[RISCV][GISel] Support G_MERGE_VALUES/G_UNMERGE_VALUES with Zfa. (#120379)

Without Zfa we use pseudos that are lowered to a stack load/store. With
Zfa we have instructions that can move a pair of registers to an FPR. Or
move the high or low half of an FPR to a GPR.

I've used a GINodeEquiv to make use of 3 of the 4 tablegen patterns. The
split case with Zfa requires 2 instructions which I'm doing through
custom isel like we do in SelectionDAG.

Author: topperc

llvm/lib/Target/RISCV/GISel/RISCVInstructionSelector.cpp

@@ -80,7 +80,6 @@ class RISCVInstructionSelector : public InstructionSelector {
   bool selectFPCompare(MachineInstr &MI, MachineIRBuilder &MIB) const;
   void emitFence(AtomicOrdering FenceOrdering, SyncScope::ID FenceSSID,
                  MachineIRBuilder &MIB) const;
-  bool selectMergeValues(MachineInstr &MI, MachineIRBuilder &MIB) const;
   bool selectUnmergeValues(MachineInstr &MI, MachineIRBuilder &MIB) const;
 
   ComplexRendererFns selectShiftMask(MachineOperand &Root,
@@ -732,35 +731,20 @@ bool RISCVInstructionSelector::select(MachineInstr &MI) {
   }
   case TargetOpcode::G_IMPLICIT_DEF:
     return selectImplicitDef(MI, MIB);
-  case TargetOpcode::G_MERGE_VALUES:
-    return selectMergeValues(MI, MIB);
   case TargetOpcode::G_UNMERGE_VALUES:
     return selectUnmergeValues(MI, MIB);
   default:
     return false;
   }
 }
 
-bool RISCVInstructionSelector::selectMergeValues(MachineInstr &MI,
-                                                 MachineIRBuilder &MIB) const {
-  assert(MI.getOpcode() == TargetOpcode::G_MERGE_VALUES);
-
-  // Build a F64 Pair from operands
-  if (MI.getNumOperands() != 3)
-    return false;
-  Register Dst = MI.getOperand(0).getReg();
-  Register Lo = MI.getOperand(1).getReg();
-  Register Hi = MI.getOperand(2).getReg();
-  if (!isRegInFprb(Dst) || !isRegInGprb(Lo) || !isRegInGprb(Hi))
-    return false;
-  MI.setDesc(TII.get(RISCV::BuildPairF64Pseudo));
-  return constrainSelectedInstRegOperands(MI, TII, TRI, RBI);
-}
-
 bool RISCVInstructionSelector::selectUnmergeValues(
     MachineInstr &MI, MachineIRBuilder &MIB) const {
   assert(MI.getOpcode() == TargetOpcode::G_UNMERGE_VALUES);
 
+  if (!Subtarget->hasStdExtZfa())
+    return false;
+
   // Split F64 Src into two s32 parts
   if (MI.getNumOperands() != 3)
     return false;
@@ -769,8 +753,17 @@ bool RISCVInstructionSelector::selectUnmergeValues(
   Register Hi = MI.getOperand(1).getReg();
   if (!isRegInFprb(Src) || !isRegInGprb(Lo) || !isRegInGprb(Hi))
     return false;
-  MI.setDesc(TII.get(RISCV::SplitF64Pseudo));
-  return constrainSelectedInstRegOperands(MI, TII, TRI, RBI);
+
+  MachineInstr *ExtractLo = MIB.buildInstr(RISCV::FMV_X_W_FPR64, {Lo}, {Src});
+  if (!constrainSelectedInstRegOperands(*ExtractLo, TII, TRI, RBI))
+    return false;
+
+  MachineInstr *ExtractHi = MIB.buildInstr(RISCV::FMVH_X_D, {Hi}, {Src});
+  if (!constrainSelectedInstRegOperands(*ExtractHi, TII, TRI, RBI))
+    return false;
+
+  MI.eraseFromParent();
+  return true;
 }
 
 bool RISCVInstructionSelector::replacePtrWithInt(MachineOperand &Op,

llvm/lib/Target/RISCV/RISCVInstrInfoD.td

@@ -23,7 +23,9 @@ def SDT_RISCVSplitF64     : SDTypeProfile<2, 1, [SDTCisVT<0, i32>,
                                                  SDTCisVT<2, f64>]>;
 
 def RISCVBuildPairF64 : SDNode<"RISCVISD::BuildPairF64", SDT_RISCVBuildPairF64>;
+def : GINodeEquiv<G_MERGE_VALUES, RISCVBuildPairF64>;
 def RISCVSplitF64     : SDNode<"RISCVISD::SplitF64", SDT_RISCVSplitF64>;
+def : GINodeEquiv<G_UNMERGE_VALUES, RISCVSplitF64>;
 
 def AddrRegImmINX : ComplexPattern<iPTR, 2, "SelectAddrRegImmRV32Zdinx">;
 

llvm/test/CodeGen/RISCV/GlobalISel/double-zfa.ll

@@ -1,9 +1,8 @@
 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 5
-
 ; RUN: llc -mtriple=riscv32 -mattr=+zfa,d -global-isel < %s \
-; RUN: | FileCheck %s
+; RUN: | FileCheck %s --check-prefixes=CHECK,RV32IDZFA
 ; RUN: llc -mtriple=riscv64 -mattr=+zfa,d -global-isel < %s \
-; RUN: | FileCheck %s
+; RUN: | FileCheck %s --check-prefixes=CHECK,RV64DZFA
 
 
 define double @fceil(double %a) {
@@ -86,3 +85,32 @@ define double @fminimum(double %a, double %b) {
   %c = call double @llvm.minimum.f64(double %a, double %b)
   ret double %c
 }
+
+define i64 @fmvh_x_d(double %fa) {
+; RV32IDZFA-LABEL: fmvh_x_d:
+; RV32IDZFA:       # %bb.0:
+; RV32IDZFA-NEXT:    fmv.x.w a0, fa0
+; RV32IDZFA-NEXT:    fmvh.x.d a1, fa0
+; RV32IDZFA-NEXT:    ret
+;
+; RV64DZFA-LABEL: fmvh_x_d:
+; RV64DZFA:       # %bb.0:
+; RV64DZFA-NEXT:    fmv.x.d a0, fa0
+; RV64DZFA-NEXT:    ret
+  %i = bitcast double %fa to i64
+  ret i64 %i
+}
+
+define double @fmvp_d_x(i64 %a) {
+; RV32IDZFA-LABEL: fmvp_d_x:
+; RV32IDZFA:       # %bb.0:
+; RV32IDZFA-NEXT:    fmvp.d.x fa0, a0, a1
+; RV32IDZFA-NEXT:    ret
+;
+; RV64DZFA-LABEL: fmvp_d_x:
+; RV64DZFA:       # %bb.0:
+; RV64DZFA-NEXT:    fmv.d.x fa0, a0
+; RV64DZFA-NEXT:    ret
+  %or = bitcast i64 %a to double
+  ret double %or
+}