Conversation
|
Tagging subscribers to this area: @BrzVlad, @kotlarmilos |
|
Add intrinsic fixed, thanks @lambdageek for spotting the bug |
|
Does this have any measurable impact ? |
We generate ~100 fewer methods during blazor startup. My timing data is very noisy, so I can't say 'this makes startup 1ms faster' with any confidence. |
|
The interp PGO tests were relying on us tiering one of these methods, oops :-) |
d80dc29 to
d970374
Compare
|
The osx interp failures look real, but I can't make sense of them. Does anyone have any ideas for why it would only be failing there? |
d970374 to
c37ad60
Compare
c37ad60 to
71cf842
Compare
|
I can't reproduce the interp failures in either debug or release configurations. I'm doing: but that failing suite along with other suites like S.R.T passes. |
Don't allocate unnecessary temp var. Don't generate mul if type size is 1.
During blazor startup these 3 methods account for a bit over 100 compiled methods if I analyzed the trace correctly. Since interp codegen dominates blazor startup (50-75% depending on how you measure it) my hope is that pruning methods out will reduce overall startup time. These three seemed like easy targets.
Add is commented out because I can't get it to work. Will be relying on help for that one unless a flash of insight helps me figure out how to fix it.Should I add a CEQ0_I8 for IsNullRef? I'm not sure it's worth adding a new opcode just for that.