[wasm] More accurate jiterpreter cfg size estimation; generate smaller dispatch tables #83759
Conversation
Generate smaller dispatch tables for traces with backward branches
…n the back branch table
… it. Just log a warning Lower the generated limit slightly
|
Tagging subscribers to 'arch-wasm': @lewing Issue DetailsThe monitoring pass's addition of a back branch flag made the cfg size estimation less accurate, so this PR mostly fixes that. There is still a bit of underestimation for a few traces I haven't tracked down but for s.r.t I see it maybe underestimate 5 traces in total out of thousands, and it's not off by too much. This PR also changes the dispatch table generator to only generate dispatch entries for back branch targets specifically (it used to generate them for all branch targets, which is wasteful and incorrect). Currently if a trace managed to hit the 4kb limit despite our best efforts, we would try to compile it anyway and then fail, which would eventually cause the jiterpreter to be disabled. This PR adds an explicit check for that and logs a warning instead with the name of the trace, which will avoid scenarios where a whole application falls off a cliff due to a couple bad traces.
|
ghost
assigned 
The monitoring pass's addition of a back branch flag made the cfg size estimation less accurate, so this PR mostly fixes that. There is still a bit of underestimation for a few traces I haven't tracked down but for s.r.t I see it maybe underestimate 5 traces in total out of thousands, and it's not off by too much.
This PR also changes the dispatch table generator to only generate dispatch entries for back branch targets specifically (it used to generate them for all branch targets, which is wasteful and incorrect).
Currently if a trace managed to hit the 4kb limit despite our best efforts, we would try to compile it anyway and then fail, which would eventually cause the jiterpreter to be disabled. This PR adds an explicit check for that and logs a warning instead with the name of the trace, which will avoid scenarios where a whole application falls off a cliff due to a couple bad traces.