[llvm-commits] Patch to safely restore optimization to testcase from PR3225 on x86

Duncan Sands baldrick at free.fr
Fri Dec 19 09:31:23 PST 2008


On Friday 19 December 2008 18:07:42 Dan Gohman wrote:
> 
> On Dec 18, 2008, at 9:49 PM, Duncan Sands wrote:
> 
> > On Friday 19 December 2008 02:31:13 Dan Gohman wrote:
> >>> +  if (KnownZero == VTBits) {
> >>
> >> If the shift count is known to be all-zero, shouldn't it
> >> have been simplified already by pre-legalize DAGCombine,
> >> if not InstCombine?
> >
> > What if it was introduced by LegalizeTypes itself?
> 
> It turns out I had misread the code.  But even if it had been this
> way, we now run DAGCombine between LegalizeTypes and
> LegalizeOps(?) so it should still catch this :-).

IIRC the above (KnownZero == VTBits) is in LegalizeTypes, so might
fire on nodes created by LegalizeTypes itself.  In any case, since
you seem happy now I am happy too :)

Ciao,

Duncan.



More information about the llvm-commits mailing list