[llvm] Promote 32bit pseudo instr that infer extsw removal to 64bit in PPCMIPeephole (PR #85451)

zhijian lin via llvm-commits llvm-commits at lists.llvm.org
Wed Oct 23 10:58:10 PDT 2024


================
@@ -5234,6 +5234,216 @@ bool PPCInstrInfo::isTOCSaveMI(const MachineInstr &MI) const {
 // We limit the max depth to track incoming values of PHIs or binary ops
 // (e.g. AND) to avoid excessive cost.
 const unsigned MAX_BINOP_DEPTH = 1;
+
+// This function will promote the instruction which defines the register `Reg`
+// in the parameter from a 32-bit to a 64-bit instruction if needed. The logic
+// used to check whether an instruction needs to be promoted or not is similar
+// to the logic used to check whether or not a defined register is sign or zero
+// extended within the function PPCInstrInfo::isSignOrZeroExtended.
+// Additionally, the `promoteInstr32To64ForElimEXTSW` function is recursive.
+// BinOpDepth does not count all of the recursions. The parameter BinOpDepth is
+// incremented  only when `promoteInstr32To64ForElimEXTSW` calls itself more
+// than once. This is done to prevent exponential recursion.
+void PPCInstrInfo::promoteInstr32To64ForElimEXTSW(const Register &Reg,
+                                                  MachineRegisterInfo *MRI,
+                                                  unsigned BinOpDepth,
+                                                  LiveVariables *LV) const {
+  if (!Reg.isVirtual())
+    return;
+
+  MachineInstr *MI = MRI->getVRegDef(Reg);
+  if (!MI)
+    return;
+
+  unsigned Opcode = MI->getOpcode();
+  bool HasNonSignedExtInstrPromoted = false;
+  int NewOpcode = -1;
+
+  std::unordered_map<unsigned, unsigned> OpcodeMap = {
+      {PPC::OR, PPC::OR8},     {PPC::ISEL, PPC::ISEL8},
+      {PPC::ORI, PPC::ORI8},   {PPC::XORI, PPC::XORI8},
+      {PPC::ORIS, PPC::ORIS8}, {PPC::XORIS, PPC::XORIS8},
+      {PPC::AND, PPC::AND8}};
+
+  // Check if the Opcode is in the map.
----------------
diggerlin wrote:

I add the comment in above 

+  // Map the opcode of instructions (which are not sign- or zero-extended
+  // themselves,but have operands that are destination registers of sign- or
+  // zero-extended instructions) to their 64-bit equivalents.

I think it is clear enough to explain it, I delete the comment here.



https://github.com/llvm/llvm-project/pull/85451


More information about the llvm-commits mailing list