cmd/evm: don't reuse state between iterations, show errors

[jwasinger]

Reusing state between benchmark iterations resulted in inconsistent results across runs, which surfaced in #30778 .

If these errors are triggered again, they should be printed to output. To ensure that the code invoking testing.Benchmark() can catch and output these errors, and then exit: I replace calls to t.B.Fatalf with setting an error which is consumed by the calling code.

The same consistency checks should probably be applied to the state test benchmarker. I am figuring that out now.

closes #30778 .

[holiman]

Some nits

[holiman]

LGTM

[holiman]

Even though you worked around the panic due to b.Fatalf, I think it would still be more "correct" to also invoke testing.Init() when you enter the if bench clause.

[xermicus]

@jwasinger @holiman

Why was this done here?

This makes it impossible to re-use the state dump. The account still gets inserted however instead of an address I now see this in the state dump:

"pre(0xdfeeda06f79105a5aa975e5ef88e1b24d0062f31b5f42850e35d9cd4596df77e)": {

Which is wrong (should be the address) and means I can no longer take the state dump and execute another call based on the state dump.

[MariusVanDerWijden]

Could you please open an issue with this, so we wont forget about it

[jwasinger]

Execution will modify the runtimeConfig.State. We wanted to ensure that the same prestate is used for executing every benchmark iteration, so we copy it here.

[xermicus]

@jwasinger thanks for pointing this out, got it. #31151 fixes this by only copying it during benchmarks.

[xermicus]

@MariusVanDerWijden since the fix looks obvious to me I decided to open a small PR instead (#31151).