summary refs log tree commit diff stats
path: root/lib/system/sysstr.nim
diff options
context:
space:
mode:
Diffstat (limited to 'lib/system/sysstr.nim')
-rw-r--r--lib/system/sysstr.nim141
1 files changed, 104 insertions, 37 deletions
diff --git a/lib/system/sysstr.nim b/lib/system/sysstr.nim
index f8b93a2c3..b1a3bee55 100644
--- a/lib/system/sysstr.nim
+++ b/lib/system/sysstr.nim
@@ -304,43 +304,37 @@ proc nimFloatToStr(f: float): string {.compilerproc.} =
 proc strtod(buf: cstring, endptr: ptr cstring): float64 {.importc,
   header: "<stdlib.h>", noSideEffect.}
 
-var decimalPoint: char
-
-proc getDecimalPoint(): char =
-  result = decimalPoint
-  if result == '\0':
-    if strtod("0,5", nil) == 0.5: result = ','
-    else: result = '.'
-    # yes this is threadsafe in practice, spare me:
-    decimalPoint = result
-
 const
   IdentChars = {'a'..'z', 'A'..'Z', '0'..'9', '_'}
+  powtens =   [ 1e0,   1e1,  1e2,  1e3,  1e4,  1e5,  1e6,  1e7,  1e8,  1e9,
+                1e10, 1e11, 1e12, 1e13, 1e14, 1e15, 1e16, 1e17, 1e18, 1e19,
+                1e20, 1e21, 1e22]
 
 proc nimParseBiggestFloat(s: string, number: var BiggestFloat,
                           start = 0): int {.compilerProc.} =
-  # This routine leverages `strtod()` for the non-trivial task of
-  # parsing floating point numbers correctly. Because `strtod()` is
-  # locale-dependent with respect to the radix character, we create
-  # a copy where the decimal point is replaced with the locale's
-  # radix character.
+  # This routine attempt to parse float that can parsed quickly.
+  # ie whose integer part can fit inside a 53bits integer.
+  # their real exponent must also be <= 22. If the float doesn't follow
+  # these restrictions, transform the float into this form:
+  #  INTEGER * 10 ^ exponent and leave the work to standard `strtod()`.
+  # This avoid the problems of decimal character portability.
+  # see: http://www.exploringbinary.com/fast-path-decimal-to-floating-point-conversion/
   var
     i = start
     sign = 1.0
-    t: array[500, char] # flaviu says: 325 is the longest reasonable literal
-    ti = 0
-    hasdigits = false
-
-  template addToBuf(c) =
-    if ti < t.high:
-      t[ti] = c; inc(ti)
+    kdigits, fdigits = 0
+    exponent: int
+    integer: uint64
+    fraction: uint64
+    frac_exponent= 0
+    exp_sign = 1
+    first_digit = 0
 
   # Sign?
   if s[i] == '+' or s[i] == '-':
     if s[i] == '-':
       sign = -1.0
-    t[ti] = s[i]
-    inc(i); inc(ti)
+    inc(i)
 
   # NaN?
   if s[i] == 'N' or s[i] == 'n':
@@ -360,40 +354,113 @@ proc nimParseBiggestFloat(s: string, number: var BiggestFloat,
           return i+3 - start
     return 0
 
+  # Skip leading zero
+  while s[i] == '0':
+    inc(i)
+
+  if s[i] in {'0'..'9'}:
+      first_digit = (s[i].ord - '0'.ord)
   # Integer part?
   while s[i] in {'0'..'9'}:
-    hasdigits = true
-    addToBuf(s[i])
+    inc(kdigits)
+    integer = integer * 10'u64 + (s[i].ord - '0'.ord).uint64
     inc(i);
     while s[i] == '_': inc(i)
 
   # Fractional part?
   if s[i] == '.':
-    addToBuf(getDecimalPoint())
     inc(i)
+    # if no integer part, Skip leading zeros
+    if kdigits <= 0:
+      while s[i] in {'_','0'}:
+        inc(i)
+        inc(frac_exponent)
+
+    if s[i] in {'0'..'9'}:
+      first_digit = (s[i].ord - '0'.ord)
+    # get fractional part
     while s[i] in {'0'..'9'}:
-      hasdigits = true
-      addToBuf(s[i])
+      inc(fdigits)
+      inc(frac_exponent)
+      integer = integer * 10'u64 + (s[i].ord - '0'.ord).uint64
       inc(i)
       while s[i] == '_': inc(i)
-  if not hasdigits:
+
+  # if has no digits: return error
+  # test for special case 0.0'f
+  if kdigits + fdigits <= 0 and i == start:
     return 0
 
-  # Exponent?
   if s[i] in {'e', 'E'}:
-    addToBuf(s[i])
     inc(i)
-    if s[i] in {'+', '-'}:
-      addToBuf(s[i])
+    if s[i] == '+' or s[i] == '-':
+      if s[i] == '-':
+        exp_sign = -1
+
       inc(i)
     if s[i] notin {'0'..'9'}:
       return 0
     while s[i] in {'0'..'9'}:
-      addToBuf(s[i])
+      exponent = exponent * 10 + (ord(s[i]) - ord('0'))
       inc(i)
-      while s[i] == '_': inc(i)
-  number = strtod(t, nil)
+      while s[i] == '_': inc(i) # underscores are allowed and ignored
+
+  var real_exponent = exp_sign*exponent - frac_exponent
+  let exp_negative = real_exponent < 0
+  var abs_exponent = abs(real_exponent)
+
+  # if exponent greater than can be represented: +/- zero or infinity
+  if abs_exponent > 999:
+    if exp_negative:
+      number = 0.0*sign
+    else:
+      number = Inf*sign
+    return i - start
+
+  # if integer is representable in 53 bits:  fast path
+  # max fast path integer is  1<<53 - 1 or  8999999999999999 (16 digits)
+  if kdigits + fdigits <= 16 and first_digit <= 8:
+    # max float power of ten with set bits above the 53th bit is 10^22
+    if abs_exponent <= 22:
+      if exp_negative:
+        number = sign * integer.float / powtens[abs_exponent]
+      else:
+        number = sign * integer.float * powtens[abs_exponent]
+      return i - start
+
+    # if exponent is greater try to fit extra exponent above 22 by multiplying
+    # integer part is there is space left.
+    let slop = 15 - kdigits - fdigits
+    if  abs_exponent <= 22 + slop and not exp_negative:
+      number = sign * integer.float * powtens[slop] * powtens[abs_exponent-slop]
+      return i - start
+
+  # if failed: slow path with strtod.
+  var t: array[500, char] # flaviu says: 325 is the longest reasonable literal
+  var ti = 0
+  let maxlen = t.high - "e+000".len # reserve enough space for exponent
+
   result = i - start
+  i = start
+  # re-parse without error checking, any error should be handled by the code above.
+  while s[i] in {'0'..'9','+','-'}:
+    if ti < maxlen:
+      t[ti] = s[i]; inc(ti)
+    inc(i)
+    while s[i] in {'.', '_'}: # skip underscore and decimal point
+      inc(i)
+
+  # insert exponent
+  t[ti] = 'E'; inc(ti)
+  t[ti] = if exp_negative: '-' else: '+'; inc(ti)
+  inc(ti, 3)
+
+  # insert adjusted exponent
+  t[ti-1] = ('0'.ord + abs_exponent mod 10).char; abs_exponent = abs_exponent div 10
+  t[ti-2] = ('0'.ord + abs_exponent mod 10).char; abs_exponent = abs_exponent div 10
+  t[ti-3] = ('0'.ord + abs_exponent mod 10).char
+
+  number = strtod(t, nil)
 
 proc nimInt64ToStr(x: int64): string {.compilerRtl.} =
   result = newString(sizeof(x)*4)
'>403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496