[PATCH] D99035: GlobalISel: Add utility function to constant fold FP ops

[Matt Arsenault via Phabricator via llvm-commits]

arsenm created this revision.
arsenm added reviewers: paquette, aemerson, bogner, dsanders.
Herald added subscribers: hiraditya, rovka.
arsenm requested review of this revision.
Herald added a subscriber: wdng.
Herald added a project: LLVM.

https://reviews.llvm.org/D99035

Files:
  llvm/include/llvm/CodeGen/GlobalISel/Utils.h
  llvm/lib/CodeGen/GlobalISel/Utils.cpp


Index: llvm/lib/CodeGen/GlobalISel/Utils.cpp
===================================================================
--- llvm/lib/CodeGen/GlobalISel/Utils.cpp
+++ llvm/lib/CodeGen/GlobalISel/Utils.cpp
@@ -490,6 +490,60 @@
   return None;
 }
 
+Optional<APFloat> llvm::ConstantFoldFPBinOp(unsigned Opcode, const Register Op1,
+                                            const Register Op2,
+                                            const MachineRegisterInfo &MRI) {
+  const ConstantFP *Op2Cst = getConstantFPVRegVal(Op2, MRI);
+  if (!Op2Cst)
+    return None;
+
+  const ConstantFP *Op1Cst = getConstantFPVRegVal(Op1, MRI);
+  if (!Op1Cst)
+    return None;
+
+  APFloat C1 = Op1Cst->getValueAPF();
+  const APFloat &C2 = Op2Cst->getValueAPF();
+  switch (Opcode) {
+  case TargetOpcode::G_FADD:
+    C1.add(C2, APFloat::rmNearestTiesToEven);
+    return C1;
+  case TargetOpcode::G_FSUB:
+    C1.subtract(C2, APFloat::rmNearestTiesToEven);
+    return C1;
+  case TargetOpcode::G_FMUL:
+    C1.multiply(C2, APFloat::rmNearestTiesToEven);
+    return C1;
+  case TargetOpcode::G_FDIV:
+    C1.divide(C2, APFloat::rmNearestTiesToEven);
+    return C1;
+  case TargetOpcode::G_FREM:
+    C1.mod(C2);
+    return C1;
+  case TargetOpcode::G_FCOPYSIGN:
+    C1.copySign(C2);
+    return C1;
+  case TargetOpcode::G_FMINNUM:
+    return minnum(C1, C2);
+  case TargetOpcode::G_FMAXNUM:
+    return maxnum(C1, C2);
+  case TargetOpcode::G_FMINIMUM:
+    return minimum(C1, C2);
+  case TargetOpcode::G_FMAXIMUM:
+    return maximum(C1, C2);
+  case TargetOpcode::G_FMINNUM_IEEE:
+  case TargetOpcode::G_FMAXNUM_IEEE:
+    // FIXME: These operations were unfortunately named. fminnum/fmaxnum do not
+    // follow the IEEE behavior for signaling nans and follow libm's fmin/fmax,
+    // and currently there isn't a nice wrapper in APFloat for the version with
+    // correct snan handling.
+    break;
+  default:
+    break;
+  }
+
+  return None;
+}
+
 bool llvm::isKnownNeverNaN(Register Val, const MachineRegisterInfo &MRI,
                            bool SNaN) {
   const MachineInstr *DefMI = MRI.getVRegDef(Val);
Index: llvm/include/llvm/CodeGen/GlobalISel/Utils.h
===================================================================
--- llvm/include/llvm/CodeGen/GlobalISel/Utils.h
+++ llvm/include/llvm/CodeGen/GlobalISel/Utils.h
@@ -204,6 +204,9 @@
 Optional<APInt> ConstantFoldBinOp(unsigned Opcode, const Register Op1,
                                   const Register Op2,
                                   const MachineRegisterInfo &MRI);
+Optional<APFloat> ConstantFoldFPBinOp(unsigned Opcode, const Register Op1,
+                                      const Register Op2,
+                                      const MachineRegisterInfo &MRI);
 
 Optional<APInt> ConstantFoldExtOp(unsigned Opcode, const Register Op1,
                                   uint64_t Imm, const MachineRegisterInfo &MRI);


-------------- next part --------------
A non-text attachment was scrubbed...
Name: D99035.332155.patch
Type: text/x-patch
Size: 2913 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20210321/7f8c6c03/attachment.bin>