LLVM emits single precision variation of double precision intrinsics

[p32blo]

The affected intrinsics are:

AVX Intrinsic	Expected Instruction	Emitted Instruction
_mm256_and_pd	vandpd	vandps
_mm256_andnot_pd	vandnpd	vandnps
_mm256_or_pd	vorpd	vorps
_mm256_xor_pd	vxorpd	vxorps

Checking in clang here confirms this behavior. Note how gcc emits the correct instructions while clang always uses the single precision variation.

The result does not seem to be affected, since these are bitwise operations on vectors of the same size. However, since the expected instructions are not generated, the assert_* tests will fail if this behavior is not taken into account.

[alexcrichton]

Ok I'm going to close this for now as I think in general we're not going to be able to fix this but we're ruling this as "it's ok". LLVM in general will performan lots of optimizations under the hood, so we can't really guarantee precise intructions unfortunately!

[avl]

Don't vandpd and vandps do exactly the same thing anyway? Are they actually emitted as different opcodes by the assembler?

[gnzlbg]

Yes and yes.