;; aarch64 instruction selection and CLIF-to-MachInst lowering.

;; The main lowering constructor term: takes a clif `Inst` and returns the
;; register(s) within which the lowered instruction's result values live.
(decl lower (Inst) ValueRegs)

;;;; Rules for `iconst` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;

(rule (lower (has_type ty (iconst (u64_from_imm64 n))))
      (imm ty n))

;;;; Rules for `bconst` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;

(rule (lower (has_type ty (bconst $false)))
      (imm ty 0))

(rule (lower (has_type ty (bconst $true)))
      (imm ty 1))

;;;; Rules for `null` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;

(rule (lower (has_type ty (null)))
      (imm ty 0))

;;;; Rules for `iadd` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;

;; `i64` and smaller

;; Base case, simply adding things in registers.
(rule (lower (has_type (fits_in_64 ty) (iadd x y)))
      (add ty  x y))

;; Special cases for when one operand is an immediate that fits in 12 bits.
(rule (lower (has_type (fits_in_64 ty) (iadd x (imm12_from_value y))))
      (add_imm ty x y))

(rule (lower (has_type (fits_in_64 ty) (iadd (imm12_from_value x) y)))
      (add_imm ty y x))

;; Same as the previous special cases, except we can switch the addition to a
;; subtraction if the negated immediate fits in 12 bits.
(rule (lower (has_type (fits_in_64 ty) (iadd x (imm12_from_negated_value y))))
      (sub_imm ty x y))

(rule (lower (has_type (fits_in_64 ty) (iadd (imm12_from_negated_value x) y)))
      (sub_imm ty y x))

;; Special cases for when we're adding an extended register where the extending
;; operation can get folded into the add itself.
(rule (lower (has_type (fits_in_64 ty) (iadd x (extended_value_from_value y))))
      (add_extend ty x y))

(rule (lower (has_type (fits_in_64 ty) (iadd (extended_value_from_value x) y)))
      (add_extend ty y x))

;; Special cases for when we're adding the shift of a different
;; register by a constant amount and the shift can get folded into the add.
(rule (lower (has_type (fits_in_64 ty)
                       (iadd x (ishl y (iconst (lshl_from_imm64 <ty amt))))))
      (add_shift ty x y amt))

(rule (lower (has_type (fits_in_64 ty)
                       (iadd (ishl x (iconst (lshl_from_imm64 <ty amt))) y)))
      (add_shift ty y x amt))

;; Fold an `iadd` and `imul` combination into a `madd` instruction.
(rule (lower (has_type (fits_in_64 ty) (iadd x (imul y z))))
      (madd ty y z x))

(rule (lower (has_type (fits_in_64 ty) (iadd (imul x y) z)))
      (madd ty x y z))

;; vectors

(rule (lower (has_type ty @ (multi_lane _ _) (iadd x y)))
      (add_vec x y (vector_size ty)))

;; `i128`
(rule (lower (has_type $I128 (iadd x y)))
      (let
          ;; Get the high/low registers for `x`.
          ((x_regs ValueRegs x)
           (x_lo Reg (value_regs_get x_regs 0))
           (x_hi Reg (value_regs_get x_regs 1))

           ;; Get the high/low registers for `y`.
           (y_regs ValueRegs y)
           (y_lo Reg (value_regs_get y_regs 0))
           (y_hi Reg (value_regs_get y_regs 1)))
        ;; the actual addition is `adds` followed by `adc` which comprises the
        ;; low/high bits of the result
        (with_flags
          (add_with_flags_paired $I64 x_lo y_lo)
          (adc_paired $I64 x_hi y_hi))))

;;;; Rules for `isub` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;

;; `i64` and smaller

;; Base case, simply subtracting things in registers.
(rule (lower (has_type (fits_in_64 ty) (isub x y)))
      (sub ty x y))

;; Special case for when one operand is an immediate that fits in 12 bits.
(rule (lower (has_type (fits_in_64 ty) (isub x (imm12_from_value y))))
      (sub_imm ty x y))

;; Same as the previous special case, except we can switch the subtraction to an
;; addition if the negated immediate fits in 12 bits.
(rule (lower (has_type (fits_in_64 ty) (isub x (imm12_from_negated_value y))))
      (add_imm ty x y))

;; Special cases for when we're subtracting an extended register where the
;; extending operation can get folded into the sub itself.
(rule (lower (has_type (fits_in_64 ty) (isub x (extended_value_from_value y))))
      (sub_extend ty x y))

;; Finally a special case for when we're subtracting the shift of a different
;; register by a constant amount and the shift can get folded into the sub.
(rule (lower (has_type (fits_in_64 ty)
                       (isub x (ishl y (iconst (lshl_from_imm64 <ty amt))))))
      (sub_shift ty x y amt))

;; vectors
(rule (lower (has_type ty @ (multi_lane _ _) (isub x y)))
      (sub_vec x y (vector_size ty)))

;; `i128`
(rule (lower (has_type $I128 (isub x y)))
      (let
          ;; Get the high/low registers for `x`.
          ((x_regs ValueRegs x)
           (x_lo Reg (value_regs_get x_regs 0))
           (x_hi Reg (value_regs_get x_regs 1))

           ;; Get the high/low registers for `y`.
           (y_regs ValueRegs y)
           (y_lo Reg (value_regs_get y_regs 0))
           (y_hi Reg (value_regs_get y_regs 1)))
        ;; the actual subtraction is `subs` followed by `sbc` which comprises
        ;; the low/high bits of the result
        (with_flags
          (sub_with_flags_paired $I64 x_lo y_lo)
          (sbc_paired $I64 x_hi y_hi))))

;;;; Rules for `uadd_sat` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;

(rule (lower (has_type (vec128 ty) (uadd_sat x y)))
      (uqadd x y (vector_size ty)))

;;;; Rules for `sadd_sat` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;

(rule (lower (has_type (vec128 ty) (sadd_sat x y)))
      (sqadd x y (vector_size ty)))

;;;; Rules for `usub_sat` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;

(rule (lower (has_type (vec128 ty) (usub_sat x y)))
      (uqsub x y (vector_size ty)))

;;;; Rules for `ssub_sat` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;

(rule (lower (has_type (vec128 ty) (ssub_sat x y)))
      (sqsub x y (vector_size ty)))

;;;; Rules for `ineg` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;

;; `i64` and smaller.
(rule (lower (has_type (fits_in_64 ty) (ineg x)))
      (sub ty (zero_reg) x))

;; vectors.
(rule (lower (has_type (vec128 ty) (ineg x)))
      (neg x (vector_size ty)))

;;;; Rules for `imul` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;

;; `i64` and smaller.
(rule (lower (has_type (fits_in_64 ty) (imul x y)))
      (madd ty x y (zero_reg)))

;; `i128`.
(rule (lower (has_type $I128 (imul x y)))
      (let
          ;; Get the high/low registers for `x`.
          ((x_regs ValueRegs x)
           (x_lo Reg (value_regs_get x_regs 0))
           (x_hi Reg (value_regs_get x_regs 1))

           ;; Get the high/low registers for `y`.
           (y_regs ValueRegs y)
           (y_lo Reg (value_regs_get y_regs 0))
           (y_hi Reg (value_regs_get y_regs 1))

           ;; 128bit mul formula:
           ;;   dst_lo = x_lo * y_lo
           ;;   dst_hi = umulhi(x_lo, y_lo) + (x_lo * y_hi) + (x_hi * y_lo)
           ;;
           ;; We can convert the above formula into the following
           ;; umulh   dst_hi, x_lo, y_lo
           ;; madd    dst_hi, x_lo, y_hi, dst_hi
           ;; madd    dst_hi, x_hi, y_lo, dst_hi
           ;; madd    dst_lo, x_lo, y_lo, zero
           (dst_hi1 Reg (umulh $I64 x_lo y_lo))
           (dst_hi2 Reg (madd64 x_lo y_hi dst_hi1))
           (dst_hi Reg (madd64 x_hi y_lo dst_hi2))
           (dst_lo Reg (madd64 x_lo y_lo (zero_reg))))
        (value_regs dst_lo dst_hi)))

;; Case for i8x16, i16x8, and i32x4.
(rule (lower (has_type (vec128 ty @ (not_i64x2)) (imul x y)))
      (mul x y (vector_size ty)))

;; Special lowering for i64x2.
;;
;; This I64X2 multiplication is performed with several 32-bit
;; operations.
;;
;; 64-bit numbers x and y, can be represented as:
;;   x = a + 2^32(b)
;;   y = c + 2^32(d)
;;
;; A 64-bit multiplication is:
;;   x * y = ac + 2^32(ad + bc) + 2^64(bd)
;; note: `2^64(bd)` can be ignored, the value is too large to fit in
;; 64 bits.
;;
;; This sequence implements a I64X2 multiply, where the registers
;; `rn` and `rm` are split up into 32-bit components:
;;   rn = |d|c|b|a|
;;   rm = |h|g|f|e|
;;
;;   rn * rm = |cg + 2^32(ch + dg)|ae + 2^32(af + be)|
;;
;;  The sequence is:
;;  rev64 rd.4s, rm.4s
;;  mul rd.4s, rd.4s, rn.4s
;;  xtn tmp1.2s, rn.2d
;;  addp rd.4s, rd.4s, rd.4s
;;  xtn tmp2.2s, rm.2d
;;  shll rd.2d, rd.2s, #32
;;  umlal rd.2d, tmp2.2s, tmp1.2s
(rule (lower (has_type $I64X2 (imul x y)))
      (let ((rn Reg x)
            (rm Reg y)
            ;; Reverse the 32-bit elements in the 64-bit words.
            ;;   rd = |g|h|e|f|
            (rev Reg (rev64 rm (VectorSize.Size32x4)))

            ;; Calculate the high half components.
            ;;   rd = |dg|ch|be|af|
            ;;
            ;; Note that this 32-bit multiply of the high half
            ;; discards the bits that would overflow, same as
            ;; if 64-bit operations were used. Also the Shll
            ;; below would shift out the overflow bits anyway.
            (mul Reg (mul rev rn (VectorSize.Size32x4)))

            ;; Extract the low half components of rn.
            ;;   tmp1 = |c|a|
            (tmp1 Reg (xtn64 rn $false))

            ;; Sum the respective high half components.
            ;;   rd = |dg+ch|be+af||dg+ch|be+af|
            (sum Reg (addp mul mul (VectorSize.Size32x4)))

            ;; Extract the low half components of rm.
            ;;   tmp2 = |g|e|
            (tmp2 Reg (xtn64 rm $false))

            ;; Shift the high half components, into the high half.
            ;;   rd = |dg+ch << 32|be+af << 32|
            (shift Reg (shll32 sum $false))

            ;; Multiply the low components together, and accumulate with the high
            ;; half.
            ;;   rd = |rd[1] + cg|rd[0] + ae|
            (result Reg (umlal32 shift tmp2 tmp1 $false)))
        result))

;; Special case for `i16x8.extmul_low_i8x16_s`.
(rule (lower (has_type $I16X8
                       (imul (swiden_low x @ (value_type $I8X16))
                             (swiden_low y @ (value_type $I8X16)))))
      (smull8 x y $false))

;; Special case for `i16x8.extmul_high_i8x16_s`.
(rule (lower (has_type $I16X8
                       (imul (swiden_high x @ (value_type $I8X16))
                             (swiden_high y @ (value_type $I8X16)))))
      (smull8 x y $true))

;; Special case for `i16x8.extmul_low_i8x16_u`.
(rule (lower (has_type $I16X8
                       (imul (uwiden_low x @ (value_type $I8X16))
                             (uwiden_low y @ (value_type $I8X16)))))
      (umull8 x y $false))

;; Special case for `i16x8.extmul_high_i8x16_u`.
(rule (lower (has_type $I16X8
                       (imul (uwiden_high x @ (value_type $I8X16))
                             (uwiden_high y @ (value_type $I8X16)))))
      (umull8 x y $true))

;; Special case for `i32x4.extmul_low_i16x8_s`.
(rule (lower (has_type $I32X4
                       (imul (swiden_low x @ (value_type $I16X8))
                             (swiden_low y @ (value_type $I16X8)))))
      (smull16 x y $false))

;; Special case for `i32x4.extmul_high_i16x8_s`.
(rule (lower (has_type $I32X4
                       (imul (swiden_high x @ (value_type $I16X8))
                             (swiden_high y @ (value_type $I16X8)))))
      (smull16 x y $true))

;; Special case for `i32x4.extmul_low_i16x8_u`.
(rule (lower (has_type $I32X4
                       (imul (uwiden_low x @ (value_type $I16X8))
                             (uwiden_low y @ (value_type $I16X8)))))
      (umull16 x y $false))

;; Special case for `i32x4.extmul_high_i16x8_u`.
(rule (lower (has_type $I32X4
                       (imul (uwiden_high x @ (value_type $I16X8))
                             (uwiden_high y @ (value_type $I16X8)))))
      (umull16 x y $true))

;; Special case for `i64x2.extmul_low_i32x4_s`.
(rule (lower (has_type $I64X2
                       (imul (swiden_low x @ (value_type $I32X4))
                             (swiden_low y @ (value_type $I32X4)))))
      (smull32 x y $false))

;; Special case for `i64x2.extmul_high_i32x4_s`.
(rule (lower (has_type $I64X2
                       (imul (swiden_high x @ (value_type $I32X4))
                             (swiden_high y @ (value_type $I32X4)))))
      (smull32 x y $true))

;; Special case for `i64x2.extmul_low_i32x4_u`.
(rule (lower (has_type $I64X2
                       (imul (uwiden_low x @ (value_type $I32X4))
                             (uwiden_low y @ (value_type $I32X4)))))
      (umull32 x y $false))

;; Special case for `i64x2.extmul_high_i32x4_u`.
(rule (lower (has_type $I64X2
                       (imul (uwiden_high x @ (value_type $I32X4))
                             (uwiden_high y @ (value_type $I32X4)))))
      (umull32 x y $true))

;;;; Rules for `smulhi` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;

(rule (lower (has_type $I64 (smulhi x y)))
      (smulh $I64 x y))

(rule (lower (has_type (fits_in_32 ty) (smulhi x y)))
      (let ((x64 Reg (put_in_reg_sext64 x))
            (y64 Reg (put_in_reg_sext64 y))
            (mul Reg (madd64 x64 y64 (zero_reg)))
            (result Reg (asr_imm $I64 mul (imm_shift_from_u8 (ty_bits ty)))))
        result))

;;;; Rules for `umulhi` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;

(rule (lower (has_type $I64 (umulhi x y)))
      (umulh $I64 x y))

(rule (lower (has_type (fits_in_32 ty) (umulhi x y)))
      (let ((x64 Reg (put_in_reg_zext64 x))
            (y64 Reg (put_in_reg_zext64 y))
            (mul Reg (madd64 x64 y64 (zero_reg)))
            (result Reg (lsr_imm $I64 mul (imm_shift_from_u8 (ty_bits ty)))))
        result))

;;;; Rules for `udiv` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;

;; TODO: Add UDiv32 to implement 32-bit directly, rather
;; than extending the input.
;;
;; Note that aarch64's `udiv` doesn't trap so to respect the semantics of
;; CLIF's `udiv` the check for zero needs to be manually performed.
(rule (lower (has_type (fits_in_64 ty) (udiv x y)))
      (a64_udiv $I64 (put_in_reg_zext64 x) (put_nonzero_in_reg_zext64 y)))

;; Helper for placing a `Value` into a `Reg` and validating that it's nonzero.
(decl put_nonzero_in_reg_zext64 (Value) Reg)
(rule (put_nonzero_in_reg_zext64 val)
      (trap_if_zero_divisor (put_in_reg_zext64 val)))

;; Special case where if a `Value` is known to be nonzero we can trivially
;; move it into a register.
(rule (put_nonzero_in_reg_zext64 (and (value_type ty)
                                      (iconst (nonzero_u64_from_imm64 n))))
      (imm ty n))

;;;; Rules for `sdiv` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;

;; TODO: Add SDiv32 to implement 32-bit directly, rather
;; than extending the input.
;;
;; The sequence of checks here should look like:
;;
;;   cbnz rm, #8
;;   udf ; divide by zero
;;   cmn rm, 1
;;   ccmp rn, 1, #nzcv, eq
;;   b.vc #8
;;   udf ; signed overflow
;;
;; Note The div instruction does not trap on divide by zero or overflow, so
;; checks need to be manually inserted.
;;
;; TODO: if `y` is -1 then a check that `x` is not INT_MIN is all that's
;; necessary, but right now `y` is checked to not be -1 as well.
(rule (lower (has_type (fits_in_64 ty) (sdiv x y)))
      (let ((x64 Reg (put_in_reg_sext64 x))
            (y64 Reg (put_nonzero_in_reg_sext64 y))
            (valid_x64 Reg (trap_if_div_overflow ty x64 y64))
            (result Reg (a64_sdiv $I64 valid_x64 y64)))
        result))

;; Helper for extracting an immediate that's not 0 and not -1 from an imm64.
(decl safe_divisor_from_imm64 (u64) Imm64)
(extern extractor safe_divisor_from_imm64 safe_divisor_from_imm64)

;; Special case for `sdiv` where no checks are needed due to division by a
;; constant meaning the checks are always passed.
(rule (lower (has_type (fits_in_64 ty) (sdiv x (iconst (safe_divisor_from_imm64 y)))))
      (a64_sdiv $I64 (put_in_reg_sext64 x) (imm ty y)))

;; Helper for placing a `Value` into a `Reg` and validating that it's nonzero.
(decl put_nonzero_in_reg_sext64 (Value) Reg)
(rule (put_nonzero_in_reg_sext64 val)
      (trap_if_zero_divisor (put_in_reg_sext64 val)))

;; Note that this has a special case where if the `Value` is a constant that's
;; not zero we can skip the zero check.
(rule (put_nonzero_in_reg_sext64 (and (value_type ty)
                                      (iconst (nonzero_u64_from_imm64 n))))
      (imm ty n))

;;;; Rules for `urem` and `srem` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;

;; Remainder (x % y) is implemented as:
;;
;;   tmp = x / y
;;   result = x - (tmp*y)
;;
;; use 'result' for tmp and you have:
;;
;;   cbnz y, #8         ; branch over trap
;;   udf                ; divide by zero
;;   div rd, x, y       ; rd = x / y
;;   msub rd, rd, y, x  ; rd = x - rd * y

(rule (lower (has_type (fits_in_64 ty) (urem x y)))
      (let ((x64 Reg (put_in_reg_zext64 x))
            (y64 Reg (put_nonzero_in_reg_zext64 y))
            (div Reg (a64_udiv $I64 x64 y64))
            (result Reg (msub64 div y64 x64)))
        result))

(rule (lower (has_type (fits_in_64 ty) (srem x y)))
      (let ((x64 Reg (put_in_reg_sext64 x))
            (y64 Reg (put_nonzero_in_reg_sext64 y))
            (div Reg (a64_sdiv $I64 x64 y64))
            (result Reg (msub64 div y64 x64)))
        result))

;;;; Rules for `uextend` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;

;; General rule for extending input to an output which fits in a single
;; register.
(rule (lower (has_type (fits_in_64 out) (uextend x @ (value_type in))))
      (extend x $false (ty_bits in) (ty_bits out)))

;; Extraction of a vector lane automatically extends as necessary, so we can
;; skip an explicit extending instruction.
(rule (lower (has_type (fits_in_64 out)
                       (uextend (extractlane vec @ (value_type in)
                                             (u8_from_uimm8 lane)))))
      (mov_from_vec (put_in_reg vec) lane (vector_size in)))

;; Atomic loads will also automatically zero their upper bits so the `uextend`
;; instruction can effectively get skipped here.
(rule (lower (has_type (fits_in_64 out)
                       (uextend (and (value_type in) (sinkable_atomic_load addr)))))
      (load_acquire in (sink_atomic_load addr)))

;; Conversion to 128-bit needs a zero-extension of the lower bits and the upper
;; bits are all zero.
(rule (lower (has_type $I128 (uextend x)))
      (value_regs (put_in_reg_zext64 x) (imm $I64 0)))

;; Like above where vector extraction automatically zero-extends extending to
;; i128 only requires generating a 0 constant for the upper bits.
(rule (lower (has_type $I128
                       (uextend (extractlane vec @ (value_type in)
                                             (u8_from_uimm8 lane)))))
      (value_regs (mov_from_vec (put_in_reg vec) lane (vector_size in)) (imm $I64 0)))

;;;; Rules for `sextend` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;

;; General rule for extending input to an output which fits in a single
;; register.
(rule (lower (has_type (fits_in_64 out) (sextend x @ (value_type in))))
      (extend x $true (ty_bits in) (ty_bits out)))

;; Extraction of a vector lane automatically extends as necessary, so we can
;; skip an explicit extending instruction.
(rule (lower (has_type (fits_in_64 out)
                       (sextend (extractlane vec @ (value_type in)
                                             (u8_from_uimm8 lane)))))
      (mov_from_vec_signed (put_in_reg vec)
                           lane
                           (vector_size in)
                           (size_from_ty out)))

;; 64-bit to 128-bit only needs to sign-extend the input to the upper bits.
(rule (lower (has_type $I128 (sextend x)))
      (let ((lo Reg (put_in_reg_sext64 x))
            (hi Reg (asr_imm $I64 lo (imm_shift_from_u8 63))))
        (value_regs lo hi)))

;; Like above where vector extraction automatically zero-extends extending to
;; i128 only requires generating a 0 constant for the upper bits.
;;
;; Note that `mov_from_vec_signed` doesn't exist for i64x2, so that's
;; specifically excluded here.
(rule (lower (has_type $I128
                       (sextend (extractlane vec @ (value_type in @ (not_i64x2))
                                             (u8_from_uimm8 lane)))))
      (let ((lo Reg (mov_from_vec_signed (put_in_reg vec)
                                         lane
                                         (vector_size in)
                                         (size_from_ty $I64)))
            (hi Reg (asr_imm $I64 lo (imm_shift_from_u8 63))))
        (value_regs lo hi)))

;; Extension from an extraction of i64x2 into i128.
(rule (lower (has_type $I128
                       (sextend (extractlane vec @ (value_type $I64X2)
                                             (u8_from_uimm8 lane)))))
      (let ((lo Reg (mov_from_vec (put_in_reg vec)
                                  lane
                                  (VectorSize.Size64x2)))
            (hi Reg (asr_imm $I64 lo (imm_shift_from_u8 63))))
        (value_regs lo hi)))

;;;; Rules for `bnot` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;

;; Base case using `orn` between two registers.
;;
;; Note that bitwise negation is implemented here as
;;
;;      NOT rd, rm ==> ORR_NOT rd, zero, rm
(rule (lower (has_type (fits_in_64 ty) (bnot x)))
      (orr_not ty (zero_reg) x))

;; Special case to use `orr_not_shift` if it's a `bnot` of a const-left-shifted
;; value.
(rule (lower (has_type (fits_in_64 ty)
                       (bnot (ishl x (iconst (lshl_from_imm64 <ty amt))))))
      (orr_not_shift ty (zero_reg) x amt))

;; Implementation of `bnot` for `i128`.
(rule (lower (has_type $I128 (bnot x)))
      (let ((x_regs ValueRegs x)
            (x_lo Reg (value_regs_get x_regs 0))
            (x_hi Reg (value_regs_get x_regs 1))
            (new_lo Reg (orr_not $I64 (zero_reg) x_lo))
            (new_hi Reg (orr_not $I64 (zero_reg) x_hi)))
        (value_regs new_lo new_hi)))

;; Implementation of `bnot` for vector types.
(rule (lower (has_type (vec128 ty) (bnot x)))
      (not x (vector_size ty)))

;;;; Rules for `band` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;

(rule (lower (has_type (fits_in_32 ty) (band x y)))
      (alu_rs_imm_logic_commutative (ALUOp.And) ty x y))

(rule (lower (has_type $I64 (band x y)))
      (alu_rs_imm_logic_commutative (ALUOp.And) $I64 x y))

(rule (lower (has_type $I128 (band x y))) (i128_alu_bitop (ALUOp.And) $I64 x y))

(rule (lower (has_type (vec128 ty) (band x y)))
      (and_vec x y (vector_size ty)))

;;;; Rules for `bor` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;

(rule (lower (has_type (fits_in_32 ty) (bor x y)))
      (alu_rs_imm_logic_commutative (ALUOp.Orr) ty x y))

(rule (lower (has_type $I64 (bor x y)))
      (alu_rs_imm_logic_commutative (ALUOp.Orr) $I64 x y))

(rule (lower (has_type $I128 (bor x y))) (i128_alu_bitop (ALUOp.Orr) $I64 x y))

(rule (lower (has_type (vec128 ty) (bor x y)))
      (orr_vec x y (vector_size ty)))

;;;; Rules for `bxor` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;

(rule (lower (has_type (fits_in_32 ty) (bxor x y)))
      (alu_rs_imm_logic_commutative (ALUOp.Eor) ty x y))

(rule (lower (has_type $I64 (bxor x y)))
      (alu_rs_imm_logic_commutative (ALUOp.Eor) $I64 x y))

(rule (lower (has_type $I128 (bxor x y))) (i128_alu_bitop (ALUOp.Eor) $I64 x y))

(rule (lower (has_type (vec128 ty) (bxor x y)))
      (eor_vec x y (vector_size ty)))

;;;; Rules for `band_not` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;

(rule (lower (has_type (fits_in_32 ty) (band_not x y)))
      (alu_rs_imm_logic (ALUOp.AndNot) ty x y))

(rule (lower (has_type $I64 (band_not x y)))
      (alu_rs_imm_logic (ALUOp.AndNot) $I64 x y))

(rule (lower (has_type $I128 (band_not x y))) (i128_alu_bitop (ALUOp.AndNot) $I64 x y))

(rule (lower (has_type (vec128 ty) (band_not x y)))
      (bic_vec x y (vector_size ty)))

;;;; Rules for `bor_not` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;

(rule (lower (has_type (fits_in_32 ty) (bor_not x y)))
      (alu_rs_imm_logic (ALUOp.OrrNot) ty x y))

(rule (lower (has_type $I64 (bor_not x y)))
      (alu_rs_imm_logic (ALUOp.OrrNot) $I64 x y))

(rule (lower (has_type $I128 (bor_not x y))) (i128_alu_bitop (ALUOp.OrrNot) $I64 x y))

;;;; Rules for `bxor_not` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;

(rule (lower (has_type (fits_in_32 ty) (bxor_not x y)))
      (alu_rs_imm_logic (ALUOp.EorNot) $I32 x y))

(rule (lower (has_type $I64 (bxor_not x y)))
      (alu_rs_imm_logic (ALUOp.EorNot) $I64 x y))

(rule (lower (has_type $I128 (bxor_not x y))) (i128_alu_bitop (ALUOp.EorNot) $I64 x y))

;;;; Rules for `ishl` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;

;; Shift for i8/i16/i32.
(rule (lower (has_type (fits_in_32 ty) (ishl x y)))
      (do_shift (ALUOp.Lsl) ty x y))

;; Shift for i64.
(rule (lower (has_type $I64 (ishl x y)))
      (do_shift (ALUOp.Lsl) $I64 x y))

;; Shift for i128.
(rule (lower (has_type $I128 (ishl x y)))
      (lower_shl128 x (value_regs_get y 0)))

;;     lsl     lo_lshift, src_lo, amt
;;     lsl     hi_lshift, src_hi, amt
;;     mvn     inv_amt, amt
;;     lsr     lo_rshift, src_lo, #1
;;     lsr     lo_rshift, lo_rshift, inv_amt
;;     orr     maybe_hi, hi_lshift, lo_rshift
;;     tst     amt, #0x40
;;     csel    dst_hi, lo_lshift, maybe_hi, ne
;;     csel    dst_lo, xzr, lo_lshift, ne
(decl lower_shl128 (ValueRegs Reg) ValueRegs)
(rule (lower_shl128 src amt)
      (let ((src_lo Reg (value_regs_get src 0))
            (src_hi Reg (value_regs_get src 1))
            (lo_lshift Reg (lsl $I64 src_lo amt))
            (hi_lshift Reg (lsl $I64 src_hi amt))
            (inv_amt Reg (orr_not $I32 (zero_reg) amt))
            (lo_rshift Reg (lsr $I64 (lsr_imm $I64 src_lo (imm_shift_from_u8 1))
                                inv_amt))
          (maybe_hi Reg (orr $I64 hi_lshift lo_rshift))
        )
        (with_flags
         (tst_imm $I64 amt (u64_into_imm_logic $I64 64))
         (consumes_flags_concat
          (csel (Cond.Ne) (zero_reg) lo_lshift)
          (csel (Cond.Ne) lo_lshift maybe_hi)))))

;; Shift for vector types.
(rule (lower (has_type (vec128 ty) (ishl x y)))
      (let ((size VectorSize (vector_size ty))
            (shift Reg (vec_dup y size)))
        (sshl x shift size)))

;; Helper function to emit a shift operation with the opcode specified and
;; the output type specified. The `Reg` provided is shifted by the `Value`
;; given.
;;
;; Note that this automatically handles the clif semantics of masking the
;; shift amount where necessary.
(decl do_shift (ALUOp Type Reg Value) Reg)

;; 8/16-bit shift base case.
;;
;; When shifting for amounts larger than the size of the type, the CLIF shift
;; instructions implement a "wrapping" behaviour, such that an i8 << 8 is
;; equivalent to i8 << 0
;;
;; On i32 and i64 types this matches what the aarch64 spec does, but on smaller
;; types (i16, i8) we need to do this manually, so we wrap the shift amount
;; with an AND instruction
(rule (do_shift op (fits_in_16 ty) x y)
      (let ((shift_amt Reg (value_regs_get y 0))
            (masked_shift_amt Reg (and_imm $I32 shift_amt (shift_mask ty))))
        (alu_rrr op $I32 x masked_shift_amt)))

(decl shift_mask (Type) ImmLogic)
(extern constructor shift_mask shift_mask)

;; 32/64-bit shift base cases.
(rule (do_shift op $I32 x y) (alu_rrr op $I32 x (value_regs_get y 0)))
(rule (do_shift op $I64 x y) (alu_rrr op $I64 x (value_regs_get y 0)))

;; Special case for shifting by a constant value where the value can fit into an
;; `ImmShift`.
;;
;; Note that this rule explicitly has a higher priority than the others
;; to ensure it's attempted first, otherwise the type-based filters on the
;; previous rules seem to take priority over this rule.
(rule 1 (do_shift op ty x (iconst (imm_shift_from_imm64 <ty shift)))
      (alu_rr_imm_shift op ty x shift))

;;;; Rules for `ushr` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;

;; Shift for i8/i16/i32.
(rule (lower (has_type (fits_in_32 ty) (ushr x y)))
      (do_shift (ALUOp.Lsr) ty (put_in_reg_zext32 x) y))

;; Shift for i64.
(rule (lower (has_type $I64 (ushr x y)))
      (do_shift (ALUOp.Lsr) $I64 (put_in_reg_zext64 x) y))

;; Shift for i128.
(rule (lower (has_type $I128 (ushr x y)))
      (lower_ushr128 x (value_regs_get y 0)))

;; Vector shifts.
(rule (lower (has_type (vec128 ty) (ushr x y)))
      (let ((size VectorSize (vector_size ty))
            (shift Reg (vec_dup (sub $I32 (zero_reg) y) size)))
        (ushl x shift size)))

;;     lsr       lo_rshift, src_lo, amt
;;     lsr       hi_rshift, src_hi, amt
;;     mvn       inv_amt, amt
;;     lsl       hi_lshift, src_hi, #1
;;     lsl       hi_lshift, hi_lshift, inv_amt
;;     tst       amt, #0x40
;;     orr       maybe_lo, lo_rshift, hi_lshift
;;     csel      dst_hi, xzr, hi_rshift, ne
;;     csel      dst_lo, hi_rshift, maybe_lo, ne
(decl lower_ushr128 (ValueRegs Reg) ValueRegs)
(rule (lower_ushr128 src amt)
      (let ((src_lo Reg (value_regs_get src 0))
            (src_hi Reg (value_regs_get src 1))
            (lo_rshift Reg (lsr $I64 src_lo amt))
            (hi_rshift Reg (lsr $I64 src_hi amt))
            
            (inv_amt Reg (orr_not $I32 (zero_reg) amt))
            (hi_lshift Reg (lsl $I64 (lsl_imm $I64 src_hi (imm_shift_from_u8 1))
                                inv_amt))
          (maybe_lo Reg (orr $I64 lo_rshift hi_lshift))
        )
        (with_flags
         (tst_imm $I64 amt (u64_into_imm_logic $I64 64))
         (consumes_flags_concat
          (csel (Cond.Ne) hi_rshift maybe_lo)
          (csel (Cond.Ne) (zero_reg) hi_rshift)))))

;;;; Rules for `sshr` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;

;; Shift for i8/i16/i32.
(rule (lower (has_type (fits_in_32 ty) (sshr x y)))
      (do_shift (ALUOp.Asr) ty (put_in_reg_sext32 x) y))

;; Shift for i64.
(rule (lower (has_type $I64 (sshr x y)))
      (do_shift (ALUOp.Asr) $I64 (put_in_reg_sext64 x) y))

;; Shift for i128.
(rule (lower (has_type $I128 (sshr x y)))
      (lower_sshr128 x (value_regs_get y 0)))

;; Vector shifts.
;;
;; Note that right shifts are implemented with a negative left shift.
(rule (lower (has_type (vec128 ty) (sshr x y)))
      (let ((size VectorSize (vector_size ty))
            (shift Reg (vec_dup (sub $I32 (zero_reg) y) size)))
        (sshl x shift size)))

;;     lsr       lo_rshift, src_lo, amt
;;     asr       hi_rshift, src_hi, amt
;;     mvn       inv_amt, amt
;;     lsl       hi_lshift, src_hi, #1
;;     lsl       hi_lshift, hi_lshift, inv_amt
;;     asr       hi_sign, src_hi, #63
;;     orr       maybe_lo, lo_rshift, hi_lshift
;;     tst       amt, #0x40
;;     csel      dst_hi, hi_sign, hi_rshift, ne
;;     csel      dst_lo, hi_rshift, maybe_lo, ne
(decl lower_sshr128 (ValueRegs Reg) ValueRegs)
(rule (lower_sshr128 src amt)
      (let ((src_lo Reg (value_regs_get src 0))
            (src_hi Reg (value_regs_get src 1))
            (lo_rshift Reg (lsr $I64 src_lo amt))
            (hi_rshift Reg (asr $I64 src_hi amt))

            (inv_amt Reg (orr_not $I32 (zero_reg) amt))
            (hi_lshift Reg (lsl $I64 (lsl_imm $I64 src_hi (imm_shift_from_u8 1))
                                inv_amt))
          (hi_sign Reg (asr_imm $I64 src_hi (imm_shift_from_u8 63)))
          (maybe_lo Reg (orr $I64 lo_rshift hi_lshift))
        )
        (with_flags
         (tst_imm $I64 amt (u64_into_imm_logic $I64 64))
         (consumes_flags_concat
          (csel (Cond.Ne) hi_rshift maybe_lo)
          (csel (Cond.Ne) hi_sign hi_rshift)))))

;;;; Rules for `rotl` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;

;; General 8/16-bit case.
(rule (lower (has_type (fits_in_16 ty) (rotl x y)))
      (let ((neg_shift Reg (sub $I32 (zero_reg) y)))
        (small_rotr ty (put_in_reg_zext32 x) neg_shift)))

;; Specialization for the 8/16-bit case when the rotation amount is an immediate.
(rule (lower (has_type (fits_in_16 ty) (rotl x (iconst (imm_shift_from_imm64 <ty n)))))
      (small_rotr_imm ty (put_in_reg_zext32 x) (negate_imm_shift ty n)))

;; aarch64 doesn't have a left-rotate instruction, but a left rotation of K
;; places is effectively a right rotation of N - K places, if N is the integer's
;; bit size. We implement left rotations with this trick.
;;
;; Note that when negating the shift amount here the upper bits are ignored
;; by the rotr instruction, meaning that we'll still left-shift by the desired
;; amount.

;; General 32-bit case.
(rule (lower (has_type $I32 (rotl x y)))
      (let ((neg_shift Reg (sub $I32 (zero_reg) y)))
        (a64_rotr $I32 x neg_shift)))

;; General 64-bit case.
(rule (lower (has_type $I64 (rotl x y)))
      (let ((neg_shift Reg (sub $I64 (zero_reg) y)))
        (a64_rotr $I64 x neg_shift)))

;; Specialization for the 32-bit case when the rotation amount is an immediate.
(rule (lower (has_type $I32 (rotl x (iconst (imm_shift_from_imm64 <$I32 n)))))
      (a64_rotr_imm $I32 x (negate_imm_shift $I32 n)))

;; Specialization for the 64-bit case when the rotation amount is an immediate.
(rule (lower (has_type $I64 (rotl x (iconst (imm_shift_from_imm64 <$I64 n)))))
      (a64_rotr_imm $I64 x (negate_imm_shift $I64 n)))

(decl negate_imm_shift (Type ImmShift) ImmShift)
(extern constructor negate_imm_shift negate_imm_shift)

;; General 128-bit case.
;;
;; TODO: much better codegen is possible with a constant amount.
(rule (lower (has_type $I128 (rotl x y)))
      (let ((val ValueRegs x)
            (amt Reg (value_regs_get y 0))
            (neg_amt Reg (sub $I64 (imm $I64 128) amt))
            (lshift ValueRegs (lower_shl128 val amt))
            (rshift ValueRegs (lower_ushr128 val neg_amt)))
        (value_regs
          (orr $I64 (value_regs_get lshift 0) (value_regs_get rshift 0))
          (orr $I64 (value_regs_get lshift 1) (value_regs_get rshift 1)))))

;;;; Rules for `rotr` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;

;; General 8/16-bit case.
(rule (lower (has_type (fits_in_16 ty) (rotr x y)))
      (small_rotr ty (put_in_reg_zext32 x) y))

;; General 32-bit case.
(rule (lower (has_type $I32 (rotr x y)))
      (a64_rotr $I32 x y))

;; General 64-bit case.
(rule (lower (has_type $I64 (rotr x y)))
      (a64_rotr $I64 x y))

;; Specialization for the 8/16-bit case when the rotation amount is an immediate.
(rule (lower (has_type (fits_in_16 ty) (rotr x (iconst (imm_shift_from_imm64 <ty n)))))
      (small_rotr_imm ty (put_in_reg_zext32 x) n))

;; Specialization for the 32-bit case when the rotation amount is an immediate.
(rule (lower (has_type $I32 (rotr x (iconst (imm_shift_from_imm64 <$I32 n)))))
      (a64_rotr_imm $I32 x n))

;; Specialization for the 64-bit case when the rotation amount is an immediate.
(rule (lower (has_type $I64 (rotr x (iconst (imm_shift_from_imm64 <$I64 n)))))
      (a64_rotr_imm $I64 x n))

;; For a < 32-bit rotate-right, we synthesize this as:
;;
;;    rotr rd, val, amt
;;
;;       =>
;;
;;    and masked_amt, amt, <bitwidth - 1>
;;    sub tmp_sub, masked_amt, <bitwidth>
;;    sub neg_amt, zero, tmp_sub  ; neg
;;    lsr val_rshift, val, masked_amt
;;    lsl val_lshift, val, neg_amt
;;    orr rd, val_lshift val_rshift
(decl small_rotr (Type Reg Reg) Reg)
(rule (small_rotr ty val amt)
      (let ((masked_amt Reg (and_imm $I32 amt (rotr_mask ty)))
            (tmp_sub Reg (sub_imm $I32 masked_amt (u8_into_imm12 (ty_bits ty))))
            (neg_amt Reg (sub $I32 (zero_reg) tmp_sub))
            (val_rshift Reg (lsr $I32 val masked_amt))
            (val_lshift Reg (lsl $I32 val neg_amt)))
        (orr $I32 val_lshift val_rshift)))

(decl rotr_mask (Type) ImmLogic)
(extern constructor rotr_mask rotr_mask)

;; For a constant amount, we can instead do:
;;
;;    rotr rd, val, #amt
;;
;;       =>
;;
;;    lsr val_rshift, val, #<amt>
;;    lsl val_lshift, val, <bitwidth - amt>
;;    orr rd, val_lshift, val_rshift
(decl small_rotr_imm (Type Reg ImmShift) Reg)
(rule (small_rotr_imm ty val amt)
      (let ((val_rshift Reg (lsr_imm $I32 val amt))
            (val_lshift Reg (lsl_imm $I32 val (rotr_opposite_amount ty amt))))
        (orr $I32 val_lshift val_rshift)))

(decl rotr_opposite_amount (Type ImmShift) ImmShift)
(extern constructor rotr_opposite_amount rotr_opposite_amount)

;; General 128-bit case.
;;
;; TODO: much better codegen is possible with a constant amount.
(rule (lower (has_type $I128 (rotr x y)))
      (let ((val ValueRegs x)
            (amt Reg (value_regs_get y 0))
            (neg_amt Reg (sub $I64 (imm $I64 128) amt))
            (rshift ValueRegs (lower_ushr128 val amt))
            (lshift ValueRegs (lower_shl128 val neg_amt))
            (hi Reg (orr $I64 (value_regs_get rshift 1) (value_regs_get lshift 1)))
            (lo Reg (orr $I64 (value_regs_get rshift 0) (value_regs_get lshift 0))))
        (value_regs lo hi)))

;;;; Rules for `bitrev` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;

;; Reversing an 8-bit value with a 32-bit bitrev instruction will place
;; the reversed result in the highest 8 bits, so we need to shift them down into
;; place.
(rule (lower (has_type $I8 (bitrev x)))
      (lsr_imm $I32 (rbit $I32 x) (imm_shift_from_u8 24)))

;; Reversing an 16-bit value with a 32-bit bitrev instruction will place
;; the reversed result in the highest 16 bits, so we need to shift them down into
;; place.
(rule (lower (has_type $I16 (bitrev x)))
      (lsr_imm $I32 (rbit $I32 x) (imm_shift_from_u8 16)))

(rule (lower (has_type $I128 (bitrev x)))
      (let ((val ValueRegs x)
            (lo_rev Reg (rbit $I64 (value_regs_get val 0)))
            (hi_rev Reg (rbit $I64 (value_regs_get val 1))))
        (value_regs hi_rev lo_rev)))

(rule (lower (has_type ty (bitrev x)))
      (rbit ty x))


;;;; Rules for `clz` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;

(rule (lower (has_type $I8 (clz x)))
      (sub_imm $I32 (a64_clz $I32 (put_in_reg_zext32 x)) (u8_into_imm12 24)))

(rule (lower (has_type $I16 (clz x)))
      (sub_imm $I32 (a64_clz $I32 (put_in_reg_zext32 x)) (u8_into_imm12 16)))

(rule (lower (has_type $I128 (clz x)))
      (lower_clz128 x))

(rule (lower (has_type ty (clz x)))
      (a64_clz ty x))

;; clz hi_clz, hi
;; clz lo_clz, lo
;; lsr tmp, hi_clz, #6
;; madd dst_lo, lo_clz, tmp, hi_clz
;; mov  dst_hi, 0
(decl lower_clz128 (ValueRegs) ValueRegs)
(rule (lower_clz128 val)
      (let ((hi_clz Reg (a64_clz $I64 (value_regs_get val 1)))
            (lo_clz Reg (a64_clz $I64 (value_regs_get val 0)))
            (tmp Reg (lsr_imm $I64 hi_clz (imm_shift_from_u8 6))))
        (value_regs (madd64 lo_clz tmp hi_clz) (imm $I64 0))))

;;;; Rules for `ctz` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;

;; Note that all `ctz` instructions are implemented by reversing the bits and
;; then using a `clz` instruction since the tail zeros are the same as the
;; leading zeros of the reversed value.

(rule (lower (has_type $I8 (ctz x)))
      (a64_clz $I32 (orr_imm $I32 (rbit $I32 x) (u64_into_imm_logic $I32 0x800000))))

(rule (lower (has_type $I16 (ctz x)))
      (a64_clz $I32 (orr_imm $I32 (rbit $I32 x) (u64_into_imm_logic $I32 0x8000))))

(rule (lower (has_type $I128 (ctz x)))
      (let ((val ValueRegs x)
            (lo Reg (rbit $I64 (value_regs_get val 0)))
            (hi Reg (rbit $I64 (value_regs_get val 1))))
        (lower_clz128 (value_regs hi lo))))

(rule (lower (has_type ty (ctz x)))
      (a64_clz ty (rbit ty x)))

;;;; Rules for `cls` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;

(rule (lower (has_type $I8 (cls x)))
      (sub_imm $I32 (a64_cls $I32 (put_in_reg_zext32 x)) (u8_into_imm12 24)))

(rule (lower (has_type $I16 (cls x)))
      (sub_imm $I32 (a64_cls $I32 (put_in_reg_zext32 x)) (u8_into_imm12 16)))

;; cls lo_cls, lo
;; cls hi_cls, hi
;; eon sign_eq_eor, hi, lo
;; lsr sign_eq, sign_eq_eor, #63
;; madd lo_sign_bits, out_lo, sign_eq, sign_eq
;; cmp hi_cls, #63
;; csel maybe_lo, lo_sign_bits, xzr, eq
;; add  out_lo, maybe_lo, hi_cls
;; mov  out_hi, 0
(rule (lower (has_type $I128 (cls x)))
      (let ((val ValueRegs x)
            (lo Reg (value_regs_get val 0))
            (hi Reg (value_regs_get val 1))
            (lo_cls Reg (a64_cls $I64 lo))
            (hi_cls Reg (a64_cls $I64 hi))
            (sign_eq_eon Reg (eon $I64 hi lo))
            (sign_eq Reg (lsr_imm $I64 sign_eq_eon (imm_shift_from_u8 63)))
            (lo_sign_bits Reg (madd64 lo_cls sign_eq sign_eq))
            (maybe_lo Reg (with_flags_reg
                           (cmp64_imm hi_cls (u8_into_imm12 63))
                           (csel (Cond.Eq) lo_sign_bits (zero_reg)))))
        (value_regs (add $I64 maybe_lo hi_cls) (imm $I64 0))))

(rule (lower (has_type ty (cls x)))
      (a64_cls ty x))

;;;; Rules for `popcnt` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;

;; The implementation of `popcnt` for scalar types is done by moving the value
;; into a vector register, using the `cnt` instruction, and then collating the
;; result back into a normal register.
;;
;; The general sequence emitted here is
;;
;;     fmov tmp, in_lo
;;     if ty == i128:
;;         mov tmp.d[1], in_hi
;;
;;     cnt tmp.16b, tmp.16b / cnt tmp.8b, tmp.8b
;;     addv tmp, tmp.16b / addv tmp, tmp.8b / addp tmp.8b, tmp.8b, tmp.8b / (no instruction for 8-bit inputs)
;;
;;     umov out_lo, tmp.b[0]
;;     if ty == i128:
;;         mov out_hi, 0

(rule (lower (has_type $I8 (popcnt x)))
      (let ((tmp Reg (mov_to_fpu x (ScalarSize.Size32)))
            (nbits Reg (vec_cnt tmp (VectorSize.Size8x8))))
        (mov_from_vec nbits 0 (VectorSize.Size8x16))))

;; Note that this uses `addp` instead of `addv` as it's usually cheaper.
(rule (lower (has_type $I16 (popcnt x)))
      (let ((tmp Reg (mov_to_fpu x (ScalarSize.Size32)))
            (nbits Reg (vec_cnt tmp (VectorSize.Size8x8)))
            (added Reg (addp nbits nbits (VectorSize.Size8x8))))
        (mov_from_vec added 0 (VectorSize.Size8x16))))

(rule (lower (has_type $I32 (popcnt x)))
      (let ((tmp Reg (mov_to_fpu x (ScalarSize.Size32)))
            (nbits Reg (vec_cnt tmp (VectorSize.Size8x8)))
            (added Reg (addv nbits (VectorSize.Size8x8))))
        (mov_from_vec added 0 (VectorSize.Size8x16))))

(rule (lower (has_type $I64 (popcnt x)))
      (let ((tmp Reg (mov_to_fpu x (ScalarSize.Size64)))
            (nbits Reg (vec_cnt tmp (VectorSize.Size8x8)))
            (added Reg (addv nbits (VectorSize.Size8x8))))
        (mov_from_vec added 0 (VectorSize.Size8x16))))

(rule (lower (has_type $I128 (popcnt x)))
      (let ((val ValueRegs x)
            (tmp_half Reg (mov_to_fpu (value_regs_get val 0) (ScalarSize.Size64)))
            (tmp Reg (mov_to_vec tmp_half (value_regs_get val 1) 1 (VectorSize.Size64x2)))
            (nbits Reg (vec_cnt tmp (VectorSize.Size8x16)))
            (added Reg (addv nbits (VectorSize.Size8x16))))
        (value_regs (mov_from_vec added 0 (VectorSize.Size8x16)) (imm $I64 0))))

(rule (lower (has_type $I8X16 (popcnt x)))
      (vec_cnt x (VectorSize.Size8x16)))