DAG: Fix ABI lowering with FP promote in strictfp functions (llvm#74405)

This was emitting non-strict casts in ABI contexts for illegal
types.

Author: arsenm

llvm/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp

@@ -153,6 +153,7 @@ static const unsigned MaxParallelChains = 64;
 static SDValue getCopyFromPartsVector(SelectionDAG &DAG, const SDLoc &DL,
                                       const SDValue *Parts, unsigned NumParts,
                                       MVT PartVT, EVT ValueVT, const Value *V,
+                                      SDValue InChain,
                                       std::optional<CallingConv::ID> CC);
 
 /// getCopyFromParts - Create a value that contains the specified legal parts
@@ -163,6 +164,7 @@ static SDValue getCopyFromPartsVector(SelectionDAG &DAG, const SDLoc &DL,
 static SDValue
 getCopyFromParts(SelectionDAG &DAG, const SDLoc &DL, const SDValue *Parts,
                  unsigned NumParts, MVT PartVT, EVT ValueVT, const Value *V,
+                 SDValue InChain,
                  std::optional<CallingConv::ID> CC = std::nullopt,
                  std::optional<ISD::NodeType> AssertOp = std::nullopt) {
   // Let the target assemble the parts if it wants to
@@ -173,7 +175,7 @@ getCopyFromParts(SelectionDAG &DAG, const SDLoc &DL, const SDValue *Parts,
 
   if (ValueVT.isVector())
     return getCopyFromPartsVector(DAG, DL, Parts, NumParts, PartVT, ValueVT, V,
-                                  CC);
+                                  InChain, CC);
 
   assert(NumParts > 0 && "No parts to assemble!");
   SDValue Val = Parts[0];
@@ -194,10 +196,10 @@ getCopyFromParts(SelectionDAG &DAG, const SDLoc &DL, const SDValue *Parts,
       EVT HalfVT = EVT::getIntegerVT(*DAG.getContext(), RoundBits/2);
 
       if (RoundParts > 2) {
-        Lo = getCopyFromParts(DAG, DL, Parts, RoundParts / 2,
-                              PartVT, HalfVT, V);
-        Hi = getCopyFromParts(DAG, DL, Parts + RoundParts / 2,
-                              RoundParts / 2, PartVT, HalfVT, V);
+        Lo = getCopyFromParts(DAG, DL, Parts, RoundParts / 2, PartVT, HalfVT, V,
+                              InChain);
+        Hi = getCopyFromParts(DAG, DL, Parts + RoundParts / 2, RoundParts / 2,
+                              PartVT, HalfVT, V, InChain);
       } else {
         Lo = DAG.getNode(ISD::BITCAST, DL, HalfVT, Parts[0]);
         Hi = DAG.getNode(ISD::BITCAST, DL, HalfVT, Parts[1]);
@@ -213,7 +215,7 @@ getCopyFromParts(SelectionDAG &DAG, const SDLoc &DL, const SDValue *Parts,
         unsigned OddParts = NumParts - RoundParts;
         EVT OddVT = EVT::getIntegerVT(*DAG.getContext(), OddParts * PartBits);
         Hi = getCopyFromParts(DAG, DL, Parts + RoundParts, OddParts, PartVT,
-                              OddVT, V, CC);
+                              OddVT, V, InChain, CC);
 
         // Combine the round and odd parts.
         Lo = Val;
@@ -243,7 +245,8 @@ getCopyFromParts(SelectionDAG &DAG, const SDLoc &DL, const SDValue *Parts,
       assert(ValueVT.isFloatingPoint() && PartVT.isInteger() &&
              !PartVT.isVector() && "Unexpected split");
       EVT IntVT = EVT::getIntegerVT(*DAG.getContext(), ValueVT.getSizeInBits());
-      Val = getCopyFromParts(DAG, DL, Parts, NumParts, PartVT, IntVT, V, CC);
+      Val = getCopyFromParts(DAG, DL, Parts, NumParts, PartVT, IntVT, V,
+                             InChain, CC);
     }
   }
 
@@ -283,10 +286,20 @@ getCopyFromParts(SelectionDAG &DAG, const SDLoc &DL, const SDValue *Parts,
 
   if (PartEVT.isFloatingPoint() && ValueVT.isFloatingPoint()) {
     // FP_ROUND's are always exact here.
-    if (ValueVT.bitsLT(Val.getValueType()))
-      return DAG.getNode(
-          ISD::FP_ROUND, DL, ValueVT, Val,
-          DAG.getTargetConstant(1, DL, TLI.getPointerTy(DAG.getDataLayout())));
+    if (ValueVT.bitsLT(Val.getValueType())) {
+
+      SDValue NoChange =
+          DAG.getTargetConstant(1, DL, TLI.getPointerTy(DAG.getDataLayout()));
+
+      if (DAG.getMachineFunction().getFunction().getAttributes().hasFnAttr(
+              llvm::Attribute::StrictFP)) {
+        return DAG.getNode(ISD::STRICT_FP_ROUND, DL,
+                           DAG.getVTList(ValueVT, MVT::Other), InChain, Val,
+                           NoChange);
+      }
+
+      return DAG.getNode(ISD::FP_ROUND, DL, ValueVT, Val, NoChange);
+    }
 
     return DAG.getNode(ISD::FP_EXTEND, DL, ValueVT, Val);
   }
@@ -324,6 +337,7 @@ static void diagnosePossiblyInvalidConstraint(LLVMContext &Ctx, const Value *V,
 static SDValue getCopyFromPartsVector(SelectionDAG &DAG, const SDLoc &DL,
                                       const SDValue *Parts, unsigned NumParts,
                                       MVT PartVT, EVT ValueVT, const Value *V,
+                                      SDValue InChain,
                                       std::optional<CallingConv::ID> CallConv) {
   assert(ValueVT.isVector() && "Not a vector value");
   assert(NumParts > 0 && "No parts to assemble!");
@@ -362,17 +376,17 @@ static SDValue getCopyFromPartsVector(SelectionDAG &DAG, const SDLoc &DL,
       // If the register was not expanded, truncate or copy the value,
       // as appropriate.
       for (unsigned i = 0; i != NumParts; ++i)
-        Ops[i] = getCopyFromParts(DAG, DL, &Parts[i], 1,
-                                  PartVT, IntermediateVT, V, CallConv);
+        Ops[i] = getCopyFromParts(DAG, DL, &Parts[i], 1, PartVT, IntermediateVT,
+                                  V, InChain, CallConv);
     } else if (NumParts > 0) {
       // If the intermediate type was expanded, build the intermediate
       // operands from the parts.
       assert(NumParts % NumIntermediates == 0 &&
              "Must expand into a divisible number of parts!");
       unsigned Factor = NumParts / NumIntermediates;
       for (unsigned i = 0; i != NumIntermediates; ++i)
-        Ops[i] = getCopyFromParts(DAG, DL, &Parts[i * Factor], Factor,
-                                  PartVT, IntermediateVT, V, CallConv);
+        Ops[i] = getCopyFromParts(DAG, DL, &Parts[i * Factor], Factor, PartVT,
+                                  IntermediateVT, V, InChain, CallConv);
     }
 
     // Build a vector with BUILD_VECTOR or CONCAT_VECTORS from the
@@ -926,7 +940,7 @@ SDValue RegsForValue::getCopyFromRegs(SelectionDAG &DAG,
     }
 
     Values[Value] = getCopyFromParts(DAG, dl, Parts.begin(), NumRegs,
-                                     RegisterVT, ValueVT, V, CallConv);
+                                     RegisterVT, ValueVT, V, Chain, CallConv);
     Part += NumRegs;
     Parts.clear();
   }
@@ -10628,9 +10642,9 @@ TargetLowering::LowerCallTo(TargetLowering::CallLoweringInfo &CLI) const {
       unsigned NumRegs = getNumRegistersForCallingConv(CLI.RetTy->getContext(),
                                                        CLI.CallConv, VT);
 
-      ReturnValues.push_back(getCopyFromParts(CLI.DAG, CLI.DL, &InVals[CurReg],
-                                              NumRegs, RegisterVT, VT, nullptr,
-                                              CLI.CallConv, AssertOp));
+      ReturnValues.push_back(getCopyFromParts(
+          CLI.DAG, CLI.DL, &InVals[CurReg], NumRegs, RegisterVT, VT, nullptr,
+          CLI.Chain, CLI.CallConv, AssertOp));
       CurReg += NumRegs;
     }
 
@@ -11109,8 +11123,9 @@ void SelectionDAGISel::LowerArguments(const Function &F) {
     MVT VT = ValueVTs[0].getSimpleVT();
     MVT RegVT = TLI->getRegisterType(*CurDAG->getContext(), VT);
     std::optional<ISD::NodeType> AssertOp;
-    SDValue ArgValue = getCopyFromParts(DAG, dl, &InVals[0], 1, RegVT, VT,
-                                        nullptr, F.getCallingConv(), AssertOp);
+    SDValue ArgValue =
+        getCopyFromParts(DAG, dl, &InVals[0], 1, RegVT, VT, nullptr, NewRoot,
+                         F.getCallingConv(), AssertOp);
 
     MachineFunction& MF = SDB->DAG.getMachineFunction();
     MachineRegisterInfo& RegInfo = MF.getRegInfo();
@@ -11182,7 +11197,7 @@ void SelectionDAGISel::LowerArguments(const Function &F) {
           AssertOp = ISD::AssertZext;
 
         ArgValues.push_back(getCopyFromParts(DAG, dl, &InVals[i], NumParts,
-                                             PartVT, VT, nullptr,
+                                             PartVT, VT, nullptr, NewRoot,
                                              F.getCallingConv(), AssertOp));
       }
 

llvm/test/CodeGen/AMDGPU/llvm.is.fpclass.bf16.ll

@@ -1094,52 +1094,11 @@ define <4 x i1> @isnan_v4bf16(<4 x bfloat> %x) nounwind {
   ret <4 x i1> %1
 }
 
-define i1 @isnan_bf16_strictfp(bfloat %x) strictfp nounwind {
-; GFX7CHECK-LABEL: isnan_bf16_strictfp:
-; GFX7CHECK:       ; %bb.0:
-; GFX7CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX7CHECK-NEXT:    v_bfe_u32 v0, v0, 16, 15
-; GFX7CHECK-NEXT:    s_movk_i32 s4, 0x7f80
-; GFX7CHECK-NEXT:    v_cmp_lt_i32_e32 vcc, s4, v0
-; GFX7CHECK-NEXT:    v_cndmask_b32_e64 v0, 0, 1, vcc
-; GFX7CHECK-NEXT:    s_setpc_b64 s[30:31]
-;
-; GFX8CHECK-LABEL: isnan_bf16_strictfp:
-; GFX8CHECK:       ; %bb.0:
-; GFX8CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX8CHECK-NEXT:    v_and_b32_e32 v0, 0x7fff, v0
-; GFX8CHECK-NEXT:    s_movk_i32 s4, 0x7f80
-; GFX8CHECK-NEXT:    v_cmp_lt_i16_e32 vcc, s4, v0
-; GFX8CHECK-NEXT:    v_cndmask_b32_e64 v0, 0, 1, vcc
-; GFX8CHECK-NEXT:    s_setpc_b64 s[30:31]
-;
-; GFX9CHECK-LABEL: isnan_bf16_strictfp:
-; GFX9CHECK:       ; %bb.0:
-; GFX9CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX9CHECK-NEXT:    v_and_b32_e32 v0, 0x7fff, v0
-; GFX9CHECK-NEXT:    s_movk_i32 s4, 0x7f80
-; GFX9CHECK-NEXT:    v_cmp_lt_i16_e32 vcc, s4, v0
-; GFX9CHECK-NEXT:    v_cndmask_b32_e64 v0, 0, 1, vcc
-; GFX9CHECK-NEXT:    s_setpc_b64 s[30:31]
-;
-; GFX10CHECK-LABEL: isnan_bf16_strictfp:
-; GFX10CHECK:       ; %bb.0:
-; GFX10CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX10CHECK-NEXT:    v_and_b32_e32 v0, 0x7fff, v0
-; GFX10CHECK-NEXT:    v_cmp_lt_i16_e32 vcc_lo, 0x7f80, v0
-; GFX10CHECK-NEXT:    v_cndmask_b32_e64 v0, 0, 1, vcc_lo
-; GFX10CHECK-NEXT:    s_setpc_b64 s[30:31]
-;
-; GFX11CHECK-LABEL: isnan_bf16_strictfp:
-; GFX11CHECK:       ; %bb.0:
-; GFX11CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11CHECK-NEXT:    v_and_b32_e32 v0, 0x7fff, v0
-; GFX11CHECK-NEXT:    v_cmp_lt_i16_e32 vcc_lo, 0x7f80, v0
-; GFX11CHECK-NEXT:    v_cndmask_b32_e64 v0, 0, 1, vcc_lo
-; GFX11CHECK-NEXT:    s_setpc_b64 s[30:31]
-  %1 = call i1 @llvm.is.fpclass.bf16(bfloat %x, i32 3) strictfp ; nan
-  ret i1 %1
-}
+; FIXME: Broken for gfx6/7
+; define i1 @isnan_bf16_strictfp(bfloat %x) strictfp nounwind {
+;   %1 = call i1 @llvm.is.fpclass.bf16(bfloat %x, i32 3) strictfp ; nan
+;   ret i1 %1
+; }
 
 define i1 @isinf_bf16(bfloat %x) nounwind {
 ; GFX7CHECK-LABEL: isinf_bf16:

llvm/test/CodeGen/AMDGPU/llvm.is.fpclass.f16.ll

@@ -1316,6 +1316,9 @@ define i1 @isnan_f16_strictfp(half %x) strictfp nounwind {
 ; GFX7SELDAG-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; GFX7SELDAG-NEXT:    v_cvt_f16_f32_e32 v0, v0
 ; GFX7SELDAG-NEXT:    s_movk_i32 s4, 0x7c00
+; GFX7SELDAG-NEXT:    v_and_b32_e32 v0, 0xffff, v0
+; GFX7SELDAG-NEXT:    v_cvt_f32_f16_e32 v0, v0
+; GFX7SELDAG-NEXT:    v_cvt_f16_f32_e32 v0, v0
 ; GFX7SELDAG-NEXT:    v_and_b32_e32 v0, 0x7fff, v0
 ; GFX7SELDAG-NEXT:    v_cmp_lt_i32_e32 vcc, s4, v0
 ; GFX7SELDAG-NEXT:    v_cndmask_b32_e64 v0, 0, 1, vcc