From 57d47c633ce692dff24a804382935f46f7074ebf Mon Sep 17 00:00:00 2001 From: Nigel Tao Date: Tue, 8 Sep 2020 16:43:31 +1000 Subject: [PATCH] Fix etc__high_prec_dec__to_f64 shift amount MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit Shifting 0.299 left by 2 doesn't put us in [½ .. 1]. --- internal/cgen/base/floatconv-submodule-code.c | 2 +- internal/cgen/data/data.go | 24 +++++++++---------- release/c/wuffs-unsupported-snapshot.c | 2 +- script/print-render-number-f64-tests.go | 2 ++ test/c/std/json.c | 21 ++++++++++++++++ 5 files changed, 37 insertions(+), 14 deletions(-) diff --git a/internal/cgen/base/floatconv-submodule-code.c b/internal/cgen/base/floatconv-submodule-code.c index a07e2328..05debab2 100644 --- a/internal/cgen/base/floatconv-submodule-code.c +++ b/internal/cgen/base/floatconv-submodule-code.c @@ -1304,7 +1304,7 @@ wuffs_base__private_implementation__high_prec_dec__to_f64( if (h->digits[0] >= 5) { break; } - shift = (h->digits[0] <= 2) ? 2 : 1; + shift = (h->digits[0] < 2) ? 2 : 1; } else { uint32_t n = (uint32_t)(-h->decimal_point); shift = (n < num_powers) diff --git a/internal/cgen/data/data.go b/internal/cgen/data/data.go index 0cdec624..b64c4a3c 100644 --- a/internal/cgen/data/data.go +++ b/internal/cgen/data/data.go @@ -399,18 +399,18 @@ const BaseFloatConvSubmoduleCodeC = "" + " ret.status.repr = NULL;\n ret.value = wuffs_base__ieee_754_bit_representation__from_u64_to_f64(\n (nan ? 0x7FFFFFFFFFFFFFFF : 0x7FF0000000000000) |\n (negative ? 0x8000000000000000 : 0));\n return ret;\n } while (0);\n\nfail:\n do {\n wuffs_base__result_f64 ret;\n ret.status.repr = wuffs_base__error__bad_argument;\n ret.value = 0;\n return ret;\n } while (0);\n}\n\nWUFFS_BASE__MAYBE_STATIC wuffs_base__result_f64 //\nwuffs_base__private_implementation__high_prec_dec__to_f64(\n wuffs_base__private_implementation__high_prec_dec* h,\n uint32_t options) {\n do {\n // powers converts decimal powers of 10 to binary powers of 2. For example,\n // (10000 >> 13) is 1. It stops before the elements exceed 60, also known\n // as WUFFS_BASE__PRIVATE_IMPLEMENTATION__HPD__SHIFT__MAX_INCL.\n static const uint32_t num_powers = 19;\n static const uint8_t powers[19] = {\n 0, 3, 6, 9, 13, 16, 19, 23, 26, 29, //\n 33, 36, 39, 43, 46, 49, 53, 56, 59, //\n };\n\n // Handl" + "e zero and obvious extremes. The largest and smallest positive\n // finite f64 values are approximately 1.8e+308 and 4.9e-324.\n if ((h->num_digits == 0) || (h->decimal_point < -326)) {\n goto zero;\n } else if (h->decimal_point > 310) {\n goto infinity;\n }\n\n // Try the fast Eisel-Lemire algorithm again. Calculating the (man, exp10)\n // pair from the high_prec_dec h is more correct but slower than the\n // approach taken in wuffs_base__parse_number_f64. The latter is optimized\n // for the common cases (e.g. assuming no underscores or a leading '+'\n // sign) rather than the full set of cases allowed by the Wuffs API.\n if (h->num_digits <= 19) {\n uint64_t man = 0;\n uint32_t i;\n for (i = 0; i < h->num_digits; i++) {\n man = (10 * man) + h->digits[i];\n }\n int32_t exp10 = h->decimal_point - ((int32_t)(h->num_digits));\n if ((man != 0) && (-326 <= exp10) && (exp10 <= 310)) {\n int64_t r =\n wuffs_base__private_implementation__parse" + "_number_f64_eisel_lemire(\n man, exp10);\n if (r >= 0) {\n wuffs_base__result_f64 ret;\n ret.status.repr = NULL;\n ret.value = wuffs_base__ieee_754_bit_representation__from_u64_to_f64(\n ((uint64_t)r) | (((uint64_t)(h->negative)) << 63));\n return ret;\n }\n }\n }\n\n // Scale by powers of 2 until we're in the range [½ .. 1], which gives us\n // our exponent (in base-2). First we shift right, possibly a little too\n // far, ending with a value certainly below 1 and possibly below ½...\n const int32_t f64_bias = -1023;\n int32_t exp2 = 0;\n while (h->decimal_point > 0) {\n uint32_t n = (uint32_t)(+h->decimal_point);\n uint32_t shift =\n (n < num_powers)\n ? powers[n]\n : WUFFS_BASE__PRIVATE_IMPLEMENTATION__HPD__SHIFT__MAX_INCL;\n\n wuffs_base__private_implementation__high_prec_dec__small_rshift(h, shift);\n if (h->decimal_point <\n -WUFFS_BASE__PRIVATE_IMPLEMENTATION__" + - "HPD__DECIMAL_POINT__RANGE) {\n goto zero;\n }\n exp2 += (int32_t)shift;\n }\n // ...then we shift left, putting us in [½ .. 1].\n while (h->decimal_point <= 0) {\n uint32_t shift;\n if (h->decimal_point == 0) {\n if (h->digits[0] >= 5) {\n break;\n }\n shift = (h->digits[0] <= 2) ? 2 : 1;\n } else {\n uint32_t n = (uint32_t)(-h->decimal_point);\n shift = (n < num_powers)\n ? powers[n]\n : WUFFS_BASE__PRIVATE_IMPLEMENTATION__HPD__SHIFT__MAX_INCL;\n }\n\n wuffs_base__private_implementation__high_prec_dec__small_lshift(h, shift);\n if (h->decimal_point >\n +WUFFS_BASE__PRIVATE_IMPLEMENTATION__HPD__DECIMAL_POINT__RANGE) {\n goto infinity;\n }\n exp2 -= (int32_t)shift;\n }\n\n // We're in the range [½ .. 1] but f64 uses [1 .. 2].\n exp2--;\n\n // The minimum normal exponent is (f64_bias + 1).\n while ((f64_bias + 1) > exp2) {\n uint32_t n = (uint32_t)((f64_bias + " + - "1) - exp2);\n if (n > WUFFS_BASE__PRIVATE_IMPLEMENTATION__HPD__SHIFT__MAX_INCL) {\n n = WUFFS_BASE__PRIVATE_IMPLEMENTATION__HPD__SHIFT__MAX_INCL;\n }\n wuffs_base__private_implementation__high_prec_dec__small_rshift(h, n);\n exp2 += (int32_t)n;\n }\n\n // Check for overflow.\n if ((exp2 - f64_bias) >= 0x07FF) { // (1 << 11) - 1.\n goto infinity;\n }\n\n // Extract 53 bits for the mantissa (in base-2).\n wuffs_base__private_implementation__high_prec_dec__small_lshift(h, 53);\n uint64_t man2 =\n wuffs_base__private_implementation__high_prec_dec__rounded_integer(h);\n\n // Rounding might have added one bit. If so, shift and re-check overflow.\n if ((man2 >> 53) != 0) {\n man2 >>= 1;\n exp2++;\n if ((exp2 - f64_bias) >= 0x07FF) { // (1 << 11) - 1.\n goto infinity;\n }\n }\n\n // Handle subnormal numbers.\n if ((man2 >> 52) == 0) {\n exp2 = f64_bias;\n }\n\n // Pack the bits and return.\n uint64_t exp2_bits =\n (uint64_t)((e" + - "xp2 - f64_bias) & 0x07FF); // (1 << 11) - 1.\n uint64_t bits = (man2 & 0x000FFFFFFFFFFFFF) | // (1 << 52) - 1.\n (exp2_bits << 52) | //\n (h->negative ? 0x8000000000000000 : 0); // (1 << 63).\n\n wuffs_base__result_f64 ret;\n ret.status.repr = NULL;\n ret.value = wuffs_base__ieee_754_bit_representation__from_u64_to_f64(bits);\n return ret;\n } while (0);\n\nzero:\n do {\n uint64_t bits = h->negative ? 0x8000000000000000 : 0;\n\n wuffs_base__result_f64 ret;\n ret.status.repr = NULL;\n ret.value = wuffs_base__ieee_754_bit_representation__from_u64_to_f64(bits);\n return ret;\n } while (0);\n\ninfinity:\n do {\n if (options & WUFFS_BASE__PARSE_NUMBER_FXX__REJECT_INF_AND_NAN) {\n wuffs_base__result_f64 ret;\n ret.status.repr = wuffs_base__error__bad_argument;\n ret.value = 0;\n return ret;\n }\n\n uint64_t bits = h->negative ? 0xFFF0000000000000 : 0x7FF0000000000000;\n\n wuffs_base__result_f64 ret;" + - "\n ret.status.repr = NULL;\n ret.value = wuffs_base__ieee_754_bit_representation__from_u64_to_f64(bits);\n return ret;\n } while (0);\n}\n\nstatic inline bool //\nwuffs_base__private_implementation__is_decimal_digit(uint8_t c) {\n return ('0' <= c) && (c <= '9');\n}\n\nWUFFS_BASE__MAYBE_STATIC wuffs_base__result_f64 //\nwuffs_base__parse_number_f64(wuffs_base__slice_u8 s, uint32_t options) {\n // In practice, almost all \"dd.ddddE±xxx\" numbers can be represented\n // losslessly by a uint64_t mantissa \"dddddd\" and an int32_t base-10\n // exponent, adjusting \"xxx\" for the position (if present) of the decimal\n // separator '.' or ','.\n //\n // This (u64 man, i32 exp10) data structure is superficially similar to the\n // \"Do It Yourself Floating Point\" type from Loitsch (†), but the exponent\n // here is base-10, not base-2.\n //\n // If s's number fits in a (man, exp10), parse that pair with the\n // Eisel-Lemire algorithm. If not, or if Eisel-Lemire fails, parsing s with\n // the fallback algorithm is slowe" + - "r but comprehensive.\n //\n // † \"Printing Floating-Point Numbers Quickly and Accurately with Integers\"\n // (https://www.cs.tufts.edu/~nr/cs257/archive/florian-loitsch/printf.pdf).\n // Florian Loitsch is also the primary contributor to\n // https://github.com/google/double-conversion\n do {\n // Calculating that (man, exp10) pair needs to stay within s's bounds.\n // Provided that s isn't extremely long, work on a NUL-terminated copy of\n // s's contents. The NUL byte isn't a valid part of \"±dd.ddddE±xxx\".\n //\n // As the pointer p walks the contents, it's faster to repeatedly check \"is\n // *p a valid digit\" than \"is p within bounds and *p a valid digit\".\n if (s.len >= 256) {\n goto fallback;\n }\n uint8_t z[256];\n memcpy(&z[0], s.ptr, s.len);\n z[s.len] = 0;\n const uint8_t* p = &z[0];\n\n // Look for a leading minus sign. Technically, we could also look for an\n // optional plus sign, but the \"script/process-json-numbers.c with -p\"\n // benchmark is noticably slowe" + - "r if we do. It's optional and, in practice,\n // usually absent. Let the fallback catch it.\n bool negative = (*p == '-');\n if (negative) {\n p++;\n }\n\n // After walking \"dd.dddd\", comparing p later with p now will produce the\n // number of \"d\"s and \".\"s.\n const uint8_t* const start_of_digits_ptr = p;\n\n // Walk the \"d\"s before a '.', 'E', NUL byte, etc. If it starts with '0',\n // it must be a single '0'. If it starts with a non-zero decimal digit, it\n // can be a sequence of decimal digits.\n //\n // Update the man variable during the walk. It's OK if man overflows now.\n // We'll detect that later.\n uint64_t man;\n if (*p == '0') {\n man = 0;\n p++;\n if (wuffs_base__private_implementation__is_decimal_digit(*p)) {\n goto fallback;\n }\n } else if (wuffs_base__private_implementation__is_decimal_digit(*p)) {\n man = ((uint8_t)(*p - '0'));\n p++;\n for (; wuffs_base__private_implementation__is_decimal_digit(*p); p++) {\n man = (" + - "10 * man) + ((uint8_t)(*p - '0'));\n }\n } else {\n goto fallback;\n }\n\n // Walk the \"d\"s after the optional decimal separator ('.' or ','),\n // updating the man and exp10 variables.\n int32_t exp10 = 0;\n if (*p ==\n ((options & WUFFS_BASE__PARSE_NUMBER_FXX__DECIMAL_SEPARATOR_IS_A_COMMA)\n ? ','\n : '.')) {\n p++;\n const uint8_t* first_after_separator_ptr = p;\n if (!wuffs_base__private_implementation__is_decimal_digit(*p)) {\n goto fallback;\n }\n man = (10 * man) + ((uint8_t)(*p - '0'));\n p++;\n for (; wuffs_base__private_implementation__is_decimal_digit(*p); p++) {\n man = (10 * man) + ((uint8_t)(*p - '0'));\n }\n exp10 = ((int32_t)(first_after_separator_ptr - p));\n }\n\n // Count the number of digits:\n // - for an input of \"314159\", digit_count is 6.\n // - for an input of \"3.14159\", digit_count is 7.\n //\n // This is off-by-one if there is a decimal separator. That's OK for now.\n // We" + - "'ll correct for that later. The \"script/process-json-numbers.c with\n // -p\" benchmark is noticably slower if we try to correct for that now.\n uint32_t digit_count = (uint32_t)(p - start_of_digits_ptr);\n\n // Update exp10 for the optional exponent, starting with 'E' or 'e'.\n if ((*p | 0x20) == 'e') {\n p++;\n int32_t exp_sign = +1;\n if (*p == '-') {\n p++;\n exp_sign = -1;\n } else if (*p == '+') {\n p++;\n }\n if (!wuffs_base__private_implementation__is_decimal_digit(*p)) {\n goto fallback;\n }\n int32_t exp_num = ((uint8_t)(*p - '0'));\n p++;\n // The rest of the exp_num walking has a peculiar control flow but, once\n // again, the \"script/process-json-numbers.c with -p\" benchmark is\n // sensitive to alternative formulations.\n if (wuffs_base__private_implementation__is_decimal_digit(*p)) {\n exp_num = (10 * exp_num) + ((uint8_t)(*p - '0'));\n p++;\n }\n if (wuffs_base__private_implementation__is_deci" + - "mal_digit(*p)) {\n exp_num = (10 * exp_num) + ((uint8_t)(*p - '0'));\n p++;\n }\n while (wuffs_base__private_implementation__is_decimal_digit(*p)) {\n if (exp_num > 0x1000000) {\n goto fallback;\n }\n exp_num = (10 * exp_num) + ((uint8_t)(*p - '0'));\n p++;\n }\n exp10 += exp_sign * exp_num;\n }\n\n // The Wuffs API is that the original slice has no trailing data. It also\n // allows underscores, which we don't catch here but the fallback should.\n if (p != &z[s.len]) {\n goto fallback;\n }\n\n // Check that the uint64_t typed man variable has not overflowed, based on\n // digit_count.\n //\n // For reference:\n // - (1 << 63) is 9223372036854775808, which has 19 decimal digits.\n // - (1 << 64) is 18446744073709551616, which has 20 decimal digits.\n // - 19 nines, 9999999999999999999, is 0x8AC7230489E7FFFF, which has 64\n // bits and 16 hexadecimal digits.\n // - 20 nines, 99999999999999999999, is 0x56BC7" + - "5E2D630FFFFF, which has 67\n // bits and 17 hexadecimal digits.\n if (digit_count > 19) {\n // Even if we have more than 19 pseudo-digits, it's not yet definitely an\n // overflow. Recall that digit_count might be off-by-one (too large) if\n // there's a decimal separator. It will also over-report the number of\n // meaningful digits if the input looks something like \"0.000dddExxx\".\n //\n // We adjust by the number of leading '0's and '.'s and re-compare to 19.\n // Once again, technically, we could skip ','s too, but that perturbs the\n // \"script/process-json-numbers.c with -p\" benchmark.\n const uint8_t* q = start_of_digits_ptr;\n for (; (*q == '0') || (*q == '.'); q++) {\n }\n digit_count -= (uint32_t)(q - start_of_digits_ptr);\n if (digit_count > 19) {\n goto fallback;\n }\n }\n\n // The wuffs_base__private_implementation__parse_number_f64_eisel_lemire\n // preconditions include that exp10 is in the range -326 ..= 310.\n if ((e" + - "xp10 < -326) || (310 < exp10)) {\n goto fallback;\n }\n\n // If man and exp10 are small enough, all three of (man), (10 ** exp10) and\n // (man ** (10 ** exp10)) are exactly representable by a double. We don't\n // need to run the Eisel-Lemire algorithm.\n if ((-22 <= exp10) && (exp10 <= 22) && ((man >> 53) == 0)) {\n double d = (double)man;\n if (exp10 >= 0) {\n d *= wuffs_base__private_implementation__f64_powers_of_10[+exp10];\n } else {\n d /= wuffs_base__private_implementation__f64_powers_of_10[-exp10];\n }\n wuffs_base__result_f64 ret;\n ret.status.repr = NULL;\n ret.value = negative ? -d : +d;\n return ret;\n }\n\n // The wuffs_base__private_implementation__parse_number_f64_eisel_lemire\n // preconditions include that man is non-zero. Parsing \"0\" should be caught\n // by the \"If man and exp10 are small enough\" above, but \"0e99\" might not.\n if (man == 0) {\n goto fallback;\n }\n\n // Our man and exp10 are in range. Run the Eisel-Le" + - "mire algorithm.\n int64_t r =\n wuffs_base__private_implementation__parse_number_f64_eisel_lemire(\n man, exp10);\n if (r < 0) {\n goto fallback;\n }\n wuffs_base__result_f64 ret;\n ret.status.repr = NULL;\n ret.value = wuffs_base__ieee_754_bit_representation__from_u64_to_f64(\n ((uint64_t)r) | (((uint64_t)negative) << 63));\n return ret;\n } while (0);\n\nfallback:\n do {\n wuffs_base__private_implementation__high_prec_dec h;\n wuffs_base__status status =\n wuffs_base__private_implementation__high_prec_dec__parse(&h, s,\n options);\n if (status.repr) {\n return wuffs_base__private_implementation__parse_number_f64_special(\n s, options);\n }\n return wuffs_base__private_implementation__high_prec_dec__to_f64(&h,\n options);\n } while (0);\n}\n\n" + + "HPD__DECIMAL_POINT__RANGE) {\n goto zero;\n }\n exp2 += (int32_t)shift;\n }\n // ...then we shift left, putting us in [½ .. 1].\n while (h->decimal_point <= 0) {\n uint32_t shift;\n if (h->decimal_point == 0) {\n if (h->digits[0] >= 5) {\n break;\n }\n shift = (h->digits[0] < 2) ? 2 : 1;\n } else {\n uint32_t n = (uint32_t)(-h->decimal_point);\n shift = (n < num_powers)\n ? powers[n]\n : WUFFS_BASE__PRIVATE_IMPLEMENTATION__HPD__SHIFT__MAX_INCL;\n }\n\n wuffs_base__private_implementation__high_prec_dec__small_lshift(h, shift);\n if (h->decimal_point >\n +WUFFS_BASE__PRIVATE_IMPLEMENTATION__HPD__DECIMAL_POINT__RANGE) {\n goto infinity;\n }\n exp2 -= (int32_t)shift;\n }\n\n // We're in the range [½ .. 1] but f64 uses [1 .. 2].\n exp2--;\n\n // The minimum normal exponent is (f64_bias + 1).\n while ((f64_bias + 1) > exp2) {\n uint32_t n = (uint32_t)((f64_bias + 1" + + ") - exp2);\n if (n > WUFFS_BASE__PRIVATE_IMPLEMENTATION__HPD__SHIFT__MAX_INCL) {\n n = WUFFS_BASE__PRIVATE_IMPLEMENTATION__HPD__SHIFT__MAX_INCL;\n }\n wuffs_base__private_implementation__high_prec_dec__small_rshift(h, n);\n exp2 += (int32_t)n;\n }\n\n // Check for overflow.\n if ((exp2 - f64_bias) >= 0x07FF) { // (1 << 11) - 1.\n goto infinity;\n }\n\n // Extract 53 bits for the mantissa (in base-2).\n wuffs_base__private_implementation__high_prec_dec__small_lshift(h, 53);\n uint64_t man2 =\n wuffs_base__private_implementation__high_prec_dec__rounded_integer(h);\n\n // Rounding might have added one bit. If so, shift and re-check overflow.\n if ((man2 >> 53) != 0) {\n man2 >>= 1;\n exp2++;\n if ((exp2 - f64_bias) >= 0x07FF) { // (1 << 11) - 1.\n goto infinity;\n }\n }\n\n // Handle subnormal numbers.\n if ((man2 >> 52) == 0) {\n exp2 = f64_bias;\n }\n\n // Pack the bits and return.\n uint64_t exp2_bits =\n (uint64_t)((ex" + + "p2 - f64_bias) & 0x07FF); // (1 << 11) - 1.\n uint64_t bits = (man2 & 0x000FFFFFFFFFFFFF) | // (1 << 52) - 1.\n (exp2_bits << 52) | //\n (h->negative ? 0x8000000000000000 : 0); // (1 << 63).\n\n wuffs_base__result_f64 ret;\n ret.status.repr = NULL;\n ret.value = wuffs_base__ieee_754_bit_representation__from_u64_to_f64(bits);\n return ret;\n } while (0);\n\nzero:\n do {\n uint64_t bits = h->negative ? 0x8000000000000000 : 0;\n\n wuffs_base__result_f64 ret;\n ret.status.repr = NULL;\n ret.value = wuffs_base__ieee_754_bit_representation__from_u64_to_f64(bits);\n return ret;\n } while (0);\n\ninfinity:\n do {\n if (options & WUFFS_BASE__PARSE_NUMBER_FXX__REJECT_INF_AND_NAN) {\n wuffs_base__result_f64 ret;\n ret.status.repr = wuffs_base__error__bad_argument;\n ret.value = 0;\n return ret;\n }\n\n uint64_t bits = h->negative ? 0xFFF0000000000000 : 0x7FF0000000000000;\n\n wuffs_base__result_f64 ret;\n" + + " ret.status.repr = NULL;\n ret.value = wuffs_base__ieee_754_bit_representation__from_u64_to_f64(bits);\n return ret;\n } while (0);\n}\n\nstatic inline bool //\nwuffs_base__private_implementation__is_decimal_digit(uint8_t c) {\n return ('0' <= c) && (c <= '9');\n}\n\nWUFFS_BASE__MAYBE_STATIC wuffs_base__result_f64 //\nwuffs_base__parse_number_f64(wuffs_base__slice_u8 s, uint32_t options) {\n // In practice, almost all \"dd.ddddE±xxx\" numbers can be represented\n // losslessly by a uint64_t mantissa \"dddddd\" and an int32_t base-10\n // exponent, adjusting \"xxx\" for the position (if present) of the decimal\n // separator '.' or ','.\n //\n // This (u64 man, i32 exp10) data structure is superficially similar to the\n // \"Do It Yourself Floating Point\" type from Loitsch (†), but the exponent\n // here is base-10, not base-2.\n //\n // If s's number fits in a (man, exp10), parse that pair with the\n // Eisel-Lemire algorithm. If not, or if Eisel-Lemire fails, parsing s with\n // the fallback algorithm is slower" + + " but comprehensive.\n //\n // † \"Printing Floating-Point Numbers Quickly and Accurately with Integers\"\n // (https://www.cs.tufts.edu/~nr/cs257/archive/florian-loitsch/printf.pdf).\n // Florian Loitsch is also the primary contributor to\n // https://github.com/google/double-conversion\n do {\n // Calculating that (man, exp10) pair needs to stay within s's bounds.\n // Provided that s isn't extremely long, work on a NUL-terminated copy of\n // s's contents. The NUL byte isn't a valid part of \"±dd.ddddE±xxx\".\n //\n // As the pointer p walks the contents, it's faster to repeatedly check \"is\n // *p a valid digit\" than \"is p within bounds and *p a valid digit\".\n if (s.len >= 256) {\n goto fallback;\n }\n uint8_t z[256];\n memcpy(&z[0], s.ptr, s.len);\n z[s.len] = 0;\n const uint8_t* p = &z[0];\n\n // Look for a leading minus sign. Technically, we could also look for an\n // optional plus sign, but the \"script/process-json-numbers.c with -p\"\n // benchmark is noticably slower" + + " if we do. It's optional and, in practice,\n // usually absent. Let the fallback catch it.\n bool negative = (*p == '-');\n if (negative) {\n p++;\n }\n\n // After walking \"dd.dddd\", comparing p later with p now will produce the\n // number of \"d\"s and \".\"s.\n const uint8_t* const start_of_digits_ptr = p;\n\n // Walk the \"d\"s before a '.', 'E', NUL byte, etc. If it starts with '0',\n // it must be a single '0'. If it starts with a non-zero decimal digit, it\n // can be a sequence of decimal digits.\n //\n // Update the man variable during the walk. It's OK if man overflows now.\n // We'll detect that later.\n uint64_t man;\n if (*p == '0') {\n man = 0;\n p++;\n if (wuffs_base__private_implementation__is_decimal_digit(*p)) {\n goto fallback;\n }\n } else if (wuffs_base__private_implementation__is_decimal_digit(*p)) {\n man = ((uint8_t)(*p - '0'));\n p++;\n for (; wuffs_base__private_implementation__is_decimal_digit(*p); p++) {\n man = (1" + + "0 * man) + ((uint8_t)(*p - '0'));\n }\n } else {\n goto fallback;\n }\n\n // Walk the \"d\"s after the optional decimal separator ('.' or ','),\n // updating the man and exp10 variables.\n int32_t exp10 = 0;\n if (*p ==\n ((options & WUFFS_BASE__PARSE_NUMBER_FXX__DECIMAL_SEPARATOR_IS_A_COMMA)\n ? ','\n : '.')) {\n p++;\n const uint8_t* first_after_separator_ptr = p;\n if (!wuffs_base__private_implementation__is_decimal_digit(*p)) {\n goto fallback;\n }\n man = (10 * man) + ((uint8_t)(*p - '0'));\n p++;\n for (; wuffs_base__private_implementation__is_decimal_digit(*p); p++) {\n man = (10 * man) + ((uint8_t)(*p - '0'));\n }\n exp10 = ((int32_t)(first_after_separator_ptr - p));\n }\n\n // Count the number of digits:\n // - for an input of \"314159\", digit_count is 6.\n // - for an input of \"3.14159\", digit_count is 7.\n //\n // This is off-by-one if there is a decimal separator. That's OK for now.\n // We'" + + "ll correct for that later. The \"script/process-json-numbers.c with\n // -p\" benchmark is noticably slower if we try to correct for that now.\n uint32_t digit_count = (uint32_t)(p - start_of_digits_ptr);\n\n // Update exp10 for the optional exponent, starting with 'E' or 'e'.\n if ((*p | 0x20) == 'e') {\n p++;\n int32_t exp_sign = +1;\n if (*p == '-') {\n p++;\n exp_sign = -1;\n } else if (*p == '+') {\n p++;\n }\n if (!wuffs_base__private_implementation__is_decimal_digit(*p)) {\n goto fallback;\n }\n int32_t exp_num = ((uint8_t)(*p - '0'));\n p++;\n // The rest of the exp_num walking has a peculiar control flow but, once\n // again, the \"script/process-json-numbers.c with -p\" benchmark is\n // sensitive to alternative formulations.\n if (wuffs_base__private_implementation__is_decimal_digit(*p)) {\n exp_num = (10 * exp_num) + ((uint8_t)(*p - '0'));\n p++;\n }\n if (wuffs_base__private_implementation__is_decim" + + "al_digit(*p)) {\n exp_num = (10 * exp_num) + ((uint8_t)(*p - '0'));\n p++;\n }\n while (wuffs_base__private_implementation__is_decimal_digit(*p)) {\n if (exp_num > 0x1000000) {\n goto fallback;\n }\n exp_num = (10 * exp_num) + ((uint8_t)(*p - '0'));\n p++;\n }\n exp10 += exp_sign * exp_num;\n }\n\n // The Wuffs API is that the original slice has no trailing data. It also\n // allows underscores, which we don't catch here but the fallback should.\n if (p != &z[s.len]) {\n goto fallback;\n }\n\n // Check that the uint64_t typed man variable has not overflowed, based on\n // digit_count.\n //\n // For reference:\n // - (1 << 63) is 9223372036854775808, which has 19 decimal digits.\n // - (1 << 64) is 18446744073709551616, which has 20 decimal digits.\n // - 19 nines, 9999999999999999999, is 0x8AC7230489E7FFFF, which has 64\n // bits and 16 hexadecimal digits.\n // - 20 nines, 99999999999999999999, is 0x56BC75" + + "E2D630FFFFF, which has 67\n // bits and 17 hexadecimal digits.\n if (digit_count > 19) {\n // Even if we have more than 19 pseudo-digits, it's not yet definitely an\n // overflow. Recall that digit_count might be off-by-one (too large) if\n // there's a decimal separator. It will also over-report the number of\n // meaningful digits if the input looks something like \"0.000dddExxx\".\n //\n // We adjust by the number of leading '0's and '.'s and re-compare to 19.\n // Once again, technically, we could skip ','s too, but that perturbs the\n // \"script/process-json-numbers.c with -p\" benchmark.\n const uint8_t* q = start_of_digits_ptr;\n for (; (*q == '0') || (*q == '.'); q++) {\n }\n digit_count -= (uint32_t)(q - start_of_digits_ptr);\n if (digit_count > 19) {\n goto fallback;\n }\n }\n\n // The wuffs_base__private_implementation__parse_number_f64_eisel_lemire\n // preconditions include that exp10 is in the range -326 ..= 310.\n if ((ex" + + "p10 < -326) || (310 < exp10)) {\n goto fallback;\n }\n\n // If man and exp10 are small enough, all three of (man), (10 ** exp10) and\n // (man ** (10 ** exp10)) are exactly representable by a double. We don't\n // need to run the Eisel-Lemire algorithm.\n if ((-22 <= exp10) && (exp10 <= 22) && ((man >> 53) == 0)) {\n double d = (double)man;\n if (exp10 >= 0) {\n d *= wuffs_base__private_implementation__f64_powers_of_10[+exp10];\n } else {\n d /= wuffs_base__private_implementation__f64_powers_of_10[-exp10];\n }\n wuffs_base__result_f64 ret;\n ret.status.repr = NULL;\n ret.value = negative ? -d : +d;\n return ret;\n }\n\n // The wuffs_base__private_implementation__parse_number_f64_eisel_lemire\n // preconditions include that man is non-zero. Parsing \"0\" should be caught\n // by the \"If man and exp10 are small enough\" above, but \"0e99\" might not.\n if (man == 0) {\n goto fallback;\n }\n\n // Our man and exp10 are in range. Run the Eisel-Lem" + + "ire algorithm.\n int64_t r =\n wuffs_base__private_implementation__parse_number_f64_eisel_lemire(\n man, exp10);\n if (r < 0) {\n goto fallback;\n }\n wuffs_base__result_f64 ret;\n ret.status.repr = NULL;\n ret.value = wuffs_base__ieee_754_bit_representation__from_u64_to_f64(\n ((uint64_t)r) | (((uint64_t)negative) << 63));\n return ret;\n } while (0);\n\nfallback:\n do {\n wuffs_base__private_implementation__high_prec_dec h;\n wuffs_base__status status =\n wuffs_base__private_implementation__high_prec_dec__parse(&h, s,\n options);\n if (status.repr) {\n return wuffs_base__private_implementation__parse_number_f64_special(\n s, options);\n }\n return wuffs_base__private_implementation__high_prec_dec__to_f64(&h,\n options);\n } while (0);\n}\n\n" + "" + "// --------\n\nstatic inline size_t //\nwuffs_base__private_implementation__render_inf(wuffs_base__slice_u8 dst,\n bool neg,\n uint32_t options) {\n if (neg) {\n if (dst.len < 4) {\n return 0;\n }\n wuffs_base__store_u32le__no_bounds_check(dst.ptr, 0x666E492D); // '-Inf'le.\n return 4;\n }\n\n if (options & WUFFS_BASE__RENDER_NUMBER_XXX__LEADING_PLUS_SIGN) {\n if (dst.len < 4) {\n return 0;\n }\n wuffs_base__store_u32le__no_bounds_check(dst.ptr, 0x666E492B); // '+Inf'le.\n return 4;\n }\n\n if (dst.len < 3) {\n return 0;\n }\n wuffs_base__store_u24le__no_bounds_check(dst.ptr, 0x666E49); // 'Inf'le.\n return 3;\n}\n\nstatic inline size_t //\nwuffs_base__private_implementation__render_nan(wuffs_base__slice_u8 dst) {\n if (dst.len < 3) {\n return 0;\n }\n wuffs_base__store_u24le__no_bounds_check(dst.ptr, 0x4E614E); // 'NaN'le.\n return 3;\n}\n\nstatic size_t //\nwuffs_base__private_implementation__high" + "_prec_dec__render_exponent_absent(\n wuffs_base__slice_u8 dst,\n wuffs_base__private_implementation__high_prec_dec* h,\n uint32_t precision,\n uint32_t options) {\n size_t n = (h->negative ||\n (options & WUFFS_BASE__RENDER_NUMBER_XXX__LEADING_PLUS_SIGN))\n ? 1\n : 0;\n if (h->decimal_point <= 0) {\n n += 1;\n } else {\n n += (size_t)(h->decimal_point);\n }\n if (precision > 0) {\n n += precision + 1; // +1 for the '.'.\n }\n\n // Don't modify dst if the formatted number won't fit.\n if (n > dst.len) {\n return 0;\n }\n\n // Align-left or align-right.\n uint8_t* ptr = (options & WUFFS_BASE__RENDER_NUMBER_XXX__ALIGN_RIGHT)\n ? &dst.ptr[dst.len - n]\n : &dst.ptr[0];\n\n // Leading \"±\".\n if (h->negative) {\n *ptr++ = '-';\n } else if (options & WUFFS_BASE__RENDER_NUMBER_XXX__LEADING_PLUS_SIGN) {\n *ptr++ = '+';\n }\n\n // Integral digits.\n if (h->decimal_point <= 0) {\n *ptr++ = '0';\n } else {\n uint32_t m =\n" + diff --git a/release/c/wuffs-unsupported-snapshot.c b/release/c/wuffs-unsupported-snapshot.c index 98fdcc1e..02cf2767 100644 --- a/release/c/wuffs-unsupported-snapshot.c +++ b/release/c/wuffs-unsupported-snapshot.c @@ -11939,7 +11939,7 @@ wuffs_base__private_implementation__high_prec_dec__to_f64( if (h->digits[0] >= 5) { break; } - shift = (h->digits[0] <= 2) ? 2 : 1; + shift = (h->digits[0] < 2) ? 2 : 1; } else { uint32_t n = (uint32_t)(-h->decimal_point); shift = (n < num_powers) diff --git a/script/print-render-number-f64-tests.go b/script/print-render-number-f64-tests.go index 7844670e..29fa3bc4 100644 --- a/script/print-render-number-f64-tests.go +++ b/script/print-render-number-f64-tests.go @@ -156,6 +156,7 @@ var u64TestCases = []uint64{ 0x4037000000000000, 0x4038000000000000, 0x40FE240C9FCB0C02, + 0x41E0246690000001, 0x4202A05F20000000, 0x4330000000000000, 0x4330000000000001, @@ -166,6 +167,7 @@ var u64TestCases = []uint64{ 0x4340000000000001, 0x4340000000000002, 0x4370000000000000, + 0x43F002F1776DDA67, 0x4415AF1D78B58C40, 0x44B52D02C7E14AF6, 0x46293E5939A08CEA, diff --git a/test/c/std/json.c b/test/c/std/json.c index d83c3b07..c1433997 100644 --- a/test/c/std/json.c +++ b/test/c/std/json.c @@ -1005,6 +1005,8 @@ test_wuffs_strconv_parse_number_f64_regular() { {.want = 0x4038000000000000, .str = "24"}, {.want = 0x4038000000000000, .str = "2400_00000_00000.00000_e-_1_2"}, {.want = 0x40FE240C9FCB0C02, .str = "123456.789012"}, + {.want = 0x41E0246690000001, + .str = "2.16656806400000023841857910156251e9"}, {.want = 0x4202A05F20000000, .str = "1e10"}, {.want = 0x4330000000000000, .str = "4503599627370496"}, // 1 << 52. {.want = 0x4330000000000000, .str = "4503599627370496.5"}, @@ -1020,6 +1022,7 @@ test_wuffs_strconv_parse_number_f64_regular() { {.want = 0x4340000000000002, .str = "9007199254740996"}, {.want = 0x4340000000000002, .str = "9_007__199_254__740_996"}, {.want = 0x4370000000000000, .str = "7.2057594037927933e+16"}, + {.want = 0x43F002F1776DDA67, .str = "18459999196907202592"}, {.want = 0x4415AF1D78B58C40, .str = "1e20"}, {.want = 0x44B52D02C7E14AF6, .str = "1e23"}, {.want = 0x46293E5939A08CEA, .str = "1e30"}, @@ -1655,6 +1658,15 @@ test_wuffs_strconv_render_number_f64() { .want_3f = "123456.789", .want_4g = "1.235e+05", }, + { + .x = 0x41E0246690000001, + .want__e = "2.1665680640000005e+09", + .want__f = "2166568064.0000005", + .want_0g = "2e+09", + .want_2e = "2.17e+09", + .want_3f = "2166568064.000", + .want_4g = "2.167e+09", + }, { .x = 0x4202A05F20000000, .want__e = "1e+10", @@ -1745,6 +1757,15 @@ test_wuffs_strconv_render_number_f64() { .want_3f = "72057594037927936.000", .want_4g = "7.206e+16", }, + { + .x = 0x43F002F1776DDA67, + .want__e = "1.8459999196907205e+19", + .want__f = "18459999196907205000", + .want_0g = "2e+19", + .want_2e = "1.85e+19", + .want_3f = "18459999196907204608.000", + .want_4g = "1.846e+19", + }, { .x = 0x4415AF1D78B58C40, .want__e = "1e+20",