tqchen commented on PR #780:
URL: https://github.com/apache/tvm-ffi/pull/780#issuecomment-5646470819

   Validation for `371a1991da04d7a37c02d323cb8d31561073982f`:
   
   - All 498 existing C++ tests passed; two preexisting tests remain disabled. 
AddressSanitizer instruments the test executable; the shared runtime is a 
normal Debug build. The local GCC build demoted the existing redundant-move 
warning from Werror.
   - The final diff preserves explicit `Optional<VisitInterrupt>` signatures 
and adds no tests. The only test-file change removes obsolete raw forwarding 
from an existing visitor layer.
   - All six optimized compiled hooks and the fixture's 17 logical 
occurrences/four Var matches were checked outside timing. AnyView pure Advance 
callback machinery erases; the typed Var-count callback retains its type probe 
and increment, while Advance result cleanup and callback-control branches 
disappear. Required child-result propagation remains.
   
   The latest GCC 11.4/O3 interleaved comparison used the same preserved 
executables and runtimes, with 12 processes per cell and all nine inner samples 
retained. Median ns per logical occurrence:
   
   | Executable source | Compiled-hook floor | Borrowed Var, Advance | Borrowed 
Var, count then Advance |
   | --- | ---: | ---: | ---: |
   | Initial optimization | 4.105 | 6.653 | 6.136 |
   | Reviewed optimization | 4.239 | 6.274 | 5.873 |
   
   The timed reviewed source precedes the final explicit-type/test cleanup. All 
11 selected engine disassembly blocks are exactly identical in the final 
source; subsequent base alignment preserved the entire source tree. These are 
not new timings of the final commit. An earlier run was slower for both 
versions and their floors; the preserved binaries recovering together supports 
transient run conditions, although the exact cause is unproven. No samples were 
trimmed.
   
   The AnyView hot paths save five registers without local result slots or 
cleanup calls; the typed fixture also retains a 32-byte local frame/canary. 
These measurements remain above the narrower compiled-hook floor. Counting and 
pure Advance have different generated bodies, so their difference is not an 
isolated counting cost.
   


-- 
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.

To unsubscribe, e-mail: [email protected]

For queries about this service, please contact Infrastructure at:
[email protected]


---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]

Reply via email to