Stage2/3 Causes 10x Drop in Performance

[ominitay]

Zig Version

0.10.0-dev.3659+e5e6eb983

Steps to Reproduce

Extract https://0x0.st/o9zs.tar.xz into a subdirectory.
Build the attached project with stage1, and then with stage2/3, with the command zig build -Drelease-fast. Run each executable, and enter the command go depth 7. The number after nps indicates the speed of the program in nodes per second.

Expected Behavior

The values for nps would be roughly equal, or stage2/3 would produce a faster executable.

Actual Behavior

Stage2/3 produces an executable approximately 10x slower than stage1 does.

[ominitay]

For reference, my output:
stage2/3

go depth 7
info depth 1 nodes 21 nps 21000 score cp 36 hashfull 0
...
info depth 7 nodes 516601 nps 210000 score cp 28 hashfull 588

stage1

go depth 7
info depth 1 nodes 21 nps 21000 score cp 36 hashfull 0
...
info depth 7 nodes 516557 nps 1666312 score cp 28 hashfull 574

[kubkon]

Could you also specify arch and os please?

[ominitay]

Linux x64

[ominitay]

If you have the time to run go depth 8, the performance loss increases to just over 10x.

[notramo]

Linux aarch64 is also affected by this.
The stdlib Argon2 implementation is ≈13 times slower both on x86_64 and aarch64.

[andrewrk]

Can you please run your performance test again now that 892fb0f has landed?