Force 2 iterations of OptRepeat #100473
Force 2 iterations of OptRepeat #100473
Conversation
dotnet-issue-labeler
bot
added
the
area-CodeGen-coreclr
|
Tagging subscribers to this area: @JulieLeeMSFT, @jakobbotsch |
|
Ouch, quite expensive (@SingleAccretion predicted similar numbers)
But of course that is quite non-trivial to guess whether you might benefit from the 2nd iter or not. Maybe if we study these diffs we'll figure out what we benefit the most from in most cases? |
|
Code size improvement from about 3 to 4.5MB. TP cost of about 20%. |
Lots of good ideas here. I agree that we wouldn't want to run it unconditionally. But running on big methods might be the most beneficial. And maybe we'd be willing to take the 20% TP cost for NAOT compiles. |
|
cc @dotnet/jit-contrib Example of enabling JitOptRepeat always, with 2 iterations. Code size improvement from about 3 to 4.5MB. TP cost of about 20%. |
BruceForstall
force-pushed
the
FixOptRepeat_ForceRepeat2InRelease
branch
from
bea3f5f to
19c5d79
Compare
BruceForstall
changed the title
BruceForstall
force-pushed
the
FixOptRepeat_ForceRepeat2InRelease
branch
from
19c5d79 to
2c9dbda
Compare
BruceForstall
added
the
NO-MERGE
|
|
|
Here is a run of asmdiffs on win-x64 between JitOptRepeat 2 and 3 iterations, showing there is some benefit to be had to run more than twice. Diffs are based on 2,384,245 contexts (931,974 MinOpts, 1,452,271 FullOpts). MISSED contexts: base: 7,980 (0.33%), diff: 8,122 (0.34%) Base JIT options: JitOptRepeat#* Diff JIT options: JitOptRepeat#*;JitOptRepeatCount#3 Overall (-149,570 bytes)
FullOpts (-149,570 bytes)
|
|
Sounds like 2 iterations are enough? 🙂 It'd be nice to have a breakdown TP report like @jakobbotsch did here for example to know what is the most expensive thing |
|
Here's a run of asmdiffs on win-x64 between JitOptRepeat 3 and 4 iterations. There's still diffs. Diffs are based on 2,384,210 contexts (931,974 MinOpts, 1,452,236 FullOpts). MISSED contexts: base: 8,122 (0.34%), diff: 8,157 (0.34%) Base JIT options: JitOptRepeat#*;JitOptRepeatCount#3 Diff JIT options: JitOptRepeat#*;JitOptRepeatCount#4 Overall (-71,360 bytes)
FullOpts (-71,360 bytes)
|
|
Here's a run of asmdiffs on win-x64 between JitOptRepeat 4 and 5 iterations. Diffs are based on 2,384,205 contexts (931,974 MinOpts, 1,452,231 FullOpts). MISSED contexts: base: 8,157 (0.34%), diff: 8,162 (0.34%) Base JIT options: JitOptRepeat#*;JitOptRepeatCount#4 Diff JIT options: JitOptRepeat#*;JitOptRepeatCount#5 Overall (-14,538 bytes)
FullOpts (-14,538 bytes)
|
|
And just "for fun", here is asmdiffs on win-x64 between JitOptRepeat 8 and 9 iterations! Yes, there are still diffs. One question: are optimization changes somehow monotonic, or is it possible the optimizations are oscillating (e.g., one iteration creates a CSE, another removes it)? Diffs are based on 2,384,201 contexts (931,974 MinOpts, 1,452,227 FullOpts). MISSED contexts: 8,166 (0.34%) Base JIT options: JitOptRepeat#*;JitOptRepeatCount#8 Diff JIT options: JitOptRepeat#*;JitOptRepeatCount#9 Overall (+4,080 bytes)
FullOpts (+4,080 bytes)
|
No description provided.