; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 6
; RUN: llc -mtriple=riscv32 -mattr=+xqcili,+xqcisls,+zba,+xqciac -verify-machineinstrs %s -o - | FileCheck %s

@sym = external global i8, align 1

define i32 @load_scaled(i32 %idx) nounwind {
; CHECK-LABEL: load_scaled:
; CHECK:       # %bb.0: # %entry
; CHECK-NEXT:    qc.e.li a1, sym+16
; CHECK-NEXT:    qc.lrw a0, a1, a0, 2
; CHECK-NEXT:    ret
entry:
  %baseptr = getelementptr i8, i8* @sym, i32 0
  %baseint = ptrtoint i8* %baseptr to i32
  %idxsh   = shl i32 %idx, 2
  %sum     = add i32 %baseint, %idxsh
  %sum_ptr = inttoptr i32 %sum to i8*
  %addr_plus_imm = getelementptr i8, i8* %sum_ptr, i32 16
  %addr_i32p = bitcast i8* %addr_plus_imm to i32*
  %val = load i32, i32* %addr_i32p, align 4
  ret i32 %val
}

define void @store_scaled(i32 %idx, i32 %val) nounwind {
; CHECK-LABEL: store_scaled:
; CHECK:       # %bb.0: # %entry
; CHECK-NEXT:    qc.e.li a2, sym+12
; CHECK-NEXT:    qc.srw a1, a2, a0, 3
; CHECK-NEXT:    ret
entry:
  %baseptr = getelementptr i8, i8* @sym, i32 0
  %baseint = ptrtoint i8* %baseptr to i32
  %idxsh   = shl i32 %idx, 3
  %sum     = add i32 %baseint, %idxsh
  %sum_ptr = inttoptr i32 %sum to i8*
  %addr_plus_imm = getelementptr i8, i8* %sum_ptr, i32 12
  %addr_i32p = bitcast i8* %addr_plus_imm to i32*
  store i32 %val, i32* %addr_i32p, align 4
  ret void
}

define i32 @load_scaled_i8(i32 %idx) nounwind {
; CHECK-LABEL: load_scaled_i8:
; CHECK:       # %bb.0: # %entry
; CHECK-NEXT:    qc.e.li a1, sym+5
; CHECK-NEXT:    qc.lrb a0, a1, a0, 2
; CHECK-NEXT:    ret
entry:
  %baseptr = getelementptr i8, i8* @sym, i32 0
  %baseint = ptrtoint i8* %baseptr to i32
  %idxsh   = shl i32 %idx, 2
  %sum     = add i32 %baseint, %idxsh
  %sum_ptr = inttoptr i32 %sum to i8*
  %addr_plus_imm = getelementptr i8, i8* %sum_ptr, i32 5
  %val8 = load i8, i8* %addr_plus_imm, align 1
  %val = sext i8 %val8 to i32
  ret i32 %val
}

define void @store_scaled_i8(i32 %idx, i32 %val) nounwind {
; CHECK-LABEL: store_scaled_i8:
; CHECK:       # %bb.0: # %entry
; CHECK-NEXT:    qc.e.li a2, sym+7
; CHECK-NEXT:    qc.srb a1, a2, a0, 3
; CHECK-NEXT:    ret
entry:
  %baseptr = getelementptr i8, i8* @sym, i32 0
  %baseint = ptrtoint i8* %baseptr to i32
  %idxsh   = shl i32 %idx, 3
  %sum     = add i32 %baseint, %idxsh
  %sum_ptr = inttoptr i32 %sum to i8*
  %addr_plus_imm = getelementptr i8, i8* %sum_ptr, i32 7
  %val8 = trunc i32 %val to i8
  store i8 %val8, i8* %addr_plus_imm, align 1
  ret void
}

define i32 @load_scaled_i16(i32 %idx) nounwind {
; CHECK-LABEL: load_scaled_i16:
; CHECK:       # %bb.0: # %entry
; CHECK-NEXT:    qc.e.li a1, sym+6
; CHECK-NEXT:    qc.lrh a0, a1, a0, 1
; CHECK-NEXT:    ret
entry:
  %baseptr = getelementptr i8, i8* @sym, i32 0
  %baseint = ptrtoint i8* %baseptr to i32
  %idxsh   = shl i32 %idx, 1
  %sum     = add i32 %baseint, %idxsh
  %sum_ptr = inttoptr i32 %sum to i16*
  %addr_plus_imm = getelementptr i16, i16* %sum_ptr, i32 3
  %val16 = load i16, i16* %addr_plus_imm, align 2
  %val = sext i16 %val16 to i32
  ret i32 %val
}

define void @store_scaled_i16(i32 %idx, i32 %val) nounwind {
; CHECK-LABEL: store_scaled_i16:
; CHECK:       # %bb.0: # %entry
; CHECK-NEXT:    qc.e.li a2, sym+10
; CHECK-NEXT:    qc.srh a1, a2, a0, 2
; CHECK-NEXT:    ret
entry:
  %baseptr = getelementptr i8, i8* @sym, i32 0
  %baseint = ptrtoint i8* %baseptr to i32
  %idxsh   = shl i32 %idx, 2
  %sum     = add i32 %baseint, %idxsh
  %sum_ptr = inttoptr i32 %sum to i16*
  %addr_plus_imm = getelementptr i16, i16* %sum_ptr, i32 5
  %val16 = trunc i32 %val to i16
  store i16 %val16, i16* %addr_plus_imm, align 2
  ret void
}

define void @no_store_scaled_i16_multiuse_base_reg(i32 %idx, i32 %val) nounwind {
; CHECK-LABEL: no_store_scaled_i16_multiuse_base_reg:
; CHECK:       # %bb.0: # %entry
; CHECK-NEXT:    qc.e.li a2, sym
; CHECK-NEXT:    sh2add a0, a0, a2
; CHECK-NEXT:    sh a1, 10(a0)
; CHECK-NEXT:    sh a1, 14(a0)
; CHECK-NEXT:    ret
entry:
  %baseptr = getelementptr i8, i8* @sym, i32 0
  %baseint = ptrtoint i8* %baseptr to i32
  %idxsh   = shl i32 %idx, 2
  %sum     = add i32 %baseint, %idxsh
  %sum_ptr = inttoptr i32 %sum to i16*
  %addr_plus_imm = getelementptr i16, i16* %sum_ptr, i32 5
  %addr_plus_imm2 = getelementptr i16, i16* %sum_ptr, i32 7
  %val16 = trunc i32 %val to i16
  store i16 %val16, i16* %addr_plus_imm, align 2
  store i16 %val16, i16* %addr_plus_imm2, align 2
  ret void
}

define void @no_store_scaled_i8_multiuse_scaled_reg(i32 %idx, i32 %val) nounwind {
; CHECK-LABEL: no_store_scaled_i8_multiuse_scaled_reg:
; CHECK:       # %bb.0: # %entry
; CHECK-NEXT:    qc.e.li a2, sym
; CHECK-NEXT:    sh3add a0, a0, a2
; CHECK-NEXT:    add a0, a0, a1
; CHECK-NEXT:    sb a1, 7(a0)
; CHECK-NEXT:    ret
entry:
  %baseptr = getelementptr i8, i8* @sym, i32 0
  %baseint = ptrtoint i8* %baseptr to i32
  %idxsh   = shl i32 %idx, 3
  %sum     = add i32 %baseint, %idxsh
  %sum2    = add i32 %sum, %val
  %sum_ptr = inttoptr i32 %sum2 to i8*
  %addr_plus_imm = getelementptr i8, i8* %sum_ptr, i32 7
  %val8 = trunc i32 %val to i8
  store i8 %val8, i8* %addr_plus_imm, align 1
  ret void
}

define i32 @load_scaled_i16_shift_gt_3(i32 %idx) nounwind {
; CHECK-LABEL: load_scaled_i16_shift_gt_3:
; CHECK:       # %bb.0: # %entry
; CHECK-NEXT:    qc.e.li a1, sym+6
; CHECK-NEXT:    qc.lrh a0, a1, a0, 5
; CHECK-NEXT:    ret
entry:
  %baseptr = getelementptr i8, i8* @sym, i32 0
  %baseint = ptrtoint i8* %baseptr to i32
  %idxsh   = shl i32 %idx, 5
  %sum     = add i32 %baseint, %idxsh
  %sum_ptr = inttoptr i32 %sum to i16*
  %addr_plus_imm = getelementptr i16, i16* %sum_ptr, i32 3
  %val16 = load i16, i16* %addr_plus_imm, align 2
  %val = sext i16 %val16 to i32
  ret i32 %val
}

define i32 @no_load_scaled_i16_shift_gt_7(i32 %idx) nounwind {
; CHECK-LABEL: no_load_scaled_i16_shift_gt_7:
; CHECK:       # %bb.0: # %entry
; CHECK-NEXT:    qc.e.li a1, sym
; CHECK-NEXT:    qc.shladd a0, a0, a1, 8
; CHECK-NEXT:    lh a0, 6(a0)
; CHECK-NEXT:    ret
entry:
  %baseptr = getelementptr i8, i8* @sym, i32 0
  %baseint = ptrtoint i8* %baseptr to i32
  %idxsh   = shl i32 %idx, 8
  %sum     = add i32 %baseint, %idxsh
  %sum_ptr = inttoptr i32 %sum to i16*
  %addr_plus_imm = getelementptr i16, i16* %sum_ptr, i32 3
  %val16 = load i16, i16* %addr_plus_imm, align 2
  %val = sext i16 %val16 to i32
  ret i32 %val
}