Port the implemention of SIMD intrinsics from Miri to const-eval #146568

sayantn · 2025-09-15T01:17:07Z

Ported the implementation of most SIMD intrinsics from Miri to rustc_const_eval. Remaining are

Math functions (as per @RalfJung's suggestions)
FMA (non-deterministic)
Funnel Shifts (not implemented in Miri yet)
Unordered reduction intrinsics (not implemented in Miri yet)

rustbot · 2025-09-15T01:17:09Z

Some changes occurred to the platform-builtins intrinsics. Make sure the
LLVM backend as well as portable-simd gets adapted for the changes.

cc @antoyo, @GuillaumeGomez, @bjorn3, @calebzulawski, @programmerjake

Some changes occurred to the CTFE / Miri interpreter

cc @rust-lang/miri

Some changes occurred to the CTFE machinery

cc @RalfJung, @oli-obk, @lcnr

Some changes occurred to the intrinsics. Make sure the CTFE / Miri interpreter
gets adapted for the changes, if necessary.

cc @rust-lang/miri, @RalfJung, @oli-obk, @lcnr

rustbot · 2025-09-15T01:17:11Z

r? @davidtwco

rustbot has assigned @davidtwco.
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

rustbot · 2025-09-15T02:03:02Z

workingjubilee is currently at their maximum review capacity.
They may take a while to respond.

programmerjake · 2025-09-15T02:42:21Z

tests/ui/consts/const-eval/simd/shuffle.rs

+}
+
+fn main() {
+    const X2: i32x2 = i32x2::from_array([20, 21]);


please add checks for length 3 and 6 since portable-simd wants to support non-powers-of-2. length 3 since it's odd and length 6 because it's the smallest where the simd type's alignment is not equal to either the simd type's size or to the element's size.

Should I also add these to the tests/ui/simd file, I just copied the test cases from there?

sure if you like. I mostly just figured while you're writing the code is a good time to add it to the code you're changing. I'll leave it up to the const-eval people to decide if they want it everywhere. shuffles do get a bunch of testing in portable-simd for miri (and normal compilation), but we don't explicitly test const-eval shuffles in portable-simd because they didn't exist yet.

Miri and const-eval will use the same implementation so in principle it is covered.

Instead of duplicating tests we usually prefer making the existing tests run both inside and outside consts. We have some infrastructure for that in the std test suite -- is there a reason this has to be a ui test?

Idts, this should be in coretests, but the tests for simd_insert and simd_extract was in this directory so I just put these there. I will port these to coretests

I have moved all the tests to coretests, and enabled const-eval tests for them

Actually, moving them to coretests is problematic because apparently there is no way to disable tests inside coretests (even for backends), so I undid the move, and modified the files in place to make them const-compatible

What do you mean by this? There's tons of ways to disable tests inside coretests, depending on what exactly you want to do.

I wanted to disable them based on the backend (clif and gcc has broken support, and I don't want this PR to be the one to fix them), couldn't figure it out, and then realized that the current design "works" 😅

We have moved a lot of tests from ui to coretests so this feels like a step in the wrong direction. I don't know how we deal with backend issues, maybe we have cfg flags for that?

We'll have to find another reviewer for the tests anyway, a 4k diffstat is just way more than I can handle currently.

compiler/rustc_const_eval/src/interpret/intrinsics.rs

RalfJung · 2025-09-15T07:04:44Z

Make simd_shuffle usable in const-contexts (I am also working on other simd intrinsics, but want to get this one over fast because a stdarch PR depends on it)

Which PR is that? Always good to cross-reference things. :) Or if the PR doesn't exist yet, what are your stdarch plans that need this?

sayantn · 2025-09-15T07:26:29Z

I want to make some stdarch functions available on const-contexts (for now, we are doing mostly constructors, but the _mm_set1 constructors use simd_shuffle, hence this. For some reason a simple Type([value; N]) doesn't optimize in darwin and ios (see #60637 (comment))

workingjubilee · 2025-09-15T18:02:56Z

I think we discussed creating a simd_splat intrinsic of some kind for simple initialization of vectors because it would optimize better.

workingjubilee · 2025-09-15T18:10:54Z

Also I know I mentioned it but please do not get into introducing such a new intrinsic here (a new PR, maybe!), as there are other good reasons to make this const so that we don't have to haggle over the many possible reasons (it is a truth universally acknowledged that std::arch is... a tad big) someone might want to use the simd_shuffle operation in their runtime API yet also const-evaluate it. At const time it's just boring operations on an array, after all, and you can already do that.

workingjubilee · 2025-09-15T18:13:23Z

Anyways, that aside, I'm basically good with this so it's mostly going to be haggling with Ralf, so

r? RalfJung

rustbot · 2025-09-15T23:28:32Z

The Miri subtree was changed

cc @rust-lang/miri

programmerjake · 2025-09-16T03:42:45Z

library/coretests/tests/simd_intrinsics.rs

+    pub const fn from_array(a: [T; N]) -> Self {
+        Simd(a)
+    }
+
+    pub const fn splat(value: T) -> Self
+    where
+        T: Copy,
+    {
+        Self([value; N])
+    }


you need to use mem::transmute instead of accessing the field inside a repr(simd) struct, since last I knew the compiler doesn't properly handle the field accesses in all cases

Suggested change

pub const fn from_array(a: [T; N]) -> Self {

Simd(a)

}

pub const fn splat(value: T) -> Self

where

T: Copy,

{

Self([value; N])

}

pub const fn from_array(a: [T; N]) -> Self {

// Safety: same shape

unsafe { mem::transmute(a) }

}

pub const fn splat(value: T) -> Self

where

T: Copy,

{

// Safety: same shape

unsafe { mem::transmute([value; N]) }

}

afaik constructing them is fine, projecting is the problematic part

library/coretests/tests/simd_intrinsics.rs

sayantn · 2025-09-25T00:00:02Z

seems spurious, let's retry

edit: huh 😕

bors · 2025-09-25T18:39:39Z

☔ The latest upstream changes (presumably #147019) made this pull request unmergeable. Please resolve the merge conflicts.

sayantn · 2025-09-25T19:20:38Z

@RalfJung as #146735 is merged now, should I also make simd_fma const in this pr, or should I postpone it for the next PR?

RalfJung · 2025-09-25T19:41:07Z

I'd prefer not expanding the scope of this PR.

rustbot · 2025-09-25T19:57:02Z

This PR was rebased onto a different master commit. Here's a range-diff highlighting what actually changed.

Rebasing is a normal part of keeping PRs up to date, so no action is needed—this note is just to help reviewers.

clarfonthey · 2025-10-05T00:20:46Z

This is probably far better performance-wise, but is there a particular reason why these are implemented directly in miri/const-eval instead of via fallback bodies? Since I had also been investigating that path to doing this, although it's clear that some versions like simd_relaxed_fma would require a direct MIRI implementation to ensure nondeterminism.

sayantn · 2025-10-05T00:27:20Z

The arguments are just random type parameters. If repr(simd) gets removed in favor of a Simd<T, N> type, then probably we can have a fallback implementation. This is in contrast to the scalar intrinsics, where we know all the possible types, and so use an extension trait or sth else to implement the fallback body

clarfonthey · 2025-10-05T00:31:42Z

I thought that repr(simd) was perma-unstable, and so, you would be able to know all possible types? I know that it's more than just Simd and also includes stuff like __m128, but I figured that you didn't have to account for some random repr(simd) struct being implemented downstream.

sayantn · 2025-10-05T00:39:42Z

Even if it is perma-unstable, there are a lot more of such structs lying around. Grepping the rustc repo I could find ~300 instances of #[repr(simd, a lot of them being in the tests

clarfonthey · 2025-10-05T02:03:11Z

Fair enough!

Sorry for interrupting, mostly just wanting to verify that this is the correct path for the implementation. While it doesn't seem unreasonable to just implement the fallback traits in those instances, you're right that it'd be a load of extra work when we could just rely on repr(simd) instead, especially when it'll probably be much more performant to directly compute things than do five loops through the traits system first.

sayantn · 2025-10-08T15:03:23Z

I (on suggestions from @RalfJung) have decided to split this PR into 2 - this one will just port the implementations, and another PR after this will update the definitions and the tests to actually make them const. So now this can be merged quickly to unblock other work such as #147355

r? @RalfJung

compiler/rustc_const_eval/src/interpret/intrinsics/simd.rs

src/tools/miri/src/intrinsics/simd.rs

compiler/rustc_const_eval/src/interpret/intrinsics/simd.rs

RalfJung · 2025-10-08T15:52:29Z

LGTM, thanks!
@bors r+

bors · 2025-10-08T15:52:33Z

📌 Commit 45ca537 has been approved by RalfJung

It is now in the queue for this repository.

rustbot added S-waiting-on-review Status: Awaiting review from the assignee but also interested parties. T-compiler Relevant to the compiler team, which will review and decide on the PR/issue. T-libs Relevant to the library team, which will review and decide on the PR/issue. labels Sep 15, 2025

rustbot assigned davidtwco Sep 15, 2025

rustbot assigned workingjubilee Sep 15, 2025

rustbot unassigned davidtwco Sep 15, 2025

programmerjake reviewed Sep 15, 2025

View reviewed changes

RalfJung reviewed Sep 15, 2025

View reviewed changes

compiler/rustc_const_eval/src/interpret/intrinsics.rs Outdated Show resolved Hide resolved

RalfJung reviewed Sep 15, 2025

View reviewed changes

compiler/rustc_const_eval/src/interpret/intrinsics.rs Show resolved Hide resolved

rustbot assigned RalfJung and unassigned workingjubilee Sep 15, 2025

sayantn force-pushed the simd-shuffle branch from 8c569d6 to ac970b6 Compare September 15, 2025 23:28

sayantn changed the title ~~Implement simd_shuffle in const_eval~~ Implement most SIMD intrinsics in const-eval Sep 15, 2025

programmerjake reviewed Sep 16, 2025

View reviewed changes

library/coretests/tests/simd_intrinsics.rs Outdated Show resolved Hide resolved

This comment has been minimized.

Sign in to view

sayantn force-pushed the simd-shuffle branch 2 times, most recently from 85efca6 to de2aa11 Compare September 16, 2025 05:54

This comment has been minimized.

Sign in to view

This comment was marked as outdated.

Sign in to view

sayantn force-pushed the simd-shuffle branch from de2aa11 to 1f632cd Compare September 18, 2025 01:33

sayantn force-pushed the simd-shuffle branch from 64804a0 to 0529a76 Compare September 24, 2025 23:27

This comment has been minimized.

Sign in to view

sayantn mentioned this pull request Sep 25, 2025

Make X86 intrinsics const rust-lang/stdarch#1925

Draft

sayantn force-pushed the simd-shuffle branch from 0529a76 to 1349bb8 Compare September 25, 2025 19:56

sayantn mentioned this pull request Oct 4, 2025

Add alignment parameter to simd_masked_{load,store} #147355

Open

3 tasks

sayantn force-pushed the simd-shuffle branch from 1349bb8 to 3d34ebc Compare October 8, 2025 14:58

sayantn changed the title ~~Implement most SIMD intrinsics in const-eval~~ Port the implemention of SIMD intrinsics from Miri to const-eval Oct 8, 2025

rustbot assigned RalfJung and unassigned tgross35 Oct 8, 2025

RalfJung reviewed Oct 8, 2025

View reviewed changes

compiler/rustc_const_eval/src/interpret/intrinsics/simd.rs Outdated Show resolved Hide resolved

RalfJung reviewed Oct 8, 2025

View reviewed changes

src/tools/miri/src/intrinsics/simd.rs Show resolved Hide resolved

RalfJung reviewed Oct 8, 2025

View reviewed changes

compiler/rustc_const_eval/src/interpret/intrinsics/simd.rs Outdated Show resolved Hide resolved

RalfJung reviewed Oct 8, 2025

View reviewed changes

compiler/rustc_const_eval/src/interpret/intrinsics/simd.rs Show resolved Hide resolved

Port the Miri implementations of SIMD intrinsics to rustc_const_eval

45ca537

sayantn force-pushed the simd-shuffle branch from 3d34ebc to 45ca537 Compare October 8, 2025 15:36

bors added S-waiting-on-bors Status: Waiting on bors to run and complete tests. Bors will change the label on completion. and removed S-waiting-on-review Status: Awaiting review from the assignee but also interested parties. labels Oct 8, 2025

Port the implemention of SIMD intrinsics from Miri to const-eval #146568

Are you sure you want to change the base?

Port the implemention of SIMD intrinsics from Miri to const-eval #146568

Conversation

sayantn commented Sep 15, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

rustbot commented Sep 15, 2025

Uh oh!

rustbot commented Sep 15, 2025

Uh oh!

rustbot commented Sep 15, 2025

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

RalfJung commented Sep 15, 2025

Uh oh!

sayantn commented Sep 15, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

workingjubilee commented Sep 15, 2025

Uh oh!

workingjubilee commented Sep 15, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

workingjubilee commented Sep 15, 2025

Uh oh!

rustbot commented Sep 15, 2025

Uh oh!

programmerjake Sep 16, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

This comment has been minimized.

This comment has been minimized.

This comment was marked as outdated.

This comment has been minimized.

sayantn commented Sep 25, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

bors commented Sep 25, 2025

Uh oh!

sayantn commented Sep 25, 2025

Uh oh!

RalfJung commented Sep 25, 2025

Uh oh!

rustbot commented Sep 25, 2025

Uh oh!

clarfonthey commented Oct 5, 2025

Uh oh!

sayantn commented Oct 5, 2025

Uh oh!

clarfonthey commented Oct 5, 2025

Uh oh!

sayantn commented Oct 5, 2025

Uh oh!

clarfonthey commented Oct 5, 2025

sayantn commented Sep 15, 2025 •

edited

Loading

sayantn commented Sep 15, 2025 •

edited

Loading

workingjubilee commented Sep 15, 2025 •

edited

Loading

programmerjake Sep 16, 2025 •

edited

Loading

sayantn commented Sep 25, 2025 •

edited

Loading