[llvm-commits] [PATCH] ARM: Custom lower v2f32 = fp_round to fix crash with NEON

[Eli Friedman]

On Wed, Nov 14, 2012 at 8:22 PM, Pete Couperus <pjcoup at gmail.com> wrote:
> Hi Eli,
>
>
> On Wed, Nov 14, 2012 at 5:58 PM, Eli Friedman <eli.friedman at gmail.com>
> wrote:
>>
>> On Tue, Nov 13, 2012 at 11:24 PM, Pete Couperus <pjcoup at gmail.com> wrote:
>> > Hello,
>> >
>> > Lowering "fptrunc <2 x double> %a to <2 x float>" is broken on ARM with
>> > NEON.
>> > This patch custom lowers this conversion using two single element
>> > vcvt's.
>> > This fixes the following PRs.
>> >
>> > http://llvm.org/bugs/show_bug.cgi?id=12540
>> > http://llvm.org/bugs/show_bug.cgi?id=13964
>> >
>> > Please review!
>> > Thanks.
>>
>> Does "setOperationAction(ISD::FP_ROUND, MVT::v2f32, Expand)" not do
>> the right thing here?
>>
>> -Eli
>
>
> No, setting FP_ROUND to expand for v2f32 gets an assert in a different
> place:
>
> llc: /home/peter/llvm/llvm-svn/lib/CodeGen/SelectionDAG/LegalizeDAG.cpp:840:
> void <anonymous
> namespace>::SelectionDAGLegalize::LegalizeStoreOps(llvm::SDNode *):
> Assertion `!StVT.isVector() && "Vector Stores are handled in
> LegalizeVectorOps"' failed.

That assertion means that VectorLegalizer::LegalizeOp is missing a
case for FP_ROUND.

-Eli