ICE when using `feature(stdsimd)` and IFMA intrinsics on non-IFMA platforms

[hdevalence]

Hi, while working on dalek-cryptography/curve25519-dalek#257 , I ran into an ICE.

On a Cannonlake CPU with target_feature=+avx512ifma, the branch in the linked PR works fine. On a Skylake CPU (without IFMA), the feature-gating should disable all of the IFMA code at compile-time. But instead, running cargo test causes an ICE:

~/c/curve25519-dalek (use_upstream_intrinsics|✔) $ cargo test --features "simd_backend"
   Compiling curve25519-dalek v1.1.4 (/home/hdevalence/code/curve25519-dalek)
warning: unused import: `ValidityCheck`
    --> src/ristretto.rs:1084:28
     |
1084 |     use traits::{Identity, ValidityCheck};
     |                            ^^^^^^^^^^^^^
     |
     = note: #[warn(unused_imports)] on by default

thread 'rustc' panicked at 'Forcing query with already existing DepNode.
- query-key: ParamEnvAnd { param_env: ParamEnv { caller_bounds: [], reveal: All, def_id: None }, value: (Const { ty: [u32; 8], val: ByRef(AllocId(4713).0x0, Allocation { bytes: [0, 0, 0, 0, 1, 0, 0, 0, 2, 0, 0, 0, 3, 0, 0, 0, 4, 0, 0, 0, 5, 0, 0, 0, 6, 0, 0, 0, 7, 0, 0, 0], relocations: Relocations(SortedMap { data: [] }), undef_mask: UndefMask { blocks: [4294967295], len: Size { raw: 32 } }, align: Align { pow2: 2 }, mutability: Immutable, extra: () }) }, field[0]) }
- dep-node: const_field(eb99858e4874f291-9a55d263e1a1abdf)', src/librustc/ty/query/plumbing.rs:548:9
note: Run with `RUST_BACKTRACE=1` environment variable to display a backtrace.

error: internal compiler error: unexpected panic

note: the compiler unexpectedly panicked. this is a bug.

note: we would appreciate a bug report: https://github.com/rust-lang/rust/blob/master/CONTRIBUTING.md#bug-reports

note: rustc 1.37.0-nightly (6ffb8f53e 2019-06-03) running on x86_64-unknown-linux-gnu

note: compiler flags: -C debuginfo=2 -C incremental -C target_cpu=skylake --crate-type lib

note: some of the compiler flags provided by cargo are hidden

error: Could not compile `curve25519-dalek`.
warning: build failed, waiting for other jobs to finish...
thread 'rustc' panicked at 'Forcing query with already existing DepNode.
- query-key: ParamEnvAnd { param_env: ParamEnv { caller_bounds: [], reveal: All, def_id: None }, value: (Const { ty: [u32; 8], val: ByRef(AllocId(6310).0x0, Allocation { bytes: [0, 0, 0, 0, 1, 0, 0, 0, 2, 0, 0, 0, 3, 0, 0, 0, 4, 0, 0, 0, 5, 0, 0, 0, 6, 0, 0, 0, 7, 0, 0, 0], relocations: Relocations(SortedMap { data: [] }), undef_mask: UndefMask { blocks: [4294967295], len: Size { raw: 32 } }, align: Align { pow2: 2 }, mutability: Immutable, extra: () }) }, field[0]) }
- dep-node: const_field(eb99858e4874f291-9a55d263e1a1abdf)', src/librustc/ty/query/plumbing.rs:548:9
note: Run with `RUST_BACKTRACE=1` environment variable to display a backtrace.

error: internal compiler error: unexpected panic

note: the compiler unexpectedly panicked. this is a bug.

note: we would appreciate a bug report: https://github.com/rust-lang/rust/blob/master/CONTRIBUTING.md#bug-reports

note: rustc 1.37.0-nightly (6ffb8f53e 2019-06-03) running on x86_64-unknown-linux-gnu

note: compiler flags: -C debuginfo=2 -C incremental -C target_cpu=skylake

note: some of the compiler flags provided by cargo are hidden

error: Could not compile `curve25519-dalek`.

To learn more, run the command again with --verbose.

(I think the reason the error appears twice is just because the crate loads itself into its own build.rs, not anything else.)

[mati865]

Could you post exact reproduction steps?
This is what I tried:

 mateusz@ubuntu  /tmp  git clone https://github.com/dalek-cryptography/curve25519-dalek.git                                          
Cloning into 'curve25519-dalek'...
remote: Enumerating objects: 89, done.
remote: Counting objects: 100% (89/89), done.
remote: Compressing objects: 100% (55/55), done.
remote: Total 5149 (delta 51), reused 59 (delta 34), pack-reused 5060
Receiving objects: 100% (5149/5149), 2.26 MiB | 3.34 MiB/s, done.
Resolving deltas: 100% (3465/3465), done.
 mateusz@ubuntu  /tmp  curve25519-dalek/
 mateusz@ubuntu  /tmp/curve25519-dalek   master  git checkout use_upstream_intrinsics 
Branch 'use_upstream_intrinsics' set up to track remote branch 'use_upstream_intrinsics' from 'origin'.
Switched to a new branch 'use_upstream_intrinsics'
 mateusz@ubuntu  /tmp/curve25519-dalek   use_upstream_intrinsics  cargo test --features "simd_backend"

[long output]

test result: ok. 25 passed; 0 failed; 0 ignored; 0 measured; 0 filtered out

CPU: Ryzen 2700X without IFMA

[hdevalence]

Hmm, following exactly those steps on my machine (i7-8650U) reproduces the ICE, using rustc 1.37.0-nightly (6ffb8f53e 2019-06-03).

My machine builds with -C target_cpu=native by default, but I can also reproduce using -C target_cpu=skylake. Does setting that change anything?

[mati865]

I was able to reproduce it on your branch with -C target_cpu=native and reduced it to https://github.com/mati865/curve25519-dalek/blob/ice/src/backend/vector/avx2/field.rs (it crashes without changing target_cpu but still depends on packed_simd).

Reproduction steps:

git clone https://github.com/mati865/curve25519-dalek -b ice
cd curve25519-dalek
cargo build

Click here for build output:

Updating crates.io index
   Compiling packed_simd v0.3.3
   Compiling cfg-if v0.1.9
   Compiling curve25519-dalek v1.1.4 (/tmp/tmp/curve25519-dalek)
thread 'rustc' panicked at 'Forcing query with already existing DepNode.
- query-key: ParamEnvAnd { param_env: ParamEnv { caller_bounds: [], reveal: All, def_id: None }, value: (Const { ty: [u32; 8], val: ByRef(AllocId(172).0x0, Allocation { bytes: [0, 0, 0, 0, 1, 0, 0, 0, 2, 0, 0, 0, 3, 0, 0, 0, 4, 0, 0, 0, 5, 0, 0, 0, 6, 0, 0, 0, 7, 0, 0, 0], relocations: Relocations(SortedMap { data: [] }), undef_mask: UndefMask { blocks: [4294967295], len: Size { raw: 32 } }, align: Align { pow2: 2 }, mutability: Immutable, extra: () }) }, field[0]) }
- dep-node: const_field(eb99858e4874f291-9a55d263e1a1abdf)', src/librustc/ty/query/plumbing.rs:548:9
note: Run with `RUST_BACKTRACE=1` environment variable to display a backtrace.

error: internal compiler error: unexpected panic

note: the compiler unexpectedly panicked. this is a bug.

note: we would appreciate a bug report: https://github.com/rust-lang/rust/blob/master/CONTRIBUTING.md#bug-reports

note: rustc 1.37.0-nightly (7cdaffd79 2019-06-05) running on x86_64-unknown-linux-gnu

note: compiler flags: -C debuginfo=2 -C incremental --crate-type lib

note: some of the compiler flags provided by cargo are hidden

error: Could not compile `curve25519-dalek`.

To learn more, run the command again with --verbose.

This is recent regression and bisection leads to #61062
cc @mark-i-m @eddyb @oli-obk

[gnzlbg]

This fails on x86_64-apple-darwin, but I can't reproduce on the playground:

#![feature(
    platform_intrinsics,
    repr_simd,
)]

extern "platform-intrinsic" {
    pub fn simd_shuffle2<T, U>(x: T, y: T, idx: [u32; 2]) -> U;
}

#[repr(simd)]
#[derive(Copy, Clone)]
pub struct T(f32, f32);

pub unsafe fn foo(b: T) -> T {
    simd_shuffle2(b, b, [0, 1])
}
pub unsafe fn bar(b: T) -> T {
    simd_shuffle2(b, b, [0, 1])
}

with

thread 'rustc' panicked at 'Forcing query with already existing DepNode.
- query-key: ParamEnvAnd { param_env: ParamEnv { caller_bounds: [], reveal: All, def_id: None }, value: (Const { ty: [u32; 2], val: ByRef(AllocId(3).0x0, Allocation { bytes: [0, 0, 0, 0, 1, 0, 0, 0], relocations: Relocations(SortedMap { data: [] }), undef_mask: UndefMask { blocks: [255], len: Size { raw: 8 } }, align: Align { pow2: 2 }, mutability: Immutable, extra: () }) }, field[0]) }
- dep-node: const_field(67586bc8d6719bbd-8210aca0f024cb22)', src/librustc/ty/query/plumbing.rs:548:9
note: run with `RUST_BACKTRACE=1` environment variable to display a backtrace.

error: internal compiler error: unexpected panic

note: the compiler unexpectedly panicked. this is a bug.

note: we would appreciate a bug report: https://github.com/rust-lang/rust/blob/master/CONTRIBUTING.md#bug-reports

note: rustc 1.37.0-nightly (5eeb567a2 2019-06-06) running on x86_64-apple-darwin

note: compiler flags: -C debuginfo=2 -C incremental --crate-type lib

note: some of the compiler flags provided by cargo are hidden

[gnzlbg]

FYI rust-lang/rust is not running stdsimd tests, and stdsimd functions are all #[inline], so they don't trigger this issue when just trying to build the crate. The moment a user tries to call one of the functions, these errors pop up.

[eddyb]

cc @Zoxc (sadly, @michaelwoerister is on leave)

[gnzlbg]

@eddyb Maybe this should be P-High, Nominated, etc. This does not compile on nightly anymore:

use core::{arch::x86_64::*, mem};
pub unsafe fn foo() {
    let x: __m128 = mem::zeroed();
    let x = _mm_cmpgt_ss(x, x);
    let x = _mm_cmpge_ss(x, x);
}

Note that this does not use any feature(...). That is, whatever broke this, broke core::arch on stable Rust.

[eddyb]

@gnzlbg Given const_field was added as a query in #61062, I suspect it's that PR.
It, and type_name, were added as eval_always and no_force, and I don't think either flag is needed.
Turns out that's not actually helping. Maybe there's a bug somewhere else ("stable hashing" vs interning for ty::Const? since this has an Allocation, maybe "stable hashing" is by value?)

[eddyb]

Self-contained reproduction:

#![feature(repr_simd, platform_intrinsics)]

#[repr(simd)]
struct I32x2(i32, i32);

extern "platform-intrinsic" {
    fn simd_shuffle2<T, U>(x: T, y: T, idx: [u32; 2]) -> U;
}

fn main() {
    unsafe {
        let _: I32x2 = simd_shuffle2(I32x2(1, 2), I32x2(3, 4), [0, 0]);
        let _: I32x2 = simd_shuffle2(I32x2(1, 2), I32x2(3, 4), [0, 0]);
    }
}

However, the tricky bit is that you need to pass -Cincremental=some-directory if you want to use rustc directly (and that also means it doesn't repro on the playground).
That's because the bit that's ICE-ing is specific to incremental (re)compilation (but it triggers right away).

This confirms my suspicion that it is const_field(a) and const_field(b) where a != b (compares AllocIds which are distinct) but hash_stable(a) == hash_stable(b) (hashes the contents).