[Core] Optimize String::replace methods

[AThousandShips]

Performs a single allocation, only when any instances are found, and avoids concatenations and other unnecessary conversions.

Haven't done any benchmarking yet but as far as I know there's no optimization that can handle the concatenations in a way that avoids repeated allocations.

Will run the code through QuickBench soon with some of the data in the PR below but have to translate some of the code first.

Related:

• Add a string replace method to handle multiple search keys efficiently. #92433

[AThousandShips]

Since the code only differs on which find method to use I created a helper method to avoid code duplication

[AThousandShips]

Gonna write up a benchmark test in the next few days when I find the time to convert the code to something benchmarkable (won't necessarily catch some of the bottlenecks but should at least catch the benefits of pre-processing and memcpy)

[AThousandShips]

There managed to write some benchmarks now in fact, using this:
https://quick-bench.com/q/HgsGkbFKTnXifEkW1BJzDe4lXP8 (adapted from #92433 (comment))

Will do some other case tests as well to see exactly how it might improve, but this shows this specific case

The improvements will likely be a bit better thanks to other behind the scenes differences this can't test for (COW stuff primarily)

[AThousandShips]

The results in some cases are hard to evaluate, might be allocation based so hard to benchmark against them and consider the differences with the engine code, will try further benchmarks and see

[Calinou]

Tested locally, it works as expected. This provides a slight speedup in a real world scenario (using print_rich() lots of times, which currently relies on string replacement a lot).

Testing project: test_print_rich.zip

Benchmark

Using optimize=speed lto=full Linux editor binaries.

Command:

$ hyperfine 'bin/godot.linuxbsd.editor.x86_64.master --path /tmp/4 &> /dev/null' 'bin/godot.linuxbsd.editor.x86_64 --path /tmp/4 &> /dev/null'

Benchmark 1: bin/godot.linuxbsd.editor.x86_64.master --path /tmp/4 &> /dev/null
  Time (mean ± σ):     15.030 s ±  0.211 s    [User: 10.756 s, System: 3.243 s]
  Range (min … max):   14.651 s … 15.292 s    10 runs
 
Benchmark 2: bin/godot.linuxbsd.editor.x86_64 --path /tmp/4 &> /dev/null  <--- This PR
  Time (mean ± σ):     14.046 s ±  0.504 s    [User: 9.959 s, System: 3.257 s]
  Range (min … max):   13.138 s … 14.656 s    10 runs
 
Summary
  bin/godot.linuxbsd.editor.x86_64 --path /tmp/4 &> /dev/null ran
    1.07 ± 0.04 times faster than bin/godot.linuxbsd.editor.x86_64.master --path /tmp/4 &> /dev/null

PC specifications

• CPU: Intel Core i9-13900K
• GPU: NVIDIA GeForce RTX 4090
• RAM: 64 GB (2×32 GB DDR5-5800 C30)
• SSD: Solidigm P44 Pro 2 TB
• OS: Linux (Fedora 39)

[akien-mga]

For the record, the unit tests for String::replace seem a bit lightweight, might be worth expanding in a follow-up (maybe for more of the String API given the number of PRs that aim to optimize it, we need to make sure it's well tested to prevent regressions).

[akien-mga]

Thanks!

[AThousandShips]

Thank you!