[X86] Post process the DAG after isel to remove vector moves that were added to zero upper bits.

[craig.topper]

Summary

We currently avoid inserting these moves during isel in a few cases which is implemented using a whitelist of opcodes. But it's too difficult to generate a perfect list of opcodes to whitelist. Especially with AVX512F without AVX512VL using 512 bit vectors to implement some 128/256 bit operations. Since isel is done bottoms up, we'd have to check the VT and opcode and subtarget in order to determine whether an EXTRACT_SUBREG would be generated for some operations.

So instead of doing that, this patch adds a post processing step that detects when the moves are unnecesssary after isel. At that point any EXTRACT_SUBREGs would have already been created and appear in the DAG. So then we just need to ensure the input to the move isn't one.

[craig.topper]

Ping

[RKSimon]

LGTM

llvm/trunk/lib/Target/X86/X86ISelDAGToDAG.cpp

bool runOnMachineFunction(MachineFunction &MF) override {
       return true;
     }
 
     void EmitFunctionEntryCode() override;
 
     bool IsProfitableToFold(SDValue N, SDNode *U, SDNode *Root) const override;
 
     void PreprocessISelDAG() override;
+    void PostprocessISelDAG() override;
 
 // Include the pieces autogenerated from the target description.
 #include "X86GenDAGISel.inc"
 
   private:
     void Select(SDNode *N) override;
 
     bool foldOffsetIntoAddress(uint64_t Offset, X86ISelAddressMode &AM);
for (SelectionDAG::allnodes_iterator I = CurDAG->allnodes_begin(),
     // Now that we did that, the node is dead.  Increment the iterator to the
     // next node to process, then delete N.
     ++I;
     CurDAG->DeleteNode(N);
   }
 }
 
 
+void X86DAGToDAGISel::PostprocessISelDAG() {
+  // Skip peepholes at -O0.
+  if (TM.getOptLevel() == CodeGenOpt::None)
+    return;
+
+  // Attempt to remove vectors moves that were inserted to zero upper bits.
+
+  SelectionDAG::allnodes_iterator Position(CurDAG->getRoot().getNode());
+  ++Position;
+
+  while (Position != CurDAG->allnodes_begin()) {
+    SDNode *N = &*--Position;
+    // Skip dead nodes and any non-machine opcodes.
+    if (N->use_empty() || !N->isMachineOpcode())
+      continue;
+
+    if (N->getMachineOpcode() != TargetOpcode::SUBREG_TO_REG)
+      continue;
+
+    unsigned SubRegIdx = N->getConstantOperandVal(2);
+    if (SubRegIdx != X86::sub_xmm && SubRegIdx != X86::sub_ymm)
+      continue;
+
+    SDValue Move = N->getOperand(1);
+    if (!Move.isMachineOpcode())
+      continue;
+
+    // Make sure its one of the move opcodes we recognize.
+    switch (Move.getMachineOpcode()) {
+    default:
+      continue;
+    case X86::VMOVAPDrr:       case X86::VMOVUPDrr:
+    case X86::VMOVAPSrr:       case X86::VMOVUPSrr:
+    case X86::VMOVDQArr:       case X86::VMOVDQUrr:
+    case X86::VMOVAPDYrr:      case X86::VMOVUPDYrr:
+    case X86::VMOVAPSYrr:      case X86::VMOVUPSYrr:
+    case X86::VMOVDQAYrr:      case X86::VMOVDQUYrr:
+    case X86::VMOVAPDZ128rr:   case X86::VMOVUPDZ128rr:
+    case X86::VMOVAPSZ128rr:   case X86::VMOVUPSZ128rr:
+    case X86::VMOVDQA32Z128rr: case X86::VMOVDQU32Z128rr:
+    case X86::VMOVDQA64Z128rr: case X86::VMOVDQU64Z128rr:
+    case X86::VMOVAPDZ256rr:   case X86::VMOVUPDZ256rr:
+    case X86::VMOVAPSZ256rr:   case X86::VMOVUPSZ256rr:
+    case X86::VMOVDQA32Z256rr: case X86::VMOVDQU32Z256rr:
+    case X86::VMOVDQA64Z256rr: case X86::VMOVDQU64Z256rr:
+      break;
+    }
+
+    SDValue In = Move.getOperand(0);
+    if (!In.isMachineOpcode() ||
+        In.getMachineOpcode() <= TargetOpcode::GENERIC_OP_END)
+      continue;
+
+    // Producing instruction is another vector instruction. We can drop the
+    // move.
+    CurDAG->UpdateNodeOperands(N, N->getOperand(0), In, N->getOperand(2));
+
+    // If the move is now dead, delete it.
+    if (Move.getNode()->use_empty())
+      CurDAG->RemoveDeadNode(Move.getNode());
+  }
+}
+
+
 /// Emit any code that needs to be executed only in the main function.
 void X86DAGToDAGISel::emitSpecialCodeForMain() {
   if (Subtarget->isTargetCygMing()) {
     TargetLowering::ArgListTy Args;
     auto &DL = CurDAG->getDataLayout();
 
     TargetLowering::CallLoweringInfo CLI(*CurDAG);
     CLI.setChain(CurDAG->getRoot())

llvm/trunk/lib/Target/X86/X86InstrVecCompiler.td

let Predicates = [HasAVX512, NoVLX] in {
   defm : subvector_zero_lowering<"DQAY", "DQUY", VR256, v16i32, v8i32,
                                  v16i32, loadv4i64, sub_ymm>;
   defm : subvector_zero_lowering<"DQAY", "DQUY", VR256, v32i16, v16i16,
                                  v16i32, loadv4i64, sub_ymm>;
   defm : subvector_zero_lowering<"DQAY", "DQUY", VR256, v64i8, v32i8,
                                  v16i32, loadv4i64, sub_ymm>;
 }
 
-// List of opcodes that guaranteed to zero the upper elements of vector regs.
-// TODO: Ideally this would be a blacklist instead of a whitelist. But SHA
-// intrinsics and some MMX->XMM move instructions that aren't VEX encoded make
-// this difficult. So starting with a couple opcodes used by reduction loops
-// where we explicitly insert zeros.
-class veczeroupper<ValueType vt, RegisterClass RC> :
-  PatLeaf<(vt RC:$src), [{
-    return N->getOpcode() == X86ISD::VPMADDWD ||
-           N->getOpcode() == X86ISD::PSADBW;
-  }]>;
-
-def zeroupperv2f64 : veczeroupper<v2f64, VR128>;
-def zeroupperv4f32 : veczeroupper<v4f32, VR128>;
-def zeroupperv2i64 : veczeroupper<v2i64, VR128>;
-def zeroupperv4i32 : veczeroupper<v4i32, VR128>;
-def zeroupperv8i16 : veczeroupper<v8i16, VR128>;
-def zeroupperv16i8 : veczeroupper<v16i8, VR128>;
-
-def zeroupperv4f64  : veczeroupper<v4f64, VR256>;
-def zeroupperv8f32  : veczeroupper<v8f32, VR256>;
-def zeroupperv4i64  : veczeroupper<v4i64, VR256>;
-def zeroupperv8i32  : veczeroupper<v8i32, VR256>;
-def zeroupperv16i16 : veczeroupper<v16i16, VR256>;
-def zeroupperv32i8  : veczeroupper<v32i8, VR256>;
-
-
-// If we can guarantee the upper elements have already been zeroed we can elide
-// an explicit zeroing.
-multiclass subvector_zero_ellision<RegisterClass RC, ValueType DstTy,
-                                   ValueType SrcTy, ValueType ZeroTy,
-                                   SubRegIndex SubIdx, PatLeaf Zeroupper> {
-  def : Pat<(DstTy (insert_subvector (bitconvert (ZeroTy immAllZerosV)),
-                                     Zeroupper:$src, (iPTR 0))),
-            (SUBREG_TO_REG (i64 0), RC:$src, SubIdx)>;
-}
-
-// 128->256
-defm: subvector_zero_ellision<VR128, v4f64,  v2f64, v8i32, sub_xmm, zeroupperv2f64>;
-defm: subvector_zero_ellision<VR128, v8f32,  v4f32, v8i32, sub_xmm, zeroupperv4f32>;
-defm: subvector_zero_ellision<VR128, v4i64,  v2i64, v8i32, sub_xmm, zeroupperv2i64>;
-defm: subvector_zero_ellision<VR128, v8i32,  v4i32, v8i32, sub_xmm, zeroupperv4i32>;
-defm: subvector_zero_ellision<VR128, v16i16, v8i16, v8i32, sub_xmm, zeroupperv8i16>;
-defm: subvector_zero_ellision<VR128, v32i8,  v16i8, v8i32, sub_xmm, zeroupperv16i8>;
-
-// 128->512
-defm: subvector_zero_ellision<VR128, v8f64,  v2f64, v16i32, sub_xmm, zeroupperv2f64>;
-defm: subvector_zero_ellision<VR128, v16f32, v4f32, v16i32, sub_xmm, zeroupperv4f32>;
-defm: subvector_zero_ellision<VR128, v8i64,  v2i64, v16i32, sub_xmm, zeroupperv2i64>;
-defm: subvector_zero_ellision<VR128, v16i32, v4i32, v16i32, sub_xmm, zeroupperv4i32>;
-defm: subvector_zero_ellision<VR128, v32i16, v8i16, v16i32, sub_xmm, zeroupperv8i16>;
-defm: subvector_zero_ellision<VR128, v64i8,  v16i8, v16i32, sub_xmm, zeroupperv16i8>;
-
-// 256->512
-defm: subvector_zero_ellision<VR256, v8f64,  v4f64,  v16i32, sub_ymm, zeroupperv4f64>;
-defm: subvector_zero_ellision<VR256, v16f32, v8f32,  v16i32, sub_ymm, zeroupperv8f32>;
-defm: subvector_zero_ellision<VR256, v8i64,  v4i64,  v16i32, sub_ymm, zeroupperv4i64>;
-defm: subvector_zero_ellision<VR256, v16i32, v8i32,  v16i32, sub_ymm, zeroupperv8i32>;
-defm: subvector_zero_ellision<VR256, v32i16, v16i16, v16i32, sub_ymm, zeroupperv16i16>;
-defm: subvector_zero_ellision<VR256, v64i8,  v32i8,  v16i32, sub_ymm, zeroupperv32i8>;
-
-
 class maskzeroupper<ValueType vt, RegisterClass RC> :
   PatLeaf<(vt RC:$src), [{
     return isMaskZeroExtended(N);
   }]>;
 
 def maskzeroupperv1i1  : maskzeroupper<v1i1,  VK1>;
 def maskzeroupperv2i1  : maskzeroupper<v2i1,  VK2>;
 def maskzeroupperv4i1  : maskzeroupper<v4i1,  VK4>;

llvm/trunk/test/CodeGen/X86/avg.ll

 ; AVX512BW-NEXT:    vpsrld $1, %zmm0, %zmm0
 ; AVX512BW-NEXT:    vpsrld $1, %zmm3, %zmm1
 ; AVX512BW-NEXT:    vpsrld $1, %zmm2, %zmm2
 ; AVX512BW-NEXT:    vpmovdw %zmm2, %ymm2
 ; AVX512BW-NEXT:    vpmovdw %zmm1, %ymm1
 ; AVX512BW-NEXT:    vinserti64x4 $1, %ymm1, %zmm2, %zmm1
 ; AVX512BW-NEXT:    vpmovwb %zmm1, %ymm1
 ; AVX512BW-NEXT:    vpmovdw %zmm0, %ymm0
-; AVX512BW-NEXT:    vmovdqa %ymm0, %ymm0
 ; AVX512BW-NEXT:    vpmovwb %zmm0, %ymm0
 ; AVX512BW-NEXT:    vinserti64x4 $1, %ymm0, %zmm1, %zmm0
 ; AVX512BW-NEXT:    vmovdqu %ymm1, (%rax)
 ; AVX512BW-NEXT:    vextracti32x4 $2, %zmm0, (%rax)
 ; AVX512BW-NEXT:    vzeroupper
 ; AVX512BW-NEXT:    retq
   %1 = load <48 x i8>, <48 x i8>* %a
   %2 = load <48 x i8>, <48 x i8>* %b

llvm/trunk/test/CodeGen/X86/bitcast-setcc-128.ll

 ; AVX1:       # %bb.0: # %entry
 ; AVX1-NEXT:    vpcmpeqb %xmm1, %xmm0, %xmm0
 ; AVX1-NEXT:    vpmovmskb %xmm0, %eax
 ; AVX1-NEXT:    retq
 ;
 ; AVX2-LABEL: v16i8_widened_with_zeroes:
 ; AVX2:       # %bb.0: # %entry
 ; AVX2-NEXT:    vpcmpeqb %xmm1, %xmm0, %xmm0
-; AVX2-NEXT:    vmovdqa %xmm0, %xmm0
 ; AVX2-NEXT:    vpmovmskb %ymm0, %eax
 ; AVX2-NEXT:    vzeroupper
 ; AVX2-NEXT:    retq
 ;
 ; AVX512F-LABEL: v16i8_widened_with_zeroes:
 ; AVX512F:       # %bb.0: # %entry
 ; AVX512F-NEXT:    vpcmpeqb %xmm1, %xmm0, %xmm0
 ; AVX512F-NEXT:    vpmovsxbd %xmm0, %zmm0

llvm/trunk/test/CodeGen/X86/merge-consecutive-loads-256.ll

 ; consecutive loads including any/all volatiles may not be combined
 ;
 
 define <4 x double> @merge_4f64_f64_34uz_volatile(double* %ptr) nounwind uwtable noinline ssp {
 ; AVX-LABEL: merge_4f64_f64_34uz_volatile:
 ; AVX:       # %bb.0:
 ; AVX-NEXT:    vmovsd {{.*#+}} xmm0 = mem[0],zero
 ; AVX-NEXT:    vmovhpd {{.*#+}} xmm0 = xmm0[0],mem[0]
-; AVX-NEXT:    vmovapd %xmm0, %xmm0
 ; AVX-NEXT:    retq
 ;
 ; X32-AVX-LABEL: merge_4f64_f64_34uz_volatile:
 ; X32-AVX:       # %bb.0:
 ; X32-AVX-NEXT:    movl {{[0-9]+}}(%esp), %eax
 ; X32-AVX-NEXT:    vmovsd {{.*#+}} xmm0 = mem[0],zero
 ; X32-AVX-NEXT:    vmovhpd {{.*#+}} xmm0 = xmm0[0],mem[0]
-; X32-AVX-NEXT:    vmovapd %xmm0, %xmm0
 ; X32-AVX-NEXT:    retl
   %ptr0 = getelementptr inbounds double, double* %ptr, i64 3
   %ptr1 = getelementptr inbounds double, double* %ptr, i64 4
   %val0 = load volatile double, double* %ptr0
   %val1 = load volatile double, double* %ptr1
   %res0 = insertelement <4 x double> undef, double %val0, i32 0
   %res1 = insertelement <4 x double> %res0, double %val1, i32 1
   %res3 = insertelement <4 x double> %res1, double   0.0, i32 3

llvm/trunk/test/CodeGen/X86/vector-shuffle-variable-256.ll

 ; ALL-NEXT:    andq $-32, %rsp
 ; ALL-NEXT:    subq $64, %rsp
 ; ALL-NEXT:    andl $3, %edi
 ; ALL-NEXT:    andl $3, %esi
 ; ALL-NEXT:    vmovaps %ymm0, (%rsp)
 ; ALL-NEXT:    vmovsd {{.*#+}} xmm0 = mem[0],zero
 ; ALL-NEXT:    vmovsd {{.*#+}} xmm1 = mem[0],zero
 ; ALL-NEXT:    vmovlhps {{.*#+}} xmm0 = xmm1[0],xmm0[0]
-; ALL-NEXT:    vmovaps %xmm0, %xmm0
 ; ALL-NEXT:    movq %rbp, %rsp
 ; ALL-NEXT:    popq %rbp
 ; ALL-NEXT:    retq
   %x0 = extractelement <4 x i64> %x, i64 %i0
   %x1 = extractelement <4 x i64> %x, i64 %i1
   %x2 = extractelement <4 x i64> %x, i64 %i2
   %x3 = extractelement <4 x i64> %x, i64 %i3
   %r0 = insertelement <4 x i64> undef, i64 %x0, i32 0