[llvm-dev] Potential bug in SelectionDAGLegalize::ConvertNodeToLibcall()?

[Justin Hibbits via llvm-dev]

Hi Nemanja,

The behavior is the same prior to my patch and with my patch, with
regard to this specific behavior.  The custom BITCAST lowerer doesn't
introduce any change with this.  In fact, it was reported on
https://reviews.llvm.org/D49754 as behavior introduced with the
original SPE additions, so is inherent in this target from the
beginning.  So, it seems to me that the new nodes added in the libcall
lowering should themselves be lowered, unless I should do as Alex
alluded to, and add new legal nodes that themselves handle the argument
splitting and reforming.  But it seems to me that the machinery to do
that already exists, but might need another pass?

- Justin

On Wed, 2 Jan 2019 11:39:59 -0500
Nemanja Ivanovic <nemanja.i.ibm at gmail.com> wrote:

> It sounds like the legalizer is lowering `fmaxnum` to a libcall
> because it is not a legal node for `f64` and in doing so, it is
> producing the `build_pair` to reassemble the results of the libcall.
> And presumably, it is assuming that the new nodes do not need
> legalization or something along those lines.
> 
> Justin, it would probably be good if you could provide the DAG before
> and after legalization both with and without your patch. Then we can
> see how it was handled before your patch and how it is handled now
> and the difference should allude to the problem.
> 
> N
> 
> On Wed, Jan 2, 2019 at 10:54 AM Justin Hibbits via llvm-dev <
> llvm-dev at lists.llvm.org> wrote:
> 
> > Hi,
> >
> > I have a custom lowering operation on ISD::BITCAST for the
> > PowerPC/SPE target, to convert 'f64 bitcast (i64 build_pair i32,
> > i32)' into a 'f64 BUILD_SPE64 i32, i32' node, which can be seen at
> > https://reviews.llvm.org/D54583. However, when building
> > compiler-rt's lib/builtins/divdc3.c an assertion is triggered that
> > BUILD_PAIR is not legal on line 24.  There should be no
> > bitcast(buildpair) anywhere in the generation, as it should all be
> > lowered.  However, this is not the case when lowering to a libcall
> > it seems.  The relevant debug output, is here:
> >
> > Creating new node: t118: i64 = build_pair t117,t116,
> > /home/chmeee/freebsd/contrib/compiler-rt/lib/builtins/divdc3.c:24:22
> > Creating new node: t119: f64 = bitcast t118,
> > /home/chmeee/freebsd/contrib/compiler-rt/lib/builtins/divdc3.c:24:22
> > Created libcall: t119: f64 = bitcast t118,
> > /home/chmeee/freebsd/contrib/compiler-rt/lib/builtins/divdc3.c:24:22
> > Successfully converted node to libcall
> >  ... replacing: t38: f64 = fmaxnum t36, t37,
> > /home/chmeee/freebsd/contrib/compiler-rt/lib/builtins/divdc3.c:24:22
> >      with:      t119: f64 = bitcast t118,
> > /home/chmeee/freebsd/contrib/compiler-rt/lib/builtins/divdc3.c:24:22
> >
> > Is this a real bug, or am I missing something in my patch?  After
> > spending quite a while on it I'm at a loss.
> >
> > Thanks,
> > Justin
> > _______________________________________________
> > LLVM Developers mailing list
> > llvm-dev at lists.llvm.org
> > http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev
> >