Re-organize `intrinsic-test` to enable seamless addition of behaviour testing for more architectures #1758

madhav-madhusoodanan · 2025-03-24T19:00:45Z

Updates so far

Moved the ARM-specific testing to the arm module, exposing only a test function at src/main.rs
Created a SupportedArchitectureTest trait that has been implemented for arm and should be implemented for other architectures too. Allows us to expose functionality like building C/Rust test files and comparing outputs.
Moved common functionality (CLI parsing, create files, compilations, compare outputs) into the common module for reuse by other architectures too.
Implemented a match block for selection of architectures using the target CLI variable.

Reasoning

The Intrinsic type may be specific to architectures, hence the implementation of a trait helps us to abstract the usage of such data types within the architecture-specific module

New architecture of `intrinsic-test`

rustbot · 2025-03-24T19:00:50Z

rustbot has assigned @Amanieu.
They will have a look at your PR within the next two weeks and either review your PR or reassign to another reviewer.

Use r? to explicitly pick a reviewer

Amanieu · 2025-03-25T00:00:26Z

The module should be called arm rather than aarch.

Amanieu · 2025-03-25T00:01:01Z

@JamieCunliffe feel free to have a look at this since you are the original author of intrinsic-test. This is part of a GSoC project to extend intrinsic-test to other architectures.

Reasoning: 1. Majority of code assumes the usage of `Intrinsic` and related types, which is derived from the JSON structure of the ARM intrinsics JSON source file 2. Further commits will start with extracting common parts of the code (eg: Create C/Rust file, Build C/Rust file, etc)

…ted for different architectures. Next steps: Move the existing ARM-specific implementation into one that fits well with this trait.

…truct args

madhav-madhusoodanan · 2025-03-26T18:30:39Z

Are we using the --a32 flag in the CLI of intrinsic-test anywhere? I was unable to find its usage.
Would it be alright if I remove it, or is there some reasoning that I may be missing?

Amanieu · 2025-03-27T16:00:30Z

--a32 is not currently tested in CI, but it is used for 32-bit ARM targets (arm* and thumb*). Some ARM intrinsics are only available on 64-bit ARM targets and this flag skips those.

…le_rust and compare_outputs

adamgemmell

Looks good! I think types.rs would map fine to x86 with a little modification.

I'd just like to add that Scalable Vectors will be added to this tool at some point - the Arm code is here and it would probably be useful for RISC-V in the future too.

Don't worry about conflicts with that PR (it's very old and needs a rework) and I don't think there's anything particularly different about how intrinsic-test needs to work but scalable vectors do add an extra dimension of complexity to the tool.

adamgemmell · 2025-04-15T15:42:28Z

crates/intrinsic-test/src/common/compare.rs

+                .arg(format!("{runner} ./c_programs/{intrinsic_name}"))
+                .output();
+
+            let rust = if SPECIAL_TARGETS


I think this can be collapsed into 1 if, it's what I did with https://github.com/rust-lang/stdarch/pull/1771/files#diff-7c34e890416eb66d4462c7f31278409c662bb808a57c3e0a4d13c232c64533a9

Got it, let me do that.

adamgemmell · 2025-04-16T14:26:55Z

crates/intrinsic-test/src/common/types.rs

+    name = "Intrinsic test tool",
+    about = "Generates Rust and C programs for intrinsics and compares the output"
+)]
+pub struct Cli {


Seems an odd file name for the CLI

I'll need to rename the files, I started off by focussing on the code refactor side of things.
Thank you for sharing this observation with me.

madhav-madhusoodanan · 2025-04-16T16:07:11Z

Over the last 4 commits I changed from the Argument type to an Argument<T, M> generic type.

This allowed me to refactor a lot more code into the common module, since Argument has a lot of code that could be reused across architectures.

Where it differed was in the usage of the IntrinsicType enum (now the IntrinsicType struct after removing the recursion-based Ptr and Type variants) and the ArgPrep struct.

At the moment, each architecture can define:

Their own Metadata type (which Constraint used to be for arm target intrinsics), represented by the generic type M.
Their own version of <arch>IntrinsicType struct (ArmIntrinsicType, etc), which wraps over the core IntrinsicType struct and can handle architecture-specific needs (represented by the generic type T).
- The common parts of the definition is now handled by a macro called base_intrinsictype_trait_def_macro.
- Reason I didn't make this into an enum is I think it might be easier to extend architecture-specific versions of IntrinsicType by adding a new module without having to update the common module, please let me know if this is a good idea.

I've made some changes in the C testfile-generation part, will pinpoint them in the next message.

After this part, I intend to work on:

Wrapping the functionality that handles the compilation commands into a struct
Moving the architecture-specific <arch>ArchitectureTest structs into an enum

madhav-madhusoodanan · 2025-04-19T18:31:08Z

Updating what I did over the last 4 commits:

Converted the way the C compilation commands were being constructed from raw format! calls into a structured manner.
Updated the execution command for Rust test files, removing the aarch64_be-specific commands (as suggested by @adamgemmell in this comment).
Implemented dynamic dispatch for the construction of <arch>ArchitectureTest structs.
- This was done since the structs are intended to hold similar functionality (viz. building C/Rust files and comparing outputs)
- The arch-specific modules accommodates any arch-specific differences
- Dynamic dispatch seems like a great way to call these functions without needing to implement match blocks.
Cleaned up any unnecessary function arguments

What I'll do next:

It seems like arm/function.rs has some code that could be made common to other arch too. Will look into it.
Files need to be renamed for clarity.

madhav-madhusoodanan · 2025-04-23T14:31:40Z

crates/intrinsic-test/src/arm/functions.rs

+fn gen_code_c(
+    indentation: Indentation,
+    intrinsic: &Intrinsic<ArmIntrinsicType, Constraint>,
+    constraints: &[&Argument<ArmIntrinsicType, Constraint>],
+    name: String,
+    target: &str,
+) -> String {


I think it might be best for the gen_code_c and gen_code_rust functions to stay module-specific. Would this be a good idea @Amanieu @adamgemmell ?

Constraints seem like a very ARM thing.

I think overall you should try to keep as much generic as possible until you find an example of an intrinsic that wouldn't work otherwise - no point trying to guess what the thousands of intrinsics will want at this point.

If an architecture doesn't have any Constraints you could pass it an empty slice and that would be fine. Constraints are the "top-level" of generated code per intrinsic so it would be a bit awkward for the generator to go to platform-specific code and then jump back to generic code that might need to be aware of what constraints were set up and how they were named.

Constraints really are for anything const (i.e. needs to end up as an immediate at the assembly level), including enums. Scanning x86_64 intrinsics I found a few examples that would require a constraint (if you were to keep this model around):

https://doc.rust-lang.org/core/arch/x86_64/fn._mm512_getmant_round_ps.html takes some enums - the test tool doesn't currently need to handle any enums but I chose a new Constraint variant to represent each enum in SVE. This might not be best overall given there's loads of enums x86.

https://doc.rust-lang.org/core/arch/x86_64/fn._mm512_fpclass_pd_mask.html takes a const IMM8 from 0 to 0xFF, not using the full i32 range of the function parameter.

I think what you end up doing depends on how your intrinsic reference is written too.

madhav-madhusoodanan · 2025-04-23T14:33:08Z

crates/intrinsic-test/src/arm/functions.rs

+    let mut command = CompilationCommandBuilder::new()
+        .add_arch_flags(vec!["armv8.6-a", "crypto", "crc", "dotprod", "fp16"])
+        .set_compiler(compiler)
+        .set_target(target)
+        .set_opt_level("2")
+        .set_cxx_toolchain_dir(cxx_toolchain_dir)
+        .set_project_root("c_programs")
+        .add_extra_flags(vec!["-ffp-contract=off", "-Wno-narrowing"]);
+
+    if !target.contains("v7") {
+        command = command.add_arch_flags(vec!["fanimalux", "lut", "sha3"]);
+    }
+
+    command = if target == "aarch64_be-unknown-linux-gnu" {
+        command
+            .set_linker(
+                cxx_toolchain_dir.unwrap_or("").to_string() + "/bin/aarch64_be-none-linux-gnu-g++",
+            )


I also wanted to ask if the if blocks are being too unclean here. I adopted a factory pattern to build out these commands.
Does this look appropriate?

The factory pattern looks good! If statements for these targets are unavoidable, don't worry about them.

Why do you have a set_target call and then manually below set the target again if it's clang?

Ohh thanks for pointing that out.

Initially I thought -target and —target had different meanings. I completely forgot to remove one of them upon clarification

…st` structs

rustbot assigned Amanieu Mar 24, 2025

madhav-madhusoodanan force-pushed the restructure-intrinsic-test-crate branch 2 times, most recently from 9900e81 to 546554d Compare March 24, 2025 19:10

madhav-madhusoodanan force-pushed the restructure-intrinsic-test-crate branch from 546554d to 3ff317f Compare March 25, 2025 09:25

madhav-madhusoodanan force-pushed the restructure-intrinsic-test-crate branch from 3ff317f to bd0a675 Compare March 25, 2025 09:32

Chore: Added SupportedArchitectureTest trait which must be implemen…

7ff8497

…ted for different architectures. Next steps: Move the existing ARM-specific implementation into one that fits well with this trait.

madhav-madhusoodanan force-pushed the restructure-intrinsic-test-crate branch 4 times, most recently from 5dce230 to 7ff8497 Compare March 26, 2025 15:32

chore: Added ProcessedCli to extract the logic to pre-process CLI s…

399f37b

…truct args

madhav-madhusoodanan force-pushed the restructure-intrinsic-test-crate branch 2 times, most recently from a12b4a4 to ff10311 Compare March 27, 2025 15:49

madhav-madhusoodanan force-pushed the restructure-intrinsic-test-crate branch 4 times, most recently from 3068c9d to 2d1e482 Compare March 27, 2025 17:52

chore: separated common logic within file creations, compile_c, compi…

3e46d80

…le_rust and compare_outputs

madhav-madhusoodanan force-pushed the restructure-intrinsic-test-crate branch 3 times, most recently from e0b0301 to 54a678c Compare March 27, 2025 18:34

chore: code consolidation

2777ceb

madhav-madhusoodanan force-pushed the restructure-intrinsic-test-crate branch from 54a678c to 2777ceb Compare March 27, 2025 18:36

chore: added match block in src/main.rs

a2ce02c

madhav-madhusoodanan force-pushed the restructure-intrinsic-test-crate branch 2 times, most recently from 448e336 to f3f4f7c Compare April 16, 2025 10:30

Added a macro to simplify <Arch>IntrinsicType definitions

358016a

madhav-madhusoodanan force-pushed the restructure-intrinsic-test-crate branch from f3f4f7c to 358016a Compare April 16, 2025 10:32

renamed a64_only data member in Intrinsic to arch_tags

cc615b6

adamgemmell reviewed Apr 16, 2025

View reviewed changes

Removed aarch64-be specific execution command for rust test files

d395e9c

madhav-madhusoodanan force-pushed the restructure-intrinsic-test-crate branch 7 times, most recently from 16c89b0 to 9eb138d Compare April 19, 2025 18:11

madhav-madhusoodanan commented Apr 23, 2025

View reviewed changes

madhav-madhusoodanan added 2 commits April 23, 2025 20:34

moved the C compilation commands into a struct for easier handling

08325f3

Added dynamic dispatch for easier management of `<arch>ArchitectureTe…

420f2ee

…st` structs

madhav-madhusoodanan force-pushed the restructure-intrinsic-test-crate branch from 17135e7 to f2923fa Compare April 23, 2025 15:06

madhav-madhusoodanan requested a review from Amanieu April 24, 2025 05:01

madhav-madhusoodanan force-pushed the restructure-intrinsic-test-crate branch 5 times, most recently from 0eeae76 to 39091eb Compare April 24, 2025 14:15

madhav-madhusoodanan added 2 commits April 24, 2025 19:47

code cleanup

b86b969

chore: file renaming

f4650c0

madhav-madhusoodanan force-pushed the restructure-intrinsic-test-crate branch from 39091eb to f4650c0 Compare April 24, 2025 14:17

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Re-organize `intrinsic-test` to enable seamless addition of behaviour testing for more architectures #1758

Re-organize `intrinsic-test` to enable seamless addition of behaviour testing for more architectures #1758

madhav-madhusoodanan commented Mar 24, 2025 •

edited

Loading

rustbot commented Mar 24, 2025

Amanieu commented Mar 25, 2025

Amanieu commented Mar 25, 2025

madhav-madhusoodanan commented Mar 26, 2025 •

edited

Loading

Amanieu commented Mar 27, 2025

adamgemmell left a comment

adamgemmell Apr 15, 2025

madhav-madhusoodanan Apr 16, 2025

adamgemmell Apr 16, 2025

madhav-madhusoodanan Apr 16, 2025

madhav-madhusoodanan commented Apr 16, 2025 •

edited

Loading

madhav-madhusoodanan commented Apr 19, 2025

madhav-madhusoodanan Apr 23, 2025 •

edited

Loading

adamgemmell Apr 24, 2025 •

edited

Loading

madhav-madhusoodanan Apr 24, 2025

madhav-madhusoodanan Apr 23, 2025 •

edited

Loading

adamgemmell Apr 24, 2025 •

edited

Loading

madhav-madhusoodanan Apr 24, 2025 •

edited

Loading

Re-organize intrinsic-test to enable seamless addition of behaviour testing for more architectures #1758

Are you sure you want to change the base?

Re-organize intrinsic-test to enable seamless addition of behaviour testing for more architectures #1758

Conversation

madhav-madhusoodanan commented Mar 24, 2025 • edited Loading

Updates so far

Reasoning

New architecture of intrinsic-test

rustbot commented Mar 24, 2025

Amanieu commented Mar 25, 2025

Amanieu commented Mar 25, 2025

madhav-madhusoodanan commented Mar 26, 2025 • edited Loading

Amanieu commented Mar 27, 2025

adamgemmell left a comment

Choose a reason for hiding this comment

adamgemmell Apr 15, 2025

Choose a reason for hiding this comment

madhav-madhusoodanan Apr 16, 2025

Choose a reason for hiding this comment

adamgemmell Apr 16, 2025

Choose a reason for hiding this comment

madhav-madhusoodanan Apr 16, 2025

Choose a reason for hiding this comment

madhav-madhusoodanan commented Apr 16, 2025 • edited Loading

madhav-madhusoodanan commented Apr 19, 2025

madhav-madhusoodanan Apr 23, 2025 • edited Loading

Choose a reason for hiding this comment

adamgemmell Apr 24, 2025 • edited Loading

Choose a reason for hiding this comment

madhav-madhusoodanan Apr 24, 2025

Choose a reason for hiding this comment

madhav-madhusoodanan Apr 23, 2025 • edited Loading

Choose a reason for hiding this comment

adamgemmell Apr 24, 2025 • edited Loading

Choose a reason for hiding this comment

madhav-madhusoodanan Apr 24, 2025 • edited Loading

Choose a reason for hiding this comment

Re-organize `intrinsic-test` to enable seamless addition of behaviour testing for more architectures #1758

Re-organize `intrinsic-test` to enable seamless addition of behaviour testing for more architectures #1758

madhav-madhusoodanan commented Mar 24, 2025 •

edited

Loading

New architecture of `intrinsic-test`

madhav-madhusoodanan commented Mar 26, 2025 •

edited

Loading

madhav-madhusoodanan commented Apr 16, 2025 •

edited

Loading

madhav-madhusoodanan Apr 23, 2025 •

edited

Loading

adamgemmell Apr 24, 2025 •

edited

Loading

madhav-madhusoodanan Apr 23, 2025 •

edited

Loading

adamgemmell Apr 24, 2025 •

edited

Loading

madhav-madhusoodanan Apr 24, 2025 •

edited

Loading