flang-new -> flang renaming article #62

DavidSpickett · 2024-11-14T10:36:25Z

Note: This article is to be published just after the release of LLVM 20, around Feburary 2025. Which will be the first release that includes LLVM Flang as flang instead of flang-new.

The original proposal being https://discourse.llvm.org/t/proposal-rename-flang-new-to-flang/69462.

I proposed writing this article (https://discourse.llvm.org/t/rfc-llvm-project-blog-post-for-flang-new-flang-renaming/80915) to mark that renaming.

It covers:

Why fortran still exists.
Why you would build a new compiler.
Flang's history.
How the community decided to rename it.
Flang's adoption of MLIR.
Quotes from users/contributors/etc.
How to contribute to Flang.

DavidSpickett · 2024-11-14T10:55:59Z

I am posting this very early relative to the likely February publishing date so we have plenty of time to gather feedback with all the holidays coming up.

It is ready for review, though the main focus at the moment is factual accuracy and general layout. I will concentrate on grammar and phrasing later, I have some reviewers in mind for that. I suggest you use the "rich" diff view so you can read the rendered Markdown document.

I am also collecting quotes for the article. Some of you I will be emailing directly, but anyone can be included and you can comment here instead. I cannot guarantee all input will be used but you of course have the right to ask that a quote be removed or corrected or put in the proper context if it is used.

(reading or reviewing the whole article is not required to give a quote, it's here for inspiration if you need it)

Quotes can be about anything related to the project:

Could be something that Flang allowed you to build
work that you enjoyed doing on Flang
something you think Flang has set a new standard for
how Flang contributes to your company/institution's strategy.

The goal is to show the breadth of the community and all the unique things people have got/continue to get from this project.

Please keep your quotes to a few sentences if possible.

If you have a particular way you wish to be cited, please let me know. My default is what I see on GitHub and the git log, plus the company name. If you want something different please tell me exactly what you prefer it to be, include:

Any special characters or formatting.
Translated forms of the name, e.g Japanese Kanji and a Romanised version, I'll include both.
Your job title if you want that included
Your company name, if you want it included

banach-space

Thanks, David. This reads really well, and it’s clear you’ve put in a lot of detective work to gather all this information - it’s greatly appreciated!

I’ve left a few minor comments and suggestions. After reading through, I see that you actually address most of these later in the text :)

I’ll aim to contribute something on the driver in the next few days.

content/posts/2024-11-05-flang-new.md

DavidSpickett · 2024-11-15T09:23:02Z

I’ve left a few minor comments and suggestions. After reading through, I see that you actually address most of these later in the text :)

These sort of comments are good to. Sometimes it's good for the reader to have questions for later, but if it's something they're going to be distracted by as they read, I should address that.

…vm flang aka f18.

DavidSpickett · 2024-11-15T09:48:51Z

FYI: big updates to the FIR and HLFIR sections coming, you can still point out my mistakes though :)

content/posts/2024-11-05-flang-new.md

luporl

I added a few nits, but the MLIR part looks good to me.

content/posts/2024-11-05-flang-new.md

DavidSpickett · 2024-11-19T10:38:10Z

I added a few nits, but the MLIR part looks good to me.

Thanks Leandro! I will be re-ordering the section soon, after I have more input from some Nvidia folks but I expect the factual content to be the same.

content/posts/2024-11-05-flang-new.md

… same.

luporl

Great explanation of OpenMP basics, illustrated with how Clang translates it to LLVM IR and contrasted with Flang, that first lowers it to MLIR.

luporl · 2024-11-21T14:51:13Z

content/posts/2024-11-05-flang-new.md

+The compiler interprets these directives and produces calls to a runtime library
+that distributes work as instructed.


At first glance, this gives me the impression that the compiler just needs to emit calls to a runtime library to distribute the work. But in most cases the block of code to be parallelized needs to be outlined to a separate function, copies of variables must be made, and other setup steps need to be performed.

Maybe something like the following would be more accurate:
The compiler interprets these directives and, by performing transformations in the marked code and producing calls to a runtime library, distributes work as instructed.

luporl · 2024-11-21T14:58:24Z

content/posts/2024-11-05-flang-new.md

+```
+([Compiler Explorer](https://godbolt.org/z/eca6vjbqe))
+
+Instead of `!`, comments start with `#` and instead of `DO` there is `for`.


C/C++ actually uses #pragma directives instead of comments.

luporl · 2024-11-21T15:05:17Z

content/posts/2024-11-05-flang-new.md

+Overall very similar to the Fortran example.
+
+LLVM IR does not know anything about OpenMP specifically, so Clang uses a
+library called `OpenMPIRBuilder` to convert OpenMP directives to LLVM IR.


Pedantic comment: the OpenMPIRBuilder class is actually part of the LLVMFrontendOpenMP library.

luporl · 2024-11-21T15:17:37Z

content/posts/2024-11-05-flang-new.md

+define dso_local void @simple(<...>) {
+<...>
+  call void <...> @__kmpc_fork_call(<...> ptr @simple(<...>) (.omp_outlined), <...>
+  ret void
+}
+
+define internal void @simple(<...>) (.omp_outlined) <...>) {
+<...>
+  call void @simple(<...>) (.omp_outlined_debug__) <...>
+  ret void
+}
+
+define internal void @simple(<...>) (.omp_outlined_debug__) <...> {
+<...>
+  call void @__kmpc_for_static_init_4(ptr @1, i32 %8, i32 34, <...>)
+<...>
+omp.inner.for.body:
+<...>
+omp.precond.end:
+  ret void
+}


It took me some time to realize that this is the output of clang and not flang, although it may be lack of attention on my side, as the previous paragraph talks about clang.

Also, why -O3 was used in the Fortran example but not in the C++ one?
With -O3, only 2 versions of simple are generated, which can simplify the example.

luporl · 2024-11-21T15:19:03Z

content/posts/2024-11-05-flang-new.md

+  ret void
+}
+
+define internal void @simple(<...>) (.omp_outlined) <...>) {


Parentheses are not balanced here.

luporl · 2024-11-21T17:18:29Z

content/posts/2024-11-05-flang-new.md

+You will see that instead of `for.do_loop` you have `omp.parallel`. `omp` is
+MLIR's [OpenMP dialect](https://mlir.llvm.org/docs/Dialects/OpenMPDialect/).
+This is a higher level and much more literal translation of the
+`parallel` directive than Clang produced.


Suggested change

`parallel` directive than Clang produced.

`parallel do` combined directive than Clang produced.

luporl · 2024-11-21T17:19:25Z

content/posts/2024-11-05-flang-new.md

+
+**Note:** Fortran arrays are [one-based](https://fortran-lang.org/en/learn/quickstart/arrays_strings/) by default. So the first element is at index 1. This example reads the previous element, so it starts `I` at 2.
+
+`!$OMP PARALLEL DO` is a directive in the form of a Fortran comment, which


Suggested change

`!$OMP PARALLEL DO` is a directive in the form of a Fortran comment, which

`!$OMP PARALLEL DO` is a combined directive in the form of a Fortran comment, which

luporl · 2024-11-21T17:30:30Z

content/posts/2024-11-05-flang-new.md

+}
+```
+
+There are some differences between this and the IR Clang produced but


Adding -O3 to Clang example reduces the differences.

luporl · 2024-11-21T17:31:26Z

content/posts/2024-11-05-flang-new.md

+
+## ClangIR and the Future
+
+It is surprising that a compiler for a language as old as Fotran got ahead of


Suggested change

It is surprising that a compiler for a language as old as Fotran got ahead of

It is surprising that a compiler for a language as old as Fortran got ahead of

luporl · 2024-11-21T17:49:29Z

content/posts/2024-11-05-flang-new.md

+above:
+
+```mlir
+define void @simple_(<...> {


nit: this is a matter of preference (so feel free to ignore it), but the omitted closing parentheses in this snippet makes it harder for me to parse it.

WIP flang release post for llvm 20

b314bb1

DavidSpickett force-pushed the flang-new-post branch from 54df17f to b314bb1 Compare November 14, 2024 10:56

DavidSpickett marked this pull request as ready for review November 14, 2024 11:15

Add some content on recent ecosystem innovations.

34f9201

banach-space reviewed Nov 14, 2024

View reviewed changes

DavidSpickett added 6 commits November 15, 2024 09:25

GFortran naming

f8396f9

Don't imply flang-new was in llvm 11.

5e06257

Be clearer that it's 10 years from the concept of a flang not from ll…

6f17580

…vm flang aka f18.

note anachronistic use of Classic Flang naming

299f6db

More specific flag to external fc wording.

31dae3a

Correct fir-dev upstreaming announcement wording

530c3a2

banach-space reviewed Nov 15, 2024

View reviewed changes

content/posts/2024-11-05-flang-new.md Outdated Show resolved Hide resolved

DavidSpickett added 5 commits November 15, 2024 16:28

Link to the review for code generation

18e3c60

more introduction of "LLVM Flang" term.

f38e86f

driver section

c53a68a

update flang-new details

73ba7e2

driver edits

a77b93e

luporl reviewed Nov 18, 2024

View reviewed changes

content/posts/2024-11-05-flang-new.md Outdated Show resolved Hide resolved

content/posts/2024-11-05-flang-new.md Outdated Show resolved Hide resolved

content/posts/2024-11-05-flang-new.md Outdated Show resolved Hide resolved

Address Leandro's comments on MLIR section

a868d26

More credits and some links I haven't used elsewhere.

38e43ee

banach-space reviewed Nov 19, 2024

View reviewed changes

DavidSpickett added 5 commits November 19, 2024 12:24

Add flang-tmp proposal to timeline.

f3dfc02

reword

de9a099

Add accent to Valentin's name

d17ae3d

Address Andrzej's comments on the driver section

447657c

adjust quote

a71b1df

DavidSpickett added 11 commits November 20, 2024 10:26

llvm weekly links

3ab2b98

pick a tense for the timeline and stick to it

09dc2c1

add AMD announcement

7b3896f

Add OpenMP events to timeline, combine entries where the month is the…

594507f

… same.

Add Arm Toolchain for Linux release to timeline

4e19efa

openmp section begins

dffdbb6

The end parallel is in fact optional.

349079c

note that mlir has a whole bunch of passes that I'm ignoring

ecf60b1

openmp continues

b321ff9

openmp edits

2a0da70

Note that OpenACC is essentially the same story as OpenMP

dadc10a

luporl reviewed Nov 21, 2024

View reviewed changes

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

flang-new -> flang renaming article #62

flang-new -> flang renaming article #62

DavidSpickett commented Nov 14, 2024 •

edited

Loading

DavidSpickett commented Nov 14, 2024 •

edited

Loading

banach-space left a comment

DavidSpickett commented Nov 15, 2024

DavidSpickett commented Nov 15, 2024

luporl left a comment

DavidSpickett commented Nov 19, 2024 •

edited

Loading

luporl left a comment

luporl Nov 21, 2024

luporl Nov 21, 2024

luporl Nov 21, 2024

luporl Nov 21, 2024

luporl Nov 21, 2024

luporl Nov 21, 2024

luporl Nov 21, 2024

luporl Nov 21, 2024

luporl Nov 21, 2024

luporl Nov 21, 2024

		The compiler interprets these directives and produces calls to a runtime library
		that distributes work as instructed.

	`parallel` directive than Clang produced.
	`parallel do` combined directive than Clang produced.


		Note: Fortran arrays are [one-based](https://fortran-lang.org/en/learn/quickstart/arrays_strings/) by default. So the first element is at index 1. This example reads the previous element, so it starts `I` at 2.

		`!$OMP PARALLEL DO` is a directive in the form of a Fortran comment, which

	`!$OMP PARALLEL DO` is a directive in the form of a Fortran comment, which
	`!$OMP PARALLEL DO` is a combined directive in the form of a Fortran comment, which


		## ClangIR and the Future

		It is surprising that a compiler for a language as old as Fotran got ahead of

flang-new -> flang renaming article #62

Are you sure you want to change the base?

flang-new -> flang renaming article #62

Conversation

DavidSpickett commented Nov 14, 2024 • edited Loading

DavidSpickett commented Nov 14, 2024 • edited Loading

banach-space left a comment

Choose a reason for hiding this comment

DavidSpickett commented Nov 15, 2024

DavidSpickett commented Nov 15, 2024

luporl left a comment

Choose a reason for hiding this comment

DavidSpickett commented Nov 19, 2024 • edited Loading

luporl left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

DavidSpickett commented Nov 14, 2024 •

edited

Loading

DavidSpickett commented Nov 14, 2024 •

edited

Loading

DavidSpickett commented Nov 19, 2024 •

edited

Loading