Post #2409978

2025-09-09 21:38 UTC

@wren6991@types.pl @fanf@mendeddrum.org @JamesWidman@mastodon.social This is only in the Tremont version. In Gracemont and later the FE essentially inserts the dummy branches by itself.

Replies (2)

@rygorous@mastodon.gamedev.place 2025-09-09 21:42

@wren6991@types.pl @fanf@mendeddrum.org @JamesWidman@mastodon.social AFAICT on first execute of a BB (so running with cold I$) you really do just get 3-wide decode, but the FE automatically inserts the dummy branches into the BTB as it's decoding long BBs, and second and later executions don't suffer this penalty. This is structurally very similar to what happens on the bigger cores for long BBs with 4-wide decoders and only the uOp cache being wide.

Open ##2409979
@wren6991@types.pl 2025-09-09 21:42

@rygorous@mastodon.gamedev.place @fanf@mendeddrum.org @JamesWidman@mastodon.social Oh that's weird, how does it know where to insert them? I assumed the point of the ping-ponging was to find multiple instruction starts without having to sum up the previous instruction lengths.

Open ##2409982