[DO NOT MERGE] Add subtraction, rounding avx instrs

[AdamNiederer]

Hi all,
I've implemented subpd256, subps256, roundpd256, ceilpd256, floorpd256, and their associated tests. It's still a bit of a work in progress

• Lines 67:78 are blatantly plagiarized from LLVM
• _mm_sub_ps256 is failing its test
• assert_instr is throwing for _mm_round_pd

I'll try to iron those out tomorrow.

[alexcrichton]

Thanks for the PR! What do the test failures look like if you add assert_instr?

[AdamNiederer]

LLVM ERROR: Cannot select: intrinsic %llvm.x86.avx.round.pd.256

My intrinsics reference says there is something for VROUNDPD, though

let TargetPrefix = "x86" in {  // All intrinsics start with "llvm.x86."
  ...
  def int_x86_avx_round_pd_256 : GCCBuiltin<"__builtin_ia32_roundpd256">,
        Intrinsic<[llvm_v4f64_ty], [llvm_v4f64_ty,
                  llvm_i32_ty], [IntrNoMem]>;
}

https://gist.githubusercontent.com/anonymous/a25d3e3b4c14ee68d63bd1dcb0e1223c/raw/fb798be53a7b29f9b28a0c7a9e2b9a16ae915f71/allintrinsics

I'm not 100% sure how I should proceed; I've tried some permutations on the name of the intrinsic.

[alexcrichton]

I think this means that the argument must be an immediate? Right now it takes an i32, but should it take a byte instead?

[AdamNiederer]

I've tried it with both an i8 and a u8, which both fail to compile. It looks like the icc intrinsic takes a c-int (http://www.felixcloutier.com/x86/ROUNDPD.html). Does the immediate have to be a constant expression?

[alexcrichton]

Ah yeah you'll find a few macros (I think constify* style) in the rest of the codebase to turn variables into immediates

[AdamNiederer]

@alexcrichton We should be good to go now; the broken appveyor build appears to be unrelated to this change.

[alexcrichton]

Looks great, thanks @AdamNiederer!