r276417 - [X86][AVX] Added support for lowering to VBROADCASTF128/VBROADCASTI128 with generic IR

[Chandler Carruth via cfe-commits]

On Fri, Jul 22, 2016 at 7:06 AM Simon Pilgrim via cfe-commits <
cfe-commits at lists.llvm.org> wrote:

> Author: rksimon
> Date: Fri Jul 22 08:58:56 2016
> New Revision: 276417
>
> URL: http://llvm.org/viewvc/llvm-project?rev=276417&view=rev
> Log:
> [X86][AVX] Added support for lowering to VBROADCASTF128/VBROADCASTI128
> with generic IR
>
> As discussed on D22460, I've updated the vbroadcastf128 pd256/ps256
> builtins to map directly to generic IR - load+splat a 128-bit vector to
> both lanes of a 256-bit vector.
>
> Fix for PR28657.
>
> Modified:
>     cfe/trunk/lib/CodeGen/CGBuiltin.cpp
>     cfe/trunk/test/CodeGen/avx-builtins.c
>
> Modified: cfe/trunk/lib/CodeGen/CGBuiltin.cpp
> URL:
> http://llvm.org/viewvc/llvm-project/cfe/trunk/lib/CodeGen/CGBuiltin.cpp?rev=276417&r1=276416&r2=276417&view=diff
>
> ==============================================================================
> --- cfe/trunk/lib/CodeGen/CGBuiltin.cpp (original)
> +++ cfe/trunk/lib/CodeGen/CGBuiltin.cpp Fri Jul 22 08:58:56 2016
> @@ -6619,6 +6619,26 @@ static Value *EmitX86MaskedLoad(CodeGenF
>    return CGF.Builder.CreateMaskedLoad(Ops[0], Align, MaskVec, Ops[1]);
>  }
>
> +static Value *EmitX86SubVectorBroadcast(CodeGenFunction &CGF,
> +                                        SmallVectorImpl<Value *> &Ops,
> +                                        llvm::Type *DstTy,
> +                                        unsigned SrcSizeInBits,
> +                                        unsigned Align) {
> +  // Load the subvector.
> +  Ops[0] = CGF.Builder.CreateAlignedLoad(Ops[0], Align);
> +
> +  // Create broadcast mask.
> +  unsigned NumDstElts = DstTy->getVectorNumElements();
> +  unsigned NumSrcElts = SrcSizeInBits / DstTy->getScalarSizeInBits();
> +
> +  SmallVector<uint32_t, 8> Mask;
> +  for (unsigned i = 0; i != NumDstElts; i += NumSrcElts)
> +    for (unsigned j = 0; j != NumSrcElts; ++j)
> +      Mask.push_back(j);
> +
> +  return CGF.Builder.CreateShuffleVector(Ops[0], Ops[0], Mask,
> "subvecbcst");
> +}
> +
>  static Value *EmitX86Select(CodeGenFunction &CGF,
>                              Value *Mask, Value *Op0, Value *Op1) {
>
> @@ -6995,6 +7015,13 @@ Value *CodeGenFunction::EmitX86BuiltinEx
>
>  getContext().getTypeAlignInChars(E->getArg(1)->getType()).getQuantity();
>      return EmitX86MaskedLoad(*this, Ops, Align);
>    }
> +
> +  case X86::BI__builtin_ia32_vbroadcastf128_pd256:
> +  case X86::BI__builtin_ia32_vbroadcastf128_ps256: {
> +    llvm::Type *DstTy = ConvertType(E->getType());
> +    return EmitX86SubVectorBroadcast(*this, Ops, DstTy, 128, 16);
>

Somewhat to my surprise, after a bunch of debugging, we found a bug in this
line.

See my fix in r278202. I wanted to mention it here in case others bisect
back to this and wonder. And because frankly, I would never have thought of
this. The broadcast instructions, even when taking a 128-bit input, don't
have an alignment requirement here. Paint me surprised.

Anyways, just FYI and in case you want to double check my fix.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/cfe-commits/attachments/20160810/d95ef966/attachment.html>