JIT: strength reduce _LOAD_FAST{_BORROW} to _LOAD_CONST_INLINE_BORROW

[Fidget-Spinner]

Feature or enhancement

Proposal:

The following code is common in some benchmarks, such as raytrace

x = 1
y = x

In raytrace, it happens in the form of the arguments being passed to the points.

We can strength reduce the y = x from a LOAD_FAST to a LOAD_CONST_INLINE_BORROW which is significantly faster in the JIT, as it eliminates a memory load from the heap.

Has this already been discussed elsewhere?

No response given

Links to previous discussion of this feature:

No response

Linked PRs

• gh-143117: Strength reduce _LOAD_FAST{_BORROW} to _LOAD_CONST_INLINE_BORROW #143118
• gh-143117: Constant fold LOAD_FAST_BORROW when possible #143631

[markshannon]

_LOAD_CONST_INLINE_BORROW could well be slower than _LOAD_FAST_BORROW_N_rXY as it needs to load data from the instruction stream instead of from the frame, making the generated code larger and increasing the icache pressure.
The local variable is guaranteed to be in the L1 dcache, so the load in LOAD_FAST_BORROW is as fast as possible.

I don't think this is a clear cut improvement, unless it enable further optimizations.

On Aarch64 this changes the jitted code size from 8 to 12 bytes. On x86-64 it changes it from 8 to 14 bytes.