cpu/gen - use hash of preprocessed source for JiT cache lookup

[zatkins-dev]

Purpose:

First attempt at actually using the cache. In my testing, the cost of preprocessing is < 10% of the compilation cost, making the preprocessed source a perfect target for hashing.

Remaining goals:

• Avoid writing .cpp to disk, instead pipe directly from memory (write to disk in debug mode?)
• Probably need to encode the compile options into the preprocessed source, maybe as a static string in the source?
• Add lock files or something for the .so files so that the first process can write and the others can wait.

Closes: #2038

LLM/GenAI Disclosure:

None

[zatkins-dev]

I feared weird compilers would have issues; ugh.

[jeremylt]

I like being able to just open the cpp for debugability so it should remain an optional feature if removed by default

[zatkins-dev]

I like being able to just open the cpp for debugability so it should remain an optional feature if removed by default

it's now only written in debug mode

[zatkins-dev]

jk there's a problem

[jeremylt]

I like being able to just open the cpp for debugability so it should remain an optional feature if removed by default

it's now only written in debug mode

Can you make CUDA and HIP mirror this behavior?

[zatkins-dev]

Can you make CUDA and HIP mirror this behavior?

I kinda blacked out for a bit and now the CUDA Clang JiT is cached and reused in the same way as the CPU. The impact is wayyyy more dramatic, since compiling for device is super slow.

[zatkins-dev]

Okay I think this is good to review, I'm happy with the current state.