[RISCV] Merge ADDI with X0 into base offset (#78940)

If offset is `addi rd, x0, imm`, merge imm into base offset.

Author: tclin914

llvm/lib/Target/RISCV/RISCVMergeBaseOffset.cpp

@@ -193,9 +193,16 @@ bool RISCVMergeBaseOffsetOpt::foldLargeOffset(MachineInstr &Hi,
     if (AddiImmOp.getTargetFlags() != RISCVII::MO_None)
       return false;
     Register AddiReg = OffsetTail.getOperand(1).getReg();
-    if (!AddiReg.isVirtual())
-      return false;
     int64_t OffLo = AddiImmOp.getImm();
+
+    // Handle rs1 of ADDI is X0.
+    if (AddiReg == RISCV::X0) {
+      LLVM_DEBUG(dbgs() << "  Offset Instrs: " << OffsetTail);
+      foldOffset(Hi, Lo, TailAdd, OffLo);
+      OffsetTail.eraseFromParent();
+      return true;
+    }
+
     MachineInstr &OffsetLui = *MRI->getVRegDef(AddiReg);
     MachineOperand &LuiImmOp = OffsetLui.getOperand(1);
     if (OffsetLui.getOpcode() != RISCV::LUI ||
@@ -206,7 +213,7 @@ bool RISCVMergeBaseOffsetOpt::foldLargeOffset(MachineInstr &Hi,
     Offset += OffLo;
     // RV32 ignores the upper 32 bits. ADDIW sign extends the result.
     if (!ST->is64Bit() || OffsetTail.getOpcode() == RISCV::ADDIW)
-       Offset = SignExtend64<32>(Offset);
+      Offset = SignExtend64<32>(Offset);
     // We can only fold simm32 offsets.
     if (!isInt<32>(Offset))
       return false;

llvm/test/CodeGen/RISCV/fold-addi-loadstore.ll

@@ -971,49 +971,37 @@ declare void @f(ptr)
 define i32 @crash() {
 ; RV32I-LABEL: crash:
 ; RV32I:       # %bb.0: # %entry
-; RV32I-NEXT:    li a0, 1
-; RV32I-NEXT:    lui a1, %hi(g)
-; RV32I-NEXT:    addi a1, a1, %lo(g)
-; RV32I-NEXT:    add a0, a1, a0
-; RV32I-NEXT:    lbu a0, 400(a0)
+; RV32I-NEXT:    lui a0, %hi(g+401)
+; RV32I-NEXT:    lbu a0, %lo(g+401)(a0)
 ; RV32I-NEXT:    seqz a0, a0
 ; RV32I-NEXT:    sw a0, 0(zero)
 ; RV32I-NEXT:    li a0, 0
 ; RV32I-NEXT:    ret
 ;
 ; RV32I-MEDIUM-LABEL: crash:
 ; RV32I-MEDIUM:       # %bb.0: # %entry
-; RV32I-MEDIUM-NEXT:    li a0, 1
 ; RV32I-MEDIUM-NEXT:  .Lpcrel_hi14:
-; RV32I-MEDIUM-NEXT:    auipc a1, %pcrel_hi(g)
-; RV32I-MEDIUM-NEXT:    addi a1, a1, %pcrel_lo(.Lpcrel_hi14)
-; RV32I-MEDIUM-NEXT:    add a0, a1, a0
-; RV32I-MEDIUM-NEXT:    lbu a0, 400(a0)
+; RV32I-MEDIUM-NEXT:    auipc a0, %pcrel_hi(g+401)
+; RV32I-MEDIUM-NEXT:    lbu a0, %pcrel_lo(.Lpcrel_hi14)(a0)
 ; RV32I-MEDIUM-NEXT:    seqz a0, a0
 ; RV32I-MEDIUM-NEXT:    sw a0, 0(zero)
 ; RV32I-MEDIUM-NEXT:    li a0, 0
 ; RV32I-MEDIUM-NEXT:    ret
 ;
 ; RV64I-LABEL: crash:
 ; RV64I:       # %bb.0: # %entry
-; RV64I-NEXT:    li a0, 1
-; RV64I-NEXT:    lui a1, %hi(g)
-; RV64I-NEXT:    addi a1, a1, %lo(g)
-; RV64I-NEXT:    add a0, a1, a0
-; RV64I-NEXT:    lbu a0, 400(a0)
+; RV64I-NEXT:    lui a0, %hi(g+401)
+; RV64I-NEXT:    lbu a0, %lo(g+401)(a0)
 ; RV64I-NEXT:    seqz a0, a0
 ; RV64I-NEXT:    sw a0, 0(zero)
 ; RV64I-NEXT:    li a0, 0
 ; RV64I-NEXT:    ret
 ;
 ; RV64I-MEDIUM-LABEL: crash:
 ; RV64I-MEDIUM:       # %bb.0: # %entry
-; RV64I-MEDIUM-NEXT:    li a0, 1
 ; RV64I-MEDIUM-NEXT:  .Lpcrel_hi14:
-; RV64I-MEDIUM-NEXT:    auipc a1, %pcrel_hi(g)
-; RV64I-MEDIUM-NEXT:    addi a1, a1, %pcrel_lo(.Lpcrel_hi14)
-; RV64I-MEDIUM-NEXT:    add a0, a1, a0
-; RV64I-MEDIUM-NEXT:    lbu a0, 400(a0)
+; RV64I-MEDIUM-NEXT:    auipc a0, %pcrel_hi(g+401)
+; RV64I-MEDIUM-NEXT:    lbu a0, %pcrel_lo(.Lpcrel_hi14)(a0)
 ; RV64I-MEDIUM-NEXT:    seqz a0, a0
 ; RV64I-MEDIUM-NEXT:    sw a0, 0(zero)
 ; RV64I-MEDIUM-NEXT:    li a0, 0