mantissa to 411 * represent a UV this summing of small low-order numbers 412 * is a waste of time (because the NV cannot preserve 413 * the low-order bits anyway): we could just remember when 414 * did we overflow and in the end just multiply value_nv by the 415 * right amount of 8-tuples. */ 416 value_nv += (NV) OCTAL_VALUE(*s); 417 continue; 418 } 419 if (*s == '_' && len && allow_underscores && isOCTAL(s[1])) { 420 --len; 421 ++s; 422 goto redo; 423 } 424 /* Allow \octal to work the DWIM way (that is, stop scanning 425 * as soon as non-octal characters are seen, complain only if 426 * someone seems to want to use the digits eight and nine. Since we 427 * know it is not octal, then if isDIGIT, must be an 8 or 9). */ 428 if (isDIGIT(*s)) { 429 if (!(*flags & PERL_SCAN_SILENT_ILLDIGIT)) 430 Perl_ck_warner(aTHX_ packWARN(WARN_DIGIT), 431 "Illegal octal digit '%c' ignored", *s); 432 } 433 break; 434 } 435 436 if ( ( overflowed && value_nv > 4294967295.0) 437 #if UVSIZE > 4 438 || (!overflowed && value > 0xffffffff 439 && ! (*flags & PERL_SCAN_SILENT_NON_PORTABLE)) 440 #endif 441 ) { 442 Perl_ck_warner(aTHX_ packWARN(WARN_PORTABLE), 443 "Octal number > 037777777777 non-portable"); 444 } 445 *len_p = s - start; 446 if (!overflowed) { 447 *flags = 0; 448 return value; 449 } 450 *flags = PERL_SCAN_GREATER_THAN_UV_MAX; 451 if (result) 452 *result = value_nv; 453 return UV_MAX; 454 } 455 456 /* 457 =for apidoc scan_bin 458 459 For backwards compatibility. Use C<grok_bin> instead. 460 461 =for apidoc scan_hex 462 463 For backwards compatibility. Use C<grok_hex> instead. 464 465 =for apidoc scan_oct 466 467 For backwards compatibility. Use C<grok_oct> instead. 468 469 =cut 470 */ 471 472 NV 473 Perl_scan_bin(pTHX_ const char *start, STRLEN len, STRLEN *retlen) 474 { 475 NV rnv; 476 I32 flags = *retlen ? PERL_SCAN_ALLOW_UNDERSCORES : 0; 477 const UV ruv = grok_bin (start, &len, &flags, &rnv); 478 479 PERL_ARGS_ASSERT_SCAN_BIN; 480 481 *retlen = len; 482 return (flags & PERL_SCAN_GREATER_THAN_UV_MAX) ? rnv : (NV)ruv; 483 } 484 485 NV 486 Perl_scan_oct(pTHX_ const char *start, STRLEN len, STRLEN *retlen) 487 { 488 NV rnv; 489 I32 flags = *retlen ? PERL_SCAN_ALLOW_UNDERSCORES : 0; 490 const UV ruv = grok_oct (start, &len, &flags, &rnv); 491 492 PERL_ARGS_ASSERT_SCAN_OCT; 493 494 *retlen = len; 495 return (flags & PERL_SCAN_GREATER_THAN_UV_MAX) ? rnv : (NV)ruv; 496 } 497 498 NV 499 Perl_scan_hex(pTHX_ const char *start, STRLEN len, STRLEN *retlen) 500 { 501 NV rnv; 502 I32 flags = *retlen ? PERL_SCAN_ALLOW_UNDERSCORES : 0; 503 const UV ruv = grok_hex (start, &len, &flags, &rnv); 504 505 PERL_ARGS_ASSERT_SCAN_HEX; 506 507 *retlen = len; 508 return (flags & PERL_SCAN_GREATER_THAN_UV_MAX) ? rnv : (NV)ruv; 509 } 510 511 /* 512 =for apidoc grok_numeric_radix 513 514 Scan and skip for a numeric decimal separator (radix). 515 516 =cut 517 */ 518 bool 519 Perl_grok_numeric_radix(pTHX_ const char **sp, const char *send) 520 { 521 #ifdef USE_LOCALE_NUMERIC 522 PERL_ARGS_ASSERT_GROK_NUMERIC_RADIX; 523 524 if (IN_LC(LC_NUMERIC)) { 525 DECLARATION_FOR_LC_NUMERIC_MANIPULATION; 526 STORE_LC_NUMERIC_SET_TO_NEEDED(); 527 if (PL_numeric_radix_sv) { 528 STRLEN len; 529 const char * const radix = SvPV(PL_numeric_radix_sv, len); 530 if (*sp + len <= send && memEQ(*sp, radix, len)) { 531 *sp += len; 532 RESTORE_LC_NUMERIC(); 533 return TRUE; 534 } 535 } 536 RESTORE_LC_NUMERIC(); 537 } 538 /* always try "." if numeric radix didn't match because 539 * we may have data from different locales mixed */ 540 #endif 541 542 PERL_ARGS_ASSERT_GROK_NUMERIC_RADIX; 543 544 if (*sp < send && **sp == '.') { 545 ++*sp; 546 return TRUE; 547 } 548 return FALSE; 549 } 550 551 /* 552 =for apidoc grok_infnan 553 554 Helper for C<grok_number()>, accepts various ways of spelling "infinity" 555 or "not a number", and returns one of the following flag combinations: 556 557 IS_NUMBER_INFINITE 558 IS_NUMBER_NAN 559 IS_NUMBER_INFINITE | IS_NUMBER_NEG 560 IS_NUMBER_NAN | IS_NUMBER_NEG 561 0 562 563 possibly |-ed with C<IS_NUMBER_TRAILING>. 564 565 If an infinity or a not-a-number is recognized, C<*sp> will point to 566 one byte past the end of the recognized string. If the recognition fails, 567 zero is returned, and C<*sp> will not move. 568 569 =cut 570 */ 571 572 int 573 Perl_grok_infnan(pTHX_ const char** sp, const char* send) 574 { 575 const char* s = *sp; 576 int flags = 0; 577 bool odh = FALSE; /* one-dot-hash: 1.#INF */ 578 579 PERL_ARGS_ASSERT_GROK_INFNAN; 580 581 if (*s == '+') { 582 s++; if (s == send) return 0; 583 } 584 else if (*s == '-') { 585 flags |= IS_NUMBER_NEG; /* Yes, -NaN happens. Incorrect but happens. */ 586 s++; if (s == send) return 0; 587 } 588 589 if (*s == '1') { 590 /* Visual C: 1.#SNAN, -1.#QNAN, 1#INF, 1.#IND (maybe also 1.#NAN) 591 * Let's keep the dot optional. */ 592 s++; if (s == send) return 0; 593 if (*s == '.') { 594 s++; if (s == send) return 0; 595 } 596 if (*s == '#') { 597 s++; if (s == send) return 0; 598 } else 599 return 0; 600 odh = TRUE; 601 } 602 603 if (isALPHA_FOLD_EQ(*s, 'I')) { 604 /* INF or IND (1.#IND is "indeterminate", a certain type of NAN) */ 605 606 s++; if (s == send || isALPHA_FOLD_NE(*s, 'N')) return 0; 607 s++; if (s == send) return 0; 608 if (isALPHA_FOLD_EQ(*s, 'F')) { 609 s++; 610 if (s < send && (isALPHA_FOLD_EQ(*s, 'I'))) { 611 int fail = 612 flags | IS_NUMBER_INFINITY | IS_NUMBER_NOT_INT | IS_NUMBER_TRAILING; 613 s++; if (s == send || isALPHA_FOLD_NE(*s, 'N')) return fail; 614 s++; if (s == send || isALPHA_FOLD_NE(*s, 'I')) return fail; 615 s++; if (s == send || isALPHA_FOLD_NE(*s, 'T')) return fail; 616 s++; if (s == send || isALPHA_FOLD_NE(*s, 'Y')) return fail; 617 s++; 618 } else if (odh) { 619 while (*s == '0') { /* 1.#INF00 */ 620 s++; 621 } 622 } 623 while (s < send && isSPACE(*s)) 624 s++; 625 if (s < send && *s) { 626 flags |= IS_NUMBER_TRAILING; 627 } 628 flags |= IS_NUMBER_INFINITY | IS_NUMBER_NOT_INT; 629 } 630 else if (isALPHA_FOLD_EQ(*s, 'D') && odh) { /* 1.#IND */ 631 s++; 632 flags |= IS_NUMBER_NAN | IS_NUMBER_NOT_INT; 633 while (*s == '0') { /* 1.#IND00 */ 634 s++; 635 } 636 if (*s) { 637 flags |= IS_NUMBER_TRAILING; 638 } 639 } else 640 return 0; 641 } 642 else { 643 /* Maybe NAN of some sort */ 644 645 if (isALPHA_FOLD_EQ(*s, 'S') || isALPHA_FOLD_EQ(*s, 'Q')) { 646 /* snan, qNaN */ 647 /* XXX do something with the snan/qnan difference */ 648 s++; if (s == send) return 0; 649 } 650 651 if (isALPHA_FOLD_EQ(*s, 'N')) { 652 s++; if (s == send || isALPHA_FOLD_NE(*s, 'A')) return 0; 653 s++; if (s == send || isALPHA_FOLD_NE(*s, 'N')) return 0; 654 s++; 655 656 flags |= IS_NUMBER_NAN | IS_NUMBER_NOT_INT; 657 658 /* NaN can be followed by various stuff (NaNQ, NaNS), but 659 * there are also multiple different NaN values, and some 660 * implementations output the "payload" values, 661 * e.g. NaN123, NAN(abc), while some legacy implementations 662 * have weird stuff like NaN%. */ 663 if (isALPHA_FOLD_EQ(*s, 'q') || 664 isALPHA_FOLD_EQ(*s, 's')) { 665 /* "nanq" or "nans" are ok, though generating 666 * these portably is tricky. */ 667 s++; 668 } 669 if (*s == '(') { 670 /* C99 style "nan(123)" or Perlish equivalent "nan($uv)". */ 671 const char *t; 672 s++; 673 if (s == send) { 674 return flags | IS_NUMBER_TRAILING; 675 } 676 t = s + 1; 677 while (t < send && *t && *t != ')') { 678 t++; 679 } 680 if (t == send) { 681 return flags | IS_NUMBER_TRAILING; 682 } 683 if (*t == ')') { 684 int nantype; 685 UV nanval; 686 if (s[0] == '0' && s + 2 < t && 687 isALPHA_FOLD_EQ(s[1], 'x') && 688 isXDIGIT(s[2])) { 689 STRLEN len = t - s; 690 I32 flags = PERL_SCAN_ALLOW_UNDERSCORES; 691 nanval = grok_hex(s, &len, &flags, NULL); 692 if ((flags & PERL_SCAN_GREATER_THAN_UV_MAX)) { 693 nantype = 0; 694 } else { 695 nantype = IS_NUMBER_IN_UV; 696 } 697 s += len; 698 } else if (s[0] == '0' && s + 2 < t && 699 isALPHA_FOLD_EQ(s[1], 'b') && 700 (s[2] == '0' || s[2] == '1')) { 701 STRLEN len = t - s; 702 I32 flags = PERL_SCAN_ALLOW_UNDERSCORES; 703 nanval = grok_bin(s, &len, &flags, NULL); 704 if ((flags & PERL_SCAN_GREATER_THAN_UV_MAX)) { 705 nantype = 0; 706 } else { 707 nantype = IS_NUMBER_IN_UV; 708 } 709 s += len; 710 } else { 711 const char *u; 712 nantype = 713 grok_number_flags(s, t - s, &nanval, 714 PERL_SCAN_TRAILING | 715 PERL_SCAN_ALLOW_UNDERSCORES); 716 /* Unfortunately grok_number_flags() doesn't 717 * tell how far we got and the ')' will always 718 * be "trailing", so we need to double-check 719 * whether we had something dubious. */ 720 for (u = s; u < t; u++) { 721 if (!isDIGIT(*u)) { 722 flags |= IS_NUMBER_TRAILING; 723 break; 724 } 725 } 726 s = u; 727 } 728 729 /* XXX Doesn't do octal: nan("0123"). 730 * Probably not a big loss. */ 731 732 if ((nantype & IS_NUMBER_NOT_INT) || 733 !(nantype && IS_NUMBER_IN_UV)) { 734 /* XXX the nanval is currently unused, that is, 735 * not inserted as the NaN payload of the NV. 736 * But the above code already parses the C99 737 * nan(...) format. See below, and see also 738 * the nan() in POSIX.xs. 739 * 740 * Certain configuration combinations where 741 * NVSIZE is greater than UVSIZE mean that 742 * a single UV cannot contain all the possible 743 * NaN payload bits. There would need to be 744 * some more generic syntax than "nan($uv)". 745 * 746 * Issues to keep in mind: 747 * 748 * (1) In most common cases there would 749 * not be an integral number of bytes that 750 * could be set, only a certain number of bits. 751 * For example for the common case of 752 * NVSIZE == UVSIZE == 8 there is room for 52 753 * bits in the payload, but the most significant 754 * bit is commonly reserved for the 755 * signaling/quiet bit, leaving 51 bits. 756 * Furthermore, the C99 nan() is supposed 757 * to generate quiet NaNs, so it is doubtful 758 * whether it should be able to generate 759 * signaling NaNs. For the x86 80-bit doubles 760 * (if building a long double Perl) there would 761 * be 62 bits (s/q bit being the 63rd). 762 * 763 * (2) Endianness of the payload bits. If the 764 * payload is specified as an UV, the low-order 765 * bits of the UV are naturally little-endianed 766 * (rightmost) bits of the payload. The endianness 767 * of UVs and NVs can be different. */ 768 return 0; 769 } 770 if (s < t) { 771 flags |= IS_NUMBER_TRAILING; 772 } 773 } else { 774 /* Looked like nan(...), but no close paren. */ 775 flags |= IS_NUMBER_TRAILING; 776 } 777 } else { 778 while (s < send && isSPACE(*s)) 779 s++; 780 if (s < send && *s) { 781 /* Note that we here implicitly accept (parse as 782 * "nan", but with warnings) also any other weird 783 * trailing stuff for "nan". In the above we just 784 * check that if we got the C99-style "nan(...)", 785 * the "..." looks sane. 786 * If in future we accept more ways of specifying 787 * the nan payload, the accepting would happen around 788 * here. */ 789 flags |= IS_NUMBER_TRAILING; 790 } 791 } 792 s = send; 793 } 794 else 795 return 0; 796 } 797 798 while (s < send && isSPACE(*s)) 799 s++; 800 801 *sp = s; 802 return flags; 803 } 804 805 /* 806 =for apidoc grok_number_flags 807 808 Recognise (or not) a number. The type of the number is returned 809 (0 if unrecognised), otherwise it is a bit-ORed combination of 810 C<IS_NUMBER_IN_UV>, C<IS_NUMBER_GREATER_THAN_UV_MAX>, C<IS_NUMBER_NOT_INT>, 811 C<IS_NUMBER_NEG>, C<IS_NUMBER_INFINITY>, C<IS_NUMBER_NAN> (defined in perl.h). 812 813 If the value of the number can fit in a UV, it is returned in C<*valuep>. 814 C<IS_NUMBER_IN_UV> will be set to indicate that C<*valuep> is valid, C<IS_NUMBER_IN_UV> 815 will never be set unless C<*valuep> is valid, but C<*valuep> may have been assigned 816 to during processing even though C<IS_NUMBER_IN_UV> is not set on return. 817 If C<valuep> is C<NULL>, C<IS_NUMBER_IN_UV> will be set for the same cases as when 818 C<valuep> is non-C<NULL>, but no actual assignment (or SEGV) will occur. 819 820 C<IS_NUMBER_NOT_INT> will be set with C<IS_NUMBER_IN_UV> if trailing decimals were 821 seen (in which case C<*valuep> gives the true value truncated to an integer), and 822 C<IS_NUMBER_NEG> if the number is negative (in which case C<*valuep> holds the 823 absolute value). C<IS_NUMBER_IN_UV> is not set if e notation was used or the 824 number is larger than a UV. 825 826 C<flags> allows only C<PERL_SCAN_TRAILING>, which allows for trailing 827 non-numeric text on an otherwise successful I<grok>, setting 828 C<IS_NUMBER_TRAILING> on the result. 829 830 =for apidoc grok_number 831 832 Identical to C<grok_number_flags()> with C<flags> set to zero. 833 834 =cut 835 */ 836 int 837 Perl_grok_number(pTHX_ const char *pv, STRLEN len, UV *valuep) 838 { 839 PERL_ARGS_ASSERT_GROK_NUMBER; 840 841 return grok_number_flags(pv, len, valuep, 0); 842 } 843 844 static const UV uv_max_div_10 = UV_MAX / 10; 845 static const U8 uv_max_mod_10 = UV_MAX % 10; 846 847 int 848 Perl_grok_number_flags(pTHX_ const char *pv, STRLEN len, UV *valuep, U32 flags) 849 { 850 const char *s = pv; 851 const char * const send = pv + len; 852 const char *d; 853 int numtype = 0; 854 855 PERL_ARGS_ASSERT_GROK_NUMBER_FLAGS; 856 857 while (s < send && isSPACE(*s)) 858 s++; 859 if (s == send) { 860 return 0; 861 } else if (*s == '-') { 862 s++; 863 numtype = IS_NUMBER_NEG; 864 } 865 else if (*s == '+') 866 s++; 867 868 if (s == send) 869 return 0; 870 871 /* The first digit (after optional sign): note that might 872 * also point to "infinity" or "nan", or "1.#INF". */ 873 d = s; 874 875 /* next must be digit or the radix separator or beginning of infinity/nan */ 876 if (isDIGIT(*s)) { 877 /* UVs are at least 32 bits, so the first 9 decimal digits cannot 878 overflow. */ 879 UV value = *s - '0'; 880 /* This construction seems to be more optimiser friendly. 881 (without it gcc does the isDIGIT test and the *s - '0' separately) 882 With it gcc on arm is managing 6 instructions (6 cycles) per digit. 883 In theory the optimiser could deduce how far to unroll the loop 884 before checking for overflow. */ 885 if (++s < send) { 886 int digit = *s - '0'; 887 if (digit >= 0 && digit <= 9) { 888 value = value * 10 + digit; 889 if (++s < send) { 890 digit = *s - '0'; 891 if (digit >= 0 && digit <= 9) { 892 value = value * 10 + digit; 893 if (++s < send) { 894 digit = *s - '0'; 895 if (digit >= 0 && digit <= 9) { 896 value = value * 10 + digit; 897 if (++s < send) { 898 digit = *s - '0'; 899 if (digit >= 0 && digit <= 9) { 900 value = value * 10 + digit; 901 if (++s < send) { 902 digit = *s - '0'; 903 if (digit >= 0 && digit <= 9) { 904 value = value * 10 + digit; 905 if (++s < send) { 906 digit = *s - '0'; 907 if (digit >= 0 && digit <= 9) { 908 value = value * 10 + digit; 909 if (++s < send) { 910 digit = *s - '0'; 911 if (digit >= 0 && digit <= 9) { 912 value = value * 10 + digit; 913 if (++s < send) { 914 digit = *s - '0'; 915 if (digit >= 0 && digit <= 9) { 916 value = value * 10 + digit; 917 if (++s < send) { 918 /* Now got 9 digits, so need to check 919 each time for overflow. */ 920 digit = *s - '0'; 921 while (digit >= 0 && digit <= 9 922 && (value < uv_max_div_10 923 || (value == uv_max_div_10 924 && digit <= uv_max_mod_10))) { 925 value = value * 10 + digit; 926 if (++s < send) 927 digit = *s - '0'; 928 else 929 break; 930 } 931 if (digit >= 0 && digit <= 9 932 && (s < send)) { 933 /* value overflowed. 934 skip the remaining digits, don't 935 worry about setting *valuep. */ 936 do { 937 s++; 938 } while (s < send && isDIGIT(*s)); 939 numtype |= 940 IS_NUMBER_GREATER_THAN_UV_MAX; 941 goto skip_value; 942 } 943 } 944 } 945 } 946 } 947 } 948 } 949 } 950 } 951 } 952 } 953 } 954 } 955 } 956 } 957 } 958 } 959 } 960 numtype |= IS_NUMBER_IN_UV; 961 if (valuep) 962 *valuep = value; 963 964 skip_value: 965 if (GROK_NUMERIC_RADIX(&s, send)) { 966 numtype |= IS_NUMBER_NOT_INT; 967 while (s < send && isDIGIT(*s)) /* optional digits after the radix */ 968 s++; 969 } 970 } 971 else if (GROK_NUMERIC_RADIX(&s, send)) { 972 numtype |= IS_NUMBER_NOT_INT | IS_NUMBER_IN_UV; /* valuep assigned below */ 973 /* no digits before the radix means we need digits after it */ 974 if (s < send && isDIGIT(*s)) { 975 do { 976 s++; 977 } while (s < send && isDIGIT(*s)); 978 if (valuep) { 979 /* integer approximation is valid - it's 0. */ 980 *valuep = 0; 981 } 982 } 983 else 984 return 0; 985 } 986 987 if (s > d && s < send) { 988 /* we can have an optional exponent part */ 989 if (isALPHA_FOLD_EQ(*s, 'e')) { 990 s++; 991 if (s < send && (*s == '-' || *s == '+')) 992 s++; 993 if (s < send && isDIGIT(*s)) { 994 do { 995 s++; 996 } while (s < send && isDIGIT(*s)); 997 } 998 else if (flags & PERL_SCAN_TRAILING) 999 return numtype | IS_NUMBER_TRAILING; 1000 else 1001 return 0; 1002 1003 /* The only flag we keep is sign. Blow away any "it's UV" */ 1004 numtype &= IS_NUMBER_NEG; 1005 numtype |= IS_NUMBER_NOT_INT; 1006 } 1007 } 1008 while (s < send && isSPACE(*s)) 1009 s++; 1010 if (s >= send) 1011 return numtype; 1012 if (len == 10 && memEQ(pv, "0 but true", 10)) { 1013 if (valuep) 1014 *valuep = 0; 1015 return IS_NUMBER_IN_UV; 1016 } 1017 /* We could be e.g. at "Inf" or "NaN", or at the "#" of "1.#INF". */ 1018 if ((s + 2 < send) && strchr("inqs#", toFOLD(*s))) { 1019 /* Really detect inf/nan. Start at d, not s, since the above 1020 * code might have already consumed the "1." or "1". */ 1021 int infnan = Perl_grok_infnan(aTHX_ &d, send); 1022 if ((infnan & IS_NUMBER_INFINITY)) { 1023 return (numtype | infnan); /* Keep sign for infinity. */ 1024 } 1025 else if ((infnan & IS_NUMBER_NAN)) { 1026 return (numtype | infnan) & ~IS_NUMBER_NEG; /* Clear sign for nan. */ 1027 } 1028 } 1029 else if (flags & PERL_SCAN_TRAILING) { 1030 return numtype | IS_NUMBER_TRAILING; 1031 } 1032 1033 return 0; 1034 } 1035 1036 /* 1037 grok_atoUV 1038 1039 grok_atoUV parses a C-style zero-byte terminated string, looking for 1040 a decimal unsigned integer. 1041 1042 Returns the unsigned integer, if a valid value can be parsed 1043 from the beginning of the string. 1044 1045 Accepts only the decimal digits '0'..'9'. 1046 1047 As opposed to atoi or strtol, grok_atoUV does NOT allow optional 1048 leading whitespace, or negative inputs. If such features are 1049 required, the calling code needs to explicitly implement those. 1050 1051 Returns true if a valid value could be parsed. In that case, valptr 1052 is set to the parsed value, and endptr (if provided) is set to point 1053 to the character after the last digit. 1054 1055 Returns false otherwise. This can happen if a) there is a leading zero 1056 followed by another digit; b) the digits would overflow a UV; or c) 1057 there are trailing non-digits AND endptr is not provided. 1058 1059 Background: atoi has severe problems with illegal inputs, it cannot be 1060 used for incremental parsing, and therefore should be avoided 1061 atoi and strtol are also affected by locale settings, which can also be 1062 seen as a bug (global state controlled by user environment). 1063 1064 */ 1065 1066 bool 1067 Perl_grok_atoUV(const char *pv, UV *valptr, const char** endptr) 1068 { 1069 const char* s = pv; 1070 const char** eptr; 1071 const char* end2; /* Used in case endptr is NULL. */ 1072 UV val = 0; /* The parsed value. */ 1073 1074 PERL_ARGS_ASSERT_GROK_ATOUV; 1075 1076 eptr = endptr ? endptr : &end2; 1077 if (isDIGIT(*s)) { 1078 /* Single-digit inputs are quite common. */ 1079 val = *s++ - '0'; 1080 if (isDIGIT(*s)) { 1081 /* Fail on extra leading zeros. */ 1082 if (val == 0) 1083 return FALSE; 1084 while (isDIGIT(*s)) { 1085 /* This could be unrolled like in grok_number(), but 1086 * the expected uses of this are not speed-needy, and 1087 * unlikely to need full 64-bitness. */ 1088 U8 digit = *s++ - '0'; 1089 if (val < uv_max_div_10 || 1090 (val == uv_max_div_10 && digit <= uv_max_mod_10)) { 1091 val = val * 10 + digit; 1092 } else { 1093 return FALSE; 1094 } 1095 } 1096 } 1097 } 1098 if (s == pv) 1099 return FALSE; 1100 if (endptr == NULL && *s) 1101 return FALSE; /* If endptr is NULL, no trailing non-digits allowed. */ 1102 *eptr = s; 1103 *valptr = val; 1104 return TRUE; 1105 } 1106 1107 #ifndef USE_QUADMATH 1108 STATIC NV 1109 S_mulexp10(NV value, I32 exponent) 1110 { 1111 NV result = 1.0; 1112 NV power = 10.0; 1113 bool negative = 0; 1114 I32 bit; 1115 1116 if (exponent == 0) 1117 return value; 1118 if (value == 0) 1119 return (NV)0; 1120 1121 /* On OpenVMS VAX we by default use the D_FLOAT double format, 1122 * and that format does not have *easy* capabilities [1] for 1123 * overflowing doubles 'silently' as IEEE fp does. We also need 1124 * to support G_FLOAT on both VAX and Alpha, and though the exponent 1125 * range is much larger than D_FLOAT it still doesn't do silent 1126 * overflow. Therefore we need to detect early whether we would 1127 * overflow (this is the behaviour of the native string-to-float 1128 * conversion routines, and therefore of native applications, too). 1129 * 1130 * [1] Trying to establish a condition handler to trap floating point 1131 * exceptions is not a good idea. */ 1132 1133 /* In UNICOS and in certain Cray models (such as T90) there is no 1134 * IEEE fp, and no way at all from C to catch fp overflows gracefully. 1135 * There is something you can do if you are willing to use some 1136 * inline assembler: the instruction is called DFI-- but that will 1137 * disable *all* floating point interrupts, a little bit too large 1138 * a hammer. Therefore we need to catch potential overflows before 1139 * it's too late. */ 1140 1141 #if ((defined(VMS) && !defined(_IEEE_FP)) || defined(_UNICOS)) && defined(NV_MAX_10_EXP) 1142 STMT_START { 1143 const NV exp_v = log10(value); 1144 if (exponent >= NV_MAX_10_EXP || exponent + exp_v >= NV_MAX_10_EXP) 1145 return NV_MAX; 1146 if (exponent < 0) { 1147 if (-(exponent + exp_v) >= NV_MAX_10_EXP) 1148 return 0.0; 1149 while (-exponent >= NV_MAX_10_EXP) { 1150 /* combination does not overflow, but 10^(-exponent) does */ 1151 value /= 10; 1152 ++exponent; 1153 } 1154 } 1155 } STMT_END; 1156 #endif 1157 1158 if (exponent < 0) { 1159 negative = 1; 1160 exponent = -exponent; 1161 #ifdef NV_MAX_10_EXP 1162 /* for something like 1234 x 10^-309, the action of calculating 1163 * the intermediate value 10^309 then returning 1234 / (10^309) 1164 * will fail, since 10^309 becomes infinity. In this case try to 1165 * refactor it as 123 / (10^308) etc. 1166 */ 1167 while (value && exponent > NV_MAX_10_EXP) { 1168 exponent--; 1169 value /= 10; 1170 } 1171 if (value == 0.0) 1172 return value; 1173 #endif 1174 } 1175 #if defined(__osf__) 1176 /* Even with cc -ieee + ieee_set_fp_control(IEEE_TRAP_ENABLE_INV) 1177 * Tru64 fp behavior on inf/nan is somewhat broken. Another way 1178 * to do this would be ieee_set_fp_control(IEEE_TRAP_ENABLE_OVF) 1179 * but that breaks another set of infnan.t tests. */ 1180 # define FP_OVERFLOWS_TO_ZERO 1181 #endif 1182 for (bit = 1; exponent; bit <<= 1) { 1183 if (exponent & bit) { 1184 exponent ^= bit; 1185 result *= power; 1186 #ifdef FP_OVERFLOWS_TO_ZERO 1187 if (result == 0) 1188 return value < 0 ? -NV_INF : NV_INF; 1189 #endif 1190 /* Floating point exceptions are supposed to be turned off, 1191 * but if we're obviously done, don't risk another iteration. 1192 */ 1193 if (exponent == 0) break; 1194 } 1195 power *= power; 1196 } 1197 return negative ? value / result : value * result; 1198 } 1199 #endif /* #ifndef USE_QUADMATH */ 1200 1201 NV 1202 Perl_my_atof(pTHX_ const char* s) 1203 { 1204 NV x = 0.0; 1205 #ifdef USE_QUADMATH 1206 Perl_my_atof2(aTHX_ s, &x); 1207 return x; 1208 #else 1209 # ifdef USE_LOCALE_NUMERIC 1210 PERL_ARGS_ASSERT_MY_ATOF; 1211 1212 { 1213 DECLARATION_FOR_LC_NUMERIC_MANIPULATION; 1214 STORE_LC_NUMERIC_SET_TO_NEEDED(); 1215 if (PL_numeric_radix_sv && IN_LC(LC_NUMERIC)) { 1216 const char *standard = NULL, *local = NULL; 1217 bool use_standard_radix; 1218 1219 /* Look through the string for the first thing that looks like a 1220 * decimal point: either the value in the current locale or the 1221 * standard fallback of '.'. The one which appears earliest in the 1222 * input string is the one that we should have atof look for. Note 1223 * that we have to determine this beforehand because on some 1224 * systems, Perl_atof2 is just a wrapper around the system's atof. 1225 * */ 1226 standard = strchr(s, '.'); 1227 local = strstr(s, SvPV_nolen(PL_numeric_radix_sv)); 1228 1229 use_standard_radix = standard && (!local || standard < local); 1230 1231 if (use_standard_radix) 1232 SET_NUMERIC_STANDARD(); 1233 1234 Perl_atof2(s, x); 1235 1236 if (use_standard_radix) 1237 SET_NUMERIC_UNDERLYING(); 1238 } 1239 else 1240 Perl_atof2(s, x); 1241 RESTORE_LC_NUMERIC(); 1242 } 1243 # else 1244 Perl_atof2(s, x); 1245 # endif 1246 #endif 1247 return x; 1248 } 1249 1250 1251 #ifdef USING_MSVC6 1252 # pragma warning(push) 1253 # pragma warning(disable:4756;disable:4056) 1254 #endif 1255 static char* 1256 S_my_atof_infnan(pTHX_ const char* s, bool negative, const char* send, NV* value) 1257 { 1258 const char *p0 = negative ? s - 1 : s; 1259 const char *p = p0; 1260 int infnan = grok_infnan(&p, send); 1261 if (infnan && p != p0) { 1262 /* If we can generate inf/nan directly, let's do so. */ 1263 #ifdef NV_INF 1264 if ((infnan & IS_NUMBER_INFINITY)) { 1265 *value = (infnan & IS_NUMBER_NEG) ? -NV_INF: NV_INF; 1266 return (char*)p; 1267 } 1268 #endif 1269 #ifdef NV_NAN 1270 if ((infnan & IS_NUMBER_NAN)) { 1271 *value = NV_NAN; 1272 return (char*)p; 1273 } 1274 #endif 1275 #ifdef Perl_strtod 1276 /* If still here, we didn't have either NV_INF or NV_NAN, 1277 * and can try falling back to native strtod/strtold. 1278 * 1279 * (Though, are our NV_INF or NV_NAN ever not defined?) 1280 * 1281 * The native interface might not recognize all the possible 1282 * inf/nan strings Perl recognizes. What we can try 1283 * is to try faking the input. We will try inf/-inf/nan 1284 * as the most promising/portable input. */ 1285 { 1286 const char* fake = NULL; 1287 char* endp; 1288 NV nv; 1289 if ((infnan & IS_NUMBER_INFINITY)) { 1290 fake = ((infnan & IS_NUMBER_NEG)) ? "-inf" : "inf"; 1291 } 1292 else if ((infnan & IS_NUMBER_NAN)) { 1293 fake = "nan"; 1294 } 1295 assert(fake); 1296 nv = Perl_strtod(fake, &endp); 1297 if (fake != endp) { 1298 if ((infnan & IS_NUMBER_INFINITY)) { 1299 #ifdef Perl_isinf 1300 if (Perl_isinf(nv)) 1301 *value = nv; 1302 #else 1303 /* last resort, may generate SIGFPE */ 1304 *value = Perl_exp((NV)1e9); 1305 if ((infnan & IS_NUMBER_NEG)) 1306 *value = -*value; 1307 #endif 1308 return (char*)p; /* p, not endp */ 1309 } 1310 else if ((infnan & IS_NUMBER_NAN)) { 1311 #ifdef Perl_isnan 1312 if (Perl_isnan(nv)) 1313 *value = nv; 1314 #else 1315 /* last resort, may generate SIGFPE */ 1316 *value = Perl_log((NV)-1.0); 1317 #endif 1318 return (char*)p; /* p, not endp */ 1319 } 1320 } 1321 } 1322 #endif /* #ifdef Perl_strtod */ 1323 } 1324 return NULL; 1325 } 1326 #ifdef USING_MSVC6 1327 # pragma warning(pop) 1328 #endif 1329 1330 char* 1331 Perl_my_atof2(pTHX_ const char* orig, NV* value) 1332 { 1333 const char* s = orig; 1334 NV result[3] = {0.0, 0.0, 0.0}; 1335 #if defined(USE_PERL_ATOF) || defined(USE_QUADMATH) 1336 const char* send = s + strlen(orig); /* one past the last */ 1337 bool negative = 0; 1338 #endif 1339 #if defined(USE_PERL_ATOF) && !defined(USE_QUADMATH) 1340 UV accumulator[2] = {0,0}; /* before/after dp */ 1341 bool seen_digit = 0; 1342 I32 exp_adjust[2] = {0,0}; 1343 I32 exp_acc[2] = {-1, -1}; 1344 /* the current exponent adjust for the accumulators */ 1345 I32 exponent = 0; 1346 I32 seen_dp = 0; 1347 I32 digit = 0; 1348 I32 old_digit = 0; 1349 I32 sig_digits = 0; /* noof significant digits seen so far */ 1350 #endif 1351 1352 #if defined(USE_PERL_ATOF) || defined(USE_QUADMATH) 1353 PERL_ARGS_ASSERT_MY_ATOF2; 1354 1355 /* leading whitespace */ 1356 while (isSPACE(*s)) 1357 ++s; 1358 1359 /* sign */ 1360 switch (*s) { 1361 case '-': 1362 negative = 1; 1363 /* FALLTHROUGH */ 1364 case '+': 1365 ++s; 1366 } 1367 #endif 1368 1369 #ifdef USE_QUADMATH 1370 { 1371 char* endp; 1372 if ((endp = S_my_atof_infnan(aTHX_ s, negative, send, value))) 1373 return endp; 1374 result[2] = strtoflt128(s, &endp); 1375 if (s != endp) { 1376 *value = negative ? -result[2] : result[2]; 1377 return endp; 1378 } 1379 return NULL; 1380 } 1381 #elif defined(USE_PERL_ATOF) 1382 1383 /* There is no point in processing more significant digits 1384 * than the NV can hold. Note that NV_DIG is a lower-bound value, 1385 * while we need an upper-bound value. We add 2 to account for this; 1386 * since it will have been conservative on both the first and last digit. 1387 * For example a 32-bit mantissa with an exponent of 4 would have 1388 * exact values in the set 1389 * 4 1390 * 8 1391 * .. 1392 * 17179869172 1393 * 17179869176 1394 * 17179869180 1395 * 1396 * where for the purposes of calculating NV_DIG we would have to discount 1397 * both the first and last digit, since neither can hold all values from 1398 * 0..9; but for calculating the value we must examine those two digits. 1399 */ 1400 #ifdef MAX_SIG_DIG_PLUS 1401 /* It is not necessarily the case that adding 2 to NV_DIG gets all the 1402 possible digits in a NV, especially if NVs are not IEEE compliant 1403 (e.g., long doubles on IRIX) - Allen <allens@cpan.org> */ 1404 # define MAX_SIG_DIGITS (NV_DIG+MAX_SIG_DIG_PLUS) 1405 #else 1406 # define MAX_SIG_DIGITS (NV_DIG+2) 1407 #endif 1408 1409 /* the max number we can accumulate in a UV, and still safely do 10*N+9 */ 1410 #define MAX_ACCUMULATE ( (UV) ((UV_MAX - 9)/10)) 1411 1412 { 1413 const char* endp; 1414 if ((endp = S_my_atof_infnan(aTHX_ s, negative, send, value))) 1415 return (char*)endp; 1416 } 1417 1418 /* we accumulate digits into an integer; when this becomes too 1419 * large, we add the total to NV and start again */ 1420 1421 while (1) { 1422 if (isDIGIT(*s)) { 1423 seen_digit = 1; 1424 old_digit = digit; 1425 digit = *s++ - '0'; 1426 if (seen_dp) 1427 exp_adjust[1]++; 1428 1429 /* don't start counting until we see the first significant 1430 * digit, eg the 5 in 0.00005... */ 1431 if (!sig_digits && digit == 0) 1432 continue; 1433 1434 if (++sig_digits > MAX_SIG_DIGITS) { 1435 /* limits of precision reached */ 1436 if (digit > 5) { 1437 ++accumulator[seen_dp]; 1438 } else if (digit == 5) { 1439 if (old_digit % 2) { /* round to even - Allen */ 1440 ++accumulator[seen_dp]; 1441 } 1442 } 1443 if (seen_dp) { 1444 exp_adjust[1]--; 1445 } else { 1446 exp_adjust[0]++; 1447 } 1448 /* skip remaining digits */ 1449 while (isDIGIT(*s)) { 1450 ++s; 1451 if (! seen_dp) { 1452 exp_adjust[0]++; 1453 } 1454 } 1455 /* warn of loss of precision? */ 1456 } 1457 else { 1458 if (accumulator[seen_dp] > MAX_ACCUMULATE) { 1459 /* add accumulator to result and start again */ 1460 result[seen_dp] = S_mulexp10(result[seen_dp], 1461 exp_acc[seen_dp]) 1462 + (NV)accumulator[seen_dp]; 1463 accumulator[seen_dp] = 0; 1464 exp_acc[seen_dp] = 0; 1465 } 1466 accumulator[seen_dp] = accumulator[seen_dp] * 10 + digit; 1467 ++exp_acc[seen_dp]; 1468 } 1469 } 1470 else if (!seen_dp && GROK_NUMERIC_RADIX(&s, send)) { 1471 seen_dp = 1; 1472 if (sig_digits > MAX_SIG_DIGITS) { 1473 do { 1474 ++s; 1475 } while (isDIGIT(*s)); 1476 break; 1477 } 1478 } 1479 else { 1480 break; 1481 } 1482 } 1483 1484 result[0] = S_mulexp10(result[0], exp_acc[0]) + (NV)accumulator[0]; 1485 if (seen_dp) { 1486 result[1] = S_mulexp10(result[1], exp_acc[1]) + (NV)accumulator[1]; 1487 } 1488 1489 if (seen_digit && (isALPHA_FOLD_EQ(*s, 'e'))) { 1490 bool expnegative = 0; 1491 1492 ++s; 1493 switch (*s) { 1494 case '-': 1495 expnegative = 1; 1496 /* FALLTHROUGH */ 1497 case '+': 1498 ++s; 1499 } 1500 while (isDIGIT(*s)) 1501 exponent = exponent * 10 + (*s++ - '0'); 1502 if (expnegative) 1503 exponent = -exponent; 1504 } 1505 1506 1507 1508 /* now apply the exponent */ 1509 1510 if (seen_dp) { 1511 result[2] = S_mulexp10(result[0],exponent+exp_adjust[0]) 1512 + S_mulexp10(result[1],exponent-exp_adjust[1]); 1513 } else { 1514 result[2] = S_mulexp10(result[0],exponent+exp_adjust[0]); 1515 } 1516 1517 /* now apply the sign */ 1518 if (negative) 1519 result[2] = -result[2]; 1520 #endif /* USE_PERL_ATOF */ 1521 *value = result[2]; 1522 return (char *)s; 1523 } 1524 1525 /* 1526 =for apidoc isinfnan 1527 1528 C<Perl_isinfnan()> is utility function that returns true if the NV 1529 argument is either an infinity or a C<NaN>, false otherwise. To test 1530 in more detail, use C<Perl_isinf()> and C<Perl_isnan()>. 1531 1532 This is also the logical inverse of Perl_isfinite(). 1533 1534 =cut 1535 */ 1536 bool 1537 Perl_isinfnan(NV nv) 1538 { 1539 #ifdef Perl_isinf 1540 if (Perl_isinf(nv)) 1541 return TRUE; 1542 #endif 1543 #ifdef Perl_isnan 1544 if (Perl_isnan(nv)) 1545 return TRUE; 1546 #endif 1547 return FALSE; 1548 } 1549 1550 /* 1551 =for apidoc 1552 1553 Checks whether the argument would be either an infinity or C<NaN> when used 1554 as a number, but is careful not to trigger non-numeric or uninitialized 1555 warnings. it assumes the caller has done C<SvGETMAGIC(sv)> already. 1556 1557 =cut 1558 */ 1559 1560 bool 1561 Perl_isinfnansv(pTHX_ SV *sv) 1562 { 1563 PERL_ARGS_ASSERT_ISINFNANSV; 1564 if (!SvOK(sv)) 1565 return FALSE; 1566 if (SvNOKp(sv)) 1567 return Perl_isinfnan(SvNVX(sv)); 1568 if (SvIOKp(sv)) 1569 return FALSE; 1570 { 1571 STRLEN len; 1572 const char *s = SvPV_nomg_const(sv, len); 1573 return cBOOL(grok_infnan(&s, s+len)); 1574 } 1575 } 1576 1577 #ifndef HAS_MODFL 1578 /* C99 has truncl, pre-C99 Solaris had aintl. We can use either with 1579 * copysignl to emulate modfl, which is in some platforms missing or 1580 * broken. */ 1581 # if defined(HAS_TRUNCL) && defined(HAS_COPYSIGNL) 1582 long double 1583 Perl_my_modfl(long double x, long double *ip) 1584 { 1585 *ip = truncl(x); 1586 return (x == *ip ? copysignl(0.0L, x) : x - *ip); 1587 } 1588 # elif defined(HAS_AINTL) && defined(HAS_COPYSIGNL) 1589 long double 1590 Perl_my_modfl(long double x, long double *ip) 1591 { 1592 *ip = aintl(x); 1593 return (x == *ip ? copysignl(0.0L, x) : x - *ip); 1594 } 1595 # endif 1596 #endif 1597 1598 /* Similarly, with ilogbl and scalbnl we can emulate frexpl. */ 1599 #if ! defined(HAS_FREXPL) && defined(HAS_ILOGBL) && defined(HAS_SCALBNL) 1600 long double 1601 Perl_my_frexpl(long double x, int *e) { 1602 *e = x == 0.0L ? 0 : ilogbl(x) + 1; 1603 return (scalbnl(x, -*e)); 1604 } 1605 #endif 1606 1607 /* 1608 =for apidoc Perl_signbit 1609 1610 Return a non-zero integer if the sign bit on an NV is set, and 0 if 1611 it is not. 1612 1613 If F<Configure> detects this system has a C<signbit()> that will work with 1614 our NVs, then we just use it via the C<#define> in F<perl.h>. Otherwise, 1615 fall back on this implementation. The main use of this function 1616 is catching C<-0.0>. 1617 1618 C<Configure> notes: This function is called C<'Perl_signbit'> instead of a 1619 plain C<'signbit'> because it is easy to imagine a system having a C<signbit()> 1620 function or macro that doesn't happen to work with our particular choice 1621 of NVs. We shouldn't just re-C<#define> C<signbit> as C<Perl_signbit> and expect 1622 the standard system headers to be happy. Also, this is a no-context 1623 function (no C<pTHX_>) because C<Perl_signbit()> is usually re-C<#defined> in 1624 F<perl.h> as a simple macro call to the system's C<signbit()>. 1625 Users should just always call C<Perl_signbit()>. 1626 1627 =cut 1628 */ 1629 #if !defined(HAS_SIGNBIT) 1630 int 1631 Perl_signbit(NV x) { 1632 # ifdef Perl_fp_class_nzero 1633 return Perl_fp_class_nzero(x); 1634 /* Try finding the high byte, and assume it's highest bit 1635 * is the sign. This assumption is probably wrong somewhere. */ 1636 # elif defined(USE_LONG_DOUBLE) && LONG_DOUBLEKIND == LONG_DOUBLE_IS_X86_80_BIT_LITTLE_ENDIAN 1637 return (((unsigned char *)&x)[9] & 0x80); 1638 # elif defined(NV_LITTLE_ENDIAN) 1639 /* Note that NVSIZE is sizeof(NV), which would make the below be 1640 * wrong if the end bytes are unused, which happens with the x86 1641 * 80-bit long doubles, which is why take care of that above. */ 1642 return (((unsigned char *)&x)[NVSIZE - 1] & 0x80); 1643 # elif defined(NV_BIG_ENDIAN) 1644 return (((unsigned char *)&x)[0] & 0x80); 1645 # else 1646 /* This last resort fallback is wrong for the negative zero. */ 1647 return (x < 0.0) ? 1 : 0; 1648 # endif 1649 } 1650 #endif 1651 1652 /* 1653 * ex: set ts=8 sts=4 sw=4 et: 1654 */ 1655