perf: use buf in Args instead of bytebufferpool

[ksw2000]

Resolve #1856

Since Args already contains a buf field, and multiple functions are designed to operate based on it, using the buf in Args results in a shorter process compared to using bytebufferpool. Additionally, based on the usage scenario, the number of calls to SetUint is unlikely to exceed 10, or even 100. Under such circumstances, using buf is significantly more efficient than bytebufferpool.

Benchmark

Call SetUint multiple times with the same key to compare the performance of the AppendUint step.

func BenchmarkSetUint1(b *testing.B) {
	benchmarkSetUint(b, 1)
}

func BenchmarkSetUint10(b *testing.B) {
	benchmarkSetUint(b, 10)
}

func BenchmarkSetUint100(b *testing.B) {
	benchmarkSetUint(b, 100)
}

func benchmarkSetUint(b *testing.B, n int) {
	data := make([]int, n)

	for i := range data {
		data[i] = i
	}

	args := AcquireArgs()
	for i := 0; i < b.N; i++ {
		for j := range data {
			args.SetUint("key", data[j])
		}
		args.Reset()
	}
	ReleaseArgs(args)
}

Result

goos: linux
goarch: amd64
pkg: github.com/valyala/fasthttp
cpu: AMD EPYC 7763 64-Core Processor                
             │  old.txt    │              new.txt                │
             │   sec/op    │   sec/op     vs base                │
SetUint1-4     34.16n ± 3%   17.74n ± 1%  -48.06% (p=0.000 n=20)
SetUint10-4    336.0n ± 1%   169.0n ± 3%  -49.68% (p=0.000 n=20)
SetUint100-4   3.288µ ± 1%   1.640µ ± 1%  -50.13% (p=0.000 n=20)
geomean        335.4n        170.1n       -49.30%
…             │   old.txt    │              new.txt                │
             │  allocs/op   │ allocs/op   vs base                 │
SetUint1-4     0.000 ± 0%     0.000 ± 0%       ~ (p=1.000 n=20) ¹
SetUint10-4    0.000 ± 0%     0.000 ± 0%       ~ (p=1.000 n=20) ¹
SetUint100-4   0.000 ± 0%     0.000 ± 0%       ~ (p=1.000 n=20) ¹
geomean                   ²               +0.00%                ²
¹ all samples are equal
² summaries must be >0 to compute geomean

The results show that the implementation using buf is approximately 50% faster than using bytebufferpool.

[erikdubbelboer]

Thanks!