[AArch64][SVE] Add initial backend support for FP splat_vector

[cameron.mcinally]

Summary

This issue was found in D73711.

We'll also need separate changes to CodeGen wrt splat_vectors. There are some incorrect uses of scalable build_vectors (int and fp) scattered there. This is obvious when an assert to check for scalable build_vectors is added to DAG.getNode(...).

Notice the TODO in the f16 immediate splat tests. There seems to be a missing lowering for f16 constants. Instead of becoming ConstantFPs, the f16 constants are being explicitly loaded from the constant pool. Does anyone know where the special f16 constant lowerings live? I don't have a lot of experience with those.

[cameron.mcinally]

Fix bad copy-and-paste.

[efriedma]

There seems to be a missing lowering for f16 constants. Instead of becoming ConstantFPs, the f16 constants are being explicitly loaded from the constant pool

See AArch64TargetLowering::isFPImmLegal ? We should generate an appropriate fmov with the right target features. I don't think "-mattr=+sve" implies those features, though.

llvm/lib/Target/AArch64/AArch64SVEInstrInfo.td

297

I'm not adding a pseudo-instruction is worthwhile if the only benefit is avoiding an INSERT_SUBREG.

310

What do we end up generating for a non-zero float immediate? We might need a pattern to avoid an extra mov in the general case.

In theory, we can generate other float immediates using the integer dup/dupm, but I guess most of them won't be useful for 32-bit or 64-bit floats. Some probably are, though; for example, you can generate 1.0 with dupm.

[cameron.mcinally]

In D74632#1877178, @efriedma wrote:

There seems to be a missing lowering for f16 constants. Instead of becoming ConstantFPs, the f16 constants are being explicitly loaded from the constant pool

See AArch64TargetLowering::isFPImmLegal ? We should generate an appropriate fmov with the right target features. I don't think "-mattr=+sve" implies those features, though.

Thanks, I'll take a look.

I also just noticed that the test file I added was renamed upstream to sve-vector-splat.ll. I'll move these new tests to that file.

llvm/lib/Target/AArch64/AArch64SVEInstrInfo.td

297

Good point. Will update it.

I'm not the original author (cherry-picked from D71712), so maybe I'm missing something though...

310

I'm fairly certain that support exists in D71712. I didn't include any of the shufflevector tests or patterns in this Diff though.

My intention was to cherry-pick something small for easy reviewing. If you'd like to see that support included in this patch, I'll add it.

[efriedma]

llvm/lib/Target/AArch64/AArch64SVEInstrInfo.td

310

Oh, that's fine, it doesn't need to be in the same patch. Just wanted to make sure you were considering it.

[cameron.mcinally]

Update with @efriedma suggestions.

[efriedma]

llvm/lib/Target/AArch64/AArch64ISelLowering.cpp

878

I think setOperationAction(ISD::ConstantFP, MVT::f16, Legal) is going to cause a fatal error for some FP constants. Not that they would be impossible to lower appropriately, but I don't think we have the necessary patterns.

[cameron.mcinally]

llvm/lib/Target/AArch64/AArch64ISelLowering.cpp

878

Ah, good call. That was also from D71712, but yeah, there are probably missing f16 patterns.

Are we're ok with not folding the f16 constants for now? Or I could wait on this patch too. Any preference?

[efriedma]

llvm/lib/Target/AArch64/AArch64ISelLowering.cpp

878

I'm okay if we miss the optimization for now...

That said, I think making +sve imply +fullfp16 solves the immediate problem you're running into, where constantfp 0.0 is getting lowered to a constant pool.

[cameron.mcinally]

Don't mark f16 as a legal ConstantFP for SVE.

That said, I think making +sve imply +fullfp16 solves the immediate problem you're running into, where constantfp 0.0 is getting lowered to a constant pool.

I don't have a lot of intuition built up around fp16, so I'm hesitant to change it. I will open another Diff linking the two feature flags though, so we can start a discussion.

[efriedma]

LGTM, assuming the regression test still passes.

llvm/lib/Target/AArch64/AArch64ISelLowering.cpp

if (Subtarget->hasSVE()) {
     // splat of 0 or undef) once vector selects supported in SVE codegen. See
     // D68877 for more details.
     for (MVT VT : MVT::integer_scalable_vector_valuetypes()) {
       if (isTypeLegal(VT))
         setOperationAction(ISD::SPLAT_VECTOR, VT, Custom);
     }
     setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::i8, Custom);
     setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::i16, Custom);
+
+    for (MVT VT : MVT::fp_scalable_vector_valuetypes()) {
+      if (isTypeLegal(VT)) {
+        setOperationAction(ISD::SPLAT_VECTOR, VT, Custom);
+      }
+    }
   }
 
   PredictableSelectIsExpensive = Subtarget->predictableSelectIsExpensive();
 }
 
 void AArch64TargetLowering::addTypeForNEON(MVT VT, MVT PromotedBitwiseVT) {
   assert(VT.isVector() && "VT should be a vector type");
 
SDValue AArch64TargetLowering::LowerSPLAT_VECTOR(SDValue Op,
   EVT VT = Op.getValueType();
   EVT ElemVT = VT.getScalarType();
 
   SDValue SplatVal = Op.getOperand(0);
 
   // Extend input splat value where needed to fit into a GPR (32b or 64b only)
   // FPRs don't have this restriction.
   switch (ElemVT.getSimpleVT().SimpleTy) {
-  case MVT::i8:
-  case MVT::i16:
-  case MVT::i32:
-    SplatVal = DAG.getAnyExtOrTrunc(SplatVal, dl, MVT::i32);
-    return DAG.getNode(AArch64ISD::DUP, dl, VT, SplatVal);
-  case MVT::i64:
-    SplatVal = DAG.getAnyExtOrTrunc(SplatVal, dl, MVT::i64);
-    return DAG.getNode(AArch64ISD::DUP, dl, VT, SplatVal);
   case MVT::i1: {
     // The general case of i1.  There isn't any natural way to do this,
     // so we use some trickery with whilelo.
     // TODO: Add special cases for splat of constant true/false.
     SplatVal = DAG.getAnyExtOrTrunc(SplatVal, dl, MVT::i64);
     SplatVal = DAG.getNode(ISD::SIGN_EXTEND_INREG, dl, MVT::i64, SplatVal,
                            DAG.getValueType(MVT::i1));
     SDValue ID = DAG.getTargetConstant(Intrinsic::aarch64_sve_whilelo, dl,
                                        MVT::i64);
     return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT, ID,
                        DAG.getConstant(0, dl, MVT::i64), SplatVal);
   }
-  // TODO: we can support float types, but haven't added patterns yet.
+  case MVT::i8:
+  case MVT::i16:
+  case MVT::i32:
+    SplatVal = DAG.getAnyExtOrTrunc(SplatVal, dl, MVT::i32);
+    break;
+  case MVT::i64:
+    SplatVal = DAG.getAnyExtOrTrunc(SplatVal, dl, MVT::i64);
+    break;
   case MVT::f16:
   case MVT::f32:
   case MVT::f64:
+    // Fine as is
+    break;
   default:
     report_fatal_error("Unsupported SPLAT_VECTOR input operand type");
   }
+
+  return DAG.getNode(AArch64ISD::DUP, dl, VT, SplatVal);
 }
 
 static bool resolveBuildVector(BuildVectorSDNode *BVN, APInt &CnstBits,
                                APInt &UndefBits) {
   EVT VT = BVN->getValueType(0);
   APInt SplatBits, SplatUndef;
   unsigned SplatBitSize;
   bool HasAnyUndefs;

llvm/lib/Target/AArch64/AArch64InstrFormats.td

 
 // Pseudo instructions (don't have encoding information)
 class Pseudo<dag oops, dag iops, list<dag> pattern, string cstr = "">
     : AArch64Inst<PseudoFrm, cstr> {
   dag OutOperandList = oops;
   dag InOperandList  = iops;
   let Pattern        = pattern;
   let isCodeGenOnly  = 1;
+  let isPseudo       = 1;
 }
 
 // Real instructions (have encoding information)
 class EncodedI<string cstr, list<dag> pattern> : AArch64Inst<NormalFrm, cstr> {
   let Pattern = pattern;
   let Size = 4;
 }
 

llvm/lib/Target/AArch64/AArch64InstrInfo.cpp

bool AArch64InstrInfo::substituteCmpToZero(
   bool succeeded = UpdateOperandRegClass(*MI);
   (void)succeeded;
   assert(succeeded && "Some operands reg class are incompatible!");
   MI->addRegisterDefined(AArch64::NZCV, TRI);
   return true;
 }
 
 bool AArch64InstrInfo::expandPostRAPseudo(MachineInstr &MI) const {
+  const TargetRegisterInfo *TRI = &getRegisterInfo();
+  MachineBasicBlock &MBB = *MI.getParent();
+  DebugLoc DL = MI.getDebugLoc();
+  unsigned Opcode = MI.getOpcode();
+
+  if ((Opcode == AArch64::DUP_ZV_H) ||
+      (Opcode == AArch64::DUP_ZV_S) ||
+      (Opcode == AArch64::DUP_ZV_D)) {
+    auto RC = &AArch64::ZPRRegClass;
+    unsigned Dst = MI.getOperand(0).getReg();
+    unsigned Src = MI.getOperand(1).getReg();
+    bool SrcIsKill = MI.getOperand(1).isKill();
+
+    unsigned NewOpcode;
+    unsigned NewSrc;
+    switch (Opcode) {
+    case AArch64::DUP_ZV_H:
+      NewOpcode = AArch64::DUP_ZZI_H;
+      NewSrc = TRI->getMatchingSuperReg(Src, AArch64::hsub, RC);
+      break;
+    case AArch64::DUP_ZV_S:
+      NewOpcode = AArch64::DUP_ZZI_S;
+      NewSrc = TRI->getMatchingSuperReg(Src, AArch64::ssub, RC);
+      break;
+    case AArch64::DUP_ZV_D:
+      NewOpcode = AArch64::DUP_ZZI_D;
+      NewSrc = TRI->getMatchingSuperReg(Src, AArch64::dsub, RC);
+      break;
+    }
+
+    BuildMI(MBB, MI, DL, get(NewOpcode), Dst)
+        .addReg(NewSrc, getKillRegState(SrcIsKill))
+        .addImm(0);
+    MBB.erase(MI);
+    return true;
+  }
+
   if (MI.getOpcode() != TargetOpcode::LOAD_STACK_GUARD &&
       MI.getOpcode() != AArch64::CATCHRET)
     return false;
 
-  MachineBasicBlock &MBB = *MI.getParent();
   auto &Subtarget = MBB.getParent()->getSubtarget<AArch64Subtarget>();
-  auto TRI = Subtarget.getRegisterInfo();
-  DebugLoc DL = MI.getDebugLoc();
 
   if (MI.getOpcode() == AArch64::CATCHRET) {
     // Skip to the first instruction before the epilog.
     const TargetInstrInfo *TII =
       MBB.getParent()->getSubtarget().getInstrInfo();
     MachineBasicBlock *TargetMBB = MI.getOperand(0).getMBB();
     auto MBBI = MachineBasicBlock::iterator(MI);
     MachineBasicBlock::iterator FirstEpilogSEH = std::prev(MBBI);

llvm/lib/Target/AArch64/AArch64SVEInstrInfo.td

let Predicates = [HasSVE] in {
   // Splat scalar register (unpredicated, GPR or vector + element index)
   defm DUP_ZR  : sve_int_perm_dup_r<"dup", AArch64dup>;
   defm DUP_ZZI : sve_int_perm_dup_i<"dup">;
 
   // Splat scalar register (predicated)
   defm CPY_ZPmR : sve_int_perm_cpy_r<"cpy", AArch64dup_pred>;
   defm CPY_ZPmV : sve_int_perm_cpy_v<"cpy", AArch64dup_pred>;
 
+  // Create variants of DUP_ZZI (I=0) that can be used without INSERT_SUBREG.
+  def DUP_ZV_H : Pseudo<(outs ZPR16:$Zd), (ins FPR16:$Vn), []>, Sched<[]>;
+  def DUP_ZV_S : Pseudo<(outs ZPR32:$Zd), (ins FPR32:$Vn), []>, Sched<[]>;
+  def DUP_ZV_D : Pseudo<(outs ZPR64:$Zd), (ins FPR64:$Vn), []>, Sched<[]>;
+
+  // Duplicate FP scalar into all vector elements
+  def : Pat<(nxv8f16 (AArch64dup (f16 FPR16:$src))), (DUP_ZV_H $src)>;
+  def : Pat<(nxv4f16 (AArch64dup (f16 FPR16:$src))), (DUP_ZV_H $src)>;
+  def : Pat<(nxv2f16 (AArch64dup (f16 FPR16:$src))), (DUP_ZV_H $src)>;
+  def : Pat<(nxv4f32 (AArch64dup (f32 FPR32:$src))), (DUP_ZV_S $src)>;
+  def : Pat<(nxv2f32 (AArch64dup (f32 FPR32:$src))), (DUP_ZV_S $src)>;
+  def : Pat<(nxv2f64 (AArch64dup (f64 FPR64:$src))), (DUP_ZV_D $src)>;
+
+  // Duplicate +0.0 into all vector elements
+  def : Pat<(nxv8f16 (AArch64dup (f16 fpimm0))), (DUP_ZI_H 0, 0)>;
+  def : Pat<(nxv4f16 (AArch64dup (f16 fpimm0))), (DUP_ZI_H 0, 0)>;
+  def : Pat<(nxv2f16 (AArch64dup (f16 fpimm0))), (DUP_ZI_H 0, 0)>;
+  def : Pat<(nxv4f32 (AArch64dup (f32 fpimm0))), (DUP_ZI_S 0, 0)>;
+  def : Pat<(nxv2f32 (AArch64dup (f32 fpimm0))), (DUP_ZI_S 0, 0)>;
+  def : Pat<(nxv2f64 (AArch64dup (f64 fpimm0))), (DUP_ZI_D 0, 0)>;
+
   // Select elements from either vector (predicated)
   defm SEL_ZPZZ    : sve_int_sel_vvv<"sel", vselect>;
 
   defm SPLICE_ZPZ : sve_int_perm_splice<"splice", int_aarch64_sve_splice>;
   defm COMPACT_ZPZ : sve_int_perm_compact<"compact", int_aarch64_sve_compact>;
   defm INSR_ZR : sve_int_perm_insrs<"insr", AArch64insr>;
   defm INSR_ZV : sve_int_perm_insrv<"insr", AArch64insr>;
   defm EXT_ZZI : sve_int_perm_extract_i<"ext", AArch64ext>;

llvm/test/CodeGen/AArch64/vector-splat-sve.ll

+; RUN: llc < %s -mtriple=aarch64--linux-gnu -mattr=+sve | FileCheck %s
+
+define <vscale x 8 x half> @splat_nxv8f16(half %val) {
+; CHECK-LABEL: splat_nxv8f16:
+; CHECK: mov z0.h, h0
+; CHECK-NEXT: ret
+  %1 = insertelement <vscale x 8 x half> undef, half %val, i32 0
+  %2 = shufflevector <vscale x 8 x half> %1, <vscale x 8 x half> undef, <vscale x 8 x i32> zeroinitializer
+  ret <vscale x 8 x half> %2
+}
+
+define <vscale x 4 x half> @splat_nxv4f16(half %val) {
+; CHECK-LABEL: splat_nxv4f16:
+; CHECK: mov z0.h, h0
+; CHECK-NEXT: ret
+  %1 = insertelement <vscale x 4 x half> undef, half %val, i32 0
+  %2 = shufflevector <vscale x 4 x half> %1, <vscale x 4 x half> undef, <vscale x 4 x i32> zeroinitializer
+  ret <vscale x 4 x half> %2
+}
+
+define <vscale x 2 x half> @splat_nxv2f16(half %val) {
+; CHECK-LABEL: splat_nxv2f16:
+; CHECK: mov z0.h, h0
+; CHECK-NEXT: ret
+  %1 = insertelement <vscale x 2 x half> undef, half %val, i32 0
+  %2 = shufflevector <vscale x 2 x half> %1, <vscale x 2 x half> undef, <vscale x 2 x i32> zeroinitializer
+  ret <vscale x 2 x half> %2
+}
+
+define <vscale x 4 x float> @splat_nxv4f32(float %val) {
+; CHECK-LABEL: splat_nxv4f32:
+; CHECK: mov z0.s, s0
+; CHECK-NEXT: ret
+  %1 = insertelement <vscale x 4 x float> undef, float %val, i32 0
+  %2 = shufflevector <vscale x 4 x float> %1, <vscale x 4 x float> undef, <vscale x 4 x i32> zeroinitializer
+  ret <vscale x 4 x float> %2
+}
+
+define <vscale x 2 x float> @splat_nxv2f32(float %val) {
+; CHECK-LABEL: splat_nxv2f32:
+; CHECK: mov z0.s, s0
+; CHECK-NEXT: ret
+  %1 = insertelement <vscale x 2 x float> undef, float %val, i32 0
+  %2 = shufflevector <vscale x 2 x float> %1, <vscale x 2 x float> undef, <vscale x 2 x i32> zeroinitializer
+  ret <vscale x 2 x float> %2
+}
+
+define <vscale x 2 x double> @splat_nxv2f64(double %val) {
+; CHECK-LABEL: splat_nxv2f64:
+; CHECK: mov z0.d, d0
+; CHECK-NEXT: ret
+  %1 = insertelement <vscale x 2 x double> undef, double %val, i32 0
+  %2 = shufflevector <vscale x 2 x double> %1, <vscale x 2 x double> undef, <vscale x 2 x i32> zeroinitializer
+  ret <vscale x 2 x double> %2
+}
+
+; TODO: The f16 constant load isn't folded correctly.
+define <vscale x 8 x half> @splat_nxv8f16_zero() {
+; CHECK-LABEL: splat_nxv8f16_zero:
+; CHECK: mov z0.h, h0
+; CHECK-NEXT: ret
+  ret <vscale x 8 x half> zeroinitializer
+}
+
+; TODO: The f16 constant load isn't folded correctly.
+define <vscale x 4 x half> @splat_nxv4f16_zero() {
+; CHECK-LABEL: splat_nxv4f16_zero:
+; CHECK: mov z0.h, h0
+; CHECK-NEXT: ret
+  ret <vscale x 4 x half> zeroinitializer
+}
+
+; TODO: The f16 constant load isn't folded correctly.
+define <vscale x 2 x half> @splat_nxv2f16_zero() {
+; CHECK-LABEL: splat_nxv2f16_zero:
+; CHECK: mov z0.h, h0
+; CHECK-NEXT: ret
+  ret <vscale x 2 x half> zeroinitializer
+}
+
+define <vscale x 4 x float> @splat_nxv4f32_zero() {
+; CHECK-LABEL: splat_nxv4f32_zero:
+; CHECK: mov z0.s, #0
+; CHECK-NEXT: ret
+  ret <vscale x 4 x float> zeroinitializer
+}
+
+define <vscale x 2 x float> @splat_nxv2f32_zero() {
+; CHECK-LABEL: splat_nxv2f32_zero:
+; CHECK: mov z0.s, #0
+; CHECK-NEXT: ret
+  ret <vscale x 2 x float> zeroinitializer
+}
+
+define <vscale x 2 x double> @splat_nxv2f64_zero() {
+; CHECK-LABEL: splat_nxv2f64_zero:
+; CHECK: mov z0.d, #0
+; CHECK-NEXT: ret
+  ret <vscale x 2 x double> zeroinitializer
+}