Improved printing for floats #563
Conversation
Testing suggests all f32s are now printed accurately.
|
Excellent work, @scurest. I'm going to wait for the tests to pass, and then merge. |
|
Ah, I forgot to ask: I couldn't get the zig files for the tests to compile in release mode. How do you do it? |
|
Are you asking how to run a test in release mode?
|
|
Well, I meant compile say If I use |
Hmm. This is really surprising, I'm looking into this right now. At first glance it seems like LLVM optimization is deleting functions that are marked with |
the optimizer was deleting compiler_rt symbols, so I changed the linkage type from LinkOnce to Weak also changed LinkOnce to mean linkonce_odr in llvm and Weak to mean weak_odr in llvm. See #563
|
I fixed the compiler-rt problem in master branch. The other problems you ran into I see the same problem, and it's a bug in zig. I'm looking into it. I'll merge this PR though, because this is better than status quo. |
All floats I've tried now print correctly.
See #375 (comment) for the first commit. The further changes are
For the last bullet, maybe we want to wait for an upstream confirmation and patch.
Tests
The tests I made are here (see the
usagefile at the bottom). All the tests I've run have passed. In particular the f32 tests are intended to show by exhaustion that we're correct for all f32s.f32 I iterate over all positive u32s, bitCasting them to f32s, skipping the NaNs and Infs, and printing the rest to stdout in zig. This is piped into a C program where I scanf with
%fand check if the nth one I get bitcasts back to n. This should exhaustively test all positive f32s.If you want to check all f32s, just change the 0x80000000s to 0xffffffffs in both files. I didn't just because it takes about 2 hours to run already.
rand This generates some random numbers with xoroshiro128+, bitcasts to f64, prints them, pipes them into C, scanfs them with
%lfand checks if they're the same. The more times you run it, the greater your confidence should be that we're correct :)