In C++, there are multiple ways to convert strings into numbers. E.g., to convert the string “1.0e2” into the number 100.

In software, we frequently have to parse numbers from strings. Numbers are typically represented in computers as 32-bit or 64-bit words whereas strings are variable-length sequences of bytes. We need to go from one representation to the other.

For example, given the 3-character string “123”, we want to recover the integer value 123, which is typically stored as a byte value 0x7b followed by some zero bytes.

Though parsing integers is not trivial, parsing floating-point numbers is trickier: e.g., the strings “1e10”, “10e9”, “100e+8” and “10000000000” all map to the same value. Typically, we store floating-point values as either 32-bit or 64-bit values using the IEEE 754 standard. Parsing may fail in various fun ways even if we set aside garbage inputs that have no reasonable number value (like “4a.14x.10.14”). For example, the string “1e500” cannot be represented, typically, in software (except as the infinity). The string “18446744073709551616” cannot be represented as either a 64-bit floating-point value or a 64-bit integer. In some instances, we need to “round” the value: the string “1.0000000000000005” cannot be represented exactly as a 64-bit floating-point value, so we need to round it down to “1.0000000000000004” while we might round “1.0000000000000006” up to 1.0000000000000007.

How expensive is it to parse numbers? To test it out, I wrote a long string containing space-separated numbers. In one case, I generated random 64-bit unsigned integers, and another I generated random 64-bit normal floating-point numbers. Then I benchmark standard C++ parsing code that simply sums up the numbers:

std::stringstream in(mystring); while(in >> x) { sum += x; } return sum;

I use long strings (100,000 numbers), and the GNU GCC 8.3 compiler on an Intel Skylake processor. I find that parsing numbers is relatively slow:

integers | 360 cycles/integer | 18 cycles/byte | 200 MB/s |

floats | 1600 cycles/integer | 66 cycles/byte | 50 MB/s |

Most disks and most networks can do much better than 50 MB/s. And good disks and good networks can beat 200 MB/s by an order of magnitude.