[llvm-branch-commits] [llvm] [LoongArch] Optimize for immediate value materialization using BSTRINS_D instruction (PR #106332)

[via llvm-branch-commits]

================
@@ -41,11 +43,82 @@ LoongArchMatInt::InstSeq LoongArchMatInt::generateInstSeq(int64_t Val) {
       Insts.push_back(Inst(LoongArch::ORI, Lo12));
   }
 
+  // hi32
+  // Higher20
   if (SignExtend32<1>(Hi20 >> 19) != SignExtend32<20>(Higher20))
     Insts.push_back(Inst(LoongArch::LU32I_D, SignExtend64<20>(Higher20)));
 
+  // Highest12
   if (SignExtend32<1>(Higher20 >> 19) != SignExtend32<12>(Highest12))
     Insts.push_back(Inst(LoongArch::LU52I_D, SignExtend64<12>(Highest12)));
 
+  size_t N = Insts.size();
+  if (N < 3)
+    return Insts;
+
+  // When the number of instruction sequences is greater than 2, we have the
+  // opportunity to optimize using the BSTRINS_D instruction. The scenario is as
+  // follows:
+  //
+  // N of Insts = 3
+  // 1. ORI + LU32I_D + LU52I_D     =>     ORI + BSTRINS_D, TmpVal = ORI
+  // 2. ADDI_W + LU32I_D + LU32I_D  =>  ADDI_W + BSTRINS_D, TmpVal = ADDI_W
----------------
heiher wrote:

ADDI_W + LU32I_D + LU{52}I_D

https://github.com/llvm/llvm-project/pull/106332