sha256 hash function on Windows in Godot 3.4 takes close to 10x as long as it does to 3.3

[rickwillcox]

Godot version

3.4

System information

Windows10

Issue description

If I run this code on 3.3 / 3.4 linux or 3.3 windows, I get about 600ms taken to finish hashit().
If I run the same code (same project even), in 3.4 windows it takes around 5550ms to finish hashit()
I have tested this on two machines and also my friend has the same result. We found the bug in our project when we tried to move to 3.4.

func _ready() -> void:
	hashit()

func hashit():
	var start : int = OS.get_system_time_msecs()
	var string : String = "string to hash"
	var ROUNDS : int = pow(2,18)
	var rounds : int = ROUNDS
	while rounds > 0:
		string = string.sha256_text()
		rounds -= 1
	print("sha256 hash run %d times and took %dms" % [ROUNDS, OS.get_system_time_msecs()-start])

Steps to reproduce

run the code on 3.4 windows
compare to time on 3.3 windows

Minimal reproduction project

testhash.zip

[Calinou]

@rickwillcox Can you reproduce this in any of the 3.4 betas and RCs to determine when the regression started?

I can't reproduce this on Linux on an i7-6700K.

3.3.4.stable (official editor build)

sha256 hash run 262144 times and took 503ms

3.4.stable (official editor build)

sha256 hash run 262144 times and took 509ms

3.5.beta.custom_build [a1ebcbc] (own editor build with target=release_debug use_lto=yes)

sha256 hash run 262144 times and took 527ms

[rickwillcox]

We tested it the day 3.4 was released. We are not continuing with this implementation but thought the godot team should know.

[rickwillcox]

This issue is only on WINDOWS 3.4. Linux 3.3 /3.4 and WINDOWS 3.3 work fine.

[akien-mga]

String.sha256_text() calls CryptoCore::sha256(), which calls into mbedtls_sha256_ret. So this might be related to a mbedtls version update, like Mbed-TLS/mbedtls#4814. CC @Faless

@rickwillcox Could you test Godot 3.2.2 and 3.2.3 too? If my hunch is right, then 3.2.3 should be the first version with this regression, 3.2.2 should be fine (mbedtls 2.16.11 and 2.16.10 respectively).

[rickwillcox]

I've uploaded the file test file and code. I dont have time to test further.

[KoBeWi]

I checked with 3.2.2, 3.2.3 and 3.5 beta5. The MRP took 790-820 ms to execute, with no big differences between versions.
Tested on Windows 10

[akien-mga]

Assuming fixed, as we updated mbedTLS from 2.16.11 to 2.28.0 between 3.4 and 3.5, which might have solved potential performance regressions in 2.16. Please comment if you can still reproduce the issue so we can investigate further.