Use PGO for Optimize Bools and If Conversion

[a74nh]

Both Optmize Bools and If Conversion use hard coded cost values to decide whether to optimize. Instead, when available, both of these should use PGO data make the decision instead.

See #112945 (comment) for discussion

[neon-sunset]

While looking at this, it would have also been great to finally get rid of if conversion / conditional moves restriction inside the loops, at the very least for simple selects - it emits jumps around single movs which is terrible. It is a sorely outdated restriction which ruins the codegen quality when compared to LLVM despite RyuJIT being capable of producing competitive output.

[a74nh]

While looking at this, it would have also been great to finally get rid of if conversion / conditional moves restriction inside the loops, at the very least for simple selects - it emits jumps around single movs which is terrible. It is a sorely outdated restriction which ruins the codegen quality when compared to LLVM despite RyuJIT being capable of producing competitive output.

Agreed. The code was written very conservatively to make sure there were no perf regressions. It'd be good to re-think the approach.

[dotnet-policy-service]

Tagging subscribers to this area: @JulieLeeMSFT, @jakobbotsch
See info in area-owners.md if you want to be subscribed.

[am11]

For NativeAOT, the cost is not a concern. dotnet publish can run a second longer if it buys us better codegen.