[x86] eliminate redundant shuffle of horizontal math ops when both inputs are the same

[spatel]

Summary

This is limited to a set of patterns based on the example in PR34111:
https://bugs.llvm.org/show_bug.cgi?id=34111
...but as I was investigating this, I see that horizontal patterns can go wrong in many, many other ways that would not be handled by this patch. Each data type may even go different in the DAG after starting with the same basic IR pattern, so even proper IR canonicalization won't fix it all.

[craig.topper]

lib/Target/X86/X86ISelLowering.cpp

29078

Take ArrayRef by value, its a pointer and size.

29079

Maybe add a TODO that this is incomplete? {1, 1}, {1, 0} would also be redundant right?

[spatel]

Patch updated:

1. Take ArrayRef by value rather than ref.
2. Add TODO about checking more shuffle masks (h-op matching doesn't currently work for the listed cases, so it probably doesn't make sense to add those unless/until we fix that).

[craig.topper]

LGTM

lib/Target/X86/X86ISelLowering.cpp

for (int Elt : SVOp->getMask())
     Mask.push_back(Elt < NumElts ? Elt : (Elt - NumElts / 2));
 
   SDLoc DL(N);
   SDValue Concat = DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, N0.getOperand(0),
                                N1.getOperand(0));
   return DAG.getVectorShuffle(VT, DL, Concat, DAG.getUNDEF(VT), Mask);
 }
 
+/// Eliminate a redundant shuffle of a horizontal math op.
+static SDValue foldShuffleOfHorizOp(SDNode *N) {
+  if (N->getOpcode() != ISD::VECTOR_SHUFFLE || !N->getOperand(1).isUndef())
+    return SDValue();
+
+  SDValue HOp = N->getOperand(0);
+  if (HOp.getOpcode() != X86ISD::HADD && HOp.getOpcode() != X86ISD::FHADD &&
+      HOp.getOpcode() != X86ISD::HSUB && HOp.getOpcode() != X86ISD::FHSUB)
+    return SDValue();
+
+  // 128-bit horizontal math instructions are defined to operate on adjacent
+  // lanes of each operand as:
+  // v4X32: A[0] + A[1] , A[2] + A[3] , B[0] + B[1] , B[2] + B[3]
+  // ...similarly for v2f64 and v8i16.
+  // TODO: 256-bit is not the same because...x86.
+  if (HOp.getOperand(0) != HOp.getOperand(1) || HOp.getValueSizeInBits() != 128)
+    return SDValue();
+
+  // When the operands of a horizontal math op are identical, the low half of
+  // the result is the same as the high half. If the shuffle is also replicating
+  // low and high halves, we don't need the shuffle.
+  // shuffle (hadd X, X), undef, [low half...high half] --> hadd X, X
+  const ArrayRef<int> &Mask = cast<ShuffleVectorSDNode>(N)->getMask();
+  if (isTargetShuffleEquivalent(Mask, { 0, 0 }) ||
+      isTargetShuffleEquivalent(Mask, { 0, 1, 0, 1 }) ||
+      isTargetShuffleEquivalent(Mask, { 0, 1, 2, 3, 0, 1, 2, 3 }))
+    return HOp;
+
+  return SDValue();
+}
+
 static SDValue combineShuffle(SDNode *N, SelectionDAG &DAG,
                               TargetLowering::DAGCombinerInfo &DCI,
                               const X86Subtarget &Subtarget) {
   SDLoc dl(N);
   EVT VT = N->getValueType(0);
   const TargetLowering &TLI = DAG.getTargetLoweringInfo();
   // If we have legalized the vector types, look for blends of FADD and FSUB
   // nodes that we can fuse into an ADDSUB node.
-  if (TLI.isTypeLegal(VT))
+  if (TLI.isTypeLegal(VT)) {
     if (SDValue AddSub = combineShuffleToAddSubOrFMAddSub(N, Subtarget, DAG))
       return AddSub;
 
+    if (SDValue HAddSub = foldShuffleOfHorizOp(N))
+      return HAddSub;
+  }
+
   // During Type Legalization, when promoting illegal vector types,
   // the backend might introduce new shuffle dag nodes and bitcasts.
   //
   // This code performs the following transformation:
   // fold: (shuffle (bitcast (BINOP A, B)), Undef, <Mask>) ->
   //       (shuffle (BINOP (bitcast A), (bitcast B)), Undef, <Mask>)
   //
   // We do this only if both the bitcast and the BINOP dag nodes have

test/CodeGen/X86/haddsub-shuf.ll

 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
 ; RUN: llc < %s -mtriple=x86_64-unknown -mattr=+ssse3 | FileCheck %s --check-prefix=SSSE3
 ; RUN: llc < %s -mtriple=x86_64-unknown -mattr=+avx | FileCheck %s --check-prefix=AVX
 
 ; The next 8 tests check for matching the horizontal op and eliminating the shuffle.
 ; PR34111 - https://bugs.llvm.org/show_bug.cgi?id=34111
 
 define <4 x float> @hadd_v4f32(<4 x float> %a) {
 ; SSSE3-LABEL: hadd_v4f32:
 ; SSSE3:       # BB#0:
 ; SSSE3-NEXT:    haddps %xmm0, %xmm0
-; SSSE3-NEXT:    movddup {{.*#+}} xmm0 = xmm0[0,0]
 ; SSSE3-NEXT:    retq
 ;
 ; AVX-LABEL: hadd_v4f32:
 ; AVX:       # BB#0:
 ; AVX-NEXT:    vhaddps %xmm0, %xmm0, %xmm0
-; AVX-NEXT:    vmovddup {{.*#+}} xmm0 = xmm0[0,0]
 ; AVX-NEXT:    retq
   %a02 = shufflevector <4 x float> %a, <4 x float> undef, <2 x i32> <i32 0, i32 2>
   %a13 = shufflevector <4 x float> %a, <4 x float> undef, <2 x i32> <i32 1, i32 3>
   %hop = fadd <2 x float> %a02, %a13
   %shuf = shufflevector <2 x float> %hop, <2 x float> undef, <4 x i32> <i32 undef, i32 undef, i32 0, i32 1>
   ret <4 x float> %shuf
 }
 
 define <4 x float> @hsub_v4f32(<4 x float> %a) {
 ; SSSE3-LABEL: hsub_v4f32:
 ; SSSE3:       # BB#0:
 ; SSSE3-NEXT:    hsubps %xmm0, %xmm0
-; SSSE3-NEXT:    movddup {{.*#+}} xmm0 = xmm0[0,0]
 ; SSSE3-NEXT:    retq
 ;
 ; AVX-LABEL: hsub_v4f32:
 ; AVX:       # BB#0:
 ; AVX-NEXT:    vhsubps %xmm0, %xmm0, %xmm0
-; AVX-NEXT:    vmovddup {{.*#+}} xmm0 = xmm0[0,0]
 ; AVX-NEXT:    retq
   %a02 = shufflevector <4 x float> %a, <4 x float> undef, <2 x i32> <i32 0, i32 2>
   %a13 = shufflevector <4 x float> %a, <4 x float> undef, <2 x i32> <i32 1, i32 3>
   %hop = fsub <2 x float> %a02, %a13
   %shuf = shufflevector <2 x float> %hop, <2 x float> undef, <4 x i32> <i32 0, i32 1, i32 0, i32 1>
   ret <4 x float> %shuf
 }
 
 define <2 x double> @hadd_v2f64(<2 x double> %a) {
 ; SSSE3-LABEL: hadd_v2f64:
 ; SSSE3:       # BB#0:
 ; SSSE3-NEXT:    haddpd %xmm0, %xmm0
-; SSSE3-NEXT:    movddup {{.*#+}} xmm0 = xmm0[0,0]
 ; SSSE3-NEXT:    retq
 ;
 ; AVX-LABEL: hadd_v2f64:
 ; AVX:       # BB#0:
 ; AVX-NEXT:    vhaddpd %xmm0, %xmm0, %xmm0
-; AVX-NEXT:    vmovddup {{.*#+}} xmm0 = xmm0[0,0]
 ; AVX-NEXT:    retq
   %a0 = shufflevector <2 x double> %a, <2 x double> undef, <2 x i32> <i32 0, i32 undef>
   %a1 = shufflevector <2 x double> %a, <2 x double> undef, <2 x i32> <i32 1, i32 undef>
   %hop = fadd <2 x double> %a0, %a1
   %shuf = shufflevector <2 x double> %hop, <2 x double> undef, <2 x i32> <i32 0, i32 0>
   ret <2 x double> %shuf
 }
 
 define <2 x double> @hsub_v2f64(<2 x double> %a) {
 ; SSSE3-LABEL: hsub_v2f64:
 ; SSSE3:       # BB#0:
 ; SSSE3-NEXT:    hsubpd %xmm0, %xmm0
-; SSSE3-NEXT:    movddup {{.*#+}} xmm0 = xmm0[0,0]
 ; SSSE3-NEXT:    retq
 ;
 ; AVX-LABEL: hsub_v2f64:
 ; AVX:       # BB#0:
 ; AVX-NEXT:    vhsubpd %xmm0, %xmm0, %xmm0
-; AVX-NEXT:    vmovddup {{.*#+}} xmm0 = xmm0[0,0]
 ; AVX-NEXT:    retq
   %a0 = shufflevector <2 x double> %a, <2 x double> undef, <2 x i32> <i32 0, i32 undef>
   %a1 = shufflevector <2 x double> %a, <2 x double> undef, <2 x i32> <i32 1, i32 undef>
   %hop = fsub <2 x double> %a0, %a1
   %shuf = shufflevector <2 x double> %hop, <2 x double> undef, <2 x i32> <i32 undef, i32 0>
   ret <2 x double> %shuf
 }
 
 define <4 x i32> @hadd_v4i32(<4 x i32> %a) {
 ; SSSE3-LABEL: hadd_v4i32:
 ; SSSE3:       # BB#0:
 ; SSSE3-NEXT:    phaddd %xmm0, %xmm0
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,1,0,1]
 ; SSSE3-NEXT:    retq
 ;
 ; AVX-LABEL: hadd_v4i32:
 ; AVX:       # BB#0:
 ; AVX-NEXT:    vphaddd %xmm0, %xmm0, %xmm0
-; AVX-NEXT:    vpshufd {{.*#+}} xmm0 = xmm0[0,1,0,1]
 ; AVX-NEXT:    retq
   %a02 = shufflevector <4 x i32> %a, <4 x i32> undef, <4 x i32> <i32 0, i32 2, i32 undef, i32 undef>
   %a13 = shufflevector <4 x i32> %a, <4 x i32> undef, <4 x i32> <i32 1, i32 3, i32 undef, i32 undef>
   %hop = add <4 x i32> %a02, %a13
   %shuf = shufflevector <4 x i32> %hop, <4 x i32> undef, <4 x i32> <i32 0, i32 undef, i32 undef, i32 1>
   ret <4 x i32> %shuf
 }
 
 define <4 x i32> @hsub_v4i32(<4 x i32> %a) {
 ; SSSE3-LABEL: hsub_v4i32:
 ; SSSE3:       # BB#0:
 ; SSSE3-NEXT:    phsubd %xmm0, %xmm0
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,1,0,1]
 ; SSSE3-NEXT:    retq
 ;
 ; AVX-LABEL: hsub_v4i32:
 ; AVX:       # BB#0:
 ; AVX-NEXT:    vphsubd %xmm0, %xmm0, %xmm0
-; AVX-NEXT:    vpshufd {{.*#+}} xmm0 = xmm0[0,1,0,1]
 ; AVX-NEXT:    retq
   %a02 = shufflevector <4 x i32> %a, <4 x i32> undef, <4 x i32> <i32 0, i32 2, i32 undef, i32 undef>
   %a13 = shufflevector <4 x i32> %a, <4 x i32> undef, <4 x i32> <i32 1, i32 3, i32 undef, i32 undef>
   %hop = sub <4 x i32> %a02, %a13
   %shuf = shufflevector <4 x i32> %hop, <4 x i32> undef, <4 x i32> <i32 undef, i32 1, i32 0, i32 undef>
   ret <4 x i32> %shuf
 }
 
 define <8 x i16> @hadd_v8i16(<8 x i16> %a) {
 ; SSSE3-LABEL: hadd_v8i16:
 ; SSSE3:       # BB#0:
 ; SSSE3-NEXT:    phaddw %xmm0, %xmm0
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,1,0,1]
 ; SSSE3-NEXT:    retq
 ;
 ; AVX-LABEL: hadd_v8i16:
 ; AVX:       # BB#0:
 ; AVX-NEXT:    vphaddw %xmm0, %xmm0, %xmm0
-; AVX-NEXT:    vpshufd {{.*#+}} xmm0 = xmm0[0,1,0,1]
 ; AVX-NEXT:    retq
   %a0246 = shufflevector <8 x i16> %a, <8 x i16> undef, <8 x i32> <i32 0, i32 2, i32 4, i32 6, i32 undef, i32 undef, i32 undef, i32 undef>
   %a1357 = shufflevector <8 x i16> %a, <8 x i16> undef, <8 x i32> <i32 1, i32 3, i32 5, i32 7, i32 undef, i32 undef, i32 undef, i32 undef>
   %hop = add <8 x i16> %a0246, %a1357
   %shuf = shufflevector <8 x i16> %hop, <8 x i16> undef, <8 x i32> <i32 undef, i32 undef, i32 undef, i32 undef, i32 0, i32 1, i32 2, i32 3>
   ret <8 x i16> %shuf
 }
 
 define <8 x i16> @hsub_v8i16(<8 x i16> %a) {
 ; SSSE3-LABEL: hsub_v8i16:
 ; SSSE3:       # BB#0:
 ; SSSE3-NEXT:    phsubw %xmm0, %xmm0
-; SSSE3-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,1,0,1]
 ; SSSE3-NEXT:    retq
 ;
 ; AVX-LABEL: hsub_v8i16:
 ; AVX:       # BB#0:
 ; AVX-NEXT:    vphsubw %xmm0, %xmm0, %xmm0
-; AVX-NEXT:    vpshufd {{.*#+}} xmm0 = xmm0[0,1,0,1]
 ; AVX-NEXT:    retq
   %a0246 = shufflevector <8 x i16> %a, <8 x i16> undef, <8 x i32> <i32 0, i32 2, i32 4, i32 6, i32 undef, i32 undef, i32 undef, i32 undef>
   %a1357 = shufflevector <8 x i16> %a, <8 x i16> undef, <8 x i32> <i32 1, i32 3, i32 5, i32 7, i32 undef, i32 undef, i32 undef, i32 undef>
   %hop = sub <8 x i16> %a0246, %a1357
   %shuf = shufflevector <8 x i16> %hop, <8 x i16> undef, <8 x i32> <i32 0, i32 undef, i32 2, i32 undef, i32 undef, i32 1, i32 undef, i32 3>
   ret <8 x i16> %shuf
 }