[llvm] [AMDGPU] Implement vop3p complex pattern optmization for gisel (PR #130234)

[via llvm-commits]

================
@@ -4282,44 +4282,364 @@ AMDGPUInstructionSelector::selectVOP3NoMods(MachineOperand &Root) const {
   }};
 }
 
-std::pair<Register, unsigned>
-AMDGPUInstructionSelector::selectVOP3PModsImpl(
-  Register Src, const MachineRegisterInfo &MRI, bool IsDOT) const {
-  unsigned Mods = 0;
-  MachineInstr *MI = MRI.getVRegDef(Src);
+enum srcStatus {
+  IS_SAME,
+  IS_UPPER_HALF,
+  IS_LOWER_HALF,
+  IS_NEG,
+  IS_UPPER_HALF_NEG,
+  IS_LOWER_HALF_NEG,
+  LAST_STAT = IS_LOWER_HALF_NEG
+};
+
+bool isTruncHalf(MachineInstr *MI, const MachineRegisterInfo &MRI) {
+  assert(MI->getOpcode() == AMDGPU::G_TRUNC);
+  unsigned dstSize = MRI.getType(MI->getOperand(0).getReg()).getSizeInBits();
+  unsigned srcSize = MRI.getType(MI->getOperand(1).getReg()).getSizeInBits();
+  return dstSize * 2 == srcSize;
+}
+
+bool isLshrHalf(MachineInstr *MI, const MachineRegisterInfo &MRI) {
+  assert(MI->getOpcode() == AMDGPU::G_LSHR);
+  Register ShiftSrc;
+  std::optional<ValueAndVReg> ShiftAmt;
+  if (mi_match(MI->getOperand(0).getReg(), MRI,
+               m_GLShr(m_Reg(ShiftSrc), m_GCst(ShiftAmt)))) {
+    unsigned srcSize = MRI.getType(MI->getOperand(1).getReg()).getSizeInBits();
+    unsigned shift = ShiftAmt->Value.getZExtValue();
+    return shift * 2 == srcSize;
+  }
+  return false;
+}
 
-  if (MI->getOpcode() == AMDGPU::G_FNEG &&
-      // It's possible to see an f32 fneg here, but unlikely.
-      // TODO: Treat f32 fneg as only high bit.
-      MRI.getType(Src) == LLT::fixed_vector(2, 16)) {
-    Mods ^= (SISrcMods::NEG | SISrcMods::NEG_HI);
-    Src = MI->getOperand(1).getReg();
-    MI = MRI.getVRegDef(Src);
+bool isShlHalf(MachineInstr *MI, const MachineRegisterInfo &MRI) {
+  assert(MI->getOpcode() == AMDGPU::G_SHL);
+  Register ShiftSrc;
+  std::optional<ValueAndVReg> ShiftAmt;
+  if (mi_match(MI->getOperand(0).getReg(), MRI,
+               m_GShl(m_Reg(ShiftSrc), m_GCst(ShiftAmt)))) {
+    unsigned srcSize = MRI.getType(MI->getOperand(1).getReg()).getSizeInBits();
+    unsigned shift = ShiftAmt->Value.getZExtValue();
+    return shift * 2 == srcSize;
+  }
+  return false;
+}
+
+bool retOpStat(MachineOperand *Op, int stat,
----------------
Shoreshen wrote:

Hi @arsenm , this is because I have to imply COPY for cases that copy from SGPR to VGPR (this happens usually after build_vector). 

While COPY can also used to copy from physical registers, I need to block physical register to be selected, since they are not SSA (e.g. used as return value and cause infinite loop).



https://github.com/llvm/llvm-project/pull/130234