[ARM][BFloat] Legalize bf16 type even without fullfp16.

[simon_tatham]

Summary

This change permits scalar bfloats to be loaded, stored, moved and
used as function call arguments and return values, whenever the bf16
feature is supported by the subtarget.

Previously that was only supported in the presence of the fullfp16
feature, because the code generation strategy depended on instructions
from that extension. This change adds alternative code generation
strategies so that those operations can be done even without fullfp16.

The strategy for loads and stores is to replace VLDRH/VSTRH with
integer LDRH/STRH plus a move between register classes. I've written
isel patterns for those, conditional on not having the fullfp16
feature (so that in the fullfp16 case, the existing patterns will
still be used).

For function arguments and returns, instead of writing isel patterns
to match VMOVhr and VMOVrh, I've avoided generating those SDNodes
in the first place, by factoring out the code that constructs them
into helper functions MoveToHPR and MoveFromHPR which have a
fallback for non-fullfp16 subtargets.

The current output code is not especially pretty: in the new test file
you can see unnecessary store/load pairs implementing no-op bitcasts,
and lots of pointless moves back and forth between FP registers and
GPRs. But it at least works, which is an improvement on the previous
situation.

[labrinea]

Hi Simon, thanks for working on this. Looks good overall. A few remarks inline.

llvm/lib/Target/ARM/ARMISelLowering.cpp

726–730

Shouldn't this and the following 4 lines be contitional to Subtarget->hasBF16() ?

llvm/lib/Target/ARM/ARMInstrVFP.td

171

Can you either define a new pattern with this predicate, or delete the FPRegs16Pat pattern and use the same syntax for uniformity? Just a personal preference.

llvm/lib/Target/ARM/Thumb2InstrInfo.cpp

595 ↗

(On Diff #272692)

Why do we need this block of changes?

[simon_tatham]

New version addressing review comments.

llvm/lib/Target/ARM/ARMISelLowering.cpp

726–730

How did I miss that?! Thanks, good catch.

llvm/lib/Target/ARM/Thumb2InstrInfo.cpp

595 ↗

(On Diff #272692)

I put it in because otherwise the Thumb invocations of llc in my new test file failed with the 'Unsupported addressing mode' assertion in this function.

However, now I look more closely, that's probably because I was using the wrong instruction – I was using LDRH for both Arm and Thumb, whereas I should have used t2LDRHi12 in Thumb. With that change, I don't seem to need this extra addressing mode support any more.

(But I couldn't tell you why it's not needed, because as far as I can tell LDRH itself still does use addressing mode 3, and is still selected in Arm state.)

[simon_tatham]

(Sorry @labrinea – removing you as a reviewer was an unintentional side effect of a bad arc command.)

[labrinea]

LGTM with a nit. Can you also remove FPRegs16Pat from ARMInstrFormats.td now that is no longer used?

[dmgreen]

llvm/lib/Target/ARM/ARMISelLowering.cpp

729

I'm not sure if setAllExpand should be guarded by !hasFullFP16? In either case they would be in pretty much the same situation - most operations are illegal with bf16. Does something go wrong always doing this?

llvm/lib/Target/ARM/ARMInstrVFP.td

166–177

I would expect HasNoFPRegs16 + f16 to never come up? If so you could alternatively move that single pattern back into the instruction above.

[simon_tatham]

It isn't unused – there are still some isel patterns using it, which I haven't touched in this commit.

[simon_tatham]

llvm/lib/Target/ARM/ARMInstrVFP.td

166–177

I'll give that a try too.

[simon_tatham]

Updated for @dmgreen's review suggestions, which both seem to work as far as I can see.

[dmgreen]

Thanks. If that setAllExpand does start causing problems we can always modify it. LGTM

[simon_tatham]

llvm/lib/Target/ARM/ARMISelLowering.cpp

729

(Huh, where did my reply to this comment go? I must have forgotten to press 'Save Draft'.)

I did it that way in order to avoid affecting the hasFullFP16 code path, because that way I was confident I wouldn't break it as a side effect :-) But I've just tried out the version with setAllExpand unconditional, and it seems to work just as well.

llvm/lib/Target/ARM/ARMISelLowering.h

private:
     SDValue LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG,
                               const ARMSubtarget *ST) const;
     SDValue LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) const;
     SDValue LowerFSINCOS(SDValue Op, SelectionDAG &DAG) const;
     SDValue LowerDivRem(SDValue Op, SelectionDAG &DAG) const;
     SDValue LowerDIV_Windows(SDValue Op, SelectionDAG &DAG, bool Signed) const;
     void ExpandDIV_Windows(SDValue Op, SelectionDAG &DAG, bool Signed,
                            SmallVectorImpl<SDValue> &Results) const;
+    SDValue ExpandBITCAST(SDNode *N, SelectionDAG &DAG,
+                          const ARMSubtarget *Subtarget) const;
     SDValue LowerWindowsDIVLibCall(SDValue Op, SelectionDAG &DAG, bool Signed,
                                    SDValue &Chain) const;
     SDValue LowerREM(SDNode *N, SelectionDAG &DAG) const;
     SDValue LowerDYNAMIC_STACKALLOC(SDValue Op, SelectionDAG &DAG) const;
     SDValue LowerFP_ROUND(SDValue Op, SelectionDAG &DAG) const;
     SDValue LowerFP_EXTEND(SDValue Op, SelectionDAG &DAG) const;
     SDValue LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG) const;
     SDValue LowerINT_TO_FP(SDValue Op, SelectionDAG &DAG) const;
     SDValue LowerFSETCC(SDValue Op, SelectionDAG &DAG) const;
     void lowerABS(SDNode *N, SmallVectorImpl<SDValue> &Results,
                   SelectionDAG &DAG) const;
     void LowerLOAD(SDNode *N, SmallVectorImpl<SDValue> &Results,
                    SelectionDAG &DAG) const;
 
     Register getRegisterByName(const char* RegName, LLT VT,
                                const MachineFunction &MF) const override;
 
     SDValue BuildSDIVPow2(SDNode *N, const APInt &Divisor, SelectionDAG &DAG,
                           SmallVectorImpl<SDNode *> &Created) const override;
 
     bool isFMAFasterThanFMulAndFAdd(const MachineFunction &MF,
                                     EVT VT) const override;
 
+    SDValue MoveToHPR(const SDLoc &dl, SelectionDAG &DAG, MVT LocVT, MVT ValVT,
+                      SDValue Val) const;
+    SDValue MoveFromHPR(const SDLoc &dl, SelectionDAG &DAG, MVT LocVT,
+                        MVT ValVT, SDValue Val) const;
+
     SDValue ReconstructShuffle(SDValue Op, SelectionDAG &DAG) const;
 
     SDValue LowerCallResult(SDValue Chain, SDValue InFlag,
                             CallingConv::ID CallConv, bool isVarArg,
                             const SmallVectorImpl<ISD::InputArg> &Ins,
                             const SDLoc &dl, SelectionDAG &DAG,
                             SmallVectorImpl<SDValue> &InVals, bool isThisReturn,
                             SDValue ThisVal) const;

llvm/lib/Target/ARM/ARMISelLowering.cpp

ARMTargetLowering::ARMTargetLowering(const TargetMachine &TM,
 
   if (Subtarget->hasFullFP16()) {
     addRegisterClass(MVT::f16, &ARM::HPRRegClass);
     setOperationAction(ISD::BITCAST, MVT::i16, Custom);
     setOperationAction(ISD::BITCAST, MVT::f16, Custom);
 
     setOperationAction(ISD::FMINNUM, MVT::f16, Legal);
     setOperationAction(ISD::FMAXNUM, MVT::f16, Legal);
+  }
 
-    // For the time being bfloat is only supported when fullfp16 is present.
-    if (Subtarget->hasBF16())
-      addRegisterClass(MVT::bf16, &ARM::HPRRegClass);
+  if (Subtarget->hasBF16()) {
+    setAllExpand(MVT::bf16);
+    if (!Subtarget->hasFullFP16())
+      setOperationAction(ISD::BITCAST, MVT::bf16, Custom);
   }
 
   for (MVT VT : MVT::fixedlen_vector_valuetypes()) {
     for (MVT InnerVT : MVT::fixedlen_vector_valuetypes()) {
       setTruncStoreAction(VT, InnerVT, Expand);
       addAllExtLoads(VT, InnerVT, Expand);
     }
 
case CallingConv::GHC:
     return (Return ? RetCC_ARM_APCS : CC_ARM_APCS_GHC);
   case CallingConv::PreserveMost:
     return (Return ? RetCC_ARM_AAPCS : CC_ARM_AAPCS);
   case CallingConv::CFGuard_Check:
     return (Return ? RetCC_ARM_AAPCS : CC_ARM_Win32_CFGuard_Check);
   }
 }
 
+SDValue ARMTargetLowering::MoveToHPR(const SDLoc &dl, SelectionDAG &DAG,
+                                     MVT LocVT, MVT ValVT, SDValue Val) const {
+  Val = DAG.getNode(ISD::BITCAST, dl, MVT::getIntegerVT(LocVT.getSizeInBits()),
+                    Val);
+  if (Subtarget->hasFullFP16()) {
+    Val = DAG.getNode(ARMISD::VMOVhr, dl, ValVT, Val);
+  } else {
+    Val = DAG.getNode(ISD::TRUNCATE, dl,
+                      MVT::getIntegerVT(ValVT.getSizeInBits()), Val);
+    Val = DAG.getNode(ISD::BITCAST, dl, ValVT, Val);
+  }
+  return Val;
+}
+
+SDValue ARMTargetLowering::MoveFromHPR(const SDLoc &dl, SelectionDAG &DAG,
+                                       MVT LocVT, MVT ValVT,
+                                       SDValue Val) const {
+  if (Subtarget->hasFullFP16()) {
+    Val = DAG.getNode(ARMISD::VMOVrh, dl,
+                      MVT::getIntegerVT(LocVT.getSizeInBits()), Val);
+  } else {
+    Val = DAG.getNode(ISD::BITCAST, dl,
+                      MVT::getIntegerVT(ValVT.getSizeInBits()), Val);
+    Val = DAG.getNode(ISD::ZERO_EXTEND, dl,
+                      MVT::getIntegerVT(LocVT.getSizeInBits()), Val);
+  }
+  return DAG.getNode(ISD::BITCAST, dl, LocVT, Val);
+}
+
 /// LowerCallResult - Lower the result values of a call into the
 /// appropriate copies out of appropriate physical registers.
 SDValue ARMTargetLowering::LowerCallResult(
     SDValue Chain, SDValue InFlag, CallingConv::ID CallConv, bool isVarArg,
     const SmallVectorImpl<ISD::InputArg> &Ins, const SDLoc &dl,
     SelectionDAG &DAG, SmallVectorImpl<SDValue> &InVals, bool isThisReturn,
     SDValue ThisVal) const {
   // Assign locations to each value returned by this call.
case CCValAssign::BCvt:
       Val = DAG.getNode(ISD::BITCAST, dl, VA.getValVT(), Val);
       break;
     }
 
     // f16 arguments have their size extended to 4 bytes and passed as if they
     // had been copied to the LSBs of a 32-bit register.
     // For that, it's passed extended to i32 (soft ABI) or to f32 (hard ABI)
     if (VA.needsCustom() &&
-        (VA.getValVT() == MVT::f16 || VA.getValVT() == MVT::bf16)) {
-      assert(Subtarget->hasFullFP16() &&
-             "Lowering half precision fp return without full fp16 support");
-      Val = DAG.getNode(ISD::BITCAST, dl,
-                        MVT::getIntegerVT(VA.getLocVT().getSizeInBits()), Val);
-      Val = DAG.getNode(ARMISD::VMOVhr, dl, VA.getValVT(), Val);
-    }
+        (VA.getValVT() == MVT::f16 || VA.getValVT() == MVT::bf16))
+      Val = MoveToHPR(dl, DAG, VA.getLocVT(), VA.getValVT(), Val);
 
     InVals.push_back(Val);
   }
 
   return Chain;
 }
 
 /// LowerMemOpCallTo - Store the argument to the stack.
case CCValAssign::BCvt:
       break;
     }
 
     // f16 arguments have their size extended to 4 bytes and passed as if they
     // had been copied to the LSBs of a 32-bit register.
     // For that, it's passed extended to i32 (soft ABI) or to f32 (hard ABI)
     if (VA.needsCustom() &&
         (VA.getValVT() == MVT::f16 || VA.getValVT() == MVT::bf16)) {
-      assert(Subtarget->hasFullFP16() &&
-             "Lowering half precision fp argument without full fp16 support");
-      Arg = DAG.getNode(ARMISD::VMOVrh, dl,
-                        MVT::getIntegerVT(VA.getLocVT().getSizeInBits()), Arg);
-      Arg = DAG.getNode(ISD::BITCAST, dl, VA.getLocVT(), Arg);
+      Arg = MoveFromHPR(dl, DAG, VA.getLocVT(), VA.getValVT(), Arg);
     } else {
       // f16 arguments could have been extended prior to argument lowering.
       // Mask them arguments if this is a CMSE nonsecure call.
       auto ArgVT = Outs[realArgIdx].ArgVT;
       if (isCmseNSCall && (ArgVT == MVT::f16)) {
         auto LocBits = VA.getLocVT().getSizeInBits();
         auto MaskValue = APInt::getLowBitsSet(LocBits, ArgVT.getSizeInBits());
         SDValue Mask =
case CCValAssign::BCvt:
         Arg = DAG.getNode(ISD::BITCAST, dl, VA.getLocVT(), Arg);
       break;
     }
 
     // Mask f16 arguments if this is a CMSE nonsecure entry.
     auto RetVT = Outs[realRVLocIdx].ArgVT;
     if (AFI->isCmseNSEntryFunction() && (RetVT == MVT::f16)) {
       if (VA.needsCustom() && VA.getValVT() == MVT::f16) {
-        assert(Subtarget->hasFullFP16() &&
-               "Lowering f16 type argument without full fp16 support");
-        Arg =
-            DAG.getNode(ARMISD::VMOVrh, dl,
-                        MVT::getIntegerVT(VA.getLocVT().getSizeInBits()), Arg);
-        Arg = DAG.getNode(ISD::BITCAST, dl, VA.getLocVT(), Arg);
+        Arg = MoveFromHPR(dl, DAG, VA.getLocVT(), VA.getValVT(), Arg);
       } else {
         auto LocBits = VA.getLocVT().getSizeInBits();
         auto MaskValue = APInt::getLowBitsSet(LocBits, RetVT.getSizeInBits());
         SDValue Mask =
             DAG.getConstant(MaskValue, dl, MVT::getIntegerVT(LocBits));
         Arg = DAG.getNode(ISD::BITCAST, dl, MVT::getIntegerVT(LocBits), Arg);
         Arg = DAG.getNode(ISD::AND, dl, MVT::getIntegerVT(LocBits), Arg, Mask);
         Arg = DAG.getNode(ISD::BITCAST, dl, VA.getLocVT(), Arg);
if (VA.isRegLoc()) {
         ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
         break;
       }
 
       // f16 arguments have their size extended to 4 bytes and passed as if they
       // had been copied to the LSBs of a 32-bit register.
       // For that, it's passed extended to i32 (soft ABI) or to f32 (hard ABI)
       if (VA.needsCustom() &&
-          (VA.getValVT() == MVT::f16 || VA.getValVT() == MVT::bf16)) {
-        assert(Subtarget->hasFullFP16() &&
-               "Lowering half precision fp argument without full fp16 support");
-        ArgValue = DAG.getNode(ISD::BITCAST, dl,
-                               MVT::getIntegerVT(VA.getLocVT().getSizeInBits()),
-                               ArgValue);
-        ArgValue = DAG.getNode(ARMISD::VMOVhr, dl, VA.getValVT(), ArgValue);
-      }
+          (VA.getValVT() == MVT::f16 || VA.getValVT() == MVT::bf16))
+        ArgValue = MoveToHPR(dl, DAG, VA.getLocVT(), VA.getValVT(), ArgValue);
 
       InVals.push_back(ArgValue);
     } else { // VA.isRegLoc()
       // sanity check
       assert(VA.isMemLoc());
       assert(VA.getValVT() != MVT::i64 && "i64 should already be lowered");
 
       int index = VA.getValNo();
return DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, DstVT, BitCast,
                      DAG.getConstant(NewIndex.getZExtValue(), dl, MVT::i32));
 }
 
 /// ExpandBITCAST - If the target supports VFP, this function is called to
 /// expand a bit convert where either the source or destination type is i64 to
 /// use a VMOVDRR or VMOVRRD node.  This should not be done when the non-i64
 /// operand type is illegal (e.g., v2f32 for a target that doesn't support
 /// vectors), since the legalizer won't know what to do with that.
-static SDValue ExpandBITCAST(SDNode *N, SelectionDAG &DAG,
-                             const ARMSubtarget *Subtarget) {
+SDValue ARMTargetLowering::ExpandBITCAST(SDNode *N, SelectionDAG &DAG,
+                                         const ARMSubtarget *Subtarget) const {
   const TargetLowering &TLI = DAG.getTargetLoweringInfo();
   SDLoc dl(N);
   SDValue Op = N->getOperand(0);
 
   // This function is only supposed to be called for i16 and i64 types, either
   // as the source or destination of the bit convert.
   EVT SrcVT = Op.getValueType();
   EVT DstVT = N->getValueType(0);
 
-  if (SrcVT == MVT::i16 && (DstVT == MVT::f16 || DstVT == MVT::bf16)) {
-    if (!Subtarget->hasFullFP16())
-      return SDValue();
-    // (b)f16 bitcast i16 -> VMOVhr
-    return DAG.getNode(ARMISD::VMOVhr, SDLoc(N), DstVT,
-                       DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N), MVT::i32, Op));
-  }
+  if ((SrcVT == MVT::i16 || SrcVT == MVT::i32) &&
+      (DstVT == MVT::f16 || DstVT == MVT::bf16))
+    return MoveToHPR(SDLoc(N), DAG, MVT::i32, DstVT.getSimpleVT(),
 
-  if ((SrcVT == MVT::f16 || SrcVT == MVT::bf16) && DstVT == MVT::i16) {
-    if (!Subtarget->hasFullFP16())
-      return SDValue();
-    // i16 bitcast (b)f16 -> VMOVrh
-    return DAG.getNode(ISD::TRUNCATE, SDLoc(N), MVT::i16,
-                       DAG.getNode(ARMISD::VMOVrh, SDLoc(N), MVT::i32, Op));
-  }
+  if ((DstVT == MVT::i16 || DstVT == MVT::i32) &&
+      (SrcVT == MVT::f16 || SrcVT == MVT::bf16))
+    return DAG.getNode(
+        ISD::TRUNCATE, SDLoc(N), DstVT,
+        MoveFromHPR(SDLoc(N), DAG, MVT::i32, SrcVT.getSimpleVT(), Op));
 
   if (!(SrcVT == MVT::i64 || DstVT == MVT::i64))
     return SDValue();
 
   // Turn i64->f64 into VMOVDRR.
   if (SrcVT == MVT::i64 && TLI.isTypeLegal(DstVT)) {
     // Do not force values to GPRs (this is what VMOVDRR does for the inputs)
     // if we can combine the bitcast with its source.

llvm/lib/Target/ARM/ARMInstrVFP.td

def VLDRS : ASI5<0b1101, 0b01, (outs SPR:$Sd), (ins addrmode5:$addr),
   // Some single precision VFP instructions may be executed on both NEON and VFP
   // pipelines.
   let D = VFPNeonDomain;
 }
 
 let isUnpredicable = 1 in
 def VLDRH : AHI5<0b1101, 0b01, (outs HPR:$Sd), (ins addrmode5fp16:$addr),
                  IIC_fpLoad16, "vldr", ".16\t$Sd, $addr",
-                 []>,
+                 [(set HPR:$Sd, (f16 (alignedload16 addrmode5fp16:$addr)))]>,
             Requires<[HasFPRegs16]>;
 
 } // End of 'let canFoldAsLoad = 1, isReMaterializable = 1 in'
 
-def : FPRegs16Pat<(f16 (alignedload16 addrmode5fp16:$addr)),
-                  (VLDRH addrmode5fp16:$addr)>;
-def : FPRegs16Pat<(bf16 (alignedload16 addrmode5fp16:$addr)),
-                  (VLDRH addrmode5fp16:$addr)>;
+def : Pat<(bf16 (alignedload16 addrmode5fp16:$addr)),
+          (VLDRH addrmode5fp16:$addr)> {
+  let Predicates = [HasFPRegs16];
+}
+def : Pat<(bf16 (alignedload16 addrmode3:$addr)),
+          (COPY_TO_REGCLASS (LDRH addrmode3:$addr), HPR)> {
+  let Predicates = [HasNoFPRegs16, IsARM];
+}
+def : Pat<(bf16 (alignedload16 t2addrmode_imm12:$addr)),
+          (COPY_TO_REGCLASS (t2LDRHi12 t2addrmode_imm12:$addr), HPR)> {
+  let Predicates = [HasNoFPRegs16, IsThumb];
+}
 
 def VSTRD : ADI5<0b1101, 0b00, (outs), (ins DPR:$Dd, addrmode5:$addr),
                  IIC_fpStore64, "vstr", "\t$Dd, $addr",
                  [(alignedstore32 (f64 DPR:$Dd), addrmode5:$addr)]>,
             Requires<[HasFPRegs]>;
 
 def VSTRS : ASI5<0b1101, 0b00, (outs), (ins SPR:$Sd, addrmode5:$addr),
                  IIC_fpStore32, "vstr", "\t$Sd, $addr",
                  [(alignedstore32 SPR:$Sd, addrmode5:$addr)]>,
             Requires<[HasFPRegs]> {
   // Some single precision VFP instructions may be executed on both NEON and VFP
   // pipelines.
   let D = VFPNeonDomain;
 }
 
 let isUnpredicable = 1 in
 def VSTRH : AHI5<0b1101, 0b00, (outs), (ins HPR:$Sd, addrmode5fp16:$addr),
                  IIC_fpStore16, "vstr", ".16\t$Sd, $addr",
-                 []>,
+                 [(alignedstore16 (f16 HPR:$Sd), addrmode5fp16:$addr)]>,
             Requires<[HasFPRegs16]>;
 
-def : FPRegs16Pat<(alignedstore16 (f16 HPR:$Sd), addrmode5fp16:$addr),
-                  (VSTRH (f16 HPR:$Sd), addrmode5fp16:$addr)>;
-def : FPRegs16Pat<(alignedstore16 (bf16 HPR:$Sd), addrmode5fp16:$addr),
-                  (VSTRH (bf16 HPR:$Sd), addrmode5fp16:$addr)>;
+def : Pat<(alignedstore16 (bf16 HPR:$Sd), addrmode5fp16:$addr),
+          (VSTRH (bf16 HPR:$Sd), addrmode5fp16:$addr)> {
+  let Predicates = [HasFPRegs16];
+}
+def : Pat<(alignedstore16 (bf16 HPR:$Sd), addrmode3:$addr),
+          (STRH (COPY_TO_REGCLASS $Sd, GPR), addrmode3:$addr)> {
+  let Predicates = [HasNoFPRegs16, IsARM];
+}
+def : Pat<(alignedstore16 (bf16 HPR:$Sd), t2addrmode_imm12:$addr),
+          (t2STRHi12 (COPY_TO_REGCLASS $Sd, GPR), t2addrmode_imm12:$addr)> {
+  let Predicates = [HasNoFPRegs16, IsThumb];
+}
 
 //===----------------------------------------------------------------------===//
 //  Load / store multiple Instructions.
 //
 
 multiclass vfp_ldst_mult<string asm, bit L_bit,
                          InstrItinClass itin, InstrItinClass itin_upd> {
   let Predicates = [HasFPRegs] in {

llvm/lib/Target/ARM/ARMPredicates.td

def HasCDE           : Predicate<"Subtarget->hasCDEOps()">,
                                  AssemblerPredicate<(all_of HasCDEOps),
                                                     "cde">;
 def HasFPRegs        : Predicate<"Subtarget->hasFPRegs()">,
                                  AssemblerPredicate<(all_of FeatureFPRegs),
                                                     "fp registers">;
 def HasFPRegs16      : Predicate<"Subtarget->hasFPRegs16()">,
                                  AssemblerPredicate<(all_of FeatureFPRegs16),
                                                     "16-bit fp registers">;
+def HasNoFPRegs16    : Predicate<"!Subtarget->hasFPRegs16()">,
+                                 AssemblerPredicate<(all_of (not FeatureFPRegs16)),
+                                                    "16-bit fp registers">;
 def HasFPRegs64      : Predicate<"Subtarget->hasFPRegs64()">,
                                  AssemblerPredicate<(all_of FeatureFPRegs64),
                                                     "64-bit fp registers">;
 def HasFPRegsV8_1M   : Predicate<"Subtarget->hasFPRegs() && Subtarget->hasV8_1MMainlineOps()">,
                                  AssemblerPredicate<(all_of FeatureFPRegs, HasV8_1MMainlineOps),
                                                     "armv8.1m.main with FP or MVE">;
 def HasV6T2          : Predicate<"Subtarget->hasV6T2Ops()">,
                                  AssemblerPredicate<(all_of HasV6T2Ops), "armv6t2">;

llvm/test/CodeGen/ARM/arm-bf16-pcs.ll

+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
+; RUN: llc -mtriple armv8.6a-arm-none-eabi -o - %s | FileCheck %s --check-prefix=BASE --check-prefix=BASE-ARM
+; RUN: llc -mtriple thumbv8.6a-arm-none-eabi -o - %s | FileCheck %s --check-prefix=BASE --check-prefix=BASE-THUMB
+; RUN: llc -mtriple armv8.6a-arm-none-eabi -mattr=+fullfp16 -o - %s | FileCheck %s --check-prefix=FULLFP16 --check-prefix=FULLFP16-ARM
+; RUN: llc -mtriple thumbv8.6a-arm-none-eabi -mattr=+fullfp16 -o - %s | FileCheck %s --check-prefix=FULLFP16 --check-prefix=FULLFP16-THUMB
+
+define bfloat @bf_load_soft(bfloat* %p) {
+; BASE-LABEL: bf_load_soft:
+; BASE:       @ %bb.0:
+; BASE-NEXT:    ldrh r0, [r0]
+; BASE-NEXT:    bx lr
+;
+; FULLFP16-LABEL: bf_load_soft:
+; FULLFP16:       @ %bb.0:
+; FULLFP16-NEXT:    vldr.16 s0, [r0]
+; FULLFP16-NEXT:    vmov r0, s0
+; FULLFP16-NEXT:    bx lr
+  %f = load bfloat, bfloat* %p, align 2
+  ret bfloat %f
+}
+
+define arm_aapcs_vfpcc bfloat @bf_load_hard(bfloat* %p) {
+; BASE-LABEL: bf_load_hard:
+; BASE:       @ %bb.0:
+; BASE-NEXT:    ldrh r0, [r0]
+; BASE-NEXT:    vmov s0, r0
+; BASE-NEXT:    bx lr
+;
+; FULLFP16-LABEL: bf_load_hard:
+; FULLFP16:       @ %bb.0:
+; FULLFP16-NEXT:    vldr.16 s0, [r0]
+; FULLFP16-NEXT:    bx lr
+  %f = load bfloat, bfloat* %p, align 2
+  ret bfloat %f
+}
+
+define void @bf_store_soft(bfloat* %p, bfloat %f) {
+; BASE-LABEL: bf_store_soft:
+; BASE:       @ %bb.0:
+; BASE-NEXT:    strh r1, [r0]
+; BASE-NEXT:    bx lr
+;
+; FULLFP16-LABEL: bf_store_soft:
+; FULLFP16:       @ %bb.0:
+; FULLFP16-NEXT:    vmov.f16 s0, r1
+; FULLFP16-NEXT:    vstr.16 s0, [r0]
+; FULLFP16-NEXT:    bx lr
+  store bfloat %f, bfloat* %p, align 2
+  ret void
+}
+
+define arm_aapcs_vfpcc void @bf_store_hard(bfloat* %p, bfloat %f) {
+; BASE-LABEL: bf_store_hard:
+; BASE:       @ %bb.0:
+; BASE-NEXT:    vmov r1, s0
+; BASE-NEXT:    strh r1, [r0]
+; BASE-NEXT:    bx lr
+;
+; FULLFP16-LABEL: bf_store_hard:
+; FULLFP16:       @ %bb.0:
+; FULLFP16-NEXT:    vstr.16 s0, [r0]
+; FULLFP16-NEXT:    bx lr
+  store bfloat %f, bfloat* %p, align 2
+  ret void
+}
+
+define i32 @bf_to_int_soft(bfloat %f) {
+; BASE-LABEL: bf_to_int_soft:
+; BASE:       @ %bb.0:
+; BASE-NEXT:    uxth r0, r0
+; BASE-NEXT:    bx lr
+;
+; FULLFP16-LABEL: bf_to_int_soft:
+; FULLFP16:       @ %bb.0:
+; FULLFP16-NEXT:    vmov.f16 s0, r0
+; FULLFP16-NEXT:    vmov.f16 r0, s0
+; FULLFP16-NEXT:    bx lr
+  %h = bitcast bfloat %f to i16
+  %w = zext i16 %h to i32
+  ret i32 %w
+}
+
+define arm_aapcs_vfpcc i32 @bf_to_int_hard(bfloat %f) {
+; BASE-LABEL: bf_to_int_hard:
+; BASE:       @ %bb.0:
+; BASE-NEXT:    vmov r0, s0
+; BASE-NEXT:    uxth r0, r0
+; BASE-NEXT:    bx lr
+;
+; FULLFP16-LABEL: bf_to_int_hard:
+; FULLFP16:       @ %bb.0:
+; FULLFP16-NEXT:    vmov.f16 r0, s0
+; FULLFP16-NEXT:    bx lr
+  %h = bitcast bfloat %f to i16
+  %w = zext i16 %h to i32
+  ret i32 %w
+}
+
+define bfloat @bf_from_int_soft(i32 %w) {
+; BASE-ARM-LABEL: bf_from_int_soft:
+; BASE-ARM:       @ %bb.0:
+; BASE-ARM-NEXT:    .pad #4
+; BASE-ARM-NEXT:    sub sp, sp, #4
+; BASE-ARM-NEXT:    strh r0, [sp, #2]
+; BASE-ARM-NEXT:    ldrh r0, [sp, #2]
+; BASE-ARM-NEXT:    add sp, sp, #4
+; BASE-ARM-NEXT:    bx lr
+;
+; BASE-THUMB-LABEL: bf_from_int_soft:
+; BASE-THUMB:       @ %bb.0:
+; BASE-THUMB-NEXT:    .pad #4
+; BASE-THUMB-NEXT:    sub sp, #4
+; BASE-THUMB-NEXT:    strh.w r0, [sp, #2]
+; BASE-THUMB-NEXT:    ldrh.w r0, [sp, #2]
+; BASE-THUMB-NEXT:    add sp, #4
+; BASE-THUMB-NEXT:    bx lr
+;
+; FULLFP16-LABEL: bf_from_int_soft:
+; FULLFP16:       @ %bb.0:
+; FULLFP16-NEXT:    vmov.f16 s0, r0
+; FULLFP16-NEXT:    vmov r0, s0
+; FULLFP16-NEXT:    bx lr
+  %h = trunc i32 %w to i16
+  %f = bitcast i16 %h to bfloat
+  ret bfloat %f
+}
+
+define arm_aapcs_vfpcc bfloat @bf_from_int_hard(i32 %w) {
+; BASE-ARM-LABEL: bf_from_int_hard:
+; BASE-ARM:       @ %bb.0:
+; BASE-ARM-NEXT:    .pad #4
+; BASE-ARM-NEXT:    sub sp, sp, #4
+; BASE-ARM-NEXT:    strh r0, [sp, #2]
+; BASE-ARM-NEXT:    ldrh r0, [sp, #2]
+; BASE-ARM-NEXT:    vmov s0, r0
+; BASE-ARM-NEXT:    add sp, sp, #4
+; BASE-ARM-NEXT:    bx lr
+;
+; BASE-THUMB-LABEL: bf_from_int_hard:
+; BASE-THUMB:       @ %bb.0:
+; BASE-THUMB-NEXT:    .pad #4
+; BASE-THUMB-NEXT:    sub sp, #4
+; BASE-THUMB-NEXT:    strh.w r0, [sp, #2]
+; BASE-THUMB-NEXT:    ldrh.w r0, [sp, #2]
+; BASE-THUMB-NEXT:    vmov s0, r0
+; BASE-THUMB-NEXT:    add sp, #4
+; BASE-THUMB-NEXT:    bx lr
+;
+; FULLFP16-LABEL: bf_from_int_hard:
+; FULLFP16:       @ %bb.0:
+; FULLFP16-NEXT:    vmov.f16 s0, r0
+; FULLFP16-NEXT:    bx lr
+  %h = trunc i32 %w to i16
+  %f = bitcast i16 %h to bfloat
+  ret bfloat %f
+}
+
+define bfloat @test_fncall_soft(bfloat %bf, bfloat (bfloat, bfloat)* %f) {
+; BASE-ARM-LABEL: test_fncall_soft:
+; BASE-ARM:       @ %bb.0:
+; BASE-ARM-NEXT:    .save {r4, r5, r11, lr}
+; BASE-ARM-NEXT:    push {r4, r5, r11, lr}
+; BASE-ARM-NEXT:    .pad #8
+; BASE-ARM-NEXT:    sub sp, sp, #8
+; BASE-ARM-NEXT:    uxth r5, r0
+; BASE-ARM-NEXT:    mov r4, r1
+; BASE-ARM-NEXT:    mov r0, r5
+; BASE-ARM-NEXT:    mov r1, r5
+; BASE-ARM-NEXT:    blx r4
+; BASE-ARM-NEXT:    strh r0, [sp, #6]
+; BASE-ARM-NEXT:    uxth r1, r0
+; BASE-ARM-NEXT:    mov r0, r5
+; BASE-ARM-NEXT:    blx r4
+; BASE-ARM-NEXT:    ldrh r0, [sp, #6]
+; BASE-ARM-NEXT:    add sp, sp, #8
+; BASE-ARM-NEXT:    pop {r4, r5, r11, pc}
+;
+; BASE-THUMB-LABEL: test_fncall_soft:
+; BASE-THUMB:       @ %bb.0:
+; BASE-THUMB-NEXT:    .save {r4, r5, r7, lr}
+; BASE-THUMB-NEXT:    push {r4, r5, r7, lr}
+; BASE-THUMB-NEXT:    .pad #8
+; BASE-THUMB-NEXT:    sub sp, #8
+; BASE-THUMB-NEXT:    uxth r5, r0
+; BASE-THUMB-NEXT:    mov r4, r1
+; BASE-THUMB-NEXT:    mov r0, r5
+; BASE-THUMB-NEXT:    mov r1, r5
+; BASE-THUMB-NEXT:    blx r4
+; BASE-THUMB-NEXT:    uxth r1, r0
+; BASE-THUMB-NEXT:    strh.w r0, [sp, #6]
+; BASE-THUMB-NEXT:    mov r0, r5
+; BASE-THUMB-NEXT:    blx r4
+; BASE-THUMB-NEXT:    ldrh.w r0, [sp, #6]
+; BASE-THUMB-NEXT:    add sp, #8
+; BASE-THUMB-NEXT:    pop {r4, r5, r7, pc}
+;
+; FULLFP16-ARM-LABEL: test_fncall_soft:
+; FULLFP16-ARM:       @ %bb.0:
+; FULLFP16-ARM-NEXT:    .save {r4, r5, r11, lr}
+; FULLFP16-ARM-NEXT:    push {r4, r5, r11, lr}
+; FULLFP16-ARM-NEXT:    .vsave {d8}
+; FULLFP16-ARM-NEXT:    vpush {d8}
+; FULLFP16-ARM-NEXT:    vmov.f16 s0, r0
+; FULLFP16-ARM-NEXT:    mov r4, r1
+; FULLFP16-ARM-NEXT:    vmov.f16 r5, s0
+; FULLFP16-ARM-NEXT:    mov r0, r5
+; FULLFP16-ARM-NEXT:    mov r1, r5
+; FULLFP16-ARM-NEXT:    blx r4
+; FULLFP16-ARM-NEXT:    vmov.f16 s16, r0
+; FULLFP16-ARM-NEXT:    mov r0, r5
+; FULLFP16-ARM-NEXT:    vmov.f16 r1, s16
+; FULLFP16-ARM-NEXT:    blx r4
+; FULLFP16-ARM-NEXT:    vmov r0, s16
+; FULLFP16-ARM-NEXT:    vpop {d8}
+; FULLFP16-ARM-NEXT:    pop {r4, r5, r11, pc}
+;
+; FULLFP16-THUMB-LABEL: test_fncall_soft:
+; FULLFP16-THUMB:       @ %bb.0:
+; FULLFP16-THUMB-NEXT:    .save {r4, r5, r7, lr}
+; FULLFP16-THUMB-NEXT:    push {r4, r5, r7, lr}
+; FULLFP16-THUMB-NEXT:    .vsave {d8}
+; FULLFP16-THUMB-NEXT:    vpush {d8}
+; FULLFP16-THUMB-NEXT:    vmov.f16 s0, r0
+; FULLFP16-THUMB-NEXT:    mov r4, r1
+; FULLFP16-THUMB-NEXT:    vmov.f16 r5, s0
+; FULLFP16-THUMB-NEXT:    mov r0, r5
+; FULLFP16-THUMB-NEXT:    mov r1, r5
+; FULLFP16-THUMB-NEXT:    blx r4
+; FULLFP16-THUMB-NEXT:    vmov.f16 s16, r0
+; FULLFP16-THUMB-NEXT:    mov r0, r5
+; FULLFP16-THUMB-NEXT:    vmov.f16 r1, s16
+; FULLFP16-THUMB-NEXT:    blx r4
+; FULLFP16-THUMB-NEXT:    vmov r0, s16
+; FULLFP16-THUMB-NEXT:    vpop {d8}
+; FULLFP16-THUMB-NEXT:    pop {r4, r5, r7, pc}
+  %call = tail call bfloat %f(bfloat %bf, bfloat %bf)
+  %call1 = tail call bfloat %f(bfloat %bf, bfloat %call)
+  ret bfloat %call
+}
+
+define arm_aapcs_vfpcc bfloat @test_fncall_hard(bfloat %bf, bfloat (bfloat, bfloat)* %f) {
+; BASE-ARM-LABEL: test_fncall_hard:
+; BASE-ARM:       @ %bb.0:
+; BASE-ARM-NEXT:    .save {r4, lr}
+; BASE-ARM-NEXT:    push {r4, lr}
+; BASE-ARM-NEXT:    .vsave {d8}
+; BASE-ARM-NEXT:    vpush {d8}
+; BASE-ARM-NEXT:    .pad #8
+; BASE-ARM-NEXT:    sub sp, sp, #8
+; BASE-ARM-NEXT:    mov r4, r0
+; BASE-ARM-NEXT:    vmov r0, s0
+; BASE-ARM-NEXT:    uxth r0, r0
+; BASE-ARM-NEXT:    vmov s16, r0
+; BASE-ARM-NEXT:    vmov.f32 s0, s16
+; BASE-ARM-NEXT:    vmov.f32 s1, s16
+; BASE-ARM-NEXT:    blx r4
+; BASE-ARM-NEXT:    vmov r0, s0
+; BASE-ARM-NEXT:    vmov.f32 s0, s16
+; BASE-ARM-NEXT:    strh r0, [sp, #6]
+; BASE-ARM-NEXT:    uxth r0, r0
+; BASE-ARM-NEXT:    vmov s1, r0
+; BASE-ARM-NEXT:    blx r4
+; BASE-ARM-NEXT:    ldrh r0, [sp, #6]
+; BASE-ARM-NEXT:    vmov s0, r0
+; BASE-ARM-NEXT:    add sp, sp, #8
+; BASE-ARM-NEXT:    vpop {d8}
+; BASE-ARM-NEXT:    pop {r4, pc}
+;
+; BASE-THUMB-LABEL: test_fncall_hard:
+; BASE-THUMB:       @ %bb.0:
+; BASE-THUMB-NEXT:    .save {r4, lr}
+; BASE-THUMB-NEXT:    push {r4, lr}
+; BASE-THUMB-NEXT:    .vsave {d8}
+; BASE-THUMB-NEXT:    vpush {d8}
+; BASE-THUMB-NEXT:    .pad #8
+; BASE-THUMB-NEXT:    sub sp, #8
+; BASE-THUMB-NEXT:    mov r4, r0
+; BASE-THUMB-NEXT:    vmov r0, s0
+; BASE-THUMB-NEXT:    uxth r0, r0
+; BASE-THUMB-NEXT:    vmov s16, r0
+; BASE-THUMB-NEXT:    vmov.f32 s0, s16
+; BASE-THUMB-NEXT:    vmov.f32 s1, s16
+; BASE-THUMB-NEXT:    blx r4
+; BASE-THUMB-NEXT:    vmov r0, s0
+; BASE-THUMB-NEXT:    vmov.f32 s0, s16
+; BASE-THUMB-NEXT:    strh.w r0, [sp, #6]
+; BASE-THUMB-NEXT:    uxth r0, r0
+; BASE-THUMB-NEXT:    vmov s1, r0
+; BASE-THUMB-NEXT:    blx r4
+; BASE-THUMB-NEXT:    ldrh.w r0, [sp, #6]
+; BASE-THUMB-NEXT:    vmov s0, r0
+; BASE-THUMB-NEXT:    add sp, #8
+; BASE-THUMB-NEXT:    vpop {d8}
+; BASE-THUMB-NEXT:    pop {r4, pc}
+;
+; FULLFP16-LABEL: test_fncall_hard:
+; FULLFP16:       @ %bb.0:
+; FULLFP16-NEXT:    .save {r4, lr}
+; FULLFP16-NEXT:    push {r4, lr}
+; FULLFP16-NEXT:    .vsave {d8, d9}
+; FULLFP16-NEXT:    vpush {d8, d9}
+; FULLFP16-NEXT:    mov r4, r0
+; FULLFP16-NEXT:    vmov.f16 r0, s0
+; FULLFP16-NEXT:    vmov s16, r0
+; FULLFP16-NEXT:    vmov.f32 s0, s16
+; FULLFP16-NEXT:    vmov.f32 s1, s16
+; FULLFP16-NEXT:    blx r4
+; FULLFP16-NEXT:    vmov.f16 r0, s0
+; FULLFP16-NEXT:    vmov.f32 s18, s0
+; FULLFP16-NEXT:    vmov.f32 s0, s16
+; FULLFP16-NEXT:    vmov s1, r0
+; FULLFP16-NEXT:    blx r4
+; FULLFP16-NEXT:    vmov.f32 s0, s18
+; FULLFP16-NEXT:    vpop {d8, d9}
+; FULLFP16-NEXT:    pop {r4, pc}
+  %call = tail call arm_aapcs_vfpcc bfloat %f(bfloat %bf, bfloat %bf)
+  %call1 = tail call arm_aapcs_vfpcc bfloat %f(bfloat %bf, bfloat %call)
+  ret bfloat %call
+}