Thanks for the comments. I forgot to "svn add" the test cases. I'll commit it tonight if I have time. Also forgot to make it work for 256-bit FP. I'll add the asserts and try to clean up the loops as well. I think there's an 80 column violation in there too.<br>
<br><div class="gmail_quote">On Wed, Nov 30, 2011 at 1:47 AM, Duncan Sands <span dir="ltr"><<a href="mailto:baldrick@free.fr">baldrick@free.fr</a>></span> wrote:<br><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex;">
Hi Craig,<br>
<div class="im"><br>
> Add instruction selection support for AVX2 horizontal add/sub instructions.<br>
<br>
</div>thanks for doing this. Please add a testcase.<br>
<div class="im"><br>
> --- llvm/trunk/lib/Target/X86/X86ISelLowering.cpp (original)<br>
> +++ llvm/trunk/lib/Target/X86/X86ISelLowering.cpp Wed Nov 30 03:10:50 2011<br>
> @@ -14329,7 +14329,9 @@<br>
> return false;<br>
><br>
> EVT VT = LHS.getValueType();<br>
> - unsigned N = VT.getVectorNumElements();<br>
> + unsigned NumElts = VT.getVectorNumElements();<br>
> + unsigned NumLanes = VT.getSizeInBits()/128;<br>
<br>
</div>Here you assume that the number of bits is always a multiple of 128. While this<br>
is currently true (and probably always will be) how about adding an assertion<br>
checking this?<br>
<div class="im"><br>
> + for (unsigned l = 0; l != NumLanes; ++l) {<br>
> + unsigned LaneStart = l*NumLaneElts;<br>
> + for (unsigned i = 0; i != NumLaneElts/2; ++i) {<br>
<br>
</div>Here you assume that NumLaneElts is even, in particular that it is not 1. Is it<br>
really impossible for it to be equal to 1? If it is impossible, please add an<br>
assertion that checks that. If it is possible, please handle it!<br>
<div class="im"><br>
> + unsigned LIdx = LMask[i+LaneStart];<br>
> + unsigned RIdx = RMask[i+LaneStart];<br>
> +<br>
> + // Ignore any UNDEF components.<br>
> + if (LIdx>= 2*NumElts || RIdx>= 2*NumElts ||<br>
</div>> + (!A.getNode()&& (LIdx< NumElts || RIdx< NumElts)) ||<br>
> + (!B.getNode()&& (LIdx>= NumElts || RIdx>= NumElts)))<br>
<div class="im">> + continue;<br>
> +<br>
> + // Check that successive elements are being operated on. If not, this is<br>
> + // not a horizontal operation.<br>
</div>> + if (!(LIdx == 2*i + LaneStart&& RIdx == 2*i + LaneStart + 1)&&<br>
> + !(isCommutative&& LIdx == 2*i + LaneStart + 1&& RIdx == 2*i + LaneStart))<br>
> + return false;<br>
> + }<br>
<br>
This loop and the one below are essentially doing the same thing. How about<br>
unifying them, but introducing an additional loop that loops over the values<br>
0 and 1:<br>
for (unsigned h = 0; h != 2; ++h)<br>
<div class="im"> for (unsigned i = 0; i != NumLaneElts/2; ++i) {<br>
</div> unsigned LIdx = LMask[i+LaneStart+h*(NumLaneElts/2)];<br>
and so on.<br>
<br>
Ciao, Duncan.<br>
<br>
PS: I didn't check the correctness of your multi-lane logic.<br>
<div class="HOEnZb"><div class="h5">_______________________________________________<br>
llvm-commits mailing list<br>
<a href="mailto:llvm-commits@cs.uiuc.edu">llvm-commits@cs.uiuc.edu</a><br>
<a href="http://lists.cs.uiuc.edu/mailman/listinfo/llvm-commits" target="_blank">http://lists.cs.uiuc.edu/mailman/listinfo/llvm-commits</a><br>
</div></div></blockquote></div><br><br clear="all"><br>-- <br>~Craig<br>