core: Print fp literals losslessly

[n-io]

mlir-opt attempts to print floats as 6dp scientific notation, but round-trips to ensure there is no precision loss and the printed number will bitwise reproduce exactly. If this fails, it will choose a different printing method to ensure bit-reproducible printing for a given precision of fp types.

This PR introduces a flag to print with full precision. This should eventually become the default.
This PR prints with full precision by default, but provides a --print-reduced-precision-fp flag to enable the legacy behaviour.
This PR takes the approach of mlir-opt: attempt to print to scientific 6dp notation iff the resulting string is losslessly reproducible, or else print to full precision using repr. This has the advantage of not affecting a vast number of filechecks, mirroring upstream, and not having a flag for either full or reduced precision.

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 90.12%. Comparing base (5e2f5da) to head (0691a1d).
Report is 12 commits behind head on main.

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #3381      +/-   ##
==========================================
+ Coverage   90.11%   90.12%   +0.01%     
==========================================
  Files         450      450              
  Lines       56798    56826      +28     
  Branches     5459     5461       +2     
==========================================
+ Hits        51182    51215      +33     
+ Misses       4174     4167       -7     
- Partials     1442     1444       +2     

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[alexarice]

I think I've missed a bit of discussion, what's wrong with having this as the default?

[n-io]

I think I've missed a bit of discussion, what's wrong with having this as the default?

I completely agree, but having this as a default requires changing a vast number of tests. Also, I suspect that full-precision is not actually very good for cross-system reproducible file checks, as this is currently breaking in CI but passing on my system.

My preference would be to have a reduced-precision flag and use this on file checks, with full-precision mode being the default.

[n-io]

I've changed the default behaviour to print with full precision.

[n-io]

Here's a minimally intrusive suggestion, which takes the same approach of mlir-opt: attempt to print to scientific 6dp notation iff the resulting string is losslessly reproducible, or else print to full precision using repr. This has the advantage of not affecting a vast number of filechecks, mirroring upstream, and not having a flag for either full or reduced precision.

[superlopuh]

Thank you!

[alexarice]

Looks good.

I feel I would still personally prefer that "simple" floats print in non scientific format, but I guess this is just too large of a filecheck diff

[n-io]

Looks good.

I feel I would still personally prefer that "simple" floats print in non scientific format, but I guess this is just too large of a filecheck diff

It can always be done in the future. I'm not sure about the performance of parsing the str back to a float to ensure there is no precision loss, and printing it straight through repr seems perfectly fine in my opinion. Though it seemed a quick way to ensure not losing precision, which was causing issues to us.