Description
Starting with last night's nightly our CI is failing on 32-bit platforms for reasons that look like:
---- coresimd::x86::sse2::tests::test_mm_add_sd stdout ----
thread 'main' panicked at '__m128d(NaN, 2.0) != __m128d(6.0, 2.0)', crates/coresimd/src/../../../coresimd/x86/test.rs:26:9
note: Run with `RUST_BACKTRACE=1` for a backtrace.
---- coresimd::x86::sse2::tests::test_mm_cvtsd_f64 stdout ----
thread 'main' panicked at 'assertion failed: `(left == right)`
left: `NaN`,
right: `-1.1`', crates/coresimd/src/../../../coresimd/x86/sse2.rs:4986:9
---- coresimd::x86::sse2::tests::test_mm_div_sd stdout ----
thread 'main' panicked at '__m128d(NaN, 2.0) != __m128d(0.2, 2.0)', crates/coresimd/src/../../../coresimd/x86/test.rs:26:9
---- coresimd::x86::sse2::tests::test_mm_mul_sd stdout ----
thread 'main' panicked at '__m128d(NaN, 2.0) != __m128d(5.0, 2.0)', crates/coresimd/src/../../../coresimd/x86/test.rs:26:9
---- coresimd::x86::sse2::tests::test_mm_sqrt_pd stdout ----
thread 'main' panicked at '__m128d(1.0, 1.4142135623730951) != __m128d(NaN, 1.4142135623730951)', crates/coresimd/src/../../../coresimd/x86/test.rs:26:9
---- coresimd::x86::sse::tests::test_mm_cvtss_f32 stdout ----
thread 'main' panicked at 'assertion failed: `(left == right)`
left: `NaN`,
right: `312.0134`', crates/coresimd/src/../../../coresimd/x86/sse.rs:3549:9
---- coresimd::x86::sse::tests::test_mm_rcp_ps stdout ----
thread 'main' panicked at 'assertion failed: `(left !== right)` (left: `NaN`, right: `0.24993896`, expect diff: `0.0009765625`, real diff: `NaN`)', crates/coresimd/src/../../../coresimd/x86/sse.rs:2670:13
---- coresimd::x86::sse::tests::test_mm_set1_ps stdout ----
thread 'main' panicked at 'assertion failed: `(left == right)`
left: `NaN`,
right: `4.25`', crates/coresimd/src/../../../coresimd/x86/sse.rs:3562:9
---- coresimd::x86::sse::tests::test_mm_stream_ps stdout ----
thread 'main' panicked at 'assertion failed: `(left == right)`
left: `7.0`,
right: `NaN`', crates/coresimd/src/../../../coresimd/x86/sse.rs:3973:13
failures:
coresimd::x86::sse2::tests::test_mm_add_sd
coresimd::x86::sse2::tests::test_mm_cvtsd_f64
coresimd::x86::sse2::tests::test_mm_div_sd
coresimd::x86::sse2::tests::test_mm_mul_sd
coresimd::x86::sse2::tests::test_mm_sqrt_pd
coresimd::x86::sse::tests::test_mm_cvtss_f32
coresimd::x86::sse::tests::test_mm_rcp_ps
coresimd::x86::sse::tests::test_mm_set1_ps
coresimd::x86::sse::tests::test_mm_stream_ps
I've done some debugging and I don't think this is a regression, but rather this has always been a bug and we just haven't exposed it until now. I've been able to reproduce some of the above failures with:
#![feature(stdsimd)]
#![feature(mmx_target_feature)]
use std::arch::x86::*;
fn main() {
unsafe {
something_mmx();
foo();
}
}
#[target_feature(enable = "mmx")]
unsafe fn something_mmx() {
another(_mm_setzero_si64());
#[target_feature(enable = "mmx")]
unsafe fn another(_: __m64) {
}
}
#[target_feature(enable = "sse2")]
unsafe fn foo() {
let a = _mm_setr_pd(1.0, 2.0);
let b = _mm_setr_pd(5.0, 10.0);
let r = _mm_add_sd(a, b);
assert_eq_m128d(r);
}
#[target_feature(enable = "sse2")]
unsafe fn assert_eq_m128d(a: __m128d) {
println!("{:?}", a);
}That program has an invalid NaN in the end, but if you comment out the call to something_mmx() it passes.
My best guess as to what's going on is that MMX is causing things to be sad. Something about floating point state and whatnot seems to be persisting across executions. I've been able to determine some bad stuff starts from a fld instruction, which apparently loads a floating point operand onto the floating point stack. The operand being loaded is 1.0 as a double, but when loaded onto the stack it loads as -NaN for whatever reasons. The fctrl register (according to GDB) differs between the program execute with the call to something_mmx and the program execution without something_mmx, and presumably that's like masking something or tweaking something?
I'm sort of lost at this point and curious if others have any idea what's going on here? Does using MMX things somehow break global state for the rest of the CPU?