Fix uses of invalid LLVM intrinsics #1807
Conversation
- The 2nd argument of the LLVM intrinsic should be IMMARG
- use correct intrinsic for unpackl
- fix invalid use of `simd_{or,and,xor}` on floating point vectors
- `vec_search_string` should require `vector-enhancements-2`
- add `-Zverify-llvm-ir` in testsuite
|
Yes those functions belong to |
| @@ -21420,7 +21420,7 @@ pub fn vrbit_s8(a: int8x8_t) -> int8x8_t { | |||
| unsafe extern "unadjusted" { | |||
| #[cfg_attr( | |||
| any(target_arch = "aarch64", target_arch = "arm64ec"), | |||
| link_name = "llvm.aarch64.neon.rbit.v8i8" | |||
| link_name = "llvm.bitreverse.v8i8" | |||
There was a problem hiding this comment.
Note to self: change these in cg_clif once this PR merges and rust-lang/rust updates stdarch.
|
Looks like llvm/llvm-project#73588 is biting us, I guess we have to use the arm-specific one |
|
Can you reproduce that failure in isolation? this seems to work fine https://godbolt.org/z/Mozez3P8f so I must be missing something |
|
an https://godbolt.org/z/b1o748T5a does it, because it uses the 32-bit version. I was confused because the example from that llvm issue does work when I put it into godbolt |
| - LLVMLink: | ||
| name: "rbit.{neon_type}" | ||
| links: | ||
| - link: "llvm.aarch64.neon.rbit.{neon_type}" | ||
| - link: "llvm.bitreverse.{neon_type}" | ||
| arch: aarch64,arm64ec |
There was a problem hiding this comment.
Can't we use simd_bitreverse directly instead? That does seem to happen for e.g. simd_and
There was a problem hiding this comment.
I don't know too much about how stdarch-gen-arm operates, particularly I don't know how I can change the impl. Still, I would try
There was a problem hiding this comment.
I mean model it after this:
- name: "vand{neon_type.no}"
doc: Vector bitwise and
arguments: ["a: {neon_type}", "b: {neon_type}"]
return_type: "{neon_type}"
attr:
- *neon-v7
- FnCall: [cfg_attr, [*test-is-arm, {FnCall: [assert_instr, [vand]]}]]
- FnCall: [cfg_attr, [*neon-target-aarch64-arm64ec, {FnCall: [assert_instr, [and]]}]]
- *neon-not-arm-stable
- *neon-cfg-arm-unstable
safety: safe
types:
- int8x8_t
- int8x16_t
- int16x4_t
- int16x8_t
- int32x2_t
- int32x4_t
- uint8x8_t
- uint8x16_t
- uint16x4_t
- uint16x8_t
- uint32x2_t
- uint32x4_t
- int64x1_t
- int64x2_t
- uint64x1_t
- uint64x2_t
compose:
- FnCall:
- simd_and
- - a
- b
see rust-lang/rust#140763 (comment) for context
This PR also starts checking for IR validity in tests, it is almost free, but adds a lot of confidence as we are dealing with LLVM intrinsics