Separate RTS Builds (Incremental and Non-Incremental GC)

[luc-blaeser]

Separate RTS Builds (Incremental and Non-Incremental GC)

Using different memory layouts determined at compile time:

• Incremental GC:
  • Extended header with forwarding pointer field.
  • Partitioned heap.
• Non-incremental GC (copying, compacting, and generational GC):
  • Small header only comprising the object tag.
  • Linear heap.

Runtime System Changes

Separate RTS builds by introducing the feature "incremental_gc" for incremental GC memory layout.

Helper macros:

• #[incremental_gc]: macro attribute, equivalent to #[cfg(feature = "incremental_gc")]
• #[non_incremental_gc]: macro attribute, equivalent to #[cfg(not(feature = "incremental_gc"))]
• is_incremental_gc!(): procedure macros, equivalent to cfg!(feature = "incremental_gc")

Different builds:

• rts.wasm: Release build with non-incremental GCs (containing the copying, compacting, and generational GC).
• rts-debug.wasm: Debug build with non-incremental GCs (containing the copying, compacting, and generational GC).
• rts-incremental.wasm: Release build with only the incremental GC (no other GCs).
• rts-incremental-debug.wasm: Debug build with only the incremental GC (no other GCs).

Compiler Changes

GC-dependent compilation:

• Conditional header layout with or without forwarding pointer.
• Linking the corresponding matching RTS build, with build-specific imports.

Switch based on the condition !Flags.gc_strategy == Flags.Incremental.

Performance

Comparing the the following designs:

• Non-Incremental RTS: Original RTS in master branch without incremental GC changes.
• Combined RTS: Combining incremental and non-incremental GC in one RTS build, PR: Incremental GC #3837
• Separate RTS: This PR.

Binary Size

Size of the release RTS binary files.

GC	Non-Incremental RTS	Combined RTS	Separate RTS
non-incremental	174 KB	194 KB	174 KB
incremental	-	194 KB	175 KB

11% reduction.

Total allocations

GC benchmark results with dfx 0.13.1.
Total amount of allocated memory (heap size + reclaimed memory) at runtime, average across benchmark cases.

GC	Non-Incremental RTS	Combined RTS	Separate RTS
copying	459 MB	496 MB	459 MB
compacting	459 MB	496 MB	459 MB
generational	480 MB	517 MB	480 MB
incremental	-	502 MB	502 MB

For non-incremental GCs: 8% reduction compared to the combined RTS, same like non-incremental RTS.

Memory Size

GC benchmark results with dfx 0.13.1.
Allocated WASM memory size at runtime, average across benchmark cases.

GC	Non-Incremental RTS	Combined RTS	Separate RTS
copying	271 MB	281 MB	271 MB
compacting	188 MB	195 MB	188 MB
generational	191 MB	201 MB	194 MB
incremental	-	294 MB	294 MB

For non-incremental GCs: 4% reduction compared to the combined RTS, same like non-incremental RTS.

Total Instructions

GC benchmark results with dfx 0.13.1.
Number of executed instructions, average across benchmark cases.

GC	Non-Incremental RTS	Combined RTS	Separate RTS
copying	2.05e10	2.12e10	2.07e10
compacting	2.20e10	2.24e10	2.21e10
generational	1.91e10	1.93e10	1.92e10
incremental	-	1.95e10	1.95e10

For non-incremental GCs: Around 2% reduction compared to combined RTS, around 1%* overhead compared to non-incremental RTS.

Conclusion

Advantages of this PR:

• Avoiding performance degrades for the classical GCs by introducing the incremental GC support.
• Smaller binary sizes.

[crusso]

But don't forget to fix those uses of physical equality