[llvm] unpack packed instructions overlapped by MFMAs post-RA scheduling (PR #157968)

[Jan Patrick Lehr via llvm-commits]

================
@@ -417,6 +455,233 @@ bool SIPreEmitPeephole::removeExeczBranch(MachineInstr &MI,
   return true;
 }
 
+bool SIPreEmitPeephole::isUnpackingSupportedInstr(MachineInstr &MI) const {
+  unsigned Opcode = MI.getOpcode();
+  switch (Opcode) {
+  case AMDGPU::V_PK_ADD_F32:
+  case AMDGPU::V_PK_MUL_F32:
+  case AMDGPU::V_PK_FMA_F32:
+    return true;
+  default:
+    return false;
+  }
+  llvm_unreachable("Fully covered switch");
+}
+
+bool SIPreEmitPeephole::hasReadWriteDependencies(const MachineInstr &PredMI,
+                                                 const MachineInstr &SuccMI) {
+  for (const MachineOperand &Pred_Ops : PredMI.operands()) {
+    if (!Pred_Ops.isReg() || !Pred_Ops.isDef())
+      continue;
+    Register Pred_Reg = Pred_Ops.getReg();
+    if (!Pred_Reg.isValid())
+      continue;
+    for (const MachineOperand &Succ_Ops : SuccMI.operands()) {
+      if (!Succ_Ops.isReg() || !Succ_Ops.isDef())
+        continue;
+      Register Succ_Reg = Succ_Ops.getReg();
+      if (!Succ_Reg.isValid())
+        continue;
+      if ((Pred_Reg == Succ_Reg) || TRI->regsOverlap(Pred_Reg, Succ_Reg)) {
+        return true;
+      }
+    }
+  }
+  return false;
+}
+
+uint16_t SIPreEmitPeephole::mapToUnpackedOpcode(MachineInstr &I) {
+  unsigned Opcode = I.getOpcode();
+  // Use 64 bit encoding to allow use of VOP3 instructions.
+  // VOP3 instructions allow VOP3P source modifiers to be translated to VOP3
+  // e32 instructions are VOP2 and don't allow source modifiers
+  switch (Opcode) {
+  case AMDGPU::V_PK_ADD_F32:
+    return AMDGPU::V_ADD_F32_e64;
+  case AMDGPU::V_PK_MUL_F32:
+    return AMDGPU::V_MUL_F32_e64;
+  case AMDGPU::V_PK_FMA_F32:
+    return AMDGPU::V_FMA_F32_e64;
+  default:
+    return std::numeric_limits<uint16_t>::max();
+  }
+  llvm_unreachable("Fully covered switch");
+}
+
+void SIPreEmitPeephole::addOperandandMods(MachineInstrBuilder NewMI,
+                                          unsigned Src_Mods,
+                                          unsigned NegModifier,
+                                          unsigned OpSelModifier,
+                                          MachineOperand &SrcMO) {
+  unsigned New_Src_Mods = 0;
+  const TargetRegisterInfo *RI = SrcMO.getParent()
+                                     ->getParent()
+                                     ->getParent()
+                                     ->getSubtarget()
+                                     .getRegisterInfo();
+  //  If NEG or NEG_HI is true, we need to negate the corresponding 32 bit
+  //  lane.
+  //  NEG_HI shares the same bit position with ABS. But packed instructions do
+  //  not support ABS. Therefore, NEG_HI must be translated to NEG source
+  //  modifier for the higher 32 bits. Unpacked VOP3 instructions do support
+  //  ABS, therefore we need to explicitly add the NEG modifier if present in
+  //  the packed instruction
+  if (Src_Mods & NegModifier) {
+    New_Src_Mods |= SISrcMods::NEG;
+  }
+  // Src modifiers. Only negative modifiers are added if needed. Unpacked
+  // operations do not have op_sel, therefore it must be handled explicitly as
+  // done below. Unpacked operations support abs, but packed instructions do
+  // not. Thus, abs is not handled.
+  NewMI.addImm(New_Src_Mods);
+  if (SrcMO.isImm()) {
+    NewMI.addImm(SrcMO.getImm());
+  } else {
----------------
jplehr wrote:

Should this be instead spelled like
```
if ( cond ) {
  NewMi.add....
  return;
}
// NO else
```
Not sure what is preferred.

https://github.com/llvm/llvm-project/pull/157968