use dragonbox algorithm; alternative to #18008 (#18139)

* use dragonbox algorithm; alternative to #18008
* removed unsafe code

changelog.md

@@ -76,6 +76,10 @@
 - `json` and `jsonutils` now serialize NaN, Inf, -Inf as strings, so that
   `%[NaN, -Inf]` is the string `["nan","-inf"]` instead of `[nan,-inf]` which was invalid json.
 
+- `system.addFloat` now uses the "Dragonbox" algorithm, which ensures correct roundtrips of floating point
+  numbers, that the minimum length representation of a floating point number is used and correct rounding.
+  Use `-d:nimLegacyAddFloat` for a transition period.
+
 - `strformat` is now part of `include std/prelude`.
 
 - Deprecated `proc reversed*[T](a: openArray[T], first: Natural, last: int): seq[T]` in `std/algorithm`.

compiler/condsyms.nim

@@ -134,3 +134,5 @@ proc initDefines*(symbols: StringTableRef) =
   defineSymbol("nimHasUnifiedTuple")
   defineSymbol("nimHasIterable")
   defineSymbol("nimHasTypeofVoid")
+
+  defineSymbol("nimHasDragonBox")

lib/system/formatfloat.nim

@@ -7,53 +7,58 @@
 #    distribution, for details about the copyright.
 #
 
-proc c_sprintf(buf, frmt: cstring): cint {.header: "<stdio.h>",
-                                    importc: "sprintf", varargs, noSideEffect.}
+when not defined(nimLegacyAddFloat) and not defined(nimscript) and
+    not defined(js) and defined(nimHasDragonBox):
+  import dragonbox
 
-proc writeToBuffer(buf: var array[65, char]; value: cstring) =
-  var i = 0
-  while value[i] != '\0':
-    buf[i] = value[i]
-    inc i
+  proc writeFloatToBuffer*(buf: var array[65, char]; value: BiggestFloat): int =
+    ## This is the implementation to format floats.
+    ##
+    ## returns the amount of bytes written to `buf` not counting the
+    ## terminating '\0' character.
+    result = toChars(buf, value, forceTrailingDotZero=true)
+    buf[result] = '\0'
 
-proc writeFloatToBuffer*(buf: var array[65, char]; value: BiggestFloat): int =
-  ## This is the implementation to format floats in the Nim
-  ## programming language. The specific format for floating point
-  ## numbers is not specified in the Nim programming language and
-  ## might change slightly in the future, but at least wherever you
-  ## format a float, it should be consistent.
-  ##
-  ## returns the amount of bytes written to `buf` not counting the
-  ## terminating '\0' character.
-  ##
-  ## * `buf` - A buffer to write into. The buffer does not need to be
-  ##           initialized and it will be overridden.
-  ##
-  var n: int = c_sprintf(addr buf, "%.16g", value)
-  var hasDot = false
-  for i in 0..n-1:
-    if buf[i] == ',':
-      buf[i] = '.'
-      hasDot = true
-    elif buf[i] in {'a'..'z', 'A'..'Z', '.'}:
-      hasDot = true
-  if not hasDot:
-    buf[n] = '.'
-    buf[n+1] = '0'
-    buf[n+2] = '\0'
-    result = n + 2
-  else:
-    result = n
-  # On Windows nice numbers like '1.#INF', '-1.#INF' or '1.#NAN' or 'nan(ind)'
-  # of '-1.#IND' are produced.
-  # We want to get rid of these here:
-  if buf[n-1] in {'n', 'N', 'D', 'd', ')'}:
-    writeToBuffer(buf, "nan")
-    result = 3
-  elif buf[n-1] == 'F':
-    if buf[0] == '-':
-      writeToBuffer(buf, "-inf")
-      result = 4
+else:
+  proc c_sprintf(buf, frmt: cstring): cint {.header: "<stdio.h>",
+                                      importc: "sprintf", varargs, noSideEffect.}
+
+  proc writeToBuffer(buf: var array[65, char]; value: cstring) =
+    var i = 0
+    while value[i] != '\0':
+      buf[i] = value[i]
+      inc i
+
+  proc writeFloatToBuffer*(buf: var array[65, char]; value: BiggestFloat): int =
+    ## This is the implementation to format floats.
+    ##
+    ## returns the amount of bytes written to `buf` not counting the
+    ## terminating '\0' character.
+    var n: int = c_sprintf(addr buf, "%.16g", value)
+    var hasDot = false
+    for i in 0..n-1:
+      if buf[i] == ',':
+        buf[i] = '.'
+        hasDot = true
+      elif buf[i] in {'a'..'z', 'A'..'Z', '.'}:
+        hasDot = true
+    if not hasDot:
+      buf[n] = '.'
+      buf[n+1] = '0'
+      buf[n+2] = '\0'
+      result = n + 2
     else:
-      writeToBuffer(buf, "inf")
+      result = n
+    # On Windows nice numbers like '1.#INF', '-1.#INF' or '1.#NAN' or 'nan(ind)'
+    # of '-1.#IND' are produced.
+    # We want to get rid of these here:
+    if buf[n-1] in {'n', 'N', 'D', 'd', ')'}:
+      writeToBuffer(buf, "nan")
       result = 3
+    elif buf[n-1] == 'F':
+      if buf[0] == '-':
+        writeToBuffer(buf, "-inf")
+        result = 4
+      else:
+        writeToBuffer(buf, "inf")
+        result = 3

tests/float/tfloat6.nim

@@ -1,10 +1,10 @@
 discard """
   output: '''
-1e-06 : 1e-06
-1e-06 : 1e-06
+0.000001 : 0.000001
+0.000001 : 0.000001
 0.001 : 0.001
-1e-06 : 1e-06
-1e-06 : 1e-06
+0.000001 : 0.000001
+0.000001 : 0.000001
 10.000001 : 10.000001
 100.000001 : 100.000001
 '''