Improve SIMD type element count validation #80652

calebzulawski · 2021-01-03T15:42:02Z

These changes are motivated by stdsimd moving in the direction of const generic vectors, e.g.:

#[repr(simd)]
struct SimdF32<const N: usize>([f32; N]);

This makes a few changes:

Establishes a maximum SIMD lane count of 2^16 (65536). This value is arbitrary, but attempts to validate lane count before hitting potential errors in the backend. It's not clear what LLVM's maximum lane count is, but cranelift's appears to be much less than usize::MAX, at least.
Expands some SIMD intrinsics to support arbitrary lane counts. This resolves the ICE in the linked issue.
Attempts to catch invalid-sized vectors during typeck when possible.

Unresolved questions:

Generic-length vectors can't be validated in typeck and are only validated after monomorphization while computing layout. This "works", but the errors simply bail out with no context beyond the name of the type. Should these errors instead return LayoutError or otherwise provide context in some way? As it stands, users of stdsimd could trivially produce monomorphization errors by making zero-length vectors.

cc @bjorn3

…limit SIMD vector lengths.

rust-highfive · 2021-01-03T15:42:06Z

r? @estebank

(rust-highfive has picked a reviewer for you, use r? to override)

workingjubilee · 2021-01-06T22:04:35Z

65536, or a 16-bit value, was chosen "arbitrarily". Allow me to provide a justification, even if it is a mere post hoc rationale:

Rust supports byte (as in octet) addressing, and it is reasonable to say something like "a SIMD vector supported in a programming language should be fully addressable at the same level as the language". In addition, most SIMD vectors let you use byte-size lanes in hardware but few will go down to bits or even nybbles, so that is another angle which suggests we should focus on byte elements as our minimum. With a 16-bit number we could index every byte in a hypothetical 524,288-bit vector. This is 512*(2^10) bits, or 10 doublings past our current capability. Even the most gratuitous interpretations of Moore's law would suggest we would not be able to reach this in commodity hardware for another 20 years, and even supercomputers would take a while. For various reasons, in practice, vectors grow even more slowly than that, doubling in commodity hardware only every ~9 years or so (SSE for 128-bit vectors in 1999, AVX-512F in 2017). Even Arm SVE2 only allows a maximum vector length of 2048 bits, for which an 8-bit number is sufficient to index every byte.

While a possible implementation of a hypothetical "extensible vector" ISA, like a further extension of SVE2 to "SVE2.1" or "SVE3", or even the current alpha of the RISCV-V design, on a... very beefy chip, even by supercomputer standards... may allow a hypothetical 2^19 bit monster-vector to live and breathe even before the next 20 years pass, I am nonetheless content with saying that this 65536 lane limitation does not impose an unnecessary limitation on the capabilities of the Rust compiler for any futures I expect Rust to remain relevant in. I would even support an 8-bit limitation for now with an eye towards eventually lifting it, as it would complicate handling 4096-bit vectors and thus could realistically be considered a limitation in meeting the capabilities of silicon that might be printed even in the lifetimes of mere mortals, as that exceeds the capacity of even the most ambitious ISAs, but by a realistic amount. Note that SVE2 itself is not publicly available on any cores I am aware of, so as far as I know, and things like masking operations will already run into amusing consequences on 2048-bit lanes as-is, so this itself is definitely not a limitation today, either.

Finally, as noted in that issue, Cranelift supports 16-bit lane lengths.

workingjubilee · 2021-01-06T22:25:04Z

Regarding errors, I believe we should try to catch zero-lane vectors specially if we can. Philosophically, the minimum number of lanes for a "SIMD" operation to make sense is not 0 or even 1, but 2.

KodrAus · 2021-01-18T10:25:18Z

r? @nagisa

nagisa

Overall looks pretty good to me I have a couple comments in-line but nothing major.

@bors try

compiler/rustc_typeck/src/check/check.rs

compiler/rustc_middle/src/ty/layout.rs

compiler/rustc_codegen_llvm/src/intrinsic.rs

nagisa · 2021-01-22T12:51:36Z

@bors try

bors · 2021-01-22T12:51:48Z

⌛ Trying commit 07db2bf with merge 84b21bfb06d318b7a01ab0f0a70d0190ffbca05e...

bors · 2021-01-22T13:36:16Z

☀️ Try build successful - checks-actions
Build commit: 84b21bfb06d318b7a01ab0f0a70d0190ffbca05e (84b21bfb06d318b7a01ab0f0a70d0190ffbca05e)

…SIMD vectors with non-power-of-two lengths.

calebzulawski · 2021-01-23T21:38:28Z

In the last stdsimd meeting we decided it would be best to limit #[repr(simd)] to only powers-of-two, since that's all cranelift supports. If/when support for non-powers-of-two is added, we can remove this restriction.

Since repr_simd is still an unstable feature I think we should be fine to change that, but I do think a crater run might be a good idea to make sure.

rust-log-analyzer · 2021-01-23T21:47:18Z

The job mingw-check failed! Check out the build log: (web) (plain)

Click to see the possible cause of the failure (guessed by this bot)

configure: rust.channel         := nightly
configure: rust.debug-assertions := True
configure: llvm.assertions      := True
configure: dist.missing-tools   := True
configure: build.configure-args := ['--enable-sccache', '--disable-manage-submodu ...
configure: writing `config.toml` in current directory
configure: 
configure: run `python /checkout/x.py --help`
configure: 
---
   Compiling regex v1.4.3
   Compiling tidy v0.1.0 (/checkout/src/tools/tidy)
    Finished release [optimized] target(s) in 10.24s
tidy check
tidy error: following path contains more than 1458 entries, you should move the test to some relevant subdirectory (current: 1464): /checkout/src/test/ui
tidy error: following path contains more than 2669 entries, you should move the test to some relevant subdirectory (current: 2671): /checkout/src/test/ui/issues
Found 435 error codes
Found 0 error codes with no tests
Done!
some tidy checks failed

rust-log-analyzer · 2021-01-23T22:32:23Z

The job mingw-check failed! Check out the build log: (web) (plain)

Click to see the possible cause of the failure (guessed by this bot)

configure: rust.channel         := nightly
configure: rust.debug-assertions := True
configure: llvm.assertions      := True
configure: dist.missing-tools   := True
configure: build.configure-args := ['--enable-sccache', '--disable-manage-submodu ...
configure: writing `config.toml` in current directory
configure: 
configure: run `python /checkout/x.py --help`
configure: 
---
   Compiling regex v1.4.3
   Compiling tidy v0.1.0 (/checkout/src/tools/tidy)
    Finished release [optimized] target(s) in 9.93s
tidy check
tidy error: following path contains more than 1458 entries, you should move the test to some relevant subdirectory (current: 1464): /checkout/src/test/ui
tidy error: following path contains more than 2669 entries, you should move the test to some relevant subdirectory (current: 2671): /checkout/src/test/ui/issues
Found 435 error codes
Found 0 error codes with no tests
Done!
some tidy checks failed

rust-log-analyzer · 2021-01-23T23:31:50Z

The job x86_64-gnu-llvm-9 failed! Check out the build log: (web) (plain)

Click to see the possible cause of the failure (guessed by this bot)

Suite("src/test/assembly") not skipped for "bootstrap::test::Assembly" -- not in ["src/tools/tidy"]
Check compiletest suite=assembly mode=assembly (x86_64-unknown-linux-gnu -> x86_64-unknown-linux-gnu)

running 29 tests
iiiiiiiiiiiiiiiiiiiiiiiiiiiii

Suite("src/test/incremental") not skipped for "bootstrap::test::Incremental" -- not in ["src/tools/tidy"]
 finished in 0.069 seconds
Check compiletest suite=incremental mode=incremental (x86_64-unknown-linux-gnu -> x86_64-unknown-linux-gnu)
---
Suite("src/test/debuginfo") not skipped for "bootstrap::test::Debuginfo" -- not in ["src/tools/tidy"]
Check compiletest suite=debuginfo mode=debuginfo (x86_64-unknown-linux-gnu -> x86_64-unknown-linux-gnu)

running 116 tests
iiiiiiiiii.i.i..i...i..ii...i.i....ii..........iiii.........i.....i...i.......ii.i.ii.....iiii.....i 100/116
test result: ok. 78 passed; 0 failed; 38 ignored; 0 measured; 0 filtered out; finished in 2.34s

 finished in 2.405 seconds
Suite("src/test/ui-fulldeps") not skipped for "bootstrap::test::UiFullDeps" -- not in ["src/tools/tidy"]
---
Testing error-index stage2
doc tests for: /checkout/obj/build/x86_64-unknown-linux-gnu/test/error-index.md


command did not execute successfully: "/checkout/obj/build/bootstrap/debug/rustdoc" "-Winvalid_codeblock_attributes" "-Dwarnings" "-Znormalize-docs" "--test" "/checkout/obj/build/x86_64-unknown-linux-gnu/test/error-index.md" "--test-args" ""

stdout ----

running 982 tests
---
test /checkout/obj/build/x86_64-unknown-linux-gnu/test/error-index.md - Rust_Compiler_Error_Index::E0779 (line 15704) ... ok

failures:

---- /checkout/obj/build/x86_64-unknown-linux-gnu/test/error-index.md - Rust_Compiler_Error_Index::E0074::_::Note__this_error_code_is_no_longer_emitted_by_the_compiler_ (line 1559) stdout ----
error[E0075]: SIMD vector length must be a power of two
 --> /checkout/obj/build/x86_64-unknown-linux-gnu/test/error-index.md:1563:1
  |
6 | struct Bad<T>(T, T, T);

error: aborting due to previous error

For more information about this error, try `rustc --explain E0075`.
For more information about this error, try `rustc --explain E0075`.
Couldn't compile the test.
---- /checkout/obj/build/x86_64-unknown-linux-gnu/test/error-index.md - Rust_Compiler_Error_Index::E0074::_::Note__this_error_code_is_no_longer_emitted_by_the_compiler_ (line 1568) stdout ----
error[E0075]: SIMD vector length must be a power of two
 --> /checkout/obj/build/x86_64-unknown-linux-gnu/test/error-index.md:1572:1
  |
6 | struct Good(u32, u32, u32);

error: aborting due to previous error

For more information about this error, try `rustc --explain E0075`.
For more information about this error, try `rustc --explain E0075`.
Couldn't compile the test.
---- /checkout/obj/build/x86_64-unknown-linux-gnu/test/error-index.md - Rust_Compiler_Error_Index::E0076 (line 1619) stdout ----
error[E0075]: SIMD vector length must be a power of two
 --> /checkout/obj/build/x86_64-unknown-linux-gnu/test/error-index.md:1623:1
  |
6 | struct Good(u32, u32, u32); // ok!

error: aborting due to previous error

For more information about this error, try `rustc --explain E0075`.
For more information about this error, try `rustc --explain E0075`.
Couldn't compile the test.
---- /checkout/obj/build/x86_64-unknown-linux-gnu/test/error-index.md - Rust_Compiler_Error_Index::E0077 (line 1644) stdout ----
error[E0075]: SIMD vector length must be a power of two
 --> /checkout/obj/build/x86_64-unknown-linux-gnu/test/error-index.md:1648:1
  |
6 | struct Good(u32, u32, u32); // ok!

error: aborting due to previous error

For more information about this error, try `rustc --explain E0075`.

KodrAus · 2021-01-25T23:39:20Z

@bors try

bors · 2021-01-25T23:39:35Z

⌛ Trying commit a4bab7c with merge aed835460397ceb814007e0d97479d7c98844d78...

bors · 2021-01-26T00:40:08Z

☀️ Try build successful - checks-actions
Build commit: aed835460397ceb814007e0d97479d7c98844d78 (aed835460397ceb814007e0d97479d7c98844d78)

KodrAus · 2021-01-26T00:51:48Z

@craterbot check

We'd like to do a crater run to see what the impact of our changes to #[repr(simd)] are here. I thought a check-only run should be sufficient, since errors are likely to surface as monomorphization failures.

craterbot · 2021-01-26T00:51:52Z

👌 Experiment pr-80652 created and queued.
🤖 Automatically detected try build aed835460397ceb814007e0d97479d7c98844d78
⚠️ Try build based on commit f4eb5d9, but latest commit is a4bab7c. Did you forget to make a new try build?
🔍 You can check out the queue and this experiment's details.

ℹ️ Crater is a tool to run experiments across parts of the Rust ecosystem. Learn more

bjorn3 · 2021-02-06T11:18:47Z

Spurious

paranoia fails to link to paranoia_sys
typ overflows while evaluating a trait bound

Real regressions

vek (various versions) and vrust (version 0.0.1) use size 3 vectors and vrust also uses a size 9 vector to represent a 3x3 matrix. vek only uses #[repr(simd)] when both the nightly and repr_simd features are enabled.

calebzulawski · 2021-02-06T16:57:02Z

vrust appears to be unmaintained and has no reverse dependencies, but vek seems quite active and a bunch of crates depend on it. Still, 2 crates doesn't seem so bad. Unsurprisingly, both are doing something that should really be a wrapper over std::simd once it exists. Thoughts on how to proceed?

bjorn3 · 2021-02-06T17:36:24Z

Given that vek only uses #[repr(simd)] when nightly features are enabled, I did suggest to inform the author(s) of vek. Once they release a new version with this fixed I think it would be fine to land this PR. If they don't release a new version within a reasonable amount of time, we will have to decide if it is acceptable to break it.

calebzulawski · 2021-02-06T17:54:50Z

That sounds like a good plan, if that sounds good to everyone I can open an issue over there.

scottmcm · 2021-02-06T19:17:29Z

Remember that this is nightly-only, so any forewarning about breakage is just about being nice, not required.

Something like opening an issue on their repos now then merging a week or later sounds more than sufficient to me.

calebzulawski · 2021-02-07T13:01:51Z

vek has been updated, so I believe we should be ready to go!

nagisa · 2021-02-07T13:35:00Z

@bors r+ rollup=never

bors · 2021-02-07T13:35:01Z

📌 Commit a4bab7c has been approved by nagisa

bors · 2021-02-07T22:25:20Z

⌛ Testing commit a4bab7c with merge bb587b1...

bors · 2021-02-08T01:13:29Z

☀️ Test successful - checks-actions
Approved by: nagisa
Pushing bb587b1 to master...

rust-highfive · 2021-02-08T01:14:42Z

📣 Toolstate changed by #80652!

Tested on commit bb587b1.
Direct link to PR: #80652

💔 miri on windows: test-pass → test-fail (cc @oli-obk @eddyb @RalfJung).
💔 miri on linux: test-pass → test-fail (cc @oli-obk @eddyb @RalfJung).

@oli-obk

Tested on commit rust-lang/rust@bb587b1. Direct link to PR: <rust-lang/rust#80652> 💔 miri on windows: test-pass → test-fail (cc @oli-obk @eddyb @RalfJung). 💔 miri on linux: test-pass → test-fail (cc @oli-obk @eddyb @RalfJung).

RalfJung · 2021-02-08T09:01:07Z

Attempts to catch invalid-sized vectors during typeck when possible.

So "invalid" here means "not a power of two or too big"?

@oli-obk @lcnr FYI -- this seems to be a new kind of post-monomorphization error (see OP).

lcnr · 2021-02-08T14:27:28Z

Are you leaking this behavior to stable or does it require unstable features to trigger?

Preventing post-monomorphization errors here is definitely possible and desirable

bjorn3 · 2021-02-08T14:33:14Z

#[repr(simd)] is unstable.

Implement floating point SIMD intrinsics over all vector widths, and …

07db2bf

…limit SIMD vector lengths.

rust-highfive assigned estebank Jan 3, 2021

rust-highfive added the S-waiting-on-review Status: Awaiting review from the assignee but also interested parties. label Jan 3, 2021

rust-highfive assigned nagisa and unassigned estebank Jan 18, 2021

nagisa reviewed Jan 22, 2021

View reviewed changes

compiler/rustc_typeck/src/check/check.rs Outdated Show resolved Hide resolved

compiler/rustc_middle/src/ty/layout.rs Outdated Show resolved Hide resolved

compiler/rustc_codegen_llvm/src/intrinsic.rs Outdated Show resolved Hide resolved

Fix maximum SIMD lane count, and expose it to other crates. Disallow …

8451656

…SIMD vectors with non-power-of-two lengths.

Make declare_cfn more flexible

4d72ed6

Move SIMD tests

f39f1a4

Update docs with powers-of-two

a4bab7c

craterbot added S-waiting-on-crater Status: Waiting on a crater run to be completed. and removed S-waiting-on-review Status: Awaiting review from the assignee but also interested parties. labels Jan 26, 2021

This was referenced Feb 7, 2021

Remove length 3 repr_simd vectors yoanlcq/vek#66

Closed

Support non-power-of-two vector lengths. rust-lang/portable-simd#63

Open

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 Feb 7, 2021

bors mentioned this pull request Feb 7, 2021

Implement --extern-location #72603

Merged

bors added the merged-by-bors This PR was explicitly merged by bors. label Feb 8, 2021

bors merged commit bb587b1 into rust-lang:master Feb 8, 2021

rustbot added this to the 1.52.0 milestone Feb 8, 2021

rust-highfive mentioned this pull request Feb 8, 2021

miri no longer builds after rust-lang/rust#80652 #81868

Closed

RalfJung mentioned this pull request Feb 8, 2021

Remove non-power-of-two SIMD vectors rust-lang/miri#1703

Merged

oli-obk mentioned this pull request Feb 8, 2021

Tracking issue for SIMD support #27731

Closed

calebzulawski deleted the simd-lanes branch February 9, 2021 04:13

calebzulawski mentioned this pull request Feb 11, 2021

ICE with unusual vector sizes rust-lang/portable-simd#53

Closed

caelunshun mentioned this pull request Feb 14, 2021

Build broken on nightly yoanlcq/vek#67

Closed

Arc-blroth mentioned this pull request Feb 20, 2021

Changes to #[repr(simd)] in nightly break XYZ structs on SPIR-V targets bitshifter/glam-rs#132

Closed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Improve SIMD type element count validation #80652

Improve SIMD type element count validation #80652

calebzulawski commented Jan 3, 2021

rust-highfive commented Jan 3, 2021

workingjubilee commented Jan 6, 2021

workingjubilee commented Jan 6, 2021

KodrAus commented Jan 18, 2021

nagisa left a comment

nagisa commented Jan 22, 2021

bors commented Jan 22, 2021

bors commented Jan 22, 2021

calebzulawski commented Jan 23, 2021

rust-log-analyzer commented Jan 23, 2021

rust-log-analyzer commented Jan 23, 2021

rust-log-analyzer commented Jan 23, 2021

KodrAus commented Jan 25, 2021

bors commented Jan 25, 2021

bors commented Jan 26, 2021

KodrAus commented Jan 26, 2021

craterbot commented Jan 26, 2021

bjorn3 commented Feb 6, 2021 •

edited

Loading

calebzulawski commented Feb 6, 2021

bjorn3 commented Feb 6, 2021

calebzulawski commented Feb 6, 2021

scottmcm commented Feb 6, 2021

calebzulawski commented Feb 7, 2021

nagisa commented Feb 7, 2021

bors commented Feb 7, 2021

bors commented Feb 7, 2021

bors commented Feb 8, 2021

rust-highfive commented Feb 8, 2021

RalfJung commented Feb 8, 2021

lcnr commented Feb 8, 2021

bjorn3 commented Feb 8, 2021

Improve SIMD type element count validation #80652

Improve SIMD type element count validation #80652

Conversation

calebzulawski commented Jan 3, 2021

rust-highfive commented Jan 3, 2021

workingjubilee commented Jan 6, 2021

workingjubilee commented Jan 6, 2021

KodrAus commented Jan 18, 2021

nagisa left a comment

Choose a reason for hiding this comment

nagisa commented Jan 22, 2021

bors commented Jan 22, 2021

bors commented Jan 22, 2021

calebzulawski commented Jan 23, 2021

rust-log-analyzer commented Jan 23, 2021

rust-log-analyzer commented Jan 23, 2021

rust-log-analyzer commented Jan 23, 2021

KodrAus commented Jan 25, 2021

bors commented Jan 25, 2021

bors commented Jan 26, 2021

KodrAus commented Jan 26, 2021

craterbot commented Jan 26, 2021

bjorn3 commented Feb 6, 2021 • edited Loading

Spurious

Real regressions

calebzulawski commented Feb 6, 2021

bjorn3 commented Feb 6, 2021

calebzulawski commented Feb 6, 2021

scottmcm commented Feb 6, 2021

calebzulawski commented Feb 7, 2021

nagisa commented Feb 7, 2021

bors commented Feb 7, 2021

bors commented Feb 7, 2021

bors commented Feb 8, 2021

rust-highfive commented Feb 8, 2021

RalfJung commented Feb 8, 2021

lcnr commented Feb 8, 2021

bjorn3 commented Feb 8, 2021

bjorn3 commented Feb 6, 2021 •

edited

Loading