[SelectionDAG] remove constant folding limitations based on FP exceptions

[spatel]

Summary

We don't have any FP exception limits in the IR constant folder (apart from strict ops), so I don't think it makes sense to have them here in the DAG either. Nothing else in the backend tries to preserve those (again outside of strict ops), so I don't see how this could have ever worked for real code that cares about FP exceptions.

[efriedma]

LGTM; this matches the direction we're going, where we use different DAG nodes for functions which care about FP exceptions. (Please wait a couple days in case I'm missing something obvious.)

[andrew.w.kaylor]

Rather than get rid of this, can we update it to handle strict FP nodes correctly? Or at least put a skeleton in place to prepare to handle it correctly?

Right now I think we're throwing away the rounding mode and exception semantics arguments when we convert constrained intrinsics to strict FP nodes, but if we still had that information here we could use it to make a smart decision. For instance, suppose we had an intrinsic like this:

%t = call double @llvm.experimental.constrained.fadd.f64(double 4.3, double 6.5,
                                                         metadata !"round.upward",
                                                         metadata !"fpexcept.ignore")

If that got to the selection DAG we would be able to fold it.

[efriedma]

Yes, we could possibly constant-fold STRICT_FADD etc. using similar logic, but I'm not sure we would actually want to reuse the existing codepath in foldConstantFPMath. In any case, it'll be in the revision history if someone does end up wanting to revive the code for that.

[andrew.w.kaylor]

Why wouldn't we want to reuse this code path?

We made the decision that we are initially willing to live with poor optimization in order to get correctness for the constrained FP implementation, but the plan all along has been that after we had correct results we wanted to go back to re-enable optimizations. I'm just trying to avoid making that path harder.

What I'm asking for here is something like this:

SDValue SelectionDAG::foldConstantFPMath(unsigned Opcode, const SDLoc &DL,
                                         EVT VT, SDValue N1, SDValue N2) {
  auto *N1CFP = dyn_cast<ConstantFPSDNode>(N1.getNode());
  auto *N2CFP = dyn_cast<ConstantFPSDNode>(N2.getNode());
  bool HasFPExceptions = false;
  APFloat::roundingMode = APFloat::round_to_nearest;
  // FIXME: Add support for strict FP rounding and exception semantics.

That would produce the same effect as this patch. To support it we'd just need extra arguments to provide the rounding mode and exception semantics. I think those are probably even available not far up the stack from where this is called.

[craig.topper]

In D61331#1484982, @andrew.w.kaylor wrote:

Why wouldn't we want to reuse this code path?

We made the decision that we are initially willing to live with poor optimization in order to get correctness for the constrained FP implementation, but the plan all along has been that after we had correct results we wanted to go back to re-enable optimizations. I'm just trying to avoid making that path harder.

What I'm asking for here is something like this:

SDValue SelectionDAG::foldConstantFPMath(unsigned Opcode, const SDLoc &DL,
                                         EVT VT, SDValue N1, SDValue N2) {
  auto *N1CFP = dyn_cast<ConstantFPSDNode>(N1.getNode());
  auto *N2CFP = dyn_cast<ConstantFPSDNode>(N2.getNode());
  bool HasFPExceptions = false;
  APFloat::roundingMode = APFloat::round_to_nearest;
  // FIXME: Add support for strict FP rounding and exception semantics.

That would produce the same effect as this patch. To support it we'd just need extra arguments to provide the rounding mode and exception semantics. I think those are probably even available not far up the stack from where this is called.

You also need to pass in the input chain and probably emit a MERGE_VALUES to connect it to the expected output chain. Probably need to call this from a different getNode function with the right number of input arguments and add STRICT_FP nodes to the switch in that function to call this.

[efriedma]

We generally try to avoid keeping dead code in-tree, but I guess I'm not strongly opposed to it here.

[andrew.w.kaylor]

In D61331#1485000, @efriedma wrote:

We generally try to avoid keeping dead code in-tree, but I guess I'm not strongly opposed to it here.

I honestly don't have strong feelings about this particular change either. I'm more concerned about the way we think about strict FP handling. We've still got a lot of work to do to finish strict FP handling, and I would hate to see any code that currently has logic to handle strict exception semantics lose it. On the other hand, looking closer at the current code, I see that the logic there today isn't actually correct for strict exception semantics. So I guess if we decide to go with my suggestion then every place that's currently checking "Status != APFloat::opInvalidOp" should really be checking "Status == APFloat::opOK".

BTW, in addition to what Craig pointed out above with regard to the chain, he visited me and mentioned additional complications related to the fact that we aren't keeping the rounding mode and exception semantics in the DAG. We really need a way to fix that. I'm inclined to believe constant folding will still end up going through this function, but it'll be more work to get here than I realized.

[spatel]

In D61331#1485112, @andrew.w.kaylor wrote:

In D61331#1485000, @efriedma wrote:

We generally try to avoid keeping dead code in-tree, but I guess I'm not strongly opposed to it here.

I honestly don't have strong feelings about this particular change either. I'm more concerned about the way we think about strict FP handling. We've still got a lot of work to do to finish strict FP handling, and I would hate to see any code that currently has logic to handle strict exception semantics lose it. On the other hand, looking closer at the current code, I see that the logic there today isn't actually correct for strict exception semantics. So I guess if we decide to go with my suggestion then every place that's currently checking "Status != APFloat::opInvalidOp" should really be checking "Status == APFloat::opOK".

That would definitely be an improvement, but I'm not sure if that's enough. For example, if we intend to track denormals here, then we'd have to check the returned constant value rather than the status value?

It seemed to me that the strict and regular cases would end up sharing just the 1 line of code that computes the constant, so it might make more sense to group strict nodes together rather than interleave those cases with the regular ops.

BTW, in addition to what Craig pointed out above with regard to the chain, he visited me and mentioned additional complications related to the fact that we aren't keeping the rounding mode and exception semantics in the DAG. We really need a way to fix that. I'm inclined to believe constant folding will still end up going through this function, but it'll be more work to get here than I realized.

Yes, I think there's a lot of plumbing to do...

[cameron.mcinally]

llvm/lib/CodeGen/SelectionDAG/SelectionDAG.cpp

4819

These break statements (FADD to FREM) are unreachable.

[spatel]

Patch updated:

1. Added a TODO comment about handling strict opcodes.
2. Removed unreachable 'break' statements in the switch.

[andrew.w.kaylor]

I'm OK with this. Thanks for the TODO comment!

[mcberg2017]

LGTM, thanks Sanjay

[cameron.mcinally]

llvm/lib/CodeGen/SelectionDAG/SelectionDAG.cpp

4221

Happened to stumble across more of these exception checks. Just a heads up...

[spatel]

llvm/lib/CodeGen/SelectionDAG/SelectionDAG.cpp

4221

Thanks. Also, I just noticed that we do not constant fold the FMA intrinsic in IR. Haven't looked at the unary ops in IR yet, but I'll try to make this consistent.

llvm/include/llvm/CodeGen/TargetLowering.h

public:
   }
 
   /// Return true if inserting a scalar into a variable element of an undef
   /// vector is more efficiently handled by splatting the scalar instead.
   virtual bool shouldSplatInsEltVarIndex(EVT) const {
     return false;
   }
 
-  /// Return true if target supports floating point exceptions.
-  bool hasFloatingPointExceptions() const {
-    return HasFloatingPointExceptions;
-  }
-
   /// Return true if target always beneficiates from combining into FMA for a
   /// given value type. This must typically return false on targets where FMA
   /// takes more cycles to execute than FADD.
   virtual bool enableAggressiveFMAFusion(EVT VT) const {
     return false;
   }
 
   /// Return the ValueType of the result of SETCC operations.
void setHasExtractBitsInsn(bool hasExtractInsn = true) {
     HasExtractBitsInsn = hasExtractInsn;
   }
 
   /// Tells the code generator not to expand logic operations on comparison
   /// predicates into separate sequences that increase the amount of flow
   /// control.
   void setJumpIsExpensive(bool isExpensive = true);
 
-  /// Tells the code generator that this target supports floating point
-  /// exceptions and cares about preserving floating point exception behavior.
-  void setHasFloatingPointExceptions(bool FPExceptions = true) {
-    HasFloatingPointExceptions = FPExceptions;
-  }
-
   /// Tells the code generator which bitwidths to bypass.
   void addBypassSlowDiv(unsigned int SlowBitWidth, unsigned int FastBitWidth) {
     BypassSlowDivWidths[SlowBitWidth] = FastBitWidth;
   }
 
   /// Add the specified register class as an available regclass for the
   /// specified value type. This indicates the selector can handle values of
   /// that class natively.
private:
   /// div/rem when the operands are positive and less than 256.
   DenseMap <unsigned int, unsigned int> BypassSlowDivWidths;
 
   /// Tells the code generator that it shouldn't generate extra flow control
   /// instructions and should attempt to combine flow control instructions via
   /// predication.
   bool JumpIsExpensive;
 
-  /// Whether the target supports or cares about preserving floating point
-  /// exception behavior.
-  bool HasFloatingPointExceptions;
-
   /// This target prefers to use _setjmp to implement llvm.setjmp.
   ///
   /// Defaults to false.
   bool UseUnderscoreSetJmp;
 
   /// This target prefers to use _longjmp to implement llvm.longjmp.
   ///
   /// Defaults to false.

llvm/lib/CodeGen/SelectionDAG/SelectionDAG.cpp

case ISD::FTRUNC: {
       APFloat::opStatus fs = V.roundToIntegral(APFloat::rmTowardZero);
       if (fs == APFloat::opOK || fs == APFloat::opInexact)
         return getConstantFP(V, DL, VT);
       break;
     }
     case ISD::FFLOOR: {
       APFloat::opStatus fs = V.roundToIntegral(APFloat::rmTowardNegative);
       if (fs == APFloat::opOK || fs == APFloat::opInexact)
         return getConstantFP(V, DL, VT);
       break;
     }
     case ISD::FP_EXTEND: {
       bool ignored;
       // This can return overflow, underflow, or inexact; we don't care.
       // FIXME need to be more flexible about rounding mode.
       (void)V.convert(EVTToAPFloatSemantics(VT),
                       APFloat::rmNearestTiesToEven, &ignored);
SDValue SelectionDAG::FoldConstantVectorArithmetic(unsigned Opcode,
   NewSDValueDbgMsg(V, "New node fold constant vector: ", this);
   return V;
 }
 
 SDValue SelectionDAG::foldConstantFPMath(unsigned Opcode, const SDLoc &DL,
                                          EVT VT, SDValue N1, SDValue N2) {
   auto *N1CFP = dyn_cast<ConstantFPSDNode>(N1.getNode());
   auto *N2CFP = dyn_cast<ConstantFPSDNode>(N2.getNode());
-  bool HasFPExceptions = TLI->hasFloatingPointExceptions();
   if (N1CFP && N2CFP) {
     APFloat C1 = N1CFP->getValueAPF(), C2 = N2CFP->getValueAPF();
-    APFloat::opStatus Status;
     switch (Opcode) {
     case ISD::FADD:
-      Status = C1.add(C2, APFloat::rmNearestTiesToEven);
-      if (!HasFPExceptions || Status != APFloat::opInvalidOp)
-        return getConstantFP(C1, DL, VT);
+      C1.add(C2, APFloat::rmNearestTiesToEven);
       break;
     case ISD::FSUB:
-      Status = C1.subtract(C2, APFloat::rmNearestTiesToEven);
-      if (!HasFPExceptions || Status != APFloat::opInvalidOp)
-        return getConstantFP(C1, DL, VT);
+      C1.subtract(C2, APFloat::rmNearestTiesToEven);
       break;
     case ISD::FMUL:
-      Status = C1.multiply(C2, APFloat::rmNearestTiesToEven);
-      if (!HasFPExceptions || Status != APFloat::opInvalidOp)
-        return getConstantFP(C1, DL, VT);
+      C1.multiply(C2, APFloat::rmNearestTiesToEven);
       break;
     case ISD::FDIV:
-      Status = C1.divide(C2, APFloat::rmNearestTiesToEven);
-      if (!HasFPExceptions || Status != APFloat::opInvalidOp)
-        return getConstantFP(C1, DL, VT);
+      C1.divide(C2, APFloat::rmNearestTiesToEven);
       break;
     case ISD::FREM:
-      Status = C1.mod(C2);
-      if (!HasFPExceptions || Status != APFloat::opInvalidOp)
-        return getConstantFP(C1, DL, VT);
+      C1.mod(C2);
       break;
     case ISD::FCOPYSIGN:
       C1.copySign(C2);
       return getConstantFP(C1, DL, VT);
     default: break;
     }
   }
   if (N1CFP && Opcode == ISD::FP_ROUND) {
     APFloat C1 = N1CFP->getValueAPF();    // make copy
assert(N1.getValueType() == VT && N2.getValueType() == VT &&
            N3.getValueType() == VT && "FMA types must match!");
     ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
     ConstantFPSDNode *N2CFP = dyn_cast<ConstantFPSDNode>(N2);
     ConstantFPSDNode *N3CFP = dyn_cast<ConstantFPSDNode>(N3);
     if (N1CFP && N2CFP && N3CFP) {
       APFloat  V1 = N1CFP->getValueAPF();
       const APFloat &V2 = N2CFP->getValueAPF();
       const APFloat &V3 = N3CFP->getValueAPF();
-      APFloat::opStatus s =
-        V1.fusedMultiplyAdd(V2, V3, APFloat::rmNearestTiesToEven);
-      if (!TLI->hasFloatingPointExceptions() || s != APFloat::opInvalidOp)
-        return getConstantFP(V1, DL, VT);
     }
     break;
   }
   case ISD::BUILD_VECTOR: {
     // Attempt to simplify BUILD_VECTOR.
     SDValue Ops[] = {N1, N2, N3};
     if (SDValue V = FoldBUILD_VECTOR(DL, VT, Ops, *this))
       return V;

llvm/lib/CodeGen/TargetLoweringBase.cpp

MaxStoresPerMemsetOptSize = MaxStoresPerMemcpyOptSize =
       MaxStoresPerMemmoveOptSize = MaxLoadsPerMemcmpOptSize = 4;
   UseUnderscoreSetJmp = false;
   UseUnderscoreLongJmp = false;
   HasMultipleConditionRegisters = false;
   HasExtractBitsInsn = false;
   JumpIsExpensive = JumpIsExpensiveOverride;
   PredictableSelectIsExpensive = false;
   EnableExtLdPromotion = false;
-  HasFloatingPointExceptions = true;
   StackPointerRegisterToSaveRestore = 0;
   BooleanContents = UndefinedBooleanContent;
   BooleanFloatContents = UndefinedBooleanContent;
   BooleanVectorContents = UndefinedBooleanContent;
   SchedPreferenceInfo = Sched::ILP;
   JumpBufSize = 0;
   JumpBufAlignment = 0;
   MinFunctionAlignment = 0;

llvm/lib/Target/AMDGPU/SIISelLowering.cpp

#endif
   setTargetDAGCombine(ISD::ATOMIC_LOAD_NAND);
   setTargetDAGCombine(ISD::ATOMIC_LOAD_MIN);
   setTargetDAGCombine(ISD::ATOMIC_LOAD_MAX);
   setTargetDAGCombine(ISD::ATOMIC_LOAD_UMIN);
   setTargetDAGCombine(ISD::ATOMIC_LOAD_UMAX);
   setTargetDAGCombine(ISD::ATOMIC_LOAD_FADD);
 
   setSchedulingPreference(Sched::RegPressure);
-
-  // SI at least has hardware support for floating point exceptions, but no way
-  // of using or handling them is implemented. They are also optional in OpenCL
-  // (Section 7.3)
-  setHasFloatingPointExceptions(Subtarget->hasFPExceptions());
 }
 
 const GCNSubtarget *SITargetLowering::getSubtarget() const {
   return Subtarget;
 }
 
 //===----------------------------------------------------------------------===//
 // TargetLowering queries

llvm/lib/Target/WebAssembly/WebAssemblyISelLowering.cpp

WebAssemblyTargetLowering::WebAssemblyTargetLowering(
     const TargetMachine &TM, const WebAssemblySubtarget &STI)
     : TargetLowering(TM), Subtarget(&STI) {
   auto MVTPtr = Subtarget->hasAddr64() ? MVT::i64 : MVT::i32;
 
   // Booleans always contain 0 or 1.
   setBooleanContents(ZeroOrOneBooleanContent);
   // Except in SIMD vectors
   setBooleanVectorContents(ZeroOrNegativeOneBooleanContent);
-  // WebAssembly does not produce floating-point exceptions on normal floating
-  // point operations.
-  setHasFloatingPointExceptions(false);
   // We don't know the microarchitecture here, so just reduce register pressure.
   setSchedulingPreference(Sched::RegPressure);
   // Tell ISel that we have a stack pointer.
   setStackPointerRegisterToSaveRestore(
       Subtarget->hasAddr64() ? WebAssembly::SP64 : WebAssembly::SP32);
   // Set up the register classes.
   addRegisterClass(MVT::i32, &WebAssembly::I32RegClass);
   addRegisterClass(MVT::i64, &WebAssembly::I64RegClass);

llvm/test/CodeGen/AArch64/fp-const-fold.ll

; CHECK-NEXT:    ret
   ret double %r
 }
 
 ; frem by 0.0 --> NaN
 
 define double @constant_fold_frem_by_zero(double* %p) {
 ; CHECK-LABEL: constant_fold_frem_by_zero:
 ; CHECK:       // %bb.0:
-; CHECK-NEXT:    mov x8, #1
-; CHECK-NEXT:    fmov d1, xzr
+; CHECK-NEXT:    mov x8, #9221120237041090560
 ; CHECK-NEXT:    fmov d0, x8
-; CHECK-NEXT:    b fmod
+; CHECK-NEXT:    ret
   %r = frem double 4.940660e-324, 0.0
   ret double %r
 }
 
 ; Inf * 0.0 --> NaN
 
 define double @constant_fold_fmul_nan(double* %p) {
 ; CHECK-LABEL: constant_fold_fmul_nan:
 ; CHECK:       // %bb.0:
-; CHECK-NEXT:    mov x8, #9218868437227405312
-; CHECK-NEXT:    fmov d0, xzr
-; CHECK-NEXT:    fmov d1, x8
-; CHECK-NEXT:    fmul d0, d1, d0
+; CHECK-NEXT:    mov x8, #9221120237041090560
+; CHECK-NEXT:    fmov d0, x8
 ; CHECK-NEXT:    ret
   %r = fmul double 0x7ff0000000000000, 0.0
   ret double %r
 }
 
 ; Inf + -Inf --> NaN
 
 define double @constant_fold_fadd_nan(double* %p) {
 ; CHECK-LABEL: constant_fold_fadd_nan:
 ; CHECK:       // %bb.0:
-; CHECK-NEXT:    mov x8, #-4503599627370496
-; CHECK-NEXT:    mov x9, #9218868437227405312
+; CHECK-NEXT:    mov x8, #9221120237041090560
 ; CHECK-NEXT:    fmov d0, x8
-; CHECK-NEXT:    fmov d1, x9
-; CHECK-NEXT:    fadd d0, d1, d0
 ; CHECK-NEXT:    ret
   %r = fadd double 0x7ff0000000000000, 0xfff0000000000000
   ret double %r
 }
 
 ; Inf - Inf --> NaN
 
 define double @constant_fold_fsub_nan(double* %p) {
 ; CHECK-LABEL: constant_fold_fsub_nan:
 ; CHECK:       // %bb.0:
-; CHECK-NEXT:    mov x8, #9218868437227405312
+; CHECK-NEXT:    mov x8, #9221120237041090560
 ; CHECK-NEXT:    fmov d0, x8
-; CHECK-NEXT:    fsub d0, d0, d0
 ; CHECK-NEXT:    ret
   %r = fsub double 0x7ff0000000000000, 0x7ff0000000000000
   ret double %r
 }
 
 ; Inf * 0.0 + ? --> NaN
 
 define double @constant_fold_fma_nan(double* %p) {
 ; CHECK-LABEL: constant_fold_fma_nan:
 ; CHECK:       // %bb.0:
-; CHECK-NEXT:    mov x8, #4631107791820423168
-; CHECK-NEXT:    mov x9, #9218868437227405312
-; CHECK-NEXT:    fmov d0, xzr
-; CHECK-NEXT:    fmov d1, x8
-; CHECK-NEXT:    fmov d2, x9
-; CHECK-NEXT:    fmadd d0, d2, d0, d1
+; CHECK-NEXT:    mov x8, #9221120237041090560
+; CHECK-NEXT:    fmov d0, x8
 ; CHECK-NEXT:    ret
   %r =  call double @llvm.fma.f64(double 0x7ff0000000000000, double 0.0, double 42.0)
   ret double %r
 }
 
 declare double @llvm.fma.f64(double, double, double)