4 ms·

Re. Go. I've turned this code into a Go benchmark and run it with both the default cmd/compile (1.11) and gccgo (8.2.0). cmd/compile: BenchmarkFib-4

Re. Go. I've turned this code into a Go benchmark and run it with both the default cmd/compile (1.11) and gccgo (8.2.0). cmd/compile: BenchmarkFib-4 1 16530409619 ns/op PASS ok command-line-arguments 16.533s gccgo: BenchmarkFib-4 2 776368644 ns/op PASS ok command-line-arguments 2.381s So yeah, if you are doing heavy-weight math stuff in Go, you might consider switching to gccgo. You might get up to 20x performance boost. Here is the gist, please tell me if I screwed up somewhere: https://gist.github.com/ainar-g/1bd363d41c441d9ebf05c0c0b9f2a529 https://gist.github.com/ainar-g/1bd363d41c441d9ebf05c0c0b9f2.... EDIT: After some disassembly and experimenting, it seems like what we see here is some clever unrolling and memoisation. If I change 46 from a constant to a variable, it becomes twice as slow. Plus there is this in disasm: /tmp/go/fibbench_test.go:13 return fib(n-1) + fib(n-2) 30b9: e8 72 ff ff ff callq 3030 <command_line_arguments.fib> 30be: bf 25 00 00 00 mov $0x25,%edi 30c3: e8 68 ff ff ff callq 3030 <command_line_arguments.fib> 30c8: bf 24 00 00 00 mov $0x24,%edi 30cd: e8 5e ff ff ff callq 3030 <command_line_arguments.fib> 30d2: bf 23 00 00 00 mov $0x23,%edi ... What puzzles me is why doesn't gcc do this for the C version. Even if I add an explicit __attribute__((const)).

FabHK 8y ago
So you're saying it would run in less than 1/4 of the time of the Nim, C, C++ implementations? Sure it's running the same algo?
- ainar-g 8y ago
  I am as confused as you are. The C version is much slower on my PC. I've added the link to the code above. Please tell me if you see something fishy.
ericpauley 8y ago
Microbenchmarks like this can be difficult to perform in practice, as gccgo can perform optimizations on pure functions that prevent Go's benchmarks from actually fully repeating a test. Also, this is a special case in which gccgo shines, specifically because it is completely cpu-bound. Generally, there isn't nearly as much of a performance difference.

FabHK 8y ago

So you're saying it would run in less than 1/4 of the time of the Nim, C, C++ implementations? Sure it's running the same algo?

ainar-g 8y ago
I am as confused as you are. The C version is much slower on my PC. I've added the link to the code above. Please tell me if you see something fishy.

ericpauley 8y ago

Microbenchmarks like this can be difficult to perform in practice, as gccgo can perform optimizations on pure functions that prevent Go's benchmarks from actually fully repeating a test. Also, this is a special case in which gccgo shines, specifically because it is completely cpu-bound. Generally, there isn't nearly as much of a performance difference.