Sorry, but I don't understand why this is useful, why there are so many {% if false %} and what's the current problem with Float#hash.

In addition to what @asterite said, if there's a LLVM intrinsic why aren't we binding and using it in instrinsics.cr.

@asterite There are two problems now:

1. Float32.hash doesn't normalize values before hash, so two representations of same floating value points to different key-value pairs. It's wrong (as seen by Python hash algorithm, for example).

2. Float64.hash has same problem and moreover, it is totally wrong due to Problem with Hash's hash codes truncation #3932. It will be solved later.

And I have a question - how to embed checks for existence of imported function in Library? llvm.canonicalize available from LLVM 5, llvm.minmax available from 3.N (I doesn't check which N).

@RX14 Yes, I have seen intrinsics, but doesn't understood criteria to place them into Intrinsics or LibM. Looks like they exists everywhere.

I think the float normalization intrinsic should go in LibM actually since they're float related. Could we just xor the lower and higher 32 bits of the float64 together?

@akzhan We don't currently have support for LLVM 5 and we won't until at least the release candidates. There's LLVM version constants defined here

Could we just xor the lower and higher 32 bits of the float64 together?

It does bad distribution.

I prefer this:

https://github.com/python/cpython/blob/master/Python/pyhash.c#L34

We don't currently have support for LLVM 5 and we won't until at least the release candidates. There's LLVM version constants defined here

It is not the problem, I will add IS_50 constant if needed.

But bin/crystal doesn't compile code with {% if LibLLVM::IS_50 ..., without recompilation of itself.

OK, will replace, recompile and try.

No. We don't support LLVM5 and compiling and linking against LLVM5 will just fail.

@ysbaddaden I want just to add

    IS_50 = {{LibLLVM::VERSION.starts_with?("5.0")}}

and test for it with

  {% if LibLLVM::IS_50 %}
  fun canonicalize_f32 = "llvm.canonicalize.f32"(value : Float32) : Float32
  fun canonicalize_f64 = "llvm.canonicalize.f64"(value : Float64) : Float64
  {% end %}

and other places.

To not rewrite code later.

@RX14 So #4560 frexp is introduced to implement Float64.hash later. And just because it is stdlib.

1. it should be IS_5 = x.starts_with?("5.") because there will only be major releases (and sometimes patch releases, numbered as minor) from now on;
2. don't bother with LLVM5 until we start supporting it (next september, at best), nobody will benefit from it until months later;
3. fix issues in a LLVM 3.8+ compatible way, and add a note that LLVM 5+ could help.

llvm.minnum was introduced in LLVM 3.6 for reference. I doubt there's any way to remove constant folding without making the method non-inlinable.

Thanks, @RX14. Looks like I want to get #4396 landed and forget about old LLVMs.

Thanks to @RX14 and @ysbaddaden for your help.

Now it's ready.

P.S.: Sorry for bad commits since I did not have a working VM during the day.

@asterite Looks like compiler cannot be built using LLVM <4.0, it raises on LLVM 3.8/Travis:

in src/llvm/abi.cr:2: class declaration can't have attributes
abstract class LLVM::ABI

I don't know, is it OK?

Looks like I must not use LibLLVM::IS_* anywhere in stdlib.

This pull request will be [WIP] until #4396 landed.

Anyway I can rework Float hashing with frexp only.