Improve ISel using across lane addition reduction

[junbuml]

Summary

In vectorized add reduction code, the final "reduce" step is sub-optimal.
This change wll combine :

ext  v1.16b, v0.16b, v0.16b, #8
add  v0.4s, v1.4s, v0.4s
dup  v1.4s, v0.s[1]
add  v0.4s, v1.4s, v0.4s

into

addv s0, v0.4s

This fixes PR21371.

[mcrosier]

Thanks for working on this, Jun. Mostly style nits with a few other comments.

lib/Target/AArch64/AArch64ISelLowering.cpp

8583

To improve readability, you could expand this into two parts:

// If the extract isn't feeding an ADD, can't do such combine.
if (N0->getOpcode() != ISD::ADD)
  return SDValue();

// Vector extract idx must constant zero. 
if (!isa<ConstantSDNode>(N1) ||  cast<ConstantSDNode>(N1)->getZExtValue())
  return SValue();

8595

The computation of NumVecElts / 2 appears to be loop invariant. Mind hoisting it out of the conditional check for clarity?

8609

Please add a space between %svn and =.

8611

I don't know the correct answer, but is the SV.hasOneUse() check necessary?

8622

Please pre-increment i.

8630

Please use DL rather than dl.

[junbuml]

Thanks Chad for the quick review. I addressed your comments.

[jmolloy]

Hi,

Thanks for working on this! I'm generally happy with the algorithm, but I have a bunch of comment and style requests.

Cheers,

James

lib/Target/AArch64/AArch64ISelLowering.cpp

8568

I'd like to see here an example using SDAG nodes. It's nice to have the assembly before/after, but what's really important is the exact pattern this function is intending to match, and what it replaces it with.

For example:

%2 = SHUFFLE_VECTOR ...
%1 = ADD %2, %3
EXTRACT_ELEMENT %1, 0

Something like that.

And specifically, this is the log2-shuffle pattern that the LoopVectorizer produces. Please make that very explicit, because this code won't match all reductions.

8583

Well actually you can - FMINNUM, FMINNAN, FMAXNUM, FMAXNAN, SMIN, SMAX, UMIN and UMAX all have lane-reduce variants in the A64 instruction set.

I understand you may not want to handle them all at this time, but:

• Please make the comment explicit about what you are and aren't handling.
• Please replace uses of "Add" later on with "binop" or something - so that it isn't perceived that this code is specific to Adds. It's not, it's just only been tested with adds. It should work for any commutative binary op.

8596

Why is this called NumMaxSubAddElts? (Where's the Sub come from?)

8600

It'd be more readable IMHO to calculate the number of expected steps beforehand (as lg(N)), and just have normal counter here. It would make the algorithm more clear.

8603

Please don't fully capitalize local variables - stick to the naming convention of UpperCamelCase (unless using an initialism, so SV is fine).

8604

... That said, it'd be nice to call this "Shuffle", to be very explicit about it.

8605

Comment what you're doing here, which is to allow the add's operands to be commutative.

8627

You've given an example here, which is good, but you haven't defined the *requirements* on the mask indices which the code below is intending to enforce.

Without that information it is difficult to know if the code has a bug in it.

Also, use "unsigned" instead of "unsigned int".

8628

I'm assuming this is saying that all elements greater than NumAddElts must be undef? better to use "Mask[i] != 0" instead of ">= 0" I think, so it is more obvious.

8629

Instead of static_casting here, just use a signed induction variable.

8633

I'm not convinced that "InputADD" is the most descriptive name here, but I don't have an explicit alternative in mind.

[junbuml]

Thanks James for the review. I will address your comments and update it at this morning.

[junbuml]

Addressed James' comments.

[junbuml]

lib/Target/AArch64/AArch64ISelLowering.cpp

8568

. Changed the example using SDAG nodes.
. Specifically mentioned that this function handles the final step of vector reduction.

8583

Add "FIXME" in comment.

8600

Use "NumExpectedStep" and "CurStep" to iterate steps.

8628

To check if an element is an UNDEF, we need to check if the mask is negative value. So, I'm checking Mask[i] < 0.

8633

Now I'm using PreOp and CurOp, instead of InputADD and ADD.

[mssimpso]

Jun/James,

While this patch is focused on the add reductions, aren't there other opportunities here for optimizing the reduce step of the other reduction types? For example, couldn't we similarly use the sminv/uminv and smaxv/umaxv instructions for min/max reductions? We currently don't do this. Min/max reductions are important for 456.hmmer.

[mssimpso]

I now see that the above comment was already made, and a FIXME was added. Please disregard.

[junbuml]

Yes, there are more across lane reduction instructions as mentioned in FIXME. We could extend it to support other types as well.
I may prefer to get this patch in first, and extend it to support other types. James, let me know your thought.

[mcrosier]

In D12325#233611, @junbuml wrote:

I may prefer to get this patch in first, and extend it to support other types. James, let me know your thought.

The LLVM community prefers incremental changes, which simplifies code review as well as eases bisection. It might be wise to land this patch before moving onto the other types of reductions.

[junbuml]

Updated comments little bit more.

[junbuml]

Confirmed that this patch was passed correctness tests for spec2000/2006, and applied several spec benchmarks including gcc, h264, hmmer, sjeng, and vpr.

[junbuml]

Hi James,

I'm planning on pushing another patch to handle other across vector reductions (SMAX/SMIN/UMAX/UMIN), which require some change in this patch. Do you prefer to have a single merged patch or separate patches? Either ways are fine with me. Please let me know your thought.

[mcrosier]

@jmolloy: This looks to be in good shape, AFAICT. Mind giving this a LGTM, if you're satisfied.

[jmolloy]

Hi,

This looks great now, thanks! LGTM with the few nitpicks below.

Cheers,

James

lib/Target/AArch64/AArch64ISelLowering.cpp

8614

Pedantic: Step -> Steps.

8618

"check" -> "check only"

8630

"Check if it forms a one" -> "Check if it forms one"

[junbuml]

Address comments.
Thanks James for the review

[junbuml]

Thanks James for the review !
Chad, would you mind landing this when you have a chance ? Thanks!

[mcrosier]

Committed in r246790.

lib/Target/AArch64/AArch64ISelLowering.cpp

AArch64TargetLowering::AArch64TargetLowering(const TargetMachine &TM,
   setTargetDAGCombine(ISD::MUL);
 
   setTargetDAGCombine(ISD::SELECT);
   setTargetDAGCombine(ISD::VSELECT);
 
   setTargetDAGCombine(ISD::INTRINSIC_VOID);
   setTargetDAGCombine(ISD::INTRINSIC_W_CHAIN);
   setTargetDAGCombine(ISD::INSERT_VECTOR_ELT);
+  setTargetDAGCombine(ISD::EXTRACT_VECTOR_ELT);
 
   MaxStoresPerMemset = MaxStoresPerMemsetOptSize = 8;
   MaxStoresPerMemcpy = MaxStoresPerMemcpyOptSize = 4;
   MaxStoresPerMemmove = MaxStoresPerMemmoveOptSize = 4;
 
   setStackPointerRegisterToSaveRestore(AArch64::SP);
 
   setSchedulingPreference(Sched::Hybrid);
for (SDNode::use_iterator UI = Addr.getNode()->use_begin(), UE =
     DCI.CombineTo(N, SDValue(UpdN.getNode(), 0));     // Dup/Inserted Result
     DCI.CombineTo(User, SDValue(UpdN.getNode(), 1));  // Write back register
 
     break;
   }
   return SDValue();
 }
 
+/// Target-specific DAG combine for the across vector reduction.
+/// This function specifically handles the final clean-up step of a vector
+/// reduction produced by the LoopVectorizer. It is the log2-shuffle pattern,
+/// consisting of log2(NumVectorElements) steps and, in each step, 2^(s)
+/// elements are reduced, where s is an induction variable from 0
+/// to log2(NumVectorElements).
+/// For example,
+///   %1 = vector_shuffle %0, <2,3,u,u>
+///   %2 = add %0, %1
+///   %3 = vector_shuffle %2, <1,u,u,u>
+///   %4 = add %2, %3
+///   %5 = extract_vector_elt %4, 0
+/// becomes :
+///   %0 = uaddv %0
+///   %1 = extract_vector_elt %0, 0
+///
+/// FIXME: Currently this function is implemented and tested specifically
+/// for the add reduction. We could also support other types of across lane
+/// reduction available in AArch64, including SMAXV, SMINV, UMAXV, UMINV,
+/// SADDLV, UADDLV, FMAXNMV, FMAXV, FMINNMV, FMINV.
+static SDValue
+performAcrossLaneReductionCombine(SDNode *N, SelectionDAG &DAG,
+                                  const AArch64Subtarget *Subtarget) {
+  if (!Subtarget->hasNEON())
+    return SDValue();
+  SDValue N0 = N->getOperand(0);
+  SDValue N1 = N->getOperand(1);
+
+  // Check if the input vector is fed by the operator we want to handle.
+  // We specifically check ADD for now.
+  if (N0->getOpcode() != ISD::ADD)
+    return SDValue();
+
+  // The vector extract idx must constant zero because we only expect the final
+  // result of the reduction is placed in lane 0.
+  if (!isa<ConstantSDNode>(N1) || cast<ConstantSDNode>(N1)->getZExtValue())
+    return SDValue();
+
+  EVT EltTy = N0.getValueType().getVectorElementType();
+  if (EltTy != MVT::i32 && EltTy != MVT::i16 && EltTy != MVT::i8)
+    return SDValue();
+
+  int NumVecElts = N0.getValueType().getVectorNumElements();
+  if (NumVecElts != 4 && NumVecElts != 8 && NumVecElts != 16)
+    return SDValue();
+
+  int NumExpectedStep = APInt(8, NumVecElts).logBase2();
+  SDValue PreOp = N0;
+  // Iterate over each step of the across vector reduction.
+  for (int CurStep = 0; CurStep != NumExpectedStep; ++CurStep) {
+    // We specifically check ADD for now.
+    if (PreOp.getOpcode() != ISD::ADD)
+      return SDValue();
+    SDValue CurOp = PreOp.getOperand(0);
+    SDValue Shuffle = PreOp.getOperand(1);
+    if (Shuffle.getOpcode() != ISD::VECTOR_SHUFFLE) {
+      // Try to swap the 1st and 2nd operand as add is commutative.
+      CurOp = PreOp.getOperand(1);
+      Shuffle = PreOp.getOperand(0);
+      if (Shuffle.getOpcode() != ISD::VECTOR_SHUFFLE)
+        return SDValue();
+    }
+    // Check if it forms a one step of the across vector reduction.
+    // E.g.,
+    //   %cur = add %1, %0
+    //   %shuffle = vector_shuffle %cur, <2, 3, u, u>
+    //   %pre = add %cur, %shuffle
+    if (Shuffle.getOperand(0) != CurOp)
+      return SDValue();
+
+    int NumMaskElts = 1 << CurStep;
+    ArrayRef<int> Mask = cast<ShuffleVectorSDNode>(Shuffle)->getMask();
+    // Check mask values in each step.
+    // We expect the shuffle mask in each step follows a specific pattern
+    // denoted here by the <M, U> form, where M is a sequence of integers
+    // starting from NumMaskElts, increasing by 1, and the number integers
+    // in M should be NumMaskElts. U is a sequence of UNDEFs and the number
+    // of undef in U should be NumVecElts - NumMaskElts.
+    // E.g., for <8 x i16>, mask values in each step should be :
+    //   step 0 : <1,u,u,u,u,u,u,u>
+    //   step 1 : <2,3,u,u,u,u,u,u>
+    //   step 2 : <4,5,6,7,u,u,u,u>
+    for (int i = 0; i < NumVecElts; ++i)
+      if ((i < NumMaskElts && Mask[i] != (NumMaskElts + i)) ||
+          (i >= NumMaskElts && !(Mask[i] < 0)))
+        return SDValue();
+
+    PreOp = CurOp;
+  }
+  SDLoc DL(N);
+  return DAG.getNode(
+      ISD::EXTRACT_VECTOR_ELT, DL, N->getValueType(0),
+      DAG.getNode(AArch64ISD::UADDV, DL, PreOp.getSimpleValueType(), PreOp),
+      DAG.getConstant(0, DL, MVT::i64));
+}
+
 /// Target-specific DAG combine function for NEON load/store intrinsics
 /// to merge base address updates.
 static SDValue performNEONPostLDSTCombine(SDNode *N,
                                           TargetLowering::DAGCombinerInfo &DCI,
                                           SelectionDAG &DAG) {
   if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
     return SDValue();
 
SDValue AArch64TargetLowering::PerformDAGCombine(SDNode *N,
   case AArch64ISD::CSEL:
     return performCONDCombine(N, DCI, DAG, 2, 3);
   case AArch64ISD::DUP:
     return performPostLD1Combine(N, DCI, false);
   case AArch64ISD::NVCAST:
     return performNVCASTCombine(N);
   case ISD::INSERT_VECTOR_ELT:
     return performPostLD1Combine(N, DCI, true);
+  case ISD::EXTRACT_VECTOR_ELT:
+    return performAcrossLaneReductionCombine(N, DAG, Subtarget);
   case ISD::INTRINSIC_VOID:
   case ISD::INTRINSIC_W_CHAIN:
     switch (cast<ConstantSDNode>(N->getOperand(1))->getZExtValue()) {
     case Intrinsic::aarch64_neon_ld2:
     case Intrinsic::aarch64_neon_ld3:
     case Intrinsic::aarch64_neon_ld4:
     case Intrinsic::aarch64_neon_ld1x2:
     case Intrinsic::aarch64_neon_ld1x3:

test/CodeGen/AArch64/aarch64-addv.ll

+; RUN: llc -march=aarch64 < %s | FileCheck %s
+
+define i8 @f_v16i8(<16 x i8>* %arr)  {
+; CHECK-LABEL: f_v16i8
+; CHECK: addv {{b[0-9]+}}, {{v[0-9]+}}.16b
+  %bin.rdx = load <16 x i8>, <16 x i8>* %arr
+  %rdx.shuf0 = shufflevector <16 x i8> %bin.rdx, <16 x i8> undef, <16 x i32> <i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 undef, i32 undef,i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>
+  %bin.rdx0 = add <16 x i8> %bin.rdx, %rdx.shuf0
+  %rdx.shuf = shufflevector <16 x i8> %bin.rdx0, <16 x i8> undef, <16 x i32> <i32 4, i32 5, i32 6, i32 7, i32 undef, i32 undef,i32 undef, i32 undef, i32 undef, i32 undef,i32 undef, i32 undef, i32 undef, i32 undef,i32 undef, i32 undef >
+  %bin.rdx11 = add <16 x i8> %bin.rdx0, %rdx.shuf
+  %rdx.shuf12 = shufflevector <16 x i8> %bin.rdx11, <16 x i8> undef, <16 x i32> <i32 2, i32 3, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef,i32 undef, i32 undef, i32 undef, i32 undef,i32 undef, i32 undef>
+  %bin.rdx13 = add <16 x i8> %bin.rdx11, %rdx.shuf12
+  %rdx.shuf13 = shufflevector <16 x i8> %bin.rdx13, <16 x i8> undef, <16 x i32> <i32 1, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef,i32 undef, i32 undef, i32 undef, i32 undef,i32 undef, i32 undef>
+  %bin.rdx14 = add <16 x i8> %bin.rdx13, %rdx.shuf13
+  %r = extractelement <16 x i8> %bin.rdx14, i32 0
+  ret i8 %r
+}
+
+define i16 @f_v8i16(<8 x i16>* %arr)  {
+; CHECK-LABEL: f_v8i16
+; CHECK: addv {{h[0-9]+}}, {{v[0-9]+}}.8h
+  %bin.rdx = load <8 x i16>, <8 x i16>* %arr
+  %rdx.shuf = shufflevector <8 x i16> %bin.rdx, <8 x i16> undef, <8 x i32> <i32 4, i32 5, i32 6, i32 7, i32 undef, i32 undef,i32 undef, i32 undef>
+  %bin.rdx11 = add <8 x i16> %bin.rdx, %rdx.shuf
+  %rdx.shuf12 = shufflevector <8 x i16> %bin.rdx11, <8 x i16> undef, <8 x i32> <i32 2, i32 3, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>
+  %bin.rdx13 = add <8 x i16> %bin.rdx11, %rdx.shuf12
+  %rdx.shuf13 = shufflevector <8 x i16> %bin.rdx13, <8 x i16> undef, <8 x i32> <i32 1, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>
+  %bin.rdx14 = add <8 x i16> %bin.rdx13, %rdx.shuf13
+  %r = extractelement <8 x i16> %bin.rdx14, i32 0
+  ret i16 %r
+}
+
+define i32 @f_v4i32( <4 x i32>* %arr)  {
+; CHECK-LABEL: f_v4i32
+; CHECK: addv {{s[0-9]+}}, {{v[0-9]+}}.4s
+  %bin.rdx = load <4 x i32>, <4 x i32>* %arr
+  %rdx.shuf = shufflevector <4 x i32> %bin.rdx, <4 x i32> undef, <4 x i32> <i32 2, i32 3, i32 undef, i32 undef>
+  %bin.rdx11 = add <4 x i32> %bin.rdx, %rdx.shuf
+  %rdx.shuf12 = shufflevector <4 x i32> %bin.rdx11, <4 x i32> undef, <4 x i32> <i32 1, i32 undef, i32 undef, i32 undef>
+  %bin.rdx13 = add <4 x i32> %bin.rdx11, %rdx.shuf12
+  %r = extractelement <4 x i32> %bin.rdx13, i32 0
+  ret i32 %r
+}
+
+define i64 @f_v2i64(<2 x i64>* %arr)  {
+; CHECK-LABEL: f_v2i64
+; CHECK-NOT: addv
+  %bin.rdx = load <2 x i64>, <2 x i64>* %arr
+  %rdx.shuf0 = shufflevector <2 x i64> %bin.rdx, <2 x i64> undef, <2 x i32> <i32 1, i32 undef>
+  %bin.rdx0 = add <2 x i64> %bin.rdx, %rdx.shuf0
+  %r = extractelement <2 x i64> %bin.rdx0, i32 0
+  ret i64 %r
+}