[LLVMdev] i64 OR instruction elimination

[Philip Reames]

I would suggest you file a bug for this if you haven't already.

This is rather odd code to be generating.  Is this coming from another 
opto pass or the frontend?  I might be tempted to see if the source is 
easy to tweak.

Philip

On 04/01/2014 04:16 PM, Mark Schimmel wrote:
>
> for.body: ; preds = %for.body.lr.ph, %for.body
>
> *_  %ACC.sroa.0.013 = phi i64 [ 0, %for.body.lr.ph ], [ %phitmp9, 
> %for.body ]_*
>
> *_  %ACC.sroa.4.012 = phi i64 [ 0, %for.body.lr.ph ], [ 
> %tmp.sroa.2.0.extract.shift, %for.body ]_*
>
>   %i.011 = phi i32 [ 0, %for.body.lr.ph ], [ %inc, %for.body ]
>
>   %arrayidx = getelementptr inbounds i16* %ary, i32 %i.011
>
>   %0 = load i16* %arrayidx, align 2, !tbaa !1
>
>   %add = add i32 %i.011, 8
>
>   %arrayidx1 = getelementptr inbounds i16* %ary, i32 %add
>
>   %1 = load i16* %arrayidx1, align 2, !tbaa !1
>
>   %2 = tail call i64 @llvm.arc.ACC40.GUARD.MUL.SQ15.SQ15(i16 %0, i16 
> %1) #1
>
> *_%L.sroa.0.0.insert.insert.i = or i64 %ACC.sroa.0.013, %ACC.sroa.4.012_*
>
>   %3 = tail call i64 @llvm.arc.ADD.ACC40(i64 
> %L.sroa.0.0.insert.insert.i, i64 %2) #1
>
> *_%tmp.sroa.2.0.extract.shift = and i64 %3, -4294967296_*
>
>   %inc = add i32 %i.011, 1
>
> *_%phitmp9 = and i64 %3, 4294967295_*
>
>   %cmp = icmp ult i32 %inc, %limit
>
>   br i1 %cmp, label %for.body, label %for.cond.for.end_crit_edge
>
> The value of the OR is simply the PHI of %3 and 0.  Nobody detects 
> this and it generates pretty ugly code in a loop.  Can you tell me 
> what phase should handle this simple transform?  That is, what phase 
> does this kind of transformation belong in?
>
> Thanks
>
>
>
> _______________________________________________
> LLVM Developers mailing list
> LLVMdev at cs.uiuc.edu         http://llvm.cs.uiuc.edu
> http://lists.cs.uiuc.edu/mailman/listinfo/llvmdev

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20140409/be10081f/attachment.html>