;; x86-64 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 partial lower (Inst) InstOutput)

;; A variant of the main lowering constructor term, used for branches.
;; The only difference is that it gets an extra argument holding a vector
;; of branch targets to be used.
(decl partial lower_branch (Inst MachLabelSlice) Unit)

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

;; `i64` and smaller.
(rule (lower (has_type (fits_in_64 ty)
                       (iconst (u64_from_imm64 x))))
      (imm ty x))

;; `i128`
(rule 1 (lower (has_type $I128
                       (iconst (u64_from_imm64 x))))
      (value_regs (imm $I64 x)
                  (imm $I64 0)))

;;;; Rules for `f32const` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;

(rule (lower (f32const (u64_from_ieee32 x)))
      (imm $F32 x))

;;;; Rules for `f64const` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;

(rule (lower (f64const (u64_from_ieee64 x)))
      (imm $F64 x))

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

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

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

;; `i64` and smaller.

;; Add two registers.
(rule -5 (lower (has_type (fits_in_64 ty)
                       (iadd x y)))
      (x64_add ty x y))

;; The above case handles when the rhs is an immediate or a sinkable load, but
;; additionally add lhs meets these criteria.

(rule -4 (lower (has_type (fits_in_64 ty)
                       (iadd (simm32_from_value x) y)))
      (x64_add ty y x))
(rule -3 (lower (has_type (fits_in_64 ty)
                       (iadd (sinkable_load x) y)))
      (x64_add ty y x))

;; SSE.

(rule (lower (has_type (multi_lane 8 16)
                       (iadd x y)))
      (x64_paddb x y))

(rule (lower (has_type (multi_lane 16 8)
                       (iadd x y)))
      (x64_paddw x y))

(rule (lower (has_type (multi_lane 32 4)
                       (iadd x y)))
      (x64_paddd x y))

(rule (lower (has_type (multi_lane 64 2)
                       (iadd x y)))
      (x64_paddq x y))

;; `i128`
(rule 1 (lower (has_type $I128 (iadd x y)))
      ;; Get the high/low registers for `x`.
      (let ((x_regs ValueRegs x)
            (x_lo Gpr (value_regs_get_gpr x_regs 0))
            (x_hi Gpr (value_regs_get_gpr x_regs 1)))
        ;; Get the high/low registers for `y`.
        (let ((y_regs ValueRegs y)
              (y_lo Gpr (value_regs_get_gpr y_regs 0))
              (y_hi Gpr (value_regs_get_gpr y_regs 1)))
          ;; Do an add followed by an add-with-carry.
          (with_flags (x64_add_with_flags_paired $I64 x_lo y_lo)
                      (x64_adc_paired $I64 x_hi y_hi)))))

;;;; Rules for `iadd_cout` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;

;; TODO: i8 and i16 support. Requires either learning how to encode ALU
;; operations on values narrower than 32-bits (better code; big change) or doing
;; the same extend-to-32-bits trick that aarch64 does (worse code; small
;; change).

(rule (lower (iadd_cout x y @ (value_type (ty_32_or_64 ty))))
      (let ((results ValueRegs (with_flags (x64_add_with_flags_paired ty x y)
                                           (x64_setcc_paired (CC.O)))))
        (output_pair (value_regs_get results 0)
                     (value_regs_get results 1))))

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

(rule (lower (has_type (multi_lane 8 16)
                       (sadd_sat x y)))
      (x64_paddsb x y))

(rule (lower (has_type (multi_lane 16 8)
                       (sadd_sat x y)))
      (x64_paddsw x y))

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

(rule (lower (has_type (multi_lane 8 16)
                       (uadd_sat x y)))
      (x64_paddusb x y))

(rule (lower (has_type (multi_lane 16 8)
                       (uadd_sat x y)))
      (x64_paddusw x y))

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

;; `i64` and smaller.

;; Sub two registers.
(rule -3 (lower (has_type (fits_in_64 ty)
                       (isub x y)))
      (x64_sub ty x y))

;; SSE.

(rule (lower (has_type (multi_lane 8 16)
                       (isub x y)))
      (x64_psubb x y))

(rule (lower (has_type (multi_lane 16 8)
                       (isub x y)))
      (x64_psubw x y))

(rule (lower (has_type (multi_lane 32 4)
                       (isub x y)))
      (x64_psubd x y))

(rule (lower (has_type (multi_lane 64 2)
                       (isub x y)))
      (x64_psubq x y))

;; `i128`
(rule 1 (lower (has_type $I128 (isub x y)))
      ;; Get the high/low registers for `x`.
      (let ((x_regs ValueRegs x)
            (x_lo Gpr (value_regs_get_gpr x_regs 0))
            (x_hi Gpr (value_regs_get_gpr x_regs 1)))
        ;; Get the high/low registers for `y`.
        (let ((y_regs ValueRegs y)
              (y_lo Gpr (value_regs_get_gpr y_regs 0))
              (y_hi Gpr (value_regs_get_gpr y_regs 1)))
          ;; Do a sub followed by an sub-with-borrow.
          (with_flags (x64_sub_with_flags_paired $I64 x_lo y_lo)
                      (x64_sbb_paired $I64 x_hi y_hi)))))

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

(rule (lower (has_type (multi_lane 8 16)
                       (ssub_sat x y)))
      (x64_psubsb x y))

(rule (lower (has_type (multi_lane 16 8)
                       (ssub_sat x y)))
      (x64_psubsw x y))

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

(rule (lower (has_type (multi_lane 8 16)
                       (usub_sat x y)))
      (x64_psubusb x y))

(rule (lower (has_type (multi_lane 16 8)
                       (usub_sat x y)))
      (x64_psubusw x y))

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

;; `{i,b}64` and smaller.

;; And two registers.
(rule 0 (lower (has_type ty (band x y)))
      (if (ty_int_ref_scalar_64 ty))
      (x64_and ty x y))

;; The above case automatically handles when the rhs is an immediate or a
;; sinkable load, but additionally handle the lhs here.

(rule 1 (lower (has_type ty (band (sinkable_load x) y)))
      (if (ty_int_ref_scalar_64 ty))
      (x64_and ty y x))

(rule 2 (lower (has_type ty (band (simm32_from_value x) y)))
      (if (ty_int_ref_scalar_64 ty))
      (x64_and ty y x))

;; f32 and f64

(rule 5 (lower (has_type (ty_scalar_float ty) (band x y)))
      (sse_and ty x y))

;; SSE.

(decl sse_and (Type Xmm XmmMem) Xmm)
(rule (sse_and $F32X4 x y) (x64_andps x y))
(rule (sse_and $F64X2 x y) (x64_andpd x y))
(rule (sse_and $F32 x y) (x64_andps x y))
(rule (sse_and $F64 x y) (x64_andpd x y))
(rule -1 (sse_and (multi_lane _bits _lanes) x y) (x64_pand x y))

(rule 6 (lower (has_type ty @ (multi_lane _bits _lanes)
                       (band x y)))
      (sse_and ty x y))

;; `i128`.

(rule 7 (lower (has_type $I128 (band x y)))
      (let ((x_regs ValueRegs x)
            (x_lo Gpr (value_regs_get_gpr x_regs 0))
            (x_hi Gpr (value_regs_get_gpr x_regs 1))
            (y_regs ValueRegs y)
            (y_lo Gpr (value_regs_get_gpr y_regs 0))
            (y_hi Gpr (value_regs_get_gpr y_regs 1)))
        (value_gprs (x64_and $I64 x_lo y_lo)
                    (x64_and $I64 x_hi y_hi))))

;; Specialized lowerings for `(band x (bnot y))` which is additionally produced
;; by Cranelift's `band_not` instruction that is legalized into the simpler
;; forms early on.

(decl sse_and_not (Type Xmm XmmMem) Xmm)
(rule (sse_and_not $F32X4 x y) (x64_andnps x y))
(rule (sse_and_not $F64X2 x y) (x64_andnpd x y))
(rule -1 (sse_and_not (multi_lane _bits _lanes) x y) (x64_pandn x y))

;; Note the flipping of operands below as we're match
;;
;;   (band x (bnot y))
;;
;; while x86 does
;;
;;   pandn(x, y) = and(not(x), y)
(rule 8 (lower (has_type ty @ (multi_lane _bits _lane) (band x (bnot y))))
      (sse_and_not ty y x))
(rule 9 (lower (has_type ty @ (multi_lane _bits _lane) (band (bnot y) x)))
      (sse_and_not ty y x))

(rule 10 (lower (has_type ty @ (use_bmi1 $true) (band x (bnot y))))
      (if (ty_int_ref_scalar_64 ty))
      ;; the first argument is the one that gets inverted with andn
      (x64_andn ty y x))
(rule 11 (lower (has_type ty @ (use_bmi1 $true) (band (bnot y) x)))
      (if (ty_int_ref_scalar_64 ty))
      (x64_andn ty y x))


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

;; `{i,b}64` and smaller.

;; Or two registers.
(rule 0 (lower (has_type ty (bor x y)))
      (if (ty_int_ref_scalar_64 ty))
      (x64_or ty x y))

;; Handle immediates/sinkable loads on the lhs in addition to the automatic
;; handling of the rhs above

(rule 1 (lower (has_type ty (bor (sinkable_load x) y)))
      (if (ty_int_ref_scalar_64 ty))
      (x64_or ty y x))

(rule 2 (lower (has_type ty (bor (simm32_from_value x) y)))
      (if (ty_int_ref_scalar_64 ty))
      (x64_or ty y x))

;; f32 and f64

(rule 5 (lower (has_type (ty_scalar_float ty) (bor x y)))
      (sse_or ty x y))

;; SSE.

(decl sse_or (Type Xmm XmmMem) Xmm)
(rule (sse_or $F32X4 x y) (x64_orps x y))
(rule (sse_or $F64X2 x y) (x64_orpd x y))
(rule (sse_or $F32 x y) (x64_orps x y))
(rule (sse_or $F64 x y) (x64_orpd x y))
(rule -1 (sse_or (multi_lane _bits _lanes) x y) (x64_por x y))

(rule 6 (lower (has_type ty @ (multi_lane _bits _lanes)
                       (bor x y)))
      (sse_or ty x y))

;; `{i,b}128`.

(decl or_i128 (ValueRegs ValueRegs) ValueRegs)
(rule (or_i128 x y)
      (let ((x_lo Gpr (value_regs_get_gpr x 0))
            (x_hi Gpr (value_regs_get_gpr x 1))
            (y_lo Gpr (value_regs_get_gpr y 0))
            (y_hi Gpr (value_regs_get_gpr y 1)))
        (value_gprs (x64_or $I64 x_lo y_lo)
                    (x64_or $I64 x_hi y_hi))))

(rule 7 (lower (has_type $I128 (bor x y)))
      (or_i128 x y))

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

;; `{i,b}64` and smaller.

;; Xor two registers.
(rule 0 (lower (has_type ty (bxor x y)))
      (if (ty_int_ref_scalar_64 ty))
      (x64_xor ty x y))

;; Handle xor with lhs immediates/sinkable loads in addition to the automatic
;; handling of the rhs above.

(rule 1 (lower (has_type ty (bxor (sinkable_load x) y)))
      (if (ty_int_ref_scalar_64 ty))
      (x64_xor ty y x))

(rule 4 (lower (has_type ty (bxor (simm32_from_value x) y)))
      (if (ty_int_ref_scalar_64 ty))
      (x64_xor ty y x))

;; f32 and f64

(rule 5 (lower (has_type (ty_scalar_float ty) (bxor x y)))
      (sse_xor ty x y))

;; SSE.

(rule 6 (lower (has_type ty @ (multi_lane _bits _lanes) (bxor x y)))
      (sse_xor ty x y))

;; `{i,b}128`.

(rule 7 (lower (has_type $I128 (bxor x y)))
      (let ((x_regs ValueRegs x)
            (x_lo Gpr (value_regs_get_gpr x_regs 0))
            (x_hi Gpr (value_regs_get_gpr x_regs 1))
            (y_regs ValueRegs y)
            (y_lo Gpr (value_regs_get_gpr y_regs 0))
            (y_hi Gpr (value_regs_get_gpr y_regs 1)))
        (value_gprs (x64_xor $I64 x_lo y_lo)
                    (x64_xor $I64 x_hi y_hi))))

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

;; `i64` and smaller.

(rule -1 (lower (has_type (fits_in_64 ty) (ishl src amt)))
      (x64_shl ty src (put_masked_in_imm8_gpr amt ty)))

;; `i128`.

(decl shl_i128 (ValueRegs Gpr) ValueRegs)
(rule (shl_i128 src amt)
      ;; Unpack the registers that make up the 128-bit value being shifted.
      (let ((src_lo Gpr (value_regs_get_gpr src 0))
            (src_hi Gpr (value_regs_get_gpr src 1))
            ;; Do two 64-bit shifts.
            (lo_shifted Gpr (x64_shl $I64 src_lo amt))
            (hi_shifted Gpr (x64_shl $I64 src_hi amt))
            ;; `src_lo >> (64 - amt)` are the bits to carry over from the lo
            ;; into the hi.
            (carry Gpr (x64_shr $I64
                            src_lo
                            (x64_sub $I64
                                 (imm $I64 64)
                                 amt)))
            (zero Gpr (imm $I64 0))
            ;; Nullify the carry if we are shifting in by a multiple of 128.
            (carry_ Gpr (with_flags_reg (x64_test (OperandSize.Size64)
                                              (RegMemImm.Imm 127)
                                              amt)
                                        (cmove $I64
                                               (CC.Z)
                                               zero
                                               carry)))
            ;; Add the carry into the high half.
            (hi_shifted_ Gpr (x64_or $I64 carry_ hi_shifted)))
        ;; Combine the two shifted halves. However, if we are shifting by >= 64
        ;; (modulo 128), then the low bits are zero and the high bits are our
        ;; low bits.
        (with_flags (x64_test (OperandSize.Size64) (RegMemImm.Imm 64) amt)
                    (consumes_flags_concat
                     (cmove $I64 (CC.Z) lo_shifted zero)
                     (cmove $I64 (CC.Z) hi_shifted_ lo_shifted)))))

(rule (lower (has_type $I128 (ishl src amt)))
      ;; NB: Only the low bits of `amt` matter since we logically mask the shift
      ;; amount to the value's bit width.
      (let ((amt_ Gpr (lo_gpr amt)))
        (shl_i128 src amt_)))

;; SSE.

;; Since the x86 instruction set does not have any 8x16 shift instructions (even
;; in higher feature sets like AVX), we lower the `ishl.i8x16` to a sequence of
;; instructions. The basic idea, whether the amount to shift by is an immediate
;; or not, is to use a 16x8 shift and then mask off the incorrect bits to 0s.
(rule (lower (has_type ty @ $I8X16 (ishl src amt)))
      (let (
            ;; Mask the amount to ensure wrapping behaviour
            (masked_amt RegMemImm (mask_xmm_shift ty amt))
            ;; Shift `src` using 16x8. Unfortunately, a 16x8 shift will only be
            ;; correct for half of the lanes; the others must be fixed up with
            ;; the mask below.
            (unmasked Xmm (x64_psllw src (mov_rmi_to_xmm masked_amt)))
            (mask_addr SyntheticAmode (ishl_i8x16_mask masked_amt))
            (mask Reg (x64_load $I8X16 mask_addr (ExtKind.None))))
        (sse_and $I8X16 unmasked (RegMem.Reg mask))))

;; Get the address of the mask to use when fixing up the lanes that weren't
;; correctly generated by the 16x8 shift.
(decl ishl_i8x16_mask (RegMemImm) SyntheticAmode)

;; When the shift amount is known, we can statically (i.e. at compile time)
;; determine the mask to use and only emit that.
(decl ishl_i8x16_mask_for_const (u32) SyntheticAmode)
(extern constructor ishl_i8x16_mask_for_const ishl_i8x16_mask_for_const)
(rule (ishl_i8x16_mask (RegMemImm.Imm amt))
      (ishl_i8x16_mask_for_const amt))

;; Otherwise, we must emit the entire mask table and dynamically (i.e. at run
;; time) find the correct mask offset in the table. We use `lea` to find the
;; base address of the mask table and then complex addressing to offset to the
;; right mask: `base_address + amt << 4`
(decl ishl_i8x16_mask_table () SyntheticAmode)
(extern constructor ishl_i8x16_mask_table ishl_i8x16_mask_table)
(rule (ishl_i8x16_mask (RegMemImm.Reg amt))
      (let ((mask_table SyntheticAmode (ishl_i8x16_mask_table))
            (base_mask_addr Gpr (x64_lea mask_table))
            (mask_offset Gpr (x64_shl $I64 amt
                                  (imm8_to_imm8_gpr 4))))
        (amode_imm_reg_reg_shift 0
                                 base_mask_addr
                                 mask_offset
                                 0)))

(rule (ishl_i8x16_mask (RegMemImm.Mem amt))
      (ishl_i8x16_mask (RegMemImm.Reg (x64_load $I64 amt (ExtKind.None)))))

;; 16x8, 32x4, and 64x2 shifts can each use a single instruction, once the shift amount is masked.

(rule (lower (has_type ty @ $I16X8 (ishl src amt)))
      (x64_psllw src (mov_rmi_to_xmm (mask_xmm_shift ty amt))))

(rule (lower (has_type ty @ $I32X4 (ishl src amt)))
      (x64_pslld src (mov_rmi_to_xmm (mask_xmm_shift ty amt))))

(rule (lower (has_type ty @ $I64X2 (ishl src amt)))
      (x64_psllq src (mov_rmi_to_xmm (mask_xmm_shift ty amt))))

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

;; `i64` and smaller.

(rule -1 (lower (has_type (fits_in_64 ty) (ushr src amt)))
      (let ((src_ Gpr (extend_to_gpr src ty (ExtendKind.Zero))))
        (x64_shr ty src_ (put_masked_in_imm8_gpr amt ty))))

;; `i128`.

(decl shr_i128 (ValueRegs Gpr) ValueRegs)
(rule (shr_i128 src amt)
      ;; Unpack the lo/hi halves of `src`.
      (let ((src_lo Gpr (value_regs_get_gpr src 0))
            (src_hi Gpr (value_regs_get_gpr src 1))
            ;; Do a shift on each half.
            (lo_shifted Gpr (x64_shr $I64 src_lo amt))
            (hi_shifted Gpr (x64_shr $I64 src_hi amt))
            ;; `src_hi << (64 - amt)` are the bits to carry over from the hi
            ;; into the lo.
            (carry Gpr (x64_shl $I64
                            src_hi
                            (x64_sub $I64
                                 (imm $I64 64)
                                 amt)))
            ;; Share the zero value to reduce register pressure
            (zero Gpr (imm $I64 0))

            ;; Nullify the carry if we are shifting by a multiple of 128.
            (carry_ Gpr (with_flags_reg (x64_test (OperandSize.Size64) (RegMemImm.Imm 127) amt)
                                        (cmove $I64 (CC.Z) zero carry)))
            ;; Add the carry bits into the lo.
            (lo_shifted_ Gpr (x64_or $I64 carry_ lo_shifted)))
        ;; Combine the two shifted halves. However, if we are shifting by >= 64
        ;; (modulo 128), then the hi bits are zero and the lo bits are what
        ;; would otherwise be our hi bits.
        (with_flags (x64_test (OperandSize.Size64) (RegMemImm.Imm 64) amt)
                    (consumes_flags_concat
                     (cmove $I64 (CC.Z) lo_shifted_ hi_shifted)
                     (cmove $I64 (CC.Z) hi_shifted zero)))))

(rule (lower (has_type $I128 (ushr src amt)))
      ;; NB: Only the low bits of `amt` matter since we logically mask the shift
      ;; amount to the value's bit width.
      (let ((amt_ Gpr (lo_gpr amt)))
        (shr_i128 src amt_)))

;; SSE.

;; There are no 8x16 shifts in x64. Do the same 16x8-shift-and-mask thing we do
;; with 8x16 `ishl`.
(rule (lower (has_type ty @ $I8X16 (ushr src amt)))
      (let (
            ;; Mask the amount to ensure wrapping behaviour
            (masked_amt RegMemImm (mask_xmm_shift ty amt))
            ;; Shift `src` using 16x8. Unfortunately, a 16x8 shift will only be
            ;; correct for half of the lanes; the others must be fixed up with
            ;; the mask below.
            (unmasked Xmm (x64_psrlw src (mov_rmi_to_xmm masked_amt)))
            (mask_addr SyntheticAmode (ushr_i8x16_mask masked_amt))
            (mask Reg (x64_load $I8X16 mask_addr (ExtKind.None))))
        (sse_and $I8X16
                 unmasked
                 (RegMem.Reg mask))))

;; Get the address of the mask to use when fixing up the lanes that weren't
;; correctly generated by the 16x8 shift.
(decl ushr_i8x16_mask (RegMemImm) SyntheticAmode)

;; When the shift amount is known, we can statically (i.e. at compile time)
;; determine the mask to use and only emit that.
(decl ushr_i8x16_mask_for_const (u32) SyntheticAmode)
(extern constructor ushr_i8x16_mask_for_const ushr_i8x16_mask_for_const)
(rule (ushr_i8x16_mask (RegMemImm.Imm amt))
      (ushr_i8x16_mask_for_const amt))

;; Otherwise, we must emit the entire mask table and dynamically (i.e. at run
;; time) find the correct mask offset in the table. We use `lea` to find the
;; base address of the mask table and then complex addressing to offset to the
;; right mask: `base_address + amt << 4`
(decl ushr_i8x16_mask_table () SyntheticAmode)
(extern constructor ushr_i8x16_mask_table ushr_i8x16_mask_table)
(rule (ushr_i8x16_mask (RegMemImm.Reg amt))
      (let ((mask_table SyntheticAmode (ushr_i8x16_mask_table))
            (base_mask_addr Gpr (x64_lea mask_table))
            (mask_offset Gpr (x64_shl $I64
                                  amt
                                  (imm8_to_imm8_gpr 4))))
        (amode_imm_reg_reg_shift 0
                                 base_mask_addr
                                 mask_offset
                                 0)))

(rule (ushr_i8x16_mask (RegMemImm.Mem amt))
      (ushr_i8x16_mask (RegMemImm.Reg (x64_load $I64 amt (ExtKind.None)))))

;; 16x8, 32x4, and 64x2 shifts can each use a single instruction, once the shift amount is masked.

(rule (lower (has_type ty @ $I16X8 (ushr src amt)))
      (x64_psrlw src (mov_rmi_to_xmm (mask_xmm_shift ty amt))))

(rule (lower (has_type ty @ $I32X4 (ushr src amt)))
      (x64_psrld src (mov_rmi_to_xmm (mask_xmm_shift ty amt))))

(rule (lower (has_type ty @ $I64X2 (ushr src amt)))
      (x64_psrlq src (mov_rmi_to_xmm (mask_xmm_shift ty amt))))

(decl mask_xmm_shift (Type Value) RegMemImm)
(rule (mask_xmm_shift ty amt)
      (gpr_to_reg (x64_and $I64 amt (RegMemImm.Imm (shift_mask ty)))))
(rule 1 (mask_xmm_shift ty (iconst n))
      (RegMemImm.Imm (shift_amount_masked ty n)))

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

;; `i64` and smaller.

(rule -1 (lower (has_type (fits_in_64 ty) (sshr src amt)))
      (let ((src_ Gpr (extend_to_gpr src ty (ExtendKind.Sign))))
        (x64_sar ty src_ (put_masked_in_imm8_gpr amt ty))))

;; `i128`.

(decl sar_i128 (ValueRegs Gpr) ValueRegs)
(rule (sar_i128 src amt)
      ;; Unpack the low/high halves of `src`.
      (let ((src_lo Gpr (value_regs_get_gpr src 0))
            (src_hi Gpr (value_regs_get_gpr src 1))
            ;; Do a shift of each half. NB: the low half uses an unsigned shift
            ;; because its MSB is not a sign bit.
            (lo_shifted Gpr (x64_shr $I64 src_lo amt))
            (hi_shifted Gpr (x64_sar $I64 src_hi amt))
            ;; `src_hi << (64 - amt)` are the bits to carry over from the low
            ;; half to the high half.
            (carry Gpr (x64_shl $I64
                            src_hi
                            (x64_sub $I64
                                 (imm $I64 64)
                                 amt)))
            ;; Nullify the carry if we are shifting by a multiple of 128.
            (carry_ Gpr (with_flags_reg (x64_test (OperandSize.Size64) (RegMemImm.Imm 127) amt)
                                        (cmove $I64 (CC.Z) (imm $I64 0) carry)))
            ;; Add the carry into the low half.
            (lo_shifted_ Gpr (x64_or $I64 lo_shifted carry_))
            ;; Get all sign bits.
            (sign_bits Gpr (x64_sar $I64 src_hi (imm8_to_imm8_gpr 63))))
        ;; Combine the two shifted halves. However, if we are shifting by >= 64
        ;; (modulo 128), then the hi bits are all sign bits and the lo bits are
        ;; what would otherwise be our hi bits.
        (with_flags (x64_test (OperandSize.Size64) (RegMemImm.Imm 64) amt)
                    (consumes_flags_concat
                     (cmove $I64 (CC.Z) lo_shifted_ hi_shifted)
                     (cmove $I64 (CC.Z) hi_shifted sign_bits)))))

(rule (lower (has_type $I128 (sshr src amt)))
      ;; NB: Only the low bits of `amt` matter since we logically mask the shift
      ;; amount to the value's bit width.
      (let ((amt_ Gpr (lo_gpr amt)))
        (sar_i128 src amt_)))

;; SSE.

;; Since the x86 instruction set does not have an 8x16 shift instruction and the
;; approach used for `ishl` and `ushr` cannot be easily used (the masks do not
;; preserve the sign), we use a different approach here: separate the low and
;; high lanes, shift them separately, and merge them into the final result.
;;
;; Visually, this looks like the following, where `src.i8x16 = [s0, s1, ...,
;; s15]:
;;
;;   lo.i16x8 = [(s0, s0), (s1, s1), ..., (s7, s7)]
;;   shifted_lo.i16x8 = shift each lane of `low`
;;   hi.i16x8 = [(s8, s8), (s9, s9), ..., (s15, s15)]
;;   shifted_hi.i16x8 = shift each lane of `high`
;;   result = [s0'', s1'', ..., s15'']
(rule (lower (has_type ty @ $I8X16 (sshr src amt @ (value_type amt_ty))))
      (let ((src_ Xmm (put_in_xmm src))
            ;; Mask the amount to ensure wrapping behaviour
            (masked_amt RegMemImm (mask_xmm_shift ty amt))
            ;; In order for `packsswb` later to only use the high byte of each
            ;; 16x8 lane, we shift right an extra 8 bits, relying on `psraw` to
            ;; fill in the upper bits appropriately.
            (lo Xmm (x64_punpcklbw src_ src_))
            (hi Xmm (x64_punpckhbw src_ src_))
            (amt_ XmmMemImm (sshr_i8x16_bigger_shift amt_ty masked_amt))
            (shifted_lo Xmm (x64_psraw lo amt_))
            (shifted_hi Xmm (x64_psraw hi amt_)))
        (x64_packsswb shifted_lo shifted_hi)))

(decl sshr_i8x16_bigger_shift (Type RegMemImm) XmmMemImm)
(rule (sshr_i8x16_bigger_shift _ty (RegMemImm.Imm i))
      (xmm_mem_imm_new (RegMemImm.Imm (u32_add i 8))))
(rule (sshr_i8x16_bigger_shift ty (RegMemImm.Reg r))
      (mov_rmi_to_xmm (RegMemImm.Reg (x64_add ty
                                          r
                                          (RegMemImm.Imm 8)))))
(rule (sshr_i8x16_bigger_shift ty rmi @ (RegMemImm.Mem _m))
      (mov_rmi_to_xmm (RegMemImm.Reg (x64_add ty
                                          (imm ty 8)
                                          rmi))))

;; `sshr.{i16x8,i32x4}` can be a simple `psra{w,d}`, we just have to make sure
;; that if the shift amount is in a register, it is in an XMM register.

(rule (lower (has_type ty @ $I16X8 (sshr src amt)))
      (x64_psraw src (mov_rmi_to_xmm (mask_xmm_shift ty amt))))

(rule (lower (has_type ty @ $I32X4 (sshr src amt)))
      (x64_psrad src (mov_rmi_to_xmm (mask_xmm_shift ty amt))))

;; The `sshr.i64x2` CLIF instruction has no single x86 instruction in the older
;; feature sets. Newer ones like AVX512VL + AVX512F include `vpsraq`, a 128-bit
;; instruction that would fit here, but this backend does not currently have
;; support for EVEX encodings. To remedy this, we extract each 64-bit lane to a
;; GPR, shift each using a scalar instruction, and insert the shifted values
;; back in the `dst` XMM register.
;;
;; (TODO: when EVEX support is available, add an alternate lowering here).
(rule (lower (has_type $I64X2 (sshr src amt)))
      (let ((src_ Xmm (put_in_xmm src))
            (lo Gpr (x64_pextrq src_ 0))
            (hi Gpr (x64_pextrq src_ 1))
            (amt_ Imm8Gpr (put_masked_in_imm8_gpr amt $I64))
            (shifted_lo Gpr (x64_sar $I64 lo amt_))
            (shifted_hi Gpr (x64_sar $I64 hi amt_)))
        (make_i64x2_from_lanes shifted_lo
                               shifted_hi)))

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

;; `i64` and smaller: we can rely on x86's rotate-amount masking since
;;  we operate on the whole register. For const's we mask the constant.

(rule -1 (lower (has_type (fits_in_64 ty) (rotl src amt)))
        (x64_rotl ty src (put_masked_in_imm8_gpr amt ty)))


;; `i128`.

(rule (lower (has_type $I128 (rotl src amt)))
      (let ((src_ ValueRegs src)
            ;; NB: Only the low bits of `amt` matter since we logically mask the
            ;; rotation amount to the value's bit width.
            (amt_ Gpr (lo_gpr amt)))
        (or_i128 (shl_i128 src_ amt_)
                 (shr_i128 src_ (x64_sub $I64
                                     (imm $I64 128)
                                     amt_)))))

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

;; `i64` and smaller: we can rely on x86's rotate-amount masking since
;;  we operate on the whole register. For const's we mask the constant.

(rule -1 (lower (has_type (fits_in_64 ty) (rotr src amt)))
        (x64_rotr ty src (put_masked_in_imm8_gpr amt ty)))


;; `i128`.

(rule (lower (has_type $I128 (rotr src amt)))
      (let ((src_ ValueRegs src)
            ;; NB: Only the low bits of `amt` matter since we logically mask the
            ;; rotation amount to the value's bit width.
            (amt_ Gpr (lo_gpr amt)))
        (or_i128 (shr_i128 src_ amt_)
                 (shl_i128 src_ (x64_sub $I64
                                     (imm $I64 128)
                                     amt_)))))

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

;; `i64` and smaller.

(rule -1 (lower (has_type (fits_in_64 ty) (ineg x)))
      (x64_neg ty x))

(rule -2 (lower (has_type $I128 (ineg x)))
      ;; Get the high/low registers for `x`.
      (let ((regs ValueRegs x)
            (lo Gpr (value_regs_get_gpr regs 0))
            (hi Gpr (value_regs_get_gpr regs 1)))
        ;; Do a neg followed by an sub-with-borrow.
        (with_flags (x64_neg_paired $I64 lo)
                    (x64_sbb_paired $I64 (imm $I64 0) hi))))

;; SSE.

(rule (lower (has_type $I8X16 (ineg x)))
      (x64_psubb (imm $I8X16 0) x))

(rule (lower (has_type $I16X8 (ineg x)))
      (x64_psubw (imm $I16X8 0) x))

(rule (lower (has_type $I32X4 (ineg x)))
      (x64_psubd (imm $I32X4 0) x))

(rule (lower (has_type $I64X2 (ineg x)))
      (x64_psubq (imm $I64X2 0) x))

;;;; Rules for `avg_round` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;

(rule (lower (has_type (multi_lane 8 16)
                       (avg_round x y)))
      (x64_pavgb x y))

(rule (lower (has_type (multi_lane 16 8)
                       (avg_round x y)))
      (x64_pavgw x y))

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

;; `i64` and smaller.

;; Multiply two registers.
(rule -5 (lower (has_type (fits_in_64 ty) (imul x y)))
      (x64_mul ty x y))

;; Handle multiplication where the lhs is an immediate or sinkable load in
;; addition to the automatic rhs handling above.

(rule -4 (lower (has_type (fits_in_64 ty)
                       (imul (simm32_from_value x) y)))
      (x64_mul ty y x))
(rule -3 (lower (has_type (fits_in_64 ty)
                       (imul (sinkable_load x) y)))
      (x64_mul ty y x))

;; `i128`.

;; mul:
;;   dst_lo = lhs_lo * rhs_lo
;;   dst_hi = umulhi(lhs_lo, rhs_lo) +
;;            lhs_lo * rhs_hi +
;;            lhs_hi * rhs_lo
;;
;; so we emit:
;;   lo_hi = mul x_lo, y_hi
;;   hi_lo = mul x_hi, y_lo
;;   hilo_hilo = add lo_hi, hi_lo
;;   dst_lo:hi_lolo = mulhi_u x_lo, y_lo
;;   dst_hi = add hilo_hilo, hi_lolo
;;   return (dst_lo, dst_hi)
(rule 2 (lower (has_type $I128 (imul x y)))
      ;; Put `x` into registers and unpack its hi/lo halves.
      (let ((x_regs ValueRegs x)
            (x_lo Gpr (value_regs_get_gpr x_regs 0))
            (x_hi Gpr (value_regs_get_gpr x_regs 1))
            ;; Put `y` into registers and unpack its hi/lo halves.
            (y_regs ValueRegs y)
            (y_lo Gpr (value_regs_get_gpr y_regs 0))
            (y_hi Gpr (value_regs_get_gpr y_regs 1))
            ;; lo_hi = mul x_lo, y_hi
            (lo_hi Gpr (x64_mul $I64 x_lo y_hi))
            ;; hi_lo = mul x_hi, y_lo
            (hi_lo Gpr (x64_mul $I64 x_hi y_lo))
            ;; hilo_hilo = add lo_hi, hi_lo
            (hilo_hilo Gpr (x64_add $I64 lo_hi hi_lo))
            ;; dst_lo:hi_lolo = mulhi_u x_lo, y_lo
            (mul_regs ValueRegs (mulhi_u $I64 x_lo y_lo))
            (dst_lo Gpr (value_regs_get_gpr mul_regs 0))
            (hi_lolo Gpr (value_regs_get_gpr mul_regs 1))
            ;; dst_hi = add hilo_hilo, hi_lolo
            (dst_hi Gpr (x64_add $I64 hilo_hilo hi_lolo)))
        (value_gprs dst_lo dst_hi)))

;; SSE.

;; (No i8x16 multiply.)

(rule (lower (has_type (multi_lane 16 8) (imul x y)))
      (x64_pmullw x y))

(rule (lower (has_type (multi_lane 32 4) (imul x y)))
      (x64_pmulld x y))

;; With AVX-512 we can implement `i64x2` multiplication with a single
;; instruction.
(rule 3 (lower (has_type (and (avx512vl_enabled $true)
                            (avx512dq_enabled $true)
                            (multi_lane 64 2))
                       (imul x y)))
      (x64_vpmullq x y))

;; Otherwise, for i64x2 multiplication we describe a lane A as being composed of
;; a 32-bit upper half "Ah" and a 32-bit lower half "Al". The 32-bit long hand
;; multiplication can then be written as:
;;
;;    Ah Al
;; *  Bh Bl
;;    -----
;;    Al * Bl
;; + (Ah * Bl) << 32
;; + (Al * Bh) << 32
;;
;; So for each lane we will compute:
;;
;;   A * B  = (Al * Bl) + ((Ah * Bl) + (Al * Bh)) << 32
;;
;; Note, the algorithm will use `pmuldq` which operates directly on the lower
;; 32-bit (`Al` or `Bl`) of a lane and writes the result to the full 64-bits of
;; the lane of the destination. For this reason we don't need shifts to isolate
;; the lower 32-bits, however, we will need to use shifts to isolate the high
;; 32-bits when doing calculations, i.e., `Ah == A >> 32`.
(rule (lower (has_type (multi_lane 64 2)
                       (imul a b)))
      (let ((a0 Xmm a)
            (b0 Xmm b)
            ;; a_hi = A >> 32
            (a_hi Xmm (x64_psrlq a0 (RegMemImm.Imm 32)))
            ;; ah_bl = Ah * Bl
            (ah_bl Xmm (x64_pmuludq a_hi b0))
            ;; b_hi = B >> 32
            (b_hi Xmm (x64_psrlq b0 (RegMemImm.Imm 32)))
            ;; al_bh = Al * Bh
            (al_bh Xmm (x64_pmuludq a0 b_hi))
            ;; aa_bb = ah_bl + al_bh
            (aa_bb Xmm (x64_paddq ah_bl al_bh))
            ;; aa_bb_shifted = aa_bb << 32
            (aa_bb_shifted Xmm (x64_psllq aa_bb (RegMemImm.Imm 32)))
            ;; al_bl = Al * Bl
            (al_bl Xmm (x64_pmuludq a0 b0)))
        ;; al_bl + aa_bb_shifted
        (x64_paddq al_bl aa_bb_shifted)))

;; Special case for `i16x8.extmul_high_i8x16_s`.
(rule 1 (lower (has_type (multi_lane 16 8)
                       (imul (swiden_high (and (value_type (multi_lane 8 16))
                                               x))
                             (swiden_high (and (value_type (multi_lane 8 16))
                                               y)))))
      (let ((x1 Xmm x)
            (x2 Xmm (x64_palignr x1 x1 8 (OperandSize.Size32)))
            (x3 Xmm (x64_pmovsxbw x2))
            (y1 Xmm y)
            (y2 Xmm (x64_palignr y1 y1 8 (OperandSize.Size32)))
            (y3 Xmm (x64_pmovsxbw y2)))
        (x64_pmullw x3 y3)))

;; Special case for `i32x4.extmul_high_i16x8_s`.
(rule 1 (lower (has_type (multi_lane 32 4)
                       (imul (swiden_high (and (value_type (multi_lane 16 8))
                                               x))
                             (swiden_high (and (value_type (multi_lane 16 8))
                                               y)))))
      (let ((x2 Xmm x)
            (y2 Xmm y)
            (lo Xmm (x64_pmullw x2 y2))
            (hi Xmm (x64_pmulhw x2 y2)))
        (x64_punpckhwd lo hi)))

;; Special case for `i64x2.extmul_high_i32x4_s`.
(rule 1 (lower (has_type (multi_lane 64 2)
                       (imul (swiden_high (and (value_type (multi_lane 32 4))
                                               x))
                             (swiden_high (and (value_type (multi_lane 32 4))
                                               y)))))
      (let ((x2 Xmm (x64_pshufd x 0xFA))
            (y2 Xmm (x64_pshufd y 0xFA)))
        (x64_pmuldq x2 y2)))

;; Special case for `i16x8.extmul_low_i8x16_s`.
(rule 1 (lower (has_type (multi_lane 16 8)
                       (imul (swiden_low (and (value_type (multi_lane 8 16))
                                              x))
                             (swiden_low (and (value_type (multi_lane 8 16))
                                              y)))))
      (let ((x2 Xmm (x64_pmovsxbw x))
            (y2 Xmm (x64_pmovsxbw y)))
        (x64_pmullw x2 y2)))

;; Special case for `i32x4.extmul_low_i16x8_s`.
(rule 1 (lower (has_type (multi_lane 32 4)
                       (imul (swiden_low (and (value_type (multi_lane 16 8))
                                              x))
                             (swiden_low (and (value_type (multi_lane 16 8))
                                              y)))))
      (let ((x2 Xmm x)
            (y2 Xmm y)
            (lo Xmm (x64_pmullw x2 y2))
            (hi Xmm (x64_pmulhw x2 y2)))
        (x64_punpcklwd lo hi)))

;; Special case for `i64x2.extmul_low_i32x4_s`.
(rule 1 (lower (has_type (multi_lane 64 2)
                       (imul (swiden_low (and (value_type (multi_lane 32 4))
                                              x))
                             (swiden_low (and (value_type (multi_lane 32 4))
                                              y)))))
      (let ((x2 Xmm (x64_pshufd x 0x50))
            (y2 Xmm (x64_pshufd y 0x50)))
        (x64_pmuldq x2 y2)))

;; Special case for `i16x8.extmul_high_i8x16_u`.
(rule 1 (lower (has_type (multi_lane 16 8)
                       (imul (uwiden_high (and (value_type (multi_lane 8 16))
                                               x))
                             (uwiden_high (and (value_type (multi_lane 8 16))
                                               y)))))
      (let ((x1 Xmm x)
            (x2 Xmm (x64_palignr x1 x1 8 (OperandSize.Size32)))
            (x3 Xmm (x64_pmovzxbw x2))
            (y1 Xmm y)
            (y2 Xmm (x64_palignr y1 y1 8 (OperandSize.Size32)))
            (y3 Xmm (x64_pmovzxbw y2)))
        (x64_pmullw x3 y3)))

;; Special case for `i32x4.extmul_high_i16x8_u`.
(rule 1 (lower (has_type (multi_lane 32 4)
                       (imul (uwiden_high (and (value_type (multi_lane 16 8))
                                               x))
                             (uwiden_high (and (value_type (multi_lane 16 8))
                                               y)))))
      (let ((x2 Xmm x)
            (y2 Xmm y)
            (lo Xmm (x64_pmullw x2 y2))
            (hi Xmm (x64_pmulhuw x2 y2)))
        (x64_punpckhwd lo hi)))

;; Special case for `i64x2.extmul_high_i32x4_u`.
(rule 1 (lower (has_type (multi_lane 64 2)
                       (imul (uwiden_high (and (value_type (multi_lane 32 4))
                                               x))
                             (uwiden_high (and (value_type (multi_lane 32 4))
                                               y)))))
      (let ((x2 Xmm (x64_pshufd x 0xFA))
            (y2 Xmm (x64_pshufd y 0xFA)))
        (x64_pmuludq x2 y2)))

;; Special case for `i16x8.extmul_low_i8x16_u`.
(rule 1 (lower (has_type (multi_lane 16 8)
                       (imul (uwiden_low (and (value_type (multi_lane 8 16))
                                              x))
                             (uwiden_low (and (value_type (multi_lane 8 16))
                                              y)))))
      (let ((x2 Xmm (x64_pmovzxbw x))
            (y2 Xmm (x64_pmovzxbw y)))
        (x64_pmullw x2 y2)))

;; Special case for `i32x4.extmul_low_i16x8_u`.
(rule 1 (lower (has_type (multi_lane 32 4)
                       (imul (uwiden_low (and (value_type (multi_lane 16 8))
                                              x))
                             (uwiden_low (and (value_type (multi_lane 16 8))
                                              y)))))
      (let ((x2 Xmm x)
            (y2 Xmm y)
            (lo Xmm (x64_pmullw x2 y2))
            (hi Xmm (x64_pmulhuw x2 y2)))
        (x64_punpcklwd lo hi)))

;; Special case for `i64x2.extmul_low_i32x4_u`.
(rule 1 (lower (has_type (multi_lane 64 2)
                       (imul (uwiden_low (and (value_type (multi_lane 32 4))
                                              x))
                             (uwiden_low (and (value_type (multi_lane 32 4))
                                              y)))))
      (let ((x2 Xmm (x64_pshufd x 0x50))
            (y2 Xmm (x64_pshufd y 0x50)))
        (x64_pmuludq x2 y2)))

;;;; Rules for `iabs` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;

(rule (lower (has_type $I8X16 (iabs x)))
      (x64_pabsb x))

(rule (lower (has_type $I16X8 (iabs x)))
      (x64_pabsw x))

(rule (lower (has_type $I32X4 (iabs x)))
      (x64_pabsd x))

;; When AVX512 is available, we can use a single `vpabsq` instruction.
(rule 1 (lower (has_type (and (avx512vl_enabled $true)
                            (avx512f_enabled $true)
                            $I64X2)
                       (iabs x)))
      (x64_vpabsq x))

;; Otherwise, we use a separate register, `neg`, to contain the results of `0 -
;; x` and then blend in those results with `blendvpd` if the MSB of `neg` was
;; set to 1 (i.e. if `neg` was negative or, conversely, if `x` was originally
;; positive).
(rule (lower (has_type $I64X2 (iabs x)))
      (let ((rx Xmm x)
            (neg Xmm (x64_psubq (imm $I64X2 0) rx)))
        (x64_blendvpd neg rx neg)))

;; `i64` and smaller.

(rule -1 (lower (has_type (fits_in_64 ty) (iabs x)))
      (let ((src Gpr x)
            (neg ProducesFlags (x64_neg_paired ty src))
            ;; Manually extract the result from the neg, then ignore
            ;; it below, since we need to pass it into the cmove
            ;; before we pass the cmove to with_flags_reg.
            (neg_result Gpr (produces_flags_get_reg neg))
            ;; When the neg instruction sets the sign flag,
            ;; takes the original (non-negative) value.
            (cmove ConsumesFlags (cmove ty (CC.S) src neg_result)))
        (with_flags_reg (produces_flags_ignore neg) cmove)))

;;;; Rules for `fabs` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;

(rule (lower (has_type $F32 (fabs x)))
      (x64_andps x (imm $F32 0x7fffffff)))

(rule (lower (has_type $F64 (fabs x)))
      (x64_andpd x (imm $F64 0x7fffffffffffffff)))

;; Special case for `f32x4.abs`.
(rule (lower (has_type $F32X4 (fabs x)))
      (x64_andps x
             (x64_psrld (vector_all_ones)
                    (RegMemImm.Imm 1))))

;; Special case for `f64x2.abs`.
(rule (lower (has_type $F64X2 (fabs x)))
      (x64_andpd x
             (x64_psrlq (vector_all_ones)
                    (RegMemImm.Imm 1))))

;;;; Rules for `fneg` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;

(rule (lower (has_type $F32 (fneg x)))
      (x64_xorps x (imm $F32 0x80000000)))

(rule (lower (has_type $F64 (fneg x)))
      (x64_xorpd x (imm $F64 0x8000000000000000)))

(rule (lower (has_type $F32X4 (fneg x)))
      (x64_xorps x
             (x64_pslld (vector_all_ones)
                    (RegMemImm.Imm 31))))

(rule (lower (has_type $F64X2 (fneg x)))
      (x64_xorpd x
             (x64_psllq (vector_all_ones)
                    (RegMemImm.Imm 63))))

;;;; Rules for `bmask` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;

(decl lower_bmask (Type Type ValueRegs) ValueRegs)

;; Values that fit in a register
;;
;; Use the neg instruction on the input which sets the CF (carry) flag
;; to 0 if the input is 0 or 1 otherwise.
;; We then subtract the output register with itself, which always gives a 0,
;; however use the carry flag from the previous negate to generate a -1 if it
;; was nonzero.
;;
;; neg in_reg
;; sbb out_reg, out_reg
(rule 0
      (lower_bmask (fits_in_64 out_ty) (fits_in_64 in_ty) val)
      (let ((reg Gpr (value_regs_get_gpr val 0))
            (out ValueRegs (with_flags
                  (x64_neg_paired in_ty reg)
                  (x64_sbb_paired out_ty reg reg))))
        ;; Extract only the output of the sbb instruction
        (value_reg (value_regs_get out 1))))


;; If the input type is I128 we can `or` the registers, and recurse to the general case.
(rule 1
      (lower_bmask (fits_in_64 out_ty) $I128 val)
      (let ((lo Gpr (value_regs_get_gpr val 0))
            (hi Gpr (value_regs_get_gpr val 1))
            (mixed Gpr (x64_or $I64 lo hi)))
        (lower_bmask out_ty $I64 (value_reg mixed))))

;; If the output type is I128 we just duplicate the result of the I64 lowering
(rule 2
      (lower_bmask $I128 in_ty val)
      (let ((res ValueRegs (lower_bmask $I64 in_ty val))
            (res Gpr (value_regs_get_gpr res 0)))
        (value_regs res res)))


;; Call the lower_bmask rule that does all the procssing
(rule (lower (has_type out_ty (bmask x @ (value_type in_ty))))
      (lower_bmask out_ty in_ty x))

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

;; `i64` and smaller.

(rule -2 (lower (has_type ty (bnot x)))
      (if (ty_int_ref_scalar_64 ty))
      (x64_not ty x))


;; `i128`.

(decl i128_not (Value) ValueRegs)
(rule (i128_not x)
      (let ((x_regs ValueRegs x)
            (x_lo Gpr (value_regs_get_gpr x_regs 0))
            (x_hi Gpr (value_regs_get_gpr x_regs 1)))
        (value_gprs (x64_not $I64 x_lo)
                    (x64_not $I64 x_hi))))

(rule (lower (has_type $I128 (bnot x)))
      (i128_not x))

;; f32 and f64

(rule -3 (lower (has_type (ty_scalar_float ty) (bnot x)))
      (sse_xor ty x (vector_all_ones)))

;; Special case for vector-types where bit-negation is an xor against an
;; all-one value
(rule -1 (lower (has_type ty @ (multi_lane _bits _lanes) (bnot x)))
      (sse_xor ty x (vector_all_ones)))

;;;; Rules for `bitselect` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;

(rule (lower (has_type ty @ (multi_lane _bits _lanes)
                       (bitselect condition
                                  if_true
                                  if_false)))
      ;; a = and if_true, condition
      ;; b = and_not condition, if_false
      ;; or b, a
      (let ((cond_xmm Xmm condition)
            (a Xmm (sse_and ty if_true cond_xmm))
            (b Xmm (sse_and_not ty cond_xmm if_false)))
        (sse_or ty b a)))

;; If every byte of the condition is guaranteed to be all ones or all zeroes,
;; we can use x86_blend like vselect does.
(rule 1 (lower (has_type ty @ (multi_lane _bits _lanes)
                         (bitselect condition
                                    if_true
                                    if_false)))
      (if (all_ones_or_all_zeros condition))
      (x64_blend ty
                 condition
                 if_true
                 if_false))

(decl pure partial all_ones_or_all_zeros (Value) bool)
(rule (all_ones_or_all_zeros (and (icmp _ _ _) (value_type (multi_lane _ _)))) $true)
(rule (all_ones_or_all_zeros (and (fcmp _ _ _) (value_type (multi_lane _ _)))) $true)
(rule (all_ones_or_all_zeros (vconst (vconst_all_ones_or_all_zeros))) $true)

(decl pure vconst_all_ones_or_all_zeros () Constant)
(extern extractor vconst_all_ones_or_all_zeros vconst_all_ones_or_all_zeros)

;;;; Rules for `vselect` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;

(rule (lower (has_type ty @ (multi_lane _bits _lanes)
                       (vselect condition if_true if_false)))
      (x64_blend ty
                 condition
                 if_true
                 if_false))

;;;; Rules for `insertlane` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;

(rule (lower (insertlane vec @ (value_type ty) val (u8_from_uimm8 idx)))
      (vec_insert_lane ty vec val idx))

;; Helper function used below for `insertlane` but also here for other
;; lowerings.
;;
;; Note that the `Type` used here is the type of vector the insertion is
;; happening into, or the type of the first `Reg` argument.
(decl vec_insert_lane (Type Xmm RegMem u8) Xmm)

;; i8x16.replace_lane
(rule (vec_insert_lane $I8X16 vec val idx)
      (x64_pinsrb vec val idx))

;; i16x8.replace_lane
(rule (vec_insert_lane $I16X8 vec val idx)
      (x64_pinsrw vec val idx))

;; i32x4.replace_lane
(rule (vec_insert_lane $I32X4 vec val idx)
      (x64_pinsrd vec val idx))

;; i64x2.replace_lane
(rule (vec_insert_lane $I64X2 vec val idx)
      (x64_pinsrq vec val idx))

;; f32x4.replace_lane
(rule (vec_insert_lane $F32X4 vec val idx)
      (x64_insertps vec val (sse_insertps_lane_imm idx)))

;; External rust code used to calculate the immediate value to `insertps`.
(decl sse_insertps_lane_imm (u8) u8)
(extern constructor sse_insertps_lane_imm sse_insertps_lane_imm)

;; f64x2.replace_lane 0
;;
;; Here the `movsd` instruction is used specifically to specialize moving
;; into the fist lane where unlike above cases we're not using the lane
;; immediate as an immediate to the instruction itself.
;;
;; Note, though, the `movsd` has different behavior with respect to the second
;; lane of the f64x2 depending on whether the RegMem operand is a register or
;; memory. When loading from a register `movsd` preserves the upper bits, but
;; when loading from memory it zeros the upper bits. We specifically want to
;; preserve the upper bits so if a `RegMem.Mem` is passed in we need to emit
;; two `movsd` instructions. The first `movsd` (used as `xmm_unary_rm_r`) will
;; load from memory into a temp register and then the second `movsd` (modeled
;; internally as `xmm_rm_r` will merge the temp register into our `vec`
;; register.
(rule 1 (vec_insert_lane $F64X2 vec (RegMem.Reg val) 0)
      (x64_movsd_regmove vec val))
(rule (vec_insert_lane $F64X2 vec mem 0)
      (x64_movsd_regmove vec (x64_movsd_load mem)))

;; f64x2.replace_lane 1
;;
;; Here the `movlhps` instruction is used specifically to specialize moving
;; into the second lane where unlike above cases we're not using the lane
;; immediate as an immediate to the instruction itself.
(rule (vec_insert_lane $F64X2 vec val 1)
      (x64_movlhps vec (reg_mem_to_xmm_mem val)))

;;;; Rules for `smin`, `smax`, `umin`, `umax` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;

;; `i64` and smaller.

(decl cmp_and_choose (Type CC Value Value) ValueRegs)
(rule (cmp_and_choose (fits_in_64 ty) cc x y)
      (let ((size OperandSize (raw_operand_size_of_type ty))
            ;; We need to put x and y in registers explicitly because
            ;; we use the values more than once. Hence, even if these
            ;; are "unique uses" at the CLIF level and would otherwise
            ;; allow for load-op merging, here we cannot do that.
            (x_reg Reg x)
            (y_reg Reg y))
        (with_flags_reg (x64_cmp size x_reg y_reg)
                        (cmove ty cc y_reg x_reg))))

(rule -1 (lower (has_type (fits_in_64 ty) (umin x y)))
      (cmp_and_choose ty (CC.B) x y))

(rule -1 (lower (has_type (fits_in_64 ty) (umax x y)))
      (cmp_and_choose ty (CC.NB) x y))

(rule -1 (lower (has_type (fits_in_64 ty) (smin x y)))
      (cmp_and_choose ty (CC.L) x y))

(rule -1 (lower (has_type (fits_in_64 ty) (smax x y)))
      (cmp_and_choose ty (CC.NL) x y))

;; SSE `smax`.

(rule (lower (has_type $I8X16 (smax x y)))
      (x64_pmaxsb x y))

(rule (lower (has_type $I16X8 (smax x y)))
      (x64_pmaxsw x y))

(rule (lower (has_type $I32X4 (smax x y)))
      (x64_pmaxsd x y))

;; SSE `smin`.

(rule (lower (has_type $I8X16 (smin x y)))
      (x64_pminsb x y))

(rule (lower (has_type $I16X8 (smin x y)))
      (x64_pminsw x y))

(rule (lower (has_type $I32X4 (smin x y)))
      (x64_pminsd x y))

;; SSE `umax`.

(rule (lower (has_type $I8X16 (umax x y)))
      (x64_pmaxub x y))

(rule (lower (has_type $I16X8 (umax x y)))
      (x64_pmaxuw x y))

(rule (lower (has_type $I32X4 (umax x y)))
      (x64_pmaxud x y))

;; SSE `umin`.

(rule (lower (has_type $I8X16 (umin x y)))
      (x64_pminub x y))

(rule (lower (has_type $I16X8 (umin x y)))
      (x64_pminuw x y))

(rule (lower (has_type $I32X4 (umin x y)))
      (x64_pminud x y))

;;;; Rules for `trap` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;

(rule (lower (trap code))
      (side_effect (x64_ud2 code)))

;;;; Rules for `uadd_overflow_trap` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;

(rule (lower (has_type (fits_in_64 ty) (uadd_overflow_trap a b tc)))
      (with_flags
        (x64_add_with_flags_paired ty a b)
        (trap_if (CC.B) tc)))

;; Handle lhs immediates/sinkable loads in addition to the automatic rhs
;; handling of above.

(rule 1 (lower (has_type (fits_in_64 ty)
                         (uadd_overflow_trap (simm32_from_value a) b tc)))
      (with_flags
        (x64_add_with_flags_paired ty b a)
        (trap_if (CC.B) tc)))

(rule 2 (lower (has_type (fits_in_64 ty)
                         (uadd_overflow_trap (sinkable_load a) b tc)))
      (with_flags
        (x64_add_with_flags_paired ty b a)
        (trap_if (CC.B) tc)))

;;;; Rules for `resumable_trap` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;

(rule (lower (resumable_trap code))
      (side_effect (x64_ud2 code)))

;;;; Rules for `return` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;

;; N.B.: the Ret itself is generated by the ABI.
(rule (lower (return args))
      (lower_return (range 0 (value_slice_len args)) args))

;;;; Rules for `icmp` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;

(rule -2 (lower (icmp cc a @ (value_type (fits_in_64 ty)) b))
      (lower_icmp_bool (emit_cmp cc a b)))

(rule -1 (lower (icmp cc a @ (value_type $I128) b))
      (lower_icmp_bool (emit_cmp cc a b)))

;; Peephole optimization for `x < 0`, when x is a signed 64 bit value
(rule 2 (lower (has_type $I8 (icmp (IntCC.SignedLessThan) x @ (value_type $I64) (u64_from_iconst 0))))
      (x64_shr $I64 x (Imm8Reg.Imm8 63)))

;; Peephole optimization for `0 > x`, when x is a signed 64 bit value
(rule 2 (lower (has_type $I8 (icmp (IntCC.SignedGreaterThan) (u64_from_iconst 0) x @ (value_type $I64))))
      (x64_shr $I64 x (Imm8Reg.Imm8 63)))

;; Peephole optimization for `0 <= x`, when x is a signed 64 bit value
(rule 2 (lower (has_type $I8 (icmp (IntCC.SignedLessThanOrEqual) (u64_from_iconst 0) x @ (value_type $I64))))
      (x64_shr $I64 (x64_not $I64 x) (Imm8Reg.Imm8 63)))

;; Peephole optimization for `x >= 0`, when x is a signed 64 bit value
(rule 2 (lower (has_type $I8 (icmp (IntCC.SignedGreaterThanOrEqual) x @ (value_type $I64) (u64_from_iconst 0))))
      (x64_shr $I64 (x64_not $I64 x) (Imm8Reg.Imm8 63)))

;; Peephole optimization for `x < 0`, when x is a signed 32 bit value
(rule 2 (lower (has_type $I8 (icmp (IntCC.SignedLessThan) x @ (value_type $I32) (u64_from_iconst 0))))
      (x64_shr $I32 x (Imm8Reg.Imm8 31)))

;; Peephole optimization for `0 > x`, when x is a signed 32 bit value
(rule 2 (lower (has_type $I8 (icmp (IntCC.SignedGreaterThan) (u64_from_iconst 0) x @ (value_type $I32))))
      (x64_shr $I32 x (Imm8Reg.Imm8 31)))

;; Peephole optimization for `0 <= x`, when x is a signed 32 bit value
(rule 2 (lower (has_type $I8 (icmp (IntCC.SignedLessThanOrEqual) (u64_from_iconst 0) x @ (value_type $I32))))
      (x64_shr $I32 (x64_not $I64 x) (Imm8Reg.Imm8 31)))

;; Peephole optimization for `x >= 0`, when x is a signed 32 bit value
(rule 2 (lower (has_type $I8 (icmp (IntCC.SignedGreaterThanOrEqual) x @ (value_type $I32) (u64_from_iconst 0))))
      (x64_shr $I32 (x64_not $I64 x) (Imm8Reg.Imm8 31)))

;; For XMM-held values, we lower to `PCMP*` instructions, sometimes more than
;; one. To note: what is different here about the output values is that each
;; lane will be filled with all 1s or all 0s according to the comparison,
;; whereas for GPR-held values, the result will be simply 0 or 1 (upper bits
;; unset).
(rule (lower (icmp (IntCC.Equal) a @ (value_type (ty_vec128 ty)) b))
      (x64_pcmpeq ty a b))

;; To lower a not-equals comparison, we perform an equality comparison
;; (PCMPEQ*) and then invert the bits (PXOR with all 1s).
(rule (lower (icmp (IntCC.NotEqual) a @ (value_type (ty_vec128 ty)) b))
      (let ((checked Xmm (x64_pcmpeq ty a b))
            (all_ones Xmm (vector_all_ones)))
           (x64_pxor checked all_ones)))
;; Signed comparisons have a single-instruction lowering, unlike their unsigned
;; counterparts. These latter instructions use the unsigned min/max
;; (PMINU*/PMAXU*) and negate the result (PXOR with all 1s).
(rule (lower (icmp (IntCC.SignedGreaterThan) a @ (value_type (ty_vec128 ty)) b))
      (x64_pcmpgt ty a b))
(rule (lower (icmp (IntCC.SignedLessThan) a @ (value_type (ty_vec128 ty)) b))
      (x64_pcmpgt ty b a))
(rule (lower (icmp (IntCC.UnsignedGreaterThan) a @ (value_type (ty_vec128 ty)) b))
      ;; N.B.: we must manually prevent load coalescing of these operands; the
      ;; register allocator gets confused otherwise. TODO:
      ;; https://github.com/bytecodealliance/wasmtime/issues/3953.
      (let ((xmm_a Xmm (put_in_xmm a))
            (xmm_b Xmm (put_in_xmm b))
            (max Xmm (x64_pmaxu ty xmm_a xmm_b))
            (eq Xmm (x64_pcmpeq ty max xmm_b))
            (all_ones Xmm (vector_all_ones)))
           (x64_pxor eq all_ones)))
(rule (lower (icmp (IntCC.UnsignedLessThan) a @ (value_type (ty_vec128 ty)) b))
      ;; N.B.: see note above.
      (let ((xmm_a Xmm (put_in_xmm a))
            (xmm_b Xmm (put_in_xmm b))
            (min Xmm (x64_pminu ty xmm_a xmm_b))
            (eq Xmm (x64_pcmpeq ty min xmm_b))
            (all_ones Xmm (vector_all_ones)))
           (x64_pxor eq all_ones)))
;; To lower signed and unsigned *-or-equals comparisons, we find the minimum
;; number (PMIN[U|S]*) and compare that to one of the terms (PCMPEQ*). Note that
;; there is no 64x2 version of this lowering (see below).
(rule (lower (icmp (IntCC.SignedGreaterThanOrEqual) a @ (value_type (ty_vec128 ty)) b))
      (let ((max Xmm (x64_pmaxs ty a b)))
           (x64_pcmpeq ty a max)))
(rule (lower (icmp (IntCC.SignedLessThanOrEqual) a @ (value_type (ty_vec128 ty)) b))
      (let ((min Xmm (x64_pmins ty a b)))
           (x64_pcmpeq ty a min)))
(rule (lower (icmp (IntCC.UnsignedGreaterThanOrEqual) a @ (value_type (ty_vec128 ty)) b))
      (let ((max Xmm (x64_pmaxu ty a b)))
           (x64_pcmpeq ty a max)))
(rule (lower (icmp (IntCC.UnsignedLessThanOrEqual) a @ (value_type (ty_vec128 ty)) b))
      (let ((min Xmm (x64_pminu ty a b)))
           (x64_pcmpeq ty a min)))
;; The PMIN[S|U]Q instruction is only available in AVX512VL/F so we must instead
;; compare with flipped operands (PCMPGT*) and negate the result (PXOR with all
;; 1s), emitting one more instruction than the smaller-lane versions.
(rule 1 (lower (icmp (IntCC.SignedGreaterThanOrEqual) a @ (value_type $I64X2) b))
      (let ((checked Xmm (x64_pcmpgt $I64X2 b a))
            (all_ones Xmm (vector_all_ones)))
           (x64_pxor checked all_ones)))
(rule 1 (lower (icmp (IntCC.SignedLessThanOrEqual) a @ (value_type $I64X2) b))
      (let ((checked Xmm (x64_pcmpgt $I64X2 a b))
            (all_ones Xmm (vector_all_ones)))
           (x64_pxor checked all_ones)))
;; TODO: not used by WebAssembly translation
;; (rule (lower (icmp (IntCC.UnsignedGreaterThanOrEqual) a @ (value_type $I64X2) b))
;; TODO: not used by WebAssembly translation
;; (rule (lower (icmp (IntCC.UnsignedLessThanOrEqual) a @ (value_type $I64X2) b))


;;;; Rules for `fcmp` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;

;; CLIF's `fcmp` instruction always operates on XMM registers--both scalar and
;; vector. For the scalar versions, we use the flag-setting behavior of the
;; `UCOMIS*` instruction to `SETcc` a 0 or 1 in a GPR register. Note that CLIF's
;; `select` uses the same kind of flag-setting behavior but chooses values other
;; than 0 or 1.
;;
;; Checking the result of `UCOMIS*` is unfortunately difficult in some cases
;; because we do not have `SETcc` instructions that explicitly check
;; simultaneously for the condition (i.e., `eq`, `le`, `gt`, etc.) *and*
;; orderedness. Instead, we must check the flags multiple times. The UCOMIS*
;; documentation (see Intel's Software Developer's Manual, volume 2, chapter 4)
;; is helpful:
;;  - unordered assigns    Z = 1, P = 1, C = 1
;;  - greater than assigns Z = 0, P = 0, C = 0
;;  - less than assigns    Z = 0, P = 0, C = 1
;;  - equal assigns        Z = 1, P = 0, C = 0

(rule -1 (lower (fcmp cc a @ (value_type (ty_scalar_float ty)) b))
      (lower_fcmp_bool (emit_fcmp cc a b)))

;; For vector lowerings, we use `CMPP*` instructions with a 3-bit operand that
;; determines the comparison to make. Note that comparisons that succeed will
;; fill the lane with 1s; comparisons that do not will fill the lane with 0s.

(rule (lower (fcmp (FloatCC.Equal) a @ (value_type (ty_vec128 ty)) b))
      (x64_cmpp ty a b (FcmpImm.Equal)))
(rule (lower (fcmp (FloatCC.NotEqual) a @ (value_type (ty_vec128 ty)) b))
      (x64_cmpp ty a b (FcmpImm.NotEqual)))
(rule (lower (fcmp (FloatCC.LessThan) a @ (value_type (ty_vec128 ty)) b))
      (x64_cmpp ty a b (FcmpImm.LessThan)))
(rule (lower (fcmp (FloatCC.LessThanOrEqual) a @ (value_type (ty_vec128 ty)) b))
      (x64_cmpp ty a b (FcmpImm.LessThanOrEqual)))
(rule (lower (fcmp (FloatCC.Ordered) a @ (value_type (ty_vec128 ty)) b))
      (x64_cmpp ty a b (FcmpImm.Ordered)))
(rule (lower (fcmp (FloatCC.Unordered) a @ (value_type (ty_vec128 ty)) b))
      (x64_cmpp ty a b (FcmpImm.Unordered)))
(rule (lower (fcmp (FloatCC.UnorderedOrGreaterThan) a @ (value_type (ty_vec128 ty)) b))
      (x64_cmpp ty a b (FcmpImm.UnorderedOrGreaterThan)))
(rule (lower (fcmp (FloatCC.UnorderedOrGreaterThanOrEqual) a @ (value_type (ty_vec128 ty)) b))
      (x64_cmpp ty a b (FcmpImm.UnorderedOrGreaterThanOrEqual)))

;; Some vector lowerings rely on flipping the operands and using a reversed
;; comparison code.

(rule (lower (fcmp (FloatCC.GreaterThan) a @ (value_type (ty_vec128 ty)) b))
      (x64_cmpp ty b a (FcmpImm.LessThan)))
(rule (lower (fcmp (FloatCC.GreaterThanOrEqual) a @ (value_type (ty_vec128 ty)) b))
      (x64_cmpp ty b a (FcmpImm.LessThanOrEqual)))
(rule (lower (fcmp (FloatCC.UnorderedOrLessThan) a @ (value_type (ty_vec128 ty)) b))
      (x64_cmpp ty b a (FcmpImm.UnorderedOrGreaterThan)))
(rule (lower (fcmp (FloatCC.UnorderedOrLessThanOrEqual) a @ (value_type (ty_vec128 ty)) b))
      (x64_cmpp ty b a (FcmpImm.UnorderedOrGreaterThanOrEqual)))

;; Some vector lowerings are simply not supported for certain codes:
;; - FloatCC::OrderedNotEqual
;; - FloatCC::UnorderedOrEqual

;;;; Rules for `select` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;

;; CLIF `select` instructions receive a testable argument (i.e. boolean or
;; integer) that determines which of the other two arguments is selected as
;; output. Since Cranelift booleans are typically generated by a comparison, the
;; lowerings in this section "look upwards in the tree" to emit the proper
;; sequence of "selection" instructions.
;;
;; The following rules--for selecting on a floating-point comparison--emit a
;; `UCOMIS*` instruction and then a conditional move, `cmove`. Note that for
;; values contained in XMM registers, `cmove` and `cmove_or` may in fact emit a
;; jump sequence, not `CMOV`. The `cmove` instruction operates on the flags set
;; by `UCOMIS*`; the key to understanding these is the UCOMIS* documentation
;; (see Intel's Software Developer's Manual, volume 2, chapter 4):
;;  - unordered assigns    Z = 1, P = 1, C = 1
;;  - greater than assigns Z = 0, P = 0, C = 0
;;  - less than assigns    Z = 0, P = 0, C = 1
;;  - equal assigns        Z = 1, P = 0, C = 0
;;
;; Note that prefixing the flag with `N` means "not," so that `CC.P -> P = 1`
;; and `CC.NP -> P = 0`. Also, x86 uses mnemonics for certain combinations of
;; flags; e.g.:
;;  - `CC.B -> C = 1` (below)
;;  - `CC.NB -> C = 0` (not below)
;;  - `CC.BE -> C = 1 OR Z = 1` (below or equal)
;;  - `CC.NBE -> C = 0 AND Z = 0` (not below or equal)

(rule (lower (has_type ty (select (maybe_uextend (fcmp (FloatCC.Ordered) a b)) x y)))
      (with_flags (x64_ucomis b a) (cmove_from_values ty (CC.NP) x y)))

(rule (lower (has_type ty (select (maybe_uextend (fcmp (FloatCC.Unordered) a b)) x y)))
      (with_flags (x64_ucomis b a) (cmove_from_values ty (CC.P) x y)))

(rule (lower (has_type ty (select (maybe_uextend (fcmp (FloatCC.GreaterThan) a b)) x y)))
      (with_flags (x64_ucomis b a) (cmove_from_values ty (CC.NBE) x y)))

(rule (lower (has_type ty (select (maybe_uextend (fcmp (FloatCC.GreaterThanOrEqual) a b)) x y)))
      (with_flags (x64_ucomis b a) (cmove_from_values ty (CC.NB) x y)))

(rule (lower (has_type ty (select (maybe_uextend (fcmp (FloatCC.UnorderedOrLessThan) a b)) x y)))
      (with_flags (x64_ucomis b a) (cmove_from_values ty (CC.B) x y)))

(rule (lower (has_type ty (select (maybe_uextend (fcmp (FloatCC.UnorderedOrLessThanOrEqual) a b)) x y)))
      (with_flags (x64_ucomis b a) (cmove_from_values ty (CC.BE) x y)))

;; Certain FloatCC variants are implemented by flipping the operands of the
;; comparison (e.g., "greater than" is lowered the same as "less than" but the
;; comparison is reversed). This allows us to use a single flag for the `cmove`,
;; which involves fewer instructions than `cmove_or`.
;;
;; But why flip at all, you may ask? Can't we just use `CC.B` (i.e., below) for
;; `FloatCC.LessThan`? Recall that in these floating-point lowerings, values may
;; be unordered and we must we want to express that `FloatCC.LessThan` is `LT`,
;; not `LT | UNO`. By flipping the operands AND inverting the comparison (e.g.,
;; to `CC.NBE`), we also avoid these unordered cases.

(rule (lower (has_type ty (select (maybe_uextend (fcmp (FloatCC.LessThan) a b)) x y)))
      (with_flags (x64_ucomis a b) (cmove_from_values ty (CC.NBE) x y)))

(rule (lower (has_type ty (select (maybe_uextend (fcmp (FloatCC.LessThanOrEqual) a b)) x y)))
      (with_flags (x64_ucomis a b) (cmove_from_values ty (CC.NB) x y)))

(rule (lower (has_type ty (select (maybe_uextend (fcmp (FloatCC.UnorderedOrGreaterThan) a b)) x y)))
      (with_flags (x64_ucomis a b) (cmove_from_values ty (CC.B) x y)))

(rule (lower (has_type ty (select (maybe_uextend (fcmp (FloatCC.UnorderedOrGreaterThanOrEqual) a b)) x y)))
      (with_flags (x64_ucomis a b) (cmove_from_values ty (CC.BE) x y)))

;; `FloatCC.Equal` and `FloatCC.NotEqual` can only be implemented with multiple
;; flag checks. Recall from the flag assignment chart above that equality, e.g.,
;; will assign `Z = 1`. But so does an unordered comparison: `Z = 1, P = 1, C =
;; 1`. In order to avoid semantics like `EQ | UNO` for equality, we must ensure
;; that the values are actually ordered, checking that `P = 0` (note that the
;; `C` flag is irrelevant here). Since we cannot find a single instruction that
;; implements a `Z = 1 AND P = 0` check, we invert the flag checks (i.e., `Z = 1
;; AND P = 0` becomes `Z = 0 OR P = 1`) and also flip the select operands, `x`
;; and `y`. The same argument applies to `FloatCC.NotEqual`.
;;
;; More details about the CLIF semantics for `fcmp` are available at
;; https://docs.rs/cranelift-codegen/latest/cranelift_codegen/ir/trait.InstBuilder.html#method.fcmp.

(rule (lower (has_type ty (select (maybe_uextend (fcmp (FloatCC.Equal) a b)) x y)))
      (with_flags (x64_ucomis a b) (cmove_or_from_values ty (CC.NZ) (CC.P) y x)))

(rule (lower (has_type ty (select (maybe_uextend (fcmp (FloatCC.NotEqual) a b)) x y)))
      (with_flags (x64_ucomis a b) (cmove_or_from_values ty (CC.NZ) (CC.P) x y)))

;; We also can lower `select`s that depend on an `icmp` test, but more simply
;; than the `fcmp` variants above. In these cases, we lower to a `CMP`
;; instruction plus a `CMOV`; recall that `cmove_from_values` here may emit more
;; than one instruction for certain types (e.g., XMM-held, I128).

(rule (lower (has_type ty (select (maybe_uextend (icmp cc a @ (value_type (fits_in_64 a_ty)) b)) x y)))
      (let ((size OperandSize (raw_operand_size_of_type a_ty)))
           (with_flags (x64_cmp size b a) (cmove_from_values ty cc x y))))

;; Finally, we lower `select` from a condition value `c`. These rules are meant
;; to be the final, default lowerings if no other patterns matched above.

(rule -1 (lower (has_type ty (select c @ (value_type (fits_in_64 a_ty)) x y)))
      (let ((size OperandSize (raw_operand_size_of_type a_ty))
            ;; N.B.: disallow load-op fusion, see above. TODO:
            ;; https://github.com/bytecodealliance/wasmtime/issues/3953.
            (gpr_c Gpr (put_in_gpr c)))
           (with_flags (x64_test size gpr_c gpr_c) (cmove_from_values ty (CC.NZ) x y))))

(rule -2 (lower (has_type ty (select c @ (value_type $I128) x y)))
      (let ((cond_result IcmpCondResult (cmp_zero_i128 (CC.Z) c)))
        (select_icmp cond_result x y)))

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

;; If available, we can use a plain lzcnt instruction here. Note no
;; special handling is required for zero inputs, because the machine
;; instruction does what the CLIF expects for zero, i.e. it returns
;; zero.
(rule 2 (lower
         (has_type (and
                    (ty_32_or_64 ty)
                    (use_lzcnt $true))
                   (clz src)))
      (x64_lzcnt ty src))

(rule 2 (lower
         (has_type (and
                    (ty_32_or_64 ty)
                    (use_lzcnt $false))
                 (clz src)))
      (do_clz ty ty src))

(rule 1 (lower
       (has_type (ty_8_or_16 ty)
                 (clz src)))
      (do_clz $I32 ty (extend_to_gpr src $I32 (ExtendKind.Zero))))

(rule 0 (lower
       (has_type $I128
                 (clz src)))
      (let ((upper Gpr (do_clz $I64 $I64 (value_regs_get_gpr src 1)))
            (lower Gpr (x64_add $I64
                            (do_clz $I64 $I64 (value_regs_get_gpr src 0))
                            (RegMemImm.Imm 64)))
            (result_lo Gpr
              (with_flags_reg
               (x64_cmp_imm (OperandSize.Size64) 64 upper)
               (cmove $I64 (CC.NZ) upper lower))))
        (value_regs result_lo (imm $I64 0))))

;; Implementation helper for clz; operates on 32 or 64-bit units.
(decl do_clz (Type Type Gpr) Gpr)
(rule (do_clz ty orig_ty src)
      (let ((highest_bit_index Reg (bsr_or_else ty src (imm_i64 $I64 -1)))
            (bits_minus_1 Reg (imm ty (u64_sub (ty_bits_u64 orig_ty) 1))))
        (x64_sub ty bits_minus_1 highest_bit_index)))

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

;; Analogous to `clz` cases above, but using mirror instructions
;; (tzcnt vs lzcnt, bsf vs bsr).

(rule 2 (lower
         (has_type (and
                    (ty_32_or_64 ty)
                    (use_bmi1 $true))
                   (ctz src)))
      (x64_tzcnt ty src))

(rule 2 (lower
          (has_type (and
                     (ty_32_or_64 ty)
                     (use_bmi1 $false))
                 (ctz src)))
      (do_ctz ty ty src))

(rule 1 (lower
       (has_type (ty_8_or_16 ty)
                 (ctz src)))
      (do_ctz $I32 ty (extend_to_gpr src $I32 (ExtendKind.Zero))))

(rule 0 (lower
       (has_type $I128
                 (ctz src)))
      (let ((lower Gpr (do_ctz $I64 $I64 (value_regs_get_gpr src 0)))
            (upper Gpr (x64_add $I64
                            (do_ctz $I64 $I64 (value_regs_get_gpr src 1))
                            (RegMemImm.Imm 64)))
            (result_lo Gpr
              (with_flags_reg
               (x64_cmp_imm (OperandSize.Size64) 64 lower)
               (cmove $I64 (CC.Z) upper lower))))
        (value_regs result_lo (imm $I64 0))))

(decl do_ctz (Type Type Gpr) Gpr)
(rule (do_ctz ty orig_ty src)
      (bsf_or_else ty src (imm $I64 (ty_bits_u64 orig_ty))))

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

(rule 3 (lower
         (has_type (and
                    (ty_32_or_64 ty)
                    (use_popcnt $true))
                   (popcnt src)))
      (x64_popcnt ty src))

(rule 2 (lower
         (has_type (and
                    (ty_8_or_16 ty)
                    (use_popcnt $true))
                   (popcnt src)))
      (x64_popcnt $I32 (extend_to_gpr src $I32 (ExtendKind.Zero))))

(rule 1 (lower
         (has_type (and
                    $I128
                    (use_popcnt $true))
                   (popcnt src)))
      (let ((lo_count Gpr (x64_popcnt $I64 (value_regs_get_gpr src 0)))
            (hi_count Gpr (x64_popcnt $I64 (value_regs_get_gpr src 1))))
        (value_regs (x64_add $I64 lo_count hi_count) (imm $I64 0))))

(rule -1 (lower
       (has_type (ty_32_or_64 ty)
                 (popcnt src)))
      (do_popcnt ty src))

(rule -2 (lower
       (has_type (ty_8_or_16 ty)
                 (popcnt src)))
      (do_popcnt $I32 (extend_to_gpr src $I32 (ExtendKind.Zero))))

(rule (lower
       (has_type $I128
                 (popcnt src)))
      (let ((lo_count Gpr (do_popcnt $I64 (value_regs_get_gpr src 0)))
            (hi_count Gpr (do_popcnt $I64 (value_regs_get_gpr src 1))))
        (value_regs (x64_add $I64 lo_count hi_count) (imm $I64 0))))

;; Implementation of popcount when we don't nave a native popcount
;; instruction.
(decl do_popcnt (Type Gpr) Gpr)
(rule (do_popcnt $I64 src)
      (let ((shifted1 Gpr (x64_shr $I64 src (Imm8Reg.Imm8 1)))
            (sevens Gpr (imm $I64 0x7777777777777777))
            (masked1 Gpr (x64_and $I64 shifted1 sevens))
            ;; diff1 := src - ((src >> 1) & 0b0111_0111_0111...)
            (diff1 Gpr (x64_sub $I64 src masked1))
            (shifted2 Gpr (x64_shr $I64 masked1 (Imm8Reg.Imm8 1)))
            (masked2 Gpr (x64_and $I64 shifted2 sevens))
            ;; diff2 := diff1 - ((diff1 >> 1) & 0b0111_0111_0111...)
            (diff2 Gpr (x64_sub $I64 diff1 masked2))
            (shifted3 Gpr (x64_shr $I64 masked2 (Imm8Reg.Imm8 1)))
            (masked3 Gpr (x64_and $I64 shifted3 sevens))
            ;; diff3 := diff2 - ((diff2 >> 1) & 0b0111_0111_0111...)
            ;;
            ;; At this point, each nibble of diff3 is the popcount of
            ;; that nibble. This works because at each step above, we
            ;; are basically subtracting floor(value / 2) from the
            ;; running value; the leftover remainder is 1 if the LSB
            ;; was 1. After three steps, we have (nibble / 8) -- 0 or
            ;; 1 for the MSB of the nibble -- plus three possible
            ;; additions for the three other bits.
            (diff3 Gpr (x64_sub $I64 diff2 masked3))
            ;; Add the two nibbles of each byte together.
            (sum1 Gpr (x64_add $I64
                           (x64_shr $I64 diff3 (Imm8Reg.Imm8 4))
                           diff3))
            ;; Mask the above sum to have the popcount for each byte
            ;; in the lower nibble of that byte.
            (ofof Gpr (imm $I64 0x0f0f0f0f0f0f0f0f))
            (masked4 Gpr (x64_and $I64 sum1 ofof))
            (ones Gpr (imm $I64 0x0101010101010101))
            ;; Use a multiply to sum all of the bytes' popcounts into
            ;; the top byte. Consider the binomial expansion for the
            ;; top byte: it is the sum of the bytes (masked4 >> 56) *
            ;; 0x01 + (masked4 >> 48) * 0x01 + (masked4 >> 40) * 0x01
            ;; + ... + (masked4 >> 0).
            (mul Gpr (x64_mul $I64 masked4 ones))
            ;; Now take that top byte and return it as the popcount.
            (final Gpr (x64_shr $I64 mul (Imm8Reg.Imm8 56))))
        final))

;; This is the 32-bit version of the above; the steps for each nibble
;; are the same, we just use constants half as wide.
(rule (do_popcnt $I32 src)
      (let ((shifted1 Gpr (x64_shr $I32 src (Imm8Reg.Imm8 1)))
            (sevens Gpr (imm $I32 0x77777777))
            (masked1 Gpr (x64_and $I32 shifted1 sevens))
            (diff1 Gpr (x64_sub $I32 src masked1))
            (shifted2 Gpr (x64_shr $I32 masked1 (Imm8Reg.Imm8 1)))
            (masked2 Gpr (x64_and $I32 shifted2 sevens))
            (diff2 Gpr (x64_sub $I32 diff1 masked2))
            (shifted3 Gpr (x64_shr $I32 masked2 (Imm8Reg.Imm8 1)))
            (masked3 Gpr (x64_and $I32 shifted3 sevens))
            (diff3 Gpr (x64_sub $I32 diff2 masked3))
            (sum1 Gpr (x64_add $I32
                           (x64_shr $I32 diff3 (Imm8Reg.Imm8 4))
                           diff3))
            (masked4 Gpr (x64_and $I32 sum1 (RegMemImm.Imm 0x0f0f0f0f)))
            (mul Gpr (x64_mul $I32 masked4 (RegMemImm.Imm 0x01010101)))
            (final Gpr (x64_shr $I32 mul (Imm8Reg.Imm8 24))))
        final))


(rule 1 (lower (has_type (and
                          $I8X16
                          (avx512vl_enabled $true)
                          (avx512bitalg_enabled $true))
                         (popcnt src)))
      (x64_vpopcntb src))



;; For SSE 4.2 we use Mula's algorithm (https://arxiv.org/pdf/1611.07612.pdf):
;;
;; __m128i count_bytes ( __m128i v) {
;;     __m128i lookup = _mm_setr_epi8(0, 1, 1, 2, 1, 2, 2, 3, 1, 2, 2, 3, 2, 3, 3, 4);
;;     __m128i low_mask = _mm_set1_epi8 (0x0f);
;;     __m128i lo = _mm_and_si128 (v, low_mask);
;;     __m128i hi = _mm_and_si128 (_mm_srli_epi16 (v, 4), low_mask);
;;     __m128i cnt1 = _mm_shuffle_epi8 (lookup, lo);
;;     __m128i cnt2 = _mm_shuffle_epi8 (lookup, hi);
;;     return _mm_add_epi8 (cnt1, cnt2);
;; }
;;
;; Details of the above algorithm can be found in the reference noted above, but the basics
;; are to create a lookup table that pre populates the popcnt values for each number [0,15].
;; The algorithm uses shifts to isolate 4 bit sections of the vector, pshufb as part of the
;; lookup process, and adds together the results.
;;
;; __m128i lookup = _mm_setr_epi8(0, 1, 1, 2, 1, 2, 2, 3, 1, 2, 2, 3, 2, 3, 3, 4);

(decl popcount_4bit_table () VCodeConstant)  ;; bits-per-nibble table `lookup` above
(extern constructor popcount_4bit_table popcount_4bit_table)

(decl popcount_low_mask () VCodeConstant)    ;; mask for low nibbles: 0x0f * 16
(extern constructor popcount_low_mask popcount_low_mask)

(rule (lower (has_type $I8X16
                       (popcnt src)))
      (let ((nibble_table_const VCodeConstant (popcount_4bit_table))
            (low_mask XmmMem (popcount_low_mask))
            (low_nibbles Xmm (sse_and $I8X16 src low_mask))
            ;; Note that this is a 16x8 shift, but that's OK; we mask
            ;; off anything that traverses from one byte to the next
            ;; with the low_mask below.
            (shifted_src Xmm (x64_psrlw src (RegMemImm.Imm 4)))
            (high_nibbles Xmm (sse_and $I8X16 shifted_src low_mask))
            (lookup Xmm (x64_xmm_load_const $I8X16 (popcount_4bit_table)))
            (bit_counts_low Xmm (x64_pshufb lookup low_nibbles))
            (bit_counts_high Xmm (x64_pshufb lookup high_nibbles)))
        (x64_paddb bit_counts_low bit_counts_high)))

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

(rule (lower (has_type $I8 (bitrev src)))
      (do_bitrev8 $I32 src))

(rule (lower (has_type $I16 (bitrev src)))
      (do_bitrev16 $I32 src))

(rule (lower (has_type $I32 (bitrev src)))
      (do_bitrev32 $I32 src))

(rule (lower (has_type $I64 (bitrev src)))
      (do_bitrev64 $I64 src))

(rule (lower (has_type $I128 (bitrev src)))
      (value_regs
       (do_bitrev64 $I64 (value_regs_get_gpr src 1))
       (do_bitrev64 $I64 (value_regs_get_gpr src 0))))

(decl do_bitrev8 (Type Gpr) Gpr)
(rule (do_bitrev8 ty src)
      (let ((tymask u64 (ty_mask ty))
            (mask1 Gpr (imm ty (u64_and tymask 0x5555555555555555)))
            (lo1 Gpr (x64_and ty src mask1))
            (hi1 Gpr (x64_and ty (x64_shr ty src (Imm8Reg.Imm8 1)) mask1))
            (swap1 Gpr (x64_or ty
                           (x64_shl ty lo1 (Imm8Reg.Imm8 1))
                           hi1))
            (mask2 Gpr (imm ty (u64_and tymask 0x3333333333333333)))
            (lo2 Gpr (x64_and ty swap1 mask2))
            (hi2 Gpr (x64_and ty (x64_shr ty swap1 (Imm8Reg.Imm8 2)) mask2))
            (swap2 Gpr (x64_or ty
                           (x64_shl ty lo2 (Imm8Reg.Imm8 2))
                           hi2))
            (mask4 Gpr (imm ty (u64_and tymask 0x0f0f0f0f0f0f0f0f)))
            (lo4 Gpr (x64_and ty swap2 mask4))
            (hi4 Gpr (x64_and ty (x64_shr ty swap2 (Imm8Reg.Imm8 4)) mask4))
            (swap4 Gpr (x64_or ty
                           (x64_shl ty lo4 (Imm8Reg.Imm8 4))
                           hi4)))
        swap4))

(decl do_bitrev16 (Type Gpr) Gpr)
(rule (do_bitrev16 ty src)
      (let ((src_ Gpr (do_bitrev8 ty src))
            (tymask u64 (ty_mask ty))
            (mask8 Gpr (imm ty (u64_and tymask 0x00ff00ff00ff00ff)))
            (lo8 Gpr (x64_and ty src_ mask8))
            (hi8 Gpr (x64_and ty (x64_shr ty src_ (Imm8Reg.Imm8 8)) mask8))
            (swap8 Gpr (x64_or ty
                           (x64_shl ty lo8 (Imm8Reg.Imm8 8))
                           hi8)))
        swap8))

(decl do_bitrev32 (Type Gpr) Gpr)
(rule (do_bitrev32 ty src)
      (let ((src_ Gpr (do_bitrev16 ty src))
            (tymask u64 (ty_mask ty))
            (mask16 Gpr (imm ty (u64_and tymask 0x0000ffff0000ffff)))
            (lo16 Gpr (x64_and ty src_ mask16))
            (hi16 Gpr (x64_and ty (x64_shr ty src_ (Imm8Reg.Imm8 16)) mask16))
            (swap16 Gpr (x64_or ty
                            (x64_shl ty lo16 (Imm8Reg.Imm8 16))
                            hi16)))
        swap16))

(decl do_bitrev64 (Type Gpr) Gpr)
(rule (do_bitrev64 ty @ $I64 src)
      (let ((src_ Gpr (do_bitrev32 ty src))
            (mask32 Gpr (imm ty 0xffffffff))
            (lo32 Gpr (x64_and ty src_ mask32))
            (hi32 Gpr (x64_shr ty src_ (Imm8Reg.Imm8 32)))
            (swap32 Gpr (x64_or ty
                            (x64_shl ty lo32 (Imm8Reg.Imm8 32))
                            hi32)))
        swap32))

;; Rules for `bswap` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;

;; x64 bswap instruction is only for 32- or 64-bit swaps
;; implement the 16-bit swap as a rotl by 8
(rule (lower (has_type $I16 (bswap src)))
      (x64_rotl $I16 src (Imm8Reg.Imm8 8)))

(rule (lower (has_type $I32 (bswap src)))
      (x64_bswap $I32 src))

(rule (lower (has_type $I64 (bswap src)))
      (x64_bswap $I64 src))

(rule (lower (has_type $I128 (bswap src)))
      (value_regs
       (x64_bswap $I64 (value_regs_get_gpr src 1))
       (x64_bswap $I64 (value_regs_get_gpr src 0))))

;; Rules for `is_null` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;

;; Null references are represented by the constant value `0`.
(rule (lower (is_null src @ (value_type $R64)))
      (with_flags
       (x64_cmp_imm (OperandSize.Size64) 0 src)
       (x64_setcc (CC.Z))))

;; Rules for `is_invalid` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;

;; Invalid references are represented by the constant value `-1`.
(rule (lower (is_invalid src @ (value_type $R64)))
      (with_flags
       (x64_cmp_imm (OperandSize.Size64) 0xffffffff src)  ;; simm32 0xffff_ffff is sign-extended to -1.
       (x64_setcc (CC.Z))))


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

;; T -> T is a no-op.
(rule 1 (lower (has_type ty (uextend src @ (value_type ty))))
      src)

;; I64 -> I128.
(rule -1 (lower (has_type $I128 (uextend src @ (value_type $I64))))
      (value_regs src (imm $I64 0)))

;; I{8,16,32} -> I128.
(rule (lower (has_type $I128 (uextend src @ (value_type (fits_in_32 src_ty)))))
      (value_regs (extend_to_gpr src $I64 (ExtendKind.Zero)) (imm $I64 0)))

;; I{8,16,32} -> I64.
(rule -1 (lower (has_type $I64 (uextend src @ (value_type (fits_in_32 src_ty)))))
      (extend_to_gpr src $I64 (ExtendKind.Zero)))

;; I8 -> I{16,32}, I16 -> I32.
(rule -2 (lower (has_type (fits_in_32 dst_ty) (uextend src @ (value_type (fits_in_32 src_ty)))))
      (extend_to_gpr src $I32 (ExtendKind.Zero)))

;; I32 -> I64 with op that produces a zero-extended value in a register.
;;
;; As a particular x64 extra-pattern matching opportunity, all the ALU
;; opcodes on 32-bits will zero-extend the upper 32-bits, so we can
;; even not generate a zero-extended move in this case.
;;
;; (Note that we unfortunately can't factor out the
;; insts-that-zero-upper-32 pattern into a separate extractor until we
;; can write internal extractors with multiple rules; and we'd rather
;; keep these here than write an external extractor containing bits of
;; the instruction pattern.s)
(rule (lower (has_type $I64
                       (uextend src @ (has_type $I32 (iadd _ _)))))
      src)
(rule (lower (has_type $I64
                       (uextend src @ (has_type $I32 (isub _ _)))))
      src)
(rule (lower (has_type $I64
                       (uextend src @ (has_type $I32 (imul _ _)))))
      src)
(rule (lower (has_type $I64
                       (uextend src @ (has_type $I32 (band _ _)))))
      src)
(rule (lower (has_type $I64
                       (uextend src @ (has_type $I32 (bor _ _)))))
      src)
(rule (lower (has_type $I64
                       (uextend src @ (has_type $I32 (bxor _ _)))))
      src)
(rule (lower (has_type $I64
                       (uextend src @ (has_type $I32 (ishl _ _)))))
      src)
(rule (lower (has_type $I64
                       (uextend src @ (has_type $I32 (ushr _ _)))))
      src)
(rule (lower (has_type $I64
                       (uextend src @ (has_type $I32 (uload32 _ _ _)))))
      src)

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

(decl generic_sextend (Value Type Type) InstOutput)

;; T -> T is a no-op.
(rule 4 (generic_sextend src ty ty)
      src)

;; Produce upper 64 bits sign-extended from lower 64: shift right by
;; 63 bits to spread the sign bit across the result.
(decl spread_sign_bit (Gpr) Gpr)
(rule (spread_sign_bit src)
      (x64_sar $I64 src (Imm8Reg.Imm8 63)))

;; I64 -> I128.
(rule 3 (generic_sextend src $I64 $I128)
      (value_regs src (spread_sign_bit src)))

;; I{8,16,32} -> I128.
(rule 2 (generic_sextend src (fits_in_32 src_ty) $I128)
      (let ((lo Gpr (extend_to_gpr src $I64 (ExtendKind.Sign)))
            (hi Gpr (spread_sign_bit lo)))
      (value_regs lo hi)))

;; I{8,16,32} -> I64.
(rule 1 (generic_sextend src (fits_in_32 src_ty) $I64)
      (extend_to_gpr src $I64 (ExtendKind.Sign)))

;; I8 -> I{16,32}, I16 -> I32.
(rule 0 (generic_sextend src (fits_in_32 src_ty) (fits_in_32 dst_ty))
      (extend_to_gpr src $I32 (ExtendKind.Sign)))

(rule (lower
       (has_type dst_ty
                 (sextend src @ (value_type src_ty))))
      (generic_sextend src src_ty dst_ty))

;; Rules for `ireduce` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;

;; T -> T is always a no-op, even I128 -> I128.
(rule (lower (has_type ty (ireduce src @ (value_type ty))))
      src)

;; T -> I{64,32,16,8}: We can simply pass through the value: values
;; are always stored with high bits undefined, so we can just leave
;; them be.
(rule 1 (lower (has_type (fits_in_64 ty) (ireduce src)))
      (value_regs_get_gpr src 0))

;; Rules for `debugtrap` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;

(rule (lower (debugtrap))
      (side_effect (x64_hlt)))

;; Rules for `fadd` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;

(rule (lower (has_type $F32 (fadd x y)))
      (x64_addss x y))
(rule (lower (has_type $F64 (fadd x y)))
      (x64_addsd x y))
(rule (lower (has_type $F32X4 (fadd x y)))
      (x64_addps x y))
(rule (lower (has_type $F64X2 (fadd x y)))
      (x64_addpd x y))

;; The above rules automatically sink loads for rhs operands, so additionally
;; add rules for sinking loads with lhs operands.
(rule 1 (lower (has_type $F32 (fadd (sinkable_load x) y)))
      (x64_addss y x))
(rule 1 (lower (has_type $F64 (fadd (sinkable_load x) y)))
      (x64_addsd y x))
(rule 1 (lower (has_type $F32X4 (fadd (sinkable_load x) y)))
      (x64_addps y x))
(rule 1 (lower (has_type $F64X2 (fadd (sinkable_load x) y)))
      (x64_addpd y x))

;; Rules for `fsub` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;

(rule (lower (has_type $F32 (fsub x y)))
      (x64_subss x y))
(rule (lower (has_type $F64 (fsub x y)))
      (x64_subsd x y))
(rule (lower (has_type $F32X4 (fsub x y)))
      (x64_subps x y))
(rule (lower (has_type $F64X2 (fsub x y)))
      (x64_subpd x y))

;; Rules for `fmul` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;

(rule (lower (has_type $F32 (fmul x y)))
      (x64_mulss x y))
(rule (lower (has_type $F64 (fmul x y)))
      (x64_mulsd x y))
(rule (lower (has_type $F32X4 (fmul x y)))
      (x64_mulps x y))
(rule (lower (has_type $F64X2 (fmul x y)))
      (x64_mulpd x y))

;; The above rules automatically sink loads for rhs operands, so additionally
;; add rules for sinking loads with lhs operands.
(rule 1 (lower (has_type $F32 (fmul (sinkable_load x) y)))
      (x64_mulss y x))
(rule 1 (lower (has_type $F64 (fmul (sinkable_load x) y)))
      (x64_mulsd y x))
(rule 1 (lower (has_type $F32X4 (fmul (sinkable_load x) y)))
      (x64_mulps y x))
(rule 1 (lower (has_type $F64X2 (fmul (sinkable_load x) y)))
      (x64_mulpd y x))

;; Rules for `fdiv` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;

(rule (lower (has_type $F32 (fdiv x y)))
      (x64_divss x y))
(rule (lower (has_type $F64 (fdiv x y)))
      (x64_divsd x y))
(rule (lower (has_type $F32X4 (fdiv x y)))
      (x64_divps x y))
(rule (lower (has_type $F64X2 (fdiv x y)))
      (x64_divpd x y))

;; Rules for `sqrt` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
(rule (lower (has_type $F32 (sqrt x)))
      (x64_sqrtss x))
(rule (lower (has_type $F64 (sqrt x)))
      (x64_sqrtsd x))
(rule (lower (has_type $F32X4 (sqrt x)))
      (x64_sqrtps x))
(rule (lower (has_type $F64X2 (sqrt x)))
      (x64_sqrtpd x))

;; Rules for `fpromote` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
(rule (lower (has_type $F64 (fpromote x)))
      (x64_cvtss2sd x))

;; Rules for `fvpromote` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
(rule (lower (has_type $F64X2 (fvpromote_low x)))
      (x64_cvtps2pd (put_in_xmm x)))

;; Rules for `fdemote` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
(rule (lower (has_type $F32 (fdemote x)))
      (x64_cvtsd2ss x))

;; Rules for `fvdemote` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
(rule (lower (has_type $F32X4 (fvdemote x)))
      (x64_cvtpd2ps x))

;; Rules for `fmin` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;

(rule (lower (has_type $F32 (fmin x y)))
      (xmm_min_max_seq $F32 $true x y))
(rule (lower (has_type $F64 (fmin x y)))
      (xmm_min_max_seq $F64 $true x y))

;; Vector-typed version. We don't use single pseudoinstructions as
;; above, because we don't need to generate a mini-CFG. Instead, we
;; perform a branchless series of operations.
;;
;; We cannot simply use native min instructions (minps, minpd) because
;; NaN handling is different per CLIF semantics than on
;; x86. Specifically, if an argument is NaN, or the arguments are both
;; zero but of opposite signs, then the x86 instruction always
;; produces the second argument. However, per CLIF semantics, we
;; require that fmin(NaN, _) = fmin(_, NaN) = NaN, and fmin(+0, -0) =
;; fmin(-0, +0) = -0.

(rule (lower (has_type $F32X4 (fmin x y)))
      ;; Compute min(x, y) and min(y, x) with native
      ;; instructions. These will differ in one of the edge cases
      ;; above that we have to handle properly. (Conversely, if they
      ;; don't differ, then the native instruction's answer is the
      ;; right one per CLIF semantics.)
      (let ((min1 Xmm (x64_minps x y))
            (min2 Xmm (x64_minps y x))
            ;; Compute the OR of the two. Note that NaNs have an
            ;; exponent field of all-ones (0xFF for F32), so if either
            ;; result is a NaN, this OR will be. And if either is a
            ;; zero (which has an exponent of 0 and mantissa of 0),
            ;; this captures a sign-bit of 1 (negative) if either
            ;; input is negative.
            ;;
            ;; In the case where we don't have a +/-0 mismatch or
            ;; NaNs, then `min1` and `min2` are equal and `min_or` is
            ;; the correct minimum.
            (min_or Xmm (x64_orps min1 min2))
            ;; "compare unordered" produces a true mask (all ones) in
            ;; a given lane if the min is a NaN. We use this to
            ;; generate a mask to ensure quiet NaNs.
            (is_nan_mask Xmm (x64_cmpps min_or min2 (FcmpImm.Unordered)))
            ;; OR in the NaN mask.
            (min_or_2 Xmm (x64_orps min_or is_nan_mask))
            ;; Shift the NaN mask down so that it covers just the
            ;; fraction below the NaN signalling bit; we'll use this
            ;; to mask off non-canonical NaN payloads.
            ;;
            ;; All-ones for NaN, shifted down to leave 10 top bits (1
            ;; sign, 8 exponent, 1 QNaN bit that must remain set)
            ;; cleared.
            (nan_fraction_mask Xmm (x64_psrld is_nan_mask (RegMemImm.Imm 10)))
            ;; Do a NAND, so that we retain every bit not set in
            ;; `nan_fraction_mask`. This mask will be all zeroes (so
            ;; we retain every bit) in non-NaN cases, and will have
            ;; ones (so we clear those bits) in NaN-payload bits
            ;; otherwise.
            (final Xmm (x64_andnps nan_fraction_mask min_or_2)))
        final))

;; Likewise for F64 lanes, except that the right-shift is by 13 bits
;; (1 sign, 11 exponent, 1 QNaN bit).
(rule (lower (has_type $F64X2 (fmin x y)))
      (let ((min1 Xmm (x64_minpd x y))
            (min2 Xmm (x64_minpd y x))
            (min_or Xmm (x64_orpd min1 min2))
            (is_nan_mask Xmm (x64_cmppd min1 min2 (FcmpImm.Unordered)))
            (min_or_2 Xmm (x64_orpd min_or is_nan_mask))
            (nan_fraction_mask Xmm (x64_psrlq is_nan_mask (RegMemImm.Imm 13)))
            (final Xmm (x64_andnpd nan_fraction_mask min_or_2)))
        final))

;; Rules for `fmax` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;

(rule (lower (has_type $F32 (fmax x y)))
      (xmm_min_max_seq $F32 $false x y))
(rule (lower (has_type $F64 (fmax x y)))
      (xmm_min_max_seq $F64 $false x y))

;; The vector version of fmax here is a dual to the fmin sequence
;; above, almost, with a few differences.

(rule (lower (has_type $F32X4 (fmax x y)))
      ;; Compute max(x, y) and max(y, x) with native
      ;; instructions. These will differ in one of the edge cases
      ;; above that we have to handle properly. (Conversely, if they
      ;; don't differ, then the native instruction's answer is the
      ;; right one per CLIF semantics.)
      (let ((max1 Xmm (x64_maxps x y))
            (max2 Xmm (x64_maxps y x))
            ;; Compute the XOR of the two maxima. In the case
            ;; where we don't have a +/-0 mismatch or NaNs, then
            ;; `min1` and `min2` are equal and this XOR is zero.
            (max_xor Xmm (x64_xorps max1 max2))
            ;; OR the XOR into one of the original maxima. If they are
            ;; equal, this does nothing. If max2 was NaN, its exponent
            ;; bits were all-ones, so the xor's exponent bits were the
            ;; complement of max1, and the OR of max1 and max_xor has
            ;; an all-ones exponent (is a NaN). If max1 was NaN, then
            ;; its exponent bits were already all-ones, so the OR will
            ;; be a NaN as well.
            (max_blended_nan Xmm (x64_orps max1 max_xor))
            ;; Subtract the XOR. This ensures that if we had +0 and
            ;; -0, we end up with +0.
            (max_blended_nan_positive Xmm (x64_subps max_blended_nan max_xor))
            ;; "compare unordered" produces a true mask (all ones) in
            ;; a given lane if the min is a NaN. We use this to
            ;; generate a mask to ensure quiet NaNs.
            (is_nan_mask Xmm (x64_cmpps max_blended_nan max_blended_nan (FcmpImm.Unordered)))
            ;; Shift the NaN mask down so that it covers just the
            ;; fraction below the NaN signalling bit; we'll use this
            ;; to mask off non-canonical NaN payloads.
            ;;
            ;; All-ones for NaN, shifted down to leave 10 top bits (1
            ;; sign, 8 exponent, 1 QNaN bit that must remain set)
            ;; cleared.
            (nan_fraction_mask Xmm (x64_psrld is_nan_mask (RegMemImm.Imm 10)))
            ;; Do a NAND, so that we retain every bit not set in
            ;; `nan_fraction_mask`. This mask will be all zeroes (so
            ;; we retain every bit) in non-NaN cases, and will have
            ;; ones (so we clear those bits) in NaN-payload bits
            ;; otherwise.
            (final Xmm (x64_andnps nan_fraction_mask max_blended_nan_positive)))
        final))

(rule (lower (has_type $F64X2 (fmax x y)))
      ;; Compute max(x, y) and max(y, x) with native
      ;; instructions. These will differ in one of the edge cases
      ;; above that we have to handle properly. (Conversely, if they
      ;; don't differ, then the native instruction's answer is the
      ;; right one per CLIF semantics.)
      (let ((max1 Xmm (x64_maxpd x y))
            (max2 Xmm (x64_maxpd y x))
            ;; Compute the XOR of the two maxima. In the case
            ;; where we don't have a +/-0 mismatch or NaNs, then
            ;; `min1` and `min2` are equal and this XOR is zero.
            (max_xor Xmm (x64_xorpd max1 max2))
            ;; OR the XOR into one of the original maxima. If they are
            ;; equal, this does nothing. If max2 was NaN, its exponent
            ;; bits were all-ones, so the xor's exponent bits were the
            ;; complement of max1, and the OR of max1 and max_xor has
            ;; an all-ones exponent (is a NaN). If max1 was NaN, then
            ;; its exponent bits were already all-ones, so the OR will
            ;; be a NaN as well.
            (max_blended_nan Xmm (x64_orpd max1 max_xor))
            ;; Subtract the XOR. This ensures that if we had +0 and
            ;; -0, we end up with +0.
            (max_blended_nan_positive Xmm (x64_subpd max_blended_nan max_xor))
            ;; `cmpps` with predicate index `3` is `cmpunordps`, or
            ;; "compare unordered": it produces a true mask (all ones)
            ;; in a given lane if the min is a NaN. We use this to
            ;; generate a mask to ensure quiet NaNs.
            (is_nan_mask Xmm (x64_cmppd max_blended_nan max_blended_nan (FcmpImm.Unordered)))
            ;; Shift the NaN mask down so that it covers just the
            ;; fraction below the NaN signalling bit; we'll use this
            ;; to mask off non-canonical NaN payloads.
            ;;
            ;; All-ones for NaN, shifted down to leave 13 top bits (1
            ;; sign, 11 exponent, 1 QNaN bit that must remain set)
            ;; cleared.
            (nan_fraction_mask Xmm (x64_psrlq is_nan_mask (RegMemImm.Imm 13)))
            ;; Do a NAND, so that we retain every bit not set in
            ;; `nan_fraction_mask`. This mask will be all zeroes (so
            ;; we retain every bit) in non-NaN cases, and will have
            ;; ones (so we clear those bits) in NaN-payload bits
            ;; otherwise.
            (final Xmm (x64_andnpd nan_fraction_mask max_blended_nan_positive)))
        final))

;; Rules for `fmin_pseudo` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;

(rule (lower (has_type $F32 (fmin_pseudo x y)))
      (x64_minss y x))
(rule (lower (has_type $F64 (fmin_pseudo x y)))
      (x64_minsd y x))
(rule (lower (has_type $F32X4 (fmin_pseudo x y)))
      (x64_minps y x))
(rule (lower (has_type $F64X2 (fmin_pseudo x y)))
      (x64_minpd y x))

;; Rules for `fmax_pseudo` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;

(rule (lower (has_type $F32 (fmax_pseudo x y)))
      (x64_maxss y x))
(rule (lower (has_type $F64 (fmax_pseudo x y)))
      (x64_maxsd y x))
(rule (lower (has_type $F32X4 (fmax_pseudo x y)))
      (x64_maxps y x))
(rule (lower (has_type $F64X2 (fmax_pseudo x y)))
      (x64_maxpd y x))

;; Rules for `fma` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;

;; Base case for fma is to call out to one of two libcalls. For vectors they
;; need to be decomposed, handle each element individually, and then recomposed.

(rule (lower (has_type $F32 (fma x y z)))
      (libcall_3 (LibCall.FmaF32) x y z))
(rule (lower (has_type $F64 (fma x y z)))
      (libcall_3 (LibCall.FmaF64) x y z))

(rule (lower (has_type $F32X4 (fma x y z)))
      (let (
          (x Xmm (put_in_xmm x))
          (y Xmm (put_in_xmm y))
          (z Xmm (put_in_xmm z))
          (x0 Xmm (libcall_3 (LibCall.FmaF32) x y z))
          (x1 Xmm (libcall_3 (LibCall.FmaF32)
            (x64_pshufd x 1)
            (x64_pshufd y 1)
            (x64_pshufd z 1)))
          (x2 Xmm (libcall_3 (LibCall.FmaF32)
            (x64_pshufd x 2)
            (x64_pshufd y 2)
            (x64_pshufd z 2)))
          (x3 Xmm (libcall_3 (LibCall.FmaF32)
            (x64_pshufd x 3)
            (x64_pshufd y 3)
            (x64_pshufd z 3)))

          (tmp Xmm (vec_insert_lane $F32X4 x0 x1 1))
          (tmp Xmm (vec_insert_lane $F32X4 tmp x2 2))
          (tmp Xmm (vec_insert_lane $F32X4 tmp x3 3))
        )
        tmp))
(rule (lower (has_type $F64X2 (fma x y z)))
      (let (
          (x Xmm (put_in_xmm x))
          (y Xmm (put_in_xmm y))
          (z Xmm (put_in_xmm z))
          (x0 Xmm (libcall_3 (LibCall.FmaF64) x y z))
          (x1 Xmm (libcall_3 (LibCall.FmaF64)
            (x64_pshufd x 0xee)
            (x64_pshufd y 0xee)
            (x64_pshufd z 0xee)))
        )
        (vec_insert_lane $F64X2 x0 x1 1)))


;; Special case for when the `fma` feature is active and a native instruction
;; can be used.
(rule 1 (lower (has_type ty (fma x y z)))
      (if-let $true (use_fma))
      (fmadd ty x y z))

(decl fmadd (Type Value Value Value) Xmm)
(decl fnmadd (Type Value Value Value) Xmm)

;; Base case. Note that this will automatically sink a load with `z`, the value
;; to add.
(rule (fmadd ty x y z) (x64_vfmadd213 ty x y z))

;; Allow sinking loads with one of the two values being multiplied in addition
;; to the value being added. Note that both x and y can be sunk here due to
;; multiplication being commutative.
(rule 1 (fmadd ty (sinkable_load x) y z) (x64_vfmadd132 ty y z x))
(rule 2 (fmadd ty x (sinkable_load y) z) (x64_vfmadd132 ty x z y))

;; If one of the values being multiplied is negated then use a `vfnmadd*`
;; instruction instead
(rule 3 (fmadd ty (fneg x) y z) (fnmadd ty x y z))
(rule 4 (fmadd ty x (fneg y) z) (fnmadd ty x y z))

(rule (fnmadd ty x y z) (x64_vfnmadd213 ty x y z))
(rule 1 (fnmadd ty (sinkable_load x) y z) (x64_vfnmadd132 ty y z x))
(rule 2 (fnmadd ty x (sinkable_load y) z) (x64_vfnmadd132 ty x z y))

;; Like `fmadd` if one argument is negated switch which one is being codegen'd
(rule 3 (fnmadd ty (fneg x) y z) (fmadd ty x y z))
(rule 4 (fnmadd ty x (fneg y) z) (fmadd ty x y z))

;; Rules for `load*` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;

;; In order to load a value from memory to a GPR register, we may need to extend
;; the loaded value from 8-, 16-, or 32-bits to this backend's expected GPR
;; width: 64 bits. Note that `ext_mode` will load 1-bit types (booleans) as
;; 8-bit loads.
;;
;; By default, we zero-extend all sub-64-bit loads to a GPR.
(rule -4 (lower (has_type (and (fits_in_32 ty) (is_gpr_type _)) (load flags address offset)))
      (x64_movzx (ext_mode (ty_bits_u16 ty) 64) (to_amode flags address offset)))
;; But if we know that both the `from` and `to` are 64 bits, we simply load with
;; no extension.
(rule -1 (lower (has_type (ty_int_ref_64 ty) (load flags address offset)))
      (x64_mov (to_amode flags address offset)))
;; Also, certain scalar loads have a specific `from` width and extension kind
;; (signed -> `sx`, zeroed -> `zx`). We overwrite the high bits of the 64-bit
;; GPR even if the `to` type is smaller (e.g., 16-bits).
(rule (lower (has_type (is_gpr_type ty) (uload8 flags address offset)))
      (x64_movzx (ExtMode.BQ) (to_amode flags address offset)))
(rule (lower (has_type (is_gpr_type ty) (sload8 flags address offset)))
      (x64_movsx (ExtMode.BQ) (to_amode flags address offset)))
(rule (lower (has_type (is_gpr_type ty) (uload16 flags address offset)))
      (x64_movzx (ExtMode.WQ) (to_amode flags address offset)))
(rule (lower (has_type (is_gpr_type ty) (sload16 flags address offset)))
      (x64_movsx (ExtMode.WQ) (to_amode flags address offset)))
(rule (lower (has_type (is_gpr_type ty) (uload32 flags address offset)))
      (x64_movzx (ExtMode.LQ) (to_amode flags address offset)))
(rule (lower (has_type (is_gpr_type ty) (sload32 flags address offset)))
      (x64_movsx (ExtMode.LQ) (to_amode flags address offset)))

;; To load to XMM registers, we use the x64-specific instructions for each type.
;; For `$F32` and `$F64` this is important--we only want to load 32 or 64 bits.
;; But for the 128-bit types, this is not strictly necessary for performance but
;; might help with clarity during disassembly.
(rule (lower (has_type $F32 (load flags address offset)))
      (x64_movss_load (to_amode flags address offset)))
(rule (lower (has_type $F64 (load flags address offset)))
      (x64_movsd_load (to_amode flags address offset)))
(rule (lower (has_type $F32X4 (load flags address offset)))
      (x64_movups (to_amode flags address offset)))
(rule (lower (has_type $F64X2 (load flags address offset)))
      (x64_movupd (to_amode flags address offset)))
(rule -2 (lower (has_type (ty_vec128 ty) (load flags address offset)))
      (x64_movdqu (to_amode flags address offset)))

;; We can load an I128 by doing two 64-bit loads.
(rule -3 (lower (has_type $I128
                       (load flags address offset)))
      (let ((addr_lo Amode (to_amode flags address offset))
            (addr_hi Amode (amode_offset addr_lo 8))
            (value_lo Reg (x64_mov addr_lo))
            (value_hi Reg (x64_mov addr_hi)))
        (value_regs value_lo value_hi)))

;; We also include widening vector loads; these sign- or zero-extend each lane
;; to the next wider width (e.g., 16x4 -> 32x4).
(rule (lower (has_type $I16X8 (sload8x8 flags address offset)))
      (x64_pmovsxbw (to_amode flags address offset)))
(rule (lower (has_type $I16X8 (uload8x8 flags address offset)))
      (x64_pmovzxbw (to_amode flags address offset)))
(rule (lower (has_type $I32X4 (sload16x4 flags address offset)))
      (x64_pmovsxwd (to_amode flags address offset)))
(rule (lower (has_type $I32X4 (uload16x4 flags address offset)))
      (x64_pmovzxwd (to_amode flags address offset)))
(rule (lower (has_type $I64X2 (sload32x2 flags address offset)))
      (x64_pmovsxdq (to_amode flags address offset)))
(rule (lower (has_type $I64X2 (uload32x2 flags address offset)))
      (x64_pmovzxdq (to_amode flags address offset)))

;; Rules for `store*` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;

;; 8-, 16-, 32- and 64-bit GPR stores.
(rule -2 (lower (store flags
                    value @ (value_type (is_gpr_type ty))
                    address
                    offset))
      (side_effect
       (x64_movrm ty (to_amode flags address offset) value)))

;; Explicit 8/16/32-bit opcodes.
(rule (lower (istore8 flags value address offset))
      (side_effect
       (x64_movrm $I8 (to_amode flags address offset) value)))
(rule (lower (istore16 flags value address offset))
      (side_effect
       (x64_movrm $I16 (to_amode flags address offset) value)))
(rule (lower (istore32 flags value address offset))
      (side_effect
       (x64_movrm $I32 (to_amode flags address offset) value)))

;; F32 stores of values in XMM registers.
(rule 1 (lower (store flags
                    value @ (value_type $F32)
                    address
                    offset))
      (side_effect
       (x64_xmm_movrm (SseOpcode.Movss) (to_amode flags address offset) value)))

;; F64 stores of values in XMM registers.
(rule 1 (lower (store flags
                    value @ (value_type $F64)
                    address
                    offset))
      (side_effect
       (x64_xmm_movrm (SseOpcode.Movsd) (to_amode flags address offset) value)))

;; Stores of F32X4 vectors.
(rule 1 (lower (store flags
                    value @ (value_type $F32X4)
                    address
                    offset))
      (side_effect
       (x64_xmm_movrm (SseOpcode.Movups) (to_amode flags address offset) value)))

;; Stores of F64X2 vectors.
(rule 1 (lower (store flags
                    value @ (value_type $F64X2)
                    address
                    offset))
      (side_effect
       (x64_xmm_movrm (SseOpcode.Movupd) (to_amode flags address offset) value)))

;; Stores of all other 128-bit vector types with integer lanes.
(rule -1 (lower (store flags
                    value @ (value_type (ty_vec128_int _))
                    address
                    offset))
      (side_effect
       (x64_xmm_movrm (SseOpcode.Movdqu) (to_amode flags address offset) value)))

;; Stores of I128 values: store the two 64-bit halves separately.
(rule 0 (lower (store flags
                    value @ (value_type $I128)
                    address
                    offset))
      (let ((value_reg ValueRegs value)
            (value_lo Gpr (value_regs_get_gpr value_reg 0))
            (value_hi Gpr (value_regs_get_gpr value_reg 1))
            (addr_lo Amode (to_amode flags address offset))
            (addr_hi Amode (amode_offset addr_lo 8)))
      (side_effect
       (side_effect_concat
        (x64_movrm $I64 addr_lo value_lo)
        (x64_movrm $I64 addr_hi value_hi)))))

;; Rules for `load*` + ALU op + `store*` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;

;; Add mem, reg
(rule 3 (lower
       (store flags
              (has_type (ty_32_or_64 ty)
                        (iadd (and
                               (sinkable_load sink)
                               (load flags addr offset))
                              src2))
              addr
              offset))
      (let ((_ RegMemImm (sink_load sink)))
        (side_effect
         (x64_add_mem ty (to_amode flags addr offset) src2))))

;; Add mem, reg with args swapped
(rule 2 (lower
       (store flags
              (has_type (ty_32_or_64 ty)
                        (iadd src2
                              (and
                               (sinkable_load sink)
                               (load flags addr offset))))
              addr
              offset))
      (let ((_ RegMemImm (sink_load sink)))
        (side_effect
         (x64_add_mem ty (to_amode flags addr offset) src2))))

;; Sub mem, reg
(rule 2 (lower
       (store flags
              (has_type (ty_32_or_64 ty)
                        (isub (and
                               (sinkable_load sink)
                               (load flags addr offset))
                              src2))
              addr
              offset))
      (let ((_ RegMemImm (sink_load sink)))
        (side_effect
         (x64_sub_mem ty (to_amode flags addr offset) src2))))

;; And mem, reg
(rule 3 (lower
       (store flags
              (has_type (ty_32_or_64 ty)
                        (band (and
                               (sinkable_load sink)
                               (load flags addr offset))
                              src2))
              addr
              offset))
      (let ((_ RegMemImm (sink_load sink)))
        (side_effect
         (x64_and_mem ty (to_amode flags addr offset) src2))))

;; And mem, reg with args swapped
(rule 2 (lower
       (store flags
              (has_type (ty_32_or_64 ty)
                        (band src2
                              (and
                               (sinkable_load sink)
                               (load flags addr offset))))
              addr
              offset))
      (let ((_ RegMemImm (sink_load sink)))
        (side_effect
         (x64_and_mem ty (to_amode flags addr offset) src2))))

;; Or mem, reg
(rule 3 (lower
       (store flags
              (has_type (ty_32_or_64 ty)
                        (bor (and
                               (sinkable_load sink)
                               (load flags addr offset))
                              src2))
              addr
              offset))
      (let ((_ RegMemImm (sink_load sink)))
        (side_effect
         (x64_or_mem ty (to_amode flags addr offset) src2))))

;; Or mem, reg with args swapped
(rule 2 (lower
       (store flags
              (has_type (ty_32_or_64 ty)
                        (bor src2
                              (and
                               (sinkable_load sink)
                               (load flags addr offset))))
              addr
              offset))
      (let ((_ RegMemImm (sink_load sink)))
        (side_effect
         (x64_or_mem ty (to_amode flags addr offset) src2))))

;; Xor mem, reg
(rule 3 (lower
       (store flags
              (has_type (ty_32_or_64 ty)
                        (bxor (and
                               (sinkable_load sink)
                               (load flags addr offset))
                              src2))
              addr
              offset))
      (let ((_ RegMemImm (sink_load sink)))
        (side_effect
         (x64_xor_mem ty (to_amode flags addr offset) src2))))

;; Xor mem, reg with args swapped
(rule 2 (lower
       (store flags
              (has_type (ty_32_or_64 ty)
                        (bxor src2
                              (and
                               (sinkable_load sink)
                               (load flags addr offset))))
              addr
              offset))
      (let ((_ RegMemImm (sink_load sink)))
        (side_effect
         (x64_xor_mem ty (to_amode flags addr offset) src2))))

;; Rules for `fence` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;

(rule (lower (fence))
      (side_effect (x64_mfence)))

;; Rules for `func_addr` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;

(rule (lower (func_addr (func_ref_data _ extname _)))
      (load_ext_name extname 0))

;; Rules for `symbol_value` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;

(rule (lower (symbol_value (symbol_value_data extname _ offset)))
      (load_ext_name extname offset))

;; Rules for `atomic_load` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;

;; This is a normal load. The x86-TSO memory model provides sufficient
;; sequencing to satisfy the CLIF synchronisation requirements for `AtomicLoad`
;; without the need for any fence instructions.
;;
;; As described in the `atomic_load` documentation, this lowering is only valid
;; for I8, I16, I32, and I64. The sub-64-bit types are zero extended, as with a
;; normal load.
(rule 1 (lower (has_type $I64 (atomic_load flags address)))
      (x64_mov (to_amode flags address (zero_offset))))
(rule (lower (has_type (and (fits_in_32 ty) (ty_int _)) (atomic_load flags address)))
      (x64_movzx (ext_mode (ty_bits_u16 ty) 64) (to_amode flags address (zero_offset))))

;; Rules for `atomic_store` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;

;; This is a normal store followed by an `mfence` instruction. As described in
;; the `atomic_load` documentation, this lowering is only valid for I8, I16,
;; I32, and I64.
(rule (lower (atomic_store flags
                           value @ (value_type (and (fits_in_64 ty) (ty_int _)))
                           address))
      (side_effect (side_effect_concat
       (x64_movrm ty (to_amode flags address (zero_offset)) value)
       (x64_mfence))))

;; Rules for `atomic_cas` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;

(rule (lower (has_type (and (fits_in_64 ty) (ty_int _))
                  (atomic_cas flags address expected replacement)))
      (x64_cmpxchg ty expected replacement (to_amode flags address (zero_offset))))

;; Rules for `atomic_rmw` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;

;; This is a simple, general-case atomic update, based on a loop involving
;; `cmpxchg`.  Note that we could do much better than this in the case where the
;; old value at the location (that is to say, the SSA `Value` computed by this
;; CLIF instruction) is not required.  In that case, we could instead implement
;; this using a single `lock`-prefixed x64 read-modify-write instruction.  Also,
;; even in the case where the old value is required, for the `add` and `sub`
;; cases, we can use the single instruction `lock xadd`.  However, those
;; improvements have been left for another day. TODO: filed as
;; https://github.com/bytecodealliance/wasmtime/issues/2153.

(rule (lower (has_type (and (fits_in_64 ty) (ty_int _))
                  (atomic_rmw flags op address input)))
      (x64_atomic_rmw_seq ty op (to_amode flags address (zero_offset)) input))

;; Rules for `call` and `call_indirect` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;

(rule (lower (call (func_ref_data sig_ref extname dist) inputs))
      (gen_call sig_ref extname dist inputs))

(rule (lower (call_indirect sig_ref val inputs))
      (gen_call_indirect sig_ref val inputs))

;;;; Rules for `get_{frame,stack}_pointer` and `get_return_address` ;;;;;;;;;;;;

(rule (lower (get_frame_pointer))
      (x64_rbp))

(rule (lower (get_stack_pointer))
      (x64_rsp))

(rule (lower (get_return_address))
      (x64_load $I64
                (Amode.ImmReg 8 (x64_rbp) (mem_flags_trusted))
                (ExtKind.None)))

;; Rules for `jump` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;

(rule (lower_branch (jump _) (single_target target))
      (emit_side_effect (jmp_known target)))

;; Rules for `brif` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;

(rule 2 (lower_branch (brif (maybe_uextend (icmp cc a b)) _ _) (two_targets then else))
        (emit_side_effect (jmp_cond_icmp (emit_cmp cc a b) then else)))

(rule 2 (lower_branch (brif (maybe_uextend (fcmp cc a b)) _ _) (two_targets then else))
        (emit_side_effect (jmp_cond_fcmp (emit_fcmp cc a b) then else)))

(rule 1 (lower_branch (brif val @ (value_type $I128) _ _)
                      (two_targets then else))
      (emit_side_effect (jmp_cond_icmp (cmp_zero_i128 (CC.Z) val) then else)))

(rule (lower_branch (brif val @ (value_type (ty_int_bool_or_ref)) _ _)
                    (two_targets then else))
      (emit_side_effect (with_flags_side_effect
                          (cmp_zero_int_bool_ref val)
                          (jmp_cond (CC.NZ) then else))))


;; Compare an I128 value to zero, returning a flags result suitable for making a
;; jump decision. The comparison is implemented as `(hi == 0) && (low == 0)`,
;; and the result can be interpreted as follows
;; * CC.Z indicates that the value was non-zero, as one or both of the halves of
;;   the value were non-zero
;; * CC.NZ indicates that both halves of the value were 0
(decl cmp_zero_i128 (CC ValueRegs) IcmpCondResult)
(rule (cmp_zero_i128 (cc_nz_or_z cc) val)
      (let ((lo Gpr (value_regs_get_gpr val 0))
            (hi Gpr (value_regs_get_gpr val 1))
            (lo_z Gpr (with_flags_reg (x64_cmp (OperandSize.Size64) (RegMemImm.Imm 0) lo)
                                      (x64_setcc (CC.Z))))
            (hi_z Gpr (with_flags_reg (x64_cmp (OperandSize.Size64) (RegMemImm.Imm 0) hi)
                                      (x64_setcc (CC.Z)))))
          (icmp_cond_result (x64_test (OperandSize.Size8) lo_z hi_z) cc)))


(decl cmp_zero_int_bool_ref (Value) ProducesFlags)
(rule (cmp_zero_int_bool_ref val @ (value_type ty))
      (let ((size OperandSize (raw_operand_size_of_type ty))
            (src Gpr val))
        (x64_test size src src)))

;; Rules for `br_table` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;

(rule (lower_branch (br_table idx @ (value_type ty) _) (jump_table_targets default_target jt_targets))
      (let ((size OperandSize (raw_operand_size_of_type ty))
            (jt_size u32 (jump_table_size jt_targets))
            (size_reg Reg (imm ty (u32_as_u64 jt_size)))
            (idx_reg Gpr (extend_to_gpr idx $I64 (ExtendKind.Zero)))
            (clamped_idx Reg (with_flags_reg
              (x64_cmp size size_reg idx_reg)
              (cmove ty (CC.B) idx_reg size_reg))))
      (emit_side_effect (jmp_table_seq ty clamped_idx default_target jt_targets))))

;; Rules for `select_spectre_guard` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;

(rule (lower (select_spectre_guard (icmp cc a b) x y))
      (select_icmp (emit_cmp cc a b) x y))

(rule -1 (lower (has_type ty (select_spectre_guard c @ (value_type (fits_in_64 a_ty)) x y)))
      (let ((size OperandSize (raw_operand_size_of_type a_ty))
            (gpr_c Gpr (put_in_gpr c)))
        (with_flags (x64_test size gpr_c gpr_c) (cmove_from_values ty (CC.NZ) x y))))

(rule -2 (lower (has_type ty (select_spectre_guard c @ (value_type $I128) x y)))
      (let ((cond_result IcmpCondResult (cmp_zero_i128 (CC.Z) c)))
        (select_icmp cond_result x y)))

;; Rules for `fcvt_from_sint` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;

(rule 2 (lower (has_type $F32 (fcvt_from_sint a @ (value_type $I8))))
      (x64_cvtsi2ss $I32 (extend_to_gpr a $I32 (ExtendKind.Sign))))

(rule 2 (lower (has_type $F32 (fcvt_from_sint a @ (value_type $I16))))
      (x64_cvtsi2ss $I32 (extend_to_gpr a $I32 (ExtendKind.Sign))))

(rule 1 (lower (has_type $F32 (fcvt_from_sint a @ (value_type (ty_int (fits_in_64 ty))))))
      (x64_cvtsi2ss ty a))

(rule 2 (lower (has_type $F64 (fcvt_from_sint a @ (value_type $I8))))
      (x64_cvtsi2sd $I32 (extend_to_gpr a $I32 (ExtendKind.Sign))))

(rule 2 (lower (has_type $F64 (fcvt_from_sint a @ (value_type $I16))))
      (x64_cvtsi2sd $I32 (extend_to_gpr a $I32 (ExtendKind.Sign))))

(rule 1 (lower (has_type $F64 (fcvt_from_sint a @ (value_type (ty_int (fits_in_64 ty))))))
      (x64_cvtsi2sd ty a))

(rule 0 (lower (fcvt_from_sint a @ (value_type $I32X4)))
      (x64_cvtdq2ps a))

;; Rules for `fcvt_low_from_sint` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;

(rule (lower (fcvt_low_from_sint a))
      (x64_cvtdq2pd a))

;; Rules for `fcvt_from_uint` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;

(rule 1 (lower (has_type $F32 (fcvt_from_uint val @ (value_type (fits_in_32 (ty_int ty))))))
      (x64_cvtsi2ss $I64 (extend_to_gpr val $I64 (ExtendKind.Zero))))

(rule 1 (lower (has_type $F64 (fcvt_from_uint val @ (value_type (fits_in_32 (ty_int ty))))))
      (x64_cvtsi2sd $I64 (extend_to_gpr val $I64 (ExtendKind.Zero))))

(rule (lower (has_type ty (fcvt_from_uint val @ (value_type $I64))))
      (cvt_u64_to_float_seq ty val))

;; Algorithm uses unpcklps to help create a float that is equivalent
;; 0x1.0p52 + double(src). 0x1.0p52 is unique because at this exponent
;; every value of the mantissa represents a corresponding uint32 number.
;; When we subtract 0x1.0p52 we are left with double(src).
(rule 1 (lower (has_type $F64X2 (fcvt_from_uint (uwiden_low val @ (value_type $I32X4)))))
      (let ((uint_mask XmmMem (fcvt_uint_mask_const))
            (res Xmm (x64_unpcklps val uint_mask))
            (uint_mask_high XmmMem (fcvt_uint_mask_high_const)))
        (x64_subpd res uint_mask_high)))

;; When AVX512VL and AVX512F are available,
;; `fcvt_from_uint` can be lowered to a single instruction.
(rule 2 (lower (has_type (and (avx512vl_enabled $true) (avx512f_enabled $true) $F32X4)
                         (fcvt_from_uint src)))
      (x64_vcvtudq2ps src))

;; Converting packed unsigned integers to packed floats
;; requires a few steps. There is no single instruction
;; lowering for converting unsigned floats but there is for
;; converting packed signed integers to float (cvtdq2ps). In
;; the steps below we isolate the upper half (16 bits) and
;; lower half (16 bits) of each lane and then we convert
;; each half separately using cvtdq2ps meant for signed
;; integers. In order for this to work for the upper half
;; bits we must shift right by 1 (divide by 2) these bits in
;; order to ensure the most significant bit is 0 not signed,
;; and then after the conversion we double the value.
;; Finally we add the converted values where addition will
;; correctly round.
;;
;; Sequence:
;; -> A = 0xffffffff
;; -> Ah = 0xffff0000
;; -> Al = 0x0000ffff
;; -> Convert(Al) // Convert int to float
;; -> Ah = Ah >> 1 // Shift right 1 to assure Ah conversion isn't treated as signed
;; -> Convert(Ah) // Convert .. with no loss of significant digits from previous shift
;; -> Ah = Ah + Ah // Double Ah to account for shift right before the conversion.
;; -> dst = Ah + Al // Add the two floats together
(rule 1 (lower (has_type $F32X4 (fcvt_from_uint val)))
      (let ((a Xmm val)

            ;;  get the low 16 bits
            (a_lo Xmm (x64_pslld a (RegMemImm.Imm 16)))
            (a_lo Xmm (x64_psrld a_lo (RegMemImm.Imm 16)))

            ;; get the high 16 bits
            (a_hi Xmm (x64_psubd a a_lo))

            ;; convert the low 16 bits
            (a_lo Xmm (x64_cvtdq2ps a_lo))

            ;; shift the high bits by 1, convert, and double to get the correct
            ;; value
            (a_hi Xmm (x64_psrld a_hi (RegMemImm.Imm 1)))
            (a_hi Xmm (x64_cvtdq2ps a_hi))
            (a_hi Xmm (x64_addps a_hi a_hi)))

        ;; add together the two converted values
        (x64_addps a_hi a_lo)))

;; Rules for `fcvt_to_uint` and `fcvt_to_sint` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;

(rule (lower (has_type out_ty (fcvt_to_uint val @ (value_type (ty_scalar_float _)))))
      (cvt_float_to_uint_seq out_ty val $false))

(rule (lower (has_type out_ty (fcvt_to_uint_sat val @ (value_type (ty_scalar_float _)))))
      (cvt_float_to_uint_seq out_ty val $true))

(rule (lower (has_type out_ty (fcvt_to_sint val @ (value_type (ty_scalar_float _)))))
      (cvt_float_to_sint_seq out_ty val $false))

(rule (lower (has_type out_ty (fcvt_to_sint_sat val @ (value_type (ty_scalar_float _)))))
      (cvt_float_to_sint_seq out_ty val $true))

;; The x64 backend currently only supports these two type combinations.
(rule 1 (lower (has_type $I32X4 (fcvt_to_sint_sat val @ (value_type $F32X4))))
      (let ((src Xmm val)

            ;; Sets tmp to zero if float is NaN
            (tmp Xmm (x64_cmpps src src (FcmpImm.Equal)))
            (dst Xmm (x64_andps src tmp))

            ;; Sets top bit of tmp if float is positive
            ;; Setting up to set top bit on negative float values
            (tmp Xmm (x64_pxor tmp dst))

            ;; Convert the packed float to packed doubleword.
            (dst Xmm (x64_cvttps2dq dst))

            ;; Set top bit only if < 0
            (tmp Xmm (x64_pand dst tmp))
            (tmp Xmm (x64_psrad tmp (RegMemImm.Imm 31))))

        ;; On overflow 0x80000000 is returned to a lane.
        ;; Below sets positive overflow lanes to 0x7FFFFFFF
        ;; Keeps negative overflow lanes as is.
        (x64_pxor tmp dst)))

;; The algorithm for converting floats to unsigned ints is a little tricky. The
;; complication arises because we are converting from a signed 64-bit int with a positive
;; integer range from 1..INT_MAX (0x1..0x7FFFFFFF) to an unsigned integer with an extended
;; range from (INT_MAX+1)..UINT_MAX. It's this range from (INT_MAX+1)..UINT_MAX
;; (0x80000000..0xFFFFFFFF) that needs to be accounted for as a special case since our
;; conversion instruction (cvttps2dq) only converts as high as INT_MAX (0x7FFFFFFF), but
;; which conveniently setting underflows and overflows (smaller than MIN_INT or larger than
;; MAX_INT) to be INT_MAX+1 (0x80000000). Nothing that the range (INT_MAX+1)..UINT_MAX includes
;; precisely INT_MAX values we can correctly account for and convert every value in this range
;; if we simply subtract INT_MAX+1 before doing the cvttps2dq conversion. After the subtraction
;; every value originally (INT_MAX+1)..UINT_MAX is now the range (0..INT_MAX).
;; After the conversion we add INT_MAX+1 back to this converted value, noting again that
;; values we are trying to account for were already set to INT_MAX+1 during the original conversion.
;; We simply have to create a mask and make sure we are adding together only the lanes that need
;; to be accounted for. Digesting it all the steps then are:
;;
;; Step 1 - Account for NaN and negative floats by setting these src values to zero.
;; Step 2 - Make a copy (tmp1) of the src value since we need to convert twice for
;;          reasons described above.
;; Step 3 - Convert the original src values. This will convert properly all floats up to INT_MAX
;; Step 4 - Subtract INT_MAX from the copy set (tmp1). Note, all zero and negative values are those
;;          values that were originally in the range (0..INT_MAX). This will come in handy during
;;          step 7 when we zero negative lanes.
;; Step 5 - Create a bit mask for tmp1 that will correspond to all lanes originally less than
;;          UINT_MAX that are now less than INT_MAX thanks to the subtraction.
;; Step 6 - Convert the second set of values (tmp1)
;; Step 7 - Prep the converted second set by zeroing out negative lanes (these have already been
;;          converted correctly with the first set) and by setting overflow lanes to 0x7FFFFFFF
;;          as this will allow us to properly saturate overflow lanes when adding to 0x80000000
;; Step 8 - Add the orginal converted src and the converted tmp1 where float values originally less
;;          than and equal to INT_MAX will be unchanged, float values originally between INT_MAX+1 and
;;          UINT_MAX will add together (INT_MAX) + (SRC - INT_MAX), and float values originally
;;          greater than UINT_MAX will be saturated to UINT_MAX (0xFFFFFFFF) after adding (0x8000000 + 0x7FFFFFFF).
;;
;;
;; The table below illustrates the result after each step where it matters for the converted set.
;; Note the original value range (original src set) is the final dst in Step 8:
;;
;; Original src set:
;; | Original Value Range |    Step 1    |         Step 3         |          Step 8           |
;; |  -FLT_MIN..FLT_MAX   | 0.0..FLT_MAX | 0..INT_MAX(w/overflow) | 0..UINT_MAX(w/saturation) |
;;
;; Copied src set (tmp1):
;; |    Step 2    |                  Step 4                  |
;; | 0.0..FLT_MAX | (0.0-(INT_MAX+1))..(FLT_MAX-(INT_MAX+1)) |
;;
;; |                       Step 6                        |                 Step 7                 |
;; | (0-(INT_MAX+1))..(UINT_MAX-(INT_MAX+1))(w/overflow) | ((INT_MAX+1)-(INT_MAX+1))..(INT_MAX+1) |
(rule 1 (lower (has_type $I32X4 (fcvt_to_uint_sat val @ (value_type $F32X4))))
      (let ((src Xmm val)

            ;; Converting to unsigned int so if float src is negative or NaN
            ;; will first set to zero.
            (tmp2 Xmm (xmm_zero $F32X4))
            (dst Xmm (x64_maxps src tmp2))

            ;; Set tmp2 to INT_MAX+1. It is important to note here that after it looks
            ;; like we are only converting INT_MAX (0x7FFFFFFF) but in fact because
            ;; single precision IEEE-754 floats can only accurately represent contingous
            ;; integers up to 2^23 and outside of this range it rounds to the closest
            ;; integer that it can represent. In the case of INT_MAX, this value gets
            ;; represented as 0x4f000000 which is the integer value (INT_MAX+1).
            (tmp2 Xmm (x64_pcmpeqd tmp2 tmp2))
            (tmp2 Xmm (x64_psrld tmp2 (RegMemImm.Imm 1)))
            (tmp2 Xmm (x64_cvtdq2ps tmp2))

            ;; Make a copy of these lanes and then do the first conversion.
            ;; Overflow lanes greater than the maximum allowed signed value will
            ;; set to 0x80000000. Negative and NaN lanes will be 0x0
            (tmp1 Xmm dst)
            (dst Xmm (x64_cvttps2dq dst))

            ;; Set lanes to src - max_signed_int
            (tmp1 Xmm (x64_subps tmp1 tmp2))

            ;; Create mask for all positive lanes to saturate (i.e. greater than
            ;; or equal to the maxmimum allowable unsigned int).
            (tmp2 Xmm (x64_cmpps tmp2 tmp1 (FcmpImm.LessThanOrEqual)))

            ;; Convert those set of lanes that have the max_signed_int factored out.
            (tmp1 Xmm (x64_cvttps2dq tmp1))

            ;; Prepare converted lanes by zeroing negative lanes and prepping lanes
            ;; that have positive overflow (based on the mask) by setting these lanes
            ;; to 0x7FFFFFFF
            (tmp1 Xmm (x64_pxor tmp1 tmp2))
            (tmp2 Xmm (xmm_zero $I32X4))
            (tmp1 Xmm (x64_pmaxsd tmp1 tmp2)))

        ;; Add this second set of converted lanes to the original to properly handle
        ;; values greater than max signed int.
        (x64_paddd tmp1 dst)))

;; Rules for `iadd_pairwise` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;

;; special case for the `i16x8.extadd_pairwise_i8x16_s` wasm instruction
(rule (lower
        (has_type $I16X8 (iadd_pairwise
                           (swiden_low val @ (value_type $I8X16))
                           (swiden_high val))))
      (let ((mul_const Xmm (x64_xmm_load_const $I8X16 (iadd_pairwise_mul_const_16))))
        (x64_pmaddubsw mul_const val)))

;; special case for the `i32x4.extadd_pairwise_i16x8_s` wasm instruction
(rule (lower
        (has_type $I32X4 (iadd_pairwise
                           (swiden_low val @ (value_type $I16X8))
                           (swiden_high val))))
      (let ((mul_const XmmMem (iadd_pairwise_mul_const_32)))
        (x64_pmaddwd val mul_const)))

;; special case for the `i16x8.extadd_pairwise_i8x16_u` wasm instruction
(rule (lower
        (has_type $I16X8 (iadd_pairwise
                           (uwiden_low val @ (value_type $I8X16))
                           (uwiden_high val))))
      (let ((mul_const XmmMem (iadd_pairwise_mul_const_16)))
        (x64_pmaddubsw val mul_const)))

;; special case for the `i32x4.extadd_pairwise_i16x8_u` wasm instruction
(rule (lower
        (has_type $I32X4 (iadd_pairwise
                           (uwiden_low val @ (value_type $I16X8))
                           (uwiden_high val))))
      (let ((xor_const XmmMem (iadd_pairwise_xor_const_32))
            (dst Xmm (x64_pxor val xor_const))

            (madd_const XmmMem (iadd_pairwise_mul_const_32))
            (dst Xmm (x64_pmaddwd dst madd_const))

            (addd_const XmmMem (iadd_pairwise_addd_const_32)))
        (x64_paddd dst addd_const)))

;; special case for the `i32x4.dot_i16x8_s` wasm instruction
(rule (lower
        (has_type $I32X4 (iadd_pairwise
                           (imul (swiden_low x) (swiden_low y))
                           (imul (swiden_high x) (swiden_high y)))))
      (x64_pmaddwd x y))

;; Rules for `swiden_low` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;

(rule (lower (has_type $I16X8 (swiden_low val @ (value_type $I8X16))))
      (x64_pmovsxbw val))

(rule (lower (has_type $I32X4 (swiden_low val @ (value_type $I16X8))))
      (x64_pmovsxwd val))

(rule (lower (has_type $I64X2 (swiden_low val @ (value_type $I32X4))))
      (x64_pmovsxdq val))

;; Rules for `swiden_high` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;

(rule (lower (has_type $I16X8 (swiden_high val @ (value_type $I8X16))))
      (let ((x Xmm val))
        (x64_pmovsxbw (x64_palignr x x 8 (OperandSize.Size32)))))

(rule (lower (has_type $I32X4 (swiden_high val @ (value_type $I16X8))))
      (let ((x Xmm val))
        (x64_pmovsxwd (x64_palignr x x 8 (OperandSize.Size32)))))

(rule (lower (has_type $I64X2 (swiden_high val @ (value_type $I32X4))))
      (x64_pmovsxdq (x64_pshufd val 0xEE)))

;; Rules for `uwiden_low` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;

(rule (lower (has_type $I16X8 (uwiden_low val @ (value_type $I8X16))))
      (x64_pmovzxbw val))

(rule (lower (has_type $I32X4 (uwiden_low val @ (value_type $I16X8))))
      (x64_pmovzxwd val))

(rule (lower (has_type $I64X2 (uwiden_low val @ (value_type $I32X4))))
      (x64_pmovzxdq val))

;; Rules for `uwiden_high` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;

(rule (lower (has_type $I16X8 (uwiden_high val @ (value_type $I8X16))))
      (let ((x Xmm val))
        (x64_pmovzxbw (x64_palignr x x 8 (OperandSize.Size32)))))

(rule (lower (has_type $I32X4 (uwiden_high val @ (value_type $I16X8))))
      (let ((x Xmm val))
        (x64_pmovzxwd (x64_palignr x x 8 (OperandSize.Size32)))))

(rule (lower (has_type $I64X2 (uwiden_high val @ (value_type $I32X4))))
      (x64_pmovzxdq (x64_pshufd val 0xEE)))

;; Rules for `snarrow` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;

(rule (lower (has_type $I8X16 (snarrow a @ (value_type $I16X8) b)))
      (x64_packsswb a b))

(rule (lower (has_type $I16X8 (snarrow a @ (value_type $I32X4) b)))
      (x64_packssdw a b))

;; We're missing a `snarrow` case for $I64X2
;; https://github.com/bytecodealliance/wasmtime/issues/4734

;; This rule is a special case for handling the translation of the wasm op
;; `i32x4.trunc_sat_f64x2_s_zero`. It can be removed once we have an
;; implementation of `snarrow` for `I64X2`.
(rule (lower (has_type $I32X4 (snarrow (has_type $I64X2 (fcvt_to_sint_sat val))
                                       (vconst (u128_from_constant 0)))))
      (let ((a Xmm val)

            ;; y = i32x4.trunc_sat_f64x2_s_zero(x) is lowered to:
            ;; MOVE xmm_tmp, xmm_x
            ;; CMPEQPD xmm_tmp, xmm_x
            ;; MOVE xmm_y, xmm_x
            ;; ANDPS xmm_tmp, [wasm_f64x2_splat(2147483647.0)]
            ;; MINPD xmm_y, xmm_tmp
            ;; CVTTPD2DQ xmm_y, xmm_y

            (tmp1 Xmm (x64_cmppd a a (FcmpImm.Equal)))
            (umax_mask XmmMem (snarrow_umax_mask))

            ;; ANDPD xmm_y, [wasm_f64x2_splat(2147483647.0)]
            (tmp1 Xmm (x64_andps tmp1 umax_mask))
            (dst Xmm (x64_minpd a tmp1)))
        (x64_cvttpd2dq dst)))

;; Rules for `unarrow` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;

(rule (lower (has_type $I8X16 (unarrow a @ (value_type $I16X8) b)))
      (x64_packuswb a b))

(rule (lower (has_type $I16X8 (unarrow a @ (value_type $I32X4) b)))
      (x64_packusdw a b))

;; We're missing a `unarrow` case for $I64X2
;; https://github.com/bytecodealliance/wasmtime/issues/4734

;; Rules for `bitcast` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;

(rule (lower (has_type $I32 (bitcast _ src @ (value_type $F32))))
      (bitcast_xmm_to_gpr $F32 src))

(rule (lower (has_type $F32 (bitcast _ src @ (value_type $I32))))
      (bitcast_gpr_to_xmm $I32 src))

(rule (lower (has_type $I64 (bitcast _ src @ (value_type $F64))))
      (bitcast_xmm_to_gpr $F64 src))

(rule (lower (has_type $F64 (bitcast _ src @ (value_type $I64))))
      (bitcast_gpr_to_xmm $I64 src))

;; Bitcast between types residing in GPR registers is a no-op.
(rule 1 (lower (has_type (is_gpr_type _)
                         (bitcast _ x @ (value_type (is_gpr_type _))))) x)

;; Bitcast between types residing in XMM registers is a no-op.
(rule 2 (lower (has_type (is_xmm_type _)
                         (bitcast _ x @ (value_type (is_xmm_type _))))) x)

;; Rules for `fcopysign` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;

(rule (lower (has_type $F32 (fcopysign a @ (value_type $F32) b)))
      (let ((sign_bit Xmm (imm $F32 0x80000000)))
        (x64_orps
          (x64_andnps sign_bit a)
          (x64_andps sign_bit b))))

(rule (lower (has_type $F64 (fcopysign a @ (value_type $F64) b)))
      (let ((sign_bit Xmm (imm $F64 0x8000000000000000)))
        (x64_orpd
          (x64_andnpd sign_bit a)
          (x64_andpd sign_bit b))))

;; Rules for `ceil` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;

(rule (lower (has_type (use_sse41 $true) (ceil a @ (value_type $F32))))
      (x64_roundss a (RoundImm.RoundUp)))

(rule (lower (has_type (use_sse41 $false) (ceil a @ (value_type $F32))))
      (libcall_1 (LibCall.CeilF32) a))

(rule (lower (has_type (use_sse41 $true) (ceil a @ (value_type $F64))))
      (x64_roundsd a (RoundImm.RoundUp)))

(rule (lower (has_type (use_sse41 $false) (ceil a @ (value_type $F64))))
      (libcall_1 (LibCall.CeilF64) a))

(rule (lower (has_type (use_sse41 $true) (ceil a @ (value_type $F32X4))))
      (x64_roundps a (RoundImm.RoundUp)))

(rule (lower (has_type (use_sse41 $true) (ceil a @ (value_type $F64X2))))
      (x64_roundpd a (RoundImm.RoundUp)))

;; Rules for `floor` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;

(rule (lower (has_type (use_sse41 $true) (floor a @ (value_type $F32))))
      (x64_roundss a (RoundImm.RoundDown)))

(rule (lower (has_type (use_sse41 $false) (floor a @ (value_type $F32))))
      (libcall_1 (LibCall.FloorF32) a))

(rule (lower (has_type (use_sse41 $true) (floor a @ (value_type $F64))))
      (x64_roundsd a (RoundImm.RoundDown)))

(rule (lower (has_type (use_sse41 $false) (floor a @ (value_type $F64))))
      (libcall_1 (LibCall.FloorF64) a))

(rule (lower (has_type (use_sse41 $true) (floor a @ (value_type $F32X4))))
      (x64_roundps a (RoundImm.RoundDown)))

(rule (lower (has_type (use_sse41 $true) (floor a @ (value_type $F64X2))))
      (x64_roundpd a (RoundImm.RoundDown)))

;; Rules for `nearest` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;

(rule (lower (has_type (use_sse41 $true) (nearest a @ (value_type $F32))))
      (x64_roundss a (RoundImm.RoundNearest)))

(rule (lower (has_type (use_sse41 $false) (nearest a @ (value_type $F32))))
      (libcall_1 (LibCall.NearestF32) a))

(rule (lower (has_type (use_sse41 $true) (nearest a @ (value_type $F64))))
      (x64_roundsd a (RoundImm.RoundNearest)))

(rule (lower (has_type (use_sse41 $false) (nearest a @ (value_type $F64))))
      (libcall_1 (LibCall.NearestF64) a))

(rule (lower (has_type (use_sse41 $true) (nearest a @ (value_type $F32X4))))
      (x64_roundps a (RoundImm.RoundNearest)))

(rule (lower (has_type (use_sse41 $true) (nearest a @ (value_type $F64X2))))
      (x64_roundpd a (RoundImm.RoundNearest)))

;; Rules for `trunc` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;

(rule (lower (has_type (use_sse41 $true) (trunc a @ (value_type $F32))))
      (x64_roundss a (RoundImm.RoundZero)))

(rule (lower (has_type (use_sse41 $false) (trunc a @ (value_type $F32))))
      (libcall_1 (LibCall.TruncF32) a))

(rule (lower (has_type (use_sse41 $true) (trunc a @ (value_type $F64))))
      (x64_roundsd a (RoundImm.RoundZero)))

(rule (lower (has_type (use_sse41 $false) (trunc a @ (value_type $F64))))
      (libcall_1 (LibCall.TruncF64) a))

(rule (lower (has_type (use_sse41 $true) (trunc a @ (value_type $F32X4))))
      (x64_roundps a (RoundImm.RoundZero)))

(rule (lower (has_type (use_sse41 $true) (trunc a @ (value_type $F64X2))))
      (x64_roundpd a (RoundImm.RoundZero)))

;; Rules for `stack_addr` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;

(rule (lower (stack_addr stack_slot offset))
      (stack_addr_impl stack_slot offset))

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

(rule (lower (udiv a @ (value_type ty) b))
      (div_or_rem (DivOrRemKind.UnsignedDiv) a b))

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

(rule (lower (sdiv a @ (value_type ty) b))
      (div_or_rem (DivOrRemKind.SignedDiv) a b))

;; Rules for `urem` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;

(rule (lower (urem a @ (value_type ty) b))
      (div_or_rem (DivOrRemKind.UnsignedRem) a b))

;; Rules for `srem` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;

(rule (lower (srem a @ (value_type ty) b))
      (div_or_rem (DivOrRemKind.SignedRem) a b))

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

(rule (lower (umulhi a @ (value_type $I16) b))
      (let ((res ValueRegs (mul_hi $I16 $false a b))
            (hi Gpr (value_regs_get_gpr res 1)))
        hi))

(rule (lower (umulhi a @ (value_type $I32) b))
      (let ((res ValueRegs (mul_hi $I32 $false a b))
            (hi Gpr (value_regs_get_gpr res 1)))
        hi))

(rule (lower (umulhi a @ (value_type $I64) b))
      (let ((res ValueRegs (mul_hi $I64 $false a b))
            (hi Gpr (value_regs_get_gpr res 1)))
        hi))

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

(rule (lower (smulhi a @ (value_type $I16) b))
      (let ((res ValueRegs (mul_hi $I16 $true a b))
            (hi Gpr (value_regs_get_gpr res 1)))
        hi))

(rule (lower (smulhi a @ (value_type $I32) b))
      (let ((res ValueRegs (mul_hi $I32 $true a b))
            (hi Gpr (value_regs_get_gpr res 1)))
        hi))

(rule (lower (smulhi a @ (value_type $I64) b))
      (let ((res ValueRegs (mul_hi $I64 $true a b))
            (hi Gpr (value_regs_get_gpr res 1)))
        hi))

;; Rules for `get_pinned_reg` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;

(rule (lower (get_pinned_reg))
      (read_pinned_gpr))

;; Rules for `set_pinned_reg` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;

(rule (lower (set_pinned_reg a @ (value_type ty)))
      (side_effect (write_pinned_gpr a)))

;; Rules for `vconst` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;

(rule (lower (has_type ty (vconst const)))
      ;; TODO use Inst::gen_constant() instead.
      (x64_xmm_load_const ty (const_to_vconst const)))

;; Rules for `shuffle` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;

;; If `lhs` and `rhs` are the same we can use a single PSHUFB to shuffle the XMM
;; register. We statically build `constructed_mask` to zero out any unknown lane
;; indices (may not be completely necessary: verification could fail incorrect
;; mask values) and fix the indexes to all point to the `dst` vector.
(rule 3 (lower (shuffle a a (vec_mask_from_immediate mask)))
      (x64_pshufb a (shuffle_0_31_mask mask)))

;; For the case where the shuffle mask contains out-of-bounds values (values
;; greater than 31) we must mask off those resulting values in the result of
;; `vpermi2b`.
(rule 2 (lower (has_type (and (avx512vl_enabled $true) (avx512vbmi_enabled $true))
                         (shuffle a b (vec_mask_from_immediate
                                        (perm_from_mask_with_zeros mask zeros)))))
      (x64_andps (x64_vpermi2b b a (x64_xmm_load_const $I8X16 mask)) zeros))

;; However, if the shuffle mask contains no out-of-bounds values, we can use
;; `vpermi2b` without any masking.
(rule 1 (lower (has_type (and (avx512vl_enabled $true) (avx512vbmi_enabled $true))
                       (shuffle a b (vec_mask_from_immediate mask))))
      (x64_vpermi2b b a (x64_xmm_load_const $I8X16 (perm_from_mask mask))))

;; If `lhs` and `rhs` are different, we must shuffle each separately and then OR
;; them together. This is necessary due to PSHUFB semantics. As in the case
;; above, we build the `constructed_mask` for each case statically.
(rule (lower (shuffle a b (vec_mask_from_immediate mask)))
      (x64_por
        (x64_pshufb a (shuffle_0_15_mask mask))
        (x64_pshufb b (shuffle_16_31_mask mask))))

;; Rules for `swizzle` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;

;; SIMD swizzle; the following inefficient implementation is due to the Wasm
;; SIMD spec requiring mask indexes greater than 15 to have the same semantics
;; as a 0 index. For the spec discussion, see
;; https://github.com/WebAssembly/simd/issues/93. The CLIF semantics match the
;; Wasm SIMD semantics for this instruction. The instruction format maps to
;; variables like: %dst = swizzle %src, %mask
(rule (lower (swizzle src mask))
      (let ((mask Xmm (x64_paddusb mask (swizzle_zero_mask))))
        (x64_pshufb src mask)))

;; Rules for `extractlane` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;

;; Remove the extractlane instruction, leaving the float where it is. The upper
;; bits will remain unchanged; for correctness, this relies on Cranelift type
;; checking to avoid using those bits.
(rule 2 (lower (has_type (ty_scalar_float _) (extractlane val (u8_from_uimm8 0))))
      val)

;; Cases 2-4 for an F32X4
(rule 1 (lower (has_type $F32 (extractlane val @ (value_type (ty_vec128 ty))
                                         (u8_from_uimm8 lane))))
      (x64_pshufd val lane))

;; This is the only remaining case for F64X2
(rule 1 (lower (has_type $F64 (extractlane val @ (value_type (ty_vec128 ty))
                                         (u8_from_uimm8 1))))
      ;; 0xee == 0b11_10_11_10
      (x64_pshufd val 0xee))

(rule 0 (lower (extractlane val @ (value_type ty @ (multi_lane 8 16)) (u8_from_uimm8 lane)))
      (x64_pextrb val lane))

(rule 0 (lower (extractlane val @ (value_type ty @ (multi_lane 16 8)) (u8_from_uimm8 lane)))
      (x64_pextrw val lane))

(rule 0 (lower (extractlane val @ (value_type ty @ (multi_lane 32 4)) (u8_from_uimm8 lane)))
      (x64_pextrd val lane))

(rule 0 (lower (extractlane val @ (value_type ty @ (multi_lane 64 2)) (u8_from_uimm8 lane)))
      (x64_pextrq val lane))

;; Rules for `scalar_to_vector` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;

;; Case 1: when moving a scalar float, we simply move from one XMM register
;; to another, expecting the register allocator to elide this. Here we
;; assume that the upper bits of a scalar float have not been munged with
;; (the same assumption the old backend makes).
(rule 1 (lower (scalar_to_vector src @ (value_type (ty_scalar_float _))))
      src)

;; Case 2: when moving a scalar value of any other type, use MOVD to zero
;; the upper lanes.
(rule (lower (scalar_to_vector src @ (value_type ty)))
      (bitcast_gpr_to_xmm ty src))

;; Case 3: when presented with `load + scalar_to_vector`, coalesce into a single
;; MOVSS/MOVSD instruction.
(rule 2 (lower (scalar_to_vector (and (sinkable_load src) (value_type (ty_32 _)))))
      (x64_movss_load src))
(rule 3 (lower (scalar_to_vector (and (sinkable_load src) (value_type (ty_64 _)))))
      (x64_movsd_load src))

;; Rules for `splat` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;

(rule (lower (has_type (multi_lane 8 16) (splat src)))
      (let ((vec Xmm (vec_insert_lane $I8X16 (xmm_uninit_value) src 0))
            (zeros Xmm (xmm_zero $I8X16)))
        ;; Shuffle the lowest byte lane to all other lanes.
        (x64_pshufb vec zeros)))

(rule (lower (has_type (multi_lane 16 8) (splat src)))
      (let (;; Force the input into a register so that we don't create a
            ;; VCodeConstant.
            (src RegMem (RegMem.Reg src))
            (vec Xmm (vec_insert_lane $I16X8 (xmm_uninit_value) src 0))
            (vec Xmm (vec_insert_lane $I16X8 vec src 1)))
        ;; Shuffle the lowest two lanes to all other lanes.
        (x64_pshufd vec 0)))

(rule 1 (lower (has_type (multi_lane 32 4) (splat src @ (value_type (ty_scalar_float _)))))
      (lower_splat_32x4 $F32X4 src))

(rule (lower (has_type (multi_lane 32 4) (splat src)))
      (lower_splat_32x4 $I32X4 src))

(decl lower_splat_32x4 (Type Value) Xmm)
(rule (lower_splat_32x4 ty src)
      (let ((src RegMem src)
            (vec Xmm (vec_insert_lane ty (xmm_uninit_value) src 0)))
        ;; Shuffle the lowest lane to all other lanes.
        (x64_pshufd vec 0)))

(rule 1 (lower (has_type (multi_lane 64 2) (splat src @ (value_type (ty_scalar_float _)))))
      (lower_splat_64x2 $F64X2 src))

(rule (lower (has_type (multi_lane 64 2) (splat src)))
      (lower_splat_64x2 $I64X2 src))

(decl lower_splat_64x2 (Type Value) Xmm)
(rule (lower_splat_64x2 ty src)
      (let (;; Force the input into a register so that we don't create a
            ;; VCodeConstant.
            (src RegMem (RegMem.Reg src))
            (vec Xmm (vec_insert_lane ty (xmm_uninit_value) src 0)))
        (vec_insert_lane ty vec src 1)))

;; Rules for `vany_true` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;

(rule (lower (vany_true val))
      (let ((val Xmm val))
        (with_flags (x64_ptest val val) (x64_setcc (CC.NZ)))))

;; Rules for `vall_true` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;

(rule (lower (vall_true val @ (value_type ty)))
      (let ((src Xmm val)
            (zeros Xmm (xmm_zero ty))
            (cmp Xmm (x64_pcmpeq (vec_int_type ty) src zeros)))
        (with_flags (x64_ptest cmp cmp) (x64_setcc (CC.Z)))))

;; Rules for `vhigh_bits` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;

;; The Intel specification allows using both 32-bit and 64-bit GPRs as
;; destination for the "move mask" instructions. This is controlled by the REX.R
;; bit: "In 64-bit mode, the instruction can access additional registers when
;; used with a REX.R prefix. The default operand size is 64-bit in 64-bit mode"
;; (PMOVMSKB in IA Software Development Manual, vol. 2). This being the case, we
;; will always clear REX.W since its use is unnecessary (`OperandSize` is used
;; for setting/clearing REX.W) as we need at most 16 bits of output for
;; `vhigh_bits`.

(rule (lower (vhigh_bits val @ (value_type (multi_lane 8 16))))
      (x64_pmovmskb (OperandSize.Size32) val))

(rule (lower (vhigh_bits val @ (value_type (multi_lane 32 4))))
      (x64_movmskps (OperandSize.Size32) val))

(rule (lower (vhigh_bits val @ (value_type (multi_lane 64 2))))
      (x64_movmskpd (OperandSize.Size32) val))

;; There is no x86 instruction for extracting the high bit of 16-bit lanes so
;; here we:
;; - duplicate the 16-bit lanes of `src` into 8-bit lanes:
;;     PACKSSWB([x1, x2, ...], [x1, x2, ...]) = [x1', x2', ..., x1', x2', ...]
;; - use PMOVMSKB to gather the high bits; now we have duplicates, though
;; - shift away the bottom 8 high bits to remove the duplicates.
(rule (lower (vhigh_bits val @ (value_type (multi_lane 16 8))))
      (let ((src Xmm val)
            (tmp Xmm (x64_packsswb src src))
            (tmp Gpr (x64_pmovmskb (OperandSize.Size32) tmp)))
        (x64_shr $I64 tmp (Imm8Reg.Imm8 8))))

;; Rules for `iconcat` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;

(rule (lower (iconcat lo @ (value_type $I64) hi))
      (value_regs lo hi))

;; Rules for `isplit` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;

(rule (lower (isplit val @ (value_type $I128)))
      (let ((regs ValueRegs val)
            (lo Reg (value_regs_get regs 0))
            (hi Reg (value_regs_get regs 1)))
        (output_pair lo hi)))

;; Rules for `tls_value` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;

(rule (lower (has_type (tls_model (TlsModel.ElfGd)) (tls_value (symbol_value_data name _ _))))
      (elf_tls_get_addr name))

(rule (lower (has_type (tls_model (TlsModel.Macho)) (tls_value (symbol_value_data name _ _))))
      (macho_tls_get_addr name))

(rule (lower (has_type (tls_model (TlsModel.Coff)) (tls_value (symbol_value_data name _ _))))
      (coff_tls_get_addr name))

;; Rules for `sqmul_round_sat` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;

(rule (lower (sqmul_round_sat qx @ (value_type $I16X8) qy))
      (let ((src1 Xmm qx)
            (src2 Xmm qy)

            (mask XmmMem (sqmul_round_sat_mask))
            (dst Xmm (x64_pmulhrsw src1 src2))
            (cmp Xmm (x64_pcmpeqw dst mask)))
        (x64_pxor dst cmp)))

;; Rules for `sqmul_round_sat` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;

;; TODO: currently we only lower a special case of `uunarrow` needed to support
;; the translation of wasm's i32x4.trunc_sat_f64x2_u_zero operation.
;; https://github.com/bytecodealliance/wasmtime/issues/4791
;;
;; y = i32x4.trunc_sat_f64x2_u_zero(x) is lowered to:
;; MOVAPD xmm_y, xmm_x
;; XORPD xmm_tmp, xmm_tmp
;; MAXPD xmm_y, xmm_tmp
;; MINPD xmm_y, [wasm_f64x2_splat(4294967295.0)]
;; ROUNDPD xmm_y, xmm_y, 0x0B
;; ADDPD xmm_y, [wasm_f64x2_splat(0x1.0p+52)]
;; SHUFPS xmm_y, xmm_xmp, 0x88
(rule (lower (uunarrow (fcvt_to_uint_sat src @ (value_type $F64X2))
                       (vconst (u128_from_constant 0))))
      (let ((src Xmm src)

            ;; MOVAPD xmm_y, xmm_x
            ;; XORPD xmm_tmp, xmm_tmp
            (zeros Xmm (xmm_zero $F64X2))
            (dst Xmm (x64_maxpd src zeros))

            (umax_mask XmmMem (uunarrow_umax_mask))

            ;; MINPD xmm_y, [wasm_f64x2_splat(4294967295.0)]
            (dst Xmm (x64_minpd dst umax_mask))

            ;; ROUNDPD xmm_y, xmm_y, 0x0B
            (dst Xmm (x64_roundpd dst (RoundImm.RoundZero)))

            ;; ADDPD xmm_y, [wasm_f64x2_splat(0x1.0p+52)]
            (uint_mask XmmMem (uunarrow_uint_mask))
            (dst Xmm (x64_addpd dst uint_mask)))

        ;; SHUFPS xmm_y, xmm_xmp, 0x88
        (x64_shufps dst zeros 0x88)))

;; Rules for `nop` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;

(rule (lower (nop))
      (invalid_reg))