Repository navigation
Top-of-stack caching in the interpreter #131498
Description
Activity
- added a commit that references this issue
on Apr 14, 2025 - added a commit that references this issue
on Apr 18, 2025 It turns out that caching a single value in the interpreter has no positive effect on performance. Probably what is happening is that any speedup from improved IPC is being undone by increased overhead of spilling around escapes.
However, all the work done is necessary for #135379 which will improve performance.
Top-of-stack caching is technique where one or more values on the top of the stack are kept in local variables, with the assumption that the C compiler will put them in registers.
We should implement this for the interpreter (tier 1).
We cannot afford to increase the number of instructions by much, so we'll have to use a fixed size cache of 1.
This will likely only produce a modest speedup, as the total memory traffic is unlikely to decrease. There should be some ILP improvements.
The main benefit of TOS caching is likely to be in the JIT where we can use a variable sized cache.
We want to do it in the interpreter first though, as it will be simpler and requires a subset of the changes needed for the JIT.
Linked PRs
bytecodes.c#132615