[CIR] introduce CIR floating-point types

[Lancern]

This PR adds a dedicated cir.float type for representing floating-point types. There are several issues linked to this PR: #5, #78, and #90.

The new cir.float type is parameterized by an enum FloatFormat which defines the format of the floating point values. Currently 7 formats are defined:

• FF_Binary16, which represents IEEE-754 binary16 format;
• FF_Binary32, which represents IEEE-754 binary32 format, and corresponds to the float type in C/C++;
• FF_Binary64, which represents IEEE-754 binary64 format, and corresponds to to the double and long double type in C/C++;
• FF_Binary128, which represents IEEE-754 binary128 format, and corresponds to the long double type in C/C++.
• FF_X86Fp80, which represents x86 80-bit floating point format, and also corresponds to the long double type in C/C++;
• FF_BFloat16, which represents brain float format;
• FF_PpcFp128, which represents PowerPC 128-bit floating point format. It is also known as the double double format. It also corresponds to the long double type in C/C++.

What floating-point format the long double type uses depends both on the implementation and on the target platform.

TODOs

• Add type cir.float and attribute #cir.float, and replace all existing usages of builtins with the new type and attribute;
• Update the tests
  • [CIR] fix constant type verification #386

Updated on Jan 8, 2024

• Add FF_Binary16 and FF_BFloat16 formats.

Updated on Jan 15, 2024

• Rename FF_Binary64E to FF_X86Fp80 to avoid confusion.
• Add ppc_fp128 format.

[philnik777]

Is there a reason to not also add support for float16 and bfloat16?

[Lancern]

Is there a reason to not also add support for float16 and bfloat16?

@philnik777 Thanks for your comment.

I thought there aren't any standard types corresponding to float16 and bfloat16. But seems that C++23 now supports fixed width floating-point types and there are now std::float16_t and std::bfloat16_t. I'll add support to these two formats in later commits.

[bcardosolopes]

Hi, this is awesome, happy to see progress in this direction, it's been a while since we want to have CIR specific float support. Some initial feedback in the comments.

Please also add new testcases to test aliases and cir to cir.

[bcardosolopes]

@sitio-couto because you did cir.int it would be great to get your review here too!

[bcardosolopes]

@Lancern when you update the PR, please also make sure tests pass, so reviewers can play around with it locally. I know this is a Draft, but no problem making it a first class PR already - incremental work is fine and encouraged!

[sitio-couto]

@Lancern thanks for tackling this! It's long overdue.

I just added a suggestion and a discussion with @bcardosolopes.

[github-actions]

✅ With the latest revision this PR passed the C/C++ code formatter.

[Lancern]

I met a problem trying to lower cir.const that yields a floating-point value in FF_PpcFp128 format to LLVMIR. The operation is lowered to a llvm.mlir.constant operation, which takes an attribute that may be an integer attribute, floating point attribute, dense or sparse attribute containing integers or floats. However, the builtin mlir::FloatAttr does not support PPC fp128 format floating-point values. How should we deal with this?

PPC fp128 type is lowered to a special !llvm.ppc_fp128 type provided by the LLVM dialect.

[Lancern]

However, the builtin mlir::FloatAttr does not support PPC fp128 format floating-point values.

I have asked in the MLIR channel on Discord. Seems that support for !llvm.ppc_fp128 constant is missing in the LLVMIR and we have to wait for upstream to resolve it before the PPC FP128 format can be lowered to LLVMIR from CIR.

[bcardosolopes]

Thanks for the changes @Lancern, more feedback for you inline. Couple of big picture comments:

• Please make this first PR as simple as possible, focus on float and double and we could go over the rest with incremental work.
• Apply @sitio-couto changes and create individual types for each floating point type.
• Remove changes (mostly formatting) that are not related to the functionality you are changing - the amount of noise makes reviewing pretty hard.

[Lancern]

Thanks for all your comments and early reviews here! According to the early reviews, I'm going to further update this PR with the following TODOs list:

• Remove all whitespace-only changes from this PR.
• Focus on float and double types only, and remove other formats from this PR. These additional formats will be implemented in future PRs.
• Instead of using a parameterized !cir.float type to represent all floating-point types, invent dedicated type for each floating format: !cir.float, !cir.double, etc.

[Lancern]

Rebased onto main and removed any whitespace-only changes.

[bcardosolopes]

@Lancern thanks for the fast fixes and the work you've been putting up into the PRs. I'm currently working on an RFC to upstream clangir, so I'm a bit behind on reviews, I'll take a look into this one and others as soon as possible.

[Lancern]

@bcardosolopes Just take your time! IMO your review latency is already good enough and it's understandable to slow down a bit especially if you have more important thing to do first.

[bcardosolopes]

Sound like many of the changes requested haven't applied yet, I still see lots of things around that are out of the scope of this patch. Let me know once you go over all of them.

[Lancern]

Thanks for everyone's early comments and reviews! Now I should have resolved all confirmed problems during the early review and there're no more TODOs on my list. So I believe this PR should be ready.

CC @bcardosolopes

[bcardosolopes]

@lanza just rebased and this has the side effect of breaking GH PR workflow, making it impossible to review, apologies. Please rebase!

[Lancern]

Please rebase!

@bcardosolopes Rebased.

[bcardosolopes]

Thanks for your patience, getting closer!

[bcardosolopes]

Let me know when you're done applying all reviews (without conflicts) and ready for another round of review!

[Lancern]

@bcardosolopes Hey I should have resolved all reviews yet. The primary change is replacing the mlir::cir::FloatType type with an interface that provides the same methods. Have a look when you have time :)

[Lancern]

Rebased onto the latest main. Updated some newly merged to start using the new !cir.float and !cir.double type.

[bcardosolopes]

Thanks for the hard work, only a few nit remaining!

[bcardosolopes]

Awesome, thanks for putting solid work here and for your patience. LGTM