LegalizeDAG: Implement expansion of f16 = FP_TO_FP16 f64

[arsenm]

LGTM

[• tstellarAMD]

Move the expansion into the AMDGPU backend.

I wanted to implement this as a target independent expansion, however when
targets say they want to expand FP_TO_FP16 what they actually want is
the unsafe math expansion when possible and expansion to a libcall in all
other cases.

The only way to make this work as a target independent would be to add logic
to target's TargetLowering construction to mark theses nodes as Expand when
LegalizeDAG can use the unsafe expansion and mark them as LibCall when it
cannot. I think this would be possible, but I think it would be too fragile
and complex as it would require targets to keep their expansion logic up
to date with the code in LegalizeDAG.

[arsenm]

LGTM with the test fixed

test/CodeGen/AMDGPU/trunc-store-f64-to-f16.ll

1

Can you remove the -mcpu and add -verify-machineinstrs, also the FileCheck part is missing

[jvesely]

llvm/trunk/test/CodeGen/AMDGPU/fptrunc.ll

4 ↗

(On Diff #76580)

GCN-UNSAFE is never checked. did you mean GCN-FAST (or use GCN-UNSAFE in the tests)?

17 ↗

(On Diff #76580)

Is there a reason to use i16 instead of half? can the half test be moved to this file?

lib/Target/AMDGPU/AMDGPUISelLowering.h

protected:
   SDValue LowerCTLZ(SDValue Op, SelectionDAG &DAG) const;
 
   SDValue LowerINT_TO_FP32(SDValue Op, SelectionDAG &DAG, bool Signed) const;
   SDValue LowerINT_TO_FP64(SDValue Op, SelectionDAG &DAG, bool Signed) const;
   SDValue LowerUINT_TO_FP(SDValue Op, SelectionDAG &DAG) const;
   SDValue LowerSINT_TO_FP(SDValue Op, SelectionDAG &DAG) const;
 
   SDValue LowerFP64_TO_INT(SDValue Op, SelectionDAG &DAG, bool Signed) const;
+  SDValue LowerFP_TO_FP16(SDValue Op, SelectionDAG &DAG) const;
   SDValue LowerFP_TO_UINT(SDValue Op, SelectionDAG &DAG) const;
   SDValue LowerFP_TO_SINT(SDValue Op, SelectionDAG &DAG) const;
 
   SDValue LowerSIGN_EXTEND_INREG(SDValue Op, SelectionDAG &DAG) const;
 
 protected:
   bool shouldCombineMemoryType(EVT VT) const;
   SDValue performLoadCombine(SDNode *N, DAGCombinerInfo &DCI) const;

lib/Target/AMDGPU/AMDGPUISelLowering.cpp

AMDGPUTargetLowering::AMDGPUTargetLowering(const TargetMachine &TM,
 
   if (!Subtarget->hasBFI()) {
     // fcopysign can be done in a single instruction with BFI.
     setOperationAction(ISD::FCOPYSIGN, MVT::f32, Expand);
     setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
   }
 
   setOperationAction(ISD::FP16_TO_FP, MVT::f64, Expand);
+  setOperationAction(ISD::FP_TO_FP16, MVT::f64, Custom);
 
   const MVT ScalarIntVTs[] = { MVT::i32, MVT::i64 };
   for (MVT VT : ScalarIntVTs) {
     // These should use [SU]DIVREM, so set them to expand
     setOperationAction(ISD::SDIV, VT, Expand);
     setOperationAction(ISD::UDIV, VT, Expand);
     setOperationAction(ISD::SREM, VT, Expand);
     setOperationAction(ISD::UREM, VT, Expand);
SDValue AMDGPUTargetLowering::LowerOperation(SDValue Op,
   case ISD::FCEIL: return LowerFCEIL(Op, DAG);
   case ISD::FTRUNC: return LowerFTRUNC(Op, DAG);
   case ISD::FRINT: return LowerFRINT(Op, DAG);
   case ISD::FNEARBYINT: return LowerFNEARBYINT(Op, DAG);
   case ISD::FROUND: return LowerFROUND(Op, DAG);
   case ISD::FFLOOR: return LowerFFLOOR(Op, DAG);
   case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
   case ISD::UINT_TO_FP: return LowerUINT_TO_FP(Op, DAG);
+  case ISD::FP_TO_FP16: return LowerFP_TO_FP16(Op, DAG);
   case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
   case ISD::FP_TO_UINT: return LowerFP_TO_UINT(Op, DAG);
   case ISD::CTLZ:
   case ISD::CTLZ_ZERO_UNDEF:
     return LowerCTLZ(Op, DAG);
   case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
   }
   return Op;
SDValue Hi = DAG.getNode(Signed ? ISD::FP_TO_SINT : ISD::FP_TO_UINT, SL,
                            MVT::i32, FloorMul);
   SDValue Lo = DAG.getNode(ISD::FP_TO_UINT, SL, MVT::i32, Fma);
 
   SDValue Result = DAG.getBuildVector(MVT::v2i32, SL, {Lo, Hi});
 
   return DAG.getNode(ISD::BITCAST, SL, MVT::i64, Result);
 }
 
+SDValue AMDGPUTargetLowering::LowerFP_TO_FP16(SDValue Op, SelectionDAG &DAG) const {
+
+  if (getTargetMachine().Options.UnsafeFPMath) {
+    // There is a generic expand for FP_TO_FP16 with unsafe fast math.
+    return SDValue();
+  }
+
+  SDLoc DL(Op);
+  SDValue N0 = Op.getOperand(0);
+  MVT SVT = N0.getSimpleValueType();
+  assert(SVT == MVT::f64);
+
+  // f64 -> f16 conversion using round-to-nearest-even rounding mode.
+  const unsigned ExpMask = 0x7ff;
+  const unsigned ExpBiasf64 = 1023;
+  const unsigned ExpBiasf16 = 15;
+  SDValue Zero = DAG.getConstant(0, DL, MVT::i32);
+  SDValue One = DAG.getConstant(1, DL, MVT::i32);
+  SDValue U = DAG.getNode(ISD::BITCAST, DL, MVT::i64, N0);
+  SDValue UH = DAG.getNode(ISD::SRL, DL, MVT::i64, U,
+                           DAG.getConstant(32, DL, MVT::i64));
+  UH = DAG.getZExtOrTrunc(UH, DL, MVT::i32);
+  U = DAG.getZExtOrTrunc(U, DL, MVT::i32);
+  SDValue E = DAG.getNode(ISD::SRL, DL, MVT::i32, UH,
+                          DAG.getConstant(20, DL, MVT::i64));
+  E = DAG.getNode(ISD::AND, DL, MVT::i32, E,
+                  DAG.getConstant(ExpMask, DL, MVT::i32));
+  // Subtract the fp64 exponent bias (1023) to get the real exponent and
+  // add the f16 bias (15) to get the biased exponent for the f16 format.
+  E = DAG.getNode(ISD::ADD, DL, MVT::i32, E,
+                  DAG.getConstant(-ExpBiasf64 + ExpBiasf16, DL, MVT::i32));
+
+  SDValue M = DAG.getNode(ISD::SRL, DL, MVT::i32, UH,
+                          DAG.getConstant(8, DL, MVT::i32));
+  M = DAG.getNode(ISD::AND, DL, MVT::i32, M,
+                  DAG.getConstant(0xffe, DL, MVT::i32));
+
+  SDValue MaskedSig = DAG.getNode(ISD::AND, DL, MVT::i32, UH,
+                                  DAG.getConstant(0x1ff, DL, MVT::i32));
+  MaskedSig = DAG.getNode(ISD::OR, DL, MVT::i32, MaskedSig, U);
+
+  SDValue Lo40Set = DAG.getSelectCC(DL, MaskedSig, Zero, Zero, One, ISD::SETEQ);
+  M = DAG.getNode(ISD::OR, DL, MVT::i32, M, Lo40Set);
+
+  // (M != 0 ? 0x0200 : 0) | 0x7c00;
+  SDValue I = DAG.getNode(ISD::OR, DL, MVT::i32,
+      DAG.getSelectCC(DL, M, Zero, DAG.getConstant(0x0200, DL, MVT::i32),
+                      Zero, ISD::SETNE), DAG.getConstant(0x7c00, DL, MVT::i32));
+
+  // N = M | (E << 12);
+  SDValue N = DAG.getNode(ISD::OR, DL, MVT::i32, M,
+      DAG.getNode(ISD::SHL, DL, MVT::i32, E,
+                  DAG.getConstant(12, DL, MVT::i32)));
+
+  // B = clamp(1-E, 0, 13);
+  SDValue OneSubExp = DAG.getNode(ISD::SUB, DL, MVT::i32,
+                                  One, E);
+  SDValue B = DAG.getNode(ISD::SMAX, DL, MVT::i32, OneSubExp, Zero);
+  B = DAG.getNode(ISD::SMIN, DL, MVT::i32, B,
+                  DAG.getConstant(13, DL, MVT::i32));
+
+  SDValue SigSetHigh = DAG.getNode(ISD::OR, DL, MVT::i32, M,
+                                   DAG.getConstant(0x1000, DL, MVT::i32));
+
+  SDValue D = DAG.getNode(ISD::SRL, DL, MVT::i32, SigSetHigh, B);
+  SDValue D0 = DAG.getNode(ISD::SHL, DL, MVT::i32, D, B);
+  SDValue D1 = DAG.getSelectCC(DL, D0, SigSetHigh, One, Zero, ISD::SETNE);
+  D = DAG.getNode(ISD::OR, DL, MVT::i32, D, D1);
+
+  SDValue V = DAG.getSelectCC(DL, E, One, D, N, ISD::SETLT);
+  SDValue VLow3 = DAG.getNode(ISD::AND, DL, MVT::i32, V,
+                              DAG.getConstant(0x7, DL, MVT::i32));
+  V = DAG.getNode(ISD::SRL, DL, MVT::i32, V,
+                  DAG.getConstant(2, DL, MVT::i32));
+  SDValue V0 = DAG.getSelectCC(DL, VLow3, DAG.getConstant(3, DL, MVT::i32),
+                               One, Zero, ISD::SETEQ);
+  SDValue V1 = DAG.getSelectCC(DL, VLow3, DAG.getConstant(5, DL, MVT::i32),
+                               One, Zero, ISD::SETGT);
+  V1 = DAG.getNode(ISD::OR, DL, MVT::i32, V0, V1);
+  V = DAG.getNode(ISD::ADD, DL, MVT::i32, V, V1);
+
+  V = DAG.getSelectCC(DL, E, DAG.getConstant(30, DL, MVT::i32),
+                      DAG.getConstant(0x7c00, DL, MVT::i32), V, ISD::SETGT);
+  V = DAG.getSelectCC(DL, E, DAG.getConstant(1039, DL, MVT::i32),
+                      I, V, ISD::SETEQ);
+
+  // Extract the sign bit.
+  SDValue Sign = DAG.getNode(ISD::SRL, DL, MVT::i32, UH,
+                            DAG.getConstant(16, DL, MVT::i32));
+  Sign = DAG.getNode(ISD::AND, DL, MVT::i32, Sign,
+                     DAG.getConstant(0x8000, DL, MVT::i32));
+
+  V = DAG.getNode(ISD::OR, DL, MVT::i32, Sign, V);
+  return DAG.getZExtOrTrunc(V, DL, Op.getValueType());
+}
+
 SDValue AMDGPUTargetLowering::LowerFP_TO_SINT(SDValue Op,
                                               SelectionDAG &DAG) const {
   SDValue Src = Op.getOperand(0);
 
   if (Op.getValueType() == MVT::i64 && Src.getValueType() == MVT::f64)
     return LowerFP64_TO_INT(Op, DAG, true);
 
   return SDValue();

test/CodeGen/AMDGPU/fptrunc.ll

-; RUN: llc -march=amdgcn -mcpu=SI -verify-machineinstrs < %s | FileCheck -check-prefix=SI -check-prefix=FUNC %s
-; RUN: llc -march=amdgcn -mcpu=tonga -verify-machineinstrs < %s | FileCheck -check-prefix=SI -check-prefix=FUNC %s
+; RUN: llc -march=amdgcn -mcpu=SI -verify-machineinstrs < %s | FileCheck -check-prefix=GCN %s
+; RUN: llc -march=amdgcn -mcpu=tonga -verify-machineinstrs < %s | FileCheck -check-prefix=GCN %s
+; RUN: llc -march=amdgcn -mcpu=tonga -enable-unsafe-fp-math -verify-machineinstrs < %s | FileCheck -check-prefixes=GCN,GCN-UNSAFE %s
 
 ; FUNC-LABEL: {{^}}fptrunc_f64_to_f32:
-; SI: v_cvt_f32_f64_e32 {{v[0-9]+}}, {{s\[[0-9]+:[0-9]+\]}}
+; GCN: v_cvt_f32_f64_e32 {{v[0-9]+}}, {{s\[[0-9]+:[0-9]+\]}}
 define void @fptrunc_f64_to_f32(float addrspace(1)* %out, double %in) {
   %result = fptrunc double %in to float
   store float %result, float addrspace(1)* %out
   ret void
 }
 
+; FUNC-LABEL: {{^}}fptrunc_f64_to_f16:
+; GCN-NOT: v_cvt
+; GCN-FAST: v_cvt_f32_f64_e32 [[F32:v[0-9]+]]
+; GCN-FAST: v_cvt_f16_f32_e32 v[0-9]+, [[F32]]
+define void @fptrunc_f64_to_f16(i16 addrspace(1)* %out, double %in) {
+  %result = fptrunc double %in to half
+  %result_i16 = bitcast half %result to i16
+  store i16 %result_i16, i16 addrspace(1)* %out
+  ret void
+}
+
 ; FUNC-LABEL: {{^}}fptrunc_v2f64_to_v2f32:
-; SI: v_cvt_f32_f64_e32
-; SI: v_cvt_f32_f64_e32
+; GCN: v_cvt_f32_f64_e32
+; GCN: v_cvt_f32_f64_e32
 define void @fptrunc_v2f64_to_v2f32(<2 x float> addrspace(1)* %out, <2 x double> %in) {
   %result = fptrunc <2 x double> %in to <2 x float>
   store <2 x float> %result, <2 x float> addrspace(1)* %out
   ret void
 }
 
 ; FUNC-LABEL: {{^}}fptrunc_v4f64_to_v4f32:
-; SI: v_cvt_f32_f64_e32
-; SI: v_cvt_f32_f64_e32
-; SI: v_cvt_f32_f64_e32
-; SI: v_cvt_f32_f64_e32
+; GCN: v_cvt_f32_f64_e32
+; GCN: v_cvt_f32_f64_e32
+; GCN: v_cvt_f32_f64_e32
+; GCN: v_cvt_f32_f64_e32
 define void @fptrunc_v4f64_to_v4f32(<4 x float> addrspace(1)* %out, <4 x double> %in) {
   %result = fptrunc <4 x double> %in to <4 x float>
   store <4 x float> %result, <4 x float> addrspace(1)* %out
   ret void
 }
 
 ; FUNC-LABEL: {{^}}fptrunc_v8f64_to_v8f32:
-; SI: v_cvt_f32_f64_e32
-; SI: v_cvt_f32_f64_e32
-; SI: v_cvt_f32_f64_e32
-; SI: v_cvt_f32_f64_e32
-; SI: v_cvt_f32_f64_e32
-; SI: v_cvt_f32_f64_e32
-; SI: v_cvt_f32_f64_e32
-; SI: v_cvt_f32_f64_e32
+; GCN: v_cvt_f32_f64_e32
+; GCN: v_cvt_f32_f64_e32
+; GCN: v_cvt_f32_f64_e32
+; GCN: v_cvt_f32_f64_e32
+; GCN: v_cvt_f32_f64_e32
+; GCN: v_cvt_f32_f64_e32
+; GCN: v_cvt_f32_f64_e32
+; GCN: v_cvt_f32_f64_e32
 define void @fptrunc_v8f64_to_v8f32(<8 x float> addrspace(1)* %out, <8 x double> %in) {
   %result = fptrunc <8 x double> %in to <8 x float>
   store <8 x float> %result, <8 x float> addrspace(1)* %out
   ret void
 }

test/CodeGen/AMDGPU/trunc-store-f64-to-f16.ll

-; XFAIL: *
 ; RUN: llc -march=amdgcn -mcpu=SI < %s
 
 ; GCN-LABEL: {{^}}global_truncstore_f64_to_f16:
 ; GCN: s_endpgm
 define void @global_truncstore_f64_to_f16(half addrspace(1)* %out, double addrspace(1)* %in) #0 {
   %val = load double, double addrspace(1)* %in
   %cvt = fptrunc double %val to half
   store half %cvt, half addrspace(1)* %out
   ret void