Improper usage of #[inline(always)] and #[target_feature] results in LLVM instruction selection error

[shepmaster]

#![feature(stdsimd)]
#![feature(target_feature)]

use std::arch::x86_64::*;

#[inline(always)]
#[target_feature(enable = "sse4.2")]
unsafe fn alpha(needle: &[u8], haystack: &[u8]) -> i32 {
    let a = _mm_loadu_si128(needle.as_ptr() as *const _);
    let b = _mm_loadu_si128(haystack.as_ptr() as *const _);

    _mm_cmpestri(a, 3, b, 15, _SIDD_CMP_EQUAL_ORDERED)
}

fn main() {
    let haystack = b"Split \r\n\t line  ";
    let needle = b"\r\n\t ignore this ";

    let idx = unsafe { alpha(needle, haystack) };
    assert_eq!(idx, 6);
}

LLVM ERROR: Cannot select: 0x7fa0f8e25ea0: i32,i32 = X86ISD::PCMPESTRI 0x7fa0f8e25410, Constant:i32<3>, 0x7fa0f8e259c0, Constant:i32<15>, Constant:i8<12>
  0x7fa0f8e25410: v16i8 = bitcast 0x7fa0f8e257b8
    0x7fa0f8e257b8: v2i64,ch = load<LD16[ConstantPool]> 0x7fa0f8e7b040, 0x7fa0f8e25208, undef:i64
      0x7fa0f8e25208: i64 = X86ISD::WrapperRIP TargetConstantPool:i64<<16 x i8> <i8 13, i8 10, i8 9, i8 32, i8 105, i8 103, i8 110, i8 111, i8 114, i8 101, i8 32, i8 116, i8 104, i8 105, i8 115, i8 32>> 0
        0x7fa0f98fa3a8: i64 = TargetConstantPool<<16 x i8> <i8 13, i8 10, i8 9, i8 32, i8 105, i8 103, i8 110, i8 111, i8 114, i8 101, i8 32, i8 116, i8 104, i8 105, i8 115, i8 32>> 0
      0x7fa0f8e25d00: i64 = undef
  0x7fa0f8e25138: i32 = Constant<3>
  0x7fa0f8e259c0: v16i8 = bitcast 0x7fa0f8e25618
    0x7fa0f8e25618: v2i64,ch = load<LD16[ConstantPool]> 0x7fa0f8e7b040, 0x7fa0f8e25c30, undef:i64
      0x7fa0f8e25c30: i64 = X86ISD::WrapperRIP TargetConstantPool:i64<<16 x i8> <i8 83, i8 112, i8 108, i8 105, i8 116, i8 32, i8 13, i8 10, i8 9, i8 32, i8 108, i8 105, i8 110, i8 101, i8 32, i8 32>> 0
        0x7fa0f98fa138: i64 = TargetConstantPool<<16 x i8> <i8 83, i8 112, i8 108, i8 105, i8 116, i8 32, i8 13, i8 10, i8 9, i8 32, i8 108, i8 105, i8 110, i8 101, i8 32, i8 32>> 0
      0x7fa0f8e25d00: i64 = undef
  0x7fa0f8e25d68: i32 = Constant<15>
  0x7fa0f8e25750: i8 = Constant<12>
In function: _ZN10playground4main17ha4aeba89c162948dE

1.26.0-nightly (2018-03-25 482a913fb337855072a53c0d602cd19947f45285)

[alexcrichton]

Thanks! I'll poke around at this soon

[gnzlbg]

The RFC2045 target_feature states that inlining a function into another function that matches or extends its feature set is ok, but "inlining is forbidden in the opposite case". The Example 1 shows:

#[target_feature(enable = "avx")] unsafe fn foo();
#[target_feature(enable = "avx")] #[inline] unsafe fn baz(); // OK
#[target_feature(enable = "avx")] #[inline(always)] unsafe fn bar(); // OK  

#[target_feature(enable = "sse3")]
unsafe fn moo() {
  // This function supports SSE3 but not AVX
  if cfg_feature_enabled!("avx") {
      foo(); // OK: foo is not inlined into moo
      baz(); // OK: baz is not inlined into moo
      bar(); 
      // ^ ERROR: bar cannot be inlined across mismatching features
      // did you meant to make bar #[inline] instead of #[inline(always)]?
      // Note: the logic to detect this is the same as for the call 
      // to baz, but in this case rustc must emit an error because an
      // #[inline(always)] function cannot be inlined in this call site. 
  }
}

We should probably suggest doing both things:

• making bar #[inline] instead of #[inline(always)] if bar is the part of the current crate
• making moo #[target_feature = "avx"] (independently of whether bar is part of the current crate or not)

[alexcrichton]

I've applied a hammer for this rust-lang/rust#49425 which is different than RFC 2045 (but easier to implement)

[gnzlbg]

@alexcrichton I've skimmed through the RFC comments, and it looks like the RFC wasn't properly updated. In the comments you argued for the fix that you proposed in this PR, and it was agreed upon that it is the right step to take. Allowing #[inline(always)] + #[target_feature] can be done in a backwards compatible way, so it can be resolved later.

[Amanieu]

coresimd doesn't build on the latest nightly because of this change.

[gnzlbg]

@Amanieu which coresimd ?

[sunchao]

I'm also getting a bunch of errors when trying to compile coresimd with nightly rustc, possibly due to the change rust-lang/rust#49425

......
error: cannot use #[inline(always)] with #[target_feature]
  --> /Users/chao/.cargo/registry/src/github.jparrowsec.cn-1ecc6299db9ec823/coresimd-0.0.4/src/x86/x86_64/xsave.rs:45:1
   |
45 | #[inline(always)]
   | ^^^^^^^^^^^^^^^^^

error: cannot use #[inline(always)] with #[target_feature]
  --> /Users/chao/.cargo/registry/src/github.jparrowsec.cn-1ecc6299db9ec823/coresimd-0.0.4/src/x86/x86_64/xsave.rs:59:1
   |
59 | #[inline(always)]
   | ^^^^^^^^^^^^^^^^^

error: cannot use #[inline(always)] with #[target_feature]
  --> /Users/chao/.cargo/registry/src/github.jparrowsec.cn-1ecc6299db9ec823/coresimd-0.0.4/src/x86/x86_64/xsave.rs:72:1
   |
72 | #[inline(always)]
   | ^^^^^^^^^^^^^^^^^

error: cannot use #[inline(always)] with #[target_feature]
  --> /Users/chao/.cargo/registry/src/github.jparrowsec.cn-1ecc6299db9ec823/coresimd-0.0.4/src/x86/x86_64/xsave.rs:86:1
   |
86 | #[inline(always)]
   | ^^^^^^^^^^^^^^^^^

error: cannot use #[inline(always)] with #[target_feature]
   --> /Users/chao/.cargo/registry/src/github.jparrowsec.cn-1ecc6299db9ec823/coresimd-0.0.4/src/x86/x86_64/xsave.rs:102:1

[gnzlbg]

The answer is probably to migrate to @AdamNiederer 's coresimd fork here: https://github.com/AdamNiederer/stdsimd/tree/fix-bft-build for the time being till we do a new 0.0.5 release, or to migrate to core::arch / core::simd.