Evaluate using Profile-Guided Optimization (PGO) for the compiler

[zamazan4ik]

Evaluate using Profile-Guided Optimization (PGO) for the compiler

[developedby]

I had no idea that this was a builtin option in the Rust compiler, very cool.

Some thoughts about this:

• Adding PGO to the Bend build looks very promissing, probably will be done at some point.
• Bend goes out of the way to have as few dependencies as possible, especially at its core. So I doubt link-time optimizations will improve too much.
• The Bend compiler is NOT optimized, almost at all. Our focus so far has been on correctness and simplicity. Before investigating compile optimizations, we'll have to profile and optimize the logic of Bend itself, which I imagine will bring much greater improvements.
• I'm much more interested in the performance this can bring to the runtime. We should try doing this to the C runtime. I imagine that for CUDA, PGO is not really a thing?

[zamazan4ik]

Bend goes out of the way to have as few dependencies as possible, especially at its core. So I doubt link-time optimizations will improve too much.

Probably - at least it should improve the binary size. This optimization is already enabled (so I changed nothing for this in the build scripts).

The Bend compiler is NOT optimized, almost at all. Our focus so far has been on correctness and simplicity. Before investigating compile optimizations, we'll have to profile and optimize the logic of Bend itself, which I imagine will bring much greater improvements.

Yep. However, even after performing manual optimizations with a higher probability PGO still will be useful. All modern "mature" compilers like Clang, Rustc, GCC, and many others still have large performance benefits from enabling PGO for them.

I'm much more interested in the performance this can bring to the runtime. We should try doing this to the C runtime. I imagine that for CUDA, PGO is not really a thing?

It depends. If we talk about GPU code - yes, PGO is not a thing yet for it. If we talk about the CPU part - PGO should work as for regular non-CUDA code.