[LLVMdev] Vector promotion broken for <2 x [i8|i16]>

Mon Jul 30 12:55:58 PDT 2012

I'm not so sure, the code never checks if v4i8 is legal and it shouldn't as that is the type we want to convert away from. The problem here is v4i8 is legal, but v4i8 on an ISD::AND node is not and I want v4i8 on ISD::AND to be promoted to anything larger. Currently the code doesn't seem to do what it should based on our email chain.

So to summarize, what the code should do is find the smallest vector type that is larger than the current type and perform the operation on that type with the proper extension/truncation afterwards. What the code currently does is find the next value type that is legal and performs a bitcast to that type, performs the operand on the type and then bitcasts back. This doesn't work if the two types are of different sizes. Since this occurs after legalize type, my approach about adding v1xi[16|32] won't work since those types are illegal.

What would you think is the best way to fix this?

Micah

> -----Original Message-----
> From: Rotem, Nadav [mailto:nadav.rotem at intel.com]
> Sent: Monday, July 30, 2012 12:14 PM
> To: Villmow, Micah; Developers Mailing List
> Subject: RE: Vector promotion broken for <2 x [i8|i16]>
> 
> If v4i8 is a legal type then getTypeToPromoteTo should return the pair
> v4i8 and 'legal'. This looks like the root of the problem.
> 
> -----Original Message-----
> From: Villmow, Micah [mailto:Micah.Villmow at amd.com]
> Sent: Monday, July 30, 2012 22:10
> To: Rotem, Nadav; Developers Mailing List
> Subject: RE: Vector promotion broken for <2 x [i8|i16]>
> 
> v4i8 itself is a legal type, just not on the 'AND' operation.
> 
> So there seems to be multiple problems here.
> 1) PromoteVectorOp doesn't handle the case where the types are not the
> same size, this occurs because #2
> 2) getTypeToPromoteTo doesn't actual check to see if the type it should
> promote to makes any sense.
> 3) PromoteVectorOp also doesn't handle the case where
> getTypeToPromoteTo returns an invalid type.
> 
> Micah
> 
> > -----Original Message-----
> > From: Rotem, Nadav [mailto:nadav.rotem at intel.com]
> > Sent: Monday, July 30, 2012 12:06 PM
> > To: Villmow, Micah; Developers Mailing List
> > Subject: RE: Vector promotion broken for <2 x [i8|i16]>
> >
> >
> > Notice that PromoteVectorOp is called after the type legalization
> > legalized all of the types in the program. It legalizes the
> > *operations*, not the types.  So, you should only see legal types
> > (Legal types are types that fit into your registers).  So, if your
> > target has v2i32, I suspect that v4i8 is an illegal because it has a
> > different size.
> >
> >
> > -----Original Message-----
> > From: Villmow, Micah [mailto:Micah.Villmow at amd.com]
> > Sent: Monday, July 30, 2012 21:26
> > To: Rotem, Nadav; Developers Mailing List
> > Subject: RE: Vector promotion broken for <2 x [i8|i16]>
> >
> > Hrmm.... PromoteVectorOp doesn't seem to follow this at all.
> > http://llvm.org/svn/llvm-
> > project/llvm/trunk/lib/CodeGen/SelectionDAG/LegalizeVectorOps.cpp
> > SDValue VectorLegalizer::PromoteVectorOp(SDValue Op) {
> >   // Vector "promotion" is basically just bitcasting and doing the
> > operation
> >   // in a different type.  For example, x86 promotes ISD::AND on
> v2i32
> > to
> >   // v1i64.
> >   EVT VT = Op.getValueType();
> >   assert(Op.getNode()->getNumValues() == 1 &&
> >          "Can't promote a vector with multiple results!");
> >   EVT NVT = TLI.getTypeToPromoteTo(Op.getOpcode(), VT);
> >   DebugLoc dl = Op.getDebugLoc();
> >   SmallVector<SDValue, 4> Operands(Op.getNumOperands());
> >
> >   for (unsigned j = 0; j != Op.getNumOperands(); ++j) {
> >     if (Op.getOperand(j).getValueType().isVector())
> >       Operands[j] = DAG.getNode(ISD::BITCAST, dl, NVT,
> > Op.getOperand(j));
> >     else
> >       Operands[j] = Op.getOperand(j);
> >   }
> >
> >   Op = DAG.getNode(Op.getOpcode(), dl, NVT, &Operands[0],
> > Operands.size());
> >
> >   return DAG.getNode(ISD::BITCAST, dl, VT, Op); }
> >
> > The input Op is <4 x i8> = and <4 x i8>, <4 x i8>.
> > The result of TLI.getTypeToPromoteTo(ISD::AND, MVT::v4i8) is
> > MVT::v2i32;
> >
> > The reason why this occurs is:
> > // See if this has an explicit type specified.
> >     std::map<std::pair<unsigned, MVT::SimpleValueType>,
> >              MVT::SimpleValueType>::const_iterator PTTI =
> >       PromoteToType.find(std::make_pair(Op,
> > VT.getSimpleVT().SimpleTy));
> >     if (PTTI != PromoteToType.end()) return PTTI->second;
> >
> >     assert((VT.isInteger() || VT.isFloatingPoint()) &&
> >            "Cannot autopromote this type, add it with
> > AddPromotedToType.");
> >
> >     EVT NVT = VT;
> >     do {
> >       NVT = (MVT::SimpleValueType)(NVT.getSimpleVT().SimpleTy+1);
> >       assert(NVT.isInteger() == VT.isInteger() && NVT != MVT::isVoid
> &&
> >              "Didn't find type to promote to!");
> >     } while (!isTypeLegal(NVT) ||
> >               getOperationAction(Op, NVT) == Promote);
> >
> > The first line in the do while loop is important, it just increments
> > the type, starting at MVT::v4i8 until it hits a legal type.
> >
> > This seems broken to me.
> > Here is what TOT LLVM has for its MVT list:
> > v4i8           =  14,   //  4 x i8
> >       v8i8           =  15,   //  8 x i8
> >       v16i8          =  16,   // 16 x i8
> >       v32i8          =  17,   // 32 x i8
> >       v2i16          =  18,   //  2 x i16
> >       v4i16          =  19,   //  4 x i16
> >       v8i16          =  20,   //  8 x i16
> >       v16i16         =  21,   // 16 x i16
> >       v2i32          =  22,   //  2 x i32
> >
> > So, for my platform with the 'and' I promote all i8 and i16 types, so
> > the first type that is legal is v2i32.
> >
> > If I add the v1i32 then it works, however, it breaks when I added
> > v1i16(which I need for the v2i8 case).
> >
> > So I set AddPromotedType(ISD::AND, MVT::v4i8, MVT::v1i32) to get
> > around it. So it seems in this case someone has hit this issue before
> > and added the ability to override promotion rules.
> >
> > > -----Original Message-----
> > > From: Rotem, Nadav [mailto:nadav.rotem at intel.com]
> > > Sent: Monday, July 30, 2012 11:17 AM
> > > To: Villmow, Micah; Developers Mailing List
> > > Subject: RE: Vector promotion broken for <2 x [i8|i16]>
> > >
> > > I don't know how your target architecture looks like, but I suspect
> > > that <4 x i8> should not be legalized to <1 x i32>.  I think that
> > what
> > > you are seeing is that <4 x i8> is first split into <2 x i8>, and
> > > later promoted to <2 x i32>.  At the moment different targets can
> > only
> > > affect type-legalization by declaring different legal types. A
> > > number of us discussed the possibility of allowing different
> targets
> > > to override the decisions for different types. But at the moment
> > > this is
> > only a plan.
> > >
> > > -----Original Message-----
> > > From: Villmow, Micah [mailto:Micah.Villmow at amd.com]
> > > Sent: Monday, July 30, 2012 21:05
> > > To: Rotem, Nadav; Developers Mailing List
> > > Subject: RE: Vector promotion broken for <2 x [i8|i16]>
> > >
> > > Sorry, <4 x i8> should convert to a <1 x i32>. What currently is
> > > happening is that it is returning a <2 x i32> because <1 x i32>
> does
> > > not exist.
> > >
> > > Micah
> > >
> > > > -----Original Message-----
> > > > From: Rotem, Nadav [mailto:nadav.rotem at intel.com]
> > > > Sent: Monday, July 30, 2012 10:51 AM
> > > > To: Villmow, Micah; Developers Mailing List
> > > > Subject: RE: Vector promotion broken for <2 x [i8|i16]>
> > > >
> > > >
> > > > >Though I am having trouble in understanding how  x86 supports
> the
> > > > ><1 x
> > > > i64> type. Based on looking at the code, it should fail because
> > > > i64> v1i64
> > > > is not supported on the x86 platform as >far as I can tell.
> > > >
> > > > The Type-Legalizer can handle vector types in the following ways:
> > > > 1. Split - this splits vectors into two halves.  For example on
> > > > SSE4,
> > > > <4 x i64>  is split to <2 x i64>
> > > > 2. Widen - this methods adds additional vector elements, but
> keeps
> > > the
> > > > element type. For example <3 x float> is legalized to <4 x float>
> > 3.
> > > > Promote - this method widens each element in the vector. For
> > example
> > > > SSE masks are promoted from <4 x i1> to <4 x i32> 4. Scalarize -
> > > > this method coverts vectors with a single element into a scalar.
> > For
> > > > example, <1 x i64> into i64.
> > > >
> > > > >> The function getTypeToPromoteTo will return a <2 x i32> for a
> > > > >> <2 x
> > > > i8>, <2 x i16> or <4 x i8>.
> > > >
> > > > The first two conversions look correct, and I assume that your
> > > > target declares v2i32 as a legal type. I am not sure how <4 x i8>
> > got there.
> > > > Maybe it was first split, and after that promoted ?
> > > >
> > > >
> > > >
> > > > From: Rotem, Nadav [mailto:nadav.rotem at intel.com]
> > > > Sent: Saturday, July 28, 2012 12:56 PM
> > > > To: Villmow, Micah; Developers Mailing List
> > > > Subject: RE: Vector promotion broken for <2 x [i8|i16]>
> > > >
> > > > I think that you attached the wrong patch. The attached patch is
> > the
> > > > one which adds the new MVT types.
> > > >
> > > >
> > > >
> > > > From: llvmdev-bounces at cs.uiuc.edu [mailto:llvmdev-
> > > bounces at cs.uiuc.edu]
> > > > On Behalf Of Villmow, Micah
> > > > Sent: Saturday, July 28, 2012 01:54
> > > > To: Developers Mailing List
> > > > Subject: [LLVMdev] Vector promotion broken for <2 x [i8|i16]>
> > > >
> > > > Vector promotion which is new in LLVM 3.1 is broken for sub32 bit
> > > > types. The problem is in the VectorLegalizer::PromoteVectorOp.
> > > > The function getTypeToPromoteTo will return a <2 x i32> for a <2
> x
> > > > i8>,
> > > > <2 x i16> or <4 x i8>. The problem is that there are no vectors
> of
> > > > size
> > > > 1 defined for i32 or i16. The attached patch fixes these issues.
> > > >
> > > > This can be reproduced by setting in any target:
> > > > setOperationAction(ISD::AND, MVT::i8, Promote);
> > > > setOperationAction(ISD::AND, MVT::v2i8, Promote);
> > > > setOperationAction(ISD::AND, MVT::i16, Promote);
> > > >
> > > > Let me know if this is good,
> > > > Micah
> > > >
> > > > -----------------------------------------------------------------
> -
> > > > -
> > -
> > > > -
> > > > Intel Israel (74) Limited
> > > >
> > > > This e-mail and any attachments may contain confidential material
> > > > for the sole use of the intended recipient(s). Any review or
> > > > distribution by others is strictly prohibited. If you are not the
> > > > intended recipient, please contact the sender and delete all
> > copies.
> > > > -----------------------------------------------------------------
> -
> > > > -
> > -
> > > > -
> > > > Intel Israel (74) Limited
> > > >
> > > > This e-mail and any attachments may contain confidential material
> > > > for the sole use of the intended recipient(s). Any review or
> > > > distribution by others is strictly prohibited. If you are not the
> > > > intended recipient, please contact the sender and delete all
> > copies.
> > >
> > > -------------------------------------------------------------------
> -
> > > -
> > > Intel Israel (74) Limited
> > >
> > > This e-mail and any attachments may contain confidential material
> > > for the sole use of the intended recipient(s). Any review or
> > > distribution by others is strictly prohibited. If you are not the
> > > intended recipient, please contact the sender and delete all
> copies.
> >
> > ---------------------------------------------------------------------
> > Intel Israel (74) Limited
> >
> > This e-mail and any attachments may contain confidential material for
> > the sole use of the intended recipient(s). Any review or distribution
> > by others is strictly prohibited. If you are not the intended
> > recipient, please contact the sender and delete all copies.
> 
> ---------------------------------------------------------------------
> Intel Israel (74) Limited
> 
> This e-mail and any attachments may contain confidential material for
> the sole use of the intended recipient(s). Any review or distribution
> by others is strictly prohibited. If you are not the intended
> recipient, please contact the sender and delete all copies.