[llvm-commits] [PATCH 2/2] FIx PR12312

[Michael Liao]

On Tue, 2012-08-14 at 13:53 -0700, Evan Cheng wrote:
> +  if (VT.is128BitVector())
> +    VecIn = DAG.getNode(ISD::BITCAST, DL, MVT::v2i64, VecIn);
> +  else // FIXME: only 128-bit vector is supported so far.
> +    return SDValue();
> 
> It seems like this check can happen earlier to avoid doing a bunch of work for no good reason.

I am fixing the similar issue on 256-bit as well but the
target-independent code generation needs fixing. So far, this patch only
address 128-bit issue.

As commented in code, I will fix it soon.

Yours
- Michael

> 
> Evan
> 
> On Aug 13, 2012, at 1:25 PM, Michael Liao <michael.liao at intel.com> wrote:
> 
> > Here is another revision of the 2nd part. The new change fixes an
> > potential issue of stale iterator being used. The new loop BFS
> > traversing OR'd tree is now over index instead of iterator as we may add
> > elements in that loop.
> > 
> > Yours
> > - Michael
> > 
> > On Sun, 2012-08-12 at 14:28 -0700, Michael Liao wrote:
> >> On Fri, 2012-08-10 at 14:27 -0700, Evan Cheng wrote:
> >>> +static SDValue FlaggedOrCombine(SDValue Or, X86::CondCode &CC, SelectionDAG &DAG,
> >>> +                                const X86Subtarget *Subtarget) {
> >>> 
> >>> 
> >>> The current naming convention calls for lower cased first letter.
> >> 
> >> The revised patch is attached. I checked the function naming in the 1st
> >> part of patch as well.
> >> 
> >>> 
> >>> 
> >>> +  if (VecIn != DAG.getVectorShuffle(VT, DL, VecIn, VecIn, &Mask[0]))
> >>> +    return SDValue();
> >>> 
> >>> What is this trying to do? Is it potentially creating a vectorshuffle node and then throwing it away? Please don't do that.
> >>> 
> >> 
> >> It is checking whether the all vector elements are used in that OR'd
> >> tree. I revised the code to use a bit mask for this purpose and simplify
> >> both this checking and other parts.
> >> 
> >> Please review the revised patch attached.
> >> 
> >> Yours
> >> - Michael
> >> 
> >>> Evan
> >>> 
> >>> On Aug 1, 2012, at 12:47 PM, Michael Liao <michael.liao at intel.com> wrote:
> >>> 
> >>>> Hi
> >>>> 
> >>>> This patch is to fix http://llvm.org/bugs/show_bug.cgi?id=12312, where a
> >>>> special use of i128 needs efficient code generation with PTEST from
> >>>> SSE4.1.
> >>>> 
> >>>> To fix this issue, 2 patches are developed. The second part is to
> >>>> recognize that special use, casting from <4 x i32> to i128 followed by a
> >>>> zero (or non-zero) checking. That sequence is lowered into OR
> >>>> (X86ISD::OR at the root) tree with all leaves extracted from that vector
> >>>> because i128 is mapped onto GPR pair. The attached patch adds a
> >>>> X86-specific DAG optimization to recognize this OR'd tree and convert it
> >>>> a PTEST if only the flag value is used from that X86ISD::OR root node.
> >>>> 
> >>>> Yours
> >>>> - Michael
> >>>> <0002-add-X86-DAG-optimization-to-recognize-a-pttern-PTEST.patch>_______________________________________________
> >>>> llvm-commits mailing list
> >>>> llvm-commits at cs.uiuc.edu
> >>>> http://lists.cs.uiuc.edu/mailman/listinfo/llvm-commits
> >>> 
> >> 
> >> _______________________________________________
> >> llvm-commits mailing list
> >> llvm-commits at cs.uiuc.edu
> >> http://lists.cs.uiuc.edu/mailman/listinfo/llvm-commits
> > 
> > <0001-add-X86-DAG-optimization-to-recognize-a-pttern-PTEST.patch>
>