Improve MemoryMarshal.Cast

[mikedn]

Avoid unnecessary checked and signed arithmetic. Handle special cases such as cast between same size types and from byte sized types, the JIT is unable to optimize these cases currently.

Contributes to #15492

It would be nice to address some of the problems this code exposes in the JIT. But that will take a while so for now let's fix the source code...

[mikedn]

This brings down the size of Marvin.ComputeHash from 445 to 310 bytes

Diff/before/after disassembly: https://gist.github.com/mikedn/026e692f8c5357b56f24dc2192d1abd9

[jkotas]

cc @AndyAyersMS

[jkotas]

@mikedn Thanks for doing this!

[AndyAyersMS]

Looks like this gives us maybe another 5-10% improvement in the longer hash computations. My machine currently shows shorter computations a bit slower but I can't do a completely clean perf test right now. Will rerun these later today.

[AndyAyersMS]

Test failures (example) look unrelated, so am going to retry.

@dotnet-bot retest Ubuntu x64 Checked Innerloop Build and Test
@dotnet-bot retest CentOS7.1 x64 Checked Innerloop Build and Test
@dotnet-bot retest OSX10.12 x64 Checked Innerloop Build and Test
@dotnet-bot retest Tizen armel Cross Checked Innerloop Build and Tes

[AndyAyersMS]

@dotnet-bot retest Tizen armel Cross Checked Innerloop Build and Test

[AndyAyersMS]

Still measure a bit of a slowdown for some cases but when looking at diffs it doesn't make a lot of sense. So perhaps my baseline data (which came from the runs I did for #16654) is stale.

(edit: reran the baseline and now it all looks good; note new results are now the leftmost column)

Old data below

[mikedn]

Cool, need to figure out how to run these tests myself :)

[AndyAyersMS]

Just for the record, here's what I did -- only works if the bits you are changing live entirely in CoreCLR:

• build release bits for diff and base (I usually use two repos). usually build checked too.
• in base repo, selectively build one perf test from the test tree
• from repo root, run tests\scripts\run-xunit-perf.py -testBinLoc <path to that test's dir> to setup the perf test sandbox (which will be bin\sandbox)
• compile your new custom perf test (I copied the one from corefx).
• copy the test binary into bin\sandbox
• cd to the sandbox and corerun perfharness.dll myperftest.exe
• switch to diff repo and repeat above to get diff data (or swap diff parts into base sandbox)
• hoover the two csv that perfharness creates into excel and pivot to get comparative view (need to add a custom item to get ratio, it's not a built in option)

[adamsitnik]

Hi @mikedn

I am currently profiling XmlSerializer deserialization and I can see that your change has improved the perf for my particular scenario (https://github.com/dotnet/corefx/issues/27636) by 2,8%

I have marked the Cast method with black circles, on the left is after, on the right before your change . Sorry for unnatural order ;)

Nice improvement!