more float to string performance improvements

- some tuples used in grisu were heap allocated. the `buffer`
  element was not really needed.
- fixed a type instability in Grisu.normalize
- provide a good size hint for floats in print_to_string to
  avoid excess reallocation.

before:

```
julia> @time for i=1:10^7;string(3.141592653);end
  6.402762 seconds (80.00 M allocations: 2.831 GB, 3.91% gc time)
```

after:

```
julia> @time for i=1:10^7;string(3.141592653);end
  4.038174 seconds (30.00 M allocations: 1.639 GB, 2.29% gc time)
```

base/grisu.jl

@@ -32,13 +32,13 @@ function grisu(v::AbstractFloat,mode,requested_digits,buffer=DIGITS,bignums=BIGN
     if mode == PRECISION && requested_digits == 0
         buffer[1] = 0x00
         len = 0
-        return 0, 0, neg, buffer
+        return 0, 0, neg
     end
     if v == 0.0
         buffer[1] = 0x30
         buffer[2] = 0x00
         len = point = 1
-        return len, point, neg, buffer
+        return len, point, neg
     end
     if mode == SHORTEST
         status,len,point = fastshortest(v,buffer)
@@ -47,9 +47,9 @@ function grisu(v::AbstractFloat,mode,requested_digits,buffer=DIGITS,bignums=BIGN
     elseif mode == PRECISION
         status,len,point = fastprecision(v,requested_digits,buffer)
     end
-    status && return len-1, point, neg, buffer
+    status && return len-1, point, neg
     status, len, point = bignumdtoa(v,mode,requested_digits,buffer,bignums)
-    return len-1, point, neg, buffer
+    return len-1, point, neg
 end
 
 _show(io::IO, x::AbstractFloat, mode, n::Int, t) =
@@ -67,7 +67,7 @@ function _show(io::IO, x::AbstractFloat, mode, n::Int, typed, nanstr, infstr)
         return
     end
     typed && isa(x,Float16) && write(io, "Float16(")
-    len,pt,neg,buffer = grisu(x,mode,n)
+    (len,pt,neg),buffer = grisu(x,mode,n),DIGITS
     pdigits = pointer(buffer)
     if mode == PRECISION
         while len > 1 && buffer[len] == 0x30
@@ -138,7 +138,7 @@ function _print_shortest(io::IO, x::AbstractFloat, dot::Bool, mode, n::Int)
     isnan(x) && return write(io, "NaN")
     x < 0 && write(io,'-')
     isinf(x) && return write(io, "Inf")
-    len,pt,neg,buffer = grisu(x,mode,n)
+    (len,pt,neg),buffer = grisu(x,mode,n),DIGITS
     pdigits = pointer(buffer)
     e = pt-len
     k = -9<=e<=9 ? 1 : 2

base/grisu/float.jl

@@ -45,7 +45,7 @@ const FloatSignificandSize = Int32(64)
 
 function normalize(v::Float)
     f = v.s
-    e = v.e
+    e::Int32 = v.e
     while (f & Float10MSBits) == 0
         f <<= 10
         e -= 10

base/printf.jl

@@ -909,7 +909,7 @@ function decode_dec(x::SmallFloatingPoint)
         DIGITS[1] = '0'
         return (Int32(1), Int32(1), false)
     end
-    len,pt,neg,buffer = grisu(x,Grisu.FIXED,0)
+    len,pt,neg = grisu(x,Grisu.FIXED,0)
     if len == 0
         DIGITS[1] = '0'
         return (Int32(1), Int32(1), false)
@@ -935,7 +935,7 @@ fix_dec(x::Integer, n::Int) = decode_dec(x)
 
 function fix_dec(x::SmallFloatingPoint, n::Int)
     if n > length(DIGITS)-1; n = length(DIGITS)-1; end
-    len,pt,neg,buffer = grisu(x,Grisu.FIXED,n)
+    len,pt,neg = grisu(x,Grisu.FIXED,n)
     if len == 0
         DIGITS[1] = '0'
         return (Int32(1), Int32(1), neg)
@@ -989,7 +989,7 @@ function ini_dec(x::SmallFloatingPoint, n::Int)
         ccall(:memset, Ptr{Void}, (Ptr{Void}, Cint, Csize_t), DIGITS, '0', n)
         return Int32(1), Int32(1), signbit(x)
     else
-        len,pt,neg,buffer = grisu(x,Grisu.PRECISION,n)
+        len,pt,neg = grisu(x,Grisu.PRECISION,n)
     end
     return Int32(len), Int32(pt), neg
 end

base/strings/io.jl

@@ -43,9 +43,14 @@ function sprint(size::Integer, f::Function, args...; env=nothing)
 end
 sprint(f::Function, args...) = sprint(0, f, args...)
 
+tostr_sizehint(x) = 0
+tostr_sizehint(x::AbstractString) = endof(x)
+tostr_sizehint(x::Float64) = 20
+tostr_sizehint(x::Float32) = 12
+
 function print_to_string(xs...; env=nothing)
     # specialized for performance reasons
-    s = IOBuffer(Array(UInt8,isa(xs[1],AbstractString) ? endof(xs[1]) : 0), true, true)
+    s = IOBuffer(Array(UInt8,tostr_sizehint(xs[1])), true, true)
     # specialized version of truncate(s,0)
     s.size = 0
     s.ptr = 1